diff --git a/.gitattributes b/.gitattributes index 29e566bcb125d03948762e3b798dc121a5210197..12991440c78741dfc5a4d0745c1e0c8cc3239157 100644 --- a/.gitattributes +++ b/.gitattributes @@ -3557,3 +3557,66 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text 2024/GPS-Gaussian_[[:space:]]Generalizable[[:space:]]Pixel-wise[[:space:]]3D[[:space:]]Gaussian[[:space:]]Splatting[[:space:]]for[[:space:]]Real-time[[:space:]]Human[[:space:]]Novel[[:space:]]View[[:space:]]Synthesis/d80f2432-62bb-4d59-a543-d7a12c972968_origin.pdf filter=lfs diff=lfs merge=lfs -text 2024/GPT-4V(ision)[[:space:]]is[[:space:]]a[[:space:]]Human-Aligned[[:space:]]Evaluator[[:space:]]for[[:space:]]Text-to-3D[[:space:]]Generation/a3660d3f-9c59-4d9f-b7f2-795fb06ae9a5_origin.pdf filter=lfs diff=lfs merge=lfs -text 2024/GPT4Point_[[:space:]]A[[:space:]]Unified[[:space:]]Framework[[:space:]]for[[:space:]]Point-Language[[:space:]]Understanding[[:space:]]and[[:space:]]Generation/f8542736-8911-4c2d-be22-bd501d54c8bd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/GRAM_[[:space:]]Global[[:space:]]Reasoning[[:space:]]for[[:space:]]Multi-Page[[:space:]]VQA/bd3538ed-fd76-4274-b7c4-f2ff80c7e1bb_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/GROUNDHOG_[[:space:]]Grounding[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]to[[:space:]]Holistic[[:space:]]Segmentation/bdfcf2de-41e8-4d27-af7d-65821e6b79f8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/GS-IR_[[:space:]]3D[[:space:]]Gaussian[[:space:]]Splatting[[:space:]]for[[:space:]]Inverse[[:space:]]Rendering/d0b63333-3b55-4d7f-93c1-cfbd4adae3a1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/GS-SLAM_[[:space:]]Dense[[:space:]]Visual[[:space:]]SLAM[[:space:]]with[[:space:]]3D[[:space:]]Gaussian[[:space:]]Splatting/93048e35-cf42-49e6-8897-d6803305631d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/GSNeRF_[[:space:]]Generalizable[[:space:]]Semantic[[:space:]]Neural[[:space:]]Radiance[[:space:]]Fields[[:space:]]with[[:space:]]Enhanced[[:space:]]3D[[:space:]]Scene[[:space:]]Understanding/e670ccdd-f8c3-4912-9d6e-b86fa014dde2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/GSVA_[[:space:]]Generalized[[:space:]]Segmentation[[:space:]]via[[:space:]]Multimodal[[:space:]]Large[[:space:]]Language[[:space:]]Models/12e5e8f5-7030-48a1-9213-1240824636f7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/G^3-LQ_[[:space:]]Marrying[[:space:]]Hyperbolic[[:space:]]Alignment[[:space:]]with[[:space:]]Explicit[[:space:]]Semantic-Geometric[[:space:]]Modeling[[:space:]]for[[:space:]]3D[[:space:]]Visual[[:space:]]Grounding/7e2af7bd-e579-4e10-8895-d1ef28fe0cc7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Garment[[:space:]]Recovery[[:space:]]with[[:space:]]Shape[[:space:]]and[[:space:]]Deformation[[:space:]]Priors/83e37ee5-1c15-40f2-ac5d-8e3141ed428d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Gated[[:space:]]Fields_[[:space:]]Learning[[:space:]]Scene[[:space:]]Reconstruction[[:space:]]from[[:space:]]Gated[[:space:]]Videos/babd13ae-5bce-44f5-bd0f-203ff88804c8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/GauHuman_[[:space:]]Articulated[[:space:]]Gaussian[[:space:]]Splatting[[:space:]]from[[:space:]]Monocular[[:space:]]Human[[:space:]]Videos/8c2dc361-8082-42c7-a19c-2ac9cac55b4b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Gaussian[[:space:]]Head[[:space:]]Avatar_[[:space:]]Ultra[[:space:]]High-fidelity[[:space:]]Head[[:space:]]Avatar[[:space:]]via[[:space:]]Dynamic[[:space:]]Gaussians/24a06516-83b5-41f4-8bb5-a4da984c01e1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Gaussian[[:space:]]Shading_[[:space:]]Provable[[:space:]]Performance-Lossless[[:space:]]Image[[:space:]]Watermarking[[:space:]]for[[:space:]]Diffusion[[:space:]]Models/cf093428-bde9-4e71-b54b-2506850cecda_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Gaussian[[:space:]]Shadow[[:space:]]Casting[[:space:]]for[[:space:]]Neural[[:space:]]Characters/e6fedb1c-caf9-484c-be44-7dbe074386f9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Gaussian[[:space:]]Shell[[:space:]]Maps[[:space:]]for[[:space:]]Efficient[[:space:]]3D[[:space:]]Human[[:space:]]Generation/c6318955-2d79-4389-9712-ec90ab24e73b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Gaussian[[:space:]]Splatting[[:space:]]SLAM/32db8396-043e-45f8-8764-d01308949b32_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Gaussian-Flow_[[:space:]]4D[[:space:]]Reconstruction[[:space:]]with[[:space:]]Dynamic[[:space:]]3D[[:space:]]Gaussian[[:space:]]Particle/3e2469b7-4042-4a31-bb56-8ba156724067_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/GaussianAvatar_[[:space:]]Towards[[:space:]]Realistic[[:space:]]Human[[:space:]]Avatar[[:space:]]Modeling[[:space:]]from[[:space:]]a[[:space:]]Single[[:space:]]Video[[:space:]]via[[:space:]]Animatable[[:space:]]3D[[:space:]]Gaussians/30b3a92a-861e-47e6-8246-2e90d1765e38_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/GaussianAvatars_[[:space:]]Photorealistic[[:space:]]Head[[:space:]]Avatars[[:space:]]with[[:space:]]Rigged[[:space:]]3D[[:space:]]Gaussians/41411f0e-3b8e-40c3-bbbd-d16b793f1be0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/GaussianDreamer_[[:space:]]Fast[[:space:]]Generation[[:space:]]from[[:space:]]Text[[:space:]]to[[:space:]]3D[[:space:]]Gaussians[[:space:]]by[[:space:]]Bridging[[:space:]]2D[[:space:]]and[[:space:]]3D[[:space:]]Diffusion[[:space:]]Models/68c13cd6-fa8d-4333-a8a0-12d60a7c6cac_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/GaussianEditor_[[:space:]]Editing[[:space:]]3D[[:space:]]Gaussians[[:space:]]Delicately[[:space:]]with[[:space:]]Text[[:space:]]Instructions/22c39c85-c3be-4622-b7a4-f891bcbdf14e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/GaussianEditor_[[:space:]]Swift[[:space:]]and[[:space:]]Controllable[[:space:]]3D[[:space:]]Editing[[:space:]]with[[:space:]]Gaussian[[:space:]]Splatting/f458e425-2ba9-468a-94e0-95be47f99560_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/GaussianShader_[[:space:]]3D[[:space:]]Gaussian[[:space:]]Splatting[[:space:]]with[[:space:]]Shading[[:space:]]Functions[[:space:]]for[[:space:]]Reflective[[:space:]]Surfaces/e3ac7488-6672-4bab-abdc-38e188c8adc6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Gear-NeRF_[[:space:]]Free-Viewpoint[[:space:]]Rendering[[:space:]]and[[:space:]]Tracking[[:space:]]with[[:space:]]Motion-aware[[:space:]]Spatio-Temporal[[:space:]]Sampling/8ba31972-a909-4d60-aef6-d492fd479dbb_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/GenFlow_[[:space:]]Generalizable[[:space:]]Recurrent[[:space:]]Flow[[:space:]]for[[:space:]]6D[[:space:]]Pose[[:space:]]Refinement[[:space:]]of[[:space:]]Novel[[:space:]]Objects/ff0e87f9-2cf2-45f0-8b79-ad267112b5a6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/GenH2R_[[:space:]]Learning[[:space:]]Generalizable[[:space:]]Human-to-Robot[[:space:]]Handover[[:space:]]via[[:space:]]Scalable[[:space:]]Simulation[[:space:]]Demonstration[[:space:]]and[[:space:]]Imitation/40314ff6-3af8-45b8-b8fc-6f9ebdc7d993_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/GenN2N_[[:space:]]Generative[[:space:]]NeRF2NeRF[[:space:]]Translation/720fe634-ca22-4125-bab8-c49c422e71c2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/GenNBV_[[:space:]]Generalizable[[:space:]]Next-Best-View[[:space:]]Policy[[:space:]]for[[:space:]]Active[[:space:]]3D[[:space:]]Reconstruction/171b2e0e-83b0-4811-a381-9c85825e51ab_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/GenTron_[[:space:]]Diffusion[[:space:]]Transformers[[:space:]]for[[:space:]]Image[[:space:]]and[[:space:]]Video[[:space:]]Generation/2bb6a697-7387-4091-9b21-cb7c548eb341_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/GenZI_[[:space:]]Zero-Shot[[:space:]]3D[[:space:]]Human-Scene[[:space:]]Interaction[[:space:]]Generation/3823a91d-f3af-49d1-a8df-4cb9a1c22737_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/GeneAvatar_[[:space:]]Generic[[:space:]]Expression-Aware[[:space:]]Volumetric[[:space:]]Head[[:space:]]Avatar[[:space:]]Editing[[:space:]]from[[:space:]]a[[:space:]]Single[[:space:]]Image/2d50005a-8c8d-4327-bf7a-e48503a1d824_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/General[[:space:]]Object[[:space:]]Foundation[[:space:]]Model[[:space:]]for[[:space:]]Images[[:space:]]and[[:space:]]Videos[[:space:]]at[[:space:]]Scale/0bc99999-be81-4577-bff4-eaa67b1ee51f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/General[[:space:]]Point[[:space:]]Model[[:space:]]Pretraining[[:space:]]with[[:space:]]Autoencoding[[:space:]]and[[:space:]]Autoregressive/a64a8093-c9f6-4338-aa51-464da1052ad9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Generalizable[[:space:]]Face[[:space:]]Landmarking[[:space:]]Guided[[:space:]]by[[:space:]]Conditional[[:space:]]Face[[:space:]]Warping/33b95235-e8cb-4b6b-a79e-5de2b0524965_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Generalizable[[:space:]]Novel-View[[:space:]]Synthesis[[:space:]]using[[:space:]]a[[:space:]]Stereo[[:space:]]Camera/516ec60a-018b-4106-92f4-813c127e6028_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Generalizable[[:space:]]Whole[[:space:]]Slide[[:space:]]Image[[:space:]]Classification[[:space:]]with[[:space:]]Fine-Grained[[:space:]]Visual-Semantic[[:space:]]Interaction/c4c5691b-0982-48b4-b41b-bc14bcf8455f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Generalized[[:space:]]Event[[:space:]]Cameras/e213eaf6-87b7-49bd-b232-571683e80e75_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Generalized[[:space:]]Large-Scale[[:space:]]Data[[:space:]]Condensation[[:space:]]via[[:space:]]Various[[:space:]]Backbone[[:space:]]and[[:space:]]Statistical[[:space:]]Matching/1c0e887f-135a-4bea-8bad-3d161e8c65c8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Generalized[[:space:]]Predictive[[:space:]]Model[[:space:]]for[[:space:]]Autonomous[[:space:]]Driving/554856ec-712f-446a-b09c-0cce647f2af6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Generalizing[[:space:]]6-DoF[[:space:]]Grasp[[:space:]]Detection[[:space:]]via[[:space:]]Domain[[:space:]]Prior[[:space:]]Knowledge/65c67c8c-a659-4259-9daf-ad24fd3765f3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Generate[[:space:]]Like[[:space:]]Experts_[[:space:]]Multi-Stage[[:space:]]Font[[:space:]]Generation[[:space:]]by[[:space:]]Incorporating[[:space:]]Font[[:space:]]Transfer[[:space:]]Process[[:space:]]into[[:space:]]Diffusion[[:space:]]Models/bfb6c4d4-ce31-4c6d-ba23-750f3d411ecc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Generate[[:space:]]Subgoal[[:space:]]Images[[:space:]]before[[:space:]]Act_[[:space:]]Unlocking[[:space:]]the[[:space:]]Chain-of-Thought[[:space:]]Reasoning[[:space:]]in[[:space:]]Diffusion[[:space:]]Model[[:space:]]for[[:space:]]Robot[[:space:]]Manipulation[[:space:]]with[[:space:]]Multimodal[[:space:]]Prompts/d0cc6546-2994-4121-92a3-5e7ff79756f6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Generating[[:space:]]Content[[:space:]]for[[:space:]]HDR[[:space:]]Deghosting[[:space:]]from[[:space:]]Frequency[[:space:]]View/2b6f8409-576d-47cf-ab70-c1b4b9e76809_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Generating[[:space:]]Enhanced[[:space:]]Negatives[[:space:]]for[[:space:]]Training[[:space:]]Language-Based[[:space:]]Object[[:space:]]Detectors/2970baed-7bcf-45c6-83b6-e79faa1f9b77_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Generating[[:space:]]Handwritten[[:space:]]Mathematical[[:space:]]Expressions[[:space:]]From[[:space:]]Symbol[[:space:]]Graphs_[[:space:]]An[[:space:]]End-to-End[[:space:]]Pipeline/f14645bb-d996-46da-9c4b-81a14daf5714_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Generating[[:space:]]Human[[:space:]]Motion[[:space:]]in[[:space:]]3D[[:space:]]Scenes[[:space:]]from[[:space:]]Text[[:space:]]Descriptions/507f8570-ad9c-455e-838f-8fa1bda75b31_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Generating[[:space:]]Illustrated[[:space:]]Instructions/dd71c86a-390f-4c46-b6ae-0a13f2787fa0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Generating[[:space:]]Non-Stationary[[:space:]]Textures[[:space:]]using[[:space:]]Self-Rectification/3ddc2997-ec2e-480d-b380-c7012481f817_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Generative[[:space:]]3D[[:space:]]Part[[:space:]]Assembly[[:space:]]via[[:space:]]Part-Whole-Hierarchy[[:space:]]Message[[:space:]]Passing/40f117f5-c539-42c1-b970-aab5fe53da07_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Generative[[:space:]]Image[[:space:]]Dynamics/34d8fada-6dfb-4124-8738-6ec5f099992d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Generative[[:space:]]Latent[[:space:]]Coding[[:space:]]for[[:space:]]Ultra-Low[[:space:]]Bitrate[[:space:]]Image[[:space:]]Compression/a71ce275-c18c-4862-b210-a735f3bb6944_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Generative[[:space:]]Multi-modal[[:space:]]Models[[:space:]]are[[:space:]]Good[[:space:]]Class[[:space:]]Incremental[[:space:]]Learners/a19d9ccd-1dff-4a86-83d4-a672741b4e0a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Generative[[:space:]]Multimodal[[:space:]]Models[[:space:]]are[[:space:]]In-Context[[:space:]]Learners/1a0461d9-1d47-4e83-8d07-3f1f380c508b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Generative[[:space:]]Powers[[:space:]]of[[:space:]]Ten/2ccba12d-1f5a-40cf-8686-79a45cb775c2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Generative[[:space:]]Proxemics_[[:space:]]A[[:space:]]Prior[[:space:]]for[[:space:]]3D[[:space:]]Social[[:space:]]Interaction[[:space:]]from[[:space:]]Images/b086fde4-b439-4f25-92c2-a2dc5214a0a3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Generative[[:space:]]Quanta[[:space:]]Color[[:space:]]Imaging/9125a07a-d19f-455b-bcb9-5f2fd8cc25e9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Generative[[:space:]]Region-Language[[:space:]]Pretraining[[:space:]]for[[:space:]]Open-Ended[[:space:]]Object[[:space:]]Detection/bb1e00ce-acc5-43f7-b817-49b140b640c0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Generative[[:space:]]Rendering_[[:space:]]Controllable[[:space:]]4D-Guided[[:space:]]Video[[:space:]]Generation[[:space:]]with[[:space:]]2D[[:space:]]Diffusion[[:space:]]Models/038dcdfb-1eda-4a35-919e-6cd09ccad282_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Generative[[:space:]]Unlearning[[:space:]]for[[:space:]]Any[[:space:]]Identity/c3c2d3ee-c7c2-4590-9220-bed4fbf9b764_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/GenesisTex_[[:space:]]Adapting[[:space:]]Image[[:space:]]Denoising[[:space:]]Diffusion[[:space:]]to[[:space:]]Texture[[:space:]]Space/ba009eff-00cf-485a-8674-0d1aa83b6a01_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Genuine[[:space:]]Knowledge[[:space:]]from[[:space:]]Practice_[[:space:]]Diffusion[[:space:]]Test-Time[[:space:]]Adaptation[[:space:]]for[[:space:]]Video[[:space:]]Adverse[[:space:]]Weather[[:space:]]Removal/254c789f-0e30-45ba-9906-00c14a08b4b6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/GeoAuxNet_[[:space:]]Towards[[:space:]]Universal[[:space:]]3D[[:space:]]Representation[[:space:]]Learning[[:space:]]for[[:space:]]Multi-sensor[[:space:]]Point[[:space:]]Clouds/f223747a-d407-4fbb-9660-d95973ceefec_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/GeoChat_[[:space:]]Grounded[[:space:]]Large[[:space:]]Vision-Language[[:space:]]Model[[:space:]]for[[:space:]]Remote[[:space:]]Sensing/d9275ce8-58db-4c85-a4fc-e4300f4bc853_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/GeoReF_[[:space:]]Geometric[[:space:]]Alignment[[:space:]]Across[[:space:]]Shape[[:space:]]Variation[[:space:]]for[[:space:]]Category-level[[:space:]]Object[[:space:]]Pose[[:space:]]Refinement/966fabe9-a4db-4529-94bf-c710cc01e232_origin.pdf filter=lfs diff=lfs merge=lfs -text diff --git a/2024/GRAM_ Global Reasoning for Multi-Page VQA/bd3538ed-fd76-4274-b7c4-f2ff80c7e1bb_content_list.json b/2024/GRAM_ Global Reasoning for Multi-Page VQA/bd3538ed-fd76-4274-b7c4-f2ff80c7e1bb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9d50fc3972052e9d9396e3ab10232fe189e44eb2 --- /dev/null +++ b/2024/GRAM_ Global Reasoning for Multi-Page VQA/bd3538ed-fd76-4274-b7c4-f2ff80c7e1bb_content_list.json @@ -0,0 +1,1405 @@ +[ + { + "type": "text", + "text": "GRAM: Global Reasoning for Multi-Page VQA", + "text_level": 1, + "bbox": [ + 243, + 130, + 725, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tsachi Blau* Technion, Israel", + "bbox": [ + 106, + 180, + 236, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sharon Fogel \nAWS AI Labs", + "bbox": [ + 276, + 181, + 388, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Roi Ronen* Technion, Israel", + "bbox": [ + 426, + 181, + 555, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Alona Golts† \nAWS AI Labs", + "bbox": [ + 596, + 181, + 710, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shahar Tsiper \nAWS AI Labs", + "bbox": [ + 748, + 181, + 862, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Elad Ben Avraham \nAWS AI Labs", + "bbox": [ + 151, + 241, + 305, + 275 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Aviad Aberdam \nAWS AI Labs", + "bbox": [ + 351, + 242, + 478, + 275 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Roy Ganz* Technion, Israel", + "bbox": [ + 524, + 241, + 656, + 275 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ron Litman\\* AWS AI Labs", + "bbox": [ + 702, + 241, + 815, + 275 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 328, + 312, + 344 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The increasing use of transformer-based large language models brings forward the challenge of processing long sequences. In document visual question answering (DocVQA), leading methods focus on the single-page setting, while documents can span hundreds of pages. We present GRAM, a method that seamlessly extends pretrained single-page models to the multi-page setting, without requiring computationally-heavy pretraining. To do so, we leverage a single-page encoder for local page-level understanding, and enhance it with document-level designated layers and learnable tokens, facilitating the flow of information across pages for global reasoning. To enforce our model to utilize the newly introduced document tokens, we propose a tailored bias adaptation method. For additional computational savings during decoding, we introduce an optional compression stage using our compression-transformer(C-Former), reducing the encoded sequence length, thereby allowing a tradeoff between quality and latency. Extensive experiments showcase GRAM's state-of-the-art performance on the benchmarks for multi-page DocVQA, demonstrating the effectiveness of our approach.", + "bbox": [ + 75, + 359, + 473, + 680 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 720, + 209, + 738 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Document understanding, particularly in the context of DocVQA, has gained substantial research interest [5, 6, 16, 25, 36, 37] and offers a wide array of practical applications, focusing on data extraction and analysis of single page documents. However, Multi/Page DocVQA (MPDocVQA) poses a more realistic challenge, considering that the majority of documents, including contracts, manuals", + "bbox": [ + 75, + 747, + 468, + 853 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b2ea6080c27f9abf9ddb071d36d1e077d7390878872fde0dff61d4993c023edd.jpg", + "image_caption": [ + "Figure 1. An Overview of GRAM. We suggest an interleaved encoder architecture combining page- with document-attention layers, allowing information to propagate between different pages. An optional compression transformer (C-former) is introduced to allow a trade-off between quality and latency." + ], + "image_footnote": [], + "bbox": [ + 504, + 330, + 890, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and scientific papers, often extend well beyond a single page. Despite the practical relevance of MPDocVQA, it has received limited attention, primarily due to the absence of suitable datasets. Two recently introduced datasets, MPDocVQA [33] and DUDE [18], have opened up new avenues for MP-DocVQA research.", + "bbox": [ + 496, + 685, + 890, + 775 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent DocVQA approaches rely on transformers [35], at the heart of their architecture. While transformers are a powerful tool, they face challenges when dealing with long input sequences [4, 7, 10-12, 27, 38]. This difficulty stems from the self-attention mechanism, which scales quadratically in terms of computation and memory, with respect to the input sequence length. Addressing this limitation has attracted significant research efforts, primarily in the field", + "bbox": [ + 496, + 779, + 892, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work conducted during an internship at Amazon.", + "bbox": [ + 93, + 862, + 362, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author: alongolt@amazon.com", + "bbox": [ + 96, + 875, + 349, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\ddagger$ Corresponding author: litmanr@amazon.com", + "bbox": [ + 96, + 887, + 344, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15598", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of natural language processing (NLP). Proposed NLP-based solutions can be divided into two main directions: The former aims to modify the attention mechanism to cut computational costs [4, 7, 38]. The latter involves altering the positional embedding mechanism to improve performance on longer sequences, with minimal fine-tuning [11, 12, 27].", + "bbox": [ + 75, + 90, + 468, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A possible option of tackling MPDocVQA is to extend NLP-based approaches to handle multi-modal document data, including visual representations, along with OCR text and corresponding 2D locations and relative page position. However, this requires extensive pre-training, with relatively scarce multi-page document data, and thus is also sub-optimal in terms of performance. Instead, we opt for leveraging powerful single-page DocVQA models, especially pretrained on millions of single-page documents, and finetuning them to the multi-page scenario. For this purpose, we combine concepts of local (page) and global (document) tokens, which promote an exchange of information within and across pages, while keeping computational cost in check. We choose pages as atomic units in our proposed scheme, as page structure often represent a semantic unit in DocVQA.", + "bbox": [ + 75, + 185, + 470, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We present GRAM (Global ReAsoning for Multi-page VQA), a novel approach for endowing multi-page processing capabilities to existing single-page DocVQA models. Alongside page tokens that encapsulate both textual and visual contents of each page, we introduce doc(ument) learnable tokens, which aim is dispersing global information across all pages. These two sets of tokens interact within our newly-devised two-stage encoder blocks. The initial stage utilizes an existing single-page layer and enhances it by including both page and doc tokens as input, allowing them to freely interact. In the second stage, we prioritize computational efficiency by restricting self-attention solely to the global doc tokens. This global reasoning layer captures collective information from multiple pages, enabling the system to respond to cross-page inquiries, as illustrated in Fig. 1. Considering that doc tokens did not appear in pretraining, to boost their significance during finetune, we employ a designated bias adaptation mechanism which strikes a balance between local and global learnable tokens.", + "bbox": [ + 75, + 429, + 468, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While our method inherently deals with long sequences, we circumvent a quadratic reliance on sequence length by segmenting the document into pages — its semantically logical parts. We restrict interaction solely among doc learnable tokens, across all pages, thereby mitigating the computational burden of depending quadratically on the page count. Apart from encoding, the auto-regressive decoding stage poses a computational burden in long sequences. To this end, we introduce a compression stage that precedes the decoder, implemented with a compression transformer, termed CFormer. The CFormer receives the concatenated output of all pages and compresses it to a", + "bbox": [ + 75, + 719, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "much shorter sequence, distilling the most pertinent information in the document. Our key contributions are:", + "bbox": [ + 498, + 90, + 888, + 119 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose GRAM, an approach to endow single-page DocVQA methods with multi-page capabilities, without pretraining, allowing the model to process multi-page documents, while preserving single-page performance.", + "- We introduce document learnable tokens and bias adaptation that enable an effective communication and collaboration between individual pages to support reasoning over multiple page documents.", + "- Our C-Former module suggests a trade-off between accuracy and compute, distilling information from multi-page sequences into more compact representations.", + "- We obtain SOTA results over the MPDocVQA and DUDE datasets, and provide extensive ablations to each component in our method." + ], + "bbox": [ + 500, + 121, + 888, + 329 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 345, + 640, + 361 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Long Sequence Approaches are an active field of research in NLP, aiming to improve the design of chat-systems [24] and image instruction tasks [19, 21]. In these applications, the ability to manage and process long sequences is vital, as conversations cannot be cut short, or limited to just a few interactions. Common approaches to tackle long sequences include sparse attention mechanisms [4, 7, 38] and methods to improve results on long sequences during inference [11, 27, 28]. 'Sliding window' approaches of limiting the range of neighbors each token can attend to, lead to a significant reduction in computation and memory consumption. Prominent works of this kind include Long-Former [7], where each token attends to a set of its nearest neighbors, along with additional global tokens. The work of Big-Bird [38] adds additional non-neighboring tokens at random, whereas Colt5 [4] uses the same sliding window approach, but performs heavier computations for important tokens and shallow operations for filler words or punctuation. Although Tito et. al. [33] have demonstrated that the above approaches do not perform as well on the task of MPDocVQA, we do incorporate the ideas of combining both local and global tokens throughout encoding to expand the attention onto additional pages.", + "bbox": [ + 496, + 372, + 890, + 718 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "DocVQA has attracted increasing attention [5, 6, 9, 16, 25, 26, 31, 36, 37] with the introduction of the DocVQA dataset by Mathew et al. [22]. Most methods in DocVQA leverage OCR [1-3, 20, 23] to input both text and layout information (bounding box coordinates and possibly font type) into the model, where some further explore different techniques to combine the two types of data streams, or alternatively, clever schemes of pretraining. DocVQA methods can be roughly divided into two categories: extractive and abstractive. Extractive methods [16, 25, 36, 37] rely on the fact that the explicit answer resides in the written text, thus only output a corresponding text span within the input", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15599", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4a1cf9abb9b129b296325ad8d617ec7270c8af4ec0258570380b4af9be277ec6.jpg", + "image_caption": [ + "Figure 2. GRAM Architecture. (a) Depicts a high-level architecture overview. For each page, the visual, textual and question tokens are concatenated together with learnable doc tokens (darker color shade). The processed information is fed into the multi-page encoder. The encoder output can be fed directly into the decoder to create the final prediction. Optionally, a compression model, C-Former, can be used between the encoder and the decoder to compress the encoder output into a predetermined length, thus reducing overall latency for long documents. (b) Shows a global-local encoder layer, containing two sub-layers. The first sub-layer uses self-attention that operates on each page separately, while the second applies a self-attention step on the doc tokens to fuse information between the different pages. The corresponding tokens are then routed back to their respective page and go into the next global-local encoder layer." + ], + "image_footnote": [], + "bbox": [ + 143, + 97, + 421, + 443 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0d71025096650b205a82653e1b09dd7a52889303abd1cfb63fbac9d494a5e79a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 442, + 108, + 805, + 445 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sequence. Abstractive methods [5, 6, 14, 26, 31], on the other hand, have the capacity to generate free-form answers which do not necessarily appear in the text, thus providing flexibility in real-world applications. Notably, existing research in DocVQA does not scale in a straightforward way to deal with the more realistic multi-page scenario.", + "bbox": [ + 75, + 580, + 468, + 671 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "MPDocVQA has recently gained momentum with the launch of two new multi-page datasets: MP-DocVQA [33] and DUDE [18], offering two separate recipes to tackle longer documents. The first approach of Tito et. al., referred to as HiVT5 [33], suggests compressing the encoding of each page separately, and feeding the decoder with the concatenation of the compressed outputs from each single-page encoder. While this approach is advantageous in terms of computation, we later show the compression may severely hinder the results. In addition, there is no communication between the single-page encoders until the final stage of decoding, whereas in our method, we allow for exchange of page and document information throughout all stages of encoding. Another prominent approach, proposed by Landeghem et. al. [18], which relatively preserves qual", + "bbox": [ + 75, + 674, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ity, involves concatenating all the pages into one long sequence and feeding it to a standard encoder-decoder structure. This, however, poses a heavy computational burden as transformers' self-attention component scales quadratically with input sequence length.", + "bbox": [ + 496, + 580, + 890, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. GRAM", + "text_level": 1, + "bbox": [ + 500, + 670, + 586, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Base Architecture", + "text_level": 1, + "bbox": [ + 500, + 696, + 671, + 709 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The underlying idea in our approach is using existing encoder-decoder single-page models for document understanding and extending them to multi-page scenarios, without additional pretraining. In this work, we provide such a recipe over the notable DocFormerv2 [6]. For the sake of completeness, DocFormerv2 is a T5-based [28] encoder-decoder transformer model which operates over both visual and textual features to support document understanding. Each page is represented by textual features $\\mathbf{T} \\in R^{N_t \\times d}$ which encapsulate OCR tokens and their corresponding 2D bounding box positions [36], along with visual $\\mathbf{V} \\in R^{N_v \\times d}$ and question $\\mathbf{Q} \\in R^{N_q \\times d}$ embeddings. Where $N_t$ , $N_v$ and", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15600", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$N_{q}$ are the lengths of the OCR, visual features and the question. The output result $\\mathbf{Y}$ is obtained by passing the concatenated inputs through the encoder-decoder model,", + "bbox": [ + 76, + 90, + 468, + 137 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Y} = \\mathbf {D} (\\mathbf {E} (\\operatorname {c o n c a t} (\\mathbf {T}, \\mathbf {V}, \\mathbf {Q}))), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 148, + 468, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{E}$ and $\\mathbf{D}$ , are the encoder and decoder, respectively.", + "bbox": [ + 76, + 178, + 459, + 193 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our method uses these basic building blocks in designing a multi-page solution. To this end, we introduce a bi-level global-local encoder, as illustrated in Fig. 2. At the local page-level of each block, we utilize the layers of the existing single-page encoder $\\mathbf{E}$ to process each page separately, together with learnable doc tokens. Next, we introduce a slim global layer in each block that facilitates communication between doc tokens across all pages. This bi-level localized processing ensures the model can understand the content of each page effectively, while also combining information across pages in the document. After $M$ such blocks, we feed the encoded features from all pages into the existing decoder $\\mathbf{D}$ to produce the overall output.", + "bbox": [ + 75, + 194, + 470, + 390 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Global-Local Reasoning", + "text_level": 1, + "bbox": [ + 76, + 400, + 299, + 416 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To operate on multiple pages we break down the document to $K$ pages, and the single-page encoder to $M$ encoder layers, $\\mathbf{E}^j, j = 0,\\dots,M - 1$ . We then construct $M$ blocks, with two sub-layers each. The first page sub-layer originates from the existing pretrained encoder layer, referred to as $\\mathbf{E}_{page}^{j}$ , and operates in parallel, with shared weights, for all pages in the document. This layer receives both page and doc tokens. The second, newly introduced, document layer $\\mathbf{E}_{doc}^{j}$ collects only the doc tokens from all pages and promotes sharing information across all of the document.", + "bbox": [ + 75, + 424, + 468, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Formally, we augment the input of the standard single-page encoder with page-specific indexing $(\\mathbf{T}_i,\\mathbf{V}_i,\\mathbf{Q})$ and incorporate page-positional embedding $\\mathbf{P}_i$ to both text and visual features, where $i = 0,\\dots,K - 1$ :", + "bbox": [ + 75, + 575, + 468, + 635 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {T}} _ {i} = \\mathbf {T} _ {i} + \\mathbf {P} _ {i}, \\quad \\tilde {\\mathbf {V}} _ {i} = \\mathbf {V} _ {i} + \\mathbf {P} _ {i}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 648, + 468, + 666 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Next, we formulate our bi-level global-local block. The input to the first page-level sub-layer in each block is the concatenation of the textual, visual and question features, denoted $\\mathbf{X}_i^j = \\mathrm{concat}(\\tilde{\\mathbf{T}}_i,\\tilde{\\mathbf{V}}_i,\\mathbf{Q}_i)$ , along with page-specific doc tokens $\\mathbf{G}_i^j\\in R^{N_g\\times d}$", + "bbox": [ + 75, + 678, + 468, + 756 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {i} ^ {j + 1}, \\tilde {\\mathbf {G}} _ {i} ^ {j + 1} = \\mathbf {E} _ {p a g e} ^ {j} (\\operatorname {c o n c a t} \\left(\\mathbf {X} _ {i} ^ {j}, \\mathbf {G} _ {i} ^ {j}\\right)). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 768, + 468, + 789 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, the features undergo self-attention, normalization and feed-forward layers. The layer output $\\mathbf{X}_i^{j + 1}$ is passed on as input to the next bi-level block, whereas only the doc tokens $\\tilde{\\mathbf{G}}_i^{j + 1}$ , enter the second doc sub-layer, which again includes self-attention, normalization and feed-forward", + "bbox": [ + 75, + 796, + 468, + 869 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\mathbf {G} _ {i} ^ {j + 1} \\right\\} _ {i = 0} ^ {K - 1} = \\mathbf {E} _ {d o c} ^ {j} (\\operatorname {c o n c a t} \\left(\\left\\{\\tilde {\\mathbf {G}} _ {i} ^ {j + 1} \\right\\} _ {i = 0} ^ {K - 1}\\right)). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 882, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/980915e4e5e177aa55bc25e81b3ce9b2634298bf14abc038ccfc2dc77cbaf0fc.jpg", + "image_caption": [ + "Figure 3. Global-Local Attention: In long sequence approaches (a), attention is applied jointly to the entire sequence of concatenated local and global tokens. Our method, separates the computation into two steps — page-level (b) and document-level (c)— leveraging the natural division of documents into pages." + ], + "image_footnote": [], + "bbox": [ + 524, + 90, + 870, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this stage, the doc tokens can interact and pass information from page to page, after which being passed on to the next block, as depicted in Eq. (3). This design allows information to flow between pages while keeping computational costs in check. When concluding the traversal over $M$ such layers, the outputs across all pages are concatenated and fed to the decoder $\\mathbf{D}$ ,", + "bbox": [ + 498, + 364, + 890, + 469 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Y} _ {m u l t i} = \\mathbf {D} (\\operatorname {c o n c a t} \\left(\\left\\{\\mathbf {X} _ {i} ^ {M}, \\mathbf {G} _ {i} ^ {M} \\right\\} _ {i = 0} ^ {K - 1}\\right)). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 478, + 890, + 497 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To visualize the difference between the attention masks in our method, we compare it with previous long sequence approaches [4, 7, 38] in Fig. 3. These prior methods optimize computation by using attention masking on nearby tokens and allowing limited global connections. However, naively applying such methods to multi-page documents will treat it as a single stream, which does not consider the division into pages. Our global-local blocks, with a two-stage attention-masking mechanism, better suit multi-page documents. In addition, our two-level design benefits from existing, extensively pretrained single page models.", + "bbox": [ + 496, + 506, + 890, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Bias Adaptation", + "text_level": 1, + "bbox": [ + 500, + 681, + 660, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "An already-pretrained model, introduced with a new stream of data, might disregard it altogether [13, 14, 34, 39]. To overcome this, we force the system to account for the newly-introduced doc tokens by modifying the encoder's bias method. Originally, the bias method intervenes in the attention mechanism, diminishing the relationships between distant tokens. However, in our specific case, the distance between doc and page tokens does not represent their actual relevance. To enforce the encoder to pay closer attention to the doc tokens, we assign them a positive constant bias value. Particularly, we replace the values in the bias matrix, corresponding with the doc tokens, with fixed ones. Instead of a single bias value, we utilize a different value", + "bbox": [ + 496, + 704, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15601", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "for each attention head, as performed in ALiBi [27], enabling more fine-grained control of the global features in each head. Specifically, the constant doc bias value is set to $c \\cdot \\frac{1}{2^a}$ , where $c$ is a constant and $a$ is the attention head index. This yields a decaying bias value across different attention heads, resulting in hierarchical importance of the document information, where the first heads are more oriented towards doc tokens and the last towards page tokens.", + "bbox": [ + 75, + 90, + 472, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Compression Transformer", + "text_level": 1, + "bbox": [ + 76, + 220, + 316, + 234 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our global-local solution to MP-DocVQA resolves the problematic quadratic dependency on the number of pages $K$ during encoding. However, the auto-regressive decoding complexity scaling linearly with $K$ also poses a practical challenge during inference time, as we later discuss in Sec. 3.5. To alleviate this burden, we place an optional transformer-based model, named C-Former (Compression TransFormer), between the encoder outputs and the decoder, as depicted in Fig. 2. The C-Former has the ability to revise the information across all pages and distill only the important details, required to correctly answer the question.", + "bbox": [ + 75, + 243, + 468, + 409 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, the C-Former is a light-weight transformer-based decoder [28], denoted as $\\mathbf{D}_C$ , featuring crossattention, layer norm and feed-forward layers in each block. The input to C-Former includes $N_{c}$ learnable tokens $\\mathbf{C} \\in R^{N_c \\times d}$ , concatenated with the input question $\\tilde{\\mathbf{C}} = \\mathrm{concat}(\\mathbf{C}, \\mathbf{Q})$ . In addition, we feed it with the outputs of the global-local interlaced encoder, concatenated to one long sequence, referred as $\\mathbf{O}$ , where $\\mathbf{O} = \\mathrm{concat}(\\{\\mathbf{X}_i^M, \\mathbf{G}_i^M\\}_{i=0}^{K-1})$ . The output of C-Former is thus", + "bbox": [ + 75, + 409, + 468, + 547 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {O} _ {C} = \\mathbf {D} _ {C} (Q = \\tilde {\\mathbf {C}}, K = \\mathbf {O}, V = \\mathbf {O}),\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 556, + 385, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where we pass forward only the first set of $N_{c}$ output embeddings and ignore the rest, setting the output sequence dimension to $N_{c}$ . C-Former offers flexibility in controlling the tradeoff between ANLS quality and computational efficiency by controlling the output sequence length $N_{c}$ .", + "bbox": [ + 75, + 582, + 468, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Computation Analysis", + "text_level": 1, + "bbox": [ + 76, + 665, + 285, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, we turn to provide a thorough computational complexity analysis. We consider a document that comprises of $K$ pages, each with $N$ tokens, and the maximum answer length is $L$ . For simplicity, we assume that all encoders and decoders have one layer. The naive way to support multi-page documents is using an existing single-page encoder-decoder model, fusing all of the textual page inputs together, and feeding them as one long sequence. We refer to this approach as 'concat'. The self-attention complexity of such a configuration scales quadratically with the sequence length, $O((N \\cdot K)^2)$ . Conversely in our method, we operate on the document pages with two alternating encoding stages in each layer. The first stage performs a self-attention over both the page and doc tokens. Hence, the complexity of", + "bbox": [ + 75, + 688, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "such sub-layer is $O((N + N_g)^2 \\cdot K)$ , where $N_g$ is the number of doc tokens. The second stage features a self-attention operation over the doc tokens, across all pages in the document. The complexity of this operation is $O((N_g \\cdot K)^2)$ . Overall, the total complexity for one global-local encoder block is $O((N + N_g)^2 \\cdot K + (N_g \\cdot K)^2)$ . Since $N_g$ is a constant, and the number of pages is usually less than the number of words in each page ( $K < N$ ), we obtain a complexity of $O(N^2 \\cdot K)$ , which is not quadratic in $K$ .", + "bbox": [ + 496, + 90, + 890, + 226 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Prior to decoding, the outputs of all per-page encoders are concatenated, thus the output sequence length is $(N + N_{g})\\cdot K$ Since the decoder is auto-regressive, its complexity depends quadratically on the maximum output length, $L$ ,namely, $O((N + N_g)\\cdot K\\cdot L^2) = O(N\\cdot K\\cdot L^2)$ Since this operation of decoding is performed iteratively during inference, the combined sequence length $(N + N_{g})\\cdot K$ becomes computationally heavy. To alleviate this concern, we propose an optional C-Former model, which performs compression prior to decoding. The overall complexity in this decoding scheme includes passing through the C-Former and then through the decoder, leading to $O((N + N_g)\\cdot K\\cdot N_c) + N_c\\cdot L^2)$ which is equivalent to $O(N\\cdot K + L^2)$ , since $N_{c}$ is a constant, denoting the number of compression tokens in C-Former.", + "bbox": [ + 496, + 227, + 892, + 469 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 498, + 483, + 633, + 501 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 500, + 508, + 705, + 525 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets and Metrics The MPDocVQA dataset [33] features 46K questions, spanning over 48K images, and includes layout elements as figures, tables, lists and diagrams, with printed, handwritten and typewritten text. MPDocVQA contains mostly extractive questions, for which answers are present in the given text. DUDE is smaller in size (23.7K questions over 3K documents), but offers complex questions that require a reader to rationalize beyond the written text content. We report our results using the ANLS metric, introduced in [8], computing a generalized accuracy. Results for DUDE can be broken apart to several types of questions, categorized to four groups: 'extractive' – for which the answer is found directly in the text; 'abstractive' – requiring a free-form answer that does not necessarily appear in the document; 'list of answers' – requiring a list of answers, as opposed to a single one, and 'unanswerable' – where the result cannot be determined using the text.", + "bbox": [ + 496, + 532, + 890, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details Our underlying architecture is based on Docformerv2 [6]. Recall, our interlaced encoder features $M$ blocks (12 in 'base' and 24 in 'large'), where each block contains a page sub-layer which originates from an extension of Docformerv2's encoder layer. Every structure contains self-attention, normalization and feed-forward", + "bbox": [ + 496, + 809, + 892, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15602", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/9fe7fa4024986a36d84a9c523ece8f2ed6b31860035b2b819ac2fdf2e218ed55.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodParamsMPDocVQADUDE
ANLSANLSExtractiveANLS per Question TypeUnanswerable
AbstractiveList of answers
Longformer [7]148M55.0627.1443.588.5510.6210.78
BigBird [38]131M58.5426.2740.267.118.4612.75
LayoutLMv3 [38]125M55.1320.3132.608.107.828.82
Hi-VT5†beamsearch [15]316M-35.7428.3132.9810.6062.90
Hi-VT5[33]316M62.0123.0617.6033.946.8361.67
Hi-VT5*257M60.7823.867.2116.563.5372.77
DocFormerv2concat [6]257M69.6744.2141.6641.8615.1365.19
GRAMC-Former286M70.8040.0740.4339.6111.4252.55
GRAM281M73.6846.1546.0744.8215.2762.18
T5-2D [18]770M-46.0655.6550.815.4368.62
DocGptVQA [30]>3.5B-50.0251.8648.3228.2262.04
DocBlipVQA [29]>3.5B-47.6250.6946.3130.7355.22
Hi-VT5* [33]784M71.3528.8918.2126.176.8458.99
DocFormerv2concat [6]784M76.4048.4450.8248.0617.6759.04
GRAMC-Former864M77.6045.4747.6344.9114.3456.99
GRAM859M80.3251.1553.6750.3518.4063.23
Hi-VT5*†[33]784M73.5149.1849.2948.3513.3065.95
DocFormerv2†concat [6]784M76.7750.7952.7049.6117.3365.14
GRAM†C-Former864M78.1250.9755.1550.4617.2661.04
GRAM†859M79.6753.3656.8352.3219.9665.43
", + "bbox": [ + 76, + 88, + 888, + 439 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/df831a3e36baa8f2ebe3e2ed6f7caee0ce3a77c9095766b2b27b45d080cb233d.jpg", + "table_caption": [ + "Table 1. Quantitative Results. We present ANLS results for the MPDocVQA [33] and DUDE [18] test sets. The methods are grouped according to the model type and size, starting from encoder-only models (top), T5-base models (middle) and T5-large models (bottom). ${}^{ \\dagger }$ denotes training with both MPDocVQA and DUDE." + ], + "table_footnote": [], + "table_body": "
MethodTraining DataANLS
DocVQAMPDocVQADocVQAMPDocVQA
DocFormerv2concat✓✗86.6072.73
✗✓85.2876.40
✓✓86.4775.37
GRAM✓✗86.7073.12
✗✓85.2980.32
✓✓86.3278.66
", + "bbox": [ + 81, + 516, + 465, + 627 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. DocVQA vs. MPDocVQA Performance. Test results over both datasets using the large model variants. A checkmark denotes whether a dataset was included in training or not.", + "bbox": [ + 75, + 638, + 468, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "layers. We extend the page layer from Docformerv2 to feature also the doc learnable embeddings. The second doc sub-layer is similar in structure to the first sub-layer, only it is initialized from scratch, with the following specification: $d_{ff} = 1024$ , $d_{kv} = 64$ , $nheads = 4$ , $d = 256$ . We implement 32 doc learnable tokens for each page, uniformly initialized to random values. For bias adaptation, the initial bias value is set to $c = 20$ , with variations between encoder heads, as described in Sec. 3.3. We incorporate an additional optional compression stage using C-Former - a randomly-initialized T5 [28] tiny decoder, with an encoder mask instead of a causal one. The output sequence extracted", + "bbox": [ + 75, + 719, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "from C-Former is $N_{c} = 256$ . Finally the decoder is initialized with pretrained weights from DocformerV2.", + "bbox": [ + 496, + 518, + 890, + 549 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The model is trained with the Hugging Face Trainer [17] for $200k$ steps, starting with a warm-up of $1k$ steps, with linear learning rate decay. We use learning rates of $3e^{-5}$ and $1e^{-4}$ for the already pretrained encoder and decoder weights, versus the newly initialized doc sub-layer weights. Training is performed on a cluster of $8 \\times A100$ GPUs, each with 40GB of RAM. During training, each page encoder receives 800 tokens, dealing with up to 4 pages. During testing, we increase the maximum length of tokens to 8,000.", + "bbox": [ + 496, + 549, + 892, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines We report the results of previous work on both MP-DocVQA and DUDE datasets (if those exist), including the NLP-based Longformer [7] and BigBird [38], which were adapted to MPDocVQA by [33]; LayoutLMv3 [16], originally designed for DocVQA; and Hi-VT5 [33] and T5-2D [32], specifically suggested for MP-DocVQA Task. We also add for reference the results of methods published in the leader-boards of MPDocVQA and DUDE, which do not have corresponding papers, including DocGptVQA [30], DocBlipVQA [30], and Hi-VT5_beamsearch [15] ([15] was trained on both MP-DocVQA and DUDE). In our approach, we present two variations: GRAM and $\\mathrm{GRAM}_{C - \\text{Former}}$ . While GRAM utilizes the full length of the encoder output,", + "bbox": [ + 496, + 703, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15603", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/58c8aa6b0d7b67224e400bd0a24a675c85c4896550ca2195790edc7174995db2.jpg", + "image_caption": [ + "How many diagrams are there?", + "HIVT5:\"2\"", + "GRAM:\"3\"", + "Which month shows the hurricane?", + "HiVT5:\"not answerable\"", + "Figure 4. Qualitative comparison between our approach and Hi-VT5 [33] indicate that the integration of our global-local encoder enhances reasoning capabilities, especially when the inquiries require multi-page context." + ], + "image_footnote": [], + "bbox": [ + 127, + 130, + 370, + 263 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a464058a112ba5afcf517b166a284d65eb3ee38e011249701fe9d365f0757b85.jpg", + "image_caption": [ + "GRAM:\"July\"" + ], + "image_footnote": [], + "bbox": [ + 387, + 135, + 576, + 260 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f6c8b74b3eaf938c1c59278d2c880c02180af5a810fd497fbfb52742cbac40a2.jpg", + "image_caption": [ + "What is the eps code for little rock?", + "HiVT5:\"EPS\"", + "GRAM:\"AR01\"" + ], + "image_footnote": [], + "bbox": [ + 588, + 130, + 846, + 258 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$\\mathrm{GRAM}_{C - \\text{Former}}$ allows the user to control the trade-off between performance and latency.", + "bbox": [ + 75, + 328, + 470, + 359 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To ensure a fair comparison, since we use the pretrained model of DocFormerv2 [6], we implement two additional baselines, referred to as Hi-VT5* and DocFormerv2concat. The first follows a similar structure as Hi-VT5 [33], with the encoder originating from DocFormerv2, however without the page answer prediction, as it does not exist in DUDE. The second recreates the approach of [18], where only the textual tokens of all pages are concatenated to one long sequence, then passing through the DocFormerv2 model. The second approach poses a computational burden, thus we use only 600 tokens during training per page, with up to 4 pages, and during test only 400 tokens.", + "bbox": [ + 75, + 359, + 468, + 541 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Results", + "text_level": 1, + "bbox": [ + 76, + 555, + 171, + 570 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We present the performance of our method over the MPDocVQA [33] and DUDE [18] datasets in Tab. 1. The methods are divided into three groups: the top contains encoder-only, and methods that rely on the T5-base model (up to $316M$ parameters); the middle section, approaches that use the T5-large model (over $770M$ parameters), and finally the bottom, T5-large models, trained on both datasets.", + "bbox": [ + 75, + 580, + 468, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As can be seen, in the first group, the encoder-only NLP methods, LongFormer [7], BigBird [38] and LayoutLMv3 [16] can only handle relatively well 'extractive' style tasks as in MPDocVQA dataset [33], but often struggle with 'abstractive' questions that are more abundant in DUDE [18]. As to T5-'base' models, versus our best competitor Docformerv2concat, we obtain an improvement of $(+4\\%, +1.9\\%)$ on MP-DocVQA and DUDE datasets. As to methods that combine an additional compression before decoding (Hi-VT5, Hi-VT5*), our C-Former achieves an increase in $(+8.8\\%, +16.2\\%)$ over the best candidates on the MP-DocVQA and DUDE datasets.", + "bbox": [ + 75, + 686, + 468, + 867 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As for the group of 'large' models, we include the results of T5-2D [18] DocGptVQA [30] and", + "bbox": [ + 75, + 869, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "DocBlipVQA [29]. Note that our model surpasses DocFormerconcat, the primary baseline, achieving improvements of $(+3.9\\%, +2.7\\%)$ on MP-DocVQA and DUDE, respectively. We also outperform DocGptVQA [30], a method that appears in the leaderboard of DUDE, by $+1.1\\%$ , thereby obtaining SOTA results for GRAM 'large'.", + "bbox": [ + 496, + 328, + 892, + 419 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The final category showcases large encoder-decoder models, fine-tuned on both MP-DocVQA and DUDE training sets, showcasing the benefits of augmented training data. GRAM consistently demonstrates performance gains over the baseline, illustrating its robustness across different datasets and training scenarios. Next, we present in Tab. 2 the effect of training on DocVQA vs. MPDocVQA. Our method achieves performance on-par on the single page task, while enhancing performance on the multi-page scenario by $+3.3\\%$ , compared to the baseline.", + "bbox": [ + 496, + 422, + 892, + 573 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Fig. 4, We show qualitative results on the DUDE dataset of GRAM versus Hi-VT5* [33]. Our method demonstrates proficiency in addressing questions that involve attention over multiple pages ('how many diagrams are there'), an increased visual analysis capability ('Which month shows the hurricane?'), and heightened abstractive ability ('What is the EPS code for Little Rock?').", + "bbox": [ + 496, + 575, + 890, + 683 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 705, + 651, + 722 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We perform an ablation study on our approach, evaluating the influence of each constituent component using DUDE's validation set [18]. This validation set enables the grouping of documents by their respective page counts: 1, 2-4, 5-10, 11-end, encompassing 1747, 2259, 1062, 1241 samples in each category, respectively. Our investigation delves into the impact of the number of doc tokens and the bias adaptation methods. Moreover, we employ the C-Former for sequence compression, adjusting the compression ratio and examining the balance between performance and latency (see supplementary for more details).", + "bbox": [ + 496, + 734, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15604", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7c525b2d9791bd906cfef1141145ec50a053883b2824ca8f450e6b8a35ab3d49.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
#Doc TokensBias TypeCompression DimensionANLS by Number of Pages DUDE validation dataset
All12-45-1011-end
XXX46.1647.1848.6643.3442.57
16X46.3948.3549.0643.1641.56
32DecayingX47.8849.2949.9045.9043.94
64X46.7047.9849.2244.0042.60
32XX47.5249.8549.9344.9042.10
ConstantX46.1447.4148.1344.4442.19
DecayingX47.8849.2949.9045.9043.94
32Decaying839.8339.7341.4136.9839.52
3240.4240.9541.6438.1239.39
25641.9942.5743.7738.4041.01
102442.5642.9744.3038.9441.93
409643.5944.7544.6440.5442.67
", + "bbox": [ + 81, + 88, + 467, + 292 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "GRAM Components We focus our initial exploration on the impact of the number of doc tokens $N_{g}$ . As can be seen in Tab. 3, while $N_{g} = 16$ leads to performance on par with not using doc tokens at all, for the optimal value of $N_{g} = 32$ , we obtain an increase of $+1.7\\%$ in ANLS. Shifting our focus to bias adaptation methods, Tab. 3 shows that using constant bias has a negative effect on the results, suggesting this method is not flexible enough in maintaining a balance between the page and doc tokens. However, our decaying bias-adaptation approach does improve results overall, versus no-bias $(+0.36\\%)$ , especially for longer documents $(+1\\%$ improvement for 5-10 pages and $+1.84\\%$ for 11 pages and more). This is to be expected, since incorporating new doc tokens and increasing their importance can potentially affect single-page performance. Finally, in Tab. 4, we reinforce our choice of pages as semantic logical units for MPDocVQA. We first ablate our method with and without page embedding. Next, we compare our page-based division with varying fixed-length division of tokens for encoder. Results in Tab. 4 clearly demonstrate an advantage towards page-level encoding in MPDocVQA. This aligns with our initial assumption that structured documents are often designed with page-division in mind.", + "bbox": [ + 75, + 376, + 470, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Performance-Latency Trade-off We assess the impact of C-Former on performance, considering compression output lengths of 8, 32, 256, 1024, 4096. Note, performance gradually improves with an increase in the compression output length. However, longer output lengths correspond to heightened model latency. Note that using C-Former for shorter documents can be redundant, as there is little to no compression compared to the input sequence length and results decrease. In Fig. 5, we scrutinize the trade-off between computational efficiency and compression rate", + "bbox": [ + 75, + 750, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1b65f40706a96aa77ea17d9ec97973110ae15eaeae3341626ff31ca4e2167582.jpg", + "table_caption": [ + "Table 3. GRAM Ablation Study. Results on DUDE validation set ablating over (a) the dimension of doc tokens, (b) the attention bias employed and (c) the C-former input dimension." + ], + "table_footnote": [], + "table_body": "
Page EmbeddingSegment LengthANLS by Number of Pages DUDE validation dataset
All12-45-1011-end
✓✗47.8849.2949.9045.9043.94
✗✗46.1248.7448.1143.5940.99
✓25645.2246.3846.6944.1341.83
✓51245.0945.9047.3242.6541.98
✓102444.3944.9846.6341.6941.78
", + "bbox": [ + 504, + 88, + 890, + 210 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. The Significance of Pages as Semantic Units. Results on DUDE validation set ablating over (a) utilization of page-embedding, (b) segment length for fixed-size encoding inputs.", + "bbox": [ + 498, + 220, + 890, + 263 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/148ed5b5337a0df77c963769acb9ace696da8e29227d9d1bcf6b5fd25c5b981c.jpg", + "image_caption": [ + "Figure 5. Latency comparison. We compare the dependency between overall latency and the number of pages in input document for GRAM, $\\mathrm{GRAM}_{C - \\text{Former}}$ , DocFormerv2concat and Hi-VT5." + ], + "image_footnote": [], + "bbox": [ + 524, + 290, + 870, + 431 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "by comparing to DocFormerv2concat [6] and Hi-VT5* [33]. We discover that DocFormerv2concat reaches a memory limit at approximately 20 pages, due to its quadratic memory increase with sequence length. At this juncture, GRAM $_{C-Former}$ surpasses DocFormerv2concat by performing 3.5 seconds faster. Notably, GRAM $_{C-Former}$ can gracefully handle documents surpassing 300 pages, effectively bridging the gap between performance and latency.", + "bbox": [ + 496, + 522, + 890, + 645 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 672, + 627, + 689 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our method, termed GRAM, extends existing single-page document models to efficiently handle multi-page documents without necessitating computationally-intensive pretraining. Leveraging the single-page encoder for local page-level comprehension, we introduce document learnable tokens and designated layers, enabling seamless information exchange across pages. Additionally, our proposed bias adaptation method enforces effective utilization of our newly introduced document tokens. The incorporation of a C-Former model reduces sequence length, balancing quality with latency in the decoding step. Extensive experiments demonstrate GRAM's state-of-the-art performance across multi-page DocVQA benchmarks.", + "bbox": [ + 496, + 703, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15605", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Aviad Aberdam, Ron Litman, Shahar Tsiper, Oron Anschel, Ron Slossberg, Shai Mazor, R Manmatha, and Pietro Perona. Sequence-to-sequence contrastive learning for text recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15302-15312, 2021. 2", + "[2] Aviad Aberdam, Roy Ganz, Shai Mazor, and Ron Litman. Multimodal semi-supervised learning for text recognition. arXiv preprint arXiv:2205.03873, 2022.", + "[3] Aviad Aberdam, David Bensaïd, Alona Golts, Roy Ganz, Oren Nuriel, Royee Tichauer, Shai Mazor, and Ron Litman. Clipter: Looking at the bigger picture in scene text recognition. arXiv preprint arXiv:2301.07464, 2023. 2", + "[4] Joshua Ainslie, Tao Lei, Michiel de Jong, Santiago Ontañón, Siddhartha Brahma, Yury Zemlyanskiy, David Uthus, Mandy Guo, James Lee-Thorp, Yi Tay, et al. Colt5: Faster long-range transformers with conditional computation. arXiv preprint arXiv:2303.09752, 2023. 1, 2, 4", + "[5] Srikar Appalaraju, Bhavan Jasani, Bhargava Urala Kota, Yusheng Xie, and R Manmatha. Docformer: End-to-end transformer for document understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 993-1003, 2021. 1, 2, 3", + "[6] Srikar Appalaraju, Peng Tang, Qi Dong, Nishant Sankaran, Yichu Zhou, and R Manmatha. Docformerv2: Local features for document understanding. arXiv preprint arXiv:2306.01733, 2023. 1, 2, 3, 5, 6, 7, 8", + "[7] Iz Beltagy, Matthew E Peters, and Arman Cohan. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150, 2020. 1, 2, 4, 6, 7", + "[8] Ali Furkan Biten, Ruben Tito, Andres Mafla, Lluis Gomez, Marçal Rusinol, Ernest Valveny, CV Jawahar, and Dimosthenis Karatzas. Scene text visual question answering. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4291-4301, 2019. 5", + "[9] Ali Furkan Biten, Ron Litman, Yusheng Xie, Srikar Appalaraju, and R Manmatha. Latr: Layout-aware transformer for scene-text vqa. pages 16548-16558, 2022. 2", + "[10] Aydar Bulatov, Yuri Kuratov, and Mikhail S Burtsev. Scaling transformer to 1m tokens and beyond with rmt. arXiv preprint arXiv:2304.11062, 2023. 1", + "[11] Shouyuan Chen, Sherman Wong, Liangjian Chen, and Yuandong Tian. Extending context window of large language models via positional interpolation. arXiv preprint arXiv:2306.15595, 2023. 2", + "[12] Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc V Le, and Ruslan Salakhutdinov. Transformer-xl: Attentive language models beyond a fixed-length context. arXiv preprint arXiv:1901.02860, 2019. 1, 2", + "[13] Robert M French. Catastrophic forgetting in connectionist networks. Trends in cognitive sciences, 3(4):128-135, 1999. 4", + "[14] Roy Ganz, Oren Nuriel, Aviad Aberdam, Yair Kittenplon, Shai Mazor, and Ron Litman. Towards models that can see and read. arXiv preprint arXiv:2301.07389, 2023. 3, 4" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] JiangLong He, Mamatha N, Shiv Vignesh, and Deepak Kumar. Hivt5beam. Hi-VT5 model pretrained with private custom document collection using span masking objective. Pretrained model is then trained with DUDE dataset and MultiPage DocVQA dataset, 2023. 6", + "[16] Yupan Huang, Tengchao Lv, Lei Cui, Yutong Lu, and Furu Wei. Layoutmv3: Pre-training for document ai with unified text and image masking. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4083-4091, 2022. 1, 2, 6, 7", + "[17] HuggingFace. Huggingface trainer. https://huggingface.co/docs/transformers/training. 6", + "[18] Jordy Landeghem, Rubén Tito, Łukasz Borchmann, Michal Pietruszka, Paweł Józiak, Rafał Powalski, Dawid Jurkiewicz, Mickaël Coustaty, Bertrand Ackaert, Ernest Valveny, et al. Document understanding dataset and evaluation (dude). arXiv preprint arXiv:2305.08455, 2023. 1, 3, 6, 7", + "[19] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 2", + "[20] Ron Litman, Oron Anschel, Shahar Tsiper, Roeit Litman, Shai Mazor, and R Manmatha. Scatter: selective context attentional scene text recognizer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11962-11972, 2020. 2", + "[21] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 2", + "[22] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 2", + "[23] Oren Nuriel, Sharon Fogel, and Ron Litman. Textadain: Paying attention to shortcut learning in text recognizers. In European Conference on Computer Vision, pages 427-445. Springer, 2022. 2", + "[24] R OpenAI. Gpt-4 technical report. arXiv, pages 2303–08774, 2023. 2", + "[25] Qiming Peng, Yinxu Pan, Wenjin Wang, Bin Luo, Zhenyu Zhang, Zhengjie Huang, Teng Hu, Weichong Yin, Yongfeng Chen, Yin Zhang, et al. Ernie-Layout: Layout knowledge enhanced pre-training for visually-rich document understanding. arXiv preprint arXiv:2210.06155, 2022. 1, 2", + "[26] Rafal Powalski, Łukasz Borchmann, Dawid Jurkiewicz, Tomasz Dwojak, Michal Pietruszka, and Gabriela Pałka. Going full-tilt boogie on document understanding with text-image layout transformer. In Document Analysis and Recognition-ICDAR 2021: 16th International Conference, Lausanne, Switzerland, September 5–10, 2021, Proceedings, Part II 16, pages 732–747. Springer, 2021. 2, 3", + "[27] Ofir Press, Noah A Smith, and Mike Lewis. Train short, test long: Attention with linear biases enables input length extrapolation. arXiv preprint arXiv:2108.12409, 2021. 1, 2, 5", + "[28] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15606", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485-5551, 2020. 2, 3, 5, 6", + "[29] RenZhou, QiaolingDeng, XinfengChang, LuyanWang, XiaochenHu, HuiLi, and YaqiangWu. Docclipvqa. We integrated the prediction outputs from the UDOP model and Blip2 to enhance our results, and we optimized the image encoder and included page number features to address the challenge of multi-page documents. GPT to generate python-like modular programs., 2023. 6, 7", + "[30] RenZhou, QiaolingDeng, XinfengChang, LuyanWang, XiaochenHu, HuiLi, and YaqiangWu. Docgptvqa. We integrated the prediction outputs from the UDOP model and Blip2 to enhance our results, and we optimized the image encoder and included page number features to address the challenge of multi-page documents. GPT to generate python-like modular programs., 2023. 6, 7", + "[31] Zineng Tang, Ziyi Yang, Guoxin Wang, Yuwei Fang, Yang Liu, Chenguang Zhu, Michael Zeng, Cha Zhang, and Mohit Bansal. Unifying vision, text, and layout for universal document processing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19254-19264, 2023. 2, 3", + "[32] Rubén Tito, Dimosthenis Karatzas, and Ernest Valveny. Document collection visual question answering. In Document Analysis and Recognition-ICDAR 2021: 16th International Conference, Lausanne, Switzerland, September 5–10, 2021, Proceedings, Part II 16, pages 778–792. Springer, 2021. 6", + "[33] Rubén Tito, Dimosthenis Karatzas, and Ernest Valveny. Hierarchical multimodal transformers for multi-page docvqa. arXiv preprint arXiv:2212.05935, 2022. 1, 2, 3, 5, 6, 7, 8", + "[34] Maria Tsimpoukelli, Jacob L Menick, Serkan Cabi, SM Eslami, Oriol Vinyls, and Felix Hill. Multimodal few-shot learning with frozen language models. Advances in Neural Information Processing Systems, 34:200-212, 2021. 4", + "[35] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1", + "[36] Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, and Ming Zhou. Layoutm: Pre-training of text and layout for document image understanding. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 1192-1200, 2020. 1, 2, 3", + "[37] Yang Xu, Yiheng Xu, Tengchao Lv, Lei Cui, Furu Wei, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Wanxiang Che, et al. Layoutmv2: Multi-modal pre-training for visually-rich document understanding. arXiv preprint arXiv:2012.14740, 2020. 1, 2", + "[38] Manzil Zaheer, Guru Guruganesh, Kumar Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontanon, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, et al. Big bird: Transformers for longer sequences. Advances in neural information processing systems, 33:17283-17297, 2020. 1, 2, 4, 6, 7", + "[39] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer." + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Lit: Zero-shot transfer with locked-image text tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18123-18133, 2022. 4", + "bbox": [ + 532, + 90, + 890, + 133 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15607", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/GRAM_ Global Reasoning for Multi-Page VQA/bd3538ed-fd76-4274-b7c4-f2ff80c7e1bb_model.json b/2024/GRAM_ Global Reasoning for Multi-Page VQA/bd3538ed-fd76-4274-b7c4-f2ff80c7e1bb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..fce045ac9c6417b7b180f5e63de8c500613fec83 --- /dev/null +++ b/2024/GRAM_ Global Reasoning for Multi-Page VQA/bd3538ed-fd76-4274-b7c4-f2ff80c7e1bb_model.json @@ -0,0 +1,1969 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.244, + 0.131, + 0.726, + 0.154 + ], + "angle": 0, + "content": "GRAM: Global Reasoning for Multi-Page VQA" + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.181, + 0.238, + 0.216 + ], + "angle": 0, + "content": "Tsachi Blau* Technion, Israel" + }, + { + "type": "text", + "bbox": [ + 0.277, + 0.182, + 0.39, + 0.216 + ], + "angle": 0, + "content": "Sharon Fogel \nAWS AI Labs" + }, + { + "type": "text", + "bbox": [ + 0.428, + 0.182, + 0.557, + 0.216 + ], + "angle": 0, + "content": "Roi Ronen* Technion, Israel" + }, + { + "type": "text", + "bbox": [ + 0.598, + 0.182, + 0.711, + 0.216 + ], + "angle": 0, + "content": "Alona Golts† \nAWS AI Labs" + }, + { + "type": "text", + "bbox": [ + 0.749, + 0.182, + 0.863, + 0.216 + ], + "angle": 0, + "content": "Shahar Tsiper \nAWS AI Labs" + }, + { + "type": "text", + "bbox": [ + 0.153, + 0.242, + 0.307, + 0.276 + ], + "angle": 0, + "content": "Elad Ben Avraham \nAWS AI Labs" + }, + { + "type": "text", + "bbox": [ + 0.352, + 0.243, + 0.48, + 0.276 + ], + "angle": 0, + "content": "Aviad Aberdam \nAWS AI Labs" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.242, + 0.657, + 0.276 + ], + "angle": 0, + "content": "Roy Ganz* Technion, Israel" + }, + { + "type": "text", + "bbox": [ + 0.703, + 0.242, + 0.816, + 0.276 + ], + "angle": 0, + "content": "Ron Litman\\* AWS AI Labs" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.329, + 0.313, + 0.345 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.361, + 0.474, + 0.681 + ], + "angle": 0, + "content": "The increasing use of transformer-based large language models brings forward the challenge of processing long sequences. In document visual question answering (DocVQA), leading methods focus on the single-page setting, while documents can span hundreds of pages. We present GRAM, a method that seamlessly extends pretrained single-page models to the multi-page setting, without requiring computationally-heavy pretraining. To do so, we leverage a single-page encoder for local page-level understanding, and enhance it with document-level designated layers and learnable tokens, facilitating the flow of information across pages for global reasoning. To enforce our model to utilize the newly introduced document tokens, we propose a tailored bias adaptation method. For additional computational savings during decoding, we introduce an optional compression stage using our compression-transformer(C-Former), reducing the encoded sequence length, thereby allowing a tradeoff between quality and latency. Extensive experiments showcase GRAM's state-of-the-art performance on the benchmarks for multi-page DocVQA, demonstrating the effectiveness of our approach." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.722, + 0.21, + 0.739 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.748, + 0.47, + 0.854 + ], + "angle": 0, + "content": "Document understanding, particularly in the context of DocVQA, has gained substantial research interest [5, 6, 16, 25, 36, 37] and offers a wide array of practical applications, focusing on data extraction and analysis of single page documents. However, Multi/Page DocVQA (MPDocVQA) poses a more realistic challenge, considering that the majority of documents, including contracts, manuals" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.331, + 0.891, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.583, + 0.894, + 0.655 + ], + "angle": 0, + "content": "Figure 1. An Overview of GRAM. We suggest an interleaved encoder architecture combining page- with document-attention layers, allowing information to propagate between different pages. An optional compression transformer (C-former) is introduced to allow a trade-off between quality and latency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.686, + 0.892, + 0.776 + ], + "angle": 0, + "content": "and scientific papers, often extend well beyond a single page. Despite the practical relevance of MPDocVQA, it has received limited attention, primarily due to the absence of suitable datasets. Two recently introduced datasets, MPDocVQA [33] and DUDE [18], have opened up new avenues for MP-DocVQA research." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Recent DocVQA approaches rely on transformers [35], at the heart of their architecture. While transformers are a powerful tool, they face challenges when dealing with long input sequences [4, 7, 10-12, 27, 38]. This difficulty stems from the self-attention mechanism, which scales quadratically in terms of computation and memory, with respect to the input sequence length. Addressing this limitation has attracted significant research efforts, primarily in the field" + }, + { + "type": "page_footnote", + "bbox": [ + 0.094, + 0.863, + 0.363, + 0.876 + ], + "angle": 0, + "content": "*Work conducted during an internship at Amazon." + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.876, + 0.35, + 0.888 + ], + "angle": 0, + "content": "† Corresponding author: alongolt@amazon.com" + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.888, + 0.345, + 0.9 + ], + "angle": 0, + "content": "\\(\\ddagger\\) Corresponding author: litmanr@amazon.com" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.863, + 0.363, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15598" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.183 + ], + "angle": 0, + "content": "of natural language processing (NLP). Proposed NLP-based solutions can be divided into two main directions: The former aims to modify the attention mechanism to cut computational costs [4, 7, 38]. The latter involves altering the positional embedding mechanism to improve performance on longer sequences, with minimal fine-tuning [11, 12, 27]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.186, + 0.471, + 0.426 + ], + "angle": 0, + "content": "A possible option of tackling MPDocVQA is to extend NLP-based approaches to handle multi-modal document data, including visual representations, along with OCR text and corresponding 2D locations and relative page position. However, this requires extensive pre-training, with relatively scarce multi-page document data, and thus is also sub-optimal in terms of performance. Instead, we opt for leveraging powerful single-page DocVQA models, especially pretrained on millions of single-page documents, and finetuning them to the multi-page scenario. For this purpose, we combine concepts of local (page) and global (document) tokens, which promote an exchange of information within and across pages, while keeping computational cost in check. We choose pages as atomic units in our proposed scheme, as page structure often represent a semantic unit in DocVQA." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.43, + 0.47, + 0.718 + ], + "angle": 0, + "content": "We present GRAM (Global ReAsoning for Multi-page VQA), a novel approach for endowing multi-page processing capabilities to existing single-page DocVQA models. Alongside page tokens that encapsulate both textual and visual contents of each page, we introduce doc(ument) learnable tokens, which aim is dispersing global information across all pages. These two sets of tokens interact within our newly-devised two-stage encoder blocks. The initial stage utilizes an existing single-page layer and enhances it by including both page and doc tokens as input, allowing them to freely interact. In the second stage, we prioritize computational efficiency by restricting self-attention solely to the global doc tokens. This global reasoning layer captures collective information from multiple pages, enabling the system to respond to cross-page inquiries, as illustrated in Fig. 1. Considering that doc tokens did not appear in pretraining, to boost their significance during finetune, we employ a designated bias adaptation mechanism which strikes a balance between local and global learnable tokens." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.902 + ], + "angle": 0, + "content": "While our method inherently deals with long sequences, we circumvent a quadratic reliance on sequence length by segmenting the document into pages — its semantically logical parts. We restrict interaction solely among doc learnable tokens, across all pages, thereby mitigating the computational burden of depending quadratically on the page count. Apart from encoding, the auto-regressive decoding stage poses a computational burden in long sequences. To this end, we introduce a compression stage that precedes the decoder, implemented with a compression transformer, termed CFormer. The CFormer receives the concatenated output of all pages and compresses it to a" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.89, + 0.121 + ], + "angle": 0, + "content": "much shorter sequence, distilling the most pertinent information in the document. Our key contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.122, + 0.89, + 0.181 + ], + "angle": 0, + "content": "- We propose GRAM, an approach to endow single-page DocVQA methods with multi-page capabilities, without pretraining, allowing the model to process multi-page documents, while preserving single-page performance." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.183, + 0.89, + 0.24 + ], + "angle": 0, + "content": "- We introduce document learnable tokens and bias adaptation that enable an effective communication and collaboration between individual pages to support reasoning over multiple page documents." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.243, + 0.89, + 0.285 + ], + "angle": 0, + "content": "- Our C-Former module suggests a trade-off between accuracy and compute, distilling information from multi-page sequences into more compact representations." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.288, + 0.89, + 0.33 + ], + "angle": 0, + "content": "- We obtain SOTA results over the MPDocVQA and DUDE datasets, and provide extensive ablations to each component in our method." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.122, + 0.89, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.347, + 0.642, + 0.362 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.373, + 0.892, + 0.719 + ], + "angle": 0, + "content": "Long Sequence Approaches are an active field of research in NLP, aiming to improve the design of chat-systems [24] and image instruction tasks [19, 21]. In these applications, the ability to manage and process long sequences is vital, as conversations cannot be cut short, or limited to just a few interactions. Common approaches to tackle long sequences include sparse attention mechanisms [4, 7, 38] and methods to improve results on long sequences during inference [11, 27, 28]. 'Sliding window' approaches of limiting the range of neighbors each token can attend to, lead to a significant reduction in computation and memory consumption. Prominent works of this kind include Long-Former [7], where each token attends to a set of its nearest neighbors, along with additional global tokens. The work of Big-Bird [38] adds additional non-neighboring tokens at random, whereas Colt5 [4] uses the same sliding window approach, but performs heavier computations for important tokens and shallow operations for filler words or punctuation. Although Tito et. al. [33] have demonstrated that the above approaches do not perform as well on the task of MPDocVQA, we do incorporate the ideas of combining both local and global tokens throughout encoding to expand the attention onto additional pages." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "DocVQA has attracted increasing attention [5, 6, 9, 16, 25, 26, 31, 36, 37] with the introduction of the DocVQA dataset by Mathew et al. [22]. Most methods in DocVQA leverage OCR [1-3, 20, 23] to input both text and layout information (bounding box coordinates and possibly font type) into the model, where some further explore different techniques to combine the two types of data streams, or alternatively, clever schemes of pretraining. DocVQA methods can be roughly divided into two categories: extractive and abstractive. Extractive methods [16, 25, 36, 37] rely on the fact that the explicit answer resides in the written text, thus only output a corresponding text span within the input" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15599" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.144, + 0.098, + 0.422, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.443, + 0.109, + 0.807, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.457, + 0.892, + 0.553 + ], + "angle": 0, + "content": "Figure 2. GRAM Architecture. (a) Depicts a high-level architecture overview. For each page, the visual, textual and question tokens are concatenated together with learnable doc tokens (darker color shade). The processed information is fed into the multi-page encoder. The encoder output can be fed directly into the decoder to create the final prediction. Optionally, a compression model, C-Former, can be used between the encoder and the decoder to compress the encoder output into a predetermined length, thus reducing overall latency for long documents. (b) Shows a global-local encoder layer, containing two sub-layers. The first sub-layer uses self-attention that operates on each page separately, while the second applies a self-attention step on the doc tokens to fuse information between the different pages. The corresponding tokens are then routed back to their respective page and go into the next global-local encoder layer." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.581, + 0.469, + 0.672 + ], + "angle": 0, + "content": "sequence. Abstractive methods [5, 6, 14, 26, 31], on the other hand, have the capacity to generate free-form answers which do not necessarily appear in the text, thus providing flexibility in real-world applications. Notably, existing research in DocVQA does not scale in a straightforward way to deal with the more realistic multi-page scenario." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.47, + 0.903 + ], + "angle": 0, + "content": "MPDocVQA has recently gained momentum with the launch of two new multi-page datasets: MP-DocVQA [33] and DUDE [18], offering two separate recipes to tackle longer documents. The first approach of Tito et. al., referred to as HiVT5 [33], suggests compressing the encoding of each page separately, and feeding the decoder with the concatenation of the compressed outputs from each single-page encoder. While this approach is advantageous in terms of computation, we later show the compression may severely hinder the results. In addition, there is no communication between the single-page encoders until the final stage of decoding, whereas in our method, we allow for exchange of page and document information throughout all stages of encoding. Another prominent approach, proposed by Landeghem et. al. [18], which relatively preserves qual" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.892, + 0.658 + ], + "angle": 0, + "content": "ity, involves concatenating all the pages into one long sequence and feeding it to a standard encoder-decoder structure. This, however, poses a heavy computational burden as transformers' self-attention component scales quadratically with input sequence length." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.671, + 0.587, + 0.687 + ], + "angle": 0, + "content": "3. GRAM" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.697, + 0.673, + 0.71 + ], + "angle": 0, + "content": "3.1. Base Architecture" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The underlying idea in our approach is using existing encoder-decoder single-page models for document understanding and extending them to multi-page scenarios, without additional pretraining. In this work, we provide such a recipe over the notable DocFormerv2 [6]. For the sake of completeness, DocFormerv2 is a T5-based [28] encoder-decoder transformer model which operates over both visual and textual features to support document understanding. Each page is represented by textual features \\(\\mathbf{T} \\in R^{N_t \\times d}\\) which encapsulate OCR tokens and their corresponding 2D bounding box positions [36], along with visual \\(\\mathbf{V} \\in R^{N_v \\times d}\\) and question \\(\\mathbf{Q} \\in R^{N_q \\times d}\\) embeddings. Where \\(N_t\\), \\(N_v\\) and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15600" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.138 + ], + "angle": 0, + "content": "\\(N_{q}\\) are the lengths of the OCR, visual features and the question. The output result \\(\\mathbf{Y}\\) is obtained by passing the concatenated inputs through the encoder-decoder model," + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.15, + 0.469, + 0.168 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Y} = \\mathbf {D} (\\mathbf {E} (\\operatorname {c o n c a t} (\\mathbf {T}, \\mathbf {V}, \\mathbf {Q}))), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.179, + 0.46, + 0.194 + ], + "angle": 0, + "content": "where \\(\\mathbf{E}\\) and \\(\\mathbf{D}\\), are the encoder and decoder, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.195, + 0.471, + 0.391 + ], + "angle": 0, + "content": "Our method uses these basic building blocks in designing a multi-page solution. To this end, we introduce a bi-level global-local encoder, as illustrated in Fig. 2. At the local page-level of each block, we utilize the layers of the existing single-page encoder \\(\\mathbf{E}\\) to process each page separately, together with learnable doc tokens. Next, we introduce a slim global layer in each block that facilitates communication between doc tokens across all pages. This bi-level localized processing ensures the model can understand the content of each page effectively, while also combining information across pages in the document. After \\(M\\) such blocks, we feed the encoded features from all pages into the existing decoder \\(\\mathbf{D}\\) to produce the overall output." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.401, + 0.3, + 0.417 + ], + "angle": 0, + "content": "3.2. Global-Local Reasoning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.425, + 0.47, + 0.576 + ], + "angle": 0, + "content": "To operate on multiple pages we break down the document to \\(K\\) pages, and the single-page encoder to \\(M\\) encoder layers, \\(\\mathbf{E}^j, j = 0,\\dots,M - 1\\). We then construct \\(M\\) blocks, with two sub-layers each. The first page sub-layer originates from the existing pretrained encoder layer, referred to as \\(\\mathbf{E}_{page}^{j}\\), and operates in parallel, with shared weights, for all pages in the document. This layer receives both page and doc tokens. The second, newly introduced, document layer \\(\\mathbf{E}_{doc}^{j}\\) collects only the doc tokens from all pages and promotes sharing information across all of the document." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.577, + 0.469, + 0.636 + ], + "angle": 0, + "content": "Formally, we augment the input of the standard single-page encoder with page-specific indexing \\((\\mathbf{T}_i,\\mathbf{V}_i,\\mathbf{Q})\\) and incorporate page-positional embedding \\(\\mathbf{P}_i\\) to both text and visual features, where \\(i = 0,\\dots,K - 1\\):" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.649, + 0.469, + 0.667 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {T}} _ {i} = \\mathbf {T} _ {i} + \\mathbf {P} _ {i}, \\quad \\tilde {\\mathbf {V}} _ {i} = \\mathbf {V} _ {i} + \\mathbf {P} _ {i}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.679, + 0.469, + 0.757 + ], + "angle": 0, + "content": "Next, we formulate our bi-level global-local block. The input to the first page-level sub-layer in each block is the concatenation of the textual, visual and question features, denoted \\(\\mathbf{X}_i^j = \\mathrm{concat}(\\tilde{\\mathbf{T}}_i,\\tilde{\\mathbf{V}}_i,\\mathbf{Q}_i)\\), along with page-specific doc tokens \\(\\mathbf{G}_i^j\\in R^{N_g\\times d}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.77, + 0.469, + 0.79 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {i} ^ {j + 1}, \\tilde {\\mathbf {G}} _ {i} ^ {j + 1} = \\mathbf {E} _ {p a g e} ^ {j} (\\operatorname {c o n c a t} \\left(\\mathbf {X} _ {i} ^ {j}, \\mathbf {G} _ {i} ^ {j}\\right)). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.797, + 0.469, + 0.871 + ], + "angle": 0, + "content": "Here, the features undergo self-attention, normalization and feed-forward layers. The layer output \\(\\mathbf{X}_i^{j + 1}\\) is passed on as input to the next bi-level block, whereas only the doc tokens \\(\\tilde{\\mathbf{G}}_i^{j + 1}\\), enter the second doc sub-layer, which again includes self-attention, normalization and feed-forward" + }, + { + "type": "equation", + "bbox": [ + 0.122, + 0.883, + 0.469, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\mathbf {G} _ {i} ^ {j + 1} \\right\\} _ {i = 0} ^ {K - 1} = \\mathbf {E} _ {d o c} ^ {j} (\\operatorname {c o n c a t} \\left(\\left\\{\\tilde {\\mathbf {G}} _ {i} ^ {j + 1} \\right\\} _ {i = 0} ^ {K - 1}\\right)). \\tag {4}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.092, + 0.871, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.27, + 0.892, + 0.34 + ], + "angle": 0, + "content": "Figure 3. Global-Local Attention: In long sequence approaches (a), attention is applied jointly to the entire sequence of concatenated local and global tokens. Our method, separates the computation into two steps — page-level (b) and document-level (c)— leveraging the natural division of documents into pages." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.365, + 0.892, + 0.47 + ], + "angle": 0, + "content": "In this stage, the doc tokens can interact and pass information from page to page, after which being passed on to the next block, as depicted in Eq. (3). This design allows information to flow between pages while keeping computational costs in check. When concluding the traversal over \\(M\\) such layers, the outputs across all pages are concatenated and fed to the decoder \\(\\mathbf{D}\\)," + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.479, + 0.891, + 0.498 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Y} _ {m u l t i} = \\mathbf {D} (\\operatorname {c o n c a t} \\left(\\left\\{\\mathbf {X} _ {i} ^ {M}, \\mathbf {G} _ {i} ^ {M} \\right\\} _ {i = 0} ^ {K - 1}\\right)). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.507, + 0.892, + 0.674 + ], + "angle": 0, + "content": "To visualize the difference between the attention masks in our method, we compare it with previous long sequence approaches [4, 7, 38] in Fig. 3. These prior methods optimize computation by using attention masking on nearby tokens and allowing limited global connections. However, naively applying such methods to multi-page documents will treat it as a single stream, which does not consider the division into pages. Our global-local blocks, with a two-stage attention-masking mechanism, better suit multi-page documents. In addition, our two-level design benefits from existing, extensively pretrained single page models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.682, + 0.661, + 0.698 + ], + "angle": 0, + "content": "3.3. Bias Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.901 + ], + "angle": 0, + "content": "An already-pretrained model, introduced with a new stream of data, might disregard it altogether [13, 14, 34, 39]. To overcome this, we force the system to account for the newly-introduced doc tokens by modifying the encoder's bias method. Originally, the bias method intervenes in the attention mechanism, diminishing the relationships between distant tokens. However, in our specific case, the distance between doc and page tokens does not represent their actual relevance. To enforce the encoder to pay closer attention to the doc tokens, we assign them a positive constant bias value. Particularly, we replace the values in the bias matrix, corresponding with the doc tokens, with fixed ones. Instead of a single bias value, we utilize a different value" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15601" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.213 + ], + "angle": 0, + "content": "for each attention head, as performed in ALiBi [27], enabling more fine-grained control of the global features in each head. Specifically, the constant doc bias value is set to \\( c \\cdot \\frac{1}{2^a} \\), where \\( c \\) is a constant and \\( a \\) is the attention head index. This yields a decaying bias value across different attention heads, resulting in hierarchical importance of the document information, where the first heads are more oriented towards doc tokens and the last towards page tokens." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.221, + 0.317, + 0.236 + ], + "angle": 0, + "content": "3.4. Compression Transformer" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.244, + 0.47, + 0.41 + ], + "angle": 0, + "content": "Our global-local solution to MP-DocVQA resolves the problematic quadratic dependency on the number of pages \\( K \\) during encoding. However, the auto-regressive decoding complexity scaling linearly with \\( K \\) also poses a practical challenge during inference time, as we later discuss in Sec. 3.5. To alleviate this burden, we place an optional transformer-based model, named C-Former (Compression TransFormer), between the encoder outputs and the decoder, as depicted in Fig. 2. The C-Former has the ability to revise the information across all pages and distill only the important details, required to correctly answer the question." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.41, + 0.47, + 0.548 + ], + "angle": 0, + "content": "Specifically, the C-Former is a light-weight transformer-based decoder [28], denoted as \\(\\mathbf{D}_C\\), featuring crossattention, layer norm and feed-forward layers in each block. The input to C-Former includes \\(N_{c}\\) learnable tokens \\(\\mathbf{C} \\in R^{N_c \\times d}\\), concatenated with the input question \\(\\tilde{\\mathbf{C}} = \\mathrm{concat}(\\mathbf{C}, \\mathbf{Q})\\). In addition, we feed it with the outputs of the global-local interlaced encoder, concatenated to one long sequence, referred as \\(\\mathbf{O}\\), where \\(\\mathbf{O} = \\mathrm{concat}(\\{\\mathbf{X}_i^M, \\mathbf{G}_i^M\\}_{i=0}^{K-1})\\). The output of C-Former is thus" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.557, + 0.386, + 0.573 + ], + "angle": 0, + "content": "\\[\n\\mathbf {O} _ {C} = \\mathbf {D} _ {C} (Q = \\tilde {\\mathbf {C}}, K = \\mathbf {O}, V = \\mathbf {O}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.47, + 0.659 + ], + "angle": 0, + "content": "where we pass forward only the first set of \\( N_{c} \\) output embeddings and ignore the rest, setting the output sequence dimension to \\( N_{c} \\). C-Former offers flexibility in controlling the tradeoff between ANLS quality and computational efficiency by controlling the output sequence length \\( N_{c} \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.666, + 0.287, + 0.683 + ], + "angle": 0, + "content": "3.5. Computation Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Next, we turn to provide a thorough computational complexity analysis. We consider a document that comprises of \\( K \\) pages, each with \\( N \\) tokens, and the maximum answer length is \\( L \\). For simplicity, we assume that all encoders and decoders have one layer. The naive way to support multi-page documents is using an existing single-page encoder-decoder model, fusing all of the textual page inputs together, and feeding them as one long sequence. We refer to this approach as 'concat'. The self-attention complexity of such a configuration scales quadratically with the sequence length, \\( O((N \\cdot K)^2) \\). Conversely in our method, we operate on the document pages with two alternating encoding stages in each layer. The first stage performs a self-attention over both the page and doc tokens. Hence, the complexity of" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.091, + 0.892, + 0.227 + ], + "angle": 0, + "content": "such sub-layer is \\( O((N + N_g)^2 \\cdot K) \\), where \\( N_g \\) is the number of doc tokens. The second stage features a self-attention operation over the doc tokens, across all pages in the document. The complexity of this operation is \\( O((N_g \\cdot K)^2) \\). Overall, the total complexity for one global-local encoder block is \\( O((N + N_g)^2 \\cdot K + (N_g \\cdot K)^2) \\). Since \\( N_g \\) is a constant, and the number of pages is usually less than the number of words in each page (\\( K < N \\)), we obtain a complexity of \\( O(N^2 \\cdot K) \\), which is not quadratic in \\( K \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.228, + 0.893, + 0.47 + ], + "angle": 0, + "content": "Prior to decoding, the outputs of all per-page encoders are concatenated, thus the output sequence length is \\((N + N_{g})\\cdot K\\) Since the decoder is auto-regressive, its complexity depends quadratically on the maximum output length, \\(L\\) ,namely, \\(O((N + N_g)\\cdot K\\cdot L^2) = O(N\\cdot K\\cdot L^2)\\) Since this operation of decoding is performed iteratively during inference, the combined sequence length \\((N + N_{g})\\cdot K\\) becomes computationally heavy. To alleviate this concern, we propose an optional C-Former model, which performs compression prior to decoding. The overall complexity in this decoding scheme includes passing through the C-Former and then through the decoder, leading to \\(O((N + N_g)\\cdot K\\cdot N_c) + N_c\\cdot L^2)\\) which is equivalent to \\(O(N\\cdot K + L^2)\\) , since \\(N_{c}\\) is a constant, denoting the number of compression tokens in C-Former." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.484, + 0.634, + 0.502 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.51, + 0.707, + 0.526 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.533, + 0.892, + 0.79 + ], + "angle": 0, + "content": "Datasets and Metrics The MPDocVQA dataset [33] features 46K questions, spanning over 48K images, and includes layout elements as figures, tables, lists and diagrams, with printed, handwritten and typewritten text. MPDocVQA contains mostly extractive questions, for which answers are present in the given text. DUDE is smaller in size (23.7K questions over 3K documents), but offers complex questions that require a reader to rationalize beyond the written text content. We report our results using the ANLS metric, introduced in [8], computing a generalized accuracy. Results for DUDE can be broken apart to several types of questions, categorized to four groups: 'extractive' – for which the answer is found directly in the text; 'abstractive' – requiring a free-form answer that does not necessarily appear in the document; 'list of answers' – requiring a list of answers, as opposed to a single one, and 'unanswerable' – where the result cannot be determined using the text." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Implementation Details Our underlying architecture is based on Docformerv2 [6]. Recall, our interlaced encoder features \\(M\\) blocks (12 in 'base' and 24 in 'large'), where each block contains a page sub-layer which originates from an extension of Docformerv2's encoder layer. Every structure contains self-attention, normalization and feed-forward" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15602" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.078, + 0.089, + 0.89, + 0.44 + ], + "angle": 0, + "content": "
MethodParamsMPDocVQADUDE
ANLSANLSExtractiveANLS per Question TypeUnanswerable
AbstractiveList of answers
Longformer [7]148M55.0627.1443.588.5510.6210.78
BigBird [38]131M58.5426.2740.267.118.4612.75
LayoutLMv3 [38]125M55.1320.3132.608.107.828.82
Hi-VT5†beamsearch [15]316M-35.7428.3132.9810.6062.90
Hi-VT5[33]316M62.0123.0617.6033.946.8361.67
Hi-VT5*257M60.7823.867.2116.563.5372.77
DocFormerv2concat [6]257M69.6744.2141.6641.8615.1365.19
GRAMC-Former286M70.8040.0740.4339.6111.4252.55
GRAM281M73.6846.1546.0744.8215.2762.18
T5-2D [18]770M-46.0655.6550.815.4368.62
DocGptVQA [30]>3.5B-50.0251.8648.3228.2262.04
DocBlipVQA [29]>3.5B-47.6250.6946.3130.7355.22
Hi-VT5* [33]784M71.3528.8918.2126.176.8458.99
DocFormerv2concat [6]784M76.4048.4450.8248.0617.6759.04
GRAMC-Former864M77.6045.4747.6344.9114.3456.99
GRAM859M80.3251.1553.6750.3518.4063.23
Hi-VT5*†[33]784M73.5149.1849.2948.3513.3065.95
DocFormerv2†concat [6]784M76.7750.7952.7049.6117.3365.14
GRAM†C-Former864M78.1250.9755.1550.4617.2661.04
GRAM†859M79.6753.3656.8352.3219.9665.43
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.451, + 0.893, + 0.493 + ], + "angle": 0, + "content": "Table 1. Quantitative Results. We present ANLS results for the MPDocVQA [33] and DUDE [18] test sets. The methods are grouped according to the model type and size, starting from encoder-only models (top), T5-base models (middle) and T5-large models (bottom). \\( {}^{ \\dagger } \\) denotes training with both MPDocVQA and DUDE." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.517, + 0.466, + 0.628 + ], + "angle": 0, + "content": "
MethodTraining DataANLS
DocVQAMPDocVQADocVQAMPDocVQA
DocFormerv2concat✓✗86.6072.73
✗✓85.2876.40
✓✓86.4775.37
GRAM✓✗86.7073.12
✗✓85.2980.32
✓✓86.3278.66
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.64, + 0.47, + 0.682 + ], + "angle": 0, + "content": "Table 2. DocVQA vs. MPDocVQA Performance. Test results over both datasets using the large model variants. A checkmark denotes whether a dataset was included in training or not." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.473, + 0.903 + ], + "angle": 0, + "content": "layers. We extend the page layer from Docformerv2 to feature also the doc learnable embeddings. The second doc sub-layer is similar in structure to the first sub-layer, only it is initialized from scratch, with the following specification: \\( d_{ff} = 1024 \\), \\( d_{kv} = 64 \\), \\( nheads = 4 \\), \\( d = 256 \\). We implement 32 doc learnable tokens for each page, uniformly initialized to random values. For bias adaptation, the initial bias value is set to \\( c = 20 \\), with variations between encoder heads, as described in Sec. 3.3. We incorporate an additional optional compression stage using C-Former - a randomly-initialized T5 [28] tiny decoder, with an encoder mask instead of a causal one. The output sequence extracted" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.519, + 0.892, + 0.55 + ], + "angle": 0, + "content": "from C-Former is \\( N_{c} = 256 \\). Finally the decoder is initialized with pretrained weights from DocformerV2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.55, + 0.893, + 0.687 + ], + "angle": 0, + "content": "The model is trained with the Hugging Face Trainer [17] for \\(200k\\) steps, starting with a warm-up of \\(1k\\) steps, with linear learning rate decay. We use learning rates of \\(3e^{-5}\\) and \\(1e^{-4}\\) for the already pretrained encoder and decoder weights, versus the newly initialized doc sub-layer weights. Training is performed on a cluster of \\(8 \\times A100\\) GPUs, each with 40GB of RAM. During training, each page encoder receives 800 tokens, dealing with up to 4 pages. During testing, we increase the maximum length of tokens to 8,000." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Baselines We report the results of previous work on both MP-DocVQA and DUDE datasets (if those exist), including the NLP-based Longformer [7] and BigBird [38], which were adapted to MPDocVQA by [33]; LayoutLMv3 [16], originally designed for DocVQA; and Hi-VT5 [33] and T5-2D [32], specifically suggested for MP-DocVQA Task. We also add for reference the results of methods published in the leader-boards of MPDocVQA and DUDE, which do not have corresponding papers, including DocGptVQA [30], DocBlipVQA [30], and Hi-VT5_beamsearch [15] ([15] was trained on both MP-DocVQA and DUDE). In our approach, we present two variations: GRAM and \\(\\mathrm{GRAM}_{C - \\text{Former}}\\). While GRAM utilizes the full length of the encoder output," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15603" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.12, + 0.09, + 0.258, + 0.105 + ], + "angle": 0, + "content": "How many diagrams are there?" + }, + { + "type": "image_caption", + "bbox": [ + 0.122, + 0.107, + 0.167, + 0.115 + ], + "angle": 0, + "content": "HIVT5:\"2\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.122, + 0.116, + 0.167, + 0.125 + ], + "angle": 0, + "content": "GRAM:\"3\"" + }, + { + "type": "image", + "bbox": [ + 0.128, + 0.131, + 0.371, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.387, + 0.091, + 0.54, + 0.104 + ], + "angle": 0, + "content": "Which month shows the hurricane?" + }, + { + "type": "image_caption", + "bbox": [ + 0.387, + 0.106, + 0.491, + 0.115 + ], + "angle": 0, + "content": "HiVT5:\"not answerable\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.387, + 0.116, + 0.444, + 0.126 + ], + "angle": 0, + "content": "GRAM:\"July\"" + }, + { + "type": "image", + "bbox": [ + 0.388, + 0.136, + 0.577, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.587, + 0.091, + 0.74, + 0.106 + ], + "angle": 0, + "content": "What is the eps code for little rock?" + }, + { + "type": "image_caption", + "bbox": [ + 0.589, + 0.107, + 0.643, + 0.116 + ], + "angle": 0, + "content": "HiVT5:\"EPS\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.589, + 0.117, + 0.649, + 0.126 + ], + "angle": 0, + "content": "GRAM:\"AR01\"" + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.131, + 0.848, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.275, + 0.893, + 0.304 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison between our approach and Hi-VT5 [33] indicate that the integration of our global-local encoder enhances reasoning capabilities, especially when the inquiries require multi-page context." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.329, + 0.471, + 0.36 + ], + "angle": 0, + "content": "\\(\\mathrm{GRAM}_{C - \\text{Former}}\\) allows the user to control the trade-off between performance and latency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.361, + 0.47, + 0.542 + ], + "angle": 0, + "content": "To ensure a fair comparison, since we use the pretrained model of DocFormerv2 [6], we implement two additional baselines, referred to as Hi-VT5* and DocFormerv2concat. The first follows a similar structure as Hi-VT5 [33], with the encoder originating from DocFormerv2, however without the page answer prediction, as it does not exist in DUDE. The second recreates the approach of [18], where only the textual tokens of all pages are concatenated to one long sequence, then passing through the DocFormerv2 model. The second approach poses a computational burden, thus we use only 600 tokens during training per page, with up to 4 pages, and during test only 400 tokens." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.556, + 0.172, + 0.571 + ], + "angle": 0, + "content": "4.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.581, + 0.47, + 0.687 + ], + "angle": 0, + "content": "We present the performance of our method over the MPDocVQA [33] and DUDE [18] datasets in Tab. 1. The methods are divided into three groups: the top contains encoder-only, and methods that rely on the T5-base model (up to \\(316M\\) parameters); the middle section, approaches that use the T5-large model (over \\(770M\\) parameters), and finally the bottom, T5-large models, trained on both datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.688, + 0.47, + 0.868 + ], + "angle": 0, + "content": "As can be seen, in the first group, the encoder-only NLP methods, LongFormer [7], BigBird [38] and LayoutLMv3 [16] can only handle relatively well 'extractive' style tasks as in MPDocVQA dataset [33], but often struggle with 'abstractive' questions that are more abundant in DUDE [18]. As to T5-'base' models, versus our best competitor Docformerv2concat, we obtain an improvement of \\((+4\\%, +1.9\\%)\\) on MP-DocVQA and DUDE datasets. As to methods that combine an additional compression before decoding (Hi-VT5, Hi-VT5*), our C-Former achieves an increase in \\((+8.8\\%, +16.2\\%)\\) over the best candidates on the MP-DocVQA and DUDE datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.471, + 0.902 + ], + "angle": 0, + "content": "As for the group of 'large' models, we include the results of T5-2D [18] DocGptVQA [30] and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.329, + 0.893, + 0.42 + ], + "angle": 0, + "content": "DocBlipVQA [29]. Note that our model surpasses DocFormerconcat, the primary baseline, achieving improvements of \\((+3.9\\%, +2.7\\%)\\) on MP-DocVQA and DUDE, respectively. We also outperform DocGptVQA [30], a method that appears in the leaderboard of DUDE, by \\(+1.1\\%\\), thereby obtaining SOTA results for GRAM 'large'." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.423, + 0.893, + 0.574 + ], + "angle": 0, + "content": "The final category showcases large encoder-decoder models, fine-tuned on both MP-DocVQA and DUDE training sets, showcasing the benefits of augmented training data. GRAM consistently demonstrates performance gains over the baseline, illustrating its robustness across different datasets and training scenarios. Next, we present in Tab. 2 the effect of training on DocVQA vs. MPDocVQA. Our method achieves performance on-par on the single page task, while enhancing performance on the multi-page scenario by \\(+3.3\\%\\), compared to the baseline." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.577, + 0.892, + 0.684 + ], + "angle": 0, + "content": "In Fig. 4, We show qualitative results on the DUDE dataset of GRAM versus Hi-VT5* [33]. Our method demonstrates proficiency in addressing questions that involve attention over multiple pages ('how many diagrams are there'), an increased visual analysis capability ('Which month shows the hurricane?'), and heightened abstractive ability ('What is the EPS code for Little Rock?')." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.706, + 0.652, + 0.723 + ], + "angle": 0, + "content": "5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.893, + 0.901 + ], + "angle": 0, + "content": "We perform an ablation study on our approach, evaluating the influence of each constituent component using DUDE's validation set [18]. This validation set enables the grouping of documents by their respective page counts: 1, 2-4, 5-10, 11-end, encompassing 1747, 2259, 1062, 1241 samples in each category, respectively. Our investigation delves into the impact of the number of doc tokens and the bias adaptation methods. Moreover, we employ the C-Former for sequence compression, adjusting the compression ratio and examining the balance between performance and latency (see supplementary for more details)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15604" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.468, + 0.294 + ], + "angle": 0, + "content": "
#Doc TokensBias TypeCompression DimensionANLS by Number of Pages DUDE validation dataset
All12-45-1011-end
XXX46.1647.1848.6643.3442.57
16X46.3948.3549.0643.1641.56
32DecayingX47.8849.2949.9045.9043.94
64X46.7047.9849.2244.0042.60
32XX47.5249.8549.9344.9042.10
ConstantX46.1447.4148.1344.4442.19
DecayingX47.8849.2949.9045.9043.94
32Decaying839.8339.7341.4136.9839.52
3240.4240.9541.6438.1239.39
25641.9942.5743.7738.4041.01
102442.5642.9744.3038.9441.93
409643.5944.7544.6440.5442.67
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.305, + 0.472, + 0.348 + ], + "angle": 0, + "content": "Table 3. GRAM Ablation Study. Results on DUDE validation set ablating over (a) the dimension of doc tokens, (b) the attention bias employed and (c) the C-former input dimension." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.377, + 0.472, + 0.725 + ], + "angle": 0, + "content": "GRAM Components We focus our initial exploration on the impact of the number of doc tokens \\( N_{g} \\). As can be seen in Tab. 3, while \\( N_{g} = 16 \\) leads to performance on par with not using doc tokens at all, for the optimal value of \\( N_{g} = 32 \\), we obtain an increase of \\( +1.7\\% \\) in ANLS. Shifting our focus to bias adaptation methods, Tab. 3 shows that using constant bias has a negative effect on the results, suggesting this method is not flexible enough in maintaining a balance between the page and doc tokens. However, our decaying bias-adaptation approach does improve results overall, versus no-bias \\( (+0.36\\%) \\), especially for longer documents \\( (+1\\% \\) improvement for 5-10 pages and \\( +1.84\\% \\) for 11 pages and more). This is to be expected, since incorporating new doc tokens and increasing their importance can potentially affect single-page performance. Finally, in Tab. 4, we reinforce our choice of pages as semantic logical units for MPDocVQA. We first ablate our method with and without page embedding. Next, we compare our page-based division with varying fixed-length division of tokens for encoder. Results in Tab. 4 clearly demonstrate an advantage towards page-level encoding in MPDocVQA. This aligns with our initial assumption that structured documents are often designed with page-division in mind." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Performance-Latency Trade-off We assess the impact of C-Former on performance, considering compression output lengths of 8, 32, 256, 1024, 4096. Note, performance gradually improves with an increase in the compression output length. However, longer output lengths correspond to heightened model latency. Note that using C-Former for shorter documents can be redundant, as there is little to no compression compared to the input sequence length and results decrease. In Fig. 5, we scrutinize the trade-off between computational efficiency and compression rate" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.089, + 0.891, + 0.212 + ], + "angle": 0, + "content": "
Page EmbeddingSegment LengthANLS by Number of Pages DUDE validation dataset
All12-45-1011-end
✓✗47.8849.2949.9045.9043.94
✗✗46.1248.7448.1143.5940.99
✓25645.2246.3846.6944.1341.83
✓51245.0945.9047.3242.6541.98
✓102444.3944.9846.6341.6941.78
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.221, + 0.892, + 0.264 + ], + "angle": 0, + "content": "Table 4. The Significance of Pages as Semantic Units. Results on DUDE validation set ablating over (a) utilization of page-embedding, (b) segment length for fixed-size encoding inputs." + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.291, + 0.871, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.445, + 0.892, + 0.487 + ], + "angle": 0, + "content": "Figure 5. Latency comparison. We compare the dependency between overall latency and the number of pages in input document for GRAM, \\(\\mathrm{GRAM}_{C - \\text{Former}}\\), DocFormerv2concat and Hi-VT5." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.523, + 0.892, + 0.646 + ], + "angle": 0, + "content": "by comparing to DocFormerv2concat [6] and Hi-VT5* [33]. We discover that DocFormerv2concat reaches a memory limit at approximately 20 pages, due to its quadratic memory increase with sequence length. At this juncture, GRAM\\(_{C-Former}\\) surpasses DocFormerv2concat by performing 3.5 seconds faster. Notably, GRAM\\(_{C-Former}\\) can gracefully handle documents surpassing 300 pages, effectively bridging the gap between performance and latency." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.674, + 0.628, + 0.69 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Our method, termed GRAM, extends existing single-page document models to efficiently handle multi-page documents without necessitating computationally-intensive pretraining. Leveraging the single-page encoder for local page-level comprehension, we introduce document learnable tokens and designated layers, enabling seamless information exchange across pages. Additionally, our proposed bias adaptation method enforces effective utilization of our newly introduced document tokens. The incorporation of a C-Former model reduces sequence length, balancing quality with latency in the decoding step. Extensive experiments demonstrate GRAM's state-of-the-art performance across multi-page DocVQA benchmarks." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15605" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.197 + ], + "angle": 0, + "content": "[1] Aviad Aberdam, Ron Litman, Shahar Tsiper, Oron Anschel, Ron Slossberg, Shai Mazor, R Manmatha, and Pietro Perona. Sequence-to-sequence contrastive learning for text recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15302-15312, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.47, + 0.242 + ], + "angle": 0, + "content": "[2] Aviad Aberdam, Roy Ganz, Shai Mazor, and Ron Litman. Multimodal semi-supervised learning for text recognition. arXiv preprint arXiv:2205.03873, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.244, + 0.469, + 0.299 + ], + "angle": 0, + "content": "[3] Aviad Aberdam, David Bensaïd, Alona Golts, Roy Ganz, Oren Nuriel, Royee Tichauer, Shai Mazor, and Ron Litman. Clipter: Looking at the bigger picture in scene text recognition. arXiv preprint arXiv:2301.07464, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.301, + 0.469, + 0.37 + ], + "angle": 0, + "content": "[4] Joshua Ainslie, Tao Lei, Michiel de Jong, Santiago Ontañón, Siddhartha Brahma, Yury Zemlyanskiy, David Uthus, Mandy Guo, James Lee-Thorp, Yi Tay, et al. Colt5: Faster long-range transformers with conditional computation. arXiv preprint arXiv:2303.09752, 2023. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.372, + 0.47, + 0.442 + ], + "angle": 0, + "content": "[5] Srikar Appalaraju, Bhavan Jasani, Bhargava Urala Kota, Yusheng Xie, and R Manmatha. Docformer: End-to-end transformer for document understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 993-1003, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.444, + 0.469, + 0.498 + ], + "angle": 0, + "content": "[6] Srikar Appalaraju, Peng Tang, Qi Dong, Nishant Sankaran, Yichu Zhou, and R Manmatha. Docformerv2: Local features for document understanding. arXiv preprint arXiv:2306.01733, 2023. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.5, + 0.469, + 0.542 + ], + "angle": 0, + "content": "[7] Iz Beltagy, Matthew E Peters, and Arman Cohan. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150, 2020. 1, 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.544, + 0.469, + 0.613 + ], + "angle": 0, + "content": "[8] Ali Furkan Biten, Ruben Tito, Andres Mafla, Lluis Gomez, Marçal Rusinol, Ernest Valveny, CV Jawahar, and Dimosthenis Karatzas. Scene text visual question answering. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4291-4301, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.615, + 0.469, + 0.657 + ], + "angle": 0, + "content": "[9] Ali Furkan Biten, Ron Litman, Yusheng Xie, Srikar Appalaraju, and R Manmatha. Latr: Layout-aware transformer for scene-text vqa. pages 16548-16558, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.659, + 0.469, + 0.699 + ], + "angle": 0, + "content": "[10] Aydar Bulatov, Yuri Kuratov, and Mikhail S Burtsev. Scaling transformer to 1m tokens and beyond with rmt. arXiv preprint arXiv:2304.11062, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.701, + 0.469, + 0.757 + ], + "angle": 0, + "content": "[11] Shouyuan Chen, Sherman Wong, Liangjian Chen, and Yuandong Tian. Extending context window of large language models via positional interpolation. arXiv preprint arXiv:2306.15595, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.758, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[12] Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc V Le, and Ruslan Salakhutdinov. Transformer-xl: Attentive language models beyond a fixed-length context. arXiv preprint arXiv:1901.02860, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.469, + 0.856 + ], + "angle": 0, + "content": "[13] Robert M French. Catastrophic forgetting in connectionist networks. Trends in cognitive sciences, 3(4):128-135, 1999. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[14] Roy Ganz, Oren Nuriel, Aviad Aberdam, Yair Kittenplon, Shai Mazor, and Ron Litman. Towards models that can see and read. arXiv preprint arXiv:2301.07389, 2023. 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[15] JiangLong He, Mamatha N, Shiv Vignesh, and Deepak Kumar. Hivt5beam. Hi-VT5 model pretrained with private custom document collection using span masking objective. Pretrained model is then trained with DUDE dataset and MultiPage DocVQA dataset, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[16] Yupan Huang, Tengchao Lv, Lei Cui, Yutong Lu, and Furu Wei. Layoutmv3: Pre-training for document ai with unified text and image masking. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4083-4091, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[17] HuggingFace. Huggingface trainer. https://huggingface.co/docs/transformers/training. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.264, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[18] Jordy Landeghem, Rubén Tito, Łukasz Borchmann, Michal Pietruszka, Paweł Józiak, Rafał Powalski, Dawid Jurkiewicz, Mickaël Coustaty, Bertrand Ackaert, Ernest Valveny, et al. Document understanding dataset and evaluation (dude). arXiv preprint arXiv:2305.08455, 2023. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[19] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.391, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[20] Ron Litman, Oron Anschel, Shahar Tsiper, Roeit Litman, Shai Mazor, and R Manmatha. Scatter: selective context attentional scene text recognizer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11962-11972, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.462, + 0.892, + 0.502 + ], + "angle": 0, + "content": "[21] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.504, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[22] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.561, + 0.892, + 0.617 + ], + "angle": 0, + "content": "[23] Oren Nuriel, Sharon Fogel, and Ron Litman. Textadain: Paying attention to shortcut learning in text recognizers. In European Conference on Computer Vision, pages 427-445. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.619, + 0.892, + 0.645 + ], + "angle": 0, + "content": "[24] R OpenAI. Gpt-4 technical report. arXiv, pages 2303–08774, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.648, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[25] Qiming Peng, Yinxu Pan, Wenjin Wang, Bin Luo, Zhenyu Zhang, Zhengjie Huang, Teng Hu, Weichong Yin, Yongfeng Chen, Yin Zhang, et al. Ernie-Layout: Layout knowledge enhanced pre-training for visually-rich document understanding. arXiv preprint arXiv:2210.06155, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[26] Rafal Powalski, Łukasz Borchmann, Dawid Jurkiewicz, Tomasz Dwojak, Michal Pietruszka, and Gabriela Pałka. Going full-tilt boogie on document understanding with text-image layout transformer. In Document Analysis and Recognition-ICDAR 2021: 16th International Conference, Lausanne, Switzerland, September 5–10, 2021, Proceedings, Part II 16, pages 732–747. Springer, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.817, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[27] Ofir Press, Noah A Smith, and Mike Lewis. Train short, test long: Attention with linear biases enables input length extrapolation. arXiv preprint arXiv:2108.12409, 2021. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[28] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15606" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485-5551, 2020. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.234 + ], + "angle": 0, + "content": "[29] RenZhou, QiaolingDeng, XinfengChang, LuyanWang, XiaochenHu, HuiLi, and YaqiangWu. Docclipvqa. We integrated the prediction outputs from the UDOP model and Blip2 to enhance our results, and we optimized the image encoder and included page number features to address the challenge of multi-page documents. GPT to generate python-like modular programs., 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.471, + 0.331 + ], + "angle": 0, + "content": "[30] RenZhou, QiaolingDeng, XinfengChang, LuyanWang, XiaochenHu, HuiLi, and YaqiangWu. Docgptvqa. We integrated the prediction outputs from the UDOP model and Blip2 to enhance our results, and we optimized the image encoder and included page number features to address the challenge of multi-page documents. GPT to generate python-like modular programs., 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.333, + 0.47, + 0.416 + ], + "angle": 0, + "content": "[31] Zineng Tang, Ziyi Yang, Guoxin Wang, Yuwei Fang, Yang Liu, Chenguang Zhu, Michael Zeng, Cha Zhang, and Mohit Bansal. Unifying vision, text, and layout for universal document processing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19254-19264, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.418, + 0.47, + 0.488 + ], + "angle": 0, + "content": "[32] Rubén Tito, Dimosthenis Karatzas, and Ernest Valveny. Document collection visual question answering. In Document Analysis and Recognition-ICDAR 2021: 16th International Conference, Lausanne, Switzerland, September 5–10, 2021, Proceedings, Part II 16, pages 778–792. Springer, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.489, + 0.469, + 0.531 + ], + "angle": 0, + "content": "[33] Rubén Tito, Dimosthenis Karatzas, and Ernest Valveny. Hierarchical multimodal transformers for multi-page docvqa. arXiv preprint arXiv:2212.05935, 2022. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.533, + 0.47, + 0.588 + ], + "angle": 0, + "content": "[34] Maria Tsimpoukelli, Jacob L Menick, Serkan Cabi, SM Eslami, Oriol Vinyls, and Felix Hill. Multimodal few-shot learning with frozen language models. Advances in Neural Information Processing Systems, 34:200-212, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.59, + 0.47, + 0.645 + ], + "angle": 0, + "content": "[35] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.47, + 0.716 + ], + "angle": 0, + "content": "[36] Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, and Ming Zhou. Layoutm: Pre-training of text and layout for document image understanding. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 1192-1200, 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.47, + 0.786 + ], + "angle": 0, + "content": "[37] Yang Xu, Yiheng Xu, Tengchao Lv, Lei Cui, Furu Wei, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Wanxiang Che, et al. Layoutmv2: Multi-modal pre-training for visually-rich document understanding. arXiv preprint arXiv:2012.14740, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.47, + 0.87 + ], + "angle": 0, + "content": "[38] Manzil Zaheer, Guru Guruganesh, Kumar Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontanon, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, et al. Big bird: Transformers for longer sequences. Advances in neural information processing systems, 33:17283-17297, 2020. 1, 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[39] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.092, + 0.892, + 0.135 + ], + "angle": 0, + "content": "Lit: Zero-shot transfer with locked-image text tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18123-18133, 2022. 4" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15607" + } + ] +] \ No newline at end of file diff --git a/2024/GRAM_ Global Reasoning for Multi-Page VQA/bd3538ed-fd76-4274-b7c4-f2ff80c7e1bb_origin.pdf b/2024/GRAM_ Global Reasoning for Multi-Page VQA/bd3538ed-fd76-4274-b7c4-f2ff80c7e1bb_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ca7f7e83633232d7bb0d9e243ea5914374adc968 --- /dev/null +++ b/2024/GRAM_ Global Reasoning for Multi-Page VQA/bd3538ed-fd76-4274-b7c4-f2ff80c7e1bb_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c2ecfa9045d305286f1866b67042796ac26e175a964c4e5e5701e36370408ae1 +size 8920483 diff --git a/2024/GRAM_ Global Reasoning for Multi-Page VQA/full.md b/2024/GRAM_ Global Reasoning for Multi-Page VQA/full.md new file mode 100644 index 0000000000000000000000000000000000000000..765e4fcaee2cd7b83e09dfeb4c08ea720c77125a --- /dev/null +++ b/2024/GRAM_ Global Reasoning for Multi-Page VQA/full.md @@ -0,0 +1,276 @@ +# GRAM: Global Reasoning for Multi-Page VQA + +Tsachi Blau* Technion, Israel + +Sharon Fogel +AWS AI Labs + +Roi Ronen* Technion, Israel + +Alona Golts† +AWS AI Labs + +Shahar Tsiper +AWS AI Labs + +Elad Ben Avraham +AWS AI Labs + +Aviad Aberdam +AWS AI Labs + +Roy Ganz* Technion, Israel + +Ron Litman\* AWS AI Labs + +# Abstract + +The increasing use of transformer-based large language models brings forward the challenge of processing long sequences. In document visual question answering (DocVQA), leading methods focus on the single-page setting, while documents can span hundreds of pages. We present GRAM, a method that seamlessly extends pretrained single-page models to the multi-page setting, without requiring computationally-heavy pretraining. To do so, we leverage a single-page encoder for local page-level understanding, and enhance it with document-level designated layers and learnable tokens, facilitating the flow of information across pages for global reasoning. To enforce our model to utilize the newly introduced document tokens, we propose a tailored bias adaptation method. For additional computational savings during decoding, we introduce an optional compression stage using our compression-transformer(C-Former), reducing the encoded sequence length, thereby allowing a tradeoff between quality and latency. Extensive experiments showcase GRAM's state-of-the-art performance on the benchmarks for multi-page DocVQA, demonstrating the effectiveness of our approach. + +# 1. Introduction + +Document understanding, particularly in the context of DocVQA, has gained substantial research interest [5, 6, 16, 25, 36, 37] and offers a wide array of practical applications, focusing on data extraction and analysis of single page documents. However, Multi/Page DocVQA (MPDocVQA) poses a more realistic challenge, considering that the majority of documents, including contracts, manuals + +![](images/b2ea6080c27f9abf9ddb071d36d1e077d7390878872fde0dff61d4993c023edd.jpg) +Figure 1. An Overview of GRAM. We suggest an interleaved encoder architecture combining page- with document-attention layers, allowing information to propagate between different pages. An optional compression transformer (C-former) is introduced to allow a trade-off between quality and latency. + +and scientific papers, often extend well beyond a single page. Despite the practical relevance of MPDocVQA, it has received limited attention, primarily due to the absence of suitable datasets. Two recently introduced datasets, MPDocVQA [33] and DUDE [18], have opened up new avenues for MP-DocVQA research. + +Recent DocVQA approaches rely on transformers [35], at the heart of their architecture. While transformers are a powerful tool, they face challenges when dealing with long input sequences [4, 7, 10-12, 27, 38]. This difficulty stems from the self-attention mechanism, which scales quadratically in terms of computation and memory, with respect to the input sequence length. Addressing this limitation has attracted significant research efforts, primarily in the field + +of natural language processing (NLP). Proposed NLP-based solutions can be divided into two main directions: The former aims to modify the attention mechanism to cut computational costs [4, 7, 38]. The latter involves altering the positional embedding mechanism to improve performance on longer sequences, with minimal fine-tuning [11, 12, 27]. + +A possible option of tackling MPDocVQA is to extend NLP-based approaches to handle multi-modal document data, including visual representations, along with OCR text and corresponding 2D locations and relative page position. However, this requires extensive pre-training, with relatively scarce multi-page document data, and thus is also sub-optimal in terms of performance. Instead, we opt for leveraging powerful single-page DocVQA models, especially pretrained on millions of single-page documents, and finetuning them to the multi-page scenario. For this purpose, we combine concepts of local (page) and global (document) tokens, which promote an exchange of information within and across pages, while keeping computational cost in check. We choose pages as atomic units in our proposed scheme, as page structure often represent a semantic unit in DocVQA. + +We present GRAM (Global ReAsoning for Multi-page VQA), a novel approach for endowing multi-page processing capabilities to existing single-page DocVQA models. Alongside page tokens that encapsulate both textual and visual contents of each page, we introduce doc(ument) learnable tokens, which aim is dispersing global information across all pages. These two sets of tokens interact within our newly-devised two-stage encoder blocks. The initial stage utilizes an existing single-page layer and enhances it by including both page and doc tokens as input, allowing them to freely interact. In the second stage, we prioritize computational efficiency by restricting self-attention solely to the global doc tokens. This global reasoning layer captures collective information from multiple pages, enabling the system to respond to cross-page inquiries, as illustrated in Fig. 1. Considering that doc tokens did not appear in pretraining, to boost their significance during finetune, we employ a designated bias adaptation mechanism which strikes a balance between local and global learnable tokens. + +While our method inherently deals with long sequences, we circumvent a quadratic reliance on sequence length by segmenting the document into pages — its semantically logical parts. We restrict interaction solely among doc learnable tokens, across all pages, thereby mitigating the computational burden of depending quadratically on the page count. Apart from encoding, the auto-regressive decoding stage poses a computational burden in long sequences. To this end, we introduce a compression stage that precedes the decoder, implemented with a compression transformer, termed CFormer. The CFormer receives the concatenated output of all pages and compresses it to a + +much shorter sequence, distilling the most pertinent information in the document. Our key contributions are: + +- We propose GRAM, an approach to endow single-page DocVQA methods with multi-page capabilities, without pretraining, allowing the model to process multi-page documents, while preserving single-page performance. +- We introduce document learnable tokens and bias adaptation that enable an effective communication and collaboration between individual pages to support reasoning over multiple page documents. +- Our C-Former module suggests a trade-off between accuracy and compute, distilling information from multi-page sequences into more compact representations. +- We obtain SOTA results over the MPDocVQA and DUDE datasets, and provide extensive ablations to each component in our method. + +# 2. Related Work + +Long Sequence Approaches are an active field of research in NLP, aiming to improve the design of chat-systems [24] and image instruction tasks [19, 21]. In these applications, the ability to manage and process long sequences is vital, as conversations cannot be cut short, or limited to just a few interactions. Common approaches to tackle long sequences include sparse attention mechanisms [4, 7, 38] and methods to improve results on long sequences during inference [11, 27, 28]. 'Sliding window' approaches of limiting the range of neighbors each token can attend to, lead to a significant reduction in computation and memory consumption. Prominent works of this kind include Long-Former [7], where each token attends to a set of its nearest neighbors, along with additional global tokens. The work of Big-Bird [38] adds additional non-neighboring tokens at random, whereas Colt5 [4] uses the same sliding window approach, but performs heavier computations for important tokens and shallow operations for filler words or punctuation. Although Tito et. al. [33] have demonstrated that the above approaches do not perform as well on the task of MPDocVQA, we do incorporate the ideas of combining both local and global tokens throughout encoding to expand the attention onto additional pages. + +DocVQA has attracted increasing attention [5, 6, 9, 16, 25, 26, 31, 36, 37] with the introduction of the DocVQA dataset by Mathew et al. [22]. Most methods in DocVQA leverage OCR [1-3, 20, 23] to input both text and layout information (bounding box coordinates and possibly font type) into the model, where some further explore different techniques to combine the two types of data streams, or alternatively, clever schemes of pretraining. DocVQA methods can be roughly divided into two categories: extractive and abstractive. Extractive methods [16, 25, 36, 37] rely on the fact that the explicit answer resides in the written text, thus only output a corresponding text span within the input + +![](images/4a1cf9abb9b129b296325ad8d617ec7270c8af4ec0258570380b4af9be277ec6.jpg) +Figure 2. GRAM Architecture. (a) Depicts a high-level architecture overview. For each page, the visual, textual and question tokens are concatenated together with learnable doc tokens (darker color shade). The processed information is fed into the multi-page encoder. The encoder output can be fed directly into the decoder to create the final prediction. Optionally, a compression model, C-Former, can be used between the encoder and the decoder to compress the encoder output into a predetermined length, thus reducing overall latency for long documents. (b) Shows a global-local encoder layer, containing two sub-layers. The first sub-layer uses self-attention that operates on each page separately, while the second applies a self-attention step on the doc tokens to fuse information between the different pages. The corresponding tokens are then routed back to their respective page and go into the next global-local encoder layer. + +![](images/0d71025096650b205a82653e1b09dd7a52889303abd1cfb63fbac9d494a5e79a.jpg) + +sequence. Abstractive methods [5, 6, 14, 26, 31], on the other hand, have the capacity to generate free-form answers which do not necessarily appear in the text, thus providing flexibility in real-world applications. Notably, existing research in DocVQA does not scale in a straightforward way to deal with the more realistic multi-page scenario. + +MPDocVQA has recently gained momentum with the launch of two new multi-page datasets: MP-DocVQA [33] and DUDE [18], offering two separate recipes to tackle longer documents. The first approach of Tito et. al., referred to as HiVT5 [33], suggests compressing the encoding of each page separately, and feeding the decoder with the concatenation of the compressed outputs from each single-page encoder. While this approach is advantageous in terms of computation, we later show the compression may severely hinder the results. In addition, there is no communication between the single-page encoders until the final stage of decoding, whereas in our method, we allow for exchange of page and document information throughout all stages of encoding. Another prominent approach, proposed by Landeghem et. al. [18], which relatively preserves qual + +ity, involves concatenating all the pages into one long sequence and feeding it to a standard encoder-decoder structure. This, however, poses a heavy computational burden as transformers' self-attention component scales quadratically with input sequence length. + +# 3. GRAM + +# 3.1. Base Architecture + +The underlying idea in our approach is using existing encoder-decoder single-page models for document understanding and extending them to multi-page scenarios, without additional pretraining. In this work, we provide such a recipe over the notable DocFormerv2 [6]. For the sake of completeness, DocFormerv2 is a T5-based [28] encoder-decoder transformer model which operates over both visual and textual features to support document understanding. Each page is represented by textual features $\mathbf{T} \in R^{N_t \times d}$ which encapsulate OCR tokens and their corresponding 2D bounding box positions [36], along with visual $\mathbf{V} \in R^{N_v \times d}$ and question $\mathbf{Q} \in R^{N_q \times d}$ embeddings. Where $N_t$ , $N_v$ and + +$N_{q}$ are the lengths of the OCR, visual features and the question. The output result $\mathbf{Y}$ is obtained by passing the concatenated inputs through the encoder-decoder model, + +$$ +\mathbf {Y} = \mathbf {D} (\mathbf {E} (\operatorname {c o n c a t} (\mathbf {T}, \mathbf {V}, \mathbf {Q}))), \tag {1} +$$ + +where $\mathbf{E}$ and $\mathbf{D}$ , are the encoder and decoder, respectively. + +Our method uses these basic building blocks in designing a multi-page solution. To this end, we introduce a bi-level global-local encoder, as illustrated in Fig. 2. At the local page-level of each block, we utilize the layers of the existing single-page encoder $\mathbf{E}$ to process each page separately, together with learnable doc tokens. Next, we introduce a slim global layer in each block that facilitates communication between doc tokens across all pages. This bi-level localized processing ensures the model can understand the content of each page effectively, while also combining information across pages in the document. After $M$ such blocks, we feed the encoded features from all pages into the existing decoder $\mathbf{D}$ to produce the overall output. + +# 3.2. Global-Local Reasoning + +To operate on multiple pages we break down the document to $K$ pages, and the single-page encoder to $M$ encoder layers, $\mathbf{E}^j, j = 0,\dots,M - 1$ . We then construct $M$ blocks, with two sub-layers each. The first page sub-layer originates from the existing pretrained encoder layer, referred to as $\mathbf{E}_{page}^{j}$ , and operates in parallel, with shared weights, for all pages in the document. This layer receives both page and doc tokens. The second, newly introduced, document layer $\mathbf{E}_{doc}^{j}$ collects only the doc tokens from all pages and promotes sharing information across all of the document. + +Formally, we augment the input of the standard single-page encoder with page-specific indexing $(\mathbf{T}_i,\mathbf{V}_i,\mathbf{Q})$ and incorporate page-positional embedding $\mathbf{P}_i$ to both text and visual features, where $i = 0,\dots,K - 1$ : + +$$ +\tilde {\mathbf {T}} _ {i} = \mathbf {T} _ {i} + \mathbf {P} _ {i}, \quad \tilde {\mathbf {V}} _ {i} = \mathbf {V} _ {i} + \mathbf {P} _ {i}. \tag {2} +$$ + +Next, we formulate our bi-level global-local block. The input to the first page-level sub-layer in each block is the concatenation of the textual, visual and question features, denoted $\mathbf{X}_i^j = \mathrm{concat}(\tilde{\mathbf{T}}_i,\tilde{\mathbf{V}}_i,\mathbf{Q}_i)$ , along with page-specific doc tokens $\mathbf{G}_i^j\in R^{N_g\times d}$ + +$$ +\mathbf {X} _ {i} ^ {j + 1}, \tilde {\mathbf {G}} _ {i} ^ {j + 1} = \mathbf {E} _ {p a g e} ^ {j} (\operatorname {c o n c a t} \left(\mathbf {X} _ {i} ^ {j}, \mathbf {G} _ {i} ^ {j}\right)). \tag {3} +$$ + +Here, the features undergo self-attention, normalization and feed-forward layers. The layer output $\mathbf{X}_i^{j + 1}$ is passed on as input to the next bi-level block, whereas only the doc tokens $\tilde{\mathbf{G}}_i^{j + 1}$ , enter the second doc sub-layer, which again includes self-attention, normalization and feed-forward + +$$ +\left\{\mathbf {G} _ {i} ^ {j + 1} \right\} _ {i = 0} ^ {K - 1} = \mathbf {E} _ {d o c} ^ {j} (\operatorname {c o n c a t} \left(\left\{\tilde {\mathbf {G}} _ {i} ^ {j + 1} \right\} _ {i = 0} ^ {K - 1}\right)). \tag {4} +$$ + +![](images/980915e4e5e177aa55bc25e81b3ce9b2634298bf14abc038ccfc2dc77cbaf0fc.jpg) +Figure 3. Global-Local Attention: In long sequence approaches (a), attention is applied jointly to the entire sequence of concatenated local and global tokens. Our method, separates the computation into two steps — page-level (b) and document-level (c)— leveraging the natural division of documents into pages. + +In this stage, the doc tokens can interact and pass information from page to page, after which being passed on to the next block, as depicted in Eq. (3). This design allows information to flow between pages while keeping computational costs in check. When concluding the traversal over $M$ such layers, the outputs across all pages are concatenated and fed to the decoder $\mathbf{D}$ , + +$$ +\mathbf {Y} _ {m u l t i} = \mathbf {D} (\operatorname {c o n c a t} \left(\left\{\mathbf {X} _ {i} ^ {M}, \mathbf {G} _ {i} ^ {M} \right\} _ {i = 0} ^ {K - 1}\right)). \tag {5} +$$ + +To visualize the difference between the attention masks in our method, we compare it with previous long sequence approaches [4, 7, 38] in Fig. 3. These prior methods optimize computation by using attention masking on nearby tokens and allowing limited global connections. However, naively applying such methods to multi-page documents will treat it as a single stream, which does not consider the division into pages. Our global-local blocks, with a two-stage attention-masking mechanism, better suit multi-page documents. In addition, our two-level design benefits from existing, extensively pretrained single page models. + +# 3.3. Bias Adaptation + +An already-pretrained model, introduced with a new stream of data, might disregard it altogether [13, 14, 34, 39]. To overcome this, we force the system to account for the newly-introduced doc tokens by modifying the encoder's bias method. Originally, the bias method intervenes in the attention mechanism, diminishing the relationships between distant tokens. However, in our specific case, the distance between doc and page tokens does not represent their actual relevance. To enforce the encoder to pay closer attention to the doc tokens, we assign them a positive constant bias value. Particularly, we replace the values in the bias matrix, corresponding with the doc tokens, with fixed ones. Instead of a single bias value, we utilize a different value + +for each attention head, as performed in ALiBi [27], enabling more fine-grained control of the global features in each head. Specifically, the constant doc bias value is set to $c \cdot \frac{1}{2^a}$ , where $c$ is a constant and $a$ is the attention head index. This yields a decaying bias value across different attention heads, resulting in hierarchical importance of the document information, where the first heads are more oriented towards doc tokens and the last towards page tokens. + +# 3.4. Compression Transformer + +Our global-local solution to MP-DocVQA resolves the problematic quadratic dependency on the number of pages $K$ during encoding. However, the auto-regressive decoding complexity scaling linearly with $K$ also poses a practical challenge during inference time, as we later discuss in Sec. 3.5. To alleviate this burden, we place an optional transformer-based model, named C-Former (Compression TransFormer), between the encoder outputs and the decoder, as depicted in Fig. 2. The C-Former has the ability to revise the information across all pages and distill only the important details, required to correctly answer the question. + +Specifically, the C-Former is a light-weight transformer-based decoder [28], denoted as $\mathbf{D}_C$ , featuring crossattention, layer norm and feed-forward layers in each block. The input to C-Former includes $N_{c}$ learnable tokens $\mathbf{C} \in R^{N_c \times d}$ , concatenated with the input question $\tilde{\mathbf{C}} = \mathrm{concat}(\mathbf{C}, \mathbf{Q})$ . In addition, we feed it with the outputs of the global-local interlaced encoder, concatenated to one long sequence, referred as $\mathbf{O}$ , where $\mathbf{O} = \mathrm{concat}(\{\mathbf{X}_i^M, \mathbf{G}_i^M\}_{i=0}^{K-1})$ . The output of C-Former is thus + +$$ +\mathbf {O} _ {C} = \mathbf {D} _ {C} (Q = \tilde {\mathbf {C}}, K = \mathbf {O}, V = \mathbf {O}), +$$ + +where we pass forward only the first set of $N_{c}$ output embeddings and ignore the rest, setting the output sequence dimension to $N_{c}$ . C-Former offers flexibility in controlling the tradeoff between ANLS quality and computational efficiency by controlling the output sequence length $N_{c}$ . + +# 3.5. Computation Analysis + +Next, we turn to provide a thorough computational complexity analysis. We consider a document that comprises of $K$ pages, each with $N$ tokens, and the maximum answer length is $L$ . For simplicity, we assume that all encoders and decoders have one layer. The naive way to support multi-page documents is using an existing single-page encoder-decoder model, fusing all of the textual page inputs together, and feeding them as one long sequence. We refer to this approach as 'concat'. The self-attention complexity of such a configuration scales quadratically with the sequence length, $O((N \cdot K)^2)$ . Conversely in our method, we operate on the document pages with two alternating encoding stages in each layer. The first stage performs a self-attention over both the page and doc tokens. Hence, the complexity of + +such sub-layer is $O((N + N_g)^2 \cdot K)$ , where $N_g$ is the number of doc tokens. The second stage features a self-attention operation over the doc tokens, across all pages in the document. The complexity of this operation is $O((N_g \cdot K)^2)$ . Overall, the total complexity for one global-local encoder block is $O((N + N_g)^2 \cdot K + (N_g \cdot K)^2)$ . Since $N_g$ is a constant, and the number of pages is usually less than the number of words in each page ( $K < N$ ), we obtain a complexity of $O(N^2 \cdot K)$ , which is not quadratic in $K$ . + +Prior to decoding, the outputs of all per-page encoders are concatenated, thus the output sequence length is $(N + N_{g})\cdot K$ Since the decoder is auto-regressive, its complexity depends quadratically on the maximum output length, $L$ ,namely, $O((N + N_g)\cdot K\cdot L^2) = O(N\cdot K\cdot L^2)$ Since this operation of decoding is performed iteratively during inference, the combined sequence length $(N + N_{g})\cdot K$ becomes computationally heavy. To alleviate this concern, we propose an optional C-Former model, which performs compression prior to decoding. The overall complexity in this decoding scheme includes passing through the C-Former and then through the decoder, leading to $O((N + N_g)\cdot K\cdot N_c) + N_c\cdot L^2)$ which is equivalent to $O(N\cdot K + L^2)$ , since $N_{c}$ is a constant, denoting the number of compression tokens in C-Former. + +# 4. Experiments + +# 4.1. Experimental Settings + +Datasets and Metrics The MPDocVQA dataset [33] features 46K questions, spanning over 48K images, and includes layout elements as figures, tables, lists and diagrams, with printed, handwritten and typewritten text. MPDocVQA contains mostly extractive questions, for which answers are present in the given text. DUDE is smaller in size (23.7K questions over 3K documents), but offers complex questions that require a reader to rationalize beyond the written text content. We report our results using the ANLS metric, introduced in [8], computing a generalized accuracy. Results for DUDE can be broken apart to several types of questions, categorized to four groups: 'extractive' – for which the answer is found directly in the text; 'abstractive' – requiring a free-form answer that does not necessarily appear in the document; 'list of answers' – requiring a list of answers, as opposed to a single one, and 'unanswerable' – where the result cannot be determined using the text. + +Implementation Details Our underlying architecture is based on Docformerv2 [6]. Recall, our interlaced encoder features $M$ blocks (12 in 'base' and 24 in 'large'), where each block contains a page sub-layer which originates from an extension of Docformerv2's encoder layer. Every structure contains self-attention, normalization and feed-forward + +
MethodParamsMPDocVQADUDE
ANLSANLSExtractiveANLS per Question TypeUnanswerable
AbstractiveList of answers
Longformer [7]148M55.0627.1443.588.5510.6210.78
BigBird [38]131M58.5426.2740.267.118.4612.75
LayoutLMv3 [38]125M55.1320.3132.608.107.828.82
Hi-VT5†beamsearch [15]316M-35.7428.3132.9810.6062.90
Hi-VT5[33]316M62.0123.0617.6033.946.8361.67
Hi-VT5*257M60.7823.867.2116.563.5372.77
DocFormerv2concat [6]257M69.6744.2141.6641.8615.1365.19
GRAMC-Former286M70.8040.0740.4339.6111.4252.55
GRAM281M73.6846.1546.0744.8215.2762.18
T5-2D [18]770M-46.0655.6550.815.4368.62
DocGptVQA [30]>3.5B-50.0251.8648.3228.2262.04
DocBlipVQA [29]>3.5B-47.6250.6946.3130.7355.22
Hi-VT5* [33]784M71.3528.8918.2126.176.8458.99
DocFormerv2concat [6]784M76.4048.4450.8248.0617.6759.04
GRAMC-Former864M77.6045.4747.6344.9114.3456.99
GRAM859M80.3251.1553.6750.3518.4063.23
Hi-VT5*†[33]784M73.5149.1849.2948.3513.3065.95
DocFormerv2†concat [6]784M76.7750.7952.7049.6117.3365.14
GRAM†C-Former864M78.1250.9755.1550.4617.2661.04
GRAM†859M79.6753.3656.8352.3219.9665.43
+ +Table 1. Quantitative Results. We present ANLS results for the MPDocVQA [33] and DUDE [18] test sets. The methods are grouped according to the model type and size, starting from encoder-only models (top), T5-base models (middle) and T5-large models (bottom). ${}^{ \dagger }$ denotes training with both MPDocVQA and DUDE. + +
MethodTraining DataANLS
DocVQAMPDocVQADocVQAMPDocVQA
DocFormerv2concat✓✗86.6072.73
✗✓85.2876.40
✓✓86.4775.37
GRAM✓✗86.7073.12
✗✓85.2980.32
✓✓86.3278.66
+ +Table 2. DocVQA vs. MPDocVQA Performance. Test results over both datasets using the large model variants. A checkmark denotes whether a dataset was included in training or not. + +layers. We extend the page layer from Docformerv2 to feature also the doc learnable embeddings. The second doc sub-layer is similar in structure to the first sub-layer, only it is initialized from scratch, with the following specification: $d_{ff} = 1024$ , $d_{kv} = 64$ , $nheads = 4$ , $d = 256$ . We implement 32 doc learnable tokens for each page, uniformly initialized to random values. For bias adaptation, the initial bias value is set to $c = 20$ , with variations between encoder heads, as described in Sec. 3.3. We incorporate an additional optional compression stage using C-Former - a randomly-initialized T5 [28] tiny decoder, with an encoder mask instead of a causal one. The output sequence extracted + +from C-Former is $N_{c} = 256$ . Finally the decoder is initialized with pretrained weights from DocformerV2. + +The model is trained with the Hugging Face Trainer [17] for $200k$ steps, starting with a warm-up of $1k$ steps, with linear learning rate decay. We use learning rates of $3e^{-5}$ and $1e^{-4}$ for the already pretrained encoder and decoder weights, versus the newly initialized doc sub-layer weights. Training is performed on a cluster of $8 \times A100$ GPUs, each with 40GB of RAM. During training, each page encoder receives 800 tokens, dealing with up to 4 pages. During testing, we increase the maximum length of tokens to 8,000. + +Baselines We report the results of previous work on both MP-DocVQA and DUDE datasets (if those exist), including the NLP-based Longformer [7] and BigBird [38], which were adapted to MPDocVQA by [33]; LayoutLMv3 [16], originally designed for DocVQA; and Hi-VT5 [33] and T5-2D [32], specifically suggested for MP-DocVQA Task. We also add for reference the results of methods published in the leader-boards of MPDocVQA and DUDE, which do not have corresponding papers, including DocGptVQA [30], DocBlipVQA [30], and Hi-VT5_beamsearch [15] ([15] was trained on both MP-DocVQA and DUDE). In our approach, we present two variations: GRAM and $\mathrm{GRAM}_{C - \text{Former}}$ . While GRAM utilizes the full length of the encoder output, + +![](images/58c8aa6b0d7b67224e400bd0a24a675c85c4896550ca2195790edc7174995db2.jpg) +How many diagrams are there? +HIVT5:"2" +GRAM:"3" +Which month shows the hurricane? +HiVT5:"not answerable" +Figure 4. Qualitative comparison between our approach and Hi-VT5 [33] indicate that the integration of our global-local encoder enhances reasoning capabilities, especially when the inquiries require multi-page context. + +![](images/a464058a112ba5afcf517b166a284d65eb3ee38e011249701fe9d365f0757b85.jpg) +GRAM:"July" + +![](images/f6c8b74b3eaf938c1c59278d2c880c02180af5a810fd497fbfb52742cbac40a2.jpg) +What is the eps code for little rock? +HiVT5:"EPS" +GRAM:"AR01" + +$\mathrm{GRAM}_{C - \text{Former}}$ allows the user to control the trade-off between performance and latency. + +To ensure a fair comparison, since we use the pretrained model of DocFormerv2 [6], we implement two additional baselines, referred to as Hi-VT5* and DocFormerv2concat. The first follows a similar structure as Hi-VT5 [33], with the encoder originating from DocFormerv2, however without the page answer prediction, as it does not exist in DUDE. The second recreates the approach of [18], where only the textual tokens of all pages are concatenated to one long sequence, then passing through the DocFormerv2 model. The second approach poses a computational burden, thus we use only 600 tokens during training per page, with up to 4 pages, and during test only 400 tokens. + +# 4.2. Results + +We present the performance of our method over the MPDocVQA [33] and DUDE [18] datasets in Tab. 1. The methods are divided into three groups: the top contains encoder-only, and methods that rely on the T5-base model (up to $316M$ parameters); the middle section, approaches that use the T5-large model (over $770M$ parameters), and finally the bottom, T5-large models, trained on both datasets. + +As can be seen, in the first group, the encoder-only NLP methods, LongFormer [7], BigBird [38] and LayoutLMv3 [16] can only handle relatively well 'extractive' style tasks as in MPDocVQA dataset [33], but often struggle with 'abstractive' questions that are more abundant in DUDE [18]. As to T5-'base' models, versus our best competitor Docformerv2concat, we obtain an improvement of $(+4\%, +1.9\%)$ on MP-DocVQA and DUDE datasets. As to methods that combine an additional compression before decoding (Hi-VT5, Hi-VT5*), our C-Former achieves an increase in $(+8.8\%, +16.2\%)$ over the best candidates on the MP-DocVQA and DUDE datasets. + +As for the group of 'large' models, we include the results of T5-2D [18] DocGptVQA [30] and + +DocBlipVQA [29]. Note that our model surpasses DocFormerconcat, the primary baseline, achieving improvements of $(+3.9\%, +2.7\%)$ on MP-DocVQA and DUDE, respectively. We also outperform DocGptVQA [30], a method that appears in the leaderboard of DUDE, by $+1.1\%$ , thereby obtaining SOTA results for GRAM 'large'. + +The final category showcases large encoder-decoder models, fine-tuned on both MP-DocVQA and DUDE training sets, showcasing the benefits of augmented training data. GRAM consistently demonstrates performance gains over the baseline, illustrating its robustness across different datasets and training scenarios. Next, we present in Tab. 2 the effect of training on DocVQA vs. MPDocVQA. Our method achieves performance on-par on the single page task, while enhancing performance on the multi-page scenario by $+3.3\%$ , compared to the baseline. + +In Fig. 4, We show qualitative results on the DUDE dataset of GRAM versus Hi-VT5* [33]. Our method demonstrates proficiency in addressing questions that involve attention over multiple pages ('how many diagrams are there'), an increased visual analysis capability ('Which month shows the hurricane?'), and heightened abstractive ability ('What is the EPS code for Little Rock?'). + +# 5. Ablation Study + +We perform an ablation study on our approach, evaluating the influence of each constituent component using DUDE's validation set [18]. This validation set enables the grouping of documents by their respective page counts: 1, 2-4, 5-10, 11-end, encompassing 1747, 2259, 1062, 1241 samples in each category, respectively. Our investigation delves into the impact of the number of doc tokens and the bias adaptation methods. Moreover, we employ the C-Former for sequence compression, adjusting the compression ratio and examining the balance between performance and latency (see supplementary for more details). + +
#Doc TokensBias TypeCompression DimensionANLS by Number of Pages DUDE validation dataset
All12-45-1011-end
XXX46.1647.1848.6643.3442.57
16X46.3948.3549.0643.1641.56
32DecayingX47.8849.2949.9045.9043.94
64X46.7047.9849.2244.0042.60
32XX47.5249.8549.9344.9042.10
ConstantX46.1447.4148.1344.4442.19
DecayingX47.8849.2949.9045.9043.94
32Decaying839.8339.7341.4136.9839.52
3240.4240.9541.6438.1239.39
25641.9942.5743.7738.4041.01
102442.5642.9744.3038.9441.93
409643.5944.7544.6440.5442.67
+ +GRAM Components We focus our initial exploration on the impact of the number of doc tokens $N_{g}$ . As can be seen in Tab. 3, while $N_{g} = 16$ leads to performance on par with not using doc tokens at all, for the optimal value of $N_{g} = 32$ , we obtain an increase of $+1.7\%$ in ANLS. Shifting our focus to bias adaptation methods, Tab. 3 shows that using constant bias has a negative effect on the results, suggesting this method is not flexible enough in maintaining a balance between the page and doc tokens. However, our decaying bias-adaptation approach does improve results overall, versus no-bias $(+0.36\%)$ , especially for longer documents $(+1\%$ improvement for 5-10 pages and $+1.84\%$ for 11 pages and more). This is to be expected, since incorporating new doc tokens and increasing their importance can potentially affect single-page performance. Finally, in Tab. 4, we reinforce our choice of pages as semantic logical units for MPDocVQA. We first ablate our method with and without page embedding. Next, we compare our page-based division with varying fixed-length division of tokens for encoder. Results in Tab. 4 clearly demonstrate an advantage towards page-level encoding in MPDocVQA. This aligns with our initial assumption that structured documents are often designed with page-division in mind. + +Performance-Latency Trade-off We assess the impact of C-Former on performance, considering compression output lengths of 8, 32, 256, 1024, 4096. Note, performance gradually improves with an increase in the compression output length. However, longer output lengths correspond to heightened model latency. Note that using C-Former for shorter documents can be redundant, as there is little to no compression compared to the input sequence length and results decrease. In Fig. 5, we scrutinize the trade-off between computational efficiency and compression rate + +Table 3. GRAM Ablation Study. Results on DUDE validation set ablating over (a) the dimension of doc tokens, (b) the attention bias employed and (c) the C-former input dimension. + +
Page EmbeddingSegment LengthANLS by Number of Pages DUDE validation dataset
All12-45-1011-end
✓✗47.8849.2949.9045.9043.94
✗✗46.1248.7448.1143.5940.99
✓25645.2246.3846.6944.1341.83
✓51245.0945.9047.3242.6541.98
✓102444.3944.9846.6341.6941.78
+ +Table 4. The Significance of Pages as Semantic Units. Results on DUDE validation set ablating over (a) utilization of page-embedding, (b) segment length for fixed-size encoding inputs. + +![](images/148ed5b5337a0df77c963769acb9ace696da8e29227d9d1bcf6b5fd25c5b981c.jpg) +Figure 5. Latency comparison. We compare the dependency between overall latency and the number of pages in input document for GRAM, $\mathrm{GRAM}_{C - \text{Former}}$ , DocFormerv2concat and Hi-VT5. + +by comparing to DocFormerv2concat [6] and Hi-VT5* [33]. We discover that DocFormerv2concat reaches a memory limit at approximately 20 pages, due to its quadratic memory increase with sequence length. At this juncture, GRAM $_{C-Former}$ surpasses DocFormerv2concat by performing 3.5 seconds faster. Notably, GRAM $_{C-Former}$ can gracefully handle documents surpassing 300 pages, effectively bridging the gap between performance and latency. + +# 6. Conclusions + +Our method, termed GRAM, extends existing single-page document models to efficiently handle multi-page documents without necessitating computationally-intensive pretraining. Leveraging the single-page encoder for local page-level comprehension, we introduce document learnable tokens and designated layers, enabling seamless information exchange across pages. Additionally, our proposed bias adaptation method enforces effective utilization of our newly introduced document tokens. The incorporation of a C-Former model reduces sequence length, balancing quality with latency in the decoding step. Extensive experiments demonstrate GRAM's state-of-the-art performance across multi-page DocVQA benchmarks. + +# References + +[1] Aviad Aberdam, Ron Litman, Shahar Tsiper, Oron Anschel, Ron Slossberg, Shai Mazor, R Manmatha, and Pietro Perona. Sequence-to-sequence contrastive learning for text recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15302-15312, 2021. 2 +[2] Aviad Aberdam, Roy Ganz, Shai Mazor, and Ron Litman. Multimodal semi-supervised learning for text recognition. arXiv preprint arXiv:2205.03873, 2022. +[3] Aviad Aberdam, David Bensaïd, Alona Golts, Roy Ganz, Oren Nuriel, Royee Tichauer, Shai Mazor, and Ron Litman. Clipter: Looking at the bigger picture in scene text recognition. arXiv preprint arXiv:2301.07464, 2023. 2 +[4] Joshua Ainslie, Tao Lei, Michiel de Jong, Santiago Ontañón, Siddhartha Brahma, Yury Zemlyanskiy, David Uthus, Mandy Guo, James Lee-Thorp, Yi Tay, et al. Colt5: Faster long-range transformers with conditional computation. arXiv preprint arXiv:2303.09752, 2023. 1, 2, 4 +[5] Srikar Appalaraju, Bhavan Jasani, Bhargava Urala Kota, Yusheng Xie, and R Manmatha. Docformer: End-to-end transformer for document understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 993-1003, 2021. 1, 2, 3 +[6] Srikar Appalaraju, Peng Tang, Qi Dong, Nishant Sankaran, Yichu Zhou, and R Manmatha. Docformerv2: Local features for document understanding. arXiv preprint arXiv:2306.01733, 2023. 1, 2, 3, 5, 6, 7, 8 +[7] Iz Beltagy, Matthew E Peters, and Arman Cohan. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150, 2020. 1, 2, 4, 6, 7 +[8] Ali Furkan Biten, Ruben Tito, Andres Mafla, Lluis Gomez, Marçal Rusinol, Ernest Valveny, CV Jawahar, and Dimosthenis Karatzas. Scene text visual question answering. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4291-4301, 2019. 5 +[9] Ali Furkan Biten, Ron Litman, Yusheng Xie, Srikar Appalaraju, and R Manmatha. Latr: Layout-aware transformer for scene-text vqa. pages 16548-16558, 2022. 2 +[10] Aydar Bulatov, Yuri Kuratov, and Mikhail S Burtsev. Scaling transformer to 1m tokens and beyond with rmt. arXiv preprint arXiv:2304.11062, 2023. 1 +[11] Shouyuan Chen, Sherman Wong, Liangjian Chen, and Yuandong Tian. Extending context window of large language models via positional interpolation. arXiv preprint arXiv:2306.15595, 2023. 2 +[12] Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc V Le, and Ruslan Salakhutdinov. Transformer-xl: Attentive language models beyond a fixed-length context. arXiv preprint arXiv:1901.02860, 2019. 1, 2 +[13] Robert M French. Catastrophic forgetting in connectionist networks. Trends in cognitive sciences, 3(4):128-135, 1999. 4 +[14] Roy Ganz, Oren Nuriel, Aviad Aberdam, Yair Kittenplon, Shai Mazor, and Ron Litman. Towards models that can see and read. arXiv preprint arXiv:2301.07389, 2023. 3, 4 + +[15] JiangLong He, Mamatha N, Shiv Vignesh, and Deepak Kumar. Hivt5beam. Hi-VT5 model pretrained with private custom document collection using span masking objective. Pretrained model is then trained with DUDE dataset and MultiPage DocVQA dataset, 2023. 6 +[16] Yupan Huang, Tengchao Lv, Lei Cui, Yutong Lu, and Furu Wei. Layoutmv3: Pre-training for document ai with unified text and image masking. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4083-4091, 2022. 1, 2, 6, 7 +[17] HuggingFace. Huggingface trainer. https://huggingface.co/docs/transformers/training. 6 +[18] Jordy Landeghem, Rubén Tito, Łukasz Borchmann, Michal Pietruszka, Paweł Józiak, Rafał Powalski, Dawid Jurkiewicz, Mickaël Coustaty, Bertrand Ackaert, Ernest Valveny, et al. Document understanding dataset and evaluation (dude). arXiv preprint arXiv:2305.08455, 2023. 1, 3, 6, 7 +[19] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 2 +[20] Ron Litman, Oron Anschel, Shahar Tsiper, Roeit Litman, Shai Mazor, and R Manmatha. Scatter: selective context attentional scene text recognizer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11962-11972, 2020. 2 +[21] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 2 +[22] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 2 +[23] Oren Nuriel, Sharon Fogel, and Ron Litman. Textadain: Paying attention to shortcut learning in text recognizers. In European Conference on Computer Vision, pages 427-445. Springer, 2022. 2 +[24] R OpenAI. Gpt-4 technical report. arXiv, pages 2303–08774, 2023. 2 +[25] Qiming Peng, Yinxu Pan, Wenjin Wang, Bin Luo, Zhenyu Zhang, Zhengjie Huang, Teng Hu, Weichong Yin, Yongfeng Chen, Yin Zhang, et al. Ernie-Layout: Layout knowledge enhanced pre-training for visually-rich document understanding. arXiv preprint arXiv:2210.06155, 2022. 1, 2 +[26] Rafal Powalski, Łukasz Borchmann, Dawid Jurkiewicz, Tomasz Dwojak, Michal Pietruszka, and Gabriela Pałka. Going full-tilt boogie on document understanding with text-image layout transformer. In Document Analysis and Recognition-ICDAR 2021: 16th International Conference, Lausanne, Switzerland, September 5–10, 2021, Proceedings, Part II 16, pages 732–747. Springer, 2021. 2, 3 +[27] Ofir Press, Noah A Smith, and Mike Lewis. Train short, test long: Attention with linear biases enables input length extrapolation. arXiv preprint arXiv:2108.12409, 2021. 1, 2, 5 +[28] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and + +Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485-5551, 2020. 2, 3, 5, 6 +[29] RenZhou, QiaolingDeng, XinfengChang, LuyanWang, XiaochenHu, HuiLi, and YaqiangWu. Docclipvqa. We integrated the prediction outputs from the UDOP model and Blip2 to enhance our results, and we optimized the image encoder and included page number features to address the challenge of multi-page documents. GPT to generate python-like modular programs., 2023. 6, 7 +[30] RenZhou, QiaolingDeng, XinfengChang, LuyanWang, XiaochenHu, HuiLi, and YaqiangWu. Docgptvqa. We integrated the prediction outputs from the UDOP model and Blip2 to enhance our results, and we optimized the image encoder and included page number features to address the challenge of multi-page documents. GPT to generate python-like modular programs., 2023. 6, 7 +[31] Zineng Tang, Ziyi Yang, Guoxin Wang, Yuwei Fang, Yang Liu, Chenguang Zhu, Michael Zeng, Cha Zhang, and Mohit Bansal. Unifying vision, text, and layout for universal document processing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19254-19264, 2023. 2, 3 +[32] Rubén Tito, Dimosthenis Karatzas, and Ernest Valveny. Document collection visual question answering. In Document Analysis and Recognition-ICDAR 2021: 16th International Conference, Lausanne, Switzerland, September 5–10, 2021, Proceedings, Part II 16, pages 778–792. Springer, 2021. 6 +[33] Rubén Tito, Dimosthenis Karatzas, and Ernest Valveny. Hierarchical multimodal transformers for multi-page docvqa. arXiv preprint arXiv:2212.05935, 2022. 1, 2, 3, 5, 6, 7, 8 +[34] Maria Tsimpoukelli, Jacob L Menick, Serkan Cabi, SM Eslami, Oriol Vinyls, and Felix Hill. Multimodal few-shot learning with frozen language models. Advances in Neural Information Processing Systems, 34:200-212, 2021. 4 +[35] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1 +[36] Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, and Ming Zhou. Layoutm: Pre-training of text and layout for document image understanding. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 1192-1200, 2020. 1, 2, 3 +[37] Yang Xu, Yiheng Xu, Tengchao Lv, Lei Cui, Furu Wei, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Wanxiang Che, et al. Layoutmv2: Multi-modal pre-training for visually-rich document understanding. arXiv preprint arXiv:2012.14740, 2020. 1, 2 +[38] Manzil Zaheer, Guru Guruganesh, Kumar Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontanon, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, et al. Big bird: Transformers for longer sequences. Advances in neural information processing systems, 33:17283-17297, 2020. 1, 2, 4, 6, 7 +[39] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. + +Lit: Zero-shot transfer with locked-image text tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18123-18133, 2022. 4 \ No newline at end of file diff --git a/2024/GRAM_ Global Reasoning for Multi-Page VQA/images.zip b/2024/GRAM_ Global Reasoning for Multi-Page VQA/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8711d7b2f667368e5cfebd611242f4452a2f9e88 --- /dev/null +++ b/2024/GRAM_ Global Reasoning for Multi-Page VQA/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:969496b80775789812a7407b4aff3abfea654dc0f17acf5f5cfdeed020566a49 +size 525256 diff --git a/2024/GRAM_ Global Reasoning for Multi-Page VQA/layout.json b/2024/GRAM_ Global Reasoning for Multi-Page VQA/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..30ddb88fdaa8692220b3769696222771e291dfc7 --- /dev/null +++ b/2024/GRAM_ Global Reasoning for Multi-Page VQA/layout.json @@ -0,0 +1,7822 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 149, + 103, + 444, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 103, + 444, + 121 + ], + "spans": [ + { + "bbox": [ + 149, + 103, + 444, + 121 + ], + "type": "text", + "content": "GRAM: Global Reasoning for Multi-Page VQA" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 65, + 143, + 145, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 143, + 145, + 171 + ], + "spans": [ + { + "bbox": [ + 65, + 143, + 145, + 171 + ], + "type": "text", + "content": "Tsachi Blau* Technion, Israel" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 169, + 144, + 238, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 144, + 238, + 171 + ], + "spans": [ + { + "bbox": [ + 169, + 144, + 238, + 171 + ], + "type": "text", + "content": "Sharon Fogel \nAWS AI Labs" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 261, + 144, + 340, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 144, + 340, + 171 + ], + "spans": [ + { + "bbox": [ + 261, + 144, + 340, + 171 + ], + "type": "text", + "content": "Roi Ronen* Technion, Israel" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 365, + 144, + 435, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 144, + 435, + 171 + ], + "spans": [ + { + "bbox": [ + 365, + 144, + 435, + 171 + ], + "type": "text", + "content": "Alona Golts† \nAWS AI Labs" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 458, + 144, + 528, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 458, + 144, + 528, + 171 + ], + "spans": [ + { + "bbox": [ + 458, + 144, + 528, + 171 + ], + "type": "text", + "content": "Shahar Tsiper \nAWS AI Labs" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 93, + 191, + 187, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 191, + 187, + 218 + ], + "spans": [ + { + "bbox": [ + 93, + 191, + 187, + 218 + ], + "type": "text", + "content": "Elad Ben Avraham \nAWS AI Labs" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 215, + 192, + 293, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 192, + 293, + 218 + ], + "spans": [ + { + "bbox": [ + 215, + 192, + 293, + 218 + ], + "type": "text", + "content": "Aviad Aberdam \nAWS AI Labs" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 321, + 191, + 402, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 191, + 402, + 218 + ], + "spans": [ + { + "bbox": [ + 321, + 191, + 402, + 218 + ], + "type": "text", + "content": "Roy Ganz* Technion, Israel" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 430, + 191, + 499, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 430, + 191, + 499, + 218 + ], + "spans": [ + { + "bbox": [ + 430, + 191, + 499, + 218 + ], + "type": "text", + "content": "Ron Litman\\* AWS AI Labs" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 143, + 260, + 191, + 273 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 260, + 191, + 273 + ], + "spans": [ + { + "bbox": [ + 143, + 260, + 191, + 273 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 285, + 290, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 285, + 290, + 539 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 290, + 539 + ], + "type": "text", + "content": "The increasing use of transformer-based large language models brings forward the challenge of processing long sequences. In document visual question answering (DocVQA), leading methods focus on the single-page setting, while documents can span hundreds of pages. We present GRAM, a method that seamlessly extends pretrained single-page models to the multi-page setting, without requiring computationally-heavy pretraining. To do so, we leverage a single-page encoder for local page-level understanding, and enhance it with document-level designated layers and learnable tokens, facilitating the flow of information across pages for global reasoning. To enforce our model to utilize the newly introduced document tokens, we propose a tailored bias adaptation method. For additional computational savings during decoding, we introduce an optional compression stage using our compression-transformer(C-Former), reducing the encoded sequence length, thereby allowing a tradeoff between quality and latency. Extensive experiments showcase GRAM's state-of-the-art performance on the benchmarks for multi-page DocVQA, demonstrating the effectiveness of our approach." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 571, + 128, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 571, + 128, + 585 + ], + "spans": [ + { + "bbox": [ + 47, + 571, + 128, + 585 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 592, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 592, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 592, + 287, + 676 + ], + "type": "text", + "content": "Document understanding, particularly in the context of DocVQA, has gained substantial research interest [5, 6, 16, 25, 36, 37] and offers a wide array of practical applications, focusing on data extraction and analysis of single page documents. However, Multi/Page DocVQA (MPDocVQA) poses a more realistic challenge, considering that the majority of documents, including contracts, manuals" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 309, + 262, + 545, + 453 + ], + "blocks": [ + { + "bbox": [ + 309, + 262, + 545, + 453 + ], + "lines": [ + { + "bbox": [ + 309, + 262, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 309, + 262, + 545, + 453 + ], + "type": "image", + "image_path": "b2ea6080c27f9abf9ddb071d36d1e077d7390878872fde0dff61d4993c023edd.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 461, + 547, + 518 + ], + "lines": [ + { + "bbox": [ + 304, + 461, + 547, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 547, + 518 + ], + "type": "text", + "content": "Figure 1. An Overview of GRAM. We suggest an interleaved encoder architecture combining page- with document-attention layers, allowing information to propagate between different pages. An optional compression transformer (C-former) is introduced to allow a trade-off between quality and latency." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 543, + 545, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 543, + 545, + 614 + ], + "spans": [ + { + "bbox": [ + 304, + 543, + 545, + 614 + ], + "type": "text", + "content": "and scientific papers, often extend well beyond a single page. Despite the practical relevance of MPDocVQA, it has received limited attention, primarily due to the absence of suitable datasets. Two recently introduced datasets, MPDocVQA [33] and DUDE [18], have opened up new avenues for MP-DocVQA research." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "type": "text", + "content": "Recent DocVQA approaches rely on transformers [35], at the heart of their architecture. While transformers are a powerful tool, they face challenges when dealing with long input sequences [4, 7, 10-12, 27, 38]. This difficulty stems from the self-attention mechanism, which scales quadratically in terms of computation and memory, with respect to the input sequence length. Addressing this limitation has attracted significant research efforts, primarily in the field" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 683, + 222, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 683, + 222, + 693 + ], + "spans": [ + { + "bbox": [ + 57, + 683, + 222, + 693 + ], + "type": "text", + "content": "*Work conducted during an internship at Amazon." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 59, + 693, + 214, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 693, + 214, + 703 + ], + "spans": [ + { + "bbox": [ + 59, + 693, + 214, + 703 + ], + "type": "text", + "content": "† Corresponding author: alongolt@amazon.com" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 59, + 703, + 211, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 211, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 211, + 712 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 59, + 703, + 211, + 712 + ], + "type": "text", + "content": " Corresponding author: litmanr@amazon.com" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15598" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": "of natural language processing (NLP). Proposed NLP-based solutions can be divided into two main directions: The former aims to modify the attention mechanism to cut computational costs [4, 7, 38]. The latter involves altering the positional embedding mechanism to improve performance on longer sequences, with minimal fine-tuning [11, 12, 27]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 147, + 288, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 147, + 288, + 337 + ], + "spans": [ + { + "bbox": [ + 46, + 147, + 288, + 337 + ], + "type": "text", + "content": "A possible option of tackling MPDocVQA is to extend NLP-based approaches to handle multi-modal document data, including visual representations, along with OCR text and corresponding 2D locations and relative page position. However, this requires extensive pre-training, with relatively scarce multi-page document data, and thus is also sub-optimal in terms of performance. Instead, we opt for leveraging powerful single-page DocVQA models, especially pretrained on millions of single-page documents, and finetuning them to the multi-page scenario. For this purpose, we combine concepts of local (page) and global (document) tokens, which promote an exchange of information within and across pages, while keeping computational cost in check. We choose pages as atomic units in our proposed scheme, as page structure often represent a semantic unit in DocVQA." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 340, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 340, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 340, + 287, + 568 + ], + "type": "text", + "content": "We present GRAM (Global ReAsoning for Multi-page VQA), a novel approach for endowing multi-page processing capabilities to existing single-page DocVQA models. Alongside page tokens that encapsulate both textual and visual contents of each page, we introduce doc(ument) learnable tokens, which aim is dispersing global information across all pages. These two sets of tokens interact within our newly-devised two-stage encoder blocks. The initial stage utilizes an existing single-page layer and enhances it by including both page and doc tokens as input, allowing them to freely interact. In the second stage, we prioritize computational efficiency by restricting self-attention solely to the global doc tokens. This global reasoning layer captures collective information from multiple pages, enabling the system to respond to cross-page inquiries, as illustrated in Fig. 1. Considering that doc tokens did not appear in pretraining, to boost their significance during finetune, we employ a designated bias adaptation mechanism which strikes a balance between local and global learnable tokens." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "content": "While our method inherently deals with long sequences, we circumvent a quadratic reliance on sequence length by segmenting the document into pages — its semantically logical parts. We restrict interaction solely among doc learnable tokens, across all pages, thereby mitigating the computational burden of depending quadratically on the page count. Apart from encoding, the auto-regressive decoding stage poses a computational burden in long sequences. To this end, we introduce a compression stage that precedes the decoder, implemented with a compression transformer, termed CFormer. The CFormer receives the concatenated output of all pages and compresses it to a" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "type": "text", + "content": "much shorter sequence, distilling the most pertinent information in the document. Our key contributions are:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 96, + 544, + 261 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 306, + 96, + 544, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 96, + 544, + 143 + ], + "spans": [ + { + "bbox": [ + 306, + 96, + 544, + 143 + ], + "type": "text", + "content": "- We propose GRAM, an approach to endow single-page DocVQA methods with multi-page capabilities, without pretraining, allowing the model to process multi-page documents, while preserving single-page performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 144, + 544, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 144, + 544, + 190 + ], + "spans": [ + { + "bbox": [ + 306, + 144, + 544, + 190 + ], + "type": "text", + "content": "- We introduce document learnable tokens and bias adaptation that enable an effective communication and collaboration between individual pages to support reasoning over multiple page documents." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 192, + 544, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 192, + 544, + 225 + ], + "spans": [ + { + "bbox": [ + 306, + 192, + 544, + 225 + ], + "type": "text", + "content": "- Our C-Former module suggests a trade-off between accuracy and compute, distilling information from multi-page sequences into more compact representations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 228, + 544, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 228, + 544, + 261 + ], + "spans": [ + { + "bbox": [ + 306, + 228, + 544, + 261 + ], + "type": "text", + "content": "- We obtain SOTA results over the MPDocVQA and DUDE datasets, and provide extensive ablations to each component in our method." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 274, + 392, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 274, + 392, + 286 + ], + "spans": [ + { + "bbox": [ + 306, + 274, + 392, + 286 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 295, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 295, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 295, + 545, + 569 + ], + "type": "text", + "content": "Long Sequence Approaches are an active field of research in NLP, aiming to improve the design of chat-systems [24] and image instruction tasks [19, 21]. In these applications, the ability to manage and process long sequences is vital, as conversations cannot be cut short, or limited to just a few interactions. Common approaches to tackle long sequences include sparse attention mechanisms [4, 7, 38] and methods to improve results on long sequences during inference [11, 27, 28]. 'Sliding window' approaches of limiting the range of neighbors each token can attend to, lead to a significant reduction in computation and memory consumption. Prominent works of this kind include Long-Former [7], where each token attends to a set of its nearest neighbors, along with additional global tokens. The work of Big-Bird [38] adds additional non-neighboring tokens at random, whereas Colt5 [4] uses the same sliding window approach, but performs heavier computations for important tokens and shallow operations for filler words or punctuation. Although Tito et. al. [33] have demonstrated that the above approaches do not perform as well on the task of MPDocVQA, we do incorporate the ideas of combining both local and global tokens throughout encoding to expand the attention onto additional pages." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "DocVQA has attracted increasing attention [5, 6, 9, 16, 25, 26, 31, 36, 37] with the introduction of the DocVQA dataset by Mathew et al. [22]. Most methods in DocVQA leverage OCR [1-3, 20, 23] to input both text and layout information (bounding box coordinates and possibly font type) into the model, where some further explore different techniques to combine the two types of data streams, or alternatively, clever schemes of pretraining. DocVQA methods can be roughly divided into two categories: extractive and abstractive. Extractive methods [16, 25, 36, 37] rely on the fact that the explicit answer resides in the written text, thus only output a corresponding text span within the input" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15599" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 88, + 77, + 258, + 351 + ], + "blocks": [ + { + "bbox": [ + 88, + 77, + 258, + 351 + ], + "lines": [ + { + "bbox": [ + 88, + 77, + 258, + 351 + ], + "spans": [ + { + "bbox": [ + 88, + 77, + 258, + 351 + ], + "type": "image", + "image_path": "4a1cf9abb9b129b296325ad8d617ec7270c8af4ec0258570380b4af9be277ec6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 361, + 545, + 437 + ], + "lines": [ + { + "bbox": [ + 46, + 361, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 545, + 437 + ], + "type": "text", + "content": "Figure 2. GRAM Architecture. (a) Depicts a high-level architecture overview. For each page, the visual, textual and question tokens are concatenated together with learnable doc tokens (darker color shade). The processed information is fed into the multi-page encoder. The encoder output can be fed directly into the decoder to create the final prediction. Optionally, a compression model, C-Former, can be used between the encoder and the decoder to compress the encoder output into a predetermined length, thus reducing overall latency for long documents. (b) Shows a global-local encoder layer, containing two sub-layers. The first sub-layer uses self-attention that operates on each page separately, while the second applies a self-attention step on the doc tokens to fuse information between the different pages. The corresponding tokens are then routed back to their respective page and go into the next global-local encoder layer." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 271, + 86, + 493, + 353 + ], + "blocks": [ + { + "bbox": [ + 271, + 86, + 493, + 353 + ], + "lines": [ + { + "bbox": [ + 271, + 86, + 493, + 353 + ], + "spans": [ + { + "bbox": [ + 271, + 86, + 493, + 353 + ], + "type": "image", + "image_path": "0d71025096650b205a82653e1b09dd7a52889303abd1cfb63fbac9d494a5e79a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 460, + 287, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 532 + ], + "type": "text", + "content": "sequence. Abstractive methods [5, 6, 14, 26, 31], on the other hand, have the capacity to generate free-form answers which do not necessarily appear in the text, thus providing flexibility in real-world applications. Notably, existing research in DocVQA does not scale in a straightforward way to deal with the more realistic multi-page scenario." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 534, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 287, + 715 + ], + "type": "text", + "content": "MPDocVQA has recently gained momentum with the launch of two new multi-page datasets: MP-DocVQA [33] and DUDE [18], offering two separate recipes to tackle longer documents. The first approach of Tito et. al., referred to as HiVT5 [33], suggests compressing the encoding of each page separately, and feeding the decoder with the concatenation of the compressed outputs from each single-page encoder. While this approach is advantageous in terms of computation, we later show the compression may severely hinder the results. In addition, there is no communication between the single-page encoders until the final stage of decoding, whereas in our method, we allow for exchange of page and document information throughout all stages of encoding. Another prominent approach, proposed by Landeghem et. al. [18], which relatively preserves qual" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 460, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 521 + ], + "type": "text", + "content": "ity, involves concatenating all the pages into one long sequence and feeding it to a standard encoder-decoder structure. This, however, poses a heavy computational burden as transformers' self-attention component scales quadratically with input sequence length." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 531, + 359, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 531, + 359, + 544 + ], + "spans": [ + { + "bbox": [ + 306, + 531, + 359, + 544 + ], + "type": "text", + "content": "3. GRAM" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 552, + 411, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 552, + 411, + 562 + ], + "spans": [ + { + "bbox": [ + 306, + 552, + 411, + 562 + ], + "type": "text", + "content": "3.1. Base Architecture" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "The underlying idea in our approach is using existing encoder-decoder single-page models for document understanding and extending them to multi-page scenarios, without additional pretraining. In this work, we provide such a recipe over the notable DocFormerv2 [6]. For the sake of completeness, DocFormerv2 is a T5-based [28] encoder-decoder transformer model which operates over both visual and textual features to support document understanding. Each page is represented by textual features " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{T} \\in R^{N_t \\times d}" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " which encapsulate OCR tokens and their corresponding 2D bounding box positions [36], along with visual " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{V} \\in R^{N_v \\times d}" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " and question " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{Q} \\in R^{N_q \\times d}" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " embeddings. Where " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "N_t" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "N_v" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15600" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "inline_equation", + "content": "N_{q}" + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": " are the lengths of the OCR, visual features and the question. The output result " + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": " is obtained by passing the concatenated inputs through the encoder-decoder model," + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 118, + 287, + 133 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 118, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 105, + 118, + 287, + 133 + ], + "type": "interline_equation", + "content": "\\mathbf {Y} = \\mathbf {D} (\\mathbf {E} (\\operatorname {c o n c a t} (\\mathbf {T}, \\mathbf {V}, \\mathbf {Q}))), \\tag {1}", + "image_path": "ee5fad28aa3324f065208f6867129977f4e2d47893a510b13ed1638d69c913da.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 141, + 281, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 141, + 281, + 153 + ], + "spans": [ + { + "bbox": [ + 47, + 141, + 281, + 153 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 141, + 281, + 153 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 47, + 141, + 281, + 153 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 141, + 281, + 153 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 47, + 141, + 281, + 153 + ], + "type": "text", + "content": ", are the encoder and decoder, respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 154, + 288, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 154, + 288, + 309 + ], + "spans": [ + { + "bbox": [ + 46, + 154, + 288, + 309 + ], + "type": "text", + "content": "Our method uses these basic building blocks in designing a multi-page solution. To this end, we introduce a bi-level global-local encoder, as illustrated in Fig. 2. At the local page-level of each block, we utilize the layers of the existing single-page encoder " + }, + { + "bbox": [ + 46, + 154, + 288, + 309 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 46, + 154, + 288, + 309 + ], + "type": "text", + "content": " to process each page separately, together with learnable doc tokens. Next, we introduce a slim global layer in each block that facilitates communication between doc tokens across all pages. This bi-level localized processing ensures the model can understand the content of each page effectively, while also combining information across pages in the document. After " + }, + { + "bbox": [ + 46, + 154, + 288, + 309 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 154, + 288, + 309 + ], + "type": "text", + "content": " such blocks, we feed the encoded features from all pages into the existing decoder " + }, + { + "bbox": [ + 46, + 154, + 288, + 309 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 46, + 154, + 288, + 309 + ], + "type": "text", + "content": " to produce the overall output." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 317, + 183, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 317, + 183, + 330 + ], + "spans": [ + { + "bbox": [ + 47, + 317, + 183, + 330 + ], + "type": "text", + "content": "3.2. Global-Local Reasoning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 336, + 287, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 336, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 336, + 287, + 456 + ], + "type": "text", + "content": "To operate on multiple pages we break down the document to " + }, + { + "bbox": [ + 46, + 336, + 287, + 456 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 336, + 287, + 456 + ], + "type": "text", + "content": " pages, and the single-page encoder to " + }, + { + "bbox": [ + 46, + 336, + 287, + 456 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 336, + 287, + 456 + ], + "type": "text", + "content": " encoder layers, " + }, + { + "bbox": [ + 46, + 336, + 287, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{E}^j, j = 0,\\dots,M - 1" + }, + { + "bbox": [ + 46, + 336, + 287, + 456 + ], + "type": "text", + "content": ". We then construct " + }, + { + "bbox": [ + 46, + 336, + 287, + 456 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 336, + 287, + 456 + ], + "type": "text", + "content": " blocks, with two sub-layers each. The first page sub-layer originates from the existing pretrained encoder layer, referred to as " + }, + { + "bbox": [ + 46, + 336, + 287, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{page}^{j}" + }, + { + "bbox": [ + 46, + 336, + 287, + 456 + ], + "type": "text", + "content": ", and operates in parallel, with shared weights, for all pages in the document. This layer receives both page and doc tokens. The second, newly introduced, document layer " + }, + { + "bbox": [ + 46, + 336, + 287, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{doc}^{j}" + }, + { + "bbox": [ + 46, + 336, + 287, + 456 + ], + "type": "text", + "content": " collects only the doc tokens from all pages and promotes sharing information across all of the document." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 456, + 287, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 287, + 503 + ], + "type": "text", + "content": "Formally, we augment the input of the standard single-page encoder with page-specific indexing " + }, + { + "bbox": [ + 46, + 456, + 287, + 503 + ], + "type": "inline_equation", + "content": "(\\mathbf{T}_i,\\mathbf{V}_i,\\mathbf{Q})" + }, + { + "bbox": [ + 46, + 456, + 287, + 503 + ], + "type": "text", + "content": " and incorporate page-positional embedding " + }, + { + "bbox": [ + 46, + 456, + 287, + 503 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_i" + }, + { + "bbox": [ + 46, + 456, + 287, + 503 + ], + "type": "text", + "content": " to both text and visual features, where " + }, + { + "bbox": [ + 46, + 456, + 287, + 503 + ], + "type": "inline_equation", + "content": "i = 0,\\dots,K - 1" + }, + { + "bbox": [ + 46, + 456, + 287, + 503 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 94, + 514, + 287, + 528 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 514, + 287, + 528 + ], + "spans": [ + { + "bbox": [ + 94, + 514, + 287, + 528 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {T}} _ {i} = \\mathbf {T} _ {i} + \\mathbf {P} _ {i}, \\quad \\tilde {\\mathbf {V}} _ {i} = \\mathbf {V} _ {i} + \\mathbf {P} _ {i}. \\tag {2}", + "image_path": "499fdc989fe98c9eb901ad19c0c31ca3a7a5bae1f58e0a198cd6791ecf139276.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 537, + 287, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 537, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 46, + 537, + 287, + 599 + ], + "type": "text", + "content": "Next, we formulate our bi-level global-local block. The input to the first page-level sub-layer in each block is the concatenation of the textual, visual and question features, denoted " + }, + { + "bbox": [ + 46, + 537, + 287, + 599 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_i^j = \\mathrm{concat}(\\tilde{\\mathbf{T}}_i,\\tilde{\\mathbf{V}}_i,\\mathbf{Q}_i)" + }, + { + "bbox": [ + 46, + 537, + 287, + 599 + ], + "type": "text", + "content": ", along with page-specific doc tokens " + }, + { + "bbox": [ + 46, + 537, + 287, + 599 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_i^j\\in R^{N_g\\times d}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 84, + 609, + 287, + 625 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 609, + 287, + 625 + ], + "spans": [ + { + "bbox": [ + 84, + 609, + 287, + 625 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {i} ^ {j + 1}, \\tilde {\\mathbf {G}} _ {i} ^ {j + 1} = \\mathbf {E} _ {p a g e} ^ {j} (\\operatorname {c o n c a t} \\left(\\mathbf {X} _ {i} ^ {j}, \\mathbf {G} _ {i} ^ {j}\\right)). \\tag {3}", + "image_path": "043b5565e37ec767bc4b84e10276ba017eb0a4aff88af366354d6d708a4916a1.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 631, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 631, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 631, + 287, + 689 + ], + "type": "text", + "content": "Here, the features undergo self-attention, normalization and feed-forward layers. The layer output " + }, + { + "bbox": [ + 46, + 631, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_i^{j + 1}" + }, + { + "bbox": [ + 46, + 631, + 287, + 689 + ], + "type": "text", + "content": " is passed on as input to the next bi-level block, whereas only the doc tokens " + }, + { + "bbox": [ + 46, + 631, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{G}}_i^{j + 1}" + }, + { + "bbox": [ + 46, + 631, + 287, + 689 + ], + "type": "text", + "content": ", enter the second doc sub-layer, which again includes self-attention, normalization and feed-forward" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 74, + 699, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 699, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 74, + 699, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\left\\{\\mathbf {G} _ {i} ^ {j + 1} \\right\\} _ {i = 0} ^ {K - 1} = \\mathbf {E} _ {d o c} ^ {j} (\\operatorname {c o n c a t} \\left(\\left\\{\\tilde {\\mathbf {G}} _ {i} ^ {j + 1} \\right\\} _ {i = 0} ^ {K - 1}\\right)). \\tag {4}", + "image_path": "44aa60bbcba3054a6163306e88fe7f7391c0e77c4135f111da6dfa842c342669.jpg" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 321, + 72, + 533, + 205 + ], + "blocks": [ + { + "bbox": [ + 321, + 72, + 533, + 205 + ], + "lines": [ + { + "bbox": [ + 321, + 72, + 533, + 205 + ], + "spans": [ + { + "bbox": [ + 321, + 72, + 533, + 205 + ], + "type": "image", + "image_path": "980915e4e5e177aa55bc25e81b3ce9b2634298bf14abc038ccfc2dc77cbaf0fc.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 213, + 545, + 269 + ], + "lines": [ + { + "bbox": [ + 305, + 213, + 545, + 269 + ], + "spans": [ + { + "bbox": [ + 305, + 213, + 545, + 269 + ], + "type": "text", + "content": "Figure 3. Global-Local Attention: In long sequence approaches (a), attention is applied jointly to the entire sequence of concatenated local and global tokens. Our method, separates the computation into two steps — page-level (b) and document-level (c)— leveraging the natural division of documents into pages." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 289, + 545, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 289, + 545, + 372 + ], + "spans": [ + { + "bbox": [ + 305, + 289, + 545, + 372 + ], + "type": "text", + "content": "In this stage, the doc tokens can interact and pass information from page to page, after which being passed on to the next block, as depicted in Eq. (3). This design allows information to flow between pages while keeping computational costs in check. When concluding the traversal over " + }, + { + "bbox": [ + 305, + 289, + 545, + 372 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 305, + 289, + 545, + 372 + ], + "type": "text", + "content": " such layers, the outputs across all pages are concatenated and fed to the decoder " + }, + { + "bbox": [ + 305, + 289, + 545, + 372 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 305, + 289, + 545, + 372 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 341, + 379, + 545, + 394 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 379, + 545, + 394 + ], + "spans": [ + { + "bbox": [ + 341, + 379, + 545, + 394 + ], + "type": "interline_equation", + "content": "\\mathbf {Y} _ {m u l t i} = \\mathbf {D} (\\operatorname {c o n c a t} \\left(\\left\\{\\mathbf {X} _ {i} ^ {M}, \\mathbf {G} _ {i} ^ {M} \\right\\} _ {i = 0} ^ {K - 1}\\right)). \\tag {5}", + "image_path": "c8f81b7e2e14516dedc4407e86387d8d111d4405c918efd8e9edd462151f54f3.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 401, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 401, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 545, + 533 + ], + "type": "text", + "content": "To visualize the difference between the attention masks in our method, we compare it with previous long sequence approaches [4, 7, 38] in Fig. 3. These prior methods optimize computation by using attention masking on nearby tokens and allowing limited global connections. However, naively applying such methods to multi-page documents will treat it as a single stream, which does not consider the division into pages. Our global-local blocks, with a two-stage attention-masking mechanism, better suit multi-page documents. In addition, our two-level design benefits from existing, extensively pretrained single page models." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 540, + 404, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 540, + 404, + 552 + ], + "spans": [ + { + "bbox": [ + 306, + 540, + 404, + 552 + ], + "type": "text", + "content": "3.3. Bias Adaptation" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": "An already-pretrained model, introduced with a new stream of data, might disregard it altogether [13, 14, 34, 39]. To overcome this, we force the system to account for the newly-introduced doc tokens by modifying the encoder's bias method. Originally, the bias method intervenes in the attention mechanism, diminishing the relationships between distant tokens. However, in our specific case, the distance between doc and page tokens does not represent their actual relevance. To enforce the encoder to pay closer attention to the doc tokens, we assign them a positive constant bias value. Particularly, we replace the values in the bias matrix, corresponding with the doc tokens, with fixed ones. Instead of a single bias value, we utilize a different value" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15601" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": "for each attention head, as performed in ALiBi [27], enabling more fine-grained control of the global features in each head. Specifically, the constant doc bias value is set to " + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "c \\cdot \\frac{1}{2^a}" + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": " is a constant and " + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": " is the attention head index. This yields a decaying bias value across different attention heads, resulting in hierarchical importance of the document information, where the first heads are more oriented towards doc tokens and the last towards page tokens." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 175, + 194, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 175, + 194, + 186 + ], + "spans": [ + { + "bbox": [ + 47, + 175, + 194, + 186 + ], + "type": "text", + "content": "3.4. Compression Transformer" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 193, + 287, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 193, + 287, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 193, + 287, + 324 + ], + "type": "text", + "content": "Our global-local solution to MP-DocVQA resolves the problematic quadratic dependency on the number of pages " + }, + { + "bbox": [ + 46, + 193, + 287, + 324 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 193, + 287, + 324 + ], + "type": "text", + "content": " during encoding. However, the auto-regressive decoding complexity scaling linearly with " + }, + { + "bbox": [ + 46, + 193, + 287, + 324 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 193, + 287, + 324 + ], + "type": "text", + "content": " also poses a practical challenge during inference time, as we later discuss in Sec. 3.5. To alleviate this burden, we place an optional transformer-based model, named C-Former (Compression TransFormer), between the encoder outputs and the decoder, as depicted in Fig. 2. The C-Former has the ability to revise the information across all pages and distill only the important details, required to correctly answer the question." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 324, + 287, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 324, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 324, + 287, + 434 + ], + "type": "text", + "content": "Specifically, the C-Former is a light-weight transformer-based decoder [28], denoted as " + }, + { + "bbox": [ + 46, + 324, + 287, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_C" + }, + { + "bbox": [ + 46, + 324, + 287, + 434 + ], + "type": "text", + "content": ", featuring crossattention, layer norm and feed-forward layers in each block. The input to C-Former includes " + }, + { + "bbox": [ + 46, + 324, + 287, + 434 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 46, + 324, + 287, + 434 + ], + "type": "text", + "content": " learnable tokens " + }, + { + "bbox": [ + 46, + 324, + 287, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{C} \\in R^{N_c \\times d}" + }, + { + "bbox": [ + 46, + 324, + 287, + 434 + ], + "type": "text", + "content": ", concatenated with the input question " + }, + { + "bbox": [ + 46, + 324, + 287, + 434 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{C}} = \\mathrm{concat}(\\mathbf{C}, \\mathbf{Q})" + }, + { + "bbox": [ + 46, + 324, + 287, + 434 + ], + "type": "text", + "content": ". In addition, we feed it with the outputs of the global-local interlaced encoder, concatenated to one long sequence, referred as " + }, + { + "bbox": [ + 46, + 324, + 287, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{O}" + }, + { + "bbox": [ + 46, + 324, + 287, + 434 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 324, + 287, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{O} = \\mathrm{concat}(\\{\\mathbf{X}_i^M, \\mathbf{G}_i^M\\}_{i=0}^{K-1})" + }, + { + "bbox": [ + 46, + 324, + 287, + 434 + ], + "type": "text", + "content": ". The output of C-Former is thus" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 97, + 441, + 236, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 441, + 236, + 453 + ], + "spans": [ + { + "bbox": [ + 97, + 441, + 236, + 453 + ], + "type": "interline_equation", + "content": "\\mathbf {O} _ {C} = \\mathbf {D} _ {C} (Q = \\tilde {\\mathbf {C}}, K = \\mathbf {O}, V = \\mathbf {O}),", + "image_path": "20451565c1932645faf1f6e5284e24c40400a7a46fec838cae0ede45adbb9e8c.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 461, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 287, + 521 + ], + "type": "text", + "content": "where we pass forward only the first set of " + }, + { + "bbox": [ + 46, + 461, + 287, + 521 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 46, + 461, + 287, + 521 + ], + "type": "text", + "content": " output embeddings and ignore the rest, setting the output sequence dimension to " + }, + { + "bbox": [ + 46, + 461, + 287, + 521 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 46, + 461, + 287, + 521 + ], + "type": "text", + "content": ". C-Former offers flexibility in controlling the tradeoff between ANLS quality and computational efficiency by controlling the output sequence length " + }, + { + "bbox": [ + 46, + 461, + 287, + 521 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 46, + 461, + 287, + 521 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 527, + 175, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 527, + 175, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 527, + 175, + 540 + ], + "type": "text", + "content": "3.5. Computation Analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": "Next, we turn to provide a thorough computational complexity analysis. We consider a document that comprises of " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": " pages, each with " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": " tokens, and the maximum answer length is " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": ". For simplicity, we assume that all encoders and decoders have one layer. The naive way to support multi-page documents is using an existing single-page encoder-decoder model, fusing all of the textual page inputs together, and feeding them as one long sequence. We refer to this approach as 'concat'. The self-attention complexity of such a configuration scales quadratically with the sequence length, " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "O((N \\cdot K)^2)" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": ". Conversely in our method, we operate on the document pages with two alternating encoding stages in each layer. The first stage performs a self-attention over both the page and doc tokens. Hence, the complexity of" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": "such sub-layer is " + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "O((N + N_g)^2 \\cdot K)" + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "N_g" + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": " is the number of doc tokens. The second stage features a self-attention operation over the doc tokens, across all pages in the document. The complexity of this operation is " + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "O((N_g \\cdot K)^2)" + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": ". Overall, the total complexity for one global-local encoder block is " + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "O((N + N_g)^2 \\cdot K + (N_g \\cdot K)^2)" + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": ". Since " + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "N_g" + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": " is a constant, and the number of pages is usually less than the number of words in each page (" + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "K < N" + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": "), we obtain a complexity of " + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "O(N^2 \\cdot K)" + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": ", which is not quadratic in " + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 180, + 546, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 180, + 546, + 372 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 546, + 372 + ], + "type": "text", + "content": "Prior to decoding, the outputs of all per-page encoders are concatenated, thus the output sequence length is " + }, + { + "bbox": [ + 304, + 180, + 546, + 372 + ], + "type": "inline_equation", + "content": "(N + N_{g})\\cdot K" + }, + { + "bbox": [ + 304, + 180, + 546, + 372 + ], + "type": "text", + "content": " Since the decoder is auto-regressive, its complexity depends quadratically on the maximum output length, " + }, + { + "bbox": [ + 304, + 180, + 546, + 372 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 180, + 546, + 372 + ], + "type": "text", + "content": " ,namely, " + }, + { + "bbox": [ + 304, + 180, + 546, + 372 + ], + "type": "inline_equation", + "content": "O((N + N_g)\\cdot K\\cdot L^2) = O(N\\cdot K\\cdot L^2)" + }, + { + "bbox": [ + 304, + 180, + 546, + 372 + ], + "type": "text", + "content": " Since this operation of decoding is performed iteratively during inference, the combined sequence length " + }, + { + "bbox": [ + 304, + 180, + 546, + 372 + ], + "type": "inline_equation", + "content": "(N + N_{g})\\cdot K" + }, + { + "bbox": [ + 304, + 180, + 546, + 372 + ], + "type": "text", + "content": " becomes computationally heavy. To alleviate this concern, we propose an optional C-Former model, which performs compression prior to decoding. The overall complexity in this decoding scheme includes passing through the C-Former and then through the decoder, leading to " + }, + { + "bbox": [ + 304, + 180, + 546, + 372 + ], + "type": "inline_equation", + "content": "O((N + N_g)\\cdot K\\cdot N_c) + N_c\\cdot L^2)" + }, + { + "bbox": [ + 304, + 180, + 546, + 372 + ], + "type": "text", + "content": " which is equivalent to " + }, + { + "bbox": [ + 304, + 180, + 546, + 372 + ], + "type": "inline_equation", + "content": "O(N\\cdot K + L^2)" + }, + { + "bbox": [ + 304, + 180, + 546, + 372 + ], + "type": "text", + "content": " , since " + }, + { + "bbox": [ + 304, + 180, + 546, + 372 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 304, + 180, + 546, + 372 + ], + "type": "text", + "content": " is a constant, denoting the number of compression tokens in C-Former." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 383, + 388, + 397 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 383, + 388, + 397 + ], + "spans": [ + { + "bbox": [ + 305, + 383, + 388, + 397 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 403, + 432, + 416 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 403, + 432, + 416 + ], + "spans": [ + { + "bbox": [ + 306, + 403, + 432, + 416 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 422, + 545, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 545, + 625 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 545, + 625 + ], + "type": "text", + "content": "Datasets and Metrics The MPDocVQA dataset [33] features 46K questions, spanning over 48K images, and includes layout elements as figures, tables, lists and diagrams, with printed, handwritten and typewritten text. MPDocVQA contains mostly extractive questions, for which answers are present in the given text. DUDE is smaller in size (23.7K questions over 3K documents), but offers complex questions that require a reader to rationalize beyond the written text content. We report our results using the ANLS metric, introduced in [8], computing a generalized accuracy. Results for DUDE can be broken apart to several types of questions, categorized to four groups: 'extractive' – for which the answer is found directly in the text; 'abstractive' – requiring a free-form answer that does not necessarily appear in the document; 'list of answers' – requiring a list of answers, as opposed to a single one, and 'unanswerable' – where the result cannot be determined using the text." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 641, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 546, + 713 + ], + "type": "text", + "content": "Implementation Details Our underlying architecture is based on Docformerv2 [6]. Recall, our interlaced encoder features " + }, + { + "bbox": [ + 304, + 641, + 546, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 641, + 546, + 713 + ], + "type": "text", + "content": " blocks (12 in 'base' and 24 in 'large'), where each block contains a page sub-layer which originates from an extension of Docformerv2's encoder layer. Every structure contains self-attention, normalization and feed-forward" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15602" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 70, + 544, + 348 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 544, + 348 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 544, + 348 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 544, + 348 + ], + "type": "table", + "html": "
MethodParamsMPDocVQADUDE
ANLSANLSExtractiveANLS per Question TypeUnanswerable
AbstractiveList of answers
Longformer [7]148M55.0627.1443.588.5510.6210.78
BigBird [38]131M58.5426.2740.267.118.4612.75
LayoutLMv3 [38]125M55.1320.3132.608.107.828.82
Hi-VT5†beamsearch [15]316M-35.7428.3132.9810.6062.90
Hi-VT5[33]316M62.0123.0617.6033.946.8361.67
Hi-VT5*257M60.7823.867.2116.563.5372.77
DocFormerv2concat [6]257M69.6744.2141.6641.8615.1365.19
GRAMC-Former286M70.8040.0740.4339.6111.4252.55
GRAM281M73.6846.1546.0744.8215.2762.18
T5-2D [18]770M-46.0655.6550.815.4368.62
DocGptVQA [30]>3.5B-50.0251.8648.3228.2262.04
DocBlipVQA [29]>3.5B-47.6250.6946.3130.7355.22
Hi-VT5* [33]784M71.3528.8918.2126.176.8458.99
DocFormerv2concat [6]784M76.4048.4450.8248.0617.6759.04
GRAMC-Former864M77.6045.4747.6344.9114.3456.99
GRAM859M80.3251.1553.6750.3518.4063.23
Hi-VT5*†[33]784M73.5149.1849.2948.3513.3065.95
DocFormerv2†concat [6]784M76.7750.7952.7049.6117.3365.14
GRAM†C-Former864M78.1250.9755.1550.4617.2661.04
GRAM†859M79.6753.3656.8352.3219.9665.43
", + "image_path": "9fe7fa4024986a36d84a9c523ece8f2ed6b31860035b2b819ac2fdf2e218ed55.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 409, + 285, + 497 + ], + "blocks": [ + { + "bbox": [ + 46, + 357, + 546, + 390 + ], + "lines": [ + { + "bbox": [ + 46, + 357, + 546, + 390 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 546, + 390 + ], + "type": "text", + "content": "Table 1. Quantitative Results. We present ANLS results for the MPDocVQA [33] and DUDE [18] test sets. The methods are grouped according to the model type and size, starting from encoder-only models (top), T5-base models (middle) and T5-large models (bottom). " + }, + { + "bbox": [ + 46, + 357, + 546, + 390 + ], + "type": "inline_equation", + "content": "{}^{ \\dagger }" + }, + { + "bbox": [ + 46, + 357, + 546, + 390 + ], + "type": "text", + "content": " denotes training with both MPDocVQA and DUDE." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 409, + 285, + 497 + ], + "lines": [ + { + "bbox": [ + 50, + 409, + 285, + 497 + ], + "spans": [ + { + "bbox": [ + 50, + 409, + 285, + 497 + ], + "type": "table", + "html": "
MethodTraining DataANLS
DocVQAMPDocVQADocVQAMPDocVQA
DocFormerv2concat✓✗86.6072.73
✗✓85.2876.40
✓✓86.4775.37
GRAM✓✗86.7073.12
✗✓85.2980.32
✓✓86.3278.66
", + "image_path": "df831a3e36baa8f2ebe3e2ed6f7caee0ce3a77c9095766b2b27b45d080cb233d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 506, + 287, + 540 + ], + "lines": [ + { + "bbox": [ + 46, + 506, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 287, + 540 + ], + "type": "text", + "content": "Table 2. DocVQA vs. MPDocVQA Performance. Test results over both datasets using the large model variants. A checkmark denotes whether a dataset was included in training or not." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": "layers. We extend the page layer from Docformerv2 to feature also the doc learnable embeddings. The second doc sub-layer is similar in structure to the first sub-layer, only it is initialized from scratch, with the following specification: " + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "inline_equation", + "content": "d_{ff} = 1024" + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "inline_equation", + "content": "d_{kv} = 64" + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "inline_equation", + "content": "nheads = 4" + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "inline_equation", + "content": "d = 256" + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": ". We implement 32 doc learnable tokens for each page, uniformly initialized to random values. For bias adaptation, the initial bias value is set to " + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "inline_equation", + "content": "c = 20" + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": ", with variations between encoder heads, as described in Sec. 3.3. We incorporate an additional optional compression stage using C-Former - a randomly-initialized T5 [28] tiny decoder, with an encoder mask instead of a causal one. The output sequence extracted" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 411, + 545, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 545, + 435 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 545, + 435 + ], + "type": "text", + "content": "from C-Former is " + }, + { + "bbox": [ + 304, + 411, + 545, + 435 + ], + "type": "inline_equation", + "content": "N_{c} = 256" + }, + { + "bbox": [ + 304, + 411, + 545, + 435 + ], + "type": "text", + "content": ". Finally the decoder is initialized with pretrained weights from DocformerV2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 435, + 546, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 435, + 546, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 546, + 544 + ], + "type": "text", + "content": "The model is trained with the Hugging Face Trainer [17] for " + }, + { + "bbox": [ + 304, + 435, + 546, + 544 + ], + "type": "inline_equation", + "content": "200k" + }, + { + "bbox": [ + 304, + 435, + 546, + 544 + ], + "type": "text", + "content": " steps, starting with a warm-up of " + }, + { + "bbox": [ + 304, + 435, + 546, + 544 + ], + "type": "inline_equation", + "content": "1k" + }, + { + "bbox": [ + 304, + 435, + 546, + 544 + ], + "type": "text", + "content": " steps, with linear learning rate decay. We use learning rates of " + }, + { + "bbox": [ + 304, + 435, + 546, + 544 + ], + "type": "inline_equation", + "content": "3e^{-5}" + }, + { + "bbox": [ + 304, + 435, + 546, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 435, + 546, + 544 + ], + "type": "inline_equation", + "content": "1e^{-4}" + }, + { + "bbox": [ + 304, + 435, + 546, + 544 + ], + "type": "text", + "content": " for the already pretrained encoder and decoder weights, versus the newly initialized doc sub-layer weights. Training is performed on a cluster of " + }, + { + "bbox": [ + 304, + 435, + 546, + 544 + ], + "type": "inline_equation", + "content": "8 \\times A100" + }, + { + "bbox": [ + 304, + 435, + 546, + 544 + ], + "type": "text", + "content": " GPUs, each with 40GB of RAM. During training, each page encoder receives 800 tokens, dealing with up to 4 pages. During testing, we increase the maximum length of tokens to 8,000." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "content": "Baselines We report the results of previous work on both MP-DocVQA and DUDE datasets (if those exist), including the NLP-based Longformer [7] and BigBird [38], which were adapted to MPDocVQA by [33]; LayoutLMv3 [16], originally designed for DocVQA; and Hi-VT5 [33] and T5-2D [32], specifically suggested for MP-DocVQA Task. We also add for reference the results of methods published in the leader-boards of MPDocVQA and DUDE, which do not have corresponding papers, including DocGptVQA [30], DocBlipVQA [30], and Hi-VT5_beamsearch [15] ([15] was trained on both MP-DocVQA and DUDE). In our approach, we present two variations: GRAM and " + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{GRAM}_{C - \\text{Former}}" + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "content": ". While GRAM utilizes the full length of the encoder output," + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15603" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 78, + 103, + 227, + 209 + ], + "blocks": [ + { + "bbox": [ + 73, + 71, + 157, + 83 + ], + "lines": [ + { + "bbox": [ + 73, + 71, + 157, + 83 + ], + "spans": [ + { + "bbox": [ + 73, + 71, + 157, + 83 + ], + "type": "text", + "content": "How many diagrams are there?" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 74, + 84, + 102, + 91 + ], + "lines": [ + { + "bbox": [ + 74, + 84, + 102, + 91 + ], + "spans": [ + { + "bbox": [ + 74, + 84, + 102, + 91 + ], + "type": "text", + "content": "HIVT5:\"2\"" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 74, + 91, + 102, + 99 + ], + "lines": [ + { + "bbox": [ + 74, + 91, + 102, + 99 + ], + "spans": [ + { + "bbox": [ + 74, + 91, + 102, + 99 + ], + "type": "text", + "content": "GRAM:\"3\"" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 78, + 103, + 227, + 209 + ], + "lines": [ + { + "bbox": [ + 78, + 103, + 227, + 209 + ], + "spans": [ + { + "bbox": [ + 78, + 103, + 227, + 209 + ], + "type": "image", + "image_path": "58c8aa6b0d7b67224e400bd0a24a675c85c4896550ca2195790edc7174995db2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 236, + 72, + 330, + 82 + ], + "lines": [ + { + "bbox": [ + 236, + 72, + 330, + 82 + ], + "spans": [ + { + "bbox": [ + 236, + 72, + 330, + 82 + ], + "type": "text", + "content": "Which month shows the hurricane?" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 236, + 83, + 300, + 91 + ], + "lines": [ + { + "bbox": [ + 236, + 83, + 300, + 91 + ], + "spans": [ + { + "bbox": [ + 236, + 83, + 300, + 91 + ], + "type": "text", + "content": "HiVT5:\"not answerable\"" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 217, + 546, + 240 + ], + "lines": [ + { + "bbox": [ + 46, + 217, + 546, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 546, + 240 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison between our approach and Hi-VT5 [33] indicate that the integration of our global-local encoder enhances reasoning capabilities, especially when the inquiries require multi-page context." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 237, + 107, + 353, + 206 + ], + "blocks": [ + { + "bbox": [ + 236, + 91, + 271, + 99 + ], + "lines": [ + { + "bbox": [ + 236, + 91, + 271, + 99 + ], + "spans": [ + { + "bbox": [ + 236, + 91, + 271, + 99 + ], + "type": "text", + "content": "GRAM:\"July\"" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 237, + 107, + 353, + 206 + ], + "lines": [ + { + "bbox": [ + 237, + 107, + 353, + 206 + ], + "spans": [ + { + "bbox": [ + 237, + 107, + 353, + 206 + ], + "type": "image", + "image_path": "a464058a112ba5afcf517b166a284d65eb3ee38e011249701fe9d365f0757b85.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 360, + 103, + 518, + 205 + ], + "blocks": [ + { + "bbox": [ + 359, + 72, + 452, + 83 + ], + "lines": [ + { + "bbox": [ + 359, + 72, + 452, + 83 + ], + "spans": [ + { + "bbox": [ + 359, + 72, + 452, + 83 + ], + "type": "text", + "content": "What is the eps code for little rock?" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 360, + 84, + 393, + 91 + ], + "lines": [ + { + "bbox": [ + 360, + 84, + 393, + 91 + ], + "spans": [ + { + "bbox": [ + 360, + 84, + 393, + 91 + ], + "type": "text", + "content": "HiVT5:\"EPS\"" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 360, + 92, + 397, + 99 + ], + "lines": [ + { + "bbox": [ + 360, + 92, + 397, + 99 + ], + "spans": [ + { + "bbox": [ + 360, + 92, + 397, + 99 + ], + "type": "text", + "content": "GRAM:\"AR01\"" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 360, + 103, + 518, + 205 + ], + "lines": [ + { + "bbox": [ + 360, + 103, + 518, + 205 + ], + "spans": [ + { + "bbox": [ + 360, + 103, + 518, + 205 + ], + "type": "image", + "image_path": "f6c8b74b3eaf938c1c59278d2c880c02180af5a810fd497fbfb52742cbac40a2.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 260, + 288, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 260, + 288, + 285 + ], + "spans": [ + { + "bbox": [ + 46, + 260, + 288, + 285 + ], + "type": "inline_equation", + "content": "\\mathrm{GRAM}_{C - \\text{Former}}" + }, + { + "bbox": [ + 46, + 260, + 288, + 285 + ], + "type": "text", + "content": " allows the user to control the trade-off between performance and latency." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 285, + 287, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 285, + 287, + 429 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 287, + 429 + ], + "type": "text", + "content": "To ensure a fair comparison, since we use the pretrained model of DocFormerv2 [6], we implement two additional baselines, referred to as Hi-VT5* and DocFormerv2concat. The first follows a similar structure as Hi-VT5 [33], with the encoder originating from DocFormerv2, however without the page answer prediction, as it does not exist in DUDE. The second recreates the approach of [18], where only the textual tokens of all pages are concatenated to one long sequence, then passing through the DocFormerv2 model. The second approach poses a computational burden, thus we use only 600 tokens during training per page, with up to 4 pages, and during test only 400 tokens." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 440, + 105, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 440, + 105, + 452 + ], + "spans": [ + { + "bbox": [ + 47, + 440, + 105, + 452 + ], + "type": "text", + "content": "4.2. Results" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 460, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 544 + ], + "type": "text", + "content": "We present the performance of our method over the MPDocVQA [33] and DUDE [18] datasets in Tab. 1. The methods are divided into three groups: the top contains encoder-only, and methods that rely on the T5-base model (up to " + }, + { + "bbox": [ + 46, + 460, + 287, + 544 + ], + "type": "inline_equation", + "content": "316M" + }, + { + "bbox": [ + 46, + 460, + 287, + 544 + ], + "type": "text", + "content": " parameters); the middle section, approaches that use the T5-large model (over " + }, + { + "bbox": [ + 46, + 460, + 287, + 544 + ], + "type": "inline_equation", + "content": "770M" + }, + { + "bbox": [ + 46, + 460, + 287, + 544 + ], + "type": "text", + "content": " parameters), and finally the bottom, T5-large models, trained on both datasets." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 544, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 687 + ], + "type": "text", + "content": "As can be seen, in the first group, the encoder-only NLP methods, LongFormer [7], BigBird [38] and LayoutLMv3 [16] can only handle relatively well 'extractive' style tasks as in MPDocVQA dataset [33], but often struggle with 'abstractive' questions that are more abundant in DUDE [18]. As to T5-'base' models, versus our best competitor Docformerv2concat, we obtain an improvement of " + }, + { + "bbox": [ + 46, + 544, + 287, + 687 + ], + "type": "inline_equation", + "content": "(+4\\%, +1.9\\%)" + }, + { + "bbox": [ + 46, + 544, + 287, + 687 + ], + "type": "text", + "content": " on MP-DocVQA and DUDE datasets. As to methods that combine an additional compression before decoding (Hi-VT5, Hi-VT5*), our C-Former achieves an increase in " + }, + { + "bbox": [ + 46, + 544, + 287, + 687 + ], + "type": "inline_equation", + "content": "(+8.8\\%, +16.2\\%)" + }, + { + "bbox": [ + 46, + 544, + 287, + 687 + ], + "type": "text", + "content": " over the best candidates on the MP-DocVQA and DUDE datasets." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 714 + ], + "type": "text", + "content": "As for the group of 'large' models, we include the results of T5-2D [18] DocGptVQA [30] and" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 260, + 546, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 260, + 546, + 332 + ], + "spans": [ + { + "bbox": [ + 304, + 260, + 546, + 332 + ], + "type": "text", + "content": "DocBlipVQA [29]. Note that our model surpasses DocFormerconcat, the primary baseline, achieving improvements of " + }, + { + "bbox": [ + 304, + 260, + 546, + 332 + ], + "type": "inline_equation", + "content": "(+3.9\\%, +2.7\\%)" + }, + { + "bbox": [ + 304, + 260, + 546, + 332 + ], + "type": "text", + "content": " on MP-DocVQA and DUDE, respectively. We also outperform DocGptVQA [30], a method that appears in the leaderboard of DUDE, by " + }, + { + "bbox": [ + 304, + 260, + 546, + 332 + ], + "type": "inline_equation", + "content": "+1.1\\%" + }, + { + "bbox": [ + 304, + 260, + 546, + 332 + ], + "type": "text", + "content": ", thereby obtaining SOTA results for GRAM 'large'." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 335, + 546, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 546, + 454 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 546, + 454 + ], + "type": "text", + "content": "The final category showcases large encoder-decoder models, fine-tuned on both MP-DocVQA and DUDE training sets, showcasing the benefits of augmented training data. GRAM consistently demonstrates performance gains over the baseline, illustrating its robustness across different datasets and training scenarios. Next, we present in Tab. 2 the effect of training on DocVQA vs. MPDocVQA. Our method achieves performance on-par on the single page task, while enhancing performance on the multi-page scenario by " + }, + { + "bbox": [ + 304, + 335, + 546, + 454 + ], + "type": "inline_equation", + "content": "+3.3\\%" + }, + { + "bbox": [ + 304, + 335, + 546, + 454 + ], + "type": "text", + "content": ", compared to the baseline." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 456, + 545, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 456, + 545, + 541 + ], + "spans": [ + { + "bbox": [ + 304, + 456, + 545, + 541 + ], + "type": "text", + "content": "In Fig. 4, We show qualitative results on the DUDE dataset of GRAM versus Hi-VT5* [33]. Our method demonstrates proficiency in addressing questions that involve attention over multiple pages ('how many diagrams are there'), an increased visual analysis capability ('Which month shows the hurricane?'), and heightened abstractive ability ('What is the EPS code for Little Rock?')." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 559, + 399, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 559, + 399, + 572 + ], + "spans": [ + { + "bbox": [ + 306, + 559, + 399, + 572 + ], + "type": "text", + "content": "5. Ablation Study" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "text", + "content": "We perform an ablation study on our approach, evaluating the influence of each constituent component using DUDE's validation set [18]. This validation set enables the grouping of documents by their respective page counts: 1, 2-4, 5-10, 11-end, encompassing 1747, 2259, 1062, 1241 samples in each category, respectively. Our investigation delves into the impact of the number of doc tokens and the bias adaptation methods. Moreover, we employ the C-Former for sequence compression, adjusting the compression ratio and examining the balance between performance and latency (see supplementary for more details)." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15604" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 286, + 232 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 286, + 232 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 286, + 232 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 286, + 232 + ], + "type": "table", + "html": "
#Doc TokensBias TypeCompression DimensionANLS by Number of Pages DUDE validation dataset
All12-45-1011-end
XXX46.1647.1848.6643.3442.57
16X46.3948.3549.0643.1641.56
32DecayingX47.8849.2949.9045.9043.94
64X46.7047.9849.2244.0042.60
32XX47.5249.8549.9344.9042.10
ConstantX46.1447.4148.1344.4442.19
DecayingX47.8849.2949.9045.9043.94
32Decaying839.8339.7341.4136.9839.52
3240.4240.9541.6438.1239.39
25641.9942.5743.7738.4041.01
102442.5642.9744.3038.9441.93
409643.5944.7544.6440.5442.67
", + "image_path": "7c525b2d9791bd906cfef1141145ec50a053883b2824ca8f450e6b8a35ab3d49.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 298, + 288, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 298, + 288, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 298, + 288, + 574 + ], + "type": "text", + "content": "GRAM Components We focus our initial exploration on the impact of the number of doc tokens " + }, + { + "bbox": [ + 46, + 298, + 288, + 574 + ], + "type": "inline_equation", + "content": "N_{g}" + }, + { + "bbox": [ + 46, + 298, + 288, + 574 + ], + "type": "text", + "content": ". As can be seen in Tab. 3, while " + }, + { + "bbox": [ + 46, + 298, + 288, + 574 + ], + "type": "inline_equation", + "content": "N_{g} = 16" + }, + { + "bbox": [ + 46, + 298, + 288, + 574 + ], + "type": "text", + "content": " leads to performance on par with not using doc tokens at all, for the optimal value of " + }, + { + "bbox": [ + 46, + 298, + 288, + 574 + ], + "type": "inline_equation", + "content": "N_{g} = 32" + }, + { + "bbox": [ + 46, + 298, + 288, + 574 + ], + "type": "text", + "content": ", we obtain an increase of " + }, + { + "bbox": [ + 46, + 298, + 288, + 574 + ], + "type": "inline_equation", + "content": "+1.7\\%" + }, + { + "bbox": [ + 46, + 298, + 288, + 574 + ], + "type": "text", + "content": " in ANLS. Shifting our focus to bias adaptation methods, Tab. 3 shows that using constant bias has a negative effect on the results, suggesting this method is not flexible enough in maintaining a balance between the page and doc tokens. However, our decaying bias-adaptation approach does improve results overall, versus no-bias " + }, + { + "bbox": [ + 46, + 298, + 288, + 574 + ], + "type": "inline_equation", + "content": "(+0.36\\%)" + }, + { + "bbox": [ + 46, + 298, + 288, + 574 + ], + "type": "text", + "content": ", especially for longer documents " + }, + { + "bbox": [ + 46, + 298, + 288, + 574 + ], + "type": "inline_equation", + "content": "(+1\\%" + }, + { + "bbox": [ + 46, + 298, + 288, + 574 + ], + "type": "text", + "content": " improvement for 5-10 pages and " + }, + { + "bbox": [ + 46, + 298, + 288, + 574 + ], + "type": "inline_equation", + "content": "+1.84\\%" + }, + { + "bbox": [ + 46, + 298, + 288, + 574 + ], + "type": "text", + "content": " for 11 pages and more). This is to be expected, since incorporating new doc tokens and increasing their importance can potentially affect single-page performance. Finally, in Tab. 4, we reinforce our choice of pages as semantic logical units for MPDocVQA. We first ablate our method with and without page embedding. Next, we compare our page-based division with varying fixed-length division of tokens for encoder. Results in Tab. 4 clearly demonstrate an advantage towards page-level encoding in MPDocVQA. This aligns with our initial assumption that structured documents are often designed with page-division in mind." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": "Performance-Latency Trade-off We assess the impact of C-Former on performance, considering compression output lengths of 8, 32, 256, 1024, 4096. Note, performance gradually improves with an increase in the compression output length. However, longer output lengths correspond to heightened model latency. Note that using C-Former for shorter documents can be redundant, as there is little to no compression compared to the input sequence length and results decrease. In Fig. 5, we scrutinize the trade-off between computational efficiency and compression rate" + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 309, + 70, + 545, + 167 + ], + "blocks": [ + { + "bbox": [ + 46, + 241, + 288, + 275 + ], + "lines": [ + { + "bbox": [ + 46, + 241, + 288, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 241, + 288, + 275 + ], + "type": "text", + "content": "Table 3. GRAM Ablation Study. Results on DUDE validation set ablating over (a) the dimension of doc tokens, (b) the attention bias employed and (c) the C-former input dimension." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 70, + 545, + 167 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 545, + 167 + ], + "type": "table", + "html": "
Page EmbeddingSegment LengthANLS by Number of Pages DUDE validation dataset
All12-45-1011-end
✓✗47.8849.2949.9045.9043.94
✗✗46.1248.7448.1143.5940.99
✓25645.2246.3846.6944.1341.83
✓51245.0945.9047.3242.6541.98
✓102444.3944.9846.6341.6941.78
", + "image_path": "1b65f40706a96aa77ea17d9ec97973110ae15eaeae3341626ff31ca4e2167582.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 175, + 545, + 209 + ], + "lines": [ + { + "bbox": [ + 305, + 175, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 305, + 175, + 545, + 209 + ], + "type": "text", + "content": "Table 4. The Significance of Pages as Semantic Units. Results on DUDE validation set ablating over (a) utilization of page-embedding, (b) segment length for fixed-size encoding inputs." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 321, + 230, + 533, + 342 + ], + "blocks": [ + { + "bbox": [ + 321, + 230, + 533, + 342 + ], + "lines": [ + { + "bbox": [ + 321, + 230, + 533, + 342 + ], + "spans": [ + { + "bbox": [ + 321, + 230, + 533, + 342 + ], + "type": "image", + "image_path": "148ed5b5337a0df77c963769acb9ace696da8e29227d9d1bcf6b5fd25c5b981c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 352, + 545, + 385 + ], + "lines": [ + { + "bbox": [ + 305, + 352, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 305, + 352, + 545, + 385 + ], + "type": "text", + "content": "Figure 5. Latency comparison. We compare the dependency between overall latency and the number of pages in input document for GRAM, " + }, + { + "bbox": [ + 305, + 352, + 545, + 385 + ], + "type": "inline_equation", + "content": "\\mathrm{GRAM}_{C - \\text{Former}}" + }, + { + "bbox": [ + 305, + 352, + 545, + 385 + ], + "type": "text", + "content": ", DocFormerv2concat and Hi-VT5." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 414, + 545, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 414, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 304, + 414, + 545, + 511 + ], + "type": "text", + "content": "by comparing to DocFormerv2concat [6] and Hi-VT5* [33]. We discover that DocFormerv2concat reaches a memory limit at approximately 20 pages, due to its quadratic memory increase with sequence length. At this juncture, GRAM" + }, + { + "bbox": [ + 304, + 414, + 545, + 511 + ], + "type": "inline_equation", + "content": "_{C-Former}" + }, + { + "bbox": [ + 304, + 414, + 545, + 511 + ], + "type": "text", + "content": " surpasses DocFormerv2concat by performing 3.5 seconds faster. Notably, GRAM" + }, + { + "bbox": [ + 304, + 414, + 545, + 511 + ], + "type": "inline_equation", + "content": "_{C-Former}" + }, + { + "bbox": [ + 304, + 414, + 545, + 511 + ], + "type": "text", + "content": " can gracefully handle documents surpassing 300 pages, effectively bridging the gap between performance and latency." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 533, + 384, + 546 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 533, + 384, + 546 + ], + "spans": [ + { + "bbox": [ + 306, + 533, + 384, + 546 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "type": "text", + "content": "Our method, termed GRAM, extends existing single-page document models to efficiently handle multi-page documents without necessitating computationally-intensive pretraining. Leveraging the single-page encoder for local page-level comprehension, we introduce document learnable tokens and designated layers, enabling seamless information exchange across pages. Additionally, our proposed bias adaptation method enforces effective utilization of our newly introduced document tokens. The incorporation of a C-Former model reduces sequence length, balancing quality with latency in the decoding step. Extensive experiments demonstrate GRAM's state-of-the-art performance across multi-page DocVQA benchmarks." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15605" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "text", + "content": "[1] Aviad Aberdam, Ron Litman, Shahar Tsiper, Oron Anschel, Ron Slossberg, Shai Mazor, R Manmatha, and Pietro Perona. Sequence-to-sequence contrastive learning for text recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15302-15312, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 158, + 287, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 287, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 287, + 191 + ], + "type": "text", + "content": "[2] Aviad Aberdam, Roy Ganz, Shai Mazor, and Ron Litman. Multimodal semi-supervised learning for text recognition. arXiv preprint arXiv:2205.03873, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 193, + 287, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 193, + 287, + 236 + ], + "spans": [ + { + "bbox": [ + 53, + 193, + 287, + 236 + ], + "type": "text", + "content": "[3] Aviad Aberdam, David Bensaïd, Alona Golts, Roy Ganz, Oren Nuriel, Royee Tichauer, Shai Mazor, and Ron Litman. Clipter: Looking at the bigger picture in scene text recognition. arXiv preprint arXiv:2301.07464, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 238, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 238, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 238, + 287, + 293 + ], + "type": "text", + "content": "[4] Joshua Ainslie, Tao Lei, Michiel de Jong, Santiago Ontañón, Siddhartha Brahma, Yury Zemlyanskiy, David Uthus, Mandy Guo, James Lee-Thorp, Yi Tay, et al. Colt5: Faster long-range transformers with conditional computation. arXiv preprint arXiv:2303.09752, 2023. 1, 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 294, + 287, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 294, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 53, + 294, + 287, + 350 + ], + "type": "text", + "content": "[5] Srikar Appalaraju, Bhavan Jasani, Bhargava Urala Kota, Yusheng Xie, and R Manmatha. Docformer: End-to-end transformer for document understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 993-1003, 2021. 1, 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 351, + 287, + 394 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 351, + 287, + 394 + ], + "spans": [ + { + "bbox": [ + 53, + 351, + 287, + 394 + ], + "type": "text", + "content": "[6] Srikar Appalaraju, Peng Tang, Qi Dong, Nishant Sankaran, Yichu Zhou, and R Manmatha. Docformerv2: Local features for document understanding. arXiv preprint arXiv:2306.01733, 2023. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 396, + 287, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 396, + 287, + 429 + ], + "spans": [ + { + "bbox": [ + 53, + 396, + 287, + 429 + ], + "type": "text", + "content": "[7] Iz Beltagy, Matthew E Peters, and Arman Cohan. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150, 2020. 1, 2, 4, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 430, + 287, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 430, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 53, + 430, + 287, + 485 + ], + "type": "text", + "content": "[8] Ali Furkan Biten, Ruben Tito, Andres Mafla, Lluis Gomez, Marçal Rusinol, Ernest Valveny, CV Jawahar, and Dimosthenis Karatzas. Scene text visual question answering. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4291-4301, 2019. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 487, + 287, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 487, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 53, + 487, + 287, + 520 + ], + "type": "text", + "content": "[9] Ali Furkan Biten, Ron Litman, Yusheng Xie, Srikar Appalaraju, and R Manmatha. Latr: Layout-aware transformer for scene-text vqa. pages 16548-16558, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 521, + 287, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 521, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 48, + 521, + 287, + 553 + ], + "type": "text", + "content": "[10] Aydar Bulatov, Yuri Kuratov, and Mikhail S Burtsev. Scaling transformer to 1m tokens and beyond with rmt. arXiv preprint arXiv:2304.11062, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 555, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 599 + ], + "type": "text", + "content": "[11] Shouyuan Chen, Sherman Wong, Liangjian Chen, and Yuandong Tian. Extending context window of large language models via positional interpolation. arXiv preprint arXiv:2306.15595, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 600, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 600, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 600, + 287, + 644 + ], + "type": "text", + "content": "[12] Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc V Le, and Ruslan Salakhutdinov. Transformer-xl: Attentive language models beyond a fixed-length context. arXiv preprint arXiv:1901.02860, 2019. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 646, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 677 + ], + "type": "text", + "content": "[13] Robert M French. Catastrophic forgetting in connectionist networks. Trends in cognitive sciences, 3(4):128-135, 1999. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[14] Roy Ganz, Oren Nuriel, Aviad Aberdam, Yair Kittenplon, Shai Mazor, and Ron Litman. Towards models that can see and read. arXiv preprint arXiv:2301.07389, 2023. 3, 4" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "text", + "content": "[15] JiangLong He, Mamatha N, Shiv Vignesh, and Deepak Kumar. Hivt5beam. Hi-VT5 model pretrained with private custom document collection using span masking objective. Pretrained model is then trained with DUDE dataset and MultiPage DocVQA dataset, 2023. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 129, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 184 + ], + "type": "text", + "content": "[16] Yupan Huang, Tengchao Lv, Lei Cui, Yutong Lu, and Furu Wei. Layoutmv3: Pre-training for document ai with unified text and image masking. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4083-4091, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 186, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 207 + ], + "type": "text", + "content": "[17] HuggingFace. Huggingface trainer. https://huggingface.co/docs/transformers/training. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 209, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 262 + ], + "type": "text", + "content": "[18] Jordy Landeghem, Rubén Tito, Łukasz Borchmann, Michal Pietruszka, Paweł Józiak, Rafał Powalski, Dawid Jurkiewicz, Mickaël Coustaty, Bertrand Ackaert, Ernest Valveny, et al. Document understanding dataset and evaluation (dude). arXiv preprint arXiv:2305.08455, 2023. 1, 3, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 264, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 308 + ], + "type": "text", + "content": "[19] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 309, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 309, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 308, + 309, + 545, + 364 + ], + "type": "text", + "content": "[20] Ron Litman, Oron Anschel, Shahar Tsiper, Roeit Litman, Shai Mazor, and R Manmatha. Scatter: selective context attentional scene text recognizer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11962-11972, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 365, + 545, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 545, + 397 + ], + "type": "text", + "content": "[21] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 399, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 443 + ], + "type": "text", + "content": "[22] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 444, + 545, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 444, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 308, + 444, + 545, + 488 + ], + "type": "text", + "content": "[23] Oren Nuriel, Sharon Fogel, and Ron Litman. Textadain: Paying attention to shortcut learning in text recognizers. In European Conference on Computer Vision, pages 427-445. Springer, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 490, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 490, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 490, + 545, + 510 + ], + "type": "text", + "content": "[24] R OpenAI. Gpt-4 technical report. arXiv, pages 2303–08774, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 513, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 513, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 513, + 545, + 567 + ], + "type": "text", + "content": "[25] Qiming Peng, Yinxu Pan, Wenjin Wang, Bin Luo, Zhenyu Zhang, Zhengjie Huang, Teng Hu, Weichong Yin, Yongfeng Chen, Yin Zhang, et al. Ernie-Layout: Layout knowledge enhanced pre-training for visually-rich document understanding. arXiv preprint arXiv:2210.06155, 2022. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 568, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 545, + 645 + ], + "type": "text", + "content": "[26] Rafal Powalski, Łukasz Borchmann, Dawid Jurkiewicz, Tomasz Dwojak, Michal Pietruszka, and Gabriela Pałka. Going full-tilt boogie on document understanding with text-image layout transformer. In Document Analysis and Recognition-ICDAR 2021: 16th International Conference, Lausanne, Switzerland, September 5–10, 2021, Proceedings, Part II 16, pages 732–747. Springer, 2021. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "text", + "content": "[27] Ofir Press, Noah A Smith, and Mike Lewis. Train short, test long: Attention with linear biases enables input length extrapolation. arXiv preprint arXiv:2108.12409, 2021. 1, 2, 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[28] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15606" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485-5551, 2020. 2, 3, 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 185 + ], + "type": "text", + "content": "[29] RenZhou, QiaolingDeng, XinfengChang, LuyanWang, XiaochenHu, HuiLi, and YaqiangWu. Docclipvqa. We integrated the prediction outputs from the UDOP model and Blip2 to enhance our results, and we optimized the image encoder and included page number features to address the challenge of multi-page documents. GPT to generate python-like modular programs., 2023. 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 186, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 288, + 262 + ], + "type": "text", + "content": "[30] RenZhou, QiaolingDeng, XinfengChang, LuyanWang, XiaochenHu, HuiLi, and YaqiangWu. Docgptvqa. We integrated the prediction outputs from the UDOP model and Blip2 to enhance our results, and we optimized the image encoder and included page number features to address the challenge of multi-page documents. GPT to generate python-like modular programs., 2023. 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 263, + 287, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 263, + 287, + 329 + ], + "spans": [ + { + "bbox": [ + 48, + 263, + 287, + 329 + ], + "type": "text", + "content": "[31] Zineng Tang, Ziyi Yang, Guoxin Wang, Yuwei Fang, Yang Liu, Chenguang Zhu, Michael Zeng, Cha Zhang, and Mohit Bansal. Unifying vision, text, and layout for universal document processing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19254-19264, 2023. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 331, + 287, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 287, + 386 + ], + "type": "text", + "content": "[32] Rubén Tito, Dimosthenis Karatzas, and Ernest Valveny. Document collection visual question answering. In Document Analysis and Recognition-ICDAR 2021: 16th International Conference, Lausanne, Switzerland, September 5–10, 2021, Proceedings, Part II 16, pages 778–792. Springer, 2021. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 387, + 287, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 387, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 48, + 387, + 287, + 420 + ], + "type": "text", + "content": "[33] Rubén Tito, Dimosthenis Karatzas, and Ernest Valveny. Hierarchical multimodal transformers for multi-page docvqa. arXiv preprint arXiv:2212.05935, 2022. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 422, + 287, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 465 + ], + "type": "text", + "content": "[34] Maria Tsimpoukelli, Jacob L Menick, Serkan Cabi, SM Eslami, Oriol Vinyls, and Felix Hill. Multimodal few-shot learning with frozen language models. Advances in Neural Information Processing Systems, 34:200-212, 2021. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 467, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 467, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 287, + 510 + ], + "type": "text", + "content": "[35] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 512, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 287, + 567 + ], + "type": "text", + "content": "[36] Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, and Ming Zhou. Layoutm: Pre-training of text and layout for document image understanding. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 1192-1200, 2020. 1, 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 568, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 622 + ], + "type": "text", + "content": "[37] Yang Xu, Yiheng Xu, Tengchao Lv, Lei Cui, Furu Wei, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Wanxiang Che, et al. Layoutmv2: Multi-modal pre-training for visually-rich document understanding. arXiv preprint arXiv:2012.14740, 2020. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 624, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 689 + ], + "type": "text", + "content": "[38] Manzil Zaheer, Guru Guruganesh, Kumar Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontanon, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, et al. Big bird: Transformers for longer sequences. Advances in neural information processing systems, 33:17283-17297, 2020. 1, 2, 4, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[39] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 326, + 72, + 545, + 106 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 72, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 545, + 106 + ], + "type": "text", + "content": "Lit: Zero-shot transfer with locked-image text tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18123-18133, 2022. 4" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15607" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/GROUNDHOG_ Grounding Large Language Models to Holistic Segmentation/bdfcf2de-41e8-4d27-af7d-65821e6b79f8_content_list.json b/2024/GROUNDHOG_ Grounding Large Language Models to Holistic Segmentation/bdfcf2de-41e8-4d27-af7d-65821e6b79f8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ab73ffd21a0eacdc2071bcafea9775d62665ce1f --- /dev/null +++ b/2024/GROUNDHOG_ Grounding Large Language Models to Holistic Segmentation/bdfcf2de-41e8-4d27-af7d-65821e6b79f8_content_list.json @@ -0,0 +1,1757 @@ +[ + { + "type": "text", + "text": "GROUNDHOG: Grounding Large Language Models to Holistic Segmentation", + "text_level": 1, + "bbox": [ + 84, + 114, + 883, + 140 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yichi Zhang $^{1\\dagger}$ , Ziqiao Ma $^{1\\dagger}$ , Xiaofeng Gao $^{2}$ , Suhaila Shakiah $^{2}$ , Qiaozi Gao $^{2}$ , Joyce Chai $^{1}$", + "bbox": [ + 135, + 154, + 831, + 172 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Michigan, 2Amazon AGI", + "bbox": [ + 328, + 172, + 643, + 190 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zhangyic@umich.edu", + "bbox": [ + 377, + 191, + 591, + 208 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://groundhog-ml1m.github.io/", + "bbox": [ + 287, + 210, + 676, + 226 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/817673d7b306725870a94ad6aeeaef6db83f1b4c43b730771fabd7dec2fffb7c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 234, + 282, + 352 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/54bd5a7be146835c9c4ce0f5c2ba8ebb098ebef2ac63383612da255f4fbb58c2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 84, + 354, + 104, + 375 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Describe the image briefly.", + "bbox": [ + 125, + 358, + 272, + 371 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A man and a little girl are sitting in a shopping cart.", + "bbox": [ + 89, + 380, + 230, + 404 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ccbb31425917613174b95d2ad4b065a5c65a3dbc814f2029381d7d512e7f56c9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 258, + 386, + 284, + 406 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4a1b2c5bb92f244b2bcbd940a39ec34b08e3c2be425deb32a6c218fe0fce444a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 236, + 485, + 352 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4c68b29cd239bf69488cce1fd1020619a1d355e9b22234fbd3d25a5795fbf0f2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 354, + 307, + 375 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generate a short caption.", + "bbox": [ + 328, + 358, + 467, + 371 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Two dogs are playing with a stick in a field.", + "bbox": [ + 292, + 378, + 437, + 404 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5429eeab0ac8ed394c150f376deea831fafd34399aeb88b0b3d9dd27b40df257.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 485, + 236, + 687, + 406 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e322558dc6e91ba19a2b8c04497025b653f07fbde7cde1448ef382f777bef8c9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 691, + 268, + 712, + 289 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Describe the given picture in very detail.", + "bbox": [ + 730, + 263, + 849, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this image, we can see a boat on the water. There are few people inside the boat. There are also few people on the rocks. In the background, there are trees, hills, and sky.", + "bbox": [ + 694, + 309, + 844, + 390 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/02a2638a71e499bc68c7a9ef19695b53f19802294ce4a234f52ee779038b4446.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 861, + 378, + 883, + 397 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5f446983be845abc58bf5800a0a7fa28bc511843f6a4d7208223e581cd2d0421.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 84, + 422, + 367, + 580 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d18ad8ab055918dc33ea563474a1b189b9950d7a8ad89f0170ffd432ceefe8c6.jpg", + "image_caption": [ + "Figure 1. We propose GROUNDHOG, a multimodal large language model that enhances its text output with pixel-level phrase grounding across diverse semantic granularities. The figure demonstrates outputs from our model on the four task types we considered in this work." + ], + "image_footnote": [], + "bbox": [ + 84, + 583, + 104, + 612 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Could you please segment out \"laptop barely onscreen not apple\" in the image?", + "bbox": [ + 125, + 585, + 348, + 608 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Laptop barely onscreen not apple.", + "bbox": [ + 91, + 617, + 277, + 628 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d08af1999556258e367d3ae3b17b97e5b763df55800993414ee9d20b69ac081d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 343, + 613, + 366, + 631 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/17ff19cbf741628e010d8814aee4f465abf5a21a5d15d5ba303fca2f1315db21.jpg", + "image_caption": [ + "(a) Grounded Image Captioning (GIC)." + ], + "image_footnote": [], + "bbox": [ + 405, + 422, + 609, + 578 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0c37a129018dcdce8c406b75452fcca98cc9a672fc1c9bff5644eb8b01b36352.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 405, + 585, + 424, + 612 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "What date is shown on the calendar?", + "bbox": [ + 437, + 585, + 584, + 607 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Feb 2.", + "bbox": [ + 403, + 617, + 439, + 628 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/54b342a5cfb2ddaa4077cad87215fd00418b68af00a4e8c4bf7dcc06fbd442c5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 614, + 614, + 632 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cc47e31e680bef21ebce829533cb6b09ba7c79499bd4f0a8c6a546ba1a97f8b5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 422, + 885, + 579 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ad2fbb69656e9111eb7e474fea7b1d498f28be11097c5c19d527dbf7c82bc28a.jpg", + "image_caption": [ + "(d) Referential Dialogue (RD)." + ], + "image_footnote": [], + "bbox": [ + 663, + 584, + 684, + 611 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Provide a distinct description for that $< \\text{PTR}>$", + "bbox": [ + 692, + 584, + 872, + 607 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Blue container with apples.", + "bbox": [ + 666, + 617, + 816, + 630 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9c1d57fcf0661fc141f0ebe98ac24998aabf81883db00c4a04f8ac274d27dea9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 859, + 613, + 883, + 631 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 686, + 312, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Most multimodal large language models (MLLMs) learn language-to-object grounding through causal language modeling where grounded objects are captured by bounding boxes as sequences of location tokens. This paradigm lacks pixel-level representations that are important for fine-grained visual understanding and diagnosis. In this work, we introduce GROUNDHOG, an MLLM developed by grounding Large Language Models to holistic segmentation. GROUNDHOG incorporates a masked feature extractor and converts extracted features into visual entity tokens for the MLLM backbone, which then con", + "bbox": [ + 75, + 705, + 470, + 872 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "nects groundable phrases to unified grounding masks by retrieving and merging the entity masks. To train GROUNDHOG, we carefully curated M3G2, a grounded visual instruction tuning dataset with Multi-Modal Multi-Grained Grounding, by harvesting a collection of segmentation-grounded datasets with rich annotations. Our experimental results show that GROUNDHOG achieves superior performance on various language grounding tasks without task-specific fine-tuning, and significantly reduces object hallucination. GROUNDHOG also demonstrates better grounding towards complex forms of visual input and provides easy-to-understand diagnosis in failure cases.", + "bbox": [ + 496, + 688, + 890, + 869 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 236, + 0, + 807, + 18 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 325, + 16, + 722, + 30 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 292, + 31, + 753, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Work done during internship at Amazon AGI.", + "bbox": [ + 94, + 886, + 346, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14227", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 80, + 90, + 207, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multimodal large language models (MLLMs) have received an increasing amount of attention to address tasks that necessitate non-linguistic knowledge, e.g., perception and reasoning about the visual world [39, 84]. For fine-grained visual understanding, grounded MLLMs often learn language-to-object grounding by causal language modeling, where grounded objects are captured by bounding boxes as sequences of location tokens. However, bounding boxes are insufficient in indicating amorphous stuff [5], semantic parts of objects [23], finer-grained regions with irregular shapes [26], or groups of instances at the same time. As a result, a single bounding box can often include other irrelevant semantics in order to engulf the target entities, leading to ambiguity in detection. In addition, the generated box coordinate lacks interpretability. When the model hallucinates, such as incorrectly predicting the association between objects and language, it is hard to diagnose whether the problem is due to the model's failure to detect the object, or its incorrect alignment of the object with language.", + "bbox": [ + 80, + 114, + 467, + 400 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these issues, in this work, we introduce GROUNDHOG, an MLLM developed by grounding Large Language Models to holistic segmentation. Our goal of language grounding is to connect text spans that refer to or can be deduced from visual information, termed as groundable phrases [50], to their corresponding regions of visual entities (Figure 1). GROUNDHOG incorporates a masked feature extractor that takes an input image and a set of class-agnostic entity mask proposals, and converts each mask's features into visual entity tokens for an MLLM backbone. This MLLM then connects groundable phrases to unified grounding masks by retrieving and merging the entity masks. Compared to previous grounded MLLMs, GROUNDHOG unlocks unprecedented pixel-level vision-language alignment. It naturally supports visual pointers as input, and can plug-in-and-play with any choice of mask proposal networks, e.g., Segment Anything Model (SAM) [35], domain-specific semantic segmentation models, or user-provided mask candidates. We introduce an enhanced Mask2Former [10] as our default mask proposal network, which detects regions at multiple granularities, e.g., instances (things and stuff), semantic parts, and visual text, leading to a holistic coverage of visual semantics.", + "bbox": [ + 80, + 402, + 467, + 747 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To train GROUNDHOG, we curated a Multi-Modal Multi-Grained Grounding (M3G2) dataset consisting of 2.5M text-image pairs for visually grounded instruction tuning, consisting of 36 sub-problems derived and augmented from 27 existing datasets. We present extensive experiments on vision-language tasks that require grounding, including grounded language generation with minimal object hallucination, language-guided segmentation, visual question answering with answer grounding, and referential dialog with spatial pointer inputs (Figure 1). Our empirical results show", + "bbox": [ + 80, + 750, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "that GROUNDHOG, without task-specific fine-tuning, can achieve superior or comparable performance with previous models that either require fine-tuning or are specialized only for that dataset. In addition, GROUNDHOG has supports easy-to-understand diagnosis when grounding fails.", + "bbox": [ + 503, + 92, + 888, + 166 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Our Method: GROUNDHOG", + "text_level": 1, + "bbox": [ + 503, + 180, + 756, + 194 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The language grounding task can be succinctly delineated into two fundamental components: localization and recognition, as established in the literature [50, 68, 92]. Such categorization not only aids in the identification of object presence (objectness) without reliance on specific object classes, but also sets the stage for models to be robust in open-vocabulary settings. Building upon this framework, we formulate the grounding process as an entity segment selection problem, which involves (1) proposing entity segmentation masks where the masks encapsulate regions with discernible semantic content, and (2) recognizing the retrieved entities through the understanding of both visual and language context. Concurrently performing both tasks is where MLLMs bring a distinct advantage. This decoupled design of entity mask proposal and language-guided grounding brings several advantages. First, it allows independent improvement of the mask proposal model and MLLM, where specialized data, training, and inference setups can be applied. Second, by decoupling language grounding, it becomes straightforward to determine if a failure is due to the model's inability to propose the entity segment, or its misalignment of the object with the language, thus improving the interpretability of the whole framework. Third, as shown later, when connecting the two parts to work in tandem in a model-independent manner, the MLLM can benefit from multiple different vision specialist models in a plug-and-play fashion. In the remainder of this section, we give details of our model design.", + "bbox": [ + 503, + 205, + 888, + 626 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Building Entity Features from Masks", + "text_level": 1, + "bbox": [ + 503, + 636, + 818, + 650 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our approach assumes the availability of a mask proposal model, which is capable of generating a set of class-agnostic entity masks from an image with high coverage. In contrast to prior studies that relied on low-level features [8, 13, 45, 56], GROUNDHOG interprets the image as a collection of entities. The primary challenge then becomes the derivation of effective visual features to accurately represent these entities. To achieve a complete decoupling of the MLLM from the mask proposal model responsible for providing the masks, we propose to condition the entity features solely on the binary masks without using any embeddings from the mask proposal model. Specifically, the mask corresponding to each entity is employed to extract patch features from pretrained vision foundation models, such as CLIP [62] and DINOv2 [54], through a convolutional mask pooling layer [12]. Given that the feature map dimensions", + "bbox": [ + 503, + 659, + 888, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14228", + "bbox": [ + 480, + 946, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/371348875dd9fbb95ccac0fca9a5c92056d4e70990a409abfe0486c28873d11c.jpg", + "image_caption": [ + "Figure 2. The model architecture of GROUNDHOG model. Given a set of class-agnostic entity mask proposals, the masked feature extractor first extracts the feature of each entity as the visual input of the multi-modal large language model (left). The output hidden states of the grounding tokens are averaged and used to retrieve the entities to ground, which will be merged into a single grounding mask for the phrase. Modules are colored by their trainability: parameter-free operators (grey), frozen (blue), trainable (orange), and partially trainable (mix)." + ], + "image_footnote": [], + "bbox": [ + 91, + 90, + 883, + 340 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d4774ca35a652f1e2e464c2a4831bcc83771df0a366352dfd27e92fd20d61e16.jpg", + "image_caption": [ + "Figure 3. GROUNDHOG can take arbitrary spatial prompts that can be resolved by an interactive segmentation model, such as SAM. The placeholder pointer token $<\\mathrm{PTR}>$ will be replaced by the extracted entity features and fed as input to the model." + ], + "image_footnote": [], + "bbox": [ + 81, + 412, + 468, + 536 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "are usually smaller than those of the mask proposals, we resize the masks to match the size of the feature maps prior to pooling. The pooled features are then fed into a Multi-Layer Perceptron (MLP) network to align with the input embeddings of the MLLM. We empirically find the combination of CLIP and DINOv2 features yields the best result, and these features are added to obtain the final input visual entity tokens to the MLLM.", + "bbox": [ + 75, + 604, + 468, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Spatial Prompts Furthermore, for grounded MLLMs to be more broadly applicable, they must be capable of interpreting multi-modal user inputs, including spatial prompts. Thanks to the mask model agnostic design, GROUND-HOGcan seamlessly support such inputs. As demonstrated in Figure 3, by applying an interactive segmentation model such as Segment-Anything (SAM) [35], arbitrary spatial prompts can be translated into binary masks and processed by the same masked feature extractor we just introduced. This extracted feature for the pointed entity will replace the pointer token placeholder in the textual input.", + "bbox": [ + 75, + 734, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Language Grounding to Entity Segmentation", + "text_level": 1, + "bbox": [ + 500, + 412, + 880, + 429 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Existing box-grounded MLLMs typically append location tokens after the groundable phrases [8, 9, 56, 85]. However, this method is not readily interpretable. To alleviate this disconnect, we introduce a pair of grounding tokens $\\langle \\mathrm{GRD} \\rangle$ and $\\langle \\mathrm{/GRD} \\rangle$ to indicate the start and end of groundable phrases, with the assumption that grounding these phrases requires mapping to certain representations of visual entities irrespective of the visual modality. In Figure 2, a sentence can be represented as I see $\\langle \\mathrm{GRD} \\rangle$ two dogs $\\langle \\mathrm{/GRD} \\rangle$ on $\\langle \\mathrm{GRD} \\rangle$ the beach $\\langle \\mathrm{/GRD} \\rangle$ , with two distinct visual entities grounded. The representation of each groundable phrase, termed as the grounding query, is obtained by adding $\\langle \\mathrm{GRD} \\rangle$ and $\\langle \\mathrm{/GRD} \\rangle$ 's output embedding from the last transformer layer of the MLLM. The representation is then used to retrieve the entities that the phrase should be grounded to. In particular, we concatenate the grounding query with the last layer output of each visual entity token, and use an MLP to predict a scalar score for each entity. Finally, we merge all the mask proposals into one single mask with pixel-wise maximization:", + "bbox": [ + 496, + 436, + 890, + 738 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} _ {h, w} = \\max _ {q} \\left(\\mathcal {S} _ {q} \\cdot \\widehat {\\mathcal {M}} _ {q, h, w}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 747, + 792, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $S_{q}$ is the normalized score of the $q$ -th mask ranging from 0 to 1, and $\\widehat{\\mathcal{M}}_{q,h,w}$ denotes the pixel probability at position $(h,w)$ for the $q$ -th mask. Note that a phrase may ground to multiple entities, thus multiple mask proposals may get a high score simultaneously and be selected in conjunction. One of the primary benefits of this decoupled design is its transparency in the selection of entities. Users can easily visualize both the mask proposals and their respective", + "bbox": [ + 496, + 777, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14229", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "scores, providing a clear understanding of how a grounding mask is predicted. This level of clarity and interpretability is a significant advantage, offering users a tangible insight into the model's grounding process.", + "bbox": [ + 75, + 90, + 470, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3. Towards Holistic Entity Mask Proposals", + "text_level": 1, + "bbox": [ + 76, + 165, + 421, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In order to support holistic language grounding to arbitrary segmentations, the entity proposal should have two essential properties. First, the proposals should strike a delicate balance in terms of semantic atomicity. While it is possible to merge multiple proposals later to form multi-entity segmentations, the reverse, i.e., dividing a single proposal into smaller segments, is not feasible. Therefore, instance segmentation is generally preferred over semantic segmentation. However, the segmentation should not be excessively fine-grained to the extent that it compromises basic semantic integrity. Over-segmentation can lead to a loss of the coherent concept of an entity, which is detrimental to the grounding process. Second, the entity proposals should have a high coverage of entities, encompassing a diverse range of granularities. This includes not only tangible objects (things) and amorphous concepts (stuff) but also extends to sub-components of objects (parts of things) and structured regions such as areas containing visual text. The ability to propose entities across this spectrum of granularity is pivotal, as it directly determines the upper bound of the grounding capability of MLLM.", + "bbox": [ + 75, + 188, + 468, + 505 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We initiated our study with a Mask2Former model pretrained on the COCO panoptic segmentation dataset. However, preliminary experiments revealed its limitations in semantic coverage and adaptability to open-world scenarios. To enhance this, we developed Mask2Former+, an upgraded version designed for multi-grained segmentation. This upgrade involved creating a diverse dataset by merging annotations from various sources, including COCO [5], LVIS [25], Entity-v2 [60], Pascal [16], PACO [63] (Figure 7); MHP-v2 [40] for human part parsing; and TextOCR [67] for text segmentation. We expanded the model's capabilities by adding 50 expert queries each for semantic parts and visual text regions, alongside the original 200 entity queries. We assessed Mask2Former+’s performance on 1000 images from validation splits from 4 grounding benchmarks, RefCOCO+ [86], PhraseCut [76], ReasonSeg [37], and TextVQA-X [66]. We use the Any-IoU [30] metric for evaluation, i.e., for each ground truth mask, we extract the most overlapped mask proposals and compute the IoU, then take the average. As Table 1 demonstrates, Mask2Former+ shows consistent improvements across all domains, particularly in those significantly divergent from COCO. This highlights its enhanced adaptability and precision in a broader range of segmentation challenges, providing a good mask proposal model for GROUNDHOG. We refer to Appendix A for more details of the model and data.", + "bbox": [ + 75, + 507, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/a5cb2f90c4666ea955fcb8d1d0e48086351920939ed6aa074b34b395429ad0da.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelRefCOCO+PhraseCutReasonSegTextVQA-X
Mask2Former0.8670.5630.6020.137
Mask2Former+0.8730.6240.7450.446
", + "bbox": [ + 504, + 88, + 890, + 133 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/e694e0b942828c5896ff70b6b9ec998b90cfa069724e8a877289d4f42f8ea706.jpg", + "table_caption": [ + "Table 1. The average Any-IoU of the proposals on each dataset. The vanilla Mask2Former is trained on the COCO-Panoptic dataset and our Mask2Former+ is trained on our combined dataset. Mask2Former+ obtains a consistent improvement in all scenarios, especially in non-COCO domains." + ], + "table_footnote": [], + "table_body": "
TaskDatasetGr. Ann.Sem. Gran.# Pairs
MBPoSThPaGTx
GCAPPNG✓✓✓✓✓132k
Flickr30K-Entity✓✓✓✓✓149k
RESRefCOCO✓✓✓113k
RefCOCO+✓✓✓112k
RefCOCOg✓✓✓80k
RefCLEF✓✓✓105k
gRefCOCO✓✓✓194k
PhraseCut✓✓✓✓✓✓85k
D-Cube✓✓✓✓10k
ReasonSeg✓✓✓✓✓✓1k
RIO✓✓✓✓28k
SK-VG✓✓23k
GVQAVizWiz-G✓✓✓✓✓6k
TextVQA-X✓✓✓15k
GQA✓✓✓✓✓302k
VQS✓✓20k
Shikra-BinaryQA✓✓✓✓✓4k
EntityCount✓✓✓✓✓✓11k
FoodSeg-QA✓✓✓✓7k
LVIS-QA✓✓✓✓✓95k
RDRefCOCO-REG✓✓✓✓17k
RefCOCO+-REG✓✓✓✓17k
RefCOCOg-REG✓✓✓✓22k
gRefCOCO-REG✓✓✓✓20k
VG-SpotCap✓✓✓✓✓✓247k
V7W✓✓✓23k
PointQA✓✓64k
VCR✓✓✓156k
ShikraRD✓✓✓✓✓✓2k
SVIT-RD✓✓✓✓✓✓33k
Guesswhat✓✓✓✓193k
VG-RefMatch✓✓✓✓✓✓247k
HierText✓✓✓✓6k
M3G2 (Total)2.5M
", + "bbox": [ + 506, + 214, + 888, + 623 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 2. Summary of datasets included in M3G2. We show the availability of Grounding Annotations (Box, Mask, and Pointer inputs), the Semantic Granularity (Stuff, Things, Parts, Groups, and Text), and the number of text-image pairs for training.", + "bbox": [ + 498, + 628, + 890, + 685 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Our Dataset: M3G2", + "text_level": 1, + "bbox": [ + 500, + 699, + 692, + 714 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we introduce M3G2, a Multi-Modal MultiGrained Grounding dataset consisting of 2.5M text-image pairs for visually grounded instruction tuning, consisting of 36 sub-problems derived and augmented from 27 existing datasets. We re-organize and augment public datasets of language grounding, visual question answering, referring expression segmentation, and referring expression generation into various forms of visually grounded dialogue for grounded instruction tuning, outlined briefly in Table 2. The dataset is categorized into four main types: (1) Grounded Image Captioning (GIC), (2) Referential Expression Segmentation (RES), (3) Grounded Visual Question Answer", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14230", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ing (GVQA), and (4) Referential Dialog (RD). We provide illustrated descriptions of our prompt design, accompanied by examples of each task type as depicted in Figure 4. We detail the task schema in the following sections and provide the complete sets of templates in Appendix B.", + "bbox": [ + 75, + 90, + 472, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1. Grounded Image Captioning (GIC)", + "text_level": 1, + "bbox": [ + 76, + 176, + 385, + 194 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The task of grounded image captioning requires the model to produce a narrative for the visual scene, and accurately identify and associate the groundable phrases with their respective binary segmentation masks. The objective of this task is to empower the model to articulate the scene while acknowledging various visual elements and their spatial interrelations. We incorporate the Panoptic Narrative Grounding (PNG) dataset [34] for dense and detailed scene descriptions, as well as the Flickr30K-Entity dataset [58] for concise descriptions of the salient contents in the image. We create a collection of task prompt templates that instruct the model to describe the image either in detail or briefly.", + "bbox": [ + 75, + 200, + 472, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Referring Expression Segmentation (RES)", + "text_level": 1, + "bbox": [ + 76, + 391, + 437, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In contrast to previous tasks, the referring expression segmentation task requires that the model generates a segmentation mask based on a given referring expression. Besides the RefCOCO series [43, 52, 86], we have further leveraged existing RES benchmarks [37, 65, 76, 78, 79] for this purpose. As shown in Figure 4, we instruct the model to localize the referred objects, with the expected output being a repetition of the referring expression whose grounding mask is the target object.", + "bbox": [ + 75, + 412, + 468, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Grounded Visual Question Answering (GVQA)", + "text_level": 1, + "bbox": [ + 76, + 559, + 468, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The task of grounded visual question answering requires the model to comprehend a question (with optional pointers) and to produce an answer that is grounded to a binary segmentation mask that justifies the answer. The goal of this task is to enable natural QA-based interaction with users and reasoning in the model with grounded explanations. Specifically, we harvest and adapt a collection of public VQA datasets with grounding annotations [6, 8, 19, 29], QA benchmarks on visual text [64], and create populated QA pairs from segmentation datasets [25, 60]. Our prompt templates instruct the model to respond either as open-ended answers or by selecting from multiple choices, with the response anchored to a segmentation mask (Figure 4).", + "bbox": [ + 75, + 583, + 470, + 781 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Referential Dialogue (RD)", + "text_level": 1, + "bbox": [ + 76, + 787, + 313, + 804 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The task of referential dialogue requires the model to conduct dialogue communication with users, especially when conditioned on user-provided spatial prompts. This includes existing RD datasets [8, 51, 88, 90, 94], multi-turn augmentations from segmentation datasets [17, 36, 47] as well as the referring expression generation (REG) task the", + "bbox": [ + 75, + 809, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f4fd885ed5f3526f41a055fa0bb8365a1ae69932f147bc99f70830caea9e9194.jpg", + "image_caption": [ + "Figure 4. The M3G2 dataset for grounded visual instruction tuning. M3G2 is a diverse dataset of multiple granularities, unifying 4 different task types with visually grounded dialogue." + ], + "image_footnote": [], + "bbox": [ + 503, + 90, + 888, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "RefCOCO series [43, 52, 86]. The REG task differs from the region captioning task in that it demands the description to be a referring expression that distinctly identifies the targeted object. Effective REG calls for the model to engage in dialogue interactions cooperatively, adhering to the Gricean Maxims [24] which dictate that communication should be as informative, truthful, relevant, and clear as necessary.", + "bbox": [ + 498, + 393, + 892, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment and Analysis", + "text_level": 1, + "bbox": [ + 500, + 513, + 736, + 531 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We adopt the LLaMA2-7B model [70] as our base LLM, and initialized the weight from LLaVA-1.5 [44]. For the vision encoders, we use the OpenAI CLIP@336 [62] model and DINOv2-L/14-reg [15] pretrained checkpoints. Please refer to Appendix C for more implementation details.", + "bbox": [ + 498, + 537, + 890, + 614 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Generalist in Grounded Vision-Language Tasks We first demonstrate GROUNDHOG's capabilities as a generalist model for three different types of grounded vision-language tasks. It's worth noting that, unlike previous work that needs dataset-specific fine-tuning on each of the tasks, GROUNDHOG can achieve comparable performance on all the tasks directly after training on M3G2, i.e., all the reported results from our model are from a single set of weights without any dataset-specific fine-tuning.", + "bbox": [ + 498, + 619, + 892, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Language Grounding To Segmentation. We start by evaluating the model on language grounding tasks, which takes text as input and generates segmentation masks as output. We assess GROUNDHOG on Referential Expression Segmentation (RES) [32] and Caption Phrase Grounding (CPG) tasks. While traditional RES benchmarks [32, 53] focus on single-instance referents requiring primarily visual understanding, we expanded our evaluation to include complex scenarios involving multi-instance or negative queries", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14231", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/412f68b1e5796260afa5028b891a73e88f793be5f2625d8d66d3028bfebf654c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelSingle InstanceMulti-/No InstanceReasoning
RefCOCORefCOCO+RefCOCOggRefCOCOPhraseCutReasonSegRIO
valtest-Atest-Bvaltest-Atest-Bval-utest-uvaltestvaltest-ctest-u
Specialist
MDETR [30]---------53.7-44.122.0
CRIS [75]70.573.266.162.368.153.759.960.455.3----
LAVT [82]72.775.868.862.168.455.161.262.158.4----
ReLA [43]73.876.570.266.071.057.765.066.063.6-22.4--
PolyFormer [46]76.078.373.369.374.661.969.270.2---48.826.8
UNINEXT-H [80]82.283.481.372.576.466.274.776.4-----
Generalist
LISA7B [37]74.176.571.162.467.456.566.468.5--44.0--
LISA7B (FT) [37]74.979.172.365.170.858.167.970.6--52.9--
7B78.579.975.770.575.064.974.174.666.754.556.257.933.9
", + "bbox": [ + 80, + 88, + 893, + 291 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5497b0b2f4d7c3dd396a279776935c989546f8c00039e53af1560da0e48f25f1.jpg", + "table_caption": [ + "Table 3. Results on 7 Referring Expression Segmentation (RES) benchmarks with single instance queries [32, 53], multi-/null instance queries [43, 76] and reasoning-based queries [37, 61]. We report cIoU for RefCOCO+/g and mIoU for other benchmarks, respectively." + ], + "table_footnote": [], + "table_body": "
ModelFlickr30K-E
R@1_valR@1_test
Shikra13B77.478.4
Ferret13B81.184.8
Shikra7B75.876.5
Ferret7B80.482.2
7B79.279.8
", + "bbox": [ + 80, + 327, + 321, + 463 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8f5cf507453a4539b127a442c4df6477465458d7394383c93cb78c5d4bb9be52.jpg", + "table_caption": [ + "Table 4. Top-1 box recall results on Flickr30K-Entity [58]." + ], + "table_footnote": [], + "table_body": "
ModelAR\\( \\mathbf{A}\\mathbf{R}_{\\mathrm {th}} \\)PNG \\( \\mathbf{A}\\mathbf{R}_{\\mathrm {st}} \\)\\( \\mathbf{A}\\mathbf{R}_{\\mathrm {s}} \\)\\( \\mathbf{A}\\mathbf{R}_{\\mathrm {p}} \\)
PiGLET65.964.068.667.254.5
\\( 7_{\\text{B}} \\)66.865.069.470.457.7
Table 5. Phrase grounding results on PNG [21].
ModelTextVQA-X [mIoU]
SAB29.0
\\( 7_{\\text{B}} \\)39.8
", + "bbox": [ + 326, + 327, + 609, + 463 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/99328dbfa3addd0fa7016d284ebc8d617c8fd7fbe5a58c6a445a7e2950e8d0db.jpg", + "table_caption": [ + "Table 6. Visual text QA results on the TextVQA-X [66] validation set." + ], + "table_footnote": [], + "table_body": "
ModelPointQA Twice AccV7W Acc
Shikra13B70.385.3
GPT4RoI13B-84.8
Shikra7B--
GPT4RoI7B-81.8
7B72.485.5
", + "bbox": [ + 616, + 327, + 893, + 463 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "[43, 76], and those necessitating common sense reasoning [37, 61]. The results in Table 3 show GROUNDHOG outperforming the generalist model LISA across all benchmarks and achieving significant improvements over specialist models in multi-instance, null, and reasoning-based RES tasks. It also performs comparably on the competitive RefCOCO series. For CPG tasks, which involve grounding all phrases in a caption and demand a deep understanding of the context for coreference resolution, we first evaluated GROUNDHOG on the Flickr30K-Entity dataset [58]. Since this dataset only has box annotations, we convert the mask predictions of our model to box and compute the top-1 box recall following the merged-box protocol [30]. Despite not specializing in predicting boxes, GROUNDHOG still outperforms Shikra 7B/13B [8] and is on par with Ferret-7B [85] in a concurrent work (Table 4). Additionally, on the PNG dataset [34] which tests phrase grounding in longer narratives, GROUNDHOG surpasses the previous state-of-the-art model, PiGLET [22], in all metrics including average recall of grounding masks and detailed scores for things, stuffs, and singular and plural entities (Table 5).", + "bbox": [ + 75, + 500, + 472, + 819 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Grounded Language Generation. Our model excels in generating language that accurately grounds to segmentation masks during user conversations. Quantitatively, we assess grounded captioning on the Flickr30K-Entity dataset [58], employing standard text generation metrics", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/bdde4fefe2e41c3207281a342ebfd0005f924ec393d29047c40940a8472fa04c.jpg", + "table_caption": [ + "Table 7. Results on PointQATwice [51] and V7W [94] test sets." + ], + "table_footnote": [], + "table_body": "
ModelBleu-4METEORCIDErSPICEF1all
Shikra13B--73.9--
Ferret13B37.025.576.118.315.1
Ferret7B35.124.674.818.015.0
7B36.726.591.320.432.1
", + "bbox": [ + 504, + 498, + 890, + 588 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 8. Grounded Captioning on Flickr30K-Entity [58].", + "bbox": [ + 524, + 588, + 862, + 601 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "such as Bleu-4 [55], METEOR [4], CIDEr [72], and SPICE [2] for language quality; and the $\\mathrm{F1}_{\\mathrm{all}}$ score for grounding accuracy following You et al. [85]. As shown in Table 8, GROUNDHOG significantly surpasses existing box-based grounded MLLMs, even their 13B versions, in both language quality and grounding accuracy. This improvement is hypothesized to stem from the diverse task distribution in our M3G2 dataset. For groundable question answering, we evaluate on the TextVQA-X benchmark [64]. GROUNDHOG outperforms the state-of-the-art specialist model SAB [33] by a significant margin, as measured by the mean IoU of the predicted mask (Table 6).", + "bbox": [ + 496, + 609, + 890, + 790 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Spatial Prompt Understanding. For grounded MLLMs, accurately interpreting multimodal instructions is essential, particularly in interactive tasks. We evaluated its performance on two pointer-based QA benchmarks, PointerQATwice [51] and V7W [94], which require the model to answer questions guided by spatial prompts, such as bounding boxes. The model is tasked to generate free-", + "bbox": [ + 496, + 794, + 892, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14232", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dfe920c57005f4d9d6b538244aeaf7653366145867b38e7e3bbdce9d5d0e890f.jpg", + "image_caption": [ + "Figure 5. Region caption using the best match proposal from Mask2Former+ versus from SAM. Mask2Former+ fails to propose the exact mask of the spire, leading to a less precise caption." + ], + "image_footnote": [], + "bbox": [ + 86, + 108, + 207, + 229 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f44c2995b738f414c001eb4996302b0a3f3e927bcbdbbb4870b7e119721bd49c.jpg", + "image_caption": [ + "Input Prompt M2F+ Best Match SAM" + ], + "image_footnote": [], + "bbox": [ + 210, + 108, + 331, + 229 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b503064ea84dd4572d44b2a4073a2f5a60f5c664633c3b3f77788d798bdf0886.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 108, + 455, + 229 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a99b56cdcb3679d9a612a7a78dfc3529ef504a9550cc9202fe2643846d33ed4b.jpg", + "image_caption": [ + "Figure 6. Illustration of a partially correct grounding. The grounding phrase and the ground truth mask are shown on the left. The top-4 mask proposals are presented, with highly-scored masks (green) selected for the merged mask, and low-scored masks (red) excluded. This illustrates the failure to recognize the word \"KWIK\" by the MLLM, despite its successful proposal." + ], + "image_footnote": [], + "bbox": [ + 83, + 277, + 459, + 412 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "form textual answers in PointerQA $_{\\text{Twice}}$ , and selects from multiple-choice options in V7W. GROUNDHOG demonstrates superior performance in these benchmarks, outperforming previous models as shown in Table 7. This highlights its effectiveness in spatial understanding and response accuracy. To further demonstrate the effectiveness of using SAM for the pointer-to-mask conversion, we show the best-matched mask proposal from our Mask2Former+ model in comparison to the mask from SAM in Figure 5. While the best match proposal from the Mask2Former+ model includes a broader area, the SAM-generated mask offers a more precise representation of the specified region, potentially leading to a more accurate caption.", + "bbox": [ + 75, + 503, + 467, + 699 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Trustworthiness and Transparency", + "text_level": 1, + "bbox": [ + 76, + 705, + 380, + 720 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Beyond its superior performance as a grounding generalist, we highlight two key improvements for creating a more trustworthy and transparent agent.", + "bbox": [ + 76, + 729, + 467, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Reduced Object Hallucination. Thanks to the varied task distribution and the inclusion of negative question-answering samples in M3G2 dataset, GROUNDHOG significantly reduces object hallucination. We assessed this using the POPE [42] benchmark, which includes binary questions about object existence across three splits, each with a different object distribution (with an order of difficulty Random $<$ Popular $<$ Adversarial). Remarkably, GROUND", + "bbox": [ + 76, + 780, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c4a82f23b38194e3150f7ce7b3c8b5ed7e415055a324d7d03f62e3e122ca7450.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelAccuracyPrecisionRecallF1 ScoreYes (%)
Random
mPLUG-Owl53.3051.7199.5368.0696.23
LLaVA54.4352.3299.8068.6595.37
MultiModal-GPT50.0350.02100.0066.6899.97
MiniGPT-477.8375.3882.6778.8654.83
InstructBLIP88.7385.0893.9389.2955.20
Shikra-13B86.9094.4079.2686.1943.26
Ferret-13B90.2497.7283.0089.7643.26
7B91.0385.8096.4090.7945.88
Popular
mPLUG-Owl50.6350.3299.2766.7998.63
LLaVA52.4351.2599.8067.7297.37
MultiModal-GPT50.0050.00100.0066.67100.00
MiniGPT-468.3064.2782.4072.2164.10
InstructBLIP81.3775.0793.9383.4562.57
Shikra-13B83.9787.5579.2083.1645.23
Ferret-13B84.9088.2480.5384.2145.63
7B90.1385.9393.8189.7045.80
Adversarial
mPLUG-Owl50.6750.3499.3366.8298.67
LLaVA50.7750.3999.8766.9899.10
MultiModal-GPT50.0050.00100.0066.67100.00
MiniGPT-466.6062.4583.2771.3766.67
InstructBLIP74.3767.6793.3378.4568.97
Shikra-13B83.1085.6079.6082.4946.50
Ferret-13B82.3683.6080.5382.0048.18
7B86.3385.9386.6386.2849.60
", + "bbox": [ + 506, + 89, + 888, + 431 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 9. Object hallucination results on the POPE [42] benchmark.", + "bbox": [ + 500, + 435, + 888, + 449 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "HOG consistently outperforms other models in both accuracy and F1 score across all splits, particularly on the more challenging ones. It shows an absolute improvement of $5.2\\%$ in accuracy for Popular and $4.0\\%$ for Adversarial over the previously best-performing model. This suggests that our model's enhanced grounding capability plays a significant role in mitigating the object hallucination problem.", + "bbox": [ + 498, + 460, + 890, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Explainability and Diagnosability. Another important highlight of GROUNDHOG is its enhancement of explainability through the decoupled design of entity proposal and selection, as outlined earlier in section 2.2. This is exemplified in the case study illustrated in Figure 6, which illustrates the mask proposal scoring and selective merging process of our model. We show the top-4 masks, where the higher-score masks are labeled in green while the lower-score masks are labeled in red. Users can easily interpret that the failure is due to the incapability of MLLM to recognize the word \"KWIK\", despite it being successfully localized and proposed as an entity candidate.", + "bbox": [ + 496, + 577, + 890, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 770, + 663, + 784 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We performed ablation studies to validate our design decisions, training, and evaluating a subset of the M3G2 dataset that includes RefCOCO+, Flickr30K, and TextVQA. These cover a range of visual entities from various image sources and granularities. We start by comparing our Mask2Former+ with the original Mask2Former for mask proposal effectiveness. As indicated in Table 10, the orig-", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14233", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/dee71b92735563a8da406e3d73573ae39ab3562c439978d9ec1ea25ac6e234d2.jpg", + "table_caption": [], + "table_footnote": [ + "Table 10. Ablation study on model design choices and evaluation setups. Models are trained on RefCOCO+, Flickr30K, TextVQA-X and tested on corresponding validation sets." + ], + "table_body": "
SetupRefCOCO+Flickr30KTextVQA-X
Mask Proposal Models
Mask2Former67.169.09.8
Mask2Former+66.677.234.0
Entity Features
CLIP59.875.032.0
DINOv262.376.328.4
CLIP+DINOv266.677.234.0
Grounding Query
<GRD> only64.467.534.2
</GRD> only64.477.233.5
Sum66.677.234.0
Eval Input Resolution
224–48054.767.227.6
480–64065.576.727.6
800–102466.677.234.0
", + "bbox": [ + 80, + 88, + 470, + 311 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "inal Mask2Former performs slightly better on RefCOCO, as it is developed specificity on COCO object categories. However, Mask2Former+ significantly surpasses the original in domains with non-COCO entities. Our second set of experiments examined the choice of visual entity features. Although using either CLIP or DINOv2 features alone shows advantages in specific datasets, their combination consistently yields the best results across all datasets. To obtain a robust grounding query representation, we experimented with using the output embedding of the token, the token, and their sum. We found that the latter approach achieves the best overall results. Finally, we demonstrate that our decoupling design of the mask proposal model and MLLM allows for training at a lower resolution (320px) to expedite grounding training, while scaling up the resolution during evaluation enhances performance.", + "bbox": [ + 75, + 366, + 468, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Related Work", + "text_level": 1, + "bbox": [ + 76, + 618, + 218, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multimodal Large Language Models. Building on the recent advance of large language models (LLMs), there is an increasing effort in adapting pretrained large language models for multimodal tasks, such as understanding and interpreting visual information [1, 71]. More recently, visual instruction tuning has gained much interest due to its surprising performance with a modest amount of data and computing resources. Various models have been developed, noticeably MiniGPT4[93], LLaVA [44, 45] and concurrent models [13, 20, 38, 74, 83]. Despite their promising performances, MLLMs often produce objects that are not presented in the given images, a phenomenon referred to as the object hallucination problem [14, 31, 42].", + "bbox": [ + 75, + 637, + 468, + 834 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "MLLM with Language Grounding The ability to connect language to their corresponding visual elements in the physical world, known as grounding [27], is crucial in everyday human communication about our shared sur", + "bbox": [ + 75, + 839, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "roundings. Grounding datasets have been shown to benefit vision-language pre-training, both in terms of object-level recognition [41] and language learning [50]. Recent works unify text and grounding regions into token sequences [49, 73, 81] in casual language modeling. Based on such paradigm, researchers have developed a family of grounded MLLM, including GPT4ROI [89], Kosmos-2 [56], Shikra [8], PVIT [7], BuboGPT [91], Qwen-VL [3], and Ferret [85]. Despite their promising performance, these models focus on object grounding to bounding boxes, which cannot handle pixel-level grounding across various semantic granularities. Furthermore, it lacks the diagnosability and explainability in failure cases. Our model fills this gap.", + "bbox": [ + 496, + 90, + 890, + 287 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Language-Guided Semantic Localization The field of language-guided semantic localization has a long history in the vision-language research community, requiring that the model localize a given referring expression with bounding boxes or segmentation masks. This task has evolved from early attempts to understand simple referring expressions within images, such as the well-known RefCOCO series [52, 86] and their generalized variant [43] that takes no-target and multi-target into account. The integration of advanced language reasoning from LLMs has enabled research to tackle even more nuanced reasoning tasks that involve complex language contexts [37, 57, 87]. Notably, LISA [37] formulates a reasoning segmentation task to bring language-informed reasoning into semantic segmentation, and contributes a powerful baseline. Our model builds on these developments, but is designed to be more universally applicable as a grounded MLLM.", + "bbox": [ + 496, + 294, + 892, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 556, + 617, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this study, we introduce GROUNDHOG, a novel framework designed to enable pixel-level explainable grounding in large language models, leveraging holistic segmentation. The system builds upon a pre-trained mask proposal network to provide pixel-level visual features for the large language models, allowing them to retrieve segmentation mask proposals that can be used for grounding. We also present M3G2, a dataset of 1.9M training text-image pairs with 36 sub-problems derived from 27 existing datasets for visually grounded instruction tuning, facilitating precise vision-language alignment at the pixel level. We show that after training on M3G2, GROUNDHOG achieves superior performance on various grounding tasks. Through extensive case studies, we further show that GROUNDHOG unlocks explainability and diagnosability, and demonstrates better grounding towards occluded objects, groups of multiple instances, amorphous background regions, semantic parts of objects, and objects with irregular shapes.", + "bbox": [ + 496, + 575, + 890, + 848 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements This work was supported by Amazon and NSF IIS-1949634. We thank the anonymous reviewers for their valuable comments and suggestions.", + "bbox": [ + 496, + 854, + 890, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14234", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 8", + "[2] Peter Anderson, Basura Fernando, Mark Johnson, and Stephen Gould. Spice: Semantic propositional image caption evaluation. In Proceedings of the 14th European Conference on Computer Vision, pages 382-398, 2016. 6", + "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 8", + "[4] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 6", + "[5] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocosstuff: Thing and stuff classes in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1209-1218, 2018. 2, 4, 1", + "[6] Chongyan Chen, Samreen Anjum, and Danna Gurari. Grounding answers for visual questions asked by visually impaired people. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19098-19107, 2022. 5, 2", + "[7] Chi Chen, Ruoyu Qin, Fuwen Luo, Xiaoyue Mi, Peng Li, Maosong Sun, and Yang Liu. Position-enhanced visual instruction tuning for multimodal large language models. arXiv preprint arXiv:2308.13437, 2023. 8", + "[8] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 2, 3, 5, 6, 8", + "[9] Ting Chen, Saurabh Saxena, Lala Li, David J Fleet, and Geoffrey Hinton. Pix2seq: A language modeling framework for object detection. In International Conference on Learning Representations, 2021. 3", + "[10] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 2, 1, 4", + "[11] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ * chatgpt quality. https://vicuna.lmsys.org, 2023. 4", + "[12] Jifeng Dai, Kaiming He, and Jian Sun. Convolutional feature masking for joint object and stuff segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3992-4000, 2015. 2", + "[13] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale" + ], + "bbox": [ + 78, + 114, + 470, + 902 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500, 2023. 2, 8", + "[14] Wenliang Dai, Zihan Liu, Ziwei Ji, Dan Su, and Pascale Fung. Plausible may not be faithful: Probing object hallucination in vision-language pre-training. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 2128-2140, 2023. 8", + "[15] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. arXiv preprint arXiv:2309.16588, 2023. 5, 4", + "[16] Daan de Geus, Panagiotis Meletis, Chenyang Lu, Xiaoxiao Wen, and Gijs Dubbelman. Part-aware panoptic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5485-5494, 2021. 4, 1", + "[17] Harm De Vries, Florian Strub, Sarath Chandar, Olivier Pietquin, Hugo Larochelle, and Aaron Courville. Guess-what?! visual object discovery through multi-modal dialogue. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5503–5512, 2017. 5, 3", + "[18] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 1", + "[19] Chuang Gan, Yandong Li, Haoxiang Li, Chen Sun, and Boqing Gong. Vqs: Linking segmentations to questions and answers for supervised attention in vqa and question-focused semantic segmentation. In Proceedings of the IEEE international conference on computer vision, pages 1811-1820, 2017. 5, 2", + "[20] Tao Gong, Chengqi Lyu, Shilong Zhang, Yudong Wang, Miao Zheng, Qian Zhao, Kuikun Liu, Wenwei Zhang, Ping Luo, and Kai Chen. Multimodal-gpt: A vision and language model for dialogue with humans. arXiv preprint arXiv:2305.04790, 2023. 8", + "[21] Cristina González, Nicolás Ayobi, Isabela Hernández, José Hernández, Jordi Pont-Tuset, and Pablo Arbeláez. Panoptic narrative grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1364-1373, 2021. 6", + "[22] Cristina González, Nicolás Ayobi, Isabela Hernández, Jordi Pont-Tuset, and Pablo Arbeláez. Piglet: Pixel-level grounding of language expressions with transformers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 6", + "[23] Abel Gonzalez-Garcia, Davide Modolo, and Vittorio Ferrari. Objects as context for detecting their semantic parts. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6907-6916, 2018. 2", + "[24] Herbert P Grice. Logic and conversation. In Speech acts, pages 41-58. Brill, 1975. 5", + "[25] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 4, 5, 1, 3" + ], + "bbox": [ + 503, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14235", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Kai Han, Yunhe Wang, Jianyuan Guo, Yehui Tang, and Enhua Wu. Vision gnn: An image is worth graph of nodes. Advances in Neural Information Processing Systems, 35:8291-8303, 2022. 2", + "[27] Stevan Harnad. The symbol grounding problem. Physica D: Nonlinear Phenomena, 42(1-3):335-346, 1990. 8", + "[28] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2021. 4", + "[29] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 5, 2", + "[30] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1780-1790, 2021. 4, 6", + "[31] Osman Semih Kayhan, Bart Vredebregt, and Jan C Van Gemert. Hallucination in object detection—a study in visual part verification. In 2021 IEEE International Conference on Image Processing (ICIP), pages 2234–2238. IEEE, 2021. 8", + "[32] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014. 5, 6, 2", + "[33] Seyedalireza Khoshkirat and Chandra Kambhametty. Sentence attention blocks for answer grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6080-6090, 2023. 6", + "[34] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dólar. Panoptic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9404-9413, 2019. 5, 6, 1", + "[35] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4015-4026, 2023. 2, 3", + "[36] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 5, 3", + "[37] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692, 2023. 4, 5, 6, 8, 2", + "[38] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726, 2023. 8" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Chunyuan Li, Zhe Gan, Zhengyuan Yang, Jianwei Yang, Linjie Li, Lijuan Wang, and Jianfeng Gao. Multimodal foundation models: From specialists to general-purpose assistants. arXiv preprint arXiv:2309.10020, 2023. 2", + "[40] Jianshu Li, Jian Zhao, Yunchao Wei, Congyan Lang, Yidong Li, Terence Sim, Shuicheng Yan, and Jiashi Feng. Multiple-human parsing in the wild. arXiv preprint arXiv:1705.07206, 2017. 4, 1", + "[41] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022. 8", + "[42] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, 2023. 7, 8", + "[43] Chang Liu, Henghui Ding, and Xudong Jiang. Gres: Generalized referring expression segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23592-23601, 2023. 5, 6, 8, 2, 3", + "[44] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 5, 8, 4", + "[45] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 2, 8", + "[46] Jiang Liu, Hui Ding, Zhaowei Cai, Yuting Zhang, Ravi Kumar Satzoda, Vijay Mahadevan, and R Manmatha. Polyformer: Referring image segmentation as sequential polygon generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18653-18663, 2023. 6", + "[47] Shangbang Long, Siyang Qin, Dmitry Panteleev, Alessandro Bissacco, Yasuhisa Fujii, and Michalis Raptis. Towards end-to-end unified scene text detection and layout analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1049-1059, 2022. 5, 3", + "[48] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 4", + "[49] Jiasen Lu, Christopher Clark, Rowan Zellers, Roozbeh Mottaghi, and Aniruddha Kembhavi. Unified-io: A unified model for vision, language, and multi-modal tasks. In The Eleventh International Conference on Learning Representations, 2022. 8", + "[50] Ziqiao Ma, Jiayi Pan, and Joyce Chai. World-to-words: Grounded open vocabulary acquisition through fast mapping in vision-language models. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 524–544, 2023. 2, 8", + "[51] Arjun Mani, Nobline Yoo, Will Hinthorn, and Olga Russakovsky. Point and ask: Incorporating pointing into visual question answering. arXiv preprint arXiv:2011.13681, 2020. 5, 6, 3" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14236", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 5, 8, 3", + "[53] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 5, 6, 2", + "[54] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2", + "[55] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 6", + "[56] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 2, 3, 8", + "[57] Renjie Pi, Jiahui Gao, Shizhe Diao, Rui Pan, Hanze Dong, Jipeng Zhang, Lewei Yao, Jianhua Han, Hang Xu, and Lingpeng Kong Tong Zhang. Detgpt: Detect what you need via reasoning. arXiv preprint arXiv:2305.14167, 2023. 8", + "[58] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In Proceedings of the IEEE international conference on computer vision, pages 2641–2649, 2015. 5, 6, 1", + "[59] Jordi Pont-Tuset, Jasper Uijlings, Soravit Changpinyo, Radu Soricut, and Vittorio Ferrari. Connecting vision and language with localized narratives. In Proceedings of the 16th European Conference on Computer Vision, pages 647-664, 2020. 1", + "[60] Lu Qi, Jason Kuen, Tiancheng Shen, Jiuxiang Gu, Wenbo Li, Weidong Guo, Jiaya Jia, Zhe Lin, and Ming-Hsuan Yang. High quality entity segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4047-4056, 2023. 4, 5, 1, 3", + "[61] Mengxue Qu, Yu Wu, Wu Liu, Xiaodan Liang, Jingkuan Song, Yao Zhao, and Yunchao Wei. RIO: A benchmark for reasoning intention-oriented objects in open environments. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 6, 2", + "[62] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 5, 4", + "[63] Vignesh Ramanathan, Anmol Kalia, Vladan Petrovic, Yi Wen, Baixue Zheng, Baishan Guo, Rui Wang, Aaron Mar" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "quez, Rama Kovvuri, Abhishek Kadian, et al. Paco: Parts and attributes of common objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7141-7151, 2023. 4, 1", + "[64] Varun Nagaraj Rao, Xingjian Zhen, Karen Hovsepian, and Mingwei Shen. A first look: Towards explainable textvqa models via visual and textual explanations. arXiv preprint arXiv:2105.02626, 2021. 5, 6, 2", + "[65] Anna Rohrbach, Marcus Rohrbach, Ronghang Hu, Trevor Darrell, and Bernt Schiele. Grounding of textual phrases in images by reconstruction. In Proceedings of the 14th European Conference on Computer Vision, pages 817-834. Springer, 2016. 5, 2", + "[66] Amanpreet Singh, Vivek Natarjan, Meet Shah, Yu Jiang, Xinlei Chen, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8317-8326, 2019. 4, 6", + "[67] Amanpreet Singh, Guan Pang, Mandy Toh, Jing Huang, Wojciech Galuba, and Tal Hassner. TextOCR: Towards large-scale end-to-end reasoning for arbitrary-shaped scene text. 2021. 4, 1", + "[68] Bharat Singh, Hengduo Li, Abhishek Sharma, and Larry S Davis. R-fcn-3000 at 30fps: Decoupling detection and classification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1081–1090, 2018. 2", + "[69] Zhi Tian, Chunhua Shen, Xinlong Wang, and Hao Chen. Boxinst: High-performance instance segmentation with box annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5443-5452, 2021. 4", + "[70] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 5, 4", + "[71] Maria Tsimpoukelli, Jacob L Menick, Serkan Cabi, SM Eslami, Oriol Vinyls, and Felix Hill. Multimodal few-shot learning with frozen language models. Advances in Neural Information Processing Systems, 34:200-212, 2021. 8", + "[72] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 6", + "[73] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In International Conference on Machine Learning, pages 23318-23340. PMLR, 2022. 8", + "[74] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, et al. Vision: Large language model is also an open-ended decoder for vision-centric tasks. arXiv preprint arXiv:2305.11175, 2023. 8", + "[75] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clip" + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "14237", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "driven referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11686-11695, 2022. 6", + "[76] Chenyun Wu, Zhe Lin, Scott Cohen, Trung Bui, and Subhransu Maji. Phrasescut: Language-based image segmentation in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10216-10225, 2020. 4, 5, 6, 2", + "[77] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.1", + "[78] Yixuan Wu, Zhao Zhang, Chi Xie, Feng Zhu, and Rui Zhao. Advancing referring expression segmentation beyond single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2628-2638, 2023. 5, 2", + "[79] Chi Xie, Zhao Zhang, Yixuan Wu, Feng Zhu, Rui Zhao, and Shuang Liang. Described object detection: Liberating object detection with flexible expressions. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 5, 2", + "[80] Bin Yan, Yi Jiang, Jiannan Wu, Dong Wang, Ping Luo, Zehuan Yuan, and Huchuan Lu. Universal instance perception as object discovery and retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15325-15336, 2023. 6", + "[81] Zhengyuan Yang, Zhe Gan, Jianfeng Wang, Xiaowei Hu, Faisal Ahmed, Zicheng Liu, Yumao Lu, and Lijuan Wang. Unitab: Unifying text and box outputs for grounded vision-language modeling. 2022. 8", + "[82] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Levt: Language-aware vision transformer for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18155-18165, 2022. 6", + "[83] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. MPLug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 8", + "[84] Shukang Yin, Chaoyou Fu, Sirui Zhao, Ke Li, Xing Sun, Tong Xu, and Enhong Chen. A survey on multimodal large language models. arXiv preprint arXiv:2306.13549, 2023. 2", + "[85] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704, 2023. 3, 6, 8", + "[86] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. 4, 5, 8, 3", + "[87] Yuhang Zang, Wei Li, Jun Han, Kaiyang Zhou, and Chen Change Loy. Contextual object detection with multimodal large language models. arXiv preprint arXiv:2305.18279, 2023. 8" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[88] Rowan Zellers, Yonatan Bisk, Ali Farhadi, and Yejin Choi. From recognition to cognition: Visual commonsense reasoning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6720-6731, 2019. 5, 3", + "[89] Shilong Zhang, Peize Sun, Shoufa Chen, Min Xiao, Wenqi Shao, Wenwei Zhang, Kai Chen, and Ping Luo. Gpt4roi: Instruction tuning large language model on region-of-interest. arXiv preprint arXiv:2307.03601, 2023. 8", + "[90] Bo Zhao, Boya Wu, and Tiejun Huang. Svit: Scaling up visual instruction tuning. arXiv preprint arXiv:2307.04087, 2023.5,3", + "[91] Yang Zhao, Zhijie Lin, Daquan Zhou, Zilong Huang, Jiashi Feng, and Bingyi Kang. Bubogpt: Enabling visual grounding in multi-modal llms. arXiv preprint arXiv:2307.08581, 2023. 8", + "[92] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16793-16803, 2022. 2", + "[93] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 8", + "[94] Yuke Zhu, Oliver Groth, Michael Bernstein, and Li Fei-Fei. Visual7w: Grounded question answering in images. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4995–5004, 2016. 5, 6, 3" + ], + "bbox": [ + 501, + 92, + 890, + 513 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "14238", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2024/GROUNDHOG_ Grounding Large Language Models to Holistic Segmentation/bdfcf2de-41e8-4d27-af7d-65821e6b79f8_model.json b/2024/GROUNDHOG_ Grounding Large Language Models to Holistic Segmentation/bdfcf2de-41e8-4d27-af7d-65821e6b79f8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..019f949757bdc3c7b5aac148529347f2a7d0146a --- /dev/null +++ b/2024/GROUNDHOG_ Grounding Large Language Models to Holistic Segmentation/bdfcf2de-41e8-4d27-af7d-65821e6b79f8_model.json @@ -0,0 +1,2776 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.02 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.326, + 0.017, + 0.723, + 0.031 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.294, + 0.032, + 0.754, + 0.047 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.085, + 0.116, + 0.885, + 0.141 + ], + "angle": 0, + "content": "GROUNDHOG: Grounding Large Language Models to Holistic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.136, + 0.155, + 0.833, + 0.174 + ], + "angle": 0, + "content": "Yichi Zhang\\(^{1\\dagger}\\), Ziqiao Ma\\(^{1\\dagger}\\), Xiaofeng Gao\\(^{2}\\), Suhaila Shakiah\\(^{2}\\), Qiaozi Gao\\(^{2}\\), Joyce Chai\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.329, + 0.174, + 0.644, + 0.191 + ], + "angle": 0, + "content": "1University of Michigan, 2Amazon AGI" + }, + { + "type": "text", + "bbox": [ + 0.378, + 0.193, + 0.593, + 0.209 + ], + "angle": 0, + "content": "zhangyic@umich.edu" + }, + { + "type": "text", + "bbox": [ + 0.288, + 0.211, + 0.678, + 0.227 + ], + "angle": 0, + "content": "https://groundhog-ml1m.github.io/" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.236, + 0.284, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.355, + 0.105, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.359, + 0.274, + 0.372 + ], + "angle": 0, + "content": "Describe the image briefly." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.381, + 0.232, + 0.405 + ], + "angle": 0, + "content": "A man and a little girl are sitting in a shopping cart." + }, + { + "type": "image", + "bbox": [ + 0.259, + 0.387, + 0.285, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.237, + 0.486, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.355, + 0.308, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.329, + 0.359, + 0.468, + 0.372 + ], + "angle": 0, + "content": "Generate a short caption." + }, + { + "type": "text", + "bbox": [ + 0.293, + 0.38, + 0.438, + 0.405 + ], + "angle": 0, + "content": "Two dogs are playing with a stick in a field." + }, + { + "type": "image", + "bbox": [ + 0.486, + 0.237, + 0.688, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.692, + 0.27, + 0.714, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.731, + 0.264, + 0.851, + 0.288 + ], + "angle": 0, + "content": "Describe the given picture in very detail." + }, + { + "type": "text", + "bbox": [ + 0.696, + 0.31, + 0.845, + 0.391 + ], + "angle": 0, + "content": "In this image, we can see a boat on the water. There are few people inside the boat. There are also few people on the rocks. In the background, there are trees, hills, and sky." + }, + { + "type": "image", + "bbox": [ + 0.862, + 0.38, + 0.885, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.378, + 0.409, + 0.592, + 0.422 + ], + "angle": 0, + "content": "(a) Grounded Image Captioning (GIC)." + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.423, + 0.368, + 0.582 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.584, + 0.105, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.586, + 0.349, + 0.609 + ], + "angle": 0, + "content": "Could you please segment out \"laptop barely onscreen not apple\" in the image?" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.618, + 0.279, + 0.63 + ], + "angle": 0, + "content": "Laptop barely onscreen not apple." + }, + { + "type": "image", + "bbox": [ + 0.344, + 0.614, + 0.367, + 0.632 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.423, + 0.611, + 0.579 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.586, + 0.425, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.438, + 0.586, + 0.585, + 0.608 + ], + "angle": 0, + "content": "What date is shown on the calendar?" + }, + { + "type": "text", + "bbox": [ + 0.405, + 0.618, + 0.44, + 0.63 + ], + "angle": 0, + "content": "Feb 2." + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.615, + 0.615, + 0.633 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.644, + 0.423, + 0.886, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.664, + 0.585, + 0.685, + 0.612 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.693, + 0.585, + 0.874, + 0.608 + ], + "angle": 0, + "content": "Provide a distinct description for that \\(< \\text{PTR}>\\)" + }, + { + "type": "text", + "bbox": [ + 0.668, + 0.618, + 0.817, + 0.631 + ], + "angle": 0, + "content": "Blue container with apples." + }, + { + "type": "image", + "bbox": [ + 0.861, + 0.614, + 0.885, + 0.632 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.681, + 0.636, + 0.848, + 0.648 + ], + "angle": 0, + "content": "(d) Referential Dialogue (RD)." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.65, + 0.892, + 0.678 + ], + "angle": 0, + "content": "Figure 1. We propose GROUNDHOG, a multimodal large language model that enhances its text output with pixel-level phrase grounding across diverse semantic granularities. The figure demonstrates outputs from our model on the four task types we considered in this work." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.687, + 0.313, + 0.702 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.707, + 0.471, + 0.873 + ], + "angle": 0, + "content": "Most multimodal large language models (MLLMs) learn language-to-object grounding through causal language modeling where grounded objects are captured by bounding boxes as sequences of location tokens. This paradigm lacks pixel-level representations that are important for fine-grained visual understanding and diagnosis. In this work, we introduce GROUNDHOG, an MLLM developed by grounding Large Language Models to holistic segmentation. GROUNDHOG incorporates a masked feature extractor and converts extracted features into visual entity tokens for the MLLM backbone, which then con" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.87 + ], + "angle": 0, + "content": "nects groundable phrases to unified grounding masks by retrieving and merging the entity masks. To train GROUNDHOG, we carefully curated M3G2, a grounded visual instruction tuning dataset with Multi-Modal Multi-Grained Grounding, by harvesting a collection of segmentation-grounded datasets with rich annotations. Our experimental results show that GROUNDHOG achieves superior performance on various language grounding tasks without task-specific fine-tuning, and significantly reduces object hallucination. GROUNDHOG also demonstrates better grounding towards complex forms of visual input and provides easy-to-understand diagnosis in failure cases." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.348, + 0.9 + ], + "angle": 0, + "content": "† Work done during internship at Amazon AGI." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14227" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.081, + 0.091, + 0.208, + 0.105 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.116, + 0.468, + 0.401 + ], + "angle": 0, + "content": "Multimodal large language models (MLLMs) have received an increasing amount of attention to address tasks that necessitate non-linguistic knowledge, e.g., perception and reasoning about the visual world [39, 84]. For fine-grained visual understanding, grounded MLLMs often learn language-to-object grounding by causal language modeling, where grounded objects are captured by bounding boxes as sequences of location tokens. However, bounding boxes are insufficient in indicating amorphous stuff [5], semantic parts of objects [23], finer-grained regions with irregular shapes [26], or groups of instances at the same time. As a result, a single bounding box can often include other irrelevant semantics in order to engulf the target entities, leading to ambiguity in detection. In addition, the generated box coordinate lacks interpretability. When the model hallucinates, such as incorrectly predicting the association between objects and language, it is hard to diagnose whether the problem is due to the model's failure to detect the object, or its incorrect alignment of the object with language." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.403, + 0.468, + 0.748 + ], + "angle": 0, + "content": "To address these issues, in this work, we introduce GROUNDHOG, an MLLM developed by grounding Large Language Models to holistic segmentation. Our goal of language grounding is to connect text spans that refer to or can be deduced from visual information, termed as groundable phrases [50], to their corresponding regions of visual entities (Figure 1). GROUNDHOG incorporates a masked feature extractor that takes an input image and a set of class-agnostic entity mask proposals, and converts each mask's features into visual entity tokens for an MLLM backbone. This MLLM then connects groundable phrases to unified grounding masks by retrieving and merging the entity masks. Compared to previous grounded MLLMs, GROUNDHOG unlocks unprecedented pixel-level vision-language alignment. It naturally supports visual pointers as input, and can plug-in-and-play with any choice of mask proposal networks, e.g., Segment Anything Model (SAM) [35], domain-specific semantic segmentation models, or user-provided mask candidates. We introduce an enhanced Mask2Former [10] as our default mask proposal network, which detects regions at multiple granularities, e.g., instances (things and stuff), semantic parts, and visual text, leading to a holistic coverage of visual semantics." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.75, + 0.468, + 0.9 + ], + "angle": 0, + "content": "To train GROUNDHOG, we curated a Multi-Modal Multi-Grained Grounding (M3G2) dataset consisting of 2.5M text-image pairs for visually grounded instruction tuning, consisting of 36 sub-problems derived and augmented from 27 existing datasets. We present extensive experiments on vision-language tasks that require grounding, including grounded language generation with minimal object hallucination, language-guided segmentation, visual question answering with answer grounding, and referential dialog with spatial pointer inputs (Figure 1). Our empirical results show" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.093, + 0.89, + 0.167 + ], + "angle": 0, + "content": "that GROUNDHOG, without task-specific fine-tuning, can achieve superior or comparable performance with previous models that either require fine-tuning or are specialized only for that dataset. In addition, GROUNDHOG has supports easy-to-understand diagnosis when grounding fails." + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.181, + 0.757, + 0.195 + ], + "angle": 0, + "content": "2. Our Method: GROUNDHOG" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.206, + 0.89, + 0.627 + ], + "angle": 0, + "content": "The language grounding task can be succinctly delineated into two fundamental components: localization and recognition, as established in the literature [50, 68, 92]. Such categorization not only aids in the identification of object presence (objectness) without reliance on specific object classes, but also sets the stage for models to be robust in open-vocabulary settings. Building upon this framework, we formulate the grounding process as an entity segment selection problem, which involves (1) proposing entity segmentation masks where the masks encapsulate regions with discernible semantic content, and (2) recognizing the retrieved entities through the understanding of both visual and language context. Concurrently performing both tasks is where MLLMs bring a distinct advantage. This decoupled design of entity mask proposal and language-guided grounding brings several advantages. First, it allows independent improvement of the mask proposal model and MLLM, where specialized data, training, and inference setups can be applied. Second, by decoupling language grounding, it becomes straightforward to determine if a failure is due to the model's inability to propose the entity segment, or its misalignment of the object with the language, thus improving the interpretability of the whole framework. Third, as shown later, when connecting the two parts to work in tandem in a model-independent manner, the MLLM can benefit from multiple different vision specialist models in a plug-and-play fashion. In the remainder of this section, we give details of our model design." + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.637, + 0.819, + 0.651 + ], + "angle": 0, + "content": "2.1. Building Entity Features from Masks" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.66, + 0.89, + 0.9 + ], + "angle": 0, + "content": "Our approach assumes the availability of a mask proposal model, which is capable of generating a set of class-agnostic entity masks from an image with high coverage. In contrast to prior studies that relied on low-level features [8, 13, 45, 56], GROUNDHOG interprets the image as a collection of entities. The primary challenge then becomes the derivation of effective visual features to accurately represent these entities. To achieve a complete decoupling of the MLLM from the mask proposal model responsible for providing the masks, we propose to condition the entity features solely on the binary masks without using any embeddings from the mask proposal model. Specifically, the mask corresponding to each entity is employed to extract patch features from pretrained vision foundation models, such as CLIP [62] and DINOv2 [54], through a convolutional mask pooling layer [12]. Given that the feature map dimensions" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.947, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14228" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.092, + 0.885, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.347, + 0.895, + 0.404 + ], + "angle": 0, + "content": "Figure 2. The model architecture of GROUNDHOG model. Given a set of class-agnostic entity mask proposals, the masked feature extractor first extracts the feature of each entity as the visual input of the multi-modal large language model (left). The output hidden states of the grounding tokens are averaged and used to retrieve the entities to ground, which will be merged into a single grounding mask for the phrase. Modules are colored by their trainability: parameter-free operators (grey), frozen (blue), trainable (orange), and partially trainable (mix)." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.414, + 0.469, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.54, + 0.471, + 0.596 + ], + "angle": 0, + "content": "Figure 3. GROUNDHOG can take arbitrary spatial prompts that can be resolved by an interactive segmentation model, such as SAM. The placeholder pointer token \\(<\\mathrm{PTR}>\\) will be replaced by the extracted entity features and fed as input to the model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.605, + 0.47, + 0.726 + ], + "angle": 0, + "content": "are usually smaller than those of the mask proposals, we resize the masks to match the size of the feature maps prior to pooling. The pooled features are then fed into a Multi-Layer Perceptron (MLP) network to align with the input embeddings of the MLLM. We empirically find the combination of CLIP and DINOv2 features yields the best result, and these features are added to obtain the final input visual entity tokens to the MLLM." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Spatial Prompts Furthermore, for grounded MLLMs to be more broadly applicable, they must be capable of interpreting multi-modal user inputs, including spatial prompts. Thanks to the mask model agnostic design, GROUND-HOGcan seamlessly support such inputs. As demonstrated in Figure 3, by applying an interactive segmentation model such as Segment-Anything (SAM) [35], arbitrary spatial prompts can be translated into binary masks and processed by the same masked feature extractor we just introduced. This extracted feature for the pointed entity will replace the pointer token placeholder in the textual input." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.414, + 0.882, + 0.43 + ], + "angle": 0, + "content": "2.2. Language Grounding to Entity Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.437, + 0.892, + 0.739 + ], + "angle": 0, + "content": "Existing box-grounded MLLMs typically append location tokens after the groundable phrases [8, 9, 56, 85]. However, this method is not readily interpretable. To alleviate this disconnect, we introduce a pair of grounding tokens \\( \\langle \\mathrm{GRD} \\rangle \\) and \\( \\langle \\mathrm{/GRD} \\rangle \\) to indicate the start and end of groundable phrases, with the assumption that grounding these phrases requires mapping to certain representations of visual entities irrespective of the visual modality. In Figure 2, a sentence can be represented as I see \\( \\langle \\mathrm{GRD} \\rangle \\) two dogs \\( \\langle \\mathrm{/GRD} \\rangle \\) on \\( \\langle \\mathrm{GRD} \\rangle \\) the beach \\( \\langle \\mathrm{/GRD} \\rangle \\), with two distinct visual entities grounded. The representation of each groundable phrase, termed as the grounding query, is obtained by adding \\( \\langle \\mathrm{GRD} \\rangle \\) and \\( \\langle \\mathrm{/GRD} \\rangle \\)'s output embedding from the last transformer layer of the MLLM. The representation is then used to retrieve the entities that the phrase should be grounded to. In particular, we concatenate the grounding query with the last layer output of each visual entity token, and use an MLP to predict a scalar score for each entity. Finally, we merge all the mask proposals into one single mask with pixel-wise maximization:" + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.748, + 0.794, + 0.775 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} _ {h, w} = \\max _ {q} \\left(\\mathcal {S} _ {q} \\cdot \\widehat {\\mathcal {M}} _ {q, h, w}\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.778, + 0.892, + 0.903 + ], + "angle": 0, + "content": "where \\( S_{q} \\) is the normalized score of the \\( q \\)-th mask ranging from 0 to 1, and \\( \\widehat{\\mathcal{M}}_{q,h,w} \\) denotes the pixel probability at position \\( (h,w) \\) for the \\( q \\)-th mask. Note that a phrase may ground to multiple entities, thus multiple mask proposals may get a high score simultaneously and be selected in conjunction. One of the primary benefits of this decoupled design is its transparency in the selection of entities. Users can easily visualize both the mask proposals and their respective" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14229" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.152 + ], + "angle": 0, + "content": "scores, providing a clear understanding of how a grounding mask is predicted. This level of clarity and interpretability is a significant advantage, offering users a tangible insight into the model's grounding process." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.166, + 0.422, + 0.182 + ], + "angle": 0, + "content": "2.3. Towards Holistic Entity Mask Proposals" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.189, + 0.47, + 0.506 + ], + "angle": 0, + "content": "In order to support holistic language grounding to arbitrary segmentations, the entity proposal should have two essential properties. First, the proposals should strike a delicate balance in terms of semantic atomicity. While it is possible to merge multiple proposals later to form multi-entity segmentations, the reverse, i.e., dividing a single proposal into smaller segments, is not feasible. Therefore, instance segmentation is generally preferred over semantic segmentation. However, the segmentation should not be excessively fine-grained to the extent that it compromises basic semantic integrity. Over-segmentation can lead to a loss of the coherent concept of an entity, which is detrimental to the grounding process. Second, the entity proposals should have a high coverage of entities, encompassing a diverse range of granularities. This includes not only tangible objects (things) and amorphous concepts (stuff) but also extends to sub-components of objects (parts of things) and structured regions such as areas containing visual text. The ability to propose entities across this spectrum of granularity is pivotal, as it directly determines the upper bound of the grounding capability of MLLM." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We initiated our study with a Mask2Former model pretrained on the COCO panoptic segmentation dataset. However, preliminary experiments revealed its limitations in semantic coverage and adaptability to open-world scenarios. To enhance this, we developed Mask2Former+, an upgraded version designed for multi-grained segmentation. This upgrade involved creating a diverse dataset by merging annotations from various sources, including COCO [5], LVIS [25], Entity-v2 [60], Pascal [16], PACO [63] (Figure 7); MHP-v2 [40] for human part parsing; and TextOCR [67] for text segmentation. We expanded the model's capabilities by adding 50 expert queries each for semantic parts and visual text regions, alongside the original 200 entity queries. We assessed Mask2Former+’s performance on 1000 images from validation splits from 4 grounding benchmarks, RefCOCO+ [86], PhraseCut [76], ReasonSeg [37], and TextVQA-X [66]. We use the Any-IoU [30] metric for evaluation, i.e., for each ground truth mask, we extract the most overlapped mask proposals and compute the IoU, then take the average. As Table 1 demonstrates, Mask2Former+ shows consistent improvements across all domains, particularly in those significantly divergent from COCO. This highlights its enhanced adaptability and precision in a broader range of segmentation challenges, providing a good mask proposal model for GROUNDHOG. We refer to Appendix A for more details of the model and data." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.089, + 0.891, + 0.135 + ], + "angle": 0, + "content": "
ModelRefCOCO+PhraseCutReasonSegTextVQA-X
Mask2Former0.8670.5630.6020.137
Mask2Former+0.8730.6240.7450.446
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.141, + 0.892, + 0.21 + ], + "angle": 0, + "content": "Table 1. The average Any-IoU of the proposals on each dataset. The vanilla Mask2Former is trained on the COCO-Panoptic dataset and our Mask2Former+ is trained on our combined dataset. Mask2Former+ obtains a consistent improvement in all scenarios, especially in non-COCO domains." + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.215, + 0.89, + 0.624 + ], + "angle": 0, + "content": "
TaskDatasetGr. Ann.Sem. Gran.# Pairs
MBPoSThPaGTx
GCAPPNG✓✓✓✓✓132k
Flickr30K-Entity✓✓✓✓✓149k
RESRefCOCO✓✓✓113k
RefCOCO+✓✓✓112k
RefCOCOg✓✓✓80k
RefCLEF✓✓✓105k
gRefCOCO✓✓✓194k
PhraseCut✓✓✓✓✓✓85k
D-Cube✓✓✓✓10k
ReasonSeg✓✓✓✓✓✓1k
RIO✓✓✓✓28k
SK-VG✓✓23k
GVQAVizWiz-G✓✓✓✓✓6k
TextVQA-X✓✓✓15k
GQA✓✓✓✓✓302k
VQS✓✓20k
Shikra-BinaryQA✓✓✓✓✓4k
EntityCount✓✓✓✓✓✓11k
FoodSeg-QA✓✓✓✓7k
LVIS-QA✓✓✓✓✓95k
RDRefCOCO-REG✓✓✓✓17k
RefCOCO+-REG✓✓✓✓17k
RefCOCOg-REG✓✓✓✓22k
gRefCOCO-REG✓✓✓✓20k
VG-SpotCap✓✓✓✓✓✓247k
V7W✓✓✓23k
PointQA✓✓64k
VCR✓✓✓156k
ShikraRD✓✓✓✓✓✓2k
SVIT-RD✓✓✓✓✓✓33k
Guesswhat✓✓✓✓193k
VG-RefMatch✓✓✓✓✓✓247k
HierText✓✓✓✓6k
M3G2 (Total)2.5M
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.629, + 0.892, + 0.686 + ], + "angle": 0, + "content": "Table 2. Summary of datasets included in M3G2. We show the availability of Grounding Annotations (Box, Mask, and Pointer inputs), the Semantic Granularity (Stuff, Things, Parts, Groups, and Text), and the number of text-image pairs for training." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.7, + 0.694, + 0.715 + ], + "angle": 0, + "content": "3. Our Dataset: M3G2" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this section, we introduce M3G2, a Multi-Modal MultiGrained Grounding dataset consisting of 2.5M text-image pairs for visually grounded instruction tuning, consisting of 36 sub-problems derived and augmented from 27 existing datasets. We re-organize and augment public datasets of language grounding, visual question answering, referring expression segmentation, and referring expression generation into various forms of visually grounded dialogue for grounded instruction tuning, outlined briefly in Table 2. The dataset is categorized into four main types: (1) Grounded Image Captioning (GIC), (2) Referential Expression Segmentation (RES), (3) Grounded Visual Question Answer" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14230" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.168 + ], + "angle": 0, + "content": "ing (GVQA), and (4) Referential Dialog (RD). We provide illustrated descriptions of our prompt design, accompanied by examples of each task type as depicted in Figure 4. We detail the task schema in the following sections and provide the complete sets of templates in Appendix B." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.178, + 0.387, + 0.195 + ], + "angle": 0, + "content": "3.1. Grounded Image Captioning (GIC)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.202, + 0.473, + 0.385 + ], + "angle": 0, + "content": "The task of grounded image captioning requires the model to produce a narrative for the visual scene, and accurately identify and associate the groundable phrases with their respective binary segmentation masks. The objective of this task is to empower the model to articulate the scene while acknowledging various visual elements and their spatial interrelations. We incorporate the Panoptic Narrative Grounding (PNG) dataset [34] for dense and detailed scene descriptions, as well as the Flickr30K-Entity dataset [58] for concise descriptions of the salient contents in the image. We create a collection of task prompt templates that instruct the model to describe the image either in detail or briefly." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.392, + 0.439, + 0.409 + ], + "angle": 0, + "content": "3.2. Referring Expression Segmentation (RES)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.413, + 0.47, + 0.549 + ], + "angle": 0, + "content": "In contrast to previous tasks, the referring expression segmentation task requires that the model generates a segmentation mask based on a given referring expression. Besides the RefCOCO series [43, 52, 86], we have further leveraged existing RES benchmarks [37, 65, 76, 78, 79] for this purpose. As shown in Figure 4, we instruct the model to localize the referred objects, with the expected output being a repetition of the referring expression whose grounding mask is the target object." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.56, + 0.47, + 0.577 + ], + "angle": 0, + "content": "3.3. Grounded Visual Question Answering (GVQA)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.471, + 0.782 + ], + "angle": 0, + "content": "The task of grounded visual question answering requires the model to comprehend a question (with optional pointers) and to produce an answer that is grounded to a binary segmentation mask that justifies the answer. The goal of this task is to enable natural QA-based interaction with users and reasoning in the model with grounded explanations. Specifically, we harvest and adapt a collection of public VQA datasets with grounding annotations [6, 8, 19, 29], QA benchmarks on visual text [64], and create populated QA pairs from segmentation datasets [25, 60]. Our prompt templates instruct the model to respond either as open-ended answers or by selecting from multiple choices, with the response anchored to a segmentation mask (Figure 4)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.789, + 0.315, + 0.805 + ], + "angle": 0, + "content": "3.4. Referential Dialogue (RD)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.901 + ], + "angle": 0, + "content": "The task of referential dialogue requires the model to conduct dialogue communication with users, especially when conditioned on user-provided spatial prompts. This includes existing RD datasets [8, 51, 88, 90, 94], multi-turn augmentations from segmentation datasets [17, 36, 47] as well as the referring expression generation (REG) task the" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.091, + 0.89, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.345, + 0.893, + 0.387 + ], + "angle": 0, + "content": "Figure 4. The M3G2 dataset for grounded visual instruction tuning. M3G2 is a diverse dataset of multiple granularities, unifying 4 different task types with visually grounded dialogue." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.394, + 0.893, + 0.501 + ], + "angle": 0, + "content": "RefCOCO series [43, 52, 86]. The REG task differs from the region captioning task in that it demands the description to be a referring expression that distinctly identifies the targeted object. Effective REG calls for the model to engage in dialogue interactions cooperatively, adhering to the Gricean Maxims [24] which dictate that communication should be as informative, truthful, relevant, and clear as necessary." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.514, + 0.738, + 0.532 + ], + "angle": 0, + "content": "4. Experiment and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.539, + 0.892, + 0.615 + ], + "angle": 0, + "content": "We adopt the LLaMA2-7B model [70] as our base LLM, and initialized the weight from LLaVA-1.5 [44]. For the vision encoders, we use the OpenAI CLIP@336 [62] model and DINOv2-L/14-reg [15] pretrained checkpoints. Please refer to Appendix C for more implementation details." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.621, + 0.893, + 0.76 + ], + "angle": 0, + "content": "4.1. Generalist in Grounded Vision-Language Tasks We first demonstrate GROUNDHOG's capabilities as a generalist model for three different types of grounded vision-language tasks. It's worth noting that, unlike previous work that needs dataset-specific fine-tuning on each of the tasks, GROUNDHOG can achieve comparable performance on all the tasks directly after training on M3G2, i.e., all the reported results from our model are from a single set of weights without any dataset-specific fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Language Grounding To Segmentation. We start by evaluating the model on language grounding tasks, which takes text as input and generates segmentation masks as output. We assess GROUNDHOG on Referential Expression Segmentation (RES) [32] and Caption Phrase Grounding (CPG) tasks. While traditional RES benchmarks [32, 53] focus on single-instance referents requiring primarily visual understanding, we expanded our evaluation to include complex scenarios involving multi-instance or negative queries" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14231" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.089, + 0.894, + 0.292 + ], + "angle": 0, + "content": "
ModelSingle InstanceMulti-/No InstanceReasoning
RefCOCORefCOCO+RefCOCOggRefCOCOPhraseCutReasonSegRIO
valtest-Atest-Bvaltest-Atest-Bval-utest-uvaltestvaltest-ctest-u
Specialist
MDETR [30]---------53.7-44.122.0
CRIS [75]70.573.266.162.368.153.759.960.455.3----
LAVT [82]72.775.868.862.168.455.161.262.158.4----
ReLA [43]73.876.570.266.071.057.765.066.063.6-22.4--
PolyFormer [46]76.078.373.369.374.661.969.270.2---48.826.8
UNINEXT-H [80]82.283.481.372.576.466.274.776.4-----
Generalist
LISA7B [37]74.176.571.162.467.456.566.468.5--44.0--
LISA7B (FT) [37]74.979.172.365.170.858.167.970.6--52.9--
7B78.579.975.770.575.064.974.174.666.754.556.257.933.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.293, + 0.893, + 0.322 + ], + "angle": 0, + "content": "Table 3. Results on 7 Referring Expression Segmentation (RES) benchmarks with single instance queries [32, 53], multi-/null instance queries [43, 76] and reasoning-based queries [37, 61]. We report cIoU for RefCOCO+/g and mIoU for other benchmarks, respectively." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.328, + 0.323, + 0.464 + ], + "angle": 0, + "content": "
ModelFlickr30K-E
R@1_valR@1_test
Shikra13B77.478.4
Ferret13B81.184.8
Shikra7B75.876.5
Ferret7B80.482.2
7B79.279.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.467, + 0.315, + 0.496 + ], + "angle": 0, + "content": "Table 4. Top-1 box recall results on Flickr30K-Entity [58]." + }, + { + "type": "table", + "bbox": [ + 0.327, + 0.328, + 0.611, + 0.464 + ], + "angle": 0, + "content": "
ModelAR\\( \\mathbf{A}\\mathbf{R}_{\\mathrm {th}} \\)PNG \\( \\mathbf{A}\\mathbf{R}_{\\mathrm {st}} \\)\\( \\mathbf{A}\\mathbf{R}_{\\mathrm {s}} \\)\\( \\mathbf{A}\\mathbf{R}_{\\mathrm {p}} \\)
PiGLET65.964.068.667.254.5
\\( 7_{\\text{B}} \\)66.865.069.470.457.7
Table 5. Phrase grounding results on PNG [21].
ModelTextVQA-X [mIoU]
SAB29.0
\\( 7_{\\text{B}} \\)39.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.324, + 0.465, + 0.616, + 0.494 + ], + "angle": 0, + "content": "Table 6. Visual text QA results on the TextVQA-X [66] validation set." + }, + { + "type": "table", + "bbox": [ + 0.617, + 0.328, + 0.895, + 0.464 + ], + "angle": 0, + "content": "
ModelPointQA Twice AccV7W Acc
Shikra13B70.385.3
GPT4RoI13B-84.8
Shikra7B--
GPT4RoI7B-81.8
7B72.485.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.65, + 0.467, + 0.888, + 0.495 + ], + "angle": 0, + "content": "Table 7. Results on PointQATwice [51] and V7W [94] test sets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.501, + 0.473, + 0.82 + ], + "angle": 0, + "content": "[43, 76], and those necessitating common sense reasoning [37, 61]. The results in Table 3 show GROUNDHOG outperforming the generalist model LISA across all benchmarks and achieving significant improvements over specialist models in multi-instance, null, and reasoning-based RES tasks. It also performs comparably on the competitive RefCOCO series. For CPG tasks, which involve grounding all phrases in a caption and demand a deep understanding of the context for coreference resolution, we first evaluated GROUNDHOG on the Flickr30K-Entity dataset [58]. Since this dataset only has box annotations, we convert the mask predictions of our model to box and compute the top-1 box recall following the merged-box protocol [30]. Despite not specializing in predicting boxes, GROUNDHOG still outperforms Shikra 7B/13B [8] and is on par with Ferret-7B [85] in a concurrent work (Table 4). Additionally, on the PNG dataset [34] which tests phrase grounding in longer narratives, GROUNDHOG surpasses the previous state-of-the-art model, PiGLET [22], in all metrics including average recall of grounding masks and detailed scores for things, stuffs, and singular and plural entities (Table 5)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Grounded Language Generation. Our model excels in generating language that accurately grounds to segmentation masks during user conversations. Quantitatively, we assess grounded captioning on the Flickr30K-Entity dataset [58], employing standard text generation metrics" + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.499, + 0.892, + 0.589 + ], + "angle": 0, + "content": "
ModelBleu-4METEORCIDErSPICEF1all
Shikra13B--73.9--
Ferret13B37.025.576.118.315.1
Ferret7B35.124.674.818.015.0
7B36.726.591.320.432.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.525, + 0.589, + 0.863, + 0.602 + ], + "angle": 0, + "content": "Table 8. Grounded Captioning on Flickr30K-Entity [58]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.61, + 0.892, + 0.791 + ], + "angle": 0, + "content": "such as Bleu-4 [55], METEOR [4], CIDEr [72], and SPICE [2] for language quality; and the \\(\\mathrm{F1}_{\\mathrm{all}}\\) score for grounding accuracy following You et al. [85]. As shown in Table 8, GROUNDHOG significantly surpasses existing box-based grounded MLLMs, even their 13B versions, in both language quality and grounding accuracy. This improvement is hypothesized to stem from the diverse task distribution in our M3G2 dataset. For groundable question answering, we evaluate on the TextVQA-X benchmark [64]. GROUNDHOG outperforms the state-of-the-art specialist model SAB [33] by a significant margin, as measured by the mean IoU of the predicted mask (Table 6)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Spatial Prompt Understanding. For grounded MLLMs, accurately interpreting multimodal instructions is essential, particularly in interactive tasks. We evaluated its performance on two pointer-based QA benchmarks, PointerQATwice [51] and V7W [94], which require the model to answer questions guided by spatial prompts, such as bounding boxes. The model is tasked to generate free-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14232" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.093, + 0.414, + 0.107 + ], + "angle": 0, + "content": "Input Prompt M2F+ Best Match SAM" + }, + { + "type": "image", + "bbox": [ + 0.087, + 0.109, + 0.208, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.212, + 0.109, + 0.332, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.109, + 0.457, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.078, + 0.232, + 0.468, + 0.273 + ], + "angle": 0, + "content": "Figure 5. Region caption using the best match proposal from Mask2Former+ versus from SAM. Mask2Former+ fails to propose the exact mask of the spire, leading to a less precise caption." + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.278, + 0.46, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.414, + 0.468, + 0.496 + ], + "angle": 0, + "content": "Figure 6. Illustration of a partially correct grounding. The grounding phrase and the ground truth mask are shown on the left. The top-4 mask proposals are presented, with highly-scored masks (green) selected for the merged mask, and low-scored masks (red) excluded. This illustrates the failure to recognize the word \"KWIK\" by the MLLM, despite its successful proposal." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.504, + 0.468, + 0.7 + ], + "angle": 0, + "content": "form textual answers in PointerQA\\(_{\\text{Twice}}\\), and selects from multiple-choice options in V7W. GROUNDHOG demonstrates superior performance in these benchmarks, outperforming previous models as shown in Table 7. This highlights its effectiveness in spatial understanding and response accuracy. To further demonstrate the effectiveness of using SAM for the pointer-to-mask conversion, we show the best-matched mask proposal from our Mask2Former+ model in comparison to the mask from SAM in Figure 5. While the best match proposal from the Mask2Former+ model includes a broader area, the SAM-generated mask offers a more precise representation of the specified region, potentially leading to a more accurate caption." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.706, + 0.382, + 0.722 + ], + "angle": 0, + "content": "4.2. Trustworthiness and Transparency" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.73, + 0.468, + 0.774 + ], + "angle": 0, + "content": "Beyond its superior performance as a grounding generalist, we highlight two key improvements for creating a more trustworthy and transparent agent." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.781, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Reduced Object Hallucination. Thanks to the varied task distribution and the inclusion of negative question-answering samples in M3G2 dataset, GROUNDHOG significantly reduces object hallucination. We assessed this using the POPE [42] benchmark, which includes binary questions about object existence across three splits, each with a different object distribution (with an order of difficulty Random \\(<\\) Popular \\(<\\) Adversarial). Remarkably, GROUND" + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.09, + 0.89, + 0.432 + ], + "angle": 0, + "content": "
ModelAccuracyPrecisionRecallF1 ScoreYes (%)
Random
mPLUG-Owl53.3051.7199.5368.0696.23
LLaVA54.4352.3299.8068.6595.37
MultiModal-GPT50.0350.02100.0066.6899.97
MiniGPT-477.8375.3882.6778.8654.83
InstructBLIP88.7385.0893.9389.2955.20
Shikra-13B86.9094.4079.2686.1943.26
Ferret-13B90.2497.7283.0089.7643.26
7B91.0385.8096.4090.7945.88
Popular
mPLUG-Owl50.6350.3299.2766.7998.63
LLaVA52.4351.2599.8067.7297.37
MultiModal-GPT50.0050.00100.0066.67100.00
MiniGPT-468.3064.2782.4072.2164.10
InstructBLIP81.3775.0793.9383.4562.57
Shikra-13B83.9787.5579.2083.1645.23
Ferret-13B84.9088.2480.5384.2145.63
7B90.1385.9393.8189.7045.80
Adversarial
mPLUG-Owl50.6750.3499.3366.8298.67
LLaVA50.7750.3999.8766.9899.10
MultiModal-GPT50.0050.00100.0066.67100.00
MiniGPT-466.6062.4583.2771.3766.67
InstructBLIP74.3767.6793.3378.4568.97
Shikra-13B83.1085.6079.6082.4946.50
Ferret-13B82.3683.6080.5382.0048.18
7B86.3385.9386.6386.2849.60
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.436, + 0.89, + 0.45 + ], + "angle": 0, + "content": "Table 9. Object hallucination results on the POPE [42] benchmark." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.461, + 0.892, + 0.568 + ], + "angle": 0, + "content": "HOG consistently outperforms other models in both accuracy and F1 score across all splits, particularly on the more challenging ones. It shows an absolute improvement of \\(5.2\\%\\) in accuracy for Popular and \\(4.0\\%\\) for Adversarial over the previously best-performing model. This suggests that our model's enhanced grounding capability plays a significant role in mitigating the object hallucination problem." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.578, + 0.892, + 0.76 + ], + "angle": 0, + "content": "Explainability and Diagnosability. Another important highlight of GROUNDHOG is its enhancement of explainability through the decoupled design of entity proposal and selection, as outlined earlier in section 2.2. This is exemplified in the case study illustrated in Figure 6, which illustrates the mask proposal scoring and selective merging process of our model. We show the top-4 masks, where the higher-score masks are labeled in green while the lower-score masks are labeled in red. Users can easily interpret that the failure is due to the incapability of MLLM to recognize the word \"KWIK\", despite it being successfully localized and proposed as an entity candidate." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.771, + 0.664, + 0.785 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We performed ablation studies to validate our design decisions, training, and evaluating a subset of the M3G2 dataset that includes RefCOCO+, Flickr30K, and TextVQA. These cover a range of visual entities from various image sources and granularities. We start by comparing our Mask2Former+ with the original Mask2Former for mask proposal effectiveness. As indicated in Table 10, the orig-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14233" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.089, + 0.472, + 0.312 + ], + "angle": 0, + "content": "
SetupRefCOCO+Flickr30KTextVQA-X
Mask Proposal Models
Mask2Former67.169.09.8
Mask2Former+66.677.234.0
Entity Features
CLIP59.875.032.0
DINOv262.376.328.4
CLIP+DINOv266.677.234.0
Grounding Query
<GRD> only64.467.534.2
</GRD> only64.477.233.5
Sum66.677.234.0
Eval Input Resolution
224–48054.767.227.6
480–64065.576.727.6
800–102466.677.234.0
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.077, + 0.318, + 0.47, + 0.36 + ], + "angle": 0, + "content": "Table 10. Ablation study on model design choices and evaluation setups. Models are trained on RefCOCO+, Flickr30K, TextVQA-X and tested on corresponding validation sets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.367, + 0.47, + 0.61 + ], + "angle": 0, + "content": "inal Mask2Former performs slightly better on RefCOCO, as it is developed specificity on COCO object categories. However, Mask2Former+ significantly surpasses the original in domains with non-COCO entities. Our second set of experiments examined the choice of visual entity features. Although using either CLIP or DINOv2 features alone shows advantages in specific datasets, their combination consistently yields the best results across all datasets. To obtain a robust grounding query representation, we experimented with using the output embedding of the token, the token, and their sum. We found that the latter approach achieves the best overall results. Finally, we demonstrate that our decoupling design of the mask proposal model and MLLM allows for training at a lower resolution (320px) to expedite grounding training, while scaling up the resolution during evaluation enhances performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.619, + 0.22, + 0.635 + ], + "angle": 0, + "content": "5. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.638, + 0.47, + 0.835 + ], + "angle": 0, + "content": "Multimodal Large Language Models. Building on the recent advance of large language models (LLMs), there is an increasing effort in adapting pretrained large language models for multimodal tasks, such as understanding and interpreting visual information [1, 71]. More recently, visual instruction tuning has gained much interest due to its surprising performance with a modest amount of data and computing resources. Various models have been developed, noticeably MiniGPT4[93], LLaVA [44, 45] and concurrent models [13, 20, 38, 74, 83]. Despite their promising performances, MLLMs often produce objects that are not presented in the given images, a phenomenon referred to as the object hallucination problem [14, 31, 42]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.903 + ], + "angle": 0, + "content": "MLLM with Language Grounding The ability to connect language to their corresponding visual elements in the physical world, known as grounding [27], is crucial in everyday human communication about our shared sur" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.289 + ], + "angle": 0, + "content": "roundings. Grounding datasets have been shown to benefit vision-language pre-training, both in terms of object-level recognition [41] and language learning [50]. Recent works unify text and grounding regions into token sequences [49, 73, 81] in casual language modeling. Based on such paradigm, researchers have developed a family of grounded MLLM, including GPT4ROI [89], Kosmos-2 [56], Shikra [8], PVIT [7], BuboGPT [91], Qwen-VL [3], and Ferret [85]. Despite their promising performance, these models focus on object grounding to bounding boxes, which cannot handle pixel-level grounding across various semantic granularities. Furthermore, it lacks the diagnosability and explainability in failure cases. Our model fills this gap." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.295, + 0.893, + 0.552 + ], + "angle": 0, + "content": "Language-Guided Semantic Localization The field of language-guided semantic localization has a long history in the vision-language research community, requiring that the model localize a given referring expression with bounding boxes or segmentation masks. This task has evolved from early attempts to understand simple referring expressions within images, such as the well-known RefCOCO series [52, 86] and their generalized variant [43] that takes no-target and multi-target into account. The integration of advanced language reasoning from LLMs has enabled research to tackle even more nuanced reasoning tasks that involve complex language contexts [37, 57, 87]. Notably, LISA [37] formulates a reasoning segmentation task to bring language-informed reasoning into semantic segmentation, and contributes a powerful baseline. Our model builds on these developments, but is designed to be more universally applicable as a grounded MLLM." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.558, + 0.619, + 0.572 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.577, + 0.892, + 0.849 + ], + "angle": 0, + "content": "In this study, we introduce GROUNDHOG, a novel framework designed to enable pixel-level explainable grounding in large language models, leveraging holistic segmentation. The system builds upon a pre-trained mask proposal network to provide pixel-level visual features for the large language models, allowing them to retrieve segmentation mask proposals that can be used for grounding. We also present M3G2, a dataset of 1.9M training text-image pairs with 36 sub-problems derived from 27 existing datasets for visually grounded instruction tuning, facilitating precise vision-language alignment at the pixel level. We show that after training on M3G2, GROUNDHOG achieves superior performance on various grounding tasks. Through extensive case studies, we further show that GROUNDHOG unlocks explainability and diagnosability, and demonstrates better grounding towards occluded objects, groups of multiple instances, amorphous background regions, semantic parts of objects, and objects with irregular shapes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Acknowledgements This work was supported by Amazon and NSF IIS-1949634. We thank the anonymous reviewers for their valuable comments and suggestions." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14234" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.197 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.472, + 0.255 + ], + "angle": 0, + "content": "[2] Peter Anderson, Basura Fernando, Mark Johnson, and Stephen Gould. Spice: Semantic propositional image caption evaluation. In Proceedings of the 14th European Conference on Computer Vision, pages 382-398, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.256, + 0.471, + 0.311 + ], + "angle": 0, + "content": "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.312, + 0.471, + 0.38 + ], + "angle": 0, + "content": "[4] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.472, + 0.437 + ], + "angle": 0, + "content": "[5] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocosstuff: Thing and stuff classes in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1209-1218, 2018. 2, 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.471, + 0.507 + ], + "angle": 0, + "content": "[6] Chongyan Chen, Samreen Anjum, and Danna Gurari. Grounding answers for visual questions asked by visually impaired people. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19098-19107, 2022. 5, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.508, + 0.471, + 0.564 + ], + "angle": 0, + "content": "[7] Chi Chen, Ruoyu Qin, Fuwen Luo, Xiaoyue Mi, Peng Li, Maosong Sun, and Yang Liu. Position-enhanced visual instruction tuning for multimodal large language models. arXiv preprint arXiv:2308.13437, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.471, + 0.619 + ], + "angle": 0, + "content": "[8] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.621, + 0.471, + 0.675 + ], + "angle": 0, + "content": "[9] Ting Chen, Saurabh Saxena, Lala Li, David J Fleet, and Geoffrey Hinton. Pix2seq: A language modeling framework for object detection. In International Conference on Learning Representations, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.677, + 0.471, + 0.746 + ], + "angle": 0, + "content": "[10] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 2, 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.747, + 0.471, + 0.816 + ], + "angle": 0, + "content": "[11] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) * chatgpt quality. https://vicuna.lmsys.org, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.817, + 0.471, + 0.873 + ], + "angle": 0, + "content": "[12] Jifeng Dai, Kaiming He, and Jian Sun. Convolutional feature masking for joint object and stuff segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3992-4000, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.873, + 0.471, + 0.903 + ], + "angle": 0, + "content": "[13] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500, 2023. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.135, + 0.894, + 0.217 + ], + "angle": 0, + "content": "[14] Wenliang Dai, Zihan Liu, Ziwei Ji, Dan Su, and Pascale Fung. Plausible may not be faithful: Probing object hallucination in vision-language pre-training. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 2128-2140, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.218, + 0.892, + 0.26 + ], + "angle": 0, + "content": "[15] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. arXiv preprint arXiv:2309.16588, 2023. 5, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[16] Daan de Geus, Panagiotis Meletis, Chenyang Lu, Xiaoxiao Wen, and Gijs Dubbelman. Part-aware panoptic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5485-5494, 2021. 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.41 + ], + "angle": 0, + "content": "[17] Harm De Vries, Florian Strub, Sarath Chandar, Olivier Pietquin, Hugo Larochelle, and Aaron Courville. Guess-what?! visual object discovery through multi-modal dialogue. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5503–5512, 2017. 5, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.413, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[18] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.469, + 0.892, + 0.551 + ], + "angle": 0, + "content": "[19] Chuang Gan, Yandong Li, Haoxiang Li, Chen Sun, and Boqing Gong. Vqs: Linking segmentations to questions and answers for supervised attention in vqa and question-focused semantic segmentation. In Proceedings of the IEEE international conference on computer vision, pages 1811-1820, 2017. 5, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.553, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[20] Tao Gong, Chengqi Lyu, Shilong Zhang, Yudong Wang, Miao Zheng, Qian Zhao, Kuikun Liu, Wenwei Zhang, Ping Luo, and Kai Chen. Multimodal-gpt: A vision and language model for dialogue with humans. arXiv preprint arXiv:2305.04790, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.622, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[21] Cristina González, Nicolás Ayobi, Isabela Hernández, José Hernández, Jordi Pont-Tuset, and Pablo Arbeláez. Panoptic narrative grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1364-1373, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.892, + 0.76 + ], + "angle": 0, + "content": "[22] Cristina González, Nicolás Ayobi, Isabela Hernández, Jordi Pont-Tuset, and Pablo Arbeláez. Piglet: Pixel-level grounding of language expressions with transformers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.761, + 0.892, + 0.817 + ], + "angle": 0, + "content": "[23] Abel Gonzalez-Garcia, Davide Modolo, and Vittorio Ferrari. Objects as context for detecting their semantic parts. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6907-6916, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[24] Herbert P Grice. Logic and conversation. In Speech acts, pages 41-58. Brill, 1975. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[25] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 4, 5, 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14235" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[26] Kai Han, Yunhe Wang, Jianyuan Guo, Yehui Tang, and Enhua Wu. Vision gnn: An image is worth graph of nodes. Advances in Neural Information Processing Systems, 35:8291-8303, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.468, + 0.175 + ], + "angle": 0, + "content": "[27] Stevan Harnad. The symbol grounding problem. Physica D: Nonlinear Phenomena, 42(1-3):335-346, 1990. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.177, + 0.469, + 0.232 + ], + "angle": 0, + "content": "[28] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.233, + 0.469, + 0.3 + ], + "angle": 0, + "content": "[29] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 5, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.302, + 0.469, + 0.371 + ], + "angle": 0, + "content": "[30] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1780-1790, 2021. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.372, + 0.469, + 0.439 + ], + "angle": 0, + "content": "[31] Osman Semih Kayhan, Bart Vredebregt, and Jan C Van Gemert. Hallucination in object detection—a study in visual part verification. In 2021 IEEE International Conference on Image Processing (ICIP), pages 2234–2238. IEEE, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.442, + 0.469, + 0.51 + ], + "angle": 0, + "content": "[32] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014. 5, 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.511, + 0.469, + 0.566 + ], + "angle": 0, + "content": "[33] Seyedalireza Khoshkirat and Chandra Kambhametty. Sentence attention blocks for answer grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6080-6090, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.567, + 0.469, + 0.622 + ], + "angle": 0, + "content": "[34] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dólar. Panoptic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9404-9413, 2019. 5, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.623, + 0.469, + 0.706 + ], + "angle": 0, + "content": "[35] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4015-4026, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.707, + 0.469, + 0.788 + ], + "angle": 0, + "content": "[36] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 5, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.79, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[37] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692, 2023. 4, 5, 6, 8, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[38] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726, 2023. 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[39] Chunyuan Li, Zhe Gan, Zhengyuan Yang, Jianwei Yang, Linjie Li, Lijuan Wang, and Jianfeng Gao. Multimodal foundation models: From specialists to general-purpose assistants. arXiv preprint arXiv:2309.10020, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.149, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[40] Jianshu Li, Jian Zhao, Yunchao Wei, Congyan Lang, Yidong Li, Terence Sim, Shuicheng Yan, and Jiashi Feng. Multiple-human parsing in the wild. arXiv preprint arXiv:1705.07206, 2017. 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.206, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[41] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.291, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[42] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.362, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[43] Chang Liu, Henghui Ding, and Xudong Jiang. Gres: Generalized referring expression segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23592-23601, 2023. 5, 6, 8, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.419, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[44] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 5, 8, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.462, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[45] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.505, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[46] Jiang Liu, Hui Ding, Zhaowei Cai, Yuting Zhang, Ravi Kumar Satzoda, Vijay Mahadevan, and R Manmatha. Polyformer: Referring image segmentation as sequential polygon generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18653-18663, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.59, + 0.892, + 0.671 + ], + "angle": 0, + "content": "[47] Shangbang Long, Siyang Qin, Dmitry Panteleev, Alessandro Bissacco, Yasuhisa Fujii, and Michalis Raptis. Towards end-to-end unified scene text detection and layout analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1049-1059, 2022. 5, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.674, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[48] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.704, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[49] Jiasen Lu, Christopher Clark, Rowan Zellers, Roozbeh Mottaghi, and Aniruddha Kembhavi. Unified-io: A unified model for vision, language, and multi-modal tasks. In The Eleventh International Conference on Learning Representations, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.774, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[50] Ziqiao Ma, Jiayi Pan, and Joyce Chai. World-to-words: Grounded open vocabulary acquisition through fast mapping in vision-language models. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 524–544, 2023. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[51] Arjun Mani, Nobline Yoo, Will Hinthorn, and Olga Russakovsky. Point and ask: Incorporating pointing into visual question answering. arXiv preprint arXiv:2011.13681, 2020. 5, 6, 3" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14236" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[52] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 5, 8, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.471, + 0.232 + ], + "angle": 0, + "content": "[53] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 5, 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.471, + 0.302 + ], + "angle": 0, + "content": "[54] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.471, + 0.37 + ], + "angle": 0, + "content": "[55] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.371, + 0.471, + 0.427 + ], + "angle": 0, + "content": "[56] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.428, + 0.471, + 0.483 + ], + "angle": 0, + "content": "[57] Renjie Pi, Jiahui Gao, Shizhe Diao, Rui Pan, Hanze Dong, Jipeng Zhang, Lewei Yao, Jianhua Han, Hang Xu, and Lingpeng Kong Tong Zhang. Detgpt: Detect what you need via reasoning. arXiv preprint arXiv:2305.14167, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.471, + 0.567 + ], + "angle": 0, + "content": "[58] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In Proceedings of the IEEE international conference on computer vision, pages 2641–2649, 2015. 5, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.471, + 0.635 + ], + "angle": 0, + "content": "[59] Jordi Pont-Tuset, Jasper Uijlings, Soravit Changpinyo, Radu Soricut, and Vittorio Ferrari. Connecting vision and language with localized narratives. In Proceedings of the 16th European Conference on Computer Vision, pages 647-664, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.636, + 0.471, + 0.706 + ], + "angle": 0, + "content": "[60] Lu Qi, Jason Kuen, Tiancheng Shen, Jiuxiang Gu, Wenbo Li, Weidong Guo, Jiaya Jia, Zhe Lin, and Ming-Hsuan Yang. High quality entity segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4047-4056, 2023. 4, 5, 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.471, + 0.788 + ], + "angle": 0, + "content": "[61] Mengxue Qu, Yu Wu, Wu Liu, Xiaodan Liang, Jingkuan Song, Yao Zhao, and Yunchao Wei. RIO: A benchmark for reasoning intention-oriented objects in open environments. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.471, + 0.872 + ], + "angle": 0, + "content": "[62] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 5, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.872, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[63] Vignesh Ramanathan, Anmol Kalia, Vladan Petrovic, Yi Wen, Baixue Zheng, Baishan Guo, Rui Wang, Aaron Mar" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "quez, Rama Kovvuri, Abhishek Kadian, et al. Paco: Parts and attributes of common objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7141-7151, 2023. 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[64] Varun Nagaraj Rao, Xingjian Zhen, Karen Hovsepian, and Mingwei Shen. A first look: Towards explainable textvqa models via visual and textual explanations. arXiv preprint arXiv:2105.02626, 2021. 5, 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.205, + 0.892, + 0.274 + ], + "angle": 0, + "content": "[65] Anna Rohrbach, Marcus Rohrbach, Ronghang Hu, Trevor Darrell, and Bernt Schiele. Grounding of textual phrases in images by reconstruction. In Proceedings of the 14th European Conference on Computer Vision, pages 817-834. Springer, 2016. 5, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.274, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[66] Amanpreet Singh, Vivek Natarjan, Meet Shah, Yu Jiang, Xinlei Chen, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8317-8326, 2019. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.343, + 0.892, + 0.398 + ], + "angle": 0, + "content": "[67] Amanpreet Singh, Guan Pang, Mandy Toh, Jing Huang, Wojciech Galuba, and Tal Hassner. TextOCR: Towards large-scale end-to-end reasoning for arbitrary-shaped scene text. 2021. 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.399, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[68] Bharat Singh, Hengduo Li, Abhishek Sharma, and Larry S Davis. R-fcn-3000 at 30fps: Decoupling detection and classification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1081–1090, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.469, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[69] Zhi Tian, Chunhua Shen, Xinlong Wang, and Hao Chen. Boxinst: High-performance instance segmentation with box annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5443-5452, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.538, + 0.892, + 0.609 + ], + "angle": 0, + "content": "[70] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 5, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.609, + 0.892, + 0.664 + ], + "angle": 0, + "content": "[71] Maria Tsimpoukelli, Jacob L Menick, Serkan Cabi, SM Eslami, Oriol Vinyls, and Felix Hill. Multimodal few-shot learning with frozen language models. Advances in Neural Information Processing Systems, 34:200-212, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.664, + 0.892, + 0.72 + ], + "angle": 0, + "content": "[72] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.72, + 0.892, + 0.803 + ], + "angle": 0, + "content": "[73] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In International Conference on Machine Learning, pages 23318-23340. PMLR, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.804, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[74] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, et al. Vision: Large language model is also an open-ended decoder for vision-centric tasks. arXiv preprint arXiv:2305.11175, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.872, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[75] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clip" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14237" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.472, + 0.134 + ], + "angle": 0, + "content": "driven referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11686-11695, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[76] Chenyun Wu, Zhe Lin, Scott Cohen, Trung Bui, and Subhransu Maji. Phrasescut: Language-based image segmentation in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10216-10225, 2020. 4, 5, 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.469, + 0.248 + ], + "angle": 0, + "content": "[77] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.25, + 0.469, + 0.305 + ], + "angle": 0, + "content": "[78] Yixuan Wu, Zhao Zhang, Chi Xie, Feng Zhu, and Rui Zhao. Advancing referring expression segmentation beyond single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2628-2638, 2023. 5, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.307, + 0.469, + 0.374 + ], + "angle": 0, + "content": "[79] Chi Xie, Zhao Zhang, Yixuan Wu, Feng Zhu, Rui Zhao, and Shuang Liang. Described object detection: Liberating object detection with flexible expressions. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 5, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.378, + 0.47, + 0.446 + ], + "angle": 0, + "content": "[80] Bin Yan, Yi Jiang, Jiannan Wu, Dong Wang, Ping Luo, Zehuan Yuan, and Huchuan Lu. Universal instance perception as object discovery and retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15325-15336, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.449, + 0.469, + 0.504 + ], + "angle": 0, + "content": "[81] Zhengyuan Yang, Zhe Gan, Jianfeng Wang, Xiaowei Hu, Faisal Ahmed, Zicheng Liu, Yumao Lu, and Lijuan Wang. Unitab: Unifying text and box outputs for grounded vision-language modeling. 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.506, + 0.469, + 0.575 + ], + "angle": 0, + "content": "[82] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Levt: Language-aware vision transformer for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18155-18165, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.577, + 0.469, + 0.645 + ], + "angle": 0, + "content": "[83] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. MPLug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.469, + 0.689 + ], + "angle": 0, + "content": "[84] Shukang Yin, Chaoyou Fu, Sirui Zhao, Ke Li, Xing Sun, Tong Xu, and Enhong Chen. A survey on multimodal large language models. arXiv preprint arXiv:2306.13549, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.69, + 0.469, + 0.758 + ], + "angle": 0, + "content": "[85] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704, 2023. 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[86] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. 4, 5, 8, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[87] Yuhang Zang, Wei Li, Jun Han, Kaiyang Zhou, and Chen Change Loy. Contextual object detection with multimodal large language models. arXiv preprint arXiv:2305.18279, 2023. 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[88] Rowan Zellers, Yonatan Bisk, Ali Farhadi, and Yejin Choi. From recognition to cognition: Visual commonsense reasoning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6720-6731, 2019. 5, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[89] Shilong Zhang, Peize Sun, Shoufa Chen, Min Xiao, Wenqi Shao, Wenwei Zhang, Kai Chen, and Ping Luo. Gpt4roi: Instruction tuning large language model on region-of-interest. arXiv preprint arXiv:2307.03601, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.26 + ], + "angle": 0, + "content": "[90] Bo Zhao, Boya Wu, and Tiejun Huang. Svit: Scaling up visual instruction tuning. arXiv preprint arXiv:2307.04087, 2023.5,3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[91] Yang Zhao, Zhijie Lin, Daquan Zhou, Zilong Huang, Jiashi Feng, and Bingyi Kang. Bubogpt: Enabling visual grounding in multi-modal llms. arXiv preprint arXiv:2307.08581, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.892, + 0.401 + ], + "angle": 0, + "content": "[92] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16793-16803, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.404, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[93] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.461, + 0.892, + 0.515 + ], + "angle": 0, + "content": "[94] Yuke Zhu, Oliver Groth, Michael Bernstein, and Li Fei-Fei. Visual7w: Grounded question answering in images. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4995–5004, 2016. 5, 6, 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14238" + } + ] +] \ No newline at end of file diff --git a/2024/GROUNDHOG_ Grounding Large Language Models to Holistic Segmentation/bdfcf2de-41e8-4d27-af7d-65821e6b79f8_origin.pdf b/2024/GROUNDHOG_ Grounding Large Language Models to Holistic Segmentation/bdfcf2de-41e8-4d27-af7d-65821e6b79f8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8260294511b02113d68b45e90c08f3d133e5f637 --- /dev/null +++ b/2024/GROUNDHOG_ Grounding Large Language Models to Holistic Segmentation/bdfcf2de-41e8-4d27-af7d-65821e6b79f8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fd338f009be8c280c9707a90fec6f021bcff07dd0964d91eef5e7edc6045def0 +size 2084552 diff --git a/2024/GROUNDHOG_ Grounding Large Language Models to Holistic Segmentation/full.md b/2024/GROUNDHOG_ Grounding Large Language Models to Holistic Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e561ca4a3c1d71ba10e44a6815406c934ce82b58 --- /dev/null +++ b/2024/GROUNDHOG_ Grounding Large Language Models to Holistic Segmentation/full.md @@ -0,0 +1,353 @@ +# GROUNDHOG: Grounding Large Language Models to Holistic Segmentation + +Yichi Zhang $^{1\dagger}$ , Ziqiao Ma $^{1\dagger}$ , Xiaofeng Gao $^{2}$ , Suhaila Shakiah $^{2}$ , Qiaozi Gao $^{2}$ , Joyce Chai $^{1}$ + +1University of Michigan, 2Amazon AGI + +zhangyic@umich.edu + +https://groundhog-ml1m.github.io/ + +![](images/817673d7b306725870a94ad6aeeaef6db83f1b4c43b730771fabd7dec2fffb7c.jpg) + +![](images/54bd5a7be146835c9c4ce0f5c2ba8ebb098ebef2ac63383612da255f4fbb58c2.jpg) + +Describe the image briefly. + +A man and a little girl are sitting in a shopping cart. + +![](images/ccbb31425917613174b95d2ad4b065a5c65a3dbc814f2029381d7d512e7f56c9.jpg) + +![](images/4a1b2c5bb92f244b2bcbd940a39ec34b08e3c2be425deb32a6c218fe0fce444a.jpg) + +![](images/4c68b29cd239bf69488cce1fd1020619a1d355e9b22234fbd3d25a5795fbf0f2.jpg) + +Generate a short caption. + +Two dogs are playing with a stick in a field. + +![](images/5429eeab0ac8ed394c150f376deea831fafd34399aeb88b0b3d9dd27b40df257.jpg) + +![](images/e322558dc6e91ba19a2b8c04497025b653f07fbde7cde1448ef382f777bef8c9.jpg) + +Describe the given picture in very detail. + +In this image, we can see a boat on the water. There are few people inside the boat. There are also few people on the rocks. In the background, there are trees, hills, and sky. + +![](images/02a2638a71e499bc68c7a9ef19695b53f19802294ce4a234f52ee779038b4446.jpg) + +![](images/5f446983be845abc58bf5800a0a7fa28bc511843f6a4d7208223e581cd2d0421.jpg) + +![](images/d18ad8ab055918dc33ea563474a1b189b9950d7a8ad89f0170ffd432ceefe8c6.jpg) +Figure 1. We propose GROUNDHOG, a multimodal large language model that enhances its text output with pixel-level phrase grounding across diverse semantic granularities. The figure demonstrates outputs from our model on the four task types we considered in this work. + +Could you please segment out "laptop barely onscreen not apple" in the image? + +Laptop barely onscreen not apple. + +![](images/d08af1999556258e367d3ae3b17b97e5b763df55800993414ee9d20b69ac081d.jpg) + +![](images/17ff19cbf741628e010d8814aee4f465abf5a21a5d15d5ba303fca2f1315db21.jpg) +(a) Grounded Image Captioning (GIC). + +![](images/0c37a129018dcdce8c406b75452fcca98cc9a672fc1c9bff5644eb8b01b36352.jpg) + +What date is shown on the calendar? + +Feb 2. + +![](images/54b342a5cfb2ddaa4077cad87215fd00418b68af00a4e8c4bf7dcc06fbd442c5.jpg) + +![](images/cc47e31e680bef21ebce829533cb6b09ba7c79499bd4f0a8c6a546ba1a97f8b5.jpg) + +![](images/ad2fbb69656e9111eb7e474fea7b1d498f28be11097c5c19d527dbf7c82bc28a.jpg) +(d) Referential Dialogue (RD). + +Provide a distinct description for that $< \text{PTR}>$ + +Blue container with apples. + +![](images/9c1d57fcf0661fc141f0ebe98ac24998aabf81883db00c4a04f8ac274d27dea9.jpg) + +# Abstract + +Most multimodal large language models (MLLMs) learn language-to-object grounding through causal language modeling where grounded objects are captured by bounding boxes as sequences of location tokens. This paradigm lacks pixel-level representations that are important for fine-grained visual understanding and diagnosis. In this work, we introduce GROUNDHOG, an MLLM developed by grounding Large Language Models to holistic segmentation. GROUNDHOG incorporates a masked feature extractor and converts extracted features into visual entity tokens for the MLLM backbone, which then con + +nects groundable phrases to unified grounding masks by retrieving and merging the entity masks. To train GROUNDHOG, we carefully curated M3G2, a grounded visual instruction tuning dataset with Multi-Modal Multi-Grained Grounding, by harvesting a collection of segmentation-grounded datasets with rich annotations. Our experimental results show that GROUNDHOG achieves superior performance on various language grounding tasks without task-specific fine-tuning, and significantly reduces object hallucination. GROUNDHOG also demonstrates better grounding towards complex forms of visual input and provides easy-to-understand diagnosis in failure cases. + +# 1. Introduction + +Multimodal large language models (MLLMs) have received an increasing amount of attention to address tasks that necessitate non-linguistic knowledge, e.g., perception and reasoning about the visual world [39, 84]. For fine-grained visual understanding, grounded MLLMs often learn language-to-object grounding by causal language modeling, where grounded objects are captured by bounding boxes as sequences of location tokens. However, bounding boxes are insufficient in indicating amorphous stuff [5], semantic parts of objects [23], finer-grained regions with irregular shapes [26], or groups of instances at the same time. As a result, a single bounding box can often include other irrelevant semantics in order to engulf the target entities, leading to ambiguity in detection. In addition, the generated box coordinate lacks interpretability. When the model hallucinates, such as incorrectly predicting the association between objects and language, it is hard to diagnose whether the problem is due to the model's failure to detect the object, or its incorrect alignment of the object with language. + +To address these issues, in this work, we introduce GROUNDHOG, an MLLM developed by grounding Large Language Models to holistic segmentation. Our goal of language grounding is to connect text spans that refer to or can be deduced from visual information, termed as groundable phrases [50], to their corresponding regions of visual entities (Figure 1). GROUNDHOG incorporates a masked feature extractor that takes an input image and a set of class-agnostic entity mask proposals, and converts each mask's features into visual entity tokens for an MLLM backbone. This MLLM then connects groundable phrases to unified grounding masks by retrieving and merging the entity masks. Compared to previous grounded MLLMs, GROUNDHOG unlocks unprecedented pixel-level vision-language alignment. It naturally supports visual pointers as input, and can plug-in-and-play with any choice of mask proposal networks, e.g., Segment Anything Model (SAM) [35], domain-specific semantic segmentation models, or user-provided mask candidates. We introduce an enhanced Mask2Former [10] as our default mask proposal network, which detects regions at multiple granularities, e.g., instances (things and stuff), semantic parts, and visual text, leading to a holistic coverage of visual semantics. + +To train GROUNDHOG, we curated a Multi-Modal Multi-Grained Grounding (M3G2) dataset consisting of 2.5M text-image pairs for visually grounded instruction tuning, consisting of 36 sub-problems derived and augmented from 27 existing datasets. We present extensive experiments on vision-language tasks that require grounding, including grounded language generation with minimal object hallucination, language-guided segmentation, visual question answering with answer grounding, and referential dialog with spatial pointer inputs (Figure 1). Our empirical results show + +that GROUNDHOG, without task-specific fine-tuning, can achieve superior or comparable performance with previous models that either require fine-tuning or are specialized only for that dataset. In addition, GROUNDHOG has supports easy-to-understand diagnosis when grounding fails. + +# 2. Our Method: GROUNDHOG + +The language grounding task can be succinctly delineated into two fundamental components: localization and recognition, as established in the literature [50, 68, 92]. Such categorization not only aids in the identification of object presence (objectness) without reliance on specific object classes, but also sets the stage for models to be robust in open-vocabulary settings. Building upon this framework, we formulate the grounding process as an entity segment selection problem, which involves (1) proposing entity segmentation masks where the masks encapsulate regions with discernible semantic content, and (2) recognizing the retrieved entities through the understanding of both visual and language context. Concurrently performing both tasks is where MLLMs bring a distinct advantage. This decoupled design of entity mask proposal and language-guided grounding brings several advantages. First, it allows independent improvement of the mask proposal model and MLLM, where specialized data, training, and inference setups can be applied. Second, by decoupling language grounding, it becomes straightforward to determine if a failure is due to the model's inability to propose the entity segment, or its misalignment of the object with the language, thus improving the interpretability of the whole framework. Third, as shown later, when connecting the two parts to work in tandem in a model-independent manner, the MLLM can benefit from multiple different vision specialist models in a plug-and-play fashion. In the remainder of this section, we give details of our model design. + +# 2.1. Building Entity Features from Masks + +Our approach assumes the availability of a mask proposal model, which is capable of generating a set of class-agnostic entity masks from an image with high coverage. In contrast to prior studies that relied on low-level features [8, 13, 45, 56], GROUNDHOG interprets the image as a collection of entities. The primary challenge then becomes the derivation of effective visual features to accurately represent these entities. To achieve a complete decoupling of the MLLM from the mask proposal model responsible for providing the masks, we propose to condition the entity features solely on the binary masks without using any embeddings from the mask proposal model. Specifically, the mask corresponding to each entity is employed to extract patch features from pretrained vision foundation models, such as CLIP [62] and DINOv2 [54], through a convolutional mask pooling layer [12]. Given that the feature map dimensions + +![](images/371348875dd9fbb95ccac0fca9a5c92056d4e70990a409abfe0486c28873d11c.jpg) +Figure 2. The model architecture of GROUNDHOG model. Given a set of class-agnostic entity mask proposals, the masked feature extractor first extracts the feature of each entity as the visual input of the multi-modal large language model (left). The output hidden states of the grounding tokens are averaged and used to retrieve the entities to ground, which will be merged into a single grounding mask for the phrase. Modules are colored by their trainability: parameter-free operators (grey), frozen (blue), trainable (orange), and partially trainable (mix). + +![](images/d4774ca35a652f1e2e464c2a4831bcc83771df0a366352dfd27e92fd20d61e16.jpg) +Figure 3. GROUNDHOG can take arbitrary spatial prompts that can be resolved by an interactive segmentation model, such as SAM. The placeholder pointer token $<\mathrm{PTR}>$ will be replaced by the extracted entity features and fed as input to the model. + +are usually smaller than those of the mask proposals, we resize the masks to match the size of the feature maps prior to pooling. The pooled features are then fed into a Multi-Layer Perceptron (MLP) network to align with the input embeddings of the MLLM. We empirically find the combination of CLIP and DINOv2 features yields the best result, and these features are added to obtain the final input visual entity tokens to the MLLM. + +Spatial Prompts Furthermore, for grounded MLLMs to be more broadly applicable, they must be capable of interpreting multi-modal user inputs, including spatial prompts. Thanks to the mask model agnostic design, GROUND-HOGcan seamlessly support such inputs. As demonstrated in Figure 3, by applying an interactive segmentation model such as Segment-Anything (SAM) [35], arbitrary spatial prompts can be translated into binary masks and processed by the same masked feature extractor we just introduced. This extracted feature for the pointed entity will replace the pointer token placeholder in the textual input. + +# 2.2. Language Grounding to Entity Segmentation + +Existing box-grounded MLLMs typically append location tokens after the groundable phrases [8, 9, 56, 85]. However, this method is not readily interpretable. To alleviate this disconnect, we introduce a pair of grounding tokens $\langle \mathrm{GRD} \rangle$ and $\langle \mathrm{/GRD} \rangle$ to indicate the start and end of groundable phrases, with the assumption that grounding these phrases requires mapping to certain representations of visual entities irrespective of the visual modality. In Figure 2, a sentence can be represented as I see $\langle \mathrm{GRD} \rangle$ two dogs $\langle \mathrm{/GRD} \rangle$ on $\langle \mathrm{GRD} \rangle$ the beach $\langle \mathrm{/GRD} \rangle$ , with two distinct visual entities grounded. The representation of each groundable phrase, termed as the grounding query, is obtained by adding $\langle \mathrm{GRD} \rangle$ and $\langle \mathrm{/GRD} \rangle$ 's output embedding from the last transformer layer of the MLLM. The representation is then used to retrieve the entities that the phrase should be grounded to. In particular, we concatenate the grounding query with the last layer output of each visual entity token, and use an MLP to predict a scalar score for each entity. Finally, we merge all the mask proposals into one single mask with pixel-wise maximization: + +$$ +\mathcal {M} _ {h, w} = \max _ {q} \left(\mathcal {S} _ {q} \cdot \widehat {\mathcal {M}} _ {q, h, w}\right) +$$ + +where $S_{q}$ is the normalized score of the $q$ -th mask ranging from 0 to 1, and $\widehat{\mathcal{M}}_{q,h,w}$ denotes the pixel probability at position $(h,w)$ for the $q$ -th mask. Note that a phrase may ground to multiple entities, thus multiple mask proposals may get a high score simultaneously and be selected in conjunction. One of the primary benefits of this decoupled design is its transparency in the selection of entities. Users can easily visualize both the mask proposals and their respective + +scores, providing a clear understanding of how a grounding mask is predicted. This level of clarity and interpretability is a significant advantage, offering users a tangible insight into the model's grounding process. + +# 2.3. Towards Holistic Entity Mask Proposals + +In order to support holistic language grounding to arbitrary segmentations, the entity proposal should have two essential properties. First, the proposals should strike a delicate balance in terms of semantic atomicity. While it is possible to merge multiple proposals later to form multi-entity segmentations, the reverse, i.e., dividing a single proposal into smaller segments, is not feasible. Therefore, instance segmentation is generally preferred over semantic segmentation. However, the segmentation should not be excessively fine-grained to the extent that it compromises basic semantic integrity. Over-segmentation can lead to a loss of the coherent concept of an entity, which is detrimental to the grounding process. Second, the entity proposals should have a high coverage of entities, encompassing a diverse range of granularities. This includes not only tangible objects (things) and amorphous concepts (stuff) but also extends to sub-components of objects (parts of things) and structured regions such as areas containing visual text. The ability to propose entities across this spectrum of granularity is pivotal, as it directly determines the upper bound of the grounding capability of MLLM. + +We initiated our study with a Mask2Former model pretrained on the COCO panoptic segmentation dataset. However, preliminary experiments revealed its limitations in semantic coverage and adaptability to open-world scenarios. To enhance this, we developed Mask2Former+, an upgraded version designed for multi-grained segmentation. This upgrade involved creating a diverse dataset by merging annotations from various sources, including COCO [5], LVIS [25], Entity-v2 [60], Pascal [16], PACO [63] (Figure 7); MHP-v2 [40] for human part parsing; and TextOCR [67] for text segmentation. We expanded the model's capabilities by adding 50 expert queries each for semantic parts and visual text regions, alongside the original 200 entity queries. We assessed Mask2Former+’s performance on 1000 images from validation splits from 4 grounding benchmarks, RefCOCO+ [86], PhraseCut [76], ReasonSeg [37], and TextVQA-X [66]. We use the Any-IoU [30] metric for evaluation, i.e., for each ground truth mask, we extract the most overlapped mask proposals and compute the IoU, then take the average. As Table 1 demonstrates, Mask2Former+ shows consistent improvements across all domains, particularly in those significantly divergent from COCO. This highlights its enhanced adaptability and precision in a broader range of segmentation challenges, providing a good mask proposal model for GROUNDHOG. We refer to Appendix A for more details of the model and data. + +
ModelRefCOCO+PhraseCutReasonSegTextVQA-X
Mask2Former0.8670.5630.6020.137
Mask2Former+0.8730.6240.7450.446
+ +Table 1. The average Any-IoU of the proposals on each dataset. The vanilla Mask2Former is trained on the COCO-Panoptic dataset and our Mask2Former+ is trained on our combined dataset. Mask2Former+ obtains a consistent improvement in all scenarios, especially in non-COCO domains. + +
TaskDatasetGr. Ann.Sem. Gran.# Pairs
MBPoSThPaGTx
GCAPPNG✓✓✓✓✓132k
Flickr30K-Entity✓✓✓✓✓149k
RESRefCOCO✓✓✓113k
RefCOCO+✓✓✓112k
RefCOCOg✓✓✓80k
RefCLEF✓✓✓105k
gRefCOCO✓✓✓194k
PhraseCut✓✓✓✓✓✓85k
D-Cube✓✓✓✓10k
ReasonSeg✓✓✓✓✓✓1k
RIO✓✓✓✓28k
SK-VG✓✓23k
GVQAVizWiz-G✓✓✓✓✓6k
TextVQA-X✓✓✓15k
GQA✓✓✓✓✓302k
VQS✓✓20k
Shikra-BinaryQA✓✓✓✓✓4k
EntityCount✓✓✓✓✓✓11k
FoodSeg-QA✓✓✓✓7k
LVIS-QA✓✓✓✓✓95k
RDRefCOCO-REG✓✓✓✓17k
RefCOCO+-REG✓✓✓✓17k
RefCOCOg-REG✓✓✓✓22k
gRefCOCO-REG✓✓✓✓20k
VG-SpotCap✓✓✓✓✓✓247k
V7W✓✓✓23k
PointQA✓✓64k
VCR✓✓✓156k
ShikraRD✓✓✓✓✓✓2k
SVIT-RD✓✓✓✓✓✓33k
Guesswhat✓✓✓✓193k
VG-RefMatch✓✓✓✓✓✓247k
HierText✓✓✓✓6k
M3G2 (Total)2.5M
+ +Table 2. Summary of datasets included in M3G2. We show the availability of Grounding Annotations (Box, Mask, and Pointer inputs), the Semantic Granularity (Stuff, Things, Parts, Groups, and Text), and the number of text-image pairs for training. + +# 3. Our Dataset: M3G2 + +In this section, we introduce M3G2, a Multi-Modal MultiGrained Grounding dataset consisting of 2.5M text-image pairs for visually grounded instruction tuning, consisting of 36 sub-problems derived and augmented from 27 existing datasets. We re-organize and augment public datasets of language grounding, visual question answering, referring expression segmentation, and referring expression generation into various forms of visually grounded dialogue for grounded instruction tuning, outlined briefly in Table 2. The dataset is categorized into four main types: (1) Grounded Image Captioning (GIC), (2) Referential Expression Segmentation (RES), (3) Grounded Visual Question Answer + +ing (GVQA), and (4) Referential Dialog (RD). We provide illustrated descriptions of our prompt design, accompanied by examples of each task type as depicted in Figure 4. We detail the task schema in the following sections and provide the complete sets of templates in Appendix B. + +# 3.1. Grounded Image Captioning (GIC) + +The task of grounded image captioning requires the model to produce a narrative for the visual scene, and accurately identify and associate the groundable phrases with their respective binary segmentation masks. The objective of this task is to empower the model to articulate the scene while acknowledging various visual elements and their spatial interrelations. We incorporate the Panoptic Narrative Grounding (PNG) dataset [34] for dense and detailed scene descriptions, as well as the Flickr30K-Entity dataset [58] for concise descriptions of the salient contents in the image. We create a collection of task prompt templates that instruct the model to describe the image either in detail or briefly. + +# 3.2. Referring Expression Segmentation (RES) + +In contrast to previous tasks, the referring expression segmentation task requires that the model generates a segmentation mask based on a given referring expression. Besides the RefCOCO series [43, 52, 86], we have further leveraged existing RES benchmarks [37, 65, 76, 78, 79] for this purpose. As shown in Figure 4, we instruct the model to localize the referred objects, with the expected output being a repetition of the referring expression whose grounding mask is the target object. + +# 3.3. Grounded Visual Question Answering (GVQA) + +The task of grounded visual question answering requires the model to comprehend a question (with optional pointers) and to produce an answer that is grounded to a binary segmentation mask that justifies the answer. The goal of this task is to enable natural QA-based interaction with users and reasoning in the model with grounded explanations. Specifically, we harvest and adapt a collection of public VQA datasets with grounding annotations [6, 8, 19, 29], QA benchmarks on visual text [64], and create populated QA pairs from segmentation datasets [25, 60]. Our prompt templates instruct the model to respond either as open-ended answers or by selecting from multiple choices, with the response anchored to a segmentation mask (Figure 4). + +# 3.4. Referential Dialogue (RD) + +The task of referential dialogue requires the model to conduct dialogue communication with users, especially when conditioned on user-provided spatial prompts. This includes existing RD datasets [8, 51, 88, 90, 94], multi-turn augmentations from segmentation datasets [17, 36, 47] as well as the referring expression generation (REG) task the + +![](images/f4fd885ed5f3526f41a055fa0bb8365a1ae69932f147bc99f70830caea9e9194.jpg) +Figure 4. The M3G2 dataset for grounded visual instruction tuning. M3G2 is a diverse dataset of multiple granularities, unifying 4 different task types with visually grounded dialogue. + +RefCOCO series [43, 52, 86]. The REG task differs from the region captioning task in that it demands the description to be a referring expression that distinctly identifies the targeted object. Effective REG calls for the model to engage in dialogue interactions cooperatively, adhering to the Gricean Maxims [24] which dictate that communication should be as informative, truthful, relevant, and clear as necessary. + +# 4. Experiment and Analysis + +We adopt the LLaMA2-7B model [70] as our base LLM, and initialized the weight from LLaVA-1.5 [44]. For the vision encoders, we use the OpenAI CLIP@336 [62] model and DINOv2-L/14-reg [15] pretrained checkpoints. Please refer to Appendix C for more implementation details. + +4.1. Generalist in Grounded Vision-Language Tasks We first demonstrate GROUNDHOG's capabilities as a generalist model for three different types of grounded vision-language tasks. It's worth noting that, unlike previous work that needs dataset-specific fine-tuning on each of the tasks, GROUNDHOG can achieve comparable performance on all the tasks directly after training on M3G2, i.e., all the reported results from our model are from a single set of weights without any dataset-specific fine-tuning. + +Language Grounding To Segmentation. We start by evaluating the model on language grounding tasks, which takes text as input and generates segmentation masks as output. We assess GROUNDHOG on Referential Expression Segmentation (RES) [32] and Caption Phrase Grounding (CPG) tasks. While traditional RES benchmarks [32, 53] focus on single-instance referents requiring primarily visual understanding, we expanded our evaluation to include complex scenarios involving multi-instance or negative queries + +
ModelSingle InstanceMulti-/No InstanceReasoning
RefCOCORefCOCO+RefCOCOggRefCOCOPhraseCutReasonSegRIO
valtest-Atest-Bvaltest-Atest-Bval-utest-uvaltestvaltest-ctest-u
Specialist
MDETR [30]---------53.7-44.122.0
CRIS [75]70.573.266.162.368.153.759.960.455.3----
LAVT [82]72.775.868.862.168.455.161.262.158.4----
ReLA [43]73.876.570.266.071.057.765.066.063.6-22.4--
PolyFormer [46]76.078.373.369.374.661.969.270.2---48.826.8
UNINEXT-H [80]82.283.481.372.576.466.274.776.4-----
Generalist
LISA7B [37]74.176.571.162.467.456.566.468.5--44.0--
LISA7B (FT) [37]74.979.172.365.170.858.167.970.6--52.9--
7B78.579.975.770.575.064.974.174.666.754.556.257.933.9
+ +Table 3. Results on 7 Referring Expression Segmentation (RES) benchmarks with single instance queries [32, 53], multi-/null instance queries [43, 76] and reasoning-based queries [37, 61]. We report cIoU for RefCOCO+/g and mIoU for other benchmarks, respectively. + +
ModelFlickr30K-E
R@1_valR@1_test
Shikra13B77.478.4
Ferret13B81.184.8
Shikra7B75.876.5
Ferret7B80.482.2
7B79.279.8
+ +Table 4. Top-1 box recall results on Flickr30K-Entity [58]. + +
ModelAR\( \mathbf{A}\mathbf{R}_{\mathrm {th}} \)PNG \( \mathbf{A}\mathbf{R}_{\mathrm {st}} \)\( \mathbf{A}\mathbf{R}_{\mathrm {s}} \)\( \mathbf{A}\mathbf{R}_{\mathrm {p}} \)
PiGLET65.964.068.667.254.5
\( 7_{\text{B}} \)66.865.069.470.457.7
Table 5. Phrase grounding results on PNG [21].
ModelTextVQA-X [mIoU]
SAB29.0
\( 7_{\text{B}} \)39.8
+ +Table 6. Visual text QA results on the TextVQA-X [66] validation set. + +
ModelPointQA Twice AccV7W Acc
Shikra13B70.385.3
GPT4RoI13B-84.8
Shikra7B--
GPT4RoI7B-81.8
7B72.485.5
+ +[43, 76], and those necessitating common sense reasoning [37, 61]. The results in Table 3 show GROUNDHOG outperforming the generalist model LISA across all benchmarks and achieving significant improvements over specialist models in multi-instance, null, and reasoning-based RES tasks. It also performs comparably on the competitive RefCOCO series. For CPG tasks, which involve grounding all phrases in a caption and demand a deep understanding of the context for coreference resolution, we first evaluated GROUNDHOG on the Flickr30K-Entity dataset [58]. Since this dataset only has box annotations, we convert the mask predictions of our model to box and compute the top-1 box recall following the merged-box protocol [30]. Despite not specializing in predicting boxes, GROUNDHOG still outperforms Shikra 7B/13B [8] and is on par with Ferret-7B [85] in a concurrent work (Table 4). Additionally, on the PNG dataset [34] which tests phrase grounding in longer narratives, GROUNDHOG surpasses the previous state-of-the-art model, PiGLET [22], in all metrics including average recall of grounding masks and detailed scores for things, stuffs, and singular and plural entities (Table 5). + +Grounded Language Generation. Our model excels in generating language that accurately grounds to segmentation masks during user conversations. Quantitatively, we assess grounded captioning on the Flickr30K-Entity dataset [58], employing standard text generation metrics + +Table 7. Results on PointQATwice [51] and V7W [94] test sets. + +
ModelBleu-4METEORCIDErSPICEF1all
Shikra13B--73.9--
Ferret13B37.025.576.118.315.1
Ferret7B35.124.674.818.015.0
7B36.726.591.320.432.1
+ +Table 8. Grounded Captioning on Flickr30K-Entity [58]. + +such as Bleu-4 [55], METEOR [4], CIDEr [72], and SPICE [2] for language quality; and the $\mathrm{F1}_{\mathrm{all}}$ score for grounding accuracy following You et al. [85]. As shown in Table 8, GROUNDHOG significantly surpasses existing box-based grounded MLLMs, even their 13B versions, in both language quality and grounding accuracy. This improvement is hypothesized to stem from the diverse task distribution in our M3G2 dataset. For groundable question answering, we evaluate on the TextVQA-X benchmark [64]. GROUNDHOG outperforms the state-of-the-art specialist model SAB [33] by a significant margin, as measured by the mean IoU of the predicted mask (Table 6). + +Spatial Prompt Understanding. For grounded MLLMs, accurately interpreting multimodal instructions is essential, particularly in interactive tasks. We evaluated its performance on two pointer-based QA benchmarks, PointerQATwice [51] and V7W [94], which require the model to answer questions guided by spatial prompts, such as bounding boxes. The model is tasked to generate free- + +![](images/dfe920c57005f4d9d6b538244aeaf7653366145867b38e7e3bbdce9d5d0e890f.jpg) +Figure 5. Region caption using the best match proposal from Mask2Former+ versus from SAM. Mask2Former+ fails to propose the exact mask of the spire, leading to a less precise caption. + +![](images/f44c2995b738f414c001eb4996302b0a3f3e927bcbdbbb4870b7e119721bd49c.jpg) +Input Prompt M2F+ Best Match SAM + +![](images/b503064ea84dd4572d44b2a4073a2f5a60f5c664633c3b3f77788d798bdf0886.jpg) + +![](images/a99b56cdcb3679d9a612a7a78dfc3529ef504a9550cc9202fe2643846d33ed4b.jpg) +Figure 6. Illustration of a partially correct grounding. The grounding phrase and the ground truth mask are shown on the left. The top-4 mask proposals are presented, with highly-scored masks (green) selected for the merged mask, and low-scored masks (red) excluded. This illustrates the failure to recognize the word "KWIK" by the MLLM, despite its successful proposal. + +form textual answers in PointerQA $_{\text{Twice}}$ , and selects from multiple-choice options in V7W. GROUNDHOG demonstrates superior performance in these benchmarks, outperforming previous models as shown in Table 7. This highlights its effectiveness in spatial understanding and response accuracy. To further demonstrate the effectiveness of using SAM for the pointer-to-mask conversion, we show the best-matched mask proposal from our Mask2Former+ model in comparison to the mask from SAM in Figure 5. While the best match proposal from the Mask2Former+ model includes a broader area, the SAM-generated mask offers a more precise representation of the specified region, potentially leading to a more accurate caption. + +# 4.2. Trustworthiness and Transparency + +Beyond its superior performance as a grounding generalist, we highlight two key improvements for creating a more trustworthy and transparent agent. + +Reduced Object Hallucination. Thanks to the varied task distribution and the inclusion of negative question-answering samples in M3G2 dataset, GROUNDHOG significantly reduces object hallucination. We assessed this using the POPE [42] benchmark, which includes binary questions about object existence across three splits, each with a different object distribution (with an order of difficulty Random $<$ Popular $<$ Adversarial). Remarkably, GROUND + +
ModelAccuracyPrecisionRecallF1 ScoreYes (%)
Random
mPLUG-Owl53.3051.7199.5368.0696.23
LLaVA54.4352.3299.8068.6595.37
MultiModal-GPT50.0350.02100.0066.6899.97
MiniGPT-477.8375.3882.6778.8654.83
InstructBLIP88.7385.0893.9389.2955.20
Shikra-13B86.9094.4079.2686.1943.26
Ferret-13B90.2497.7283.0089.7643.26
7B91.0385.8096.4090.7945.88
Popular
mPLUG-Owl50.6350.3299.2766.7998.63
LLaVA52.4351.2599.8067.7297.37
MultiModal-GPT50.0050.00100.0066.67100.00
MiniGPT-468.3064.2782.4072.2164.10
InstructBLIP81.3775.0793.9383.4562.57
Shikra-13B83.9787.5579.2083.1645.23
Ferret-13B84.9088.2480.5384.2145.63
7B90.1385.9393.8189.7045.80
Adversarial
mPLUG-Owl50.6750.3499.3366.8298.67
LLaVA50.7750.3999.8766.9899.10
MultiModal-GPT50.0050.00100.0066.67100.00
MiniGPT-466.6062.4583.2771.3766.67
InstructBLIP74.3767.6793.3378.4568.97
Shikra-13B83.1085.6079.6082.4946.50
Ferret-13B82.3683.6080.5382.0048.18
7B86.3385.9386.6386.2849.60
+ +Table 9. Object hallucination results on the POPE [42] benchmark. + +HOG consistently outperforms other models in both accuracy and F1 score across all splits, particularly on the more challenging ones. It shows an absolute improvement of $5.2\%$ in accuracy for Popular and $4.0\%$ for Adversarial over the previously best-performing model. This suggests that our model's enhanced grounding capability plays a significant role in mitigating the object hallucination problem. + +Explainability and Diagnosability. Another important highlight of GROUNDHOG is its enhancement of explainability through the decoupled design of entity proposal and selection, as outlined earlier in section 2.2. This is exemplified in the case study illustrated in Figure 6, which illustrates the mask proposal scoring and selective merging process of our model. We show the top-4 masks, where the higher-score masks are labeled in green while the lower-score masks are labeled in red. Users can easily interpret that the failure is due to the incapability of MLLM to recognize the word "KWIK", despite it being successfully localized and proposed as an entity candidate. + +# 4.3. Ablation Studies + +We performed ablation studies to validate our design decisions, training, and evaluating a subset of the M3G2 dataset that includes RefCOCO+, Flickr30K, and TextVQA. These cover a range of visual entities from various image sources and granularities. We start by comparing our Mask2Former+ with the original Mask2Former for mask proposal effectiveness. As indicated in Table 10, the orig- + +
SetupRefCOCO+Flickr30KTextVQA-X
Mask Proposal Models
Mask2Former67.169.09.8
Mask2Former+66.677.234.0
Entity Features
CLIP59.875.032.0
DINOv262.376.328.4
CLIP+DINOv266.677.234.0
Grounding Query
<GRD> only64.467.534.2
</GRD> only64.477.233.5
Sum66.677.234.0
Eval Input Resolution
224–48054.767.227.6
480–64065.576.727.6
800–102466.677.234.0
+ +Table 10. Ablation study on model design choices and evaluation setups. Models are trained on RefCOCO+, Flickr30K, TextVQA-X and tested on corresponding validation sets. + +inal Mask2Former performs slightly better on RefCOCO, as it is developed specificity on COCO object categories. However, Mask2Former+ significantly surpasses the original in domains with non-COCO entities. Our second set of experiments examined the choice of visual entity features. Although using either CLIP or DINOv2 features alone shows advantages in specific datasets, their combination consistently yields the best results across all datasets. To obtain a robust grounding query representation, we experimented with using the output embedding of the token, the token, and their sum. We found that the latter approach achieves the best overall results. Finally, we demonstrate that our decoupling design of the mask proposal model and MLLM allows for training at a lower resolution (320px) to expedite grounding training, while scaling up the resolution during evaluation enhances performance. + +# 5. Related Work + +Multimodal Large Language Models. Building on the recent advance of large language models (LLMs), there is an increasing effort in adapting pretrained large language models for multimodal tasks, such as understanding and interpreting visual information [1, 71]. More recently, visual instruction tuning has gained much interest due to its surprising performance with a modest amount of data and computing resources. Various models have been developed, noticeably MiniGPT4[93], LLaVA [44, 45] and concurrent models [13, 20, 38, 74, 83]. Despite their promising performances, MLLMs often produce objects that are not presented in the given images, a phenomenon referred to as the object hallucination problem [14, 31, 42]. + +MLLM with Language Grounding The ability to connect language to their corresponding visual elements in the physical world, known as grounding [27], is crucial in everyday human communication about our shared sur + +roundings. Grounding datasets have been shown to benefit vision-language pre-training, both in terms of object-level recognition [41] and language learning [50]. Recent works unify text and grounding regions into token sequences [49, 73, 81] in casual language modeling. Based on such paradigm, researchers have developed a family of grounded MLLM, including GPT4ROI [89], Kosmos-2 [56], Shikra [8], PVIT [7], BuboGPT [91], Qwen-VL [3], and Ferret [85]. Despite their promising performance, these models focus on object grounding to bounding boxes, which cannot handle pixel-level grounding across various semantic granularities. Furthermore, it lacks the diagnosability and explainability in failure cases. Our model fills this gap. + +Language-Guided Semantic Localization The field of language-guided semantic localization has a long history in the vision-language research community, requiring that the model localize a given referring expression with bounding boxes or segmentation masks. This task has evolved from early attempts to understand simple referring expressions within images, such as the well-known RefCOCO series [52, 86] and their generalized variant [43] that takes no-target and multi-target into account. The integration of advanced language reasoning from LLMs has enabled research to tackle even more nuanced reasoning tasks that involve complex language contexts [37, 57, 87]. Notably, LISA [37] formulates a reasoning segmentation task to bring language-informed reasoning into semantic segmentation, and contributes a powerful baseline. Our model builds on these developments, but is designed to be more universally applicable as a grounded MLLM. + +# 6. Conclusion + +In this study, we introduce GROUNDHOG, a novel framework designed to enable pixel-level explainable grounding in large language models, leveraging holistic segmentation. The system builds upon a pre-trained mask proposal network to provide pixel-level visual features for the large language models, allowing them to retrieve segmentation mask proposals that can be used for grounding. We also present M3G2, a dataset of 1.9M training text-image pairs with 36 sub-problems derived from 27 existing datasets for visually grounded instruction tuning, facilitating precise vision-language alignment at the pixel level. We show that after training on M3G2, GROUNDHOG achieves superior performance on various grounding tasks. Through extensive case studies, we further show that GROUNDHOG unlocks explainability and diagnosability, and demonstrates better grounding towards occluded objects, groups of multiple instances, amorphous background regions, semantic parts of objects, and objects with irregular shapes. + +Acknowledgements This work was supported by Amazon and NSF IIS-1949634. We thank the anonymous reviewers for their valuable comments and suggestions. + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 8 +[2] Peter Anderson, Basura Fernando, Mark Johnson, and Stephen Gould. Spice: Semantic propositional image caption evaluation. In Proceedings of the 14th European Conference on Computer Vision, pages 382-398, 2016. 6 +[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 8 +[4] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 6 +[5] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocosstuff: Thing and stuff classes in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1209-1218, 2018. 2, 4, 1 +[6] Chongyan Chen, Samreen Anjum, and Danna Gurari. Grounding answers for visual questions asked by visually impaired people. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19098-19107, 2022. 5, 2 +[7] Chi Chen, Ruoyu Qin, Fuwen Luo, Xiaoyue Mi, Peng Li, Maosong Sun, and Yang Liu. Position-enhanced visual instruction tuning for multimodal large language models. arXiv preprint arXiv:2308.13437, 2023. 8 +[8] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 2, 3, 5, 6, 8 +[9] Ting Chen, Saurabh Saxena, Lala Li, David J Fleet, and Geoffrey Hinton. Pix2seq: A language modeling framework for object detection. In International Conference on Learning Representations, 2021. 3 +[10] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 2, 1, 4 +[11] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ * chatgpt quality. https://vicuna.lmsys.org, 2023. 4 +[12] Jifeng Dai, Kaiming He, and Jian Sun. Convolutional feature masking for joint object and stuff segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3992-4000, 2015. 2 +[13] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale + +Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500, 2023. 2, 8 +[14] Wenliang Dai, Zihan Liu, Ziwei Ji, Dan Su, and Pascale Fung. Plausible may not be faithful: Probing object hallucination in vision-language pre-training. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 2128-2140, 2023. 8 +[15] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. arXiv preprint arXiv:2309.16588, 2023. 5, 4 +[16] Daan de Geus, Panagiotis Meletis, Chenyang Lu, Xiaoxiao Wen, and Gijs Dubbelman. Part-aware panoptic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5485-5494, 2021. 4, 1 +[17] Harm De Vries, Florian Strub, Sarath Chandar, Olivier Pietquin, Hugo Larochelle, and Aaron Courville. Guess-what?! visual object discovery through multi-modal dialogue. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5503–5512, 2017. 5, 3 +[18] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 1 +[19] Chuang Gan, Yandong Li, Haoxiang Li, Chen Sun, and Boqing Gong. Vqs: Linking segmentations to questions and answers for supervised attention in vqa and question-focused semantic segmentation. In Proceedings of the IEEE international conference on computer vision, pages 1811-1820, 2017. 5, 2 +[20] Tao Gong, Chengqi Lyu, Shilong Zhang, Yudong Wang, Miao Zheng, Qian Zhao, Kuikun Liu, Wenwei Zhang, Ping Luo, and Kai Chen. Multimodal-gpt: A vision and language model for dialogue with humans. arXiv preprint arXiv:2305.04790, 2023. 8 +[21] Cristina González, Nicolás Ayobi, Isabela Hernández, José Hernández, Jordi Pont-Tuset, and Pablo Arbeláez. Panoptic narrative grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1364-1373, 2021. 6 +[22] Cristina González, Nicolás Ayobi, Isabela Hernández, Jordi Pont-Tuset, and Pablo Arbeláez. Piglet: Pixel-level grounding of language expressions with transformers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 6 +[23] Abel Gonzalez-Garcia, Davide Modolo, and Vittorio Ferrari. Objects as context for detecting their semantic parts. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6907-6916, 2018. 2 +[24] Herbert P Grice. Logic and conversation. In Speech acts, pages 41-58. Brill, 1975. 5 +[25] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 4, 5, 1, 3 + +[26] Kai Han, Yunhe Wang, Jianyuan Guo, Yehui Tang, and Enhua Wu. Vision gnn: An image is worth graph of nodes. Advances in Neural Information Processing Systems, 35:8291-8303, 2022. 2 +[27] Stevan Harnad. The symbol grounding problem. Physica D: Nonlinear Phenomena, 42(1-3):335-346, 1990. 8 +[28] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2021. 4 +[29] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 5, 2 +[30] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1780-1790, 2021. 4, 6 +[31] Osman Semih Kayhan, Bart Vredebregt, and Jan C Van Gemert. Hallucination in object detection—a study in visual part verification. In 2021 IEEE International Conference on Image Processing (ICIP), pages 2234–2238. IEEE, 2021. 8 +[32] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014. 5, 6, 2 +[33] Seyedalireza Khoshkirat and Chandra Kambhametty. Sentence attention blocks for answer grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6080-6090, 2023. 6 +[34] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dólar. Panoptic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9404-9413, 2019. 5, 6, 1 +[35] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4015-4026, 2023. 2, 3 +[36] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 5, 3 +[37] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692, 2023. 4, 5, 6, 8, 2 +[38] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726, 2023. 8 + +[39] Chunyuan Li, Zhe Gan, Zhengyuan Yang, Jianwei Yang, Linjie Li, Lijuan Wang, and Jianfeng Gao. Multimodal foundation models: From specialists to general-purpose assistants. arXiv preprint arXiv:2309.10020, 2023. 2 +[40] Jianshu Li, Jian Zhao, Yunchao Wei, Congyan Lang, Yidong Li, Terence Sim, Shuicheng Yan, and Jiashi Feng. Multiple-human parsing in the wild. arXiv preprint arXiv:1705.07206, 2017. 4, 1 +[41] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022. 8 +[42] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, 2023. 7, 8 +[43] Chang Liu, Henghui Ding, and Xudong Jiang. Gres: Generalized referring expression segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23592-23601, 2023. 5, 6, 8, 2, 3 +[44] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 5, 8, 4 +[45] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 2, 8 +[46] Jiang Liu, Hui Ding, Zhaowei Cai, Yuting Zhang, Ravi Kumar Satzoda, Vijay Mahadevan, and R Manmatha. Polyformer: Referring image segmentation as sequential polygon generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18653-18663, 2023. 6 +[47] Shangbang Long, Siyang Qin, Dmitry Panteleev, Alessandro Bissacco, Yasuhisa Fujii, and Michalis Raptis. Towards end-to-end unified scene text detection and layout analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1049-1059, 2022. 5, 3 +[48] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 4 +[49] Jiasen Lu, Christopher Clark, Rowan Zellers, Roozbeh Mottaghi, and Aniruddha Kembhavi. Unified-io: A unified model for vision, language, and multi-modal tasks. In The Eleventh International Conference on Learning Representations, 2022. 8 +[50] Ziqiao Ma, Jiayi Pan, and Joyce Chai. World-to-words: Grounded open vocabulary acquisition through fast mapping in vision-language models. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 524–544, 2023. 2, 8 +[51] Arjun Mani, Nobline Yoo, Will Hinthorn, and Olga Russakovsky. Point and ask: Incorporating pointing into visual question answering. arXiv preprint arXiv:2011.13681, 2020. 5, 6, 3 + +[52] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 5, 8, 3 +[53] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 5, 6, 2 +[54] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2 +[55] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 6 +[56] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 2, 3, 8 +[57] Renjie Pi, Jiahui Gao, Shizhe Diao, Rui Pan, Hanze Dong, Jipeng Zhang, Lewei Yao, Jianhua Han, Hang Xu, and Lingpeng Kong Tong Zhang. Detgpt: Detect what you need via reasoning. arXiv preprint arXiv:2305.14167, 2023. 8 +[58] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In Proceedings of the IEEE international conference on computer vision, pages 2641–2649, 2015. 5, 6, 1 +[59] Jordi Pont-Tuset, Jasper Uijlings, Soravit Changpinyo, Radu Soricut, and Vittorio Ferrari. Connecting vision and language with localized narratives. In Proceedings of the 16th European Conference on Computer Vision, pages 647-664, 2020. 1 +[60] Lu Qi, Jason Kuen, Tiancheng Shen, Jiuxiang Gu, Wenbo Li, Weidong Guo, Jiaya Jia, Zhe Lin, and Ming-Hsuan Yang. High quality entity segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4047-4056, 2023. 4, 5, 1, 3 +[61] Mengxue Qu, Yu Wu, Wu Liu, Xiaodan Liang, Jingkuan Song, Yao Zhao, and Yunchao Wei. RIO: A benchmark for reasoning intention-oriented objects in open environments. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 6, 2 +[62] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 5, 4 +[63] Vignesh Ramanathan, Anmol Kalia, Vladan Petrovic, Yi Wen, Baixue Zheng, Baishan Guo, Rui Wang, Aaron Mar + +quez, Rama Kovvuri, Abhishek Kadian, et al. Paco: Parts and attributes of common objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7141-7151, 2023. 4, 1 +[64] Varun Nagaraj Rao, Xingjian Zhen, Karen Hovsepian, and Mingwei Shen. A first look: Towards explainable textvqa models via visual and textual explanations. arXiv preprint arXiv:2105.02626, 2021. 5, 6, 2 +[65] Anna Rohrbach, Marcus Rohrbach, Ronghang Hu, Trevor Darrell, and Bernt Schiele. Grounding of textual phrases in images by reconstruction. In Proceedings of the 14th European Conference on Computer Vision, pages 817-834. Springer, 2016. 5, 2 +[66] Amanpreet Singh, Vivek Natarjan, Meet Shah, Yu Jiang, Xinlei Chen, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8317-8326, 2019. 4, 6 +[67] Amanpreet Singh, Guan Pang, Mandy Toh, Jing Huang, Wojciech Galuba, and Tal Hassner. TextOCR: Towards large-scale end-to-end reasoning for arbitrary-shaped scene text. 2021. 4, 1 +[68] Bharat Singh, Hengduo Li, Abhishek Sharma, and Larry S Davis. R-fcn-3000 at 30fps: Decoupling detection and classification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1081–1090, 2018. 2 +[69] Zhi Tian, Chunhua Shen, Xinlong Wang, and Hao Chen. Boxinst: High-performance instance segmentation with box annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5443-5452, 2021. 4 +[70] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 5, 4 +[71] Maria Tsimpoukelli, Jacob L Menick, Serkan Cabi, SM Eslami, Oriol Vinyls, and Felix Hill. Multimodal few-shot learning with frozen language models. Advances in Neural Information Processing Systems, 34:200-212, 2021. 8 +[72] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 6 +[73] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In International Conference on Machine Learning, pages 23318-23340. PMLR, 2022. 8 +[74] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, et al. Vision: Large language model is also an open-ended decoder for vision-centric tasks. arXiv preprint arXiv:2305.11175, 2023. 8 +[75] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clip + +driven referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11686-11695, 2022. 6 +[76] Chenyun Wu, Zhe Lin, Scott Cohen, Trung Bui, and Subhransu Maji. Phrasescut: Language-based image segmentation in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10216-10225, 2020. 4, 5, 6, 2 +[77] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.1 +[78] Yixuan Wu, Zhao Zhang, Chi Xie, Feng Zhu, and Rui Zhao. Advancing referring expression segmentation beyond single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2628-2638, 2023. 5, 2 +[79] Chi Xie, Zhao Zhang, Yixuan Wu, Feng Zhu, Rui Zhao, and Shuang Liang. Described object detection: Liberating object detection with flexible expressions. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 5, 2 +[80] Bin Yan, Yi Jiang, Jiannan Wu, Dong Wang, Ping Luo, Zehuan Yuan, and Huchuan Lu. Universal instance perception as object discovery and retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15325-15336, 2023. 6 +[81] Zhengyuan Yang, Zhe Gan, Jianfeng Wang, Xiaowei Hu, Faisal Ahmed, Zicheng Liu, Yumao Lu, and Lijuan Wang. Unitab: Unifying text and box outputs for grounded vision-language modeling. 2022. 8 +[82] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Levt: Language-aware vision transformer for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18155-18165, 2022. 6 +[83] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. MPLug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 8 +[84] Shukang Yin, Chaoyou Fu, Sirui Zhao, Ke Li, Xing Sun, Tong Xu, and Enhong Chen. A survey on multimodal large language models. arXiv preprint arXiv:2306.13549, 2023. 2 +[85] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704, 2023. 3, 6, 8 +[86] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. 4, 5, 8, 3 +[87] Yuhang Zang, Wei Li, Jun Han, Kaiyang Zhou, and Chen Change Loy. Contextual object detection with multimodal large language models. arXiv preprint arXiv:2305.18279, 2023. 8 + +[88] Rowan Zellers, Yonatan Bisk, Ali Farhadi, and Yejin Choi. From recognition to cognition: Visual commonsense reasoning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6720-6731, 2019. 5, 3 +[89] Shilong Zhang, Peize Sun, Shoufa Chen, Min Xiao, Wenqi Shao, Wenwei Zhang, Kai Chen, and Ping Luo. Gpt4roi: Instruction tuning large language model on region-of-interest. arXiv preprint arXiv:2307.03601, 2023. 8 +[90] Bo Zhao, Boya Wu, and Tiejun Huang. Svit: Scaling up visual instruction tuning. arXiv preprint arXiv:2307.04087, 2023.5,3 +[91] Yang Zhao, Zhijie Lin, Daquan Zhou, Zilong Huang, Jiashi Feng, and Bingyi Kang. Bubogpt: Enabling visual grounding in multi-modal llms. arXiv preprint arXiv:2307.08581, 2023. 8 +[92] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16793-16803, 2022. 2 +[93] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 8 +[94] Yuke Zhu, Oliver Groth, Michael Bernstein, and Li Fei-Fei. Visual7w: Grounded question answering in images. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4995–5004, 2016. 5, 6, 3 \ No newline at end of file diff --git a/2024/GROUNDHOG_ Grounding Large Language Models to Holistic Segmentation/images.zip b/2024/GROUNDHOG_ Grounding Large Language Models to Holistic Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7125cbc4f7717877bfee0b06638e2d424ebed079 --- /dev/null +++ b/2024/GROUNDHOG_ Grounding Large Language Models to Holistic Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5a68d29beb1993a397b803879ade5133a43516e3ff92344dca3f780ab34d8191 +size 770601 diff --git a/2024/GROUNDHOG_ Grounding Large Language Models to Holistic Segmentation/layout.json b/2024/GROUNDHOG_ Grounding Large Language Models to Holistic Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..288fcd437ca7ac2b6503fab6bf47b5219e2c8341 --- /dev/null +++ b/2024/GROUNDHOG_ Grounding Large Language Models to Holistic Segmentation/layout.json @@ -0,0 +1,9182 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 52, + 91, + 541, + 111 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 91, + 541, + 111 + ], + "spans": [ + { + "bbox": [ + 52, + 91, + 541, + 111 + ], + "type": "text", + "content": "GROUNDHOG: Grounding Large Language Models to Holistic Segmentation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 83, + 122, + 509, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 122, + 509, + 137 + ], + "spans": [ + { + "bbox": [ + 83, + 122, + 509, + 137 + ], + "type": "text", + "content": "Yichi Zhang" + }, + { + "bbox": [ + 83, + 122, + 509, + 137 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 83, + 122, + 509, + 137 + ], + "type": "text", + "content": ", Ziqiao Ma" + }, + { + "bbox": [ + 83, + 122, + 509, + 137 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 83, + 122, + 509, + 137 + ], + "type": "text", + "content": ", Xiaofeng Gao" + }, + { + "bbox": [ + 83, + 122, + 509, + 137 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 83, + 122, + 509, + 137 + ], + "type": "text", + "content": ", Suhaila Shakiah" + }, + { + "bbox": [ + 83, + 122, + 509, + 137 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 83, + 122, + 509, + 137 + ], + "type": "text", + "content": ", Qiaozi Gao" + }, + { + "bbox": [ + 83, + 122, + 509, + 137 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 83, + 122, + 509, + 137 + ], + "type": "text", + "content": ", Joyce Chai" + }, + { + "bbox": [ + 83, + 122, + 509, + 137 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 201, + 137, + 394, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 137, + 394, + 151 + ], + "spans": [ + { + "bbox": [ + 201, + 137, + 394, + 151 + ], + "type": "text", + "content": "1University of Michigan, 2Amazon AGI" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 231, + 152, + 362, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 152, + 362, + 165 + ], + "spans": [ + { + "bbox": [ + 231, + 152, + 362, + 165 + ], + "type": "text", + "content": "zhangyic@umich.edu" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 176, + 167, + 414, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 167, + 414, + 179 + ], + "spans": [ + { + "bbox": [ + 176, + 167, + 414, + 179 + ], + "type": "text", + "content": "https://groundhog-ml1m.github.io/" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 51, + 186, + 173, + 279 + ], + "blocks": [ + { + "bbox": [ + 51, + 186, + 173, + 279 + ], + "lines": [ + { + "bbox": [ + 51, + 186, + 173, + 279 + ], + "spans": [ + { + "bbox": [ + 51, + 186, + 173, + 279 + ], + "type": "image", + "image_path": "817673d7b306725870a94ad6aeeaef6db83f1b4c43b730771fabd7dec2fffb7c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 52, + 281, + 64, + 297 + ], + "blocks": [ + { + "bbox": [ + 52, + 281, + 64, + 297 + ], + "lines": [ + { + "bbox": [ + 52, + 281, + 64, + 297 + ], + "spans": [ + { + "bbox": [ + 52, + 281, + 64, + 297 + ], + "type": "image", + "image_path": "54bd5a7be146835c9c4ce0f5c2ba8ebb098ebef2ac63383612da255f4fbb58c2.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 77, + 284, + 167, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 284, + 167, + 294 + ], + "spans": [ + { + "bbox": [ + 77, + 284, + 167, + 294 + ], + "type": "text", + "content": "Describe the image briefly." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 301, + 141, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 301, + 141, + 320 + ], + "spans": [ + { + "bbox": [ + 55, + 301, + 141, + 320 + ], + "type": "text", + "content": "A man and a little girl are sitting in a shopping cart." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 158, + 306, + 174, + 322 + ], + "blocks": [ + { + "bbox": [ + 158, + 306, + 174, + 322 + ], + "lines": [ + { + "bbox": [ + 158, + 306, + 174, + 322 + ], + "spans": [ + { + "bbox": [ + 158, + 306, + 174, + 322 + ], + "type": "image", + "image_path": "ccbb31425917613174b95d2ad4b065a5c65a3dbc814f2029381d7d512e7f56c9.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 176, + 187, + 297, + 279 + ], + "blocks": [ + { + "bbox": [ + 176, + 187, + 297, + 279 + ], + "lines": [ + { + "bbox": [ + 176, + 187, + 297, + 279 + ], + "spans": [ + { + "bbox": [ + 176, + 187, + 297, + 279 + ], + "type": "image", + "image_path": "4a1b2c5bb92f244b2bcbd940a39ec34b08e3c2be425deb32a6c218fe0fce444a.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 176, + 281, + 188, + 297 + ], + "blocks": [ + { + "bbox": [ + 176, + 281, + 188, + 297 + ], + "lines": [ + { + "bbox": [ + 176, + 281, + 188, + 297 + ], + "spans": [ + { + "bbox": [ + 176, + 281, + 188, + 297 + ], + "type": "image", + "image_path": "4c68b29cd239bf69488cce1fd1020619a1d355e9b22234fbd3d25a5795fbf0f2.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 201, + 284, + 286, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 284, + 286, + 294 + ], + "spans": [ + { + "bbox": [ + 201, + 284, + 286, + 294 + ], + "type": "text", + "content": "Generate a short caption." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 179, + 300, + 268, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 300, + 268, + 320 + ], + "spans": [ + { + "bbox": [ + 179, + 300, + 268, + 320 + ], + "type": "text", + "content": "Two dogs are playing with a stick in a field." + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 297, + 187, + 421, + 322 + ], + "blocks": [ + { + "bbox": [ + 297, + 187, + 421, + 322 + ], + "lines": [ + { + "bbox": [ + 297, + 187, + 421, + 322 + ], + "spans": [ + { + "bbox": [ + 297, + 187, + 421, + 322 + ], + "type": "image", + "image_path": "5429eeab0ac8ed394c150f376deea831fafd34399aeb88b0b3d9dd27b40df257.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 423, + 213, + 436, + 229 + ], + "blocks": [ + { + "bbox": [ + 423, + 213, + 436, + 229 + ], + "lines": [ + { + "bbox": [ + 423, + 213, + 436, + 229 + ], + "spans": [ + { + "bbox": [ + 423, + 213, + 436, + 229 + ], + "type": "image", + "image_path": "e322558dc6e91ba19a2b8c04497025b653f07fbde7cde1448ef382f777bef8c9.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 447, + 209, + 520, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 447, + 209, + 520, + 228 + ], + "spans": [ + { + "bbox": [ + 447, + 209, + 520, + 228 + ], + "type": "text", + "content": "Describe the given picture in very detail." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 425, + 245, + 517, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 425, + 245, + 517, + 309 + ], + "spans": [ + { + "bbox": [ + 425, + 245, + 517, + 309 + ], + "type": "text", + "content": "In this image, we can see a boat on the water. There are few people inside the boat. There are also few people on the rocks. In the background, there are trees, hills, and sky." + } + ] + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 527, + 300, + 541, + 315 + ], + "blocks": [ + { + "bbox": [ + 527, + 300, + 541, + 315 + ], + "lines": [ + { + "bbox": [ + 527, + 300, + 541, + 315 + ], + "spans": [ + { + "bbox": [ + 527, + 300, + 541, + 315 + ], + "type": "image", + "image_path": "02a2638a71e499bc68c7a9ef19695b53f19802294ce4a234f52ee779038b4446.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 52, + 335, + 225, + 460 + ], + "blocks": [ + { + "bbox": [ + 52, + 335, + 225, + 460 + ], + "lines": [ + { + "bbox": [ + 52, + 335, + 225, + 460 + ], + "spans": [ + { + "bbox": [ + 52, + 335, + 225, + 460 + ], + "type": "image", + "image_path": "5f446983be845abc58bf5800a0a7fa28bc511843f6a4d7208223e581cd2d0421.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 52, + 462, + 64, + 485 + ], + "blocks": [ + { + "bbox": [ + 52, + 462, + 64, + 485 + ], + "lines": [ + { + "bbox": [ + 52, + 462, + 64, + 485 + ], + "spans": [ + { + "bbox": [ + 52, + 462, + 64, + 485 + ], + "type": "image", + "image_path": "d18ad8ab055918dc33ea563474a1b189b9950d7a8ad89f0170ffd432ceefe8c6.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 514, + 545, + 536 + ], + "lines": [ + { + "bbox": [ + 46, + 514, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 46, + 514, + 545, + 536 + ], + "type": "text", + "content": "Figure 1. We propose GROUNDHOG, a multimodal large language model that enhances its text output with pixel-level phrase grounding across diverse semantic granularities. The figure demonstrates outputs from our model on the four task types we considered in this work." + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "bbox": [ + 77, + 464, + 213, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 464, + 213, + 482 + ], + "spans": [ + { + "bbox": [ + 77, + 464, + 213, + 482 + ], + "type": "text", + "content": "Could you please segment out \"laptop barely onscreen not apple\" in the image?" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 56, + 489, + 170, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 170, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 170, + 498 + ], + "type": "text", + "content": "Laptop barely onscreen not apple." + } + ] + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 210, + 486, + 224, + 500 + ], + "blocks": [ + { + "bbox": [ + 210, + 486, + 224, + 500 + ], + "lines": [ + { + "bbox": [ + 210, + 486, + 224, + 500 + ], + "spans": [ + { + "bbox": [ + 210, + 486, + 224, + 500 + ], + "type": "image", + "image_path": "d08af1999556258e367d3ae3b17b97e5b763df55800993414ee9d20b69ac081d.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 248, + 335, + 373, + 458 + ], + "blocks": [ + { + "bbox": [ + 231, + 323, + 362, + 334 + ], + "lines": [ + { + "bbox": [ + 231, + 323, + 362, + 334 + ], + "spans": [ + { + "bbox": [ + 231, + 323, + 362, + 334 + ], + "type": "text", + "content": "(a) Grounded Image Captioning (GIC)." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 248, + 335, + 373, + 458 + ], + "lines": [ + { + "bbox": [ + 248, + 335, + 373, + 458 + ], + "spans": [ + { + "bbox": [ + 248, + 335, + 373, + 458 + ], + "type": "image", + "image_path": "17ff19cbf741628e010d8814aee4f465abf5a21a5d15d5ba303fca2f1315db21.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 248, + 464, + 260, + 485 + ], + "blocks": [ + { + "bbox": [ + 248, + 464, + 260, + 485 + ], + "lines": [ + { + "bbox": [ + 248, + 464, + 260, + 485 + ], + "spans": [ + { + "bbox": [ + 248, + 464, + 260, + 485 + ], + "type": "image", + "image_path": "0c37a129018dcdce8c406b75452fcca98cc9a672fc1c9bff5644eb8b01b36352.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "bbox": [ + 268, + 464, + 358, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 464, + 358, + 481 + ], + "spans": [ + { + "bbox": [ + 268, + 464, + 358, + 481 + ], + "type": "text", + "content": "What date is shown on the calendar?" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 247, + 489, + 269, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 489, + 269, + 498 + ], + "spans": [ + { + "bbox": [ + 247, + 489, + 269, + 498 + ], + "type": "text", + "content": "Feb 2." + } + ] + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 361, + 487, + 376, + 501 + ], + "blocks": [ + { + "bbox": [ + 361, + 487, + 376, + 501 + ], + "lines": [ + { + "bbox": [ + 361, + 487, + 376, + 501 + ], + "spans": [ + { + "bbox": [ + 361, + 487, + 376, + 501 + ], + "type": "image", + "image_path": "54b342a5cfb2ddaa4077cad87215fd00418b68af00a4e8c4bf7dcc06fbd442c5.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 394, + 335, + 542, + 459 + ], + "blocks": [ + { + "bbox": [ + 394, + 335, + 542, + 459 + ], + "lines": [ + { + "bbox": [ + 394, + 335, + 542, + 459 + ], + "spans": [ + { + "bbox": [ + 394, + 335, + 542, + 459 + ], + "type": "image", + "image_path": "cc47e31e680bef21ebce829533cb6b09ba7c79499bd4f0a8c6a546ba1a97f8b5.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 406, + 463, + 419, + 484 + ], + "blocks": [ + { + "bbox": [ + 406, + 463, + 419, + 484 + ], + "lines": [ + { + "bbox": [ + 406, + 463, + 419, + 484 + ], + "spans": [ + { + "bbox": [ + 406, + 463, + 419, + 484 + ], + "type": "image", + "image_path": "ad2fbb69656e9111eb7e474fea7b1d498f28be11097c5c19d527dbf7c82bc28a.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 416, + 503, + 518, + 513 + ], + "lines": [ + { + "bbox": [ + 416, + 503, + 518, + 513 + ], + "spans": [ + { + "bbox": [ + 416, + 503, + 518, + 513 + ], + "type": "text", + "content": "(d) Referential Dialogue (RD)." + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "bbox": [ + 424, + 463, + 534, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 424, + 463, + 534, + 481 + ], + "spans": [ + { + "bbox": [ + 424, + 463, + 534, + 481 + ], + "type": "text", + "content": "Provide a distinct description for that " + }, + { + "bbox": [ + 424, + 463, + 534, + 481 + ], + "type": "inline_equation", + "content": "< \\text{PTR}>" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 408, + 489, + 500, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 489, + 500, + 499 + ], + "spans": [ + { + "bbox": [ + 408, + 489, + 500, + 499 + ], + "type": "text", + "content": "Blue container with apples." + } + ] + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 526, + 486, + 541, + 500 + ], + "blocks": [ + { + "bbox": [ + 526, + 486, + 541, + 500 + ], + "lines": [ + { + "bbox": [ + 526, + 486, + 541, + 500 + ], + "spans": [ + { + "bbox": [ + 526, + 486, + 541, + 500 + ], + "type": "image", + "image_path": "9c1d57fcf0661fc141f0ebe98ac24998aabf81883db00c4a04f8ac274d27dea9.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "bbox": [ + 143, + 544, + 191, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 544, + 191, + 555 + ], + "spans": [ + { + "bbox": [ + 143, + 544, + 191, + 555 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 46, + 559, + 288, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 559, + 288, + 691 + ], + "spans": [ + { + "bbox": [ + 46, + 559, + 288, + 691 + ], + "type": "text", + "content": "Most multimodal large language models (MLLMs) learn language-to-object grounding through causal language modeling where grounded objects are captured by bounding boxes as sequences of location tokens. This paradigm lacks pixel-level representations that are important for fine-grained visual understanding and diagnosis. In this work, we introduce GROUNDHOG, an MLLM developed by grounding Large Language Models to holistic segmentation. GROUNDHOG incorporates a masked feature extractor and converts extracted features into visual entity tokens for the MLLM backbone, which then con" + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 304, + 545, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 689 + ], + "type": "text", + "content": "nects groundable phrases to unified grounding masks by retrieving and merging the entity masks. To train GROUNDHOG, we carefully curated M3G2, a grounded visual instruction tuning dataset with Multi-Modal Multi-Grained Grounding, by harvesting a collection of segmentation-grounded datasets with rich annotations. Our experimental results show that GROUNDHOG achieves superior performance on various language grounding tasks without task-specific fine-tuning, and significantly reduces object hallucination. GROUNDHOG also demonstrates better grounding towards complex forms of visual input and provides easy-to-understand diagnosis in failure cases." + } + ] + } + ], + "index": 43 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 15 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 15 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 15 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 199, + 13, + 442, + 24 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 13, + 442, + 24 + ], + "spans": [ + { + "bbox": [ + 199, + 13, + 442, + 24 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 25, + 461, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 25, + 461, + 37 + ], + "spans": [ + { + "bbox": [ + 179, + 25, + 461, + 37 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 702, + 212, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 212, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 212, + 712 + ], + "type": "text", + "content": "† Work done during internship at Amazon AGI." + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14227" + } + ] + } + ], + "index": 45 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 49, + 72, + 127, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 72, + 127, + 83 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 127, + 83 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 91, + 286, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 91, + 286, + 317 + ], + "spans": [ + { + "bbox": [ + 49, + 91, + 286, + 317 + ], + "type": "text", + "content": "Multimodal large language models (MLLMs) have received an increasing amount of attention to address tasks that necessitate non-linguistic knowledge, e.g., perception and reasoning about the visual world [39, 84]. For fine-grained visual understanding, grounded MLLMs often learn language-to-object grounding by causal language modeling, where grounded objects are captured by bounding boxes as sequences of location tokens. However, bounding boxes are insufficient in indicating amorphous stuff [5], semantic parts of objects [23], finer-grained regions with irregular shapes [26], or groups of instances at the same time. As a result, a single bounding box can often include other irrelevant semantics in order to engulf the target entities, leading to ambiguity in detection. In addition, the generated box coordinate lacks interpretability. When the model hallucinates, such as incorrectly predicting the association between objects and language, it is hard to diagnose whether the problem is due to the model's failure to detect the object, or its incorrect alignment of the object with language." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 319, + 286, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 319, + 286, + 592 + ], + "spans": [ + { + "bbox": [ + 49, + 319, + 286, + 592 + ], + "type": "text", + "content": "To address these issues, in this work, we introduce GROUNDHOG, an MLLM developed by grounding Large Language Models to holistic segmentation. Our goal of language grounding is to connect text spans that refer to or can be deduced from visual information, termed as groundable phrases [50], to their corresponding regions of visual entities (Figure 1). GROUNDHOG incorporates a masked feature extractor that takes an input image and a set of class-agnostic entity mask proposals, and converts each mask's features into visual entity tokens for an MLLM backbone. This MLLM then connects groundable phrases to unified grounding masks by retrieving and merging the entity masks. Compared to previous grounded MLLMs, GROUNDHOG unlocks unprecedented pixel-level vision-language alignment. It naturally supports visual pointers as input, and can plug-in-and-play with any choice of mask proposal networks, e.g., Segment Anything Model (SAM) [35], domain-specific semantic segmentation models, or user-provided mask candidates. We introduce an enhanced Mask2Former [10] as our default mask proposal network, which detects regions at multiple granularities, e.g., instances (things and stuff), semantic parts, and visual text, leading to a holistic coverage of visual semantics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 594, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 594, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 594, + 286, + 712 + ], + "type": "text", + "content": "To train GROUNDHOG, we curated a Multi-Modal Multi-Grained Grounding (M3G2) dataset consisting of 2.5M text-image pairs for visually grounded instruction tuning, consisting of 36 sub-problems derived and augmented from 27 existing datasets. We present extensive experiments on vision-language tasks that require grounding, including grounded language generation with minimal object hallucination, language-guided segmentation, visual question answering with answer grounding, and referential dialog with spatial pointer inputs (Figure 1). Our empirical results show" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 308, + 73, + 544, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 73, + 544, + 132 + ], + "spans": [ + { + "bbox": [ + 308, + 73, + 544, + 132 + ], + "type": "text", + "content": "that GROUNDHOG, without task-specific fine-tuning, can achieve superior or comparable performance with previous models that either require fine-tuning or are specialized only for that dataset. In addition, GROUNDHOG has supports easy-to-understand diagnosis when grounding fails." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 308, + 143, + 463, + 154 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 143, + 463, + 154 + ], + "spans": [ + { + "bbox": [ + 308, + 143, + 463, + 154 + ], + "type": "text", + "content": "2. Our Method: GROUNDHOG" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 308, + 163, + 544, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 544, + 496 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 544, + 496 + ], + "type": "text", + "content": "The language grounding task can be succinctly delineated into two fundamental components: localization and recognition, as established in the literature [50, 68, 92]. Such categorization not only aids in the identification of object presence (objectness) without reliance on specific object classes, but also sets the stage for models to be robust in open-vocabulary settings. Building upon this framework, we formulate the grounding process as an entity segment selection problem, which involves (1) proposing entity segmentation masks where the masks encapsulate regions with discernible semantic content, and (2) recognizing the retrieved entities through the understanding of both visual and language context. Concurrently performing both tasks is where MLLMs bring a distinct advantage. This decoupled design of entity mask proposal and language-guided grounding brings several advantages. First, it allows independent improvement of the mask proposal model and MLLM, where specialized data, training, and inference setups can be applied. Second, by decoupling language grounding, it becomes straightforward to determine if a failure is due to the model's inability to propose the entity segment, or its misalignment of the object with the language, thus improving the interpretability of the whole framework. Third, as shown later, when connecting the two parts to work in tandem in a model-independent manner, the MLLM can benefit from multiple different vision specialist models in a plug-and-play fashion. In the remainder of this section, we give details of our model design." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 308, + 504, + 501, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 501, + 515 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 501, + 515 + ], + "type": "text", + "content": "2.1. Building Entity Features from Masks" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 308, + 522, + 544, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 522, + 544, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 522, + 544, + 712 + ], + "type": "text", + "content": "Our approach assumes the availability of a mask proposal model, which is capable of generating a set of class-agnostic entity masks from an image with high coverage. In contrast to prior studies that relied on low-level features [8, 13, 45, 56], GROUNDHOG interprets the image as a collection of entities. The primary challenge then becomes the derivation of effective visual features to accurately represent these entities. To achieve a complete decoupling of the MLLM from the mask proposal model responsible for providing the masks, we propose to condition the entity features solely on the binary masks without using any embeddings from the mask proposal model. Specifically, the mask corresponding to each entity is employed to extract patch features from pretrained vision foundation models, such as CLIP [62] and DINOv2 [54], through a convolutional mask pooling layer [12]. Given that the feature map dimensions" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "type": "text", + "content": "14228" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 72, + 541, + 270 + ], + "blocks": [ + { + "bbox": [ + 56, + 72, + 541, + 270 + ], + "lines": [ + { + "bbox": [ + 56, + 72, + 541, + 270 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 541, + 270 + ], + "type": "image", + "image_path": "371348875dd9fbb95ccac0fca9a5c92056d4e70990a409abfe0486c28873d11c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 274, + 547, + 319 + ], + "lines": [ + { + "bbox": [ + 46, + 274, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 547, + 319 + ], + "type": "text", + "content": "Figure 2. The model architecture of GROUNDHOG model. Given a set of class-agnostic entity mask proposals, the masked feature extractor first extracts the feature of each entity as the visual input of the multi-modal large language model (left). The output hidden states of the grounding tokens are averaged and used to retrieve the entities to ground, which will be merged into a single grounding mask for the phrase. Modules are colored by their trainability: parameter-free operators (grey), frozen (blue), trainable (orange), and partially trainable (mix)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 327, + 287, + 425 + ], + "blocks": [ + { + "bbox": [ + 50, + 327, + 287, + 425 + ], + "lines": [ + { + "bbox": [ + 50, + 327, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 50, + 327, + 287, + 425 + ], + "type": "image", + "image_path": "d4774ca35a652f1e2e464c2a4831bcc83771df0a366352dfd27e92fd20d61e16.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 427, + 288, + 472 + ], + "lines": [ + { + "bbox": [ + 46, + 427, + 288, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 427, + 288, + 472 + ], + "type": "text", + "content": "Figure 3. GROUNDHOG can take arbitrary spatial prompts that can be resolved by an interactive segmentation model, such as SAM. The placeholder pointer token " + }, + { + "bbox": [ + 46, + 427, + 288, + 472 + ], + "type": "inline_equation", + "content": "<\\mathrm{PTR}>" + }, + { + "bbox": [ + 46, + 427, + 288, + 472 + ], + "type": "text", + "content": " will be replaced by the extracted entity features and fed as input to the model." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 479, + 287, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 287, + 574 + ], + "type": "text", + "content": "are usually smaller than those of the mask proposals, we resize the masks to match the size of the feature maps prior to pooling. The pooled features are then fed into a Multi-Layer Perceptron (MLP) network to align with the input embeddings of the MLLM. We empirically find the combination of CLIP and DINOv2 features yields the best result, and these features are added to obtain the final input visual entity tokens to the MLLM." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": "Spatial Prompts Furthermore, for grounded MLLMs to be more broadly applicable, they must be capable of interpreting multi-modal user inputs, including spatial prompts. Thanks to the mask model agnostic design, GROUND-HOGcan seamlessly support such inputs. As demonstrated in Figure 3, by applying an interactive segmentation model such as Segment-Anything (SAM) [35], arbitrary spatial prompts can be translated into binary masks and processed by the same masked feature extractor we just introduced. This extracted feature for the pointed entity will replace the pointer token placeholder in the textual input." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 327, + 539, + 340 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 327, + 539, + 340 + ], + "spans": [ + { + "bbox": [ + 306, + 327, + 539, + 340 + ], + "type": "text", + "content": "2.2. Language Grounding to Entity Segmentation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 346, + 545, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 346, + 545, + 585 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 545, + 585 + ], + "type": "text", + "content": "Existing box-grounded MLLMs typically append location tokens after the groundable phrases [8, 9, 56, 85]. However, this method is not readily interpretable. To alleviate this disconnect, we introduce a pair of grounding tokens " + }, + { + "bbox": [ + 304, + 346, + 545, + 585 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{GRD} \\rangle" + }, + { + "bbox": [ + 304, + 346, + 545, + 585 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 346, + 545, + 585 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{/GRD} \\rangle" + }, + { + "bbox": [ + 304, + 346, + 545, + 585 + ], + "type": "text", + "content": " to indicate the start and end of groundable phrases, with the assumption that grounding these phrases requires mapping to certain representations of visual entities irrespective of the visual modality. In Figure 2, a sentence can be represented as I see " + }, + { + "bbox": [ + 304, + 346, + 545, + 585 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{GRD} \\rangle" + }, + { + "bbox": [ + 304, + 346, + 545, + 585 + ], + "type": "text", + "content": " two dogs " + }, + { + "bbox": [ + 304, + 346, + 545, + 585 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{/GRD} \\rangle" + }, + { + "bbox": [ + 304, + 346, + 545, + 585 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 304, + 346, + 545, + 585 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{GRD} \\rangle" + }, + { + "bbox": [ + 304, + 346, + 545, + 585 + ], + "type": "text", + "content": " the beach " + }, + { + "bbox": [ + 304, + 346, + 545, + 585 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{/GRD} \\rangle" + }, + { + "bbox": [ + 304, + 346, + 545, + 585 + ], + "type": "text", + "content": ", with two distinct visual entities grounded. The representation of each groundable phrase, termed as the grounding query, is obtained by adding " + }, + { + "bbox": [ + 304, + 346, + 545, + 585 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{GRD} \\rangle" + }, + { + "bbox": [ + 304, + 346, + 545, + 585 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 346, + 545, + 585 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{/GRD} \\rangle" + }, + { + "bbox": [ + 304, + 346, + 545, + 585 + ], + "type": "text", + "content": "'s output embedding from the last transformer layer of the MLLM. The representation is then used to retrieve the entities that the phrase should be grounded to. In particular, we concatenate the grounding query with the last layer output of each visual entity token, and use an MLP to predict a scalar score for each entity. Finally, we merge all the mask proposals into one single mask with pixel-wise maximization:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 362, + 592, + 485, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 592, + 485, + 613 + ], + "spans": [ + { + "bbox": [ + 362, + 592, + 485, + 613 + ], + "type": "interline_equation", + "content": "\\mathcal {M} _ {h, w} = \\max _ {q} \\left(\\mathcal {S} _ {q} \\cdot \\widehat {\\mathcal {M}} _ {q, h, w}\\right)", + "image_path": "69215c7f69a2410545d84f82cbb81c8cf0baab1155b5876dd2c4939576adb751.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 616, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 616, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 616, + 545, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 616, + 545, + 715 + ], + "type": "inline_equation", + "content": "S_{q}" + }, + { + "bbox": [ + 304, + 616, + 545, + 715 + ], + "type": "text", + "content": " is the normalized score of the " + }, + { + "bbox": [ + 304, + 616, + 545, + 715 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 304, + 616, + 545, + 715 + ], + "type": "text", + "content": "-th mask ranging from 0 to 1, and " + }, + { + "bbox": [ + 304, + 616, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathcal{M}}_{q,h,w}" + }, + { + "bbox": [ + 304, + 616, + 545, + 715 + ], + "type": "text", + "content": " denotes the pixel probability at position " + }, + { + "bbox": [ + 304, + 616, + 545, + 715 + ], + "type": "inline_equation", + "content": "(h,w)" + }, + { + "bbox": [ + 304, + 616, + 545, + 715 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 304, + 616, + 545, + 715 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 304, + 616, + 545, + 715 + ], + "type": "text", + "content": "-th mask. Note that a phrase may ground to multiple entities, thus multiple mask proposals may get a high score simultaneously and be selected in conjunction. One of the primary benefits of this decoupled design is its transparency in the selection of entities. Users can easily visualize both the mask proposals and their respective" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14229" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "type": "text", + "content": "scores, providing a clear understanding of how a grounding mask is predicted. This level of clarity and interpretability is a significant advantage, offering users a tangible insight into the model's grounding process." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 131, + 258, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 131, + 258, + 144 + ], + "spans": [ + { + "bbox": [ + 47, + 131, + 258, + 144 + ], + "type": "text", + "content": "2.3. Towards Holistic Entity Mask Proposals" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 149, + 287, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 149, + 287, + 400 + ], + "spans": [ + { + "bbox": [ + 46, + 149, + 287, + 400 + ], + "type": "text", + "content": "In order to support holistic language grounding to arbitrary segmentations, the entity proposal should have two essential properties. First, the proposals should strike a delicate balance in terms of semantic atomicity. While it is possible to merge multiple proposals later to form multi-entity segmentations, the reverse, i.e., dividing a single proposal into smaller segments, is not feasible. Therefore, instance segmentation is generally preferred over semantic segmentation. However, the segmentation should not be excessively fine-grained to the extent that it compromises basic semantic integrity. Over-segmentation can lead to a loss of the coherent concept of an entity, which is detrimental to the grounding process. Second, the entity proposals should have a high coverage of entities, encompassing a diverse range of granularities. This includes not only tangible objects (things) and amorphous concepts (stuff) but also extends to sub-components of objects (parts of things) and structured regions such as areas containing visual text. The ability to propose entities across this spectrum of granularity is pivotal, as it directly determines the upper bound of the grounding capability of MLLM." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 402, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 288, + 714 + ], + "type": "text", + "content": "We initiated our study with a Mask2Former model pretrained on the COCO panoptic segmentation dataset. However, preliminary experiments revealed its limitations in semantic coverage and adaptability to open-world scenarios. To enhance this, we developed Mask2Former+, an upgraded version designed for multi-grained segmentation. This upgrade involved creating a diverse dataset by merging annotations from various sources, including COCO [5], LVIS [25], Entity-v2 [60], Pascal [16], PACO [63] (Figure 7); MHP-v2 [40] for human part parsing; and TextOCR [67] for text segmentation. We expanded the model's capabilities by adding 50 expert queries each for semantic parts and visual text regions, alongside the original 200 entity queries. We assessed Mask2Former+’s performance on 1000 images from validation splits from 4 grounding benchmarks, RefCOCO+ [86], PhraseCut [76], ReasonSeg [37], and TextVQA-X [66]. We use the Any-IoU [30] metric for evaluation, i.e., for each ground truth mask, we extract the most overlapped mask proposals and compute the IoU, then take the average. As Table 1 demonstrates, Mask2Former+ shows consistent improvements across all domains, particularly in those significantly divergent from COCO. This highlights its enhanced adaptability and precision in a broader range of segmentation challenges, providing a good mask proposal model for GROUNDHOG. We refer to Appendix A for more details of the model and data." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 309, + 70, + 545, + 106 + ], + "blocks": [ + { + "bbox": [ + 309, + 70, + 545, + 106 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 545, + 106 + ], + "type": "table", + "html": "
ModelRefCOCO+PhraseCutReasonSegTextVQA-X
Mask2Former0.8670.5630.6020.137
Mask2Former+0.8730.6240.7450.446
", + "image_path": "a5cb2f90c4666ea955fcb8d1d0e48086351920939ed6aa074b34b395429ad0da.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 310, + 170, + 544, + 494 + ], + "blocks": [ + { + "bbox": [ + 305, + 111, + 545, + 166 + ], + "lines": [ + { + "bbox": [ + 305, + 111, + 545, + 166 + ], + "spans": [ + { + "bbox": [ + 305, + 111, + 545, + 166 + ], + "type": "text", + "content": "Table 1. The average Any-IoU of the proposals on each dataset. The vanilla Mask2Former is trained on the COCO-Panoptic dataset and our Mask2Former+ is trained on our combined dataset. Mask2Former+ obtains a consistent improvement in all scenarios, especially in non-COCO domains." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 170, + 544, + 494 + ], + "lines": [ + { + "bbox": [ + 310, + 170, + 544, + 494 + ], + "spans": [ + { + "bbox": [ + 310, + 170, + 544, + 494 + ], + "type": "table", + "html": "
TaskDatasetGr. Ann.Sem. Gran.# Pairs
MBPoSThPaGTx
GCAPPNG✓✓✓✓✓132k
Flickr30K-Entity✓✓✓✓✓149k
RESRefCOCO✓✓✓113k
RefCOCO+✓✓✓112k
RefCOCOg✓✓✓80k
RefCLEF✓✓✓105k
gRefCOCO✓✓✓194k
PhraseCut✓✓✓✓✓✓85k
D-Cube✓✓✓✓10k
ReasonSeg✓✓✓✓✓✓1k
RIO✓✓✓✓28k
SK-VG✓✓23k
GVQAVizWiz-G✓✓✓✓✓6k
TextVQA-X✓✓✓15k
GQA✓✓✓✓✓302k
VQS✓✓20k
Shikra-BinaryQA✓✓✓✓✓4k
EntityCount✓✓✓✓✓✓11k
FoodSeg-QA✓✓✓✓7k
LVIS-QA✓✓✓✓✓95k
RDRefCOCO-REG✓✓✓✓17k
RefCOCO+-REG✓✓✓✓17k
RefCOCOg-REG✓✓✓✓22k
gRefCOCO-REG✓✓✓✓20k
VG-SpotCap✓✓✓✓✓✓247k
V7W✓✓✓23k
PointQA✓✓64k
VCR✓✓✓156k
ShikraRD✓✓✓✓✓✓2k
SVIT-RD✓✓✓✓✓✓33k
Guesswhat✓✓✓✓193k
VG-RefMatch✓✓✓✓✓✓247k
HierText✓✓✓✓6k
M3G2 (Total)2.5M
", + "image_path": "e694e0b942828c5896ff70b6b9ec998b90cfa069724e8a877289d4f42f8ea706.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 498, + 545, + 543 + ], + "lines": [ + { + "bbox": [ + 305, + 498, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 305, + 498, + 545, + 543 + ], + "type": "text", + "content": "Table 2. Summary of datasets included in M3G2. We show the availability of Grounding Annotations (Box, Mask, and Pointer inputs), the Semantic Granularity (Stuff, Things, Parts, Groups, and Text), and the number of text-image pairs for training." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 554, + 424, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 554, + 424, + 566 + ], + "spans": [ + { + "bbox": [ + 306, + 554, + 424, + 566 + ], + "type": "text", + "content": "3. Our Dataset: M3G2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "In this section, we introduce M3G2, a Multi-Modal MultiGrained Grounding dataset consisting of 2.5M text-image pairs for visually grounded instruction tuning, consisting of 36 sub-problems derived and augmented from 27 existing datasets. We re-organize and augment public datasets of language grounding, visual question answering, referring expression segmentation, and referring expression generation into various forms of visually grounded dialogue for grounded instruction tuning, outlined briefly in Table 2. The dataset is categorized into four main types: (1) Grounded Image Captioning (GIC), (2) Referential Expression Segmentation (RES), (3) Grounded Visual Question Answer" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14230" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 133 + ], + "type": "text", + "content": "ing (GVQA), and (4) Referential Dialog (RD). We provide illustrated descriptions of our prompt design, accompanied by examples of each task type as depicted in Figure 4. We detail the task schema in the following sections and provide the complete sets of templates in Appendix B." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 140, + 236, + 154 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 140, + 236, + 154 + ], + "spans": [ + { + "bbox": [ + 47, + 140, + 236, + 154 + ], + "type": "text", + "content": "3.1. Grounded Image Captioning (GIC)" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 159, + 289, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 159, + 289, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 159, + 289, + 304 + ], + "type": "text", + "content": "The task of grounded image captioning requires the model to produce a narrative for the visual scene, and accurately identify and associate the groundable phrases with their respective binary segmentation masks. The objective of this task is to empower the model to articulate the scene while acknowledging various visual elements and their spatial interrelations. We incorporate the Panoptic Narrative Grounding (PNG) dataset [34] for dense and detailed scene descriptions, as well as the Flickr30K-Entity dataset [58] for concise descriptions of the salient contents in the image. We create a collection of task prompt templates that instruct the model to describe the image either in detail or briefly." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 310, + 268, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 310, + 268, + 323 + ], + "spans": [ + { + "bbox": [ + 47, + 310, + 268, + 323 + ], + "type": "text", + "content": "3.2. Referring Expression Segmentation (RES)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 327, + 287, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 287, + 434 + ], + "type": "text", + "content": "In contrast to previous tasks, the referring expression segmentation task requires that the model generates a segmentation mask based on a given referring expression. Besides the RefCOCO series [43, 52, 86], we have further leveraged existing RES benchmarks [37, 65, 76, 78, 79] for this purpose. As shown in Figure 4, we instruct the model to localize the referred objects, with the expected output being a repetition of the referring expression whose grounding mask is the target object." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 443, + 287, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 443, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 47, + 443, + 287, + 456 + ], + "type": "text", + "content": "3.3. Grounded Visual Question Answering (GVQA)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 462, + 288, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 288, + 619 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 288, + 619 + ], + "type": "text", + "content": "The task of grounded visual question answering requires the model to comprehend a question (with optional pointers) and to produce an answer that is grounded to a binary segmentation mask that justifies the answer. The goal of this task is to enable natural QA-based interaction with users and reasoning in the model with grounded explanations. Specifically, we harvest and adapt a collection of public VQA datasets with grounding annotations [6, 8, 19, 29], QA benchmarks on visual text [64], and create populated QA pairs from segmentation datasets [25, 60]. Our prompt templates instruct the model to respond either as open-ended answers or by selecting from multiple choices, with the response anchored to a segmentation mask (Figure 4)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 624, + 192, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 624, + 192, + 637 + ], + "spans": [ + { + "bbox": [ + 47, + 624, + 192, + 637 + ], + "type": "text", + "content": "3.4. Referential Dialogue (RD)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 641, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 713 + ], + "type": "text", + "content": "The task of referential dialogue requires the model to conduct dialogue communication with users, especially when conditioned on user-provided spatial prompts. This includes existing RD datasets [8, 51, 88, 90, 94], multi-turn augmentations from segmentation datasets [17, 36, 47] as well as the referring expression generation (REG) task the" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 308, + 72, + 544, + 266 + ], + "blocks": [ + { + "bbox": [ + 308, + 72, + 544, + 266 + ], + "lines": [ + { + "bbox": [ + 308, + 72, + 544, + 266 + ], + "spans": [ + { + "bbox": [ + 308, + 72, + 544, + 266 + ], + "type": "image", + "image_path": "f4fd885ed5f3526f41a055fa0bb8365a1ae69932f147bc99f70830caea9e9194.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 273, + 546, + 306 + ], + "lines": [ + { + "bbox": [ + 305, + 273, + 546, + 306 + ], + "spans": [ + { + "bbox": [ + 305, + 273, + 546, + 306 + ], + "type": "text", + "content": "Figure 4. The M3G2 dataset for grounded visual instruction tuning. M3G2 is a diverse dataset of multiple granularities, unifying 4 different task types with visually grounded dialogue." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 312, + 546, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 312, + 546, + 396 + ], + "spans": [ + { + "bbox": [ + 305, + 312, + 546, + 396 + ], + "type": "text", + "content": "RefCOCO series [43, 52, 86]. The REG task differs from the region captioning task in that it demands the description to be a referring expression that distinctly identifies the targeted object. Effective REG calls for the model to engage in dialogue interactions cooperatively, adhering to the Gricean Maxims [24] which dictate that communication should be as informative, truthful, relevant, and clear as necessary." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 407, + 451, + 421 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 407, + 451, + 421 + ], + "spans": [ + { + "bbox": [ + 306, + 407, + 451, + 421 + ], + "type": "text", + "content": "4. Experiment and Analysis" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 426, + 545, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 426, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 305, + 426, + 545, + 487 + ], + "type": "text", + "content": "We adopt the LLaMA2-7B model [70] as our base LLM, and initialized the weight from LLaVA-1.5 [44]. For the vision encoders, we use the OpenAI CLIP@336 [62] model and DINOv2-L/14-reg [15] pretrained checkpoints. Please refer to Appendix C for more implementation details." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 491, + 546, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 491, + 546, + 601 + ], + "spans": [ + { + "bbox": [ + 305, + 491, + 546, + 601 + ], + "type": "text", + "content": "4.1. Generalist in Grounded Vision-Language Tasks We first demonstrate GROUNDHOG's capabilities as a generalist model for three different types of grounded vision-language tasks. It's worth noting that, unlike previous work that needs dataset-specific fine-tuning on each of the tasks, GROUNDHOG can achieve comparable performance on all the tasks directly after training on M3G2, i.e., all the reported results from our model are from a single set of weights without any dataset-specific fine-tuning." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "Language Grounding To Segmentation. We start by evaluating the model on language grounding tasks, which takes text as input and generates segmentation masks as output. We assess GROUNDHOG on Referential Expression Segmentation (RES) [32] and Caption Phrase Grounding (CPG) tasks. While traditional RES benchmarks [32, 53] focus on single-instance referents requiring primarily visual understanding, we expanded our evaluation to include complex scenarios involving multi-instance or negative queries" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14231" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 70, + 547, + 231 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 547, + 231 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 547, + 231 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 547, + 231 + ], + "type": "table", + "html": "
ModelSingle InstanceMulti-/No InstanceReasoning
RefCOCORefCOCO+RefCOCOggRefCOCOPhraseCutReasonSegRIO
valtest-Atest-Bvaltest-Atest-Bval-utest-uvaltestvaltest-ctest-u
Specialist
MDETR [30]---------53.7-44.122.0
CRIS [75]70.573.266.162.368.153.759.960.455.3----
LAVT [82]72.775.868.862.168.455.161.262.158.4----
ReLA [43]73.876.570.266.071.057.765.066.063.6-22.4--
PolyFormer [46]76.078.373.369.374.661.969.270.2---48.826.8
UNINEXT-H [80]82.283.481.372.576.466.274.776.4-----
Generalist
LISA7B [37]74.176.571.162.467.456.566.468.5--44.0--
LISA7B (FT) [37]74.979.172.365.170.858.167.970.6--52.9--
7B78.579.975.770.575.064.974.174.666.754.556.257.933.9
", + "image_path": "412f68b1e5796260afa5028b891a73e88f793be5f2625d8d66d3028bfebf654c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 49, + 259, + 197, + 367 + ], + "blocks": [ + { + "bbox": [ + 47, + 232, + 546, + 255 + ], + "lines": [ + { + "bbox": [ + 47, + 232, + 546, + 255 + ], + "spans": [ + { + "bbox": [ + 47, + 232, + 546, + 255 + ], + "type": "text", + "content": "Table 3. Results on 7 Referring Expression Segmentation (RES) benchmarks with single instance queries [32, 53], multi-/null instance queries [43, 76] and reasoning-based queries [37, 61]. We report cIoU for RefCOCO+/g and mIoU for other benchmarks, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 259, + 197, + 367 + ], + "lines": [ + { + "bbox": [ + 49, + 259, + 197, + 367 + ], + "spans": [ + { + "bbox": [ + 49, + 259, + 197, + 367 + ], + "type": "table", + "html": "
ModelFlickr30K-E
R@1_valR@1_test
Shikra13B77.478.4
Ferret13B81.184.8
Shikra7B75.876.5
Ferret7B80.482.2
7B79.279.8
", + "image_path": "5497b0b2f4d7c3dd396a279776935c989546f8c00039e53af1560da0e48f25f1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 200, + 259, + 373, + 367 + ], + "blocks": [ + { + "bbox": [ + 47, + 369, + 192, + 392 + ], + "lines": [ + { + "bbox": [ + 47, + 369, + 192, + 392 + ], + "spans": [ + { + "bbox": [ + 47, + 369, + 192, + 392 + ], + "type": "text", + "content": "Table 4. Top-1 box recall results on Flickr30K-Entity [58]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 200, + 259, + 373, + 367 + ], + "lines": [ + { + "bbox": [ + 200, + 259, + 373, + 367 + ], + "spans": [ + { + "bbox": [ + 200, + 259, + 373, + 367 + ], + "type": "table", + "html": "
ModelAR\\( \\mathbf{A}\\mathbf{R}_{\\mathrm {th}} \\)PNG \\( \\mathbf{A}\\mathbf{R}_{\\mathrm {st}} \\)\\( \\mathbf{A}\\mathbf{R}_{\\mathrm {s}} \\)\\( \\mathbf{A}\\mathbf{R}_{\\mathrm {p}} \\)
PiGLET65.964.068.667.254.5
\\( 7_{\\text{B}} \\)66.865.069.470.457.7
Table 5. Phrase grounding results on PNG [21].
ModelTextVQA-X [mIoU]
SAB29.0
\\( 7_{\\text{B}} \\)39.8
", + "image_path": "8f5cf507453a4539b127a442c4df6477465458d7394383c93cb78c5d4bb9be52.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 377, + 259, + 547, + 367 + ], + "blocks": [ + { + "bbox": [ + 198, + 368, + 376, + 391 + ], + "lines": [ + { + "bbox": [ + 198, + 368, + 376, + 391 + ], + "spans": [ + { + "bbox": [ + 198, + 368, + 376, + 391 + ], + "type": "text", + "content": "Table 6. Visual text QA results on the TextVQA-X [66] validation set." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 377, + 259, + 547, + 367 + ], + "lines": [ + { + "bbox": [ + 377, + 259, + 547, + 367 + ], + "spans": [ + { + "bbox": [ + 377, + 259, + 547, + 367 + ], + "type": "table", + "html": "
ModelPointQA Twice AccV7W Acc
Shikra13B70.385.3
GPT4RoI13B-84.8
Shikra7B--
GPT4RoI7B-81.8
7B72.485.5
", + "image_path": "99328dbfa3addd0fa7016d284ebc8d617c8fd7fbe5a58c6a445a7e2950e8d0db.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 396, + 289, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 289, + 649 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 289, + 649 + ], + "type": "text", + "content": "[43, 76], and those necessitating common sense reasoning [37, 61]. The results in Table 3 show GROUNDHOG outperforming the generalist model LISA across all benchmarks and achieving significant improvements over specialist models in multi-instance, null, and reasoning-based RES tasks. It also performs comparably on the competitive RefCOCO series. For CPG tasks, which involve grounding all phrases in a caption and demand a deep understanding of the context for coreference resolution, we first evaluated GROUNDHOG on the Flickr30K-Entity dataset [58]. Since this dataset only has box annotations, we convert the mask predictions of our model to box and compute the top-1 box recall following the merged-box protocol [30]. Despite not specializing in predicting boxes, GROUNDHOG still outperforms Shikra 7B/13B [8] and is on par with Ferret-7B [85] in a concurrent work (Table 4). Additionally, on the PNG dataset [34] which tests phrase grounding in longer narratives, GROUNDHOG surpasses the previous state-of-the-art model, PiGLET [22], in all metrics including average recall of grounding masks and detailed scores for things, stuffs, and singular and plural entities (Table 5)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "Grounded Language Generation. Our model excels in generating language that accurately grounds to segmentation masks during user conversations. Quantitatively, we assess grounded captioning on the Flickr30K-Entity dataset [58], employing standard text generation metrics" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 309, + 395, + 545, + 466 + ], + "blocks": [ + { + "bbox": [ + 397, + 369, + 543, + 392 + ], + "lines": [ + { + "bbox": [ + 397, + 369, + 543, + 392 + ], + "spans": [ + { + "bbox": [ + 397, + 369, + 543, + 392 + ], + "type": "text", + "content": "Table 7. Results on PointQATwice [51] and V7W [94] test sets." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 395, + 545, + 466 + ], + "lines": [ + { + "bbox": [ + 309, + 395, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 309, + 395, + 545, + 466 + ], + "type": "table", + "html": "
ModelBleu-4METEORCIDErSPICEF1all
Shikra13B--73.9--
Ferret13B37.025.576.118.315.1
Ferret7B35.124.674.818.015.0
7B36.726.591.320.432.1
", + "image_path": "bdde4fefe2e41c3207281a342ebfd0005f924ec393d29047c40940a8472fa04c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 321, + 466, + 528, + 476 + ], + "lines": [ + { + "bbox": [ + 321, + 466, + 528, + 476 + ], + "spans": [ + { + "bbox": [ + 321, + 466, + 528, + 476 + ], + "type": "text", + "content": "Table 8. Grounded Captioning on Flickr30K-Entity [58]." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 483, + 545, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 545, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 545, + 626 + ], + "type": "text", + "content": "such as Bleu-4 [55], METEOR [4], CIDEr [72], and SPICE [2] for language quality; and the " + }, + { + "bbox": [ + 304, + 483, + 545, + 626 + ], + "type": "inline_equation", + "content": "\\mathrm{F1}_{\\mathrm{all}}" + }, + { + "bbox": [ + 304, + 483, + 545, + 626 + ], + "type": "text", + "content": " score for grounding accuracy following You et al. [85]. As shown in Table 8, GROUNDHOG significantly surpasses existing box-based grounded MLLMs, even their 13B versions, in both language quality and grounding accuracy. This improvement is hypothesized to stem from the diverse task distribution in our M3G2 dataset. For groundable question answering, we evaluate on the TextVQA-X benchmark [64]. GROUNDHOG outperforms the state-of-the-art specialist model SAB [33] by a significant margin, as measured by the mean IoU of the predicted mask (Table 6)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 629, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 546, + 713 + ], + "type": "text", + "content": "Spatial Prompt Understanding. For grounded MLLMs, accurately interpreting multimodal instructions is essential, particularly in interactive tasks. We evaluated its performance on two pointer-based QA benchmarks, PointerQATwice [51] and V7W [94], which require the model to answer questions guided by spatial prompts, such as bounding boxes. The model is tasked to generate free-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14232" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 86, + 127, + 182 + ], + "blocks": [ + { + "bbox": [ + 53, + 86, + 127, + 182 + ], + "lines": [ + { + "bbox": [ + 53, + 86, + 127, + 182 + ], + "spans": [ + { + "bbox": [ + 53, + 86, + 127, + 182 + ], + "type": "image", + "image_path": "dfe920c57005f4d9d6b538244aeaf7653366145867b38e7e3bbdce9d5d0e890f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 183, + 286, + 216 + ], + "lines": [ + { + "bbox": [ + 47, + 183, + 286, + 216 + ], + "spans": [ + { + "bbox": [ + 47, + 183, + 286, + 216 + ], + "type": "text", + "content": "Figure 5. Region caption using the best match proposal from Mask2Former+ versus from SAM. Mask2Former+ fails to propose the exact mask of the spire, leading to a less precise caption." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 129, + 86, + 203, + 182 + ], + "blocks": [ + { + "bbox": [ + 62, + 73, + 253, + 84 + ], + "lines": [ + { + "bbox": [ + 62, + 73, + 253, + 84 + ], + "spans": [ + { + "bbox": [ + 62, + 73, + 253, + 84 + ], + "type": "text", + "content": "Input Prompt M2F+ Best Match SAM" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 129, + 86, + 203, + 182 + ], + "lines": [ + { + "bbox": [ + 129, + 86, + 203, + 182 + ], + "spans": [ + { + "bbox": [ + 129, + 86, + 203, + 182 + ], + "type": "image", + "image_path": "f44c2995b738f414c001eb4996302b0a3f3e927bcbdbbb4870b7e119721bd49c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 205, + 86, + 279, + 182 + ], + "blocks": [ + { + "bbox": [ + 205, + 86, + 279, + 182 + ], + "lines": [ + { + "bbox": [ + 205, + 86, + 279, + 182 + ], + "spans": [ + { + "bbox": [ + 205, + 86, + 279, + 182 + ], + "type": "image", + "image_path": "b503064ea84dd4572d44b2a4073a2f5a60f5c664633c3b3f77788d798bdf0886.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 51, + 220, + 281, + 327 + ], + "blocks": [ + { + "bbox": [ + 51, + 220, + 281, + 327 + ], + "lines": [ + { + "bbox": [ + 51, + 220, + 281, + 327 + ], + "spans": [ + { + "bbox": [ + 51, + 220, + 281, + 327 + ], + "type": "image", + "image_path": "a99b56cdcb3679d9a612a7a78dfc3529ef504a9550cc9202fe2643846d33ed4b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 327, + 286, + 392 + ], + "lines": [ + { + "bbox": [ + 47, + 327, + 286, + 392 + ], + "spans": [ + { + "bbox": [ + 47, + 327, + 286, + 392 + ], + "type": "text", + "content": "Figure 6. Illustration of a partially correct grounding. The grounding phrase and the ground truth mask are shown on the left. The top-4 mask proposals are presented, with highly-scored masks (green) selected for the merged mask, and low-scored masks (red) excluded. This illustrates the failure to recognize the word \"KWIK\" by the MLLM, despite its successful proposal." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 399, + 286, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 286, + 554 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 286, + 554 + ], + "type": "text", + "content": "form textual answers in PointerQA" + }, + { + "bbox": [ + 46, + 399, + 286, + 554 + ], + "type": "inline_equation", + "content": "_{\\text{Twice}}" + }, + { + "bbox": [ + 46, + 399, + 286, + 554 + ], + "type": "text", + "content": ", and selects from multiple-choice options in V7W. GROUNDHOG demonstrates superior performance in these benchmarks, outperforming previous models as shown in Table 7. This highlights its effectiveness in spatial understanding and response accuracy. To further demonstrate the effectiveness of using SAM for the pointer-to-mask conversion, we show the best-matched mask proposal from our Mask2Former+ model in comparison to the mask from SAM in Figure 5. While the best match proposal from the Mask2Former+ model includes a broader area, the SAM-generated mask offers a more precise representation of the specified region, potentially leading to a more accurate caption." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 559, + 233, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 559, + 233, + 571 + ], + "spans": [ + { + "bbox": [ + 47, + 559, + 233, + 571 + ], + "type": "text", + "content": "4.2. Trustworthiness and Transparency" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 578, + 286, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 578, + 286, + 613 + ], + "spans": [ + { + "bbox": [ + 47, + 578, + 286, + 613 + ], + "type": "text", + "content": "Beyond its superior performance as a grounding generalist, we highlight two key improvements for creating a more trustworthy and transparent agent." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 618, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 618, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 618, + 286, + 713 + ], + "type": "text", + "content": "Reduced Object Hallucination. Thanks to the varied task distribution and the inclusion of negative question-answering samples in M3G2 dataset, GROUNDHOG significantly reduces object hallucination. We assessed this using the POPE [42] benchmark, which includes binary questions about object existence across three splits, each with a different object distribution (with an order of difficulty Random " + }, + { + "bbox": [ + 47, + 618, + 286, + 713 + ], + "type": "inline_equation", + "content": "<" + }, + { + "bbox": [ + 47, + 618, + 286, + 713 + ], + "type": "text", + "content": " Popular " + }, + { + "bbox": [ + 47, + 618, + 286, + 713 + ], + "type": "inline_equation", + "content": "<" + }, + { + "bbox": [ + 47, + 618, + 286, + 713 + ], + "type": "text", + "content": " Adversarial). Remarkably, GROUND" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 310, + 71, + 544, + 342 + ], + "blocks": [ + { + "bbox": [ + 310, + 71, + 544, + 342 + ], + "lines": [ + { + "bbox": [ + 310, + 71, + 544, + 342 + ], + "spans": [ + { + "bbox": [ + 310, + 71, + 544, + 342 + ], + "type": "table", + "html": "
ModelAccuracyPrecisionRecallF1 ScoreYes (%)
Random
mPLUG-Owl53.3051.7199.5368.0696.23
LLaVA54.4352.3299.8068.6595.37
MultiModal-GPT50.0350.02100.0066.6899.97
MiniGPT-477.8375.3882.6778.8654.83
InstructBLIP88.7385.0893.9389.2955.20
Shikra-13B86.9094.4079.2686.1943.26
Ferret-13B90.2497.7283.0089.7643.26
7B91.0385.8096.4090.7945.88
Popular
mPLUG-Owl50.6350.3299.2766.7998.63
LLaVA52.4351.2599.8067.7297.37
MultiModal-GPT50.0050.00100.0066.67100.00
MiniGPT-468.3064.2782.4072.2164.10
InstructBLIP81.3775.0793.9383.4562.57
Shikra-13B83.9787.5579.2083.1645.23
Ferret-13B84.9088.2480.5384.2145.63
7B90.1385.9393.8189.7045.80
Adversarial
mPLUG-Owl50.6750.3499.3366.8298.67
LLaVA50.7750.3999.8766.9899.10
MultiModal-GPT50.0050.00100.0066.67100.00
MiniGPT-466.6062.4583.2771.3766.67
InstructBLIP74.3767.6793.3378.4568.97
Shikra-13B83.1085.6079.6082.4946.50
Ferret-13B82.3683.6080.5382.0048.18
7B86.3385.9386.6386.2849.60
", + "image_path": "c4a82f23b38194e3150f7ce7b3c8b5ed7e415055a324d7d03f62e3e122ca7450.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 345, + 544, + 356 + ], + "lines": [ + { + "bbox": [ + 306, + 345, + 544, + 356 + ], + "spans": [ + { + "bbox": [ + 306, + 345, + 544, + 356 + ], + "type": "text", + "content": "Table 9. Object hallucination results on the POPE [42] benchmark." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 365, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 365, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 305, + 365, + 545, + 449 + ], + "type": "text", + "content": "HOG consistently outperforms other models in both accuracy and F1 score across all splits, particularly on the more challenging ones. It shows an absolute improvement of " + }, + { + "bbox": [ + 305, + 365, + 545, + 449 + ], + "type": "inline_equation", + "content": "5.2\\%" + }, + { + "bbox": [ + 305, + 365, + 545, + 449 + ], + "type": "text", + "content": " in accuracy for Popular and " + }, + { + "bbox": [ + 305, + 365, + 545, + 449 + ], + "type": "inline_equation", + "content": "4.0\\%" + }, + { + "bbox": [ + 305, + 365, + 545, + 449 + ], + "type": "text", + "content": " for Adversarial over the previously best-performing model. This suggests that our model's enhanced grounding capability plays a significant role in mitigating the object hallucination problem." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 457, + 545, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 457, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 457, + 545, + 601 + ], + "type": "text", + "content": "Explainability and Diagnosability. Another important highlight of GROUNDHOG is its enhancement of explainability through the decoupled design of entity proposal and selection, as outlined earlier in section 2.2. This is exemplified in the case study illustrated in Figure 6, which illustrates the mask proposal scoring and selective merging process of our model. We show the top-4 masks, where the higher-score masks are labeled in green while the lower-score masks are labeled in red. Users can easily interpret that the failure is due to the incapability of MLLM to recognize the word \"KWIK\", despite it being successfully localized and proposed as an entity candidate." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 610, + 406, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 610, + 406, + 621 + ], + "spans": [ + { + "bbox": [ + 306, + 610, + 406, + 621 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "We performed ablation studies to validate our design decisions, training, and evaluating a subset of the M3G2 dataset that includes RefCOCO+, Flickr30K, and TextVQA. These cover a range of visual entities from various image sources and granularities. We start by comparing our Mask2Former+ with the original Mask2Former for mask proposal effectiveness. As indicated in Table 10, the orig-" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14233" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 70, + 288, + 247 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 288, + 247 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 288, + 247 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 288, + 247 + ], + "type": "table", + "html": "
SetupRefCOCO+Flickr30KTextVQA-X
Mask Proposal Models
Mask2Former67.169.09.8
Mask2Former+66.677.234.0
Entity Features
CLIP59.875.032.0
DINOv262.376.328.4
CLIP+DINOv266.677.234.0
Grounding Query
<GRD> only64.467.534.2
</GRD> only64.477.233.5
Sum66.677.234.0
Eval Input Resolution
224–48054.767.227.6
480–64065.576.727.6
800–102466.677.234.0
", + "image_path": "dee71b92735563a8da406e3d73573ae39ab3562c439978d9ec1ea25ac6e234d2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 47, + 251, + 287, + 285 + ], + "lines": [ + { + "bbox": [ + 47, + 251, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 47, + 251, + 287, + 285 + ], + "type": "text", + "content": "Table 10. Ablation study on model design choices and evaluation setups. Models are trained on RefCOCO+, Flickr30K, TextVQA-X and tested on corresponding validation sets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 290, + 287, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 290, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 290, + 287, + 483 + ], + "type": "text", + "content": "inal Mask2Former performs slightly better on RefCOCO, as it is developed specificity on COCO object categories. However, Mask2Former+ significantly surpasses the original in domains with non-COCO entities. Our second set of experiments examined the choice of visual entity features. Although using either CLIP or DINOv2 features alone shows advantages in specific datasets, their combination consistently yields the best results across all datasets. To obtain a robust grounding query representation, we experimented with using the output embedding of the token, the token, and their sum. We found that the latter approach achieves the best overall results. Finally, we demonstrate that our decoupling design of the mask proposal model and MLLM allows for training at a lower resolution (320px) to expedite grounding training, while scaling up the resolution during evaluation enhances performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 490, + 134, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 490, + 134, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 490, + 134, + 502 + ], + "type": "text", + "content": "5. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 505, + 287, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 505, + 287, + 661 + ], + "spans": [ + { + "bbox": [ + 46, + 505, + 287, + 661 + ], + "type": "text", + "content": "Multimodal Large Language Models. Building on the recent advance of large language models (LLMs), there is an increasing effort in adapting pretrained large language models for multimodal tasks, such as understanding and interpreting visual information [1, 71]. More recently, visual instruction tuning has gained much interest due to its surprising performance with a modest amount of data and computing resources. Various models have been developed, noticeably MiniGPT4[93], LLaVA [44, 45] and concurrent models [13, 20, 38, 74, 83]. Despite their promising performances, MLLMs often produce objects that are not presented in the given images, a phenomenon referred to as the object hallucination problem [14, 31, 42]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "content": "MLLM with Language Grounding The ability to connect language to their corresponding visual elements in the physical world, known as grounding [27], is crucial in everyday human communication about our shared sur" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": "roundings. Grounding datasets have been shown to benefit vision-language pre-training, both in terms of object-level recognition [41] and language learning [50]. Recent works unify text and grounding regions into token sequences [49, 73, 81] in casual language modeling. Based on such paradigm, researchers have developed a family of grounded MLLM, including GPT4ROI [89], Kosmos-2 [56], Shikra [8], PVIT [7], BuboGPT [91], Qwen-VL [3], and Ferret [85]. Despite their promising performance, these models focus on object grounding to bounding boxes, which cannot handle pixel-level grounding across various semantic granularities. Furthermore, it lacks the diagnosability and explainability in failure cases. Our model fills this gap." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 233, + 546, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 233, + 546, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 233, + 546, + 437 + ], + "type": "text", + "content": "Language-Guided Semantic Localization The field of language-guided semantic localization has a long history in the vision-language research community, requiring that the model localize a given referring expression with bounding boxes or segmentation masks. This task has evolved from early attempts to understand simple referring expressions within images, such as the well-known RefCOCO series [52, 86] and their generalized variant [43] that takes no-target and multi-target into account. The integration of advanced language reasoning from LLMs has enabled research to tackle even more nuanced reasoning tasks that involve complex language contexts [37, 57, 87]. Notably, LISA [37] formulates a reasoning segmentation task to bring language-informed reasoning into semantic segmentation, and contributes a powerful baseline. Our model builds on these developments, but is designed to be more universally applicable as a grounded MLLM." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 441, + 378, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 441, + 378, + 453 + ], + "spans": [ + { + "bbox": [ + 306, + 441, + 378, + 453 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 456, + 545, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 456, + 545, + 672 + ], + "spans": [ + { + "bbox": [ + 304, + 456, + 545, + 672 + ], + "type": "text", + "content": "In this study, we introduce GROUNDHOG, a novel framework designed to enable pixel-level explainable grounding in large language models, leveraging holistic segmentation. The system builds upon a pre-trained mask proposal network to provide pixel-level visual features for the large language models, allowing them to retrieve segmentation mask proposals that can be used for grounding. We also present M3G2, a dataset of 1.9M training text-image pairs with 36 sub-problems derived from 27 existing datasets for visually grounded instruction tuning, facilitating precise vision-language alignment at the pixel level. We show that after training on M3G2, GROUNDHOG achieves superior performance on various grounding tasks. Through extensive case studies, we further show that GROUNDHOG unlocks explainability and diagnosability, and demonstrates better grounding towards occluded objects, groups of multiple instances, amorphous background regions, semantic parts of objects, and objects with irregular shapes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 677, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 715 + ], + "type": "text", + "content": "Acknowledgements This work was supported by Amazon and NSF IIS-1949634. We thank the anonymous reviewers for their valuable comments and suggestions." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14234" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 158, + 288, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 288, + 201 + ], + "type": "text", + "content": "[2] Peter Anderson, Basura Fernando, Mark Johnson, and Stephen Gould. Spice: Semantic propositional image caption evaluation. In Proceedings of the 14th European Conference on Computer Vision, pages 382-398, 2016. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 202, + 288, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 288, + 246 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 288, + 246 + ], + "type": "text", + "content": "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 247, + 288, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 288, + 300 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 288, + 300 + ], + "type": "text", + "content": "[4] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 302, + 288, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 288, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 288, + 346 + ], + "type": "text", + "content": "[5] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocosstuff: Thing and stuff classes in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1209-1218, 2018. 2, 4, 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 346, + 288, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 288, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 288, + 401 + ], + "type": "text", + "content": "[6] Chongyan Chen, Samreen Anjum, and Danna Gurari. Grounding answers for visual questions asked by visually impaired people. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19098-19107, 2022. 5, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 402, + 288, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 402, + 288, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 402, + 288, + 446 + ], + "type": "text", + "content": "[7] Chi Chen, Ruoyu Qin, Fuwen Luo, Xiaoyue Mi, Peng Li, Maosong Sun, and Yang Liu. Position-enhanced visual instruction tuning for multimodal large language models. arXiv preprint arXiv:2308.13437, 2023. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 447, + 288, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 288, + 490 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 288, + 490 + ], + "type": "text", + "content": "[8] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 491, + 288, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 491, + 288, + 534 + ], + "spans": [ + { + "bbox": [ + 53, + 491, + 288, + 534 + ], + "type": "text", + "content": "[9] Ting Chen, Saurabh Saxena, Lala Li, David J Fleet, and Geoffrey Hinton. Pix2seq: A language modeling framework for object detection. In International Conference on Learning Representations, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 536, + 288, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 288, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 288, + 590 + ], + "type": "text", + "content": "[10] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 2, 1, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 591, + 288, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 288, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 288, + 646 + ], + "type": "text", + "content": "[11] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 48, + 591, + 288, + 646 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 48, + 591, + 288, + 646 + ], + "type": "text", + "content": " * chatgpt quality. https://vicuna.lmsys.org, 2023. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 288, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 691 + ], + "type": "text", + "content": "[12] Jifeng Dai, Kaiming He, and Jian Sun. Convolutional feature masking for joint object and stuff segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3992-4000, 2015. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "text", + "content": "[13] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500, 2023. 2, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 106, + 547, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 106, + 547, + 171 + ], + "spans": [ + { + "bbox": [ + 308, + 106, + 547, + 171 + ], + "type": "text", + "content": "[14] Wenliang Dai, Zihan Liu, Ziwei Ji, Dan Su, and Pascale Fung. Plausible may not be faithful: Probing object hallucination in vision-language pre-training. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 2128-2140, 2023. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 172, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 172, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 172, + 545, + 205 + ], + "type": "text", + "content": "[15] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. arXiv preprint arXiv:2309.16588, 2023. 5, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 206, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 259 + ], + "type": "text", + "content": "[16] Daan de Geus, Panagiotis Meletis, Chenyang Lu, Xiaoxiao Wen, and Gijs Dubbelman. Part-aware panoptic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5485-5494, 2021. 4, 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 261, + 545, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 324 + ], + "type": "text", + "content": "[17] Harm De Vries, Florian Strub, Sarath Chandar, Olivier Pietquin, Hugo Larochelle, and Aaron Courville. Guess-what?! visual object discovery through multi-modal dialogue. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5503–5512, 2017. 5, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 327, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 370 + ], + "type": "text", + "content": "[18] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 371, + 545, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 371, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 308, + 371, + 545, + 436 + ], + "type": "text", + "content": "[19] Chuang Gan, Yandong Li, Haoxiang Li, Chen Sun, and Boqing Gong. Vqs: Linking segmentations to questions and answers for supervised attention in vqa and question-focused semantic segmentation. In Proceedings of the IEEE international conference on computer vision, pages 1811-1820, 2017. 5, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 437, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 545, + 491 + ], + "type": "text", + "content": "[20] Tao Gong, Chengqi Lyu, Shilong Zhang, Yudong Wang, Miao Zheng, Qian Zhao, Kuikun Liu, Wenwei Zhang, Ping Luo, and Kai Chen. Multimodal-gpt: A vision and language model for dialogue with humans. arXiv preprint arXiv:2305.04790, 2023. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 492, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 492, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 308, + 492, + 545, + 547 + ], + "type": "text", + "content": "[21] Cristina González, Nicolás Ayobi, Isabela Hernández, José Hernández, Jordi Pont-Tuset, and Pablo Arbeláez. Panoptic narrative grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1364-1373, 2021. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 548, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 601 + ], + "type": "text", + "content": "[22] Cristina González, Nicolás Ayobi, Isabela Hernández, Jordi Pont-Tuset, and Pablo Arbeláez. Piglet: Pixel-level grounding of language expressions with transformers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 602, + 545, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 602, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 308, + 602, + 545, + 647 + ], + "type": "text", + "content": "[23] Abel Gonzalez-Garcia, Davide Modolo, and Vittorio Ferrari. Objects as context for detecting their semantic parts. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6907-6916, 2018. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 647, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 669 + ], + "type": "text", + "content": "[24] Herbert P Grice. Logic and conversation. In Speech acts, pages 41-58. Brill, 1975. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 670, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 714 + ], + "type": "text", + "content": "[25] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 4, 5, 1, 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14235" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[26] Kai Han, Yunhe Wang, Jianyuan Guo, Yehui Tang, and Enhua Wu. Vision gnn: An image is worth graph of nodes. Advances in Neural Information Processing Systems, 35:8291-8303, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 286, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 286, + 138 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 286, + 138 + ], + "type": "text", + "content": "[27] Stevan Harnad. The symbol grounding problem. Physica D: Nonlinear Phenomena, 42(1-3):335-346, 1990. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 140, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 140, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 49, + 140, + 287, + 183 + ], + "type": "text", + "content": "[28] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2021. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 184, + 287, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 184, + 287, + 237 + ], + "spans": [ + { + "bbox": [ + 49, + 184, + 287, + 237 + ], + "type": "text", + "content": "[29] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 5, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 239, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 239, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 49, + 239, + 287, + 293 + ], + "type": "text", + "content": "[30] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1780-1790, 2021. 4, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 294, + 287, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 294, + 287, + 347 + ], + "spans": [ + { + "bbox": [ + 49, + 294, + 287, + 347 + ], + "type": "text", + "content": "[31] Osman Semih Kayhan, Bart Vredebregt, and Jan C Van Gemert. Hallucination in object detection—a study in visual part verification. In 2021 IEEE International Conference on Image Processing (ICIP), pages 2234–2238. IEEE, 2021. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 350, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 350, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 49, + 350, + 287, + 403 + ], + "type": "text", + "content": "[32] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014. 5, 6, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 404, + 287, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 404, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 49, + 404, + 287, + 448 + ], + "type": "text", + "content": "[33] Seyedalireza Khoshkirat and Chandra Kambhametty. Sentence attention blocks for answer grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6080-6090, 2023. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 449, + 287, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 449, + 287, + 492 + ], + "spans": [ + { + "bbox": [ + 49, + 449, + 287, + 492 + ], + "type": "text", + "content": "[34] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dólar. Panoptic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9404-9413, 2019. 5, 6, 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 493, + 287, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 493, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 49, + 493, + 287, + 559 + ], + "type": "text", + "content": "[35] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4015-4026, 2023. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 559, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 559, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 49, + 559, + 287, + 624 + ], + "type": "text", + "content": "[36] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 5, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 625, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 625, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 49, + 625, + 287, + 668 + ], + "type": "text", + "content": "[37] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692, 2023. 4, 5, 6, 8, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "type": "text", + "content": "[38] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726, 2023. 8" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 308, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 308, + 73, + 545, + 116 + ], + "type": "text", + "content": "[39] Chunyuan Li, Zhe Gan, Zhengyuan Yang, Jianwei Yang, Linjie Li, Lijuan Wang, and Jianfeng Gao. Multimodal foundation models: From specialists to general-purpose assistants. arXiv preprint arXiv:2309.10020, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 118, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 545, + 161 + ], + "type": "text", + "content": "[40] Jianshu Li, Jian Zhao, Yunchao Wei, Congyan Lang, Yidong Li, Terence Sim, Shuicheng Yan, and Jiashi Feng. Multiple-human parsing in the wild. arXiv preprint arXiv:1705.07206, 2017. 4, 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 163, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 229 + ], + "type": "text", + "content": "[41] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 230, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 230, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 230, + 545, + 285 + ], + "type": "text", + "content": "[42] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, 2023. 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 286, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 545, + 330 + ], + "type": "text", + "content": "[43] Chang Liu, Henghui Ding, and Xudong Jiang. Gres: Generalized referring expression segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23592-23601, 2023. 5, 6, 8, 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 331, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 545, + 364 + ], + "type": "text", + "content": "[44] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 5, 8, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 365, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 545, + 398 + ], + "type": "text", + "content": "[45] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 2, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 399, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 464 + ], + "type": "text", + "content": "[46] Jiang Liu, Hui Ding, Zhaowei Cai, Yuting Zhang, Ravi Kumar Satzoda, Vijay Mahadevan, and R Manmatha. Polyformer: Referring image segmentation as sequential polygon generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18653-18663, 2023. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 467, + 545, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 467, + 545, + 531 + ], + "spans": [ + { + "bbox": [ + 308, + 467, + 545, + 531 + ], + "type": "text", + "content": "[47] Shangbang Long, Siyang Qin, Dmitry Panteleev, Alessandro Bissacco, Yasuhisa Fujii, and Michalis Raptis. Towards end-to-end unified scene text detection and layout analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1049-1059, 2022. 5, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 533, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 533, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 533, + 545, + 555 + ], + "type": "text", + "content": "[48] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 557, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 557, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 308, + 557, + 545, + 611 + ], + "type": "text", + "content": "[49] Jiasen Lu, Christopher Clark, Rowan Zellers, Roozbeh Mottaghi, and Aniruddha Kembhavi. Unified-io: A unified model for vision, language, and multi-modal tasks. In The Eleventh International Conference on Learning Representations, 2022. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 613, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 668 + ], + "type": "text", + "content": "[50] Ziqiao Ma, Jiayi Pan, and Joyce Chai. World-to-words: Grounded open vocabulary acquisition through fast mapping in vision-language models. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 524–544, 2023. 2, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "text", + "content": "[51] Arjun Mani, Nobline Yoo, Will Hinthorn, and Olga Russakovsky. Point and ask: Incorporating pointing into visual question answering. arXiv preprint arXiv:2011.13681, 2020. 5, 6, 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "14236" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[52] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 5, 8, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 288, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 288, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 288, + 183 + ], + "type": "text", + "content": "[53] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 5, 6, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 184, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 288, + 239 + ], + "type": "text", + "content": "[54] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 239, + 288, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 288, + 293 + ], + "type": "text", + "content": "[55] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 293, + 288, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 293, + 288, + 338 + ], + "spans": [ + { + "bbox": [ + 48, + 293, + 288, + 338 + ], + "type": "text", + "content": "[56] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 2, 3, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 338, + 288, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 288, + 382 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 288, + 382 + ], + "type": "text", + "content": "[57] Renjie Pi, Jiahui Gao, Shizhe Diao, Rui Pan, Hanze Dong, Jipeng Zhang, Lewei Yao, Jianhua Han, Hang Xu, and Lingpeng Kong Tong Zhang. Detgpt: Detect what you need via reasoning. arXiv preprint arXiv:2305.14167, 2023. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 383, + 288, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 288, + 449 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 288, + 449 + ], + "type": "text", + "content": "[58] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In Proceedings of the IEEE international conference on computer vision, pages 2641–2649, 2015. 5, 6, 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 449, + 288, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 288, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 288, + 502 + ], + "type": "text", + "content": "[59] Jordi Pont-Tuset, Jasper Uijlings, Soravit Changpinyo, Radu Soricut, and Vittorio Ferrari. Connecting vision and language with localized narratives. In Proceedings of the 16th European Conference on Computer Vision, pages 647-664, 2020. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 503, + 288, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 503, + 288, + 559 + ], + "spans": [ + { + "bbox": [ + 48, + 503, + 288, + 559 + ], + "type": "text", + "content": "[60] Lu Qi, Jason Kuen, Tiancheng Shen, Jiuxiang Gu, Wenbo Li, Weidong Guo, Jiaya Jia, Zhe Lin, and Ming-Hsuan Yang. High quality entity segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4047-4056, 2023. 4, 5, 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 559, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 288, + 624 + ], + "type": "text", + "content": "[61] Mengxue Qu, Yu Wu, Wu Liu, Xiaodan Liang, Jingkuan Song, Yao Zhao, and Yunchao Wei. RIO: A benchmark for reasoning intention-oriented objects in open environments. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 6, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 625, + 288, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 288, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 288, + 690 + ], + "type": "text", + "content": "[62] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 5, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 690, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 690, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 690, + 288, + 714 + ], + "type": "text", + "content": "[63] Vignesh Ramanathan, Anmol Kalia, Vladan Petrovic, Yi Wen, Baixue Zheng, Baishan Guo, Rui Wang, Aaron Mar" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 117 + ], + "type": "text", + "content": "quez, Rama Kovvuri, Abhishek Kadian, et al. Paco: Parts and attributes of common objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7141-7151, 2023. 4, 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "text", + "content": "[64] Varun Nagaraj Rao, Xingjian Zhen, Karen Hovsepian, and Mingwei Shen. A first look: Towards explainable textvqa models via visual and textual explanations. arXiv preprint arXiv:2105.02626, 2021. 5, 6, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 162, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 162, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 162, + 545, + 217 + ], + "type": "text", + "content": "[65] Anna Rohrbach, Marcus Rohrbach, Ronghang Hu, Trevor Darrell, and Bernt Schiele. Grounding of textual phrases in images by reconstruction. In Proceedings of the 14th European Conference on Computer Vision, pages 817-834. Springer, 2016. 5, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 217, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 217, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 307, + 217, + 545, + 271 + ], + "type": "text", + "content": "[66] Amanpreet Singh, Vivek Natarjan, Meet Shah, Yu Jiang, Xinlei Chen, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8317-8326, 2019. 4, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 271, + 545, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 271, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 307, + 271, + 545, + 315 + ], + "type": "text", + "content": "[67] Amanpreet Singh, Guan Pang, Mandy Toh, Jing Huang, Wojciech Galuba, and Tal Hassner. TextOCR: Towards large-scale end-to-end reasoning for arbitrary-shaped scene text. 2021. 4, 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 316, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 316, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 307, + 316, + 545, + 370 + ], + "type": "text", + "content": "[68] Bharat Singh, Hengduo Li, Abhishek Sharma, and Larry S Davis. R-fcn-3000 at 30fps: Decoupling detection and classification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1081–1090, 2018. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 371, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 371, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 307, + 371, + 545, + 425 + ], + "type": "text", + "content": "[69] Zhi Tian, Chunhua Shen, Xinlong Wang, and Hao Chen. Boxinst: High-performance instance segmentation with box annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5443-5452, 2021. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 426, + 545, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 426, + 545, + 482 + ], + "spans": [ + { + "bbox": [ + 307, + 426, + 545, + 482 + ], + "type": "text", + "content": "[70] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 5, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 482, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 482, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 307, + 482, + 545, + 525 + ], + "type": "text", + "content": "[71] Maria Tsimpoukelli, Jacob L Menick, Serkan Cabi, SM Eslami, Oriol Vinyls, and Felix Hill. Multimodal few-shot learning with frozen language models. Advances in Neural Information Processing Systems, 34:200-212, 2021. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 525, + 545, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 525, + 545, + 570 + ], + "spans": [ + { + "bbox": [ + 307, + 525, + 545, + 570 + ], + "type": "text", + "content": "[72] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 570, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 570, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 570, + 545, + 635 + ], + "type": "text", + "content": "[73] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In International Conference on Machine Learning, pages 23318-23340. PMLR, 2022. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 636, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 636, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 636, + 545, + 690 + ], + "type": "text", + "content": "[74] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, et al. Vision: Large language model is also an open-ended decoder for vision-centric tasks. arXiv preprint arXiv:2305.11175, 2023. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 690, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 690, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 690, + 545, + 714 + ], + "type": "text", + "content": "[75] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clip" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14237" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 288, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 288, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 288, + 106 + ], + "type": "text", + "content": "driven referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11686-11695, 2022. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "text", + "content": "[76] Chenyun Wu, Zhe Lin, Scott Cohen, Trung Bui, and Subhransu Maji. Phrasescut: Language-based image segmentation in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10216-10225, 2020. 4, 5, 6, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 196 + ], + "type": "text", + "content": "[77] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 198, + 287, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 287, + 241 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 287, + 241 + ], + "type": "text", + "content": "[78] Yixuan Wu, Zhao Zhang, Chi Xie, Feng Zhu, and Rui Zhao. Advancing referring expression segmentation beyond single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2628-2638, 2023. 5, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 243, + 287, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 243, + 287, + 296 + ], + "spans": [ + { + "bbox": [ + 48, + 243, + 287, + 296 + ], + "type": "text", + "content": "[79] Chi Xie, Zhao Zhang, Yixuan Wu, Feng Zhu, Rui Zhao, and Shuang Liang. Described object detection: Liberating object detection with flexible expressions. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 5, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 299, + 287, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 299, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 299, + 287, + 353 + ], + "type": "text", + "content": "[80] Bin Yan, Yi Jiang, Jiannan Wu, Dong Wang, Ping Luo, Zehuan Yuan, and Huchuan Lu. Universal instance perception as object discovery and retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15325-15336, 2023. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 355, + 287, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 355, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 48, + 355, + 287, + 399 + ], + "type": "text", + "content": "[81] Zhengyuan Yang, Zhe Gan, Jianfeng Wang, Xiaowei Hu, Faisal Ahmed, Zicheng Liu, Yumao Lu, and Lijuan Wang. Unitab: Unifying text and box outputs for grounded vision-language modeling. 2022. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 400, + 287, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 400, + 287, + 455 + ], + "spans": [ + { + "bbox": [ + 48, + 400, + 287, + 455 + ], + "type": "text", + "content": "[82] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Levt: Language-aware vision transformer for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18155-18165, 2022. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 456, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 510 + ], + "type": "text", + "content": "[83] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. MPLug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 512, + 287, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 287, + 545 + ], + "type": "text", + "content": "[84] Shukang Yin, Chaoyou Fu, Sirui Zhao, Ke Li, Xing Sun, Tong Xu, and Enhong Chen. A survey on multimodal large language models. arXiv preprint arXiv:2306.13549, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 546, + 287, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 546, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 48, + 546, + 287, + 600 + ], + "type": "text", + "content": "[85] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704, 2023. 3, 6, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 602, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 667 + ], + "type": "text", + "content": "[86] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. 4, 5, 8, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[87] Yuhang Zang, Wei Li, Jun Han, Kaiyang Zhou, and Chen Change Loy. Contextual object detection with multimodal large language models. arXiv preprint arXiv:2305.18279, 2023. 8" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 407 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[88] Rowan Zellers, Yonatan Bisk, Ali Farhadi, and Yejin Choi. From recognition to cognition: Visual commonsense reasoning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6720-6731, 2019. 5, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "type": "text", + "content": "[89] Shilong Zhang, Peize Sun, Shoufa Chen, Min Xiao, Wenqi Shao, Wenwei Zhang, Kai Chen, and Ping Luo. Gpt4roi: Instruction tuning large language model on region-of-interest. arXiv preprint arXiv:2307.03601, 2023. 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 175, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 205 + ], + "type": "text", + "content": "[90] Bo Zhao, Boya Wu, and Tiejun Huang. Svit: Scaling up visual instruction tuning. arXiv preprint arXiv:2307.04087, 2023.5,3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "type": "text", + "content": "[91] Yang Zhao, Zhijie Lin, Daquan Zhou, Zilong Huang, Jiashi Feng, and Bingyi Kang. Bubogpt: Enabling visual grounding in multi-modal llms. arXiv preprint arXiv:2307.08581, 2023. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 253, + 545, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 545, + 317 + ], + "type": "text", + "content": "[92] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16793-16803, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 319, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 319, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 307, + 319, + 545, + 363 + ], + "type": "text", + "content": "[93] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 365, + 545, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 407 + ], + "type": "text", + "content": "[94] Yuke Zhu, Oliver Groth, Michael Bernstein, and Li Fei-Fei. Visual7w: Grounded question answering in images. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4995–5004, 2016. 5, 6, 3" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14238" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/GS-IR_ 3D Gaussian Splatting for Inverse Rendering/d0b63333-3b55-4d7f-93c1-cfbd4adae3a1_content_list.json b/2024/GS-IR_ 3D Gaussian Splatting for Inverse Rendering/d0b63333-3b55-4d7f-93c1-cfbd4adae3a1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4861ee10eb887bb84946bfa3ca5a54bd2b797e68 --- /dev/null +++ b/2024/GS-IR_ 3D Gaussian Splatting for Inverse Rendering/d0b63333-3b55-4d7f-93c1-cfbd4adae3a1_content_list.json @@ -0,0 +1,1707 @@ +[ + { + "type": "text", + "text": "GS-IR: 3D Gaussian Splating for Inverse Rendering", + "text_level": 1, + "bbox": [ + 217, + 130, + 754, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhihao Liang $^{1,\\ast}$ , Qi Zhang $^{2,\\ast}$ , Ying Feng $^{2}$ , Ying Shan $^{2}$ , Kui Jia $^{3,\\dagger}$", + "bbox": [ + 227, + 179, + 738, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ South China University of Technology, $^{2}$ Tencent AI Lab,", + "bbox": [ + 245, + 198, + 722, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ School of Data Science, The Chinese University of Hong Kong, Shenzhen", + "bbox": [ + 186, + 215, + 782, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "eezhihaoliang@mail.scut.edu.cn, nwpuqzhang@gmail.com,", + "bbox": [ + 245, + 237, + 720, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "vonyfeng@gmail.com, yingsshan@tencent.com, kuijia@cuhk.edu.cn", + "bbox": [ + 202, + 255, + 759, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 333, + 313, + 349 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose GS-IR, a novel inverse rendering approach based on 3D Gaussian Splatting (3DGS) that leverages forward mapping volume rendering to achieve photorealistic novel view synthesis and relighting results. Unlike previous works that use implicit neural representations and volume rendering (e.g. NeRF), which suffer from low expressive power and high computational complexity, we extend 3DGS, a top-performance representation for novel view synthesis, to estimate scene geometry, surface material, and environment illumination from multi-view images captured under unknown lighting conditions. There are two main problems when introducing 3DGS to inverse rendering: 1) 3DGS does not support producing plausible normal natively; 2) forward mapping (e.g. rasterization and splatting) cannot trace the occlusion like backward mapping (e.g. ray tracing). To address these challenges, our GS-IR proposes an efficient optimization scheme incorporating a depth-derivation-based regularization for normal estimation and a baking-based occlusion to model indirect lighting. The flexible and expressive 3DGS representation allows us to achieve fast and compact geometry reconstruction, photorealistic novel view synthesis, and effective physically-based rendering. We demonstrate the superiority of our method over baseline methods through qualitative and quantitative evaluations of various challenging scenes. The source code is available at https://github.com/lzhnb/GS-IR.", + "bbox": [ + 75, + 366, + 473, + 773 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 780, + 209, + 795 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Inverse rendering is a long-standing task, seeking to answer the question: \"How can we deduce physical attributes (e.g. geometry, material, and lighting) of a 3D scene from multiview images?\" This problem is inherently challenging and", + "bbox": [ + 75, + 805, + 468, + 867 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f7d7891f2e1c9fb294c30870f7ea42156d5f26935de4c8f00b8fde40d7d9910a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 303, + 890, + 424 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/85913a2826953e4b36701c8b55b70f16eadf47508bed410b937ef3e389a60a27.jpg", + "image_caption": [ + "Figure 1. Given multi-view captured images of a complex scene, we propose GS-IR (3D Gaussian Splatting for Inverse Rendering), which utilizes 3D Gaussian and forward mapping splatting to recover high-quality physical properties (e.g., normal, material, illumination). This enables us to perform relighting and material editing, resulting in outstanding inverse rendering results. Better viewed on screen with zoom in, especially the remarkable material decomposition and normal reconstruction of bicycle axle." + ], + "image_footnote": [], + "bbox": [ + 506, + 426, + 750, + 487 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bb1d0adc675c004c79921e17e488466f8e935f6e5053f2e95f2521b0cd04bb26.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 426, + 888, + 487 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ill-posed, particularly when input images are captured in uncontrolled environments with unknown illumination. Recent research [9, 10, 35, 46] has sought to address this issue by employing implicit neural representations akin to NeRF [30] that utilizes multi-layer perceptrons (MLPs). However, current methods incorporating MLP face challenges in terms of their low expressive capacity and high computational demands, which significantly limits the effectiveness and efficiency of inverse rendering, especially when it cannot be rendered at interactive rates.", + "bbox": [ + 496, + 625, + 893, + 775 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D Gaussian Splatting (3DGS) [24] has recently emerged as a promising technique to model 3D static scenes and significantly boost the rendering speed to a real-time level. It makes the scene representation more compact and achieves fast and top performance for novel view synthesis. Introducing it to the inverse rendering pipeline is natural and essential, including geometry reconstruction, materials decomposition, and illumination estimation. Unlike ray trac", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* indicates equal contribution.", + "bbox": [ + 99, + 875, + 263, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{\\dagger}$ Correspondence to Kui Jia .", + "bbox": [ + 99, + 887, + 382, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21644", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing in NeRF, 3DGS produces a set of 3D Gaussians around sparse points. During the 3DGS optimization, the adaptive control of the Gaussian density may lead to loose geometry, making it difficult to estimate accurate scene's normal. Consequently, it is necessary to introduce a well-designed strategy to regularize 3DGS's normal estimation.", + "bbox": [ + 75, + 90, + 470, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our goal is to use 3D Gaussians as the scene representation for inverse rendering from multi-view images captured under unknown lighting conditions. However, capturing observations under natural illumination often shows complex effects such as soft shadows and interreflections. TensoIR [22] leverages the ray tracing of NeRF to directly model occlusion and indirect illumination. In contrast, 3DGS replaces the ray tracing in the NeRF with differentiable forward mapping volume rendering, which directly projects 3D Gaussians onto the 2D plane. This strategy improves the rendering efficiency but makes it difficult to calculate occlusion. Inspired by the \"Indirect Lighting Cache\" used in real-time rendering [4], we attempt to bake the occlusion into volumes for caching.", + "bbox": [ + 75, + 191, + 470, + 402 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we present a novel 3D Gaussian-based inverse rendering framework called GS-IR (3D Gaussian Splatting for Inverse Rendering) that leverages forward mapping splatting to deduce the physical attributes of a complex scene. To the best of our knowledge, our method is the first work to introduce the 3DGS technique for inverse rendering, which can simultaneously estimate scene geometry, materials, and illumination from multi-view images. Our GS-IR addresses two main issues when using 3DGS for inverse rendering. Firstly, we develop an intuitive and well-designed regularization to estimate the scene's normal. Secondly, we use a baking-based method embedded in GS-IR to cache occlusions, obtaining an efficient indirect illumination model. As shown in Fig. 1, our approach can reconstruct high-fidelity geometry and materials of a complex real scene under unknown natural illumination, enabling state-of-the-art rendering of novel view synthesis and additional applications like relighting. Our technical contributions are summarized as follows:", + "bbox": [ + 75, + 412, + 468, + 698 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present GS-IR that models a scene as a set of 3D Gaussians to achieve physically-based rendering and state-of-the-art decomposition results for both objects and scenes;", + "- We propose an efficient optimization scheme with regularization to concentrate depth gradient around 3DGS and produce reliable normals for GS-IR;", + "- We develop a baking-based method embedded in GS-IR to handle the occlusion in modeling indirect lighting;" + ], + "bbox": [ + 76, + 709, + 467, + 829 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We demonstrate the superiority of our method to baseline methods qualitatively and quantitatively on various challenging scenes, including the TensoIR synthesis dataset [22] and Mip-NeRF 360 real dataset [5].", + "bbox": [ + 75, + 839, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 89, + 648, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural Representation Recently, neural rendering techniques, exemplified by Neural Radiance Field (NeRF) [30], have achieved impressive success in addressing visual computing problems, giving rise to numerous neural representations [14, 21, 31, 36, 39, 40] tailored for different tasks [11-13, 18, 27, 28, 34, 43]. The vanilla NeRF models a continuous radiance field implicitly in MLPs, which requires massive repeated queries during training and inference. To address the computational inefficiencies, many neural scene representations are proposed with more discretized geometry proxies such as voxel grids [17, 20, 36], hash grids [31], tri-planes [14] or points [39]. Neural features are stored in a structured manner, allowing for efficient storage and retrieval. The computational cost can thus be significantly reduced by introducing interpolation techniques, however, with an inevitable loss in image quality. 3D Gaussians are introduced as an unstructured scene representation to strike a balance between efficiency and quality [24]. With the specially designed tile-based rasterizer for Gaussian splats, this method achieves real-time rendering with high quality for novel-view synthesis. In this work, 3D Gaussian representation is combined with the physical-based rendering (PBR) model for inverse rendering.", + "bbox": [ + 496, + 114, + 890, + 460 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inverse Rendering Inverse rendering aims to decompose the image's appearance into the geometry, material, and lighting conditions. Considering the inherent ambiguity between observed images and underlying properties, many methods are proposed with different constrained settings, such as capturing images with fixed lighting and rotating object [16, 38], capturing with moving camera and colocated lighting [7, 8, 29, 33]. Combined with neural representations, inverse rendering models the scene simulating how the light interacts with the neural volume with various material properties, and estimates the lighting and material parameters during optimization [6, 9, 10, 19, 22, 35, 42, 44, 46, 47]. Neural Reflectance Fields [6] assumes a known point light source and represents the scene as a field of volume density, surface normals, and bi-directional reflectance distribution functions(BRDFs) with one bounce direct illumination. NeRV [35] and InvRender [47] extend to arbitrary known lighting conditions and train an additional MLP to model the light visibility. PhySG [42] assumes full light source visibility without shadow simulation, and represents the lighting and scene BRDFs with spherical Gaussians for acceleration. TensoIR [22] adopts the efficient TensoRF [14] representation which enables the computation of visibility and indirect lighting by raytracing, while limited to object-level. For modeling surface geometry using point clouds, Fuzzy Metaballs (FMs) [25, 26] offers a great way to render depth from 3D Gaussian using Order Independent Transparency (OIT) and approximate intersection. However, it requires silhouettes as input and struggles to handle", + "bbox": [ + 496, + 463, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21645", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "intricate geometry (e.g. Lego and Ficus) let alone complex scenes. In this work, we propose a 3DGS-based pipeline to recover the geometry, material, and lighting that is available for both objects and unbounded scenes.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminary", + "text_level": 1, + "bbox": [ + 76, + 162, + 202, + 180 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we give the technical backgrounds and math symbols that are necessary for the presentation of our proposed method in subsequent sections.", + "bbox": [ + 75, + 186, + 468, + 232 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Gaussian Splatting (3DGS) [24] is an explicit 3D scene representation in the form of point clouds. Each point is represented as a Gaussian function $g$ that approximates the shape of a bell curve, which is defined as,", + "bbox": [ + 75, + 233, + 468, + 292 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ng (\\boldsymbol {x} | \\boldsymbol {\\mu}, \\boldsymbol {\\Sigma}) = e ^ {- \\frac {1}{2} (\\boldsymbol {x} - \\boldsymbol {\\mu}) ^ {T} \\boldsymbol {\\Sigma} ^ {- 1} (\\boldsymbol {x} - \\boldsymbol {\\mu})}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 160, + 297, + 468, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\pmb{\\mu} \\in \\mathbb{R}^3$ is its mean vector, and $\\pmb{\\Sigma} \\in \\mathbb{R}^{3 \\times 3}$ is an anisotropic covariance matrix. The mean vector $\\pmb{\\mu}$ of a 3D Gaussian is parameterized as $\\pmb{\\mu} = (\\mu_x, \\mu_y, \\mu_z)$ , and the covariance matrix $\\pmb{\\Sigma}$ is factorized into a scaling matrix $\\pmb{S}$ and a rotation matrix $\\pmb{R}$ as $\\pmb{\\Sigma} = R\\pmb{S}\\pmb{S}^\\top\\pmb{R}^\\top$ . $\\pmb{S}$ and $\\pmb{R}$ refer to a diagonal matrix $\\mathrm{diag}(s_x, s_y, s_z)$ and a rotation matrix constructed from a unit quaternion $\\pmb{q}$ . Given a viewing transformation with extrinsic matrix $\\pmb{T}$ and intrinsic matrix $\\pmb{K}$ , the mean vector $\\pmb{\\mu}$ and covariance matrix $\\pmb{\\Sigma}'$ from the 3D point $\\pmb{x}$ to 2D pixel $\\pmb{u}$ is defined as,", + "bbox": [ + 75, + 316, + 468, + 469 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\mu} ^ {\\prime} = \\boldsymbol {K} \\boldsymbol {T} [ \\boldsymbol {\\mu}, 1 ] ^ {\\top}, \\boldsymbol {\\Sigma} ^ {\\prime} = \\boldsymbol {J} \\boldsymbol {T} \\boldsymbol {\\Sigma} \\boldsymbol {T} ^ {\\top} \\boldsymbol {J} ^ {\\top}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 472, + 468, + 489 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\pmb{J}$ is the Jacobian matrix of the affine approximation of the perspective projection. Besides, each Gaussian represents the view-dependent color $c_{i}$ via a set of coefficients of spherical harmonics (SH), which is then multiplied by opacity $\\alpha$ for volume rendering. We finally obtain the color $\\hat{\\pmb{C}}$ at pixel $\\pmb{u}$ based on Eq. (1) and Eq. (2),", + "bbox": [ + 75, + 491, + 468, + 582 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {C}} = \\sum_ {i \\in N} T _ {i} g _ {i} \\left(\\boldsymbol {u} \\mid \\boldsymbol {\\mu} ^ {\\prime}, \\boldsymbol {\\Sigma} ^ {\\prime}\\right) \\alpha_ {i} \\boldsymbol {c} _ {i}, \\quad T _ {i} = \\prod_ {j = 1} ^ {i - 1} \\left(1 - g _ {j} \\left(\\boldsymbol {u} \\mid \\boldsymbol {\\mu} ^ {\\prime}, \\boldsymbol {\\Sigma} ^ {\\prime}\\right) \\alpha_ {j}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 587, + 468, + 635 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where accumulated transmittance $T_{i}$ quantifies the probability density of $i$ -th Gaussian at pixel $\\mathbf{u}$ .", + "bbox": [ + 75, + 633, + 467, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The Rendering Equation In GS-IR, we leverage the classic rendering equation to formulate the outgoing radiance of a surface point $\\mathbf{x}$ with normal $\\mathbf{n}$ :", + "bbox": [ + 75, + 662, + 468, + 708 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {o} (\\boldsymbol {x}, \\boldsymbol {v}) = \\int_ {\\Omega} L _ {i} (\\boldsymbol {x}, \\boldsymbol {l}) f _ {r} (\\boldsymbol {l}, \\boldsymbol {v}) (\\boldsymbol {l} \\cdot \\boldsymbol {n}) d \\boldsymbol {l}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 713, + 468, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\Omega$ denotes the upper hemisphere centered at $\\pmb{x}$ , $\\pmb{l}$ and $\\pmb{v}$ denote incident and view directions respectively. $L_{i}(\\pmb{x},\\pmb{l})$ denotes the radiance received at $\\pmb{x}$ from $\\pmb{l}$ . Notably, we follow Cook-Torrance microfacet model [15, 37] and formulate the bidirectional reflectance distribution function (BRDF) $f_{r}$ as a function of albedo $\\pmb{a} \\in [0,1]^3$ , metallic $m \\in [0,1]$ , and roughness $\\rho \\in [0,1]$ :", + "bbox": [ + 75, + 744, + 468, + 851 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {r} (\\boldsymbol {l}, \\boldsymbol {v}) = \\underbrace {(1 - m) \\frac {\\boldsymbol {a}}{\\pi}} _ {\\text {d i f f u s e}} + \\underbrace {\\frac {D F G}{4 (\\boldsymbol {n} \\cdot \\boldsymbol {l}) (\\boldsymbol {n} \\cdot \\boldsymbol {v})}} _ {\\text {s p e c u l a r}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 854, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where microfacet distribution function $D$ , Fresnel reflection $F$ , and geometric shadowing factor $G$ are related to the surface roughness $\\rho$ . We use 3D Gaussians to store these material properties in GS-IR.", + "bbox": [ + 498, + 90, + 890, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 498, + 164, + 589, + 180 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a set of calibrated RGB images $\\{I_m\\}_{m=1}^M$ of a target scene captured from multiple views under static, yet unknown illumination, inverse rendering aims to decompose the scene's intrinsic properties, including normal, materials, and illumination. This decomposition facilitates the recovery and subsequent edition of the target scene. Motivated by the remarkable performance in quality and speed of 3DGS [24], we present a novel framework GS-IR consisting of three well-designed stage strategies, as shown in Fig. 2. In the initial stage, we leverage differentiable splating to optimize 3D Gaussians. Concurrently, we utilize the gradient derived from the rendered depth map to supervise the normal stored in 3D Gaussians (cf. Sec. 4.1). In the second stage, we precompute the occlusion based on the learned geometric information (i.e. depth and normal) and store it in an efficient spherical harmonics-based architecture to model indirect illumination (cf. Sec. 4.2). In the final stage, we combine a differentiable splating with the physical-based rendering (PBR) pipeline to optimize illumination and material-aware 3D Gaussians (cf. Sec. 4.3).", + "bbox": [ + 496, + 186, + 890, + 492 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Normal Reconstruction", + "text_level": 1, + "bbox": [ + 498, + 500, + 715, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "During the initial stage, we optimize 3D Gaussians for geometry reconstruction from observed images, denoted as $\\mathcal{G}$ . The optimized $\\mathcal{G}$ functions as a geometric proxy for surface points and their corresponding normals $\\pmb{n}$ , which are crucial for successful inverse rendering. As highlighted in Sec. 1, generating reasonable normals within the 3DGS-based framework poses a significant challenge. To address this obstacle, we introduce an intuitive strategy that improves depth $\\hat{D}$ and leverages the depth gradient to derive pseudo normals $\\hat{\\pmb{n}}_{\\hat{D}} = \\nabla_{\\mathrm{uv}}\\hat{D}$ . These pseudo normals then guide the optimization of normals within the 3D Gaussians.", + "bbox": [ + 496, + 523, + 890, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Depth Generation Given a pretrained 3D Gaussians $\\mathcal{G}$ and a view designated for rendering, the pixel's shading results in that view can be obtained by Eq. (3). Consequently, it is reasonable to utilize the same volumetric accumulation to compute the depth $\\hat{D} = \\sum_{i=1}^{N} T_i \\alpha_i d_i$ , where $d_i$ denotes the distance from the corresponding 3D Gaussian to the image plane. However, we observed the floating problem during volumetric accumulation, unlike the backward mapping volume rendering used in NeRF. During the 3DGS optimization, the adaptive control of the Gaussian density may result in the depth falling in front of the 3D Gaussians, thereby posing challenges in accurately predicting the depth. Specifically, the backward mapping methods can obtain an accurate depth by considering only peak samples,", + "bbox": [ + 496, + 688, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21646", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/520935b331613e8ca20236530cf0e59333c0c3f429eea69a2269ca4991750034.jpg", + "image_caption": [ + "Figure 2. GS-IR Pipeline. We propose a novel GS-based inverse rendering framework, called GS-IR, to reconstruct scene geometry, materials, and unknown natural illumination from multi-view captured images. Our GS-IR consists of three well-designed stage strategies using 3D Gaussian and differentiable forward mapping splatting to achieve physical-based rendering. In our approach, the Gaussian stores not only the basic 3DGS information but also the normal and material properties, enhancing its capabilities for inverse rendering tasks." + ], + "image_footnote": [], + "bbox": [ + 101, + 88, + 869, + 415 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ca483dfb406e9c50d41bba16dfb9a89fa5f2196c656b2ec5fc2847d0a156fea0.jpg", + "image_caption": [ + "Figure 3. Depth Illustration. By considering the depth as a linear interpolation of the distances from 3D Gaussians to the image plane, and ensuring it lies between the minimum and maximum distance, our method could produce accurate depth." + ], + "image_footnote": [], + "bbox": [ + 101, + 493, + 447, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "that is $\\hat{D} = d_{i^*}$ , where $i^* = \\arg \\max_i T_i \\alpha_i$ . However, for 3DGS, a typical forward mapping method, the peak selection results in disc aliasing within 3DGS. To overcome this limitation, we consider that depth $\\hat{D}$ must be between the minimum and maximum distance of 3D Gaussians to the image plane, as illustrated in Fig. 3. We then treat the depth as a linear interpolation of the distances from 3D Gaussians to the image plane:", + "bbox": [ + 75, + 669, + 470, + 790 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {D} = \\sum_ {i = 1} ^ {N} \\hat {w} _ {i} d _ {i}, \\quad \\hat {w} _ {i} = \\frac {T _ {i} \\alpha_ {i}}{\\sum_ {i = 1} ^ {N} T _ {i} \\alpha_ {i}}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 799, + 468, + 837 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Normal Derivation While the accurate prediction of depth within Gaussians provides better guidance for the normal reconstruction, directly using depth gradient to produce normals has two limitations that still cannot meet the require", + "bbox": [ + 76, + 839, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ment for effective inverse rendering. First, the depth gradient estimation is highly sensitive to noise, making the predicted normal often extremely noisy; Second, the normals derived individually from each view's depth map do not satisfy multi-view consistency. To address these issues, we use Gaussian $\\mathcal{G}$ as a proxy for normal estimation instead of directly from the depth gradient. Benefiting from the efficiency of 3DGS, we obtain the depth $\\hat{D}$ and normals $\\hat{\\pmb{n}}$ of the observed view after performing a single rendering pass. We then tie these predicted pseudo normal to the underlying depth gradient normal $\\hat{\\pmb{n}}_{\\hat{D}}$ using a simple penalty:", + "bbox": [ + 496, + 496, + 893, + 665 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {n - p} = \\left\\| \\hat {\\boldsymbol {n}} - \\hat {\\boldsymbol {n}} _ {\\hat {D}} \\right\\|, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 632, + 670, + 890, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{\\pmb{n}} = \\sum_{i=1}^{N} T_i \\alpha_i \\pmb{n}_i$ , and $\\pmb{n}_i$ is the normal stored in the 3D Gaussian. Secondly, unlike the MLP-based normal estimation [22] acting MLP as a low-pass filter, the predicted normal of Gaussian $\\mathcal{G}$ is rough, so smoothness regularization should be included. We introduce the TV term $TV_{\\text{normal}}$ to smooth the predicted normal $\\hat{\\pmb{n}}$ . For more details, please refer to the supplement.", + "bbox": [ + 496, + 689, + 890, + 796 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In optimization of the first stage, we optimize 3D Gaussians $\\mathcal{G}$ (storing SH coefficients for view-dependent color $c$ , opacity $\\alpha$ , and normal $\\hat{n}$ ) by using the color reconstruction loss $\\mathcal{L}_c$ , which is the same as 3DGS [24], and the proposed normal loss $\\mathcal{L}_n$ ,", + "bbox": [ + 496, + 797, + 890, + 871 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {n} = \\mathcal {L} _ {n - p} + \\lambda_ {n - T V} T V _ {\\text {n o r m a l}}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 878, + 890, + 895 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21647", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/122d068955408441c595b1a88809c97880f2d58b74013a64e5351d41b7f52f34.jpg", + "image_caption": [ + "Figure 4. Baking. We employ the spherical harmonics (SH) architecture to bake occlusion volumes for modeling indirect illumination. For each grid of occlusion volumes, we initially use 3D Gaussians to compute the depth cubemap by performing six forward mapping splatting passes. Next, we convert the depth cubemap into a binary occlusion cubemap based on a distance threshold. Finally, the occlusion cubemap is baked as SH coefficients, enabling efficient interpolation of the occlusion cubemap at any point within the scene." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 450, + 268 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Indirect Illumination Modeling", + "text_level": 1, + "bbox": [ + 76, + 415, + 354, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Drawing inspiration from the successful implementation of precomputation techniques in the video game industry (e.g. Irradiance Volume in Blender [3], Lightmass Volume in Unreal [1], and Light Probes in Unity [2]), we introduce spherical harmonics (SH) architectures to store occlusion information and model indirect illumination.", + "bbox": [ + 75, + 441, + 468, + 531 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the optimized 3D Gaussians $\\mathcal{G}$ from one stage (cf. Sec. 4.1), we freeze $\\mathcal{G}$ and regularly place occlusion volumes $\\nu^{\\mathrm{occl}}$ in the 3D space. For each volume $\\pmb{v}_i^{\\mathrm{occl}}\\subset \\nu^{\\mathrm{occl}}$ we then cache the occlusion in the form of SH coefficients $\\pmb{f}_i^{\\mathrm{o}}$ . Consequently, the formula of the occlusion $O(\\cdot)$ of $\\pmb{v}_i^{\\mathrm{occl}}$ with respect to the direction $(\\theta ,\\phi)$ is expressed as:", + "bbox": [ + 76, + 534, + 468, + 626 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nO (\\theta , \\phi) = \\sum_ {l = 0} ^ {\\text {d e g}} \\sum_ {m = - l} ^ {l} \\mathbf {f} _ {i (l m)} ^ {0} Y _ {l m} (\\theta , \\phi), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 643, + 468, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $deg$ denotes the degree of SH, and $\\{Y_{lm}(\\cdot)\\}$ is a set of real basis of SH.", + "bbox": [ + 75, + 696, + 468, + 727 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As discussed in Sec. 4.1, the 3DGS technique employs a forward mapping approach that projects 3D points to the 2D plane, in contrast to the backward mapping volume rendering utilized in NeRF, which means it cannot use ray marching to calculate occlusions. To precompute the SH coefficients $\\pmb{f}_i^{\\mathrm{o}}$ of occlusion volume $\\pmb{v}_i^{\\mathrm{occl}}$ , we obtain the depth cubemap $\\{\\hat{D}_p^i\\}_{p=1}^6$ by performing six times rendering passes, once for each face of the cubemap. We then convert it into a binary occlusion cubemap $\\{\\hat{O}_p^i\\}_{p=1}^6$ based on a manually set distance threshold.", + "bbox": [ + 75, + 729, + 470, + 882 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we convolve the the occlusion cubemap", + "bbox": [ + 96, + 885, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\{\\hat{O}_p^i\\}_{p = 1}^6$ using SH bases and get the SH coefficients $\\pmb{f}_i^0$ ..", + "bbox": [ + 500, + 89, + 875, + 109 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\boldsymbol {f} _ {i (l m)} ^ {0} = \\int_ {S ^ {2}} \\hat {\\boldsymbol {O}} (\\boldsymbol {\\omega}) Y _ {l m} (\\boldsymbol {\\omega}) d \\boldsymbol {\\omega} \\tag {10} \\\\ = \\int_ {0} ^ {2 \\pi} \\int_ {0} ^ {\\pi} \\sin \\theta \\hat {\\boldsymbol {O}} (\\theta , \\phi) Y _ {l m} (\\theta , \\phi) d \\theta d \\phi , \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 112, + 890, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $S^2$ denotes the unit sphere, and $\\hat{O} (\\theta ,\\phi)$ denotes the occlusion query from the occlusion cubemap $\\{\\hat{O}_p^i\\}_{p = 1}^6$ Note that we numerically calculate the convolution in Eq. (10) in parallel. The caching process is shown in Fig. 4.", + "bbox": [ + 496, + 170, + 890, + 232 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To handle the indirect illumination in occlusion regions, we also maintain illumination volumes $\\mathcal{V}^{\\mathrm{illu}}$ to cache the indirect illumination. Similar to the caching process of occlusion volumes, the caching target of illumination volumes changes from the occlusion cubemap $\\{\\hat{O}_p^i\\}_{p=1}^6$ to the captured environment cubemap $\\{\\hat{I}_p^i\\}_{p=1}^6$ . These cubemaps can be obtained simultaneously by conducting six rendering passes. For more details, please refer to the supplement.", + "bbox": [ + 496, + 233, + 892, + 356 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Intrinsic Decomposition", + "text_level": 1, + "bbox": [ + 498, + 364, + 720, + 382 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the final stage, we employ differentiable splatting in conjunction with a PBR pipeline to accomplish the intrinsic decomposition. According to Eq. (5), the rendering equation Eq. (4) is rewritten as diffuse $L_{\\mathrm{d}}$ and specular $L_{\\mathrm{s}}$ components:", + "bbox": [ + 496, + 388, + 890, + 462 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {o} (\\boldsymbol {x}, \\boldsymbol {v}) = \\int_ {\\Omega} \\left[ (1 - m) \\frac {\\boldsymbol {a}}{\\pi} + \\frac {D F G}{4 (\\boldsymbol {n} \\cdot \\boldsymbol {l}) (\\boldsymbol {n} \\cdot \\boldsymbol {v})} \\right] L _ {i} (\\boldsymbol {x}, \\boldsymbol {l}) (\\boldsymbol {l} \\cdot \\boldsymbol {n}) d \\boldsymbol {l} \\\\ L _ {\\mathrm {d}} = (1 - m) \\frac {\\boldsymbol {a}}{\\pi} \\int_ {\\Omega} L _ {i} (\\boldsymbol {x}, \\boldsymbol {l}) (\\boldsymbol {l} \\cdot \\boldsymbol {n}) d \\boldsymbol {l} \\tag {11} \\\\ L _ {\\mathrm {s}} = \\int_ {\\Omega} \\frac {D F G}{4 (\\boldsymbol {n} \\cdot \\boldsymbol {l}) (\\boldsymbol {n} \\cdot \\boldsymbol {v})} L _ {i} (\\boldsymbol {x}, \\boldsymbol {l}) (\\boldsymbol {l} \\cdot \\boldsymbol {n}) d \\boldsymbol {l}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 468, + 890, + 546 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In GS-IR, we adopt an image-based lighting (IBL) model and split-sum approximation [23] to tackle the intractable integral. To calculate the diffuse component $L_{\\mathrm{d}}$ , the illumination $I_{\\mathrm{d}}$ is defined as:", + "bbox": [ + 496, + 553, + 890, + 613 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} I _ {\\mathrm {d}} (\\boldsymbol {x}) = \\int_ {\\Omega} L _ {i} (\\boldsymbol {x}, \\boldsymbol {l}) (\\boldsymbol {l} \\cdot \\boldsymbol {n}) d \\boldsymbol {l} \\\\ = \\int_ {\\Omega_ {\\mathrm {V i s}}} L _ {i} ^ {\\operatorname {d i r}} (\\boldsymbol {x}, \\boldsymbol {l}) (\\boldsymbol {l} \\cdot \\boldsymbol {n}) d \\boldsymbol {l} + \\int_ {\\Omega_ {\\mathrm {O c c l}}} L _ {i} ^ {\\operatorname {i n d i r}} (\\boldsymbol {x}, \\boldsymbol {l}) (\\boldsymbol {l} \\cdot \\boldsymbol {n}) d \\boldsymbol {l} \\tag {12} \\\\ \\approx \\left(1 - \\mathrm {O} (\\boldsymbol {x})\\right) I _ {\\mathrm {d}} ^ {\\operatorname {d i r}} (\\boldsymbol {x}) + \\mathrm {O} (\\boldsymbol {x}) I _ {\\mathrm {d}} ^ {\\operatorname {i n d i r}} (\\boldsymbol {x}), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 617, + 890, + 690 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the first component indicates direct illumination and the second is indirect illumination. Notably, our baking-based indirect illumination model enables us to calculate the occlusion and illumination online. This means that our GS-IR achieves intrinsic decomposition while maintaining real-time rendering performance. For the specular component $L_{\\mathrm{s}}$ , we follow split-sum approximation and treat the integral as two separate integrals:", + "bbox": [ + 496, + 693, + 890, + 814 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {\\mathrm {s}} = \\int_ {\\Omega} \\frac {D F G}{4 \\langle \\boldsymbol {n} \\cdot \\boldsymbol {l} \\rangle \\langle \\boldsymbol {n} \\cdot \\boldsymbol {v} \\rangle} L _ {i} (\\boldsymbol {l}) \\langle \\boldsymbol {l} \\cdot \\boldsymbol {n} \\rangle d \\boldsymbol {l} \\\\ \\approx \\underbrace {\\int_ {\\Omega} \\frac {D F G}{4 \\left\\langle \\boldsymbol {n} \\cdot \\boldsymbol {l} \\right\\rangle \\left\\langle \\boldsymbol {n} \\cdot \\boldsymbol {v} \\right\\rangle}} _ {\\text {E n v i r o n m e n t B R D F -} R} \\left\\langle \\boldsymbol {l} \\cdot \\boldsymbol {n} \\right\\rangle d \\boldsymbol {l} \\underbrace {\\int_ {\\Omega} D L _ {i} (\\boldsymbol {l}) \\left\\langle \\boldsymbol {l} \\cdot \\boldsymbol {n} \\right\\rangle d \\boldsymbol {l}} _ {\\text {P r e - F i l t e r e d E n v i r o n m e n t M a p -} I _ {s}}, \\tag {13} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 829, + 890, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21648", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fec20d9f6c6a5555aabe002a85419f33cc928bbe4f577ec908c4c453d5265c6d.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison on TensoIR Synthetic. We visualize the estimated normal, albedo, and rendering results of our GS-IR and baseline methods on two scenes. By utilizing the efficient 3D Gaussian representation and a robust tile-based rasterizer, GS-IR achieves rapid convergence and supports real-time rendering. This performance advantage underscores the effectiveness of our method in addressing complex inverse rendering tasks, thereby surpassing existing state-of-the-art approaches. (For albedo reconstruction results, we follow NeRFactor [46] and scale each RGB channel by a global scalar.)" + ], + "image_footnote": [], + "bbox": [ + 76, + 85, + 893, + 333 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where both $R$ and $I_{\\mathrm{s}}$ can be precomputed in advance and stored in look-up tables. With Eq. (12) and Eq. (13), the rendering results of Eq. (11) can be represented as:", + "bbox": [ + 75, + 439, + 470, + 484 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {o} (\\boldsymbol {x}, \\boldsymbol {v}) = L _ {\\mathrm {d}} + L _ {\\mathrm {s}} \\\\ \\approx (1 - m) \\frac {\\boldsymbol {a}}{\\pi} \\left[ (1 - \\mathrm {O} (\\boldsymbol {x})) I _ {\\mathrm {d}} ^ {\\operatorname {d i r}} (\\boldsymbol {x}) + \\mathrm {O} (\\boldsymbol {x}) I _ {\\mathrm {d}} ^ {\\operatorname {i n d i r}} (\\boldsymbol {x}) \\right] + R I _ {\\mathrm {s}}. \\tag {14} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 488, + 468, + 536 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For intrinsic decomposition, we optimize the material $\\hat{M}$ (i.e. albedo $a$ , metallic value $m$ , and roughness $\\rho$ ) stored in 3D Gaussians $\\mathcal{G}$ , environment map $\\hat{\\pmb{E}}$ , and illumination volumes $\\mathcal{V}^{\\mathrm{illu}}$ by minimizing the decomposition loss $\\mathcal{L}_d$ :", + "bbox": [ + 76, + 537, + 468, + 599 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d} = \\underbrace {\\left\\| \\boldsymbol {I} - \\hat {\\boldsymbol {I}} ^ {\\text {s h a d e}} \\left(\\hat {\\boldsymbol {M}} , \\hat {\\boldsymbol {E}} , \\mathcal {V} ^ {\\text {i l l u}}\\right) \\right\\|} _ {\\mathcal {L} _ {\\text {s h a d e}}} + \\underbrace {\\lambda_ {M} T V _ {\\mathrm {m a t}}} _ {\\mathcal {L} _ {\\text {m a t e r i a l}}} + \\underbrace {\\lambda_ {\\boldsymbol {E}} T V _ {\\text {l i g h t}}} _ {\\mathcal {L} _ {\\text {l i g h t}}}, \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 599, + 468, + 651 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{shade}}$ indicate the shade loss. $\\hat{I}^{\\mathrm{shade}}(\\hat{M},\\hat{E},\\hat{\\mathcal{V}}^{\\mathrm{illu}})$ is the recovered image that uses the PBR pipeline defined by Eq. (14). Please refer to the supplement for more details about the material TV loss $\\mathcal{L}_{\\mathrm{material}}$ and lighting TV loss $\\mathcal{L}_{\\mathrm{light}}$ .", + "bbox": [ + 75, + 651, + 468, + 713 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 724, + 209, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset & Metrics We conduct experiments using benchmark datasets of TensoIR Synthetic [22] and Mip-NeRF 360 [5] for decomposing both objects and scenes. They contain 4 objects with reference materials and 7 publicly available scenes, respectively. To verify the efficacy of our normal reconstruction, we evaluate the normal quality on the TensoIR Synthetic [22] dataset using mean angular error (MAE). We further assess our reconstructed albedo quality on this synthetic dataset. More generally, we evaluate the synthesized novel view on both datasets in terms of Peak", + "bbox": [ + 75, + 750, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/adbe248cba838b012773c77d3e24a54dc606e0633c32d31505fa1ee3bf19c23c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 524, + 436, + 689, + 520 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0c29f848a8fb7d39cef3e865df897a311b5552f97d19c22e4630297fd7751282.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 700, + 436, + 856, + 520 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7981f9ed67a6b735312df99f5e779e5291f3247653e2de0a592343aa1f587c43.jpg", + "image_caption": [ + "Figure 6. Novel view synthesis results on Mip-NeRF 360. GS-IR can reconstruct scene details including geometric normals and high-frequency appearance, rendering high-fidelity appearance and recovering fine geometric details such as those on leaves and bicycle axles. Better viewed on screen with zoom in." + ], + "image_footnote": [], + "bbox": [ + 524, + 521, + 694, + 603 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f009744cc300cc761df0f99cd28e47e2da474757ec2ff377fada9bdc1c075103.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 700, + 521, + 870, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS) [45]. Note that albedo quality assessment uses the same metrics as novel view synthesis.", + "bbox": [ + 496, + 709, + 893, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Comparisons", + "text_level": 1, + "bbox": [ + 500, + 784, + 638, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct a comprehensive comparison against state-of-the-art neural field-based inverse rendering methods on the public TensoIR Synthetic dataset [22]. All the methods utilize multi-view images captured under unknown lighting conditions. Our evaluation encompasses normal quality (measured by MAE), novel view synthesis, albedo fidelity,", + "bbox": [ + 496, + 810, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21649", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8ed8845edb740bc015a3be856500b1080dd49dc3094270b64a24d7f23300dea7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodNormal ↓ MAE ↓Novel View SynthesisAlbedoRelightRuntime
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
NeRFactor [46]6.31424.6790.9220.12025.1250.9400.10923.3830.9080.131>100 hrs
InvRender [47]5.07427.3670.9340.08927.3410.9330.10023.9730.9010.10115 hrs
NVDiffrec [32]6.07830.6960.9620.05229.1740.9080.11519.8800.8790.104<1 hr
TensoIR [22]4.10035.0880.9760.04029.2750.9500.08528.5800.9440.0815 hrs
Ours4.94835.3330.9740.03930.2860.9410.08424.3740.8850.096<1 hr
", + "bbox": [ + 83, + 88, + 888, + 185 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/71b7a9668d53c3d808d7b6d288dfc0e716e83625e32dc65c7a9f0408cbde6fae.jpg", + "table_caption": [ + "Table 1. Quantitative Comparison on TensoIR Synthetic dataset. Our method outperforms baseline methods in terms of novel view synthesis and albedo quality, showcasing the effectiveness of material decomposition and PBR rendering. This is particularly noteworthy considering that our normal reconstruction is slightly inferior to TensoIR. In terms of relighting performance, we rank second, trailing only behind TensoIR. Importantly, the average training time of our GS-IR is accelerated by a factor of $5\\mathrm{x}$ , making its performance acceptable and further demonstrating the effectiveness of our approach in handling complex inverse rendering tasks." + ], + "table_footnote": [], + "table_body": "
MethodPSNR ↑SSIM ↑LPIPS ↓Runtime ↓
NeRF++ [41]25.1120.6960.375≈ 20h
Plenoxels [17]23.0790.6250.462≈ 30m
INGP-Base [31]25.3030.6710.371≈ 5m
INGP-Big [31]25.5870.6990.331≈ 8m
Mip-NeRF 360 [31]27.5690.7930.234≈ 48h
3DGS [24]27.210.8150.214≈ 35m
Ours25.3810.7570.267≈ 45m
", + "bbox": [ + 84, + 280, + 460, + 381 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "relighting effects (measured by PSNR, SSIM, and LPIPS), and efficiency. Tab. 1 summarizes the quantitative comparisons on the synthesis dataset. Our method achieves superior performance in novel view synthesis and albedo quality compared to the baseline methods, demonstrating the effectiveness of material decomposition and PBR rendering, particularly given that our normal reconstruction is slightly inferior to TensoIR. Our relighting performance ranks second, only behind TensoIR. Notably, the average training time of our GS-IR is accelerated by a factor of $5\\mathrm{x}$ , making its performance acceptable and demonstrating the effectiveness of our approach in handling complex inverse rendering tasks. We also include qualitative comparisons in Fig. 5, which show that our GS-IR produces reasonable albedo and photorealistic renderings that are closer to the ground truth than most methods.", + "bbox": [ + 75, + 459, + 468, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Meanwhile, owing to our more efficient and compact representation with powerful expressiveness, our method showcases remarkable performance on complex real unbounded scenes [5]. Tab. 2 presents the quantitative comparisons on the real dataset. Fig. 6 demonstrates the normal reconstruction and novel view synthesis on the real dataset. Our method renders a high-fidelity appearance and recovers fine geometric details, such as those on the leaves and bicycle axil. In summary, by leveraging the efficient 3D Gaussian representation and a powerful tile-based rasterizer, GS-IR achieves fast convergence and supports real-time rendering. This performance advantage highlights the effectiveness of our method in handling complex inverse rendering", + "bbox": [ + 75, + 704, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8339ae5be5c08400ffe63c823387ed024298440ff26d166425b38926c90df8c9.jpg", + "table_caption": [ + "Table 2. Quantitative Comparison on Mip-NeRF 360. The results show that our inverse rendering approach even surpasses some NeRF variants dedicated to novel view synthesis." + ], + "table_footnote": [], + "table_body": "
MethodTensoIR Synthetic [22]Mip-NeRF 360 [5]
Normal MAE↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Vol. Accum.16.34725.7560.8550.13122.0520.6100.394
Peak Selec.9.46628.7500.9270.08423.0930.7190.317
Linear Interp.6.21828.9830.9390.06623.1190.7070.319
Vol. Accum.†9.31530.0910.9400.07124.6640.7470.277
Peak Selec.†7.98631.0640.9500.06025.1430.7530.281
Linear Interp.†4.94835.3330.9740.03925.3810.7570.267
", + "bbox": [ + 504, + 281, + 888, + 368 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Analyses on the impact of different depth generation strategies on normals. Methods without $\\dagger$ marks directly use the normals derived from the depth map; Methods marked with $\\dagger$ use depth derivation to optimize the normals stored in 3D Gaussians.", + "bbox": [ + 498, + 372, + 892, + 429 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a4b22106eecb0f719527926e096a12f57a24dbe6c81dccde32433f0e29266c8c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 521, + 435, + 869, + 537 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/08f294b7e157d0d152b10f8dd8f7bd192431da5a1166f18478cbb891b6b016a7.jpg", + "table_caption": [ + "Figure 7. Visual comparison of depth produced by different strategies. The linear interpolation adopted in GS-IR overcomes the floating problem and disc aliasing." + ], + "table_footnote": [], + "table_body": "
MethodTensoIR Synthetic [22]Mip-NeRF 360 [5]
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
w/o occlusion34.9970.9620.04125.0600.7530.270
w/o indirect illum.35.1860.9650.04424.8980.7490.272
Ours35.3330.9740.03925.3810.7570.267
", + "bbox": [ + 504, + 582, + 883, + 636 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Analyses on the occlusion and indirect illumination. Physically modeling indirect illumination improves the inverse rendering of objects and scenes.", + "bbox": [ + 498, + 638, + 890, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tasks, outperforming existing state-of-the-art approaches.", + "bbox": [ + 500, + 700, + 877, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 726, + 663, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We initially introduce the 3DGS technique for inverse rendering in GS-IR and propose depth-derivation-based normal regularization and a baking-based method to address the challenges encountered during the process. To evaluate the efficacy of our proposed schemes, we design elaborate experiments on both TensoIR Synthetic [22] and MipNeRF 360 [5] datasets, providing comprehensive insights into the effectiveness of our approach in handling complex inverse rendering tasks. Below is the detailed ablation study on Normal Regularization and Indirect Illumination.", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21650", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/06e1ddad6069322fee46aaec46d4d73b83ac3a589b44435883ab6aac4844ac13.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 117, + 90, + 251, + 143 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/26d62ae8ab69950521c2423c6f6751c682f5cb5bf22a2798efacd7f2e1699494.jpg", + "image_caption": [ + "Figure 8. Relighting Visualization. We perform relighting experiments on both synthetic and real scenes using the recovered geometry, material, and illumination properties from our GS-IR method. We test our method under different lighting conditions and directions." + ], + "image_footnote": [], + "bbox": [ + 127, + 146, + 243, + 222 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/770b9ac4012b5565b75c44c29eedb055bb460b045d5b9948f5edc2fd4001eaa1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 267, + 98, + 400, + 143 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/40924564f4e82eb75884a7b39760ffaf3502d68ed7d3d77378de34087d413b56.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 276, + 146, + 393, + 222 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6d57db72969d48d2f949c3fe9796e4a458fe0c6d606f2a4a23d5b7d23eaedb3d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 416, + 90, + 547, + 143 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b9c42674746ef5fc3aed59a7c0e80fd2c1baae462b5123fad4059003e6c3f4b1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 146, + 542, + 222 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7b1d79b19403f75c45aa69811fc2c11d876d60a054edadf0af4641b008fdb050.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 566, + 90, + 697, + 143 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8e3aae9e547181498b68f34deddc2cd236575c22f65383ecb0f7eef436873df2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 573, + 146, + 689, + 222 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f7364b1a86329aacf374d3cf64b5294ff2be41accc5b74c3206494d9c5b9cce7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 715, + 90, + 846, + 143 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/05a653bd301302f501abb5f081f5f6f0506a8fe2181bc5001bf638f5aac93307.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 723, + 146, + 839, + 222 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4738bfbdf4bb0161b98a7a8d26f6fc3f5f82b475312889a06f42cc0fbd88aae8.jpg", + "image_caption": [ + "Figure 9. Ambient Occlusion Visualization. The visualization highlights the intricate shadowing and occlusion details captured by our GS-IR method, emphasizing the performance of our approach in modeling indirect illumination." + ], + "image_footnote": [], + "bbox": [ + 81, + 276, + 166, + 420 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/af86a9520be3f710f0f85515276be486e606e7e8b99cffe40415e6fd78b857c7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 178, + 277, + 316, + 420 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4bb38e441cdac8b70d210c4ae655298395ad7c438da73ef80c52740d1d31b4f0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 325, + 280, + 462, + 420 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis on the Normal Regularization Reliable normal estimation is critical for conducting inverse rendering. To this end, we present depth-derivation-based regularization to facilitate 3D Gaussian-based normal estimation as stated in Sec. 4.1. In this section, we explore the impact of different acquisition schemes on the final normal quality and inverse rendering results. The quantitative results shown in Tab. 3 demonstrate that using 3D Gaussians as a normal proxy and adopting the linear interpolation strategy significantly improves the normal estimation and inverse rendering results. In addition, Fig. 7 qualitatively shows that conducting volumetric accumulation results in the floating problem (cf. Fig. 7 (b)). Despite peak selection overcomes this problem, it introduces disc aliasing (cf. Fig. 7 (c)). Compared with them, the linear interpolation adopted in GS-IR robustly produces accurate depth (cf. Fig. 7 (d)).", + "bbox": [ + 75, + 506, + 468, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis on the Indirect Illumination To demonstrate the effectiveness of our indirect illumination model, we compare our method with two variants: a model without occlusion volume (w/o occlusion) and a model without indirect illumination (w/o indirect illum.). The quantitative comparisons in Tab. 4 indicate that each component is crucial for estimating accurate material decomposition and generating photorealistic rendering results. Additionally, Fig. 9 showcases the ambient occlusion visualization in both synthetic and real scenes. The visualization highlights the intra-", + "bbox": [ + 75, + 750, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "cate shadowing and occlusion details captured by our GS-IR method, emphasizing the performance of our approach in modeling indirect illumination. This analysis further supports the effectiveness of using occlusion volume and introducing indirect illumination in enhancing the decomposition capabilities of our GS-IR.", + "bbox": [ + 498, + 275, + 890, + 367 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Application", + "text_level": 1, + "bbox": [ + 500, + 383, + 625, + 400 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We perform relighting experiments using the recovered geometry, material, and illumination from our GS-IR method. We test GS-IR under different lighting conditions and directions, observing how the reconstructed scene responds to the changes in lighting. The results of these experiments demonstrate that our GS-IR method can effectively handle relighting applications, producing photorealistic renderings under various lighting conditions. More results can be found in the supplement.", + "bbox": [ + 498, + 409, + 890, + 546 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 566, + 617, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present GS-IR, a novel inverse rendering approach based on 3D Gaussian Splatting (3DGS), which employs forward mapping volume rendering to achieve photorealistic novel view synthesis and relighting results. Our GS-IR proposes an optimization scheme with depth-derivation-based regularization for normal estimation and a baking-based occlusion to model indirect lighting. These components are eventually employed to decompose material and illumination. Our extensive experiments demonstrate the effectiveness of GS-IR in achieving state-of-the-art inverse rendering results, surpassing previous neural methods in terms of both reconstruction quality and efficiency.", + "bbox": [ + 496, + 595, + 890, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitation Spherical Harmonics (SH) is only suitable for representing low-frequency, and we only use the occlusion represented by SH to model the diffuse term of indirect illumination. Modeling the specular term of indirect illumination remains a limitation of GS-IR, and has been a challenging problem in computer graphics. We believe it would be valuable to address this limitation in future work and suggest screen space global illumination (SSGI) techniques.", + "bbox": [ + 496, + 779, + 890, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21651", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Unreal engine. https://www.unrealengine.com/. 5", + "[2] Unity. https://unity.com/.5", + "[3] Blender. https://www.blender.org/.5", + "[4] Tomas Akenine-Mo, Eric Haines, Nate Hoffman, et al. Real-time rendering. 2018. 2", + "[5] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 2, 6, 7", + "[6] Sai Bi, Zexiang Xu, Pratul Srinivasan, Ben Mildenhall, Kalyan Sunkavalli, Miloš Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Neural reflectance fields for appearance acquisition. arXiv preprint arXiv:2008.03824, 2020. 2", + "[7] Sai Bi, Zexiang Xu, Kalyan Sunkavalli, Miloš Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Deep reflectance volumes: Relightable reconstructions from multi-view photometric images. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part III 16, pages 294–311. Springer, 2020. 2", + "[8] Sai Bi, Zexiang Xu, Kalyan Sunkavalli, David Kriegman, and Ravi Ramamoorthi. Deep 3d capture: Geometry and reflectance from sparse multi-view images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5960-5969, 2020. 2", + "[9] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T Barron, Ce Liu, and Hendrik Lensch. Nerd: Neural reflectance decomposition from image collections. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12684-12694, 2021. 1, 2", + "[10] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan Barron, and Hendrik Lensch. Neural-pil: Neural pre-integrated lighting for reflectance decomposition. Advances in Neural Information Processing Systems, 34: 10691-10704, 2021. 1, 2", + "[11] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 130-141, 2023. 2", + "[12] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5799-5809, 2021.", + "[13] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 2", + "[14] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European" + ], + "bbox": [ + 78, + 114, + 467, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Computer Vision, pages 333-350. Springer, 2022. 2", + "[15] Robert L Cook and Kenneth E. Torrance. A reflectance model for computer graphics. ACM Transactions on Graphics (ToG), 1(1):7-24, 1982. 3", + "[16] Yue Dong, Guojun Chen, Pieter Peers, Jiawan Zhang, and Xin Tong. Appearance-from-motion: Recovering spatially varying surface reflectance under unknown lighting. ACM Transactions on Graphics (TOG), 33(6):1-12, 2014. 2", + "[17] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 2, 7", + "[18] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbek Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12479–12488, 2023. 2", + "[19] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, light, and material decomposition from images using monte carlo rendering and denoising. Advances in Neural Information Processing Systems, 35:22856-22869, 2022. 2", + "[20] Peter Hedman, Pratul P Srinivasan, Ben Mildenhall, Jonathan T Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5875-5884, 2021. 2", + "[21] Zhangjin Huang, Zhihao Liang, Haojie Zhang, Yangkai Lin, and Kui Jia. Sur2f: A hybrid representation for high-quality and efficient surface reconstruction from multi-view images. arXiv preprint arXiv:2401.03704, 2024. 2", + "[22] Haian Jin, Isabella Liu, Peijia Xu, Xiaoshuai Zhang, Songfang Han, Sai Bi, Xiaowei Zhou, Zexiang Xu, and Hao Su. Tensoir: Tensorial inverse rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 165-174, 2023. 2, 4, 6, 7", + "[23] Brian Karis and Epic Games. Real shading in unreal engine 4. Proc. Physically Based Shading Theory Practice, 4(3):1, 2013. 5", + "[24] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 42(4):1-14, 2023. 1, 2, 3, 4, 7", + "[25] Leonid Keselman and Martial Hebert. Approximate differentiable rendering with algebraic surfaces. In European Conference on Computer Vision, pages 596-614. Springer, 2022. 2", + "[26] Leonid Keselman and Martial Hebert. Flexible techniques for differentiable rendering with 3d gaussians. arXiv preprint arXiv:2308.14737, 2023. 2", + "[27] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21652", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Zhihao Liang, Zhangjin Huang, Changxing Ding, and Kui Jia. Helixsurf: A robust and efficient neural implicit surface learning of indoor scenes with iterative intertwined regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13165-13174, 2023. 2", + "[29] Fujun Luan, Shuang Zhao, Kavita Bala, and Zhao Dong. Unified shape and svbrdf recovery using differentiable monte carlo rendering. In Computer Graphics Forum, pages 101-113. Wiley Online Library, 2021. 2", + "[30] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision, pages 405-421, 2020. 1, 2", + "[31] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 2, 7", + "[32] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3d models, materials, and lighting from images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8280-8290, 2022. 7", + "[33] Giljoo Nam, Joo Ho Lee, Diego Gutierrez, and Min H Kim. Practical svbrdf acquisition of 3d objects with unstructured flash photography. ACM Transactions on Graphics (TOG), 37(6):1-12, 2018. 2", + "[34] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 2", + "[35] Pratul P Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T Barron. Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7495-7504, 2021. 1, 2", + "[36] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 2", + "[37] Bruce Walter, Stephen R Marschner, Hongsong Li, and Kenneth E Torrance. Microfacet models for refraction through rough surfaces. In Proceedings of the 18th Eurographics conference on Rendering Techniques, pages 195-206, 2007. 3", + "[38] Rui Xia, Yue Dong, Pieter Peers, and Xin Tong. Recovering shape and spatially-varying surface reflectance under unknown illumination. ACM Transactions on Graphics (TOG), 35(6):1-12, 2016. 2", + "[39] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5438-5448, 2022. 2", + "[40] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems, 33:2492-2502, 2020. 2", + "[41] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 7", + "[42] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snavely. Physg: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5453-5462, 2021. 2", + "[43] Kai Zhang, Nick Kolkin, Sai Bi, Fujun Luan, Zexiang Xu, Eli Shechtman, and Noah Snavely. Arf: Artistic radiance fields. In European Conference on Computer Vision, pages 717-733. Springer, 2022. 2", + "[44] Kai Zhang, Fujun Luan, Zhengqi Li, and Noah Snavely. Iron: Inverse rendering by optimizing neural sdfs and materials from photometric images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5565-5574, 2022. 2", + "[45] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6", + "[46] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (ToG), 40(6):1-18, 2021. 1, 2, 6, 7", + "[47] Yuanqing Zhang, Jiaming Sun, Xingyi He, Huan Fu, Rongfei Jia, and Xiaowei Zhou. Modeling indirect illumination for inverse rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18643-18652, 2022. 2, 7" + ], + "bbox": [ + 501, + 92, + 890, + 641 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21653", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/GS-IR_ 3D Gaussian Splatting for Inverse Rendering/d0b63333-3b55-4d7f-93c1-cfbd4adae3a1_model.json b/2024/GS-IR_ 3D Gaussian Splatting for Inverse Rendering/d0b63333-3b55-4d7f-93c1-cfbd4adae3a1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1aa36680f0e2755a4155e47f7d8cbb3a68710b51 --- /dev/null +++ b/2024/GS-IR_ 3D Gaussian Splatting for Inverse Rendering/d0b63333-3b55-4d7f-93c1-cfbd4adae3a1_model.json @@ -0,0 +1,2255 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.131, + 0.755, + 0.154 + ], + "angle": 0, + "content": "GS-IR: 3D Gaussian Splating for Inverse Rendering" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.18, + 0.74, + 0.199 + ], + "angle": 0, + "content": "Zhihao Liang\\(^{1,\\ast}\\), Qi Zhang\\(^{2,\\ast}\\), Ying Feng\\(^{2}\\), Ying Shan\\(^{2}\\), Kui Jia\\(^{3,\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.246, + 0.199, + 0.723, + 0.217 + ], + "angle": 0, + "content": "\\(^{1}\\)South China University of Technology, \\(^{2}\\)Tencent AI Lab," + }, + { + "type": "text", + "bbox": [ + 0.187, + 0.217, + 0.784, + 0.235 + ], + "angle": 0, + "content": "\\(^{3}\\) School of Data Science, The Chinese University of Hong Kong, Shenzhen" + }, + { + "type": "text", + "bbox": [ + 0.246, + 0.238, + 0.721, + 0.252 + ], + "angle": 0, + "content": "eezhihaoliang@mail.scut.edu.cn, nwpuqzhang@gmail.com," + }, + { + "type": "text", + "bbox": [ + 0.204, + 0.256, + 0.76, + 0.269 + ], + "angle": 0, + "content": "vonyfeng@gmail.com, yingsshan@tencent.com, kuijia@cuhk.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.334, + 0.314, + 0.35 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.367, + 0.474, + 0.775 + ], + "angle": 0, + "content": "We propose GS-IR, a novel inverse rendering approach based on 3D Gaussian Splatting (3DGS) that leverages forward mapping volume rendering to achieve photorealistic novel view synthesis and relighting results. Unlike previous works that use implicit neural representations and volume rendering (e.g. NeRF), which suffer from low expressive power and high computational complexity, we extend 3DGS, a top-performance representation for novel view synthesis, to estimate scene geometry, surface material, and environment illumination from multi-view images captured under unknown lighting conditions. There are two main problems when introducing 3DGS to inverse rendering: 1) 3DGS does not support producing plausible normal natively; 2) forward mapping (e.g. rasterization and splatting) cannot trace the occlusion like backward mapping (e.g. ray tracing). To address these challenges, our GS-IR proposes an efficient optimization scheme incorporating a depth-derivation-based regularization for normal estimation and a baking-based occlusion to model indirect lighting. The flexible and expressive 3DGS representation allows us to achieve fast and compact geometry reconstruction, photorealistic novel view synthesis, and effective physically-based rendering. We demonstrate the superiority of our method over baseline methods through qualitative and quantitative evaluations of various challenging scenes. The source code is available at https://github.com/lzhnb/GS-IR." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.781, + 0.21, + 0.796 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.806, + 0.47, + 0.868 + ], + "angle": 0, + "content": "Inverse rendering is a long-standing task, seeking to answer the question: \"How can we deduce physical attributes (e.g. geometry, material, and lighting) of a 3D scene from multiview images?\" This problem is inherently challenging and" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.304, + 0.891, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.428, + 0.751, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.756, + 0.427, + 0.89, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.5, + 0.895, + 0.613 + ], + "angle": 0, + "content": "Figure 1. Given multi-view captured images of a complex scene, we propose GS-IR (3D Gaussian Splatting for Inverse Rendering), which utilizes 3D Gaussian and forward mapping splatting to recover high-quality physical properties (e.g., normal, material, illumination). This enables us to perform relighting and material editing, resulting in outstanding inverse rendering results. Better viewed on screen with zoom in, especially the remarkable material decomposition and normal reconstruction of bicycle axle." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.625, + 0.895, + 0.776 + ], + "angle": 0, + "content": "ill-posed, particularly when input images are captured in uncontrolled environments with unknown illumination. Recent research [9, 10, 35, 46] has sought to address this issue by employing implicit neural representations akin to NeRF [30] that utilizes multi-layer perceptrons (MLPs). However, current methods incorporating MLP face challenges in terms of their low expressive capacity and high computational demands, which significantly limits the effectiveness and efficiency of inverse rendering, especially when it cannot be rendered at interactive rates." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.903 + ], + "angle": 0, + "content": "3D Gaussian Splatting (3DGS) [24] has recently emerged as a promising technique to model 3D static scenes and significantly boost the rendering speed to a real-time level. It makes the scene representation more compact and achieves fast and top performance for novel view synthesis. Introducing it to the inverse rendering pipeline is natural and essential, including geometry reconstruction, materials decomposition, and illumination estimation. Unlike ray trac" + }, + { + "type": "page_footnote", + "bbox": [ + 0.1, + 0.875, + 0.264, + 0.888 + ], + "angle": 0, + "content": "* indicates equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.1, + 0.888, + 0.383, + 0.901 + ], + "angle": 0, + "content": "\\(^{\\dagger}\\) Correspondence to Kui Jia ." + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.875, + 0.383, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21644" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.471, + 0.182 + ], + "angle": 0, + "content": "ing in NeRF, 3DGS produces a set of 3D Gaussians around sparse points. During the 3DGS optimization, the adaptive control of the Gaussian density may lead to loose geometry, making it difficult to estimate accurate scene's normal. Consequently, it is necessary to introduce a well-designed strategy to regularize 3DGS's normal estimation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.192, + 0.471, + 0.403 + ], + "angle": 0, + "content": "Our goal is to use 3D Gaussians as the scene representation for inverse rendering from multi-view images captured under unknown lighting conditions. However, capturing observations under natural illumination often shows complex effects such as soft shadows and interreflections. TensoIR [22] leverages the ray tracing of NeRF to directly model occlusion and indirect illumination. In contrast, 3DGS replaces the ray tracing in the NeRF with differentiable forward mapping volume rendering, which directly projects 3D Gaussians onto the 2D plane. This strategy improves the rendering efficiency but makes it difficult to calculate occlusion. Inspired by the \"Indirect Lighting Cache\" used in real-time rendering [4], we attempt to bake the occlusion into volumes for caching." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.413, + 0.47, + 0.699 + ], + "angle": 0, + "content": "In this paper, we present a novel 3D Gaussian-based inverse rendering framework called GS-IR (3D Gaussian Splatting for Inverse Rendering) that leverages forward mapping splatting to deduce the physical attributes of a complex scene. To the best of our knowledge, our method is the first work to introduce the 3DGS technique for inverse rendering, which can simultaneously estimate scene geometry, materials, and illumination from multi-view images. Our GS-IR addresses two main issues when using 3DGS for inverse rendering. Firstly, we develop an intuitive and well-designed regularization to estimate the scene's normal. Secondly, we use a baking-based method embedded in GS-IR to cache occlusions, obtaining an efficient indirect illumination model. As shown in Fig. 1, our approach can reconstruct high-fidelity geometry and materials of a complex real scene under unknown natural illumination, enabling state-of-the-art rendering of novel view synthesis and additional applications like relighting. Our technical contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.71, + 0.468, + 0.754 + ], + "angle": 0, + "content": "- We present GS-IR that models a scene as a set of 3D Gaussians to achieve physically-based rendering and state-of-the-art decomposition results for both objects and scenes;" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.756, + 0.468, + 0.799 + ], + "angle": 0, + "content": "- We propose an efficient optimization scheme with regularization to concentrate depth gradient around 3DGS and produce reliable normals for GS-IR;" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.8, + 0.468, + 0.83 + ], + "angle": 0, + "content": "- We develop a baking-based method embedded in GS-IR to handle the occlusion in modeling indirect lighting;" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.71, + 0.468, + 0.83 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.468, + 0.901 + ], + "angle": 0, + "content": "We demonstrate the superiority of our method to baseline methods qualitatively and quantitatively on various challenging scenes, including the TensoIR synthesis dataset [22] and Mip-NeRF 360 real dataset [5]." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.09, + 0.65, + 0.105 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.892, + 0.462 + ], + "angle": 0, + "content": "Neural Representation Recently, neural rendering techniques, exemplified by Neural Radiance Field (NeRF) [30], have achieved impressive success in addressing visual computing problems, giving rise to numerous neural representations [14, 21, 31, 36, 39, 40] tailored for different tasks [11-13, 18, 27, 28, 34, 43]. The vanilla NeRF models a continuous radiance field implicitly in MLPs, which requires massive repeated queries during training and inference. To address the computational inefficiencies, many neural scene representations are proposed with more discretized geometry proxies such as voxel grids [17, 20, 36], hash grids [31], tri-planes [14] or points [39]. Neural features are stored in a structured manner, allowing for efficient storage and retrieval. The computational cost can thus be significantly reduced by introducing interpolation techniques, however, with an inevitable loss in image quality. 3D Gaussians are introduced as an unstructured scene representation to strike a balance between efficiency and quality [24]. With the specially designed tile-based rasterizer for Gaussian splats, this method achieves real-time rendering with high quality for novel-view synthesis. In this work, 3D Gaussian representation is combined with the physical-based rendering (PBR) model for inverse rendering." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.464, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Inverse Rendering Inverse rendering aims to decompose the image's appearance into the geometry, material, and lighting conditions. Considering the inherent ambiguity between observed images and underlying properties, many methods are proposed with different constrained settings, such as capturing images with fixed lighting and rotating object [16, 38], capturing with moving camera and colocated lighting [7, 8, 29, 33]. Combined with neural representations, inverse rendering models the scene simulating how the light interacts with the neural volume with various material properties, and estimates the lighting and material parameters during optimization [6, 9, 10, 19, 22, 35, 42, 44, 46, 47]. Neural Reflectance Fields [6] assumes a known point light source and represents the scene as a field of volume density, surface normals, and bi-directional reflectance distribution functions(BRDFs) with one bounce direct illumination. NeRV [35] and InvRender [47] extend to arbitrary known lighting conditions and train an additional MLP to model the light visibility. PhySG [42] assumes full light source visibility without shadow simulation, and represents the lighting and scene BRDFs with spherical Gaussians for acceleration. TensoIR [22] adopts the efficient TensoRF [14] representation which enables the computation of visibility and indirect lighting by raytracing, while limited to object-level. For modeling surface geometry using point clouds, Fuzzy Metaballs (FMs) [25, 26] offers a great way to render depth from 3D Gaussian using Order Independent Transparency (OIT) and approximate intersection. However, it requires silhouettes as input and struggles to handle" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21645" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "intricate geometry (e.g. Lego and Ficus) let alone complex scenes. In this work, we propose a 3DGS-based pipeline to recover the geometry, material, and lighting that is available for both objects and unbounded scenes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.163, + 0.204, + 0.181 + ], + "angle": 0, + "content": "3. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.188, + 0.469, + 0.233 + ], + "angle": 0, + "content": "In this section, we give the technical backgrounds and math symbols that are necessary for the presentation of our proposed method in subsequent sections." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.234, + 0.469, + 0.294 + ], + "angle": 0, + "content": "3D Gaussian Splatting (3DGS) [24] is an explicit 3D scene representation in the form of point clouds. Each point is represented as a Gaussian function \\( g \\) that approximates the shape of a bell curve, which is defined as," + }, + { + "type": "equation", + "bbox": [ + 0.161, + 0.298, + 0.469, + 0.318 + ], + "angle": 0, + "content": "\\[\ng (\\boldsymbol {x} | \\boldsymbol {\\mu}, \\boldsymbol {\\Sigma}) = e ^ {- \\frac {1}{2} (\\boldsymbol {x} - \\boldsymbol {\\mu}) ^ {T} \\boldsymbol {\\Sigma} ^ {- 1} (\\boldsymbol {x} - \\boldsymbol {\\mu})}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.318, + 0.469, + 0.47 + ], + "angle": 0, + "content": "where \\(\\pmb{\\mu} \\in \\mathbb{R}^3\\) is its mean vector, and \\(\\pmb{\\Sigma} \\in \\mathbb{R}^{3 \\times 3}\\) is an anisotropic covariance matrix. The mean vector \\(\\pmb{\\mu}\\) of a 3D Gaussian is parameterized as \\(\\pmb{\\mu} = (\\mu_x, \\mu_y, \\mu_z)\\), and the covariance matrix \\(\\pmb{\\Sigma}\\) is factorized into a scaling matrix \\(\\pmb{S}\\) and a rotation matrix \\(\\pmb{R}\\) as \\(\\pmb{\\Sigma} = R\\pmb{S}\\pmb{S}^\\top\\pmb{R}^\\top\\). \\(\\pmb{S}\\) and \\(\\pmb{R}\\) refer to a diagonal matrix \\(\\mathrm{diag}(s_x, s_y, s_z)\\) and a rotation matrix constructed from a unit quaternion \\(\\pmb{q}\\). Given a viewing transformation with extrinsic matrix \\(\\pmb{T}\\) and intrinsic matrix \\(\\pmb{K}\\), the mean vector \\(\\pmb{\\mu}\\) and covariance matrix \\(\\pmb{\\Sigma}'\\) from the 3D point \\(\\pmb{x}\\) to 2D pixel \\(\\pmb{u}\\) is defined as," + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.473, + 0.469, + 0.49 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\mu} ^ {\\prime} = \\boldsymbol {K} \\boldsymbol {T} [ \\boldsymbol {\\mu}, 1 ] ^ {\\top}, \\boldsymbol {\\Sigma} ^ {\\prime} = \\boldsymbol {J} \\boldsymbol {T} \\boldsymbol {\\Sigma} \\boldsymbol {T} ^ {\\top} \\boldsymbol {J} ^ {\\top}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.492, + 0.469, + 0.583 + ], + "angle": 0, + "content": "where \\(\\pmb{J}\\) is the Jacobian matrix of the affine approximation of the perspective projection. Besides, each Gaussian represents the view-dependent color \\(c_{i}\\) via a set of coefficients of spherical harmonics (SH), which is then multiplied by opacity \\(\\alpha\\) for volume rendering. We finally obtain the color \\(\\hat{\\pmb{C}}\\) at pixel \\(\\pmb{u}\\) based on Eq. (1) and Eq. (2)," + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.588, + 0.469, + 0.636 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {C}} = \\sum_ {i \\in N} T _ {i} g _ {i} \\left(\\boldsymbol {u} \\mid \\boldsymbol {\\mu} ^ {\\prime}, \\boldsymbol {\\Sigma} ^ {\\prime}\\right) \\alpha_ {i} \\boldsymbol {c} _ {i}, \\quad T _ {i} = \\prod_ {j = 1} ^ {i - 1} \\left(1 - g _ {j} \\left(\\boldsymbol {u} \\mid \\boldsymbol {\\mu} ^ {\\prime}, \\boldsymbol {\\Sigma} ^ {\\prime}\\right) \\alpha_ {j}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.635, + 0.468, + 0.664 + ], + "angle": 0, + "content": "where accumulated transmittance \\( T_{i} \\) quantifies the probability density of \\( i \\)-th Gaussian at pixel \\( \\mathbf{u} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.664, + 0.469, + 0.709 + ], + "angle": 0, + "content": "The Rendering Equation In GS-IR, we leverage the classic rendering equation to formulate the outgoing radiance of a surface point \\( \\mathbf{x} \\) with normal \\( \\mathbf{n} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.147, + 0.714, + 0.469, + 0.743 + ], + "angle": 0, + "content": "\\[\nL _ {o} (\\boldsymbol {x}, \\boldsymbol {v}) = \\int_ {\\Omega} L _ {i} (\\boldsymbol {x}, \\boldsymbol {l}) f _ {r} (\\boldsymbol {l}, \\boldsymbol {v}) (\\boldsymbol {l} \\cdot \\boldsymbol {n}) d \\boldsymbol {l}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.746, + 0.469, + 0.852 + ], + "angle": 0, + "content": "\\(\\Omega\\) denotes the upper hemisphere centered at \\(\\pmb{x}\\), \\(\\pmb{l}\\) and \\(\\pmb{v}\\) denote incident and view directions respectively. \\(L_{i}(\\pmb{x},\\pmb{l})\\) denotes the radiance received at \\(\\pmb{x}\\) from \\(\\pmb{l}\\). Notably, we follow Cook-Torrance microfacet model [15, 37] and formulate the bidirectional reflectance distribution function (BRDF) \\(f_{r}\\) as a function of albedo \\(\\pmb{a} \\in [0,1]^3\\), metallic \\(m \\in [0,1]\\), and roughness \\(\\rho \\in [0,1]\\):" + }, + { + "type": "equation", + "bbox": [ + 0.147, + 0.856, + 0.469, + 0.904 + ], + "angle": 0, + "content": "\\[\nf _ {r} (\\boldsymbol {l}, \\boldsymbol {v}) = \\underbrace {(1 - m) \\frac {\\boldsymbol {a}}{\\pi}} _ {\\text {d i f f u s e}} + \\underbrace {\\frac {D F G}{4 (\\boldsymbol {n} \\cdot \\boldsymbol {l}) (\\boldsymbol {n} \\cdot \\boldsymbol {v})}} _ {\\text {s p e c u l a r}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "where microfacet distribution function \\( D \\), Fresnel reflection \\( F \\), and geometric shadowing factor \\( G \\) are related to the surface roughness \\( \\rho \\). We use 3D Gaussians to store these material properties in GS-IR." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.165, + 0.591, + 0.181 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.188, + 0.892, + 0.493 + ], + "angle": 0, + "content": "Given a set of calibrated RGB images \\(\\{I_m\\}_{m=1}^M\\) of a target scene captured from multiple views under static, yet unknown illumination, inverse rendering aims to decompose the scene's intrinsic properties, including normal, materials, and illumination. This decomposition facilitates the recovery and subsequent edition of the target scene. Motivated by the remarkable performance in quality and speed of 3DGS [24], we present a novel framework GS-IR consisting of three well-designed stage strategies, as shown in Fig. 2. In the initial stage, we leverage differentiable splating to optimize 3D Gaussians. Concurrently, we utilize the gradient derived from the rendered depth map to supervise the normal stored in 3D Gaussians (cf. Sec. 4.1). In the second stage, we precompute the occlusion based on the learned geometric information (i.e. depth and normal) and store it in an efficient spherical harmonics-based architecture to model indirect illumination (cf. Sec. 4.2). In the final stage, we combine a differentiable splating with the physical-based rendering (PBR) pipeline to optimize illumination and material-aware 3D Gaussians (cf. Sec. 4.3)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.5, + 0.716, + 0.515 + ], + "angle": 0, + "content": "4.1. Normal Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.892, + 0.689 + ], + "angle": 0, + "content": "During the initial stage, we optimize 3D Gaussians for geometry reconstruction from observed images, denoted as \\(\\mathcal{G}\\). The optimized \\(\\mathcal{G}\\) functions as a geometric proxy for surface points and their corresponding normals \\(\\pmb{n}\\), which are crucial for successful inverse rendering. As highlighted in Sec. 1, generating reasonable normals within the 3DGS-based framework poses a significant challenge. To address this obstacle, we introduce an intuitive strategy that improves depth \\(\\hat{D}\\) and leverages the depth gradient to derive pseudo normals \\(\\hat{\\pmb{n}}_{\\hat{D}} = \\nabla_{\\mathrm{uv}}\\hat{D}\\). These pseudo normals then guide the optimization of normals within the 3D Gaussians." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Depth Generation Given a pretrained 3D Gaussians \\(\\mathcal{G}\\) and a view designated for rendering, the pixel's shading results in that view can be obtained by Eq. (3). Consequently, it is reasonable to utilize the same volumetric accumulation to compute the depth \\(\\hat{D} = \\sum_{i=1}^{N} T_i \\alpha_i d_i\\), where \\(d_i\\) denotes the distance from the corresponding 3D Gaussian to the image plane. However, we observed the floating problem during volumetric accumulation, unlike the backward mapping volume rendering used in NeRF. During the 3DGS optimization, the adaptive control of the Gaussian density may result in the depth falling in front of the 3D Gaussians, thereby posing challenges in accurately predicting the depth. Specifically, the backward mapping methods can obtain an accurate depth by considering only peak samples," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21646" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.089, + 0.87, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.429, + 0.896, + 0.488 + ], + "angle": 0, + "content": "Figure 2. GS-IR Pipeline. We propose a novel GS-based inverse rendering framework, called GS-IR, to reconstruct scene geometry, materials, and unknown natural illumination from multi-view captured images. Our GS-IR consists of three well-designed stage strategies using 3D Gaussian and differentiable forward mapping splatting to achieve physical-based rendering. In our approach, the Gaussian stores not only the basic 3DGS information but also the normal and material properties, enhancing its capabilities for inverse rendering tasks." + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.494, + 0.449, + 0.601 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.61, + 0.47, + 0.667 + ], + "angle": 0, + "content": "Figure 3. Depth Illustration. By considering the depth as a linear interpolation of the distances from 3D Gaussians to the image plane, and ensuring it lies between the minimum and maximum distance, our method could produce accurate depth." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.67, + 0.471, + 0.791 + ], + "angle": 0, + "content": "that is \\(\\hat{D} = d_{i^*}\\), where \\(i^* = \\arg \\max_i T_i \\alpha_i\\). However, for 3DGS, a typical forward mapping method, the peak selection results in disc aliasing within 3DGS. To overcome this limitation, we consider that depth \\(\\hat{D}\\) must be between the minimum and maximum distance of 3D Gaussians to the image plane, as illustrated in Fig. 3. We then treat the depth as a linear interpolation of the distances from 3D Gaussians to the image plane:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.8, + 0.47, + 0.838 + ], + "angle": 0, + "content": "\\[\n\\hat {D} = \\sum_ {i = 1} ^ {N} \\hat {w} _ {i} d _ {i}, \\quad \\hat {w} _ {i} = \\frac {T _ {i} \\alpha_ {i}}{\\sum_ {i = 1} ^ {N} T _ {i} \\alpha_ {i}}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Normal Derivation While the accurate prediction of depth within Gaussians provides better guidance for the normal reconstruction, directly using depth gradient to produce normals has two limitations that still cannot meet the require" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.497, + 0.895, + 0.666 + ], + "angle": 0, + "content": "ment for effective inverse rendering. First, the depth gradient estimation is highly sensitive to noise, making the predicted normal often extremely noisy; Second, the normals derived individually from each view's depth map do not satisfy multi-view consistency. To address these issues, we use Gaussian \\(\\mathcal{G}\\) as a proxy for normal estimation instead of directly from the depth gradient. Benefiting from the efficiency of 3DGS, we obtain the depth \\(\\hat{D}\\) and normals \\(\\hat{\\pmb{n}}\\) of the observed view after performing a single rendering pass. We then tie these predicted pseudo normal to the underlying depth gradient normal \\(\\hat{\\pmb{n}}_{\\hat{D}}\\) using a simple penalty:" + }, + { + "type": "equation", + "bbox": [ + 0.633, + 0.671, + 0.891, + 0.687 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {n - p} = \\left\\| \\hat {\\boldsymbol {n}} - \\hat {\\boldsymbol {n}} _ {\\hat {D}} \\right\\|, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.892, + 0.797 + ], + "angle": 0, + "content": "where \\(\\hat{\\pmb{n}} = \\sum_{i=1}^{N} T_i \\alpha_i \\pmb{n}_i\\), and \\(\\pmb{n}_i\\) is the normal stored in the 3D Gaussian. Secondly, unlike the MLP-based normal estimation [22] acting MLP as a low-pass filter, the predicted normal of Gaussian \\(\\mathcal{G}\\) is rough, so smoothness regularization should be included. We introduce the TV term \\(TV_{\\text{normal}}\\) to smooth the predicted normal \\(\\hat{\\pmb{n}}\\). For more details, please refer to the supplement." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.798, + 0.892, + 0.872 + ], + "angle": 0, + "content": "In optimization of the first stage, we optimize 3D Gaussians \\(\\mathcal{G}\\) (storing SH coefficients for view-dependent color \\(c\\), opacity \\(\\alpha\\), and normal \\(\\hat{n}\\)) by using the color reconstruction loss \\(\\mathcal{L}_c\\), which is the same as 3DGS [24], and the proposed normal loss \\(\\mathcal{L}_n\\)," + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.88, + 0.891, + 0.896 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {n} = \\mathcal {L} _ {n - p} + \\lambda_ {n - T V} T V _ {\\text {n o r m a l}}. \\tag {8}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21647" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.089, + 0.452, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.28, + 0.473, + 0.406 + ], + "angle": 0, + "content": "Figure 4. Baking. We employ the spherical harmonics (SH) architecture to bake occlusion volumes for modeling indirect illumination. For each grid of occlusion volumes, we initially use 3D Gaussians to compute the depth cubemap by performing six forward mapping splatting passes. Next, we convert the depth cubemap into a binary occlusion cubemap based on a distance threshold. Finally, the occlusion cubemap is baked as SH coefficients, enabling efficient interpolation of the occlusion cubemap at any point within the scene." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.416, + 0.355, + 0.433 + ], + "angle": 0, + "content": "4.2. Indirect Illumination Modeling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.442, + 0.47, + 0.532 + ], + "angle": 0, + "content": "Drawing inspiration from the successful implementation of precomputation techniques in the video game industry (e.g. Irradiance Volume in Blender [3], Lightmass Volume in Unreal [1], and Light Probes in Unity [2]), we introduce spherical harmonics (SH) architectures to store occlusion information and model indirect illumination." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.535, + 0.47, + 0.627 + ], + "angle": 0, + "content": "Given the optimized 3D Gaussians \\(\\mathcal{G}\\) from one stage (cf. Sec. 4.1), we freeze \\(\\mathcal{G}\\) and regularly place occlusion volumes \\(\\nu^{\\mathrm{occl}}\\) in the 3D space. For each volume \\(\\pmb{v}_i^{\\mathrm{occl}}\\subset \\nu^{\\mathrm{occl}}\\) we then cache the occlusion in the form of SH coefficients \\(\\pmb{f}_i^{\\mathrm{o}}\\). Consequently, the formula of the occlusion \\(O(\\cdot)\\) of \\(\\pmb{v}_i^{\\mathrm{occl}}\\) with respect to the direction \\((\\theta ,\\phi)\\) is expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.645, + 0.47, + 0.682 + ], + "angle": 0, + "content": "\\[\nO (\\theta , \\phi) = \\sum_ {l = 0} ^ {\\text {d e g}} \\sum_ {m = - l} ^ {l} \\mathbf {f} _ {i (l m)} ^ {0} Y _ {l m} (\\theta , \\phi), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.697, + 0.47, + 0.728 + ], + "angle": 0, + "content": "where \\(deg\\) denotes the degree of SH, and \\(\\{Y_{lm}(\\cdot)\\}\\) is a set of real basis of SH." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.73, + 0.471, + 0.883 + ], + "angle": 0, + "content": "As discussed in Sec. 4.1, the 3DGS technique employs a forward mapping approach that projects 3D points to the 2D plane, in contrast to the backward mapping volume rendering utilized in NeRF, which means it cannot use ray marching to calculate occlusions. To precompute the SH coefficients \\( \\pmb{f}_i^{\\mathrm{o}} \\) of occlusion volume \\( \\pmb{v}_i^{\\mathrm{occl}} \\), we obtain the depth cubemap \\( \\{\\hat{D}_p^i\\}_{p=1}^6 \\) by performing six times rendering passes, once for each face of the cubemap. We then convert it into a binary occlusion cubemap \\( \\{\\hat{O}_p^i\\}_{p=1}^6 \\) based on a manually set distance threshold." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Finally, we convolve the the occlusion cubemap" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.09, + 0.877, + 0.11 + ], + "angle": 0, + "content": "\\(\\{\\hat{O}_p^i\\}_{p = 1}^6\\) using SH bases and get the SH coefficients \\(\\pmb{f}_i^0\\) .." + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.113, + 0.891, + 0.169 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\boldsymbol {f} _ {i (l m)} ^ {0} = \\int_ {S ^ {2}} \\hat {\\boldsymbol {O}} (\\boldsymbol {\\omega}) Y _ {l m} (\\boldsymbol {\\omega}) d \\boldsymbol {\\omega} \\tag {10} \\\\ = \\int_ {0} ^ {2 \\pi} \\int_ {0} ^ {\\pi} \\sin \\theta \\hat {\\boldsymbol {O}} (\\theta , \\phi) Y _ {l m} (\\theta , \\phi) d \\theta d \\phi , \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.171, + 0.892, + 0.233 + ], + "angle": 0, + "content": "where \\(S^2\\) denotes the unit sphere, and \\(\\hat{O} (\\theta ,\\phi)\\) denotes the occlusion query from the occlusion cubemap \\(\\{\\hat{O}_p^i\\}_{p = 1}^6\\) Note that we numerically calculate the convolution in Eq. (10) in parallel. The caching process is shown in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.234, + 0.893, + 0.357 + ], + "angle": 0, + "content": "To handle the indirect illumination in occlusion regions, we also maintain illumination volumes \\(\\mathcal{V}^{\\mathrm{illu}}\\) to cache the indirect illumination. Similar to the caching process of occlusion volumes, the caching target of illumination volumes changes from the occlusion cubemap \\(\\{\\hat{O}_p^i\\}_{p=1}^6\\) to the captured environment cubemap \\(\\{\\hat{I}_p^i\\}_{p=1}^6\\). These cubemaps can be obtained simultaneously by conducting six rendering passes. For more details, please refer to the supplement." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.366, + 0.722, + 0.383 + ], + "angle": 0, + "content": "4.3. Intrinsic Decomposition" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.389, + 0.892, + 0.463 + ], + "angle": 0, + "content": "In the final stage, we employ differentiable splatting in conjunction with a PBR pipeline to accomplish the intrinsic decomposition. According to Eq. (5), the rendering equation Eq. (4) is rewritten as diffuse \\( L_{\\mathrm{d}} \\) and specular \\( L_{\\mathrm{s}} \\) components:" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.469, + 0.892, + 0.547 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {o} (\\boldsymbol {x}, \\boldsymbol {v}) = \\int_ {\\Omega} \\left[ (1 - m) \\frac {\\boldsymbol {a}}{\\pi} + \\frac {D F G}{4 (\\boldsymbol {n} \\cdot \\boldsymbol {l}) (\\boldsymbol {n} \\cdot \\boldsymbol {v})} \\right] L _ {i} (\\boldsymbol {x}, \\boldsymbol {l}) (\\boldsymbol {l} \\cdot \\boldsymbol {n}) d \\boldsymbol {l} \\\\ L _ {\\mathrm {d}} = (1 - m) \\frac {\\boldsymbol {a}}{\\pi} \\int_ {\\Omega} L _ {i} (\\boldsymbol {x}, \\boldsymbol {l}) (\\boldsymbol {l} \\cdot \\boldsymbol {n}) d \\boldsymbol {l} \\tag {11} \\\\ L _ {\\mathrm {s}} = \\int_ {\\Omega} \\frac {D F G}{4 (\\boldsymbol {n} \\cdot \\boldsymbol {l}) (\\boldsymbol {n} \\cdot \\boldsymbol {v})} L _ {i} (\\boldsymbol {x}, \\boldsymbol {l}) (\\boldsymbol {l} \\cdot \\boldsymbol {n}) d \\boldsymbol {l}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.892, + 0.614 + ], + "angle": 0, + "content": "In GS-IR, we adopt an image-based lighting (IBL) model and split-sum approximation [23] to tackle the intractable integral. To calculate the diffuse component \\( L_{\\mathrm{d}} \\), the illumination \\( I_{\\mathrm{d}} \\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.618, + 0.891, + 0.691 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} I _ {\\mathrm {d}} (\\boldsymbol {x}) = \\int_ {\\Omega} L _ {i} (\\boldsymbol {x}, \\boldsymbol {l}) (\\boldsymbol {l} \\cdot \\boldsymbol {n}) d \\boldsymbol {l} \\\\ = \\int_ {\\Omega_ {\\mathrm {V i s}}} L _ {i} ^ {\\operatorname {d i r}} (\\boldsymbol {x}, \\boldsymbol {l}) (\\boldsymbol {l} \\cdot \\boldsymbol {n}) d \\boldsymbol {l} + \\int_ {\\Omega_ {\\mathrm {O c c l}}} L _ {i} ^ {\\operatorname {i n d i r}} (\\boldsymbol {x}, \\boldsymbol {l}) (\\boldsymbol {l} \\cdot \\boldsymbol {n}) d \\boldsymbol {l} \\tag {12} \\\\ \\approx \\left(1 - \\mathrm {O} (\\boldsymbol {x})\\right) I _ {\\mathrm {d}} ^ {\\operatorname {d i r}} (\\boldsymbol {x}) + \\mathrm {O} (\\boldsymbol {x}) I _ {\\mathrm {d}} ^ {\\operatorname {i n d i r}} (\\boldsymbol {x}), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.694, + 0.892, + 0.815 + ], + "angle": 0, + "content": "where the first component indicates direct illumination and the second is indirect illumination. Notably, our baking-based indirect illumination model enables us to calculate the occlusion and illumination online. This means that our GS-IR achieves intrinsic decomposition while maintaining real-time rendering performance. For the specular component \\( L_{\\mathrm{s}} \\), we follow split-sum approximation and treat the integral as two separate integrals:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.83, + 0.892, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {\\mathrm {s}} = \\int_ {\\Omega} \\frac {D F G}{4 \\langle \\boldsymbol {n} \\cdot \\boldsymbol {l} \\rangle \\langle \\boldsymbol {n} \\cdot \\boldsymbol {v} \\rangle} L _ {i} (\\boldsymbol {l}) \\langle \\boldsymbol {l} \\cdot \\boldsymbol {n} \\rangle d \\boldsymbol {l} \\\\ \\approx \\underbrace {\\int_ {\\Omega} \\frac {D F G}{4 \\left\\langle \\boldsymbol {n} \\cdot \\boldsymbol {l} \\right\\rangle \\left\\langle \\boldsymbol {n} \\cdot \\boldsymbol {v} \\right\\rangle}} _ {\\text {E n v i r o n m e n t B R D F -} R} \\left\\langle \\boldsymbol {l} \\cdot \\boldsymbol {n} \\right\\rangle d \\boldsymbol {l} \\underbrace {\\int_ {\\Omega} D L _ {i} (\\boldsymbol {l}) \\left\\langle \\boldsymbol {l} \\cdot \\boldsymbol {n} \\right\\rangle d \\boldsymbol {l}} _ {\\text {P r e - F i l t e r e d E n v i r o n m e n t M a p -} I _ {s}}, \\tag {13} \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21648" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.086, + 0.895, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.343, + 0.895, + 0.415 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison on TensoIR Synthetic. We visualize the estimated normal, albedo, and rendering results of our GS-IR and baseline methods on two scenes. By utilizing the efficient 3D Gaussian representation and a robust tile-based rasterizer, GS-IR achieves rapid convergence and supports real-time rendering. This performance advantage underscores the effectiveness of our method in addressing complex inverse rendering tasks, thereby surpassing existing state-of-the-art approaches. (For albedo reconstruction results, we follow NeRFactor [46] and scale each RGB channel by a global scalar.)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.44, + 0.471, + 0.486 + ], + "angle": 0, + "content": "where both \\( R \\) and \\( I_{\\mathrm{s}} \\) can be precomputed in advance and stored in look-up tables. With Eq. (12) and Eq. (13), the rendering results of Eq. (11) can be represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.489, + 0.469, + 0.537 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {o} (\\boldsymbol {x}, \\boldsymbol {v}) = L _ {\\mathrm {d}} + L _ {\\mathrm {s}} \\\\ \\approx (1 - m) \\frac {\\boldsymbol {a}}{\\pi} \\left[ (1 - \\mathrm {O} (\\boldsymbol {x})) I _ {\\mathrm {d}} ^ {\\operatorname {d i r}} (\\boldsymbol {x}) + \\mathrm {O} (\\boldsymbol {x}) I _ {\\mathrm {d}} ^ {\\operatorname {i n d i r}} (\\boldsymbol {x}) \\right] + R I _ {\\mathrm {s}}. \\tag {14} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.539, + 0.47, + 0.6 + ], + "angle": 0, + "content": "For intrinsic decomposition, we optimize the material \\(\\hat{M}\\) (i.e. albedo \\(a\\), metallic value \\(m\\), and roughness \\(\\rho\\)) stored in 3D Gaussians \\(\\mathcal{G}\\), environment map \\(\\hat{\\pmb{E}}\\), and illumination volumes \\(\\mathcal{V}^{\\mathrm{illu}}\\) by minimizing the decomposition loss \\(\\mathcal{L}_d\\):" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.601, + 0.47, + 0.652 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d} = \\underbrace {\\left\\| \\boldsymbol {I} - \\hat {\\boldsymbol {I}} ^ {\\text {s h a d e}} \\left(\\hat {\\boldsymbol {M}} , \\hat {\\boldsymbol {E}} , \\mathcal {V} ^ {\\text {i l l u}}\\right) \\right\\|} _ {\\mathcal {L} _ {\\text {s h a d e}}} + \\underbrace {\\lambda_ {M} T V _ {\\mathrm {m a t}}} _ {\\mathcal {L} _ {\\text {m a t e r i a l}}} + \\underbrace {\\lambda_ {\\boldsymbol {E}} T V _ {\\text {l i g h t}}} _ {\\mathcal {L} _ {\\text {l i g h t}}}, \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.652, + 0.47, + 0.714 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{shade}}\\) indicate the shade loss. \\(\\hat{I}^{\\mathrm{shade}}(\\hat{M},\\hat{E},\\hat{\\mathcal{V}}^{\\mathrm{illu}})\\) is the recovered image that uses the PBR pipeline defined by Eq. (14). Please refer to the supplement for more details about the material TV loss \\(\\mathcal{L}_{\\mathrm{material}}\\) and lighting TV loss \\(\\mathcal{L}_{\\mathrm{light}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.725, + 0.21, + 0.743 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Dataset & Metrics We conduct experiments using benchmark datasets of TensoIR Synthetic [22] and Mip-NeRF 360 [5] for decomposing both objects and scenes. They contain 4 objects with reference materials and 7 publicly available scenes, respectively. To verify the efficacy of our normal reconstruction, we evaluate the normal quality on the TensoIR Synthetic [22] dataset using mean angular error (MAE). We further assess our reconstructed albedo quality on this synthetic dataset. More generally, we evaluate the synthesized novel view on both datasets in terms of Peak" + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.438, + 0.69, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.438, + 0.857, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.522, + 0.696, + 0.604 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.522, + 0.871, + 0.605 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.617, + 0.894, + 0.687 + ], + "angle": 0, + "content": "Figure 6. Novel view synthesis results on Mip-NeRF 360. GS-IR can reconstruct scene details including geometric normals and high-frequency appearance, rendering high-fidelity appearance and recovering fine geometric details such as those on leaves and bicycle axles. Better viewed on screen with zoom in." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.71, + 0.894, + 0.772 + ], + "angle": 0, + "content": "Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS) [45]. Note that albedo quality assessment uses the same metrics as novel view synthesis." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.785, + 0.64, + 0.802 + ], + "angle": 0, + "content": "5.1. Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.894, + 0.903 + ], + "angle": 0, + "content": "We conduct a comprehensive comparison against state-of-the-art neural field-based inverse rendering methods on the public TensoIR Synthetic dataset [22]. All the methods utilize multi-view images captured under unknown lighting conditions. Our evaluation encompasses normal quality (measured by MAE), novel view synthesis, albedo fidelity," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21649" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.084, + 0.089, + 0.89, + 0.186 + ], + "angle": 0, + "content": "
MethodNormal ↓ MAE ↓Novel View SynthesisAlbedoRelightRuntime
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
NeRFactor [46]6.31424.6790.9220.12025.1250.9400.10923.3830.9080.131>100 hrs
InvRender [47]5.07427.3670.9340.08927.3410.9330.10023.9730.9010.10115 hrs
NVDiffrec [32]6.07830.6960.9620.05229.1740.9080.11519.8800.8790.104<1 hr
TensoIR [22]4.10035.0880.9760.04029.2750.9500.08528.5800.9440.0815 hrs
Ours4.94835.3330.9740.03930.2860.9410.08424.3740.8850.096<1 hr
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.195, + 0.893, + 0.268 + ], + "angle": 0, + "content": "Table 1. Quantitative Comparison on TensoIR Synthetic dataset. Our method outperforms baseline methods in terms of novel view synthesis and albedo quality, showcasing the effectiveness of material decomposition and PBR rendering. This is particularly noteworthy considering that our normal reconstruction is slightly inferior to TensoIR. In terms of relighting performance, we rank second, trailing only behind TensoIR. Importantly, the average training time of our GS-IR is accelerated by a factor of \\(5\\mathrm{x}\\), making its performance acceptable and further demonstrating the effectiveness of our approach in handling complex inverse rendering tasks." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.281, + 0.462, + 0.382 + ], + "angle": 0, + "content": "
MethodPSNR ↑SSIM ↑LPIPS ↓Runtime ↓
NeRF++ [41]25.1120.6960.375≈ 20h
Plenoxels [17]23.0790.6250.462≈ 30m
INGP-Base [31]25.3030.6710.371≈ 5m
INGP-Big [31]25.5870.6990.331≈ 8m
Mip-NeRF 360 [31]27.5690.7930.234≈ 48h
3DGS [24]27.210.8150.214≈ 35m
Ours25.3810.7570.267≈ 45m
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.393, + 0.47, + 0.435 + ], + "angle": 0, + "content": "Table 2. Quantitative Comparison on Mip-NeRF 360. The results show that our inverse rendering approach even surpasses some NeRF variants dedicated to novel view synthesis." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.46, + 0.47, + 0.701 + ], + "angle": 0, + "content": "relighting effects (measured by PSNR, SSIM, and LPIPS), and efficiency. Tab. 1 summarizes the quantitative comparisons on the synthesis dataset. Our method achieves superior performance in novel view synthesis and albedo quality compared to the baseline methods, demonstrating the effectiveness of material decomposition and PBR rendering, particularly given that our normal reconstruction is slightly inferior to TensoIR. Our relighting performance ranks second, only behind TensoIR. Notably, the average training time of our GS-IR is accelerated by a factor of \\(5\\mathrm{x}\\), making its performance acceptable and demonstrating the effectiveness of our approach in handling complex inverse rendering tasks. We also include qualitative comparisons in Fig. 5, which show that our GS-IR produces reasonable albedo and photorealistic renderings that are closer to the ground truth than most methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Meanwhile, owing to our more efficient and compact representation with powerful expressiveness, our method showcases remarkable performance on complex real unbounded scenes [5]. Tab. 2 presents the quantitative comparisons on the real dataset. Fig. 6 demonstrates the normal reconstruction and novel view synthesis on the real dataset. Our method renders a high-fidelity appearance and recovers fine geometric details, such as those on the leaves and bicycle axil. In summary, by leveraging the efficient 3D Gaussian representation and a powerful tile-based rasterizer, GS-IR achieves fast convergence and supports real-time rendering. This performance advantage highlights the effectiveness of our method in handling complex inverse rendering" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.282, + 0.889, + 0.369 + ], + "angle": 0, + "content": "
MethodTensoIR Synthetic [22]Mip-NeRF 360 [5]
Normal MAE↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Vol. Accum.16.34725.7560.8550.13122.0520.6100.394
Peak Selec.9.46628.7500.9270.08423.0930.7190.317
Linear Interp.6.21828.9830.9390.06623.1190.7070.319
Vol. Accum.†9.31530.0910.9400.07124.6640.7470.277
Peak Selec.†7.98631.0640.9500.06025.1430.7530.281
Linear Interp.†4.94835.3330.9740.03925.3810.7570.267
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.373, + 0.893, + 0.43 + ], + "angle": 0, + "content": "Table 3. Analyses on the impact of different depth generation strategies on normals. Methods without \\(\\dagger\\) marks directly use the normals derived from the depth map; Methods marked with \\(\\dagger\\) use depth derivation to optimize the normals stored in 3D Gaussians." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.436, + 0.87, + 0.539 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.541, + 0.892, + 0.582 + ], + "angle": 0, + "content": "Figure 7. Visual comparison of depth produced by different strategies. The linear interpolation adopted in GS-IR overcomes the floating problem and disc aliasing." + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.583, + 0.885, + 0.637 + ], + "angle": 0, + "content": "
MethodTensoIR Synthetic [22]Mip-NeRF 360 [5]
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
w/o occlusion34.9970.9620.04125.0600.7530.270
w/o indirect illum.35.1860.9650.04424.8980.7490.272
Ours35.3330.9740.03925.3810.7570.267
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.64, + 0.892, + 0.682 + ], + "angle": 0, + "content": "Table 4. Analyses on the occlusion and indirect illumination. Physically modeling indirect illumination improves the inverse rendering of objects and scenes." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.702, + 0.878, + 0.717 + ], + "angle": 0, + "content": "tasks, outperforming existing state-of-the-art approaches." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.727, + 0.665, + 0.741 + ], + "angle": 0, + "content": "5.2. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We initially introduce the 3DGS technique for inverse rendering in GS-IR and propose depth-derivation-based normal regularization and a baking-based method to address the challenges encountered during the process. To evaluate the efficacy of our proposed schemes, we design elaborate experiments on both TensoIR Synthetic [22] and MipNeRF 360 [5] datasets, providing comprehensive insights into the effectiveness of our approach in handling complex inverse rendering tasks. Below is the detailed ablation study on Normal Regularization and Indirect Illumination." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21650" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.119, + 0.091, + 0.252, + 0.144 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.128, + 0.147, + 0.245, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.269, + 0.099, + 0.401, + 0.144 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.147, + 0.395, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.418, + 0.091, + 0.549, + 0.144 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.147, + 0.543, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.091, + 0.698, + 0.144 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.575, + 0.147, + 0.691, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.717, + 0.091, + 0.847, + 0.144 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.724, + 0.147, + 0.84, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.228, + 0.892, + 0.258 + ], + "angle": 0, + "content": "Figure 8. Relighting Visualization. We perform relighting experiments on both synthetic and real scenes using the recovered geometry, material, and illumination properties from our GS-IR method. We test our method under different lighting conditions and directions." + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.277, + 0.168, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.179, + 0.278, + 0.317, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.326, + 0.281, + 0.464, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.444, + 0.47, + 0.5 + ], + "angle": 0, + "content": "Figure 9. Ambient Occlusion Visualization. The visualization highlights the intricate shadowing and occlusion details captured by our GS-IR method, emphasizing the performance of our approach in modeling indirect illumination." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.47, + 0.749 + ], + "angle": 0, + "content": "Analysis on the Normal Regularization Reliable normal estimation is critical for conducting inverse rendering. To this end, we present depth-derivation-based regularization to facilitate 3D Gaussian-based normal estimation as stated in Sec. 4.1. In this section, we explore the impact of different acquisition schemes on the final normal quality and inverse rendering results. The quantitative results shown in Tab. 3 demonstrate that using 3D Gaussians as a normal proxy and adopting the linear interpolation strategy significantly improves the normal estimation and inverse rendering results. In addition, Fig. 7 qualitatively shows that conducting volumetric accumulation results in the floating problem (cf. Fig. 7 (b)). Despite peak selection overcomes this problem, it introduces disc aliasing (cf. Fig. 7 (c)). Compared with them, the linear interpolation adopted in GS-IR robustly produces accurate depth (cf. Fig. 7 (d))." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Analysis on the Indirect Illumination To demonstrate the effectiveness of our indirect illumination model, we compare our method with two variants: a model without occlusion volume (w/o occlusion) and a model without indirect illumination (w/o indirect illum.). The quantitative comparisons in Tab. 4 indicate that each component is crucial for estimating accurate material decomposition and generating photorealistic rendering results. Additionally, Fig. 9 showcases the ambient occlusion visualization in both synthetic and real scenes. The visualization highlights the intra-" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.276, + 0.892, + 0.368 + ], + "angle": 0, + "content": "cate shadowing and occlusion details captured by our GS-IR method, emphasizing the performance of our approach in modeling indirect illumination. This analysis further supports the effectiveness of using occlusion volume and introducing indirect illumination in enhancing the decomposition capabilities of our GS-IR." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.384, + 0.627, + 0.401 + ], + "angle": 0, + "content": "5.3. Application" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.41, + 0.892, + 0.547 + ], + "angle": 0, + "content": "We perform relighting experiments using the recovered geometry, material, and illumination from our GS-IR method. We test GS-IR under different lighting conditions and directions, observing how the reconstructed scene responds to the changes in lighting. The results of these experiments demonstrate that our GS-IR method can effectively handle relighting applications, producing photorealistic renderings under various lighting conditions. More results can be found in the supplement." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.568, + 0.619, + 0.584 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.596, + 0.892, + 0.778 + ], + "angle": 0, + "content": "We present GS-IR, a novel inverse rendering approach based on 3D Gaussian Splatting (3DGS), which employs forward mapping volume rendering to achieve photorealistic novel view synthesis and relighting results. Our GS-IR proposes an optimization scheme with depth-derivation-based regularization for normal estimation and a baking-based occlusion to model indirect lighting. These components are eventually employed to decompose material and illumination. Our extensive experiments demonstrate the effectiveness of GS-IR in achieving state-of-the-art inverse rendering results, surpassing previous neural methods in terms of both reconstruction quality and efficiency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Limitation Spherical Harmonics (SH) is only suitable for representing low-frequency, and we only use the occlusion represented by SH to model the diffuse term of indirect illumination. Modeling the specular term of indirect illumination remains a limitation of GS-IR, and has been a challenging problem in computer graphics. We believe it would be valuable to address this limitation in future work and suggest screen space global illumination (SSGI) techniques." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21651" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.468, + 0.141 + ], + "angle": 0, + "content": "[1] Unreal engine. https://www.unrealengine.com/. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.145, + 0.332, + 0.158 + ], + "angle": 0, + "content": "[2] Unity. https://unity.com/.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.161, + 0.397, + 0.174 + ], + "angle": 0, + "content": "[3] Blender. https://www.blender.org/.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.177, + 0.468, + 0.204 + ], + "angle": 0, + "content": "[4] Tomas Akenine-Mo, Eric Haines, Nate Hoffman, et al. Real-time rendering. 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.206, + 0.468, + 0.274 + ], + "angle": 0, + "content": "[5] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.277, + 0.468, + 0.345 + ], + "angle": 0, + "content": "[6] Sai Bi, Zexiang Xu, Pratul Srinivasan, Ben Mildenhall, Kalyan Sunkavalli, Miloš Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Neural reflectance fields for appearance acquisition. arXiv preprint arXiv:2008.03824, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.348, + 0.468, + 0.445 + ], + "angle": 0, + "content": "[7] Sai Bi, Zexiang Xu, Kalyan Sunkavalli, Miloš Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Deep reflectance volumes: Relightable reconstructions from multi-view photometric images. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part III 16, pages 294–311. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.447, + 0.468, + 0.516 + ], + "angle": 0, + "content": "[8] Sai Bi, Zexiang Xu, Kalyan Sunkavalli, David Kriegman, and Ravi Ramamoorthi. Deep 3d capture: Geometry and reflectance from sparse multi-view images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5960-5969, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.518, + 0.468, + 0.587 + ], + "angle": 0, + "content": "[9] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T Barron, Ce Liu, and Hendrik Lensch. Nerd: Neural reflectance decomposition from image collections. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12684-12694, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.589, + 0.468, + 0.657 + ], + "angle": 0, + "content": "[10] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan Barron, and Hendrik Lensch. Neural-pil: Neural pre-integrated lighting for reflectance decomposition. Advances in Neural Information Processing Systems, 34: 10691-10704, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.468, + 0.715 + ], + "angle": 0, + "content": "[11] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 130-141, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.468, + 0.786 + ], + "angle": 0, + "content": "[12] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5799-5809, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.468, + 0.87 + ], + "angle": 0, + "content": "[13] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[14] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Conference on Computer Vision, pages 333-350. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.123, + 0.892, + 0.164 + ], + "angle": 0, + "content": "[15] Robert L Cook and Kenneth E. Torrance. A reflectance model for computer graphics. ACM Transactions on Graphics (ToG), 1(1):7-24, 1982. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.167, + 0.892, + 0.222 + ], + "angle": 0, + "content": "[16] Yue Dong, Guojun Chen, Pieter Peers, Jiawan Zhang, and Xin Tong. Appearance-from-motion: Recovering spatially varying surface reflectance under unknown lighting. ACM Transactions on Graphics (TOG), 33(6):1-12, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.225, + 0.892, + 0.293 + ], + "angle": 0, + "content": "[17] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.296, + 0.892, + 0.366 + ], + "angle": 0, + "content": "[18] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbek Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12479–12488, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.369, + 0.892, + 0.423 + ], + "angle": 0, + "content": "[19] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, light, and material decomposition from images using monte carlo rendering and denoising. Advances in Neural Information Processing Systems, 35:22856-22869, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.426, + 0.892, + 0.495 + ], + "angle": 0, + "content": "[20] Peter Hedman, Pratul P Srinivasan, Ben Mildenhall, Jonathan T Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5875-5884, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.498, + 0.892, + 0.553 + ], + "angle": 0, + "content": "[21] Zhangjin Huang, Zhihao Liang, Haojie Zhang, Yangkai Lin, and Kui Jia. Sur2f: A hybrid representation for high-quality and efficient surface reconstruction from multi-view images. arXiv preprint arXiv:2401.03704, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.556, + 0.892, + 0.625 + ], + "angle": 0, + "content": "[22] Haian Jin, Isabella Liu, Peijia Xu, Xiaoshuai Zhang, Songfang Han, Sai Bi, Xiaowei Zhou, Zexiang Xu, and Hao Su. Tensoir: Tensorial inverse rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 165-174, 2023. 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.628, + 0.892, + 0.667 + ], + "angle": 0, + "content": "[23] Brian Karis and Epic Games. Real shading in unreal engine 4. Proc. Physically Based Shading Theory Practice, 4(3):1, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.672, + 0.892, + 0.727 + ], + "angle": 0, + "content": "[24] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 42(4):1-14, 2023. 1, 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.73, + 0.892, + 0.783 + ], + "angle": 0, + "content": "[25] Leonid Keselman and Martial Hebert. Approximate differentiable rendering with algebraic surfaces. In European Conference on Computer Vision, pages 596-614. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.787, + 0.892, + 0.828 + ], + "angle": 0, + "content": "[26] Leonid Keselman and Martial Hebert. Flexible techniques for differentiable rendering with 3d gaussians. arXiv preprint arXiv:2308.14737, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.831, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[27] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21652" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.175 + ], + "angle": 0, + "content": "[28] Zhihao Liang, Zhangjin Huang, Changxing Ding, and Kui Jia. Helixsurf: A robust and efficient neural implicit surface learning of indoor scenes with iterative intertwined regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13165-13174, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.178, + 0.472, + 0.234 + ], + "angle": 0, + "content": "[29] Fujun Luan, Shuang Zhao, Kavita Bala, and Zhao Dong. Unified shape and svbrdf recovery using differentiable monte carlo rendering. In Computer Graphics Forum, pages 101-113. Wiley Online Library, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.47, + 0.303 + ], + "angle": 0, + "content": "[30] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision, pages 405-421, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.47, + 0.361 + ], + "angle": 0, + "content": "[31] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.47, + 0.445 + ], + "angle": 0, + "content": "[32] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3d models, materials, and lighting from images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8280-8290, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.447, + 0.469, + 0.502 + ], + "angle": 0, + "content": "[33] Giljoo Nam, Joo Ho Lee, Diego Gutierrez, and Min H Kim. Practical svbrdf acquisition of 3d objects with unstructured flash photography. ACM Transactions on Graphics (TOG), 37(6):1-12, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.47, + 0.574 + ], + "angle": 0, + "content": "[34] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.469, + 0.658 + ], + "angle": 0, + "content": "[35] Pratul P Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T Barron. Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7495-7504, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.469, + 0.729 + ], + "angle": 0, + "content": "[36] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.731, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[37] Bruce Walter, Stephen R Marschner, Hongsong Li, and Kenneth E Torrance. Microfacet models for refraction through rough surfaces. In Proceedings of the 18th Eurographics conference on Rendering Techniques, pages 195-206, 2007. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[38] Rui Xia, Yue Dong, Pieter Peers, and Xin Tong. Recovering shape and spatially-varying surface reflectance under unknown illumination. ACM Transactions on Graphics (TOG), 35(6):1-12, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[39] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5438-5448, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[40] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems, 33:2492-2502, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[41] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[42] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snavely. Physg: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5453-5462, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.305, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[43] Kai Zhang, Nick Kolkin, Sai Bi, Fujun Luan, Zexiang Xu, Eli Shechtman, and Noah Snavely. Arf: Artistic radiance fields. In European Conference on Computer Vision, pages 717-733. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.362, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[44] Kai Zhang, Fujun Luan, Zhengqi Li, and Noah Snavely. Iron: Inverse rendering by optimizing neural sdfs and materials from photometric images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5565-5574, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.433, + 0.892, + 0.502 + ], + "angle": 0, + "content": "[45] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.503, + 0.892, + 0.572 + ], + "angle": 0, + "content": "[46] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (ToG), 40(6):1-18, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.574, + 0.892, + 0.642 + ], + "angle": 0, + "content": "[47] Yuanqing Zhang, Jiaming Sun, Xingyi He, Huan Fu, Rongfei Jia, and Xiaowei Zhou. Modeling indirect illumination for inverse rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18643-18652, 2022. 2, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.642 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21653" + } + ] +] \ No newline at end of file diff --git a/2024/GS-IR_ 3D Gaussian Splatting for Inverse Rendering/d0b63333-3b55-4d7f-93c1-cfbd4adae3a1_origin.pdf b/2024/GS-IR_ 3D Gaussian Splatting for Inverse Rendering/d0b63333-3b55-4d7f-93c1-cfbd4adae3a1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..df30da568569b7746f7cb4109ea45b5a21248891 --- /dev/null +++ b/2024/GS-IR_ 3D Gaussian Splatting for Inverse Rendering/d0b63333-3b55-4d7f-93c1-cfbd4adae3a1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3be00bf10d91af8b8d214d062d3a398d883ec665300f45524caa5f101acab667 +size 8428970 diff --git a/2024/GS-IR_ 3D Gaussian Splatting for Inverse Rendering/full.md b/2024/GS-IR_ 3D Gaussian Splatting for Inverse Rendering/full.md new file mode 100644 index 0000000000000000000000000000000000000000..75270fa6590197a2f3d8f2434890d31041129dcb --- /dev/null +++ b/2024/GS-IR_ 3D Gaussian Splatting for Inverse Rendering/full.md @@ -0,0 +1,340 @@ +# GS-IR: 3D Gaussian Splating for Inverse Rendering + +Zhihao Liang $^{1,\ast}$ , Qi Zhang $^{2,\ast}$ , Ying Feng $^{2}$ , Ying Shan $^{2}$ , Kui Jia $^{3,\dagger}$ + +$^{1}$ South China University of Technology, $^{2}$ Tencent AI Lab, + +$^{3}$ School of Data Science, The Chinese University of Hong Kong, Shenzhen + +eezhihaoliang@mail.scut.edu.cn, nwpuqzhang@gmail.com, + +vonyfeng@gmail.com, yingsshan@tencent.com, kuijia@cuhk.edu.cn + +# Abstract + +We propose GS-IR, a novel inverse rendering approach based on 3D Gaussian Splatting (3DGS) that leverages forward mapping volume rendering to achieve photorealistic novel view synthesis and relighting results. Unlike previous works that use implicit neural representations and volume rendering (e.g. NeRF), which suffer from low expressive power and high computational complexity, we extend 3DGS, a top-performance representation for novel view synthesis, to estimate scene geometry, surface material, and environment illumination from multi-view images captured under unknown lighting conditions. There are two main problems when introducing 3DGS to inverse rendering: 1) 3DGS does not support producing plausible normal natively; 2) forward mapping (e.g. rasterization and splatting) cannot trace the occlusion like backward mapping (e.g. ray tracing). To address these challenges, our GS-IR proposes an efficient optimization scheme incorporating a depth-derivation-based regularization for normal estimation and a baking-based occlusion to model indirect lighting. The flexible and expressive 3DGS representation allows us to achieve fast and compact geometry reconstruction, photorealistic novel view synthesis, and effective physically-based rendering. We demonstrate the superiority of our method over baseline methods through qualitative and quantitative evaluations of various challenging scenes. The source code is available at https://github.com/lzhnb/GS-IR. + +# 1. Introduction + +Inverse rendering is a long-standing task, seeking to answer the question: "How can we deduce physical attributes (e.g. geometry, material, and lighting) of a 3D scene from multiview images?" This problem is inherently challenging and + +![](images/f7d7891f2e1c9fb294c30870f7ea42156d5f26935de4c8f00b8fde40d7d9910a.jpg) + +![](images/85913a2826953e4b36701c8b55b70f16eadf47508bed410b937ef3e389a60a27.jpg) +Figure 1. Given multi-view captured images of a complex scene, we propose GS-IR (3D Gaussian Splatting for Inverse Rendering), which utilizes 3D Gaussian and forward mapping splatting to recover high-quality physical properties (e.g., normal, material, illumination). This enables us to perform relighting and material editing, resulting in outstanding inverse rendering results. Better viewed on screen with zoom in, especially the remarkable material decomposition and normal reconstruction of bicycle axle. + +![](images/bb1d0adc675c004c79921e17e488466f8e935f6e5053f2e95f2521b0cd04bb26.jpg) + +ill-posed, particularly when input images are captured in uncontrolled environments with unknown illumination. Recent research [9, 10, 35, 46] has sought to address this issue by employing implicit neural representations akin to NeRF [30] that utilizes multi-layer perceptrons (MLPs). However, current methods incorporating MLP face challenges in terms of their low expressive capacity and high computational demands, which significantly limits the effectiveness and efficiency of inverse rendering, especially when it cannot be rendered at interactive rates. + +3D Gaussian Splatting (3DGS) [24] has recently emerged as a promising technique to model 3D static scenes and significantly boost the rendering speed to a real-time level. It makes the scene representation more compact and achieves fast and top performance for novel view synthesis. Introducing it to the inverse rendering pipeline is natural and essential, including geometry reconstruction, materials decomposition, and illumination estimation. Unlike ray trac + +ing in NeRF, 3DGS produces a set of 3D Gaussians around sparse points. During the 3DGS optimization, the adaptive control of the Gaussian density may lead to loose geometry, making it difficult to estimate accurate scene's normal. Consequently, it is necessary to introduce a well-designed strategy to regularize 3DGS's normal estimation. + +Our goal is to use 3D Gaussians as the scene representation for inverse rendering from multi-view images captured under unknown lighting conditions. However, capturing observations under natural illumination often shows complex effects such as soft shadows and interreflections. TensoIR [22] leverages the ray tracing of NeRF to directly model occlusion and indirect illumination. In contrast, 3DGS replaces the ray tracing in the NeRF with differentiable forward mapping volume rendering, which directly projects 3D Gaussians onto the 2D plane. This strategy improves the rendering efficiency but makes it difficult to calculate occlusion. Inspired by the "Indirect Lighting Cache" used in real-time rendering [4], we attempt to bake the occlusion into volumes for caching. + +In this paper, we present a novel 3D Gaussian-based inverse rendering framework called GS-IR (3D Gaussian Splatting for Inverse Rendering) that leverages forward mapping splatting to deduce the physical attributes of a complex scene. To the best of our knowledge, our method is the first work to introduce the 3DGS technique for inverse rendering, which can simultaneously estimate scene geometry, materials, and illumination from multi-view images. Our GS-IR addresses two main issues when using 3DGS for inverse rendering. Firstly, we develop an intuitive and well-designed regularization to estimate the scene's normal. Secondly, we use a baking-based method embedded in GS-IR to cache occlusions, obtaining an efficient indirect illumination model. As shown in Fig. 1, our approach can reconstruct high-fidelity geometry and materials of a complex real scene under unknown natural illumination, enabling state-of-the-art rendering of novel view synthesis and additional applications like relighting. Our technical contributions are summarized as follows: + +- We present GS-IR that models a scene as a set of 3D Gaussians to achieve physically-based rendering and state-of-the-art decomposition results for both objects and scenes; +- We propose an efficient optimization scheme with regularization to concentrate depth gradient around 3DGS and produce reliable normals for GS-IR; +- We develop a baking-based method embedded in GS-IR to handle the occlusion in modeling indirect lighting; + +We demonstrate the superiority of our method to baseline methods qualitatively and quantitatively on various challenging scenes, including the TensoIR synthesis dataset [22] and Mip-NeRF 360 real dataset [5]. + +# 2. Related Works + +Neural Representation Recently, neural rendering techniques, exemplified by Neural Radiance Field (NeRF) [30], have achieved impressive success in addressing visual computing problems, giving rise to numerous neural representations [14, 21, 31, 36, 39, 40] tailored for different tasks [11-13, 18, 27, 28, 34, 43]. The vanilla NeRF models a continuous radiance field implicitly in MLPs, which requires massive repeated queries during training and inference. To address the computational inefficiencies, many neural scene representations are proposed with more discretized geometry proxies such as voxel grids [17, 20, 36], hash grids [31], tri-planes [14] or points [39]. Neural features are stored in a structured manner, allowing for efficient storage and retrieval. The computational cost can thus be significantly reduced by introducing interpolation techniques, however, with an inevitable loss in image quality. 3D Gaussians are introduced as an unstructured scene representation to strike a balance between efficiency and quality [24]. With the specially designed tile-based rasterizer for Gaussian splats, this method achieves real-time rendering with high quality for novel-view synthesis. In this work, 3D Gaussian representation is combined with the physical-based rendering (PBR) model for inverse rendering. + +Inverse Rendering Inverse rendering aims to decompose the image's appearance into the geometry, material, and lighting conditions. Considering the inherent ambiguity between observed images and underlying properties, many methods are proposed with different constrained settings, such as capturing images with fixed lighting and rotating object [16, 38], capturing with moving camera and colocated lighting [7, 8, 29, 33]. Combined with neural representations, inverse rendering models the scene simulating how the light interacts with the neural volume with various material properties, and estimates the lighting and material parameters during optimization [6, 9, 10, 19, 22, 35, 42, 44, 46, 47]. Neural Reflectance Fields [6] assumes a known point light source and represents the scene as a field of volume density, surface normals, and bi-directional reflectance distribution functions(BRDFs) with one bounce direct illumination. NeRV [35] and InvRender [47] extend to arbitrary known lighting conditions and train an additional MLP to model the light visibility. PhySG [42] assumes full light source visibility without shadow simulation, and represents the lighting and scene BRDFs with spherical Gaussians for acceleration. TensoIR [22] adopts the efficient TensoRF [14] representation which enables the computation of visibility and indirect lighting by raytracing, while limited to object-level. For modeling surface geometry using point clouds, Fuzzy Metaballs (FMs) [25, 26] offers a great way to render depth from 3D Gaussian using Order Independent Transparency (OIT) and approximate intersection. However, it requires silhouettes as input and struggles to handle + +intricate geometry (e.g. Lego and Ficus) let alone complex scenes. In this work, we propose a 3DGS-based pipeline to recover the geometry, material, and lighting that is available for both objects and unbounded scenes. + +# 3. Preliminary + +In this section, we give the technical backgrounds and math symbols that are necessary for the presentation of our proposed method in subsequent sections. + +3D Gaussian Splatting (3DGS) [24] is an explicit 3D scene representation in the form of point clouds. Each point is represented as a Gaussian function $g$ that approximates the shape of a bell curve, which is defined as, + +$$ +g (\boldsymbol {x} | \boldsymbol {\mu}, \boldsymbol {\Sigma}) = e ^ {- \frac {1}{2} (\boldsymbol {x} - \boldsymbol {\mu}) ^ {T} \boldsymbol {\Sigma} ^ {- 1} (\boldsymbol {x} - \boldsymbol {\mu})}, \tag {1} +$$ + +where $\pmb{\mu} \in \mathbb{R}^3$ is its mean vector, and $\pmb{\Sigma} \in \mathbb{R}^{3 \times 3}$ is an anisotropic covariance matrix. The mean vector $\pmb{\mu}$ of a 3D Gaussian is parameterized as $\pmb{\mu} = (\mu_x, \mu_y, \mu_z)$ , and the covariance matrix $\pmb{\Sigma}$ is factorized into a scaling matrix $\pmb{S}$ and a rotation matrix $\pmb{R}$ as $\pmb{\Sigma} = R\pmb{S}\pmb{S}^\top\pmb{R}^\top$ . $\pmb{S}$ and $\pmb{R}$ refer to a diagonal matrix $\mathrm{diag}(s_x, s_y, s_z)$ and a rotation matrix constructed from a unit quaternion $\pmb{q}$ . Given a viewing transformation with extrinsic matrix $\pmb{T}$ and intrinsic matrix $\pmb{K}$ , the mean vector $\pmb{\mu}$ and covariance matrix $\pmb{\Sigma}'$ from the 3D point $\pmb{x}$ to 2D pixel $\pmb{u}$ is defined as, + +$$ +\boldsymbol {\mu} ^ {\prime} = \boldsymbol {K} \boldsymbol {T} [ \boldsymbol {\mu}, 1 ] ^ {\top}, \boldsymbol {\Sigma} ^ {\prime} = \boldsymbol {J} \boldsymbol {T} \boldsymbol {\Sigma} \boldsymbol {T} ^ {\top} \boldsymbol {J} ^ {\top}, \tag {2} +$$ + +where $\pmb{J}$ is the Jacobian matrix of the affine approximation of the perspective projection. Besides, each Gaussian represents the view-dependent color $c_{i}$ via a set of coefficients of spherical harmonics (SH), which is then multiplied by opacity $\alpha$ for volume rendering. We finally obtain the color $\hat{\pmb{C}}$ at pixel $\pmb{u}$ based on Eq. (1) and Eq. (2), + +$$ +\hat {\boldsymbol {C}} = \sum_ {i \in N} T _ {i} g _ {i} \left(\boldsymbol {u} \mid \boldsymbol {\mu} ^ {\prime}, \boldsymbol {\Sigma} ^ {\prime}\right) \alpha_ {i} \boldsymbol {c} _ {i}, \quad T _ {i} = \prod_ {j = 1} ^ {i - 1} \left(1 - g _ {j} \left(\boldsymbol {u} \mid \boldsymbol {\mu} ^ {\prime}, \boldsymbol {\Sigma} ^ {\prime}\right) \alpha_ {j}\right), \tag {3} +$$ + +where accumulated transmittance $T_{i}$ quantifies the probability density of $i$ -th Gaussian at pixel $\mathbf{u}$ . + +The Rendering Equation In GS-IR, we leverage the classic rendering equation to formulate the outgoing radiance of a surface point $\mathbf{x}$ with normal $\mathbf{n}$ : + +$$ +L _ {o} (\boldsymbol {x}, \boldsymbol {v}) = \int_ {\Omega} L _ {i} (\boldsymbol {x}, \boldsymbol {l}) f _ {r} (\boldsymbol {l}, \boldsymbol {v}) (\boldsymbol {l} \cdot \boldsymbol {n}) d \boldsymbol {l}, \tag {4} +$$ + +$\Omega$ denotes the upper hemisphere centered at $\pmb{x}$ , $\pmb{l}$ and $\pmb{v}$ denote incident and view directions respectively. $L_{i}(\pmb{x},\pmb{l})$ denotes the radiance received at $\pmb{x}$ from $\pmb{l}$ . Notably, we follow Cook-Torrance microfacet model [15, 37] and formulate the bidirectional reflectance distribution function (BRDF) $f_{r}$ as a function of albedo $\pmb{a} \in [0,1]^3$ , metallic $m \in [0,1]$ , and roughness $\rho \in [0,1]$ : + +$$ +f _ {r} (\boldsymbol {l}, \boldsymbol {v}) = \underbrace {(1 - m) \frac {\boldsymbol {a}}{\pi}} _ {\text {d i f f u s e}} + \underbrace {\frac {D F G}{4 (\boldsymbol {n} \cdot \boldsymbol {l}) (\boldsymbol {n} \cdot \boldsymbol {v})}} _ {\text {s p e c u l a r}}, \tag {5} +$$ + +where microfacet distribution function $D$ , Fresnel reflection $F$ , and geometric shadowing factor $G$ are related to the surface roughness $\rho$ . We use 3D Gaussians to store these material properties in GS-IR. + +# 4. Method + +Given a set of calibrated RGB images $\{I_m\}_{m=1}^M$ of a target scene captured from multiple views under static, yet unknown illumination, inverse rendering aims to decompose the scene's intrinsic properties, including normal, materials, and illumination. This decomposition facilitates the recovery and subsequent edition of the target scene. Motivated by the remarkable performance in quality and speed of 3DGS [24], we present a novel framework GS-IR consisting of three well-designed stage strategies, as shown in Fig. 2. In the initial stage, we leverage differentiable splating to optimize 3D Gaussians. Concurrently, we utilize the gradient derived from the rendered depth map to supervise the normal stored in 3D Gaussians (cf. Sec. 4.1). In the second stage, we precompute the occlusion based on the learned geometric information (i.e. depth and normal) and store it in an efficient spherical harmonics-based architecture to model indirect illumination (cf. Sec. 4.2). In the final stage, we combine a differentiable splating with the physical-based rendering (PBR) pipeline to optimize illumination and material-aware 3D Gaussians (cf. Sec. 4.3). + +# 4.1. Normal Reconstruction + +During the initial stage, we optimize 3D Gaussians for geometry reconstruction from observed images, denoted as $\mathcal{G}$ . The optimized $\mathcal{G}$ functions as a geometric proxy for surface points and their corresponding normals $\pmb{n}$ , which are crucial for successful inverse rendering. As highlighted in Sec. 1, generating reasonable normals within the 3DGS-based framework poses a significant challenge. To address this obstacle, we introduce an intuitive strategy that improves depth $\hat{D}$ and leverages the depth gradient to derive pseudo normals $\hat{\pmb{n}}_{\hat{D}} = \nabla_{\mathrm{uv}}\hat{D}$ . These pseudo normals then guide the optimization of normals within the 3D Gaussians. + +Depth Generation Given a pretrained 3D Gaussians $\mathcal{G}$ and a view designated for rendering, the pixel's shading results in that view can be obtained by Eq. (3). Consequently, it is reasonable to utilize the same volumetric accumulation to compute the depth $\hat{D} = \sum_{i=1}^{N} T_i \alpha_i d_i$ , where $d_i$ denotes the distance from the corresponding 3D Gaussian to the image plane. However, we observed the floating problem during volumetric accumulation, unlike the backward mapping volume rendering used in NeRF. During the 3DGS optimization, the adaptive control of the Gaussian density may result in the depth falling in front of the 3D Gaussians, thereby posing challenges in accurately predicting the depth. Specifically, the backward mapping methods can obtain an accurate depth by considering only peak samples, + +![](images/520935b331613e8ca20236530cf0e59333c0c3f429eea69a2269ca4991750034.jpg) +Figure 2. GS-IR Pipeline. We propose a novel GS-based inverse rendering framework, called GS-IR, to reconstruct scene geometry, materials, and unknown natural illumination from multi-view captured images. Our GS-IR consists of three well-designed stage strategies using 3D Gaussian and differentiable forward mapping splatting to achieve physical-based rendering. In our approach, the Gaussian stores not only the basic 3DGS information but also the normal and material properties, enhancing its capabilities for inverse rendering tasks. + +![](images/ca483dfb406e9c50d41bba16dfb9a89fa5f2196c656b2ec5fc2847d0a156fea0.jpg) +Figure 3. Depth Illustration. By considering the depth as a linear interpolation of the distances from 3D Gaussians to the image plane, and ensuring it lies between the minimum and maximum distance, our method could produce accurate depth. + +that is $\hat{D} = d_{i^*}$ , where $i^* = \arg \max_i T_i \alpha_i$ . However, for 3DGS, a typical forward mapping method, the peak selection results in disc aliasing within 3DGS. To overcome this limitation, we consider that depth $\hat{D}$ must be between the minimum and maximum distance of 3D Gaussians to the image plane, as illustrated in Fig. 3. We then treat the depth as a linear interpolation of the distances from 3D Gaussians to the image plane: + +$$ +\hat {D} = \sum_ {i = 1} ^ {N} \hat {w} _ {i} d _ {i}, \quad \hat {w} _ {i} = \frac {T _ {i} \alpha_ {i}}{\sum_ {i = 1} ^ {N} T _ {i} \alpha_ {i}}. \tag {6} +$$ + +Normal Derivation While the accurate prediction of depth within Gaussians provides better guidance for the normal reconstruction, directly using depth gradient to produce normals has two limitations that still cannot meet the require + +ment for effective inverse rendering. First, the depth gradient estimation is highly sensitive to noise, making the predicted normal often extremely noisy; Second, the normals derived individually from each view's depth map do not satisfy multi-view consistency. To address these issues, we use Gaussian $\mathcal{G}$ as a proxy for normal estimation instead of directly from the depth gradient. Benefiting from the efficiency of 3DGS, we obtain the depth $\hat{D}$ and normals $\hat{\pmb{n}}$ of the observed view after performing a single rendering pass. We then tie these predicted pseudo normal to the underlying depth gradient normal $\hat{\pmb{n}}_{\hat{D}}$ using a simple penalty: + +$$ +\mathcal {L} _ {n - p} = \left\| \hat {\boldsymbol {n}} - \hat {\boldsymbol {n}} _ {\hat {D}} \right\|, \tag {7} +$$ + +where $\hat{\pmb{n}} = \sum_{i=1}^{N} T_i \alpha_i \pmb{n}_i$ , and $\pmb{n}_i$ is the normal stored in the 3D Gaussian. Secondly, unlike the MLP-based normal estimation [22] acting MLP as a low-pass filter, the predicted normal of Gaussian $\mathcal{G}$ is rough, so smoothness regularization should be included. We introduce the TV term $TV_{\text{normal}}$ to smooth the predicted normal $\hat{\pmb{n}}$ . For more details, please refer to the supplement. + +In optimization of the first stage, we optimize 3D Gaussians $\mathcal{G}$ (storing SH coefficients for view-dependent color $c$ , opacity $\alpha$ , and normal $\hat{n}$ ) by using the color reconstruction loss $\mathcal{L}_c$ , which is the same as 3DGS [24], and the proposed normal loss $\mathcal{L}_n$ , + +$$ +\mathcal {L} _ {n} = \mathcal {L} _ {n - p} + \lambda_ {n - T V} T V _ {\text {n o r m a l}}. \tag {8} +$$ + +![](images/122d068955408441c595b1a88809c97880f2d58b74013a64e5351d41b7f52f34.jpg) +Figure 4. Baking. We employ the spherical harmonics (SH) architecture to bake occlusion volumes for modeling indirect illumination. For each grid of occlusion volumes, we initially use 3D Gaussians to compute the depth cubemap by performing six forward mapping splatting passes. Next, we convert the depth cubemap into a binary occlusion cubemap based on a distance threshold. Finally, the occlusion cubemap is baked as SH coefficients, enabling efficient interpolation of the occlusion cubemap at any point within the scene. + +# 4.2. Indirect Illumination Modeling + +Drawing inspiration from the successful implementation of precomputation techniques in the video game industry (e.g. Irradiance Volume in Blender [3], Lightmass Volume in Unreal [1], and Light Probes in Unity [2]), we introduce spherical harmonics (SH) architectures to store occlusion information and model indirect illumination. + +Given the optimized 3D Gaussians $\mathcal{G}$ from one stage (cf. Sec. 4.1), we freeze $\mathcal{G}$ and regularly place occlusion volumes $\nu^{\mathrm{occl}}$ in the 3D space. For each volume $\pmb{v}_i^{\mathrm{occl}}\subset \nu^{\mathrm{occl}}$ we then cache the occlusion in the form of SH coefficients $\pmb{f}_i^{\mathrm{o}}$ . Consequently, the formula of the occlusion $O(\cdot)$ of $\pmb{v}_i^{\mathrm{occl}}$ with respect to the direction $(\theta ,\phi)$ is expressed as: + +$$ +O (\theta , \phi) = \sum_ {l = 0} ^ {\text {d e g}} \sum_ {m = - l} ^ {l} \mathbf {f} _ {i (l m)} ^ {0} Y _ {l m} (\theta , \phi), \tag {9} +$$ + +where $deg$ denotes the degree of SH, and $\{Y_{lm}(\cdot)\}$ is a set of real basis of SH. + +As discussed in Sec. 4.1, the 3DGS technique employs a forward mapping approach that projects 3D points to the 2D plane, in contrast to the backward mapping volume rendering utilized in NeRF, which means it cannot use ray marching to calculate occlusions. To precompute the SH coefficients $\pmb{f}_i^{\mathrm{o}}$ of occlusion volume $\pmb{v}_i^{\mathrm{occl}}$ , we obtain the depth cubemap $\{\hat{D}_p^i\}_{p=1}^6$ by performing six times rendering passes, once for each face of the cubemap. We then convert it into a binary occlusion cubemap $\{\hat{O}_p^i\}_{p=1}^6$ based on a manually set distance threshold. + +Finally, we convolve the the occlusion cubemap + +$\{\hat{O}_p^i\}_{p = 1}^6$ using SH bases and get the SH coefficients $\pmb{f}_i^0$ .. + +$$ +\begin{array}{l} \boldsymbol {f} _ {i (l m)} ^ {0} = \int_ {S ^ {2}} \hat {\boldsymbol {O}} (\boldsymbol {\omega}) Y _ {l m} (\boldsymbol {\omega}) d \boldsymbol {\omega} \tag {10} \\ = \int_ {0} ^ {2 \pi} \int_ {0} ^ {\pi} \sin \theta \hat {\boldsymbol {O}} (\theta , \phi) Y _ {l m} (\theta , \phi) d \theta d \phi , \\ \end{array} +$$ + +where $S^2$ denotes the unit sphere, and $\hat{O} (\theta ,\phi)$ denotes the occlusion query from the occlusion cubemap $\{\hat{O}_p^i\}_{p = 1}^6$ Note that we numerically calculate the convolution in Eq. (10) in parallel. The caching process is shown in Fig. 4. + +To handle the indirect illumination in occlusion regions, we also maintain illumination volumes $\mathcal{V}^{\mathrm{illu}}$ to cache the indirect illumination. Similar to the caching process of occlusion volumes, the caching target of illumination volumes changes from the occlusion cubemap $\{\hat{O}_p^i\}_{p=1}^6$ to the captured environment cubemap $\{\hat{I}_p^i\}_{p=1}^6$ . These cubemaps can be obtained simultaneously by conducting six rendering passes. For more details, please refer to the supplement. + +# 4.3. Intrinsic Decomposition + +In the final stage, we employ differentiable splatting in conjunction with a PBR pipeline to accomplish the intrinsic decomposition. According to Eq. (5), the rendering equation Eq. (4) is rewritten as diffuse $L_{\mathrm{d}}$ and specular $L_{\mathrm{s}}$ components: + +$$ +\begin{array}{l} L _ {o} (\boldsymbol {x}, \boldsymbol {v}) = \int_ {\Omega} \left[ (1 - m) \frac {\boldsymbol {a}}{\pi} + \frac {D F G}{4 (\boldsymbol {n} \cdot \boldsymbol {l}) (\boldsymbol {n} \cdot \boldsymbol {v})} \right] L _ {i} (\boldsymbol {x}, \boldsymbol {l}) (\boldsymbol {l} \cdot \boldsymbol {n}) d \boldsymbol {l} \\ L _ {\mathrm {d}} = (1 - m) \frac {\boldsymbol {a}}{\pi} \int_ {\Omega} L _ {i} (\boldsymbol {x}, \boldsymbol {l}) (\boldsymbol {l} \cdot \boldsymbol {n}) d \boldsymbol {l} \tag {11} \\ L _ {\mathrm {s}} = \int_ {\Omega} \frac {D F G}{4 (\boldsymbol {n} \cdot \boldsymbol {l}) (\boldsymbol {n} \cdot \boldsymbol {v})} L _ {i} (\boldsymbol {x}, \boldsymbol {l}) (\boldsymbol {l} \cdot \boldsymbol {n}) d \boldsymbol {l}. \\ \end{array} +$$ + +In GS-IR, we adopt an image-based lighting (IBL) model and split-sum approximation [23] to tackle the intractable integral. To calculate the diffuse component $L_{\mathrm{d}}$ , the illumination $I_{\mathrm{d}}$ is defined as: + +$$ +\begin{array}{l} I _ {\mathrm {d}} (\boldsymbol {x}) = \int_ {\Omega} L _ {i} (\boldsymbol {x}, \boldsymbol {l}) (\boldsymbol {l} \cdot \boldsymbol {n}) d \boldsymbol {l} \\ = \int_ {\Omega_ {\mathrm {V i s}}} L _ {i} ^ {\operatorname {d i r}} (\boldsymbol {x}, \boldsymbol {l}) (\boldsymbol {l} \cdot \boldsymbol {n}) d \boldsymbol {l} + \int_ {\Omega_ {\mathrm {O c c l}}} L _ {i} ^ {\operatorname {i n d i r}} (\boldsymbol {x}, \boldsymbol {l}) (\boldsymbol {l} \cdot \boldsymbol {n}) d \boldsymbol {l} \tag {12} \\ \approx \left(1 - \mathrm {O} (\boldsymbol {x})\right) I _ {\mathrm {d}} ^ {\operatorname {d i r}} (\boldsymbol {x}) + \mathrm {O} (\boldsymbol {x}) I _ {\mathrm {d}} ^ {\operatorname {i n d i r}} (\boldsymbol {x}), \\ \end{array} +$$ + +where the first component indicates direct illumination and the second is indirect illumination. Notably, our baking-based indirect illumination model enables us to calculate the occlusion and illumination online. This means that our GS-IR achieves intrinsic decomposition while maintaining real-time rendering performance. For the specular component $L_{\mathrm{s}}$ , we follow split-sum approximation and treat the integral as two separate integrals: + +$$ +\begin{array}{l} L _ {\mathrm {s}} = \int_ {\Omega} \frac {D F G}{4 \langle \boldsymbol {n} \cdot \boldsymbol {l} \rangle \langle \boldsymbol {n} \cdot \boldsymbol {v} \rangle} L _ {i} (\boldsymbol {l}) \langle \boldsymbol {l} \cdot \boldsymbol {n} \rangle d \boldsymbol {l} \\ \approx \underbrace {\int_ {\Omega} \frac {D F G}{4 \left\langle \boldsymbol {n} \cdot \boldsymbol {l} \right\rangle \left\langle \boldsymbol {n} \cdot \boldsymbol {v} \right\rangle}} _ {\text {E n v i r o n m e n t B R D F -} R} \left\langle \boldsymbol {l} \cdot \boldsymbol {n} \right\rangle d \boldsymbol {l} \underbrace {\int_ {\Omega} D L _ {i} (\boldsymbol {l}) \left\langle \boldsymbol {l} \cdot \boldsymbol {n} \right\rangle d \boldsymbol {l}} _ {\text {P r e - F i l t e r e d E n v i r o n m e n t M a p -} I _ {s}}, \tag {13} \\ \end{array} +$$ + +![](images/fec20d9f6c6a5555aabe002a85419f33cc928bbe4f577ec908c4c453d5265c6d.jpg) +Figure 5. Qualitative comparison on TensoIR Synthetic. We visualize the estimated normal, albedo, and rendering results of our GS-IR and baseline methods on two scenes. By utilizing the efficient 3D Gaussian representation and a robust tile-based rasterizer, GS-IR achieves rapid convergence and supports real-time rendering. This performance advantage underscores the effectiveness of our method in addressing complex inverse rendering tasks, thereby surpassing existing state-of-the-art approaches. (For albedo reconstruction results, we follow NeRFactor [46] and scale each RGB channel by a global scalar.) + +where both $R$ and $I_{\mathrm{s}}$ can be precomputed in advance and stored in look-up tables. With Eq. (12) and Eq. (13), the rendering results of Eq. (11) can be represented as: + +$$ +\begin{array}{l} L _ {o} (\boldsymbol {x}, \boldsymbol {v}) = L _ {\mathrm {d}} + L _ {\mathrm {s}} \\ \approx (1 - m) \frac {\boldsymbol {a}}{\pi} \left[ (1 - \mathrm {O} (\boldsymbol {x})) I _ {\mathrm {d}} ^ {\operatorname {d i r}} (\boldsymbol {x}) + \mathrm {O} (\boldsymbol {x}) I _ {\mathrm {d}} ^ {\operatorname {i n d i r}} (\boldsymbol {x}) \right] + R I _ {\mathrm {s}}. \tag {14} \\ \end{array} +$$ + +For intrinsic decomposition, we optimize the material $\hat{M}$ (i.e. albedo $a$ , metallic value $m$ , and roughness $\rho$ ) stored in 3D Gaussians $\mathcal{G}$ , environment map $\hat{\pmb{E}}$ , and illumination volumes $\mathcal{V}^{\mathrm{illu}}$ by minimizing the decomposition loss $\mathcal{L}_d$ : + +$$ +\mathcal {L} _ {d} = \underbrace {\left\| \boldsymbol {I} - \hat {\boldsymbol {I}} ^ {\text {s h a d e}} \left(\hat {\boldsymbol {M}} , \hat {\boldsymbol {E}} , \mathcal {V} ^ {\text {i l l u}}\right) \right\|} _ {\mathcal {L} _ {\text {s h a d e}}} + \underbrace {\lambda_ {M} T V _ {\mathrm {m a t}}} _ {\mathcal {L} _ {\text {m a t e r i a l}}} + \underbrace {\lambda_ {\boldsymbol {E}} T V _ {\text {l i g h t}}} _ {\mathcal {L} _ {\text {l i g h t}}}, \tag {15} +$$ + +where $\mathcal{L}_{\mathrm{shade}}$ indicate the shade loss. $\hat{I}^{\mathrm{shade}}(\hat{M},\hat{E},\hat{\mathcal{V}}^{\mathrm{illu}})$ is the recovered image that uses the PBR pipeline defined by Eq. (14). Please refer to the supplement for more details about the material TV loss $\mathcal{L}_{\mathrm{material}}$ and lighting TV loss $\mathcal{L}_{\mathrm{light}}$ . + +# 5. Experiments + +Dataset & Metrics We conduct experiments using benchmark datasets of TensoIR Synthetic [22] and Mip-NeRF 360 [5] for decomposing both objects and scenes. They contain 4 objects with reference materials and 7 publicly available scenes, respectively. To verify the efficacy of our normal reconstruction, we evaluate the normal quality on the TensoIR Synthetic [22] dataset using mean angular error (MAE). We further assess our reconstructed albedo quality on this synthetic dataset. More generally, we evaluate the synthesized novel view on both datasets in terms of Peak + +![](images/adbe248cba838b012773c77d3e24a54dc606e0633c32d31505fa1ee3bf19c23c.jpg) + +![](images/0c29f848a8fb7d39cef3e865df897a311b5552f97d19c22e4630297fd7751282.jpg) + +![](images/7981f9ed67a6b735312df99f5e779e5291f3247653e2de0a592343aa1f587c43.jpg) +Figure 6. Novel view synthesis results on Mip-NeRF 360. GS-IR can reconstruct scene details including geometric normals and high-frequency appearance, rendering high-fidelity appearance and recovering fine geometric details such as those on leaves and bicycle axles. Better viewed on screen with zoom in. + +![](images/f009744cc300cc761df0f99cd28e47e2da474757ec2ff377fada9bdc1c075103.jpg) + +Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS) [45]. Note that albedo quality assessment uses the same metrics as novel view synthesis. + +# 5.1. Comparisons + +We conduct a comprehensive comparison against state-of-the-art neural field-based inverse rendering methods on the public TensoIR Synthetic dataset [22]. All the methods utilize multi-view images captured under unknown lighting conditions. Our evaluation encompasses normal quality (measured by MAE), novel view synthesis, albedo fidelity, + +
MethodNormal ↓ MAE ↓Novel View SynthesisAlbedoRelightRuntime
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
NeRFactor [46]6.31424.6790.9220.12025.1250.9400.10923.3830.9080.131>100 hrs
InvRender [47]5.07427.3670.9340.08927.3410.9330.10023.9730.9010.10115 hrs
NVDiffrec [32]6.07830.6960.9620.05229.1740.9080.11519.8800.8790.104<1 hr
TensoIR [22]4.10035.0880.9760.04029.2750.9500.08528.5800.9440.0815 hrs
Ours4.94835.3330.9740.03930.2860.9410.08424.3740.8850.096<1 hr
+ +Table 1. Quantitative Comparison on TensoIR Synthetic dataset. Our method outperforms baseline methods in terms of novel view synthesis and albedo quality, showcasing the effectiveness of material decomposition and PBR rendering. This is particularly noteworthy considering that our normal reconstruction is slightly inferior to TensoIR. In terms of relighting performance, we rank second, trailing only behind TensoIR. Importantly, the average training time of our GS-IR is accelerated by a factor of $5\mathrm{x}$ , making its performance acceptable and further demonstrating the effectiveness of our approach in handling complex inverse rendering tasks. + +
MethodPSNR ↑SSIM ↑LPIPS ↓Runtime ↓
NeRF++ [41]25.1120.6960.375≈ 20h
Plenoxels [17]23.0790.6250.462≈ 30m
INGP-Base [31]25.3030.6710.371≈ 5m
INGP-Big [31]25.5870.6990.331≈ 8m
Mip-NeRF 360 [31]27.5690.7930.234≈ 48h
3DGS [24]27.210.8150.214≈ 35m
Ours25.3810.7570.267≈ 45m
+ +relighting effects (measured by PSNR, SSIM, and LPIPS), and efficiency. Tab. 1 summarizes the quantitative comparisons on the synthesis dataset. Our method achieves superior performance in novel view synthesis and albedo quality compared to the baseline methods, demonstrating the effectiveness of material decomposition and PBR rendering, particularly given that our normal reconstruction is slightly inferior to TensoIR. Our relighting performance ranks second, only behind TensoIR. Notably, the average training time of our GS-IR is accelerated by a factor of $5\mathrm{x}$ , making its performance acceptable and demonstrating the effectiveness of our approach in handling complex inverse rendering tasks. We also include qualitative comparisons in Fig. 5, which show that our GS-IR produces reasonable albedo and photorealistic renderings that are closer to the ground truth than most methods. + +Meanwhile, owing to our more efficient and compact representation with powerful expressiveness, our method showcases remarkable performance on complex real unbounded scenes [5]. Tab. 2 presents the quantitative comparisons on the real dataset. Fig. 6 demonstrates the normal reconstruction and novel view synthesis on the real dataset. Our method renders a high-fidelity appearance and recovers fine geometric details, such as those on the leaves and bicycle axil. In summary, by leveraging the efficient 3D Gaussian representation and a powerful tile-based rasterizer, GS-IR achieves fast convergence and supports real-time rendering. This performance advantage highlights the effectiveness of our method in handling complex inverse rendering + +Table 2. Quantitative Comparison on Mip-NeRF 360. The results show that our inverse rendering approach even surpasses some NeRF variants dedicated to novel view synthesis. + +
MethodTensoIR Synthetic [22]Mip-NeRF 360 [5]
Normal MAE↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Vol. Accum.16.34725.7560.8550.13122.0520.6100.394
Peak Selec.9.46628.7500.9270.08423.0930.7190.317
Linear Interp.6.21828.9830.9390.06623.1190.7070.319
Vol. Accum.†9.31530.0910.9400.07124.6640.7470.277
Peak Selec.†7.98631.0640.9500.06025.1430.7530.281
Linear Interp.†4.94835.3330.9740.03925.3810.7570.267
+ +Table 3. Analyses on the impact of different depth generation strategies on normals. Methods without $\dagger$ marks directly use the normals derived from the depth map; Methods marked with $\dagger$ use depth derivation to optimize the normals stored in 3D Gaussians. + +![](images/a4b22106eecb0f719527926e096a12f57a24dbe6c81dccde32433f0e29266c8c.jpg) + +Figure 7. Visual comparison of depth produced by different strategies. The linear interpolation adopted in GS-IR overcomes the floating problem and disc aliasing. + +
MethodTensoIR Synthetic [22]Mip-NeRF 360 [5]
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
w/o occlusion34.9970.9620.04125.0600.7530.270
w/o indirect illum.35.1860.9650.04424.8980.7490.272
Ours35.3330.9740.03925.3810.7570.267
+ +Table 4. Analyses on the occlusion and indirect illumination. Physically modeling indirect illumination improves the inverse rendering of objects and scenes. + +tasks, outperforming existing state-of-the-art approaches. + +# 5.2. Ablation Studies + +We initially introduce the 3DGS technique for inverse rendering in GS-IR and propose depth-derivation-based normal regularization and a baking-based method to address the challenges encountered during the process. To evaluate the efficacy of our proposed schemes, we design elaborate experiments on both TensoIR Synthetic [22] and MipNeRF 360 [5] datasets, providing comprehensive insights into the effectiveness of our approach in handling complex inverse rendering tasks. Below is the detailed ablation study on Normal Regularization and Indirect Illumination. + +![](images/06e1ddad6069322fee46aaec46d4d73b83ac3a589b44435883ab6aac4844ac13.jpg) + +![](images/26d62ae8ab69950521c2423c6f6751c682f5cb5bf22a2798efacd7f2e1699494.jpg) +Figure 8. Relighting Visualization. We perform relighting experiments on both synthetic and real scenes using the recovered geometry, material, and illumination properties from our GS-IR method. We test our method under different lighting conditions and directions. + +![](images/770b9ac4012b5565b75c44c29eedb055bb460b045d5b9948f5edc2fd4001eaa1.jpg) + +![](images/40924564f4e82eb75884a7b39760ffaf3502d68ed7d3d77378de34087d413b56.jpg) + +![](images/6d57db72969d48d2f949c3fe9796e4a458fe0c6d606f2a4a23d5b7d23eaedb3d.jpg) + +![](images/b9c42674746ef5fc3aed59a7c0e80fd2c1baae462b5123fad4059003e6c3f4b1.jpg) + +![](images/7b1d79b19403f75c45aa69811fc2c11d876d60a054edadf0af4641b008fdb050.jpg) + +![](images/8e3aae9e547181498b68f34deddc2cd236575c22f65383ecb0f7eef436873df2.jpg) + +![](images/f7364b1a86329aacf374d3cf64b5294ff2be41accc5b74c3206494d9c5b9cce7.jpg) + +![](images/05a653bd301302f501abb5f081f5f6f0506a8fe2181bc5001bf638f5aac93307.jpg) + +![](images/4738bfbdf4bb0161b98a7a8d26f6fc3f5f82b475312889a06f42cc0fbd88aae8.jpg) +Figure 9. Ambient Occlusion Visualization. The visualization highlights the intricate shadowing and occlusion details captured by our GS-IR method, emphasizing the performance of our approach in modeling indirect illumination. + +![](images/af86a9520be3f710f0f85515276be486e606e7e8b99cffe40415e6fd78b857c7.jpg) + +![](images/4bb38e441cdac8b70d210c4ae655298395ad7c438da73ef80c52740d1d31b4f0.jpg) + +Analysis on the Normal Regularization Reliable normal estimation is critical for conducting inverse rendering. To this end, we present depth-derivation-based regularization to facilitate 3D Gaussian-based normal estimation as stated in Sec. 4.1. In this section, we explore the impact of different acquisition schemes on the final normal quality and inverse rendering results. The quantitative results shown in Tab. 3 demonstrate that using 3D Gaussians as a normal proxy and adopting the linear interpolation strategy significantly improves the normal estimation and inverse rendering results. In addition, Fig. 7 qualitatively shows that conducting volumetric accumulation results in the floating problem (cf. Fig. 7 (b)). Despite peak selection overcomes this problem, it introduces disc aliasing (cf. Fig. 7 (c)). Compared with them, the linear interpolation adopted in GS-IR robustly produces accurate depth (cf. Fig. 7 (d)). + +Analysis on the Indirect Illumination To demonstrate the effectiveness of our indirect illumination model, we compare our method with two variants: a model without occlusion volume (w/o occlusion) and a model without indirect illumination (w/o indirect illum.). The quantitative comparisons in Tab. 4 indicate that each component is crucial for estimating accurate material decomposition and generating photorealistic rendering results. Additionally, Fig. 9 showcases the ambient occlusion visualization in both synthetic and real scenes. The visualization highlights the intra- + +cate shadowing and occlusion details captured by our GS-IR method, emphasizing the performance of our approach in modeling indirect illumination. This analysis further supports the effectiveness of using occlusion volume and introducing indirect illumination in enhancing the decomposition capabilities of our GS-IR. + +# 5.3. Application + +We perform relighting experiments using the recovered geometry, material, and illumination from our GS-IR method. We test GS-IR under different lighting conditions and directions, observing how the reconstructed scene responds to the changes in lighting. The results of these experiments demonstrate that our GS-IR method can effectively handle relighting applications, producing photorealistic renderings under various lighting conditions. More results can be found in the supplement. + +# 6. Conclusion + +We present GS-IR, a novel inverse rendering approach based on 3D Gaussian Splatting (3DGS), which employs forward mapping volume rendering to achieve photorealistic novel view synthesis and relighting results. Our GS-IR proposes an optimization scheme with depth-derivation-based regularization for normal estimation and a baking-based occlusion to model indirect lighting. These components are eventually employed to decompose material and illumination. Our extensive experiments demonstrate the effectiveness of GS-IR in achieving state-of-the-art inverse rendering results, surpassing previous neural methods in terms of both reconstruction quality and efficiency. + +Limitation Spherical Harmonics (SH) is only suitable for representing low-frequency, and we only use the occlusion represented by SH to model the diffuse term of indirect illumination. Modeling the specular term of indirect illumination remains a limitation of GS-IR, and has been a challenging problem in computer graphics. We believe it would be valuable to address this limitation in future work and suggest screen space global illumination (SSGI) techniques. + +# References + +[1] Unreal engine. https://www.unrealengine.com/. 5 +[2] Unity. https://unity.com/.5 +[3] Blender. https://www.blender.org/.5 +[4] Tomas Akenine-Mo, Eric Haines, Nate Hoffman, et al. Real-time rendering. 2018. 2 +[5] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 2, 6, 7 +[6] Sai Bi, Zexiang Xu, Pratul Srinivasan, Ben Mildenhall, Kalyan Sunkavalli, Miloš Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Neural reflectance fields for appearance acquisition. arXiv preprint arXiv:2008.03824, 2020. 2 +[7] Sai Bi, Zexiang Xu, Kalyan Sunkavalli, Miloš Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Deep reflectance volumes: Relightable reconstructions from multi-view photometric images. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part III 16, pages 294–311. Springer, 2020. 2 +[8] Sai Bi, Zexiang Xu, Kalyan Sunkavalli, David Kriegman, and Ravi Ramamoorthi. Deep 3d capture: Geometry and reflectance from sparse multi-view images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5960-5969, 2020. 2 +[9] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T Barron, Ce Liu, and Hendrik Lensch. Nerd: Neural reflectance decomposition from image collections. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12684-12694, 2021. 1, 2 +[10] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan Barron, and Hendrik Lensch. Neural-pil: Neural pre-integrated lighting for reflectance decomposition. Advances in Neural Information Processing Systems, 34: 10691-10704, 2021. 1, 2 +[11] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 130-141, 2023. 2 +[12] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5799-5809, 2021. +[13] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 2 +[14] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European + +Conference on Computer Vision, pages 333-350. Springer, 2022. 2 +[15] Robert L Cook and Kenneth E. Torrance. A reflectance model for computer graphics. ACM Transactions on Graphics (ToG), 1(1):7-24, 1982. 3 +[16] Yue Dong, Guojun Chen, Pieter Peers, Jiawan Zhang, and Xin Tong. Appearance-from-motion: Recovering spatially varying surface reflectance under unknown lighting. ACM Transactions on Graphics (TOG), 33(6):1-12, 2014. 2 +[17] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 2, 7 +[18] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbek Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12479–12488, 2023. 2 +[19] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, light, and material decomposition from images using monte carlo rendering and denoising. Advances in Neural Information Processing Systems, 35:22856-22869, 2022. 2 +[20] Peter Hedman, Pratul P Srinivasan, Ben Mildenhall, Jonathan T Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5875-5884, 2021. 2 +[21] Zhangjin Huang, Zhihao Liang, Haojie Zhang, Yangkai Lin, and Kui Jia. Sur2f: A hybrid representation for high-quality and efficient surface reconstruction from multi-view images. arXiv preprint arXiv:2401.03704, 2024. 2 +[22] Haian Jin, Isabella Liu, Peijia Xu, Xiaoshuai Zhang, Songfang Han, Sai Bi, Xiaowei Zhou, Zexiang Xu, and Hao Su. Tensoir: Tensorial inverse rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 165-174, 2023. 2, 4, 6, 7 +[23] Brian Karis and Epic Games. Real shading in unreal engine 4. Proc. Physically Based Shading Theory Practice, 4(3):1, 2013. 5 +[24] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 42(4):1-14, 2023. 1, 2, 3, 4, 7 +[25] Leonid Keselman and Martial Hebert. Approximate differentiable rendering with algebraic surfaces. In European Conference on Computer Vision, pages 596-614. Springer, 2022. 2 +[26] Leonid Keselman and Martial Hebert. Flexible techniques for differentiable rendering with 3d gaussians. arXiv preprint arXiv:2308.14737, 2023. 2 +[27] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021. 2 + +[28] Zhihao Liang, Zhangjin Huang, Changxing Ding, and Kui Jia. Helixsurf: A robust and efficient neural implicit surface learning of indoor scenes with iterative intertwined regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13165-13174, 2023. 2 +[29] Fujun Luan, Shuang Zhao, Kavita Bala, and Zhao Dong. Unified shape and svbrdf recovery using differentiable monte carlo rendering. In Computer Graphics Forum, pages 101-113. Wiley Online Library, 2021. 2 +[30] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision, pages 405-421, 2020. 1, 2 +[31] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 2, 7 +[32] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3d models, materials, and lighting from images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8280-8290, 2022. 7 +[33] Giljoo Nam, Joo Ho Lee, Diego Gutierrez, and Min H Kim. Practical svbrdf acquisition of 3d objects with unstructured flash photography. ACM Transactions on Graphics (TOG), 37(6):1-12, 2018. 2 +[34] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 2 +[35] Pratul P Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T Barron. Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7495-7504, 2021. 1, 2 +[36] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 2 +[37] Bruce Walter, Stephen R Marschner, Hongsong Li, and Kenneth E Torrance. Microfacet models for refraction through rough surfaces. In Proceedings of the 18th Eurographics conference on Rendering Techniques, pages 195-206, 2007. 3 +[38] Rui Xia, Yue Dong, Pieter Peers, and Xin Tong. Recovering shape and spatially-varying surface reflectance under unknown illumination. ACM Transactions on Graphics (TOG), 35(6):1-12, 2016. 2 +[39] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the + +IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5438-5448, 2022. 2 +[40] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems, 33:2492-2502, 2020. 2 +[41] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 7 +[42] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snavely. Physg: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5453-5462, 2021. 2 +[43] Kai Zhang, Nick Kolkin, Sai Bi, Fujun Luan, Zexiang Xu, Eli Shechtman, and Noah Snavely. Arf: Artistic radiance fields. In European Conference on Computer Vision, pages 717-733. Springer, 2022. 2 +[44] Kai Zhang, Fujun Luan, Zhengqi Li, and Noah Snavely. Iron: Inverse rendering by optimizing neural sdfs and materials from photometric images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5565-5574, 2022. 2 +[45] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6 +[46] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (ToG), 40(6):1-18, 2021. 1, 2, 6, 7 +[47] Yuanqing Zhang, Jiaming Sun, Xingyi He, Huan Fu, Rongfei Jia, and Xiaowei Zhou. Modeling indirect illumination for inverse rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18643-18652, 2022. 2, 7 \ No newline at end of file diff --git a/2024/GS-IR_ 3D Gaussian Splatting for Inverse Rendering/images.zip b/2024/GS-IR_ 3D Gaussian Splatting for Inverse Rendering/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..eadad44ba3da745e6c00829407ccb76184f5d3e7 --- /dev/null +++ b/2024/GS-IR_ 3D Gaussian Splatting for Inverse Rendering/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2c381ff83b57ba0c2811ae9cb026d8ffc15bf8ecea0ee67145b932ce7a87495d +size 732691 diff --git a/2024/GS-IR_ 3D Gaussian Splatting for Inverse Rendering/layout.json b/2024/GS-IR_ 3D Gaussian Splatting for Inverse Rendering/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..25a9a60a43343e7639af30742fc5a06a351dce8b --- /dev/null +++ b/2024/GS-IR_ 3D Gaussian Splatting for Inverse Rendering/layout.json @@ -0,0 +1,9494 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 133, + 103, + 462, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 103, + 462, + 121 + ], + "spans": [ + { + "bbox": [ + 133, + 103, + 462, + 121 + ], + "type": "text", + "content": "GS-IR: 3D Gaussian Splating for Inverse Rendering" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 139, + 142, + 452, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 142, + 452, + 157 + ], + "spans": [ + { + "bbox": [ + 139, + 142, + 452, + 157 + ], + "type": "text", + "content": "Zhihao Liang" + }, + { + "bbox": [ + 139, + 142, + 452, + 157 + ], + "type": "inline_equation", + "content": "^{1,\\ast}" + }, + { + "bbox": [ + 139, + 142, + 452, + 157 + ], + "type": "text", + "content": ", Qi Zhang" + }, + { + "bbox": [ + 139, + 142, + 452, + 157 + ], + "type": "inline_equation", + "content": "^{2,\\ast}" + }, + { + "bbox": [ + 139, + 142, + 452, + 157 + ], + "type": "text", + "content": ", Ying Feng" + }, + { + "bbox": [ + 139, + 142, + 452, + 157 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 139, + 142, + 452, + 157 + ], + "type": "text", + "content": ", Ying Shan" + }, + { + "bbox": [ + 139, + 142, + 452, + 157 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 139, + 142, + 452, + 157 + ], + "type": "text", + "content": ", Kui Jia" + }, + { + "bbox": [ + 139, + 142, + 452, + 157 + ], + "type": "inline_equation", + "content": "^{3,\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 150, + 157, + 442, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 157, + 442, + 171 + ], + "spans": [ + { + "bbox": [ + 150, + 157, + 442, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 150, + 157, + 442, + 171 + ], + "type": "text", + "content": "South China University of Technology, " + }, + { + "bbox": [ + 150, + 157, + 442, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 150, + 157, + 442, + 171 + ], + "type": "text", + "content": "Tencent AI Lab," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 114, + 171, + 479, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 171, + 479, + 186 + ], + "spans": [ + { + "bbox": [ + 114, + 171, + 479, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 114, + 171, + 479, + 186 + ], + "type": "text", + "content": " School of Data Science, The Chinese University of Hong Kong, Shenzhen" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 150, + 188, + 441, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 188, + 441, + 199 + ], + "spans": [ + { + "bbox": [ + 150, + 188, + 441, + 199 + ], + "type": "text", + "content": "eezhihaoliang@mail.scut.edu.cn, nwpuqzhang@gmail.com," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 124, + 202, + 465, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 202, + 465, + 213 + ], + "spans": [ + { + "bbox": [ + 124, + 202, + 465, + 213 + ], + "type": "text", + "content": "vonyfeng@gmail.com, yingsshan@tencent.com, kuijia@cuhk.edu.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 264, + 192, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 264, + 192, + 277 + ], + "spans": [ + { + "bbox": [ + 143, + 264, + 192, + 277 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 290, + 290, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 290, + 290, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 290, + 290, + 613 + ], + "type": "text", + "content": "We propose GS-IR, a novel inverse rendering approach based on 3D Gaussian Splatting (3DGS) that leverages forward mapping volume rendering to achieve photorealistic novel view synthesis and relighting results. Unlike previous works that use implicit neural representations and volume rendering (e.g. NeRF), which suffer from low expressive power and high computational complexity, we extend 3DGS, a top-performance representation for novel view synthesis, to estimate scene geometry, surface material, and environment illumination from multi-view images captured under unknown lighting conditions. There are two main problems when introducing 3DGS to inverse rendering: 1) 3DGS does not support producing plausible normal natively; 2) forward mapping (e.g. rasterization and splatting) cannot trace the occlusion like backward mapping (e.g. ray tracing). To address these challenges, our GS-IR proposes an efficient optimization scheme incorporating a depth-derivation-based regularization for normal estimation and a baking-based occlusion to model indirect lighting. The flexible and expressive 3DGS representation allows us to achieve fast and compact geometry reconstruction, photorealistic novel view synthesis, and effective physically-based rendering. We demonstrate the superiority of our method over baseline methods through qualitative and quantitative evaluations of various challenging scenes. The source code is available at https://github.com/lzhnb/GS-IR." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 618, + 128, + 630 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 618, + 128, + 630 + ], + "spans": [ + { + "bbox": [ + 47, + 618, + 128, + 630 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 638, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 638, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 638, + 287, + 687 + ], + "type": "text", + "content": "Inverse rendering is a long-standing task, seeking to answer the question: \"How can we deduce physical attributes (e.g. geometry, material, and lighting) of a 3D scene from multiview images?\" This problem is inherently challenging and" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 309, + 240, + 545, + 336 + ], + "blocks": [ + { + "bbox": [ + 309, + 240, + 545, + 336 + ], + "lines": [ + { + "bbox": [ + 309, + 240, + 545, + 336 + ], + "spans": [ + { + "bbox": [ + 309, + 240, + 545, + 336 + ], + "type": "image", + "image_path": "f7d7891f2e1c9fb294c30870f7ea42156d5f26935de4c8f00b8fde40d7d9910a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 310, + 338, + 459, + 386 + ], + "blocks": [ + { + "bbox": [ + 310, + 338, + 459, + 386 + ], + "lines": [ + { + "bbox": [ + 310, + 338, + 459, + 386 + ], + "spans": [ + { + "bbox": [ + 310, + 338, + 459, + 386 + ], + "type": "image", + "image_path": "85913a2826953e4b36701c8b55b70f16eadf47508bed410b937ef3e389a60a27.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 396, + 547, + 485 + ], + "lines": [ + { + "bbox": [ + 304, + 396, + 547, + 485 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 547, + 485 + ], + "type": "text", + "content": "Figure 1. Given multi-view captured images of a complex scene, we propose GS-IR (3D Gaussian Splatting for Inverse Rendering), which utilizes 3D Gaussian and forward mapping splatting to recover high-quality physical properties (e.g., normal, material, illumination). This enables us to perform relighting and material editing, resulting in outstanding inverse rendering results. Better viewed on screen with zoom in, especially the remarkable material decomposition and normal reconstruction of bicycle axle." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 462, + 338, + 544, + 386 + ], + "blocks": [ + { + "bbox": [ + 462, + 338, + 544, + 386 + ], + "lines": [ + { + "bbox": [ + 462, + 338, + 544, + 386 + ], + "spans": [ + { + "bbox": [ + 462, + 338, + 544, + 386 + ], + "type": "image", + "image_path": "bb1d0adc675c004c79921e17e488466f8e935f6e5053f2e95f2521b0cd04bb26.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 495, + 547, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 547, + 614 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 547, + 614 + ], + "type": "text", + "content": "ill-posed, particularly when input images are captured in uncontrolled environments with unknown illumination. Recent research [9, 10, 35, 46] has sought to address this issue by employing implicit neural representations akin to NeRF [30] that utilizes multi-layer perceptrons (MLPs). However, current methods incorporating MLP face challenges in terms of their low expressive capacity and high computational demands, which significantly limits the effectiveness and efficiency of inverse rendering, especially when it cannot be rendered at interactive rates." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "3D Gaussian Splatting (3DGS) [24] has recently emerged as a promising technique to model 3D static scenes and significantly boost the rendering speed to a real-time level. It makes the scene representation more compact and achieves fast and top performance for novel view synthesis. Introducing it to the inverse rendering pipeline is natural and essential, including geometry reconstruction, materials decomposition, and illumination estimation. Unlike ray trac" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 693, + 161, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 693, + 161, + 703 + ], + "spans": [ + { + "bbox": [ + 61, + 693, + 161, + 703 + ], + "type": "text", + "content": "* indicates equal contribution." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 61, + 703, + 234, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 703, + 234, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 703, + 234, + 713 + ], + "type": "inline_equation", + "content": "^{\\dagger}" + }, + { + "bbox": [ + 61, + 703, + 234, + 713 + ], + "type": "text", + "content": " Correspondence to Kui Jia ." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21644" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": "ing in NeRF, 3DGS produces a set of 3D Gaussians around sparse points. During the 3DGS optimization, the adaptive control of the Gaussian density may lead to loose geometry, making it difficult to estimate accurate scene's normal. Consequently, it is necessary to introduce a well-designed strategy to regularize 3DGS's normal estimation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 152, + 288, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 152, + 288, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 152, + 288, + 319 + ], + "type": "text", + "content": "Our goal is to use 3D Gaussians as the scene representation for inverse rendering from multi-view images captured under unknown lighting conditions. However, capturing observations under natural illumination often shows complex effects such as soft shadows and interreflections. TensoIR [22] leverages the ray tracing of NeRF to directly model occlusion and indirect illumination. In contrast, 3DGS replaces the ray tracing in the NeRF with differentiable forward mapping volume rendering, which directly projects 3D Gaussians onto the 2D plane. This strategy improves the rendering efficiency but makes it difficult to calculate occlusion. Inspired by the \"Indirect Lighting Cache\" used in real-time rendering [4], we attempt to bake the occlusion into volumes for caching." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 327, + 287, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 287, + 553 + ], + "type": "text", + "content": "In this paper, we present a novel 3D Gaussian-based inverse rendering framework called GS-IR (3D Gaussian Splatting for Inverse Rendering) that leverages forward mapping splatting to deduce the physical attributes of a complex scene. To the best of our knowledge, our method is the first work to introduce the 3DGS technique for inverse rendering, which can simultaneously estimate scene geometry, materials, and illumination from multi-view images. Our GS-IR addresses two main issues when using 3DGS for inverse rendering. Firstly, we develop an intuitive and well-designed regularization to estimate the scene's normal. Secondly, we use a baking-based method embedded in GS-IR to cache occlusions, obtaining an efficient indirect illumination model. As shown in Fig. 1, our approach can reconstruct high-fidelity geometry and materials of a complex real scene under unknown natural illumination, enabling state-of-the-art rendering of novel view synthesis and additional applications like relighting. Our technical contributions are summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 562, + 286, + 657 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 562, + 286, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 562, + 286, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 562, + 286, + 597 + ], + "type": "text", + "content": "- We present GS-IR that models a scene as a set of 3D Gaussians to achieve physically-based rendering and state-of-the-art decomposition results for both objects and scenes;" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 598, + 286, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 598, + 286, + 632 + ], + "spans": [ + { + "bbox": [ + 47, + 598, + 286, + 632 + ], + "type": "text", + "content": "- We propose an efficient optimization scheme with regularization to concentrate depth gradient around 3DGS and produce reliable normals for GS-IR;" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 633, + 286, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 633, + 286, + 657 + ], + "spans": [ + { + "bbox": [ + 47, + 633, + 286, + 657 + ], + "type": "text", + "content": "- We develop a baking-based method embedded in GS-IR to handle the occlusion in modeling indirect lighting;" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "content": "We demonstrate the superiority of our method to baseline methods qualitatively and quantitatively on various challenging scenes, including the TensoIR synthesis dataset [22] and Mip-NeRF 360 real dataset [5]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 71, + 397, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 397, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 397, + 83 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 91, + 545, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 545, + 365 + ], + "type": "text", + "content": "Neural Representation Recently, neural rendering techniques, exemplified by Neural Radiance Field (NeRF) [30], have achieved impressive success in addressing visual computing problems, giving rise to numerous neural representations [14, 21, 31, 36, 39, 40] tailored for different tasks [11-13, 18, 27, 28, 34, 43]. The vanilla NeRF models a continuous radiance field implicitly in MLPs, which requires massive repeated queries during training and inference. To address the computational inefficiencies, many neural scene representations are proposed with more discretized geometry proxies such as voxel grids [17, 20, 36], hash grids [31], tri-planes [14] or points [39]. Neural features are stored in a structured manner, allowing for efficient storage and retrieval. The computational cost can thus be significantly reduced by introducing interpolation techniques, however, with an inevitable loss in image quality. 3D Gaussians are introduced as an unstructured scene representation to strike a balance between efficiency and quality [24]. With the specially designed tile-based rasterizer for Gaussian splats, this method achieves real-time rendering with high quality for novel-view synthesis. In this work, 3D Gaussian representation is combined with the physical-based rendering (PBR) model for inverse rendering." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 367, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 367, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 367, + 545, + 713 + ], + "type": "text", + "content": "Inverse Rendering Inverse rendering aims to decompose the image's appearance into the geometry, material, and lighting conditions. Considering the inherent ambiguity between observed images and underlying properties, many methods are proposed with different constrained settings, such as capturing images with fixed lighting and rotating object [16, 38], capturing with moving camera and colocated lighting [7, 8, 29, 33]. Combined with neural representations, inverse rendering models the scene simulating how the light interacts with the neural volume with various material properties, and estimates the lighting and material parameters during optimization [6, 9, 10, 19, 22, 35, 42, 44, 46, 47]. Neural Reflectance Fields [6] assumes a known point light source and represents the scene as a field of volume density, surface normals, and bi-directional reflectance distribution functions(BRDFs) with one bounce direct illumination. NeRV [35] and InvRender [47] extend to arbitrary known lighting conditions and train an additional MLP to model the light visibility. PhySG [42] assumes full light source visibility without shadow simulation, and represents the lighting and scene BRDFs with spherical Gaussians for acceleration. TensoIR [22] adopts the efficient TensoRF [14] representation which enables the computation of visibility and indirect lighting by raytracing, while limited to object-level. For modeling surface geometry using point clouds, Fuzzy Metaballs (FMs) [25, 26] offers a great way to render depth from 3D Gaussian using Order Independent Transparency (OIT) and approximate intersection. However, it requires silhouettes as input and struggles to handle" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21645" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "intricate geometry (e.g. Lego and Ficus) let alone complex scenes. In this work, we propose a 3DGS-based pipeline to recover the geometry, material, and lighting that is available for both objects and unbounded scenes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 129, + 124, + 143 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 129, + 124, + 143 + ], + "spans": [ + { + "bbox": [ + 47, + 129, + 124, + 143 + ], + "type": "text", + "content": "3. Preliminary" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 148, + 287, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 148, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 46, + 148, + 287, + 184 + ], + "type": "text", + "content": "In this section, we give the technical backgrounds and math symbols that are necessary for the presentation of our proposed method in subsequent sections." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 185, + 287, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 185, + 287, + 232 + ], + "spans": [ + { + "bbox": [ + 46, + 185, + 287, + 232 + ], + "type": "text", + "content": "3D Gaussian Splatting (3DGS) [24] is an explicit 3D scene representation in the form of point clouds. Each point is represented as a Gaussian function " + }, + { + "bbox": [ + 46, + 185, + 287, + 232 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 185, + 287, + 232 + ], + "type": "text", + "content": " that approximates the shape of a bell curve, which is defined as," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 98, + 236, + 287, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 236, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 98, + 236, + 287, + 251 + ], + "type": "interline_equation", + "content": "g (\\boldsymbol {x} | \\boldsymbol {\\mu}, \\boldsymbol {\\Sigma}) = e ^ {- \\frac {1}{2} (\\boldsymbol {x} - \\boldsymbol {\\mu}) ^ {T} \\boldsymbol {\\Sigma} ^ {- 1} (\\boldsymbol {x} - \\boldsymbol {\\mu})}, \\tag {1}", + "image_path": "7449caef61da805925b804db5bda4d7e125e39d18b8877e41befd310f7f0f8f3.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 251, + 287, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 287, + 372 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 287, + 372 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 251, + 287, + 372 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 251, + 287, + 372 + ], + "type": "text", + "content": " is its mean vector, and " + }, + { + "bbox": [ + 46, + 251, + 287, + 372 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma} \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 46, + 251, + 287, + 372 + ], + "type": "text", + "content": " is an anisotropic covariance matrix. The mean vector " + }, + { + "bbox": [ + 46, + 251, + 287, + 372 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}" + }, + { + "bbox": [ + 46, + 251, + 287, + 372 + ], + "type": "text", + "content": " of a 3D Gaussian is parameterized as " + }, + { + "bbox": [ + 46, + 251, + 287, + 372 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu} = (\\mu_x, \\mu_y, \\mu_z)" + }, + { + "bbox": [ + 46, + 251, + 287, + 372 + ], + "type": "text", + "content": ", and the covariance matrix " + }, + { + "bbox": [ + 46, + 251, + 287, + 372 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}" + }, + { + "bbox": [ + 46, + 251, + 287, + 372 + ], + "type": "text", + "content": " is factorized into a scaling matrix " + }, + { + "bbox": [ + 46, + 251, + 287, + 372 + ], + "type": "inline_equation", + "content": "\\pmb{S}" + }, + { + "bbox": [ + 46, + 251, + 287, + 372 + ], + "type": "text", + "content": " and a rotation matrix " + }, + { + "bbox": [ + 46, + 251, + 287, + 372 + ], + "type": "inline_equation", + "content": "\\pmb{R}" + }, + { + "bbox": [ + 46, + 251, + 287, + 372 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 251, + 287, + 372 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma} = R\\pmb{S}\\pmb{S}^\\top\\pmb{R}^\\top" + }, + { + "bbox": [ + 46, + 251, + 287, + 372 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 251, + 287, + 372 + ], + "type": "inline_equation", + "content": "\\pmb{S}" + }, + { + "bbox": [ + 46, + 251, + 287, + 372 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 251, + 287, + 372 + ], + "type": "inline_equation", + "content": "\\pmb{R}" + }, + { + "bbox": [ + 46, + 251, + 287, + 372 + ], + "type": "text", + "content": " refer to a diagonal matrix " + }, + { + "bbox": [ + 46, + 251, + 287, + 372 + ], + "type": "inline_equation", + "content": "\\mathrm{diag}(s_x, s_y, s_z)" + }, + { + "bbox": [ + 46, + 251, + 287, + 372 + ], + "type": "text", + "content": " and a rotation matrix constructed from a unit quaternion " + }, + { + "bbox": [ + 46, + 251, + 287, + 372 + ], + "type": "inline_equation", + "content": "\\pmb{q}" + }, + { + "bbox": [ + 46, + 251, + 287, + 372 + ], + "type": "text", + "content": ". Given a viewing transformation with extrinsic matrix " + }, + { + "bbox": [ + 46, + 251, + 287, + 372 + ], + "type": "inline_equation", + "content": "\\pmb{T}" + }, + { + "bbox": [ + 46, + 251, + 287, + 372 + ], + "type": "text", + "content": " and intrinsic matrix " + }, + { + "bbox": [ + 46, + 251, + 287, + 372 + ], + "type": "inline_equation", + "content": "\\pmb{K}" + }, + { + "bbox": [ + 46, + 251, + 287, + 372 + ], + "type": "text", + "content": ", the mean vector " + }, + { + "bbox": [ + 46, + 251, + 287, + 372 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}" + }, + { + "bbox": [ + 46, + 251, + 287, + 372 + ], + "type": "text", + "content": " and covariance matrix " + }, + { + "bbox": [ + 46, + 251, + 287, + 372 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}'" + }, + { + "bbox": [ + 46, + 251, + 287, + 372 + ], + "type": "text", + "content": " from the 3D point " + }, + { + "bbox": [ + 46, + 251, + 287, + 372 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 46, + 251, + 287, + 372 + ], + "type": "text", + "content": " to 2D pixel " + }, + { + "bbox": [ + 46, + 251, + 287, + 372 + ], + "type": "inline_equation", + "content": "\\pmb{u}" + }, + { + "bbox": [ + 46, + 251, + 287, + 372 + ], + "type": "text", + "content": " is defined as," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 91, + 374, + 287, + 388 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 374, + 287, + 388 + ], + "spans": [ + { + "bbox": [ + 91, + 374, + 287, + 388 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\mu} ^ {\\prime} = \\boldsymbol {K} \\boldsymbol {T} [ \\boldsymbol {\\mu}, 1 ] ^ {\\top}, \\boldsymbol {\\Sigma} ^ {\\prime} = \\boldsymbol {J} \\boldsymbol {T} \\boldsymbol {\\Sigma} \\boldsymbol {T} ^ {\\top} \\boldsymbol {J} ^ {\\top}, \\tag {2}", + "image_path": "453a6a0e13b1d75508f44313b812a78ae28b1286c90d32ea170e21f559486ecb.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 389, + 287, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 389, + 287, + 461 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 287, + 461 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 389, + 287, + 461 + ], + "type": "inline_equation", + "content": "\\pmb{J}" + }, + { + "bbox": [ + 46, + 389, + 287, + 461 + ], + "type": "text", + "content": " is the Jacobian matrix of the affine approximation of the perspective projection. Besides, each Gaussian represents the view-dependent color " + }, + { + "bbox": [ + 46, + 389, + 287, + 461 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 46, + 389, + 287, + 461 + ], + "type": "text", + "content": " via a set of coefficients of spherical harmonics (SH), which is then multiplied by opacity " + }, + { + "bbox": [ + 46, + 389, + 287, + 461 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 389, + 287, + 461 + ], + "type": "text", + "content": " for volume rendering. We finally obtain the color " + }, + { + "bbox": [ + 46, + 389, + 287, + 461 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{C}}" + }, + { + "bbox": [ + 46, + 389, + 287, + 461 + ], + "type": "text", + "content": " at pixel " + }, + { + "bbox": [ + 46, + 389, + 287, + 461 + ], + "type": "inline_equation", + "content": "\\pmb{u}" + }, + { + "bbox": [ + 46, + 389, + 287, + 461 + ], + "type": "text", + "content": " based on Eq. (1) and Eq. (2)," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 465, + 287, + 503 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 465, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 47, + 465, + 287, + 503 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {C}} = \\sum_ {i \\in N} T _ {i} g _ {i} \\left(\\boldsymbol {u} \\mid \\boldsymbol {\\mu} ^ {\\prime}, \\boldsymbol {\\Sigma} ^ {\\prime}\\right) \\alpha_ {i} \\boldsymbol {c} _ {i}, \\quad T _ {i} = \\prod_ {j = 1} ^ {i - 1} \\left(1 - g _ {j} \\left(\\boldsymbol {u} \\mid \\boldsymbol {\\mu} ^ {\\prime}, \\boldsymbol {\\Sigma} ^ {\\prime}\\right) \\alpha_ {j}\\right), \\tag {3}", + "image_path": "1c268dac3147d392f1948b2fda7e46e43846628d51d757bdb912188b0202862d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 502, + 286, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 502, + 286, + 525 + ], + "spans": [ + { + "bbox": [ + 46, + 502, + 286, + 525 + ], + "type": "text", + "content": "where accumulated transmittance " + }, + { + "bbox": [ + 46, + 502, + 286, + 525 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 46, + 502, + 286, + 525 + ], + "type": "text", + "content": " quantifies the probability density of " + }, + { + "bbox": [ + 46, + 502, + 286, + 525 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 502, + 286, + 525 + ], + "type": "text", + "content": "-th Gaussian at pixel " + }, + { + "bbox": [ + 46, + 502, + 286, + 525 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 46, + 502, + 286, + 525 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 525, + 287, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 525, + 287, + 561 + ], + "spans": [ + { + "bbox": [ + 46, + 525, + 287, + 561 + ], + "type": "text", + "content": "The Rendering Equation In GS-IR, we leverage the classic rendering equation to formulate the outgoing radiance of a surface point " + }, + { + "bbox": [ + 46, + 525, + 287, + 561 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 525, + 287, + 561 + ], + "type": "text", + "content": " with normal " + }, + { + "bbox": [ + 46, + 525, + 287, + 561 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 46, + 525, + 287, + 561 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 89, + 565, + 287, + 588 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 565, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 89, + 565, + 287, + 588 + ], + "type": "interline_equation", + "content": "L _ {o} (\\boldsymbol {x}, \\boldsymbol {v}) = \\int_ {\\Omega} L _ {i} (\\boldsymbol {x}, \\boldsymbol {l}) f _ {r} (\\boldsymbol {l}, \\boldsymbol {v}) (\\boldsymbol {l} \\cdot \\boldsymbol {n}) d \\boldsymbol {l}, \\tag {4}", + "image_path": "e61228136ad523ef441e148b78461724a896eaa573b8ee9c6be37dae79795e5e.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 590, + 287, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 590, + 287, + 674 + ], + "spans": [ + { + "bbox": [ + 46, + 590, + 287, + 674 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 46, + 590, + 287, + 674 + ], + "type": "text", + "content": " denotes the upper hemisphere centered at " + }, + { + "bbox": [ + 46, + 590, + 287, + 674 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 46, + 590, + 287, + 674 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 590, + 287, + 674 + ], + "type": "inline_equation", + "content": "\\pmb{l}" + }, + { + "bbox": [ + 46, + 590, + 287, + 674 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 590, + 287, + 674 + ], + "type": "inline_equation", + "content": "\\pmb{v}" + }, + { + "bbox": [ + 46, + 590, + 287, + 674 + ], + "type": "text", + "content": " denote incident and view directions respectively. " + }, + { + "bbox": [ + 46, + 590, + 287, + 674 + ], + "type": "inline_equation", + "content": "L_{i}(\\pmb{x},\\pmb{l})" + }, + { + "bbox": [ + 46, + 590, + 287, + 674 + ], + "type": "text", + "content": " denotes the radiance received at " + }, + { + "bbox": [ + 46, + 590, + 287, + 674 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 46, + 590, + 287, + 674 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 590, + 287, + 674 + ], + "type": "inline_equation", + "content": "\\pmb{l}" + }, + { + "bbox": [ + 46, + 590, + 287, + 674 + ], + "type": "text", + "content": ". Notably, we follow Cook-Torrance microfacet model [15, 37] and formulate the bidirectional reflectance distribution function (BRDF) " + }, + { + "bbox": [ + 46, + 590, + 287, + 674 + ], + "type": "inline_equation", + "content": "f_{r}" + }, + { + "bbox": [ + 46, + 590, + 287, + 674 + ], + "type": "text", + "content": " as a function of albedo " + }, + { + "bbox": [ + 46, + 590, + 287, + 674 + ], + "type": "inline_equation", + "content": "\\pmb{a} \\in [0,1]^3" + }, + { + "bbox": [ + 46, + 590, + 287, + 674 + ], + "type": "text", + "content": ", metallic " + }, + { + "bbox": [ + 46, + 590, + 287, + 674 + ], + "type": "inline_equation", + "content": "m \\in [0,1]" + }, + { + "bbox": [ + 46, + 590, + 287, + 674 + ], + "type": "text", + "content": ", and roughness " + }, + { + "bbox": [ + 46, + 590, + 287, + 674 + ], + "type": "inline_equation", + "content": "\\rho \\in [0,1]" + }, + { + "bbox": [ + 46, + 590, + 287, + 674 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 89, + 677, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 89, + 677, + 287, + 715 + ], + "type": "interline_equation", + "content": "f _ {r} (\\boldsymbol {l}, \\boldsymbol {v}) = \\underbrace {(1 - m) \\frac {\\boldsymbol {a}}{\\pi}} _ {\\text {d i f f u s e}} + \\underbrace {\\frac {D F G}{4 (\\boldsymbol {n} \\cdot \\boldsymbol {l}) (\\boldsymbol {n} \\cdot \\boldsymbol {v})}} _ {\\text {s p e c u l a r}}, \\tag {5}", + "image_path": "20402fcbcf7302be11aa1fb8bc232c8d775004b04cd87e0ed0387148a9182e1b.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": "where microfacet distribution function " + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": ", Fresnel reflection " + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": ", and geometric shadowing factor " + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": " are related to the surface roughness " + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": ". We use 3D Gaussians to store these material properties in GS-IR." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 130, + 361, + 143 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 130, + 361, + 143 + ], + "spans": [ + { + "bbox": [ + 305, + 130, + 361, + 143 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 148, + 545, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 148, + 545, + 390 + ], + "spans": [ + { + "bbox": [ + 304, + 148, + 545, + 390 + ], + "type": "text", + "content": "Given a set of calibrated RGB images " + }, + { + "bbox": [ + 304, + 148, + 545, + 390 + ], + "type": "inline_equation", + "content": "\\{I_m\\}_{m=1}^M" + }, + { + "bbox": [ + 304, + 148, + 545, + 390 + ], + "type": "text", + "content": " of a target scene captured from multiple views under static, yet unknown illumination, inverse rendering aims to decompose the scene's intrinsic properties, including normal, materials, and illumination. This decomposition facilitates the recovery and subsequent edition of the target scene. Motivated by the remarkable performance in quality and speed of 3DGS [24], we present a novel framework GS-IR consisting of three well-designed stage strategies, as shown in Fig. 2. In the initial stage, we leverage differentiable splating to optimize 3D Gaussians. Concurrently, we utilize the gradient derived from the rendered depth map to supervise the normal stored in 3D Gaussians (cf. Sec. 4.1). In the second stage, we precompute the occlusion based on the learned geometric information (i.e. depth and normal) and store it in an efficient spherical harmonics-based architecture to model indirect illumination (cf. Sec. 4.2). In the final stage, we combine a differentiable splating with the physical-based rendering (PBR) pipeline to optimize illumination and material-aware 3D Gaussians (cf. Sec. 4.3)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 396, + 438, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 396, + 438, + 407 + ], + "spans": [ + { + "bbox": [ + 305, + 396, + 438, + 407 + ], + "type": "text", + "content": "4.1. Normal Reconstruction" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 415, + 545, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 545, + 545 + ], + "type": "text", + "content": "During the initial stage, we optimize 3D Gaussians for geometry reconstruction from observed images, denoted as " + }, + { + "bbox": [ + 304, + 415, + 545, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 304, + 415, + 545, + 545 + ], + "type": "text", + "content": ". The optimized " + }, + { + "bbox": [ + 304, + 415, + 545, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 304, + 415, + 545, + 545 + ], + "type": "text", + "content": " functions as a geometric proxy for surface points and their corresponding normals " + }, + { + "bbox": [ + 304, + 415, + 545, + 545 + ], + "type": "inline_equation", + "content": "\\pmb{n}" + }, + { + "bbox": [ + 304, + 415, + 545, + 545 + ], + "type": "text", + "content": ", which are crucial for successful inverse rendering. As highlighted in Sec. 1, generating reasonable normals within the 3DGS-based framework poses a significant challenge. To address this obstacle, we introduce an intuitive strategy that improves depth " + }, + { + "bbox": [ + 304, + 415, + 545, + 545 + ], + "type": "inline_equation", + "content": "\\hat{D}" + }, + { + "bbox": [ + 304, + 415, + 545, + 545 + ], + "type": "text", + "content": " and leverages the depth gradient to derive pseudo normals " + }, + { + "bbox": [ + 304, + 415, + 545, + 545 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{n}}_{\\hat{D}} = \\nabla_{\\mathrm{uv}}\\hat{D}" + }, + { + "bbox": [ + 304, + 415, + 545, + 545 + ], + "type": "text", + "content": ". These pseudo normals then guide the optimization of normals within the 3D Gaussians." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": "Depth Generation Given a pretrained 3D Gaussians " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": " and a view designated for rendering, the pixel's shading results in that view can be obtained by Eq. (3). Consequently, it is reasonable to utilize the same volumetric accumulation to compute the depth " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{D} = \\sum_{i=1}^{N} T_i \\alpha_i d_i" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "d_i" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": " denotes the distance from the corresponding 3D Gaussian to the image plane. However, we observed the floating problem during volumetric accumulation, unlike the backward mapping volume rendering used in NeRF. During the 3DGS optimization, the adaptive control of the Gaussian density may result in the depth falling in front of the 3D Gaussians, thereby posing challenges in accurately predicting the depth. Specifically, the backward mapping methods can obtain an accurate depth by considering only peak samples," + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21646" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 70, + 532, + 329 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 532, + 329 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 532, + 329 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 532, + 329 + ], + "type": "image", + "image_path": "520935b331613e8ca20236530cf0e59333c0c3f429eea69a2269ca4991750034.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 339, + 548, + 386 + ], + "lines": [ + { + "bbox": [ + 46, + 339, + 548, + 386 + ], + "spans": [ + { + "bbox": [ + 46, + 339, + 548, + 386 + ], + "type": "text", + "content": "Figure 2. GS-IR Pipeline. We propose a novel GS-based inverse rendering framework, called GS-IR, to reconstruct scene geometry, materials, and unknown natural illumination from multi-view captured images. Our GS-IR consists of three well-designed stage strategies using 3D Gaussian and differentiable forward mapping splatting to achieve physical-based rendering. In our approach, the Gaussian stores not only the basic 3DGS information but also the normal and material properties, enhancing its capabilities for inverse rendering tasks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 62, + 391, + 274, + 475 + ], + "blocks": [ + { + "bbox": [ + 62, + 391, + 274, + 475 + ], + "lines": [ + { + "bbox": [ + 62, + 391, + 274, + 475 + ], + "spans": [ + { + "bbox": [ + 62, + 391, + 274, + 475 + ], + "type": "image", + "image_path": "ca483dfb406e9c50d41bba16dfb9a89fa5f2196c656b2ec5fc2847d0a156fea0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 483, + 287, + 528 + ], + "lines": [ + { + "bbox": [ + 46, + 483, + 287, + 528 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 287, + 528 + ], + "type": "text", + "content": "Figure 3. Depth Illustration. By considering the depth as a linear interpolation of the distances from 3D Gaussians to the image plane, and ensuring it lies between the minimum and maximum distance, our method could produce accurate depth." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 530, + 288, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 530, + 288, + 626 + ], + "spans": [ + { + "bbox": [ + 46, + 530, + 288, + 626 + ], + "type": "text", + "content": "that is " + }, + { + "bbox": [ + 46, + 530, + 288, + 626 + ], + "type": "inline_equation", + "content": "\\hat{D} = d_{i^*}" + }, + { + "bbox": [ + 46, + 530, + 288, + 626 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 530, + 288, + 626 + ], + "type": "inline_equation", + "content": "i^* = \\arg \\max_i T_i \\alpha_i" + }, + { + "bbox": [ + 46, + 530, + 288, + 626 + ], + "type": "text", + "content": ". However, for 3DGS, a typical forward mapping method, the peak selection results in disc aliasing within 3DGS. To overcome this limitation, we consider that depth " + }, + { + "bbox": [ + 46, + 530, + 288, + 626 + ], + "type": "inline_equation", + "content": "\\hat{D}" + }, + { + "bbox": [ + 46, + 530, + 288, + 626 + ], + "type": "text", + "content": " must be between the minimum and maximum distance of 3D Gaussians to the image plane, as illustrated in Fig. 3. We then treat the depth as a linear interpolation of the distances from 3D Gaussians to the image plane:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 97, + 633, + 287, + 663 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 633, + 287, + 663 + ], + "spans": [ + { + "bbox": [ + 97, + 633, + 287, + 663 + ], + "type": "interline_equation", + "content": "\\hat {D} = \\sum_ {i = 1} ^ {N} \\hat {w} _ {i} d _ {i}, \\quad \\hat {w} _ {i} = \\frac {T _ {i} \\alpha_ {i}}{\\sum_ {i = 1} ^ {N} T _ {i} \\alpha_ {i}}. \\tag {6}", + "image_path": "e82288197c1a12ab0a9b0a7b2cf1a9bcbfa4761f8b72e2840f902f2e9e614696.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 288, + 714 + ], + "type": "text", + "content": "Normal Derivation While the accurate prediction of depth within Gaussians provides better guidance for the normal reconstruction, directly using depth gradient to produce normals has two limitations that still cannot meet the require" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 393, + 547, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 393, + 547, + 527 + ], + "spans": [ + { + "bbox": [ + 304, + 393, + 547, + 527 + ], + "type": "text", + "content": "ment for effective inverse rendering. First, the depth gradient estimation is highly sensitive to noise, making the predicted normal often extremely noisy; Second, the normals derived individually from each view's depth map do not satisfy multi-view consistency. To address these issues, we use Gaussian " + }, + { + "bbox": [ + 304, + 393, + 547, + 527 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 304, + 393, + 547, + 527 + ], + "type": "text", + "content": " as a proxy for normal estimation instead of directly from the depth gradient. Benefiting from the efficiency of 3DGS, we obtain the depth " + }, + { + "bbox": [ + 304, + 393, + 547, + 527 + ], + "type": "inline_equation", + "content": "\\hat{D}" + }, + { + "bbox": [ + 304, + 393, + 547, + 527 + ], + "type": "text", + "content": " and normals " + }, + { + "bbox": [ + 304, + 393, + 547, + 527 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{n}}" + }, + { + "bbox": [ + 304, + 393, + 547, + 527 + ], + "type": "text", + "content": " of the observed view after performing a single rendering pass. We then tie these predicted pseudo normal to the underlying depth gradient normal " + }, + { + "bbox": [ + 304, + 393, + 547, + 527 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{n}}_{\\hat{D}}" + }, + { + "bbox": [ + 304, + 393, + 547, + 527 + ], + "type": "text", + "content": " using a simple penalty:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 387, + 531, + 545, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 531, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 387, + 531, + 545, + 544 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {n - p} = \\left\\| \\hat {\\boldsymbol {n}} - \\hat {\\boldsymbol {n}} _ {\\hat {D}} \\right\\|, \\tag {7}", + "image_path": "c14ff2d90a69ebdf0102289368066e7e774f109649942c7621c83488d6dfd0ed.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 546, + 545, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 545, + 631 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 545, + 631 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 546, + 545, + 631 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{n}} = \\sum_{i=1}^{N} T_i \\alpha_i \\pmb{n}_i" + }, + { + "bbox": [ + 304, + 546, + 545, + 631 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 546, + 545, + 631 + ], + "type": "inline_equation", + "content": "\\pmb{n}_i" + }, + { + "bbox": [ + 304, + 546, + 545, + 631 + ], + "type": "text", + "content": " is the normal stored in the 3D Gaussian. Secondly, unlike the MLP-based normal estimation [22] acting MLP as a low-pass filter, the predicted normal of Gaussian " + }, + { + "bbox": [ + 304, + 546, + 545, + 631 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 304, + 546, + 545, + 631 + ], + "type": "text", + "content": " is rough, so smoothness regularization should be included. We introduce the TV term " + }, + { + "bbox": [ + 304, + 546, + 545, + 631 + ], + "type": "inline_equation", + "content": "TV_{\\text{normal}}" + }, + { + "bbox": [ + 304, + 546, + 545, + 631 + ], + "type": "text", + "content": " to smooth the predicted normal " + }, + { + "bbox": [ + 304, + 546, + 545, + 631 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{n}}" + }, + { + "bbox": [ + 304, + 546, + 545, + 631 + ], + "type": "text", + "content": ". For more details, please refer to the supplement." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 632, + 545, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 632, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 304, + 632, + 545, + 690 + ], + "type": "text", + "content": "In optimization of the first stage, we optimize 3D Gaussians " + }, + { + "bbox": [ + 304, + 632, + 545, + 690 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 304, + 632, + 545, + 690 + ], + "type": "text", + "content": " (storing SH coefficients for view-dependent color " + }, + { + "bbox": [ + 304, + 632, + 545, + 690 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 632, + 545, + 690 + ], + "type": "text", + "content": ", opacity " + }, + { + "bbox": [ + 304, + 632, + 545, + 690 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 632, + 545, + 690 + ], + "type": "text", + "content": ", and normal " + }, + { + "bbox": [ + 304, + 632, + 545, + 690 + ], + "type": "inline_equation", + "content": "\\hat{n}" + }, + { + "bbox": [ + 304, + 632, + 545, + 690 + ], + "type": "text", + "content": ") by using the color reconstruction loss " + }, + { + "bbox": [ + 304, + 632, + 545, + 690 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 304, + 632, + 545, + 690 + ], + "type": "text", + "content": ", which is the same as 3DGS [24], and the proposed normal loss " + }, + { + "bbox": [ + 304, + 632, + 545, + 690 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_n" + }, + { + "bbox": [ + 304, + 632, + 545, + 690 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 367, + 696, + 545, + 709 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 696, + 545, + 709 + ], + "spans": [ + { + "bbox": [ + 367, + 696, + 545, + 709 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {n} = \\mathcal {L} _ {n - p} + \\lambda_ {n - T V} T V _ {\\text {n o r m a l}}. \\tag {8}", + "image_path": "c05365534ca0f5a1ecd7149a71aa2f4ca7eca86a4913b2468dc6f249bb5a1f63.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21647" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 276, + 213 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 276, + 213 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 276, + 213 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 276, + 213 + ], + "type": "image", + "image_path": "122d068955408441c595b1a88809c97880f2d58b74013a64e5351d41b7f52f34.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 221, + 289, + 321 + ], + "lines": [ + { + "bbox": [ + 46, + 221, + 289, + 321 + ], + "spans": [ + { + "bbox": [ + 46, + 221, + 289, + 321 + ], + "type": "text", + "content": "Figure 4. Baking. We employ the spherical harmonics (SH) architecture to bake occlusion volumes for modeling indirect illumination. For each grid of occlusion volumes, we initially use 3D Gaussians to compute the depth cubemap by performing six forward mapping splatting passes. Next, we convert the depth cubemap into a binary occlusion cubemap based on a distance threshold. Finally, the occlusion cubemap is baked as SH coefficients, enabling efficient interpolation of the occlusion cubemap at any point within the scene." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 329, + 217, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 329, + 217, + 342 + ], + "spans": [ + { + "bbox": [ + 47, + 329, + 217, + 342 + ], + "type": "text", + "content": "4.2. Indirect Illumination Modeling" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 350, + 287, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 350, + 287, + 421 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 287, + 421 + ], + "type": "text", + "content": "Drawing inspiration from the successful implementation of precomputation techniques in the video game industry (e.g. Irradiance Volume in Blender [3], Lightmass Volume in Unreal [1], and Light Probes in Unity [2]), we introduce spherical harmonics (SH) architectures to store occlusion information and model indirect illumination." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 423, + 287, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 423, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 47, + 423, + 287, + 496 + ], + "type": "text", + "content": "Given the optimized 3D Gaussians " + }, + { + "bbox": [ + 47, + 423, + 287, + 496 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 47, + 423, + 287, + 496 + ], + "type": "text", + "content": " from one stage (cf. Sec. 4.1), we freeze " + }, + { + "bbox": [ + 47, + 423, + 287, + 496 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 47, + 423, + 287, + 496 + ], + "type": "text", + "content": " and regularly place occlusion volumes " + }, + { + "bbox": [ + 47, + 423, + 287, + 496 + ], + "type": "inline_equation", + "content": "\\nu^{\\mathrm{occl}}" + }, + { + "bbox": [ + 47, + 423, + 287, + 496 + ], + "type": "text", + "content": " in the 3D space. For each volume " + }, + { + "bbox": [ + 47, + 423, + 287, + 496 + ], + "type": "inline_equation", + "content": "\\pmb{v}_i^{\\mathrm{occl}}\\subset \\nu^{\\mathrm{occl}}" + }, + { + "bbox": [ + 47, + 423, + 287, + 496 + ], + "type": "text", + "content": " we then cache the occlusion in the form of SH coefficients " + }, + { + "bbox": [ + 47, + 423, + 287, + 496 + ], + "type": "inline_equation", + "content": "\\pmb{f}_i^{\\mathrm{o}}" + }, + { + "bbox": [ + 47, + 423, + 287, + 496 + ], + "type": "text", + "content": ". Consequently, the formula of the occlusion " + }, + { + "bbox": [ + 47, + 423, + 287, + 496 + ], + "type": "inline_equation", + "content": "O(\\cdot)" + }, + { + "bbox": [ + 47, + 423, + 287, + 496 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 47, + 423, + 287, + 496 + ], + "type": "inline_equation", + "content": "\\pmb{v}_i^{\\mathrm{occl}}" + }, + { + "bbox": [ + 47, + 423, + 287, + 496 + ], + "type": "text", + "content": " with respect to the direction " + }, + { + "bbox": [ + 47, + 423, + 287, + 496 + ], + "type": "inline_equation", + "content": "(\\theta ,\\phi)" + }, + { + "bbox": [ + 47, + 423, + 287, + 496 + ], + "type": "text", + "content": " is expressed as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 95, + 510, + 287, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 510, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 95, + 510, + 287, + 540 + ], + "type": "interline_equation", + "content": "O (\\theta , \\phi) = \\sum_ {l = 0} ^ {\\text {d e g}} \\sum_ {m = - l} ^ {l} \\mathbf {f} _ {i (l m)} ^ {0} Y _ {l m} (\\theta , \\phi), \\tag {9}", + "image_path": "44ded620b1cd9fd84c27e79d36a22199e10e67948d36da308ca35a4fad1ef8f6.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 552, + 287, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 552, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 46, + 552, + 287, + 576 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 552, + 287, + 576 + ], + "type": "inline_equation", + "content": "deg" + }, + { + "bbox": [ + 46, + 552, + 287, + 576 + ], + "type": "text", + "content": " denotes the degree of SH, and " + }, + { + "bbox": [ + 46, + 552, + 287, + 576 + ], + "type": "inline_equation", + "content": "\\{Y_{lm}(\\cdot)\\}" + }, + { + "bbox": [ + 46, + 552, + 287, + 576 + ], + "type": "text", + "content": " is a set of real basis of SH." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 578, + 288, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 578, + 288, + 699 + ], + "spans": [ + { + "bbox": [ + 46, + 578, + 288, + 699 + ], + "type": "text", + "content": "As discussed in Sec. 4.1, the 3DGS technique employs a forward mapping approach that projects 3D points to the 2D plane, in contrast to the backward mapping volume rendering utilized in NeRF, which means it cannot use ray marching to calculate occlusions. To precompute the SH coefficients " + }, + { + "bbox": [ + 46, + 578, + 288, + 699 + ], + "type": "inline_equation", + "content": "\\pmb{f}_i^{\\mathrm{o}}" + }, + { + "bbox": [ + 46, + 578, + 288, + 699 + ], + "type": "text", + "content": " of occlusion volume " + }, + { + "bbox": [ + 46, + 578, + 288, + 699 + ], + "type": "inline_equation", + "content": "\\pmb{v}_i^{\\mathrm{occl}}" + }, + { + "bbox": [ + 46, + 578, + 288, + 699 + ], + "type": "text", + "content": ", we obtain the depth cubemap " + }, + { + "bbox": [ + 46, + 578, + 288, + 699 + ], + "type": "inline_equation", + "content": "\\{\\hat{D}_p^i\\}_{p=1}^6" + }, + { + "bbox": [ + 46, + 578, + 288, + 699 + ], + "type": "text", + "content": " by performing six times rendering passes, once for each face of the cubemap. We then convert it into a binary occlusion cubemap " + }, + { + "bbox": [ + 46, + 578, + 288, + 699 + ], + "type": "inline_equation", + "content": "\\{\\hat{O}_p^i\\}_{p=1}^6" + }, + { + "bbox": [ + 46, + 578, + 288, + 699 + ], + "type": "text", + "content": " based on a manually set distance threshold." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 59, + 701, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 288, + 714 + ], + "type": "text", + "content": "Finally, we convolve the the occlusion cubemap" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 71, + 536, + 87 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 536, + 87 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 536, + 87 + ], + "type": "inline_equation", + "content": "\\{\\hat{O}_p^i\\}_{p = 1}^6" + }, + { + "bbox": [ + 306, + 71, + 536, + 87 + ], + "type": "text", + "content": " using SH bases and get the SH coefficients " + }, + { + "bbox": [ + 306, + 71, + 536, + 87 + ], + "type": "inline_equation", + "content": "\\pmb{f}_i^0" + }, + { + "bbox": [ + 306, + 71, + 536, + 87 + ], + "type": "text", + "content": " .." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 341, + 89, + 545, + 133 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 89, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 341, + 89, + 545, + 133 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\boldsymbol {f} _ {i (l m)} ^ {0} = \\int_ {S ^ {2}} \\hat {\\boldsymbol {O}} (\\boldsymbol {\\omega}) Y _ {l m} (\\boldsymbol {\\omega}) d \\boldsymbol {\\omega} \\tag {10} \\\\ = \\int_ {0} ^ {2 \\pi} \\int_ {0} ^ {\\pi} \\sin \\theta \\hat {\\boldsymbol {O}} (\\theta , \\phi) Y _ {l m} (\\theta , \\phi) d \\theta d \\phi , \\\\ \\end{array}", + "image_path": "3dbc1454860019438b9c3984ea6677b1e538e91dd12a8a0e6d8cf3fdb5d17035.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 135, + 545, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 135, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 304, + 135, + 545, + 184 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 135, + 545, + 184 + ], + "type": "inline_equation", + "content": "S^2" + }, + { + "bbox": [ + 304, + 135, + 545, + 184 + ], + "type": "text", + "content": " denotes the unit sphere, and " + }, + { + "bbox": [ + 304, + 135, + 545, + 184 + ], + "type": "inline_equation", + "content": "\\hat{O} (\\theta ,\\phi)" + }, + { + "bbox": [ + 304, + 135, + 545, + 184 + ], + "type": "text", + "content": " denotes the occlusion query from the occlusion cubemap " + }, + { + "bbox": [ + 304, + 135, + 545, + 184 + ], + "type": "inline_equation", + "content": "\\{\\hat{O}_p^i\\}_{p = 1}^6" + }, + { + "bbox": [ + 304, + 135, + 545, + 184 + ], + "type": "text", + "content": " Note that we numerically calculate the convolution in Eq. (10) in parallel. The caching process is shown in Fig. 4." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 185, + 546, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 185, + 546, + 282 + ], + "spans": [ + { + "bbox": [ + 304, + 185, + 546, + 282 + ], + "type": "text", + "content": "To handle the indirect illumination in occlusion regions, we also maintain illumination volumes " + }, + { + "bbox": [ + 304, + 185, + 546, + 282 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^{\\mathrm{illu}}" + }, + { + "bbox": [ + 304, + 185, + 546, + 282 + ], + "type": "text", + "content": " to cache the indirect illumination. Similar to the caching process of occlusion volumes, the caching target of illumination volumes changes from the occlusion cubemap " + }, + { + "bbox": [ + 304, + 185, + 546, + 282 + ], + "type": "inline_equation", + "content": "\\{\\hat{O}_p^i\\}_{p=1}^6" + }, + { + "bbox": [ + 304, + 185, + 546, + 282 + ], + "type": "text", + "content": " to the captured environment cubemap " + }, + { + "bbox": [ + 304, + 185, + 546, + 282 + ], + "type": "inline_equation", + "content": "\\{\\hat{I}_p^i\\}_{p=1}^6" + }, + { + "bbox": [ + 304, + 185, + 546, + 282 + ], + "type": "text", + "content": ". These cubemaps can be obtained simultaneously by conducting six rendering passes. For more details, please refer to the supplement." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 289, + 441, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 289, + 441, + 303 + ], + "spans": [ + { + "bbox": [ + 305, + 289, + 441, + 303 + ], + "type": "text", + "content": "4.3. Intrinsic Decomposition" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 308, + 545, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 308, + 545, + 366 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 545, + 366 + ], + "type": "text", + "content": "In the final stage, we employ differentiable splatting in conjunction with a PBR pipeline to accomplish the intrinsic decomposition. According to Eq. (5), the rendering equation Eq. (4) is rewritten as diffuse " + }, + { + "bbox": [ + 304, + 308, + 545, + 366 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{d}}" + }, + { + "bbox": [ + 304, + 308, + 545, + 366 + ], + "type": "text", + "content": " and specular " + }, + { + "bbox": [ + 304, + 308, + 545, + 366 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{s}}" + }, + { + "bbox": [ + 304, + 308, + 545, + 366 + ], + "type": "text", + "content": " components:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 371, + 545, + 433 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 545, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 545, + 433 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {o} (\\boldsymbol {x}, \\boldsymbol {v}) = \\int_ {\\Omega} \\left[ (1 - m) \\frac {\\boldsymbol {a}}{\\pi} + \\frac {D F G}{4 (\\boldsymbol {n} \\cdot \\boldsymbol {l}) (\\boldsymbol {n} \\cdot \\boldsymbol {v})} \\right] L _ {i} (\\boldsymbol {x}, \\boldsymbol {l}) (\\boldsymbol {l} \\cdot \\boldsymbol {n}) d \\boldsymbol {l} \\\\ L _ {\\mathrm {d}} = (1 - m) \\frac {\\boldsymbol {a}}{\\pi} \\int_ {\\Omega} L _ {i} (\\boldsymbol {x}, \\boldsymbol {l}) (\\boldsymbol {l} \\cdot \\boldsymbol {n}) d \\boldsymbol {l} \\tag {11} \\\\ L _ {\\mathrm {s}} = \\int_ {\\Omega} \\frac {D F G}{4 (\\boldsymbol {n} \\cdot \\boldsymbol {l}) (\\boldsymbol {n} \\cdot \\boldsymbol {v})} L _ {i} (\\boldsymbol {x}, \\boldsymbol {l}) (\\boldsymbol {l} \\cdot \\boldsymbol {n}) d \\boldsymbol {l}. \\\\ \\end{array}", + "image_path": "13c05a655cf0fd525b1d1026faa8e9bb0b957b98e111014ccce98a358b2b747e.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 438, + 545, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 545, + 486 + ], + "type": "text", + "content": "In GS-IR, we adopt an image-based lighting (IBL) model and split-sum approximation [23] to tackle the intractable integral. To calculate the diffuse component " + }, + { + "bbox": [ + 304, + 438, + 545, + 486 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{d}}" + }, + { + "bbox": [ + 304, + 438, + 545, + 486 + ], + "type": "text", + "content": ", the illumination " + }, + { + "bbox": [ + 304, + 438, + 545, + 486 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{d}}" + }, + { + "bbox": [ + 304, + 438, + 545, + 486 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 489, + 545, + 547 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 489, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 315, + 489, + 545, + 547 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} I _ {\\mathrm {d}} (\\boldsymbol {x}) = \\int_ {\\Omega} L _ {i} (\\boldsymbol {x}, \\boldsymbol {l}) (\\boldsymbol {l} \\cdot \\boldsymbol {n}) d \\boldsymbol {l} \\\\ = \\int_ {\\Omega_ {\\mathrm {V i s}}} L _ {i} ^ {\\operatorname {d i r}} (\\boldsymbol {x}, \\boldsymbol {l}) (\\boldsymbol {l} \\cdot \\boldsymbol {n}) d \\boldsymbol {l} + \\int_ {\\Omega_ {\\mathrm {O c c l}}} L _ {i} ^ {\\operatorname {i n d i r}} (\\boldsymbol {x}, \\boldsymbol {l}) (\\boldsymbol {l} \\cdot \\boldsymbol {n}) d \\boldsymbol {l} \\tag {12} \\\\ \\approx \\left(1 - \\mathrm {O} (\\boldsymbol {x})\\right) I _ {\\mathrm {d}} ^ {\\operatorname {d i r}} (\\boldsymbol {x}) + \\mathrm {O} (\\boldsymbol {x}) I _ {\\mathrm {d}} ^ {\\operatorname {i n d i r}} (\\boldsymbol {x}), \\\\ \\end{array}", + "image_path": "e162cddaa450989c54a9606c1decd63d8e4150d3f7b6c518ba52365932a0bc84.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 549, + 545, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 549, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 304, + 549, + 545, + 645 + ], + "type": "text", + "content": "where the first component indicates direct illumination and the second is indirect illumination. Notably, our baking-based indirect illumination model enables us to calculate the occlusion and illumination online. This means that our GS-IR achieves intrinsic decomposition while maintaining real-time rendering performance. For the specular component " + }, + { + "bbox": [ + 304, + 549, + 545, + 645 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{s}}" + }, + { + "bbox": [ + 304, + 549, + 545, + 645 + ], + "type": "text", + "content": ", we follow split-sum approximation and treat the integral as two separate integrals:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 657, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 657, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 321, + 657, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {\\mathrm {s}} = \\int_ {\\Omega} \\frac {D F G}{4 \\langle \\boldsymbol {n} \\cdot \\boldsymbol {l} \\rangle \\langle \\boldsymbol {n} \\cdot \\boldsymbol {v} \\rangle} L _ {i} (\\boldsymbol {l}) \\langle \\boldsymbol {l} \\cdot \\boldsymbol {n} \\rangle d \\boldsymbol {l} \\\\ \\approx \\underbrace {\\int_ {\\Omega} \\frac {D F G}{4 \\left\\langle \\boldsymbol {n} \\cdot \\boldsymbol {l} \\right\\rangle \\left\\langle \\boldsymbol {n} \\cdot \\boldsymbol {v} \\right\\rangle}} _ {\\text {E n v i r o n m e n t B R D F -} R} \\left\\langle \\boldsymbol {l} \\cdot \\boldsymbol {n} \\right\\rangle d \\boldsymbol {l} \\underbrace {\\int_ {\\Omega} D L _ {i} (\\boldsymbol {l}) \\left\\langle \\boldsymbol {l} \\cdot \\boldsymbol {n} \\right\\rangle d \\boldsymbol {l}} _ {\\text {P r e - F i l t e r e d E n v i r o n m e n t M a p -} I _ {s}}, \\tag {13} \\\\ \\end{array}", + "image_path": "8fcad3aacb634cf21892e2d689234f889d211a2f2dfcf9b2e2efb6efe46efb4c.jpg" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21648" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 547, + 264 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 547, + 264 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 547, + 264 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 547, + 264 + ], + "type": "image", + "image_path": "fec20d9f6c6a5555aabe002a85419f33cc928bbe4f577ec908c4c453d5265c6d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 271, + 547, + 328 + ], + "lines": [ + { + "bbox": [ + 46, + 271, + 547, + 328 + ], + "spans": [ + { + "bbox": [ + 46, + 271, + 547, + 328 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison on TensoIR Synthetic. We visualize the estimated normal, albedo, and rendering results of our GS-IR and baseline methods on two scenes. By utilizing the efficient 3D Gaussian representation and a robust tile-based rasterizer, GS-IR achieves rapid convergence and supports real-time rendering. This performance advantage underscores the effectiveness of our method in addressing complex inverse rendering tasks, thereby surpassing existing state-of-the-art approaches. (For albedo reconstruction results, we follow NeRFactor [46] and scale each RGB channel by a global scalar.)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 348, + 288, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 348, + 288, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 288, + 384 + ], + "type": "text", + "content": "where both " + }, + { + "bbox": [ + 46, + 348, + 288, + 384 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 348, + 288, + 384 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 348, + 288, + 384 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{s}}" + }, + { + "bbox": [ + 46, + 348, + 288, + 384 + ], + "type": "text", + "content": " can be precomputed in advance and stored in look-up tables. With Eq. (12) and Eq. (13), the rendering results of Eq. (11) can be represented as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 387, + 287, + 425 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 387, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 56, + 387, + 287, + 425 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {o} (\\boldsymbol {x}, \\boldsymbol {v}) = L _ {\\mathrm {d}} + L _ {\\mathrm {s}} \\\\ \\approx (1 - m) \\frac {\\boldsymbol {a}}{\\pi} \\left[ (1 - \\mathrm {O} (\\boldsymbol {x})) I _ {\\mathrm {d}} ^ {\\operatorname {d i r}} (\\boldsymbol {x}) + \\mathrm {O} (\\boldsymbol {x}) I _ {\\mathrm {d}} ^ {\\operatorname {i n d i r}} (\\boldsymbol {x}) \\right] + R I _ {\\mathrm {s}}. \\tag {14} \\\\ \\end{array}", + "image_path": "105afe62c1591142a611952d2fde0f0198c65133800de4d261e10bc5ed92cadd.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 426, + 287, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 426, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 47, + 426, + 287, + 475 + ], + "type": "text", + "content": "For intrinsic decomposition, we optimize the material " + }, + { + "bbox": [ + 47, + 426, + 287, + 475 + ], + "type": "inline_equation", + "content": "\\hat{M}" + }, + { + "bbox": [ + 47, + 426, + 287, + 475 + ], + "type": "text", + "content": " (i.e. albedo " + }, + { + "bbox": [ + 47, + 426, + 287, + 475 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 47, + 426, + 287, + 475 + ], + "type": "text", + "content": ", metallic value " + }, + { + "bbox": [ + 47, + 426, + 287, + 475 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 47, + 426, + 287, + 475 + ], + "type": "text", + "content": ", and roughness " + }, + { + "bbox": [ + 47, + 426, + 287, + 475 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 47, + 426, + 287, + 475 + ], + "type": "text", + "content": ") stored in 3D Gaussians " + }, + { + "bbox": [ + 47, + 426, + 287, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 47, + 426, + 287, + 475 + ], + "type": "text", + "content": ", environment map " + }, + { + "bbox": [ + 47, + 426, + 287, + 475 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{E}}" + }, + { + "bbox": [ + 47, + 426, + 287, + 475 + ], + "type": "text", + "content": ", and illumination volumes " + }, + { + "bbox": [ + 47, + 426, + 287, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^{\\mathrm{illu}}" + }, + { + "bbox": [ + 47, + 426, + 287, + 475 + ], + "type": "text", + "content": " by minimizing the decomposition loss " + }, + { + "bbox": [ + 47, + 426, + 287, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_d" + }, + { + "bbox": [ + 47, + 426, + 287, + 475 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 475, + 287, + 516 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 287, + 516 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 287, + 516 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d} = \\underbrace {\\left\\| \\boldsymbol {I} - \\hat {\\boldsymbol {I}} ^ {\\text {s h a d e}} \\left(\\hat {\\boldsymbol {M}} , \\hat {\\boldsymbol {E}} , \\mathcal {V} ^ {\\text {i l l u}}\\right) \\right\\|} _ {\\mathcal {L} _ {\\text {s h a d e}}} + \\underbrace {\\lambda_ {M} T V _ {\\mathrm {m a t}}} _ {\\mathcal {L} _ {\\text {m a t e r i a l}}} + \\underbrace {\\lambda_ {\\boldsymbol {E}} T V _ {\\text {l i g h t}}} _ {\\mathcal {L} _ {\\text {l i g h t}}}, \\tag {15}", + "image_path": "93a3d8826a93d19cdc84564d64e84888cc449a3ffef622a5f1e9123e73cbcfee.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 516, + 287, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 516, + 287, + 565 + ], + "spans": [ + { + "bbox": [ + 46, + 516, + 287, + 565 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 516, + 287, + 565 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{shade}}" + }, + { + "bbox": [ + 46, + 516, + 287, + 565 + ], + "type": "text", + "content": " indicate the shade loss. " + }, + { + "bbox": [ + 46, + 516, + 287, + 565 + ], + "type": "inline_equation", + "content": "\\hat{I}^{\\mathrm{shade}}(\\hat{M},\\hat{E},\\hat{\\mathcal{V}}^{\\mathrm{illu}})" + }, + { + "bbox": [ + 46, + 516, + 287, + 565 + ], + "type": "text", + "content": " is the recovered image that uses the PBR pipeline defined by Eq. (14). Please refer to the supplement for more details about the material TV loss " + }, + { + "bbox": [ + 46, + 516, + 287, + 565 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{material}}" + }, + { + "bbox": [ + 46, + 516, + 287, + 565 + ], + "type": "text", + "content": " and lighting TV loss " + }, + { + "bbox": [ + 46, + 516, + 287, + 565 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{light}}" + }, + { + "bbox": [ + 46, + 516, + 287, + 565 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 574, + 128, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 128, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 128, + 588 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": "Dataset & Metrics We conduct experiments using benchmark datasets of TensoIR Synthetic [22] and Mip-NeRF 360 [5] for decomposing both objects and scenes. They contain 4 objects with reference materials and 7 publicly available scenes, respectively. To verify the efficacy of our normal reconstruction, we evaluate the normal quality on the TensoIR Synthetic [22] dataset using mean angular error (MAE). We further assess our reconstructed albedo quality on this synthetic dataset. More generally, we evaluate the synthesized novel view on both datasets in terms of Peak" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 321, + 346, + 422, + 412 + ], + "blocks": [ + { + "bbox": [ + 321, + 346, + 422, + 412 + ], + "lines": [ + { + "bbox": [ + 321, + 346, + 422, + 412 + ], + "spans": [ + { + "bbox": [ + 321, + 346, + 422, + 412 + ], + "type": "image", + "image_path": "adbe248cba838b012773c77d3e24a54dc606e0633c32d31505fa1ee3bf19c23c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 429, + 346, + 524, + 412 + ], + "blocks": [ + { + "bbox": [ + 429, + 346, + 524, + 412 + ], + "lines": [ + { + "bbox": [ + 429, + 346, + 524, + 412 + ], + "spans": [ + { + "bbox": [ + 429, + 346, + 524, + 412 + ], + "type": "image", + "image_path": "0c29f848a8fb7d39cef3e865df897a311b5552f97d19c22e4630297fd7751282.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 321, + 413, + 425, + 478 + ], + "blocks": [ + { + "bbox": [ + 321, + 413, + 425, + 478 + ], + "lines": [ + { + "bbox": [ + 321, + 413, + 425, + 478 + ], + "spans": [ + { + "bbox": [ + 321, + 413, + 425, + 478 + ], + "type": "image", + "image_path": "7981f9ed67a6b735312df99f5e779e5291f3247653e2de0a592343aa1f587c43.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 488, + 547, + 544 + ], + "lines": [ + { + "bbox": [ + 304, + 488, + 547, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 488, + 547, + 544 + ], + "type": "text", + "content": "Figure 6. Novel view synthesis results on Mip-NeRF 360. GS-IR can reconstruct scene details including geometric normals and high-frequency appearance, rendering high-fidelity appearance and recovering fine geometric details such as those on leaves and bicycle axles. Better viewed on screen with zoom in." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 429, + 413, + 533, + 479 + ], + "blocks": [ + { + "bbox": [ + 429, + 413, + 533, + 479 + ], + "lines": [ + { + "bbox": [ + 429, + 413, + 533, + 479 + ], + "spans": [ + { + "bbox": [ + 429, + 413, + 533, + 479 + ], + "type": "image", + "image_path": "f009744cc300cc761df0f99cd28e47e2da474757ec2ff377fada9bdc1c075103.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 562, + 547, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 562, + 547, + 611 + ], + "spans": [ + { + "bbox": [ + 304, + 562, + 547, + 611 + ], + "type": "text", + "content": "Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS) [45]. Note that albedo quality assessment uses the same metrics as novel view synthesis." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 621, + 391, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 621, + 391, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 621, + 391, + 635 + ], + "type": "text", + "content": "5.1. Comparisons" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "content": "We conduct a comprehensive comparison against state-of-the-art neural field-based inverse rendering methods on the public TensoIR Synthetic dataset [22]. All the methods utilize multi-view images captured under unknown lighting conditions. Our evaluation encompasses normal quality (measured by MAE), novel view synthesis, albedo fidelity," + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21649" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 70, + 544, + 147 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 544, + 147 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 544, + 147 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 544, + 147 + ], + "type": "table", + "html": "
MethodNormal ↓ MAE ↓Novel View SynthesisAlbedoRelightRuntime
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
NeRFactor [46]6.31424.6790.9220.12025.1250.9400.10923.3830.9080.131>100 hrs
InvRender [47]5.07427.3670.9340.08927.3410.9330.10023.9730.9010.10115 hrs
NVDiffrec [32]6.07830.6960.9620.05229.1740.9080.11519.8800.8790.104<1 hr
TensoIR [22]4.10035.0880.9760.04029.2750.9500.08528.5800.9440.0815 hrs
Ours4.94835.3330.9740.03930.2860.9410.08424.3740.8850.096<1 hr
", + "image_path": "8ed8845edb740bc015a3be856500b1080dd49dc3094270b64a24d7f23300dea7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 52, + 222, + 282, + 302 + ], + "blocks": [ + { + "bbox": [ + 46, + 154, + 546, + 212 + ], + "lines": [ + { + "bbox": [ + 46, + 154, + 546, + 212 + ], + "spans": [ + { + "bbox": [ + 46, + 154, + 546, + 212 + ], + "type": "text", + "content": "Table 1. Quantitative Comparison on TensoIR Synthetic dataset. Our method outperforms baseline methods in terms of novel view synthesis and albedo quality, showcasing the effectiveness of material decomposition and PBR rendering. This is particularly noteworthy considering that our normal reconstruction is slightly inferior to TensoIR. In terms of relighting performance, we rank second, trailing only behind TensoIR. Importantly, the average training time of our GS-IR is accelerated by a factor of " + }, + { + "bbox": [ + 46, + 154, + 546, + 212 + ], + "type": "inline_equation", + "content": "5\\mathrm{x}" + }, + { + "bbox": [ + 46, + 154, + 546, + 212 + ], + "type": "text", + "content": ", making its performance acceptable and further demonstrating the effectiveness of our approach in handling complex inverse rendering tasks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 222, + 282, + 302 + ], + "lines": [ + { + "bbox": [ + 52, + 222, + 282, + 302 + ], + "spans": [ + { + "bbox": [ + 52, + 222, + 282, + 302 + ], + "type": "table", + "html": "
MethodPSNR ↑SSIM ↑LPIPS ↓Runtime ↓
NeRF++ [41]25.1120.6960.375≈ 20h
Plenoxels [17]23.0790.6250.462≈ 30m
INGP-Base [31]25.3030.6710.371≈ 5m
INGP-Big [31]25.5870.6990.331≈ 8m
Mip-NeRF 360 [31]27.5690.7930.234≈ 48h
3DGS [24]27.210.8150.214≈ 35m
Ours25.3810.7570.267≈ 45m
", + "image_path": "71b7a9668d53c3d808d7b6d288dfc0e716e83625e32dc65c7a9f0408cbde6fae.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 364, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 364, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 364, + 287, + 555 + ], + "type": "text", + "content": "relighting effects (measured by PSNR, SSIM, and LPIPS), and efficiency. Tab. 1 summarizes the quantitative comparisons on the synthesis dataset. Our method achieves superior performance in novel view synthesis and albedo quality compared to the baseline methods, demonstrating the effectiveness of material decomposition and PBR rendering, particularly given that our normal reconstruction is slightly inferior to TensoIR. Our relighting performance ranks second, only behind TensoIR. Notably, the average training time of our GS-IR is accelerated by a factor of " + }, + { + "bbox": [ + 46, + 364, + 287, + 555 + ], + "type": "inline_equation", + "content": "5\\mathrm{x}" + }, + { + "bbox": [ + 46, + 364, + 287, + 555 + ], + "type": "text", + "content": ", making its performance acceptable and demonstrating the effectiveness of our approach in handling complex inverse rendering tasks. We also include qualitative comparisons in Fig. 5, which show that our GS-IR produces reasonable albedo and photorealistic renderings that are closer to the ground truth than most methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": "Meanwhile, owing to our more efficient and compact representation with powerful expressiveness, our method showcases remarkable performance on complex real unbounded scenes [5]. Tab. 2 presents the quantitative comparisons on the real dataset. Fig. 6 demonstrates the normal reconstruction and novel view synthesis on the real dataset. Our method renders a high-fidelity appearance and recovers fine geometric details, such as those on the leaves and bicycle axil. In summary, by leveraging the efficient 3D Gaussian representation and a powerful tile-based rasterizer, GS-IR achieves fast convergence and supports real-time rendering. This performance advantage highlights the effectiveness of our method in handling complex inverse rendering" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 309, + 223, + 544, + 292 + ], + "blocks": [ + { + "bbox": [ + 46, + 311, + 287, + 344 + ], + "lines": [ + { + "bbox": [ + 46, + 311, + 287, + 344 + ], + "spans": [ + { + "bbox": [ + 46, + 311, + 287, + 344 + ], + "type": "text", + "content": "Table 2. Quantitative Comparison on Mip-NeRF 360. The results show that our inverse rendering approach even surpasses some NeRF variants dedicated to novel view synthesis." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 223, + 544, + 292 + ], + "lines": [ + { + "bbox": [ + 309, + 223, + 544, + 292 + ], + "spans": [ + { + "bbox": [ + 309, + 223, + 544, + 292 + ], + "type": "table", + "html": "
MethodTensoIR Synthetic [22]Mip-NeRF 360 [5]
Normal MAE↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Vol. Accum.16.34725.7560.8550.13122.0520.6100.394
Peak Selec.9.46628.7500.9270.08423.0930.7190.317
Linear Interp.6.21828.9830.9390.06623.1190.7070.319
Vol. Accum.†9.31530.0910.9400.07124.6640.7470.277
Peak Selec.†7.98631.0640.9500.06025.1430.7530.281
Linear Interp.†4.94835.3330.9740.03925.3810.7570.267
", + "image_path": "8339ae5be5c08400ffe63c823387ed024298440ff26d166425b38926c90df8c9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 295, + 546, + 340 + ], + "lines": [ + { + "bbox": [ + 305, + 295, + 546, + 340 + ], + "spans": [ + { + "bbox": [ + 305, + 295, + 546, + 340 + ], + "type": "text", + "content": "Table 3. Analyses on the impact of different depth generation strategies on normals. Methods without " + }, + { + "bbox": [ + 305, + 295, + 546, + 340 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 305, + 295, + 546, + 340 + ], + "type": "text", + "content": " marks directly use the normals derived from the depth map; Methods marked with " + }, + { + "bbox": [ + 305, + 295, + 546, + 340 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 305, + 295, + 546, + 340 + ], + "type": "text", + "content": " use depth derivation to optimize the normals stored in 3D Gaussians." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 319, + 345, + 532, + 426 + ], + "blocks": [ + { + "bbox": [ + 319, + 345, + 532, + 426 + ], + "lines": [ + { + "bbox": [ + 319, + 345, + 532, + 426 + ], + "spans": [ + { + "bbox": [ + 319, + 345, + 532, + 426 + ], + "type": "image", + "image_path": "a4b22106eecb0f719527926e096a12f57a24dbe6c81dccde32433f0e29266c8c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 309, + 461, + 541, + 504 + ], + "blocks": [ + { + "bbox": [ + 305, + 428, + 545, + 460 + ], + "lines": [ + { + "bbox": [ + 305, + 428, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 305, + 428, + 545, + 460 + ], + "type": "text", + "content": "Figure 7. Visual comparison of depth produced by different strategies. The linear interpolation adopted in GS-IR overcomes the floating problem and disc aliasing." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 461, + 541, + 504 + ], + "lines": [ + { + "bbox": [ + 309, + 461, + 541, + 504 + ], + "spans": [ + { + "bbox": [ + 309, + 461, + 541, + 504 + ], + "type": "table", + "html": "
MethodTensoIR Synthetic [22]Mip-NeRF 360 [5]
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
w/o occlusion34.9970.9620.04125.0600.7530.270
w/o indirect illum.35.1860.9650.04424.8980.7490.272
Ours35.3330.9740.03925.3810.7570.267
", + "image_path": "08f294b7e157d0d152b10f8dd8f7bd192431da5a1166f18478cbb891b6b016a7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 506, + 545, + 540 + ], + "lines": [ + { + "bbox": [ + 305, + 506, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 305, + 506, + 545, + 540 + ], + "type": "text", + "content": "Table 4. Analyses on the occlusion and indirect illumination. Physically modeling indirect illumination improves the inverse rendering of objects and scenes." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 555, + 537, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 555, + 537, + 567 + ], + "spans": [ + { + "bbox": [ + 306, + 555, + 537, + 567 + ], + "type": "text", + "content": "tasks, outperforming existing state-of-the-art approaches." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 575, + 406, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 575, + 406, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 575, + 406, + 586 + ], + "type": "text", + "content": "5.2. Ablation Studies" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "We initially introduce the 3DGS technique for inverse rendering in GS-IR and propose depth-derivation-based normal regularization and a baking-based method to address the challenges encountered during the process. To evaluate the efficacy of our proposed schemes, we design elaborate experiments on both TensoIR Synthetic [22] and MipNeRF 360 [5] datasets, providing comprehensive insights into the effectiveness of our approach in handling complex inverse rendering tasks. Below is the detailed ablation study on Normal Regularization and Indirect Illumination." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21650" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 72, + 154, + 114 + ], + "blocks": [ + { + "bbox": [ + 72, + 72, + 154, + 114 + ], + "lines": [ + { + "bbox": [ + 72, + 72, + 154, + 114 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 154, + 114 + ], + "type": "image", + "image_path": "06e1ddad6069322fee46aaec46d4d73b83ac3a589b44435883ab6aac4844ac13.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 78, + 116, + 149, + 176 + ], + "blocks": [ + { + "bbox": [ + 78, + 116, + 149, + 176 + ], + "lines": [ + { + "bbox": [ + 78, + 116, + 149, + 176 + ], + "spans": [ + { + "bbox": [ + 78, + 116, + 149, + 176 + ], + "type": "image", + "image_path": "26d62ae8ab69950521c2423c6f6751c682f5cb5bf22a2798efacd7f2e1699494.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 180, + 545, + 204 + ], + "lines": [ + { + "bbox": [ + 46, + 180, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 545, + 204 + ], + "type": "text", + "content": "Figure 8. Relighting Visualization. We perform relighting experiments on both synthetic and real scenes using the recovered geometry, material, and illumination properties from our GS-IR method. We test our method under different lighting conditions and directions." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 164, + 78, + 245, + 114 + ], + "blocks": [ + { + "bbox": [ + 164, + 78, + 245, + 114 + ], + "lines": [ + { + "bbox": [ + 164, + 78, + 245, + 114 + ], + "spans": [ + { + "bbox": [ + 164, + 78, + 245, + 114 + ], + "type": "image", + "image_path": "770b9ac4012b5565b75c44c29eedb055bb460b045d5b9948f5edc2fd4001eaa1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 169, + 116, + 241, + 176 + ], + "blocks": [ + { + "bbox": [ + 169, + 116, + 241, + 176 + ], + "lines": [ + { + "bbox": [ + 169, + 116, + 241, + 176 + ], + "spans": [ + { + "bbox": [ + 169, + 116, + 241, + 176 + ], + "type": "image", + "image_path": "40924564f4e82eb75884a7b39760ffaf3502d68ed7d3d77378de34087d413b56.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 255, + 72, + 335, + 114 + ], + "blocks": [ + { + "bbox": [ + 255, + 72, + 335, + 114 + ], + "lines": [ + { + "bbox": [ + 255, + 72, + 335, + 114 + ], + "spans": [ + { + "bbox": [ + 255, + 72, + 335, + 114 + ], + "type": "image", + "image_path": "6d57db72969d48d2f949c3fe9796e4a458fe0c6d606f2a4a23d5b7d23eaedb3d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 260, + 116, + 332, + 176 + ], + "blocks": [ + { + "bbox": [ + 260, + 116, + 332, + 176 + ], + "lines": [ + { + "bbox": [ + 260, + 116, + 332, + 176 + ], + "spans": [ + { + "bbox": [ + 260, + 116, + 332, + 176 + ], + "type": "image", + "image_path": "b9c42674746ef5fc3aed59a7c0e80fd2c1baae462b5123fad4059003e6c3f4b1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 347, + 72, + 427, + 114 + ], + "blocks": [ + { + "bbox": [ + 347, + 72, + 427, + 114 + ], + "lines": [ + { + "bbox": [ + 347, + 72, + 427, + 114 + ], + "spans": [ + { + "bbox": [ + 347, + 72, + 427, + 114 + ], + "type": "image", + "image_path": "7b1d79b19403f75c45aa69811fc2c11d876d60a054edadf0af4641b008fdb050.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 351, + 116, + 422, + 176 + ], + "blocks": [ + { + "bbox": [ + 351, + 116, + 422, + 176 + ], + "lines": [ + { + "bbox": [ + 351, + 116, + 422, + 176 + ], + "spans": [ + { + "bbox": [ + 351, + 116, + 422, + 176 + ], + "type": "image", + "image_path": "8e3aae9e547181498b68f34deddc2cd236575c22f65383ecb0f7eef436873df2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 438, + 72, + 518, + 114 + ], + "blocks": [ + { + "bbox": [ + 438, + 72, + 518, + 114 + ], + "lines": [ + { + "bbox": [ + 438, + 72, + 518, + 114 + ], + "spans": [ + { + "bbox": [ + 438, + 72, + 518, + 114 + ], + "type": "image", + "image_path": "f7364b1a86329aacf374d3cf64b5294ff2be41accc5b74c3206494d9c5b9cce7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 443, + 116, + 514, + 176 + ], + "blocks": [ + { + "bbox": [ + 443, + 116, + 514, + 176 + ], + "lines": [ + { + "bbox": [ + 443, + 116, + 514, + 176 + ], + "spans": [ + { + "bbox": [ + 443, + 116, + 514, + 176 + ], + "type": "image", + "image_path": "05a653bd301302f501abb5f081f5f6f0506a8fe2181bc5001bf638f5aac93307.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 50, + 219, + 102, + 333 + ], + "blocks": [ + { + "bbox": [ + 50, + 219, + 102, + 333 + ], + "lines": [ + { + "bbox": [ + 50, + 219, + 102, + 333 + ], + "spans": [ + { + "bbox": [ + 50, + 219, + 102, + 333 + ], + "type": "image", + "image_path": "4738bfbdf4bb0161b98a7a8d26f6fc3f5f82b475312889a06f42cc0fbd88aae8.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 351, + 287, + 396 + ], + "lines": [ + { + "bbox": [ + 46, + 351, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 351, + 287, + 396 + ], + "type": "text", + "content": "Figure 9. Ambient Occlusion Visualization. The visualization highlights the intricate shadowing and occlusion details captured by our GS-IR method, emphasizing the performance of our approach in modeling indirect illumination." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 109, + 220, + 194, + 333 + ], + "blocks": [ + { + "bbox": [ + 109, + 220, + 194, + 333 + ], + "lines": [ + { + "bbox": [ + 109, + 220, + 194, + 333 + ], + "spans": [ + { + "bbox": [ + 109, + 220, + 194, + 333 + ], + "type": "image", + "image_path": "af86a9520be3f710f0f85515276be486e606e7e8b99cffe40415e6fd78b857c7.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 199, + 222, + 283, + 333 + ], + "blocks": [ + { + "bbox": [ + 199, + 222, + 283, + 333 + ], + "lines": [ + { + "bbox": [ + 199, + 222, + 283, + 333 + ], + "spans": [ + { + "bbox": [ + 199, + 222, + 283, + 333 + ], + "type": "image", + "image_path": "4bb38e441cdac8b70d210c4ae655298395ad7c438da73ef80c52740d1d31b4f0.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 401, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 287, + 593 + ], + "type": "text", + "content": "Analysis on the Normal Regularization Reliable normal estimation is critical for conducting inverse rendering. To this end, we present depth-derivation-based regularization to facilitate 3D Gaussian-based normal estimation as stated in Sec. 4.1. In this section, we explore the impact of different acquisition schemes on the final normal quality and inverse rendering results. The quantitative results shown in Tab. 3 demonstrate that using 3D Gaussians as a normal proxy and adopting the linear interpolation strategy significantly improves the normal estimation and inverse rendering results. In addition, Fig. 7 qualitatively shows that conducting volumetric accumulation results in the floating problem (cf. Fig. 7 (b)). Despite peak selection overcomes this problem, it introduces disc aliasing (cf. Fig. 7 (c)). Compared with them, the linear interpolation adopted in GS-IR robustly produces accurate depth (cf. Fig. 7 (d))." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": "Analysis on the Indirect Illumination To demonstrate the effectiveness of our indirect illumination model, we compare our method with two variants: a model without occlusion volume (w/o occlusion) and a model without indirect illumination (w/o indirect illum.). The quantitative comparisons in Tab. 4 indicate that each component is crucial for estimating accurate material decomposition and generating photorealistic rendering results. Additionally, Fig. 9 showcases the ambient occlusion visualization in both synthetic and real scenes. The visualization highlights the intra-" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 218, + 545, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 218, + 545, + 291 + ], + "spans": [ + { + "bbox": [ + 305, + 218, + 545, + 291 + ], + "type": "text", + "content": "cate shadowing and occlusion details captured by our GS-IR method, emphasizing the performance of our approach in modeling indirect illumination. This analysis further supports the effectiveness of using occlusion volume and introducing indirect illumination in enhancing the decomposition capabilities of our GS-IR." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 304, + 383, + 317 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 304, + 383, + 317 + ], + "spans": [ + { + "bbox": [ + 306, + 304, + 383, + 317 + ], + "type": "text", + "content": "5.3. Application" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 324, + 545, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 324, + 545, + 433 + ], + "spans": [ + { + "bbox": [ + 305, + 324, + 545, + 433 + ], + "type": "text", + "content": "We perform relighting experiments using the recovered geometry, material, and illumination from our GS-IR method. We test GS-IR under different lighting conditions and directions, observing how the reconstructed scene responds to the changes in lighting. The results of these experiments demonstrate that our GS-IR method can effectively handle relighting applications, producing photorealistic renderings under various lighting conditions. More results can be found in the supplement." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 449, + 378, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 449, + 378, + 462 + ], + "spans": [ + { + "bbox": [ + 306, + 449, + 378, + 462 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 472, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 545, + 616 + ], + "type": "text", + "content": "We present GS-IR, a novel inverse rendering approach based on 3D Gaussian Splatting (3DGS), which employs forward mapping volume rendering to achieve photorealistic novel view synthesis and relighting results. Our GS-IR proposes an optimization scheme with depth-derivation-based regularization for normal estimation and a baking-based occlusion to model indirect lighting. These components are eventually employed to decompose material and illumination. Our extensive experiments demonstrate the effectiveness of GS-IR in achieving state-of-the-art inverse rendering results, surpassing previous neural methods in terms of both reconstruction quality and efficiency." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "text", + "content": "Limitation Spherical Harmonics (SH) is only suitable for representing low-frequency, and we only use the occlusion represented by SH to model the diffuse term of indirect illumination. Modeling the specular term of indirect illumination remains a limitation of GS-IR, and has been a challenging problem in computer graphics. We believe it would be valuable to address this limitation in future work and suggest screen space global illumination (SSGI) techniques." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21651" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 111 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 111 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 111 + ], + "type": "text", + "content": "[1] Unreal engine. https://www.unrealengine.com/. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 114, + 203, + 125 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 203, + 125 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 203, + 125 + ], + "type": "text", + "content": "[2] Unity. https://unity.com/.5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 127, + 242, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 127, + 242, + 137 + ], + "spans": [ + { + "bbox": [ + 53, + 127, + 242, + 137 + ], + "type": "text", + "content": "[3] Blender. https://www.blender.org/.5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 140, + 286, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 140, + 286, + 161 + ], + "spans": [ + { + "bbox": [ + 53, + 140, + 286, + 161 + ], + "type": "text", + "content": "[4] Tomas Akenine-Mo, Eric Haines, Nate Hoffman, et al. Real-time rendering. 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 163, + 286, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 163, + 286, + 217 + ], + "spans": [ + { + "bbox": [ + 53, + 163, + 286, + 217 + ], + "type": "text", + "content": "[5] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 2, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 219, + 286, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 219, + 286, + 273 + ], + "spans": [ + { + "bbox": [ + 53, + 219, + 286, + 273 + ], + "type": "text", + "content": "[6] Sai Bi, Zexiang Xu, Pratul Srinivasan, Ben Mildenhall, Kalyan Sunkavalli, Miloš Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Neural reflectance fields for appearance acquisition. arXiv preprint arXiv:2008.03824, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 275, + 286, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 275, + 286, + 352 + ], + "spans": [ + { + "bbox": [ + 53, + 275, + 286, + 352 + ], + "type": "text", + "content": "[7] Sai Bi, Zexiang Xu, Kalyan Sunkavalli, Miloš Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Deep reflectance volumes: Relightable reconstructions from multi-view photometric images. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part III 16, pages 294–311. Springer, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 354, + 286, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 354, + 286, + 408 + ], + "spans": [ + { + "bbox": [ + 53, + 354, + 286, + 408 + ], + "type": "text", + "content": "[8] Sai Bi, Zexiang Xu, Kalyan Sunkavalli, David Kriegman, and Ravi Ramamoorthi. Deep 3d capture: Geometry and reflectance from sparse multi-view images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5960-5969, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 410, + 286, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 410, + 286, + 464 + ], + "spans": [ + { + "bbox": [ + 53, + 410, + 286, + 464 + ], + "type": "text", + "content": "[9] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T Barron, Ce Liu, and Hendrik Lensch. Nerd: Neural reflectance decomposition from image collections. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12684-12694, 2021. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 466, + 286, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 466, + 286, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 466, + 286, + 520 + ], + "type": "text", + "content": "[10] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan Barron, and Hendrik Lensch. Neural-pil: Neural pre-integrated lighting for reflectance decomposition. Advances in Neural Information Processing Systems, 34: 10691-10704, 2021. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 522, + 286, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 286, + 566 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 286, + 566 + ], + "type": "text", + "content": "[11] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 130-141, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 567, + 286, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 286, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 286, + 622 + ], + "type": "text", + "content": "[12] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5799-5809, 2021." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 624, + 286, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 286, + 689 + ], + "type": "text", + "content": "[13] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "type": "text", + "content": "[14] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "Conference on Computer Vision, pages 333-350. Springer, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 97, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 97, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 308, + 97, + 545, + 129 + ], + "type": "text", + "content": "[15] Robert L Cook and Kenneth E. Torrance. A reflectance model for computer graphics. ACM Transactions on Graphics (ToG), 1(1):7-24, 1982. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 132, + 545, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 132, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 307, + 132, + 545, + 175 + ], + "type": "text", + "content": "[16] Yue Dong, Guojun Chen, Pieter Peers, Jiawan Zhang, and Xin Tong. Appearance-from-motion: Recovering spatially varying surface reflectance under unknown lighting. ACM Transactions on Graphics (TOG), 33(6):1-12, 2014. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 178, + 545, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 178, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 308, + 178, + 545, + 232 + ], + "type": "text", + "content": "[17] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 2, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 234, + 545, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 234, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 308, + 234, + 545, + 289 + ], + "type": "text", + "content": "[18] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbek Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12479–12488, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 292, + 545, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 292, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 308, + 292, + 545, + 335 + ], + "type": "text", + "content": "[19] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, light, and material decomposition from images using monte carlo rendering and denoising. Advances in Neural Information Processing Systems, 35:22856-22869, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 337, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 337, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 337, + 545, + 392 + ], + "type": "text", + "content": "[20] Peter Hedman, Pratul P Srinivasan, Ben Mildenhall, Jonathan T Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5875-5884, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 394, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 394, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 394, + 545, + 437 + ], + "type": "text", + "content": "[21] Zhangjin Huang, Zhihao Liang, Haojie Zhang, Yangkai Lin, and Kui Jia. Sur2f: A hybrid representation for high-quality and efficient surface reconstruction from multi-view images. arXiv preprint arXiv:2401.03704, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 440, + 545, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 440, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 308, + 440, + 545, + 495 + ], + "type": "text", + "content": "[22] Haian Jin, Isabella Liu, Peijia Xu, Xiaoshuai Zhang, Songfang Han, Sai Bi, Xiaowei Zhou, Zexiang Xu, and Hao Su. Tensoir: Tensorial inverse rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 165-174, 2023. 2, 4, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 497, + 545, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 497, + 545, + 528 + ], + "spans": [ + { + "bbox": [ + 308, + 497, + 545, + 528 + ], + "type": "text", + "content": "[23] Brian Karis and Epic Games. Real shading in unreal engine 4. Proc. Physically Based Shading Theory Practice, 4(3):1, 2013. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 532, + 545, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 532, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 308, + 532, + 545, + 575 + ], + "type": "text", + "content": "[24] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 42(4):1-14, 2023. 1, 2, 3, 4, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 578, + 545, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 578, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 308, + 578, + 545, + 620 + ], + "type": "text", + "content": "[25] Leonid Keselman and Martial Hebert. Approximate differentiable rendering with algebraic surfaces. In European Conference on Computer Vision, pages 596-614. Springer, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 623, + 545, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 623, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 308, + 623, + 545, + 655 + ], + "type": "text", + "content": "[26] Leonid Keselman and Martial Hebert. Flexible techniques for differentiable rendering with 3d gaussians. arXiv preprint arXiv:2308.14737, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "text", + "content": "[27] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21652" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 138 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 138 + ], + "type": "text", + "content": "[28] Zhihao Liang, Zhangjin Huang, Changxing Ding, and Kui Jia. Helixsurf: A robust and efficient neural implicit surface learning of indoor scenes with iterative intertwined regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13165-13174, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 140, + 288, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 288, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 288, + 185 + ], + "type": "text", + "content": "[29] Fujun Luan, Shuang Zhao, Kavita Bala, and Zhao Dong. Unified shape and svbrdf recovery using differentiable monte carlo rendering. In Computer Graphics Forum, pages 101-113. Wiley Online Library, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 186, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 287, + 239 + ], + "type": "text", + "content": "[30] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision, pages 405-421, 2020. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 241, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 287, + 285 + ], + "type": "text", + "content": "[31] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 2, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 286, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 287, + 352 + ], + "type": "text", + "content": "[32] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3d models, materials, and lighting from images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8280-8290, 2022. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 354, + 287, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 397 + ], + "type": "text", + "content": "[33] Giljoo Nam, Joo Ho Lee, Diego Gutierrez, and Min H Kim. Practical svbrdf acquisition of 3d objects with unstructured flash photography. ACM Transactions on Graphics (TOG), 37(6):1-12, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 399, + 287, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 454 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 454 + ], + "type": "text", + "content": "[34] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 456, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 521 + ], + "type": "text", + "content": "[35] Pratul P Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T Barron. Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7495-7504, 2021. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 522, + 287, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 287, + 577 + ], + "type": "text", + "content": "[36] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 578, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 633 + ], + "type": "text", + "content": "[37] Bruce Walter, Stephen R Marschner, Hongsong Li, and Kenneth E Torrance. Microfacet models for refraction through rough surfaces. In Proceedings of the 18th Eurographics conference on Rendering Techniques, pages 195-206, 2007. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "text", + "content": "[38] Rui Xia, Yue Dong, Pieter Peers, and Xin Tong. Recovering shape and spatially-varying surface reflectance under unknown illumination. ACM Transactions on Graphics (TOG), 35(6):1-12, 2016. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[39] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 508 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5438-5448, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 96, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 150 + ], + "type": "text", + "content": "[40] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems, 33:2492-2502, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "type": "text", + "content": "[41] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 186, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 240 + ], + "type": "text", + "content": "[42] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snavely. Physg: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5453-5462, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 241, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 241, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 241, + 545, + 285 + ], + "type": "text", + "content": "[43] Kai Zhang, Nick Kolkin, Sai Bi, Fujun Luan, Zexiang Xu, Eli Shechtman, and Noah Snavely. Arf: Artistic radiance fields. In European Conference on Computer Vision, pages 717-733. Springer, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 286, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 545, + 342 + ], + "type": "text", + "content": "[44] Kai Zhang, Fujun Luan, Zhengqi Li, and Noah Snavely. Iron: Inverse rendering by optimizing neural sdfs and materials from photometric images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5565-5574, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 342, + 545, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 342, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 308, + 342, + 545, + 397 + ], + "type": "text", + "content": "[45] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 398, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 398, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 308, + 398, + 545, + 453 + ], + "type": "text", + "content": "[46] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (ToG), 40(6):1-18, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 454, + 545, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 454, + 545, + 508 + ], + "spans": [ + { + "bbox": [ + 308, + 454, + 545, + 508 + ], + "type": "text", + "content": "[47] Yuanqing Zhang, Jiaming Sun, Xingyi He, Huan Fu, Rongfei Jia, and Xiaowei Zhou. Modeling indirect illumination for inverse rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18643-18652, 2022. 2, 7" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21653" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/GS-SLAM_ Dense Visual SLAM with 3D Gaussian Splatting/93048e35-cf42-49e6-8897-d6803305631d_content_list.json b/2024/GS-SLAM_ Dense Visual SLAM with 3D Gaussian Splatting/93048e35-cf42-49e6-8897-d6803305631d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..cabf5ee4db5e187bb0685b80d01433b7e866891a --- /dev/null +++ b/2024/GS-SLAM_ Dense Visual SLAM with 3D Gaussian Splatting/93048e35-cf42-49e6-8897-d6803305631d_content_list.json @@ -0,0 +1,1650 @@ +[ + { + "type": "text", + "text": "GS-SLAM: Dense Visual SLAM with 3D Gaussian Splatting", + "text_level": 1, + "bbox": [ + 179, + 130, + 789, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chi Yan $^{1,3*}$", + "bbox": [ + 217, + 178, + 313, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Delin $\\mathbf{Q}\\mathbf{u}^{1,2*}$", + "bbox": [ + 370, + 179, + 470, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dan $\\mathrm{Xu}^3$", + "bbox": [ + 527, + 180, + 602, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bin Zhao $^{1,4}$", + "bbox": [ + 666, + 180, + 759, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhigang Wang", + "bbox": [ + 251, + 198, + 379, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dong", + "bbox": [ + 444, + 198, + 491, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xuelong Li $^{1,5}$", + "bbox": [ + 614, + 198, + 725, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Shanghai AI Laboratory $^{2}$ Fudan University $^{4}$ Northwestern Polytechnical University", + "bbox": [ + 88, + 215, + 462, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Hong Kong University of Science and Technology $^{5}$ TeleAI, China Telecom Corp Ltd", + "bbox": [ + 488, + 215, + 893, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 312, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we introduce GS-SLAM that first utilizes 3D Gaussian representation in the Simultaneous Localization and Mapping (SLAM) system. It facilitates a better balance between efficiency and accuracy. Compared to recent SLAM methods employing neural implicit representations, our method utilizes a real-time differentiable splatting rendering pipeline that offers significant speedup to map optimization and RGB-D rendering. Specifically, we propose an adaptive expansion strategy that adds new or deletes noisy 3D Gaussians in order to efficiently reconstruct new observed scene geometry and improve the mapping of previously observed areas. This strategy is essential to extend 3D Gaussian representation to reconstruct the whole scene rather than synthesize a static object in existing methods. Moreover, in the pose tracking process, an effective coarse-to-fine technique is designed to select reliable 3D Gaussian representations to optimize camera pose, resulting in runtime reduction and robust estimation. Our method achieves competitive performance compared with existing state-of-the-art real-time methods on the Replica, TUM-RGBD datasets. Project page: https://gs-slam.github.io/.", + "bbox": [ + 76, + 316, + 473, + 636 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 660, + 209, + 676 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Simultaneous localization and mapping (SLAM) has emerged as a pivotal technology in fields such as robotics [6], virtual reality [10], and augmented reality [25, 39]. The goal of SLAM is to construct a dense/sparse map of an unknown environment while simultaneously tracking the camera pose. Traditional SLAM methods employ point/surfel clouds [20, 32, 42, 46], mesh representations [26], voxel hashing [12, 18, 23] or voxel grids [21] as scene representations to construct dense mapping, and have made considerable progress on localization accuracy. However, these methods face serious challenges in obtaining fine-grained dense maps.", + "bbox": [ + 75, + 685, + 468, + 867 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/22fe11d95c4291fb95410f5b65e9fc36f5340c37a7d6b3c833bfdee4c95ddd87.jpg", + "image_caption": [ + "Figure 1. The illustration of the proposed GS-SLAM. It first utilizes the 3D Gaussian representation and differentiable splatting rasterization pipeline in SLAM, achieving real-time tracking and mapping performance on GPU. Besides, benefiting from the splatting rasterization pipeline, GS-SLAM achieves a $100\\times$ faster rendering FPS and more high-quality full image results than the other SOTA methods." + ], + "image_footnote": [], + "bbox": [ + 501, + 284, + 890, + 508 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, Neural Radiance Fields (NeRF) [19] have been explored to enhance SLAM methodologies and exhibit strengths in generating high-quality, dense maps with low memory consumption [35]. In particular, iMAP [35] uses a single multi-layer perceptron (MLP) to represent the entire scene, which is updated globally with the loss between volume-rendered RGB-D image and ground-truth observations. NICE-SLAM [55] utilizes a hierarchical neural implicit grid as scene map representation to allow local updates for reconstructing large scenes. Moreover, ESLAM [11], CoSLAM [41] and EN-SLAM [24] utilize axis-aligned feature planes and joint coordinate-parametric encoding to improve the capability of scene representation, achieving efficient and high-quality surface map reconstruction. In practical mapping and tracking steps, these methods only render a small set of pixels to reduce optimization time, which leads to the reconstructed dense maps lacking", + "bbox": [ + 496, + 643, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution: yanchi@pjlab.org.cn.", + "bbox": [ + 94, + 875, + 321, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding author: dongwang.dw93@gmail.com.", + "bbox": [ + 96, + 887, + 382, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19595", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the richness and intricacy of details. In essence, it is a tradeoff for the efficiency and accuracy of NeRF-based SLAM since obtaining high-resolution images with the ray-based volume rendering technique is time-consuming and unacceptable.", + "bbox": [ + 75, + 90, + 467, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Fortunately, recent work [13, 17, 47] with 3D Gaussian representation and tile-based splatting techniques has shown great superiority in the efficiency of high-resolution image rendering. It is applied to synthesize novel view RGB images of static objects, achieving state-of-the-art visual quality for $1080\\mathrm{p}$ resolution at real-time speed. Inspired by this, we extend the rendering superiority of 3D Gaussian scene representation and real-time differentiable splatting rendering pipeline for the task of dense RGB-D SLAM and manage to jointly promote the speed and accuracy of NeRF-based dense SLAM, as shown in Fig. 1.", + "bbox": [ + 75, + 166, + 467, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we propose GS-SLAM, the first RGB-D dense SLAM system that first utilizes 3D Gaussian scene representation coupled with the splatting rendering technique to better balance speed and accuracy. Our system optimizes camera tracking and mapping with a novel RGB-D rendering approach that processes 3D Gaussians quickly and accurately through sorting and $\\alpha$ -blending. We enhance scene reconstruction by introducing an adaptive strategy for managing 3D Gaussian elements, which optimizes mapping by focusing on current observations and minimizes errors in dense maps and images. Moreover, we propose a coarse-to-fine approach, starting with low-resolution image analysis for initial pose estimation and refining it with high-resolution rendering using select 3D Gaussians, to boost speed and accuracy. We perform extensive evaluations on a selection of indoor RGB-D datasets and demonstrate state-of-the-art performance on dense neural RGB-D SLAM in terms of tracking, rendering, and mapping. Overall, our contributions include:", + "bbox": [ + 75, + 332, + 467, + 618 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose GS-SLAM, the first 3D Gaussian Splatting(3DGS)-based dense RGB-D SLAM approach, which takes advantage of the fast splatting rendering technique to boost the mapping optimizing and pose tracking, achieving real-time and photo-realistic reconstruction performance.", + "- We present an adaptive 3D Gaussian expansion strategy to efficiently reconstruct new observed scene geometry and develop a coarse-to-fine technique to select reliable 3D Gaussians to improve camera pose estimation.", + "- Our approach achieves competitive performance on Replica and TUM-RGBD datasets in terms of tracking, and mapping and runs at 8.43 FPS, resulting in a better balance between efficiency and accuracy." + ], + "bbox": [ + 78, + 621, + 467, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 843, + 217, + 858 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Dense Visual SLAM. The existing real-time dense visual SLAM systems are typically based on discrete handcrafted", + "bbox": [ + 76, + 869, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "features or deep-learning embeddings, and follow the mapping and tracking architecture in [16]. DTAM [22] first introduces a dense SLAM system that uses photometric consistency to track a handheld camera and represent the scene as a cost volume. KinectFusion [44] performs camera tracking by iterative-closest-point and updates the scene via TSDF-Fusion. BAD-SLAM [29] proposes to jointly optimize the keyframe poses and 3D scene geometry via a direct bundle adjustment (BA) technique. In contrast, recent works integrate deep learning with the traditional geometry framework for more accurate and robust camera tracking and mapping, such as DROID-SLAM [37], CodeSLAM [1], SceneCode [54], and NodeSLAM [34], have made significant advances in the field, achieving more accurate and robust camera tracking and mapping performance.", + "bbox": [ + 496, + 90, + 890, + 318 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural Implicit Radiance Field based SLAM. For NeRF-based SLAM, existing methods can be divided into three main types: MLP-based methods, Hybrid representation methods, and Explicit methods. MLP-based method iMAP [35] offers scalable and memory-efficient map representations but faces challenges with catastrophic forgetting in larger scenes. Hybrid representation methods combine the advantages of implicit MLPs and structure features, significantly enhancing the scene scalability and precision. For example, NICE-SLAM [55] integrates MLPs with multi-resolution voxel grids, enabling large scene reconstruction, and Vox-Fusion [48] employs octree expansion for dynamic map scalability, while ESLAM [11] and Point-SLAM [27] utilize tri-planes and neural point clouds respectively to improve the mapping capability. As for the explicit method proposed in [38], it stores map features in voxel directly, without any MLPs, enabling faster optimization. Instead of representing maps with implicit features, GS-SLAM utilizes the 3D Gaussian representation, efficiently renders images using splatting-based rasterization, and optimizes parameters directly with backward propagation.", + "bbox": [ + 496, + 321, + 890, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Gaussian Representation. Several recent approaches have used 3D Gaussians for shape reconstruction, such as Fuzzy Metaballs [14, 15], VoGE [40], 3DGS [13]. Notably, 3DGS [13] demonstrates great superiorities in high-quality real-time novel-view synthesis. This work represents the scene with 3D Gaussians and develops a NeRF-style fast rendering algorithm to support anisotropic splating, achieving SOTA visual quality and fast high-resolution rendering performance. Beyond the rendering superiorities, Gaussian splating holds an explicit geometry scene structure and appearance, benefiting from the exact modeling of scenes representation [50]. This promising technology has been rapidly applied in several fields, including 3D generation [3, 36, 51], dynamic scene modeling [17][47][49], and photorealistic drivable avatars [56]. However, currently, there is no research addressing camera pose estimation or real-time mapping using 3D Gaussian models due to the in", + "bbox": [ + 496, + 643, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19596", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/53be036ebc19fd8fa446cd8da14749f0bad356f997a5cb2bd86001c01bec7a5d.jpg", + "image_caption": [ + "Figure 2. Overview of the proposed method. We aim to use 3D Gaussians to represent the scene and use the rendered RGB-D image for inverse camera tracking. GS-SLAM proposes a novel Gaussian expansion strategy to make the 3D Gaussian feasible to reconstruct the whole scene and can achieve real-time tracking, mapping, and rendering performance on GPU." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 893, + 236 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "herent limitations of the prime pipeline [13], i.e., prerequisites of initialized point clouds or camera pose inputs [28]. In contrast, we derive the analytical derivative equations for pose estimation in the Gaussian representation and implement efficient CUDA optimization.", + "bbox": [ + 75, + 308, + 470, + 383 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 400, + 210, + 417 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fig. 2 shows the overview of the proposed GS-SLAM. We aim to estimate the camera poses $\\{\\mathbf{P}_i\\}_{i=1}^N$ of every frame and simultaneously reconstruct a dense scene map by giving an input sequential RGB-D stream $\\{\\mathbf{I}_i, \\mathbf{D}_i\\}_{i=1}^M$ with known camera intrinsic $\\mathbf{K} \\in \\mathbb{R}^{3 \\times 3}$ . In Sec. 3.1, we first introduce 3D Gaussian as the scene representation $\\mathbf{S}$ and the RGB-D render by differentiable splattering rasterization. With the estimated camera pose of the keyframe, in Sec. 3.2, an adaptive expansion strategy is proposed to add new or delete noisy 3D Gaussians to efficiently reconstruct new observed scene geometry while improving the mapping of the previously observed areas. For camera tracking of every input frame, we derive an analytical formula for backward optimization with rendering RGB-D loss. We further introduce an effective coarse-to-fine technique to minimize rendering losses to achieve efficient and accurate pose estimation in Sec. 3.3.", + "bbox": [ + 75, + 426, + 473, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. 3D Gaussian Scene Representation", + "text_level": 1, + "bbox": [ + 76, + 696, + 382, + 713 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to optimize a scene representation that captures the geometry and appearance of the scene, resulting in a detailed dense map and high-quality novel view synthesis. To do this, we model the scene as a set of 3D Gaussians coupled with opacity and spherical harmonics", + "bbox": [ + 75, + 720, + 468, + 796 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {G} = \\left\\{G _ {i}: \\left(\\mathbf {X} _ {i}, \\boldsymbol {\\Sigma} _ {i}, \\Lambda_ {i}, \\mathbf {Y} _ {i}\\right) | i = 1, \\dots , N \\right\\}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 809, + 468, + 825 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Each 3D Gaussian scene representation $G_{i}$ is defined by position $\\mathbf{X}_i\\in \\mathbb{R}^3$ , 3D covariance matrix $\\pmb{\\Sigma}_{i}\\in \\mathbb{R}^{3\\times 3}$ , opacity $\\Lambda_{i}\\in \\mathbb{R}$ and 1-degree spherical harmonics $(\\mathbf{Y})$ per color channel, a total of 12 coefficients for $\\mathbf{Y}_i\\in \\mathbb{R}^{12}$ . To reduce", + "bbox": [ + 76, + 839, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the learning difficulty of the 3D Gaussians [57], we parameterize the 3D Gaussian's covariance as:", + "bbox": [ + 498, + 308, + 890, + 338 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Sigma} = \\mathbf {R} \\mathbf {S} \\mathbf {S} ^ {T} \\mathbf {R} ^ {T}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 638, + 347, + 890, + 367 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{S} \\in \\mathbb{R}^3$ is a 3D scale vector, $\\mathbf{R} \\in \\mathbb{R}^{3 \\times 3}$ is the rotation matrix, storing as a 4D quaternion.", + "bbox": [ + 498, + 375, + 890, + 406 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Color and depth splatting rendering. With the optimized 3D Gaussian scene representation parameters, given the camera pose $\\mathbf{P} = \\{\\mathbf{R},\\mathbf{t}\\}$ , the 3D Gaussians $G$ are projected into 2D image plane for rendering with:", + "bbox": [ + 498, + 407, + 890, + 468 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Sigma} ^ {\\prime} = \\mathbf {J} \\mathbf {P} ^ {- 1} \\boldsymbol {\\Sigma} \\mathbf {P} ^ {- T} \\mathbf {J} ^ {T}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 476, + 890, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{J}$ is the Jacobian of the affine approximation of the projective function. After projecting 3D Gaussians to the image plane, the color of one pixel is rendered by sorting the Gaussians in depth order and performing front-to-back $\\alpha$ -blending rendering as follows:", + "bbox": [ + 496, + 505, + 893, + 580 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {C}} = \\sum_ {i \\in N} \\mathbf {c} _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 590, + 890, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{c}_i$ represents the color of the $i$ -th 3D Gaussian obtained by learned spherical harmonics coefficients $\\mathbf{Y}$ , $\\alpha_i$ is the density computed by learned opacity $\\Lambda_i$ and 2D Gaussian with covariance $\\Sigma'$ . Similarly, the depth is rendered by", + "bbox": [ + 496, + 643, + 893, + 719 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {D} = \\sum_ {i \\in N} d _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 729, + 890, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $d_{i}$ denotes the depth of the center of the $i$ -th 3D Gaussian, which is obtained by projecting to $z$ -axis in the camera coordinate.", + "bbox": [ + 498, + 777, + 890, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Adaptive 3D Gaussian Expanding Mapping", + "text_level": 1, + "bbox": [ + 498, + 832, + 870, + 849 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The 3D Gaussian scene representations are updated and optimized on each selected keyframe for stable mapping. Given the estimated pose of each selected keyframe, we first", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19597", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "apply the proposed adaptive expansion strategy to add new or delete noisy 3D Gaussians from the whole scene representations to render RGB-D images with resolution $H \\times W$ , and then the updated 3D Gaussian scene representations are optimized by minimizing the geometric depth loss $\\mathcal{L}_d$ and the photometric color loss $\\mathcal{L}_{\\mathrm{c}}$ to the sensor observation depth $D$ and color $\\mathbf{C}$ ,", + "bbox": [ + 75, + 90, + 472, + 196 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathbf {c}} = \\sum_ {m = 1} ^ {H W} \\left| \\mathbf {C} _ {m} - \\hat {\\mathbf {C}} _ {m} \\right|, \\quad \\mathcal {L} _ {d} = \\sum_ {m = 1} ^ {H W} \\left| D _ {m} - \\hat {D} _ {m} \\right|. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 208, + 468, + 251 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The loss optimizes the parameters of all 3D Gaussians that contribute to the rendering of these keyframe images.", + "bbox": [ + 75, + 261, + 468, + 291 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Adaptive 3D Gaussian Expansion Strategy. At the first frame of the RGB-D sequence, we first uniformly sample half pixels from a whole image with $H \\times W$ resolution and back-projecting them into 3D points $\\mathbf{X}$ with corresponding depth observation $D$ . The 3D Gaussian scene representations are created by setting position as $\\mathbf{X}$ and initializing zero degree Spherical Harmonics coefficients with RGB color $\\mathbf{C}_i$ . The opacities are set to pre-defined values, and the covariance is set depending on the spatial point density, i.e.,", + "bbox": [ + 75, + 292, + 470, + 429 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{G _ {i} = \\left(\\mathbf {P} _ {i}, \\boldsymbol {\\Sigma} _ {\\text {i n i t}}, \\Lambda_ {\\text {i n i t}}, \\mathbf {C} _ {i}\\right) | i = 1, \\dots , M \\right\\}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 441, + 468, + 459 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $M$ equals to $HW / 2$ . The 3D Gaussians are initialized and then optimized using the first RGB-D image with rendering loss. Note that only half of the pixels are used to initialize the scene, leaving space to conduct adaptive density control of Gaussians that splits large points into smaller ones and clones them with different directions to capture missing geometric details.", + "bbox": [ + 75, + 470, + 468, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Adding step: To obtain a complete map of the environment, the 3D Gaussian scene representations should be able to model the geometry and appearance of newly observed areas. Specifically, at every keyframe, we add first rendered RGB-D images using historical 3D Gaussians and calculate cumulative opacity $T = \\sum_{i\\in N}\\alpha_i\\prod_{j = 1}^{i - 1}(1 - \\alpha_j)$ for each pixel. We label one pixel as un-reliable $x^{un}$ if its cumulative opacity $T$ is too low or its rendered depth $\\hat{D}$ is far away from observed depth $D$ , i.e.,", + "bbox": [ + 75, + 577, + 468, + 713 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nT < \\tau_ {T} \\quad \\text {o r} \\quad | D - \\hat {D} | > \\tau_ {D}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 724, + 468, + 742 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "These selected un-reliable pixels mostly capture new observed areas. Then we back-project these un-reliable pixels to 3D points $\\mathbf{P}^{un}$ , and a set of new 3D Gaussians at $\\mathbf{P}^{un}$ initialized as Eq. 7 are added into scene representations to model the new observed areas.", + "bbox": [ + 75, + 750, + 468, + 823 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "**Deleting step:** As shown in Fig. 3, there are some floating 3D Gaussians due to the unstable adaptive control of Gaussians after optimization with Eq. 6. These floating 3D Gaussians will result in a low-quality dense map and a rendered image containing lots of artifacts. To address this", + "bbox": [ + 75, + 824, + 470, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d4d5a6ae1d5f38d2d2e048d028616ebf9d872090c545a009861906a8433fbe83.jpg", + "image_caption": [ + "Figure 3. Illustration of the proposed adaptive 3D Gaussian expansion strategy. GS-SLAM inhibits the low-quality 3D Gaussian floaters in the current frustum according to depth." + ], + "image_footnote": [], + "bbox": [ + 521, + 84, + 874, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "issue, after adding new 3D Gaussians, we check all visible 3D Gaussians in the current camera frustum and significantly decrease opacity $\\Lambda_{i}$ of 3D Gaussians whose position is not near the scene surfaces. Formally, for each visible 3D Gaussian, we draw a ray $\\mathbf{r}(t)$ from camera origin $\\mathbf{o}$ and its position $\\mathbf{X}_i = (x_i,y_i,z_i)$ , i.e., $\\mathbf{r}(t) = \\mathbf{o} + t(\\mathbf{X}_i - \\mathbf{o})$ . Then, we can find a pixel with coordinate $(u,v)$ where this ray intersects the image plane and corresponding depth observation $D$ . The 3D Gaussians are deleted by degenerating its opacity as follows:", + "bbox": [ + 498, + 316, + 890, + 468 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nG _ {i}: \\Lambda_ {i} \\Rightarrow G _ {i}: \\eta \\Lambda_ {i}, \\text {i f} D - d i s t (\\mathbf {X} _ {i}, \\mathbf {P} _ {u v}) > \\gamma , \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 483, + 890, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $P_{uv}$ is the world coordinates of the intersected pixel calculated with the camera intrinsic and extrinsic. $dist(\\cdot, \\cdot)$ is the Euclidean distance, and $\\eta$ (much smaller than 1) and $\\gamma$ are the hyper-parameters. Note that we decrease the opacity of floating 3D Gaussians in front of the scene surfaces to make our newly added 3D Gaussians well-optimized.", + "bbox": [ + 498, + 507, + 890, + 598 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Tracking and Bundle Adjustment", + "text_level": 1, + "bbox": [ + 500, + 609, + 794, + 626 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the parallel camera tracking phase of our work, we first employ a common straightforward constant velocity assumption to initialize new poses. This assumption transforms the last known pose based on the relative transformation between the second-to-last pose and the last pose. Then, the accurate camera pose $\\mathbf{P}$ is optimized by minimizing rendered color loss, i.e.,", + "bbox": [ + 498, + 633, + 890, + 741 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t r a c k}} = \\sum_ {m = 1} ^ {M} \\left| \\mathbf {C} _ {m} - \\hat {\\mathbf {C}} _ {m} \\right| _ {1}, \\quad \\min _ {\\mathbf {R}, \\mathbf {t}} (\\mathcal {L} _ {\\text {t r a c k}}), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 753, + 890, + 795 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $M$ is the number of sampled pixels for rendering.", + "bbox": [ + 500, + 806, + 870, + 823 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Differentiable pose estimation. According to Eqs. (3) and (4), we observe that the gradient of the camera pose $\\mathbf{P}$ is related to three intermediate variables: $\\pmb{\\Sigma}^{\\prime}$ , $\\mathbf{c}_i$ , and the projected coordinate $\\mathbf{m}_i$ of Gaussian $G_{i}$ . By applying the chain rule of derivation, we obtain the analytical formula", + "bbox": [ + 498, + 824, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19598", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tion of camera pose P:", + "bbox": [ + 76, + 90, + 230, + 104 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\frac {\\partial \\mathcal {L} _ {\\mathbf {c}}}{\\partial \\mathbf {P}} = \\frac {\\partial \\mathcal {L} _ {\\mathbf {c}}}{\\partial \\mathbf {C}} \\frac {\\partial \\mathbf {C}}{\\partial \\mathbf {P}} = \\frac {\\partial \\mathcal {L} _ {\\mathbf {c}}}{\\partial \\mathbf {C}} \\left(\\frac {\\partial \\mathbf {C}}{\\partial \\mathbf {c} _ {i}} \\frac {\\partial \\mathbf {c} _ {i}}{\\partial \\mathbf {P}} + \\frac {\\partial \\mathbf {C}}{\\partial \\alpha_ {i}} \\frac {\\partial \\alpha_ {i}}{\\partial \\mathbf {P}}\\right) \\\\ = \\frac {\\partial \\mathcal {L} _ {\\mathrm {c}}}{\\partial \\overline {{\\mathbf {C}}}} \\frac {\\partial \\mathbf {C}}{\\partial \\alpha_ {i}} \\left(\\frac {\\partial \\alpha_ {i}}{\\partial \\boldsymbol {\\Sigma} ^ {\\prime}} \\frac {\\partial \\boldsymbol {\\Sigma} ^ {\\prime}}{\\partial \\mathbf {P}} + \\frac {\\partial \\alpha_ {i}}{\\partial \\mathbf {m} _ {\\mathrm {i}}} \\frac {\\partial \\mathbf {m} _ {\\mathrm {i}}}{\\partial \\mathbf {P}}\\right) \\tag {11} \\\\ = \\frac {\\partial \\mathcal {L} _ {\\mathbf {c}}}{\\partial \\mathbf {C}} \\frac {\\partial \\mathbf {C}}{\\partial \\alpha_ {i}} \\left(\\frac {\\partial \\alpha_ {i}}{\\partial \\boldsymbol {\\Sigma} ^ {\\prime}} \\frac {\\partial (\\mathbf {J P} ^ {- 1} \\boldsymbol {\\Sigma} \\mathbf {P} ^ {- T} \\mathbf {J} ^ {T})}{\\partial \\mathbf {P}} + \\frac {\\partial \\alpha_ {i}}{\\partial \\mathbf {m} _ {\\mathbf {i}}} \\frac {\\partial (\\mathbf {K P X} _ {i})}{\\partial \\mathbf {P} d _ {i}}\\right) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 116, + 468, + 202 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $d_{i}$ denotes the $z$ -axis coordinate of projection $\\mathbf{m}_i$ . The item $\\frac{\\partial\\mathbf{C}}{\\partial\\mathbf{c}_i}\\frac{\\partial\\mathbf{c}_i}{\\partial\\mathbf{P}}$ can be eliminated because we are only concerned about the view-independent color in our tracking implementation. In addition, we find that the intermediate gradient $\\frac{\\partial(\\mathbf{KP}\\mathbf{X}_i)}{\\partial\\mathbf{P}d_i}$ is the deterministic component for the camera pose $\\mathbf{P}$ . So we simply ignore the backpropagation of $\\frac{\\partial(\\mathbf{JP}^{-1}\\pmb{\\Sigma}\\mathbf{P}^{-T}\\mathbf{J}^{T})}{\\partial\\mathbf{P}}$ for efficiency. More details can be found in the supplemental materials.", + "bbox": [ + 76, + 212, + 467, + 337 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Coarse-to-fine camera tracking. It would be problematic to optimize the camera pose with all image pixels since artifacts in images can cause drifted camera tracking. To address this issue, as shown in Fig. 2, in the differentiable pose estimation step for each frame, we first take advantage of image regularity to render only a sparse set of pixels and optimize tracking loss to obtain a coarse camera pose. This coarse optimization step significantly eases the influence of detailed artifacts. Further, we use this coarse camera pose and depth observation to select reliable 3D Gaussians, which guides GS-SLAM to render informative areas with clear geometric structures to refine coarse camera pose via further optimizing tracking loss on new rendering pixels.", + "bbox": [ + 75, + 339, + 467, + 535 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, in the coarse stage, we first render a coarse image $\\hat{\\mathbf{I}}_c$ with resolution $H / 2\\times W / 2$ at uniformly sampled image coordinates and optimize tracking loss in Eq. 10 for $T_{c}$ iterations, and the obtained camera pose is denoted as $\\mathbf{P}_c$ . In the fine stage, we use a similar technique with adaptive 3D Gaussian expansion strategy in Section 3.2 to select reliable 3D Gaussian to render full-resolution images while ignoring noisy 3D Gaussians that cause artifacts. In detail, we check all visible 3D Gaussians under coarse camera pose $\\mathbf{P}_c$ , and remove 3D Gaussians whose position is far away from the scene surface. Formally, for each visible 3D Gaussians $G_{i}$ with position $\\mathbf{X}_i$ , we project it to the camera plane using coarse camera pose $\\mathbf{P}_c$ and camera intrinsic. Given the projected pixel's depth observation $D_{i}$ and the distance $d_{i}$ that is between 3D Gaussians $G_{i}$ and the camera image plane, the reliable 3D Gaussians are selected as follows:", + "bbox": [ + 75, + 536, + 467, + 777 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {G} _ {\\text {s e l e c t e d}} = \\left\\{G _ {i} \\mid G _ {i} \\in \\mathbf {G} \\text {a n d} a b s \\left(D _ {i} - d _ {i}\\right) \\leq \\varepsilon \\right\\}, \\tag {12} \\\\ \\tilde {\\mathbf {I}} _ {f} = \\mathcal {F} (u, v, \\mathbf {G} _ {s e l e c t e d}), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 787, + 468, + 827 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where we use the selected reliable 3D Gaussians to render full-resolution images $\\hat{\\mathbf{I}}_f$ . $u, v$ denote the pixel coordinates in $\\hat{\\mathbf{I}}_f$ , and $\\mathcal{F}$ represents the color splattering rendering function. The final camera pose $\\mathbf{P}$ is obtained by optimizing", + "bbox": [ + 76, + 838, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tracking loss in Eq. 10 with $\\hat{\\mathbf{I}}_f$ for other $T_{f}$ iterations. Note that $\\hat{\\mathbf{I}}_c$ and $\\hat{\\mathbf{I}}_f$ are only rendered at previously observed areas, avoiding rendering areas where 3D scene representations have not been optimized in the mapping process. Also, we add keyframes based on the proportion of the currently observed image's reliable region to the overall image. At the same time, when the current tracking frame and most recent keyframe differ by more than a threshold value $\\mu_{k}$ , this frame will be inserted as a keyframe.", + "bbox": [ + 496, + 90, + 890, + 228 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "**Bundle adjustment.** In the bundle adjustment (BA) phase, we optimize the camera poses $\\mathbf{P}$ and the 3D Gaussian scene representation $\\mathbf{S}$ jointly. We randomly select $K$ keyframes from the keyframe database for optimization, using the loss function similar to the mapping part. For pose optimization stability, we only optimize the scene representation $\\mathbf{S}$ in the first half of the iterations. In the other half of the iterations, we simultaneously optimize the map and the poses. Then, the accurate camera pose $\\mathbf{P}$ is optimized by minimizing rendering color loss, i.e.,", + "bbox": [ + 496, + 229, + 890, + 380 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {b a} = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\sum_ {m = 1} ^ {H W} \\left| D _ {m} - \\hat {D} _ {m} \\right| _ {1} + \\lambda_ {m} \\left| \\mathbf {C} _ {m} - \\hat {\\mathbf {C}} _ {m} \\right| _ {1}, \\min _ {\\mathbf {R}, \\mathbf {t}, \\mathbf {S}} (\\mathcal {L} _ {b a}). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 388, + 890, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 500, + 430, + 624, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 500, + 454, + 689, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. To evaluate the performance of GS-SLAM, we conduct experiments on the Replica [31], and TUM-RGBD [33]. Following [11, 27, 41, 48, 55], we use 8 scenes from the Replica dataset for localization, mesh reconstruction, and rendering quality comparison. The selected three subsets of TUM-RGBD datasets are used for localization.", + "bbox": [ + 496, + 477, + 890, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We compare our method with existing SOTA NeRF-based dense visual SLAM: NICE-SLAM [55], VoxFusion [48], CoSLAM [41], ESLAM [11] and Point-SLAM [27]. The rendering performance of CoSLAM [41] and ESLAM [11] is conducted from the open source code with the same configuration in [27].", + "bbox": [ + 496, + 568, + 890, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metric. For mesh reconstruction, we use the 2D Depth L1 (cm) [55], the Precision $(\\mathbb{P},\\%)$ , Recall $(\\mathbb{R},\\%)$ , and F-score with a threshold of $1\\mathrm{cm}$ to measure the scene geometry. For localization, we use the absolute trajectory (ATE, cm) error [33] to measure the accuracy of the estimated camera poses. We further evaluate the rendering performance using the peak signal-to-noise ratio (PSNR), SSIM [43], and LPIPS [52] by following [27]. To be fair, we run all the methods on a dataset 10 times and report the average results. More details can be found in the supplemental materials.", + "bbox": [ + 496, + 659, + 890, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. GS-SLAM is implemented in Python using the PyTorch framework, incorporating CUDA code for Gaussian splatting and trained on a desktop PC with a 5.50GHz Intel Core i9-13900K CPU and NVIDIA RTX 4090 GPU. We extended the existing code for differentiable Gaussian splatting rasterization with additional func", + "bbox": [ + 496, + 810, + 890, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19599", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tionality for handling depth, pose, and cumulative opacity during both forward and backward propagation. More details can be found in the supplemental materials.", + "bbox": [ + 75, + 90, + 468, + 137 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Evaluation of Localization and Mapping", + "text_level": 1, + "bbox": [ + 76, + 146, + 423, + 162 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation on Replica. Tracking ATE: Tab. 1 illustrates the tracking performance of our method and the state-of-the-art methods on the Replica dataset. Our method achieves the best or second performance in 7 of 8 scenes and outperforms the second-best method Point-SLAM [27] by $0.4\\mathrm{cm}$ on average at 8.34 FPS. It is noticeable that the second best method, Point-SLAM [27] runs at 0.42 FPS, which is $20\\times$ slower than our method, indicating that GS-SLAM achieves a better trade-off between the tracking accuracy and the runtime efficiency. Mapping ACC: Tab. 3 report the mapping evaluation results of our method with other current state-of-the-art visual SLAM methods. GS-SLAM achieves the best performance in Depth L1 (1.16cm) and Precision $(74.0\\%)$ metrics on average. For Recall and F1 scores, GS-SLAM performs comparably to the second best method CoSLAM [41]. The visualization results in Fig. 4 show that GS-SLAM achieves satisfying construction mesh with clear boundaries and details.", + "bbox": [ + 75, + 170, + 468, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation on TUM-RGBD. Tab. 2 compares GS-SLAM with the other SLAM systems in TUM-RGBD dataset. Our method surpasses iMAP [35], NICE-SLAM [55] and Voxfusion [48], and achieves a comparable performance, average $3.7\\mathrm{cm}$ ATE RSME, with the SOTA methods. A gap to traditional methods still exists between the neural vSLAM and the traditional SLAM systems, which employ more sophisticated tracking schemes [27].", + "bbox": [ + 75, + 444, + 468, + 565 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Tracking comparison (ATE RMSE [cm]) of the proposed method vs. the SOTA methods on the Replica dataset. The running speed of methods in the upper part is lower than 5 FPS, * denotes the reproduced results by running officially released code.", + "bbox": [ + 75, + 570, + 468, + 626 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ebaacfcd04b4e7c1c8485faeb94dabce369a34c211e452d8eb12edcea2513134.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodRm0Rm1Rm2Off0Off1Off2Off3Off4avg
Point-SLAM [27]0.560.470.300.350.620.550.720.730.54
NICE-SLAM [55]0.971.311.070.881.001.061.101.131.06
Vox-Fusion* [48]1.374.701.478.482.042.581.112.943.09
ESLAM [11]0.710.700.520.570.550.580.720.630.63
CoSLAM [41]0.700.951.350.590.552.031.560.721.00
Ours0.480.530.330.520.410.590.460.70.50
", + "bbox": [ + 78, + 628, + 468, + 698 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d8143725206c2ca42e876e971b760f321256520e7b29be9c0fbcdc030081115c.jpg", + "table_caption": [ + "Table 2. Tracking ATE [cm] on TUM-RGBD [33]. Our method achieves a comparable performance among the neural vSLAMs. * denotes the reproduced results by running officially released code." + ], + "table_footnote": [], + "table_body": "
Methodfr1Deskfr2_xyzfr3_offAvg.Methodfr1Deskfr2_xyzfr3_offAvg.
DI-Fusion [9]4.42.05.84.1NICE-SLAM [55]4.331.73.913.3
Elastic-Fusion [46]2.51.22.52.1Vox-Fusion* [48]3.51.526.010.3
BAD-SLAM [30]1.71.11.71.5CoSLAM [41]2.71.92.62.4
Kintinous [45]3.72.93.03.2ESLAM [11]2.31.12.42.0
ORB-SLAM2 [20]1.60.41.01.0Point-SLAM2.61.33.22.4
iMAP* [35]7.22.19.06.1Ours3.31.36.63.7
", + "bbox": [ + 78, + 753, + 467, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Rendering Evaluation", + "text_level": 1, + "bbox": [ + 76, + 830, + 282, + 847 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare the rendering performance of the proposed GS-SLAM with the neural visual SLAM methods in Tab. 6. The results show that GS-SLAM achieves the best performance", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d37d4a79788887ecfeb1728517e76e4b8abb690ff93cac0de321fabed79cb399.jpg", + "table_caption": [ + "Table 3. Reconstruction comparison of the proposed method vs. the SOTA methods on Replica dataset." + ], + "table_footnote": [], + "table_body": "
MethodMetricRm 0Rm 1Rm 2Off 0Off 1Off 2Off 3Off 4Avg.
NICESL AM [55]Depth L1 ↓1.811.442.041.391.768.334.992.012.97
Precision ↑45.8643.7644.3851.4050.8038.3740.8537.3544.10
Recall↑44.1046.1242.7848.6653.0839.9839.0435.7743.69
F1↑44.9644.8443.5649.9951.9139.1639.9236.5443.86
VoxFusion [48]Depth L1 ↓1.091.902.212.323.404.192.961.612.46
Precision↑75.8335.8863.1048.5143.5054.4869.1155.4055.73
Recall↑64.8933.0756.6244.7638.4447.8560.6146.7949.13
F1↑69.9334.3859.6746.5440.8150.9564.5650.7252.20
CoSLAM [41]Depth L1 ↓0.990.822.281.241.617.704.651.432.59
Precision↑81.7177.9573.3079.4180.6755.6457.6379.7673.26
Recall↑74.0370.7965.7371.4670.3552.9656.0671.2266.58
F1↑77.6874.2069.3175.2375.1654.2756.8375.2569.74
ESL AM [11]Depth L1 ↓0.630.620.980.571.667.323.940.882.08
Precision↑74.3375.9482.4872.2065.7470.7372.4872.2473.27
Recall↑87.3787.0184.9988.3684.3881.9279.1880.6384.23
F1↑80.3281.1083.7279.4773.9075.9275.6876.2178.29
OursDepth L1 ↓1.310.821.260.810.961.411.531.081.16
Precision↑64.5883.1170.1383.4387.7770.9163.1868.8874.00
Recall↑61.2976.8363.8476.9076.1561.6362.9161.5067.63
F1↑62.8979.8566.8480.0381.5565.9559.1764.9870.15
", + "bbox": [ + 508, + 123, + 890, + 342 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "in all the metrics. Our method significantly outperforms the second-best methods CoSLAM [41], ESLAM [11] and NICE-SLAM [55] by 1.52 dB in PSNR, 0.027 in SSIM and 0.12 in LPIPS, respectively. It is noticeable that GS-SLAM achieves 386 FPS rendering speed on average, which is $100 \\times$ faster than the second-best method Vox-Fusion [48]. This excellent rendering performance is attributed to the efficient 3D Gaussian rendering pipeline and can be further applied to real-time downstream tasks, such as VR [5], robot navigation [7] and autonomous driving [2]. The visualization results in Fig. 5 show that GS-SLAM can generate much more high-quality and realistic images than the other methods, especially in edge areas with detailed structures. While NICE-SLAM [55] causes severe artifacts and blurs, CoSLAM [41] and ESLAM [11] generate blur around the image boundaries.", + "bbox": [ + 496, + 363, + 890, + 604 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fec0ca7d56512019e46711d86a69001801b6665ef0be59affaa59b06322d9b83.jpg", + "image_caption": [ + "Figure 4. Reconstruction performance comparison of the proposed GS-SLAM and SOTA methods on the Replica dataset." + ], + "image_footnote": [], + "bbox": [ + 501, + 617, + 888, + 869 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19600", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fdcd81bd916aabd7dbf1de73f9b520bda2fa382663cb7aaaf6a756593cc4751d.jpg", + "image_caption": [ + "Figure 5. The render visualization results on the Replica dataset of the proposed GS-SLAM and SOTA methods. GS-SLAM can generate much more high-quality and realistic images than the other methods, especially around the object boundaries." + ], + "image_footnote": [], + "bbox": [ + 76, + 85, + 893, + 383 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Runtime Analysis", + "text_level": 1, + "bbox": [ + 76, + 421, + 250, + 436 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 4 and Tab. 5 illustrate the runtime and memory usage of GS-SLAM and the state-of-the-art methods on the Replica and TUM-RGBD, respectively. We report the parameters of the neural networks and the memory usage of the scene representation. Note that Point-SLAM uses extra memory dynamic radius to improve performance (mark as $\\dagger$ ). The results show that GS-SLAM achieves a competitive running speed with 8.34 FPS compared to the other Radiance Fields-based vSLAMs. Note that we do not use any neural network decoder in our system, which results in the zero learnable parameter. However, the 3D Gaussian scene representations of GS-SLAM consume 198.04 MB memory, $4\\times$ larger than the second large method NICE-SLAM [55]. Memory usage is mainly caused by spherical harmonic coefficients in training, which is a common constraint among Gaussian splatting-based reconstruction methods. Despite this, we still achieve a $20\\times$ faster FPS compared to the similar point-based method Point-SLAM [27]. Besides, we also provide a light version of GS-SLAM with zero-order spherical harmonic coefficients, significantly reducing memory usage while maintaining stable performance.", + "bbox": [ + 75, + 444, + 472, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 771, + 230, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We perform the ablation of GS-SLAM on the Replica dataset #Room0 subset to evaluate the effectiveness of coarse-to-fine tracking, and expansion mapping strategy.", + "bbox": [ + 75, + 794, + 468, + 840 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of our expansion strategy for mapping. Tab. 7 shows the ablation of our proposed expansion strategy for mapping. The results illustrate that the expansion strategy can significantly improve the tracking and mapping per", + "bbox": [ + 75, + 839, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f8e304c563eaade375dcbd49ed189484cdc240d3ccf7e6140946a15b92b24604.jpg", + "table_caption": [ + "Table 4. Runtime and memory usage on Replica #Room0. The decoder parameters and embedding denote the parameter number of MLPs and the memory usage of the scene representation." + ], + "table_footnote": [], + "table_body": "
MethodTracking [ms×it] ↓Mapping [ms×it] ↓System FPS ↑Decoder param ↓Scene Embedding↓
Point-SLAM [27]0.06 × 4034.81 × 3000.420.127 M55.42 (+12453.2)†MB
NICE-SLAM [55]6.64 × 1028.63 × 602.910.06 M48.48 MB
Vox-Fusion [48]0.03 × 3066.53 × 101.280.054 M1.49 MB
CoSLAM [55]6.01 × 1013.18 × 1016.641.671 M—
ESLAM [11]6.85 × 819.87 × 1513.420.003 M27.12 MB
GS-SLAM11.9 × 1012.8 × 1008.340 M198.04 MB
", + "bbox": [ + 503, + 468, + 890, + 555 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/96303eb6c0537cdc8fabb06d7fbe1332a903fc5d8086c07a0b59d0fc0d4b708a.jpg", + "table_caption": [ + "Table 5. Runtime and memory usage on TUM-RGBD dataset #fr1Desk and #fr2_xyz." + ], + "table_footnote": [], + "table_body": "
Method#fr1Desk#fr2.xyz
FPS ↑Memory ↓FPS ↑Memory ↓
Point-SLAM0.1018.3 (+160.7)†MB0.1214.2 (+7687.4)†MB
NICE-SLAM [55]0.11178.8MB0.12484.0MB
ESLAM [11]0.3127.2MB0.3151.6MB
GS-SLAM1.83(ATE:3.3)40.8MB1.51(ATE:1.3)48.4MB
GS-SLAM (light)1.92(ATE:4.3)18.8MB1.68(ATE:2.7)22.3MB
", + "bbox": [ + 503, + 592, + 890, + 664 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "formance. The implementation w/o adding means that we only initialize 3D Gaussians in the first frame and optimize the scene without adding new points. However, this strategy completely crashes because the density control in [13] can not handle real-time mapping tasks without an accurate point cloud input. Besides, the implementation w/o deletion suffers from a large number of redundant and noisy 3D Gaussian, which causes undesirable supervision. In contrast, the proposed expansion strategy effectively improves the tracking and mapping performance by 0.1 in ATE and 11.97 in Recall by adding more accurate constraints for the optimization. According to the visualization results in Fig. 6, our full implementation achieves more high-quality and detailed rendering and reconstruction results than the w/o delete strategy.", + "bbox": [ + 496, + 674, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19601", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e4de8bbd01b3a458bb8d5cf6142d9bfff8b35bb3cd95e372d5dd9fe42c1a393f.jpg", + "table_caption": [ + "Table 6. Rendering performance on Replica dataset. We outperform existing dense neural RGB-D methods on the commonly reported rendering metrics. Note that GS-SLAM achieves 386 FPS on average, benefiting from the efficient Gaussian scene representation." + ], + "table_footnote": [], + "table_body": "
MethodMetricRoom 0Room 1Room 2Office 0Office 1Office 2Office 3Office 4Avg.FPS.
NICE-SLAM [55]PSNR [dB] ↑22.1222.4724.5229.0730.3419.6622.2324.9424.42
SSIM ↑0.6890.7570.8140.8740.8860.7970.8010.8560.8090.30
LPIPS ↓0.3300.2710.2080.2290.1810.2350.2090.1980.233
Vox-Fusion* [48]PSNR [dB] ↑22.3922.3623.9227.7929.8320.3323.4725.2124.413.88
SSIM ↑0.6830.7510.7980.8570.8760.7940.8030.8470.801
LPIPS ↓0.3030.2690.2340.2410.1840.2430.2130.1990.236
CoSLAM [41]PSNR [dB] ↑27.2728.4529.0634.1434.8728.4328.7630.9130.243.68
SSIM ↑0.9100.9090.9320.9610.9690.9380.9410.9550.939
LPIPS ↓0.3240.2940.2660.2090.1960.2580.2290.2360.252
ESLAM [11]PSNR [dB] ↑25.3227.7729.0833.7130.2028.0928.7729.7129.082.82
SSIM ↑0.8750.9020.9320.9600.9230.9430.9480.9450.929
LPIPS ↓0.3130.2980.2480.1840.2280.2410.1960.2040.336
OursPSNR [dB] ↑31.5632.8632.5938.7041.1732.3632.0332.9234.27386.91
SSIM ↑0.9680.9730.9710.9860.9930.9780.9700.9680.975
LPIPS ↓0.0940.0750.0930.0500.0330.0940.1100.1120.082
", + "bbox": [ + 76, + 99, + 893, + 297 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/060edc2b93575b53a5c40e29b19007b057268823217e8702d5f14500fd7c8567.jpg", + "table_caption": [ + "Table 7. Ablation of the adaptive 3D Gaussian expansion strategy on Replica #Room0." + ], + "table_footnote": [], + "table_body": "
Setting#Room0
ATE↓Depth L1↓Precision↑Recall ↑F1↑PSNR↑SSIM↑LPIPS↓
w/o addXXXXXXXX
w/o delete0.581.6853.5549.3251.3531.220.9670.094
w/ add & delete0.481.3164.5861.2962.8931.560.9680.094
", + "bbox": [ + 91, + 329, + 455, + 388 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8638d8a9fdd7b756aed4e5f8639dace2d83ce64aa5d26879543034ead65db6e6.jpg", + "table_caption": [ + "Table 8. Ablation of the coarse-to-fine tracking strategy on Replica #Room0." + ], + "table_footnote": [], + "table_body": "
Setting#Room0
ATE↓Depth L1↓Precision↑Recall ↑F1↑PSNR↑SSIM↑LPIPS↓
Coarse0.911.4859.6857.5456.5029.130.9540.120
Fine0.491.3962.6159.1861.2930.840.9640.096
Coarse-to-fine0.481.3164.5861.2962.8931.560.9680.094
", + "bbox": [ + 91, + 417, + 455, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of coarse-to-fine tracking. According to the results in Tab. 8, the proposed coarse-to-fine tracking strategy performs best in all tracking, mapping, and rendering metrics. Compared with fine tracking, the coarse-to-fine tracking strategy significantly improves the performance by 0.01 in tracking ATE, 2.11 in Recall, and 0.72 in PSNR. Although the fine strategy surpasses the coarse strategy in precision, it suffers from the artifacts and noise in the reconstructed scene, leading to a fluctuation optimization. The coarse-to-fine strategy effectively avoids noise reconstruction and improves accuracy and robustness.", + "bbox": [ + 75, + 479, + 468, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Efficiency-to-accuracy trade-off.", + "text_level": 1, + "bbox": [ + 76, + 656, + 361, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3DGS-based SLAM trade-off focuses not only on the efficiency of tracking and mapping but also emphasizes high-quality ultra-real-time rendering. As shown in Fig. 7b, GS-SLAM achieves $\\approx 400$ FPS ultra-fast speed and highest PSNR in map rendering. At the same time, our method remains a competitive system FPS and lowest tracking ATE in Fig. 7a. Moreover, GS-SLAM shows great potential in memory reduction in Tab. 5, and comparable in mesh reconstruction. Note that baselines directly use 3DGS in Fig. 7, resulting in inferior performances.", + "bbox": [ + 75, + 679, + 468, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Limitations", + "text_level": 1, + "bbox": [ + 76, + 844, + 333, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduced GS-SLAM, a novel dense visual SLAM method leveraging 3D Gaussian Splatting for efficient map", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e6ec1207e2f1c7667e3b982ea4481f5291ef8ba4b49e1dce14416e5d3f772787.jpg", + "image_caption": [ + "Figure 6. Rendering and mesh visualization of the adaptive 3D Gaussian expansion ablation on Replica #Room0." + ], + "image_footnote": [], + "bbox": [ + 522, + 300, + 697, + 455 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/089e70d5edcd2ffc6b016ee37ec8ac84ec8232cb7c3d3e4e29e2b47e2b3178e3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 301, + 864, + 455 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/db25191b988ea444aa2308b4496a313ed24b0dcc14cad4d704ffcfb923089b5c.jpg", + "image_caption": [ + "(a) Tracking performance", + "Figure 7. Bi-criteria figure of tracking/render performance and system FPS on Replica #Office0." + ], + "image_footnote": [], + "bbox": [ + 500, + 492, + 679, + 611 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/346e7ac0b3ae2b5afc7b9731206c46252952a57766f796d562e794810a2f9413.jpg", + "image_caption": [ + "(b) Render performance" + ], + "image_footnote": [], + "bbox": [ + 707, + 493, + 887, + 611 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ping and accurate camera pose estimation, striking a better speed-accuracy balance. However, its reliance on high-quality depth data may limit performance in certain conditions. Additionally, the approach's high memory requirements for large scenes suggest future improvements could focus on optimizing memory use, potentially via techniques such as quantization and clustering. We believe GS-SLAM has the potential to extend to larger scale with some improvements and will explore this in future work.", + "bbox": [ + 496, + 657, + 890, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work is supported by the Shanghai AI Laboratory, National Key R&D Program of China (2022ZD0160101), the National Natural Science Foundation of China (62376222), Young Elite Scientists Sponsorship Program by CAST (2023QNRC001) and the Early Career Scheme of the Research Grants Council (RGC) of the Hong Kong SAR under grant No. 26202321.", + "bbox": [ + 496, + 795, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19602", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Michael Bloesch, Jan Czarnowski, Ronald Clark, Stefan Leutenegger, and Andrew J. Davison. Codeslam - learning a compact, optimisable representation for dense visual slam. CVPR, pages 2560-2568, 2018. 2", + "[2] Guillaume Bresson, Zayed Alsayed, Li Yu, and Sébastien Glaser. Simultaneous localization and mapping: A survey of current trends in autonomous driving. TIV, 2:194-220, 2017. 6", + "[3] Zilong Chen, Feng Wang, and Huaping Liu. Text-to-3d using gaussian splatting. ArXiv, abs/2309.16585, 2023. 2", + "[4] Brian Curless and Marc Levoy. Volumetric method for building complex models from range images. In SIGGRAPH, ACM, 1996. 4", + "[5] Parth Rajesh Desai, Pooja Nikhil Desai, Komal Deepak Ajmera, and Khushbu Mehta. A review paper on oculus rift-a virtual reality headset. ArXiv, abs/1408.1173, 2014. 6", + "[6] Hugh F. Durrant-Whyte and Tim Bailey. Simultaneous localization and mapping: part i. RAM, 13:99–110, 2006. 1", + "[7] Christian Hane, Christopher Zach, Jongwoo Lim, Ananth Ranganathan, and Marc Pollefeys. Stereo depth map fusion for robot navigation. IROS, pages 1618-1625, 2011. 6", + "[8] Caner Hazirbas, Andreas Wiedemann, Robert Maier, Laura Leal-Taixe, and Daniel Cremers. Tum rgb-d scribble-based segmentation benchmark. https://github.com/tum-vision/rgbdscribble_benchmark, 2018.6", + "[9] Jiahui Huang, Shi-Sheng Huang, Haoxuan Song, and ShiMin Hu. Di-fusion: Online implicit 3d reconstruction with deep priors. In CVPR, pages 8932-8941, 2021. 6", + "[10] Xudong Jiang, Lifeng Zhu, Jia Liu, and Aiguo Song. A slam-based 6 dof controller with smooth auto-calibration for virtual reality. TVC, 39:3873 - 3886, 2022. 1", + "[11] Mohammad Mahdi Johari, Camilla Carta, and Francois Fleuret. Eslam: Efficient dense slam system based on hybrid representation of signed distance fields. CVPR, 2023. 1, 2, 5, 6, 7, 8, 4", + "[12] Olaf Kähler, Victor Adrian Prisacariu, Julien P. C. Valentin, and David William Murray. Hierarchical voxel block hashing for efficient integration of depth images. RAL, 1:192-197, 2016. 1", + "[13] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. TOG, 42(4), 2023. 2, 3, 7, 4", + "[14] Leonid Keselman and Martial Hebert. Approximate differentiable rendering with algebraic surfaces. In ECCV, 2022. 2", + "[15] Leonid Keselman and Martial Hebert. Flexible techniques for differentiable rendering with 3d gaussians. arXiv preprint arXiv:2308.14737, 2023. 2", + "[16] Georg S. W. Klein and David William Murray. Parallel tracking and mapping on a camera phone. ISMAR, pages 83-86, 2009. 2", + "[17] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. ArXiv, abs/2308.09713, 2023. 2" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] Robert Maier, Raphael Schaller, and Daniel Cremers. Efficient online surface correction for real-time large-scale 3d reconstruction. ArXiv, abs/1709.03763, 2017. 1", + "[19] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1", + "[20] Raul Mur-Artal and Juan D. Tardós. Orb-slam2: An opensource slam system for monocular, stereo, and rgb-d cameras. *TRO*, 33:1255–1262, 2016. 1, 6", + "[21] Richard A. Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J. Davison, Pushmeet Kohli, Jamie Shotton, Steve Hodges, and Andrew William Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. ISMAR, pages 127-136, 2011. 1", + "[22] Richard A. Newcombe, S. Lovegrove, and Andrew J. Davison. Dtam: Dense tracking and mapping in real-time. ICCV, pages 2320-2327, 2011. 2", + "[23] Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Marc Stamminger. Real-time 3d reconstruction at scale using voxel hashing. TOG, 32:1-11, 2013. 1", + "[24] Delin Qu, Chi Yan, Dong Wang, Jie Yin, Dan Xu, Bin Zhao, and Xuelong Li. Implicit event-rgbd neural slam. CVPR, 2024. 1", + "[25] Gerhard Reitmayr, Tobias Langlotz, Daniel Wagner, Alessandro Mulloni, Gerhard Schall, Dieter Schmalstieg, and Qi Pan. Simultaneous localization and mapping for augmented reality. ISUVR, pages 5-8, 2010. 1", + "[26] Fabio Ruetz, Emili Hernández, Mark Pfeiffer, Helen Oleynikova, Mark Cox, Thomas Lowe, and Paulo Vinicius Koerich Borges. Ovpc mesh: 3d free-space representation for local ground vehicle navigation. ICRA, pages 8648-8654, 2018. 1", + "[27] Erik Sandström, Yue Li, Luc Van Gool, and Martin R. Oswald. Point-slam: Dense neural point cloud-based slam. In ICCV, 2023. 2, 5, 6, 7, 4", + "[28] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 3", + "[29] Thomas Schöps, Torsten Sattler, and Marc Pollefeys. Bad slam: Bundle adjusted direct rgb-d slam. CVPR, pages 134-144, 2019. 2", + "[30] Thomas Schops, Torsten Sattler, and Marc Pollefeys. BAD SLAM: Bundle adjusted direct RGB-D SLAM. In CVPR, 2019. 6", + "[31] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Yuheng Ren, Shobhit Verma, Anton Clarkson, Ming Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Briales, Tyler Gillingham, Elias Mueggler, Luis Pesqueira, Manolis Savva, Dhruv Batra, Hauke Malte Strasdat, Renzo De Nardi, Michael Goesele, S. Lovegrove, and Richard A. Newcombe. The replica dataset: A digital replica of indoor spaces. ArXiv, abs/1906.05797, 2019. 5, 3, 4", + "[32] J. Stuckler and Sven Behnke. Multi-resolution surfel maps for efficient dense 3d modeling and tracking. JVCIR, 25: 137-147, 2014. 1" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19603", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Jürgen Sturm, Nikolas Engelhard, Felix Endres, Wolfram Burgard, and Daniel Cremers. A benchmark for the evaluation of RGB-D SLAM systems. In IROS. IEEE/RSJ, 2012. 5, 6", + "[34] Edgar Sucar, Kentaro Wada, and Andrew J. Davison. Nodeslam: Neural object descriptors for multi-view shape reconstruction. 3DV, pages 949–958, 2020. 2", + "[35] Edgar Sucar, Shikun Liu, Joseph Ortiz, and Andrew J. Davison. imap: Implicit mapping and positioning in real-time. ICCV, 2021. 1, 2, 6", + "[36] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. ArXiv, abs/2309.16653, 2023. 2, 5", + "[37] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. In NIPS, 2021. 2", + "[38] Andreas Langeland Teigen, Yeonsoo Park, Annette Stahl, and Rudolf Mester. Rgb-d mapping and tracking in a plenoxel radiance field. ArXiv, abs/2307.03404, 2023. 2", + "[39] Charalambos Theodorou, Vladan Velisavljevic, Vladimir Dyo, and Fredi Nonyelu. Visual slam algorithms and their application for ar, mapping, localization and wayfinding. Array, 15:100222, 2022. 1", + "[40] Angtian Wang, Peng Wang, Jian Sun, Adam Kortylewski, and Alan Yuille. Voge: a differentiable volume renderer using gaussian ellipsoids for analysis-by-synthesis. arXiv preprint arXiv:2205.15401, 2022. 2", + "[41] Hengyi Wang, Jingwen Wang, and Lourdes de Agapito. Coslam: Joint coordinate and sparse parametric encodings for neural real-time slam. CVPR, 2023. 1, 5, 6, 8, 2, 4", + "[42] Kaixuan Wang, Fei Gao, and Shaojie Shen. Real-time scalable dense surfel mapping. ICRA, pages 6919-6925, 2019. 1", + "[43] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. TIP, 13(4):600-612, 2004. 5", + "[44] Thomas Whelan, Michael Kaess, Maurice F. Fallon, Hordur Johannsson, John J. Leonard, and John B. McDonald. Kintinous: Spatially extended kinectfusion. In AAAI, 2012. 2", + "[45] Thomas Whelan, John McDonald, Michael Kaess, Maurice Fallon, Hordur Johannsson, and John J. Leonard. Kintinuous: Spatially extended kinectfusion. In RSS Workshop on RGB-D, 2012. 6", + "[46] Thomas Whelan, Stefan Leutenegger, Renato Salas-Moreno, Ben Glocker, and Andrew Davison. Elasticfusion: Dense slam without a pose graph. In RSS, 2015. 1, 6", + "[47] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. ArXiv, abs/2310.08528, 2023. 2", + "[48] Xingrui Yang, Hai Li, Hongjia Zhai, Yuhang Ming, Yuqian Liu, and Guofeng Zhang. Vox-fusion: Dense tracking and mapping with voxel-based neural implicit representation. IS-MAR, pages 499-507, 2022. 2, 5, 6, 7, 8", + "[49] Ziyi Yang, Xinyu Gao, Wenming Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for high-fidelity monocular dynamic scene reconstruction. ArXiv, abs/2309.13101, 2023. 2", + "[50] Zeyu Yang, Hongye Yang, Zijie Pan, Xiatian Zhu, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. ArXiv, abs/2310.10642, 2023. 2", + "[51] Taoran Yi, Jiemin Fang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussiandreamer: Fast generation from text to 3d gaussian splattering with point cloud priors. ArXiv, abs/2310.08529, 2023. 2", + "[52] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 5", + "[53] Youmin Zhang, Fabio Tosi, Stefano Mattoccia, and Matteo Poggi. Go-slam: Global optimization for consistent 3d instant reconstruction. In ICCV, 2023. 2", + "[54] Shuaifeng Zhi, Michael Bloesch, Stefan Leutenegger, and Andrew J. Davison. Scenocode: Monocular dense semantic reconstruction using learned encoded scene representations. CVPR, pages 11768-11777, 2019. 2", + "[55] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R. Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. CVPR, 2021. 1, 2, 5, 6, 7, 8", + "[56] Wojciech Zielonka, Timur M. Bagautdinov, Shunsuke Saito, Michael Zollhofer, Justus Thies, and Javier Romero. Drivable 3d gaussian avatars. 2023. 2", + "[57] Matthias Zwicker, Hanspeter Pfister, Jeroen van Baar, and Markus H. Gross. Ewa volume splatting. Proceedings Visualization, 2001. VIS '01., pages 29-538, 2001. 3" + ], + "bbox": [ + 503, + 92, + 890, + 531 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19604", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/GS-SLAM_ Dense Visual SLAM with 3D Gaussian Splatting/93048e35-cf42-49e6-8897-d6803305631d_model.json b/2024/GS-SLAM_ Dense Visual SLAM with 3D Gaussian Splatting/93048e35-cf42-49e6-8897-d6803305631d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3e149af6fa21f4a52e48e7107d78d8c34c4dbbec --- /dev/null +++ b/2024/GS-SLAM_ Dense Visual SLAM with 3D Gaussian Splatting/93048e35-cf42-49e6-8897-d6803305631d_model.json @@ -0,0 +1,2354 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.18, + 0.131, + 0.79, + 0.154 + ], + "angle": 0, + "content": "GS-SLAM: Dense Visual SLAM with 3D Gaussian Splatting" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.179, + 0.314, + 0.198 + ], + "angle": 0, + "content": "Chi Yan\\(^{1,3*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.371, + 0.18, + 0.472, + 0.199 + ], + "angle": 0, + "content": "Delin \\(\\mathbf{Q}\\mathbf{u}^{1,2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.529, + 0.181, + 0.603, + 0.198 + ], + "angle": 0, + "content": "Dan \\(\\mathrm{Xu}^3\\)" + }, + { + "type": "text", + "bbox": [ + 0.667, + 0.181, + 0.761, + 0.197 + ], + "angle": 0, + "content": "Bin Zhao\\(^{1,4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.253, + 0.199, + 0.38, + 0.217 + ], + "angle": 0, + "content": "Zhigang Wang" + }, + { + "type": "text", + "bbox": [ + 0.446, + 0.199, + 0.493, + 0.217 + ], + "angle": 0, + "content": "Dong" + }, + { + "type": "text", + "bbox": [ + 0.616, + 0.199, + 0.726, + 0.217 + ], + "angle": 0, + "content": "Xuelong Li\\(^{1,5}\\)" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.217, + 0.464, + 0.253 + ], + "angle": 0, + "content": "\\(^{1}\\)Shanghai AI Laboratory \\(^{2}\\)Fudan University \\(^{4}\\)Northwestern Polytechnical University" + }, + { + "type": "text", + "bbox": [ + 0.489, + 0.217, + 0.895, + 0.253 + ], + "angle": 0, + "content": "\\(^{3}\\)Hong Kong University of Science and Technology \\(^{5}\\)TeleAI, China Telecom Corp Ltd" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.313, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.318, + 0.474, + 0.637 + ], + "angle": 0, + "content": "In this paper, we introduce GS-SLAM that first utilizes 3D Gaussian representation in the Simultaneous Localization and Mapping (SLAM) system. It facilitates a better balance between efficiency and accuracy. Compared to recent SLAM methods employing neural implicit representations, our method utilizes a real-time differentiable splatting rendering pipeline that offers significant speedup to map optimization and RGB-D rendering. Specifically, we propose an adaptive expansion strategy that adds new or deletes noisy 3D Gaussians in order to efficiently reconstruct new observed scene geometry and improve the mapping of previously observed areas. This strategy is essential to extend 3D Gaussian representation to reconstruct the whole scene rather than synthesize a static object in existing methods. Moreover, in the pose tracking process, an effective coarse-to-fine technique is designed to select reliable 3D Gaussian representations to optimize camera pose, resulting in runtime reduction and robust estimation. Our method achieves competitive performance compared with existing state-of-the-art real-time methods on the Replica, TUM-RGBD datasets. Project page: https://gs-slam.github.io/." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.661, + 0.21, + 0.677 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.686, + 0.47, + 0.868 + ], + "angle": 0, + "content": "Simultaneous localization and mapping (SLAM) has emerged as a pivotal technology in fields such as robotics [6], virtual reality [10], and augmented reality [25, 39]. The goal of SLAM is to construct a dense/sparse map of an unknown environment while simultaneously tracking the camera pose. Traditional SLAM methods employ point/surfel clouds [20, 32, 42, 46], mesh representations [26], voxel hashing [12, 18, 23] or voxel grids [21] as scene representations to construct dense mapping, and have made considerable progress on localization accuracy. However, these methods face serious challenges in obtaining fine-grained dense maps." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.285, + 0.892, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.522, + 0.895, + 0.62 + ], + "angle": 0, + "content": "Figure 1. The illustration of the proposed GS-SLAM. It first utilizes the 3D Gaussian representation and differentiable splatting rasterization pipeline in SLAM, achieving real-time tracking and mapping performance on GPU. Besides, benefiting from the splatting rasterization pipeline, GS-SLAM achieves a \\(100\\times\\) faster rendering FPS and more high-quality full image results than the other SOTA methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Recently, Neural Radiance Fields (NeRF) [19] have been explored to enhance SLAM methodologies and exhibit strengths in generating high-quality, dense maps with low memory consumption [35]. In particular, iMAP [35] uses a single multi-layer perceptron (MLP) to represent the entire scene, which is updated globally with the loss between volume-rendered RGB-D image and ground-truth observations. NICE-SLAM [55] utilizes a hierarchical neural implicit grid as scene map representation to allow local updates for reconstructing large scenes. Moreover, ESLAM [11], CoSLAM [41] and EN-SLAM [24] utilize axis-aligned feature planes and joint coordinate-parametric encoding to improve the capability of scene representation, achieving efficient and high-quality surface map reconstruction. In practical mapping and tracking steps, these methods only render a small set of pixels to reduce optimization time, which leads to the reconstructed dense maps lacking" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.323, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution: yanchi@pjlab.org.cn." + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.888, + 0.383, + 0.9 + ], + "angle": 0, + "content": "†Corresponding author: dongwang.dw93@gmail.com." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.383, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19595" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.166 + ], + "angle": 0, + "content": "the richness and intricacy of details. In essence, it is a tradeoff for the efficiency and accuracy of NeRF-based SLAM since obtaining high-resolution images with the ray-based volume rendering technique is time-consuming and unacceptable." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.167, + 0.468, + 0.332 + ], + "angle": 0, + "content": "Fortunately, recent work [13, 17, 47] with 3D Gaussian representation and tile-based splatting techniques has shown great superiority in the efficiency of high-resolution image rendering. It is applied to synthesize novel view RGB images of static objects, achieving state-of-the-art visual quality for \\(1080\\mathrm{p}\\) resolution at real-time speed. Inspired by this, we extend the rendering superiority of 3D Gaussian scene representation and real-time differentiable splatting rendering pipeline for the task of dense RGB-D SLAM and manage to jointly promote the speed and accuracy of NeRF-based dense SLAM, as shown in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.333, + 0.468, + 0.619 + ], + "angle": 0, + "content": "To this end, we propose GS-SLAM, the first RGB-D dense SLAM system that first utilizes 3D Gaussian scene representation coupled with the splatting rendering technique to better balance speed and accuracy. Our system optimizes camera tracking and mapping with a novel RGB-D rendering approach that processes 3D Gaussians quickly and accurately through sorting and \\(\\alpha\\)-blending. We enhance scene reconstruction by introducing an adaptive strategy for managing 3D Gaussian elements, which optimizes mapping by focusing on current observations and minimizes errors in dense maps and images. Moreover, we propose a coarse-to-fine approach, starting with low-resolution image analysis for initial pose estimation and refining it with high-resolution rendering using select 3D Gaussians, to boost speed and accuracy. We perform extensive evaluations on a selection of indoor RGB-D datasets and demonstrate state-of-the-art performance on dense neural RGB-D SLAM in terms of tracking, rendering, and mapping. Overall, our contributions include:" + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.622, + 0.468, + 0.711 + ], + "angle": 0, + "content": "- We propose GS-SLAM, the first 3D Gaussian Splatting(3DGS)-based dense RGB-D SLAM approach, which takes advantage of the fast splatting rendering technique to boost the mapping optimizing and pose tracking, achieving real-time and photo-realistic reconstruction performance." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.713, + 0.468, + 0.771 + ], + "angle": 0, + "content": "- We present an adaptive 3D Gaussian expansion strategy to efficiently reconstruct new observed scene geometry and develop a coarse-to-fine technique to select reliable 3D Gaussians to improve camera pose estimation." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.773, + 0.468, + 0.832 + ], + "angle": 0, + "content": "- Our approach achieves competitive performance on Replica and TUM-RGBD datasets in terms of tracking, and mapping and runs at 8.43 FPS, resulting in a better balance between efficiency and accuracy." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.622, + 0.468, + 0.832 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.844, + 0.218, + 0.859 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "- Dense Visual SLAM. The existing real-time dense visual SLAM systems are typically based on discrete handcrafted" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.319 + ], + "angle": 0, + "content": "features or deep-learning embeddings, and follow the mapping and tracking architecture in [16]. DTAM [22] first introduces a dense SLAM system that uses photometric consistency to track a handheld camera and represent the scene as a cost volume. KinectFusion [44] performs camera tracking by iterative-closest-point and updates the scene via TSDF-Fusion. BAD-SLAM [29] proposes to jointly optimize the keyframe poses and 3D scene geometry via a direct bundle adjustment (BA) technique. In contrast, recent works integrate deep learning with the traditional geometry framework for more accurate and robust camera tracking and mapping, such as DROID-SLAM [37], CodeSLAM [1], SceneCode [54], and NodeSLAM [34], have made significant advances in the field, achieving more accurate and robust camera tracking and mapping performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.322, + 0.892, + 0.64 + ], + "angle": 0, + "content": "Neural Implicit Radiance Field based SLAM. For NeRF-based SLAM, existing methods can be divided into three main types: MLP-based methods, Hybrid representation methods, and Explicit methods. MLP-based method iMAP [35] offers scalable and memory-efficient map representations but faces challenges with catastrophic forgetting in larger scenes. Hybrid representation methods combine the advantages of implicit MLPs and structure features, significantly enhancing the scene scalability and precision. For example, NICE-SLAM [55] integrates MLPs with multi-resolution voxel grids, enabling large scene reconstruction, and Vox-Fusion [48] employs octree expansion for dynamic map scalability, while ESLAM [11] and Point-SLAM [27] utilize tri-planes and neural point clouds respectively to improve the mapping capability. As for the explicit method proposed in [38], it stores map features in voxel directly, without any MLPs, enabling faster optimization. Instead of representing maps with implicit features, GS-SLAM utilizes the 3D Gaussian representation, efficiently renders images using splatting-based rasterization, and optimizes parameters directly with backward propagation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.892, + 0.901 + ], + "angle": 0, + "content": "3D Gaussian Representation. Several recent approaches have used 3D Gaussians for shape reconstruction, such as Fuzzy Metaballs [14, 15], VoGE [40], 3DGS [13]. Notably, 3DGS [13] demonstrates great superiorities in high-quality real-time novel-view synthesis. This work represents the scene with 3D Gaussians and develops a NeRF-style fast rendering algorithm to support anisotropic splating, achieving SOTA visual quality and fast high-resolution rendering performance. Beyond the rendering superiorities, Gaussian splating holds an explicit geometry scene structure and appearance, benefiting from the exact modeling of scenes representation [50]. This promising technology has been rapidly applied in several fields, including 3D generation [3, 36, 51], dynamic scene modeling [17][47][49], and photorealistic drivable avatars [56]. However, currently, there is no research addressing camera pose estimation or real-time mapping using 3D Gaussian models due to the in" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19596" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.895, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.248, + 0.895, + 0.292 + ], + "angle": 0, + "content": "Figure 2. Overview of the proposed method. We aim to use 3D Gaussians to represent the scene and use the rendered RGB-D image for inverse camera tracking. GS-SLAM proposes a novel Gaussian expansion strategy to make the 3D Gaussian feasible to reconstruct the whole scene and can achieve real-time tracking, mapping, and rendering performance on GPU." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.309, + 0.471, + 0.384 + ], + "angle": 0, + "content": "herent limitations of the prime pipeline [13], i.e., prerequisites of initialized point clouds or camera pose inputs [28]. In contrast, we derive the analytical derivative equations for pose estimation in the Gaussian representation and implement efficient CUDA optimization." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.401, + 0.212, + 0.418 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.427, + 0.474, + 0.683 + ], + "angle": 0, + "content": "Fig. 2 shows the overview of the proposed GS-SLAM. We aim to estimate the camera poses \\(\\{\\mathbf{P}_i\\}_{i=1}^N\\) of every frame and simultaneously reconstruct a dense scene map by giving an input sequential RGB-D stream \\(\\{\\mathbf{I}_i, \\mathbf{D}_i\\}_{i=1}^M\\) with known camera intrinsic \\(\\mathbf{K} \\in \\mathbb{R}^{3 \\times 3}\\). In Sec. 3.1, we first introduce 3D Gaussian as the scene representation \\(\\mathbf{S}\\) and the RGB-D render by differentiable splattering rasterization. With the estimated camera pose of the keyframe, in Sec. 3.2, an adaptive expansion strategy is proposed to add new or delete noisy 3D Gaussians to efficiently reconstruct new observed scene geometry while improving the mapping of the previously observed areas. For camera tracking of every input frame, we derive an analytical formula for backward optimization with rendering RGB-D loss. We further introduce an effective coarse-to-fine technique to minimize rendering losses to achieve efficient and accurate pose estimation in Sec. 3.3." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.697, + 0.383, + 0.714 + ], + "angle": 0, + "content": "3.1. 3D Gaussian Scene Representation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.721, + 0.47, + 0.797 + ], + "angle": 0, + "content": "Our goal is to optimize a scene representation that captures the geometry and appearance of the scene, resulting in a detailed dense map and high-quality novel view synthesis. To do this, we model the scene as a set of 3D Gaussians coupled with opacity and spherical harmonics" + }, + { + "type": "equation", + "bbox": [ + 0.129, + 0.81, + 0.47, + 0.827 + ], + "angle": 0, + "content": "\\[\n\\mathbf {G} = \\left\\{G _ {i}: \\left(\\mathbf {X} _ {i}, \\boldsymbol {\\Sigma} _ {i}, \\Lambda_ {i}, \\mathbf {Y} _ {i}\\right) | i = 1, \\dots , N \\right\\}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Each 3D Gaussian scene representation \\( G_{i} \\) is defined by position \\( \\mathbf{X}_i\\in \\mathbb{R}^3 \\), 3D covariance matrix \\( \\pmb{\\Sigma}_{i}\\in \\mathbb{R}^{3\\times 3} \\), opacity \\( \\Lambda_{i}\\in \\mathbb{R} \\) and 1-degree spherical harmonics \\( (\\mathbf{Y}) \\) per color channel, a total of 12 coefficients for \\( \\mathbf{Y}_i\\in \\mathbb{R}^{12} \\). To reduce" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.309, + 0.892, + 0.339 + ], + "angle": 0, + "content": "the learning difficulty of the 3D Gaussians [57], we parameterize the 3D Gaussian's covariance as:" + }, + { + "type": "equation", + "bbox": [ + 0.639, + 0.348, + 0.892, + 0.368 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Sigma} = \\mathbf {R} \\mathbf {S} \\mathbf {S} ^ {T} \\mathbf {R} ^ {T}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.376, + 0.892, + 0.407 + ], + "angle": 0, + "content": "where \\(\\mathbf{S} \\in \\mathbb{R}^3\\) is a 3D scale vector, \\(\\mathbf{R} \\in \\mathbb{R}^{3 \\times 3}\\) is the rotation matrix, storing as a 4D quaternion." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.408, + 0.892, + 0.469 + ], + "angle": 0, + "content": "Color and depth splatting rendering. With the optimized 3D Gaussian scene representation parameters, given the camera pose \\(\\mathbf{P} = \\{\\mathbf{R},\\mathbf{t}\\}\\), the 3D Gaussians \\(G\\) are projected into 2D image plane for rendering with:" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.477, + 0.892, + 0.496 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Sigma} ^ {\\prime} = \\mathbf {J} \\mathbf {P} ^ {- 1} \\boldsymbol {\\Sigma} \\mathbf {P} ^ {- T} \\mathbf {J} ^ {T}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.506, + 0.894, + 0.582 + ], + "angle": 0, + "content": "where \\(\\mathbf{J}\\) is the Jacobian of the affine approximation of the projective function. After projecting 3D Gaussians to the image plane, the color of one pixel is rendered by sorting the Gaussians in depth order and performing front-to-back \\(\\alpha\\)-blending rendering as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.592, + 0.892, + 0.635 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {C}} = \\sum_ {i \\in N} \\mathbf {c} _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.894, + 0.72 + ], + "angle": 0, + "content": "where \\(\\mathbf{c}_i\\) represents the color of the \\(i\\)-th 3D Gaussian obtained by learned spherical harmonics coefficients \\(\\mathbf{Y}\\), \\(\\alpha_i\\) is the density computed by learned opacity \\(\\Lambda_i\\) and 2D Gaussian with covariance \\(\\Sigma'\\). Similarly, the depth is rendered by" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.731, + 0.892, + 0.773 + ], + "angle": 0, + "content": "\\[\n\\hat {D} = \\sum_ {i \\in N} d _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.778, + 0.892, + 0.823 + ], + "angle": 0, + "content": "where \\( d_{i} \\) denotes the depth of the center of the \\( i \\)-th 3D Gaussian, which is obtained by projecting to \\( z \\)-axis in the camera coordinate." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.833, + 0.871, + 0.85 + ], + "angle": 0, + "content": "3.2. Adaptive 3D Gaussian Expanding Mapping" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "The 3D Gaussian scene representations are updated and optimized on each selected keyframe for stable mapping. Given the estimated pose of each selected keyframe, we first" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19597" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.198 + ], + "angle": 0, + "content": "apply the proposed adaptive expansion strategy to add new or delete noisy 3D Gaussians from the whole scene representations to render RGB-D images with resolution \\( H \\times W \\), and then the updated 3D Gaussian scene representations are optimized by minimizing the geometric depth loss \\( \\mathcal{L}_d \\) and the photometric color loss \\( \\mathcal{L}_{\\mathrm{c}} \\) to the sensor observation depth \\( D \\) and color \\( \\mathbf{C} \\)," + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.209, + 0.47, + 0.252 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathbf {c}} = \\sum_ {m = 1} ^ {H W} \\left| \\mathbf {C} _ {m} - \\hat {\\mathbf {C}} _ {m} \\right|, \\quad \\mathcal {L} _ {d} = \\sum_ {m = 1} ^ {H W} \\left| D _ {m} - \\hat {D} _ {m} \\right|. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.262, + 0.47, + 0.292 + ], + "angle": 0, + "content": "The loss optimizes the parameters of all 3D Gaussians that contribute to the rendering of these keyframe images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.294, + 0.471, + 0.43 + ], + "angle": 0, + "content": "Adaptive 3D Gaussian Expansion Strategy. At the first frame of the RGB-D sequence, we first uniformly sample half pixels from a whole image with \\( H \\times W \\) resolution and back-projecting them into 3D points \\( \\mathbf{X} \\) with corresponding depth observation \\( D \\). The 3D Gaussian scene representations are created by setting position as \\( \\mathbf{X} \\) and initializing zero degree Spherical Harmonics coefficients with RGB color \\( \\mathbf{C}_i \\). The opacities are set to pre-defined values, and the covariance is set depending on the spatial point density, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.127, + 0.442, + 0.47, + 0.46 + ], + "angle": 0, + "content": "\\[\n\\left\\{G _ {i} = \\left(\\mathbf {P} _ {i}, \\boldsymbol {\\Sigma} _ {\\text {i n i t}}, \\Lambda_ {\\text {i n i t}}, \\mathbf {C} _ {i}\\right) | i = 1, \\dots , M \\right\\}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.471, + 0.47, + 0.576 + ], + "angle": 0, + "content": "where \\( M \\) equals to \\( HW / 2 \\). The 3D Gaussians are initialized and then optimized using the first RGB-D image with rendering loss. Note that only half of the pixels are used to initialize the scene, leaving space to conduct adaptive density control of Gaussians that splits large points into smaller ones and clones them with different directions to capture missing geometric details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.578, + 0.47, + 0.714 + ], + "angle": 0, + "content": "Adding step: To obtain a complete map of the environment, the 3D Gaussian scene representations should be able to model the geometry and appearance of newly observed areas. Specifically, at every keyframe, we add first rendered RGB-D images using historical 3D Gaussians and calculate cumulative opacity \\( T = \\sum_{i\\in N}\\alpha_i\\prod_{j = 1}^{i - 1}(1 - \\alpha_j) \\) for each pixel. We label one pixel as un-reliable \\( x^{un} \\) if its cumulative opacity \\( T \\) is too low or its rendered depth \\( \\hat{D} \\) is far away from observed depth \\( D \\), i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.725, + 0.469, + 0.743 + ], + "angle": 0, + "content": "\\[\nT < \\tau_ {T} \\quad \\text {o r} \\quad | D - \\hat {D} | > \\tau_ {D}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.824 + ], + "angle": 0, + "content": "These selected un-reliable pixels mostly capture new observed areas. Then we back-project these un-reliable pixels to 3D points \\(\\mathbf{P}^{un}\\), and a set of new 3D Gaussians at \\(\\mathbf{P}^{un}\\) initialized as Eq. 7 are added into scene representations to model the new observed areas." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.901 + ], + "angle": 0, + "content": "**Deleting step:** As shown in Fig. 3, there are some floating 3D Gaussians due to the unstable adaptive control of Gaussians after optimization with Eq. 6. These floating 3D Gaussians will result in a low-quality dense map and a rendered image containing lots of artifacts. To address this" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.085, + 0.875, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.261, + 0.892, + 0.304 + ], + "angle": 0, + "content": "Figure 3. Illustration of the proposed adaptive 3D Gaussian expansion strategy. GS-SLAM inhibits the low-quality 3D Gaussian floaters in the current frustum according to depth." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.318, + 0.892, + 0.469 + ], + "angle": 0, + "content": "issue, after adding new 3D Gaussians, we check all visible 3D Gaussians in the current camera frustum and significantly decrease opacity \\(\\Lambda_{i}\\) of 3D Gaussians whose position is not near the scene surfaces. Formally, for each visible 3D Gaussian, we draw a ray \\(\\mathbf{r}(t)\\) from camera origin \\(\\mathbf{o}\\) and its position \\(\\mathbf{X}_i = (x_i,y_i,z_i)\\), i.e., \\(\\mathbf{r}(t) = \\mathbf{o} + t(\\mathbf{X}_i - \\mathbf{o})\\). Then, we can find a pixel with coordinate \\((u,v)\\) where this ray intersects the image plane and corresponding depth observation \\(D\\). The 3D Gaussians are deleted by degenerating its opacity as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.484, + 0.892, + 0.501 + ], + "angle": 0, + "content": "\\[\nG _ {i}: \\Lambda_ {i} \\Rightarrow G _ {i}: \\eta \\Lambda_ {i}, \\text {i f} D - d i s t (\\mathbf {X} _ {i}, \\mathbf {P} _ {u v}) > \\gamma , \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.508, + 0.892, + 0.599 + ], + "angle": 0, + "content": "where \\( P_{uv} \\) is the world coordinates of the intersected pixel calculated with the camera intrinsic and extrinsic. \\( dist(\\cdot, \\cdot) \\) is the Euclidean distance, and \\( \\eta \\) (much smaller than 1) and \\( \\gamma \\) are the hyper-parameters. Note that we decrease the opacity of floating 3D Gaussians in front of the scene surfaces to make our newly added 3D Gaussians well-optimized." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.611, + 0.795, + 0.627 + ], + "angle": 0, + "content": "3.3. Tracking and Bundle Adjustment" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.635, + 0.892, + 0.742 + ], + "angle": 0, + "content": "In the parallel camera tracking phase of our work, we first employ a common straightforward constant velocity assumption to initialize new poses. This assumption transforms the last known pose based on the relative transformation between the second-to-last pose and the last pose. Then, the accurate camera pose \\(\\mathbf{P}\\) is optimized by minimizing rendered color loss, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.755, + 0.892, + 0.796 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t r a c k}} = \\sum_ {m = 1} ^ {M} \\left| \\mathbf {C} _ {m} - \\hat {\\mathbf {C}} _ {m} \\right| _ {1}, \\quad \\min _ {\\mathbf {R}, \\mathbf {t}} (\\mathcal {L} _ {\\text {t r a c k}}), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.808, + 0.871, + 0.824 + ], + "angle": 0, + "content": "where \\(M\\) is the number of sampled pixels for rendering." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Differentiable pose estimation. According to Eqs. (3) and (4), we observe that the gradient of the camera pose \\(\\mathbf{P}\\) is related to three intermediate variables: \\(\\pmb{\\Sigma}^{\\prime}\\), \\(\\mathbf{c}_i\\), and the projected coordinate \\(\\mathbf{m}_i\\) of Gaussian \\(G_{i}\\). By applying the chain rule of derivation, we obtain the analytical formula" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19598" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.232, + 0.106 + ], + "angle": 0, + "content": "tion of camera pose P:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.117, + 0.469, + 0.203 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\frac {\\partial \\mathcal {L} _ {\\mathbf {c}}}{\\partial \\mathbf {P}} = \\frac {\\partial \\mathcal {L} _ {\\mathbf {c}}}{\\partial \\mathbf {C}} \\frac {\\partial \\mathbf {C}}{\\partial \\mathbf {P}} = \\frac {\\partial \\mathcal {L} _ {\\mathbf {c}}}{\\partial \\mathbf {C}} \\left(\\frac {\\partial \\mathbf {C}}{\\partial \\mathbf {c} _ {i}} \\frac {\\partial \\mathbf {c} _ {i}}{\\partial \\mathbf {P}} + \\frac {\\partial \\mathbf {C}}{\\partial \\alpha_ {i}} \\frac {\\partial \\alpha_ {i}}{\\partial \\mathbf {P}}\\right) \\\\ = \\frac {\\partial \\mathcal {L} _ {\\mathrm {c}}}{\\partial \\overline {{\\mathbf {C}}}} \\frac {\\partial \\mathbf {C}}{\\partial \\alpha_ {i}} \\left(\\frac {\\partial \\alpha_ {i}}{\\partial \\boldsymbol {\\Sigma} ^ {\\prime}} \\frac {\\partial \\boldsymbol {\\Sigma} ^ {\\prime}}{\\partial \\mathbf {P}} + \\frac {\\partial \\alpha_ {i}}{\\partial \\mathbf {m} _ {\\mathrm {i}}} \\frac {\\partial \\mathbf {m} _ {\\mathrm {i}}}{\\partial \\mathbf {P}}\\right) \\tag {11} \\\\ = \\frac {\\partial \\mathcal {L} _ {\\mathbf {c}}}{\\partial \\mathbf {C}} \\frac {\\partial \\mathbf {C}}{\\partial \\alpha_ {i}} \\left(\\frac {\\partial \\alpha_ {i}}{\\partial \\boldsymbol {\\Sigma} ^ {\\prime}} \\frac {\\partial (\\mathbf {J P} ^ {- 1} \\boldsymbol {\\Sigma} \\mathbf {P} ^ {- T} \\mathbf {J} ^ {T})}{\\partial \\mathbf {P}} + \\frac {\\partial \\alpha_ {i}}{\\partial \\mathbf {m} _ {\\mathbf {i}}} \\frac {\\partial (\\mathbf {K P X} _ {i})}{\\partial \\mathbf {P} d _ {i}}\\right) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.213, + 0.468, + 0.338 + ], + "angle": 0, + "content": "where \\(d_{i}\\) denotes the \\(z\\)-axis coordinate of projection \\(\\mathbf{m}_i\\). The item \\(\\frac{\\partial\\mathbf{C}}{\\partial\\mathbf{c}_i}\\frac{\\partial\\mathbf{c}_i}{\\partial\\mathbf{P}}\\) can be eliminated because we are only concerned about the view-independent color in our tracking implementation. In addition, we find that the intermediate gradient \\(\\frac{\\partial(\\mathbf{KP}\\mathbf{X}_i)}{\\partial\\mathbf{P}d_i}\\) is the deterministic component for the camera pose \\(\\mathbf{P}\\). So we simply ignore the backpropagation of \\(\\frac{\\partial(\\mathbf{JP}^{-1}\\pmb{\\Sigma}\\mathbf{P}^{-T}\\mathbf{J}^{T})}{\\partial\\mathbf{P}}\\) for efficiency. More details can be found in the supplemental materials." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.34, + 0.468, + 0.536 + ], + "angle": 0, + "content": "Coarse-to-fine camera tracking. It would be problematic to optimize the camera pose with all image pixels since artifacts in images can cause drifted camera tracking. To address this issue, as shown in Fig. 2, in the differentiable pose estimation step for each frame, we first take advantage of image regularity to render only a sparse set of pixels and optimize tracking loss to obtain a coarse camera pose. This coarse optimization step significantly eases the influence of detailed artifacts. Further, we use this coarse camera pose and depth observation to select reliable 3D Gaussians, which guides GS-SLAM to render informative areas with clear geometric structures to refine coarse camera pose via further optimizing tracking loss on new rendering pixels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.537, + 0.468, + 0.779 + ], + "angle": 0, + "content": "Specifically, in the coarse stage, we first render a coarse image \\(\\hat{\\mathbf{I}}_c\\) with resolution \\(H / 2\\times W / 2\\) at uniformly sampled image coordinates and optimize tracking loss in Eq. 10 for \\(T_{c}\\) iterations, and the obtained camera pose is denoted as \\(\\mathbf{P}_c\\). In the fine stage, we use a similar technique with adaptive 3D Gaussian expansion strategy in Section 3.2 to select reliable 3D Gaussian to render full-resolution images while ignoring noisy 3D Gaussians that cause artifacts. In detail, we check all visible 3D Gaussians under coarse camera pose \\(\\mathbf{P}_c\\), and remove 3D Gaussians whose position is far away from the scene surface. Formally, for each visible 3D Gaussians \\(G_{i}\\) with position \\(\\mathbf{X}_i\\), we project it to the camera plane using coarse camera pose \\(\\mathbf{P}_c\\) and camera intrinsic. Given the projected pixel's depth observation \\(D_{i}\\) and the distance \\(d_{i}\\) that is between 3D Gaussians \\(G_{i}\\) and the camera image plane, the reliable 3D Gaussians are selected as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.789, + 0.469, + 0.828 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {G} _ {\\text {s e l e c t e d}} = \\left\\{G _ {i} \\mid G _ {i} \\in \\mathbf {G} \\text {a n d} a b s \\left(D _ {i} - d _ {i}\\right) \\leq \\varepsilon \\right\\}, \\tag {12} \\\\ \\tilde {\\mathbf {I}} _ {f} = \\mathcal {F} (u, v, \\mathbf {G} _ {s e l e c t e d}), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.839, + 0.469, + 0.902 + ], + "angle": 0, + "content": "where we use the selected reliable 3D Gaussians to render full-resolution images \\(\\hat{\\mathbf{I}}_f\\). \\(u, v\\) denote the pixel coordinates in \\(\\hat{\\mathbf{I}}_f\\), and \\(\\mathcal{F}\\) represents the color splattering rendering function. The final camera pose \\(\\mathbf{P}\\) is obtained by optimizing" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.091, + 0.892, + 0.229 + ], + "angle": 0, + "content": "tracking loss in Eq. 10 with \\(\\hat{\\mathbf{I}}_f\\) for other \\(T_{f}\\) iterations. Note that \\(\\hat{\\mathbf{I}}_c\\) and \\(\\hat{\\mathbf{I}}_f\\) are only rendered at previously observed areas, avoiding rendering areas where 3D scene representations have not been optimized in the mapping process. Also, we add keyframes based on the proportion of the currently observed image's reliable region to the overall image. At the same time, when the current tracking frame and most recent keyframe differ by more than a threshold value \\(\\mu_{k}\\), this frame will be inserted as a keyframe." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.23, + 0.892, + 0.381 + ], + "angle": 0, + "content": "**Bundle adjustment.** In the bundle adjustment (BA) phase, we optimize the camera poses \\(\\mathbf{P}\\) and the 3D Gaussian scene representation \\(\\mathbf{S}\\) jointly. We randomly select \\(K\\) keyframes from the keyframe database for optimization, using the loss function similar to the mapping part. For pose optimization stability, we only optimize the scene representation \\(\\mathbf{S}\\) in the first half of the iterations. In the other half of the iterations, we simultaneously optimize the map and the poses. Then, the accurate camera pose \\(\\mathbf{P}\\) is optimized by minimizing rendering color loss, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.389, + 0.892, + 0.423 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {b a} = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\sum_ {m = 1} ^ {H W} \\left| D _ {m} - \\hat {D} _ {m} \\right| _ {1} + \\lambda_ {m} \\left| \\mathbf {C} _ {m} - \\hat {\\mathbf {C}} _ {m} \\right| _ {1}, \\min _ {\\mathbf {R}, \\mathbf {t}, \\mathbf {S}} (\\mathcal {L} _ {b a}). \\tag {13}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.431, + 0.625, + 0.447 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.455, + 0.691, + 0.471 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.478, + 0.892, + 0.568 + ], + "angle": 0, + "content": "Dataset. To evaluate the performance of GS-SLAM, we conduct experiments on the Replica [31], and TUM-RGBD [33]. Following [11, 27, 41, 48, 55], we use 8 scenes from the Replica dataset for localization, mesh reconstruction, and rendering quality comparison. The selected three subsets of TUM-RGBD datasets are used for localization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.892, + 0.659 + ], + "angle": 0, + "content": "Baselines. We compare our method with existing SOTA NeRF-based dense visual SLAM: NICE-SLAM [55], VoxFusion [48], CoSLAM [41], ESLAM [11] and Point-SLAM [27]. The rendering performance of CoSLAM [41] and ESLAM [11] is conducted from the open source code with the same configuration in [27]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.892, + 0.81 + ], + "angle": 0, + "content": "Metric. For mesh reconstruction, we use the 2D Depth L1 (cm) [55], the Precision \\((\\mathbb{P},\\%)\\), Recall \\((\\mathbb{R},\\%)\\), and F-score with a threshold of \\(1\\mathrm{cm}\\) to measure the scene geometry. For localization, we use the absolute trajectory (ATE, cm) error [33] to measure the accuracy of the estimated camera poses. We further evaluate the rendering performance using the peak signal-to-noise ratio (PSNR), SSIM [43], and LPIPS [52] by following [27]. To be fair, we run all the methods on a dataset 10 times and report the average results. More details can be found in the supplemental materials." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Implementation details. GS-SLAM is implemented in Python using the PyTorch framework, incorporating CUDA code for Gaussian splatting and trained on a desktop PC with a 5.50GHz Intel Core i9-13900K CPU and NVIDIA RTX 4090 GPU. We extended the existing code for differentiable Gaussian splatting rasterization with additional func" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19599" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.138 + ], + "angle": 0, + "content": "tionality for handling depth, pose, and cumulative opacity during both forward and backward propagation. More details can be found in the supplemental materials." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.147, + 0.424, + 0.164 + ], + "angle": 0, + "content": "4.2. Evaluation of Localization and Mapping" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.171, + 0.47, + 0.442 + ], + "angle": 0, + "content": "Evaluation on Replica. Tracking ATE: Tab. 1 illustrates the tracking performance of our method and the state-of-the-art methods on the Replica dataset. Our method achieves the best or second performance in 7 of 8 scenes and outperforms the second-best method Point-SLAM [27] by \\(0.4\\mathrm{cm}\\) on average at 8.34 FPS. It is noticeable that the second best method, Point-SLAM [27] runs at 0.42 FPS, which is \\(20\\times\\) slower than our method, indicating that GS-SLAM achieves a better trade-off between the tracking accuracy and the runtime efficiency. Mapping ACC: Tab. 3 report the mapping evaluation results of our method with other current state-of-the-art visual SLAM methods. GS-SLAM achieves the best performance in Depth L1 (1.16cm) and Precision \\((74.0\\%)\\) metrics on average. For Recall and F1 scores, GS-SLAM performs comparably to the second best method CoSLAM [41]. The visualization results in Fig. 4 show that GS-SLAM achieves satisfying construction mesh with clear boundaries and details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.445, + 0.469, + 0.566 + ], + "angle": 0, + "content": "Evaluation on TUM-RGBD. Tab. 2 compares GS-SLAM with the other SLAM systems in TUM-RGBD dataset. Our method surpasses iMAP [35], NICE-SLAM [55] and Voxfusion [48], and achieves a comparable performance, average \\(3.7\\mathrm{cm}\\) ATE RSME, with the SOTA methods. A gap to traditional methods still exists between the neural vSLAM and the traditional SLAM systems, which employ more sophisticated tracking schemes [27]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.571, + 0.469, + 0.627 + ], + "angle": 0, + "content": "Table 1. Tracking comparison (ATE RMSE [cm]) of the proposed method vs. the SOTA methods on the Replica dataset. The running speed of methods in the upper part is lower than 5 FPS, * denotes the reproduced results by running officially released code." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.63, + 0.469, + 0.699 + ], + "angle": 0, + "content": "
MethodRm0Rm1Rm2Off0Off1Off2Off3Off4avg
Point-SLAM [27]0.560.470.300.350.620.550.720.730.54
NICE-SLAM [55]0.971.311.070.881.001.061.101.131.06
Vox-Fusion* [48]1.374.701.478.482.042.581.112.943.09
ESLAM [11]0.710.700.520.570.550.580.720.630.63
CoSLAM [41]0.700.951.350.590.552.031.560.721.00
Ours0.480.530.330.520.410.590.460.70.50
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.71, + 0.468, + 0.751 + ], + "angle": 0, + "content": "Table 2. Tracking ATE [cm] on TUM-RGBD [33]. Our method achieves a comparable performance among the neural vSLAMs. * denotes the reproduced results by running officially released code." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.755, + 0.468, + 0.817 + ], + "angle": 0, + "content": "
Methodfr1Deskfr2_xyzfr3_offAvg.Methodfr1Deskfr2_xyzfr3_offAvg.
DI-Fusion [9]4.42.05.84.1NICE-SLAM [55]4.331.73.913.3
Elastic-Fusion [46]2.51.22.52.1Vox-Fusion* [48]3.51.526.010.3
BAD-SLAM [30]1.71.11.71.5CoSLAM [41]2.71.92.62.4
Kintinous [45]3.72.93.03.2ESLAM [11]2.31.12.42.0
ORB-SLAM2 [20]1.60.41.01.0Point-SLAM2.61.33.22.4
iMAP* [35]7.22.19.06.1Ours3.31.36.63.7
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.283, + 0.848 + ], + "angle": 0, + "content": "4.3. Rendering Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We compare the rendering performance of the proposed GS-SLAM with the neural visual SLAM methods in Tab. 6. The results show that GS-SLAM achieves the best performance" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.094, + 0.89, + 0.121 + ], + "angle": 0, + "content": "Table 3. Reconstruction comparison of the proposed method vs. the SOTA methods on Replica dataset." + }, + { + "type": "table", + "bbox": [ + 0.509, + 0.124, + 0.892, + 0.343 + ], + "angle": 0, + "content": "
MethodMetricRm 0Rm 1Rm 2Off 0Off 1Off 2Off 3Off 4Avg.
NICESL AM [55]Depth L1 ↓1.811.442.041.391.768.334.992.012.97
Precision ↑45.8643.7644.3851.4050.8038.3740.8537.3544.10
Recall↑44.1046.1242.7848.6653.0839.9839.0435.7743.69
F1↑44.9644.8443.5649.9951.9139.1639.9236.5443.86
VoxFusion [48]Depth L1 ↓1.091.902.212.323.404.192.961.612.46
Precision↑75.8335.8863.1048.5143.5054.4869.1155.4055.73
Recall↑64.8933.0756.6244.7638.4447.8560.6146.7949.13
F1↑69.9334.3859.6746.5440.8150.9564.5650.7252.20
CoSLAM [41]Depth L1 ↓0.990.822.281.241.617.704.651.432.59
Precision↑81.7177.9573.3079.4180.6755.6457.6379.7673.26
Recall↑74.0370.7965.7371.4670.3552.9656.0671.2266.58
F1↑77.6874.2069.3175.2375.1654.2756.8375.2569.74
ESL AM [11]Depth L1 ↓0.630.620.980.571.667.323.940.882.08
Precision↑74.3375.9482.4872.2065.7470.7372.4872.2473.27
Recall↑87.3787.0184.9988.3684.3881.9279.1880.6384.23
F1↑80.3281.1083.7279.4773.9075.9275.6876.2178.29
OursDepth L1 ↓1.310.821.260.810.961.411.531.081.16
Precision↑64.5883.1170.1383.4387.7770.9163.1868.8874.00
Recall↑61.2976.8363.8476.9076.1561.6362.9161.5067.63
F1↑62.8979.8566.8480.0381.5565.9559.1764.9870.15
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.364, + 0.892, + 0.606 + ], + "angle": 0, + "content": "in all the metrics. Our method significantly outperforms the second-best methods CoSLAM [41], ESLAM [11] and NICE-SLAM [55] by 1.52 dB in PSNR, 0.027 in SSIM and 0.12 in LPIPS, respectively. It is noticeable that GS-SLAM achieves 386 FPS rendering speed on average, which is \\(100 \\times\\) faster than the second-best method Vox-Fusion [48]. This excellent rendering performance is attributed to the efficient 3D Gaussian rendering pipeline and can be further applied to real-time downstream tasks, such as VR [5], robot navigation [7] and autonomous driving [2]. The visualization results in Fig. 5 show that GS-SLAM can generate much more high-quality and realistic images than the other methods, especially in edge areas with detailed structures. While NICE-SLAM [55] causes severe artifacts and blurs, CoSLAM [41] and ESLAM [11] generate blur around the image boundaries." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.618, + 0.889, + 0.87 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.873, + 0.89, + 0.901 + ], + "angle": 0, + "content": "Figure 4. Reconstruction performance comparison of the proposed GS-SLAM and SOTA methods on the Replica dataset." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19600" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.086, + 0.895, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.387, + 0.895, + 0.416 + ], + "angle": 0, + "content": "Figure 5. The render visualization results on the Replica dataset of the proposed GS-SLAM and SOTA methods. GS-SLAM can generate much more high-quality and realistic images than the other methods, especially around the object boundaries." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.422, + 0.251, + 0.437 + ], + "angle": 0, + "content": "4.4. Runtime Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.445, + 0.473, + 0.763 + ], + "angle": 0, + "content": "Tab. 4 and Tab. 5 illustrate the runtime and memory usage of GS-SLAM and the state-of-the-art methods on the Replica and TUM-RGBD, respectively. We report the parameters of the neural networks and the memory usage of the scene representation. Note that Point-SLAM uses extra memory dynamic radius to improve performance (mark as \\(\\dagger\\)). The results show that GS-SLAM achieves a competitive running speed with 8.34 FPS compared to the other Radiance Fields-based vSLAMs. Note that we do not use any neural network decoder in our system, which results in the zero learnable parameter. However, the 3D Gaussian scene representations of GS-SLAM consume 198.04 MB memory, \\(4\\times\\) larger than the second large method NICE-SLAM [55]. Memory usage is mainly caused by spherical harmonic coefficients in training, which is a common constraint among Gaussian splatting-based reconstruction methods. Despite this, we still achieve a \\(20\\times\\) faster FPS compared to the similar point-based method Point-SLAM [27]. Besides, we also provide a light version of GS-SLAM with zero-order spherical harmonic coefficients, significantly reducing memory usage while maintaining stable performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.772, + 0.231, + 0.789 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.841 + ], + "angle": 0, + "content": "We perform the ablation of GS-SLAM on the Replica dataset #Room0 subset to evaluate the effectiveness of coarse-to-fine tracking, and expansion mapping strategy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Effect of our expansion strategy for mapping. Tab. 7 shows the ablation of our proposed expansion strategy for mapping. The results illustrate that the expansion strategy can significantly improve the tracking and mapping per" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.425, + 0.892, + 0.466 + ], + "angle": 0, + "content": "Table 4. Runtime and memory usage on Replica #Room0. The decoder parameters and embedding denote the parameter number of MLPs and the memory usage of the scene representation." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.469, + 0.891, + 0.556 + ], + "angle": 0, + "content": "
MethodTracking [ms×it] ↓Mapping [ms×it] ↓System FPS ↑Decoder param ↓Scene Embedding↓
Point-SLAM [27]0.06 × 4034.81 × 3000.420.127 M55.42 (+12453.2)†MB
NICE-SLAM [55]6.64 × 1028.63 × 602.910.06 M48.48 MB
Vox-Fusion [48]0.03 × 3066.53 × 101.280.054 M1.49 MB
CoSLAM [55]6.01 × 1013.18 × 1016.641.671 M—
ESLAM [11]6.85 × 819.87 × 1513.420.003 M27.12 MB
GS-SLAM11.9 × 1012.8 × 1008.340 M198.04 MB
" + }, + { + "type": "table_caption", + "bbox": [ + 0.501, + 0.562, + 0.892, + 0.59 + ], + "angle": 0, + "content": "Table 5. Runtime and memory usage on TUM-RGBD dataset #fr1Desk and #fr2_xyz." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.593, + 0.891, + 0.665 + ], + "angle": 0, + "content": "
Method#fr1Desk#fr2.xyz
FPS ↑Memory ↓FPS ↑Memory ↓
Point-SLAM0.1018.3 (+160.7)†MB0.1214.2 (+7687.4)†MB
NICE-SLAM [55]0.11178.8MB0.12484.0MB
ESLAM [11]0.3127.2MB0.3151.6MB
GS-SLAM1.83(ATE:3.3)40.8MB1.51(ATE:1.3)48.4MB
GS-SLAM (light)1.92(ATE:4.3)18.8MB1.68(ATE:2.7)22.3MB
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.901 + ], + "angle": 0, + "content": "formance. The implementation w/o adding means that we only initialize 3D Gaussians in the first frame and optimize the scene without adding new points. However, this strategy completely crashes because the density control in [13] can not handle real-time mapping tasks without an accurate point cloud input. Besides, the implementation w/o deletion suffers from a large number of redundant and noisy 3D Gaussian, which causes undesirable supervision. In contrast, the proposed expansion strategy effectively improves the tracking and mapping performance by 0.1 in ATE and 11.97 in Recall by adding more accurate constraints for the optimization. According to the visualization results in Fig. 6, our full implementation achieves more high-quality and detailed rendering and reconstruction results than the w/o delete strategy." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19601" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.071, + 0.891, + 0.099 + ], + "angle": 0, + "content": "Table 6. Rendering performance on Replica dataset. We outperform existing dense neural RGB-D methods on the commonly reported rendering metrics. Note that GS-SLAM achieves 386 FPS on average, benefiting from the efficient Gaussian scene representation." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.101, + 0.895, + 0.299 + ], + "angle": 0, + "content": "
MethodMetricRoom 0Room 1Room 2Office 0Office 1Office 2Office 3Office 4Avg.FPS.
NICE-SLAM [55]PSNR [dB] ↑22.1222.4724.5229.0730.3419.6622.2324.9424.42
SSIM ↑0.6890.7570.8140.8740.8860.7970.8010.8560.8090.30
LPIPS ↓0.3300.2710.2080.2290.1810.2350.2090.1980.233
Vox-Fusion* [48]PSNR [dB] ↑22.3922.3623.9227.7929.8320.3323.4725.2124.413.88
SSIM ↑0.6830.7510.7980.8570.8760.7940.8030.8470.801
LPIPS ↓0.3030.2690.2340.2410.1840.2430.2130.1990.236
CoSLAM [41]PSNR [dB] ↑27.2728.4529.0634.1434.8728.4328.7630.9130.243.68
SSIM ↑0.9100.9090.9320.9610.9690.9380.9410.9550.939
LPIPS ↓0.3240.2940.2660.2090.1960.2580.2290.2360.252
ESLAM [11]PSNR [dB] ↑25.3227.7729.0833.7130.2028.0928.7729.7129.082.82
SSIM ↑0.8750.9020.9320.9600.9230.9430.9480.9450.929
LPIPS ↓0.3130.2980.2480.1840.2280.2410.1960.2040.336
OursPSNR [dB] ↑31.5632.8632.5938.7041.1732.3632.0332.9234.27386.91
SSIM ↑0.9680.9730.9710.9860.9930.9780.9700.9680.975
LPIPS ↓0.0940.0750.0930.0500.0330.0940.1100.1120.082
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.303, + 0.468, + 0.329 + ], + "angle": 0, + "content": "Table 7. Ablation of the adaptive 3D Gaussian expansion strategy on Replica #Room0." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.33, + 0.456, + 0.389 + ], + "angle": 0, + "content": "
Setting#Room0
ATE↓Depth L1↓Precision↑Recall ↑F1↑PSNR↑SSIM↑LPIPS↓
w/o addXXXXXXXX
w/o delete0.581.6853.5549.3251.3531.220.9670.094
w/ add & delete0.481.3164.5861.2962.8931.560.9680.094
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.392, + 0.468, + 0.417 + ], + "angle": 0, + "content": "Table 8. Ablation of the coarse-to-fine tracking strategy on Replica #Room0." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.418, + 0.456, + 0.475 + ], + "angle": 0, + "content": "
Setting#Room0
ATE↓Depth L1↓Precision↑Recall ↑F1↑PSNR↑SSIM↑LPIPS↓
Coarse0.911.4859.6857.5456.5029.130.9540.120
Fine0.491.3962.6159.1861.2930.840.9640.096
Coarse-to-fine0.481.3164.5861.2962.8931.560.9680.094
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.481, + 0.47, + 0.648 + ], + "angle": 0, + "content": "Effect of coarse-to-fine tracking. According to the results in Tab. 8, the proposed coarse-to-fine tracking strategy performs best in all tracking, mapping, and rendering metrics. Compared with fine tracking, the coarse-to-fine tracking strategy significantly improves the performance by 0.01 in tracking ATE, 2.11 in Recall, and 0.72 in PSNR. Although the fine strategy surpasses the coarse strategy in precision, it suffers from the artifacts and noise in the reconstructed scene, leading to a fluctuation optimization. The coarse-to-fine strategy effectively avoids noise reconstruction and improves accuracy and robustness." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.657, + 0.362, + 0.673 + ], + "angle": 0, + "content": "4.6. Efficiency-to-accuracy trade-off." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.68, + 0.469, + 0.831 + ], + "angle": 0, + "content": "3DGS-based SLAM trade-off focuses not only on the efficiency of tracking and mapping but also emphasizes high-quality ultra-real-time rendering. As shown in Fig. 7b, GS-SLAM achieves \\(\\approx 400\\) FPS ultra-fast speed and highest PSNR in map rendering. At the same time, our method remains a competitive system FPS and lowest tracking ATE in Fig. 7a. Moreover, GS-SLAM shows great potential in memory reduction in Tab. 5, and comparable in mesh reconstruction. Note that baselines directly use 3DGS in Fig. 7, resulting in inferior performances." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.845, + 0.334, + 0.861 + ], + "angle": 0, + "content": "5. Conclusion and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.87, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We introduced GS-SLAM, a novel dense visual SLAM method leveraging 3D Gaussian Splatting for efficient map" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.301, + 0.699, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.302, + 0.865, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.46, + 0.891, + 0.487 + ], + "angle": 0, + "content": "Figure 6. Rendering and mesh visualization of the adaptive 3D Gaussian expansion ablation on Replica #Room0." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.493, + 0.681, + 0.612 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.523, + 0.612, + 0.659, + 0.623 + ], + "angle": 0, + "content": "(a) Tracking performance" + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.494, + 0.888, + 0.612 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.736, + 0.612, + 0.862, + 0.623 + ], + "angle": 0, + "content": "(b) Render performance" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.626, + 0.891, + 0.653 + ], + "angle": 0, + "content": "Figure 7. Bi-criteria figure of tracking/render performance and system FPS on Replica #Office0." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.892, + 0.794 + ], + "angle": 0, + "content": "ping and accurate camera pose estimation, striking a better speed-accuracy balance. However, its reliance on high-quality depth data may limit performance in certain conditions. Additionally, the approach's high memory requirements for large scenes suggest future improvements could focus on optimizing memory use, potentially via techniques such as quantization and clustering. We believe GS-SLAM has the potential to extend to larger scale with some improvements and will explore this in future work." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgements. This work is supported by the Shanghai AI Laboratory, National Key R&D Program of China (2022ZD0160101), the National Natural Science Foundation of China (62376222), Young Elite Scientists Sponsorship Program by CAST (2023QNRC001) and the Early Career Scheme of the Research Grants Council (RGC) of the Hong Kong SAR under grant No. 26202321." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19602" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Michael Bloesch, Jan Czarnowski, Ronald Clark, Stefan Leutenegger, and Andrew J. Davison. Codeslam - learning a compact, optimisable representation for dense visual slam. CVPR, pages 2560-2568, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.226 + ], + "angle": 0, + "content": "[2] Guillaume Bresson, Zayed Alsayed, Li Yu, and Sébastien Glaser. Simultaneous localization and mapping: A survey of current trends in autonomous driving. TIV, 2:194-220, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.229, + 0.47, + 0.257 + ], + "angle": 0, + "content": "[3] Zilong Chen, Feng Wang, and Huaping Liu. Text-to-3d using gaussian splatting. ArXiv, abs/2309.16585, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.47, + 0.299 + ], + "angle": 0, + "content": "[4] Brian Curless and Marc Levoy. Volumetric method for building complex models from range images. In SIGGRAPH, ACM, 1996. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.302, + 0.47, + 0.343 + ], + "angle": 0, + "content": "[5] Parth Rajesh Desai, Pooja Nikhil Desai, Komal Deepak Ajmera, and Khushbu Mehta. A review paper on oculus rift-a virtual reality headset. ArXiv, abs/1408.1173, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.345, + 0.47, + 0.373 + ], + "angle": 0, + "content": "[6] Hugh F. Durrant-Whyte and Tim Bailey. Simultaneous localization and mapping: part i. RAM, 13:99–110, 2006. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.375, + 0.47, + 0.415 + ], + "angle": 0, + "content": "[7] Christian Hane, Christopher Zach, Jongwoo Lim, Ananth Ranganathan, and Marc Pollefeys. Stereo depth map fusion for robot navigation. IROS, pages 1618-1625, 2011. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.417, + 0.47, + 0.473 + ], + "angle": 0, + "content": "[8] Caner Hazirbas, Andreas Wiedemann, Robert Maier, Laura Leal-Taixe, and Daniel Cremers. Tum rgb-d scribble-based segmentation benchmark. https://github.com/tum-vision/rgbdscribble_benchmark, 2018.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.475, + 0.47, + 0.515 + ], + "angle": 0, + "content": "[9] Jiahui Huang, Shi-Sheng Huang, Haoxuan Song, and ShiMin Hu. Di-fusion: Online implicit 3d reconstruction with deep priors. In CVPR, pages 8932-8941, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.517, + 0.47, + 0.558 + ], + "angle": 0, + "content": "[10] Xudong Jiang, Lifeng Zhu, Jia Liu, and Aiguo Song. A slam-based 6 dof controller with smooth auto-calibration for virtual reality. TVC, 39:3873 - 3886, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.56, + 0.47, + 0.614 + ], + "angle": 0, + "content": "[11] Mohammad Mahdi Johari, Camilla Carta, and Francois Fleuret. Eslam: Efficient dense slam system based on hybrid representation of signed distance fields. CVPR, 2023. 1, 2, 5, 6, 7, 8, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.617, + 0.47, + 0.671 + ], + "angle": 0, + "content": "[12] Olaf Kähler, Victor Adrian Prisacariu, Julien P. C. Valentin, and David William Murray. Hierarchical voxel block hashing for efficient integration of depth images. RAL, 1:192-197, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.673, + 0.47, + 0.715 + ], + "angle": 0, + "content": "[13] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. TOG, 42(4), 2023. 2, 3, 7, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.47, + 0.757 + ], + "angle": 0, + "content": "[14] Leonid Keselman and Martial Hebert. Approximate differentiable rendering with algebraic surfaces. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.759, + 0.47, + 0.8 + ], + "angle": 0, + "content": "[15] Leonid Keselman and Martial Hebert. Flexible techniques for differentiable rendering with 3d gaussians. arXiv preprint arXiv:2308.14737, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[16] Georg S. W. Klein and David William Murray. Parallel tracking and mapping on a camera phone. ISMAR, pages 83-86, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.899 + ], + "angle": 0, + "content": "[17] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. ArXiv, abs/2308.09713, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[18] Robert Maier, Raphael Schaller, and Daniel Cremers. Efficient online surface correction for real-time large-scale 3d reconstruction. ArXiv, abs/1709.03763, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[19] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.191, + 0.892, + 0.231 + ], + "angle": 0, + "content": "[20] Raul Mur-Artal and Juan D. Tardós. Orb-slam2: An opensource slam system for monocular, stereo, and rgb-d cameras. *TRO*, 33:1255–1262, 2016. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.233, + 0.892, + 0.314 + ], + "angle": 0, + "content": "[21] Richard A. Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J. Davison, Pushmeet Kohli, Jamie Shotton, Steve Hodges, and Andrew William Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. ISMAR, pages 127-136, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.316, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[22] Richard A. Newcombe, S. Lovegrove, and Andrew J. Davison. Dtam: Dense tracking and mapping in real-time. ICCV, pages 2320-2327, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.358, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[23] Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Marc Stamminger. Real-time 3d reconstruction at scale using voxel hashing. TOG, 32:1-11, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.401, + 0.892, + 0.439 + ], + "angle": 0, + "content": "[24] Delin Qu, Chi Yan, Dong Wang, Jie Yin, Dan Xu, Bin Zhao, and Xuelong Li. Implicit event-rgbd neural slam. CVPR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.442, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[25] Gerhard Reitmayr, Tobias Langlotz, Daniel Wagner, Alessandro Mulloni, Gerhard Schall, Dieter Schmalstieg, and Qi Pan. Simultaneous localization and mapping for augmented reality. ISUVR, pages 5-8, 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.498, + 0.892, + 0.566 + ], + "angle": 0, + "content": "[26] Fabio Ruetz, Emili Hernández, Mark Pfeiffer, Helen Oleynikova, Mark Cox, Thomas Lowe, and Paulo Vinicius Koerich Borges. Ovpc mesh: 3d free-space representation for local ground vehicle navigation. ICRA, pages 8648-8654, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.568, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[27] Erik Sandström, Yue Li, Luc Van Gool, and Martin R. Oswald. Point-slam: Dense neural point cloud-based slam. In ICCV, 2023. 2, 5, 6, 7, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.609, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[28] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.637, + 0.892, + 0.677 + ], + "angle": 0, + "content": "[29] Thomas Schöps, Torsten Sattler, and Marc Pollefeys. Bad slam: Bundle adjusted direct rgb-d slam. CVPR, pages 134-144, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.679, + 0.892, + 0.719 + ], + "angle": 0, + "content": "[30] Thomas Schops, Torsten Sattler, and Marc Pollefeys. BAD SLAM: Bundle adjusted direct RGB-D SLAM. In CVPR, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.721, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[31] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Yuheng Ren, Shobhit Verma, Anton Clarkson, Ming Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Briales, Tyler Gillingham, Elias Mueggler, Luis Pesqueira, Manolis Savva, Dhruv Batra, Hauke Malte Strasdat, Renzo De Nardi, Michael Goesele, S. Lovegrove, and Richard A. Newcombe. The replica dataset: A digital replica of indoor spaces. ArXiv, abs/1906.05797, 2019. 5, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[32] J. Stuckler and Sven Behnke. Multi-resolution surfel maps for efficient dense 3d modeling and tracking. JVCIR, 25: 137-147, 2014. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19603" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[33] Jürgen Sturm, Nikolas Engelhard, Felix Endres, Wolfram Burgard, and Daniel Cremers. A benchmark for the evaluation of RGB-D SLAM systems. In IROS. IEEE/RSJ, 2012. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.468, + 0.19 + ], + "angle": 0, + "content": "[34] Edgar Sucar, Kentaro Wada, and Andrew J. Davison. Nodeslam: Neural object descriptors for multi-view shape reconstruction. 3DV, pages 949–958, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.468, + 0.233 + ], + "angle": 0, + "content": "[35] Edgar Sucar, Shikun Liu, Joseph Ortiz, and Andrew J. Davison. imap: Implicit mapping and positioning in real-time. ICCV, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.468, + 0.288 + ], + "angle": 0, + "content": "[36] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. ArXiv, abs/2309.16653, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.468, + 0.319 + ], + "angle": 0, + "content": "[37] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. In NIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.468, + 0.362 + ], + "angle": 0, + "content": "[38] Andreas Langeland Teigen, Yeonsoo Park, Annette Stahl, and Rudolf Mester. Rgb-d mapping and tracking in a plenoxel radiance field. ArXiv, abs/2307.03404, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.363, + 0.468, + 0.417 + ], + "angle": 0, + "content": "[39] Charalambos Theodorou, Vladan Velisavljevic, Vladimir Dyo, and Fredi Nonyelu. Visual slam algorithms and their application for ar, mapping, localization and wayfinding. Array, 15:100222, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.419, + 0.468, + 0.475 + ], + "angle": 0, + "content": "[40] Angtian Wang, Peng Wang, Jian Sun, Adam Kortylewski, and Alan Yuille. Voge: a differentiable volume renderer using gaussian ellipsoids for analysis-by-synthesis. arXiv preprint arXiv:2205.15401, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.476, + 0.468, + 0.517 + ], + "angle": 0, + "content": "[41] Hengyi Wang, Jingwen Wang, and Lourdes de Agapito. Coslam: Joint coordinate and sparse parametric encodings for neural real-time slam. CVPR, 2023. 1, 5, 6, 8, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.519, + 0.468, + 0.558 + ], + "angle": 0, + "content": "[42] Kaixuan Wang, Fei Gao, and Shaojie Shen. Real-time scalable dense surfel mapping. ICRA, pages 6919-6925, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.562, + 0.468, + 0.603 + ], + "angle": 0, + "content": "[43] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. TIP, 13(4):600-612, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.604, + 0.468, + 0.657 + ], + "angle": 0, + "content": "[44] Thomas Whelan, Michael Kaess, Maurice F. Fallon, Hordur Johannsson, John J. Leonard, and John B. McDonald. Kintinous: Spatially extended kinectfusion. In AAAI, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.468, + 0.715 + ], + "angle": 0, + "content": "[45] Thomas Whelan, John McDonald, Michael Kaess, Maurice Fallon, Hordur Johannsson, and John J. Leonard. Kintinuous: Spatially extended kinectfusion. In RSS Workshop on RGB-D, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.468, + 0.759 + ], + "angle": 0, + "content": "[46] Thomas Whelan, Stefan Leutenegger, Renato Salas-Moreno, Ben Glocker, and Andrew Davison. Elasticfusion: Dense slam without a pose graph. In RSS, 2015. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.468, + 0.814 + ], + "angle": 0, + "content": "[47] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. ArXiv, abs/2310.08528, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.468, + 0.871 + ], + "angle": 0, + "content": "[48] Xingrui Yang, Hai Li, Hongjia Zhai, Yuhang Ming, Yuqian Liu, and Guofeng Zhang. Vox-fusion: Dense tracking and mapping with voxel-based neural implicit representation. IS-MAR, pages 499-507, 2022. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[49] Ziyi Yang, Xinyu Gao, Wenming Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.89, + 0.12 + ], + "angle": 0, + "content": "for high-fidelity monocular dynamic scene reconstruction. ArXiv, abs/2309.13101, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[50] Zeyu Yang, Hongye Yang, Zijie Pan, Xiatian Zhu, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. ArXiv, abs/2310.10642, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[51] Taoran Yi, Jiemin Fang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussiandreamer: Fast generation from text to 3d gaussian splattering with point cloud priors. ArXiv, abs/2310.08529, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.235, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[52] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.291, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[53] Youmin Zhang, Fabio Tosi, Stefano Mattoccia, and Matteo Poggi. Go-slam: Global optimization for consistent 3d instant reconstruction. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[54] Shuaifeng Zhi, Michael Bloesch, Stefan Leutenegger, and Andrew J. Davison. Scenocode: Monocular dense semantic reconstruction using learned encoded scene representations. CVPR, pages 11768-11777, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.391, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[55] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R. Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. CVPR, 2021. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.447, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[56] Wojciech Zielonka, Timur M. Bagautdinov, Shunsuke Saito, Michael Zollhofer, Justus Thies, and Javier Romero. Drivable 3d gaussian avatars. 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.49, + 0.892, + 0.532 + ], + "angle": 0, + "content": "[57] Matthias Zwicker, Hanspeter Pfister, Jeroen van Baar, and Markus H. Gross. Ewa volume splatting. Proceedings Visualization, 2001. VIS '01., pages 29-538, 2001. 3" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.532 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19604" + } + ] +] \ No newline at end of file diff --git a/2024/GS-SLAM_ Dense Visual SLAM with 3D Gaussian Splatting/93048e35-cf42-49e6-8897-d6803305631d_origin.pdf b/2024/GS-SLAM_ Dense Visual SLAM with 3D Gaussian Splatting/93048e35-cf42-49e6-8897-d6803305631d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..12fec5222b272719fd8f3590fd955eabc8d4d390 --- /dev/null +++ b/2024/GS-SLAM_ Dense Visual SLAM with 3D Gaussian Splatting/93048e35-cf42-49e6-8897-d6803305631d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8df976cc5594f2874ea1f4451909ffbbf7c53a9bb1af98111b024131ac7b0386 +size 13829510 diff --git a/2024/GS-SLAM_ Dense Visual SLAM with 3D Gaussian Splatting/full.md b/2024/GS-SLAM_ Dense Visual SLAM with 3D Gaussian Splatting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..10f23a815a32640819635ca34b48ff4162008171 --- /dev/null +++ b/2024/GS-SLAM_ Dense Visual SLAM with 3D Gaussian Splatting/full.md @@ -0,0 +1,344 @@ +# GS-SLAM: Dense Visual SLAM with 3D Gaussian Splatting + +Chi Yan $^{1,3*}$ + +Delin $\mathbf{Q}\mathbf{u}^{1,2*}$ + +Dan $\mathrm{Xu}^3$ + +Bin Zhao $^{1,4}$ + +Zhigang Wang + +Dong + +Xuelong Li $^{1,5}$ + +$^{1}$ Shanghai AI Laboratory $^{2}$ Fudan University $^{4}$ Northwestern Polytechnical University + +$^{3}$ Hong Kong University of Science and Technology $^{5}$ TeleAI, China Telecom Corp Ltd + +# Abstract + +In this paper, we introduce GS-SLAM that first utilizes 3D Gaussian representation in the Simultaneous Localization and Mapping (SLAM) system. It facilitates a better balance between efficiency and accuracy. Compared to recent SLAM methods employing neural implicit representations, our method utilizes a real-time differentiable splatting rendering pipeline that offers significant speedup to map optimization and RGB-D rendering. Specifically, we propose an adaptive expansion strategy that adds new or deletes noisy 3D Gaussians in order to efficiently reconstruct new observed scene geometry and improve the mapping of previously observed areas. This strategy is essential to extend 3D Gaussian representation to reconstruct the whole scene rather than synthesize a static object in existing methods. Moreover, in the pose tracking process, an effective coarse-to-fine technique is designed to select reliable 3D Gaussian representations to optimize camera pose, resulting in runtime reduction and robust estimation. Our method achieves competitive performance compared with existing state-of-the-art real-time methods on the Replica, TUM-RGBD datasets. Project page: https://gs-slam.github.io/. + +# 1. Introduction + +Simultaneous localization and mapping (SLAM) has emerged as a pivotal technology in fields such as robotics [6], virtual reality [10], and augmented reality [25, 39]. The goal of SLAM is to construct a dense/sparse map of an unknown environment while simultaneously tracking the camera pose. Traditional SLAM methods employ point/surfel clouds [20, 32, 42, 46], mesh representations [26], voxel hashing [12, 18, 23] or voxel grids [21] as scene representations to construct dense mapping, and have made considerable progress on localization accuracy. However, these methods face serious challenges in obtaining fine-grained dense maps. + +![](images/22fe11d95c4291fb95410f5b65e9fc36f5340c37a7d6b3c833bfdee4c95ddd87.jpg) +Figure 1. The illustration of the proposed GS-SLAM. It first utilizes the 3D Gaussian representation and differentiable splatting rasterization pipeline in SLAM, achieving real-time tracking and mapping performance on GPU. Besides, benefiting from the splatting rasterization pipeline, GS-SLAM achieves a $100\times$ faster rendering FPS and more high-quality full image results than the other SOTA methods. + +Recently, Neural Radiance Fields (NeRF) [19] have been explored to enhance SLAM methodologies and exhibit strengths in generating high-quality, dense maps with low memory consumption [35]. In particular, iMAP [35] uses a single multi-layer perceptron (MLP) to represent the entire scene, which is updated globally with the loss between volume-rendered RGB-D image and ground-truth observations. NICE-SLAM [55] utilizes a hierarchical neural implicit grid as scene map representation to allow local updates for reconstructing large scenes. Moreover, ESLAM [11], CoSLAM [41] and EN-SLAM [24] utilize axis-aligned feature planes and joint coordinate-parametric encoding to improve the capability of scene representation, achieving efficient and high-quality surface map reconstruction. In practical mapping and tracking steps, these methods only render a small set of pixels to reduce optimization time, which leads to the reconstructed dense maps lacking + +the richness and intricacy of details. In essence, it is a tradeoff for the efficiency and accuracy of NeRF-based SLAM since obtaining high-resolution images with the ray-based volume rendering technique is time-consuming and unacceptable. + +Fortunately, recent work [13, 17, 47] with 3D Gaussian representation and tile-based splatting techniques has shown great superiority in the efficiency of high-resolution image rendering. It is applied to synthesize novel view RGB images of static objects, achieving state-of-the-art visual quality for $1080\mathrm{p}$ resolution at real-time speed. Inspired by this, we extend the rendering superiority of 3D Gaussian scene representation and real-time differentiable splatting rendering pipeline for the task of dense RGB-D SLAM and manage to jointly promote the speed and accuracy of NeRF-based dense SLAM, as shown in Fig. 1. + +To this end, we propose GS-SLAM, the first RGB-D dense SLAM system that first utilizes 3D Gaussian scene representation coupled with the splatting rendering technique to better balance speed and accuracy. Our system optimizes camera tracking and mapping with a novel RGB-D rendering approach that processes 3D Gaussians quickly and accurately through sorting and $\alpha$ -blending. We enhance scene reconstruction by introducing an adaptive strategy for managing 3D Gaussian elements, which optimizes mapping by focusing on current observations and minimizes errors in dense maps and images. Moreover, we propose a coarse-to-fine approach, starting with low-resolution image analysis for initial pose estimation and refining it with high-resolution rendering using select 3D Gaussians, to boost speed and accuracy. We perform extensive evaluations on a selection of indoor RGB-D datasets and demonstrate state-of-the-art performance on dense neural RGB-D SLAM in terms of tracking, rendering, and mapping. Overall, our contributions include: + +- We propose GS-SLAM, the first 3D Gaussian Splatting(3DGS)-based dense RGB-D SLAM approach, which takes advantage of the fast splatting rendering technique to boost the mapping optimizing and pose tracking, achieving real-time and photo-realistic reconstruction performance. +- We present an adaptive 3D Gaussian expansion strategy to efficiently reconstruct new observed scene geometry and develop a coarse-to-fine technique to select reliable 3D Gaussians to improve camera pose estimation. +- Our approach achieves competitive performance on Replica and TUM-RGBD datasets in terms of tracking, and mapping and runs at 8.43 FPS, resulting in a better balance between efficiency and accuracy. + +# 2. Related Work + +- Dense Visual SLAM. The existing real-time dense visual SLAM systems are typically based on discrete handcrafted + +features or deep-learning embeddings, and follow the mapping and tracking architecture in [16]. DTAM [22] first introduces a dense SLAM system that uses photometric consistency to track a handheld camera and represent the scene as a cost volume. KinectFusion [44] performs camera tracking by iterative-closest-point and updates the scene via TSDF-Fusion. BAD-SLAM [29] proposes to jointly optimize the keyframe poses and 3D scene geometry via a direct bundle adjustment (BA) technique. In contrast, recent works integrate deep learning with the traditional geometry framework for more accurate and robust camera tracking and mapping, such as DROID-SLAM [37], CodeSLAM [1], SceneCode [54], and NodeSLAM [34], have made significant advances in the field, achieving more accurate and robust camera tracking and mapping performance. + +Neural Implicit Radiance Field based SLAM. For NeRF-based SLAM, existing methods can be divided into three main types: MLP-based methods, Hybrid representation methods, and Explicit methods. MLP-based method iMAP [35] offers scalable and memory-efficient map representations but faces challenges with catastrophic forgetting in larger scenes. Hybrid representation methods combine the advantages of implicit MLPs and structure features, significantly enhancing the scene scalability and precision. For example, NICE-SLAM [55] integrates MLPs with multi-resolution voxel grids, enabling large scene reconstruction, and Vox-Fusion [48] employs octree expansion for dynamic map scalability, while ESLAM [11] and Point-SLAM [27] utilize tri-planes and neural point clouds respectively to improve the mapping capability. As for the explicit method proposed in [38], it stores map features in voxel directly, without any MLPs, enabling faster optimization. Instead of representing maps with implicit features, GS-SLAM utilizes the 3D Gaussian representation, efficiently renders images using splatting-based rasterization, and optimizes parameters directly with backward propagation. + +3D Gaussian Representation. Several recent approaches have used 3D Gaussians for shape reconstruction, such as Fuzzy Metaballs [14, 15], VoGE [40], 3DGS [13]. Notably, 3DGS [13] demonstrates great superiorities in high-quality real-time novel-view synthesis. This work represents the scene with 3D Gaussians and develops a NeRF-style fast rendering algorithm to support anisotropic splating, achieving SOTA visual quality and fast high-resolution rendering performance. Beyond the rendering superiorities, Gaussian splating holds an explicit geometry scene structure and appearance, benefiting from the exact modeling of scenes representation [50]. This promising technology has been rapidly applied in several fields, including 3D generation [3, 36, 51], dynamic scene modeling [17][47][49], and photorealistic drivable avatars [56]. However, currently, there is no research addressing camera pose estimation or real-time mapping using 3D Gaussian models due to the in + +![](images/53be036ebc19fd8fa446cd8da14749f0bad356f997a5cb2bd86001c01bec7a5d.jpg) +Figure 2. Overview of the proposed method. We aim to use 3D Gaussians to represent the scene and use the rendered RGB-D image for inverse camera tracking. GS-SLAM proposes a novel Gaussian expansion strategy to make the 3D Gaussian feasible to reconstruct the whole scene and can achieve real-time tracking, mapping, and rendering performance on GPU. + +herent limitations of the prime pipeline [13], i.e., prerequisites of initialized point clouds or camera pose inputs [28]. In contrast, we derive the analytical derivative equations for pose estimation in the Gaussian representation and implement efficient CUDA optimization. + +# 3. Methodology + +Fig. 2 shows the overview of the proposed GS-SLAM. We aim to estimate the camera poses $\{\mathbf{P}_i\}_{i=1}^N$ of every frame and simultaneously reconstruct a dense scene map by giving an input sequential RGB-D stream $\{\mathbf{I}_i, \mathbf{D}_i\}_{i=1}^M$ with known camera intrinsic $\mathbf{K} \in \mathbb{R}^{3 \times 3}$ . In Sec. 3.1, we first introduce 3D Gaussian as the scene representation $\mathbf{S}$ and the RGB-D render by differentiable splattering rasterization. With the estimated camera pose of the keyframe, in Sec. 3.2, an adaptive expansion strategy is proposed to add new or delete noisy 3D Gaussians to efficiently reconstruct new observed scene geometry while improving the mapping of the previously observed areas. For camera tracking of every input frame, we derive an analytical formula for backward optimization with rendering RGB-D loss. We further introduce an effective coarse-to-fine technique to minimize rendering losses to achieve efficient and accurate pose estimation in Sec. 3.3. + +# 3.1. 3D Gaussian Scene Representation + +Our goal is to optimize a scene representation that captures the geometry and appearance of the scene, resulting in a detailed dense map and high-quality novel view synthesis. To do this, we model the scene as a set of 3D Gaussians coupled with opacity and spherical harmonics + +$$ +\mathbf {G} = \left\{G _ {i}: \left(\mathbf {X} _ {i}, \boldsymbol {\Sigma} _ {i}, \Lambda_ {i}, \mathbf {Y} _ {i}\right) | i = 1, \dots , N \right\}. \tag {1} +$$ + +Each 3D Gaussian scene representation $G_{i}$ is defined by position $\mathbf{X}_i\in \mathbb{R}^3$ , 3D covariance matrix $\pmb{\Sigma}_{i}\in \mathbb{R}^{3\times 3}$ , opacity $\Lambda_{i}\in \mathbb{R}$ and 1-degree spherical harmonics $(\mathbf{Y})$ per color channel, a total of 12 coefficients for $\mathbf{Y}_i\in \mathbb{R}^{12}$ . To reduce + +the learning difficulty of the 3D Gaussians [57], we parameterize the 3D Gaussian's covariance as: + +$$ +\boldsymbol {\Sigma} = \mathbf {R} \mathbf {S} \mathbf {S} ^ {T} \mathbf {R} ^ {T}, \tag {2} +$$ + +where $\mathbf{S} \in \mathbb{R}^3$ is a 3D scale vector, $\mathbf{R} \in \mathbb{R}^{3 \times 3}$ is the rotation matrix, storing as a 4D quaternion. + +Color and depth splatting rendering. With the optimized 3D Gaussian scene representation parameters, given the camera pose $\mathbf{P} = \{\mathbf{R},\mathbf{t}\}$ , the 3D Gaussians $G$ are projected into 2D image plane for rendering with: + +$$ +\boldsymbol {\Sigma} ^ {\prime} = \mathbf {J} \mathbf {P} ^ {- 1} \boldsymbol {\Sigma} \mathbf {P} ^ {- T} \mathbf {J} ^ {T}, \tag {3} +$$ + +where $\mathbf{J}$ is the Jacobian of the affine approximation of the projective function. After projecting 3D Gaussians to the image plane, the color of one pixel is rendered by sorting the Gaussians in depth order and performing front-to-back $\alpha$ -blending rendering as follows: + +$$ +\hat {\mathbf {C}} = \sum_ {i \in N} \mathbf {c} _ {i} \alpha_ {i} \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {j}), \tag {4} +$$ + +where $\mathbf{c}_i$ represents the color of the $i$ -th 3D Gaussian obtained by learned spherical harmonics coefficients $\mathbf{Y}$ , $\alpha_i$ is the density computed by learned opacity $\Lambda_i$ and 2D Gaussian with covariance $\Sigma'$ . Similarly, the depth is rendered by + +$$ +\hat {D} = \sum_ {i \in N} d _ {i} \alpha_ {i} \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {j}), \tag {5} +$$ + +where $d_{i}$ denotes the depth of the center of the $i$ -th 3D Gaussian, which is obtained by projecting to $z$ -axis in the camera coordinate. + +# 3.2. Adaptive 3D Gaussian Expanding Mapping + +The 3D Gaussian scene representations are updated and optimized on each selected keyframe for stable mapping. Given the estimated pose of each selected keyframe, we first + +apply the proposed adaptive expansion strategy to add new or delete noisy 3D Gaussians from the whole scene representations to render RGB-D images with resolution $H \times W$ , and then the updated 3D Gaussian scene representations are optimized by minimizing the geometric depth loss $\mathcal{L}_d$ and the photometric color loss $\mathcal{L}_{\mathrm{c}}$ to the sensor observation depth $D$ and color $\mathbf{C}$ , + +$$ +\mathcal {L} _ {\mathbf {c}} = \sum_ {m = 1} ^ {H W} \left| \mathbf {C} _ {m} - \hat {\mathbf {C}} _ {m} \right|, \quad \mathcal {L} _ {d} = \sum_ {m = 1} ^ {H W} \left| D _ {m} - \hat {D} _ {m} \right|. \tag {6} +$$ + +The loss optimizes the parameters of all 3D Gaussians that contribute to the rendering of these keyframe images. + +Adaptive 3D Gaussian Expansion Strategy. At the first frame of the RGB-D sequence, we first uniformly sample half pixels from a whole image with $H \times W$ resolution and back-projecting them into 3D points $\mathbf{X}$ with corresponding depth observation $D$ . The 3D Gaussian scene representations are created by setting position as $\mathbf{X}$ and initializing zero degree Spherical Harmonics coefficients with RGB color $\mathbf{C}_i$ . The opacities are set to pre-defined values, and the covariance is set depending on the spatial point density, i.e., + +$$ +\left\{G _ {i} = \left(\mathbf {P} _ {i}, \boldsymbol {\Sigma} _ {\text {i n i t}}, \Lambda_ {\text {i n i t}}, \mathbf {C} _ {i}\right) | i = 1, \dots , M \right\}, \tag {7} +$$ + +where $M$ equals to $HW / 2$ . The 3D Gaussians are initialized and then optimized using the first RGB-D image with rendering loss. Note that only half of the pixels are used to initialize the scene, leaving space to conduct adaptive density control of Gaussians that splits large points into smaller ones and clones them with different directions to capture missing geometric details. + +Adding step: To obtain a complete map of the environment, the 3D Gaussian scene representations should be able to model the geometry and appearance of newly observed areas. Specifically, at every keyframe, we add first rendered RGB-D images using historical 3D Gaussians and calculate cumulative opacity $T = \sum_{i\in N}\alpha_i\prod_{j = 1}^{i - 1}(1 - \alpha_j)$ for each pixel. We label one pixel as un-reliable $x^{un}$ if its cumulative opacity $T$ is too low or its rendered depth $\hat{D}$ is far away from observed depth $D$ , i.e., + +$$ +T < \tau_ {T} \quad \text {o r} \quad | D - \hat {D} | > \tau_ {D}. \tag {8} +$$ + +These selected un-reliable pixels mostly capture new observed areas. Then we back-project these un-reliable pixels to 3D points $\mathbf{P}^{un}$ , and a set of new 3D Gaussians at $\mathbf{P}^{un}$ initialized as Eq. 7 are added into scene representations to model the new observed areas. + +**Deleting step:** As shown in Fig. 3, there are some floating 3D Gaussians due to the unstable adaptive control of Gaussians after optimization with Eq. 6. These floating 3D Gaussians will result in a low-quality dense map and a rendered image containing lots of artifacts. To address this + +![](images/d4d5a6ae1d5f38d2d2e048d028616ebf9d872090c545a009861906a8433fbe83.jpg) +Figure 3. Illustration of the proposed adaptive 3D Gaussian expansion strategy. GS-SLAM inhibits the low-quality 3D Gaussian floaters in the current frustum according to depth. + +issue, after adding new 3D Gaussians, we check all visible 3D Gaussians in the current camera frustum and significantly decrease opacity $\Lambda_{i}$ of 3D Gaussians whose position is not near the scene surfaces. Formally, for each visible 3D Gaussian, we draw a ray $\mathbf{r}(t)$ from camera origin $\mathbf{o}$ and its position $\mathbf{X}_i = (x_i,y_i,z_i)$ , i.e., $\mathbf{r}(t) = \mathbf{o} + t(\mathbf{X}_i - \mathbf{o})$ . Then, we can find a pixel with coordinate $(u,v)$ where this ray intersects the image plane and corresponding depth observation $D$ . The 3D Gaussians are deleted by degenerating its opacity as follows: + +$$ +G _ {i}: \Lambda_ {i} \Rightarrow G _ {i}: \eta \Lambda_ {i}, \text {i f} D - d i s t (\mathbf {X} _ {i}, \mathbf {P} _ {u v}) > \gamma , \tag {9} +$$ + +where $P_{uv}$ is the world coordinates of the intersected pixel calculated with the camera intrinsic and extrinsic. $dist(\cdot, \cdot)$ is the Euclidean distance, and $\eta$ (much smaller than 1) and $\gamma$ are the hyper-parameters. Note that we decrease the opacity of floating 3D Gaussians in front of the scene surfaces to make our newly added 3D Gaussians well-optimized. + +# 3.3. Tracking and Bundle Adjustment + +In the parallel camera tracking phase of our work, we first employ a common straightforward constant velocity assumption to initialize new poses. This assumption transforms the last known pose based on the relative transformation between the second-to-last pose and the last pose. Then, the accurate camera pose $\mathbf{P}$ is optimized by minimizing rendered color loss, i.e., + +$$ +\mathcal {L} _ {\text {t r a c k}} = \sum_ {m = 1} ^ {M} \left| \mathbf {C} _ {m} - \hat {\mathbf {C}} _ {m} \right| _ {1}, \quad \min _ {\mathbf {R}, \mathbf {t}} (\mathcal {L} _ {\text {t r a c k}}), \tag {10} +$$ + +where $M$ is the number of sampled pixels for rendering. + +Differentiable pose estimation. According to Eqs. (3) and (4), we observe that the gradient of the camera pose $\mathbf{P}$ is related to three intermediate variables: $\pmb{\Sigma}^{\prime}$ , $\mathbf{c}_i$ , and the projected coordinate $\mathbf{m}_i$ of Gaussian $G_{i}$ . By applying the chain rule of derivation, we obtain the analytical formula + +tion of camera pose P: + +$$ +\begin{array}{l} \frac {\partial \mathcal {L} _ {\mathbf {c}}}{\partial \mathbf {P}} = \frac {\partial \mathcal {L} _ {\mathbf {c}}}{\partial \mathbf {C}} \frac {\partial \mathbf {C}}{\partial \mathbf {P}} = \frac {\partial \mathcal {L} _ {\mathbf {c}}}{\partial \mathbf {C}} \left(\frac {\partial \mathbf {C}}{\partial \mathbf {c} _ {i}} \frac {\partial \mathbf {c} _ {i}}{\partial \mathbf {P}} + \frac {\partial \mathbf {C}}{\partial \alpha_ {i}} \frac {\partial \alpha_ {i}}{\partial \mathbf {P}}\right) \\ = \frac {\partial \mathcal {L} _ {\mathrm {c}}}{\partial \overline {{\mathbf {C}}}} \frac {\partial \mathbf {C}}{\partial \alpha_ {i}} \left(\frac {\partial \alpha_ {i}}{\partial \boldsymbol {\Sigma} ^ {\prime}} \frac {\partial \boldsymbol {\Sigma} ^ {\prime}}{\partial \mathbf {P}} + \frac {\partial \alpha_ {i}}{\partial \mathbf {m} _ {\mathrm {i}}} \frac {\partial \mathbf {m} _ {\mathrm {i}}}{\partial \mathbf {P}}\right) \tag {11} \\ = \frac {\partial \mathcal {L} _ {\mathbf {c}}}{\partial \mathbf {C}} \frac {\partial \mathbf {C}}{\partial \alpha_ {i}} \left(\frac {\partial \alpha_ {i}}{\partial \boldsymbol {\Sigma} ^ {\prime}} \frac {\partial (\mathbf {J P} ^ {- 1} \boldsymbol {\Sigma} \mathbf {P} ^ {- T} \mathbf {J} ^ {T})}{\partial \mathbf {P}} + \frac {\partial \alpha_ {i}}{\partial \mathbf {m} _ {\mathbf {i}}} \frac {\partial (\mathbf {K P X} _ {i})}{\partial \mathbf {P} d _ {i}}\right) \\ \end{array} +$$ + +where $d_{i}$ denotes the $z$ -axis coordinate of projection $\mathbf{m}_i$ . The item $\frac{\partial\mathbf{C}}{\partial\mathbf{c}_i}\frac{\partial\mathbf{c}_i}{\partial\mathbf{P}}$ can be eliminated because we are only concerned about the view-independent color in our tracking implementation. In addition, we find that the intermediate gradient $\frac{\partial(\mathbf{KP}\mathbf{X}_i)}{\partial\mathbf{P}d_i}$ is the deterministic component for the camera pose $\mathbf{P}$ . So we simply ignore the backpropagation of $\frac{\partial(\mathbf{JP}^{-1}\pmb{\Sigma}\mathbf{P}^{-T}\mathbf{J}^{T})}{\partial\mathbf{P}}$ for efficiency. More details can be found in the supplemental materials. + +Coarse-to-fine camera tracking. It would be problematic to optimize the camera pose with all image pixels since artifacts in images can cause drifted camera tracking. To address this issue, as shown in Fig. 2, in the differentiable pose estimation step for each frame, we first take advantage of image regularity to render only a sparse set of pixels and optimize tracking loss to obtain a coarse camera pose. This coarse optimization step significantly eases the influence of detailed artifacts. Further, we use this coarse camera pose and depth observation to select reliable 3D Gaussians, which guides GS-SLAM to render informative areas with clear geometric structures to refine coarse camera pose via further optimizing tracking loss on new rendering pixels. + +Specifically, in the coarse stage, we first render a coarse image $\hat{\mathbf{I}}_c$ with resolution $H / 2\times W / 2$ at uniformly sampled image coordinates and optimize tracking loss in Eq. 10 for $T_{c}$ iterations, and the obtained camera pose is denoted as $\mathbf{P}_c$ . In the fine stage, we use a similar technique with adaptive 3D Gaussian expansion strategy in Section 3.2 to select reliable 3D Gaussian to render full-resolution images while ignoring noisy 3D Gaussians that cause artifacts. In detail, we check all visible 3D Gaussians under coarse camera pose $\mathbf{P}_c$ , and remove 3D Gaussians whose position is far away from the scene surface. Formally, for each visible 3D Gaussians $G_{i}$ with position $\mathbf{X}_i$ , we project it to the camera plane using coarse camera pose $\mathbf{P}_c$ and camera intrinsic. Given the projected pixel's depth observation $D_{i}$ and the distance $d_{i}$ that is between 3D Gaussians $G_{i}$ and the camera image plane, the reliable 3D Gaussians are selected as follows: + +$$ +\begin{array}{l} \mathbf {G} _ {\text {s e l e c t e d}} = \left\{G _ {i} \mid G _ {i} \in \mathbf {G} \text {a n d} a b s \left(D _ {i} - d _ {i}\right) \leq \varepsilon \right\}, \tag {12} \\ \tilde {\mathbf {I}} _ {f} = \mathcal {F} (u, v, \mathbf {G} _ {s e l e c t e d}), \\ \end{array} +$$ + +where we use the selected reliable 3D Gaussians to render full-resolution images $\hat{\mathbf{I}}_f$ . $u, v$ denote the pixel coordinates in $\hat{\mathbf{I}}_f$ , and $\mathcal{F}$ represents the color splattering rendering function. The final camera pose $\mathbf{P}$ is obtained by optimizing + +tracking loss in Eq. 10 with $\hat{\mathbf{I}}_f$ for other $T_{f}$ iterations. Note that $\hat{\mathbf{I}}_c$ and $\hat{\mathbf{I}}_f$ are only rendered at previously observed areas, avoiding rendering areas where 3D scene representations have not been optimized in the mapping process. Also, we add keyframes based on the proportion of the currently observed image's reliable region to the overall image. At the same time, when the current tracking frame and most recent keyframe differ by more than a threshold value $\mu_{k}$ , this frame will be inserted as a keyframe. + +**Bundle adjustment.** In the bundle adjustment (BA) phase, we optimize the camera poses $\mathbf{P}$ and the 3D Gaussian scene representation $\mathbf{S}$ jointly. We randomly select $K$ keyframes from the keyframe database for optimization, using the loss function similar to the mapping part. For pose optimization stability, we only optimize the scene representation $\mathbf{S}$ in the first half of the iterations. In the other half of the iterations, we simultaneously optimize the map and the poses. Then, the accurate camera pose $\mathbf{P}$ is optimized by minimizing rendering color loss, i.e., + +$$ +\mathcal {L} _ {b a} = \frac {1}{K} \sum_ {k = 1} ^ {K} \sum_ {m = 1} ^ {H W} \left| D _ {m} - \hat {D} _ {m} \right| _ {1} + \lambda_ {m} \left| \mathbf {C} _ {m} - \hat {\mathbf {C}} _ {m} \right| _ {1}, \min _ {\mathbf {R}, \mathbf {t}, \mathbf {S}} (\mathcal {L} _ {b a}). \tag {13} +$$ + +# 4. Experiment + +# 4.1. Experimental Setup + +Dataset. To evaluate the performance of GS-SLAM, we conduct experiments on the Replica [31], and TUM-RGBD [33]. Following [11, 27, 41, 48, 55], we use 8 scenes from the Replica dataset for localization, mesh reconstruction, and rendering quality comparison. The selected three subsets of TUM-RGBD datasets are used for localization. + +Baselines. We compare our method with existing SOTA NeRF-based dense visual SLAM: NICE-SLAM [55], VoxFusion [48], CoSLAM [41], ESLAM [11] and Point-SLAM [27]. The rendering performance of CoSLAM [41] and ESLAM [11] is conducted from the open source code with the same configuration in [27]. + +Metric. For mesh reconstruction, we use the 2D Depth L1 (cm) [55], the Precision $(\mathbb{P},\%)$ , Recall $(\mathbb{R},\%)$ , and F-score with a threshold of $1\mathrm{cm}$ to measure the scene geometry. For localization, we use the absolute trajectory (ATE, cm) error [33] to measure the accuracy of the estimated camera poses. We further evaluate the rendering performance using the peak signal-to-noise ratio (PSNR), SSIM [43], and LPIPS [52] by following [27]. To be fair, we run all the methods on a dataset 10 times and report the average results. More details can be found in the supplemental materials. + +Implementation details. GS-SLAM is implemented in Python using the PyTorch framework, incorporating CUDA code for Gaussian splatting and trained on a desktop PC with a 5.50GHz Intel Core i9-13900K CPU and NVIDIA RTX 4090 GPU. We extended the existing code for differentiable Gaussian splatting rasterization with additional func + +tionality for handling depth, pose, and cumulative opacity during both forward and backward propagation. More details can be found in the supplemental materials. + +# 4.2. Evaluation of Localization and Mapping + +Evaluation on Replica. Tracking ATE: Tab. 1 illustrates the tracking performance of our method and the state-of-the-art methods on the Replica dataset. Our method achieves the best or second performance in 7 of 8 scenes and outperforms the second-best method Point-SLAM [27] by $0.4\mathrm{cm}$ on average at 8.34 FPS. It is noticeable that the second best method, Point-SLAM [27] runs at 0.42 FPS, which is $20\times$ slower than our method, indicating that GS-SLAM achieves a better trade-off between the tracking accuracy and the runtime efficiency. Mapping ACC: Tab. 3 report the mapping evaluation results of our method with other current state-of-the-art visual SLAM methods. GS-SLAM achieves the best performance in Depth L1 (1.16cm) and Precision $(74.0\%)$ metrics on average. For Recall and F1 scores, GS-SLAM performs comparably to the second best method CoSLAM [41]. The visualization results in Fig. 4 show that GS-SLAM achieves satisfying construction mesh with clear boundaries and details. + +Evaluation on TUM-RGBD. Tab. 2 compares GS-SLAM with the other SLAM systems in TUM-RGBD dataset. Our method surpasses iMAP [35], NICE-SLAM [55] and Voxfusion [48], and achieves a comparable performance, average $3.7\mathrm{cm}$ ATE RSME, with the SOTA methods. A gap to traditional methods still exists between the neural vSLAM and the traditional SLAM systems, which employ more sophisticated tracking schemes [27]. + +Table 1. Tracking comparison (ATE RMSE [cm]) of the proposed method vs. the SOTA methods on the Replica dataset. The running speed of methods in the upper part is lower than 5 FPS, * denotes the reproduced results by running officially released code. + +
MethodRm0Rm1Rm2Off0Off1Off2Off3Off4avg
Point-SLAM [27]0.560.470.300.350.620.550.720.730.54
NICE-SLAM [55]0.971.311.070.881.001.061.101.131.06
Vox-Fusion* [48]1.374.701.478.482.042.581.112.943.09
ESLAM [11]0.710.700.520.570.550.580.720.630.63
CoSLAM [41]0.700.951.350.590.552.031.560.721.00
Ours0.480.530.330.520.410.590.460.70.50
+ +Table 2. Tracking ATE [cm] on TUM-RGBD [33]. Our method achieves a comparable performance among the neural vSLAMs. * denotes the reproduced results by running officially released code. + +
Methodfr1Deskfr2_xyzfr3_offAvg.Methodfr1Deskfr2_xyzfr3_offAvg.
DI-Fusion [9]4.42.05.84.1NICE-SLAM [55]4.331.73.913.3
Elastic-Fusion [46]2.51.22.52.1Vox-Fusion* [48]3.51.526.010.3
BAD-SLAM [30]1.71.11.71.5CoSLAM [41]2.71.92.62.4
Kintinous [45]3.72.93.03.2ESLAM [11]2.31.12.42.0
ORB-SLAM2 [20]1.60.41.01.0Point-SLAM2.61.33.22.4
iMAP* [35]7.22.19.06.1Ours3.31.36.63.7
+ +# 4.3. Rendering Evaluation + +We compare the rendering performance of the proposed GS-SLAM with the neural visual SLAM methods in Tab. 6. The results show that GS-SLAM achieves the best performance + +Table 3. Reconstruction comparison of the proposed method vs. the SOTA methods on Replica dataset. + +
MethodMetricRm 0Rm 1Rm 2Off 0Off 1Off 2Off 3Off 4Avg.
NICESL AM [55]Depth L1 ↓1.811.442.041.391.768.334.992.012.97
Precision ↑45.8643.7644.3851.4050.8038.3740.8537.3544.10
Recall↑44.1046.1242.7848.6653.0839.9839.0435.7743.69
F1↑44.9644.8443.5649.9951.9139.1639.9236.5443.86
VoxFusion [48]Depth L1 ↓1.091.902.212.323.404.192.961.612.46
Precision↑75.8335.8863.1048.5143.5054.4869.1155.4055.73
Recall↑64.8933.0756.6244.7638.4447.8560.6146.7949.13
F1↑69.9334.3859.6746.5440.8150.9564.5650.7252.20
CoSLAM [41]Depth L1 ↓0.990.822.281.241.617.704.651.432.59
Precision↑81.7177.9573.3079.4180.6755.6457.6379.7673.26
Recall↑74.0370.7965.7371.4670.3552.9656.0671.2266.58
F1↑77.6874.2069.3175.2375.1654.2756.8375.2569.74
ESL AM [11]Depth L1 ↓0.630.620.980.571.667.323.940.882.08
Precision↑74.3375.9482.4872.2065.7470.7372.4872.2473.27
Recall↑87.3787.0184.9988.3684.3881.9279.1880.6384.23
F1↑80.3281.1083.7279.4773.9075.9275.6876.2178.29
OursDepth L1 ↓1.310.821.260.810.961.411.531.081.16
Precision↑64.5883.1170.1383.4387.7770.9163.1868.8874.00
Recall↑61.2976.8363.8476.9076.1561.6362.9161.5067.63
F1↑62.8979.8566.8480.0381.5565.9559.1764.9870.15
+ +in all the metrics. Our method significantly outperforms the second-best methods CoSLAM [41], ESLAM [11] and NICE-SLAM [55] by 1.52 dB in PSNR, 0.027 in SSIM and 0.12 in LPIPS, respectively. It is noticeable that GS-SLAM achieves 386 FPS rendering speed on average, which is $100 \times$ faster than the second-best method Vox-Fusion [48]. This excellent rendering performance is attributed to the efficient 3D Gaussian rendering pipeline and can be further applied to real-time downstream tasks, such as VR [5], robot navigation [7] and autonomous driving [2]. The visualization results in Fig. 5 show that GS-SLAM can generate much more high-quality and realistic images than the other methods, especially in edge areas with detailed structures. While NICE-SLAM [55] causes severe artifacts and blurs, CoSLAM [41] and ESLAM [11] generate blur around the image boundaries. + +![](images/fec0ca7d56512019e46711d86a69001801b6665ef0be59affaa59b06322d9b83.jpg) +Figure 4. Reconstruction performance comparison of the proposed GS-SLAM and SOTA methods on the Replica dataset. + +![](images/fdcd81bd916aabd7dbf1de73f9b520bda2fa382663cb7aaaf6a756593cc4751d.jpg) +Figure 5. The render visualization results on the Replica dataset of the proposed GS-SLAM and SOTA methods. GS-SLAM can generate much more high-quality and realistic images than the other methods, especially around the object boundaries. + +# 4.4. Runtime Analysis + +Tab. 4 and Tab. 5 illustrate the runtime and memory usage of GS-SLAM and the state-of-the-art methods on the Replica and TUM-RGBD, respectively. We report the parameters of the neural networks and the memory usage of the scene representation. Note that Point-SLAM uses extra memory dynamic radius to improve performance (mark as $\dagger$ ). The results show that GS-SLAM achieves a competitive running speed with 8.34 FPS compared to the other Radiance Fields-based vSLAMs. Note that we do not use any neural network decoder in our system, which results in the zero learnable parameter. However, the 3D Gaussian scene representations of GS-SLAM consume 198.04 MB memory, $4\times$ larger than the second large method NICE-SLAM [55]. Memory usage is mainly caused by spherical harmonic coefficients in training, which is a common constraint among Gaussian splatting-based reconstruction methods. Despite this, we still achieve a $20\times$ faster FPS compared to the similar point-based method Point-SLAM [27]. Besides, we also provide a light version of GS-SLAM with zero-order spherical harmonic coefficients, significantly reducing memory usage while maintaining stable performance. + +# 4.5. Ablation Study + +We perform the ablation of GS-SLAM on the Replica dataset #Room0 subset to evaluate the effectiveness of coarse-to-fine tracking, and expansion mapping strategy. + +Effect of our expansion strategy for mapping. Tab. 7 shows the ablation of our proposed expansion strategy for mapping. The results illustrate that the expansion strategy can significantly improve the tracking and mapping per + +Table 4. Runtime and memory usage on Replica #Room0. The decoder parameters and embedding denote the parameter number of MLPs and the memory usage of the scene representation. + +
MethodTracking [ms×it] ↓Mapping [ms×it] ↓System FPS ↑Decoder param ↓Scene Embedding↓
Point-SLAM [27]0.06 × 4034.81 × 3000.420.127 M55.42 (+12453.2)†MB
NICE-SLAM [55]6.64 × 1028.63 × 602.910.06 M48.48 MB
Vox-Fusion [48]0.03 × 3066.53 × 101.280.054 M1.49 MB
CoSLAM [55]6.01 × 1013.18 × 1016.641.671 M—
ESLAM [11]6.85 × 819.87 × 1513.420.003 M27.12 MB
GS-SLAM11.9 × 1012.8 × 1008.340 M198.04 MB
+ +Table 5. Runtime and memory usage on TUM-RGBD dataset #fr1Desk and #fr2_xyz. + +
Method#fr1Desk#fr2.xyz
FPS ↑Memory ↓FPS ↑Memory ↓
Point-SLAM0.1018.3 (+160.7)†MB0.1214.2 (+7687.4)†MB
NICE-SLAM [55]0.11178.8MB0.12484.0MB
ESLAM [11]0.3127.2MB0.3151.6MB
GS-SLAM1.83(ATE:3.3)40.8MB1.51(ATE:1.3)48.4MB
GS-SLAM (light)1.92(ATE:4.3)18.8MB1.68(ATE:2.7)22.3MB
+ +formance. The implementation w/o adding means that we only initialize 3D Gaussians in the first frame and optimize the scene without adding new points. However, this strategy completely crashes because the density control in [13] can not handle real-time mapping tasks without an accurate point cloud input. Besides, the implementation w/o deletion suffers from a large number of redundant and noisy 3D Gaussian, which causes undesirable supervision. In contrast, the proposed expansion strategy effectively improves the tracking and mapping performance by 0.1 in ATE and 11.97 in Recall by adding more accurate constraints for the optimization. According to the visualization results in Fig. 6, our full implementation achieves more high-quality and detailed rendering and reconstruction results than the w/o delete strategy. + +Table 6. Rendering performance on Replica dataset. We outperform existing dense neural RGB-D methods on the commonly reported rendering metrics. Note that GS-SLAM achieves 386 FPS on average, benefiting from the efficient Gaussian scene representation. + +
MethodMetricRoom 0Room 1Room 2Office 0Office 1Office 2Office 3Office 4Avg.FPS.
NICE-SLAM [55]PSNR [dB] ↑22.1222.4724.5229.0730.3419.6622.2324.9424.42
SSIM ↑0.6890.7570.8140.8740.8860.7970.8010.8560.8090.30
LPIPS ↓0.3300.2710.2080.2290.1810.2350.2090.1980.233
Vox-Fusion* [48]PSNR [dB] ↑22.3922.3623.9227.7929.8320.3323.4725.2124.413.88
SSIM ↑0.6830.7510.7980.8570.8760.7940.8030.8470.801
LPIPS ↓0.3030.2690.2340.2410.1840.2430.2130.1990.236
CoSLAM [41]PSNR [dB] ↑27.2728.4529.0634.1434.8728.4328.7630.9130.243.68
SSIM ↑0.9100.9090.9320.9610.9690.9380.9410.9550.939
LPIPS ↓0.3240.2940.2660.2090.1960.2580.2290.2360.252
ESLAM [11]PSNR [dB] ↑25.3227.7729.0833.7130.2028.0928.7729.7129.082.82
SSIM ↑0.8750.9020.9320.9600.9230.9430.9480.9450.929
LPIPS ↓0.3130.2980.2480.1840.2280.2410.1960.2040.336
OursPSNR [dB] ↑31.5632.8632.5938.7041.1732.3632.0332.9234.27386.91
SSIM ↑0.9680.9730.9710.9860.9930.9780.9700.9680.975
LPIPS ↓0.0940.0750.0930.0500.0330.0940.1100.1120.082
+ +Table 7. Ablation of the adaptive 3D Gaussian expansion strategy on Replica #Room0. + +
Setting#Room0
ATE↓Depth L1↓Precision↑Recall ↑F1↑PSNR↑SSIM↑LPIPS↓
w/o addXXXXXXXX
w/o delete0.581.6853.5549.3251.3531.220.9670.094
w/ add & delete0.481.3164.5861.2962.8931.560.9680.094
+ +Table 8. Ablation of the coarse-to-fine tracking strategy on Replica #Room0. + +
Setting#Room0
ATE↓Depth L1↓Precision↑Recall ↑F1↑PSNR↑SSIM↑LPIPS↓
Coarse0.911.4859.6857.5456.5029.130.9540.120
Fine0.491.3962.6159.1861.2930.840.9640.096
Coarse-to-fine0.481.3164.5861.2962.8931.560.9680.094
+ +Effect of coarse-to-fine tracking. According to the results in Tab. 8, the proposed coarse-to-fine tracking strategy performs best in all tracking, mapping, and rendering metrics. Compared with fine tracking, the coarse-to-fine tracking strategy significantly improves the performance by 0.01 in tracking ATE, 2.11 in Recall, and 0.72 in PSNR. Although the fine strategy surpasses the coarse strategy in precision, it suffers from the artifacts and noise in the reconstructed scene, leading to a fluctuation optimization. The coarse-to-fine strategy effectively avoids noise reconstruction and improves accuracy and robustness. + +# 4.6. Efficiency-to-accuracy trade-off. + +3DGS-based SLAM trade-off focuses not only on the efficiency of tracking and mapping but also emphasizes high-quality ultra-real-time rendering. As shown in Fig. 7b, GS-SLAM achieves $\approx 400$ FPS ultra-fast speed and highest PSNR in map rendering. At the same time, our method remains a competitive system FPS and lowest tracking ATE in Fig. 7a. Moreover, GS-SLAM shows great potential in memory reduction in Tab. 5, and comparable in mesh reconstruction. Note that baselines directly use 3DGS in Fig. 7, resulting in inferior performances. + +# 5. Conclusion and Limitations + +We introduced GS-SLAM, a novel dense visual SLAM method leveraging 3D Gaussian Splatting for efficient map + +![](images/e6ec1207e2f1c7667e3b982ea4481f5291ef8ba4b49e1dce14416e5d3f772787.jpg) +Figure 6. Rendering and mesh visualization of the adaptive 3D Gaussian expansion ablation on Replica #Room0. + +![](images/089e70d5edcd2ffc6b016ee37ec8ac84ec8232cb7c3d3e4e29e2b47e2b3178e3.jpg) + +![](images/db25191b988ea444aa2308b4496a313ed24b0dcc14cad4d704ffcfb923089b5c.jpg) +(a) Tracking performance +Figure 7. Bi-criteria figure of tracking/render performance and system FPS on Replica #Office0. + +![](images/346e7ac0b3ae2b5afc7b9731206c46252952a57766f796d562e794810a2f9413.jpg) +(b) Render performance + +ping and accurate camera pose estimation, striking a better speed-accuracy balance. However, its reliance on high-quality depth data may limit performance in certain conditions. Additionally, the approach's high memory requirements for large scenes suggest future improvements could focus on optimizing memory use, potentially via techniques such as quantization and clustering. We believe GS-SLAM has the potential to extend to larger scale with some improvements and will explore this in future work. + +Acknowledgements. This work is supported by the Shanghai AI Laboratory, National Key R&D Program of China (2022ZD0160101), the National Natural Science Foundation of China (62376222), Young Elite Scientists Sponsorship Program by CAST (2023QNRC001) and the Early Career Scheme of the Research Grants Council (RGC) of the Hong Kong SAR under grant No. 26202321. + +# References + +[1] Michael Bloesch, Jan Czarnowski, Ronald Clark, Stefan Leutenegger, and Andrew J. Davison. Codeslam - learning a compact, optimisable representation for dense visual slam. CVPR, pages 2560-2568, 2018. 2 +[2] Guillaume Bresson, Zayed Alsayed, Li Yu, and Sébastien Glaser. Simultaneous localization and mapping: A survey of current trends in autonomous driving. TIV, 2:194-220, 2017. 6 +[3] Zilong Chen, Feng Wang, and Huaping Liu. Text-to-3d using gaussian splatting. ArXiv, abs/2309.16585, 2023. 2 +[4] Brian Curless and Marc Levoy. Volumetric method for building complex models from range images. In SIGGRAPH, ACM, 1996. 4 +[5] Parth Rajesh Desai, Pooja Nikhil Desai, Komal Deepak Ajmera, and Khushbu Mehta. A review paper on oculus rift-a virtual reality headset. ArXiv, abs/1408.1173, 2014. 6 +[6] Hugh F. Durrant-Whyte and Tim Bailey. Simultaneous localization and mapping: part i. RAM, 13:99–110, 2006. 1 +[7] Christian Hane, Christopher Zach, Jongwoo Lim, Ananth Ranganathan, and Marc Pollefeys. Stereo depth map fusion for robot navigation. IROS, pages 1618-1625, 2011. 6 +[8] Caner Hazirbas, Andreas Wiedemann, Robert Maier, Laura Leal-Taixe, and Daniel Cremers. Tum rgb-d scribble-based segmentation benchmark. https://github.com/tum-vision/rgbdscribble_benchmark, 2018.6 +[9] Jiahui Huang, Shi-Sheng Huang, Haoxuan Song, and ShiMin Hu. Di-fusion: Online implicit 3d reconstruction with deep priors. In CVPR, pages 8932-8941, 2021. 6 +[10] Xudong Jiang, Lifeng Zhu, Jia Liu, and Aiguo Song. A slam-based 6 dof controller with smooth auto-calibration for virtual reality. TVC, 39:3873 - 3886, 2022. 1 +[11] Mohammad Mahdi Johari, Camilla Carta, and Francois Fleuret. Eslam: Efficient dense slam system based on hybrid representation of signed distance fields. CVPR, 2023. 1, 2, 5, 6, 7, 8, 4 +[12] Olaf Kähler, Victor Adrian Prisacariu, Julien P. C. Valentin, and David William Murray. Hierarchical voxel block hashing for efficient integration of depth images. RAL, 1:192-197, 2016. 1 +[13] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. TOG, 42(4), 2023. 2, 3, 7, 4 +[14] Leonid Keselman and Martial Hebert. Approximate differentiable rendering with algebraic surfaces. In ECCV, 2022. 2 +[15] Leonid Keselman and Martial Hebert. Flexible techniques for differentiable rendering with 3d gaussians. arXiv preprint arXiv:2308.14737, 2023. 2 +[16] Georg S. W. Klein and David William Murray. Parallel tracking and mapping on a camera phone. ISMAR, pages 83-86, 2009. 2 +[17] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. ArXiv, abs/2308.09713, 2023. 2 + +[18] Robert Maier, Raphael Schaller, and Daniel Cremers. Efficient online surface correction for real-time large-scale 3d reconstruction. ArXiv, abs/1709.03763, 2017. 1 +[19] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1 +[20] Raul Mur-Artal and Juan D. Tardós. Orb-slam2: An opensource slam system for monocular, stereo, and rgb-d cameras. *TRO*, 33:1255–1262, 2016. 1, 6 +[21] Richard A. Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J. Davison, Pushmeet Kohli, Jamie Shotton, Steve Hodges, and Andrew William Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. ISMAR, pages 127-136, 2011. 1 +[22] Richard A. Newcombe, S. Lovegrove, and Andrew J. Davison. Dtam: Dense tracking and mapping in real-time. ICCV, pages 2320-2327, 2011. 2 +[23] Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Marc Stamminger. Real-time 3d reconstruction at scale using voxel hashing. TOG, 32:1-11, 2013. 1 +[24] Delin Qu, Chi Yan, Dong Wang, Jie Yin, Dan Xu, Bin Zhao, and Xuelong Li. Implicit event-rgbd neural slam. CVPR, 2024. 1 +[25] Gerhard Reitmayr, Tobias Langlotz, Daniel Wagner, Alessandro Mulloni, Gerhard Schall, Dieter Schmalstieg, and Qi Pan. Simultaneous localization and mapping for augmented reality. ISUVR, pages 5-8, 2010. 1 +[26] Fabio Ruetz, Emili Hernández, Mark Pfeiffer, Helen Oleynikova, Mark Cox, Thomas Lowe, and Paulo Vinicius Koerich Borges. Ovpc mesh: 3d free-space representation for local ground vehicle navigation. ICRA, pages 8648-8654, 2018. 1 +[27] Erik Sandström, Yue Li, Luc Van Gool, and Martin R. Oswald. Point-slam: Dense neural point cloud-based slam. In ICCV, 2023. 2, 5, 6, 7, 4 +[28] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 3 +[29] Thomas Schöps, Torsten Sattler, and Marc Pollefeys. Bad slam: Bundle adjusted direct rgb-d slam. CVPR, pages 134-144, 2019. 2 +[30] Thomas Schops, Torsten Sattler, and Marc Pollefeys. BAD SLAM: Bundle adjusted direct RGB-D SLAM. In CVPR, 2019. 6 +[31] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Yuheng Ren, Shobhit Verma, Anton Clarkson, Ming Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Briales, Tyler Gillingham, Elias Mueggler, Luis Pesqueira, Manolis Savva, Dhruv Batra, Hauke Malte Strasdat, Renzo De Nardi, Michael Goesele, S. Lovegrove, and Richard A. Newcombe. The replica dataset: A digital replica of indoor spaces. ArXiv, abs/1906.05797, 2019. 5, 3, 4 +[32] J. Stuckler and Sven Behnke. Multi-resolution surfel maps for efficient dense 3d modeling and tracking. JVCIR, 25: 137-147, 2014. 1 + +[33] Jürgen Sturm, Nikolas Engelhard, Felix Endres, Wolfram Burgard, and Daniel Cremers. A benchmark for the evaluation of RGB-D SLAM systems. In IROS. IEEE/RSJ, 2012. 5, 6 +[34] Edgar Sucar, Kentaro Wada, and Andrew J. Davison. Nodeslam: Neural object descriptors for multi-view shape reconstruction. 3DV, pages 949–958, 2020. 2 +[35] Edgar Sucar, Shikun Liu, Joseph Ortiz, and Andrew J. Davison. imap: Implicit mapping and positioning in real-time. ICCV, 2021. 1, 2, 6 +[36] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. ArXiv, abs/2309.16653, 2023. 2, 5 +[37] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. In NIPS, 2021. 2 +[38] Andreas Langeland Teigen, Yeonsoo Park, Annette Stahl, and Rudolf Mester. Rgb-d mapping and tracking in a plenoxel radiance field. ArXiv, abs/2307.03404, 2023. 2 +[39] Charalambos Theodorou, Vladan Velisavljevic, Vladimir Dyo, and Fredi Nonyelu. Visual slam algorithms and their application for ar, mapping, localization and wayfinding. Array, 15:100222, 2022. 1 +[40] Angtian Wang, Peng Wang, Jian Sun, Adam Kortylewski, and Alan Yuille. Voge: a differentiable volume renderer using gaussian ellipsoids for analysis-by-synthesis. arXiv preprint arXiv:2205.15401, 2022. 2 +[41] Hengyi Wang, Jingwen Wang, and Lourdes de Agapito. Coslam: Joint coordinate and sparse parametric encodings for neural real-time slam. CVPR, 2023. 1, 5, 6, 8, 2, 4 +[42] Kaixuan Wang, Fei Gao, and Shaojie Shen. Real-time scalable dense surfel mapping. ICRA, pages 6919-6925, 2019. 1 +[43] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. TIP, 13(4):600-612, 2004. 5 +[44] Thomas Whelan, Michael Kaess, Maurice F. Fallon, Hordur Johannsson, John J. Leonard, and John B. McDonald. Kintinous: Spatially extended kinectfusion. In AAAI, 2012. 2 +[45] Thomas Whelan, John McDonald, Michael Kaess, Maurice Fallon, Hordur Johannsson, and John J. Leonard. Kintinuous: Spatially extended kinectfusion. In RSS Workshop on RGB-D, 2012. 6 +[46] Thomas Whelan, Stefan Leutenegger, Renato Salas-Moreno, Ben Glocker, and Andrew Davison. Elasticfusion: Dense slam without a pose graph. In RSS, 2015. 1, 6 +[47] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. ArXiv, abs/2310.08528, 2023. 2 +[48] Xingrui Yang, Hai Li, Hongjia Zhai, Yuhang Ming, Yuqian Liu, and Guofeng Zhang. Vox-fusion: Dense tracking and mapping with voxel-based neural implicit representation. IS-MAR, pages 499-507, 2022. 2, 5, 6, 7, 8 +[49] Ziyi Yang, Xinyu Gao, Wenming Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians + +for high-fidelity monocular dynamic scene reconstruction. ArXiv, abs/2309.13101, 2023. 2 +[50] Zeyu Yang, Hongye Yang, Zijie Pan, Xiatian Zhu, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. ArXiv, abs/2310.10642, 2023. 2 +[51] Taoran Yi, Jiemin Fang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussiandreamer: Fast generation from text to 3d gaussian splattering with point cloud priors. ArXiv, abs/2310.08529, 2023. 2 +[52] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 5 +[53] Youmin Zhang, Fabio Tosi, Stefano Mattoccia, and Matteo Poggi. Go-slam: Global optimization for consistent 3d instant reconstruction. In ICCV, 2023. 2 +[54] Shuaifeng Zhi, Michael Bloesch, Stefan Leutenegger, and Andrew J. Davison. Scenocode: Monocular dense semantic reconstruction using learned encoded scene representations. CVPR, pages 11768-11777, 2019. 2 +[55] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R. Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. CVPR, 2021. 1, 2, 5, 6, 7, 8 +[56] Wojciech Zielonka, Timur M. Bagautdinov, Shunsuke Saito, Michael Zollhofer, Justus Thies, and Javier Romero. Drivable 3d gaussian avatars. 2023. 2 +[57] Matthias Zwicker, Hanspeter Pfister, Jeroen van Baar, and Markus H. Gross. Ewa volume splatting. Proceedings Visualization, 2001. VIS '01., pages 29-538, 2001. 3 \ No newline at end of file diff --git a/2024/GS-SLAM_ Dense Visual SLAM with 3D Gaussian Splatting/images.zip b/2024/GS-SLAM_ Dense Visual SLAM with 3D Gaussian Splatting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6b0a934aecaf737250d305ab48743d5f55981943 --- /dev/null +++ b/2024/GS-SLAM_ Dense Visual SLAM with 3D Gaussian Splatting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9590e02174754695d7380e5716ef65622cc7c1b756016b349450d487566dcc75 +size 850977 diff --git a/2024/GS-SLAM_ Dense Visual SLAM with 3D Gaussian Splatting/layout.json b/2024/GS-SLAM_ Dense Visual SLAM with 3D Gaussian Splatting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9f909c97b6f6e9440104340b640bd1840c1a4e77 --- /dev/null +++ b/2024/GS-SLAM_ Dense Visual SLAM with 3D Gaussian Splatting/layout.json @@ -0,0 +1,9539 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 110, + 103, + 483, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 103, + 483, + 121 + ], + "spans": [ + { + "bbox": [ + 110, + 103, + 483, + 121 + ], + "type": "text", + "content": "GS-SLAM: Dense Visual SLAM with 3D Gaussian Splatting" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 141, + 192, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 141, + 192, + 156 + ], + "spans": [ + { + "bbox": [ + 133, + 141, + 192, + 156 + ], + "type": "text", + "content": "Chi Yan" + }, + { + "bbox": [ + 133, + 141, + 192, + 156 + ], + "type": "inline_equation", + "content": "^{1,3*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 227, + 142, + 288, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 142, + 288, + 157 + ], + "spans": [ + { + "bbox": [ + 227, + 142, + 288, + 157 + ], + "type": "text", + "content": "Delin " + }, + { + "bbox": [ + 227, + 142, + 288, + 157 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}\\mathbf{u}^{1,2*}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 323, + 143, + 369, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 143, + 369, + 156 + ], + "spans": [ + { + "bbox": [ + 323, + 143, + 369, + 156 + ], + "type": "text", + "content": "Dan " + }, + { + "bbox": [ + 323, + 143, + 369, + 156 + ], + "type": "inline_equation", + "content": "\\mathrm{Xu}^3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 408, + 143, + 465, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 143, + 465, + 156 + ], + "spans": [ + { + "bbox": [ + 408, + 143, + 465, + 156 + ], + "type": "text", + "content": "Bin Zhao" + }, + { + "bbox": [ + 408, + 143, + 465, + 156 + ], + "type": "inline_equation", + "content": "^{1,4}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 154, + 157, + 232, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 157, + 232, + 171 + ], + "spans": [ + { + "bbox": [ + 154, + 157, + 232, + 171 + ], + "type": "text", + "content": "Zhigang Wang" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 272, + 157, + 301, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 157, + 301, + 171 + ], + "spans": [ + { + "bbox": [ + 272, + 157, + 301, + 171 + ], + "type": "text", + "content": "Dong" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 376, + 157, + 444, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 157, + 444, + 171 + ], + "spans": [ + { + "bbox": [ + 376, + 157, + 444, + 171 + ], + "type": "text", + "content": "Xuelong Li" + }, + { + "bbox": [ + 376, + 157, + 444, + 171 + ], + "type": "inline_equation", + "content": "^{1,5}" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 171, + 283, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 171, + 283, + 200 + ], + "spans": [ + { + "bbox": [ + 54, + 171, + 283, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 54, + 171, + 283, + 200 + ], + "type": "text", + "content": "Shanghai AI Laboratory " + }, + { + "bbox": [ + 54, + 171, + 283, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 54, + 171, + 283, + 200 + ], + "type": "text", + "content": "Fudan University " + }, + { + "bbox": [ + 54, + 171, + 283, + 200 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 54, + 171, + 283, + 200 + ], + "type": "text", + "content": "Northwestern Polytechnical University" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 299, + 171, + 547, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 171, + 547, + 200 + ], + "spans": [ + { + "bbox": [ + 299, + 171, + 547, + 200 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 299, + 171, + 547, + 200 + ], + "type": "text", + "content": "Hong Kong University of Science and Technology " + }, + { + "bbox": [ + 299, + 171, + 547, + 200 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 299, + 171, + 547, + 200 + ], + "type": "text", + "content": "TeleAI, China Telecom Corp Ltd" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 143, + 227, + 191, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 191, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 191, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 251, + 290, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 251, + 290, + 504 + ], + "spans": [ + { + "bbox": [ + 47, + 251, + 290, + 504 + ], + "type": "text", + "content": "In this paper, we introduce GS-SLAM that first utilizes 3D Gaussian representation in the Simultaneous Localization and Mapping (SLAM) system. It facilitates a better balance between efficiency and accuracy. Compared to recent SLAM methods employing neural implicit representations, our method utilizes a real-time differentiable splatting rendering pipeline that offers significant speedup to map optimization and RGB-D rendering. Specifically, we propose an adaptive expansion strategy that adds new or deletes noisy 3D Gaussians in order to efficiently reconstruct new observed scene geometry and improve the mapping of previously observed areas. This strategy is essential to extend 3D Gaussian representation to reconstruct the whole scene rather than synthesize a static object in existing methods. Moreover, in the pose tracking process, an effective coarse-to-fine technique is designed to select reliable 3D Gaussian representations to optimize camera pose, resulting in runtime reduction and robust estimation. Our method achieves competitive performance compared with existing state-of-the-art real-time methods on the Replica, TUM-RGBD datasets. Project page: https://gs-slam.github.io/." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 523, + 128, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 523, + 128, + 536 + ], + "spans": [ + { + "bbox": [ + 47, + 523, + 128, + 536 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 543, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 543, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 543, + 287, + 687 + ], + "type": "text", + "content": "Simultaneous localization and mapping (SLAM) has emerged as a pivotal technology in fields such as robotics [6], virtual reality [10], and augmented reality [25, 39]. The goal of SLAM is to construct a dense/sparse map of an unknown environment while simultaneously tracking the camera pose. Traditional SLAM methods employ point/surfel clouds [20, 32, 42, 46], mesh representations [26], voxel hashing [12, 18, 23] or voxel grids [21] as scene representations to construct dense mapping, and have made considerable progress on localization accuracy. However, these methods face serious challenges in obtaining fine-grained dense maps." + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 307, + 225, + 545, + 403 + ], + "blocks": [ + { + "bbox": [ + 307, + 225, + 545, + 403 + ], + "lines": [ + { + "bbox": [ + 307, + 225, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 225, + 545, + 403 + ], + "type": "image", + "image_path": "22fe11d95c4291fb95410f5b65e9fc36f5340c37a7d6b3c833bfdee4c95ddd87.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 413, + 547, + 491 + ], + "lines": [ + { + "bbox": [ + 304, + 413, + 547, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 547, + 491 + ], + "type": "text", + "content": "Figure 1. The illustration of the proposed GS-SLAM. It first utilizes the 3D Gaussian representation and differentiable splatting rasterization pipeline in SLAM, achieving real-time tracking and mapping performance on GPU. Besides, benefiting from the splatting rasterization pipeline, GS-SLAM achieves a " + }, + { + "bbox": [ + 304, + 413, + 547, + 491 + ], + "type": "inline_equation", + "content": "100\\times" + }, + { + "bbox": [ + 304, + 413, + 547, + 491 + ], + "type": "text", + "content": " faster rendering FPS and more high-quality full image results than the other SOTA methods." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "text", + "content": "Recently, Neural Radiance Fields (NeRF) [19] have been explored to enhance SLAM methodologies and exhibit strengths in generating high-quality, dense maps with low memory consumption [35]. In particular, iMAP [35] uses a single multi-layer perceptron (MLP) to represent the entire scene, which is updated globally with the loss between volume-rendered RGB-D image and ground-truth observations. NICE-SLAM [55] utilizes a hierarchical neural implicit grid as scene map representation to allow local updates for reconstructing large scenes. Moreover, ESLAM [11], CoSLAM [41] and EN-SLAM [24] utilize axis-aligned feature planes and joint coordinate-parametric encoding to improve the capability of scene representation, achieving efficient and high-quality surface map reconstruction. In practical mapping and tracking steps, these methods only render a small set of pixels to reduce optimization time, which leads to the reconstructed dense maps lacking" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 197, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 197, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 197, + 703 + ], + "type": "text", + "content": "*Equal contribution: yanchi@pjlab.org.cn." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 59, + 703, + 234, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 234, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 234, + 712 + ], + "type": "text", + "content": "†Corresponding author: dongwang.dw93@gmail.com." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19595" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "type": "text", + "content": "the richness and intricacy of details. In essence, it is a tradeoff for the efficiency and accuracy of NeRF-based SLAM since obtaining high-resolution images with the ray-based volume rendering technique is time-consuming and unacceptable." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 132, + 286, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 132, + 286, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 132, + 286, + 262 + ], + "type": "text", + "content": "Fortunately, recent work [13, 17, 47] with 3D Gaussian representation and tile-based splatting techniques has shown great superiority in the efficiency of high-resolution image rendering. It is applied to synthesize novel view RGB images of static objects, achieving state-of-the-art visual quality for " + }, + { + "bbox": [ + 46, + 132, + 286, + 262 + ], + "type": "inline_equation", + "content": "1080\\mathrm{p}" + }, + { + "bbox": [ + 46, + 132, + 286, + 262 + ], + "type": "text", + "content": " resolution at real-time speed. Inspired by this, we extend the rendering superiority of 3D Gaussian scene representation and real-time differentiable splatting rendering pipeline for the task of dense RGB-D SLAM and manage to jointly promote the speed and accuracy of NeRF-based dense SLAM, as shown in Fig. 1." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 263, + 286, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 263, + 286, + 490 + ], + "spans": [ + { + "bbox": [ + 46, + 263, + 286, + 490 + ], + "type": "text", + "content": "To this end, we propose GS-SLAM, the first RGB-D dense SLAM system that first utilizes 3D Gaussian scene representation coupled with the splatting rendering technique to better balance speed and accuracy. Our system optimizes camera tracking and mapping with a novel RGB-D rendering approach that processes 3D Gaussians quickly and accurately through sorting and " + }, + { + "bbox": [ + 46, + 263, + 286, + 490 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 263, + 286, + 490 + ], + "type": "text", + "content": "-blending. We enhance scene reconstruction by introducing an adaptive strategy for managing 3D Gaussian elements, which optimizes mapping by focusing on current observations and minimizes errors in dense maps and images. Moreover, we propose a coarse-to-fine approach, starting with low-resolution image analysis for initial pose estimation and refining it with high-resolution rendering using select 3D Gaussians, to boost speed and accuracy. We perform extensive evaluations on a selection of indoor RGB-D datasets and demonstrate state-of-the-art performance on dense neural RGB-D SLAM in terms of tracking, rendering, and mapping. Overall, our contributions include:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 492, + 286, + 658 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 48, + 492, + 286, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 492, + 286, + 563 + ], + "spans": [ + { + "bbox": [ + 48, + 492, + 286, + 563 + ], + "type": "text", + "content": "- We propose GS-SLAM, the first 3D Gaussian Splatting(3DGS)-based dense RGB-D SLAM approach, which takes advantage of the fast splatting rendering technique to boost the mapping optimizing and pose tracking, achieving real-time and photo-realistic reconstruction performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 564, + 286, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 564, + 286, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 564, + 286, + 610 + ], + "type": "text", + "content": "- We present an adaptive 3D Gaussian expansion strategy to efficiently reconstruct new observed scene geometry and develop a coarse-to-fine technique to select reliable 3D Gaussians to improve camera pose estimation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 612, + 286, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 612, + 286, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 612, + 286, + 658 + ], + "type": "text", + "content": "- Our approach achieves competitive performance on Replica and TUM-RGBD datasets in terms of tracking, and mapping and runs at 8.43 FPS, resulting in a better balance between efficiency and accuracy." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 668, + 133, + 680 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 668, + 133, + 680 + ], + "spans": [ + { + "bbox": [ + 47, + 668, + 133, + 680 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "content": "- Dense Visual SLAM. The existing real-time dense visual SLAM systems are typically based on discrete handcrafted" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 252 + ], + "type": "text", + "content": "features or deep-learning embeddings, and follow the mapping and tracking architecture in [16]. DTAM [22] first introduces a dense SLAM system that uses photometric consistency to track a handheld camera and represent the scene as a cost volume. KinectFusion [44] performs camera tracking by iterative-closest-point and updates the scene via TSDF-Fusion. BAD-SLAM [29] proposes to jointly optimize the keyframe poses and 3D scene geometry via a direct bundle adjustment (BA) technique. In contrast, recent works integrate deep learning with the traditional geometry framework for more accurate and robust camera tracking and mapping, such as DROID-SLAM [37], CodeSLAM [1], SceneCode [54], and NodeSLAM [34], have made significant advances in the field, achieving more accurate and robust camera tracking and mapping performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 255, + 545, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 545, + 506 + ], + "type": "text", + "content": "Neural Implicit Radiance Field based SLAM. For NeRF-based SLAM, existing methods can be divided into three main types: MLP-based methods, Hybrid representation methods, and Explicit methods. MLP-based method iMAP [35] offers scalable and memory-efficient map representations but faces challenges with catastrophic forgetting in larger scenes. Hybrid representation methods combine the advantages of implicit MLPs and structure features, significantly enhancing the scene scalability and precision. For example, NICE-SLAM [55] integrates MLPs with multi-resolution voxel grids, enabling large scene reconstruction, and Vox-Fusion [48] employs octree expansion for dynamic map scalability, while ESLAM [11] and Point-SLAM [27] utilize tri-planes and neural point clouds respectively to improve the mapping capability. As for the explicit method proposed in [38], it stores map features in voxel directly, without any MLPs, enabling faster optimization. Instead of representing maps with implicit features, GS-SLAM utilizes the 3D Gaussian representation, efficiently renders images using splatting-based rasterization, and optimizes parameters directly with backward propagation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "3D Gaussian Representation. Several recent approaches have used 3D Gaussians for shape reconstruction, such as Fuzzy Metaballs [14, 15], VoGE [40], 3DGS [13]. Notably, 3DGS [13] demonstrates great superiorities in high-quality real-time novel-view synthesis. This work represents the scene with 3D Gaussians and develops a NeRF-style fast rendering algorithm to support anisotropic splating, achieving SOTA visual quality and fast high-resolution rendering performance. Beyond the rendering superiorities, Gaussian splating holds an explicit geometry scene structure and appearance, benefiting from the exact modeling of scenes representation [50]. This promising technology has been rapidly applied in several fields, including 3D generation [3, 36, 51], dynamic scene modeling [17][47][49], and photorealistic drivable avatars [56]. However, currently, there is no research addressing camera pose estimation or real-time mapping using 3D Gaussian models due to the in" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19596" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 547, + 187 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 547, + 187 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 547, + 187 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 547, + 187 + ], + "type": "image", + "image_path": "53be036ebc19fd8fa446cd8da14749f0bad356f997a5cb2bd86001c01bec7a5d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 196, + 547, + 231 + ], + "lines": [ + { + "bbox": [ + 46, + 196, + 547, + 231 + ], + "spans": [ + { + "bbox": [ + 46, + 196, + 547, + 231 + ], + "type": "text", + "content": "Figure 2. Overview of the proposed method. We aim to use 3D Gaussians to represent the scene and use the rendered RGB-D image for inverse camera tracking. GS-SLAM proposes a novel Gaussian expansion strategy to make the 3D Gaussian feasible to reconstruct the whole scene and can achieve real-time tracking, mapping, and rendering performance on GPU." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 244, + 288, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 244, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 244, + 288, + 304 + ], + "type": "text", + "content": "herent limitations of the prime pipeline [13], i.e., prerequisites of initialized point clouds or camera pose inputs [28]. In contrast, we derive the analytical derivative equations for pose estimation in the Gaussian representation and implement efficient CUDA optimization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 317, + 129, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 317, + 129, + 331 + ], + "spans": [ + { + "bbox": [ + 47, + 317, + 129, + 331 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 338, + 290, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 338, + 290, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 290, + 540 + ], + "type": "text", + "content": "Fig. 2 shows the overview of the proposed GS-SLAM. We aim to estimate the camera poses " + }, + { + "bbox": [ + 46, + 338, + 290, + 540 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{P}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 46, + 338, + 290, + 540 + ], + "type": "text", + "content": " of every frame and simultaneously reconstruct a dense scene map by giving an input sequential RGB-D stream " + }, + { + "bbox": [ + 46, + 338, + 290, + 540 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{I}_i, \\mathbf{D}_i\\}_{i=1}^M" + }, + { + "bbox": [ + 46, + 338, + 290, + 540 + ], + "type": "text", + "content": " with known camera intrinsic " + }, + { + "bbox": [ + 46, + 338, + 290, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{K} \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 46, + 338, + 290, + 540 + ], + "type": "text", + "content": ". In Sec. 3.1, we first introduce 3D Gaussian as the scene representation " + }, + { + "bbox": [ + 46, + 338, + 290, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 46, + 338, + 290, + 540 + ], + "type": "text", + "content": " and the RGB-D render by differentiable splattering rasterization. With the estimated camera pose of the keyframe, in Sec. 3.2, an adaptive expansion strategy is proposed to add new or delete noisy 3D Gaussians to efficiently reconstruct new observed scene geometry while improving the mapping of the previously observed areas. For camera tracking of every input frame, we derive an analytical formula for backward optimization with rendering RGB-D loss. We further introduce an effective coarse-to-fine technique to minimize rendering losses to achieve efficient and accurate pose estimation in Sec. 3.3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 552, + 234, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 234, + 565 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 234, + 565 + ], + "type": "text", + "content": "3.1. 3D Gaussian Scene Representation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 571, + 287, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 571, + 287, + 631 + ], + "spans": [ + { + "bbox": [ + 46, + 571, + 287, + 631 + ], + "type": "text", + "content": "Our goal is to optimize a scene representation that captures the geometry and appearance of the scene, resulting in a detailed dense map and high-quality novel view synthesis. To do this, we model the scene as a set of 3D Gaussians coupled with opacity and spherical harmonics" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 78, + 641, + 287, + 654 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 641, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 78, + 641, + 287, + 654 + ], + "type": "interline_equation", + "content": "\\mathbf {G} = \\left\\{G _ {i}: \\left(\\mathbf {X} _ {i}, \\boldsymbol {\\Sigma} _ {i}, \\Lambda_ {i}, \\mathbf {Y} _ {i}\\right) | i = 1, \\dots , N \\right\\}. \\tag {1}", + "image_path": "546bf9eac42740023c644910f94a7be6fbe68c7cdc153ce7ce996bf8d672d06d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 665, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 288, + 713 + ], + "type": "text", + "content": "Each 3D Gaussian scene representation " + }, + { + "bbox": [ + 47, + 665, + 288, + 713 + ], + "type": "inline_equation", + "content": "G_{i}" + }, + { + "bbox": [ + 47, + 665, + 288, + 713 + ], + "type": "text", + "content": " is defined by position " + }, + { + "bbox": [ + 47, + 665, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_i\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 47, + 665, + 288, + 713 + ], + "type": "text", + "content": ", 3D covariance matrix " + }, + { + "bbox": [ + 47, + 665, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}_{i}\\in \\mathbb{R}^{3\\times 3}" + }, + { + "bbox": [ + 47, + 665, + 288, + 713 + ], + "type": "text", + "content": ", opacity " + }, + { + "bbox": [ + 47, + 665, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\Lambda_{i}\\in \\mathbb{R}" + }, + { + "bbox": [ + 47, + 665, + 288, + 713 + ], + "type": "text", + "content": " and 1-degree spherical harmonics " + }, + { + "bbox": [ + 47, + 665, + 288, + 713 + ], + "type": "inline_equation", + "content": "(\\mathbf{Y})" + }, + { + "bbox": [ + 47, + 665, + 288, + 713 + ], + "type": "text", + "content": " per color channel, a total of 12 coefficients for " + }, + { + "bbox": [ + 47, + 665, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_i\\in \\mathbb{R}^{12}" + }, + { + "bbox": [ + 47, + 665, + 288, + 713 + ], + "type": "text", + "content": ". To reduce" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 244, + 545, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 244, + 545, + 268 + ], + "spans": [ + { + "bbox": [ + 305, + 244, + 545, + 268 + ], + "type": "text", + "content": "the learning difficulty of the 3D Gaussians [57], we parameterize the 3D Gaussian's covariance as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 391, + 275, + 545, + 291 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 275, + 545, + 291 + ], + "spans": [ + { + "bbox": [ + 391, + 275, + 545, + 291 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Sigma} = \\mathbf {R} \\mathbf {S} \\mathbf {S} ^ {T} \\mathbf {R} ^ {T}, \\tag {2}", + "image_path": "236aba05fd690c587f829e2ca36bbdd581e59be53cedb3a89e0eee810a1b2567.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 297, + 545, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 297, + 545, + 322 + ], + "spans": [ + { + "bbox": [ + 305, + 297, + 545, + 322 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 297, + 545, + 322 + ], + "type": "inline_equation", + "content": "\\mathbf{S} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 305, + 297, + 545, + 322 + ], + "type": "text", + "content": " is a 3D scale vector, " + }, + { + "bbox": [ + 305, + 297, + 545, + 322 + ], + "type": "inline_equation", + "content": "\\mathbf{R} \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 305, + 297, + 545, + 322 + ], + "type": "text", + "content": " is the rotation matrix, storing as a 4D quaternion." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 323, + 545, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 323, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 305, + 323, + 545, + 371 + ], + "type": "text", + "content": "Color and depth splatting rendering. With the optimized 3D Gaussian scene representation parameters, given the camera pose " + }, + { + "bbox": [ + 305, + 323, + 545, + 371 + ], + "type": "inline_equation", + "content": "\\mathbf{P} = \\{\\mathbf{R},\\mathbf{t}\\}" + }, + { + "bbox": [ + 305, + 323, + 545, + 371 + ], + "type": "text", + "content": ", the 3D Gaussians " + }, + { + "bbox": [ + 305, + 323, + 545, + 371 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 305, + 323, + 545, + 371 + ], + "type": "text", + "content": " are projected into 2D image plane for rendering with:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 377, + 377, + 545, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 377, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 377, + 377, + 545, + 392 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Sigma} ^ {\\prime} = \\mathbf {J} \\mathbf {P} ^ {- 1} \\boldsymbol {\\Sigma} \\mathbf {P} ^ {- T} \\mathbf {J} ^ {T}, \\tag {3}", + "image_path": "af9922820372689d796b535abdfcf80f27d9076ead59b1136a325fb97b43c5c5.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 400, + 547, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 400, + 547, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 400, + 547, + 460 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 400, + 547, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{J}" + }, + { + "bbox": [ + 304, + 400, + 547, + 460 + ], + "type": "text", + "content": " is the Jacobian of the affine approximation of the projective function. After projecting 3D Gaussians to the image plane, the color of one pixel is rendered by sorting the Gaussians in depth order and performing front-to-back " + }, + { + "bbox": [ + 304, + 400, + 547, + 460 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 400, + 547, + 460 + ], + "type": "text", + "content": "-blending rendering as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 366, + 468, + 545, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 468, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 366, + 468, + 545, + 502 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {C}} = \\sum_ {i \\in N} \\mathbf {c} _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {4}", + "image_path": "6aa1ddc43a67b22ebca0567c887712834594e6d6ecda13891d799213704b4fc6.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 510, + 547, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 547, + 570 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 547, + 570 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 510, + 547, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i" + }, + { + "bbox": [ + 304, + 510, + 547, + 570 + ], + "type": "text", + "content": " represents the color of the " + }, + { + "bbox": [ + 304, + 510, + 547, + 570 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 510, + 547, + 570 + ], + "type": "text", + "content": "-th 3D Gaussian obtained by learned spherical harmonics coefficients " + }, + { + "bbox": [ + 304, + 510, + 547, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 304, + 510, + 547, + 570 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 510, + 547, + 570 + ], + "type": "inline_equation", + "content": "\\alpha_i" + }, + { + "bbox": [ + 304, + 510, + 547, + 570 + ], + "type": "text", + "content": " is the density computed by learned opacity " + }, + { + "bbox": [ + 304, + 510, + 547, + 570 + ], + "type": "inline_equation", + "content": "\\Lambda_i" + }, + { + "bbox": [ + 304, + 510, + 547, + 570 + ], + "type": "text", + "content": " and 2D Gaussian with covariance " + }, + { + "bbox": [ + 304, + 510, + 547, + 570 + ], + "type": "inline_equation", + "content": "\\Sigma'" + }, + { + "bbox": [ + 304, + 510, + 547, + 570 + ], + "type": "text", + "content": ". Similarly, the depth is rendered by" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 366, + 578, + 545, + 612 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 578, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 366, + 578, + 545, + 612 + ], + "type": "interline_equation", + "content": "\\hat {D} = \\sum_ {i \\in N} d _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {5}", + "image_path": "bbbf29b0ca69358da24236b918cd8033dc69409ed8bf6a42c361949459e42f6c.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 616, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 616, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 305, + 616, + 545, + 651 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 616, + 545, + 651 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 305, + 616, + 545, + 651 + ], + "type": "text", + "content": " denotes the depth of the center of the " + }, + { + "bbox": [ + 305, + 616, + 545, + 651 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 616, + 545, + 651 + ], + "type": "text", + "content": "-th 3D Gaussian, which is obtained by projecting to " + }, + { + "bbox": [ + 305, + 616, + 545, + 651 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 305, + 616, + 545, + 651 + ], + "type": "text", + "content": "-axis in the camera coordinate." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 659, + 533, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 659, + 533, + 673 + ], + "spans": [ + { + "bbox": [ + 305, + 659, + 533, + 673 + ], + "type": "text", + "content": "3.2. Adaptive 3D Gaussian Expanding Mapping" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "The 3D Gaussian scene representations are updated and optimized on each selected keyframe for stable mapping. Given the estimated pose of each selected keyframe, we first" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "19597" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": "apply the proposed adaptive expansion strategy to add new or delete noisy 3D Gaussians from the whole scene representations to render RGB-D images with resolution " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": ", and then the updated 3D Gaussian scene representations are optimized by minimizing the geometric depth loss " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_d" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " and the photometric color loss " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{c}}" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " to the sensor observation depth " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " and color " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 58, + 165, + 287, + 199 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 165, + 287, + 199 + ], + "spans": [ + { + "bbox": [ + 58, + 165, + 287, + 199 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathbf {c}} = \\sum_ {m = 1} ^ {H W} \\left| \\mathbf {C} _ {m} - \\hat {\\mathbf {C}} _ {m} \\right|, \\quad \\mathcal {L} _ {d} = \\sum_ {m = 1} ^ {H W} \\left| D _ {m} - \\hat {D} _ {m} \\right|. \\tag {6}", + "image_path": "a2b3d5dc21e390fa1b6ef32d8a5beaff37e5e90741846978519e81a12af55f82.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 207, + 287, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 207, + 287, + 231 + ], + "spans": [ + { + "bbox": [ + 46, + 207, + 287, + 231 + ], + "type": "text", + "content": "The loss optimizes the parameters of all 3D Gaussians that contribute to the rendering of these keyframe images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 232, + 288, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 232, + 288, + 340 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 288, + 340 + ], + "type": "text", + "content": "Adaptive 3D Gaussian Expansion Strategy. At the first frame of the RGB-D sequence, we first uniformly sample half pixels from a whole image with " + }, + { + "bbox": [ + 46, + 232, + 288, + 340 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 46, + 232, + 288, + 340 + ], + "type": "text", + "content": " resolution and back-projecting them into 3D points " + }, + { + "bbox": [ + 46, + 232, + 288, + 340 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 46, + 232, + 288, + 340 + ], + "type": "text", + "content": " with corresponding depth observation " + }, + { + "bbox": [ + 46, + 232, + 288, + 340 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 232, + 288, + 340 + ], + "type": "text", + "content": ". The 3D Gaussian scene representations are created by setting position as " + }, + { + "bbox": [ + 46, + 232, + 288, + 340 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 46, + 232, + 288, + 340 + ], + "type": "text", + "content": " and initializing zero degree Spherical Harmonics coefficients with RGB color " + }, + { + "bbox": [ + 46, + 232, + 288, + 340 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_i" + }, + { + "bbox": [ + 46, + 232, + 288, + 340 + ], + "type": "text", + "content": ". The opacities are set to pre-defined values, and the covariance is set depending on the spatial point density, i.e.," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 350, + 287, + 364 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 350, + 287, + 364 + ], + "spans": [ + { + "bbox": [ + 77, + 350, + 287, + 364 + ], + "type": "interline_equation", + "content": "\\left\\{G _ {i} = \\left(\\mathbf {P} _ {i}, \\boldsymbol {\\Sigma} _ {\\text {i n i t}}, \\Lambda_ {\\text {i n i t}}, \\mathbf {C} _ {i}\\right) | i = 1, \\dots , M \\right\\}, \\tag {7}", + "image_path": "3232b230b533182bcad76835086f54ee2d518e7ae7df54ee52fd311dc78320cc.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 373, + 287, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 287, + 456 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 373, + 287, + 456 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 373, + 287, + 456 + ], + "type": "text", + "content": " equals to " + }, + { + "bbox": [ + 46, + 373, + 287, + 456 + ], + "type": "inline_equation", + "content": "HW / 2" + }, + { + "bbox": [ + 46, + 373, + 287, + 456 + ], + "type": "text", + "content": ". The 3D Gaussians are initialized and then optimized using the first RGB-D image with rendering loss. Note that only half of the pixels are used to initialize the scene, leaving space to conduct adaptive density control of Gaussians that splits large points into smaller ones and clones them with different directions to capture missing geometric details." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 457, + 287, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 457, + 287, + 565 + ], + "spans": [ + { + "bbox": [ + 46, + 457, + 287, + 565 + ], + "type": "text", + "content": "Adding step: To obtain a complete map of the environment, the 3D Gaussian scene representations should be able to model the geometry and appearance of newly observed areas. Specifically, at every keyframe, we add first rendered RGB-D images using historical 3D Gaussians and calculate cumulative opacity " + }, + { + "bbox": [ + 46, + 457, + 287, + 565 + ], + "type": "inline_equation", + "content": "T = \\sum_{i\\in N}\\alpha_i\\prod_{j = 1}^{i - 1}(1 - \\alpha_j)" + }, + { + "bbox": [ + 46, + 457, + 287, + 565 + ], + "type": "text", + "content": " for each pixel. We label one pixel as un-reliable " + }, + { + "bbox": [ + 46, + 457, + 287, + 565 + ], + "type": "inline_equation", + "content": "x^{un}" + }, + { + "bbox": [ + 46, + 457, + 287, + 565 + ], + "type": "text", + "content": " if its cumulative opacity " + }, + { + "bbox": [ + 46, + 457, + 287, + 565 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 457, + 287, + 565 + ], + "type": "text", + "content": " is too low or its rendered depth " + }, + { + "bbox": [ + 46, + 457, + 287, + 565 + ], + "type": "inline_equation", + "content": "\\hat{D}" + }, + { + "bbox": [ + 46, + 457, + 287, + 565 + ], + "type": "text", + "content": " is far away from observed depth " + }, + { + "bbox": [ + 46, + 457, + 287, + 565 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 457, + 287, + 565 + ], + "type": "text", + "content": ", i.e.," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 574, + 287, + 588 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 574, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 104, + 574, + 287, + 588 + ], + "type": "interline_equation", + "content": "T < \\tau_ {T} \\quad \\text {o r} \\quad | D - \\hat {D} | > \\tau_ {D}. \\tag {8}", + "image_path": "d4bd98ba79da484507a72224db6fa45328738edb83024363c879dcaf63b74283.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 594, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 652 + ], + "type": "text", + "content": "These selected un-reliable pixels mostly capture new observed areas. Then we back-project these un-reliable pixels to 3D points " + }, + { + "bbox": [ + 46, + 594, + 287, + 652 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^{un}" + }, + { + "bbox": [ + 46, + 594, + 287, + 652 + ], + "type": "text", + "content": ", and a set of new 3D Gaussians at " + }, + { + "bbox": [ + 46, + 594, + 287, + 652 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^{un}" + }, + { + "bbox": [ + 46, + 594, + 287, + 652 + ], + "type": "text", + "content": " initialized as Eq. 7 are added into scene representations to model the new observed areas." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 653, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 713 + ], + "type": "text", + "content": "**Deleting step:** As shown in Fig. 3, there are some floating 3D Gaussians due to the unstable adaptive control of Gaussians after optimization with Eq. 6. These floating 3D Gaussians will result in a low-quality dense map and a rendered image containing lots of artifacts. To address this" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 319, + 67, + 535, + 205 + ], + "blocks": [ + { + "bbox": [ + 319, + 67, + 535, + 205 + ], + "lines": [ + { + "bbox": [ + 319, + 67, + 535, + 205 + ], + "spans": [ + { + "bbox": [ + 319, + 67, + 535, + 205 + ], + "type": "image", + "image_path": "d4d5a6ae1d5f38d2d2e048d028616ebf9d872090c545a009861906a8433fbe83.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 206, + 545, + 240 + ], + "lines": [ + { + "bbox": [ + 305, + 206, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 305, + 206, + 545, + 240 + ], + "type": "text", + "content": "Figure 3. Illustration of the proposed adaptive 3D Gaussian expansion strategy. GS-SLAM inhibits the low-quality 3D Gaussian floaters in the current frustum according to depth." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 251, + 545, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 251, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 305, + 251, + 545, + 371 + ], + "type": "text", + "content": "issue, after adding new 3D Gaussians, we check all visible 3D Gaussians in the current camera frustum and significantly decrease opacity " + }, + { + "bbox": [ + 305, + 251, + 545, + 371 + ], + "type": "inline_equation", + "content": "\\Lambda_{i}" + }, + { + "bbox": [ + 305, + 251, + 545, + 371 + ], + "type": "text", + "content": " of 3D Gaussians whose position is not near the scene surfaces. Formally, for each visible 3D Gaussian, we draw a ray " + }, + { + "bbox": [ + 305, + 251, + 545, + 371 + ], + "type": "inline_equation", + "content": "\\mathbf{r}(t)" + }, + { + "bbox": [ + 305, + 251, + 545, + 371 + ], + "type": "text", + "content": " from camera origin " + }, + { + "bbox": [ + 305, + 251, + 545, + 371 + ], + "type": "inline_equation", + "content": "\\mathbf{o}" + }, + { + "bbox": [ + 305, + 251, + 545, + 371 + ], + "type": "text", + "content": " and its position " + }, + { + "bbox": [ + 305, + 251, + 545, + 371 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_i = (x_i,y_i,z_i)" + }, + { + "bbox": [ + 305, + 251, + 545, + 371 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 305, + 251, + 545, + 371 + ], + "type": "inline_equation", + "content": "\\mathbf{r}(t) = \\mathbf{o} + t(\\mathbf{X}_i - \\mathbf{o})" + }, + { + "bbox": [ + 305, + 251, + 545, + 371 + ], + "type": "text", + "content": ". Then, we can find a pixel with coordinate " + }, + { + "bbox": [ + 305, + 251, + 545, + 371 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 305, + 251, + 545, + 371 + ], + "type": "text", + "content": " where this ray intersects the image plane and corresponding depth observation " + }, + { + "bbox": [ + 305, + 251, + 545, + 371 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 305, + 251, + 545, + 371 + ], + "type": "text", + "content": ". The 3D Gaussians are deleted by degenerating its opacity as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 319, + 383, + 545, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 383, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 319, + 383, + 545, + 396 + ], + "type": "interline_equation", + "content": "G _ {i}: \\Lambda_ {i} \\Rightarrow G _ {i}: \\eta \\Lambda_ {i}, \\text {i f} D - d i s t (\\mathbf {X} _ {i}, \\mathbf {P} _ {u v}) > \\gamma , \\tag {9}", + "image_path": "eafc556fb35ef375c38e7bda7ef03b6cf7a5c3c19024e7c210d15dac79fdc460.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 402, + 545, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 402, + 545, + 474 + ], + "spans": [ + { + "bbox": [ + 305, + 402, + 545, + 474 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 402, + 545, + 474 + ], + "type": "inline_equation", + "content": "P_{uv}" + }, + { + "bbox": [ + 305, + 402, + 545, + 474 + ], + "type": "text", + "content": " is the world coordinates of the intersected pixel calculated with the camera intrinsic and extrinsic. " + }, + { + "bbox": [ + 305, + 402, + 545, + 474 + ], + "type": "inline_equation", + "content": "dist(\\cdot, \\cdot)" + }, + { + "bbox": [ + 305, + 402, + 545, + 474 + ], + "type": "text", + "content": " is the Euclidean distance, and " + }, + { + "bbox": [ + 305, + 402, + 545, + 474 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 305, + 402, + 545, + 474 + ], + "type": "text", + "content": " (much smaller than 1) and " + }, + { + "bbox": [ + 305, + 402, + 545, + 474 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 305, + 402, + 545, + 474 + ], + "type": "text", + "content": " are the hyper-parameters. Note that we decrease the opacity of floating 3D Gaussians in front of the scene surfaces to make our newly added 3D Gaussians well-optimized." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 483, + 486, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 483, + 486, + 496 + ], + "spans": [ + { + "bbox": [ + 306, + 483, + 486, + 496 + ], + "type": "text", + "content": "3.3. Tracking and Bundle Adjustment" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 502, + 545, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 502, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 305, + 502, + 545, + 587 + ], + "type": "text", + "content": "In the parallel camera tracking phase of our work, we first employ a common straightforward constant velocity assumption to initialize new poses. This assumption transforms the last known pose based on the relative transformation between the second-to-last pose and the last pose. Then, the accurate camera pose " + }, + { + "bbox": [ + 305, + 502, + 545, + 587 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 305, + 502, + 545, + 587 + ], + "type": "text", + "content": " is optimized by minimizing rendered color loss, i.e.," + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 326, + 597, + 545, + 630 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 597, + 545, + 630 + ], + "spans": [ + { + "bbox": [ + 326, + 597, + 545, + 630 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t r a c k}} = \\sum_ {m = 1} ^ {M} \\left| \\mathbf {C} _ {m} - \\hat {\\mathbf {C}} _ {m} \\right| _ {1}, \\quad \\min _ {\\mathbf {R}, \\mathbf {t}} (\\mathcal {L} _ {\\text {t r a c k}}), \\tag {10}", + "image_path": "02409c7327cb56d0db22ceb850290da14efe1c944e902ede525d711a5faba85a.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 639, + 533, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 639, + 533, + 652 + ], + "spans": [ + { + "bbox": [ + 306, + 639, + 533, + 652 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 639, + 533, + 652 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 306, + 639, + 533, + 652 + ], + "type": "text", + "content": " is the number of sampled pixels for rendering." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 653, + 545, + 713 + ], + "type": "text", + "content": "Differentiable pose estimation. According to Eqs. (3) and (4), we observe that the gradient of the camera pose " + }, + { + "bbox": [ + 305, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 305, + 653, + 545, + 713 + ], + "type": "text", + "content": " is related to three intermediate variables: " + }, + { + "bbox": [ + 305, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}^{\\prime}" + }, + { + "bbox": [ + 305, + 653, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i" + }, + { + "bbox": [ + 305, + 653, + 545, + 713 + ], + "type": "text", + "content": ", and the projected coordinate " + }, + { + "bbox": [ + 305, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_i" + }, + { + "bbox": [ + 305, + 653, + 545, + 713 + ], + "type": "text", + "content": " of Gaussian " + }, + { + "bbox": [ + 305, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "G_{i}" + }, + { + "bbox": [ + 305, + 653, + 545, + 713 + ], + "type": "text", + "content": ". By applying the chain rule of derivation, we obtain the analytical formula" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19598" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 141, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 141, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 141, + 83 + ], + "type": "text", + "content": "tion of camera pose P:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 92, + 287, + 160 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 92, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 55, + 92, + 287, + 160 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\frac {\\partial \\mathcal {L} _ {\\mathbf {c}}}{\\partial \\mathbf {P}} = \\frac {\\partial \\mathcal {L} _ {\\mathbf {c}}}{\\partial \\mathbf {C}} \\frac {\\partial \\mathbf {C}}{\\partial \\mathbf {P}} = \\frac {\\partial \\mathcal {L} _ {\\mathbf {c}}}{\\partial \\mathbf {C}} \\left(\\frac {\\partial \\mathbf {C}}{\\partial \\mathbf {c} _ {i}} \\frac {\\partial \\mathbf {c} _ {i}}{\\partial \\mathbf {P}} + \\frac {\\partial \\mathbf {C}}{\\partial \\alpha_ {i}} \\frac {\\partial \\alpha_ {i}}{\\partial \\mathbf {P}}\\right) \\\\ = \\frac {\\partial \\mathcal {L} _ {\\mathrm {c}}}{\\partial \\overline {{\\mathbf {C}}}} \\frac {\\partial \\mathbf {C}}{\\partial \\alpha_ {i}} \\left(\\frac {\\partial \\alpha_ {i}}{\\partial \\boldsymbol {\\Sigma} ^ {\\prime}} \\frac {\\partial \\boldsymbol {\\Sigma} ^ {\\prime}}{\\partial \\mathbf {P}} + \\frac {\\partial \\alpha_ {i}}{\\partial \\mathbf {m} _ {\\mathrm {i}}} \\frac {\\partial \\mathbf {m} _ {\\mathrm {i}}}{\\partial \\mathbf {P}}\\right) \\tag {11} \\\\ = \\frac {\\partial \\mathcal {L} _ {\\mathbf {c}}}{\\partial \\mathbf {C}} \\frac {\\partial \\mathbf {C}}{\\partial \\alpha_ {i}} \\left(\\frac {\\partial \\alpha_ {i}}{\\partial \\boldsymbol {\\Sigma} ^ {\\prime}} \\frac {\\partial (\\mathbf {J P} ^ {- 1} \\boldsymbol {\\Sigma} \\mathbf {P} ^ {- T} \\mathbf {J} ^ {T})}{\\partial \\mathbf {P}} + \\frac {\\partial \\alpha_ {i}}{\\partial \\mathbf {m} _ {\\mathbf {i}}} \\frac {\\partial (\\mathbf {K P X} _ {i})}{\\partial \\mathbf {P} d _ {i}}\\right) \\\\ \\end{array}", + "image_path": "8e85ab64ff14a908a75ae23dfbcb3451cf181ec1a60e5cd7892cfb48c229e912.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 168, + 286, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 168, + 286, + 267 + ], + "spans": [ + { + "bbox": [ + 47, + 168, + 286, + 267 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 168, + 286, + 267 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 47, + 168, + 286, + 267 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 47, + 168, + 286, + 267 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 47, + 168, + 286, + 267 + ], + "type": "text", + "content": "-axis coordinate of projection " + }, + { + "bbox": [ + 47, + 168, + 286, + 267 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_i" + }, + { + "bbox": [ + 47, + 168, + 286, + 267 + ], + "type": "text", + "content": ". The item " + }, + { + "bbox": [ + 47, + 168, + 286, + 267 + ], + "type": "inline_equation", + "content": "\\frac{\\partial\\mathbf{C}}{\\partial\\mathbf{c}_i}\\frac{\\partial\\mathbf{c}_i}{\\partial\\mathbf{P}}" + }, + { + "bbox": [ + 47, + 168, + 286, + 267 + ], + "type": "text", + "content": " can be eliminated because we are only concerned about the view-independent color in our tracking implementation. In addition, we find that the intermediate gradient " + }, + { + "bbox": [ + 47, + 168, + 286, + 267 + ], + "type": "inline_equation", + "content": "\\frac{\\partial(\\mathbf{KP}\\mathbf{X}_i)}{\\partial\\mathbf{P}d_i}" + }, + { + "bbox": [ + 47, + 168, + 286, + 267 + ], + "type": "text", + "content": " is the deterministic component for the camera pose " + }, + { + "bbox": [ + 47, + 168, + 286, + 267 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 47, + 168, + 286, + 267 + ], + "type": "text", + "content": ". So we simply ignore the backpropagation of " + }, + { + "bbox": [ + 47, + 168, + 286, + 267 + ], + "type": "inline_equation", + "content": "\\frac{\\partial(\\mathbf{JP}^{-1}\\pmb{\\Sigma}\\mathbf{P}^{-T}\\mathbf{J}^{T})}{\\partial\\mathbf{P}}" + }, + { + "bbox": [ + 47, + 168, + 286, + 267 + ], + "type": "text", + "content": " for efficiency. More details can be found in the supplemental materials." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 269, + 286, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 269, + 286, + 424 + ], + "spans": [ + { + "bbox": [ + 46, + 269, + 286, + 424 + ], + "type": "text", + "content": "Coarse-to-fine camera tracking. It would be problematic to optimize the camera pose with all image pixels since artifacts in images can cause drifted camera tracking. To address this issue, as shown in Fig. 2, in the differentiable pose estimation step for each frame, we first take advantage of image regularity to render only a sparse set of pixels and optimize tracking loss to obtain a coarse camera pose. This coarse optimization step significantly eases the influence of detailed artifacts. Further, we use this coarse camera pose and depth observation to select reliable 3D Gaussians, which guides GS-SLAM to render informative areas with clear geometric structures to refine coarse camera pose via further optimizing tracking loss on new rendering pixels." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 425, + 286, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 425, + 286, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 286, + 616 + ], + "type": "text", + "content": "Specifically, in the coarse stage, we first render a coarse image " + }, + { + "bbox": [ + 46, + 425, + 286, + 616 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}_c" + }, + { + "bbox": [ + 46, + 425, + 286, + 616 + ], + "type": "text", + "content": " with resolution " + }, + { + "bbox": [ + 46, + 425, + 286, + 616 + ], + "type": "inline_equation", + "content": "H / 2\\times W / 2" + }, + { + "bbox": [ + 46, + 425, + 286, + 616 + ], + "type": "text", + "content": " at uniformly sampled image coordinates and optimize tracking loss in Eq. 10 for " + }, + { + "bbox": [ + 46, + 425, + 286, + 616 + ], + "type": "inline_equation", + "content": "T_{c}" + }, + { + "bbox": [ + 46, + 425, + 286, + 616 + ], + "type": "text", + "content": " iterations, and the obtained camera pose is denoted as " + }, + { + "bbox": [ + 46, + 425, + 286, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_c" + }, + { + "bbox": [ + 46, + 425, + 286, + 616 + ], + "type": "text", + "content": ". In the fine stage, we use a similar technique with adaptive 3D Gaussian expansion strategy in Section 3.2 to select reliable 3D Gaussian to render full-resolution images while ignoring noisy 3D Gaussians that cause artifacts. In detail, we check all visible 3D Gaussians under coarse camera pose " + }, + { + "bbox": [ + 46, + 425, + 286, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_c" + }, + { + "bbox": [ + 46, + 425, + 286, + 616 + ], + "type": "text", + "content": ", and remove 3D Gaussians whose position is far away from the scene surface. Formally, for each visible 3D Gaussians " + }, + { + "bbox": [ + 46, + 425, + 286, + 616 + ], + "type": "inline_equation", + "content": "G_{i}" + }, + { + "bbox": [ + 46, + 425, + 286, + 616 + ], + "type": "text", + "content": " with position " + }, + { + "bbox": [ + 46, + 425, + 286, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_i" + }, + { + "bbox": [ + 46, + 425, + 286, + 616 + ], + "type": "text", + "content": ", we project it to the camera plane using coarse camera pose " + }, + { + "bbox": [ + 46, + 425, + 286, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_c" + }, + { + "bbox": [ + 46, + 425, + 286, + 616 + ], + "type": "text", + "content": " and camera intrinsic. Given the projected pixel's depth observation " + }, + { + "bbox": [ + 46, + 425, + 286, + 616 + ], + "type": "inline_equation", + "content": "D_{i}" + }, + { + "bbox": [ + 46, + 425, + 286, + 616 + ], + "type": "text", + "content": " and the distance " + }, + { + "bbox": [ + 46, + 425, + 286, + 616 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 46, + 425, + 286, + 616 + ], + "type": "text", + "content": " that is between 3D Gaussians " + }, + { + "bbox": [ + 46, + 425, + 286, + 616 + ], + "type": "inline_equation", + "content": "G_{i}" + }, + { + "bbox": [ + 46, + 425, + 286, + 616 + ], + "type": "text", + "content": " and the camera image plane, the reliable 3D Gaussians are selected as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 624, + 287, + 655 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 287, + 655 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {G} _ {\\text {s e l e c t e d}} = \\left\\{G _ {i} \\mid G _ {i} \\in \\mathbf {G} \\text {a n d} a b s \\left(D _ {i} - d _ {i}\\right) \\leq \\varepsilon \\right\\}, \\tag {12} \\\\ \\tilde {\\mathbf {I}} _ {f} = \\mathcal {F} (u, v, \\mathbf {G} _ {s e l e c t e d}), \\\\ \\end{array}", + "image_path": "7df5811f2e896cbb3372610cf5684b2d9fa5ab621fe97fa2f21b7bf335d743c7.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 664, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 664, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 664, + 287, + 714 + ], + "type": "text", + "content": "where we use the selected reliable 3D Gaussians to render full-resolution images " + }, + { + "bbox": [ + 47, + 664, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}_f" + }, + { + "bbox": [ + 47, + 664, + 287, + 714 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 664, + 287, + 714 + ], + "type": "inline_equation", + "content": "u, v" + }, + { + "bbox": [ + 47, + 664, + 287, + 714 + ], + "type": "text", + "content": " denote the pixel coordinates in " + }, + { + "bbox": [ + 47, + 664, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}_f" + }, + { + "bbox": [ + 47, + 664, + 287, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 664, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 47, + 664, + 287, + 714 + ], + "type": "text", + "content": " represents the color splattering rendering function. The final camera pose " + }, + { + "bbox": [ + 47, + 664, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 47, + 664, + 287, + 714 + ], + "type": "text", + "content": " is obtained by optimizing" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 181 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 181 + ], + "type": "text", + "content": "tracking loss in Eq. 10 with " + }, + { + "bbox": [ + 304, + 72, + 545, + 181 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}_f" + }, + { + "bbox": [ + 304, + 72, + 545, + 181 + ], + "type": "text", + "content": " for other " + }, + { + "bbox": [ + 304, + 72, + 545, + 181 + ], + "type": "inline_equation", + "content": "T_{f}" + }, + { + "bbox": [ + 304, + 72, + 545, + 181 + ], + "type": "text", + "content": " iterations. Note that " + }, + { + "bbox": [ + 304, + 72, + 545, + 181 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}_c" + }, + { + "bbox": [ + 304, + 72, + 545, + 181 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 181 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}_f" + }, + { + "bbox": [ + 304, + 72, + 545, + 181 + ], + "type": "text", + "content": " are only rendered at previously observed areas, avoiding rendering areas where 3D scene representations have not been optimized in the mapping process. Also, we add keyframes based on the proportion of the currently observed image's reliable region to the overall image. At the same time, when the current tracking frame and most recent keyframe differ by more than a threshold value " + }, + { + "bbox": [ + 304, + 72, + 545, + 181 + ], + "type": "inline_equation", + "content": "\\mu_{k}" + }, + { + "bbox": [ + 304, + 72, + 545, + 181 + ], + "type": "text", + "content": ", this frame will be inserted as a keyframe." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 182, + 545, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 182, + 545, + 301 + ], + "spans": [ + { + "bbox": [ + 304, + 182, + 545, + 301 + ], + "type": "text", + "content": "**Bundle adjustment.** In the bundle adjustment (BA) phase, we optimize the camera poses " + }, + { + "bbox": [ + 304, + 182, + 545, + 301 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 304, + 182, + 545, + 301 + ], + "type": "text", + "content": " and the 3D Gaussian scene representation " + }, + { + "bbox": [ + 304, + 182, + 545, + 301 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 304, + 182, + 545, + 301 + ], + "type": "text", + "content": " jointly. We randomly select " + }, + { + "bbox": [ + 304, + 182, + 545, + 301 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 182, + 545, + 301 + ], + "type": "text", + "content": " keyframes from the keyframe database for optimization, using the loss function similar to the mapping part. For pose optimization stability, we only optimize the scene representation " + }, + { + "bbox": [ + 304, + 182, + 545, + 301 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 304, + 182, + 545, + 301 + ], + "type": "text", + "content": " in the first half of the iterations. In the other half of the iterations, we simultaneously optimize the map and the poses. Then, the accurate camera pose " + }, + { + "bbox": [ + 304, + 182, + 545, + 301 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 304, + 182, + 545, + 301 + ], + "type": "text", + "content": " is optimized by minimizing rendering color loss, i.e.," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 308, + 545, + 335 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 308, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 315, + 308, + 545, + 335 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {b a} = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\sum_ {m = 1} ^ {H W} \\left| D _ {m} - \\hat {D} _ {m} \\right| _ {1} + \\lambda_ {m} \\left| \\mathbf {C} _ {m} - \\hat {\\mathbf {C}} _ {m} \\right| _ {1}, \\min _ {\\mathbf {R}, \\mathbf {t}, \\mathbf {S}} (\\mathcal {L} _ {b a}). \\tag {13}", + "image_path": "b166c5711bbb4368fac6a0b17a57a1710fcc66a276bacfa7651c905d70d219e8.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 341, + 382, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 341, + 382, + 354 + ], + "spans": [ + { + "bbox": [ + 306, + 341, + 382, + 354 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 360, + 422, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 360, + 422, + 373 + ], + "spans": [ + { + "bbox": [ + 306, + 360, + 422, + 373 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 378, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 378, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 378, + 545, + 449 + ], + "type": "text", + "content": "Dataset. To evaluate the performance of GS-SLAM, we conduct experiments on the Replica [31], and TUM-RGBD [33]. Following [11, 27, 41, 48, 55], we use 8 scenes from the Replica dataset for localization, mesh reconstruction, and rendering quality comparison. The selected three subsets of TUM-RGBD datasets are used for localization." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 450, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 545, + 521 + ], + "type": "text", + "content": "Baselines. We compare our method with existing SOTA NeRF-based dense visual SLAM: NICE-SLAM [55], VoxFusion [48], CoSLAM [41], ESLAM [11] and Point-SLAM [27]. The rendering performance of CoSLAM [41] and ESLAM [11] is conducted from the open source code with the same configuration in [27]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 522, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 545, + 641 + ], + "type": "text", + "content": "Metric. For mesh reconstruction, we use the 2D Depth L1 (cm) [55], the Precision " + }, + { + "bbox": [ + 304, + 522, + 545, + 641 + ], + "type": "inline_equation", + "content": "(\\mathbb{P},\\%)" + }, + { + "bbox": [ + 304, + 522, + 545, + 641 + ], + "type": "text", + "content": ", Recall " + }, + { + "bbox": [ + 304, + 522, + 545, + 641 + ], + "type": "inline_equation", + "content": "(\\mathbb{R},\\%)" + }, + { + "bbox": [ + 304, + 522, + 545, + 641 + ], + "type": "text", + "content": ", and F-score with a threshold of " + }, + { + "bbox": [ + 304, + 522, + 545, + 641 + ], + "type": "inline_equation", + "content": "1\\mathrm{cm}" + }, + { + "bbox": [ + 304, + 522, + 545, + 641 + ], + "type": "text", + "content": " to measure the scene geometry. For localization, we use the absolute trajectory (ATE, cm) error [33] to measure the accuracy of the estimated camera poses. We further evaluate the rendering performance using the peak signal-to-noise ratio (PSNR), SSIM [43], and LPIPS [52] by following [27]. To be fair, we run all the methods on a dataset 10 times and report the average results. More details can be found in the supplemental materials." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "text", + "content": "Implementation details. GS-SLAM is implemented in Python using the PyTorch framework, incorporating CUDA code for Gaussian splatting and trained on a desktop PC with a 5.50GHz Intel Core i9-13900K CPU and NVIDIA RTX 4090 GPU. We extended the existing code for differentiable Gaussian splatting rasterization with additional func" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19599" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "content": "tionality for handling depth, pose, and cumulative opacity during both forward and backward propagation. More details can be found in the supplemental materials." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 116, + 259, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 116, + 259, + 129 + ], + "spans": [ + { + "bbox": [ + 47, + 116, + 259, + 129 + ], + "type": "text", + "content": "4.2. Evaluation of Localization and Mapping" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 135, + 287, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 135, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 135, + 287, + 350 + ], + "type": "text", + "content": "Evaluation on Replica. Tracking ATE: Tab. 1 illustrates the tracking performance of our method and the state-of-the-art methods on the Replica dataset. Our method achieves the best or second performance in 7 of 8 scenes and outperforms the second-best method Point-SLAM [27] by " + }, + { + "bbox": [ + 46, + 135, + 287, + 350 + ], + "type": "inline_equation", + "content": "0.4\\mathrm{cm}" + }, + { + "bbox": [ + 46, + 135, + 287, + 350 + ], + "type": "text", + "content": " on average at 8.34 FPS. It is noticeable that the second best method, Point-SLAM [27] runs at 0.42 FPS, which is " + }, + { + "bbox": [ + 46, + 135, + 287, + 350 + ], + "type": "inline_equation", + "content": "20\\times" + }, + { + "bbox": [ + 46, + 135, + 287, + 350 + ], + "type": "text", + "content": " slower than our method, indicating that GS-SLAM achieves a better trade-off between the tracking accuracy and the runtime efficiency. Mapping ACC: Tab. 3 report the mapping evaluation results of our method with other current state-of-the-art visual SLAM methods. GS-SLAM achieves the best performance in Depth L1 (1.16cm) and Precision " + }, + { + "bbox": [ + 46, + 135, + 287, + 350 + ], + "type": "inline_equation", + "content": "(74.0\\%)" + }, + { + "bbox": [ + 46, + 135, + 287, + 350 + ], + "type": "text", + "content": " metrics on average. For Recall and F1 scores, GS-SLAM performs comparably to the second best method CoSLAM [41]. The visualization results in Fig. 4 show that GS-SLAM achieves satisfying construction mesh with clear boundaries and details." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 352, + 287, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 352, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 352, + 287, + 448 + ], + "type": "text", + "content": "Evaluation on TUM-RGBD. Tab. 2 compares GS-SLAM with the other SLAM systems in TUM-RGBD dataset. Our method surpasses iMAP [35], NICE-SLAM [55] and Voxfusion [48], and achieves a comparable performance, average " + }, + { + "bbox": [ + 46, + 352, + 287, + 448 + ], + "type": "inline_equation", + "content": "3.7\\mathrm{cm}" + }, + { + "bbox": [ + 46, + 352, + 287, + 448 + ], + "type": "text", + "content": " ATE RSME, with the SOTA methods. A gap to traditional methods still exists between the neural vSLAM and the traditional SLAM systems, which employ more sophisticated tracking schemes [27]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 452, + 287, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 452, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 46, + 452, + 287, + 496 + ], + "type": "text", + "content": "Table 1. Tracking comparison (ATE RMSE [cm]) of the proposed method vs. the SOTA methods on the Replica dataset. The running speed of methods in the upper part is lower than 5 FPS, * denotes the reproduced results by running officially released code." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 48, + 498, + 287, + 553 + ], + "blocks": [ + { + "bbox": [ + 48, + 498, + 287, + 553 + ], + "lines": [ + { + "bbox": [ + 48, + 498, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 48, + 498, + 287, + 553 + ], + "type": "table", + "html": "
MethodRm0Rm1Rm2Off0Off1Off2Off3Off4avg
Point-SLAM [27]0.560.470.300.350.620.550.720.730.54
NICE-SLAM [55]0.971.311.070.881.001.061.101.131.06
Vox-Fusion* [48]1.374.701.478.482.042.581.112.943.09
ESLAM [11]0.710.700.520.570.550.580.720.630.63
CoSLAM [41]0.700.951.350.590.552.031.560.721.00
Ours0.480.530.330.520.410.590.460.70.50
", + "image_path": "ebaacfcd04b4e7c1c8485faeb94dabce369a34c211e452d8eb12edcea2513134.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 48, + 597, + 286, + 647 + ], + "blocks": [ + { + "bbox": [ + 46, + 562, + 286, + 594 + ], + "lines": [ + { + "bbox": [ + 46, + 562, + 286, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 562, + 286, + 594 + ], + "type": "text", + "content": "Table 2. Tracking ATE [cm] on TUM-RGBD [33]. Our method achieves a comparable performance among the neural vSLAMs. * denotes the reproduced results by running officially released code." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 597, + 286, + 647 + ], + "lines": [ + { + "bbox": [ + 48, + 597, + 286, + 647 + ], + "spans": [ + { + "bbox": [ + 48, + 597, + 286, + 647 + ], + "type": "table", + "html": "
Methodfr1Deskfr2_xyzfr3_offAvg.Methodfr1Deskfr2_xyzfr3_offAvg.
DI-Fusion [9]4.42.05.84.1NICE-SLAM [55]4.331.73.913.3
Elastic-Fusion [46]2.51.22.52.1Vox-Fusion* [48]3.51.526.010.3
BAD-SLAM [30]1.71.11.71.5CoSLAM [41]2.71.92.62.4
Kintinous [45]3.72.93.03.2ESLAM [11]2.31.12.42.0
ORB-SLAM2 [20]1.60.41.01.0Point-SLAM2.61.33.22.4
iMAP* [35]7.22.19.06.1Ours3.31.36.63.7
", + "image_path": "d8143725206c2ca42e876e971b760f321256520e7b29be9c0fbcdc030081115c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 658, + 173, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 173, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 173, + 671 + ], + "type": "text", + "content": "4.3. Rendering Evaluation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "We compare the rendering performance of the proposed GS-SLAM with the neural visual SLAM methods in Tab. 6. The results show that GS-SLAM achieves the best performance" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 311, + 98, + 545, + 271 + ], + "blocks": [ + { + "bbox": [ + 305, + 74, + 544, + 95 + ], + "lines": [ + { + "bbox": [ + 305, + 74, + 544, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 74, + 544, + 95 + ], + "type": "text", + "content": "Table 3. Reconstruction comparison of the proposed method vs. the SOTA methods on Replica dataset." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 311, + 98, + 545, + 271 + ], + "lines": [ + { + "bbox": [ + 311, + 98, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 311, + 98, + 545, + 271 + ], + "type": "table", + "html": "
MethodMetricRm 0Rm 1Rm 2Off 0Off 1Off 2Off 3Off 4Avg.
NICESL AM [55]Depth L1 ↓1.811.442.041.391.768.334.992.012.97
Precision ↑45.8643.7644.3851.4050.8038.3740.8537.3544.10
Recall↑44.1046.1242.7848.6653.0839.9839.0435.7743.69
F1↑44.9644.8443.5649.9951.9139.1639.9236.5443.86
VoxFusion [48]Depth L1 ↓1.091.902.212.323.404.192.961.612.46
Precision↑75.8335.8863.1048.5143.5054.4869.1155.4055.73
Recall↑64.8933.0756.6244.7638.4447.8560.6146.7949.13
F1↑69.9334.3859.6746.5440.8150.9564.5650.7252.20
CoSLAM [41]Depth L1 ↓0.990.822.281.241.617.704.651.432.59
Precision↑81.7177.9573.3079.4180.6755.6457.6379.7673.26
Recall↑74.0370.7965.7371.4670.3552.9656.0671.2266.58
F1↑77.6874.2069.3175.2375.1654.2756.8375.2569.74
ESL AM [11]Depth L1 ↓0.630.620.980.571.667.323.940.882.08
Precision↑74.3375.9482.4872.2065.7470.7372.4872.2473.27
Recall↑87.3787.0184.9988.3684.3881.9279.1880.6384.23
F1↑80.3281.1083.7279.4773.9075.9275.6876.2178.29
OursDepth L1 ↓1.310.821.260.810.961.411.531.081.16
Precision↑64.5883.1170.1383.4387.7770.9163.1868.8874.00
Recall↑61.2976.8363.8476.9076.1561.6362.9161.5067.63
F1↑62.8979.8566.8480.0381.5565.9559.1764.9870.15
", + "image_path": "d37d4a79788887ecfeb1728517e76e4b8abb690ff93cac0de321fabed79cb399.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 288, + 545, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 288, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 288, + 545, + 479 + ], + "type": "text", + "content": "in all the metrics. Our method significantly outperforms the second-best methods CoSLAM [41], ESLAM [11] and NICE-SLAM [55] by 1.52 dB in PSNR, 0.027 in SSIM and 0.12 in LPIPS, respectively. It is noticeable that GS-SLAM achieves 386 FPS rendering speed on average, which is " + }, + { + "bbox": [ + 304, + 288, + 545, + 479 + ], + "type": "inline_equation", + "content": "100 \\times" + }, + { + "bbox": [ + 304, + 288, + 545, + 479 + ], + "type": "text", + "content": " faster than the second-best method Vox-Fusion [48]. This excellent rendering performance is attributed to the efficient 3D Gaussian rendering pipeline and can be further applied to real-time downstream tasks, such as VR [5], robot navigation [7] and autonomous driving [2]. The visualization results in Fig. 5 show that GS-SLAM can generate much more high-quality and realistic images than the other methods, especially in edge areas with detailed structures. While NICE-SLAM [55] causes severe artifacts and blurs, CoSLAM [41] and ESLAM [11] generate blur around the image boundaries." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 307, + 489, + 544, + 689 + ], + "blocks": [ + { + "bbox": [ + 307, + 489, + 544, + 689 + ], + "lines": [ + { + "bbox": [ + 307, + 489, + 544, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 544, + 689 + ], + "type": "image", + "image_path": "fec0ca7d56512019e46711d86a69001801b6665ef0be59affaa59b06322d9b83.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 691, + 544, + 713 + ], + "lines": [ + { + "bbox": [ + 305, + 691, + 544, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 691, + 544, + 713 + ], + "type": "text", + "content": "Figure 4. Reconstruction performance comparison of the proposed GS-SLAM and SOTA methods on the Replica dataset." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19600" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 547, + 304 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 547, + 304 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 547, + 304 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 547, + 304 + ], + "type": "image", + "image_path": "fdcd81bd916aabd7dbf1de73f9b520bda2fa382663cb7aaaf6a756593cc4751d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 306, + 547, + 329 + ], + "lines": [ + { + "bbox": [ + 46, + 306, + 547, + 329 + ], + "spans": [ + { + "bbox": [ + 46, + 306, + 547, + 329 + ], + "type": "text", + "content": "Figure 5. The render visualization results on the Replica dataset of the proposed GS-SLAM and SOTA methods. GS-SLAM can generate much more high-quality and realistic images than the other methods, especially around the object boundaries." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 334, + 153, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 334, + 153, + 346 + ], + "spans": [ + { + "bbox": [ + 47, + 334, + 153, + 346 + ], + "type": "text", + "content": "4.4. Runtime Analysis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 352, + 289, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 352, + 289, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 352, + 289, + 604 + ], + "type": "text", + "content": "Tab. 4 and Tab. 5 illustrate the runtime and memory usage of GS-SLAM and the state-of-the-art methods on the Replica and TUM-RGBD, respectively. We report the parameters of the neural networks and the memory usage of the scene representation. Note that Point-SLAM uses extra memory dynamic radius to improve performance (mark as " + }, + { + "bbox": [ + 46, + 352, + 289, + 604 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 46, + 352, + 289, + 604 + ], + "type": "text", + "content": "). The results show that GS-SLAM achieves a competitive running speed with 8.34 FPS compared to the other Radiance Fields-based vSLAMs. Note that we do not use any neural network decoder in our system, which results in the zero learnable parameter. However, the 3D Gaussian scene representations of GS-SLAM consume 198.04 MB memory, " + }, + { + "bbox": [ + 46, + 352, + 289, + 604 + ], + "type": "inline_equation", + "content": "4\\times" + }, + { + "bbox": [ + 46, + 352, + 289, + 604 + ], + "type": "text", + "content": " larger than the second large method NICE-SLAM [55]. Memory usage is mainly caused by spherical harmonic coefficients in training, which is a common constraint among Gaussian splatting-based reconstruction methods. Despite this, we still achieve a " + }, + { + "bbox": [ + 46, + 352, + 289, + 604 + ], + "type": "inline_equation", + "content": "20\\times" + }, + { + "bbox": [ + 46, + 352, + 289, + 604 + ], + "type": "text", + "content": " faster FPS compared to the similar point-based method Point-SLAM [27]. Besides, we also provide a light version of GS-SLAM with zero-order spherical harmonic coefficients, significantly reducing memory usage while maintaining stable performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 611, + 141, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 141, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 141, + 624 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 629, + 287, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 666 + ], + "type": "text", + "content": "We perform the ablation of GS-SLAM on the Replica dataset #Room0 subset to evaluate the effectiveness of coarse-to-fine tracking, and expansion mapping strategy." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "content": "Effect of our expansion strategy for mapping. Tab. 7 shows the ablation of our proposed expansion strategy for mapping. The results illustrate that the expansion strategy can significantly improve the tracking and mapping per" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 308, + 371, + 545, + 440 + ], + "blocks": [ + { + "bbox": [ + 305, + 336, + 545, + 369 + ], + "lines": [ + { + "bbox": [ + 305, + 336, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 305, + 336, + 545, + 369 + ], + "type": "text", + "content": "Table 4. Runtime and memory usage on Replica #Room0. The decoder parameters and embedding denote the parameter number of MLPs and the memory usage of the scene representation." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 371, + 545, + 440 + ], + "lines": [ + { + "bbox": [ + 308, + 371, + 545, + 440 + ], + "spans": [ + { + "bbox": [ + 308, + 371, + 545, + 440 + ], + "type": "table", + "html": "
MethodTracking [ms×it] ↓Mapping [ms×it] ↓System FPS ↑Decoder param ↓Scene Embedding↓
Point-SLAM [27]0.06 × 4034.81 × 3000.420.127 M55.42 (+12453.2)†MB
NICE-SLAM [55]6.64 × 1028.63 × 602.910.06 M48.48 MB
Vox-Fusion [48]0.03 × 3066.53 × 101.280.054 M1.49 MB
CoSLAM [55]6.01 × 1013.18 × 1016.641.671 M—
ESLAM [11]6.85 × 819.87 × 1513.420.003 M27.12 MB
GS-SLAM11.9 × 1012.8 × 1008.340 M198.04 MB
", + "image_path": "f8e304c563eaade375dcbd49ed189484cdc240d3ccf7e6140946a15b92b24604.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 308, + 469, + 545, + 526 + ], + "blocks": [ + { + "bbox": [ + 306, + 445, + 545, + 467 + ], + "lines": [ + { + "bbox": [ + 306, + 445, + 545, + 467 + ], + "spans": [ + { + "bbox": [ + 306, + 445, + 545, + 467 + ], + "type": "text", + "content": "Table 5. Runtime and memory usage on TUM-RGBD dataset #fr1Desk and #fr2_xyz." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 469, + 545, + 526 + ], + "lines": [ + { + "bbox": [ + 308, + 469, + 545, + 526 + ], + "spans": [ + { + "bbox": [ + 308, + 469, + 545, + 526 + ], + "type": "table", + "html": "
Method#fr1Desk#fr2.xyz
FPS ↑Memory ↓FPS ↑Memory ↓
Point-SLAM0.1018.3 (+160.7)†MB0.1214.2 (+7687.4)†MB
NICE-SLAM [55]0.11178.8MB0.12484.0MB
ESLAM [11]0.3127.2MB0.3151.6MB
GS-SLAM1.83(ATE:3.3)40.8MB1.51(ATE:1.3)48.4MB
GS-SLAM (light)1.92(ATE:4.3)18.8MB1.68(ATE:2.7)22.3MB
", + "image_path": "96303eb6c0537cdc8fabb06d7fbe1332a903fc5d8086c07a0b59d0fc0d4b708a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": "formance. The implementation w/o adding means that we only initialize 3D Gaussians in the first frame and optimize the scene without adding new points. However, this strategy completely crashes because the density control in [13] can not handle real-time mapping tasks without an accurate point cloud input. Besides, the implementation w/o deletion suffers from a large number of redundant and noisy 3D Gaussian, which causes undesirable supervision. In contrast, the proposed expansion strategy effectively improves the tracking and mapping performance by 0.1 in ATE and 11.97 in Recall by adding more accurate constraints for the optimization. According to the visualization results in Fig. 6, our full implementation achieves more high-quality and detailed rendering and reconstruction results than the w/o delete strategy." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19601" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 79, + 547, + 236 + ], + "blocks": [ + { + "bbox": [ + 47, + 56, + 545, + 78 + ], + "lines": [ + { + "bbox": [ + 47, + 56, + 545, + 78 + ], + "spans": [ + { + "bbox": [ + 47, + 56, + 545, + 78 + ], + "type": "text", + "content": "Table 6. Rendering performance on Replica dataset. We outperform existing dense neural RGB-D methods on the commonly reported rendering metrics. Note that GS-SLAM achieves 386 FPS on average, benefiting from the efficient Gaussian scene representation." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 79, + 547, + 236 + ], + "lines": [ + { + "bbox": [ + 47, + 79, + 547, + 236 + ], + "spans": [ + { + "bbox": [ + 47, + 79, + 547, + 236 + ], + "type": "table", + "html": "
MethodMetricRoom 0Room 1Room 2Office 0Office 1Office 2Office 3Office 4Avg.FPS.
NICE-SLAM [55]PSNR [dB] ↑22.1222.4724.5229.0730.3419.6622.2324.9424.42
SSIM ↑0.6890.7570.8140.8740.8860.7970.8010.8560.8090.30
LPIPS ↓0.3300.2710.2080.2290.1810.2350.2090.1980.233
Vox-Fusion* [48]PSNR [dB] ↑22.3922.3623.9227.7929.8320.3323.4725.2124.413.88
SSIM ↑0.6830.7510.7980.8570.8760.7940.8030.8470.801
LPIPS ↓0.3030.2690.2340.2410.1840.2430.2130.1990.236
CoSLAM [41]PSNR [dB] ↑27.2728.4529.0634.1434.8728.4328.7630.9130.243.68
SSIM ↑0.9100.9090.9320.9610.9690.9380.9410.9550.939
LPIPS ↓0.3240.2940.2660.2090.1960.2580.2290.2360.252
ESLAM [11]PSNR [dB] ↑25.3227.7729.0833.7130.2028.0928.7729.7129.082.82
SSIM ↑0.8750.9020.9320.9600.9230.9430.9480.9450.929
LPIPS ↓0.3130.2980.2480.1840.2280.2410.1960.2040.336
OursPSNR [dB] ↑31.5632.8632.5938.7041.1732.3632.0332.9234.27386.91
SSIM ↑0.9680.9730.9710.9860.9930.9780.9700.9680.975
LPIPS ↓0.0940.0750.0930.0500.0330.0940.1100.1120.082
", + "image_path": "e4de8bbd01b3a458bb8d5cf6142d9bfff8b35bb3cd95e372d5dd9fe42c1a393f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 261, + 279, + 308 + ], + "blocks": [ + { + "bbox": [ + 47, + 239, + 286, + 260 + ], + "lines": [ + { + "bbox": [ + 47, + 239, + 286, + 260 + ], + "spans": [ + { + "bbox": [ + 47, + 239, + 286, + 260 + ], + "type": "text", + "content": "Table 7. Ablation of the adaptive 3D Gaussian expansion strategy on Replica #Room0." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 261, + 279, + 308 + ], + "lines": [ + { + "bbox": [ + 56, + 261, + 279, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 279, + 308 + ], + "type": "table", + "html": "
Setting#Room0
ATE↓Depth L1↓Precision↑Recall ↑F1↑PSNR↑SSIM↑LPIPS↓
w/o addXXXXXXXX
w/o delete0.581.6853.5549.3251.3531.220.9670.094
w/ add & delete0.481.3164.5861.2962.8931.560.9680.094
", + "image_path": "060edc2b93575b53a5c40e29b19007b057268823217e8702d5f14500fd7c8567.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 56, + 331, + 279, + 376 + ], + "blocks": [ + { + "bbox": [ + 47, + 310, + 286, + 330 + ], + "lines": [ + { + "bbox": [ + 47, + 310, + 286, + 330 + ], + "spans": [ + { + "bbox": [ + 47, + 310, + 286, + 330 + ], + "type": "text", + "content": "Table 8. Ablation of the coarse-to-fine tracking strategy on Replica #Room0." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 331, + 279, + 376 + ], + "lines": [ + { + "bbox": [ + 56, + 331, + 279, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 279, + 376 + ], + "type": "table", + "html": "
Setting#Room0
ATE↓Depth L1↓Precision↑Recall ↑F1↑PSNR↑SSIM↑LPIPS↓
Coarse0.911.4859.6857.5456.5029.130.9540.120
Fine0.491.3962.6159.1861.2930.840.9640.096
Coarse-to-fine0.481.3164.5861.2962.8931.560.9680.094
", + "image_path": "8638d8a9fdd7b756aed4e5f8639dace2d83ce64aa5d26879543034ead65db6e6.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 380, + 287, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 380, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 287, + 513 + ], + "type": "text", + "content": "Effect of coarse-to-fine tracking. According to the results in Tab. 8, the proposed coarse-to-fine tracking strategy performs best in all tracking, mapping, and rendering metrics. Compared with fine tracking, the coarse-to-fine tracking strategy significantly improves the performance by 0.01 in tracking ATE, 2.11 in Recall, and 0.72 in PSNR. Although the fine strategy surpasses the coarse strategy in precision, it suffers from the artifacts and noise in the reconstructed scene, leading to a fluctuation optimization. The coarse-to-fine strategy effectively avoids noise reconstruction and improves accuracy and robustness." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 520, + 221, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 520, + 221, + 533 + ], + "spans": [ + { + "bbox": [ + 47, + 520, + 221, + 533 + ], + "type": "text", + "content": "4.6. Efficiency-to-accuracy trade-off." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 538, + 287, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 538, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 46, + 538, + 287, + 658 + ], + "type": "text", + "content": "3DGS-based SLAM trade-off focuses not only on the efficiency of tracking and mapping but also emphasizes high-quality ultra-real-time rendering. As shown in Fig. 7b, GS-SLAM achieves " + }, + { + "bbox": [ + 46, + 538, + 287, + 658 + ], + "type": "inline_equation", + "content": "\\approx 400" + }, + { + "bbox": [ + 46, + 538, + 287, + 658 + ], + "type": "text", + "content": " FPS ultra-fast speed and highest PSNR in map rendering. At the same time, our method remains a competitive system FPS and lowest tracking ATE in Fig. 7a. Moreover, GS-SLAM shows great potential in memory reduction in Tab. 5, and comparable in mesh reconstruction. Note that baselines directly use 3DGS in Fig. 7, resulting in inferior performances." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 669, + 204, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 204, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 204, + 681 + ], + "type": "text", + "content": "5. Conclusion and Limitations" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "We introduced GS-SLAM, a novel dense visual SLAM method leveraging 3D Gaussian Splatting for efficient map" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 320, + 238, + 427, + 361 + ], + "blocks": [ + { + "bbox": [ + 320, + 238, + 427, + 361 + ], + "lines": [ + { + "bbox": [ + 320, + 238, + 427, + 361 + ], + "spans": [ + { + "bbox": [ + 320, + 238, + 427, + 361 + ], + "type": "image", + "image_path": "e6ec1207e2f1c7667e3b982ea4481f5291ef8ba4b49e1dce14416e5d3f772787.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 364, + 545, + 385 + ], + "lines": [ + { + "bbox": [ + 306, + 364, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 306, + 364, + 545, + 385 + ], + "type": "text", + "content": "Figure 6. Rendering and mesh visualization of the adaptive 3D Gaussian expansion ablation on Replica #Room0." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 430, + 239, + 529, + 361 + ], + "blocks": [ + { + "bbox": [ + 430, + 239, + 529, + 361 + ], + "lines": [ + { + "bbox": [ + 430, + 239, + 529, + 361 + ], + "spans": [ + { + "bbox": [ + 430, + 239, + 529, + 361 + ], + "type": "image", + "image_path": "089e70d5edcd2ffc6b016ee37ec8ac84ec8232cb7c3d3e4e29e2b47e2b3178e3.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 306, + 390, + 416, + 484 + ], + "blocks": [ + { + "bbox": [ + 306, + 390, + 416, + 484 + ], + "lines": [ + { + "bbox": [ + 306, + 390, + 416, + 484 + ], + "spans": [ + { + "bbox": [ + 306, + 390, + 416, + 484 + ], + "type": "image", + "image_path": "db25191b988ea444aa2308b4496a313ed24b0dcc14cad4d704ffcfb923089b5c.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 484, + 403, + 493 + ], + "lines": [ + { + "bbox": [ + 320, + 484, + 403, + 493 + ], + "spans": [ + { + "bbox": [ + 320, + 484, + 403, + 493 + ], + "type": "text", + "content": "(a) Tracking performance" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 495, + 545, + 517 + ], + "lines": [ + { + "bbox": [ + 306, + 495, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 306, + 495, + 545, + 517 + ], + "type": "text", + "content": "Figure 7. Bi-criteria figure of tracking/render performance and system FPS on Replica #Office0." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 433, + 391, + 543, + 484 + ], + "blocks": [ + { + "bbox": [ + 433, + 391, + 543, + 484 + ], + "lines": [ + { + "bbox": [ + 433, + 391, + 543, + 484 + ], + "spans": [ + { + "bbox": [ + 433, + 391, + 543, + 484 + ], + "type": "image", + "image_path": "346e7ac0b3ae2b5afc7b9731206c46252952a57766f796d562e794810a2f9413.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 450, + 484, + 527, + 493 + ], + "lines": [ + { + "bbox": [ + 450, + 484, + 527, + 493 + ], + "spans": [ + { + "bbox": [ + 450, + 484, + 527, + 493 + ], + "type": "text", + "content": "(b) Render performance" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 521, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 628 + ], + "type": "text", + "content": "ping and accurate camera pose estimation, striking a better speed-accuracy balance. However, its reliance on high-quality depth data may limit performance in certain conditions. Additionally, the approach's high memory requirements for large scenes suggest future improvements could focus on optimizing memory use, potentially via techniques such as quantization and clustering. We believe GS-SLAM has the potential to extend to larger scale with some improvements and will explore this in future work." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 630, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgements. This work is supported by the Shanghai AI Laboratory, National Key R&D Program of China (2022ZD0160101), the National Natural Science Foundation of China (62376222), Young Elite Scientists Sponsorship Program by CAST (2023QNRC001) and the Early Career Scheme of the Research Grants Council (RGC) of the Hong Kong SAR under grant No. 26202321." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19602" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Michael Bloesch, Jan Czarnowski, Ronald Clark, Stefan Leutenegger, and Andrew J. Davison. Codeslam - learning a compact, optimisable representation for dense visual slam. CVPR, pages 2560-2568, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 178 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 178 + ], + "type": "text", + "content": "[2] Guillaume Bresson, Zayed Alsayed, Li Yu, and Sébastien Glaser. Simultaneous localization and mapping: A survey of current trends in autonomous driving. TIV, 2:194-220, 2017. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 181, + 287, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 181, + 287, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 181, + 287, + 203 + ], + "type": "text", + "content": "[3] Zilong Chen, Feng Wang, and Huaping Liu. Text-to-3d using gaussian splatting. ArXiv, abs/2309.16585, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 205, + 287, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 287, + 236 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 287, + 236 + ], + "type": "text", + "content": "[4] Brian Curless and Marc Levoy. Volumetric method for building complex models from range images. In SIGGRAPH, ACM, 1996. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 239, + 287, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 287, + 271 + ], + "type": "text", + "content": "[5] Parth Rajesh Desai, Pooja Nikhil Desai, Komal Deepak Ajmera, and Khushbu Mehta. A review paper on oculus rift-a virtual reality headset. ArXiv, abs/1408.1173, 2014. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 273, + 287, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 273, + 287, + 295 + ], + "spans": [ + { + "bbox": [ + 53, + 273, + 287, + 295 + ], + "type": "text", + "content": "[6] Hugh F. Durrant-Whyte and Tim Bailey. Simultaneous localization and mapping: part i. RAM, 13:99–110, 2006. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 297, + 287, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 297, + 287, + 328 + ], + "spans": [ + { + "bbox": [ + 53, + 297, + 287, + 328 + ], + "type": "text", + "content": "[7] Christian Hane, Christopher Zach, Jongwoo Lim, Ananth Ranganathan, and Marc Pollefeys. Stereo depth map fusion for robot navigation. IROS, pages 1618-1625, 2011. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 330, + 287, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 330, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 53, + 330, + 287, + 374 + ], + "type": "text", + "content": "[8] Caner Hazirbas, Andreas Wiedemann, Robert Maier, Laura Leal-Taixe, and Daniel Cremers. Tum rgb-d scribble-based segmentation benchmark. https://github.com/tum-vision/rgbdscribble_benchmark, 2018.6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 376, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 376, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 376, + 287, + 407 + ], + "type": "text", + "content": "[9] Jiahui Huang, Shi-Sheng Huang, Haoxuan Song, and ShiMin Hu. Di-fusion: Online implicit 3d reconstruction with deep priors. In CVPR, pages 8932-8941, 2021. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 409, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 409, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 48, + 409, + 287, + 441 + ], + "type": "text", + "content": "[10] Xudong Jiang, Lifeng Zhu, Jia Liu, and Aiguo Song. A slam-based 6 dof controller with smooth auto-calibration for virtual reality. TVC, 39:3873 - 3886, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 443, + 287, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 443, + 287, + 486 + ], + "spans": [ + { + "bbox": [ + 48, + 443, + 287, + 486 + ], + "type": "text", + "content": "[11] Mohammad Mahdi Johari, Camilla Carta, and Francois Fleuret. Eslam: Efficient dense slam system based on hybrid representation of signed distance fields. CVPR, 2023. 1, 2, 5, 6, 7, 8, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 488, + 287, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 488, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 48, + 488, + 287, + 531 + ], + "type": "text", + "content": "[12] Olaf Kähler, Victor Adrian Prisacariu, Julien P. C. Valentin, and David William Murray. Hierarchical voxel block hashing for efficient integration of depth images. RAL, 1:192-197, 2016. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 533, + 287, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 287, + 566 + ], + "type": "text", + "content": "[13] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. TOG, 42(4), 2023. 2, 3, 7, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 567, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 599 + ], + "type": "text", + "content": "[14] Leonid Keselman and Martial Hebert. Approximate differentiable rendering with algebraic surfaces. In ECCV, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 601, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 633 + ], + "type": "text", + "content": "[15] Leonid Keselman and Martial Hebert. Flexible techniques for differentiable rendering with 3d gaussians. arXiv preprint arXiv:2308.14737, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 635, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 667 + ], + "type": "text", + "content": "[16] Georg S. W. Klein and David William Murray. Parallel tracking and mapping on a camera phone. ISMAR, pages 83-86, 2009. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[17] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. ArXiv, abs/2308.09713, 2023. 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[18] Robert Maier, Raphael Schaller, and Daniel Cremers. Efficient online surface correction for real-time large-scale 3d reconstruction. ArXiv, abs/1709.03763, 2017. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 107, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 149 + ], + "type": "text", + "content": "[19] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 151, + 545, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 151, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 307, + 151, + 545, + 182 + ], + "type": "text", + "content": "[20] Raul Mur-Artal and Juan D. Tardós. Orb-slam2: An opensource slam system for monocular, stereo, and rgb-d cameras. *TRO*, 33:1255–1262, 2016. 1, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 184, + 545, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 184, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 307, + 184, + 545, + 248 + ], + "type": "text", + "content": "[21] Richard A. Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J. Davison, Pushmeet Kohli, Jamie Shotton, Steve Hodges, and Andrew William Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. ISMAR, pages 127-136, 2011. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 250, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 250, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 307, + 250, + 545, + 282 + ], + "type": "text", + "content": "[22] Richard A. Newcombe, S. Lovegrove, and Andrew J. Davison. Dtam: Dense tracking and mapping in real-time. ICCV, pages 2320-2327, 2011. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 283, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 283, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 307, + 283, + 545, + 316 + ], + "type": "text", + "content": "[23] Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Marc Stamminger. Real-time 3d reconstruction at scale using voxel hashing. TOG, 32:1-11, 2013. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 317, + 545, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 317, + 545, + 347 + ], + "spans": [ + { + "bbox": [ + 307, + 317, + 545, + 347 + ], + "type": "text", + "content": "[24] Delin Qu, Chi Yan, Dong Wang, Jie Yin, Dan Xu, Bin Zhao, and Xuelong Li. Implicit event-rgbd neural slam. CVPR, 2024. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 350, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 350, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 350, + 545, + 392 + ], + "type": "text", + "content": "[25] Gerhard Reitmayr, Tobias Langlotz, Daniel Wagner, Alessandro Mulloni, Gerhard Schall, Dieter Schmalstieg, and Qi Pan. Simultaneous localization and mapping for augmented reality. ISUVR, pages 5-8, 2010. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 394, + 545, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 394, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 307, + 394, + 545, + 448 + ], + "type": "text", + "content": "[26] Fabio Ruetz, Emili Hernández, Mark Pfeiffer, Helen Oleynikova, Mark Cox, Thomas Lowe, and Paulo Vinicius Koerich Borges. Ovpc mesh: 3d free-space representation for local ground vehicle navigation. ICRA, pages 8648-8654, 2018. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 449, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 545, + 480 + ], + "type": "text", + "content": "[27] Erik Sandström, Yue Li, Luc Van Gool, and Martin R. Oswald. Point-slam: Dense neural point cloud-based slam. In ICCV, 2023. 2, 5, 6, 7, 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 482, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 482, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 307, + 482, + 545, + 503 + ], + "type": "text", + "content": "[28] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 504, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 504, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 307, + 504, + 545, + 536 + ], + "type": "text", + "content": "[29] Thomas Schöps, Torsten Sattler, and Marc Pollefeys. Bad slam: Bundle adjusted direct rgb-d slam. CVPR, pages 134-144, 2019. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 537, + 545, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 537, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 307, + 537, + 545, + 569 + ], + "type": "text", + "content": "[30] Thomas Schops, Torsten Sattler, and Marc Pollefeys. BAD SLAM: Bundle adjusted direct RGB-D SLAM. In CVPR, 2019. 6" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 571, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 571, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 571, + 545, + 679 + ], + "type": "text", + "content": "[31] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Yuheng Ren, Shobhit Verma, Anton Clarkson, Ming Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Briales, Tyler Gillingham, Elias Mueggler, Luis Pesqueira, Manolis Savva, Dhruv Batra, Hauke Malte Strasdat, Renzo De Nardi, Michael Goesele, S. Lovegrove, and Richard A. Newcombe. The replica dataset: A digital replica of indoor spaces. ArXiv, abs/1906.05797, 2019. 5, 3, 4" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 681, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 545, + 712 + ], + "type": "text", + "content": "[32] J. Stuckler and Sven Behnke. Multi-resolution surfel maps for efficient dense 3d modeling and tracking. JVCIR, 25: 137-147, 2014. 1" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19603" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[33] Jürgen Sturm, Nikolas Engelhard, Felix Endres, Wolfram Burgard, and Daniel Cremers. A benchmark for the evaluation of RGB-D SLAM systems. In IROS. IEEE/RSJ, 2012. 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 286, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 286, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 286, + 150 + ], + "type": "text", + "content": "[34] Edgar Sucar, Kentaro Wada, and Andrew J. Davison. Nodeslam: Neural object descriptors for multi-view shape reconstruction. 3DV, pages 949–958, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 286, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 286, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 286, + 184 + ], + "type": "text", + "content": "[35] Edgar Sucar, Shikun Liu, Joseph Ortiz, and Andrew J. Davison. imap: Implicit mapping and positioning in real-time. ICCV, 2021. 1, 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 286, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 286, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 286, + 228 + ], + "type": "text", + "content": "[36] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. ArXiv, abs/2309.16653, 2023. 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 230, + 286, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 286, + 252 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 286, + 252 + ], + "type": "text", + "content": "[37] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. In NIPS, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 254, + 286, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 286, + 286 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 286, + 286 + ], + "type": "text", + "content": "[38] Andreas Langeland Teigen, Yeonsoo Park, Annette Stahl, and Rudolf Mester. Rgb-d mapping and tracking in a plenoxel radiance field. ArXiv, abs/2307.03404, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 287, + 286, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 287, + 286, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 287, + 286, + 330 + ], + "type": "text", + "content": "[39] Charalambos Theodorou, Vladan Velisavljevic, Vladimir Dyo, and Fredi Nonyelu. Visual slam algorithms and their application for ar, mapping, localization and wayfinding. Array, 15:100222, 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 331, + 286, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 286, + 376 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 286, + 376 + ], + "type": "text", + "content": "[40] Angtian Wang, Peng Wang, Jian Sun, Adam Kortylewski, and Alan Yuille. Voge: a differentiable volume renderer using gaussian ellipsoids for analysis-by-synthesis. arXiv preprint arXiv:2205.15401, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 376, + 286, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 286, + 409 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 286, + 409 + ], + "type": "text", + "content": "[41] Hengyi Wang, Jingwen Wang, and Lourdes de Agapito. Coslam: Joint coordinate and sparse parametric encodings for neural real-time slam. CVPR, 2023. 1, 5, 6, 8, 2, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 411, + 286, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 286, + 441 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 286, + 441 + ], + "type": "text", + "content": "[42] Kaixuan Wang, Fei Gao, and Shaojie Shen. Real-time scalable dense surfel mapping. ICRA, pages 6919-6925, 2019. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 445, + 286, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 286, + 477 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 286, + 477 + ], + "type": "text", + "content": "[43] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. TIP, 13(4):600-612, 2004. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 478, + 286, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 478, + 286, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 478, + 286, + 520 + ], + "type": "text", + "content": "[44] Thomas Whelan, Michael Kaess, Maurice F. Fallon, Hordur Johannsson, John J. Leonard, and John B. McDonald. Kintinous: Spatially extended kinectfusion. In AAAI, 2012. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 522, + 286, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 286, + 566 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 286, + 566 + ], + "type": "text", + "content": "[45] Thomas Whelan, John McDonald, Michael Kaess, Maurice Fallon, Hordur Johannsson, and John J. Leonard. Kintinuous: Spatially extended kinectfusion. In RSS Workshop on RGB-D, 2012. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 567, + 286, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 286, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 286, + 601 + ], + "type": "text", + "content": "[46] Thomas Whelan, Stefan Leutenegger, Renato Salas-Moreno, Ben Glocker, and Andrew Davison. Elasticfusion: Dense slam without a pose graph. In RSS, 2015. 1, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 601, + 286, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 286, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 286, + 644 + ], + "type": "text", + "content": "[47] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. ArXiv, abs/2310.08528, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 647, + 286, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 286, + 689 + ], + "type": "text", + "content": "[48] Xingrui Yang, Hai Li, Hongjia Zhai, Yuhang Ming, Yuqian Liu, and Guofeng Zhang. Vox-fusion: Dense tracking and mapping with voxel-based neural implicit representation. IS-MAR, pages 499-507, 2022. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "type": "text", + "content": "[49] Ziyi Yang, Xinyu Gao, Wenming Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 421 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 327, + 73, + 544, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 544, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 544, + 95 + ], + "type": "text", + "content": "for high-fidelity monocular dynamic scene reconstruction. ArXiv, abs/2309.13101, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "type": "text", + "content": "[50] Zeyu Yang, Hongye Yang, Zijie Pan, Xiatian Zhu, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. ArXiv, abs/2310.10642, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 141, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 545, + 184 + ], + "type": "text", + "content": "[51] Taoran Yi, Jiemin Fang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussiandreamer: Fast generation from text to 3d gaussian splattering with point cloud priors. ArXiv, abs/2310.08529, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 186, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 545, + 228 + ], + "type": "text", + "content": "[52] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 230, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 230, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 230, + 545, + 262 + ], + "type": "text", + "content": "[53] Youmin Zhang, Fabio Tosi, Stefano Mattoccia, and Matteo Poggi. Go-slam: Global optimization for consistent 3d instant reconstruction. In ICCV, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 264, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 308 + ], + "type": "text", + "content": "[54] Shuaifeng Zhi, Michael Bloesch, Stefan Leutenegger, and Andrew J. Davison. Scenocode: Monocular dense semantic reconstruction using learned encoded scene representations. CVPR, pages 11768-11777, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 309, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 309, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 308, + 309, + 545, + 352 + ], + "type": "text", + "content": "[55] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R. Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. CVPR, 2021. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 354, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 354, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 308, + 354, + 545, + 387 + ], + "type": "text", + "content": "[56] Wojciech Zielonka, Timur M. Bagautdinov, Shunsuke Saito, Michael Zollhofer, Justus Thies, and Javier Romero. Drivable 3d gaussian avatars. 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 388, + 545, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 545, + 421 + ], + "type": "text", + "content": "[57] Matthias Zwicker, Hanspeter Pfister, Jeroen van Baar, and Markus H. Gross. Ewa volume splatting. Proceedings Visualization, 2001. VIS '01., pages 29-538, 2001. 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "19604" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/GSNeRF_ Generalizable Semantic Neural Radiance Fields with Enhanced 3D Scene Understanding/e670ccdd-f8c3-4912-9d6e-b86fa014dde2_content_list.json b/2024/GSNeRF_ Generalizable Semantic Neural Radiance Fields with Enhanced 3D Scene Understanding/e670ccdd-f8c3-4912-9d6e-b86fa014dde2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a71504c68396dd4ff82eb2aa950c846b4218b79e --- /dev/null +++ b/2024/GSNeRF_ Generalizable Semantic Neural Radiance Fields with Enhanced 3D Scene Understanding/e670ccdd-f8c3-4912-9d6e-b86fa014dde2_content_list.json @@ -0,0 +1,1982 @@ +[ + { + "type": "text", + "text": "GSNeRF: Generalizable Semantic Neural Radiance Fields with Enhanced 3D Scene Understanding", + "text_level": 1, + "bbox": [ + 192, + 130, + 779, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zi-Ting Chou\\*1, Sheng-Yu Huang\\*1, I-Jieh Liu1, Yu-Chiang Frank Wang\\*2, Graduate Institute of Communication Engineering, National Taiwan University NVIDIA, Taiwan {r11942101, f08942095, r11942087}@ntu.edu.tw frankwang@nvidia.com", + "bbox": [ + 147, + 202, + 823, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Utilizing multi-view inputs to synthesize novel-view images, Neural Radiance Fields (NeRF) have emerged as a popular research topic in 3D vision. In this work, we introduce a Generalizable Semantic Neural Radiance Fields (GSNeRF), which uniquely takes image semantics into the synthesis process so that both novel view image and the associated semantic maps can be produced for unseen scenes. Our GSNeRF is composed of two stages: Semantic Geo-Reasoning and Depth-Guided Visual rendering. The former is able to observe multi-view image inputs to extract semantic and geometry features from a scene. Guided by the resulting image geometry information, the latter performs both image and semantic rendering with improved performances. Our experiments not only confirm that GSNeRF performs favorably against prior works on both novel-view image and semantic segmentation synthesis but the effectiveness of our sampling strategy for visual rendering is further verified.", + "bbox": [ + 75, + 321, + 472, + 595 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 625, + 209, + 641 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D scene understanding plays a pivotal role in many vision-related tasks, including 3D reconstruction and 3D reasoning. The former domain focuses on low-level understanding, including developing efficient models that interpret and reconstruct observations from various data sources, such as RGB, RGBD, and physical sensors. In contrast, the latter emphasizes high-level understanding, like encoding, segmentation, and common sense knowledge learning. For real-world applications like robotic navigation or augmented reality, both these facets - reconstruction and reasoning - are crucial to enhance interactions with the real world [10]. Among the challenges in the reconstruction domain, novel view synthesis has consistently been a challenging objective. Being able to generate a novel view image based on a few existing views shows that a model com", + "bbox": [ + 75, + 650, + 472, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "prehends the scene's geometry and possesses a foundational understanding akin to reconstruction.", + "bbox": [ + 500, + 292, + 890, + 321 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The Neural Radiance Field (NeRF) [1, 4, 8, 19-22, 28, 35] has recently risen as an exciting research area, providing a novel way to tackle the task of novel view synthesis. By encoding the density and emitted radiance at each spatial location, NeRF is able to compress a scene into a learnable model given several images and the corresponding camera poses of the scene. By incorporating a volumetric rendering skill, images of unseen camera views can be generated with convincing quality. However, NeRF primarily focuses on reconstructing the color information of novel views and understanding the associated high-level semantic information (e.g., semantic segmentation or object detection) remains a significant challenge.", + "bbox": [ + 496, + 323, + 892, + 520 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent works [9, 15, 25, 30, 31, 37] aim to amplify the high-level scene understanding ability of NeRF by integrating NeRF with semantic segmentation. By sharing the information between the semantic object class and their corresponding appearances, these two tasks are capable of benefiting from each other's insights [17]. Nevertheless, these semantic understanding methods stick closely to the original NeRF paradigm, focusing on exploiting the model to represent a specific scene. As mentioned in [17], such a strategy requires additional annotations of semantic segmentation maps when applied to a new scene, limiting real-world applicability and generalizability.", + "bbox": [ + 496, + 521, + 893, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To tackle this problem, generalizable NeRFs [3, 14, 18, 27, 32, 33, 36] have emerged as a promising solution. They adopt an on-the-fly approach for building a neural radiance field conditioned on extracted features from input images of different scenes rather than encoding the scene representation directly into the model. For instance, PixelNeRF [36] introduces the idea of conditioning a NeRF model with multi-view images across multiple scenes. This innovation boosts NeRF's ability to generalize to unseen scenes and avoids retraining for each individual scene. Yet, it leaves uncertainties in its application to semantic understanding.", + "bbox": [ + 496, + 702, + 893, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As a pioneer in integrating a generalizable NeRF with semantic segmentation capabilities, Semantic-Ray (S-", + "bbox": [ + 500, + 869, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal Contribution", + "bbox": [ + 94, + 886, + 207, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20806", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/ff3bb270e289c0ea0cd6fc84343f028d1326465d1125dad7fdbe9dabf0c3a872.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Methodgeneralized NeRFgeneralized segmentationsampling strategyefficient sampling
Semantic NeRF [37]--hierarchical✓
Panoptic Neural Fields [15]--uniform-
Panoptic Lifting [25]--uniform-
DM-NeRF [31]--hierarchical✓
NeSF [30]-✓hierarchical✓
Semantic Ray [17]✓✓hierarchical✓
Ours✓✓depth-guided✓
", + "bbox": [ + 158, + 88, + 812, + 214 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. Comparisons of semantic-aware NeRFs (i.e., NeRFs, which jointly perform novel view synthesis and semantic segmentation). Note that NeSF only performs generalizable segmentation, yet its view synthesis remains restricted to the scene-specific training scheme. And, Semantic Ray requires ground truth depth as additional inputs for both training and inference.", + "bbox": [ + 75, + 224, + 893, + 267 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Ray) [17] made notable advancements by introducing a generalizable semantic field concept. By incorporating a cross-reprojection attention module, S-Ray aggregates information of multi-view images from all points sampled along a ray and predicts RGB values and semantic labels. Although this approach appears to capture the necessary contextual information for semantic segmentation, aggregating details of all points introduces noisy features that could interfere with segmentation tasks, as segmentation is mainly concerned with the first object class that the ray encounters. Additionally, their hierarchical sampling approach further increases additional computation and model complexity in view synthesis. Thus, a properly designed rendering strategy for a generalizable semantic field would be desirable.", + "bbox": [ + 75, + 290, + 472, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose Generalizable Semantic Neural Raisance Fields (GSNeRF) to jointly tackle the problems of generalized novel view synthesis and semantic segmentation. Given a set of source images of a scene with corresponding camera views and a novel target camera view, our GSNeRF is able to derive visual features and depth map predictions of each source view, which are used to predict the depth map of the target view. With the depth map of the target view and properly derived visual features of source views, the novel view RGB image and semantic segmentation can be derived accordingly. Our GSNeRF consists of two key learning stages: Semantic Geo-Reasoning and Depth-Guided Visual Rendering. The former is to derive visual features and aggregate the predicted depth information of each source view to estimate the depth of the novel view, while the latter renders the RGB image and semantic segmentation map of the target view. We further design two distinct sampling strategies to minimize noisy features and enhance rendering efficiency. Through conducting extensive experiments on both real-world and synthetic datasets, we show that our method outperforms current state-of-the-art generalizable NeRF methods in novel view synthesis and semantic segmentation.", + "bbox": [ + 75, + 503, + 472, + 851 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The key contributions of our work are as follows:", + "bbox": [ + 94, + 853, + 423, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We propose GSNeRF for jointly rendering novel view images and producing the associated semantic segmentation", + "bbox": [ + 76, + 869, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "mask on unseen scenes.", + "bbox": [ + 511, + 291, + 673, + 303 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- The proposed Semantic Geo-Reasoning stage learns color, geometry, and semantic information of the input scene, introducing generalization ability of our GSNeRF.", + "- Based on the inferred geometry information, the introduced Depth-Guided Visual Rendering stage customizes two different sampling strategies according to the predicted target view depth map, so that image and semantic map rendering can be performed simultaneously." + ], + "bbox": [ + 500, + 305, + 890, + 426 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 440, + 640, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Neural Radiance Fields", + "text_level": 1, + "bbox": [ + 500, + 465, + 717, + 479 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The Neural Radiance Field [20] has emerged as a widely embraced implicit representation by encoding a 3D scene within a neural network. Nonetheless, as mentioned in [7], the original NeRF method demands extensive training time, ranging from hours to days, and relies on dozens of multiview images as input. As a result, a large number of subsequent works [8, 21, 28, 35] are put forward to confront these issues. To shorten the training process, methods such as Instant NGP [21] and DVGO [28] opt for a balance between speed and memory, using hash encoding and voxel encoding to minimize the training time to minutes. On the other hand, to address the demand for multiple input views, some methods [7, 14, 23] introduce depth as additional supervision, a notable example being DS-NeRF [7]. DS-NeRF applies extra depth supervision to ensure that the NeRF accurately encodes a scene's geometry, thus enabling the construction of neural radiance fields with fewer images. This method shows how incorporating depth information can enhance the quality of novel view synthesis while reducing the number of required source images.", + "bbox": [ + 496, + 488, + 890, + 790 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Generalizable Novel View Synthesis", + "text_level": 1, + "bbox": [ + 500, + 801, + 808, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Since NeRF directly encodes information of a scene into a neural network, it is only applicable to the training scene of interest. In response to this limitation, some methods [3, 14, 16, 18, 27, 29, 32, 36] propose on-the-fly construction of NeRF, allowing trained NeRFs applied for syn", + "bbox": [ + 496, + 824, + 890, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20807", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "thesizing novel views in unseen scenes. The concept is to extract features from each source view image and condition the NeRF renderer on these scene-specific features, typically through [36] 2D CNN or [3, 14] cost volume techniques that aggregate features across different images. While these methods realize generalizable novel view synthesis, extensions of such NeRFs to high-level scene understanding tasks remain an open challenge.", + "bbox": [ + 75, + 90, + 472, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Multi-tasking NeRF", + "text_level": 1, + "bbox": [ + 76, + 220, + 267, + 234 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, several methods [9, 15, 17, 25, 30, 31, 37] seek to enhance NeRF with higher-level understanding abilities. Semantic NeRF [37] stands as the first method to integrate semantic segmentation with NeRF by designing two projection heads for predicting both semantics and color simultaneously. Although it achieves satisfactory performance and robustness on semantic segmentation tasks, it requires retraining for each scene. This implies a need for not only RGB images but also corresponding semantic segmentation maps to represent each new scene, hindering their generalizability. NeSF [30] has made an attempt to improve this aspect by proposing a half-generalizable method that includes a generalizable semantic segmentation module. That is, after training an original NeRF of each scene, they extract a 3D voxel-like grid density from the neural radiance field and train a generalizable 3D UNet to do a semantic segmentation task. However, this learning scheme still requires additional time to retrain NeRFs for novel scenes.", + "bbox": [ + 75, + 243, + 472, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "S-Ray [17] advances further by proposing a generalized semantic neural field. Following the backbone of Neuray [18], they propose a cross-reprojection attention module. Given a ray emitted from the target viewpoint, they uniformly sample points along the ray, and hierarchically sample more points on the second forward pass according to the density predicted from the first pass. With the cross-reprojection module, they aggregate the contextual information on all sampled points. Despite achieving good results on both real-world and synthetic datasets, the sampling strategy is time-consuming. Also, aggregating all information along the entire ray can introduce noisy features that may hamper segmentation tasks, as segmentation primarily focuses on the first object class encountered by the ray.", + "bbox": [ + 75, + 515, + 468, + 726 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Contrarily, our proposed GSNeRF aims to predict the depth map of the target view and utilize it to sample points along the ray more efficiently. This unique strategy enables us to sample points focusing on the first encountered object and filter out unrelated noisy points for semantic segmentation prediction. In Table 1, we compare the characteristics of recent semantic-aware NeRFs with GSNeRF.", + "bbox": [ + 75, + 727, + 470, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Brief Review of Generalizable NeRFs", + "text_level": 1, + "bbox": [ + 76, + 844, + 415, + 861 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For the sake of completeness, we provide a brief review of generalizable NeRFs [3, 14, 18, 29, 36]. Given $K$ multi", + "bbox": [ + 75, + 869, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "view images $I_{1:K} = \\{I_1, I_2, \\dots, I_K\\}$ with corresponding camera poses $\\xi_{1:K} = \\{\\xi_1, \\xi_2, \\dots, \\xi_K\\}$ of a scene, and a target camera pose $\\xi_T$ , the goal is to synthesize the novel view image $I_T$ . Rays are sampled across the scene from the target camera pose to render the novel view image. Along each ray, $r$ , $N_c$ points (coarse points) are sampled within the given near bound $b_l$ and far bound $b_u$ . Existing methods partition $[b_l, b_u]$ into $N_c$ bins uniformly and draw one sample within each bin as defined by:", + "bbox": [ + 496, + 90, + 890, + 227 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nt _ {n} ^ {r} \\sim \\mathcal {U} \\left[ b _ {l} + \\frac {n - 1}{N _ {c}} \\left(b _ {u} - b _ {l}\\right), b _ {l} + \\frac {n}{N _ {c}} \\left(b _ {u} - b _ {l}\\right) \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 234, + 890, + 266 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $t_n^r$ is the $n$ -th sample along $r$ . With the corresponding camera position $\\sigma^r$ and viewing directions $\\phi^r$ , the $n$ -th sampled point $x_n^r$ can be derived as $x_n^r = o^r + t_n^r\\phi^r$ , with $x_n^r \\in \\mathbb{R}^3$ . Several methods [3, 18, 29, 36] also use hierarchical sampling, conducting a second forward pass to sample $N_f$ points (fine points) where the first $N_c$ uniformly sampled points show higher predicted density.", + "bbox": [ + 496, + 272, + 890, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With the above information, a generalized NeRF can be represented by a learnable continuous function $A_{\\theta}$ , where $\\theta$ denotes the learnable parameters. $A_{\\theta}$ computes a volumetric density $\\sigma_n^r \\in \\mathbb{R}^+$ and a view-dependent color representation $\\mathbf{c}_n^r \\in \\mathbb{R}^3$ for each point $x_n^r$ . That is,", + "bbox": [ + 496, + 378, + 890, + 455 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma_ {n} ^ {r}, \\mathbf {c} _ {n} ^ {r} = A _ {\\theta} \\left(x _ {n} ^ {r}, \\phi^ {r}, f _ {1: K} \\left(w \\left(x _ {n} ^ {r}\\right)\\right)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 462, + 890, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $f_{1:K}$ are features of $I_{1:K}$ and $w(.)$ is the function that reprojects the 3D position $x_{n}^{r}$ onto the image plane with known camera poses.", + "bbox": [ + 496, + 488, + 890, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To this end, the color value $\\mathbf{C}(r)$ of the pixel corresponding to the ray $r$ is formulated as:", + "bbox": [ + 496, + 534, + 890, + 564 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {C} (r) = \\sum_ {n = 1} ^ {N _ {c}} T _ {n} ^ {r} \\alpha_ {n} ^ {r} \\mathbf {c} _ {n} ^ {r}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 571, + 890, + 613 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nT _ {n} ^ {r} = \\exp \\left(- \\sum_ {i = 1} ^ {n - 1} \\sigma_ {i} ^ {r} \\delta_ {i} ^ {r}\\right), \\quad \\alpha_ {n} ^ {r} = 1 - \\exp \\left(- \\sigma_ {n} ^ {r} \\delta_ {n} ^ {r}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 633, + 890, + 674 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $T_{n}^{r}$ is the accumulated transmittance and $\\delta_n^r = ||x_{n + 1}^r -x_n^r ||_2$ represents the distance between every two adjacent points along $r$ .", + "bbox": [ + 496, + 681, + 890, + 727 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As is the case with other NeRF methods, a generalized NeRF also optimizes the model based on the rendering loss between the rendered image and the GT image:", + "bbox": [ + 496, + 728, + 890, + 773 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {i m a g e}} = \\sum_ {r \\in R} \\| \\mathbf {C} (r) - \\hat {\\mathbf {C}} (r) \\| _ {2} ^ {2}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 781, + 890, + 813 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{C}}(r)$ represents the GT color of the renderer image corresponding to a ray $r$ .", + "bbox": [ + 496, + 824, + 890, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our method is built on top of generalized NeRF methods but with a more efficient sampling strategy that can deal with both semantic segmentation and image rendering.", + "bbox": [ + 496, + 854, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "footer", + "text": "20808", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/86ad533e35e25e87828d7c3f4f9d4c75fe009278798a4a6dfca7ae4bb27bfa51.jpg", + "image_caption": [ + "Figure 1. Overview of GSNeRF: including Semantic Geo-Reasoning and Depth-Guided Visual Rendering. Given K multi-view image $I_{1:K}$ of a scene, the Semantic Geo-Reasoner predicts the depth map $D_{1:K}$ for each source image, which is aggregated to estimate the target view depth map $D_T$ . With $D_T$ as key geometric guidance, we design Depth-Guided Visual Rendering to render target view image $I_T$ and semantic segmentation $S_T$ , by Volume Renderer $R_\\theta$ and Semantic Renderer $P_\\theta$ , respectively." + ], + "image_footnote": [], + "bbox": [ + 125, + 95, + 831, + 300 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/69d8a02cbd00b997aafe84f1ab8d005e8f358daef319f3d1ed35fd2f6bd8bd34.jpg", + "image_caption": [ + "Figure 2. Semantic Geo-Reasoning. The Semantic Geo-Reasoner contains a shared Encoder $E_{\\theta}$ , a shared Decoder $D_{\\theta}$ and a cost-volume aggregator $C_{\\theta}$ , producing image and volume features $(f_{1:K}^{I}$ and $f_{1:K}^{V})$ , with the associated depth maps $D_{1:K}$ . The depth map $D_{T}$ of the target view is estimated from $D_{1:K}$ ." + ], + "image_footnote": [], + "bbox": [ + 78, + 393, + 472, + 565 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 76, + 669, + 169, + 684 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Problem Formulation and Model Overview", + "text_level": 1, + "bbox": [ + 76, + 694, + 444, + 709 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first define the setting and notations used in this paper. Given a scene with $K$ multi-view images $I_{1:K}$ and their corresponding camera poses $\\xi_{1:K}$ along with a target camera pose as input, we aim to achieve novel-view synthesis and semantic segmentation simultaneously. During training, we synthesize both the novel view image $I_T$ and the semantic segmentation map $S_T$ from $\\xi_T$ , given ground truth (GT) image $\\hat{I}_T$ and semantic segmentation map $\\hat{S}_T$ of the target view $\\xi_T$ . Once the model completes training on a given set of training scenes, it is able to directly generalize to unseen scenes without finetuning.", + "bbox": [ + 75, + 718, + 468, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To achieve this goal, we propose Generalizable Seman", + "bbox": [ + 96, + 885, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tic Neural Radiance Fields (GSNeRF), as illustrated in Figure 1. GSNeRF consists of two key learning stages: Semantic Geo-Reasoning and Depth-Guided Visual Rendering. For Semantic Geo-Reasoning, we employ a Semantic Geo-Reasoner $G_{\\theta}$ on each input source image $I_{k} \\in \\mathbb{R}^{H \\times W \\times 3}$ to create 2D feature $f_{k}^{I} \\in \\mathbb{R}^{H \\times W \\times d}$ , semantic feature $f_{k}^{S} \\in \\mathbb{R}^{H \\times W \\times d}$ , 3D volume feature $f_{k}^{V} \\in \\mathbb{R}^{H \\times W \\times L \\times d}$ , and depth predictions $D_{k} \\in \\mathbb{R}^{H \\times W}$ , where $H$ and $W$ represent the height and width of the image, $L$ is the depth sample number defined in [34] to determine the volume size for the 3D feature, and $d$ is the feature dimension. The source view depth maps are integrated to estimate the depth for the target view $D_{T}$ . As for Depth-Guided Visual Rendering, we conduct a unique sampling strategy based on $D_{T}$ to minimize noisy features and enhance rendering efficiency. The sampled points, along with the previously extracted feature, are respectively input into the volume renderer $R_{\\theta}$ and semantic renderer $P_{\\theta}$ , synthesizing the target view's image and semantic segmentation map. In the following subsection, we will explain the Semantic Geo-Reasoning and Depth-guided Visual Rendering process within our GSNeRF framework.", + "bbox": [ + 496, + 396, + 893, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Generalizable Semantic NeRF", + "text_level": 1, + "bbox": [ + 498, + 743, + 767, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2.1 Semantic Geo-Reasoning", + "text_level": 1, + "bbox": [ + 500, + 768, + 728, + 782 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given $K$ multi-view source images $I_{1:K}$ of a scene, our Semantic Geo-Reasoner $G_{\\theta}$ in Figure 1 is designed to achieve two goals. First, geometric clues are extracted from $G_{\\theta}$ , comprising 3D Volume features $f_{1:K}^{V}$ and depth maps $D_{1:K}$ for each source image. In addition, 2D image features $f_{1:K}^{I}$ and semantic features $f_{1:K}^{S}$ for each source image can be jointly produced. Secondly, our $G_{\\theta}$ is learned to predict the", + "bbox": [ + 496, + 794, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20809", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "target view depth map $D_{T}$ , which is realized by deterministic aggregation and reprojection of all source view depth predictions $D_{1:K}$ . We now detail this learning stage.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As depicted in Figure 2, our Semantic Geo-Reasoner consists of a shared 2D CNN encoder $E_{\\theta}$ that extracts $f_{1:K}^{I}$ from every image $I_{1:K}$ . To construct geometry-aware features, a cost volume aggregator $C_{\\theta}$ [6, 12, 13, 34] is applied to gather 2D features $f_{1:K}^{I}$ across different images. As a result, $C_{\\theta}$ produces 3D feature $f_{1:K}^{V}$ , which are ultimately used to predict the depth map $D_{1:K}$ for every source view image. To obtain semantic-aware features for semantic prediction, a shared semantic feature decoder $D_{\\theta}$ is additionally applied to derive 2D semantic features $f_{1:K}^{S}$ from $f_{1:K}^{I}$ .", + "bbox": [ + 75, + 137, + 467, + 289 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since $D_{1:K}$ is a vital intermediate element that assists in estimating $D_T$ , which allows the subsequent depth-guided rendering process, we strive to ensure the precise prediction of $D_{1:K}$ , whether or not the ground truth depth images are accessible. To achieve this, we integrate ideas from DSNeRF [7] and GeoNeRF [14] to supervise our $G_{\\theta}$ with either a depth-supervised loss or a self-supervised depth loss. If the ground-truth depth maps of $D_{1:K}$ are available, our depth-supervised loss $\\mathcal{L}_D$ is given as:", + "bbox": [ + 75, + 289, + 467, + 424 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {D} = \\frac {1}{K} \\left(\\sum_ {k = 1} ^ {K} \\| D _ {k} - \\hat {D} _ {k} \\| _ {s 1}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 439, + 468, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $\\hat{D}_k$ represents the GT depth map of view $k$ , $\\| \\cdot \\|_{s1}$ represents smooth L1 loss [11]. If $\\hat{D}_k$ is unavailable, we use a self-supervised depth loss $\\mathcal{L}_{ssl}$ to regularize our depth predictions by considering cross-view depth consistency between all source views. Following the self-supervised depth estimation approach of RC-MVSNet [2], $\\mathcal{L}_{ssl}$ is defined as:", + "bbox": [ + 75, + 489, + 467, + 583 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s s l} = \\lambda_ {1} \\mathcal {L} _ {R C} + \\lambda_ {2} \\mathcal {L} _ {S S I M} + \\lambda_ {3} \\mathcal {L} _ {S m o o t h}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 599, + 468, + 616 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{RC}$ , $\\mathcal{L}_{SSIM}$ , and $\\mathcal{L}_{Smooth}$ represent reconstruction, structure similarity, and depth smoothness losses [2]. The hyper-parameters $\\lambda_1$ , $\\lambda_2$ , and $\\lambda_3$ are also selected following [2] (see supplementary material for details).", + "bbox": [ + 75, + 625, + 467, + 685 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To this end, the target view depth map $D_{T}$ is calculated using $D_{1:K}$ through a deterministic algorithm. In simple terms, we directly project the pixels of each depth map into 3D space and then reproject them back to the target view. More details about the self-supervised depth loss and corresponding pseudo-code of target view depth estimation are available in the supplementary material.", + "bbox": [ + 75, + 686, + 467, + 791 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.2 Depth-Guided Visual Rendering", + "text_level": 1, + "bbox": [ + 76, + 814, + 356, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We now discuss how we perform volume and semantic rendering with the guidance of the predicted $D_{T}$ . That is, given a target view $\\xi_{T}$ , a ray is sampled for each pixel on the image plane to render the corresponding RGB value for $I_{T}$", + "bbox": [ + 75, + 839, + 467, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and the semantic segmentation output for $S_{T}$ via the Volume Renderer $R_{\\theta}$ and the Semantic Renderer $P_{\\theta}$ .", + "bbox": [ + 498, + 90, + 888, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Volume Rendering For volume rendering of $I_{T}$ , existing semantic-NeRF approaches [17, 30, 37] follow conventional NeRFs [20] by applying either uniform sampling or hierarchical sampling along each ray $r$ , as depicted in Table 1. However, both strategies require dense sampling to locate the surface of objects in the given scene, limiting the sampling efficiency. To promote the efficiency of our sampling process, we conduct a depth-guided sampling strategy by focusing on sampling points near the depth values of $D_{T}$ . More specifically, this is achieved by modifying the sampling strategy in Eq. 1 as follows:", + "bbox": [ + 496, + 138, + 890, + 306 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nt _ {n} ^ {r} \\sim \\mathcal {G} (z, v ^ {2}), \\quad v = \\frac {\\operatorname* {m i n} \\left(\\| z - b _ {u} \\| , \\| z - b _ {l} \\|\\right)}{3}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 316, + 890, + 347 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $t_n^r$ is the $n$ -th sample along the ray $r$ , and $\\mathcal{G}$ represents Gaussian (normal) distribution. Note that $z$ and $v$ represent the mean and standard deviation of the distribution, respectively, while the value of $z$ is determined as the predicted depth for the pixel $p$ that emits the ray $r$ (i.e., $z = D_T(p)$ ). By these sampling point values, we derive $N$ sampled points $x_{1:N}^r$ (i.e., $x_n^r = o^r + t_n^r\\phi^r$ ) closer to the surface of the first encountered object by $r$ , avoiding considering noisy points that are far from the object surface. Please see the supplementary material for further details.", + "bbox": [ + 496, + 356, + 890, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We feed all these points alongside the 2D feature $f_{1:K}^{I}$ and 3D volume feature $f_{1:K}^{V}$ as additional conditions into the volume renderer $R_{\\theta}$ to predict the novel view image of the scene. Specifically, we reproject the point $x_{n}^{r}$ onto each source view to get aggregated features from these condition features. Thus, we write the aggregated feature $f_{n,k} \\in \\mathbb{R}^{2d}$ of the $n$ -th sampled point $x_{n}^{r}$ from the $k$ -th view as:", + "bbox": [ + 496, + 507, + 890, + 613 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {n, k} = \\left[ f _ {k} ^ {I} \\left(w \\left(x _ {n} ^ {r}\\right)\\right); f _ {k} ^ {V} \\left(x _ {n} ^ {r}\\right) \\right], \\forall k \\in [ 1: K ] \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 625, + 890, + 642 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $w(\\cdot)$ is the reprojection function used in Eq. 2 and $[\\cdot; \\cdot]$ denotes concatenation along feature dimension. We further define the global feature $f_{n,0}$ as guidance across all $K$ source views as:", + "bbox": [ + 498, + 652, + 890, + 712 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {n, 0} = \\left[ \\operatorname {m e a n} \\left(\\left\\{f _ {n, k} \\right\\} _ {k = 1} ^ {K}\\right); \\operatorname {v a r} \\left(\\left\\{f _ {n, k} \\right\\} _ {k = 1} ^ {K}\\right) \\right]. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 724, + 890, + 743 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following GeoNeRF [14] methodology, the volume renderer $R_{\\theta}$ (with learnable parameters $\\theta$ ) predicts per point density and radiance based on the feature $f_{n} = \\{f_{n,k}\\}_{k=0}^{K} \\in \\mathbb{R}^{(K+1) \\times 2d}$ mentioned above:", + "bbox": [ + 498, + 753, + 890, + 815 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma_ {n} ^ {r}, \\mathbf {c} _ {n} ^ {r} = R _ {\\theta} \\left(x _ {n} ^ {r}, \\phi^ {r}, f _ {n}, M _ {n}\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 827, + 890, + 843 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $M_{n} = \\{M_{n,k}\\}_{k = 0}^{K}$ denotes the mask used to mask out unrelated features of a view. If the sampled point $x_{n}^{r}$ lies behind the predicted depth map $D_{k}$ or if the reprojected", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "footer", + "text": "20810", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/2448e76cd4d968d11a98d4ec4ce7424b5e4046e2f1fcfc146511197338dd10e5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Generalized methodGT DepthScanNet [5]Replica [26]
Train / TestmIoUacc. / class acc.PSNR↑SSIM↑LPIPS↓mIoUacc. / class acc.PSNR↑SSIM↑LPIPS↓
Neuray [18] + semhead✓ / ✓52.0967.81 / 61.9825.0183.0731.6344.3779.93 / 54.2526.2187.3730.93
GeoNeRF [14] + semhead✓ /53.7876.18 / 61.9032.5590.8812.6945.1281.67 / 52.3628.7088.9420.42
S-Ray [17]✓ / ✓55.5377.79 / 60.9225.1983.6630.9845.3080.48 / 53.7226.3888.1330.04
GSNeRF (Ours)✓ /58.3079.79 / 65.9331.3390.7312.5351.5283.41 / 61.2931.1692.4412.54
MVSNeRF [3] + semhead43.0666.90 / 53.6324.1480.3634.6330.2169.35 / 39.7523.6884.3728.08
GeoNeRF [14] + semhead45.1167.12 / 53.4430.7588.2716.4840.3574.63 / 49.1829.9291.1417.60
GNT [32] + semhead43.4962.06 / 51.8424.3982.3728.3638.1471.44 / 47.4624.5687.3120.97
Neuray [18] + semhead46.0966.39 / 53.7925.2484.3931.3340.9176.23 / 50.1527.8089.5523.68
S-Ray [17]47.6964.90 / 54.4725.1384.1830.4443.2777.63 / 52.8526.7788.5422.81
GSNeRF (Ours)52.2174.71 / 60.1431.4990.3913.8751.2383.06 / 61.1031.7192.8912.93
", + "bbox": [ + 78, + 88, + 893, + 292 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Quantitative results on ScanNet & Replica. Note that methods in the first four rows take GT depth as inputs or training supervision, while the methods in the last six rows do not observe GT depth during training/testing.", + "bbox": [ + 75, + 303, + 892, + 333 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "location is outside the source image, $M_{n,k}$ is set to 0. However, the mask for the global feature $M_{n,0}$ is always set to 1.", + "bbox": [ + 75, + 349, + 468, + 380 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To this end, we apply the same volume rendering as in Eq. 3, predicting the target view Image $I_{T}$ . We also apply $\\mathcal{L}_{image}$ as in Eq. 5 to supervise the rendered image quality.", + "bbox": [ + 76, + 380, + 468, + 428 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Semantic Rendering Semantic rendering of $S_{T}$ , on the other hand, is different from the above volume rendering process. Specifically, image rendering considers the color of both the surface and nearby objects because of possible semi-transparent or transparent objects. Conversely, semantic segmentation focuses only on the first object a ray touches. Therefore, for each ray $r$ , we sample only one surface point $x_{z}^{r}$ estimated by the target depth map for each ray $r$ to render its semantic value (i.e., $x_{z}^{r} = o^{r} + z\\phi^{r}$ , where $z$ is the same depth value as mentioned in Eq. 8). More specifically, with the aggregated semantic feature $f_{z} = \\{f_{k}^{S}(w(x_{z}^{r}))\\}_{k=1}^{K} \\in \\mathbb{R}^{K \\times d}$ , our semantic renderer $P_{\\theta}$ with learnable parameters $\\theta$ predicts the semantic logits $\\mathbf{S}(r)$ of ray $r$ as follows:", + "bbox": [ + 75, + 446, + 472, + 657 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S} (r) = P _ {\\theta} \\left(x _ {z} ^ {r}, f _ {z} ^ {\\prime}, M\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 670, + 468, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 76, + 699, + 122, + 712 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nf _ {z} ^ {\\prime} = f _ {z} + \\left\\{f _ {0} \\right\\} \\in \\mathbb {R} ^ {(K + 1) \\times d}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 712, + 468, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Note that we follow volume rendering in Eq. 10 to define a global semantic feature $f_{0}$ as guidance across all views:", + "bbox": [ + 75, + 738, + 468, + 767 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nf _ {0} = \\left[ \\operatorname {m e a n} \\left(f _ {z}\\right); \\operatorname {v a r} \\left(f _ {z}\\right) \\right]. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 781, + 468, + 797 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As for the mask $M = \\{M_k\\}_{k=0}^K$ in Eqn. 12, it is similar to the one used in Eq. 11, where $M_k$ is to mask out the unrelated features of a view if the sample point $x_z^r$ is behind the corresponding depth value of $D_k$ .", + "bbox": [ + 75, + 808, + 468, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Eventually, we utilize a semantic loss to supervise the novel view semantic prediction. Here, we use a multi-class", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "cross-entropy loss to reduce the semantic prediction error:", + "bbox": [ + 498, + 351, + 883, + 366 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s e m}} = \\sum_ {r \\in R} (\\mathbf {S} (r) \\log \\hat {\\mathbf {S}} (r)), \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 372, + 890, + 404 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\hat{S}(r)$ represent the ground truth segmentation score along the ray $r$ .", + "bbox": [ + 498, + 412, + 890, + 444 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Training and Inference", + "text_level": 1, + "bbox": [ + 498, + 450, + 714, + 465 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.1 Training", + "text_level": 1, + "bbox": [ + 500, + 473, + 616, + 488 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "At the training stage, the total loss is calculated by combining the losses that are mentioned above. The aggregated loss function for our model, when ground-truth depth is available, is given by:", + "bbox": [ + 498, + 497, + 890, + 556 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {i m a g e}} + \\mathcal {L} _ {D} + \\lambda \\mathcal {L} _ {\\text {s e m}}. \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 565, + 890, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "If the GT depth is not available, our model is trained by:", + "bbox": [ + 500, + 587, + 872, + 603 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {i m a g e}} + \\mathcal {L} _ {\\text {s s l}} + \\lambda \\mathcal {L} _ {\\text {s e m}}, \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 609, + 890, + 626 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where the coefficient $\\lambda$ is set to 0.5 for simplicity.", + "bbox": [ + 498, + 633, + 828, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In essence, this optimization process enables our model to estimate the depth map of the target camera view accurately. Utilizing this depth information, the model is capable of efficiently sampling points to render new view images and create corresponding semantic segmentation maps.", + "bbox": [ + 496, + 648, + 890, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.2 Inference", + "text_level": 1, + "bbox": [ + 500, + 739, + 620, + 753 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "At the inference stage, our GSNeRF is able to generalize to unseen scenes without the need for retraining from scratch. Due to the design of the conditioning renderer on scene-specific features $f^{I}, f^{V}$ , and $f^{S}$ , our model is able to build a semantic neural radiance field on the fly. Once we finish training our model on training scenes, we are able to directly infer novel view images and semantic segmentation maps on any scene given multi-view images and corresponding camera poses of that scene.", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20811", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/37ce1e2a58a14d0a872f185a5689cd39934055bf3de1d4e018c4d999c06e603f.jpg", + "image_caption": [ + "S-Ray" + ], + "image_footnote": [], + "bbox": [ + 81, + 102, + 214, + 178 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5c24ddee3802bce84eb65fba8426bedb3783ff72f10ee6f6023595d48cc8626f.jpg", + "image_caption": [ + "GSNeRF (Ours)" + ], + "image_footnote": [], + "bbox": [ + 217, + 101, + 349, + 178 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/adca324a7d265da1989de05806fac0a90dd19393c80b86694077ad0580ce1e09.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 351, + 101, + 483, + 178 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5fdff6ab6616ccf481068ac8634c2de626700c6198729f6cfc91e35d11f361cd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 180, + 212, + 257 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/98165cff2b6872e4831bf67a033862d09337d4470474012290b0916d03e7ff23.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 180, + 348, + 257 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5d058faf9ab067ba3527fa6e6d42c8f5f0bd00716bedd3f9cc26f90ae26c7c22.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 180, + 483, + 257 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/867343889916a3245a7372838ae59394e849320e15a1ead84509b9dffaf4d86b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 258, + 214, + 335 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/93473372901ed42cfe50c05344e8833c22ca4ce590aa3ff3642eda09af6cd48e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 258, + 349, + 335 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/03d2bafb9ac62658a427d26d38a641188c2668c3318a95ce07420fabe3b57165.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 258, + 483, + 335 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a4fdd178526e91fdfef30c9d05f9ec231e39bc78574bca49fd3c260489a061ef.jpg", + "image_caption": [ + "Figure 3. Qualitative evaluation. We compare the visual quality of the rendered novel view images (the first three columns) and semantic segmentation maps (the last three columns) with S-Ray [17]." + ], + "image_footnote": [], + "bbox": [ + 81, + 338, + 212, + 415 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a5812dd1dde51ef924929c95f0772304723a2d93679fd7d5d56dd147d74de4e3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 338, + 348, + 415 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/906b2e6bd9ef7a8ae4cef7a1e3fe13e7b8297ba28a420c5a5a5326ca1826cf5b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 338, + 483, + 415 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/91a37656f6cc47463d6d7b7b66cbba290e2e94ee27f5fd4e6b26e5508a97c70f.jpg", + "image_caption": [ + "S-Ray" + ], + "image_footnote": [], + "bbox": [ + 486, + 101, + 617, + 178 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f56b2e55d694112d99164465d9fafccc1026991198270e93b88f9bb776105f1f.jpg", + "image_caption": [ + "GSNeRF (Ours)" + ], + "image_footnote": [], + "bbox": [ + 620, + 101, + 753, + 178 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f30efb61ab6e7221665b7a6fd565cc74898baadea65cbab7b2ee3181edb5ef42.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 756, + 101, + 888, + 178 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/91a3d33427c6a72490ca7ecc468f9148c46a497a789fc3afe4ab09d4064bdc6d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 180, + 617, + 257 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/71cfb630bbd1a8159080a0afd2b0ccfe22a451447cf64b575a10839dcc66ac71.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 620, + 180, + 751, + 257 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a6def87bcfc5f5dfb8013bc30605d2a15f464414920e701037191d8c39f21578.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 756, + 180, + 887, + 257 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a21c11f1df7afe9baa5f807204463ae112d13f77c1ab5b571302d7758e67a94b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 258, + 617, + 335 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d60d961c28cc91b4060875a072ec7ee0e85ee35c15f661654f96e5a9541b387f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 620, + 258, + 751, + 335 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fa1609161ee7cd5030a1f826aa903550acc0a2b75fc3cbb6e1df635a3d59b640.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 756, + 258, + 887, + 335 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f39a52bde205f78178bee4a35a45c1cf592e62604fea8b2a136ac0dc588d7998.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 338, + 617, + 415 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9b9894c1e7716630c3baa2909ca11040d9fecec218db523fc44cb709646acb97.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 620, + 338, + 751, + 415 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/59c725c48e516ed597bbcf9a2f02c52d55bac3ca74ec9faae489c1d8bde03654.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 756, + 338, + 887, + 415 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 482, + 207, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Datasets", + "text_level": 1, + "bbox": [ + 76, + 508, + 178, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate the effectiveness of our proposed method, we conduct experiments on both real-world and synthetic datasets. For real-world data, we use ScanNet [5], a large-scale, indoor RGB-D video dataset with over 2.5 million views from 1513 distinct scenes, including semantic annotations and camera poses. Following the setting presented in S-Ray [17], we train our model on 60 scenes and test its generalization on 10 new, unseen scenes. For synthetic data, we utilize Replica [26], a 3D reconstruction-based indoor dataset comprising 18 high-quality scenes with dense geometry, HDR textures, and semantic labels. Following the Semantic NeRF [37] supplied rendered data, we train our model on 6 distinct scenes across 12 video sequences and test on 2 novel scenes across 4 video sequences. Please see the supplementary material for details.", + "bbox": [ + 75, + 534, + 468, + 760 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Results and analysis", + "text_level": 1, + "bbox": [ + 76, + 773, + 267, + 789 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2.1 Quantitative Results", + "text_level": 1, + "bbox": [ + 76, + 797, + 272, + 813 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2 compares our method against several baselines, in which S-Ray [17] serves as a primary baseline since it is the first to address semantic segmentation with NeRF in a generalized setting. We further conduct experiments on MVS-NeRF [3], GeoNeRF [14], GNT [32] and NeuRay [18] with", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the semantic head as described in [17]. It is worth noting that, even though S-Ray does not explicitly mention using a depth map of each source view image as input in their paper, their official implementation takes the depth maps $D_{1:K}$ as additional input along with $I_{1:K}$ during training and testing. Therefore, we follow their official implementation to reproduce S-Ray results.", + "bbox": [ + 496, + 484, + 890, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "From Table 2, it can be seen that our model generalizes well to unseen scenes. To show that our GSNeRF is also capable of producing impressive results without depth supervision, we conduct an additional experiment in which we only use the loss function in Eq. 17 to train our GSNeRF in Table 2. We also conduct experiments of S-Ray, Neuray with only RGB images as input, and GeoNeRF without GT depth as supervision for a fair comparison. From Table 2, we see that even without access to GT depth and with the target depth information regularized in a self-supervised manner, our method still outperforms all the other baseline approaches. This suggests that our model can be trained without requiring depth map supervision, confirming the effectiveness and practicality of our proposed approach. We also note that, although GeoNeRF + semhead (with depth supervision) slightly outperforms our approach regarding PSNR and SSIM for the rendered RGB images on ScanNet, our method excels in all semantic segmentation metrics by approximately $5\\%$ . More comprehensive discussions are presented in the supplementary materials.", + "bbox": [ + 496, + 598, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20812", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6d76ad9ec1e799fe0e7e3f60baa5f6c257308470bf4355842435e43db00d5e60.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ITST
ModelDepth-guidedrenderer PθDepth-guidedPSNRmIoU
A---28.4837.88
B✓--30.5641.72
C✓✓-30.8849.47
Ours✓✓✓31.4952.21
", + "bbox": [ + 78, + 88, + 470, + 204 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Ablation studies on GSNeRF. We verify the effectiveness of using depth-guided rendering for rendering $I_{T}$ and $S_{T}$ . Note that, without $P_{\\theta}$ , we simply take a pre-trained 2D segmentor on $I_{T}$ to produce $S_{T}$ .", + "bbox": [ + 75, + 213, + 470, + 270 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1e7d1138daa3c614d285f06e394d0251d581c1fc5e5008ed8e7ccff8f9e949c6.jpg", + "image_caption": [ + "Figure 4. Sampling efficiency on ScanNet. Compared to SOTAs, our GSNeRF is able to achieve significantly improved rendering performance especially when the number of sampling points is small. Even with increasing numbers of sampling points, GSNeRF still performs favorably against existing models." + ], + "image_footnote": [], + "bbox": [ + 78, + 277, + 472, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2.2 Qualitative Results", + "text_level": 1, + "bbox": [ + 76, + 563, + 264, + 578 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The qualitative results of our approach compared to S-Ray [17] are depicted in Fig. 3. This comparison illustrates that while the existing method can approximate the contours of objects, it is unable to capture the geometrical details, causing segmentation inconsistencies. In contrast, our GSNeRF accurately interprets scene geometry through our targeted view depth estimation design, allowing us to not only precisely segment objects but also render novel view images more realistically. This confirms that our depth-guided sampling technique effectively captures the correct geometry and that the sampled point features are appropriately aggregated from features of the Semantic Geo-Reasoner.", + "bbox": [ + 75, + 585, + 468, + 782 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2.3 Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 801, + 246, + 815 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To further analyze the effectiveness of our designed modules, we conduct ablation studies on ScanNet, as shown in Table 3. Baseline model A employs our Semantic Geo-Reasoner $G_{\\theta}$ and volume renderer $R_{\\theta}$ to create the target view image $I_{T}$ , using uniform sampling instead of depth", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "guided sampling $x_{1:N}$ . Also, a pre-trained 2D segmentor [24] is applied on the rendered image $I_T$ to produce $S_T$ without our semantic renderer $P_\\theta$ . Model B enhances visual quality by replacing uniform sampling with depth-guided sampling. We further replace the 2D segmentor with our semantic renderer $P_\\theta$ using uniform sampling for semantic rendering in model C, illustrating mutual benefits for both tasks. Our full model in the last row uses depth-guided sampling exclusively for surface points in the semantic renderer, achieving optimal results. This verifies the success of our proposed modules and sampling strategies.", + "bbox": [ + 496, + 90, + 890, + 258 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2.4 Sampling Efficiency", + "text_level": 1, + "bbox": [ + 500, + 275, + 692, + 291 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Another benefit of our depth-guided sampling is that the performance of our GSNeRF is less sensitive to the number of sampled points along each ray. In Fig. 4, we conduct experiments of reducing the number of sampling points $N$ , showing PSNR of the synthesized novel view images as evaluation. We find that by employing depth-guided sampling, our method accurately preserves visual details with less noise (compared to Neuray and S-Ray with 128 sampled points along each ray). It can be seen that GeoNeRF + semhead shows a catastrophic drop in PSNR with fewer points, while ours leveraging depth-guided sampling maintains satisfactory visual quality even with a significantly reduced number of sampled points (e.g., when sampling only 4 points along a ray, the PSNR remains near 28). This result further attests to the success of our design in estimating target depth and implementing the corresponding sampling strategy with better novel view synthesis performance.", + "bbox": [ + 496, + 299, + 890, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 570, + 617, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we proposed Generalizable Semantic Neural Raidance Fields (GSNeRF) to achieve generalized novel view synthesis and semantic segmentation. Our GSNeRF is trained to derive visual features and perform depth map prediction of each source view so that the depth map for a novel target view can be estimated. With such target-view depth information observed, the associated RGB image and semantic segmentation can be jointly produced via depth-guide rendering. In our experiments, we quantitatively and qualitatively confirm that our GSNeRF performs favorably against existing generalizable semantic-aware NeRF methods on both real-world and synthetic datasets.", + "bbox": [ + 496, + 594, + 890, + 776 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement This work is supported in part by the National Science and Technology Council via grant NSCT112-2634-F-002-007 and the NTU Center of Data Intelligence: Technologies, Applications, and Systems via grant NTU-113L900902. We also thank the National Center for High-performance Computing (NCHC) for providing computational and storage resources.", + "bbox": [ + 496, + 794, + 890, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20813", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 1", + "[2] Di Chang, Aljaž Božić, Tong Zhang, Qingsong Yan, Yingcong Chen, Sabine Susstrunk, and Matthias Nießner. Rcmvsnet: unsupervised multi-view stereo with neural rendering. In European Conference on Computer Vision, pages 665-680. Springer, 2022. 5", + "[3] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14124-14133, 2021. 1, 2, 3, 6, 7", + "[4] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 1", + "[5] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 6, 7", + "[6] Yuchao Dai, Zhidong Zhu, Zhibo Rao, and Bo Li. Mvs2: Deep unsupervised multi-view stereo with multi-view symmetry. In 2019 International Conference on 3D Vision (3DV), pages 1-8. IEEE, 2019. 5", + "[7] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12882-12891, 2022. 2, 5", + "[8] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 1, 2", + "[9] Xiao Fu, Shangzhan Zhang, Tianrun Chen, Yichong Lu, Lanyun Zhu, Xiaowei Zhou, Andreas Geiger, and Yiyi Liao. Panoptic nerf: 3d-to-2d label transfer for panoptic urban scene segmentation. In 2022 International Conference on 3D Vision (3DV), pages 1-11. IEEE, 2022. 1, 3", + "[10] Sourav Garg, Niko Sünderhauf, Feras Dayoub, Douglas Morrison, Akansel Cosgun, Gustavo Carneiro, Qi Wu, Tat-Jun Chin, Ian Reid, Stephen Gould, et al. Semantics for robotic mapping, perception and interaction: A survey. Foundations and Trends® in Robotics, 8(1-2):1-224, 2020. 1", + "[11] Ross Girshick. Fast r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 1440-1448, 2015. 5", + "[12] Xiaodong Gu, Zhiwen Fan, Siyu Zhu, Zuozhuo Dai, Feitong Tan, and Ping Tan. Cascade cost volume for high-resolution multi-view stereo and stereo matching. In Proceedings of" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF conference on computer vision and pattern recognition, pages 2495-2504, 2020. 5", + "[13] Baichuan Huang, Hongwei Yi, Can Huang, Yijia He, Jingbin Liu, and Xiao Liu. M3vsnet: Unsupervised multi-metric multi-view stereo network. In 2021 IEEE International Conference on Image Processing (ICIP), pages 3163-3167. IEEE, 2021. 5", + "[14] Mohammad Mahdi Johari, Yann Lepoittevin, and François Fleuret. Geonerf: Generalizing nerf with geometry priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18365-18375, 2022. 1, 2, 3, 5, 6, 7", + "[15] Abhijit Kundu, Kyle Genova, Xiaqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12871-12881, 2022. 1, 2, 3", + "[16] Haotong Lin, Sida Peng, Zhen Xu, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Efficient neural radiance fields for interactive free-viewpoint video. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 2", + "[17] Fangfu Liu, Chubin Zhang, Yu Zheng, and Yueqi Duan. Semantic ray: Learning a generalizable semantic field with cross-reprojection attention. In CVPR, pages 17386-17396, 2023. 1, 2, 3, 5, 6, 7, 8", + "[18] Yuan Liu, Sida Peng, Lingjie Liu, Qianqian Wang, Peng Wang, Christian Theobalt, Xiaowei Zhou, and Wenping Wang. Neural rays for occlusion-aware image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7824-7833, 2022. 1, 2, 3, 6, 7", + "[19] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7210-7219, 2021. 1", + "[20] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 5", + "[21] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 2", + "[22] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14335-14345, 2021. 1", + "[23] Barbara Roessle, Jonathan T Barron, Ben Mildenhall, Pratul P Srinivasan, and Matthias Nießner. Dense depth priors for neural radiance fields from sparse input views. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12892-12901, 2022. 2" + ], + "bbox": [ + 503, + 93, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20814", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 8", + "[25] Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulò, Norman Müller, Matthias Nießner, Angela Dai, and Peter Kontschieder. Panoptic lifting for 3d scene understanding with neural fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9043-9052, 2023. 1, 2, 3", + "[26] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 6, 7", + "[27] Mohammed Suhail, Carlos Esteves, Leonid Sigal, and Ameesh Makadia. Generalizable patch-based neural rendering. In European Conference on Computer Vision, pages 156-174. Springer, 2022. 1, 2", + "[28] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 1, 2", + "[29] Alex Trevithick and Bo Yang. Grf: Learning a general radiance field for 3d representation and rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15182-15192, 2021. 2, 3", + "[30] Suhani Vora, Noha Radwan, Klaus Greff, Henning Meyer, Kyle Genova, Mehdi SM Sajjadi, Etienne Pot, Andrea Tagliasacchi, and Daniel Duckworth. Nesf: Neural semantic fields for generalizable semantic segmentation of 3d scenes. arXiv preprint arXiv:2111.13260, 2021. 1, 2, 3, 5", + "[31] Bing Wang, Lu Chen, and Bo Yang. Dm-nerf: 3d scene geometry decomposition and manipulation from 2d images. arXiv preprint arXiv:2208.07227, 2022. 1, 2, 3", + "[32] Peihao Wang, Xuxi Chen, Tianlong Chen, Subhashini Venugopalan, Zhangyang Wang, et al. Is attention all nerf needs? arXiv preprint arXiv:2207.13298, 2022. 1, 2, 6, 7", + "[33] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690-4699, 2021. 1", + "[34] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vision (ECCV), pages 767-783, 2018. 4, 5", + "[35] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5752-5761, 2021. 1, 2", + "[36] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images." + ], + "bbox": [ + 78, + 90, + 470, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021. 1, 2, 3", + "[37] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J Davison. In-place scene labelling and understanding with implicit scene representation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15838-15847, 2021. 1, 2, 3, 5, 7" + ], + "bbox": [ + 501, + 92, + 890, + 204 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20815", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/GSNeRF_ Generalizable Semantic Neural Radiance Fields with Enhanced 3D Scene Understanding/e670ccdd-f8c3-4912-9d6e-b86fa014dde2_model.json b/2024/GSNeRF_ Generalizable Semantic Neural Radiance Fields with Enhanced 3D Scene Understanding/e670ccdd-f8c3-4912-9d6e-b86fa014dde2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..40b5aeedc656d86bef49edc61601b192db75f36a --- /dev/null +++ b/2024/GSNeRF_ Generalizable Semantic Neural Radiance Fields with Enhanced 3D Scene Understanding/e670ccdd-f8c3-4912-9d6e-b86fa014dde2_model.json @@ -0,0 +1,2387 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.193, + 0.131, + 0.78, + 0.177 + ], + "angle": 0, + "content": "GSNeRF: Generalizable Semantic Neural Radiance Fields with Enhanced 3D Scene Understanding" + }, + { + "type": "text", + "bbox": [ + 0.148, + 0.203, + 0.825, + 0.258 + ], + "angle": 0, + "content": "Zi-Ting Chou\\*1, Sheng-Yu Huang\\*1, I-Jieh Liu1, Yu-Chiang Frank Wang\\*2, Graduate Institute of Communication Engineering, National Taiwan University NVIDIA, Taiwan {r11942101, f08942095, r11942087}@ntu.edu.tw frankwang@nvidia.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.323, + 0.473, + 0.596 + ], + "angle": 0, + "content": "Utilizing multi-view inputs to synthesize novel-view images, Neural Radiance Fields (NeRF) have emerged as a popular research topic in 3D vision. In this work, we introduce a Generalizable Semantic Neural Radiance Fields (GSNeRF), which uniquely takes image semantics into the synthesis process so that both novel view image and the associated semantic maps can be produced for unseen scenes. Our GSNeRF is composed of two stages: Semantic Geo-Reasoning and Depth-Guided Visual rendering. The former is able to observe multi-view image inputs to extract semantic and geometry features from a scene. Guided by the resulting image geometry information, the latter performs both image and semantic rendering with improved performances. Our experiments not only confirm that GSNeRF performs favorably against prior works on both novel-view image and semantic segmentation synthesis but the effectiveness of our sampling strategy for visual rendering is further verified." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.625, + 0.21, + 0.642 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.651, + 0.473, + 0.879 + ], + "angle": 0, + "content": "3D scene understanding plays a pivotal role in many vision-related tasks, including 3D reconstruction and 3D reasoning. The former domain focuses on low-level understanding, including developing efficient models that interpret and reconstruct observations from various data sources, such as RGB, RGBD, and physical sensors. In contrast, the latter emphasizes high-level understanding, like encoding, segmentation, and common sense knowledge learning. For real-world applications like robotic navigation or augmented reality, both these facets - reconstruction and reasoning - are crucial to enhance interactions with the real world [10]. Among the challenges in the reconstruction domain, novel view synthesis has consistently been a challenging objective. Being able to generate a novel view image based on a few existing views shows that a model com" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.293, + 0.892, + 0.323 + ], + "angle": 0, + "content": "prehends the scene's geometry and possesses a foundational understanding akin to reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.324, + 0.893, + 0.521 + ], + "angle": 0, + "content": "The Neural Radiance Field (NeRF) [1, 4, 8, 19-22, 28, 35] has recently risen as an exciting research area, providing a novel way to tackle the task of novel view synthesis. By encoding the density and emitted radiance at each spatial location, NeRF is able to compress a scene into a learnable model given several images and the corresponding camera poses of the scene. By incorporating a volumetric rendering skill, images of unseen camera views can be generated with convincing quality. However, NeRF primarily focuses on reconstructing the color information of novel views and understanding the associated high-level semantic information (e.g., semantic segmentation or object detection) remains a significant challenge." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.895, + 0.702 + ], + "angle": 0, + "content": "Recent works [9, 15, 25, 30, 31, 37] aim to amplify the high-level scene understanding ability of NeRF by integrating NeRF with semantic segmentation. By sharing the information between the semantic object class and their corresponding appearances, these two tasks are capable of benefiting from each other's insights [17]. Nevertheless, these semantic understanding methods stick closely to the original NeRF paradigm, focusing on exploiting the model to represent a specific scene. As mentioned in [17], such a strategy requires additional annotations of semantic segmentation maps when applied to a new scene, limiting real-world applicability and generalizability." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.895, + 0.87 + ], + "angle": 0, + "content": "To tackle this problem, generalizable NeRFs [3, 14, 18, 27, 32, 33, 36] have emerged as a promising solution. They adopt an on-the-fly approach for building a neural radiance field conditioned on extracted features from input images of different scenes rather than encoding the scene representation directly into the model. For instance, PixelNeRF [36] introduces the idea of conditioning a NeRF model with multi-view images across multiple scenes. This innovation boosts NeRF's ability to generalize to unseen scenes and avoids retraining for each individual scene. Yet, it leaves uncertainties in its application to semantic understanding." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.895, + 0.902 + ], + "angle": 0, + "content": "As a pioneer in integrating a generalizable NeRF with semantic segmentation capabilities, Semantic-Ray (S-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.208, + 0.901 + ], + "angle": 0, + "content": "*Equal Contribution" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20806" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.159, + 0.089, + 0.813, + 0.215 + ], + "angle": 0, + "content": "
Methodgeneralized NeRFgeneralized segmentationsampling strategyefficient sampling
Semantic NeRF [37]--hierarchical✓
Panoptic Neural Fields [15]--uniform-
Panoptic Lifting [25]--uniform-
DM-NeRF [31]--hierarchical✓
NeSF [30]-✓hierarchical✓
Semantic Ray [17]✓✓hierarchical✓
Ours✓✓depth-guided✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.225, + 0.895, + 0.268 + ], + "angle": 0, + "content": "Table 1. Comparisons of semantic-aware NeRFs (i.e., NeRFs, which jointly perform novel view synthesis and semantic segmentation). Note that NeSF only performs generalizable segmentation, yet its view synthesis remains restricted to the scene-specific training scheme. And, Semantic Ray requires ground truth depth as additional inputs for both training and inference." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.291, + 0.473, + 0.502 + ], + "angle": 0, + "content": "Ray) [17] made notable advancements by introducing a generalizable semantic field concept. By incorporating a cross-reprojection attention module, S-Ray aggregates information of multi-view images from all points sampled along a ray and predicts RGB values and semantic labels. Although this approach appears to capture the necessary contextual information for semantic segmentation, aggregating details of all points introduces noisy features that could interfere with segmentation tasks, as segmentation is mainly concerned with the first object class that the ray encounters. Additionally, their hierarchical sampling approach further increases additional computation and model complexity in view synthesis. Thus, a properly designed rendering strategy for a generalizable semantic field would be desirable." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.504, + 0.473, + 0.852 + ], + "angle": 0, + "content": "In this paper, we propose Generalizable Semantic Neural Raisance Fields (GSNeRF) to jointly tackle the problems of generalized novel view synthesis and semantic segmentation. Given a set of source images of a scene with corresponding camera views and a novel target camera view, our GSNeRF is able to derive visual features and depth map predictions of each source view, which are used to predict the depth map of the target view. With the depth map of the target view and properly derived visual features of source views, the novel view RGB image and semantic segmentation can be derived accordingly. Our GSNeRF consists of two key learning stages: Semantic Geo-Reasoning and Depth-Guided Visual Rendering. The former is to derive visual features and aggregate the predicted depth information of each source view to estimate the depth of the novel view, while the latter renders the RGB image and semantic segmentation map of the target view. We further design two distinct sampling strategies to minimize noisy features and enhance rendering efficiency. Through conducting extensive experiments on both real-world and synthetic datasets, we show that our method outperforms current state-of-the-art generalizable NeRF methods in novel view synthesis and semantic segmentation." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.854, + 0.424, + 0.868 + ], + "angle": 0, + "content": "The key contributions of our work are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.871, + 0.471, + 0.902 + ], + "angle": 0, + "content": "- We propose GSNeRF for jointly rendering novel view images and producing the associated semantic segmentation" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.292, + 0.674, + 0.304 + ], + "angle": 0, + "content": "mask on unseen scenes." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.306, + 0.891, + 0.351 + ], + "angle": 0, + "content": "- The proposed Semantic Geo-Reasoning stage learns color, geometry, and semantic information of the input scene, introducing generalization ability of our GSNeRF." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.352, + 0.892, + 0.427 + ], + "angle": 0, + "content": "- Based on the inferred geometry information, the introduced Depth-Guided Visual Rendering stage customizes two different sampling strategies according to the predicted target view depth map, so that image and semantic map rendering can be performed simultaneously." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.306, + 0.892, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.441, + 0.642, + 0.456 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.466, + 0.718, + 0.481 + ], + "angle": 0, + "content": "2.1. Neural Radiance Fields" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.489, + 0.892, + 0.791 + ], + "angle": 0, + "content": "The Neural Radiance Field [20] has emerged as a widely embraced implicit representation by encoding a 3D scene within a neural network. Nonetheless, as mentioned in [7], the original NeRF method demands extensive training time, ranging from hours to days, and relies on dozens of multiview images as input. As a result, a large number of subsequent works [8, 21, 28, 35] are put forward to confront these issues. To shorten the training process, methods such as Instant NGP [21] and DVGO [28] opt for a balance between speed and memory, using hash encoding and voxel encoding to minimize the training time to minutes. On the other hand, to address the demand for multiple input views, some methods [7, 14, 23] introduce depth as additional supervision, a notable example being DS-NeRF [7]. DS-NeRF applies extra depth supervision to ensure that the NeRF accurately encodes a scene's geometry, thus enabling the construction of neural radiance fields with fewer images. This method shows how incorporating depth information can enhance the quality of novel view synthesis while reducing the number of required source images." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.81, + 0.818 + ], + "angle": 0, + "content": "2.2. Generalizable Novel View Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Since NeRF directly encodes information of a scene into a neural network, it is only applicable to the training scene of interest. In response to this limitation, some methods [3, 14, 16, 18, 27, 29, 32, 36] propose on-the-fly construction of NeRF, allowing trained NeRFs applied for syn" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20807" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.213 + ], + "angle": 0, + "content": "thesizing novel views in unseen scenes. The concept is to extract features from each source view image and condition the NeRF renderer on these scene-specific features, typically through [36] 2D CNN or [3, 14] cost volume techniques that aggregate features across different images. While these methods realize generalizable novel view synthesis, extensions of such NeRFs to high-level scene understanding tasks remain an open challenge." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.221, + 0.269, + 0.236 + ], + "angle": 0, + "content": "2.3. Multi-tasking NeRF" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.244, + 0.473, + 0.515 + ], + "angle": 0, + "content": "Recently, several methods [9, 15, 17, 25, 30, 31, 37] seek to enhance NeRF with higher-level understanding abilities. Semantic NeRF [37] stands as the first method to integrate semantic segmentation with NeRF by designing two projection heads for predicting both semantics and color simultaneously. Although it achieves satisfactory performance and robustness on semantic segmentation tasks, it requires retraining for each scene. This implies a need for not only RGB images but also corresponding semantic segmentation maps to represent each new scene, hindering their generalizability. NeSF [30] has made an attempt to improve this aspect by proposing a half-generalizable method that includes a generalizable semantic segmentation module. That is, after training an original NeRF of each scene, they extract a 3D voxel-like grid density from the neural radiance field and train a generalizable 3D UNet to do a semantic segmentation task. However, this learning scheme still requires additional time to retrain NeRFs for novel scenes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.516, + 0.47, + 0.727 + ], + "angle": 0, + "content": "S-Ray [17] advances further by proposing a generalized semantic neural field. Following the backbone of Neuray [18], they propose a cross-reprojection attention module. Given a ray emitted from the target viewpoint, they uniformly sample points along the ray, and hierarchically sample more points on the second forward pass according to the density predicted from the first pass. With the cross-reprojection module, they aggregate the contextual information on all sampled points. Despite achieving good results on both real-world and synthetic datasets, the sampling strategy is time-consuming. Also, aggregating all information along the entire ray can introduce noisy features that may hamper segmentation tasks, as segmentation primarily focuses on the first object class encountered by the ray." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.728, + 0.471, + 0.833 + ], + "angle": 0, + "content": "Contrarily, our proposed GSNeRF aims to predict the depth map of the target view and utilize it to sample points along the ray more efficiently. This unique strategy enables us to sample points focusing on the first encountered object and filter out unrelated noisy points for semantic segmentation prediction. In Table 1, we compare the characteristics of recent semantic-aware NeRFs with GSNeRF." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.845, + 0.416, + 0.862 + ], + "angle": 0, + "content": "3. Brief Review of Generalizable NeRFs" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.472, + 0.903 + ], + "angle": 0, + "content": "For the sake of completeness, we provide a brief review of generalizable NeRFs [3, 14, 18, 29, 36]. Given \\(K\\) multi" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.228 + ], + "angle": 0, + "content": "view images \\( I_{1:K} = \\{I_1, I_2, \\dots, I_K\\} \\) with corresponding camera poses \\( \\xi_{1:K} = \\{\\xi_1, \\xi_2, \\dots, \\xi_K\\} \\) of a scene, and a target camera pose \\( \\xi_T \\), the goal is to synthesize the novel view image \\( I_T \\). Rays are sampled across the scene from the target camera pose to render the novel view image. Along each ray, \\( r \\), \\( N_c \\) points (coarse points) are sampled within the given near bound \\( b_l \\) and far bound \\( b_u \\). Existing methods partition \\( [b_l, b_u] \\) into \\( N_c \\) bins uniformly and draw one sample within each bin as defined by:" + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.236, + 0.892, + 0.267 + ], + "angle": 0, + "content": "\\[\nt _ {n} ^ {r} \\sim \\mathcal {U} \\left[ b _ {l} + \\frac {n - 1}{N _ {c}} \\left(b _ {u} - b _ {l}\\right), b _ {l} + \\frac {n}{N _ {c}} \\left(b _ {u} - b _ {l}\\right) \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.273, + 0.892, + 0.379 + ], + "angle": 0, + "content": "where \\(t_n^r\\) is the \\(n\\)-th sample along \\(r\\). With the corresponding camera position \\(\\sigma^r\\) and viewing directions \\(\\phi^r\\), the \\(n\\)-th sampled point \\(x_n^r\\) can be derived as \\(x_n^r = o^r + t_n^r\\phi^r\\), with \\(x_n^r \\in \\mathbb{R}^3\\). Several methods [3, 18, 29, 36] also use hierarchical sampling, conducting a second forward pass to sample \\(N_f\\) points (fine points) where the first \\(N_c\\) uniformly sampled points show higher predicted density." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.379, + 0.892, + 0.456 + ], + "angle": 0, + "content": "With the above information, a generalized NeRF can be represented by a learnable continuous function \\(A_{\\theta}\\), where \\(\\theta\\) denotes the learnable parameters. \\(A_{\\theta}\\) computes a volumetric density \\(\\sigma_n^r \\in \\mathbb{R}^+\\) and a view-dependent color representation \\(\\mathbf{c}_n^r \\in \\mathbb{R}^3\\) for each point \\(x_n^r\\). That is," + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.463, + 0.892, + 0.481 + ], + "angle": 0, + "content": "\\[\n\\sigma_ {n} ^ {r}, \\mathbf {c} _ {n} ^ {r} = A _ {\\theta} \\left(x _ {n} ^ {r}, \\phi^ {r}, f _ {1: K} \\left(w \\left(x _ {n} ^ {r}\\right)\\right)\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.489, + 0.892, + 0.534 + ], + "angle": 0, + "content": "where \\( f_{1:K} \\) are features of \\( I_{1:K} \\) and \\( w(.) \\) is the function that reprojects the 3D position \\( x_{n}^{r} \\) onto the image plane with known camera poses." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.892, + 0.565 + ], + "angle": 0, + "content": "To this end, the color value \\(\\mathbf{C}(r)\\) of the pixel corresponding to the ray \\(r\\) is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.618, + 0.573, + 0.892, + 0.614 + ], + "angle": 0, + "content": "\\[\n\\mathbf {C} (r) = \\sum_ {n = 1} ^ {N _ {c}} T _ {n} ^ {r} \\alpha_ {n} ^ {r} \\mathbf {c} _ {n} ^ {r}. \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.635, + 0.892, + 0.675 + ], + "angle": 0, + "content": "\\[\nT _ {n} ^ {r} = \\exp \\left(- \\sum_ {i = 1} ^ {n - 1} \\sigma_ {i} ^ {r} \\delta_ {i} ^ {r}\\right), \\quad \\alpha_ {n} ^ {r} = 1 - \\exp \\left(- \\sigma_ {n} ^ {r} \\delta_ {n} ^ {r}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.683, + 0.892, + 0.728 + ], + "angle": 0, + "content": "where \\( T_{n}^{r} \\) is the accumulated transmittance and \\( \\delta_n^r = ||x_{n + 1}^r -x_n^r ||_2 \\) represents the distance between every two adjacent points along \\( r \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.729, + 0.892, + 0.774 + ], + "angle": 0, + "content": "As is the case with other NeRF methods, a generalized NeRF also optimizes the model based on the rendering loss between the rendered image and the GT image:" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.782, + 0.892, + 0.814 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {i m a g e}} = \\sum_ {r \\in R} \\| \\mathbf {C} (r) - \\hat {\\mathbf {C}} (r) \\| _ {2} ^ {2}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.855 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{C}}(r)\\) represents the GT color of the renderer image corresponding to a ray \\(r\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Our method is built on top of generalized NeRF methods but with a more efficient sampling strategy that can deal with both semantic segmentation and image rendering." + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20808" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.127, + 0.096, + 0.832, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.318, + 0.895, + 0.377 + ], + "angle": 0, + "content": "Figure 1. Overview of GSNeRF: including Semantic Geo-Reasoning and Depth-Guided Visual Rendering. Given K multi-view image \\( I_{1:K} \\) of a scene, the Semantic Geo-Reasoner predicts the depth map \\( D_{1:K} \\) for each source image, which is aggregated to estimate the target view depth map \\( D_T \\). With \\( D_T \\) as key geometric guidance, we design Depth-Guided Visual Rendering to render target view image \\( I_T \\) and semantic segmentation \\( S_T \\), by Volume Renderer \\( R_\\theta \\) and Semantic Renderer \\( P_\\theta \\), respectively." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.394, + 0.473, + 0.566 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.575, + 0.47, + 0.646 + ], + "angle": 0, + "content": "Figure 2. Semantic Geo-Reasoning. The Semantic Geo-Reasoner contains a shared Encoder \\( E_{\\theta} \\), a shared Decoder \\( D_{\\theta} \\) and a cost-volume aggregator \\( C_{\\theta} \\), producing image and volume features \\( (f_{1:K}^{I} \\) and \\( f_{1:K}^{V}) \\), with the associated depth maps \\( D_{1:K} \\). The depth map \\( D_{T} \\) of the target view is estimated from \\( D_{1:K} \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.67, + 0.17, + 0.685 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.695, + 0.445, + 0.71 + ], + "angle": 0, + "content": "4.1. Problem Formulation and Model Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.719, + 0.47, + 0.884 + ], + "angle": 0, + "content": "We first define the setting and notations used in this paper. Given a scene with \\(K\\) multi-view images \\(I_{1:K}\\) and their corresponding camera poses \\(\\xi_{1:K}\\) along with a target camera pose as input, we aim to achieve novel-view synthesis and semantic segmentation simultaneously. During training, we synthesize both the novel view image \\(I_T\\) and the semantic segmentation map \\(S_T\\) from \\(\\xi_T\\), given ground truth (GT) image \\(\\hat{I}_T\\) and semantic segmentation map \\(\\hat{S}_T\\) of the target view \\(\\xi_T\\). Once the model completes training on a given set of training scenes, it is able to directly generalize to unseen scenes without finetuning." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.469, + 0.901 + ], + "angle": 0, + "content": "To achieve this goal, we propose Generalizable Seman" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.397, + 0.895, + 0.73 + ], + "angle": 0, + "content": "tic Neural Radiance Fields (GSNeRF), as illustrated in Figure 1. GSNeRF consists of two key learning stages: Semantic Geo-Reasoning and Depth-Guided Visual Rendering. For Semantic Geo-Reasoning, we employ a Semantic Geo-Reasoner \\( G_{\\theta} \\) on each input source image \\( I_{k} \\in \\mathbb{R}^{H \\times W \\times 3} \\) to create 2D feature \\( f_{k}^{I} \\in \\mathbb{R}^{H \\times W \\times d} \\), semantic feature \\( f_{k}^{S} \\in \\mathbb{R}^{H \\times W \\times d} \\), 3D volume feature \\( f_{k}^{V} \\in \\mathbb{R}^{H \\times W \\times L \\times d} \\), and depth predictions \\( D_{k} \\in \\mathbb{R}^{H \\times W} \\), where \\( H \\) and \\( W \\) represent the height and width of the image, \\( L \\) is the depth sample number defined in [34] to determine the volume size for the 3D feature, and \\( d \\) is the feature dimension. The source view depth maps are integrated to estimate the depth for the target view \\( D_{T} \\). As for Depth-Guided Visual Rendering, we conduct a unique sampling strategy based on \\( D_{T} \\) to minimize noisy features and enhance rendering efficiency. The sampled points, along with the previously extracted feature, are respectively input into the volume renderer \\( R_{\\theta} \\) and semantic renderer \\( P_{\\theta} \\), synthesizing the target view's image and semantic segmentation map. In the following subsection, we will explain the Semantic Geo-Reasoning and Depth-guided Visual Rendering process within our GSNeRF framework." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.744, + 0.769, + 0.759 + ], + "angle": 0, + "content": "4.2. Generalizable Semantic NeRF" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.769, + 0.73, + 0.784 + ], + "angle": 0, + "content": "4.2.1 Semantic Geo-Reasoning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Given \\(K\\) multi-view source images \\(I_{1:K}\\) of a scene, our Semantic Geo-Reasoner \\(G_{\\theta}\\) in Figure 1 is designed to achieve two goals. First, geometric clues are extracted from \\(G_{\\theta}\\), comprising 3D Volume features \\(f_{1:K}^{V}\\) and depth maps \\(D_{1:K}\\) for each source image. In addition, 2D image features \\(f_{1:K}^{I}\\) and semantic features \\(f_{1:K}^{S}\\) for each source image can be jointly produced. Secondly, our \\(G_{\\theta}\\) is learned to predict the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20809" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "target view depth map \\( D_{T} \\), which is realized by deterministic aggregation and reprojection of all source view depth predictions \\( D_{1:K} \\). We now detail this learning stage." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.138, + 0.468, + 0.29 + ], + "angle": 0, + "content": "As depicted in Figure 2, our Semantic Geo-Reasoner consists of a shared 2D CNN encoder \\( E_{\\theta} \\) that extracts \\( f_{1:K}^{I} \\) from every image \\( I_{1:K} \\). To construct geometry-aware features, a cost volume aggregator \\( C_{\\theta} \\) [6, 12, 13, 34] is applied to gather 2D features \\( f_{1:K}^{I} \\) across different images. As a result, \\( C_{\\theta} \\) produces 3D feature \\( f_{1:K}^{V} \\), which are ultimately used to predict the depth map \\( D_{1:K} \\) for every source view image. To obtain semantic-aware features for semantic prediction, a shared semantic feature decoder \\( D_{\\theta} \\) is additionally applied to derive 2D semantic features \\( f_{1:K}^{S} \\) from \\( f_{1:K}^{I} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.29, + 0.468, + 0.425 + ], + "angle": 0, + "content": "Since \\(D_{1:K}\\) is a vital intermediate element that assists in estimating \\(D_T\\), which allows the subsequent depth-guided rendering process, we strive to ensure the precise prediction of \\(D_{1:K}\\), whether or not the ground truth depth images are accessible. To achieve this, we integrate ideas from DSNeRF [7] and GeoNeRF [14] to supervise our \\(G_{\\theta}\\) with either a depth-supervised loss or a self-supervised depth loss. If the ground-truth depth maps of \\(D_{1:K}\\) are available, our depth-supervised loss \\(\\mathcal{L}_D\\) is given as:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.44, + 0.469, + 0.481 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {D} = \\frac {1}{K} \\left(\\sum_ {k = 1} ^ {K} \\| D _ {k} - \\hat {D} _ {k} \\| _ {s 1}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.491, + 0.468, + 0.584 + ], + "angle": 0, + "content": "Here, \\(\\hat{D}_k\\) represents the GT depth map of view \\(k\\), \\(\\| \\cdot \\|_{s1}\\) represents smooth L1 loss [11]. If \\(\\hat{D}_k\\) is unavailable, we use a self-supervised depth loss \\(\\mathcal{L}_{ssl}\\) to regularize our depth predictions by considering cross-view depth consistency between all source views. Following the self-supervised depth estimation approach of RC-MVSNet [2], \\(\\mathcal{L}_{ssl}\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.128, + 0.601, + 0.469, + 0.617 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s s l} = \\lambda_ {1} \\mathcal {L} _ {R C} + \\lambda_ {2} \\mathcal {L} _ {S S I M} + \\lambda_ {3} \\mathcal {L} _ {S m o o t h}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.626, + 0.468, + 0.686 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{RC}\\), \\(\\mathcal{L}_{SSIM}\\), and \\(\\mathcal{L}_{Smooth}\\) represent reconstruction, structure similarity, and depth smoothness losses [2]. The hyper-parameters \\(\\lambda_1\\), \\(\\lambda_2\\), and \\(\\lambda_3\\) are also selected following [2] (see supplementary material for details)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.687, + 0.468, + 0.792 + ], + "angle": 0, + "content": "To this end, the target view depth map \\( D_{T} \\) is calculated using \\( D_{1:K} \\) through a deterministic algorithm. In simple terms, we directly project the pixels of each depth map into 3D space and then reproject them back to the target view. More details about the self-supervised depth loss and corresponding pseudo-code of target view depth estimation are available in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.815, + 0.357, + 0.831 + ], + "angle": 0, + "content": "4.2.2 Depth-Guided Visual Rendering" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.468, + 0.901 + ], + "angle": 0, + "content": "We now discuss how we perform volume and semantic rendering with the guidance of the predicted \\( D_{T} \\). That is, given a target view \\( \\xi_{T} \\), a ray is sampled for each pixel on the image plane to render the corresponding RGB value for \\( I_{T} \\)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.89, + 0.121 + ], + "angle": 0, + "content": "and the semantic segmentation output for \\( S_{T} \\) via the Volume Renderer \\( R_{\\theta} \\) and the Semantic Renderer \\( P_{\\theta} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.14, + 0.892, + 0.307 + ], + "angle": 0, + "content": "Volume Rendering For volume rendering of \\( I_{T} \\), existing semantic-NeRF approaches [17, 30, 37] follow conventional NeRFs [20] by applying either uniform sampling or hierarchical sampling along each ray \\( r \\), as depicted in Table 1. However, both strategies require dense sampling to locate the surface of objects in the given scene, limiting the sampling efficiency. To promote the efficiency of our sampling process, we conduct a depth-guided sampling strategy by focusing on sampling points near the depth values of \\( D_{T} \\). More specifically, this is achieved by modifying the sampling strategy in Eq. 1 as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.317, + 0.892, + 0.348 + ], + "angle": 0, + "content": "\\[\nt _ {n} ^ {r} \\sim \\mathcal {G} (z, v ^ {2}), \\quad v = \\frac {\\operatorname* {m i n} \\left(\\| z - b _ {u} \\| , \\| z - b _ {l} \\|\\right)}{3}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.357, + 0.892, + 0.508 + ], + "angle": 0, + "content": "where \\( t_n^r \\) is the \\( n \\)-th sample along the ray \\( r \\), and \\( \\mathcal{G} \\) represents Gaussian (normal) distribution. Note that \\( z \\) and \\( v \\) represent the mean and standard deviation of the distribution, respectively, while the value of \\( z \\) is determined as the predicted depth for the pixel \\( p \\) that emits the ray \\( r \\) (i.e., \\( z = D_T(p) \\)). By these sampling point values, we derive \\( N \\) sampled points \\( x_{1:N}^r \\) (i.e., \\( x_n^r = o^r + t_n^r\\phi^r \\)) closer to the surface of the first encountered object by \\( r \\), avoiding considering noisy points that are far from the object surface. Please see the supplementary material for further details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.508, + 0.892, + 0.614 + ], + "angle": 0, + "content": "We feed all these points alongside the 2D feature \\( f_{1:K}^{I} \\) and 3D volume feature \\( f_{1:K}^{V} \\) as additional conditions into the volume renderer \\( R_{\\theta} \\) to predict the novel view image of the scene. Specifically, we reproject the point \\( x_{n}^{r} \\) onto each source view to get aggregated features from these condition features. Thus, we write the aggregated feature \\( f_{n,k} \\in \\mathbb{R}^{2d} \\) of the \\( n \\)-th sampled point \\( x_{n}^{r} \\) from the \\( k \\)-th view as:" + }, + { + "type": "equation", + "bbox": [ + 0.552, + 0.625, + 0.891, + 0.643 + ], + "angle": 0, + "content": "\\[\nf _ {n, k} = \\left[ f _ {k} ^ {I} \\left(w \\left(x _ {n} ^ {r}\\right)\\right); f _ {k} ^ {V} \\left(x _ {n} ^ {r}\\right) \\right], \\forall k \\in [ 1: K ] \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.654, + 0.892, + 0.713 + ], + "angle": 0, + "content": "where \\( w(\\cdot) \\) is the reprojection function used in Eq. 2 and \\( [\\cdot; \\cdot] \\) denotes concatenation along feature dimension. We further define the global feature \\( f_{n,0} \\) as guidance across all \\( K \\) source views as:" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.725, + 0.891, + 0.744 + ], + "angle": 0, + "content": "\\[\nf _ {n, 0} = \\left[ \\operatorname {m e a n} \\left(\\left\\{f _ {n, k} \\right\\} _ {k = 1} ^ {K}\\right); \\operatorname {v a r} \\left(\\left\\{f _ {n, k} \\right\\} _ {k = 1} ^ {K}\\right) \\right]. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.755, + 0.892, + 0.816 + ], + "angle": 0, + "content": "Following GeoNeRF [14] methodology, the volume renderer \\(R_{\\theta}\\) (with learnable parameters \\(\\theta\\)) predicts per point density and radiance based on the feature \\(f_{n} = \\{f_{n,k}\\}_{k=0}^{K} \\in \\mathbb{R}^{(K+1) \\times 2d}\\) mentioned above:" + }, + { + "type": "equation", + "bbox": [ + 0.593, + 0.828, + 0.891, + 0.844 + ], + "angle": 0, + "content": "\\[\n\\sigma_ {n} ^ {r}, \\mathbf {c} _ {n} ^ {r} = R _ {\\theta} \\left(x _ {n} ^ {r}, \\phi^ {r}, f _ {n}, M _ {n}\\right), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.855, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(M_{n} = \\{M_{n,k}\\}_{k = 0}^{K}\\) denotes the mask used to mask out unrelated features of a view. If the sampled point \\(x_{n}^{r}\\) lies behind the predicted depth map \\(D_{k}\\) or if the reprojected" + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20810" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.089, + 0.895, + 0.294 + ], + "angle": 0, + "content": "
Generalized methodGT DepthScanNet [5]Replica [26]
Train / TestmIoUacc. / class acc.PSNR↑SSIM↑LPIPS↓mIoUacc. / class acc.PSNR↑SSIM↑LPIPS↓
Neuray [18] + semhead✓ / ✓52.0967.81 / 61.9825.0183.0731.6344.3779.93 / 54.2526.2187.3730.93
GeoNeRF [14] + semhead✓ /53.7876.18 / 61.9032.5590.8812.6945.1281.67 / 52.3628.7088.9420.42
S-Ray [17]✓ / ✓55.5377.79 / 60.9225.1983.6630.9845.3080.48 / 53.7226.3888.1330.04
GSNeRF (Ours)✓ /58.3079.79 / 65.9331.3390.7312.5351.5283.41 / 61.2931.1692.4412.54
MVSNeRF [3] + semhead43.0666.90 / 53.6324.1480.3634.6330.2169.35 / 39.7523.6884.3728.08
GeoNeRF [14] + semhead45.1167.12 / 53.4430.7588.2716.4840.3574.63 / 49.1829.9291.1417.60
GNT [32] + semhead43.4962.06 / 51.8424.3982.3728.3638.1471.44 / 47.4624.5687.3120.97
Neuray [18] + semhead46.0966.39 / 53.7925.2484.3931.3340.9176.23 / 50.1527.8089.5523.68
S-Ray [17]47.6964.90 / 54.4725.1384.1830.4443.2777.63 / 52.8526.7788.5422.81
GSNeRF (Ours)52.2174.71 / 60.1431.4990.3913.8751.2383.06 / 61.1031.7192.8912.93
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.304, + 0.893, + 0.334 + ], + "angle": 0, + "content": "Table 2. Quantitative results on ScanNet & Replica. Note that methods in the first four rows take GT depth as inputs or training supervision, while the methods in the last six rows do not observe GT depth during training/testing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.351, + 0.469, + 0.381 + ], + "angle": 0, + "content": "location is outside the source image, \\(M_{n,k}\\) is set to 0. However, the mask for the global feature \\(M_{n,0}\\) is always set to 1." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.381, + 0.469, + 0.429 + ], + "angle": 0, + "content": "To this end, we apply the same volume rendering as in Eq. 3, predicting the target view Image \\( I_{T} \\). We also apply \\( \\mathcal{L}_{image} \\) as in Eq. 5 to supervise the rendered image quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.447, + 0.473, + 0.659 + ], + "angle": 0, + "content": "Semantic Rendering Semantic rendering of \\( S_{T} \\), on the other hand, is different from the above volume rendering process. Specifically, image rendering considers the color of both the surface and nearby objects because of possible semi-transparent or transparent objects. Conversely, semantic segmentation focuses only on the first object a ray touches. Therefore, for each ray \\( r \\), we sample only one surface point \\( x_{z}^{r} \\) estimated by the target depth map for each ray \\( r \\) to render its semantic value (i.e., \\( x_{z}^{r} = o^{r} + z\\phi^{r} \\), where \\( z \\) is the same depth value as mentioned in Eq. 8). More specifically, with the aggregated semantic feature \\( f_{z} = \\{f_{k}^{S}(w(x_{z}^{r}))\\}_{k=1}^{K} \\in \\mathbb{R}^{K \\times d} \\), our semantic renderer \\( P_{\\theta} \\) with learnable parameters \\( \\theta \\) predicts the semantic logits \\( \\mathbf{S}(r) \\) of ray \\( r \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.671, + 0.47, + 0.689 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S} (r) = P _ {\\theta} \\left(x _ {z} ^ {r}, f _ {z} ^ {\\prime}, M\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.7, + 0.124, + 0.713 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.713, + 0.469, + 0.732 + ], + "angle": 0, + "content": "\\[\nf _ {z} ^ {\\prime} = f _ {z} + \\left\\{f _ {0} \\right\\} \\in \\mathbb {R} ^ {(K + 1) \\times d}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.739, + 0.47, + 0.768 + ], + "angle": 0, + "content": "Note that we follow volume rendering in Eq. 10 to define a global semantic feature \\( f_{0} \\) as guidance across all views:" + }, + { + "type": "equation", + "bbox": [ + 0.181, + 0.782, + 0.47, + 0.799 + ], + "angle": 0, + "content": "\\[\nf _ {0} = \\left[ \\operatorname {m e a n} \\left(f _ {z}\\right); \\operatorname {v a r} \\left(f _ {z}\\right) \\right]. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.809, + 0.469, + 0.87 + ], + "angle": 0, + "content": "As for the mask \\( M = \\{M_k\\}_{k=0}^K \\) in Eqn. 12, it is similar to the one used in Eq. 11, where \\( M_k \\) is to mask out the unrelated features of a view if the sample point \\( x_z^r \\) is behind the corresponding depth value of \\( D_k \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Eventually, we utilize a semantic loss to supervise the novel view semantic prediction. Here, we use a multi-class" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.352, + 0.884, + 0.367 + ], + "angle": 0, + "content": "cross-entropy loss to reduce the semantic prediction error:" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.373, + 0.892, + 0.405 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s e m}} = \\sum_ {r \\in R} (\\mathbf {S} (r) \\log \\hat {\\mathbf {S}} (r)), \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.413, + 0.892, + 0.445 + ], + "angle": 0, + "content": "where \\(\\hat{S}(r)\\) represent the ground truth segmentation score along the ray \\(r\\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.451, + 0.715, + 0.467 + ], + "angle": 0, + "content": "4.3. Training and Inference" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.474, + 0.617, + 0.489 + ], + "angle": 0, + "content": "4.3.1 Training" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.498, + 0.892, + 0.558 + ], + "angle": 0, + "content": "At the training stage, the total loss is calculated by combining the losses that are mentioned above. The aggregated loss function for our model, when ground-truth depth is available, is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.566, + 0.891, + 0.582 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {i m a g e}} + \\mathcal {L} _ {D} + \\lambda \\mathcal {L} _ {\\text {s e m}}. \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.588, + 0.874, + 0.604 + ], + "angle": 0, + "content": "If the GT depth is not available, our model is trained by:" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.611, + 0.891, + 0.627 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {i m a g e}} + \\mathcal {L} _ {\\text {s s l}} + \\lambda \\mathcal {L} _ {\\text {s e m}}, \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.634, + 0.829, + 0.649 + ], + "angle": 0, + "content": "where the coefficient \\(\\lambda\\) is set to 0.5 for simplicity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.65, + 0.892, + 0.725 + ], + "angle": 0, + "content": "In essence, this optimization process enables our model to estimate the depth map of the target camera view accurately. Utilizing this depth information, the model is capable of efficiently sampling points to render new view images and create corresponding semantic segmentation maps." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.741, + 0.622, + 0.755 + ], + "angle": 0, + "content": "4.3.2 Inference" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "At the inference stage, our GSNeRF is able to generalize to unseen scenes without the need for retraining from scratch. Due to the design of the conditioning renderer on scene-specific features \\( f^{I}, f^{V} \\), and \\( f^{S} \\), our model is able to build a semantic neural radiance field on the fly. Once we finish training our model on training scenes, we are able to directly infer novel view images and semantic segmentation maps on any scene given multi-view images and corresponding camera poses of that scene." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20811" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.127, + 0.089, + 0.169, + 0.101 + ], + "angle": 0, + "content": "S-Ray" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.103, + 0.215, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.229, + 0.089, + 0.338, + 0.102 + ], + "angle": 0, + "content": "GSNeRF (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.102, + 0.35, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.406, + 0.089, + 0.43, + 0.1 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.102, + 0.484, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.181, + 0.214, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.181, + 0.349, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.181, + 0.484, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.26, + 0.215, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.26, + 0.35, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.26, + 0.484, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.339, + 0.214, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.339, + 0.349, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.339, + 0.484, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.089, + 0.574, + 0.102 + ], + "angle": 0, + "content": "S-Ray" + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.102, + 0.619, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.634, + 0.089, + 0.741, + 0.102 + ], + "angle": 0, + "content": "GSNeRF (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.622, + 0.102, + 0.754, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.811, + 0.089, + 0.835, + 0.1 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.757, + 0.102, + 0.889, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.181, + 0.618, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.621, + 0.181, + 0.753, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.757, + 0.181, + 0.888, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.26, + 0.618, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.621, + 0.26, + 0.753, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.757, + 0.26, + 0.888, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.339, + 0.618, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.621, + 0.339, + 0.753, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.757, + 0.339, + 0.888, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.43, + 0.893, + 0.459 + ], + "angle": 0, + "content": "Figure 3. Qualitative evaluation. We compare the visual quality of the rendered novel view images (the first three columns) and semantic segmentation maps (the last three columns) with S-Ray [17]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.483, + 0.209, + 0.5 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.51, + 0.179, + 0.525 + ], + "angle": 0, + "content": "5.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.535, + 0.47, + 0.761 + ], + "angle": 0, + "content": "To evaluate the effectiveness of our proposed method, we conduct experiments on both real-world and synthetic datasets. For real-world data, we use ScanNet [5], a large-scale, indoor RGB-D video dataset with over 2.5 million views from 1513 distinct scenes, including semantic annotations and camera poses. Following the setting presented in S-Ray [17], we train our model on 60 scenes and test its generalization on 10 new, unseen scenes. For synthetic data, we utilize Replica [26], a 3D reconstruction-based indoor dataset comprising 18 high-quality scenes with dense geometry, HDR textures, and semantic labels. Following the Semantic NeRF [37] supplied rendered data, we train our model on 6 distinct scenes across 12 video sequences and test on 2 novel scenes across 4 video sequences. Please see the supplementary material for details." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.774, + 0.268, + 0.79 + ], + "angle": 0, + "content": "5.2. Results and analysis" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.799, + 0.274, + 0.814 + ], + "angle": 0, + "content": "5.2.1 Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Table 2 compares our method against several baselines, in which S-Ray [17] serves as a primary baseline since it is the first to address semantic segmentation with NeRF in a generalized setting. We further conduct experiments on MVS-NeRF [3], GeoNeRF [14], GNT [32] and NeuRay [18] with" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.485, + 0.892, + 0.591 + ], + "angle": 0, + "content": "the semantic head as described in [17]. It is worth noting that, even though S-Ray does not explicitly mention using a depth map of each source view image as input in their paper, their official implementation takes the depth maps \\( D_{1:K} \\) as additional input along with \\( I_{1:K} \\) during training and testing. Therefore, we follow their official implementation to reproduce S-Ray results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.893, + 0.901 + ], + "angle": 0, + "content": "From Table 2, it can be seen that our model generalizes well to unseen scenes. To show that our GSNeRF is also capable of producing impressive results without depth supervision, we conduct an additional experiment in which we only use the loss function in Eq. 17 to train our GSNeRF in Table 2. We also conduct experiments of S-Ray, Neuray with only RGB images as input, and GeoNeRF without GT depth as supervision for a fair comparison. From Table 2, we see that even without access to GT depth and with the target depth information regularized in a self-supervised manner, our method still outperforms all the other baseline approaches. This suggests that our model can be trained without requiring depth map supervision, confirming the effectiveness and practicality of our proposed approach. We also note that, although GeoNeRF + semhead (with depth supervision) slightly outperforms our approach regarding PSNR and SSIM for the rendered RGB images on ScanNet, our method excels in all semantic segmentation metrics by approximately \\(5\\%\\). More comprehensive discussions are presented in the supplementary materials." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20812" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.079, + 0.089, + 0.472, + 0.205 + ], + "angle": 0, + "content": "
ITST
ModelDepth-guidedrenderer PθDepth-guidedPSNRmIoU
A---28.4837.88
B✓--30.5641.72
C✓✓-30.8849.47
Ours✓✓✓31.4952.21
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.214, + 0.472, + 0.271 + ], + "angle": 0, + "content": "Table 3. Ablation studies on GSNeRF. We verify the effectiveness of using depth-guided rendering for rendering \\( I_{T} \\) and \\( S_{T} \\). Note that, without \\( P_{\\theta} \\), we simply take a pre-trained 2D segmentor on \\( I_{T} \\) to produce \\( S_{T} \\)." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.279, + 0.473, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.475, + 0.47, + 0.545 + ], + "angle": 0, + "content": "Figure 4. Sampling efficiency on ScanNet. Compared to SOTAs, our GSNeRF is able to achieve significantly improved rendering performance especially when the number of sampling points is small. Even with increasing numbers of sampling points, GSNeRF still performs favorably against existing models." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.564, + 0.265, + 0.579 + ], + "angle": 0, + "content": "5.2.2 Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.587, + 0.47, + 0.784 + ], + "angle": 0, + "content": "The qualitative results of our approach compared to S-Ray [17] are depicted in Fig. 3. This comparison illustrates that while the existing method can approximate the contours of objects, it is unable to capture the geometrical details, causing segmentation inconsistencies. In contrast, our GSNeRF accurately interprets scene geometry through our targeted view depth estimation design, allowing us to not only precisely segment objects but also render novel view images more realistically. This confirms that our depth-guided sampling technique effectively captures the correct geometry and that the sampled point features are appropriately aggregated from features of the Semantic Geo-Reasoner." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.248, + 0.816 + ], + "angle": 0, + "content": "5.2.3 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "To further analyze the effectiveness of our designed modules, we conduct ablation studies on ScanNet, as shown in Table 3. Baseline model A employs our Semantic Geo-Reasoner \\( G_{\\theta} \\) and volume renderer \\( R_{\\theta} \\) to create the target view image \\( I_{T} \\), using uniform sampling instead of depth" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.259 + ], + "angle": 0, + "content": "guided sampling \\( x_{1:N} \\). Also, a pre-trained 2D segmentor [24] is applied on the rendered image \\( I_T \\) to produce \\( S_T \\) without our semantic renderer \\( P_\\theta \\). Model B enhances visual quality by replacing uniform sampling with depth-guided sampling. We further replace the 2D segmentor with our semantic renderer \\( P_\\theta \\) using uniform sampling for semantic rendering in model C, illustrating mutual benefits for both tasks. Our full model in the last row uses depth-guided sampling exclusively for surface points in the semantic renderer, achieving optimal results. This verifies the success of our proposed modules and sampling strategies." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.276, + 0.694, + 0.292 + ], + "angle": 0, + "content": "5.2.4 Sampling Efficiency" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.3, + 0.892, + 0.557 + ], + "angle": 0, + "content": "Another benefit of our depth-guided sampling is that the performance of our GSNeRF is less sensitive to the number of sampled points along each ray. In Fig. 4, we conduct experiments of reducing the number of sampling points \\( N \\), showing PSNR of the synthesized novel view images as evaluation. We find that by employing depth-guided sampling, our method accurately preserves visual details with less noise (compared to Neuray and S-Ray with 128 sampled points along each ray). It can be seen that GeoNeRF + semhead shows a catastrophic drop in PSNR with fewer points, while ours leveraging depth-guided sampling maintains satisfactory visual quality even with a significantly reduced number of sampled points (e.g., when sampling only 4 points along a ray, the PSNR remains near 28). This result further attests to the success of our design in estimating target depth and implementing the corresponding sampling strategy with better novel view synthesis performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.571, + 0.619, + 0.586 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.595, + 0.892, + 0.777 + ], + "angle": 0, + "content": "In this paper, we proposed Generalizable Semantic Neural Raidance Fields (GSNeRF) to achieve generalized novel view synthesis and semantic segmentation. Our GSNeRF is trained to derive visual features and perform depth map prediction of each source view so that the depth map for a novel target view can be estimated. With such target-view depth information observed, the associated RGB image and semantic segmentation can be jointly produced via depth-guide rendering. In our experiments, we quantitatively and qualitatively confirm that our GSNeRF performs favorably against existing generalizable semantic-aware NeRF methods on both real-world and synthetic datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Acknowledgement This work is supported in part by the National Science and Technology Council via grant NSCT112-2634-F-002-007 and the NTU Center of Data Intelligence: Technologies, Applications, and Systems via grant NTU-113L900902. We also thank the National Center for High-performance Computing (NCHC) for providing computational and storage resources." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20813" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.197 + ], + "angle": 0, + "content": "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.472, + 0.268 + ], + "angle": 0, + "content": "[2] Di Chang, Aljaž Božić, Tong Zhang, Qingsong Yan, Yingcong Chen, Sabine Susstrunk, and Matthias Nießner. Rcmvsnet: unsupervised multi-view stereo with neural rendering. In European Conference on Computer Vision, pages 665-680. Springer, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.27, + 0.47, + 0.338 + ], + "angle": 0, + "content": "[3] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14124-14133, 2021. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.34, + 0.47, + 0.394 + ], + "angle": 0, + "content": "[4] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.397, + 0.47, + 0.466 + ], + "angle": 0, + "content": "[5] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.467, + 0.47, + 0.522 + ], + "angle": 0, + "content": "[6] Yuchao Dai, Zhidong Zhu, Zhibo Rao, and Bo Li. Mvs2: Deep unsupervised multi-view stereo with multi-view symmetry. In 2019 International Conference on 3D Vision (3DV), pages 1-8. IEEE, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.523, + 0.47, + 0.59 + ], + "angle": 0, + "content": "[7] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12882-12891, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.593, + 0.47, + 0.661 + ], + "angle": 0, + "content": "[8] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.663, + 0.469, + 0.732 + ], + "angle": 0, + "content": "[9] Xiao Fu, Shangzhan Zhang, Tianrun Chen, Yichong Lu, Lanyun Zhu, Xiaowei Zhou, Andreas Geiger, and Yiyi Liao. Panoptic nerf: 3d-to-2d label transfer for panoptic urban scene segmentation. In 2022 International Conference on 3D Vision (3DV), pages 1-11. IEEE, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[10] Sourav Garg, Niko Sünderhauf, Feras Dayoub, Douglas Morrison, Akansel Cosgun, Gustavo Carneiro, Qi Wu, Tat-Jun Chin, Ian Reid, Stephen Gould, et al. Semantics for robotic mapping, perception and interaction: A survey. Foundations and Trends® in Robotics, 8(1-2):1-224, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[11] Ross Girshick. Fast r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 1440-1448, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[12] Xiaodong Gu, Zhiwen Fan, Siyu Zhu, Zuozhuo Dai, Feitong Tan, and Ping Tan. Cascade cost volume for high-resolution multi-view stereo and stereo matching. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.094, + 0.892, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF conference on computer vision and pattern recognition, pages 2495-2504, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.188 + ], + "angle": 0, + "content": "[13] Baichuan Huang, Hongwei Yi, Can Huang, Yijia He, Jingbin Liu, and Xiao Liu. M3vsnet: Unsupervised multi-metric multi-view stereo network. In 2021 IEEE International Conference on Image Processing (ICIP), pages 3163-3167. IEEE, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.191, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[14] Mohammad Mahdi Johari, Yann Lepoittevin, and François Fleuret. Geonerf: Generalizing nerf with geometry priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18365-18375, 2022. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.356 + ], + "angle": 0, + "content": "[15] Abhijit Kundu, Kyle Genova, Xiaqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12871-12881, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.358, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[16] Haotong Lin, Sida Peng, Zhen Xu, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Efficient neural radiance fields for interactive free-viewpoint video. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[17] Fangfu Liu, Chubin Zhang, Yu Zheng, and Yueqi Duan. Semantic ray: Learning a generalizable semantic field with cross-reprojection attention. In CVPR, pages 17386-17396, 2023. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.47, + 0.892, + 0.551 + ], + "angle": 0, + "content": "[18] Yuan Liu, Sida Peng, Lingjie Liu, Qianqian Wang, Peng Wang, Christian Theobalt, Xiaowei Zhou, and Wenping Wang. Neural rays for occlusion-aware image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7824-7833, 2022. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.553, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[19] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7210-7219, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.892, + 0.704 + ], + "angle": 0, + "content": "[20] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[21] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.763, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[22] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14335-14345, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[23] Barbara Roessle, Jonathan T Barron, Ben Mildenhall, Pratul P Srinivasan, and Matthias Nießner. Dense depth priors for neural radiance fields from sparse input views. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12892-12901, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.094, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20814" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[24] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.177, + 0.471, + 0.259 + ], + "angle": 0, + "content": "[25] Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulò, Norman Müller, Matthias Nießner, Angela Dai, and Peter Kontschieder. Panoptic lifting for 3d scene understanding with neural fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9043-9052, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.261, + 0.469, + 0.328 + ], + "angle": 0, + "content": "[26] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.469, + 0.385 + ], + "angle": 0, + "content": "[27] Mohammed Suhail, Carlos Esteves, Leonid Sigal, and Ameesh Makadia. Generalizable patch-based neural rendering. In European Conference on Computer Vision, pages 156-174. Springer, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.386, + 0.469, + 0.453 + ], + "angle": 0, + "content": "[28] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.469, + 0.511 + ], + "angle": 0, + "content": "[29] Alex Trevithick and Bo Yang. Grf: Learning a general radiance field for 3d representation and rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15182-15192, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.469, + 0.58 + ], + "angle": 0, + "content": "[30] Suhani Vora, Noha Radwan, Klaus Greff, Henning Meyer, Kyle Genova, Mehdi SM Sajjadi, Etienne Pot, Andrea Tagliasacchi, and Daniel Duckworth. Nesf: Neural semantic fields for generalizable semantic segmentation of 3d scenes. arXiv preprint arXiv:2111.13260, 2021. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.58, + 0.469, + 0.622 + ], + "angle": 0, + "content": "[31] Bing Wang, Lu Chen, and Bo Yang. Dm-nerf: 3d scene geometry decomposition and manipulation from 2d images. arXiv preprint arXiv:2208.07227, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.469, + 0.663 + ], + "angle": 0, + "content": "[32] Peihao Wang, Xuxi Chen, Tianlong Chen, Subhashini Venugopalan, Zhangyang Wang, et al. Is attention all nerf needs? arXiv preprint arXiv:2207.13298, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.469, + 0.748 + ], + "angle": 0, + "content": "[33] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690-4699, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.749, + 0.469, + 0.803 + ], + "angle": 0, + "content": "[34] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vision (ECCV), pages 767-783, 2018. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.469, + 0.872 + ], + "angle": 0, + "content": "[35] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5752-5761, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.903 + ], + "angle": 0, + "content": "[36] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[37] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J Davison. In-place scene labelling and understanding with implicit scene representation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15838-15847, 2021. 1, 2, 3, 5, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20815" + } + ] +] \ No newline at end of file diff --git a/2024/GSNeRF_ Generalizable Semantic Neural Radiance Fields with Enhanced 3D Scene Understanding/e670ccdd-f8c3-4912-9d6e-b86fa014dde2_origin.pdf b/2024/GSNeRF_ Generalizable Semantic Neural Radiance Fields with Enhanced 3D Scene Understanding/e670ccdd-f8c3-4912-9d6e-b86fa014dde2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6aa55013ef7cae51b6cd3dcc025be242b5c3ff4d --- /dev/null +++ b/2024/GSNeRF_ Generalizable Semantic Neural Radiance Fields with Enhanced 3D Scene Understanding/e670ccdd-f8c3-4912-9d6e-b86fa014dde2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:43df55ef14805cb15783fd7330da645f2b3056d185e31c1e24325a8950658d52 +size 860058 diff --git a/2024/GSNeRF_ Generalizable Semantic Neural Radiance Fields with Enhanced 3D Scene Understanding/full.md b/2024/GSNeRF_ Generalizable Semantic Neural Radiance Fields with Enhanced 3D Scene Understanding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..bf3e3497911de8ca8872b0a73fe0b5093a8d42e4 --- /dev/null +++ b/2024/GSNeRF_ Generalizable Semantic Neural Radiance Fields with Enhanced 3D Scene Understanding/full.md @@ -0,0 +1,381 @@ +# GSNeRF: Generalizable Semantic Neural Radiance Fields with Enhanced 3D Scene Understanding + +Zi-Ting Chou\*1, Sheng-Yu Huang\*1, I-Jieh Liu1, Yu-Chiang Frank Wang\*2, Graduate Institute of Communication Engineering, National Taiwan University NVIDIA, Taiwan {r11942101, f08942095, r11942087}@ntu.edu.tw frankwang@nvidia.com + +# Abstract + +Utilizing multi-view inputs to synthesize novel-view images, Neural Radiance Fields (NeRF) have emerged as a popular research topic in 3D vision. In this work, we introduce a Generalizable Semantic Neural Radiance Fields (GSNeRF), which uniquely takes image semantics into the synthesis process so that both novel view image and the associated semantic maps can be produced for unseen scenes. Our GSNeRF is composed of two stages: Semantic Geo-Reasoning and Depth-Guided Visual rendering. The former is able to observe multi-view image inputs to extract semantic and geometry features from a scene. Guided by the resulting image geometry information, the latter performs both image and semantic rendering with improved performances. Our experiments not only confirm that GSNeRF performs favorably against prior works on both novel-view image and semantic segmentation synthesis but the effectiveness of our sampling strategy for visual rendering is further verified. + +# 1. Introduction + +3D scene understanding plays a pivotal role in many vision-related tasks, including 3D reconstruction and 3D reasoning. The former domain focuses on low-level understanding, including developing efficient models that interpret and reconstruct observations from various data sources, such as RGB, RGBD, and physical sensors. In contrast, the latter emphasizes high-level understanding, like encoding, segmentation, and common sense knowledge learning. For real-world applications like robotic navigation or augmented reality, both these facets - reconstruction and reasoning - are crucial to enhance interactions with the real world [10]. Among the challenges in the reconstruction domain, novel view synthesis has consistently been a challenging objective. Being able to generate a novel view image based on a few existing views shows that a model com + +prehends the scene's geometry and possesses a foundational understanding akin to reconstruction. + +The Neural Radiance Field (NeRF) [1, 4, 8, 19-22, 28, 35] has recently risen as an exciting research area, providing a novel way to tackle the task of novel view synthesis. By encoding the density and emitted radiance at each spatial location, NeRF is able to compress a scene into a learnable model given several images and the corresponding camera poses of the scene. By incorporating a volumetric rendering skill, images of unseen camera views can be generated with convincing quality. However, NeRF primarily focuses on reconstructing the color information of novel views and understanding the associated high-level semantic information (e.g., semantic segmentation or object detection) remains a significant challenge. + +Recent works [9, 15, 25, 30, 31, 37] aim to amplify the high-level scene understanding ability of NeRF by integrating NeRF with semantic segmentation. By sharing the information between the semantic object class and their corresponding appearances, these two tasks are capable of benefiting from each other's insights [17]. Nevertheless, these semantic understanding methods stick closely to the original NeRF paradigm, focusing on exploiting the model to represent a specific scene. As mentioned in [17], such a strategy requires additional annotations of semantic segmentation maps when applied to a new scene, limiting real-world applicability and generalizability. + +To tackle this problem, generalizable NeRFs [3, 14, 18, 27, 32, 33, 36] have emerged as a promising solution. They adopt an on-the-fly approach for building a neural radiance field conditioned on extracted features from input images of different scenes rather than encoding the scene representation directly into the model. For instance, PixelNeRF [36] introduces the idea of conditioning a NeRF model with multi-view images across multiple scenes. This innovation boosts NeRF's ability to generalize to unseen scenes and avoids retraining for each individual scene. Yet, it leaves uncertainties in its application to semantic understanding. + +As a pioneer in integrating a generalizable NeRF with semantic segmentation capabilities, Semantic-Ray (S- + +
Methodgeneralized NeRFgeneralized segmentationsampling strategyefficient sampling
Semantic NeRF [37]--hierarchical✓
Panoptic Neural Fields [15]--uniform-
Panoptic Lifting [25]--uniform-
DM-NeRF [31]--hierarchical✓
NeSF [30]-✓hierarchical✓
Semantic Ray [17]✓✓hierarchical✓
Ours✓✓depth-guided✓
+ +Table 1. Comparisons of semantic-aware NeRFs (i.e., NeRFs, which jointly perform novel view synthesis and semantic segmentation). Note that NeSF only performs generalizable segmentation, yet its view synthesis remains restricted to the scene-specific training scheme. And, Semantic Ray requires ground truth depth as additional inputs for both training and inference. + +Ray) [17] made notable advancements by introducing a generalizable semantic field concept. By incorporating a cross-reprojection attention module, S-Ray aggregates information of multi-view images from all points sampled along a ray and predicts RGB values and semantic labels. Although this approach appears to capture the necessary contextual information for semantic segmentation, aggregating details of all points introduces noisy features that could interfere with segmentation tasks, as segmentation is mainly concerned with the first object class that the ray encounters. Additionally, their hierarchical sampling approach further increases additional computation and model complexity in view synthesis. Thus, a properly designed rendering strategy for a generalizable semantic field would be desirable. + +In this paper, we propose Generalizable Semantic Neural Raisance Fields (GSNeRF) to jointly tackle the problems of generalized novel view synthesis and semantic segmentation. Given a set of source images of a scene with corresponding camera views and a novel target camera view, our GSNeRF is able to derive visual features and depth map predictions of each source view, which are used to predict the depth map of the target view. With the depth map of the target view and properly derived visual features of source views, the novel view RGB image and semantic segmentation can be derived accordingly. Our GSNeRF consists of two key learning stages: Semantic Geo-Reasoning and Depth-Guided Visual Rendering. The former is to derive visual features and aggregate the predicted depth information of each source view to estimate the depth of the novel view, while the latter renders the RGB image and semantic segmentation map of the target view. We further design two distinct sampling strategies to minimize noisy features and enhance rendering efficiency. Through conducting extensive experiments on both real-world and synthetic datasets, we show that our method outperforms current state-of-the-art generalizable NeRF methods in novel view synthesis and semantic segmentation. + +The key contributions of our work are as follows: + +- We propose GSNeRF for jointly rendering novel view images and producing the associated semantic segmentation + +mask on unseen scenes. + +- The proposed Semantic Geo-Reasoning stage learns color, geometry, and semantic information of the input scene, introducing generalization ability of our GSNeRF. +- Based on the inferred geometry information, the introduced Depth-Guided Visual Rendering stage customizes two different sampling strategies according to the predicted target view depth map, so that image and semantic map rendering can be performed simultaneously. + +# 2. Related Work + +# 2.1. Neural Radiance Fields + +The Neural Radiance Field [20] has emerged as a widely embraced implicit representation by encoding a 3D scene within a neural network. Nonetheless, as mentioned in [7], the original NeRF method demands extensive training time, ranging from hours to days, and relies on dozens of multiview images as input. As a result, a large number of subsequent works [8, 21, 28, 35] are put forward to confront these issues. To shorten the training process, methods such as Instant NGP [21] and DVGO [28] opt for a balance between speed and memory, using hash encoding and voxel encoding to minimize the training time to minutes. On the other hand, to address the demand for multiple input views, some methods [7, 14, 23] introduce depth as additional supervision, a notable example being DS-NeRF [7]. DS-NeRF applies extra depth supervision to ensure that the NeRF accurately encodes a scene's geometry, thus enabling the construction of neural radiance fields with fewer images. This method shows how incorporating depth information can enhance the quality of novel view synthesis while reducing the number of required source images. + +# 2.2. Generalizable Novel View Synthesis + +Since NeRF directly encodes information of a scene into a neural network, it is only applicable to the training scene of interest. In response to this limitation, some methods [3, 14, 16, 18, 27, 29, 32, 36] propose on-the-fly construction of NeRF, allowing trained NeRFs applied for syn + +thesizing novel views in unseen scenes. The concept is to extract features from each source view image and condition the NeRF renderer on these scene-specific features, typically through [36] 2D CNN or [3, 14] cost volume techniques that aggregate features across different images. While these methods realize generalizable novel view synthesis, extensions of such NeRFs to high-level scene understanding tasks remain an open challenge. + +# 2.3. Multi-tasking NeRF + +Recently, several methods [9, 15, 17, 25, 30, 31, 37] seek to enhance NeRF with higher-level understanding abilities. Semantic NeRF [37] stands as the first method to integrate semantic segmentation with NeRF by designing two projection heads for predicting both semantics and color simultaneously. Although it achieves satisfactory performance and robustness on semantic segmentation tasks, it requires retraining for each scene. This implies a need for not only RGB images but also corresponding semantic segmentation maps to represent each new scene, hindering their generalizability. NeSF [30] has made an attempt to improve this aspect by proposing a half-generalizable method that includes a generalizable semantic segmentation module. That is, after training an original NeRF of each scene, they extract a 3D voxel-like grid density from the neural radiance field and train a generalizable 3D UNet to do a semantic segmentation task. However, this learning scheme still requires additional time to retrain NeRFs for novel scenes. + +S-Ray [17] advances further by proposing a generalized semantic neural field. Following the backbone of Neuray [18], they propose a cross-reprojection attention module. Given a ray emitted from the target viewpoint, they uniformly sample points along the ray, and hierarchically sample more points on the second forward pass according to the density predicted from the first pass. With the cross-reprojection module, they aggregate the contextual information on all sampled points. Despite achieving good results on both real-world and synthetic datasets, the sampling strategy is time-consuming. Also, aggregating all information along the entire ray can introduce noisy features that may hamper segmentation tasks, as segmentation primarily focuses on the first object class encountered by the ray. + +Contrarily, our proposed GSNeRF aims to predict the depth map of the target view and utilize it to sample points along the ray more efficiently. This unique strategy enables us to sample points focusing on the first encountered object and filter out unrelated noisy points for semantic segmentation prediction. In Table 1, we compare the characteristics of recent semantic-aware NeRFs with GSNeRF. + +# 3. Brief Review of Generalizable NeRFs + +For the sake of completeness, we provide a brief review of generalizable NeRFs [3, 14, 18, 29, 36]. Given $K$ multi + +view images $I_{1:K} = \{I_1, I_2, \dots, I_K\}$ with corresponding camera poses $\xi_{1:K} = \{\xi_1, \xi_2, \dots, \xi_K\}$ of a scene, and a target camera pose $\xi_T$ , the goal is to synthesize the novel view image $I_T$ . Rays are sampled across the scene from the target camera pose to render the novel view image. Along each ray, $r$ , $N_c$ points (coarse points) are sampled within the given near bound $b_l$ and far bound $b_u$ . Existing methods partition $[b_l, b_u]$ into $N_c$ bins uniformly and draw one sample within each bin as defined by: + +$$ +t _ {n} ^ {r} \sim \mathcal {U} \left[ b _ {l} + \frac {n - 1}{N _ {c}} \left(b _ {u} - b _ {l}\right), b _ {l} + \frac {n}{N _ {c}} \left(b _ {u} - b _ {l}\right) \right], \tag {1} +$$ + +where $t_n^r$ is the $n$ -th sample along $r$ . With the corresponding camera position $\sigma^r$ and viewing directions $\phi^r$ , the $n$ -th sampled point $x_n^r$ can be derived as $x_n^r = o^r + t_n^r\phi^r$ , with $x_n^r \in \mathbb{R}^3$ . Several methods [3, 18, 29, 36] also use hierarchical sampling, conducting a second forward pass to sample $N_f$ points (fine points) where the first $N_c$ uniformly sampled points show higher predicted density. + +With the above information, a generalized NeRF can be represented by a learnable continuous function $A_{\theta}$ , where $\theta$ denotes the learnable parameters. $A_{\theta}$ computes a volumetric density $\sigma_n^r \in \mathbb{R}^+$ and a view-dependent color representation $\mathbf{c}_n^r \in \mathbb{R}^3$ for each point $x_n^r$ . That is, + +$$ +\sigma_ {n} ^ {r}, \mathbf {c} _ {n} ^ {r} = A _ {\theta} \left(x _ {n} ^ {r}, \phi^ {r}, f _ {1: K} \left(w \left(x _ {n} ^ {r}\right)\right)\right), \tag {2} +$$ + +where $f_{1:K}$ are features of $I_{1:K}$ and $w(.)$ is the function that reprojects the 3D position $x_{n}^{r}$ onto the image plane with known camera poses. + +To this end, the color value $\mathbf{C}(r)$ of the pixel corresponding to the ray $r$ is formulated as: + +$$ +\mathbf {C} (r) = \sum_ {n = 1} ^ {N _ {c}} T _ {n} ^ {r} \alpha_ {n} ^ {r} \mathbf {c} _ {n} ^ {r}. \tag {3} +$$ + +$$ +T _ {n} ^ {r} = \exp \left(- \sum_ {i = 1} ^ {n - 1} \sigma_ {i} ^ {r} \delta_ {i} ^ {r}\right), \quad \alpha_ {n} ^ {r} = 1 - \exp \left(- \sigma_ {n} ^ {r} \delta_ {n} ^ {r}\right), \tag {4} +$$ + +where $T_{n}^{r}$ is the accumulated transmittance and $\delta_n^r = ||x_{n + 1}^r -x_n^r ||_2$ represents the distance between every two adjacent points along $r$ . + +As is the case with other NeRF methods, a generalized NeRF also optimizes the model based on the rendering loss between the rendered image and the GT image: + +$$ +\mathcal {L} _ {\text {i m a g e}} = \sum_ {r \in R} \| \mathbf {C} (r) - \hat {\mathbf {C}} (r) \| _ {2} ^ {2}, \tag {5} +$$ + +where $\hat{\mathbf{C}}(r)$ represents the GT color of the renderer image corresponding to a ray $r$ . + +Our method is built on top of generalized NeRF methods but with a more efficient sampling strategy that can deal with both semantic segmentation and image rendering. + +![](images/86ad533e35e25e87828d7c3f4f9d4c75fe009278798a4a6dfca7ae4bb27bfa51.jpg) +Figure 1. Overview of GSNeRF: including Semantic Geo-Reasoning and Depth-Guided Visual Rendering. Given K multi-view image $I_{1:K}$ of a scene, the Semantic Geo-Reasoner predicts the depth map $D_{1:K}$ for each source image, which is aggregated to estimate the target view depth map $D_T$ . With $D_T$ as key geometric guidance, we design Depth-Guided Visual Rendering to render target view image $I_T$ and semantic segmentation $S_T$ , by Volume Renderer $R_\theta$ and Semantic Renderer $P_\theta$ , respectively. + +![](images/69d8a02cbd00b997aafe84f1ab8d005e8f358daef319f3d1ed35fd2f6bd8bd34.jpg) +Figure 2. Semantic Geo-Reasoning. The Semantic Geo-Reasoner contains a shared Encoder $E_{\theta}$ , a shared Decoder $D_{\theta}$ and a cost-volume aggregator $C_{\theta}$ , producing image and volume features $(f_{1:K}^{I}$ and $f_{1:K}^{V})$ , with the associated depth maps $D_{1:K}$ . The depth map $D_{T}$ of the target view is estimated from $D_{1:K}$ . + +# 4. Method + +# 4.1. Problem Formulation and Model Overview + +We first define the setting and notations used in this paper. Given a scene with $K$ multi-view images $I_{1:K}$ and their corresponding camera poses $\xi_{1:K}$ along with a target camera pose as input, we aim to achieve novel-view synthesis and semantic segmentation simultaneously. During training, we synthesize both the novel view image $I_T$ and the semantic segmentation map $S_T$ from $\xi_T$ , given ground truth (GT) image $\hat{I}_T$ and semantic segmentation map $\hat{S}_T$ of the target view $\xi_T$ . Once the model completes training on a given set of training scenes, it is able to directly generalize to unseen scenes without finetuning. + +To achieve this goal, we propose Generalizable Seman + +tic Neural Radiance Fields (GSNeRF), as illustrated in Figure 1. GSNeRF consists of two key learning stages: Semantic Geo-Reasoning and Depth-Guided Visual Rendering. For Semantic Geo-Reasoning, we employ a Semantic Geo-Reasoner $G_{\theta}$ on each input source image $I_{k} \in \mathbb{R}^{H \times W \times 3}$ to create 2D feature $f_{k}^{I} \in \mathbb{R}^{H \times W \times d}$ , semantic feature $f_{k}^{S} \in \mathbb{R}^{H \times W \times d}$ , 3D volume feature $f_{k}^{V} \in \mathbb{R}^{H \times W \times L \times d}$ , and depth predictions $D_{k} \in \mathbb{R}^{H \times W}$ , where $H$ and $W$ represent the height and width of the image, $L$ is the depth sample number defined in [34] to determine the volume size for the 3D feature, and $d$ is the feature dimension. The source view depth maps are integrated to estimate the depth for the target view $D_{T}$ . As for Depth-Guided Visual Rendering, we conduct a unique sampling strategy based on $D_{T}$ to minimize noisy features and enhance rendering efficiency. The sampled points, along with the previously extracted feature, are respectively input into the volume renderer $R_{\theta}$ and semantic renderer $P_{\theta}$ , synthesizing the target view's image and semantic segmentation map. In the following subsection, we will explain the Semantic Geo-Reasoning and Depth-guided Visual Rendering process within our GSNeRF framework. + +# 4.2. Generalizable Semantic NeRF + +# 4.2.1 Semantic Geo-Reasoning + +Given $K$ multi-view source images $I_{1:K}$ of a scene, our Semantic Geo-Reasoner $G_{\theta}$ in Figure 1 is designed to achieve two goals. First, geometric clues are extracted from $G_{\theta}$ , comprising 3D Volume features $f_{1:K}^{V}$ and depth maps $D_{1:K}$ for each source image. In addition, 2D image features $f_{1:K}^{I}$ and semantic features $f_{1:K}^{S}$ for each source image can be jointly produced. Secondly, our $G_{\theta}$ is learned to predict the + +target view depth map $D_{T}$ , which is realized by deterministic aggregation and reprojection of all source view depth predictions $D_{1:K}$ . We now detail this learning stage. + +As depicted in Figure 2, our Semantic Geo-Reasoner consists of a shared 2D CNN encoder $E_{\theta}$ that extracts $f_{1:K}^{I}$ from every image $I_{1:K}$ . To construct geometry-aware features, a cost volume aggregator $C_{\theta}$ [6, 12, 13, 34] is applied to gather 2D features $f_{1:K}^{I}$ across different images. As a result, $C_{\theta}$ produces 3D feature $f_{1:K}^{V}$ , which are ultimately used to predict the depth map $D_{1:K}$ for every source view image. To obtain semantic-aware features for semantic prediction, a shared semantic feature decoder $D_{\theta}$ is additionally applied to derive 2D semantic features $f_{1:K}^{S}$ from $f_{1:K}^{I}$ . + +Since $D_{1:K}$ is a vital intermediate element that assists in estimating $D_T$ , which allows the subsequent depth-guided rendering process, we strive to ensure the precise prediction of $D_{1:K}$ , whether or not the ground truth depth images are accessible. To achieve this, we integrate ideas from DSNeRF [7] and GeoNeRF [14] to supervise our $G_{\theta}$ with either a depth-supervised loss or a self-supervised depth loss. If the ground-truth depth maps of $D_{1:K}$ are available, our depth-supervised loss $\mathcal{L}_D$ is given as: + +$$ +\mathcal {L} _ {D} = \frac {1}{K} \left(\sum_ {k = 1} ^ {K} \| D _ {k} - \hat {D} _ {k} \| _ {s 1}\right), \tag {6} +$$ + +Here, $\hat{D}_k$ represents the GT depth map of view $k$ , $\| \cdot \|_{s1}$ represents smooth L1 loss [11]. If $\hat{D}_k$ is unavailable, we use a self-supervised depth loss $\mathcal{L}_{ssl}$ to regularize our depth predictions by considering cross-view depth consistency between all source views. Following the self-supervised depth estimation approach of RC-MVSNet [2], $\mathcal{L}_{ssl}$ is defined as: + +$$ +\mathcal {L} _ {s s l} = \lambda_ {1} \mathcal {L} _ {R C} + \lambda_ {2} \mathcal {L} _ {S S I M} + \lambda_ {3} \mathcal {L} _ {S m o o t h}, \tag {7} +$$ + +where $\mathcal{L}_{RC}$ , $\mathcal{L}_{SSIM}$ , and $\mathcal{L}_{Smooth}$ represent reconstruction, structure similarity, and depth smoothness losses [2]. The hyper-parameters $\lambda_1$ , $\lambda_2$ , and $\lambda_3$ are also selected following [2] (see supplementary material for details). + +To this end, the target view depth map $D_{T}$ is calculated using $D_{1:K}$ through a deterministic algorithm. In simple terms, we directly project the pixels of each depth map into 3D space and then reproject them back to the target view. More details about the self-supervised depth loss and corresponding pseudo-code of target view depth estimation are available in the supplementary material. + +# 4.2.2 Depth-Guided Visual Rendering + +We now discuss how we perform volume and semantic rendering with the guidance of the predicted $D_{T}$ . That is, given a target view $\xi_{T}$ , a ray is sampled for each pixel on the image plane to render the corresponding RGB value for $I_{T}$ + +and the semantic segmentation output for $S_{T}$ via the Volume Renderer $R_{\theta}$ and the Semantic Renderer $P_{\theta}$ . + +Volume Rendering For volume rendering of $I_{T}$ , existing semantic-NeRF approaches [17, 30, 37] follow conventional NeRFs [20] by applying either uniform sampling or hierarchical sampling along each ray $r$ , as depicted in Table 1. However, both strategies require dense sampling to locate the surface of objects in the given scene, limiting the sampling efficiency. To promote the efficiency of our sampling process, we conduct a depth-guided sampling strategy by focusing on sampling points near the depth values of $D_{T}$ . More specifically, this is achieved by modifying the sampling strategy in Eq. 1 as follows: + +$$ +t _ {n} ^ {r} \sim \mathcal {G} (z, v ^ {2}), \quad v = \frac {\operatorname* {m i n} \left(\| z - b _ {u} \| , \| z - b _ {l} \|\right)}{3}, \tag {8} +$$ + +where $t_n^r$ is the $n$ -th sample along the ray $r$ , and $\mathcal{G}$ represents Gaussian (normal) distribution. Note that $z$ and $v$ represent the mean and standard deviation of the distribution, respectively, while the value of $z$ is determined as the predicted depth for the pixel $p$ that emits the ray $r$ (i.e., $z = D_T(p)$ ). By these sampling point values, we derive $N$ sampled points $x_{1:N}^r$ (i.e., $x_n^r = o^r + t_n^r\phi^r$ ) closer to the surface of the first encountered object by $r$ , avoiding considering noisy points that are far from the object surface. Please see the supplementary material for further details. + +We feed all these points alongside the 2D feature $f_{1:K}^{I}$ and 3D volume feature $f_{1:K}^{V}$ as additional conditions into the volume renderer $R_{\theta}$ to predict the novel view image of the scene. Specifically, we reproject the point $x_{n}^{r}$ onto each source view to get aggregated features from these condition features. Thus, we write the aggregated feature $f_{n,k} \in \mathbb{R}^{2d}$ of the $n$ -th sampled point $x_{n}^{r}$ from the $k$ -th view as: + +$$ +f _ {n, k} = \left[ f _ {k} ^ {I} \left(w \left(x _ {n} ^ {r}\right)\right); f _ {k} ^ {V} \left(x _ {n} ^ {r}\right) \right], \forall k \in [ 1: K ] \tag {9} +$$ + +where $w(\cdot)$ is the reprojection function used in Eq. 2 and $[\cdot; \cdot]$ denotes concatenation along feature dimension. We further define the global feature $f_{n,0}$ as guidance across all $K$ source views as: + +$$ +f _ {n, 0} = \left[ \operatorname {m e a n} \left(\left\{f _ {n, k} \right\} _ {k = 1} ^ {K}\right); \operatorname {v a r} \left(\left\{f _ {n, k} \right\} _ {k = 1} ^ {K}\right) \right]. \tag {10} +$$ + +Following GeoNeRF [14] methodology, the volume renderer $R_{\theta}$ (with learnable parameters $\theta$ ) predicts per point density and radiance based on the feature $f_{n} = \{f_{n,k}\}_{k=0}^{K} \in \mathbb{R}^{(K+1) \times 2d}$ mentioned above: + +$$ +\sigma_ {n} ^ {r}, \mathbf {c} _ {n} ^ {r} = R _ {\theta} \left(x _ {n} ^ {r}, \phi^ {r}, f _ {n}, M _ {n}\right), \tag {11} +$$ + +where $M_{n} = \{M_{n,k}\}_{k = 0}^{K}$ denotes the mask used to mask out unrelated features of a view. If the sampled point $x_{n}^{r}$ lies behind the predicted depth map $D_{k}$ or if the reprojected + +
Generalized methodGT DepthScanNet [5]Replica [26]
Train / TestmIoUacc. / class acc.PSNR↑SSIM↑LPIPS↓mIoUacc. / class acc.PSNR↑SSIM↑LPIPS↓
Neuray [18] + semhead✓ / ✓52.0967.81 / 61.9825.0183.0731.6344.3779.93 / 54.2526.2187.3730.93
GeoNeRF [14] + semhead✓ /53.7876.18 / 61.9032.5590.8812.6945.1281.67 / 52.3628.7088.9420.42
S-Ray [17]✓ / ✓55.5377.79 / 60.9225.1983.6630.9845.3080.48 / 53.7226.3888.1330.04
GSNeRF (Ours)✓ /58.3079.79 / 65.9331.3390.7312.5351.5283.41 / 61.2931.1692.4412.54
MVSNeRF [3] + semhead43.0666.90 / 53.6324.1480.3634.6330.2169.35 / 39.7523.6884.3728.08
GeoNeRF [14] + semhead45.1167.12 / 53.4430.7588.2716.4840.3574.63 / 49.1829.9291.1417.60
GNT [32] + semhead43.4962.06 / 51.8424.3982.3728.3638.1471.44 / 47.4624.5687.3120.97
Neuray [18] + semhead46.0966.39 / 53.7925.2484.3931.3340.9176.23 / 50.1527.8089.5523.68
S-Ray [17]47.6964.90 / 54.4725.1384.1830.4443.2777.63 / 52.8526.7788.5422.81
GSNeRF (Ours)52.2174.71 / 60.1431.4990.3913.8751.2383.06 / 61.1031.7192.8912.93
+ +Table 2. Quantitative results on ScanNet & Replica. Note that methods in the first four rows take GT depth as inputs or training supervision, while the methods in the last six rows do not observe GT depth during training/testing. + +location is outside the source image, $M_{n,k}$ is set to 0. However, the mask for the global feature $M_{n,0}$ is always set to 1. + +To this end, we apply the same volume rendering as in Eq. 3, predicting the target view Image $I_{T}$ . We also apply $\mathcal{L}_{image}$ as in Eq. 5 to supervise the rendered image quality. + +Semantic Rendering Semantic rendering of $S_{T}$ , on the other hand, is different from the above volume rendering process. Specifically, image rendering considers the color of both the surface and nearby objects because of possible semi-transparent or transparent objects. Conversely, semantic segmentation focuses only on the first object a ray touches. Therefore, for each ray $r$ , we sample only one surface point $x_{z}^{r}$ estimated by the target depth map for each ray $r$ to render its semantic value (i.e., $x_{z}^{r} = o^{r} + z\phi^{r}$ , where $z$ is the same depth value as mentioned in Eq. 8). More specifically, with the aggregated semantic feature $f_{z} = \{f_{k}^{S}(w(x_{z}^{r}))\}_{k=1}^{K} \in \mathbb{R}^{K \times d}$ , our semantic renderer $P_{\theta}$ with learnable parameters $\theta$ predicts the semantic logits $\mathbf{S}(r)$ of ray $r$ as follows: + +$$ +\mathbf {S} (r) = P _ {\theta} \left(x _ {z} ^ {r}, f _ {z} ^ {\prime}, M\right), \tag {12} +$$ + +where + +$$ +f _ {z} ^ {\prime} = f _ {z} + \left\{f _ {0} \right\} \in \mathbb {R} ^ {(K + 1) \times d}. \tag {13} +$$ + +Note that we follow volume rendering in Eq. 10 to define a global semantic feature $f_{0}$ as guidance across all views: + +$$ +f _ {0} = \left[ \operatorname {m e a n} \left(f _ {z}\right); \operatorname {v a r} \left(f _ {z}\right) \right]. \tag {14} +$$ + +As for the mask $M = \{M_k\}_{k=0}^K$ in Eqn. 12, it is similar to the one used in Eq. 11, where $M_k$ is to mask out the unrelated features of a view if the sample point $x_z^r$ is behind the corresponding depth value of $D_k$ . + +Eventually, we utilize a semantic loss to supervise the novel view semantic prediction. Here, we use a multi-class + +cross-entropy loss to reduce the semantic prediction error: + +$$ +\mathcal {L} _ {\text {s e m}} = \sum_ {r \in R} (\mathbf {S} (r) \log \hat {\mathbf {S}} (r)), \tag {15} +$$ + +where $\hat{S}(r)$ represent the ground truth segmentation score along the ray $r$ . + +# 4.3. Training and Inference + +# 4.3.1 Training + +At the training stage, the total loss is calculated by combining the losses that are mentioned above. The aggregated loss function for our model, when ground-truth depth is available, is given by: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {i m a g e}} + \mathcal {L} _ {D} + \lambda \mathcal {L} _ {\text {s e m}}. \tag {16} +$$ + +If the GT depth is not available, our model is trained by: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {i m a g e}} + \mathcal {L} _ {\text {s s l}} + \lambda \mathcal {L} _ {\text {s e m}}, \tag {17} +$$ + +where the coefficient $\lambda$ is set to 0.5 for simplicity. + +In essence, this optimization process enables our model to estimate the depth map of the target camera view accurately. Utilizing this depth information, the model is capable of efficiently sampling points to render new view images and create corresponding semantic segmentation maps. + +# 4.3.2 Inference + +At the inference stage, our GSNeRF is able to generalize to unseen scenes without the need for retraining from scratch. Due to the design of the conditioning renderer on scene-specific features $f^{I}, f^{V}$ , and $f^{S}$ , our model is able to build a semantic neural radiance field on the fly. Once we finish training our model on training scenes, we are able to directly infer novel view images and semantic segmentation maps on any scene given multi-view images and corresponding camera poses of that scene. + +![](images/37ce1e2a58a14d0a872f185a5689cd39934055bf3de1d4e018c4d999c06e603f.jpg) +S-Ray + +![](images/5c24ddee3802bce84eb65fba8426bedb3783ff72f10ee6f6023595d48cc8626f.jpg) +GSNeRF (Ours) + +![](images/adca324a7d265da1989de05806fac0a90dd19393c80b86694077ad0580ce1e09.jpg) +GT + +![](images/5fdff6ab6616ccf481068ac8634c2de626700c6198729f6cfc91e35d11f361cd.jpg) + +![](images/98165cff2b6872e4831bf67a033862d09337d4470474012290b0916d03e7ff23.jpg) + +![](images/5d058faf9ab067ba3527fa6e6d42c8f5f0bd00716bedd3f9cc26f90ae26c7c22.jpg) + +![](images/867343889916a3245a7372838ae59394e849320e15a1ead84509b9dffaf4d86b.jpg) + +![](images/93473372901ed42cfe50c05344e8833c22ca4ce590aa3ff3642eda09af6cd48e.jpg) + +![](images/03d2bafb9ac62658a427d26d38a641188c2668c3318a95ce07420fabe3b57165.jpg) + +![](images/a4fdd178526e91fdfef30c9d05f9ec231e39bc78574bca49fd3c260489a061ef.jpg) +Figure 3. Qualitative evaluation. We compare the visual quality of the rendered novel view images (the first three columns) and semantic segmentation maps (the last three columns) with S-Ray [17]. + +![](images/a5812dd1dde51ef924929c95f0772304723a2d93679fd7d5d56dd147d74de4e3.jpg) + +![](images/906b2e6bd9ef7a8ae4cef7a1e3fe13e7b8297ba28a420c5a5a5326ca1826cf5b.jpg) + +![](images/91a37656f6cc47463d6d7b7b66cbba290e2e94ee27f5fd4e6b26e5508a97c70f.jpg) +S-Ray + +![](images/f56b2e55d694112d99164465d9fafccc1026991198270e93b88f9bb776105f1f.jpg) +GSNeRF (Ours) + +![](images/f30efb61ab6e7221665b7a6fd565cc74898baadea65cbab7b2ee3181edb5ef42.jpg) +GT + +![](images/91a3d33427c6a72490ca7ecc468f9148c46a497a789fc3afe4ab09d4064bdc6d.jpg) + +![](images/71cfb630bbd1a8159080a0afd2b0ccfe22a451447cf64b575a10839dcc66ac71.jpg) + +![](images/a6def87bcfc5f5dfb8013bc30605d2a15f464414920e701037191d8c39f21578.jpg) + +![](images/a21c11f1df7afe9baa5f807204463ae112d13f77c1ab5b571302d7758e67a94b.jpg) + +![](images/d60d961c28cc91b4060875a072ec7ee0e85ee35c15f661654f96e5a9541b387f.jpg) + +![](images/fa1609161ee7cd5030a1f826aa903550acc0a2b75fc3cbb6e1df635a3d59b640.jpg) + +![](images/f39a52bde205f78178bee4a35a45c1cf592e62604fea8b2a136ac0dc588d7998.jpg) + +![](images/9b9894c1e7716630c3baa2909ca11040d9fecec218db523fc44cb709646acb97.jpg) + +![](images/59c725c48e516ed597bbcf9a2f02c52d55bac3ca74ec9faae489c1d8bde03654.jpg) + +# 5. Experiments + +# 5.1. Datasets + +To evaluate the effectiveness of our proposed method, we conduct experiments on both real-world and synthetic datasets. For real-world data, we use ScanNet [5], a large-scale, indoor RGB-D video dataset with over 2.5 million views from 1513 distinct scenes, including semantic annotations and camera poses. Following the setting presented in S-Ray [17], we train our model on 60 scenes and test its generalization on 10 new, unseen scenes. For synthetic data, we utilize Replica [26], a 3D reconstruction-based indoor dataset comprising 18 high-quality scenes with dense geometry, HDR textures, and semantic labels. Following the Semantic NeRF [37] supplied rendered data, we train our model on 6 distinct scenes across 12 video sequences and test on 2 novel scenes across 4 video sequences. Please see the supplementary material for details. + +# 5.2. Results and analysis + +# 5.2.1 Quantitative Results + +Table 2 compares our method against several baselines, in which S-Ray [17] serves as a primary baseline since it is the first to address semantic segmentation with NeRF in a generalized setting. We further conduct experiments on MVS-NeRF [3], GeoNeRF [14], GNT [32] and NeuRay [18] with + +the semantic head as described in [17]. It is worth noting that, even though S-Ray does not explicitly mention using a depth map of each source view image as input in their paper, their official implementation takes the depth maps $D_{1:K}$ as additional input along with $I_{1:K}$ during training and testing. Therefore, we follow their official implementation to reproduce S-Ray results. + +From Table 2, it can be seen that our model generalizes well to unseen scenes. To show that our GSNeRF is also capable of producing impressive results without depth supervision, we conduct an additional experiment in which we only use the loss function in Eq. 17 to train our GSNeRF in Table 2. We also conduct experiments of S-Ray, Neuray with only RGB images as input, and GeoNeRF without GT depth as supervision for a fair comparison. From Table 2, we see that even without access to GT depth and with the target depth information regularized in a self-supervised manner, our method still outperforms all the other baseline approaches. This suggests that our model can be trained without requiring depth map supervision, confirming the effectiveness and practicality of our proposed approach. We also note that, although GeoNeRF + semhead (with depth supervision) slightly outperforms our approach regarding PSNR and SSIM for the rendered RGB images on ScanNet, our method excels in all semantic segmentation metrics by approximately $5\%$ . More comprehensive discussions are presented in the supplementary materials. + +
ITST
ModelDepth-guidedrenderer PθDepth-guidedPSNRmIoU
A---28.4837.88
B✓--30.5641.72
C✓✓-30.8849.47
Ours✓✓✓31.4952.21
+ +Table 3. Ablation studies on GSNeRF. We verify the effectiveness of using depth-guided rendering for rendering $I_{T}$ and $S_{T}$ . Note that, without $P_{\theta}$ , we simply take a pre-trained 2D segmentor on $I_{T}$ to produce $S_{T}$ . + +![](images/1e7d1138daa3c614d285f06e394d0251d581c1fc5e5008ed8e7ccff8f9e949c6.jpg) +Figure 4. Sampling efficiency on ScanNet. Compared to SOTAs, our GSNeRF is able to achieve significantly improved rendering performance especially when the number of sampling points is small. Even with increasing numbers of sampling points, GSNeRF still performs favorably against existing models. + +# 5.2.2 Qualitative Results + +The qualitative results of our approach compared to S-Ray [17] are depicted in Fig. 3. This comparison illustrates that while the existing method can approximate the contours of objects, it is unable to capture the geometrical details, causing segmentation inconsistencies. In contrast, our GSNeRF accurately interprets scene geometry through our targeted view depth estimation design, allowing us to not only precisely segment objects but also render novel view images more realistically. This confirms that our depth-guided sampling technique effectively captures the correct geometry and that the sampled point features are appropriately aggregated from features of the Semantic Geo-Reasoner. + +# 5.2.3 Ablation Studies + +To further analyze the effectiveness of our designed modules, we conduct ablation studies on ScanNet, as shown in Table 3. Baseline model A employs our Semantic Geo-Reasoner $G_{\theta}$ and volume renderer $R_{\theta}$ to create the target view image $I_{T}$ , using uniform sampling instead of depth + +guided sampling $x_{1:N}$ . Also, a pre-trained 2D segmentor [24] is applied on the rendered image $I_T$ to produce $S_T$ without our semantic renderer $P_\theta$ . Model B enhances visual quality by replacing uniform sampling with depth-guided sampling. We further replace the 2D segmentor with our semantic renderer $P_\theta$ using uniform sampling for semantic rendering in model C, illustrating mutual benefits for both tasks. Our full model in the last row uses depth-guided sampling exclusively for surface points in the semantic renderer, achieving optimal results. This verifies the success of our proposed modules and sampling strategies. + +# 5.2.4 Sampling Efficiency + +Another benefit of our depth-guided sampling is that the performance of our GSNeRF is less sensitive to the number of sampled points along each ray. In Fig. 4, we conduct experiments of reducing the number of sampling points $N$ , showing PSNR of the synthesized novel view images as evaluation. We find that by employing depth-guided sampling, our method accurately preserves visual details with less noise (compared to Neuray and S-Ray with 128 sampled points along each ray). It can be seen that GeoNeRF + semhead shows a catastrophic drop in PSNR with fewer points, while ours leveraging depth-guided sampling maintains satisfactory visual quality even with a significantly reduced number of sampled points (e.g., when sampling only 4 points along a ray, the PSNR remains near 28). This result further attests to the success of our design in estimating target depth and implementing the corresponding sampling strategy with better novel view synthesis performance. + +# 6. Conclusion + +In this paper, we proposed Generalizable Semantic Neural Raidance Fields (GSNeRF) to achieve generalized novel view synthesis and semantic segmentation. Our GSNeRF is trained to derive visual features and perform depth map prediction of each source view so that the depth map for a novel target view can be estimated. With such target-view depth information observed, the associated RGB image and semantic segmentation can be jointly produced via depth-guide rendering. In our experiments, we quantitatively and qualitatively confirm that our GSNeRF performs favorably against existing generalizable semantic-aware NeRF methods on both real-world and synthetic datasets. + +Acknowledgement This work is supported in part by the National Science and Technology Council via grant NSCT112-2634-F-002-007 and the NTU Center of Data Intelligence: Technologies, Applications, and Systems via grant NTU-113L900902. We also thank the National Center for High-performance Computing (NCHC) for providing computational and storage resources. + +# References + +[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 1 +[2] Di Chang, Aljaž Božić, Tong Zhang, Qingsong Yan, Yingcong Chen, Sabine Susstrunk, and Matthias Nießner. Rcmvsnet: unsupervised multi-view stereo with neural rendering. In European Conference on Computer Vision, pages 665-680. Springer, 2022. 5 +[3] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14124-14133, 2021. 1, 2, 3, 6, 7 +[4] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 1 +[5] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 6, 7 +[6] Yuchao Dai, Zhidong Zhu, Zhibo Rao, and Bo Li. Mvs2: Deep unsupervised multi-view stereo with multi-view symmetry. In 2019 International Conference on 3D Vision (3DV), pages 1-8. IEEE, 2019. 5 +[7] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12882-12891, 2022. 2, 5 +[8] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 1, 2 +[9] Xiao Fu, Shangzhan Zhang, Tianrun Chen, Yichong Lu, Lanyun Zhu, Xiaowei Zhou, Andreas Geiger, and Yiyi Liao. Panoptic nerf: 3d-to-2d label transfer for panoptic urban scene segmentation. In 2022 International Conference on 3D Vision (3DV), pages 1-11. IEEE, 2022. 1, 3 +[10] Sourav Garg, Niko Sünderhauf, Feras Dayoub, Douglas Morrison, Akansel Cosgun, Gustavo Carneiro, Qi Wu, Tat-Jun Chin, Ian Reid, Stephen Gould, et al. Semantics for robotic mapping, perception and interaction: A survey. Foundations and Trends® in Robotics, 8(1-2):1-224, 2020. 1 +[11] Ross Girshick. Fast r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 1440-1448, 2015. 5 +[12] Xiaodong Gu, Zhiwen Fan, Siyu Zhu, Zuozhuo Dai, Feitong Tan, and Ping Tan. Cascade cost volume for high-resolution multi-view stereo and stereo matching. In Proceedings of + +the IEEE/CVF conference on computer vision and pattern recognition, pages 2495-2504, 2020. 5 +[13] Baichuan Huang, Hongwei Yi, Can Huang, Yijia He, Jingbin Liu, and Xiao Liu. M3vsnet: Unsupervised multi-metric multi-view stereo network. In 2021 IEEE International Conference on Image Processing (ICIP), pages 3163-3167. IEEE, 2021. 5 +[14] Mohammad Mahdi Johari, Yann Lepoittevin, and François Fleuret. Geonerf: Generalizing nerf with geometry priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18365-18375, 2022. 1, 2, 3, 5, 6, 7 +[15] Abhijit Kundu, Kyle Genova, Xiaqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12871-12881, 2022. 1, 2, 3 +[16] Haotong Lin, Sida Peng, Zhen Xu, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Efficient neural radiance fields for interactive free-viewpoint video. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 2 +[17] Fangfu Liu, Chubin Zhang, Yu Zheng, and Yueqi Duan. Semantic ray: Learning a generalizable semantic field with cross-reprojection attention. In CVPR, pages 17386-17396, 2023. 1, 2, 3, 5, 6, 7, 8 +[18] Yuan Liu, Sida Peng, Lingjie Liu, Qianqian Wang, Peng Wang, Christian Theobalt, Xiaowei Zhou, and Wenping Wang. Neural rays for occlusion-aware image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7824-7833, 2022. 1, 2, 3, 6, 7 +[19] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7210-7219, 2021. 1 +[20] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 5 +[21] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 2 +[22] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14335-14345, 2021. 1 +[23] Barbara Roessle, Jonathan T Barron, Ben Mildenhall, Pratul P Srinivasan, and Matthias Nießner. Dense depth priors for neural radiance fields from sparse input views. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12892-12901, 2022. 2 + +[24] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 8 +[25] Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulò, Norman Müller, Matthias Nießner, Angela Dai, and Peter Kontschieder. Panoptic lifting for 3d scene understanding with neural fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9043-9052, 2023. 1, 2, 3 +[26] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 6, 7 +[27] Mohammed Suhail, Carlos Esteves, Leonid Sigal, and Ameesh Makadia. Generalizable patch-based neural rendering. In European Conference on Computer Vision, pages 156-174. Springer, 2022. 1, 2 +[28] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 1, 2 +[29] Alex Trevithick and Bo Yang. Grf: Learning a general radiance field for 3d representation and rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15182-15192, 2021. 2, 3 +[30] Suhani Vora, Noha Radwan, Klaus Greff, Henning Meyer, Kyle Genova, Mehdi SM Sajjadi, Etienne Pot, Andrea Tagliasacchi, and Daniel Duckworth. Nesf: Neural semantic fields for generalizable semantic segmentation of 3d scenes. arXiv preprint arXiv:2111.13260, 2021. 1, 2, 3, 5 +[31] Bing Wang, Lu Chen, and Bo Yang. Dm-nerf: 3d scene geometry decomposition and manipulation from 2d images. arXiv preprint arXiv:2208.07227, 2022. 1, 2, 3 +[32] Peihao Wang, Xuxi Chen, Tianlong Chen, Subhashini Venugopalan, Zhangyang Wang, et al. Is attention all nerf needs? arXiv preprint arXiv:2207.13298, 2022. 1, 2, 6, 7 +[33] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690-4699, 2021. 1 +[34] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vision (ECCV), pages 767-783, 2018. 4, 5 +[35] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5752-5761, 2021. 1, 2 +[36] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. + +In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021. 1, 2, 3 +[37] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J Davison. In-place scene labelling and understanding with implicit scene representation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15838-15847, 2021. 1, 2, 3, 5, 7 \ No newline at end of file diff --git a/2024/GSNeRF_ Generalizable Semantic Neural Radiance Fields with Enhanced 3D Scene Understanding/images.zip b/2024/GSNeRF_ Generalizable Semantic Neural Radiance Fields with Enhanced 3D Scene Understanding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3098cd29d325cde1f3bdaba8159821875d14ec00 --- /dev/null +++ b/2024/GSNeRF_ Generalizable Semantic Neural Radiance Fields with Enhanced 3D Scene Understanding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ccec919561e4d52c358f0f60e094d78f724792c58beabf378ada66a99364d1d +size 444186 diff --git a/2024/GSNeRF_ Generalizable Semantic Neural Radiance Fields with Enhanced 3D Scene Understanding/layout.json b/2024/GSNeRF_ Generalizable Semantic Neural Radiance Fields with Enhanced 3D Scene Understanding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a7bbccb55c87c14b48e96f528524b04cb1b114ea --- /dev/null +++ b/2024/GSNeRF_ Generalizable Semantic Neural Radiance Fields with Enhanced 3D Scene Understanding/layout.json @@ -0,0 +1,11546 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 118, + 103, + 477, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 103, + 477, + 140 + ], + "spans": [ + { + "bbox": [ + 118, + 103, + 477, + 140 + ], + "type": "text", + "content": "GSNeRF: Generalizable Semantic Neural Radiance Fields with Enhanced 3D Scene Understanding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 90, + 160, + 504, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 160, + 504, + 204 + ], + "spans": [ + { + "bbox": [ + 90, + 160, + 504, + 204 + ], + "type": "text", + "content": "Zi-Ting Chou\\*1, Sheng-Yu Huang\\*1, I-Jieh Liu1, Yu-Chiang Frank Wang\\*2, Graduate Institute of Communication Engineering, National Taiwan University NVIDIA, Taiwan {r11942101, f08942095, r11942087}@ntu.edu.tw frankwang@nvidia.com" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 255, + 289, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 289, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 289, + 472 + ], + "type": "text", + "content": "Utilizing multi-view inputs to synthesize novel-view images, Neural Radiance Fields (NeRF) have emerged as a popular research topic in 3D vision. In this work, we introduce a Generalizable Semantic Neural Radiance Fields (GSNeRF), which uniquely takes image semantics into the synthesis process so that both novel view image and the associated semantic maps can be produced for unseen scenes. Our GSNeRF is composed of two stages: Semantic Geo-Reasoning and Depth-Guided Visual rendering. The former is able to observe multi-view image inputs to extract semantic and geometry features from a scene. Guided by the resulting image geometry information, the latter performs both image and semantic rendering with improved performances. Our experiments not only confirm that GSNeRF performs favorably against prior works on both novel-view image and semantic segmentation synthesis but the effectiveness of our sampling strategy for visual rendering is further verified." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 495, + 128, + 508 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 495, + 128, + 508 + ], + "spans": [ + { + "bbox": [ + 47, + 495, + 128, + 508 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 515, + 289, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 515, + 289, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 515, + 289, + 696 + ], + "type": "text", + "content": "3D scene understanding plays a pivotal role in many vision-related tasks, including 3D reconstruction and 3D reasoning. The former domain focuses on low-level understanding, including developing efficient models that interpret and reconstruct observations from various data sources, such as RGB, RGBD, and physical sensors. In contrast, the latter emphasizes high-level understanding, like encoding, segmentation, and common sense knowledge learning. For real-world applications like robotic navigation or augmented reality, both these facets - reconstruction and reasoning - are crucial to enhance interactions with the real world [10]. Among the challenges in the reconstruction domain, novel view synthesis has consistently been a challenging objective. Being able to generate a novel view image based on a few existing views shows that a model com" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 232, + 545, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 232, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 306, + 232, + 545, + 255 + ], + "type": "text", + "content": "prehends the scene's geometry and possesses a foundational understanding akin to reconstruction." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 256, + 546, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 256, + 546, + 412 + ], + "spans": [ + { + "bbox": [ + 304, + 256, + 546, + 412 + ], + "type": "text", + "content": "The Neural Radiance Field (NeRF) [1, 4, 8, 19-22, 28, 35] has recently risen as an exciting research area, providing a novel way to tackle the task of novel view synthesis. By encoding the density and emitted radiance at each spatial location, NeRF is able to compress a scene into a learnable model given several images and the corresponding camera poses of the scene. By incorporating a volumetric rendering skill, images of unseen camera views can be generated with convincing quality. However, NeRF primarily focuses on reconstructing the color information of novel views and understanding the associated high-level semantic information (e.g., semantic segmentation or object detection) remains a significant challenge." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 413, + 547, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 547, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 547, + 555 + ], + "type": "text", + "content": "Recent works [9, 15, 25, 30, 31, 37] aim to amplify the high-level scene understanding ability of NeRF by integrating NeRF with semantic segmentation. By sharing the information between the semantic object class and their corresponding appearances, these two tasks are capable of benefiting from each other's insights [17]. Nevertheless, these semantic understanding methods stick closely to the original NeRF paradigm, focusing on exploiting the model to represent a specific scene. As mentioned in [17], such a strategy requires additional annotations of semantic segmentation maps when applied to a new scene, limiting real-world applicability and generalizability." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 556, + 547, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 547, + 689 + ], + "type": "text", + "content": "To tackle this problem, generalizable NeRFs [3, 14, 18, 27, 32, 33, 36] have emerged as a promising solution. They adopt an on-the-fly approach for building a neural radiance field conditioned on extracted features from input images of different scenes rather than encoding the scene representation directly into the model. For instance, PixelNeRF [36] introduces the idea of conditioning a NeRF model with multi-view images across multiple scenes. This innovation boosts NeRF's ability to generalize to unseen scenes and avoids retraining for each individual scene. Yet, it leaves uncertainties in its application to semantic understanding." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "content": "As a pioneer in integrating a generalizable NeRF with semantic segmentation capabilities, Semantic-Ray (S-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 127, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 127, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 127, + 713 + ], + "type": "text", + "content": "*Equal Contribution" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20806" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 97, + 70, + 497, + 170 + ], + "blocks": [ + { + "bbox": [ + 97, + 70, + 497, + 170 + ], + "lines": [ + { + "bbox": [ + 97, + 70, + 497, + 170 + ], + "spans": [ + { + "bbox": [ + 97, + 70, + 497, + 170 + ], + "type": "table", + "html": "
Methodgeneralized NeRFgeneralized segmentationsampling strategyefficient sampling
Semantic NeRF [37]--hierarchical✓
Panoptic Neural Fields [15]--uniform-
Panoptic Lifting [25]--uniform-
DM-NeRF [31]--hierarchical✓
NeSF [30]-✓hierarchical✓
Semantic Ray [17]✓✓hierarchical✓
Ours✓✓depth-guided✓
", + "image_path": "ff3bb270e289c0ea0cd6fc84343f028d1326465d1125dad7fdbe9dabf0c3a872.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 178, + 547, + 212 + ], + "lines": [ + { + "bbox": [ + 46, + 178, + 547, + 212 + ], + "spans": [ + { + "bbox": [ + 46, + 178, + 547, + 212 + ], + "type": "text", + "content": "Table 1. Comparisons of semantic-aware NeRFs (i.e., NeRFs, which jointly perform novel view synthesis and semantic segmentation). Note that NeSF only performs generalizable segmentation, yet its view synthesis remains restricted to the scene-specific training scheme. And, Semantic Ray requires ground truth depth as additional inputs for both training and inference." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 230, + 289, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 230, + 289, + 397 + ], + "spans": [ + { + "bbox": [ + 46, + 230, + 289, + 397 + ], + "type": "text", + "content": "Ray) [17] made notable advancements by introducing a generalizable semantic field concept. By incorporating a cross-reprojection attention module, S-Ray aggregates information of multi-view images from all points sampled along a ray and predicts RGB values and semantic labels. Although this approach appears to capture the necessary contextual information for semantic segmentation, aggregating details of all points introduces noisy features that could interfere with segmentation tasks, as segmentation is mainly concerned with the first object class that the ray encounters. Additionally, their hierarchical sampling approach further increases additional computation and model complexity in view synthesis. Thus, a properly designed rendering strategy for a generalizable semantic field would be desirable." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 399, + 289, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 289, + 674 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 289, + 674 + ], + "type": "text", + "content": "In this paper, we propose Generalizable Semantic Neural Raisance Fields (GSNeRF) to jointly tackle the problems of generalized novel view synthesis and semantic segmentation. Given a set of source images of a scene with corresponding camera views and a novel target camera view, our GSNeRF is able to derive visual features and depth map predictions of each source view, which are used to predict the depth map of the target view. With the depth map of the target view and properly derived visual features of source views, the novel view RGB image and semantic segmentation can be derived accordingly. Our GSNeRF consists of two key learning stages: Semantic Geo-Reasoning and Depth-Guided Visual Rendering. The former is to derive visual features and aggregate the predicted depth information of each source view to estimate the depth of the novel view, while the latter renders the RGB image and semantic segmentation map of the target view. We further design two distinct sampling strategies to minimize noisy features and enhance rendering efficiency. Through conducting extensive experiments on both real-world and synthetic datasets, we show that our method outperforms current state-of-the-art generalizable NeRF methods in novel view synthesis and semantic segmentation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 676, + 259, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 676, + 259, + 687 + ], + "spans": [ + { + "bbox": [ + 58, + 676, + 259, + 687 + ], + "type": "text", + "content": "The key contributions of our work are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": "- We propose GSNeRF for jointly rendering novel view images and producing the associated semantic segmentation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 231, + 412, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 231, + 412, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 231, + 412, + 240 + ], + "type": "text", + "content": "mask on unseen scenes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 242, + 545, + 338 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 306, + 242, + 545, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 242, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 306, + 242, + 545, + 277 + ], + "type": "text", + "content": "- The proposed Semantic Geo-Reasoning stage learns color, geometry, and semantic information of the input scene, introducing generalization ability of our GSNeRF." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 278, + 545, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 278, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 306, + 278, + 545, + 338 + ], + "type": "text", + "content": "- Based on the inferred geometry information, the introduced Depth-Guided Visual Rendering stage customizes two different sampling strategies according to the predicted target view depth map, so that image and semantic map rendering can be performed simultaneously." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 349, + 392, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 349, + 392, + 361 + ], + "spans": [ + { + "bbox": [ + 306, + 349, + 392, + 361 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 369, + 439, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 369, + 439, + 380 + ], + "spans": [ + { + "bbox": [ + 306, + 369, + 439, + 380 + ], + "type": "text", + "content": "2.1. Neural Radiance Fields" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 387, + 545, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 387, + 545, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 387, + 545, + 626 + ], + "type": "text", + "content": "The Neural Radiance Field [20] has emerged as a widely embraced implicit representation by encoding a 3D scene within a neural network. Nonetheless, as mentioned in [7], the original NeRF method demands extensive training time, ranging from hours to days, and relies on dozens of multiview images as input. As a result, a large number of subsequent works [8, 21, 28, 35] are put forward to confront these issues. To shorten the training process, methods such as Instant NGP [21] and DVGO [28] opt for a balance between speed and memory, using hash encoding and voxel encoding to minimize the training time to minutes. On the other hand, to address the demand for multiple input views, some methods [7, 14, 23] introduce depth as additional supervision, a notable example being DS-NeRF [7]. DS-NeRF applies extra depth supervision to ensure that the NeRF accurately encodes a scene's geometry, thus enabling the construction of neural radiance fields with fewer images. This method shows how incorporating depth information can enhance the quality of novel view synthesis while reducing the number of required source images." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 635, + 495, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 495, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 495, + 647 + ], + "type": "text", + "content": "2.2. Generalizable Novel View Synthesis" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "type": "text", + "content": "Since NeRF directly encodes information of a scene into a neural network, it is only applicable to the training scene of interest. In response to this limitation, some methods [3, 14, 16, 18, 27, 29, 32, 36] propose on-the-fly construction of NeRF, allowing trained NeRFs applied for syn" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20807" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": "thesizing novel views in unseen scenes. The concept is to extract features from each source view image and condition the NeRF renderer on these scene-specific features, typically through [36] 2D CNN or [3, 14] cost volume techniques that aggregate features across different images. While these methods realize generalizable novel view synthesis, extensions of such NeRFs to high-level scene understanding tasks remain an open challenge." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 175, + 164, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 175, + 164, + 186 + ], + "spans": [ + { + "bbox": [ + 47, + 175, + 164, + 186 + ], + "type": "text", + "content": "2.3. Multi-tasking NeRF" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 193, + 289, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 193, + 289, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 193, + 289, + 407 + ], + "type": "text", + "content": "Recently, several methods [9, 15, 17, 25, 30, 31, 37] seek to enhance NeRF with higher-level understanding abilities. Semantic NeRF [37] stands as the first method to integrate semantic segmentation with NeRF by designing two projection heads for predicting both semantics and color simultaneously. Although it achieves satisfactory performance and robustness on semantic segmentation tasks, it requires retraining for each scene. This implies a need for not only RGB images but also corresponding semantic segmentation maps to represent each new scene, hindering their generalizability. NeSF [30] has made an attempt to improve this aspect by proposing a half-generalizable method that includes a generalizable semantic segmentation module. That is, after training an original NeRF of each scene, they extract a 3D voxel-like grid density from the neural radiance field and train a generalizable 3D UNet to do a semantic segmentation task. However, this learning scheme still requires additional time to retrain NeRFs for novel scenes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 408, + 287, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 408, + 287, + 575 + ], + "spans": [ + { + "bbox": [ + 46, + 408, + 287, + 575 + ], + "type": "text", + "content": "S-Ray [17] advances further by proposing a generalized semantic neural field. Following the backbone of Neuray [18], they propose a cross-reprojection attention module. Given a ray emitted from the target viewpoint, they uniformly sample points along the ray, and hierarchically sample more points on the second forward pass according to the density predicted from the first pass. With the cross-reprojection module, they aggregate the contextual information on all sampled points. Despite achieving good results on both real-world and synthetic datasets, the sampling strategy is time-consuming. Also, aggregating all information along the entire ray can introduce noisy features that may hamper segmentation tasks, as segmentation primarily focuses on the first object class encountered by the ray." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 576, + 288, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 576, + 288, + 659 + ], + "spans": [ + { + "bbox": [ + 46, + 576, + 288, + 659 + ], + "type": "text", + "content": "Contrarily, our proposed GSNeRF aims to predict the depth map of the target view and utilize it to sample points along the ray more efficiently. This unique strategy enables us to sample points focusing on the first encountered object and filter out unrelated noisy points for semantic segmentation prediction. In Table 1, we compare the characteristics of recent semantic-aware NeRFs with GSNeRF." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 669, + 254, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 254, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 254, + 682 + ], + "type": "text", + "content": "3. Brief Review of Generalizable NeRFs" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "content": "For the sake of completeness, we provide a brief review of generalizable NeRFs [3, 14, 18, 29, 36]. Given " + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "content": " multi" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": "view images " + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "I_{1:K} = \\{I_1, I_2, \\dots, I_K\\}" + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": " with corresponding camera poses " + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "\\xi_{1:K} = \\{\\xi_1, \\xi_2, \\dots, \\xi_K\\}" + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": " of a scene, and a target camera pose " + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "\\xi_T" + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": ", the goal is to synthesize the novel view image " + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "I_T" + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": ". Rays are sampled across the scene from the target camera pose to render the novel view image. Along each ray, " + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "N_c" + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": " points (coarse points) are sampled within the given near bound " + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "b_l" + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": " and far bound " + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "b_u" + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": ". Existing methods partition " + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "[b_l, b_u]" + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "N_c" + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": " bins uniformly and draw one sample within each bin as defined by:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 320, + 186, + 545, + 211 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 186, + 545, + 211 + ], + "spans": [ + { + "bbox": [ + 320, + 186, + 545, + 211 + ], + "type": "interline_equation", + "content": "t _ {n} ^ {r} \\sim \\mathcal {U} \\left[ b _ {l} + \\frac {n - 1}{N _ {c}} \\left(b _ {u} - b _ {l}\\right), b _ {l} + \\frac {n}{N _ {c}} \\left(b _ {u} - b _ {l}\\right) \\right], \\tag {1}", + "image_path": "ce19dfd8d8a2c75e8e2443b24c89db07eb1a0a002a87d9edfec9c630e130aa1d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 216, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 216, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 216, + 545, + 300 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 216, + 545, + 300 + ], + "type": "inline_equation", + "content": "t_n^r" + }, + { + "bbox": [ + 304, + 216, + 545, + 300 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 304, + 216, + 545, + 300 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 216, + 545, + 300 + ], + "type": "text", + "content": "-th sample along " + }, + { + "bbox": [ + 304, + 216, + 545, + 300 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 216, + 545, + 300 + ], + "type": "text", + "content": ". With the corresponding camera position " + }, + { + "bbox": [ + 304, + 216, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\sigma^r" + }, + { + "bbox": [ + 304, + 216, + 545, + 300 + ], + "type": "text", + "content": " and viewing directions " + }, + { + "bbox": [ + 304, + 216, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\phi^r" + }, + { + "bbox": [ + 304, + 216, + 545, + 300 + ], + "type": "text", + "content": ", the " + }, + { + "bbox": [ + 304, + 216, + 545, + 300 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 216, + 545, + 300 + ], + "type": "text", + "content": "-th sampled point " + }, + { + "bbox": [ + 304, + 216, + 545, + 300 + ], + "type": "inline_equation", + "content": "x_n^r" + }, + { + "bbox": [ + 304, + 216, + 545, + 300 + ], + "type": "text", + "content": " can be derived as " + }, + { + "bbox": [ + 304, + 216, + 545, + 300 + ], + "type": "inline_equation", + "content": "x_n^r = o^r + t_n^r\\phi^r" + }, + { + "bbox": [ + 304, + 216, + 545, + 300 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 304, + 216, + 545, + 300 + ], + "type": "inline_equation", + "content": "x_n^r \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 216, + 545, + 300 + ], + "type": "text", + "content": ". Several methods [3, 18, 29, 36] also use hierarchical sampling, conducting a second forward pass to sample " + }, + { + "bbox": [ + 304, + 216, + 545, + 300 + ], + "type": "inline_equation", + "content": "N_f" + }, + { + "bbox": [ + 304, + 216, + 545, + 300 + ], + "type": "text", + "content": " points (fine points) where the first " + }, + { + "bbox": [ + 304, + 216, + 545, + 300 + ], + "type": "inline_equation", + "content": "N_c" + }, + { + "bbox": [ + 304, + 216, + 545, + 300 + ], + "type": "text", + "content": " uniformly sampled points show higher predicted density." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 300, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 545, + 361 + ], + "type": "text", + "content": "With the above information, a generalized NeRF can be represented by a learnable continuous function " + }, + { + "bbox": [ + 304, + 300, + 545, + 361 + ], + "type": "inline_equation", + "content": "A_{\\theta}" + }, + { + "bbox": [ + 304, + 300, + 545, + 361 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 300, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 300, + 545, + 361 + ], + "type": "text", + "content": " denotes the learnable parameters. " + }, + { + "bbox": [ + 304, + 300, + 545, + 361 + ], + "type": "inline_equation", + "content": "A_{\\theta}" + }, + { + "bbox": [ + 304, + 300, + 545, + 361 + ], + "type": "text", + "content": " computes a volumetric density " + }, + { + "bbox": [ + 304, + 300, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\sigma_n^r \\in \\mathbb{R}^+" + }, + { + "bbox": [ + 304, + 300, + 545, + 361 + ], + "type": "text", + "content": " and a view-dependent color representation " + }, + { + "bbox": [ + 304, + 300, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_n^r \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 300, + 545, + 361 + ], + "type": "text", + "content": " for each point " + }, + { + "bbox": [ + 304, + 300, + 545, + 361 + ], + "type": "inline_equation", + "content": "x_n^r" + }, + { + "bbox": [ + 304, + 300, + 545, + 361 + ], + "type": "text", + "content": ". That is," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 351, + 366, + 545, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 366, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 351, + 366, + 545, + 380 + ], + "type": "interline_equation", + "content": "\\sigma_ {n} ^ {r}, \\mathbf {c} _ {n} ^ {r} = A _ {\\theta} \\left(x _ {n} ^ {r}, \\phi^ {r}, f _ {1: K} \\left(w \\left(x _ {n} ^ {r}\\right)\\right)\\right), \\tag {2}", + "image_path": "b4033eedfd3daab9d2e1ab17fb1fd21cf857988235aecc1cff838e2cbda425ee.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 387, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 387, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 387, + 545, + 422 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 387, + 545, + 422 + ], + "type": "inline_equation", + "content": "f_{1:K}" + }, + { + "bbox": [ + 304, + 387, + 545, + 422 + ], + "type": "text", + "content": " are features of " + }, + { + "bbox": [ + 304, + 387, + 545, + 422 + ], + "type": "inline_equation", + "content": "I_{1:K}" + }, + { + "bbox": [ + 304, + 387, + 545, + 422 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 387, + 545, + 422 + ], + "type": "inline_equation", + "content": "w(.)" + }, + { + "bbox": [ + 304, + 387, + 545, + 422 + ], + "type": "text", + "content": " is the function that reprojects the 3D position " + }, + { + "bbox": [ + 304, + 387, + 545, + 422 + ], + "type": "inline_equation", + "content": "x_{n}^{r}" + }, + { + "bbox": [ + 304, + 387, + 545, + 422 + ], + "type": "text", + "content": " onto the image plane with known camera poses." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 423, + 545, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 545, + 447 + ], + "type": "text", + "content": "To this end, the color value " + }, + { + "bbox": [ + 304, + 423, + 545, + 447 + ], + "type": "inline_equation", + "content": "\\mathbf{C}(r)" + }, + { + "bbox": [ + 304, + 423, + 545, + 447 + ], + "type": "text", + "content": " of the pixel corresponding to the ray " + }, + { + "bbox": [ + 304, + 423, + 545, + 447 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 423, + 545, + 447 + ], + "type": "text", + "content": " is formulated as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 378, + 453, + 545, + 486 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 453, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 378, + 453, + 545, + 486 + ], + "type": "interline_equation", + "content": "\\mathbf {C} (r) = \\sum_ {n = 1} ^ {N _ {c}} T _ {n} ^ {r} \\alpha_ {n} ^ {r} \\mathbf {c} _ {n} ^ {r}. \\tag {3}", + "image_path": "f8685bd2267f0ad5eb0e6b3cc64460c348a8010077171806c6281ea805431427.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 310, + 502, + 545, + 534 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 502, + 545, + 534 + ], + "spans": [ + { + "bbox": [ + 310, + 502, + 545, + 534 + ], + "type": "interline_equation", + "content": "T _ {n} ^ {r} = \\exp \\left(- \\sum_ {i = 1} ^ {n - 1} \\sigma_ {i} ^ {r} \\delta_ {i} ^ {r}\\right), \\quad \\alpha_ {n} ^ {r} = 1 - \\exp \\left(- \\sigma_ {n} ^ {r} \\delta_ {n} ^ {r}\\right), \\tag {4}", + "image_path": "ae7cbb438fb67edede19f23b7a8cdc4f8efad7136d7cc16a822e1fe50bb2a32f.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 540, + 545, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 540, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 304, + 540, + 545, + 576 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 540, + 545, + 576 + ], + "type": "inline_equation", + "content": "T_{n}^{r}" + }, + { + "bbox": [ + 304, + 540, + 545, + 576 + ], + "type": "text", + "content": " is the accumulated transmittance and " + }, + { + "bbox": [ + 304, + 540, + 545, + 576 + ], + "type": "inline_equation", + "content": "\\delta_n^r = ||x_{n + 1}^r -x_n^r ||_2" + }, + { + "bbox": [ + 304, + 540, + 545, + 576 + ], + "type": "text", + "content": " represents the distance between every two adjacent points along " + }, + { + "bbox": [ + 304, + 540, + 545, + 576 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 540, + 545, + 576 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 577, + 545, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 577, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 304, + 577, + 545, + 613 + ], + "type": "text", + "content": "As is the case with other NeRF methods, a generalized NeRF also optimizes the model based on the rendering loss between the rendered image and the GT image:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 358, + 619, + 545, + 644 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 619, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 358, + 619, + 545, + 644 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {i m a g e}} = \\sum_ {r \\in R} \\| \\mathbf {C} (r) - \\hat {\\mathbf {C}} (r) \\| _ {2} ^ {2}, \\tag {5}", + "image_path": "cd9d7b78ae440f069025cba6cc49940916caf8af67505aee7f354dc29154aa0a.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 653, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 653, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{C}}(r)" + }, + { + "bbox": [ + 304, + 653, + 545, + 677 + ], + "type": "text", + "content": " represents the GT color of the renderer image corresponding to a ray " + }, + { + "bbox": [ + 304, + 653, + 545, + 677 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 653, + 545, + 677 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": "Our method is built on top of generalized NeRF methods but with a more efficient sampling strategy that can deal with both semantic segmentation and image rendering." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20808" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 76, + 509, + 238 + ], + "blocks": [ + { + "bbox": [ + 77, + 76, + 509, + 238 + ], + "lines": [ + { + "bbox": [ + 77, + 76, + 509, + 238 + ], + "spans": [ + { + "bbox": [ + 77, + 76, + 509, + 238 + ], + "type": "image", + "image_path": "86ad533e35e25e87828d7c3f4f9d4c75fe009278798a4a6dfca7ae4bb27bfa51.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 251, + 547, + 298 + ], + "lines": [ + { + "bbox": [ + 46, + 251, + 547, + 298 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 547, + 298 + ], + "type": "text", + "content": "Figure 1. Overview of GSNeRF: including Semantic Geo-Reasoning and Depth-Guided Visual Rendering. Given K multi-view image " + }, + { + "bbox": [ + 46, + 251, + 547, + 298 + ], + "type": "inline_equation", + "content": "I_{1:K}" + }, + { + "bbox": [ + 46, + 251, + 547, + 298 + ], + "type": "text", + "content": " of a scene, the Semantic Geo-Reasoner predicts the depth map " + }, + { + "bbox": [ + 46, + 251, + 547, + 298 + ], + "type": "inline_equation", + "content": "D_{1:K}" + }, + { + "bbox": [ + 46, + 251, + 547, + 298 + ], + "type": "text", + "content": " for each source image, which is aggregated to estimate the target view depth map " + }, + { + "bbox": [ + 46, + 251, + 547, + 298 + ], + "type": "inline_equation", + "content": "D_T" + }, + { + "bbox": [ + 46, + 251, + 547, + 298 + ], + "type": "text", + "content": ". With " + }, + { + "bbox": [ + 46, + 251, + 547, + 298 + ], + "type": "inline_equation", + "content": "D_T" + }, + { + "bbox": [ + 46, + 251, + 547, + 298 + ], + "type": "text", + "content": " as key geometric guidance, we design Depth-Guided Visual Rendering to render target view image " + }, + { + "bbox": [ + 46, + 251, + 547, + 298 + ], + "type": "inline_equation", + "content": "I_T" + }, + { + "bbox": [ + 46, + 251, + 547, + 298 + ], + "type": "text", + "content": " and semantic segmentation " + }, + { + "bbox": [ + 46, + 251, + 547, + 298 + ], + "type": "inline_equation", + "content": "S_T" + }, + { + "bbox": [ + 46, + 251, + 547, + 298 + ], + "type": "text", + "content": ", by Volume Renderer " + }, + { + "bbox": [ + 46, + 251, + 547, + 298 + ], + "type": "inline_equation", + "content": "R_\\theta" + }, + { + "bbox": [ + 46, + 251, + 547, + 298 + ], + "type": "text", + "content": " and Semantic Renderer " + }, + { + "bbox": [ + 46, + 251, + 547, + 298 + ], + "type": "inline_equation", + "content": "P_\\theta" + }, + { + "bbox": [ + 46, + 251, + 547, + 298 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 312, + 289, + 448 + ], + "blocks": [ + { + "bbox": [ + 48, + 312, + 289, + 448 + ], + "lines": [ + { + "bbox": [ + 48, + 312, + 289, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 312, + 289, + 448 + ], + "type": "image", + "image_path": "69d8a02cbd00b997aafe84f1ab8d005e8f358daef319f3d1ed35fd2f6bd8bd34.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 455, + 287, + 511 + ], + "lines": [ + { + "bbox": [ + 46, + 455, + 287, + 511 + ], + "spans": [ + { + "bbox": [ + 46, + 455, + 287, + 511 + ], + "type": "text", + "content": "Figure 2. Semantic Geo-Reasoning. The Semantic Geo-Reasoner contains a shared Encoder " + }, + { + "bbox": [ + 46, + 455, + 287, + 511 + ], + "type": "inline_equation", + "content": "E_{\\theta}" + }, + { + "bbox": [ + 46, + 455, + 287, + 511 + ], + "type": "text", + "content": ", a shared Decoder " + }, + { + "bbox": [ + 46, + 455, + 287, + 511 + ], + "type": "inline_equation", + "content": "D_{\\theta}" + }, + { + "bbox": [ + 46, + 455, + 287, + 511 + ], + "type": "text", + "content": " and a cost-volume aggregator " + }, + { + "bbox": [ + 46, + 455, + 287, + 511 + ], + "type": "inline_equation", + "content": "C_{\\theta}" + }, + { + "bbox": [ + 46, + 455, + 287, + 511 + ], + "type": "text", + "content": ", producing image and volume features " + }, + { + "bbox": [ + 46, + 455, + 287, + 511 + ], + "type": "inline_equation", + "content": "(f_{1:K}^{I}" + }, + { + "bbox": [ + 46, + 455, + 287, + 511 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 455, + 287, + 511 + ], + "type": "inline_equation", + "content": "f_{1:K}^{V})" + }, + { + "bbox": [ + 46, + 455, + 287, + 511 + ], + "type": "text", + "content": ", with the associated depth maps " + }, + { + "bbox": [ + 46, + 455, + 287, + 511 + ], + "type": "inline_equation", + "content": "D_{1:K}" + }, + { + "bbox": [ + 46, + 455, + 287, + 511 + ], + "type": "text", + "content": ". The depth map " + }, + { + "bbox": [ + 46, + 455, + 287, + 511 + ], + "type": "inline_equation", + "content": "D_{T}" + }, + { + "bbox": [ + 46, + 455, + 287, + 511 + ], + "type": "text", + "content": " of the target view is estimated from " + }, + { + "bbox": [ + 46, + 455, + 287, + 511 + ], + "type": "inline_equation", + "content": "D_{1:K}" + }, + { + "bbox": [ + 46, + 455, + 287, + 511 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 530, + 104, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 530, + 104, + 542 + ], + "spans": [ + { + "bbox": [ + 47, + 530, + 104, + 542 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 550, + 272, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 550, + 272, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 550, + 272, + 562 + ], + "type": "text", + "content": "4.1. Problem Formulation and Model Overview" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 569, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 569, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 569, + 287, + 700 + ], + "type": "text", + "content": "We first define the setting and notations used in this paper. Given a scene with " + }, + { + "bbox": [ + 46, + 569, + 287, + 700 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 569, + 287, + 700 + ], + "type": "text", + "content": " multi-view images " + }, + { + "bbox": [ + 46, + 569, + 287, + 700 + ], + "type": "inline_equation", + "content": "I_{1:K}" + }, + { + "bbox": [ + 46, + 569, + 287, + 700 + ], + "type": "text", + "content": " and their corresponding camera poses " + }, + { + "bbox": [ + 46, + 569, + 287, + 700 + ], + "type": "inline_equation", + "content": "\\xi_{1:K}" + }, + { + "bbox": [ + 46, + 569, + 287, + 700 + ], + "type": "text", + "content": " along with a target camera pose as input, we aim to achieve novel-view synthesis and semantic segmentation simultaneously. During training, we synthesize both the novel view image " + }, + { + "bbox": [ + 46, + 569, + 287, + 700 + ], + "type": "inline_equation", + "content": "I_T" + }, + { + "bbox": [ + 46, + 569, + 287, + 700 + ], + "type": "text", + "content": " and the semantic segmentation map " + }, + { + "bbox": [ + 46, + 569, + 287, + 700 + ], + "type": "inline_equation", + "content": "S_T" + }, + { + "bbox": [ + 46, + 569, + 287, + 700 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 569, + 287, + 700 + ], + "type": "inline_equation", + "content": "\\xi_T" + }, + { + "bbox": [ + 46, + 569, + 287, + 700 + ], + "type": "text", + "content": ", given ground truth (GT) image " + }, + { + "bbox": [ + 46, + 569, + 287, + 700 + ], + "type": "inline_equation", + "content": "\\hat{I}_T" + }, + { + "bbox": [ + 46, + 569, + 287, + 700 + ], + "type": "text", + "content": " and semantic segmentation map " + }, + { + "bbox": [ + 46, + 569, + 287, + 700 + ], + "type": "inline_equation", + "content": "\\hat{S}_T" + }, + { + "bbox": [ + 46, + 569, + 287, + 700 + ], + "type": "text", + "content": " of the target view " + }, + { + "bbox": [ + 46, + 569, + 287, + 700 + ], + "type": "inline_equation", + "content": "\\xi_T" + }, + { + "bbox": [ + 46, + 569, + 287, + 700 + ], + "type": "text", + "content": ". Once the model completes training on a given set of training scenes, it is able to directly generalize to unseen scenes without finetuning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "content": "To achieve this goal, we propose Generalizable Seman" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 314, + 547, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 314, + 547, + 578 + ], + "spans": [ + { + "bbox": [ + 304, + 314, + 547, + 578 + ], + "type": "text", + "content": "tic Neural Radiance Fields (GSNeRF), as illustrated in Figure 1. GSNeRF consists of two key learning stages: Semantic Geo-Reasoning and Depth-Guided Visual Rendering. For Semantic Geo-Reasoning, we employ a Semantic Geo-Reasoner " + }, + { + "bbox": [ + 304, + 314, + 547, + 578 + ], + "type": "inline_equation", + "content": "G_{\\theta}" + }, + { + "bbox": [ + 304, + 314, + 547, + 578 + ], + "type": "text", + "content": " on each input source image " + }, + { + "bbox": [ + 304, + 314, + 547, + 578 + ], + "type": "inline_equation", + "content": "I_{k} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 304, + 314, + 547, + 578 + ], + "type": "text", + "content": " to create 2D feature " + }, + { + "bbox": [ + 304, + 314, + 547, + 578 + ], + "type": "inline_equation", + "content": "f_{k}^{I} \\in \\mathbb{R}^{H \\times W \\times d}" + }, + { + "bbox": [ + 304, + 314, + 547, + 578 + ], + "type": "text", + "content": ", semantic feature " + }, + { + "bbox": [ + 304, + 314, + 547, + 578 + ], + "type": "inline_equation", + "content": "f_{k}^{S} \\in \\mathbb{R}^{H \\times W \\times d}" + }, + { + "bbox": [ + 304, + 314, + 547, + 578 + ], + "type": "text", + "content": ", 3D volume feature " + }, + { + "bbox": [ + 304, + 314, + 547, + 578 + ], + "type": "inline_equation", + "content": "f_{k}^{V} \\in \\mathbb{R}^{H \\times W \\times L \\times d}" + }, + { + "bbox": [ + 304, + 314, + 547, + 578 + ], + "type": "text", + "content": ", and depth predictions " + }, + { + "bbox": [ + 304, + 314, + 547, + 578 + ], + "type": "inline_equation", + "content": "D_{k} \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 304, + 314, + 547, + 578 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 314, + 547, + 578 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 314, + 547, + 578 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 314, + 547, + 578 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 314, + 547, + 578 + ], + "type": "text", + "content": " represent the height and width of the image, " + }, + { + "bbox": [ + 304, + 314, + 547, + 578 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 314, + 547, + 578 + ], + "type": "text", + "content": " is the depth sample number defined in [34] to determine the volume size for the 3D feature, and " + }, + { + "bbox": [ + 304, + 314, + 547, + 578 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 314, + 547, + 578 + ], + "type": "text", + "content": " is the feature dimension. The source view depth maps are integrated to estimate the depth for the target view " + }, + { + "bbox": [ + 304, + 314, + 547, + 578 + ], + "type": "inline_equation", + "content": "D_{T}" + }, + { + "bbox": [ + 304, + 314, + 547, + 578 + ], + "type": "text", + "content": ". As for Depth-Guided Visual Rendering, we conduct a unique sampling strategy based on " + }, + { + "bbox": [ + 304, + 314, + 547, + 578 + ], + "type": "inline_equation", + "content": "D_{T}" + }, + { + "bbox": [ + 304, + 314, + 547, + 578 + ], + "type": "text", + "content": " to minimize noisy features and enhance rendering efficiency. The sampled points, along with the previously extracted feature, are respectively input into the volume renderer " + }, + { + "bbox": [ + 304, + 314, + 547, + 578 + ], + "type": "inline_equation", + "content": "R_{\\theta}" + }, + { + "bbox": [ + 304, + 314, + 547, + 578 + ], + "type": "text", + "content": " and semantic renderer " + }, + { + "bbox": [ + 304, + 314, + 547, + 578 + ], + "type": "inline_equation", + "content": "P_{\\theta}" + }, + { + "bbox": [ + 304, + 314, + 547, + 578 + ], + "type": "text", + "content": ", synthesizing the target view's image and semantic segmentation map. In the following subsection, we will explain the Semantic Geo-Reasoning and Depth-guided Visual Rendering process within our GSNeRF framework." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 589, + 470, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 589, + 470, + 601 + ], + "spans": [ + { + "bbox": [ + 305, + 589, + 470, + 601 + ], + "type": "text", + "content": "4.2. Generalizable Semantic NeRF" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 609, + 446, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 609, + 446, + 620 + ], + "spans": [ + { + "bbox": [ + 306, + 609, + 446, + 620 + ], + "type": "text", + "content": "4.2.1 Semantic Geo-Reasoning" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "text", + "content": "Given " + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "text", + "content": " multi-view source images " + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "inline_equation", + "content": "I_{1:K}" + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "text", + "content": " of a scene, our Semantic Geo-Reasoner " + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "inline_equation", + "content": "G_{\\theta}" + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "text", + "content": " in Figure 1 is designed to achieve two goals. First, geometric clues are extracted from " + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "inline_equation", + "content": "G_{\\theta}" + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "text", + "content": ", comprising 3D Volume features " + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "inline_equation", + "content": "f_{1:K}^{V}" + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "text", + "content": " and depth maps " + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "inline_equation", + "content": "D_{1:K}" + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "text", + "content": " for each source image. In addition, 2D image features " + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "inline_equation", + "content": "f_{1:K}^{I}" + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "text", + "content": " and semantic features " + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "inline_equation", + "content": "f_{1:K}^{S}" + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "text", + "content": " for each source image can be jointly produced. Secondly, our " + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "inline_equation", + "content": "G_{\\theta}" + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "text", + "content": " is learned to predict the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20809" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "target view depth map " + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "D_{T}" + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": ", which is realized by deterministic aggregation and reprojection of all source view depth predictions " + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "D_{1:K}" + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": ". We now detail this learning stage." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 109, + 286, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 109, + 286, + 229 + ], + "spans": [ + { + "bbox": [ + 46, + 109, + 286, + 229 + ], + "type": "text", + "content": "As depicted in Figure 2, our Semantic Geo-Reasoner consists of a shared 2D CNN encoder " + }, + { + "bbox": [ + 46, + 109, + 286, + 229 + ], + "type": "inline_equation", + "content": "E_{\\theta}" + }, + { + "bbox": [ + 46, + 109, + 286, + 229 + ], + "type": "text", + "content": " that extracts " + }, + { + "bbox": [ + 46, + 109, + 286, + 229 + ], + "type": "inline_equation", + "content": "f_{1:K}^{I}" + }, + { + "bbox": [ + 46, + 109, + 286, + 229 + ], + "type": "text", + "content": " from every image " + }, + { + "bbox": [ + 46, + 109, + 286, + 229 + ], + "type": "inline_equation", + "content": "I_{1:K}" + }, + { + "bbox": [ + 46, + 109, + 286, + 229 + ], + "type": "text", + "content": ". To construct geometry-aware features, a cost volume aggregator " + }, + { + "bbox": [ + 46, + 109, + 286, + 229 + ], + "type": "inline_equation", + "content": "C_{\\theta}" + }, + { + "bbox": [ + 46, + 109, + 286, + 229 + ], + "type": "text", + "content": " [6, 12, 13, 34] is applied to gather 2D features " + }, + { + "bbox": [ + 46, + 109, + 286, + 229 + ], + "type": "inline_equation", + "content": "f_{1:K}^{I}" + }, + { + "bbox": [ + 46, + 109, + 286, + 229 + ], + "type": "text", + "content": " across different images. As a result, " + }, + { + "bbox": [ + 46, + 109, + 286, + 229 + ], + "type": "inline_equation", + "content": "C_{\\theta}" + }, + { + "bbox": [ + 46, + 109, + 286, + 229 + ], + "type": "text", + "content": " produces 3D feature " + }, + { + "bbox": [ + 46, + 109, + 286, + 229 + ], + "type": "inline_equation", + "content": "f_{1:K}^{V}" + }, + { + "bbox": [ + 46, + 109, + 286, + 229 + ], + "type": "text", + "content": ", which are ultimately used to predict the depth map " + }, + { + "bbox": [ + 46, + 109, + 286, + 229 + ], + "type": "inline_equation", + "content": "D_{1:K}" + }, + { + "bbox": [ + 46, + 109, + 286, + 229 + ], + "type": "text", + "content": " for every source view image. To obtain semantic-aware features for semantic prediction, a shared semantic feature decoder " + }, + { + "bbox": [ + 46, + 109, + 286, + 229 + ], + "type": "inline_equation", + "content": "D_{\\theta}" + }, + { + "bbox": [ + 46, + 109, + 286, + 229 + ], + "type": "text", + "content": " is additionally applied to derive 2D semantic features " + }, + { + "bbox": [ + 46, + 109, + 286, + 229 + ], + "type": "inline_equation", + "content": "f_{1:K}^{S}" + }, + { + "bbox": [ + 46, + 109, + 286, + 229 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 109, + 286, + 229 + ], + "type": "inline_equation", + "content": "f_{1:K}^{I}" + }, + { + "bbox": [ + 46, + 109, + 286, + 229 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 229, + 286, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 229, + 286, + 336 + ], + "spans": [ + { + "bbox": [ + 46, + 229, + 286, + 336 + ], + "type": "text", + "content": "Since " + }, + { + "bbox": [ + 46, + 229, + 286, + 336 + ], + "type": "inline_equation", + "content": "D_{1:K}" + }, + { + "bbox": [ + 46, + 229, + 286, + 336 + ], + "type": "text", + "content": " is a vital intermediate element that assists in estimating " + }, + { + "bbox": [ + 46, + 229, + 286, + 336 + ], + "type": "inline_equation", + "content": "D_T" + }, + { + "bbox": [ + 46, + 229, + 286, + 336 + ], + "type": "text", + "content": ", which allows the subsequent depth-guided rendering process, we strive to ensure the precise prediction of " + }, + { + "bbox": [ + 46, + 229, + 286, + 336 + ], + "type": "inline_equation", + "content": "D_{1:K}" + }, + { + "bbox": [ + 46, + 229, + 286, + 336 + ], + "type": "text", + "content": ", whether or not the ground truth depth images are accessible. To achieve this, we integrate ideas from DSNeRF [7] and GeoNeRF [14] to supervise our " + }, + { + "bbox": [ + 46, + 229, + 286, + 336 + ], + "type": "inline_equation", + "content": "G_{\\theta}" + }, + { + "bbox": [ + 46, + 229, + 286, + 336 + ], + "type": "text", + "content": " with either a depth-supervised loss or a self-supervised depth loss. If the ground-truth depth maps of " + }, + { + "bbox": [ + 46, + 229, + 286, + 336 + ], + "type": "inline_equation", + "content": "D_{1:K}" + }, + { + "bbox": [ + 46, + 229, + 286, + 336 + ], + "type": "text", + "content": " are available, our depth-supervised loss " + }, + { + "bbox": [ + 46, + 229, + 286, + 336 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_D" + }, + { + "bbox": [ + 46, + 229, + 286, + 336 + ], + "type": "text", + "content": " is given as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 348, + 287, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 348, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 104, + 348, + 287, + 380 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {D} = \\frac {1}{K} \\left(\\sum_ {k = 1} ^ {K} \\| D _ {k} - \\hat {D} _ {k} \\| _ {s 1}\\right), \\tag {6}", + "image_path": "f0a8374a1320510fd763a94f83b1ee65f522daaa86736eb03cae85e5a87da646.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 388, + 286, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 286, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 286, + 462 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 46, + 388, + 286, + 462 + ], + "type": "inline_equation", + "content": "\\hat{D}_k" + }, + { + "bbox": [ + 46, + 388, + 286, + 462 + ], + "type": "text", + "content": " represents the GT depth map of view " + }, + { + "bbox": [ + 46, + 388, + 286, + 462 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 388, + 286, + 462 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 388, + 286, + 462 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|_{s1}" + }, + { + "bbox": [ + 46, + 388, + 286, + 462 + ], + "type": "text", + "content": " represents smooth L1 loss [11]. If " + }, + { + "bbox": [ + 46, + 388, + 286, + 462 + ], + "type": "inline_equation", + "content": "\\hat{D}_k" + }, + { + "bbox": [ + 46, + 388, + 286, + 462 + ], + "type": "text", + "content": " is unavailable, we use a self-supervised depth loss " + }, + { + "bbox": [ + 46, + 388, + 286, + 462 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ssl}" + }, + { + "bbox": [ + 46, + 388, + 286, + 462 + ], + "type": "text", + "content": " to regularize our depth predictions by considering cross-view depth consistency between all source views. Following the self-supervised depth estimation approach of RC-MVSNet [2], " + }, + { + "bbox": [ + 46, + 388, + 286, + 462 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ssl}" + }, + { + "bbox": [ + 46, + 388, + 286, + 462 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 78, + 475, + 287, + 488 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 475, + 287, + 488 + ], + "spans": [ + { + "bbox": [ + 78, + 475, + 287, + 488 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s s l} = \\lambda_ {1} \\mathcal {L} _ {R C} + \\lambda_ {2} \\mathcal {L} _ {S S I M} + \\lambda_ {3} \\mathcal {L} _ {S m o o t h}, \\tag {7}", + "image_path": "e34f19da8a25bd860f0071f93ab8ae31b03d296ee2ea0960c6bf4b30d4dd4f56.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 495, + 286, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 286, + 543 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 286, + 543 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 495, + 286, + 543 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{RC}" + }, + { + "bbox": [ + 46, + 495, + 286, + 543 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 495, + 286, + 543 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SSIM}" + }, + { + "bbox": [ + 46, + 495, + 286, + 543 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 495, + 286, + 543 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Smooth}" + }, + { + "bbox": [ + 46, + 495, + 286, + 543 + ], + "type": "text", + "content": " represent reconstruction, structure similarity, and depth smoothness losses [2]. The hyper-parameters " + }, + { + "bbox": [ + 46, + 495, + 286, + 543 + ], + "type": "inline_equation", + "content": "\\lambda_1" + }, + { + "bbox": [ + 46, + 495, + 286, + 543 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 495, + 286, + 543 + ], + "type": "inline_equation", + "content": "\\lambda_2" + }, + { + "bbox": [ + 46, + 495, + 286, + 543 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 495, + 286, + 543 + ], + "type": "inline_equation", + "content": "\\lambda_3" + }, + { + "bbox": [ + 46, + 495, + 286, + 543 + ], + "type": "text", + "content": " are also selected following [2] (see supplementary material for details)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 544, + 286, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 286, + 627 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 286, + 627 + ], + "type": "text", + "content": "To this end, the target view depth map " + }, + { + "bbox": [ + 46, + 544, + 286, + 627 + ], + "type": "inline_equation", + "content": "D_{T}" + }, + { + "bbox": [ + 46, + 544, + 286, + 627 + ], + "type": "text", + "content": " is calculated using " + }, + { + "bbox": [ + 46, + 544, + 286, + 627 + ], + "type": "inline_equation", + "content": "D_{1:K}" + }, + { + "bbox": [ + 46, + 544, + 286, + 627 + ], + "type": "text", + "content": " through a deterministic algorithm. In simple terms, we directly project the pixels of each depth map into 3D space and then reproject them back to the target view. More details about the self-supervised depth loss and corresponding pseudo-code of target view depth estimation are available in the supplementary material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 645, + 218, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 645, + 218, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 645, + 218, + 658 + ], + "type": "text", + "content": "4.2.2 Depth-Guided Visual Rendering" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "content": "We now discuss how we perform volume and semantic rendering with the guidance of the predicted " + }, + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "inline_equation", + "content": "D_{T}" + }, + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "content": ". That is, given a target view " + }, + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\xi_{T}" + }, + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "content": ", a ray is sampled for each pixel on the image plane to render the corresponding RGB value for " + }, + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "inline_equation", + "content": "I_{T}" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "type": "text", + "content": "and the semantic segmentation output for " + }, + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "type": "inline_equation", + "content": "S_{T}" + }, + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "type": "text", + "content": " via the Volume Renderer " + }, + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "type": "inline_equation", + "content": "R_{\\theta}" + }, + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "type": "text", + "content": " and the Semantic Renderer " + }, + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "type": "inline_equation", + "content": "P_{\\theta}" + }, + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 110, + 545, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 110, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 304, + 110, + 545, + 243 + ], + "type": "text", + "content": "Volume Rendering For volume rendering of " + }, + { + "bbox": [ + 304, + 110, + 545, + 243 + ], + "type": "inline_equation", + "content": "I_{T}" + }, + { + "bbox": [ + 304, + 110, + 545, + 243 + ], + "type": "text", + "content": ", existing semantic-NeRF approaches [17, 30, 37] follow conventional NeRFs [20] by applying either uniform sampling or hierarchical sampling along each ray " + }, + { + "bbox": [ + 304, + 110, + 545, + 243 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 110, + 545, + 243 + ], + "type": "text", + "content": ", as depicted in Table 1. However, both strategies require dense sampling to locate the surface of objects in the given scene, limiting the sampling efficiency. To promote the efficiency of our sampling process, we conduct a depth-guided sampling strategy by focusing on sampling points near the depth values of " + }, + { + "bbox": [ + 304, + 110, + 545, + 243 + ], + "type": "inline_equation", + "content": "D_{T}" + }, + { + "bbox": [ + 304, + 110, + 545, + 243 + ], + "type": "text", + "content": ". More specifically, this is achieved by modifying the sampling strategy in Eq. 1 as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 321, + 251, + 545, + 275 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 251, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 321, + 251, + 545, + 275 + ], + "type": "interline_equation", + "content": "t _ {n} ^ {r} \\sim \\mathcal {G} (z, v ^ {2}), \\quad v = \\frac {\\operatorname* {m i n} \\left(\\| z - b _ {u} \\| , \\| z - b _ {l} \\|\\right)}{3}, \\tag {8}", + "image_path": "b2af90473447e5ee5150c60629f683c49333eb7aef99227412a5ef2872888d97.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 282, + 545, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 282, + 545, + 402 + ], + "spans": [ + { + "bbox": [ + 304, + 282, + 545, + 402 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 282, + 545, + 402 + ], + "type": "inline_equation", + "content": "t_n^r" + }, + { + "bbox": [ + 304, + 282, + 545, + 402 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 304, + 282, + 545, + 402 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 282, + 545, + 402 + ], + "type": "text", + "content": "-th sample along the ray " + }, + { + "bbox": [ + 304, + 282, + 545, + 402 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 282, + 545, + 402 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 282, + 545, + 402 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 304, + 282, + 545, + 402 + ], + "type": "text", + "content": " represents Gaussian (normal) distribution. Note that " + }, + { + "bbox": [ + 304, + 282, + 545, + 402 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 282, + 545, + 402 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 282, + 545, + 402 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 304, + 282, + 545, + 402 + ], + "type": "text", + "content": " represent the mean and standard deviation of the distribution, respectively, while the value of " + }, + { + "bbox": [ + 304, + 282, + 545, + 402 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 282, + 545, + 402 + ], + "type": "text", + "content": " is determined as the predicted depth for the pixel " + }, + { + "bbox": [ + 304, + 282, + 545, + 402 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 282, + 545, + 402 + ], + "type": "text", + "content": " that emits the ray " + }, + { + "bbox": [ + 304, + 282, + 545, + 402 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 282, + 545, + 402 + ], + "type": "text", + "content": " (i.e., " + }, + { + "bbox": [ + 304, + 282, + 545, + 402 + ], + "type": "inline_equation", + "content": "z = D_T(p)" + }, + { + "bbox": [ + 304, + 282, + 545, + 402 + ], + "type": "text", + "content": "). By these sampling point values, we derive " + }, + { + "bbox": [ + 304, + 282, + 545, + 402 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 282, + 545, + 402 + ], + "type": "text", + "content": " sampled points " + }, + { + "bbox": [ + 304, + 282, + 545, + 402 + ], + "type": "inline_equation", + "content": "x_{1:N}^r" + }, + { + "bbox": [ + 304, + 282, + 545, + 402 + ], + "type": "text", + "content": " (i.e., " + }, + { + "bbox": [ + 304, + 282, + 545, + 402 + ], + "type": "inline_equation", + "content": "x_n^r = o^r + t_n^r\\phi^r" + }, + { + "bbox": [ + 304, + 282, + 545, + 402 + ], + "type": "text", + "content": ") closer to the surface of the first encountered object by " + }, + { + "bbox": [ + 304, + 282, + 545, + 402 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 282, + 545, + 402 + ], + "type": "text", + "content": ", avoiding considering noisy points that are far from the object surface. Please see the supplementary material for further details." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 402, + 545, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 402, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 545, + 486 + ], + "type": "text", + "content": "We feed all these points alongside the 2D feature " + }, + { + "bbox": [ + 304, + 402, + 545, + 486 + ], + "type": "inline_equation", + "content": "f_{1:K}^{I}" + }, + { + "bbox": [ + 304, + 402, + 545, + 486 + ], + "type": "text", + "content": " and 3D volume feature " + }, + { + "bbox": [ + 304, + 402, + 545, + 486 + ], + "type": "inline_equation", + "content": "f_{1:K}^{V}" + }, + { + "bbox": [ + 304, + 402, + 545, + 486 + ], + "type": "text", + "content": " as additional conditions into the volume renderer " + }, + { + "bbox": [ + 304, + 402, + 545, + 486 + ], + "type": "inline_equation", + "content": "R_{\\theta}" + }, + { + "bbox": [ + 304, + 402, + 545, + 486 + ], + "type": "text", + "content": " to predict the novel view image of the scene. Specifically, we reproject the point " + }, + { + "bbox": [ + 304, + 402, + 545, + 486 + ], + "type": "inline_equation", + "content": "x_{n}^{r}" + }, + { + "bbox": [ + 304, + 402, + 545, + 486 + ], + "type": "text", + "content": " onto each source view to get aggregated features from these condition features. Thus, we write the aggregated feature " + }, + { + "bbox": [ + 304, + 402, + 545, + 486 + ], + "type": "inline_equation", + "content": "f_{n,k} \\in \\mathbb{R}^{2d}" + }, + { + "bbox": [ + 304, + 402, + 545, + 486 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 304, + 402, + 545, + 486 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 402, + 545, + 486 + ], + "type": "text", + "content": "-th sampled point " + }, + { + "bbox": [ + 304, + 402, + 545, + 486 + ], + "type": "inline_equation", + "content": "x_{n}^{r}" + }, + { + "bbox": [ + 304, + 402, + 545, + 486 + ], + "type": "text", + "content": " from the " + }, + { + "bbox": [ + 304, + 402, + 545, + 486 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 402, + 545, + 486 + ], + "type": "text", + "content": "-th view as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 337, + 495, + 545, + 509 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 495, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 337, + 495, + 545, + 509 + ], + "type": "interline_equation", + "content": "f _ {n, k} = \\left[ f _ {k} ^ {I} \\left(w \\left(x _ {n} ^ {r}\\right)\\right); f _ {k} ^ {V} \\left(x _ {n} ^ {r}\\right) \\right], \\forall k \\in [ 1: K ] \\tag {9}", + "image_path": "44af905ea37fb635360d6b82aeaf5cf869b45e39481852e9b243f8b5d3a07d69.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 517, + 545, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 517, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 305, + 517, + 545, + 564 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 517, + 545, + 564 + ], + "type": "inline_equation", + "content": "w(\\cdot)" + }, + { + "bbox": [ + 305, + 517, + 545, + 564 + ], + "type": "text", + "content": " is the reprojection function used in Eq. 2 and " + }, + { + "bbox": [ + 305, + 517, + 545, + 564 + ], + "type": "inline_equation", + "content": "[\\cdot; \\cdot]" + }, + { + "bbox": [ + 305, + 517, + 545, + 564 + ], + "type": "text", + "content": " denotes concatenation along feature dimension. We further define the global feature " + }, + { + "bbox": [ + 305, + 517, + 545, + 564 + ], + "type": "inline_equation", + "content": "f_{n,0}" + }, + { + "bbox": [ + 305, + 517, + 545, + 564 + ], + "type": "text", + "content": " as guidance across all " + }, + { + "bbox": [ + 305, + 517, + 545, + 564 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 305, + 517, + 545, + 564 + ], + "type": "text", + "content": " source views as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 324, + 574, + 545, + 589 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 574, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 324, + 574, + 545, + 589 + ], + "type": "interline_equation", + "content": "f _ {n, 0} = \\left[ \\operatorname {m e a n} \\left(\\left\\{f _ {n, k} \\right\\} _ {k = 1} ^ {K}\\right); \\operatorname {v a r} \\left(\\left\\{f _ {n, k} \\right\\} _ {k = 1} ^ {K}\\right) \\right]. \\tag {10}", + "image_path": "57d4356cc1bd07ee493bbf9f7ad600b6eb0c1091540e57c962b7614c641994ea.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 597, + 545, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 597, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 305, + 597, + 545, + 646 + ], + "type": "text", + "content": "Following GeoNeRF [14] methodology, the volume renderer " + }, + { + "bbox": [ + 305, + 597, + 545, + 646 + ], + "type": "inline_equation", + "content": "R_{\\theta}" + }, + { + "bbox": [ + 305, + 597, + 545, + 646 + ], + "type": "text", + "content": " (with learnable parameters " + }, + { + "bbox": [ + 305, + 597, + 545, + 646 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 305, + 597, + 545, + 646 + ], + "type": "text", + "content": ") predicts per point density and radiance based on the feature " + }, + { + "bbox": [ + 305, + 597, + 545, + 646 + ], + "type": "inline_equation", + "content": "f_{n} = \\{f_{n,k}\\}_{k=0}^{K} \\in \\mathbb{R}^{(K+1) \\times 2d}" + }, + { + "bbox": [ + 305, + 597, + 545, + 646 + ], + "type": "text", + "content": " mentioned above:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 362, + 655, + 545, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 655, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 362, + 655, + 545, + 668 + ], + "type": "interline_equation", + "content": "\\sigma_ {n} ^ {r}, \\mathbf {c} _ {n} ^ {r} = R _ {\\theta} \\left(x _ {n} ^ {r}, \\phi^ {r}, f _ {n}, M _ {n}\\right), \\tag {11}", + "image_path": "ba8ff2ddc34927a52d2fbffa2cfd25c04e71b8c547cc681ca103b227033f7ef5.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "M_{n} = \\{M_{n,k}\\}_{k = 0}^{K}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " denotes the mask used to mask out unrelated features of a view. If the sampled point " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "x_{n}^{r}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " lies behind the predicted depth map " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "D_{k}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " or if the reprojected" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20810" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 547, + 232 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 547, + 232 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 547, + 232 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 547, + 232 + ], + "type": "table", + "html": "
Generalized methodGT DepthScanNet [5]Replica [26]
Train / TestmIoUacc. / class acc.PSNR↑SSIM↑LPIPS↓mIoUacc. / class acc.PSNR↑SSIM↑LPIPS↓
Neuray [18] + semhead✓ / ✓52.0967.81 / 61.9825.0183.0731.6344.3779.93 / 54.2526.2187.3730.93
GeoNeRF [14] + semhead✓ /53.7876.18 / 61.9032.5590.8812.6945.1281.67 / 52.3628.7088.9420.42
S-Ray [17]✓ / ✓55.5377.79 / 60.9225.1983.6630.9845.3080.48 / 53.7226.3888.1330.04
GSNeRF (Ours)✓ /58.3079.79 / 65.9331.3390.7312.5351.5283.41 / 61.2931.1692.4412.54
MVSNeRF [3] + semhead43.0666.90 / 53.6324.1480.3634.6330.2169.35 / 39.7523.6884.3728.08
GeoNeRF [14] + semhead45.1167.12 / 53.4430.7588.2716.4840.3574.63 / 49.1829.9291.1417.60
GNT [32] + semhead43.4962.06 / 51.8424.3982.3728.3638.1471.44 / 47.4624.5687.3120.97
Neuray [18] + semhead46.0966.39 / 53.7925.2484.3931.3340.9176.23 / 50.1527.8089.5523.68
S-Ray [17]47.6964.90 / 54.4725.1384.1830.4443.2777.63 / 52.8526.7788.5422.81
GSNeRF (Ours)52.2174.71 / 60.1431.4990.3913.8751.2383.06 / 61.1031.7192.8912.93
", + "image_path": "2448e76cd4d968d11a98d4ec4ce7424b5e4046e2f1fcfc146511197338dd10e5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 240, + 546, + 264 + ], + "lines": [ + { + "bbox": [ + 46, + 240, + 546, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 546, + 264 + ], + "type": "text", + "content": "Table 2. Quantitative results on ScanNet & Replica. Note that methods in the first four rows take GT depth as inputs or training supervision, while the methods in the last six rows do not observe GT depth during training/testing." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 277, + 287, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 287, + 301 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 287, + 301 + ], + "type": "text", + "content": "location is outside the source image, " + }, + { + "bbox": [ + 46, + 277, + 287, + 301 + ], + "type": "inline_equation", + "content": "M_{n,k}" + }, + { + "bbox": [ + 46, + 277, + 287, + 301 + ], + "type": "text", + "content": " is set to 0. However, the mask for the global feature " + }, + { + "bbox": [ + 46, + 277, + 287, + 301 + ], + "type": "inline_equation", + "content": "M_{n,0}" + }, + { + "bbox": [ + 46, + 277, + 287, + 301 + ], + "type": "text", + "content": " is always set to 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 301, + 287, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 301, + 287, + 339 + ], + "spans": [ + { + "bbox": [ + 47, + 301, + 287, + 339 + ], + "type": "text", + "content": "To this end, we apply the same volume rendering as in Eq. 3, predicting the target view Image " + }, + { + "bbox": [ + 47, + 301, + 287, + 339 + ], + "type": "inline_equation", + "content": "I_{T}" + }, + { + "bbox": [ + 47, + 301, + 287, + 339 + ], + "type": "text", + "content": ". We also apply " + }, + { + "bbox": [ + 47, + 301, + 287, + 339 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{image}" + }, + { + "bbox": [ + 47, + 301, + 287, + 339 + ], + "type": "text", + "content": " as in Eq. 5 to supervise the rendered image quality." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 354, + 289, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 289, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 289, + 521 + ], + "type": "text", + "content": "Semantic Rendering Semantic rendering of " + }, + { + "bbox": [ + 46, + 354, + 289, + 521 + ], + "type": "inline_equation", + "content": "S_{T}" + }, + { + "bbox": [ + 46, + 354, + 289, + 521 + ], + "type": "text", + "content": ", on the other hand, is different from the above volume rendering process. Specifically, image rendering considers the color of both the surface and nearby objects because of possible semi-transparent or transparent objects. Conversely, semantic segmentation focuses only on the first object a ray touches. Therefore, for each ray " + }, + { + "bbox": [ + 46, + 354, + 289, + 521 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 354, + 289, + 521 + ], + "type": "text", + "content": ", we sample only one surface point " + }, + { + "bbox": [ + 46, + 354, + 289, + 521 + ], + "type": "inline_equation", + "content": "x_{z}^{r}" + }, + { + "bbox": [ + 46, + 354, + 289, + 521 + ], + "type": "text", + "content": " estimated by the target depth map for each ray " + }, + { + "bbox": [ + 46, + 354, + 289, + 521 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 354, + 289, + 521 + ], + "type": "text", + "content": " to render its semantic value (i.e., " + }, + { + "bbox": [ + 46, + 354, + 289, + 521 + ], + "type": "inline_equation", + "content": "x_{z}^{r} = o^{r} + z\\phi^{r}" + }, + { + "bbox": [ + 46, + 354, + 289, + 521 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 354, + 289, + 521 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 354, + 289, + 521 + ], + "type": "text", + "content": " is the same depth value as mentioned in Eq. 8). More specifically, with the aggregated semantic feature " + }, + { + "bbox": [ + 46, + 354, + 289, + 521 + ], + "type": "inline_equation", + "content": "f_{z} = \\{f_{k}^{S}(w(x_{z}^{r}))\\}_{k=1}^{K} \\in \\mathbb{R}^{K \\times d}" + }, + { + "bbox": [ + 46, + 354, + 289, + 521 + ], + "type": "text", + "content": ", our semantic renderer " + }, + { + "bbox": [ + 46, + 354, + 289, + 521 + ], + "type": "inline_equation", + "content": "P_{\\theta}" + }, + { + "bbox": [ + 46, + 354, + 289, + 521 + ], + "type": "text", + "content": " with learnable parameters " + }, + { + "bbox": [ + 46, + 354, + 289, + 521 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 354, + 289, + 521 + ], + "type": "text", + "content": " predicts the semantic logits " + }, + { + "bbox": [ + 46, + 354, + 289, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{S}(r)" + }, + { + "bbox": [ + 46, + 354, + 289, + 521 + ], + "type": "text", + "content": " of ray " + }, + { + "bbox": [ + 46, + 354, + 289, + 521 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 354, + 289, + 521 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 118, + 531, + 287, + 545 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 531, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 118, + 531, + 287, + 545 + ], + "type": "interline_equation", + "content": "\\mathbf {S} (r) = P _ {\\theta} \\left(x _ {z} ^ {r}, f _ {z} ^ {\\prime}, M\\right), \\tag {12}", + "image_path": "e541116696b6f90d2a4e3081e709f729951054cfb24481bbca01b4b64664db1c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 554, + 75, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 554, + 75, + 564 + ], + "spans": [ + { + "bbox": [ + 47, + 554, + 75, + 564 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 564, + 287, + 579 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 564, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 287, + 579 + ], + "type": "interline_equation", + "content": "f _ {z} ^ {\\prime} = f _ {z} + \\left\\{f _ {0} \\right\\} \\in \\mathbb {R} ^ {(K + 1) \\times d}. \\tag {13}", + "image_path": "5445525ae1f9d7c6bf5beb9aa30fbee242a9714448ae73b087042da899584284.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 585, + 287, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 585, + 287, + 608 + ], + "spans": [ + { + "bbox": [ + 46, + 585, + 287, + 608 + ], + "type": "text", + "content": "Note that we follow volume rendering in Eq. 10 to define a global semantic feature " + }, + { + "bbox": [ + 46, + 585, + 287, + 608 + ], + "type": "inline_equation", + "content": "f_{0}" + }, + { + "bbox": [ + 46, + 585, + 287, + 608 + ], + "type": "text", + "content": " as guidance across all views:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 110, + 619, + 287, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 619, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 110, + 619, + 287, + 632 + ], + "type": "interline_equation", + "content": "f _ {0} = \\left[ \\operatorname {m e a n} \\left(f _ {z}\\right); \\operatorname {v a r} \\left(f _ {z}\\right) \\right]. \\tag {14}", + "image_path": "0a077babb231c0961003b9267bbb38a901c81ddcdb8cbf0dc6ce48b1aa170644.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 640, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 640, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 640, + 287, + 689 + ], + "type": "text", + "content": "As for the mask " + }, + { + "bbox": [ + 46, + 640, + 287, + 689 + ], + "type": "inline_equation", + "content": "M = \\{M_k\\}_{k=0}^K" + }, + { + "bbox": [ + 46, + 640, + 287, + 689 + ], + "type": "text", + "content": " in Eqn. 12, it is similar to the one used in Eq. 11, where " + }, + { + "bbox": [ + 46, + 640, + 287, + 689 + ], + "type": "inline_equation", + "content": "M_k" + }, + { + "bbox": [ + 46, + 640, + 287, + 689 + ], + "type": "text", + "content": " is to mask out the unrelated features of a view if the sample point " + }, + { + "bbox": [ + 46, + 640, + 287, + 689 + ], + "type": "inline_equation", + "content": "x_z^r" + }, + { + "bbox": [ + 46, + 640, + 287, + 689 + ], + "type": "text", + "content": " is behind the corresponding depth value of " + }, + { + "bbox": [ + 46, + 640, + 287, + 689 + ], + "type": "inline_equation", + "content": "D_k" + }, + { + "bbox": [ + 46, + 640, + 287, + 689 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Eventually, we utilize a semantic loss to supervise the novel view semantic prediction. Here, we use a multi-class" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 278, + 541, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 278, + 541, + 290 + ], + "spans": [ + { + "bbox": [ + 305, + 278, + 541, + 290 + ], + "type": "text", + "content": "cross-entropy loss to reduce the semantic prediction error:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 368, + 295, + 545, + 320 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 295, + 545, + 320 + ], + "spans": [ + { + "bbox": [ + 368, + 295, + 545, + 320 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s e m}} = \\sum_ {r \\in R} (\\mathbf {S} (r) \\log \\hat {\\mathbf {S}} (r)), \\tag {15}", + "image_path": "5173a3dc9ba3c5ad0689b984cbce84963772715efb845c1162aba8ba27344e1f.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 327, + 545, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 327, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 305, + 327, + 545, + 352 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 327, + 545, + 352 + ], + "type": "inline_equation", + "content": "\\hat{S}(r)" + }, + { + "bbox": [ + 305, + 327, + 545, + 352 + ], + "type": "text", + "content": " represent the ground truth segmentation score along the ray " + }, + { + "bbox": [ + 305, + 327, + 545, + 352 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 305, + 327, + 545, + 352 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 357, + 437, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 357, + 437, + 369 + ], + "spans": [ + { + "bbox": [ + 305, + 357, + 437, + 369 + ], + "type": "text", + "content": "4.3. Training and Inference" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 375, + 377, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 375, + 377, + 387 + ], + "spans": [ + { + "bbox": [ + 306, + 375, + 377, + 387 + ], + "type": "text", + "content": "4.3.1 Training" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 394, + 545, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 394, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 305, + 394, + 545, + 441 + ], + "type": "text", + "content": "At the training stage, the total loss is calculated by combining the losses that are mentioned above. The aggregated loss function for our model, when ground-truth depth is available, is given by:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 365, + 448, + 545, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 448, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 365, + 448, + 545, + 460 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {i m a g e}} + \\mathcal {L} _ {D} + \\lambda \\mathcal {L} _ {\\text {s e m}}. \\tag {16}", + "image_path": "2ca9f1ee28ed74eecd70fcda4a6bbc0a4ba4fad5605f1ff7858207e09008ce3d.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 465, + 534, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 465, + 534, + 478 + ], + "spans": [ + { + "bbox": [ + 306, + 465, + 534, + 478 + ], + "type": "text", + "content": "If the GT depth is not available, our model is trained by:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 363, + 483, + 545, + 496 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 483, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 363, + 483, + 545, + 496 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {i m a g e}} + \\mathcal {L} _ {\\text {s s l}} + \\lambda \\mathcal {L} _ {\\text {s e m}}, \\tag {17}", + "image_path": "6b111b949c005c13811f59da60c0f521c4a942d820f09c58d26f753fbba538e8.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 502, + 507, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 502, + 507, + 514 + ], + "spans": [ + { + "bbox": [ + 305, + 502, + 507, + 514 + ], + "type": "text", + "content": "where the coefficient " + }, + { + "bbox": [ + 305, + 502, + 507, + 514 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 305, + 502, + 507, + 514 + ], + "type": "text", + "content": " is set to 0.5 for simplicity." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 514, + 545, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 514, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 304, + 514, + 545, + 574 + ], + "type": "text", + "content": "In essence, this optimization process enables our model to estimate the depth map of the target camera view accurately. Utilizing this depth information, the model is capable of efficiently sampling points to render new view images and create corresponding semantic segmentation maps." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 586, + 380, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 380, + 597 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 380, + 597 + ], + "type": "text", + "content": "4.3.2 Inference" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "At the inference stage, our GSNeRF is able to generalize to unseen scenes without the need for retraining from scratch. Due to the design of the conditioning renderer on scene-specific features " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "f^{I}, f^{V}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "f^{S}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": ", our model is able to build a semantic neural radiance field on the fly. Once we finish training our model on training scenes, we are able to directly infer novel view images and semantic segmentation maps on any scene given multi-view images and corresponding camera poses of that scene." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20811" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 81, + 131, + 141 + ], + "blocks": [ + { + "bbox": [ + 77, + 70, + 103, + 79 + ], + "lines": [ + { + "bbox": [ + 77, + 70, + 103, + 79 + ], + "spans": [ + { + "bbox": [ + 77, + 70, + 103, + 79 + ], + "type": "text", + "content": "S-Ray" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 81, + 131, + 141 + ], + "lines": [ + { + "bbox": [ + 50, + 81, + 131, + 141 + ], + "spans": [ + { + "bbox": [ + 50, + 81, + 131, + 141 + ], + "type": "image", + "image_path": "37ce1e2a58a14d0a872f185a5689cd39934055bf3de1d4e018c4d999c06e603f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 133, + 80, + 214, + 141 + ], + "blocks": [ + { + "bbox": [ + 140, + 70, + 206, + 80 + ], + "lines": [ + { + "bbox": [ + 140, + 70, + 206, + 80 + ], + "spans": [ + { + "bbox": [ + 140, + 70, + 206, + 80 + ], + "type": "text", + "content": "GSNeRF (Ours)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 133, + 80, + 214, + 141 + ], + "lines": [ + { + "bbox": [ + 133, + 80, + 214, + 141 + ], + "spans": [ + { + "bbox": [ + 133, + 80, + 214, + 141 + ], + "type": "image", + "image_path": "5c24ddee3802bce84eb65fba8426bedb3783ff72f10ee6f6023595d48cc8626f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 215, + 80, + 296, + 141 + ], + "blocks": [ + { + "bbox": [ + 248, + 70, + 263, + 79 + ], + "lines": [ + { + "bbox": [ + 248, + 70, + 263, + 79 + ], + "spans": [ + { + "bbox": [ + 248, + 70, + 263, + 79 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 215, + 80, + 296, + 141 + ], + "lines": [ + { + "bbox": [ + 215, + 80, + 296, + 141 + ], + "spans": [ + { + "bbox": [ + 215, + 80, + 296, + 141 + ], + "type": "image", + "image_path": "adca324a7d265da1989de05806fac0a90dd19393c80b86694077ad0580ce1e09.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 50, + 143, + 130, + 204 + ], + "blocks": [ + { + "bbox": [ + 50, + 143, + 130, + 204 + ], + "lines": [ + { + "bbox": [ + 50, + 143, + 130, + 204 + ], + "spans": [ + { + "bbox": [ + 50, + 143, + 130, + 204 + ], + "type": "image", + "image_path": "5fdff6ab6616ccf481068ac8634c2de626700c6198729f6cfc91e35d11f361cd.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 133, + 143, + 213, + 204 + ], + "blocks": [ + { + "bbox": [ + 133, + 143, + 213, + 204 + ], + "lines": [ + { + "bbox": [ + 133, + 143, + 213, + 204 + ], + "spans": [ + { + "bbox": [ + 133, + 143, + 213, + 204 + ], + "type": "image", + "image_path": "98165cff2b6872e4831bf67a033862d09337d4470474012290b0916d03e7ff23.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 215, + 143, + 296, + 204 + ], + "blocks": [ + { + "bbox": [ + 215, + 143, + 296, + 204 + ], + "lines": [ + { + "bbox": [ + 215, + 143, + 296, + 204 + ], + "spans": [ + { + "bbox": [ + 215, + 143, + 296, + 204 + ], + "type": "image", + "image_path": "5d058faf9ab067ba3527fa6e6d42c8f5f0bd00716bedd3f9cc26f90ae26c7c22.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 50, + 205, + 131, + 266 + ], + "blocks": [ + { + "bbox": [ + 50, + 205, + 131, + 266 + ], + "lines": [ + { + "bbox": [ + 50, + 205, + 131, + 266 + ], + "spans": [ + { + "bbox": [ + 50, + 205, + 131, + 266 + ], + "type": "image", + "image_path": "867343889916a3245a7372838ae59394e849320e15a1ead84509b9dffaf4d86b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 133, + 205, + 214, + 266 + ], + "blocks": [ + { + "bbox": [ + 133, + 205, + 214, + 266 + ], + "lines": [ + { + "bbox": [ + 133, + 205, + 214, + 266 + ], + "spans": [ + { + "bbox": [ + 133, + 205, + 214, + 266 + ], + "type": "image", + "image_path": "93473372901ed42cfe50c05344e8833c22ca4ce590aa3ff3642eda09af6cd48e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 215, + 205, + 296, + 266 + ], + "blocks": [ + { + "bbox": [ + 215, + 205, + 296, + 266 + ], + "lines": [ + { + "bbox": [ + 215, + 205, + 296, + 266 + ], + "spans": [ + { + "bbox": [ + 215, + 205, + 296, + 266 + ], + "type": "image", + "image_path": "03d2bafb9ac62658a427d26d38a641188c2668c3318a95ce07420fabe3b57165.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 50, + 268, + 130, + 329 + ], + "blocks": [ + { + "bbox": [ + 50, + 268, + 130, + 329 + ], + "lines": [ + { + "bbox": [ + 50, + 268, + 130, + 329 + ], + "spans": [ + { + "bbox": [ + 50, + 268, + 130, + 329 + ], + "type": "image", + "image_path": "a4fdd178526e91fdfef30c9d05f9ec231e39bc78574bca49fd3c260489a061ef.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 340, + 546, + 363 + ], + "lines": [ + { + "bbox": [ + 46, + 340, + 546, + 363 + ], + "spans": [ + { + "bbox": [ + 46, + 340, + 546, + 363 + ], + "type": "text", + "content": "Figure 3. Qualitative evaluation. We compare the visual quality of the rendered novel view images (the first three columns) and semantic segmentation maps (the last three columns) with S-Ray [17]." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 133, + 268, + 213, + 329 + ], + "blocks": [ + { + "bbox": [ + 133, + 268, + 213, + 329 + ], + "lines": [ + { + "bbox": [ + 133, + 268, + 213, + 329 + ], + "spans": [ + { + "bbox": [ + 133, + 268, + 213, + 329 + ], + "type": "image", + "image_path": "a5812dd1dde51ef924929c95f0772304723a2d93679fd7d5d56dd147d74de4e3.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 215, + 268, + 296, + 329 + ], + "blocks": [ + { + "bbox": [ + 215, + 268, + 296, + 329 + ], + "lines": [ + { + "bbox": [ + 215, + 268, + 296, + 329 + ], + "spans": [ + { + "bbox": [ + 215, + 268, + 296, + 329 + ], + "type": "image", + "image_path": "906b2e6bd9ef7a8ae4cef7a1e3fe13e7b8297ba28a420c5a5a5326ca1826cf5b.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 298, + 80, + 378, + 141 + ], + "blocks": [ + { + "bbox": [ + 325, + 70, + 351, + 80 + ], + "lines": [ + { + "bbox": [ + 325, + 70, + 351, + 80 + ], + "spans": [ + { + "bbox": [ + 325, + 70, + 351, + 80 + ], + "type": "text", + "content": "S-Ray" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 298, + 80, + 378, + 141 + ], + "lines": [ + { + "bbox": [ + 298, + 80, + 378, + 141 + ], + "spans": [ + { + "bbox": [ + 298, + 80, + 378, + 141 + ], + "type": "image", + "image_path": "91a37656f6cc47463d6d7b7b66cbba290e2e94ee27f5fd4e6b26e5508a97c70f.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 380, + 80, + 461, + 141 + ], + "blocks": [ + { + "bbox": [ + 388, + 70, + 453, + 80 + ], + "lines": [ + { + "bbox": [ + 388, + 70, + 453, + 80 + ], + "spans": [ + { + "bbox": [ + 388, + 70, + 453, + 80 + ], + "type": "text", + "content": "GSNeRF (Ours)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 380, + 80, + 461, + 141 + ], + "lines": [ + { + "bbox": [ + 380, + 80, + 461, + 141 + ], + "spans": [ + { + "bbox": [ + 380, + 80, + 461, + 141 + ], + "type": "image", + "image_path": "f56b2e55d694112d99164465d9fafccc1026991198270e93b88f9bb776105f1f.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 463, + 80, + 544, + 141 + ], + "blocks": [ + { + "bbox": [ + 496, + 70, + 511, + 79 + ], + "lines": [ + { + "bbox": [ + 496, + 70, + 511, + 79 + ], + "spans": [ + { + "bbox": [ + 496, + 70, + 511, + 79 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 463, + 80, + 544, + 141 + ], + "lines": [ + { + "bbox": [ + 463, + 80, + 544, + 141 + ], + "spans": [ + { + "bbox": [ + 463, + 80, + 544, + 141 + ], + "type": "image", + "image_path": "f30efb61ab6e7221665b7a6fd565cc74898baadea65cbab7b2ee3181edb5ef42.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 298, + 143, + 378, + 204 + ], + "blocks": [ + { + "bbox": [ + 298, + 143, + 378, + 204 + ], + "lines": [ + { + "bbox": [ + 298, + 143, + 378, + 204 + ], + "spans": [ + { + "bbox": [ + 298, + 143, + 378, + 204 + ], + "type": "image", + "image_path": "91a3d33427c6a72490ca7ecc468f9148c46a497a789fc3afe4ab09d4064bdc6d.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 380, + 143, + 460, + 204 + ], + "blocks": [ + { + "bbox": [ + 380, + 143, + 460, + 204 + ], + "lines": [ + { + "bbox": [ + 380, + 143, + 460, + 204 + ], + "spans": [ + { + "bbox": [ + 380, + 143, + 460, + 204 + ], + "type": "image", + "image_path": "71cfb630bbd1a8159080a0afd2b0ccfe22a451447cf64b575a10839dcc66ac71.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 463, + 143, + 543, + 204 + ], + "blocks": [ + { + "bbox": [ + 463, + 143, + 543, + 204 + ], + "lines": [ + { + "bbox": [ + 463, + 143, + 543, + 204 + ], + "spans": [ + { + "bbox": [ + 463, + 143, + 543, + 204 + ], + "type": "image", + "image_path": "a6def87bcfc5f5dfb8013bc30605d2a15f464414920e701037191d8c39f21578.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 298, + 205, + 378, + 266 + ], + "blocks": [ + { + "bbox": [ + 298, + 205, + 378, + 266 + ], + "lines": [ + { + "bbox": [ + 298, + 205, + 378, + 266 + ], + "spans": [ + { + "bbox": [ + 298, + 205, + 378, + 266 + ], + "type": "image", + "image_path": "a21c11f1df7afe9baa5f807204463ae112d13f77c1ab5b571302d7758e67a94b.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 380, + 205, + 460, + 266 + ], + "blocks": [ + { + "bbox": [ + 380, + 205, + 460, + 266 + ], + "lines": [ + { + "bbox": [ + 380, + 205, + 460, + 266 + ], + "spans": [ + { + "bbox": [ + 380, + 205, + 460, + 266 + ], + "type": "image", + "image_path": "d60d961c28cc91b4060875a072ec7ee0e85ee35c15f661654f96e5a9541b387f.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 463, + 205, + 543, + 266 + ], + "blocks": [ + { + "bbox": [ + 463, + 205, + 543, + 266 + ], + "lines": [ + { + "bbox": [ + 463, + 205, + 543, + 266 + ], + "spans": [ + { + "bbox": [ + 463, + 205, + 543, + 266 + ], + "type": "image", + "image_path": "fa1609161ee7cd5030a1f826aa903550acc0a2b75fc3cbb6e1df635a3d59b640.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 298, + 268, + 378, + 329 + ], + "blocks": [ + { + "bbox": [ + 298, + 268, + 378, + 329 + ], + "lines": [ + { + "bbox": [ + 298, + 268, + 378, + 329 + ], + "spans": [ + { + "bbox": [ + 298, + 268, + 378, + 329 + ], + "type": "image", + "image_path": "f39a52bde205f78178bee4a35a45c1cf592e62604fea8b2a136ac0dc588d7998.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 380, + 268, + 460, + 329 + ], + "blocks": [ + { + "bbox": [ + 380, + 268, + 460, + 329 + ], + "lines": [ + { + "bbox": [ + 380, + 268, + 460, + 329 + ], + "spans": [ + { + "bbox": [ + 380, + 268, + 460, + 329 + ], + "type": "image", + "image_path": "9b9894c1e7716630c3baa2909ca11040d9fecec218db523fc44cb709646acb97.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 463, + 268, + 543, + 329 + ], + "blocks": [ + { + "bbox": [ + 463, + 268, + 543, + 329 + ], + "lines": [ + { + "bbox": [ + 463, + 268, + 543, + 329 + ], + "spans": [ + { + "bbox": [ + 463, + 268, + 543, + 329 + ], + "type": "image", + "image_path": "59c725c48e516ed597bbcf9a2f02c52d55bac3ca74ec9faae489c1d8bde03654.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "bbox": [ + 47, + 382, + 127, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 382, + 127, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 382, + 127, + 396 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 47, + 403, + 109, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 403, + 109, + 415 + ], + "spans": [ + { + "bbox": [ + 47, + 403, + 109, + 415 + ], + "type": "text", + "content": "5.1. Datasets" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 46, + 423, + 287, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 423, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 287, + 602 + ], + "type": "text", + "content": "To evaluate the effectiveness of our proposed method, we conduct experiments on both real-world and synthetic datasets. For real-world data, we use ScanNet [5], a large-scale, indoor RGB-D video dataset with over 2.5 million views from 1513 distinct scenes, including semantic annotations and camera poses. Following the setting presented in S-Ray [17], we train our model on 60 scenes and test its generalization on 10 new, unseen scenes. For synthetic data, we utilize Replica [26], a 3D reconstruction-based indoor dataset comprising 18 high-quality scenes with dense geometry, HDR textures, and semantic labels. Following the Semantic NeRF [37] supplied rendered data, we train our model on 6 distinct scenes across 12 video sequences and test on 2 novel scenes across 4 video sequences. Please see the supplementary material for details." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 47, + 613, + 164, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 613, + 164, + 625 + ], + "spans": [ + { + "bbox": [ + 47, + 613, + 164, + 625 + ], + "type": "text", + "content": "5.2. Results and analysis" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 47, + 632, + 167, + 644 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 632, + 167, + 644 + ], + "spans": [ + { + "bbox": [ + 47, + 632, + 167, + 644 + ], + "type": "text", + "content": "5.2.1 Quantitative Results" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "Table 2 compares our method against several baselines, in which S-Ray [17] serves as a primary baseline since it is the first to address semantic segmentation with NeRF in a generalized setting. We further conduct experiments on MVS-NeRF [3], GeoNeRF [14], GNT [32] and NeuRay [18] with" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 304, + 384, + 545, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 545, + 468 + ], + "type": "text", + "content": "the semantic head as described in [17]. It is worth noting that, even though S-Ray does not explicitly mention using a depth map of each source view image as input in their paper, their official implementation takes the depth maps " + }, + { + "bbox": [ + 304, + 384, + 545, + 468 + ], + "type": "inline_equation", + "content": "D_{1:K}" + }, + { + "bbox": [ + 304, + 384, + 545, + 468 + ], + "type": "text", + "content": " as additional input along with " + }, + { + "bbox": [ + 304, + 384, + 545, + 468 + ], + "type": "inline_equation", + "content": "I_{1:K}" + }, + { + "bbox": [ + 304, + 384, + 545, + 468 + ], + "type": "text", + "content": " during training and testing. Therefore, we follow their official implementation to reproduce S-Ray results." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 304, + 474, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 546, + 713 + ], + "type": "text", + "content": "From Table 2, it can be seen that our model generalizes well to unseen scenes. To show that our GSNeRF is also capable of producing impressive results without depth supervision, we conduct an additional experiment in which we only use the loss function in Eq. 17 to train our GSNeRF in Table 2. We also conduct experiments of S-Ray, Neuray with only RGB images as input, and GeoNeRF without GT depth as supervision for a fair comparison. From Table 2, we see that even without access to GT depth and with the target depth information regularized in a self-supervised manner, our method still outperforms all the other baseline approaches. This suggests that our model can be trained without requiring depth map supervision, confirming the effectiveness and practicality of our proposed approach. We also note that, although GeoNeRF + semhead (with depth supervision) slightly outperforms our approach regarding PSNR and SSIM for the rendered RGB images on ScanNet, our method excels in all semantic segmentation metrics by approximately " + }, + { + "bbox": [ + 304, + 474, + 546, + 713 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 304, + 474, + 546, + 713 + ], + "type": "text", + "content": ". More comprehensive discussions are presented in the supplementary materials." + } + ] + } + ], + "index": 38 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20812" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 288, + 162 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 288, + 162 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 288, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 288, + 162 + ], + "type": "table", + "html": "
ITST
ModelDepth-guidedrenderer PθDepth-guidedPSNRmIoU
A---28.4837.88
B✓--30.5641.72
C✓✓-30.8849.47
Ours✓✓✓31.4952.21
", + "image_path": "6d76ad9ec1e799fe0e7e3f60baa5f6c257308470bf4355842435e43db00d5e60.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 169, + 288, + 214 + ], + "lines": [ + { + "bbox": [ + 46, + 169, + 288, + 214 + ], + "spans": [ + { + "bbox": [ + 46, + 169, + 288, + 214 + ], + "type": "text", + "content": "Table 3. Ablation studies on GSNeRF. We verify the effectiveness of using depth-guided rendering for rendering " + }, + { + "bbox": [ + 46, + 169, + 288, + 214 + ], + "type": "inline_equation", + "content": "I_{T}" + }, + { + "bbox": [ + 46, + 169, + 288, + 214 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 169, + 288, + 214 + ], + "type": "inline_equation", + "content": "S_{T}" + }, + { + "bbox": [ + 46, + 169, + 288, + 214 + ], + "type": "text", + "content": ". Note that, without " + }, + { + "bbox": [ + 46, + 169, + 288, + 214 + ], + "type": "inline_equation", + "content": "P_{\\theta}" + }, + { + "bbox": [ + 46, + 169, + 288, + 214 + ], + "type": "text", + "content": ", we simply take a pre-trained 2D segmentor on " + }, + { + "bbox": [ + 46, + 169, + 288, + 214 + ], + "type": "inline_equation", + "content": "I_{T}" + }, + { + "bbox": [ + 46, + 169, + 288, + 214 + ], + "type": "text", + "content": " to produce " + }, + { + "bbox": [ + 46, + 169, + 288, + 214 + ], + "type": "inline_equation", + "content": "S_{T}" + }, + { + "bbox": [ + 46, + 169, + 288, + 214 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 48, + 220, + 289, + 369 + ], + "blocks": [ + { + "bbox": [ + 48, + 220, + 289, + 369 + ], + "lines": [ + { + "bbox": [ + 48, + 220, + 289, + 369 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 289, + 369 + ], + "type": "image", + "image_path": "1e7d1138daa3c614d285f06e394d0251d581c1fc5e5008ed8e7ccff8f9e949c6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 376, + 287, + 431 + ], + "lines": [ + { + "bbox": [ + 46, + 376, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 287, + 431 + ], + "type": "text", + "content": "Figure 4. Sampling efficiency on ScanNet. Compared to SOTAs, our GSNeRF is able to achieve significantly improved rendering performance especially when the number of sampling points is small. Even with increasing numbers of sampling points, GSNeRF still performs favorably against existing models." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 446, + 162, + 458 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 446, + 162, + 458 + ], + "spans": [ + { + "bbox": [ + 47, + 446, + 162, + 458 + ], + "type": "text", + "content": "5.2.2 Qualitative Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 464, + 287, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 464, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 46, + 464, + 287, + 620 + ], + "type": "text", + "content": "The qualitative results of our approach compared to S-Ray [17] are depicted in Fig. 3. This comparison illustrates that while the existing method can approximate the contours of objects, it is unable to capture the geometrical details, causing segmentation inconsistencies. In contrast, our GSNeRF accurately interprets scene geometry through our targeted view depth estimation design, allowing us to not only precisely segment objects but also render novel view images more realistically. This confirms that our depth-guided sampling technique effectively captures the correct geometry and that the sampled point features are appropriately aggregated from features of the Semantic Geo-Reasoner." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 635, + 151, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 151, + 646 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 151, + 646 + ], + "type": "text", + "content": "5.2.3 Ablation Studies" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "To further analyze the effectiveness of our designed modules, we conduct ablation studies on ScanNet, as shown in Table 3. Baseline model A employs our Semantic Geo-Reasoner " + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "G_{\\theta}" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": " and volume renderer " + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "R_{\\theta}" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": " to create the target view image " + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "I_{T}" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": ", using uniform sampling instead of depth" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 205 + ], + "type": "text", + "content": "guided sampling " + }, + { + "bbox": [ + 304, + 72, + 545, + 205 + ], + "type": "inline_equation", + "content": "x_{1:N}" + }, + { + "bbox": [ + 304, + 72, + 545, + 205 + ], + "type": "text", + "content": ". Also, a pre-trained 2D segmentor [24] is applied on the rendered image " + }, + { + "bbox": [ + 304, + 72, + 545, + 205 + ], + "type": "inline_equation", + "content": "I_T" + }, + { + "bbox": [ + 304, + 72, + 545, + 205 + ], + "type": "text", + "content": " to produce " + }, + { + "bbox": [ + 304, + 72, + 545, + 205 + ], + "type": "inline_equation", + "content": "S_T" + }, + { + "bbox": [ + 304, + 72, + 545, + 205 + ], + "type": "text", + "content": " without our semantic renderer " + }, + { + "bbox": [ + 304, + 72, + 545, + 205 + ], + "type": "inline_equation", + "content": "P_\\theta" + }, + { + "bbox": [ + 304, + 72, + 545, + 205 + ], + "type": "text", + "content": ". Model B enhances visual quality by replacing uniform sampling with depth-guided sampling. We further replace the 2D segmentor with our semantic renderer " + }, + { + "bbox": [ + 304, + 72, + 545, + 205 + ], + "type": "inline_equation", + "content": "P_\\theta" + }, + { + "bbox": [ + 304, + 72, + 545, + 205 + ], + "type": "text", + "content": " using uniform sampling for semantic rendering in model C, illustrating mutual benefits for both tasks. Our full model in the last row uses depth-guided sampling exclusively for surface points in the semantic renderer, achieving optimal results. This verifies the success of our proposed modules and sampling strategies." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 218, + 424, + 231 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 218, + 424, + 231 + ], + "spans": [ + { + "bbox": [ + 306, + 218, + 424, + 231 + ], + "type": "text", + "content": "5.2.4 Sampling Efficiency" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 237, + 545, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 237, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 304, + 237, + 545, + 441 + ], + "type": "text", + "content": "Another benefit of our depth-guided sampling is that the performance of our GSNeRF is less sensitive to the number of sampled points along each ray. In Fig. 4, we conduct experiments of reducing the number of sampling points " + }, + { + "bbox": [ + 304, + 237, + 545, + 441 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 237, + 545, + 441 + ], + "type": "text", + "content": ", showing PSNR of the synthesized novel view images as evaluation. We find that by employing depth-guided sampling, our method accurately preserves visual details with less noise (compared to Neuray and S-Ray with 128 sampled points along each ray). It can be seen that GeoNeRF + semhead shows a catastrophic drop in PSNR with fewer points, while ours leveraging depth-guided sampling maintains satisfactory visual quality even with a significantly reduced number of sampled points (e.g., when sampling only 4 points along a ray, the PSNR remains near 28). This result further attests to the success of our design in estimating target depth and implementing the corresponding sampling strategy with better novel view synthesis performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 452, + 378, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 452, + 378, + 464 + ], + "spans": [ + { + "bbox": [ + 306, + 452, + 378, + 464 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 471, + 545, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 471, + 545, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 471, + 545, + 615 + ], + "type": "text", + "content": "In this paper, we proposed Generalizable Semantic Neural Raidance Fields (GSNeRF) to achieve generalized novel view synthesis and semantic segmentation. Our GSNeRF is trained to derive visual features and perform depth map prediction of each source view so that the depth map for a novel target view can be estimated. With such target-view depth information observed, the associated RGB image and semantic segmentation can be jointly produced via depth-guide rendering. In our experiments, we quantitatively and qualitatively confirm that our GSNeRF performs favorably against existing generalizable semantic-aware NeRF methods on both real-world and synthetic datasets." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "text", + "content": "Acknowledgement This work is supported in part by the National Science and Technology Council via grant NSCT112-2634-F-002-007 and the NTU Center of Data Intelligence: Technologies, Applications, and Systems via grant NTU-113L900902. We also thank the National Center for High-performance Computing (NCHC) for providing computational and storage resources." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20813" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "text", + "content": "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 158, + 288, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 288, + 212 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 288, + 212 + ], + "type": "text", + "content": "[2] Di Chang, Aljaž Božić, Tong Zhang, Qingsong Yan, Yingcong Chen, Sabine Susstrunk, and Matthias Nießner. Rcmvsnet: unsupervised multi-view stereo with neural rendering. In European Conference on Computer Vision, pages 665-680. Springer, 2022. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 213, + 287, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 213, + 287, + 267 + ], + "spans": [ + { + "bbox": [ + 53, + 213, + 287, + 267 + ], + "type": "text", + "content": "[3] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14124-14133, 2021. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 269, + 287, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 269, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 269, + 287, + 312 + ], + "type": "text", + "content": "[4] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 314, + 287, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 314, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 53, + 314, + 287, + 369 + ], + "type": "text", + "content": "[5] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 369, + 287, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 287, + 413 + ], + "type": "text", + "content": "[6] Yuchao Dai, Zhidong Zhu, Zhibo Rao, and Bo Li. Mvs2: Deep unsupervised multi-view stereo with multi-view symmetry. In 2019 International Conference on 3D Vision (3DV), pages 1-8. IEEE, 2019. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 414, + 287, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 414, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 53, + 414, + 287, + 467 + ], + "type": "text", + "content": "[7] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12882-12891, 2022. 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 469, + 287, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 469, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 53, + 469, + 287, + 523 + ], + "type": "text", + "content": "[8] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 525, + 287, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 525, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 53, + 525, + 287, + 579 + ], + "type": "text", + "content": "[9] Xiao Fu, Shangzhan Zhang, Tianrun Chen, Yichong Lu, Lanyun Zhu, Xiaowei Zhou, Andreas Geiger, and Yiyi Liao. Panoptic nerf: 3d-to-2d label transfer for panoptic urban scene segmentation. In 2022 International Conference on 3D Vision (3DV), pages 1-11. IEEE, 2022. 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 581, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 287, + 644 + ], + "type": "text", + "content": "[10] Sourav Garg, Niko Sünderhauf, Feras Dayoub, Douglas Morrison, Akansel Cosgun, Gustavo Carneiro, Qi Wu, Tat-Jun Chin, Ian Reid, Stephen Gould, et al. Semantics for robotic mapping, perception and interaction: A survey. Foundations and Trends® in Robotics, 8(1-2):1-224, 2020. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 647, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 678 + ], + "type": "text", + "content": "[11] Ross Girshick. Fast r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 1440-1448, 2015. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[12] Xiaodong Gu, Zhiwen Fan, Siyu Zhu, Zuozhuo Dai, Feitong Tan, and Ping Tan. Cascade cost volume for high-resolution multi-view stereo and stereo matching. In Proceedings of" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 74, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 326, + 74, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 74, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 74, + 545, + 95 + ], + "type": "text", + "content": "the IEEE/CVF conference on computer vision and pattern recognition, pages 2495-2504, 2020. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 96, + 545, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 148 + ], + "type": "text", + "content": "[13] Baichuan Huang, Hongwei Yi, Can Huang, Yijia He, Jingbin Liu, and Xiao Liu. M3vsnet: Unsupervised multi-metric multi-view stereo network. In 2021 IEEE International Conference on Image Processing (ICIP), pages 3163-3167. IEEE, 2021. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 151, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 151, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 151, + 545, + 205 + ], + "type": "text", + "content": "[14] Mohammad Mahdi Johari, Yann Lepoittevin, and François Fleuret. Geonerf: Generalizing nerf with geometry priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18365-18375, 2022. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 206, + 545, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 281 + ], + "type": "text", + "content": "[15] Abhijit Kundu, Kyle Genova, Xiaqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12871-12881, 2022. 1, 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 283, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 283, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 308, + 283, + 545, + 327 + ], + "type": "text", + "content": "[16] Haotong Lin, Sida Peng, Zhen Xu, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Efficient neural radiance fields for interactive free-viewpoint video. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 327, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 370 + ], + "type": "text", + "content": "[17] Fangfu Liu, Chubin Zhang, Yu Zheng, and Yueqi Duan. Semantic ray: Learning a generalizable semantic field with cross-reprojection attention. In CVPR, pages 17386-17396, 2023. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 372, + 545, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 372, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 308, + 372, + 545, + 436 + ], + "type": "text", + "content": "[18] Yuan Liu, Sida Peng, Lingjie Liu, Qianqian Wang, Peng Wang, Christian Theobalt, Xiaowei Zhou, and Wenping Wang. Neural rays for occlusion-aware image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7824-7833, 2022. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 437, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 545, + 503 + ], + "type": "text", + "content": "[19] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7210-7219, 2021. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 504, + 545, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 545, + 557 + ], + "type": "text", + "content": "[20] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 559, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 602 + ], + "type": "text", + "content": "[21] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 604, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 604, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 604, + 545, + 657 + ], + "type": "text", + "content": "[22] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14335-14345, 2021. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "text", + "content": "[23] Barbara Roessle, Jonathan T Barron, Ben Mildenhall, Pratul P Srinivasan, and Matthias Nießner. Dense depth priors for neural radiance fields from sparse input views. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12892-12901, 2022. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "20814" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 139 + ], + "type": "text", + "content": "[24] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 140, + 288, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 288, + 205 + ], + "type": "text", + "content": "[25] Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulò, Norman Müller, Matthias Nießner, Angela Dai, and Peter Kontschieder. Panoptic lifting for 3d scene understanding with neural fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9043-9052, 2023. 1, 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 206, + 287, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 206, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 206, + 287, + 259 + ], + "type": "text", + "content": "[26] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "type": "text", + "content": "[27] Mohammed Suhail, Carlos Esteves, Leonid Sigal, and Ameesh Makadia. Generalizable patch-based neural rendering. In European Conference on Computer Vision, pages 156-174. Springer, 2022. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 305, + 287, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 305, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 48, + 305, + 287, + 358 + ], + "type": "text", + "content": "[28] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 360, + 287, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 287, + 404 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 287, + 404 + ], + "type": "text", + "content": "[29] Alex Trevithick and Bo Yang. Grf: Learning a general radiance field for 3d representation and rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15182-15192, 2021. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 404, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 287, + 459 + ], + "type": "text", + "content": "[30] Suhani Vora, Noha Radwan, Klaus Greff, Henning Meyer, Kyle Genova, Mehdi SM Sajjadi, Etienne Pot, Andrea Tagliasacchi, and Daniel Duckworth. Nesf: Neural semantic fields for generalizable semantic segmentation of 3d scenes. arXiv preprint arXiv:2111.13260, 2021. 1, 2, 3, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 459, + 287, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 459, + 287, + 492 + ], + "spans": [ + { + "bbox": [ + 48, + 459, + 287, + 492 + ], + "type": "text", + "content": "[31] Bing Wang, Lu Chen, and Bo Yang. Dm-nerf: 3d scene geometry decomposition and manipulation from 2d images. arXiv preprint arXiv:2208.07227, 2022. 1, 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 492, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 492, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 492, + 287, + 525 + ], + "type": "text", + "content": "[32] Peihao Wang, Xuxi Chen, Tianlong Chen, Subhashini Venugopalan, Zhangyang Wang, et al. Is attention all nerf needs? arXiv preprint arXiv:2207.13298, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 525, + 287, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 592 + ], + "type": "text", + "content": "[33] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690-4699, 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 593, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 593, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 593, + 287, + 635 + ], + "type": "text", + "content": "[34] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vision (ECCV), pages 767-783, 2018. 4, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 636, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 690 + ], + "type": "text", + "content": "[35] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5752-5761, 2021. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "text", + "content": "[36] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 162 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "text", + "content": "In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021. 1, 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 107, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 162 + ], + "type": "text", + "content": "[37] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J Davison. In-place scene labelling and understanding with implicit scene representation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15838-15847, 2021. 1, 2, 3, 5, 7" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "20815" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/GSVA_ Generalized Segmentation via Multimodal Large Language Models/12e5e8f5-7030-48a1-9213-1240824636f7_content_list.json b/2024/GSVA_ Generalized Segmentation via Multimodal Large Language Models/12e5e8f5-7030-48a1-9213-1240824636f7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..cb7436198c6e326fed57a365555fe5b9f8daae39 --- /dev/null +++ b/2024/GSVA_ Generalized Segmentation via Multimodal Large Language Models/12e5e8f5-7030-48a1-9213-1240824636f7_content_list.json @@ -0,0 +1,1753 @@ +[ + { + "type": "text", + "text": "GSVA: Generalized Segmentation via Multimodal Large Language Models", + "text_level": 1, + "bbox": [ + 107, + 130, + 861, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhuofan Xia* Dongchen Han* Yizeng Han Xuran Pan Shiji Song Gao Huang†", + "bbox": [ + 137, + 180, + 836, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Department of Automation, BNRist, Tsinghua University", + "bbox": [ + 292, + 200, + 671, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 251, + 313, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generalized Referring Expression Segmentation (GRES) extends the scope of classic RES to refer to multiple objects in one expression or identify the empty targets absent in the image. GRES poses challenges in modeling the complex spatial relationships of the instances in the image and identifying non-existing referents. Multimodal Large Language Models (MLLMs) have recently shown tremendous progress in these complicated vision-language tasks. Connecting Large Language Models (LLMs) and vision models, MLLMs are proficient in understanding contexts with visual inputs. Among them, LISA, as a representative, adopts a special [SEG] token to prompt a segmentation mask decoder; e.g., SAM, to enable MLLMs in the RES task. However, existing solutions to GRES remain unsatisfactory since current segmentation MLLMs cannot correctly handle the cases where users might reference multiple subjects in a singular prompt or provide descriptions incongruent with any image target. In this paper, we propose Generalized Segmentation Vision Assistant (GSVA) to address this gap. Specifically, GSVA reuses the [SEG] token to prompt the segmentation model towards supporting multiple mask references simultaneously and innovatively learns to generate a [REJ] token to reject the null targets explicitly. Experiments validate GSVA's efficacy in resolving the GRES issue, marking a notable enhancement and setting a new record on the GRES benchmark gRefCOCO dataset. GSVA also proves effective across various classic referring segmentation and comprehension tasks. Code is available at https://github.com/LeapLabTHU/GSVA.", + "bbox": [ + 75, + 282, + 472, + 720 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 750, + 207, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Referring Expression Segmentation (RES) [5, 25] is an emerging vision-language (VL) task predicting the masks of the interested objects referred to in the language expression. RES has great potential in many areas, especially embodied AI [17, 28, 54, 58, 66], including VL navigation, humanrobot interaction, etc. Nevertheless, the simplification in", + "bbox": [ + 75, + 773, + 468, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e2387831eaffe94f5cd97cbfd1215f1b02e9a6c9b6249555ee5eca4f28302746.jpg", + "image_caption": [ + "Figure 1. Comparison of the segmentation masks by LISA [32] and GSVA, facing the challenges in Generalized Referring Expression Segmentation (GRES) [38]. (a) LISA fails to segment the correct targets when multiple targets are requested due to the single [SEG] token restriction. GSVA successfully generates all target masks via learning multiple [SEG] tokens. (b) When the referent does not exist in the image, i.e., the empty target is requested, LISA reluctantly produces the wrong mask because of the compulsive [SEG] token output. In contrast, GSVA can reject the empty targets by predicting [REJ] tokens in the output sequence." + ], + "image_footnote": [], + "bbox": [ + 506, + 252, + 883, + 450 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "RES formulation that one referring expression must match an individual object in the image [30] has left a gap between current RES algorithms and real-world applications, neglecting multiple-target and empty-target cases.", + "bbox": [ + 496, + 609, + 890, + 671 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To bridge this gap, Generalized Referring Expression Segmentation (GRES) [38] has recently been proposed to support multiple-target and empty-target cases. Practically, users refer to multiple subjects within a single instruction or provide descriptions that do not correspond to any targets in the image. As an example shown in Figure 1 (a), the left man and the second man from the right are targeted simultaneously, while a banana is referred to in a scene of apples in Figure 1 (b). RES takes no account of these cases, which GRES handles. In addition to the multimodal correspondences between images and text prompts in classic RES, GRES poses new challenges in handling more complicated multiple-target and empty-target cases. Therefore, the models must handle complex spatial relationships of the instances in the image [38] to segment the targets at various", + "bbox": [ + 496, + 674, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 94, + 875, + 207, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 96, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3858", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "locations and reject the empty targets in the wrong places.", + "bbox": [ + 76, + 90, + 460, + 107 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The recent blooming Multimodal Large Language Models (MLLMs) [1, 34, 39, 83, 87] meet the requirements of GRES since they show excellency in complex reasoning [8] and instruction following [49] with visual inputs by aligning the LLMs [2, 6, 48, 60-62] and Visual Foundation Models (VFMs) [15, 55, 63, 79] which are typically various Vision Transformers [14, 18, 23, 50, 51, 76, 78] to perceive image or video inputs. To support segmentation output, many works [74, 75, 86] link an MLLM (e.g., LLaVA [39]) and a segmentation model (e.g., SAM [31]) by prompting the decoder with special token embeddings (e.g., [SEG] in LISA [32]) to generate masks of the referents in the user's instructions. Although these models manage to handle RES, GRES is still beyond their reach. As shown in Figure 1, LISA fails to work well in GRES where the multiple-target and empty-target challenges remain uncharted.", + "bbox": [ + 75, + 109, + 468, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the above challenges, we propose Generalized Segmentation Vision Assistant (GSVA). We attribute the vulnerability of other segmentation MLLMs in GRES to the single constant [SEG] token that restricts its flexibility. Therefore, we present two pivotal designs in GSVA: (1) learning to predict multiple [SEG] tokens to segment multiple targets; (2) rejecting empty targets in referring expressions by predicting [REJ] tokens. Specifically, when multiple targets are requested in the referring expression, we place multiple weight-sharing [SEG] tokens corresponding to the entities in the expression, encouraging the MLLM to learn to output multiple [SEG] tokens. To distinguish each [SEG] token and avoid ambiguity, we add the expression of each entity in front of the corresponding [SEG] token, hinting each [SEG] token to focus on the specific target, which can be regarded as implicit In-Context Learning, and dynamic neural network [19]. Meanwhile, if the referents are absent in the image, the corresponding [SEG] tokens after the prompts are altered to [REJ] tokens to identify empty targets. The predicted [REJ] tokens are directly assigned with empty masks without decoded, which liberates the segmentation model from seeking non-existing targets in the image. This Benefiting from these novel designs, GSVA takes a big step forward in addressing GRES challenges, as shown in the second row of Figure 1.", + "bbox": [ + 75, + 354, + 472, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows: (1) We propose GSVA to solve the GRES problem with MLLM by handling the spatial relationships among targets, and study the GRES problem systematically in the context of LLM for the first time. (2) We propose the non-trivial shared-weight multiple [SEG] tokens guided by each referent prompt to address the multiple-target problem. (3) We firstly propose a clean solution, the [REJ] token, to reject the empty targets, which can be seamlessly applied to various models. (4) GSVA is intuitive and effective, achieving state-of-the-art performance on the GRES benchmarks.", + "bbox": [ + 75, + 734, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 89, + 650, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "RES and GRES. Referring Expression Segmentation (RES) [4, 5, 25, 44] assumes that one expression matches one existing target, and many works explore fusing image and language [3, 16, 35, 37, 81] to segment objects under instructions. Currently, most RES methods adopt the cross-attention module or cross-modal alignment to bridge the modality gap [7, 57, 73, 82, 89]. Another line of research enables the text prompts for segmentation model with a unified decoder [31, 41, 93, 94], offering more flexible outputs. To break the jail for the arbitrary number of targets, DMMI [26] focuses on the one-to-many setting where text expression refers to varying numbers of targets. ReLA [38] proposes the Generalized Referring Expression Segmentation (GRES) task, supporting both the multi-target and empty-target scenarios, which is our main research scope.", + "bbox": [ + 496, + 114, + 890, + 343 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "MLLM. Multimodal Large Language Models (MLLMs) align the vision and language modalities by various techniques, including cross-attention module [1], prompt tuning tokens [87], Q-Former [9, 34], linear projection layers [39], and unified model architectures [52, 79]. Endowed with unprecedented capabilities in reasoning with world-knowledge and following instructions of users, MLLM shows extraordinary performances in various vision language tasks [43, 65]. Equipped with the segmentation decoders or detection heads, MLLMs can also excel in the vision-centric tasks, such as object detection and segmentation [36, 56, 64, 80, 86]. Among them, LISA [32] makes the most of the reasoning ability with a [SEG] token to address the Reasoning Segmentation problem. However, LISA fails to tackle the challenge in GRES due to the inflexible [SEG] token, which is addressed by our proposed GSVA.", + "bbox": [ + 496, + 343, + 892, + 584 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dynamic Networks. Dynamic neural networks [19] can adapt their architectures [20-22, 24, 27, 67-70] or parameters [11, 12, 53] to different inputs or switch the computation architecture in adjustment to different different time steps [46, 47, 71, 77], in order to achieve better accuracy and efficiency. In GSVA, the weight-sharing [SEG] tokens adapt to multiple targets under the hint of the prepended target prompts and dynamically reject empty targets with an individual prediction of rejection tokens.", + "bbox": [ + 496, + 585, + 890, + 722 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Generalized Segmentation Vision Assistant", + "text_level": 1, + "bbox": [ + 498, + 738, + 883, + 755 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we initiate with the introduction of the model design of Generalized Segmentation Vision Assistant (GSVA), which is followed by the analysis of some certain limitations of LISA in Generalized Referring Segmentation (GRES). Subsequently, we delve into the introduction of two pivotal elements of GSVA, segmenting multiple targets and learning the rejection token. These components are fundamental in the conceptualization and construction of GSVA.", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3859", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3e61033a07e86c5ac476931d0362596077bece25caa38a012ce7bc873e60cd01.jpg", + "image_caption": [ + "Figure 2. Overview of GSVA. At the bottom of the figure, the MLLM encodes the input image and concatenates the tokenized text tokens to follow instructions. GSVA generates multiple [SEG] tokens to handle multiple referred targets and rejects the objects absent in the image through [REJ] tokens. At the top of the figure, the SFM also encodes the image for segmentation and selects all [SEG] tokens in the output sequence to prompt the mask decoder to segment the target objects referred to in the instructions." + ], + "image_footnote": [], + "bbox": [ + 76, + 89, + 890, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Model Architecture", + "text_level": 1, + "bbox": [ + 76, + 424, + 264, + 439 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The architecture of GSVA is illustrated in Figure 2, resembling LISA [32], which enables high-fidelity segmentation outputs by integrating two types of foundation models: (1) Multimodal Large Language Model (MLLM) as an aligned vision-language cognitive module; (2) Segmentation Foundation Model (SFM) to segment the target out of the input image based on user's instruction. To connect these two modules, LISA proposes a paradigm named embedding as mask where an extra [SEG] token is appended to the vocabularies of the MLLM and serves as the prompt of the SFM to segment the target following the intention of the user.", + "bbox": [ + 75, + 448, + 468, + 613 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multimodal Large Language Model. The MLLM consists of a decoder-based language model $F_{\\mathrm{LLM}}$ to auto-regressively generate text responses following the user's inputs, a vision encoder $F_{\\mathrm{V1}}$ to extract features from the input image, and a linear projector $\\phi$ to align the representations between image and text modalities. Specifically, the pretrained LLaVA [39] variants with CLIP-ViT-L/14 [55] and Vicuna-7B/13B [6] are employed. Given an input image $\\mathbf{x}_{\\mathrm{img}}$ , the vision encoder $F_{\\mathrm{V1}}$ first encodes it into image features, and then the projector $\\phi$ maps the features into the visual token embeddings in the LLM input space:", + "bbox": [ + 75, + 616, + 468, + 782 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {h} _ {\\mathrm {i m g}} = \\phi \\left(F _ {\\mathrm {V} 1} \\left(\\mathbf {x} _ {\\mathrm {i m g}}\\right)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 787, + 468, + 805 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the input image $\\mathbf{x}_{\\mathrm{img}}$ is typically resized to $h\\times w\\times 3$ and the image tokens $\\mathbf{h}_{\\mathrm{img}}\\in \\mathbb{R}^{n_{\\mathrm{img}}\\times d}$ is aligned with the language modality. For CLIP-ViT-L/14, the input image with $h = w = 224$ is encoded with ViT of patch size in 14, therefore the length of tokens $n_{\\mathrm{img}} = hw / 14^2 = 256$ and the LLM dimensions $d$ are 4096 and 5120 for Vicuna", + "bbox": [ + 73, + 810, + 470, + 898 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "7B/13B, respectively. Along with the input image, the text instructions describing the targets to segment are tokenized into text tokens by the LLM tokenizer $T$ :", + "bbox": [ + 498, + 425, + 892, + 469 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {h} _ {\\mathrm {t x t}} = T (\\mathbf {x} _ {\\mathrm {t x t}}). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 473, + 890, + 489 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The image tokens and text tokens are concatenated together and then fed into the LLM after prepending a sequence of fixed prompt tokens $\\mathbf{h}_{\\mathrm{prompt}}$ (omitted in the figure). The output token embeddings $\\tilde{\\mathbf{y}}_{\\mathrm{txt}}$ are generated autoregressively:", + "bbox": [ + 496, + 493, + 890, + 568 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {y}} _ {\\mathrm {t x t}} = F _ {\\mathrm {L L M}} \\left(\\left[ \\mathbf {h} _ {\\text {p r o m p t}} \\right| \\left| \\mathbf {h} _ {\\text {i m g}} \\right| \\left| \\mathbf {h} _ {\\text {t x t}} \\right|\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 575, + 890, + 592 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $||$ is concatenation operation. The text responses are obtained from $\\tilde{\\mathbf{y}}_{\\mathrm{txt}}$ by applying a linear classifier to predict the next words in the vocabulary.", + "bbox": [ + 496, + 595, + 890, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In LISA, a special token [SEG] is appended in the vocabulary to activate the segmentation ability of MLLM. The model learns to predict [SEG] token in the output sequence to indicate there is a target to segment. LISA then selects the output embedding of the [SEG] token $\\tilde{\\mathbf{y}}_{\\mathrm{xt}}[\\mathrm{SEG}]$ , and projects it into the prompt space of the SFM by an MLP projector $\\psi$ :", + "bbox": [ + 496, + 641, + 890, + 744 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {h} _ {\\mathrm {s e g}} = \\psi (\\tilde {\\mathcal {Y}} _ {\\mathrm {t x t}} [ \\mathrm {S E G} ]). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 744, + 890, + 761 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The segmentation model is hence ready to decode the target mask from the query token $\\mathbf{h}_{\\mathrm{seg}}$", + "bbox": [ + 496, + 763, + 890, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Segmentation Foundation Model. The SFM is a query-based segmentation model, where a frozen vision encoder $F_{\\mathrm{V2}}$ takes in images of higher resolution than the vision encoder in MLLM to keep more details, followed by a trainable mask decoder $F_{\\mathrm{mask}}$ to decode masks from the queries. The pretrained SAM [31] with ViT-H backbone is instantiated as the SFM to produce high-quality masks. The given", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3860", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "input image $\\mathbf{x}_{\\mathrm{img}}$ is resized to a larger resolution $H\\times W\\times 3$ with $H = W = 1024$ and encoded with $F_{\\mathrm{V2}}$ to extract features $\\mathbf{f}_{\\mathrm{seg}}$ for segmentation:", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {\\text {s e g}} = F _ {\\mathrm {V} 2} \\left(\\mathbf {x} _ {\\text {i m g}}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 138, + 468, + 156 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Condition on the features $\\mathbf{f}_{\\mathrm{seg}} \\in \\mathbb{R}^{\\frac{H}{16} \\times \\frac{W}{16} \\times C}$ with $C = 256$ for SAM, the mask decoder $F_{\\mathrm{mask}}$ decodes the segmentation masks from the segmentation queries $\\mathbf{h}_{\\mathrm{seg}} \\in \\mathbb{R}^{N_{\\mathrm{seg}} \\times C}$ :", + "bbox": [ + 76, + 160, + 468, + 207 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {y}} _ {\\text {m a s k}} = F _ {\\text {m a s k}} \\left(\\mathbf {h} _ {\\text {s e g}} | \\mathbf {f} _ {\\text {s e g}}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 210, + 468, + 227 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where each query in $\\mathbf{h}_{\\mathrm{seg}}$ corresponds to one segmentation mask in $\\tilde{\\mathbf{y}}_{\\mathrm{mask}}$ .", + "bbox": [ + 76, + 231, + 467, + 260 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "LISA assumes that only one target exists to segment in the input image and its corresponding instructions. However, in GSVA, we extend it to a new scenario with multiple targets and empty targets, including multiple [SEG] tokens to invoke segmentation and [REJ] tokens to reject unreasonable instructed targets absent in the image. As shown in Figure 2, GSVA supports multiple [SEG]/[REJ] tokens in the output sequence and selects all the [SEG] tokens and discards every [REJ] token after attaining $\\tilde{\\mathbf{y}}_{\\mathrm{txt}}$ in Eq. (3). Therefore, there is more than one query in $\\mathbf{h}_{\\mathrm{seg}}$ , thus enabling the SFM to segment multiple targets. These designs make GSVA competent in the GRES task, where the awareness of multiple and empty targets is of vital importance.", + "bbox": [ + 75, + 262, + 468, + 458 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. GRES: Task and Challenges", + "text_level": 1, + "bbox": [ + 76, + 468, + 331, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Task. Generalized Referring Expression Segmentation (GRES) [38] removes the constraint on the number of referred targets in the expression in the conventional Referring Expression Segmentation (RES) [30, 44, 84]. Different from that one expression only refers to one instance or region in RES, GRES allows arbitrary numbers of referred targets, including multiple instances or no target circumstances. In GRES, the user can refer to many instances simultaneously or include the objects that do not exist in the image. For instance, there are three referring expressions in Figure 2, including a plate of broccoli and a spoon in the bowl of sauce, vase bottom left corner, and soup in bowl right and bowl bottom left corner. In the GRES case, the model ought to segment the masks of the objects referred to in the $1^{\\text{st}}$ and $3^{\\text{rd}}$ expression, meanwhile producing an empty mask for the $2^{\\text{nd}}$ expressions since there are no vases present at the bottom left corner.", + "bbox": [ + 75, + 491, + 468, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Challenges. The challenges in GRES are common in practice, especially in embodied AI [17, 54, 58, 66]. The one challenge is multiple targets. Take robot navigation and planning [4, 45, 66] as an example, a robot may be asked to perceive multiple targets in the surrounding environment, e.g., to bring the two bowls of soup in Figure 2. The vision system of the robot needs to locate and segment the containers holding the referred food one at a time. The other challenge is empty target. Suppose the robot is ordered to cut an apple with a knife in the scene of Figure 2, whereas", + "bbox": [ + 75, + 750, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "no apple is in the view of the camera, the vision system has to identify that the referred object is not in the scene. If it relies on some conventional RES methods which assume the expression must match something in the image, the output of the vision system could be undefined and potentially dangerous in some real-world cases.", + "bbox": [ + 496, + 90, + 890, + 180 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Differences from ReasonSeg. Reasoning Segmentation (ReasonSeg) proposed by LISA [32] emphasizes the complex text instructions in RES. In ReasonSeg, the instructions are more implicit and sophisticated, forcing models to reason using world knowledge. Besides, the logic chain is usually longer and more challenging in ReasonSeg, which requires the model to deduce the final target object referred to in the image. In contrast, GRES increases complexity in another dimension by involving complicated spatial relationships. Hence, the model has to learn to handle this spatial information and understand the relationships between the instances. To meet these requirements, LISA tunes MLLMs with complex instructions paired with masks, while GSVA arouses the spatial modeling capabilities of MLLMs by learning multiple targets and rejecting empty targets.", + "bbox": [ + 496, + 181, + 890, + 409 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Multiple [SEG] Tokens for Multiple Targets", + "text_level": 1, + "bbox": [ + 498, + 416, + 870, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Single [SEG] token. LISA [32] follows the classic RES methods to generate a segmentation mask under given instructions by adding the [SEG] token into the answer. The prompt is formatted as:", + "bbox": [ + 496, + 439, + 890, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "User: What is {obj} in this image? Please output segmentation mask. Assistant: Sure, it is [SEG].", + "bbox": [ + 521, + 502, + 890, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the above prompt, $\\{obj\\}$ represents an instance referred to or some semantic area to segment, and the embeddings of the [SEG] token output from the LLM are projected to prompt the SFM. When multiple instances are requested simultaneously, this prompt would confuse the model since only one [SEG] token is forced to match several targets. As shown in Figure 1 (a), LISA coercively predicts the masks of the left man and the guy second from the right, leading to a tattered mask and masking the wrong location.", + "bbox": [ + 496, + 534, + 890, + 669 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multiple [SEG] tokens. To mitigate this issue in GRES, we relax this constraint for GSVA to support multiple target outputs via learning multiple [SEG] tokens. To avoid the ambiguity between the [SEG] tokens and the corresponding objects, we prepend the referring expression before each [SEG] token, i.e., each segmentation prompt in the output text is $\\{obj\\_n\\} :[SEG]$ , as shown in Figure 3 (a). Streaming all the pairs of object and [SEG] token into one sentence, the question-answer prompt is formatted as (e.g., two targets):", + "bbox": [ + 496, + 670, + 890, + 806 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "User: What are {obj1}, {obj2} in this image? Please output the segmentation masks.", + "Assistant: Sure, $\\{obj1\\} :[SEG]$ $\\{obj2\\} :[SEG]$" + ], + "bbox": [ + 521, + 808, + 890, + 852 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This prompt requests the MLLM to identify and distinguish different objects in the image based on the instructions and infuse the corresponding location information", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3861", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c06c0cf7c3251335e0eb02ab8135f9741a1798f23205e9fe1cd6f93724455a27.jpg", + "image_caption": [ + "(a) Multiple Targets", + "Figure 3. Example of the prompts and predicted masks of GSVA-Vicuna-7B drawn from gRefCOCO validation set. (a) depicts the multiple-target case, in which two zebras referred to are handled with two separate [SEG] tokens. (b) shows the empty-target case, where no apple is in the bowl. Thus, the null referent is rejected with a [REJ] token, and no segmentation mask will be generated." + ], + "image_footnote": [], + "bbox": [ + 80, + 106, + 269, + 224 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "User:", + "text_level": 1, + "bbox": [ + 81, + 234, + 114, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "What are front zebra, right zebra in this image? Please output segmentation masks.", + "bbox": [ + 81, + 246, + 261, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Assistant:", + "text_level": 1, + "bbox": [ + 81, + 297, + 138, + 309 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sure, front zebra:[SEG], right zebra:[SEG].", + "bbox": [ + 81, + 310, + 246, + 334 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a4c5d6ec3e13b820e727079189eb8de2c1e9b89b887bc43f7cffbbf63dc46067.jpg", + "image_caption": [ + "(b) Empty Target" + ], + "image_footnote": [], + "bbox": [ + 276, + 106, + 467, + 224 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "User:", + "text_level": 1, + "bbox": [ + 279, + 233, + 313, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "What are the apple in the bowl, partial orange bottom left in this image? Please output segmentation masks.", + "bbox": [ + 279, + 246, + 457, + 294 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Assistant:", + "text_level": 1, + "bbox": [ + 279, + 297, + 336, + 309 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sure, the apple in the bowl: [REJ], partial orange bottom left:[SEG].", + "bbox": [ + 279, + 310, + 439, + 347 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of each target into the associated [SEG] token. We regard this ability as an implicit multimodal version of In-Context Learning (ICL), which is demonstrated by many prior works [2, 13, 33, 59, 90]. GSVA takes the target description preceding each [SEG] token as a hint to link this token to the object requested in the image through auto-regressive decoding.", + "bbox": [ + 75, + 438, + 468, + 545 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Rejecting Empty Targets via [REJ] Tokens", + "text_level": 1, + "bbox": [ + 76, + 553, + 439, + 570 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Empty targets. In GRES, many expressions match no targets in the image, including absence, incorrect attributes, inaccurate locations, etc. These expressions should be treated as empty targets for models to predict all-negative masks. LISA [32] falls short of predicting masks with all-zeros seamlessly since the [SEG] token always calls for a segmentation mask. As shown in Figure 1 (b), LISA incorrectly marks a piece of green apple as the empty target banana.", + "bbox": [ + 75, + 577, + 468, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "[REJ] token. We let the MLLM of GSVA predict a [REJ] token for each object that does not exist in the image but is referred to in the instructions, as shown in Figure 3 (b). GSVA predicts the targets marked with [REJ] tokens as empty targets, therefore setting all-zero masks for them. The involvement of [REJ] tokens directly rejects the empty target, liberating the mask decoder of the SFM from learning to segment the inconsistent targets. An example prompt with one existing target and one empty target is as follows:", + "bbox": [ + 75, + 698, + 468, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "User: What are {obj1} (absent), {obj2} (absent), {obj3} in this image? Please output the segmentation masks.", + "bbox": [ + 98, + 837, + 468, + 866 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Assistant: Sure, $\\{obj1\\} :[REJ],\\{obj2\\} :[REJ],\\{obj3\\} :[SEG].$", + "bbox": [ + 98, + 867, + 467, + 882 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The [REJ] token prediction can also be seen as a variant", + "bbox": [ + 76, + 885, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of VQA task, where the specified object and its position in the image need to be considered. Thanks to the unprecedented capabilities of MLLM in understanding the images [9, 34, 39, 92] and reasoning the spatial relationships of the referring objects [52, 88, 91], we make the most of the MLLM in GSVA to unleash the burden of the segmentation model. The proposed empty-target-aware mechanism both improves the quality of masks and ameliorates the errors of identifying nonexistent objects.", + "bbox": [ + 496, + 90, + 890, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 239, + 632, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we conduct comprehensive experiments to validate the efficacy of GSVA. First, we show the results on gRefCOCO [38] dataset to show the superiority of GSVA in GRES tasks. Then we verify GSVA that is also competent with other baselines in classic RES, REC tasks. We move on to ablate some important design choices of GSVA, followed by some qualitative visualization of GRES results.", + "bbox": [ + 496, + 265, + 890, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. GRES", + "text_level": 1, + "bbox": [ + 500, + 378, + 586, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Settings. We adopt gRefCOCO [38] dataset to validate GSVA and LISA [32] on GRES, which contains 278,232 expressions, including 80,022 multi-target and 32,202 empty-target ones, referring to the objects in 19,994 images. The images are split into four subsets: training, validation, test-A, and test-B, following the same UNC partition of RefCOCO [84]. We first add the gRefCOCO training set into the mixed training dataset in LISA to pretrain GSVA and LISA for 50,000 steps and then finetune the models on the gRefCOCO training dataset for another 10 epochs. We evaluate the pretrained and finetuned models on the remaining validation set, test set A, and test set B, respectively. We adopt the IoU, cIoU metrics for the segmentation mask outputs. Following the implementation in Liu et al. [38], IoU averages the IoU for each mask, whereas cIoU computes the cumulative intersection area over the cumulative union area across the whole dataset. As for the empty target, we compute the No-target-accuracy (N-acc.), which is the ratio of the correctly classified empty-target expressions over all the empty-target expressions in the dataset. For a correctly classified empty target, the gIoU is set to 1.0, and the cIoU does not take them into account, while for a misclassified empty target, the gIoU is set to 0.0 and the union area is accumulated in the cIoU. Following the criteria in Liu et al. [38], a predicted mask is regarded as empty for LISA if the positive pixels are less than 50, whereas GSVA predicts [REJ] tokens to identify empty targets.", + "bbox": [ + 496, + 402, + 890, + 810 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Results. We report the GRES segmentation results of GSVA and LISA [32] in Table 1. Three variants of GSVA, including GSVA-Vicuna-7B, GSVA-Vicuna-13B, and GSVA-Llama2-13B show competitive performance without finetuning on gIoU to the strongest non-LLM baseline ReLA [38]. However, LISA models fail to handle the", + "bbox": [ + 496, + 810, + 892, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3862", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/6e3c0a02185afd8b429b419df3b4f3768bf1c96a108c180021e587f97c6e008a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodGeneralized Referring Expression Segmentation on gRefCOCO
Validation SetTest Set ATest Set B
gIoUcIoUN-acc.gIoUcIoUN-acc.gIoUcIoUN-acc.
MattNet [85]48.2447.5141.1559.3058.6644.0446.1445.3341.32
LTS [29]52.7052.30-62.6461.87-50.4249.96-
VLT [10]52.0052.5147.1763.2062.1948.7450.8850.5247.82
CRIS [72]56.2755.34-63.4263.82-51.7951.04-
LAVT [81]58.4057.6449.3265.9065.3249.2555.8355.0448.46
ReLA [38]63.6062.4256.3770.0369.2659.0261.0259.8858.40
LISA-Vicuna-7B [32]32.2138.722.7148.5452.556.3739.6544.795.00
GSVA-Vicuna-7B63.3261.7056.4570.1169.2363.5061.3460.2658.42
LISA-Vicuna-7B [32] (ft)61.6361.7654.6766.2768.5050.0158.8460.6351.91
GSVA-Vicuna-7B (ft)66.4763.2962.4371.0869.9365.3162.2360.4760.56
LISA-Vicuna-13B [32]32.7339.853.6648.7653.624.8939.4945.354.41
GSVA-Vicuna-13B62.9760.1858.4467.1767.5954.6058.0657.2852.22
LISA-Vicuna-13B [32] (ft)63.4562.9955.2568.1869.6552.1661.8462.2456.15
GSVA-Vicuna-13B (ft)68.0164.0565.3671.7570.5167.2563.8361.2863.11
LISA-Llama2-13B [32]33.2639.643.2749.7653.807.2840.4945.415.73
GSVA-Llama2-13B63.2062.3854.5169.5269.8657.8462.0660.7758.30
LISA-Llama2-13B [32] (ft)65.2463.9657.4969.9971.0055.4362.1162.2956.34
GSVA-Llama2-13B (ft)70.0466.3866.0273.2972.7964.7265.4563.2062.47
", + "bbox": [ + 81, + 90, + 890, + 391 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Generalized referring expression segmentation (GRES) results on gRefCOCO [38] dataset. gIoU and cIoU are IoU metrics averaged by each example and accumulated over whole dataset, respectively. N-acc. is short for the accuracy of correctly classifying null targets. Baselines are copied from Liu et al. [38]. (ft) denotes the model is finetuned on the training set of gRefCOCO.", + "bbox": [ + 78, + 397, + 890, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "GRES task without finetuning, showing degradation in both gIoU and cIoU in each model variant. Especially the low N-acc indicates that LISA is unable to correctly reject the empty targets. When finetuned on gRefCOCO training set, GSVA-Vicuna-7B performs better than the finetuned LISA counterpart, with about $4\\%$ improvement in gIoU and over $5\\%$ in N-acc on all three evaluation splits. GSVA variants with larger LLM incorporated further push the limits, achieving over $70\\%$ in gIoU on the validation set, $73\\%$ on test set A, and $65\\%$ on test set B. The 13B models also consistently outperform LISA by large margins, demonstrating the superiority of GSVA in GRES task.", + "bbox": [ + 75, + 443, + 468, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Referring Expression Segmentation", + "text_level": 1, + "bbox": [ + 76, + 635, + 385, + 651 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Settings. To validate the abilities to handle various tasks, we evaluate GSVA in the classic RES task. Following the common evaluation protocols, we test variants of GSVA and LISA equipped with different LLMs on RefCOCO, RefCOCO+ [30], and RefCOCOg [44]. We follow the UNC split to perform experiments on RefCOCO and RefCOCO+, and UMD split for RefCOCOg. The models are firstly pretrained as in GRES, and then finetuned for 10 epochs with a joint dataset of these three RES training set. The cIoU metric is adopted to measure the model performances.", + "bbox": [ + 75, + 657, + 468, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results. Table 2 shows the RES results of GSVA. For the pretrained models, all the three variants of GSVA achieve higher cIoU than LISA [32] by clear margins. Our 7B model outperforms LISA-Vicuna about $2\\%$ cIoU on almost every data split. After finetuning, the preponderance of GSVA over LISA keeps and even enlarges. Specifically,", + "bbox": [ + 75, + 810, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the cIoU metrics of GSVA-Llama2-13B on 8 sets surpass LISA by at least $2.8\\%$ . For the test set of RefCOCOg, the margin even grows to $5.9\\%$ , which exhibits GSVA is also competitive in the classic RES task.", + "bbox": [ + 496, + 443, + 890, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Referring Expression Comprehension", + "text_level": 1, + "bbox": [ + 498, + 532, + 823, + 549 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Settings. Since GSVA is capable for RES tasks, it is natural to transfer to Referring Expression Comprehension (REC) tasks, by simply computing the bounding boxes of the masks. The datasets of REC are the same as RES, in which we evaluate GSVA. If the IoU of a predicted bounding box and the ground truth is greater than 0.5, this prediction is marked as correct. We use the same models in RES to evaluate the phrase grounding capability in REC tasks.", + "bbox": [ + 496, + 561, + 890, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results. As shown in Table 3, we mainly compare our GSVA to LISA in the three variants. Without finetuning, GSVA-Vicuna-7B outperforms LISA with a large margin over $7\\%$ in almost all evaluation sets. The similar trends also hold for the Vicuna-13B and Llama2-13B variants. GSVA also benefits a lot from finetuning, e.g., the 7B variant achieves consistently higher Prec@0.5 than uLLaVA-7B [80] with LoRA finetuning, which is another strong baseline that adopts the \"mask2bbox\" pipelines without direct bounding box supervision. The finetuned GSVA also shows competitive performances to the fully finetuned uLLaVA-7B, suggesting the strong potential of our method. With larger LLMs incorporated, the performances of GSVA continue with over $3\\%$ increments over all datasets.", + "bbox": [ + 496, + 688, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3863", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1cf1669ea7bb7e68411f85ea48c926479fb37e27ef50bc0fbc05deae57806d83.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Referring Expression Segmentation on RefCOCO, RefCOCO+, and RefCOCOg
MethodRefCOCO (UNC)RefCOCO+ (UNC)RefCOCOg (UMD)
Val.Test-ATest-BVal.Test-ATest-BVal.Test
MCN [42]62.464.259.750.655.044.749.249.4
VLT [10]67.570.565.256.361.050.155.057.7
CRIS [72]70.573.266.162.368.153.759.960.4
LAVT [81]72.775.868.862.168.455.161.262.1
ReLA [38]73.876.570.266.071.057.765.066.0
X-Decoder [93]------64.6-
SEEM [94]------65.7-
PolyFormer-L [40]76.078.373.369.374.661.969.270.2
LISA-Vicuna-7B* [32]74.176.571.162.467.456.566.468.5
GSVA-Vicuna-7B76.477.472.864.567.758.671.172.0
LISA-Vicuna-7B* [32] (ft)74.979.172.365.170.858.167.970.6
GSVA-Vicuna-7B (ft)77.278.973.565.969.659.872.773.3
LISA-Vicuna-13B [32]71.774.768.159.464.252.965.266.1
GSVA-Vicuna-13B74.677.570.562.566.555.569.671.2
LISA-Vicuna-13B [32] (ft)76.078.872.965.070.258.169.570.5
GSVA-Vicuna-13B (ft)78.280.474.267.471.560.974.275.6
LISA-Llama2-13B [32]73.476.269.562.366.656.368.268.5
GSVA-Llama2-13B77.779.974.268.071.561.573.273.9
LISA-Llama2-13B [32] (ft)76.378.772.466.271.059.370.171.1
GSVA-Llama2-13B (ft)79.281.777.170.373.863.675.777.0
", + "bbox": [ + 81, + 89, + 890, + 409 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1c6bb18d46211a074d3f1008f020e8265506f56316a3b7241ee220bfcdf77d6b.jpg", + "table_caption": [ + "Table 2. Referring expression segmentation results on RefCOCO, RefCOCO+ [30] and RefCOCOg [44] dataset. The cIoU metrics of each split are reported. Baselines are excerpted from Lai et al. [32]. (ft) denotes the models are finetuned on the joint training set of the referring expression segmentation datasets. * means the results are excerpted from the original paper." + ], + "table_footnote": [], + "table_body": "
MethodReferring Expression Comprehension on RefCOCO, RefCOCO+, and RefCOCOg
Val.RefCOCO Test-ATest-BVal.RefCOCO+ Test-ATest-BRefCOCOg Val.Test
u-LLaVA-7B [80] (LoRA)83.4787.1380.2168.7476.3260.9876.1978.24
u-LLaVA-7B [80] (full-ft)86.0489.4782.2674.0981.1666.6179.8781.68
LISA-Vicuna-7B [32]78.6881.7275.7462.9268.9356.4970.1072.47
GSVA-Vicuna-7B85.5088.0182.4970.2175.6265.1179.0079.21
LISA-Vicuna-7B [32] (ft)85.3988.8482.5974.2379.4668.4079.3480.42
GSVA-Vicuna-7B (ft)86.2789.2283.7772.8178.7868.0181.5881.83
LISA-Vicuna-13B [32]80.0183.2676.2663.7770.2457.4271.7973.34
GSVA-Vicuna-13B83.1287.0180.5468.1473.9062.0077.0878.89
LISA-Vicuna-13B [32] (ft)85.9289.0583.1674.8681.0868.8780.0981.48
GSVA-Vicuna-13B (ft)87.7190.4984.5776.5281.6970.3583.9084.85
LISA-Llama2-13B [32]82.5285.5678.8267.9173.7762.2575.3776.83
GSVA-Llama2-13B86.9989.5484.0873.8979.1069.3880.6882.07
LISA-Llama2-13B [32] (ft)85.9188.8481.7374.4680.5668.2680.0981.27
GSVA-Llama2-13B (ft)89.1692.0887.1779.7484.4573.4185.4786.18
", + "bbox": [ + 81, + 460, + 890, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Referring expression comprehension results on RefCOCO, RefCOCO+ [30] and RefCOCOg [44] dataset. The metric is precision @ 0.5 IoU threshold. (LoRA) means the LLM in u-LLaVA [80] is finetuned with LoRA adapter, as LISA and GSVA, while (full-ft) represents the LLM in u-LLaVA is fully trained. Results of u-LLaVA-7B with \"mask2bbox\" strategy are reported for fair comparison.", + "bbox": [ + 76, + 696, + 893, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 746, + 230, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The involvement of the [REJ] token. [REJ] token plays a rather important role in GSVA. We study the effect of the [REJ] token by removing it from the vocabulary in GSVA, yielding a variant unable to reject a target from the text outputs. As shown in the $2^{\\text{nd}}$ row of Table 4, after removing [REJ] token, there is a sharp N-acc drop over $25\\%$ relatively, followed by the decline of gIoU at about $10\\%$ on gRefCOCO validation set. This performance degradation", + "bbox": [ + 76, + 779, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "indicates the significance of LLM learning a special token to reject the referred instances absent in the image.", + "bbox": [ + 498, + 747, + 890, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Learning multiple [SEG] tokens. We continue to ablate the multiple [SEG] tokens, which is another core design of GSVA. After removing [REJ] token, we then reduce the number of [SEG] tokens to 1, which is identical to LISA [32] with the referring expression added before the only one [SEG] token: Assistant: Sure, {obj}:[SEG]. In the early experiments, we have found that stacking multiple expressions before one [SEG] token would result in diver", + "bbox": [ + 496, + 779, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3864", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/daf746cea54f9db87410b369a5fded0fe5eb165343920dc34d492be4a37d8a60.jpg", + "image_caption": [ + "(a) Multiple Target Cases" + ], + "image_footnote": [], + "bbox": [ + 80, + 107, + 215, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4ec7b27483a384f355e27e2cfe72aaa48709e263fe2ea74aea250884f6c9d306.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 107, + 359, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/26a9add807664ac57a40ce031e8cfa0cbfcb8a0cabd1c34520fc53fb4a2ca992.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 107, + 477, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ed01be79d76617b81ca32630c63cc333515e0ff7504186c73807b006edfddbb7.jpg", + "image_caption": [ + "Elephant in back, Left elephant", + "Kid on left in blue, Boy on right, Blond boy on right, Girl back left pink", + "Figure 4. Visualizations of GSVA and LISA [32] in the GRES task. The first row shows LISA's segmentation results, the second row is the masks and rejections of GSVA, and the third row shows the referring expressions in the instructions. In (a) multiple target cases, each target is colored with a specific color. In (b) empty target cases, the images turn darker to highlight the incorrect predictions of LISA. The examples are selected from the gRefCOCO validation set. The masks are generated by the 7B models. Zoom in for the best view." + ], + "image_footnote": [], + "bbox": [ + 80, + 178, + 215, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d172daaca66210060b87ca57f49113019b04b76b1f2a651ffd246bea7c3e41f9.jpg", + "image_caption": [ + "Tan vase, Bottom vase, Right vase" + ], + "image_footnote": [], + "bbox": [ + 223, + 178, + 359, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ec875e834067d302b43ef1aa075dabff7c15cec5af9e6b2f3eaf5532e0bd984e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 178, + 477, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/465f562d5740bd84758eb8a00204d29617174eaf3afd6cc7f9881ea051ad3854.jpg", + "image_caption": [ + "(b) Empty Target Cases" + ], + "image_footnote": [], + "bbox": [ + 483, + 107, + 620, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0a39cdcf3e583ce5afec93306543ceeef6189aa388dc51808c9ceff98d38844a.jpg", + "image_caption": [ + "Lady on left" + ], + "image_footnote": [], + "bbox": [ + 485, + 178, + 620, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/11400d90a2e5f11667b43e3559b917794630bd1b2dcba7d6930afd4bdf80229d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 627, + 107, + 766, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/638204a0c6903c24c88310a1b409bd3dd8a2eb86eb758446f56062d76e6b39c8.jpg", + "image_caption": [ + "Teddy with X eye" + ], + "image_footnote": [], + "bbox": [ + 629, + 178, + 766, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fa154cc8e320f6229a187f7766e439434cfccf48d12a0602cc780105dcb7b891.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 107, + 880, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dbb082b818b53a9e2883bff956ccc92ac2f7ebb36a68ddbaf3d98fe4e3d0284b.jpg", + "image_caption": [ + "The dog on the suitcase" + ], + "image_footnote": [], + "bbox": [ + 774, + 178, + 880, + 244 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ead0c4ade1b9d61b92890db88b80ec3a6e2745d6c83c37111648289f5bce54da.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Model w/Vicuna-7BModificationsgRefCOCO Val.
RefExp. +[SEG]Multiple [SEG][REJ] TokengIoUcIoUN-acc.
GSVA✓✓✓63.3261.7056.45
✓✓✗51.5760.9530.32
✓✗✗44.8659.3711.96
LISA [32]✗✗✗32.2138.722.71
✗✓✓21.8327.220.00
", + "bbox": [ + 81, + 339, + 468, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablation study on the core designs of GSVA. $\\checkmark$ means the employment of the component while $X$ means not. \"Ref-Exp.+[SEG]\", \"Multiple [SEG]\", and \"[REJ] Token\" are short for adding referring expression before [SEG] in the answer prompt, using multiple [SEG] tokens, involving [REJ] token, respectively.", + "bbox": [ + 75, + 452, + 468, + 522 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "gence. Therefore we separate multiple targets to prompt the model with one expression at a time. The sharp decrements of gIoU by nearly $7\\%$ and N-acc by over $15\\%$ in the $3^{\\mathrm{rd}}$ row demonstrate the significance of the multiple-[SEG]-token.", + "bbox": [ + 75, + 532, + 468, + 593 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Answers without referring expression. To examine the efficacy of the hinting prompts, we remove all the referring expressions before the [SEG] tokens. Based on the removal of multiple [SEG]s and [REJ]s, erasing the added referring expression falls back to the original LISA model, as shown in the $4^{\\text{th}}$ row. We further choose only to remove it from GSVA model, keeping other configurations unchanged. Specifically, if there are two referents, the prompts in the answer will turn to Assistant: Sure, [SEG], [SEG], whose results are in the last row. The zero N-acc shows the model fails to identify any empty target without the help of the expressions, meanwhile the poor gIoU and cIoU indicates the segmentation ability is damaged. This phenomenon also suggests that the added referring expression hint GSVA to associate each [SEG] token to its corresponding target, which is in coherence with our hypothesis.", + "bbox": [ + 75, + 593, + 468, + 835 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Visualization", + "text_level": 1, + "bbox": [ + 76, + 845, + 212, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We visualize some qualitative results of GSVA to verify its effectiveness. As shown in Figure 4, we present two groups", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "of examples from the validation set of gRefCOCO [38] to see how GSVA outperforms LISA in the face of the two main challenges in GRES: multiple targets and empty targets. In part (a), GSVA has managed to segment all the targets referred to, while LISA could only segment one of the requested instances. For example, in the third column, LISA only predicts the mask of the rightmost vase. On the contrary, GSVA separately segments all three targeted vases. In part (b), LISA mistakenly segments the instances in the image that disagree with the referring expression, e.g., in the sixth column, LISA proposes a mask of a cat in response to the request of the dog on the suitcase, whereas GSVA successfully reject all the empty targets.", + "bbox": [ + 496, + 340, + 890, + 539 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 556, + 617, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper introduces a novel multimodal large language model dubbed Generalized Segmentation Vision Assistant (GSVA). By introducing multiple [SEG] tokens and the new [REJ] token, GSVA effectively achieves multi-objective segmentation and empty target rejection, which addresses the challenging segmentation problems in practical application scenarios, e.g., Generalized Referring Expression Segmentation (GRES). Extensive experiments on GRES, classic RES, and REC tasks fully demonstrate the superior performance of our method, highlighting its significance for future research and applications.", + "bbox": [ + 496, + 583, + 890, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 500, + 768, + 658, + 784 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is supported in part by the National Key R&D Program of China under Grant 2021ZD0140407, the National Natural Science Foundation of China under Grants 62321005 and 62276150. We also thank Dr. Yuan Yao and Prof. Zhiyuan Liu for their insightful and valuable comments on this research project. We also appreciate their generous support of computational resources.", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3865", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022. 2", + "[2] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, 2020. 2, 5", + "[3] Ding-Jie Chen, Songhao Jia, Yi-Chen Lo, Hwann-Tzong Chen, and Tyng-Luh Liu. See-through-text grouping for referring image segmentation. In IEEE ICCV, 2019. 2", + "[4] Yixin Chen, Qing Li, Deqian Kong, Yik Lun Kei, Song-Chun Zhu, Tao Gao, Yixin Zhu, and Siyuan Huang. Yourefit: Embodied reference understanding with language and gesture. In IEEE ICCV, 2021. 2, 4", + "[5] Ming-Ming Cheng, Shuai Zheng, Wen-Yan Lin, Vibhav Vineet, Paul Sturgess, Nigel Crook, Niloy J Mitra, and Philip Torr. Imagesprit: Verbal guided image parsing. ACM ToG, 2014. 1, 2", + "[6] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ * chatgpt quality, 2023. 2, 3", + "[7] Yong Xien Chng, Henry Zheng, Yizeng Han, Xuchong Qiu, and Gao Huang. Mask grounding for referring image segmentation. In IEEE CVPR, 2024. 2", + "[8] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416, 2022. 2", + "[9] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. In NeurIPS, 2023. 2, 5", + "[10] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vision-language transformer and query generation for referring segmentation. In IEEE ICCV, 2021. 6, 7", + "[11] Xiaohan Ding, Xiangyu Zhang, Ningning Ma, Jungong Han, Guiguang Ding, and Jian Sun. Repvgg: Making vgg-style convnets great again. In IEEE ICCV, 2021. 2", + "[12] Xiaohan Ding, Xiangyu Zhang, Jungong Han, and Guiguang Ding. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In IEEE CVPR, 2022. 2", + "[13] Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Zhiyong Wu, Baobao Chang, Xu Sun, Jingjing Xu, and Zhifang Sui. A survey for in-context learning. arXiv preprint arXiv:2301.00234, 2022.5", + "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale. In IEEE CVPR, 2023. 2", + "[16] Guang Feng, Zhiwei Hu, Lihe Zhang, and Huchuan Lu. Encoder fusion network with co-attention embedding for referring image segmentation. In IEEE CVPR, 2021. 2", + "[17] Chen Gao, Jinyu Chen, Si Liu, Luting Wang, Qiong Zhang, and Qi Wu. Room-and-object aware knowledge reasoning for remote embodied referring expression. In IEEE CVPR, 2021. 1, 4", + "[18] Dongchen Han, Tianzhu Ye, Yizeng Han, Zhuofan Xia, Shiji Song, and Gao Huang. Agent attention: On the integration of softmax and linear attention. arXiv preprint arXiv:2312.08874, 2023. 2", + "[19] Yizeng Han, Gao Huang, Shiji Song, Le Yang, Honghui Wang, and Yulin Wang. Dynamic neural networks: A survey. IEEE TPAMI, 2021. 2", + "[20] Yizeng Han, Gao Huang, Shiji Song, Le Yang, Yitian Zhang, and Haojun Jiang. Spatially adaptive feature refinement for efficient inference. IEEE TIP, 2021. 2", + "[21] Yizeng Han, Yifan Pu, Zihang Lai, Chaofei Wang, Shiji Song, Junfeng Cao, Wenhui Huang, Chao Deng, and Gao Huang. Learning to weight samples for dynamic early-exiting networks. In ECCV, 2022.", + "[22] Yizeng Han, Zhihang Yuan, Yifan Pu, Chenhao Xue, Shiji Song, Guangyu Sun, and Gao Huang. Latency-aware spatialwise dynamic networks. In NeurIPS, 2022. 2", + "[23] Yizeng Han, Dongchen Han, Zeyu Liu, Yulin Wang, Xuran Pan, Yifan Pu, Chao Deng, Junlan Feng, Shiji Song, and Gao Huang. Dynamic perceiver for efficient visual recognition. In IEEE ICCV, 2023. 2", + "[24] Yizeng Han, Zeyu Liu, Zhihang Yuan, Yifan Pu, Chaofei Wang, Shiji Song, and Gao Huang. Latency-aware unified dynamic networks for efficient image recognition. arXiv preprint arXiv:2308.15949, 2023. 2", + "[25] Ronghang Hu, Marcus Rohrbach, and Trevor Darrell. Segmentation from natural language expressions. In ECCV, 2016. 1, 2", + "[26] Yutao Hu, Qixiong Wang, Wenqi Shao, Enze Xie, Zhenguo Li, Jungong Han, and Ping Luo. Beyond one-to-one: Rethinking the referring image segmentation. In IEEE ICCV, 2023. 2", + "[27] Gao Huang, Yulin Wang, Kangchen Lv, Haojun Jiang, Wenhui Huang, Pengfei Qi, and Shiji Song. Glance and focus networks for dynamic visual recognition. IEEE TPAMI, 2022. 2", + "[28] Rui Huang, Xuran Pan, Henry Zheng, Haojun Jiang, Zhifeng Xie, Cheng Wu, Shiji Song, and Gao Huang. Joint representation learning for text and 3d point cloud. Pattern Recognition, 2024. 1", + "[29] Ya Jing, Tao Kong, Wei Wang, Liang Wang, Lei Li, and Tie-niu Tan. Locate then segment: A strong pipeline for referring image segmentation. In IEEE CVPR, 2021. 6", + "[30] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In EMNLP, 2014. 1, 4, 6, 7" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3866", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In IEEE ICCV, 2023. 2, 3", + "[32] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692, 2023. 1, 2, 3, 4, 5, 6, 7, 8", + "[33] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Fanyi Pu, Jingkang Yang, Chunyuan Li, and Ziwei Liu. Mimicit: Multi-modal in-context instruction tuning. arXiv preprint arXiv:2306.05425, 2023. 5", + "[34] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 2, 5", + "[35] Ruiyu Li, Kaican Li, Yi-Chun Kuo, Michelle Shu, Xiaojuan Qi, Xiaoyong Shen, and Jiaya Jia. Referring image segmentation via recurrent refinement networks. In IEEE CVPR, 2018. 2", + "[36] Ziyi Lin, Chris Liu, Renrui Zhang, Peng Gao, Longtian Qiu, Han Xiao, Han Qiu, Chen Lin, Wenqi Shao, Keqin Chen, et al. Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models. arXiv preprint arXiv:2311.07575, 2023. 2", + "[37] Chenxi Liu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, and Alan Yuille. Recurrent multimodal interaction for referring image segmentation. In IEEE CVPR, 2017. 2", + "[38] Chang Liu, Henghui Ding, and Xudong Jiang. Gres: Generalized referring expression segmentation. In IEEE CVPR, 2023. 1, 2, 4, 5, 6, 7, 8", + "[39] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 2, 3, 5", + "[40] Jiang Liu, Hui Ding, Zhaowei Cai, Yuting Zhang, Ravi Kumar Satzoda, Vijay Mahadevan, and R Manmatha. Polyformer: Referring image segmentation as sequential polygon generation. In IEEE CVPR, 2023. 7", + "[41] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 2", + "[42] Gen Luo, Yiyi Zhou, Xiaoshuai Sun, Liujuan Cao, Chenglin Wu, Cheng Deng, and Rongrong Ji. Multi-task collaborative network for joint referring expression comprehension and segmentation. In IEEE CVPR, 2020. 7", + "[43] Tengchao Lv, Yupan Huang, Jingye Chen, Lei Cui, Shuming Ma, Yaoyao Chang, Shaohan Huang, Wenhui Wang, Li Dong, Weiyao Luo, et al. Kosmos-2.5: A multimodal literate model. arXiv preprint arXiv:2309.11419, 2023. 2", + "[44] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In IEEE CVPR, 2016. 2, 4, 6, 7", + "[45] Khanh Nguyen, Debadeepta Dey, Chris Brockett, and Bill Dolan. Vision-based navigation with language-based assistance via imitation learning with indirect intervention. In IEEE CVPR, 2019. 4" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Zanlin Ni, Yulin Wang, Jiangwei Yu, Haojun Jiang, Yue Cao, and Gao Huang. Deep incubation: Training large models by divide-and-conquering. In IEEE CVPR, 2023. 2", + "[47] Zanlin Ni, Yulin Wang, Renping Zhou, Jiayi Guo, Jinyi Hu, Zhiyuan Liu, Shiji Song, Yuan Yao, and Gao Huang. Revisiting non-autoregressive transformers for efficient image synthesis. In CVPR, 2024. 2", + "[48] OpenAI. Gpt-4 technical report, 2023. 2", + "[49] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. In NeurIPS, 2022. 2", + "[50] Xuran Pan, Chunjiang Ge, Rui Lu, Shiji Song, Guanfu Chen, Zeyi Huang, and Gao Huang. On the integration of self-attention and convolution. In IEEE CVPR, 2022. 2", + "[51] Xuran Pan, Tianzhu Ye, Zhuofan Xia, Shiji Song, and Gao Huang. Slide-transformer: Hierarchical vision transformer with local self-attention. In IEEE CVPR, 2023. 2", + "[52] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 2, 5", + "[53] Yifan Pu, Yiru Wang, Zhuofan Xia, Yizeng Han, Yulin Wang, Weihao Gan, Zidong Wang, Shiji Song, and Gao Huang. Adaptive rotated convolution for rotated object detection. In IEEE ICCV, 2023. 2", + "[54] Yuankai Qi, Qi Wu, Peter Anderson, Xin Wang, William Yang Wang, Chunhua Shen, and Anton van den Hengel. Reverie: Remote embodied visual referring expression in real indoor environments. In IEEE CVPR, 2020. 1, 4", + "[55] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3", + "[56] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M Anwer, Erix Xing, Ming-Hsuan Yang, and Fahad S Khan. Glamm: Pixel grounding large multimodal model. arXiv preprint arXiv:2311.03356, 2023. 2", + "[57] Hengcan Shi, Hongliang Li, Fanman Meng, and Qingbo Wu. Key-word-aware network for referring expression image segmentation. In ECCV, 2018. 2", + "[58] Qie Sima, Sinan Tan, Huaping Liu, Fuchun Sun, Weifeng Xu, and Ling Fu. Embodied referring expression for manipulation question answering in interactive environment. In IEEE ICRA, 2023. 1, 4", + "[59] Yan Tai, Weichen Fan, Zhao Zhang, Feng Zhu, Rui Zhao, and Ziwei Liu. Link-context learning for multimodal llms. arXiv preprint arXiv:2308.07891, 2023. 5", + "[60] Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. Stanford alpaca: An instruction-following llama model, 2023. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3867", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[61] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023.", + "[62] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2", + "[63] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, et al. Image as a foreign language: Beit pretraining for vision and vision-language tasks. In IEEE CVPR, 2023. 2", + "[64] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, et al. Vision: Large language model is also an open-ended decoder for vision-centric tasks. In NeurlPS, 2023. 2", + "[65] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, et al. Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079, 2023. 2", + "[66] Xin Wang, Qiuyuan Huang, Asli Celikyilmaz, Jianfeng Gao, Dinghan Shen, Yuan-Fang Wang, William Yang Wang, and Lei Zhang. Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation. In IEEE CVPR, 2019. 1, 4", + "[67] Yulin Wang, Zhaoxi Chen, Haojun Jiang, Shiji Song, Yizeng Han, and Gao Huang. Adaptive focus for efficient video recognition. In IEEE ICCV, 2021. 2", + "[68] Yulin Wang, Rui Huang, Shiji Song, Zeyi Huang, and Gao Huang. Not all images are worth 16x16 words: Dynamic transformers for efficient image recognition. In NeurIPS, 2021.", + "[69] Yulin Wang, Yang Yue, Yuanze Lin, Haojun Jiang, Zihang Lai, Victor Kulikov, Nikita Orlov, Humphrey Shi, and Gao Huang. Adafocus v2: End-to-end training of spatial dynamic networks for video recognition. In IEEE CVPR, 2022.", + "[70] Yulin Wang, Yang Yue, Xinhong Xu, Ali Hassani, Victor Kulikov, Nikita Orlov, Shiji Song, Humphrey Shi, and Gao Huang. Adafocusv3: On unified spatial-temporal dynamic video recognition. In ECCV, 2022. 2", + "[71] Yulin Wang, Yang Yue, Rui Lu, Tianjiao Liu, Zhao Zhong, Shiji Song, and Gao Huang. Efficienttrain: Exploring generalized curriculum learning for training visual backbones. In IEEE ICCV, 2023. 2", + "[72] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clip-driven referring image segmentation. In IEEE CVPR, 2022. 6, 7", + "[73] Zhichao Wei, Xiaohao Chen, Mingqiang Chen, and Siyu Zhu. Learning aligned cross-modal representations for referring image segmentation. arXiv preprint arXiv:2301.06429, 2023. 2", + "[74] Penghao Wu and Saining Xie. V*: Guided visual search as a core mechanism in multimodal llms. arXiv preprint arXiv:2312.14135, 2023. 2" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[75] Tsung-Han Wu, Giscard Biamby, David Chan, Lisa Dunlap, Ritwik Gupta, Xudong Wang, Joseph E Gonzalez, and Trevor Darrell. See, say, and segment: Teaching Imms to overcome false premises. arXiv preprint arXiv:2312.08366, 2023. 2", + "[76] Zhuofan Xia, Xuran Pan, Shiji Song, Li Erran Li, and Gao Huang. Vision transformer with deformable attention. In IEEE CVPR, 2022. 2", + "[77] Zhuofan Xia, Xuran Pan, Xuan Jin, Yuan He, Hui Xue, Shiji Song, and Gao Huang. Budgeted training for vision transformer. In ICLR, 2023. 2", + "[78] Zhuofan Xia, Xuran Pan, Shiji Song, Li Erran Li, and Gao Huang. Dat++: Spatially dynamic vision transformer with deformable attention. arXiv preprint arXiv:2309.01430, 2023.2", + "[79] Bin Xiao, Haiping Wu, Weijian Xu, Xiyang Dai, Houdong Hu, Yumao Lu, Michael Zeng, Ce Liu, and Lu Yuan. Florence-2: Advancing a unified representation for a variety of vision tasks. arXiv preprint arXiv:2311.06242, 2023. 2", + "[80] Jinjin Xu, Liwu Xu, Yuzhe Yang, Xiang Li, Yanchun Xie, Yi-Jie Huang, and Yaqian Li. u-llava: Unifying multimodal tasks via large language model. arXiv preprint arXiv:2311.05348, 2023. 2, 6, 7", + "[81] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Lavt: Language-aware vision transformer for referring image segmentation. In IEEE CVPR, 2022. 2, 6, 7", + "[82] Linwei Ye, Mrigank Rochan, Zhi Liu, and Yang Wang. Cross-modal self-attention network for referring image segmentation. In IEEE CVPR, 2019. 2", + "[83] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 2", + "[84] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In ECCV, 2016. 4, 5", + "[85] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In IEEE CVPR, 2018. 6", + "[86] Ao Zhang, Liming Zhao, Chen-Wei Xie, Yun Zheng, Wei Ji, and Tat-Seng Chua. Next-chat: An Imm for chat, detection and segmentation. arXiv preprint arXiv:2311.04498, 2023. 2", + "[87] Renrui Zhang, Jiaming Han, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, Peng Gao, and Yu Qiao. Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023. 2", + "[88] Shilong Zhang, Peize Sun, Shoufa Chen, Min Xiao, Wenqi Shao, Wenwei Zhang, Kai Chen, and Ping Luo. Gpt4roi: Instruction tuning large language model on region-of-interest. arXiv preprint arXiv:2307.03601, 2023. 5" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "3868", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[89] Zicheng Zhang, Yi Zhu, Jianzhuang Liu, Xiaodan Liang, and Wei Ke. Coupalign: Coupling word-pixel with sentence-mask alignments for referring image segmentation. In NeurIPS, 2022. 2", + "[90] Haozhe Zhao, Zefan Cai, Shuzheng Si, Xiaojian Ma, Kaikai An, Liang Chen, Zixuan Liu, Sheng Wang, Wenjuan Han, and Baobao Chang. Mmicl: Empowering vision-language model with multi-modal in-context learning. arXiv preprint arXiv:2309.07915, 2023. 5", + "[91] Yang Zhao, Zhijie Lin, Daquan Zhou, Zilong Huang, Jiashi Feng, and Bingyi Kang. Bubogpt: Enabling visual grounding in multi-modal llms. arXiv preprint arXiv:2307.08581, 2023. 5", + "[92] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 5", + "[93] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In IEEE CVPR, 2023. 2, 7", + "[94] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. In NeurIPS, 2023. 2, 7" + ], + "bbox": [ + 78, + 90, + 468, + 430 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "3869", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2024/GSVA_ Generalized Segmentation via Multimodal Large Language Models/12e5e8f5-7030-48a1-9213-1240824636f7_model.json b/2024/GSVA_ Generalized Segmentation via Multimodal Large Language Models/12e5e8f5-7030-48a1-9213-1240824636f7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4390529660137f74bc004518de6fff71df0b9f42 --- /dev/null +++ b/2024/GSVA_ Generalized Segmentation via Multimodal Large Language Models/12e5e8f5-7030-48a1-9213-1240824636f7_model.json @@ -0,0 +1,2798 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.108, + 0.131, + 0.862, + 0.154 + ], + "angle": 0, + "content": "GSVA: Generalized Segmentation via Multimodal Large Language Models" + }, + { + "type": "text", + "bbox": [ + 0.138, + 0.181, + 0.837, + 0.201 + ], + "angle": 0, + "content": "Zhuofan Xia* Dongchen Han* Yizeng Han Xuran Pan Shiji Song Gao Huang†" + }, + { + "type": "text", + "bbox": [ + 0.294, + 0.202, + 0.673, + 0.217 + ], + "angle": 0, + "content": "Department of Automation, BNRist, Tsinghua University" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.252, + 0.314, + 0.267 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.283, + 0.473, + 0.722 + ], + "angle": 0, + "content": "Generalized Referring Expression Segmentation (GRES) extends the scope of classic RES to refer to multiple objects in one expression or identify the empty targets absent in the image. GRES poses challenges in modeling the complex spatial relationships of the instances in the image and identifying non-existing referents. Multimodal Large Language Models (MLLMs) have recently shown tremendous progress in these complicated vision-language tasks. Connecting Large Language Models (LLMs) and vision models, MLLMs are proficient in understanding contexts with visual inputs. Among them, LISA, as a representative, adopts a special [SEG] token to prompt a segmentation mask decoder; e.g., SAM, to enable MLLMs in the RES task. However, existing solutions to GRES remain unsatisfactory since current segmentation MLLMs cannot correctly handle the cases where users might reference multiple subjects in a singular prompt or provide descriptions incongruent with any image target. In this paper, we propose Generalized Segmentation Vision Assistant (GSVA) to address this gap. Specifically, GSVA reuses the [SEG] token to prompt the segmentation model towards supporting multiple mask references simultaneously and innovatively learns to generate a [REJ] token to reject the null targets explicitly. Experiments validate GSVA's efficacy in resolving the GRES issue, marking a notable enhancement and setting a new record on the GRES benchmark gRefCOCO dataset. GSVA also proves effective across various classic referring segmentation and comprehension tasks. Code is available at https://github.com/LeapLabTHU/GSVA." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.75, + 0.208, + 0.765 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.775, + 0.47, + 0.866 + ], + "angle": 0, + "content": "Referring Expression Segmentation (RES) [5, 25] is an emerging vision-language (VL) task predicting the masks of the interested objects referred to in the language expression. RES has great potential in many areas, especially embodied AI [17, 28, 54, 58, 66], including VL navigation, humanrobot interaction, etc. Nevertheless, the simplification in" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.253, + 0.885, + 0.451 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.462, + 0.892, + 0.602 + ], + "angle": 0, + "content": "Figure 1. Comparison of the segmentation masks by LISA [32] and GSVA, facing the challenges in Generalized Referring Expression Segmentation (GRES) [38]. (a) LISA fails to segment the correct targets when multiple targets are requested due to the single [SEG] token restriction. GSVA successfully generates all target masks via learning multiple [SEG] tokens. (b) When the referent does not exist in the image, i.e., the empty target is requested, LISA reluctantly produces the wrong mask because of the compulsive [SEG] token output. In contrast, GSVA can reject the empty targets by predicting [REJ] tokens in the output sequence." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.611, + 0.892, + 0.672 + ], + "angle": 0, + "content": "RES formulation that one referring expression must match an individual object in the image [30] has left a gap between current RES algorithms and real-world applications, neglecting multiple-target and empty-target cases." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.893, + 0.901 + ], + "angle": 0, + "content": "To bridge this gap, Generalized Referring Expression Segmentation (GRES) [38] has recently been proposed to support multiple-target and empty-target cases. Practically, users refer to multiple subjects within a single instruction or provide descriptions that do not correspond to any targets in the image. As an example shown in Figure 1 (a), the left man and the second man from the right are targeted simultaneously, while a banana is referred to in a scene of apples in Figure 1 (b). RES takes no account of these cases, which GRES handles. In addition to the multimodal correspondences between images and text prompts in classic RES, GRES poses new challenges in handling more complicated multiple-target and empty-target cases. Therefore, the models must handle complex spatial relationships of the instances in the image [38] to segment the targets at various" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.208, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.888, + 0.222, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.222, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3858" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.461, + 0.108 + ], + "angle": 0, + "content": "locations and reject the empty targets in the wrong places." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.11, + 0.47, + 0.351 + ], + "angle": 0, + "content": "The recent blooming Multimodal Large Language Models (MLLMs) [1, 34, 39, 83, 87] meet the requirements of GRES since they show excellency in complex reasoning [8] and instruction following [49] with visual inputs by aligning the LLMs [2, 6, 48, 60-62] and Visual Foundation Models (VFMs) [15, 55, 63, 79] which are typically various Vision Transformers [14, 18, 23, 50, 51, 76, 78] to perceive image or video inputs. To support segmentation output, many works [74, 75, 86] link an MLLM (e.g., LLaVA [39]) and a segmentation model (e.g., SAM [31]) by prompting the decoder with special token embeddings (e.g., [SEG] in LISA [32]) to generate masks of the referents in the user's instructions. Although these models manage to handle RES, GRES is still beyond their reach. As shown in Figure 1, LISA fails to work well in GRES where the multiple-target and empty-target challenges remain uncharted." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.355, + 0.473, + 0.732 + ], + "angle": 0, + "content": "To address the above challenges, we propose Generalized Segmentation Vision Assistant (GSVA). We attribute the vulnerability of other segmentation MLLMs in GRES to the single constant [SEG] token that restricts its flexibility. Therefore, we present two pivotal designs in GSVA: (1) learning to predict multiple [SEG] tokens to segment multiple targets; (2) rejecting empty targets in referring expressions by predicting [REJ] tokens. Specifically, when multiple targets are requested in the referring expression, we place multiple weight-sharing [SEG] tokens corresponding to the entities in the expression, encouraging the MLLM to learn to output multiple [SEG] tokens. To distinguish each [SEG] token and avoid ambiguity, we add the expression of each entity in front of the corresponding [SEG] token, hinting each [SEG] token to focus on the specific target, which can be regarded as implicit In-Context Learning, and dynamic neural network [19]. Meanwhile, if the referents are absent in the image, the corresponding [SEG] tokens after the prompts are altered to [REJ] tokens to identify empty targets. The predicted [REJ] tokens are directly assigned with empty masks without decoded, which liberates the segmentation model from seeking non-existing targets in the image. This Benefiting from these novel designs, GSVA takes a big step forward in addressing GRES challenges, as shown in the second row of Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Our contributions are summarized as follows: (1) We propose GSVA to solve the GRES problem with MLLM by handling the spatial relationships among targets, and study the GRES problem systematically in the context of LLM for the first time. (2) We propose the non-trivial shared-weight multiple [SEG] tokens guided by each referent prompt to address the multiple-target problem. (3) We firstly propose a clean solution, the [REJ] token, to reject the empty targets, which can be seamlessly applied to various models. (4) GSVA is intuitive and effective, achieving state-of-the-art performance on the GRES benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.09, + 0.651, + 0.107 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.116, + 0.892, + 0.344 + ], + "angle": 0, + "content": "RES and GRES. Referring Expression Segmentation (RES) [4, 5, 25, 44] assumes that one expression matches one existing target, and many works explore fusing image and language [3, 16, 35, 37, 81] to segment objects under instructions. Currently, most RES methods adopt the cross-attention module or cross-modal alignment to bridge the modality gap [7, 57, 73, 82, 89]. Another line of research enables the text prompts for segmentation model with a unified decoder [31, 41, 93, 94], offering more flexible outputs. To break the jail for the arbitrary number of targets, DMMI [26] focuses on the one-to-many setting where text expression refers to varying numbers of targets. ReLA [38] proposes the Generalized Referring Expression Segmentation (GRES) task, supporting both the multi-target and empty-target scenarios, which is our main research scope." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.344, + 0.893, + 0.585 + ], + "angle": 0, + "content": "MLLM. Multimodal Large Language Models (MLLMs) align the vision and language modalities by various techniques, including cross-attention module [1], prompt tuning tokens [87], Q-Former [9, 34], linear projection layers [39], and unified model architectures [52, 79]. Endowed with unprecedented capabilities in reasoning with world-knowledge and following instructions of users, MLLM shows extraordinary performances in various vision language tasks [43, 65]. Equipped with the segmentation decoders or detection heads, MLLMs can also excel in the vision-centric tasks, such as object detection and segmentation [36, 56, 64, 80, 86]. Among them, LISA [32] makes the most of the reasoning ability with a [SEG] token to address the Reasoning Segmentation problem. However, LISA fails to tackle the challenge in GRES due to the inflexible [SEG] token, which is addressed by our proposed GSVA." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.587, + 0.892, + 0.723 + ], + "angle": 0, + "content": "Dynamic Networks. Dynamic neural networks [19] can adapt their architectures [20-22, 24, 27, 67-70] or parameters [11, 12, 53] to different inputs or switch the computation architecture in adjustment to different different time steps [46, 47, 71, 77], in order to achieve better accuracy and efficiency. In GSVA, the weight-sharing [SEG] tokens adapt to multiple targets under the hint of the prepended target prompts and dynamically reject empty targets with an individual prediction of rejection tokens." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.739, + 0.885, + 0.756 + ], + "angle": 0, + "content": "3. Generalized Segmentation Vision Assistant" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this section, we initiate with the introduction of the model design of Generalized Segmentation Vision Assistant (GSVA), which is followed by the analysis of some certain limitations of LISA in Generalized Referring Segmentation (GRES). Subsequently, we delve into the introduction of two pivotal elements of GSVA, segmenting multiple targets and learning the rejection token. These components are fundamental in the conceptualization and construction of GSVA." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3859" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.09, + 0.891, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.361, + 0.893, + 0.418 + ], + "angle": 0, + "content": "Figure 2. Overview of GSVA. At the bottom of the figure, the MLLM encodes the input image and concatenates the tokenized text tokens to follow instructions. GSVA generates multiple [SEG] tokens to handle multiple referred targets and rejects the objects absent in the image through [REJ] tokens. At the top of the figure, the SFM also encodes the image for segmentation and selects all [SEG] tokens in the output sequence to prompt the mask decoder to segment the target objects referred to in the instructions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.425, + 0.266, + 0.44 + ], + "angle": 0, + "content": "3.1. Model Architecture" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.449, + 0.47, + 0.614 + ], + "angle": 0, + "content": "The architecture of GSVA is illustrated in Figure 2, resembling LISA [32], which enables high-fidelity segmentation outputs by integrating two types of foundation models: (1) Multimodal Large Language Model (MLLM) as an aligned vision-language cognitive module; (2) Segmentation Foundation Model (SFM) to segment the target out of the input image based on user's instruction. To connect these two modules, LISA proposes a paradigm named embedding as mask where an extra [SEG] token is appended to the vocabularies of the MLLM and serves as the prompt of the SFM to segment the target following the intention of the user." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.617, + 0.47, + 0.784 + ], + "angle": 0, + "content": "Multimodal Large Language Model. The MLLM consists of a decoder-based language model \\( F_{\\mathrm{LLM}} \\) to auto-regressively generate text responses following the user's inputs, a vision encoder \\( F_{\\mathrm{V1}} \\) to extract features from the input image, and a linear projector \\( \\phi \\) to align the representations between image and text modalities. Specifically, the pretrained LLaVA [39] variants with CLIP-ViT-L/14 [55] and Vicuna-7B/13B [6] are employed. Given an input image \\( \\mathbf{x}_{\\mathrm{img}} \\), the vision encoder \\( F_{\\mathrm{V1}} \\) first encodes it into image features, and then the projector \\( \\phi \\) maps the features into the visual token embeddings in the LLM input space:" + }, + { + "type": "equation", + "bbox": [ + 0.198, + 0.788, + 0.47, + 0.806 + ], + "angle": 0, + "content": "\\[\n\\mathbf {h} _ {\\mathrm {i m g}} = \\phi \\left(F _ {\\mathrm {V} 1} \\left(\\mathbf {x} _ {\\mathrm {i m g}}\\right)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.811, + 0.471, + 0.9 + ], + "angle": 0, + "content": "where the input image \\(\\mathbf{x}_{\\mathrm{img}}\\) is typically resized to \\(h\\times w\\times 3\\) and the image tokens \\(\\mathbf{h}_{\\mathrm{img}}\\in \\mathbb{R}^{n_{\\mathrm{img}}\\times d}\\) is aligned with the language modality. For CLIP-ViT-L/14, the input image with \\(h = w = 224\\) is encoded with ViT of patch size in 14, therefore the length of tokens \\(n_{\\mathrm{img}} = hw / 14^2 = 256\\) and the LLM dimensions \\(d\\) are 4096 and 5120 for Vicuna" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.426, + 0.893, + 0.47 + ], + "angle": 0, + "content": "7B/13B, respectively. Along with the input image, the text instructions describing the targets to segment are tokenized into text tokens by the LLM tokenizer \\( T \\):" + }, + { + "type": "equation", + "bbox": [ + 0.644, + 0.474, + 0.891, + 0.49 + ], + "angle": 0, + "content": "\\[\n\\mathbf {h} _ {\\mathrm {t x t}} = T (\\mathbf {x} _ {\\mathrm {t x t}}). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.494, + 0.892, + 0.569 + ], + "angle": 0, + "content": "The image tokens and text tokens are concatenated together and then fed into the LLM after prepending a sequence of fixed prompt tokens \\(\\mathbf{h}_{\\mathrm{prompt}}\\) (omitted in the figure). The output token embeddings \\(\\tilde{\\mathbf{y}}_{\\mathrm{txt}}\\) are generated autoregressively:" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.576, + 0.891, + 0.593 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {y}} _ {\\mathrm {t x t}} = F _ {\\mathrm {L L M}} \\left(\\left[ \\mathbf {h} _ {\\text {p r o m p t}} \\right| \\left| \\mathbf {h} _ {\\text {i m g}} \\right| \\left| \\mathbf {h} _ {\\text {t x t}} \\right|\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.596, + 0.891, + 0.64 + ], + "angle": 0, + "content": "where \\( || \\) is concatenation operation. The text responses are obtained from \\( \\tilde{\\mathbf{y}}_{\\mathrm{txt}} \\) by applying a linear classifier to predict the next words in the vocabulary." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.642, + 0.892, + 0.746 + ], + "angle": 0, + "content": "In LISA, a special token [SEG] is appended in the vocabulary to activate the segmentation ability of MLLM. The model learns to predict [SEG] token in the output sequence to indicate there is a target to segment. LISA then selects the output embedding of the [SEG] token \\(\\tilde{\\mathbf{y}}_{\\mathrm{xt}}[\\mathrm{SEG}]\\), and projects it into the prompt space of the SFM by an MLP projector \\(\\psi\\):" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.746, + 0.891, + 0.762 + ], + "angle": 0, + "content": "\\[\n\\mathbf {h} _ {\\mathrm {s e g}} = \\psi (\\tilde {\\mathcal {Y}} _ {\\mathrm {t x t}} [ \\mathrm {S E G} ]). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.891, + 0.795 + ], + "angle": 0, + "content": "The segmentation model is hence ready to decode the target mask from the query token \\(\\mathbf{h}_{\\mathrm{seg}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Segmentation Foundation Model. The SFM is a query-based segmentation model, where a frozen vision encoder \\( F_{\\mathrm{V2}} \\) takes in images of higher resolution than the vision encoder in MLLM to keep more details, followed by a trainable mask decoder \\( F_{\\mathrm{mask}} \\) to decode masks from the queries. The pretrained SAM [31] with ViT-H backbone is instantiated as the SFM to produce high-quality masks. The given" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3860" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "input image \\(\\mathbf{x}_{\\mathrm{img}}\\) is resized to a larger resolution \\(H\\times W\\times 3\\) with \\(H = W = 1024\\) and encoded with \\(F_{\\mathrm{V2}}\\) to extract features \\(\\mathbf{f}_{\\mathrm{seg}}\\) for segmentation:" + }, + { + "type": "equation", + "bbox": [ + 0.213, + 0.14, + 0.469, + 0.157 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {\\text {s e g}} = F _ {\\mathrm {V} 2} \\left(\\mathbf {x} _ {\\text {i m g}}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.161, + 0.469, + 0.208 + ], + "angle": 0, + "content": "Condition on the features \\(\\mathbf{f}_{\\mathrm{seg}} \\in \\mathbb{R}^{\\frac{H}{16} \\times \\frac{W}{16} \\times C}\\) with \\(C = 256\\) for SAM, the mask decoder \\(F_{\\mathrm{mask}}\\) decodes the segmentation masks from the segmentation queries \\(\\mathbf{h}_{\\mathrm{seg}} \\in \\mathbb{R}^{N_{\\mathrm{seg}} \\times C}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.211, + 0.469, + 0.228 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {y}} _ {\\text {m a s k}} = F _ {\\text {m a s k}} \\left(\\mathbf {h} _ {\\text {s e g}} | \\mathbf {f} _ {\\text {s e g}}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.232, + 0.468, + 0.261 + ], + "angle": 0, + "content": "where each query in \\(\\mathbf{h}_{\\mathrm{seg}}\\) corresponds to one segmentation mask in \\(\\tilde{\\mathbf{y}}_{\\mathrm{mask}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.263, + 0.469, + 0.459 + ], + "angle": 0, + "content": "LISA assumes that only one target exists to segment in the input image and its corresponding instructions. However, in GSVA, we extend it to a new scenario with multiple targets and empty targets, including multiple [SEG] tokens to invoke segmentation and [REJ] tokens to reject unreasonable instructed targets absent in the image. As shown in Figure 2, GSVA supports multiple [SEG]/[REJ] tokens in the output sequence and selects all the [SEG] tokens and discards every [REJ] token after attaining \\(\\tilde{\\mathbf{y}}_{\\mathrm{txt}}\\) in Eq. (3). Therefore, there is more than one query in \\(\\mathbf{h}_{\\mathrm{seg}}\\), thus enabling the SFM to segment multiple targets. These designs make GSVA competent in the GRES task, where the awareness of multiple and empty targets is of vital importance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.469, + 0.332, + 0.485 + ], + "angle": 0, + "content": "3.2. GRES: Task and Challenges" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.492, + 0.469, + 0.749 + ], + "angle": 0, + "content": "Task. Generalized Referring Expression Segmentation (GRES) [38] removes the constraint on the number of referred targets in the expression in the conventional Referring Expression Segmentation (RES) [30, 44, 84]. Different from that one expression only refers to one instance or region in RES, GRES allows arbitrary numbers of referred targets, including multiple instances or no target circumstances. In GRES, the user can refer to many instances simultaneously or include the objects that do not exist in the image. For instance, there are three referring expressions in Figure 2, including a plate of broccoli and a spoon in the bowl of sauce, vase bottom left corner, and soup in bowl right and bowl bottom left corner. In the GRES case, the model ought to segment the masks of the objects referred to in the \\(1^{\\text{st}}\\) and \\(3^{\\text{rd}}\\) expression, meanwhile producing an empty mask for the \\(2^{\\text{nd}}\\) expressions since there are no vases present at the bottom left corner." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Challenges. The challenges in GRES are common in practice, especially in embodied AI [17, 54, 58, 66]. The one challenge is multiple targets. Take robot navigation and planning [4, 45, 66] as an example, a robot may be asked to perceive multiple targets in the surrounding environment, e.g., to bring the two bowls of soup in Figure 2. The vision system of the robot needs to locate and segment the containers holding the referred food one at a time. The other challenge is empty target. Suppose the robot is ordered to cut an apple with a knife in the scene of Figure 2, whereas" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.181 + ], + "angle": 0, + "content": "no apple is in the view of the camera, the vision system has to identify that the referred object is not in the scene. If it relies on some conventional RES methods which assume the expression must match something in the image, the output of the vision system could be undefined and potentially dangerous in some real-world cases." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.183, + 0.892, + 0.41 + ], + "angle": 0, + "content": "Differences from ReasonSeg. Reasoning Segmentation (ReasonSeg) proposed by LISA [32] emphasizes the complex text instructions in RES. In ReasonSeg, the instructions are more implicit and sophisticated, forcing models to reason using world knowledge. Besides, the logic chain is usually longer and more challenging in ReasonSeg, which requires the model to deduce the final target object referred to in the image. In contrast, GRES increases complexity in another dimension by involving complicated spatial relationships. Hence, the model has to learn to handle this spatial information and understand the relationships between the instances. To meet these requirements, LISA tunes MLLMs with complex instructions paired with masks, while GSVA arouses the spatial modeling capabilities of MLLMs by learning multiple targets and rejecting empty targets." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.417, + 0.871, + 0.433 + ], + "angle": 0, + "content": "3.3. Multiple [SEG] Tokens for Multiple Targets" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.44, + 0.892, + 0.501 + ], + "angle": 0, + "content": "Single [SEG] token. LISA [32] follows the classic RES methods to generate a segmentation mask under given instructions by adding the [SEG] token into the answer. The prompt is formatted as:" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.503, + 0.892, + 0.533 + ], + "angle": 0, + "content": "User: What is {obj} in this image? Please output segmentation mask. Assistant: Sure, it is [SEG]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.892, + 0.67 + ], + "angle": 0, + "content": "In the above prompt, \\(\\{obj\\}\\) represents an instance referred to or some semantic area to segment, and the embeddings of the [SEG] token output from the LLM are projected to prompt the SFM. When multiple instances are requested simultaneously, this prompt would confuse the model since only one [SEG] token is forced to match several targets. As shown in Figure 1 (a), LISA coercively predicts the masks of the left man and the guy second from the right, leading to a tattered mask and masking the wrong location." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.671, + 0.892, + 0.807 + ], + "angle": 0, + "content": "Multiple [SEG] tokens. To mitigate this issue in GRES, we relax this constraint for GSVA to support multiple target outputs via learning multiple [SEG] tokens. To avoid the ambiguity between the [SEG] tokens and the corresponding objects, we prepend the referring expression before each [SEG] token, i.e., each segmentation prompt in the output text is \\(\\{obj\\_n\\} :[SEG]\\), as shown in Figure 3 (a). Streaming all the pairs of object and [SEG] token into one sentence, the question-answer prompt is formatted as (e.g., two targets):" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.809, + 0.892, + 0.837 + ], + "angle": 0, + "content": "User: What are {obj1}, {obj2} in this image? Please output the segmentation masks." + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.839, + 0.803, + 0.853 + ], + "angle": 0, + "content": "Assistant: Sure, \\(\\{obj1\\} :[SEG]\\) \\(\\{obj2\\} :[SEG]\\)" + }, + { + "type": "list", + "bbox": [ + 0.522, + 0.809, + 0.892, + 0.853 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This prompt requests the MLLM to identify and distinguish different objects in the image based on the instructions and infuse the corresponding location information" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3861" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.12, + 0.091, + 0.228, + 0.103 + ], + "angle": 0, + "content": "(a) Multiple Targets" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.107, + 0.27, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.082, + 0.235, + 0.115, + 0.246 + ], + "angle": 0, + "content": "User:" + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.247, + 0.262, + 0.283 + ], + "angle": 0, + "content": "What are front zebra, right zebra in this image? Please output segmentation masks." + }, + { + "type": "title", + "bbox": [ + 0.082, + 0.299, + 0.139, + 0.31 + ], + "angle": 0, + "content": "Assistant:" + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.311, + 0.248, + 0.335 + ], + "angle": 0, + "content": "Sure, front zebra:[SEG], right zebra:[SEG]." + }, + { + "type": "image_caption", + "bbox": [ + 0.326, + 0.091, + 0.42, + 0.104 + ], + "angle": 0, + "content": "(b) Empty Target" + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.107, + 0.468, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.281, + 0.234, + 0.314, + 0.245 + ], + "angle": 0, + "content": "User:" + }, + { + "type": "text", + "bbox": [ + 0.28, + 0.247, + 0.458, + 0.295 + ], + "angle": 0, + "content": "What are the apple in the bowl, partial orange bottom left in this image? Please output segmentation masks." + }, + { + "type": "title", + "bbox": [ + 0.28, + 0.299, + 0.337, + 0.31 + ], + "angle": 0, + "content": "Assistant:" + }, + { + "type": "text", + "bbox": [ + 0.28, + 0.311, + 0.441, + 0.348 + ], + "angle": 0, + "content": "Sure, the apple in the bowl: [REJ], partial orange bottom left:[SEG]." + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.352, + 0.47, + 0.436 + ], + "angle": 0, + "content": "Figure 3. Example of the prompts and predicted masks of GSVA-Vicuna-7B drawn from gRefCOCO validation set. (a) depicts the multiple-target case, in which two zebras referred to are handled with two separate [SEG] tokens. (b) shows the empty-target case, where no apple is in the bowl. Thus, the null referent is rejected with a [REJ] token, and no segmentation mask will be generated." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.439, + 0.469, + 0.546 + ], + "angle": 0, + "content": "of each target into the associated [SEG] token. We regard this ability as an implicit multimodal version of In-Context Learning (ICL), which is demonstrated by many prior works [2, 13, 33, 59, 90]. GSVA takes the target description preceding each [SEG] token as a hint to link this token to the object requested in the image through auto-regressive decoding." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.554, + 0.441, + 0.571 + ], + "angle": 0, + "content": "3.4. Rejecting Empty Targets via [REJ] Tokens" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.578, + 0.469, + 0.699 + ], + "angle": 0, + "content": "Empty targets. In GRES, many expressions match no targets in the image, including absence, incorrect attributes, inaccurate locations, etc. These expressions should be treated as empty targets for models to predict all-negative masks. LISA [32] falls short of predicting masks with all-zeros seamlessly since the [SEG] token always calls for a segmentation mask. As shown in Figure 1 (b), LISA incorrectly marks a piece of green apple as the empty target banana." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.699, + 0.469, + 0.835 + ], + "angle": 0, + "content": "[REJ] token. We let the MLLM of GSVA predict a [REJ] token for each object that does not exist in the image but is referred to in the instructions, as shown in Figure 3 (b). GSVA predicts the targets marked with [REJ] tokens as empty targets, therefore setting all-zero masks for them. The involvement of [REJ] tokens directly rejects the empty target, liberating the mask decoder of the SFM from learning to segment the inconsistent targets. An example prompt with one existing target and one empty target is as follows:" + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.838, + 0.469, + 0.867 + ], + "angle": 0, + "content": "User: What are {obj1} (absent), {obj2} (absent), {obj3} in this image? Please output the segmentation masks." + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.868, + 0.468, + 0.883 + ], + "angle": 0, + "content": "Assistant: Sure, \\(\\{obj1\\} :[REJ],\\{obj2\\} :[REJ],\\{obj3\\} :[SEG].\\)" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.886, + 0.469, + 0.901 + ], + "angle": 0, + "content": "The [REJ] token prediction can also be seen as a variant" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.228 + ], + "angle": 0, + "content": "of VQA task, where the specified object and its position in the image need to be considered. Thanks to the unprecedented capabilities of MLLM in understanding the images [9, 34, 39, 92] and reasoning the spatial relationships of the referring objects [52, 88, 91], we make the most of the MLLM in GSVA to unleash the burden of the segmentation model. The proposed empty-target-aware mechanism both improves the quality of masks and ameliorates the errors of identifying nonexistent objects." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.24, + 0.633, + 0.258 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.266, + 0.892, + 0.371 + ], + "angle": 0, + "content": "In this section, we conduct comprehensive experiments to validate the efficacy of GSVA. First, we show the results on gRefCOCO [38] dataset to show the superiority of GSVA in GRES tasks. Then we verify GSVA that is also competent with other baselines in classic RES, REC tasks. We move on to ablate some important design choices of GSVA, followed by some qualitative visualization of GRES results." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.38, + 0.587, + 0.394 + ], + "angle": 0, + "content": "4.1. GRES" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.403, + 0.892, + 0.811 + ], + "angle": 0, + "content": "Settings. We adopt gRefCOCO [38] dataset to validate GSVA and LISA [32] on GRES, which contains 278,232 expressions, including 80,022 multi-target and 32,202 empty-target ones, referring to the objects in 19,994 images. The images are split into four subsets: training, validation, test-A, and test-B, following the same UNC partition of RefCOCO [84]. We first add the gRefCOCO training set into the mixed training dataset in LISA to pretrain GSVA and LISA for 50,000 steps and then finetune the models on the gRefCOCO training dataset for another 10 epochs. We evaluate the pretrained and finetuned models on the remaining validation set, test set A, and test set B, respectively. We adopt the IoU, cIoU metrics for the segmentation mask outputs. Following the implementation in Liu et al. [38], IoU averages the IoU for each mask, whereas cIoU computes the cumulative intersection area over the cumulative union area across the whole dataset. As for the empty target, we compute the No-target-accuracy (N-acc.), which is the ratio of the correctly classified empty-target expressions over all the empty-target expressions in the dataset. For a correctly classified empty target, the gIoU is set to 1.0, and the cIoU does not take them into account, while for a misclassified empty target, the gIoU is set to 0.0 and the union area is accumulated in the cIoU. Following the criteria in Liu et al. [38], a predicted mask is regarded as empty for LISA if the positive pixels are less than 50, whereas GSVA predicts [REJ] tokens to identify empty targets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Results. We report the GRES segmentation results of GSVA and LISA [32] in Table 1. Three variants of GSVA, including GSVA-Vicuna-7B, GSVA-Vicuna-13B, and GSVA-Llama2-13B show competitive performance without finetuning on gIoU to the strongest non-LLM baseline ReLA [38]. However, LISA models fail to handle the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3862" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.091, + 0.892, + 0.392 + ], + "angle": 0, + "content": "
MethodGeneralized Referring Expression Segmentation on gRefCOCO
Validation SetTest Set ATest Set B
gIoUcIoUN-acc.gIoUcIoUN-acc.gIoUcIoUN-acc.
MattNet [85]48.2447.5141.1559.3058.6644.0446.1445.3341.32
LTS [29]52.7052.30-62.6461.87-50.4249.96-
VLT [10]52.0052.5147.1763.2062.1948.7450.8850.5247.82
CRIS [72]56.2755.34-63.4263.82-51.7951.04-
LAVT [81]58.4057.6449.3265.9065.3249.2555.8355.0448.46
ReLA [38]63.6062.4256.3770.0369.2659.0261.0259.8858.40
LISA-Vicuna-7B [32]32.2138.722.7148.5452.556.3739.6544.795.00
GSVA-Vicuna-7B63.3261.7056.4570.1169.2363.5061.3460.2658.42
LISA-Vicuna-7B [32] (ft)61.6361.7654.6766.2768.5050.0158.8460.6351.91
GSVA-Vicuna-7B (ft)66.4763.2962.4371.0869.9365.3162.2360.4760.56
LISA-Vicuna-13B [32]32.7339.853.6648.7653.624.8939.4945.354.41
GSVA-Vicuna-13B62.9760.1858.4467.1767.5954.6058.0657.2852.22
LISA-Vicuna-13B [32] (ft)63.4562.9955.2568.1869.6552.1661.8462.2456.15
GSVA-Vicuna-13B (ft)68.0164.0565.3671.7570.5167.2563.8361.2863.11
LISA-Llama2-13B [32]33.2639.643.2749.7653.807.2840.4945.415.73
GSVA-Llama2-13B63.2062.3854.5169.5269.8657.8462.0660.7758.30
LISA-Llama2-13B [32] (ft)65.2463.9657.4969.9971.0055.4362.1162.2956.34
GSVA-Llama2-13B (ft)70.0466.3866.0273.2972.7964.7265.4563.2062.47
" + }, + { + "type": "table_caption", + "bbox": [ + 0.08, + 0.398, + 0.892, + 0.44 + ], + "angle": 0, + "content": "Table 1. Generalized referring expression segmentation (GRES) results on gRefCOCO [38] dataset. gIoU and cIoU are IoU metrics averaged by each example and accumulated over whole dataset, respectively. N-acc. is short for the accuracy of correctly classifying null targets. Baselines are copied from Liu et al. [38]. (ft) denotes the model is finetuned on the training set of gRefCOCO." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.444, + 0.47, + 0.626 + ], + "angle": 0, + "content": "GRES task without finetuning, showing degradation in both gIoU and cIoU in each model variant. Especially the low N-acc indicates that LISA is unable to correctly reject the empty targets. When finetuned on gRefCOCO training set, GSVA-Vicuna-7B performs better than the finetuned LISA counterpart, with about \\(4\\%\\) improvement in gIoU and over \\(5\\%\\) in N-acc on all three evaluation splits. GSVA variants with larger LLM incorporated further push the limits, achieving over \\(70\\%\\) in gIoU on the validation set, \\(73\\%\\) on test set A, and \\(65\\%\\) on test set B. The 13B models also consistently outperform LISA by large margins, demonstrating the superiority of GSVA in GRES task." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.636, + 0.387, + 0.652 + ], + "angle": 0, + "content": "4.2. Referring Expression Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.469, + 0.809 + ], + "angle": 0, + "content": "Settings. To validate the abilities to handle various tasks, we evaluate GSVA in the classic RES task. Following the common evaluation protocols, we test variants of GSVA and LISA equipped with different LLMs on RefCOCO, RefCOCO+ [30], and RefCOCOg [44]. We follow the UNC split to perform experiments on RefCOCO and RefCOCO+, and UMD split for RefCOCOg. The models are firstly pretrained as in GRES, and then finetuned for 10 epochs with a joint dataset of these three RES training set. The cIoU metric is adopted to measure the model performances." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Results. Table 2 shows the RES results of GSVA. For the pretrained models, all the three variants of GSVA achieve higher cIoU than LISA [32] by clear margins. Our 7B model outperforms LISA-Vicuna about \\(2\\%\\) cIoU on almost every data split. After finetuning, the preponderance of GSVA over LISA keeps and even enlarges. Specifically," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.444, + 0.892, + 0.504 + ], + "angle": 0, + "content": "the cIoU metrics of GSVA-Llama2-13B on 8 sets surpass LISA by at least \\(2.8\\%\\). For the test set of RefCOCOg, the margin even grows to \\(5.9\\%\\), which exhibits GSVA is also competitive in the classic RES task." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.533, + 0.825, + 0.55 + ], + "angle": 0, + "content": "4.3. Referring Expression Comprehension" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.562, + 0.892, + 0.683 + ], + "angle": 0, + "content": "Settings. Since GSVA is capable for RES tasks, it is natural to transfer to Referring Expression Comprehension (REC) tasks, by simply computing the bounding boxes of the masks. The datasets of REC are the same as RES, in which we evaluate GSVA. If the IoU of a predicted bounding box and the ground truth is greater than 0.5, this prediction is marked as correct. We use the same models in RES to evaluate the phrase grounding capability in REC tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Results. As shown in Table 3, we mainly compare our GSVA to LISA in the three variants. Without finetuning, GSVA-Vicuna-7B outperforms LISA with a large margin over \\(7\\%\\) in almost all evaluation sets. The similar trends also hold for the Vicuna-13B and Llama2-13B variants. GSVA also benefits a lot from finetuning, e.g., the 7B variant achieves consistently higher Prec@0.5 than uLLaVA-7B [80] with LoRA finetuning, which is another strong baseline that adopts the \"mask2bbox\" pipelines without direct bounding box supervision. The finetuned GSVA also shows competitive performances to the fully finetuned uLLaVA-7B, suggesting the strong potential of our method. With larger LLMs incorporated, the performances of GSVA continue with over \\(3\\%\\) increments over all datasets." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3863" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.09, + 0.892, + 0.41 + ], + "angle": 0, + "content": "
Referring Expression Segmentation on RefCOCO, RefCOCO+, and RefCOCOg
MethodRefCOCO (UNC)RefCOCO+ (UNC)RefCOCOg (UMD)
Val.Test-ATest-BVal.Test-ATest-BVal.Test
MCN [42]62.464.259.750.655.044.749.249.4
VLT [10]67.570.565.256.361.050.155.057.7
CRIS [72]70.573.266.162.368.153.759.960.4
LAVT [81]72.775.868.862.168.455.161.262.1
ReLA [38]73.876.570.266.071.057.765.066.0
X-Decoder [93]------64.6-
SEEM [94]------65.7-
PolyFormer-L [40]76.078.373.369.374.661.969.270.2
LISA-Vicuna-7B* [32]74.176.571.162.467.456.566.468.5
GSVA-Vicuna-7B76.477.472.864.567.758.671.172.0
LISA-Vicuna-7B* [32] (ft)74.979.172.365.170.858.167.970.6
GSVA-Vicuna-7B (ft)77.278.973.565.969.659.872.773.3
LISA-Vicuna-13B [32]71.774.768.159.464.252.965.266.1
GSVA-Vicuna-13B74.677.570.562.566.555.569.671.2
LISA-Vicuna-13B [32] (ft)76.078.872.965.070.258.169.570.5
GSVA-Vicuna-13B (ft)78.280.474.267.471.560.974.275.6
LISA-Llama2-13B [32]73.476.269.562.366.656.368.268.5
GSVA-Llama2-13B77.779.974.268.071.561.573.273.9
LISA-Llama2-13B [32] (ft)76.378.772.466.271.059.370.171.1
GSVA-Llama2-13B (ft)79.281.777.170.373.863.675.777.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.416, + 0.893, + 0.459 + ], + "angle": 0, + "content": "Table 2. Referring expression segmentation results on RefCOCO, RefCOCO+ [30] and RefCOCOg [44] dataset. The cIoU metrics of each split are reported. Baselines are excerpted from Lai et al. [32]. (ft) denotes the models are finetuned on the joint training set of the referring expression segmentation datasets. * means the results are excerpted from the original paper." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.461, + 0.892, + 0.692 + ], + "angle": 0, + "content": "
MethodReferring Expression Comprehension on RefCOCO, RefCOCO+, and RefCOCOg
Val.RefCOCO Test-ATest-BVal.RefCOCO+ Test-ATest-BRefCOCOg Val.Test
u-LLaVA-7B [80] (LoRA)83.4787.1380.2168.7476.3260.9876.1978.24
u-LLaVA-7B [80] (full-ft)86.0489.4782.2674.0981.1666.6179.8781.68
LISA-Vicuna-7B [32]78.6881.7275.7462.9268.9356.4970.1072.47
GSVA-Vicuna-7B85.5088.0182.4970.2175.6265.1179.0079.21
LISA-Vicuna-7B [32] (ft)85.3988.8482.5974.2379.4668.4079.3480.42
GSVA-Vicuna-7B (ft)86.2789.2283.7772.8178.7868.0181.5881.83
LISA-Vicuna-13B [32]80.0183.2676.2663.7770.2457.4271.7973.34
GSVA-Vicuna-13B83.1287.0180.5468.1473.9062.0077.0878.89
LISA-Vicuna-13B [32] (ft)85.9289.0583.1674.8681.0868.8780.0981.48
GSVA-Vicuna-13B (ft)87.7190.4984.5776.5281.6970.3583.9084.85
LISA-Llama2-13B [32]82.5285.5678.8267.9173.7762.2575.3776.83
GSVA-Llama2-13B86.9989.5484.0873.8979.1069.3880.6882.07
LISA-Llama2-13B [32] (ft)85.9188.8481.7374.4680.5668.2680.0981.27
GSVA-Llama2-13B (ft)89.1692.0887.1779.7484.4573.4185.4786.18
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.697, + 0.895, + 0.741 + ], + "angle": 0, + "content": "Table 3. Referring expression comprehension results on RefCOCO, RefCOCO+ [30] and RefCOCOg [44] dataset. The metric is precision @ 0.5 IoU threshold. (LoRA) means the LLM in u-LLaVA [80] is finetuned with LoRA adapter, as LISA and GSVA, while (full-ft) represents the LLM in u-LLaVA is fully trained. Results of u-LLaVA-7B with \"mask2bbox\" strategy are reported for fair comparison." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.747, + 0.232, + 0.763 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.78, + 0.471, + 0.902 + ], + "angle": 0, + "content": "The involvement of the [REJ] token. [REJ] token plays a rather important role in GSVA. We study the effect of the [REJ] token by removing it from the vocabulary in GSVA, yielding a variant unable to reject a target from the text outputs. As shown in the \\(2^{\\text{nd}}\\) row of Table 4, after removing [REJ] token, there is a sharp N-acc drop over \\(25\\%\\) relatively, followed by the decline of gIoU at about \\(10\\%\\) on gRefCOCO validation set. This performance degradation" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.748, + 0.892, + 0.778 + ], + "angle": 0, + "content": "indicates the significance of LLM learning a special token to reject the referred instances absent in the image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Learning multiple [SEG] tokens. We continue to ablate the multiple [SEG] tokens, which is another core design of GSVA. After removing [REJ] token, we then reduce the number of [SEG] tokens to 1, which is identical to LISA [32] with the referring expression added before the only one [SEG] token: Assistant: Sure, {obj}:[SEG]. In the early experiments, we have found that stacking multiple expressions before one [SEG] token would result in diver" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3864" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.092, + 0.359, + 0.104 + ], + "angle": 0, + "content": "(a) Multiple Target Cases" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.108, + 0.217, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.108, + 0.361, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.369, + 0.108, + 0.478, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.179, + 0.217, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.179, + 0.361, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.369, + 0.179, + 0.478, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.085, + 0.258, + 0.21, + 0.267 + ], + "angle": 0, + "content": "Elephant in back, Left elephant" + }, + { + "type": "image_caption", + "bbox": [ + 0.221, + 0.253, + 0.371, + 0.271 + ], + "angle": 0, + "content": "Kid on left in blue, Boy on right, Blond boy on right, Girl back left pink" + }, + { + "type": "image_caption", + "bbox": [ + 0.375, + 0.254, + 0.465, + 0.272 + ], + "angle": 0, + "content": "Tan vase, Bottom vase, Right vase" + }, + { + "type": "image_caption", + "bbox": [ + 0.618, + 0.092, + 0.746, + 0.104 + ], + "angle": 0, + "content": "(b) Empty Target Cases" + }, + { + "type": "image", + "bbox": [ + 0.485, + 0.108, + 0.622, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.486, + 0.179, + 0.622, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.259, + 0.578, + 0.268 + ], + "angle": 0, + "content": "Lady on left" + }, + { + "type": "image", + "bbox": [ + 0.629, + 0.108, + 0.767, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.63, + 0.179, + 0.767, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.259, + 0.732, + 0.268 + ], + "angle": 0, + "content": "Teddy with X eye" + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.108, + 0.882, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.179, + 0.882, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.783, + 0.259, + 0.875, + 0.268 + ], + "angle": 0, + "content": "The dog on the suitcase" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.278, + 0.892, + 0.334 + ], + "angle": 0, + "content": "Figure 4. Visualizations of GSVA and LISA [32] in the GRES task. The first row shows LISA's segmentation results, the second row is the masks and rejections of GSVA, and the third row shows the referring expressions in the instructions. In (a) multiple target cases, each target is colored with a specific color. In (b) empty target cases, the images turn darker to highlight the incorrect predictions of LISA. The examples are selected from the gRefCOCO validation set. The masks are generated by the 7B models. Zoom in for the best view." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.34, + 0.469, + 0.448 + ], + "angle": 0, + "content": "
Model w/Vicuna-7BModificationsgRefCOCO Val.
RefExp. +[SEG]Multiple [SEG][REJ] TokengIoUcIoUN-acc.
GSVA✓✓✓63.3261.7056.45
✓✓✗51.5760.9530.32
✓✗✗44.8659.3711.96
LISA [32]✗✗✗32.2138.722.71
✗✓✓21.8327.220.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.453, + 0.469, + 0.523 + ], + "angle": 0, + "content": "Table 4. Ablation study on the core designs of GSVA. \\(\\checkmark\\) means the employment of the component while \\(X\\) means not. \"Ref-Exp.+[SEG]\", \"Multiple [SEG]\", and \"[REJ] Token\" are short for adding referring expression before [SEG] in the answer prompt, using multiple [SEG] tokens, involving [REJ] token, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.533, + 0.469, + 0.594 + ], + "angle": 0, + "content": "gence. Therefore we separate multiple targets to prompt the model with one expression at a time. The sharp decrements of gIoU by nearly \\(7\\%\\) and N-acc by over \\(15\\%\\) in the \\(3^{\\mathrm{rd}}\\) row demonstrate the significance of the multiple-[SEG]-token." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.594, + 0.469, + 0.837 + ], + "angle": 0, + "content": "Answers without referring expression. To examine the efficacy of the hinting prompts, we remove all the referring expressions before the [SEG] tokens. Based on the removal of multiple [SEG]s and [REJ]s, erasing the added referring expression falls back to the original LISA model, as shown in the \\(4^{\\text{th}}\\) row. We further choose only to remove it from GSVA model, keeping other configurations unchanged. Specifically, if there are two referents, the prompts in the answer will turn to Assistant: Sure, [SEG], [SEG], whose results are in the last row. The zero N-acc shows the model fails to identify any empty target without the help of the expressions, meanwhile the poor gIoU and cIoU indicates the segmentation ability is damaged. This phenomenon also suggests that the added referring expression hint GSVA to associate each [SEG] token to its corresponding target, which is in coherence with our hypothesis." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.847, + 0.214, + 0.86 + ], + "angle": 0, + "content": "4.5. Visualization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "We visualize some qualitative results of GSVA to verify its effectiveness. As shown in Figure 4, we present two groups" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.342, + 0.892, + 0.54 + ], + "angle": 0, + "content": "of examples from the validation set of gRefCOCO [38] to see how GSVA outperforms LISA in the face of the two main challenges in GRES: multiple targets and empty targets. In part (a), GSVA has managed to segment all the targets referred to, while LISA could only segment one of the requested instances. For example, in the third column, LISA only predicts the mask of the rightmost vase. On the contrary, GSVA separately segments all three targeted vases. In part (b), LISA mistakenly segments the instances in the image that disagree with the referring expression, e.g., in the sixth column, LISA proposes a mask of a cat in response to the request of the dog on the suitcase, whereas GSVA successfully reject all the empty targets." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.557, + 0.618, + 0.573 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.751 + ], + "angle": 0, + "content": "This paper introduces a novel multimodal large language model dubbed Generalized Segmentation Vision Assistant (GSVA). By introducing multiple [SEG] tokens and the new [REJ] token, GSVA effectively achieves multi-objective segmentation and empty target rejection, which addresses the challenging segmentation problems in practical application scenarios, e.g., Generalized Referring Expression Segmentation (GRES). Extensive experiments on GRES, classic RES, and REC tasks fully demonstrate the superior performance of our method, highlighting its significance for future research and applications." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.769, + 0.659, + 0.785 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This work is supported in part by the National Key R&D Program of China under Grant 2021ZD0140407, the National Natural Science Foundation of China under Grants 62321005 and 62276150. We also thank Dr. Yuan Yao and Prof. Zhiyuan Liu for their insightful and valuable comments on this research project. We also appreciate their generous support of computational resources." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.957 + ], + "angle": 0, + "content": "3865" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.183 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.185, + 0.472, + 0.242 + ], + "angle": 0, + "content": "[2] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.471, + 0.284 + ], + "angle": 0, + "content": "[3] Ding-Jie Chen, Songhao Jia, Yi-Chen Lo, Hwann-Tzong Chen, and Tyng-Luh Liu. See-through-text grouping for referring image segmentation. In IEEE ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.284, + 0.471, + 0.338 + ], + "angle": 0, + "content": "[4] Yixin Chen, Qing Li, Deqian Kong, Yik Lun Kei, Song-Chun Zhu, Tao Gao, Yixin Zhu, and Siyuan Huang. Yourefit: Embodied reference understanding with language and gesture. In IEEE ICCV, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.34, + 0.471, + 0.395 + ], + "angle": 0, + "content": "[5] Ming-Ming Cheng, Shuai Zheng, Wen-Yan Lin, Vibhav Vineet, Paul Sturgess, Nigel Crook, Niloy J Mitra, and Philip Torr. Imagesprit: Verbal guided image parsing. ACM ToG, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.396, + 0.471, + 0.466 + ], + "angle": 0, + "content": "[6] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) * chatgpt quality, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.467, + 0.47, + 0.507 + ], + "angle": 0, + "content": "[7] Yong Xien Chng, Henry Zheng, Yizeng Han, Xuchong Qiu, and Gao Huang. Mask grounding for referring image segmentation. In IEEE CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.509, + 0.471, + 0.577 + ], + "angle": 0, + "content": "[8] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.579, + 0.471, + 0.647 + ], + "angle": 0, + "content": "[9] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. In NeurIPS, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.649, + 0.468, + 0.69 + ], + "angle": 0, + "content": "[10] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vision-language transformer and query generation for referring segmentation. In IEEE ICCV, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.691, + 0.468, + 0.733 + ], + "angle": 0, + "content": "[11] Xiaohan Ding, Xiangyu Zhang, Ningning Ma, Jungong Han, Guiguang Ding, and Jian Sun. Repvgg: Making vgg-style convnets great again. In IEEE ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.468, + 0.774 + ], + "angle": 0, + "content": "[12] Xiaohan Ding, Xiangyu Zhang, Jungong Han, and Guiguang Ding. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In IEEE CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.468, + 0.83 + ], + "angle": 0, + "content": "[13] Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Zhiyong Wu, Baobao Chang, Xu Sun, Jingjing Xu, and Zhifang Sui. A survey for in-context learning. arXiv preprint arXiv:2301.00234, 2022.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[15] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale. In IEEE CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[16] Guang Feng, Zhiwei Hu, Lihe Zhang, and Huchuan Lu. Encoder fusion network with co-attention embedding for referring image segmentation. In IEEE CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[17] Chen Gao, Jinyu Chen, Si Liu, Luting Wang, Qiong Zhang, and Qi Wu. Room-and-object aware knowledge reasoning for remote embodied referring expression. In IEEE CVPR, 2021. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[18] Dongchen Han, Tianzhu Ye, Yizeng Han, Zhuofan Xia, Shiji Song, and Gao Huang. Agent attention: On the integration of softmax and linear attention. arXiv preprint arXiv:2312.08874, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[19] Yizeng Han, Gao Huang, Shiji Song, Le Yang, Honghui Wang, and Yulin Wang. Dynamic neural networks: A survey. IEEE TPAMI, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[20] Yizeng Han, Gao Huang, Shiji Song, Le Yang, Yitian Zhang, and Haojun Jiang. Spatially adaptive feature refinement for efficient inference. IEEE TIP, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[21] Yizeng Han, Yifan Pu, Zihang Lai, Chaofei Wang, Shiji Song, Junfeng Cao, Wenhui Huang, Chao Deng, and Gao Huang. Learning to weight samples for dynamic early-exiting networks. In ECCV, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.448, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[22] Yizeng Han, Zhihang Yuan, Yifan Pu, Chenhao Xue, Shiji Song, Guangyu Sun, and Gao Huang. Latency-aware spatialwise dynamic networks. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.491, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[23] Yizeng Han, Dongchen Han, Zeyu Liu, Yulin Wang, Xuran Pan, Yifan Pu, Chao Deng, Junlan Feng, Shiji Song, and Gao Huang. Dynamic perceiver for efficient visual recognition. In IEEE ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.547, + 0.892, + 0.603 + ], + "angle": 0, + "content": "[24] Yizeng Han, Zeyu Liu, Zhihang Yuan, Yifan Pu, Chaofei Wang, Shiji Song, and Gao Huang. Latency-aware unified dynamic networks for efficient image recognition. arXiv preprint arXiv:2308.15949, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.604, + 0.892, + 0.644 + ], + "angle": 0, + "content": "[25] Ronghang Hu, Marcus Rohrbach, and Trevor Darrell. Segmentation from natural language expressions. In ECCV, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.646, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[26] Yutao Hu, Qixiong Wang, Wenqi Shao, Enze Xie, Zhenguo Li, Jungong Han, and Ping Luo. Beyond one-to-one: Rethinking the referring image segmentation. In IEEE ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.703, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[27] Gao Huang, Yulin Wang, Kangchen Lv, Haojun Jiang, Wenhui Huang, Pengfei Qi, and Shiji Song. Glance and focus networks for dynamic visual recognition. IEEE TPAMI, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.76, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[28] Rui Huang, Xuran Pan, Henry Zheng, Haojun Jiang, Zhifeng Xie, Cheng Wu, Shiji Song, and Gao Huang. Joint representation learning for text and 3d point cloud. Pattern Recognition, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[29] Ya Jing, Tao Kong, Wei Wang, Liang Wang, Lei Li, and Tie-niu Tan. Locate then segment: A strong pipeline for referring image segmentation. In IEEE CVPR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[30] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In EMNLP, 2014. 1, 4, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3866" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[31] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In IEEE ICCV, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.148, + 0.47, + 0.203 + ], + "angle": 0, + "content": "[32] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692, 2023. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.204, + 0.471, + 0.259 + ], + "angle": 0, + "content": "[33] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Fanyi Pu, Jingkang Yang, Chunyuan Li, and Ziwei Liu. Mimicit: Multi-modal in-context instruction tuning. arXiv preprint arXiv:2306.05425, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.26, + 0.469, + 0.314 + ], + "angle": 0, + "content": "[34] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.315, + 0.469, + 0.37 + ], + "angle": 0, + "content": "[35] Ruiyu Li, Kaican Li, Yi-Chun Kuo, Michelle Shu, Xiaojuan Qi, Xiaoyong Shen, and Jiaya Jia. Referring image segmentation via recurrent refinement networks. In IEEE CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.371, + 0.469, + 0.44 + ], + "angle": 0, + "content": "[36] Ziyi Lin, Chris Liu, Renrui Zhang, Peng Gao, Longtian Qiu, Han Xiao, Han Qiu, Chen Lin, Wenqi Shao, Keqin Chen, et al. Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models. arXiv preprint arXiv:2311.07575, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.441, + 0.469, + 0.482 + ], + "angle": 0, + "content": "[37] Chenxi Liu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, and Alan Yuille. Recurrent multimodal interaction for referring image segmentation. In IEEE CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.469, + 0.523 + ], + "angle": 0, + "content": "[38] Chang Liu, Henghui Ding, and Xudong Jiang. Gres: Generalized referring expression segmentation. In IEEE CVPR, 2023. 1, 2, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.469, + 0.552 + ], + "angle": 0, + "content": "[39] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.553, + 0.469, + 0.607 + ], + "angle": 0, + "content": "[40] Jiang Liu, Hui Ding, Zhaowei Cai, Yuting Zhang, Ravi Kumar Satzoda, Vijay Mahadevan, and R Manmatha. Polyformer: Referring image segmentation as sequential polygon generation. In IEEE CVPR, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.609, + 0.469, + 0.676 + ], + "angle": 0, + "content": "[41] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.678, + 0.469, + 0.733 + ], + "angle": 0, + "content": "[42] Gen Luo, Yiyi Zhou, Xiaoshuai Sun, Liujuan Cao, Chenglin Wu, Cheng Deng, and Rongrong Ji. Multi-task collaborative network for joint referring expression comprehension and segmentation. In IEEE CVPR, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.469, + 0.789 + ], + "angle": 0, + "content": "[43] Tengchao Lv, Yupan Huang, Jingye Chen, Lei Cui, Shuming Ma, Yaoyao Chang, Shaohan Huang, Wenhui Wang, Li Dong, Weiyao Luo, et al. Kosmos-2.5: A multimodal literate model. arXiv preprint arXiv:2309.11419, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[44] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In IEEE CVPR, 2016. 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[45] Khanh Nguyen, Debadeepta Dey, Chris Brockett, and Bill Dolan. Vision-based navigation with language-based assistance via imitation learning with indirect intervention. In IEEE CVPR, 2019. 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[46] Zanlin Ni, Yulin Wang, Jiangwei Yu, Haojun Jiang, Yue Cao, and Gao Huang. Deep incubation: Training large models by divide-and-conquering. In IEEE CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.192 + ], + "angle": 0, + "content": "[47] Zanlin Ni, Yulin Wang, Renping Zhou, Jiayi Guo, Jinyi Hu, Zhiyuan Liu, Shiji Song, Yuan Yao, and Gao Huang. Revisiting non-autoregressive transformers for efficient image synthesis. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.194, + 0.777, + 0.209 + ], + "angle": 0, + "content": "[48] OpenAI. Gpt-4 technical report, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.211, + 0.892, + 0.279 + ], + "angle": 0, + "content": "[49] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.282, + 0.892, + 0.323 + ], + "angle": 0, + "content": "[50] Xuran Pan, Chunjiang Ge, Rui Lu, Shiji Song, Guanfu Chen, Zeyi Huang, and Gao Huang. On the integration of self-attention and convolution. In IEEE CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.326, + 0.892, + 0.367 + ], + "angle": 0, + "content": "[51] Xuran Pan, Tianzhu Ye, Zhuofan Xia, Shiji Song, and Gao Huang. Slide-transformer: Hierarchical vision transformer with local self-attention. In IEEE CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.369, + 0.892, + 0.425 + ], + "angle": 0, + "content": "[52] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.427, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[53] Yifan Pu, Yiru Wang, Zhuofan Xia, Yizeng Han, Yulin Wang, Weihao Gan, Zidong Wang, Shiji Song, and Gao Huang. Adaptive rotated convolution for rotated object detection. In IEEE ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.485, + 0.892, + 0.553 + ], + "angle": 0, + "content": "[54] Yuankai Qi, Qi Wu, Peter Anderson, Xin Wang, William Yang Wang, Chunhua Shen, and Anton van den Hengel. Reverie: Remote embodied visual referring expression in real indoor environments. In IEEE CVPR, 2020. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.556, + 0.892, + 0.625 + ], + "angle": 0, + "content": "[55] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.628, + 0.892, + 0.697 + ], + "angle": 0, + "content": "[56] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M Anwer, Erix Xing, Ming-Hsuan Yang, and Fahad S Khan. Glamm: Pixel grounding large multimodal model. arXiv preprint arXiv:2311.03356, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.7, + 0.892, + 0.74 + ], + "angle": 0, + "content": "[57] Hengcan Shi, Hongliang Li, Fanman Meng, and Qingbo Wu. Key-word-aware network for referring expression image segmentation. In ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.743, + 0.892, + 0.798 + ], + "angle": 0, + "content": "[58] Qie Sima, Sinan Tan, Huaping Liu, Fuchun Sun, Weifeng Xu, and Ling Fu. Embodied referring expression for manipulation question answering in interactive environment. In IEEE ICRA, 2023. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.801, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[59] Yan Tai, Weichen Fan, Zhao Zhang, Feng Zhu, Rui Zhao, and Ziwei Liu. Link-context learning for multimodal llms. arXiv preprint arXiv:2308.07891, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.845, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[60] Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. Stanford alpaca: An instruction-following llama model, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3867" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[61] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.164, + 0.47, + 0.233 + ], + "angle": 0, + "content": "[62] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.47, + 0.303 + ], + "angle": 0, + "content": "[63] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, et al. Image as a foreign language: Beit pretraining for vision and vision-language tasks. In IEEE CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.469, + 0.359 + ], + "angle": 0, + "content": "[64] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, et al. Vision: Large language model is also an open-ended decoder for vision-centric tasks. In NeurlPS, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.361, + 0.469, + 0.416 + ], + "angle": 0, + "content": "[65] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, et al. Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.418, + 0.469, + 0.487 + ], + "angle": 0, + "content": "[66] Xin Wang, Qiuyuan Huang, Asli Celikyilmaz, Jianfeng Gao, Dinghan Shen, Yuan-Fang Wang, William Yang Wang, and Lei Zhang. Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation. In IEEE CVPR, 2019. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.489, + 0.469, + 0.531 + ], + "angle": 0, + "content": "[67] Yulin Wang, Zhaoxi Chen, Haojun Jiang, Shiji Song, Yizeng Han, and Gao Huang. Adaptive focus for efficient video recognition. In IEEE ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.532, + 0.469, + 0.587 + ], + "angle": 0, + "content": "[68] Yulin Wang, Rui Huang, Shiji Song, Zeyi Huang, and Gao Huang. Not all images are worth 16x16 words: Dynamic transformers for efficient image recognition. In NeurIPS, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.589, + 0.469, + 0.644 + ], + "angle": 0, + "content": "[69] Yulin Wang, Yang Yue, Yuanze Lin, Haojun Jiang, Zihang Lai, Victor Kulikov, Nikita Orlov, Humphrey Shi, and Gao Huang. Adafocus v2: End-to-end training of spatial dynamic networks for video recognition. In IEEE CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.645, + 0.469, + 0.701 + ], + "angle": 0, + "content": "[70] Yulin Wang, Yang Yue, Xinhong Xu, Ali Hassani, Victor Kulikov, Nikita Orlov, Shiji Song, Humphrey Shi, and Gao Huang. Adafocusv3: On unified spatial-temporal dynamic video recognition. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.702, + 0.469, + 0.757 + ], + "angle": 0, + "content": "[71] Yulin Wang, Yang Yue, Rui Lu, Tianjiao Liu, Zhao Zhong, Shiji Song, and Gao Huang. Efficienttrain: Exploring generalized curriculum learning for training visual backbones. In IEEE ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.759, + 0.469, + 0.801 + ], + "angle": 0, + "content": "[72] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clip-driven referring image segmentation. In IEEE CVPR, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[73] Zhichao Wei, Xiaohao Chen, Mingqiang Chen, and Siyu Zhu. Learning aligned cross-modal representations for referring image segmentation. arXiv preprint arXiv:2301.06429, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[74] Penghao Wu and Saining Xie. V*: Guided visual search as a core mechanism in multimodal llms. arXiv preprint arXiv:2312.14135, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[75] Tsung-Han Wu, Giscard Biamby, David Chan, Lisa Dunlap, Ritwik Gupta, Xudong Wang, Joseph E Gonzalez, and Trevor Darrell. See, say, and segment: Teaching Imms to overcome false premises. arXiv preprint arXiv:2312.08366, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[76] Zhuofan Xia, Xuran Pan, Shiji Song, Li Erran Li, and Gao Huang. Vision transformer with deformable attention. In IEEE CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.208, + 0.892, + 0.249 + ], + "angle": 0, + "content": "[77] Zhuofan Xia, Xuran Pan, Xuan Jin, Yuan He, Hui Xue, Shiji Song, and Gao Huang. Budgeted training for vision transformer. In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.252, + 0.892, + 0.307 + ], + "angle": 0, + "content": "[78] Zhuofan Xia, Xuran Pan, Shiji Song, Li Erran Li, and Gao Huang. Dat++: Spatially dynamic vision transformer with deformable attention. arXiv preprint arXiv:2309.01430, 2023.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.31, + 0.892, + 0.378 + ], + "angle": 0, + "content": "[79] Bin Xiao, Haiping Wu, Weijian Xu, Xiyang Dai, Houdong Hu, Yumao Lu, Michael Zeng, Ce Liu, and Lu Yuan. Florence-2: Advancing a unified representation for a variety of vision tasks. arXiv preprint arXiv:2311.06242, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.382, + 0.892, + 0.437 + ], + "angle": 0, + "content": "[80] Jinjin Xu, Liwu Xu, Yuzhe Yang, Xiang Li, Yanchun Xie, Yi-Jie Huang, and Yaqian Li. u-llava: Unifying multimodal tasks via large language model. arXiv preprint arXiv:2311.05348, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.44, + 0.892, + 0.495 + ], + "angle": 0, + "content": "[81] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Lavt: Language-aware vision transformer for referring image segmentation. In IEEE CVPR, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.498, + 0.892, + 0.539 + ], + "angle": 0, + "content": "[82] Linwei Ye, Mrigank Rochan, Zhi Liu, and Yang Wang. Cross-modal self-attention network for referring image segmentation. In IEEE CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.541, + 0.892, + 0.61 + ], + "angle": 0, + "content": "[83] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.613, + 0.892, + 0.655 + ], + "angle": 0, + "content": "[84] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In ECCV, 2016. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.657, + 0.892, + 0.713 + ], + "angle": 0, + "content": "[85] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In IEEE CVPR, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.715, + 0.892, + 0.77 + ], + "angle": 0, + "content": "[86] Ao Zhang, Liming Zhao, Chen-Wei Xie, Yun Zheng, Wei Ji, and Tat-Seng Chua. Next-chat: An Imm for chat, detection and segmentation. arXiv preprint arXiv:2311.04498, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.773, + 0.892, + 0.842 + ], + "angle": 0, + "content": "[87] Renrui Zhang, Jiaming Han, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, Peng Gao, and Yu Qiao. Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.845, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[88] Shilong Zhang, Peize Sun, Shoufa Chen, Min Xiao, Wenqi Shao, Wenwei Zhang, Kai Chen, and Ping Luo. Gpt4roi: Instruction tuning large language model on region-of-interest. arXiv preprint arXiv:2307.03601, 2023. 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3868" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[89] Zicheng Zhang, Yi Zhu, Jianzhuang Liu, Xiaodan Liang, and Wei Ke. Coupalign: Coupling word-pixel with sentence-mask alignments for referring image segmentation. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.218 + ], + "angle": 0, + "content": "[90] Haozhe Zhao, Zefan Cai, Shuzheng Si, Xiaojian Ma, Kaikai An, Liang Chen, Zixuan Liu, Sheng Wang, Wenjuan Han, and Baobao Chang. Mmicl: Empowering vision-language model with multi-modal in-context learning. arXiv preprint arXiv:2309.07915, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.22, + 0.47, + 0.274 + ], + "angle": 0, + "content": "[91] Yang Zhao, Zhijie Lin, Daquan Zhou, Zilong Huang, Jiashi Feng, and Bingyi Kang. Bubogpt: Enabling visual grounding in multi-modal llms. arXiv preprint arXiv:2307.08581, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.277, + 0.47, + 0.331 + ], + "angle": 0, + "content": "[92] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.333, + 0.47, + 0.388 + ], + "angle": 0, + "content": "[93] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In IEEE CVPR, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.39, + 0.469, + 0.431 + ], + "angle": 0, + "content": "[94] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. In NeurIPS, 2023. 2, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3869" + } + ] +] \ No newline at end of file diff --git a/2024/GSVA_ Generalized Segmentation via Multimodal Large Language Models/12e5e8f5-7030-48a1-9213-1240824636f7_origin.pdf b/2024/GSVA_ Generalized Segmentation via Multimodal Large Language Models/12e5e8f5-7030-48a1-9213-1240824636f7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b8fb2617dca3d37bbba94aabd7fefc0d4b5d9414 --- /dev/null +++ b/2024/GSVA_ Generalized Segmentation via Multimodal Large Language Models/12e5e8f5-7030-48a1-9213-1240824636f7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a96e20cf00a163cb4c79f894eadbef353d91fa8ae81d7813ae4d44a7a6bd33a6 +size 8630866 diff --git a/2024/GSVA_ Generalized Segmentation via Multimodal Large Language Models/full.md b/2024/GSVA_ Generalized Segmentation via Multimodal Large Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c0864a90bd4407f971c3ba722763389f25a5fb64 --- /dev/null +++ b/2024/GSVA_ Generalized Segmentation via Multimodal Large Language Models/full.md @@ -0,0 +1,363 @@ +# GSVA: Generalized Segmentation via Multimodal Large Language Models + +Zhuofan Xia* Dongchen Han* Yizeng Han Xuran Pan Shiji Song Gao Huang† + +Department of Automation, BNRist, Tsinghua University + +# Abstract + +Generalized Referring Expression Segmentation (GRES) extends the scope of classic RES to refer to multiple objects in one expression or identify the empty targets absent in the image. GRES poses challenges in modeling the complex spatial relationships of the instances in the image and identifying non-existing referents. Multimodal Large Language Models (MLLMs) have recently shown tremendous progress in these complicated vision-language tasks. Connecting Large Language Models (LLMs) and vision models, MLLMs are proficient in understanding contexts with visual inputs. Among them, LISA, as a representative, adopts a special [SEG] token to prompt a segmentation mask decoder; e.g., SAM, to enable MLLMs in the RES task. However, existing solutions to GRES remain unsatisfactory since current segmentation MLLMs cannot correctly handle the cases where users might reference multiple subjects in a singular prompt or provide descriptions incongruent with any image target. In this paper, we propose Generalized Segmentation Vision Assistant (GSVA) to address this gap. Specifically, GSVA reuses the [SEG] token to prompt the segmentation model towards supporting multiple mask references simultaneously and innovatively learns to generate a [REJ] token to reject the null targets explicitly. Experiments validate GSVA's efficacy in resolving the GRES issue, marking a notable enhancement and setting a new record on the GRES benchmark gRefCOCO dataset. GSVA also proves effective across various classic referring segmentation and comprehension tasks. Code is available at https://github.com/LeapLabTHU/GSVA. + +# 1. Introduction + +Referring Expression Segmentation (RES) [5, 25] is an emerging vision-language (VL) task predicting the masks of the interested objects referred to in the language expression. RES has great potential in many areas, especially embodied AI [17, 28, 54, 58, 66], including VL navigation, humanrobot interaction, etc. Nevertheless, the simplification in + +![](images/e2387831eaffe94f5cd97cbfd1215f1b02e9a6c9b6249555ee5eca4f28302746.jpg) +Figure 1. Comparison of the segmentation masks by LISA [32] and GSVA, facing the challenges in Generalized Referring Expression Segmentation (GRES) [38]. (a) LISA fails to segment the correct targets when multiple targets are requested due to the single [SEG] token restriction. GSVA successfully generates all target masks via learning multiple [SEG] tokens. (b) When the referent does not exist in the image, i.e., the empty target is requested, LISA reluctantly produces the wrong mask because of the compulsive [SEG] token output. In contrast, GSVA can reject the empty targets by predicting [REJ] tokens in the output sequence. + +RES formulation that one referring expression must match an individual object in the image [30] has left a gap between current RES algorithms and real-world applications, neglecting multiple-target and empty-target cases. + +To bridge this gap, Generalized Referring Expression Segmentation (GRES) [38] has recently been proposed to support multiple-target and empty-target cases. Practically, users refer to multiple subjects within a single instruction or provide descriptions that do not correspond to any targets in the image. As an example shown in Figure 1 (a), the left man and the second man from the right are targeted simultaneously, while a banana is referred to in a scene of apples in Figure 1 (b). RES takes no account of these cases, which GRES handles. In addition to the multimodal correspondences between images and text prompts in classic RES, GRES poses new challenges in handling more complicated multiple-target and empty-target cases. Therefore, the models must handle complex spatial relationships of the instances in the image [38] to segment the targets at various + +locations and reject the empty targets in the wrong places. + +The recent blooming Multimodal Large Language Models (MLLMs) [1, 34, 39, 83, 87] meet the requirements of GRES since they show excellency in complex reasoning [8] and instruction following [49] with visual inputs by aligning the LLMs [2, 6, 48, 60-62] and Visual Foundation Models (VFMs) [15, 55, 63, 79] which are typically various Vision Transformers [14, 18, 23, 50, 51, 76, 78] to perceive image or video inputs. To support segmentation output, many works [74, 75, 86] link an MLLM (e.g., LLaVA [39]) and a segmentation model (e.g., SAM [31]) by prompting the decoder with special token embeddings (e.g., [SEG] in LISA [32]) to generate masks of the referents in the user's instructions. Although these models manage to handle RES, GRES is still beyond their reach. As shown in Figure 1, LISA fails to work well in GRES where the multiple-target and empty-target challenges remain uncharted. + +To address the above challenges, we propose Generalized Segmentation Vision Assistant (GSVA). We attribute the vulnerability of other segmentation MLLMs in GRES to the single constant [SEG] token that restricts its flexibility. Therefore, we present two pivotal designs in GSVA: (1) learning to predict multiple [SEG] tokens to segment multiple targets; (2) rejecting empty targets in referring expressions by predicting [REJ] tokens. Specifically, when multiple targets are requested in the referring expression, we place multiple weight-sharing [SEG] tokens corresponding to the entities in the expression, encouraging the MLLM to learn to output multiple [SEG] tokens. To distinguish each [SEG] token and avoid ambiguity, we add the expression of each entity in front of the corresponding [SEG] token, hinting each [SEG] token to focus on the specific target, which can be regarded as implicit In-Context Learning, and dynamic neural network [19]. Meanwhile, if the referents are absent in the image, the corresponding [SEG] tokens after the prompts are altered to [REJ] tokens to identify empty targets. The predicted [REJ] tokens are directly assigned with empty masks without decoded, which liberates the segmentation model from seeking non-existing targets in the image. This Benefiting from these novel designs, GSVA takes a big step forward in addressing GRES challenges, as shown in the second row of Figure 1. + +Our contributions are summarized as follows: (1) We propose GSVA to solve the GRES problem with MLLM by handling the spatial relationships among targets, and study the GRES problem systematically in the context of LLM for the first time. (2) We propose the non-trivial shared-weight multiple [SEG] tokens guided by each referent prompt to address the multiple-target problem. (3) We firstly propose a clean solution, the [REJ] token, to reject the empty targets, which can be seamlessly applied to various models. (4) GSVA is intuitive and effective, achieving state-of-the-art performance on the GRES benchmarks. + +# 2. Related Works + +RES and GRES. Referring Expression Segmentation (RES) [4, 5, 25, 44] assumes that one expression matches one existing target, and many works explore fusing image and language [3, 16, 35, 37, 81] to segment objects under instructions. Currently, most RES methods adopt the cross-attention module or cross-modal alignment to bridge the modality gap [7, 57, 73, 82, 89]. Another line of research enables the text prompts for segmentation model with a unified decoder [31, 41, 93, 94], offering more flexible outputs. To break the jail for the arbitrary number of targets, DMMI [26] focuses on the one-to-many setting where text expression refers to varying numbers of targets. ReLA [38] proposes the Generalized Referring Expression Segmentation (GRES) task, supporting both the multi-target and empty-target scenarios, which is our main research scope. + +MLLM. Multimodal Large Language Models (MLLMs) align the vision and language modalities by various techniques, including cross-attention module [1], prompt tuning tokens [87], Q-Former [9, 34], linear projection layers [39], and unified model architectures [52, 79]. Endowed with unprecedented capabilities in reasoning with world-knowledge and following instructions of users, MLLM shows extraordinary performances in various vision language tasks [43, 65]. Equipped with the segmentation decoders or detection heads, MLLMs can also excel in the vision-centric tasks, such as object detection and segmentation [36, 56, 64, 80, 86]. Among them, LISA [32] makes the most of the reasoning ability with a [SEG] token to address the Reasoning Segmentation problem. However, LISA fails to tackle the challenge in GRES due to the inflexible [SEG] token, which is addressed by our proposed GSVA. + +Dynamic Networks. Dynamic neural networks [19] can adapt their architectures [20-22, 24, 27, 67-70] or parameters [11, 12, 53] to different inputs or switch the computation architecture in adjustment to different different time steps [46, 47, 71, 77], in order to achieve better accuracy and efficiency. In GSVA, the weight-sharing [SEG] tokens adapt to multiple targets under the hint of the prepended target prompts and dynamically reject empty targets with an individual prediction of rejection tokens. + +# 3. Generalized Segmentation Vision Assistant + +In this section, we initiate with the introduction of the model design of Generalized Segmentation Vision Assistant (GSVA), which is followed by the analysis of some certain limitations of LISA in Generalized Referring Segmentation (GRES). Subsequently, we delve into the introduction of two pivotal elements of GSVA, segmenting multiple targets and learning the rejection token. These components are fundamental in the conceptualization and construction of GSVA. + +![](images/3e61033a07e86c5ac476931d0362596077bece25caa38a012ce7bc873e60cd01.jpg) +Figure 2. Overview of GSVA. At the bottom of the figure, the MLLM encodes the input image and concatenates the tokenized text tokens to follow instructions. GSVA generates multiple [SEG] tokens to handle multiple referred targets and rejects the objects absent in the image through [REJ] tokens. At the top of the figure, the SFM also encodes the image for segmentation and selects all [SEG] tokens in the output sequence to prompt the mask decoder to segment the target objects referred to in the instructions. + +# 3.1. Model Architecture + +The architecture of GSVA is illustrated in Figure 2, resembling LISA [32], which enables high-fidelity segmentation outputs by integrating two types of foundation models: (1) Multimodal Large Language Model (MLLM) as an aligned vision-language cognitive module; (2) Segmentation Foundation Model (SFM) to segment the target out of the input image based on user's instruction. To connect these two modules, LISA proposes a paradigm named embedding as mask where an extra [SEG] token is appended to the vocabularies of the MLLM and serves as the prompt of the SFM to segment the target following the intention of the user. + +Multimodal Large Language Model. The MLLM consists of a decoder-based language model $F_{\mathrm{LLM}}$ to auto-regressively generate text responses following the user's inputs, a vision encoder $F_{\mathrm{V1}}$ to extract features from the input image, and a linear projector $\phi$ to align the representations between image and text modalities. Specifically, the pretrained LLaVA [39] variants with CLIP-ViT-L/14 [55] and Vicuna-7B/13B [6] are employed. Given an input image $\mathbf{x}_{\mathrm{img}}$ , the vision encoder $F_{\mathrm{V1}}$ first encodes it into image features, and then the projector $\phi$ maps the features into the visual token embeddings in the LLM input space: + +$$ +\mathbf {h} _ {\mathrm {i m g}} = \phi \left(F _ {\mathrm {V} 1} \left(\mathbf {x} _ {\mathrm {i m g}}\right)\right), \tag {1} +$$ + +where the input image $\mathbf{x}_{\mathrm{img}}$ is typically resized to $h\times w\times 3$ and the image tokens $\mathbf{h}_{\mathrm{img}}\in \mathbb{R}^{n_{\mathrm{img}}\times d}$ is aligned with the language modality. For CLIP-ViT-L/14, the input image with $h = w = 224$ is encoded with ViT of patch size in 14, therefore the length of tokens $n_{\mathrm{img}} = hw / 14^2 = 256$ and the LLM dimensions $d$ are 4096 and 5120 for Vicuna + +7B/13B, respectively. Along with the input image, the text instructions describing the targets to segment are tokenized into text tokens by the LLM tokenizer $T$ : + +$$ +\mathbf {h} _ {\mathrm {t x t}} = T (\mathbf {x} _ {\mathrm {t x t}}). \tag {2} +$$ + +The image tokens and text tokens are concatenated together and then fed into the LLM after prepending a sequence of fixed prompt tokens $\mathbf{h}_{\mathrm{prompt}}$ (omitted in the figure). The output token embeddings $\tilde{\mathbf{y}}_{\mathrm{txt}}$ are generated autoregressively: + +$$ +\tilde {\mathbf {y}} _ {\mathrm {t x t}} = F _ {\mathrm {L L M}} \left(\left[ \mathbf {h} _ {\text {p r o m p t}} \right| \left| \mathbf {h} _ {\text {i m g}} \right| \left| \mathbf {h} _ {\text {t x t}} \right|\right), \tag {3} +$$ + +where $||$ is concatenation operation. The text responses are obtained from $\tilde{\mathbf{y}}_{\mathrm{txt}}$ by applying a linear classifier to predict the next words in the vocabulary. + +In LISA, a special token [SEG] is appended in the vocabulary to activate the segmentation ability of MLLM. The model learns to predict [SEG] token in the output sequence to indicate there is a target to segment. LISA then selects the output embedding of the [SEG] token $\tilde{\mathbf{y}}_{\mathrm{xt}}[\mathrm{SEG}]$ , and projects it into the prompt space of the SFM by an MLP projector $\psi$ : + +$$ +\mathbf {h} _ {\mathrm {s e g}} = \psi (\tilde {\mathcal {Y}} _ {\mathrm {t x t}} [ \mathrm {S E G} ]). \tag {4} +$$ + +The segmentation model is hence ready to decode the target mask from the query token $\mathbf{h}_{\mathrm{seg}}$ + +Segmentation Foundation Model. The SFM is a query-based segmentation model, where a frozen vision encoder $F_{\mathrm{V2}}$ takes in images of higher resolution than the vision encoder in MLLM to keep more details, followed by a trainable mask decoder $F_{\mathrm{mask}}$ to decode masks from the queries. The pretrained SAM [31] with ViT-H backbone is instantiated as the SFM to produce high-quality masks. The given + +input image $\mathbf{x}_{\mathrm{img}}$ is resized to a larger resolution $H\times W\times 3$ with $H = W = 1024$ and encoded with $F_{\mathrm{V2}}$ to extract features $\mathbf{f}_{\mathrm{seg}}$ for segmentation: + +$$ +\mathbf {f} _ {\text {s e g}} = F _ {\mathrm {V} 2} \left(\mathbf {x} _ {\text {i m g}}\right). \tag {5} +$$ + +Condition on the features $\mathbf{f}_{\mathrm{seg}} \in \mathbb{R}^{\frac{H}{16} \times \frac{W}{16} \times C}$ with $C = 256$ for SAM, the mask decoder $F_{\mathrm{mask}}$ decodes the segmentation masks from the segmentation queries $\mathbf{h}_{\mathrm{seg}} \in \mathbb{R}^{N_{\mathrm{seg}} \times C}$ : + +$$ +\tilde {\mathbf {y}} _ {\text {m a s k}} = F _ {\text {m a s k}} \left(\mathbf {h} _ {\text {s e g}} | \mathbf {f} _ {\text {s e g}}\right), \tag {6} +$$ + +where each query in $\mathbf{h}_{\mathrm{seg}}$ corresponds to one segmentation mask in $\tilde{\mathbf{y}}_{\mathrm{mask}}$ . + +LISA assumes that only one target exists to segment in the input image and its corresponding instructions. However, in GSVA, we extend it to a new scenario with multiple targets and empty targets, including multiple [SEG] tokens to invoke segmentation and [REJ] tokens to reject unreasonable instructed targets absent in the image. As shown in Figure 2, GSVA supports multiple [SEG]/[REJ] tokens in the output sequence and selects all the [SEG] tokens and discards every [REJ] token after attaining $\tilde{\mathbf{y}}_{\mathrm{txt}}$ in Eq. (3). Therefore, there is more than one query in $\mathbf{h}_{\mathrm{seg}}$ , thus enabling the SFM to segment multiple targets. These designs make GSVA competent in the GRES task, where the awareness of multiple and empty targets is of vital importance. + +# 3.2. GRES: Task and Challenges + +Task. Generalized Referring Expression Segmentation (GRES) [38] removes the constraint on the number of referred targets in the expression in the conventional Referring Expression Segmentation (RES) [30, 44, 84]. Different from that one expression only refers to one instance or region in RES, GRES allows arbitrary numbers of referred targets, including multiple instances or no target circumstances. In GRES, the user can refer to many instances simultaneously or include the objects that do not exist in the image. For instance, there are three referring expressions in Figure 2, including a plate of broccoli and a spoon in the bowl of sauce, vase bottom left corner, and soup in bowl right and bowl bottom left corner. In the GRES case, the model ought to segment the masks of the objects referred to in the $1^{\text{st}}$ and $3^{\text{rd}}$ expression, meanwhile producing an empty mask for the $2^{\text{nd}}$ expressions since there are no vases present at the bottom left corner. + +Challenges. The challenges in GRES are common in practice, especially in embodied AI [17, 54, 58, 66]. The one challenge is multiple targets. Take robot navigation and planning [4, 45, 66] as an example, a robot may be asked to perceive multiple targets in the surrounding environment, e.g., to bring the two bowls of soup in Figure 2. The vision system of the robot needs to locate and segment the containers holding the referred food one at a time. The other challenge is empty target. Suppose the robot is ordered to cut an apple with a knife in the scene of Figure 2, whereas + +no apple is in the view of the camera, the vision system has to identify that the referred object is not in the scene. If it relies on some conventional RES methods which assume the expression must match something in the image, the output of the vision system could be undefined and potentially dangerous in some real-world cases. + +Differences from ReasonSeg. Reasoning Segmentation (ReasonSeg) proposed by LISA [32] emphasizes the complex text instructions in RES. In ReasonSeg, the instructions are more implicit and sophisticated, forcing models to reason using world knowledge. Besides, the logic chain is usually longer and more challenging in ReasonSeg, which requires the model to deduce the final target object referred to in the image. In contrast, GRES increases complexity in another dimension by involving complicated spatial relationships. Hence, the model has to learn to handle this spatial information and understand the relationships between the instances. To meet these requirements, LISA tunes MLLMs with complex instructions paired with masks, while GSVA arouses the spatial modeling capabilities of MLLMs by learning multiple targets and rejecting empty targets. + +# 3.3. Multiple [SEG] Tokens for Multiple Targets + +Single [SEG] token. LISA [32] follows the classic RES methods to generate a segmentation mask under given instructions by adding the [SEG] token into the answer. The prompt is formatted as: + +User: What is {obj} in this image? Please output segmentation mask. Assistant: Sure, it is [SEG]. + +In the above prompt, $\{obj\}$ represents an instance referred to or some semantic area to segment, and the embeddings of the [SEG] token output from the LLM are projected to prompt the SFM. When multiple instances are requested simultaneously, this prompt would confuse the model since only one [SEG] token is forced to match several targets. As shown in Figure 1 (a), LISA coercively predicts the masks of the left man and the guy second from the right, leading to a tattered mask and masking the wrong location. + +Multiple [SEG] tokens. To mitigate this issue in GRES, we relax this constraint for GSVA to support multiple target outputs via learning multiple [SEG] tokens. To avoid the ambiguity between the [SEG] tokens and the corresponding objects, we prepend the referring expression before each [SEG] token, i.e., each segmentation prompt in the output text is $\{obj\_n\} :[SEG]$ , as shown in Figure 3 (a). Streaming all the pairs of object and [SEG] token into one sentence, the question-answer prompt is formatted as (e.g., two targets): + +User: What are {obj1}, {obj2} in this image? Please output the segmentation masks. +Assistant: Sure, $\{obj1\} :[SEG]$ $\{obj2\} :[SEG]$ + +This prompt requests the MLLM to identify and distinguish different objects in the image based on the instructions and infuse the corresponding location information + +![](images/c06c0cf7c3251335e0eb02ab8135f9741a1798f23205e9fe1cd6f93724455a27.jpg) +(a) Multiple Targets +Figure 3. Example of the prompts and predicted masks of GSVA-Vicuna-7B drawn from gRefCOCO validation set. (a) depicts the multiple-target case, in which two zebras referred to are handled with two separate [SEG] tokens. (b) shows the empty-target case, where no apple is in the bowl. Thus, the null referent is rejected with a [REJ] token, and no segmentation mask will be generated. + +# User: + +What are front zebra, right zebra in this image? Please output segmentation masks. + +# Assistant: + +Sure, front zebra:[SEG], right zebra:[SEG]. + +![](images/a4c5d6ec3e13b820e727079189eb8de2c1e9b89b887bc43f7cffbbf63dc46067.jpg) +(b) Empty Target + +# User: + +What are the apple in the bowl, partial orange bottom left in this image? Please output segmentation masks. + +# Assistant: + +Sure, the apple in the bowl: [REJ], partial orange bottom left:[SEG]. + +of each target into the associated [SEG] token. We regard this ability as an implicit multimodal version of In-Context Learning (ICL), which is demonstrated by many prior works [2, 13, 33, 59, 90]. GSVA takes the target description preceding each [SEG] token as a hint to link this token to the object requested in the image through auto-regressive decoding. + +# 3.4. Rejecting Empty Targets via [REJ] Tokens + +Empty targets. In GRES, many expressions match no targets in the image, including absence, incorrect attributes, inaccurate locations, etc. These expressions should be treated as empty targets for models to predict all-negative masks. LISA [32] falls short of predicting masks with all-zeros seamlessly since the [SEG] token always calls for a segmentation mask. As shown in Figure 1 (b), LISA incorrectly marks a piece of green apple as the empty target banana. + +[REJ] token. We let the MLLM of GSVA predict a [REJ] token for each object that does not exist in the image but is referred to in the instructions, as shown in Figure 3 (b). GSVA predicts the targets marked with [REJ] tokens as empty targets, therefore setting all-zero masks for them. The involvement of [REJ] tokens directly rejects the empty target, liberating the mask decoder of the SFM from learning to segment the inconsistent targets. An example prompt with one existing target and one empty target is as follows: + +User: What are {obj1} (absent), {obj2} (absent), {obj3} in this image? Please output the segmentation masks. + +Assistant: Sure, $\{obj1\} :[REJ],\{obj2\} :[REJ],\{obj3\} :[SEG].$ + +The [REJ] token prediction can also be seen as a variant + +of VQA task, where the specified object and its position in the image need to be considered. Thanks to the unprecedented capabilities of MLLM in understanding the images [9, 34, 39, 92] and reasoning the spatial relationships of the referring objects [52, 88, 91], we make the most of the MLLM in GSVA to unleash the burden of the segmentation model. The proposed empty-target-aware mechanism both improves the quality of masks and ameliorates the errors of identifying nonexistent objects. + +# 4. Experiments + +In this section, we conduct comprehensive experiments to validate the efficacy of GSVA. First, we show the results on gRefCOCO [38] dataset to show the superiority of GSVA in GRES tasks. Then we verify GSVA that is also competent with other baselines in classic RES, REC tasks. We move on to ablate some important design choices of GSVA, followed by some qualitative visualization of GRES results. + +# 4.1. GRES + +Settings. We adopt gRefCOCO [38] dataset to validate GSVA and LISA [32] on GRES, which contains 278,232 expressions, including 80,022 multi-target and 32,202 empty-target ones, referring to the objects in 19,994 images. The images are split into four subsets: training, validation, test-A, and test-B, following the same UNC partition of RefCOCO [84]. We first add the gRefCOCO training set into the mixed training dataset in LISA to pretrain GSVA and LISA for 50,000 steps and then finetune the models on the gRefCOCO training dataset for another 10 epochs. We evaluate the pretrained and finetuned models on the remaining validation set, test set A, and test set B, respectively. We adopt the IoU, cIoU metrics for the segmentation mask outputs. Following the implementation in Liu et al. [38], IoU averages the IoU for each mask, whereas cIoU computes the cumulative intersection area over the cumulative union area across the whole dataset. As for the empty target, we compute the No-target-accuracy (N-acc.), which is the ratio of the correctly classified empty-target expressions over all the empty-target expressions in the dataset. For a correctly classified empty target, the gIoU is set to 1.0, and the cIoU does not take them into account, while for a misclassified empty target, the gIoU is set to 0.0 and the union area is accumulated in the cIoU. Following the criteria in Liu et al. [38], a predicted mask is regarded as empty for LISA if the positive pixels are less than 50, whereas GSVA predicts [REJ] tokens to identify empty targets. + +Results. We report the GRES segmentation results of GSVA and LISA [32] in Table 1. Three variants of GSVA, including GSVA-Vicuna-7B, GSVA-Vicuna-13B, and GSVA-Llama2-13B show competitive performance without finetuning on gIoU to the strongest non-LLM baseline ReLA [38]. However, LISA models fail to handle the + +
MethodGeneralized Referring Expression Segmentation on gRefCOCO
Validation SetTest Set ATest Set B
gIoUcIoUN-acc.gIoUcIoUN-acc.gIoUcIoUN-acc.
MattNet [85]48.2447.5141.1559.3058.6644.0446.1445.3341.32
LTS [29]52.7052.30-62.6461.87-50.4249.96-
VLT [10]52.0052.5147.1763.2062.1948.7450.8850.5247.82
CRIS [72]56.2755.34-63.4263.82-51.7951.04-
LAVT [81]58.4057.6449.3265.9065.3249.2555.8355.0448.46
ReLA [38]63.6062.4256.3770.0369.2659.0261.0259.8858.40
LISA-Vicuna-7B [32]32.2138.722.7148.5452.556.3739.6544.795.00
GSVA-Vicuna-7B63.3261.7056.4570.1169.2363.5061.3460.2658.42
LISA-Vicuna-7B [32] (ft)61.6361.7654.6766.2768.5050.0158.8460.6351.91
GSVA-Vicuna-7B (ft)66.4763.2962.4371.0869.9365.3162.2360.4760.56
LISA-Vicuna-13B [32]32.7339.853.6648.7653.624.8939.4945.354.41
GSVA-Vicuna-13B62.9760.1858.4467.1767.5954.6058.0657.2852.22
LISA-Vicuna-13B [32] (ft)63.4562.9955.2568.1869.6552.1661.8462.2456.15
GSVA-Vicuna-13B (ft)68.0164.0565.3671.7570.5167.2563.8361.2863.11
LISA-Llama2-13B [32]33.2639.643.2749.7653.807.2840.4945.415.73
GSVA-Llama2-13B63.2062.3854.5169.5269.8657.8462.0660.7758.30
LISA-Llama2-13B [32] (ft)65.2463.9657.4969.9971.0055.4362.1162.2956.34
GSVA-Llama2-13B (ft)70.0466.3866.0273.2972.7964.7265.4563.2062.47
+ +Table 1. Generalized referring expression segmentation (GRES) results on gRefCOCO [38] dataset. gIoU and cIoU are IoU metrics averaged by each example and accumulated over whole dataset, respectively. N-acc. is short for the accuracy of correctly classifying null targets. Baselines are copied from Liu et al. [38]. (ft) denotes the model is finetuned on the training set of gRefCOCO. + +GRES task without finetuning, showing degradation in both gIoU and cIoU in each model variant. Especially the low N-acc indicates that LISA is unable to correctly reject the empty targets. When finetuned on gRefCOCO training set, GSVA-Vicuna-7B performs better than the finetuned LISA counterpart, with about $4\%$ improvement in gIoU and over $5\%$ in N-acc on all three evaluation splits. GSVA variants with larger LLM incorporated further push the limits, achieving over $70\%$ in gIoU on the validation set, $73\%$ on test set A, and $65\%$ on test set B. The 13B models also consistently outperform LISA by large margins, demonstrating the superiority of GSVA in GRES task. + +# 4.2. Referring Expression Segmentation + +Settings. To validate the abilities to handle various tasks, we evaluate GSVA in the classic RES task. Following the common evaluation protocols, we test variants of GSVA and LISA equipped with different LLMs on RefCOCO, RefCOCO+ [30], and RefCOCOg [44]. We follow the UNC split to perform experiments on RefCOCO and RefCOCO+, and UMD split for RefCOCOg. The models are firstly pretrained as in GRES, and then finetuned for 10 epochs with a joint dataset of these three RES training set. The cIoU metric is adopted to measure the model performances. + +Results. Table 2 shows the RES results of GSVA. For the pretrained models, all the three variants of GSVA achieve higher cIoU than LISA [32] by clear margins. Our 7B model outperforms LISA-Vicuna about $2\%$ cIoU on almost every data split. After finetuning, the preponderance of GSVA over LISA keeps and even enlarges. Specifically, + +the cIoU metrics of GSVA-Llama2-13B on 8 sets surpass LISA by at least $2.8\%$ . For the test set of RefCOCOg, the margin even grows to $5.9\%$ , which exhibits GSVA is also competitive in the classic RES task. + +# 4.3. Referring Expression Comprehension + +Settings. Since GSVA is capable for RES tasks, it is natural to transfer to Referring Expression Comprehension (REC) tasks, by simply computing the bounding boxes of the masks. The datasets of REC are the same as RES, in which we evaluate GSVA. If the IoU of a predicted bounding box and the ground truth is greater than 0.5, this prediction is marked as correct. We use the same models in RES to evaluate the phrase grounding capability in REC tasks. + +Results. As shown in Table 3, we mainly compare our GSVA to LISA in the three variants. Without finetuning, GSVA-Vicuna-7B outperforms LISA with a large margin over $7\%$ in almost all evaluation sets. The similar trends also hold for the Vicuna-13B and Llama2-13B variants. GSVA also benefits a lot from finetuning, e.g., the 7B variant achieves consistently higher Prec@0.5 than uLLaVA-7B [80] with LoRA finetuning, which is another strong baseline that adopts the "mask2bbox" pipelines without direct bounding box supervision. The finetuned GSVA also shows competitive performances to the fully finetuned uLLaVA-7B, suggesting the strong potential of our method. With larger LLMs incorporated, the performances of GSVA continue with over $3\%$ increments over all datasets. + +
Referring Expression Segmentation on RefCOCO, RefCOCO+, and RefCOCOg
MethodRefCOCO (UNC)RefCOCO+ (UNC)RefCOCOg (UMD)
Val.Test-ATest-BVal.Test-ATest-BVal.Test
MCN [42]62.464.259.750.655.044.749.249.4
VLT [10]67.570.565.256.361.050.155.057.7
CRIS [72]70.573.266.162.368.153.759.960.4
LAVT [81]72.775.868.862.168.455.161.262.1
ReLA [38]73.876.570.266.071.057.765.066.0
X-Decoder [93]------64.6-
SEEM [94]------65.7-
PolyFormer-L [40]76.078.373.369.374.661.969.270.2
LISA-Vicuna-7B* [32]74.176.571.162.467.456.566.468.5
GSVA-Vicuna-7B76.477.472.864.567.758.671.172.0
LISA-Vicuna-7B* [32] (ft)74.979.172.365.170.858.167.970.6
GSVA-Vicuna-7B (ft)77.278.973.565.969.659.872.773.3
LISA-Vicuna-13B [32]71.774.768.159.464.252.965.266.1
GSVA-Vicuna-13B74.677.570.562.566.555.569.671.2
LISA-Vicuna-13B [32] (ft)76.078.872.965.070.258.169.570.5
GSVA-Vicuna-13B (ft)78.280.474.267.471.560.974.275.6
LISA-Llama2-13B [32]73.476.269.562.366.656.368.268.5
GSVA-Llama2-13B77.779.974.268.071.561.573.273.9
LISA-Llama2-13B [32] (ft)76.378.772.466.271.059.370.171.1
GSVA-Llama2-13B (ft)79.281.777.170.373.863.675.777.0
+ +Table 2. Referring expression segmentation results on RefCOCO, RefCOCO+ [30] and RefCOCOg [44] dataset. The cIoU metrics of each split are reported. Baselines are excerpted from Lai et al. [32]. (ft) denotes the models are finetuned on the joint training set of the referring expression segmentation datasets. * means the results are excerpted from the original paper. + +
MethodReferring Expression Comprehension on RefCOCO, RefCOCO+, and RefCOCOg
Val.RefCOCO Test-ATest-BVal.RefCOCO+ Test-ATest-BRefCOCOg Val.Test
u-LLaVA-7B [80] (LoRA)83.4787.1380.2168.7476.3260.9876.1978.24
u-LLaVA-7B [80] (full-ft)86.0489.4782.2674.0981.1666.6179.8781.68
LISA-Vicuna-7B [32]78.6881.7275.7462.9268.9356.4970.1072.47
GSVA-Vicuna-7B85.5088.0182.4970.2175.6265.1179.0079.21
LISA-Vicuna-7B [32] (ft)85.3988.8482.5974.2379.4668.4079.3480.42
GSVA-Vicuna-7B (ft)86.2789.2283.7772.8178.7868.0181.5881.83
LISA-Vicuna-13B [32]80.0183.2676.2663.7770.2457.4271.7973.34
GSVA-Vicuna-13B83.1287.0180.5468.1473.9062.0077.0878.89
LISA-Vicuna-13B [32] (ft)85.9289.0583.1674.8681.0868.8780.0981.48
GSVA-Vicuna-13B (ft)87.7190.4984.5776.5281.6970.3583.9084.85
LISA-Llama2-13B [32]82.5285.5678.8267.9173.7762.2575.3776.83
GSVA-Llama2-13B86.9989.5484.0873.8979.1069.3880.6882.07
LISA-Llama2-13B [32] (ft)85.9188.8481.7374.4680.5668.2680.0981.27
GSVA-Llama2-13B (ft)89.1692.0887.1779.7484.4573.4185.4786.18
+ +Table 3. Referring expression comprehension results on RefCOCO, RefCOCO+ [30] and RefCOCOg [44] dataset. The metric is precision @ 0.5 IoU threshold. (LoRA) means the LLM in u-LLaVA [80] is finetuned with LoRA adapter, as LISA and GSVA, while (full-ft) represents the LLM in u-LLaVA is fully trained. Results of u-LLaVA-7B with "mask2bbox" strategy are reported for fair comparison. + +# 4.4. Ablation Study + +The involvement of the [REJ] token. [REJ] token plays a rather important role in GSVA. We study the effect of the [REJ] token by removing it from the vocabulary in GSVA, yielding a variant unable to reject a target from the text outputs. As shown in the $2^{\text{nd}}$ row of Table 4, after removing [REJ] token, there is a sharp N-acc drop over $25\%$ relatively, followed by the decline of gIoU at about $10\%$ on gRefCOCO validation set. This performance degradation + +indicates the significance of LLM learning a special token to reject the referred instances absent in the image. + +Learning multiple [SEG] tokens. We continue to ablate the multiple [SEG] tokens, which is another core design of GSVA. After removing [REJ] token, we then reduce the number of [SEG] tokens to 1, which is identical to LISA [32] with the referring expression added before the only one [SEG] token: Assistant: Sure, {obj}:[SEG]. In the early experiments, we have found that stacking multiple expressions before one [SEG] token would result in diver + +![](images/daf746cea54f9db87410b369a5fded0fe5eb165343920dc34d492be4a37d8a60.jpg) +(a) Multiple Target Cases + +![](images/4ec7b27483a384f355e27e2cfe72aaa48709e263fe2ea74aea250884f6c9d306.jpg) + +![](images/26a9add807664ac57a40ce031e8cfa0cbfcb8a0cabd1c34520fc53fb4a2ca992.jpg) + +![](images/ed01be79d76617b81ca32630c63cc333515e0ff7504186c73807b006edfddbb7.jpg) +Elephant in back, Left elephant +Kid on left in blue, Boy on right, Blond boy on right, Girl back left pink +Figure 4. Visualizations of GSVA and LISA [32] in the GRES task. The first row shows LISA's segmentation results, the second row is the masks and rejections of GSVA, and the third row shows the referring expressions in the instructions. In (a) multiple target cases, each target is colored with a specific color. In (b) empty target cases, the images turn darker to highlight the incorrect predictions of LISA. The examples are selected from the gRefCOCO validation set. The masks are generated by the 7B models. Zoom in for the best view. + +![](images/d172daaca66210060b87ca57f49113019b04b76b1f2a651ffd246bea7c3e41f9.jpg) +Tan vase, Bottom vase, Right vase + +![](images/ec875e834067d302b43ef1aa075dabff7c15cec5af9e6b2f3eaf5532e0bd984e.jpg) + +![](images/465f562d5740bd84758eb8a00204d29617174eaf3afd6cc7f9881ea051ad3854.jpg) +(b) Empty Target Cases + +![](images/0a39cdcf3e583ce5afec93306543ceeef6189aa388dc51808c9ceff98d38844a.jpg) +Lady on left + +![](images/11400d90a2e5f11667b43e3559b917794630bd1b2dcba7d6930afd4bdf80229d.jpg) + +![](images/638204a0c6903c24c88310a1b409bd3dd8a2eb86eb758446f56062d76e6b39c8.jpg) +Teddy with X eye + +![](images/fa154cc8e320f6229a187f7766e439434cfccf48d12a0602cc780105dcb7b891.jpg) + +![](images/dbb082b818b53a9e2883bff956ccc92ac2f7ebb36a68ddbaf3d98fe4e3d0284b.jpg) +The dog on the suitcase + +
Model w/Vicuna-7BModificationsgRefCOCO Val.
RefExp. +[SEG]Multiple [SEG][REJ] TokengIoUcIoUN-acc.
GSVA✓✓✓63.3261.7056.45
✓✓✗51.5760.9530.32
✓✗✗44.8659.3711.96
LISA [32]✗✗✗32.2138.722.71
✗✓✓21.8327.220.00
+ +Table 4. Ablation study on the core designs of GSVA. $\checkmark$ means the employment of the component while $X$ means not. "Ref-Exp.+[SEG]", "Multiple [SEG]", and "[REJ] Token" are short for adding referring expression before [SEG] in the answer prompt, using multiple [SEG] tokens, involving [REJ] token, respectively. + +gence. Therefore we separate multiple targets to prompt the model with one expression at a time. The sharp decrements of gIoU by nearly $7\%$ and N-acc by over $15\%$ in the $3^{\mathrm{rd}}$ row demonstrate the significance of the multiple-[SEG]-token. + +Answers without referring expression. To examine the efficacy of the hinting prompts, we remove all the referring expressions before the [SEG] tokens. Based on the removal of multiple [SEG]s and [REJ]s, erasing the added referring expression falls back to the original LISA model, as shown in the $4^{\text{th}}$ row. We further choose only to remove it from GSVA model, keeping other configurations unchanged. Specifically, if there are two referents, the prompts in the answer will turn to Assistant: Sure, [SEG], [SEG], whose results are in the last row. The zero N-acc shows the model fails to identify any empty target without the help of the expressions, meanwhile the poor gIoU and cIoU indicates the segmentation ability is damaged. This phenomenon also suggests that the added referring expression hint GSVA to associate each [SEG] token to its corresponding target, which is in coherence with our hypothesis. + +# 4.5. Visualization + +We visualize some qualitative results of GSVA to verify its effectiveness. As shown in Figure 4, we present two groups + +of examples from the validation set of gRefCOCO [38] to see how GSVA outperforms LISA in the face of the two main challenges in GRES: multiple targets and empty targets. In part (a), GSVA has managed to segment all the targets referred to, while LISA could only segment one of the requested instances. For example, in the third column, LISA only predicts the mask of the rightmost vase. On the contrary, GSVA separately segments all three targeted vases. In part (b), LISA mistakenly segments the instances in the image that disagree with the referring expression, e.g., in the sixth column, LISA proposes a mask of a cat in response to the request of the dog on the suitcase, whereas GSVA successfully reject all the empty targets. + +# 5. Conclusion + +This paper introduces a novel multimodal large language model dubbed Generalized Segmentation Vision Assistant (GSVA). By introducing multiple [SEG] tokens and the new [REJ] token, GSVA effectively achieves multi-objective segmentation and empty target rejection, which addresses the challenging segmentation problems in practical application scenarios, e.g., Generalized Referring Expression Segmentation (GRES). Extensive experiments on GRES, classic RES, and REC tasks fully demonstrate the superior performance of our method, highlighting its significance for future research and applications. + +# Acknowledgments + +This work is supported in part by the National Key R&D Program of China under Grant 2021ZD0140407, the National Natural Science Foundation of China under Grants 62321005 and 62276150. We also thank Dr. Yuan Yao and Prof. Zhiyuan Liu for their insightful and valuable comments on this research project. We also appreciate their generous support of computational resources. + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022. 2 +[2] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, 2020. 2, 5 +[3] Ding-Jie Chen, Songhao Jia, Yi-Chen Lo, Hwann-Tzong Chen, and Tyng-Luh Liu. See-through-text grouping for referring image segmentation. In IEEE ICCV, 2019. 2 +[4] Yixin Chen, Qing Li, Deqian Kong, Yik Lun Kei, Song-Chun Zhu, Tao Gao, Yixin Zhu, and Siyuan Huang. Yourefit: Embodied reference understanding with language and gesture. In IEEE ICCV, 2021. 2, 4 +[5] Ming-Ming Cheng, Shuai Zheng, Wen-Yan Lin, Vibhav Vineet, Paul Sturgess, Nigel Crook, Niloy J Mitra, and Philip Torr. Imagesprit: Verbal guided image parsing. ACM ToG, 2014. 1, 2 +[6] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ * chatgpt quality, 2023. 2, 3 +[7] Yong Xien Chng, Henry Zheng, Yizeng Han, Xuchong Qiu, and Gao Huang. Mask grounding for referring image segmentation. In IEEE CVPR, 2024. 2 +[8] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416, 2022. 2 +[9] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. In NeurIPS, 2023. 2, 5 +[10] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vision-language transformer and query generation for referring segmentation. In IEEE ICCV, 2021. 6, 7 +[11] Xiaohan Ding, Xiangyu Zhang, Ningning Ma, Jungong Han, Guiguang Ding, and Jian Sun. Repvgg: Making vgg-style convnets great again. In IEEE ICCV, 2021. 2 +[12] Xiaohan Ding, Xiangyu Zhang, Jungong Han, and Guiguang Ding. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In IEEE CVPR, 2022. 2 +[13] Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Zhiyong Wu, Baobao Chang, Xu Sun, Jingjing Xu, and Zhifang Sui. A survey for in-context learning. arXiv preprint arXiv:2301.00234, 2022.5 +[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2 + +[15] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale. In IEEE CVPR, 2023. 2 +[16] Guang Feng, Zhiwei Hu, Lihe Zhang, and Huchuan Lu. Encoder fusion network with co-attention embedding for referring image segmentation. In IEEE CVPR, 2021. 2 +[17] Chen Gao, Jinyu Chen, Si Liu, Luting Wang, Qiong Zhang, and Qi Wu. Room-and-object aware knowledge reasoning for remote embodied referring expression. In IEEE CVPR, 2021. 1, 4 +[18] Dongchen Han, Tianzhu Ye, Yizeng Han, Zhuofan Xia, Shiji Song, and Gao Huang. Agent attention: On the integration of softmax and linear attention. arXiv preprint arXiv:2312.08874, 2023. 2 +[19] Yizeng Han, Gao Huang, Shiji Song, Le Yang, Honghui Wang, and Yulin Wang. Dynamic neural networks: A survey. IEEE TPAMI, 2021. 2 +[20] Yizeng Han, Gao Huang, Shiji Song, Le Yang, Yitian Zhang, and Haojun Jiang. Spatially adaptive feature refinement for efficient inference. IEEE TIP, 2021. 2 +[21] Yizeng Han, Yifan Pu, Zihang Lai, Chaofei Wang, Shiji Song, Junfeng Cao, Wenhui Huang, Chao Deng, and Gao Huang. Learning to weight samples for dynamic early-exiting networks. In ECCV, 2022. +[22] Yizeng Han, Zhihang Yuan, Yifan Pu, Chenhao Xue, Shiji Song, Guangyu Sun, and Gao Huang. Latency-aware spatialwise dynamic networks. In NeurIPS, 2022. 2 +[23] Yizeng Han, Dongchen Han, Zeyu Liu, Yulin Wang, Xuran Pan, Yifan Pu, Chao Deng, Junlan Feng, Shiji Song, and Gao Huang. Dynamic perceiver for efficient visual recognition. In IEEE ICCV, 2023. 2 +[24] Yizeng Han, Zeyu Liu, Zhihang Yuan, Yifan Pu, Chaofei Wang, Shiji Song, and Gao Huang. Latency-aware unified dynamic networks for efficient image recognition. arXiv preprint arXiv:2308.15949, 2023. 2 +[25] Ronghang Hu, Marcus Rohrbach, and Trevor Darrell. Segmentation from natural language expressions. In ECCV, 2016. 1, 2 +[26] Yutao Hu, Qixiong Wang, Wenqi Shao, Enze Xie, Zhenguo Li, Jungong Han, and Ping Luo. Beyond one-to-one: Rethinking the referring image segmentation. In IEEE ICCV, 2023. 2 +[27] Gao Huang, Yulin Wang, Kangchen Lv, Haojun Jiang, Wenhui Huang, Pengfei Qi, and Shiji Song. Glance and focus networks for dynamic visual recognition. IEEE TPAMI, 2022. 2 +[28] Rui Huang, Xuran Pan, Henry Zheng, Haojun Jiang, Zhifeng Xie, Cheng Wu, Shiji Song, and Gao Huang. Joint representation learning for text and 3d point cloud. Pattern Recognition, 2024. 1 +[29] Ya Jing, Tao Kong, Wei Wang, Liang Wang, Lei Li, and Tie-niu Tan. Locate then segment: A strong pipeline for referring image segmentation. In IEEE CVPR, 2021. 6 +[30] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In EMNLP, 2014. 1, 4, 6, 7 + +[31] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In IEEE ICCV, 2023. 2, 3 +[32] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692, 2023. 1, 2, 3, 4, 5, 6, 7, 8 +[33] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Fanyi Pu, Jingkang Yang, Chunyuan Li, and Ziwei Liu. Mimicit: Multi-modal in-context instruction tuning. arXiv preprint arXiv:2306.05425, 2023. 5 +[34] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 2, 5 +[35] Ruiyu Li, Kaican Li, Yi-Chun Kuo, Michelle Shu, Xiaojuan Qi, Xiaoyong Shen, and Jiaya Jia. Referring image segmentation via recurrent refinement networks. In IEEE CVPR, 2018. 2 +[36] Ziyi Lin, Chris Liu, Renrui Zhang, Peng Gao, Longtian Qiu, Han Xiao, Han Qiu, Chen Lin, Wenqi Shao, Keqin Chen, et al. Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models. arXiv preprint arXiv:2311.07575, 2023. 2 +[37] Chenxi Liu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, and Alan Yuille. Recurrent multimodal interaction for referring image segmentation. In IEEE CVPR, 2017. 2 +[38] Chang Liu, Henghui Ding, and Xudong Jiang. Gres: Generalized referring expression segmentation. In IEEE CVPR, 2023. 1, 2, 4, 5, 6, 7, 8 +[39] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 2, 3, 5 +[40] Jiang Liu, Hui Ding, Zhaowei Cai, Yuting Zhang, Ravi Kumar Satzoda, Vijay Mahadevan, and R Manmatha. Polyformer: Referring image segmentation as sequential polygon generation. In IEEE CVPR, 2023. 7 +[41] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 2 +[42] Gen Luo, Yiyi Zhou, Xiaoshuai Sun, Liujuan Cao, Chenglin Wu, Cheng Deng, and Rongrong Ji. Multi-task collaborative network for joint referring expression comprehension and segmentation. In IEEE CVPR, 2020. 7 +[43] Tengchao Lv, Yupan Huang, Jingye Chen, Lei Cui, Shuming Ma, Yaoyao Chang, Shaohan Huang, Wenhui Wang, Li Dong, Weiyao Luo, et al. Kosmos-2.5: A multimodal literate model. arXiv preprint arXiv:2309.11419, 2023. 2 +[44] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In IEEE CVPR, 2016. 2, 4, 6, 7 +[45] Khanh Nguyen, Debadeepta Dey, Chris Brockett, and Bill Dolan. Vision-based navigation with language-based assistance via imitation learning with indirect intervention. In IEEE CVPR, 2019. 4 + +[46] Zanlin Ni, Yulin Wang, Jiangwei Yu, Haojun Jiang, Yue Cao, and Gao Huang. Deep incubation: Training large models by divide-and-conquering. In IEEE CVPR, 2023. 2 +[47] Zanlin Ni, Yulin Wang, Renping Zhou, Jiayi Guo, Jinyi Hu, Zhiyuan Liu, Shiji Song, Yuan Yao, and Gao Huang. Revisiting non-autoregressive transformers for efficient image synthesis. In CVPR, 2024. 2 +[48] OpenAI. Gpt-4 technical report, 2023. 2 +[49] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. In NeurIPS, 2022. 2 +[50] Xuran Pan, Chunjiang Ge, Rui Lu, Shiji Song, Guanfu Chen, Zeyi Huang, and Gao Huang. On the integration of self-attention and convolution. In IEEE CVPR, 2022. 2 +[51] Xuran Pan, Tianzhu Ye, Zhuofan Xia, Shiji Song, and Gao Huang. Slide-transformer: Hierarchical vision transformer with local self-attention. In IEEE CVPR, 2023. 2 +[52] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 2, 5 +[53] Yifan Pu, Yiru Wang, Zhuofan Xia, Yizeng Han, Yulin Wang, Weihao Gan, Zidong Wang, Shiji Song, and Gao Huang. Adaptive rotated convolution for rotated object detection. In IEEE ICCV, 2023. 2 +[54] Yuankai Qi, Qi Wu, Peter Anderson, Xin Wang, William Yang Wang, Chunhua Shen, and Anton van den Hengel. Reverie: Remote embodied visual referring expression in real indoor environments. In IEEE CVPR, 2020. 1, 4 +[55] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3 +[56] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M Anwer, Erix Xing, Ming-Hsuan Yang, and Fahad S Khan. Glamm: Pixel grounding large multimodal model. arXiv preprint arXiv:2311.03356, 2023. 2 +[57] Hengcan Shi, Hongliang Li, Fanman Meng, and Qingbo Wu. Key-word-aware network for referring expression image segmentation. In ECCV, 2018. 2 +[58] Qie Sima, Sinan Tan, Huaping Liu, Fuchun Sun, Weifeng Xu, and Ling Fu. Embodied referring expression for manipulation question answering in interactive environment. In IEEE ICRA, 2023. 1, 4 +[59] Yan Tai, Weichen Fan, Zhao Zhang, Feng Zhu, Rui Zhao, and Ziwei Liu. Link-context learning for multimodal llms. arXiv preprint arXiv:2308.07891, 2023. 5 +[60] Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. Stanford alpaca: An instruction-following llama model, 2023. 2 + +[61] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. +[62] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2 +[63] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, et al. Image as a foreign language: Beit pretraining for vision and vision-language tasks. In IEEE CVPR, 2023. 2 +[64] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, et al. Vision: Large language model is also an open-ended decoder for vision-centric tasks. In NeurlPS, 2023. 2 +[65] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, et al. Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079, 2023. 2 +[66] Xin Wang, Qiuyuan Huang, Asli Celikyilmaz, Jianfeng Gao, Dinghan Shen, Yuan-Fang Wang, William Yang Wang, and Lei Zhang. Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation. In IEEE CVPR, 2019. 1, 4 +[67] Yulin Wang, Zhaoxi Chen, Haojun Jiang, Shiji Song, Yizeng Han, and Gao Huang. Adaptive focus for efficient video recognition. In IEEE ICCV, 2021. 2 +[68] Yulin Wang, Rui Huang, Shiji Song, Zeyi Huang, and Gao Huang. Not all images are worth 16x16 words: Dynamic transformers for efficient image recognition. In NeurIPS, 2021. +[69] Yulin Wang, Yang Yue, Yuanze Lin, Haojun Jiang, Zihang Lai, Victor Kulikov, Nikita Orlov, Humphrey Shi, and Gao Huang. Adafocus v2: End-to-end training of spatial dynamic networks for video recognition. In IEEE CVPR, 2022. +[70] Yulin Wang, Yang Yue, Xinhong Xu, Ali Hassani, Victor Kulikov, Nikita Orlov, Shiji Song, Humphrey Shi, and Gao Huang. Adafocusv3: On unified spatial-temporal dynamic video recognition. In ECCV, 2022. 2 +[71] Yulin Wang, Yang Yue, Rui Lu, Tianjiao Liu, Zhao Zhong, Shiji Song, and Gao Huang. Efficienttrain: Exploring generalized curriculum learning for training visual backbones. In IEEE ICCV, 2023. 2 +[72] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clip-driven referring image segmentation. In IEEE CVPR, 2022. 6, 7 +[73] Zhichao Wei, Xiaohao Chen, Mingqiang Chen, and Siyu Zhu. Learning aligned cross-modal representations for referring image segmentation. arXiv preprint arXiv:2301.06429, 2023. 2 +[74] Penghao Wu and Saining Xie. V*: Guided visual search as a core mechanism in multimodal llms. arXiv preprint arXiv:2312.14135, 2023. 2 + +[75] Tsung-Han Wu, Giscard Biamby, David Chan, Lisa Dunlap, Ritwik Gupta, Xudong Wang, Joseph E Gonzalez, and Trevor Darrell. See, say, and segment: Teaching Imms to overcome false premises. arXiv preprint arXiv:2312.08366, 2023. 2 +[76] Zhuofan Xia, Xuran Pan, Shiji Song, Li Erran Li, and Gao Huang. Vision transformer with deformable attention. In IEEE CVPR, 2022. 2 +[77] Zhuofan Xia, Xuran Pan, Xuan Jin, Yuan He, Hui Xue, Shiji Song, and Gao Huang. Budgeted training for vision transformer. In ICLR, 2023. 2 +[78] Zhuofan Xia, Xuran Pan, Shiji Song, Li Erran Li, and Gao Huang. Dat++: Spatially dynamic vision transformer with deformable attention. arXiv preprint arXiv:2309.01430, 2023.2 +[79] Bin Xiao, Haiping Wu, Weijian Xu, Xiyang Dai, Houdong Hu, Yumao Lu, Michael Zeng, Ce Liu, and Lu Yuan. Florence-2: Advancing a unified representation for a variety of vision tasks. arXiv preprint arXiv:2311.06242, 2023. 2 +[80] Jinjin Xu, Liwu Xu, Yuzhe Yang, Xiang Li, Yanchun Xie, Yi-Jie Huang, and Yaqian Li. u-llava: Unifying multimodal tasks via large language model. arXiv preprint arXiv:2311.05348, 2023. 2, 6, 7 +[81] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Lavt: Language-aware vision transformer for referring image segmentation. In IEEE CVPR, 2022. 2, 6, 7 +[82] Linwei Ye, Mrigank Rochan, Zhi Liu, and Yang Wang. Cross-modal self-attention network for referring image segmentation. In IEEE CVPR, 2019. 2 +[83] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 2 +[84] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In ECCV, 2016. 4, 5 +[85] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In IEEE CVPR, 2018. 6 +[86] Ao Zhang, Liming Zhao, Chen-Wei Xie, Yun Zheng, Wei Ji, and Tat-Seng Chua. Next-chat: An Imm for chat, detection and segmentation. arXiv preprint arXiv:2311.04498, 2023. 2 +[87] Renrui Zhang, Jiaming Han, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, Peng Gao, and Yu Qiao. Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023. 2 +[88] Shilong Zhang, Peize Sun, Shoufa Chen, Min Xiao, Wenqi Shao, Wenwei Zhang, Kai Chen, and Ping Luo. Gpt4roi: Instruction tuning large language model on region-of-interest. arXiv preprint arXiv:2307.03601, 2023. 5 + +[89] Zicheng Zhang, Yi Zhu, Jianzhuang Liu, Xiaodan Liang, and Wei Ke. Coupalign: Coupling word-pixel with sentence-mask alignments for referring image segmentation. In NeurIPS, 2022. 2 +[90] Haozhe Zhao, Zefan Cai, Shuzheng Si, Xiaojian Ma, Kaikai An, Liang Chen, Zixuan Liu, Sheng Wang, Wenjuan Han, and Baobao Chang. Mmicl: Empowering vision-language model with multi-modal in-context learning. arXiv preprint arXiv:2309.07915, 2023. 5 +[91] Yang Zhao, Zhijie Lin, Daquan Zhou, Zilong Huang, Jiashi Feng, and Bingyi Kang. Bubogpt: Enabling visual grounding in multi-modal llms. arXiv preprint arXiv:2307.08581, 2023. 5 +[92] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 5 +[93] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In IEEE CVPR, 2023. 2, 7 +[94] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. In NeurIPS, 2023. 2, 7 \ No newline at end of file diff --git a/2024/GSVA_ Generalized Segmentation via Multimodal Large Language Models/images.zip b/2024/GSVA_ Generalized Segmentation via Multimodal Large Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..cb561b78900203a8114f25bc4232a0be2fcbc3e8 --- /dev/null +++ b/2024/GSVA_ Generalized Segmentation via Multimodal Large Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0b8ebaff5399e8c285d921f760c19d1a02f68ba6975ae320fdbc2f94481b61ff +size 726053 diff --git a/2024/GSVA_ Generalized Segmentation via Multimodal Large Language Models/layout.json b/2024/GSVA_ Generalized Segmentation via Multimodal Large Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0ac6ac814e3f9d4dd986a25f0446958a7bbbbacd --- /dev/null +++ b/2024/GSVA_ Generalized Segmentation via Multimodal Large Language Models/layout.json @@ -0,0 +1,9739 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 66, + 103, + 527, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 103, + 527, + 121 + ], + "spans": [ + { + "bbox": [ + 66, + 103, + 527, + 121 + ], + "type": "text", + "content": "GSVA: Generalized Segmentation via Multimodal Large Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 84, + 143, + 512, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 143, + 512, + 159 + ], + "spans": [ + { + "bbox": [ + 84, + 143, + 512, + 159 + ], + "type": "text", + "content": "Zhuofan Xia* Dongchen Han* Yizeng Han Xuran Pan Shiji Song Gao Huang†" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 179, + 159, + 411, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 159, + 411, + 171 + ], + "spans": [ + { + "bbox": [ + 179, + 159, + 411, + 171 + ], + "type": "text", + "content": "Department of Automation, BNRist, Tsinghua University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 199, + 192, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 199, + 192, + 211 + ], + "spans": [ + { + "bbox": [ + 143, + 199, + 192, + 211 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 224, + 289, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 224, + 289, + 571 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 289, + 571 + ], + "type": "text", + "content": "Generalized Referring Expression Segmentation (GRES) extends the scope of classic RES to refer to multiple objects in one expression or identify the empty targets absent in the image. GRES poses challenges in modeling the complex spatial relationships of the instances in the image and identifying non-existing referents. Multimodal Large Language Models (MLLMs) have recently shown tremendous progress in these complicated vision-language tasks. Connecting Large Language Models (LLMs) and vision models, MLLMs are proficient in understanding contexts with visual inputs. Among them, LISA, as a representative, adopts a special [SEG] token to prompt a segmentation mask decoder; e.g., SAM, to enable MLLMs in the RES task. However, existing solutions to GRES remain unsatisfactory since current segmentation MLLMs cannot correctly handle the cases where users might reference multiple subjects in a singular prompt or provide descriptions incongruent with any image target. In this paper, we propose Generalized Segmentation Vision Assistant (GSVA) to address this gap. Specifically, GSVA reuses the [SEG] token to prompt the segmentation model towards supporting multiple mask references simultaneously and innovatively learns to generate a [REJ] token to reject the null targets explicitly. Experiments validate GSVA's efficacy in resolving the GRES issue, marking a notable enhancement and setting a new record on the GRES benchmark gRefCOCO dataset. GSVA also proves effective across various classic referring segmentation and comprehension tasks. Code is available at https://github.com/LeapLabTHU/GSVA." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 594, + 127, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 594, + 127, + 605 + ], + "spans": [ + { + "bbox": [ + 47, + 594, + 127, + 605 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 613, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 613, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 613, + 287, + 685 + ], + "type": "text", + "content": "Referring Expression Segmentation (RES) [5, 25] is an emerging vision-language (VL) task predicting the masks of the interested objects referred to in the language expression. RES has great potential in many areas, especially embodied AI [17, 28, 54, 58, 66], including VL navigation, humanrobot interaction, etc. Nevertheless, the simplification in" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 200, + 541, + 357 + ], + "blocks": [ + { + "bbox": [ + 310, + 200, + 541, + 357 + ], + "lines": [ + { + "bbox": [ + 310, + 200, + 541, + 357 + ], + "spans": [ + { + "bbox": [ + 310, + 200, + 541, + 357 + ], + "type": "image", + "image_path": "e2387831eaffe94f5cd97cbfd1215f1b02e9a6c9b6249555ee5eca4f28302746.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 365, + 545, + 476 + ], + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 476 + ], + "type": "text", + "content": "Figure 1. Comparison of the segmentation masks by LISA [32] and GSVA, facing the challenges in Generalized Referring Expression Segmentation (GRES) [38]. (a) LISA fails to segment the correct targets when multiple targets are requested due to the single [SEG] token restriction. GSVA successfully generates all target masks via learning multiple [SEG] tokens. (b) When the referent does not exist in the image, i.e., the empty target is requested, LISA reluctantly produces the wrong mask because of the compulsive [SEG] token output. In contrast, GSVA can reject the empty targets by predicting [REJ] tokens in the output sequence." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 483, + 545, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 545, + 532 + ], + "type": "text", + "content": "RES formulation that one referring expression must match an individual object in the image [30] has left a gap between current RES algorithms and real-world applications, neglecting multiple-target and empty-target cases." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": "To bridge this gap, Generalized Referring Expression Segmentation (GRES) [38] has recently been proposed to support multiple-target and empty-target cases. Practically, users refer to multiple subjects within a single instruction or provide descriptions that do not correspond to any targets in the image. As an example shown in Figure 1 (a), the left man and the second man from the right are targeted simultaneously, while a banana is referred to in a scene of apples in Figure 1 (b). RES takes no account of these cases, which GRES handles. In addition to the multimodal correspondences between images and text prompts in classic RES, GRES poses new challenges in handling more complicated multiple-target and empty-target cases. Therefore, the models must handle complex spatial relationships of the instances in the image [38] to segment the targets at various" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 127, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 127, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 127, + 703 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3858" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 282, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 282, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 282, + 85 + ], + "type": "text", + "content": "locations and reject the empty targets in the wrong places." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 87, + 287, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 87, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 87, + 287, + 277 + ], + "type": "text", + "content": "The recent blooming Multimodal Large Language Models (MLLMs) [1, 34, 39, 83, 87] meet the requirements of GRES since they show excellency in complex reasoning [8] and instruction following [49] with visual inputs by aligning the LLMs [2, 6, 48, 60-62] and Visual Foundation Models (VFMs) [15, 55, 63, 79] which are typically various Vision Transformers [14, 18, 23, 50, 51, 76, 78] to perceive image or video inputs. To support segmentation output, many works [74, 75, 86] link an MLLM (e.g., LLaVA [39]) and a segmentation model (e.g., SAM [31]) by prompting the decoder with special token embeddings (e.g., [SEG] in LISA [32]) to generate masks of the referents in the user's instructions. Although these models manage to handle RES, GRES is still beyond their reach. As shown in Figure 1, LISA fails to work well in GRES where the multiple-target and empty-target challenges remain uncharted." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 281, + 289, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 289, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 289, + 579 + ], + "type": "text", + "content": "To address the above challenges, we propose Generalized Segmentation Vision Assistant (GSVA). We attribute the vulnerability of other segmentation MLLMs in GRES to the single constant [SEG] token that restricts its flexibility. Therefore, we present two pivotal designs in GSVA: (1) learning to predict multiple [SEG] tokens to segment multiple targets; (2) rejecting empty targets in referring expressions by predicting [REJ] tokens. Specifically, when multiple targets are requested in the referring expression, we place multiple weight-sharing [SEG] tokens corresponding to the entities in the expression, encouraging the MLLM to learn to output multiple [SEG] tokens. To distinguish each [SEG] token and avoid ambiguity, we add the expression of each entity in front of the corresponding [SEG] token, hinting each [SEG] token to focus on the specific target, which can be regarded as implicit In-Context Learning, and dynamic neural network [19]. Meanwhile, if the referents are absent in the image, the corresponding [SEG] tokens after the prompts are altered to [REJ] tokens to identify empty targets. The predicted [REJ] tokens are directly assigned with empty masks without decoded, which liberates the segmentation model from seeking non-existing targets in the image. This Benefiting from these novel designs, GSVA takes a big step forward in addressing GRES challenges, as shown in the second row of Figure 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": "Our contributions are summarized as follows: (1) We propose GSVA to solve the GRES problem with MLLM by handling the spatial relationships among targets, and study the GRES problem systematically in the context of LLM for the first time. (2) We propose the non-trivial shared-weight multiple [SEG] tokens guided by each referent prompt to address the multiple-target problem. (3) We firstly propose a clean solution, the [REJ] token, to reject the empty targets, which can be seamlessly applied to various models. (4) GSVA is intuitive and effective, achieving state-of-the-art performance on the GRES benchmarks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 71, + 398, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 398, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 398, + 84 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 91, + 545, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 545, + 272 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 545, + 272 + ], + "type": "text", + "content": "RES and GRES. Referring Expression Segmentation (RES) [4, 5, 25, 44] assumes that one expression matches one existing target, and many works explore fusing image and language [3, 16, 35, 37, 81] to segment objects under instructions. Currently, most RES methods adopt the cross-attention module or cross-modal alignment to bridge the modality gap [7, 57, 73, 82, 89]. Another line of research enables the text prompts for segmentation model with a unified decoder [31, 41, 93, 94], offering more flexible outputs. To break the jail for the arbitrary number of targets, DMMI [26] focuses on the one-to-many setting where text expression refers to varying numbers of targets. ReLA [38] proposes the Generalized Referring Expression Segmentation (GRES) task, supporting both the multi-target and empty-target scenarios, which is our main research scope." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 272, + 546, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 272, + 546, + 463 + ], + "spans": [ + { + "bbox": [ + 304, + 272, + 546, + 463 + ], + "type": "text", + "content": "MLLM. Multimodal Large Language Models (MLLMs) align the vision and language modalities by various techniques, including cross-attention module [1], prompt tuning tokens [87], Q-Former [9, 34], linear projection layers [39], and unified model architectures [52, 79]. Endowed with unprecedented capabilities in reasoning with world-knowledge and following instructions of users, MLLM shows extraordinary performances in various vision language tasks [43, 65]. Equipped with the segmentation decoders or detection heads, MLLMs can also excel in the vision-centric tasks, such as object detection and segmentation [36, 56, 64, 80, 86]. Among them, LISA [32] makes the most of the reasoning ability with a [SEG] token to address the Reasoning Segmentation problem. However, LISA fails to tackle the challenge in GRES due to the inflexible [SEG] token, which is addressed by our proposed GSVA." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 464, + 545, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 464, + 545, + 572 + ], + "spans": [ + { + "bbox": [ + 304, + 464, + 545, + 572 + ], + "type": "text", + "content": "Dynamic Networks. Dynamic neural networks [19] can adapt their architectures [20-22, 24, 27, 67-70] or parameters [11, 12, 53] to different inputs or switch the computation architecture in adjustment to different different time steps [46, 47, 71, 77], in order to achieve better accuracy and efficiency. In GSVA, the weight-sharing [SEG] tokens adapt to multiple targets under the hint of the prepended target prompts and dynamically reject empty targets with an individual prediction of rejection tokens." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 585, + 541, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 585, + 541, + 598 + ], + "spans": [ + { + "bbox": [ + 305, + 585, + 541, + 598 + ], + "type": "text", + "content": "3. Generalized Segmentation Vision Assistant" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "In this section, we initiate with the introduction of the model design of Generalized Segmentation Vision Assistant (GSVA), which is followed by the analysis of some certain limitations of LISA in Generalized Referring Segmentation (GRES). Subsequently, we delve into the introduction of two pivotal elements of GSVA, segmenting multiple targets and learning the rejection token. These components are fundamental in the conceptualization and construction of GSVA." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3859" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 71, + 545, + 277 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 545, + 277 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 545, + 277 + ], + "type": "image", + "image_path": "3e61033a07e86c5ac476931d0362596077bece25caa38a012ce7bc873e60cd01.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 285, + 546, + 331 + ], + "lines": [ + { + "bbox": [ + 46, + 285, + 546, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 546, + 331 + ], + "type": "text", + "content": "Figure 2. Overview of GSVA. At the bottom of the figure, the MLLM encodes the input image and concatenates the tokenized text tokens to follow instructions. GSVA generates multiple [SEG] tokens to handle multiple referred targets and rejects the objects absent in the image through [REJ] tokens. At the top of the figure, the SFM also encodes the image for segmentation and selects all [SEG] tokens in the output sequence to prompt the mask decoder to segment the target objects referred to in the instructions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 336, + 162, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 336, + 162, + 348 + ], + "spans": [ + { + "bbox": [ + 47, + 336, + 162, + 348 + ], + "type": "text", + "content": "3.1. Model Architecture" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 355, + 287, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 355, + 287, + 486 + ], + "spans": [ + { + "bbox": [ + 46, + 355, + 287, + 486 + ], + "type": "text", + "content": "The architecture of GSVA is illustrated in Figure 2, resembling LISA [32], which enables high-fidelity segmentation outputs by integrating two types of foundation models: (1) Multimodal Large Language Model (MLLM) as an aligned vision-language cognitive module; (2) Segmentation Foundation Model (SFM) to segment the target out of the input image based on user's instruction. To connect these two modules, LISA proposes a paradigm named embedding as mask where an extra [SEG] token is appended to the vocabularies of the MLLM and serves as the prompt of the SFM to segment the target following the intention of the user." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 488, + 287, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 488, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 46, + 488, + 287, + 620 + ], + "type": "text", + "content": "Multimodal Large Language Model. The MLLM consists of a decoder-based language model " + }, + { + "bbox": [ + 46, + 488, + 287, + 620 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{LLM}}" + }, + { + "bbox": [ + 46, + 488, + 287, + 620 + ], + "type": "text", + "content": " to auto-regressively generate text responses following the user's inputs, a vision encoder " + }, + { + "bbox": [ + 46, + 488, + 287, + 620 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{V1}}" + }, + { + "bbox": [ + 46, + 488, + 287, + 620 + ], + "type": "text", + "content": " to extract features from the input image, and a linear projector " + }, + { + "bbox": [ + 46, + 488, + 287, + 620 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 488, + 287, + 620 + ], + "type": "text", + "content": " to align the representations between image and text modalities. Specifically, the pretrained LLaVA [39] variants with CLIP-ViT-L/14 [55] and Vicuna-7B/13B [6] are employed. Given an input image " + }, + { + "bbox": [ + 46, + 488, + 287, + 620 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{img}}" + }, + { + "bbox": [ + 46, + 488, + 287, + 620 + ], + "type": "text", + "content": ", the vision encoder " + }, + { + "bbox": [ + 46, + 488, + 287, + 620 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{V1}}" + }, + { + "bbox": [ + 46, + 488, + 287, + 620 + ], + "type": "text", + "content": " first encodes it into image features, and then the projector " + }, + { + "bbox": [ + 46, + 488, + 287, + 620 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 488, + 287, + 620 + ], + "type": "text", + "content": " maps the features into the visual token embeddings in the LLM input space:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 121, + 624, + 287, + 638 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 624, + 287, + 638 + ], + "spans": [ + { + "bbox": [ + 121, + 624, + 287, + 638 + ], + "type": "interline_equation", + "content": "\\mathbf {h} _ {\\mathrm {i m g}} = \\phi \\left(F _ {\\mathrm {V} 1} \\left(\\mathbf {x} _ {\\mathrm {i m g}}\\right)\\right), \\tag {1}", + "image_path": "773c4514d57dcdef7365d370d5790c1aa390c409c0379d833fc01b5ec7ff5e3b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 642, + 288, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 642, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 45, + 642, + 288, + 712 + ], + "type": "text", + "content": "where the input image " + }, + { + "bbox": [ + 45, + 642, + 288, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{img}}" + }, + { + "bbox": [ + 45, + 642, + 288, + 712 + ], + "type": "text", + "content": " is typically resized to " + }, + { + "bbox": [ + 45, + 642, + 288, + 712 + ], + "type": "inline_equation", + "content": "h\\times w\\times 3" + }, + { + "bbox": [ + 45, + 642, + 288, + 712 + ], + "type": "text", + "content": " and the image tokens " + }, + { + "bbox": [ + 45, + 642, + 288, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_{\\mathrm{img}}\\in \\mathbb{R}^{n_{\\mathrm{img}}\\times d}" + }, + { + "bbox": [ + 45, + 642, + 288, + 712 + ], + "type": "text", + "content": " is aligned with the language modality. For CLIP-ViT-L/14, the input image with " + }, + { + "bbox": [ + 45, + 642, + 288, + 712 + ], + "type": "inline_equation", + "content": "h = w = 224" + }, + { + "bbox": [ + 45, + 642, + 288, + 712 + ], + "type": "text", + "content": " is encoded with ViT of patch size in 14, therefore the length of tokens " + }, + { + "bbox": [ + 45, + 642, + 288, + 712 + ], + "type": "inline_equation", + "content": "n_{\\mathrm{img}} = hw / 14^2 = 256" + }, + { + "bbox": [ + 45, + 642, + 288, + 712 + ], + "type": "text", + "content": " and the LLM dimensions " + }, + { + "bbox": [ + 45, + 642, + 288, + 712 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 45, + 642, + 288, + 712 + ], + "type": "text", + "content": " are 4096 and 5120 for Vicuna" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 337, + 546, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 337, + 546, + 372 + ], + "spans": [ + { + "bbox": [ + 305, + 337, + 546, + 372 + ], + "type": "text", + "content": "7B/13B, respectively. Along with the input image, the text instructions describing the targets to segment are tokenized into text tokens by the LLM tokenizer " + }, + { + "bbox": [ + 305, + 337, + 546, + 372 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 305, + 337, + 546, + 372 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 394, + 375, + 545, + 388 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 375, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 394, + 375, + 545, + 388 + ], + "type": "interline_equation", + "content": "\\mathbf {h} _ {\\mathrm {t x t}} = T (\\mathbf {x} _ {\\mathrm {t x t}}). \\tag {2}", + "image_path": "810b1bad85992c0010f47cb8d40fd5526fe8041443f059e4bcda6e103da0bb07.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 391, + 545, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 391, + 545, + 450 + ], + "spans": [ + { + "bbox": [ + 304, + 391, + 545, + 450 + ], + "type": "text", + "content": "The image tokens and text tokens are concatenated together and then fed into the LLM after prepending a sequence of fixed prompt tokens " + }, + { + "bbox": [ + 304, + 391, + 545, + 450 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_{\\mathrm{prompt}}" + }, + { + "bbox": [ + 304, + 391, + 545, + 450 + ], + "type": "text", + "content": " (omitted in the figure). The output token embeddings " + }, + { + "bbox": [ + 304, + 391, + 545, + 450 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{y}}_{\\mathrm{txt}}" + }, + { + "bbox": [ + 304, + 391, + 545, + 450 + ], + "type": "text", + "content": " are generated autoregressively:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 354, + 456, + 545, + 469 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 456, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 354, + 456, + 545, + 469 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {y}} _ {\\mathrm {t x t}} = F _ {\\mathrm {L L M}} \\left(\\left[ \\mathbf {h} _ {\\text {p r o m p t}} \\right| \\left| \\mathbf {h} _ {\\text {i m g}} \\right| \\left| \\mathbf {h} _ {\\text {t x t}} \\right|\\right), \\tag {3}", + "image_path": "7c8d528e79a05fe3e8737a5a3befab215a0703fdf443fbee0ea6659aa9fe0c0e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 472, + 545, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 545, + 506 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 472, + 545, + 506 + ], + "type": "inline_equation", + "content": "||" + }, + { + "bbox": [ + 304, + 472, + 545, + 506 + ], + "type": "text", + "content": " is concatenation operation. The text responses are obtained from " + }, + { + "bbox": [ + 304, + 472, + 545, + 506 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{y}}_{\\mathrm{txt}}" + }, + { + "bbox": [ + 304, + 472, + 545, + 506 + ], + "type": "text", + "content": " by applying a linear classifier to predict the next words in the vocabulary." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 508, + 545, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 508, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 508, + 545, + 590 + ], + "type": "text", + "content": "In LISA, a special token [SEG] is appended in the vocabulary to activate the segmentation ability of MLLM. The model learns to predict [SEG] token in the output sequence to indicate there is a target to segment. LISA then selects the output embedding of the [SEG] token " + }, + { + "bbox": [ + 304, + 508, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{y}}_{\\mathrm{xt}}[\\mathrm{SEG}]" + }, + { + "bbox": [ + 304, + 508, + 545, + 590 + ], + "type": "text", + "content": ", and projects it into the prompt space of the SFM by an MLP projector " + }, + { + "bbox": [ + 304, + 508, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 304, + 508, + 545, + 590 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 380, + 590, + 545, + 603 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 590, + 545, + 603 + ], + "spans": [ + { + "bbox": [ + 380, + 590, + 545, + 603 + ], + "type": "interline_equation", + "content": "\\mathbf {h} _ {\\mathrm {s e g}} = \\psi (\\tilde {\\mathcal {Y}} _ {\\mathrm {t x t}} [ \\mathrm {S E G} ]). \\tag {4}", + "image_path": "b494b90386f4d674d8aff7ff8a92ac2ad47ddd3a133f780348aff58a50f2b0e6.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 605, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 629 + ], + "type": "text", + "content": "The segmentation model is hence ready to decode the target mask from the query token " + }, + { + "bbox": [ + 304, + 605, + 545, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_{\\mathrm{seg}}" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Segmentation Foundation Model. The SFM is a query-based segmentation model, where a frozen vision encoder " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{V2}}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " takes in images of higher resolution than the vision encoder in MLLM to keep more details, followed by a trainable mask decoder " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{mask}}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " to decode masks from the queries. The pretrained SAM [31] with ViT-H backbone is instantiated as the SFM to produce high-quality masks. The given" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3860" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "input image " + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{img}}" + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": " is resized to a larger resolution " + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "H\\times W\\times 3" + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "H = W = 1024" + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": " and encoded with " + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{V2}}" + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": " to extract features " + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathrm{seg}}" + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": " for segmentation:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 130, + 110, + 287, + 124 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 110, + 287, + 124 + ], + "spans": [ + { + "bbox": [ + 130, + 110, + 287, + 124 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {\\text {s e g}} = F _ {\\mathrm {V} 2} \\left(\\mathbf {x} _ {\\text {i m g}}\\right). \\tag {5}", + "image_path": "6b65c5d2fc1310dc98b1b6e5de2da54f0b25ce5ea9baabb25dce8f598381a00b.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 127, + 287, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 127, + 287, + 164 + ], + "spans": [ + { + "bbox": [ + 47, + 127, + 287, + 164 + ], + "type": "text", + "content": "Condition on the features " + }, + { + "bbox": [ + 47, + 127, + 287, + 164 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathrm{seg}} \\in \\mathbb{R}^{\\frac{H}{16} \\times \\frac{W}{16} \\times C}" + }, + { + "bbox": [ + 47, + 127, + 287, + 164 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 47, + 127, + 287, + 164 + ], + "type": "inline_equation", + "content": "C = 256" + }, + { + "bbox": [ + 47, + 127, + 287, + 164 + ], + "type": "text", + "content": " for SAM, the mask decoder " + }, + { + "bbox": [ + 47, + 127, + 287, + 164 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{mask}}" + }, + { + "bbox": [ + 47, + 127, + 287, + 164 + ], + "type": "text", + "content": " decodes the segmentation masks from the segmentation queries " + }, + { + "bbox": [ + 47, + 127, + 287, + 164 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_{\\mathrm{seg}} \\in \\mathbb{R}^{N_{\\mathrm{seg}} \\times C}" + }, + { + "bbox": [ + 47, + 127, + 287, + 164 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 167, + 287, + 180 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 167, + 287, + 180 + ], + "spans": [ + { + "bbox": [ + 116, + 167, + 287, + 180 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {y}} _ {\\text {m a s k}} = F _ {\\text {m a s k}} \\left(\\mathbf {h} _ {\\text {s e g}} | \\mathbf {f} _ {\\text {s e g}}\\right), \\tag {6}", + "image_path": "c7b646cbfc4ac827c8c2c5c953147219fb60d650cf93616cf8003aa1300d084c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 183, + 286, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 183, + 286, + 206 + ], + "spans": [ + { + "bbox": [ + 47, + 183, + 286, + 206 + ], + "type": "text", + "content": "where each query in " + }, + { + "bbox": [ + 47, + 183, + 286, + 206 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_{\\mathrm{seg}}" + }, + { + "bbox": [ + 47, + 183, + 286, + 206 + ], + "type": "text", + "content": " corresponds to one segmentation mask in " + }, + { + "bbox": [ + 47, + 183, + 286, + 206 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{y}}_{\\mathrm{mask}}" + }, + { + "bbox": [ + 47, + 183, + 286, + 206 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 208, + 287, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 208, + 287, + 363 + ], + "spans": [ + { + "bbox": [ + 46, + 208, + 287, + 363 + ], + "type": "text", + "content": "LISA assumes that only one target exists to segment in the input image and its corresponding instructions. However, in GSVA, we extend it to a new scenario with multiple targets and empty targets, including multiple [SEG] tokens to invoke segmentation and [REJ] tokens to reject unreasonable instructed targets absent in the image. As shown in Figure 2, GSVA supports multiple [SEG]/[REJ] tokens in the output sequence and selects all the [SEG] tokens and discards every [REJ] token after attaining " + }, + { + "bbox": [ + 46, + 208, + 287, + 363 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{y}}_{\\mathrm{txt}}" + }, + { + "bbox": [ + 46, + 208, + 287, + 363 + ], + "type": "text", + "content": " in Eq. (3). Therefore, there is more than one query in " + }, + { + "bbox": [ + 46, + 208, + 287, + 363 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_{\\mathrm{seg}}" + }, + { + "bbox": [ + 46, + 208, + 287, + 363 + ], + "type": "text", + "content": ", thus enabling the SFM to segment multiple targets. These designs make GSVA competent in the GRES task, where the awareness of multiple and empty targets is of vital importance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 371, + 203, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 371, + 203, + 384 + ], + "spans": [ + { + "bbox": [ + 47, + 371, + 203, + 384 + ], + "type": "text", + "content": "3.2. GRES: Task and Challenges" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 389, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 389, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 287, + 593 + ], + "type": "text", + "content": "Task. Generalized Referring Expression Segmentation (GRES) [38] removes the constraint on the number of referred targets in the expression in the conventional Referring Expression Segmentation (RES) [30, 44, 84]. Different from that one expression only refers to one instance or region in RES, GRES allows arbitrary numbers of referred targets, including multiple instances or no target circumstances. In GRES, the user can refer to many instances simultaneously or include the objects that do not exist in the image. For instance, there are three referring expressions in Figure 2, including a plate of broccoli and a spoon in the bowl of sauce, vase bottom left corner, and soup in bowl right and bowl bottom left corner. In the GRES case, the model ought to segment the masks of the objects referred to in the " + }, + { + "bbox": [ + 46, + 389, + 287, + 593 + ], + "type": "inline_equation", + "content": "1^{\\text{st}}" + }, + { + "bbox": [ + 46, + 389, + 287, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 389, + 287, + 593 + ], + "type": "inline_equation", + "content": "3^{\\text{rd}}" + }, + { + "bbox": [ + 46, + 389, + 287, + 593 + ], + "type": "text", + "content": " expression, meanwhile producing an empty mask for the " + }, + { + "bbox": [ + 46, + 389, + 287, + 593 + ], + "type": "inline_equation", + "content": "2^{\\text{nd}}" + }, + { + "bbox": [ + 46, + 389, + 287, + 593 + ], + "type": "text", + "content": " expressions since there are no vases present at the bottom left corner." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": "Challenges. The challenges in GRES are common in practice, especially in embodied AI [17, 54, 58, 66]. The one challenge is multiple targets. Take robot navigation and planning [4, 45, 66] as an example, a robot may be asked to perceive multiple targets in the surrounding environment, e.g., to bring the two bowls of soup in Figure 2. The vision system of the robot needs to locate and segment the containers holding the referred food one at a time. The other challenge is empty target. Suppose the robot is ordered to cut an apple with a knife in the scene of Figure 2, whereas" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 143 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 143 + ], + "type": "text", + "content": "no apple is in the view of the camera, the vision system has to identify that the referred object is not in the scene. If it relies on some conventional RES methods which assume the expression must match something in the image, the output of the vision system could be undefined and potentially dangerous in some real-world cases." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 144, + 545, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 144, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 304, + 144, + 545, + 324 + ], + "type": "text", + "content": "Differences from ReasonSeg. Reasoning Segmentation (ReasonSeg) proposed by LISA [32] emphasizes the complex text instructions in RES. In ReasonSeg, the instructions are more implicit and sophisticated, forcing models to reason using world knowledge. Besides, the logic chain is usually longer and more challenging in ReasonSeg, which requires the model to deduce the final target object referred to in the image. In contrast, GRES increases complexity in another dimension by involving complicated spatial relationships. Hence, the model has to learn to handle this spatial information and understand the relationships between the instances. To meet these requirements, LISA tunes MLLMs with complex instructions paired with masks, while GSVA arouses the spatial modeling capabilities of MLLMs by learning multiple targets and rejecting empty targets." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 330, + 533, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 330, + 533, + 342 + ], + "spans": [ + { + "bbox": [ + 305, + 330, + 533, + 342 + ], + "type": "text", + "content": "3.3. Multiple [SEG] Tokens for Multiple Targets" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 348, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 348, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 348, + 545, + 396 + ], + "type": "text", + "content": "Single [SEG] token. LISA [32] follows the classic RES methods to generate a segmentation mask under given instructions by adding the [SEG] token into the answer. The prompt is formatted as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 319, + 398, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 398, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 319, + 398, + 545, + 422 + ], + "type": "text", + "content": "User: What is {obj} in this image? Please output segmentation mask. Assistant: Sure, it is [SEG]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 423, + 545, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 545, + 530 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 545, + 530 + ], + "type": "text", + "content": "In the above prompt, " + }, + { + "bbox": [ + 304, + 423, + 545, + 530 + ], + "type": "inline_equation", + "content": "\\{obj\\}" + }, + { + "bbox": [ + 304, + 423, + 545, + 530 + ], + "type": "text", + "content": " represents an instance referred to or some semantic area to segment, and the embeddings of the [SEG] token output from the LLM are projected to prompt the SFM. When multiple instances are requested simultaneously, this prompt would confuse the model since only one [SEG] token is forced to match several targets. As shown in Figure 1 (a), LISA coercively predicts the masks of the left man and the guy second from the right, leading to a tattered mask and masking the wrong location." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 531, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 531, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 545, + 639 + ], + "type": "text", + "content": "Multiple [SEG] tokens. To mitigate this issue in GRES, we relax this constraint for GSVA to support multiple target outputs via learning multiple [SEG] tokens. To avoid the ambiguity between the [SEG] tokens and the corresponding objects, we prepend the referring expression before each [SEG] token, i.e., each segmentation prompt in the output text is " + }, + { + "bbox": [ + 304, + 531, + 545, + 639 + ], + "type": "inline_equation", + "content": "\\{obj\\_n\\} :[SEG]" + }, + { + "bbox": [ + 304, + 531, + 545, + 639 + ], + "type": "text", + "content": ", as shown in Figure 3 (a). Streaming all the pairs of object and [SEG] token into one sentence, the question-answer prompt is formatted as (e.g., two targets):" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 319, + 640, + 545, + 675 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 319, + 640, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 640, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 319, + 640, + 545, + 662 + ], + "type": "text", + "content": "User: What are {obj1}, {obj2} in this image? Please output the segmentation masks." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 319, + 664, + 491, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 664, + 491, + 675 + ], + "spans": [ + { + "bbox": [ + 319, + 664, + 491, + 675 + ], + "type": "text", + "content": "Assistant: Sure, " + }, + { + "bbox": [ + 319, + 664, + 491, + 675 + ], + "type": "inline_equation", + "content": "\\{obj1\\} :[SEG]" + }, + { + "bbox": [ + 319, + 664, + 491, + 675 + ], + "type": "inline_equation", + "content": "\\{obj2\\} :[SEG]" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "This prompt requests the MLLM to identify and distinguish different objects in the image based on the instructions and infuse the corresponding location information" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3861" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 84, + 165, + 178 + ], + "blocks": [ + { + "bbox": [ + 73, + 72, + 139, + 81 + ], + "lines": [ + { + "bbox": [ + 73, + 72, + 139, + 81 + ], + "spans": [ + { + "bbox": [ + 73, + 72, + 139, + 81 + ], + "type": "text", + "content": "(a) Multiple Targets" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 84, + 165, + 178 + ], + "lines": [ + { + "bbox": [ + 49, + 84, + 165, + 178 + ], + "spans": [ + { + "bbox": [ + 49, + 84, + 165, + 178 + ], + "type": "image", + "image_path": "c06c0cf7c3251335e0eb02ab8135f9741a1798f23205e9fe1cd6f93724455a27.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 278, + 287, + 345 + ], + "lines": [ + { + "bbox": [ + 47, + 278, + 287, + 345 + ], + "spans": [ + { + "bbox": [ + 47, + 278, + 287, + 345 + ], + "type": "text", + "content": "Figure 3. Example of the prompts and predicted masks of GSVA-Vicuna-7B drawn from gRefCOCO validation set. (a) depicts the multiple-target case, in which two zebras referred to are handled with two separate [SEG] tokens. (b) shows the empty-target case, where no apple is in the bowl. Thus, the null referent is rejected with a [REJ] token, and no segmentation mask will be generated." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 50, + 186, + 70, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 186, + 70, + 194 + ], + "spans": [ + { + "bbox": [ + 50, + 186, + 70, + 194 + ], + "type": "text", + "content": "User:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 50, + 195, + 160, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 195, + 160, + 224 + ], + "spans": [ + { + "bbox": [ + 50, + 195, + 160, + 224 + ], + "type": "text", + "content": "What are front zebra, right zebra in this image? Please output segmentation masks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 50, + 236, + 85, + 245 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 236, + 85, + 245 + ], + "spans": [ + { + "bbox": [ + 50, + 236, + 85, + 245 + ], + "type": "text", + "content": "Assistant:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 50, + 246, + 151, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 246, + 151, + 265 + ], + "spans": [ + { + "bbox": [ + 50, + 246, + 151, + 265 + ], + "type": "text", + "content": "Sure, front zebra:[SEG], right zebra:[SEG]." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 169, + 84, + 286, + 178 + ], + "blocks": [ + { + "bbox": [ + 199, + 72, + 257, + 82 + ], + "lines": [ + { + "bbox": [ + 199, + 72, + 257, + 82 + ], + "spans": [ + { + "bbox": [ + 199, + 72, + 257, + 82 + ], + "type": "text", + "content": "(b) Empty Target" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 169, + 84, + 286, + 178 + ], + "lines": [ + { + "bbox": [ + 169, + 84, + 286, + 178 + ], + "spans": [ + { + "bbox": [ + 169, + 84, + 286, + 178 + ], + "type": "image", + "image_path": "a4c5d6ec3e13b820e727079189eb8de2c1e9b89b887bc43f7cffbbf63dc46067.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 171, + 185, + 192, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 185, + 192, + 194 + ], + "spans": [ + { + "bbox": [ + 171, + 185, + 192, + 194 + ], + "type": "text", + "content": "User:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 171, + 195, + 280, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 195, + 280, + 233 + ], + "spans": [ + { + "bbox": [ + 171, + 195, + 280, + 233 + ], + "type": "text", + "content": "What are the apple in the bowl, partial orange bottom left in this image? Please output segmentation masks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 171, + 236, + 206, + 245 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 236, + 206, + 245 + ], + "spans": [ + { + "bbox": [ + 171, + 236, + 206, + 245 + ], + "type": "text", + "content": "Assistant:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 171, + 246, + 269, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 246, + 269, + 275 + ], + "spans": [ + { + "bbox": [ + 171, + 246, + 269, + 275 + ], + "type": "text", + "content": "Sure, the apple in the bowl: [REJ], partial orange bottom left:[SEG]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 347, + 287, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 347, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 46, + 347, + 287, + 432 + ], + "type": "text", + "content": "of each target into the associated [SEG] token. We regard this ability as an implicit multimodal version of In-Context Learning (ICL), which is demonstrated by many prior works [2, 13, 33, 59, 90]. GSVA takes the target description preceding each [SEG] token as a hint to link this token to the object requested in the image through auto-regressive decoding." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 438, + 269, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 438, + 269, + 452 + ], + "spans": [ + { + "bbox": [ + 47, + 438, + 269, + 452 + ], + "type": "text", + "content": "3.4. Rejecting Empty Targets via [REJ] Tokens" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 457, + 287, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 457, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 46, + 457, + 287, + 553 + ], + "type": "text", + "content": "Empty targets. In GRES, many expressions match no targets in the image, including absence, incorrect attributes, inaccurate locations, etc. These expressions should be treated as empty targets for models to predict all-negative masks. LISA [32] falls short of predicting masks with all-zeros seamlessly since the [SEG] token always calls for a segmentation mask. As shown in Figure 1 (b), LISA incorrectly marks a piece of green apple as the empty target banana." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 553, + 287, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 553, + 287, + 661 + ], + "spans": [ + { + "bbox": [ + 46, + 553, + 287, + 661 + ], + "type": "text", + "content": "[REJ] token. We let the MLLM of GSVA predict a [REJ] token for each object that does not exist in the image but is referred to in the instructions, as shown in Figure 3 (b). GSVA predicts the targets marked with [REJ] tokens as empty targets, therefore setting all-zero masks for them. The involvement of [REJ] tokens directly rejects the empty target, liberating the mask decoder of the SFM from learning to segment the inconsistent targets. An example prompt with one existing target and one empty target is as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 60, + 663, + 287, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 663, + 287, + 686 + ], + "spans": [ + { + "bbox": [ + 60, + 663, + 287, + 686 + ], + "type": "text", + "content": "User: What are {obj1} (absent), {obj2} (absent), {obj3} in this image? Please output the segmentation masks." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 60, + 687, + 286, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 687, + 286, + 699 + ], + "spans": [ + { + "bbox": [ + 60, + 687, + 286, + 699 + ], + "type": "text", + "content": "Assistant: Sure, " + }, + { + "bbox": [ + 60, + 687, + 286, + 699 + ], + "type": "inline_equation", + "content": "\\{obj1\\} :[REJ],\\{obj2\\} :[REJ],\\{obj3\\} :[SEG]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "type": "text", + "content": "The [REJ] token prediction can also be seen as a variant" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": "of VQA task, where the specified object and its position in the image need to be considered. Thanks to the unprecedented capabilities of MLLM in understanding the images [9, 34, 39, 92] and reasoning the spatial relationships of the referring objects [52, 88, 91], we make the most of the MLLM in GSVA to unleash the burden of the segmentation model. The proposed empty-target-aware mechanism both improves the quality of masks and ameliorates the errors of identifying nonexistent objects." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 190, + 387, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 190, + 387, + 204 + ], + "spans": [ + { + "bbox": [ + 306, + 190, + 387, + 204 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 210, + 545, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 210, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 304, + 210, + 545, + 293 + ], + "type": "text", + "content": "In this section, we conduct comprehensive experiments to validate the efficacy of GSVA. First, we show the results on gRefCOCO [38] dataset to show the superiority of GSVA in GRES tasks. Then we verify GSVA that is also competent with other baselines in classic RES, REC tasks. We move on to ablate some important design choices of GSVA, followed by some qualitative visualization of GRES results." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 300, + 359, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 300, + 359, + 312 + ], + "spans": [ + { + "bbox": [ + 306, + 300, + 359, + 312 + ], + "type": "text", + "content": "4.1. GRES" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 319, + 545, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 545, + 642 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 545, + 642 + ], + "type": "text", + "content": "Settings. We adopt gRefCOCO [38] dataset to validate GSVA and LISA [32] on GRES, which contains 278,232 expressions, including 80,022 multi-target and 32,202 empty-target ones, referring to the objects in 19,994 images. The images are split into four subsets: training, validation, test-A, and test-B, following the same UNC partition of RefCOCO [84]. We first add the gRefCOCO training set into the mixed training dataset in LISA to pretrain GSVA and LISA for 50,000 steps and then finetune the models on the gRefCOCO training dataset for another 10 epochs. We evaluate the pretrained and finetuned models on the remaining validation set, test set A, and test set B, respectively. We adopt the IoU, cIoU metrics for the segmentation mask outputs. Following the implementation in Liu et al. [38], IoU averages the IoU for each mask, whereas cIoU computes the cumulative intersection area over the cumulative union area across the whole dataset. As for the empty target, we compute the No-target-accuracy (N-acc.), which is the ratio of the correctly classified empty-target expressions over all the empty-target expressions in the dataset. For a correctly classified empty target, the gIoU is set to 1.0, and the cIoU does not take them into account, while for a misclassified empty target, the gIoU is set to 0.0 and the union area is accumulated in the cIoU. Following the criteria in Liu et al. [38], a predicted mask is regarded as empty for LISA if the positive pixels are less than 50, whereas GSVA predicts [REJ] tokens to identify empty targets." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 642, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 546, + 713 + ], + "type": "text", + "content": "Results. We report the GRES segmentation results of GSVA and LISA [32] in Table 1. Three variants of GSVA, including GSVA-Vicuna-7B, GSVA-Vicuna-13B, and GSVA-Llama2-13B show competitive performance without finetuning on gIoU to the strongest non-LLM baseline ReLA [38]. However, LISA models fail to handle the" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3862" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 72, + 545, + 310 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 545, + 310 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 545, + 310 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 545, + 310 + ], + "type": "table", + "html": "
MethodGeneralized Referring Expression Segmentation on gRefCOCO
Validation SetTest Set ATest Set B
gIoUcIoUN-acc.gIoUcIoUN-acc.gIoUcIoUN-acc.
MattNet [85]48.2447.5141.1559.3058.6644.0446.1445.3341.32
LTS [29]52.7052.30-62.6461.87-50.4249.96-
VLT [10]52.0052.5147.1763.2062.1948.7450.8850.5247.82
CRIS [72]56.2755.34-63.4263.82-51.7951.04-
LAVT [81]58.4057.6449.3265.9065.3249.2555.8355.0448.46
ReLA [38]63.6062.4256.3770.0369.2659.0261.0259.8858.40
LISA-Vicuna-7B [32]32.2138.722.7148.5452.556.3739.6544.795.00
GSVA-Vicuna-7B63.3261.7056.4570.1169.2363.5061.3460.2658.42
LISA-Vicuna-7B [32] (ft)61.6361.7654.6766.2768.5050.0158.8460.6351.91
GSVA-Vicuna-7B (ft)66.4763.2962.4371.0869.9365.3162.2360.4760.56
LISA-Vicuna-13B [32]32.7339.853.6648.7653.624.8939.4945.354.41
GSVA-Vicuna-13B62.9760.1858.4467.1767.5954.6058.0657.2852.22
LISA-Vicuna-13B [32] (ft)63.4562.9955.2568.1869.6552.1661.8462.2456.15
GSVA-Vicuna-13B (ft)68.0164.0565.3671.7570.5167.2563.8361.2863.11
LISA-Llama2-13B [32]33.2639.643.2749.7653.807.2840.4945.415.73
GSVA-Llama2-13B63.2062.3854.5169.5269.8657.8462.0660.7758.30
LISA-Llama2-13B [32] (ft)65.2463.9657.4969.9971.0055.4362.1162.2956.34
GSVA-Llama2-13B (ft)70.0466.3866.0273.2972.7964.7265.4563.2062.47
", + "image_path": "6e3c0a02185afd8b429b419df3b4f3768bf1c96a108c180021e587f97c6e008a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 315, + 545, + 348 + ], + "lines": [ + { + "bbox": [ + 48, + 315, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 315, + 545, + 348 + ], + "type": "text", + "content": "Table 1. Generalized referring expression segmentation (GRES) results on gRefCOCO [38] dataset. gIoU and cIoU are IoU metrics averaged by each example and accumulated over whole dataset, respectively. N-acc. is short for the accuracy of correctly classifying null targets. Baselines are copied from Liu et al. [38]. (ft) denotes the model is finetuned on the training set of gRefCOCO." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 351, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 351, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 351, + 287, + 495 + ], + "type": "text", + "content": "GRES task without finetuning, showing degradation in both gIoU and cIoU in each model variant. Especially the low N-acc indicates that LISA is unable to correctly reject the empty targets. When finetuned on gRefCOCO training set, GSVA-Vicuna-7B performs better than the finetuned LISA counterpart, with about " + }, + { + "bbox": [ + 46, + 351, + 287, + 495 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 46, + 351, + 287, + 495 + ], + "type": "text", + "content": " improvement in gIoU and over " + }, + { + "bbox": [ + 46, + 351, + 287, + 495 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 46, + 351, + 287, + 495 + ], + "type": "text", + "content": " in N-acc on all three evaluation splits. GSVA variants with larger LLM incorporated further push the limits, achieving over " + }, + { + "bbox": [ + 46, + 351, + 287, + 495 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 46, + 351, + 287, + 495 + ], + "type": "text", + "content": " in gIoU on the validation set, " + }, + { + "bbox": [ + 46, + 351, + 287, + 495 + ], + "type": "inline_equation", + "content": "73\\%" + }, + { + "bbox": [ + 46, + 351, + 287, + 495 + ], + "type": "text", + "content": " on test set A, and " + }, + { + "bbox": [ + 46, + 351, + 287, + 495 + ], + "type": "inline_equation", + "content": "65\\%" + }, + { + "bbox": [ + 46, + 351, + 287, + 495 + ], + "type": "text", + "content": " on test set B. The 13B models also consistently outperform LISA by large margins, demonstrating the superiority of GSVA in GRES task." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 503, + 236, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 503, + 236, + 516 + ], + "spans": [ + { + "bbox": [ + 47, + 503, + 236, + 516 + ], + "type": "text", + "content": "4.2. Referring Expression Segmentation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 521, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 640 + ], + "type": "text", + "content": "Settings. To validate the abilities to handle various tasks, we evaluate GSVA in the classic RES task. Following the common evaluation protocols, we test variants of GSVA and LISA equipped with different LLMs on RefCOCO, RefCOCO+ [30], and RefCOCOg [44]. We follow the UNC split to perform experiments on RefCOCO and RefCOCO+, and UMD split for RefCOCOg. The models are firstly pretrained as in GRES, and then finetuned for 10 epochs with a joint dataset of these three RES training set. The cIoU metric is adopted to measure the model performances." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": "Results. Table 2 shows the RES results of GSVA. For the pretrained models, all the three variants of GSVA achieve higher cIoU than LISA [32] by clear margins. Our 7B model outperforms LISA-Vicuna about " + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": " cIoU on almost every data split. After finetuning, the preponderance of GSVA over LISA keeps and even enlarges. Specifically," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 351, + 545, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 351, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 351, + 545, + 399 + ], + "type": "text", + "content": "the cIoU metrics of GSVA-Llama2-13B on 8 sets surpass LISA by at least " + }, + { + "bbox": [ + 304, + 351, + 545, + 399 + ], + "type": "inline_equation", + "content": "2.8\\%" + }, + { + "bbox": [ + 304, + 351, + 545, + 399 + ], + "type": "text", + "content": ". For the test set of RefCOCOg, the margin even grows to " + }, + { + "bbox": [ + 304, + 351, + 545, + 399 + ], + "type": "inline_equation", + "content": "5.9\\%" + }, + { + "bbox": [ + 304, + 351, + 545, + 399 + ], + "type": "text", + "content": ", which exhibits GSVA is also competitive in the classic RES task." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 422, + 504, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 422, + 504, + 435 + ], + "spans": [ + { + "bbox": [ + 305, + 422, + 504, + 435 + ], + "type": "text", + "content": "4.3. Referring Expression Comprehension" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 445, + 545, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 545, + 540 + ], + "type": "text", + "content": "Settings. Since GSVA is capable for RES tasks, it is natural to transfer to Referring Expression Comprehension (REC) tasks, by simply computing the bounding boxes of the masks. The datasets of REC are the same as RES, in which we evaluate GSVA. If the IoU of a predicted bounding box and the ground truth is greater than 0.5, this prediction is marked as correct. We use the same models in RES to evaluate the phrase grounding capability in REC tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": "Results. As shown in Table 3, we mainly compare our GSVA to LISA in the three variants. Without finetuning, GSVA-Vicuna-7B outperforms LISA with a large margin over " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": " in almost all evaluation sets. The similar trends also hold for the Vicuna-13B and Llama2-13B variants. GSVA also benefits a lot from finetuning, e.g., the 7B variant achieves consistently higher Prec@0.5 than uLLaVA-7B [80] with LoRA finetuning, which is another strong baseline that adopts the \"mask2bbox\" pipelines without direct bounding box supervision. The finetuned GSVA also shows competitive performances to the fully finetuned uLLaVA-7B, suggesting the strong potential of our method. With larger LLMs incorporated, the performances of GSVA continue with over " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": " increments over all datasets." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3863" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 71, + 545, + 324 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 545, + 324 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 545, + 324 + ], + "type": "table", + "html": "
Referring Expression Segmentation on RefCOCO, RefCOCO+, and RefCOCOg
MethodRefCOCO (UNC)RefCOCO+ (UNC)RefCOCOg (UMD)
Val.Test-ATest-BVal.Test-ATest-BVal.Test
MCN [42]62.464.259.750.655.044.749.249.4
VLT [10]67.570.565.256.361.050.155.057.7
CRIS [72]70.573.266.162.368.153.759.960.4
LAVT [81]72.775.868.862.168.455.161.262.1
ReLA [38]73.876.570.266.071.057.765.066.0
X-Decoder [93]------64.6-
SEEM [94]------65.7-
PolyFormer-L [40]76.078.373.369.374.661.969.270.2
LISA-Vicuna-7B* [32]74.176.571.162.467.456.566.468.5
GSVA-Vicuna-7B76.477.472.864.567.758.671.172.0
LISA-Vicuna-7B* [32] (ft)74.979.172.365.170.858.167.970.6
GSVA-Vicuna-7B (ft)77.278.973.565.969.659.872.773.3
LISA-Vicuna-13B [32]71.774.768.159.464.252.965.266.1
GSVA-Vicuna-13B74.677.570.562.566.555.569.671.2
LISA-Vicuna-13B [32] (ft)76.078.872.965.070.258.169.570.5
GSVA-Vicuna-13B (ft)78.280.474.267.471.560.974.275.6
LISA-Llama2-13B [32]73.476.269.562.366.656.368.268.5
GSVA-Llama2-13B77.779.974.268.071.561.573.273.9
LISA-Llama2-13B [32] (ft)76.378.772.466.271.059.370.171.1
GSVA-Llama2-13B (ft)79.281.777.170.373.863.675.777.0
", + "image_path": "1cf1669ea7bb7e68411f85ea48c926479fb37e27ef50bc0fbc05deae57806d83.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 365, + 545, + 548 + ], + "blocks": [ + { + "bbox": [ + 47, + 329, + 546, + 363 + ], + "lines": [ + { + "bbox": [ + 47, + 329, + 546, + 363 + ], + "spans": [ + { + "bbox": [ + 47, + 329, + 546, + 363 + ], + "type": "text", + "content": "Table 2. Referring expression segmentation results on RefCOCO, RefCOCO+ [30] and RefCOCOg [44] dataset. The cIoU metrics of each split are reported. Baselines are excerpted from Lai et al. [32]. (ft) denotes the models are finetuned on the joint training set of the referring expression segmentation datasets. * means the results are excerpted from the original paper." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 365, + 545, + 548 + ], + "lines": [ + { + "bbox": [ + 50, + 365, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 50, + 365, + 545, + 548 + ], + "type": "table", + "html": "
MethodReferring Expression Comprehension on RefCOCO, RefCOCO+, and RefCOCOg
Val.RefCOCO Test-ATest-BVal.RefCOCO+ Test-ATest-BRefCOCOg Val.Test
u-LLaVA-7B [80] (LoRA)83.4787.1380.2168.7476.3260.9876.1978.24
u-LLaVA-7B [80] (full-ft)86.0489.4782.2674.0981.1666.6179.8781.68
LISA-Vicuna-7B [32]78.6881.7275.7462.9268.9356.4970.1072.47
GSVA-Vicuna-7B85.5088.0182.4970.2175.6265.1179.0079.21
LISA-Vicuna-7B [32] (ft)85.3988.8482.5974.2379.4668.4079.3480.42
GSVA-Vicuna-7B (ft)86.2789.2283.7772.8178.7868.0181.5881.83
LISA-Vicuna-13B [32]80.0183.2676.2663.7770.2457.4271.7973.34
GSVA-Vicuna-13B83.1287.0180.5468.1473.9062.0077.0878.89
LISA-Vicuna-13B [32] (ft)85.9289.0583.1674.8681.0868.8780.0981.48
GSVA-Vicuna-13B (ft)87.7190.4984.5776.5281.6970.3583.9084.85
LISA-Llama2-13B [32]82.5285.5678.8267.9173.7762.2575.3776.83
GSVA-Llama2-13B86.9989.5484.0873.8979.1069.3880.6882.07
LISA-Llama2-13B [32] (ft)85.9188.8481.7374.4680.5668.2680.0981.27
GSVA-Llama2-13B (ft)89.1692.0887.1779.7484.4573.4185.4786.18
", + "image_path": "1c6bb18d46211a074d3f1008f020e8265506f56316a3b7241ee220bfcdf77d6b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 552, + 547, + 586 + ], + "lines": [ + { + "bbox": [ + 47, + 552, + 547, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 547, + 586 + ], + "type": "text", + "content": "Table 3. Referring expression comprehension results on RefCOCO, RefCOCO+ [30] and RefCOCOg [44] dataset. The metric is precision @ 0.5 IoU threshold. (LoRA) means the LLM in u-LLaVA [80] is finetuned with LoRA adapter, as LISA and GSVA, while (full-ft) represents the LLM in u-LLaVA is fully trained. Results of u-LLaVA-7B with \"mask2bbox\" strategy are reported for fair comparison." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 591, + 141, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 591, + 141, + 604 + ], + "spans": [ + { + "bbox": [ + 47, + 591, + 141, + 604 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 617, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 617, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 617, + 288, + 714 + ], + "type": "text", + "content": "The involvement of the [REJ] token. [REJ] token plays a rather important role in GSVA. We study the effect of the [REJ] token by removing it from the vocabulary in GSVA, yielding a variant unable to reject a target from the text outputs. As shown in the " + }, + { + "bbox": [ + 47, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "2^{\\text{nd}}" + }, + { + "bbox": [ + 47, + 617, + 288, + 714 + ], + "type": "text", + "content": " row of Table 4, after removing [REJ] token, there is a sharp N-acc drop over " + }, + { + "bbox": [ + 47, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 47, + 617, + 288, + 714 + ], + "type": "text", + "content": " relatively, followed by the decline of gIoU at about " + }, + { + "bbox": [ + 47, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 47, + 617, + 288, + 714 + ], + "type": "text", + "content": " on gRefCOCO validation set. This performance degradation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 592, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 592, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 305, + 592, + 545, + 616 + ], + "type": "text", + "content": "indicates the significance of LLM learning a special token to reject the referred instances absent in the image." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "content": "Learning multiple [SEG] tokens. We continue to ablate the multiple [SEG] tokens, which is another core design of GSVA. After removing [REJ] token, we then reduce the number of [SEG] tokens to 1, which is identical to LISA [32] with the referring expression added before the only one [SEG] token: Assistant: Sure, {obj}:[SEG]. In the early experiments, we have found that stacking multiple expressions before one [SEG] token would result in diver" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3864" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 85, + 132, + 137 + ], + "blocks": [ + { + "bbox": [ + 136, + 72, + 219, + 82 + ], + "lines": [ + { + "bbox": [ + 136, + 72, + 219, + 82 + ], + "spans": [ + { + "bbox": [ + 136, + 72, + 219, + 82 + ], + "type": "text", + "content": "(a) Multiple Target Cases" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 85, + 132, + 137 + ], + "lines": [ + { + "bbox": [ + 49, + 85, + 132, + 137 + ], + "spans": [ + { + "bbox": [ + 49, + 85, + 132, + 137 + ], + "type": "image", + "image_path": "daf746cea54f9db87410b369a5fded0fe5eb165343920dc34d492be4a37d8a60.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 137, + 85, + 220, + 137 + ], + "blocks": [ + { + "bbox": [ + 137, + 85, + 220, + 137 + ], + "lines": [ + { + "bbox": [ + 137, + 85, + 220, + 137 + ], + "spans": [ + { + "bbox": [ + 137, + 85, + 220, + 137 + ], + "type": "image", + "image_path": "4ec7b27483a384f355e27e2cfe72aaa48709e263fe2ea74aea250884f6c9d306.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 225, + 85, + 292, + 137 + ], + "blocks": [ + { + "bbox": [ + 225, + 85, + 292, + 137 + ], + "lines": [ + { + "bbox": [ + 225, + 85, + 292, + 137 + ], + "spans": [ + { + "bbox": [ + 225, + 85, + 292, + 137 + ], + "type": "image", + "image_path": "26a9add807664ac57a40ce031e8cfa0cbfcb8a0cabd1c34520fc53fb4a2ca992.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 49, + 141, + 132, + 194 + ], + "blocks": [ + { + "bbox": [ + 49, + 141, + 132, + 194 + ], + "lines": [ + { + "bbox": [ + 49, + 141, + 132, + 194 + ], + "spans": [ + { + "bbox": [ + 49, + 141, + 132, + 194 + ], + "type": "image", + "image_path": "ed01be79d76617b81ca32630c63cc333515e0ff7504186c73807b006edfddbb7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 52, + 204, + 128, + 211 + ], + "lines": [ + { + "bbox": [ + 52, + 204, + 128, + 211 + ], + "spans": [ + { + "bbox": [ + 52, + 204, + 128, + 211 + ], + "type": "text", + "content": "Elephant in back, Left elephant" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 135, + 200, + 227, + 214 + ], + "lines": [ + { + "bbox": [ + 135, + 200, + 227, + 214 + ], + "spans": [ + { + "bbox": [ + 135, + 200, + 227, + 214 + ], + "type": "text", + "content": "Kid on left in blue, Boy on right, Blond boy on right, Girl back left pink" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 220, + 545, + 264 + ], + "lines": [ + { + "bbox": [ + 46, + 220, + 545, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 545, + 264 + ], + "type": "text", + "content": "Figure 4. Visualizations of GSVA and LISA [32] in the GRES task. The first row shows LISA's segmentation results, the second row is the masks and rejections of GSVA, and the third row shows the referring expressions in the instructions. In (a) multiple target cases, each target is colored with a specific color. In (b) empty target cases, the images turn darker to highlight the incorrect predictions of LISA. The examples are selected from the gRefCOCO validation set. The masks are generated by the 7B models. Zoom in for the best view." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 137, + 141, + 220, + 194 + ], + "blocks": [ + { + "bbox": [ + 137, + 141, + 220, + 194 + ], + "lines": [ + { + "bbox": [ + 137, + 141, + 220, + 194 + ], + "spans": [ + { + "bbox": [ + 137, + 141, + 220, + 194 + ], + "type": "image", + "image_path": "d172daaca66210060b87ca57f49113019b04b76b1f2a651ffd246bea7c3e41f9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 229, + 201, + 284, + 215 + ], + "lines": [ + { + "bbox": [ + 229, + 201, + 284, + 215 + ], + "spans": [ + { + "bbox": [ + 229, + 201, + 284, + 215 + ], + "type": "text", + "content": "Tan vase, Bottom vase, Right vase" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 225, + 141, + 292, + 194 + ], + "blocks": [ + { + "bbox": [ + 225, + 141, + 292, + 194 + ], + "lines": [ + { + "bbox": [ + 225, + 141, + 292, + 194 + ], + "spans": [ + { + "bbox": [ + 225, + 141, + 292, + 194 + ], + "type": "image", + "image_path": "ec875e834067d302b43ef1aa075dabff7c15cec5af9e6b2f3eaf5532e0bd984e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 296, + 85, + 380, + 137 + ], + "blocks": [ + { + "bbox": [ + 378, + 72, + 456, + 82 + ], + "lines": [ + { + "bbox": [ + 378, + 72, + 456, + 82 + ], + "spans": [ + { + "bbox": [ + 378, + 72, + 456, + 82 + ], + "type": "text", + "content": "(b) Empty Target Cases" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 296, + 85, + 380, + 137 + ], + "lines": [ + { + "bbox": [ + 296, + 85, + 380, + 137 + ], + "spans": [ + { + "bbox": [ + 296, + 85, + 380, + 137 + ], + "type": "image", + "image_path": "465f562d5740bd84758eb8a00204d29617174eaf3afd6cc7f9881ea051ad3854.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 297, + 141, + 380, + 194 + ], + "blocks": [ + { + "bbox": [ + 297, + 141, + 380, + 194 + ], + "lines": [ + { + "bbox": [ + 297, + 141, + 380, + 194 + ], + "spans": [ + { + "bbox": [ + 297, + 141, + 380, + 194 + ], + "type": "image", + "image_path": "0a39cdcf3e583ce5afec93306543ceeef6189aa388dc51808c9ceff98d38844a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 205, + 353, + 212 + ], + "lines": [ + { + "bbox": [ + 323, + 205, + 353, + 212 + ], + "spans": [ + { + "bbox": [ + 323, + 205, + 353, + 212 + ], + "type": "text", + "content": "Lady on left" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 384, + 85, + 469, + 137 + ], + "blocks": [ + { + "bbox": [ + 384, + 85, + 469, + 137 + ], + "lines": [ + { + "bbox": [ + 384, + 85, + 469, + 137 + ], + "spans": [ + { + "bbox": [ + 384, + 85, + 469, + 137 + ], + "type": "image", + "image_path": "11400d90a2e5f11667b43e3559b917794630bd1b2dcba7d6930afd4bdf80229d.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 385, + 141, + 469, + 194 + ], + "blocks": [ + { + "bbox": [ + 385, + 141, + 469, + 194 + ], + "lines": [ + { + "bbox": [ + 385, + 141, + 469, + 194 + ], + "spans": [ + { + "bbox": [ + 385, + 141, + 469, + 194 + ], + "type": "image", + "image_path": "638204a0c6903c24c88310a1b409bd3dd8a2eb86eb758446f56062d76e6b39c8.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 205, + 447, + 212 + ], + "lines": [ + { + "bbox": [ + 406, + 205, + 447, + 212 + ], + "spans": [ + { + "bbox": [ + 406, + 205, + 447, + 212 + ], + "type": "text", + "content": "Teddy with X eye" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 474, + 85, + 539, + 137 + ], + "blocks": [ + { + "bbox": [ + 474, + 85, + 539, + 137 + ], + "lines": [ + { + "bbox": [ + 474, + 85, + 539, + 137 + ], + "spans": [ + { + "bbox": [ + 474, + 85, + 539, + 137 + ], + "type": "image", + "image_path": "fa154cc8e320f6229a187f7766e439434cfccf48d12a0602cc780105dcb7b891.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 474, + 141, + 539, + 194 + ], + "blocks": [ + { + "bbox": [ + 474, + 141, + 539, + 194 + ], + "lines": [ + { + "bbox": [ + 474, + 141, + 539, + 194 + ], + "spans": [ + { + "bbox": [ + 474, + 141, + 539, + 194 + ], + "type": "image", + "image_path": "dbb082b818b53a9e2883bff956ccc92ac2f7ebb36a68ddbaf3d98fe4e3d0284b.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 479, + 205, + 535, + 212 + ], + "lines": [ + { + "bbox": [ + 479, + 205, + 535, + 212 + ], + "spans": [ + { + "bbox": [ + 479, + 205, + 535, + 212 + ], + "type": "text", + "content": "The dog on the suitcase" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "table", + "bbox": [ + 50, + 269, + 287, + 354 + ], + "blocks": [ + { + "bbox": [ + 50, + 269, + 287, + 354 + ], + "lines": [ + { + "bbox": [ + 50, + 269, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 50, + 269, + 287, + 354 + ], + "type": "table", + "html": "
Model w/Vicuna-7BModificationsgRefCOCO Val.
RefExp. +[SEG]Multiple [SEG][REJ] TokengIoUcIoUN-acc.
GSVA✓✓✓63.3261.7056.45
✓✓✗51.5760.9530.32
✓✗✗44.8659.3711.96
LISA [32]✗✗✗32.2138.722.71
✗✓✓21.8327.220.00
", + "image_path": "ead0c4ade1b9d61b92890db88b80ec3a6e2745d6c83c37111648289f5bce54da.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 358, + 287, + 414 + ], + "lines": [ + { + "bbox": [ + 46, + 358, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 46, + 358, + 287, + 414 + ], + "type": "text", + "content": "Table 4. Ablation study on the core designs of GSVA. " + }, + { + "bbox": [ + 46, + 358, + 287, + 414 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 46, + 358, + 287, + 414 + ], + "type": "text", + "content": " means the employment of the component while " + }, + { + "bbox": [ + 46, + 358, + 287, + 414 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 358, + 287, + 414 + ], + "type": "text", + "content": " means not. \"Ref-Exp.+[SEG]\", \"Multiple [SEG]\", and \"[REJ] Token\" are short for adding referring expression before [SEG] in the answer prompt, using multiple [SEG] tokens, involving [REJ] token, respectively." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 422, + 287, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 287, + 470 + ], + "type": "text", + "content": "gence. Therefore we separate multiple targets to prompt the model with one expression at a time. The sharp decrements of gIoU by nearly " + }, + { + "bbox": [ + 46, + 422, + 287, + 470 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 46, + 422, + 287, + 470 + ], + "type": "text", + "content": " and N-acc by over " + }, + { + "bbox": [ + 46, + 422, + 287, + 470 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 46, + 422, + 287, + 470 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 46, + 422, + 287, + 470 + ], + "type": "inline_equation", + "content": "3^{\\mathrm{rd}}" + }, + { + "bbox": [ + 46, + 422, + 287, + 470 + ], + "type": "text", + "content": " row demonstrate the significance of the multiple-[SEG]-token." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 46, + 470, + 287, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 470, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 470, + 287, + 662 + ], + "type": "text", + "content": "Answers without referring expression. To examine the efficacy of the hinting prompts, we remove all the referring expressions before the [SEG] tokens. Based on the removal of multiple [SEG]s and [REJ]s, erasing the added referring expression falls back to the original LISA model, as shown in the " + }, + { + "bbox": [ + 46, + 470, + 287, + 662 + ], + "type": "inline_equation", + "content": "4^{\\text{th}}" + }, + { + "bbox": [ + 46, + 470, + 287, + 662 + ], + "type": "text", + "content": " row. We further choose only to remove it from GSVA model, keeping other configurations unchanged. Specifically, if there are two referents, the prompts in the answer will turn to Assistant: Sure, [SEG], [SEG], whose results are in the last row. The zero N-acc shows the model fails to identify any empty target without the help of the expressions, meanwhile the poor gIoU and cIoU indicates the segmentation ability is damaged. This phenomenon also suggests that the added referring expression hint GSVA to associate each [SEG] token to its corresponding target, which is in coherence with our hypothesis." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 47, + 670, + 130, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 130, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 130, + 681 + ], + "type": "text", + "content": "4.5. Visualization" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "We visualize some qualitative results of GSVA to verify its effectiveness. As shown in Figure 4, we present two groups" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 270, + 545, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 270, + 545, + 427 + ], + "spans": [ + { + "bbox": [ + 304, + 270, + 545, + 427 + ], + "type": "text", + "content": "of examples from the validation set of gRefCOCO [38] to see how GSVA outperforms LISA in the face of the two main challenges in GRES: multiple targets and empty targets. In part (a), GSVA has managed to segment all the targets referred to, while LISA could only segment one of the requested instances. For example, in the third column, LISA only predicts the mask of the rightmost vase. On the contrary, GSVA separately segments all three targeted vases. In part (b), LISA mistakenly segments the instances in the image that disagree with the referring expression, e.g., in the sixth column, LISA proposes a mask of a cat in response to the request of the dog on the suitcase, whereas GSVA successfully reject all the empty targets." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 441, + 378, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 441, + 378, + 453 + ], + "spans": [ + { + "bbox": [ + 306, + 441, + 378, + 453 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 462, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 594 + ], + "type": "text", + "content": "This paper introduces a novel multimodal large language model dubbed Generalized Segmentation Vision Assistant (GSVA). By introducing multiple [SEG] tokens and the new [REJ] token, GSVA effectively achieves multi-objective segmentation and empty target rejection, which addresses the challenging segmentation problems in practical application scenarios, e.g., Generalized Referring Expression Segmentation (GRES). Extensive experiments on GRES, classic RES, and REC tasks fully demonstrate the superior performance of our method, highlighting its significance for future research and applications." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 306, + 609, + 403, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 609, + 403, + 621 + ], + "spans": [ + { + "bbox": [ + 306, + 609, + 403, + 621 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "This work is supported in part by the National Key R&D Program of China under Grant 2021ZD0140407, the National Natural Science Foundation of China under Grants 62321005 and 62276150. We also thank Dr. Yuan Yao and Prof. Zhiyuan Liu for their insightful and valuable comments on this research project. We also appreciate their generous support of computational resources." + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3865" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 144 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 144 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 146, + 288, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 146, + 288, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 146, + 288, + 191 + ], + "type": "text", + "content": "[2] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, 2020. 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 288, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 288, + 224 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 288, + 224 + ], + "type": "text", + "content": "[3] Ding-Jie Chen, Songhao Jia, Yi-Chen Lo, Hwann-Tzong Chen, and Tyng-Luh Liu. See-through-text grouping for referring image segmentation. In IEEE ICCV, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 224, + 288, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 288, + 267 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 288, + 267 + ], + "type": "text", + "content": "[4] Yixin Chen, Qing Li, Deqian Kong, Yik Lun Kei, Song-Chun Zhu, Tao Gao, Yixin Zhu, and Siyuan Huang. Yourefit: Embodied reference understanding with language and gesture. In IEEE ICCV, 2021. 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 269, + 288, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 269, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 269, + 288, + 312 + ], + "type": "text", + "content": "[5] Ming-Ming Cheng, Shuai Zheng, Wen-Yan Lin, Vibhav Vineet, Paul Sturgess, Nigel Crook, Niloy J Mitra, and Philip Torr. Imagesprit: Verbal guided image parsing. ACM ToG, 2014. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 313, + 288, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 313, + 288, + 369 + ], + "spans": [ + { + "bbox": [ + 53, + 313, + 288, + 369 + ], + "type": "text", + "content": "[6] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 53, + 313, + 288, + 369 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 53, + 313, + 288, + 369 + ], + "type": "text", + "content": " * chatgpt quality, 2023. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 369, + 287, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 287, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 287, + 401 + ], + "type": "text", + "content": "[7] Yong Xien Chng, Henry Zheng, Yizeng Han, Xuchong Qiu, and Gao Huang. Mask grounding for referring image segmentation. In IEEE CVPR, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 403, + 288, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 288, + 456 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 288, + 456 + ], + "type": "text", + "content": "[8] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 458, + 288, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 458, + 288, + 512 + ], + "spans": [ + { + "bbox": [ + 53, + 458, + 288, + 512 + ], + "type": "text", + "content": "[9] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. In NeurIPS, 2023. 2, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 514, + 286, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 286, + 546 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 286, + 546 + ], + "type": "text", + "content": "[10] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vision-language transformer and query generation for referring segmentation. In IEEE ICCV, 2021. 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 547, + 286, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 286, + 580 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 286, + 580 + ], + "type": "text", + "content": "[11] Xiaohan Ding, Xiangyu Zhang, Ningning Ma, Jungong Han, Guiguang Ding, and Jian Sun. Repvgg: Making vgg-style convnets great again. In IEEE ICCV, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 581, + 286, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 286, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 286, + 613 + ], + "type": "text", + "content": "[12] Xiaohan Ding, Xiangyu Zhang, Jungong Han, and Guiguang Ding. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In IEEE CVPR, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 614, + 286, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 286, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 286, + 657 + ], + "type": "text", + "content": "[13] Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Zhiyong Wu, Baobao Chang, Xu Sun, Jingjing Xu, and Zhifang Sui. A survey for in-context learning. arXiv preprint arXiv:2301.00234, 2022.5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 658, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 286, + 713 + ], + "type": "text", + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[15] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale. In IEEE CVPR, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "text", + "content": "[16] Guang Feng, Zhiwei Hu, Lihe Zhang, and Huchuan Lu. Encoder fusion network with co-attention embedding for referring image segmentation. In IEEE CVPR, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "text", + "content": "[17] Chen Gao, Jinyu Chen, Si Liu, Luting Wang, Qiong Zhang, and Qi Wu. Room-and-object aware knowledge reasoning for remote embodied referring expression. In IEEE CVPR, 2021. 1, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "type": "text", + "content": "[18] Dongchen Han, Tianzhu Ye, Yizeng Han, Zhuofan Xia, Shiji Song, and Gao Huang. Agent attention: On the integration of softmax and linear attention. arXiv preprint arXiv:2312.08874, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 241, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 274 + ], + "type": "text", + "content": "[19] Yizeng Han, Gao Huang, Shiji Song, Le Yang, Honghui Wang, and Yulin Wang. Dynamic neural networks: A survey. IEEE TPAMI, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 276, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 308 + ], + "type": "text", + "content": "[20] Yizeng Han, Gao Huang, Shiji Song, Le Yang, Yitian Zhang, and Haojun Jiang. Spatially adaptive feature refinement for efficient inference. IEEE TIP, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 309, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 353 + ], + "type": "text", + "content": "[21] Yizeng Han, Yifan Pu, Zihang Lai, Chaofei Wang, Shiji Song, Junfeng Cao, Wenhui Huang, Chao Deng, and Gao Huang. Learning to weight samples for dynamic early-exiting networks. In ECCV, 2022." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 354, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 387 + ], + "type": "text", + "content": "[22] Yizeng Han, Zhihang Yuan, Yifan Pu, Chenhao Xue, Shiji Song, Guangyu Sun, and Gao Huang. Latency-aware spatialwise dynamic networks. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 388, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 431 + ], + "type": "text", + "content": "[23] Yizeng Han, Dongchen Han, Zeyu Liu, Yulin Wang, Xuran Pan, Yifan Pu, Chao Deng, Junlan Feng, Shiji Song, and Gao Huang. Dynamic perceiver for efficient visual recognition. In IEEE ICCV, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 433, + 545, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 433, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 307, + 433, + 545, + 477 + ], + "type": "text", + "content": "[24] Yizeng Han, Zeyu Liu, Zhihang Yuan, Yifan Pu, Chaofei Wang, Shiji Song, and Gao Huang. Latency-aware unified dynamic networks for efficient image recognition. arXiv preprint arXiv:2308.15949, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 478, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 478, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 478, + 545, + 510 + ], + "type": "text", + "content": "[25] Ronghang Hu, Marcus Rohrbach, and Trevor Darrell. Segmentation from natural language expressions. In ECCV, 2016. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 511, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 511, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 511, + 545, + 555 + ], + "type": "text", + "content": "[26] Yutao Hu, Qixiong Wang, Wenqi Shao, Enze Xie, Zhenguo Li, Jungong Han, and Ping Luo. Beyond one-to-one: Rethinking the referring image segmentation. In IEEE ICCV, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 556, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 556, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 307, + 556, + 545, + 600 + ], + "type": "text", + "content": "[27] Gao Huang, Yulin Wang, Kangchen Lv, Haojun Jiang, Wenhui Huang, Pengfei Qi, and Shiji Song. Glance and focus networks for dynamic visual recognition. IEEE TPAMI, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 601, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 645 + ], + "type": "text", + "content": "[28] Rui Huang, Xuran Pan, Henry Zheng, Haojun Jiang, Zhifeng Xie, Cheng Wu, Shiji Song, and Gao Huang. Joint representation learning for text and 3d point cloud. Pattern Recognition, 2024. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 647, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 679 + ], + "type": "text", + "content": "[29] Ya Jing, Tao Kong, Wei Wang, Liang Wang, Lei Li, and Tie-niu Tan. Locate then segment: A strong pipeline for referring image segmentation. In IEEE CVPR, 2021. 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "text", + "content": "[30] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In EMNLP, 2014. 1, 4, 6, 7" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "3866" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[31] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In IEEE ICCV, 2023. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 117, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 117, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 117, + 287, + 160 + ], + "type": "text", + "content": "[32] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692, 2023. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 161, + 288, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 161, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 161, + 288, + 205 + ], + "type": "text", + "content": "[33] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Fanyi Pu, Jingkang Yang, Chunyuan Li, and Ziwei Liu. Mimicit: Multi-modal in-context instruction tuning. arXiv preprint arXiv:2306.05425, 2023. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 205, + 287, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 205, + 287, + 248 + ], + "spans": [ + { + "bbox": [ + 48, + 205, + 287, + 248 + ], + "type": "text", + "content": "[34] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 249, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 249, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 249, + 287, + 293 + ], + "type": "text", + "content": "[35] Ruiyu Li, Kaican Li, Yi-Chun Kuo, Michelle Shu, Xiaojuan Qi, Xiaoyong Shen, and Jiaya Jia. Referring image segmentation via recurrent refinement networks. In IEEE CVPR, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 293, + 287, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 293, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 293, + 287, + 348 + ], + "type": "text", + "content": "[36] Ziyi Lin, Chris Liu, Renrui Zhang, Peng Gao, Longtian Qiu, Han Xiao, Han Qiu, Chen Lin, Wenqi Shao, Keqin Chen, et al. Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models. arXiv preprint arXiv:2311.07575, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 349, + 287, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 349, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 349, + 287, + 381 + ], + "type": "text", + "content": "[37] Chenxi Liu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, and Alan Yuille. Recurrent multimodal interaction for referring image segmentation. In IEEE CVPR, 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 383, + 287, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 287, + 414 + ], + "type": "text", + "content": "[38] Chang Liu, Henghui Ding, and Xudong Jiang. Gres: Generalized referring expression segmentation. In IEEE CVPR, 2023. 1, 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 415, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 437 + ], + "type": "text", + "content": "[39] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 2, 3, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 437, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 287, + 480 + ], + "type": "text", + "content": "[40] Jiang Liu, Hui Ding, Zhaowei Cai, Yuting Zhang, Ravi Kumar Satzoda, Vijay Mahadevan, and R Manmatha. Polyformer: Referring image segmentation as sequential polygon generation. In IEEE CVPR, 2023. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 482, + 287, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 482, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 48, + 482, + 287, + 535 + ], + "type": "text", + "content": "[41] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 536, + 287, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 580 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 580 + ], + "type": "text", + "content": "[42] Gen Luo, Yiyi Zhou, Xiaoshuai Sun, Liujuan Cao, Chenglin Wu, Cheng Deng, and Rongrong Ji. Multi-task collaborative network for joint referring expression comprehension and segmentation. In IEEE CVPR, 2020. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 581, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 287, + 624 + ], + "type": "text", + "content": "[43] Tengchao Lv, Yupan Huang, Jingye Chen, Lei Cui, Shuming Ma, Yaoyao Chang, Shaohan Huang, Wenhui Wang, Li Dong, Weiyao Luo, et al. Kosmos-2.5: A multimodal literate model. arXiv preprint arXiv:2309.11419, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 625, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 668 + ], + "type": "text", + "content": "[44] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In IEEE CVPR, 2016. 2, 4, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "text", + "content": "[45] Khanh Nguyen, Debadeepta Dey, Chris Brockett, and Bill Dolan. Vision-based navigation with language-based assistance via imitation learning with indirect intervention. In IEEE CVPR, 2019. 4" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[46] Zanlin Ni, Yulin Wang, Jiangwei Yu, Haojun Jiang, Yue Cao, and Gao Huang. Deep incubation: Training large models by divide-and-conquering. In IEEE CVPR, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 108, + 545, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 152 + ], + "type": "text", + "content": "[47] Zanlin Ni, Yulin Wang, Renping Zhou, Jiayi Guo, Jinyi Hu, Zhiyuan Liu, Shiji Song, Yuan Yao, and Gao Huang. Revisiting non-autoregressive transformers for efficient image synthesis. In CVPR, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 153, + 475, + 165 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 153, + 475, + 165 + ], + "spans": [ + { + "bbox": [ + 307, + 153, + 475, + 165 + ], + "type": "text", + "content": "[48] OpenAI. Gpt-4 technical report, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 167, + 545, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 167, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 308, + 167, + 545, + 220 + ], + "type": "text", + "content": "[49] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 223, + 545, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 223, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 308, + 223, + 545, + 255 + ], + "type": "text", + "content": "[50] Xuran Pan, Chunjiang Ge, Rui Lu, Shiji Song, Guanfu Chen, Zeyi Huang, and Gao Huang. On the integration of self-attention and convolution. In IEEE CVPR, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 258, + 545, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 258, + 545, + 290 + ], + "spans": [ + { + "bbox": [ + 308, + 258, + 545, + 290 + ], + "type": "text", + "content": "[51] Xuran Pan, Tianzhu Ye, Zhuofan Xia, Shiji Song, and Gao Huang. Slide-transformer: Hierarchical vision transformer with local self-attention. In IEEE CVPR, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 292, + 545, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 292, + 545, + 336 + ], + "spans": [ + { + "bbox": [ + 308, + 292, + 545, + 336 + ], + "type": "text", + "content": "[52] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 2, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 338, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 545, + 381 + ], + "type": "text", + "content": "[53] Yifan Pu, Yiru Wang, Zhuofan Xia, Yizeng Han, Yulin Wang, Weihao Gan, Zidong Wang, Shiji Song, and Gao Huang. Adaptive rotated convolution for rotated object detection. In IEEE ICCV, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 384, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 384, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 384, + 545, + 437 + ], + "type": "text", + "content": "[54] Yuankai Qi, Qi Wu, Peter Anderson, Xin Wang, William Yang Wang, Chunhua Shen, and Anton van den Hengel. Reverie: Remote embodied visual referring expression in real indoor environments. In IEEE CVPR, 2020. 1, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 440, + 545, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 440, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 308, + 440, + 545, + 495 + ], + "type": "text", + "content": "[55] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 497, + 545, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 497, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 308, + 497, + 545, + 552 + ], + "type": "text", + "content": "[56] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M Anwer, Erix Xing, Ming-Hsuan Yang, and Fahad S Khan. Glamm: Pixel grounding large multimodal model. arXiv preprint arXiv:2311.03356, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 554, + 545, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 554, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 308, + 554, + 545, + 586 + ], + "type": "text", + "content": "[57] Hengcan Shi, Hongliang Li, Fanman Meng, and Qingbo Wu. Key-word-aware network for referring expression image segmentation. In ECCV, 2018. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 588, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 588, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 308, + 588, + 545, + 632 + ], + "type": "text", + "content": "[58] Qie Sima, Sinan Tan, Huaping Liu, Fuchun Sun, Weifeng Xu, and Ling Fu. Embodied referring expression for manipulation question answering in interactive environment. In IEEE ICRA, 2023. 1, 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 634, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 634, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 308, + 634, + 545, + 667 + ], + "type": "text", + "content": "[59] Yan Tai, Weichen Fan, Zhao Zhang, Feng Zhu, Rui Zhao, and Ziwei Liu. Link-context learning for multimodal llms. arXiv preprint arXiv:2308.07891, 2023. 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "type": "text", + "content": "[60] Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. Stanford alpaca: An instruction-following llama model, 2023. 2" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3867" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[61] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 184 + ], + "type": "text", + "content": "[62] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 185, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 287, + 239 + ], + "type": "text", + "content": "[63] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, et al. Image as a foreign language: Beit pretraining for vision and vision-language tasks. In IEEE CVPR, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 241, + 287, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 287, + 284 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 287, + 284 + ], + "type": "text", + "content": "[64] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, et al. Vision: Large language model is also an open-ended decoder for vision-centric tasks. In NeurlPS, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 285, + 287, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 285, + 287, + 329 + ], + "spans": [ + { + "bbox": [ + 48, + 285, + 287, + 329 + ], + "type": "text", + "content": "[65] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, et al. Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 331, + 287, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 287, + 385 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 287, + 385 + ], + "type": "text", + "content": "[66] Xin Wang, Qiuyuan Huang, Asli Celikyilmaz, Jianfeng Gao, Dinghan Shen, Yuan-Fang Wang, William Yang Wang, and Lei Zhang. Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation. In IEEE CVPR, 2019. 1, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 387, + 287, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 387, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 48, + 387, + 287, + 420 + ], + "type": "text", + "content": "[67] Yulin Wang, Zhaoxi Chen, Haojun Jiang, Shiji Song, Yizeng Han, and Gao Huang. Adaptive focus for efficient video recognition. In IEEE ICCV, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 421, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 421, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 421, + 287, + 464 + ], + "type": "text", + "content": "[68] Yulin Wang, Rui Huang, Shiji Song, Zeyi Huang, and Gao Huang. Not all images are worth 16x16 words: Dynamic transformers for efficient image recognition. In NeurIPS, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 466, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 466, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 466, + 287, + 510 + ], + "type": "text", + "content": "[69] Yulin Wang, Yang Yue, Yuanze Lin, Haojun Jiang, Zihang Lai, Victor Kulikov, Nikita Orlov, Humphrey Shi, and Gao Huang. Adafocus v2: End-to-end training of spatial dynamic networks for video recognition. In IEEE CVPR, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 510, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 287, + 555 + ], + "type": "text", + "content": "[70] Yulin Wang, Yang Yue, Xinhong Xu, Ali Hassani, Victor Kulikov, Nikita Orlov, Shiji Song, Humphrey Shi, and Gao Huang. Adafocusv3: On unified spatial-temporal dynamic video recognition. In ECCV, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 555, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 599 + ], + "type": "text", + "content": "[71] Yulin Wang, Yang Yue, Rui Lu, Tianjiao Liu, Zhao Zhong, Shiji Song, and Gao Huang. Efficienttrain: Exploring generalized curriculum learning for training visual backbones. In IEEE ICCV, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 601, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 634 + ], + "type": "text", + "content": "[72] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clip-driven referring image segmentation. In IEEE CVPR, 2022. 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "text", + "content": "[73] Zhichao Wei, Xiaohao Chen, Mingqiang Chen, and Siyu Zhu. Learning aligned cross-modal representations for referring image segmentation. arXiv preprint arXiv:2301.06429, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "text", + "content": "[74] Penghao Wu and Saining Xie. V*: Guided visual search as a core mechanism in multimodal llms. arXiv preprint arXiv:2312.14135, 2023. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[75] Tsung-Han Wu, Giscard Biamby, David Chan, Lisa Dunlap, Ritwik Gupta, Xudong Wang, Joseph E Gonzalez, and Trevor Darrell. See, say, and segment: Teaching Imms to overcome false premises. arXiv preprint arXiv:2312.08366, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 162 + ], + "type": "text", + "content": "[76] Zhuofan Xia, Xuran Pan, Shiji Song, Li Erran Li, and Gao Huang. Vision transformer with deformable attention. In IEEE CVPR, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 164, + 545, + 197 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 164, + 545, + 197 + ], + "spans": [ + { + "bbox": [ + 307, + 164, + 545, + 197 + ], + "type": "text", + "content": "[77] Zhuofan Xia, Xuran Pan, Xuan Jin, Yuan He, Hui Xue, Shiji Song, and Gao Huang. Budgeted training for vision transformer. In ICLR, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 199, + 545, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 199, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 308, + 199, + 545, + 243 + ], + "type": "text", + "content": "[78] Zhuofan Xia, Xuran Pan, Shiji Song, Li Erran Li, and Gao Huang. Dat++: Spatially dynamic vision transformer with deformable attention. arXiv preprint arXiv:2309.01430, 2023.2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 245, + 545, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 245, + 545, + 299 + ], + "spans": [ + { + "bbox": [ + 308, + 245, + 545, + 299 + ], + "type": "text", + "content": "[79] Bin Xiao, Haiping Wu, Weijian Xu, Xiyang Dai, Houdong Hu, Yumao Lu, Michael Zeng, Ce Liu, and Lu Yuan. Florence-2: Advancing a unified representation for a variety of vision tasks. arXiv preprint arXiv:2311.06242, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 302, + 545, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 302, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 308, + 302, + 545, + 346 + ], + "type": "text", + "content": "[80] Jinjin Xu, Liwu Xu, Yuzhe Yang, Xiang Li, Yanchun Xie, Yi-Jie Huang, and Yaqian Li. u-llava: Unifying multimodal tasks via large language model. arXiv preprint arXiv:2311.05348, 2023. 2, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 348, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 348, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 348, + 545, + 392 + ], + "type": "text", + "content": "[81] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Lavt: Language-aware vision transformer for referring image segmentation. In IEEE CVPR, 2022. 2, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 394, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 394, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 308, + 394, + 545, + 426 + ], + "type": "text", + "content": "[82] Linwei Ye, Mrigank Rochan, Zhi Liu, and Yang Wang. Cross-modal self-attention network for referring image segmentation. In IEEE CVPR, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 428, + 545, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 428, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 308, + 428, + 545, + 483 + ], + "type": "text", + "content": "[83] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 485, + 545, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 485, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 308, + 485, + 545, + 518 + ], + "type": "text", + "content": "[84] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In ECCV, 2016. 4, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 520, + 545, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 520, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 308, + 520, + 545, + 564 + ], + "type": "text", + "content": "[85] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In IEEE CVPR, 2018. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 566, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 566, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 308, + 566, + 545, + 609 + ], + "type": "text", + "content": "[86] Ao Zhang, Liming Zhao, Chen-Wei Xie, Yun Zheng, Wei Ji, and Tat-Seng Chua. Next-chat: An Imm for chat, detection and segmentation. arXiv preprint arXiv:2311.04498, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 612, + 545, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 612, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 308, + 612, + 545, + 666 + ], + "type": "text", + "content": "[87] Renrui Zhang, Jiaming Han, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, Peng Gao, and Yu Qiao. Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "type": "text", + "content": "[88] Shilong Zhang, Peize Sun, Shoufa Chen, Min Xiao, Wenqi Shao, Wenwei Zhang, Kai Chen, and Ping Luo. Gpt4roi: Instruction tuning large language model on region-of-interest. arXiv preprint arXiv:2307.03601, 2023. 5" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3868" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 341 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[89] Zicheng Zhang, Yi Zhu, Jianzhuang Liu, Xiaodan Liang, and Wei Ke. Coupalign: Coupling word-pixel with sentence-mask alignments for referring image segmentation. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 172 + ], + "type": "text", + "content": "[90] Haozhe Zhao, Zefan Cai, Shuzheng Si, Xiaojian Ma, Kaikai An, Liang Chen, Zixuan Liu, Sheng Wang, Wenjuan Han, and Baobao Chang. Mmicl: Empowering vision-language model with multi-modal in-context learning. arXiv preprint arXiv:2309.07915, 2023. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 174, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 174, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 174, + 287, + 217 + ], + "type": "text", + "content": "[91] Yang Zhao, Zhijie Lin, Daquan Zhou, Zilong Huang, Jiashi Feng, and Bingyi Kang. Bubogpt: Enabling visual grounding in multi-modal llms. arXiv preprint arXiv:2307.08581, 2023. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 219, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 219, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 287, + 262 + ], + "type": "text", + "content": "[92] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 263, + 287, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 263, + 287, + 307 + ], + "spans": [ + { + "bbox": [ + 48, + 263, + 287, + 307 + ], + "type": "text", + "content": "[93] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In IEEE CVPR, 2023. 2, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 308, + 287, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 308, + 287, + 341 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 287, + 341 + ], + "type": "text", + "content": "[94] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. In NeurIPS, 2023. 2, 7" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "3869" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/G^3-LQ_ Marrying Hyperbolic Alignment with Explicit Semantic-Geometric Modeling for 3D Visual Grounding/7e2af7bd-e579-4e10-8895-d1ef28fe0cc7_content_list.json b/2024/G^3-LQ_ Marrying Hyperbolic Alignment with Explicit Semantic-Geometric Modeling for 3D Visual Grounding/7e2af7bd-e579-4e10-8895-d1ef28fe0cc7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..578a369260268b2b4f0e9a97ca7407d1fe604bdc --- /dev/null +++ b/2024/G^3-LQ_ Marrying Hyperbolic Alignment with Explicit Semantic-Geometric Modeling for 3D Visual Grounding/7e2af7bd-e579-4e10-8895-d1ef28fe0cc7_content_list.json @@ -0,0 +1,1695 @@ +[ + { + "type": "text", + "text": "$\\mathbf{G}^{3}$ -LQ: Marrying Hyperbolic Alignment with Explicit Semantic-Geometric Modeling for 3D Visual Grounding", + "text_level": 1, + "bbox": [ + 104, + 128, + 866, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuan Wang $^{1,2}$ Yali Li $^{1,2}$ Shengjin Wang $^{1,2\\ddagger}$", + "bbox": [ + 300, + 202, + 668, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Department of Electronic Engineering, Tsinghua University, China", + "bbox": [ + 217, + 220, + 751, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Beijing National Research Center for Information Science and Technology (BNRist), China wy23@mails.tsinghua.edu.cn {liyali13, wgsgj} @tsinghua.edu.cn $\\ddagger$ Corresponding Author", + "bbox": [ + 106, + 238, + 857, + 276 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/797fc1d01a607893704a400430e6eea3bd193b90bcef460b710919ce8241dddd.jpg", + "image_caption": [ + "Figure 1. Illustration of the proposed G $^3$ -LQ method on the ScanRefer benchmark. The versatile scene-level and object-level understanding properties empower G $^3$ -LQ to comprehend spatial relationship, relative distance, geometric attribute as well as intricate references." + ], + "image_footnote": [], + "bbox": [ + 93, + 311, + 879, + 478 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 532, + 313, + 547 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Grounding referred objects in 3D scenes is a burgeoning vision-language task pivotal for propelling Embodied AI, as it endeavors to connect the 3D physical world with free-form descriptions. Compared to the 2D counterparts, challenges posed by the variability of 3D visual grounding remain relatively unsolved in existing studies: 1) the underlying geometric and complex spatial relationships in 3D scene. 2) the inherent complexity of 3D grounded language. 3) the inconsistencies between text and geometric features. To tackle these issues, we propose $G^3$ -LQ, a DEtection Transformer-based model tailored for 3D visual grounding task. $G^3$ -LQ explicitly models Geometric-aware visual representations and Generates fine-Grained Language-guided object Queries in an overarching framework, which comprises two dedicated modules. Specifically, the Position Adaptive Geometric Exploring (PAGE) unearths underlying information of 3D objects in the geometric details and spatial relationships perspectives. The Fine-grained Language-guided Query Selection (Flan-QS) delves into syntactic structure of texts and generates object queries that exhibit higher relevance towards fine-grained text features. Finally, a pioneering Poincaré Semantic Alignment (PSA) loss establishes semantic-geometry consistencies by mod", + "bbox": [ + 73, + 553, + 472, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "eling non-linear vision-text feature mappings and aligning them on a hyperbolic prototype—Poincaré ball. Extensive experiments verify the superiority of our $G^3$ - $LQ$ method, trumping the state-of-the-arts by a considerable margin.", + "bbox": [ + 498, + 532, + 892, + 594 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 603, + 632, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Grounding 3D physical realm with natural languages has propelled to the forefront in advancing Embodied AI [10, 18, 35], an ability that empowers agents to comprehend human instructions in real-world contexts. 3D Visual Grounding (3D VG) [6, 31, 43, 48] has garnered substantial attention as a crucial cross-modal 3D perception task in defacto applications, e.g., assistive robots, AR/VR, and metaverse.", + "bbox": [ + 496, + 628, + 892, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Compared with extensive explorations of 2D counterpart [20, 28, 29, 49], 3D VG poses prominent challenge known as \"Semantic-Geometric Coupling\". As shown in Fig. 1, we offer distinctive glimpse into it: 1) Intricate geometric details and spatial relations. 3D scenes encounter multiple objects with intricate layout, multi-level occlusion, and multifaceted spatial relation. 2) Heightened flexibility and complexity of descriptions. Free-form utterances with complex syntax convey precise spatial information, intricate references, and detailed relational terms. 3) Inconsistency of the semantic-geometric features. A single semantic concep", + "bbox": [ + 496, + 734, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13917", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tion yields multiple potential correspondences within point cloud or exhibit diverse geometry in varied contexts.", + "bbox": [ + 75, + 90, + 468, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To ameliorate these issues, some pioneering methods have been proposed. They have largely focused on enhancing the representation capacity of 3D point clouds with 2D image priors [44]; capturing spatial relationships of 3D objects with powerful transformer [5, 15, 48] or graph convolution network [3, 12]; generating discriminative object proposals through instance segmentation [16] or linguistic guidance [31]. Despite commendable advancements be achieved, an equally important yet underexplored problem is explicit semantic-geometric modeling. Prior studies exhibit a significant dearth in explicitly modeling the intrinsic geometric attributes and spatial layouts of 3D objects, accentuating the dilemma to associate text descriptions with grounding objects. Moreover, object queries contain rich instance characteristics. But the object proposal [15, 48] and query generation [19, 31, 43] in current DETR-like models overlook explicit fine-grained interaction between text and object features, which in turn limit the model's capacity to refer potential language clues, e.g., appearances, shapes, or textures. Finally, major 3D VG systems [19, 21, 43, 51] strive to establish the alignment between point cloud and semantic concepts in canonical Euclidean space. However, their efficacy falters in unraveling the complex non-linear relationships between multi-modal features and upholding the intrinsic hierarchy natures in 3D scenes and texts.", + "bbox": [ + 75, + 123, + 472, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, our primary goal endeavors to explicitly delve into fine-grained representations of geometric and semantic concepts, then aligns two modalities in hyperbolic space. We thus devise an overarching G $^3$ -LQ framework:", + "bbox": [ + 75, + 503, + 468, + 564 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) For geometric feature modeling, we firstly propose a Proxy Adaptive Geometric Exploring (PAGE) module. PAGE is composed of the Point-Proxy Geometric Extraction and Proxy Adaptive Geometric Refinement component. The former explicitly capture the local geometric attributes (e.g., distance and orientation) in a 3D local proxy, while the latter adaptively refine the geometric structure according to the position of point proxy. The geometric features are fed into the geometric encoder for high-level embedding and further achieve the interaction with text features for cross-modality enhancement. In this fashion, the PAGE module facilitates a seamless alignment of text descriptions with intrinsic geometric properties of 3D objects.", + "2) For complex utterance understanding, inspired by the dependency parsing, we aim to explore context dependency and syntactic structure of texts for guiding query selection. To this end, we propose a Fine-grained Language-guided Query Selection (Flan-QS) module, comprising the Language Scene Graph and Fine-grained Query Selection components. We decouple the free-form utterances to tree-like semantic components with directed dependency relationships as edges. Then, we establish and update the lan" + ], + "bbox": [ + 75, + 566, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "guage scene graph to represent the context-aware entry features via node/edge embedding and graph context learning. Finally, we employ the constructed language scene graph to steer the generation of subsequent object candidates that are densely-aligned with fine-grained linguistic attributes.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3) For semantic-geometric consistency, we embed vision and text features to a hyperbolic prototype—Poincaré ball. The emerging appreciation [11, 13, 34] suggest that the hyperbolic space offers overwhelming superiority to handle data correlations. 3D scene (free-form text) exhibits an inherent hierarchical nature ranging from scene, object, part (from sentence, phrase, word). The hyperbolic space with negative curvature is the only space that can successfully embed the tree-shaped hierarchy [33]. Moreover, the compact Poincaré ball [13, 34] holds suitable geometric properties, enhancing its capacity to model non-linear mappings of vision and text features. Therefore, we design a Poincaré Semantic Alignment loss (PSA) to probe the complex mappings between vision and text features, facilitating the semantics-geometry consistency.", + "bbox": [ + 496, + 167, + 892, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our chief contributions are threefold:", + "bbox": [ + 517, + 393, + 767, + 407 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We roundly analysis the Semantic-Geometric Coupling of 3D VG task. Towards this issue, we propose a unified $\\mathbf{G}^3$ -LQ framework to model intrinsic geometric features (PAGE module) and explicitly generate fine-grained object queries for 3D visual grounding (Flan-QS module).", + "- We propose a novel PSA loss to capture the non-linear relationship and underlying hierarchy of the vision-text features in the hyperbolic space, which further paves the way for semantics-geometry consistency learning.", + "- A battery of experiments on ScanRefer and Nr3d/Sr3d datasets demonstrate the effectiveness of our G $^{3}$ -LQ method. G $^{3}$ -LQ has proven its mettle, outperforming all existing SOTA methods by a considerable margin." + ], + "bbox": [ + 500, + 407, + 890, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 609, + 650, + 626 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1.3D Visual Grounding", + "text_level": 1, + "bbox": [ + 500, + 635, + 700, + 651 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Visual Grounding has been an area of intense investigation that dedicates to pinpointing the target object in a 3D scene based on the text description. Prevailing methods are typically two-stage, which conforms to the detection-then-matching paradigm. Firstly, the 3D object detectors [30, 37, 48] are harnessed to yield candidate 3D object proposals, while the text features are encoded by language models [9, 27]. Secondly, these advanced methods perform a pivotal endeavor to align the vision and text features for target objects grounding. Among these, TGNN [16] utilizes the graph neural network to deduce spatial relationships among 3D object proposals, which are further enriched by text features. FFL-3DOG [12] captures the intramodal and cross-modal relationships of text descriptions and 3D point cloud via scene graphs interaction. Recent endeavors embrace the formidable transformers [40] to model", + "bbox": [ + 496, + 657, + 893, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13918", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the relationship of proposals and text features. 3DVG [48] and TransRefer3D [15] achieve exceptional performance by establishing self-attention and cross-attention protocols for context perception. LanguageRefer [39] seamlessly converts the cross-modal task into a unified language modeling challenge, underpinned by predicted object labels.", + "bbox": [ + 75, + 90, + 468, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, two-stage methods encounter a noteworthy impediment: the detection stage neglects to leverage language contexts to prioritize the objects that are essential to the referring task. An alternative approach to the problem is one-stage methods [19, 31, 43], where extracted vision features are directly fused with the language features for object grounding. For instance, 3D-SPS [31] leverages text features to guide visual keypoints selection, enabling the progressive object grounding. BUTD-DETR [19] pioneers a DETR-like model, which fuses the vision-text features via a co-attention paradigm and then decodes objects in the utterance from the contextualized features. Building upon this, EDA [43] proposes a text-decoupled strategy and performs dense alignment of 3D objects and associated texts.", + "bbox": [ + 75, + 183, + 470, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Despite one-stage methods unfold prowess performance, inherent limitations have been yet underexplored: 1) They overlook the intrinsic geometric attributes (distance, orientation, and shape) in 3D point clouds, leading to ambiguous visual representations. Conversely, our proposed PAGE module explicitly captures geometric features to avoid ambiguity. 2) The queries generation lack the fine-grained language guidance, limiting the model's capacity to infer potential language clues. Our tailored Flan-QS module explores language prior and context dependency to generate precise proposals. 3) They establish the vision-text alignment in Euclidean space, incapable of modeling the nonlinear relation of vision and text features. Our method embeds multimodal features on the Poincaré ball then leverage the PSA loss to foster the semantic-geometric alignment.", + "bbox": [ + 75, + 395, + 470, + 623 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Geometric Extraction in 3D Point Clouds", + "text_level": 1, + "bbox": [ + 76, + 633, + 429, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The explicit modeling of geometric structures has recently garnered considerable attention in the realms of 3D point cloud, e.g., point cloud understanding [8, 26, 32, 41], point cloud completion [46], point cloud registration [38], and 3D object detection [45]. DGCNN [41] proposes an EdgeConv operator, by which the point features are aggregated through graph convolution to capture intricate local geometric structures. RSCNN [26] explicitly encodes geometric relationship of points with relation-aware convolution. PointMixer [8] embeds geometric relationships of 3D point in a MLP-Mixer's streamline. PoinTr [46] devises a geometry-aware transformer to learn structural knowledge and local geometric relationships for point cloud completion. HGNet [45] delves into localized shape information to generate the proposals with a hierarchical graph model for 3D object detection. Our cutting-edge $\\mathrm{G}^3$ -LQ incorporates", + "bbox": [ + 75, + 659, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the geometric attributes of point cloud into the 3D VG task, thereby enhancing the discriminability of vision features.", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 127, + 591, + 142 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 500, + 151, + 609, + 165 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Provided a point cloud $\\mathcal{P} = \\{\\pmb{p}_i\\}_{i=1}^N \\in \\mathbb{R}^{N \\times (3 + F)}$ with $F$ -dim auxiliary features (e.g., RGB, normals, or multi-view features) of $N$ points, and a free-form text $\\mathcal{T}$ with $L$ words, the primary goal of 3D VG is to establish a mapping $\\mathcal{M}$ that links $\\mathcal{P}$ and $\\mathcal{T}$ to the target object $\\pmb{o}$ , i.e., $\\mathcal{M}(\\mathcal{P}, \\mathcal{T}) \\to \\pmb{o}$ .", + "bbox": [ + 496, + 172, + 890, + 250 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fig. 2 gives the outline of our proposed $\\mathrm{G}^3$ -LQ method. Firstly, we leverage the well-accepted PointNet++ [36] tockenize the point cloud into vision token $\\mathbf{V}\\in \\mathbb{R}^{n\\times d}$ . The text is encoded by the pre-trained RoBERTa [27] and yields the vanilla text token $\\mathbf{T}\\in \\mathbb{R}^{t\\times d}$ . The vision token $\\mathbf{V}$ is fed into the PAGE module to capture underlying geometric features $\\mathbf{V}_g\\in \\mathbb{R}^{n\\times d}$ , which adjusts according to the point positions, thereby furnishing precise geometric clues. Further, the geometric visual encoder embeds $\\mathbf{V}_g$ via self-attention and interact with text features $\\mathbf{T}$ via cross-attention.", + "bbox": [ + 496, + 250, + 890, + 400 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Finally, we build the language scene graph $\\mathcal{G}$ and obtain the updated fine-grained text features, which will explicitly guide the queries selection. Like the object queries in most DETR-like models [22, 25, 43], the selected queries $\\mathbf{Q} \\in \\mathbb{R}^{K \\times d}$ are fed into a cross-modal decoder to probe desired features and update themselves. The decoded queries $\\mathbf{Q}' \\in \\mathbb{R}^{K \\times l}$ are fed into an MLP to predict object boxes.", + "bbox": [ + 496, + 401, + 890, + 507 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Geometric Exploring and Embedding", + "text_level": 1, + "bbox": [ + 500, + 515, + 825, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we propose a PAGE module to explicitly capture geometric properties of 3D scenes, which serves to encode holistic visual representations for 3D VG tasks.", + "bbox": [ + 496, + 539, + 890, + 583 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Point-Proxy Geometric Extraction. To facilitate local geometric features modeling, we firstly instantiate point proxies predicated on the proximity of the points, thereby serving as the local regions of the point cloud. As illustrated on Fig 2, we search $k$ neighbors of $p_i \\in \\mathcal{P}$ as a point proxy with the KNN algorithm, denoted as: $\\mathcal{X}_i = k\\mathrm{NN}(\\mathcal{P},p_i)$ .", + "bbox": [ + 496, + 584, + 890, + 674 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The point proxy $\\mathcal{X}_i$ denotes index set with $k$ closest points of $\\pmb{p}_i$ and $\\pmb{v}_{m_i}:\\{\\pmb{v}_{m_{ij}}|j\\in \\mathcal{X}_i\\} \\in \\mathbb{R}^{k\\times d}$ is the corresponding features. With the Point-Proxy Geometric Extraction component, we capture the local structure by learning the geometric topology features within the point proxy $\\mathcal{X}_i$ . In this fashion, it is effective to achieve an inductive local representations that integrate explicit reasoning pertaining to geometric structures and objects spatial layout. We primarily compute pre-defined geometric prior $\\widetilde{\\pmb{p}}_{ij}\\in \\mathbb{R}^7$ :", + "bbox": [ + 496, + 675, + 890, + 811 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {\\boldsymbol {p}} _ {i j} = \\operatorname {c o n c a t} \\left(\\boldsymbol {p} _ {i} - \\boldsymbol {p} _ {i j}, \\boldsymbol {p} _ {i}, \\left\\| \\boldsymbol {p} _ {i} - \\boldsymbol {p} _ {i j} \\right\\|\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 816, + 890, + 837 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\widetilde{\\pmb{p}}_{ij}$ embeds relative orientation and distance of different objects. Further, the low-level geometric features are encoded to high-level features $\\pmb{h}_{ij} \\in \\mathbb{R}^d$ with a linear layer $\\phi[\\cdot]$ to obtain abstract geometric conception and excavate", + "bbox": [ + 496, + 840, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13919", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b67b4ac56f727ff4ffbe50de3d23ae65a9b69c8a024e406fcf3168d59955884b.jpg", + "image_caption": [ + "Text: $\\mathcal{T} = \\{t_i\\}_{i = 1}^L$", + "Point: $\\mathcal{P} = \\{\\pmb {p}_i\\}_{i = 1}^N$" + ], + "image_footnote": [], + "bbox": [ + 106, + 94, + 395, + 426 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1cdd38271661ca0d94a0c929951da1fe2b660e3702bc82dfe7edaa945e4be369.jpg", + "image_caption": [ + "Figure 2. Illustration of (a) our proposed $\\mathbf{G}^3$ -LQ framework. The red dashed frames denote our novel modules. (b) PAGE moulder to model the underlying geometric structure of 3D point cloud and (c) the Geometric Visual Encoder further to embed the geometric topology features. (d) Flan-QS module to capture fine-grained language context with (e) Language Scene Graph for guiding query selection." + ], + "image_footnote": [], + "bbox": [ + 397, + 95, + 867, + 452 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "discriminative geometric structures, which is described as:", + "bbox": [ + 76, + 520, + 460, + 535 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nh _ {i j} = \\widetilde {\\boldsymbol {p}} _ {i j} + \\phi [ \\widetilde {\\boldsymbol {p}} _ {i j} ], \\forall j \\in \\mathcal {X} _ {i} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 540, + 468, + 560 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proxy-Adaptive Geometric Refinement. The presence of diverse geometric structures across distinct point proxies necessitate anisotropic extractors. Further, we flesh out this intuition and engineer a Proxy-Adaptive Geometric Refinement component, poised to model refined geometric features effectively of diverse point proxies. Given the geometric prior $\\{h_{ij}|j\\in \\mathcal{X}_i\\} \\in \\mathbb{R}^{k\\times d}$ in $\\mathcal{X}_i$ we establish local geometric refinement by the following streamline:", + "bbox": [ + 75, + 571, + 468, + 691 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{g _ {i j} \\right\\} = \\alpha \\odot \\frac {\\left\\{h _ {i j} \\right\\} - h _ {i}}{\\Omega + \\varepsilon} + \\beta \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 696, + 468, + 728 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\varepsilon$ is a small constant for numerical stability and $\\alpha, \\beta$ are learnable parameters for features refinement. $\\Omega$ is the standard deviation across all point proxies and channels. $h_i$ is the center of $\\mathcal{X}_i$ . $g_i = \\{g_{ij} | j \\in \\mathcal{X}_i\\} \\in \\mathbb{R}^{k \\times d}$ represents the anisotropic geometric structures in point proxy $\\mathcal{X}_i$ .", + "bbox": [ + 75, + 731, + 468, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, we opt for a softmax aggregation strategy to densely relate point features interaction in each $\\mathcal{X}_i$ :", + "bbox": [ + 76, + 815, + 468, + 845 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {v} _ {g _ {i}} = \\operatorname {L i n e a r} \\left(\\operatorname {c o n c a t} \\left(v _ {m _ {i}}, g _ {i}\\right)\\right) \\in \\mathbb {R} ^ {k \\times d}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 848, + 403, + 867 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nv _ {g _ {i}} = \\sum_ {j \\in \\mathcal {X} _ {i}} \\operatorname {s o f t m a x} \\left(\\widetilde {v} _ {g _ {i j}}\\right) \\cdot \\widetilde {v} _ {g _ {i j}} \\in \\mathbb {R} ^ {d}, \\forall j \\in \\mathcal {X} _ {i}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 869, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where Linear $(\\cdot)$ denotes the linear projection, $\\mathbf{V}_g = \\{v_{g_i}\\}_{i=1}^n$ is the output token with geometric representations.", + "bbox": [ + 500, + 520, + 890, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Geometric Visual Encoder. Geometric perception enables to strengthen shape features of 3D objects and demonstrate enhanced viewpoint sensitivity, which bestows flexibility in capturing spatial relation and improves precise 3D objects grounding. Hence, we embed geometric topology features $\\mathbf{V}_g$ via a tailored Geometric Encoder. As shown in Fig. 2, the core of Geometric Encoder is Geometric self-attention layer. Given the high-level vision features $\\mathbf{V}_m\\in \\mathbb{R}^{n\\times d}$ and the explicit geometric features $\\mathbf{V}_g$ , the output $\\mathbf{V}_{emb}\\in \\mathbb{R}^{n\\times d}$ is a weighted sum of projected features:", + "bbox": [ + 498, + 550, + 890, + 700 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nv _ {i} ^ {e m b} = \\sum_ {j = 1} ^ {n} a _ {i j} \\left(v _ {m _ {j}} \\mathbf {W} ^ {V}\\right) \\in \\mathbb {R} ^ {d} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 700, + 890, + 741 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The attention weight $a_{ij}$ is computed by a row-wise softmax, which can be formulated as:", + "bbox": [ + 498, + 742, + 890, + 771 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\na _ {i j} = \\sigma \\left[ \\frac {\\left(v _ {g _ {i}} \\mathbf {W} _ {g} ^ {Q} + v _ {m _ {i}} \\mathbf {W} _ {m} ^ {Q}\\right) \\left(v _ {g _ {j}} \\mathbf {W} _ {g} ^ {K} + v _ {m _ {j}} \\mathbf {W} _ {m} ^ {K}\\right) ^ {T}}{\\sqrt {d}} \\right] \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 775, + 890, + 815 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathbf{W}_g^Q,\\mathbf{W}_g^K$ are geometric projection matrices of query and key, $\\mathbf{W}_m^Q,\\mathbf{W}_m^K$ and $\\mathbf{W}_m^V$ are semantic projection matrices of query, key and value. $\\sigma$ is the row-wise softmax operator.", + "bbox": [ + 498, + 818, + 890, + 868 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Further, the geometric-enhanced features $\\mathbf{V}_{emb}$ pass through the cross-attention layer to interact with the text", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13920", + "bbox": [ + 480, + 989, + 517, + 1000 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "features and box features (used in two stage), for obtaining the cross-modal features $\\mathbf{V}' \\in \\mathbb{R}^{n \\times d}$ and $\\mathbf{T}' \\in \\mathbb{R}^{t \\times d}$ . Finally, $\\mathbf{V}'$ is linearly projected as the candidate query $\\mathbf{Q}_c$ .", + "bbox": [ + 76, + 90, + 468, + 137 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Query Selection", + "text_level": 1, + "bbox": [ + 76, + 146, + 235, + 161 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the DETR-like model [24, 42, 50], object queries play essential roles to determine potential regions of targets, which are equally prominent for 3G VG systems, decoded into a box center and size that conform to vision tokens. Assisted by global text features, GroundingDINO [25] employs a language-guided query selection protocol to generate query proposals. However, such text features are relatively coarse-grained, leading to inaccurate query generation. Actually, free-form texts harbor intricate syntactic structure and possess a context-dependency nature. It further serves as inspiration for us to generate precise vision queries via improving fine-grained semantic understanding. Language Scene Graph. We intend to delve into fine-grained text information, which is defined as decoupled semantic components (e.g., object, attributes, and relations) and syntactic structures. To this end, we perform a syntactic dependency parsing to construct language graph $\\mathcal{G} = \\{\\mathcal{O},\\mathcal{R}\\}$ by the off-the-shelf Scene Graph Parser [1], which serves as a reasoning inductive bias of 3D VG tasks. $\\mathcal{O} = \\{\\pmb{o}_i\\}$ and $\\mathcal{R} = \\{\\pmb{r}_{ij}\\}$ are nodes and edges set. Each object $o_i$ is denoted as a phrase with a set of attributes (appearance, shape, texture). Harnessing the final layer of the pre-trained RoBERTa model, we then acquire the object phrase embedding $\\pmb{x}_{o_i}\\in \\mathbb{R}^d$ as the graph node $o_i\\in \\mathcal{O}$ . In a similar vein, the relation phrases are also encoded into the context-aware embeddings $\\pmb{x}_{r_{ij}}\\in \\mathbb{R}^d$ akin to the graph edges $\\pmb{r}_{ij}\\in \\mathcal{R}$ .", + "bbox": [ + 75, + 170, + 468, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inspired by the message passing mechanism [14, 23], we proceed to enhance relation and object embeddings via language graph convolution. Firstly, to enrich the contextual representations of relation embedding in connection with its affiliated nodes, we aggregate messages and update its features with the subject $\\pmb{x}_{o_i}$ and object node $\\pmb{x}_{o_j}$ :", + "bbox": [ + 75, + 566, + 468, + 657 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {\\boldsymbol {x}} _ {r _ {i j}} = \\boldsymbol {x} _ {r _ {i j}} + F _ {r} \\left(\\left[ \\boldsymbol {x} _ {o _ {i}}; \\boldsymbol {x} _ {r _ {i j}}; \\boldsymbol {x} _ {o _ {j}} \\right]\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 661, + 468, + 680 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $F_{r}$ is a linear projection layer and $\\widetilde{\\pmb{x}}_{r_{ij}}$ is the context-aware relation embedding. We then update the phrase embedding $\\pmb{x}_{o_i}$ by passing messages from all connected nodes $\\Gamma (i)$ and the edges with a tailored attention mechanism:", + "bbox": [ + 76, + 684, + 468, + 744 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {\\boldsymbol {x}} _ {o _ {i}} = \\boldsymbol {x} _ {o _ {j}} + \\sum_ {j \\in \\Gamma (i)} \\boldsymbol {w} _ {o _ {i j}} F _ {o} \\left(\\left[ \\boldsymbol {x} _ {o _ {j}}; \\boldsymbol {x} _ {r _ {i j}} \\right]\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 750, + 468, + 782 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $F_{o}$ is another linear layer and $\\widetilde{\\pmb{x}}_{o_i}$ denotes context-aware object features. $\\pmb{w}_{o_{ij}}$ is the tailored attention weight:", + "bbox": [ + 76, + 792, + 468, + 824 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {w} _ {o _ {i j}} = \\operatorname {s o f t m a x} \\left[ F _ {o} \\left(\\left[ \\boldsymbol {x} _ {o _ {i}}; \\widetilde {\\boldsymbol {x}} _ {r _ {i j}} \\right]\\right) ^ {T} F _ {o} \\left(\\left[ \\boldsymbol {x} _ {o _ {j}}; \\widetilde {\\boldsymbol {x}} _ {r _ {i j}} \\right]\\right) \\right] \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 839, + 468, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The attention protocol manifests the capability to discern crucial contextual features from language graph, serving as", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "potent language priors for query selection.", + "bbox": [ + 498, + 92, + 779, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fine-grained Query Selection. We input the candidate query features $\\mathbf{Q}_c\\in \\mathbb{R}^{n\\times d}$ from the cross-modal encoding layer into the Query Embedding component, which consists of several stacked Fully-connected (FC) layers and BN layers. Further, we leverage the fine-grained text features $T_{f}$ to guide the query selection through a cross-attention fusion layer. The resulting query tokens are fed into an MLP layer for confidence score prediction. Through selecting the top- $K$ highest scoring tokens for the following decoder, we derive the updated object queries $\\mathbf{Q}_s\\in \\mathbb{R}^{K\\times d}$ , which are utilized to predict a box center, height and width.", + "bbox": [ + 496, + 106, + 890, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Poincaré Semantic Alignment loss", + "text_level": 1, + "bbox": [ + 500, + 277, + 795, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We linearly embed the queries $\\mathbf{Q}_s$ to the predicted object features $\\pmb{q} \\in \\mathbb{R}^{K \\times 64}$ , which aligns with the embedded text features $\\pmb{t} \\in \\mathbb{R}^{t \\times 64}$ . Noteworthy, point clouds (spanning from scenes, objects to parts) and texts (ranging from sentences, phrases to words) encompass inherent hierarchy. Further, granularity disparities of geometric and semantic representations complicate the intricate correlations. However, it is the Achilles' heel to model hierarchical structures and complex relationships in canonical Euclidean space.", + "bbox": [ + 496, + 301, + 890, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To tackle this issue, we provide the insight to map vision-text features to a special hyperbolic space—Poincaré ball, for its powerful natural aptitude in modeling tree-shaped features. Based on this, we carefully devise a PSA loss to pave the way for semantic-geometric consistency. Benefiting from exponential volume expansion of hyperbolic embedding, PSA explicitly models complex semantic-geometric relationship for capturing hierarchical structures.", + "bbox": [ + 496, + 438, + 890, + 558 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since the Poincaré ball is a Riemannian manifold, we project the vision $\\pmb{q}$ and text features $\\pmb{t}$ to the Poincaré ball $\\mathbb{D}_c^d$ with the curvature $c$ via an exponential mapping function $\\exp_{\\pmb{x}}^{c}: \\mathbb{R}^{d} \\to \\mathbb{D}_{c}^{d}$ , which can be defined as follows:", + "bbox": [ + 498, + 558, + 890, + 619 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {q} ^ {\\mathrm {c}} = \\exp_ {\\boldsymbol {x}} ^ {\\mathrm {c}} (\\boldsymbol {q}) = \\boldsymbol {x} \\oplus_ {c} \\left(\\tanh \\left(\\sqrt {c} \\frac {\\lambda_ {\\boldsymbol {x}} ^ {\\mathrm {c}} \\| \\boldsymbol {q} \\|}{2}\\right) \\frac {\\boldsymbol {q}}{\\sqrt {c} \\| \\boldsymbol {q} \\|}\\right) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 635, + 890, + 666 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pmb{x}$ is a fixed base point to make formulas less cumbersome and $\\lambda_{\\pmb{x}}^{c} = 2 / (1 - c\\| \\pmb{x}\\|^{2})$ . $t^c$ is defined in a similar approach. $\\oplus_{c}$ denotes the addition operator in the Mobius gyrovector space. The geodesic distance between the a vision-text feature pair $(\\pmb{q}_i^c,\\pmb{t}_j^c)$ on the Poincaré ball $\\mathbb{D}_c^d$ is:", + "bbox": [ + 496, + 670, + 890, + 747 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nd _ {c} \\left(\\boldsymbol {q} _ {i} ^ {c}, \\boldsymbol {t} _ {j} ^ {c}\\right) = \\frac {2}{\\sqrt {c}} \\operatorname {a r c t a n h} \\left(\\sqrt {c} \\| - \\boldsymbol {q} _ {i} ^ {c} \\oplus_ {c} \\boldsymbol {t} _ {j} ^ {c} \\|\\right) \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 534, + 750, + 890, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inspired by [43], we term the PSA loss to densely align the fine-grained vision-text features via contrastive learning on Poincaré ball. The query loss is defined as:", + "bbox": [ + 496, + 781, + 890, + 825 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {p s a}} ^ {q} = \\sum_ {i = 1} ^ {K} \\frac {- 1}{\\left| \\mathcal {T} _ {i} ^ {+} \\right|} \\sum_ {\\boldsymbol {t} _ {i} ^ {c} \\in \\mathcal {T} _ {i} ^ {+}} \\log \\left[ \\right. \\frac {\\exp \\left[ \\boldsymbol {w} _ {+} \\left(- d _ {c} \\left(\\boldsymbol {q} _ {i} ^ {c} , \\boldsymbol {t} _ {i} ^ {c}\\right) / \\tau\\right)\\right]}{\\sum_ {j = 1} ^ {l} \\exp \\left[ \\boldsymbol {w} _ {-} \\left(- d _ {c} \\left(\\boldsymbol {q} _ {i} ^ {c} , \\boldsymbol {t} _ {j} ^ {c}\\right) / \\tau\\right)} \\right] \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 833, + 890, + 880 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $k$ and $l$ are the number of object tokens and text to-", + "bbox": [ + 500, + 885, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13921", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/04a528d765c1989069c791d625ba58086399593d204637a02fade71bd9dee0dd.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodVenueInputUnique (~19%)Multiple (~81%)Overall
0.250.50.250.50.250.5
ScanRefer [6]ECCV'203D+2D76.3353.5132.7321.1141.1927.40
TGNN [16]AAAI'213D68.6156.8029.8423.1837.3729.70
InstanceRefer [47]ICCV'213D77.4566.8331.2724.7740.2332.93
SAT [44]ICCV'213D+2D73.2150.8337.6425.1644.5430.14
FFL-3DOG [12]ICCV'213D78.8067.9435.1925.741.3334.01
3DVG [48]ICCV'213D+2D81.9360.6439.328.4247.5734.67
3D-SPS [31]CVPR'223D+2D84.1266.7240.3229.8248.8236.98
BUTD-DETR [19]ECCV'223D82.8864.9844.7333.9750.4238.60
ViL3DRel [7]NeurlPS'223D81.5868.6240.3030.7147.9437.73
EDA [43]CVPR'233D85.7668.5749.1337.6454.5942.26
3DJCG [5]CVPR'223D+2D83.3764.3441.3930.8249.5637.33
3D-VLP [21]CVPR'233D+2D84.2364.6143.5133.4151.4139.46
3D-VisTA(Scratch) [51]ICCV'233D77.4070.9038.7034.8045.9041.50
G3-LQ(Two-Stage)——3D88.0972.7351.4840.8056.9045.58
3D-SPS(One-stage) [31]CVPR'223D81.6364.7739.4829.6147.6536.43
BUTD-DETR(One-stage) [19]ECCV'223D81.4761.2444.2032.8149.7637.05
EDA(One-stage) [43]CVPR'233D86.4069.4248.1136.8253.8341.70
G3-LQ(One-stage)——3D88.5973.2850.2339.7255.9544.72
", + "bbox": [ + 78, + 87, + 893, + 371 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "kens, $\\tau$ is the temperature coefficient. $\\pmb{w}^{+}$ $w^{-}$ are weights of positive and negative terms. Similar to EDA, $\\pmb{t}_i$ is the positive text features of the $i$ -th object query and $\\mathcal{T}_i^+$ are the positive features set of $\\pmb{t}_i$ , which primarily includes the object entries, attributes, relations, and pronouns. Moreover, the text loss can be formulated similarly:", + "bbox": [ + 75, + 438, + 468, + 529 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {p s a}} ^ {t} = \\sum_ {i = 1} ^ {l} \\frac {- 1}{\\left| \\mathcal {Q} _ {i} ^ {+} \\right|} \\sum_ {\\boldsymbol {q} _ {i} ^ {c} \\in \\mathcal {Q} _ {i} ^ {+}} \\log \\left[ \\right. \\frac {\\exp \\left[ \\boldsymbol {w} _ {+} \\left(- d _ {c} \\left(\\boldsymbol {t} _ {i} ^ {c} , \\boldsymbol {q} _ {i} ^ {c}\\right) / \\tau\\right)\\right]}{\\sum_ {j = 1} ^ {K} \\exp \\left[ \\boldsymbol {w} _ {-} \\left(- d _ {c} \\left(\\boldsymbol {t} _ {i} ^ {c} , \\boldsymbol {q} _ {j} ^ {c}\\right) / \\tau\\right)} \\right] \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 537, + 468, + 583 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\pmb{q}_i^c\\in \\mathcal{Q}_i^+$ is the positive object features of $\\pmb{t}_i^c$ . The final loss $\\mathcal{L}_{\\mathrm{psa}}$ is the average of the above two terms.", + "bbox": [ + 76, + 590, + 468, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following [43], the total loss of our method also includes the box regression loss $\\mathcal{L}_{\\mathrm{loc}}$ , the position alignment loss $\\mathcal{L}_{\\mathrm{pa}}$ which details in the Supplementary Material.", + "bbox": [ + 76, + 621, + 468, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 76, + 674, + 200, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Firstly, we carry out comparisons with SOTA methods on ScanRefer and $\\mathrm{Nr3D / Sr3D}$ dataset in Sec. 4.1. Further, we delve into ablation studies in Sec. 4.2. Finally, we visualize and analyze the 3D VG results in Sec. 4.3.", + "bbox": [ + 75, + 699, + 468, + 761 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Quantitative Comparisons", + "text_level": 1, + "bbox": [ + 76, + 770, + 316, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Performance on the ScanRefer. As illustrated in Table. 1, our method consistently exhibits admirably performance, trumping all competitors by a substantial margin on all test subsets. Our method achieves $56.90\\%$ and $45.58\\%$ performance in terms of the overall accuracy, with a remarkable improvement compared with EDA [43], $2.31\\%$ and $2.34\\%$ . 1) Compared with 2D assistance methods [17, 44]", + "bbox": [ + 75, + 794, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1848c3b839dd84d9fdd4714d75b6ba643df9c12e457e327ea873821c17bd41da.jpg", + "table_caption": [ + "Table 1. Comparison results on the ScanRefer dataset, in terms of the accuracy evaluated by IoU 0.25 and IoU 0.5. The unique denotes samples devoid of distracting objects, while multiple applies to remaining samples. In the one-stage setting, there is no reliance on the additional 3D object detection step. Our proposed $\\mathrm{G}^3 -\\mathrm{LQ}$ performs favorably over current state-of-the-art approach EDA [43]." + ], + "table_footnote": [], + "table_body": "
MethodNr3DSr3D
OverallHardOverallHard
TGNN [16]37.330.645.036.9
InstanceRefer [47]38.831.848.040.5
3DVG [48]40.834.851.444.9
LanguageRefer [39]43.936.656.049.3
TransRefer3D [15]48.039.657.450.2
SAT [44]49.242.457.950.0
LAR [4]48.942.359.451.2
3DRef [2]47.038.339.032.0
3D-SPS [31]51.545.162.665.4
MVT [17]55.149.164.558.8
BUTD-DETR [19]54.648.467.063.2
EDA [43]52.146.168.162.9
3D-VisTA [51]57.549.469.663.6
G3-LQ58.450.773.166.3
", + "bbox": [ + 501, + 436, + 890, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Quantitative comparisons on $Nr3D$ and $Sr3D$ dataset. We have highlighted the top-performing three methods in purple.", + "bbox": [ + 498, + 684, + 890, + 712 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "that furnish 2D image priors imbued with elaborate semantics and textures, the proposed $\\mathrm{G}^3$ -LQ method showcases its superiority in performance. This serves as further evidence that our method explicitly captures geometric features, thus facilitating the shape understanding of target object and the vision-text alignment. 2) Our approach, featuring fine-grained language-guided query selection, surpasses all recently published DETR-like models [19, 43]. To explain, the proposed Flan-QS module thoroughly exploits fine-grained language prior and captures global context dependency, which effectively mitigates the issues of grounding ambiguity. 3) Compared with the 3D pre-trained", + "bbox": [ + 496, + 719, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13922", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5a3f526b7fdd65c1142166422caac2bfc9c7f302a77d76abe21feb503bdc3d97.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
IDPAGEFlan-QSPSAUniqueMultiple
0.250.50.250.5
(a)———85.6768.5749.1337.64
(b)✓——86.8970.2650.4638.61
(c)—✓—86.7569.7650.1138.58
(d)——✓86.3269.2049.5838.18
(e)✓—✓87.4571.9551.0840.49
(f)—✓✓87.3170.7550.9040.19
(g)✓✓—87.6672.1650.8740.05
(h)✓✓✓88.0972.7351.4840.80
", + "bbox": [ + 78, + 89, + 470, + 243 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "models [21, 51], $\\mathrm{G}^3$ -LQ achieves SOTA accuracy, because of the semantic-geometric alignment in Poincaré space with the tailored PSA loss, which aids in modeling the complex relationships of vision-text features. 4) The \"multiple\" setting entails a 3D scene where the description distinguishes the target object amidst numerous distractors. Under this setting, we reach a promising performance of $51.14\\%$ and $40.08\\%$ , providing further validation for the efficacy of $\\mathrm{G}^3$ -LQ in addressing the Semantic-Geometric Coupling issue.", + "bbox": [ + 75, + 291, + 468, + 428 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Performance on the Nr3D/Sr3D. The Nr3D/Sr3D task is geared towards locating the target object among all provided ground truth candidate boxes, representing a departure from the ScanRefer dataset. The overall accuracy of the proposed G $^3$ -LQ along with other exceptional methods are reported in Table. 2. We attain peak performance of $58.4\\%$ and $73.1\\%$ on Nr3D and Sr3D. In Nr3D, descriptions exhibit noteworthy intricacy and detail, inducing additional challenges to 3D VG task. However, our method demonstrates a noteworthy superiority over alternative methods that rely on 2D images priors [44] or 3D vision-language pre-training models [51]. In the more challenging \"Hard\" subset, our method notably improves the accuracy by $1.3\\%$ in Nr3D and $2.7\\%$ in Sr3D, which underscores our method is beneficial for distinguishing ambiguous objects by effectively modeling the underlying geometric shape and unraveling the complex relation between vision-text features.", + "bbox": [ + 75, + 428, + 468, + 685 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Ablation Study and Analysis", + "text_level": 1, + "bbox": [ + 76, + 695, + 333, + 710 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of Proposed Components. We develop several alternative designs and undertake experiments to highlight the advantages yielded by various components. We herein train our $\\mathrm{G}^3$ -LQ model on the ScanRefer dataset. (a) outlines the performance demonstrated by EDA [43], while (h) is our proposed method. 1) Comparing (a) and (b), we witness that the PAGE module delivers a remarkable gain over the baseline model. The comparison further accentuates the paramount role of explicit 3D geometric features in modeling object shape and spatial relationship, conferring benefits to 3D visual grounding. 2) The comparative result of (b) and (g) furnishes compelling evidence that the Flan", + "bbox": [ + 75, + 719, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8e975b4519dcf5d66feef533930c04b9678fc00aeb4f282547f7ab258a7c7653.jpg", + "table_caption": [ + "Table 3. Ablation study on the effectiveness of the proposed PAGE, Flan-QS and PSA. Evaluated on the ScanRefer dataset." + ], + "table_footnote": [], + "table_body": "
pipi - pijpi - pijUnique(0.25)Multiple(0.25)
✓——85.6749.13
✓✓—85.8349.19
✓—maxpool85.9849.32
✓—softmax86.0449.37
✓—Afine86.5750.06
✓✓Afine86.8950.46
", + "bbox": [ + 503, + 89, + 890, + 202 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8c58b4aad3741167102286b889d57daa69d2b0895c35c34aa9340b1622cc75f5.jpg", + "table_caption": [ + "Table 4. Ablation study of the geometric priors selection in the proposed PAGE module on the ScanRefer dataset." + ], + "table_footnote": [], + "table_body": "
Query SelectionUnique(~19%)Multiple(~81%)
0.250.50.250.5
Parametric81.4060.1338.0228.05
Top-K(EDA)85.6768.5749.1337.64
Language87.1071.5950.6339.31
Language Graph87.6672.1650.8740.05
", + "bbox": [ + 501, + 256, + 893, + 356 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Ablation study of the query selection. \"Parametric\" represents the selection strategy with random parametric queries. \"Language\" denotes the guidance of sentence-level language.", + "bbox": [ + 498, + 366, + 890, + 409 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "QS module exploiting language prior and context dependency, helps to alleviate the problem of grounding ambiguity and generate precise query proposals. 3) Comparing (h) and (g), it showcases the effectiveness of the proposed PSA loss to capture complex non-linear mappings of point cloud and text embeddings on Poincaré ball. When we integrate all proposed modules, a discernible surge is observed, yielding an impressive overall accuracy of $56.90\\%$ and $45.58\\%$ . Selection of Geometric Priors. The fundamental objective of the PAGE module lies in embedding low-level geometric priors into high-level geometric features, thus the definition of $\\tilde{p}$ emerges as a topic deserving exploration. We undertake ablations with $p_i, p_i - p_{ij}, \\| p_i - p_{ij} \\|$ as illustrative examples, which unveils the inherent relative position, shapes, and distance (see Fig. 3) among 3D objects. As shown in Table. 4, using $\\| p_i - p_{ij} \\|$ or $p_i - p_{ij}$ alone yields inferior performance compared with their gradual integration, the accuracy of which reach $86.89\\%$ (Uniqu) and $50.46\\%$ (Multiple) respectively. Further, when compared with max pooling and softmax operators for features aggregation in $\\mathcal{X}_i$ , our geometric refinement strategy attests to an improved performance by $0.59\\%$ and $0.59\\%$ (Unique). The noteworthy results accentuate the significance of the geometric refinement in the PAGE module, given its potential to capture position-adaptive geometric structures.", + "bbox": [ + 496, + 415, + 890, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Methods of Query Selection. We carry out several experiments to validate the proposed query selection method. We delineate the overall performance on the ScanRefer dataset in Table. 5. 1) We adopt the parametric queries used in MDETR [22]. However, we observe that the random parametric queries in performance decline for its incapability to capture contextual information of the text and 3D scene. 2)", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13923", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/afa2de50b5088c37cdffed8fbad6598f569c7524c74f7b598ac6ae9a0aabd718.jpg", + "image_caption": [ + "(a) Ordinal Numbers", + "this is a chair by the wall. It is the fifth chair from the right wall surface." + ], + "image_footnote": [], + "bbox": [ + 101, + 109, + 267, + 185 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2f43e6854e8502f292069b36f782546485f86068fa07d523221d3719171f87c6.jpg", + "image_caption": [ + "(b) Distance and Size" + ], + "image_footnote": [], + "bbox": [ + 282, + 113, + 475, + 179 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/16400971a747609d2f1074db44477d3041827196f210d758be476fb7920178de.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 109, + 205, + 267, + 287 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ae0582df79992fb5178ef53b9d48cd314cb0a681bdacb7883bbc7e1f1e67db1e.jpg", + "image_caption": [ + "this is a grey chair. its at the end of the shorter table closest to the window.", + "we are looking for the chair on the right. we are not looking for the chair on the left. the chair on the left is closest to the bed." + ], + "image_footnote": [], + "bbox": [ + 313, + 205, + 470, + 291 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f568db9e3a1ffd529047e19383418b3901b3e3172830242fef75bbefa271bd67.jpg", + "image_caption": [ + "it is a black office chair. The black office chair is the fourth office chair on the left side of the table.", + "there is a gray chair. it is the second closest one to a trash can." + ], + "image_footnote": [], + "bbox": [ + 101, + 321, + 292, + 385 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8d92fcdc30592e09814481c6db15b0a894a2c67532829771b5a2f34cc790ece2.jpg", + "image_caption": [ + "the table is the left-most one in the center of the room. The table is dark orange and has four legs." + ], + "image_footnote": [], + "bbox": [ + 297, + 321, + 473, + 386 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3e39a366eda7b48c0f14f033612ddafe87c082bbfb98d25a3cc5dda28c4e36ce.jpg", + "image_caption": [ + "(c) Geometric shape", + "standing semi-circular wooden table with flat surface for holding objects. object can be found in center of image", + "in front of a brown wooden desk are two white chairs. we are looking for the chair on the right. we are not looking for the chair on the left. the chair on the left is closest to the bed." + ], + "image_footnote": [], + "bbox": [ + 485, + 113, + 683, + 181 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4cbc3c1692187fd558fc3991e33f777105f30c9d3a86cf10ab7f55e7e2edc167.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 473, + 212, + 648, + 291 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cc430a091aaf8b0aa4e914b62a54bc497e4ef0cf530ade1159608d2ed4ee9fce.jpg", + "image_caption": [ + "The towel is between the entrance door and the sink. The towel is olive green and wavy.", + "the end table is the closest one to the window. The end table is a round cylinder.", + "Figure 3. Qualitative results of the EDA [43] and our proposed $\\mathrm{G}^3$ -LQ on the ScanRefer dataset. Our $\\mathrm{G}^3$ -LQ manifests excellent grounding performance in terms of ordinal numbers, spatial distance or object size, geometric attribute and complex utterances." + ], + "image_footnote": [], + "bbox": [ + 496, + 316, + 650, + 391 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b17cc1dc1cc08e9c56d214e641f83e7bf89334ce1eb4550da8c2a45d5122ef74.jpg", + "image_caption": [ + "(d) Complex Utterance" + ], + "image_footnote": [], + "bbox": [ + 697, + 109, + 821, + 175 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/aa76ba1a2154fb8d23ab8babd93593659a8921ce769d955f5eb2a025fa5b6c9c.jpg", + "image_caption": [ + "this is the brown arm chair that is at the end of the chalk board, to the left of it there is a small bookshelf, to the right of it and at the table is a blue chair. The brown chair is not under the table it is just next to the table." + ], + "image_footnote": [], + "bbox": [ + 696, + 210, + 828, + 279 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a228d9520a5e582e20fe4624395155954ed62515275a16cc27de00e1472903a4.jpg", + "image_caption": [ + "walking into the room, a large office desk is in the middle of the room. On the right side of the desk, there is a computer monitor on the table, the one on the left side, closest to the door." + ], + "image_footnote": [], + "bbox": [ + 684, + 325, + 862, + 380 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We generate the query proposals guided by the sentence-level features. The performance of the language-guided query selection mechanism stands out prominently, underscoring the crucial role of text features in the query generation. 3) Our Flan-QS module showcases superior performance over the prevailing top- $K$ selection and sentence-level methods. The results give explanation of the efficacy of explicit fine-grained text features. They precisely model the attributes and positions of target objects, thereby facilitating a more accurate generation of query proposals.", + "bbox": [ + 75, + 460, + 468, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Qualitative Comparisons", + "text_level": 1, + "bbox": [ + 76, + 623, + 305, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To offer profound insight into the effectiveness of the proposed $\\mathrm{G}^3$ -LQ, we visualize the grounding results of our method alongside the SOTA EDA [43]. As shown in Fig. 3, the predicted boxes of our $\\mathrm{G}^3$ -LQ method and EDA are visually highlighted in yellow and red, with ground-truth boxes marked in green. The localization ability of EDA is found wanting, encountering challenges in handling the 3D objects described by ordinal numbers, relative distance, and geometric shapes (see Fig. 3(a)-(c)). Conversely, our $\\mathrm{G}^3$ -LQ incorporates intrinsic geometric features and spatial relation, fostering an enriched understanding of both the scene and objects, manifests commendable grounding prowess. Secondly, Fig. 3(d) exemplifies the commendable capabilities of our method in grounding 3D objects narrated by complex utterances. To elucidate, the $\\mathrm{G}^3$ -LQ model with the tailored Flan-QS, excels in harnessing language priors and capturing global context dependencies, generat-", + "bbox": [ + 75, + 643, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ing high-quality semantic-aware query representations.", + "bbox": [ + 500, + 460, + 862, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 483, + 617, + 498 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we analysis the Semantic-Geometric Coupling issue of 3D VG tasks in three aspects: intricate geometric details, complex text descriptions, and semantic-geometric inconsistency. To this end, we propose a $\\mathrm{G}^3$ -LQ framework to explicitly capture the fine-grained geometric and semantic concepts, which includes three crucial novelties. Firstly, a Position Adaptive Geometric Exploring module is designed to capture explicit geometric features. Secondly, with decoupled multiple semantic components, a Fine-grained Language-guided Query Selection module generates object queries densely aligned by fine-grained text features. Finally, a Poincaré Semantic Alignment loss capitalizes on vision-text hierarchy natures and achieves alignment in Poincaré space, encouraging semantic-geometric consistency. Experimental findings showcase the superior performance of our $\\mathrm{G}^3$ -LQ across the ScanRefer and Nr3D/Sr3D benchmarks, establishing a new state-of-the-art standard.", + "bbox": [ + 496, + 503, + 890, + 760 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 768, + 679, + 786 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is supported by the state key development program in 14th Five-Year under Grant No. 2021YFF0602103, 2021YFF0602102, 2021QY1702. Thanks for the research fund under Grant No. 2019GQG0001 from the Institute for Guo Qiang, Tsinghua University. This study is also supported by Tsinghua University Toyota Joint Research Center for AI Technology of Automated Vehicle(TTAD 2023-07).", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13924", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Scene graph parser. https://github.com/vacancy/SceneGraphParser, 2019.5", + "[2] Ahmed Abdelreheem, Ujjwal Upadhyay, Ivan Skorokhodov, Rawan Al Yahya, Jun Chen, and Mohamed Elhoseiny. 3dref-transformer: Fine-grained object identification in real-world scenes using natural language. In Proceedings of the IEEE Winter Conference on Applications of Computer Vision, pages 3941-3950, 2022. 6", + "[3] Panos Achlioptas, Ahmed Abdelreehem, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In Proceedings of the European Conference on Computer Vision, pages 422-440. Springer, 2020. 2", + "[4] Eslam Bakr, Yasmeen Alsaeda, and Mohamed Elhoseiny. Look around and refer: 2d synthetic semantics knowledge distillation for 3d visual grounding. Advances in Neural Information Processing Systems, 35:37146-37158, 2022. 6", + "[5] Daigang Cai, Lichen Zhao, Jing Zhang, Lu Sheng, and Dong Xu. 3djcg: A unified framework for joint dense captioning and visual grounding on 3d point clouds. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 16464-16473, 2022. 2, 6", + "[6] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In Proceedings of the European Conference on Computer Vision, pages 202-221. Springer, 2020. 1, 6", + "[7] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Language conditioned spatial relation reasoning for 3d object grounding. Advances in Neural Information Processing Systems, 35:20522-20535, 2022. 6", + "[8] Jaesung Choe, Chunghyun Park, Francois Rameau, and Jaesik Park. Pointmixer: Mlp-mixer for point cloud understanding. In Proceedings of the European Conference on Computer Vision, pages 620-640. Springer, 2022. 3", + "[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2", + "[10] Jiafei Duan, Samson Yu, Hui Li Tan, Hongyuan Zhu, and Cheston Tan. A survey of embodied ai: From simulators to research tasks. IEEE Transactions on Emerging Topics in Computational Intelligence, 6(2):230-244, 2022. 1", + "[11] Aleksandr Ermolov, Leyla Mirvakhabova, Valentin Khrulkov, Nicu Sebe, and Ivan Oseledets. Hyperbolic vision transformers: Combining improvements in metric learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7409-7419, 2022. 2", + "[12] Mingtao Feng, Zhen Li, Qi Li, Liang Zhang, XiangDong Zhang, Guangming Zhu, Hui Zhang, Yaonan Wang, and Ajmal Mian. Free-form description guided 3d visual graph network for object grounding in point cloud. In Proceedings of the IEEE International Conference on Computer Vision, pages 3722-3731, 2021. 2, 6" + ], + "bbox": [ + 78, + 114, + 467, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Octavian Ganea, Gary Bécigneul, and Thomas Hofmann. Hyperbolic neural networks. Advances in Neural Information Processing Systems, 31, 2018. 2", + "[14] William L Hamilton, Rex Ying, and Jure Leskovec. Inductive representation learning on large graphs. In Advances in Neural Information Processing Systems, pages 1025-1035, 2017. 5", + "[15] Dailan He, Yusheng Zhao, Junyu Luo, Tianrui Hui, Shaofei Huang, Aixi Zhang, and Si Liu. Transrefer3d: Entity-andrelation aware transformer for fine-grained 3d visual grounding. In Proceedings of the 29th ACM International Conference on Multimedia, pages 2344-2352, 2021. 2, 3, 6", + "[16] Pin-Hao Huang, Han-Hung Lee, Hwann-Tzong Chen, and Tyng-Luh Liu. Text-guided graph neural networks for referring 3d instance segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1610-1618, 2021. 2, 6", + "[17] Shijia Huang, Yilun Chen, Jiaya Jia, and Liwei Wang. Multiview transformer for 3d visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15524-15533, 2022. 6", + "[18] Wenlong Huang, Pieter Abbeel, Deepak Pathak, and Igor Mordatch. Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. In International Conference on Machine Learning, pages 9118-9147. PMLR, 2022. 1", + "[19] Ayush Jain, Nikolaos Gkanatsios, Ishita Mediratta, and Katerina Fragkiadaki. Bottom up top down detection transformers for language grounding in images and point clouds. In Proceedings of the European Conference on Computer Vision, pages 417-433. Springer, 2022. 2, 3, 6", + "[20] Haojun Jiang, Yuanze Lin, Dongchen Han, Shiji Song, and Gao Huang. Pseudo-q: Generating pseudo language queries for visual grounding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 15513-15523, 2022. 1", + "[21] Zhao Jin, Munawar Hayat, Yuwei Yang, Yulan Guo, and Yinjie Lei. Context-aware alignment and mutual masking for 3d-language pre-training. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 10984-10994, 2023. 2, 6, 7", + "[22] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE International Conference on Computer Vision, pages 1780-1790, 2021. 3, 7", + "[23] Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In International Conference on Learning Representations, 2017. 5", + "[24] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329, 2022. 5", + "[25] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 3, 5" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13925", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Yongcheng Liu, Bin Fan, Shiming Xiang, and Chunhong Pan. Relation-shape convolutional neural network for point cloud analysis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8895-8904, 2019. 3", + "[27] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 2, 3", + "[28] Yongfei Liu, Bo Wan, Xiaodan Zhu, and Xuming He. Learning cross-modal context graph for visual grounding. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11645-11652, 2020. 1", + "[29] Yongfei Liu, Bo Wan, Lin Ma, and Xuming He. Relation-aware instance refinement for weakly supervised visual grounding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5612-5621, 2021. 1", + "[30] Ze Liu, Zheng Zhang, Yue Cao, Han Hu, and Xin Tong. Group-free 3d object detection via transformers. In Proceedings of the IEEE International Conference on Computer Vision, pages 2949-2958, 2021. 2", + "[31] Junyu Luo, Jiahui Fu, Xianghao Kong, Chen Gao, Haibing Ren, Hao Shen, Huaxia Xia, and Si Liu. 3d-sps: Single-stage 3d visual grounding via referred point progressive selection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 16454-16463, 2022. 1, 2, 3, 6", + "[32] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. In International Conference on Learning Representations, 2021. 3", + "[33] Antonio Montanaro, Diego Valsesia, and Enrico Magli. Rethinking the compositionality of point clouds through regularization in the hyperbolic space. Advances in Neural Information Processing Systems, 35:33741-33753, 2022. 2", + "[34] Maximillian Nickel and Douwe Kiela. Poincaré embeddings for learning hierarchical representations. Advances in Neural Information Processing Systems, 30, 2017. 2", + "[35] Aishwarya Padmakumar, Jesse Thomason, Ayush Shrivastava, Patrick Lange, Anjali Narayan-Chen, Spandana Gella, Robinson Piramuthu, Gokhan Tur, and Dilek Hakkani-Tur. Teach: Task-driven embodied agents that chat. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2017–2025, 2022. 1", + "[36] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in Neural Information Processing Systems, 30, 2017. 3", + "[37] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In proceedings of the IEEE International Conference on Computer Vision, pages 9277-9286, 2019. 2", + "[38] Zheng Qin, Hao Yu, Changjian Wang, Yulan Guo, Yuxing Peng, and Kai Xu. Geometric transformer for fast and robust point cloud registration. In Proceedings of the IEEE Con-" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ference on Computer Vision and Pattern Recognition, pages 11143-11152, 2022. 3", + "[39] Junha Roh, Karthik Desingh, Ali Farhadi, and Dieter Fox. Languagerefer: Spatial-language model for 3d visual grounding. In Conference on Robot Learning, pages 1046-1056. PMLR, 2022. 3, 6", + "[40] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 30, 2017. 2", + "[41] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (TOG), 38(5):1-12, 2019. 3", + "[42] Yingming Wang, Xiangyu Zhang, Tong Yang, and Jian Sun. Anchor detr: Query design for transformer-based detector. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2567-2575, 2022. 5", + "[43] Yanmin Wu, Xinhua Cheng, Renrui Zhang, Zesen Cheng, and Jian Zhang. Eda: Explicit text-decoupling and dense alignment for 3d visual and language learning. arXiv preprint arXiv:2209.14941, 2022. 1, 2, 3, 5, 6, 7, 8", + "[44] Zhengyuan Yang, Songyang Zhang, and Jiebo Luo. Sat: 2d semantics assisted training for 3d visual grounding. In Proceedings of the IEEE International Conference on Computer Vision, pages 1856-1866, 2021. 2, 6, 7", + "[45] Ting Yao, Yehao Li, and Tao Mei. Hgnet: Learning hierarchical geometry from points, edges, and surfaces. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 21846-21855, 2023. 3", + "[46] Xumin Yu, Yongming Rao, Ziyi Wang, Zuyan Liu, Jiwen Lu, and Jie Zhou. Pointr: Diverse point cloud completion with geometry-aware transformers. In Proceedings of the IEEE International Conference on Computer Vision, pages 12498-12507, 2021. 3", + "[47] Zhihao Yuan, Xu Yan, Yinghong Liao, Ruimao Zhang, Sheng Wang, Zhen Li, and Shuguang Cui. Instancerefer: Cooperative holistic understanding for visual grounding on point clouds through instance multi-level contextual referring. In Proceedings of the IEEE International Conference on Computer Vision, pages 1791-1800, 2021. 6", + "[48] Lichen Zhao, Daigang Cai, Lu Sheng, and Dong Xu. 3dvg-transformer: relation modeling for visual grounding on point clouds. In Proceedings of the IEEE International Conference on Computer Vision, pages 2928-2937, 2021. 1, 2, 3, 6", + "[49] Chaoyang Zhu, Yiyi Zhou, Yunhang Shen, Gen Luo, Xingjia Pan, Mingbao Lin, Chao Chen, Liujuan Cao, and Xiaoshuai Sun. Seqtr: A simple yet universal network for visual grounding. In Proceedings of the European Conference on Computer Vision, pages 598-615. Springer, 2022. 1", + "[50] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 5", + "[51] Ziyu Zhu, Xiaojian Ma, Yixin Chen, Zhidong Deng, Siyuan Huang, and Qing Li. 3d-vista: Pre-trained transformer for 3d vision and text alignment. arXiv preprint arXiv:2308.04352, 2023. 2, 6, 7" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13926", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/G^3-LQ_ Marrying Hyperbolic Alignment with Explicit Semantic-Geometric Modeling for 3D Visual Grounding/7e2af7bd-e579-4e10-8895-d1ef28fe0cc7_model.json b/2024/G^3-LQ_ Marrying Hyperbolic Alignment with Explicit Semantic-Geometric Modeling for 3D Visual Grounding/7e2af7bd-e579-4e10-8895-d1ef28fe0cc7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4edca1833a24ad731db9dff1d305dbd731df616f --- /dev/null +++ b/2024/G^3-LQ_ Marrying Hyperbolic Alignment with Explicit Semantic-Geometric Modeling for 3D Visual Grounding/7e2af7bd-e579-4e10-8895-d1ef28fe0cc7_model.json @@ -0,0 +1,2420 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.105, + 0.13, + 0.867, + 0.177 + ], + "angle": 0, + "content": "\\(\\mathbf{G}^{3}\\)-LQ: Marrying Hyperbolic Alignment with Explicit Semantic-Geometric Modeling for 3D Visual Grounding" + }, + { + "type": "text", + "bbox": [ + 0.302, + 0.203, + 0.669, + 0.222 + ], + "angle": 0, + "content": "Yuan Wang\\(^{1,2}\\) Yali Li\\(^{1,2}\\) Shengjin Wang\\(^{1,2\\ddagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.222, + 0.753, + 0.24 + ], + "angle": 0, + "content": "\\(^{1}\\)Department of Electronic Engineering, Tsinghua University, China" + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.239, + 0.858, + 0.277 + ], + "angle": 0, + "content": "\\(^{2}\\)Beijing National Research Center for Information Science and Technology (BNRist), China wy23@mails.tsinghua.edu.cn {liyali13, wgsgj} @tsinghua.edu.cn \\(\\ddagger\\)Corresponding Author" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.313, + 0.88, + 0.479 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.491, + 0.893, + 0.521 + ], + "angle": 0, + "content": "Figure 1. Illustration of the proposed G\\(^3\\)-LQ method on the ScanRefer benchmark. The versatile scene-level and object-level understanding properties empower G\\(^3\\)-LQ to comprehend spatial relationship, relative distance, geometric attribute as well as intricate references." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.533, + 0.314, + 0.548 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.554, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Grounding referred objects in 3D scenes is a burgeoning vision-language task pivotal for propelling Embodied AI, as it endeavors to connect the 3D physical world with free-form descriptions. Compared to the 2D counterparts, challenges posed by the variability of 3D visual grounding remain relatively unsolved in existing studies: 1) the underlying geometric and complex spatial relationships in 3D scene. 2) the inherent complexity of 3D grounded language. 3) the inconsistencies between text and geometric features. To tackle these issues, we propose \\( G^3 \\)-LQ, a DEtection Transformer-based model tailored for 3D visual grounding task. \\( G^3 \\)-LQ explicitly models Geometric-aware visual representations and Generates fine-Grained Language-guided object Queries in an overarching framework, which comprises two dedicated modules. Specifically, the Position Adaptive Geometric Exploring (PAGE) unearths underlying information of 3D objects in the geometric details and spatial relationships perspectives. The Fine-grained Language-guided Query Selection (Flan-QS) delves into syntactic structure of texts and generates object queries that exhibit higher relevance towards fine-grained text features. Finally, a pioneering Poincaré Semantic Alignment (PSA) loss establishes semantic-geometry consistencies by mod" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.534, + 0.893, + 0.595 + ], + "angle": 0, + "content": "eling non-linear vision-text feature mappings and aligning them on a hyperbolic prototype—Poincaré ball. Extensive experiments verify the superiority of our \\( G^3 \\)-\\( LQ \\) method, trumping the state-of-the-arts by a considerable margin." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.604, + 0.633, + 0.62 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.893, + 0.735 + ], + "angle": 0, + "content": "Grounding 3D physical realm with natural languages has propelled to the forefront in advancing Embodied AI [10, 18, 35], an ability that empowers agents to comprehend human instructions in real-world contexts. 3D Visual Grounding (3D VG) [6, 31, 43, 48] has garnered substantial attention as a crucial cross-modal 3D perception task in defacto applications, e.g., assistive robots, AR/VR, and metaverse." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Compared with extensive explorations of 2D counterpart [20, 28, 29, 49], 3D VG poses prominent challenge known as \"Semantic-Geometric Coupling\". As shown in Fig. 1, we offer distinctive glimpse into it: 1) Intricate geometric details and spatial relations. 3D scenes encounter multiple objects with intricate layout, multi-level occlusion, and multifaceted spatial relation. 2) Heightened flexibility and complexity of descriptions. Free-form utterances with complex syntax convey precise spatial information, intricate references, and detailed relational terms. 3) Inconsistency of the semantic-geometric features. A single semantic concep" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13917" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "tion yields multiple potential correspondences within point cloud or exhibit diverse geometry in varied contexts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.124, + 0.473, + 0.502 + ], + "angle": 0, + "content": "To ameliorate these issues, some pioneering methods have been proposed. They have largely focused on enhancing the representation capacity of 3D point clouds with 2D image priors [44]; capturing spatial relationships of 3D objects with powerful transformer [5, 15, 48] or graph convolution network [3, 12]; generating discriminative object proposals through instance segmentation [16] or linguistic guidance [31]. Despite commendable advancements be achieved, an equally important yet underexplored problem is explicit semantic-geometric modeling. Prior studies exhibit a significant dearth in explicitly modeling the intrinsic geometric attributes and spatial layouts of 3D objects, accentuating the dilemma to associate text descriptions with grounding objects. Moreover, object queries contain rich instance characteristics. But the object proposal [15, 48] and query generation [19, 31, 43] in current DETR-like models overlook explicit fine-grained interaction between text and object features, which in turn limit the model's capacity to refer potential language clues, e.g., appearances, shapes, or textures. Finally, major 3D VG systems [19, 21, 43, 51] strive to establish the alignment between point cloud and semantic concepts in canonical Euclidean space. However, their efficacy falters in unraveling the complex non-linear relationships between multi-modal features and upholding the intrinsic hierarchy natures in 3D scenes and texts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.504, + 0.47, + 0.565 + ], + "angle": 0, + "content": "In this paper, our primary goal endeavors to explicitly delve into fine-grained representations of geometric and semantic concepts, then aligns two modalities in hyperbolic space. We thus devise an overarching G\\(^3\\)-LQ framework:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.567, + 0.473, + 0.764 + ], + "angle": 0, + "content": "1) For geometric feature modeling, we firstly propose a Proxy Adaptive Geometric Exploring (PAGE) module. PAGE is composed of the Point-Proxy Geometric Extraction and Proxy Adaptive Geometric Refinement component. The former explicitly capture the local geometric attributes (e.g., distance and orientation) in a 3D local proxy, while the latter adaptively refine the geometric structure according to the position of point proxy. The geometric features are fed into the geometric encoder for high-level embedding and further achieve the interaction with text features for cross-modality enhancement. In this fashion, the PAGE module facilitates a seamless alignment of text descriptions with intrinsic geometric properties of 3D objects." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.473, + 0.903 + ], + "angle": 0, + "content": "2) For complex utterance understanding, inspired by the dependency parsing, we aim to explore context dependency and syntactic structure of texts for guiding query selection. To this end, we propose a Fine-grained Language-guided Query Selection (Flan-QS) module, comprising the Language Scene Graph and Fine-grained Query Selection components. We decouple the free-form utterances to tree-like semantic components with directed dependency relationships as edges. Then, we establish and update the lan" + }, + { + "type": "list", + "bbox": [ + 0.076, + 0.567, + 0.473, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "guage scene graph to represent the context-aware entry features via node/edge embedding and graph context learning. Finally, we employ the constructed language scene graph to steer the generation of subsequent object candidates that are densely-aligned with fine-grained linguistic attributes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.168, + 0.893, + 0.394 + ], + "angle": 0, + "content": "3) For semantic-geometric consistency, we embed vision and text features to a hyperbolic prototype—Poincaré ball. The emerging appreciation [11, 13, 34] suggest that the hyperbolic space offers overwhelming superiority to handle data correlations. 3D scene (free-form text) exhibits an inherent hierarchical nature ranging from scene, object, part (from sentence, phrase, word). The hyperbolic space with negative curvature is the only space that can successfully embed the tree-shaped hierarchy [33]. Moreover, the compact Poincaré ball [13, 34] holds suitable geometric properties, enhancing its capacity to model non-linear mappings of vision and text features. Therefore, we design a Poincaré Semantic Alignment loss (PSA) to probe the complex mappings between vision and text features, facilitating the semantics-geometry consistency." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.394, + 0.768, + 0.408 + ], + "angle": 0, + "content": "Our chief contributions are threefold:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.409, + 0.892, + 0.484 + ], + "angle": 0, + "content": "- We roundly analysis the Semantic-Geometric Coupling of 3D VG task. Towards this issue, we propose a unified \\(\\mathbf{G}^3\\)-LQ framework to model intrinsic geometric features (PAGE module) and explicitly generate fine-grained object queries for 3D visual grounding (Flan-QS module)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.484, + 0.892, + 0.545 + ], + "angle": 0, + "content": "- We propose a novel PSA loss to capture the non-linear relationship and underlying hierarchy of the vision-text features in the hyperbolic space, which further paves the way for semantics-geometry consistency learning." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.545, + 0.892, + 0.606 + ], + "angle": 0, + "content": "- A battery of experiments on ScanRefer and Nr3d/Sr3d datasets demonstrate the effectiveness of our G\\(^{3}\\)-LQ method. G\\(^{3}\\)-LQ has proven its mettle, outperforming all existing SOTA methods by a considerable margin." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.409, + 0.892, + 0.606 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.611, + 0.651, + 0.627 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.636, + 0.702, + 0.652 + ], + "angle": 0, + "content": "2.1.3D Visual Grounding" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.895, + 0.902 + ], + "angle": 0, + "content": "3D Visual Grounding has been an area of intense investigation that dedicates to pinpointing the target object in a 3D scene based on the text description. Prevailing methods are typically two-stage, which conforms to the detection-then-matching paradigm. Firstly, the 3D object detectors [30, 37, 48] are harnessed to yield candidate 3D object proposals, while the text features are encoded by language models [9, 27]. Secondly, these advanced methods perform a pivotal endeavor to align the vision and text features for target objects grounding. Among these, TGNN [16] utilizes the graph neural network to deduce spatial relationships among 3D object proposals, which are further enriched by text features. FFL-3DOG [12] captures the intramodal and cross-modal relationships of text descriptions and 3D point cloud via scene graphs interaction. Recent endeavors embrace the formidable transformers [40] to model" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13918" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.47, + 0.182 + ], + "angle": 0, + "content": "the relationship of proposals and text features. 3DVG [48] and TransRefer3D [15] achieve exceptional performance by establishing self-attention and cross-attention protocols for context perception. LanguageRefer [39] seamlessly converts the cross-modal task into a unified language modeling challenge, underpinned by predicted object labels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.184, + 0.471, + 0.395 + ], + "angle": 0, + "content": "However, two-stage methods encounter a noteworthy impediment: the detection stage neglects to leverage language contexts to prioritize the objects that are essential to the referring task. An alternative approach to the problem is one-stage methods [19, 31, 43], where extracted vision features are directly fused with the language features for object grounding. For instance, 3D-SPS [31] leverages text features to guide visual keypoints selection, enabling the progressive object grounding. BUTD-DETR [19] pioneers a DETR-like model, which fuses the vision-text features via a co-attention paradigm and then decodes objects in the utterance from the contextualized features. Building upon this, EDA [43] proposes a text-decoupled strategy and performs dense alignment of 3D objects and associated texts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.396, + 0.471, + 0.624 + ], + "angle": 0, + "content": "Despite one-stage methods unfold prowess performance, inherent limitations have been yet underexplored: 1) They overlook the intrinsic geometric attributes (distance, orientation, and shape) in 3D point clouds, leading to ambiguous visual representations. Conversely, our proposed PAGE module explicitly captures geometric features to avoid ambiguity. 2) The queries generation lack the fine-grained language guidance, limiting the model's capacity to infer potential language clues. Our tailored Flan-QS module explores language prior and context dependency to generate precise proposals. 3) They establish the vision-text alignment in Euclidean space, incapable of modeling the nonlinear relation of vision and text features. Our method embeds multimodal features on the Poincaré ball then leverage the PSA loss to foster the semantic-geometric alignment." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.635, + 0.431, + 0.65 + ], + "angle": 0, + "content": "2.2. Geometric Extraction in 3D Point Clouds" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.47, + 0.902 + ], + "angle": 0, + "content": "The explicit modeling of geometric structures has recently garnered considerable attention in the realms of 3D point cloud, e.g., point cloud understanding [8, 26, 32, 41], point cloud completion [46], point cloud registration [38], and 3D object detection [45]. DGCNN [41] proposes an EdgeConv operator, by which the point features are aggregated through graph convolution to capture intricate local geometric structures. RSCNN [26] explicitly encodes geometric relationship of points with relation-aware convolution. PointMixer [8] embeds geometric relationships of 3D point in a MLP-Mixer's streamline. PoinTr [46] devises a geometry-aware transformer to learn structural knowledge and local geometric relationships for point cloud completion. HGNet [45] delves into localized shape information to generate the proposals with a hierarchical graph model for 3D object detection. Our cutting-edge \\(\\mathrm{G}^3\\)-LQ incorporates" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "the geometric attributes of point cloud into the 3D VG task, thereby enhancing the discriminability of vision features." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.128, + 0.593, + 0.143 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.152, + 0.611, + 0.166 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.174, + 0.892, + 0.251 + ], + "angle": 0, + "content": "Provided a point cloud \\(\\mathcal{P} = \\{\\pmb{p}_i\\}_{i=1}^N \\in \\mathbb{R}^{N \\times (3 + F)}\\) with \\(F\\)-dim auxiliary features (e.g., RGB, normals, or multi-view features) of \\(N\\) points, and a free-form text \\(\\mathcal{T}\\) with \\(L\\) words, the primary goal of 3D VG is to establish a mapping \\(\\mathcal{M}\\) that links \\(\\mathcal{P}\\) and \\(\\mathcal{T}\\) to the target object \\(\\pmb{o}\\), i.e., \\(\\mathcal{M}(\\mathcal{P}, \\mathcal{T}) \\to \\pmb{o}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.251, + 0.892, + 0.401 + ], + "angle": 0, + "content": "Fig. 2 gives the outline of our proposed \\(\\mathrm{G}^3\\)-LQ method. Firstly, we leverage the well-accepted PointNet++ [36] tockenize the point cloud into vision token \\(\\mathbf{V}\\in \\mathbb{R}^{n\\times d}\\). The text is encoded by the pre-trained RoBERTa [27] and yields the vanilla text token \\(\\mathbf{T}\\in \\mathbb{R}^{t\\times d}\\). The vision token \\(\\mathbf{V}\\) is fed into the PAGE module to capture underlying geometric features \\(\\mathbf{V}_g\\in \\mathbb{R}^{n\\times d}\\), which adjusts according to the point positions, thereby furnishing precise geometric clues. Further, the geometric visual encoder embeds \\(\\mathbf{V}_g\\) via self-attention and interact with text features \\(\\mathbf{T}\\) via cross-attention." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.402, + 0.892, + 0.508 + ], + "angle": 0, + "content": "Finally, we build the language scene graph \\(\\mathcal{G}\\) and obtain the updated fine-grained text features, which will explicitly guide the queries selection. Like the object queries in most DETR-like models [22, 25, 43], the selected queries \\(\\mathbf{Q} \\in \\mathbb{R}^{K \\times d}\\) are fed into a cross-modal decoder to probe desired features and update themselves. The decoded queries \\(\\mathbf{Q}' \\in \\mathbb{R}^{K \\times l}\\) are fed into an MLP to predict object boxes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.516, + 0.826, + 0.533 + ], + "angle": 0, + "content": "3.2. Geometric Exploring and Embedding" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.54, + 0.892, + 0.584 + ], + "angle": 0, + "content": "In this section, we propose a PAGE module to explicitly capture geometric properties of 3D scenes, which serves to encode holistic visual representations for 3D VG tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.585, + 0.892, + 0.675 + ], + "angle": 0, + "content": "Point-Proxy Geometric Extraction. To facilitate local geometric features modeling, we firstly instantiate point proxies predicated on the proximity of the points, thereby serving as the local regions of the point cloud. As illustrated on Fig 2, we search \\( k \\) neighbors of \\( p_i \\in \\mathcal{P} \\) as a point proxy with the KNN algorithm, denoted as: \\( \\mathcal{X}_i = k\\mathrm{NN}(\\mathcal{P},p_i) \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.676, + 0.892, + 0.813 + ], + "angle": 0, + "content": "The point proxy \\(\\mathcal{X}_i\\) denotes index set with \\(k\\) closest points of \\(\\pmb{p}_i\\) and \\(\\pmb{v}_{m_i}:\\{\\pmb{v}_{m_{ij}}|j\\in \\mathcal{X}_i\\} \\in \\mathbb{R}^{k\\times d}\\) is the corresponding features. With the Point-Proxy Geometric Extraction component, we capture the local structure by learning the geometric topology features within the point proxy \\(\\mathcal{X}_i\\). In this fashion, it is effective to achieve an inductive local representations that integrate explicit reasoning pertaining to geometric structures and objects spatial layout. We primarily compute pre-defined geometric prior \\(\\widetilde{\\pmb{p}}_{ij}\\in \\mathbb{R}^7\\):" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.818, + 0.892, + 0.838 + ], + "angle": 0, + "content": "\\[\n\\widetilde {\\boldsymbol {p}} _ {i j} = \\operatorname {c o n c a t} \\left(\\boldsymbol {p} _ {i} - \\boldsymbol {p} _ {i j}, \\boldsymbol {p} _ {i}, \\left\\| \\boldsymbol {p} _ {i} - \\boldsymbol {p} _ {i j} \\right\\|\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.892, + 0.902 + ], + "angle": 0, + "content": "\\(\\widetilde{\\pmb{p}}_{ij}\\) embeds relative orientation and distance of different objects. Further, the low-level geometric features are encoded to high-level features \\(\\pmb{h}_{ij} \\in \\mathbb{R}^d\\) with a linear layer \\(\\phi[\\cdot]\\) to obtain abstract geometric conception and excavate" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13919" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.095, + 0.396, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.43, + 0.244, + 0.445 + ], + "angle": 0, + "content": "Text: \\(\\mathcal{T} = \\{t_i\\}_{i = 1}^L\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.256, + 0.43, + 0.386, + 0.445 + ], + "angle": 0, + "content": "Point: \\(\\mathcal{P} = \\{\\pmb {p}_i\\}_{i = 1}^N\\)" + }, + { + "type": "image", + "bbox": [ + 0.398, + 0.096, + 0.868, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.466, + 0.893, + 0.511 + ], + "angle": 0, + "content": "Figure 2. Illustration of (a) our proposed \\(\\mathbf{G}^3\\)-LQ framework. The red dashed frames denote our novel modules. (b) PAGE moulder to model the underlying geometric structure of 3D point cloud and (c) the Geometric Visual Encoder further to embed the geometric topology features. (d) Flan-QS module to capture fine-grained language context with (e) Language Scene Graph for guiding query selection." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.521, + 0.462, + 0.536 + ], + "angle": 0, + "content": "discriminative geometric structures, which is described as:" + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.541, + 0.469, + 0.561 + ], + "angle": 0, + "content": "\\[\nh _ {i j} = \\widetilde {\\boldsymbol {p}} _ {i j} + \\phi [ \\widetilde {\\boldsymbol {p}} _ {i j} ], \\forall j \\in \\mathcal {X} _ {i} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.572, + 0.47, + 0.693 + ], + "angle": 0, + "content": "Proxy-Adaptive Geometric Refinement. The presence of diverse geometric structures across distinct point proxies necessitate anisotropic extractors. Further, we flesh out this intuition and engineer a Proxy-Adaptive Geometric Refinement component, poised to model refined geometric features effectively of diverse point proxies. Given the geometric prior \\(\\{h_{ij}|j\\in \\mathcal{X}_i\\} \\in \\mathbb{R}^{k\\times d}\\) in \\(\\mathcal{X}_i\\) we establish local geometric refinement by the following streamline:" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.697, + 0.469, + 0.729 + ], + "angle": 0, + "content": "\\[\n\\left\\{g _ {i j} \\right\\} = \\alpha \\odot \\frac {\\left\\{h _ {i j} \\right\\} - h _ {i}}{\\Omega + \\varepsilon} + \\beta \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.732, + 0.47, + 0.809 + ], + "angle": 0, + "content": "where \\(\\varepsilon\\) is a small constant for numerical stability and \\(\\alpha, \\beta\\) are learnable parameters for features refinement. \\(\\Omega\\) is the standard deviation across all point proxies and channels. \\(h_i\\) is the center of \\(\\mathcal{X}_i\\). \\(g_i = \\{g_{ij} | j \\in \\mathcal{X}_i\\} \\in \\mathbb{R}^{k \\times d}\\) represents the anisotropic geometric structures in point proxy \\(\\mathcal{X}_i\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.816, + 0.469, + 0.846 + ], + "angle": 0, + "content": "Finally, we opt for a softmax aggregation strategy to densely relate point features interaction in each \\(\\mathcal{X}_i\\):" + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.849, + 0.404, + 0.868 + ], + "angle": 0, + "content": "\\[\n\\widetilde {v} _ {g _ {i}} = \\operatorname {L i n e a r} \\left(\\operatorname {c o n c a t} \\left(v _ {m _ {i}}, g _ {i}\\right)\\right) \\in \\mathbb {R} ^ {k \\times d}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.871, + 0.469, + 0.904 + ], + "angle": 0, + "content": "\\[\nv _ {g _ {i}} = \\sum_ {j \\in \\mathcal {X} _ {i}} \\operatorname {s o f t m a x} \\left(\\widetilde {v} _ {g _ {i j}}\\right) \\cdot \\widetilde {v} _ {g _ {i j}} \\in \\mathbb {R} ^ {d}, \\forall j \\in \\mathcal {X} _ {i}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.521, + 0.892, + 0.552 + ], + "angle": 0, + "content": "where Linear \\((\\cdot)\\) denotes the linear projection, \\(\\mathbf{V}_g = \\{v_{g_i}\\}_{i=1}^n\\) is the output token with geometric representations." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.551, + 0.892, + 0.701 + ], + "angle": 0, + "content": "Geometric Visual Encoder. Geometric perception enables to strengthen shape features of 3D objects and demonstrate enhanced viewpoint sensitivity, which bestows flexibility in capturing spatial relation and improves precise 3D objects grounding. Hence, we embed geometric topology features \\(\\mathbf{V}_g\\) via a tailored Geometric Encoder. As shown in Fig. 2, the core of Geometric Encoder is Geometric self-attention layer. Given the high-level vision features \\(\\mathbf{V}_m\\in \\mathbb{R}^{n\\times d}\\) and the explicit geometric features \\(\\mathbf{V}_g\\), the output \\(\\mathbf{V}_{emb}\\in \\mathbb{R}^{n\\times d}\\) is a weighted sum of projected features:" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.702, + 0.892, + 0.742 + ], + "angle": 0, + "content": "\\[\nv _ {i} ^ {e m b} = \\sum_ {j = 1} ^ {n} a _ {i j} \\left(v _ {m _ {j}} \\mathbf {W} ^ {V}\\right) \\in \\mathbb {R} ^ {d} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.743, + 0.891, + 0.772 + ], + "angle": 0, + "content": "The attention weight \\( a_{ij} \\) is computed by a row-wise softmax, which can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.776, + 0.892, + 0.816 + ], + "angle": 0, + "content": "\\[\na _ {i j} = \\sigma \\left[ \\frac {\\left(v _ {g _ {i}} \\mathbf {W} _ {g} ^ {Q} + v _ {m _ {i}} \\mathbf {W} _ {m} ^ {Q}\\right) \\left(v _ {g _ {j}} \\mathbf {W} _ {g} ^ {K} + v _ {m _ {j}} \\mathbf {W} _ {m} ^ {K}\\right) ^ {T}}{\\sqrt {d}} \\right] \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.819, + 0.892, + 0.869 + ], + "angle": 0, + "content": "\\(\\mathbf{W}_g^Q,\\mathbf{W}_g^K\\) are geometric projection matrices of query and key, \\(\\mathbf{W}_m^Q,\\mathbf{W}_m^K\\) and \\(\\mathbf{W}_m^V\\) are semantic projection matrices of query, key and value. \\(\\sigma\\) is the row-wise softmax operator." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.87, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Further, the geometric-enhanced features \\(\\mathbf{V}_{emb}\\) pass through the cross-attention layer to interact with the text" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.99, + 0.518, + 1.0 + ], + "angle": 0, + "content": "13920" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.138 + ], + "angle": 0, + "content": "features and box features (used in two stage), for obtaining the cross-modal features \\(\\mathbf{V}' \\in \\mathbb{R}^{n \\times d}\\) and \\(\\mathbf{T}' \\in \\mathbb{R}^{t \\times d}\\). Finally, \\(\\mathbf{V}'\\) is linearly projected as the candidate query \\(\\mathbf{Q}_c\\)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.147, + 0.236, + 0.162 + ], + "angle": 0, + "content": "3.3. Query Selection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.171, + 0.47, + 0.568 + ], + "angle": 0, + "content": "In the DETR-like model [24, 42, 50], object queries play essential roles to determine potential regions of targets, which are equally prominent for 3G VG systems, decoded into a box center and size that conform to vision tokens. Assisted by global text features, GroundingDINO [25] employs a language-guided query selection protocol to generate query proposals. However, such text features are relatively coarse-grained, leading to inaccurate query generation. Actually, free-form texts harbor intricate syntactic structure and possess a context-dependency nature. It further serves as inspiration for us to generate precise vision queries via improving fine-grained semantic understanding. Language Scene Graph. We intend to delve into fine-grained text information, which is defined as decoupled semantic components (e.g., object, attributes, and relations) and syntactic structures. To this end, we perform a syntactic dependency parsing to construct language graph \\(\\mathcal{G} = \\{\\mathcal{O},\\mathcal{R}\\}\\) by the off-the-shelf Scene Graph Parser [1], which serves as a reasoning inductive bias of 3D VG tasks. \\(\\mathcal{O} = \\{\\pmb{o}_i\\}\\) and \\(\\mathcal{R} = \\{\\pmb{r}_{ij}\\}\\) are nodes and edges set. Each object \\(o_i\\) is denoted as a phrase with a set of attributes (appearance, shape, texture). Harnessing the final layer of the pre-trained RoBERTa model, we then acquire the object phrase embedding \\(\\pmb{x}_{o_i}\\in \\mathbb{R}^d\\) as the graph node \\(o_i\\in \\mathcal{O}\\). In a similar vein, the relation phrases are also encoded into the context-aware embeddings \\(\\pmb{x}_{r_{ij}}\\in \\mathbb{R}^d\\) akin to the graph edges \\(\\pmb{r}_{ij}\\in \\mathcal{R}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.567, + 0.47, + 0.659 + ], + "angle": 0, + "content": "Inspired by the message passing mechanism [14, 23], we proceed to enhance relation and object embeddings via language graph convolution. Firstly, to enrich the contextual representations of relation embedding in connection with its affiliated nodes, we aggregate messages and update its features with the subject \\( \\pmb{x}_{o_i} \\) and object node \\( \\pmb{x}_{o_j} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.662, + 0.469, + 0.681 + ], + "angle": 0, + "content": "\\[\n\\widetilde {\\boldsymbol {x}} _ {r _ {i j}} = \\boldsymbol {x} _ {r _ {i j}} + F _ {r} \\left(\\left[ \\boldsymbol {x} _ {o _ {i}}; \\boldsymbol {x} _ {r _ {i j}}; \\boldsymbol {x} _ {o _ {j}} \\right]\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.685, + 0.469, + 0.745 + ], + "angle": 0, + "content": "where \\(F_{r}\\) is a linear projection layer and \\(\\widetilde{\\pmb{x}}_{r_{ij}}\\) is the context-aware relation embedding. We then update the phrase embedding \\(\\pmb{x}_{o_i}\\) by passing messages from all connected nodes \\(\\Gamma (i)\\) and the edges with a tailored attention mechanism:" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.751, + 0.469, + 0.784 + ], + "angle": 0, + "content": "\\[\n\\widetilde {\\boldsymbol {x}} _ {o _ {i}} = \\boldsymbol {x} _ {o _ {j}} + \\sum_ {j \\in \\Gamma (i)} \\boldsymbol {w} _ {o _ {i j}} F _ {o} \\left(\\left[ \\boldsymbol {x} _ {o _ {j}}; \\boldsymbol {x} _ {r _ {i j}} \\right]\\right) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.794, + 0.469, + 0.825 + ], + "angle": 0, + "content": "where \\(F_{o}\\) is another linear layer and \\(\\widetilde{\\pmb{x}}_{o_i}\\) denotes context-aware object features. \\(\\pmb{w}_{o_{ij}}\\) is the tailored attention weight:" + }, + { + "type": "equation", + "bbox": [ + 0.087, + 0.84, + 0.469, + 0.866 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {w} _ {o _ {i j}} = \\operatorname {s o f t m a x} \\left[ F _ {o} \\left(\\left[ \\boldsymbol {x} _ {o _ {i}}; \\widetilde {\\boldsymbol {x}} _ {r _ {i j}} \\right]\\right) ^ {T} F _ {o} \\left(\\left[ \\boldsymbol {x} _ {o _ {j}}; \\widetilde {\\boldsymbol {x}} _ {r _ {i j}} \\right]\\right) \\right] \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "The attention protocol manifests the capability to discern crucial contextual features from language graph, serving as" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.093, + 0.78, + 0.107 + ], + "angle": 0, + "content": "potent language priors for query selection." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.107, + 0.892, + 0.273 + ], + "angle": 0, + "content": "Fine-grained Query Selection. We input the candidate query features \\(\\mathbf{Q}_c\\in \\mathbb{R}^{n\\times d}\\) from the cross-modal encoding layer into the Query Embedding component, which consists of several stacked Fully-connected (FC) layers and BN layers. Further, we leverage the fine-grained text features \\(T_{f}\\) to guide the query selection through a cross-attention fusion layer. The resulting query tokens are fed into an MLP layer for confidence score prediction. Through selecting the top-\\(K\\) highest scoring tokens for the following decoder, we derive the updated object queries \\(\\mathbf{Q}_s\\in \\mathbb{R}^{K\\times d}\\), which are utilized to predict a box center, height and width." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.279, + 0.797, + 0.294 + ], + "angle": 0, + "content": "3.4. Poincaré Semantic Alignment loss" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.302, + 0.892, + 0.438 + ], + "angle": 0, + "content": "We linearly embed the queries \\(\\mathbf{Q}_s\\) to the predicted object features \\(\\pmb{q} \\in \\mathbb{R}^{K \\times 64}\\), which aligns with the embedded text features \\(\\pmb{t} \\in \\mathbb{R}^{t \\times 64}\\). Noteworthy, point clouds (spanning from scenes, objects to parts) and texts (ranging from sentences, phrases to words) encompass inherent hierarchy. Further, granularity disparities of geometric and semantic representations complicate the intricate correlations. However, it is the Achilles' heel to model hierarchical structures and complex relationships in canonical Euclidean space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.439, + 0.892, + 0.559 + ], + "angle": 0, + "content": "To tackle this issue, we provide the insight to map vision-text features to a special hyperbolic space—Poincaré ball, for its powerful natural aptitude in modeling tree-shaped features. Based on this, we carefully devise a PSA loss to pave the way for semantic-geometric consistency. Benefiting from exponential volume expansion of hyperbolic embedding, PSA explicitly models complex semantic-geometric relationship for capturing hierarchical structures." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.559, + 0.892, + 0.62 + ], + "angle": 0, + "content": "Since the Poincaré ball is a Riemannian manifold, we project the vision \\( \\pmb{q} \\) and text features \\( \\pmb{t} \\) to the Poincaré ball \\( \\mathbb{D}_c^d \\) with the curvature \\( c \\) via an exponential mapping function \\( \\exp_{\\pmb{x}}^{c}: \\mathbb{R}^{d} \\to \\mathbb{D}_{c}^{d} \\), which can be defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.636, + 0.892, + 0.667 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {q} ^ {\\mathrm {c}} = \\exp_ {\\boldsymbol {x}} ^ {\\mathrm {c}} (\\boldsymbol {q}) = \\boldsymbol {x} \\oplus_ {c} \\left(\\tanh \\left(\\sqrt {c} \\frac {\\lambda_ {\\boldsymbol {x}} ^ {\\mathrm {c}} \\| \\boldsymbol {q} \\|}{2}\\right) \\frac {\\boldsymbol {q}}{\\sqrt {c} \\| \\boldsymbol {q} \\|}\\right) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.671, + 0.892, + 0.748 + ], + "angle": 0, + "content": "where \\(\\pmb{x}\\) is a fixed base point to make formulas less cumbersome and \\(\\lambda_{\\pmb{x}}^{c} = 2 / (1 - c\\| \\pmb{x}\\|^{2})\\). \\(t^c\\) is defined in a similar approach. \\(\\oplus_{c}\\) denotes the addition operator in the Mobius gyrovector space. The geodesic distance between the a vision-text feature pair \\((\\pmb{q}_i^c,\\pmb{t}_j^c)\\) on the Poincaré ball \\(\\mathbb{D}_c^d\\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.535, + 0.75, + 0.891, + 0.78 + ], + "angle": 0, + "content": "\\[\nd _ {c} \\left(\\boldsymbol {q} _ {i} ^ {c}, \\boldsymbol {t} _ {j} ^ {c}\\right) = \\frac {2}{\\sqrt {c}} \\operatorname {a r c t a n h} \\left(\\sqrt {c} \\| - \\boldsymbol {q} _ {i} ^ {c} \\oplus_ {c} \\boldsymbol {t} _ {j} ^ {c} \\|\\right) \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.782, + 0.892, + 0.827 + ], + "angle": 0, + "content": "Inspired by [43], we term the PSA loss to densely align the fine-grained vision-text features via contrastive learning on Poincaré ball. The query loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.51, + 0.834, + 0.891, + 0.881 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {p s a}} ^ {q} = \\sum_ {i = 1} ^ {K} \\frac {- 1}{\\left| \\mathcal {T} _ {i} ^ {+} \\right|} \\sum_ {\\boldsymbol {t} _ {i} ^ {c} \\in \\mathcal {T} _ {i} ^ {+}} \\log \\left[ \\right. \\frac {\\exp \\left[ \\boldsymbol {w} _ {+} \\left(- d _ {c} \\left(\\boldsymbol {q} _ {i} ^ {c} , \\boldsymbol {t} _ {i} ^ {c}\\right) / \\tau\\right)\\right]}{\\sum_ {j = 1} ^ {l} \\exp \\left[ \\boldsymbol {w} _ {-} \\left(- d _ {c} \\left(\\boldsymbol {q} _ {i} ^ {c} , \\boldsymbol {t} _ {j} ^ {c}\\right) / \\tau\\right)} \\right] \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\( k \\) and \\( l \\) are the number of object tokens and text to-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13921" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.088, + 0.895, + 0.372 + ], + "angle": 0, + "content": "
MethodVenueInputUnique (~19%)Multiple (~81%)Overall
0.250.50.250.50.250.5
ScanRefer [6]ECCV'203D+2D76.3353.5132.7321.1141.1927.40
TGNN [16]AAAI'213D68.6156.8029.8423.1837.3729.70
InstanceRefer [47]ICCV'213D77.4566.8331.2724.7740.2332.93
SAT [44]ICCV'213D+2D73.2150.8337.6425.1644.5430.14
FFL-3DOG [12]ICCV'213D78.8067.9435.1925.741.3334.01
3DVG [48]ICCV'213D+2D81.9360.6439.328.4247.5734.67
3D-SPS [31]CVPR'223D+2D84.1266.7240.3229.8248.8236.98
BUTD-DETR [19]ECCV'223D82.8864.9844.7333.9750.4238.60
ViL3DRel [7]NeurlPS'223D81.5868.6240.3030.7147.9437.73
EDA [43]CVPR'233D85.7668.5749.1337.6454.5942.26
3DJCG [5]CVPR'223D+2D83.3764.3441.3930.8249.5637.33
3D-VLP [21]CVPR'233D+2D84.2364.6143.5133.4151.4139.46
3D-VisTA(Scratch) [51]ICCV'233D77.4070.9038.7034.8045.9041.50
G3-LQ(Two-Stage)——3D88.0972.7351.4840.8056.9045.58
3D-SPS(One-stage) [31]CVPR'223D81.6364.7739.4829.6147.6536.43
BUTD-DETR(One-stage) [19]ECCV'223D81.4761.2444.2032.8149.7637.05
EDA(One-stage) [43]CVPR'233D86.4069.4248.1136.8253.8341.70
G3-LQ(One-stage)——3D88.5973.2850.2339.7255.9544.72
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.38, + 0.893, + 0.424 + ], + "angle": 0, + "content": "Table 1. Comparison results on the ScanRefer dataset, in terms of the accuracy evaluated by IoU 0.25 and IoU 0.5. The unique denotes samples devoid of distracting objects, while multiple applies to remaining samples. In the one-stage setting, there is no reliance on the additional 3D object detection step. Our proposed \\( \\mathrm{G}^3 -\\mathrm{LQ} \\) performs favorably over current state-of-the-art approach EDA [43]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.439, + 0.47, + 0.53 + ], + "angle": 0, + "content": "kens, \\(\\tau\\) is the temperature coefficient. \\(\\pmb{w}^{+}\\) \\(w^{-}\\) are weights of positive and negative terms. Similar to EDA, \\(\\pmb{t}_i\\) is the positive text features of the \\(i\\)-th object query and \\(\\mathcal{T}_i^+\\) are the positive features set of \\(\\pmb{t}_i\\), which primarily includes the object entries, attributes, relations, and pronouns. Moreover, the text loss can be formulated similarly:" + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.538, + 0.47, + 0.584 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {p s a}} ^ {t} = \\sum_ {i = 1} ^ {l} \\frac {- 1}{\\left| \\mathcal {Q} _ {i} ^ {+} \\right|} \\sum_ {\\boldsymbol {q} _ {i} ^ {c} \\in \\mathcal {Q} _ {i} ^ {+}} \\log \\left[ \\right. \\frac {\\exp \\left[ \\boldsymbol {w} _ {+} \\left(- d _ {c} \\left(\\boldsymbol {t} _ {i} ^ {c} , \\boldsymbol {q} _ {i} ^ {c}\\right) / \\tau\\right)\\right]}{\\sum_ {j = 1} ^ {K} \\exp \\left[ \\boldsymbol {w} _ {-} \\left(- d _ {c} \\left(\\boldsymbol {t} _ {i} ^ {c} , \\boldsymbol {q} _ {j} ^ {c}\\right) / \\tau\\right)} \\right] \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.591, + 0.469, + 0.621 + ], + "angle": 0, + "content": "where \\(\\pmb{q}_i^c\\in \\mathcal{Q}_i^+\\) is the positive object features of \\(\\pmb{t}_i^c\\) . The final loss \\(\\mathcal{L}_{\\mathrm{psa}}\\) is the average of the above two terms." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.622, + 0.469, + 0.667 + ], + "angle": 0, + "content": "Following [43], the total loss of our method also includes the box regression loss \\(\\mathcal{L}_{\\mathrm{loc}}\\), the position alignment loss \\(\\mathcal{L}_{\\mathrm{pa}}\\) which details in the Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.675, + 0.202, + 0.692 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.7, + 0.469, + 0.762 + ], + "angle": 0, + "content": "Firstly, we carry out comparisons with SOTA methods on ScanRefer and \\(\\mathrm{Nr3D / Sr3D}\\) dataset in Sec. 4.1. Further, we delve into ablation studies in Sec. 4.2. Finally, we visualize and analyze the 3D VG results in Sec. 4.3." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.771, + 0.317, + 0.789 + ], + "angle": 0, + "content": "4.1. Quantitative Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Performance on the ScanRefer. As illustrated in Table. 1, our method consistently exhibits admirably performance, trumping all competitors by a substantial margin on all test subsets. Our method achieves \\(56.90\\%\\) and \\(45.58\\%\\) performance in terms of the overall accuracy, with a remarkable improvement compared with EDA [43], \\(2.31\\%\\) and \\(2.34\\%\\). 1) Compared with 2D assistance methods [17, 44]" + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.437, + 0.892, + 0.673 + ], + "angle": 0, + "content": "
MethodNr3DSr3D
OverallHardOverallHard
TGNN [16]37.330.645.036.9
InstanceRefer [47]38.831.848.040.5
3DVG [48]40.834.851.444.9
LanguageRefer [39]43.936.656.049.3
TransRefer3D [15]48.039.657.450.2
SAT [44]49.242.457.950.0
LAR [4]48.942.359.451.2
3DRef [2]47.038.339.032.0
3D-SPS [31]51.545.162.665.4
MVT [17]55.149.164.558.8
BUTD-DETR [19]54.648.467.063.2
EDA [43]52.146.168.162.9
3D-VisTA [51]57.549.469.663.6
G3-LQ58.450.773.166.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.685, + 0.892, + 0.713 + ], + "angle": 0, + "content": "Table 2. Quantitative comparisons on \\(Nr3D\\) and \\(Sr3D\\) dataset. We have highlighted the top-performing three methods in purple." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.894, + 0.901 + ], + "angle": 0, + "content": "that furnish 2D image priors imbued with elaborate semantics and textures, the proposed \\(\\mathrm{G}^3\\)-LQ method showcases its superiority in performance. This serves as further evidence that our method explicitly captures geometric features, thus facilitating the shape understanding of target object and the vision-text alignment. 2) Our approach, featuring fine-grained language-guided query selection, surpasses all recently published DETR-like models [19, 43]. To explain, the proposed Flan-QS module thoroughly exploits fine-grained language prior and captures global context dependency, which effectively mitigates the issues of grounding ambiguity. 3) Compared with the 3D pre-trained" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13922" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.09, + 0.472, + 0.244 + ], + "angle": 0, + "content": "
IDPAGEFlan-QSPSAUniqueMultiple
0.250.50.250.5
(a)———85.6768.5749.1337.64
(b)✓——86.8970.2650.4638.61
(c)—✓—86.7569.7650.1138.58
(d)——✓86.3269.2049.5838.18
(e)✓—✓87.4571.9551.0840.49
(f)—✓✓87.3170.7550.9040.19
(g)✓✓—87.6672.1650.8740.05
(h)✓✓✓88.0972.7351.4840.80
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.255, + 0.47, + 0.283 + ], + "angle": 0, + "content": "Table 3. Ablation study on the effectiveness of the proposed PAGE, Flan-QS and PSA. Evaluated on the ScanRefer dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.292, + 0.469, + 0.429 + ], + "angle": 0, + "content": "models [21, 51], \\(\\mathrm{G}^3\\)-LQ achieves SOTA accuracy, because of the semantic-geometric alignment in Poincaré space with the tailored PSA loss, which aids in modeling the complex relationships of vision-text features. 4) The \"multiple\" setting entails a 3D scene where the description distinguishes the target object amidst numerous distractors. Under this setting, we reach a promising performance of \\(51.14\\%\\) and \\(40.08\\%\\), providing further validation for the efficacy of \\(\\mathrm{G}^3\\)-LQ in addressing the Semantic-Geometric Coupling issue." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.429, + 0.47, + 0.686 + ], + "angle": 0, + "content": "Performance on the Nr3D/Sr3D. The Nr3D/Sr3D task is geared towards locating the target object among all provided ground truth candidate boxes, representing a departure from the ScanRefer dataset. The overall accuracy of the proposed G\\(^3\\)-LQ along with other exceptional methods are reported in Table. 2. We attain peak performance of \\(58.4\\%\\) and \\(73.1\\%\\) on Nr3D and Sr3D. In Nr3D, descriptions exhibit noteworthy intricacy and detail, inducing additional challenges to 3D VG task. However, our method demonstrates a noteworthy superiority over alternative methods that rely on 2D images priors [44] or 3D vision-language pre-training models [51]. In the more challenging \"Hard\" subset, our method notably improves the accuracy by \\(1.3\\%\\) in Nr3D and \\(2.7\\%\\) in Sr3D, which underscores our method is beneficial for distinguishing ambiguous objects by effectively modeling the underlying geometric shape and unraveling the complex relation between vision-text features." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.696, + 0.334, + 0.712 + ], + "angle": 0, + "content": "4.2. Ablation Study and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Effectiveness of Proposed Components. We develop several alternative designs and undertake experiments to highlight the advantages yielded by various components. We herein train our \\(\\mathrm{G}^3\\)-LQ model on the ScanRefer dataset. (a) outlines the performance demonstrated by EDA [43], while (h) is our proposed method. 1) Comparing (a) and (b), we witness that the PAGE module delivers a remarkable gain over the baseline model. The comparison further accentuates the paramount role of explicit 3D geometric features in modeling object shape and spatial relationship, conferring benefits to 3D visual grounding. 2) The comparative result of (b) and (g) furnishes compelling evidence that the Flan" + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.09, + 0.891, + 0.203 + ], + "angle": 0, + "content": "
pipi - pijpi - pijUnique(0.25)Multiple(0.25)
✓——85.6749.13
✓✓—85.8349.19
✓—maxpool85.9849.32
✓—softmax86.0449.37
✓—Afine86.5750.06
✓✓Afine86.8950.46
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.213, + 0.892, + 0.241 + ], + "angle": 0, + "content": "Table 4. Ablation study of the geometric priors selection in the proposed PAGE module on the ScanRefer dataset." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.257, + 0.895, + 0.357 + ], + "angle": 0, + "content": "
Query SelectionUnique(~19%)Multiple(~81%)
0.250.50.250.5
Parametric81.4060.1338.0228.05
Top-K(EDA)85.6768.5749.1337.64
Language87.1071.5950.6339.31
Language Graph87.6672.1650.8740.05
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.367, + 0.892, + 0.41 + ], + "angle": 0, + "content": "Table 5. Ablation study of the query selection. \"Parametric\" represents the selection strategy with random parametric queries. \"Language\" denotes the guidance of sentence-level language." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.416, + 0.892, + 0.793 + ], + "angle": 0, + "content": "QS module exploiting language prior and context dependency, helps to alleviate the problem of grounding ambiguity and generate precise query proposals. 3) Comparing (h) and (g), it showcases the effectiveness of the proposed PSA loss to capture complex non-linear mappings of point cloud and text embeddings on Poincaré ball. When we integrate all proposed modules, a discernible surge is observed, yielding an impressive overall accuracy of \\(56.90\\%\\) and \\(45.58\\%\\). Selection of Geometric Priors. The fundamental objective of the PAGE module lies in embedding low-level geometric priors into high-level geometric features, thus the definition of \\(\\tilde{p}\\) emerges as a topic deserving exploration. We undertake ablations with \\(p_i, p_i - p_{ij}, \\| p_i - p_{ij} \\|\\) as illustrative examples, which unveils the inherent relative position, shapes, and distance (see Fig. 3) among 3D objects. As shown in Table. 4, using \\(\\| p_i - p_{ij} \\|\\) or \\(p_i - p_{ij}\\) alone yields inferior performance compared with their gradual integration, the accuracy of which reach \\(86.89\\%\\) (Uniqu) and \\(50.46\\%\\) (Multiple) respectively. Further, when compared with max pooling and softmax operators for features aggregation in \\(\\mathcal{X}_i\\), our geometric refinement strategy attests to an improved performance by \\(0.59\\%\\) and \\(0.59\\%\\) (Unique). The noteworthy results accentuate the significance of the geometric refinement in the PAGE module, given its potential to capture position-adaptive geometric structures." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Methods of Query Selection. We carry out several experiments to validate the proposed query selection method. We delineate the overall performance on the ScanRefer dataset in Table. 5. 1) We adopt the parametric queries used in MDETR [22]. However, we observe that the random parametric queries in performance decline for its incapability to capture contextual information of the text and 3D scene. 2)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13923" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.131, + 0.098, + 0.255, + 0.11 + ], + "angle": 0, + "content": "(a) Ordinal Numbers" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.111, + 0.268, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.187, + 0.273, + 0.204 + ], + "angle": 0, + "content": "this is a chair by the wall. It is the fifth chair from the right wall surface." + }, + { + "type": "image_caption", + "bbox": [ + 0.314, + 0.098, + 0.444, + 0.11 + ], + "angle": 0, + "content": "(b) Distance and Size" + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.114, + 0.477, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.306, + 0.187, + 0.458, + 0.204 + ], + "angle": 0, + "content": "this is a grey chair. its at the end of the shorter table closest to the window." + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.207, + 0.269, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.129, + 0.294, + 0.266, + 0.321 + ], + "angle": 0, + "content": "it is a black office chair. The black office chair is the fourth office chair on the left side of the table." + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.206, + 0.471, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.3, + 0.295, + 0.465, + 0.32 + ], + "angle": 0, + "content": "we are looking for the chair on the right. we are not looking for the chair on the left. the chair on the left is closest to the bed." + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.323, + 0.293, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.391, + 0.27, + 0.409 + ], + "angle": 0, + "content": "there is a gray chair. it is the second closest one to a trash can." + }, + { + "type": "image", + "bbox": [ + 0.298, + 0.323, + 0.475, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.316, + 0.388, + 0.456, + 0.413 + ], + "angle": 0, + "content": "the table is the left-most one in the center of the room. The table is dark orange and has four legs." + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.098, + 0.626, + 0.11 + ], + "angle": 0, + "content": "(c) Geometric shape" + }, + { + "type": "image", + "bbox": [ + 0.486, + 0.114, + 0.684, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.183, + 0.646, + 0.209 + ], + "angle": 0, + "content": "standing semi-circular wooden table with flat surface for holding objects. object can be found in center of image" + }, + { + "type": "image", + "bbox": [ + 0.475, + 0.213, + 0.649, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.491, + 0.297, + 0.659, + 0.314 + ], + "angle": 0, + "content": "The towel is between the entrance door and the sink. The towel is olive green and wavy." + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.318, + 0.651, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.49, + 0.393, + 0.654, + 0.409 + ], + "angle": 0, + "content": "the end table is the closest one to the window. The end table is a round cylinder." + }, + { + "type": "image_caption", + "bbox": [ + 0.691, + 0.098, + 0.827, + 0.11 + ], + "angle": 0, + "content": "(d) Complex Utterance" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.111, + 0.822, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.679, + 0.177, + 0.858, + 0.209 + ], + "angle": 0, + "content": "in front of a brown wooden desk are two white chairs. we are looking for the chair on the right. we are not looking for the chair on the left. the chair on the left is closest to the bed." + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.211, + 0.829, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.686, + 0.281, + 0.859, + 0.323 + ], + "angle": 0, + "content": "this is the brown arm chair that is at the end of the chalk board, to the left of it there is a small bookshelf, to the right of it and at the table is a blue chair. The brown chair is not under the table it is just next to the table." + }, + { + "type": "image", + "bbox": [ + 0.685, + 0.326, + 0.863, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.686, + 0.381, + 0.867, + 0.415 + ], + "angle": 0, + "content": "walking into the room, a large office desk is in the middle of the room. On the right side of the desk, there is a computer monitor on the table, the one on the left side, closest to the door." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.42, + 0.892, + 0.449 + ], + "angle": 0, + "content": "Figure 3. Qualitative results of the EDA [43] and our proposed \\(\\mathrm{G}^3\\)-LQ on the ScanRefer dataset. Our \\(\\mathrm{G}^3\\)-LQ manifests excellent grounding performance in terms of ordinal numbers, spatial distance or object size, geometric attribute and complex utterances." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.461, + 0.47, + 0.613 + ], + "angle": 0, + "content": "We generate the query proposals guided by the sentence-level features. The performance of the language-guided query selection mechanism stands out prominently, underscoring the crucial role of text features in the query generation. 3) Our Flan-QS module showcases superior performance over the prevailing top- \\( K \\) selection and sentence-level methods. The results give explanation of the efficacy of explicit fine-grained text features. They precisely model the attributes and positions of target objects, thereby facilitating a more accurate generation of query proposals." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.624, + 0.307, + 0.64 + ], + "angle": 0, + "content": "4.3. Qualitative Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.471, + 0.903 + ], + "angle": 0, + "content": "To offer profound insight into the effectiveness of the proposed \\(\\mathrm{G}^3\\)-LQ, we visualize the grounding results of our method alongside the SOTA EDA [43]. As shown in Fig. 3, the predicted boxes of our \\(\\mathrm{G}^3\\)-LQ method and EDA are visually highlighted in yellow and red, with ground-truth boxes marked in green. The localization ability of EDA is found wanting, encountering challenges in handling the 3D objects described by ordinal numbers, relative distance, and geometric shapes (see Fig. 3(a)-(c)). Conversely, our \\(\\mathrm{G}^3\\)-LQ incorporates intrinsic geometric features and spatial relation, fostering an enriched understanding of both the scene and objects, manifests commendable grounding prowess. Secondly, Fig. 3(d) exemplifies the commendable capabilities of our method in grounding 3D objects narrated by complex utterances. To elucidate, the \\(\\mathrm{G}^3\\)-LQ model with the tailored Flan-QS, excels in harnessing language priors and capturing global context dependencies, generat-" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.461, + 0.863, + 0.476 + ], + "angle": 0, + "content": "ing high-quality semantic-aware query representations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.484, + 0.619, + 0.499 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.504, + 0.892, + 0.761 + ], + "angle": 0, + "content": "In this paper, we analysis the Semantic-Geometric Coupling issue of 3D VG tasks in three aspects: intricate geometric details, complex text descriptions, and semantic-geometric inconsistency. To this end, we propose a \\(\\mathrm{G}^3\\)-LQ framework to explicitly capture the fine-grained geometric and semantic concepts, which includes three crucial novelties. Firstly, a Position Adaptive Geometric Exploring module is designed to capture explicit geometric features. Secondly, with decoupled multiple semantic components, a Fine-grained Language-guided Query Selection module generates object queries densely aligned by fine-grained text features. Finally, a Poincaré Semantic Alignment loss capitalizes on vision-text hierarchy natures and achieves alignment in Poincaré space, encouraging semantic-geometric consistency. Experimental findings showcase the superior performance of our \\(\\mathrm{G}^3\\)-LQ across the ScanRefer and Nr3D/Sr3D benchmarks, establishing a new state-of-the-art standard." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.77, + 0.681, + 0.787 + ], + "angle": 0, + "content": "6. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This work is supported by the state key development program in 14th Five-Year under Grant No. 2021YFF0602103, 2021YFF0602102, 2021QY1702. Thanks for the research fund under Grant No. 2019GQG0001 from the Institute for Guo Qiang, Tsinghua University. This study is also supported by Tsinghua University Toyota Joint Research Center for AI Technology of Automated Vehicle(TTAD 2023-07)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13924" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.142 + ], + "angle": 0, + "content": "[1] Scene graph parser. https://github.com/vacancy/SceneGraphParser, 2019.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.145, + 0.468, + 0.228 + ], + "angle": 0, + "content": "[2] Ahmed Abdelreheem, Ujjwal Upadhyay, Ivan Skorokhodov, Rawan Al Yahya, Jun Chen, and Mohamed Elhoseiny. 3dref-transformer: Fine-grained object identification in real-world scenes using natural language. In Proceedings of the IEEE Winter Conference on Applications of Computer Vision, pages 3941-3950, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.231, + 0.468, + 0.3 + ], + "angle": 0, + "content": "[3] Panos Achlioptas, Ahmed Abdelreehem, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In Proceedings of the European Conference on Computer Vision, pages 422-440. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.302, + 0.468, + 0.357 + ], + "angle": 0, + "content": "[4] Eslam Bakr, Yasmeen Alsaeda, and Mohamed Elhoseiny. Look around and refer: 2d synthetic semantics knowledge distillation for 3d visual grounding. Advances in Neural Information Processing Systems, 35:37146-37158, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.36, + 0.468, + 0.429 + ], + "angle": 0, + "content": "[5] Daigang Cai, Lichen Zhao, Jing Zhang, Lu Sheng, and Dong Xu. 3djcg: A unified framework for joint dense captioning and visual grounding on 3d point clouds. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 16464-16473, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.431, + 0.468, + 0.487 + ], + "angle": 0, + "content": "[6] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In Proceedings of the European Conference on Computer Vision, pages 202-221. Springer, 2020. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.488, + 0.468, + 0.557 + ], + "angle": 0, + "content": "[7] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Language conditioned spatial relation reasoning for 3d object grounding. Advances in Neural Information Processing Systems, 35:20522-20535, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.56, + 0.468, + 0.616 + ], + "angle": 0, + "content": "[8] Jaesung Choe, Chunghyun Park, Francois Rameau, and Jaesik Park. Pointmixer: Mlp-mixer for point cloud understanding. In Proceedings of the European Conference on Computer Vision, pages 620-640. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.618, + 0.468, + 0.672 + ], + "angle": 0, + "content": "[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.468, + 0.731 + ], + "angle": 0, + "content": "[10] Jiafei Duan, Samson Yu, Hui Li Tan, Hongyuan Zhu, and Cheston Tan. A survey of embodied ai: From simulators to research tasks. IEEE Transactions on Emerging Topics in Computational Intelligence, 6(2):230-244, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.468, + 0.814 + ], + "angle": 0, + "content": "[11] Aleksandr Ermolov, Leyla Mirvakhabova, Valentin Khrulkov, Nicu Sebe, and Ivan Oseledets. Hyperbolic vision transformers: Combining improvements in metric learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7409-7419, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[12] Mingtao Feng, Zhen Li, Qi Li, Liang Zhang, XiangDong Zhang, Guangming Zhu, Hui Zhang, Yaonan Wang, and Ajmal Mian. Free-form description guided 3d visual graph network for object grounding in point cloud. In Proceedings of the IEEE International Conference on Computer Vision, pages 3722-3731, 2021. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[13] Octavian Ganea, Gary Bécigneul, and Thomas Hofmann. Hyperbolic neural networks. Advances in Neural Information Processing Systems, 31, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.188 + ], + "angle": 0, + "content": "[14] William L Hamilton, Rex Ying, and Jure Leskovec. Inductive representation learning on large graphs. In Advances in Neural Information Processing Systems, pages 1025-1035, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.19, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[15] Dailan He, Yusheng Zhao, Junyu Luo, Tianrui Hui, Shaofei Huang, Aixi Zhang, and Si Liu. Transrefer3d: Entity-andrelation aware transformer for fine-grained 3d visual grounding. In Proceedings of the 29th ACM International Conference on Multimedia, pages 2344-2352, 2021. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.261, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[16] Pin-Hao Huang, Han-Hung Lee, Hwann-Tzong Chen, and Tyng-Luh Liu. Text-guided graph neural networks for referring 3d instance segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1610-1618, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.33, + 0.892, + 0.385 + ], + "angle": 0, + "content": "[17] Shijia Huang, Yilun Chen, Jiaya Jia, and Liwei Wang. Multiview transformer for 3d visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15524-15533, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.386, + 0.892, + 0.453 + ], + "angle": 0, + "content": "[18] Wenlong Huang, Pieter Abbeel, Deepak Pathak, and Igor Mordatch. Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. In International Conference on Machine Learning, pages 9118-9147. PMLR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.455, + 0.892, + 0.524 + ], + "angle": 0, + "content": "[19] Ayush Jain, Nikolaos Gkanatsios, Ishita Mediratta, and Katerina Fragkiadaki. Bottom up top down detection transformers for language grounding in images and point clouds. In Proceedings of the European Conference on Computer Vision, pages 417-433. Springer, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.525, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[20] Haojun Jiang, Yuanze Lin, Dongchen Han, Shiji Song, and Gao Huang. Pseudo-q: Generating pseudo language queries for visual grounding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 15513-15523, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.595, + 0.892, + 0.662 + ], + "angle": 0, + "content": "[21] Zhao Jin, Munawar Hayat, Yuwei Yang, Yulan Guo, and Yinjie Lei. Context-aware alignment and mutual masking for 3d-language pre-training. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 10984-10994, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.664, + 0.892, + 0.733 + ], + "angle": 0, + "content": "[22] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE International Conference on Computer Vision, pages 1780-1790, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.735, + 0.892, + 0.775 + ], + "angle": 0, + "content": "[23] Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In International Conference on Learning Representations, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.776, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[24] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[25] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13925" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[26] Yongcheng Liu, Bin Fan, Shiming Xiang, and Chunhong Pan. Relation-shape convolutional neural network for point cloud analysis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8895-8904, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.471, + 0.233 + ], + "angle": 0, + "content": "[27] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.471, + 0.29 + ], + "angle": 0, + "content": "[28] Yongfei Liu, Bo Wan, Xiaodan Zhu, and Xuming He. Learning cross-modal context graph for visual grounding. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11645-11652, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.471, + 0.359 + ], + "angle": 0, + "content": "[29] Yongfei Liu, Bo Wan, Lin Ma, and Xuming He. Relation-aware instance refinement for weakly supervised visual grounding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5612-5621, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.361, + 0.471, + 0.417 + ], + "angle": 0, + "content": "[30] Ze Liu, Zheng Zhang, Yue Cao, Han Hu, and Xin Tong. Group-free 3d object detection via transformers. In Proceedings of the IEEE International Conference on Computer Vision, pages 2949-2958, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.419, + 0.471, + 0.501 + ], + "angle": 0, + "content": "[31] Junyu Luo, Jiahui Fu, Xianghao Kong, Chen Gao, Haibing Ren, Hao Shen, Huaxia Xia, and Si Liu. 3d-sps: Single-stage 3d visual grounding via referred point progressive selection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 16454-16463, 2022. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.503, + 0.471, + 0.56 + ], + "angle": 0, + "content": "[32] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. In International Conference on Learning Representations, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.561, + 0.471, + 0.616 + ], + "angle": 0, + "content": "[33] Antonio Montanaro, Diego Valsesia, and Enrico Magli. Rethinking the compositionality of point clouds through regularization in the hyperbolic space. Advances in Neural Information Processing Systems, 35:33741-33753, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.617, + 0.471, + 0.659 + ], + "angle": 0, + "content": "[34] Maximillian Nickel and Douwe Kiela. Poincaré embeddings for learning hierarchical representations. Advances in Neural Information Processing Systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.471, + 0.743 + ], + "angle": 0, + "content": "[35] Aishwarya Padmakumar, Jesse Thomason, Ayush Shrivastava, Patrick Lange, Anjali Narayan-Chen, Spandana Gella, Robinson Piramuthu, Gokhan Tur, and Dilek Hakkani-Tur. Teach: Task-driven embodied agents that chat. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2017–2025, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.745, + 0.471, + 0.801 + ], + "angle": 0, + "content": "[36] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in Neural Information Processing Systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.471, + 0.858 + ], + "angle": 0, + "content": "[37] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In proceedings of the IEEE International Conference on Computer Vision, pages 9277-9286, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[38] Zheng Qin, Hao Yu, Changjian Wang, Yulan Guo, Yuxing Peng, and Kai Xu. Geometric transformer for fast and robust point cloud registration. In Proceedings of the IEEE Con-" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "ference on Computer Vision and Pattern Recognition, pages 11143-11152, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[39] Junha Roh, Karthik Desingh, Ali Farhadi, and Dieter Fox. Languagerefer: Spatial-language model for 3d visual grounding. In Conference on Robot Learning, pages 1046-1056. PMLR, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.177, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[40] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.892, + 0.288 + ], + "angle": 0, + "content": "[41] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (TOG), 38(5):1-12, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.289, + 0.892, + 0.344 + ], + "angle": 0, + "content": "[42] Yingming Wang, Xiangyu Zhang, Tong Yang, and Jian Sun. Anchor detr: Query design for transformer-based detector. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2567-2575, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.344, + 0.892, + 0.4 + ], + "angle": 0, + "content": "[43] Yanmin Wu, Xinhua Cheng, Renrui Zhang, Zesen Cheng, and Jian Zhang. Eda: Explicit text-decoupling and dense alignment for 3d visual and language learning. arXiv preprint arXiv:2209.14941, 2022. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.4, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[44] Zhengyuan Yang, Songyang Zhang, and Jiebo Luo. Sat: 2d semantics assisted training for 3d visual grounding. In Proceedings of the IEEE International Conference on Computer Vision, pages 1856-1866, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.456, + 0.892, + 0.512 + ], + "angle": 0, + "content": "[45] Ting Yao, Yehao Li, and Tao Mei. Hgnet: Learning hierarchical geometry from points, edges, and surfaces. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 21846-21855, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.512, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[46] Xumin Yu, Yongming Rao, Ziyi Wang, Zuyan Liu, Jiwen Lu, and Jie Zhou. Pointr: Diverse point cloud completion with geometry-aware transformers. In Proceedings of the IEEE International Conference on Computer Vision, pages 12498-12507, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.58, + 0.892, + 0.664 + ], + "angle": 0, + "content": "[47] Zhihao Yuan, Xu Yan, Yinghong Liao, Ruimao Zhang, Sheng Wang, Zhen Li, and Shuguang Cui. Instancerefer: Cooperative holistic understanding for visual grounding on point clouds through instance multi-level contextual referring. In Proceedings of the IEEE International Conference on Computer Vision, pages 1791-1800, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.664, + 0.892, + 0.72 + ], + "angle": 0, + "content": "[48] Lichen Zhao, Daigang Cai, Lu Sheng, and Dong Xu. 3dvg-transformer: relation modeling for visual grounding on point clouds. In Proceedings of the IEEE International Conference on Computer Vision, pages 2928-2937, 2021. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.72, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[49] Chaoyang Zhu, Yiyi Zhou, Yunhang Shen, Gen Luo, Xingjia Pan, Mingbao Lin, Chao Chen, Liujuan Cao, and Xiaoshuai Sun. Seqtr: A simple yet universal network for visual grounding. In Proceedings of the European Conference on Computer Vision, pages 598-615. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.79, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[50] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[51] Ziyu Zhu, Xiaojian Ma, Yixin Chen, Zhidong Deng, Siyuan Huang, and Qing Li. 3d-vista: Pre-trained transformer for 3d vision and text alignment. arXiv preprint arXiv:2308.04352, 2023. 2, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13926" + } + ] +] \ No newline at end of file diff --git a/2024/G^3-LQ_ Marrying Hyperbolic Alignment with Explicit Semantic-Geometric Modeling for 3D Visual Grounding/7e2af7bd-e579-4e10-8895-d1ef28fe0cc7_origin.pdf b/2024/G^3-LQ_ Marrying Hyperbolic Alignment with Explicit Semantic-Geometric Modeling for 3D Visual Grounding/7e2af7bd-e579-4e10-8895-d1ef28fe0cc7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..77c4436cf1fff3b917c9a845ac16a25f1ccc6e89 --- /dev/null +++ b/2024/G^3-LQ_ Marrying Hyperbolic Alignment with Explicit Semantic-Geometric Modeling for 3D Visual Grounding/7e2af7bd-e579-4e10-8895-d1ef28fe0cc7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9662ebbe10ef9cc4b50bbeee049b958cf25f9451d65991fb5d3fed255ee4ebbb +size 2624189 diff --git a/2024/G^3-LQ_ Marrying Hyperbolic Alignment with Explicit Semantic-Geometric Modeling for 3D Visual Grounding/full.md b/2024/G^3-LQ_ Marrying Hyperbolic Alignment with Explicit Semantic-Geometric Modeling for 3D Visual Grounding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ad3d616f7dc71698e1dfc042ec99efad92172f65 --- /dev/null +++ b/2024/G^3-LQ_ Marrying Hyperbolic Alignment with Explicit Semantic-Geometric Modeling for 3D Visual Grounding/full.md @@ -0,0 +1,355 @@ +# $\mathbf{G}^{3}$ -LQ: Marrying Hyperbolic Alignment with Explicit Semantic-Geometric Modeling for 3D Visual Grounding + +Yuan Wang $^{1,2}$ Yali Li $^{1,2}$ Shengjin Wang $^{1,2\ddagger}$ + +$^{1}$ Department of Electronic Engineering, Tsinghua University, China + +$^{2}$ Beijing National Research Center for Information Science and Technology (BNRist), China wy23@mails.tsinghua.edu.cn {liyali13, wgsgj} @tsinghua.edu.cn $\ddagger$ Corresponding Author + +![](images/797fc1d01a607893704a400430e6eea3bd193b90bcef460b710919ce8241dddd.jpg) +Figure 1. Illustration of the proposed G $^3$ -LQ method on the ScanRefer benchmark. The versatile scene-level and object-level understanding properties empower G $^3$ -LQ to comprehend spatial relationship, relative distance, geometric attribute as well as intricate references. + +# Abstract + +Grounding referred objects in 3D scenes is a burgeoning vision-language task pivotal for propelling Embodied AI, as it endeavors to connect the 3D physical world with free-form descriptions. Compared to the 2D counterparts, challenges posed by the variability of 3D visual grounding remain relatively unsolved in existing studies: 1) the underlying geometric and complex spatial relationships in 3D scene. 2) the inherent complexity of 3D grounded language. 3) the inconsistencies between text and geometric features. To tackle these issues, we propose $G^3$ -LQ, a DEtection Transformer-based model tailored for 3D visual grounding task. $G^3$ -LQ explicitly models Geometric-aware visual representations and Generates fine-Grained Language-guided object Queries in an overarching framework, which comprises two dedicated modules. Specifically, the Position Adaptive Geometric Exploring (PAGE) unearths underlying information of 3D objects in the geometric details and spatial relationships perspectives. The Fine-grained Language-guided Query Selection (Flan-QS) delves into syntactic structure of texts and generates object queries that exhibit higher relevance towards fine-grained text features. Finally, a pioneering Poincaré Semantic Alignment (PSA) loss establishes semantic-geometry consistencies by mod + +eling non-linear vision-text feature mappings and aligning them on a hyperbolic prototype—Poincaré ball. Extensive experiments verify the superiority of our $G^3$ - $LQ$ method, trumping the state-of-the-arts by a considerable margin. + +# 1. Introduction + +Grounding 3D physical realm with natural languages has propelled to the forefront in advancing Embodied AI [10, 18, 35], an ability that empowers agents to comprehend human instructions in real-world contexts. 3D Visual Grounding (3D VG) [6, 31, 43, 48] has garnered substantial attention as a crucial cross-modal 3D perception task in defacto applications, e.g., assistive robots, AR/VR, and metaverse. + +Compared with extensive explorations of 2D counterpart [20, 28, 29, 49], 3D VG poses prominent challenge known as "Semantic-Geometric Coupling". As shown in Fig. 1, we offer distinctive glimpse into it: 1) Intricate geometric details and spatial relations. 3D scenes encounter multiple objects with intricate layout, multi-level occlusion, and multifaceted spatial relation. 2) Heightened flexibility and complexity of descriptions. Free-form utterances with complex syntax convey precise spatial information, intricate references, and detailed relational terms. 3) Inconsistency of the semantic-geometric features. A single semantic concep + +tion yields multiple potential correspondences within point cloud or exhibit diverse geometry in varied contexts. + +To ameliorate these issues, some pioneering methods have been proposed. They have largely focused on enhancing the representation capacity of 3D point clouds with 2D image priors [44]; capturing spatial relationships of 3D objects with powerful transformer [5, 15, 48] or graph convolution network [3, 12]; generating discriminative object proposals through instance segmentation [16] or linguistic guidance [31]. Despite commendable advancements be achieved, an equally important yet underexplored problem is explicit semantic-geometric modeling. Prior studies exhibit a significant dearth in explicitly modeling the intrinsic geometric attributes and spatial layouts of 3D objects, accentuating the dilemma to associate text descriptions with grounding objects. Moreover, object queries contain rich instance characteristics. But the object proposal [15, 48] and query generation [19, 31, 43] in current DETR-like models overlook explicit fine-grained interaction between text and object features, which in turn limit the model's capacity to refer potential language clues, e.g., appearances, shapes, or textures. Finally, major 3D VG systems [19, 21, 43, 51] strive to establish the alignment between point cloud and semantic concepts in canonical Euclidean space. However, their efficacy falters in unraveling the complex non-linear relationships between multi-modal features and upholding the intrinsic hierarchy natures in 3D scenes and texts. + +In this paper, our primary goal endeavors to explicitly delve into fine-grained representations of geometric and semantic concepts, then aligns two modalities in hyperbolic space. We thus devise an overarching G $^3$ -LQ framework: + +1) For geometric feature modeling, we firstly propose a Proxy Adaptive Geometric Exploring (PAGE) module. PAGE is composed of the Point-Proxy Geometric Extraction and Proxy Adaptive Geometric Refinement component. The former explicitly capture the local geometric attributes (e.g., distance and orientation) in a 3D local proxy, while the latter adaptively refine the geometric structure according to the position of point proxy. The geometric features are fed into the geometric encoder for high-level embedding and further achieve the interaction with text features for cross-modality enhancement. In this fashion, the PAGE module facilitates a seamless alignment of text descriptions with intrinsic geometric properties of 3D objects. +2) For complex utterance understanding, inspired by the dependency parsing, we aim to explore context dependency and syntactic structure of texts for guiding query selection. To this end, we propose a Fine-grained Language-guided Query Selection (Flan-QS) module, comprising the Language Scene Graph and Fine-grained Query Selection components. We decouple the free-form utterances to tree-like semantic components with directed dependency relationships as edges. Then, we establish and update the lan + +guage scene graph to represent the context-aware entry features via node/edge embedding and graph context learning. Finally, we employ the constructed language scene graph to steer the generation of subsequent object candidates that are densely-aligned with fine-grained linguistic attributes. + +3) For semantic-geometric consistency, we embed vision and text features to a hyperbolic prototype—Poincaré ball. The emerging appreciation [11, 13, 34] suggest that the hyperbolic space offers overwhelming superiority to handle data correlations. 3D scene (free-form text) exhibits an inherent hierarchical nature ranging from scene, object, part (from sentence, phrase, word). The hyperbolic space with negative curvature is the only space that can successfully embed the tree-shaped hierarchy [33]. Moreover, the compact Poincaré ball [13, 34] holds suitable geometric properties, enhancing its capacity to model non-linear mappings of vision and text features. Therefore, we design a Poincaré Semantic Alignment loss (PSA) to probe the complex mappings between vision and text features, facilitating the semantics-geometry consistency. + +Our chief contributions are threefold: + +- We roundly analysis the Semantic-Geometric Coupling of 3D VG task. Towards this issue, we propose a unified $\mathbf{G}^3$ -LQ framework to model intrinsic geometric features (PAGE module) and explicitly generate fine-grained object queries for 3D visual grounding (Flan-QS module). +- We propose a novel PSA loss to capture the non-linear relationship and underlying hierarchy of the vision-text features in the hyperbolic space, which further paves the way for semantics-geometry consistency learning. +- A battery of experiments on ScanRefer and Nr3d/Sr3d datasets demonstrate the effectiveness of our G $^{3}$ -LQ method. G $^{3}$ -LQ has proven its mettle, outperforming all existing SOTA methods by a considerable margin. + +# 2. Related Works + +# 2.1.3D Visual Grounding + +3D Visual Grounding has been an area of intense investigation that dedicates to pinpointing the target object in a 3D scene based on the text description. Prevailing methods are typically two-stage, which conforms to the detection-then-matching paradigm. Firstly, the 3D object detectors [30, 37, 48] are harnessed to yield candidate 3D object proposals, while the text features are encoded by language models [9, 27]. Secondly, these advanced methods perform a pivotal endeavor to align the vision and text features for target objects grounding. Among these, TGNN [16] utilizes the graph neural network to deduce spatial relationships among 3D object proposals, which are further enriched by text features. FFL-3DOG [12] captures the intramodal and cross-modal relationships of text descriptions and 3D point cloud via scene graphs interaction. Recent endeavors embrace the formidable transformers [40] to model + +the relationship of proposals and text features. 3DVG [48] and TransRefer3D [15] achieve exceptional performance by establishing self-attention and cross-attention protocols for context perception. LanguageRefer [39] seamlessly converts the cross-modal task into a unified language modeling challenge, underpinned by predicted object labels. + +However, two-stage methods encounter a noteworthy impediment: the detection stage neglects to leverage language contexts to prioritize the objects that are essential to the referring task. An alternative approach to the problem is one-stage methods [19, 31, 43], where extracted vision features are directly fused with the language features for object grounding. For instance, 3D-SPS [31] leverages text features to guide visual keypoints selection, enabling the progressive object grounding. BUTD-DETR [19] pioneers a DETR-like model, which fuses the vision-text features via a co-attention paradigm and then decodes objects in the utterance from the contextualized features. Building upon this, EDA [43] proposes a text-decoupled strategy and performs dense alignment of 3D objects and associated texts. + +Despite one-stage methods unfold prowess performance, inherent limitations have been yet underexplored: 1) They overlook the intrinsic geometric attributes (distance, orientation, and shape) in 3D point clouds, leading to ambiguous visual representations. Conversely, our proposed PAGE module explicitly captures geometric features to avoid ambiguity. 2) The queries generation lack the fine-grained language guidance, limiting the model's capacity to infer potential language clues. Our tailored Flan-QS module explores language prior and context dependency to generate precise proposals. 3) They establish the vision-text alignment in Euclidean space, incapable of modeling the nonlinear relation of vision and text features. Our method embeds multimodal features on the Poincaré ball then leverage the PSA loss to foster the semantic-geometric alignment. + +# 2.2. Geometric Extraction in 3D Point Clouds + +The explicit modeling of geometric structures has recently garnered considerable attention in the realms of 3D point cloud, e.g., point cloud understanding [8, 26, 32, 41], point cloud completion [46], point cloud registration [38], and 3D object detection [45]. DGCNN [41] proposes an EdgeConv operator, by which the point features are aggregated through graph convolution to capture intricate local geometric structures. RSCNN [26] explicitly encodes geometric relationship of points with relation-aware convolution. PointMixer [8] embeds geometric relationships of 3D point in a MLP-Mixer's streamline. PoinTr [46] devises a geometry-aware transformer to learn structural knowledge and local geometric relationships for point cloud completion. HGNet [45] delves into localized shape information to generate the proposals with a hierarchical graph model for 3D object detection. Our cutting-edge $\mathrm{G}^3$ -LQ incorporates + +the geometric attributes of point cloud into the 3D VG task, thereby enhancing the discriminability of vision features. + +# 3. Method + +# 3.1. Overview + +Provided a point cloud $\mathcal{P} = \{\pmb{p}_i\}_{i=1}^N \in \mathbb{R}^{N \times (3 + F)}$ with $F$ -dim auxiliary features (e.g., RGB, normals, or multi-view features) of $N$ points, and a free-form text $\mathcal{T}$ with $L$ words, the primary goal of 3D VG is to establish a mapping $\mathcal{M}$ that links $\mathcal{P}$ and $\mathcal{T}$ to the target object $\pmb{o}$ , i.e., $\mathcal{M}(\mathcal{P}, \mathcal{T}) \to \pmb{o}$ . + +Fig. 2 gives the outline of our proposed $\mathrm{G}^3$ -LQ method. Firstly, we leverage the well-accepted PointNet++ [36] tockenize the point cloud into vision token $\mathbf{V}\in \mathbb{R}^{n\times d}$ . The text is encoded by the pre-trained RoBERTa [27] and yields the vanilla text token $\mathbf{T}\in \mathbb{R}^{t\times d}$ . The vision token $\mathbf{V}$ is fed into the PAGE module to capture underlying geometric features $\mathbf{V}_g\in \mathbb{R}^{n\times d}$ , which adjusts according to the point positions, thereby furnishing precise geometric clues. Further, the geometric visual encoder embeds $\mathbf{V}_g$ via self-attention and interact with text features $\mathbf{T}$ via cross-attention. + +Finally, we build the language scene graph $\mathcal{G}$ and obtain the updated fine-grained text features, which will explicitly guide the queries selection. Like the object queries in most DETR-like models [22, 25, 43], the selected queries $\mathbf{Q} \in \mathbb{R}^{K \times d}$ are fed into a cross-modal decoder to probe desired features and update themselves. The decoded queries $\mathbf{Q}' \in \mathbb{R}^{K \times l}$ are fed into an MLP to predict object boxes. + +# 3.2. Geometric Exploring and Embedding + +In this section, we propose a PAGE module to explicitly capture geometric properties of 3D scenes, which serves to encode holistic visual representations for 3D VG tasks. + +Point-Proxy Geometric Extraction. To facilitate local geometric features modeling, we firstly instantiate point proxies predicated on the proximity of the points, thereby serving as the local regions of the point cloud. As illustrated on Fig 2, we search $k$ neighbors of $p_i \in \mathcal{P}$ as a point proxy with the KNN algorithm, denoted as: $\mathcal{X}_i = k\mathrm{NN}(\mathcal{P},p_i)$ . + +The point proxy $\mathcal{X}_i$ denotes index set with $k$ closest points of $\pmb{p}_i$ and $\pmb{v}_{m_i}:\{\pmb{v}_{m_{ij}}|j\in \mathcal{X}_i\} \in \mathbb{R}^{k\times d}$ is the corresponding features. With the Point-Proxy Geometric Extraction component, we capture the local structure by learning the geometric topology features within the point proxy $\mathcal{X}_i$ . In this fashion, it is effective to achieve an inductive local representations that integrate explicit reasoning pertaining to geometric structures and objects spatial layout. We primarily compute pre-defined geometric prior $\widetilde{\pmb{p}}_{ij}\in \mathbb{R}^7$ : + +$$ +\widetilde {\boldsymbol {p}} _ {i j} = \operatorname {c o n c a t} \left(\boldsymbol {p} _ {i} - \boldsymbol {p} _ {i j}, \boldsymbol {p} _ {i}, \left\| \boldsymbol {p} _ {i} - \boldsymbol {p} _ {i j} \right\|\right) \tag {1} +$$ + +$\widetilde{\pmb{p}}_{ij}$ embeds relative orientation and distance of different objects. Further, the low-level geometric features are encoded to high-level features $\pmb{h}_{ij} \in \mathbb{R}^d$ with a linear layer $\phi[\cdot]$ to obtain abstract geometric conception and excavate + +![](images/b67b4ac56f727ff4ffbe50de3d23ae65a9b69c8a024e406fcf3168d59955884b.jpg) +Text: $\mathcal{T} = \{t_i\}_{i = 1}^L$ +Point: $\mathcal{P} = \{\pmb {p}_i\}_{i = 1}^N$ + +![](images/1cdd38271661ca0d94a0c929951da1fe2b660e3702bc82dfe7edaa945e4be369.jpg) +Figure 2. Illustration of (a) our proposed $\mathbf{G}^3$ -LQ framework. The red dashed frames denote our novel modules. (b) PAGE moulder to model the underlying geometric structure of 3D point cloud and (c) the Geometric Visual Encoder further to embed the geometric topology features. (d) Flan-QS module to capture fine-grained language context with (e) Language Scene Graph for guiding query selection. + +discriminative geometric structures, which is described as: + +$$ +h _ {i j} = \widetilde {\boldsymbol {p}} _ {i j} + \phi [ \widetilde {\boldsymbol {p}} _ {i j} ], \forall j \in \mathcal {X} _ {i} \tag {2} +$$ + +Proxy-Adaptive Geometric Refinement. The presence of diverse geometric structures across distinct point proxies necessitate anisotropic extractors. Further, we flesh out this intuition and engineer a Proxy-Adaptive Geometric Refinement component, poised to model refined geometric features effectively of diverse point proxies. Given the geometric prior $\{h_{ij}|j\in \mathcal{X}_i\} \in \mathbb{R}^{k\times d}$ in $\mathcal{X}_i$ we establish local geometric refinement by the following streamline: + +$$ +\left\{g _ {i j} \right\} = \alpha \odot \frac {\left\{h _ {i j} \right\} - h _ {i}}{\Omega + \varepsilon} + \beta \tag {3} +$$ + +where $\varepsilon$ is a small constant for numerical stability and $\alpha, \beta$ are learnable parameters for features refinement. $\Omega$ is the standard deviation across all point proxies and channels. $h_i$ is the center of $\mathcal{X}_i$ . $g_i = \{g_{ij} | j \in \mathcal{X}_i\} \in \mathbb{R}^{k \times d}$ represents the anisotropic geometric structures in point proxy $\mathcal{X}_i$ . + +Finally, we opt for a softmax aggregation strategy to densely relate point features interaction in each $\mathcal{X}_i$ : + +$$ +\widetilde {v} _ {g _ {i}} = \operatorname {L i n e a r} \left(\operatorname {c o n c a t} \left(v _ {m _ {i}}, g _ {i}\right)\right) \in \mathbb {R} ^ {k \times d} +$$ + +$$ +v _ {g _ {i}} = \sum_ {j \in \mathcal {X} _ {i}} \operatorname {s o f t m a x} \left(\widetilde {v} _ {g _ {i j}}\right) \cdot \widetilde {v} _ {g _ {i j}} \in \mathbb {R} ^ {d}, \forall j \in \mathcal {X} _ {i}, \tag {4} +$$ + +where Linear $(\cdot)$ denotes the linear projection, $\mathbf{V}_g = \{v_{g_i}\}_{i=1}^n$ is the output token with geometric representations. + +Geometric Visual Encoder. Geometric perception enables to strengthen shape features of 3D objects and demonstrate enhanced viewpoint sensitivity, which bestows flexibility in capturing spatial relation and improves precise 3D objects grounding. Hence, we embed geometric topology features $\mathbf{V}_g$ via a tailored Geometric Encoder. As shown in Fig. 2, the core of Geometric Encoder is Geometric self-attention layer. Given the high-level vision features $\mathbf{V}_m\in \mathbb{R}^{n\times d}$ and the explicit geometric features $\mathbf{V}_g$ , the output $\mathbf{V}_{emb}\in \mathbb{R}^{n\times d}$ is a weighted sum of projected features: + +$$ +v _ {i} ^ {e m b} = \sum_ {j = 1} ^ {n} a _ {i j} \left(v _ {m _ {j}} \mathbf {W} ^ {V}\right) \in \mathbb {R} ^ {d} \tag {5} +$$ + +The attention weight $a_{ij}$ is computed by a row-wise softmax, which can be formulated as: + +$$ +a _ {i j} = \sigma \left[ \frac {\left(v _ {g _ {i}} \mathbf {W} _ {g} ^ {Q} + v _ {m _ {i}} \mathbf {W} _ {m} ^ {Q}\right) \left(v _ {g _ {j}} \mathbf {W} _ {g} ^ {K} + v _ {m _ {j}} \mathbf {W} _ {m} ^ {K}\right) ^ {T}}{\sqrt {d}} \right] \tag {6} +$$ + +$\mathbf{W}_g^Q,\mathbf{W}_g^K$ are geometric projection matrices of query and key, $\mathbf{W}_m^Q,\mathbf{W}_m^K$ and $\mathbf{W}_m^V$ are semantic projection matrices of query, key and value. $\sigma$ is the row-wise softmax operator. + +Further, the geometric-enhanced features $\mathbf{V}_{emb}$ pass through the cross-attention layer to interact with the text + +features and box features (used in two stage), for obtaining the cross-modal features $\mathbf{V}' \in \mathbb{R}^{n \times d}$ and $\mathbf{T}' \in \mathbb{R}^{t \times d}$ . Finally, $\mathbf{V}'$ is linearly projected as the candidate query $\mathbf{Q}_c$ . + +# 3.3. Query Selection + +In the DETR-like model [24, 42, 50], object queries play essential roles to determine potential regions of targets, which are equally prominent for 3G VG systems, decoded into a box center and size that conform to vision tokens. Assisted by global text features, GroundingDINO [25] employs a language-guided query selection protocol to generate query proposals. However, such text features are relatively coarse-grained, leading to inaccurate query generation. Actually, free-form texts harbor intricate syntactic structure and possess a context-dependency nature. It further serves as inspiration for us to generate precise vision queries via improving fine-grained semantic understanding. Language Scene Graph. We intend to delve into fine-grained text information, which is defined as decoupled semantic components (e.g., object, attributes, and relations) and syntactic structures. To this end, we perform a syntactic dependency parsing to construct language graph $\mathcal{G} = \{\mathcal{O},\mathcal{R}\}$ by the off-the-shelf Scene Graph Parser [1], which serves as a reasoning inductive bias of 3D VG tasks. $\mathcal{O} = \{\pmb{o}_i\}$ and $\mathcal{R} = \{\pmb{r}_{ij}\}$ are nodes and edges set. Each object $o_i$ is denoted as a phrase with a set of attributes (appearance, shape, texture). Harnessing the final layer of the pre-trained RoBERTa model, we then acquire the object phrase embedding $\pmb{x}_{o_i}\in \mathbb{R}^d$ as the graph node $o_i\in \mathcal{O}$ . In a similar vein, the relation phrases are also encoded into the context-aware embeddings $\pmb{x}_{r_{ij}}\in \mathbb{R}^d$ akin to the graph edges $\pmb{r}_{ij}\in \mathcal{R}$ . + +Inspired by the message passing mechanism [14, 23], we proceed to enhance relation and object embeddings via language graph convolution. Firstly, to enrich the contextual representations of relation embedding in connection with its affiliated nodes, we aggregate messages and update its features with the subject $\pmb{x}_{o_i}$ and object node $\pmb{x}_{o_j}$ : + +$$ +\widetilde {\boldsymbol {x}} _ {r _ {i j}} = \boldsymbol {x} _ {r _ {i j}} + F _ {r} \left(\left[ \boldsymbol {x} _ {o _ {i}}; \boldsymbol {x} _ {r _ {i j}}; \boldsymbol {x} _ {o _ {j}} \right]\right) \tag {7} +$$ + +where $F_{r}$ is a linear projection layer and $\widetilde{\pmb{x}}_{r_{ij}}$ is the context-aware relation embedding. We then update the phrase embedding $\pmb{x}_{o_i}$ by passing messages from all connected nodes $\Gamma (i)$ and the edges with a tailored attention mechanism: + +$$ +\widetilde {\boldsymbol {x}} _ {o _ {i}} = \boldsymbol {x} _ {o _ {j}} + \sum_ {j \in \Gamma (i)} \boldsymbol {w} _ {o _ {i j}} F _ {o} \left(\left[ \boldsymbol {x} _ {o _ {j}}; \boldsymbol {x} _ {r _ {i j}} \right]\right) \tag {8} +$$ + +where $F_{o}$ is another linear layer and $\widetilde{\pmb{x}}_{o_i}$ denotes context-aware object features. $\pmb{w}_{o_{ij}}$ is the tailored attention weight: + +$$ +\boldsymbol {w} _ {o _ {i j}} = \operatorname {s o f t m a x} \left[ F _ {o} \left(\left[ \boldsymbol {x} _ {o _ {i}}; \widetilde {\boldsymbol {x}} _ {r _ {i j}} \right]\right) ^ {T} F _ {o} \left(\left[ \boldsymbol {x} _ {o _ {j}}; \widetilde {\boldsymbol {x}} _ {r _ {i j}} \right]\right) \right] \tag {9} +$$ + +The attention protocol manifests the capability to discern crucial contextual features from language graph, serving as + +potent language priors for query selection. + +Fine-grained Query Selection. We input the candidate query features $\mathbf{Q}_c\in \mathbb{R}^{n\times d}$ from the cross-modal encoding layer into the Query Embedding component, which consists of several stacked Fully-connected (FC) layers and BN layers. Further, we leverage the fine-grained text features $T_{f}$ to guide the query selection through a cross-attention fusion layer. The resulting query tokens are fed into an MLP layer for confidence score prediction. Through selecting the top- $K$ highest scoring tokens for the following decoder, we derive the updated object queries $\mathbf{Q}_s\in \mathbb{R}^{K\times d}$ , which are utilized to predict a box center, height and width. + +# 3.4. Poincaré Semantic Alignment loss + +We linearly embed the queries $\mathbf{Q}_s$ to the predicted object features $\pmb{q} \in \mathbb{R}^{K \times 64}$ , which aligns with the embedded text features $\pmb{t} \in \mathbb{R}^{t \times 64}$ . Noteworthy, point clouds (spanning from scenes, objects to parts) and texts (ranging from sentences, phrases to words) encompass inherent hierarchy. Further, granularity disparities of geometric and semantic representations complicate the intricate correlations. However, it is the Achilles' heel to model hierarchical structures and complex relationships in canonical Euclidean space. + +To tackle this issue, we provide the insight to map vision-text features to a special hyperbolic space—Poincaré ball, for its powerful natural aptitude in modeling tree-shaped features. Based on this, we carefully devise a PSA loss to pave the way for semantic-geometric consistency. Benefiting from exponential volume expansion of hyperbolic embedding, PSA explicitly models complex semantic-geometric relationship for capturing hierarchical structures. + +Since the Poincaré ball is a Riemannian manifold, we project the vision $\pmb{q}$ and text features $\pmb{t}$ to the Poincaré ball $\mathbb{D}_c^d$ with the curvature $c$ via an exponential mapping function $\exp_{\pmb{x}}^{c}: \mathbb{R}^{d} \to \mathbb{D}_{c}^{d}$ , which can be defined as follows: + +$$ +\boldsymbol {q} ^ {\mathrm {c}} = \exp_ {\boldsymbol {x}} ^ {\mathrm {c}} (\boldsymbol {q}) = \boldsymbol {x} \oplus_ {c} \left(\tanh \left(\sqrt {c} \frac {\lambda_ {\boldsymbol {x}} ^ {\mathrm {c}} \| \boldsymbol {q} \|}{2}\right) \frac {\boldsymbol {q}}{\sqrt {c} \| \boldsymbol {q} \|}\right) \tag {10} +$$ + +where $\pmb{x}$ is a fixed base point to make formulas less cumbersome and $\lambda_{\pmb{x}}^{c} = 2 / (1 - c\| \pmb{x}\|^{2})$ . $t^c$ is defined in a similar approach. $\oplus_{c}$ denotes the addition operator in the Mobius gyrovector space. The geodesic distance between the a vision-text feature pair $(\pmb{q}_i^c,\pmb{t}_j^c)$ on the Poincaré ball $\mathbb{D}_c^d$ is: + +$$ +d _ {c} \left(\boldsymbol {q} _ {i} ^ {c}, \boldsymbol {t} _ {j} ^ {c}\right) = \frac {2}{\sqrt {c}} \operatorname {a r c t a n h} \left(\sqrt {c} \| - \boldsymbol {q} _ {i} ^ {c} \oplus_ {c} \boldsymbol {t} _ {j} ^ {c} \|\right) \tag {11} +$$ + +Inspired by [43], we term the PSA loss to densely align the fine-grained vision-text features via contrastive learning on Poincaré ball. The query loss is defined as: + +$$ +\mathcal {L} _ {\mathrm {p s a}} ^ {q} = \sum_ {i = 1} ^ {K} \frac {- 1}{\left| \mathcal {T} _ {i} ^ {+} \right|} \sum_ {\boldsymbol {t} _ {i} ^ {c} \in \mathcal {T} _ {i} ^ {+}} \log \left[ \right. \frac {\exp \left[ \boldsymbol {w} _ {+} \left(- d _ {c} \left(\boldsymbol {q} _ {i} ^ {c} , \boldsymbol {t} _ {i} ^ {c}\right) / \tau\right)\right]}{\sum_ {j = 1} ^ {l} \exp \left[ \boldsymbol {w} _ {-} \left(- d _ {c} \left(\boldsymbol {q} _ {i} ^ {c} , \boldsymbol {t} _ {j} ^ {c}\right) / \tau\right)} \right] \tag {12} +$$ + +where $k$ and $l$ are the number of object tokens and text to- + +
MethodVenueInputUnique (~19%)Multiple (~81%)Overall
0.250.50.250.50.250.5
ScanRefer [6]ECCV'203D+2D76.3353.5132.7321.1141.1927.40
TGNN [16]AAAI'213D68.6156.8029.8423.1837.3729.70
InstanceRefer [47]ICCV'213D77.4566.8331.2724.7740.2332.93
SAT [44]ICCV'213D+2D73.2150.8337.6425.1644.5430.14
FFL-3DOG [12]ICCV'213D78.8067.9435.1925.741.3334.01
3DVG [48]ICCV'213D+2D81.9360.6439.328.4247.5734.67
3D-SPS [31]CVPR'223D+2D84.1266.7240.3229.8248.8236.98
BUTD-DETR [19]ECCV'223D82.8864.9844.7333.9750.4238.60
ViL3DRel [7]NeurlPS'223D81.5868.6240.3030.7147.9437.73
EDA [43]CVPR'233D85.7668.5749.1337.6454.5942.26
3DJCG [5]CVPR'223D+2D83.3764.3441.3930.8249.5637.33
3D-VLP [21]CVPR'233D+2D84.2364.6143.5133.4151.4139.46
3D-VisTA(Scratch) [51]ICCV'233D77.4070.9038.7034.8045.9041.50
G3-LQ(Two-Stage)——3D88.0972.7351.4840.8056.9045.58
3D-SPS(One-stage) [31]CVPR'223D81.6364.7739.4829.6147.6536.43
BUTD-DETR(One-stage) [19]ECCV'223D81.4761.2444.2032.8149.7637.05
EDA(One-stage) [43]CVPR'233D86.4069.4248.1136.8253.8341.70
G3-LQ(One-stage)——3D88.5973.2850.2339.7255.9544.72
+ +kens, $\tau$ is the temperature coefficient. $\pmb{w}^{+}$ $w^{-}$ are weights of positive and negative terms. Similar to EDA, $\pmb{t}_i$ is the positive text features of the $i$ -th object query and $\mathcal{T}_i^+$ are the positive features set of $\pmb{t}_i$ , which primarily includes the object entries, attributes, relations, and pronouns. Moreover, the text loss can be formulated similarly: + +$$ +\mathcal {L} _ {\mathrm {p s a}} ^ {t} = \sum_ {i = 1} ^ {l} \frac {- 1}{\left| \mathcal {Q} _ {i} ^ {+} \right|} \sum_ {\boldsymbol {q} _ {i} ^ {c} \in \mathcal {Q} _ {i} ^ {+}} \log \left[ \right. \frac {\exp \left[ \boldsymbol {w} _ {+} \left(- d _ {c} \left(\boldsymbol {t} _ {i} ^ {c} , \boldsymbol {q} _ {i} ^ {c}\right) / \tau\right)\right]}{\sum_ {j = 1} ^ {K} \exp \left[ \boldsymbol {w} _ {-} \left(- d _ {c} \left(\boldsymbol {t} _ {i} ^ {c} , \boldsymbol {q} _ {j} ^ {c}\right) / \tau\right)} \right] \tag {13} +$$ + +where $\pmb{q}_i^c\in \mathcal{Q}_i^+$ is the positive object features of $\pmb{t}_i^c$ . The final loss $\mathcal{L}_{\mathrm{psa}}$ is the average of the above two terms. + +Following [43], the total loss of our method also includes the box regression loss $\mathcal{L}_{\mathrm{loc}}$ , the position alignment loss $\mathcal{L}_{\mathrm{pa}}$ which details in the Supplementary Material. + +# 4. Experiment + +Firstly, we carry out comparisons with SOTA methods on ScanRefer and $\mathrm{Nr3D / Sr3D}$ dataset in Sec. 4.1. Further, we delve into ablation studies in Sec. 4.2. Finally, we visualize and analyze the 3D VG results in Sec. 4.3. + +# 4.1. Quantitative Comparisons + +Performance on the ScanRefer. As illustrated in Table. 1, our method consistently exhibits admirably performance, trumping all competitors by a substantial margin on all test subsets. Our method achieves $56.90\%$ and $45.58\%$ performance in terms of the overall accuracy, with a remarkable improvement compared with EDA [43], $2.31\%$ and $2.34\%$ . 1) Compared with 2D assistance methods [17, 44] + +Table 1. Comparison results on the ScanRefer dataset, in terms of the accuracy evaluated by IoU 0.25 and IoU 0.5. The unique denotes samples devoid of distracting objects, while multiple applies to remaining samples. In the one-stage setting, there is no reliance on the additional 3D object detection step. Our proposed $\mathrm{G}^3 -\mathrm{LQ}$ performs favorably over current state-of-the-art approach EDA [43]. + +
MethodNr3DSr3D
OverallHardOverallHard
TGNN [16]37.330.645.036.9
InstanceRefer [47]38.831.848.040.5
3DVG [48]40.834.851.444.9
LanguageRefer [39]43.936.656.049.3
TransRefer3D [15]48.039.657.450.2
SAT [44]49.242.457.950.0
LAR [4]48.942.359.451.2
3DRef [2]47.038.339.032.0
3D-SPS [31]51.545.162.665.4
MVT [17]55.149.164.558.8
BUTD-DETR [19]54.648.467.063.2
EDA [43]52.146.168.162.9
3D-VisTA [51]57.549.469.663.6
G3-LQ58.450.773.166.3
+ +Table 2. Quantitative comparisons on $Nr3D$ and $Sr3D$ dataset. We have highlighted the top-performing three methods in purple. + +that furnish 2D image priors imbued with elaborate semantics and textures, the proposed $\mathrm{G}^3$ -LQ method showcases its superiority in performance. This serves as further evidence that our method explicitly captures geometric features, thus facilitating the shape understanding of target object and the vision-text alignment. 2) Our approach, featuring fine-grained language-guided query selection, surpasses all recently published DETR-like models [19, 43]. To explain, the proposed Flan-QS module thoroughly exploits fine-grained language prior and captures global context dependency, which effectively mitigates the issues of grounding ambiguity. 3) Compared with the 3D pre-trained + +
IDPAGEFlan-QSPSAUniqueMultiple
0.250.50.250.5
(a)———85.6768.5749.1337.64
(b)✓——86.8970.2650.4638.61
(c)—✓—86.7569.7650.1138.58
(d)——✓86.3269.2049.5838.18
(e)✓—✓87.4571.9551.0840.49
(f)—✓✓87.3170.7550.9040.19
(g)✓✓—87.6672.1650.8740.05
(h)✓✓✓88.0972.7351.4840.80
+ +models [21, 51], $\mathrm{G}^3$ -LQ achieves SOTA accuracy, because of the semantic-geometric alignment in Poincaré space with the tailored PSA loss, which aids in modeling the complex relationships of vision-text features. 4) The "multiple" setting entails a 3D scene where the description distinguishes the target object amidst numerous distractors. Under this setting, we reach a promising performance of $51.14\%$ and $40.08\%$ , providing further validation for the efficacy of $\mathrm{G}^3$ -LQ in addressing the Semantic-Geometric Coupling issue. + +Performance on the Nr3D/Sr3D. The Nr3D/Sr3D task is geared towards locating the target object among all provided ground truth candidate boxes, representing a departure from the ScanRefer dataset. The overall accuracy of the proposed G $^3$ -LQ along with other exceptional methods are reported in Table. 2. We attain peak performance of $58.4\%$ and $73.1\%$ on Nr3D and Sr3D. In Nr3D, descriptions exhibit noteworthy intricacy and detail, inducing additional challenges to 3D VG task. However, our method demonstrates a noteworthy superiority over alternative methods that rely on 2D images priors [44] or 3D vision-language pre-training models [51]. In the more challenging "Hard" subset, our method notably improves the accuracy by $1.3\%$ in Nr3D and $2.7\%$ in Sr3D, which underscores our method is beneficial for distinguishing ambiguous objects by effectively modeling the underlying geometric shape and unraveling the complex relation between vision-text features. + +# 4.2. Ablation Study and Analysis + +Effectiveness of Proposed Components. We develop several alternative designs and undertake experiments to highlight the advantages yielded by various components. We herein train our $\mathrm{G}^3$ -LQ model on the ScanRefer dataset. (a) outlines the performance demonstrated by EDA [43], while (h) is our proposed method. 1) Comparing (a) and (b), we witness that the PAGE module delivers a remarkable gain over the baseline model. The comparison further accentuates the paramount role of explicit 3D geometric features in modeling object shape and spatial relationship, conferring benefits to 3D visual grounding. 2) The comparative result of (b) and (g) furnishes compelling evidence that the Flan + +Table 3. Ablation study on the effectiveness of the proposed PAGE, Flan-QS and PSA. Evaluated on the ScanRefer dataset. + +
pipi - pijpi - pijUnique(0.25)Multiple(0.25)
✓——85.6749.13
✓✓—85.8349.19
✓—maxpool85.9849.32
✓—softmax86.0449.37
✓—Afine86.5750.06
✓✓Afine86.8950.46
+ +Table 4. Ablation study of the geometric priors selection in the proposed PAGE module on the ScanRefer dataset. + +
Query SelectionUnique(~19%)Multiple(~81%)
0.250.50.250.5
Parametric81.4060.1338.0228.05
Top-K(EDA)85.6768.5749.1337.64
Language87.1071.5950.6339.31
Language Graph87.6672.1650.8740.05
+ +Table 5. Ablation study of the query selection. "Parametric" represents the selection strategy with random parametric queries. "Language" denotes the guidance of sentence-level language. + +QS module exploiting language prior and context dependency, helps to alleviate the problem of grounding ambiguity and generate precise query proposals. 3) Comparing (h) and (g), it showcases the effectiveness of the proposed PSA loss to capture complex non-linear mappings of point cloud and text embeddings on Poincaré ball. When we integrate all proposed modules, a discernible surge is observed, yielding an impressive overall accuracy of $56.90\%$ and $45.58\%$ . Selection of Geometric Priors. The fundamental objective of the PAGE module lies in embedding low-level geometric priors into high-level geometric features, thus the definition of $\tilde{p}$ emerges as a topic deserving exploration. We undertake ablations with $p_i, p_i - p_{ij}, \| p_i - p_{ij} \|$ as illustrative examples, which unveils the inherent relative position, shapes, and distance (see Fig. 3) among 3D objects. As shown in Table. 4, using $\| p_i - p_{ij} \|$ or $p_i - p_{ij}$ alone yields inferior performance compared with their gradual integration, the accuracy of which reach $86.89\%$ (Uniqu) and $50.46\%$ (Multiple) respectively. Further, when compared with max pooling and softmax operators for features aggregation in $\mathcal{X}_i$ , our geometric refinement strategy attests to an improved performance by $0.59\%$ and $0.59\%$ (Unique). The noteworthy results accentuate the significance of the geometric refinement in the PAGE module, given its potential to capture position-adaptive geometric structures. + +Methods of Query Selection. We carry out several experiments to validate the proposed query selection method. We delineate the overall performance on the ScanRefer dataset in Table. 5. 1) We adopt the parametric queries used in MDETR [22]. However, we observe that the random parametric queries in performance decline for its incapability to capture contextual information of the text and 3D scene. 2) + +![](images/afa2de50b5088c37cdffed8fbad6598f569c7524c74f7b598ac6ae9a0aabd718.jpg) +(a) Ordinal Numbers +this is a chair by the wall. It is the fifth chair from the right wall surface. + +![](images/2f43e6854e8502f292069b36f782546485f86068fa07d523221d3719171f87c6.jpg) +(b) Distance and Size + +![](images/16400971a747609d2f1074db44477d3041827196f210d758be476fb7920178de.jpg) + +![](images/ae0582df79992fb5178ef53b9d48cd314cb0a681bdacb7883bbc7e1f1e67db1e.jpg) +this is a grey chair. its at the end of the shorter table closest to the window. +we are looking for the chair on the right. we are not looking for the chair on the left. the chair on the left is closest to the bed. + +![](images/f568db9e3a1ffd529047e19383418b3901b3e3172830242fef75bbefa271bd67.jpg) +it is a black office chair. The black office chair is the fourth office chair on the left side of the table. +there is a gray chair. it is the second closest one to a trash can. + +![](images/8d92fcdc30592e09814481c6db15b0a894a2c67532829771b5a2f34cc790ece2.jpg) +the table is the left-most one in the center of the room. The table is dark orange and has four legs. + +![](images/3e39a366eda7b48c0f14f033612ddafe87c082bbfb98d25a3cc5dda28c4e36ce.jpg) +(c) Geometric shape +standing semi-circular wooden table with flat surface for holding objects. object can be found in center of image +in front of a brown wooden desk are two white chairs. we are looking for the chair on the right. we are not looking for the chair on the left. the chair on the left is closest to the bed. + +![](images/4cbc3c1692187fd558fc3991e33f777105f30c9d3a86cf10ab7f55e7e2edc167.jpg) + +![](images/cc430a091aaf8b0aa4e914b62a54bc497e4ef0cf530ade1159608d2ed4ee9fce.jpg) +The towel is between the entrance door and the sink. The towel is olive green and wavy. +the end table is the closest one to the window. The end table is a round cylinder. +Figure 3. Qualitative results of the EDA [43] and our proposed $\mathrm{G}^3$ -LQ on the ScanRefer dataset. Our $\mathrm{G}^3$ -LQ manifests excellent grounding performance in terms of ordinal numbers, spatial distance or object size, geometric attribute and complex utterances. + +![](images/b17cc1dc1cc08e9c56d214e641f83e7bf89334ce1eb4550da8c2a45d5122ef74.jpg) +(d) Complex Utterance + +![](images/aa76ba1a2154fb8d23ab8babd93593659a8921ce769d955f5eb2a025fa5b6c9c.jpg) +this is the brown arm chair that is at the end of the chalk board, to the left of it there is a small bookshelf, to the right of it and at the table is a blue chair. The brown chair is not under the table it is just next to the table. + +![](images/a228d9520a5e582e20fe4624395155954ed62515275a16cc27de00e1472903a4.jpg) +walking into the room, a large office desk is in the middle of the room. On the right side of the desk, there is a computer monitor on the table, the one on the left side, closest to the door. + +We generate the query proposals guided by the sentence-level features. The performance of the language-guided query selection mechanism stands out prominently, underscoring the crucial role of text features in the query generation. 3) Our Flan-QS module showcases superior performance over the prevailing top- $K$ selection and sentence-level methods. The results give explanation of the efficacy of explicit fine-grained text features. They precisely model the attributes and positions of target objects, thereby facilitating a more accurate generation of query proposals. + +# 4.3. Qualitative Comparisons + +To offer profound insight into the effectiveness of the proposed $\mathrm{G}^3$ -LQ, we visualize the grounding results of our method alongside the SOTA EDA [43]. As shown in Fig. 3, the predicted boxes of our $\mathrm{G}^3$ -LQ method and EDA are visually highlighted in yellow and red, with ground-truth boxes marked in green. The localization ability of EDA is found wanting, encountering challenges in handling the 3D objects described by ordinal numbers, relative distance, and geometric shapes (see Fig. 3(a)-(c)). Conversely, our $\mathrm{G}^3$ -LQ incorporates intrinsic geometric features and spatial relation, fostering an enriched understanding of both the scene and objects, manifests commendable grounding prowess. Secondly, Fig. 3(d) exemplifies the commendable capabilities of our method in grounding 3D objects narrated by complex utterances. To elucidate, the $\mathrm{G}^3$ -LQ model with the tailored Flan-QS, excels in harnessing language priors and capturing global context dependencies, generat- + +ing high-quality semantic-aware query representations. + +# 5. Conclusion + +In this paper, we analysis the Semantic-Geometric Coupling issue of 3D VG tasks in three aspects: intricate geometric details, complex text descriptions, and semantic-geometric inconsistency. To this end, we propose a $\mathrm{G}^3$ -LQ framework to explicitly capture the fine-grained geometric and semantic concepts, which includes three crucial novelties. Firstly, a Position Adaptive Geometric Exploring module is designed to capture explicit geometric features. Secondly, with decoupled multiple semantic components, a Fine-grained Language-guided Query Selection module generates object queries densely aligned by fine-grained text features. Finally, a Poincaré Semantic Alignment loss capitalizes on vision-text hierarchy natures and achieves alignment in Poincaré space, encouraging semantic-geometric consistency. Experimental findings showcase the superior performance of our $\mathrm{G}^3$ -LQ across the ScanRefer and Nr3D/Sr3D benchmarks, establishing a new state-of-the-art standard. + +# 6. Acknowledgement + +This work is supported by the state key development program in 14th Five-Year under Grant No. 2021YFF0602103, 2021YFF0602102, 2021QY1702. Thanks for the research fund under Grant No. 2019GQG0001 from the Institute for Guo Qiang, Tsinghua University. This study is also supported by Tsinghua University Toyota Joint Research Center for AI Technology of Automated Vehicle(TTAD 2023-07). + +# References + +[1] Scene graph parser. https://github.com/vacancy/SceneGraphParser, 2019.5 +[2] Ahmed Abdelreheem, Ujjwal Upadhyay, Ivan Skorokhodov, Rawan Al Yahya, Jun Chen, and Mohamed Elhoseiny. 3dref-transformer: Fine-grained object identification in real-world scenes using natural language. In Proceedings of the IEEE Winter Conference on Applications of Computer Vision, pages 3941-3950, 2022. 6 +[3] Panos Achlioptas, Ahmed Abdelreehem, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In Proceedings of the European Conference on Computer Vision, pages 422-440. Springer, 2020. 2 +[4] Eslam Bakr, Yasmeen Alsaeda, and Mohamed Elhoseiny. Look around and refer: 2d synthetic semantics knowledge distillation for 3d visual grounding. Advances in Neural Information Processing Systems, 35:37146-37158, 2022. 6 +[5] Daigang Cai, Lichen Zhao, Jing Zhang, Lu Sheng, and Dong Xu. 3djcg: A unified framework for joint dense captioning and visual grounding on 3d point clouds. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 16464-16473, 2022. 2, 6 +[6] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In Proceedings of the European Conference on Computer Vision, pages 202-221. Springer, 2020. 1, 6 +[7] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Language conditioned spatial relation reasoning for 3d object grounding. Advances in Neural Information Processing Systems, 35:20522-20535, 2022. 6 +[8] Jaesung Choe, Chunghyun Park, Francois Rameau, and Jaesik Park. Pointmixer: Mlp-mixer for point cloud understanding. In Proceedings of the European Conference on Computer Vision, pages 620-640. Springer, 2022. 3 +[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2 +[10] Jiafei Duan, Samson Yu, Hui Li Tan, Hongyuan Zhu, and Cheston Tan. A survey of embodied ai: From simulators to research tasks. IEEE Transactions on Emerging Topics in Computational Intelligence, 6(2):230-244, 2022. 1 +[11] Aleksandr Ermolov, Leyla Mirvakhabova, Valentin Khrulkov, Nicu Sebe, and Ivan Oseledets. Hyperbolic vision transformers: Combining improvements in metric learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7409-7419, 2022. 2 +[12] Mingtao Feng, Zhen Li, Qi Li, Liang Zhang, XiangDong Zhang, Guangming Zhu, Hui Zhang, Yaonan Wang, and Ajmal Mian. Free-form description guided 3d visual graph network for object grounding in point cloud. In Proceedings of the IEEE International Conference on Computer Vision, pages 3722-3731, 2021. 2, 6 + +[13] Octavian Ganea, Gary Bécigneul, and Thomas Hofmann. Hyperbolic neural networks. Advances in Neural Information Processing Systems, 31, 2018. 2 +[14] William L Hamilton, Rex Ying, and Jure Leskovec. Inductive representation learning on large graphs. In Advances in Neural Information Processing Systems, pages 1025-1035, 2017. 5 +[15] Dailan He, Yusheng Zhao, Junyu Luo, Tianrui Hui, Shaofei Huang, Aixi Zhang, and Si Liu. Transrefer3d: Entity-andrelation aware transformer for fine-grained 3d visual grounding. In Proceedings of the 29th ACM International Conference on Multimedia, pages 2344-2352, 2021. 2, 3, 6 +[16] Pin-Hao Huang, Han-Hung Lee, Hwann-Tzong Chen, and Tyng-Luh Liu. Text-guided graph neural networks for referring 3d instance segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1610-1618, 2021. 2, 6 +[17] Shijia Huang, Yilun Chen, Jiaya Jia, and Liwei Wang. Multiview transformer for 3d visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15524-15533, 2022. 6 +[18] Wenlong Huang, Pieter Abbeel, Deepak Pathak, and Igor Mordatch. Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. In International Conference on Machine Learning, pages 9118-9147. PMLR, 2022. 1 +[19] Ayush Jain, Nikolaos Gkanatsios, Ishita Mediratta, and Katerina Fragkiadaki. Bottom up top down detection transformers for language grounding in images and point clouds. In Proceedings of the European Conference on Computer Vision, pages 417-433. Springer, 2022. 2, 3, 6 +[20] Haojun Jiang, Yuanze Lin, Dongchen Han, Shiji Song, and Gao Huang. Pseudo-q: Generating pseudo language queries for visual grounding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 15513-15523, 2022. 1 +[21] Zhao Jin, Munawar Hayat, Yuwei Yang, Yulan Guo, and Yinjie Lei. Context-aware alignment and mutual masking for 3d-language pre-training. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 10984-10994, 2023. 2, 6, 7 +[22] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE International Conference on Computer Vision, pages 1780-1790, 2021. 3, 7 +[23] Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In International Conference on Learning Representations, 2017. 5 +[24] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329, 2022. 5 +[25] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 3, 5 + +[26] Yongcheng Liu, Bin Fan, Shiming Xiang, and Chunhong Pan. Relation-shape convolutional neural network for point cloud analysis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8895-8904, 2019. 3 +[27] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 2, 3 +[28] Yongfei Liu, Bo Wan, Xiaodan Zhu, and Xuming He. Learning cross-modal context graph for visual grounding. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11645-11652, 2020. 1 +[29] Yongfei Liu, Bo Wan, Lin Ma, and Xuming He. Relation-aware instance refinement for weakly supervised visual grounding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5612-5621, 2021. 1 +[30] Ze Liu, Zheng Zhang, Yue Cao, Han Hu, and Xin Tong. Group-free 3d object detection via transformers. In Proceedings of the IEEE International Conference on Computer Vision, pages 2949-2958, 2021. 2 +[31] Junyu Luo, Jiahui Fu, Xianghao Kong, Chen Gao, Haibing Ren, Hao Shen, Huaxia Xia, and Si Liu. 3d-sps: Single-stage 3d visual grounding via referred point progressive selection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 16454-16463, 2022. 1, 2, 3, 6 +[32] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. In International Conference on Learning Representations, 2021. 3 +[33] Antonio Montanaro, Diego Valsesia, and Enrico Magli. Rethinking the compositionality of point clouds through regularization in the hyperbolic space. Advances in Neural Information Processing Systems, 35:33741-33753, 2022. 2 +[34] Maximillian Nickel and Douwe Kiela. Poincaré embeddings for learning hierarchical representations. Advances in Neural Information Processing Systems, 30, 2017. 2 +[35] Aishwarya Padmakumar, Jesse Thomason, Ayush Shrivastava, Patrick Lange, Anjali Narayan-Chen, Spandana Gella, Robinson Piramuthu, Gokhan Tur, and Dilek Hakkani-Tur. Teach: Task-driven embodied agents that chat. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2017–2025, 2022. 1 +[36] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in Neural Information Processing Systems, 30, 2017. 3 +[37] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In proceedings of the IEEE International Conference on Computer Vision, pages 9277-9286, 2019. 2 +[38] Zheng Qin, Hao Yu, Changjian Wang, Yulan Guo, Yuxing Peng, and Kai Xu. Geometric transformer for fast and robust point cloud registration. In Proceedings of the IEEE Con- + +ference on Computer Vision and Pattern Recognition, pages 11143-11152, 2022. 3 +[39] Junha Roh, Karthik Desingh, Ali Farhadi, and Dieter Fox. Languagerefer: Spatial-language model for 3d visual grounding. In Conference on Robot Learning, pages 1046-1056. PMLR, 2022. 3, 6 +[40] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 30, 2017. 2 +[41] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (TOG), 38(5):1-12, 2019. 3 +[42] Yingming Wang, Xiangyu Zhang, Tong Yang, and Jian Sun. Anchor detr: Query design for transformer-based detector. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2567-2575, 2022. 5 +[43] Yanmin Wu, Xinhua Cheng, Renrui Zhang, Zesen Cheng, and Jian Zhang. Eda: Explicit text-decoupling and dense alignment for 3d visual and language learning. arXiv preprint arXiv:2209.14941, 2022. 1, 2, 3, 5, 6, 7, 8 +[44] Zhengyuan Yang, Songyang Zhang, and Jiebo Luo. Sat: 2d semantics assisted training for 3d visual grounding. In Proceedings of the IEEE International Conference on Computer Vision, pages 1856-1866, 2021. 2, 6, 7 +[45] Ting Yao, Yehao Li, and Tao Mei. Hgnet: Learning hierarchical geometry from points, edges, and surfaces. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 21846-21855, 2023. 3 +[46] Xumin Yu, Yongming Rao, Ziyi Wang, Zuyan Liu, Jiwen Lu, and Jie Zhou. Pointr: Diverse point cloud completion with geometry-aware transformers. In Proceedings of the IEEE International Conference on Computer Vision, pages 12498-12507, 2021. 3 +[47] Zhihao Yuan, Xu Yan, Yinghong Liao, Ruimao Zhang, Sheng Wang, Zhen Li, and Shuguang Cui. Instancerefer: Cooperative holistic understanding for visual grounding on point clouds through instance multi-level contextual referring. In Proceedings of the IEEE International Conference on Computer Vision, pages 1791-1800, 2021. 6 +[48] Lichen Zhao, Daigang Cai, Lu Sheng, and Dong Xu. 3dvg-transformer: relation modeling for visual grounding on point clouds. In Proceedings of the IEEE International Conference on Computer Vision, pages 2928-2937, 2021. 1, 2, 3, 6 +[49] Chaoyang Zhu, Yiyi Zhou, Yunhang Shen, Gen Luo, Xingjia Pan, Mingbao Lin, Chao Chen, Liujuan Cao, and Xiaoshuai Sun. Seqtr: A simple yet universal network for visual grounding. In Proceedings of the European Conference on Computer Vision, pages 598-615. Springer, 2022. 1 +[50] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 5 +[51] Ziyu Zhu, Xiaojian Ma, Yixin Chen, Zhidong Deng, Siyuan Huang, and Qing Li. 3d-vista: Pre-trained transformer for 3d vision and text alignment. arXiv preprint arXiv:2308.04352, 2023. 2, 6, 7 \ No newline at end of file diff --git a/2024/G^3-LQ_ Marrying Hyperbolic Alignment with Explicit Semantic-Geometric Modeling for 3D Visual Grounding/images.zip b/2024/G^3-LQ_ Marrying Hyperbolic Alignment with Explicit Semantic-Geometric Modeling for 3D Visual Grounding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8937b2625321927c318d2314a9da5091fdb24e02 --- /dev/null +++ b/2024/G^3-LQ_ Marrying Hyperbolic Alignment with Explicit Semantic-Geometric Modeling for 3D Visual Grounding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5fef80c37134d0104d2cf64a82ee2e579df0d86be39796651365e34a58527afd +size 780815 diff --git a/2024/G^3-LQ_ Marrying Hyperbolic Alignment with Explicit Semantic-Geometric Modeling for 3D Visual Grounding/layout.json b/2024/G^3-LQ_ Marrying Hyperbolic Alignment with Explicit Semantic-Geometric Modeling for 3D Visual Grounding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ef8eec7cd3cd98605570a6e811d85e86735ba6b9 --- /dev/null +++ b/2024/G^3-LQ_ Marrying Hyperbolic Alignment with Explicit Semantic-Geometric Modeling for 3D Visual Grounding/layout.json @@ -0,0 +1,10705 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 64, + 102, + 530, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 102, + 530, + 140 + ], + "spans": [ + { + "bbox": [ + 64, + 102, + 530, + 140 + ], + "type": "inline_equation", + "content": "\\mathbf{G}^{3}" + }, + { + "bbox": [ + 64, + 102, + 530, + 140 + ], + "type": "text", + "content": "-LQ: Marrying Hyperbolic Alignment with Explicit Semantic-Geometric Modeling for 3D Visual Grounding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 184, + 160, + 409, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 160, + 409, + 175 + ], + "spans": [ + { + "bbox": [ + 184, + 160, + 409, + 175 + ], + "type": "text", + "content": "Yuan Wang" + }, + { + "bbox": [ + 184, + 160, + 409, + 175 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 184, + 160, + 409, + 175 + ], + "type": "text", + "content": " Yali Li" + }, + { + "bbox": [ + 184, + 160, + 409, + 175 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 184, + 160, + 409, + 175 + ], + "type": "text", + "content": " Shengjin Wang" + }, + { + "bbox": [ + 184, + 160, + 409, + 175 + ], + "type": "inline_equation", + "content": "^{1,2\\ddagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 175, + 460, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 175, + 460, + 190 + ], + "spans": [ + { + "bbox": [ + 133, + 175, + 460, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 133, + 175, + 460, + 190 + ], + "type": "text", + "content": "Department of Electronic Engineering, Tsinghua University, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 65, + 189, + 525, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 189, + 525, + 219 + ], + "spans": [ + { + "bbox": [ + 65, + 189, + 525, + 219 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 65, + 189, + 525, + 219 + ], + "type": "text", + "content": "Beijing National Research Center for Information Science and Technology (BNRist), China wy23@mails.tsinghua.edu.cn {liyali13, wgsgj} @tsinghua.edu.cn " + }, + { + "bbox": [ + 65, + 189, + 525, + 219 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 65, + 189, + 525, + 219 + ], + "type": "text", + "content": "Corresponding Author" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 57, + 247, + 538, + 379 + ], + "blocks": [ + { + "bbox": [ + 57, + 247, + 538, + 379 + ], + "lines": [ + { + "bbox": [ + 57, + 247, + 538, + 379 + ], + "spans": [ + { + "bbox": [ + 57, + 247, + 538, + 379 + ], + "type": "image", + "image_path": "797fc1d01a607893704a400430e6eea3bd193b90bcef460b710919ce8241dddd.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 388, + 546, + 412 + ], + "lines": [ + { + "bbox": [ + 46, + 388, + 546, + 412 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 546, + 412 + ], + "type": "text", + "content": "Figure 1. Illustration of the proposed G" + }, + { + "bbox": [ + 46, + 388, + 546, + 412 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 46, + 388, + 546, + 412 + ], + "type": "text", + "content": "-LQ method on the ScanRefer benchmark. The versatile scene-level and object-level understanding properties empower G" + }, + { + "bbox": [ + 46, + 388, + 546, + 412 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 46, + 388, + 546, + 412 + ], + "type": "text", + "content": "-LQ to comprehend spatial relationship, relative distance, geometric attribute as well as intricate references." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 422, + 192, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 422, + 192, + 434 + ], + "spans": [ + { + "bbox": [ + 143, + 422, + 192, + 434 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 45, + 438, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 438, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 45, + 438, + 289, + 714 + ], + "type": "text", + "content": "Grounding referred objects in 3D scenes is a burgeoning vision-language task pivotal for propelling Embodied AI, as it endeavors to connect the 3D physical world with free-form descriptions. Compared to the 2D counterparts, challenges posed by the variability of 3D visual grounding remain relatively unsolved in existing studies: 1) the underlying geometric and complex spatial relationships in 3D scene. 2) the inherent complexity of 3D grounded language. 3) the inconsistencies between text and geometric features. To tackle these issues, we propose " + }, + { + "bbox": [ + 45, + 438, + 289, + 714 + ], + "type": "inline_equation", + "content": "G^3" + }, + { + "bbox": [ + 45, + 438, + 289, + 714 + ], + "type": "text", + "content": "-LQ, a DEtection Transformer-based model tailored for 3D visual grounding task. " + }, + { + "bbox": [ + 45, + 438, + 289, + 714 + ], + "type": "inline_equation", + "content": "G^3" + }, + { + "bbox": [ + 45, + 438, + 289, + 714 + ], + "type": "text", + "content": "-LQ explicitly models Geometric-aware visual representations and Generates fine-Grained Language-guided object Queries in an overarching framework, which comprises two dedicated modules. Specifically, the Position Adaptive Geometric Exploring (PAGE) unearths underlying information of 3D objects in the geometric details and spatial relationships perspectives. The Fine-grained Language-guided Query Selection (Flan-QS) delves into syntactic structure of texts and generates object queries that exhibit higher relevance towards fine-grained text features. Finally, a pioneering Poincaré Semantic Alignment (PSA) loss establishes semantic-geometry consistencies by mod" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 422, + 546, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 422, + 546, + 471 + ], + "spans": [ + { + "bbox": [ + 305, + 422, + 546, + 471 + ], + "type": "text", + "content": "eling non-linear vision-text feature mappings and aligning them on a hyperbolic prototype—Poincaré ball. Extensive experiments verify the superiority of our " + }, + { + "bbox": [ + 305, + 422, + 546, + 471 + ], + "type": "inline_equation", + "content": "G^3" + }, + { + "bbox": [ + 305, + 422, + 546, + 471 + ], + "type": "text", + "content": "-" + }, + { + "bbox": [ + 305, + 422, + 546, + 471 + ], + "type": "inline_equation", + "content": "LQ" + }, + { + "bbox": [ + 305, + 422, + 546, + 471 + ], + "type": "text", + "content": " method, trumping the state-of-the-arts by a considerable margin." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 478, + 387, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 478, + 387, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 478, + 387, + 491 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 498, + 546, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 546, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 546, + 582 + ], + "type": "text", + "content": "Grounding 3D physical realm with natural languages has propelled to the forefront in advancing Embodied AI [10, 18, 35], an ability that empowers agents to comprehend human instructions in real-world contexts. 3D Visual Grounding (3D VG) [6, 31, 43, 48] has garnered substantial attention as a crucial cross-modal 3D perception task in defacto applications, e.g., assistive robots, AR/VR, and metaverse." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "text", + "content": "Compared with extensive explorations of 2D counterpart [20, 28, 29, 49], 3D VG poses prominent challenge known as \"Semantic-Geometric Coupling\". As shown in Fig. 1, we offer distinctive glimpse into it: 1) Intricate geometric details and spatial relations. 3D scenes encounter multiple objects with intricate layout, multi-level occlusion, and multifaceted spatial relation. 2) Heightened flexibility and complexity of descriptions. Free-form utterances with complex syntax convey precise spatial information, intricate references, and detailed relational terms. 3) Inconsistency of the semantic-geometric features. A single semantic concep" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13917" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "text", + "content": "tion yields multiple potential correspondences within point cloud or exhibit diverse geometry in varied contexts." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 98, + 289, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 98, + 289, + 397 + ], + "spans": [ + { + "bbox": [ + 46, + 98, + 289, + 397 + ], + "type": "text", + "content": "To ameliorate these issues, some pioneering methods have been proposed. They have largely focused on enhancing the representation capacity of 3D point clouds with 2D image priors [44]; capturing spatial relationships of 3D objects with powerful transformer [5, 15, 48] or graph convolution network [3, 12]; generating discriminative object proposals through instance segmentation [16] or linguistic guidance [31]. Despite commendable advancements be achieved, an equally important yet underexplored problem is explicit semantic-geometric modeling. Prior studies exhibit a significant dearth in explicitly modeling the intrinsic geometric attributes and spatial layouts of 3D objects, accentuating the dilemma to associate text descriptions with grounding objects. Moreover, object queries contain rich instance characteristics. But the object proposal [15, 48] and query generation [19, 31, 43] in current DETR-like models overlook explicit fine-grained interaction between text and object features, which in turn limit the model's capacity to refer potential language clues, e.g., appearances, shapes, or textures. Finally, major 3D VG systems [19, 21, 43, 51] strive to establish the alignment between point cloud and semantic concepts in canonical Euclidean space. However, their efficacy falters in unraveling the complex non-linear relationships between multi-modal features and upholding the intrinsic hierarchy natures in 3D scenes and texts." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 399, + 287, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 287, + 447 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 287, + 447 + ], + "type": "text", + "content": "In this paper, our primary goal endeavors to explicitly delve into fine-grained representations of geometric and semantic concepts, then aligns two modalities in hyperbolic space. We thus devise an overarching G" + }, + { + "bbox": [ + 46, + 399, + 287, + 447 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 46, + 399, + 287, + 447 + ], + "type": "text", + "content": "-LQ framework:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 449, + 289, + 715 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 46, + 449, + 289, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 289, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 289, + 605 + ], + "type": "text", + "content": "1) For geometric feature modeling, we firstly propose a Proxy Adaptive Geometric Exploring (PAGE) module. PAGE is composed of the Point-Proxy Geometric Extraction and Proxy Adaptive Geometric Refinement component. The former explicitly capture the local geometric attributes (e.g., distance and orientation) in a 3D local proxy, while the latter adaptively refine the geometric structure according to the position of point proxy. The geometric features are fed into the geometric encoder for high-level embedding and further achieve the interaction with text features for cross-modality enhancement. In this fashion, the PAGE module facilitates a seamless alignment of text descriptions with intrinsic geometric properties of 3D objects." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "content": "2) For complex utterance understanding, inspired by the dependency parsing, we aim to explore context dependency and syntactic structure of texts for guiding query selection. To this end, we propose a Fine-grained Language-guided Query Selection (Flan-QS) module, comprising the Language Scene Graph and Fine-grained Query Selection components. We decouple the free-form utterances to tree-like semantic components with directed dependency relationships as edges. Then, we establish and update the lan" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "guage scene graph to represent the context-aware entry features via node/edge embedding and graph context learning. Finally, we employ the constructed language scene graph to steer the generation of subsequent object candidates that are densely-aligned with fine-grained linguistic attributes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 133, + 546, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 546, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 546, + 312 + ], + "type": "text", + "content": "3) For semantic-geometric consistency, we embed vision and text features to a hyperbolic prototype—Poincaré ball. The emerging appreciation [11, 13, 34] suggest that the hyperbolic space offers overwhelming superiority to handle data correlations. 3D scene (free-form text) exhibits an inherent hierarchical nature ranging from scene, object, part (from sentence, phrase, word). The hyperbolic space with negative curvature is the only space that can successfully embed the tree-shaped hierarchy [33]. Moreover, the compact Poincaré ball [13, 34] holds suitable geometric properties, enhancing its capacity to model non-linear mappings of vision and text features. Therefore, we design a Poincaré Semantic Alignment loss (PSA) to probe the complex mappings between vision and text features, facilitating the semantics-geometry consistency." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 312, + 470, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 312, + 470, + 323 + ], + "spans": [ + { + "bbox": [ + 317, + 312, + 470, + 323 + ], + "type": "text", + "content": "Our chief contributions are threefold:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 323, + 545, + 479 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 306, + 323, + 545, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 323, + 545, + 383 + ], + "spans": [ + { + "bbox": [ + 306, + 323, + 545, + 383 + ], + "type": "text", + "content": "- We roundly analysis the Semantic-Geometric Coupling of 3D VG task. Towards this issue, we propose a unified " + }, + { + "bbox": [ + 306, + 323, + 545, + 383 + ], + "type": "inline_equation", + "content": "\\mathbf{G}^3" + }, + { + "bbox": [ + 306, + 323, + 545, + 383 + ], + "type": "text", + "content": "-LQ framework to model intrinsic geometric features (PAGE module) and explicitly generate fine-grained object queries for 3D visual grounding (Flan-QS module)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 383, + 545, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 383, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 306, + 383, + 545, + 431 + ], + "type": "text", + "content": "- We propose a novel PSA loss to capture the non-linear relationship and underlying hierarchy of the vision-text features in the hyperbolic space, which further paves the way for semantics-geometry consistency learning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 431, + 545, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 431, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 306, + 431, + 545, + 479 + ], + "type": "text", + "content": "- A battery of experiments on ScanRefer and Nr3d/Sr3d datasets demonstrate the effectiveness of our G" + }, + { + "bbox": [ + 306, + 431, + 545, + 479 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 306, + 431, + 545, + 479 + ], + "type": "text", + "content": "-LQ method. G" + }, + { + "bbox": [ + 306, + 431, + 545, + 479 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 306, + 431, + 545, + 479 + ], + "type": "text", + "content": "-LQ has proven its mettle, outperforming all existing SOTA methods by a considerable margin." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 483, + 398, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 483, + 398, + 496 + ], + "spans": [ + { + "bbox": [ + 306, + 483, + 398, + 496 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 503, + 429, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 503, + 429, + 516 + ], + "spans": [ + { + "bbox": [ + 306, + 503, + 429, + 516 + ], + "type": "text", + "content": "2.1.3D Visual Grounding" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "text", + "content": "3D Visual Grounding has been an area of intense investigation that dedicates to pinpointing the target object in a 3D scene based on the text description. Prevailing methods are typically two-stage, which conforms to the detection-then-matching paradigm. Firstly, the 3D object detectors [30, 37, 48] are harnessed to yield candidate 3D object proposals, while the text features are encoded by language models [9, 27]. Secondly, these advanced methods perform a pivotal endeavor to align the vision and text features for target objects grounding. Among these, TGNN [16] utilizes the graph neural network to deduce spatial relationships among 3D object proposals, which are further enriched by text features. FFL-3DOG [12] captures the intramodal and cross-modal relationships of text descriptions and 3D point cloud via scene graphs interaction. Recent endeavors embrace the formidable transformers [40] to model" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13918" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": "the relationship of proposals and text features. 3DVG [48] and TransRefer3D [15] achieve exceptional performance by establishing self-attention and cross-attention protocols for context perception. LanguageRefer [39] seamlessly converts the cross-modal task into a unified language modeling challenge, underpinned by predicted object labels." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 145, + 288, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 145, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 145, + 288, + 312 + ], + "type": "text", + "content": "However, two-stage methods encounter a noteworthy impediment: the detection stage neglects to leverage language contexts to prioritize the objects that are essential to the referring task. An alternative approach to the problem is one-stage methods [19, 31, 43], where extracted vision features are directly fused with the language features for object grounding. For instance, 3D-SPS [31] leverages text features to guide visual keypoints selection, enabling the progressive object grounding. BUTD-DETR [19] pioneers a DETR-like model, which fuses the vision-text features via a co-attention paradigm and then decodes objects in the utterance from the contextualized features. Building upon this, EDA [43] proposes a text-decoupled strategy and performs dense alignment of 3D objects and associated texts." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 313, + 288, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 313, + 288, + 494 + ], + "spans": [ + { + "bbox": [ + 46, + 313, + 288, + 494 + ], + "type": "text", + "content": "Despite one-stage methods unfold prowess performance, inherent limitations have been yet underexplored: 1) They overlook the intrinsic geometric attributes (distance, orientation, and shape) in 3D point clouds, leading to ambiguous visual representations. Conversely, our proposed PAGE module explicitly captures geometric features to avoid ambiguity. 2) The queries generation lack the fine-grained language guidance, limiting the model's capacity to infer potential language clues. Our tailored Flan-QS module explores language prior and context dependency to generate precise proposals. 3) They establish the vision-text alignment in Euclidean space, incapable of modeling the nonlinear relation of vision and text features. Our method embeds multimodal features on the Poincaré ball then leverage the PSA loss to foster the semantic-geometric alignment." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 502, + 263, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 502, + 263, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 502, + 263, + 514 + ], + "type": "text", + "content": "2.2. Geometric Extraction in 3D Point Clouds" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 522, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 714 + ], + "type": "text", + "content": "The explicit modeling of geometric structures has recently garnered considerable attention in the realms of 3D point cloud, e.g., point cloud understanding [8, 26, 32, 41], point cloud completion [46], point cloud registration [38], and 3D object detection [45]. DGCNN [41] proposes an EdgeConv operator, by which the point features are aggregated through graph convolution to capture intricate local geometric structures. RSCNN [26] explicitly encodes geometric relationship of points with relation-aware convolution. PointMixer [8] embeds geometric relationships of 3D point in a MLP-Mixer's streamline. PoinTr [46] devises a geometry-aware transformer to learn structural knowledge and local geometric relationships for point cloud completion. HGNet [45] delves into localized shape information to generate the proposals with a hierarchical graph model for 3D object detection. Our cutting-edge " + }, + { + "bbox": [ + 46, + 522, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{G}^3" + }, + { + "bbox": [ + 46, + 522, + 287, + 714 + ], + "type": "text", + "content": "-LQ incorporates" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "the geometric attributes of point cloud into the 3D VG task, thereby enhancing the discriminability of vision features." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 101, + 362, + 113 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 101, + 362, + 113 + ], + "spans": [ + { + "bbox": [ + 306, + 101, + 362, + 113 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 120, + 373, + 131 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 120, + 373, + 131 + ], + "spans": [ + { + "bbox": [ + 306, + 120, + 373, + 131 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 137, + 545, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 137, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 304, + 137, + 545, + 198 + ], + "type": "text", + "content": "Provided a point cloud " + }, + { + "bbox": [ + 304, + 137, + 545, + 198 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\{\\pmb{p}_i\\}_{i=1}^N \\in \\mathbb{R}^{N \\times (3 + F)}" + }, + { + "bbox": [ + 304, + 137, + 545, + 198 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 137, + 545, + 198 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 137, + 545, + 198 + ], + "type": "text", + "content": "-dim auxiliary features (e.g., RGB, normals, or multi-view features) of " + }, + { + "bbox": [ + 304, + 137, + 545, + 198 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 137, + 545, + 198 + ], + "type": "text", + "content": " points, and a free-form text " + }, + { + "bbox": [ + 304, + 137, + 545, + 198 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 304, + 137, + 545, + 198 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 137, + 545, + 198 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 137, + 545, + 198 + ], + "type": "text", + "content": " words, the primary goal of 3D VG is to establish a mapping " + }, + { + "bbox": [ + 304, + 137, + 545, + 198 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 137, + 545, + 198 + ], + "type": "text", + "content": " that links " + }, + { + "bbox": [ + 304, + 137, + 545, + 198 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 304, + 137, + 545, + 198 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 137, + 545, + 198 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 304, + 137, + 545, + 198 + ], + "type": "text", + "content": " to the target object " + }, + { + "bbox": [ + 304, + 137, + 545, + 198 + ], + "type": "inline_equation", + "content": "\\pmb{o}" + }, + { + "bbox": [ + 304, + 137, + 545, + 198 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 304, + 137, + 545, + 198 + ], + "type": "inline_equation", + "content": "\\mathcal{M}(\\mathcal{P}, \\mathcal{T}) \\to \\pmb{o}" + }, + { + "bbox": [ + 304, + 137, + 545, + 198 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 198, + 545, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 198, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 304, + 198, + 545, + 317 + ], + "type": "text", + "content": "Fig. 2 gives the outline of our proposed " + }, + { + "bbox": [ + 304, + 198, + 545, + 317 + ], + "type": "inline_equation", + "content": "\\mathrm{G}^3" + }, + { + "bbox": [ + 304, + 198, + 545, + 317 + ], + "type": "text", + "content": "-LQ method. Firstly, we leverage the well-accepted PointNet++ [36] tockenize the point cloud into vision token " + }, + { + "bbox": [ + 304, + 198, + 545, + 317 + ], + "type": "inline_equation", + "content": "\\mathbf{V}\\in \\mathbb{R}^{n\\times d}" + }, + { + "bbox": [ + 304, + 198, + 545, + 317 + ], + "type": "text", + "content": ". The text is encoded by the pre-trained RoBERTa [27] and yields the vanilla text token " + }, + { + "bbox": [ + 304, + 198, + 545, + 317 + ], + "type": "inline_equation", + "content": "\\mathbf{T}\\in \\mathbb{R}^{t\\times d}" + }, + { + "bbox": [ + 304, + 198, + 545, + 317 + ], + "type": "text", + "content": ". The vision token " + }, + { + "bbox": [ + 304, + 198, + 545, + 317 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 304, + 198, + 545, + 317 + ], + "type": "text", + "content": " is fed into the PAGE module to capture underlying geometric features " + }, + { + "bbox": [ + 304, + 198, + 545, + 317 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_g\\in \\mathbb{R}^{n\\times d}" + }, + { + "bbox": [ + 304, + 198, + 545, + 317 + ], + "type": "text", + "content": ", which adjusts according to the point positions, thereby furnishing precise geometric clues. Further, the geometric visual encoder embeds " + }, + { + "bbox": [ + 304, + 198, + 545, + 317 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_g" + }, + { + "bbox": [ + 304, + 198, + 545, + 317 + ], + "type": "text", + "content": " via self-attention and interact with text features " + }, + { + "bbox": [ + 304, + 198, + 545, + 317 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 304, + 198, + 545, + 317 + ], + "type": "text", + "content": " via cross-attention." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 318, + 545, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 318, + 545, + 402 + ], + "spans": [ + { + "bbox": [ + 304, + 318, + 545, + 402 + ], + "type": "text", + "content": "Finally, we build the language scene graph " + }, + { + "bbox": [ + 304, + 318, + 545, + 402 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 304, + 318, + 545, + 402 + ], + "type": "text", + "content": " and obtain the updated fine-grained text features, which will explicitly guide the queries selection. Like the object queries in most DETR-like models [22, 25, 43], the selected queries " + }, + { + "bbox": [ + 304, + 318, + 545, + 402 + ], + "type": "inline_equation", + "content": "\\mathbf{Q} \\in \\mathbb{R}^{K \\times d}" + }, + { + "bbox": [ + 304, + 318, + 545, + 402 + ], + "type": "text", + "content": " are fed into a cross-modal decoder to probe desired features and update themselves. The decoded queries " + }, + { + "bbox": [ + 304, + 318, + 545, + 402 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}' \\in \\mathbb{R}^{K \\times l}" + }, + { + "bbox": [ + 304, + 318, + 545, + 402 + ], + "type": "text", + "content": " are fed into an MLP to predict object boxes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 408, + 505, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 408, + 505, + 422 + ], + "spans": [ + { + "bbox": [ + 306, + 408, + 505, + 422 + ], + "type": "text", + "content": "3.2. Geometric Exploring and Embedding" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 427, + 545, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 427, + 545, + 462 + ], + "spans": [ + { + "bbox": [ + 304, + 427, + 545, + 462 + ], + "type": "text", + "content": "In this section, we propose a PAGE module to explicitly capture geometric properties of 3D scenes, which serves to encode holistic visual representations for 3D VG tasks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 463, + 545, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 463, + 545, + 534 + ], + "spans": [ + { + "bbox": [ + 304, + 463, + 545, + 534 + ], + "type": "text", + "content": "Point-Proxy Geometric Extraction. To facilitate local geometric features modeling, we firstly instantiate point proxies predicated on the proximity of the points, thereby serving as the local regions of the point cloud. As illustrated on Fig 2, we search " + }, + { + "bbox": [ + 304, + 463, + 545, + 534 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 463, + 545, + 534 + ], + "type": "text", + "content": " neighbors of " + }, + { + "bbox": [ + 304, + 463, + 545, + 534 + ], + "type": "inline_equation", + "content": "p_i \\in \\mathcal{P}" + }, + { + "bbox": [ + 304, + 463, + 545, + 534 + ], + "type": "text", + "content": " as a point proxy with the KNN algorithm, denoted as: " + }, + { + "bbox": [ + 304, + 463, + 545, + 534 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_i = k\\mathrm{NN}(\\mathcal{P},p_i)" + }, + { + "bbox": [ + 304, + 463, + 545, + 534 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 535, + 545, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 535, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 304, + 535, + 545, + 643 + ], + "type": "text", + "content": "The point proxy " + }, + { + "bbox": [ + 304, + 535, + 545, + 643 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_i" + }, + { + "bbox": [ + 304, + 535, + 545, + 643 + ], + "type": "text", + "content": " denotes index set with " + }, + { + "bbox": [ + 304, + 535, + 545, + 643 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 535, + 545, + 643 + ], + "type": "text", + "content": " closest points of " + }, + { + "bbox": [ + 304, + 535, + 545, + 643 + ], + "type": "inline_equation", + "content": "\\pmb{p}_i" + }, + { + "bbox": [ + 304, + 535, + 545, + 643 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 535, + 545, + 643 + ], + "type": "inline_equation", + "content": "\\pmb{v}_{m_i}:\\{\\pmb{v}_{m_{ij}}|j\\in \\mathcal{X}_i\\} \\in \\mathbb{R}^{k\\times d}" + }, + { + "bbox": [ + 304, + 535, + 545, + 643 + ], + "type": "text", + "content": " is the corresponding features. With the Point-Proxy Geometric Extraction component, we capture the local structure by learning the geometric topology features within the point proxy " + }, + { + "bbox": [ + 304, + 535, + 545, + 643 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_i" + }, + { + "bbox": [ + 304, + 535, + 545, + 643 + ], + "type": "text", + "content": ". In this fashion, it is effective to achieve an inductive local representations that integrate explicit reasoning pertaining to geometric structures and objects spatial layout. We primarily compute pre-defined geometric prior " + }, + { + "bbox": [ + 304, + 535, + 545, + 643 + ], + "type": "inline_equation", + "content": "\\widetilde{\\pmb{p}}_{ij}\\in \\mathbb{R}^7" + }, + { + "bbox": [ + 304, + 535, + 545, + 643 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 341, + 647, + 545, + 663 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 647, + 545, + 663 + ], + "spans": [ + { + "bbox": [ + 341, + 647, + 545, + 663 + ], + "type": "interline_equation", + "content": "\\widetilde {\\boldsymbol {p}} _ {i j} = \\operatorname {c o n c a t} \\left(\\boldsymbol {p} _ {i} - \\boldsymbol {p} _ {i j}, \\boldsymbol {p} _ {i}, \\left\\| \\boldsymbol {p} _ {i} - \\boldsymbol {p} _ {i j} \\right\\|\\right) \\tag {1}", + "image_path": "a2f1f51dda74988f265e577acae67e4c416ea7d6bbf5403cbfa9c731b4c64efd.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 666, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\widetilde{\\pmb{p}}_{ij}" + }, + { + "bbox": [ + 304, + 666, + 545, + 714 + ], + "type": "text", + "content": " embeds relative orientation and distance of different objects. Further, the low-level geometric features are encoded to high-level features " + }, + { + "bbox": [ + 304, + 666, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{h}_{ij} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 304, + 666, + 545, + 714 + ], + "type": "text", + "content": " with a linear layer " + }, + { + "bbox": [ + 304, + 666, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\phi[\\cdot]" + }, + { + "bbox": [ + 304, + 666, + 545, + 714 + ], + "type": "text", + "content": " to obtain abstract geometric conception and excavate" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13919" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 75, + 242, + 338 + ], + "blocks": [ + { + "bbox": [ + 65, + 75, + 242, + 338 + ], + "lines": [ + { + "bbox": [ + 65, + 75, + 242, + 338 + ], + "spans": [ + { + "bbox": [ + 65, + 75, + 242, + 338 + ], + "type": "image", + "image_path": "b67b4ac56f727ff4ffbe50de3d23ae65a9b69c8a024e406fcf3168d59955884b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 76, + 340, + 149, + 352 + ], + "lines": [ + { + "bbox": [ + 76, + 340, + 149, + 352 + ], + "spans": [ + { + "bbox": [ + 76, + 340, + 149, + 352 + ], + "type": "text", + "content": "Text: " + }, + { + "bbox": [ + 76, + 340, + 149, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = \\{t_i\\}_{i = 1}^L" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 156, + 340, + 236, + 352 + ], + "lines": [ + { + "bbox": [ + 156, + 340, + 236, + 352 + ], + "spans": [ + { + "bbox": [ + 156, + 340, + 236, + 352 + ], + "type": "text", + "content": "Point: " + }, + { + "bbox": [ + 156, + 340, + 236, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\{\\pmb {p}_i\\}_{i = 1}^N" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 243, + 76, + 531, + 358 + ], + "blocks": [ + { + "bbox": [ + 243, + 76, + 531, + 358 + ], + "lines": [ + { + "bbox": [ + 243, + 76, + 531, + 358 + ], + "spans": [ + { + "bbox": [ + 243, + 76, + 531, + 358 + ], + "type": "image", + "image_path": "1cdd38271661ca0d94a0c929951da1fe2b660e3702bc82dfe7edaa945e4be369.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 369, + 546, + 404 + ], + "lines": [ + { + "bbox": [ + 46, + 369, + 546, + 404 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 546, + 404 + ], + "type": "text", + "content": "Figure 2. Illustration of (a) our proposed " + }, + { + "bbox": [ + 46, + 369, + 546, + 404 + ], + "type": "inline_equation", + "content": "\\mathbf{G}^3" + }, + { + "bbox": [ + 46, + 369, + 546, + 404 + ], + "type": "text", + "content": "-LQ framework. The red dashed frames denote our novel modules. (b) PAGE moulder to model the underlying geometric structure of 3D point cloud and (c) the Geometric Visual Encoder further to embed the geometric topology features. (d) Flan-QS module to capture fine-grained language context with (e) Language Scene Graph for guiding query selection." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 412, + 282, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 412, + 282, + 424 + ], + "spans": [ + { + "bbox": [ + 47, + 412, + 282, + 424 + ], + "type": "text", + "content": "discriminative geometric structures, which is described as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 428, + 287, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 428, + 287, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 287, + 444 + ], + "type": "interline_equation", + "content": "h _ {i j} = \\widetilde {\\boldsymbol {p}} _ {i j} + \\phi [ \\widetilde {\\boldsymbol {p}} _ {i j} ], \\forall j \\in \\mathcal {X} _ {i} \\tag {2}", + "image_path": "c4bc4db98fc874ce0855ac79aee9cb0e380298ed465ea6a2e29bd6015e96ff8b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 453, + 287, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 453, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 453, + 287, + 548 + ], + "type": "text", + "content": "Proxy-Adaptive Geometric Refinement. The presence of diverse geometric structures across distinct point proxies necessitate anisotropic extractors. Further, we flesh out this intuition and engineer a Proxy-Adaptive Geometric Refinement component, poised to model refined geometric features effectively of diverse point proxies. Given the geometric prior " + }, + { + "bbox": [ + 46, + 453, + 287, + 548 + ], + "type": "inline_equation", + "content": "\\{h_{ij}|j\\in \\mathcal{X}_i\\} \\in \\mathbb{R}^{k\\times d}" + }, + { + "bbox": [ + 46, + 453, + 287, + 548 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 46, + 453, + 287, + 548 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_i" + }, + { + "bbox": [ + 46, + 453, + 287, + 548 + ], + "type": "text", + "content": " we establish local geometric refinement by the following streamline:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 104, + 552, + 287, + 577 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 552, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 104, + 552, + 287, + 577 + ], + "type": "interline_equation", + "content": "\\left\\{g _ {i j} \\right\\} = \\alpha \\odot \\frac {\\left\\{h _ {i j} \\right\\} - h _ {i}}{\\Omega + \\varepsilon} + \\beta \\tag {3}", + "image_path": "2053cf3bcca96244a10b67aa1f66d018bf453051c9610793fbfdadcd493651fa.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 579, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 579, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 579, + 287, + 640 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 579, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 46, + 579, + 287, + 640 + ], + "type": "text", + "content": " is a small constant for numerical stability and " + }, + { + "bbox": [ + 46, + 579, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\alpha, \\beta" + }, + { + "bbox": [ + 46, + 579, + 287, + 640 + ], + "type": "text", + "content": " are learnable parameters for features refinement. " + }, + { + "bbox": [ + 46, + 579, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 46, + 579, + 287, + 640 + ], + "type": "text", + "content": " is the standard deviation across all point proxies and channels. " + }, + { + "bbox": [ + 46, + 579, + 287, + 640 + ], + "type": "inline_equation", + "content": "h_i" + }, + { + "bbox": [ + 46, + 579, + 287, + 640 + ], + "type": "text", + "content": " is the center of " + }, + { + "bbox": [ + 46, + 579, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_i" + }, + { + "bbox": [ + 46, + 579, + 287, + 640 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 579, + 287, + 640 + ], + "type": "inline_equation", + "content": "g_i = \\{g_{ij} | j \\in \\mathcal{X}_i\\} \\in \\mathbb{R}^{k \\times d}" + }, + { + "bbox": [ + 46, + 579, + 287, + 640 + ], + "type": "text", + "content": " represents the anisotropic geometric structures in point proxy " + }, + { + "bbox": [ + 46, + 579, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_i" + }, + { + "bbox": [ + 46, + 579, + 287, + 640 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 646, + 287, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 646, + 287, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 646, + 287, + 670 + ], + "type": "text", + "content": "Finally, we opt for a softmax aggregation strategy to densely relate point features interaction in each " + }, + { + "bbox": [ + 47, + 646, + 287, + 670 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_i" + }, + { + "bbox": [ + 47, + 646, + 287, + 670 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 75, + 672, + 247, + 687 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 672, + 247, + 687 + ], + "spans": [ + { + "bbox": [ + 75, + 672, + 247, + 687 + ], + "type": "interline_equation", + "content": "\\widetilde {v} _ {g _ {i}} = \\operatorname {L i n e a r} \\left(\\operatorname {c o n c a t} \\left(v _ {m _ {i}}, g _ {i}\\right)\\right) \\in \\mathbb {R} ^ {k \\times d}", + "image_path": "18e18702ae2b64b75754af44bcc6cf046aed26caafa274a255b439dd1d6c8f91.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 61, + 689, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 61, + 689, + 287, + 715 + ], + "type": "interline_equation", + "content": "v _ {g _ {i}} = \\sum_ {j \\in \\mathcal {X} _ {i}} \\operatorname {s o f t m a x} \\left(\\widetilde {v} _ {g _ {i j}}\\right) \\cdot \\widetilde {v} _ {g _ {i j}} \\in \\mathbb {R} ^ {d}, \\forall j \\in \\mathcal {X} _ {i}, \\tag {4}", + "image_path": "c9aace666845a5a244b2b733e8d8e5748bc98f91da076e4acfa839ce3d3f03d9.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 412, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 412, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 306, + 412, + 545, + 437 + ], + "type": "text", + "content": "where Linear " + }, + { + "bbox": [ + 306, + 412, + 545, + 437 + ], + "type": "inline_equation", + "content": "(\\cdot)" + }, + { + "bbox": [ + 306, + 412, + 545, + 437 + ], + "type": "text", + "content": " denotes the linear projection, " + }, + { + "bbox": [ + 306, + 412, + 545, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_g = \\{v_{g_i}\\}_{i=1}^n" + }, + { + "bbox": [ + 306, + 412, + 545, + 437 + ], + "type": "text", + "content": " is the output token with geometric representations." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 436, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 436, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 305, + 436, + 545, + 555 + ], + "type": "text", + "content": "Geometric Visual Encoder. Geometric perception enables to strengthen shape features of 3D objects and demonstrate enhanced viewpoint sensitivity, which bestows flexibility in capturing spatial relation and improves precise 3D objects grounding. Hence, we embed geometric topology features " + }, + { + "bbox": [ + 305, + 436, + 545, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_g" + }, + { + "bbox": [ + 305, + 436, + 545, + 555 + ], + "type": "text", + "content": " via a tailored Geometric Encoder. As shown in Fig. 2, the core of Geometric Encoder is Geometric self-attention layer. Given the high-level vision features " + }, + { + "bbox": [ + 305, + 436, + 545, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_m\\in \\mathbb{R}^{n\\times d}" + }, + { + "bbox": [ + 305, + 436, + 545, + 555 + ], + "type": "text", + "content": " and the explicit geometric features " + }, + { + "bbox": [ + 305, + 436, + 545, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_g" + }, + { + "bbox": [ + 305, + 436, + 545, + 555 + ], + "type": "text", + "content": ", the output " + }, + { + "bbox": [ + 305, + 436, + 545, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{emb}\\in \\mathbb{R}^{n\\times d}" + }, + { + "bbox": [ + 305, + 436, + 545, + 555 + ], + "type": "text", + "content": " is a weighted sum of projected features:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 358, + 555, + 545, + 587 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 555, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 358, + 555, + 545, + 587 + ], + "type": "interline_equation", + "content": "v _ {i} ^ {e m b} = \\sum_ {j = 1} ^ {n} a _ {i j} \\left(v _ {m _ {j}} \\mathbf {W} ^ {V}\\right) \\in \\mathbb {R} ^ {d} \\tag {5}", + "image_path": "fbc7ab6590f4d66719442c669076be039f8a0b28e519443a2f9d8310adccb88c.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 588, + 545, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 588, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 305, + 588, + 545, + 611 + ], + "type": "text", + "content": "The attention weight " + }, + { + "bbox": [ + 305, + 588, + 545, + 611 + ], + "type": "inline_equation", + "content": "a_{ij}" + }, + { + "bbox": [ + 305, + 588, + 545, + 611 + ], + "type": "text", + "content": " is computed by a row-wise softmax, which can be formulated as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 311, + 614, + 545, + 646 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 614, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 311, + 614, + 545, + 646 + ], + "type": "interline_equation", + "content": "a _ {i j} = \\sigma \\left[ \\frac {\\left(v _ {g _ {i}} \\mathbf {W} _ {g} ^ {Q} + v _ {m _ {i}} \\mathbf {W} _ {m} ^ {Q}\\right) \\left(v _ {g _ {j}} \\mathbf {W} _ {g} ^ {K} + v _ {m _ {j}} \\mathbf {W} _ {m} ^ {K}\\right) ^ {T}}{\\sqrt {d}} \\right] \\tag {6}", + "image_path": "eea849bbdc850f29097df400d724c94db5869eaaa2b5c793ddac377309b8afd3.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 648, + 545, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 648, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 305, + 648, + 545, + 688 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_g^Q,\\mathbf{W}_g^K" + }, + { + "bbox": [ + 305, + 648, + 545, + 688 + ], + "type": "text", + "content": " are geometric projection matrices of query and key, " + }, + { + "bbox": [ + 305, + 648, + 545, + 688 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_m^Q,\\mathbf{W}_m^K" + }, + { + "bbox": [ + 305, + 648, + 545, + 688 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 648, + 545, + 688 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_m^V" + }, + { + "bbox": [ + 305, + 648, + 545, + 688 + ], + "type": "text", + "content": " are semantic projection matrices of query, key and value. " + }, + { + "bbox": [ + 305, + 648, + 545, + 688 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 305, + 648, + 545, + 688 + ], + "type": "text", + "content": " is the row-wise softmax operator." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Further, the geometric-enhanced features " + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{emb}" + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": " pass through the cross-attention layer to interact with the text" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 784, + 317, + 792 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 784, + 317, + 792 + ], + "spans": [ + { + "bbox": [ + 294, + 784, + 317, + 792 + ], + "type": "text", + "content": "13920" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": "features and box features (used in two stage), for obtaining the cross-modal features " + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "inline_equation", + "content": "\\mathbf{V}' \\in \\mathbb{R}^{n \\times d}" + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "inline_equation", + "content": "\\mathbf{T}' \\in \\mathbb{R}^{t \\times d}" + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": ". Finally, " + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "inline_equation", + "content": "\\mathbf{V}'" + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": " is linearly projected as the candidate query " + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_c" + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 116, + 144, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 116, + 144, + 128 + ], + "spans": [ + { + "bbox": [ + 47, + 116, + 144, + 128 + ], + "type": "text", + "content": "3.3. Query Selection" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 135, + 287, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 135, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 135, + 287, + 449 + ], + "type": "text", + "content": "In the DETR-like model [24, 42, 50], object queries play essential roles to determine potential regions of targets, which are equally prominent for 3G VG systems, decoded into a box center and size that conform to vision tokens. Assisted by global text features, GroundingDINO [25] employs a language-guided query selection protocol to generate query proposals. However, such text features are relatively coarse-grained, leading to inaccurate query generation. Actually, free-form texts harbor intricate syntactic structure and possess a context-dependency nature. It further serves as inspiration for us to generate precise vision queries via improving fine-grained semantic understanding. Language Scene Graph. We intend to delve into fine-grained text information, which is defined as decoupled semantic components (e.g., object, attributes, and relations) and syntactic structures. To this end, we perform a syntactic dependency parsing to construct language graph " + }, + { + "bbox": [ + 46, + 135, + 287, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{G} = \\{\\mathcal{O},\\mathcal{R}\\}" + }, + { + "bbox": [ + 46, + 135, + 287, + 449 + ], + "type": "text", + "content": " by the off-the-shelf Scene Graph Parser [1], which serves as a reasoning inductive bias of 3D VG tasks. " + }, + { + "bbox": [ + 46, + 135, + 287, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{O} = \\{\\pmb{o}_i\\}" + }, + { + "bbox": [ + 46, + 135, + 287, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 135, + 287, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{R} = \\{\\pmb{r}_{ij}\\}" + }, + { + "bbox": [ + 46, + 135, + 287, + 449 + ], + "type": "text", + "content": " are nodes and edges set. Each object " + }, + { + "bbox": [ + 46, + 135, + 287, + 449 + ], + "type": "inline_equation", + "content": "o_i" + }, + { + "bbox": [ + 46, + 135, + 287, + 449 + ], + "type": "text", + "content": " is denoted as a phrase with a set of attributes (appearance, shape, texture). Harnessing the final layer of the pre-trained RoBERTa model, we then acquire the object phrase embedding " + }, + { + "bbox": [ + 46, + 135, + 287, + 449 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{o_i}\\in \\mathbb{R}^d" + }, + { + "bbox": [ + 46, + 135, + 287, + 449 + ], + "type": "text", + "content": " as the graph node " + }, + { + "bbox": [ + 46, + 135, + 287, + 449 + ], + "type": "inline_equation", + "content": "o_i\\in \\mathcal{O}" + }, + { + "bbox": [ + 46, + 135, + 287, + 449 + ], + "type": "text", + "content": ". In a similar vein, the relation phrases are also encoded into the context-aware embeddings " + }, + { + "bbox": [ + 46, + 135, + 287, + 449 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{r_{ij}}\\in \\mathbb{R}^d" + }, + { + "bbox": [ + 46, + 135, + 287, + 449 + ], + "type": "text", + "content": " akin to the graph edges " + }, + { + "bbox": [ + 46, + 135, + 287, + 449 + ], + "type": "inline_equation", + "content": "\\pmb{r}_{ij}\\in \\mathcal{R}" + }, + { + "bbox": [ + 46, + 135, + 287, + 449 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 449, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 521 + ], + "type": "text", + "content": "Inspired by the message passing mechanism [14, 23], we proceed to enhance relation and object embeddings via language graph convolution. Firstly, to enrich the contextual representations of relation embedding in connection with its affiliated nodes, we aggregate messages and update its features with the subject " + }, + { + "bbox": [ + 46, + 449, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{o_i}" + }, + { + "bbox": [ + 46, + 449, + 287, + 521 + ], + "type": "text", + "content": " and object node " + }, + { + "bbox": [ + 46, + 449, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{o_j}" + }, + { + "bbox": [ + 46, + 449, + 287, + 521 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 93, + 524, + 287, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 524, + 287, + 539 + ], + "spans": [ + { + "bbox": [ + 93, + 524, + 287, + 539 + ], + "type": "interline_equation", + "content": "\\widetilde {\\boldsymbol {x}} _ {r _ {i j}} = \\boldsymbol {x} _ {r _ {i j}} + F _ {r} \\left(\\left[ \\boldsymbol {x} _ {o _ {i}}; \\boldsymbol {x} _ {r _ {i j}}; \\boldsymbol {x} _ {o _ {j}} \\right]\\right) \\tag {7}", + "image_path": "140d45ab0a8f6302f79625fb46790cdfe996383a1920437f7e2348c545d192a2.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 542, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 542, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 47, + 542, + 287, + 590 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 542, + 287, + 590 + ], + "type": "inline_equation", + "content": "F_{r}" + }, + { + "bbox": [ + 47, + 542, + 287, + 590 + ], + "type": "text", + "content": " is a linear projection layer and " + }, + { + "bbox": [ + 47, + 542, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\widetilde{\\pmb{x}}_{r_{ij}}" + }, + { + "bbox": [ + 47, + 542, + 287, + 590 + ], + "type": "text", + "content": " is the context-aware relation embedding. We then update the phrase embedding " + }, + { + "bbox": [ + 47, + 542, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{o_i}" + }, + { + "bbox": [ + 47, + 542, + 287, + 590 + ], + "type": "text", + "content": " by passing messages from all connected nodes " + }, + { + "bbox": [ + 47, + 542, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\Gamma (i)" + }, + { + "bbox": [ + 47, + 542, + 287, + 590 + ], + "type": "text", + "content": " and the edges with a tailored attention mechanism:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 82, + 594, + 287, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 594, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 82, + 594, + 287, + 620 + ], + "type": "interline_equation", + "content": "\\widetilde {\\boldsymbol {x}} _ {o _ {i}} = \\boldsymbol {x} _ {o _ {j}} + \\sum_ {j \\in \\Gamma (i)} \\boldsymbol {w} _ {o _ {i j}} F _ {o} \\left(\\left[ \\boldsymbol {x} _ {o _ {j}}; \\boldsymbol {x} _ {r _ {i j}} \\right]\\right) \\tag {8}", + "image_path": "194f2313b9bfc3a6a7f7e1c4ac4e0761b17a959dd3875c9024f709038fa43655.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 628, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 628, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 47, + 628, + 287, + 653 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 628, + 287, + 653 + ], + "type": "inline_equation", + "content": "F_{o}" + }, + { + "bbox": [ + 47, + 628, + 287, + 653 + ], + "type": "text", + "content": " is another linear layer and " + }, + { + "bbox": [ + 47, + 628, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\widetilde{\\pmb{x}}_{o_i}" + }, + { + "bbox": [ + 47, + 628, + 287, + 653 + ], + "type": "text", + "content": " denotes context-aware object features. " + }, + { + "bbox": [ + 47, + 628, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\pmb{w}_{o_{ij}}" + }, + { + "bbox": [ + 47, + 628, + 287, + 653 + ], + "type": "text", + "content": " is the tailored attention weight:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 665, + 287, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 665, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 53, + 665, + 287, + 685 + ], + "type": "interline_equation", + "content": "\\boldsymbol {w} _ {o _ {i j}} = \\operatorname {s o f t m a x} \\left[ F _ {o} \\left(\\left[ \\boldsymbol {x} _ {o _ {i}}; \\widetilde {\\boldsymbol {x}} _ {r _ {i j}} \\right]\\right) ^ {T} F _ {o} \\left(\\left[ \\boldsymbol {x} _ {o _ {j}}; \\widetilde {\\boldsymbol {x}} _ {r _ {i j}} \\right]\\right) \\right] \\tag {9}", + "image_path": "f05d56a8b9e01a8c831cc10b5c12192332a051083bed6ca0c9b870faea76e75a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "The attention protocol manifests the capability to discern crucial contextual features from language graph, serving as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 73, + 477, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 73, + 477, + 84 + ], + "spans": [ + { + "bbox": [ + 305, + 73, + 477, + 84 + ], + "type": "text", + "content": "potent language priors for query selection." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 84, + 545, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 84, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 304, + 84, + 545, + 216 + ], + "type": "text", + "content": "Fine-grained Query Selection. We input the candidate query features " + }, + { + "bbox": [ + 304, + 84, + 545, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_c\\in \\mathbb{R}^{n\\times d}" + }, + { + "bbox": [ + 304, + 84, + 545, + 216 + ], + "type": "text", + "content": " from the cross-modal encoding layer into the Query Embedding component, which consists of several stacked Fully-connected (FC) layers and BN layers. Further, we leverage the fine-grained text features " + }, + { + "bbox": [ + 304, + 84, + 545, + 216 + ], + "type": "inline_equation", + "content": "T_{f}" + }, + { + "bbox": [ + 304, + 84, + 545, + 216 + ], + "type": "text", + "content": " to guide the query selection through a cross-attention fusion layer. The resulting query tokens are fed into an MLP layer for confidence score prediction. Through selecting the top-" + }, + { + "bbox": [ + 304, + 84, + 545, + 216 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 84, + 545, + 216 + ], + "type": "text", + "content": " highest scoring tokens for the following decoder, we derive the updated object queries " + }, + { + "bbox": [ + 304, + 84, + 545, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_s\\in \\mathbb{R}^{K\\times d}" + }, + { + "bbox": [ + 304, + 84, + 545, + 216 + ], + "type": "text", + "content": ", which are utilized to predict a box center, height and width." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 220, + 487, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 220, + 487, + 232 + ], + "spans": [ + { + "bbox": [ + 306, + 220, + 487, + 232 + ], + "type": "text", + "content": "3.4. Poincaré Semantic Alignment loss" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 239, + 545, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 239, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 304, + 239, + 545, + 346 + ], + "type": "text", + "content": "We linearly embed the queries " + }, + { + "bbox": [ + 304, + 239, + 545, + 346 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_s" + }, + { + "bbox": [ + 304, + 239, + 545, + 346 + ], + "type": "text", + "content": " to the predicted object features " + }, + { + "bbox": [ + 304, + 239, + 545, + 346 + ], + "type": "inline_equation", + "content": "\\pmb{q} \\in \\mathbb{R}^{K \\times 64}" + }, + { + "bbox": [ + 304, + 239, + 545, + 346 + ], + "type": "text", + "content": ", which aligns with the embedded text features " + }, + { + "bbox": [ + 304, + 239, + 545, + 346 + ], + "type": "inline_equation", + "content": "\\pmb{t} \\in \\mathbb{R}^{t \\times 64}" + }, + { + "bbox": [ + 304, + 239, + 545, + 346 + ], + "type": "text", + "content": ". Noteworthy, point clouds (spanning from scenes, objects to parts) and texts (ranging from sentences, phrases to words) encompass inherent hierarchy. Further, granularity disparities of geometric and semantic representations complicate the intricate correlations. However, it is the Achilles' heel to model hierarchical structures and complex relationships in canonical Euclidean space." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 347, + 545, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 347, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 304, + 347, + 545, + 442 + ], + "type": "text", + "content": "To tackle this issue, we provide the insight to map vision-text features to a special hyperbolic space—Poincaré ball, for its powerful natural aptitude in modeling tree-shaped features. Based on this, we carefully devise a PSA loss to pave the way for semantic-geometric consistency. Benefiting from exponential volume expansion of hyperbolic embedding, PSA explicitly models complex semantic-geometric relationship for capturing hierarchical structures." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 442, + 545, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 442, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 305, + 442, + 545, + 491 + ], + "type": "text", + "content": "Since the Poincaré ball is a Riemannian manifold, we project the vision " + }, + { + "bbox": [ + 305, + 442, + 545, + 491 + ], + "type": "inline_equation", + "content": "\\pmb{q}" + }, + { + "bbox": [ + 305, + 442, + 545, + 491 + ], + "type": "text", + "content": " and text features " + }, + { + "bbox": [ + 305, + 442, + 545, + 491 + ], + "type": "inline_equation", + "content": "\\pmb{t}" + }, + { + "bbox": [ + 305, + 442, + 545, + 491 + ], + "type": "text", + "content": " to the Poincaré ball " + }, + { + "bbox": [ + 305, + 442, + 545, + 491 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_c^d" + }, + { + "bbox": [ + 305, + 442, + 545, + 491 + ], + "type": "text", + "content": " with the curvature " + }, + { + "bbox": [ + 305, + 442, + 545, + 491 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 305, + 442, + 545, + 491 + ], + "type": "text", + "content": " via an exponential mapping function " + }, + { + "bbox": [ + 305, + 442, + 545, + 491 + ], + "type": "inline_equation", + "content": "\\exp_{\\pmb{x}}^{c}: \\mathbb{R}^{d} \\to \\mathbb{D}_{c}^{d}" + }, + { + "bbox": [ + 305, + 442, + 545, + 491 + ], + "type": "text", + "content": ", which can be defined as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 503, + 545, + 528 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 503, + 545, + 528 + ], + "spans": [ + { + "bbox": [ + 314, + 503, + 545, + 528 + ], + "type": "interline_equation", + "content": "\\boldsymbol {q} ^ {\\mathrm {c}} = \\exp_ {\\boldsymbol {x}} ^ {\\mathrm {c}} (\\boldsymbol {q}) = \\boldsymbol {x} \\oplus_ {c} \\left(\\tanh \\left(\\sqrt {c} \\frac {\\lambda_ {\\boldsymbol {x}} ^ {\\mathrm {c}} \\| \\boldsymbol {q} \\|}{2}\\right) \\frac {\\boldsymbol {q}}{\\sqrt {c} \\| \\boldsymbol {q} \\|}\\right) \\tag {10}", + "image_path": "9f98f7fc55e18de9f7ddce969bcff8e08d43acf2a97b97c1142e1caad8b5c843.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 531, + 545, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 531, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 545, + 592 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 531, + 545, + 592 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 304, + 531, + 545, + 592 + ], + "type": "text", + "content": " is a fixed base point to make formulas less cumbersome and " + }, + { + "bbox": [ + 304, + 531, + 545, + 592 + ], + "type": "inline_equation", + "content": "\\lambda_{\\pmb{x}}^{c} = 2 / (1 - c\\| \\pmb{x}\\|^{2})" + }, + { + "bbox": [ + 304, + 531, + 545, + 592 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 531, + 545, + 592 + ], + "type": "inline_equation", + "content": "t^c" + }, + { + "bbox": [ + 304, + 531, + 545, + 592 + ], + "type": "text", + "content": " is defined in a similar approach. " + }, + { + "bbox": [ + 304, + 531, + 545, + 592 + ], + "type": "inline_equation", + "content": "\\oplus_{c}" + }, + { + "bbox": [ + 304, + 531, + 545, + 592 + ], + "type": "text", + "content": " denotes the addition operator in the Mobius gyrovector space. The geodesic distance between the a vision-text feature pair " + }, + { + "bbox": [ + 304, + 531, + 545, + 592 + ], + "type": "inline_equation", + "content": "(\\pmb{q}_i^c,\\pmb{t}_j^c)" + }, + { + "bbox": [ + 304, + 531, + 545, + 592 + ], + "type": "text", + "content": " on the Poincaré ball " + }, + { + "bbox": [ + 304, + 531, + 545, + 592 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_c^d" + }, + { + "bbox": [ + 304, + 531, + 545, + 592 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 327, + 594, + 545, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 594, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 327, + 594, + 545, + 617 + ], + "type": "interline_equation", + "content": "d _ {c} \\left(\\boldsymbol {q} _ {i} ^ {c}, \\boldsymbol {t} _ {j} ^ {c}\\right) = \\frac {2}{\\sqrt {c}} \\operatorname {a r c t a n h} \\left(\\sqrt {c} \\| - \\boldsymbol {q} _ {i} ^ {c} \\oplus_ {c} \\boldsymbol {t} _ {j} ^ {c} \\|\\right) \\tag {11}", + "image_path": "afef64250903e215de22106275e11afb739b396dbd1183236d5340c0bce662db.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 619, + 545, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 619, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 304, + 619, + 545, + 654 + ], + "type": "text", + "content": "Inspired by [43], we term the PSA loss to densely align the fine-grained vision-text features via contrastive learning on Poincaré ball. The query loss is defined as:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 660, + 545, + 697 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 660, + 545, + 697 + ], + "spans": [ + { + "bbox": [ + 312, + 660, + 545, + 697 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {p s a}} ^ {q} = \\sum_ {i = 1} ^ {K} \\frac {- 1}{\\left| \\mathcal {T} _ {i} ^ {+} \\right|} \\sum_ {\\boldsymbol {t} _ {i} ^ {c} \\in \\mathcal {T} _ {i} ^ {+}} \\log \\left[ \\right. \\frac {\\exp \\left[ \\boldsymbol {w} _ {+} \\left(- d _ {c} \\left(\\boldsymbol {q} _ {i} ^ {c} , \\boldsymbol {t} _ {i} ^ {c}\\right) / \\tau\\right)\\right]}{\\sum_ {j = 1} ^ {l} \\exp \\left[ \\boldsymbol {w} _ {-} \\left(- d _ {c} \\left(\\boldsymbol {q} _ {i} ^ {c} , \\boldsymbol {t} _ {j} ^ {c}\\right) / \\tau\\right)} \\right] \\tag {12}", + "image_path": "3fec73273855e94b0b9282da4b2eddf156964c2c88febcac19ccf25a6a81fd25.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "content": " are the number of object tokens and text to-" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13921" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 69, + 547, + 294 + ], + "blocks": [ + { + "bbox": [ + 48, + 69, + 547, + 294 + ], + "lines": [ + { + "bbox": [ + 48, + 69, + 547, + 294 + ], + "spans": [ + { + "bbox": [ + 48, + 69, + 547, + 294 + ], + "type": "table", + "html": "
MethodVenueInputUnique (~19%)Multiple (~81%)Overall
0.250.50.250.50.250.5
ScanRefer [6]ECCV'203D+2D76.3353.5132.7321.1141.1927.40
TGNN [16]AAAI'213D68.6156.8029.8423.1837.3729.70
InstanceRefer [47]ICCV'213D77.4566.8331.2724.7740.2332.93
SAT [44]ICCV'213D+2D73.2150.8337.6425.1644.5430.14
FFL-3DOG [12]ICCV'213D78.8067.9435.1925.741.3334.01
3DVG [48]ICCV'213D+2D81.9360.6439.328.4247.5734.67
3D-SPS [31]CVPR'223D+2D84.1266.7240.3229.8248.8236.98
BUTD-DETR [19]ECCV'223D82.8864.9844.7333.9750.4238.60
ViL3DRel [7]NeurlPS'223D81.5868.6240.3030.7147.9437.73
EDA [43]CVPR'233D85.7668.5749.1337.6454.5942.26
3DJCG [5]CVPR'223D+2D83.3764.3441.3930.8249.5637.33
3D-VLP [21]CVPR'233D+2D84.2364.6143.5133.4151.4139.46
3D-VisTA(Scratch) [51]ICCV'233D77.4070.9038.7034.8045.9041.50
G3-LQ(Two-Stage)——3D88.0972.7351.4840.8056.9045.58
3D-SPS(One-stage) [31]CVPR'223D81.6364.7739.4829.6147.6536.43
BUTD-DETR(One-stage) [19]ECCV'223D81.4761.2444.2032.8149.7637.05
EDA(One-stage) [43]CVPR'233D86.4069.4248.1136.8253.8341.70
G3-LQ(One-stage)——3D88.5973.2850.2339.7255.9544.72
", + "image_path": "04a528d765c1989069c791d625ba58086399593d204637a02fade71bd9dee0dd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 347, + 287, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 347, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 46, + 347, + 287, + 419 + ], + "type": "text", + "content": "kens, " + }, + { + "bbox": [ + 46, + 347, + 287, + 419 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 347, + 287, + 419 + ], + "type": "text", + "content": " is the temperature coefficient. " + }, + { + "bbox": [ + 46, + 347, + 287, + 419 + ], + "type": "inline_equation", + "content": "\\pmb{w}^{+}" + }, + { + "bbox": [ + 46, + 347, + 287, + 419 + ], + "type": "inline_equation", + "content": "w^{-}" + }, + { + "bbox": [ + 46, + 347, + 287, + 419 + ], + "type": "text", + "content": " are weights of positive and negative terms. Similar to EDA, " + }, + { + "bbox": [ + 46, + 347, + 287, + 419 + ], + "type": "inline_equation", + "content": "\\pmb{t}_i" + }, + { + "bbox": [ + 46, + 347, + 287, + 419 + ], + "type": "text", + "content": " is the positive text features of the " + }, + { + "bbox": [ + 46, + 347, + 287, + 419 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 347, + 287, + 419 + ], + "type": "text", + "content": "-th object query and " + }, + { + "bbox": [ + 46, + 347, + 287, + 419 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i^+" + }, + { + "bbox": [ + 46, + 347, + 287, + 419 + ], + "type": "text", + "content": " are the positive features set of " + }, + { + "bbox": [ + 46, + 347, + 287, + 419 + ], + "type": "inline_equation", + "content": "\\pmb{t}_i" + }, + { + "bbox": [ + 46, + 347, + 287, + 419 + ], + "type": "text", + "content": ", which primarily includes the object entries, attributes, relations, and pronouns. Moreover, the text loss can be formulated similarly:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 52, + 426, + 287, + 462 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 426, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 52, + 426, + 287, + 462 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {p s a}} ^ {t} = \\sum_ {i = 1} ^ {l} \\frac {- 1}{\\left| \\mathcal {Q} _ {i} ^ {+} \\right|} \\sum_ {\\boldsymbol {q} _ {i} ^ {c} \\in \\mathcal {Q} _ {i} ^ {+}} \\log \\left[ \\right. \\frac {\\exp \\left[ \\boldsymbol {w} _ {+} \\left(- d _ {c} \\left(\\boldsymbol {t} _ {i} ^ {c} , \\boldsymbol {q} _ {i} ^ {c}\\right) / \\tau\\right)\\right]}{\\sum_ {j = 1} ^ {K} \\exp \\left[ \\boldsymbol {w} _ {-} \\left(- d _ {c} \\left(\\boldsymbol {t} _ {i} ^ {c} , \\boldsymbol {q} _ {j} ^ {c}\\right) / \\tau\\right)} \\right] \\tag {13}", + "image_path": "440102b4f69a7bca4b9ab531f07c290ec54b525128a3cf8e48def8f76d37a03b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 468, + 287, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 468, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 468, + 287, + 491 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 468, + 287, + 491 + ], + "type": "inline_equation", + "content": "\\pmb{q}_i^c\\in \\mathcal{Q}_i^+" + }, + { + "bbox": [ + 47, + 468, + 287, + 491 + ], + "type": "text", + "content": " is the positive object features of " + }, + { + "bbox": [ + 47, + 468, + 287, + 491 + ], + "type": "inline_equation", + "content": "\\pmb{t}_i^c" + }, + { + "bbox": [ + 47, + 468, + 287, + 491 + ], + "type": "text", + "content": " . The final loss " + }, + { + "bbox": [ + 47, + 468, + 287, + 491 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{psa}}" + }, + { + "bbox": [ + 47, + 468, + 287, + 491 + ], + "type": "text", + "content": " is the average of the above two terms." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 492, + 287, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 492, + 287, + 528 + ], + "spans": [ + { + "bbox": [ + 47, + 492, + 287, + 528 + ], + "type": "text", + "content": "Following [43], the total loss of our method also includes the box regression loss " + }, + { + "bbox": [ + 47, + 492, + 287, + 528 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{loc}}" + }, + { + "bbox": [ + 47, + 492, + 287, + 528 + ], + "type": "text", + "content": ", the position alignment loss " + }, + { + "bbox": [ + 47, + 492, + 287, + 528 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{pa}}" + }, + { + "bbox": [ + 47, + 492, + 287, + 528 + ], + "type": "text", + "content": " which details in the Supplementary Material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 534, + 123, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 534, + 123, + 548 + ], + "spans": [ + { + "bbox": [ + 47, + 534, + 123, + 548 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 554, + 287, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 554, + 287, + 603 + ], + "spans": [ + { + "bbox": [ + 46, + 554, + 287, + 603 + ], + "type": "text", + "content": "Firstly, we carry out comparisons with SOTA methods on ScanRefer and " + }, + { + "bbox": [ + 46, + 554, + 287, + 603 + ], + "type": "inline_equation", + "content": "\\mathrm{Nr3D / Sr3D}" + }, + { + "bbox": [ + 46, + 554, + 287, + 603 + ], + "type": "text", + "content": " dataset in Sec. 4.1. Further, we delve into ablation studies in Sec. 4.2. Finally, we visualize and analyze the 3D VG results in Sec. 4.3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 610, + 194, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 610, + 194, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 610, + 194, + 624 + ], + "type": "text", + "content": "4.1. Quantitative Comparisons" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": "Performance on the ScanRefer. As illustrated in Table. 1, our method consistently exhibits admirably performance, trumping all competitors by a substantial margin on all test subsets. Our method achieves " + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "inline_equation", + "content": "56.90\\%" + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "inline_equation", + "content": "45.58\\%" + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": " performance in terms of the overall accuracy, with a remarkable improvement compared with EDA [43], " + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "inline_equation", + "content": "2.31\\%" + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "inline_equation", + "content": "2.34\\%" + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": ". 1) Compared with 2D assistance methods [17, 44]" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 307, + 346, + 545, + 533 + ], + "blocks": [ + { + "bbox": [ + 46, + 300, + 546, + 335 + ], + "lines": [ + { + "bbox": [ + 46, + 300, + 546, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 546, + 335 + ], + "type": "text", + "content": "Table 1. Comparison results on the ScanRefer dataset, in terms of the accuracy evaluated by IoU 0.25 and IoU 0.5. The unique denotes samples devoid of distracting objects, while multiple applies to remaining samples. In the one-stage setting, there is no reliance on the additional 3D object detection step. Our proposed " + }, + { + "bbox": [ + 46, + 300, + 546, + 335 + ], + "type": "inline_equation", + "content": "\\mathrm{G}^3 -\\mathrm{LQ}" + }, + { + "bbox": [ + 46, + 300, + 546, + 335 + ], + "type": "text", + "content": " performs favorably over current state-of-the-art approach EDA [43]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 346, + 545, + 533 + ], + "lines": [ + { + "bbox": [ + 307, + 346, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 346, + 545, + 533 + ], + "type": "table", + "html": "
MethodNr3DSr3D
OverallHardOverallHard
TGNN [16]37.330.645.036.9
InstanceRefer [47]38.831.848.040.5
3DVG [48]40.834.851.444.9
LanguageRefer [39]43.936.656.049.3
TransRefer3D [15]48.039.657.450.2
SAT [44]49.242.457.950.0
LAR [4]48.942.359.451.2
3DRef [2]47.038.339.032.0
3D-SPS [31]51.545.162.665.4
MVT [17]55.149.164.558.8
BUTD-DETR [19]54.648.467.063.2
EDA [43]52.146.168.162.9
3D-VisTA [51]57.549.469.663.6
G3-LQ58.450.773.166.3
", + "image_path": "1848c3b839dd84d9fdd4714d75b6ba643df9c12e457e327ea873821c17bd41da.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 542, + 545, + 564 + ], + "lines": [ + { + "bbox": [ + 305, + 542, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 305, + 542, + 545, + 564 + ], + "type": "text", + "content": "Table 2. Quantitative comparisons on " + }, + { + "bbox": [ + 305, + 542, + 545, + 564 + ], + "type": "inline_equation", + "content": "Nr3D" + }, + { + "bbox": [ + 305, + 542, + 545, + 564 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 542, + 545, + 564 + ], + "type": "inline_equation", + "content": "Sr3D" + }, + { + "bbox": [ + 305, + 542, + 545, + 564 + ], + "type": "text", + "content": " dataset. We have highlighted the top-performing three methods in purple." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": "that furnish 2D image priors imbued with elaborate semantics and textures, the proposed " + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{G}^3" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": "-LQ method showcases its superiority in performance. This serves as further evidence that our method explicitly captures geometric features, thus facilitating the shape understanding of target object and the vision-text alignment. 2) Our approach, featuring fine-grained language-guided query selection, surpasses all recently published DETR-like models [19, 43]. To explain, the proposed Flan-QS module thoroughly exploits fine-grained language prior and captures global context dependency, which effectively mitigates the issues of grounding ambiguity. 3) Compared with the 3D pre-trained" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "13922" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 71, + 288, + 193 + ], + "blocks": [ + { + "bbox": [ + 48, + 71, + 288, + 193 + ], + "lines": [ + { + "bbox": [ + 48, + 71, + 288, + 193 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 288, + 193 + ], + "type": "table", + "html": "
IDPAGEFlan-QSPSAUniqueMultiple
0.250.50.250.5
(a)———85.6768.5749.1337.64
(b)✓——86.8970.2650.4638.61
(c)—✓—86.7569.7650.1138.58
(d)——✓86.3269.2049.5838.18
(e)✓—✓87.4571.9551.0840.49
(f)—✓✓87.3170.7550.9040.19
(g)✓✓—87.6672.1650.8740.05
(h)✓✓✓88.0972.7351.4840.80
", + "image_path": "5a3f526b7fdd65c1142166422caac2bfc9c7f302a77d76abe21feb503bdc3d97.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 231, + 287, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 231, + 287, + 339 + ], + "spans": [ + { + "bbox": [ + 46, + 231, + 287, + 339 + ], + "type": "text", + "content": "models [21, 51], " + }, + { + "bbox": [ + 46, + 231, + 287, + 339 + ], + "type": "inline_equation", + "content": "\\mathrm{G}^3" + }, + { + "bbox": [ + 46, + 231, + 287, + 339 + ], + "type": "text", + "content": "-LQ achieves SOTA accuracy, because of the semantic-geometric alignment in Poincaré space with the tailored PSA loss, which aids in modeling the complex relationships of vision-text features. 4) The \"multiple\" setting entails a 3D scene where the description distinguishes the target object amidst numerous distractors. Under this setting, we reach a promising performance of " + }, + { + "bbox": [ + 46, + 231, + 287, + 339 + ], + "type": "inline_equation", + "content": "51.14\\%" + }, + { + "bbox": [ + 46, + 231, + 287, + 339 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 231, + 287, + 339 + ], + "type": "inline_equation", + "content": "40.08\\%" + }, + { + "bbox": [ + 46, + 231, + 287, + 339 + ], + "type": "text", + "content": ", providing further validation for the efficacy of " + }, + { + "bbox": [ + 46, + 231, + 287, + 339 + ], + "type": "inline_equation", + "content": "\\mathrm{G}^3" + }, + { + "bbox": [ + 46, + 231, + 287, + 339 + ], + "type": "text", + "content": "-LQ in addressing the Semantic-Geometric Coupling issue." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 339, + 287, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 339, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 46, + 339, + 287, + 543 + ], + "type": "text", + "content": "Performance on the Nr3D/Sr3D. The Nr3D/Sr3D task is geared towards locating the target object among all provided ground truth candidate boxes, representing a departure from the ScanRefer dataset. The overall accuracy of the proposed G" + }, + { + "bbox": [ + 46, + 339, + 287, + 543 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 46, + 339, + 287, + 543 + ], + "type": "text", + "content": "-LQ along with other exceptional methods are reported in Table. 2. We attain peak performance of " + }, + { + "bbox": [ + 46, + 339, + 287, + 543 + ], + "type": "inline_equation", + "content": "58.4\\%" + }, + { + "bbox": [ + 46, + 339, + 287, + 543 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 339, + 287, + 543 + ], + "type": "inline_equation", + "content": "73.1\\%" + }, + { + "bbox": [ + 46, + 339, + 287, + 543 + ], + "type": "text", + "content": " on Nr3D and Sr3D. In Nr3D, descriptions exhibit noteworthy intricacy and detail, inducing additional challenges to 3D VG task. However, our method demonstrates a noteworthy superiority over alternative methods that rely on 2D images priors [44] or 3D vision-language pre-training models [51]. In the more challenging \"Hard\" subset, our method notably improves the accuracy by " + }, + { + "bbox": [ + 46, + 339, + 287, + 543 + ], + "type": "inline_equation", + "content": "1.3\\%" + }, + { + "bbox": [ + 46, + 339, + 287, + 543 + ], + "type": "text", + "content": " in Nr3D and " + }, + { + "bbox": [ + 46, + 339, + 287, + 543 + ], + "type": "inline_equation", + "content": "2.7\\%" + }, + { + "bbox": [ + 46, + 339, + 287, + 543 + ], + "type": "text", + "content": " in Sr3D, which underscores our method is beneficial for distinguishing ambiguous objects by effectively modeling the underlying geometric shape and unraveling the complex relation between vision-text features." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 551, + 204, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 551, + 204, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 551, + 204, + 563 + ], + "type": "text", + "content": "4.2. Ablation Study and Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": "Effectiveness of Proposed Components. We develop several alternative designs and undertake experiments to highlight the advantages yielded by various components. We herein train our " + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{G}^3" + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": "-LQ model on the ScanRefer dataset. (a) outlines the performance demonstrated by EDA [43], while (h) is our proposed method. 1) Comparing (a) and (b), we witness that the PAGE module delivers a remarkable gain over the baseline model. The comparison further accentuates the paramount role of explicit 3D geometric features in modeling object shape and spatial relationship, conferring benefits to 3D visual grounding. 2) The comparative result of (b) and (g) furnishes compelling evidence that the Flan" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 308, + 71, + 545, + 160 + ], + "blocks": [ + { + "bbox": [ + 46, + 201, + 287, + 224 + ], + "lines": [ + { + "bbox": [ + 46, + 201, + 287, + 224 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 287, + 224 + ], + "type": "text", + "content": "Table 3. Ablation study on the effectiveness of the proposed PAGE, Flan-QS and PSA. Evaluated on the ScanRefer dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 71, + 545, + 160 + ], + "lines": [ + { + "bbox": [ + 308, + 71, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 308, + 71, + 545, + 160 + ], + "type": "table", + "html": "
pipi - pijpi - pijUnique(0.25)Multiple(0.25)
✓——85.6749.13
✓✓—85.8349.19
✓—maxpool85.9849.32
✓—softmax86.0449.37
✓—Afine86.5750.06
✓✓Afine86.8950.46
", + "image_path": "8e975b4519dcf5d66feef533930c04b9678fc00aeb4f282547f7ab258a7c7653.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 307, + 203, + 547, + 282 + ], + "blocks": [ + { + "bbox": [ + 305, + 168, + 545, + 190 + ], + "lines": [ + { + "bbox": [ + 305, + 168, + 545, + 190 + ], + "spans": [ + { + "bbox": [ + 305, + 168, + 545, + 190 + ], + "type": "text", + "content": "Table 4. Ablation study of the geometric priors selection in the proposed PAGE module on the ScanRefer dataset." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 203, + 547, + 282 + ], + "lines": [ + { + "bbox": [ + 307, + 203, + 547, + 282 + ], + "spans": [ + { + "bbox": [ + 307, + 203, + 547, + 282 + ], + "type": "table", + "html": "
Query SelectionUnique(~19%)Multiple(~81%)
0.250.50.250.5
Parametric81.4060.1338.0228.05
Top-K(EDA)85.6768.5749.1337.64
Language87.1071.5950.6339.31
Language Graph87.6672.1650.8740.05
", + "image_path": "8c58b4aad3741167102286b889d57daa69d2b0895c35c34aa9340b1622cc75f5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 290, + 545, + 324 + ], + "lines": [ + { + "bbox": [ + 305, + 290, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 305, + 290, + 545, + 324 + ], + "type": "text", + "content": "Table 5. Ablation study of the query selection. \"Parametric\" represents the selection strategy with random parametric queries. \"Language\" denotes the guidance of sentence-level language." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 329, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 329, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 329, + 545, + 628 + ], + "type": "text", + "content": "QS module exploiting language prior and context dependency, helps to alleviate the problem of grounding ambiguity and generate precise query proposals. 3) Comparing (h) and (g), it showcases the effectiveness of the proposed PSA loss to capture complex non-linear mappings of point cloud and text embeddings on Poincaré ball. When we integrate all proposed modules, a discernible surge is observed, yielding an impressive overall accuracy of " + }, + { + "bbox": [ + 304, + 329, + 545, + 628 + ], + "type": "inline_equation", + "content": "56.90\\%" + }, + { + "bbox": [ + 304, + 329, + 545, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 329, + 545, + 628 + ], + "type": "inline_equation", + "content": "45.58\\%" + }, + { + "bbox": [ + 304, + 329, + 545, + 628 + ], + "type": "text", + "content": ". Selection of Geometric Priors. The fundamental objective of the PAGE module lies in embedding low-level geometric priors into high-level geometric features, thus the definition of " + }, + { + "bbox": [ + 304, + 329, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\tilde{p}" + }, + { + "bbox": [ + 304, + 329, + 545, + 628 + ], + "type": "text", + "content": " emerges as a topic deserving exploration. We undertake ablations with " + }, + { + "bbox": [ + 304, + 329, + 545, + 628 + ], + "type": "inline_equation", + "content": "p_i, p_i - p_{ij}, \\| p_i - p_{ij} \\|" + }, + { + "bbox": [ + 304, + 329, + 545, + 628 + ], + "type": "text", + "content": " as illustrative examples, which unveils the inherent relative position, shapes, and distance (see Fig. 3) among 3D objects. As shown in Table. 4, using " + }, + { + "bbox": [ + 304, + 329, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\| p_i - p_{ij} \\|" + }, + { + "bbox": [ + 304, + 329, + 545, + 628 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 304, + 329, + 545, + 628 + ], + "type": "inline_equation", + "content": "p_i - p_{ij}" + }, + { + "bbox": [ + 304, + 329, + 545, + 628 + ], + "type": "text", + "content": " alone yields inferior performance compared with their gradual integration, the accuracy of which reach " + }, + { + "bbox": [ + 304, + 329, + 545, + 628 + ], + "type": "inline_equation", + "content": "86.89\\%" + }, + { + "bbox": [ + 304, + 329, + 545, + 628 + ], + "type": "text", + "content": " (Uniqu) and " + }, + { + "bbox": [ + 304, + 329, + 545, + 628 + ], + "type": "inline_equation", + "content": "50.46\\%" + }, + { + "bbox": [ + 304, + 329, + 545, + 628 + ], + "type": "text", + "content": " (Multiple) respectively. Further, when compared with max pooling and softmax operators for features aggregation in " + }, + { + "bbox": [ + 304, + 329, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_i" + }, + { + "bbox": [ + 304, + 329, + 545, + 628 + ], + "type": "text", + "content": ", our geometric refinement strategy attests to an improved performance by " + }, + { + "bbox": [ + 304, + 329, + 545, + 628 + ], + "type": "inline_equation", + "content": "0.59\\%" + }, + { + "bbox": [ + 304, + 329, + 545, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 329, + 545, + 628 + ], + "type": "inline_equation", + "content": "0.59\\%" + }, + { + "bbox": [ + 304, + 329, + 545, + 628 + ], + "type": "text", + "content": " (Unique). The noteworthy results accentuate the significance of the geometric refinement in the PAGE module, given its potential to capture position-adaptive geometric structures." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "Methods of Query Selection. We carry out several experiments to validate the proposed query selection method. We delineate the overall performance on the ScanRefer dataset in Table. 5. 1) We adopt the parametric queries used in MDETR [22]. However, we observe that the random parametric queries in performance decline for its incapability to capture contextual information of the text and 3D scene. 2)" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13923" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 87, + 164, + 147 + ], + "blocks": [ + { + "bbox": [ + 80, + 77, + 156, + 87 + ], + "lines": [ + { + "bbox": [ + 80, + 77, + 156, + 87 + ], + "spans": [ + { + "bbox": [ + 80, + 77, + 156, + 87 + ], + "type": "text", + "content": "(a) Ordinal Numbers" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 62, + 87, + 164, + 147 + ], + "lines": [ + { + "bbox": [ + 62, + 87, + 164, + 147 + ], + "spans": [ + { + "bbox": [ + 62, + 87, + 164, + 147 + ], + "type": "image", + "image_path": "afa2de50b5088c37cdffed8fbad6598f569c7524c74f7b598ac6ae9a0aabd718.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 148, + 167, + 161 + ], + "lines": [ + { + "bbox": [ + 74, + 148, + 167, + 161 + ], + "spans": [ + { + "bbox": [ + 74, + 148, + 167, + 161 + ], + "type": "text", + "content": "this is a chair by the wall. It is the fifth chair from the right wall surface." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 173, + 90, + 291, + 142 + ], + "blocks": [ + { + "bbox": [ + 192, + 77, + 271, + 87 + ], + "lines": [ + { + "bbox": [ + 192, + 77, + 271, + 87 + ], + "spans": [ + { + "bbox": [ + 192, + 77, + 271, + 87 + ], + "type": "text", + "content": "(b) Distance and Size" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 173, + 90, + 291, + 142 + ], + "lines": [ + { + "bbox": [ + 173, + 90, + 291, + 142 + ], + "spans": [ + { + "bbox": [ + 173, + 90, + 291, + 142 + ], + "type": "image", + "image_path": "2f43e6854e8502f292069b36f782546485f86068fa07d523221d3719171f87c6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 67, + 163, + 164, + 228 + ], + "blocks": [ + { + "bbox": [ + 67, + 163, + 164, + 228 + ], + "lines": [ + { + "bbox": [ + 67, + 163, + 164, + 228 + ], + "spans": [ + { + "bbox": [ + 67, + 163, + 164, + 228 + ], + "type": "image", + "image_path": "16400971a747609d2f1074db44477d3041827196f210d758be476fb7920178de.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 192, + 163, + 288, + 231 + ], + "blocks": [ + { + "bbox": [ + 187, + 148, + 280, + 161 + ], + "lines": [ + { + "bbox": [ + 187, + 148, + 280, + 161 + ], + "spans": [ + { + "bbox": [ + 187, + 148, + 280, + 161 + ], + "type": "text", + "content": "this is a grey chair. its at the end of the shorter table closest to the window." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 192, + 163, + 288, + 231 + ], + "lines": [ + { + "bbox": [ + 192, + 163, + 288, + 231 + ], + "spans": [ + { + "bbox": [ + 192, + 163, + 288, + 231 + ], + "type": "image", + "image_path": "ae0582df79992fb5178ef53b9d48cd314cb0a681bdacb7883bbc7e1f1e67db1e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 183, + 233, + 284, + 253 + ], + "lines": [ + { + "bbox": [ + 183, + 233, + 284, + 253 + ], + "spans": [ + { + "bbox": [ + 183, + 233, + 284, + 253 + ], + "type": "text", + "content": "we are looking for the chair on the right. we are not looking for the chair on the left. the chair on the left is closest to the bed." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 62, + 255, + 179, + 305 + ], + "blocks": [ + { + "bbox": [ + 78, + 232, + 162, + 254 + ], + "lines": [ + { + "bbox": [ + 78, + 232, + 162, + 254 + ], + "spans": [ + { + "bbox": [ + 78, + 232, + 162, + 254 + ], + "type": "text", + "content": "it is a black office chair. The black office chair is the fourth office chair on the left side of the table." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 62, + 255, + 179, + 305 + ], + "lines": [ + { + "bbox": [ + 62, + 255, + 179, + 305 + ], + "spans": [ + { + "bbox": [ + 62, + 255, + 179, + 305 + ], + "type": "image", + "image_path": "f568db9e3a1ffd529047e19383418b3901b3e3172830242fef75bbefa271bd67.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 76, + 309, + 165, + 323 + ], + "lines": [ + { + "bbox": [ + 76, + 309, + 165, + 323 + ], + "spans": [ + { + "bbox": [ + 76, + 309, + 165, + 323 + ], + "type": "text", + "content": "there is a gray chair. it is the second closest one to a trash can." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 182, + 255, + 290, + 306 + ], + "blocks": [ + { + "bbox": [ + 182, + 255, + 290, + 306 + ], + "lines": [ + { + "bbox": [ + 182, + 255, + 290, + 306 + ], + "spans": [ + { + "bbox": [ + 182, + 255, + 290, + 306 + ], + "type": "image", + "image_path": "8d92fcdc30592e09814481c6db15b0a894a2c67532829771b5a2f34cc790ece2.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 193, + 307, + 279, + 327 + ], + "lines": [ + { + "bbox": [ + 193, + 307, + 279, + 327 + ], + "spans": [ + { + "bbox": [ + 193, + 307, + 279, + 327 + ], + "type": "text", + "content": "the table is the left-most one in the center of the room. The table is dark orange and has four legs." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 297, + 90, + 418, + 144 + ], + "blocks": [ + { + "bbox": [ + 307, + 77, + 383, + 87 + ], + "lines": [ + { + "bbox": [ + 307, + 77, + 383, + 87 + ], + "spans": [ + { + "bbox": [ + 307, + 77, + 383, + 87 + ], + "type": "text", + "content": "(c) Geometric shape" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 297, + 90, + 418, + 144 + ], + "lines": [ + { + "bbox": [ + 297, + 90, + 418, + 144 + ], + "spans": [ + { + "bbox": [ + 297, + 90, + 418, + 144 + ], + "type": "image", + "image_path": "3e39a366eda7b48c0f14f033612ddafe87c082bbfb98d25a3cc5dda28c4e36ce.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 144, + 395, + 165 + ], + "lines": [ + { + "bbox": [ + 304, + 144, + 395, + 165 + ], + "spans": [ + { + "bbox": [ + 304, + 144, + 395, + 165 + ], + "type": "text", + "content": "standing semi-circular wooden table with flat surface for holding objects. object can be found in center of image" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 415, + 140, + 525, + 165 + ], + "lines": [ + { + "bbox": [ + 415, + 140, + 525, + 165 + ], + "spans": [ + { + "bbox": [ + 415, + 140, + 525, + 165 + ], + "type": "text", + "content": "in front of a brown wooden desk are two white chairs. we are looking for the chair on the right. we are not looking for the chair on the left. the chair on the left is closest to the bed." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 290, + 168, + 397, + 231 + ], + "blocks": [ + { + "bbox": [ + 290, + 168, + 397, + 231 + ], + "lines": [ + { + "bbox": [ + 290, + 168, + 397, + 231 + ], + "spans": [ + { + "bbox": [ + 290, + 168, + 397, + 231 + ], + "type": "image", + "image_path": "4cbc3c1692187fd558fc3991e33f777105f30c9d3a86cf10ab7f55e7e2edc167.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 304, + 251, + 398, + 310 + ], + "blocks": [ + { + "bbox": [ + 300, + 235, + 403, + 248 + ], + "lines": [ + { + "bbox": [ + 300, + 235, + 403, + 248 + ], + "spans": [ + { + "bbox": [ + 300, + 235, + 403, + 248 + ], + "type": "text", + "content": "The towel is between the entrance door and the sink. The towel is olive green and wavy." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 251, + 398, + 310 + ], + "lines": [ + { + "bbox": [ + 304, + 251, + 398, + 310 + ], + "spans": [ + { + "bbox": [ + 304, + 251, + 398, + 310 + ], + "type": "image", + "image_path": "cc430a091aaf8b0aa4e914b62a54bc497e4ef0cf530ade1159608d2ed4ee9fce.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 299, + 311, + 400, + 323 + ], + "lines": [ + { + "bbox": [ + 299, + 311, + 400, + 323 + ], + "spans": [ + { + "bbox": [ + 299, + 311, + 400, + 323 + ], + "type": "text", + "content": "the end table is the closest one to the window. The end table is a round cylinder." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 332, + 545, + 355 + ], + "lines": [ + { + "bbox": [ + 46, + 332, + 545, + 355 + ], + "spans": [ + { + "bbox": [ + 46, + 332, + 545, + 355 + ], + "type": "text", + "content": "Figure 3. Qualitative results of the EDA [43] and our proposed " + }, + { + "bbox": [ + 46, + 332, + 545, + 355 + ], + "type": "inline_equation", + "content": "\\mathrm{G}^3" + }, + { + "bbox": [ + 46, + 332, + 545, + 355 + ], + "type": "text", + "content": "-LQ on the ScanRefer dataset. Our " + }, + { + "bbox": [ + 46, + 332, + 545, + 355 + ], + "type": "inline_equation", + "content": "\\mathrm{G}^3" + }, + { + "bbox": [ + 46, + 332, + 545, + 355 + ], + "type": "text", + "content": "-LQ manifests excellent grounding performance in terms of ordinal numbers, spatial distance or object size, geometric attribute and complex utterances." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 427, + 87, + 503, + 139 + ], + "blocks": [ + { + "bbox": [ + 422, + 77, + 506, + 87 + ], + "lines": [ + { + "bbox": [ + 422, + 77, + 506, + 87 + ], + "spans": [ + { + "bbox": [ + 422, + 77, + 506, + 87 + ], + "type": "text", + "content": "(d) Complex Utterance" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 427, + 87, + 503, + 139 + ], + "lines": [ + { + "bbox": [ + 427, + 87, + 503, + 139 + ], + "spans": [ + { + "bbox": [ + 427, + 87, + 503, + 139 + ], + "type": "image", + "image_path": "b17cc1dc1cc08e9c56d214e641f83e7bf89334ce1eb4550da8c2a45d5122ef74.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 426, + 167, + 507, + 221 + ], + "blocks": [ + { + "bbox": [ + 426, + 167, + 507, + 221 + ], + "lines": [ + { + "bbox": [ + 426, + 167, + 507, + 221 + ], + "spans": [ + { + "bbox": [ + 426, + 167, + 507, + 221 + ], + "type": "image", + "image_path": "aa76ba1a2154fb8d23ab8babd93593659a8921ce769d955f5eb2a025fa5b6c9c.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 419, + 222, + 525, + 255 + ], + "lines": [ + { + "bbox": [ + 419, + 222, + 525, + 255 + ], + "spans": [ + { + "bbox": [ + 419, + 222, + 525, + 255 + ], + "type": "text", + "content": "this is the brown arm chair that is at the end of the chalk board, to the left of it there is a small bookshelf, to the right of it and at the table is a blue chair. The brown chair is not under the table it is just next to the table." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 419, + 258, + 528, + 301 + ], + "blocks": [ + { + "bbox": [ + 419, + 258, + 528, + 301 + ], + "lines": [ + { + "bbox": [ + 419, + 258, + 528, + 301 + ], + "spans": [ + { + "bbox": [ + 419, + 258, + 528, + 301 + ], + "type": "image", + "image_path": "a228d9520a5e582e20fe4624395155954ed62515275a16cc27de00e1472903a4.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 419, + 301, + 530, + 328 + ], + "lines": [ + { + "bbox": [ + 419, + 301, + 530, + 328 + ], + "spans": [ + { + "bbox": [ + 419, + 301, + 530, + 328 + ], + "type": "text", + "content": "walking into the room, a large office desk is in the middle of the room. On the right side of the desk, there is a computer monitor on the table, the one on the left side, closest to the door." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "bbox": [ + 46, + 365, + 287, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 287, + 485 + ], + "type": "text", + "content": "We generate the query proposals guided by the sentence-level features. The performance of the language-guided query selection mechanism stands out prominently, underscoring the crucial role of text features in the query generation. 3) Our Flan-QS module showcases superior performance over the prevailing top- " + }, + { + "bbox": [ + 46, + 365, + 287, + 485 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 365, + 287, + 485 + ], + "type": "text", + "content": " selection and sentence-level methods. The results give explanation of the efficacy of explicit fine-grained text features. They precisely model the attributes and positions of target objects, thereby facilitating a more accurate generation of query proposals." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 47, + 494, + 187, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 494, + 187, + 506 + ], + "spans": [ + { + "bbox": [ + 47, + 494, + 187, + 506 + ], + "type": "text", + "content": "4.3. Qualitative Comparisons" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 46, + 510, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 288, + 715 + ], + "type": "text", + "content": "To offer profound insight into the effectiveness of the proposed " + }, + { + "bbox": [ + 46, + 510, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{G}^3" + }, + { + "bbox": [ + 46, + 510, + 288, + 715 + ], + "type": "text", + "content": "-LQ, we visualize the grounding results of our method alongside the SOTA EDA [43]. As shown in Fig. 3, the predicted boxes of our " + }, + { + "bbox": [ + 46, + 510, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{G}^3" + }, + { + "bbox": [ + 46, + 510, + 288, + 715 + ], + "type": "text", + "content": "-LQ method and EDA are visually highlighted in yellow and red, with ground-truth boxes marked in green. The localization ability of EDA is found wanting, encountering challenges in handling the 3D objects described by ordinal numbers, relative distance, and geometric shapes (see Fig. 3(a)-(c)). Conversely, our " + }, + { + "bbox": [ + 46, + 510, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{G}^3" + }, + { + "bbox": [ + 46, + 510, + 288, + 715 + ], + "type": "text", + "content": "-LQ incorporates intrinsic geometric features and spatial relation, fostering an enriched understanding of both the scene and objects, manifests commendable grounding prowess. Secondly, Fig. 3(d) exemplifies the commendable capabilities of our method in grounding 3D objects narrated by complex utterances. To elucidate, the " + }, + { + "bbox": [ + 46, + 510, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{G}^3" + }, + { + "bbox": [ + 46, + 510, + 288, + 715 + ], + "type": "text", + "content": "-LQ model with the tailored Flan-QS, excels in harnessing language priors and capturing global context dependencies, generat-" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 306, + 365, + 528, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 365, + 528, + 376 + ], + "spans": [ + { + "bbox": [ + 306, + 365, + 528, + 376 + ], + "type": "text", + "content": "ing high-quality semantic-aware query representations." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 306, + 383, + 378, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 383, + 378, + 395 + ], + "spans": [ + { + "bbox": [ + 306, + 383, + 378, + 395 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 304, + 399, + 545, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 545, + 602 + ], + "type": "text", + "content": "In this paper, we analysis the Semantic-Geometric Coupling issue of 3D VG tasks in three aspects: intricate geometric details, complex text descriptions, and semantic-geometric inconsistency. To this end, we propose a " + }, + { + "bbox": [ + 304, + 399, + 545, + 602 + ], + "type": "inline_equation", + "content": "\\mathrm{G}^3" + }, + { + "bbox": [ + 304, + 399, + 545, + 602 + ], + "type": "text", + "content": "-LQ framework to explicitly capture the fine-grained geometric and semantic concepts, which includes three crucial novelties. Firstly, a Position Adaptive Geometric Exploring module is designed to capture explicit geometric features. Secondly, with decoupled multiple semantic components, a Fine-grained Language-guided Query Selection module generates object queries densely aligned by fine-grained text features. Finally, a Poincaré Semantic Alignment loss capitalizes on vision-text hierarchy natures and achieves alignment in Poincaré space, encouraging semantic-geometric consistency. Experimental findings showcase the superior performance of our " + }, + { + "bbox": [ + 304, + 399, + 545, + 602 + ], + "type": "inline_equation", + "content": "\\mathrm{G}^3" + }, + { + "bbox": [ + 304, + 399, + 545, + 602 + ], + "type": "text", + "content": "-LQ across the ScanRefer and Nr3D/Sr3D benchmarks, establishing a new state-of-the-art standard." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 306, + 609, + 416, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 609, + 416, + 623 + ], + "spans": [ + { + "bbox": [ + 306, + 609, + 416, + 623 + ], + "type": "text", + "content": "6. Acknowledgement" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "This work is supported by the state key development program in 14th Five-Year under Grant No. 2021YFF0602103, 2021YFF0602102, 2021QY1702. Thanks for the research fund under Grant No. 2019GQG0001 from the Institute for Guo Qiang, Tsinghua University. This study is also supported by Tsinghua University Toyota Joint Research Center for AI Technology of Automated Vehicle(TTAD 2023-07)." + } + ] + } + ], + "index": 36 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13924" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 112 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 112 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 112 + ], + "type": "text", + "content": "[1] Scene graph parser. https://github.com/vacancy/SceneGraphParser, 2019.5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 114, + 286, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 286, + 180 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 286, + 180 + ], + "type": "text", + "content": "[2] Ahmed Abdelreheem, Ujjwal Upadhyay, Ivan Skorokhodov, Rawan Al Yahya, Jun Chen, and Mohamed Elhoseiny. 3dref-transformer: Fine-grained object identification in real-world scenes using natural language. In Proceedings of the IEEE Winter Conference on Applications of Computer Vision, pages 3941-3950, 2022. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 286, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 286, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 286, + 237 + ], + "type": "text", + "content": "[3] Panos Achlioptas, Ahmed Abdelreehem, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In Proceedings of the European Conference on Computer Vision, pages 422-440. Springer, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 239, + 286, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 286, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 286, + 282 + ], + "type": "text", + "content": "[4] Eslam Bakr, Yasmeen Alsaeda, and Mohamed Elhoseiny. Look around and refer: 2d synthetic semantics knowledge distillation for 3d visual grounding. Advances in Neural Information Processing Systems, 35:37146-37158, 2022. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 285, + 286, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 285, + 286, + 339 + ], + "spans": [ + { + "bbox": [ + 53, + 285, + 286, + 339 + ], + "type": "text", + "content": "[5] Daigang Cai, Lichen Zhao, Jing Zhang, Lu Sheng, and Dong Xu. 3djcg: A unified framework for joint dense captioning and visual grounding on 3d point clouds. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 16464-16473, 2022. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 341, + 286, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 341, + 286, + 385 + ], + "spans": [ + { + "bbox": [ + 53, + 341, + 286, + 385 + ], + "type": "text", + "content": "[6] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In Proceedings of the European Conference on Computer Vision, pages 202-221. Springer, 2020. 1, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 386, + 286, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 386, + 286, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 386, + 286, + 441 + ], + "type": "text", + "content": "[7] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Language conditioned spatial relation reasoning for 3d object grounding. Advances in Neural Information Processing Systems, 35:20522-20535, 2022. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 443, + 286, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 443, + 286, + 487 + ], + "spans": [ + { + "bbox": [ + 53, + 443, + 286, + 487 + ], + "type": "text", + "content": "[8] Jaesung Choe, Chunghyun Park, Francois Rameau, and Jaesik Park. Pointmixer: Mlp-mixer for point cloud understanding. In Proceedings of the European Conference on Computer Vision, pages 620-640. Springer, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 489, + 286, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 489, + 286, + 532 + ], + "spans": [ + { + "bbox": [ + 53, + 489, + 286, + 532 + ], + "type": "text", + "content": "[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 534, + 286, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 286, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 286, + 578 + ], + "type": "text", + "content": "[10] Jiafei Duan, Samson Yu, Hui Li Tan, Hongyuan Zhu, and Cheston Tan. A survey of embodied ai: From simulators to research tasks. IEEE Transactions on Emerging Topics in Computational Intelligence, 6(2):230-244, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 580, + 286, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 286, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 286, + 644 + ], + "type": "text", + "content": "[11] Aleksandr Ermolov, Leyla Mirvakhabova, Valentin Khrulkov, Nicu Sebe, and Ivan Oseledets. Hyperbolic vision transformers: Combining improvements in metric learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7409-7419, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 286, + 713 + ], + "type": "text", + "content": "[12] Mingtao Feng, Zhen Li, Qi Li, Liang Zhang, XiangDong Zhang, Guangming Zhu, Hui Zhang, Yaonan Wang, and Ajmal Mian. Free-form description guided 3d visual graph network for object grounding in point cloud. In Proceedings of the IEEE International Conference on Computer Vision, pages 3722-3731, 2021. 2, 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[13] Octavian Ganea, Gary Bécigneul, and Thomas Hofmann. Hyperbolic neural networks. Advances in Neural Information Processing Systems, 31, 2018. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 107, + 545, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 148 + ], + "type": "text", + "content": "[14] William L Hamilton, Rex Ying, and Jure Leskovec. Inductive representation learning on large graphs. In Advances in Neural Information Processing Systems, pages 1025-1035, 2017. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 150, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 150, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 150, + 545, + 205 + ], + "type": "text", + "content": "[15] Dailan He, Yusheng Zhao, Junyu Luo, Tianrui Hui, Shaofei Huang, Aixi Zhang, and Si Liu. Transrefer3d: Entity-andrelation aware transformer for fine-grained 3d visual grounding. In Proceedings of the 29th ACM International Conference on Multimedia, pages 2344-2352, 2021. 2, 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 206, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 206, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 307, + 206, + 545, + 259 + ], + "type": "text", + "content": "[16] Pin-Hao Huang, Han-Hung Lee, Hwann-Tzong Chen, and Tyng-Luh Liu. Text-guided graph neural networks for referring 3d instance segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1610-1618, 2021. 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 261, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 261, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 261, + 545, + 304 + ], + "type": "text", + "content": "[17] Shijia Huang, Yilun Chen, Jiaya Jia, and Liwei Wang. Multiview transformer for 3d visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15524-15533, 2022. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 305, + 545, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 305, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 307, + 305, + 545, + 358 + ], + "type": "text", + "content": "[18] Wenlong Huang, Pieter Abbeel, Deepak Pathak, and Igor Mordatch. Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. In International Conference on Machine Learning, pages 9118-9147. PMLR, 2022. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 360, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 360, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 360, + 545, + 415 + ], + "type": "text", + "content": "[19] Ayush Jain, Nikolaos Gkanatsios, Ishita Mediratta, and Katerina Fragkiadaki. Bottom up top down detection transformers for language grounding in images and point clouds. In Proceedings of the European Conference on Computer Vision, pages 417-433. Springer, 2022. 2, 3, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 415, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 415, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 307, + 415, + 545, + 469 + ], + "type": "text", + "content": "[20] Haojun Jiang, Yuanze Lin, Dongchen Han, Shiji Song, and Gao Huang. Pseudo-q: Generating pseudo language queries for visual grounding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 15513-15523, 2022. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 471, + 545, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 471, + 545, + 524 + ], + "spans": [ + { + "bbox": [ + 307, + 471, + 545, + 524 + ], + "type": "text", + "content": "[21] Zhao Jin, Munawar Hayat, Yuwei Yang, Yulan Guo, and Yinjie Lei. Context-aware alignment and mutual masking for 3d-language pre-training. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 10984-10994, 2023. 2, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 525, + 545, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 525, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 307, + 525, + 545, + 580 + ], + "type": "text", + "content": "[22] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE International Conference on Computer Vision, pages 1780-1790, 2021. 3, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 582, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 582, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 307, + 582, + 545, + 613 + ], + "type": "text", + "content": "[23] Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In International Conference on Learning Representations, 2017. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 614, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 614, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 307, + 614, + 545, + 657 + ], + "type": "text", + "content": "[24] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329, 2022. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "text", + "content": "[25] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 3, 5" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13925" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[26] Yongcheng Liu, Bin Fan, Shiming Xiang, and Chunhong Pan. Relation-shape convolutional neural network for point cloud analysis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8895-8904, 2019. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 288, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 184 + ], + "type": "text", + "content": "[27] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 185, + 288, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 288, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 288, + 229 + ], + "type": "text", + "content": "[28] Yongfei Liu, Bo Wan, Xiaodan Zhu, and Xuming He. Learning cross-modal context graph for visual grounding. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11645-11652, 2020. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 230, + 288, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 288, + 284 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 288, + 284 + ], + "type": "text", + "content": "[29] Yongfei Liu, Bo Wan, Lin Ma, and Xuming He. Relation-aware instance refinement for weakly supervised visual grounding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5612-5621, 2021. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 285, + 288, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 285, + 288, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 285, + 288, + 330 + ], + "type": "text", + "content": "[30] Ze Liu, Zheng Zhang, Yue Cao, Han Hu, and Xin Tong. Group-free 3d object detection via transformers. In Proceedings of the IEEE International Conference on Computer Vision, pages 2949-2958, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 331, + 288, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 288, + 396 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 288, + 396 + ], + "type": "text", + "content": "[31] Junyu Luo, Jiahui Fu, Xianghao Kong, Chen Gao, Haibing Ren, Hao Shen, Huaxia Xia, and Si Liu. 3d-sps: Single-stage 3d visual grounding via referred point progressive selection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 16454-16463, 2022. 1, 2, 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 398, + 288, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 398, + 288, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 398, + 288, + 443 + ], + "type": "text", + "content": "[32] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. In International Conference on Learning Representations, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 444, + 288, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 444, + 288, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 444, + 288, + 487 + ], + "type": "text", + "content": "[33] Antonio Montanaro, Diego Valsesia, and Enrico Magli. Rethinking the compositionality of point clouds through regularization in the hyperbolic space. Advances in Neural Information Processing Systems, 35:33741-33753, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 488, + 288, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 488, + 288, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 488, + 288, + 521 + ], + "type": "text", + "content": "[34] Maximillian Nickel and Douwe Kiela. Poincaré embeddings for learning hierarchical representations. Advances in Neural Information Processing Systems, 30, 2017. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 522, + 288, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 288, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 288, + 588 + ], + "type": "text", + "content": "[35] Aishwarya Padmakumar, Jesse Thomason, Ayush Shrivastava, Patrick Lange, Anjali Narayan-Chen, Spandana Gella, Robinson Piramuthu, Gokhan Tur, and Dilek Hakkani-Tur. Teach: Task-driven embodied agents that chat. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2017–2025, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 590, + 288, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 288, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 288, + 634 + ], + "type": "text", + "content": "[36] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in Neural Information Processing Systems, 30, 2017. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 288, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 679 + ], + "type": "text", + "content": "[37] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In proceedings of the IEEE International Conference on Computer Vision, pages 9277-9286, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "text", + "content": "[38] Zheng Qin, Hao Yu, Changjian Wang, Yulan Guo, Yuxing Peng, and Kai Xu. Geometric transformer for fast and robust point cloud registration. In Proceedings of the IEEE Con-" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "ference on Computer Vision and Pattern Recognition, pages 11143-11152, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 95, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 545, + 139 + ], + "type": "text", + "content": "[39] Junha Roh, Karthik Desingh, Ali Farhadi, and Dieter Fox. Languagerefer: Spatial-language model for 3d visual grounding. In Conference on Robot Learning, pages 1046-1056. PMLR, 2022. 3, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 140, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 184 + ], + "type": "text", + "content": "[40] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 30, 2017. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 185, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 545, + 228 + ], + "type": "text", + "content": "[41] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (TOG), 38(5):1-12, 2019. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 228, + 545, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 545, + 272 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 545, + 272 + ], + "type": "text", + "content": "[42] Yingming Wang, Xiangyu Zhang, Tong Yang, and Jian Sun. Anchor detr: Query design for transformer-based detector. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2567-2575, 2022. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 272, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 272, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 307, + 272, + 545, + 316 + ], + "type": "text", + "content": "[43] Yanmin Wu, Xinhua Cheng, Renrui Zhang, Zesen Cheng, and Jian Zhang. Eda: Explicit text-decoupling and dense alignment for 3d visual and language learning. arXiv preprint arXiv:2209.14941, 2022. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 316, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 316, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 307, + 316, + 545, + 360 + ], + "type": "text", + "content": "[44] Zhengyuan Yang, Songyang Zhang, and Jiebo Luo. Sat: 2d semantics assisted training for 3d visual grounding. In Proceedings of the IEEE International Conference on Computer Vision, pages 1856-1866, 2021. 2, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 361, + 545, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 361, + 545, + 405 + ], + "spans": [ + { + "bbox": [ + 307, + 361, + 545, + 405 + ], + "type": "text", + "content": "[45] Ting Yao, Yehao Li, and Tao Mei. Hgnet: Learning hierarchical geometry from points, edges, and surfaces. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 21846-21855, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 405, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 405, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 307, + 405, + 545, + 459 + ], + "type": "text", + "content": "[46] Xumin Yu, Yongming Rao, Ziyi Wang, Zuyan Liu, Jiwen Lu, and Jie Zhou. Pointr: Diverse point cloud completion with geometry-aware transformers. In Proceedings of the IEEE International Conference on Computer Vision, pages 12498-12507, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 459, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 459, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 307, + 459, + 545, + 525 + ], + "type": "text", + "content": "[47] Zhihao Yuan, Xu Yan, Yinghong Liao, Ruimao Zhang, Sheng Wang, Zhen Li, and Shuguang Cui. Instancerefer: Cooperative holistic understanding for visual grounding on point clouds through instance multi-level contextual referring. In Proceedings of the IEEE International Conference on Computer Vision, pages 1791-1800, 2021. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 525, + 545, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 525, + 545, + 570 + ], + "spans": [ + { + "bbox": [ + 307, + 525, + 545, + 570 + ], + "type": "text", + "content": "[48] Lichen Zhao, Daigang Cai, Lu Sheng, and Dong Xu. 3dvg-transformer: relation modeling for visual grounding on point clouds. In Proceedings of the IEEE International Conference on Computer Vision, pages 2928-2937, 2021. 1, 2, 3, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 570, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 570, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 570, + 545, + 624 + ], + "type": "text", + "content": "[49] Chaoyang Zhu, Yiyi Zhou, Yunhang Shen, Gen Luo, Xingjia Pan, Mingbao Lin, Chao Chen, Liujuan Cao, and Xiaoshuai Sun. Seqtr: A simple yet universal network for visual grounding. In Proceedings of the European Conference on Computer Vision, pages 598-615. Springer, 2022. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 625, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 625, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 545, + 669 + ], + "type": "text", + "content": "[50] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "text", + "content": "[51] Ziyu Zhu, Xiaojian Ma, Yixin Chen, Zhidong Deng, Siyuan Huang, and Qing Li. 3d-vista: Pre-trained transformer for 3d vision and text alignment. arXiv preprint arXiv:2308.04352, 2023. 2, 6, 7" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "13926" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Garment Recovery with Shape and Deformation Priors/83e37ee5-1c15-40f2-ac5d-8e3141ed428d_content_list.json b/2024/Garment Recovery with Shape and Deformation Priors/83e37ee5-1c15-40f2-ac5d-8e3141ed428d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..004542b4e43a8cf539caab7a2c92c8bb15361f5e --- /dev/null +++ b/2024/Garment Recovery with Shape and Deformation Priors/83e37ee5-1c15-40f2-ac5d-8e3141ed428d_content_list.json @@ -0,0 +1,1541 @@ +[ + { + "type": "text", + "text": "Garment Recovery with Shape and Deformation Priors", + "text_level": 1, + "bbox": [ + 205, + 130, + 764, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ren Li Corentin Dumery Benoit Guillard Pascal Fua CVLab,EPFL", + "bbox": [ + 245, + 180, + 723, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{name.surname}@epfl.ch", + "bbox": [ + 385, + 219, + 583, + 233 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/95a9792aeb4fd3cb1239f78ae40ab874c16733acbb8f0c4be1f2d1ec0a8a3629.jpg", + "image_caption": [ + "Figure 1. We propose a fitting method that leverages shape and deformation priors trained on synthetic data to recover realistic 3D garment mesh from in-the-wild images. We produce triangulated meshes that are directly usable for animation and simulation." + ], + "image_footnote": [], + "bbox": [ + 89, + 247, + 890, + 415 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 455, + 313, + 472 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While modeling people wearing tight-fitting clothing has made great strides in recent years, loose-fitting clothing remains a challenge. We propose a method that delivers realistic garment models from real-world images, regardless of garment shape or deformation. To this end, we introduce a fitting approach that utilizes shape and deformation priors learned from synthetic data to accurately capture garment shapes and deformations, including large ones. Not only does our approach recover the garment geometry accurately, it also yields models that can be directly used by downstream applications such as animation and simulation.", + "bbox": [ + 75, + 489, + 470, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 691, + 209, + 707 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D clothed human recovery aims to reconstruct the body shape, pose, and clothing of many different people from images. It is key to applications such as fashion design, virtual try-on, 3D avatars, along with virtual and augmented reality. Recent years have seen tremendous progress in modeling people wearing tight-fitting clothing both in terms of body poses [6, 14, 21, 22, 25, 26, 28, 30, 38, 41, 54] and 3D shape of the clothes [5, 9-11, 20, 31, 32, 39]. However, loose-fitting clothing remains a challenge. Existing approaches either rely on mesh templates with limited generality, or produce models expressed in terms of 3D point clouds [37] or as a single watertight mesh that tightly binds the body", + "bbox": [ + 75, + 719, + 468, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and garment together [53], neither of which is straightforward to integrate into downstream applications.", + "bbox": [ + 498, + 458, + 890, + 488 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose a method that overcomes these limitations and can effectively recover the shape of loose fitting garments from single images. The recovered garments can then be animated without any additional processing, as shown in Fig. 1. Starting from the Implicit Sewing Patterns (ISP) model [32] that represents garments in terms of a set of individual 2D panels and 3D surfaces associated to these panels, we introduce a deformation model that we apply to the 3D surfaces so that they can deviate substantially from the body shape. These deformations are conditioned on normals estimated from an input image of the target garment. They are learned from synthetic mesh data featuring loose clothing, where the deformations are taken to be those required to fit individual ISP 3D surfaces to the ground-truth 3D meshes.", + "bbox": [ + 496, + 489, + 892, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Given the trained deformation model, we designed a two-stage fitting process to recover the 3D garment from in-the-wild images. First, the parameters of the pre-trained deformation model are optimized to produce a shape that minimizes the distance between garment outlines and segmented garment regions, the differences between garment normals and the normals estimated in the images by off-the-shelf-algorithms [46, 53], and a physics-based loss to promote physical plausibility of the results. Then, fine local details are recovered by directly optimizing the vertex positions of the reconstructed mesh with the same loss. Our fitting process does not require external 3D annotations, other", + "bbox": [ + 496, + 719, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1586", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "than the estimated normals of the target garment.", + "bbox": [ + 76, + 90, + 400, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We demonstrate that our method can recover garments that go from tight- to loose-fitting and outperforms existing approaches [9, 11, 20, 32, 39, 53] in terms of reconstruction accuracy. Furthermore, our reconstructed meshes are directly usable for virtual try-on or animation, unlike [37, 53]. Our implementation and model weights are available at https://github.com/liren2515/GarmentRecovery.", + "bbox": [ + 75, + 106, + 467, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 239, + 218, + 253 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Before the advent of Deep Learning, garment shape recovery from images depended mostly on user defined outlines and shape-from-shading techniques [59]. Since then, data-driven techniques have become dominant.", + "bbox": [ + 75, + 263, + 467, + 325 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Tight-Fitting Clothing. The majority of methods developed in recent years focus on clothing that clings relatively closely to the body. These can be classified into two main categories.", + "bbox": [ + 75, + 340, + 467, + 400 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the first category, are methods that model garments as surfaces that are distinct from the body surface and interact with it. DeepGarment [10], MGN [5], and BCNet [20] train neural networks on synthetic images to predict the vertex positions of specific mesh templates. [8] and [34] leverage normal estimation to optimize the vertex position of the template and recover wrinkle details. Such methods are inherently limited in the range of shapes they can handle, and those being trained on synthetic data can easily fail when facing real images. To overcome these limitations, SMPLicit [9], DIG [31], and ClothWild [39] leverage Signed Distance Functions (SDF) to recover a wide array of garment meshes from RGB images and the corresponding segmentation masks. However, to represent non-watertight garment surfaces using an SDF, one has to wrap around them a watertight surface with a minimum thickness, which reduces accuracy. This can be addressed by using Unsigned Distance Functions (UDFs) instead [11, 15] but creates robustness issues: if the UDF is even slightly inaccurate, the value of the surface is never exactly zero and holes can appear in the reconstructed models. In our experience, the Implicit Sewing Patterns (ISP) model of [32] effectively addresses the issues of generality, accuracy, and robustness. The garments consist of flat 2D panels whose boundary is defined by a 2D SDF. To each panel is associated a 3D surface parameterized by the 2D panel coordinates. Hence, different articles of clothing are represented in a standardized way, which allows the recovery of various garments from single images. This is why we choose it as the basis for our approach.", + "bbox": [ + 75, + 401, + 467, + 852 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the second category, are the many methods that represent body and garment using a single model. For example, in [19, 57] a volumetric regression network yields a voxel", + "bbox": [ + 75, + 854, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "representation of 3D clothed humans given a single image. Other works [3, 16, 45, 46] employ a pixel-aligned implicit function that defines 3D occupancy fields or signed distance fields for clothed humans. In [1, 2], displacement vectors or UV maps are used to represent deviations from a SMPL parametric body model [36]. Similarly, in [17, 18, 52, 58] parametric body models are combined with implicit representations to achieve robustness to pose changes. While effective, all these methods suffer from significant limitations, because they cannot separate the surface of the garment from that of the body, and they are at a disadvantage when it comes to modeling loose garments whose motion can be relatively independent from the body.", + "bbox": [ + 496, + 90, + 890, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Loose-Fitting Clothing. There is a more limited number of methods designed to handle free-flowing garments. Some recent works [55, 60, 61] rely on complex physics simulation steps or feature line estimation to align the surface reconstruction with the input image. However, their dependence on garment templates limits their generality, in the same way it did for other template-based methods discussed above. Point-based methods that can reconstruct generic clothes have been proposed [50, 56] to overcome this. Unfortunately, point clouds are not straightforward to integrate into downstream applications. As a result, the method of [50] resorts to modified Poisson Surface Reconstruction (PSR) to create a garment surface from the point cloud, which can result in incorrect geometry. Another point-based representation is introduced in [37]. While being successful at modeling and animating humans wearing loose garments, it also relies on PSR to infer the mesh from the point cloud, yielding a single mesh that represents body and garment jointly. Furthermore, [37] does not explore how this representation can be fitted to images.", + "bbox": [ + 496, + 304, + 890, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ECON [53] is a method specifically designed for clothed human recovery from images. By leveraging techniques such as normal integration and shape completion, it achieves visually appealing results for individuals wearing loose clothing. However, as [37], ECON produces a single watertight mesh that tightly binds the body and garment together, precluding easy use for applications such as cloth simulation and re-animation.", + "bbox": [ + 496, + 607, + 890, + 727 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 739, + 589, + 753 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given an image of a clothed person and a body model extracted from it using existing techniques, our goal is to recover accurate 3D models of the garments matching the image. To this end, we add to the Implicit Sewing Pattern (ISP) garment model [32], which provides us with a shape prior for garment in its rest state, a deformation model that allows us to recover its potentially large deformations, as illustrated by Fig. 2. As a result, whereas the original ISP, like most current clothes-recovery algorithms, is limited to", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1587", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/751dad8a40f9c353c2654b0b7e8d6d1d462ccdd694d384c348e0fc10f9a2abc4.jpg", + "image_caption": [ + "Figure 2. Framework. Given an image, (1) we first estimate the normal map of the target garment and the SMPL body parameters $(\\beta, \\theta)$ , which are used to compute the body part segmentation and position maps. (2) The maximum coverage garment shape $\\bar{\\mathcal{M}}$ is then skinned to closely fit to the body, yielding $\\mathcal{M}$ . Leveraging (3) pixel-aligned image features, our deformation model (4) predicts occupancy and position maps to correct $\\mathcal{M}$ for large deformations. (5) The 3D garment mesh is recovered using ISP and further refined." + ], + "image_footnote": [], + "bbox": [ + 86, + 89, + 883, + 375 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tight-fitting clothing, our approach can handle both tight- and loose-fitting garments, such as skirts and open jackets.", + "bbox": [ + 75, + 452, + 468, + 482 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first describe briefly the ISP model upon which we build our approach. We then introduce the deformation model that underpins the main contribution of this paper, going from tight fitting clothes to loosely fitting and free flowing ones. Finally, we present our approach to fitting this model to real-world images.", + "bbox": [ + 75, + 482, + 470, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. ISP Garment Model", + "text_level": 1, + "bbox": [ + 76, + 580, + 269, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ISP is a garment model inspired by the sewing patterns that fashion designers use to represent clothes. A sewing pattern is made of several 2D panels along with information about how to stitch them into a complete garment. ISP implicitly models patterns using a 2D signed distance field and a 2D label field.", + "bbox": [ + 75, + 604, + 468, + 695 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Formalization. Given the latent code $\\mathbf{z}$ of a garment and a point $\\mathbf{u}$ in the 2D UV space $\\Omega = [-1,1]^2$ of a 2D panel of that garment, ISP outputs the signed distance $s$ to the panel boundary and a label $l$ as", + "bbox": [ + 75, + 713, + 470, + 773 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n(s, l) = \\mathcal {I} _ {\\Theta} (\\mathbf {u}, \\mathbf {z}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 784, + 468, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{I}_{\\Theta}$ is a fully connected network. The zero crossing of the SDF defines the shape of the panel, with $s < 0$ indicating that $\\mathbf{u}$ is within the panel and $s > 0$ indicating that $\\mathbf{u}$ is outside the panel. The stitch information is encoded in $l$ , where panel boundaries with the same label should be stitched together. To transform the 2D sewing patterns into", + "bbox": [ + 75, + 809, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D surfaces, a UV parameterization function $\\mathcal{A}_{\\Phi}$ is learned to perform the 2D-to-3D mapping", + "bbox": [ + 498, + 452, + 892, + 483 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} = \\mathcal {A} _ {\\Phi} (\\mathbf {u}, \\mathbf {z}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 638, + 494, + 892, + 511 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{X} \\in \\mathbb{R}^3$ represents the 3D position of $\\mathbf{u}$ . Essentially, ISP registers each garment onto a unified 2D space $\\Omega$ , and represents it using UV maps that record 2D-to-3D mapping, as shown in Fig. 3(b). Given the paired 2D sewing patterns and their 3D meshes, the pattern parameterization network $\\mathcal{I}_{\\Theta}$ and the UV parameterization network $\\mathcal{A}_{\\Phi}$ are trained by minimizing the losses", + "bbox": [ + 496, + 521, + 892, + 628 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathcal {I}} = L _ {S D F} + L _ {C E} + \\left\\| \\mathbf {z} \\right\\| _ {2} ^ {2}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 638, + 890, + 656 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathcal {A}} = L _ {M S E} + L _ {\\text {c o n s i s t}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 657, + 890, + 674 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "respectively. $L_{SDF}$ is the mean absolute error for the predicted SDF value $s$ , $L_{CE}$ is the cross-entropy loss for the predicted label $l$ , $L_{MSE}$ is the mean squared error of the predicted 3D position $\\mathbf{X}$ , and $L_{consist}$ is the loss to reduce the gap between the front and back panels. More details can be found in [32].", + "bbox": [ + 496, + 686, + 890, + 776 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Training. Training ISP requires the 2D sewing patterns of 3D garments in a rest state, that is, draped on a T-posed neutral body. However, patterns are absent in most large scale garment datasets, including CLOTH3D [4]. We use the garment flattening algorithm of [42] to generate the required 2D patterns. The garment mesh is first divided into several pieces given predefined cutting rules. These pieces", + "bbox": [ + 496, + 794, + 892, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1588", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1a9d6399310b40c3c791d39ec14b692fff3f23438cbe26cf5d379baa695f8b55.jpg", + "image_caption": [ + "Figure 3. Cutting and flattening. (a) The front (top) and back (bottom) surfaces after cutting. (b) The flattened panels and UV maps generated by ISP for (a). (c) The maximum-coverage UV maps and its represented 3D shape." + ], + "image_footnote": [], + "bbox": [ + 122, + 85, + 349, + 250 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/563b4d353c8fcbc230500c43fdf75b433791452abcc1142940fe5cb335915269.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 352, + 85, + 607, + 250 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2baa67f8e87409764137dc0da8127f966c221578180e84921cdc99a86402806c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 609, + 85, + 851, + 250 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "are then unfolded into 2D panels by minimizing an as-rigid-as-possible [33] energy, ensuring local area preservation between the 3D and 2D parameterizations. While doing this, we constrain the boundary vertices at places such as waist and sleeves to have a constant value along a specific axis. This enhances pattern consistency across the dataset. Fig. 3 illustrates this and shows the panels generated for a shirt. We provide more details in the supplementary material.", + "bbox": [ + 75, + 313, + 468, + 434 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We generate a front and a back panels as the sewing pattern for each garment in our dataset (shirt, skirt, and trousers). Once ISP has been trained on these, we compute the maximum-coverage UV maps $\\bar{\\mathcal{M}}$ over the UV maps of each garment for each category, as shown in Fig. 3(c). The values of $\\bar{\\mathcal{M}}$ are taken to be", + "bbox": [ + 75, + 434, + 468, + 523 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\mathcal {M}} [ u, v ] = \\frac {\\sum_ {i} \\mathbb {1} _ {s _ {\\mathbf {u}} ^ {i} \\leq 0} \\cdot m _ {i} [ u , v ]}{\\sum_ {i} \\mathbb {1} _ {s _ {\\mathbf {u}} ^ {i} \\leq 0}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 526, + 468, + 563 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $m_{i}$ is the UV maps of garment $i$ , $\\mathbb{1}$ is the indicator function, $[\\cdot ,\\cdot ]$ denotes array addressing, and $s_{\\mathbf{u}}^{i}$ is the SDF value of ISP at $\\mathbf{u} = (u,v)$ . The maximum coverage map $\\bar{\\mathcal{M}}$ encompasses information from all the patterns in the dataset. It represents the smallest possible map that covers all garments in a category, with the 3D position of each $uv$ -point $\\mathbf{u}$ being the average of all garments that include $\\mathbf{u}$ . We use it as a prototype for a garment category and to compute an initial guess of the deformed garment given the body pose, as discussed below.", + "bbox": [ + 75, + 566, + 468, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "**Skinning.** As in many prior work [11, 31, 32], we use SMPL [36] to parameterize the body in terms of shape and pose parameters $(\\beta, \\theta)$ , and its extended skinning procedure for the 3D volume around the body to initially deform the 3D shape represented as $\\bar{\\mathcal{M}}$ . More specifically, given a 2D point $\\mathbf{u} = (u, v)$ in UV space $\\Omega$ , we get its actual 3D position as $\\bar{\\mathbf{X}}_{\\mathbf{u}} = \\bar{\\mathcal{M}}[u, v]$ . We then deform it by computing", + "bbox": [ + 75, + 734, + 468, + 842 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {\\mathbf {u}} = W \\left(\\mathbf {X} _ {(\\beta , \\theta)}, \\beta , \\theta , w (\\bar {\\mathbf {X}} _ {\\mathbf {u}}) \\mathcal {W}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 859, + 468, + 878 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {(\\beta , \\theta)} = \\bar {\\mathbf {X}} _ {\\mathbf {u}} + w (\\bar {\\mathbf {X}} _ {\\mathbf {u}}) B _ {\\beta} + w (\\bar {\\mathbf {X}} _ {\\mathbf {u}}) B _ {\\theta},\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 880, + 411, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $W(\\cdot)$ is the SMPL skinning function with skinning weights $\\mathcal{W} \\in \\mathbb{R}^{N_B \\times 24}$ , with $N_B$ being the number of vertices of the SMPL body mesh, and $B_\\beta \\in \\mathbb{R}^{N_B \\times 3}$ and $B_\\theta \\in \\mathbb{R}^{N_B \\times 3}$ are the shape and pose displacements of SMPL, respectively. The diffused weights $w(\\cdot) \\in \\mathbb{R}^{N_B}$ are computed by a neural network, which generalizes the SMPL skinning to any point in 3D space. By repeating this for all points in a panel, we obtain the vertex position map $\\mathcal{M}[u,v] = \\mathbf{X}_{\\mathbf{u}}$ .", + "bbox": [ + 500, + 313, + 890, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Modeling Large Deformations", + "text_level": 1, + "bbox": [ + 500, + 458, + 767, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 2, the ISP approach described above generates a garment prototype $\\mathcal{M}$ that closely fits the underlying body. To a point $\\mathbf{u} = (u,v)$ that belongs to a given garment panel, it associates the vertex position $\\mathbf{X}_{\\mathbf{u}} = \\mathcal{M}[u,v]$ of Eq. 6, which is usually relatively close to the body. To model loose clothing, we now need to compute a potentially larger displacement $\\Delta \\mathbf{X}_{\\mathbf{u}}$ to be added to $\\mathbf{X}_{\\mathbf{u}}$ .", + "bbox": [ + 496, + 481, + 890, + 587 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To evaluate $\\Delta \\mathbf{X}_{\\mathbf{u}}$ we train a network $\\mathcal{D}$ that consists of an MLP to estimate the occupancy value and the corrective displacement value, and two CNNs to extract image features $F_{n}$ and $F_{b}$ from the image of normals and the segmentation and vertex position images of the SMPL body, respectively. Fig. 2 depicts this architecture. We obtain the pixel-aligned image feature for $\\mathbf{u}$ by computing", + "bbox": [ + 496, + 587, + 890, + 693 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF \\left(\\mathbf {x} _ {\\mathbf {u}}\\right) = F _ {n} \\oplus F _ {b} \\left(\\mathbf {x} _ {\\mathbf {u}}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 611, + 704, + 890, + 720 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x _ {u}} = P (\\mathbf {X _ {u}}) ,\n$$\n", + "text_format": "latex", + "bbox": [ + 637, + 723, + 740, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $P(\\cdot)$ denotes the projection into the image and $\\oplus$ concatenation. The MLP takes as input $\\mathbf{u}$ , its 3D position and image features to predict the occupancy $O_{\\mathbf{u}} \\in \\{0,1\\}$ and the corrective displacement $\\Delta \\mathbf{X}_{\\mathbf{u}}$ . By assembling the results of each point in the UV space, we obtain the final occupancy maps $\\mathcal{O}$ and the vertex position maps $\\hat{\\mathcal{M}}$ , where $\\hat{\\mathcal{M}}[u,v] = \\mathbf{X}_{\\mathbf{u}} + \\Delta \\mathbf{X}_{\\mathbf{u}}$ . In essence, $\\mathcal{O}$ is the binarized SDF of the garment 2D panels, which implicitly defines the garment shape and geometry in the rest state, while $\\hat{\\mathcal{M}}$ encodes the deformed state for that garment. The 3D garment mesh", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1589", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "can be recovered from these 2D maps with ISP, as discussed below.", + "bbox": [ + 75, + 90, + 468, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use a single network $\\mathcal{D}$ for the front and the back panels, and encode the points $\\mathbf{u}$ of the front as $[u,v,+1]$ and those of the back as $[u,v,-1]$ . We learn the parameters of $\\mathcal{D}$ by minimizing", + "bbox": [ + 75, + 121, + 467, + 181 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL = \\sum_ {\\mathbf {u} \\in \\Omega} | | \\mathbf {X} _ {\\mathbf {u}} + \\Delta \\mathbf {X} _ {\\mathbf {u}} - \\tilde {\\mathbf {X}} _ {\\mathbf {u}} | | _ {2} + \\lambda \\sum_ {\\mathbf {u} \\in \\Omega} B E (O _ {\\mathbf {u}}, \\tilde {O} _ {\\mathbf {u}}), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 191, + 468, + 223 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tilde{\\mathbf{X}}_{\\mathbf{u}}$ and $\\tilde{O}_{\\mathbf{u}}$ are the ground truth vertex positions and occupancy values, $BE(\\cdot)$ is the binary cross entropy, and $\\lambda$ is a weighting constant.", + "bbox": [ + 75, + 234, + 468, + 281 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "From 2D maps to 3D mesh. The occupancy value $O_{\\mathbf{u}}$ indicates whether $\\mathbf{u}$ falls within the panels of the garment. To convert the generated 2D occupancy maps $\\mathcal{O}$ into a 3D garment mesh in its rest shape, we need to recover an ISP latent code $\\mathbf{z}$ as defined in Eq. 1. To this end, we find the vector $\\mathbf{z}$ such that the corresponding SDF of ISP best matches the produced occupancy maps by minimizing", + "bbox": [ + 75, + 297, + 468, + 397 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} ^ {*} = \\underset {\\mathbf {z}} {\\operatorname {a r g m i n}} \\sum_ {\\mathbf {u} \\in \\Omega_ {-}} R \\left(s _ {\\mathbf {u}} (\\mathbf {z})\\right) + \\sum_ {\\mathbf {u} \\in \\Omega_ {+}} R \\left(- s _ {\\mathbf {u}} (\\mathbf {z})\\right) + \\lambda_ {\\mathbf {z}} \\| \\mathbf {z} \\| _ {2}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 83, + 404, + 468, + 433 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Omega_{-} = \\{\\mathbf{u}|O_{\\mathbf{u}} = 1,\\mathbf{u}\\in \\Omega \\}$ , $\\Omega_{+} = \\{\\mathbf{u}|O_{\\mathbf{u}} = 0,\\mathbf{u}\\in \\Omega \\}$ , $R(\\cdot)$ is the ReLU function, $s_{\\mathbf{u}}(\\mathbf{z})$ is the SDF value of $\\mathbf{u}$ computed by ISP, and $\\lambda_{\\mathbf{z}}$ is the weighting constant. With $\\mathbf{z}^*$ , the rest garment mesh is inferred through ISP's meshing and sewing process. The deformed garment mesh can be obtained by simply replacing the vertex position of the recovered mesh with the values stored in $\\hat{\\mathcal{M}}$ . This yields 3D garment meshes in both the rest and deformed states as shown in the top-right of Fig. 2, which are required for the application such as cloth simulation and can be used for further refinement as discussed below.", + "bbox": [ + 75, + 441, + 468, + 608 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Fitting the Models to Images", + "text_level": 1, + "bbox": [ + 76, + 618, + 333, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For practical reasons, the range of garment materials, external forces, and body motions present in the training data is limited. As a result, given in-the-wild images as input, the trained model can produce inaccurate results as shown in Fig. 4(b). To remedy this and to leverage the deformation prior that the network $\\mathcal{D}$ captures, we refine the result by minimizing a loss function with respect to the pretrained deformation parameters of $\\mathcal{D}$ as in [13, 21, 51]. $L$ is designed to promote a good match between the garment mesh and image observations. We take it to be", + "bbox": [ + 75, + 641, + 468, + 791 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL = \\lambda_ {C} L _ {C D} + \\lambda_ {n} L _ {\\text {n o r m a l}} + \\lambda_ {p} L _ {\\text {p h y s i c s}}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 804, + 467, + 819 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {C D} = d \\left(x _ {c} ^ {f}, x _ {I _ {n}}\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 823, + 467, + 840 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {n o r m a l}} = \\sum_ {i \\in f} 1 - \\cos \\left(n ^ {i}, I _ {n} \\left(x _ {c} ^ {i}\\right)\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 844, + 467, + 875 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {p h y s i c s}} = L _ {\\text {s t r a i n}} + L _ {\\text {b e n d}} + L _ {\\text {g r a v i t y}} + L _ {\\text {c o l}}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 881, + 467, + 897 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5346d41e7e76a011f27ff50cb66777b03bbedfce4702431c284fc30e2c64c5b6.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 545, + 88, + 694, + 200 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a5f9a37d13f87c2eb6001c2531b4222c42f3a93e1ec14548fdbdbe3620642b56.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 718, + 107, + 833, + 190 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/efd48e2fc2c792631ebcbf5c5ad46423b275e817bf956c97732261ddab285c74.jpg", + "image_caption": [ + "(c)", + "Figure 4. Fitting results. Given (a) the normal estimation of an in-the-wild image, (b) is the inference result with ISP recovered geometry. (c) is obtained by optimizing the parameters of the pretrained deformation model. Further refinement of the mesh vertex positions yields (d)." + ], + "image_footnote": [], + "bbox": [ + 547, + 220, + 691, + 306 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/08b64de78b05e863885fd64134115ba86f91f72b6ff20c8a404cc02ec95ec968.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 700, + 220, + 844, + 305 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $x_{c}^{f}$ is the 2D projection of the centers of visible faces $f$ after mesh rasterization, $x_{I_n}$ denotes the coordinates of foreground pixels, $n^i$ is the normal of face $i$ , $I_{n}(x_{c}^{i})$ is the normal image values at $x_{c}^{i}$ , and $\\lambda_{C}$ , $\\lambda_{n}$ and $\\lambda_{p}$ are the balancing scalars. $d(\\cdot)$ and $\\cos (\\cdot)$ are the functions measuring the 2D Chamfer Distance and the cosine similarity, respectively. $L_{\\mathrm{physics}}$ is a physics-based loss derived from [40, 48], which computes the membrane strain energy $L_{strain}$ caused by the deformation, the bending energy $L_{bend}$ resulting from the folding of adjacent faces, the gravitational potential energy $L_{gravity}$ and the penalty for body-garment collision $L_{col}$ . Minimizing $L_{CD}$ induces an external force of stretching or compression on the garment mesh to align its 2D projection with the given image, while minimizing $L_{\\mathrm{physics}}$ ensures that the mesh exhibits physically plausible deformation adhering to the shape constraints of the rest-state mesh recovered by ISP.", + "bbox": [ + 498, + 419, + 890, + 676 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Minimizing the loss $L$ with respect to the deformation parameters yields a mesh whose overall shape matches the input image, as illustrated in Fig. 4(c). However, since neural networks tend to learn low-frequency functions [43], the result might be too smooth. To recover fine surface details, we perform a refinement step by minimizing $L$ directly with respect to the coordinates of the garment mesh vertices. This generates realistic local details, such as wrinkles on the surface, as shown in Fig. 4(d).", + "bbox": [ + 498, + 678, + 890, + 813 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 829, + 630, + 845 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In short, our method begins by inferring the shape and the deformation of the garment in terms of the occupancy and position maps. Leveraging the shape prior of ISP, we then", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1590", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "recover the garment geometry from the occupancy maps and deform it using the position maps. Next, we refine the initial deformed mesh to better align with image observations by fine-tuning the pre-trained network $\\mathcal{D}$ that captures the deformation prior. Finally, we recover fine details through vertex-level optimization of the garment mesh. In this section, we demonstrate the effectiveness of this process and compare it to that of other state-of-the-art methods.", + "bbox": [ + 75, + 90, + 472, + 212 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 223, + 294, + 239 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following the implementation of [32], both the pattern parameterization model $\\mathcal{I}_{\\Theta}$ and the UV parameterization model $\\mathcal{A}_{\\Phi}$ of ISP have two separate MLPs for the front and back panels. Each MLP has 7 layers with Softplus activations. The dimension of latent code $\\mathbf{z}$ is 32. The skinning weight model $w$ is a 9-layer MLP with leaky ReLU activations, whose output is normalized by a final Softmax layer. $\\mathcal{I}_{\\Theta}$ and $\\mathcal{A}_{\\Phi}$ are trained jointly for 9000 iterations with a batch size of 50. $w$ is trained with the same parameters as [32]. For the image feature extraction, we use two separate ConvNeXt [35] networks to extract multi-scale garment and body features of sizes 96, 192, 384, 768, which are concatenated as the final features. The point UV coordinates $\\mathbf{u}$ , 3D position $\\mathbf{X_u}$ , and image features $F(\\mathbf{x_u})$ are projected separately to 384 dimensions by three linear layers, which are then concatenated as the input of the MLP of the deformation model. The MLP of the deformation model has 10 layers with a skip connection from the input layer to the middle, and uses Gaussian functions as the activation layer following [44]. The CNNs and MLP of the deformation model are trained jointly for 40 epochs with the Adam optimizer [23] and a learning rate of $10^{-4}$ . For real images, we use [53] and [12] to obtain their normal and SMPL body parameter estimations, respectively. The garment segmentation masks are generated by leveraging the segmentation of SAM [24] and the semantic labels of [29].", + "bbox": [ + 75, + 246, + 472, + 640 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Dataset, Evaluation Metrics, and Baseline", + "text_level": 1, + "bbox": [ + 76, + 650, + 433, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our models are trained on CLOTH3D [4], which is a synthetic dataset with motion sequences of 3D clothed human. It contains garment in the rest state, and deformed states caused by the motion of underlying body. For each garment, it has a single simulated sequence up to 10 seconds. It covers a large variety of garment in different shapes, types and topologies. For the training of ISP, we randomly select 400 shirts, 200 skirts and 200 pairs of trousers, and generate their sewing patterns by the method described in Sec. 3.1. The deformation model is trained on the corresponding simulated sequences. For each frame, we render 11 normal images for the garment mesh with random rotations around the Y-axis, which produces 40K, 40K and 20K training images for shirt, skirt and trousers respectively. During training, we augment the data with image flipping and rotation.", + "bbox": [ + 75, + 672, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate the garment reconstruction quality, we use the Chamfer Distance (CD) between the ground truth and the recovered garment mesh, and Intersection over Union (IoU) between the ground truth mask and the rendered mask of reconstructed garment mesh.", + "bbox": [ + 496, + 90, + 890, + 165 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our method against state-of-the-art methods BCNet [20], SMPLicit [9], ClothWild [39], DrapeNet [11] and ISP [32]. SMPLicit, DrapeNet and ISP use the garment segmentation mask for reconstruction, while BCNet and ClothWild take the RGB images as input.", + "bbox": [ + 496, + 167, + 893, + 243 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparison with State-of-the-Art Methods", + "text_level": 1, + "bbox": [ + 498, + 250, + 867, + 266 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/61d265aa5c602ed56aabbf35a114ffeac37626f97ab442484dec454ca31efaa7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CD (×103) ↓SkirtShirtTrousersIoU ↑SkirtShirtTrousers
SMPLicit8.053.490.81SMPLicit0.6570.5270.782
DrapeNetn/a1.540.84DrapeNetn/a0.7150.845
ISP7.262.020.91ISP0.7090.6650.781
Ours2.870.440.28Ours0.9400.9390.946
Ours-GT1.670.310.23Ours-GT0.9530.9420.954
", + "bbox": [ + 495, + 284, + 893, + 371 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparisons. Our method outperforms SMPLicit, DrapeNet, and ISP in terms of CD and IoU on all three garment categories, as shown in the second-to-last row of both tables (Ours). These results were obtained using normals estimated from the images using [53]. In the last row (Ours-GT), we provide the results we obtained using ground-truth normals instead.", + "bbox": [ + 496, + 375, + 893, + 458 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Results. Due to the absence of publicly available real dataset for evaluating garment reconstruction, we utilize synthetic data consisting of 30 unseen shirts, 30 unseen skirts, and 30 pairs of unseen trousers from CLOTH3D for quantitative evaluation. We compare our method with SMPLicit, DrapeNet, and ISP, which recover garment meshes from segmentation masks. The segmentation masks required by the baselines are rendered from the ground truth mesh data, while the normal images required by our method are estimated with [53]. As shown in Tab. 1, our approach significantly outperforms the baselines in terms of Chamfer Distance (CD) and Intersection over Union (IoU) across all garment categories.", + "bbox": [ + 496, + 489, + 890, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For comparison purposes, we re-ran our algorithm using the ground-truth normals and report the results in the last row of Tab. 1. As expected, this leads to an improvement in reconstruction accuracy, but the increase is only modest. This highlights the robustness of our approach to the slight inaccuracies that can be expected from a normal-estimation algorithm.", + "bbox": [ + 496, + 686, + 893, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Results. Fig. 5 shows the qualitative comparison for the results reconstructed from in-the-wild images. Since BCNet is trained only on synthetic RGB data, it is not able to predict accurate body and garment results. Both SMPLicit and ClothWild generate unrealistic watertight meshes that do not accurately represent real garments.", + "bbox": [ + 496, + 809, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1591", + "bbox": [ + 483, + 945, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2f0fb335fe2dea224549700b3f523ff6976a06bda21bbe3ddd9544024b086391.jpg", + "image_caption": [ + "Figure 5. Comparison against SOTA methods. From left to right, we show the input image and the 3D garment meshes recovered by our method and SOTA methods: BCNet, SMPLcit, ClothWild, DrapeNet, ISP. (Since skirt is unavailable for DrapeNet, a random pair of trousers is put on its result in the second row.)" + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 885, + 424 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/705d6698165a902bddbd4698090b63240d9b8d195faea797b5817ab50d080a5d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CD (×103) ↓GTEst.
(-D*,-v*)2.043.58
(-D*,+v*)1.803.39
(+D*,-v*)1.752.93
(+D*,+v*)1.672.87
", + "bbox": [ + 81, + 492, + 267, + 580 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8068dd4d0314ac8d413b66ad4748e66ebad32ed8b160004d6f68c3e95f2ca9c6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
IoU ↑GTEst.
(-D*,-v*)0.8480.751
(-D*,+v*)0.9470.904
(+D*,-v*)0.9410.923
(+D*,+v*)0.9530.940
", + "bbox": [ + 279, + 492, + 465, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Ablation study. GT and Est. mean using ground-truth and estimated normal images as input, respectively. $+ / - {\\mathcal{D}}^{ * }$ denote with/without the finetuning of the deformation model. $+ / - {v}^{ * }$ denote with/without the optimization of vertex positions.", + "bbox": [ + 75, + 603, + 470, + 660 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "While DrapeNet and ISP are able to recover open surfaces for garments, their results remain closely adhered to the body, similar to SMPLlicit and ClothWild. In contrast, our method can faithfully recover garment mesh from input images, as clearly observed in the depiction of the open jacket and the flowing long skirt that stand away from the body.", + "bbox": [ + 75, + 686, + 468, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 786, + 230, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2 presents the evaluation results of our method with and without fitting on the test set of skirts. The results of row 1, 3 and 4 indicate that optimizing the parameters of the pretrained deformation network $(+D^{*}, - v^{*})$ improves the quality of the raw inference results, and optimizing the vertex positions $(+D^{*}, + v^{*})$ further enhances the outcome,", + "bbox": [ + 75, + 809, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7478b15f133d96704f58df0b2f2bcbb1666c21e39f9fa5c8d7d6457986c66f81.jpg", + "image_caption": [ + "Figure 6. Fitting strategy comparison between (a) our fitting method and (b) only vertex optimization." + ], + "image_footnote": [], + "bbox": [ + 504, + 492, + 638, + 623 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/01769d869b91a12282eb349c458d95ff8d6aeb1fbee41764ffecfa781005fd17.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 645, + 500, + 746, + 623 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ba54e37c27845e8ac3e46b1829ec92bcb5a36c730677597d999e9ceeb686e7fc.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 769, + 501, + 867, + 622 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "yielding the lowest CD and highest IoU.", + "bbox": [ + 498, + 703, + 767, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "However, as demonstrated in the second row, directly optimizing the vertex positions on the raw inference results without optimizing the deformation model $(-\\mathcal{D}^{*}, + v^{*})$ is not as effective. The deformation model encapsulates a continuous deformation field. When its weights are optimized based on partial observations, the entire field undergoes modification, thereby influencing all mesh vertices globally. On the contrary, a direct vertex optimization with partial observations predominantly affects the mesh vertices locally. While this can capture localized details like wrinkles, it struggles to resolve discrepancies in the overall shape. Fig. 6 displays the results of our methods and an ablation with", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1592", + "bbox": [ + 483, + 944, + 514, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a84fb222c2ae35f5f1b04f7b7c0ef7c7e718caba24661b33ebab34379c1b9cfa.jpg", + "image_caption": [ + "Figure 7. Recovering from in-the-wild images. Our method is able to recover realistic meshes for garments with diverse shapes and deformations." + ], + "image_footnote": [], + "bbox": [ + 76, + 89, + 893, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "vertex optimization only. Unsurprisingly, the latter fails at recovering the correct shape and produces implausible deformation on the mesh surface.", + "bbox": [ + 75, + 705, + 470, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. More Results", + "text_level": 1, + "bbox": [ + 76, + 791, + 215, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 7 shows a collection of results reconstructed by our method from in-the-wild images. Our method can produce realistic 3D meshes with fine details across a wide range of garment types, from tight-fitting attire to more relaxed and flowing outfits.", + "bbox": [ + 75, + 824, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 704, + 619, + 719 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have presented a novel approach to recovering realistic 3D garment meshes from in-the-wild images featuring loose fitting clothing. It relies on a fitting process that imposes shape and deformation priors learned on synthetic data to accurately capture garment shape and deformations. In future work, we will extend our approach to modeling deformations over time from video sequences while enforcing temporal consistency of the reconstructions.", + "bbox": [ + 498, + 729, + 890, + 849 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This project was supported in part by the Swiss National Science Foundation.", + "bbox": [ + 498, + 869, + 890, + 897 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1593", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] T. Alldieck, M. Magnor, B. Bhatnagar, C. Theobalt, and G. Pons-Moll. Learning to reconstruct people in clothing from a single RGB camera. In Conference on Computer Vision and Pattern Recognition, pages 1175-1186, 2019. 2", + "[2] T. Alldieck, G. Pons-Moll, C. Theobalt, and M. Magnor. Tex2shape: Detailed full human body geometry from a single image. In International Conference on Computer Vision, pages 2293-2303, 2019. 2", + "[3] T. Alldieck, M. Zanfir, and C. Sminchisescu. Photorealistic monocular 3d reconstruction of humans wearing clothing. In Conference on Computer Vision and Pattern Recognition, pages 1506-1515, 2022. 2", + "[4] H. Bertiche, M. Madadi, and S. Escalera. CLOTH3D: Clothed 3D Humans. In European Conference on Computer Vision, pages 344-359, 2020. 3, 6, 1", + "[5] B. L. Bhatnagar, G. Tiwari, C. Theobalt, and G. Pons-Moll. Multi-Garment Net: Learning to Dress 3D People from Images. In International Conference on Computer Vision, 2019. 1, 2", + "[6] F. Bogo, A. Kanazawa, C. Lassner, P. Gehler, J. Romero, and M. J. Black. Keep It SMPL: Automatic Estimation of 3D Human Pose and Shape from a Single Image. In European Conference on Computer Vision, 2016. 1", + "[7] R. Brouet, A. Sheffer, L. Boissieux, and M.P. Cani. Design Preserving Garment Transfer. ACM Transactions on Graphics, 31(4):361-3611, 2012. 1", + "[8] A. Casado-Elvira, M. Trinidad, and D. Casas. PERGAMO: Personalized 3d garments from monocular video. In Computer Graphics Forum, pages 293–304, 2022. 2", + "[9] E. Corona, A. Pumarola, G. Alenya, G. Pons-Moll, and F. Moreno-Noguer. SMPlicit: Topology-Aware Generative Model for Clothed People. In Conference on Computer Vision and Pattern Recognition, 2021. 1, 2, 6", + "[10] R. Danerek, E. Dibra, C. Öztireli, R. Ziegler, and M. Gross. Deepgarment : 3D Garment Shape Estimation from a Single Image. Eurographics, 2017. 2", + "[11] Luca DeLuigi, Ren Li, Benoit Guillard, Mathieu Salzmann, and Pascal Fua. DrapeNet: Generating Garments and Draping them with Self-Supervision. In Conference on Computer Vision and Pattern Recognition, 2023. 1, 2, 4, 6", + "[12] Y. Feng, V. Choutas, T. Bolkart, D. Tzionas, and M. J. Black. Collaborative Regression of Expressive Bodies using Moderation. In International Conference on Computer Vision, 2021. 6", + "[13] M. Gadelha, R. Wang, and S. Maji. Deep manifold prior. In International Conference on Computer Vision, pages 1107-1116, 2021. 5", + "[14] G. Georgakis, R. Li and S. Karanam, T. Chen, J. Košecka, and Z. Wu. Hierarchical kinematic human mesh recovery. In European Conference on Computer Vision, pages 768-784, 2020. 1", + "[15] B. Guillard, F. Stella, and P. Fua. MeshUDF: Fast and Differentiable Meshing of Unsigned Distance Field Networks. In European Conference on Computer Vision, 2022. 2", + "[16] T. He, J. Collomosse, H. Jin, and S. Soatto. Geo-pifu: Geometry and pixel aligned implicit functions for single-view" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "human reconstruction. Advances in Neural Information Processing Systems, 33:9276-9287, 2020. 2", + "[17] T. He, Y. Xu, S. Saito, S. Soatto, and T. Tung. Arch++: Animation-ready clothed human reconstruction revisited. In International Conference on Computer Vision, pages 11046-11056, 2021. 2", + "[18] Z. Huang, Y. Xu, C. Lassner, H. Li, and T. Tung. Arch: Animatable Reconstruction of Clothed Humans. In Conference on Computer Vision and Pattern Recognition, pages 3093-3102, 2020. 2", + "[19] A. Jackson, C. Manafas, and G. Tzimiropoulos. 3d human body reconstruction from a single image via volumetric regression. In European Conference on Computer Vision Workshops, 2018. 2", + "[20] B. Jiang, J. Zhang, Y. Hong, J. Luo, L. Liu, and H. Bao. Bcnet: Learning body and cloth shape from a single image. In European Conference on Computer Vision, 2020. 1, 2, 6", + "[21] H. Joo, N. Neverova, and A. Vedaldi. Exemplar Fine-Tuning for 3D Human Pose Fitting Towards In-the-Wild 3D Human Pose Estimation. In International Conference on 3D Vision, 2020. 1, 5", + "[22] A. Kanazawa, M. J. Black, D. W. Jacobs, and J. Malik. End-To-End Recovery of Human Shape and Pose. In Conference on Computer Vision and Pattern Recognition, 2018. 1", + "[23] D. P. Kingma and J. Ba. Adam: A Method for Stochastic Optimisation. In International Conference on Learning Representations, 2015. 6", + "[24] A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. Berg, W. Lo, P. Dollar, and R. Girshick. Segment anything. In International Conference on Computer Vision, 2023. 6", + "[25] N. Kolotouros, G. Pavlakos, M. J. Black, and K. Daniilidis. Learning to Reconstruct 3D Human Pose and Shape via Model-Fitting in the Loop. In Conference on Computer Vision and Pattern Recognition, 2019. 1", + "[26] C. Lassner, J. Romero, M. Kiefel, F. Bogo, M.J. Black, and P.V. Gehler. Unite the People: Closing the Loop Between 3D and 2D Human Representations. In Conference on Computer Vision and Pattern Recognition, 2017. 1", + "[27] B. Lévy, S. Petitjean, N. Ray, and J. Maillot. Least Squares Conformal Maps for Automatic Texture Atlas Generation. In ACM SIGGRAPH, 2002. 1", + "[28] J. Li, C. Xu, Z. Chen, S. Bian, L. Yang, and C. Lu. Hybrid: A hybrid analytical-neural inverse kinematics solution for 3d human pose and shape estimation. In Conference on Computer Vision and Pattern Recognition, 2021. 1", + "[29] P. Li, Y. Xu, Y. Wei, and Y. Yang. Self-Correction for Human Parsing. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(6):3260-3271, 2020. 6", + "[30] R. Li, M. Zheng, S. Karanam, T. Chen, and Z. Wu. Everybody Is Unique: Towards Unbiased Human Mesh Recovery. In British Machine Vision Conference, 2021. 1", + "[31] R. Li, B. Guillard, E. Remelli, and P. Fua. DIG: Draping Implicit Garment over the Human Body. In Asian Conference on Computer Vision, 2022. 1, 2, 4", + "[32] Ren Li, Benoit Guillard, and Pascal Fua. ISP: Multi-Layered Garment Draping with Implicit Sewing Patterns. In Ad-" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1594", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "vances in Neural Information Processing Systems, 2023. 1, 2, 3, 4, 6", + "[33] L. Liu, L. Zhang, Y. Xu, C. Gotsman, and S. J. Gortler. A Local/Global Approach to Mesh Parameterization. In Proceedings of the Symposium on Geometry Processing, pages 1495-1504, 2008. 4, 1", + "[34] X. Liu, J. Li, and G. Lu. Modeling Realistic Clothing from a Single Image under Normal Guide. IEEE Transactions on Visualization and Computer Graphics, 2023. 2", + "[35] Z. Liu, H. Mao, C. Wu, C. Feichtenhofer, T. Darrell, and S. Xie. A ConvNet for the 2020s. In Conference on Computer Vision and Pattern Recognition, 2022. 6", + "[36] M. Loper, N. Mahmood, J. Romero, G. Pons-Moll, and M.J. Black. SMPL: A Skinned Multi-Person Linear Model. ACM SIGGRAPH Asia, 34(6), 2015. 2, 4", + "[37] Qianli Ma, Jinlong Yang, Michael J. Black, and Siyu Tang. Neural Point-based Shape Modeling of Humans in Challenging Clothing. In International Conference on 3D Vision, 2022. 1, 2", + "[38] G. Moon and K.M. Lee. I2L-MeshNet: Image-to-Lixel Prediction Network for Accurate 3D Human Pose and Mesh Estimation from a Single RGB Image. In European Conference on Computer Vision, 2020. 1", + "[39] G. Moon, H. Nam, T. Shiratori, and K.M. Lee. 3d clothed human reconstruction in the wild. In European Conference on Computer Vision, 2022. 1, 2, 6", + "[40] R. Narain, A. Samii, and J.F. O'brien. Adaptive anisotropic remeshing for cloth simulation. ACM Transactions on Graphics, 2012. 5", + "[41] M. Omran, C. Lassner, G. Pons-Moll, P. Gehler, and B. Schiele. Neural Body Fitting: Unifying Deep Learning and Model-Based Human Pose and Shape Estimation. In International Conference on 3D Vision, 2018. 1", + "[42] N. Pietroni, C. Dumery, R. Falque, M. Liu, T. Vidal-Calleja, and O. Sorkine-Hornung. Computational pattern making from 3D garment models. ACM Transactions on Graphics, 41(4):1-14, 2022. 3, 1", + "[43] N. Rahaman, A. Baratin, D. Arpit, F. Draxler, M. Lin, F. Hamprecht, Y. Bengio, and A. Courville. On the spectral bias of neural networks. In International Conference on Machine Learning, 2019. 5", + "[44] S. Ramasinghe and S. Lucey. Beyond Periodicity: Towards a Unifying Framework for Activations in Coordinate-MLPs. In European Conference on Computer Vision, pages 142–158, 2022. 6", + "[45] S. Saito, Z. Huang, R. Natsume, S. Morishima, A. Kanazawa, and H. Li. PIFu: Pixel-Aligned Implicit Function for High-Resolution Clothed Human Digitization. In International Conference on Computer Vision, 2019. 2", + "[46] S. Saito, T. Simon, J. Saragih, and H. Joo. Pifuuhd: Multi-Level Pixel-Aligned Implicit Function for High-Resolution 3D Human Digitization. In Conference on Computer Vision and Pattern Recognition, pages 84–93, 2020. 1, 2", + "[47] I. Santesteban, N. Thuerey, M. A. Otaduy, and D. Casas. Self-Supervised Collision Handling via Generative 3D Garment Models for Virtual Try-On. In Conference on Computer Vision and Pattern Recognition, 2021. 1" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[48] I. Santesteban, M.A. Otaduy, and D. Casas. SNUG: Self-Supervised Neural Dynamic Garments. In Conference on Computer Vision and Pattern Recognition, 2022. 5", + "[49] O. Sorkine-Hornung and M. Rabinovich. Least-squares rigid motion using svd. Computing, 1(1):1-5, 2017. 1", + "[50] A. Srivastava, C. Pokhariya, S. Jinka, and A. Sharma. xCloth: Extracting Template-free Textured 3D Clothes from a Monocular Image. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2504-2512, 2022. 2", + "[51] D. Ulyanov, A. Vedaldi, and V. Lempitsky. Deep Image Prior. In Conference on Computer Vision and Pattern Recognition, pages 9446-9454, 2018. 5", + "[52] Y. Xiu, J. Yang, D. Tzionas, and M. J Black. Icon: Implicit clothed humans obtained from normals. In Conference on Computer Vision and Pattern Recognition, pages 13286-13296, 2022. 2", + "[53] Y. Xiu, J. Yang, X. Cao, D. Tzionas, and M. J Black. ECON: Explicit Clothed humans Optimized via Normal integration. In Conference on Computer Vision and Pattern Recognition, pages 512-523, 2023. 1, 2, 6, 3, 4", + "[54] F. Yang, R. Li, G. Georgakis, S. Karanam, T. Chen, H. Ling, and Z. Wu. Robust multi-modal 3d patient body modeling. In Conference on Medical Image Computing and Computer Assisted Intervention, pages 86-95, 2020. 1", + "[55] S. Yang, Z. Pan, T. Amert, K. Wang, L. Yu, T. Berg, and M. Lin. Physics-inspired garment recovery from a single-view image. ACM Transactions on Graphics, 37(5):1-14, 2018. 2", + "[56] I. Zakharkin, K. Mazur, A. Grigorev, and V. Lempitsky. Point-based modeling of human clothing. In International Conference on Computer Vision, 2021. 2", + "[57] Z. Zheng, T. Yu, Y. Wei, Q. Dai, and Y. Liu. Deephuman: 3d human reconstruction from a single image. In International Conference on Computer Vision, pages 7739-7749, 2019. 2", + "[58] Z. Zheng, T. Yu, Y. Liu, and Q. Dai. Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(6):3170-3184, 2021. 2", + "[59] B. Zhou, X. Chen, Q. Fu, K. Guo, and P. Tan. Garment modeling from a single image. In Computer Graphics Forum, pages 85–91, 2013. 2", + "[60] H. Zhu, Y. Cao, H. Jin, W. Chen, D. Du, Z. Wang, S. Cui, and X. Han. Deep fashion3d: A dataset and benchmark for 3d garment reconstruction from single images. In European Conference on Computer Vision, pages 512-530, 2020. 2", + "[61] H. Zhu, L. Qiu, Y. Qiu, and X. Han. Registering explicit to implicit: Towards high-fidelity garment mesh reconstruction from single images. In Conference on Computer Vision and Pattern Recognition, pages 3845-3854, 2022. 2" + ], + "bbox": [ + 501, + 92, + 890, + 786 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1595", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Garment Recovery with Shape and Deformation Priors/83e37ee5-1c15-40f2-ac5d-8e3141ed428d_model.json b/2024/Garment Recovery with Shape and Deformation Priors/83e37ee5-1c15-40f2-ac5d-8e3141ed428d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2f7d5ac68b9f9390d5cfc4135cb350bc22cdfe13 --- /dev/null +++ b/2024/Garment Recovery with Shape and Deformation Priors/83e37ee5-1c15-40f2-ac5d-8e3141ed428d_model.json @@ -0,0 +1,2233 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.207, + 0.131, + 0.766, + 0.154 + ], + "angle": 0, + "content": "Garment Recovery with Shape and Deformation Priors" + }, + { + "type": "text", + "bbox": [ + 0.246, + 0.181, + 0.724, + 0.216 + ], + "angle": 0, + "content": "Ren Li Corentin Dumery Benoit Guillard Pascal Fua CVLab,EPFL" + }, + { + "type": "text", + "bbox": [ + 0.386, + 0.22, + 0.584, + 0.234 + ], + "angle": 0, + "content": "{name.surname}@epfl.ch" + }, + { + "type": "image", + "bbox": [ + 0.09, + 0.248, + 0.891, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.42, + 0.893, + 0.448 + ], + "angle": 0, + "content": "Figure 1. We propose a fitting method that leverages shape and deformation priors trained on synthetic data to recover realistic 3D garment mesh from in-the-wild images. We produce triangulated meshes that are directly usable for animation and simulation." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.457, + 0.314, + 0.473 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.491, + 0.471, + 0.658 + ], + "angle": 0, + "content": "While modeling people wearing tight-fitting clothing has made great strides in recent years, loose-fitting clothing remains a challenge. We propose a method that delivers realistic garment models from real-world images, regardless of garment shape or deformation. To this end, we introduce a fitting approach that utilizes shape and deformation priors learned from synthetic data to accurately capture garment shapes and deformations, including large ones. Not only does our approach recover the garment geometry accurately, it also yields models that can be directly used by downstream applications such as animation and simulation." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.692, + 0.21, + 0.708 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.902 + ], + "angle": 0, + "content": "3D clothed human recovery aims to reconstruct the body shape, pose, and clothing of many different people from images. It is key to applications such as fashion design, virtual try-on, 3D avatars, along with virtual and augmented reality. Recent years have seen tremendous progress in modeling people wearing tight-fitting clothing both in terms of body poses [6, 14, 21, 22, 25, 26, 28, 30, 38, 41, 54] and 3D shape of the clothes [5, 9-11, 20, 31, 32, 39]. However, loose-fitting clothing remains a challenge. Existing approaches either rely on mesh templates with limited generality, or produce models expressed in terms of 3D point clouds [37] or as a single watertight mesh that tightly binds the body" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.459, + 0.892, + 0.489 + ], + "angle": 0, + "content": "and garment together [53], neither of which is straightforward to integrate into downstream applications." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.491, + 0.893, + 0.717 + ], + "angle": 0, + "content": "In this paper, we propose a method that overcomes these limitations and can effectively recover the shape of loose fitting garments from single images. The recovered garments can then be animated without any additional processing, as shown in Fig. 1. Starting from the Implicit Sewing Patterns (ISP) model [32] that represents garments in terms of a set of individual 2D panels and 3D surfaces associated to these panels, we introduce a deformation model that we apply to the 3D surfaces so that they can deviate substantially from the body shape. These deformations are conditioned on normals estimated from an input image of the target garment. They are learned from synthetic mesh data featuring loose clothing, where the deformations are taken to be those required to fit individual ISP 3D surfaces to the ground-truth 3D meshes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Given the trained deformation model, we designed a two-stage fitting process to recover the 3D garment from in-the-wild images. First, the parameters of the pre-trained deformation model are optimized to produce a shape that minimizes the distance between garment outlines and segmented garment regions, the differences between garment normals and the normals estimated in the images by off-the-shelf-algorithms [46, 53], and a physics-based loss to promote physical plausibility of the results. Then, fine local details are recovered by directly optimizing the vertex positions of the reconstructed mesh with the same loss. Our fitting process does not require external 3D annotations, other" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1586" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.401, + 0.106 + ], + "angle": 0, + "content": "than the estimated normals of the target garment." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.107, + 0.468, + 0.227 + ], + "angle": 0, + "content": "We demonstrate that our method can recover garments that go from tight- to loose-fitting and outperforms existing approaches [9, 11, 20, 32, 39, 53] in terms of reconstruction accuracy. Furthermore, our reconstructed meshes are directly usable for virtual try-on or animation, unlike [37, 53]. Our implementation and model weights are available at https://github.com/liren2515/GarmentRecovery." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.24, + 0.219, + 0.255 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.265, + 0.468, + 0.326 + ], + "angle": 0, + "content": "Before the advent of Deep Learning, garment shape recovery from images depended mostly on user defined outlines and shape-from-shading techniques [59]. Since then, data-driven techniques have become dominant." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.342, + 0.468, + 0.401 + ], + "angle": 0, + "content": "Tight-Fitting Clothing. The majority of methods developed in recent years focus on clothing that clings relatively closely to the body. These can be classified into two main categories." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.402, + 0.468, + 0.853 + ], + "angle": 0, + "content": "In the first category, are methods that model garments as surfaces that are distinct from the body surface and interact with it. DeepGarment [10], MGN [5], and BCNet [20] train neural networks on synthetic images to predict the vertex positions of specific mesh templates. [8] and [34] leverage normal estimation to optimize the vertex position of the template and recover wrinkle details. Such methods are inherently limited in the range of shapes they can handle, and those being trained on synthetic data can easily fail when facing real images. To overcome these limitations, SMPLicit [9], DIG [31], and ClothWild [39] leverage Signed Distance Functions (SDF) to recover a wide array of garment meshes from RGB images and the corresponding segmentation masks. However, to represent non-watertight garment surfaces using an SDF, one has to wrap around them a watertight surface with a minimum thickness, which reduces accuracy. This can be addressed by using Unsigned Distance Functions (UDFs) instead [11, 15] but creates robustness issues: if the UDF is even slightly inaccurate, the value of the surface is never exactly zero and holes can appear in the reconstructed models. In our experience, the Implicit Sewing Patterns (ISP) model of [32] effectively addresses the issues of generality, accuracy, and robustness. The garments consist of flat 2D panels whose boundary is defined by a 2D SDF. To each panel is associated a 3D surface parameterized by the 2D panel coordinates. Hence, different articles of clothing are represented in a standardized way, which allows the recovery of various garments from single images. This is why we choose it as the basis for our approach." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "In the second category, are the many methods that represent body and garment using a single model. For example, in [19, 57] a volumetric regression network yields a voxel" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.289 + ], + "angle": 0, + "content": "representation of 3D clothed humans given a single image. Other works [3, 16, 45, 46] employ a pixel-aligned implicit function that defines 3D occupancy fields or signed distance fields for clothed humans. In [1, 2], displacement vectors or UV maps are used to represent deviations from a SMPL parametric body model [36]. Similarly, in [17, 18, 52, 58] parametric body models are combined with implicit representations to achieve robustness to pose changes. While effective, all these methods suffer from significant limitations, because they cannot separate the surface of the garment from that of the body, and they are at a disadvantage when it comes to modeling loose garments whose motion can be relatively independent from the body." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.305, + 0.892, + 0.606 + ], + "angle": 0, + "content": "Loose-Fitting Clothing. There is a more limited number of methods designed to handle free-flowing garments. Some recent works [55, 60, 61] rely on complex physics simulation steps or feature line estimation to align the surface reconstruction with the input image. However, their dependence on garment templates limits their generality, in the same way it did for other template-based methods discussed above. Point-based methods that can reconstruct generic clothes have been proposed [50, 56] to overcome this. Unfortunately, point clouds are not straightforward to integrate into downstream applications. As a result, the method of [50] resorts to modified Poisson Surface Reconstruction (PSR) to create a garment surface from the point cloud, which can result in incorrect geometry. Another point-based representation is introduced in [37]. While being successful at modeling and animating humans wearing loose garments, it also relies on PSR to infer the mesh from the point cloud, yielding a single mesh that represents body and garment jointly. Furthermore, [37] does not explore how this representation can be fitted to images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.608, + 0.892, + 0.728 + ], + "angle": 0, + "content": "ECON [53] is a method specifically designed for clothed human recovery from images. By leveraging techniques such as normal integration and shape completion, it achieves visually appealing results for individuals wearing loose clothing. However, as [37], ECON produces a single watertight mesh that tightly binds the body and garment together, precluding easy use for applications such as cloth simulation and re-animation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.74, + 0.591, + 0.755 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Given an image of a clothed person and a body model extracted from it using existing techniques, our goal is to recover accurate 3D models of the garments matching the image. To this end, we add to the Implicit Sewing Pattern (ISP) garment model [32], which provides us with a shape prior for garment in its rest state, a deformation model that allows us to recover its potentially large deformations, as illustrated by Fig. 2. As a result, whereas the original ISP, like most current clothes-recovery algorithms, is limited to" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1587" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.09, + 0.885, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.379, + 0.893, + 0.437 + ], + "angle": 0, + "content": "Figure 2. Framework. Given an image, (1) we first estimate the normal map of the target garment and the SMPL body parameters \\((\\beta, \\theta)\\), which are used to compute the body part segmentation and position maps. (2) The maximum coverage garment shape \\(\\bar{\\mathcal{M}}\\) is then skinned to closely fit to the body, yielding \\(\\mathcal{M}\\). Leveraging (3) pixel-aligned image features, our deformation model (4) predicts occupancy and position maps to correct \\(\\mathcal{M}\\) for large deformations. (5) The 3D garment mesh is recovered using ISP and further refined." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.453, + 0.47, + 0.483 + ], + "angle": 0, + "content": "tight-fitting clothing, our approach can handle both tight- and loose-fitting garments, such as skirts and open jackets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.483, + 0.471, + 0.575 + ], + "angle": 0, + "content": "In this section, we first describe briefly the ISP model upon which we build our approach. We then introduce the deformation model that underpins the main contribution of this paper, going from tight fitting clothes to loosely fitting and free flowing ones. Finally, we present our approach to fitting this model to real-world images." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.582, + 0.271, + 0.597 + ], + "angle": 0, + "content": "3.1. ISP Garment Model" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.606, + 0.47, + 0.696 + ], + "angle": 0, + "content": "ISP is a garment model inspired by the sewing patterns that fashion designers use to represent clothes. A sewing pattern is made of several 2D panels along with information about how to stitch them into a complete garment. ISP implicitly models patterns using a 2D signed distance field and a 2D label field." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.714, + 0.471, + 0.775 + ], + "angle": 0, + "content": "Formalization. Given the latent code \\(\\mathbf{z}\\) of a garment and a point \\(\\mathbf{u}\\) in the 2D UV space \\(\\Omega = [-1,1]^2\\) of a 2D panel of that garment, ISP outputs the signed distance \\(s\\) to the panel boundary and a label \\(l\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.209, + 0.785, + 0.47, + 0.802 + ], + "angle": 0, + "content": "\\[\n(s, l) = \\mathcal {I} _ {\\Theta} (\\mathbf {u}, \\mathbf {z}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathcal{I}_{\\Theta}\\) is a fully connected network. The zero crossing of the SDF defines the shape of the panel, with \\(s < 0\\) indicating that \\(\\mathbf{u}\\) is within the panel and \\(s > 0\\) indicating that \\(\\mathbf{u}\\) is outside the panel. The stitch information is encoded in \\(l\\), where panel boundaries with the same label should be stitched together. To transform the 2D sewing patterns into" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.453, + 0.893, + 0.484 + ], + "angle": 0, + "content": "3D surfaces, a UV parameterization function \\(\\mathcal{A}_{\\Phi}\\) is learned to perform the 2D-to-3D mapping" + }, + { + "type": "equation", + "bbox": [ + 0.64, + 0.495, + 0.893, + 0.512 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} = \\mathcal {A} _ {\\Phi} (\\mathbf {u}, \\mathbf {z}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.893, + 0.629 + ], + "angle": 0, + "content": "where \\(\\mathbf{X} \\in \\mathbb{R}^3\\) represents the 3D position of \\(\\mathbf{u}\\). Essentially, ISP registers each garment onto a unified 2D space \\(\\Omega\\), and represents it using UV maps that record 2D-to-3D mapping, as shown in Fig. 3(b). Given the paired 2D sewing patterns and their 3D meshes, the pattern parameterization network \\(\\mathcal{I}_{\\Theta}\\) and the UV parameterization network \\(\\mathcal{A}_{\\Phi}\\) are trained by minimizing the losses" + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.639, + 0.891, + 0.657 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathcal {I}} = L _ {S D F} + L _ {C E} + \\left\\| \\mathbf {z} \\right\\| _ {2} ^ {2}, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.659, + 0.892, + 0.675 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathcal {A}} = L _ {M S E} + L _ {\\text {c o n s i s t}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.687, + 0.892, + 0.777 + ], + "angle": 0, + "content": "respectively. \\(L_{SDF}\\) is the mean absolute error for the predicted SDF value \\(s\\), \\(L_{CE}\\) is the cross-entropy loss for the predicted label \\(l\\), \\(L_{MSE}\\) is the mean squared error of the predicted 3D position \\(\\mathbf{X}\\), and \\(L_{consist}\\) is the loss to reduce the gap between the front and back panels. More details can be found in [32]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Training. Training ISP requires the 2D sewing patterns of 3D garments in a rest state, that is, draped on a T-posed neutral body. However, patterns are absent in most large scale garment datasets, including CLOTH3D [4]. We use the garment flattening algorithm of [42] to generate the required 2D patterns. The garment mesh is first divided into several pieces given predefined cutting rules. These pieces" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1588" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.123, + 0.087, + 0.351, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.087, + 0.608, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.087, + 0.852, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.259, + 0.894, + 0.288 + ], + "angle": 0, + "content": "Figure 3. Cutting and flattening. (a) The front (top) and back (bottom) surfaces after cutting. (b) The flattened panels and UV maps generated by ISP for (a). (c) The maximum-coverage UV maps and its represented 3D shape." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.314, + 0.469, + 0.435 + ], + "angle": 0, + "content": "are then unfolded into 2D panels by minimizing an as-rigid-as-possible [33] energy, ensuring local area preservation between the 3D and 2D parameterizations. While doing this, we constrain the boundary vertices at places such as waist and sleeves to have a constant value along a specific axis. This enhances pattern consistency across the dataset. Fig. 3 illustrates this and shows the panels generated for a shirt. We provide more details in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.435, + 0.47, + 0.524 + ], + "angle": 0, + "content": "We generate a front and a back panels as the sewing pattern for each garment in our dataset (shirt, skirt, and trousers). Once ISP has been trained on these, we compute the maximum-coverage UV maps \\(\\bar{\\mathcal{M}}\\) over the UV maps of each garment for each category, as shown in Fig. 3(c). The values of \\(\\bar{\\mathcal{M}}\\) are taken to be" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.527, + 0.469, + 0.564 + ], + "angle": 0, + "content": "\\[\n\\bar {\\mathcal {M}} [ u, v ] = \\frac {\\sum_ {i} \\mathbb {1} _ {s _ {\\mathbf {u}} ^ {i} \\leq 0} \\cdot m _ {i} [ u , v ]}{\\sum_ {i} \\mathbb {1} _ {s _ {\\mathbf {u}} ^ {i} \\leq 0}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.567, + 0.47, + 0.718 + ], + "angle": 0, + "content": "where \\( m_{i} \\) is the UV maps of garment \\( i \\), \\( \\mathbb{1} \\) is the indicator function, \\( [\\cdot ,\\cdot ] \\) denotes array addressing, and \\( s_{\\mathbf{u}}^{i} \\) is the SDF value of ISP at \\( \\mathbf{u} = (u,v) \\). The maximum coverage map \\( \\bar{\\mathcal{M}} \\) encompasses information from all the patterns in the dataset. It represents the smallest possible map that covers all garments in a category, with the 3D position of each \\( uv \\)-point \\( \\mathbf{u} \\) being the average of all garments that include \\( \\mathbf{u} \\). We use it as a prototype for a garment category and to compute an initial guess of the deformed garment given the body pose, as discussed below." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.843 + ], + "angle": 0, + "content": "**Skinning.** As in many prior work [11, 31, 32], we use SMPL [36] to parameterize the body in terms of shape and pose parameters \\((\\beta, \\theta)\\), and its extended skinning procedure for the 3D volume around the body to initially deform the 3D shape represented as \\(\\bar{\\mathcal{M}}\\). More specifically, given a 2D point \\(\\mathbf{u} = (u, v)\\) in UV space \\(\\Omega\\), we get its actual 3D position as \\(\\bar{\\mathbf{X}}_{\\mathbf{u}} = \\bar{\\mathcal{M}}[u, v]\\). We then deform it by computing" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.861, + 0.469, + 0.88 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {\\mathbf {u}} = W \\left(\\mathbf {X} _ {(\\beta , \\theta)}, \\beta , \\theta , w (\\bar {\\mathbf {X}} _ {\\mathbf {u}}) \\mathcal {W}\\right), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.881, + 0.413, + 0.899 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {(\\beta , \\theta)} = \\bar {\\mathbf {X}} _ {\\mathbf {u}} + w (\\bar {\\mathbf {X}} _ {\\mathbf {u}}) B _ {\\beta} + w (\\bar {\\mathbf {X}} _ {\\mathbf {u}}) B _ {\\theta},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.314, + 0.892, + 0.451 + ], + "angle": 0, + "content": "where \\(W(\\cdot)\\) is the SMPL skinning function with skinning weights \\(\\mathcal{W} \\in \\mathbb{R}^{N_B \\times 24}\\), with \\(N_B\\) being the number of vertices of the SMPL body mesh, and \\(B_\\beta \\in \\mathbb{R}^{N_B \\times 3}\\) and \\(B_\\theta \\in \\mathbb{R}^{N_B \\times 3}\\) are the shape and pose displacements of SMPL, respectively. The diffused weights \\(w(\\cdot) \\in \\mathbb{R}^{N_B}\\) are computed by a neural network, which generalizes the SMPL skinning to any point in 3D space. By repeating this for all points in a panel, we obtain the vertex position map \\(\\mathcal{M}[u,v] = \\mathbf{X}_{\\mathbf{u}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.459, + 0.768, + 0.475 + ], + "angle": 0, + "content": "3.2. Modeling Large Deformations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.482, + 0.892, + 0.588 + ], + "angle": 0, + "content": "As shown in Fig. 2, the ISP approach described above generates a garment prototype \\(\\mathcal{M}\\) that closely fits the underlying body. To a point \\(\\mathbf{u} = (u,v)\\) that belongs to a given garment panel, it associates the vertex position \\(\\mathbf{X}_{\\mathbf{u}} = \\mathcal{M}[u,v]\\) of Eq. 6, which is usually relatively close to the body. To model loose clothing, we now need to compute a potentially larger displacement \\(\\Delta \\mathbf{X}_{\\mathbf{u}}\\) to be added to \\(\\mathbf{X}_{\\mathbf{u}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.588, + 0.892, + 0.694 + ], + "angle": 0, + "content": "To evaluate \\(\\Delta \\mathbf{X}_{\\mathbf{u}}\\) we train a network \\(\\mathcal{D}\\) that consists of an MLP to estimate the occupancy value and the corrective displacement value, and two CNNs to extract image features \\(F_{n}\\) and \\(F_{b}\\) from the image of normals and the segmentation and vertex position images of the SMPL body, respectively. Fig. 2 depicts this architecture. We obtain the pixel-aligned image feature for \\(\\mathbf{u}\\) by computing" + }, + { + "type": "equation", + "bbox": [ + 0.612, + 0.705, + 0.891, + 0.721 + ], + "angle": 0, + "content": "\\[\nF \\left(\\mathbf {x} _ {\\mathbf {u}}\\right) = F _ {n} \\oplus F _ {b} \\left(\\mathbf {x} _ {\\mathbf {u}}\\right), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.638, + 0.724, + 0.741, + 0.74 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x _ {u}} = P (\\mathbf {X _ {u}}) ,\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(P(\\cdot)\\) denotes the projection into the image and \\(\\oplus\\) concatenation. The MLP takes as input \\(\\mathbf{u}\\), its 3D position and image features to predict the occupancy \\(O_{\\mathbf{u}} \\in \\{0,1\\}\\) and the corrective displacement \\(\\Delta \\mathbf{X}_{\\mathbf{u}}\\). By assembling the results of each point in the UV space, we obtain the final occupancy maps \\(\\mathcal{O}\\) and the vertex position maps \\(\\hat{\\mathcal{M}}\\), where \\(\\hat{\\mathcal{M}}[u,v] = \\mathbf{X}_{\\mathbf{u}} + \\Delta \\mathbf{X}_{\\mathbf{u}}\\). In essence, \\(\\mathcal{O}\\) is the binarized SDF of the garment 2D panels, which implicitly defines the garment shape and geometry in the rest state, while \\(\\hat{\\mathcal{M}}\\) encodes the deformed state for that garment. The 3D garment mesh" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1589" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "can be recovered from these 2D maps with ISP, as discussed below." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.122, + 0.468, + 0.182 + ], + "angle": 0, + "content": "We use a single network \\(\\mathcal{D}\\) for the front and the back panels, and encode the points \\(\\mathbf{u}\\) of the front as \\([u,v,+1]\\) and those of the back as \\([u,v,-1]\\). We learn the parameters of \\(\\mathcal{D}\\) by minimizing" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.192, + 0.469, + 0.224 + ], + "angle": 0, + "content": "\\[\nL = \\sum_ {\\mathbf {u} \\in \\Omega} | | \\mathbf {X} _ {\\mathbf {u}} + \\Delta \\mathbf {X} _ {\\mathbf {u}} - \\tilde {\\mathbf {X}} _ {\\mathbf {u}} | | _ {2} + \\lambda \\sum_ {\\mathbf {u} \\in \\Omega} B E (O _ {\\mathbf {u}}, \\tilde {O} _ {\\mathbf {u}}), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.236, + 0.469, + 0.282 + ], + "angle": 0, + "content": "where \\(\\tilde{\\mathbf{X}}_{\\mathbf{u}}\\) and \\(\\tilde{O}_{\\mathbf{u}}\\) are the ground truth vertex positions and occupancy values, \\(BE(\\cdot)\\) is the binary cross entropy, and \\(\\lambda\\) is a weighting constant." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.299, + 0.469, + 0.398 + ], + "angle": 0, + "content": "From 2D maps to 3D mesh. The occupancy value \\( O_{\\mathbf{u}} \\) indicates whether \\( \\mathbf{u} \\) falls within the panels of the garment. To convert the generated 2D occupancy maps \\( \\mathcal{O} \\) into a 3D garment mesh in its rest shape, we need to recover an ISP latent code \\( \\mathbf{z} \\) as defined in Eq. 1. To this end, we find the vector \\( \\mathbf{z} \\) such that the corresponding SDF of ISP best matches the produced occupancy maps by minimizing" + }, + { + "type": "equation", + "bbox": [ + 0.084, + 0.405, + 0.469, + 0.434 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} ^ {*} = \\underset {\\mathbf {z}} {\\operatorname {a r g m i n}} \\sum_ {\\mathbf {u} \\in \\Omega_ {-}} R \\left(s _ {\\mathbf {u}} (\\mathbf {z})\\right) + \\sum_ {\\mathbf {u} \\in \\Omega_ {+}} R \\left(- s _ {\\mathbf {u}} (\\mathbf {z})\\right) + \\lambda_ {\\mathbf {z}} \\| \\mathbf {z} \\| _ {2}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.443, + 0.469, + 0.609 + ], + "angle": 0, + "content": "where \\(\\Omega_{-} = \\{\\mathbf{u}|O_{\\mathbf{u}} = 1,\\mathbf{u}\\in \\Omega \\}\\), \\(\\Omega_{+} = \\{\\mathbf{u}|O_{\\mathbf{u}} = 0,\\mathbf{u}\\in \\Omega \\}\\), \\(R(\\cdot)\\) is the ReLU function, \\(s_{\\mathbf{u}}(\\mathbf{z})\\) is the SDF value of \\(\\mathbf{u}\\) computed by ISP, and \\(\\lambda_{\\mathbf{z}}\\) is the weighting constant. With \\(\\mathbf{z}^*\\), the rest garment mesh is inferred through ISP's meshing and sewing process. The deformed garment mesh can be obtained by simply replacing the vertex position of the recovered mesh with the values stored in \\(\\hat{\\mathcal{M}}\\). This yields 3D garment meshes in both the rest and deformed states as shown in the top-right of Fig. 2, which are required for the application such as cloth simulation and can be used for further refinement as discussed below." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.619, + 0.334, + 0.635 + ], + "angle": 0, + "content": "3.3. Fitting the Models to Images" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.642, + 0.469, + 0.792 + ], + "angle": 0, + "content": "For practical reasons, the range of garment materials, external forces, and body motions present in the training data is limited. As a result, given in-the-wild images as input, the trained model can produce inaccurate results as shown in Fig. 4(b). To remedy this and to leverage the deformation prior that the network \\(\\mathcal{D}\\) captures, we refine the result by minimizing a loss function with respect to the pretrained deformation parameters of \\(\\mathcal{D}\\) as in [13, 21, 51]. \\(L\\) is designed to promote a good match between the garment mesh and image observations. We take it to be" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.805, + 0.468, + 0.82 + ], + "angle": 0, + "content": "\\[\nL = \\lambda_ {C} L _ {C D} + \\lambda_ {n} L _ {\\text {n o r m a l}} + \\lambda_ {p} L _ {\\text {p h y s i c s}}, \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.119, + 0.824, + 0.468, + 0.842 + ], + "angle": 0, + "content": "\\[\nL _ {C D} = d \\left(x _ {c} ^ {f}, x _ {I _ {n}}\\right), \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.096, + 0.845, + 0.468, + 0.876 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {n o r m a l}} = \\sum_ {i \\in f} 1 - \\cos \\left(n ^ {i}, I _ {n} \\left(x _ {c} ^ {i}\\right)\\right), \\tag {12}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.096, + 0.882, + 0.468, + 0.898 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {p h y s i c s}} = L _ {\\text {s t r a i n}} + L _ {\\text {b e n d}} + L _ {\\text {g r a v i t y}} + L _ {\\text {c o l}}, \\tag {13}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.547, + 0.089, + 0.695, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.612, + 0.202, + 0.629, + 0.213 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.108, + 0.834, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.765, + 0.2, + 0.783, + 0.213 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.549, + 0.221, + 0.692, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.612, + 0.31, + 0.628, + 0.321 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.221, + 0.845, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.765, + 0.309, + 0.783, + 0.322 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.333, + 0.892, + 0.402 + ], + "angle": 0, + "content": "Figure 4. Fitting results. Given (a) the normal estimation of an in-the-wild image, (b) is the inference result with ISP recovered geometry. (c) is obtained by optimizing the parameters of the pretrained deformation model. Further refinement of the mesh vertex positions yields (d)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.42, + 0.892, + 0.678 + ], + "angle": 0, + "content": "where \\( x_{c}^{f} \\) is the 2D projection of the centers of visible faces \\( f \\) after mesh rasterization, \\( x_{I_n} \\) denotes the coordinates of foreground pixels, \\( n^i \\) is the normal of face \\( i \\), \\( I_{n}(x_{c}^{i}) \\) is the normal image values at \\( x_{c}^{i} \\), and \\( \\lambda_{C} \\), \\( \\lambda_{n} \\) and \\( \\lambda_{p} \\) are the balancing scalars. \\( d(\\cdot) \\) and \\( \\cos (\\cdot) \\) are the functions measuring the 2D Chamfer Distance and the cosine similarity, respectively. \\( L_{\\mathrm{physics}} \\) is a physics-based loss derived from [40, 48], which computes the membrane strain energy \\( L_{strain} \\) caused by the deformation, the bending energy \\( L_{bend} \\) resulting from the folding of adjacent faces, the gravitational potential energy \\( L_{gravity} \\) and the penalty for body-garment collision \\( L_{col} \\). Minimizing \\( L_{CD} \\) induces an external force of stretching or compression on the garment mesh to align its 2D projection with the given image, while minimizing \\( L_{\\mathrm{physics}} \\) ensures that the mesh exhibits physically plausible deformation adhering to the shape constraints of the rest-state mesh recovered by ISP." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.679, + 0.891, + 0.814 + ], + "angle": 0, + "content": "Minimizing the loss \\( L \\) with respect to the deformation parameters yields a mesh whose overall shape matches the input image, as illustrated in Fig. 4(c). However, since neural networks tend to learn low-frequency functions [43], the result might be too smooth. To recover fine surface details, we perform a refinement step by minimizing \\( L \\) directly with respect to the coordinates of the garment mesh vertices. This generates realistic local details, such as wrinkles on the surface, as shown in Fig. 4(d)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.83, + 0.632, + 0.847 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In short, our method begins by inferring the shape and the deformation of the garment in terms of the occupancy and position maps. Leveraging the shape prior of ISP, we then" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1590" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.213 + ], + "angle": 0, + "content": "recover the garment geometry from the occupancy maps and deform it using the position maps. Next, we refine the initial deformed mesh to better align with image observations by fine-tuning the pre-trained network \\(\\mathcal{D}\\) that captures the deformation prior. Finally, we recover fine details through vertex-level optimization of the garment mesh. In this section, we demonstrate the effectiveness of this process and compare it to that of other state-of-the-art methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.224, + 0.295, + 0.24 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.247, + 0.473, + 0.641 + ], + "angle": 0, + "content": "Following the implementation of [32], both the pattern parameterization model \\(\\mathcal{I}_{\\Theta}\\) and the UV parameterization model \\(\\mathcal{A}_{\\Phi}\\) of ISP have two separate MLPs for the front and back panels. Each MLP has 7 layers with Softplus activations. The dimension of latent code \\(\\mathbf{z}\\) is 32. The skinning weight model \\(w\\) is a 9-layer MLP with leaky ReLU activations, whose output is normalized by a final Softmax layer. \\(\\mathcal{I}_{\\Theta}\\) and \\(\\mathcal{A}_{\\Phi}\\) are trained jointly for 9000 iterations with a batch size of 50. \\(w\\) is trained with the same parameters as [32]. For the image feature extraction, we use two separate ConvNeXt [35] networks to extract multi-scale garment and body features of sizes 96, 192, 384, 768, which are concatenated as the final features. The point UV coordinates \\(\\mathbf{u}\\), 3D position \\(\\mathbf{X_u}\\), and image features \\(F(\\mathbf{x_u})\\) are projected separately to 384 dimensions by three linear layers, which are then concatenated as the input of the MLP of the deformation model. The MLP of the deformation model has 10 layers with a skip connection from the input layer to the middle, and uses Gaussian functions as the activation layer following [44]. The CNNs and MLP of the deformation model are trained jointly for 40 epochs with the Adam optimizer [23] and a learning rate of \\(10^{-4}\\). For real images, we use [53] and [12] to obtain their normal and SMPL body parameter estimations, respectively. The garment segmentation masks are generated by leveraging the segmentation of SAM [24] and the semantic labels of [29]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.651, + 0.434, + 0.667 + ], + "angle": 0, + "content": "4.2. Dataset, Evaluation Metrics, and Baseline" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Our models are trained on CLOTH3D [4], which is a synthetic dataset with motion sequences of 3D clothed human. It contains garment in the rest state, and deformed states caused by the motion of underlying body. For each garment, it has a single simulated sequence up to 10 seconds. It covers a large variety of garment in different shapes, types and topologies. For the training of ISP, we randomly select 400 shirts, 200 skirts and 200 pairs of trousers, and generate their sewing patterns by the method described in Sec. 3.1. The deformation model is trained on the corresponding simulated sequences. For each frame, we render 11 normal images for the garment mesh with random rotations around the Y-axis, which produces 40K, 40K and 20K training images for shirt, skirt and trousers respectively. During training, we augment the data with image flipping and rotation." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.092, + 0.892, + 0.166 + ], + "angle": 0, + "content": "To evaluate the garment reconstruction quality, we use the Chamfer Distance (CD) between the ground truth and the recovered garment mesh, and Intersection over Union (IoU) between the ground truth mask and the rendered mask of reconstructed garment mesh." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.168, + 0.894, + 0.244 + ], + "angle": 0, + "content": "We compare our method against state-of-the-art methods BCNet [20], SMPLicit [9], ClothWild [39], DrapeNet [11] and ISP [32]. SMPLicit, DrapeNet and ISP use the garment segmentation mask for reconstruction, while BCNet and ClothWild take the RGB images as input." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.25, + 0.869, + 0.267 + ], + "angle": 0, + "content": "4.3. Comparison with State-of-the-Art Methods" + }, + { + "type": "table", + "bbox": [ + 0.496, + 0.285, + 0.895, + 0.372 + ], + "angle": 0, + "content": "
CD (×103) ↓SkirtShirtTrousersIoU ↑SkirtShirtTrousers
SMPLicit8.053.490.81SMPLicit0.6570.5270.782
DrapeNetn/a1.540.84DrapeNetn/a0.7150.845
ISP7.262.020.91ISP0.7090.6650.781
Ours2.870.440.28Ours0.9400.9390.946
Ours-GT1.670.310.23Ours-GT0.9530.9420.954
" + }, + { + "type": "table_caption", + "bbox": [ + 0.497, + 0.375, + 0.894, + 0.459 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons. Our method outperforms SMPLicit, DrapeNet, and ISP in terms of CD and IoU on all three garment categories, as shown in the second-to-last row of both tables (Ours). These results were obtained using normals estimated from the images using [53]. In the last row (Ours-GT), we provide the results we obtained using ground-truth normals instead." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.491, + 0.892, + 0.687 + ], + "angle": 0, + "content": "Quantitative Results. Due to the absence of publicly available real dataset for evaluating garment reconstruction, we utilize synthetic data consisting of 30 unseen shirts, 30 unseen skirts, and 30 pairs of unseen trousers from CLOTH3D for quantitative evaluation. We compare our method with SMPLicit, DrapeNet, and ISP, which recover garment meshes from segmentation masks. The segmentation masks required by the baselines are rendered from the ground truth mesh data, while the normal images required by our method are estimated with [53]. As shown in Tab. 1, our approach significantly outperforms the baselines in terms of Chamfer Distance (CD) and Intersection over Union (IoU) across all garment categories." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.688, + 0.894, + 0.793 + ], + "angle": 0, + "content": "For comparison purposes, we re-ran our algorithm using the ground-truth normals and report the results in the last row of Tab. 1. As expected, this leads to an improvement in reconstruction accuracy, but the increase is only modest. This highlights the robustness of our approach to the slight inaccuracies that can be expected from a normal-estimation algorithm." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.81, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Qualitative Results. Fig. 5 shows the qualitative comparison for the results reconstructed from in-the-wild images. Since BCNet is trained only on synthetic RGB data, it is not able to predict accurate body and garment results. Both SMPLicit and ClothWild generate unrealistic watertight meshes that do not accurately represent real garments." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1591" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.887, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.427, + 0.895, + 0.471 + ], + "angle": 0, + "content": "Figure 5. Comparison against SOTA methods. From left to right, we show the input image and the 3D garment meshes recovered by our method and SOTA methods: BCNet, SMPLcit, ClothWild, DrapeNet, ISP. (Since skirt is unavailable for DrapeNet, a random pair of trousers is put on its result in the second row.)" + }, + { + "type": "table", + "bbox": [ + 0.083, + 0.493, + 0.268, + 0.582 + ], + "angle": 0, + "content": "
CD (×103) ↓GTEst.
(-D*,-v*)2.043.58
(-D*,+v*)1.803.39
(+D*,-v*)1.752.93
(+D*,+v*)1.672.87
" + }, + { + "type": "table", + "bbox": [ + 0.281, + 0.493, + 0.467, + 0.582 + ], + "angle": 0, + "content": "
IoU ↑GTEst.
(-D*,-v*)0.8480.751
(-D*,+v*)0.9470.904
(+D*,-v*)0.9410.923
(+D*,+v*)0.9530.940
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.604, + 0.471, + 0.661 + ], + "angle": 0, + "content": "Table 2. Ablation study. GT and Est. mean using ground-truth and estimated normal images as input, respectively. \\( + / - {\\mathcal{D}}^{ * } \\) denote with/without the finetuning of the deformation model. \\( + / - {v}^{ * } \\) denote with/without the optimization of vertex positions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.687, + 0.47, + 0.779 + ], + "angle": 0, + "content": "While DrapeNet and ISP are able to recover open surfaces for garments, their results remain closely adhered to the body, similar to SMPLlicit and ClothWild. In contrast, our method can faithfully recover garment mesh from input images, as clearly observed in the depiction of the open jacket and the flowing long skirt that stand away from the body." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.231, + 0.803 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Table 2 presents the evaluation results of our method with and without fitting on the test set of skirts. The results of row 1, 3 and 4 indicate that optimizing the parameters of the pretrained deformation network \\((+D^{*}, - v^{*})\\) improves the quality of the raw inference results, and optimizing the vertex positions \\((+D^{*}, + v^{*})\\) further enhances the outcome," + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.493, + 0.639, + 0.624 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.5, + 0.747, + 0.624 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.697, + 0.624, + 0.716, + 0.637 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.771, + 0.502, + 0.868, + 0.623 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.817, + 0.624, + 0.837, + 0.637 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.648, + 0.892, + 0.676 + ], + "angle": 0, + "content": "Figure 6. Fitting strategy comparison between (a) our fitting method and (b) only vertex optimization." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.704, + 0.768, + 0.719 + ], + "angle": 0, + "content": "yielding the lowest CD and highest IoU." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "However, as demonstrated in the second row, directly optimizing the vertex positions on the raw inference results without optimizing the deformation model \\((-\\mathcal{D}^{*}, + v^{*})\\) is not as effective. The deformation model encapsulates a continuous deformation field. When its weights are optimized based on partial observations, the entire field undergoes modification, thereby influencing all mesh vertices globally. On the contrary, a direct vertex optimization with partial observations predominantly affects the mesh vertices locally. While this can capture localized details like wrinkles, it struggles to resolve discrepancies in the overall shape. Fig. 6 displays the results of our methods and an ablation with" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "1592" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.09, + 0.895, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.66, + 0.893, + 0.688 + ], + "angle": 0, + "content": "Figure 7. Recovering from in-the-wild images. Our method is able to recover realistic meshes for garments with diverse shapes and deformations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.706, + 0.471, + 0.751 + ], + "angle": 0, + "content": "vertex optimization only. Unsurprisingly, the latter fails at recovering the correct shape and produces implausible deformation on the mesh surface." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.792, + 0.217, + 0.807 + ], + "angle": 0, + "content": "4.5. More Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Fig. 7 shows a collection of results reconstructed by our method from in-the-wild images. Our method can produce realistic 3D meshes with fine details across a wide range of garment types, from tight-fitting attire to more relaxed and flowing outfits." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.705, + 0.62, + 0.72 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.73, + 0.892, + 0.851 + ], + "angle": 0, + "content": "We have presented a novel approach to recovering realistic 3D garment meshes from in-the-wild images featuring loose fitting clothing. It relies on a fitting process that imposes shape and deformation priors learned on synthetic data to accurately capture garment shape and deformations. In future work, we will extend our approach to modeling deformations over time from video sequences while enforcing temporal consistency of the reconstructions." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.87, + 0.892, + 0.898 + ], + "angle": 0, + "content": "Acknowledgement. This project was supported in part by the Swiss National Science Foundation." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1593" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] T. Alldieck, M. Magnor, B. Bhatnagar, C. Theobalt, and G. Pons-Moll. Learning to reconstruct people in clothing from a single RGB camera. In Conference on Computer Vision and Pattern Recognition, pages 1175-1186, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.469, + 0.227 + ], + "angle": 0, + "content": "[2] T. Alldieck, G. Pons-Moll, C. Theobalt, and M. Magnor. Tex2shape: Detailed full human body geometry from a single image. In International Conference on Computer Vision, pages 2293-2303, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.469, + 0.282 + ], + "angle": 0, + "content": "[3] T. Alldieck, M. Zanfir, and C. Sminchisescu. Photorealistic monocular 3d reconstruction of humans wearing clothing. In Conference on Computer Vision and Pattern Recognition, pages 1506-1515, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.284, + 0.469, + 0.325 + ], + "angle": 0, + "content": "[4] H. Bertiche, M. Madadi, and S. Escalera. CLOTH3D: Clothed 3D Humans. In European Conference on Computer Vision, pages 344-359, 2020. 3, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.469, + 0.38 + ], + "angle": 0, + "content": "[5] B. L. Bhatnagar, G. Tiwari, C. Theobalt, and G. Pons-Moll. Multi-Garment Net: Learning to Dress 3D People from Images. In International Conference on Computer Vision, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.469, + 0.437 + ], + "angle": 0, + "content": "[6] F. Bogo, A. Kanazawa, C. Lassner, P. Gehler, J. Romero, and M. J. Black. Keep It SMPL: Automatic Estimation of 3D Human Pose and Shape from a Single Image. In European Conference on Computer Vision, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.469, + 0.479 + ], + "angle": 0, + "content": "[7] R. Brouet, A. Sheffer, L. Boissieux, and M.P. Cani. Design Preserving Garment Transfer. ACM Transactions on Graphics, 31(4):361-3611, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.48, + 0.469, + 0.522 + ], + "angle": 0, + "content": "[8] A. Casado-Elvira, M. Trinidad, and D. Casas. PERGAMO: Personalized 3d garments from monocular video. In Computer Graphics Forum, pages 293–304, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.523, + 0.469, + 0.578 + ], + "angle": 0, + "content": "[9] E. Corona, A. Pumarola, G. Alenya, G. Pons-Moll, and F. Moreno-Noguer. SMPlicit: Topology-Aware Generative Model for Clothed People. In Conference on Computer Vision and Pattern Recognition, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.579, + 0.469, + 0.62 + ], + "angle": 0, + "content": "[10] R. Danerek, E. Dibra, C. Öztireli, R. Ziegler, and M. Gross. Deepgarment : 3D Garment Shape Estimation from a Single Image. Eurographics, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.621, + 0.469, + 0.676 + ], + "angle": 0, + "content": "[11] Luca DeLuigi, Ren Li, Benoit Guillard, Mathieu Salzmann, and Pascal Fua. DrapeNet: Generating Garments and Draping them with Self-Supervision. In Conference on Computer Vision and Pattern Recognition, 2023. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.469, + 0.731 + ], + "angle": 0, + "content": "[12] Y. Feng, V. Choutas, T. Bolkart, D. Tzionas, and M. J. Black. Collaborative Regression of Expressive Bodies using Moderation. In International Conference on Computer Vision, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.469, + 0.773 + ], + "angle": 0, + "content": "[13] M. Gadelha, R. Wang, and S. Maji. Deep manifold prior. In International Conference on Computer Vision, pages 1107-1116, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.469, + 0.829 + ], + "angle": 0, + "content": "[14] G. Georgakis, R. Li and S. Karanam, T. Chen, J. Košecka, and Z. Wu. Hierarchical kinematic human mesh recovery. In European Conference on Computer Vision, pages 768-784, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.469, + 0.873 + ], + "angle": 0, + "content": "[15] B. Guillard, F. Stella, and P. Fua. MeshUDF: Fast and Differentiable Meshing of Unsigned Distance Field Networks. In European Conference on Computer Vision, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[16] T. He, J. Collomosse, H. Jin, and S. Soatto. Geo-pifu: Geometry and pixel aligned implicit functions for single-view" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "human reconstruction. Advances in Neural Information Processing Systems, 33:9276-9287, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[17] T. He, Y. Xu, S. Saito, S. Soatto, and T. Tung. Arch++: Animation-ready clothed human reconstruction revisited. In International Conference on Computer Vision, pages 11046-11056, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[18] Z. Huang, Y. Xu, C. Lassner, H. Li, and T. Tung. Arch: Animatable Reconstruction of Clothed Humans. In Conference on Computer Vision and Pattern Recognition, pages 3093-3102, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.235, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[19] A. Jackson, C. Manafas, and G. Tzimiropoulos. 3d human body reconstruction from a single image via volumetric regression. In European Conference on Computer Vision Workshops, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.291, + 0.892, + 0.333 + ], + "angle": 0, + "content": "[20] B. Jiang, J. Zhang, Y. Hong, J. Luo, L. Liu, and H. Bao. Bcnet: Learning body and cloth shape from a single image. In European Conference on Computer Vision, 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.387 + ], + "angle": 0, + "content": "[21] H. Joo, N. Neverova, and A. Vedaldi. Exemplar Fine-Tuning for 3D Human Pose Fitting Towards In-the-Wild 3D Human Pose Estimation. In International Conference on 3D Vision, 2020. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.39, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[22] A. Kanazawa, M. J. Black, D. W. Jacobs, and J. Malik. End-To-End Recovery of Human Shape and Pose. In Conference on Computer Vision and Pattern Recognition, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.434, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[23] D. P. Kingma and J. Ba. Adam: A Method for Stochastic Optimisation. In International Conference on Learning Representations, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.476, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[24] A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. Berg, W. Lo, P. Dollar, and R. Girshick. Segment anything. In International Conference on Computer Vision, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.532, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[25] N. Kolotouros, G. Pavlakos, M. J. Black, and K. Daniilidis. Learning to Reconstruct 3D Human Pose and Shape via Model-Fitting in the Loop. In Conference on Computer Vision and Pattern Recognition, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.589, + 0.892, + 0.644 + ], + "angle": 0, + "content": "[26] C. Lassner, J. Romero, M. Kiefel, F. Bogo, M.J. Black, and P.V. Gehler. Unite the People: Closing the Loop Between 3D and 2D Human Representations. In Conference on Computer Vision and Pattern Recognition, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.646, + 0.892, + 0.686 + ], + "angle": 0, + "content": "[27] B. Lévy, S. Petitjean, N. Ray, and J. Maillot. Least Squares Conformal Maps for Automatic Texture Atlas Generation. In ACM SIGGRAPH, 2002. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.688, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[28] J. Li, C. Xu, Z. Chen, S. Bian, L. Yang, and C. Lu. Hybrid: A hybrid analytical-neural inverse kinematics solution for 3d human pose and shape estimation. In Conference on Computer Vision and Pattern Recognition, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.745, + 0.892, + 0.786 + ], + "angle": 0, + "content": "[29] P. Li, Y. Xu, Y. Wei, and Y. Yang. Self-Correction for Human Parsing. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(6):3260-3271, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.788, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[30] R. Li, M. Zheng, S. Karanam, T. Chen, and Z. Wu. Everybody Is Unique: Towards Unbiased Human Mesh Recovery. In British Machine Vision Conference, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.831, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[31] R. Li, B. Guillard, E. Remelli, and P. Fua. DIG: Draping Implicit Garment over the Human Body. In Asian Conference on Computer Vision, 2022. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[32] Ren Li, Benoit Guillard, and Pascal Fua. ISP: Multi-Layered Garment Draping with Implicit Sewing Patterns. In Ad-" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1594" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.119 + ], + "angle": 0, + "content": "vances in Neural Information Processing Systems, 2023. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.177 + ], + "angle": 0, + "content": "[33] L. Liu, L. Zhang, Y. Xu, C. Gotsman, and S. J. Gortler. A Local/Global Approach to Mesh Parameterization. In Proceedings of the Symposium on Geometry Processing, pages 1495-1504, 2008. 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.22 + ], + "angle": 0, + "content": "[34] X. Liu, J. Li, and G. Lu. Modeling Realistic Clothing from a Single Image under Normal Guide. IEEE Transactions on Visualization and Computer Graphics, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.222, + 0.469, + 0.262 + ], + "angle": 0, + "content": "[35] Z. Liu, H. Mao, C. Wu, C. Feichtenhofer, T. Darrell, and S. Xie. A ConvNet for the 2020s. In Conference on Computer Vision and Pattern Recognition, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.469, + 0.304 + ], + "angle": 0, + "content": "[36] M. Loper, N. Mahmood, J. Romero, G. Pons-Moll, and M.J. Black. SMPL: A Skinned Multi-Person Linear Model. ACM SIGGRAPH Asia, 34(6), 2015. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.307, + 0.469, + 0.36 + ], + "angle": 0, + "content": "[37] Qianli Ma, Jinlong Yang, Michael J. Black, and Siyu Tang. Neural Point-based Shape Modeling of Humans in Challenging Clothing. In International Conference on 3D Vision, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.364, + 0.469, + 0.418 + ], + "angle": 0, + "content": "[38] G. Moon and K.M. Lee. I2L-MeshNet: Image-to-Lixel Prediction Network for Accurate 3D Human Pose and Mesh Estimation from a Single RGB Image. In European Conference on Computer Vision, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.421, + 0.469, + 0.461 + ], + "angle": 0, + "content": "[39] G. Moon, H. Nam, T. Shiratori, and K.M. Lee. 3d clothed human reconstruction in the wild. In European Conference on Computer Vision, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.464, + 0.469, + 0.504 + ], + "angle": 0, + "content": "[40] R. Narain, A. Samii, and J.F. O'brien. Adaptive anisotropic remeshing for cloth simulation. ACM Transactions on Graphics, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.507, + 0.469, + 0.56 + ], + "angle": 0, + "content": "[41] M. Omran, C. Lassner, G. Pons-Moll, P. Gehler, and B. Schiele. Neural Body Fitting: Unifying Deep Learning and Model-Based Human Pose and Shape Estimation. In International Conference on 3D Vision, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.563, + 0.469, + 0.617 + ], + "angle": 0, + "content": "[42] N. Pietroni, C. Dumery, R. Falque, M. Liu, T. Vidal-Calleja, and O. Sorkine-Hornung. Computational pattern making from 3D garment models. ACM Transactions on Graphics, 41(4):1-14, 2022. 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.62, + 0.469, + 0.674 + ], + "angle": 0, + "content": "[43] N. Rahaman, A. Baratin, D. Arpit, F. Draxler, M. Lin, F. Hamprecht, Y. Bengio, and A. Courville. On the spectral bias of neural networks. In International Conference on Machine Learning, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.469, + 0.73 + ], + "angle": 0, + "content": "[44] S. Ramasinghe and S. Lucey. Beyond Periodicity: Towards a Unifying Framework for Activations in Coordinate-MLPs. In European Conference on Computer Vision, pages 142–158, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.469, + 0.787 + ], + "angle": 0, + "content": "[45] S. Saito, Z. Huang, R. Natsume, S. Morishima, A. Kanazawa, and H. Li. PIFu: Pixel-Aligned Implicit Function for High-Resolution Clothed Human Digitization. In International Conference on Computer Vision, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[46] S. Saito, T. Simon, J. Saragih, and H. Joo. Pifuuhd: Multi-Level Pixel-Aligned Implicit Function for High-Resolution 3D Human Digitization. In Conference on Computer Vision and Pattern Recognition, pages 84–93, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.847, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[47] I. Santesteban, N. Thuerey, M. A. Otaduy, and D. Casas. Self-Supervised Collision Handling via Generative 3D Garment Models for Virtual Try-On. In Conference on Computer Vision and Pattern Recognition, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[48] I. Santesteban, M.A. Otaduy, and D. Casas. SNUG: Self-Supervised Neural Dynamic Garments. In Conference on Computer Vision and Pattern Recognition, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.163 + ], + "angle": 0, + "content": "[49] O. Sorkine-Hornung and M. Rabinovich. Least-squares rigid motion using svd. Computing, 1(1):1-5, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.165, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[50] A. Srivastava, C. Pokhariya, S. Jinka, and A. Sharma. xCloth: Extracting Template-free Textured 3D Clothes from a Monocular Image. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2504-2512, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.276 + ], + "angle": 0, + "content": "[51] D. Ulyanov, A. Vedaldi, and V. Lempitsky. Deep Image Prior. In Conference on Computer Vision and Pattern Recognition, pages 9446-9454, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[52] Y. Xiu, J. Yang, D. Tzionas, and M. J Black. Icon: Implicit clothed humans obtained from normals. In Conference on Computer Vision and Pattern Recognition, pages 13286-13296, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[53] Y. Xiu, J. Yang, X. Cao, D. Tzionas, and M. J Black. ECON: Explicit Clothed humans Optimized via Normal integration. In Conference on Computer Vision and Pattern Recognition, pages 512-523, 2023. 1, 2, 6, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[54] F. Yang, R. Li, G. Georgakis, S. Karanam, T. Chen, H. Ling, and Z. Wu. Robust multi-modal 3d patient body modeling. In Conference on Medical Image Computing and Computer Assisted Intervention, pages 86-95, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.448, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[55] S. Yang, Z. Pan, T. Amert, K. Wang, L. Yu, T. Berg, and M. Lin. Physics-inspired garment recovery from a single-view image. ACM Transactions on Graphics, 37(5):1-14, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.491, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[56] I. Zakharkin, K. Mazur, A. Grigorev, and V. Lempitsky. Point-based modeling of human clothing. In International Conference on Computer Vision, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.533, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[57] Z. Zheng, T. Yu, Y. Wei, Q. Dai, and Y. Liu. Deephuman: 3d human reconstruction from a single image. In International Conference on Computer Vision, pages 7739-7749, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.577, + 0.892, + 0.631 + ], + "angle": 0, + "content": "[58] Z. Zheng, T. Yu, Y. Liu, and Q. Dai. Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(6):3170-3184, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.633, + 0.892, + 0.674 + ], + "angle": 0, + "content": "[59] B. Zhou, X. Chen, Q. Fu, K. Guo, and P. Tan. Garment modeling from a single image. In Computer Graphics Forum, pages 85–91, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.676, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[60] H. Zhu, Y. Cao, H. Jin, W. Chen, D. Du, Z. Wang, S. Cui, and X. Han. Deep fashion3d: A dataset and benchmark for 3d garment reconstruction from single images. In European Conference on Computer Vision, pages 512-530, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.733, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[61] H. Zhu, L. Qiu, Y. Qiu, and X. Han. Registering explicit to implicit: Towards high-fidelity garment mesh reconstruction from single images. In Conference on Computer Vision and Pattern Recognition, pages 3845-3854, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.787 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1595" + } + ] +] \ No newline at end of file diff --git a/2024/Garment Recovery with Shape and Deformation Priors/83e37ee5-1c15-40f2-ac5d-8e3141ed428d_origin.pdf b/2024/Garment Recovery with Shape and Deformation Priors/83e37ee5-1c15-40f2-ac5d-8e3141ed428d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..72ab713340b26c2aa698f7390d9f27d64fa74563 --- /dev/null +++ b/2024/Garment Recovery with Shape and Deformation Priors/83e37ee5-1c15-40f2-ac5d-8e3141ed428d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:953d0ec99d2b8c25d1f6386e6825606a43d616e29fa700d725b09083f46a5c22 +size 2940960 diff --git a/2024/Garment Recovery with Shape and Deformation Priors/full.md b/2024/Garment Recovery with Shape and Deformation Priors/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4177059e12c420bc235594ee1f0458867c1d24c0 --- /dev/null +++ b/2024/Garment Recovery with Shape and Deformation Priors/full.md @@ -0,0 +1,327 @@ +# Garment Recovery with Shape and Deformation Priors + +Ren Li Corentin Dumery Benoit Guillard Pascal Fua CVLab,EPFL + +{name.surname}@epfl.ch + +![](images/95a9792aeb4fd3cb1239f78ae40ab874c16733acbb8f0c4be1f2d1ec0a8a3629.jpg) +Figure 1. We propose a fitting method that leverages shape and deformation priors trained on synthetic data to recover realistic 3D garment mesh from in-the-wild images. We produce triangulated meshes that are directly usable for animation and simulation. + +# Abstract + +While modeling people wearing tight-fitting clothing has made great strides in recent years, loose-fitting clothing remains a challenge. We propose a method that delivers realistic garment models from real-world images, regardless of garment shape or deformation. To this end, we introduce a fitting approach that utilizes shape and deformation priors learned from synthetic data to accurately capture garment shapes and deformations, including large ones. Not only does our approach recover the garment geometry accurately, it also yields models that can be directly used by downstream applications such as animation and simulation. + +# 1. Introduction + +3D clothed human recovery aims to reconstruct the body shape, pose, and clothing of many different people from images. It is key to applications such as fashion design, virtual try-on, 3D avatars, along with virtual and augmented reality. Recent years have seen tremendous progress in modeling people wearing tight-fitting clothing both in terms of body poses [6, 14, 21, 22, 25, 26, 28, 30, 38, 41, 54] and 3D shape of the clothes [5, 9-11, 20, 31, 32, 39]. However, loose-fitting clothing remains a challenge. Existing approaches either rely on mesh templates with limited generality, or produce models expressed in terms of 3D point clouds [37] or as a single watertight mesh that tightly binds the body + +and garment together [53], neither of which is straightforward to integrate into downstream applications. + +In this paper, we propose a method that overcomes these limitations and can effectively recover the shape of loose fitting garments from single images. The recovered garments can then be animated without any additional processing, as shown in Fig. 1. Starting from the Implicit Sewing Patterns (ISP) model [32] that represents garments in terms of a set of individual 2D panels and 3D surfaces associated to these panels, we introduce a deformation model that we apply to the 3D surfaces so that they can deviate substantially from the body shape. These deformations are conditioned on normals estimated from an input image of the target garment. They are learned from synthetic mesh data featuring loose clothing, where the deformations are taken to be those required to fit individual ISP 3D surfaces to the ground-truth 3D meshes. + +Given the trained deformation model, we designed a two-stage fitting process to recover the 3D garment from in-the-wild images. First, the parameters of the pre-trained deformation model are optimized to produce a shape that minimizes the distance between garment outlines and segmented garment regions, the differences between garment normals and the normals estimated in the images by off-the-shelf-algorithms [46, 53], and a physics-based loss to promote physical plausibility of the results. Then, fine local details are recovered by directly optimizing the vertex positions of the reconstructed mesh with the same loss. Our fitting process does not require external 3D annotations, other + +than the estimated normals of the target garment. + +We demonstrate that our method can recover garments that go from tight- to loose-fitting and outperforms existing approaches [9, 11, 20, 32, 39, 53] in terms of reconstruction accuracy. Furthermore, our reconstructed meshes are directly usable for virtual try-on or animation, unlike [37, 53]. Our implementation and model weights are available at https://github.com/liren2515/GarmentRecovery. + +# 2. Related Work + +Before the advent of Deep Learning, garment shape recovery from images depended mostly on user defined outlines and shape-from-shading techniques [59]. Since then, data-driven techniques have become dominant. + +Tight-Fitting Clothing. The majority of methods developed in recent years focus on clothing that clings relatively closely to the body. These can be classified into two main categories. + +In the first category, are methods that model garments as surfaces that are distinct from the body surface and interact with it. DeepGarment [10], MGN [5], and BCNet [20] train neural networks on synthetic images to predict the vertex positions of specific mesh templates. [8] and [34] leverage normal estimation to optimize the vertex position of the template and recover wrinkle details. Such methods are inherently limited in the range of shapes they can handle, and those being trained on synthetic data can easily fail when facing real images. To overcome these limitations, SMPLicit [9], DIG [31], and ClothWild [39] leverage Signed Distance Functions (SDF) to recover a wide array of garment meshes from RGB images and the corresponding segmentation masks. However, to represent non-watertight garment surfaces using an SDF, one has to wrap around them a watertight surface with a minimum thickness, which reduces accuracy. This can be addressed by using Unsigned Distance Functions (UDFs) instead [11, 15] but creates robustness issues: if the UDF is even slightly inaccurate, the value of the surface is never exactly zero and holes can appear in the reconstructed models. In our experience, the Implicit Sewing Patterns (ISP) model of [32] effectively addresses the issues of generality, accuracy, and robustness. The garments consist of flat 2D panels whose boundary is defined by a 2D SDF. To each panel is associated a 3D surface parameterized by the 2D panel coordinates. Hence, different articles of clothing are represented in a standardized way, which allows the recovery of various garments from single images. This is why we choose it as the basis for our approach. + +In the second category, are the many methods that represent body and garment using a single model. For example, in [19, 57] a volumetric regression network yields a voxel + +representation of 3D clothed humans given a single image. Other works [3, 16, 45, 46] employ a pixel-aligned implicit function that defines 3D occupancy fields or signed distance fields for clothed humans. In [1, 2], displacement vectors or UV maps are used to represent deviations from a SMPL parametric body model [36]. Similarly, in [17, 18, 52, 58] parametric body models are combined with implicit representations to achieve robustness to pose changes. While effective, all these methods suffer from significant limitations, because they cannot separate the surface of the garment from that of the body, and they are at a disadvantage when it comes to modeling loose garments whose motion can be relatively independent from the body. + +Loose-Fitting Clothing. There is a more limited number of methods designed to handle free-flowing garments. Some recent works [55, 60, 61] rely on complex physics simulation steps or feature line estimation to align the surface reconstruction with the input image. However, their dependence on garment templates limits their generality, in the same way it did for other template-based methods discussed above. Point-based methods that can reconstruct generic clothes have been proposed [50, 56] to overcome this. Unfortunately, point clouds are not straightforward to integrate into downstream applications. As a result, the method of [50] resorts to modified Poisson Surface Reconstruction (PSR) to create a garment surface from the point cloud, which can result in incorrect geometry. Another point-based representation is introduced in [37]. While being successful at modeling and animating humans wearing loose garments, it also relies on PSR to infer the mesh from the point cloud, yielding a single mesh that represents body and garment jointly. Furthermore, [37] does not explore how this representation can be fitted to images. + +ECON [53] is a method specifically designed for clothed human recovery from images. By leveraging techniques such as normal integration and shape completion, it achieves visually appealing results for individuals wearing loose clothing. However, as [37], ECON produces a single watertight mesh that tightly binds the body and garment together, precluding easy use for applications such as cloth simulation and re-animation. + +# 3. Method + +Given an image of a clothed person and a body model extracted from it using existing techniques, our goal is to recover accurate 3D models of the garments matching the image. To this end, we add to the Implicit Sewing Pattern (ISP) garment model [32], which provides us with a shape prior for garment in its rest state, a deformation model that allows us to recover its potentially large deformations, as illustrated by Fig. 2. As a result, whereas the original ISP, like most current clothes-recovery algorithms, is limited to + +![](images/751dad8a40f9c353c2654b0b7e8d6d1d462ccdd694d384c348e0fc10f9a2abc4.jpg) +Figure 2. Framework. Given an image, (1) we first estimate the normal map of the target garment and the SMPL body parameters $(\beta, \theta)$ , which are used to compute the body part segmentation and position maps. (2) The maximum coverage garment shape $\bar{\mathcal{M}}$ is then skinned to closely fit to the body, yielding $\mathcal{M}$ . Leveraging (3) pixel-aligned image features, our deformation model (4) predicts occupancy and position maps to correct $\mathcal{M}$ for large deformations. (5) The 3D garment mesh is recovered using ISP and further refined. + +tight-fitting clothing, our approach can handle both tight- and loose-fitting garments, such as skirts and open jackets. + +In this section, we first describe briefly the ISP model upon which we build our approach. We then introduce the deformation model that underpins the main contribution of this paper, going from tight fitting clothes to loosely fitting and free flowing ones. Finally, we present our approach to fitting this model to real-world images. + +# 3.1. ISP Garment Model + +ISP is a garment model inspired by the sewing patterns that fashion designers use to represent clothes. A sewing pattern is made of several 2D panels along with information about how to stitch them into a complete garment. ISP implicitly models patterns using a 2D signed distance field and a 2D label field. + +Formalization. Given the latent code $\mathbf{z}$ of a garment and a point $\mathbf{u}$ in the 2D UV space $\Omega = [-1,1]^2$ of a 2D panel of that garment, ISP outputs the signed distance $s$ to the panel boundary and a label $l$ as + +$$ +(s, l) = \mathcal {I} _ {\Theta} (\mathbf {u}, \mathbf {z}), \tag {1} +$$ + +where $\mathcal{I}_{\Theta}$ is a fully connected network. The zero crossing of the SDF defines the shape of the panel, with $s < 0$ indicating that $\mathbf{u}$ is within the panel and $s > 0$ indicating that $\mathbf{u}$ is outside the panel. The stitch information is encoded in $l$ , where panel boundaries with the same label should be stitched together. To transform the 2D sewing patterns into + +3D surfaces, a UV parameterization function $\mathcal{A}_{\Phi}$ is learned to perform the 2D-to-3D mapping + +$$ +\mathbf {X} = \mathcal {A} _ {\Phi} (\mathbf {u}, \mathbf {z}), \tag {2} +$$ + +where $\mathbf{X} \in \mathbb{R}^3$ represents the 3D position of $\mathbf{u}$ . Essentially, ISP registers each garment onto a unified 2D space $\Omega$ , and represents it using UV maps that record 2D-to-3D mapping, as shown in Fig. 3(b). Given the paired 2D sewing patterns and their 3D meshes, the pattern parameterization network $\mathcal{I}_{\Theta}$ and the UV parameterization network $\mathcal{A}_{\Phi}$ are trained by minimizing the losses + +$$ +L _ {\mathcal {I}} = L _ {S D F} + L _ {C E} + \left\| \mathbf {z} \right\| _ {2} ^ {2}, \tag {3} +$$ + +$$ +L _ {\mathcal {A}} = L _ {M S E} + L _ {\text {c o n s i s t}}, \tag {4} +$$ + +respectively. $L_{SDF}$ is the mean absolute error for the predicted SDF value $s$ , $L_{CE}$ is the cross-entropy loss for the predicted label $l$ , $L_{MSE}$ is the mean squared error of the predicted 3D position $\mathbf{X}$ , and $L_{consist}$ is the loss to reduce the gap between the front and back panels. More details can be found in [32]. + +Training. Training ISP requires the 2D sewing patterns of 3D garments in a rest state, that is, draped on a T-posed neutral body. However, patterns are absent in most large scale garment datasets, including CLOTH3D [4]. We use the garment flattening algorithm of [42] to generate the required 2D patterns. The garment mesh is first divided into several pieces given predefined cutting rules. These pieces + +![](images/1a9d6399310b40c3c791d39ec14b692fff3f23438cbe26cf5d379baa695f8b55.jpg) +Figure 3. Cutting and flattening. (a) The front (top) and back (bottom) surfaces after cutting. (b) The flattened panels and UV maps generated by ISP for (a). (c) The maximum-coverage UV maps and its represented 3D shape. + +![](images/563b4d353c8fcbc230500c43fdf75b433791452abcc1142940fe5cb335915269.jpg) + +![](images/2baa67f8e87409764137dc0da8127f966c221578180e84921cdc99a86402806c.jpg) + +are then unfolded into 2D panels by minimizing an as-rigid-as-possible [33] energy, ensuring local area preservation between the 3D and 2D parameterizations. While doing this, we constrain the boundary vertices at places such as waist and sleeves to have a constant value along a specific axis. This enhances pattern consistency across the dataset. Fig. 3 illustrates this and shows the panels generated for a shirt. We provide more details in the supplementary material. + +We generate a front and a back panels as the sewing pattern for each garment in our dataset (shirt, skirt, and trousers). Once ISP has been trained on these, we compute the maximum-coverage UV maps $\bar{\mathcal{M}}$ over the UV maps of each garment for each category, as shown in Fig. 3(c). The values of $\bar{\mathcal{M}}$ are taken to be + +$$ +\bar {\mathcal {M}} [ u, v ] = \frac {\sum_ {i} \mathbb {1} _ {s _ {\mathbf {u}} ^ {i} \leq 0} \cdot m _ {i} [ u , v ]}{\sum_ {i} \mathbb {1} _ {s _ {\mathbf {u}} ^ {i} \leq 0}}, \tag {5} +$$ + +where $m_{i}$ is the UV maps of garment $i$ , $\mathbb{1}$ is the indicator function, $[\cdot ,\cdot ]$ denotes array addressing, and $s_{\mathbf{u}}^{i}$ is the SDF value of ISP at $\mathbf{u} = (u,v)$ . The maximum coverage map $\bar{\mathcal{M}}$ encompasses information from all the patterns in the dataset. It represents the smallest possible map that covers all garments in a category, with the 3D position of each $uv$ -point $\mathbf{u}$ being the average of all garments that include $\mathbf{u}$ . We use it as a prototype for a garment category and to compute an initial guess of the deformed garment given the body pose, as discussed below. + +**Skinning.** As in many prior work [11, 31, 32], we use SMPL [36] to parameterize the body in terms of shape and pose parameters $(\beta, \theta)$ , and its extended skinning procedure for the 3D volume around the body to initially deform the 3D shape represented as $\bar{\mathcal{M}}$ . More specifically, given a 2D point $\mathbf{u} = (u, v)$ in UV space $\Omega$ , we get its actual 3D position as $\bar{\mathbf{X}}_{\mathbf{u}} = \bar{\mathcal{M}}[u, v]$ . We then deform it by computing + +$$ +\mathbf {X} _ {\mathbf {u}} = W \left(\mathbf {X} _ {(\beta , \theta)}, \beta , \theta , w (\bar {\mathbf {X}} _ {\mathbf {u}}) \mathcal {W}\right), \tag {6} +$$ + +$$ +\mathbf {X} _ {(\beta , \theta)} = \bar {\mathbf {X}} _ {\mathbf {u}} + w (\bar {\mathbf {X}} _ {\mathbf {u}}) B _ {\beta} + w (\bar {\mathbf {X}} _ {\mathbf {u}}) B _ {\theta}, +$$ + +where $W(\cdot)$ is the SMPL skinning function with skinning weights $\mathcal{W} \in \mathbb{R}^{N_B \times 24}$ , with $N_B$ being the number of vertices of the SMPL body mesh, and $B_\beta \in \mathbb{R}^{N_B \times 3}$ and $B_\theta \in \mathbb{R}^{N_B \times 3}$ are the shape and pose displacements of SMPL, respectively. The diffused weights $w(\cdot) \in \mathbb{R}^{N_B}$ are computed by a neural network, which generalizes the SMPL skinning to any point in 3D space. By repeating this for all points in a panel, we obtain the vertex position map $\mathcal{M}[u,v] = \mathbf{X}_{\mathbf{u}}$ . + +# 3.2. Modeling Large Deformations + +As shown in Fig. 2, the ISP approach described above generates a garment prototype $\mathcal{M}$ that closely fits the underlying body. To a point $\mathbf{u} = (u,v)$ that belongs to a given garment panel, it associates the vertex position $\mathbf{X}_{\mathbf{u}} = \mathcal{M}[u,v]$ of Eq. 6, which is usually relatively close to the body. To model loose clothing, we now need to compute a potentially larger displacement $\Delta \mathbf{X}_{\mathbf{u}}$ to be added to $\mathbf{X}_{\mathbf{u}}$ . + +To evaluate $\Delta \mathbf{X}_{\mathbf{u}}$ we train a network $\mathcal{D}$ that consists of an MLP to estimate the occupancy value and the corrective displacement value, and two CNNs to extract image features $F_{n}$ and $F_{b}$ from the image of normals and the segmentation and vertex position images of the SMPL body, respectively. Fig. 2 depicts this architecture. We obtain the pixel-aligned image feature for $\mathbf{u}$ by computing + +$$ +F \left(\mathbf {x} _ {\mathbf {u}}\right) = F _ {n} \oplus F _ {b} \left(\mathbf {x} _ {\mathbf {u}}\right), \tag {7} +$$ + +$$ +\mathbf {x _ {u}} = P (\mathbf {X _ {u}}) , +$$ + +where $P(\cdot)$ denotes the projection into the image and $\oplus$ concatenation. The MLP takes as input $\mathbf{u}$ , its 3D position and image features to predict the occupancy $O_{\mathbf{u}} \in \{0,1\}$ and the corrective displacement $\Delta \mathbf{X}_{\mathbf{u}}$ . By assembling the results of each point in the UV space, we obtain the final occupancy maps $\mathcal{O}$ and the vertex position maps $\hat{\mathcal{M}}$ , where $\hat{\mathcal{M}}[u,v] = \mathbf{X}_{\mathbf{u}} + \Delta \mathbf{X}_{\mathbf{u}}$ . In essence, $\mathcal{O}$ is the binarized SDF of the garment 2D panels, which implicitly defines the garment shape and geometry in the rest state, while $\hat{\mathcal{M}}$ encodes the deformed state for that garment. The 3D garment mesh + +can be recovered from these 2D maps with ISP, as discussed below. + +We use a single network $\mathcal{D}$ for the front and the back panels, and encode the points $\mathbf{u}$ of the front as $[u,v,+1]$ and those of the back as $[u,v,-1]$ . We learn the parameters of $\mathcal{D}$ by minimizing + +$$ +L = \sum_ {\mathbf {u} \in \Omega} | | \mathbf {X} _ {\mathbf {u}} + \Delta \mathbf {X} _ {\mathbf {u}} - \tilde {\mathbf {X}} _ {\mathbf {u}} | | _ {2} + \lambda \sum_ {\mathbf {u} \in \Omega} B E (O _ {\mathbf {u}}, \tilde {O} _ {\mathbf {u}}), \tag {8} +$$ + +where $\tilde{\mathbf{X}}_{\mathbf{u}}$ and $\tilde{O}_{\mathbf{u}}$ are the ground truth vertex positions and occupancy values, $BE(\cdot)$ is the binary cross entropy, and $\lambda$ is a weighting constant. + +From 2D maps to 3D mesh. The occupancy value $O_{\mathbf{u}}$ indicates whether $\mathbf{u}$ falls within the panels of the garment. To convert the generated 2D occupancy maps $\mathcal{O}$ into a 3D garment mesh in its rest shape, we need to recover an ISP latent code $\mathbf{z}$ as defined in Eq. 1. To this end, we find the vector $\mathbf{z}$ such that the corresponding SDF of ISP best matches the produced occupancy maps by minimizing + +$$ +\mathbf {z} ^ {*} = \underset {\mathbf {z}} {\operatorname {a r g m i n}} \sum_ {\mathbf {u} \in \Omega_ {-}} R \left(s _ {\mathbf {u}} (\mathbf {z})\right) + \sum_ {\mathbf {u} \in \Omega_ {+}} R \left(- s _ {\mathbf {u}} (\mathbf {z})\right) + \lambda_ {\mathbf {z}} \| \mathbf {z} \| _ {2}, \tag {9} +$$ + +where $\Omega_{-} = \{\mathbf{u}|O_{\mathbf{u}} = 1,\mathbf{u}\in \Omega \}$ , $\Omega_{+} = \{\mathbf{u}|O_{\mathbf{u}} = 0,\mathbf{u}\in \Omega \}$ , $R(\cdot)$ is the ReLU function, $s_{\mathbf{u}}(\mathbf{z})$ is the SDF value of $\mathbf{u}$ computed by ISP, and $\lambda_{\mathbf{z}}$ is the weighting constant. With $\mathbf{z}^*$ , the rest garment mesh is inferred through ISP's meshing and sewing process. The deformed garment mesh can be obtained by simply replacing the vertex position of the recovered mesh with the values stored in $\hat{\mathcal{M}}$ . This yields 3D garment meshes in both the rest and deformed states as shown in the top-right of Fig. 2, which are required for the application such as cloth simulation and can be used for further refinement as discussed below. + +# 3.3. Fitting the Models to Images + +For practical reasons, the range of garment materials, external forces, and body motions present in the training data is limited. As a result, given in-the-wild images as input, the trained model can produce inaccurate results as shown in Fig. 4(b). To remedy this and to leverage the deformation prior that the network $\mathcal{D}$ captures, we refine the result by minimizing a loss function with respect to the pretrained deformation parameters of $\mathcal{D}$ as in [13, 21, 51]. $L$ is designed to promote a good match between the garment mesh and image observations. We take it to be + +$$ +L = \lambda_ {C} L _ {C D} + \lambda_ {n} L _ {\text {n o r m a l}} + \lambda_ {p} L _ {\text {p h y s i c s}}, \tag {10} +$$ + +$$ +L _ {C D} = d \left(x _ {c} ^ {f}, x _ {I _ {n}}\right), \tag {11} +$$ + +$$ +L _ {\text {n o r m a l}} = \sum_ {i \in f} 1 - \cos \left(n ^ {i}, I _ {n} \left(x _ {c} ^ {i}\right)\right), \tag {12} +$$ + +$$ +L _ {\text {p h y s i c s}} = L _ {\text {s t r a i n}} + L _ {\text {b e n d}} + L _ {\text {g r a v i t y}} + L _ {\text {c o l}}, \tag {13} +$$ + +![](images/5346d41e7e76a011f27ff50cb66777b03bbedfce4702431c284fc30e2c64c5b6.jpg) +(a) + +![](images/a5f9a37d13f87c2eb6001c2531b4222c42f3a93e1ec14548fdbdbe3620642b56.jpg) +(b) + +![](images/efd48e2fc2c792631ebcbf5c5ad46423b275e817bf956c97732261ddab285c74.jpg) +(c) +Figure 4. Fitting results. Given (a) the normal estimation of an in-the-wild image, (b) is the inference result with ISP recovered geometry. (c) is obtained by optimizing the parameters of the pretrained deformation model. Further refinement of the mesh vertex positions yields (d). + +![](images/08b64de78b05e863885fd64134115ba86f91f72b6ff20c8a404cc02ec95ec968.jpg) +(d) + +where $x_{c}^{f}$ is the 2D projection of the centers of visible faces $f$ after mesh rasterization, $x_{I_n}$ denotes the coordinates of foreground pixels, $n^i$ is the normal of face $i$ , $I_{n}(x_{c}^{i})$ is the normal image values at $x_{c}^{i}$ , and $\lambda_{C}$ , $\lambda_{n}$ and $\lambda_{p}$ are the balancing scalars. $d(\cdot)$ and $\cos (\cdot)$ are the functions measuring the 2D Chamfer Distance and the cosine similarity, respectively. $L_{\mathrm{physics}}$ is a physics-based loss derived from [40, 48], which computes the membrane strain energy $L_{strain}$ caused by the deformation, the bending energy $L_{bend}$ resulting from the folding of adjacent faces, the gravitational potential energy $L_{gravity}$ and the penalty for body-garment collision $L_{col}$ . Minimizing $L_{CD}$ induces an external force of stretching or compression on the garment mesh to align its 2D projection with the given image, while minimizing $L_{\mathrm{physics}}$ ensures that the mesh exhibits physically plausible deformation adhering to the shape constraints of the rest-state mesh recovered by ISP. + +Minimizing the loss $L$ with respect to the deformation parameters yields a mesh whose overall shape matches the input image, as illustrated in Fig. 4(c). However, since neural networks tend to learn low-frequency functions [43], the result might be too smooth. To recover fine surface details, we perform a refinement step by minimizing $L$ directly with respect to the coordinates of the garment mesh vertices. This generates realistic local details, such as wrinkles on the surface, as shown in Fig. 4(d). + +# 4. Experiments + +In short, our method begins by inferring the shape and the deformation of the garment in terms of the occupancy and position maps. Leveraging the shape prior of ISP, we then + +recover the garment geometry from the occupancy maps and deform it using the position maps. Next, we refine the initial deformed mesh to better align with image observations by fine-tuning the pre-trained network $\mathcal{D}$ that captures the deformation prior. Finally, we recover fine details through vertex-level optimization of the garment mesh. In this section, we demonstrate the effectiveness of this process and compare it to that of other state-of-the-art methods. + +# 4.1. Implementation Details + +Following the implementation of [32], both the pattern parameterization model $\mathcal{I}_{\Theta}$ and the UV parameterization model $\mathcal{A}_{\Phi}$ of ISP have two separate MLPs for the front and back panels. Each MLP has 7 layers with Softplus activations. The dimension of latent code $\mathbf{z}$ is 32. The skinning weight model $w$ is a 9-layer MLP with leaky ReLU activations, whose output is normalized by a final Softmax layer. $\mathcal{I}_{\Theta}$ and $\mathcal{A}_{\Phi}$ are trained jointly for 9000 iterations with a batch size of 50. $w$ is trained with the same parameters as [32]. For the image feature extraction, we use two separate ConvNeXt [35] networks to extract multi-scale garment and body features of sizes 96, 192, 384, 768, which are concatenated as the final features. The point UV coordinates $\mathbf{u}$ , 3D position $\mathbf{X_u}$ , and image features $F(\mathbf{x_u})$ are projected separately to 384 dimensions by three linear layers, which are then concatenated as the input of the MLP of the deformation model. The MLP of the deformation model has 10 layers with a skip connection from the input layer to the middle, and uses Gaussian functions as the activation layer following [44]. The CNNs and MLP of the deformation model are trained jointly for 40 epochs with the Adam optimizer [23] and a learning rate of $10^{-4}$ . For real images, we use [53] and [12] to obtain their normal and SMPL body parameter estimations, respectively. The garment segmentation masks are generated by leveraging the segmentation of SAM [24] and the semantic labels of [29]. + +# 4.2. Dataset, Evaluation Metrics, and Baseline + +Our models are trained on CLOTH3D [4], which is a synthetic dataset with motion sequences of 3D clothed human. It contains garment in the rest state, and deformed states caused by the motion of underlying body. For each garment, it has a single simulated sequence up to 10 seconds. It covers a large variety of garment in different shapes, types and topologies. For the training of ISP, we randomly select 400 shirts, 200 skirts and 200 pairs of trousers, and generate their sewing patterns by the method described in Sec. 3.1. The deformation model is trained on the corresponding simulated sequences. For each frame, we render 11 normal images for the garment mesh with random rotations around the Y-axis, which produces 40K, 40K and 20K training images for shirt, skirt and trousers respectively. During training, we augment the data with image flipping and rotation. + +To evaluate the garment reconstruction quality, we use the Chamfer Distance (CD) between the ground truth and the recovered garment mesh, and Intersection over Union (IoU) between the ground truth mask and the rendered mask of reconstructed garment mesh. + +We compare our method against state-of-the-art methods BCNet [20], SMPLicit [9], ClothWild [39], DrapeNet [11] and ISP [32]. SMPLicit, DrapeNet and ISP use the garment segmentation mask for reconstruction, while BCNet and ClothWild take the RGB images as input. + +# 4.3. Comparison with State-of-the-Art Methods + +
CD (×103) ↓SkirtShirtTrousersIoU ↑SkirtShirtTrousers
SMPLicit8.053.490.81SMPLicit0.6570.5270.782
DrapeNetn/a1.540.84DrapeNetn/a0.7150.845
ISP7.262.020.91ISP0.7090.6650.781
Ours2.870.440.28Ours0.9400.9390.946
Ours-GT1.670.310.23Ours-GT0.9530.9420.954
+ +Table 1. Quantitative comparisons. Our method outperforms SMPLicit, DrapeNet, and ISP in terms of CD and IoU on all three garment categories, as shown in the second-to-last row of both tables (Ours). These results were obtained using normals estimated from the images using [53]. In the last row (Ours-GT), we provide the results we obtained using ground-truth normals instead. + +Quantitative Results. Due to the absence of publicly available real dataset for evaluating garment reconstruction, we utilize synthetic data consisting of 30 unseen shirts, 30 unseen skirts, and 30 pairs of unseen trousers from CLOTH3D for quantitative evaluation. We compare our method with SMPLicit, DrapeNet, and ISP, which recover garment meshes from segmentation masks. The segmentation masks required by the baselines are rendered from the ground truth mesh data, while the normal images required by our method are estimated with [53]. As shown in Tab. 1, our approach significantly outperforms the baselines in terms of Chamfer Distance (CD) and Intersection over Union (IoU) across all garment categories. + +For comparison purposes, we re-ran our algorithm using the ground-truth normals and report the results in the last row of Tab. 1. As expected, this leads to an improvement in reconstruction accuracy, but the increase is only modest. This highlights the robustness of our approach to the slight inaccuracies that can be expected from a normal-estimation algorithm. + +Qualitative Results. Fig. 5 shows the qualitative comparison for the results reconstructed from in-the-wild images. Since BCNet is trained only on synthetic RGB data, it is not able to predict accurate body and garment results. Both SMPLicit and ClothWild generate unrealistic watertight meshes that do not accurately represent real garments. + +![](images/2f0fb335fe2dea224549700b3f523ff6976a06bda21bbe3ddd9544024b086391.jpg) +Figure 5. Comparison against SOTA methods. From left to right, we show the input image and the 3D garment meshes recovered by our method and SOTA methods: BCNet, SMPLcit, ClothWild, DrapeNet, ISP. (Since skirt is unavailable for DrapeNet, a random pair of trousers is put on its result in the second row.) + +
CD (×103) ↓GTEst.
(-D*,-v*)2.043.58
(-D*,+v*)1.803.39
(+D*,-v*)1.752.93
(+D*,+v*)1.672.87
+ +
IoU ↑GTEst.
(-D*,-v*)0.8480.751
(-D*,+v*)0.9470.904
(+D*,-v*)0.9410.923
(+D*,+v*)0.9530.940
+ +Table 2. Ablation study. GT and Est. mean using ground-truth and estimated normal images as input, respectively. $+ / - {\mathcal{D}}^{ * }$ denote with/without the finetuning of the deformation model. $+ / - {v}^{ * }$ denote with/without the optimization of vertex positions. + +While DrapeNet and ISP are able to recover open surfaces for garments, their results remain closely adhered to the body, similar to SMPLlicit and ClothWild. In contrast, our method can faithfully recover garment mesh from input images, as clearly observed in the depiction of the open jacket and the flowing long skirt that stand away from the body. + +# 4.4. Ablation Study + +Table 2 presents the evaluation results of our method with and without fitting on the test set of skirts. The results of row 1, 3 and 4 indicate that optimizing the parameters of the pretrained deformation network $(+D^{*}, - v^{*})$ improves the quality of the raw inference results, and optimizing the vertex positions $(+D^{*}, + v^{*})$ further enhances the outcome, + +![](images/7478b15f133d96704f58df0b2f2bcbb1666c21e39f9fa5c8d7d6457986c66f81.jpg) +Figure 6. Fitting strategy comparison between (a) our fitting method and (b) only vertex optimization. + +![](images/01769d869b91a12282eb349c458d95ff8d6aeb1fbee41764ffecfa781005fd17.jpg) +(a) + +![](images/ba54e37c27845e8ac3e46b1829ec92bcb5a36c730677597d999e9ceeb686e7fc.jpg) +(b) + +yielding the lowest CD and highest IoU. + +However, as demonstrated in the second row, directly optimizing the vertex positions on the raw inference results without optimizing the deformation model $(-\mathcal{D}^{*}, + v^{*})$ is not as effective. The deformation model encapsulates a continuous deformation field. When its weights are optimized based on partial observations, the entire field undergoes modification, thereby influencing all mesh vertices globally. On the contrary, a direct vertex optimization with partial observations predominantly affects the mesh vertices locally. While this can capture localized details like wrinkles, it struggles to resolve discrepancies in the overall shape. Fig. 6 displays the results of our methods and an ablation with + +![](images/a84fb222c2ae35f5f1b04f7b7c0ef7c7e718caba24661b33ebab34379c1b9cfa.jpg) +Figure 7. Recovering from in-the-wild images. Our method is able to recover realistic meshes for garments with diverse shapes and deformations. + +vertex optimization only. Unsurprisingly, the latter fails at recovering the correct shape and produces implausible deformation on the mesh surface. + +# 4.5. More Results + +Fig. 7 shows a collection of results reconstructed by our method from in-the-wild images. Our method can produce realistic 3D meshes with fine details across a wide range of garment types, from tight-fitting attire to more relaxed and flowing outfits. + +# 5. Conclusion + +We have presented a novel approach to recovering realistic 3D garment meshes from in-the-wild images featuring loose fitting clothing. It relies on a fitting process that imposes shape and deformation priors learned on synthetic data to accurately capture garment shape and deformations. In future work, we will extend our approach to modeling deformations over time from video sequences while enforcing temporal consistency of the reconstructions. + +Acknowledgement. This project was supported in part by the Swiss National Science Foundation. + +# References + +[1] T. Alldieck, M. Magnor, B. Bhatnagar, C. Theobalt, and G. Pons-Moll. Learning to reconstruct people in clothing from a single RGB camera. In Conference on Computer Vision and Pattern Recognition, pages 1175-1186, 2019. 2 +[2] T. Alldieck, G. Pons-Moll, C. Theobalt, and M. Magnor. Tex2shape: Detailed full human body geometry from a single image. In International Conference on Computer Vision, pages 2293-2303, 2019. 2 +[3] T. Alldieck, M. Zanfir, and C. Sminchisescu. Photorealistic monocular 3d reconstruction of humans wearing clothing. In Conference on Computer Vision and Pattern Recognition, pages 1506-1515, 2022. 2 +[4] H. Bertiche, M. Madadi, and S. Escalera. CLOTH3D: Clothed 3D Humans. In European Conference on Computer Vision, pages 344-359, 2020. 3, 6, 1 +[5] B. L. Bhatnagar, G. Tiwari, C. Theobalt, and G. Pons-Moll. Multi-Garment Net: Learning to Dress 3D People from Images. In International Conference on Computer Vision, 2019. 1, 2 +[6] F. Bogo, A. Kanazawa, C. Lassner, P. Gehler, J. Romero, and M. J. Black. Keep It SMPL: Automatic Estimation of 3D Human Pose and Shape from a Single Image. In European Conference on Computer Vision, 2016. 1 +[7] R. Brouet, A. Sheffer, L. Boissieux, and M.P. Cani. Design Preserving Garment Transfer. ACM Transactions on Graphics, 31(4):361-3611, 2012. 1 +[8] A. Casado-Elvira, M. Trinidad, and D. Casas. PERGAMO: Personalized 3d garments from monocular video. In Computer Graphics Forum, pages 293–304, 2022. 2 +[9] E. Corona, A. Pumarola, G. Alenya, G. Pons-Moll, and F. Moreno-Noguer. SMPlicit: Topology-Aware Generative Model for Clothed People. In Conference on Computer Vision and Pattern Recognition, 2021. 1, 2, 6 +[10] R. Danerek, E. Dibra, C. Öztireli, R. Ziegler, and M. Gross. Deepgarment : 3D Garment Shape Estimation from a Single Image. Eurographics, 2017. 2 +[11] Luca DeLuigi, Ren Li, Benoit Guillard, Mathieu Salzmann, and Pascal Fua. DrapeNet: Generating Garments and Draping them with Self-Supervision. In Conference on Computer Vision and Pattern Recognition, 2023. 1, 2, 4, 6 +[12] Y. Feng, V. Choutas, T. Bolkart, D. Tzionas, and M. J. Black. Collaborative Regression of Expressive Bodies using Moderation. In International Conference on Computer Vision, 2021. 6 +[13] M. Gadelha, R. Wang, and S. Maji. Deep manifold prior. In International Conference on Computer Vision, pages 1107-1116, 2021. 5 +[14] G. Georgakis, R. Li and S. Karanam, T. Chen, J. Košecka, and Z. Wu. Hierarchical kinematic human mesh recovery. In European Conference on Computer Vision, pages 768-784, 2020. 1 +[15] B. Guillard, F. Stella, and P. Fua. MeshUDF: Fast and Differentiable Meshing of Unsigned Distance Field Networks. In European Conference on Computer Vision, 2022. 2 +[16] T. He, J. Collomosse, H. Jin, and S. Soatto. Geo-pifu: Geometry and pixel aligned implicit functions for single-view + +human reconstruction. Advances in Neural Information Processing Systems, 33:9276-9287, 2020. 2 +[17] T. He, Y. Xu, S. Saito, S. Soatto, and T. Tung. Arch++: Animation-ready clothed human reconstruction revisited. In International Conference on Computer Vision, pages 11046-11056, 2021. 2 +[18] Z. Huang, Y. Xu, C. Lassner, H. Li, and T. Tung. Arch: Animatable Reconstruction of Clothed Humans. In Conference on Computer Vision and Pattern Recognition, pages 3093-3102, 2020. 2 +[19] A. Jackson, C. Manafas, and G. Tzimiropoulos. 3d human body reconstruction from a single image via volumetric regression. In European Conference on Computer Vision Workshops, 2018. 2 +[20] B. Jiang, J. Zhang, Y. Hong, J. Luo, L. Liu, and H. Bao. Bcnet: Learning body and cloth shape from a single image. In European Conference on Computer Vision, 2020. 1, 2, 6 +[21] H. Joo, N. Neverova, and A. Vedaldi. Exemplar Fine-Tuning for 3D Human Pose Fitting Towards In-the-Wild 3D Human Pose Estimation. In International Conference on 3D Vision, 2020. 1, 5 +[22] A. Kanazawa, M. J. Black, D. W. Jacobs, and J. Malik. End-To-End Recovery of Human Shape and Pose. In Conference on Computer Vision and Pattern Recognition, 2018. 1 +[23] D. P. Kingma and J. Ba. Adam: A Method for Stochastic Optimisation. In International Conference on Learning Representations, 2015. 6 +[24] A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. Berg, W. Lo, P. Dollar, and R. Girshick. Segment anything. In International Conference on Computer Vision, 2023. 6 +[25] N. Kolotouros, G. Pavlakos, M. J. Black, and K. Daniilidis. Learning to Reconstruct 3D Human Pose and Shape via Model-Fitting in the Loop. In Conference on Computer Vision and Pattern Recognition, 2019. 1 +[26] C. Lassner, J. Romero, M. Kiefel, F. Bogo, M.J. Black, and P.V. Gehler. Unite the People: Closing the Loop Between 3D and 2D Human Representations. In Conference on Computer Vision and Pattern Recognition, 2017. 1 +[27] B. Lévy, S. Petitjean, N. Ray, and J. Maillot. Least Squares Conformal Maps for Automatic Texture Atlas Generation. In ACM SIGGRAPH, 2002. 1 +[28] J. Li, C. Xu, Z. Chen, S. Bian, L. Yang, and C. Lu. Hybrid: A hybrid analytical-neural inverse kinematics solution for 3d human pose and shape estimation. In Conference on Computer Vision and Pattern Recognition, 2021. 1 +[29] P. Li, Y. Xu, Y. Wei, and Y. Yang. Self-Correction for Human Parsing. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(6):3260-3271, 2020. 6 +[30] R. Li, M. Zheng, S. Karanam, T. Chen, and Z. Wu. Everybody Is Unique: Towards Unbiased Human Mesh Recovery. In British Machine Vision Conference, 2021. 1 +[31] R. Li, B. Guillard, E. Remelli, and P. Fua. DIG: Draping Implicit Garment over the Human Body. In Asian Conference on Computer Vision, 2022. 1, 2, 4 +[32] Ren Li, Benoit Guillard, and Pascal Fua. ISP: Multi-Layered Garment Draping with Implicit Sewing Patterns. In Ad- + +vances in Neural Information Processing Systems, 2023. 1, 2, 3, 4, 6 +[33] L. Liu, L. Zhang, Y. Xu, C. Gotsman, and S. J. Gortler. A Local/Global Approach to Mesh Parameterization. In Proceedings of the Symposium on Geometry Processing, pages 1495-1504, 2008. 4, 1 +[34] X. Liu, J. Li, and G. Lu. Modeling Realistic Clothing from a Single Image under Normal Guide. IEEE Transactions on Visualization and Computer Graphics, 2023. 2 +[35] Z. Liu, H. Mao, C. Wu, C. Feichtenhofer, T. Darrell, and S. Xie. A ConvNet for the 2020s. In Conference on Computer Vision and Pattern Recognition, 2022. 6 +[36] M. Loper, N. Mahmood, J. Romero, G. Pons-Moll, and M.J. Black. SMPL: A Skinned Multi-Person Linear Model. ACM SIGGRAPH Asia, 34(6), 2015. 2, 4 +[37] Qianli Ma, Jinlong Yang, Michael J. Black, and Siyu Tang. Neural Point-based Shape Modeling of Humans in Challenging Clothing. In International Conference on 3D Vision, 2022. 1, 2 +[38] G. Moon and K.M. Lee. I2L-MeshNet: Image-to-Lixel Prediction Network for Accurate 3D Human Pose and Mesh Estimation from a Single RGB Image. In European Conference on Computer Vision, 2020. 1 +[39] G. Moon, H. Nam, T. Shiratori, and K.M. Lee. 3d clothed human reconstruction in the wild. In European Conference on Computer Vision, 2022. 1, 2, 6 +[40] R. Narain, A. Samii, and J.F. O'brien. Adaptive anisotropic remeshing for cloth simulation. ACM Transactions on Graphics, 2012. 5 +[41] M. Omran, C. Lassner, G. Pons-Moll, P. Gehler, and B. Schiele. Neural Body Fitting: Unifying Deep Learning and Model-Based Human Pose and Shape Estimation. In International Conference on 3D Vision, 2018. 1 +[42] N. Pietroni, C. Dumery, R. Falque, M. Liu, T. Vidal-Calleja, and O. Sorkine-Hornung. Computational pattern making from 3D garment models. ACM Transactions on Graphics, 41(4):1-14, 2022. 3, 1 +[43] N. Rahaman, A. Baratin, D. Arpit, F. Draxler, M. Lin, F. Hamprecht, Y. Bengio, and A. Courville. On the spectral bias of neural networks. In International Conference on Machine Learning, 2019. 5 +[44] S. Ramasinghe and S. Lucey. Beyond Periodicity: Towards a Unifying Framework for Activations in Coordinate-MLPs. In European Conference on Computer Vision, pages 142–158, 2022. 6 +[45] S. Saito, Z. Huang, R. Natsume, S. Morishima, A. Kanazawa, and H. Li. PIFu: Pixel-Aligned Implicit Function for High-Resolution Clothed Human Digitization. In International Conference on Computer Vision, 2019. 2 +[46] S. Saito, T. Simon, J. Saragih, and H. Joo. Pifuuhd: Multi-Level Pixel-Aligned Implicit Function for High-Resolution 3D Human Digitization. In Conference on Computer Vision and Pattern Recognition, pages 84–93, 2020. 1, 2 +[47] I. Santesteban, N. Thuerey, M. A. Otaduy, and D. Casas. Self-Supervised Collision Handling via Generative 3D Garment Models for Virtual Try-On. In Conference on Computer Vision and Pattern Recognition, 2021. 1 + +[48] I. Santesteban, M.A. Otaduy, and D. Casas. SNUG: Self-Supervised Neural Dynamic Garments. In Conference on Computer Vision and Pattern Recognition, 2022. 5 +[49] O. Sorkine-Hornung and M. Rabinovich. Least-squares rigid motion using svd. Computing, 1(1):1-5, 2017. 1 +[50] A. Srivastava, C. Pokhariya, S. Jinka, and A. Sharma. xCloth: Extracting Template-free Textured 3D Clothes from a Monocular Image. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2504-2512, 2022. 2 +[51] D. Ulyanov, A. Vedaldi, and V. Lempitsky. Deep Image Prior. In Conference on Computer Vision and Pattern Recognition, pages 9446-9454, 2018. 5 +[52] Y. Xiu, J. Yang, D. Tzionas, and M. J Black. Icon: Implicit clothed humans obtained from normals. In Conference on Computer Vision and Pattern Recognition, pages 13286-13296, 2022. 2 +[53] Y. Xiu, J. Yang, X. Cao, D. Tzionas, and M. J Black. ECON: Explicit Clothed humans Optimized via Normal integration. In Conference on Computer Vision and Pattern Recognition, pages 512-523, 2023. 1, 2, 6, 3, 4 +[54] F. Yang, R. Li, G. Georgakis, S. Karanam, T. Chen, H. Ling, and Z. Wu. Robust multi-modal 3d patient body modeling. In Conference on Medical Image Computing and Computer Assisted Intervention, pages 86-95, 2020. 1 +[55] S. Yang, Z. Pan, T. Amert, K. Wang, L. Yu, T. Berg, and M. Lin. Physics-inspired garment recovery from a single-view image. ACM Transactions on Graphics, 37(5):1-14, 2018. 2 +[56] I. Zakharkin, K. Mazur, A. Grigorev, and V. Lempitsky. Point-based modeling of human clothing. In International Conference on Computer Vision, 2021. 2 +[57] Z. Zheng, T. Yu, Y. Wei, Q. Dai, and Y. Liu. Deephuman: 3d human reconstruction from a single image. In International Conference on Computer Vision, pages 7739-7749, 2019. 2 +[58] Z. Zheng, T. Yu, Y. Liu, and Q. Dai. Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(6):3170-3184, 2021. 2 +[59] B. Zhou, X. Chen, Q. Fu, K. Guo, and P. Tan. Garment modeling from a single image. In Computer Graphics Forum, pages 85–91, 2013. 2 +[60] H. Zhu, Y. Cao, H. Jin, W. Chen, D. Du, Z. Wang, S. Cui, and X. Han. Deep fashion3d: A dataset and benchmark for 3d garment reconstruction from single images. In European Conference on Computer Vision, pages 512-530, 2020. 2 +[61] H. Zhu, L. Qiu, Y. Qiu, and X. Han. Registering explicit to implicit: Towards high-fidelity garment mesh reconstruction from single images. In Conference on Computer Vision and Pattern Recognition, pages 3845-3854, 2022. 2 \ No newline at end of file diff --git a/2024/Garment Recovery with Shape and Deformation Priors/images.zip b/2024/Garment Recovery with Shape and Deformation Priors/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..863a5b7682d08e5cb14f858689d275a5b2934417 --- /dev/null +++ b/2024/Garment Recovery with Shape and Deformation Priors/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0ed67e40bc3647c9283388c04dac55a6833b959dc70bd16e2646d777456f6cc3 +size 657488 diff --git a/2024/Garment Recovery with Shape and Deformation Priors/layout.json b/2024/Garment Recovery with Shape and Deformation Priors/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..582ada5c737071ef5f2b7ce9c8847979d2843010 --- /dev/null +++ b/2024/Garment Recovery with Shape and Deformation Priors/layout.json @@ -0,0 +1,9595 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 126, + 103, + 468, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 103, + 468, + 121 + ], + "spans": [ + { + "bbox": [ + 126, + 103, + 468, + 121 + ], + "type": "text", + "content": "Garment Recovery with Shape and Deformation Priors" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 150, + 143, + 443, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 143, + 443, + 171 + ], + "spans": [ + { + "bbox": [ + 150, + 143, + 443, + 171 + ], + "type": "text", + "content": "Ren Li Corentin Dumery Benoit Guillard Pascal Fua CVLab,EPFL" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 236, + 174, + 357, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 174, + 357, + 185 + ], + "spans": [ + { + "bbox": [ + 236, + 174, + 357, + 185 + ], + "type": "text", + "content": "{name.surname}@epfl.ch" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 55, + 196, + 545, + 329 + ], + "blocks": [ + { + "bbox": [ + 55, + 196, + 545, + 329 + ], + "lines": [ + { + "bbox": [ + 55, + 196, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 55, + 196, + 545, + 329 + ], + "type": "image", + "image_path": "95a9792aeb4fd3cb1239f78ae40ab874c16733acbb8f0c4be1f2d1ec0a8a3629.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 332, + 546, + 354 + ], + "lines": [ + { + "bbox": [ + 46, + 332, + 546, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 332, + 546, + 354 + ], + "type": "text", + "content": "Figure 1. We propose a fitting method that leverages shape and deformation priors trained on synthetic data to recover realistic 3D garment mesh from in-the-wild images. We produce triangulated meshes that are directly usable for animation and simulation." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 361, + 192, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 361, + 192, + 374 + ], + "spans": [ + { + "bbox": [ + 143, + 361, + 192, + 374 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 388, + 288, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 288, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 288, + 521 + ], + "type": "text", + "content": "While modeling people wearing tight-fitting clothing has made great strides in recent years, loose-fitting clothing remains a challenge. We propose a method that delivers realistic garment models from real-world images, regardless of garment shape or deformation. To this end, we introduce a fitting approach that utilizes shape and deformation priors learned from synthetic data to accurately capture garment shapes and deformations, including large ones. Not only does our approach recover the garment geometry accurately, it also yields models that can be directly used by downstream applications such as animation and simulation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 548, + 128, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 548, + 128, + 560 + ], + "spans": [ + { + "bbox": [ + 47, + 548, + 128, + 560 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "content": "3D clothed human recovery aims to reconstruct the body shape, pose, and clothing of many different people from images. It is key to applications such as fashion design, virtual try-on, 3D avatars, along with virtual and augmented reality. Recent years have seen tremendous progress in modeling people wearing tight-fitting clothing both in terms of body poses [6, 14, 21, 22, 25, 26, 28, 30, 38, 41, 54] and 3D shape of the clothes [5, 9-11, 20, 31, 32, 39]. However, loose-fitting clothing remains a challenge. Existing approaches either rely on mesh templates with limited generality, or produce models expressed in terms of 3D point clouds [37] or as a single watertight mesh that tightly binds the body" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 363, + 545, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 363, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 305, + 363, + 545, + 387 + ], + "type": "text", + "content": "and garment together [53], neither of which is straightforward to integrate into downstream applications." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 388, + 546, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 546, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 546, + 567 + ], + "type": "text", + "content": "In this paper, we propose a method that overcomes these limitations and can effectively recover the shape of loose fitting garments from single images. The recovered garments can then be animated without any additional processing, as shown in Fig. 1. Starting from the Implicit Sewing Patterns (ISP) model [32] that represents garments in terms of a set of individual 2D panels and 3D surfaces associated to these panels, we introduce a deformation model that we apply to the 3D surfaces so that they can deviate substantially from the body shape. These deformations are conditioned on normals estimated from an input image of the target garment. They are learned from synthetic mesh data featuring loose clothing, where the deformations are taken to be those required to fit individual ISP 3D surfaces to the ground-truth 3D meshes." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": "Given the trained deformation model, we designed a two-stage fitting process to recover the 3D garment from in-the-wild images. First, the parameters of the pre-trained deformation model are optimized to produce a shape that minimizes the distance between garment outlines and segmented garment regions, the differences between garment normals and the normals estimated in the images by off-the-shelf-algorithms [46, 53], and a physics-based loss to promote physical plausibility of the results. Then, fine local details are recovered by directly optimizing the vertex positions of the reconstructed mesh with the same loss. Our fitting process does not require external 3D annotations, other" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1586" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 245, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 245, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 245, + 83 + ], + "type": "text", + "content": "than the estimated normals of the target garment." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 84, + 286, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 84, + 286, + 179 + ], + "spans": [ + { + "bbox": [ + 46, + 84, + 286, + 179 + ], + "type": "text", + "content": "We demonstrate that our method can recover garments that go from tight- to loose-fitting and outperforms existing approaches [9, 11, 20, 32, 39, 53] in terms of reconstruction accuracy. Furthermore, our reconstructed meshes are directly usable for virtual try-on or animation, unlike [37, 53]. Our implementation and model weights are available at https://github.com/liren2515/GarmentRecovery." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 190, + 134, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 190, + 134, + 201 + ], + "spans": [ + { + "bbox": [ + 47, + 190, + 134, + 201 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 209, + 286, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 209, + 286, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 286, + 258 + ], + "type": "text", + "content": "Before the advent of Deep Learning, garment shape recovery from images depended mostly on user defined outlines and shape-from-shading techniques [59]. Since then, data-driven techniques have become dominant." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 270, + 286, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 286, + 317 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 286, + 317 + ], + "type": "text", + "content": "Tight-Fitting Clothing. The majority of methods developed in recent years focus on clothing that clings relatively closely to the body. These can be classified into two main categories." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 318, + 286, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 318, + 286, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 318, + 286, + 675 + ], + "type": "text", + "content": "In the first category, are methods that model garments as surfaces that are distinct from the body surface and interact with it. DeepGarment [10], MGN [5], and BCNet [20] train neural networks on synthetic images to predict the vertex positions of specific mesh templates. [8] and [34] leverage normal estimation to optimize the vertex position of the template and recover wrinkle details. Such methods are inherently limited in the range of shapes they can handle, and those being trained on synthetic data can easily fail when facing real images. To overcome these limitations, SMPLicit [9], DIG [31], and ClothWild [39] leverage Signed Distance Functions (SDF) to recover a wide array of garment meshes from RGB images and the corresponding segmentation masks. However, to represent non-watertight garment surfaces using an SDF, one has to wrap around them a watertight surface with a minimum thickness, which reduces accuracy. This can be addressed by using Unsigned Distance Functions (UDFs) instead [11, 15] but creates robustness issues: if the UDF is even slightly inaccurate, the value of the surface is never exactly zero and holes can appear in the reconstructed models. In our experience, the Implicit Sewing Patterns (ISP) model of [32] effectively addresses the issues of generality, accuracy, and robustness. The garments consist of flat 2D panels whose boundary is defined by a 2D SDF. To each panel is associated a 3D surface parameterized by the 2D panel coordinates. Hence, different articles of clothing are represented in a standardized way, which allows the recovery of various garments from single images. This is why we choose it as the basis for our approach." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": "In the second category, are the many methods that represent body and garment using a single model. For example, in [19, 57] a volumetric regression network yields a voxel" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": "representation of 3D clothed humans given a single image. Other works [3, 16, 45, 46] employ a pixel-aligned implicit function that defines 3D occupancy fields or signed distance fields for clothed humans. In [1, 2], displacement vectors or UV maps are used to represent deviations from a SMPL parametric body model [36]. Similarly, in [17, 18, 52, 58] parametric body models are combined with implicit representations to achieve robustness to pose changes. While effective, all these methods suffer from significant limitations, because they cannot separate the surface of the garment from that of the body, and they are at a disadvantage when it comes to modeling loose garments whose motion can be relatively independent from the body." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 241, + 545, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 241, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 241, + 545, + 479 + ], + "type": "text", + "content": "Loose-Fitting Clothing. There is a more limited number of methods designed to handle free-flowing garments. Some recent works [55, 60, 61] rely on complex physics simulation steps or feature line estimation to align the surface reconstruction with the input image. However, their dependence on garment templates limits their generality, in the same way it did for other template-based methods discussed above. Point-based methods that can reconstruct generic clothes have been proposed [50, 56] to overcome this. Unfortunately, point clouds are not straightforward to integrate into downstream applications. As a result, the method of [50] resorts to modified Poisson Surface Reconstruction (PSR) to create a garment surface from the point cloud, which can result in incorrect geometry. Another point-based representation is introduced in [37]. While being successful at modeling and animating humans wearing loose garments, it also relies on PSR to infer the mesh from the point cloud, yielding a single mesh that represents body and garment jointly. Furthermore, [37] does not explore how this representation can be fitted to images." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 481, + 545, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 481, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 304, + 481, + 545, + 576 + ], + "type": "text", + "content": "ECON [53] is a method specifically designed for clothed human recovery from images. By leveraging techniques such as normal integration and shape completion, it achieves visually appealing results for individuals wearing loose clothing. However, as [37], ECON produces a single watertight mesh that tightly binds the body and garment together, precluding easy use for applications such as cloth simulation and re-animation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 586, + 361, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 361, + 597 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 361, + 597 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Given an image of a clothed person and a body model extracted from it using existing techniques, our goal is to recover accurate 3D models of the garments matching the image. To this end, we add to the Implicit Sewing Pattern (ISP) garment model [32], which provides us with a shape prior for garment in its rest state, a deformation model that allows us to recover its potentially large deformations, as illustrated by Fig. 2. As a result, whereas the original ISP, like most current clothes-recovery algorithms, is limited to" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1587" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 71, + 541, + 297 + ], + "blocks": [ + { + "bbox": [ + 53, + 71, + 541, + 297 + ], + "lines": [ + { + "bbox": [ + 53, + 71, + 541, + 297 + ], + "spans": [ + { + "bbox": [ + 53, + 71, + 541, + 297 + ], + "type": "image", + "image_path": "751dad8a40f9c353c2654b0b7e8d6d1d462ccdd694d384c348e0fc10f9a2abc4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 300, + 546, + 346 + ], + "lines": [ + { + "bbox": [ + 46, + 300, + 546, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 546, + 346 + ], + "type": "text", + "content": "Figure 2. Framework. Given an image, (1) we first estimate the normal map of the target garment and the SMPL body parameters " + }, + { + "bbox": [ + 46, + 300, + 546, + 346 + ], + "type": "inline_equation", + "content": "(\\beta, \\theta)" + }, + { + "bbox": [ + 46, + 300, + 546, + 346 + ], + "type": "text", + "content": ", which are used to compute the body part segmentation and position maps. (2) The maximum coverage garment shape " + }, + { + "bbox": [ + 46, + 300, + 546, + 346 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{M}}" + }, + { + "bbox": [ + 46, + 300, + 546, + 346 + ], + "type": "text", + "content": " is then skinned to closely fit to the body, yielding " + }, + { + "bbox": [ + 46, + 300, + 546, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 300, + 546, + 346 + ], + "type": "text", + "content": ". Leveraging (3) pixel-aligned image features, our deformation model (4) predicts occupancy and position maps to correct " + }, + { + "bbox": [ + 46, + 300, + 546, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 300, + 546, + 346 + ], + "type": "text", + "content": " for large deformations. (5) The 3D garment mesh is recovered using ISP and further refined." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 358, + 287, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 358, + 287, + 382 + ], + "spans": [ + { + "bbox": [ + 46, + 358, + 287, + 382 + ], + "type": "text", + "content": "tight-fitting clothing, our approach can handle both tight- and loose-fitting garments, such as skirts and open jackets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 382, + 288, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 382, + 288, + 455 + ], + "spans": [ + { + "bbox": [ + 46, + 382, + 288, + 455 + ], + "type": "text", + "content": "In this section, we first describe briefly the ISP model upon which we build our approach. We then introduce the deformation model that underpins the main contribution of this paper, going from tight fitting clothes to loosely fitting and free flowing ones. Finally, we present our approach to fitting this model to real-world images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 460, + 165, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 460, + 165, + 472 + ], + "spans": [ + { + "bbox": [ + 47, + 460, + 165, + 472 + ], + "type": "text", + "content": "3.1. ISP Garment Model" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 479, + 287, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 287, + 551 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 287, + 551 + ], + "type": "text", + "content": "ISP is a garment model inspired by the sewing patterns that fashion designers use to represent clothes. A sewing pattern is made of several 2D panels along with information about how to stitch them into a complete garment. ISP implicitly models patterns using a 2D signed distance field and a 2D label field." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 565, + 288, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 565, + 288, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 565, + 288, + 613 + ], + "type": "text", + "content": "Formalization. Given the latent code " + }, + { + "bbox": [ + 46, + 565, + 288, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 46, + 565, + 288, + 613 + ], + "type": "text", + "content": " of a garment and a point " + }, + { + "bbox": [ + 46, + 565, + 288, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 46, + 565, + 288, + 613 + ], + "type": "text", + "content": " in the 2D UV space " + }, + { + "bbox": [ + 46, + 565, + 288, + 613 + ], + "type": "inline_equation", + "content": "\\Omega = [-1,1]^2" + }, + { + "bbox": [ + 46, + 565, + 288, + 613 + ], + "type": "text", + "content": " of a 2D panel of that garment, ISP outputs the signed distance " + }, + { + "bbox": [ + 46, + 565, + 288, + 613 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 565, + 288, + 613 + ], + "type": "text", + "content": " to the panel boundary and a label " + }, + { + "bbox": [ + 46, + 565, + 288, + 613 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 565, + 288, + 613 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 127, + 621, + 287, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 621, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 127, + 621, + 287, + 635 + ], + "type": "interline_equation", + "content": "(s, l) = \\mathcal {I} _ {\\Theta} (\\mathbf {u}, \\mathbf {z}), \\tag {1}", + "image_path": "4215009b0f6576f64bc156fed785e70a2117eb7cb8bfc986ee8c0f513855547b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\Theta}" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": " is a fully connected network. The zero crossing of the SDF defines the shape of the panel, with " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "s < 0" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": " indicating that " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": " is within the panel and " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "s > 0" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": " indicating that " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": " is outside the panel. The stitch information is encoded in " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": ", where panel boundaries with the same label should be stitched together. To transform the 2D sewing patterns into" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 358, + 546, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 358, + 546, + 383 + ], + "spans": [ + { + "bbox": [ + 305, + 358, + 546, + 383 + ], + "type": "text", + "content": "3D surfaces, a UV parameterization function " + }, + { + "bbox": [ + 305, + 358, + 546, + 383 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\Phi}" + }, + { + "bbox": [ + 305, + 358, + 546, + 383 + ], + "type": "text", + "content": " is learned to perform the 2D-to-3D mapping" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 391, + 392, + 546, + 405 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 392, + 546, + 405 + ], + "spans": [ + { + "bbox": [ + 391, + 392, + 546, + 405 + ], + "type": "interline_equation", + "content": "\\mathbf {X} = \\mathcal {A} _ {\\Phi} (\\mathbf {u}, \\mathbf {z}), \\tag {2}", + "image_path": "30c8069364095fe89728157caec8832fabc18a3f04d931939afe7e0d9baea140.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 413, + 546, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 546, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 546, + 498 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 413, + 546, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 413, + 546, + 498 + ], + "type": "text", + "content": " represents the 3D position of " + }, + { + "bbox": [ + 304, + 413, + 546, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 304, + 413, + 546, + 498 + ], + "type": "text", + "content": ". Essentially, ISP registers each garment onto a unified 2D space " + }, + { + "bbox": [ + 304, + 413, + 546, + 498 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 304, + 413, + 546, + 498 + ], + "type": "text", + "content": ", and represents it using UV maps that record 2D-to-3D mapping, as shown in Fig. 3(b). Given the paired 2D sewing patterns and their 3D meshes, the pattern parameterization network " + }, + { + "bbox": [ + 304, + 413, + 546, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\Theta}" + }, + { + "bbox": [ + 304, + 413, + 546, + 498 + ], + "type": "text", + "content": " and the UV parameterization network " + }, + { + "bbox": [ + 304, + 413, + 546, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\Phi}" + }, + { + "bbox": [ + 304, + 413, + 546, + 498 + ], + "type": "text", + "content": " are trained by minimizing the losses" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 362, + 506, + 545, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 506, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 362, + 506, + 545, + 520 + ], + "type": "interline_equation", + "content": "L _ {\\mathcal {I}} = L _ {S D F} + L _ {C E} + \\left\\| \\mathbf {z} \\right\\| _ {2} ^ {2}, \\tag {3}", + "image_path": "961dc15d22682a190d91a0c54c0260ba2f34e5b21c3f9033dc32f6e17c810307.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 362, + 521, + 545, + 534 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 521, + 545, + 534 + ], + "spans": [ + { + "bbox": [ + 362, + 521, + 545, + 534 + ], + "type": "interline_equation", + "content": "L _ {\\mathcal {A}} = L _ {M S E} + L _ {\\text {c o n s i s t}}, \\tag {4}", + "image_path": "f221885dc7b103dd275f4ee7357143f6c931ea8da7cf2eb6251607425dcbcc57.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 544, + 545, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 545, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 545, + 615 + ], + "type": "text", + "content": "respectively. " + }, + { + "bbox": [ + 304, + 544, + 545, + 615 + ], + "type": "inline_equation", + "content": "L_{SDF}" + }, + { + "bbox": [ + 304, + 544, + 545, + 615 + ], + "type": "text", + "content": " is the mean absolute error for the predicted SDF value " + }, + { + "bbox": [ + 304, + 544, + 545, + 615 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 544, + 545, + 615 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 544, + 545, + 615 + ], + "type": "inline_equation", + "content": "L_{CE}" + }, + { + "bbox": [ + 304, + 544, + 545, + 615 + ], + "type": "text", + "content": " is the cross-entropy loss for the predicted label " + }, + { + "bbox": [ + 304, + 544, + 545, + 615 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 544, + 545, + 615 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 544, + 545, + 615 + ], + "type": "inline_equation", + "content": "L_{MSE}" + }, + { + "bbox": [ + 304, + 544, + 545, + 615 + ], + "type": "text", + "content": " is the mean squared error of the predicted 3D position " + }, + { + "bbox": [ + 304, + 544, + 545, + 615 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 304, + 544, + 545, + 615 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 544, + 545, + 615 + ], + "type": "inline_equation", + "content": "L_{consist}" + }, + { + "bbox": [ + 304, + 544, + 545, + 615 + ], + "type": "text", + "content": " is the loss to reduce the gap between the front and back panels. More details can be found in [32]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 629, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 546, + 714 + ], + "type": "text", + "content": "Training. Training ISP requires the 2D sewing patterns of 3D garments in a rest state, that is, draped on a T-posed neutral body. However, patterns are absent in most large scale garment datasets, including CLOTH3D [4]. We use the garment flattening algorithm of [42] to generate the required 2D patterns. The garment mesh is first divided into several pieces given predefined cutting rules. These pieces" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1588" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 68, + 214, + 198 + ], + "blocks": [ + { + "bbox": [ + 75, + 68, + 214, + 198 + ], + "lines": [ + { + "bbox": [ + 75, + 68, + 214, + 198 + ], + "spans": [ + { + "bbox": [ + 75, + 68, + 214, + 198 + ], + "type": "image", + "image_path": "1a9d6399310b40c3c791d39ec14b692fff3f23438cbe26cf5d379baa695f8b55.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 205, + 547, + 228 + ], + "lines": [ + { + "bbox": [ + 46, + 205, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 547, + 228 + ], + "type": "text", + "content": "Figure 3. Cutting and flattening. (a) The front (top) and back (bottom) surfaces after cutting. (b) The flattened panels and UV maps generated by ISP for (a). (c) The maximum-coverage UV maps and its represented 3D shape." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 216, + 68, + 372, + 198 + ], + "blocks": [ + { + "bbox": [ + 216, + 68, + 372, + 198 + ], + "lines": [ + { + "bbox": [ + 216, + 68, + 372, + 198 + ], + "spans": [ + { + "bbox": [ + 216, + 68, + 372, + 198 + ], + "type": "image", + "image_path": "563b4d353c8fcbc230500c43fdf75b433791452abcc1142940fe5cb335915269.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 373, + 68, + 521, + 198 + ], + "blocks": [ + { + "bbox": [ + 373, + 68, + 521, + 198 + ], + "lines": [ + { + "bbox": [ + 373, + 68, + 521, + 198 + ], + "spans": [ + { + "bbox": [ + 373, + 68, + 521, + 198 + ], + "type": "image", + "image_path": "2baa67f8e87409764137dc0da8127f966c221578180e84921cdc99a86402806c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 248, + 287, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 248, + 287, + 344 + ], + "spans": [ + { + "bbox": [ + 46, + 248, + 287, + 344 + ], + "type": "text", + "content": "are then unfolded into 2D panels by minimizing an as-rigid-as-possible [33] energy, ensuring local area preservation between the 3D and 2D parameterizations. While doing this, we constrain the boundary vertices at places such as waist and sleeves to have a constant value along a specific axis. This enhances pattern consistency across the dataset. Fig. 3 illustrates this and shows the panels generated for a shirt. We provide more details in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 344, + 287, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 344, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 46, + 344, + 287, + 415 + ], + "type": "text", + "content": "We generate a front and a back panels as the sewing pattern for each garment in our dataset (shirt, skirt, and trousers). Once ISP has been trained on these, we compute the maximum-coverage UV maps " + }, + { + "bbox": [ + 46, + 344, + 287, + 415 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{M}}" + }, + { + "bbox": [ + 46, + 344, + 287, + 415 + ], + "type": "text", + "content": " over the UV maps of each garment for each category, as shown in Fig. 3(c). The values of " + }, + { + "bbox": [ + 46, + 344, + 287, + 415 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{M}}" + }, + { + "bbox": [ + 46, + 344, + 287, + 415 + ], + "type": "text", + "content": " are taken to be" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 99, + 417, + 287, + 446 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 417, + 287, + 446 + ], + "spans": [ + { + "bbox": [ + 99, + 417, + 287, + 446 + ], + "type": "interline_equation", + "content": "\\bar {\\mathcal {M}} [ u, v ] = \\frac {\\sum_ {i} \\mathbb {1} _ {s _ {\\mathbf {u}} ^ {i} \\leq 0} \\cdot m _ {i} [ u , v ]}{\\sum_ {i} \\mathbb {1} _ {s _ {\\mathbf {u}} ^ {i} \\leq 0}}, \\tag {5}", + "image_path": "3c6be7021fc19654ca91698c2dbdc84a5240ad2fbd58e8d6440a82a7675548b0.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "inline_equation", + "content": "m_{i}" + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "text", + "content": " is the UV maps of garment " + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\mathbb{1}" + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "text", + "content": " is the indicator function, " + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "inline_equation", + "content": "[\\cdot ,\\cdot ]" + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "text", + "content": " denotes array addressing, and " + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "inline_equation", + "content": "s_{\\mathbf{u}}^{i}" + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "text", + "content": " is the SDF value of ISP at " + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{u} = (u,v)" + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "text", + "content": ". The maximum coverage map " + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{M}}" + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "text", + "content": " encompasses information from all the patterns in the dataset. It represents the smallest possible map that covers all garments in a category, with the 3D position of each " + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "inline_equation", + "content": "uv" + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "text", + "content": "-point " + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "text", + "content": " being the average of all garments that include " + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "text", + "content": ". We use it as a prototype for a garment category and to compute an initial guess of the deformed garment given the body pose, as discussed below." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 582, + 287, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 667 + ], + "type": "text", + "content": "**Skinning.** As in many prior work [11, 31, 32], we use SMPL [36] to parameterize the body in terms of shape and pose parameters " + }, + { + "bbox": [ + 46, + 582, + 287, + 667 + ], + "type": "inline_equation", + "content": "(\\beta, \\theta)" + }, + { + "bbox": [ + 46, + 582, + 287, + 667 + ], + "type": "text", + "content": ", and its extended skinning procedure for the 3D volume around the body to initially deform the 3D shape represented as " + }, + { + "bbox": [ + 46, + 582, + 287, + 667 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{M}}" + }, + { + "bbox": [ + 46, + 582, + 287, + 667 + ], + "type": "text", + "content": ". More specifically, given a 2D point " + }, + { + "bbox": [ + 46, + 582, + 287, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{u} = (u, v)" + }, + { + "bbox": [ + 46, + 582, + 287, + 667 + ], + "type": "text", + "content": " in UV space " + }, + { + "bbox": [ + 46, + 582, + 287, + 667 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 46, + 582, + 287, + 667 + ], + "type": "text", + "content": ", we get its actual 3D position as " + }, + { + "bbox": [ + 46, + 582, + 287, + 667 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{X}}_{\\mathbf{u}} = \\bar{\\mathcal{M}}[u, v]" + }, + { + "bbox": [ + 46, + 582, + 287, + 667 + ], + "type": "text", + "content": ". We then deform it by computing" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 94, + 681, + 287, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 681, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 94, + 681, + 287, + 696 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {\\mathbf {u}} = W \\left(\\mathbf {X} _ {(\\beta , \\theta)}, \\beta , \\theta , w (\\bar {\\mathbf {X}} _ {\\mathbf {u}}) \\mathcal {W}\\right), \\tag {6}", + "image_path": "56b43daadc677b0982864ffff97b5ce0c31ebff1fab7a4efa3a6bcc2a8de4660.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 83, + 697, + 252, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 697, + 252, + 712 + ], + "spans": [ + { + "bbox": [ + 83, + 697, + 252, + 712 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {(\\beta , \\theta)} = \\bar {\\mathbf {X}} _ {\\mathbf {u}} + w (\\bar {\\mathbf {X}} _ {\\mathbf {u}}) B _ {\\beta} + w (\\bar {\\mathbf {X}} _ {\\mathbf {u}}) B _ {\\theta},", + "image_path": "bca6ef296c430792a344fdae2450c65ffb706f922b7df98f189cd80cf955af18.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 248, + 545, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 248, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 306, + 248, + 545, + 357 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 248, + 545, + 357 + ], + "type": "inline_equation", + "content": "W(\\cdot)" + }, + { + "bbox": [ + 306, + 248, + 545, + 357 + ], + "type": "text", + "content": " is the SMPL skinning function with skinning weights " + }, + { + "bbox": [ + 306, + 248, + 545, + 357 + ], + "type": "inline_equation", + "content": "\\mathcal{W} \\in \\mathbb{R}^{N_B \\times 24}" + }, + { + "bbox": [ + 306, + 248, + 545, + 357 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 306, + 248, + 545, + 357 + ], + "type": "inline_equation", + "content": "N_B" + }, + { + "bbox": [ + 306, + 248, + 545, + 357 + ], + "type": "text", + "content": " being the number of vertices of the SMPL body mesh, and " + }, + { + "bbox": [ + 306, + 248, + 545, + 357 + ], + "type": "inline_equation", + "content": "B_\\beta \\in \\mathbb{R}^{N_B \\times 3}" + }, + { + "bbox": [ + 306, + 248, + 545, + 357 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 248, + 545, + 357 + ], + "type": "inline_equation", + "content": "B_\\theta \\in \\mathbb{R}^{N_B \\times 3}" + }, + { + "bbox": [ + 306, + 248, + 545, + 357 + ], + "type": "text", + "content": " are the shape and pose displacements of SMPL, respectively. The diffused weights " + }, + { + "bbox": [ + 306, + 248, + 545, + 357 + ], + "type": "inline_equation", + "content": "w(\\cdot) \\in \\mathbb{R}^{N_B}" + }, + { + "bbox": [ + 306, + 248, + 545, + 357 + ], + "type": "text", + "content": " are computed by a neural network, which generalizes the SMPL skinning to any point in 3D space. By repeating this for all points in a panel, we obtain the vertex position map " + }, + { + "bbox": [ + 306, + 248, + 545, + 357 + ], + "type": "inline_equation", + "content": "\\mathcal{M}[u,v] = \\mathbf{X}_{\\mathbf{u}}" + }, + { + "bbox": [ + 306, + 248, + 545, + 357 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 363, + 470, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 363, + 470, + 376 + ], + "spans": [ + { + "bbox": [ + 306, + 363, + 470, + 376 + ], + "type": "text", + "content": "3.2. Modeling Large Deformations" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 381, + 545, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 381, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 304, + 381, + 545, + 465 + ], + "type": "text", + "content": "As shown in Fig. 2, the ISP approach described above generates a garment prototype " + }, + { + "bbox": [ + 304, + 381, + 545, + 465 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 381, + 545, + 465 + ], + "type": "text", + "content": " that closely fits the underlying body. To a point " + }, + { + "bbox": [ + 304, + 381, + 545, + 465 + ], + "type": "inline_equation", + "content": "\\mathbf{u} = (u,v)" + }, + { + "bbox": [ + 304, + 381, + 545, + 465 + ], + "type": "text", + "content": " that belongs to a given garment panel, it associates the vertex position " + }, + { + "bbox": [ + 304, + 381, + 545, + 465 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\mathbf{u}} = \\mathcal{M}[u,v]" + }, + { + "bbox": [ + 304, + 381, + 545, + 465 + ], + "type": "text", + "content": " of Eq. 6, which is usually relatively close to the body. To model loose clothing, we now need to compute a potentially larger displacement " + }, + { + "bbox": [ + 304, + 381, + 545, + 465 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{X}_{\\mathbf{u}}" + }, + { + "bbox": [ + 304, + 381, + 545, + 465 + ], + "type": "text", + "content": " to be added to " + }, + { + "bbox": [ + 304, + 381, + 545, + 465 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\mathbf{u}}" + }, + { + "bbox": [ + 304, + 381, + 545, + 465 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 465, + 545, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 465, + 545, + 549 + ], + "spans": [ + { + "bbox": [ + 304, + 465, + 545, + 549 + ], + "type": "text", + "content": "To evaluate " + }, + { + "bbox": [ + 304, + 465, + 545, + 549 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{X}_{\\mathbf{u}}" + }, + { + "bbox": [ + 304, + 465, + 545, + 549 + ], + "type": "text", + "content": " we train a network " + }, + { + "bbox": [ + 304, + 465, + 545, + 549 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 465, + 545, + 549 + ], + "type": "text", + "content": " that consists of an MLP to estimate the occupancy value and the corrective displacement value, and two CNNs to extract image features " + }, + { + "bbox": [ + 304, + 465, + 545, + 549 + ], + "type": "inline_equation", + "content": "F_{n}" + }, + { + "bbox": [ + 304, + 465, + 545, + 549 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 465, + 545, + 549 + ], + "type": "inline_equation", + "content": "F_{b}" + }, + { + "bbox": [ + 304, + 465, + 545, + 549 + ], + "type": "text", + "content": " from the image of normals and the segmentation and vertex position images of the SMPL body, respectively. Fig. 2 depicts this architecture. We obtain the pixel-aligned image feature for " + }, + { + "bbox": [ + 304, + 465, + 545, + 549 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 304, + 465, + 545, + 549 + ], + "type": "text", + "content": " by computing" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 374, + 558, + 545, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 558, + 545, + 571 + ], + "spans": [ + { + "bbox": [ + 374, + 558, + 545, + 571 + ], + "type": "interline_equation", + "content": "F \\left(\\mathbf {x} _ {\\mathbf {u}}\\right) = F _ {n} \\oplus F _ {b} \\left(\\mathbf {x} _ {\\mathbf {u}}\\right), \\tag {7}", + "image_path": "7649eccb0f54f80cea2d5cc4bb1c9c28797d854bced952e2571ddf7c7849a913.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 390, + 573, + 453, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 573, + 453, + 586 + ], + "spans": [ + { + "bbox": [ + 390, + 573, + 453, + 586 + ], + "type": "interline_equation", + "content": "\\mathbf {x _ {u}} = P (\\mathbf {X _ {u}}) ,", + "image_path": "63190bf32cc27b01fff3c0ca807a7b83e6caf0c93e0c9dd5445811200c21fb31.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "P(\\cdot)" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " denotes the projection into the image and " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " concatenation. The MLP takes as input " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": ", its 3D position and image features to predict the occupancy " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "O_{\\mathbf{u}} \\in \\{0,1\\}" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " and the corrective displacement " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{X}_{\\mathbf{u}}" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": ". By assembling the results of each point in the UV space, we obtain the final occupancy maps " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " and the vertex position maps " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{M}}" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{M}}[u,v] = \\mathbf{X}_{\\mathbf{u}} + \\Delta \\mathbf{X}_{\\mathbf{u}}" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": ". In essence, " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " is the binarized SDF of the garment 2D panels, which implicitly defines the garment shape and geometry in the rest state, while " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{M}}" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " encodes the deformed state for that garment. The 3D garment mesh" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1589" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": "can be recovered from these 2D maps with ISP, as discussed below." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 96, + 286, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 96, + 286, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 96, + 286, + 144 + ], + "type": "text", + "content": "We use a single network " + }, + { + "bbox": [ + 46, + 96, + 286, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 96, + 286, + 144 + ], + "type": "text", + "content": " for the front and the back panels, and encode the points " + }, + { + "bbox": [ + 46, + 96, + 286, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 46, + 96, + 286, + 144 + ], + "type": "text", + "content": " of the front as " + }, + { + "bbox": [ + 46, + 96, + 286, + 144 + ], + "type": "inline_equation", + "content": "[u,v,+1]" + }, + { + "bbox": [ + 46, + 96, + 286, + 144 + ], + "type": "text", + "content": " and those of the back as " + }, + { + "bbox": [ + 46, + 96, + 286, + 144 + ], + "type": "inline_equation", + "content": "[u,v,-1]" + }, + { + "bbox": [ + 46, + 96, + 286, + 144 + ], + "type": "text", + "content": ". We learn the parameters of " + }, + { + "bbox": [ + 46, + 96, + 286, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 96, + 286, + 144 + ], + "type": "text", + "content": " by minimizing" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 52, + 152, + 287, + 177 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 152, + 287, + 177 + ], + "spans": [ + { + "bbox": [ + 52, + 152, + 287, + 177 + ], + "type": "interline_equation", + "content": "L = \\sum_ {\\mathbf {u} \\in \\Omega} | | \\mathbf {X} _ {\\mathbf {u}} + \\Delta \\mathbf {X} _ {\\mathbf {u}} - \\tilde {\\mathbf {X}} _ {\\mathbf {u}} | | _ {2} + \\lambda \\sum_ {\\mathbf {u} \\in \\Omega} B E (O _ {\\mathbf {u}}, \\tilde {O} _ {\\mathbf {u}}), \\tag {8}", + "image_path": "32511e7e3f5664e4b7cbfa835c93f123d3545ddcc9ef58bb70eed0b6863aa928.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 186, + 287, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 186, + 287, + 223 + ], + "spans": [ + { + "bbox": [ + 46, + 186, + 287, + 223 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 186, + 287, + 223 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{X}}_{\\mathbf{u}}" + }, + { + "bbox": [ + 46, + 186, + 287, + 223 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 186, + 287, + 223 + ], + "type": "inline_equation", + "content": "\\tilde{O}_{\\mathbf{u}}" + }, + { + "bbox": [ + 46, + 186, + 287, + 223 + ], + "type": "text", + "content": " are the ground truth vertex positions and occupancy values, " + }, + { + "bbox": [ + 46, + 186, + 287, + 223 + ], + "type": "inline_equation", + "content": "BE(\\cdot)" + }, + { + "bbox": [ + 46, + 186, + 287, + 223 + ], + "type": "text", + "content": " is the binary cross entropy, and " + }, + { + "bbox": [ + 46, + 186, + 287, + 223 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 186, + 287, + 223 + ], + "type": "text", + "content": " is a weighting constant." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 236, + 287, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 236, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 287, + 315 + ], + "type": "text", + "content": "From 2D maps to 3D mesh. The occupancy value " + }, + { + "bbox": [ + 46, + 236, + 287, + 315 + ], + "type": "inline_equation", + "content": "O_{\\mathbf{u}}" + }, + { + "bbox": [ + 46, + 236, + 287, + 315 + ], + "type": "text", + "content": " indicates whether " + }, + { + "bbox": [ + 46, + 236, + 287, + 315 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 46, + 236, + 287, + 315 + ], + "type": "text", + "content": " falls within the panels of the garment. To convert the generated 2D occupancy maps " + }, + { + "bbox": [ + 46, + 236, + 287, + 315 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 46, + 236, + 287, + 315 + ], + "type": "text", + "content": " into a 3D garment mesh in its rest shape, we need to recover an ISP latent code " + }, + { + "bbox": [ + 46, + 236, + 287, + 315 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 46, + 236, + 287, + 315 + ], + "type": "text", + "content": " as defined in Eq. 1. To this end, we find the vector " + }, + { + "bbox": [ + 46, + 236, + 287, + 315 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 46, + 236, + 287, + 315 + ], + "type": "text", + "content": " such that the corresponding SDF of ISP best matches the produced occupancy maps by minimizing" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 51, + 320, + 287, + 343 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 320, + 287, + 343 + ], + "spans": [ + { + "bbox": [ + 51, + 320, + 287, + 343 + ], + "type": "interline_equation", + "content": "\\mathbf {z} ^ {*} = \\underset {\\mathbf {z}} {\\operatorname {a r g m i n}} \\sum_ {\\mathbf {u} \\in \\Omega_ {-}} R \\left(s _ {\\mathbf {u}} (\\mathbf {z})\\right) + \\sum_ {\\mathbf {u} \\in \\Omega_ {+}} R \\left(- s _ {\\mathbf {u}} (\\mathbf {z})\\right) + \\lambda_ {\\mathbf {z}} \\| \\mathbf {z} \\| _ {2}, \\tag {9}", + "image_path": "bff62a6f48101461eaad2e87b51b60120fe0ae81b2cb53bbc9baa9afc68b2b70.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 350, + 287, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 350, + 287, + 482 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 287, + 482 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 350, + 287, + 482 + ], + "type": "inline_equation", + "content": "\\Omega_{-} = \\{\\mathbf{u}|O_{\\mathbf{u}} = 1,\\mathbf{u}\\in \\Omega \\}" + }, + { + "bbox": [ + 46, + 350, + 287, + 482 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 350, + 287, + 482 + ], + "type": "inline_equation", + "content": "\\Omega_{+} = \\{\\mathbf{u}|O_{\\mathbf{u}} = 0,\\mathbf{u}\\in \\Omega \\}" + }, + { + "bbox": [ + 46, + 350, + 287, + 482 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 350, + 287, + 482 + ], + "type": "inline_equation", + "content": "R(\\cdot)" + }, + { + "bbox": [ + 46, + 350, + 287, + 482 + ], + "type": "text", + "content": " is the ReLU function, " + }, + { + "bbox": [ + 46, + 350, + 287, + 482 + ], + "type": "inline_equation", + "content": "s_{\\mathbf{u}}(\\mathbf{z})" + }, + { + "bbox": [ + 46, + 350, + 287, + 482 + ], + "type": "text", + "content": " is the SDF value of " + }, + { + "bbox": [ + 46, + 350, + 287, + 482 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 46, + 350, + 287, + 482 + ], + "type": "text", + "content": " computed by ISP, and " + }, + { + "bbox": [ + 46, + 350, + 287, + 482 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathbf{z}}" + }, + { + "bbox": [ + 46, + 350, + 287, + 482 + ], + "type": "text", + "content": " is the weighting constant. With " + }, + { + "bbox": [ + 46, + 350, + 287, + 482 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^*" + }, + { + "bbox": [ + 46, + 350, + 287, + 482 + ], + "type": "text", + "content": ", the rest garment mesh is inferred through ISP's meshing and sewing process. The deformed garment mesh can be obtained by simply replacing the vertex position of the recovered mesh with the values stored in " + }, + { + "bbox": [ + 46, + 350, + 287, + 482 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{M}}" + }, + { + "bbox": [ + 46, + 350, + 287, + 482 + ], + "type": "text", + "content": ". This yields 3D garment meshes in both the rest and deformed states as shown in the top-right of Fig. 2, which are required for the application such as cloth simulation and can be used for further refinement as discussed below." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 490, + 204, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 490, + 204, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 490, + 204, + 502 + ], + "type": "text", + "content": "3.3. Fitting the Models to Images" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 508, + 287, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 508, + 287, + 627 + ], + "spans": [ + { + "bbox": [ + 46, + 508, + 287, + 627 + ], + "type": "text", + "content": "For practical reasons, the range of garment materials, external forces, and body motions present in the training data is limited. As a result, given in-the-wild images as input, the trained model can produce inaccurate results as shown in Fig. 4(b). To remedy this and to leverage the deformation prior that the network " + }, + { + "bbox": [ + 46, + 508, + 287, + 627 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 508, + 287, + 627 + ], + "type": "text", + "content": " captures, we refine the result by minimizing a loss function with respect to the pretrained deformation parameters of " + }, + { + "bbox": [ + 46, + 508, + 287, + 627 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 508, + 287, + 627 + ], + "type": "text", + "content": " as in [13, 21, 51]. " + }, + { + "bbox": [ + 46, + 508, + 287, + 627 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 508, + 287, + 627 + ], + "type": "text", + "content": " is designed to promote a good match between the garment mesh and image observations. We take it to be" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 86, + 637, + 286, + 649 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 637, + 286, + 649 + ], + "spans": [ + { + "bbox": [ + 86, + 637, + 286, + 649 + ], + "type": "interline_equation", + "content": "L = \\lambda_ {C} L _ {C D} + \\lambda_ {n} L _ {\\text {n o r m a l}} + \\lambda_ {p} L _ {\\text {p h y s i c s}}, \\tag {10}", + "image_path": "fbd7333e0f1c54d178f57c6ae660fc68a291f9f5944f0cb7918c58daf2e76190.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 72, + 652, + 286, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 652, + 286, + 666 + ], + "spans": [ + { + "bbox": [ + 72, + 652, + 286, + 666 + ], + "type": "interline_equation", + "content": "L _ {C D} = d \\left(x _ {c} ^ {f}, x _ {I _ {n}}\\right), \\tag {11}", + "image_path": "9a4c84676786d3c6833afa6b50635f0a17287d51663d994abec33ca135e4aa70.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 58, + 669, + 286, + 693 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 669, + 286, + 693 + ], + "spans": [ + { + "bbox": [ + 58, + 669, + 286, + 693 + ], + "type": "interline_equation", + "content": "L _ {\\text {n o r m a l}} = \\sum_ {i \\in f} 1 - \\cos \\left(n ^ {i}, I _ {n} \\left(x _ {c} ^ {i}\\right)\\right), \\tag {12}", + "image_path": "b5e74fc31d9f84db9ad1d5f883564d44536066a817ae0c7d17e5ead179a0fa8f.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 58, + 698, + 286, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 698, + 286, + 711 + ], + "spans": [ + { + "bbox": [ + 58, + 698, + 286, + 711 + ], + "type": "interline_equation", + "content": "L _ {\\text {p h y s i c s}} = L _ {\\text {s t r a i n}} + L _ {\\text {b e n d}} + L _ {\\text {g r a v i t y}} + L _ {\\text {c o l}}, \\tag {13}", + "image_path": "f7344fef98a0a734aea7733cd080a4e0a4c6975a4842aef55bbe6f629c5c80a4.jpg" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 334, + 70, + 425, + 159 + ], + "blocks": [ + { + "bbox": [ + 334, + 70, + 425, + 159 + ], + "lines": [ + { + "bbox": [ + 334, + 70, + 425, + 159 + ], + "spans": [ + { + "bbox": [ + 334, + 70, + 425, + 159 + ], + "type": "image", + "image_path": "5346d41e7e76a011f27ff50cb66777b03bbedfce4702431c284fc30e2c64c5b6.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 374, + 159, + 384, + 168 + ], + "lines": [ + { + "bbox": [ + 374, + 159, + 384, + 168 + ], + "spans": [ + { + "bbox": [ + 374, + 159, + 384, + 168 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 440, + 85, + 510, + 151 + ], + "blocks": [ + { + "bbox": [ + 440, + 85, + 510, + 151 + ], + "lines": [ + { + "bbox": [ + 440, + 85, + 510, + 151 + ], + "spans": [ + { + "bbox": [ + 440, + 85, + 510, + 151 + ], + "type": "image", + "image_path": "a5f9a37d13f87c2eb6001c2531b4222c42f3a93e1ec14548fdbdbe3620642b56.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 468, + 158, + 479, + 168 + ], + "lines": [ + { + "bbox": [ + 468, + 158, + 479, + 168 + ], + "spans": [ + { + "bbox": [ + 468, + 158, + 479, + 168 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 335, + 175, + 423, + 243 + ], + "blocks": [ + { + "bbox": [ + 335, + 175, + 423, + 243 + ], + "lines": [ + { + "bbox": [ + 335, + 175, + 423, + 243 + ], + "spans": [ + { + "bbox": [ + 335, + 175, + 423, + 243 + ], + "type": "image", + "image_path": "efd48e2fc2c792631ebcbf5c5ad46423b275e817bf956c97732261ddab285c74.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 374, + 245, + 384, + 254 + ], + "lines": [ + { + "bbox": [ + 374, + 245, + 384, + 254 + ], + "spans": [ + { + "bbox": [ + 374, + 245, + 384, + 254 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 263, + 545, + 318 + ], + "lines": [ + { + "bbox": [ + 305, + 263, + 545, + 318 + ], + "spans": [ + { + "bbox": [ + 305, + 263, + 545, + 318 + ], + "type": "text", + "content": "Figure 4. Fitting results. Given (a) the normal estimation of an in-the-wild image, (b) is the inference result with ISP recovered geometry. (c) is obtained by optimizing the parameters of the pretrained deformation model. Further refinement of the mesh vertex positions yields (d)." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 429, + 175, + 517, + 242 + ], + "blocks": [ + { + "bbox": [ + 429, + 175, + 517, + 242 + ], + "lines": [ + { + "bbox": [ + 429, + 175, + 517, + 242 + ], + "spans": [ + { + "bbox": [ + 429, + 175, + 517, + 242 + ], + "type": "image", + "image_path": "08b64de78b05e863885fd64134115ba86f91f72b6ff20c8a404cc02ec95ec968.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 468, + 244, + 479, + 255 + ], + "lines": [ + { + "bbox": [ + 468, + 244, + 479, + 255 + ], + "spans": [ + { + "bbox": [ + 468, + 244, + 479, + 255 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 332, + 545, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 332, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 305, + 332, + 545, + 536 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 332, + 545, + 536 + ], + "type": "inline_equation", + "content": "x_{c}^{f}" + }, + { + "bbox": [ + 305, + 332, + 545, + 536 + ], + "type": "text", + "content": " is the 2D projection of the centers of visible faces " + }, + { + "bbox": [ + 305, + 332, + 545, + 536 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 305, + 332, + 545, + 536 + ], + "type": "text", + "content": " after mesh rasterization, " + }, + { + "bbox": [ + 305, + 332, + 545, + 536 + ], + "type": "inline_equation", + "content": "x_{I_n}" + }, + { + "bbox": [ + 305, + 332, + 545, + 536 + ], + "type": "text", + "content": " denotes the coordinates of foreground pixels, " + }, + { + "bbox": [ + 305, + 332, + 545, + 536 + ], + "type": "inline_equation", + "content": "n^i" + }, + { + "bbox": [ + 305, + 332, + 545, + 536 + ], + "type": "text", + "content": " is the normal of face " + }, + { + "bbox": [ + 305, + 332, + 545, + 536 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 332, + 545, + 536 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 332, + 545, + 536 + ], + "type": "inline_equation", + "content": "I_{n}(x_{c}^{i})" + }, + { + "bbox": [ + 305, + 332, + 545, + 536 + ], + "type": "text", + "content": " is the normal image values at " + }, + { + "bbox": [ + 305, + 332, + 545, + 536 + ], + "type": "inline_equation", + "content": "x_{c}^{i}" + }, + { + "bbox": [ + 305, + 332, + 545, + 536 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 332, + 545, + 536 + ], + "type": "inline_equation", + "content": "\\lambda_{C}" + }, + { + "bbox": [ + 305, + 332, + 545, + 536 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 332, + 545, + 536 + ], + "type": "inline_equation", + "content": "\\lambda_{n}" + }, + { + "bbox": [ + 305, + 332, + 545, + 536 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 332, + 545, + 536 + ], + "type": "inline_equation", + "content": "\\lambda_{p}" + }, + { + "bbox": [ + 305, + 332, + 545, + 536 + ], + "type": "text", + "content": " are the balancing scalars. " + }, + { + "bbox": [ + 305, + 332, + 545, + 536 + ], + "type": "inline_equation", + "content": "d(\\cdot)" + }, + { + "bbox": [ + 305, + 332, + 545, + 536 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 332, + 545, + 536 + ], + "type": "inline_equation", + "content": "\\cos (\\cdot)" + }, + { + "bbox": [ + 305, + 332, + 545, + 536 + ], + "type": "text", + "content": " are the functions measuring the 2D Chamfer Distance and the cosine similarity, respectively. " + }, + { + "bbox": [ + 305, + 332, + 545, + 536 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{physics}}" + }, + { + "bbox": [ + 305, + 332, + 545, + 536 + ], + "type": "text", + "content": " is a physics-based loss derived from [40, 48], which computes the membrane strain energy " + }, + { + "bbox": [ + 305, + 332, + 545, + 536 + ], + "type": "inline_equation", + "content": "L_{strain}" + }, + { + "bbox": [ + 305, + 332, + 545, + 536 + ], + "type": "text", + "content": " caused by the deformation, the bending energy " + }, + { + "bbox": [ + 305, + 332, + 545, + 536 + ], + "type": "inline_equation", + "content": "L_{bend}" + }, + { + "bbox": [ + 305, + 332, + 545, + 536 + ], + "type": "text", + "content": " resulting from the folding of adjacent faces, the gravitational potential energy " + }, + { + "bbox": [ + 305, + 332, + 545, + 536 + ], + "type": "inline_equation", + "content": "L_{gravity}" + }, + { + "bbox": [ + 305, + 332, + 545, + 536 + ], + "type": "text", + "content": " and the penalty for body-garment collision " + }, + { + "bbox": [ + 305, + 332, + 545, + 536 + ], + "type": "inline_equation", + "content": "L_{col}" + }, + { + "bbox": [ + 305, + 332, + 545, + 536 + ], + "type": "text", + "content": ". Minimizing " + }, + { + "bbox": [ + 305, + 332, + 545, + 536 + ], + "type": "inline_equation", + "content": "L_{CD}" + }, + { + "bbox": [ + 305, + 332, + 545, + 536 + ], + "type": "text", + "content": " induces an external force of stretching or compression on the garment mesh to align its 2D projection with the given image, while minimizing " + }, + { + "bbox": [ + 305, + 332, + 545, + 536 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{physics}}" + }, + { + "bbox": [ + 305, + 332, + 545, + 536 + ], + "type": "text", + "content": " ensures that the mesh exhibits physically plausible deformation adhering to the shape constraints of the rest-state mesh recovered by ISP." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 537, + 545, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 537, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 305, + 537, + 545, + 644 + ], + "type": "text", + "content": "Minimizing the loss " + }, + { + "bbox": [ + 305, + 537, + 545, + 644 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 305, + 537, + 545, + 644 + ], + "type": "text", + "content": " with respect to the deformation parameters yields a mesh whose overall shape matches the input image, as illustrated in Fig. 4(c). However, since neural networks tend to learn low-frequency functions [43], the result might be too smooth. To recover fine surface details, we perform a refinement step by minimizing " + }, + { + "bbox": [ + 305, + 537, + 545, + 644 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 305, + 537, + 545, + 644 + ], + "type": "text", + "content": " directly with respect to the coordinates of the garment mesh vertices. This generates realistic local details, such as wrinkles on the surface, as shown in Fig. 4(d)." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 657, + 386, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 657, + 386, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 657, + 386, + 670 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "In short, our method begins by inferring the shape and the deformation of the garment in terms of the occupancy and position maps. Leveraging the shape prior of ISP, we then" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1590" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": "recover the garment geometry from the occupancy maps and deform it using the position maps. Next, we refine the initial deformed mesh to better align with image observations by fine-tuning the pre-trained network " + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": " that captures the deformation prior. Finally, we recover fine details through vertex-level optimization of the garment mesh. In this section, we demonstrate the effectiveness of this process and compare it to that of other state-of-the-art methods." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 177, + 180, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 177, + 180, + 190 + ], + "spans": [ + { + "bbox": [ + 47, + 177, + 180, + 190 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 195, + 289, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 195, + 289, + 507 + ], + "spans": [ + { + "bbox": [ + 46, + 195, + 289, + 507 + ], + "type": "text", + "content": "Following the implementation of [32], both the pattern parameterization model " + }, + { + "bbox": [ + 46, + 195, + 289, + 507 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\Theta}" + }, + { + "bbox": [ + 46, + 195, + 289, + 507 + ], + "type": "text", + "content": " and the UV parameterization model " + }, + { + "bbox": [ + 46, + 195, + 289, + 507 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\Phi}" + }, + { + "bbox": [ + 46, + 195, + 289, + 507 + ], + "type": "text", + "content": " of ISP have two separate MLPs for the front and back panels. Each MLP has 7 layers with Softplus activations. The dimension of latent code " + }, + { + "bbox": [ + 46, + 195, + 289, + 507 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 46, + 195, + 289, + 507 + ], + "type": "text", + "content": " is 32. The skinning weight model " + }, + { + "bbox": [ + 46, + 195, + 289, + 507 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 195, + 289, + 507 + ], + "type": "text", + "content": " is a 9-layer MLP with leaky ReLU activations, whose output is normalized by a final Softmax layer. " + }, + { + "bbox": [ + 46, + 195, + 289, + 507 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\Theta}" + }, + { + "bbox": [ + 46, + 195, + 289, + 507 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 195, + 289, + 507 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\Phi}" + }, + { + "bbox": [ + 46, + 195, + 289, + 507 + ], + "type": "text", + "content": " are trained jointly for 9000 iterations with a batch size of 50. " + }, + { + "bbox": [ + 46, + 195, + 289, + 507 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 195, + 289, + 507 + ], + "type": "text", + "content": " is trained with the same parameters as [32]. For the image feature extraction, we use two separate ConvNeXt [35] networks to extract multi-scale garment and body features of sizes 96, 192, 384, 768, which are concatenated as the final features. The point UV coordinates " + }, + { + "bbox": [ + 46, + 195, + 289, + 507 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 46, + 195, + 289, + 507 + ], + "type": "text", + "content": ", 3D position " + }, + { + "bbox": [ + 46, + 195, + 289, + 507 + ], + "type": "inline_equation", + "content": "\\mathbf{X_u}" + }, + { + "bbox": [ + 46, + 195, + 289, + 507 + ], + "type": "text", + "content": ", and image features " + }, + { + "bbox": [ + 46, + 195, + 289, + 507 + ], + "type": "inline_equation", + "content": "F(\\mathbf{x_u})" + }, + { + "bbox": [ + 46, + 195, + 289, + 507 + ], + "type": "text", + "content": " are projected separately to 384 dimensions by three linear layers, which are then concatenated as the input of the MLP of the deformation model. The MLP of the deformation model has 10 layers with a skip connection from the input layer to the middle, and uses Gaussian functions as the activation layer following [44]. The CNNs and MLP of the deformation model are trained jointly for 40 epochs with the Adam optimizer [23] and a learning rate of " + }, + { + "bbox": [ + 46, + 195, + 289, + 507 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 46, + 195, + 289, + 507 + ], + "type": "text", + "content": ". For real images, we use [53] and [12] to obtain their normal and SMPL body parameter estimations, respectively. The garment segmentation masks are generated by leveraging the segmentation of SAM [24] and the semantic labels of [29]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 515, + 265, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 515, + 265, + 528 + ], + "spans": [ + { + "bbox": [ + 47, + 515, + 265, + 528 + ], + "type": "text", + "content": "4.2. Dataset, Evaluation Metrics, and Baseline" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "text", + "content": "Our models are trained on CLOTH3D [4], which is a synthetic dataset with motion sequences of 3D clothed human. It contains garment in the rest state, and deformed states caused by the motion of underlying body. For each garment, it has a single simulated sequence up to 10 seconds. It covers a large variety of garment in different shapes, types and topologies. For the training of ISP, we randomly select 400 shirts, 200 skirts and 200 pairs of trousers, and generate their sewing patterns by the method described in Sec. 3.1. The deformation model is trained on the corresponding simulated sequences. For each frame, we render 11 normal images for the garment mesh with random rotations around the Y-axis, which produces 40K, 40K and 20K training images for shirt, skirt and trousers respectively. During training, we augment the data with image flipping and rotation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "content": "To evaluate the garment reconstruction quality, we use the Chamfer Distance (CD) between the ground truth and the recovered garment mesh, and Intersection over Union (IoU) between the ground truth mask and the rendered mask of reconstructed garment mesh." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 133, + 547, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 547, + 193 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 547, + 193 + ], + "type": "text", + "content": "We compare our method against state-of-the-art methods BCNet [20], SMPLicit [9], ClothWild [39], DrapeNet [11] and ISP [32]. SMPLicit, DrapeNet and ISP use the garment segmentation mask for reconstruction, while BCNet and ClothWild take the RGB images as input." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 198, + 531, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 198, + 531, + 211 + ], + "spans": [ + { + "bbox": [ + 305, + 198, + 531, + 211 + ], + "type": "text", + "content": "4.3. Comparison with State-of-the-Art Methods" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 303, + 225, + 547, + 294 + ], + "blocks": [ + { + "bbox": [ + 303, + 225, + 547, + 294 + ], + "lines": [ + { + "bbox": [ + 303, + 225, + 547, + 294 + ], + "spans": [ + { + "bbox": [ + 303, + 225, + 547, + 294 + ], + "type": "table", + "html": "
CD (×103) ↓SkirtShirtTrousersIoU ↑SkirtShirtTrousers
SMPLicit8.053.490.81SMPLicit0.6570.5270.782
DrapeNetn/a1.540.84DrapeNetn/a0.7150.845
ISP7.262.020.91ISP0.7090.6650.781
Ours2.870.440.28Ours0.9400.9390.946
Ours-GT1.670.310.23Ours-GT0.9530.9420.954
", + "image_path": "61d265aa5c602ed56aabbf35a114ffeac37626f97ab442484dec454ca31efaa7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 297, + 547, + 363 + ], + "lines": [ + { + "bbox": [ + 304, + 297, + 547, + 363 + ], + "spans": [ + { + "bbox": [ + 304, + 297, + 547, + 363 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons. Our method outperforms SMPLicit, DrapeNet, and ISP in terms of CD and IoU on all three garment categories, as shown in the second-to-last row of both tables (Ours). These results were obtained using normals estimated from the images using [53]. In the last row (Ours-GT), we provide the results we obtained using ground-truth normals instead." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 388, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 545, + 544 + ], + "type": "text", + "content": "Quantitative Results. Due to the absence of publicly available real dataset for evaluating garment reconstruction, we utilize synthetic data consisting of 30 unseen shirts, 30 unseen skirts, and 30 pairs of unseen trousers from CLOTH3D for quantitative evaluation. We compare our method with SMPLicit, DrapeNet, and ISP, which recover garment meshes from segmentation masks. The segmentation masks required by the baselines are rendered from the ground truth mesh data, while the normal images required by our method are estimated with [53]. As shown in Tab. 1, our approach significantly outperforms the baselines in terms of Chamfer Distance (CD) and Intersection over Union (IoU) across all garment categories." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 544, + 547, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 547, + 628 + ], + "type": "text", + "content": "For comparison purposes, we re-ran our algorithm using the ground-truth normals and report the results in the last row of Tab. 1. As expected, this leads to an improvement in reconstruction accuracy, but the increase is only modest. This highlights the robustness of our approach to the slight inaccuracies that can be expected from a normal-estimation algorithm." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": "Qualitative Results. Fig. 5 shows the qualitative comparison for the results reconstructed from in-the-wild images. Since BCNet is trained only on synthetic RGB data, it is not able to predict accurate body and garment results. Both SMPLicit and ClothWild generate unrealistic watertight meshes that do not accurately represent real garments." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 314, + 757 + ], + "type": "text", + "content": "1591" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 542, + 336 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 542, + 336 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 542, + 336 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 542, + 336 + ], + "type": "image", + "image_path": "2f0fb335fe2dea224549700b3f523ff6976a06bda21bbe3ddd9544024b086391.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 338, + 547, + 373 + ], + "lines": [ + { + "bbox": [ + 46, + 338, + 547, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 547, + 373 + ], + "type": "text", + "content": "Figure 5. Comparison against SOTA methods. From left to right, we show the input image and the 3D garment meshes recovered by our method and SOTA methods: BCNet, SMPLcit, ClothWild, DrapeNet, ISP. (Since skirt is unavailable for DrapeNet, a random pair of trousers is put on its result in the second row.)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 390, + 164, + 460 + ], + "blocks": [ + { + "bbox": [ + 50, + 390, + 164, + 460 + ], + "lines": [ + { + "bbox": [ + 50, + 390, + 164, + 460 + ], + "spans": [ + { + "bbox": [ + 50, + 390, + 164, + 460 + ], + "type": "table", + "html": "
CD (×103) ↓GTEst.
(-D*,-v*)2.043.58
(-D*,+v*)1.803.39
(+D*,-v*)1.752.93
(+D*,+v*)1.672.87
", + "image_path": "705d6698165a902bddbd4698090b63240d9b8d195faea797b5817ab50d080a5d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 171, + 390, + 285, + 460 + ], + "blocks": [ + { + "bbox": [ + 171, + 390, + 285, + 460 + ], + "lines": [ + { + "bbox": [ + 171, + 390, + 285, + 460 + ], + "spans": [ + { + "bbox": [ + 171, + 390, + 285, + 460 + ], + "type": "table", + "html": "
IoU ↑GTEst.
(-D*,-v*)0.8480.751
(-D*,+v*)0.9470.904
(+D*,-v*)0.9410.923
(+D*,+v*)0.9530.940
", + "image_path": "8068dd4d0314ac8d413b66ad4748e66ebad32ed8b160004d6f68c3e95f2ca9c6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 478, + 288, + 523 + ], + "lines": [ + { + "bbox": [ + 46, + 478, + 288, + 523 + ], + "spans": [ + { + "bbox": [ + 46, + 478, + 288, + 523 + ], + "type": "text", + "content": "Table 2. Ablation study. GT and Est. mean using ground-truth and estimated normal images as input, respectively. " + }, + { + "bbox": [ + 46, + 478, + 288, + 523 + ], + "type": "inline_equation", + "content": "+ / - {\\mathcal{D}}^{ * }" + }, + { + "bbox": [ + 46, + 478, + 288, + 523 + ], + "type": "text", + "content": " denote with/without the finetuning of the deformation model. " + }, + { + "bbox": [ + 46, + 478, + 288, + 523 + ], + "type": "inline_equation", + "content": "+ / - {v}^{ * }" + }, + { + "bbox": [ + 46, + 478, + 288, + 523 + ], + "type": "text", + "content": " denote with/without the optimization of vertex positions." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 544, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 616 + ], + "type": "text", + "content": "While DrapeNet and ISP are able to recover open surfaces for garments, their results remain closely adhered to the body, similar to SMPLlicit and ClothWild. In contrast, our method can faithfully recover garment mesh from input images, as clearly observed in the depiction of the open jacket and the flowing long skirt that stand away from the body." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 623, + 141, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 141, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 141, + 635 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": "Table 2 presents the evaluation results of our method with and without fitting on the test set of skirts. The results of row 1, 3 and 4 indicate that optimizing the parameters of the pretrained deformation network " + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "(+D^{*}, - v^{*})" + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": " improves the quality of the raw inference results, and optimizing the vertex positions " + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "(+D^{*}, + v^{*})" + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": " further enhances the outcome," + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 309, + 390, + 391, + 494 + ], + "blocks": [ + { + "bbox": [ + 309, + 390, + 391, + 494 + ], + "lines": [ + { + "bbox": [ + 309, + 390, + 391, + 494 + ], + "spans": [ + { + "bbox": [ + 309, + 390, + 391, + 494 + ], + "type": "image", + "image_path": "7478b15f133d96704f58df0b2f2bcbb1666c21e39f9fa5c8d7d6457986c66f81.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 513, + 545, + 535 + ], + "lines": [ + { + "bbox": [ + 305, + 513, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 305, + 513, + 545, + 535 + ], + "type": "text", + "content": "Figure 6. Fitting strategy comparison between (a) our fitting method and (b) only vertex optimization." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 395, + 396, + 457, + 494 + ], + "blocks": [ + { + "bbox": [ + 395, + 396, + 457, + 494 + ], + "lines": [ + { + "bbox": [ + 395, + 396, + 457, + 494 + ], + "spans": [ + { + "bbox": [ + 395, + 396, + 457, + 494 + ], + "type": "image", + "image_path": "01769d869b91a12282eb349c458d95ff8d6aeb1fbee41764ffecfa781005fd17.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 426, + 494, + 438, + 504 + ], + "lines": [ + { + "bbox": [ + 426, + 494, + 438, + 504 + ], + "spans": [ + { + "bbox": [ + 426, + 494, + 438, + 504 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 471, + 397, + 531, + 493 + ], + "blocks": [ + { + "bbox": [ + 471, + 397, + 531, + 493 + ], + "lines": [ + { + "bbox": [ + 471, + 397, + 531, + 493 + ], + "spans": [ + { + "bbox": [ + 471, + 397, + 531, + 493 + ], + "type": "image", + "image_path": "ba54e37c27845e8ac3e46b1829ec92bcb5a36c730677597d999e9ceeb686e7fc.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 500, + 494, + 512, + 504 + ], + "lines": [ + { + "bbox": [ + 500, + 494, + 512, + 504 + ], + "spans": [ + { + "bbox": [ + 500, + 494, + 512, + 504 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 557, + 470, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 557, + 470, + 569 + ], + "spans": [ + { + "bbox": [ + 305, + 557, + 470, + 569 + ], + "type": "text", + "content": "yielding the lowest CD and highest IoU." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "However, as demonstrated in the second row, directly optimizing the vertex positions on the raw inference results without optimizing the deformation model " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "(-\\mathcal{D}^{*}, + v^{*})" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " is not as effective. The deformation model encapsulates a continuous deformation field. When its weights are optimized based on partial observations, the entire field undergoes modification, thereby influencing all mesh vertices globally. On the contrary, a direct vertex optimization with partial observations predominantly affects the mesh vertices locally. While this can capture localized details like wrinkles, it struggles to resolve discrepancies in the overall shape. Fig. 6 displays the results of our methods and an ablation with" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "type": "text", + "content": "1592" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 71, + 547, + 513 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 547, + 513 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 547, + 513 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 547, + 513 + ], + "type": "image", + "image_path": "a84fb222c2ae35f5f1b04f7b7c0ef7c7e718caba24661b33ebab34379c1b9cfa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 522, + 546, + 544 + ], + "lines": [ + { + "bbox": [ + 46, + 522, + 546, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 546, + 544 + ], + "type": "text", + "content": "Figure 7. Recovering from in-the-wild images. Our method is able to recover realistic meshes for garments with diverse shapes and deformations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 559, + 288, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 559, + 288, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 559, + 288, + 594 + ], + "type": "text", + "content": "vertex optimization only. Unsurprisingly, the latter fails at recovering the correct shape and produces implausible deformation on the mesh surface." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 627, + 132, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 627, + 132, + 639 + ], + "spans": [ + { + "bbox": [ + 47, + 627, + 132, + 639 + ], + "type": "text", + "content": "4.5. More Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 653, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 713 + ], + "type": "text", + "content": "Fig. 7 shows a collection of results reconstructed by our method from in-the-wild images. Our method can produce realistic 3D meshes with fine details across a wide range of garment types, from tight-fitting attire to more relaxed and flowing outfits." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 558, + 379, + 570 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 558, + 379, + 570 + ], + "spans": [ + { + "bbox": [ + 306, + 558, + 379, + 570 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 578, + 545, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 578, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 305, + 578, + 545, + 673 + ], + "type": "text", + "content": "We have presented a novel approach to recovering realistic 3D garment meshes from in-the-wild images featuring loose fitting clothing. It relies on a fitting process that imposes shape and deformation priors learned on synthetic data to accurately capture garment shape and deformations. In future work, we will extend our approach to modeling deformations over time from video sequences while enforcing temporal consistency of the reconstructions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 689, + 545, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 711 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 711 + ], + "type": "text", + "content": "Acknowledgement. This project was supported in part by the Swiss National Science Foundation." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1593" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] T. Alldieck, M. Magnor, B. Bhatnagar, C. Theobalt, and G. Pons-Moll. Learning to reconstruct people in clothing from a single RGB camera. In Conference on Computer Vision and Pattern Recognition, pages 1175-1186, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 287, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 287, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 287, + 179 + ], + "type": "text", + "content": "[2] T. Alldieck, G. Pons-Moll, C. Theobalt, and M. Magnor. Tex2shape: Detailed full human body geometry from a single image. In International Conference on Computer Vision, pages 2293-2303, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 287, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 287, + 223 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 287, + 223 + ], + "type": "text", + "content": "[3] T. Alldieck, M. Zanfir, and C. Sminchisescu. Photorealistic monocular 3d reconstruction of humans wearing clothing. In Conference on Computer Vision and Pattern Recognition, pages 1506-1515, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 224, + 287, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 287, + 257 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 287, + 257 + ], + "type": "text", + "content": "[4] H. Bertiche, M. Madadi, and S. Escalera. CLOTH3D: Clothed 3D Humans. In European Conference on Computer Vision, pages 344-359, 2020. 3, 6, 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 258, + 287, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 287, + 300 + ], + "type": "text", + "content": "[5] B. L. Bhatnagar, G. Tiwari, C. Theobalt, and G. Pons-Moll. Multi-Garment Net: Learning to Dress 3D People from Images. In International Conference on Computer Vision, 2019. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 302, + 287, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 287, + 346 + ], + "type": "text", + "content": "[6] F. Bogo, A. Kanazawa, C. Lassner, P. Gehler, J. Romero, and M. J. Black. Keep It SMPL: Automatic Estimation of 3D Human Pose and Shape from a Single Image. In European Conference on Computer Vision, 2016. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 346, + 287, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 287, + 379 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 287, + 379 + ], + "type": "text", + "content": "[7] R. Brouet, A. Sheffer, L. Boissieux, and M.P. Cani. Design Preserving Garment Transfer. ACM Transactions on Graphics, 31(4):361-3611, 2012. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 380, + 287, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 287, + 413 + ], + "type": "text", + "content": "[8] A. Casado-Elvira, M. Trinidad, and D. Casas. PERGAMO: Personalized 3d garments from monocular video. In Computer Graphics Forum, pages 293–304, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 414, + 287, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 414, + 287, + 457 + ], + "spans": [ + { + "bbox": [ + 53, + 414, + 287, + 457 + ], + "type": "text", + "content": "[9] E. Corona, A. Pumarola, G. Alenya, G. Pons-Moll, and F. Moreno-Noguer. SMPlicit: Topology-Aware Generative Model for Clothed People. In Conference on Computer Vision and Pattern Recognition, 2021. 1, 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 458, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 458, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 458, + 287, + 491 + ], + "type": "text", + "content": "[10] R. Danerek, E. Dibra, C. Öztireli, R. Ziegler, and M. Gross. Deepgarment : 3D Garment Shape Estimation from a Single Image. Eurographics, 2017. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 491, + 287, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 287, + 535 + ], + "type": "text", + "content": "[11] Luca DeLuigi, Ren Li, Benoit Guillard, Mathieu Salzmann, and Pascal Fua. DrapeNet: Generating Garments and Draping them with Self-Supervision. In Conference on Computer Vision and Pattern Recognition, 2023. 1, 2, 4, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 536, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 578 + ], + "type": "text", + "content": "[12] Y. Feng, V. Choutas, T. Bolkart, D. Tzionas, and M. J. Black. Collaborative Regression of Expressive Bodies using Moderation. In International Conference on Computer Vision, 2021. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 579, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 287, + 612 + ], + "type": "text", + "content": "[13] M. Gadelha, R. Wang, and S. Maji. Deep manifold prior. In International Conference on Computer Vision, pages 1107-1116, 2021. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "type": "text", + "content": "[14] G. Georgakis, R. Li and S. Karanam, T. Chen, J. Košecka, and Z. Wu. Hierarchical kinematic human mesh recovery. In European Conference on Computer Vision, pages 768-784, 2020. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 658, + 287, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 691 + ], + "type": "text", + "content": "[15] B. Guillard, F. Stella, and P. Fua. MeshUDF: Fast and Differentiable Meshing of Unsigned Distance Field Networks. In European Conference on Computer Vision, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "text", + "content": "[16] T. He, J. Collomosse, H. Jin, and S. Soatto. Geo-pifu: Geometry and pixel aligned implicit functions for single-view" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "human reconstruction. Advances in Neural Information Processing Systems, 33:9276-9287, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "type": "text", + "content": "[17] T. He, Y. Xu, S. Saito, S. Soatto, and T. Tung. Arch++: Animation-ready clothed human reconstruction revisited. In International Conference on Computer Vision, pages 11046-11056, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 141, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 545, + 184 + ], + "type": "text", + "content": "[18] Z. Huang, Y. Xu, C. Lassner, H. Li, and T. Tung. Arch: Animatable Reconstruction of Clothed Humans. In Conference on Computer Vision and Pattern Recognition, pages 3093-3102, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 186, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 545, + 228 + ], + "type": "text", + "content": "[19] A. Jackson, C. Manafas, and G. Tzimiropoulos. 3d human body reconstruction from a single image via volumetric regression. In European Conference on Computer Vision Workshops, 2018. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 230, + 545, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 230, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 308, + 230, + 545, + 263 + ], + "type": "text", + "content": "[20] B. Jiang, J. Zhang, Y. Hong, J. Luo, L. Liu, and H. Bao. Bcnet: Learning body and cloth shape from a single image. In European Conference on Computer Vision, 2020. 1, 2, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 264, + 545, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 306 + ], + "type": "text", + "content": "[21] H. Joo, N. Neverova, and A. Vedaldi. Exemplar Fine-Tuning for 3D Human Pose Fitting Towards In-the-Wild 3D Human Pose Estimation. In International Conference on 3D Vision, 2020. 1, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 308, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 308, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 308, + 308, + 545, + 342 + ], + "type": "text", + "content": "[22] A. Kanazawa, M. J. Black, D. W. Jacobs, and J. Malik. End-To-End Recovery of Human Shape and Pose. In Conference on Computer Vision and Pattern Recognition, 2018. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 343, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 343, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 308, + 343, + 545, + 374 + ], + "type": "text", + "content": "[23] D. P. Kingma and J. Ba. Adam: A Method for Stochastic Optimisation. In International Conference on Learning Representations, 2015. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 376, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 545, + 420 + ], + "type": "text", + "content": "[24] A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. Berg, W. Lo, P. Dollar, and R. Girshick. Segment anything. In International Conference on Computer Vision, 2023. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 421, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 421, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 308, + 421, + 545, + 464 + ], + "type": "text", + "content": "[25] N. Kolotouros, G. Pavlakos, M. J. Black, and K. Daniilidis. Learning to Reconstruct 3D Human Pose and Shape via Model-Fitting in the Loop. In Conference on Computer Vision and Pattern Recognition, 2019. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 466, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 466, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 466, + 545, + 510 + ], + "type": "text", + "content": "[26] C. Lassner, J. Romero, M. Kiefel, F. Bogo, M.J. Black, and P.V. Gehler. Unite the People: Closing the Loop Between 3D and 2D Human Representations. In Conference on Computer Vision and Pattern Recognition, 2017. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 511, + 545, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 511, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 308, + 511, + 545, + 543 + ], + "type": "text", + "content": "[27] B. Lévy, S. Petitjean, N. Ray, and J. Maillot. Least Squares Conformal Maps for Automatic Texture Atlas Generation. In ACM SIGGRAPH, 2002. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 544, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 544, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 308, + 544, + 545, + 589 + ], + "type": "text", + "content": "[28] J. Li, C. Xu, Z. Chen, S. Bian, L. Yang, and C. Lu. Hybrid: A hybrid analytical-neural inverse kinematics solution for 3d human pose and shape estimation. In Conference on Computer Vision and Pattern Recognition, 2021. 1" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 590, + 545, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 545, + 622 + ], + "type": "text", + "content": "[29] P. Li, Y. Xu, Y. Wei, and Y. Yang. Self-Correction for Human Parsing. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(6):3260-3271, 2020. 6" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 624, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 656 + ], + "type": "text", + "content": "[30] R. Li, M. Zheng, S. Karanam, T. Chen, and Z. Wu. Everybody Is Unique: Towards Unbiased Human Mesh Recovery. In British Machine Vision Conference, 2021. 1" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "type": "text", + "content": "[31] R. Li, B. Guillard, E. Remelli, and P. Fua. DIG: Draping Implicit Garment over the Human Body. In Asian Conference on Computer Vision, 2022. 1, 2, 4" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[32] Ren Li, Benoit Guillard, and Pascal Fua. ISP: Multi-Layered Garment Draping with Implicit Sewing Patterns. In Ad-" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1594" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 94 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 94 + ], + "type": "text", + "content": "vances in Neural Information Processing Systems, 2023. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "text", + "content": "[33] L. Liu, L. Zhang, Y. Xu, C. Gotsman, and S. J. Gortler. A Local/Global Approach to Mesh Parameterization. In Proceedings of the Symposium on Geometry Processing, pages 1495-1504, 2008. 4, 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "type": "text", + "content": "[34] X. Liu, J. Li, and G. Lu. Modeling Realistic Clothing from a Single Image under Normal Guide. IEEE Transactions on Visualization and Computer Graphics, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 175, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 207 + ], + "type": "text", + "content": "[35] Z. Liu, H. Mao, C. Wu, C. Feichtenhofer, T. Darrell, and S. Xie. A ConvNet for the 2020s. In Conference on Computer Vision and Pattern Recognition, 2022. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 209, + 287, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 240 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 240 + ], + "type": "text", + "content": "[36] M. Loper, N. Mahmood, J. Romero, G. Pons-Moll, and M.J. Black. SMPL: A Skinned Multi-Person Linear Model. ACM SIGGRAPH Asia, 34(6), 2015. 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 243, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 243, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 243, + 287, + 285 + ], + "type": "text", + "content": "[37] Qianli Ma, Jinlong Yang, Michael J. Black, and Siyu Tang. Neural Point-based Shape Modeling of Humans in Challenging Clothing. In International Conference on 3D Vision, 2022. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 288, + 287, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 288, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 48, + 288, + 287, + 331 + ], + "type": "text", + "content": "[38] G. Moon and K.M. Lee. I2L-MeshNet: Image-to-Lixel Prediction Network for Accurate 3D Human Pose and Mesh Estimation from a Single RGB Image. In European Conference on Computer Vision, 2020. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 333, + 287, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 333, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 48, + 333, + 287, + 365 + ], + "type": "text", + "content": "[39] G. Moon, H. Nam, T. Shiratori, and K.M. Lee. 3d clothed human reconstruction in the wild. In European Conference on Computer Vision, 2022. 1, 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 367, + 287, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 367, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 48, + 367, + 287, + 399 + ], + "type": "text", + "content": "[40] R. Narain, A. Samii, and J.F. O'brien. Adaptive anisotropic remeshing for cloth simulation. ACM Transactions on Graphics, 2012. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 401, + 287, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 401, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 401, + 287, + 443 + ], + "type": "text", + "content": "[41] M. Omran, C. Lassner, G. Pons-Moll, P. Gehler, and B. Schiele. Neural Body Fitting: Unifying Deep Learning and Model-Based Human Pose and Shape Estimation. In International Conference on 3D Vision, 2018. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 445, + 287, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 287, + 488 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 287, + 488 + ], + "type": "text", + "content": "[42] N. Pietroni, C. Dumery, R. Falque, M. Liu, T. Vidal-Calleja, and O. Sorkine-Hornung. Computational pattern making from 3D garment models. ACM Transactions on Graphics, 41(4):1-14, 2022. 3, 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 491, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 287, + 533 + ], + "type": "text", + "content": "[43] N. Rahaman, A. Baratin, D. Arpit, F. Draxler, M. Lin, F. Hamprecht, Y. Bengio, and A. Courville. On the spectral bias of neural networks. In International Conference on Machine Learning, 2019. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 536, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 578 + ], + "type": "text", + "content": "[44] S. Ramasinghe and S. Lucey. Beyond Periodicity: Towards a Unifying Framework for Activations in Coordinate-MLPs. In European Conference on Computer Vision, pages 142–158, 2022. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 580, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 287, + 623 + ], + "type": "text", + "content": "[45] S. Saito, Z. Huang, R. Natsume, S. Morishima, A. Kanazawa, and H. Li. PIFu: Pixel-Aligned Implicit Function for High-Resolution Clothed Human Digitization. In International Conference on Computer Vision, 2019. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "type": "text", + "content": "[46] S. Saito, T. Simon, J. Saragih, and H. Joo. Pifuuhd: Multi-Level Pixel-Aligned Implicit Function for High-Resolution 3D Human Digitization. In Conference on Computer Vision and Pattern Recognition, pages 84–93, 2020. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[47] I. Santesteban, N. Thuerey, M. A. Otaduy, and D. Casas. Self-Supervised Collision Handling via Generative 3D Garment Models for Virtual Try-On. In Conference on Computer Vision and Pattern Recognition, 2021. 1" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 623 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[48] I. Santesteban, M.A. Otaduy, and D. Casas. SNUG: Self-Supervised Neural Dynamic Garments. In Conference on Computer Vision and Pattern Recognition, 2022. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 108, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 129 + ], + "type": "text", + "content": "[49] O. Sorkine-Hornung and M. Rabinovich. Least-squares rigid motion using svd. Computing, 1(1):1-5, 2017. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 130, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 130, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 130, + 545, + 183 + ], + "type": "text", + "content": "[50] A. Srivastava, C. Pokhariya, S. Jinka, and A. Sharma. xCloth: Extracting Template-free Textured 3D Clothes from a Monocular Image. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2504-2512, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 186, + 545, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 218 + ], + "type": "text", + "content": "[51] D. Ulyanov, A. Vedaldi, and V. Lempitsky. Deep Image Prior. In Conference on Computer Vision and Pattern Recognition, pages 9446-9454, 2018. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 220, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 262 + ], + "type": "text", + "content": "[52] Y. Xiu, J. Yang, D. Tzionas, and M. J Black. Icon: Implicit clothed humans obtained from normals. In Conference on Computer Vision and Pattern Recognition, pages 13286-13296, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "type": "text", + "content": "[53] Y. Xiu, J. Yang, X. Cao, D. Tzionas, and M. J Black. ECON: Explicit Clothed humans Optimized via Normal integration. In Conference on Computer Vision and Pattern Recognition, pages 512-523, 2023. 1, 2, 6, 3, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 309, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 353 + ], + "type": "text", + "content": "[54] F. Yang, R. Li, G. Georgakis, S. Karanam, T. Chen, H. Ling, and Z. Wu. Robust multi-modal 3d patient body modeling. In Conference on Medical Image Computing and Computer Assisted Intervention, pages 86-95, 2020. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 354, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 387 + ], + "type": "text", + "content": "[55] S. Yang, Z. Pan, T. Amert, K. Wang, L. Yu, T. Berg, and M. Lin. Physics-inspired garment recovery from a single-view image. ACM Transactions on Graphics, 37(5):1-14, 2018. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 388, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 420 + ], + "type": "text", + "content": "[56] I. Zakharkin, K. Mazur, A. Grigorev, and V. Lempitsky. Point-based modeling of human clothing. In International Conference on Computer Vision, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 422, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 545, + 454 + ], + "type": "text", + "content": "[57] Z. Zheng, T. Yu, Y. Wei, Q. Dai, and Y. Liu. Deephuman: 3d human reconstruction from a single image. In International Conference on Computer Vision, pages 7739-7749, 2019. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 456, + 545, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 499 + ], + "type": "text", + "content": "[58] Z. Zheng, T. Yu, Y. Liu, and Q. Dai. Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(6):3170-3184, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 501, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 501, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 501, + 545, + 533 + ], + "type": "text", + "content": "[59] B. Zhou, X. Chen, Q. Fu, K. Guo, and P. Tan. Garment modeling from a single image. In Computer Graphics Forum, pages 85–91, 2013. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 535, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 535, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 535, + 545, + 578 + ], + "type": "text", + "content": "[60] H. Zhu, Y. Cao, H. Jin, W. Chen, D. Du, Z. Wang, S. Cui, and X. Han. Deep fashion3d: A dataset and benchmark for 3d garment reconstruction from single images. In European Conference on Computer Vision, pages 512-530, 2020. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 580, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 580, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 307, + 580, + 545, + 623 + ], + "type": "text", + "content": "[61] H. Zhu, L. Qiu, Y. Qiu, and X. Han. Registering explicit to implicit: Towards high-fidelity garment mesh reconstruction from single images. In Conference on Computer Vision and Pattern Recognition, pages 3845-3854, 2022. 2" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1595" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Gated Fields_ Learning Scene Reconstruction from Gated Videos/babd13ae-5bce-44f5-bd0f-203ff88804c8_content_list.json b/2024/Gated Fields_ Learning Scene Reconstruction from Gated Videos/babd13ae-5bce-44f5-bd0f-203ff88804c8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b7f3b2cf291b386a844d91ff36e1dac6d0d4e7b3 --- /dev/null +++ b/2024/Gated Fields_ Learning Scene Reconstruction from Gated Videos/babd13ae-5bce-44f5-bd0f-203ff88804c8_content_list.json @@ -0,0 +1,1695 @@ +[ + { + "type": "text", + "text": "Gated Fields: Learning Scene Reconstruction from Gated Videos", + "text_level": 1, + "bbox": [ + 156, + 92, + 816, + 113 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Andrea Ramazzina $^{1*}$ Stefanie Walz $^{1,2*}$ Pragyan Dahal $^{3}$ Mario Bijelic $^{4,5}$ Felix Heide $^{4,5}$", + "bbox": [ + 127, + 142, + 866, + 162 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Mercedes-Benz $^{2}$ Saarland University $^{3}$ Politecnico di Milano $^{4}$ Torch Robotics $^{5}$ Princeton University", + "bbox": [ + 168, + 164, + 820, + 181 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 215, + 313, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Reconstructing outdoor 3D scenes from temporal observations is a challenge that recent work on neural fields has offered a new avenue for. However, existing methods that recover scene properties, such as geometry, appearance, or radiance, solely from RGB captures often fail when handling poorly-lit or texture-deficient regions. Similarly, recovering scenes with scanning LiDAR sensors is also difficult due to their low angular sampling rate which makes recovering expansive real-world scenes difficult. Tackling these gaps, we introduce Gated Fields - a neural scene reconstruction method that utilizes active gated video sequences. To this end, we propose a neural rendering approach that seamlessly incorporates time-gated capture and illumination. Our method exploits the intrinsic depth cues in the gated videos, achieving precise and dense geometry reconstruction irrespective of ambient illumination conditions. We validate the method across day and night scenarios and find that Gated Fields compares favorably to RGB and LiDAR reconstruction methods. Our code and datasets are available here1.", + "bbox": [ + 75, + 250, + 473, + 551 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 580, + 209, + 595 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large-scale outdoor scene reconstruction is essential for advancing autonomous robotics, drones, and driver-assistance systems, serving as the foundation for scene understanding, safe navigation, dataset generation and validation. Existing works in this domain [68, 76, 96] have typically adopted a two-step approach. Initially, they infer depth maps from different poses, utilizing time-of-flight sensors or RGB captures. Subsequently, these depth estimates are fused to produce a coherent 3D representation, using either classical methodologies [25] or learned-based representations [72]. In contrast, more recent studies [16, 65, 79] have proposed end-to-end strategies that bypass the estimation of local depth maps as an intermediate representation. Instead, they directly regress a Truncated Signed Distance Function (TSDF) [65, 79] or an occupancy volume [16]. A rapidly growing body of work on neural rendering [62, 92] offers not only geometrically-accurate scene recon", + "bbox": [ + 75, + 606, + 470, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ea50763dced457e9ea678ff4d3a3ba539dccfd54e79a4d0efab815d8df49d087.jpg", + "image_caption": [ + "Figure 1. From a single video of gated captures (top-row), we reconstruct an accurate scene representation and render depth projections (mid-row, right) as accurate as LiDAR scans (mid-row, left), and we recover 3D geometry and normals (bottom-row)." + ], + "image_footnote": [], + "bbox": [ + 501, + 214, + 893, + 404 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "struction from posed RGB images but also to generate novel perspectives from unobserved angles. Hinging on implicit coordinate-based neural representations, RGB-based methods [7, 8, 55, 100] have been adapted to large open outdoor environments. A recent line of work [37, 71, 86, 93] includes LiDAR scans for auxiliary depth supervision and to improve scene reconstruction for urban environments. However, recovery based on the RGB images exhibit is fundamentally limited in the presence of low light [63, 91] or in the presence of scattering such as fog [49, 69].", + "bbox": [ + 496, + 474, + 893, + 626 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A parallel direction of research investigates neural rendering techniques tailored to Time-of-Flight (ToF) sensors as opposed to the conventional RGB cameras. Existing methods [41, 84] tackle scene understanding with posed LiDAR scans, and have modeled the raw output from a single-photon LiDAR system [59] or continuous-wave ToF sensor [4] as additional depth supervision. However, all of these existing methods struggle with recovering large unbounded outdoor scenes: while continuous-wave ToF sensors [4] offer signal only in room-sized scenes, methods based on scanning LiDAR suffer from low angular sampling which mandates temporal aggregation. Specifically, even today's LiDAR sensors boasting 200 scan lines, lag drastically in resolution compared to current HDR cameras that offer two orders of magnitude higher vertical pixel counts nearing $10\\mathrm{k}$ and three orders of magnitude higher total resolution.", + "bbox": [ + 496, + 627, + 895, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Addressing these challenges, our work explores scene reconstruction using active gated imaging. Gated imag-", + "bbox": [ + 500, + 869, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*These authors contributed equally to this work.", + "bbox": [ + 93, + 875, + 351, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1https://light.princeton.edu/gatedfields/", + "bbox": [ + 93, + 887, + 379, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10530", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing functions by integrating the transient response from a scene that has been flash-illuminated by a synchronized light source [19]. This imaging technique is robust to adverse weather conditions – temporal gating allows us to filter out backscatter – and provides signal in poorly-lit scenes [13]. Existing work has exploited this sensing modality to achieve state-of-the-art depth estimation [89, 90] as well as object detection [44]. In our approach, we train a neural field-based representation of the scene, concurrently learning its geometry, illumination, and material properties. This is accomplished by integrating the gated imaging formation model with a neural rendering framework, that jointly learns the associated gating parameters along with the scene reconstruction. By leveraging the implicit depth cues present in gated video captures, we are able to reconstruct a detailed 3D geometric model of the scene, as shown in Fig. 1. Compared to LiDAR-based approaches, our method offers distinct advantages, as LiDAR systems are inherently constrained by their resolution, necessitating additional time-multiplexed scene captures to aggregate points. This results in an extended acquisition process for LiDAR-based methods or, conversely, compromises the geometric detail of the final estimate. Specifically, for a fixed time acquisition budget the scene reconstruction from a LiDAR sensor is less supervised, although offering highly accurate depth information, the sensor yields data at a volume one order of magnitude less than that of a camera stereo pair. This disparity in data quantity means that while the LiDAR provides precise depth points, it is unable to provide dense and fine detailed predictions.", + "bbox": [ + 76, + 90, + 472, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To assess our method, we captured a dataset of varied scenes at day and night conditions, using a vehicle test setup comprising of LiDAR, RGB and Gated sensors. We compare our method with feed-forward and 3D reconstruction methods, and assess its superiority in novel depth synthesis beating the next best method by $21.87\\%$ MAE, 3D reconstruction improving on the baseline by $11\\%$ IoU, and performs novel view synthesis with a PSNR of $32.28~\\mathrm{dB}$ .", + "bbox": [ + 76, + 547, + 472, + 670 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In our work, we make the following contributions:", + "bbox": [ + 76, + 679, + 413, + 694 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel neural rendering method and scene representation that is capable of reconstructing scene geometry and radiance from active gated camera videos.", + "- By modeling the gated image formation process and integrating into differentiable volume rendering, our approach is able to reconstruct and decompose both passive and active light transport components conditioned on the scene parameters in a physically accurate way.", + "- We validate our method on large outdoor scenes, captured across different scenarios in both day and night, achieving a reduction of MAE error in depth precision of $59.8\\%$ to the next best RGB+LiDAR method and $31.7\\%$ to the next best methods using gated images." + ], + "bbox": [ + 76, + 704, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 89, + 640, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Monocular and Stereo Depth Estimation Depth estimation tasks, from a single image [32, 36, 51, 54], from stereo image pairs [5, 22, 52, 95], or single/stereo images with a LiDAR scan [24, 40, 67, 82, 83, 94, 102] have been at the center of a large body of work. Single CMOS sensor-based depth estimation from RGB color images is inherently limited by scale ambiguity. Additional measurements from LiDAR [9] or ego-vehicle speed [36] can resolve this ambiguity at the cost of an additional sensor. Similarly, stereo methods rely on an additional camera sensor to resolve ambiguity by triangulating between two camera views [22]. Training approaches for learned depth estimation methods using intensity images cover both unsupervised methods [30, 32, 33, 36, 103], which harnesses multi-view geometry consistency, and supervised techniques [22, 27, 42, 45, 51, 54, 58, 60] relying primarily on multi-view datasets [45, 51, 60] or time-of-flight captures [22, 27, 42, 58]. In particular, LiDAR measurements have been proven as a ground-truth signal for depth supervision [2, 22, 27, 42, 54, 58]. Several methods [31, 87] have mitigated the sparsity and range limitations of scanning LiDAR by accumulating scans. However, adverse weather [14] can make LiDAR ground truth unreliable, and methods that rely on consistency between camera and LiDAR [24, 40, 67, 82, 83, 94, 102] suffer from degradations, including scan pattern artifacts and temporal distortions.", + "bbox": [ + 496, + 122, + 893, + 515 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Time-gated Depth Sensing Time-of-Flight (ToF) sensors determine depth by emitting light into a scene and calculating the distance based on the round trip time of the light. Successful sensing methods can be categorized into three classes: correlation ToF cameras [38, 46, 48], pulsed ToF sensors [77], and gated imaging [34, 39]. Correlation ToF sensors[38, 46, 48] estimate depth by continuously flood-illuminating the scene and assessing the phase shift between the emitted and received light. While these sensors can provide high-resolution depth information, their use is primarily confined to indoor settings due to their susceptibility to external light interference. Pulsed ToF sensors [77] function by emitting light pulses toward specific scene points and measure the total travel time to estimate depth. Emitting collimated light, this approach is robust against ambient illumination and allows for outdoor depth measurements. However, its spatial resolution is fundamentally limited owing to its scanning illumination technique, and its efficacy can be compromised in adverse weather due to backscatter [11, 21, 43]. In contrast, gated cameras [12, 34, 39] capture light from a scene over brief intervals, essentially constraining the observable depth to specific range segments. The inherent gating mechanism of these cameras offers resistance to backscattering, and they allow for the recovery of detailed depth maps when using a large number of short", + "bbox": [ + 496, + 523, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10531", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "gates [3, 17, 18] Subsequent works have improved gated depth estimation with few gates by adopting Bayesian approaches [1, 75] or deep neural networks [35, 89, 90], and they achieve accurate gated depth estimation for dynamic outdoor scenes, even under challenging conditions. Recently, Gated Stereo [90] reached state-of-the-art results using a stereo-gated setup and self-supervision [89].", + "bbox": [ + 75, + 90, + 468, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Neural Scene Reconstruction Recent research has amalgamated sets of single-sensor measurements to recover comprehensive scene representations. This synthesis has led to advancements in both novel-view generation [6, 23, 62, 64] and depth estimation [85], with neural radiance field methods [6, 23, 62, 64] emerging as a pivotal approach for representing scenes as continuous volumetric fields of radiance. These methods combine this representation with volumetric rendering as a forward model in a test-time optimization approach. Specific representations that these methods explore include coordinate-based networks [6, 7, 62, 101], 3D voxel-grid representation [23, 28, 98], or hybrid approaches [8, 64, 81]. Subsequent works have extended this representation to large outdoor scenes [7, 100], and increased the efficiency at training and test time [8, 23, 64, 98]. Other works departed from radiance-based representation and explicitly learned the scene illumination, geometry, and material properties [15, 74, 101]. A particular challenge within the field is reconstructing large urban terrains based on imagery captured from vehicles [37, 47, 56, 66, 71, 80, 86, 93, 97], given that a significant portion of the scene is seen from only a narrow range of viewpoints. This problem is tackled by additional supervision cues from sparse LiDAR [37, 66, 71, 86], pre-estimated depths [26, 37, 73], optical flow [61, 86] and semantic segmentation [47, 86, 93]. Departing from RGB-based approaches, recent works have investigated neural reconstruction methods using ToF sensors [41, 59, 84]. Existing methods [41, 84, 99] learn a neural field from posed LiDAR scans, allowing for the synthesis of realistic LiDAR scans from novel views. Recently, Malik et al. [59] represent the time-resolved photon count acquired by a single-photon LiDAR system with a neural reconstruction method.", + "bbox": [ + 75, + 199, + 470, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Gated Imaging Model", + "text_level": 1, + "bbox": [ + 76, + 709, + 285, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we provide a brief background on gated imaging as presented in [35] and introduce the proposed gated image formation model. Unlike prior work, this model incorporates shadow effects and features self-calibrated parameter learning.", + "bbox": [ + 75, + 734, + 468, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A gated imaging system, illustrated in Fig. 2, utilizes a pulsed flood-light illumination source $p$ with a synchronized imager that operates with a nanosecond (ns) gated exposure $g$ that is delayed by $\\xi$ compared to the pulse. This allows us to capture only photons with round-trip times inside the gates, hence specific distance segments in the scene.", + "bbox": [ + 75, + 810, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c84e875ac23349b146648aeea382ea7ea8a76e3c33dcaebf4a3f7b5271090e64.jpg", + "image_caption": [ + "Figure 2. Gated Image Formation and Bi-Directional Sampling. Top-row: Our test vehicle is equipped with a synchronized stereo camera setup and illuminator that flood-lits the scene with a light pulse and FoV $\\gamma$ . Using different gating profiles $C(z)$ , we capture three slices with intensity visualised here in red, green and blue. Illustrated in the middle row, the gating profiles describe pixel intensity for a point at sensor distance $z$ . The first slice (red) accounts for close ranges, the green for mid-ranges, and the blue for far ranges. Bottom-row: we show the ray sampling employed in our method, based on a bidirectional sampling strategy. We cast the rays from the illuminator view to explore the occluded areas, while the rays casted from the camera integrate the reflected scene response. The shadowed areas are marked in gray." + ], + "image_footnote": [], + "bbox": [ + 511, + 83, + 880, + 244 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We formalize this using so-called range intensity profiles $C_k(2z_c)$ given distance $z_{c}$ from the camera, time $t$ , and a parameter set $k$ [35], that is", + "bbox": [ + 498, + 450, + 890, + 494 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nC _ {k} \\left(2 z _ {c}\\right) = \\int_ {- \\infty} ^ {\\infty} g _ {k} (t - \\xi) p _ {k} \\left(t - \\frac {2 z _ {c}}{c}\\right) \\beta \\left(2 z _ {c}\\right) d t, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 516, + 507, + 890, + 553 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $c$ is the speed of light and $\\beta (\\cdot)$ models the distance-dependent decay of the reflected light pulse. The resulting gated pixel value is", + "bbox": [ + 498, + 566, + 890, + 612 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI _ {k} \\left(z _ {c}\\right) = \\alpha \\iota C _ {k} \\left(2 z _ {c}\\right) + \\Lambda + \\mathcal {D} _ {k}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 625, + 890, + 640 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\Lambda$ represents the passive ambient contribution, $\\alpha$ is the scene reflection, $\\iota$ the laser illumination, and $\\mathcal{D}_k$ is an additive noise term.", + "bbox": [ + 496, + 652, + 890, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This model assumes camera position $\\mathbf{o}_c$ and the illuminator position $\\mathbf{o}_i$ are collocated. To allow for noncollocated positions, we express the travel time as $z = z_c + z_i$ , where $z_i$ denotes the distance between the illuminator and the point on the surface impacted by the light beam, represented as $z_i = |\\mathbf{x} - \\mathbf{o}_i|_2$ . Additionally, there may be areas visible to the camera that remain dark due to potential occlusions. Modeling shadow effects and attenuation due to incident angle $\\omega$ results in the following image formation", + "bbox": [ + 496, + 698, + 890, + 834 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI _ {k} (z) = \\alpha \\iota \\psi | \\mathbf {n} \\cdot \\omega | C _ {k} (z) + \\Lambda + \\mathcal {D} _ {k}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 844, + 890, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $\\psi \\in [0,1]$ serves as a shadow indicator for the pixel, and $\\omega$ is the direction of the incident light at that point.", + "bbox": [ + 496, + 869, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10532", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We extend this model to fit the range intensity profiles during optimization, thereby eliminating the need for their direct measurement. This approach overcomes potential calibration inaccuracies encountered in previous approaches [89, 90]. We model both the laser pulse $p_k$ and the gate $g_k$ as rectangular functions with durations $t_{l,k}$ and $t_{g,k}$ , respectively. This simplification permits the analytical computation of the integral in Eq. (1), that is", + "bbox": [ + 75, + 90, + 472, + 212 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {C} _ {k} = \\left\\{ \\begin{array}{l l} \\frac {2 z}{c} - \\xi_ {k} + t _ {l, k} & \\text {i f} \\xi_ {k} - t _ {l, k} < \\frac {2 z}{c} < \\xi_ {k} \\\\ t _ {l, k} & \\text {i f} \\xi_ {k} < \\frac {2 z}{c} < \\xi_ {k} + t _ {g, k} - t _ {l, k} \\\\ - \\frac {2 z}{c} + \\xi_ {k} + t _ {g, k} & \\text {i f} \\xi_ {k} + l _ {g, k} - t _ {l, k} < \\frac {2 z}{c} < \\xi_ {k} + t _ {g, k} \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 222, + 477, + 308 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Gated Field", + "text_level": 1, + "bbox": [ + 76, + 311, + 200, + 327 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We reconstruct a scene by fitting a neural field representation to gated videos. We collect videos of three active gated slices $I_{k \\in \\{1,2,3\\}}$ with different gating parameters and one passive slice $I_P$ . We model active illumination by jointly estimating light and material properties, and separately represent the ambient light as a radiance field. The proposed reconstruction method relies on both photometric reconstruction cues and scene priors and is described in the following.", + "bbox": [ + 75, + 335, + 468, + 458 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Neural Gated Fields", + "text_level": 1, + "bbox": [ + 76, + 465, + 267, + 482 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We describe the scene properties using two neural fields $f_{Gp}$ and $f_{G\\alpha}$ , respectively representing the ambient light scattered in the scene and the reflection of the scene surfaces, conditioned on a spatial embedding $\\chi$ . Moreover, the laser illumination contribution is represented by a physics-based model, while shadow effects are simulated through ray-tracing using the volumetric density field $f_{Gd}$ .", + "bbox": [ + 75, + 489, + 468, + 597 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Neural Ambient and Reflection Field We represent a scene as a neural field $f_{G}:\\{\\mathbf{x},\\mathbf{d},\\omega \\} \\to \\{\\sigma ,\\alpha ,\\Lambda ,\\mathbf{n}\\}$ mapping each point in space $\\mathbf{x}$ viewed from a direction $\\mathbf{d}$ and laser direction $\\omega$ to its volumetric density $\\sigma$ , normal vector $\\mathbf{n}$ , scene reflection $\\alpha$ and the passive component $\\Lambda$ , that is", + "bbox": [ + 75, + 614, + 468, + 690 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {G p}: \\{\\mathbf {d}, \\chi \\} \\longrightarrow \\{\\Lambda \\}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 712, + 272, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Ambient Component", + "bbox": [ + 334, + 712, + 477, + 728 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {G n}: \\{\\mathbf {x}, \\chi \\} \\longrightarrow \\{\\mathbf {n} \\}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 731, + 269, + 747 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Surface Normal", + "bbox": [ + 367, + 732, + 475, + 744 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {G \\alpha}: \\{\\mathbf {d}, \\omega , \\chi \\} \\rightarrow \\{\\alpha \\}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 750, + 271, + 766 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Surface Reflection", + "bbox": [ + 351, + 751, + 475, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with $f_{Gd}:\\{\\mathbf{x}\\} \\longrightarrow \\{\\sigma ,\\chi \\}$ Vol. Density and Embedding", + "bbox": [ + 76, + 768, + 475, + 785 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We condition here normal, ambient and reflectance on a volumetric embedding via the field $f_{Gd}:\\{\\mathbf{x}\\} \\to \\{\\sigma ,\\chi \\}$ estimating the density $\\sigma$ and embedding $\\chi$ . This embedding is being shared by the network branches to estimate the normal with $f_{Gn}:\\{\\mathbf{x},\\chi \\} \\to \\{\\mathbf{n}\\}$ , scene reflection $f_{G\\alpha}:\\{\\mathbf{d},\\omega ,\\chi \\} \\to \\{\\alpha \\}$ , and ambient light component with $f_{Gp}:\\{\\mathbf{d},\\chi \\} \\to \\{\\Lambda \\}$ . An overview of the overall Neural", + "bbox": [ + 76, + 794, + 472, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Gated Fields is shown in Fig. 3. We also use a proposal sampler $f_{P}$ as in [7] for efficiency. Both $f_{G}$ and $f_{P}$ are MLPs (of different size) with multi-resolution hash encoding [64].", + "bbox": [ + 498, + 90, + 890, + 137 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Shadow and Illumination Model As the light pulse emitted by the illuminator is a diverging light beam, we model it as cone of light with irradiance maximum at the cross-section center and exponentially decreasing as it diverges from the center by the angles $\\gamma$ . As such, we express the illumination intensity as a 2D higher-order Gaussian $\\mathcal{G}$ with mean $\\Xi$ , standard deviation $\\Omega$ and power $\\Theta$", + "bbox": [ + 496, + 155, + 890, + 261 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\iota = \\eta \\mathcal {G} _ {i} (\\gamma ; \\boldsymbol {\\Xi}, \\boldsymbol {\\Omega}, \\boldsymbol {\\Theta}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 273, + 890, + 290 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\eta$ is a scaling parameter.", + "bbox": [ + 498, + 301, + 705, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Instead of predicting the shadow indicator $\\psi (\\mathbf{x})$ , we can directly estimate it using the density field, by computing the accumulated transmittance along the ray from the pixel to the point $\\mathbf{r}_{ill}(l) = \\mathbf{o}_i + \\omega l$", + "bbox": [ + 496, + 316, + 890, + 378 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\psi (\\mathbf {x}) = \\exp \\left(- \\int_ {0} ^ {l _ {\\mathbf {x}}} \\sigma \\left(\\mathbf {r} _ {i l l} (l)\\right) \\mathrm {d} l\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 388, + 890, + 429 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The illuminator origin and direction $\\mathbf{o}_i, \\omega$ are obtained from the camera as $[\\mathbf{o}_i, \\omega] = \\mathbf{R}[\\mathbf{o}_c, \\mathbf{d}_c] + \\mathbf{T}$ . During training, we jointly fine-tune the translation and rotation matrices $\\mathbf{T}, \\mathbf{R}$ , as well as $\\mathbf{o}_c$ . We also treat illuminator profile properties as learnable parameters, namely $\\eta, \\Xi, \\Omega, \\Theta$ and the gating parameters, i.e. number of accumulated laser pulses $m_k$ before read-out, laser pulse duration $t_{l,k}$ , camera exposure $t_{g,k}$ , and delay $\\xi_k$ between laser pulse emission and gated exposure for all three slices $k \\in \\{0,1,2\\}$ . In addition, we optimize a general distance offset $d_0$ for the range intensity profiles to compensate for internal signal processing delays.", + "bbox": [ + 496, + 439, + 890, + 607 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Gated Field Learning", + "text_level": 1, + "bbox": [ + 498, + 616, + 702, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We learn to acquire a gated capture with camera origin $\\mathbf{o}_c$ and direction $\\mathbf{d}$ by casting a ray $\\mathbf{r}(l) = \\mathbf{o}_c + l\\mathbf{d}$ for each pixel into the scene and computing the intensity $\\tilde{I}_k(\\mathbf{r})$ through volume rendering. Using the gated imaging formation model from Sec. 3, we define volume rendering as", + "bbox": [ + 496, + 638, + 890, + 715 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\tilde{I}_{k}(\\mathbf{r}) = \\int_{0}^{\\infty}T(l)\\sigma (\\mathbf{x})\\int_{-\\infty}^{\\infty}g_{k}(t - \\xi_{k})\\beta (l) \\\\ \\cdot \\left(\\alpha \\iota \\psi | \\mathbf {n} \\cdot \\omega | p _ {k} \\left(t - \\frac {l + l _ {i}}{c}\\right) + \\kappa\\right) d t d l + \\mathcal {D} _ {k}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 726, + 846, + 795 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\kappa$ is the ambient level, $T(l) = \\exp (-\\int_0^l\\sigma (u)du)$ is the accumulated transmittance along the ray and $l_{i}$ is the distance of $\\mathbf{x}(l)$ from the illuminator origin $\\mathbf{o}_i$ . As illustrated in Fig. 2, in our volume rendering formulation the pixel intensity contribution of a point along the ray depends not only on its accumulated transmittance and volumetric", + "bbox": [ + 496, + 808, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10533", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/579a4ac62f052897a054d43a79cababa260500aac920db1757c83436ee8c2a32.jpg", + "image_caption": [ + "Figure 3. Neural Gated Fields. For any point in space $\\mathbf{x}$ , we learn its volumetric density $\\sigma$ , normal $\\mathbf{n}$ , reflectance $\\alpha$ and ambient lighting $\\Lambda$ through four neural fields, conditioned on direction $\\mathbf{d}$ , incident laser light direction $\\omega$ and spatial embedding $\\chi$ . The illuminator light $\\iota$ is represented by a physics-based model dependent on the displacement angle $\\gamma$ , while the gating imaging process is described by the range intensity profiles $\\tilde{C}(z)$ using as input the camera-point-laser distance $z$ , as explained in Sec. 3. With this information, we reconstruct a gated image $I_k$ through the gated volume rendering formulation introduced in Sec. 4.2. As this process is fully differentiable, we simultaneously fit neural fields and physical parameters through image reconstruction together with other regularization losses discussed in Sec. 4.3." + ], + "image_footnote": [], + "bbox": [ + 138, + 75, + 831, + 233 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "density, but also on its distance from the illuminator and camera origins through $C(z)$ , as well as on its relative position to the illuminator source via $\\iota$ and $\\psi$ .", + "bbox": [ + 75, + 334, + 468, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We simplify the time-dependent integral using Eq. (4) as", + "bbox": [ + 76, + 380, + 452, + 393 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\tilde {I} _ {k} (\\mathbf {r}) = \\int_ {0} ^ {\\infty} T (l) \\sigma (\\mathbf {x}) \\left(\\alpha (\\mathbf {x}, \\mathbf {d}, \\omega) \\tilde {C} _ {k} \\left(l + l _ {i}\\right) \\psi (\\mathbf {x}) \\right. \\tag {8} \\\\ \\cdot | \\mathbf {n} \\cdot \\omega | \\iota (\\gamma) + \\Lambda (\\mathbf {x}, \\mathbf {d})) \\mathrm {d} l + \\mathcal {D} _ {k}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 393, + 468, + 445 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We numerically estimate this spatial integral by numerical quadrature [62, 88], approximating it with a set of points $\\mathbf{X}_{ray}$ . Specifically, for each point $\\mathbf{x}_j \\in \\mathbf{X}_{ray}$ , we query the neural field $f_{G}$ to infer the normal vector $\\mathbf{n}_j$ , scene reflection $\\alpha_j$ , volumetric density $\\sigma_j$ and ambient light component $\\Lambda_j$ . The laser illumination intensity $\\iota_j$ is instead computed following the physics-based model defined in Eq. (5). The gated intensity $\\tilde{I}_k$ is then expressed as", + "bbox": [ + 76, + 450, + 468, + 571 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\tilde {I} _ {k} (\\mathbf {r}) = \\sum_ {j = 0} ^ {N} w _ {j} \\left(\\underbrace {\\alpha_ {j} \\tilde {C} _ {j} \\psi_ {j} | \\mathbf {n} _ {j} \\cdot \\omega_ {j} | \\iota_ {j}} _ {\\text {A c t i v e C o m p o n e n t}} + \\underbrace {\\Lambda_ {j}} _ {\\text {P a s s i v e C o m p o n e n t}}\\right) + \\mathcal {D} _ {k}, (9) \\\\ w _ {j} = \\exp \\left(- \\sum_ {k = 1} ^ {j - 1} \\sigma_ {k} \\delta_ {k}\\right) \\left(1 - \\exp \\left(- \\sigma_ {j} \\delta_ {j}\\right)\\right). (9) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 570, + 468, + 671 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The shadow indicator $\\psi_{j}$ from Eq. (6) is similarly approximated by sampling on $\\mathbf{r}_{ill} = \\mathbf{o}_i + \\omega_j l$ a set of points $\\mathbf{X}_{ill}$ bounded between $\\mathbf{o}_i$ and $\\mathbf{x}_j$", + "bbox": [ + 76, + 675, + 468, + 719 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\psi_ {j} = \\exp \\left(- \\sum_ {k} ^ {\\mathbf {X} _ {i l l}} \\sigma_ {k} \\delta_ {k}\\right). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 718, + 468, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the passive slice, the active component is null, further simplifying to", + "bbox": [ + 76, + 765, + 468, + 794 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {I} _ {P} (\\mathbf {r}) = \\sum_ {j = 0} ^ {N} w _ {j} \\Lambda_ {j} + \\mathcal {D} _ {P}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 792, + 468, + 833 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Both $\\mathbf{X}_{ray}$ and $\\mathbf{X}_{ill}$ are sampled using a proposal network [7] $f_{P}$ that, analogously to $f_{Gd}$ , predicts point-wise densities converted with Eq. (10) to proposal weights $\\hat{w}$ for sampling with piece-wise-constant probabilities.", + "bbox": [ + 76, + 839, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Training Supervision", + "text_level": 1, + "bbox": [ + 498, + 334, + 697, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We supervise the predicted passive and active gated frames applying a photometric loss $\\mathcal{L}_c$ , regularize the volumetric density with a depth loss $\\mathcal{L}_d$ and by supervising the shadow estimate with $\\mathcal{L}_s$ . We regularize normal and reflectance estimates through $\\mathcal{L}_{nc}$ and $\\mathcal{L}_{\\alpha}$ , respectively.", + "bbox": [ + 496, + 349, + 890, + 425 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Photometric Loss We supervise with ground truth captures for active and passive gated slice reconstruction as", + "bbox": [ + 498, + 426, + 890, + 455 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c} = \\sum_ {k, r} \\| \\tilde {I} _ {k} (\\mathbf {r}) - I _ {k} (\\mathbf {r}) \\| _ {2} + \\sum_ {r} \\| \\tilde {I} _ {P} (\\mathbf {r}) - I _ {P} (\\mathbf {r}) \\| _ {2} \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 457, + 890, + 488 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Volume Density Regularization As additional training supervision, we use the depth estimate $\\hat{D}(\\mathbf{r})$ of a pretrained stereo depth estimation algorithm [90] as pseudo ground-truth to regularize the ray termination distribution [26]", + "bbox": [ + 496, + 497, + 890, + 558 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d} = \\sum_ {\\mathbf {r}} \\sum_ {j} \\log w _ {j} \\exp \\left(- \\frac {\\left(l _ {j} - \\hat {D} (\\mathbf {r})\\right) ^ {2}}{2 s ^ {2}}\\right) \\delta_ {i} \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 558, + 890, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We regularize the density field by partially supervising the shadow indicator $\\psi$ . Each pixel whose active intensity $I_{kA} = I_k - I_P$ in any of the three gated slices is above a certain threshold $\\epsilon_i$ is considered as visible from the illuminator. We hence supervise the expected shadow value for such rays $\\mathbf{r}_v \\in \\{\\mathbf{r} | \\forall k \\in \\{1, 2, 3\\} : I_{kA}(\\mathbf{r}) > \\epsilon_i\\}$ as", + "bbox": [ + 496, + 607, + 890, + 698 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s} = \\sum_ {\\mathbf {r} _ {v}} \\| 1 - \\int T (l) \\sigma (\\mathbf {x}) \\psi (\\mathbf {x}) d l \\| _ {2} \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 699, + 890, + 732 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Normals Consistency Following [88], for each sampled point $\\mathbf{x}$ we enforce a consistency between the predicted normal $\\mathbf{n}$ and the density gradient $\\hat{\\mathbf{n}} (\\mathbf{x}) = -\\frac{\\nabla\\mathbf{x}}{||\\nabla\\mathbf{x}||}$ , and we penalize normals which are back-facing the camera as", + "bbox": [ + 496, + 734, + 890, + 796 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {n c} = \\sum_ {\\mathbf {x}} w (\\mathbf {x}) \\left(\\| \\mathbf {n} (\\mathbf {x}) - \\hat {\\mathbf {n}} (\\mathbf {x}) \\| _ {2} + \\max (0, \\hat {\\mathbf {n}} (\\mathbf {x}) \\cdot \\mathbf {d}) ^ {2}\\right) \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 797, + 890, + 827 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Reflectance Regularization We enforce the predicted scene reflection $\\alpha$ to be spatially consistent within $\\epsilon_{\\mathbf{x}}$ , i.e.", + "bbox": [ + 498, + 829, + 890, + 859 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\alpha} = \\sum_ {\\mathbf {x}} w (\\mathbf {x}) \\left(\\| \\alpha (\\mathbf {x}, \\mathbf {d}) - \\alpha \\left(\\mathbf {x} + \\epsilon_ {\\mathbf {x}}, \\mathbf {d} + \\epsilon_ {\\mathbf {d}}\\right) \\| _ {2}\\right). \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 872, + 890, + 904 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10534", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here the $\\omega$ dependency of $\\alpha$ is omitted here for brevity. We also include an angular noise $\\epsilon_{\\mathrm{d}}$ that we set high at the beginning of the training and then decrease it exponentially. By forcing the reflectance to behave as fully diffuse at the beginning of the training, we disincentive it to bake in the effects from lighting or shadowing, hence improving the disjoint learning of the scene components.", + "bbox": [ + 75, + 90, + 472, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Total Training Loss Combining the different losses we obtain the following loss formulation:", + "bbox": [ + 76, + 199, + 468, + 229 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {1} \\mathcal {L} _ {c} + \\lambda_ {2} \\mathcal {L} _ {d} + \\lambda_ {3} \\mathcal {L} _ {s} + \\lambda_ {4} \\mathcal {L} _ {n c} + \\lambda_ {5} \\mathcal {L} _ {\\alpha}, \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 241, + 468, + 258 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "see $\\lambda_{1,\\dots,5}$ hyperparameters in the Supplementary Material.", + "bbox": [ + 76, + 268, + 468, + 284 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 289, + 299, + 306 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We train for 35,000 steps and a batch size of 4096 rays. As optimizer we use ADAMW [57] with $\\beta_{1} = 0.9$ , $\\beta_{2} = 0.999$ , learning rate $10^{-2}$ for $f_{P}$ and $f_{G}$ , $10^{-4}$ for the camera poses optimization, $10^{-4}$ for the laser profile and gated parameters. We train on two NVIDIA V100 GPUs, for approximately 3 hours. The proposal network $f_{P}$ is comprised of two MLPs and trained following [7]. Additional architecture details, training procedures, and hyper-parameters are found in the Supplementary Material.", + "bbox": [ + 75, + 308, + 468, + 444 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Dataset", + "text_level": 1, + "bbox": [ + 76, + 450, + 165, + 465 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To conduct this work, we have collected a diverse set of 10 static sequences, recorded in both day and night conditions across North America. To this goal, we equipped a test vehicle with a NIR gated stereo camera setup (Bright-Way Vision), an automotive RGB stereo camera (OnSemi AR0230), a LiDAR sensor (Velodyne VLS128) and a GNSS with IMU (Xsens MTi-7), as shown in Fig. 4. Each gated camera has a resolution of $1280 \\times 720$ pixels, 10 bit depth and runs at $120 \\mathrm{~Hz}$ , split up to collect the three active and one passive slice. The illuminator source consists of two vertical-cavity surface-emitting laser (VCSEL) modules, which illuminate the scene with a laser pulse with duration of 240-370 ns and a wavelength of $808 \\mathrm{~nm}$ . The RGB cameras provide 12 bit HDR images with resolution of $1920 \\times 1080$ pixels and $30 \\mathrm{~Hz}$ frame-rate. The LiDAR has a vertical resolution of 128 lines and $10 \\mathrm{~Hz}$ framerate, while the GNSS sensor runs at $4 \\mathrm{~Hz}$ . Example captures from the dataset are being visualized in Fig. 4. In total, we collect 2650 samples, captured in both day (1223 samples) and night (1427 samples). We divide it in training, validation and test splits with a 50-25-25 split.", + "bbox": [ + 75, + 468, + 472, + 785 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As ground-truth, we construct a large-scale ground-truth pointcloud by aggregating LiDAR scans with LIO-SAM [78] and removing noisy points. Additional details on the accumulation are provided in the Supplementary Material.", + "bbox": [ + 76, + 786, + 468, + 845 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "7. Assessment", + "text_level": 1, + "bbox": [ + 76, + 852, + 197, + 867 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we validate the proposed method quantitatively and qualitatively. Specifically, we investigate scene", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3aa04838ebe91debfd13421c7886c9598b95b0ec86127b16edc92eeada4ab79a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 498, + 75, + 903, + 256 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5643129c4f14fc3124cc751a35972b2d7735a1efa25bf35b89dcc5d3f45a9d71.jpg", + "image_caption": [ + "Figure 4. Top: Example captures from our collected dataset across different urban and suburban areas in North America. From left to right: RGB image, active gated slices (with red for slice 1, green for slice 2 and blue for slice 3), passive slice, projected LiDAR scan, accumulated LiDAR. Bottom: Sensors setup with LiDAR, stereo Gated camera, stereo RGB camera, IMU and GNSS." + ], + "image_footnote": [], + "bbox": [ + 511, + 258, + 898, + 401 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "reconstruction at both day and night, using novel depth and view synthesis for 2D evaluation, and surface reconstruction for the 3D evaluation. To this end, we compare our approach to state-of-the-art feed-forward depth estimation algorithms and neural scene reconstruction methods. We also conduct ablation experiments to validate our design choices.", + "bbox": [ + 496, + 508, + 890, + 601 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Depth Reconstruction We assess the quality of depth synthesis of Gated Fields for camera poses unseen during training. We use as ground truth the accumulated and filtered LiDAR pointcloud. Unlike previous works relying on single LiDAR scans for evaluation [35, 89, 90], we use as ground truth an accumulated LiDAR pointcloud as described in Sec. 6. This allows us to evaluate the depth reconstruction up to $160\\mathrm{m}$ accurately and without bias. We follow previous works [89, 90] and use as depth evaluation metrics RMSE, MAE, ARD, $\\sigma_{i} < 1.25^{i}, i \\in \\{1,2,3\\}$ . We compare our method against 9 feed-forward depth estimation methods, namely SimIPU [53], AdaBins [10], DPT [70], DepthFormer [54] and CREStereo [50] for monocular and stereo RGB methods, Gated2Gated [89] and GatedStereo [90] for monocular and stereo gated estimation methods. We also compare our method against depths rendered with other neural reconstruction algorithms, using RGB images [7],[63],[29],[26] LiDAR [84], RGB+LiDAR [86], [37], and a varying-appearance method [29] for gated", + "bbox": [ + 496, + 613, + 892, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10535", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ab0701162834ba87697bca1bb1490b5b0aebca8011d2d528ffea7c1506e19800.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison of the proposed Gated Fields and state-of-the-art depth estimation approaches, including LiDAR-NeRF [84], StreetSurf [37], NLSPN [67], and Gated Stereo [90]. Compared to baseline methods, we are able to reconstruct fine geometry details like branches or poles, also for far distances. Unlike RGB methods, Gated Fields is unaffected by poor ambient lighting, and unlike LiDAR-based methods it is able to reconstruct sharp object discontinuities. The active gated slices are visualized in red for slice 1, green for slice 2 and blue for slice 3." + ], + "image_footnote": [], + "bbox": [ + 76, + 66, + 898, + 483 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "captures. Results for day and night sequences are presented in Tab. 1. We outperform the next best neural field method by $21.87\\%$ MAE and $30.35\\%$ in RMSE. For night sequences the performance difference sharpens, with Gated Fields outperforming the best RGB-based method [50] by $3.14\\mathrm{m}$ MAE. This performance decline is to be attributed to the limited pixel information present in RGB captures taken at night time, making impossible to learn a meaningful 3D representation of the scene, as shown qualitatively in Fig. 5. LiDAR-based methods are unaffected by the change in illumination, but suffer from the limited sensor resolution. On the other hand, gated cameras retrieve information-rich captures both day and night, which Gated Fields can explicitly leverage during training. We confirm that employing state-of-the-art neural field methods on gated captures does not yield accurate results, as they are not able to model the gated imaging formation and can only fit the ambient light component.", + "bbox": [ + 75, + 578, + 472, + 851 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3D Reconstruction We evaluate the 3D scene reconstruction capabilities of Gated Fields using the accumulated Li-DAR pointcloud as ground truth. We follow [20] and extract", + "bbox": [ + 76, + 854, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "for both the 3D ground truth pointcloud and different neural field-based methods a voxelized occupancy grid of the scene, and compute intersection over union (IoU), Precision and Recall between ground truth voxels and estimated ones. Quantitative results are shown in Tab. 3. Our method outperforms RGB baselines [7, 29] by an average of $15\\%$ IoU. RGB baselines using additional LiDAR sensor [37, 86] data partially improve the results, but such methods are still unable to reconstruct finer surfaces details and struggle at night. On the other hand, Gated Fields is able to recover finer geometries using gated, illumination and depth cues, and the quality does not degrade with diminishing ambient light, as shown in Fig. 5. See details on evaluation and further qualitative results in the Supplementary Material.", + "bbox": [ + 496, + 578, + 890, + 790 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Novel View Synthesis For novel view synthesis, we compare our method with Mip-NeRF360 [7], a state-of-the-art neural radiance field-based method, and K-Planes [29], to implicitly model the time-varying appearance of the static scene. Mip-NeRF [7] struggles to reconstruct novel views due to the inherent difficulty of modeling the gating imaging effects, resulting in a PSNR of 17.16dB. By learning", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10536", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a2cef6d30e7386f3bcf1d101e0dfd3edfd5170e0082db277465bc71addf5309a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodModalityRMSE [m]ARDMAE [m]δ1 [%]δ2 [%]δ3 [%]
Test Data - Night (Evaluated on Accumulated LiDAR Ground-Truth Points)
2D DEPTH COMPARISONGATED2GATED [89]Gated13.330.288.5761.7890.3094.90
GATEDSTEREO [90]Stereo-Gated10.100.205.9782.8693.3596.52
SIMIPU [53]RGB19.330.4414.2140.6777.9989.80
ADABINS [10]RGB21.140.3814.2851.7280.0090.64
DPT [70]RGB14.170.289.9062.8288.3294.42
DEPTHFORMER [54]RGB14.320.2910.0861.1987.7294.57
CRESTEREO [50]Stereo-RGB14.220.137.2782.2190.1094.32
NLSPN [67]RGB+LiDAR11.120.186.4877.1590.8196.00
MIPNERF360 [7]RGB23.800.5116.4341.6061.7576.16
K-PLANES [29]RGB19.700.4113.6644.3966.2081.46
RAWNERF [63]RGB27.460.6419.7534.3254.3369.41
DEPTH-NERF [26]RGB15.230.2610.0461.6786.2393.87
SUDS [86]RGB+LiDAR11.070.176.1779.4988.4195.31
STREETSURF [37]RGB+LiDAR10.860.155.9082.1693.5796.96
LiDAR-NERF [50]LiDAR10.210.124.7287.7195.0597.78
GATED FIELDS [50]Gated7.920.124.1390.6195.7697.90
Test Data - Day (Evaluated on Accumulated LiDAR Ground-Truth Points)
2D DEPTH COMPARISONGATED2GATED [89]Gated9.260.226.6958.4693.7097.38
GATEDSTEREO [90]Stereo-Gated6.320.093.3092.8697.3198.53
SIMIPU [53]RGB13.540.3110.0852.9086.4995.65
ADABINS [10]RGB12.740.258.3969.8489.1395.52
DPT [70]RGB10.340.217.0877.6194.2997.24
DEPTHFORMER [54]RGB9.060.196.0981.1994.1497.42
CRESTEREO [50]Stereo-RGB7.350.093.4594.4897.2398.50
NLSPN [67]RGB+LiDAR10.340.175.9777.8391.1696.11
MIPNERF360 [7]RGB16.910.319.5370.8984.4390.92
K-PLANES [29]RGB12.370.248.5563.1679.6891.70
RAWNERF [63]RGB15.100.239.3865.9084.9092.43
DEPTH-NERF [26]RGB10.340.176.0778.9790.7596.53
SUDS [86]RGB+LiDAR9.110.175.8480.9695.0898.33
STREETSURF [37]RGB+LiDAR9.600.135.3583.9495.3298.56
LiDAR-NERF [50]LiDAR8.130.103.8688.9995.4998.06
GATED FIELDS [50]Gated6.150.092.9193.8897.3298.75
", + "bbox": [ + 81, + 79, + 460, + 496 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e5de8884bfc78b9911c769f2f3914d7afef4cb26af82c59abc4b045ff63b49b8.jpg", + "table_caption": [ + "Table 1. Comparison of our proposed approach and state-of-the-art approaches on depth synthesis. Best results in each category are in bold and second best are underlined." + ], + "table_footnote": [], + "table_body": "
Im. Formation ModelDepth Sup.Normal Active Illum.RMSE [m]MAE [m]PSNR [dB]SSIM
Test Data - Night
ABLATIONEnd2End [62]XXXX27.3419.9517.430.633
GatedXXXX9.348.3523.470.889
Gated✓XXX6.725.4325.840.908
Gated✓✓✓X3.782.8230.660.946
Gated✓✓✓✓3.392.5130.910.95
Test Data - Day
ABLATIONEnd2End [62]XXXX30.0319.2216.770.66
GatedXXXX11.646.3426.770.915
Gated✓XXX9.204.2226.880.922
Gated✓✓✓X9.454.3132.150.946
Gated✓✓✓✓8.884.1232.280.948
", + "bbox": [ + 99, + 558, + 444, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "a time-varying appearance, K-Planes improves the quality reaching 27.42dB PSNR for day but only 19.35dB for night, as the model fails to learn an accurate scene geometry representation without ambient light information. Gated Fields outperforms these baselines in both day and night, reaching a PSNR of 32.28dB.", + "bbox": [ + 75, + 776, + 467, + 864 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation Experiments To assess the role and contribution of the different components of our method, we conduct", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b3ae9d67f26df991927204f6b3d71488813fd1a7602935decd83d7567ea55106.jpg", + "table_caption": [ + "Table 2. Ablation studies of the Gated Fields contributions, on a subset of the test dataset. We investigate different image formation models, neural fields components and supervision losses." + ], + "table_footnote": [], + "table_body": "
MethodModalityIoU [%]Precision [%]Recall [%]
3D REC.MIPNERF360 [7]RGB6.327.3431.21
STREETSURF [37]RGB+LiDAR5.416.3127.35
SUDS [86]RGB+LiDAR8.969.8849.09
LiDAR-NERF [84]LiDAR20.0332.3834.44
Gated Fields (ours)Gated22.2525.0166.51
", + "bbox": [ + 535, + 77, + 851, + 162 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Comparison of Gated Fields and state-of-the-art scene reconstruction methods. We evaluate over 3D occupancy reconstruction, using as ground truth the voxelized accumulated LiDAR pointcloud. Best results in each category are in bold and second best are underlined.", + "bbox": [ + 498, + 167, + 890, + 237 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "an ablation study in Tab. 2. In particular, we consider as starting point a single neural field directly inferring one intensity value for each of the four slices (3 active + 1 passive), and obtain an average MAE of $19.58\\mathrm{m}$ . By separately predicting ambient light and reflectance, and reconstructing the gated image as in Eq. (2), we significantly improve the MAE to $7.34\\mathrm{m}$ . However, this approach still performs poorly on flat-color areas during the day and in unilluminated areas during the night due to lack of any depth cue. By adding the depth supervision, we are able to supervise also such areas and the PSNR improves by $4.83\\mathrm{dB}$ . By adding the angular-dependent attenuation and regularizing the reflectance in Eq. (17), the model is able to disentangle the material properties from other spurious effects. Finally, by explicitly modeling the shadow, casted by the illuminator, we improve final depth reconstruction to $3.32\\mathrm{m}$ MAE.", + "bbox": [ + 496, + 250, + 890, + 492 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 506, + 617, + 522 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce Gated Fields, a neural rendering method capable of reconstructing scene geometry from video captures of active time-gated cameras. The method hinges on a differentiable gated image formation as part of the rendering formulation, and it jointly learns geometry, ambient light and surface proprieties, represented implicitly as neural field components, alongside illumination and gating parameters, represented with physics-based models. Extensive experiments on real-world large-scale scenes validate that our method is able to precisely reconstruct a 3D scene both in day and night-time conditions. Our approach outperforms existing RGB and LiDAR methods by $21.87\\%$ on MAE, as well as baseline methods using gated captures by $31.67\\%$ . In the future, we hope to extend the proposed approach by \"closing the loop\" and providing dynamic feedback to the gated acquisition, allowing for adaptive gated scene reconstructions.", + "bbox": [ + 496, + 530, + 890, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments This work was supported by the AI-SEE project with funding from the FFG, BMBF, and NRC-IRA. Felix Heide was supported by an NSF CAREER Award (2047359), a Packard Foundation Fellowship, a Sloan Research Fellowship, a Sony Young Faculty Award, a Project X Innovation Award, and an Amazon Science Research Award.", + "bbox": [ + 496, + 792, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10537", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Amit Adam, Christoph Dann, Omer Yair, Shai Mazor, and Sebastian Nowozin. Bayesian time-of-flight for realtime shape, illumination and albedo. 39(5):851-864, 2017. 3", + "[2] Ashutosh Agarwal and Chetan Arora. Attention attention everywhere: Monocular depth prediction with skip attention. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5861-5870, 2023. 2", + "[3] Pierre Andersson. Long-range three-dimensional imaging using range-gated laser radar images. 45(3):034301, 2006. 3", + "[4] Benjamin Attal, Eliot Laidlaw, Aaron Gokaslan, Changil Kim, Christian Richardt, James Tompkin, and Matthew O'Toole. Törf: Time-of-flight radiance fields for dynamic scene view synthesis. Advances in neural information processing systems, 34:26289-26301, 2021. 1", + "[5] Abhishek Badki, Alejandro Troccoli, Kihwan Kim, Jan Kautz, Pradeep Sen, and Orazio Gallo. Bi3D: Stereo depth estimation via binary classifications. In arXiv preprint arXiv:2005.07274, 2020. 2", + "[6] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. ICCV, 2021. 3", + "[7] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 1, 3, 4, 5, 6, 7, 8", + "[8] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. ICCV, 2023. 1, 3", + "[9] Florent Bartoccioni, Éloi Zablocki, Patrick Pérez, Matthieu Cord, and Karteek Alahari. Lidartouch: Monocular metric depth estimation with a few-beam lidar. Computer Vision and Image Understanding, 227:103601, 2023. 2", + "[10] Shariq Farooq Bhat, Ibraheem Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4009-4018, 2021. 6, 8", + "[11] Mario Bijelic, Tobias Gruber, and Werner Ritter. A benchmark for lidar sensors in fog: Is detection breaking down? In 2018 IEEE Intelligent Vehicles Symposium (IV), pages 760-767, 2018. 2", + "[12] Mario Bijelic, Tobias Gruber, and Werner Ritter. Benchmarking image sensors under adverse weather conditions for autonomous driving. 2018. 2", + "[13] Mario Bijelic, Tobias Gruber, and Werner Ritter. Benchmarking image sensors under adverse weather conditions for autonomous driving. In 2018 IEEE Intelligent Vehicles Symposium (IV), pages 1773-1779. IEEE, 2018. 2", + "[14] Mario Bijelic, Tobias Gruber, Fahim Mannan, Florian Kraus, Werner Ritter, Klaus Dietmayer, and Felix Heide. Seeing through fog without seeing fog: Deep multimodal sensor fusion in unseen adverse weather. In The IEEE" + ], + "bbox": [ + 86, + 114, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2", + "[15] Mark Boss, Andreas Engelhardt, Abhishek Kar, Yuanzhen Li, Deqing Sun, Jonathan Barron, Hendrik Lensch, and Varun Jampani. *Samurai: Shape and material from unconstrained real-world arbitrary image collections*. Advances in Neural Information Processing Systems, 35:26389-26403, 2022. 3", + "[16] Aljaz Bozic, Pablo Palafox, Justus Thies, Angela Dai, and Matthias Nießner. Transformerfusion: Monocular rgb scene reconstruction using transformers. Advances in Neural Information Processing Systems, 34:1403-1414, 2021. 1", + "[17] Jens Busck. Underwater 3-D optical imaging with a gated viewing laser radar. 2005. 3", + "[18] Jens Busck and Henning Heiselberg. Gated viewing and high-accuracy three-dimensional laser radar. 43(24):4705–10, 2004. 3", + "[19] Jens Busck and Henning Heiselberg. Gated viewing and high-accuracy three-dimensional laser radar. Applied optics, 43(24):4705-4710, 2004. 2", + "[20] Anh-Quan Cao and Raoul de Charette. Scenerf: Self-supervised monocular 3d scene reconstruction with radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9387-9398, 2023. 7", + "[21] A. Carballo, J. Lambert, A. Monrroy, D. Wong, P. Narksri, Y. Kitsukawa, E. Takeuchi, S. Kato, and K. Takeda. Libre: The multiple 3d lidar dataset. In IEEE Intelligent Vehicles Symposium (IV), 2020. 2", + "[22] Jia-Ren Chang and Yong-Sheng Chen. Pyramid stereo matching network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5410-5418, 2018. 2", + "[23] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 3", + "[24] Jaesung Choe, Kyungdon Joo, Tooba Imtiaz, and In So Kweon. Volumetric propagation network: Stereo-lidar fusion for long-range depth estimation. IEEE Robotics and Automation Letters, 6(3):4672-4679, 2021. 2", + "[25] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 303-312, 1996. 1", + "[26] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nef: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12882-12891, 2022. 3, 5, 6, 8", + "[27] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. In NeurIPS, pages 2366-2374, 2014. 2", + "[28] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of" + ], + "bbox": [ + 509, + 93, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10538", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501-5510, 2022. 3", + "[29] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12479–12488, 2023. 6, 7, 8", + "[30] Ravi Garg, B.G. Vijay Kumar, Gustavo Carneiro, and Ian Reid. Unsupervised CNN for single view depth estimation: Geometry to the rescue. In ECCV, pages 740-756, 2016. 2", + "[31] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, pages 3354-3361, 2012. 2", + "[32] Clément Godard, Oisin Mac Aodha, and Gabriel J Bros-tow. Unsupervised monocular depth estimation with left-right consistency. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 270-279, 2017. 2", + "[33] Clément Godard, Oisin Mac Aodha, Michael Firman, and Gabriel J Brostow. Digging into self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3828-3838, 2019. 2", + "[34] Yoav Grauer. Active gated imaging in driver assistance system. Advanced Optical Technologies, 3(2):151-160, 2014. 2", + "[35] Tobias Gruber, Frank Julca-Aguilar, Mario Bijelic, and Felix Heide. Gated2depth: Real-time dense lidar from gated images. In The IEEE International Conference on Computer Vision (ICCV), 2019. 3, 6", + "[36] Vitor Guizilini, Rares Ambrus, Sudeep Pillai, Allan Ravens \ntos, and Adrien Gaidon. 3d packing for self-supervised \nmonocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern \nRecognition, pages 2485-2494, 2020. 2", + "[37] Jianfei Guo, Nianchen Deng, Xinyang Li, Yeqi Bai, Botian Shi, Chiyu Wang, Chenjing Ding, Dongliang Wang, and Yikang Li. Streetsurf: Extending multi-view implicit surface reconstruction to street views. arXiv preprint arXiv:2306.04988, 2023. 1, 3, 6, 7, 8", + "[38] Miles Hansard, Seungkyu Lee, Ouk Choi, and Radu Patrice Horaud. Time-of-flight cameras: principles, methods and applications. Springer Science & Business Media, 2012. 2", + "[39] Paul Heckman and Robert T. Hodgson. Underwater optical range gating. 3(11):445-448, 1967. 2", + "[40] Mu Hu, Shuling Wang, Bin Li, Shiyu Ning, Li Fan, and Xiaojin Gong. Towards precise and efficient image guided depth completion. 2021. 2", + "[41] Shengyu Huang, Zan Gojcic, Zian Wang, Francis Williams, Yoni Kasten, Sanja Fidler, Konrad Schindler, and Or Litany. Neural lidar fields for novel view synthesis. 2023. 1, 3", + "[42] Maximilian Jaritz, Raoul De Charette, Emilie Wirbel, Xavier Perrotton, and Fawzi Nashashibi. Sparse and dense data with cnns: Depth completion and semantic segmentation. pages 52-60, 2018. 2" + ], + "bbox": [ + 86, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Maria Jokela, Matti Kutila, and Pasi Pyykönen. Testing and validation of automotive point-cloud sensors in adverse weather conditions. Applied Sciences, 9, 2019. 2", + "[44] Frank Julca-Aguilar, Jason Taylor, Mario Bijelic, Fahim Mannan, Ethan Tseng, and Felix Heide. Gated3d: Monocular 3d object detection from temporal illumination cues. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2938-2948, 2021. 2", + "[45] Alex Kendall, Hayk Martirosyan, Saumitro Dasgupta, Peter Henry, Ryan Kennedy, Abraham Bachrach, and Adam Bry. End-to-end learning of geometry and context for deep stereo regression. In ICCV, 2017. 2", + "[46] Andreas Kolb, Erhardt Barth, Reinhard Koch, and Rasmus Larsen. Time-of-flight cameras in computer graphics. In Computer Graphics Forum, pages 141-159. Wiley Online Library, 2010. 2", + "[47] Abhijit Kundu, Kyle Genova, Xiaqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12871-12881, 2022. 3", + "[48] Robert Lange. 3D time-of-flight distance measurement with custom solid-state image sensors in CMOS/CCD-technology. 2000. 2", + "[49] Deborah Levy, Amit Peleg, Naama Pearl, Dan Rosenbaum, Derya Akkaynak, Simon Korman, and Tali Treibitz. Seathru-nerf: Neural radiance fields in scattering media. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 56–65, 2023. 1", + "[50] Jiankun Li, Peisen Wang, Pengfei Xiong, Tao Cai, Ziwei Yan, Lei Yang, Jiangyu Liu, Haoqiang Fan, and Shuaicheng Liu. Practical stereo matching via cascaded recurrent network with adaptive correlation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16263-16272, 2022. 6, 7, 8", + "[51] Zhengqi Li, Tali Dekel, Forrester Cole, Richard Tucker, Noah Snavely, Ce Liu, and William T. Freeman. Manequinchallenge: Learning the depths of moving people by watching frozen people. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(12):4229-4241, 2021. 2", + "[52] Zhaoshuo Li, Xingtong Liu, Nathan Drenkow, Andy Ding, Francis X. Creighton, Russell H. Taylor, and Mathias Unberath. Revisiting stereo depth estimation from a sequence-to-sequence perspective with transformers. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6177-6186, 2021. 2", + "[53] Zhenyu Li, Zehui Chen, Ang Li, Liangji Fang, Qinhong Jiang, Xianming Liu, Junjun Jiang, Bolei Zhou, and Hang Zhao. Simipu: Simple 2d image and 3d point cloud unsupervised pre-training for spatial-aware visual representations. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1500-1508, 2022. 6, 8", + "[54] Zhenyu Li, Zehui Chen, Xianming Liu, and Junjun Jiang. Depthformer: Depthformer: Exploiting long-range corre" + ], + "bbox": [ + 509, + 90, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10539", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "lation and local information for accurate monocular depth estimation. arXiv preprint arXiv:2203.14211, 2022. 2, 6, 8", + "[55] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1", + "[56] Jeffrey Yunfan Liu, Yun Chen, Ze Yang, Jingkang Wang, Sivabalan Manivasagam, and Raquel Urtasun. Real-time neural rasterization for large scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8416-8427, 2023. 3", + "[57] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. International Conference on Learning Representations, 2018. 6", + "[58] Fangchang Ma and Sertac Karaman. Sparse-to-dense: Depth prediction from sparse depth samples and a single image. pages 1-8, 2018. 2", + "[59] Anagh Malik, Parsa Mirdehghan, Sotiris Nousias, Kiriakos N Kutulakos, and David B Lindell. Transient neural radiance fields for lidar view synthesis and 3d reconstruction. arXiv preprint arXiv:2307.09555, 2023. 1, 3", + "[60] N. Mayer, E. Ilg, P. Häusser, P. Fischer, D. Cremers, A. Dosovitskiy, and T. Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In CVPR, 2016. 2", + "[61] Andreas Meuleman, Yu-Lun Liu, Chen Gao, Jia-Bin Huang, Changil Kim, Min H. Kim, and Johannes Kopf. Progressively optimized local radiance fields for robust view synthesis. In CVPR, 2023. 3", + "[62] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 3, 5, 8", + "[63] Ben Mildenhall, Peter Hedman, Ricardo Martin-Brualla, Pratul P Srinivasan, and Jonathan T Barron. Nerf in the dark: High dynamic range view synthesis from noisy raw images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16190-16199, 2022. 1, 6, 8", + "[64] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 3, 4", + "[65] Zak Murez, Tarrence Van As, James Bartolozzi, Ayan Sinha, Vijay Badrinarayanan, and Andrew Rabinovich. Atlas: End-to-end 3d scene reconstruction from posed images. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VII 16, pages 414-431. Springer, 2020. 1", + "[66] Julian Ost, Issam Laradji, Alejandro Newell, Yuval Bahat, and Felix Heide. Neural point light fields. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3", + "[67] Jinsun Park, Kyungdon Joo, Zhe Hu, Chi-Kuei Liu, and In So Kweon. Non-local spatial propagation network for depth completion. In Proc. of European Conference on Computer Vision (ECCV), 2020. 2, 7, 8" + ], + "bbox": [ + 86, + 90, + 468, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[68] Vivek Pradeep, Christoph Rhemann, Shahram Izadi, Christopher Zach, Michael Bleyer, and Steven Bathiche. Monofusion: Real-time 3d reconstruction of small scenes with a single web camera. In 2013 IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 83-88. IEEE, 2013. 1", + "[69] Andrea Ramazzina, Mario Bijelic, Stefanie Walz, Alessandro Sanvito, Dominik Scheuble, and Felix Heide. Scatternerf: Seeing through fog with physically-based inverse neural rendering. arXiv preprint arXiv:2305.02103, 2023. 1", + "[70] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12179-12188, 2021. 6, 8", + "[71] Konstantinos Rematas, Andrew Liu, Pratul P. Srinivasan, Jonathan T. Barron, Andrea Tagliasacchi, Tom Funkhouser, and Vittorio Ferrari. Urban radiance fields. CVPR, 2022. 1, 3", + "[72] Gernot Riegler, Ali Osman Ulusoy, Horst Bischof, and Andreas Geiger. Octnetfusion: Learning depth fusion from data. In 2017 International Conference on 3D Vision (3DV), pages 57-66. IEEE, 2017. 1", + "[73] Barbara Roessle, Jonathan T Barron, Ben Mildenhall, Pratul P Srinivasan, and Matthias Nießner. Dense depth priors for neural radiance fields from sparse input views. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12892-12901, 2022. 3", + "[74] Viktor Rudnev, Mohamed Elgharib, William Smith, Lingjie Liu, Vladislav Golyanik, and Christian Theobalt. Nerf for outdoor scene relighting. In European Conference on Computer Vision, pages 615-631. Springer, 2022. 3", + "[75] Michael Schober, Amit Adam, Omer Yair, Shai Mazor, and Sebastian Nowozin. Dynamic time-of-flight. In CVPR, pages 6109-6118, 2017. 3", + "[76] Thomas Schöps, Torsten Sattler, Christian Hane, and Marc Pollefeys. 3d modeling on the go: Interactive 3d reconstruction of large-scale scenes on mobile devices. In 2015 International Conference on 3D Vision, pages 291-299. IEEE, 2015. 1", + "[77] Brent Schwarz. Lidar: Mapping the world in 3D. Nature Photonics, 4(7):429, 2010. 2", + "[78] Tixiao Shan, Brendan Englot, Drew Meyers, Wei Wang, Carlo Ratti, and Daniela Rus. Lio-sam: Tightly-coupled lidar inertial odometry via smoothing and mapping. In 2020 IEEE/RSJ international conference on intelligent robots and systems (IROS), pages 5135-5142. IEEE, 2020. 6", + "[79] Jiaming Sun, Yiming Xie, Linghao Chen, Xiaowei Zhou, and Hujun Bao. Neuralrecon: Real-time coherent 3d reconstruction from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15598-15607, 2021. 1", + "[80] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8248–8258, 2022. 3" + ], + "bbox": [ + 509, + 92, + 893, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "10540", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[81] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, et al. Nerfstudio: A modular framework for neural radiance field development. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-12, 2023. 3", + "[82] Jiexiong Tang, John Folkesson, and Patric Jensfelt. Sparse2dense: From direct sparse odometry to dense 3-d reconstruction. IEEE Robotics and Automation Letters, 4 (2):530-537, 2019. 2", + "[83] Jie Tang, Fei-Peng Tian, Wei Feng, Jian Li, and Ping Tan. Learning guided convolutional network for depth completion. IEEE Transactions on Image Processing, 30:1116-1129, 2020. 2", + "[84] Tang Tao, Longfei Gao, Guangrun Wang, Peng Chen, Dayang Hao, Xiaodan Liang, Mathieu Salzmann, and Kaicheng Yu. Lidar-nerf: Novel lidar view synthesis via neural radiance fields. arXiv preprint arXiv:2304.10406, 2023. 1, 3, 6, 7, 8", + "[85] Fabio Tosi, Alessio Tonioni, Daniele De Gregorio, and Matteo Poggi. Nerf-supervised deep stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 855–866, 2023. 3", + "[86] Haithem Turki, Jason Y Zhang, Francesco Ferroni, and Deva Ramanan. Suits: Scalable urban dynamic scenes. In Computer Vision and Pattern Recognition (CVPR), 2023. 1, 3, 6, 7, 8", + "[87] Jonas Uhrig, Nick Schneider, Lukas Schneider, Uwe Franke, Thomas Brox, and Andreas Geiger. Sparsity invariant cnns. In International Conference on 3D Vision (3DV), 2017. 2", + "[88] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481–5490. IEEE, 2022. 5", + "[89] Amanpreet Walia, Stefanie Walz, Mario Bijelic, Fahim Mannan, Frank Julca-Aguilar, Michael Langer, Werner Ritter, and Felix Heide. Gated2gated: Self-supervised depth estimation from gated images. 2022. 2, 3, 4, 6, 8", + "[90] Stefanie Walz, Mario Bijelic, Andrea Ramazzina, Amanpreet Walia, Fahim Mannan, and Felix Heide. Gated stereo: Joint depth estimation from gated and wide-baseline active stereo cues. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13252-13262, 2023. 2, 3, 4, 5, 6, 7, 8", + "[91] Haoyuan Wang, Xiaogang Xu, Ke Xu, and Rynson WH Lau. Lighting up nerf via unsupervised decomposition and enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12632-12641, 2023. 1", + "[92] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 1", + "[93] Zian Wang, Tianchang Shen, Jun Gao, Shengyu Huang, Jacob Munkberg, Jon Hasselgren, Zan Gojcic, Wenzheng" + ], + "bbox": [ + 86, + 92, + 470, + 902 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Chen, and Sanja Fidler. Neural fields meet explicit geometric representations for inverse rendering of urban scenes. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 3", + "[94] Alex Wong and Stefano Soatto. Unsupervised depth completion with calibrated backprojection layers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12747-12756, 2021. 2", + "[95] Gengshan Yang, Joshua Manela, Michael Happold, and Deva Ramanan. Hierarchical deep stereo matching on high-resolution images. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2", + "[96] Zhenfei Yang, Fei Gao, and Shaojie Shen. Real-time monocular dense mapping on aerial robots using visual-inertial fusion. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 4552-4559. IEEE, 2017. 1", + "[97] Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasun. Unisim: A neural closed-loop sensor simulator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1389-1399, 2023. 3", + "[98] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoptrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5752-5761, 2021. 3", + "[99] Junge Zhang, Feihu Zhang, Shaochen Kuang, and Li Zhang. Nerf-lidar: Generating realistic lidar point clouds with neural radiance fields. arXiv preprint arXiv:2304.14811, 2023. 3", + "[100] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 1, 3", + "[101] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Ner-factor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (ToG), 40(6):1-18, 2021. 3", + "[102] Yongjian Zhang, Longguang Wang, Kunhong Li, Zhiheng Fu, and Yulan Guo. Slfnet: A stereo and lidar fusion network for depth completion. IEEE Robotics and Automation Letters, 7(4):10605-10612, 2022. 2", + "[103] Tinghui Zhou, Matthew Brown, Noah Snavely, and David G. Lowe. Unsupervised learning of depth and egomotion from video. In CVPR, 2017. 2" + ], + "bbox": [ + 501, + 92, + 890, + 741 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "10541", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2024/Gated Fields_ Learning Scene Reconstruction from Gated Videos/babd13ae-5bce-44f5-bd0f-203ff88804c8_model.json b/2024/Gated Fields_ Learning Scene Reconstruction from Gated Videos/babd13ae-5bce-44f5-bd0f-203ff88804c8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..930fa08e7ef2807d56e81b86931b6fc85f006c9b --- /dev/null +++ b/2024/Gated Fields_ Learning Scene Reconstruction from Gated Videos/babd13ae-5bce-44f5-bd0f-203ff88804c8_model.json @@ -0,0 +1,2820 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.157, + 0.093, + 0.817, + 0.114 + ], + "angle": 0, + "content": "Gated Fields: Learning Scene Reconstruction from Gated Videos" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.143, + 0.867, + 0.163 + ], + "angle": 0, + "content": "Andrea Ramazzina\\(^{1*}\\) Stefanie Walz\\(^{1,2*}\\) Pragyan Dahal\\(^{3}\\) Mario Bijelic\\(^{4,5}\\) Felix Heide\\(^{4,5}\\)" + }, + { + "type": "text", + "bbox": [ + 0.169, + 0.165, + 0.821, + 0.183 + ], + "angle": 0, + "content": "\\(^{1}\\)Mercedes-Benz \\(^{2}\\)Saarland University \\(^{3}\\)Politecnico di Milano \\(^{4}\\)Torch Robotics \\(^{5}\\)Princeton University" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.217, + 0.314, + 0.233 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.25, + 0.474, + 0.552 + ], + "angle": 0, + "content": "Reconstructing outdoor 3D scenes from temporal observations is a challenge that recent work on neural fields has offered a new avenue for. However, existing methods that recover scene properties, such as geometry, appearance, or radiance, solely from RGB captures often fail when handling poorly-lit or texture-deficient regions. Similarly, recovering scenes with scanning LiDAR sensors is also difficult due to their low angular sampling rate which makes recovering expansive real-world scenes difficult. Tackling these gaps, we introduce Gated Fields - a neural scene reconstruction method that utilizes active gated video sequences. To this end, we propose a neural rendering approach that seamlessly incorporates time-gated capture and illumination. Our method exploits the intrinsic depth cues in the gated videos, achieving precise and dense geometry reconstruction irrespective of ambient illumination conditions. We validate the method across day and night scenarios and find that Gated Fields compares favorably to RGB and LiDAR reconstruction methods. Our code and datasets are available here1." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.581, + 0.21, + 0.597 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.607, + 0.471, + 0.865 + ], + "angle": 0, + "content": "Large-scale outdoor scene reconstruction is essential for advancing autonomous robotics, drones, and driver-assistance systems, serving as the foundation for scene understanding, safe navigation, dataset generation and validation. Existing works in this domain [68, 76, 96] have typically adopted a two-step approach. Initially, they infer depth maps from different poses, utilizing time-of-flight sensors or RGB captures. Subsequently, these depth estimates are fused to produce a coherent 3D representation, using either classical methodologies [25] or learned-based representations [72]. In contrast, more recent studies [16, 65, 79] have proposed end-to-end strategies that bypass the estimation of local depth maps as an intermediate representation. Instead, they directly regress a Truncated Signed Distance Function (TSDF) [65, 79] or an occupancy volume [16]. A rapidly growing body of work on neural rendering [62, 92] offers not only geometrically-accurate scene recon" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.215, + 0.895, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.407, + 0.895, + 0.464 + ], + "angle": 0, + "content": "Figure 1. From a single video of gated captures (top-row), we reconstruct an accurate scene representation and render depth projections (mid-row, right) as accurate as LiDAR scans (mid-row, left), and we recover 3D geometry and normals (bottom-row)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.475, + 0.895, + 0.627 + ], + "angle": 0, + "content": "struction from posed RGB images but also to generate novel perspectives from unobserved angles. Hinging on implicit coordinate-based neural representations, RGB-based methods [7, 8, 55, 100] have been adapted to large open outdoor environments. A recent line of work [37, 71, 86, 93] includes LiDAR scans for auxiliary depth supervision and to improve scene reconstruction for urban environments. However, recovery based on the RGB images exhibit is fundamentally limited in the presence of low light [63, 91] or in the presence of scattering such as fog [49, 69]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.896, + 0.869 + ], + "angle": 0, + "content": "A parallel direction of research investigates neural rendering techniques tailored to Time-of-Flight (ToF) sensors as opposed to the conventional RGB cameras. Existing methods [41, 84] tackle scene understanding with posed LiDAR scans, and have modeled the raw output from a single-photon LiDAR system [59] or continuous-wave ToF sensor [4] as additional depth supervision. However, all of these existing methods struggle with recovering large unbounded outdoor scenes: while continuous-wave ToF sensors [4] offer signal only in room-sized scenes, methods based on scanning LiDAR suffer from low angular sampling which mandates temporal aggregation. Specifically, even today's LiDAR sensors boasting 200 scan lines, lag drastically in resolution compared to current HDR cameras that offer two orders of magnitude higher vertical pixel counts nearing \\(10\\mathrm{k}\\) and three orders of magnitude higher total resolution." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Addressing these challenges, our work explores scene reconstruction using active gated imaging. Gated imag-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.094, + 0.875, + 0.352, + 0.888 + ], + "angle": 0, + "content": "*These authors contributed equally to this work." + }, + { + "type": "page_footnote", + "bbox": [ + 0.094, + 0.888, + 0.38, + 0.9 + ], + "angle": 0, + "content": "1https://light.princeton.edu/gatedfields/" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.871, + 0.38, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10530" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.544 + ], + "angle": 0, + "content": "ing functions by integrating the transient response from a scene that has been flash-illuminated by a synchronized light source [19]. This imaging technique is robust to adverse weather conditions – temporal gating allows us to filter out backscatter – and provides signal in poorly-lit scenes [13]. Existing work has exploited this sensing modality to achieve state-of-the-art depth estimation [89, 90] as well as object detection [44]. In our approach, we train a neural field-based representation of the scene, concurrently learning its geometry, illumination, and material properties. This is accomplished by integrating the gated imaging formation model with a neural rendering framework, that jointly learns the associated gating parameters along with the scene reconstruction. By leveraging the implicit depth cues present in gated video captures, we are able to reconstruct a detailed 3D geometric model of the scene, as shown in Fig. 1. Compared to LiDAR-based approaches, our method offers distinct advantages, as LiDAR systems are inherently constrained by their resolution, necessitating additional time-multiplexed scene captures to aggregate points. This results in an extended acquisition process for LiDAR-based methods or, conversely, compromises the geometric detail of the final estimate. Specifically, for a fixed time acquisition budget the scene reconstruction from a LiDAR sensor is less supervised, although offering highly accurate depth information, the sensor yields data at a volume one order of magnitude less than that of a camera stereo pair. This disparity in data quantity means that while the LiDAR provides precise depth points, it is unable to provide dense and fine detailed predictions." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.549, + 0.473, + 0.671 + ], + "angle": 0, + "content": "To assess our method, we captured a dataset of varied scenes at day and night conditions, using a vehicle test setup comprising of LiDAR, RGB and Gated sensors. We compare our method with feed-forward and 3D reconstruction methods, and assess its superiority in novel depth synthesis beating the next best method by \\(21.87\\%\\) MAE, 3D reconstruction improving on the baseline by \\(11\\%\\) IoU, and performs novel view synthesis with a PSNR of \\(32.28~\\mathrm{dB}\\)." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.68, + 0.414, + 0.695 + ], + "angle": 0, + "content": "In our work, we make the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.705, + 0.468, + 0.749 + ], + "angle": 0, + "content": "- We propose a novel neural rendering method and scene representation that is capable of reconstructing scene geometry and radiance from active gated camera videos." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.751, + 0.468, + 0.825 + ], + "angle": 0, + "content": "- By modeling the gated image formation process and integrating into differentiable volume rendering, our approach is able to reconstruct and decompose both passive and active light transport components conditioned on the scene parameters in a physically accurate way." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.826, + 0.468, + 0.901 + ], + "angle": 0, + "content": "- We validate our method on large outdoor scenes, captured across different scenarios in both day and night, achieving a reduction of MAE error in depth precision of \\(59.8\\%\\) to the next best RGB+LiDAR method and \\(31.7\\%\\) to the next best methods using gated images." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.705, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.09, + 0.642, + 0.107 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.123, + 0.895, + 0.516 + ], + "angle": 0, + "content": "Monocular and Stereo Depth Estimation Depth estimation tasks, from a single image [32, 36, 51, 54], from stereo image pairs [5, 22, 52, 95], or single/stereo images with a LiDAR scan [24, 40, 67, 82, 83, 94, 102] have been at the center of a large body of work. Single CMOS sensor-based depth estimation from RGB color images is inherently limited by scale ambiguity. Additional measurements from LiDAR [9] or ego-vehicle speed [36] can resolve this ambiguity at the cost of an additional sensor. Similarly, stereo methods rely on an additional camera sensor to resolve ambiguity by triangulating between two camera views [22]. Training approaches for learned depth estimation methods using intensity images cover both unsupervised methods [30, 32, 33, 36, 103], which harnesses multi-view geometry consistency, and supervised techniques [22, 27, 42, 45, 51, 54, 58, 60] relying primarily on multi-view datasets [45, 51, 60] or time-of-flight captures [22, 27, 42, 58]. In particular, LiDAR measurements have been proven as a ground-truth signal for depth supervision [2, 22, 27, 42, 54, 58]. Several methods [31, 87] have mitigated the sparsity and range limitations of scanning LiDAR by accumulating scans. However, adverse weather [14] can make LiDAR ground truth unreliable, and methods that rely on consistency between camera and LiDAR [24, 40, 67, 82, 83, 94, 102] suffer from degradations, including scan pattern artifacts and temporal distortions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Time-gated Depth Sensing Time-of-Flight (ToF) sensors determine depth by emitting light into a scene and calculating the distance based on the round trip time of the light. Successful sensing methods can be categorized into three classes: correlation ToF cameras [38, 46, 48], pulsed ToF sensors [77], and gated imaging [34, 39]. Correlation ToF sensors[38, 46, 48] estimate depth by continuously flood-illuminating the scene and assessing the phase shift between the emitted and received light. While these sensors can provide high-resolution depth information, their use is primarily confined to indoor settings due to their susceptibility to external light interference. Pulsed ToF sensors [77] function by emitting light pulses toward specific scene points and measure the total travel time to estimate depth. Emitting collimated light, this approach is robust against ambient illumination and allows for outdoor depth measurements. However, its spatial resolution is fundamentally limited owing to its scanning illumination technique, and its efficacy can be compromised in adverse weather due to backscatter [11, 21, 43]. In contrast, gated cameras [12, 34, 39] capture light from a scene over brief intervals, essentially constraining the observable depth to specific range segments. The inherent gating mechanism of these cameras offers resistance to backscattering, and they allow for the recovery of detailed depth maps when using a large number of short" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10531" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.198 + ], + "angle": 0, + "content": "gates [3, 17, 18] Subsequent works have improved gated depth estimation with few gates by adopting Bayesian approaches [1, 75] or deep neural networks [35, 89, 90], and they achieve accurate gated depth estimation for dynamic outdoor scenes, even under challenging conditions. Recently, Gated Stereo [90] reached state-of-the-art results using a stereo-gated setup and self-supervision [89]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.2, + 0.471, + 0.697 + ], + "angle": 0, + "content": "Neural Scene Reconstruction Recent research has amalgamated sets of single-sensor measurements to recover comprehensive scene representations. This synthesis has led to advancements in both novel-view generation [6, 23, 62, 64] and depth estimation [85], with neural radiance field methods [6, 23, 62, 64] emerging as a pivotal approach for representing scenes as continuous volumetric fields of radiance. These methods combine this representation with volumetric rendering as a forward model in a test-time optimization approach. Specific representations that these methods explore include coordinate-based networks [6, 7, 62, 101], 3D voxel-grid representation [23, 28, 98], or hybrid approaches [8, 64, 81]. Subsequent works have extended this representation to large outdoor scenes [7, 100], and increased the efficiency at training and test time [8, 23, 64, 98]. Other works departed from radiance-based representation and explicitly learned the scene illumination, geometry, and material properties [15, 74, 101]. A particular challenge within the field is reconstructing large urban terrains based on imagery captured from vehicles [37, 47, 56, 66, 71, 80, 86, 93, 97], given that a significant portion of the scene is seen from only a narrow range of viewpoints. This problem is tackled by additional supervision cues from sparse LiDAR [37, 66, 71, 86], pre-estimated depths [26, 37, 73], optical flow [61, 86] and semantic segmentation [47, 86, 93]. Departing from RGB-based approaches, recent works have investigated neural reconstruction methods using ToF sensors [41, 59, 84]. Existing methods [41, 84, 99] learn a neural field from posed LiDAR scans, allowing for the synthesis of realistic LiDAR scans from novel views. Recently, Malik et al. [59] represent the time-resolved photon count acquired by a single-photon LiDAR system with a neural reconstruction method." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.71, + 0.287, + 0.726 + ], + "angle": 0, + "content": "3. Gated Imaging Model" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.469, + 0.81 + ], + "angle": 0, + "content": "In this section, we provide a brief background on gated imaging as presented in [35] and introduce the proposed gated image formation model. Unlike prior work, this model incorporates shadow effects and features self-calibrated parameter learning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.469, + 0.901 + ], + "angle": 0, + "content": "A gated imaging system, illustrated in Fig. 2, utilizes a pulsed flood-light illumination source \\( p \\) with a synchronized imager that operates with a nanosecond (ns) gated exposure \\( g \\) that is delayed by \\( \\xi \\) compared to the pulse. This allows us to capture only photons with round-trip times inside the gates, hence specific distance segments in the scene." + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.084, + 0.882, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.254, + 0.892, + 0.435 + ], + "angle": 0, + "content": "Figure 2. Gated Image Formation and Bi-Directional Sampling. Top-row: Our test vehicle is equipped with a synchronized stereo camera setup and illuminator that flood-lits the scene with a light pulse and FoV \\(\\gamma\\). Using different gating profiles \\(C(z)\\), we capture three slices with intensity visualised here in red, green and blue. Illustrated in the middle row, the gating profiles describe pixel intensity for a point at sensor distance \\(z\\). The first slice (red) accounts for close ranges, the green for mid-ranges, and the blue for far ranges. Bottom-row: we show the ray sampling employed in our method, based on a bidirectional sampling strategy. We cast the rays from the illuminator view to explore the occluded areas, while the rays casted from the camera integrate the reflected scene response. The shadowed areas are marked in gray." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.451, + 0.892, + 0.496 + ], + "angle": 0, + "content": "We formalize this using so-called range intensity profiles \\( C_k(2z_c) \\) given distance \\( z_{c} \\) from the camera, time \\( t \\), and a parameter set \\( k \\) [35], that is" + }, + { + "type": "equation", + "bbox": [ + 0.517, + 0.508, + 0.892, + 0.554 + ], + "angle": 0, + "content": "\\[\nC _ {k} \\left(2 z _ {c}\\right) = \\int_ {- \\infty} ^ {\\infty} g _ {k} (t - \\xi) p _ {k} \\left(t - \\frac {2 z _ {c}}{c}\\right) \\beta \\left(2 z _ {c}\\right) d t, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.567, + 0.891, + 0.613 + ], + "angle": 0, + "content": "where \\(c\\) is the speed of light and \\(\\beta (\\cdot)\\) models the distance-dependent decay of the reflected light pulse. The resulting gated pixel value is" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.625, + 0.891, + 0.641 + ], + "angle": 0, + "content": "\\[\nI _ {k} \\left(z _ {c}\\right) = \\alpha \\iota C _ {k} \\left(2 z _ {c}\\right) + \\Lambda + \\mathcal {D} _ {k}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.653, + 0.891, + 0.697 + ], + "angle": 0, + "content": "where \\(\\Lambda\\) represents the passive ambient contribution, \\(\\alpha\\) is the scene reflection, \\(\\iota\\) the laser illumination, and \\(\\mathcal{D}_k\\) is an additive noise term." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.699, + 0.892, + 0.835 + ], + "angle": 0, + "content": "This model assumes camera position \\(\\mathbf{o}_c\\) and the illuminator position \\(\\mathbf{o}_i\\) are collocated. To allow for noncollocated positions, we express the travel time as \\(z = z_c + z_i\\), where \\(z_i\\) denotes the distance between the illuminator and the point on the surface impacted by the light beam, represented as \\(z_i = |\\mathbf{x} - \\mathbf{o}_i|_2\\). Additionally, there may be areas visible to the camera that remain dark due to potential occlusions. Modeling shadow effects and attenuation due to incident angle \\(\\omega\\) results in the following image formation" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.845, + 0.891, + 0.86 + ], + "angle": 0, + "content": "\\[\nI _ {k} (z) = \\alpha \\iota \\psi | \\mathbf {n} \\cdot \\omega | C _ {k} (z) + \\Lambda + \\mathcal {D} _ {k}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.891, + 0.901 + ], + "angle": 0, + "content": "Here, \\(\\psi \\in [0,1]\\) serves as a shadow indicator for the pixel, and \\(\\omega\\) is the direction of the incident light at that point." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10532" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.213 + ], + "angle": 0, + "content": "We extend this model to fit the range intensity profiles during optimization, thereby eliminating the need for their direct measurement. This approach overcomes potential calibration inaccuracies encountered in previous approaches [89, 90]. We model both the laser pulse \\( p_k \\) and the gate \\( g_k \\) as rectangular functions with durations \\( t_{l,k} \\) and \\( t_{g,k} \\), respectively. This simplification permits the analytical computation of the integral in Eq. (1), that is" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.223, + 0.478, + 0.309 + ], + "angle": 0, + "content": "\\[\n\\tilde {C} _ {k} = \\left\\{ \\begin{array}{l l} \\frac {2 z}{c} - \\xi_ {k} + t _ {l, k} & \\text {i f} \\xi_ {k} - t _ {l, k} < \\frac {2 z}{c} < \\xi_ {k} \\\\ t _ {l, k} & \\text {i f} \\xi_ {k} < \\frac {2 z}{c} < \\xi_ {k} + t _ {g, k} - t _ {l, k} \\\\ - \\frac {2 z}{c} + \\xi_ {k} + t _ {g, k} & \\text {i f} \\xi_ {k} + l _ {g, k} - t _ {l, k} < \\frac {2 z}{c} < \\xi_ {k} + t _ {g, k} \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {4}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.312, + 0.202, + 0.328 + ], + "angle": 0, + "content": "4. Gated Field" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.337, + 0.47, + 0.459 + ], + "angle": 0, + "content": "We reconstruct a scene by fitting a neural field representation to gated videos. We collect videos of three active gated slices \\( I_{k \\in \\{1,2,3\\}} \\) with different gating parameters and one passive slice \\( I_P \\). We model active illumination by jointly estimating light and material properties, and separately represent the ambient light as a radiance field. The proposed reconstruction method relies on both photometric reconstruction cues and scene priors and is described in the following." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.467, + 0.269, + 0.483 + ], + "angle": 0, + "content": "4.1. Neural Gated Fields" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.49, + 0.47, + 0.598 + ], + "angle": 0, + "content": "We describe the scene properties using two neural fields \\( f_{Gp} \\) and \\( f_{G\\alpha} \\), respectively representing the ambient light scattered in the scene and the reflection of the scene surfaces, conditioned on a spatial embedding \\( \\chi \\). Moreover, the laser illumination contribution is represented by a physics-based model, while shadow effects are simulated through ray-tracing using the volumetric density field \\( f_{Gd} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.615, + 0.47, + 0.691 + ], + "angle": 0, + "content": "Neural Ambient and Reflection Field We represent a scene as a neural field \\( f_{G}:\\{\\mathbf{x},\\mathbf{d},\\omega \\} \\to \\{\\sigma ,\\alpha ,\\Lambda ,\\mathbf{n}\\} \\) mapping each point in space \\( \\mathbf{x} \\) viewed from a direction \\( \\mathbf{d} \\) and laser direction \\( \\omega \\) to its volumetric density \\( \\sigma \\), normal vector \\( \\mathbf{n} \\), scene reflection \\( \\alpha \\) and the passive component \\( \\Lambda \\), that is" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.713, + 0.273, + 0.73 + ], + "angle": 0, + "content": "\\[\nf _ {G p}: \\{\\mathbf {d}, \\chi \\} \\longrightarrow \\{\\Lambda \\}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.336, + 0.713, + 0.478, + 0.729 + ], + "angle": 0, + "content": "Ambient Component" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.732, + 0.271, + 0.748 + ], + "angle": 0, + "content": "\\[\nf _ {G n}: \\{\\mathbf {x}, \\chi \\} \\longrightarrow \\{\\mathbf {n} \\}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.369, + 0.733, + 0.476, + 0.746 + ], + "angle": 0, + "content": "Surface Normal" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.751, + 0.272, + 0.767 + ], + "angle": 0, + "content": "\\[\nf _ {G \\alpha}: \\{\\mathbf {d}, \\omega , \\chi \\} \\rightarrow \\{\\alpha \\}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.352, + 0.752, + 0.476, + 0.765 + ], + "angle": 0, + "content": "Surface Reflection" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.769, + 0.477, + 0.786 + ], + "angle": 0, + "content": "with \\(f_{Gd}:\\{\\mathbf{x}\\} \\longrightarrow \\{\\sigma ,\\chi \\}\\) Vol. Density and Embedding" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.795, + 0.473, + 0.902 + ], + "angle": 0, + "content": "We condition here normal, ambient and reflectance on a volumetric embedding via the field \\(f_{Gd}:\\{\\mathbf{x}\\} \\to \\{\\sigma ,\\chi \\}\\) estimating the density \\(\\sigma\\) and embedding \\(\\chi\\). This embedding is being shared by the network branches to estimate the normal with \\(f_{Gn}:\\{\\mathbf{x},\\chi \\} \\to \\{\\mathbf{n}\\}\\), scene reflection \\(f_{G\\alpha}:\\{\\mathbf{d},\\omega ,\\chi \\} \\to \\{\\alpha \\}\\), and ambient light component with \\(f_{Gp}:\\{\\mathbf{d},\\chi \\} \\to \\{\\Lambda \\}\\). An overview of the overall Neural" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.138 + ], + "angle": 0, + "content": "Gated Fields is shown in Fig. 3. We also use a proposal sampler \\( f_{P} \\) as in [7] for efficiency. Both \\( f_{G} \\) and \\( f_{P} \\) are MLPs (of different size) with multi-resolution hash encoding [64]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.156, + 0.892, + 0.262 + ], + "angle": 0, + "content": "Shadow and Illumination Model As the light pulse emitted by the illuminator is a diverging light beam, we model it as cone of light with irradiance maximum at the cross-section center and exponentially decreasing as it diverges from the center by the angles \\(\\gamma\\). As such, we express the illumination intensity as a 2D higher-order Gaussian \\(\\mathcal{G}\\) with mean \\(\\Xi\\), standard deviation \\(\\Omega\\) and power \\(\\Theta\\)" + }, + { + "type": "equation", + "bbox": [ + 0.623, + 0.274, + 0.892, + 0.291 + ], + "angle": 0, + "content": "\\[\n\\iota = \\eta \\mathcal {G} _ {i} (\\gamma ; \\boldsymbol {\\Xi}, \\boldsymbol {\\Omega}, \\boldsymbol {\\Theta}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.303, + 0.707, + 0.318 + ], + "angle": 0, + "content": "where \\(\\eta\\) is a scaling parameter." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.318, + 0.892, + 0.379 + ], + "angle": 0, + "content": "Instead of predicting the shadow indicator \\(\\psi (\\mathbf{x})\\), we can directly estimate it using the density field, by computing the accumulated transmittance along the ray from the pixel to the point \\(\\mathbf{r}_{ill}(l) = \\mathbf{o}_i + \\omega l\\)" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.39, + 0.892, + 0.43 + ], + "angle": 0, + "content": "\\[\n\\psi (\\mathbf {x}) = \\exp \\left(- \\int_ {0} ^ {l _ {\\mathbf {x}}} \\sigma \\left(\\mathbf {r} _ {i l l} (l)\\right) \\mathrm {d} l\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.44, + 0.892, + 0.608 + ], + "angle": 0, + "content": "The illuminator origin and direction \\(\\mathbf{o}_i, \\omega\\) are obtained from the camera as \\([\\mathbf{o}_i, \\omega] = \\mathbf{R}[\\mathbf{o}_c, \\mathbf{d}_c] + \\mathbf{T}\\). During training, we jointly fine-tune the translation and rotation matrices \\(\\mathbf{T}, \\mathbf{R}\\), as well as \\(\\mathbf{o}_c\\). We also treat illuminator profile properties as learnable parameters, namely \\(\\eta, \\Xi, \\Omega, \\Theta\\) and the gating parameters, i.e. number of accumulated laser pulses \\(m_k\\) before read-out, laser pulse duration \\(t_{l,k}\\), camera exposure \\(t_{g,k}\\), and delay \\(\\xi_k\\) between laser pulse emission and gated exposure for all three slices \\(k \\in \\{0,1,2\\}\\). In addition, we optimize a general distance offset \\(d_0\\) for the range intensity profiles to compensate for internal signal processing delays." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.617, + 0.704, + 0.634 + ], + "angle": 0, + "content": "4.2. Gated Field Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.64, + 0.892, + 0.716 + ], + "angle": 0, + "content": "We learn to acquire a gated capture with camera origin \\(\\mathbf{o}_c\\) and direction \\(\\mathbf{d}\\) by casting a ray \\(\\mathbf{r}(l) = \\mathbf{o}_c + l\\mathbf{d}\\) for each pixel into the scene and computing the intensity \\(\\tilde{I}_k(\\mathbf{r})\\) through volume rendering. Using the gated imaging formation model from Sec. 3, we define volume rendering as" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.727, + 0.848, + 0.796 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\tilde{I}_{k}(\\mathbf{r}) = \\int_{0}^{\\infty}T(l)\\sigma (\\mathbf{x})\\int_{-\\infty}^{\\infty}g_{k}(t - \\xi_{k})\\beta (l) \\\\ \\cdot \\left(\\alpha \\iota \\psi | \\mathbf {n} \\cdot \\omega | p _ {k} \\left(t - \\frac {l + l _ {i}}{c}\\right) + \\kappa\\right) d t d l + \\mathcal {D} _ {k}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.809, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\kappa\\) is the ambient level, \\(T(l) = \\exp (-\\int_0^l\\sigma (u)du)\\) is the accumulated transmittance along the ray and \\(l_{i}\\) is the distance of \\(\\mathbf{x}(l)\\) from the illuminator origin \\(\\mathbf{o}_i\\). As illustrated in Fig. 2, in our volume rendering formulation the pixel intensity contribution of a point along the ray depends not only on its accumulated transmittance and volumetric" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10533" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.139, + 0.077, + 0.833, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.236, + 0.895, + 0.321 + ], + "angle": 0, + "content": "Figure 3. Neural Gated Fields. For any point in space \\(\\mathbf{x}\\), we learn its volumetric density \\(\\sigma\\), normal \\(\\mathbf{n}\\), reflectance \\(\\alpha\\) and ambient lighting \\(\\Lambda\\) through four neural fields, conditioned on direction \\(\\mathbf{d}\\), incident laser light direction \\(\\omega\\) and spatial embedding \\(\\chi\\). The illuminator light \\(\\iota\\) is represented by a physics-based model dependent on the displacement angle \\(\\gamma\\), while the gating imaging process is described by the range intensity profiles \\(\\tilde{C}(z)\\) using as input the camera-point-laser distance \\(z\\), as explained in Sec. 3. With this information, we reconstruct a gated image \\(I_k\\) through the gated volume rendering formulation introduced in Sec. 4.2. As this process is fully differentiable, we simultaneously fit neural fields and physical parameters through image reconstruction together with other regularization losses discussed in Sec. 4.3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.335, + 0.47, + 0.379 + ], + "angle": 0, + "content": "density, but also on its distance from the illuminator and camera origins through \\( C(z) \\), as well as on its relative position to the illuminator source via \\( \\iota \\) and \\( \\psi \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.381, + 0.453, + 0.395 + ], + "angle": 0, + "content": "We simplify the time-dependent integral using Eq. (4) as" + }, + { + "type": "equation", + "bbox": [ + 0.096, + 0.395, + 0.469, + 0.446 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\tilde {I} _ {k} (\\mathbf {r}) = \\int_ {0} ^ {\\infty} T (l) \\sigma (\\mathbf {x}) \\left(\\alpha (\\mathbf {x}, \\mathbf {d}, \\omega) \\tilde {C} _ {k} \\left(l + l _ {i}\\right) \\psi (\\mathbf {x}) \\right. \\tag {8} \\\\ \\cdot | \\mathbf {n} \\cdot \\omega | \\iota (\\gamma) + \\Lambda (\\mathbf {x}, \\mathbf {d})) \\mathrm {d} l + \\mathcal {D} _ {k}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.452, + 0.47, + 0.572 + ], + "angle": 0, + "content": "We numerically estimate this spatial integral by numerical quadrature [62, 88], approximating it with a set of points \\(\\mathbf{X}_{ray}\\). Specifically, for each point \\(\\mathbf{x}_j \\in \\mathbf{X}_{ray}\\), we query the neural field \\(f_{G}\\) to infer the normal vector \\(\\mathbf{n}_j\\), scene reflection \\(\\alpha_j\\), volumetric density \\(\\sigma_j\\) and ambient light component \\(\\Lambda_j\\). The laser illumination intensity \\(\\iota_j\\) is instead computed following the physics-based model defined in Eq. (5). The gated intensity \\(\\tilde{I}_k\\) is then expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.571, + 0.469, + 0.672 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\tilde {I} _ {k} (\\mathbf {r}) = \\sum_ {j = 0} ^ {N} w _ {j} \\left(\\underbrace {\\alpha_ {j} \\tilde {C} _ {j} \\psi_ {j} | \\mathbf {n} _ {j} \\cdot \\omega_ {j} | \\iota_ {j}} _ {\\text {A c t i v e C o m p o n e n t}} + \\underbrace {\\Lambda_ {j}} _ {\\text {P a s s i v e C o m p o n e n t}}\\right) + \\mathcal {D} _ {k}, (9) \\\\ w _ {j} = \\exp \\left(- \\sum_ {k = 1} ^ {j - 1} \\sigma_ {k} \\delta_ {k}\\right) \\left(1 - \\exp \\left(- \\sigma_ {j} \\delta_ {j}\\right)\\right). (9) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.676, + 0.469, + 0.72 + ], + "angle": 0, + "content": "The shadow indicator \\(\\psi_{j}\\) from Eq. (6) is similarly approximated by sampling on \\(\\mathbf{r}_{ill} = \\mathbf{o}_i + \\omega_j l\\) a set of points \\(\\mathbf{X}_{ill}\\) bounded between \\(\\mathbf{o}_i\\) and \\(\\mathbf{x}_j\\)" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.719, + 0.469, + 0.759 + ], + "angle": 0, + "content": "\\[\n\\psi_ {j} = \\exp \\left(- \\sum_ {k} ^ {\\mathbf {X} _ {i l l}} \\sigma_ {k} \\delta_ {k}\\right). \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.766, + 0.469, + 0.795 + ], + "angle": 0, + "content": "For the passive slice, the active component is null, further simplifying to" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.793, + 0.469, + 0.834 + ], + "angle": 0, + "content": "\\[\n\\tilde {I} _ {P} (\\mathbf {r}) = \\sum_ {j = 0} ^ {N} w _ {j} \\Lambda_ {j} + \\mathcal {D} _ {P}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Both \\(\\mathbf{X}_{ray}\\) and \\(\\mathbf{X}_{ill}\\) are sampled using a proposal network [7] \\(f_{P}\\) that, analogously to \\(f_{Gd}\\), predicts point-wise densities converted with Eq. (10) to proposal weights \\(\\hat{w}\\) for sampling with piece-wise-constant probabilities." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.335, + 0.698, + 0.35 + ], + "angle": 0, + "content": "4.3. Training Supervision" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.351, + 0.892, + 0.426 + ], + "angle": 0, + "content": "We supervise the predicted passive and active gated frames applying a photometric loss \\(\\mathcal{L}_c\\), regularize the volumetric density with a depth loss \\(\\mathcal{L}_d\\) and by supervising the shadow estimate with \\(\\mathcal{L}_s\\). We regularize normal and reflectance estimates through \\(\\mathcal{L}_{nc}\\) and \\(\\mathcal{L}_{\\alpha}\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.428, + 0.892, + 0.457 + ], + "angle": 0, + "content": "Photometric Loss We supervise with ground truth captures for active and passive gated slice reconstruction as" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.458, + 0.892, + 0.489 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c} = \\sum_ {k, r} \\| \\tilde {I} _ {k} (\\mathbf {r}) - I _ {k} (\\mathbf {r}) \\| _ {2} + \\sum_ {r} \\| \\tilde {I} _ {P} (\\mathbf {r}) - I _ {P} (\\mathbf {r}) \\| _ {2} \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.498, + 0.892, + 0.559 + ], + "angle": 0, + "content": "Volume Density Regularization As additional training supervision, we use the depth estimate \\(\\hat{D}(\\mathbf{r})\\) of a pretrained stereo depth estimation algorithm [90] as pseudo ground-truth to regularize the ray termination distribution [26]" + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.559, + 0.892, + 0.6 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d} = \\sum_ {\\mathbf {r}} \\sum_ {j} \\log w _ {j} \\exp \\left(- \\frac {\\left(l _ {j} - \\hat {D} (\\mathbf {r})\\right) ^ {2}}{2 s ^ {2}}\\right) \\delta_ {i} \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.608, + 0.892, + 0.699 + ], + "angle": 0, + "content": "We regularize the density field by partially supervising the shadow indicator \\(\\psi\\). Each pixel whose active intensity \\(I_{kA} = I_k - I_P\\) in any of the three gated slices is above a certain threshold \\(\\epsilon_i\\) is considered as visible from the illuminator. We hence supervise the expected shadow value for such rays \\(\\mathbf{r}_v \\in \\{\\mathbf{r} | \\forall k \\in \\{1, 2, 3\\} : I_{kA}(\\mathbf{r}) > \\epsilon_i\\}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.7, + 0.891, + 0.733 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s} = \\sum_ {\\mathbf {r} _ {v}} \\| 1 - \\int T (l) \\sigma (\\mathbf {x}) \\psi (\\mathbf {x}) d l \\| _ {2} \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.736, + 0.892, + 0.797 + ], + "angle": 0, + "content": "Normals Consistency Following [88], for each sampled point \\(\\mathbf{x}\\) we enforce a consistency between the predicted normal \\(\\mathbf{n}\\) and the density gradient \\(\\hat{\\mathbf{n}} (\\mathbf{x}) = -\\frac{\\nabla\\mathbf{x}}{||\\nabla\\mathbf{x}||}\\), and we penalize normals which are back-facing the camera as" + }, + { + "type": "equation", + "bbox": [ + 0.501, + 0.798, + 0.891, + 0.828 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {n c} = \\sum_ {\\mathbf {x}} w (\\mathbf {x}) \\left(\\| \\mathbf {n} (\\mathbf {x}) - \\hat {\\mathbf {n}} (\\mathbf {x}) \\| _ {2} + \\max (0, \\hat {\\mathbf {n}} (\\mathbf {x}) \\cdot \\mathbf {d}) ^ {2}\\right) \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.83, + 0.892, + 0.86 + ], + "angle": 0, + "content": "Reflectance Regularization We enforce the predicted scene reflection \\(\\alpha\\) to be spatially consistent within \\(\\epsilon_{\\mathbf{x}}\\), i.e." + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.873, + 0.892, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\alpha} = \\sum_ {\\mathbf {x}} w (\\mathbf {x}) \\left(\\| \\alpha (\\mathbf {x}, \\mathbf {d}) - \\alpha \\left(\\mathbf {x} + \\epsilon_ {\\mathbf {x}}, \\mathbf {d} + \\epsilon_ {\\mathbf {d}}\\right) \\| _ {2}\\right). \\tag {17}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10534" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.198 + ], + "angle": 0, + "content": "Here the \\(\\omega\\) dependency of \\(\\alpha\\) is omitted here for brevity. We also include an angular noise \\(\\epsilon_{\\mathrm{d}}\\) that we set high at the beginning of the training and then decrease it exponentially. By forcing the reflectance to behave as fully diffuse at the beginning of the training, we disincentive it to bake in the effects from lighting or shadowing, hence improving the disjoint learning of the scene components." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.2, + 0.47, + 0.23 + ], + "angle": 0, + "content": "Total Training Loss Combining the different losses we obtain the following loss formulation:" + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.242, + 0.469, + 0.259 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {1} \\mathcal {L} _ {c} + \\lambda_ {2} \\mathcal {L} _ {d} + \\lambda_ {3} \\mathcal {L} _ {s} + \\lambda_ {4} \\mathcal {L} _ {n c} + \\lambda_ {5} \\mathcal {L} _ {\\alpha}, \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.269, + 0.469, + 0.285 + ], + "angle": 0, + "content": "see \\(\\lambda_{1,\\dots,5}\\) hyperparameters in the Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.29, + 0.3, + 0.307 + ], + "angle": 0, + "content": "5. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.309, + 0.47, + 0.445 + ], + "angle": 0, + "content": "We train for 35,000 steps and a batch size of 4096 rays. As optimizer we use ADAMW [57] with \\(\\beta_{1} = 0.9\\), \\(\\beta_{2} = 0.999\\), learning rate \\(10^{-2}\\) for \\(f_{P}\\) and \\(f_{G}\\), \\(10^{-4}\\) for the camera poses optimization, \\(10^{-4}\\) for the laser profile and gated parameters. We train on two NVIDIA V100 GPUs, for approximately 3 hours. The proposal network \\(f_{P}\\) is comprised of two MLPs and trained following [7]. Additional architecture details, training procedures, and hyper-parameters are found in the Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.451, + 0.166, + 0.466 + ], + "angle": 0, + "content": "6. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.469, + 0.473, + 0.786 + ], + "angle": 0, + "content": "To conduct this work, we have collected a diverse set of 10 static sequences, recorded in both day and night conditions across North America. To this goal, we equipped a test vehicle with a NIR gated stereo camera setup (Bright-Way Vision), an automotive RGB stereo camera (OnSemi AR0230), a LiDAR sensor (Velodyne VLS128) and a GNSS with IMU (Xsens MTi-7), as shown in Fig. 4. Each gated camera has a resolution of \\(1280 \\times 720\\) pixels, 10 bit depth and runs at \\(120 \\mathrm{~Hz}\\), split up to collect the three active and one passive slice. The illuminator source consists of two vertical-cavity surface-emitting laser (VCSEL) modules, which illuminate the scene with a laser pulse with duration of 240-370 ns and a wavelength of \\(808 \\mathrm{~nm}\\). The RGB cameras provide 12 bit HDR images with resolution of \\(1920 \\times 1080\\) pixels and \\(30 \\mathrm{~Hz}\\) frame-rate. The LiDAR has a vertical resolution of 128 lines and \\(10 \\mathrm{~Hz}\\) framerate, while the GNSS sensor runs at \\(4 \\mathrm{~Hz}\\). Example captures from the dataset are being visualized in Fig. 4. In total, we collect 2650 samples, captured in both day (1223 samples) and night (1427 samples). We divide it in training, validation and test splits with a 50-25-25 split." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.787, + 0.47, + 0.847 + ], + "angle": 0, + "content": "As ground-truth, we construct a large-scale ground-truth pointcloud by aggregating LiDAR scans with LIO-SAM [78] and removing noisy points. Additional details on the accumulation are provided in the Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.853, + 0.199, + 0.868 + ], + "angle": 0, + "content": "7. Assessment" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "In this section, we validate the proposed method quantitatively and qualitatively. Specifically, we investigate scene" + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.077, + 0.905, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.259, + 0.9, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.411, + 0.895, + 0.495 + ], + "angle": 0, + "content": "Figure 4. Top: Example captures from our collected dataset across different urban and suburban areas in North America. From left to right: RGB image, active gated slices (with red for slice 1, green for slice 2 and blue for slice 3), passive slice, projected LiDAR scan, accumulated LiDAR. Bottom: Sensors setup with LiDAR, stereo Gated camera, stereo RGB camera, IMU and GNSS." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.51, + 0.892, + 0.602 + ], + "angle": 0, + "content": "reconstruction at both day and night, using novel depth and view synthesis for 2D evaluation, and surface reconstruction for the 3D evaluation. To this end, we compare our approach to state-of-the-art feed-forward depth estimation algorithms and neural scene reconstruction methods. We also conduct ablation experiments to validate our design choices." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Depth Reconstruction We assess the quality of depth synthesis of Gated Fields for camera poses unseen during training. We use as ground truth the accumulated and filtered LiDAR pointcloud. Unlike previous works relying on single LiDAR scans for evaluation [35, 89, 90], we use as ground truth an accumulated LiDAR pointcloud as described in Sec. 6. This allows us to evaluate the depth reconstruction up to \\(160\\mathrm{m}\\) accurately and without bias. We follow previous works [89, 90] and use as depth evaluation metrics RMSE, MAE, ARD, \\(\\sigma_{i} < 1.25^{i}, i \\in \\{1,2,3\\}\\). We compare our method against 9 feed-forward depth estimation methods, namely SimIPU [53], AdaBins [10], DPT [70], DepthFormer [54] and CREStereo [50] for monocular and stereo RGB methods, Gated2Gated [89] and GatedStereo [90] for monocular and stereo gated estimation methods. We also compare our method against depths rendered with other neural reconstruction algorithms, using RGB images [7],[63],[29],[26] LiDAR [84], RGB+LiDAR [86], [37], and a varying-appearance method [29] for gated" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10535" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.068, + 0.899, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.497, + 0.894, + 0.568 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison of the proposed Gated Fields and state-of-the-art depth estimation approaches, including LiDAR-NeRF [84], StreetSurf [37], NLSPN [67], and Gated Stereo [90]. Compared to baseline methods, we are able to reconstruct fine geometry details like branches or poles, also for far distances. Unlike RGB methods, Gated Fields is unaffected by poor ambient lighting, and unlike LiDAR-based methods it is able to reconstruct sharp object discontinuities. The active gated slices are visualized in red for slice 1, green for slice 2 and blue for slice 3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.579, + 0.473, + 0.852 + ], + "angle": 0, + "content": "captures. Results for day and night sequences are presented in Tab. 1. We outperform the next best neural field method by \\(21.87\\%\\) MAE and \\(30.35\\%\\) in RMSE. For night sequences the performance difference sharpens, with Gated Fields outperforming the best RGB-based method [50] by \\(3.14\\mathrm{m}\\) MAE. This performance decline is to be attributed to the limited pixel information present in RGB captures taken at night time, making impossible to learn a meaningful 3D representation of the scene, as shown qualitatively in Fig. 5. LiDAR-based methods are unaffected by the change in illumination, but suffer from the limited sensor resolution. On the other hand, gated cameras retrieve information-rich captures both day and night, which Gated Fields can explicitly leverage during training. We confirm that employing state-of-the-art neural field methods on gated captures does not yield accurate results, as they are not able to model the gated imaging formation and can only fit the ambient light component." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "3D Reconstruction We evaluate the 3D scene reconstruction capabilities of Gated Fields using the accumulated Li-DAR pointcloud as ground truth. We follow [20] and extract" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.579, + 0.892, + 0.791 + ], + "angle": 0, + "content": "for both the 3D ground truth pointcloud and different neural field-based methods a voxelized occupancy grid of the scene, and compute intersection over union (IoU), Precision and Recall between ground truth voxels and estimated ones. Quantitative results are shown in Tab. 3. Our method outperforms RGB baselines [7, 29] by an average of \\(15\\%\\) IoU. RGB baselines using additional LiDAR sensor [37, 86] data partially improve the results, but such methods are still unable to reconstruct finer surfaces details and struggle at night. On the other hand, Gated Fields is able to recover finer geometries using gated, illumination and depth cues, and the quality does not degrade with diminishing ambient light, as shown in Fig. 5. See details on evaluation and further qualitative results in the Supplementary Material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Novel View Synthesis For novel view synthesis, we compare our method with Mip-NeRF360 [7], a state-of-the-art neural radiance field-based method, and K-Planes [29], to implicitly model the time-varying appearance of the static scene. Mip-NeRF [7] struggles to reconstruct novel views due to the inherent difficulty of modeling the gating imaging effects, resulting in a PSNR of 17.16dB. By learning" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10536" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.08, + 0.462, + 0.497 + ], + "angle": 0, + "content": "
MethodModalityRMSE [m]ARDMAE [m]δ1 [%]δ2 [%]δ3 [%]
Test Data - Night (Evaluated on Accumulated LiDAR Ground-Truth Points)
2D DEPTH COMPARISONGATED2GATED [89]Gated13.330.288.5761.7890.3094.90
GATEDSTEREO [90]Stereo-Gated10.100.205.9782.8693.3596.52
SIMIPU [53]RGB19.330.4414.2140.6777.9989.80
ADABINS [10]RGB21.140.3814.2851.7280.0090.64
DPT [70]RGB14.170.289.9062.8288.3294.42
DEPTHFORMER [54]RGB14.320.2910.0861.1987.7294.57
CRESTEREO [50]Stereo-RGB14.220.137.2782.2190.1094.32
NLSPN [67]RGB+LiDAR11.120.186.4877.1590.8196.00
MIPNERF360 [7]RGB23.800.5116.4341.6061.7576.16
K-PLANES [29]RGB19.700.4113.6644.3966.2081.46
RAWNERF [63]RGB27.460.6419.7534.3254.3369.41
DEPTH-NERF [26]RGB15.230.2610.0461.6786.2393.87
SUDS [86]RGB+LiDAR11.070.176.1779.4988.4195.31
STREETSURF [37]RGB+LiDAR10.860.155.9082.1693.5796.96
LiDAR-NERF [50]LiDAR10.210.124.7287.7195.0597.78
GATED FIELDS [50]Gated7.920.124.1390.6195.7697.90
Test Data - Day (Evaluated on Accumulated LiDAR Ground-Truth Points)
2D DEPTH COMPARISONGATED2GATED [89]Gated9.260.226.6958.4693.7097.38
GATEDSTEREO [90]Stereo-Gated6.320.093.3092.8697.3198.53
SIMIPU [53]RGB13.540.3110.0852.9086.4995.65
ADABINS [10]RGB12.740.258.3969.8489.1395.52
DPT [70]RGB10.340.217.0877.6194.2997.24
DEPTHFORMER [54]RGB9.060.196.0981.1994.1497.42
CRESTEREO [50]Stereo-RGB7.350.093.4594.4897.2398.50
NLSPN [67]RGB+LiDAR10.340.175.9777.8391.1696.11
MIPNERF360 [7]RGB16.910.319.5370.8984.4390.92
K-PLANES [29]RGB12.370.248.5563.1679.6891.70
RAWNERF [63]RGB15.100.239.3865.9084.9092.43
DEPTH-NERF [26]RGB10.340.176.0778.9790.7596.53
SUDS [86]RGB+LiDAR9.110.175.8480.9695.0898.33
STREETSURF [37]RGB+LiDAR9.600.135.3583.9495.3298.56
LiDAR-NERF [50]LiDAR8.130.103.8688.9995.4998.06
GATED FIELDS [50]Gated6.150.092.9193.8897.3298.75
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.503, + 0.47, + 0.545 + ], + "angle": 0, + "content": "Table 1. Comparison of our proposed approach and state-of-the-art approaches on depth synthesis. Best results in each category are in bold and second best are underlined." + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.559, + 0.445, + 0.716 + ], + "angle": 0, + "content": "
Im. Formation ModelDepth Sup.Normal Active Illum.RMSE [m]MAE [m]PSNR [dB]SSIM
Test Data - Night
ABLATIONEnd2End [62]XXXX27.3419.9517.430.633
GatedXXXX9.348.3523.470.889
Gated✓XXX6.725.4325.840.908
Gated✓✓✓X3.782.8230.660.946
Gated✓✓✓✓3.392.5130.910.95
Test Data - Day
ABLATIONEnd2End [62]XXXX30.0319.2216.770.66
GatedXXXX11.646.3426.770.915
Gated✓XXX9.204.2226.880.922
Gated✓✓✓X9.454.3132.150.946
Gated✓✓✓✓8.884.1232.280.948
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.721, + 0.47, + 0.763 + ], + "angle": 0, + "content": "Table 2. Ablation studies of the Gated Fields contributions, on a subset of the test dataset. We investigate different image formation models, neural fields components and supervision losses." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.777, + 0.468, + 0.866 + ], + "angle": 0, + "content": "a time-varying appearance, K-Planes improves the quality reaching 27.42dB PSNR for day but only 19.35dB for night, as the model fails to learn an accurate scene geometry representation without ambient light information. Gated Fields outperforms these baselines in both day and night, reaching a PSNR of 32.28dB." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Ablation Experiments To assess the role and contribution of the different components of our method, we conduct" + }, + { + "type": "table", + "bbox": [ + 0.537, + 0.078, + 0.852, + 0.164 + ], + "angle": 0, + "content": "
MethodModalityIoU [%]Precision [%]Recall [%]
3D REC.MIPNERF360 [7]RGB6.327.3431.21
STREETSURF [37]RGB+LiDAR5.416.3127.35
SUDS [86]RGB+LiDAR8.969.8849.09
LiDAR-NERF [84]LiDAR20.0332.3834.44
Gated Fields (ours)Gated22.2525.0166.51
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.169, + 0.892, + 0.238 + ], + "angle": 0, + "content": "Table 3. Comparison of Gated Fields and state-of-the-art scene reconstruction methods. We evaluate over 3D occupancy reconstruction, using as ground truth the voxelized accumulated LiDAR pointcloud. Best results in each category are in bold and second best are underlined." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.251, + 0.892, + 0.493 + ], + "angle": 0, + "content": "an ablation study in Tab. 2. In particular, we consider as starting point a single neural field directly inferring one intensity value for each of the four slices (3 active + 1 passive), and obtain an average MAE of \\(19.58\\mathrm{m}\\). By separately predicting ambient light and reflectance, and reconstructing the gated image as in Eq. (2), we significantly improve the MAE to \\(7.34\\mathrm{m}\\). However, this approach still performs poorly on flat-color areas during the day and in unilluminated areas during the night due to lack of any depth cue. By adding the depth supervision, we are able to supervise also such areas and the PSNR improves by \\(4.83\\mathrm{dB}\\). By adding the angular-dependent attenuation and regularizing the reflectance in Eq. (17), the model is able to disentangle the material properties from other spurious effects. Finally, by explicitly modeling the shadow, casted by the illuminator, we improve final depth reconstruction to \\(3.32\\mathrm{m}\\) MAE." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.507, + 0.619, + 0.523 + ], + "angle": 0, + "content": "8. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.531, + 0.892, + 0.788 + ], + "angle": 0, + "content": "We introduce Gated Fields, a neural rendering method capable of reconstructing scene geometry from video captures of active time-gated cameras. The method hinges on a differentiable gated image formation as part of the rendering formulation, and it jointly learns geometry, ambient light and surface proprieties, represented implicitly as neural field components, alongside illumination and gating parameters, represented with physics-based models. Extensive experiments on real-world large-scale scenes validate that our method is able to precisely reconstruct a 3D scene both in day and night-time conditions. Our approach outperforms existing RGB and LiDAR methods by \\(21.87\\%\\) on MAE, as well as baseline methods using gated captures by \\(31.67\\%\\). In the future, we hope to extend the proposed approach by \"closing the loop\" and providing dynamic feedback to the gated acquisition, allowing for adaptive gated scene reconstructions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.794, + 0.892, + 0.899 + ], + "angle": 0, + "content": "Acknowledgments This work was supported by the AI-SEE project with funding from the FFG, BMBF, and NRC-IRA. Felix Heide was supported by an NSF CAREER Award (2047359), a Packard Foundation Fellowship, a Sloan Research Fellowship, a Sony Young Faculty Award, a Project X Innovation Award, and an Amazon Science Research Award." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10537" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.116, + 0.469, + 0.157 + ], + "angle": 0, + "content": "[1] Amit Adam, Christoph Dann, Omer Yair, Shai Mazor, and Sebastian Nowozin. Bayesian time-of-flight for realtime shape, illumination and albedo. 39(5):851-864, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.16, + 0.469, + 0.227 + ], + "angle": 0, + "content": "[2] Ashutosh Agarwal and Chetan Arora. Attention attention everywhere: Monocular depth prediction with skip attention. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5861-5870, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.231, + 0.468, + 0.27 + ], + "angle": 0, + "content": "[3] Pierre Andersson. Long-range three-dimensional imaging using range-gated laser radar images. 45(3):034301, 2006. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.274, + 0.469, + 0.342 + ], + "angle": 0, + "content": "[4] Benjamin Attal, Eliot Laidlaw, Aaron Gokaslan, Changil Kim, Christian Richardt, James Tompkin, and Matthew O'Toole. Törf: Time-of-flight radiance fields for dynamic scene view synthesis. Advances in neural information processing systems, 34:26289-26301, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.345, + 0.469, + 0.399 + ], + "angle": 0, + "content": "[5] Abhishek Badki, Alejandro Troccoli, Kihwan Kim, Jan Kautz, Pradeep Sen, and Orazio Gallo. Bi3D: Stereo depth estimation via binary classifications. In arXiv preprint arXiv:2005.07274, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.403, + 0.469, + 0.456 + ], + "angle": 0, + "content": "[6] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.46, + 0.469, + 0.528 + ], + "angle": 0, + "content": "[7] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 1, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.531, + 0.469, + 0.571 + ], + "angle": 0, + "content": "[8] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. ICCV, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.574, + 0.469, + 0.628 + ], + "angle": 0, + "content": "[9] Florent Bartoccioni, Éloi Zablocki, Patrick Pérez, Matthieu Cord, and Karteek Alahari. Lidartouch: Monocular metric depth estimation with a few-beam lidar. Computer Vision and Image Understanding, 227:103601, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.632, + 0.469, + 0.686 + ], + "angle": 0, + "content": "[10] Shariq Farooq Bhat, Ibraheem Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4009-4018, 2021. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.688, + 0.469, + 0.742 + ], + "angle": 0, + "content": "[11] Mario Bijelic, Tobias Gruber, and Werner Ritter. A benchmark for lidar sensors in fog: Is detection breaking down? In 2018 IEEE Intelligent Vehicles Symposium (IV), pages 760-767, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.746, + 0.469, + 0.786 + ], + "angle": 0, + "content": "[12] Mario Bijelic, Tobias Gruber, and Werner Ritter. Benchmarking image sensors under adverse weather conditions for autonomous driving. 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.789, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[13] Mario Bijelic, Tobias Gruber, and Werner Ritter. Benchmarking image sensors under adverse weather conditions for autonomous driving. In 2018 IEEE Intelligent Vehicles Symposium (IV), pages 1773-1779. IEEE, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.846, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[14] Mario Bijelic, Tobias Gruber, Fahim Mannan, Florian Kraus, Werner Ritter, Klaus Dietmayer, and Felix Heide. Seeing through fog without seeing fog: Deep multimodal sensor fusion in unseen adverse weather. In The IEEE" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.116, + 0.469, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.543, + 0.094, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.123, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[15] Mark Boss, Andreas Engelhardt, Abhishek Kar, Yuanzhen Li, Deqing Sun, Jonathan Barron, Hendrik Lensch, and Varun Jampani. *Samurai: Shape and material from unconstrained real-world arbitrary image collections*. Advances in Neural Information Processing Systems, 35:26389-26403, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.207, + 0.892, + 0.274 + ], + "angle": 0, + "content": "[16] Aljaz Bozic, Pablo Palafox, Justus Thies, Angela Dai, and Matthias Nießner. Transformerfusion: Monocular rgb scene reconstruction using transformers. Advances in Neural Information Processing Systems, 34:1403-1414, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.278, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[17] Jens Busck. Underwater 3-D optical imaging with a gated viewing laser radar. 2005. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.307, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[18] Jens Busck and Henning Heiselberg. Gated viewing and high-accuracy three-dimensional laser radar. 43(24):4705–10, 2004. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.349, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[19] Jens Busck and Henning Heiselberg. Gated viewing and high-accuracy three-dimensional laser radar. Applied optics, 43(24):4705-4710, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.392, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[20] Anh-Quan Cao and Raoul de Charette. Scenerf: Self-supervised monocular 3d scene reconstruction with radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9387-9398, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.463, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[21] A. Carballo, J. Lambert, A. Monrroy, D. Wong, P. Narksri, Y. Kitsukawa, E. Takeuchi, S. Kato, and K. Takeda. Libre: The multiple 3d lidar dataset. In IEEE Intelligent Vehicles Symposium (IV), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.52, + 0.892, + 0.573 + ], + "angle": 0, + "content": "[22] Jia-Ren Chang and Yong-Sheng Chen. Pyramid stereo matching network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5410-5418, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.577, + 0.892, + 0.629 + ], + "angle": 0, + "content": "[23] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.633, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[24] Jaesung Choe, Kyungdon Joo, Tooba Imtiaz, and In So Kweon. Volumetric propagation network: Stereo-lidar fusion for long-range depth estimation. IEEE Robotics and Automation Letters, 6(3):4672-4679, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.69, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[25] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 303-312, 1996. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.747, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[26] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nef: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12882-12891, 2022. 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.818, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[27] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. In NeurIPS, pages 2366-2374, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.861, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[28] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.094, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10538" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501-5510, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.123, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[29] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12479–12488, 2023. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.208, + 0.471, + 0.25 + ], + "angle": 0, + "content": "[30] Ravi Garg, B.G. Vijay Kumar, Gustavo Carneiro, and Ian Reid. Unsupervised CNN for single view depth estimation: Geometry to the rescue. In ECCV, pages 740-756, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.252, + 0.471, + 0.293 + ], + "angle": 0, + "content": "[31] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, pages 3354-3361, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.296, + 0.471, + 0.364 + ], + "angle": 0, + "content": "[32] Clément Godard, Oisin Mac Aodha, and Gabriel J Bros-tow. Unsupervised monocular depth estimation with left-right consistency. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 270-279, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.367, + 0.471, + 0.436 + ], + "angle": 0, + "content": "[33] Clément Godard, Oisin Mac Aodha, Michael Firman, and Gabriel J Brostow. Digging into self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3828-3838, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.439, + 0.471, + 0.479 + ], + "angle": 0, + "content": "[34] Yoav Grauer. Active gated imaging in driver assistance system. Advanced Optical Technologies, 3(2):151-160, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.482, + 0.471, + 0.538 + ], + "angle": 0, + "content": "[35] Tobias Gruber, Frank Julca-Aguilar, Mario Bijelic, and Felix Heide. Gated2depth: Real-time dense lidar from gated images. In The IEEE International Conference on Computer Vision (ICCV), 2019. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.54, + 0.471, + 0.61 + ], + "angle": 0, + "content": "[36] Vitor Guizilini, Rares Ambrus, Sudeep Pillai, Allan Ravens \ntos, and Adrien Gaidon. 3d packing for self-supervised \nmonocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern \nRecognition, pages 2485-2494, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.612, + 0.471, + 0.681 + ], + "angle": 0, + "content": "[37] Jianfei Guo, Nianchen Deng, Xinyang Li, Yeqi Bai, Botian Shi, Chiyu Wang, Chenjing Ding, Dongliang Wang, and Yikang Li. Streetsurf: Extending multi-view implicit surface reconstruction to street views. arXiv preprint arXiv:2306.04988, 2023. 1, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.684, + 0.471, + 0.726 + ], + "angle": 0, + "content": "[38] Miles Hansard, Seungkyu Lee, Ouk Choi, and Radu Patrice Horaud. Time-of-flight cameras: principles, methods and applications. Springer Science & Business Media, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.728, + 0.471, + 0.755 + ], + "angle": 0, + "content": "[39] Paul Heckman and Robert T. Hodgson. Underwater optical range gating. 3(11):445-448, 1967. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.758, + 0.471, + 0.799 + ], + "angle": 0, + "content": "[40] Mu Hu, Shuling Wang, Bin Li, Shiyu Ning, Li Fan, and Xiaojin Gong. Towards precise and efficient image guided depth completion. 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.802, + 0.471, + 0.843 + ], + "angle": 0, + "content": "[41] Shengyu Huang, Zan Gojcic, Zian Wang, Francis Williams, Yoni Kasten, Sanja Fidler, Konrad Schindler, and Or Litany. Neural lidar fields for novel view synthesis. 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.846, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[42] Maximilian Jaritz, Raoul De Charette, Emilie Wirbel, Xavier Perrotton, and Fawzi Nashashibi. Sparse and dense data with cnns: Depth completion and semantic segmentation. pages 52-60, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.092, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[43] Maria Jokela, Matti Kutila, and Pasi Pyykönen. Testing and validation of automotive point-cloud sensors in adverse weather conditions. Applied Sciences, 9, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.137, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[44] Frank Julca-Aguilar, Jason Taylor, Mario Bijelic, Fahim Mannan, Ethan Tseng, and Felix Heide. Gated3d: Monocular 3d object detection from temporal illumination cues. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2938-2948, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.208, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[45] Alex Kendall, Hayk Martirosyan, Saumitro Dasgupta, Peter Henry, Ryan Kennedy, Abraham Bachrach, and Adam Bry. End-to-end learning of geometry and context for deep stereo regression. In ICCV, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.265, + 0.892, + 0.319 + ], + "angle": 0, + "content": "[46] Andreas Kolb, Erhardt Barth, Reinhard Koch, and Rasmus Larsen. Time-of-flight cameras in computer graphics. In Computer Graphics Forum, pages 141-159. Wiley Online Library, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.322, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[47] Abhijit Kundu, Kyle Genova, Xiaqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12871-12881, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.42, + 0.892, + 0.461 + ], + "angle": 0, + "content": "[48] Robert Lange. 3D time-of-flight distance measurement with custom solid-state image sensors in CMOS/CCD-technology. 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.463, + 0.892, + 0.532 + ], + "angle": 0, + "content": "[49] Deborah Levy, Amit Peleg, Naama Pearl, Dan Rosenbaum, Derya Akkaynak, Simon Korman, and Tali Treibitz. Seathru-nerf: Neural radiance fields in scattering media. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 56–65, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.534, + 0.892, + 0.617 + ], + "angle": 0, + "content": "[50] Jiankun Li, Peisen Wang, Pengfei Xiong, Tao Cai, Ziwei Yan, Lei Yang, Jiangyu Liu, Haoqiang Fan, and Shuaicheng Liu. Practical stereo matching via cascaded recurrent network with adaptive correlation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16263-16272, 2022. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.619, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[51] Zhengqi Li, Tali Dekel, Forrester Cole, Richard Tucker, Noah Snavely, Ce Liu, and William T. Freeman. Manequinchallenge: Learning the depths of moving people by watching frozen people. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(12):4229-4241, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.704, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[52] Zhaoshuo Li, Xingtong Liu, Nathan Drenkow, Andy Ding, Francis X. Creighton, Russell H. Taylor, and Mathias Unberath. Revisiting stereo depth estimation from a sequence-to-sequence perspective with transformers. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6177-6186, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.789, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[53] Zhenyu Li, Zehui Chen, Ang Li, Liangji Fang, Qinhong Jiang, Xianming Liu, Junjun Jiang, Bolei Zhou, and Hang Zhao. Simipu: Simple 2d image and 3d point cloud unsupervised pre-training for spatial-aware visual representations. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1500-1508, 2022. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[54] Zhenyu Li, Zehui Chen, Xianming Liu, and Junjun Jiang. Depthformer: Depthformer: Exploiting long-range corre" + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.092, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10539" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "lation and local information for accurate monocular depth estimation. arXiv preprint arXiv:2203.14211, 2022. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.121, + 0.47, + 0.189 + ], + "angle": 0, + "content": "[55] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.19, + 0.469, + 0.259 + ], + "angle": 0, + "content": "[56] Jeffrey Yunfan Liu, Yun Chen, Ze Yang, Jingkang Wang, Sivabalan Manivasagam, and Raquel Urtasun. Real-time neural rasterization for large scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8416-8427, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.261, + 0.469, + 0.302 + ], + "angle": 0, + "content": "[57] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. International Conference on Learning Representations, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.302, + 0.469, + 0.344 + ], + "angle": 0, + "content": "[58] Fangchang Ma and Sertac Karaman. Sparse-to-dense: Depth prediction from sparse depth samples and a single image. pages 1-8, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.344, + 0.469, + 0.399 + ], + "angle": 0, + "content": "[59] Anagh Malik, Parsa Mirdehghan, Sotiris Nousias, Kiriakos N Kutulakos, and David B Lindell. Transient neural radiance fields for lidar view synthesis and 3d reconstruction. arXiv preprint arXiv:2307.09555, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.401, + 0.469, + 0.453 + ], + "angle": 0, + "content": "[60] N. Mayer, E. Ilg, P. Häusser, P. Fischer, D. Cremers, A. Dosovitskiy, and T. Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In CVPR, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.456, + 0.469, + 0.51 + ], + "angle": 0, + "content": "[61] Andreas Meuleman, Yu-Lun Liu, Chen Gao, Jia-Bin Huang, Changil Kim, Min H. Kim, and Johannes Kopf. Progressively optimized local radiance fields for robust view synthesis. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.511, + 0.469, + 0.566 + ], + "angle": 0, + "content": "[62] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.567, + 0.469, + 0.649 + ], + "angle": 0, + "content": "[63] Ben Mildenhall, Peter Hedman, Ricardo Martin-Brualla, Pratul P Srinivasan, and Jonathan T Barron. Nerf in the dark: High dynamic range view synthesis from noisy raw images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16190-16199, 2022. 1, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.651, + 0.469, + 0.705 + ], + "angle": 0, + "content": "[64] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.707, + 0.469, + 0.789 + ], + "angle": 0, + "content": "[65] Zak Murez, Tarrence Van As, James Bartolozzi, Ayan Sinha, Vijay Badrinarayanan, and Andrew Rabinovich. Atlas: End-to-end 3d scene reconstruction from posed images. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VII 16, pages 414-431. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.79, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[66] Julian Ost, Issam Laradji, Alejandro Newell, Yuval Bahat, and Felix Heide. Neural point light fields. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[67] Jinsun Park, Kyungdon Joo, Zhe Hu, Chi-Kuei Liu, and In So Kweon. Non-local spatial propagation network for depth completion. In Proc. of European Conference on Computer Vision (ECCV), 2020. 2, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.175 + ], + "angle": 0, + "content": "[68] Vivek Pradeep, Christoph Rhemann, Shahram Izadi, Christopher Zach, Michael Bleyer, and Steven Bathiche. Monofusion: Real-time 3d reconstruction of small scenes with a single web camera. In 2013 IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 83-88. IEEE, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.177, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[69] Andrea Ramazzina, Mario Bijelic, Stefanie Walz, Alessandro Sanvito, Dominik Scheuble, and Felix Heide. Scatternerf: Seeing through fog with physically-based inverse neural rendering. arXiv preprint arXiv:2305.02103, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.233, + 0.894, + 0.287 + ], + "angle": 0, + "content": "[70] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12179-12188, 2021. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.288, + 0.892, + 0.342 + ], + "angle": 0, + "content": "[71] Konstantinos Rematas, Andrew Liu, Pratul P. Srinivasan, Jonathan T. Barron, Andrea Tagliasacchi, Tom Funkhouser, and Vittorio Ferrari. Urban radiance fields. CVPR, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.344, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[72] Gernot Riegler, Ali Osman Ulusoy, Horst Bischof, and Andreas Geiger. Octnetfusion: Learning depth fusion from data. In 2017 International Conference on 3D Vision (3DV), pages 57-66. IEEE, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.4, + 0.892, + 0.481 + ], + "angle": 0, + "content": "[73] Barbara Roessle, Jonathan T Barron, Ben Mildenhall, Pratul P Srinivasan, and Matthias Nießner. Dense depth priors for neural radiance fields from sparse input views. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12892-12901, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.483, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[74] Viktor Rudnev, Mohamed Elgharib, William Smith, Lingjie Liu, Vladislav Golyanik, and Christian Theobalt. Nerf for outdoor scene relighting. In European Conference on Computer Vision, pages 615-631. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.539, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[75] Michael Schober, Amit Adam, Omer Yair, Shai Mazor, and Sebastian Nowozin. Dynamic time-of-flight. In CVPR, pages 6109-6118, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.581, + 0.892, + 0.649 + ], + "angle": 0, + "content": "[76] Thomas Schöps, Torsten Sattler, Christian Hane, and Marc Pollefeys. 3d modeling on the go: Interactive 3d reconstruction of large-scale scenes on mobile devices. In 2015 International Conference on 3D Vision, pages 291-299. IEEE, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.65, + 0.892, + 0.678 + ], + "angle": 0, + "content": "[77] Brent Schwarz. Lidar: Mapping the world in 3D. Nature Photonics, 4(7):429, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.679, + 0.892, + 0.748 + ], + "angle": 0, + "content": "[78] Tixiao Shan, Brendan Englot, Drew Meyers, Wei Wang, Carlo Ratti, and Daniela Rus. Lio-sam: Tightly-coupled lidar inertial odometry via smoothing and mapping. In 2020 IEEE/RSJ international conference on intelligent robots and systems (IROS), pages 5135-5142. IEEE, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.749, + 0.894, + 0.817 + ], + "angle": 0, + "content": "[79] Jiaming Sun, Yiming Xie, Linghao Chen, Xiaowei Zhou, and Hujun Bao. Neuralrecon: Real-time coherent 3d reconstruction from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15598-15607, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.818, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[80] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8248–8258, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10540" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.093, + 0.47, + 0.174 + ], + "angle": 0, + "content": "[81] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, et al. Nerfstudio: A modular framework for neural radiance field development. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-12, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.176, + 0.471, + 0.232 + ], + "angle": 0, + "content": "[82] Jiexiong Tang, John Folkesson, and Patric Jensfelt. Sparse2dense: From direct sparse odometry to dense 3-d reconstruction. IEEE Robotics and Automation Letters, 4 (2):530-537, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.233, + 0.471, + 0.286 + ], + "angle": 0, + "content": "[83] Jie Tang, Fei-Peng Tian, Wei Feng, Jian Li, and Ping Tan. Learning guided convolutional network for depth completion. IEEE Transactions on Image Processing, 30:1116-1129, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.288, + 0.471, + 0.357 + ], + "angle": 0, + "content": "[84] Tang Tao, Longfei Gao, Guangrun Wang, Peng Chen, Dayang Hao, Xiaodan Liang, Mathieu Salzmann, and Kaicheng Yu. Lidar-nerf: Novel lidar view synthesis via neural radiance fields. arXiv preprint arXiv:2304.10406, 2023. 1, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.358, + 0.471, + 0.413 + ], + "angle": 0, + "content": "[85] Fabio Tosi, Alessio Tonioni, Daniele De Gregorio, and Matteo Poggi. Nerf-supervised deep stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 855–866, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.414, + 0.471, + 0.468 + ], + "angle": 0, + "content": "[86] Haithem Turki, Jason Y Zhang, Francesco Ferroni, and Deva Ramanan. Suits: Scalable urban dynamic scenes. In Computer Vision and Pattern Recognition (CVPR), 2023. 1, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.469, + 0.471, + 0.523 + ], + "angle": 0, + "content": "[87] Jonas Uhrig, Nick Schneider, Lukas Schneider, Uwe Franke, Thomas Brox, and Andreas Geiger. Sparsity invariant cnns. In International Conference on 3D Vision (3DV), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.525, + 0.471, + 0.607 + ], + "angle": 0, + "content": "[88] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481–5490. IEEE, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.608, + 0.471, + 0.664 + ], + "angle": 0, + "content": "[89] Amanpreet Walia, Stefanie Walz, Mario Bijelic, Fahim Mannan, Frank Julca-Aguilar, Michael Langer, Werner Ritter, and Felix Heide. Gated2gated: Self-supervised depth estimation from gated images. 2022. 2, 3, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.665, + 0.471, + 0.747 + ], + "angle": 0, + "content": "[90] Stefanie Walz, Mario Bijelic, Andrea Ramazzina, Amanpreet Walia, Fahim Mannan, and Felix Heide. Gated stereo: Joint depth estimation from gated and wide-baseline active stereo cues. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13252-13262, 2023. 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.748, + 0.471, + 0.816 + ], + "angle": 0, + "content": "[91] Haoyuan Wang, Xiaogang Xu, Ke Xu, and Rynson WH Lau. Lighting up nerf via unsupervised decomposition and enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12632-12641, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.818, + 0.471, + 0.873 + ], + "angle": 0, + "content": "[92] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.873, + 0.471, + 0.903 + ], + "angle": 0, + "content": "[93] Zian Wang, Tianchang Shen, Jun Gao, Shengyu Huang, Jacob Munkberg, Jon Hasselgren, Zan Gojcic, Wenzheng" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.093, + 0.471, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "Chen, and Sanja Fidler. Neural fields meet explicit geometric representations for inverse rendering of urban scenes. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.149, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[94] Alex Wong and Stefano Soatto. Unsupervised depth completion with calibrated backprojection layers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12747-12756, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.206, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[95] Gengshan Yang, Joshua Manela, Michael Happold, and Deva Ramanan. Hierarchical deep stereo matching on high-resolution images. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.262, + 0.892, + 0.33 + ], + "angle": 0, + "content": "[96] Zhenfei Yang, Fei Gao, and Shaojie Shen. Real-time monocular dense mapping on aerial robots using visual-inertial fusion. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 4552-4559. IEEE, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.332, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[97] Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasun. Unisim: A neural closed-loop sensor simulator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1389-1399, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.403, + 0.892, + 0.472 + ], + "angle": 0, + "content": "[98] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoptrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5752-5761, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.473, + 0.892, + 0.529 + ], + "angle": 0, + "content": "[99] Junge Zhang, Feihu Zhang, Shaochen Kuang, and Li Zhang. Nerf-lidar: Generating realistic lidar point clouds with neural radiance fields. arXiv preprint arXiv:2304.14811, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.53, + 0.892, + 0.572 + ], + "angle": 0, + "content": "[100] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.573, + 0.892, + 0.643 + ], + "angle": 0, + "content": "[101] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Ner-factor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (ToG), 40(6):1-18, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.643, + 0.892, + 0.699 + ], + "angle": 0, + "content": "[102] Yongjian Zhang, Longguang Wang, Kunhong Li, Zhiheng Fu, and Yulan Guo. Slfnet: A stereo and lidar fusion network for depth completion. IEEE Robotics and Automation Letters, 7(4):10605-10612, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.7, + 0.892, + 0.742 + ], + "angle": 0, + "content": "[103] Tinghui Zhou, Matthew Brown, Noah Snavely, and David G. Lowe. Unsupervised learning of depth and egomotion from video. In CVPR, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.742 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10541" + } + ] +] \ No newline at end of file diff --git a/2024/Gated Fields_ Learning Scene Reconstruction from Gated Videos/babd13ae-5bce-44f5-bd0f-203ff88804c8_origin.pdf b/2024/Gated Fields_ Learning Scene Reconstruction from Gated Videos/babd13ae-5bce-44f5-bd0f-203ff88804c8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9045315b4419e939c0678590124ad31c77ef767f --- /dev/null +++ b/2024/Gated Fields_ Learning Scene Reconstruction from Gated Videos/babd13ae-5bce-44f5-bd0f-203ff88804c8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2969987f3b7d55e1b5f90878c6e5e050a1398e13ba13dc54c6b899cce31cb951 +size 1673581 diff --git a/2024/Gated Fields_ Learning Scene Reconstruction from Gated Videos/full.md b/2024/Gated Fields_ Learning Scene Reconstruction from Gated Videos/full.md new file mode 100644 index 0000000000000000000000000000000000000000..481513d519c04cd380a8e060f478ea430143dd94 --- /dev/null +++ b/2024/Gated Fields_ Learning Scene Reconstruction from Gated Videos/full.md @@ -0,0 +1,392 @@ +# Gated Fields: Learning Scene Reconstruction from Gated Videos + +Andrea Ramazzina $^{1*}$ Stefanie Walz $^{1,2*}$ Pragyan Dahal $^{3}$ Mario Bijelic $^{4,5}$ Felix Heide $^{4,5}$ + +$^{1}$ Mercedes-Benz $^{2}$ Saarland University $^{3}$ Politecnico di Milano $^{4}$ Torch Robotics $^{5}$ Princeton University + +# Abstract + +Reconstructing outdoor 3D scenes from temporal observations is a challenge that recent work on neural fields has offered a new avenue for. However, existing methods that recover scene properties, such as geometry, appearance, or radiance, solely from RGB captures often fail when handling poorly-lit or texture-deficient regions. Similarly, recovering scenes with scanning LiDAR sensors is also difficult due to their low angular sampling rate which makes recovering expansive real-world scenes difficult. Tackling these gaps, we introduce Gated Fields - a neural scene reconstruction method that utilizes active gated video sequences. To this end, we propose a neural rendering approach that seamlessly incorporates time-gated capture and illumination. Our method exploits the intrinsic depth cues in the gated videos, achieving precise and dense geometry reconstruction irrespective of ambient illumination conditions. We validate the method across day and night scenarios and find that Gated Fields compares favorably to RGB and LiDAR reconstruction methods. Our code and datasets are available here1. + +# 1. Introduction + +Large-scale outdoor scene reconstruction is essential for advancing autonomous robotics, drones, and driver-assistance systems, serving as the foundation for scene understanding, safe navigation, dataset generation and validation. Existing works in this domain [68, 76, 96] have typically adopted a two-step approach. Initially, they infer depth maps from different poses, utilizing time-of-flight sensors or RGB captures. Subsequently, these depth estimates are fused to produce a coherent 3D representation, using either classical methodologies [25] or learned-based representations [72]. In contrast, more recent studies [16, 65, 79] have proposed end-to-end strategies that bypass the estimation of local depth maps as an intermediate representation. Instead, they directly regress a Truncated Signed Distance Function (TSDF) [65, 79] or an occupancy volume [16]. A rapidly growing body of work on neural rendering [62, 92] offers not only geometrically-accurate scene recon + +![](images/ea50763dced457e9ea678ff4d3a3ba539dccfd54e79a4d0efab815d8df49d087.jpg) +Figure 1. From a single video of gated captures (top-row), we reconstruct an accurate scene representation and render depth projections (mid-row, right) as accurate as LiDAR scans (mid-row, left), and we recover 3D geometry and normals (bottom-row). + +struction from posed RGB images but also to generate novel perspectives from unobserved angles. Hinging on implicit coordinate-based neural representations, RGB-based methods [7, 8, 55, 100] have been adapted to large open outdoor environments. A recent line of work [37, 71, 86, 93] includes LiDAR scans for auxiliary depth supervision and to improve scene reconstruction for urban environments. However, recovery based on the RGB images exhibit is fundamentally limited in the presence of low light [63, 91] or in the presence of scattering such as fog [49, 69]. + +A parallel direction of research investigates neural rendering techniques tailored to Time-of-Flight (ToF) sensors as opposed to the conventional RGB cameras. Existing methods [41, 84] tackle scene understanding with posed LiDAR scans, and have modeled the raw output from a single-photon LiDAR system [59] or continuous-wave ToF sensor [4] as additional depth supervision. However, all of these existing methods struggle with recovering large unbounded outdoor scenes: while continuous-wave ToF sensors [4] offer signal only in room-sized scenes, methods based on scanning LiDAR suffer from low angular sampling which mandates temporal aggregation. Specifically, even today's LiDAR sensors boasting 200 scan lines, lag drastically in resolution compared to current HDR cameras that offer two orders of magnitude higher vertical pixel counts nearing $10\mathrm{k}$ and three orders of magnitude higher total resolution. + +Addressing these challenges, our work explores scene reconstruction using active gated imaging. Gated imag- + +ing functions by integrating the transient response from a scene that has been flash-illuminated by a synchronized light source [19]. This imaging technique is robust to adverse weather conditions – temporal gating allows us to filter out backscatter – and provides signal in poorly-lit scenes [13]. Existing work has exploited this sensing modality to achieve state-of-the-art depth estimation [89, 90] as well as object detection [44]. In our approach, we train a neural field-based representation of the scene, concurrently learning its geometry, illumination, and material properties. This is accomplished by integrating the gated imaging formation model with a neural rendering framework, that jointly learns the associated gating parameters along with the scene reconstruction. By leveraging the implicit depth cues present in gated video captures, we are able to reconstruct a detailed 3D geometric model of the scene, as shown in Fig. 1. Compared to LiDAR-based approaches, our method offers distinct advantages, as LiDAR systems are inherently constrained by their resolution, necessitating additional time-multiplexed scene captures to aggregate points. This results in an extended acquisition process for LiDAR-based methods or, conversely, compromises the geometric detail of the final estimate. Specifically, for a fixed time acquisition budget the scene reconstruction from a LiDAR sensor is less supervised, although offering highly accurate depth information, the sensor yields data at a volume one order of magnitude less than that of a camera stereo pair. This disparity in data quantity means that while the LiDAR provides precise depth points, it is unable to provide dense and fine detailed predictions. + +To assess our method, we captured a dataset of varied scenes at day and night conditions, using a vehicle test setup comprising of LiDAR, RGB and Gated sensors. We compare our method with feed-forward and 3D reconstruction methods, and assess its superiority in novel depth synthesis beating the next best method by $21.87\%$ MAE, 3D reconstruction improving on the baseline by $11\%$ IoU, and performs novel view synthesis with a PSNR of $32.28~\mathrm{dB}$ . + +In our work, we make the following contributions: + +- We propose a novel neural rendering method and scene representation that is capable of reconstructing scene geometry and radiance from active gated camera videos. +- By modeling the gated image formation process and integrating into differentiable volume rendering, our approach is able to reconstruct and decompose both passive and active light transport components conditioned on the scene parameters in a physically accurate way. +- We validate our method on large outdoor scenes, captured across different scenarios in both day and night, achieving a reduction of MAE error in depth precision of $59.8\%$ to the next best RGB+LiDAR method and $31.7\%$ to the next best methods using gated images. + +# 2. Related Work + +Monocular and Stereo Depth Estimation Depth estimation tasks, from a single image [32, 36, 51, 54], from stereo image pairs [5, 22, 52, 95], or single/stereo images with a LiDAR scan [24, 40, 67, 82, 83, 94, 102] have been at the center of a large body of work. Single CMOS sensor-based depth estimation from RGB color images is inherently limited by scale ambiguity. Additional measurements from LiDAR [9] or ego-vehicle speed [36] can resolve this ambiguity at the cost of an additional sensor. Similarly, stereo methods rely on an additional camera sensor to resolve ambiguity by triangulating between two camera views [22]. Training approaches for learned depth estimation methods using intensity images cover both unsupervised methods [30, 32, 33, 36, 103], which harnesses multi-view geometry consistency, and supervised techniques [22, 27, 42, 45, 51, 54, 58, 60] relying primarily on multi-view datasets [45, 51, 60] or time-of-flight captures [22, 27, 42, 58]. In particular, LiDAR measurements have been proven as a ground-truth signal for depth supervision [2, 22, 27, 42, 54, 58]. Several methods [31, 87] have mitigated the sparsity and range limitations of scanning LiDAR by accumulating scans. However, adverse weather [14] can make LiDAR ground truth unreliable, and methods that rely on consistency between camera and LiDAR [24, 40, 67, 82, 83, 94, 102] suffer from degradations, including scan pattern artifacts and temporal distortions. + +Time-gated Depth Sensing Time-of-Flight (ToF) sensors determine depth by emitting light into a scene and calculating the distance based on the round trip time of the light. Successful sensing methods can be categorized into three classes: correlation ToF cameras [38, 46, 48], pulsed ToF sensors [77], and gated imaging [34, 39]. Correlation ToF sensors[38, 46, 48] estimate depth by continuously flood-illuminating the scene and assessing the phase shift between the emitted and received light. While these sensors can provide high-resolution depth information, their use is primarily confined to indoor settings due to their susceptibility to external light interference. Pulsed ToF sensors [77] function by emitting light pulses toward specific scene points and measure the total travel time to estimate depth. Emitting collimated light, this approach is robust against ambient illumination and allows for outdoor depth measurements. However, its spatial resolution is fundamentally limited owing to its scanning illumination technique, and its efficacy can be compromised in adverse weather due to backscatter [11, 21, 43]. In contrast, gated cameras [12, 34, 39] capture light from a scene over brief intervals, essentially constraining the observable depth to specific range segments. The inherent gating mechanism of these cameras offers resistance to backscattering, and they allow for the recovery of detailed depth maps when using a large number of short + +gates [3, 17, 18] Subsequent works have improved gated depth estimation with few gates by adopting Bayesian approaches [1, 75] or deep neural networks [35, 89, 90], and they achieve accurate gated depth estimation for dynamic outdoor scenes, even under challenging conditions. Recently, Gated Stereo [90] reached state-of-the-art results using a stereo-gated setup and self-supervision [89]. + +Neural Scene Reconstruction Recent research has amalgamated sets of single-sensor measurements to recover comprehensive scene representations. This synthesis has led to advancements in both novel-view generation [6, 23, 62, 64] and depth estimation [85], with neural radiance field methods [6, 23, 62, 64] emerging as a pivotal approach for representing scenes as continuous volumetric fields of radiance. These methods combine this representation with volumetric rendering as a forward model in a test-time optimization approach. Specific representations that these methods explore include coordinate-based networks [6, 7, 62, 101], 3D voxel-grid representation [23, 28, 98], or hybrid approaches [8, 64, 81]. Subsequent works have extended this representation to large outdoor scenes [7, 100], and increased the efficiency at training and test time [8, 23, 64, 98]. Other works departed from radiance-based representation and explicitly learned the scene illumination, geometry, and material properties [15, 74, 101]. A particular challenge within the field is reconstructing large urban terrains based on imagery captured from vehicles [37, 47, 56, 66, 71, 80, 86, 93, 97], given that a significant portion of the scene is seen from only a narrow range of viewpoints. This problem is tackled by additional supervision cues from sparse LiDAR [37, 66, 71, 86], pre-estimated depths [26, 37, 73], optical flow [61, 86] and semantic segmentation [47, 86, 93]. Departing from RGB-based approaches, recent works have investigated neural reconstruction methods using ToF sensors [41, 59, 84]. Existing methods [41, 84, 99] learn a neural field from posed LiDAR scans, allowing for the synthesis of realistic LiDAR scans from novel views. Recently, Malik et al. [59] represent the time-resolved photon count acquired by a single-photon LiDAR system with a neural reconstruction method. + +# 3. Gated Imaging Model + +In this section, we provide a brief background on gated imaging as presented in [35] and introduce the proposed gated image formation model. Unlike prior work, this model incorporates shadow effects and features self-calibrated parameter learning. + +A gated imaging system, illustrated in Fig. 2, utilizes a pulsed flood-light illumination source $p$ with a synchronized imager that operates with a nanosecond (ns) gated exposure $g$ that is delayed by $\xi$ compared to the pulse. This allows us to capture only photons with round-trip times inside the gates, hence specific distance segments in the scene. + +![](images/c84e875ac23349b146648aeea382ea7ea8a76e3c33dcaebf4a3f7b5271090e64.jpg) +Figure 2. Gated Image Formation and Bi-Directional Sampling. Top-row: Our test vehicle is equipped with a synchronized stereo camera setup and illuminator that flood-lits the scene with a light pulse and FoV $\gamma$ . Using different gating profiles $C(z)$ , we capture three slices with intensity visualised here in red, green and blue. Illustrated in the middle row, the gating profiles describe pixel intensity for a point at sensor distance $z$ . The first slice (red) accounts for close ranges, the green for mid-ranges, and the blue for far ranges. Bottom-row: we show the ray sampling employed in our method, based on a bidirectional sampling strategy. We cast the rays from the illuminator view to explore the occluded areas, while the rays casted from the camera integrate the reflected scene response. The shadowed areas are marked in gray. + +We formalize this using so-called range intensity profiles $C_k(2z_c)$ given distance $z_{c}$ from the camera, time $t$ , and a parameter set $k$ [35], that is + +$$ +C _ {k} \left(2 z _ {c}\right) = \int_ {- \infty} ^ {\infty} g _ {k} (t - \xi) p _ {k} \left(t - \frac {2 z _ {c}}{c}\right) \beta \left(2 z _ {c}\right) d t, \tag {1} +$$ + +where $c$ is the speed of light and $\beta (\cdot)$ models the distance-dependent decay of the reflected light pulse. The resulting gated pixel value is + +$$ +I _ {k} \left(z _ {c}\right) = \alpha \iota C _ {k} \left(2 z _ {c}\right) + \Lambda + \mathcal {D} _ {k}, \tag {2} +$$ + +where $\Lambda$ represents the passive ambient contribution, $\alpha$ is the scene reflection, $\iota$ the laser illumination, and $\mathcal{D}_k$ is an additive noise term. + +This model assumes camera position $\mathbf{o}_c$ and the illuminator position $\mathbf{o}_i$ are collocated. To allow for noncollocated positions, we express the travel time as $z = z_c + z_i$ , where $z_i$ denotes the distance between the illuminator and the point on the surface impacted by the light beam, represented as $z_i = |\mathbf{x} - \mathbf{o}_i|_2$ . Additionally, there may be areas visible to the camera that remain dark due to potential occlusions. Modeling shadow effects and attenuation due to incident angle $\omega$ results in the following image formation + +$$ +I _ {k} (z) = \alpha \iota \psi | \mathbf {n} \cdot \omega | C _ {k} (z) + \Lambda + \mathcal {D} _ {k}. \tag {3} +$$ + +Here, $\psi \in [0,1]$ serves as a shadow indicator for the pixel, and $\omega$ is the direction of the incident light at that point. + +We extend this model to fit the range intensity profiles during optimization, thereby eliminating the need for their direct measurement. This approach overcomes potential calibration inaccuracies encountered in previous approaches [89, 90]. We model both the laser pulse $p_k$ and the gate $g_k$ as rectangular functions with durations $t_{l,k}$ and $t_{g,k}$ , respectively. This simplification permits the analytical computation of the integral in Eq. (1), that is + +$$ +\tilde {C} _ {k} = \left\{ \begin{array}{l l} \frac {2 z}{c} - \xi_ {k} + t _ {l, k} & \text {i f} \xi_ {k} - t _ {l, k} < \frac {2 z}{c} < \xi_ {k} \\ t _ {l, k} & \text {i f} \xi_ {k} < \frac {2 z}{c} < \xi_ {k} + t _ {g, k} - t _ {l, k} \\ - \frac {2 z}{c} + \xi_ {k} + t _ {g, k} & \text {i f} \xi_ {k} + l _ {g, k} - t _ {l, k} < \frac {2 z}{c} < \xi_ {k} + t _ {g, k} \\ 0 & \text {o t h e r w i s e} \end{array} \right. \tag {4} +$$ + +# 4. Gated Field + +We reconstruct a scene by fitting a neural field representation to gated videos. We collect videos of three active gated slices $I_{k \in \{1,2,3\}}$ with different gating parameters and one passive slice $I_P$ . We model active illumination by jointly estimating light and material properties, and separately represent the ambient light as a radiance field. The proposed reconstruction method relies on both photometric reconstruction cues and scene priors and is described in the following. + +# 4.1. Neural Gated Fields + +We describe the scene properties using two neural fields $f_{Gp}$ and $f_{G\alpha}$ , respectively representing the ambient light scattered in the scene and the reflection of the scene surfaces, conditioned on a spatial embedding $\chi$ . Moreover, the laser illumination contribution is represented by a physics-based model, while shadow effects are simulated through ray-tracing using the volumetric density field $f_{Gd}$ . + +Neural Ambient and Reflection Field We represent a scene as a neural field $f_{G}:\{\mathbf{x},\mathbf{d},\omega \} \to \{\sigma ,\alpha ,\Lambda ,\mathbf{n}\}$ mapping each point in space $\mathbf{x}$ viewed from a direction $\mathbf{d}$ and laser direction $\omega$ to its volumetric density $\sigma$ , normal vector $\mathbf{n}$ , scene reflection $\alpha$ and the passive component $\Lambda$ , that is + +$$ +f _ {G p}: \{\mathbf {d}, \chi \} \longrightarrow \{\Lambda \} +$$ + +Ambient Component + +$$ +f _ {G n}: \{\mathbf {x}, \chi \} \longrightarrow \{\mathbf {n} \} +$$ + +Surface Normal + +$$ +f _ {G \alpha}: \{\mathbf {d}, \omega , \chi \} \rightarrow \{\alpha \} +$$ + +Surface Reflection + +with $f_{Gd}:\{\mathbf{x}\} \longrightarrow \{\sigma ,\chi \}$ Vol. Density and Embedding + +We condition here normal, ambient and reflectance on a volumetric embedding via the field $f_{Gd}:\{\mathbf{x}\} \to \{\sigma ,\chi \}$ estimating the density $\sigma$ and embedding $\chi$ . This embedding is being shared by the network branches to estimate the normal with $f_{Gn}:\{\mathbf{x},\chi \} \to \{\mathbf{n}\}$ , scene reflection $f_{G\alpha}:\{\mathbf{d},\omega ,\chi \} \to \{\alpha \}$ , and ambient light component with $f_{Gp}:\{\mathbf{d},\chi \} \to \{\Lambda \}$ . An overview of the overall Neural + +Gated Fields is shown in Fig. 3. We also use a proposal sampler $f_{P}$ as in [7] for efficiency. Both $f_{G}$ and $f_{P}$ are MLPs (of different size) with multi-resolution hash encoding [64]. + +Shadow and Illumination Model As the light pulse emitted by the illuminator is a diverging light beam, we model it as cone of light with irradiance maximum at the cross-section center and exponentially decreasing as it diverges from the center by the angles $\gamma$ . As such, we express the illumination intensity as a 2D higher-order Gaussian $\mathcal{G}$ with mean $\Xi$ , standard deviation $\Omega$ and power $\Theta$ + +$$ +\iota = \eta \mathcal {G} _ {i} (\gamma ; \boldsymbol {\Xi}, \boldsymbol {\Omega}, \boldsymbol {\Theta}), \tag {5} +$$ + +where $\eta$ is a scaling parameter. + +Instead of predicting the shadow indicator $\psi (\mathbf{x})$ , we can directly estimate it using the density field, by computing the accumulated transmittance along the ray from the pixel to the point $\mathbf{r}_{ill}(l) = \mathbf{o}_i + \omega l$ + +$$ +\psi (\mathbf {x}) = \exp \left(- \int_ {0} ^ {l _ {\mathbf {x}}} \sigma \left(\mathbf {r} _ {i l l} (l)\right) \mathrm {d} l\right) \tag {6} +$$ + +The illuminator origin and direction $\mathbf{o}_i, \omega$ are obtained from the camera as $[\mathbf{o}_i, \omega] = \mathbf{R}[\mathbf{o}_c, \mathbf{d}_c] + \mathbf{T}$ . During training, we jointly fine-tune the translation and rotation matrices $\mathbf{T}, \mathbf{R}$ , as well as $\mathbf{o}_c$ . We also treat illuminator profile properties as learnable parameters, namely $\eta, \Xi, \Omega, \Theta$ and the gating parameters, i.e. number of accumulated laser pulses $m_k$ before read-out, laser pulse duration $t_{l,k}$ , camera exposure $t_{g,k}$ , and delay $\xi_k$ between laser pulse emission and gated exposure for all three slices $k \in \{0,1,2\}$ . In addition, we optimize a general distance offset $d_0$ for the range intensity profiles to compensate for internal signal processing delays. + +# 4.2. Gated Field Learning + +We learn to acquire a gated capture with camera origin $\mathbf{o}_c$ and direction $\mathbf{d}$ by casting a ray $\mathbf{r}(l) = \mathbf{o}_c + l\mathbf{d}$ for each pixel into the scene and computing the intensity $\tilde{I}_k(\mathbf{r})$ through volume rendering. Using the gated imaging formation model from Sec. 3, we define volume rendering as + +$$ +\begin{array}{l} \tilde{I}_{k}(\mathbf{r}) = \int_{0}^{\infty}T(l)\sigma (\mathbf{x})\int_{-\infty}^{\infty}g_{k}(t - \xi_{k})\beta (l) \\ \cdot \left(\alpha \iota \psi | \mathbf {n} \cdot \omega | p _ {k} \left(t - \frac {l + l _ {i}}{c}\right) + \kappa\right) d t d l + \mathcal {D} _ {k}, \\ \end{array} +$$ + +where $\kappa$ is the ambient level, $T(l) = \exp (-\int_0^l\sigma (u)du)$ is the accumulated transmittance along the ray and $l_{i}$ is the distance of $\mathbf{x}(l)$ from the illuminator origin $\mathbf{o}_i$ . As illustrated in Fig. 2, in our volume rendering formulation the pixel intensity contribution of a point along the ray depends not only on its accumulated transmittance and volumetric + +![](images/579a4ac62f052897a054d43a79cababa260500aac920db1757c83436ee8c2a32.jpg) +Figure 3. Neural Gated Fields. For any point in space $\mathbf{x}$ , we learn its volumetric density $\sigma$ , normal $\mathbf{n}$ , reflectance $\alpha$ and ambient lighting $\Lambda$ through four neural fields, conditioned on direction $\mathbf{d}$ , incident laser light direction $\omega$ and spatial embedding $\chi$ . The illuminator light $\iota$ is represented by a physics-based model dependent on the displacement angle $\gamma$ , while the gating imaging process is described by the range intensity profiles $\tilde{C}(z)$ using as input the camera-point-laser distance $z$ , as explained in Sec. 3. With this information, we reconstruct a gated image $I_k$ through the gated volume rendering formulation introduced in Sec. 4.2. As this process is fully differentiable, we simultaneously fit neural fields and physical parameters through image reconstruction together with other regularization losses discussed in Sec. 4.3. + +density, but also on its distance from the illuminator and camera origins through $C(z)$ , as well as on its relative position to the illuminator source via $\iota$ and $\psi$ . + +We simplify the time-dependent integral using Eq. (4) as + +$$ +\begin{array}{l} \tilde {I} _ {k} (\mathbf {r}) = \int_ {0} ^ {\infty} T (l) \sigma (\mathbf {x}) \left(\alpha (\mathbf {x}, \mathbf {d}, \omega) \tilde {C} _ {k} \left(l + l _ {i}\right) \psi (\mathbf {x}) \right. \tag {8} \\ \cdot | \mathbf {n} \cdot \omega | \iota (\gamma) + \Lambda (\mathbf {x}, \mathbf {d})) \mathrm {d} l + \mathcal {D} _ {k}. \\ \end{array} +$$ + +We numerically estimate this spatial integral by numerical quadrature [62, 88], approximating it with a set of points $\mathbf{X}_{ray}$ . Specifically, for each point $\mathbf{x}_j \in \mathbf{X}_{ray}$ , we query the neural field $f_{G}$ to infer the normal vector $\mathbf{n}_j$ , scene reflection $\alpha_j$ , volumetric density $\sigma_j$ and ambient light component $\Lambda_j$ . The laser illumination intensity $\iota_j$ is instead computed following the physics-based model defined in Eq. (5). The gated intensity $\tilde{I}_k$ is then expressed as + +$$ +\begin{array}{l} \tilde {I} _ {k} (\mathbf {r}) = \sum_ {j = 0} ^ {N} w _ {j} \left(\underbrace {\alpha_ {j} \tilde {C} _ {j} \psi_ {j} | \mathbf {n} _ {j} \cdot \omega_ {j} | \iota_ {j}} _ {\text {A c t i v e C o m p o n e n t}} + \underbrace {\Lambda_ {j}} _ {\text {P a s s i v e C o m p o n e n t}}\right) + \mathcal {D} _ {k}, (9) \\ w _ {j} = \exp \left(- \sum_ {k = 1} ^ {j - 1} \sigma_ {k} \delta_ {k}\right) \left(1 - \exp \left(- \sigma_ {j} \delta_ {j}\right)\right). (9) \\ \end{array} +$$ + +The shadow indicator $\psi_{j}$ from Eq. (6) is similarly approximated by sampling on $\mathbf{r}_{ill} = \mathbf{o}_i + \omega_j l$ a set of points $\mathbf{X}_{ill}$ bounded between $\mathbf{o}_i$ and $\mathbf{x}_j$ + +$$ +\psi_ {j} = \exp \left(- \sum_ {k} ^ {\mathbf {X} _ {i l l}} \sigma_ {k} \delta_ {k}\right). \tag {11} +$$ + +For the passive slice, the active component is null, further simplifying to + +$$ +\tilde {I} _ {P} (\mathbf {r}) = \sum_ {j = 0} ^ {N} w _ {j} \Lambda_ {j} + \mathcal {D} _ {P}. \tag {12} +$$ + +Both $\mathbf{X}_{ray}$ and $\mathbf{X}_{ill}$ are sampled using a proposal network [7] $f_{P}$ that, analogously to $f_{Gd}$ , predicts point-wise densities converted with Eq. (10) to proposal weights $\hat{w}$ for sampling with piece-wise-constant probabilities. + +# 4.3. Training Supervision + +We supervise the predicted passive and active gated frames applying a photometric loss $\mathcal{L}_c$ , regularize the volumetric density with a depth loss $\mathcal{L}_d$ and by supervising the shadow estimate with $\mathcal{L}_s$ . We regularize normal and reflectance estimates through $\mathcal{L}_{nc}$ and $\mathcal{L}_{\alpha}$ , respectively. + +Photometric Loss We supervise with ground truth captures for active and passive gated slice reconstruction as + +$$ +\mathcal {L} _ {c} = \sum_ {k, r} \| \tilde {I} _ {k} (\mathbf {r}) - I _ {k} (\mathbf {r}) \| _ {2} + \sum_ {r} \| \tilde {I} _ {P} (\mathbf {r}) - I _ {P} (\mathbf {r}) \| _ {2} \tag {13} +$$ + +Volume Density Regularization As additional training supervision, we use the depth estimate $\hat{D}(\mathbf{r})$ of a pretrained stereo depth estimation algorithm [90] as pseudo ground-truth to regularize the ray termination distribution [26] + +$$ +\mathcal {L} _ {d} = \sum_ {\mathbf {r}} \sum_ {j} \log w _ {j} \exp \left(- \frac {\left(l _ {j} - \hat {D} (\mathbf {r})\right) ^ {2}}{2 s ^ {2}}\right) \delta_ {i} \tag {14} +$$ + +We regularize the density field by partially supervising the shadow indicator $\psi$ . Each pixel whose active intensity $I_{kA} = I_k - I_P$ in any of the three gated slices is above a certain threshold $\epsilon_i$ is considered as visible from the illuminator. We hence supervise the expected shadow value for such rays $\mathbf{r}_v \in \{\mathbf{r} | \forall k \in \{1, 2, 3\} : I_{kA}(\mathbf{r}) > \epsilon_i\}$ as + +$$ +\mathcal {L} _ {s} = \sum_ {\mathbf {r} _ {v}} \| 1 - \int T (l) \sigma (\mathbf {x}) \psi (\mathbf {x}) d l \| _ {2} \tag {15} +$$ + +Normals Consistency Following [88], for each sampled point $\mathbf{x}$ we enforce a consistency between the predicted normal $\mathbf{n}$ and the density gradient $\hat{\mathbf{n}} (\mathbf{x}) = -\frac{\nabla\mathbf{x}}{||\nabla\mathbf{x}||}$ , and we penalize normals which are back-facing the camera as + +$$ +\mathcal {L} _ {n c} = \sum_ {\mathbf {x}} w (\mathbf {x}) \left(\| \mathbf {n} (\mathbf {x}) - \hat {\mathbf {n}} (\mathbf {x}) \| _ {2} + \max (0, \hat {\mathbf {n}} (\mathbf {x}) \cdot \mathbf {d}) ^ {2}\right) \tag {16} +$$ + +Reflectance Regularization We enforce the predicted scene reflection $\alpha$ to be spatially consistent within $\epsilon_{\mathbf{x}}$ , i.e. + +$$ +\mathcal {L} _ {\alpha} = \sum_ {\mathbf {x}} w (\mathbf {x}) \left(\| \alpha (\mathbf {x}, \mathbf {d}) - \alpha \left(\mathbf {x} + \epsilon_ {\mathbf {x}}, \mathbf {d} + \epsilon_ {\mathbf {d}}\right) \| _ {2}\right). \tag {17} +$$ + +Here the $\omega$ dependency of $\alpha$ is omitted here for brevity. We also include an angular noise $\epsilon_{\mathrm{d}}$ that we set high at the beginning of the training and then decrease it exponentially. By forcing the reflectance to behave as fully diffuse at the beginning of the training, we disincentive it to bake in the effects from lighting or shadowing, hence improving the disjoint learning of the scene components. + +Total Training Loss Combining the different losses we obtain the following loss formulation: + +$$ +\mathcal {L} = \lambda_ {1} \mathcal {L} _ {c} + \lambda_ {2} \mathcal {L} _ {d} + \lambda_ {3} \mathcal {L} _ {s} + \lambda_ {4} \mathcal {L} _ {n c} + \lambda_ {5} \mathcal {L} _ {\alpha}, \tag {18} +$$ + +see $\lambda_{1,\dots,5}$ hyperparameters in the Supplementary Material. + +# 5. Implementation Details + +We train for 35,000 steps and a batch size of 4096 rays. As optimizer we use ADAMW [57] with $\beta_{1} = 0.9$ , $\beta_{2} = 0.999$ , learning rate $10^{-2}$ for $f_{P}$ and $f_{G}$ , $10^{-4}$ for the camera poses optimization, $10^{-4}$ for the laser profile and gated parameters. We train on two NVIDIA V100 GPUs, for approximately 3 hours. The proposal network $f_{P}$ is comprised of two MLPs and trained following [7]. Additional architecture details, training procedures, and hyper-parameters are found in the Supplementary Material. + +# 6. Dataset + +To conduct this work, we have collected a diverse set of 10 static sequences, recorded in both day and night conditions across North America. To this goal, we equipped a test vehicle with a NIR gated stereo camera setup (Bright-Way Vision), an automotive RGB stereo camera (OnSemi AR0230), a LiDAR sensor (Velodyne VLS128) and a GNSS with IMU (Xsens MTi-7), as shown in Fig. 4. Each gated camera has a resolution of $1280 \times 720$ pixels, 10 bit depth and runs at $120 \mathrm{~Hz}$ , split up to collect the three active and one passive slice. The illuminator source consists of two vertical-cavity surface-emitting laser (VCSEL) modules, which illuminate the scene with a laser pulse with duration of 240-370 ns and a wavelength of $808 \mathrm{~nm}$ . The RGB cameras provide 12 bit HDR images with resolution of $1920 \times 1080$ pixels and $30 \mathrm{~Hz}$ frame-rate. The LiDAR has a vertical resolution of 128 lines and $10 \mathrm{~Hz}$ framerate, while the GNSS sensor runs at $4 \mathrm{~Hz}$ . Example captures from the dataset are being visualized in Fig. 4. In total, we collect 2650 samples, captured in both day (1223 samples) and night (1427 samples). We divide it in training, validation and test splits with a 50-25-25 split. + +As ground-truth, we construct a large-scale ground-truth pointcloud by aggregating LiDAR scans with LIO-SAM [78] and removing noisy points. Additional details on the accumulation are provided in the Supplementary Material. + +# 7. Assessment + +In this section, we validate the proposed method quantitatively and qualitatively. Specifically, we investigate scene + +![](images/3aa04838ebe91debfd13421c7886c9598b95b0ec86127b16edc92eeada4ab79a.jpg) + +![](images/5643129c4f14fc3124cc751a35972b2d7735a1efa25bf35b89dcc5d3f45a9d71.jpg) +Figure 4. Top: Example captures from our collected dataset across different urban and suburban areas in North America. From left to right: RGB image, active gated slices (with red for slice 1, green for slice 2 and blue for slice 3), passive slice, projected LiDAR scan, accumulated LiDAR. Bottom: Sensors setup with LiDAR, stereo Gated camera, stereo RGB camera, IMU and GNSS. + +reconstruction at both day and night, using novel depth and view synthesis for 2D evaluation, and surface reconstruction for the 3D evaluation. To this end, we compare our approach to state-of-the-art feed-forward depth estimation algorithms and neural scene reconstruction methods. We also conduct ablation experiments to validate our design choices. + +Depth Reconstruction We assess the quality of depth synthesis of Gated Fields for camera poses unseen during training. We use as ground truth the accumulated and filtered LiDAR pointcloud. Unlike previous works relying on single LiDAR scans for evaluation [35, 89, 90], we use as ground truth an accumulated LiDAR pointcloud as described in Sec. 6. This allows us to evaluate the depth reconstruction up to $160\mathrm{m}$ accurately and without bias. We follow previous works [89, 90] and use as depth evaluation metrics RMSE, MAE, ARD, $\sigma_{i} < 1.25^{i}, i \in \{1,2,3\}$ . We compare our method against 9 feed-forward depth estimation methods, namely SimIPU [53], AdaBins [10], DPT [70], DepthFormer [54] and CREStereo [50] for monocular and stereo RGB methods, Gated2Gated [89] and GatedStereo [90] for monocular and stereo gated estimation methods. We also compare our method against depths rendered with other neural reconstruction algorithms, using RGB images [7],[63],[29],[26] LiDAR [84], RGB+LiDAR [86], [37], and a varying-appearance method [29] for gated + +![](images/ab0701162834ba87697bca1bb1490b5b0aebca8011d2d528ffea7c1506e19800.jpg) +Figure 5. Qualitative comparison of the proposed Gated Fields and state-of-the-art depth estimation approaches, including LiDAR-NeRF [84], StreetSurf [37], NLSPN [67], and Gated Stereo [90]. Compared to baseline methods, we are able to reconstruct fine geometry details like branches or poles, also for far distances. Unlike RGB methods, Gated Fields is unaffected by poor ambient lighting, and unlike LiDAR-based methods it is able to reconstruct sharp object discontinuities. The active gated slices are visualized in red for slice 1, green for slice 2 and blue for slice 3. + +captures. Results for day and night sequences are presented in Tab. 1. We outperform the next best neural field method by $21.87\%$ MAE and $30.35\%$ in RMSE. For night sequences the performance difference sharpens, with Gated Fields outperforming the best RGB-based method [50] by $3.14\mathrm{m}$ MAE. This performance decline is to be attributed to the limited pixel information present in RGB captures taken at night time, making impossible to learn a meaningful 3D representation of the scene, as shown qualitatively in Fig. 5. LiDAR-based methods are unaffected by the change in illumination, but suffer from the limited sensor resolution. On the other hand, gated cameras retrieve information-rich captures both day and night, which Gated Fields can explicitly leverage during training. We confirm that employing state-of-the-art neural field methods on gated captures does not yield accurate results, as they are not able to model the gated imaging formation and can only fit the ambient light component. + +3D Reconstruction We evaluate the 3D scene reconstruction capabilities of Gated Fields using the accumulated Li-DAR pointcloud as ground truth. We follow [20] and extract + +for both the 3D ground truth pointcloud and different neural field-based methods a voxelized occupancy grid of the scene, and compute intersection over union (IoU), Precision and Recall between ground truth voxels and estimated ones. Quantitative results are shown in Tab. 3. Our method outperforms RGB baselines [7, 29] by an average of $15\%$ IoU. RGB baselines using additional LiDAR sensor [37, 86] data partially improve the results, but such methods are still unable to reconstruct finer surfaces details and struggle at night. On the other hand, Gated Fields is able to recover finer geometries using gated, illumination and depth cues, and the quality does not degrade with diminishing ambient light, as shown in Fig. 5. See details on evaluation and further qualitative results in the Supplementary Material. + +Novel View Synthesis For novel view synthesis, we compare our method with Mip-NeRF360 [7], a state-of-the-art neural radiance field-based method, and K-Planes [29], to implicitly model the time-varying appearance of the static scene. Mip-NeRF [7] struggles to reconstruct novel views due to the inherent difficulty of modeling the gating imaging effects, resulting in a PSNR of 17.16dB. By learning + +
MethodModalityRMSE [m]ARDMAE [m]δ1 [%]δ2 [%]δ3 [%]
Test Data - Night (Evaluated on Accumulated LiDAR Ground-Truth Points)
2D DEPTH COMPARISONGATED2GATED [89]Gated13.330.288.5761.7890.3094.90
GATEDSTEREO [90]Stereo-Gated10.100.205.9782.8693.3596.52
SIMIPU [53]RGB19.330.4414.2140.6777.9989.80
ADABINS [10]RGB21.140.3814.2851.7280.0090.64
DPT [70]RGB14.170.289.9062.8288.3294.42
DEPTHFORMER [54]RGB14.320.2910.0861.1987.7294.57
CRESTEREO [50]Stereo-RGB14.220.137.2782.2190.1094.32
NLSPN [67]RGB+LiDAR11.120.186.4877.1590.8196.00
MIPNERF360 [7]RGB23.800.5116.4341.6061.7576.16
K-PLANES [29]RGB19.700.4113.6644.3966.2081.46
RAWNERF [63]RGB27.460.6419.7534.3254.3369.41
DEPTH-NERF [26]RGB15.230.2610.0461.6786.2393.87
SUDS [86]RGB+LiDAR11.070.176.1779.4988.4195.31
STREETSURF [37]RGB+LiDAR10.860.155.9082.1693.5796.96
LiDAR-NERF [50]LiDAR10.210.124.7287.7195.0597.78
GATED FIELDS [50]Gated7.920.124.1390.6195.7697.90
Test Data - Day (Evaluated on Accumulated LiDAR Ground-Truth Points)
2D DEPTH COMPARISONGATED2GATED [89]Gated9.260.226.6958.4693.7097.38
GATEDSTEREO [90]Stereo-Gated6.320.093.3092.8697.3198.53
SIMIPU [53]RGB13.540.3110.0852.9086.4995.65
ADABINS [10]RGB12.740.258.3969.8489.1395.52
DPT [70]RGB10.340.217.0877.6194.2997.24
DEPTHFORMER [54]RGB9.060.196.0981.1994.1497.42
CRESTEREO [50]Stereo-RGB7.350.093.4594.4897.2398.50
NLSPN [67]RGB+LiDAR10.340.175.9777.8391.1696.11
MIPNERF360 [7]RGB16.910.319.5370.8984.4390.92
K-PLANES [29]RGB12.370.248.5563.1679.6891.70
RAWNERF [63]RGB15.100.239.3865.9084.9092.43
DEPTH-NERF [26]RGB10.340.176.0778.9790.7596.53
SUDS [86]RGB+LiDAR9.110.175.8480.9695.0898.33
STREETSURF [37]RGB+LiDAR9.600.135.3583.9495.3298.56
LiDAR-NERF [50]LiDAR8.130.103.8688.9995.4998.06
GATED FIELDS [50]Gated6.150.092.9193.8897.3298.75
+ +Table 1. Comparison of our proposed approach and state-of-the-art approaches on depth synthesis. Best results in each category are in bold and second best are underlined. + +
Im. Formation ModelDepth Sup.Normal Active Illum.RMSE [m]MAE [m]PSNR [dB]SSIM
Test Data - Night
ABLATIONEnd2End [62]XXXX27.3419.9517.430.633
GatedXXXX9.348.3523.470.889
Gated✓XXX6.725.4325.840.908
Gated✓✓✓X3.782.8230.660.946
Gated✓✓✓✓3.392.5130.910.95
Test Data - Day
ABLATIONEnd2End [62]XXXX30.0319.2216.770.66
GatedXXXX11.646.3426.770.915
Gated✓XXX9.204.2226.880.922
Gated✓✓✓X9.454.3132.150.946
Gated✓✓✓✓8.884.1232.280.948
+ +a time-varying appearance, K-Planes improves the quality reaching 27.42dB PSNR for day but only 19.35dB for night, as the model fails to learn an accurate scene geometry representation without ambient light information. Gated Fields outperforms these baselines in both day and night, reaching a PSNR of 32.28dB. + +Ablation Experiments To assess the role and contribution of the different components of our method, we conduct + +Table 2. Ablation studies of the Gated Fields contributions, on a subset of the test dataset. We investigate different image formation models, neural fields components and supervision losses. + +
MethodModalityIoU [%]Precision [%]Recall [%]
3D REC.MIPNERF360 [7]RGB6.327.3431.21
STREETSURF [37]RGB+LiDAR5.416.3127.35
SUDS [86]RGB+LiDAR8.969.8849.09
LiDAR-NERF [84]LiDAR20.0332.3834.44
Gated Fields (ours)Gated22.2525.0166.51
+ +Table 3. Comparison of Gated Fields and state-of-the-art scene reconstruction methods. We evaluate over 3D occupancy reconstruction, using as ground truth the voxelized accumulated LiDAR pointcloud. Best results in each category are in bold and second best are underlined. + +an ablation study in Tab. 2. In particular, we consider as starting point a single neural field directly inferring one intensity value for each of the four slices (3 active + 1 passive), and obtain an average MAE of $19.58\mathrm{m}$ . By separately predicting ambient light and reflectance, and reconstructing the gated image as in Eq. (2), we significantly improve the MAE to $7.34\mathrm{m}$ . However, this approach still performs poorly on flat-color areas during the day and in unilluminated areas during the night due to lack of any depth cue. By adding the depth supervision, we are able to supervise also such areas and the PSNR improves by $4.83\mathrm{dB}$ . By adding the angular-dependent attenuation and regularizing the reflectance in Eq. (17), the model is able to disentangle the material properties from other spurious effects. Finally, by explicitly modeling the shadow, casted by the illuminator, we improve final depth reconstruction to $3.32\mathrm{m}$ MAE. + +# 8. Conclusion + +We introduce Gated Fields, a neural rendering method capable of reconstructing scene geometry from video captures of active time-gated cameras. The method hinges on a differentiable gated image formation as part of the rendering formulation, and it jointly learns geometry, ambient light and surface proprieties, represented implicitly as neural field components, alongside illumination and gating parameters, represented with physics-based models. Extensive experiments on real-world large-scale scenes validate that our method is able to precisely reconstruct a 3D scene both in day and night-time conditions. Our approach outperforms existing RGB and LiDAR methods by $21.87\%$ on MAE, as well as baseline methods using gated captures by $31.67\%$ . In the future, we hope to extend the proposed approach by "closing the loop" and providing dynamic feedback to the gated acquisition, allowing for adaptive gated scene reconstructions. + +Acknowledgments This work was supported by the AI-SEE project with funding from the FFG, BMBF, and NRC-IRA. Felix Heide was supported by an NSF CAREER Award (2047359), a Packard Foundation Fellowship, a Sloan Research Fellowship, a Sony Young Faculty Award, a Project X Innovation Award, and an Amazon Science Research Award. + +# References + +[1] Amit Adam, Christoph Dann, Omer Yair, Shai Mazor, and Sebastian Nowozin. Bayesian time-of-flight for realtime shape, illumination and albedo. 39(5):851-864, 2017. 3 +[2] Ashutosh Agarwal and Chetan Arora. Attention attention everywhere: Monocular depth prediction with skip attention. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5861-5870, 2023. 2 +[3] Pierre Andersson. Long-range three-dimensional imaging using range-gated laser radar images. 45(3):034301, 2006. 3 +[4] Benjamin Attal, Eliot Laidlaw, Aaron Gokaslan, Changil Kim, Christian Richardt, James Tompkin, and Matthew O'Toole. Törf: Time-of-flight radiance fields for dynamic scene view synthesis. Advances in neural information processing systems, 34:26289-26301, 2021. 1 +[5] Abhishek Badki, Alejandro Troccoli, Kihwan Kim, Jan Kautz, Pradeep Sen, and Orazio Gallo. Bi3D: Stereo depth estimation via binary classifications. In arXiv preprint arXiv:2005.07274, 2020. 2 +[6] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. ICCV, 2021. 3 +[7] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 1, 3, 4, 5, 6, 7, 8 +[8] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. ICCV, 2023. 1, 3 +[9] Florent Bartoccioni, Éloi Zablocki, Patrick Pérez, Matthieu Cord, and Karteek Alahari. Lidartouch: Monocular metric depth estimation with a few-beam lidar. Computer Vision and Image Understanding, 227:103601, 2023. 2 +[10] Shariq Farooq Bhat, Ibraheem Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4009-4018, 2021. 6, 8 +[11] Mario Bijelic, Tobias Gruber, and Werner Ritter. A benchmark for lidar sensors in fog: Is detection breaking down? In 2018 IEEE Intelligent Vehicles Symposium (IV), pages 760-767, 2018. 2 +[12] Mario Bijelic, Tobias Gruber, and Werner Ritter. Benchmarking image sensors under adverse weather conditions for autonomous driving. 2018. 2 +[13] Mario Bijelic, Tobias Gruber, and Werner Ritter. Benchmarking image sensors under adverse weather conditions for autonomous driving. In 2018 IEEE Intelligent Vehicles Symposium (IV), pages 1773-1779. IEEE, 2018. 2 +[14] Mario Bijelic, Tobias Gruber, Fahim Mannan, Florian Kraus, Werner Ritter, Klaus Dietmayer, and Felix Heide. Seeing through fog without seeing fog: Deep multimodal sensor fusion in unseen adverse weather. In The IEEE + +Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2 +[15] Mark Boss, Andreas Engelhardt, Abhishek Kar, Yuanzhen Li, Deqing Sun, Jonathan Barron, Hendrik Lensch, and Varun Jampani. *Samurai: Shape and material from unconstrained real-world arbitrary image collections*. Advances in Neural Information Processing Systems, 35:26389-26403, 2022. 3 +[16] Aljaz Bozic, Pablo Palafox, Justus Thies, Angela Dai, and Matthias Nießner. Transformerfusion: Monocular rgb scene reconstruction using transformers. Advances in Neural Information Processing Systems, 34:1403-1414, 2021. 1 +[17] Jens Busck. Underwater 3-D optical imaging with a gated viewing laser radar. 2005. 3 +[18] Jens Busck and Henning Heiselberg. Gated viewing and high-accuracy three-dimensional laser radar. 43(24):4705–10, 2004. 3 +[19] Jens Busck and Henning Heiselberg. Gated viewing and high-accuracy three-dimensional laser radar. Applied optics, 43(24):4705-4710, 2004. 2 +[20] Anh-Quan Cao and Raoul de Charette. Scenerf: Self-supervised monocular 3d scene reconstruction with radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9387-9398, 2023. 7 +[21] A. Carballo, J. Lambert, A. Monrroy, D. Wong, P. Narksri, Y. Kitsukawa, E. Takeuchi, S. Kato, and K. Takeda. Libre: The multiple 3d lidar dataset. In IEEE Intelligent Vehicles Symposium (IV), 2020. 2 +[22] Jia-Ren Chang and Yong-Sheng Chen. Pyramid stereo matching network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5410-5418, 2018. 2 +[23] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 3 +[24] Jaesung Choe, Kyungdon Joo, Tooba Imtiaz, and In So Kweon. Volumetric propagation network: Stereo-lidar fusion for long-range depth estimation. IEEE Robotics and Automation Letters, 6(3):4672-4679, 2021. 2 +[25] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 303-312, 1996. 1 +[26] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nef: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12882-12891, 2022. 3, 5, 6, 8 +[27] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. In NeurIPS, pages 2366-2374, 2014. 2 +[28] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of + +the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501-5510, 2022. 3 +[29] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12479–12488, 2023. 6, 7, 8 +[30] Ravi Garg, B.G. Vijay Kumar, Gustavo Carneiro, and Ian Reid. Unsupervised CNN for single view depth estimation: Geometry to the rescue. In ECCV, pages 740-756, 2016. 2 +[31] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, pages 3354-3361, 2012. 2 +[32] Clément Godard, Oisin Mac Aodha, and Gabriel J Bros-tow. Unsupervised monocular depth estimation with left-right consistency. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 270-279, 2017. 2 +[33] Clément Godard, Oisin Mac Aodha, Michael Firman, and Gabriel J Brostow. Digging into self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3828-3838, 2019. 2 +[34] Yoav Grauer. Active gated imaging in driver assistance system. Advanced Optical Technologies, 3(2):151-160, 2014. 2 +[35] Tobias Gruber, Frank Julca-Aguilar, Mario Bijelic, and Felix Heide. Gated2depth: Real-time dense lidar from gated images. In The IEEE International Conference on Computer Vision (ICCV), 2019. 3, 6 +[36] Vitor Guizilini, Rares Ambrus, Sudeep Pillai, Allan Ravens +tos, and Adrien Gaidon. 3d packing for self-supervised +monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern +Recognition, pages 2485-2494, 2020. 2 +[37] Jianfei Guo, Nianchen Deng, Xinyang Li, Yeqi Bai, Botian Shi, Chiyu Wang, Chenjing Ding, Dongliang Wang, and Yikang Li. Streetsurf: Extending multi-view implicit surface reconstruction to street views. arXiv preprint arXiv:2306.04988, 2023. 1, 3, 6, 7, 8 +[38] Miles Hansard, Seungkyu Lee, Ouk Choi, and Radu Patrice Horaud. Time-of-flight cameras: principles, methods and applications. Springer Science & Business Media, 2012. 2 +[39] Paul Heckman and Robert T. Hodgson. Underwater optical range gating. 3(11):445-448, 1967. 2 +[40] Mu Hu, Shuling Wang, Bin Li, Shiyu Ning, Li Fan, and Xiaojin Gong. Towards precise and efficient image guided depth completion. 2021. 2 +[41] Shengyu Huang, Zan Gojcic, Zian Wang, Francis Williams, Yoni Kasten, Sanja Fidler, Konrad Schindler, and Or Litany. Neural lidar fields for novel view synthesis. 2023. 1, 3 +[42] Maximilian Jaritz, Raoul De Charette, Emilie Wirbel, Xavier Perrotton, and Fawzi Nashashibi. Sparse and dense data with cnns: Depth completion and semantic segmentation. pages 52-60, 2018. 2 + +[43] Maria Jokela, Matti Kutila, and Pasi Pyykönen. Testing and validation of automotive point-cloud sensors in adverse weather conditions. Applied Sciences, 9, 2019. 2 +[44] Frank Julca-Aguilar, Jason Taylor, Mario Bijelic, Fahim Mannan, Ethan Tseng, and Felix Heide. Gated3d: Monocular 3d object detection from temporal illumination cues. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2938-2948, 2021. 2 +[45] Alex Kendall, Hayk Martirosyan, Saumitro Dasgupta, Peter Henry, Ryan Kennedy, Abraham Bachrach, and Adam Bry. End-to-end learning of geometry and context for deep stereo regression. In ICCV, 2017. 2 +[46] Andreas Kolb, Erhardt Barth, Reinhard Koch, and Rasmus Larsen. Time-of-flight cameras in computer graphics. In Computer Graphics Forum, pages 141-159. Wiley Online Library, 2010. 2 +[47] Abhijit Kundu, Kyle Genova, Xiaqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12871-12881, 2022. 3 +[48] Robert Lange. 3D time-of-flight distance measurement with custom solid-state image sensors in CMOS/CCD-technology. 2000. 2 +[49] Deborah Levy, Amit Peleg, Naama Pearl, Dan Rosenbaum, Derya Akkaynak, Simon Korman, and Tali Treibitz. Seathru-nerf: Neural radiance fields in scattering media. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 56–65, 2023. 1 +[50] Jiankun Li, Peisen Wang, Pengfei Xiong, Tao Cai, Ziwei Yan, Lei Yang, Jiangyu Liu, Haoqiang Fan, and Shuaicheng Liu. Practical stereo matching via cascaded recurrent network with adaptive correlation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16263-16272, 2022. 6, 7, 8 +[51] Zhengqi Li, Tali Dekel, Forrester Cole, Richard Tucker, Noah Snavely, Ce Liu, and William T. Freeman. Manequinchallenge: Learning the depths of moving people by watching frozen people. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(12):4229-4241, 2021. 2 +[52] Zhaoshuo Li, Xingtong Liu, Nathan Drenkow, Andy Ding, Francis X. Creighton, Russell H. Taylor, and Mathias Unberath. Revisiting stereo depth estimation from a sequence-to-sequence perspective with transformers. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6177-6186, 2021. 2 +[53] Zhenyu Li, Zehui Chen, Ang Li, Liangji Fang, Qinhong Jiang, Xianming Liu, Junjun Jiang, Bolei Zhou, and Hang Zhao. Simipu: Simple 2d image and 3d point cloud unsupervised pre-training for spatial-aware visual representations. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1500-1508, 2022. 6, 8 +[54] Zhenyu Li, Zehui Chen, Xianming Liu, and Junjun Jiang. Depthformer: Depthformer: Exploiting long-range corre + +lation and local information for accurate monocular depth estimation. arXiv preprint arXiv:2203.14211, 2022. 2, 6, 8 +[55] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1 +[56] Jeffrey Yunfan Liu, Yun Chen, Ze Yang, Jingkang Wang, Sivabalan Manivasagam, and Raquel Urtasun. Real-time neural rasterization for large scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8416-8427, 2023. 3 +[57] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. International Conference on Learning Representations, 2018. 6 +[58] Fangchang Ma and Sertac Karaman. Sparse-to-dense: Depth prediction from sparse depth samples and a single image. pages 1-8, 2018. 2 +[59] Anagh Malik, Parsa Mirdehghan, Sotiris Nousias, Kiriakos N Kutulakos, and David B Lindell. Transient neural radiance fields for lidar view synthesis and 3d reconstruction. arXiv preprint arXiv:2307.09555, 2023. 1, 3 +[60] N. Mayer, E. Ilg, P. Häusser, P. Fischer, D. Cremers, A. Dosovitskiy, and T. Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In CVPR, 2016. 2 +[61] Andreas Meuleman, Yu-Lun Liu, Chen Gao, Jia-Bin Huang, Changil Kim, Min H. Kim, and Johannes Kopf. Progressively optimized local radiance fields for robust view synthesis. In CVPR, 2023. 3 +[62] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 3, 5, 8 +[63] Ben Mildenhall, Peter Hedman, Ricardo Martin-Brualla, Pratul P Srinivasan, and Jonathan T Barron. Nerf in the dark: High dynamic range view synthesis from noisy raw images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16190-16199, 2022. 1, 6, 8 +[64] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 3, 4 +[65] Zak Murez, Tarrence Van As, James Bartolozzi, Ayan Sinha, Vijay Badrinarayanan, and Andrew Rabinovich. Atlas: End-to-end 3d scene reconstruction from posed images. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VII 16, pages 414-431. Springer, 2020. 1 +[66] Julian Ost, Issam Laradji, Alejandro Newell, Yuval Bahat, and Felix Heide. Neural point light fields. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3 +[67] Jinsun Park, Kyungdon Joo, Zhe Hu, Chi-Kuei Liu, and In So Kweon. Non-local spatial propagation network for depth completion. In Proc. of European Conference on Computer Vision (ECCV), 2020. 2, 7, 8 + +[68] Vivek Pradeep, Christoph Rhemann, Shahram Izadi, Christopher Zach, Michael Bleyer, and Steven Bathiche. Monofusion: Real-time 3d reconstruction of small scenes with a single web camera. In 2013 IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 83-88. IEEE, 2013. 1 +[69] Andrea Ramazzina, Mario Bijelic, Stefanie Walz, Alessandro Sanvito, Dominik Scheuble, and Felix Heide. Scatternerf: Seeing through fog with physically-based inverse neural rendering. arXiv preprint arXiv:2305.02103, 2023. 1 +[70] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12179-12188, 2021. 6, 8 +[71] Konstantinos Rematas, Andrew Liu, Pratul P. Srinivasan, Jonathan T. Barron, Andrea Tagliasacchi, Tom Funkhouser, and Vittorio Ferrari. Urban radiance fields. CVPR, 2022. 1, 3 +[72] Gernot Riegler, Ali Osman Ulusoy, Horst Bischof, and Andreas Geiger. Octnetfusion: Learning depth fusion from data. In 2017 International Conference on 3D Vision (3DV), pages 57-66. IEEE, 2017. 1 +[73] Barbara Roessle, Jonathan T Barron, Ben Mildenhall, Pratul P Srinivasan, and Matthias Nießner. Dense depth priors for neural radiance fields from sparse input views. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12892-12901, 2022. 3 +[74] Viktor Rudnev, Mohamed Elgharib, William Smith, Lingjie Liu, Vladislav Golyanik, and Christian Theobalt. Nerf for outdoor scene relighting. In European Conference on Computer Vision, pages 615-631. Springer, 2022. 3 +[75] Michael Schober, Amit Adam, Omer Yair, Shai Mazor, and Sebastian Nowozin. Dynamic time-of-flight. In CVPR, pages 6109-6118, 2017. 3 +[76] Thomas Schöps, Torsten Sattler, Christian Hane, and Marc Pollefeys. 3d modeling on the go: Interactive 3d reconstruction of large-scale scenes on mobile devices. In 2015 International Conference on 3D Vision, pages 291-299. IEEE, 2015. 1 +[77] Brent Schwarz. Lidar: Mapping the world in 3D. Nature Photonics, 4(7):429, 2010. 2 +[78] Tixiao Shan, Brendan Englot, Drew Meyers, Wei Wang, Carlo Ratti, and Daniela Rus. Lio-sam: Tightly-coupled lidar inertial odometry via smoothing and mapping. In 2020 IEEE/RSJ international conference on intelligent robots and systems (IROS), pages 5135-5142. IEEE, 2020. 6 +[79] Jiaming Sun, Yiming Xie, Linghao Chen, Xiaowei Zhou, and Hujun Bao. Neuralrecon: Real-time coherent 3d reconstruction from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15598-15607, 2021. 1 +[80] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8248–8258, 2022. 3 + +[81] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, et al. Nerfstudio: A modular framework for neural radiance field development. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-12, 2023. 3 +[82] Jiexiong Tang, John Folkesson, and Patric Jensfelt. Sparse2dense: From direct sparse odometry to dense 3-d reconstruction. IEEE Robotics and Automation Letters, 4 (2):530-537, 2019. 2 +[83] Jie Tang, Fei-Peng Tian, Wei Feng, Jian Li, and Ping Tan. Learning guided convolutional network for depth completion. IEEE Transactions on Image Processing, 30:1116-1129, 2020. 2 +[84] Tang Tao, Longfei Gao, Guangrun Wang, Peng Chen, Dayang Hao, Xiaodan Liang, Mathieu Salzmann, and Kaicheng Yu. Lidar-nerf: Novel lidar view synthesis via neural radiance fields. arXiv preprint arXiv:2304.10406, 2023. 1, 3, 6, 7, 8 +[85] Fabio Tosi, Alessio Tonioni, Daniele De Gregorio, and Matteo Poggi. Nerf-supervised deep stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 855–866, 2023. 3 +[86] Haithem Turki, Jason Y Zhang, Francesco Ferroni, and Deva Ramanan. Suits: Scalable urban dynamic scenes. In Computer Vision and Pattern Recognition (CVPR), 2023. 1, 3, 6, 7, 8 +[87] Jonas Uhrig, Nick Schneider, Lukas Schneider, Uwe Franke, Thomas Brox, and Andreas Geiger. Sparsity invariant cnns. In International Conference on 3D Vision (3DV), 2017. 2 +[88] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481–5490. IEEE, 2022. 5 +[89] Amanpreet Walia, Stefanie Walz, Mario Bijelic, Fahim Mannan, Frank Julca-Aguilar, Michael Langer, Werner Ritter, and Felix Heide. Gated2gated: Self-supervised depth estimation from gated images. 2022. 2, 3, 4, 6, 8 +[90] Stefanie Walz, Mario Bijelic, Andrea Ramazzina, Amanpreet Walia, Fahim Mannan, and Felix Heide. Gated stereo: Joint depth estimation from gated and wide-baseline active stereo cues. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13252-13262, 2023. 2, 3, 4, 5, 6, 7, 8 +[91] Haoyuan Wang, Xiaogang Xu, Ke Xu, and Rynson WH Lau. Lighting up nerf via unsupervised decomposition and enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12632-12641, 2023. 1 +[92] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 1 +[93] Zian Wang, Tianchang Shen, Jun Gao, Shengyu Huang, Jacob Munkberg, Jon Hasselgren, Zan Gojcic, Wenzheng + +Chen, and Sanja Fidler. Neural fields meet explicit geometric representations for inverse rendering of urban scenes. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 3 +[94] Alex Wong and Stefano Soatto. Unsupervised depth completion with calibrated backprojection layers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12747-12756, 2021. 2 +[95] Gengshan Yang, Joshua Manela, Michael Happold, and Deva Ramanan. Hierarchical deep stereo matching on high-resolution images. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2 +[96] Zhenfei Yang, Fei Gao, and Shaojie Shen. Real-time monocular dense mapping on aerial robots using visual-inertial fusion. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 4552-4559. IEEE, 2017. 1 +[97] Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasun. Unisim: A neural closed-loop sensor simulator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1389-1399, 2023. 3 +[98] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoptrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5752-5761, 2021. 3 +[99] Junge Zhang, Feihu Zhang, Shaochen Kuang, and Li Zhang. Nerf-lidar: Generating realistic lidar point clouds with neural radiance fields. arXiv preprint arXiv:2304.14811, 2023. 3 +[100] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 1, 3 +[101] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Ner-factor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (ToG), 40(6):1-18, 2021. 3 +[102] Yongjian Zhang, Longguang Wang, Kunhong Li, Zhiheng Fu, and Yulan Guo. Slfnet: A stereo and lidar fusion network for depth completion. IEEE Robotics and Automation Letters, 7(4):10605-10612, 2022. 2 +[103] Tinghui Zhou, Matthew Brown, Noah Snavely, and David G. Lowe. Unsupervised learning of depth and egomotion from video. In CVPR, 2017. 2 \ No newline at end of file diff --git a/2024/Gated Fields_ Learning Scene Reconstruction from Gated Videos/images.zip b/2024/Gated Fields_ Learning Scene Reconstruction from Gated Videos/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..69518d0491e739e197641819e8dd5efb4b53b54a --- /dev/null +++ b/2024/Gated Fields_ Learning Scene Reconstruction from Gated Videos/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:767b647e42692b668b845e3d0b160ddabdb9c559f2de67692d207bb70c0a9ceb +size 758285 diff --git a/2024/Gated Fields_ Learning Scene Reconstruction from Gated Videos/layout.json b/2024/Gated Fields_ Learning Scene Reconstruction from Gated Videos/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..037e5b036de61c26eb5213fe202e5721adfdc11d --- /dev/null +++ b/2024/Gated Fields_ Learning Scene Reconstruction from Gated Videos/layout.json @@ -0,0 +1,11937 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 96, + 73, + 500, + 90 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 73, + 500, + 90 + ], + "spans": [ + { + "bbox": [ + 96, + 73, + 500, + 90 + ], + "type": "text", + "content": "Gated Fields: Learning Scene Reconstruction from Gated Videos" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 78, + 113, + 530, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 113, + 530, + 129 + ], + "spans": [ + { + "bbox": [ + 78, + 113, + 530, + 129 + ], + "type": "text", + "content": "Andrea Ramazzina" + }, + { + "bbox": [ + 78, + 113, + 530, + 129 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 78, + 113, + 530, + 129 + ], + "type": "text", + "content": " Stefanie Walz" + }, + { + "bbox": [ + 78, + 113, + 530, + 129 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 78, + 113, + 530, + 129 + ], + "type": "text", + "content": " Pragyan Dahal" + }, + { + "bbox": [ + 78, + 113, + 530, + 129 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 78, + 113, + 530, + 129 + ], + "type": "text", + "content": " Mario Bijelic" + }, + { + "bbox": [ + 78, + 113, + 530, + 129 + ], + "type": "inline_equation", + "content": "^{4,5}" + }, + { + "bbox": [ + 78, + 113, + 530, + 129 + ], + "type": "text", + "content": " Felix Heide" + }, + { + "bbox": [ + 78, + 113, + 530, + 129 + ], + "type": "inline_equation", + "content": "^{4,5}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 103, + 130, + 502, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 130, + 502, + 144 + ], + "spans": [ + { + "bbox": [ + 103, + 130, + 502, + 144 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 103, + 130, + 502, + 144 + ], + "type": "text", + "content": "Mercedes-Benz " + }, + { + "bbox": [ + 103, + 130, + 502, + 144 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 103, + 130, + 502, + 144 + ], + "type": "text", + "content": "Saarland University " + }, + { + "bbox": [ + 103, + 130, + 502, + 144 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 103, + 130, + 502, + 144 + ], + "type": "text", + "content": "Politecnico di Milano " + }, + { + "bbox": [ + 103, + 130, + 502, + 144 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 103, + 130, + 502, + 144 + ], + "type": "text", + "content": "Torch Robotics " + }, + { + "bbox": [ + 103, + 130, + 502, + 144 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 103, + 130, + 502, + 144 + ], + "type": "text", + "content": "Princeton University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 171, + 192, + 184 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 171, + 192, + 184 + ], + "spans": [ + { + "bbox": [ + 143, + 171, + 192, + 184 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 198, + 290, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 198, + 290, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 290, + 437 + ], + "type": "text", + "content": "Reconstructing outdoor 3D scenes from temporal observations is a challenge that recent work on neural fields has offered a new avenue for. However, existing methods that recover scene properties, such as geometry, appearance, or radiance, solely from RGB captures often fail when handling poorly-lit or texture-deficient regions. Similarly, recovering scenes with scanning LiDAR sensors is also difficult due to their low angular sampling rate which makes recovering expansive real-world scenes difficult. Tackling these gaps, we introduce Gated Fields - a neural scene reconstruction method that utilizes active gated video sequences. To this end, we propose a neural rendering approach that seamlessly incorporates time-gated capture and illumination. Our method exploits the intrinsic depth cues in the gated videos, achieving precise and dense geometry reconstruction irrespective of ambient illumination conditions. We validate the method across day and night scenarios and find that Gated Fields compares favorably to RGB and LiDAR reconstruction methods. Our code and datasets are available here1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 460, + 128, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 460, + 128, + 472 + ], + "spans": [ + { + "bbox": [ + 47, + 460, + 128, + 472 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 480, + 288, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 480, + 288, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 480, + 288, + 685 + ], + "type": "text", + "content": "Large-scale outdoor scene reconstruction is essential for advancing autonomous robotics, drones, and driver-assistance systems, serving as the foundation for scene understanding, safe navigation, dataset generation and validation. Existing works in this domain [68, 76, 96] have typically adopted a two-step approach. Initially, they infer depth maps from different poses, utilizing time-of-flight sensors or RGB captures. Subsequently, these depth estimates are fused to produce a coherent 3D representation, using either classical methodologies [25] or learned-based representations [72]. In contrast, more recent studies [16, 65, 79] have proposed end-to-end strategies that bypass the estimation of local depth maps as an intermediate representation. Instead, they directly regress a Truncated Signed Distance Function (TSDF) [65, 79] or an occupancy volume [16]. A rapidly growing body of work on neural rendering [62, 92] offers not only geometrically-accurate scene recon" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 170, + 547, + 320 + ], + "blocks": [ + { + "bbox": [ + 307, + 170, + 547, + 320 + ], + "lines": [ + { + "bbox": [ + 307, + 170, + 547, + 320 + ], + "spans": [ + { + "bbox": [ + 307, + 170, + 547, + 320 + ], + "type": "image", + "image_path": "ea50763dced457e9ea678ff4d3a3ba539dccfd54e79a4d0efab815d8df49d087.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 322, + 547, + 367 + ], + "lines": [ + { + "bbox": [ + 305, + 322, + 547, + 367 + ], + "spans": [ + { + "bbox": [ + 305, + 322, + 547, + 367 + ], + "type": "text", + "content": "Figure 1. From a single video of gated captures (top-row), we reconstruct an accurate scene representation and render depth projections (mid-row, right) as accurate as LiDAR scans (mid-row, left), and we recover 3D geometry and normals (bottom-row)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 376, + 547, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 547, + 496 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 547, + 496 + ], + "type": "text", + "content": "struction from posed RGB images but also to generate novel perspectives from unobserved angles. Hinging on implicit coordinate-based neural representations, RGB-based methods [7, 8, 55, 100] have been adapted to large open outdoor environments. A recent line of work [37, 71, 86, 93] includes LiDAR scans for auxiliary depth supervision and to improve scene reconstruction for urban environments. However, recovery based on the RGB images exhibit is fundamentally limited in the presence of low light [63, 91] or in the presence of scattering such as fog [49, 69]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 497, + 548, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 548, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 548, + 688 + ], + "type": "text", + "content": "A parallel direction of research investigates neural rendering techniques tailored to Time-of-Flight (ToF) sensors as opposed to the conventional RGB cameras. Existing methods [41, 84] tackle scene understanding with posed LiDAR scans, and have modeled the raw output from a single-photon LiDAR system [59] or continuous-wave ToF sensor [4] as additional depth supervision. However, all of these existing methods struggle with recovering large unbounded outdoor scenes: while continuous-wave ToF sensors [4] offer signal only in room-sized scenes, methods based on scanning LiDAR suffer from low angular sampling which mandates temporal aggregation. Specifically, even today's LiDAR sensors boasting 200 scan lines, lag drastically in resolution compared to current HDR cameras that offer two orders of magnitude higher vertical pixel counts nearing " + }, + { + "bbox": [ + 304, + 497, + 548, + 688 + ], + "type": "inline_equation", + "content": "10\\mathrm{k}" + }, + { + "bbox": [ + 304, + 497, + 548, + 688 + ], + "type": "text", + "content": " and three orders of magnitude higher total resolution." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "content": "Addressing these challenges, our work explores scene reconstruction using active gated imaging. Gated imag-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 693, + 215, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 693, + 215, + 703 + ], + "spans": [ + { + "bbox": [ + 57, + 693, + 215, + 703 + ], + "type": "text", + "content": "*These authors contributed equally to this work." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 703, + 232, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 703, + 232, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 703, + 232, + 712 + ], + "type": "text", + "content": "1https://light.princeton.edu/gatedfields/" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10530" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 430 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 430 + ], + "type": "text", + "content": "ing functions by integrating the transient response from a scene that has been flash-illuminated by a synchronized light source [19]. This imaging technique is robust to adverse weather conditions – temporal gating allows us to filter out backscatter – and provides signal in poorly-lit scenes [13]. Existing work has exploited this sensing modality to achieve state-of-the-art depth estimation [89, 90] as well as object detection [44]. In our approach, we train a neural field-based representation of the scene, concurrently learning its geometry, illumination, and material properties. This is accomplished by integrating the gated imaging formation model with a neural rendering framework, that jointly learns the associated gating parameters along with the scene reconstruction. By leveraging the implicit depth cues present in gated video captures, we are able to reconstruct a detailed 3D geometric model of the scene, as shown in Fig. 1. Compared to LiDAR-based approaches, our method offers distinct advantages, as LiDAR systems are inherently constrained by their resolution, necessitating additional time-multiplexed scene captures to aggregate points. This results in an extended acquisition process for LiDAR-based methods or, conversely, compromises the geometric detail of the final estimate. Specifically, for a fixed time acquisition budget the scene reconstruction from a LiDAR sensor is less supervised, although offering highly accurate depth information, the sensor yields data at a volume one order of magnitude less than that of a camera stereo pair. This disparity in data quantity means that while the LiDAR provides precise depth points, it is unable to provide dense and fine detailed predictions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 434, + 289, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 434, + 289, + 531 + ], + "spans": [ + { + "bbox": [ + 47, + 434, + 289, + 531 + ], + "type": "text", + "content": "To assess our method, we captured a dataset of varied scenes at day and night conditions, using a vehicle test setup comprising of LiDAR, RGB and Gated sensors. We compare our method with feed-forward and 3D reconstruction methods, and assess its superiority in novel depth synthesis beating the next best method by " + }, + { + "bbox": [ + 47, + 434, + 289, + 531 + ], + "type": "inline_equation", + "content": "21.87\\%" + }, + { + "bbox": [ + 47, + 434, + 289, + 531 + ], + "type": "text", + "content": " MAE, 3D reconstruction improving on the baseline by " + }, + { + "bbox": [ + 47, + 434, + 289, + 531 + ], + "type": "inline_equation", + "content": "11\\%" + }, + { + "bbox": [ + 47, + 434, + 289, + 531 + ], + "type": "text", + "content": " IoU, and performs novel view synthesis with a PSNR of " + }, + { + "bbox": [ + 47, + 434, + 289, + 531 + ], + "type": "inline_equation", + "content": "32.28~\\mathrm{dB}" + }, + { + "bbox": [ + 47, + 434, + 289, + 531 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 538, + 253, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 538, + 253, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 538, + 253, + 550 + ], + "type": "text", + "content": "In our work, we make the following contributions:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 558, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 558, + 286, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 558, + 286, + 593 + ], + "spans": [ + { + "bbox": [ + 47, + 558, + 286, + 593 + ], + "type": "text", + "content": "- We propose a novel neural rendering method and scene representation that is capable of reconstructing scene geometry and radiance from active gated camera videos." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 594, + 286, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 594, + 286, + 653 + ], + "spans": [ + { + "bbox": [ + 47, + 594, + 286, + 653 + ], + "type": "text", + "content": "- By modeling the gated image formation process and integrating into differentiable volume rendering, our approach is able to reconstruct and decompose both passive and active light transport components conditioned on the scene parameters in a physically accurate way." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 654, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 286, + 713 + ], + "type": "text", + "content": "- We validate our method on large outdoor scenes, captured across different scenarios in both day and night, achieving a reduction of MAE error in depth precision of " + }, + { + "bbox": [ + 47, + 654, + 286, + 713 + ], + "type": "inline_equation", + "content": "59.8\\%" + }, + { + "bbox": [ + 47, + 654, + 286, + 713 + ], + "type": "text", + "content": " to the next best RGB+LiDAR method and " + }, + { + "bbox": [ + 47, + 654, + 286, + 713 + ], + "type": "inline_equation", + "content": "31.7\\%" + }, + { + "bbox": [ + 47, + 654, + 286, + 713 + ], + "type": "text", + "content": " to the next best methods using gated images." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 97, + 547, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 97, + 547, + 408 + ], + "spans": [ + { + "bbox": [ + 304, + 97, + 547, + 408 + ], + "type": "text", + "content": "Monocular and Stereo Depth Estimation Depth estimation tasks, from a single image [32, 36, 51, 54], from stereo image pairs [5, 22, 52, 95], or single/stereo images with a LiDAR scan [24, 40, 67, 82, 83, 94, 102] have been at the center of a large body of work. Single CMOS sensor-based depth estimation from RGB color images is inherently limited by scale ambiguity. Additional measurements from LiDAR [9] or ego-vehicle speed [36] can resolve this ambiguity at the cost of an additional sensor. Similarly, stereo methods rely on an additional camera sensor to resolve ambiguity by triangulating between two camera views [22]. Training approaches for learned depth estimation methods using intensity images cover both unsupervised methods [30, 32, 33, 36, 103], which harnesses multi-view geometry consistency, and supervised techniques [22, 27, 42, 45, 51, 54, 58, 60] relying primarily on multi-view datasets [45, 51, 60] or time-of-flight captures [22, 27, 42, 58]. In particular, LiDAR measurements have been proven as a ground-truth signal for depth supervision [2, 22, 27, 42, 54, 58]. Several methods [31, 87] have mitigated the sparsity and range limitations of scanning LiDAR by accumulating scans. However, adverse weather [14] can make LiDAR ground truth unreliable, and methods that rely on consistency between camera and LiDAR [24, 40, 67, 82, 83, 94, 102] suffer from degradations, including scan pattern artifacts and temporal distortions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "text", + "content": "Time-gated Depth Sensing Time-of-Flight (ToF) sensors determine depth by emitting light into a scene and calculating the distance based on the round trip time of the light. Successful sensing methods can be categorized into three classes: correlation ToF cameras [38, 46, 48], pulsed ToF sensors [77], and gated imaging [34, 39]. Correlation ToF sensors[38, 46, 48] estimate depth by continuously flood-illuminating the scene and assessing the phase shift between the emitted and received light. While these sensors can provide high-resolution depth information, their use is primarily confined to indoor settings due to their susceptibility to external light interference. Pulsed ToF sensors [77] function by emitting light pulses toward specific scene points and measure the total travel time to estimate depth. Emitting collimated light, this approach is robust against ambient illumination and allows for outdoor depth measurements. However, its spatial resolution is fundamentally limited owing to its scanning illumination technique, and its efficacy can be compromised in adverse weather due to backscatter [11, 21, 43]. In contrast, gated cameras [12, 34, 39] capture light from a scene over brief intervals, essentially constraining the observable depth to specific range segments. The inherent gating mechanism of these cameras offers resistance to backscattering, and they allow for the recovery of detailed depth maps when using a large number of short" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10531" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": "gates [3, 17, 18] Subsequent works have improved gated depth estimation with few gates by adopting Bayesian approaches [1, 75] or deep neural networks [35, 89, 90], and they achieve accurate gated depth estimation for dynamic outdoor scenes, even under challenging conditions. Recently, Gated Stereo [90] reached state-of-the-art results using a stereo-gated setup and self-supervision [89]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 158, + 288, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 158, + 288, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 158, + 288, + 552 + ], + "type": "text", + "content": "Neural Scene Reconstruction Recent research has amalgamated sets of single-sensor measurements to recover comprehensive scene representations. This synthesis has led to advancements in both novel-view generation [6, 23, 62, 64] and depth estimation [85], with neural radiance field methods [6, 23, 62, 64] emerging as a pivotal approach for representing scenes as continuous volumetric fields of radiance. These methods combine this representation with volumetric rendering as a forward model in a test-time optimization approach. Specific representations that these methods explore include coordinate-based networks [6, 7, 62, 101], 3D voxel-grid representation [23, 28, 98], or hybrid approaches [8, 64, 81]. Subsequent works have extended this representation to large outdoor scenes [7, 100], and increased the efficiency at training and test time [8, 23, 64, 98]. Other works departed from radiance-based representation and explicitly learned the scene illumination, geometry, and material properties [15, 74, 101]. A particular challenge within the field is reconstructing large urban terrains based on imagery captured from vehicles [37, 47, 56, 66, 71, 80, 86, 93, 97], given that a significant portion of the scene is seen from only a narrow range of viewpoints. This problem is tackled by additional supervision cues from sparse LiDAR [37, 66, 71, 86], pre-estimated depths [26, 37, 73], optical flow [61, 86] and semantic segmentation [47, 86, 93]. Departing from RGB-based approaches, recent works have investigated neural reconstruction methods using ToF sensors [41, 59, 84]. Existing methods [41, 84, 99] learn a neural field from posed LiDAR scans, allowing for the synthesis of realistic LiDAR scans from novel views. Recently, Malik et al. [59] represent the time-resolved photon count acquired by a single-photon LiDAR system with a neural reconstruction method." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 562, + 175, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 562, + 175, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 562, + 175, + 574 + ], + "type": "text", + "content": "3. Gated Imaging Model" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 582, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 641 + ], + "type": "text", + "content": "In this section, we provide a brief background on gated imaging as presented in [35] and introduce the proposed gated image formation model. Unlike prior work, this model incorporates shadow effects and features self-calibrated parameter learning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": "A gated imaging system, illustrated in Fig. 2, utilizes a pulsed flood-light illumination source " + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": " with a synchronized imager that operates with a nanosecond (ns) gated exposure " + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": " that is delayed by " + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": " compared to the pulse. This allows us to capture only photons with round-trip times inside the gates, hence specific distance segments in the scene." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 313, + 66, + 539, + 194 + ], + "blocks": [ + { + "bbox": [ + 313, + 66, + 539, + 194 + ], + "lines": [ + { + "bbox": [ + 313, + 66, + 539, + 194 + ], + "spans": [ + { + "bbox": [ + 313, + 66, + 539, + 194 + ], + "type": "image", + "image_path": "c84e875ac23349b146648aeea382ea7ea8a76e3c33dcaebf4a3f7b5271090e64.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 201, + 545, + 344 + ], + "lines": [ + { + "bbox": [ + 304, + 201, + 545, + 344 + ], + "spans": [ + { + "bbox": [ + 304, + 201, + 545, + 344 + ], + "type": "text", + "content": "Figure 2. Gated Image Formation and Bi-Directional Sampling. Top-row: Our test vehicle is equipped with a synchronized stereo camera setup and illuminator that flood-lits the scene with a light pulse and FoV " + }, + { + "bbox": [ + 304, + 201, + 545, + 344 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 201, + 545, + 344 + ], + "type": "text", + "content": ". Using different gating profiles " + }, + { + "bbox": [ + 304, + 201, + 545, + 344 + ], + "type": "inline_equation", + "content": "C(z)" + }, + { + "bbox": [ + 304, + 201, + 545, + 344 + ], + "type": "text", + "content": ", we capture three slices with intensity visualised here in red, green and blue. Illustrated in the middle row, the gating profiles describe pixel intensity for a point at sensor distance " + }, + { + "bbox": [ + 304, + 201, + 545, + 344 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 201, + 545, + 344 + ], + "type": "text", + "content": ". The first slice (red) accounts for close ranges, the green for mid-ranges, and the blue for far ranges. Bottom-row: we show the ray sampling employed in our method, based on a bidirectional sampling strategy. We cast the rays from the illuminator view to explore the occluded areas, while the rays casted from the camera integrate the reflected scene response. The shadowed areas are marked in gray." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 357, + 545, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 357, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 305, + 357, + 545, + 392 + ], + "type": "text", + "content": "We formalize this using so-called range intensity profiles " + }, + { + "bbox": [ + 305, + 357, + 545, + 392 + ], + "type": "inline_equation", + "content": "C_k(2z_c)" + }, + { + "bbox": [ + 305, + 357, + 545, + 392 + ], + "type": "text", + "content": " given distance " + }, + { + "bbox": [ + 305, + 357, + 545, + 392 + ], + "type": "inline_equation", + "content": "z_{c}" + }, + { + "bbox": [ + 305, + 357, + 545, + 392 + ], + "type": "text", + "content": " from the camera, time " + }, + { + "bbox": [ + 305, + 357, + 545, + 392 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 357, + 545, + 392 + ], + "type": "text", + "content": ", and a parameter set " + }, + { + "bbox": [ + 305, + 357, + 545, + 392 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 357, + 545, + 392 + ], + "type": "text", + "content": " [35], that is" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 316, + 402, + 545, + 438 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 402, + 545, + 438 + ], + "spans": [ + { + "bbox": [ + 316, + 402, + 545, + 438 + ], + "type": "interline_equation", + "content": "C _ {k} \\left(2 z _ {c}\\right) = \\int_ {- \\infty} ^ {\\infty} g _ {k} (t - \\xi) p _ {k} \\left(t - \\frac {2 z _ {c}}{c}\\right) \\beta \\left(2 z _ {c}\\right) d t, \\tag {1}", + "image_path": "b49984cdd66ae8680732de5481c58b2bf790a6c99b02ab1ba386ba21fe54123d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 449, + 545, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 449, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 305, + 449, + 545, + 485 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 449, + 545, + 485 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 305, + 449, + 545, + 485 + ], + "type": "text", + "content": " is the speed of light and " + }, + { + "bbox": [ + 305, + 449, + 545, + 485 + ], + "type": "inline_equation", + "content": "\\beta (\\cdot)" + }, + { + "bbox": [ + 305, + 449, + 545, + 485 + ], + "type": "text", + "content": " models the distance-dependent decay of the reflected light pulse. The resulting gated pixel value is" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 359, + 495, + 545, + 507 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 495, + 545, + 507 + ], + "spans": [ + { + "bbox": [ + 359, + 495, + 545, + 507 + ], + "type": "interline_equation", + "content": "I _ {k} \\left(z _ {c}\\right) = \\alpha \\iota C _ {k} \\left(2 z _ {c}\\right) + \\Lambda + \\mathcal {D} _ {k}, \\tag {2}", + "image_path": "c70ab56e96b1992eb58b257edef9303a17e2e9f689621d9c1e737946dc505212.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 517, + 545, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 517, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 517, + 545, + 552 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 517, + 545, + 552 + ], + "type": "inline_equation", + "content": "\\Lambda" + }, + { + "bbox": [ + 304, + 517, + 545, + 552 + ], + "type": "text", + "content": " represents the passive ambient contribution, " + }, + { + "bbox": [ + 304, + 517, + 545, + 552 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 517, + 545, + 552 + ], + "type": "text", + "content": " is the scene reflection, " + }, + { + "bbox": [ + 304, + 517, + 545, + 552 + ], + "type": "inline_equation", + "content": "\\iota" + }, + { + "bbox": [ + 304, + 517, + 545, + 552 + ], + "type": "text", + "content": " the laser illumination, and " + }, + { + "bbox": [ + 304, + 517, + 545, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_k" + }, + { + "bbox": [ + 304, + 517, + 545, + 552 + ], + "type": "text", + "content": " is an additive noise term." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 553, + 545, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 553, + 545, + 661 + ], + "spans": [ + { + "bbox": [ + 304, + 553, + 545, + 661 + ], + "type": "text", + "content": "This model assumes camera position " + }, + { + "bbox": [ + 304, + 553, + 545, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_c" + }, + { + "bbox": [ + 304, + 553, + 545, + 661 + ], + "type": "text", + "content": " and the illuminator position " + }, + { + "bbox": [ + 304, + 553, + 545, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_i" + }, + { + "bbox": [ + 304, + 553, + 545, + 661 + ], + "type": "text", + "content": " are collocated. To allow for noncollocated positions, we express the travel time as " + }, + { + "bbox": [ + 304, + 553, + 545, + 661 + ], + "type": "inline_equation", + "content": "z = z_c + z_i" + }, + { + "bbox": [ + 304, + 553, + 545, + 661 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 553, + 545, + 661 + ], + "type": "inline_equation", + "content": "z_i" + }, + { + "bbox": [ + 304, + 553, + 545, + 661 + ], + "type": "text", + "content": " denotes the distance between the illuminator and the point on the surface impacted by the light beam, represented as " + }, + { + "bbox": [ + 304, + 553, + 545, + 661 + ], + "type": "inline_equation", + "content": "z_i = |\\mathbf{x} - \\mathbf{o}_i|_2" + }, + { + "bbox": [ + 304, + 553, + 545, + 661 + ], + "type": "text", + "content": ". Additionally, there may be areas visible to the camera that remain dark due to potential occlusions. Modeling shadow effects and attenuation due to incident angle " + }, + { + "bbox": [ + 304, + 553, + 545, + 661 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 304, + 553, + 545, + 661 + ], + "type": "text", + "content": " results in the following image formation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 354, + 669, + 545, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 669, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 354, + 669, + 545, + 681 + ], + "type": "interline_equation", + "content": "I _ {k} (z) = \\alpha \\iota \\psi | \\mathbf {n} \\cdot \\omega | C _ {k} (z) + \\Lambda + \\mathcal {D} _ {k}. \\tag {3}", + "image_path": "5af132c44bbda9d7fd1397eebe42c1f4f8cfd385d04d73040c1394bfca19c925.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\psi \\in [0,1]" + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": " serves as a shadow indicator for the pixel, and " + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": " is the direction of the incident light at that point." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10532" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": "We extend this model to fit the range intensity profiles during optimization, thereby eliminating the need for their direct measurement. This approach overcomes potential calibration inaccuracies encountered in previous approaches [89, 90]. We model both the laser pulse " + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "p_k" + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": " and the gate " + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "g_k" + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": " as rectangular functions with durations " + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "t_{l,k}" + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "t_{g,k}" + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": ", respectively. This simplification permits the analytical computation of the integral in Eq. (1), that is" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 176, + 292, + 244 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 176, + 292, + 244 + ], + "spans": [ + { + "bbox": [ + 47, + 176, + 292, + 244 + ], + "type": "interline_equation", + "content": "\\tilde {C} _ {k} = \\left\\{ \\begin{array}{l l} \\frac {2 z}{c} - \\xi_ {k} + t _ {l, k} & \\text {i f} \\xi_ {k} - t _ {l, k} < \\frac {2 z}{c} < \\xi_ {k} \\\\ t _ {l, k} & \\text {i f} \\xi_ {k} < \\frac {2 z}{c} < \\xi_ {k} + t _ {g, k} - t _ {l, k} \\\\ - \\frac {2 z}{c} + \\xi_ {k} + t _ {g, k} & \\text {i f} \\xi_ {k} + l _ {g, k} - t _ {l, k} < \\frac {2 z}{c} < \\xi_ {k} + t _ {g, k} \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {4}", + "image_path": "a416d595239c0580bb611fdd9f14ce27df27e007d46fcd0044632eba3dd8c2bd.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 247, + 123, + 259 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 247, + 123, + 259 + ], + "spans": [ + { + "bbox": [ + 47, + 247, + 123, + 259 + ], + "type": "text", + "content": "4. Gated Field" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 266, + 287, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 287, + 363 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 287, + 363 + ], + "type": "text", + "content": "We reconstruct a scene by fitting a neural field representation to gated videos. We collect videos of three active gated slices " + }, + { + "bbox": [ + 46, + 266, + 287, + 363 + ], + "type": "inline_equation", + "content": "I_{k \\in \\{1,2,3\\}}" + }, + { + "bbox": [ + 46, + 266, + 287, + 363 + ], + "type": "text", + "content": " with different gating parameters and one passive slice " + }, + { + "bbox": [ + 46, + 266, + 287, + 363 + ], + "type": "inline_equation", + "content": "I_P" + }, + { + "bbox": [ + 46, + 266, + 287, + 363 + ], + "type": "text", + "content": ". We model active illumination by jointly estimating light and material properties, and separately represent the ambient light as a radiance field. The proposed reconstruction method relies on both photometric reconstruction cues and scene priors and is described in the following." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 369, + 164, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 369, + 164, + 382 + ], + "spans": [ + { + "bbox": [ + 47, + 369, + 164, + 382 + ], + "type": "text", + "content": "4.1. Neural Gated Fields" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 388, + 287, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 287, + 473 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 287, + 473 + ], + "type": "text", + "content": "We describe the scene properties using two neural fields " + }, + { + "bbox": [ + 46, + 388, + 287, + 473 + ], + "type": "inline_equation", + "content": "f_{Gp}" + }, + { + "bbox": [ + 46, + 388, + 287, + 473 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 388, + 287, + 473 + ], + "type": "inline_equation", + "content": "f_{G\\alpha}" + }, + { + "bbox": [ + 46, + 388, + 287, + 473 + ], + "type": "text", + "content": ", respectively representing the ambient light scattered in the scene and the reflection of the scene surfaces, conditioned on a spatial embedding " + }, + { + "bbox": [ + 46, + 388, + 287, + 473 + ], + "type": "inline_equation", + "content": "\\chi" + }, + { + "bbox": [ + 46, + 388, + 287, + 473 + ], + "type": "text", + "content": ". Moreover, the laser illumination contribution is represented by a physics-based model, while shadow effects are simulated through ray-tracing using the volumetric density field " + }, + { + "bbox": [ + 46, + 388, + 287, + 473 + ], + "type": "inline_equation", + "content": "f_{Gd}" + }, + { + "bbox": [ + 46, + 388, + 287, + 473 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 487, + 287, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 487, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 46, + 487, + 287, + 547 + ], + "type": "text", + "content": "Neural Ambient and Reflection Field We represent a scene as a neural field " + }, + { + "bbox": [ + 46, + 487, + 287, + 547 + ], + "type": "inline_equation", + "content": "f_{G}:\\{\\mathbf{x},\\mathbf{d},\\omega \\} \\to \\{\\sigma ,\\alpha ,\\Lambda ,\\mathbf{n}\\}" + }, + { + "bbox": [ + 46, + 487, + 287, + 547 + ], + "type": "text", + "content": " mapping each point in space " + }, + { + "bbox": [ + 46, + 487, + 287, + 547 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 487, + 287, + 547 + ], + "type": "text", + "content": " viewed from a direction " + }, + { + "bbox": [ + 46, + 487, + 287, + 547 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 46, + 487, + 287, + 547 + ], + "type": "text", + "content": " and laser direction " + }, + { + "bbox": [ + 46, + 487, + 287, + 547 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 46, + 487, + 287, + 547 + ], + "type": "text", + "content": " to its volumetric density " + }, + { + "bbox": [ + 46, + 487, + 287, + 547 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 487, + 287, + 547 + ], + "type": "text", + "content": ", normal vector " + }, + { + "bbox": [ + 46, + 487, + 287, + 547 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 46, + 487, + 287, + 547 + ], + "type": "text", + "content": ", scene reflection " + }, + { + "bbox": [ + 46, + 487, + 287, + 547 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 487, + 287, + 547 + ], + "type": "text", + "content": " and the passive component " + }, + { + "bbox": [ + 46, + 487, + 287, + 547 + ], + "type": "inline_equation", + "content": "\\Lambda" + }, + { + "bbox": [ + 46, + 487, + 287, + 547 + ], + "type": "text", + "content": ", that is" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 564, + 167, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 564, + 167, + 578 + ], + "spans": [ + { + "bbox": [ + 69, + 564, + 167, + 578 + ], + "type": "interline_equation", + "content": "f _ {G p}: \\{\\mathbf {d}, \\chi \\} \\longrightarrow \\{\\Lambda \\}", + "image_path": "8182db3ed73e597955989a66c9de4cb857b6cbf2f5088345839c584328cd887e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 205, + 564, + 292, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 564, + 292, + 577 + ], + "spans": [ + { + "bbox": [ + 205, + 564, + 292, + 577 + ], + "type": "text", + "content": "Ambient Component" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 579, + 165, + 592 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 579, + 165, + 592 + ], + "spans": [ + { + "bbox": [ + 69, + 579, + 165, + 592 + ], + "type": "interline_equation", + "content": "f _ {G n}: \\{\\mathbf {x}, \\chi \\} \\longrightarrow \\{\\mathbf {n} \\}", + "image_path": "d2e99c0124c4bae61e3cf085b26bca5dfc7a8358a4240ee94d5a07aebb148be9.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 225, + 580, + 291, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 580, + 291, + 590 + ], + "spans": [ + { + "bbox": [ + 225, + 580, + 291, + 590 + ], + "type": "text", + "content": "Surface Normal" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 69, + 594, + 166, + 607 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 594, + 166, + 607 + ], + "spans": [ + { + "bbox": [ + 69, + 594, + 166, + 607 + ], + "type": "interline_equation", + "content": "f _ {G \\alpha}: \\{\\mathbf {d}, \\omega , \\chi \\} \\rightarrow \\{\\alpha \\}", + "image_path": "f92a7ffb37a57b38ef9d0c43957fdaf769db5dd7f4a895a1ca1ac98e0a82656c.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 215, + 595, + 291, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 595, + 291, + 605 + ], + "spans": [ + { + "bbox": [ + 215, + 595, + 291, + 605 + ], + "type": "text", + "content": "Surface Reflection" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 609, + 291, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 291, + 622 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 291, + 622 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 47, + 609, + 291, + 622 + ], + "type": "inline_equation", + "content": "f_{Gd}:\\{\\mathbf{x}\\} \\longrightarrow \\{\\sigma ,\\chi \\}" + }, + { + "bbox": [ + 47, + 609, + 291, + 622 + ], + "type": "text", + "content": " Vol. Density and Embedding" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 629, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 629, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 629, + 289, + 714 + ], + "type": "text", + "content": "We condition here normal, ambient and reflectance on a volumetric embedding via the field " + }, + { + "bbox": [ + 47, + 629, + 289, + 714 + ], + "type": "inline_equation", + "content": "f_{Gd}:\\{\\mathbf{x}\\} \\to \\{\\sigma ,\\chi \\}" + }, + { + "bbox": [ + 47, + 629, + 289, + 714 + ], + "type": "text", + "content": " estimating the density " + }, + { + "bbox": [ + 47, + 629, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 47, + 629, + 289, + 714 + ], + "type": "text", + "content": " and embedding " + }, + { + "bbox": [ + 47, + 629, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\chi" + }, + { + "bbox": [ + 47, + 629, + 289, + 714 + ], + "type": "text", + "content": ". This embedding is being shared by the network branches to estimate the normal with " + }, + { + "bbox": [ + 47, + 629, + 289, + 714 + ], + "type": "inline_equation", + "content": "f_{Gn}:\\{\\mathbf{x},\\chi \\} \\to \\{\\mathbf{n}\\}" + }, + { + "bbox": [ + 47, + 629, + 289, + 714 + ], + "type": "text", + "content": ", scene reflection " + }, + { + "bbox": [ + 47, + 629, + 289, + 714 + ], + "type": "inline_equation", + "content": "f_{G\\alpha}:\\{\\mathbf{d},\\omega ,\\chi \\} \\to \\{\\alpha \\}" + }, + { + "bbox": [ + 47, + 629, + 289, + 714 + ], + "type": "text", + "content": ", and ambient light component with " + }, + { + "bbox": [ + 47, + 629, + 289, + 714 + ], + "type": "inline_equation", + "content": "f_{Gp}:\\{\\mathbf{d},\\chi \\} \\to \\{\\Lambda \\}" + }, + { + "bbox": [ + 47, + 629, + 289, + 714 + ], + "type": "text", + "content": ". An overview of the overall Neural" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "text", + "content": "Gated Fields is shown in Fig. 3. We also use a proposal sampler " + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "inline_equation", + "content": "f_{P}" + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "text", + "content": " as in [7] for efficiency. Both " + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "inline_equation", + "content": "f_{G}" + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "inline_equation", + "content": "f_{P}" + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "text", + "content": " are MLPs (of different size) with multi-resolution hash encoding [64]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 123, + 545, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 123, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 304, + 123, + 545, + 207 + ], + "type": "text", + "content": "Shadow and Illumination Model As the light pulse emitted by the illuminator is a diverging light beam, we model it as cone of light with irradiance maximum at the cross-section center and exponentially decreasing as it diverges from the center by the angles " + }, + { + "bbox": [ + 304, + 123, + 545, + 207 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 123, + 545, + 207 + ], + "type": "text", + "content": ". As such, we express the illumination intensity as a 2D higher-order Gaussian " + }, + { + "bbox": [ + 304, + 123, + 545, + 207 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 304, + 123, + 545, + 207 + ], + "type": "text", + "content": " with mean " + }, + { + "bbox": [ + 304, + 123, + 545, + 207 + ], + "type": "inline_equation", + "content": "\\Xi" + }, + { + "bbox": [ + 304, + 123, + 545, + 207 + ], + "type": "text", + "content": ", standard deviation " + }, + { + "bbox": [ + 304, + 123, + 545, + 207 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 304, + 123, + 545, + 207 + ], + "type": "text", + "content": " and power " + }, + { + "bbox": [ + 304, + 123, + 545, + 207 + ], + "type": "inline_equation", + "content": "\\Theta" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 381, + 217, + 545, + 230 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 217, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 381, + 217, + 545, + 230 + ], + "type": "interline_equation", + "content": "\\iota = \\eta \\mathcal {G} _ {i} (\\gamma ; \\boldsymbol {\\Xi}, \\boldsymbol {\\Omega}, \\boldsymbol {\\Theta}), \\tag {5}", + "image_path": "6fba34f74225e44b9afe63a964a2dbf49b6cec03c69fe4d60528c7e3ae7cdf46.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 239, + 432, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 239, + 432, + 251 + ], + "spans": [ + { + "bbox": [ + 305, + 239, + 432, + 251 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 239, + 432, + 251 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 305, + 239, + 432, + 251 + ], + "type": "text", + "content": " is a scaling parameter." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 251, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 251, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 251, + 545, + 300 + ], + "type": "text", + "content": "Instead of predicting the shadow indicator " + }, + { + "bbox": [ + 304, + 251, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\psi (\\mathbf{x})" + }, + { + "bbox": [ + 304, + 251, + 545, + 300 + ], + "type": "text", + "content": ", we can directly estimate it using the density field, by computing the accumulated transmittance along the ray from the pixel to the point " + }, + { + "bbox": [ + 304, + 251, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{ill}(l) = \\mathbf{o}_i + \\omega l" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 353, + 308, + 545, + 340 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 308, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 353, + 308, + 545, + 340 + ], + "type": "interline_equation", + "content": "\\psi (\\mathbf {x}) = \\exp \\left(- \\int_ {0} ^ {l _ {\\mathbf {x}}} \\sigma \\left(\\mathbf {r} _ {i l l} (l)\\right) \\mathrm {d} l\\right) \\tag {6}", + "image_path": "b1b1460d0939b4987d4b83fe0ed2d62e056e120ee9dc49270e4dcc77822912a9.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 348, + 545, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 348, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 304, + 348, + 545, + 481 + ], + "type": "text", + "content": "The illuminator origin and direction " + }, + { + "bbox": [ + 304, + 348, + 545, + 481 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_i, \\omega" + }, + { + "bbox": [ + 304, + 348, + 545, + 481 + ], + "type": "text", + "content": " are obtained from the camera as " + }, + { + "bbox": [ + 304, + 348, + 545, + 481 + ], + "type": "inline_equation", + "content": "[\\mathbf{o}_i, \\omega] = \\mathbf{R}[\\mathbf{o}_c, \\mathbf{d}_c] + \\mathbf{T}" + }, + { + "bbox": [ + 304, + 348, + 545, + 481 + ], + "type": "text", + "content": ". During training, we jointly fine-tune the translation and rotation matrices " + }, + { + "bbox": [ + 304, + 348, + 545, + 481 + ], + "type": "inline_equation", + "content": "\\mathbf{T}, \\mathbf{R}" + }, + { + "bbox": [ + 304, + 348, + 545, + 481 + ], + "type": "text", + "content": ", as well as " + }, + { + "bbox": [ + 304, + 348, + 545, + 481 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_c" + }, + { + "bbox": [ + 304, + 348, + 545, + 481 + ], + "type": "text", + "content": ". We also treat illuminator profile properties as learnable parameters, namely " + }, + { + "bbox": [ + 304, + 348, + 545, + 481 + ], + "type": "inline_equation", + "content": "\\eta, \\Xi, \\Omega, \\Theta" + }, + { + "bbox": [ + 304, + 348, + 545, + 481 + ], + "type": "text", + "content": " and the gating parameters, i.e. number of accumulated laser pulses " + }, + { + "bbox": [ + 304, + 348, + 545, + 481 + ], + "type": "inline_equation", + "content": "m_k" + }, + { + "bbox": [ + 304, + 348, + 545, + 481 + ], + "type": "text", + "content": " before read-out, laser pulse duration " + }, + { + "bbox": [ + 304, + 348, + 545, + 481 + ], + "type": "inline_equation", + "content": "t_{l,k}" + }, + { + "bbox": [ + 304, + 348, + 545, + 481 + ], + "type": "text", + "content": ", camera exposure " + }, + { + "bbox": [ + 304, + 348, + 545, + 481 + ], + "type": "inline_equation", + "content": "t_{g,k}" + }, + { + "bbox": [ + 304, + 348, + 545, + 481 + ], + "type": "text", + "content": ", and delay " + }, + { + "bbox": [ + 304, + 348, + 545, + 481 + ], + "type": "inline_equation", + "content": "\\xi_k" + }, + { + "bbox": [ + 304, + 348, + 545, + 481 + ], + "type": "text", + "content": " between laser pulse emission and gated exposure for all three slices " + }, + { + "bbox": [ + 304, + 348, + 545, + 481 + ], + "type": "inline_equation", + "content": "k \\in \\{0,1,2\\}" + }, + { + "bbox": [ + 304, + 348, + 545, + 481 + ], + "type": "text", + "content": ". In addition, we optimize a general distance offset " + }, + { + "bbox": [ + 304, + 348, + 545, + 481 + ], + "type": "inline_equation", + "content": "d_0" + }, + { + "bbox": [ + 304, + 348, + 545, + 481 + ], + "type": "text", + "content": " for the range intensity profiles to compensate for internal signal processing delays." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 488, + 430, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 488, + 430, + 502 + ], + "spans": [ + { + "bbox": [ + 305, + 488, + 430, + 502 + ], + "type": "text", + "content": "4.2. Gated Field Learning" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 506, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 545, + 567 + ], + "type": "text", + "content": "We learn to acquire a gated capture with camera origin " + }, + { + "bbox": [ + 304, + 506, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_c" + }, + { + "bbox": [ + 304, + 506, + 545, + 567 + ], + "type": "text", + "content": " and direction " + }, + { + "bbox": [ + 304, + 506, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 304, + 506, + 545, + 567 + ], + "type": "text", + "content": " by casting a ray " + }, + { + "bbox": [ + 304, + 506, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{r}(l) = \\mathbf{o}_c + l\\mathbf{d}" + }, + { + "bbox": [ + 304, + 506, + 545, + 567 + ], + "type": "text", + "content": " for each pixel into the scene and computing the intensity " + }, + { + "bbox": [ + 304, + 506, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\tilde{I}_k(\\mathbf{r})" + }, + { + "bbox": [ + 304, + 506, + 545, + 567 + ], + "type": "text", + "content": " through volume rendering. Using the gated imaging formation model from Sec. 3, we define volume rendering as" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 324, + 575, + 518, + 630 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 575, + 518, + 630 + ], + "spans": [ + { + "bbox": [ + 324, + 575, + 518, + 630 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\tilde{I}_{k}(\\mathbf{r}) = \\int_{0}^{\\infty}T(l)\\sigma (\\mathbf{x})\\int_{-\\infty}^{\\infty}g_{k}(t - \\xi_{k})\\beta (l) \\\\ \\cdot \\left(\\alpha \\iota \\psi | \\mathbf {n} \\cdot \\omega | p _ {k} \\left(t - \\frac {l + l _ {i}}{c}\\right) + \\kappa\\right) d t d l + \\mathcal {D} _ {k}, \\\\ \\end{array}", + "image_path": "02a490f8a3cde18897e2772c7e69fe3976194551b8d34e1602569095ef6d4a29.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 640, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 640, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 640, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 640, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 304, + 640, + 545, + 713 + ], + "type": "text", + "content": " is the ambient level, " + }, + { + "bbox": [ + 304, + 640, + 545, + 713 + ], + "type": "inline_equation", + "content": "T(l) = \\exp (-\\int_0^l\\sigma (u)du)" + }, + { + "bbox": [ + 304, + 640, + 545, + 713 + ], + "type": "text", + "content": " is the accumulated transmittance along the ray and " + }, + { + "bbox": [ + 304, + 640, + 545, + 713 + ], + "type": "inline_equation", + "content": "l_{i}" + }, + { + "bbox": [ + 304, + 640, + 545, + 713 + ], + "type": "text", + "content": " is the distance of " + }, + { + "bbox": [ + 304, + 640, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}(l)" + }, + { + "bbox": [ + 304, + 640, + 545, + 713 + ], + "type": "text", + "content": " from the illuminator origin " + }, + { + "bbox": [ + 304, + 640, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_i" + }, + { + "bbox": [ + 304, + 640, + 545, + 713 + ], + "type": "text", + "content": ". As illustrated in Fig. 2, in our volume rendering formulation the pixel intensity contribution of a point along the ray depends not only on its accumulated transmittance and volumetric" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10533" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 60, + 509, + 185 + ], + "blocks": [ + { + "bbox": [ + 85, + 60, + 509, + 185 + ], + "lines": [ + { + "bbox": [ + 85, + 60, + 509, + 185 + ], + "spans": [ + { + "bbox": [ + 85, + 60, + 509, + 185 + ], + "type": "image", + "image_path": "579a4ac62f052897a054d43a79cababa260500aac920db1757c83436ee8c2a32.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 186, + 547, + 254 + ], + "lines": [ + { + "bbox": [ + 46, + 186, + 547, + 254 + ], + "spans": [ + { + "bbox": [ + 46, + 186, + 547, + 254 + ], + "type": "text", + "content": "Figure 3. Neural Gated Fields. For any point in space " + }, + { + "bbox": [ + 46, + 186, + 547, + 254 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 186, + 547, + 254 + ], + "type": "text", + "content": ", we learn its volumetric density " + }, + { + "bbox": [ + 46, + 186, + 547, + 254 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 186, + 547, + 254 + ], + "type": "text", + "content": ", normal " + }, + { + "bbox": [ + 46, + 186, + 547, + 254 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 46, + 186, + 547, + 254 + ], + "type": "text", + "content": ", reflectance " + }, + { + "bbox": [ + 46, + 186, + 547, + 254 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 186, + 547, + 254 + ], + "type": "text", + "content": " and ambient lighting " + }, + { + "bbox": [ + 46, + 186, + 547, + 254 + ], + "type": "inline_equation", + "content": "\\Lambda" + }, + { + "bbox": [ + 46, + 186, + 547, + 254 + ], + "type": "text", + "content": " through four neural fields, conditioned on direction " + }, + { + "bbox": [ + 46, + 186, + 547, + 254 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 46, + 186, + 547, + 254 + ], + "type": "text", + "content": ", incident laser light direction " + }, + { + "bbox": [ + 46, + 186, + 547, + 254 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 46, + 186, + 547, + 254 + ], + "type": "text", + "content": " and spatial embedding " + }, + { + "bbox": [ + 46, + 186, + 547, + 254 + ], + "type": "inline_equation", + "content": "\\chi" + }, + { + "bbox": [ + 46, + 186, + 547, + 254 + ], + "type": "text", + "content": ". The illuminator light " + }, + { + "bbox": [ + 46, + 186, + 547, + 254 + ], + "type": "inline_equation", + "content": "\\iota" + }, + { + "bbox": [ + 46, + 186, + 547, + 254 + ], + "type": "text", + "content": " is represented by a physics-based model dependent on the displacement angle " + }, + { + "bbox": [ + 46, + 186, + 547, + 254 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 186, + 547, + 254 + ], + "type": "text", + "content": ", while the gating imaging process is described by the range intensity profiles " + }, + { + "bbox": [ + 46, + 186, + 547, + 254 + ], + "type": "inline_equation", + "content": "\\tilde{C}(z)" + }, + { + "bbox": [ + 46, + 186, + 547, + 254 + ], + "type": "text", + "content": " using as input the camera-point-laser distance " + }, + { + "bbox": [ + 46, + 186, + 547, + 254 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 186, + 547, + 254 + ], + "type": "text", + "content": ", as explained in Sec. 3. With this information, we reconstruct a gated image " + }, + { + "bbox": [ + 46, + 186, + 547, + 254 + ], + "type": "inline_equation", + "content": "I_k" + }, + { + "bbox": [ + 46, + 186, + 547, + 254 + ], + "type": "text", + "content": " through the gated volume rendering formulation introduced in Sec. 4.2. As this process is fully differentiable, we simultaneously fit neural fields and physical parameters through image reconstruction together with other regularization losses discussed in Sec. 4.3." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 265, + 287, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 265, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 265, + 287, + 300 + ], + "type": "text", + "content": "density, but also on its distance from the illuminator and camera origins through " + }, + { + "bbox": [ + 46, + 265, + 287, + 300 + ], + "type": "inline_equation", + "content": "C(z)" + }, + { + "bbox": [ + 46, + 265, + 287, + 300 + ], + "type": "text", + "content": ", as well as on its relative position to the illuminator source via " + }, + { + "bbox": [ + 46, + 265, + 287, + 300 + ], + "type": "inline_equation", + "content": "\\iota" + }, + { + "bbox": [ + 46, + 265, + 287, + 300 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 265, + 287, + 300 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 46, + 265, + 287, + 300 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 301, + 277, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 301, + 277, + 312 + ], + "spans": [ + { + "bbox": [ + 47, + 301, + 277, + 312 + ], + "type": "text", + "content": "We simplify the time-dependent integral using Eq. (4) as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 312, + 287, + 353 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 312, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 58, + 312, + 287, + 353 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\tilde {I} _ {k} (\\mathbf {r}) = \\int_ {0} ^ {\\infty} T (l) \\sigma (\\mathbf {x}) \\left(\\alpha (\\mathbf {x}, \\mathbf {d}, \\omega) \\tilde {C} _ {k} \\left(l + l _ {i}\\right) \\psi (\\mathbf {x}) \\right. \\tag {8} \\\\ \\cdot | \\mathbf {n} \\cdot \\omega | \\iota (\\gamma) + \\Lambda (\\mathbf {x}, \\mathbf {d})) \\mathrm {d} l + \\mathcal {D} _ {k}. \\\\ \\end{array}", + "image_path": "68eebd716009870234addd0752a248b037d51cc772076e579d48d925e939502b.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 357, + 287, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 357, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 47, + 357, + 287, + 453 + ], + "type": "text", + "content": "We numerically estimate this spatial integral by numerical quadrature [62, 88], approximating it with a set of points " + }, + { + "bbox": [ + 47, + 357, + 287, + 453 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{ray}" + }, + { + "bbox": [ + 47, + 357, + 287, + 453 + ], + "type": "text", + "content": ". Specifically, for each point " + }, + { + "bbox": [ + 47, + 357, + 287, + 453 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_j \\in \\mathbf{X}_{ray}" + }, + { + "bbox": [ + 47, + 357, + 287, + 453 + ], + "type": "text", + "content": ", we query the neural field " + }, + { + "bbox": [ + 47, + 357, + 287, + 453 + ], + "type": "inline_equation", + "content": "f_{G}" + }, + { + "bbox": [ + 47, + 357, + 287, + 453 + ], + "type": "text", + "content": " to infer the normal vector " + }, + { + "bbox": [ + 47, + 357, + 287, + 453 + ], + "type": "inline_equation", + "content": "\\mathbf{n}_j" + }, + { + "bbox": [ + 47, + 357, + 287, + 453 + ], + "type": "text", + "content": ", scene reflection " + }, + { + "bbox": [ + 47, + 357, + 287, + 453 + ], + "type": "inline_equation", + "content": "\\alpha_j" + }, + { + "bbox": [ + 47, + 357, + 287, + 453 + ], + "type": "text", + "content": ", volumetric density " + }, + { + "bbox": [ + 47, + 357, + 287, + 453 + ], + "type": "inline_equation", + "content": "\\sigma_j" + }, + { + "bbox": [ + 47, + 357, + 287, + 453 + ], + "type": "text", + "content": " and ambient light component " + }, + { + "bbox": [ + 47, + 357, + 287, + 453 + ], + "type": "inline_equation", + "content": "\\Lambda_j" + }, + { + "bbox": [ + 47, + 357, + 287, + 453 + ], + "type": "text", + "content": ". The laser illumination intensity " + }, + { + "bbox": [ + 47, + 357, + 287, + 453 + ], + "type": "inline_equation", + "content": "\\iota_j" + }, + { + "bbox": [ + 47, + 357, + 287, + 453 + ], + "type": "text", + "content": " is instead computed following the physics-based model defined in Eq. (5). The gated intensity " + }, + { + "bbox": [ + 47, + 357, + 287, + 453 + ], + "type": "inline_equation", + "content": "\\tilde{I}_k" + }, + { + "bbox": [ + 47, + 357, + 287, + 453 + ], + "type": "text", + "content": " is then expressed as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 452, + 287, + 532 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 452, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 53, + 452, + 287, + 532 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\tilde {I} _ {k} (\\mathbf {r}) = \\sum_ {j = 0} ^ {N} w _ {j} \\left(\\underbrace {\\alpha_ {j} \\tilde {C} _ {j} \\psi_ {j} | \\mathbf {n} _ {j} \\cdot \\omega_ {j} | \\iota_ {j}} _ {\\text {A c t i v e C o m p o n e n t}} + \\underbrace {\\Lambda_ {j}} _ {\\text {P a s s i v e C o m p o n e n t}}\\right) + \\mathcal {D} _ {k}, (9) \\\\ w _ {j} = \\exp \\left(- \\sum_ {k = 1} ^ {j - 1} \\sigma_ {k} \\delta_ {k}\\right) \\left(1 - \\exp \\left(- \\sigma_ {j} \\delta_ {j}\\right)\\right). (9) \\\\ \\end{array}", + "image_path": "89b6ebf793030946e6a3bf877f94dc61904049cb7e76d68813bba9d98c01507a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 535, + 287, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 535, + 287, + 570 + ], + "spans": [ + { + "bbox": [ + 47, + 535, + 287, + 570 + ], + "type": "text", + "content": "The shadow indicator " + }, + { + "bbox": [ + 47, + 535, + 287, + 570 + ], + "type": "inline_equation", + "content": "\\psi_{j}" + }, + { + "bbox": [ + 47, + 535, + 287, + 570 + ], + "type": "text", + "content": " from Eq. (6) is similarly approximated by sampling on " + }, + { + "bbox": [ + 47, + 535, + 287, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{ill} = \\mathbf{o}_i + \\omega_j l" + }, + { + "bbox": [ + 47, + 535, + 287, + 570 + ], + "type": "text", + "content": " a set of points " + }, + { + "bbox": [ + 47, + 535, + 287, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{ill}" + }, + { + "bbox": [ + 47, + 535, + 287, + 570 + ], + "type": "text", + "content": " bounded between " + }, + { + "bbox": [ + 47, + 535, + 287, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_i" + }, + { + "bbox": [ + 47, + 535, + 287, + 570 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 535, + 287, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_j" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 117, + 569, + 287, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 569, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 117, + 569, + 287, + 601 + ], + "type": "interline_equation", + "content": "\\psi_ {j} = \\exp \\left(- \\sum_ {k} ^ {\\mathbf {X} _ {i l l}} \\sigma_ {k} \\delta_ {k}\\right). \\tag {11}", + "image_path": "c7d6aa87972924cb7eb84d91a3ae09a618ee0067ca4f3f38524703fbd481a7b0.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 606, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 606, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 47, + 606, + 287, + 629 + ], + "type": "text", + "content": "For the passive slice, the active component is null, further simplifying to" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 113, + 628, + 287, + 660 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 628, + 287, + 660 + ], + "spans": [ + { + "bbox": [ + 113, + 628, + 287, + 660 + ], + "type": "interline_equation", + "content": "\\tilde {I} _ {P} (\\mathbf {r}) = \\sum_ {j = 0} ^ {N} w _ {j} \\Lambda_ {j} + \\mathcal {D} _ {P}. \\tag {12}", + "image_path": "24f97a3d9b5dd57ef275dd224be70e87cae453b0352ae567eeaf5e9ccf4f3797.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "content": "Both " + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{ray}" + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{ill}" + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "content": " are sampled using a proposal network [7] " + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "f_{P}" + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "content": " that, analogously to " + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "f_{Gd}" + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "content": ", predicts point-wise densities converted with Eq. (10) to proposal weights " + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\hat{w}" + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "content": " for sampling with piece-wise-constant probabilities." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 265, + 427, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 265, + 427, + 277 + ], + "spans": [ + { + "bbox": [ + 305, + 265, + 427, + 277 + ], + "type": "text", + "content": "4.3. Training Supervision" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 277, + 545, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 545, + 337 + ], + "type": "text", + "content": "We supervise the predicted passive and active gated frames applying a photometric loss " + }, + { + "bbox": [ + 304, + 277, + 545, + 337 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 304, + 277, + 545, + 337 + ], + "type": "text", + "content": ", regularize the volumetric density with a depth loss " + }, + { + "bbox": [ + 304, + 277, + 545, + 337 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_d" + }, + { + "bbox": [ + 304, + 277, + 545, + 337 + ], + "type": "text", + "content": " and by supervising the shadow estimate with " + }, + { + "bbox": [ + 304, + 277, + 545, + 337 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 304, + 277, + 545, + 337 + ], + "type": "text", + "content": ". We regularize normal and reflectance estimates through " + }, + { + "bbox": [ + 304, + 277, + 545, + 337 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{nc}" + }, + { + "bbox": [ + 304, + 277, + 545, + 337 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 277, + 545, + 337 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\alpha}" + }, + { + "bbox": [ + 304, + 277, + 545, + 337 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 338, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 338, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 305, + 338, + 545, + 361 + ], + "type": "text", + "content": "Photometric Loss We supervise with ground truth captures for active and passive gated slice reconstruction as" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 310, + 362, + 545, + 387 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 362, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 310, + 362, + 545, + 387 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c} = \\sum_ {k, r} \\| \\tilde {I} _ {k} (\\mathbf {r}) - I _ {k} (\\mathbf {r}) \\| _ {2} + \\sum_ {r} \\| \\tilde {I} _ {P} (\\mathbf {r}) - I _ {P} (\\mathbf {r}) \\| _ {2} \\tag {13}", + "image_path": "380a28e7246fbe5a4b8f740c923093674428860ea824cfaf72c1ca44ddb55374.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 394, + 545, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 394, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 304, + 394, + 545, + 442 + ], + "type": "text", + "content": "Volume Density Regularization As additional training supervision, we use the depth estimate " + }, + { + "bbox": [ + 304, + 394, + 545, + 442 + ], + "type": "inline_equation", + "content": "\\hat{D}(\\mathbf{r})" + }, + { + "bbox": [ + 304, + 394, + 545, + 442 + ], + "type": "text", + "content": " of a pretrained stereo depth estimation algorithm [90] as pseudo ground-truth to regularize the ray termination distribution [26]" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 442, + 545, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 442, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 321, + 442, + 545, + 475 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d} = \\sum_ {\\mathbf {r}} \\sum_ {j} \\log w _ {j} \\exp \\left(- \\frac {\\left(l _ {j} - \\hat {D} (\\mathbf {r})\\right) ^ {2}}{2 s ^ {2}}\\right) \\delta_ {i} \\tag {14}", + "image_path": "c8c993989dbb3847867546ff87d077932145ec240ef35152ff042dd941cb5381.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 481, + 545, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 481, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 304, + 481, + 545, + 553 + ], + "type": "text", + "content": "We regularize the density field by partially supervising the shadow indicator " + }, + { + "bbox": [ + 304, + 481, + 545, + 553 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 304, + 481, + 545, + 553 + ], + "type": "text", + "content": ". Each pixel whose active intensity " + }, + { + "bbox": [ + 304, + 481, + 545, + 553 + ], + "type": "inline_equation", + "content": "I_{kA} = I_k - I_P" + }, + { + "bbox": [ + 304, + 481, + 545, + 553 + ], + "type": "text", + "content": " in any of the three gated slices is above a certain threshold " + }, + { + "bbox": [ + 304, + 481, + 545, + 553 + ], + "type": "inline_equation", + "content": "\\epsilon_i" + }, + { + "bbox": [ + 304, + 481, + 545, + 553 + ], + "type": "text", + "content": " is considered as visible from the illuminator. We hence supervise the expected shadow value for such rays " + }, + { + "bbox": [ + 304, + 481, + 545, + 553 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_v \\in \\{\\mathbf{r} | \\forall k \\in \\{1, 2, 3\\} : I_{kA}(\\mathbf{r}) > \\epsilon_i\\}" + }, + { + "bbox": [ + 304, + 481, + 545, + 553 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 348, + 554, + 545, + 580 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 554, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 348, + 554, + 545, + 580 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s} = \\sum_ {\\mathbf {r} _ {v}} \\| 1 - \\int T (l) \\sigma (\\mathbf {x}) \\psi (\\mathbf {x}) d l \\| _ {2} \\tag {15}", + "image_path": "9fd7901dd6a92e033606ae1803be69c07518f04f4d130da79ac5dbade3a6ebd4.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 582, + 545, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 631 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 631 + ], + "type": "text", + "content": "Normals Consistency Following [88], for each sampled point " + }, + { + "bbox": [ + 304, + 582, + 545, + 631 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 582, + 545, + 631 + ], + "type": "text", + "content": " we enforce a consistency between the predicted normal " + }, + { + "bbox": [ + 304, + 582, + 545, + 631 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 304, + 582, + 545, + 631 + ], + "type": "text", + "content": " and the density gradient " + }, + { + "bbox": [ + 304, + 582, + 545, + 631 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{n}} (\\mathbf{x}) = -\\frac{\\nabla\\mathbf{x}}{||\\nabla\\mathbf{x}||}" + }, + { + "bbox": [ + 304, + 582, + 545, + 631 + ], + "type": "text", + "content": ", and we penalize normals which are back-facing the camera as" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 632, + 545, + 655 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 632, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 306, + 632, + 545, + 655 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {n c} = \\sum_ {\\mathbf {x}} w (\\mathbf {x}) \\left(\\| \\mathbf {n} (\\mathbf {x}) - \\hat {\\mathbf {n}} (\\mathbf {x}) \\| _ {2} + \\max (0, \\hat {\\mathbf {n}} (\\mathbf {x}) \\cdot \\mathbf {d}) ^ {2}\\right) \\tag {16}", + "image_path": "41f29746653c89f3c32eba9263b6135b71f1eb7270d9372b8381a23246c2b31d.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 657, + 545, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 657, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 305, + 657, + 545, + 681 + ], + "type": "text", + "content": "Reflectance Regularization We enforce the predicted scene reflection " + }, + { + "bbox": [ + 305, + 657, + 545, + 681 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 305, + 657, + 545, + 681 + ], + "type": "text", + "content": " to be spatially consistent within " + }, + { + "bbox": [ + 305, + 657, + 545, + 681 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\mathbf{x}}" + }, + { + "bbox": [ + 305, + 657, + 545, + 681 + ], + "type": "text", + "content": ", i.e." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 311, + 691, + 545, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 691, + 545, + 716 + ], + "spans": [ + { + "bbox": [ + 311, + 691, + 545, + 716 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\alpha} = \\sum_ {\\mathbf {x}} w (\\mathbf {x}) \\left(\\| \\alpha (\\mathbf {x}, \\mathbf {d}) - \\alpha \\left(\\mathbf {x} + \\epsilon_ {\\mathbf {x}}, \\mathbf {d} + \\epsilon_ {\\mathbf {d}}\\right) \\| _ {2}\\right). \\tag {17}", + "image_path": "0f057e376204cda3fdbcc846dcb4b5e413c91c2dc7b23d6dc18e25709b8eb4e3.jpg" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10534" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": "Here the " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " dependency of " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " is omitted here for brevity. We also include an angular noise " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\mathrm{d}}" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " that we set high at the beginning of the training and then decrease it exponentially. By forcing the reflectance to behave as fully diffuse at the beginning of the training, we disincentive it to bake in the effects from lighting or shadowing, hence improving the disjoint learning of the scene components." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 158, + 287, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 158, + 287, + 182 + ], + "spans": [ + { + "bbox": [ + 47, + 158, + 287, + 182 + ], + "type": "text", + "content": "Total Training Loss Combining the different losses we obtain the following loss formulation:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 64, + 191, + 287, + 205 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 191, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 64, + 191, + 287, + 205 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {1} \\mathcal {L} _ {c} + \\lambda_ {2} \\mathcal {L} _ {d} + \\lambda_ {3} \\mathcal {L} _ {s} + \\lambda_ {4} \\mathcal {L} _ {n c} + \\lambda_ {5} \\mathcal {L} _ {\\alpha}, \\tag {18}", + "image_path": "c2f958681e227df536df7ce351662ed4dbb30e9cee5bef21f53408d0c7f9e8a8.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 213, + 287, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 213, + 287, + 225 + ], + "spans": [ + { + "bbox": [ + 47, + 213, + 287, + 225 + ], + "type": "text", + "content": "see " + }, + { + "bbox": [ + 47, + 213, + 287, + 225 + ], + "type": "inline_equation", + "content": "\\lambda_{1,\\dots,5}" + }, + { + "bbox": [ + 47, + 213, + 287, + 225 + ], + "type": "text", + "content": " hyperparameters in the Supplementary Material." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 229, + 183, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 229, + 183, + 243 + ], + "spans": [ + { + "bbox": [ + 47, + 229, + 183, + 243 + ], + "type": "text", + "content": "5. Implementation Details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 244, + 287, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 244, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 46, + 244, + 287, + 352 + ], + "type": "text", + "content": "We train for 35,000 steps and a batch size of 4096 rays. As optimizer we use ADAMW [57] with " + }, + { + "bbox": [ + 46, + 244, + 287, + 352 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 46, + 244, + 287, + 352 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 244, + 287, + 352 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.999" + }, + { + "bbox": [ + 46, + 244, + 287, + 352 + ], + "type": "text", + "content": ", learning rate " + }, + { + "bbox": [ + 46, + 244, + 287, + 352 + ], + "type": "inline_equation", + "content": "10^{-2}" + }, + { + "bbox": [ + 46, + 244, + 287, + 352 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 244, + 287, + 352 + ], + "type": "inline_equation", + "content": "f_{P}" + }, + { + "bbox": [ + 46, + 244, + 287, + 352 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 244, + 287, + 352 + ], + "type": "inline_equation", + "content": "f_{G}" + }, + { + "bbox": [ + 46, + 244, + 287, + 352 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 244, + 287, + 352 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 46, + 244, + 287, + 352 + ], + "type": "text", + "content": " for the camera poses optimization, " + }, + { + "bbox": [ + 46, + 244, + 287, + 352 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 46, + 244, + 287, + 352 + ], + "type": "text", + "content": " for the laser profile and gated parameters. We train on two NVIDIA V100 GPUs, for approximately 3 hours. The proposal network " + }, + { + "bbox": [ + 46, + 244, + 287, + 352 + ], + "type": "inline_equation", + "content": "f_{P}" + }, + { + "bbox": [ + 46, + 244, + 287, + 352 + ], + "type": "text", + "content": " is comprised of two MLPs and trained following [7]. Additional architecture details, training procedures, and hyper-parameters are found in the Supplementary Material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 357, + 101, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 357, + 101, + 369 + ], + "spans": [ + { + "bbox": [ + 47, + 357, + 101, + 369 + ], + "type": "text", + "content": "6. Dataset" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 371, + 289, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 371, + 289, + 622 + ], + "spans": [ + { + "bbox": [ + 46, + 371, + 289, + 622 + ], + "type": "text", + "content": "To conduct this work, we have collected a diverse set of 10 static sequences, recorded in both day and night conditions across North America. To this goal, we equipped a test vehicle with a NIR gated stereo camera setup (Bright-Way Vision), an automotive RGB stereo camera (OnSemi AR0230), a LiDAR sensor (Velodyne VLS128) and a GNSS with IMU (Xsens MTi-7), as shown in Fig. 4. Each gated camera has a resolution of " + }, + { + "bbox": [ + 46, + 371, + 289, + 622 + ], + "type": "inline_equation", + "content": "1280 \\times 720" + }, + { + "bbox": [ + 46, + 371, + 289, + 622 + ], + "type": "text", + "content": " pixels, 10 bit depth and runs at " + }, + { + "bbox": [ + 46, + 371, + 289, + 622 + ], + "type": "inline_equation", + "content": "120 \\mathrm{~Hz}" + }, + { + "bbox": [ + 46, + 371, + 289, + 622 + ], + "type": "text", + "content": ", split up to collect the three active and one passive slice. The illuminator source consists of two vertical-cavity surface-emitting laser (VCSEL) modules, which illuminate the scene with a laser pulse with duration of 240-370 ns and a wavelength of " + }, + { + "bbox": [ + 46, + 371, + 289, + 622 + ], + "type": "inline_equation", + "content": "808 \\mathrm{~nm}" + }, + { + "bbox": [ + 46, + 371, + 289, + 622 + ], + "type": "text", + "content": ". The RGB cameras provide 12 bit HDR images with resolution of " + }, + { + "bbox": [ + 46, + 371, + 289, + 622 + ], + "type": "inline_equation", + "content": "1920 \\times 1080" + }, + { + "bbox": [ + 46, + 371, + 289, + 622 + ], + "type": "text", + "content": " pixels and " + }, + { + "bbox": [ + 46, + 371, + 289, + 622 + ], + "type": "inline_equation", + "content": "30 \\mathrm{~Hz}" + }, + { + "bbox": [ + 46, + 371, + 289, + 622 + ], + "type": "text", + "content": " frame-rate. The LiDAR has a vertical resolution of 128 lines and " + }, + { + "bbox": [ + 46, + 371, + 289, + 622 + ], + "type": "inline_equation", + "content": "10 \\mathrm{~Hz}" + }, + { + "bbox": [ + 46, + 371, + 289, + 622 + ], + "type": "text", + "content": " framerate, while the GNSS sensor runs at " + }, + { + "bbox": [ + 46, + 371, + 289, + 622 + ], + "type": "inline_equation", + "content": "4 \\mathrm{~Hz}" + }, + { + "bbox": [ + 46, + 371, + 289, + 622 + ], + "type": "text", + "content": ". Example captures from the dataset are being visualized in Fig. 4. In total, we collect 2650 samples, captured in both day (1223 samples) and night (1427 samples). We divide it in training, validation and test splits with a 50-25-25 split." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 623, + 287, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 287, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 287, + 670 + ], + "type": "text", + "content": "As ground-truth, we construct a large-scale ground-truth pointcloud by aggregating LiDAR scans with LIO-SAM [78] and removing noisy points. Additional details on the accumulation are provided in the Supplementary Material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 675, + 121, + 687 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 675, + 121, + 687 + ], + "spans": [ + { + "bbox": [ + 47, + 675, + 121, + 687 + ], + "type": "text", + "content": "7. Assessment" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "In this section, we validate the proposed method quantitatively and qualitatively. Specifically, we investigate scene" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 305, + 60, + 553, + 203 + ], + "blocks": [ + { + "bbox": [ + 305, + 60, + 553, + 203 + ], + "lines": [ + { + "bbox": [ + 305, + 60, + 553, + 203 + ], + "spans": [ + { + "bbox": [ + 305, + 60, + 553, + 203 + ], + "type": "image", + "image_path": "3aa04838ebe91debfd13421c7886c9598b95b0ec86127b16edc92eeada4ab79a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 313, + 205, + 550, + 318 + ], + "blocks": [ + { + "bbox": [ + 313, + 205, + 550, + 318 + ], + "lines": [ + { + "bbox": [ + 313, + 205, + 550, + 318 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 550, + 318 + ], + "type": "image", + "image_path": "5643129c4f14fc3124cc751a35972b2d7735a1efa25bf35b89dcc5d3f45a9d71.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 325, + 547, + 392 + ], + "lines": [ + { + "bbox": [ + 305, + 325, + 547, + 392 + ], + "spans": [ + { + "bbox": [ + 305, + 325, + 547, + 392 + ], + "type": "text", + "content": "Figure 4. Top: Example captures from our collected dataset across different urban and suburban areas in North America. From left to right: RGB image, active gated slices (with red for slice 1, green for slice 2 and blue for slice 3), passive slice, projected LiDAR scan, accumulated LiDAR. Bottom: Sensors setup with LiDAR, stereo Gated camera, stereo RGB camera, IMU and GNSS." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 403, + 545, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 545, + 476 + ], + "type": "text", + "content": "reconstruction at both day and night, using novel depth and view synthesis for 2D evaluation, and surface reconstruction for the 3D evaluation. To this end, we compare our approach to state-of-the-art feed-forward depth estimation algorithms and neural scene reconstruction methods. We also conduct ablation experiments to validate our design choices." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 486, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 546, + 714 + ], + "type": "text", + "content": "Depth Reconstruction We assess the quality of depth synthesis of Gated Fields for camera poses unseen during training. We use as ground truth the accumulated and filtered LiDAR pointcloud. Unlike previous works relying on single LiDAR scans for evaluation [35, 89, 90], we use as ground truth an accumulated LiDAR pointcloud as described in Sec. 6. This allows us to evaluate the depth reconstruction up to " + }, + { + "bbox": [ + 304, + 486, + 546, + 714 + ], + "type": "inline_equation", + "content": "160\\mathrm{m}" + }, + { + "bbox": [ + 304, + 486, + 546, + 714 + ], + "type": "text", + "content": " accurately and without bias. We follow previous works [89, 90] and use as depth evaluation metrics RMSE, MAE, ARD, " + }, + { + "bbox": [ + 304, + 486, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\sigma_{i} < 1.25^{i}, i \\in \\{1,2,3\\}" + }, + { + "bbox": [ + 304, + 486, + 546, + 714 + ], + "type": "text", + "content": ". We compare our method against 9 feed-forward depth estimation methods, namely SimIPU [53], AdaBins [10], DPT [70], DepthFormer [54] and CREStereo [50] for monocular and stereo RGB methods, Gated2Gated [89] and GatedStereo [90] for monocular and stereo gated estimation methods. We also compare our method against depths rendered with other neural reconstruction algorithms, using RGB images [7],[63],[29],[26] LiDAR [84], RGB+LiDAR [86], [37], and a varying-appearance method [29] for gated" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10535" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 53, + 550, + 383 + ], + "blocks": [ + { + "bbox": [ + 47, + 53, + 550, + 383 + ], + "lines": [ + { + "bbox": [ + 47, + 53, + 550, + 383 + ], + "spans": [ + { + "bbox": [ + 47, + 53, + 550, + 383 + ], + "type": "image", + "image_path": "ab0701162834ba87697bca1bb1490b5b0aebca8011d2d528ffea7c1506e19800.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 393, + 547, + 449 + ], + "lines": [ + { + "bbox": [ + 46, + 393, + 547, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 393, + 547, + 449 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison of the proposed Gated Fields and state-of-the-art depth estimation approaches, including LiDAR-NeRF [84], StreetSurf [37], NLSPN [67], and Gated Stereo [90]. Compared to baseline methods, we are able to reconstruct fine geometry details like branches or poles, also for far distances. Unlike RGB methods, Gated Fields is unaffected by poor ambient lighting, and unlike LiDAR-based methods it is able to reconstruct sharp object discontinuities. The active gated slices are visualized in red for slice 1, green for slice 2 and blue for slice 3." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 458, + 289, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 458, + 289, + 674 + ], + "spans": [ + { + "bbox": [ + 46, + 458, + 289, + 674 + ], + "type": "text", + "content": "captures. Results for day and night sequences are presented in Tab. 1. We outperform the next best neural field method by " + }, + { + "bbox": [ + 46, + 458, + 289, + 674 + ], + "type": "inline_equation", + "content": "21.87\\%" + }, + { + "bbox": [ + 46, + 458, + 289, + 674 + ], + "type": "text", + "content": " MAE and " + }, + { + "bbox": [ + 46, + 458, + 289, + 674 + ], + "type": "inline_equation", + "content": "30.35\\%" + }, + { + "bbox": [ + 46, + 458, + 289, + 674 + ], + "type": "text", + "content": " in RMSE. For night sequences the performance difference sharpens, with Gated Fields outperforming the best RGB-based method [50] by " + }, + { + "bbox": [ + 46, + 458, + 289, + 674 + ], + "type": "inline_equation", + "content": "3.14\\mathrm{m}" + }, + { + "bbox": [ + 46, + 458, + 289, + 674 + ], + "type": "text", + "content": " MAE. This performance decline is to be attributed to the limited pixel information present in RGB captures taken at night time, making impossible to learn a meaningful 3D representation of the scene, as shown qualitatively in Fig. 5. LiDAR-based methods are unaffected by the change in illumination, but suffer from the limited sensor resolution. On the other hand, gated cameras retrieve information-rich captures both day and night, which Gated Fields can explicitly leverage during training. We confirm that employing state-of-the-art neural field methods on gated captures does not yield accurate results, as they are not able to model the gated imaging formation and can only fit the ambient light component." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "content": "3D Reconstruction We evaluate the 3D scene reconstruction capabilities of Gated Fields using the accumulated Li-DAR pointcloud as ground truth. We follow [20] and extract" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 458, + 545, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 458, + 545, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 458, + 545, + 626 + ], + "type": "text", + "content": "for both the 3D ground truth pointcloud and different neural field-based methods a voxelized occupancy grid of the scene, and compute intersection over union (IoU), Precision and Recall between ground truth voxels and estimated ones. Quantitative results are shown in Tab. 3. Our method outperforms RGB baselines [7, 29] by an average of " + }, + { + "bbox": [ + 304, + 458, + 545, + 626 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 304, + 458, + 545, + 626 + ], + "type": "text", + "content": " IoU. RGB baselines using additional LiDAR sensor [37, 86] data partially improve the results, but such methods are still unable to reconstruct finer surfaces details and struggle at night. On the other hand, Gated Fields is able to recover finer geometries using gated, illumination and depth cues, and the quality does not degrade with diminishing ambient light, as shown in Fig. 5. See details on evaluation and further qualitative results in the Supplementary Material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "Novel View Synthesis For novel view synthesis, we compare our method with Mip-NeRF360 [7], a state-of-the-art neural radiance field-based method, and K-Planes [29], to implicitly model the time-varying appearance of the static scene. Mip-NeRF [7] struggles to reconstruct novel views due to the inherent difficulty of modeling the gating imaging effects, resulting in a PSNR of 17.16dB. By learning" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10536" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 63, + 282, + 393 + ], + "blocks": [ + { + "bbox": [ + 50, + 63, + 282, + 393 + ], + "lines": [ + { + "bbox": [ + 50, + 63, + 282, + 393 + ], + "spans": [ + { + "bbox": [ + 50, + 63, + 282, + 393 + ], + "type": "table", + "html": "
MethodModalityRMSE [m]ARDMAE [m]δ1 [%]δ2 [%]δ3 [%]
Test Data - Night (Evaluated on Accumulated LiDAR Ground-Truth Points)
2D DEPTH COMPARISONGATED2GATED [89]Gated13.330.288.5761.7890.3094.90
GATEDSTEREO [90]Stereo-Gated10.100.205.9782.8693.3596.52
SIMIPU [53]RGB19.330.4414.2140.6777.9989.80
ADABINS [10]RGB21.140.3814.2851.7280.0090.64
DPT [70]RGB14.170.289.9062.8288.3294.42
DEPTHFORMER [54]RGB14.320.2910.0861.1987.7294.57
CRESTEREO [50]Stereo-RGB14.220.137.2782.2190.1094.32
NLSPN [67]RGB+LiDAR11.120.186.4877.1590.8196.00
MIPNERF360 [7]RGB23.800.5116.4341.6061.7576.16
K-PLANES [29]RGB19.700.4113.6644.3966.2081.46
RAWNERF [63]RGB27.460.6419.7534.3254.3369.41
DEPTH-NERF [26]RGB15.230.2610.0461.6786.2393.87
SUDS [86]RGB+LiDAR11.070.176.1779.4988.4195.31
STREETSURF [37]RGB+LiDAR10.860.155.9082.1693.5796.96
LiDAR-NERF [50]LiDAR10.210.124.7287.7195.0597.78
GATED FIELDS [50]Gated7.920.124.1390.6195.7697.90
Test Data - Day (Evaluated on Accumulated LiDAR Ground-Truth Points)
2D DEPTH COMPARISONGATED2GATED [89]Gated9.260.226.6958.4693.7097.38
GATEDSTEREO [90]Stereo-Gated6.320.093.3092.8697.3198.53
SIMIPU [53]RGB13.540.3110.0852.9086.4995.65
ADABINS [10]RGB12.740.258.3969.8489.1395.52
DPT [70]RGB10.340.217.0877.6194.2997.24
DEPTHFORMER [54]RGB9.060.196.0981.1994.1497.42
CRESTEREO [50]Stereo-RGB7.350.093.4594.4897.2398.50
NLSPN [67]RGB+LiDAR10.340.175.9777.8391.1696.11
MIPNERF360 [7]RGB16.910.319.5370.8984.4390.92
K-PLANES [29]RGB12.370.248.5563.1679.6891.70
RAWNERF [63]RGB15.100.239.3865.9084.9092.43
DEPTH-NERF [26]RGB10.340.176.0778.9790.7596.53
SUDS [86]RGB+LiDAR9.110.175.8480.9695.0898.33
STREETSURF [37]RGB+LiDAR9.600.135.3583.9495.3298.56
LiDAR-NERF [50]LiDAR8.130.103.8688.9995.4998.06
GATED FIELDS [50]Gated6.150.092.9193.8897.3298.75
", + "image_path": "a2cef6d30e7386f3bcf1d101e0dfd3edfd5170e0082db277465bc71addf5309a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 61, + 442, + 272, + 567 + ], + "blocks": [ + { + "bbox": [ + 47, + 398, + 287, + 431 + ], + "lines": [ + { + "bbox": [ + 47, + 398, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 47, + 398, + 287, + 431 + ], + "type": "text", + "content": "Table 1. Comparison of our proposed approach and state-of-the-art approaches on depth synthesis. Best results in each category are in bold and second best are underlined." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 442, + 272, + 567 + ], + "lines": [ + { + "bbox": [ + 61, + 442, + 272, + 567 + ], + "spans": [ + { + "bbox": [ + 61, + 442, + 272, + 567 + ], + "type": "table", + "html": "
Im. Formation ModelDepth Sup.Normal Active Illum.RMSE [m]MAE [m]PSNR [dB]SSIM
Test Data - Night
ABLATIONEnd2End [62]XXXX27.3419.9517.430.633
GatedXXXX9.348.3523.470.889
Gated✓XXX6.725.4325.840.908
Gated✓✓✓X3.782.8230.660.946
Gated✓✓✓✓3.392.5130.910.95
Test Data - Day
ABLATIONEnd2End [62]XXXX30.0319.2216.770.66
GatedXXXX11.646.3426.770.915
Gated✓XXX9.204.2226.880.922
Gated✓✓✓X9.454.3132.150.946
Gated✓✓✓✓8.884.1232.280.948
", + "image_path": "e5de8884bfc78b9911c769f2f3914d7afef4cb26af82c59abc4b045ff63b49b8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 615, + 286, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 615, + 286, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 615, + 286, + 685 + ], + "type": "text", + "content": "a time-varying appearance, K-Planes improves the quality reaching 27.42dB PSNR for day but only 19.35dB for night, as the model fails to learn an accurate scene geometry representation without ambient light information. Gated Fields outperforms these baselines in both day and night, reaching a PSNR of 32.28dB." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Ablation Experiments To assess the role and contribution of the different components of our method, we conduct" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 328, + 61, + 521, + 129 + ], + "blocks": [ + { + "bbox": [ + 47, + 571, + 287, + 604 + ], + "lines": [ + { + "bbox": [ + 47, + 571, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 47, + 571, + 287, + 604 + ], + "type": "text", + "content": "Table 2. Ablation studies of the Gated Fields contributions, on a subset of the test dataset. We investigate different image formation models, neural fields components and supervision losses." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 61, + 521, + 129 + ], + "lines": [ + { + "bbox": [ + 328, + 61, + 521, + 129 + ], + "spans": [ + { + "bbox": [ + 328, + 61, + 521, + 129 + ], + "type": "table", + "html": "
MethodModalityIoU [%]Precision [%]Recall [%]
3D REC.MIPNERF360 [7]RGB6.327.3431.21
STREETSURF [37]RGB+LiDAR5.416.3127.35
SUDS [86]RGB+LiDAR8.969.8849.09
LiDAR-NERF [84]LiDAR20.0332.3834.44
Gated Fields (ours)Gated22.2525.0166.51
", + "image_path": "b3ae9d67f26df991927204f6b3d71488813fd1a7602935decd83d7567ea55106.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 133, + 545, + 188 + ], + "lines": [ + { + "bbox": [ + 305, + 133, + 545, + 188 + ], + "spans": [ + { + "bbox": [ + 305, + 133, + 545, + 188 + ], + "type": "text", + "content": "Table 3. Comparison of Gated Fields and state-of-the-art scene reconstruction methods. We evaluate over 3D occupancy reconstruction, using as ground truth the voxelized accumulated LiDAR pointcloud. Best results in each category are in bold and second best are underlined." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 198, + 545, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 198, + 545, + 390 + ], + "spans": [ + { + "bbox": [ + 304, + 198, + 545, + 390 + ], + "type": "text", + "content": "an ablation study in Tab. 2. In particular, we consider as starting point a single neural field directly inferring one intensity value for each of the four slices (3 active + 1 passive), and obtain an average MAE of " + }, + { + "bbox": [ + 304, + 198, + 545, + 390 + ], + "type": "inline_equation", + "content": "19.58\\mathrm{m}" + }, + { + "bbox": [ + 304, + 198, + 545, + 390 + ], + "type": "text", + "content": ". By separately predicting ambient light and reflectance, and reconstructing the gated image as in Eq. (2), we significantly improve the MAE to " + }, + { + "bbox": [ + 304, + 198, + 545, + 390 + ], + "type": "inline_equation", + "content": "7.34\\mathrm{m}" + }, + { + "bbox": [ + 304, + 198, + 545, + 390 + ], + "type": "text", + "content": ". However, this approach still performs poorly on flat-color areas during the day and in unilluminated areas during the night due to lack of any depth cue. By adding the depth supervision, we are able to supervise also such areas and the PSNR improves by " + }, + { + "bbox": [ + 304, + 198, + 545, + 390 + ], + "type": "inline_equation", + "content": "4.83\\mathrm{dB}" + }, + { + "bbox": [ + 304, + 198, + 545, + 390 + ], + "type": "text", + "content": ". By adding the angular-dependent attenuation and regularizing the reflectance in Eq. (17), the model is able to disentangle the material properties from other spurious effects. Finally, by explicitly modeling the shadow, casted by the illuminator, we improve final depth reconstruction to " + }, + { + "bbox": [ + 304, + 198, + 545, + 390 + ], + "type": "inline_equation", + "content": "3.32\\mathrm{m}" + }, + { + "bbox": [ + 304, + 198, + 545, + 390 + ], + "type": "text", + "content": " MAE." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 401, + 378, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 401, + 378, + 414 + ], + "spans": [ + { + "bbox": [ + 306, + 401, + 378, + 414 + ], + "type": "text", + "content": "8. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 420, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 420, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 420, + 545, + 624 + ], + "type": "text", + "content": "We introduce Gated Fields, a neural rendering method capable of reconstructing scene geometry from video captures of active time-gated cameras. The method hinges on a differentiable gated image formation as part of the rendering formulation, and it jointly learns geometry, ambient light and surface proprieties, represented implicitly as neural field components, alongside illumination and gating parameters, represented with physics-based models. Extensive experiments on real-world large-scale scenes validate that our method is able to precisely reconstruct a 3D scene both in day and night-time conditions. Our approach outperforms existing RGB and LiDAR methods by " + }, + { + "bbox": [ + 304, + 420, + 545, + 624 + ], + "type": "inline_equation", + "content": "21.87\\%" + }, + { + "bbox": [ + 304, + 420, + 545, + 624 + ], + "type": "text", + "content": " on MAE, as well as baseline methods using gated captures by " + }, + { + "bbox": [ + 304, + 420, + 545, + 624 + ], + "type": "inline_equation", + "content": "31.67\\%" + }, + { + "bbox": [ + 304, + 420, + 545, + 624 + ], + "type": "text", + "content": ". In the future, we hope to extend the proposed approach by \"closing the loop\" and providing dynamic feedback to the gated acquisition, allowing for adaptive gated scene reconstructions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 628, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 628, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 628, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgments This work was supported by the AI-SEE project with funding from the FFG, BMBF, and NRC-IRA. Felix Heide was supported by an NSF CAREER Award (2047359), a Packard Foundation Fellowship, a Sloan Research Fellowship, a Sony Young Faculty Award, a Project X Innovation Award, and an Amazon Science Research Award." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10537" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 58, + 91, + 287, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 91, + 287, + 124 + ], + "spans": [ + { + "bbox": [ + 58, + 91, + 287, + 124 + ], + "type": "text", + "content": "[1] Amit Adam, Christoph Dann, Omer Yair, Shai Mazor, and Sebastian Nowozin. Bayesian time-of-flight for realtime shape, illumination and albedo. 39(5):851-864, 2017. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 126, + 287, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 126, + 287, + 179 + ], + "spans": [ + { + "bbox": [ + 58, + 126, + 287, + 179 + ], + "type": "text", + "content": "[2] Ashutosh Agarwal and Chetan Arora. Attention attention everywhere: Monocular depth prediction with skip attention. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5861-5870, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 182, + 286, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 182, + 286, + 213 + ], + "spans": [ + { + "bbox": [ + 58, + 182, + 286, + 213 + ], + "type": "text", + "content": "[3] Pierre Andersson. Long-range three-dimensional imaging using range-gated laser radar images. 45(3):034301, 2006. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 217, + 287, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 217, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 58, + 217, + 287, + 270 + ], + "type": "text", + "content": "[4] Benjamin Attal, Eliot Laidlaw, Aaron Gokaslan, Changil Kim, Christian Richardt, James Tompkin, and Matthew O'Toole. Törf: Time-of-flight radiance fields for dynamic scene view synthesis. Advances in neural information processing systems, 34:26289-26301, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 273, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 273, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 58, + 273, + 287, + 316 + ], + "type": "text", + "content": "[5] Abhishek Badki, Alejandro Troccoli, Kihwan Kim, Jan Kautz, Pradeep Sen, and Orazio Gallo. Bi3D: Stereo depth estimation via binary classifications. In arXiv preprint arXiv:2005.07274, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 319, + 287, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 319, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 58, + 319, + 287, + 361 + ], + "type": "text", + "content": "[6] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. ICCV, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 364, + 287, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 364, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 58, + 364, + 287, + 418 + ], + "type": "text", + "content": "[7] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 1, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 58, + 420, + 287, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 420, + 287, + 452 + ], + "spans": [ + { + "bbox": [ + 58, + 420, + 287, + 452 + ], + "type": "text", + "content": "[8] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. ICCV, 2023. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 58, + 454, + 287, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 454, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 58, + 454, + 287, + 497 + ], + "type": "text", + "content": "[9] Florent Bartoccioni, Éloi Zablocki, Patrick Pérez, Matthieu Cord, and Karteek Alahari. Lidartouch: Monocular metric depth estimation with a few-beam lidar. Computer Vision and Image Understanding, 227:103601, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 500, + 287, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 500, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 53, + 500, + 287, + 543 + ], + "type": "text", + "content": "[10] Shariq Farooq Bhat, Ibraheem Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4009-4018, 2021. 6, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 544, + 287, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 544, + 287, + 587 + ], + "spans": [ + { + "bbox": [ + 53, + 544, + 287, + 587 + ], + "type": "text", + "content": "[11] Mario Bijelic, Tobias Gruber, and Werner Ritter. A benchmark for lidar sensors in fog: Is detection breaking down? In 2018 IEEE Intelligent Vehicles Symposium (IV), pages 760-767, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 590, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 590, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 53, + 590, + 287, + 622 + ], + "type": "text", + "content": "[12] Mario Bijelic, Tobias Gruber, and Werner Ritter. Benchmarking image sensors under adverse weather conditions for autonomous driving. 2018. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 624, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 624, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 53, + 624, + 287, + 667 + ], + "type": "text", + "content": "[13] Mario Bijelic, Tobias Gruber, and Werner Ritter. Benchmarking image sensors under adverse weather conditions for autonomous driving. In 2018 IEEE Intelligent Vehicles Symposium (IV), pages 1773-1779. IEEE, 2018. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 670, + 287, + 712 + ], + "type": "text", + "content": "[14] Mario Bijelic, Tobias Gruber, Fahim Mannan, Florian Kraus, Werner Ritter, Klaus Dietmayer, and Felix Heide. Seeing through fog without seeing fog: Deep multimodal sensor fusion in unseen adverse weather. In The IEEE" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 74, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 332, + 74, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 74, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 332, + 74, + 545, + 95 + ], + "type": "text", + "content": "Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 97, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 97, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 312, + 97, + 545, + 161 + ], + "type": "text", + "content": "[15] Mark Boss, Andreas Engelhardt, Abhishek Kar, Yuanzhen Li, Deqing Sun, Jonathan Barron, Hendrik Lensch, and Varun Jampani. *Samurai: Shape and material from unconstrained real-world arbitrary image collections*. Advances in Neural Information Processing Systems, 35:26389-26403, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 163, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 163, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 312, + 163, + 545, + 217 + ], + "type": "text", + "content": "[16] Aljaz Bozic, Pablo Palafox, Justus Thies, Angela Dai, and Matthias Nießner. Transformerfusion: Monocular rgb scene reconstruction using transformers. Advances in Neural Information Processing Systems, 34:1403-1414, 2021. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 220, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 220, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 312, + 220, + 545, + 240 + ], + "type": "text", + "content": "[17] Jens Busck. Underwater 3-D optical imaging with a gated viewing laser radar. 2005. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 243, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 243, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 312, + 243, + 545, + 274 + ], + "type": "text", + "content": "[18] Jens Busck and Henning Heiselberg. Gated viewing and high-accuracy three-dimensional laser radar. 43(24):4705–10, 2004. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 276, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 276, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 312, + 276, + 545, + 308 + ], + "type": "text", + "content": "[19] Jens Busck and Henning Heiselberg. Gated viewing and high-accuracy three-dimensional laser radar. Applied optics, 43(24):4705-4710, 2004. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 310, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 310, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 312, + 310, + 545, + 363 + ], + "type": "text", + "content": "[20] Anh-Quan Cao and Raoul de Charette. Scenerf: Self-supervised monocular 3d scene reconstruction with radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9387-9398, 2023. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 366, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 366, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 312, + 366, + 545, + 409 + ], + "type": "text", + "content": "[21] A. Carballo, J. Lambert, A. Monrroy, D. Wong, P. Narksri, Y. Kitsukawa, E. Takeuchi, S. Kato, and K. Takeda. Libre: The multiple 3d lidar dataset. In IEEE Intelligent Vehicles Symposium (IV), 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 411, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 411, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 312, + 411, + 545, + 453 + ], + "type": "text", + "content": "[22] Jia-Ren Chang and Yong-Sheng Chen. Pyramid stereo matching network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5410-5418, 2018. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 456, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 456, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 312, + 456, + 545, + 498 + ], + "type": "text", + "content": "[23] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 501, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 501, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 312, + 501, + 545, + 544 + ], + "type": "text", + "content": "[24] Jaesung Choe, Kyungdon Joo, Tooba Imtiaz, and In So Kweon. Volumetric propagation network: Stereo-lidar fusion for long-range depth estimation. IEEE Robotics and Automation Letters, 6(3):4672-4679, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 546, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 546, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 312, + 546, + 545, + 589 + ], + "type": "text", + "content": "[25] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 303-312, 1996. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 591, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 591, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 312, + 591, + 545, + 644 + ], + "type": "text", + "content": "[26] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nef: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12882-12891, 2022. 3, 5, 6, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 647, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 647, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 312, + 647, + 545, + 679 + ], + "type": "text", + "content": "[27] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. In NeurIPS, pages 2366-2374, 2014. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 681, + 545, + 713 + ], + "type": "text", + "content": "[28] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "10538" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 72, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 287, + 95 + ], + "type": "text", + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501-5510, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 97, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 97, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 53, + 97, + 287, + 162 + ], + "type": "text", + "content": "[29] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12479–12488, 2023. 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 164, + 288, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 164, + 288, + 198 + ], + "spans": [ + { + "bbox": [ + 53, + 164, + 288, + 198 + ], + "type": "text", + "content": "[30] Ravi Garg, B.G. Vijay Kumar, Gustavo Carneiro, and Ian Reid. Unsupervised CNN for single view depth estimation: Geometry to the rescue. In ECCV, pages 740-756, 2016. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 199, + 288, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 199, + 288, + 232 + ], + "spans": [ + { + "bbox": [ + 53, + 199, + 288, + 232 + ], + "type": "text", + "content": "[31] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, pages 3354-3361, 2012. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 234, + 288, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 234, + 288, + 288 + ], + "spans": [ + { + "bbox": [ + 53, + 234, + 288, + 288 + ], + "type": "text", + "content": "[32] Clément Godard, Oisin Mac Aodha, and Gabriel J Bros-tow. Unsupervised monocular depth estimation with left-right consistency. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 270-279, 2017. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 290, + 288, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 290, + 288, + 345 + ], + "spans": [ + { + "bbox": [ + 53, + 290, + 288, + 345 + ], + "type": "text", + "content": "[33] Clément Godard, Oisin Mac Aodha, Michael Firman, and Gabriel J Brostow. Digging into self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3828-3838, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 347, + 288, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 347, + 288, + 379 + ], + "spans": [ + { + "bbox": [ + 53, + 347, + 288, + 379 + ], + "type": "text", + "content": "[34] Yoav Grauer. Active gated imaging in driver assistance system. Advanced Optical Technologies, 3(2):151-160, 2014. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 381, + 288, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 381, + 288, + 426 + ], + "spans": [ + { + "bbox": [ + 53, + 381, + 288, + 426 + ], + "type": "text", + "content": "[35] Tobias Gruber, Frank Julca-Aguilar, Mario Bijelic, and Felix Heide. Gated2depth: Real-time dense lidar from gated images. In The IEEE International Conference on Computer Vision (ICCV), 2019. 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 427, + 288, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 427, + 288, + 483 + ], + "spans": [ + { + "bbox": [ + 53, + 427, + 288, + 483 + ], + "type": "text", + "content": "[36] Vitor Guizilini, Rares Ambrus, Sudeep Pillai, Allan Ravens \ntos, and Adrien Gaidon. 3d packing for self-supervised \nmonocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern \nRecognition, pages 2485-2494, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 484, + 288, + 539 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 484, + 288, + 539 + ], + "spans": [ + { + "bbox": [ + 53, + 484, + 288, + 539 + ], + "type": "text", + "content": "[37] Jianfei Guo, Nianchen Deng, Xinyang Li, Yeqi Bai, Botian Shi, Chiyu Wang, Chenjing Ding, Dongliang Wang, and Yikang Li. Streetsurf: Extending multi-view implicit surface reconstruction to street views. arXiv preprint arXiv:2306.04988, 2023. 1, 3, 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 541, + 288, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 541, + 288, + 574 + ], + "spans": [ + { + "bbox": [ + 53, + 541, + 288, + 574 + ], + "type": "text", + "content": "[38] Miles Hansard, Seungkyu Lee, Ouk Choi, and Radu Patrice Horaud. Time-of-flight cameras: principles, methods and applications. Springer Science & Business Media, 2012. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 576, + 288, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 576, + 288, + 597 + ], + "spans": [ + { + "bbox": [ + 53, + 576, + 288, + 597 + ], + "type": "text", + "content": "[39] Paul Heckman and Robert T. Hodgson. Underwater optical range gating. 3(11):445-448, 1967. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 600, + 288, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 600, + 288, + 632 + ], + "spans": [ + { + "bbox": [ + 53, + 600, + 288, + 632 + ], + "type": "text", + "content": "[40] Mu Hu, Shuling Wang, Bin Li, Shiyu Ning, Li Fan, and Xiaojin Gong. Towards precise and efficient image guided depth completion. 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 635, + 288, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 635, + 288, + 667 + ], + "spans": [ + { + "bbox": [ + 53, + 635, + 288, + 667 + ], + "type": "text", + "content": "[41] Shengyu Huang, Zan Gojcic, Zian Wang, Francis Williams, Yoni Kasten, Sanja Fidler, Konrad Schindler, and Or Litany. Neural lidar fields for novel view synthesis. 2023. 1, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 670, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 670, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 670, + 288, + 713 + ], + "type": "text", + "content": "[42] Maximilian Jaritz, Raoul De Charette, Emilie Wirbel, Xavier Perrotton, and Fawzi Nashashibi. Sparse and dense data with cnns: Depth completion and semantic segmentation. pages 52-60, 2018. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 72, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 312, + 72, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 72, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 312, + 72, + 545, + 106 + ], + "type": "text", + "content": "[43] Maria Jokela, Matti Kutila, and Pasi Pyykönen. Testing and validation of automotive point-cloud sensors in adverse weather conditions. Applied Sciences, 9, 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 108, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 108, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 312, + 108, + 545, + 162 + ], + "type": "text", + "content": "[44] Frank Julca-Aguilar, Jason Taylor, Mario Bijelic, Fahim Mannan, Ethan Tseng, and Felix Heide. Gated3d: Monocular 3d object detection from temporal illumination cues. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2938-2948, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 164, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 164, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 312, + 164, + 545, + 207 + ], + "type": "text", + "content": "[45] Alex Kendall, Hayk Martirosyan, Saumitro Dasgupta, Peter Henry, Ryan Kennedy, Abraham Bachrach, and Adam Bry. End-to-end learning of geometry and context for deep stereo regression. In ICCV, 2017. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 209, + 545, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 209, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 312, + 209, + 545, + 252 + ], + "type": "text", + "content": "[46] Andreas Kolb, Erhardt Barth, Reinhard Koch, and Rasmus Larsen. Time-of-flight cameras in computer graphics. In Computer Graphics Forum, pages 141-159. Wiley Online Library, 2010. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 255, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 255, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 312, + 255, + 545, + 330 + ], + "type": "text", + "content": "[47] Abhijit Kundu, Kyle Genova, Xiaqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12871-12881, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 332, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 332, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 312, + 332, + 545, + 365 + ], + "type": "text", + "content": "[48] Robert Lange. 3D time-of-flight distance measurement with custom solid-state image sensors in CMOS/CCD-technology. 2000. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 366, + 545, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 366, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 312, + 366, + 545, + 421 + ], + "type": "text", + "content": "[49] Deborah Levy, Amit Peleg, Naama Pearl, Dan Rosenbaum, Derya Akkaynak, Simon Korman, and Tali Treibitz. Seathru-nerf: Neural radiance fields in scattering media. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 56–65, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 422, + 545, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 422, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 312, + 422, + 545, + 488 + ], + "type": "text", + "content": "[50] Jiankun Li, Peisen Wang, Pengfei Xiong, Tao Cai, Ziwei Yan, Lei Yang, Jiangyu Liu, Haoqiang Fan, and Shuaicheng Liu. Practical stereo matching via cascaded recurrent network with adaptive correlation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16263-16272, 2022. 6, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 490, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 490, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 312, + 490, + 545, + 555 + ], + "type": "text", + "content": "[51] Zhengqi Li, Tali Dekel, Forrester Cole, Richard Tucker, Noah Snavely, Ce Liu, and William T. Freeman. Manequinchallenge: Learning the depths of moving people by watching frozen people. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(12):4229-4241, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 557, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 557, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 312, + 557, + 545, + 623 + ], + "type": "text", + "content": "[52] Zhaoshuo Li, Xingtong Liu, Nathan Drenkow, Andy Ding, Francis X. Creighton, Russell H. Taylor, and Mathias Unberath. Revisiting stereo depth estimation from a sequence-to-sequence perspective with transformers. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6177-6186, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 624, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 624, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 312, + 624, + 545, + 689 + ], + "type": "text", + "content": "[53] Zhenyu Li, Zehui Chen, Ang Li, Liangji Fang, Qinhong Jiang, Xianming Liu, Junjun Jiang, Bolei Zhou, and Hang Zhao. Simipu: Simple 2d image and 3d point cloud unsupervised pre-training for spatial-aware visual representations. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1500-1508, 2022. 6, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 691, + 545, + 713 + ], + "type": "text", + "content": "[54] Zhenyu Li, Zehui Chen, Xianming Liu, and Junjun Jiang. Depthformer: Depthformer: Exploiting long-range corre" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10539" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 72, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 287, + 95 + ], + "type": "text", + "content": "lation and local information for accurate monocular depth estimation. arXiv preprint arXiv:2203.14211, 2022. 2, 6, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 95, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 95, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 53, + 95, + 287, + 149 + ], + "type": "text", + "content": "[55] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 150, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 150, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 53, + 150, + 287, + 205 + ], + "type": "text", + "content": "[56] Jeffrey Yunfan Liu, Yun Chen, Ze Yang, Jingkang Wang, Sivabalan Manivasagam, and Raquel Urtasun. Real-time neural rasterization for large scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8416-8427, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 206, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 206, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 54, + 206, + 287, + 239 + ], + "type": "text", + "content": "[57] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. International Conference on Learning Representations, 2018. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 239, + 287, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 239, + 287, + 272 + ], + "spans": [ + { + "bbox": [ + 54, + 239, + 287, + 272 + ], + "type": "text", + "content": "[58] Fangchang Ma and Sertac Karaman. Sparse-to-dense: Depth prediction from sparse depth samples and a single image. pages 1-8, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 272, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 272, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 272, + 287, + 316 + ], + "type": "text", + "content": "[59] Anagh Malik, Parsa Mirdehghan, Sotiris Nousias, Kiriakos N Kutulakos, and David B Lindell. Transient neural radiance fields for lidar view synthesis and 3d reconstruction. arXiv preprint arXiv:2307.09555, 2023. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 317, + 287, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 317, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 53, + 317, + 287, + 358 + ], + "type": "text", + "content": "[60] N. Mayer, E. Ilg, P. Häusser, P. Fischer, D. Cremers, A. Dosovitskiy, and T. Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In CVPR, 2016. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 361, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 361, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 53, + 361, + 287, + 403 + ], + "type": "text", + "content": "[61] Andreas Meuleman, Yu-Lun Liu, Chen Gao, Jia-Bin Huang, Changil Kim, Min H. Kim, and Johannes Kopf. Progressively optimized local radiance fields for robust view synthesis. In CVPR, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 404, + 287, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 404, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 53, + 404, + 287, + 448 + ], + "type": "text", + "content": "[62] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 3, 5, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 449, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 449, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 53, + 449, + 287, + 514 + ], + "type": "text", + "content": "[63] Ben Mildenhall, Peter Hedman, Ricardo Martin-Brualla, Pratul P Srinivasan, and Jonathan T Barron. Nerf in the dark: High dynamic range view synthesis from noisy raw images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16190-16199, 2022. 1, 6, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 515, + 287, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 515, + 287, + 558 + ], + "spans": [ + { + "bbox": [ + 53, + 515, + 287, + 558 + ], + "type": "text", + "content": "[64] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 559, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 559, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 53, + 559, + 287, + 624 + ], + "type": "text", + "content": "[65] Zak Murez, Tarrence Van As, James Bartolozzi, Ayan Sinha, Vijay Badrinarayanan, and Andrew Rabinovich. Atlas: End-to-end 3d scene reconstruction from posed images. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VII 16, pages 414-431. Springer, 2020. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 625, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 625, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 53, + 625, + 287, + 668 + ], + "type": "text", + "content": "[66] Julian Ost, Issam Laradji, Alejandro Newell, Yuval Bahat, and Felix Heide. Neural point light fields. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 670, + 287, + 712 + ], + "type": "text", + "content": "[67] Jinsun Park, Kyungdon Joo, Zhe Hu, Chi-Kuei Liu, and In So Kweon. Non-local spatial propagation network for depth completion. In Proc. of European Conference on Computer Vision (ECCV), 2020. 2, 7, 8" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 312, + 73, + 545, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 545, + 138 + ], + "type": "text", + "content": "[68] Vivek Pradeep, Christoph Rhemann, Shahram Izadi, Christopher Zach, Michael Bleyer, and Steven Bathiche. Monofusion: Real-time 3d reconstruction of small scenes with a single web camera. In 2013 IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 83-88. IEEE, 2013. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 140, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 140, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 312, + 140, + 545, + 183 + ], + "type": "text", + "content": "[69] Andrea Ramazzina, Mario Bijelic, Stefanie Walz, Alessandro Sanvito, Dominik Scheuble, and Felix Heide. Scatternerf: Seeing through fog with physically-based inverse neural rendering. arXiv preprint arXiv:2305.02103, 2023. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 184, + 547, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 184, + 547, + 227 + ], + "spans": [ + { + "bbox": [ + 312, + 184, + 547, + 227 + ], + "type": "text", + "content": "[70] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12179-12188, 2021. 6, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 228, + 545, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 228, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 312, + 228, + 545, + 270 + ], + "type": "text", + "content": "[71] Konstantinos Rematas, Andrew Liu, Pratul P. Srinivasan, Jonathan T. Barron, Andrea Tagliasacchi, Tom Funkhouser, and Vittorio Ferrari. Urban radiance fields. CVPR, 2022. 1, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 272, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 272, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 312, + 272, + 545, + 316 + ], + "type": "text", + "content": "[72] Gernot Riegler, Ali Osman Ulusoy, Horst Bischof, and Andreas Geiger. Octnetfusion: Learning depth fusion from data. In 2017 International Conference on 3D Vision (3DV), pages 57-66. IEEE, 2017. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 316, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 316, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 312, + 316, + 545, + 380 + ], + "type": "text", + "content": "[73] Barbara Roessle, Jonathan T Barron, Ben Mildenhall, Pratul P Srinivasan, and Matthias Nießner. Dense depth priors for neural radiance fields from sparse input views. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12892-12901, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 382, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 382, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 312, + 382, + 545, + 426 + ], + "type": "text", + "content": "[74] Viktor Rudnev, Mohamed Elgharib, William Smith, Lingjie Liu, Vladislav Golyanik, and Christian Theobalt. Nerf for outdoor scene relighting. In European Conference on Computer Vision, pages 615-631. Springer, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 426, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 426, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 312, + 426, + 545, + 459 + ], + "type": "text", + "content": "[75] Michael Schober, Amit Adam, Omer Yair, Shai Mazor, and Sebastian Nowozin. Dynamic time-of-flight. In CVPR, pages 6109-6118, 2017. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 460, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 460, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 312, + 460, + 545, + 514 + ], + "type": "text", + "content": "[76] Thomas Schöps, Torsten Sattler, Christian Hane, and Marc Pollefeys. 3d modeling on the go: Interactive 3d reconstruction of large-scale scenes on mobile devices. In 2015 International Conference on 3D Vision, pages 291-299. IEEE, 2015. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 514, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 514, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 312, + 514, + 545, + 536 + ], + "type": "text", + "content": "[77] Brent Schwarz. Lidar: Mapping the world in 3D. Nature Photonics, 4(7):429, 2010. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 537, + 545, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 537, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 312, + 537, + 545, + 592 + ], + "type": "text", + "content": "[78] Tixiao Shan, Brendan Englot, Drew Meyers, Wei Wang, Carlo Ratti, and Daniela Rus. Lio-sam: Tightly-coupled lidar inertial odometry via smoothing and mapping. In 2020 IEEE/RSJ international conference on intelligent robots and systems (IROS), pages 5135-5142. IEEE, 2020. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 593, + 547, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 593, + 547, + 647 + ], + "spans": [ + { + "bbox": [ + 312, + 593, + 547, + 647 + ], + "type": "text", + "content": "[79] Jiaming Sun, Yiming Xie, Linghao Chen, Xiaowei Zhou, and Hujun Bao. Neuralrecon: Real-time coherent 3d reconstruction from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15598-15607, 2021. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 647, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 647, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 312, + 647, + 545, + 712 + ], + "type": "text", + "content": "[80] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8248–8258, 2022. 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10540" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 73, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 73, + 287, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 73, + 287, + 137 + ], + "spans": [ + { + "bbox": [ + 53, + 73, + 287, + 137 + ], + "type": "text", + "content": "[81] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, et al. Nerfstudio: A modular framework for neural radiance field development. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-12, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 139, + 288, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 139, + 288, + 183 + ], + "spans": [ + { + "bbox": [ + 53, + 139, + 288, + 183 + ], + "type": "text", + "content": "[82] Jiexiong Tang, John Folkesson, and Patric Jensfelt. Sparse2dense: From direct sparse odometry to dense 3-d reconstruction. IEEE Robotics and Automation Letters, 4 (2):530-537, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 184, + 288, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 184, + 288, + 226 + ], + "spans": [ + { + "bbox": [ + 53, + 184, + 288, + 226 + ], + "type": "text", + "content": "[83] Jie Tang, Fei-Peng Tian, Wei Feng, Jian Li, and Ping Tan. Learning guided convolutional network for depth completion. IEEE Transactions on Image Processing, 30:1116-1129, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 228, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 228, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 228, + 288, + 282 + ], + "type": "text", + "content": "[84] Tang Tao, Longfei Gao, Guangrun Wang, Peng Chen, Dayang Hao, Xiaodan Liang, Mathieu Salzmann, and Kaicheng Yu. Lidar-nerf: Novel lidar view synthesis via neural radiance fields. arXiv preprint arXiv:2304.10406, 2023. 1, 3, 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 283, + 288, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 283, + 288, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 283, + 288, + 327 + ], + "type": "text", + "content": "[85] Fabio Tosi, Alessio Tonioni, Daniele De Gregorio, and Matteo Poggi. Nerf-supervised deep stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 855–866, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 327, + 288, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 327, + 288, + 370 + ], + "spans": [ + { + "bbox": [ + 53, + 327, + 288, + 370 + ], + "type": "text", + "content": "[86] Haithem Turki, Jason Y Zhang, Francesco Ferroni, and Deva Ramanan. Suits: Scalable urban dynamic scenes. In Computer Vision and Pattern Recognition (CVPR), 2023. 1, 3, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 371, + 288, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 371, + 288, + 414 + ], + "spans": [ + { + "bbox": [ + 53, + 371, + 288, + 414 + ], + "type": "text", + "content": "[87] Jonas Uhrig, Nick Schneider, Lukas Schneider, Uwe Franke, Thomas Brox, and Andreas Geiger. Sparsity invariant cnns. In International Conference on 3D Vision (3DV), 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 415, + 288, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 415, + 288, + 480 + ], + "spans": [ + { + "bbox": [ + 53, + 415, + 288, + 480 + ], + "type": "text", + "content": "[88] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481–5490. IEEE, 2022. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 481, + 288, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 481, + 288, + 525 + ], + "spans": [ + { + "bbox": [ + 53, + 481, + 288, + 525 + ], + "type": "text", + "content": "[89] Amanpreet Walia, Stefanie Walz, Mario Bijelic, Fahim Mannan, Frank Julca-Aguilar, Michael Langer, Werner Ritter, and Felix Heide. Gated2gated: Self-supervised depth estimation from gated images. 2022. 2, 3, 4, 6, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 526, + 288, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 526, + 288, + 591 + ], + "spans": [ + { + "bbox": [ + 53, + 526, + 288, + 591 + ], + "type": "text", + "content": "[90] Stefanie Walz, Mario Bijelic, Andrea Ramazzina, Amanpreet Walia, Fahim Mannan, and Felix Heide. Gated stereo: Joint depth estimation from gated and wide-baseline active stereo cues. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13252-13262, 2023. 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 592, + 288, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 592, + 288, + 646 + ], + "spans": [ + { + "bbox": [ + 53, + 592, + 288, + 646 + ], + "type": "text", + "content": "[91] Haoyuan Wang, Xiaogang Xu, Ke Xu, and Rynson WH Lau. Lighting up nerf via unsupervised decomposition and enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12632-12641, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 647, + 288, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 647, + 288, + 691 + ], + "spans": [ + { + "bbox": [ + 53, + 647, + 288, + 691 + ], + "type": "text", + "content": "[92] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 691, + 288, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 691, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 53, + 691, + 288, + 715 + ], + "type": "text", + "content": "[93] Zian Wang, Tianchang Shen, Jun Gao, Shengyu Huang, Jacob Munkberg, Jon Hasselgren, Zan Gojcic, Wenzheng" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 587 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 315, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 545, + 117 + ], + "type": "text", + "content": "Chen, and Sanja Fidler. Neural fields meet explicit geometric representations for inverse rendering of urban scenes. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 118, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 118, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 312, + 118, + 545, + 162 + ], + "type": "text", + "content": "[94] Alex Wong and Stefano Soatto. Unsupervised depth completion with calibrated backprojection layers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12747-12756, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 163, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 163, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 312, + 163, + 545, + 206 + ], + "type": "text", + "content": "[95] Gengshan Yang, Joshua Manela, Michael Happold, and Deva Ramanan. Hierarchical deep stereo matching on high-resolution images. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 207, + 545, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 207, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 312, + 207, + 545, + 261 + ], + "type": "text", + "content": "[96] Zhenfei Yang, Fei Gao, and Shaojie Shen. Real-time monocular dense mapping on aerial robots using visual-inertial fusion. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 4552-4559. IEEE, 2017. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 262, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 262, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 312, + 262, + 545, + 319 + ], + "type": "text", + "content": "[97] Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasun. Unisim: A neural closed-loop sensor simulator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1389-1399, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 319, + 545, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 319, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 312, + 319, + 545, + 373 + ], + "type": "text", + "content": "[98] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoptrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5752-5761, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 374, + 545, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 374, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 312, + 374, + 545, + 418 + ], + "type": "text", + "content": "[99] Junge Zhang, Feihu Zhang, Shaochen Kuang, and Li Zhang. Nerf-lidar: Generating realistic lidar point clouds with neural radiance fields. arXiv preprint arXiv:2304.14811, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 419, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 419, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 419, + 545, + 453 + ], + "type": "text", + "content": "[100] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 453, + 545, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 453, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 307, + 453, + 545, + 509 + ], + "type": "text", + "content": "[101] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Ner-factor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (ToG), 40(6):1-18, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 509, + 545, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 509, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 307, + 509, + 545, + 553 + ], + "type": "text", + "content": "[102] Yongjian Zhang, Longguang Wang, Kunhong Li, Zhiheng Fu, and Yulan Guo. Slfnet: A stereo and lidar fusion network for depth completion. IEEE Robotics and Automation Letters, 7(4):10605-10612, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 554, + 545, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 554, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 307, + 554, + 545, + 587 + ], + "type": "text", + "content": "[103] Tinghui Zhou, Matthew Brown, Noah Snavely, and David G. Lowe. Unsupervised learning of depth and egomotion from video. In CVPR, 2017. 2" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10541" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/GauHuman_ Articulated Gaussian Splatting from Monocular Human Videos/8c2dc361-8082-42c7-a19c-2ac9cac55b4b_content_list.json b/2024/GauHuman_ Articulated Gaussian Splatting from Monocular Human Videos/8c2dc361-8082-42c7-a19c-2ac9cac55b4b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..84442b735ddabe52e72fed1494a3957f02d3d251 --- /dev/null +++ b/2024/GauHuman_ Articulated Gaussian Splatting from Monocular Human Videos/8c2dc361-8082-42c7-a19c-2ac9cac55b4b_content_list.json @@ -0,0 +1,1514 @@ +[ + { + "type": "text", + "text": "GauHuman: Articulated Gaussian Splitting from Monocular Human Videos", + "text_level": 1, + "bbox": [ + 96, + 130, + 872, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shoukang Hu Tao Hu Ziwei Liu S-Lab, Nanyang Technological University", + "bbox": [ + 318, + 181, + 648, + 219 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/617cdbc8d1353f1e11596549b67dfe83def79e76cab9d7154ad1b9d0ba764791.jpg", + "image_caption": [ + "Figure 1. SauHuman models 3D humans with fast training (1~2 minutes) and real-time rendering (~189 FPS) given monocular videos. For illustration purposes, we select one sequence (1024p) from MonoCap [36, 37, 83] and one sequence (1024p) from DNA-Rendering [15]." + ], + "image_footnote": [], + "bbox": [ + 99, + 224, + 872, + 448 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 489, + 313, + 506 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present, GauHuman, a 3D human model with Gaussian Splitting for both fast training (1~2 minutes) and real-time rendering (up to 189 FPS), compared with existing NeRF-based implicit representation modelling frameworks demanding hours of training and seconds of rendering per frame. Specifically, GauHuman encodes Gaussian Splitting in the canonical space and transforms 3D Gaussians from canonical space to posed space with linear blend skinning (LBS), in which effective pose and LBS refinement modules are designed to learn fine details of 3D humans under negligible computational cost. Moreover, to enable fast optimization of GauHuman, we initialize and prune 3D Gaussians with 3D human prior, while splitting/cloning via KL divergence guidance, along with a novel merge operation for further speeding up. Extensive experiments on ZJU_Mocap and MonoCap datasets demonstrate that GauHuman achieves state-of-the-art performance quantitatively and qualitatively with fast training and real-time rendering speed. Notably, without sacrificing rendering quality, GauHuman can fast model the 3D human performer with $\\sim 13k$ 3D Gaussians. Our code is available at https://github.com/skhu101/GauHuman.", + "bbox": [ + 73, + 551, + 473, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 489, + 630, + 505 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Creating high-quality 3D human performers holds extensive utility across various domains, including AR/VR, telepresence, video games, and movie production. Recent methods show 3D human avatar can be learned from sparse-view videos [12, 48, 50, 80, 82, 83, 103, 114, 116, 120] or even a single image [8, 43, 65, 117] with NeRF-based implicit representation [76]. However, these methods typically require expensive time and computational costs for both training and rendering, which hinders their applications in real-world scenarios. For example, it normally takes about 10 GPU hours to learn a 3D human performer with a rendering speed of less than 1 frame per second (FPS).", + "bbox": [ + 496, + 518, + 893, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To speed up 3D human modelling, generalizable Human NeRF methods [17, 28, 42, 60, 132] trade the quality for less training time. They usually fine-tune new human performers with models pre-trained on articulated human data. However, such an inefficient pretrain-and-finetune paradigm would normally take several hours of pre-training to obtain generalizable representations for 3D humans and an additional one hour of finetuning for each human, while achieving low rendering quality due to the incomplete information within limited conditional input images. On the contrary, a recent line of research focuses on learning efficient 3D human representations [30, 49] to improve training speed. For example, instead of learning neural radiance field (NeRF) from", + "bbox": [ + 496, + 703, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20418", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "coordinate-based neural networks, methods like [30, 49] incorporate 3D human modelling with multi-resolution hash encoding (MHE) [78] to accelerate the training. Although these methods can reduce the training time by a magnitude via human prior like part-based human representations, their rendering speed still limits their applications. A critical bottleneck hindering their training and rendering speed arises from the inefficiency of the backward mapping-based ray-cast rendering algorithm employed in these methods, especially at higher resolutions, e.g., 1080p and 4k.", + "bbox": [ + 75, + 90, + 472, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To mitigate this time-consuming dilemma, recent research [56, 131, 135] in point-based rendering can fast render both static and dynamic scenes [72, 119, 125, 126] by splatting a restricted set of 3D points. Motivated by these great successes, it is desirable to explore Gaussian Splatting for real-time rendering of articulated objects such as 3D humans with monocular videos. Yet, this promising avenue comes with two primary barriers: 1) How to properly integrate articulated human information from monocular videos in the Gaussian Splatting framework? 2) How to efficiently optimize Gaussian Splatting for fast training? In particular, one potential solution for 3D human modelling with Gaussian Splatting is to directly utilize LBS transformation to transform 3D Gaussians from canonical space into posed space. However, it is infeasible to directly apply a 3D vertex point-based LBS transformation for 3D Gaussians, with inaccurate LBS weights and pose information derived from human prior models such as SMPL [71]. Meanwhile, current Gaussian Splatting methods utilize either sparse point cloud produced from Structure-from-Motion (SfM) [98, 101] or random point cloud to initialize 3D Gaussians. This initialization is specifically designed for static scenes, rather than 3D human structures, making it unsuitable for effective and fast human modelling. Also, adaptive control of 3D Gaussians proposed in [56] mainly utilizes gradient information of 3D position to split and clone Gaussians for over-reconstruction and under-reconstruction. Despite improving the rendering performance, it trivially generates a large number of 3D Gaussians, which causes difficulty in optimization and consumes a large storage memory (up to $734\\mathrm{MB}$ in a static scene reconstruction). More importantly, the Gaussian Splatting framework contains plenty of redundant closely-positioned Gaussians, which holds a potential for additional acceleration.", + "bbox": [ + 75, + 246, + 472, + 760 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle these two challenges, we propose GauHuman, a 3D human model with articulated Gaussian Splatting representation for both fast training and rendering. Our modelling framework and fast optimization algorithm are detailed as follows. Articulated Gaussian Splatting with integrated human information. Inspired by previous Human NeRF methods, we encode Gaussian Splatting in a canonical space and wrap it to a posed space. Thanks to the Gaussian property, we can transform 3D Gaussians from a canonical space", + "bbox": [ + 75, + 763, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to a posed space via linear blend skinning (LBS). We rotate and translate the 3D position and covariance of each 3D Gaussian with the estimated LBS transformation matrix. However, directly applying multi-layer perceptron (MLP) to estimate LBS weights would be time-consuming and generate low-quality rendering results. To address this, we start with the LBS weights from SMPL [71] and employ an MLP to predict corresponding LBS weight offsets, followed by an effective pose refinement module to rectify SMPL poses for more accurate LBS transformation. For rendering an image with $1024\\mathrm{p}$ , our GauHuman only needs to splat a limited set (e.g., 13k) of 3D Gaussians, in comparison with Human NeRF methods that apply volume rendering to millions of sampled points in the whole 3D space. This significantly reduces the computational burden in both training and rendering processes. Fast Optimization of 3D Gaussian Splatting. 1) 3D Gaussian Initialization. In GauHuman, we utilize the human prior like SMPL for 3D Gaussian initialization, which enables fast optimization of articulated Gaussian Splatting as the initialization encodes human structures. 2) Split/Clone/Merge. Our analysis indicates that the splitting and cloning process overlooks a crucial metric, namely, the distance between 3D Gaussians, during the split and clone process. In light of this, we integrate the Kullback-Leibler (KL) divergence of 3D Gaussians to regulate the split and clone process. A novel merge operator is also proposed to merge 3D Gaussians that are close to each other for further speeding up. 3) Prune. As 3D human has specific structures (human prior), we prune 3D Gaussians that are far away from the human body, e.g., SMPL vertices. With the above algorithm implemented in articulated 3D Gaussian Splatting, we can fast model 3D human performers with $\\sim 13\\mathrm{k}$ 3D Gaussians (3.5MB memory storage) within one or two minutes. Our main contributions are as follows:", + "bbox": [ + 496, + 90, + 893, + 604 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We present, GauHuman, a 3D model with articulated Gaussian Splatting for both fast training (1~2 minutes) and real-time rendering (up to 189 FPS), in which pose and LBS refinement modules are designed to learn fine details of 3D humans.", + "2. To enable fast optimization of articulated 3D Gaussian Splitting, we initialize and prune 3D Gaussians with 3D human prior, while splitting/cloning via KL divergence measures, accompanied by a novel merge operation for further speeding up.", + "3. Experiments on two monocular data sets validate that our GauHuman achieves state-of-the-art novel view synthesis performance with fast training and real-time rendering." + ], + "bbox": [ + 500, + 604, + 890, + 800 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 814, + 640, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Rendering and Radiance Fields. Early novel view synthesis approaches focus on using light fields for densely sampled [32, 62] and unstructured [7] capture. In later years, the advent of Structure-from-Motion (SfM) [101] enables a new", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20419", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "domain by synthesizing novel views from a collection of photos with blending techniques [9, 23, 31, 38, 58, 107]. With the development of deep learning techniques, previous research [24, 38, 91, 108, 133] propose to adopt deep learning techniques with multi-view stereo (MVS) for novel view synthesis. Recently, deep learning techniques have been used to learn implicit neural radiance fields [39, 76, 84, 100] with volume rendering [55, 73]. NeRF [76] is one of the representative work, learning high-quality novel view synthesis results, which has inspired a series of follow-up works [4, 10, 16, 19, 21, 45, 46, 53, 69, 78, 79, 87, 110, 112, 113, 121, 123, 129, 130]. To accelerate training and rendering speed, several methods [11, 25, 29, 64, 67, 78, 86, 105, 106, 128] are proposed, including replacing multi-layer perceptrons (MLPs) with more efficient representations or skipping the empty space via the occupancy grid. In contrast to NeRF, which applies volume rendering to millions of sampled points in the whole 3D space when rendering an image, point-based rendering [2, 33, 34, 59, 61, 92, 93, 131] splats a restricted set of points. This approach provides flexibility and efficiency in reconstructing 3D scenes by optimizing the opacity and position of individual points. Furthermore, point primitives can be represented as circular, elliptic discs, ellipsoids or surfels for high quality results [6, 85, 89, 118, 127, 136]. To achieve real-time rendering of large-scale scenes, 3D Gaussians with splatting techniques are used to represent static scenes[56], dynamic scenes [72, 119, 125, 126], and other tasks [90, 102, 111]. In this work, we further extend Gaussian Splatting to articulated 3D humans with both fast training and real-time rendering.", + "bbox": [ + 76, + 90, + 472, + 560 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Human Modelling. Although 2D human modelling [1, 26, 27, 51, 52, 63, 74, 97] has made substantial progress, it is still challenging to reconstruct high-quality 3D humans for synthesizing free-viewpoint videos. Traditional methods either leverage multi-view stereo methods [35, 99] or depth fusion [18, 22, 104] to reconstruct 3D human geometry, which typically requires dense camera arrays or depth sensors. To alleviate this, previous research [3, 17, 28, 40-42, 60, 94, 95, 132] learn human priors from data sets containing either 3D ground truth human models or multiview image collections, which enables them to synthesize novel views even with a single image. However, due to the limited diversity in pre-training data sets, these methods can not generalize well to 3D humans with complex poses. Another line of works reconstruct 3D human performers with implicit neural representations from sparse-view videos [12, 48, 50, 66, 68, 80, 82, 83, 83, 103, 114, 116, 120] or even a single image [8, 43, 65, 117]. However, these methods usually require hours of training time and seconds of rendering time per frame, restricting their applications to real-world scenarios. To improve the training efficiency, recent research [30, 49] apply multi hashing encoding (MHE)", + "bbox": [ + 76, + 568, + 472, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to reduce the training time to several minutes, while their rendering remains comparatively slow. Moreover, neural volumetric primitives [14, 70, 88] have been proposed for efficient rendering of 3D humans, but they still require hours of training time. In this work, we focus on reconstructing high-quality 3D humans from monocular videos with both fast training and real-time rendering. Our concurrent works [47, 54, 57, 77, 122, 134] learn real-time rendering of 3D humans from either dense-view or monocular videos.", + "bbox": [ + 496, + 90, + 893, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Our Approach", + "text_level": 1, + "bbox": [ + 498, + 241, + 650, + 258 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The goal of this paper is to learn 3D human performers with fast training and rendering from monocular videos. We assume the calibrated camera parameters and the human region masks are known. We also assume the corresponding SMPL parameters $\\beta, \\theta$ are given.", + "bbox": [ + 496, + 266, + 893, + 343 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 498, + 352, + 630, + 369 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SMPL [71] is a parametric human model $M(\\beta, \\theta)$ , which defines $\\beta, \\theta$ to control body shapes and poses. In this work, we apply the Linear Blend Skinning (LBS) algorithm used in SMPL to transform points from a canonical space to a posed space. For example, 3D point $p^c$ in the canonical space is transformed to a posed space defined by pose $\\theta$ as $p^t = \\sum_{k=1}^{K} w_k (G_k(J, \\theta)p^c + b_k(J, \\theta, \\beta))$ , where $J$ includes $K$ joint locations, $G_k(J, \\theta)$ and $b_k(J, \\theta, \\beta)$ are the transformation matrix and translation vector of joint $k$ respectively, $w_k$ is the linear blend weight.", + "bbox": [ + 496, + 375, + 893, + 526 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Gaussian Splatting [56] In comparison with NeRF methods that learn continuous structured volumetric representation for the whole 3D space, Gaussian Splatting represents a 3D scene with a set of unstructured discrete 3D Gaussians. Each 3D Gaussian is defined with a point center $\\pmb{p}$ and a full 3D covariance matrix $\\Sigma$ in a world space, i.e.,", + "bbox": [ + 496, + 527, + 895, + 619 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nG (\\boldsymbol {x}) = \\frac {1}{(2 \\pi) ^ {\\frac {3}{2}} | \\boldsymbol {\\Sigma} | ^ {\\frac {1}{2}}} e ^ {- \\frac {1}{2} (\\boldsymbol {x} - \\boldsymbol {p}) ^ {T} \\boldsymbol {\\Sigma} ^ {- 1} (\\boldsymbol {x} - \\boldsymbol {p})}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 622, + 893, + 655 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the covariance matrix $\\boldsymbol{\\Sigma}$ is decomposed as $\\boldsymbol{\\Sigma} = \\boldsymbol{R}\\boldsymbol{S}\\boldsymbol{S}^T\\boldsymbol{R}^T$ , rotation matrix $\\boldsymbol{R}$ and scaling matrix $\\boldsymbol{S}$ are represented with a quaternion $\\boldsymbol{r} \\in \\mathcal{R}^4$ and a scaling vector $\\boldsymbol{s} \\in \\mathcal{R}^3$ respectively.", + "bbox": [ + 496, + 659, + 893, + 719 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To splot 3D Gaussians, 3D Gaussian Splatting [135, 136] first transforms them from world space to ray coordinate and then projects them to a 2D screen plane. As shown in [135], utilizing a world-to-camera transformation matrix $\\mathbf{W}$ , translation vector $\\mathbf{d}$ , and the Jacobian matrix $\\mathbf{J}$ representing an approximated projective transformation from camera coordinate to ray coordinate, 3D Gaussians in ray coordinate assume the following form:", + "bbox": [ + 496, + 719, + 895, + 840 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nG (\\boldsymbol {x}) = \\frac {1}{(2 \\pi) ^ {\\frac {3}{2}} | \\boldsymbol {\\Sigma} ^ {\\prime} | ^ {\\frac {1}{2}}} e ^ {- \\frac {1}{2} (\\boldsymbol {x} - \\boldsymbol {p} ^ {\\prime}) ^ {T} \\boldsymbol {\\Sigma} ^ {\\prime} ^ {- 1} (\\boldsymbol {x} - \\boldsymbol {p} ^ {\\prime})} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 844, + 893, + 878 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\pmb{p}^{\\prime} = m(\\pmb {W}\\pmb {p} + \\pmb {d})$ and $\\boldsymbol {\\Sigma}' = \\boldsymbol {J}\\boldsymbol {W}\\boldsymbol {\\Sigma}\\boldsymbol{W}^{T}\\boldsymbol{J}^{T}$ are", + "bbox": [ + 498, + 883, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "footer", + "text": "20420", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a8ec7681a1cda701a7a20c550df198e9af56e459ce98838a1af63b1818cd95f5.jpg", + "image_caption": [ + "Figure 2. GauHuman Framework. We first initialize 3D Gaussians' positions $\\pmb{p}^{c}$ from SMPL vertex points. Then we incorporate pose refinement module $\\mathrm{MLP}_{\\Phi_{\\mathrm{pose}}}(\\cdot)$ and LBS weight field $\\mathrm{MLP}_{\\Phi_{\\mathrm{lbs}}}(\\cdot)$ module to learn the LBS transformation to transform 3D Gaussians from canonical space to posed space. During optimization, a tile-based differentiable rasterizer is applied to enable fast rendering. To adaptively control the number of 3D Gaussians, we further propose to use human prior (e.g., SMPL) and KL divergence measure to regulate the split, clone, merge, and prune process." + ], + "image_footnote": [], + "bbox": [ + 109, + 80, + 848, + 321 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the mean vector and covariance matrix of 3D Gaussians in ray coordinate, $m(\\cdot)$ denotes the mapping from camera coordinate to ray coordinate. The projected 2D Gaussians in the 2D screen plane can then be obtained by integrating Eqn. (2) along the projection direction or utilizing a property mentioned by [135].", + "bbox": [ + 75, + 404, + 470, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After mapping 3D Gaussians to a 2D screen space, we count 2D Gaussians that overlap with each pixel and calculate their color $c_{i}$ and density $\\alpha_{i}$ contribution. The colors of each pixel can be obtained by blending $N$ ordered Gaussians:", + "bbox": [ + 75, + 496, + 470, + 556 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {C} = \\sum_ {i \\in N} c _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {i}). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 571, + 470, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During optimization, 3D Gaussians are first initialized from either Structure-from-Motion (SfM) or random sampling. Then adaptive control of Gaussians is employed to improve the rendering quality, which mainly includes three operations, i.e., split, clone, and prune. The split and clone operations are performed on 3D Gaussians with large position gradients. These 3D Gaussians would be 1) split into smaller Gaussians if the magnitude of the scaling matrix is larger than a threshold or 2) cloned if the magnitude of the scaling matrix is smaller than a threshold. While splitting and cloning augments the number of Gaussians, the prune operation selectively eliminates Gaussians with either excessively small opacity $(\\alpha)$ or overly large scaling magnitudes.", + "bbox": [ + 75, + 622, + 472, + 820 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Articulated 3D Gaussian Splatting", + "text_level": 1, + "bbox": [ + 76, + 830, + 380, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recent research can fast render static scenes [56] or dynamic scenes [72, 119, 125, 126] by splatting a restricted set of 3D points. Motivated by these great successes, it is desirable to", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "explore Gaussian Splatting for real-time rendering of articulated objects such as 3D humans with monocular videos. When employing Gaussian Splatting to articulated 3D humans, there are two main challenges, i.e., 1) How to properly integrate articulated human information from monocular videos in the Gaussian Splatting framework? ii). how to efficiently optimize Gaussian Splatting for fast training?", + "bbox": [ + 496, + 404, + 893, + 510 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired by previous Human NeRF research, as shown in Fig. 2, we represent 3D humans with 3D Gaussians in a canonical space and then wrap 3D Gaussians from the canonical space to posed space through forward LBS transformation. Thanks to the Gaussian property, we can transform 3D Gaussians from a canonical space to a posed space via linear blend skinning (LBS). Specifically, we rotate and translate the 3D position and covariance of each 3D Gaussian with the estimated LBS transformation matrix.", + "bbox": [ + 496, + 511, + 893, + 645 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {p} ^ {t} = \\boldsymbol {G} (\\boldsymbol {J} ^ {t}, \\boldsymbol {\\theta} ^ {t}) \\boldsymbol {p} ^ {c} + \\boldsymbol {b} (\\boldsymbol {J} ^ {t}, \\boldsymbol {\\theta} ^ {t}, \\beta^ {t})\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 652, + 812, + 669 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Sigma} ^ {t} = \\boldsymbol {G} (\\boldsymbol {J} ^ {t}, \\boldsymbol {\\theta} ^ {t}) \\boldsymbol {\\Sigma} ^ {c} \\boldsymbol {G} (\\boldsymbol {J} ^ {t}, \\boldsymbol {\\theta} ^ {t}) ^ {T}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 669, + 890, + 690 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pmb{p}^t, \\pmb{\\Sigma}^t, \\pmb{p}^c$ , and $\\pmb{\\Sigma}^c$ are the 3D position vector and covariance matrix in the canonical and posed space respectively. $\\pmb{G}(\\pmb{J}^t, \\pmb{\\theta}^t) = \\sum_{k=1}^{K} w_k \\pmb{G}_k(\\pmb{J}^t, \\pmb{\\theta}^t)$ , $\\pmb{b}(\\pmb{J}^t, \\pmb{\\theta}^t, \\pmb{\\beta}^t) = \\sum_{k=1}^{K} w_k \\pmb{b}_k(\\pmb{J}^t, \\pmb{\\theta}^t, \\pmb{\\beta}^t)$ are the rotation matrix and translation vector. $K$ is the joint number, $\\pmb{G}_k(\\pmb{J}^t, \\pmb{\\theta}^t)$ and $\\pmb{b}_k(\\pmb{J}^t, \\pmb{\\theta}^t, \\pmb{\\beta}^t)$ are the transformation matrix and translation vector of joint $k$ respectively, $w_k$ is the linear blend skinning (LBS) weight we need to estimate.", + "bbox": [ + 496, + 700, + 893, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "LBS Weight Field Multi-layer perceptrons (MLPs) can be used to predict the LBS weight coefficients $w_{k}$ for each 3D Gaussian in the canonical space. However, it would be time-consuming and generate low-quality rendering results. To address this, we start with the LBS weights from SMPL and", + "bbox": [ + 496, + 825, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20421", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "employ an $\\mathrm{MLP}_{\\Phi_{\\mathrm{Ibs}}}(\\cdot)$ to predict corresponding LBS weight offsets. Specifically, for each 3D Gaussian, we add the LBS weight $w_{k}^{\\mathrm{SMPL}}$ of nearest SMPL vertex with the predicted offsets $\\mathrm{MLP}_{\\Phi_{\\mathrm{Ibs}}}(\\gamma (\\pmb{p}^{c}))$ , i.e.,", + "bbox": [ + 75, + 90, + 468, + 154 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw _ {k} = \\frac {e ^ {\\log \\left(w _ {k} ^ {\\mathrm {S M P L}} + 1 0 ^ {- 8}\\right) + \\mathrm {M L P} _ {\\Phi_ {\\mathrm {l b s}}} \\left(\\gamma \\left(\\boldsymbol {p} ^ {c}\\right)\\right) [ k ]}}{\\sum_ {k = 1} ^ {K} e ^ {\\log \\left(w _ {k} ^ {\\mathrm {S M P L}} + 1 0 ^ {- 8}\\right) + \\mathrm {M L P} _ {\\Phi_ {\\mathrm {l b s}}} \\left(\\gamma \\left(\\boldsymbol {p} ^ {c}\\right)\\right) [ k ]}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 157, + 468, + 199 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where we apply sinusoidal positional encoding $\\gamma (\\cdot)$ [76] to the input (3D position $p^c$ ) of $\\mathrm{MLP}_{\\Phi_{\\mathrm{lbs}}(\\cdot)}$ . The idea of learning LBS weight field is also applied in previous human modelling works [5, 13, 20, 44, 75, 81, 96, 109, 115, 124].", + "bbox": [ + 75, + 204, + 468, + 263 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pose Refinement The body poses $\\theta$ estimated from images is often inaccurate, which degrades the rendering quality. To address this, we incorporate a pose correction module $\\mathrm{MLP}_{\\Phi_{\\mathrm{pose}}}(\\cdot)$ to learn relative updates to joint angles:", + "bbox": [ + 75, + 265, + 470, + 325 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\theta} = \\boldsymbol {\\theta} ^ {\\mathrm {S M P L}} \\otimes \\mathrm {M L P} _ {\\Phi_ {\\text {p o s e}}} (\\boldsymbol {\\theta}), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 330, + 468, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\theta^{\\mathrm{SMPL}}$ is the SMPL body pose parameters estimated from images. The body pose refinement is also investigated in [12, 49, 116].", + "bbox": [ + 75, + 354, + 468, + 400 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During inference, thanks to the 3D Gaussian-based representation defined in a canonical space, the learned LBS weight for each 3D Gaussian can be derived from the LBS weight field module in advance, which saves computation time. Meanwhile, the rectified pose parameters can also be saved from the pose refinement module at the end of training. Considering the above, our GauHuman can model high-quality 3D humans with high rendering speed.", + "bbox": [ + 75, + 401, + 468, + 522 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Fast Optimization of GauHuman", + "text_level": 1, + "bbox": [ + 76, + 530, + 366, + 545 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The current optimization scheme of Gaussian Splitting [56] is specifically designed for static scenes. To facilitate fast optimization for articulated 3D humans, we incorporate 3D human prior and KL divergence measures to speed up.", + "bbox": [ + 75, + 551, + 468, + 613 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3D Gaussian Initialization Current Gaussian Splatting methods utilize either Structure-from-Motion (SfM) [98, 101] or random point cloud to initialize 3D Gaussians, which overlooks 3D human structures. In GauHuman, we initialize 3D Gaussians with human prior like SMPL vertex points. The above initialization encodes human structure, thus accelerating optimization. Note that for 3D Gaussian covariance initialization, we follow the default setting in [56].", + "bbox": [ + 75, + 614, + 468, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Split & Clone & Merge & Prune As mentioned in Sec. 3.1, the optimization of Gaussian Splatting relies on the adaptive control of 3D Gaussians to ensure superior rendering quality. To adaptively control the number of Gaussians, three operations, i.e., split, clone, and prune, are performed based on the position gradients and magnitude of the scaling matrix. However, although split and clone operations substantially improve the performance, they trivially generate a substantial quantity of 3D Gaussians. This results in a slowdown of the optimization and a significant consumption of storage memory, reaching up to $734\\mathrm{MB}$ in static scenes. Our", + "bbox": [ + 75, + 734, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "analysis indicates that previous research overlooks a crucial metric, namely, the distance between 3D Gaussians, during the split and clone process. For example, if two Gaussians are close to each other, even if the position gradients are larger than a threshold, they should not be split or cloned as 1) these Gaussians are updating their positions to improve their performance, 2) splitting and cloning these Gaussians have negligible influence on the final performance as they are too close to each other. To regulate the split and clone process, we propose to use Kullback-Leibler (KL) divergence as a measure of the distance of 3D Gaussians, i.e.,", + "bbox": [ + 496, + 90, + 893, + 258 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} K L \\left(G \\left(\\boldsymbol {x} _ {0}\\right) \\mid G \\left(\\boldsymbol {x} _ {1}\\right)\\right) = \\frac {1}{2} \\left(t r \\left(\\boldsymbol {\\Sigma} _ {1} ^ {- 1} \\boldsymbol {\\Sigma} _ {0}\\right) + \\ln \\frac {\\det \\boldsymbol {\\Sigma} _ {1}}{\\det \\boldsymbol {\\Sigma} _ {2}} \\right. \\tag {7} \\\\ + \\left(\\boldsymbol {p} _ {1} - \\boldsymbol {p} _ {0}\\right) ^ {T} \\boldsymbol {\\Sigma} _ {1} ^ {- 1} \\left(\\boldsymbol {p} _ {1} - \\boldsymbol {p} _ {0}\\right) - 3), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 273, + 890, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pmb{p}_0, \\pmb{\\Sigma}_0, \\pmb{p}_1, \\pmb{\\Sigma}_1$ are the position and covariance matrix of two 3D Gaussians $G(\\pmb{x}_0)$ and $G(\\pmb{x}_1)$ .", + "bbox": [ + 496, + 329, + 890, + 359 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After obtaining the KL divergence of nearby 3D Gaussian pairs, we select 3D Gaussians with large KL divergence and positional gradients to perform the split and clone. However, the above requires calculating the KL divergence for each pair of 3D Gaussians, which consumes lots of time during the optimization. To solve this problem, we further adopt two approaches to simplify the whole calculation. 1) For each 3D Gaussian, we first identify its closest 3D Gaussian through the widely adopted k-nearest neighbor (k-NN) algorithm, which assesses the distance between their respective centers. Then we calculate a KL divergence for each pair of nearby 3D Gaussians. The aforementioned simplification effectively reduces the time complexity from $O(U^2)$ to $O(U)$ , where $U$ represents the total number of 3D Gaussians. 2) As the covariance matrix is decomposed into the product of rotation and scaling matrices, we further simplify the computation of matrix inverse and determinant operations in Eqn. (7) with the diagonal and orthogonal property of rotation and scaling matrices. Details are shown in the appendix.", + "bbox": [ + 496, + 364, + 892, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition to split and clone operations, we further propose a novel merge operation, which merges redundant 3D Gaussians with 1) large position gradients, 2) small scaling magnitude, and 3) KL divergence less than 0.1. Specifically, we merge two Gaussians by averaging their positions, opacity, and SH coefficients. Additionally, we scale the covariance (scaling) matrix of the first Gaussian by a factor of 1.25 as the initialization for the new Gaussian. For prune operation, we incorporate 3D human prior like SMPL to prune 3D Gaussians far away from SMPL vertex points.", + "bbox": [ + 496, + 656, + 893, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During optimization, four loss functions are used, i.e.,", + "bbox": [ + 517, + 811, + 875, + 827 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {c o l o r}} + \\lambda_ {1} \\mathcal {L} _ {\\text {m a s k}} + \\lambda_ {2} \\mathcal {L} _ {\\text {S S I M}} + \\lambda_ {3} \\mathcal {L} _ {\\text {L P I P S}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 848, + 890, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ 's are loss weights. Empirically, we set $\\lambda_1 = 0.5$ , $\\lambda_2 = \\lambda_3 = 0.01$ to ensure the same magnitude for each loss.", + "bbox": [ + 500, + 869, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20422", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/463984b8ca0818d64cb8aaf80a325b64b357a877f45cfd96b08d83303b4f59a4.jpg", + "table_caption": [ + "Table 1. Quantitative comparison of our GauHuman and baseline methods on the ZJU_MoCap and MonoCap data sets. We use bold and underlined text to denote the best and second-best results of each metric respectively. Our GauHuman achieves the state-of-the-art PSNR and LPIPS with fast training (1~2 minutes) and real-time rendering (up to 189 FPS) in both data sets. Note that PixelNeRF [129] and NHP [60] are pre-trained for an additional 10 hours. $\\dagger$ denotes the fine-tuning time for generalizable methods. LPIPS\\* = 1000 × LPIPS. Frames per second (FPS) is measured on an RTX 3090. For a fair comparison, we do not conduct test-time optimization of SMPL parameters with images from the test set on InstantAvatar [49]." + ], + "table_footnote": [], + "table_body": "
MethodZJU_MoCapMonoCap
PSNR↑SSIM↑LPIPS*↓TrainFPSPSNR↑SSIM↑LPIPS*↓TrainFPS
PixelNeRF [129]24.710.892121.86\\( 1h^† \\)1.2026.430.96043.98\\( 1h^† \\)0.75
NHP [60]28.250.95564.77\\( 1h^† \\)0.1530.510.98027.14\\( 1h^† \\)0.05
NB [82]29.030.96442.4710h1.4832.360.98616.7010h0.98
AN [81]29.770.96546.8910h1.1131.070.98519.4710h0.31
AS [83]30.380.97537.2310h0.4032.480.98813.1810h0.29
HumanNeRF [116]30.660.96933.3810h0.3032.680.98715.5210h0.08
DVA [88]29.450.95637.741.5h16.532.990.98315.831.5h10.5
InstantNVR [30]31.010.97138.455m2.2032.610.98816.6810m0.75
InstantAvatar [49]29.730.93868.413m4.1530.790.96439.756m2.54
GauHuman(Ours)31.340.96530.511m18933.450.98513.352m154
", + "bbox": [ + 91, + 154, + 874, + 351 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 361, + 209, + 378 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 76, + 386, + 266, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. Experiments are conducted on two widely used human modelling data sets, i.e., ZJU_MoCap [82] and MonoCap [36, 37, 83]. For ZJU_MoCap, same as previous work [30, 116], we select 6 human subjects (377, 386, 387, 392, 393, 394) to conduct experiments. We also adopt the same training and testing setting as [30], i.e., one camera is used for training, while the remaining cameras are used for evaluation. For each subject, we sample 1 frame every 5 frames and collect 100 frames for training. The foreground masks, camera, and SMPL parameters provided by the data set are used for evaluation purposes. MonoCap data set contains four multi-view videos collected by [83] from the DeepCap data set [36] and the DynaCap data set [37]. Similar to ZJU_MoCap, it provides human masks, cameras, and SMPL parameters. We adopt the same camera views as [83] for training and testing. For each subject, we collect 100 frames for training by sampling 1 frame every 5 frames. More details are shown in the appendix.", + "bbox": [ + 75, + 410, + 470, + 680 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison Methods. We compare our GauHuman with two categories of state-of-the-art human modelling methods, i.e., subject-specific optimization based methods [30, 49, 81-83, 88, 116] and generalizable methods [60, 129]. Following [30], we report the average metric values of all selected human subjects. For a fair comparison, we do not conduct test-time optimization of SMPL parameters with images from the test set on InstantAvatar [49]. More details are shown in the appendix.", + "bbox": [ + 75, + 683, + 470, + 819 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Quantitative Results", + "text_level": 1, + "bbox": [ + 76, + 830, + 271, + 847 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tab. 1 summarizes the overall novel view synthesis results of our GauHuman, NB [82], AN [81], AS [83], HumanNeRF [116], DVA [88], InstantNVR [30], InstantAvatar [49],", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "PixelNeRF [129], and NHP [60]. Thanks to our proposed articulated 3D human Gaussian-based representation, we can fast optimize it for around 1 or 2 minutes to achieve photorealistic rendering results, while baseline methods (i.e., NB, AN, AS, and HumanNeRF) take around 10 hours to finish the training and generalizable methods takes about 1 hour for fine-tuning after 10 hours' pre-training. To accelerate the training of 3D human modelling, InstantNVR [30] and InstantAvatar [49] reduce the training to several minutes with multi hashing encoding representation. However, their rendering speed still restricts their applications in real-world applications. When compared with these two state-of-the-art methods for fasting training, our GauHuman achieves slightly better performance with 3 times faster training and 45 times faster rendering speed. Among baseline methods, DVA achieves a fast rendering speed (up to 16.5 FPS) with efficient mixtures of volumetric primitives. However, in contrast to our GauHuman, DVA takes about $90\\mathrm{x}$ more time to converge and $11.45\\mathrm{x}$ more time to render one image.", + "bbox": [ + 496, + 362, + 893, + 650 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Qualitative Results", + "text_level": 1, + "bbox": [ + 500, + 664, + 683, + 680 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Fig. 3, our GauHuman shows better appearance and geometry details than DVA [88], NB [82] and NHP [60]. Although NB [82] and NHP [60] show impressive results with 4-view videos, they do not perform well with monocular videos. NB [82] encodes human structure information with latent codes positioned in SMPL vertex's locations, which fails to capture complex human motions. NHP [60] does not generalize well in monocular videos due to the information missing from input images. Thanks to the LBS weight field and deformation field learned in HumanNeRF [116] and AS [83], they achieve comparable visualization results as our GauHuman for $600\\mathrm{x}$ more time for training and $472\\mathrm{x}$ more time for rendering. InstantNVR [30] and InstantAvatar [49] reduce the training time to several minutes with a multi", + "bbox": [ + 496, + 688, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20423", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/eed4e7ea3430f589fe608b28c4a70010fb51cad4eab29e1ffd345e9e177286f7.jpg", + "image_caption": [ + "Figure 3. Novel view synthesis results produced by our GauHuman and baseline methods on ZJU_MoCap and MonoCap. The bottom lines show the training time and rendering speed of each method on ZJU_Mocap. Zoom in for the best view." + ], + "image_footnote": [], + "bbox": [ + 86, + 65, + 883, + 492 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "hashing encoding representation, but they produce slightly worse visualization results than ours and require $45\\mathrm{x}$ more rendering time. DVA [] improves the rendering speed to 16.5 FPS with efficient mixtures of volumetric primitives. However, it fails to produce high-quality results and produces obvious artifacts on garments and faces, due to incomplete information within conditional input images and a lack of mechanisms to adaptively control the number of voxels.", + "bbox": [ + 75, + 539, + 472, + 660 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 676, + 230, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We perform ablation studies with the sequence of 386 from the ZJU_Mocap data set. To highlight the effectiveness of each proposed component, we project the 3D human bounding box to each camera plane to obtain the bounding box mask and report these metrics based on the masked area.", + "bbox": [ + 75, + 700, + 470, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Pose Refinement and LBS Weight Field. Tab. 2 shows the results of ablating pose refinement and LBS weight field modules. All models are trained in around 1 minute. As shown in Tab. 2, pose refinement and LBS weight field modules significantly improve all metrics with negligible computational cost. The visualization results in Fig. 4 show that these two modules can help erase artifacts and learn fine details like cloth wrinkles. For example, we observe obvious", + "bbox": [ + 75, + 779, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "artifacts in hands and fewer cloth wrinkles without using either one of these two modules.", + "bbox": [ + 498, + 539, + 890, + 566 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5b05c9a560aff5d2eb1add5161fcb17d3abd9947f8a8dbb574420a89b75a7ff1.jpg", + "table_caption": [ + "Table 2. Quantitative Results of ablating pose refinement and LBS weight field modules. $\\mathrm{LPIPS^{*}} = 1000\\times$ LPIPS." + ], + "table_footnote": [], + "table_body": "
PSNR↑SSIM↑LPIPS*↓TrainFPS
Ours (full model)28.080.886103.355s189
Ours w/o pose refine27.940.884108.845s192
Ours w/o LBS field27.870.882108.952s190
", + "bbox": [ + 504, + 595, + 883, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3D Gaussian Initialization. In Gaussian Splitting, previous research [56] indicates that initialization plays an important role in the final results. For static scenes, there are two commonly adopted initialization approaches, i.e., random initialization or initialization with sparse points from Structure-from-Motion (SfM). To incorporate human articulated information into the initialization, we propose to use articulated vertex points from SMPL to initialize our 3D Gaussians in the canonical space. We compare our proposed articulated initialization with random initialization1 in Tab. 3. With the same number of 3D Gaussians at initialization, our proposed articulated initialization produces both better", + "bbox": [ + 496, + 672, + 893, + 854 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "Estimating sparse point cloud from Structure-from-Motion (SfM) requires multi-view inputs, which is infeasible with monocular videos. So we do not compare it with our articulated initialization.", + "bbox": [ + 498, + 862, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20424", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7b42fc5011e011a098790b971ec59fa731997588929a8c3cf3a387886253f1a6.jpg", + "image_caption": [ + "Figure 4. Qualitative results of ablation study on the 386 sequence of ZJU_Mocap data set. \"w/o pose refinement\" and \"w/o LBS weight field\" mean that we do not adopt the pose refinement or LBS weight field module respectively. \"w/ random Init (#Gau 6890)\" and \"w/ random Init (#Gau 30000)\" mean that we use random initialization to initialize 6890 or 30000 3D Gaussians respectively." + ], + "image_footnote": [], + "bbox": [ + 80, + 73, + 888, + 281 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "quantitative and qualitative results than random initialization as shown in Tab. 3 and Fig. 4. Notably, our articulated initialization converges x10 times much faster than the random initialization, which illustrates the strength of encoding human articulated information into the initialization. For random initialization, we need to initialize 30k Gaussians to achieve comparable performance as our proposed articulated initialization, while at cost of 30x times more training time. Table 3. Quantitative Results of ablating 3D Gaussian initialization. #Gau denotes the number of 3D Gaussians at initialization.", + "bbox": [ + 75, + 344, + 472, + 494 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2705ac9f144e3ae48cb47e838d780cf6776a8cfad1da18e94bc53eb2b6446ef2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PSNR↑SSIM↑LPIPS*↓Train
Articulated Init (w/ #Gau 6890)28.080.886103.355s
Random Init (w/ #Gau 6890)27.980.882108.95m
Random Init (w/ #Gau 30k)28.080.884104.330m
", + "bbox": [ + 78, + 496, + 464, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Split/Clone/Merge/Prune. Adaptive control of 3D Gaussians with split, clone, and prune operations substantially improves the rendering quality, but it trivially leads to a large number of 3D Gaussians (up to $100\\mathrm{k}$ ), which slows down the whole optimization process. As mentioned in Sec. 3.3, we propose to use the KL divergence measure to regulate the split and clone process. As shown in Tab. 4, our KL-based split and clone operations lead to $60\\mathrm{x}$ times faster convergence speed when compared with the original split and clone operation proposed in [56]. Furthermore, our novel merge operation merges about $8.5\\%$ of 3D Gaussians without degrading the performance. Articulated prune operation with SMPL prior introduced in Sec. 3.3 also helps prune about $0.5\\mathrm{k}$ 3D Gaussians far away from the 3D human body.", + "bbox": [ + 75, + 571, + 468, + 782 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Quantitative Results of ablating our proposed KL-based split/clone operations and a novel merge operation. #Gau denotes the final number of 3D Gaussians during optimization.", + "bbox": [ + 75, + 785, + 468, + 825 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/652903be0e8dd30354576f20fec32e0182801c9b4d19611ac581677338171a23.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PSNR↑SSIM↑LPIPS*↓#GauTrain
Ours (full model)28.080.886103.31316255s
w/o KL split/clone28.010.885103.1100k1h
w/o KL merge28.080.886103.61437670s
w/o articulated prune28.080.886103.41368757s
", + "bbox": [ + 81, + 827, + 460, + 911 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussion and Conclusion", + "text_level": 1, + "bbox": [ + 500, + 343, + 746, + 359 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To conclude, we propose GauHuman, a 3D human model with Gaussian Splating representation for fast training $(1\\sim 2$ minutes) and real-time rendering (166 FPS) of 3D humans. Specifically, we encode this representation in a canonical space and transform 3D Gaussian from canonical space to posed space with LBS transformation, in which effective pose refinement and LBS weight field modules are also designed to learn fine details of 3D humans. To enable fast optimization, we initialize and prune 3D Gaussians with 3D human prior, while splitting/cloning 3D Gaussians via KI divergence guidance, accompanied by a novel merge operation for further speeding up. Experiments on monocular videos suggest that our GauHuman achieves state-of-the-art performance while maintaining both fast training and real-time rendering speed. In addition, GauHuman can fast learn 3D human performers with $\\sim 13k$ Gaussians.", + "bbox": [ + 496, + 378, + 893, + 622 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. 1). GauHuman is composed of 3D Gaussians and current pipeline does not support extracting 3D meshes. How to extract meshes from 3D Gaussians remains a future work to be investigated. 2). Recovering 3D humans with loose/complex clothing from monocular videos is still a challenging problem to tackle. One direction is to incorporate physical simulation of garments into current pipelines.", + "bbox": [ + 496, + 632, + 893, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment. This study is supported by the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOET2EP20221-0012), NTU NAP, and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). We would like to thank Zhaoxi Chen, Wei Li, Guangcong Wang from NTU, Xurong Xie from ISCAS, Zhiyuan Yu from ZJU, and Zihao Huang from HUST for the discussions on the paper draft.", + "bbox": [ + 496, + 750, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20425", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Badour AlBahar, Jingwan Lu, Jimei Yang, Zhixin Shu, Eli Shechtman, and Jia-Bin Huang. Pose with style: Detail-preserving pose-guided image synthesis with conditional stylegan. ACM Transactions on Graphics (TOG), 40(6): 1-11, 2021. 3", + "[2] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 696-712. Springer, 2020. 3", + "[3] Thiemo Alldieck, Mihai Zanfir, and Cristian Sminchisescu. Photorealistic monocular 3d reconstruction of humans wearing clothing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1506-1515, 2022. 3", + "[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 3", + "[5] Bharat Lal Bhatnagar, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. Loopreg: Self-supervised learning of implicit surface correspondences, pose and shape for 3d human mesh registration. Advances in Neural Information Processing Systems, 33:12909-12922, 2020. 5", + "[6] Mario Botsch, Alexander Hornung, Matthias Zwicker, and Leif Kobbelt. High-quality surface splatting on today's gpus. In Proceedings Eurographics/IEEE VGTC Symposium Point-Based Graphics, 2005., pages 17-141. IEEE, 2005. 3", + "[7] Chris Buehler, Michael Bosse, Leonard McMillan, Steven Gortler, and Michael Cohen. Unstructured lumigraph rendering. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 497-504. 2023. 2", + "[8] Sihun Cha, Kwanggyoon Seo, Amirsaman Ashtari, and Jun-yong Noh. Generating texture for 3d human avatar from a single image using sampling and refinement networks. In Computer Graphics Forum, pages 385-396. Wiley Online Library, 2023. 1, 3", + "[9] Gaurav Chaurasia, Sylvain Duchene, Olga Sorkine-Hornung, and George Drettakis. Depth synthesis and local warps for plausible image-based navigation. ACM Transactions on Graphics (TOG), 32(3):1-12, 2013. 3", + "[10] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. arXiv preprint arXiv:2103.15595, 2021. 3", + "[11] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 3", + "[12] Jianchuan Chen, Ying Zhang, Di Kang, Xuefei Zhe, Linchao Bao, Xu Jia, and Huchuan Lu. Animatable neural radiance fields from monocular rgb videos. arXiv preprint arXiv:2106.13629, 2021. 1, 3, 5", + "[13] Xu Chen, Yufeng Zheng, Michael J Black, Otmar Hilliges, and Andreas Geiger. Snarf: Differentiable forward skin" + ], + "bbox": [ + 86, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ning for animating non-rigid neural implicit shapes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11594-11604, 2021. 5", + "[14] Zhaoxi Chen, Fangzhou Hong, Haiyi Mei, Guangcong Wang, Lei Yang, and Ziwei Liu. Primdiffusion: Volumetric primitives diffusion for 3d human generation. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 3", + "[15] Wei Cheng, Ruixiang Chen, Siming Fan, Wanqi Yin, Keyu Chen, Zhongang Cai, Jingbo Wang, Yang Gao, Zhengming Yu, Zhengyu Lin, et al. Dna-rendering: A diverse neural actor repository for high-fidelity human-centric rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19982-19993, 2023. 1", + "[16] Julian Chibane, Aayush Bansal, Verica Lazova, and Gerard Pons-Moll. Stereo radiance fields (srf): Learning view synthesis for sparse views of novel scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7911-7920, 2021. 3", + "[17] Hongsuk Choi, Gyeongsik Moon, Matthieu Armando, Vincent Leroy, Kyoung Mu Lee, and Grégory Rogez. Mononhr: Monocular neural human renderer. In 2022 International Conference on 3D Vision (3DV), pages 242-251. IEEE, 2022. 1, 3", + "[18] Alvaro Collet, Ming Chuang, Pat Sweeney, Don Gillett, Dennis Evseev, David Calabrese, Hugues Hoppe, Adam Kirk, and Steve Sullivan. High-quality streamable free-viewpoint video. ACM Transactions on Graphics (ToG), 34 (4):1-13, 2015. 3", + "[19] François Darmon, Bénédicte Bascle, Jean-Clement Devaux, Pascal Monasse, and Mathieu Aubry. Improving neural implicit surfaces geometry with patch warping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6260-6269, 2022. 3", + "[20] Boyang Deng, John P Lewis, Timothy Jeruzalski, Gerard Pons-Moll, Geoffrey Hinton, Mohammad Norouzi, and Andrea Tagliasacchi. Nasa neural articulated shape approximation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VII 16, pages 612-628. Springer, 2020. 5", + "[21] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nef: Fewer views and faster training for free. arXiv preprint arXiv:2107.02791, 2021. 3", + "[22] Mingsong Dou, Sameh Khamis, Yury Degtyarev, Philip Davidson, Sean Ryan Fanello, Adarsh Kowdle, Sergio Orts Escolano, Christoph Rhemann, David Kim, Jonathan Taylor, et al. Fusion4d: Real-time performance capture of challenging scenes. ACM Transactions on Graphics (ToG), 35(4): 1-13, 2016. 3", + "[23] Martin Eisemann, Bert De Decker, Marcus Magnor, Philippe Bekaert, Edilson De Aguiar, Naveed Ahmed, Christian Theobalt, and Anita Sellent. Floating textures. In Computer graphics forum, pages 409-418. Wiley Online Library, 2008. 3", + "[24] John Flynn, Ivan Neulander, James Philbin, and Noah Snavely. Deepstereo: Learning to predict new views from the world's imagery. In Proceedings of the IEEE conference" + ], + "bbox": [ + 509, + 93, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20426", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on computer vision and pattern recognition, pages 5515-5524, 2016.3", + "[25] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 3", + "[26] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Chen Qian, Chen Change Loy, Wayne Wu, and Ziwei Liu. Stylegan-human: A data-centric odyssey of human generation. In European Conference on Computer Vision, pages 1-19. Springer, 2022. 3", + "[27] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Wayne Wu, and Ziwei Liu. Unitedhuman: Harnessing multi-source data for high-resolution human generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7301-7311, 2023. 3", + "[28] Xiangjun Gao, Jiaolong Yang, Jongyoo Kim, Sida Peng, Zicheng Liu, and Xin Tong. Mps-nerf: Generalizable 3d human rendering from multiview images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1, 3", + "[29] Stephan J Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14346-14355, 2021. 3", + "[30] Chen Geng, Sida Peng, Zhen Xu, Hujun Bao, and Xiaowei Zhou. Learning neural volumetric representations of dynamic humans in minutes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8759-8770, 2023. 1, 2, 3, 6", + "[31] Michael Goesele, Noah Snavely, Brian Curless, Hugues Hoppe, and Steven M Seitz. Multi-view stereo for community photo collections. In 2007 IEEE 11th International Conference on Computer Vision, pages 1-8. IEEE, 2007. 3", + "[32] Steven J Gortler, Radek Grzeszcuk, Richard Szeliski, and Michael F Cohen. The lumigraph. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 453-464. 2023. 2", + "[33] Markus Gross and Hanspeter Pfister. Point-based graphics. Elsevier, 2011. 3", + "[34] Jeffrey P Grossman and William J Dally. Point sample rendering. In Rendering Techniques' 98: Proceedings of the Eurographics Workshop in Vienna, Austria, June 29—July 1, 1998 9, pages 181–192. Springer, 1998. 3", + "[35] Kaiwen Guo, Peter Lincoln, Philip Davidson, Jay Busch, Xueming Yu, Matt Whalen, Geoff Harvey, Sergio Orts-Escolano, Rohit Pandey, Jason Dourgarian, et al. The relightables: Volumetric performance capture of humans with realistic relighting. ACM Transactions on Graphics (ToG), 38(6):1-19, 2019. 3", + "[36] Marc Habermann, Weipeng Xu, Michael Zollhofer, Gerard Pons-Moll, and Christian Theobalt. Deepcap: Monocular human performance capture using weak supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5052-5063, 2020. 1, 6" + ], + "bbox": [ + 86, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Marc Habermann, Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Real-time deep dynamic characters. ACM Transactions on Graphics (ToG), 40(4):1-16, 2021. 1, 6", + "[38] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering. ACM Transactions on Graphics (ToG), 37(6):1-15, 2018. 3", + "[39] Philipp Henzler, Niloy J Mitra, and Tobias Ritschel. Escaping plato's cave: 3d shape from adversarial rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9984-9993, 2019. 3", + "[40] Fangzhou Hong, Zhaoxi Chen, Yushi Lan, Liang Pan, and Ziwei Liu. Eva3d: Compositional 3d human generation from 2d image collections. arXiv preprint arXiv:2210.04888, 2022.3", + "[41] Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. arXiv preprint arXiv:2205.08535, 2022.", + "[42] Shoukang Hu, Fangzhou Hong, Liang Pan, Haiyi Mei, Lei Yang, and Ziwei Liu. Sherf: Generalizable human nerf from a single image. arXiv preprint arXiv:2303.12791, 2023. 1, 3", + "[43] Yangyi Huang, Hongwei Yi, Yuliang Xiu, Tingting Liao, Jiaxiang Tang, Deng Cai, and Justus Thies. Tech: Text-guided reconstruction of lifelike clothed humans. arXiv preprint arXiv:2308.08545, 2023. 1, 3", + "[44] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. Arch: Animatable reconstruction of clothed humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3093-3102, 2020. 5", + "[45] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5885-5894, 2021. 3", + "[46] Wonbong Jang and Lourdes Agapito. Codenerf: Disentangled neural radiance fields for object categories. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12949-12958, 2021. 3", + "[47] Rohit Jena, Ganesh Subramanian Iyer, Siddharth Choudhary, Brandon Smith, Pratik Chaudhari, and James Gee. Splatarmor: Articulated gaussian splatting for animatable humans from monocular rgb videos. arXiv preprint arXiv:2311.10812, 2023. 3", + "[48] Boyi Jiang, Yang Hong, Hujun Bao, and Juyong Zhang. Self-con: Self reconstruction your digital avatar from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5605-5615, 2022. 1, 3", + "[49] Tianjian Jiang, Xu Chen, Jie Song, and Otmar Hilliges. Instantavatar: Learning avatars from monocular video in 60 seconds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16922-16932, 2023. 1, 2, 3, 5, 6", + "[50] Wei Jiang, Kwang Moo Yi, Golnoosh Samei, Oncel Tuzel, and Anurag Ranjan. Neuman: Neural human radiance field" + ], + "bbox": [ + 509, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20427", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "from a single video. arXiv preprint arXiv:2203.12575, 2022. 1, 3", + "[51] Yuming Jiang, Shuai Yang, Haonan Qiu, Wayne Wu, Chen Change Loy, and Ziwei Liu. Text2human: Text-driven controllable human image generation. ACM Transactions on Graphics (TOG), 41(4):1-11, 2022. 3", + "[52] Yuming Jiang, Shuai Yang, Tong Liang Koh, Wayne Wu, Chen Change Loy, and Ziwei Liu. Text2performer: Text-driven human video generation. arXiv preprint arXiv:2304.08483, 2023. 3", + "[53] Mohammad Mahdi Johari, Yann Lepoittevin, and François Fleuret. Geonerf: Generalizing nerf with geometry priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18365-18375, 2022. 3", + "[54] HyunJun Jung, Nikolas Brasch, Jifei Song, Eduardo Perez-Pellitero, Yiren Zhou, Zhihao Li, Nassir Navab, and Benjamin Busam. Deformable 3d gaussian splatting for animatable human avatars. arXiv preprint arXiv:2312.15059, 2023. 3", + "[55] James T Kajiya and Brian P Von Herzen. Ray tracing volume densities. ACM SIGGRAPH computer graphics, 18(3):165-174, 1984. 3", + "[56] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 42(4):1-14, 2023. 2, 3, 4, 5, 7, 8", + "[57] Muhammed Kocabas, Jen-Hao Rick Chang, James Gabriel, Oncel Tuzel, and Anurag Ranjan. Hugs: Human gaussian splats. arXiv preprint arXiv:2311.17910, 2023. 3", + "[58] Georgios Kopanas, Julien Philip, Thomas Leimkuhler, and George Drettakis. Point-based neural rendering with perview optimization. In Computer Graphics Forum, pages 29-43. Wiley Online Library, 2021. 3", + "[59] Georgios Kopanas, Thomas Leimkuhler, Gilles Rainer, Clément Jambon, and George Drettakis. Neural point catauustics for novel-view synthesis of reflections. ACM Transactions on Graphics (TOG), 41(6):1-15, 2022. 3", + "[60] Youngjoong Kwon, Dahun Kim, Duygu Ceylan, and Henry Fuchs. Neural human performer: Learning generalizable radiance fields for human performance rendering. Advances in Neural Information Processing Systems, 34:24741-24752, 2021. 1, 3, 6", + "[61] Christoph Lassner and Michael Zollhofer. Pulsar: Efficient sphere-based neural rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1440-1449, 2021. 3", + "[62] Marc Levoy and Pat Hanrahan. Light field rendering. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 441-452. 2023. 2", + "[63] Kathleen M Lewis, Srivatsan Varadharajan, and Ira Kemelmacher-Shlizerman. Tryongan: Body-aware try-on via layered interpolation. ACM Transactions on Graphics (TOG), 40(4):1-10, 2021. 3", + "[64] Ruilong Li, Matthew Tancik, and Angjoo Kanazawa. Nerve: A general nef acceleration toolbox. arXiv preprint arXiv:2210.04847, 2022. 3" + ], + "bbox": [ + 86, + 90, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[65] Tingting Liao, Xiaomei Zhang, Yuliang Xiu, Hongwei Yi, Xudong Liu, Guo-Jun Qi, Yong Zhang, Xuan Wang, Xi-angyu Zhu, and Zhen Lei. High-fidelity clothed avatar reconstruction from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8662–8672, 2023. 1, 3", + "[66] Haotong Lin, Sida Peng, Zhen Xu, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Efficient neural radiance fields for interactive free-viewpoint video. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 3", + "[67] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 3", + "[68] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. ACM transactions on graphics (TOG), 40(6):1-16, 2021. 3", + "[69] Yuan Liu, Sida Peng, Lingjie Liu, Qianqian Wang, Peng Wang, Christian Theobalt, Xiaowei Zhou, and Wenping Wang. Neural rays for occlusion-aware image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7824-7833, 2022. 3", + "[70] Stephen Lombardi, Tomas Simon, Gabriel Schwartz, Michael Zollhoefer, Yaser Sheikh, and Jason Saragih. Mixture of volumetric primitives for efficient neural rendering. ACM Transactions on Graphics (ToG), 40(4):1-13, 2021. 3", + "[71] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi-person linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia), 34(6):248:1-248:16, 2015. 2, 3", + "[72] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 2, 3, 4", + "[73] Nelson Max. Optical models for direct volume rendering. IEEE Transactions on Visualization and Computer Graphics, 1(2):99-108, 1995. 3", + "[74] Yifang Men, Yiming Mao, Yuning Jiang, Wei-Ying Ma, and Zhouhui Lian. Controllable person image synthesis with attribute-decomposed gan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5084-5093, 2020. 3", + "[75] Marko Mihajlovic, Yan Zhang, Michael J Black, and Siyu Tang. Leap: Learning articulated occupancy of people. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10461-10471, 2021. 5", + "[76] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 3, 5", + "[77] Arthur Moreau, Jifei Song, Helisa Dhamo, Richard Shaw, Yiren Zhou, and Eduardo Pérez-Pellitero. Human gaussian" + ], + "bbox": [ + 509, + 92, + 893, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "20428", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "splatting: Real-time rendering of animatable avatars. arXiv preprint arXiv:2311.17113, 2023. 3", + "[78] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 2, 3", + "[79] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. arXiv preprint arXiv:2112.00724, 2021. 3", + "[80] Atsuhiro Noguchi, Xiao Sun, Stephen Lin, and Tatsuya Harada. Neural articulated radiance field. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5762-5772, 2021. 1, 3", + "[81] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Animatable neural radiance fields for modeling dynamic human bodies. In ICCV, 2021. 5, 6", + "[82] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9054-9063, 2021. 1, 3, 6", + "[83] Sida Peng, Shangzhan Zhang, Zhen Xu, Chen Geng, Boyi Jiang, Hujun Bao, and Xiaowei Zhou. Animatable neural implicit surfaces for creating avatars from videos. arXiv preprint arXiv:2203.08133, 4(5), 2022. 1, 3, 6", + "[84] Eric Penner and Li Zhang. Soft 3d reconstruction for view synthesis. ACM Transactions on Graphics (TOG), 36(6): 1-11, 2017. 3", + "[85] Hanspeter Pfister, Matthias Zwicker, Jeroen Van Baar, and Markus Gross. Surfels: Surface elements as rendering primitives. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 335-342, 2000. 3", + "[86] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14335-14345, 2021. 3", + "[87] Konstantinos Rematas, Ricardo Martin-Brualla, and Vittorio Ferrari. Sharf: Shape-conditioned radiance fields from a single view. arXiv preprint arXiv:2102.08860, 2021. 3", + "[88] Edoardo Remelli, Timur Bagautdinov, Shunsuke Saito, Chenglei Wu, Tomas Simon, Shih-En Wei, Kaiwen Guo, Zhe Cao, Fabian Prada, Jason Saragih, et al. Drivable volumetric avatars using texel-aligned features. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1–9, 2022. 3, 6", + "[89] Liu Ren, Hanspeter Pfister, and Matthias Zwicker. Object space ewa surface splatting: A hardware accelerated approach to high quality point rendering. In Computer Graphics Forum, pages 461-470. Wiley Online Library, 2002. 3", + "[90] Helge Rhodin, Nadia Robertini, Christian Richardt, Hans-Peter Seidel, and Christian Theobalt. A versatile scene" + ], + "bbox": [ + 86, + 92, + 470, + 898 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "model with differentiable visibility applied to generative pose estimation. In Proceedings of the IEEE International Conference on Computer Vision, pages 765-773, 2015. 3", + "[91] Gernot Riegler and Vladlen Koltun. Free view synthesis. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIX 16, pages 623-640. Springer, 2020. 3", + "[92] Darius Rückert, Linus Franke, and Marc Stamminger. Adop: Approximate differentiable one-pixel point rendering. ACM Transactions on Graphics (ToG), 41(4):1-14, 2022. 3", + "[93] Miguel Sainz and Renato Pajarola. Point-based rendering techniques. Computers & Graphics, 28(6):869-879, 2004. 3", + "[94] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2304-2314, 2019. 3", + "[95] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 84-93, 2020. 3", + "[96] Shunsuke Saito, Jinlong Yang, Qianli Ma, and Michael J Black. Scanimate: Weakly supervised learning of skinned clothed avatar networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2886-2897, 2021. 5", + "[97] Kripasindhu Sarkar, Lingjie Liu, Vladislav Golyanik, and Christian Theobalt. Humangan: A generative model of human images. In 2021 International Conference on 3D Vision (3DV), pages 258-267. IEEE, 2021. 3", + "[98] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 2, 5", + "[99] Johannes L Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 501-518. Springer, 2016. 3", + "[100] Vincent Sitzmann, Justus Thies, Felix Heide, Matthias Nießner, Gordon Wetzstein, and Michael Zollhofer. Deepvoxels: Learning persistent 3d feature embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2437-2446, 2019. 3", + "[101] Noah Snavely, Steven M Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. In ACM siggraph 2006 papers, pages 835-846. 2006. 2, 5", + "[102] Carsten Stoll, Nils Hasler, Juergen Gall, Hans-Peter Seidel, and Christian Theobalt. Fast articulated motion tracking using a sums of gaussians body model. In 2011 International Conference on Computer Vision, pages 951-958. IEEE, 2011. 3", + "[103] Shih-Yang Su, Frank Yu, Michael Zollhöfer, and Helge Rhodin. A-nerf: Articulated neural radiance fields for learning human shape, appearance, and pose. Advances in Neural" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "20429", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Information Processing Systems, 34:12278-12291, 2021. 1, 3", + "[104] Zhuo Su, Lan Xu, Zerong Zheng, Tao Yu, Yebin Liu, and Lu Fang. Robustfusion: Human volumetric capture with data-driven visual cues using a rgbd camera. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16, pages 246-264. Springer, 2020. 3", + "[105] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 3", + "[106] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3d shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11358-11367, 2021. 3", + "[107] Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srinivasan, Edgar Tretschk, Wang Yifan, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lombardi, et al. Advances in neural rendering. In Computer Graphics Forum, pages 703-735. Wiley Online Library, 2022. 3", + "[108] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: Image synthesis using neural textures. Acm Transactions on Graphics (TOG), 38(4):1-12, 2019. 3", + "[109] Garvita Tiwari, Nikolaos Sarafianos, Tony Tung, and Gerard Pons-Moll. Neural-gif: Neural generalized implicit functions for animating people in clothing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11708-11718, 2021. 5", + "[110] Alex Trevithick and Bo Yang. Grf: Learning a general radiance field for 3d representation and rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15182-15192, 2021. 3", + "[111] Angtian Wang, Peng Wang, Jian Sun, Adam Kortylewski, and Alan Yuille. Voge: a differentiable volume renderer using gaussian ellipsoids for analysis-by-synthesis. In The Eleventh International Conference on Learning Representations, 2022. 3", + "[112] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 3", + "[113] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snively, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690-4699, 2021. 3", + "[114] Shaofei Wang, Katja Schwarz, Andreas Geiger, and Siyu Tang. Arah: Animatable volume rendering of articulated human sdfs. In European conference on computer vision, 2022. 1, 3" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[115] Chung-Yi Weng, Brian Curless, and Ira Kemelmacher-Shlizerman. Vid2actor: Free-viewpoint animatable person synthesis from video in the wild. arXiv preprint arXiv:2012.12884, 2020. 5", + "[116] Chung-Yi Weng, Brian Curless, Pratul P. Srinivasan, Jonathan T. Barron, and Ira Kemelmacher-Shlizerman. HumanNeRF: Free-viewpoint rendering of moving people from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16210-16220, 2022. 1, 3, 5, 6", + "[117] Zhenzhen Weng, Zeyu Wang, and Serena Yeung. Zero-)rator: Zero-shot 3d avatar generation from a single image. arXiv preprint arXiv:2305.16411, 2023. 1, 3", + "[118] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. Synsin: End-to-end view synthesis from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7467-7477, 2020. 3", + "[119] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528, 2023. 2, 3, 4", + "[120] Hongyi Xu, Thiemo Alldieck, and Cristian Sminchisescu. H-nerf: Neural radiance fields for rendering and temporal reconstruction of humans in motion. Advances in Neural Information Processing Systems, 34, 2021. 1, 3", + "[121] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5438–5448, 2022. 3", + "[122] Zhen Xu, Sida Peng, Haotong Lin, Guangzhao He, Jiaming Sun, Yujun Shen, Hujun Bao, and Xiaowei Zhou. 4k4d: Real-time 4d view synthesis at 4k resolution. arXiv preprint arXiv:2310.11448, 2023. 3", + "[123] Jiayu Yang, Wei Mao, Jose M Alvarez, and Miaomiao Liu. Cost volume pyramid based depth inference for multi-view stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4877-4886, 2020. 3", + "[124] Ze Yang, Shenlong Wang, Sivabalan Manivasagam, Zeng Huang, Wei-Chiu Ma, Xinchen Yan, Ersin Yumer, and Raquel Urtasun. S3: Neural shape, skeleton, and skinn-ning fields for 3d human modeling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13284–13293, 2021. 5", + "[125] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101, 2023. 2, 3, 4", + "[126] Zeyu Yang, Hongye Yang, Zijie Pan, Xiatian Zhu, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. arXiv preprint arXiv:2310.10642, 2023. 2, 3, 4", + "[127] Wang Yifan, Felice Serena, Shihao Wu, Cengiz Öztireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM Transactions on Graphics (TOG), 38(6):1-14, 2019. 3" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "20430", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[128] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoptrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5752-5761, 2021. 3", + "[129] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelNeRF: Neural radiance fields from one or few images. In CVPR, 2021. 3, 6", + "[130] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 3", + "[131] Qiang Zhang, Seung-Hwan Baek, Szymon Rusinkiewicz, and Felix Heide. Differentiable point-based radiance fields for efficient view synthesis. In SIGGRAPH Asia 2022 Conference Papers, pages 1-12, 2022. 2, 3", + "[132] Fuqiang Zhao, Wei Yang, Jiakai Zhang, Pei Lin, Yingliang Zhang, Jingyi Yu, and Lan Xu. Humannerf: Generalizable neural human radiance field from sparse inputs. arXiv preprint arXiv:2112.02789, 3, 2021. 1, 3", + "[133] Tinghui Zhou, Shubham Tulsiani, Weilun Sun, Jitendra Malik, and Alexei A Efros. View synthesis by appearance flow. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 286-301. Springer, 2016. 3", + "[134] Wojciech Zielonka, Timur Bagautdinov, Shunsuke Saito, Michael Zollhöfer, Justus Thies, and Javier Romero. Drivable 3d gaussian avatars. arXiv preprint arXiv:2311.08581, 2023. 3", + "[135] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Ewa volume splatting. In Proceedings Visualization, 2001. VIS'01., pages 29-538. IEEE, 2001. 2, 3, 4", + "[136] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Surface splatting. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 371-378, 2001. 3" + ], + "bbox": [ + 78, + 90, + 470, + 601 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "20431", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 13 + } +] \ No newline at end of file diff --git a/2024/GauHuman_ Articulated Gaussian Splatting from Monocular Human Videos/8c2dc361-8082-42c7-a19c-2ac9cac55b4b_model.json b/2024/GauHuman_ Articulated Gaussian Splatting from Monocular Human Videos/8c2dc361-8082-42c7-a19c-2ac9cac55b4b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c5f47c353c063f6299608626c891db1a7aeb3db0 --- /dev/null +++ b/2024/GauHuman_ Articulated Gaussian Splatting from Monocular Human Videos/8c2dc361-8082-42c7-a19c-2ac9cac55b4b_model.json @@ -0,0 +1,2978 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.098, + 0.131, + 0.873, + 0.152 + ], + "angle": 0, + "content": "GauHuman: Articulated Gaussian Splitting from Monocular Human Videos" + }, + { + "type": "text", + "bbox": [ + 0.32, + 0.182, + 0.65, + 0.22 + ], + "angle": 0, + "content": "Shoukang Hu Tao Hu Ziwei Liu S-Lab, Nanyang Technological University" + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.226, + 0.874, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.45, + 0.894, + 0.48 + ], + "angle": 0, + "content": "Figure 1. SauHuman models 3D humans with fast training (1~2 minutes) and real-time rendering (~189 FPS) given monocular videos. For illustration purposes, we select one sequence (1024p) from MonoCap [36, 37, 83] and one sequence (1024p) from DNA-Rendering [15]." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.491, + 0.314, + 0.507 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.552, + 0.474, + 0.884 + ], + "angle": 0, + "content": "We present, GauHuman, a 3D human model with Gaussian Splitting for both fast training (1~2 minutes) and real-time rendering (up to 189 FPS), compared with existing NeRF-based implicit representation modelling frameworks demanding hours of training and seconds of rendering per frame. Specifically, GauHuman encodes Gaussian Splitting in the canonical space and transforms 3D Gaussians from canonical space to posed space with linear blend skinning (LBS), in which effective pose and LBS refinement modules are designed to learn fine details of 3D humans under negligible computational cost. Moreover, to enable fast optimization of GauHuman, we initialize and prune 3D Gaussians with 3D human prior, while splitting/cloning via KL divergence guidance, along with a novel merge operation for further speeding up. Extensive experiments on ZJU_Mocap and MonoCap datasets demonstrate that GauHuman achieves state-of-the-art performance quantitatively and qualitatively with fast training and real-time rendering speed. Notably, without sacrificing rendering quality, GauHuman can fast model the 3D human performer with \\(\\sim 13k\\) 3D Gaussians. Our code is available at https://github.com/skhu101/GauHuman." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.491, + 0.631, + 0.506 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.52, + 0.895, + 0.701 + ], + "angle": 0, + "content": "Creating high-quality 3D human performers holds extensive utility across various domains, including AR/VR, telepresence, video games, and movie production. Recent methods show 3D human avatar can be learned from sparse-view videos [12, 48, 50, 80, 82, 83, 103, 114, 116, 120] or even a single image [8, 43, 65, 117] with NeRF-based implicit representation [76]. However, these methods typically require expensive time and computational costs for both training and rendering, which hinders their applications in real-world scenarios. For example, it normally takes about 10 GPU hours to learn a 3D human performer with a rendering speed of less than 1 frame per second (FPS)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.897, + 0.903 + ], + "angle": 0, + "content": "To speed up 3D human modelling, generalizable Human NeRF methods [17, 28, 42, 60, 132] trade the quality for less training time. They usually fine-tune new human performers with models pre-trained on articulated human data. However, such an inefficient pretrain-and-finetune paradigm would normally take several hours of pre-training to obtain generalizable representations for 3D humans and an additional one hour of finetuning for each human, while achieving low rendering quality due to the incomplete information within limited conditional input images. On the contrary, a recent line of research focuses on learning efficient 3D human representations [30, 49] to improve training speed. For example, instead of learning neural radiance field (NeRF) from" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20418" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.243 + ], + "angle": 0, + "content": "coordinate-based neural networks, methods like [30, 49] incorporate 3D human modelling with multi-resolution hash encoding (MHE) [78] to accelerate the training. Although these methods can reduce the training time by a magnitude via human prior like part-based human representations, their rendering speed still limits their applications. A critical bottleneck hindering their training and rendering speed arises from the inefficiency of the backward mapping-based ray-cast rendering algorithm employed in these methods, especially at higher resolutions, e.g., 1080p and 4k." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.247, + 0.473, + 0.761 + ], + "angle": 0, + "content": "To mitigate this time-consuming dilemma, recent research [56, 131, 135] in point-based rendering can fast render both static and dynamic scenes [72, 119, 125, 126] by splatting a restricted set of 3D points. Motivated by these great successes, it is desirable to explore Gaussian Splatting for real-time rendering of articulated objects such as 3D humans with monocular videos. Yet, this promising avenue comes with two primary barriers: 1) How to properly integrate articulated human information from monocular videos in the Gaussian Splatting framework? 2) How to efficiently optimize Gaussian Splatting for fast training? In particular, one potential solution for 3D human modelling with Gaussian Splatting is to directly utilize LBS transformation to transform 3D Gaussians from canonical space into posed space. However, it is infeasible to directly apply a 3D vertex point-based LBS transformation for 3D Gaussians, with inaccurate LBS weights and pose information derived from human prior models such as SMPL [71]. Meanwhile, current Gaussian Splatting methods utilize either sparse point cloud produced from Structure-from-Motion (SfM) [98, 101] or random point cloud to initialize 3D Gaussians. This initialization is specifically designed for static scenes, rather than 3D human structures, making it unsuitable for effective and fast human modelling. Also, adaptive control of 3D Gaussians proposed in [56] mainly utilizes gradient information of 3D position to split and clone Gaussians for over-reconstruction and under-reconstruction. Despite improving the rendering performance, it trivially generates a large number of 3D Gaussians, which causes difficulty in optimization and consumes a large storage memory (up to \\(734\\mathrm{MB}\\) in a static scene reconstruction). More importantly, the Gaussian Splatting framework contains plenty of redundant closely-positioned Gaussians, which holds a potential for additional acceleration." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.473, + 0.903 + ], + "angle": 0, + "content": "To tackle these two challenges, we propose GauHuman, a 3D human model with articulated Gaussian Splatting representation for both fast training and rendering. Our modelling framework and fast optimization algorithm are detailed as follows. Articulated Gaussian Splatting with integrated human information. Inspired by previous Human NeRF methods, we encode Gaussian Splatting in a canonical space and wrap it to a posed space. Thanks to the Gaussian property, we can transform 3D Gaussians from a canonical space" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.605 + ], + "angle": 0, + "content": "to a posed space via linear blend skinning (LBS). We rotate and translate the 3D position and covariance of each 3D Gaussian with the estimated LBS transformation matrix. However, directly applying multi-layer perceptron (MLP) to estimate LBS weights would be time-consuming and generate low-quality rendering results. To address this, we start with the LBS weights from SMPL [71] and employ an MLP to predict corresponding LBS weight offsets, followed by an effective pose refinement module to rectify SMPL poses for more accurate LBS transformation. For rendering an image with \\(1024\\mathrm{p}\\), our GauHuman only needs to splat a limited set (e.g., 13k) of 3D Gaussians, in comparison with Human NeRF methods that apply volume rendering to millions of sampled points in the whole 3D space. This significantly reduces the computational burden in both training and rendering processes. Fast Optimization of 3D Gaussian Splatting. 1) 3D Gaussian Initialization. In GauHuman, we utilize the human prior like SMPL for 3D Gaussian initialization, which enables fast optimization of articulated Gaussian Splatting as the initialization encodes human structures. 2) Split/Clone/Merge. Our analysis indicates that the splitting and cloning process overlooks a crucial metric, namely, the distance between 3D Gaussians, during the split and clone process. In light of this, we integrate the Kullback-Leibler (KL) divergence of 3D Gaussians to regulate the split and clone process. A novel merge operator is also proposed to merge 3D Gaussians that are close to each other for further speeding up. 3) Prune. As 3D human has specific structures (human prior), we prune 3D Gaussians that are far away from the human body, e.g., SMPL vertices. With the above algorithm implemented in articulated 3D Gaussian Splatting, we can fast model 3D human performers with \\(\\sim 13\\mathrm{k}\\) 3D Gaussians (3.5MB memory storage) within one or two minutes. Our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.606, + 0.892, + 0.679 + ], + "angle": 0, + "content": "1. We present, GauHuman, a 3D model with articulated Gaussian Splatting for both fast training (1~2 minutes) and real-time rendering (up to 189 FPS), in which pose and LBS refinement modules are designed to learn fine details of 3D humans." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.681, + 0.892, + 0.756 + ], + "angle": 0, + "content": "2. To enable fast optimization of articulated 3D Gaussian Splitting, we initialize and prune 3D Gaussians with 3D human prior, while splitting/cloning via KL divergence measures, accompanied by a novel merge operation for further speeding up." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.757, + 0.892, + 0.801 + ], + "angle": 0, + "content": "3. Experiments on two monocular data sets validate that our GauHuman achieves state-of-the-art novel view synthesis performance with fast training and real-time rendering." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.606, + 0.892, + 0.801 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.815, + 0.642, + 0.831 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Rendering and Radiance Fields. Early novel view synthesis approaches focus on using light fields for densely sampled [32, 62] and unstructured [7] capture. In later years, the advent of Structure-from-Motion (SfM) [101] enables a new" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20419" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.091, + 0.473, + 0.561 + ], + "angle": 0, + "content": "domain by synthesizing novel views from a collection of photos with blending techniques [9, 23, 31, 38, 58, 107]. With the development of deep learning techniques, previous research [24, 38, 91, 108, 133] propose to adopt deep learning techniques with multi-view stereo (MVS) for novel view synthesis. Recently, deep learning techniques have been used to learn implicit neural radiance fields [39, 76, 84, 100] with volume rendering [55, 73]. NeRF [76] is one of the representative work, learning high-quality novel view synthesis results, which has inspired a series of follow-up works [4, 10, 16, 19, 21, 45, 46, 53, 69, 78, 79, 87, 110, 112, 113, 121, 123, 129, 130]. To accelerate training and rendering speed, several methods [11, 25, 29, 64, 67, 78, 86, 105, 106, 128] are proposed, including replacing multi-layer perceptrons (MLPs) with more efficient representations or skipping the empty space via the occupancy grid. In contrast to NeRF, which applies volume rendering to millions of sampled points in the whole 3D space when rendering an image, point-based rendering [2, 33, 34, 59, 61, 92, 93, 131] splats a restricted set of points. This approach provides flexibility and efficiency in reconstructing 3D scenes by optimizing the opacity and position of individual points. Furthermore, point primitives can be represented as circular, elliptic discs, ellipsoids or surfels for high quality results [6, 85, 89, 118, 127, 136]. To achieve real-time rendering of large-scale scenes, 3D Gaussians with splatting techniques are used to represent static scenes[56], dynamic scenes [72, 119, 125, 126], and other tasks [90, 102, 111]. In this work, we further extend Gaussian Splatting to articulated 3D humans with both fast training and real-time rendering." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.569, + 0.473, + 0.902 + ], + "angle": 0, + "content": "3D Human Modelling. Although 2D human modelling [1, 26, 27, 51, 52, 63, 74, 97] has made substantial progress, it is still challenging to reconstruct high-quality 3D humans for synthesizing free-viewpoint videos. Traditional methods either leverage multi-view stereo methods [35, 99] or depth fusion [18, 22, 104] to reconstruct 3D human geometry, which typically requires dense camera arrays or depth sensors. To alleviate this, previous research [3, 17, 28, 40-42, 60, 94, 95, 132] learn human priors from data sets containing either 3D ground truth human models or multiview image collections, which enables them to synthesize novel views even with a single image. However, due to the limited diversity in pre-training data sets, these methods can not generalize well to 3D humans with complex poses. Another line of works reconstruct 3D human performers with implicit neural representations from sparse-view videos [12, 48, 50, 66, 68, 80, 82, 83, 83, 103, 114, 116, 120] or even a single image [8, 43, 65, 117]. However, these methods usually require hours of training time and seconds of rendering time per frame, restricting their applications to real-world scenarios. To improve the training efficiency, recent research [30, 49] apply multi hashing encoding (MHE)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.228 + ], + "angle": 0, + "content": "to reduce the training time to several minutes, while their rendering remains comparatively slow. Moreover, neural volumetric primitives [14, 70, 88] have been proposed for efficient rendering of 3D humans, but they still require hours of training time. In this work, we focus on reconstructing high-quality 3D humans from monocular videos with both fast training and real-time rendering. Our concurrent works [47, 54, 57, 77, 122, 134] learn real-time rendering of 3D humans from either dense-view or monocular videos." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.242, + 0.651, + 0.26 + ], + "angle": 0, + "content": "3. Our Approach" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.267, + 0.895, + 0.344 + ], + "angle": 0, + "content": "The goal of this paper is to learn 3D human performers with fast training and rendering from monocular videos. We assume the calibrated camera parameters and the human region masks are known. We also assume the corresponding SMPL parameters \\(\\beta, \\theta\\) are given." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.353, + 0.631, + 0.37 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.376, + 0.895, + 0.527 + ], + "angle": 0, + "content": "SMPL [71] is a parametric human model \\( M(\\beta, \\theta) \\), which defines \\( \\beta, \\theta \\) to control body shapes and poses. In this work, we apply the Linear Blend Skinning (LBS) algorithm used in SMPL to transform points from a canonical space to a posed space. For example, 3D point \\( p^c \\) in the canonical space is transformed to a posed space defined by pose \\( \\theta \\) as \\( p^t = \\sum_{k=1}^{K} w_k (G_k(J, \\theta)p^c + b_k(J, \\theta, \\beta)) \\), where \\( J \\) includes \\( K \\) joint locations, \\( G_k(J, \\theta) \\) and \\( b_k(J, \\theta, \\beta) \\) are the transformation matrix and translation vector of joint \\( k \\) respectively, \\( w_k \\) is the linear blend weight." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.528, + 0.897, + 0.62 + ], + "angle": 0, + "content": "3D Gaussian Splatting [56] In comparison with NeRF methods that learn continuous structured volumetric representation for the whole 3D space, Gaussian Splatting represents a 3D scene with a set of unstructured discrete 3D Gaussians. Each 3D Gaussian is defined with a point center \\( \\pmb{p} \\) and a full 3D covariance matrix \\( \\Sigma \\) in a world space, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.623, + 0.895, + 0.656 + ], + "angle": 0, + "content": "\\[\nG (\\boldsymbol {x}) = \\frac {1}{(2 \\pi) ^ {\\frac {3}{2}} | \\boldsymbol {\\Sigma} | ^ {\\frac {1}{2}}} e ^ {- \\frac {1}{2} (\\boldsymbol {x} - \\boldsymbol {p}) ^ {T} \\boldsymbol {\\Sigma} ^ {- 1} (\\boldsymbol {x} - \\boldsymbol {p})}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.895, + 0.72 + ], + "angle": 0, + "content": "where the covariance matrix \\(\\boldsymbol{\\Sigma}\\) is decomposed as \\(\\boldsymbol{\\Sigma} = \\boldsymbol{R}\\boldsymbol{S}\\boldsymbol{S}^T\\boldsymbol{R}^T\\), rotation matrix \\(\\boldsymbol{R}\\) and scaling matrix \\(\\boldsymbol{S}\\) are represented with a quaternion \\(\\boldsymbol{r} \\in \\mathcal{R}^4\\) and a scaling vector \\(\\boldsymbol{s} \\in \\mathcal{R}^3\\) respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.897, + 0.841 + ], + "angle": 0, + "content": "To splot 3D Gaussians, 3D Gaussian Splatting [135, 136] first transforms them from world space to ray coordinate and then projects them to a 2D screen plane. As shown in [135], utilizing a world-to-camera transformation matrix \\( \\mathbf{W} \\), translation vector \\( \\mathbf{d} \\), and the Jacobian matrix \\( \\mathbf{J} \\) representing an approximated projective transformation from camera coordinate to ray coordinate, 3D Gaussians in ray coordinate assume the following form:" + }, + { + "type": "equation", + "bbox": [ + 0.546, + 0.845, + 0.895, + 0.879 + ], + "angle": 0, + "content": "\\[\nG (\\boldsymbol {x}) = \\frac {1}{(2 \\pi) ^ {\\frac {3}{2}} | \\boldsymbol {\\Sigma} ^ {\\prime} | ^ {\\frac {1}{2}}} e ^ {- \\frac {1}{2} (\\boldsymbol {x} - \\boldsymbol {p} ^ {\\prime}) ^ {T} \\boldsymbol {\\Sigma} ^ {\\prime} ^ {- 1} (\\boldsymbol {x} - \\boldsymbol {p} ^ {\\prime})} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.884, + 0.895, + 0.903 + ], + "angle": 0, + "content": "where \\(\\pmb{p}^{\\prime} = m(\\pmb {W}\\pmb {p} + \\pmb {d})\\) and \\(\\boldsymbol {\\Sigma}' = \\boldsymbol {J}\\boldsymbol {W}\\boldsymbol {\\Sigma}\\boldsymbol{W}^{T}\\boldsymbol{J}^{T}\\) are" + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20420" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.081, + 0.849, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.323, + 0.896, + 0.395 + ], + "angle": 0, + "content": "Figure 2. GauHuman Framework. We first initialize 3D Gaussians' positions \\(\\pmb{p}^{c}\\) from SMPL vertex points. Then we incorporate pose refinement module \\(\\mathrm{MLP}_{\\Phi_{\\mathrm{pose}}}(\\cdot)\\) and LBS weight field \\(\\mathrm{MLP}_{\\Phi_{\\mathrm{lbs}}}(\\cdot)\\) module to learn the LBS transformation to transform 3D Gaussians from canonical space to posed space. During optimization, a tile-based differentiable rasterizer is applied to enable fast rendering. To adaptively control the number of 3D Gaussians, we further propose to use human prior (e.g., SMPL) and KL divergence measure to regulate the split, clone, merge, and prune process." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.405, + 0.471, + 0.496 + ], + "angle": 0, + "content": "the mean vector and covariance matrix of 3D Gaussians in ray coordinate, \\( m(\\cdot) \\) denotes the mapping from camera coordinate to ray coordinate. The projected 2D Gaussians in the 2D screen plane can then be obtained by integrating Eqn. (2) along the projection direction or utilizing a property mentioned by [135]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.497, + 0.472, + 0.558 + ], + "angle": 0, + "content": "After mapping 3D Gaussians to a 2D screen space, we count 2D Gaussians that overlap with each pixel and calculate their color \\( c_{i} \\) and density \\( \\alpha_{i} \\) contribution. The colors of each pixel can be obtained by blending \\( N \\) ordered Gaussians:" + }, + { + "type": "equation", + "bbox": [ + 0.181, + 0.572, + 0.472, + 0.613 + ], + "angle": 0, + "content": "\\[\n\\hat {C} = \\sum_ {i \\in N} c _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {i}). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.623, + 0.473, + 0.821 + ], + "angle": 0, + "content": "During optimization, 3D Gaussians are first initialized from either Structure-from-Motion (SfM) or random sampling. Then adaptive control of Gaussians is employed to improve the rendering quality, which mainly includes three operations, i.e., split, clone, and prune. The split and clone operations are performed on 3D Gaussians with large position gradients. These 3D Gaussians would be 1) split into smaller Gaussians if the magnitude of the scaling matrix is larger than a threshold or 2) cloned if the magnitude of the scaling matrix is smaller than a threshold. While splitting and cloning augments the number of Gaussians, the prune operation selectively eliminates Gaussians with either excessively small opacity \\((\\alpha)\\) or overly large scaling magnitudes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.381, + 0.849 + ], + "angle": 0, + "content": "3.2. Articulated 3D Gaussian Splatting" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Recent research can fast render static scenes [56] or dynamic scenes [72, 119, 125, 126] by splatting a restricted set of 3D points. Motivated by these great successes, it is desirable to" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.405, + 0.895, + 0.511 + ], + "angle": 0, + "content": "explore Gaussian Splatting for real-time rendering of articulated objects such as 3D humans with monocular videos. When employing Gaussian Splatting to articulated 3D humans, there are two main challenges, i.e., 1) How to properly integrate articulated human information from monocular videos in the Gaussian Splatting framework? ii). how to efficiently optimize Gaussian Splatting for fast training?" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.512, + 0.895, + 0.646 + ], + "angle": 0, + "content": "Inspired by previous Human NeRF research, as shown in Fig. 2, we represent 3D humans with 3D Gaussians in a canonical space and then wrap 3D Gaussians from the canonical space to posed space through forward LBS transformation. Thanks to the Gaussian property, we can transform 3D Gaussians from a canonical space to a posed space via linear blend skinning (LBS). Specifically, we rotate and translate the 3D position and covariance of each 3D Gaussian with the estimated LBS transformation matrix." + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.653, + 0.813, + 0.67 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {p} ^ {t} = \\boldsymbol {G} (\\boldsymbol {J} ^ {t}, \\boldsymbol {\\theta} ^ {t}) \\boldsymbol {p} ^ {c} + \\boldsymbol {b} (\\boldsymbol {J} ^ {t}, \\boldsymbol {\\theta} ^ {t}, \\beta^ {t})\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.67, + 0.892, + 0.691 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Sigma} ^ {t} = \\boldsymbol {G} (\\boldsymbol {J} ^ {t}, \\boldsymbol {\\theta} ^ {t}) \\boldsymbol {\\Sigma} ^ {c} \\boldsymbol {G} (\\boldsymbol {J} ^ {t}, \\boldsymbol {\\theta} ^ {t}) ^ {T}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.702, + 0.895, + 0.825 + ], + "angle": 0, + "content": "where \\(\\pmb{p}^t, \\pmb{\\Sigma}^t, \\pmb{p}^c\\), and \\(\\pmb{\\Sigma}^c\\) are the 3D position vector and covariance matrix in the canonical and posed space respectively. \\(\\pmb{G}(\\pmb{J}^t, \\pmb{\\theta}^t) = \\sum_{k=1}^{K} w_k \\pmb{G}_k(\\pmb{J}^t, \\pmb{\\theta}^t)\\), \\(\\pmb{b}(\\pmb{J}^t, \\pmb{\\theta}^t, \\pmb{\\beta}^t) = \\sum_{k=1}^{K} w_k \\pmb{b}_k(\\pmb{J}^t, \\pmb{\\theta}^t, \\pmb{\\beta}^t)\\) are the rotation matrix and translation vector. \\(K\\) is the joint number, \\(\\pmb{G}_k(\\pmb{J}^t, \\pmb{\\theta}^t)\\) and \\(\\pmb{b}_k(\\pmb{J}^t, \\pmb{\\theta}^t, \\pmb{\\beta}^t)\\) are the transformation matrix and translation vector of joint \\(k\\) respectively, \\(w_k\\) is the linear blend skinning (LBS) weight we need to estimate." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.895, + 0.901 + ], + "angle": 0, + "content": "LBS Weight Field Multi-layer perceptrons (MLPs) can be used to predict the LBS weight coefficients \\( w_{k} \\) for each 3D Gaussian in the canonical space. However, it would be time-consuming and generate low-quality rendering results. To address this, we start with the LBS weights from SMPL and" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "20421" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.155 + ], + "angle": 0, + "content": "employ an \\(\\mathrm{MLP}_{\\Phi_{\\mathrm{Ibs}}}(\\cdot)\\) to predict corresponding LBS weight offsets. Specifically, for each 3D Gaussian, we add the LBS weight \\(w_{k}^{\\mathrm{SMPL}}\\) of nearest SMPL vertex with the predicted offsets \\(\\mathrm{MLP}_{\\Phi_{\\mathrm{Ibs}}}(\\gamma (\\pmb{p}^{c}))\\), i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.159, + 0.47, + 0.2 + ], + "angle": 0, + "content": "\\[\nw _ {k} = \\frac {e ^ {\\log \\left(w _ {k} ^ {\\mathrm {S M P L}} + 1 0 ^ {- 8}\\right) + \\mathrm {M L P} _ {\\Phi_ {\\mathrm {l b s}}} \\left(\\gamma \\left(\\boldsymbol {p} ^ {c}\\right)\\right) [ k ]}}{\\sum_ {k = 1} ^ {K} e ^ {\\log \\left(w _ {k} ^ {\\mathrm {S M P L}} + 1 0 ^ {- 8}\\right) + \\mathrm {M L P} _ {\\Phi_ {\\mathrm {l b s}}} \\left(\\gamma \\left(\\boldsymbol {p} ^ {c}\\right)\\right) [ k ]}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.205, + 0.47, + 0.265 + ], + "angle": 0, + "content": "where we apply sinusoidal positional encoding \\(\\gamma (\\cdot)\\) [76] to the input (3D position \\(p^c\\)) of \\(\\mathrm{MLP}_{\\Phi_{\\mathrm{lbs}}(\\cdot)}\\). The idea of learning LBS weight field is also applied in previous human modelling works [5, 13, 20, 44, 75, 81, 96, 109, 115, 124]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.266, + 0.471, + 0.327 + ], + "angle": 0, + "content": "Pose Refinement The body poses \\(\\theta\\) estimated from images is often inaccurate, which degrades the rendering quality. To address this, we incorporate a pose correction module \\(\\mathrm{MLP}_{\\Phi_{\\mathrm{pose}}}(\\cdot)\\) to learn relative updates to joint angles:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.332, + 0.47, + 0.351 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\theta} = \\boldsymbol {\\theta} ^ {\\mathrm {S M P L}} \\otimes \\mathrm {M L P} _ {\\Phi_ {\\text {p o s e}}} (\\boldsymbol {\\theta}), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.356, + 0.469, + 0.401 + ], + "angle": 0, + "content": "where \\(\\theta^{\\mathrm{SMPL}}\\) is the SMPL body pose parameters estimated from images. The body pose refinement is also investigated in [12, 49, 116]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.402, + 0.47, + 0.523 + ], + "angle": 0, + "content": "During inference, thanks to the 3D Gaussian-based representation defined in a canonical space, the learned LBS weight for each 3D Gaussian can be derived from the LBS weight field module in advance, which saves computation time. Meanwhile, the rectified pose parameters can also be saved from the pose refinement module at the end of training. Considering the above, our GauHuman can model high-quality 3D humans with high rendering speed." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.531, + 0.367, + 0.546 + ], + "angle": 0, + "content": "3.3. Fast Optimization of GauHuman" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.469, + 0.614 + ], + "angle": 0, + "content": "The current optimization scheme of Gaussian Splitting [56] is specifically designed for static scenes. To facilitate fast optimization for articulated 3D humans, we incorporate 3D human prior and KL divergence measures to speed up." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.615, + 0.47, + 0.734 + ], + "angle": 0, + "content": "3D Gaussian Initialization Current Gaussian Splatting methods utilize either Structure-from-Motion (SfM) [98, 101] or random point cloud to initialize 3D Gaussians, which overlooks 3D human structures. In GauHuman, we initialize 3D Gaussians with human prior like SMPL vertex points. The above initialization encodes human structure, thus accelerating optimization. Note that for 3D Gaussian covariance initialization, we follow the default setting in [56]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Split & Clone & Merge & Prune As mentioned in Sec. 3.1, the optimization of Gaussian Splatting relies on the adaptive control of 3D Gaussians to ensure superior rendering quality. To adaptively control the number of Gaussians, three operations, i.e., split, clone, and prune, are performed based on the position gradients and magnitude of the scaling matrix. However, although split and clone operations substantially improve the performance, they trivially generate a substantial quantity of 3D Gaussians. This results in a slowdown of the optimization and a significant consumption of storage memory, reaching up to \\(734\\mathrm{MB}\\) in static scenes. Our" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.259 + ], + "angle": 0, + "content": "analysis indicates that previous research overlooks a crucial metric, namely, the distance between 3D Gaussians, during the split and clone process. For example, if two Gaussians are close to each other, even if the position gradients are larger than a threshold, they should not be split or cloned as 1) these Gaussians are updating their positions to improve their performance, 2) splitting and cloning these Gaussians have negligible influence on the final performance as they are too close to each other. To regulate the split and clone process, we propose to use Kullback-Leibler (KL) divergence as a measure of the distance of 3D Gaussians, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.274, + 0.892, + 0.326 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} K L \\left(G \\left(\\boldsymbol {x} _ {0}\\right) \\mid G \\left(\\boldsymbol {x} _ {1}\\right)\\right) = \\frac {1}{2} \\left(t r \\left(\\boldsymbol {\\Sigma} _ {1} ^ {- 1} \\boldsymbol {\\Sigma} _ {0}\\right) + \\ln \\frac {\\det \\boldsymbol {\\Sigma} _ {1}}{\\det \\boldsymbol {\\Sigma} _ {2}} \\right. \\tag {7} \\\\ + \\left(\\boldsymbol {p} _ {1} - \\boldsymbol {p} _ {0}\\right) ^ {T} \\boldsymbol {\\Sigma} _ {1} ^ {- 1} \\left(\\boldsymbol {p} _ {1} - \\boldsymbol {p} _ {0}\\right) - 3), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.33, + 0.892, + 0.361 + ], + "angle": 0, + "content": "where \\(\\pmb{p}_0, \\pmb{\\Sigma}_0, \\pmb{p}_1, \\pmb{\\Sigma}_1\\) are the position and covariance matrix of two 3D Gaussians \\(G(\\pmb{x}_0)\\) and \\(G(\\pmb{x}_1)\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.365, + 0.893, + 0.652 + ], + "angle": 0, + "content": "After obtaining the KL divergence of nearby 3D Gaussian pairs, we select 3D Gaussians with large KL divergence and positional gradients to perform the split and clone. However, the above requires calculating the KL divergence for each pair of 3D Gaussians, which consumes lots of time during the optimization. To solve this problem, we further adopt two approaches to simplify the whole calculation. 1) For each 3D Gaussian, we first identify its closest 3D Gaussian through the widely adopted k-nearest neighbor (k-NN) algorithm, which assesses the distance between their respective centers. Then we calculate a KL divergence for each pair of nearby 3D Gaussians. The aforementioned simplification effectively reduces the time complexity from \\( O(U^2) \\) to \\( O(U) \\), where \\( U \\) represents the total number of 3D Gaussians. 2) As the covariance matrix is decomposed into the product of rotation and scaling matrices, we further simplify the computation of matrix inverse and determinant operations in Eqn. (7) with the diagonal and orthogonal property of rotation and scaling matrices. Details are shown in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.657, + 0.894, + 0.809 + ], + "angle": 0, + "content": "In addition to split and clone operations, we further propose a novel merge operation, which merges redundant 3D Gaussians with 1) large position gradients, 2) small scaling magnitude, and 3) KL divergence less than 0.1. Specifically, we merge two Gaussians by averaging their positions, opacity, and SH coefficients. Additionally, we scale the covariance (scaling) matrix of the first Gaussian by a factor of 1.25 as the initialization for the new Gaussian. For prune operation, we incorporate 3D human prior like SMPL to prune 3D Gaussians far away from SMPL vertex points." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.813, + 0.876, + 0.828 + ], + "angle": 0, + "content": "During optimization, four loss functions are used, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.849, + 0.892, + 0.866 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {c o l o r}} + \\lambda_ {1} \\mathcal {L} _ {\\text {m a s k}} + \\lambda_ {2} \\mathcal {L} _ {\\text {S S I M}} + \\lambda_ {3} \\mathcal {L} _ {\\text {L P I P S}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.894, + 0.901 + ], + "angle": 0, + "content": "where \\(\\lambda\\)'s are loss weights. Empirically, we set \\(\\lambda_1 = 0.5\\), \\(\\lambda_2 = \\lambda_3 = 0.01\\) to ensure the same magnitude for each loss." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20422" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.071, + 0.895, + 0.154 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison of our GauHuman and baseline methods on the ZJU_MoCap and MonoCap data sets. We use bold and underlined text to denote the best and second-best results of each metric respectively. Our GauHuman achieves the state-of-the-art PSNR and LPIPS with fast training (1~2 minutes) and real-time rendering (up to 189 FPS) in both data sets. Note that PixelNeRF [129] and NHP [60] are pre-trained for an additional 10 hours. \\(\\dagger\\) denotes the fine-tuning time for generalizable methods. LPIPS\\* = 1000 × LPIPS. Frames per second (FPS) is measured on an RTX 3090. For a fair comparison, we do not conduct test-time optimization of SMPL parameters with images from the test set on InstantAvatar [49]." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.155, + 0.875, + 0.352 + ], + "angle": 0, + "content": "
MethodZJU_MoCapMonoCap
PSNR↑SSIM↑LPIPS*↓TrainFPSPSNR↑SSIM↑LPIPS*↓TrainFPS
PixelNeRF [129]24.710.892121.86\\( 1h^† \\)1.2026.430.96043.98\\( 1h^† \\)0.75
NHP [60]28.250.95564.77\\( 1h^† \\)0.1530.510.98027.14\\( 1h^† \\)0.05
NB [82]29.030.96442.4710h1.4832.360.98616.7010h0.98
AN [81]29.770.96546.8910h1.1131.070.98519.4710h0.31
AS [83]30.380.97537.2310h0.4032.480.98813.1810h0.29
HumanNeRF [116]30.660.96933.3810h0.3032.680.98715.5210h0.08
DVA [88]29.450.95637.741.5h16.532.990.98315.831.5h10.5
InstantNVR [30]31.010.97138.455m2.2032.610.98816.6810m0.75
InstantAvatar [49]29.730.93868.413m4.1530.790.96439.756m2.54
GauHuman(Ours)31.340.96530.511m18933.450.98513.352m154
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.362, + 0.21, + 0.379 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.387, + 0.267, + 0.403 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.411, + 0.471, + 0.681 + ], + "angle": 0, + "content": "Datasets. Experiments are conducted on two widely used human modelling data sets, i.e., ZJU_MoCap [82] and MonoCap [36, 37, 83]. For ZJU_MoCap, same as previous work [30, 116], we select 6 human subjects (377, 386, 387, 392, 393, 394) to conduct experiments. We also adopt the same training and testing setting as [30], i.e., one camera is used for training, while the remaining cameras are used for evaluation. For each subject, we sample 1 frame every 5 frames and collect 100 frames for training. The foreground masks, camera, and SMPL parameters provided by the data set are used for evaluation purposes. MonoCap data set contains four multi-view videos collected by [83] from the DeepCap data set [36] and the DynaCap data set [37]. Similar to ZJU_MoCap, it provides human masks, cameras, and SMPL parameters. We adopt the same camera views as [83] for training and testing. For each subject, we collect 100 frames for training by sampling 1 frame every 5 frames. More details are shown in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.684, + 0.472, + 0.82 + ], + "angle": 0, + "content": "Comparison Methods. We compare our GauHuman with two categories of state-of-the-art human modelling methods, i.e., subject-specific optimization based methods [30, 49, 81-83, 88, 116] and generalizable methods [60, 129]. Following [30], we report the average metric values of all selected human subjects. For a fair comparison, we do not conduct test-time optimization of SMPL parameters with images from the test set on InstantAvatar [49]. More details are shown in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.272, + 0.848 + ], + "angle": 0, + "content": "4.2. Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Tab. 1 summarizes the overall novel view synthesis results of our GauHuman, NB [82], AN [81], AS [83], HumanNeRF [116], DVA [88], InstantNVR [30], InstantAvatar [49]," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.363, + 0.895, + 0.651 + ], + "angle": 0, + "content": "PixelNeRF [129], and NHP [60]. Thanks to our proposed articulated 3D human Gaussian-based representation, we can fast optimize it for around 1 or 2 minutes to achieve photorealistic rendering results, while baseline methods (i.e., NB, AN, AS, and HumanNeRF) take around 10 hours to finish the training and generalizable methods takes about 1 hour for fine-tuning after 10 hours' pre-training. To accelerate the training of 3D human modelling, InstantNVR [30] and InstantAvatar [49] reduce the training to several minutes with multi hashing encoding representation. However, their rendering speed still restricts their applications in real-world applications. When compared with these two state-of-the-art methods for fasting training, our GauHuman achieves slightly better performance with 3 times faster training and 45 times faster rendering speed. Among baseline methods, DVA achieves a fast rendering speed (up to 16.5 FPS) with efficient mixtures of volumetric primitives. However, in contrast to our GauHuman, DVA takes about \\(90\\mathrm{x}\\) more time to converge and \\(11.45\\mathrm{x}\\) more time to render one image." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.665, + 0.684, + 0.681 + ], + "angle": 0, + "content": "4.3. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.901 + ], + "angle": 0, + "content": "As shown in Fig. 3, our GauHuman shows better appearance and geometry details than DVA [88], NB [82] and NHP [60]. Although NB [82] and NHP [60] show impressive results with 4-view videos, they do not perform well with monocular videos. NB [82] encodes human structure information with latent codes positioned in SMPL vertex's locations, which fails to capture complex human motions. NHP [60] does not generalize well in monocular videos due to the information missing from input images. Thanks to the LBS weight field and deformation field learned in HumanNeRF [116] and AS [83], they achieve comparable visualization results as our GauHuman for \\(600\\mathrm{x}\\) more time for training and \\(472\\mathrm{x}\\) more time for rendering. InstantNVR [30] and InstantAvatar [49] reduce the training time to several minutes with a multi" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20423" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.066, + 0.885, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.495, + 0.893, + 0.525 + ], + "angle": 0, + "content": "Figure 3. Novel view synthesis results produced by our GauHuman and baseline methods on ZJU_MoCap and MonoCap. The bottom lines show the training time and rendering speed of each method on ZJU_Mocap. Zoom in for the best view." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.54, + 0.473, + 0.661 + ], + "angle": 0, + "content": "hashing encoding representation, but they produce slightly worse visualization results than ours and require \\(45\\mathrm{x}\\) more rendering time. DVA [] improves the rendering speed to 16.5 FPS with efficient mixtures of volumetric primitives. However, it fails to produce high-quality results and produces obvious artifacts on garments and faces, due to incomplete information within conditional input images and a lack of mechanisms to adaptively control the number of voxels." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.677, + 0.232, + 0.693 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.702, + 0.471, + 0.778 + ], + "angle": 0, + "content": "We perform ablation studies with the sequence of 386 from the ZJU_Mocap data set. To highlight the effectiveness of each proposed component, we project the 3D human bounding box to each camera plane to obtain the bounding box mask and report these metrics based on the masked area." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Pose Refinement and LBS Weight Field. Tab. 2 shows the results of ablating pose refinement and LBS weight field modules. All models are trained in around 1 minute. As shown in Tab. 2, pose refinement and LBS weight field modules significantly improve all metrics with negligible computational cost. The visualization results in Fig. 4 show that these two modules can help erase artifacts and learn fine details like cloth wrinkles. For example, we observe obvious" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.54, + 0.892, + 0.568 + ], + "angle": 0, + "content": "artifacts in hands and fewer cloth wrinkles without using either one of these two modules." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.569, + 0.892, + 0.596 + ], + "angle": 0, + "content": "Table 2. Quantitative Results of ablating pose refinement and LBS weight field modules. \\(\\mathrm{LPIPS^{*}} = 1000\\times\\) LPIPS." + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.597, + 0.885, + 0.668 + ], + "angle": 0, + "content": "
PSNR↑SSIM↑LPIPS*↓TrainFPS
Ours (full model)28.080.886103.355s189
Ours w/o pose refine27.940.884108.845s192
Ours w/o LBS field27.870.882108.952s190
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.895, + 0.856 + ], + "angle": 0, + "content": "3D Gaussian Initialization. In Gaussian Splitting, previous research [56] indicates that initialization plays an important role in the final results. For static scenes, there are two commonly adopted initialization approaches, i.e., random initialization or initialization with sparse points from Structure-from-Motion (SfM). To incorporate human articulated information into the initialization, we propose to use articulated vertex points from SMPL to initialize our 3D Gaussians in the canonical space. We compare our proposed articulated initialization with random initialization1 in Tab. 3. With the same number of 3D Gaussians at initialization, our proposed articulated initialization produces both better" + }, + { + "type": "page_footnote", + "bbox": [ + 0.499, + 0.863, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Estimating sparse point cloud from Structure-from-Motion (SfM) requires multi-view inputs, which is infeasible with monocular videos. So we do not compare it with our articulated initialization." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20424" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.074, + 0.89, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.291, + 0.897, + 0.335 + ], + "angle": 0, + "content": "Figure 4. Qualitative results of ablation study on the 386 sequence of ZJU_Mocap data set. \"w/o pose refinement\" and \"w/o LBS weight field\" mean that we do not adopt the pose refinement or LBS weight field module respectively. \"w/ random Init (#Gau 6890)\" and \"w/ random Init (#Gau 30000)\" mean that we use random initialization to initialize 6890 or 30000 3D Gaussians respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.345, + 0.473, + 0.496 + ], + "angle": 0, + "content": "quantitative and qualitative results than random initialization as shown in Tab. 3 and Fig. 4. Notably, our articulated initialization converges x10 times much faster than the random initialization, which illustrates the strength of encoding human articulated information into the initialization. For random initialization, we need to initialize 30k Gaussians to achieve comparable performance as our proposed articulated initialization, while at cost of 30x times more training time. Table 3. Quantitative Results of ablating 3D Gaussian initialization. #Gau denotes the number of 3D Gaussians at initialization." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.497, + 0.465, + 0.568 + ], + "angle": 0, + "content": "
PSNR↑SSIM↑LPIPS*↓Train
Articulated Init (w/ #Gau 6890)28.080.886103.355s
Random Init (w/ #Gau 6890)27.980.882108.95m
Random Init (w/ #Gau 30k)28.080.884104.330m
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.572, + 0.47, + 0.784 + ], + "angle": 0, + "content": "Split/Clone/Merge/Prune. Adaptive control of 3D Gaussians with split, clone, and prune operations substantially improves the rendering quality, but it trivially leads to a large number of 3D Gaussians (up to \\(100\\mathrm{k}\\)), which slows down the whole optimization process. As mentioned in Sec. 3.3, we propose to use the KL divergence measure to regulate the split and clone process. As shown in Tab. 4, our KL-based split and clone operations lead to \\(60\\mathrm{x}\\) times faster convergence speed when compared with the original split and clone operation proposed in [56]. Furthermore, our novel merge operation merges about \\(8.5\\%\\) of 3D Gaussians without degrading the performance. Articulated prune operation with SMPL prior introduced in Sec. 3.3 also helps prune about \\(0.5\\mathrm{k}\\) 3D Gaussians far away from the 3D human body." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.786, + 0.47, + 0.827 + ], + "angle": 0, + "content": "Table 4. Quantitative Results of ablating our proposed KL-based split/clone operations and a novel merge operation. #Gau denotes the final number of 3D Gaussians during optimization." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.828, + 0.462, + 0.912 + ], + "angle": 0, + "content": "
PSNR↑SSIM↑LPIPS*↓#GauTrain
Ours (full model)28.080.886103.31316255s
w/o KL split/clone28.010.885103.1100k1h
w/o KL merge28.080.886103.61437670s
w/o articulated prune28.080.886103.41368757s
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.344, + 0.748, + 0.36 + ], + "angle": 0, + "content": "5. Discussion and Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.38, + 0.895, + 0.623 + ], + "angle": 0, + "content": "To conclude, we propose GauHuman, a 3D human model with Gaussian Splating representation for fast training \\((1\\sim 2\\) minutes) and real-time rendering (166 FPS) of 3D humans. Specifically, we encode this representation in a canonical space and transform 3D Gaussian from canonical space to posed space with LBS transformation, in which effective pose refinement and LBS weight field modules are also designed to learn fine details of 3D humans. To enable fast optimization, we initialize and prune 3D Gaussians with 3D human prior, while splitting/cloning 3D Gaussians via KI divergence guidance, accompanied by a novel merge operation for further speeding up. Experiments on monocular videos suggest that our GauHuman achieves state-of-the-art performance while maintaining both fast training and real-time rendering speed. In addition, GauHuman can fast learn 3D human performers with \\(\\sim 13k\\) Gaussians." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.633, + 0.894, + 0.74 + ], + "angle": 0, + "content": "Limitations. 1). GauHuman is composed of 3D Gaussians and current pipeline does not support extracting 3D meshes. How to extract meshes from 3D Gaussians remains a future work to be investigated. 2). Recovering 3D humans with loose/complex clothing from monocular videos is still a challenging problem to tackle. One direction is to incorporate physical simulation of garments into current pipelines." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Acknowledgment. This study is supported by the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOET2EP20221-0012), NTU NAP, and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). We would like to thank Zhaoxi Chen, Wei Li, Guangcong Wang from NTU, Xurong Xie from ISCAS, Zhiyuan Yu from ZJU, and Zihao Huang from HUST for the discussions on the paper draft." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20425" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.116, + 0.471, + 0.184 + ], + "angle": 0, + "content": "[1] Badour AlBahar, Jingwan Lu, Jimei Yang, Zhixin Shu, Eli Shechtman, and Jia-Bin Huang. Pose with style: Detail-preserving pose-guided image synthesis with conditional stylegan. ACM Transactions on Graphics (TOG), 40(6): 1-11, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.186, + 0.472, + 0.255 + ], + "angle": 0, + "content": "[2] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 696-712. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.256, + 0.472, + 0.324 + ], + "angle": 0, + "content": "[3] Thiemo Alldieck, Mihai Zanfir, and Cristian Sminchisescu. Photorealistic monocular 3d reconstruction of humans wearing clothing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1506-1515, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.326, + 0.471, + 0.395 + ], + "angle": 0, + "content": "[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.397, + 0.471, + 0.465 + ], + "angle": 0, + "content": "[5] Bharat Lal Bhatnagar, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. Loopreg: Self-supervised learning of implicit surface correspondences, pose and shape for 3d human mesh registration. Advances in Neural Information Processing Systems, 33:12909-12922, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.467, + 0.471, + 0.522 + ], + "angle": 0, + "content": "[6] Mario Botsch, Alexander Hornung, Matthias Zwicker, and Leif Kobbelt. High-quality surface splatting on today's gpus. In Proceedings Eurographics/IEEE VGTC Symposium Point-Based Graphics, 2005., pages 17-141. IEEE, 2005. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.523, + 0.471, + 0.578 + ], + "angle": 0, + "content": "[7] Chris Buehler, Michael Bosse, Leonard McMillan, Steven Gortler, and Michael Cohen. Unstructured lumigraph rendering. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 497-504. 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.579, + 0.471, + 0.648 + ], + "angle": 0, + "content": "[8] Sihun Cha, Kwanggyoon Seo, Amirsaman Ashtari, and Jun-yong Noh. Generating texture for 3d human avatar from a single image using sampling and refinement networks. In Computer Graphics Forum, pages 385-396. Wiley Online Library, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.649, + 0.471, + 0.703 + ], + "angle": 0, + "content": "[9] Gaurav Chaurasia, Sylvain Duchene, Olga Sorkine-Hornung, and George Drettakis. Depth synthesis and local warps for plausible image-based navigation. ACM Transactions on Graphics (TOG), 32(3):1-12, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.705, + 0.471, + 0.761 + ], + "angle": 0, + "content": "[10] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. arXiv preprint arXiv:2103.15595, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.761, + 0.471, + 0.815 + ], + "angle": 0, + "content": "[11] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.818, + 0.471, + 0.872 + ], + "angle": 0, + "content": "[12] Jianchuan Chen, Ying Zhang, Di Kang, Xuefei Zhe, Linchao Bao, Xu Jia, and Huchuan Lu. Animatable neural radiance fields from monocular rgb videos. arXiv preprint arXiv:2106.13629, 2021. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.873, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[13] Xu Chen, Yufeng Zheng, Michael J Black, Otmar Hilliges, and Andreas Geiger. Snarf: Differentiable forward skin" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.116, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.094, + 0.892, + 0.135 + ], + "angle": 0, + "content": "ning for animating non-rigid neural implicit shapes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11594-11604, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.137, + 0.894, + 0.205 + ], + "angle": 0, + "content": "[14] Zhaoxi Chen, Fangzhou Hong, Haiyi Mei, Guangcong Wang, Lei Yang, and Ziwei Liu. Primdiffusion: Volumetric primitives diffusion for 3d human generation. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.208, + 0.894, + 0.29 + ], + "angle": 0, + "content": "[15] Wei Cheng, Ruixiang Chen, Siming Fan, Wanqi Yin, Keyu Chen, Zhongang Cai, Jingbo Wang, Yang Gao, Zhengming Yu, Zhengyu Lin, et al. Dna-rendering: A diverse neural actor repository for high-fidelity human-centric rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19982-19993, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.292, + 0.894, + 0.361 + ], + "angle": 0, + "content": "[16] Julian Chibane, Aayush Bansal, Verica Lazova, and Gerard Pons-Moll. Stereo radiance fields (srf): Learning view synthesis for sparse views of novel scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7911-7920, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.363, + 0.894, + 0.431 + ], + "angle": 0, + "content": "[17] Hongsuk Choi, Gyeongsik Moon, Matthieu Armando, Vincent Leroy, Kyoung Mu Lee, and Grégory Rogez. Mononhr: Monocular neural human renderer. In 2022 International Conference on 3D Vision (3DV), pages 242-251. IEEE, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.434, + 0.894, + 0.502 + ], + "angle": 0, + "content": "[18] Alvaro Collet, Ming Chuang, Pat Sweeney, Don Gillett, Dennis Evseev, David Calabrese, Hugues Hoppe, Adam Kirk, and Steve Sullivan. High-quality streamable free-viewpoint video. ACM Transactions on Graphics (ToG), 34 (4):1-13, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.505, + 0.894, + 0.574 + ], + "angle": 0, + "content": "[19] François Darmon, Bénédicte Bascle, Jean-Clement Devaux, Pascal Monasse, and Mathieu Aubry. Improving neural implicit surfaces geometry with patch warping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6260-6269, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.576, + 0.894, + 0.659 + ], + "angle": 0, + "content": "[20] Boyang Deng, John P Lewis, Timothy Jeruzalski, Gerard Pons-Moll, Geoffrey Hinton, Mohammad Norouzi, and Andrea Tagliasacchi. Nasa neural articulated shape approximation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VII 16, pages 612-628. Springer, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.661, + 0.894, + 0.702 + ], + "angle": 0, + "content": "[21] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nef: Fewer views and faster training for free. arXiv preprint arXiv:2107.02791, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.704, + 0.894, + 0.786 + ], + "angle": 0, + "content": "[22] Mingsong Dou, Sameh Khamis, Yury Degtyarev, Philip Davidson, Sean Ryan Fanello, Adarsh Kowdle, Sergio Orts Escolano, Christoph Rhemann, David Kim, Jonathan Taylor, et al. Fusion4d: Real-time performance capture of challenging scenes. ACM Transactions on Graphics (ToG), 35(4): 1-13, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.789, + 0.894, + 0.857 + ], + "angle": 0, + "content": "[23] Martin Eisemann, Bert De Decker, Marcus Magnor, Philippe Bekaert, Edilson De Aguiar, Naveed Ahmed, Christian Theobalt, and Anita Sellent. Floating textures. In Computer graphics forum, pages 409-418. Wiley Online Library, 2008. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.86, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[24] John Flynn, Ivan Neulander, James Philbin, and Noah Snavely. Deepstereo: Learning to predict new views from the world's imagery. In Proceedings of the IEEE conference" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.094, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20426" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.092, + 0.472, + 0.12 + ], + "angle": 0, + "content": "on computer vision and pattern recognition, pages 5515-5524, 2016.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.122, + 0.472, + 0.191 + ], + "angle": 0, + "content": "[25] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.193, + 0.472, + 0.262 + ], + "angle": 0, + "content": "[26] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Chen Qian, Chen Change Loy, Wayne Wu, and Ziwei Liu. Stylegan-human: A data-centric odyssey of human generation. In European Conference on Computer Vision, pages 1-19. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.264, + 0.472, + 0.332 + ], + "angle": 0, + "content": "[27] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Wayne Wu, and Ziwei Liu. Unitedhuman: Harnessing multi-source data for high-resolution human generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7301-7311, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.334, + 0.472, + 0.39 + ], + "angle": 0, + "content": "[28] Xiangjun Gao, Jiaolong Yang, Jongyoo Kim, Sida Peng, Zicheng Liu, and Xin Tong. Mps-nerf: Generalizable 3d human rendering from multiview images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.392, + 0.472, + 0.459 + ], + "angle": 0, + "content": "[29] Stephan J Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14346-14355, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.462, + 0.472, + 0.532 + ], + "angle": 0, + "content": "[30] Chen Geng, Sida Peng, Zhen Xu, Hujun Bao, and Xiaowei Zhou. Learning neural volumetric representations of dynamic humans in minutes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8759-8770, 2023. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.533, + 0.472, + 0.589 + ], + "angle": 0, + "content": "[31] Michael Goesele, Noah Snavely, Brian Curless, Hugues Hoppe, and Steven M Seitz. Multi-view stereo for community photo collections. In 2007 IEEE 11th International Conference on Computer Vision, pages 1-8. IEEE, 2007. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.591, + 0.472, + 0.644 + ], + "angle": 0, + "content": "[32] Steven J Gortler, Radek Grzeszcuk, Richard Szeliski, and Michael F Cohen. The lumigraph. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 453-464. 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.647, + 0.472, + 0.673 + ], + "angle": 0, + "content": "[33] Markus Gross and Hanspeter Pfister. Point-based graphics. Elsevier, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.676, + 0.472, + 0.732 + ], + "angle": 0, + "content": "[34] Jeffrey P Grossman and William J Dally. Point sample rendering. In Rendering Techniques' 98: Proceedings of the Eurographics Workshop in Vienna, Austria, June 29—July 1, 1998 9, pages 181–192. Springer, 1998. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.734, + 0.472, + 0.815 + ], + "angle": 0, + "content": "[35] Kaiwen Guo, Peter Lincoln, Philip Davidson, Jay Busch, Xueming Yu, Matt Whalen, Geoff Harvey, Sergio Orts-Escolano, Rohit Pandey, Jason Dourgarian, et al. The relightables: Volumetric performance capture of humans with realistic relighting. ACM Transactions on Graphics (ToG), 38(6):1-19, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.818, + 0.472, + 0.899 + ], + "angle": 0, + "content": "[36] Marc Habermann, Weipeng Xu, Michael Zollhofer, Gerard Pons-Moll, and Christian Theobalt. Deepcap: Monocular human performance capture using weak supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5052-5063, 2020. 1, 6" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.092, + 0.472, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.093, + 0.894, + 0.147 + ], + "angle": 0, + "content": "[37] Marc Habermann, Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Real-time deep dynamic characters. ACM Transactions on Graphics (ToG), 40(4):1-16, 2021. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.15, + 0.894, + 0.205 + ], + "angle": 0, + "content": "[38] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering. ACM Transactions on Graphics (ToG), 37(6):1-15, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.207, + 0.894, + 0.262 + ], + "angle": 0, + "content": "[39] Philipp Henzler, Niloy J Mitra, and Tobias Ritschel. Escaping plato's cave: 3d shape from adversarial rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9984-9993, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.264, + 0.894, + 0.317 + ], + "angle": 0, + "content": "[40] Fangzhou Hong, Zhaoxi Chen, Yushi Lan, Liang Pan, and Ziwei Liu. Eva3d: Compositional 3d human generation from 2d image collections. arXiv preprint arXiv:2210.04888, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.321, + 0.894, + 0.374 + ], + "angle": 0, + "content": "[41] Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. arXiv preprint arXiv:2205.08535, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.377, + 0.894, + 0.418 + ], + "angle": 0, + "content": "[42] Shoukang Hu, Fangzhou Hong, Liang Pan, Haiyi Mei, Lei Yang, and Ziwei Liu. Sherf: Generalizable human nerf from a single image. arXiv preprint arXiv:2303.12791, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.42, + 0.894, + 0.475 + ], + "angle": 0, + "content": "[43] Yangyi Huang, Hongwei Yi, Yuliang Xiu, Tingting Liao, Jiaxiang Tang, Deng Cai, and Justus Thies. Tech: Text-guided reconstruction of lifelike clothed humans. arXiv preprint arXiv:2308.08545, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.477, + 0.894, + 0.545 + ], + "angle": 0, + "content": "[44] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. Arch: Animatable reconstruction of clothed humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3093-3102, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.548, + 0.894, + 0.603 + ], + "angle": 0, + "content": "[45] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5885-5894, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.605, + 0.894, + 0.659 + ], + "angle": 0, + "content": "[46] Wonbong Jang and Lourdes Agapito. Codenerf: Disentangled neural radiance fields for object categories. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12949-12958, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.661, + 0.894, + 0.73 + ], + "angle": 0, + "content": "[47] Rohit Jena, Ganesh Subramanian Iyer, Siddharth Choudhary, Brandon Smith, Pratik Chaudhari, and James Gee. Splatarmor: Articulated gaussian splatting for animatable humans from monocular rgb videos. arXiv preprint arXiv:2311.10812, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.733, + 0.894, + 0.8 + ], + "angle": 0, + "content": "[48] Boyi Jiang, Yang Hong, Hujun Bao, and Juyong Zhang. Self-con: Self reconstruction your digital avatar from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5605-5615, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.802, + 0.894, + 0.871 + ], + "angle": 0, + "content": "[49] Tianjian Jiang, Xu Chen, Jie Song, and Otmar Hilliges. Instantavatar: Learning avatars from monocular video in 60 seconds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16922-16932, 2023. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.873, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[50] Wei Jiang, Kwang Moo Yi, Golnoosh Samei, Oncel Tuzel, and Anurag Ranjan. Neuman: Neural human radiance field" + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20427" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.092, + 0.471, + 0.119 + ], + "angle": 0, + "content": "from a single video. arXiv preprint arXiv:2203.12575, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.122, + 0.472, + 0.177 + ], + "angle": 0, + "content": "[51] Yuming Jiang, Shuai Yang, Haonan Qiu, Wayne Wu, Chen Change Loy, and Ziwei Liu. Text2human: Text-driven controllable human image generation. ACM Transactions on Graphics (TOG), 41(4):1-11, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.179, + 0.472, + 0.233 + ], + "angle": 0, + "content": "[52] Yuming Jiang, Shuai Yang, Tong Liang Koh, Wayne Wu, Chen Change Loy, and Ziwei Liu. Text2performer: Text-driven human video generation. arXiv preprint arXiv:2304.08483, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.235, + 0.472, + 0.303 + ], + "angle": 0, + "content": "[53] Mohammad Mahdi Johari, Yann Lepoittevin, and François Fleuret. Geonerf: Generalizing nerf with geometry priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18365-18375, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.305, + 0.472, + 0.373 + ], + "angle": 0, + "content": "[54] HyunJun Jung, Nikolas Brasch, Jifei Song, Eduardo Perez-Pellitero, Yiren Zhou, Zhihao Li, Nassir Navab, and Benjamin Busam. Deformable 3d gaussian splatting for animatable human avatars. arXiv preprint arXiv:2312.15059, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.376, + 0.472, + 0.417 + ], + "angle": 0, + "content": "[55] James T Kajiya and Brian P Von Herzen. Ray tracing volume densities. ACM SIGGRAPH computer graphics, 18(3):165-174, 1984. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.42, + 0.472, + 0.473 + ], + "angle": 0, + "content": "[56] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 42(4):1-14, 2023. 2, 3, 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.476, + 0.472, + 0.517 + ], + "angle": 0, + "content": "[57] Muhammed Kocabas, Jen-Hao Rick Chang, James Gabriel, Oncel Tuzel, and Anurag Ranjan. Hugs: Human gaussian splats. arXiv preprint arXiv:2311.17910, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.519, + 0.472, + 0.574 + ], + "angle": 0, + "content": "[58] Georgios Kopanas, Julien Philip, Thomas Leimkuhler, and George Drettakis. Point-based neural rendering with perview optimization. In Computer Graphics Forum, pages 29-43. Wiley Online Library, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.576, + 0.472, + 0.63 + ], + "angle": 0, + "content": "[59] Georgios Kopanas, Thomas Leimkuhler, Gilles Rainer, Clément Jambon, and George Drettakis. Neural point catauustics for novel-view synthesis of reflections. ACM Transactions on Graphics (TOG), 41(6):1-15, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.633, + 0.472, + 0.7 + ], + "angle": 0, + "content": "[60] Youngjoong Kwon, Dahun Kim, Duygu Ceylan, and Henry Fuchs. Neural human performer: Learning generalizable radiance fields for human performance rendering. Advances in Neural Information Processing Systems, 34:24741-24752, 2021. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.703, + 0.472, + 0.758 + ], + "angle": 0, + "content": "[61] Christoph Lassner and Michael Zollhofer. Pulsar: Efficient sphere-based neural rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1440-1449, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.76, + 0.472, + 0.801 + ], + "angle": 0, + "content": "[62] Marc Levoy and Pat Hanrahan. Light field rendering. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 441-452. 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.803, + 0.472, + 0.857 + ], + "angle": 0, + "content": "[63] Kathleen M Lewis, Srivatsan Varadharajan, and Ira Kemelmacher-Shlizerman. Tryongan: Body-aware try-on via layered interpolation. ACM Transactions on Graphics (TOG), 40(4):1-10, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.86, + 0.472, + 0.9 + ], + "angle": 0, + "content": "[64] Ruilong Li, Matthew Tancik, and Angjoo Kanazawa. Nerve: A general nef acceleration toolbox. arXiv preprint arXiv:2210.04847, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.093, + 0.894, + 0.176 + ], + "angle": 0, + "content": "[65] Tingting Liao, Xiaomei Zhang, Yuliang Xiu, Hongwei Yi, Xudong Liu, Guo-Jun Qi, Yong Zhang, Xuan Wang, Xi-angyu Zhu, and Zhen Lei. High-fidelity clothed avatar reconstruction from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8662–8672, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.179, + 0.894, + 0.233 + ], + "angle": 0, + "content": "[66] Haotong Lin, Sida Peng, Zhen Xu, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Efficient neural radiance fields for interactive free-viewpoint video. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.235, + 0.894, + 0.289 + ], + "angle": 0, + "content": "[67] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.292, + 0.894, + 0.359 + ], + "angle": 0, + "content": "[68] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. ACM transactions on graphics (TOG), 40(6):1-16, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.362, + 0.894, + 0.444 + ], + "angle": 0, + "content": "[69] Yuan Liu, Sida Peng, Lingjie Liu, Qianqian Wang, Peng Wang, Christian Theobalt, Xiaowei Zhou, and Wenping Wang. Neural rays for occlusion-aware image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7824-7833, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.447, + 0.894, + 0.502 + ], + "angle": 0, + "content": "[70] Stephen Lombardi, Tomas Simon, Gabriel Schwartz, Michael Zollhoefer, Yaser Sheikh, and Jason Saragih. Mixture of volumetric primitives for efficient neural rendering. ACM Transactions on Graphics (ToG), 40(4):1-13, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.504, + 0.894, + 0.559 + ], + "angle": 0, + "content": "[71] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi-person linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia), 34(6):248:1-248:16, 2015. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.562, + 0.894, + 0.615 + ], + "angle": 0, + "content": "[72] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.618, + 0.894, + 0.657 + ], + "angle": 0, + "content": "[73] Nelson Max. Optical models for direct volume rendering. IEEE Transactions on Visualization and Computer Graphics, 1(2):99-108, 1995. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.661, + 0.894, + 0.73 + ], + "angle": 0, + "content": "[74] Yifang Men, Yiming Mao, Yuning Jiang, Wei-Ying Ma, and Zhouhui Lian. Controllable person image synthesis with attribute-decomposed gan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5084-5093, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.732, + 0.894, + 0.799 + ], + "angle": 0, + "content": "[75] Marko Mihajlovic, Yan Zhang, Michael J Black, and Siyu Tang. Leap: Learning articulated occupancy of people. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10461-10471, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.803, + 0.894, + 0.871 + ], + "angle": 0, + "content": "[76] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.874, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[77] Arthur Moreau, Jifei Song, Helisa Dhamo, Richard Shaw, Yiren Zhou, and Eduardo Pérez-Pellitero. Human gaussian" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20428" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.093, + 0.469, + 0.12 + ], + "angle": 0, + "content": "splatting: Real-time rendering of animatable avatars. arXiv preprint arXiv:2311.17113, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.122, + 0.472, + 0.177 + ], + "angle": 0, + "content": "[78] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.179, + 0.472, + 0.246 + ], + "angle": 0, + "content": "[79] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. arXiv preprint arXiv:2112.00724, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.249, + 0.471, + 0.305 + ], + "angle": 0, + "content": "[80] Atsuhiro Noguchi, Xiao Sun, Stephen Lin, and Tatsuya Harada. Neural articulated radiance field. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5762-5772, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.307, + 0.471, + 0.361 + ], + "angle": 0, + "content": "[81] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Animatable neural radiance fields for modeling dynamic human bodies. In ICCV, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.363, + 0.471, + 0.446 + ], + "angle": 0, + "content": "[82] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9054-9063, 2021. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.448, + 0.471, + 0.504 + ], + "angle": 0, + "content": "[83] Sida Peng, Shangzhan Zhang, Zhen Xu, Chen Geng, Boyi Jiang, Hujun Bao, and Xiaowei Zhou. Animatable neural implicit surfaces for creating avatars from videos. arXiv preprint arXiv:2203.08133, 4(5), 2022. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.505, + 0.471, + 0.546 + ], + "angle": 0, + "content": "[84] Eric Penner and Li Zhang. Soft 3d reconstruction for view synthesis. ACM Transactions on Graphics (TOG), 36(6): 1-11, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.548, + 0.471, + 0.616 + ], + "angle": 0, + "content": "[85] Hanspeter Pfister, Matthias Zwicker, Jeroen Van Baar, and Markus Gross. Surfels: Surface elements as rendering primitives. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 335-342, 2000. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.618, + 0.471, + 0.687 + ], + "angle": 0, + "content": "[86] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14335-14345, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.689, + 0.471, + 0.731 + ], + "angle": 0, + "content": "[87] Konstantinos Rematas, Ricardo Martin-Brualla, and Vittorio Ferrari. Sharf: Shape-conditioned radiance fields from a single view. arXiv preprint arXiv:2102.08860, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.733, + 0.471, + 0.813 + ], + "angle": 0, + "content": "[88] Edoardo Remelli, Timur Bagautdinov, Shunsuke Saito, Chenglei Wu, Tomas Simon, Shih-En Wei, Kaiwen Guo, Zhe Cao, Fabian Prada, Jason Saragih, et al. Drivable volumetric avatars using texel-aligned features. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1–9, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.817, + 0.471, + 0.872 + ], + "angle": 0, + "content": "[89] Liu Ren, Hanspeter Pfister, and Matthias Zwicker. Object space ewa surface splatting: A hardware accelerated approach to high quality point rendering. In Computer Graphics Forum, pages 461-470. Wiley Online Library, 2002. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.874, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[90] Helge Rhodin, Nadia Robertini, Christian Richardt, Hans-Peter Seidel, and Christian Theobalt. A versatile scene" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.093, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "model with differentiable visibility applied to generative pose estimation. In Proceedings of the IEEE International Conference on Computer Vision, pages 765-773, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.136, + 0.894, + 0.19 + ], + "angle": 0, + "content": "[91] Gernot Riegler and Vladlen Koltun. Free view synthesis. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIX 16, pages 623-640. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.192, + 0.894, + 0.233 + ], + "angle": 0, + "content": "[92] Darius Rückert, Linus Franke, and Marc Stamminger. Adop: Approximate differentiable one-pixel point rendering. ACM Transactions on Graphics (ToG), 41(4):1-14, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.234, + 0.894, + 0.272 + ], + "angle": 0, + "content": "[93] Miguel Sainz and Renato Pajarola. Point-based rendering techniques. Computers & Graphics, 28(6):869-879, 2004. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.274, + 0.894, + 0.344 + ], + "angle": 0, + "content": "[94] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2304-2314, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.345, + 0.894, + 0.413 + ], + "angle": 0, + "content": "[95] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 84-93, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.414, + 0.894, + 0.483 + ], + "angle": 0, + "content": "[96] Shunsuke Saito, Jinlong Yang, Qianli Ma, and Michael J Black. Scanimate: Weakly supervised learning of skinned clothed avatar networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2886-2897, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.484, + 0.894, + 0.539 + ], + "angle": 0, + "content": "[97] Kripasindhu Sarkar, Lingjie Liu, Vladislav Golyanik, and Christian Theobalt. Humangan: A generative model of human images. In 2021 International Conference on 3D Vision (3DV), pages 258-267. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.54, + 0.894, + 0.594 + ], + "angle": 0, + "content": "[98] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.596, + 0.894, + 0.678 + ], + "angle": 0, + "content": "[99] Johannes L Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 501-518. Springer, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.679, + 0.894, + 0.748 + ], + "angle": 0, + "content": "[100] Vincent Sitzmann, Justus Thies, Felix Heide, Matthias Nießner, Gordon Wetzstein, and Michael Zollhofer. Deepvoxels: Learning persistent 3d feature embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2437-2446, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.749, + 0.894, + 0.79 + ], + "angle": 0, + "content": "[101] Noah Snavely, Steven M Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. In ACM siggraph 2006 papers, pages 835-846. 2006. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.791, + 0.894, + 0.858 + ], + "angle": 0, + "content": "[102] Carsten Stoll, Nils Hasler, Juergen Gall, Hans-Peter Seidel, and Christian Theobalt. Fast articulated motion tracking using a sums of gaussians body model. In 2011 International Conference on Computer Vision, pages 951-958. IEEE, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[103] Shih-Yang Su, Frank Yu, Michael Zollhöfer, and Helge Rhodin. A-nerf: Articulated neural radiance fields for learning human shape, appearance, and pose. Advances in Neural" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20429" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.092, + 0.471, + 0.119 + ], + "angle": 0, + "content": "Information Processing Systems, 34:12278-12291, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.472, + 0.207 + ], + "angle": 0, + "content": "[104] Zhuo Su, Lan Xu, Zerong Zheng, Tao Yu, Yebin Liu, and Lu Fang. Robustfusion: Human volumetric capture with data-driven visual cues using a rgbd camera. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16, pages 246-264. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.208, + 0.472, + 0.276 + ], + "angle": 0, + "content": "[105] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.472, + 0.374 + ], + "angle": 0, + "content": "[106] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3d shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11358-11367, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.472, + 0.446 + ], + "angle": 0, + "content": "[107] Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srinivasan, Edgar Tretschk, Wang Yifan, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lombardi, et al. Advances in neural rendering. In Computer Graphics Forum, pages 703-735. Wiley Online Library, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.448, + 0.472, + 0.502 + ], + "angle": 0, + "content": "[108] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: Image synthesis using neural textures. Acm Transactions on Graphics (TOG), 38(4):1-12, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.472, + 0.574 + ], + "angle": 0, + "content": "[109] Garvita Tiwari, Nikolaos Sarafianos, Tony Tung, and Gerard Pons-Moll. Neural-gif: Neural generalized implicit functions for animating people in clothing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11708-11718, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.472, + 0.631 + ], + "angle": 0, + "content": "[110] Alex Trevithick and Bo Yang. Grf: Learning a general radiance field for 3d representation and rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15182-15192, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.472, + 0.701 + ], + "angle": 0, + "content": "[111] Angtian Wang, Peng Wang, Jian Sun, Adam Kortylewski, and Alan Yuille. Voge: a differentiable volume renderer using gaussian ellipsoids for analysis-by-synthesis. In The Eleventh International Conference on Learning Representations, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.472, + 0.759 + ], + "angle": 0, + "content": "[112] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.472, + 0.844 + ], + "angle": 0, + "content": "[113] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snively, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690-4699, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.472, + 0.9 + ], + "angle": 0, + "content": "[114] Shaofei Wang, Katja Schwarz, Andreas Geiger, and Siyu Tang. Arah: Animatable volume rendering of articulated human sdfs. In European conference on computer vision, 2022. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.147 + ], + "angle": 0, + "content": "[115] Chung-Yi Weng, Brian Curless, and Ira Kemelmacher-Shlizerman. Vid2actor: Free-viewpoint animatable person synthesis from video in the wild. arXiv preprint arXiv:2012.12884, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.894, + 0.232 + ], + "angle": 0, + "content": "[116] Chung-Yi Weng, Brian Curless, Pratul P. Srinivasan, Jonathan T. Barron, and Ira Kemelmacher-Shlizerman. HumanNeRF: Free-viewpoint rendering of moving people from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16210-16220, 2022. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.233, + 0.894, + 0.274 + ], + "angle": 0, + "content": "[117] Zhenzhen Weng, Zeyu Wang, and Serena Yeung. Zero-)rator: Zero-shot 3d avatar generation from a single image. arXiv preprint arXiv:2305.16411, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.275, + 0.894, + 0.342 + ], + "angle": 0, + "content": "[118] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. Synsin: End-to-end view synthesis from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7467-7477, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.344, + 0.894, + 0.399 + ], + "angle": 0, + "content": "[119] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.401, + 0.894, + 0.455 + ], + "angle": 0, + "content": "[120] Hongyi Xu, Thiemo Alldieck, and Cristian Sminchisescu. H-nerf: Neural radiance fields for rendering and temporal reconstruction of humans in motion. Advances in Neural Information Processing Systems, 34, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.456, + 0.894, + 0.525 + ], + "angle": 0, + "content": "[121] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5438–5448, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.526, + 0.894, + 0.58 + ], + "angle": 0, + "content": "[122] Zhen Xu, Sida Peng, Haotong Lin, Guangzhao He, Jiaming Sun, Yujun Shen, Hujun Bao, and Xiaowei Zhou. 4k4d: Real-time 4d view synthesis at 4k resolution. arXiv preprint arXiv:2310.11448, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.581, + 0.894, + 0.649 + ], + "angle": 0, + "content": "[123] Jiayu Yang, Wei Mao, Jose M Alvarez, and Miaomiao Liu. Cost volume pyramid based depth inference for multi-view stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4877-4886, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.65, + 0.894, + 0.734 + ], + "angle": 0, + "content": "[124] Ze Yang, Shenlong Wang, Sivabalan Manivasagam, Zeng Huang, Wei-Chiu Ma, Xinchen Yan, Ersin Yumer, and Raquel Urtasun. S3: Neural shape, skeleton, and skinn-ning fields for 3d human modeling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13284–13293, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.735, + 0.894, + 0.789 + ], + "angle": 0, + "content": "[125] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.79, + 0.894, + 0.844 + ], + "angle": 0, + "content": "[126] Zeyu Yang, Hongye Yang, Zijie Pan, Xiatian Zhu, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. arXiv preprint arXiv:2310.10642, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.894, + 0.9 + ], + "angle": 0, + "content": "[127] Wang Yifan, Felice Serena, Shihao Wu, Cengiz Öztireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM Transactions on Graphics (TOG), 38(6):1-14, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20430" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.161 + ], + "angle": 0, + "content": "[128] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoptrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5752-5761, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.163, + 0.472, + 0.204 + ], + "angle": 0, + "content": "[129] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelNeRF: Neural radiance fields from one or few images. In CVPR, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.207, + 0.472, + 0.248 + ], + "angle": 0, + "content": "[130] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.25, + 0.472, + 0.305 + ], + "angle": 0, + "content": "[131] Qiang Zhang, Seung-Hwan Baek, Szymon Rusinkiewicz, and Felix Heide. Differentiable point-based radiance fields for efficient view synthesis. In SIGGRAPH Asia 2022 Conference Papers, pages 1-12, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.306, + 0.472, + 0.361 + ], + "angle": 0, + "content": "[132] Fuqiang Zhao, Wei Yang, Jiakai Zhang, Pei Lin, Yingliang Zhang, Jingyi Yu, and Lan Xu. Humannerf: Generalizable neural human radiance field from sparse inputs. arXiv preprint arXiv:2112.02789, 3, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.363, + 0.472, + 0.432 + ], + "angle": 0, + "content": "[133] Tinghui Zhou, Shubham Tulsiani, Weilun Sun, Jitendra Malik, and Alexei A Efros. View synthesis by appearance flow. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 286-301. Springer, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.434, + 0.472, + 0.487 + ], + "angle": 0, + "content": "[134] Wojciech Zielonka, Timur Bagautdinov, Shunsuke Saito, Michael Zollhöfer, Justus Thies, and Javier Romero. Drivable 3d gaussian avatars. arXiv preprint arXiv:2311.08581, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.489, + 0.472, + 0.543 + ], + "angle": 0, + "content": "[135] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Ewa volume splatting. In Proceedings Visualization, 2001. VIS'01., pages 29-538. IEEE, 2001. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.546, + 0.472, + 0.602 + ], + "angle": 0, + "content": "[136] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Surface splatting. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 371-378, 2001. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.602 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20431" + } + ] +] \ No newline at end of file diff --git a/2024/GauHuman_ Articulated Gaussian Splatting from Monocular Human Videos/8c2dc361-8082-42c7-a19c-2ac9cac55b4b_origin.pdf b/2024/GauHuman_ Articulated Gaussian Splatting from Monocular Human Videos/8c2dc361-8082-42c7-a19c-2ac9cac55b4b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..aac70e00f92d1fd3b8ca874501bedfcb3967df4d --- /dev/null +++ b/2024/GauHuman_ Articulated Gaussian Splatting from Monocular Human Videos/8c2dc361-8082-42c7-a19c-2ac9cac55b4b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4468f27336303ee445b380490522d26eaca7219901994587a84ab3c46656a991 +size 2255472 diff --git a/2024/GauHuman_ Articulated Gaussian Splatting from Monocular Human Videos/full.md b/2024/GauHuman_ Articulated Gaussian Splatting from Monocular Human Videos/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9f66752b05c1f3a161a567bc5500c914f1af95cb --- /dev/null +++ b/2024/GauHuman_ Articulated Gaussian Splatting from Monocular Human Videos/full.md @@ -0,0 +1,358 @@ +# GauHuman: Articulated Gaussian Splitting from Monocular Human Videos + +Shoukang Hu Tao Hu Ziwei Liu S-Lab, Nanyang Technological University + +![](images/617cdbc8d1353f1e11596549b67dfe83def79e76cab9d7154ad1b9d0ba764791.jpg) +Figure 1. SauHuman models 3D humans with fast training (1~2 minutes) and real-time rendering (~189 FPS) given monocular videos. For illustration purposes, we select one sequence (1024p) from MonoCap [36, 37, 83] and one sequence (1024p) from DNA-Rendering [15]. + +# Abstract + +We present, GauHuman, a 3D human model with Gaussian Splitting for both fast training (1~2 minutes) and real-time rendering (up to 189 FPS), compared with existing NeRF-based implicit representation modelling frameworks demanding hours of training and seconds of rendering per frame. Specifically, GauHuman encodes Gaussian Splitting in the canonical space and transforms 3D Gaussians from canonical space to posed space with linear blend skinning (LBS), in which effective pose and LBS refinement modules are designed to learn fine details of 3D humans under negligible computational cost. Moreover, to enable fast optimization of GauHuman, we initialize and prune 3D Gaussians with 3D human prior, while splitting/cloning via KL divergence guidance, along with a novel merge operation for further speeding up. Extensive experiments on ZJU_Mocap and MonoCap datasets demonstrate that GauHuman achieves state-of-the-art performance quantitatively and qualitatively with fast training and real-time rendering speed. Notably, without sacrificing rendering quality, GauHuman can fast model the 3D human performer with $\sim 13k$ 3D Gaussians. Our code is available at https://github.com/skhu101/GauHuman. + +# 1. Introduction + +Creating high-quality 3D human performers holds extensive utility across various domains, including AR/VR, telepresence, video games, and movie production. Recent methods show 3D human avatar can be learned from sparse-view videos [12, 48, 50, 80, 82, 83, 103, 114, 116, 120] or even a single image [8, 43, 65, 117] with NeRF-based implicit representation [76]. However, these methods typically require expensive time and computational costs for both training and rendering, which hinders their applications in real-world scenarios. For example, it normally takes about 10 GPU hours to learn a 3D human performer with a rendering speed of less than 1 frame per second (FPS). + +To speed up 3D human modelling, generalizable Human NeRF methods [17, 28, 42, 60, 132] trade the quality for less training time. They usually fine-tune new human performers with models pre-trained on articulated human data. However, such an inefficient pretrain-and-finetune paradigm would normally take several hours of pre-training to obtain generalizable representations for 3D humans and an additional one hour of finetuning for each human, while achieving low rendering quality due to the incomplete information within limited conditional input images. On the contrary, a recent line of research focuses on learning efficient 3D human representations [30, 49] to improve training speed. For example, instead of learning neural radiance field (NeRF) from + +coordinate-based neural networks, methods like [30, 49] incorporate 3D human modelling with multi-resolution hash encoding (MHE) [78] to accelerate the training. Although these methods can reduce the training time by a magnitude via human prior like part-based human representations, their rendering speed still limits their applications. A critical bottleneck hindering their training and rendering speed arises from the inefficiency of the backward mapping-based ray-cast rendering algorithm employed in these methods, especially at higher resolutions, e.g., 1080p and 4k. + +To mitigate this time-consuming dilemma, recent research [56, 131, 135] in point-based rendering can fast render both static and dynamic scenes [72, 119, 125, 126] by splatting a restricted set of 3D points. Motivated by these great successes, it is desirable to explore Gaussian Splatting for real-time rendering of articulated objects such as 3D humans with monocular videos. Yet, this promising avenue comes with two primary barriers: 1) How to properly integrate articulated human information from monocular videos in the Gaussian Splatting framework? 2) How to efficiently optimize Gaussian Splatting for fast training? In particular, one potential solution for 3D human modelling with Gaussian Splatting is to directly utilize LBS transformation to transform 3D Gaussians from canonical space into posed space. However, it is infeasible to directly apply a 3D vertex point-based LBS transformation for 3D Gaussians, with inaccurate LBS weights and pose information derived from human prior models such as SMPL [71]. Meanwhile, current Gaussian Splatting methods utilize either sparse point cloud produced from Structure-from-Motion (SfM) [98, 101] or random point cloud to initialize 3D Gaussians. This initialization is specifically designed for static scenes, rather than 3D human structures, making it unsuitable for effective and fast human modelling. Also, adaptive control of 3D Gaussians proposed in [56] mainly utilizes gradient information of 3D position to split and clone Gaussians for over-reconstruction and under-reconstruction. Despite improving the rendering performance, it trivially generates a large number of 3D Gaussians, which causes difficulty in optimization and consumes a large storage memory (up to $734\mathrm{MB}$ in a static scene reconstruction). More importantly, the Gaussian Splatting framework contains plenty of redundant closely-positioned Gaussians, which holds a potential for additional acceleration. + +To tackle these two challenges, we propose GauHuman, a 3D human model with articulated Gaussian Splatting representation for both fast training and rendering. Our modelling framework and fast optimization algorithm are detailed as follows. Articulated Gaussian Splatting with integrated human information. Inspired by previous Human NeRF methods, we encode Gaussian Splatting in a canonical space and wrap it to a posed space. Thanks to the Gaussian property, we can transform 3D Gaussians from a canonical space + +to a posed space via linear blend skinning (LBS). We rotate and translate the 3D position and covariance of each 3D Gaussian with the estimated LBS transformation matrix. However, directly applying multi-layer perceptron (MLP) to estimate LBS weights would be time-consuming and generate low-quality rendering results. To address this, we start with the LBS weights from SMPL [71] and employ an MLP to predict corresponding LBS weight offsets, followed by an effective pose refinement module to rectify SMPL poses for more accurate LBS transformation. For rendering an image with $1024\mathrm{p}$ , our GauHuman only needs to splat a limited set (e.g., 13k) of 3D Gaussians, in comparison with Human NeRF methods that apply volume rendering to millions of sampled points in the whole 3D space. This significantly reduces the computational burden in both training and rendering processes. Fast Optimization of 3D Gaussian Splatting. 1) 3D Gaussian Initialization. In GauHuman, we utilize the human prior like SMPL for 3D Gaussian initialization, which enables fast optimization of articulated Gaussian Splatting as the initialization encodes human structures. 2) Split/Clone/Merge. Our analysis indicates that the splitting and cloning process overlooks a crucial metric, namely, the distance between 3D Gaussians, during the split and clone process. In light of this, we integrate the Kullback-Leibler (KL) divergence of 3D Gaussians to regulate the split and clone process. A novel merge operator is also proposed to merge 3D Gaussians that are close to each other for further speeding up. 3) Prune. As 3D human has specific structures (human prior), we prune 3D Gaussians that are far away from the human body, e.g., SMPL vertices. With the above algorithm implemented in articulated 3D Gaussian Splatting, we can fast model 3D human performers with $\sim 13\mathrm{k}$ 3D Gaussians (3.5MB memory storage) within one or two minutes. Our main contributions are as follows: + +1. We present, GauHuman, a 3D model with articulated Gaussian Splatting for both fast training (1~2 minutes) and real-time rendering (up to 189 FPS), in which pose and LBS refinement modules are designed to learn fine details of 3D humans. +2. To enable fast optimization of articulated 3D Gaussian Splitting, we initialize and prune 3D Gaussians with 3D human prior, while splitting/cloning via KL divergence measures, accompanied by a novel merge operation for further speeding up. +3. Experiments on two monocular data sets validate that our GauHuman achieves state-of-the-art novel view synthesis performance with fast training and real-time rendering. + +# 2. Related Work + +Rendering and Radiance Fields. Early novel view synthesis approaches focus on using light fields for densely sampled [32, 62] and unstructured [7] capture. In later years, the advent of Structure-from-Motion (SfM) [101] enables a new + +domain by synthesizing novel views from a collection of photos with blending techniques [9, 23, 31, 38, 58, 107]. With the development of deep learning techniques, previous research [24, 38, 91, 108, 133] propose to adopt deep learning techniques with multi-view stereo (MVS) for novel view synthesis. Recently, deep learning techniques have been used to learn implicit neural radiance fields [39, 76, 84, 100] with volume rendering [55, 73]. NeRF [76] is one of the representative work, learning high-quality novel view synthesis results, which has inspired a series of follow-up works [4, 10, 16, 19, 21, 45, 46, 53, 69, 78, 79, 87, 110, 112, 113, 121, 123, 129, 130]. To accelerate training and rendering speed, several methods [11, 25, 29, 64, 67, 78, 86, 105, 106, 128] are proposed, including replacing multi-layer perceptrons (MLPs) with more efficient representations or skipping the empty space via the occupancy grid. In contrast to NeRF, which applies volume rendering to millions of sampled points in the whole 3D space when rendering an image, point-based rendering [2, 33, 34, 59, 61, 92, 93, 131] splats a restricted set of points. This approach provides flexibility and efficiency in reconstructing 3D scenes by optimizing the opacity and position of individual points. Furthermore, point primitives can be represented as circular, elliptic discs, ellipsoids or surfels for high quality results [6, 85, 89, 118, 127, 136]. To achieve real-time rendering of large-scale scenes, 3D Gaussians with splatting techniques are used to represent static scenes[56], dynamic scenes [72, 119, 125, 126], and other tasks [90, 102, 111]. In this work, we further extend Gaussian Splatting to articulated 3D humans with both fast training and real-time rendering. + +3D Human Modelling. Although 2D human modelling [1, 26, 27, 51, 52, 63, 74, 97] has made substantial progress, it is still challenging to reconstruct high-quality 3D humans for synthesizing free-viewpoint videos. Traditional methods either leverage multi-view stereo methods [35, 99] or depth fusion [18, 22, 104] to reconstruct 3D human geometry, which typically requires dense camera arrays or depth sensors. To alleviate this, previous research [3, 17, 28, 40-42, 60, 94, 95, 132] learn human priors from data sets containing either 3D ground truth human models or multiview image collections, which enables them to synthesize novel views even with a single image. However, due to the limited diversity in pre-training data sets, these methods can not generalize well to 3D humans with complex poses. Another line of works reconstruct 3D human performers with implicit neural representations from sparse-view videos [12, 48, 50, 66, 68, 80, 82, 83, 83, 103, 114, 116, 120] or even a single image [8, 43, 65, 117]. However, these methods usually require hours of training time and seconds of rendering time per frame, restricting their applications to real-world scenarios. To improve the training efficiency, recent research [30, 49] apply multi hashing encoding (MHE) + +to reduce the training time to several minutes, while their rendering remains comparatively slow. Moreover, neural volumetric primitives [14, 70, 88] have been proposed for efficient rendering of 3D humans, but they still require hours of training time. In this work, we focus on reconstructing high-quality 3D humans from monocular videos with both fast training and real-time rendering. Our concurrent works [47, 54, 57, 77, 122, 134] learn real-time rendering of 3D humans from either dense-view or monocular videos. + +# 3. Our Approach + +The goal of this paper is to learn 3D human performers with fast training and rendering from monocular videos. We assume the calibrated camera parameters and the human region masks are known. We also assume the corresponding SMPL parameters $\beta, \theta$ are given. + +# 3.1. Preliminary + +SMPL [71] is a parametric human model $M(\beta, \theta)$ , which defines $\beta, \theta$ to control body shapes and poses. In this work, we apply the Linear Blend Skinning (LBS) algorithm used in SMPL to transform points from a canonical space to a posed space. For example, 3D point $p^c$ in the canonical space is transformed to a posed space defined by pose $\theta$ as $p^t = \sum_{k=1}^{K} w_k (G_k(J, \theta)p^c + b_k(J, \theta, \beta))$ , where $J$ includes $K$ joint locations, $G_k(J, \theta)$ and $b_k(J, \theta, \beta)$ are the transformation matrix and translation vector of joint $k$ respectively, $w_k$ is the linear blend weight. + +3D Gaussian Splatting [56] In comparison with NeRF methods that learn continuous structured volumetric representation for the whole 3D space, Gaussian Splatting represents a 3D scene with a set of unstructured discrete 3D Gaussians. Each 3D Gaussian is defined with a point center $\pmb{p}$ and a full 3D covariance matrix $\Sigma$ in a world space, i.e., + +$$ +G (\boldsymbol {x}) = \frac {1}{(2 \pi) ^ {\frac {3}{2}} | \boldsymbol {\Sigma} | ^ {\frac {1}{2}}} e ^ {- \frac {1}{2} (\boldsymbol {x} - \boldsymbol {p}) ^ {T} \boldsymbol {\Sigma} ^ {- 1} (\boldsymbol {x} - \boldsymbol {p})}, \tag {1} +$$ + +where the covariance matrix $\boldsymbol{\Sigma}$ is decomposed as $\boldsymbol{\Sigma} = \boldsymbol{R}\boldsymbol{S}\boldsymbol{S}^T\boldsymbol{R}^T$ , rotation matrix $\boldsymbol{R}$ and scaling matrix $\boldsymbol{S}$ are represented with a quaternion $\boldsymbol{r} \in \mathcal{R}^4$ and a scaling vector $\boldsymbol{s} \in \mathcal{R}^3$ respectively. + +To splot 3D Gaussians, 3D Gaussian Splatting [135, 136] first transforms them from world space to ray coordinate and then projects them to a 2D screen plane. As shown in [135], utilizing a world-to-camera transformation matrix $\mathbf{W}$ , translation vector $\mathbf{d}$ , and the Jacobian matrix $\mathbf{J}$ representing an approximated projective transformation from camera coordinate to ray coordinate, 3D Gaussians in ray coordinate assume the following form: + +$$ +G (\boldsymbol {x}) = \frac {1}{(2 \pi) ^ {\frac {3}{2}} | \boldsymbol {\Sigma} ^ {\prime} | ^ {\frac {1}{2}}} e ^ {- \frac {1}{2} (\boldsymbol {x} - \boldsymbol {p} ^ {\prime}) ^ {T} \boldsymbol {\Sigma} ^ {\prime} ^ {- 1} (\boldsymbol {x} - \boldsymbol {p} ^ {\prime})} \tag {2} +$$ + +where $\pmb{p}^{\prime} = m(\pmb {W}\pmb {p} + \pmb {d})$ and $\boldsymbol {\Sigma}' = \boldsymbol {J}\boldsymbol {W}\boldsymbol {\Sigma}\boldsymbol{W}^{T}\boldsymbol{J}^{T}$ are + +![](images/a8ec7681a1cda701a7a20c550df198e9af56e459ce98838a1af63b1818cd95f5.jpg) +Figure 2. GauHuman Framework. We first initialize 3D Gaussians' positions $\pmb{p}^{c}$ from SMPL vertex points. Then we incorporate pose refinement module $\mathrm{MLP}_{\Phi_{\mathrm{pose}}}(\cdot)$ and LBS weight field $\mathrm{MLP}_{\Phi_{\mathrm{lbs}}}(\cdot)$ module to learn the LBS transformation to transform 3D Gaussians from canonical space to posed space. During optimization, a tile-based differentiable rasterizer is applied to enable fast rendering. To adaptively control the number of 3D Gaussians, we further propose to use human prior (e.g., SMPL) and KL divergence measure to regulate the split, clone, merge, and prune process. + +the mean vector and covariance matrix of 3D Gaussians in ray coordinate, $m(\cdot)$ denotes the mapping from camera coordinate to ray coordinate. The projected 2D Gaussians in the 2D screen plane can then be obtained by integrating Eqn. (2) along the projection direction or utilizing a property mentioned by [135]. + +After mapping 3D Gaussians to a 2D screen space, we count 2D Gaussians that overlap with each pixel and calculate their color $c_{i}$ and density $\alpha_{i}$ contribution. The colors of each pixel can be obtained by blending $N$ ordered Gaussians: + +$$ +\hat {C} = \sum_ {i \in N} c _ {i} \alpha_ {i} \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {i}). \tag {3} +$$ + +During optimization, 3D Gaussians are first initialized from either Structure-from-Motion (SfM) or random sampling. Then adaptive control of Gaussians is employed to improve the rendering quality, which mainly includes three operations, i.e., split, clone, and prune. The split and clone operations are performed on 3D Gaussians with large position gradients. These 3D Gaussians would be 1) split into smaller Gaussians if the magnitude of the scaling matrix is larger than a threshold or 2) cloned if the magnitude of the scaling matrix is smaller than a threshold. While splitting and cloning augments the number of Gaussians, the prune operation selectively eliminates Gaussians with either excessively small opacity $(\alpha)$ or overly large scaling magnitudes. + +# 3.2. Articulated 3D Gaussian Splatting + +Recent research can fast render static scenes [56] or dynamic scenes [72, 119, 125, 126] by splatting a restricted set of 3D points. Motivated by these great successes, it is desirable to + +explore Gaussian Splatting for real-time rendering of articulated objects such as 3D humans with monocular videos. When employing Gaussian Splatting to articulated 3D humans, there are two main challenges, i.e., 1) How to properly integrate articulated human information from monocular videos in the Gaussian Splatting framework? ii). how to efficiently optimize Gaussian Splatting for fast training? + +Inspired by previous Human NeRF research, as shown in Fig. 2, we represent 3D humans with 3D Gaussians in a canonical space and then wrap 3D Gaussians from the canonical space to posed space through forward LBS transformation. Thanks to the Gaussian property, we can transform 3D Gaussians from a canonical space to a posed space via linear blend skinning (LBS). Specifically, we rotate and translate the 3D position and covariance of each 3D Gaussian with the estimated LBS transformation matrix. + +$$ +\boldsymbol {p} ^ {t} = \boldsymbol {G} (\boldsymbol {J} ^ {t}, \boldsymbol {\theta} ^ {t}) \boldsymbol {p} ^ {c} + \boldsymbol {b} (\boldsymbol {J} ^ {t}, \boldsymbol {\theta} ^ {t}, \beta^ {t}) +$$ + +$$ +\boldsymbol {\Sigma} ^ {t} = \boldsymbol {G} (\boldsymbol {J} ^ {t}, \boldsymbol {\theta} ^ {t}) \boldsymbol {\Sigma} ^ {c} \boldsymbol {G} (\boldsymbol {J} ^ {t}, \boldsymbol {\theta} ^ {t}) ^ {T}, \tag {4} +$$ + +where $\pmb{p}^t, \pmb{\Sigma}^t, \pmb{p}^c$ , and $\pmb{\Sigma}^c$ are the 3D position vector and covariance matrix in the canonical and posed space respectively. $\pmb{G}(\pmb{J}^t, \pmb{\theta}^t) = \sum_{k=1}^{K} w_k \pmb{G}_k(\pmb{J}^t, \pmb{\theta}^t)$ , $\pmb{b}(\pmb{J}^t, \pmb{\theta}^t, \pmb{\beta}^t) = \sum_{k=1}^{K} w_k \pmb{b}_k(\pmb{J}^t, \pmb{\theta}^t, \pmb{\beta}^t)$ are the rotation matrix and translation vector. $K$ is the joint number, $\pmb{G}_k(\pmb{J}^t, \pmb{\theta}^t)$ and $\pmb{b}_k(\pmb{J}^t, \pmb{\theta}^t, \pmb{\beta}^t)$ are the transformation matrix and translation vector of joint $k$ respectively, $w_k$ is the linear blend skinning (LBS) weight we need to estimate. + +LBS Weight Field Multi-layer perceptrons (MLPs) can be used to predict the LBS weight coefficients $w_{k}$ for each 3D Gaussian in the canonical space. However, it would be time-consuming and generate low-quality rendering results. To address this, we start with the LBS weights from SMPL and + +employ an $\mathrm{MLP}_{\Phi_{\mathrm{Ibs}}}(\cdot)$ to predict corresponding LBS weight offsets. Specifically, for each 3D Gaussian, we add the LBS weight $w_{k}^{\mathrm{SMPL}}$ of nearest SMPL vertex with the predicted offsets $\mathrm{MLP}_{\Phi_{\mathrm{Ibs}}}(\gamma (\pmb{p}^{c}))$ , i.e., + +$$ +w _ {k} = \frac {e ^ {\log \left(w _ {k} ^ {\mathrm {S M P L}} + 1 0 ^ {- 8}\right) + \mathrm {M L P} _ {\Phi_ {\mathrm {l b s}}} \left(\gamma \left(\boldsymbol {p} ^ {c}\right)\right) [ k ]}}{\sum_ {k = 1} ^ {K} e ^ {\log \left(w _ {k} ^ {\mathrm {S M P L}} + 1 0 ^ {- 8}\right) + \mathrm {M L P} _ {\Phi_ {\mathrm {l b s}}} \left(\gamma \left(\boldsymbol {p} ^ {c}\right)\right) [ k ]}}, \tag {5} +$$ + +where we apply sinusoidal positional encoding $\gamma (\cdot)$ [76] to the input (3D position $p^c$ ) of $\mathrm{MLP}_{\Phi_{\mathrm{lbs}}(\cdot)}$ . The idea of learning LBS weight field is also applied in previous human modelling works [5, 13, 20, 44, 75, 81, 96, 109, 115, 124]. + +Pose Refinement The body poses $\theta$ estimated from images is often inaccurate, which degrades the rendering quality. To address this, we incorporate a pose correction module $\mathrm{MLP}_{\Phi_{\mathrm{pose}}}(\cdot)$ to learn relative updates to joint angles: + +$$ +\boldsymbol {\theta} = \boldsymbol {\theta} ^ {\mathrm {S M P L}} \otimes \mathrm {M L P} _ {\Phi_ {\text {p o s e}}} (\boldsymbol {\theta}), \tag {6} +$$ + +where $\theta^{\mathrm{SMPL}}$ is the SMPL body pose parameters estimated from images. The body pose refinement is also investigated in [12, 49, 116]. + +During inference, thanks to the 3D Gaussian-based representation defined in a canonical space, the learned LBS weight for each 3D Gaussian can be derived from the LBS weight field module in advance, which saves computation time. Meanwhile, the rectified pose parameters can also be saved from the pose refinement module at the end of training. Considering the above, our GauHuman can model high-quality 3D humans with high rendering speed. + +# 3.3. Fast Optimization of GauHuman + +The current optimization scheme of Gaussian Splitting [56] is specifically designed for static scenes. To facilitate fast optimization for articulated 3D humans, we incorporate 3D human prior and KL divergence measures to speed up. + +3D Gaussian Initialization Current Gaussian Splatting methods utilize either Structure-from-Motion (SfM) [98, 101] or random point cloud to initialize 3D Gaussians, which overlooks 3D human structures. In GauHuman, we initialize 3D Gaussians with human prior like SMPL vertex points. The above initialization encodes human structure, thus accelerating optimization. Note that for 3D Gaussian covariance initialization, we follow the default setting in [56]. + +Split & Clone & Merge & Prune As mentioned in Sec. 3.1, the optimization of Gaussian Splatting relies on the adaptive control of 3D Gaussians to ensure superior rendering quality. To adaptively control the number of Gaussians, three operations, i.e., split, clone, and prune, are performed based on the position gradients and magnitude of the scaling matrix. However, although split and clone operations substantially improve the performance, they trivially generate a substantial quantity of 3D Gaussians. This results in a slowdown of the optimization and a significant consumption of storage memory, reaching up to $734\mathrm{MB}$ in static scenes. Our + +analysis indicates that previous research overlooks a crucial metric, namely, the distance between 3D Gaussians, during the split and clone process. For example, if two Gaussians are close to each other, even if the position gradients are larger than a threshold, they should not be split or cloned as 1) these Gaussians are updating their positions to improve their performance, 2) splitting and cloning these Gaussians have negligible influence on the final performance as they are too close to each other. To regulate the split and clone process, we propose to use Kullback-Leibler (KL) divergence as a measure of the distance of 3D Gaussians, i.e., + +$$ +\begin{array}{l} K L \left(G \left(\boldsymbol {x} _ {0}\right) \mid G \left(\boldsymbol {x} _ {1}\right)\right) = \frac {1}{2} \left(t r \left(\boldsymbol {\Sigma} _ {1} ^ {- 1} \boldsymbol {\Sigma} _ {0}\right) + \ln \frac {\det \boldsymbol {\Sigma} _ {1}}{\det \boldsymbol {\Sigma} _ {2}} \right. \tag {7} \\ + \left(\boldsymbol {p} _ {1} - \boldsymbol {p} _ {0}\right) ^ {T} \boldsymbol {\Sigma} _ {1} ^ {- 1} \left(\boldsymbol {p} _ {1} - \boldsymbol {p} _ {0}\right) - 3), \\ \end{array} +$$ + +where $\pmb{p}_0, \pmb{\Sigma}_0, \pmb{p}_1, \pmb{\Sigma}_1$ are the position and covariance matrix of two 3D Gaussians $G(\pmb{x}_0)$ and $G(\pmb{x}_1)$ . + +After obtaining the KL divergence of nearby 3D Gaussian pairs, we select 3D Gaussians with large KL divergence and positional gradients to perform the split and clone. However, the above requires calculating the KL divergence for each pair of 3D Gaussians, which consumes lots of time during the optimization. To solve this problem, we further adopt two approaches to simplify the whole calculation. 1) For each 3D Gaussian, we first identify its closest 3D Gaussian through the widely adopted k-nearest neighbor (k-NN) algorithm, which assesses the distance between their respective centers. Then we calculate a KL divergence for each pair of nearby 3D Gaussians. The aforementioned simplification effectively reduces the time complexity from $O(U^2)$ to $O(U)$ , where $U$ represents the total number of 3D Gaussians. 2) As the covariance matrix is decomposed into the product of rotation and scaling matrices, we further simplify the computation of matrix inverse and determinant operations in Eqn. (7) with the diagonal and orthogonal property of rotation and scaling matrices. Details are shown in the appendix. + +In addition to split and clone operations, we further propose a novel merge operation, which merges redundant 3D Gaussians with 1) large position gradients, 2) small scaling magnitude, and 3) KL divergence less than 0.1. Specifically, we merge two Gaussians by averaging their positions, opacity, and SH coefficients. Additionally, we scale the covariance (scaling) matrix of the first Gaussian by a factor of 1.25 as the initialization for the new Gaussian. For prune operation, we incorporate 3D human prior like SMPL to prune 3D Gaussians far away from SMPL vertex points. + +During optimization, four loss functions are used, i.e., + +$$ +\mathcal {L} = \mathcal {L} _ {\text {c o l o r}} + \lambda_ {1} \mathcal {L} _ {\text {m a s k}} + \lambda_ {2} \mathcal {L} _ {\text {S S I M}} + \lambda_ {3} \mathcal {L} _ {\text {L P I P S}}, \tag {8} +$$ + +where $\lambda$ 's are loss weights. Empirically, we set $\lambda_1 = 0.5$ , $\lambda_2 = \lambda_3 = 0.01$ to ensure the same magnitude for each loss. + +Table 1. Quantitative comparison of our GauHuman and baseline methods on the ZJU_MoCap and MonoCap data sets. We use bold and underlined text to denote the best and second-best results of each metric respectively. Our GauHuman achieves the state-of-the-art PSNR and LPIPS with fast training (1~2 minutes) and real-time rendering (up to 189 FPS) in both data sets. Note that PixelNeRF [129] and NHP [60] are pre-trained for an additional 10 hours. $\dagger$ denotes the fine-tuning time for generalizable methods. LPIPS\* = 1000 × LPIPS. Frames per second (FPS) is measured on an RTX 3090. For a fair comparison, we do not conduct test-time optimization of SMPL parameters with images from the test set on InstantAvatar [49]. + +
MethodZJU_MoCapMonoCap
PSNR↑SSIM↑LPIPS*↓TrainFPSPSNR↑SSIM↑LPIPS*↓TrainFPS
PixelNeRF [129]24.710.892121.86\( 1h^† \)1.2026.430.96043.98\( 1h^† \)0.75
NHP [60]28.250.95564.77\( 1h^† \)0.1530.510.98027.14\( 1h^† \)0.05
NB [82]29.030.96442.4710h1.4832.360.98616.7010h0.98
AN [81]29.770.96546.8910h1.1131.070.98519.4710h0.31
AS [83]30.380.97537.2310h0.4032.480.98813.1810h0.29
HumanNeRF [116]30.660.96933.3810h0.3032.680.98715.5210h0.08
DVA [88]29.450.95637.741.5h16.532.990.98315.831.5h10.5
InstantNVR [30]31.010.97138.455m2.2032.610.98816.6810m0.75
InstantAvatar [49]29.730.93868.413m4.1530.790.96439.756m2.54
GauHuman(Ours)31.340.96530.511m18933.450.98513.352m154
+ +# 4. Experiments + +# 4.1. Experimental Setup + +Datasets. Experiments are conducted on two widely used human modelling data sets, i.e., ZJU_MoCap [82] and MonoCap [36, 37, 83]. For ZJU_MoCap, same as previous work [30, 116], we select 6 human subjects (377, 386, 387, 392, 393, 394) to conduct experiments. We also adopt the same training and testing setting as [30], i.e., one camera is used for training, while the remaining cameras are used for evaluation. For each subject, we sample 1 frame every 5 frames and collect 100 frames for training. The foreground masks, camera, and SMPL parameters provided by the data set are used for evaluation purposes. MonoCap data set contains four multi-view videos collected by [83] from the DeepCap data set [36] and the DynaCap data set [37]. Similar to ZJU_MoCap, it provides human masks, cameras, and SMPL parameters. We adopt the same camera views as [83] for training and testing. For each subject, we collect 100 frames for training by sampling 1 frame every 5 frames. More details are shown in the appendix. + +Comparison Methods. We compare our GauHuman with two categories of state-of-the-art human modelling methods, i.e., subject-specific optimization based methods [30, 49, 81-83, 88, 116] and generalizable methods [60, 129]. Following [30], we report the average metric values of all selected human subjects. For a fair comparison, we do not conduct test-time optimization of SMPL parameters with images from the test set on InstantAvatar [49]. More details are shown in the appendix. + +# 4.2. Quantitative Results + +Tab. 1 summarizes the overall novel view synthesis results of our GauHuman, NB [82], AN [81], AS [83], HumanNeRF [116], DVA [88], InstantNVR [30], InstantAvatar [49], + +PixelNeRF [129], and NHP [60]. Thanks to our proposed articulated 3D human Gaussian-based representation, we can fast optimize it for around 1 or 2 minutes to achieve photorealistic rendering results, while baseline methods (i.e., NB, AN, AS, and HumanNeRF) take around 10 hours to finish the training and generalizable methods takes about 1 hour for fine-tuning after 10 hours' pre-training. To accelerate the training of 3D human modelling, InstantNVR [30] and InstantAvatar [49] reduce the training to several minutes with multi hashing encoding representation. However, their rendering speed still restricts their applications in real-world applications. When compared with these two state-of-the-art methods for fasting training, our GauHuman achieves slightly better performance with 3 times faster training and 45 times faster rendering speed. Among baseline methods, DVA achieves a fast rendering speed (up to 16.5 FPS) with efficient mixtures of volumetric primitives. However, in contrast to our GauHuman, DVA takes about $90\mathrm{x}$ more time to converge and $11.45\mathrm{x}$ more time to render one image. + +# 4.3. Qualitative Results + +As shown in Fig. 3, our GauHuman shows better appearance and geometry details than DVA [88], NB [82] and NHP [60]. Although NB [82] and NHP [60] show impressive results with 4-view videos, they do not perform well with monocular videos. NB [82] encodes human structure information with latent codes positioned in SMPL vertex's locations, which fails to capture complex human motions. NHP [60] does not generalize well in monocular videos due to the information missing from input images. Thanks to the LBS weight field and deformation field learned in HumanNeRF [116] and AS [83], they achieve comparable visualization results as our GauHuman for $600\mathrm{x}$ more time for training and $472\mathrm{x}$ more time for rendering. InstantNVR [30] and InstantAvatar [49] reduce the training time to several minutes with a multi + +![](images/eed4e7ea3430f589fe608b28c4a70010fb51cad4eab29e1ffd345e9e177286f7.jpg) +Figure 3. Novel view synthesis results produced by our GauHuman and baseline methods on ZJU_MoCap and MonoCap. The bottom lines show the training time and rendering speed of each method on ZJU_Mocap. Zoom in for the best view. + +hashing encoding representation, but they produce slightly worse visualization results than ours and require $45\mathrm{x}$ more rendering time. DVA [] improves the rendering speed to 16.5 FPS with efficient mixtures of volumetric primitives. However, it fails to produce high-quality results and produces obvious artifacts on garments and faces, due to incomplete information within conditional input images and a lack of mechanisms to adaptively control the number of voxels. + +# 4.4. Ablation Study + +We perform ablation studies with the sequence of 386 from the ZJU_Mocap data set. To highlight the effectiveness of each proposed component, we project the 3D human bounding box to each camera plane to obtain the bounding box mask and report these metrics based on the masked area. + +Pose Refinement and LBS Weight Field. Tab. 2 shows the results of ablating pose refinement and LBS weight field modules. All models are trained in around 1 minute. As shown in Tab. 2, pose refinement and LBS weight field modules significantly improve all metrics with negligible computational cost. The visualization results in Fig. 4 show that these two modules can help erase artifacts and learn fine details like cloth wrinkles. For example, we observe obvious + +artifacts in hands and fewer cloth wrinkles without using either one of these two modules. + +Table 2. Quantitative Results of ablating pose refinement and LBS weight field modules. $\mathrm{LPIPS^{*}} = 1000\times$ LPIPS. + +
PSNR↑SSIM↑LPIPS*↓TrainFPS
Ours (full model)28.080.886103.355s189
Ours w/o pose refine27.940.884108.845s192
Ours w/o LBS field27.870.882108.952s190
+ +3D Gaussian Initialization. In Gaussian Splitting, previous research [56] indicates that initialization plays an important role in the final results. For static scenes, there are two commonly adopted initialization approaches, i.e., random initialization or initialization with sparse points from Structure-from-Motion (SfM). To incorporate human articulated information into the initialization, we propose to use articulated vertex points from SMPL to initialize our 3D Gaussians in the canonical space. We compare our proposed articulated initialization with random initialization1 in Tab. 3. With the same number of 3D Gaussians at initialization, our proposed articulated initialization produces both better + +![](images/7b42fc5011e011a098790b971ec59fa731997588929a8c3cf3a387886253f1a6.jpg) +Figure 4. Qualitative results of ablation study on the 386 sequence of ZJU_Mocap data set. "w/o pose refinement" and "w/o LBS weight field" mean that we do not adopt the pose refinement or LBS weight field module respectively. "w/ random Init (#Gau 6890)" and "w/ random Init (#Gau 30000)" mean that we use random initialization to initialize 6890 or 30000 3D Gaussians respectively. + +quantitative and qualitative results than random initialization as shown in Tab. 3 and Fig. 4. Notably, our articulated initialization converges x10 times much faster than the random initialization, which illustrates the strength of encoding human articulated information into the initialization. For random initialization, we need to initialize 30k Gaussians to achieve comparable performance as our proposed articulated initialization, while at cost of 30x times more training time. Table 3. Quantitative Results of ablating 3D Gaussian initialization. #Gau denotes the number of 3D Gaussians at initialization. + +
PSNR↑SSIM↑LPIPS*↓Train
Articulated Init (w/ #Gau 6890)28.080.886103.355s
Random Init (w/ #Gau 6890)27.980.882108.95m
Random Init (w/ #Gau 30k)28.080.884104.330m
+ +Split/Clone/Merge/Prune. Adaptive control of 3D Gaussians with split, clone, and prune operations substantially improves the rendering quality, but it trivially leads to a large number of 3D Gaussians (up to $100\mathrm{k}$ ), which slows down the whole optimization process. As mentioned in Sec. 3.3, we propose to use the KL divergence measure to regulate the split and clone process. As shown in Tab. 4, our KL-based split and clone operations lead to $60\mathrm{x}$ times faster convergence speed when compared with the original split and clone operation proposed in [56]. Furthermore, our novel merge operation merges about $8.5\%$ of 3D Gaussians without degrading the performance. Articulated prune operation with SMPL prior introduced in Sec. 3.3 also helps prune about $0.5\mathrm{k}$ 3D Gaussians far away from the 3D human body. + +Table 4. Quantitative Results of ablating our proposed KL-based split/clone operations and a novel merge operation. #Gau denotes the final number of 3D Gaussians during optimization. + +
PSNR↑SSIM↑LPIPS*↓#GauTrain
Ours (full model)28.080.886103.31316255s
w/o KL split/clone28.010.885103.1100k1h
w/o KL merge28.080.886103.61437670s
w/o articulated prune28.080.886103.41368757s
+ +# 5. Discussion and Conclusion + +To conclude, we propose GauHuman, a 3D human model with Gaussian Splating representation for fast training $(1\sim 2$ minutes) and real-time rendering (166 FPS) of 3D humans. Specifically, we encode this representation in a canonical space and transform 3D Gaussian from canonical space to posed space with LBS transformation, in which effective pose refinement and LBS weight field modules are also designed to learn fine details of 3D humans. To enable fast optimization, we initialize and prune 3D Gaussians with 3D human prior, while splitting/cloning 3D Gaussians via KI divergence guidance, accompanied by a novel merge operation for further speeding up. Experiments on monocular videos suggest that our GauHuman achieves state-of-the-art performance while maintaining both fast training and real-time rendering speed. In addition, GauHuman can fast learn 3D human performers with $\sim 13k$ Gaussians. + +Limitations. 1). GauHuman is composed of 3D Gaussians and current pipeline does not support extracting 3D meshes. How to extract meshes from 3D Gaussians remains a future work to be investigated. 2). Recovering 3D humans with loose/complex clothing from monocular videos is still a challenging problem to tackle. One direction is to incorporate physical simulation of garments into current pipelines. + +Acknowledgment. This study is supported by the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOET2EP20221-0012), NTU NAP, and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). We would like to thank Zhaoxi Chen, Wei Li, Guangcong Wang from NTU, Xurong Xie from ISCAS, Zhiyuan Yu from ZJU, and Zihao Huang from HUST for the discussions on the paper draft. + +# References + +[1] Badour AlBahar, Jingwan Lu, Jimei Yang, Zhixin Shu, Eli Shechtman, and Jia-Bin Huang. Pose with style: Detail-preserving pose-guided image synthesis with conditional stylegan. ACM Transactions on Graphics (TOG), 40(6): 1-11, 2021. 3 +[2] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 696-712. Springer, 2020. 3 +[3] Thiemo Alldieck, Mihai Zanfir, and Cristian Sminchisescu. Photorealistic monocular 3d reconstruction of humans wearing clothing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1506-1515, 2022. 3 +[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 3 +[5] Bharat Lal Bhatnagar, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. Loopreg: Self-supervised learning of implicit surface correspondences, pose and shape for 3d human mesh registration. Advances in Neural Information Processing Systems, 33:12909-12922, 2020. 5 +[6] Mario Botsch, Alexander Hornung, Matthias Zwicker, and Leif Kobbelt. High-quality surface splatting on today's gpus. In Proceedings Eurographics/IEEE VGTC Symposium Point-Based Graphics, 2005., pages 17-141. IEEE, 2005. 3 +[7] Chris Buehler, Michael Bosse, Leonard McMillan, Steven Gortler, and Michael Cohen. Unstructured lumigraph rendering. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 497-504. 2023. 2 +[8] Sihun Cha, Kwanggyoon Seo, Amirsaman Ashtari, and Jun-yong Noh. Generating texture for 3d human avatar from a single image using sampling and refinement networks. In Computer Graphics Forum, pages 385-396. Wiley Online Library, 2023. 1, 3 +[9] Gaurav Chaurasia, Sylvain Duchene, Olga Sorkine-Hornung, and George Drettakis. Depth synthesis and local warps for plausible image-based navigation. ACM Transactions on Graphics (TOG), 32(3):1-12, 2013. 3 +[10] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. arXiv preprint arXiv:2103.15595, 2021. 3 +[11] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 3 +[12] Jianchuan Chen, Ying Zhang, Di Kang, Xuefei Zhe, Linchao Bao, Xu Jia, and Huchuan Lu. Animatable neural radiance fields from monocular rgb videos. arXiv preprint arXiv:2106.13629, 2021. 1, 3, 5 +[13] Xu Chen, Yufeng Zheng, Michael J Black, Otmar Hilliges, and Andreas Geiger. Snarf: Differentiable forward skin + +ning for animating non-rigid neural implicit shapes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11594-11604, 2021. 5 +[14] Zhaoxi Chen, Fangzhou Hong, Haiyi Mei, Guangcong Wang, Lei Yang, and Ziwei Liu. Primdiffusion: Volumetric primitives diffusion for 3d human generation. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 3 +[15] Wei Cheng, Ruixiang Chen, Siming Fan, Wanqi Yin, Keyu Chen, Zhongang Cai, Jingbo Wang, Yang Gao, Zhengming Yu, Zhengyu Lin, et al. Dna-rendering: A diverse neural actor repository for high-fidelity human-centric rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19982-19993, 2023. 1 +[16] Julian Chibane, Aayush Bansal, Verica Lazova, and Gerard Pons-Moll. Stereo radiance fields (srf): Learning view synthesis for sparse views of novel scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7911-7920, 2021. 3 +[17] Hongsuk Choi, Gyeongsik Moon, Matthieu Armando, Vincent Leroy, Kyoung Mu Lee, and Grégory Rogez. Mononhr: Monocular neural human renderer. In 2022 International Conference on 3D Vision (3DV), pages 242-251. IEEE, 2022. 1, 3 +[18] Alvaro Collet, Ming Chuang, Pat Sweeney, Don Gillett, Dennis Evseev, David Calabrese, Hugues Hoppe, Adam Kirk, and Steve Sullivan. High-quality streamable free-viewpoint video. ACM Transactions on Graphics (ToG), 34 (4):1-13, 2015. 3 +[19] François Darmon, Bénédicte Bascle, Jean-Clement Devaux, Pascal Monasse, and Mathieu Aubry. Improving neural implicit surfaces geometry with patch warping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6260-6269, 2022. 3 +[20] Boyang Deng, John P Lewis, Timothy Jeruzalski, Gerard Pons-Moll, Geoffrey Hinton, Mohammad Norouzi, and Andrea Tagliasacchi. Nasa neural articulated shape approximation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VII 16, pages 612-628. Springer, 2020. 5 +[21] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nef: Fewer views and faster training for free. arXiv preprint arXiv:2107.02791, 2021. 3 +[22] Mingsong Dou, Sameh Khamis, Yury Degtyarev, Philip Davidson, Sean Ryan Fanello, Adarsh Kowdle, Sergio Orts Escolano, Christoph Rhemann, David Kim, Jonathan Taylor, et al. Fusion4d: Real-time performance capture of challenging scenes. ACM Transactions on Graphics (ToG), 35(4): 1-13, 2016. 3 +[23] Martin Eisemann, Bert De Decker, Marcus Magnor, Philippe Bekaert, Edilson De Aguiar, Naveed Ahmed, Christian Theobalt, and Anita Sellent. Floating textures. In Computer graphics forum, pages 409-418. Wiley Online Library, 2008. 3 +[24] John Flynn, Ivan Neulander, James Philbin, and Noah Snavely. Deepstereo: Learning to predict new views from the world's imagery. In Proceedings of the IEEE conference + +on computer vision and pattern recognition, pages 5515-5524, 2016.3 +[25] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 3 +[26] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Chen Qian, Chen Change Loy, Wayne Wu, and Ziwei Liu. Stylegan-human: A data-centric odyssey of human generation. In European Conference on Computer Vision, pages 1-19. Springer, 2022. 3 +[27] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Wayne Wu, and Ziwei Liu. Unitedhuman: Harnessing multi-source data for high-resolution human generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7301-7311, 2023. 3 +[28] Xiangjun Gao, Jiaolong Yang, Jongyoo Kim, Sida Peng, Zicheng Liu, and Xin Tong. Mps-nerf: Generalizable 3d human rendering from multiview images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1, 3 +[29] Stephan J Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14346-14355, 2021. 3 +[30] Chen Geng, Sida Peng, Zhen Xu, Hujun Bao, and Xiaowei Zhou. Learning neural volumetric representations of dynamic humans in minutes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8759-8770, 2023. 1, 2, 3, 6 +[31] Michael Goesele, Noah Snavely, Brian Curless, Hugues Hoppe, and Steven M Seitz. Multi-view stereo for community photo collections. In 2007 IEEE 11th International Conference on Computer Vision, pages 1-8. IEEE, 2007. 3 +[32] Steven J Gortler, Radek Grzeszcuk, Richard Szeliski, and Michael F Cohen. The lumigraph. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 453-464. 2023. 2 +[33] Markus Gross and Hanspeter Pfister. Point-based graphics. Elsevier, 2011. 3 +[34] Jeffrey P Grossman and William J Dally. Point sample rendering. In Rendering Techniques' 98: Proceedings of the Eurographics Workshop in Vienna, Austria, June 29—July 1, 1998 9, pages 181–192. Springer, 1998. 3 +[35] Kaiwen Guo, Peter Lincoln, Philip Davidson, Jay Busch, Xueming Yu, Matt Whalen, Geoff Harvey, Sergio Orts-Escolano, Rohit Pandey, Jason Dourgarian, et al. The relightables: Volumetric performance capture of humans with realistic relighting. ACM Transactions on Graphics (ToG), 38(6):1-19, 2019. 3 +[36] Marc Habermann, Weipeng Xu, Michael Zollhofer, Gerard Pons-Moll, and Christian Theobalt. Deepcap: Monocular human performance capture using weak supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5052-5063, 2020. 1, 6 + +[37] Marc Habermann, Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Real-time deep dynamic characters. ACM Transactions on Graphics (ToG), 40(4):1-16, 2021. 1, 6 +[38] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering. ACM Transactions on Graphics (ToG), 37(6):1-15, 2018. 3 +[39] Philipp Henzler, Niloy J Mitra, and Tobias Ritschel. Escaping plato's cave: 3d shape from adversarial rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9984-9993, 2019. 3 +[40] Fangzhou Hong, Zhaoxi Chen, Yushi Lan, Liang Pan, and Ziwei Liu. Eva3d: Compositional 3d human generation from 2d image collections. arXiv preprint arXiv:2210.04888, 2022.3 +[41] Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. arXiv preprint arXiv:2205.08535, 2022. +[42] Shoukang Hu, Fangzhou Hong, Liang Pan, Haiyi Mei, Lei Yang, and Ziwei Liu. Sherf: Generalizable human nerf from a single image. arXiv preprint arXiv:2303.12791, 2023. 1, 3 +[43] Yangyi Huang, Hongwei Yi, Yuliang Xiu, Tingting Liao, Jiaxiang Tang, Deng Cai, and Justus Thies. Tech: Text-guided reconstruction of lifelike clothed humans. arXiv preprint arXiv:2308.08545, 2023. 1, 3 +[44] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. Arch: Animatable reconstruction of clothed humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3093-3102, 2020. 5 +[45] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5885-5894, 2021. 3 +[46] Wonbong Jang and Lourdes Agapito. Codenerf: Disentangled neural radiance fields for object categories. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12949-12958, 2021. 3 +[47] Rohit Jena, Ganesh Subramanian Iyer, Siddharth Choudhary, Brandon Smith, Pratik Chaudhari, and James Gee. Splatarmor: Articulated gaussian splatting for animatable humans from monocular rgb videos. arXiv preprint arXiv:2311.10812, 2023. 3 +[48] Boyi Jiang, Yang Hong, Hujun Bao, and Juyong Zhang. Self-con: Self reconstruction your digital avatar from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5605-5615, 2022. 1, 3 +[49] Tianjian Jiang, Xu Chen, Jie Song, and Otmar Hilliges. Instantavatar: Learning avatars from monocular video in 60 seconds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16922-16932, 2023. 1, 2, 3, 5, 6 +[50] Wei Jiang, Kwang Moo Yi, Golnoosh Samei, Oncel Tuzel, and Anurag Ranjan. Neuman: Neural human radiance field + +from a single video. arXiv preprint arXiv:2203.12575, 2022. 1, 3 +[51] Yuming Jiang, Shuai Yang, Haonan Qiu, Wayne Wu, Chen Change Loy, and Ziwei Liu. Text2human: Text-driven controllable human image generation. ACM Transactions on Graphics (TOG), 41(4):1-11, 2022. 3 +[52] Yuming Jiang, Shuai Yang, Tong Liang Koh, Wayne Wu, Chen Change Loy, and Ziwei Liu. Text2performer: Text-driven human video generation. arXiv preprint arXiv:2304.08483, 2023. 3 +[53] Mohammad Mahdi Johari, Yann Lepoittevin, and François Fleuret. Geonerf: Generalizing nerf with geometry priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18365-18375, 2022. 3 +[54] HyunJun Jung, Nikolas Brasch, Jifei Song, Eduardo Perez-Pellitero, Yiren Zhou, Zhihao Li, Nassir Navab, and Benjamin Busam. Deformable 3d gaussian splatting for animatable human avatars. arXiv preprint arXiv:2312.15059, 2023. 3 +[55] James T Kajiya and Brian P Von Herzen. Ray tracing volume densities. ACM SIGGRAPH computer graphics, 18(3):165-174, 1984. 3 +[56] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 42(4):1-14, 2023. 2, 3, 4, 5, 7, 8 +[57] Muhammed Kocabas, Jen-Hao Rick Chang, James Gabriel, Oncel Tuzel, and Anurag Ranjan. Hugs: Human gaussian splats. arXiv preprint arXiv:2311.17910, 2023. 3 +[58] Georgios Kopanas, Julien Philip, Thomas Leimkuhler, and George Drettakis. Point-based neural rendering with perview optimization. In Computer Graphics Forum, pages 29-43. Wiley Online Library, 2021. 3 +[59] Georgios Kopanas, Thomas Leimkuhler, Gilles Rainer, Clément Jambon, and George Drettakis. Neural point catauustics for novel-view synthesis of reflections. ACM Transactions on Graphics (TOG), 41(6):1-15, 2022. 3 +[60] Youngjoong Kwon, Dahun Kim, Duygu Ceylan, and Henry Fuchs. Neural human performer: Learning generalizable radiance fields for human performance rendering. Advances in Neural Information Processing Systems, 34:24741-24752, 2021. 1, 3, 6 +[61] Christoph Lassner and Michael Zollhofer. Pulsar: Efficient sphere-based neural rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1440-1449, 2021. 3 +[62] Marc Levoy and Pat Hanrahan. Light field rendering. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 441-452. 2023. 2 +[63] Kathleen M Lewis, Srivatsan Varadharajan, and Ira Kemelmacher-Shlizerman. Tryongan: Body-aware try-on via layered interpolation. ACM Transactions on Graphics (TOG), 40(4):1-10, 2021. 3 +[64] Ruilong Li, Matthew Tancik, and Angjoo Kanazawa. Nerve: A general nef acceleration toolbox. arXiv preprint arXiv:2210.04847, 2022. 3 + +[65] Tingting Liao, Xiaomei Zhang, Yuliang Xiu, Hongwei Yi, Xudong Liu, Guo-Jun Qi, Yong Zhang, Xuan Wang, Xi-angyu Zhu, and Zhen Lei. High-fidelity clothed avatar reconstruction from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8662–8672, 2023. 1, 3 +[66] Haotong Lin, Sida Peng, Zhen Xu, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Efficient neural radiance fields for interactive free-viewpoint video. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 3 +[67] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 3 +[68] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. ACM transactions on graphics (TOG), 40(6):1-16, 2021. 3 +[69] Yuan Liu, Sida Peng, Lingjie Liu, Qianqian Wang, Peng Wang, Christian Theobalt, Xiaowei Zhou, and Wenping Wang. Neural rays for occlusion-aware image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7824-7833, 2022. 3 +[70] Stephen Lombardi, Tomas Simon, Gabriel Schwartz, Michael Zollhoefer, Yaser Sheikh, and Jason Saragih. Mixture of volumetric primitives for efficient neural rendering. ACM Transactions on Graphics (ToG), 40(4):1-13, 2021. 3 +[71] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi-person linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia), 34(6):248:1-248:16, 2015. 2, 3 +[72] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 2, 3, 4 +[73] Nelson Max. Optical models for direct volume rendering. IEEE Transactions on Visualization and Computer Graphics, 1(2):99-108, 1995. 3 +[74] Yifang Men, Yiming Mao, Yuning Jiang, Wei-Ying Ma, and Zhouhui Lian. Controllable person image synthesis with attribute-decomposed gan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5084-5093, 2020. 3 +[75] Marko Mihajlovic, Yan Zhang, Michael J Black, and Siyu Tang. Leap: Learning articulated occupancy of people. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10461-10471, 2021. 5 +[76] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 3, 5 +[77] Arthur Moreau, Jifei Song, Helisa Dhamo, Richard Shaw, Yiren Zhou, and Eduardo Pérez-Pellitero. Human gaussian + +splatting: Real-time rendering of animatable avatars. arXiv preprint arXiv:2311.17113, 2023. 3 +[78] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 2, 3 +[79] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. arXiv preprint arXiv:2112.00724, 2021. 3 +[80] Atsuhiro Noguchi, Xiao Sun, Stephen Lin, and Tatsuya Harada. Neural articulated radiance field. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5762-5772, 2021. 1, 3 +[81] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Animatable neural radiance fields for modeling dynamic human bodies. In ICCV, 2021. 5, 6 +[82] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9054-9063, 2021. 1, 3, 6 +[83] Sida Peng, Shangzhan Zhang, Zhen Xu, Chen Geng, Boyi Jiang, Hujun Bao, and Xiaowei Zhou. Animatable neural implicit surfaces for creating avatars from videos. arXiv preprint arXiv:2203.08133, 4(5), 2022. 1, 3, 6 +[84] Eric Penner and Li Zhang. Soft 3d reconstruction for view synthesis. ACM Transactions on Graphics (TOG), 36(6): 1-11, 2017. 3 +[85] Hanspeter Pfister, Matthias Zwicker, Jeroen Van Baar, and Markus Gross. Surfels: Surface elements as rendering primitives. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 335-342, 2000. 3 +[86] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14335-14345, 2021. 3 +[87] Konstantinos Rematas, Ricardo Martin-Brualla, and Vittorio Ferrari. Sharf: Shape-conditioned radiance fields from a single view. arXiv preprint arXiv:2102.08860, 2021. 3 +[88] Edoardo Remelli, Timur Bagautdinov, Shunsuke Saito, Chenglei Wu, Tomas Simon, Shih-En Wei, Kaiwen Guo, Zhe Cao, Fabian Prada, Jason Saragih, et al. Drivable volumetric avatars using texel-aligned features. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1–9, 2022. 3, 6 +[89] Liu Ren, Hanspeter Pfister, and Matthias Zwicker. Object space ewa surface splatting: A hardware accelerated approach to high quality point rendering. In Computer Graphics Forum, pages 461-470. Wiley Online Library, 2002. 3 +[90] Helge Rhodin, Nadia Robertini, Christian Richardt, Hans-Peter Seidel, and Christian Theobalt. A versatile scene + +model with differentiable visibility applied to generative pose estimation. In Proceedings of the IEEE International Conference on Computer Vision, pages 765-773, 2015. 3 +[91] Gernot Riegler and Vladlen Koltun. Free view synthesis. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIX 16, pages 623-640. Springer, 2020. 3 +[92] Darius Rückert, Linus Franke, and Marc Stamminger. Adop: Approximate differentiable one-pixel point rendering. ACM Transactions on Graphics (ToG), 41(4):1-14, 2022. 3 +[93] Miguel Sainz and Renato Pajarola. Point-based rendering techniques. Computers & Graphics, 28(6):869-879, 2004. 3 +[94] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2304-2314, 2019. 3 +[95] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 84-93, 2020. 3 +[96] Shunsuke Saito, Jinlong Yang, Qianli Ma, and Michael J Black. Scanimate: Weakly supervised learning of skinned clothed avatar networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2886-2897, 2021. 5 +[97] Kripasindhu Sarkar, Lingjie Liu, Vladislav Golyanik, and Christian Theobalt. Humangan: A generative model of human images. In 2021 International Conference on 3D Vision (3DV), pages 258-267. IEEE, 2021. 3 +[98] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 2, 5 +[99] Johannes L Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 501-518. Springer, 2016. 3 +[100] Vincent Sitzmann, Justus Thies, Felix Heide, Matthias Nießner, Gordon Wetzstein, and Michael Zollhofer. Deepvoxels: Learning persistent 3d feature embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2437-2446, 2019. 3 +[101] Noah Snavely, Steven M Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. In ACM siggraph 2006 papers, pages 835-846. 2006. 2, 5 +[102] Carsten Stoll, Nils Hasler, Juergen Gall, Hans-Peter Seidel, and Christian Theobalt. Fast articulated motion tracking using a sums of gaussians body model. In 2011 International Conference on Computer Vision, pages 951-958. IEEE, 2011. 3 +[103] Shih-Yang Su, Frank Yu, Michael Zollhöfer, and Helge Rhodin. A-nerf: Articulated neural radiance fields for learning human shape, appearance, and pose. Advances in Neural + +Information Processing Systems, 34:12278-12291, 2021. 1, 3 +[104] Zhuo Su, Lan Xu, Zerong Zheng, Tao Yu, Yebin Liu, and Lu Fang. Robustfusion: Human volumetric capture with data-driven visual cues using a rgbd camera. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16, pages 246-264. Springer, 2020. 3 +[105] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 3 +[106] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3d shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11358-11367, 2021. 3 +[107] Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srinivasan, Edgar Tretschk, Wang Yifan, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lombardi, et al. Advances in neural rendering. In Computer Graphics Forum, pages 703-735. Wiley Online Library, 2022. 3 +[108] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: Image synthesis using neural textures. Acm Transactions on Graphics (TOG), 38(4):1-12, 2019. 3 +[109] Garvita Tiwari, Nikolaos Sarafianos, Tony Tung, and Gerard Pons-Moll. Neural-gif: Neural generalized implicit functions for animating people in clothing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11708-11718, 2021. 5 +[110] Alex Trevithick and Bo Yang. Grf: Learning a general radiance field for 3d representation and rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15182-15192, 2021. 3 +[111] Angtian Wang, Peng Wang, Jian Sun, Adam Kortylewski, and Alan Yuille. Voge: a differentiable volume renderer using gaussian ellipsoids for analysis-by-synthesis. In The Eleventh International Conference on Learning Representations, 2022. 3 +[112] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 3 +[113] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snively, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690-4699, 2021. 3 +[114] Shaofei Wang, Katja Schwarz, Andreas Geiger, and Siyu Tang. Arah: Animatable volume rendering of articulated human sdfs. In European conference on computer vision, 2022. 1, 3 + +[115] Chung-Yi Weng, Brian Curless, and Ira Kemelmacher-Shlizerman. Vid2actor: Free-viewpoint animatable person synthesis from video in the wild. arXiv preprint arXiv:2012.12884, 2020. 5 +[116] Chung-Yi Weng, Brian Curless, Pratul P. Srinivasan, Jonathan T. Barron, and Ira Kemelmacher-Shlizerman. HumanNeRF: Free-viewpoint rendering of moving people from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16210-16220, 2022. 1, 3, 5, 6 +[117] Zhenzhen Weng, Zeyu Wang, and Serena Yeung. Zero-)rator: Zero-shot 3d avatar generation from a single image. arXiv preprint arXiv:2305.16411, 2023. 1, 3 +[118] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. Synsin: End-to-end view synthesis from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7467-7477, 2020. 3 +[119] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528, 2023. 2, 3, 4 +[120] Hongyi Xu, Thiemo Alldieck, and Cristian Sminchisescu. H-nerf: Neural radiance fields for rendering and temporal reconstruction of humans in motion. Advances in Neural Information Processing Systems, 34, 2021. 1, 3 +[121] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5438–5448, 2022. 3 +[122] Zhen Xu, Sida Peng, Haotong Lin, Guangzhao He, Jiaming Sun, Yujun Shen, Hujun Bao, and Xiaowei Zhou. 4k4d: Real-time 4d view synthesis at 4k resolution. arXiv preprint arXiv:2310.11448, 2023. 3 +[123] Jiayu Yang, Wei Mao, Jose M Alvarez, and Miaomiao Liu. Cost volume pyramid based depth inference for multi-view stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4877-4886, 2020. 3 +[124] Ze Yang, Shenlong Wang, Sivabalan Manivasagam, Zeng Huang, Wei-Chiu Ma, Xinchen Yan, Ersin Yumer, and Raquel Urtasun. S3: Neural shape, skeleton, and skinn-ning fields for 3d human modeling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13284–13293, 2021. 5 +[125] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101, 2023. 2, 3, 4 +[126] Zeyu Yang, Hongye Yang, Zijie Pan, Xiatian Zhu, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. arXiv preprint arXiv:2310.10642, 2023. 2, 3, 4 +[127] Wang Yifan, Felice Serena, Shihao Wu, Cengiz Öztireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM Transactions on Graphics (TOG), 38(6):1-14, 2019. 3 + +[128] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoptrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5752-5761, 2021. 3 +[129] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelNeRF: Neural radiance fields from one or few images. In CVPR, 2021. 3, 6 +[130] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 3 +[131] Qiang Zhang, Seung-Hwan Baek, Szymon Rusinkiewicz, and Felix Heide. Differentiable point-based radiance fields for efficient view synthesis. In SIGGRAPH Asia 2022 Conference Papers, pages 1-12, 2022. 2, 3 +[132] Fuqiang Zhao, Wei Yang, Jiakai Zhang, Pei Lin, Yingliang Zhang, Jingyi Yu, and Lan Xu. Humannerf: Generalizable neural human radiance field from sparse inputs. arXiv preprint arXiv:2112.02789, 3, 2021. 1, 3 +[133] Tinghui Zhou, Shubham Tulsiani, Weilun Sun, Jitendra Malik, and Alexei A Efros. View synthesis by appearance flow. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 286-301. Springer, 2016. 3 +[134] Wojciech Zielonka, Timur Bagautdinov, Shunsuke Saito, Michael Zollhöfer, Justus Thies, and Javier Romero. Drivable 3d gaussian avatars. arXiv preprint arXiv:2311.08581, 2023. 3 +[135] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Ewa volume splatting. In Proceedings Visualization, 2001. VIS'01., pages 29-538. IEEE, 2001. 2, 3, 4 +[136] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Surface splatting. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 371-378, 2001. 3 \ No newline at end of file diff --git a/2024/GauHuman_ Articulated Gaussian Splatting from Monocular Human Videos/images.zip b/2024/GauHuman_ Articulated Gaussian Splatting from Monocular Human Videos/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..591207096c6d2814bd005a39595ebf8bf72d4681 --- /dev/null +++ b/2024/GauHuman_ Articulated Gaussian Splatting from Monocular Human Videos/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dc76000874ff0cbcdd03b5e05d9020d61983f486dc717e1b424ca7d528a6f5b7 +size 528411 diff --git a/2024/GauHuman_ Articulated Gaussian Splatting from Monocular Human Videos/layout.json b/2024/GauHuman_ Articulated Gaussian Splatting from Monocular Human Videos/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..37000a682c1927e789bee2447d73b85cc6b67229 --- /dev/null +++ b/2024/GauHuman_ Articulated Gaussian Splatting from Monocular Human Videos/layout.json @@ -0,0 +1,10465 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 59, + 103, + 534, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 103, + 534, + 120 + ], + "spans": [ + { + "bbox": [ + 59, + 103, + 534, + 120 + ], + "type": "text", + "content": "GauHuman: Articulated Gaussian Splitting from Monocular Human Videos" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 195, + 144, + 397, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 144, + 397, + 174 + ], + "spans": [ + { + "bbox": [ + 195, + 144, + 397, + 174 + ], + "type": "text", + "content": "Shoukang Hu Tao Hu Ziwei Liu S-Lab, Nanyang Technological University" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 61, + 178, + 534, + 355 + ], + "blocks": [ + { + "bbox": [ + 61, + 178, + 534, + 355 + ], + "lines": [ + { + "bbox": [ + 61, + 178, + 534, + 355 + ], + "spans": [ + { + "bbox": [ + 61, + 178, + 534, + 355 + ], + "type": "image", + "image_path": "617cdbc8d1353f1e11596549b67dfe83def79e76cab9d7154ad1b9d0ba764791.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 356, + 547, + 380 + ], + "lines": [ + { + "bbox": [ + 46, + 356, + 547, + 380 + ], + "spans": [ + { + "bbox": [ + 46, + 356, + 547, + 380 + ], + "type": "text", + "content": "Figure 1. SauHuman models 3D humans with fast training (1~2 minutes) and real-time rendering (~189 FPS) given monocular videos. For illustration purposes, we select one sequence (1024p) from MonoCap [36, 37, 83] and one sequence (1024p) from DNA-Rendering [15]." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 388, + 192, + 401 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 388, + 192, + 401 + ], + "spans": [ + { + "bbox": [ + 143, + 388, + 192, + 401 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 437, + 290, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 437, + 290, + 700 + ], + "spans": [ + { + "bbox": [ + 45, + 437, + 290, + 700 + ], + "type": "text", + "content": "We present, GauHuman, a 3D human model with Gaussian Splitting for both fast training (1~2 minutes) and real-time rendering (up to 189 FPS), compared with existing NeRF-based implicit representation modelling frameworks demanding hours of training and seconds of rendering per frame. Specifically, GauHuman encodes Gaussian Splitting in the canonical space and transforms 3D Gaussians from canonical space to posed space with linear blend skinning (LBS), in which effective pose and LBS refinement modules are designed to learn fine details of 3D humans under negligible computational cost. Moreover, to enable fast optimization of GauHuman, we initialize and prune 3D Gaussians with 3D human prior, while splitting/cloning via KL divergence guidance, along with a novel merge operation for further speeding up. Extensive experiments on ZJU_Mocap and MonoCap datasets demonstrate that GauHuman achieves state-of-the-art performance quantitatively and qualitatively with fast training and real-time rendering speed. Notably, without sacrificing rendering quality, GauHuman can fast model the 3D human performer with " + }, + { + "bbox": [ + 45, + 437, + 290, + 700 + ], + "type": "inline_equation", + "content": "\\sim 13k" + }, + { + "bbox": [ + 45, + 437, + 290, + 700 + ], + "type": "text", + "content": " 3D Gaussians. Our code is available at https://github.com/skhu101/GauHuman." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 307, + 388, + 386, + 400 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 386, + 400 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 386, + 400 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 411, + 547, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 547, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 547, + 555 + ], + "type": "text", + "content": "Creating high-quality 3D human performers holds extensive utility across various domains, including AR/VR, telepresence, video games, and movie production. Recent methods show 3D human avatar can be learned from sparse-view videos [12, 48, 50, 80, 82, 83, 103, 114, 116, 120] or even a single image [8, 43, 65, 117] with NeRF-based implicit representation [76]. However, these methods typically require expensive time and computational costs for both training and rendering, which hinders their applications in real-world scenarios. For example, it normally takes about 10 GPU hours to learn a 3D human performer with a rendering speed of less than 1 frame per second (FPS)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 557, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 548, + 715 + ], + "type": "text", + "content": "To speed up 3D human modelling, generalizable Human NeRF methods [17, 28, 42, 60, 132] trade the quality for less training time. They usually fine-tune new human performers with models pre-trained on articulated human data. However, such an inefficient pretrain-and-finetune paradigm would normally take several hours of pre-training to obtain generalizable representations for 3D humans and an additional one hour of finetuning for each human, while achieving low rendering quality due to the incomplete information within limited conditional input images. On the contrary, a recent line of research focuses on learning efficient 3D human representations [30, 49] to improve training speed. For example, instead of learning neural radiance field (NeRF) from" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20418" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "content": "coordinate-based neural networks, methods like [30, 49] incorporate 3D human modelling with multi-resolution hash encoding (MHE) [78] to accelerate the training. Although these methods can reduce the training time by a magnitude via human prior like part-based human representations, their rendering speed still limits their applications. A critical bottleneck hindering their training and rendering speed arises from the inefficiency of the backward mapping-based ray-cast rendering algorithm employed in these methods, especially at higher resolutions, e.g., 1080p and 4k." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 195, + 289, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 195, + 289, + 602 + ], + "spans": [ + { + "bbox": [ + 46, + 195, + 289, + 602 + ], + "type": "text", + "content": "To mitigate this time-consuming dilemma, recent research [56, 131, 135] in point-based rendering can fast render both static and dynamic scenes [72, 119, 125, 126] by splatting a restricted set of 3D points. Motivated by these great successes, it is desirable to explore Gaussian Splatting for real-time rendering of articulated objects such as 3D humans with monocular videos. Yet, this promising avenue comes with two primary barriers: 1) How to properly integrate articulated human information from monocular videos in the Gaussian Splatting framework? 2) How to efficiently optimize Gaussian Splatting for fast training? In particular, one potential solution for 3D human modelling with Gaussian Splatting is to directly utilize LBS transformation to transform 3D Gaussians from canonical space into posed space. However, it is infeasible to directly apply a 3D vertex point-based LBS transformation for 3D Gaussians, with inaccurate LBS weights and pose information derived from human prior models such as SMPL [71]. Meanwhile, current Gaussian Splatting methods utilize either sparse point cloud produced from Structure-from-Motion (SfM) [98, 101] or random point cloud to initialize 3D Gaussians. This initialization is specifically designed for static scenes, rather than 3D human structures, making it unsuitable for effective and fast human modelling. Also, adaptive control of 3D Gaussians proposed in [56] mainly utilizes gradient information of 3D position to split and clone Gaussians for over-reconstruction and under-reconstruction. Despite improving the rendering performance, it trivially generates a large number of 3D Gaussians, which causes difficulty in optimization and consumes a large storage memory (up to " + }, + { + "bbox": [ + 46, + 195, + 289, + 602 + ], + "type": "inline_equation", + "content": "734\\mathrm{MB}" + }, + { + "bbox": [ + 46, + 195, + 289, + 602 + ], + "type": "text", + "content": " in a static scene reconstruction). More importantly, the Gaussian Splatting framework contains plenty of redundant closely-positioned Gaussians, which holds a potential for additional acceleration." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "content": "To tackle these two challenges, we propose GauHuman, a 3D human model with articulated Gaussian Splatting representation for both fast training and rendering. Our modelling framework and fast optimization algorithm are detailed as follows. Articulated Gaussian Splatting with integrated human information. Inspired by previous Human NeRF methods, we encode Gaussian Splatting in a canonical space and wrap it to a posed space. Thanks to the Gaussian property, we can transform 3D Gaussians from a canonical space" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 547, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 479 + ], + "type": "text", + "content": "to a posed space via linear blend skinning (LBS). We rotate and translate the 3D position and covariance of each 3D Gaussian with the estimated LBS transformation matrix. However, directly applying multi-layer perceptron (MLP) to estimate LBS weights would be time-consuming and generate low-quality rendering results. To address this, we start with the LBS weights from SMPL [71] and employ an MLP to predict corresponding LBS weight offsets, followed by an effective pose refinement module to rectify SMPL poses for more accurate LBS transformation. For rendering an image with " + }, + { + "bbox": [ + 304, + 72, + 547, + 479 + ], + "type": "inline_equation", + "content": "1024\\mathrm{p}" + }, + { + "bbox": [ + 304, + 72, + 547, + 479 + ], + "type": "text", + "content": ", our GauHuman only needs to splat a limited set (e.g., 13k) of 3D Gaussians, in comparison with Human NeRF methods that apply volume rendering to millions of sampled points in the whole 3D space. This significantly reduces the computational burden in both training and rendering processes. Fast Optimization of 3D Gaussian Splatting. 1) 3D Gaussian Initialization. In GauHuman, we utilize the human prior like SMPL for 3D Gaussian initialization, which enables fast optimization of articulated Gaussian Splatting as the initialization encodes human structures. 2) Split/Clone/Merge. Our analysis indicates that the splitting and cloning process overlooks a crucial metric, namely, the distance between 3D Gaussians, during the split and clone process. In light of this, we integrate the Kullback-Leibler (KL) divergence of 3D Gaussians to regulate the split and clone process. A novel merge operator is also proposed to merge 3D Gaussians that are close to each other for further speeding up. 3) Prune. As 3D human has specific structures (human prior), we prune 3D Gaussians that are far away from the human body, e.g., SMPL vertices. With the above algorithm implemented in articulated 3D Gaussian Splatting, we can fast model 3D human performers with " + }, + { + "bbox": [ + 304, + 72, + 547, + 479 + ], + "type": "inline_equation", + "content": "\\sim 13\\mathrm{k}" + }, + { + "bbox": [ + 304, + 72, + 547, + 479 + ], + "type": "text", + "content": " 3D Gaussians (3.5MB memory storage) within one or two minutes. Our main contributions are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 479, + 545, + 634 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 307, + 479, + 545, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 479, + 545, + 537 + ], + "spans": [ + { + "bbox": [ + 307, + 479, + 545, + 537 + ], + "type": "text", + "content": "1. We present, GauHuman, a 3D model with articulated Gaussian Splatting for both fast training (1~2 minutes) and real-time rendering (up to 189 FPS), in which pose and LBS refinement modules are designed to learn fine details of 3D humans." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 539, + 545, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 539, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 306, + 539, + 545, + 598 + ], + "type": "text", + "content": "2. To enable fast optimization of articulated 3D Gaussian Splitting, we initialize and prune 3D Gaussians with 3D human prior, while splitting/cloning via KL divergence measures, accompanied by a novel merge operation for further speeding up." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 599, + 545, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 599, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 306, + 599, + 545, + 634 + ], + "type": "text", + "content": "3. Experiments on two monocular data sets validate that our GauHuman achieves state-of-the-art novel view synthesis performance with fast training and real-time rendering." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 645, + 392, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 645, + 392, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 645, + 392, + 658 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "Rendering and Radiance Fields. Early novel view synthesis approaches focus on using light fields for densely sampled [32, 62] and unstructured [7] capture. In later years, the advent of Structure-from-Motion (SfM) [101] enables a new" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20419" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 444 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 444 + ], + "type": "text", + "content": "domain by synthesizing novel views from a collection of photos with blending techniques [9, 23, 31, 38, 58, 107]. With the development of deep learning techniques, previous research [24, 38, 91, 108, 133] propose to adopt deep learning techniques with multi-view stereo (MVS) for novel view synthesis. Recently, deep learning techniques have been used to learn implicit neural radiance fields [39, 76, 84, 100] with volume rendering [55, 73]. NeRF [76] is one of the representative work, learning high-quality novel view synthesis results, which has inspired a series of follow-up works [4, 10, 16, 19, 21, 45, 46, 53, 69, 78, 79, 87, 110, 112, 113, 121, 123, 129, 130]. To accelerate training and rendering speed, several methods [11, 25, 29, 64, 67, 78, 86, 105, 106, 128] are proposed, including replacing multi-layer perceptrons (MLPs) with more efficient representations or skipping the empty space via the occupancy grid. In contrast to NeRF, which applies volume rendering to millions of sampled points in the whole 3D space when rendering an image, point-based rendering [2, 33, 34, 59, 61, 92, 93, 131] splats a restricted set of points. This approach provides flexibility and efficiency in reconstructing 3D scenes by optimizing the opacity and position of individual points. Furthermore, point primitives can be represented as circular, elliptic discs, ellipsoids or surfels for high quality results [6, 85, 89, 118, 127, 136]. To achieve real-time rendering of large-scale scenes, 3D Gaussians with splatting techniques are used to represent static scenes[56], dynamic scenes [72, 119, 125, 126], and other tasks [90, 102, 111]. In this work, we further extend Gaussian Splatting to articulated 3D humans with both fast training and real-time rendering." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 450, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 450, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 450, + 289, + 714 + ], + "type": "text", + "content": "3D Human Modelling. Although 2D human modelling [1, 26, 27, 51, 52, 63, 74, 97] has made substantial progress, it is still challenging to reconstruct high-quality 3D humans for synthesizing free-viewpoint videos. Traditional methods either leverage multi-view stereo methods [35, 99] or depth fusion [18, 22, 104] to reconstruct 3D human geometry, which typically requires dense camera arrays or depth sensors. To alleviate this, previous research [3, 17, 28, 40-42, 60, 94, 95, 132] learn human priors from data sets containing either 3D ground truth human models or multiview image collections, which enables them to synthesize novel views even with a single image. However, due to the limited diversity in pre-training data sets, these methods can not generalize well to 3D humans with complex poses. Another line of works reconstruct 3D human performers with implicit neural representations from sparse-view videos [12, 48, 50, 66, 68, 80, 82, 83, 83, 103, 114, 116, 120] or even a single image [8, 43, 65, 117]. However, these methods usually require hours of training time and seconds of rendering time per frame, restricting their applications to real-world scenarios. To improve the training efficiency, recent research [30, 49] apply multi hashing encoding (MHE)" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "content": "to reduce the training time to several minutes, while their rendering remains comparatively slow. Moreover, neural volumetric primitives [14, 70, 88] have been proposed for efficient rendering of 3D humans, but they still require hours of training time. In this work, we focus on reconstructing high-quality 3D humans from monocular videos with both fast training and real-time rendering. Our concurrent works [47, 54, 57, 77, 122, 134] learn real-time rendering of 3D humans from either dense-view or monocular videos." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 305, + 191, + 398, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 191, + 398, + 205 + ], + "spans": [ + { + "bbox": [ + 305, + 191, + 398, + 205 + ], + "type": "text", + "content": "3. Our Approach" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 211, + 547, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 211, + 547, + 272 + ], + "spans": [ + { + "bbox": [ + 304, + 211, + 547, + 272 + ], + "type": "text", + "content": "The goal of this paper is to learn 3D human performers with fast training and rendering from monocular videos. We assume the calibrated camera parameters and the human region masks are known. We also assume the corresponding SMPL parameters " + }, + { + "bbox": [ + 304, + 211, + 547, + 272 + ], + "type": "inline_equation", + "content": "\\beta, \\theta" + }, + { + "bbox": [ + 304, + 211, + 547, + 272 + ], + "type": "text", + "content": " are given." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 279, + 386, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 279, + 386, + 293 + ], + "spans": [ + { + "bbox": [ + 305, + 279, + 386, + 293 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 297, + 547, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 297, + 547, + 417 + ], + "spans": [ + { + "bbox": [ + 304, + 297, + 547, + 417 + ], + "type": "text", + "content": "SMPL [71] is a parametric human model " + }, + { + "bbox": [ + 304, + 297, + 547, + 417 + ], + "type": "inline_equation", + "content": "M(\\beta, \\theta)" + }, + { + "bbox": [ + 304, + 297, + 547, + 417 + ], + "type": "text", + "content": ", which defines " + }, + { + "bbox": [ + 304, + 297, + 547, + 417 + ], + "type": "inline_equation", + "content": "\\beta, \\theta" + }, + { + "bbox": [ + 304, + 297, + 547, + 417 + ], + "type": "text", + "content": " to control body shapes and poses. In this work, we apply the Linear Blend Skinning (LBS) algorithm used in SMPL to transform points from a canonical space to a posed space. For example, 3D point " + }, + { + "bbox": [ + 304, + 297, + 547, + 417 + ], + "type": "inline_equation", + "content": "p^c" + }, + { + "bbox": [ + 304, + 297, + 547, + 417 + ], + "type": "text", + "content": " in the canonical space is transformed to a posed space defined by pose " + }, + { + "bbox": [ + 304, + 297, + 547, + 417 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 297, + 547, + 417 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 297, + 547, + 417 + ], + "type": "inline_equation", + "content": "p^t = \\sum_{k=1}^{K} w_k (G_k(J, \\theta)p^c + b_k(J, \\theta, \\beta))" + }, + { + "bbox": [ + 304, + 297, + 547, + 417 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 297, + 547, + 417 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 304, + 297, + 547, + 417 + ], + "type": "text", + "content": " includes " + }, + { + "bbox": [ + 304, + 297, + 547, + 417 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 297, + 547, + 417 + ], + "type": "text", + "content": " joint locations, " + }, + { + "bbox": [ + 304, + 297, + 547, + 417 + ], + "type": "inline_equation", + "content": "G_k(J, \\theta)" + }, + { + "bbox": [ + 304, + 297, + 547, + 417 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 297, + 547, + 417 + ], + "type": "inline_equation", + "content": "b_k(J, \\theta, \\beta)" + }, + { + "bbox": [ + 304, + 297, + 547, + 417 + ], + "type": "text", + "content": " are the transformation matrix and translation vector of joint " + }, + { + "bbox": [ + 304, + 297, + 547, + 417 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 297, + 547, + 417 + ], + "type": "text", + "content": " respectively, " + }, + { + "bbox": [ + 304, + 297, + 547, + 417 + ], + "type": "inline_equation", + "content": "w_k" + }, + { + "bbox": [ + 304, + 297, + 547, + 417 + ], + "type": "text", + "content": " is the linear blend weight." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 418, + 548, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 548, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 548, + 491 + ], + "type": "text", + "content": "3D Gaussian Splatting [56] In comparison with NeRF methods that learn continuous structured volumetric representation for the whole 3D space, Gaussian Splatting represents a 3D scene with a set of unstructured discrete 3D Gaussians. Each 3D Gaussian is defined with a point center " + }, + { + "bbox": [ + 304, + 418, + 548, + 491 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 304, + 418, + 548, + 491 + ], + "type": "text", + "content": " and a full 3D covariance matrix " + }, + { + "bbox": [ + 304, + 418, + 548, + 491 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 304, + 418, + 548, + 491 + ], + "type": "text", + "content": " in a world space, i.e.," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 339, + 493, + 547, + 519 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 493, + 547, + 519 + ], + "spans": [ + { + "bbox": [ + 339, + 493, + 547, + 519 + ], + "type": "interline_equation", + "content": "G (\\boldsymbol {x}) = \\frac {1}{(2 \\pi) ^ {\\frac {3}{2}} | \\boldsymbol {\\Sigma} | ^ {\\frac {1}{2}}} e ^ {- \\frac {1}{2} (\\boldsymbol {x} - \\boldsymbol {p}) ^ {T} \\boldsymbol {\\Sigma} ^ {- 1} (\\boldsymbol {x} - \\boldsymbol {p})}, \\tag {1}", + "image_path": "049778254875d4ea235c7a479934a08b9c49744d1cc8e97ad88db3004ee7091c.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 522, + 547, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 547, + 570 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 547, + 570 + ], + "type": "text", + "content": "where the covariance matrix " + }, + { + "bbox": [ + 304, + 522, + 547, + 570 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\Sigma}" + }, + { + "bbox": [ + 304, + 522, + 547, + 570 + ], + "type": "text", + "content": " is decomposed as " + }, + { + "bbox": [ + 304, + 522, + 547, + 570 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\Sigma} = \\boldsymbol{R}\\boldsymbol{S}\\boldsymbol{S}^T\\boldsymbol{R}^T" + }, + { + "bbox": [ + 304, + 522, + 547, + 570 + ], + "type": "text", + "content": ", rotation matrix " + }, + { + "bbox": [ + 304, + 522, + 547, + 570 + ], + "type": "inline_equation", + "content": "\\boldsymbol{R}" + }, + { + "bbox": [ + 304, + 522, + 547, + 570 + ], + "type": "text", + "content": " and scaling matrix " + }, + { + "bbox": [ + 304, + 522, + 547, + 570 + ], + "type": "inline_equation", + "content": "\\boldsymbol{S}" + }, + { + "bbox": [ + 304, + 522, + 547, + 570 + ], + "type": "text", + "content": " are represented with a quaternion " + }, + { + "bbox": [ + 304, + 522, + 547, + 570 + ], + "type": "inline_equation", + "content": "\\boldsymbol{r} \\in \\mathcal{R}^4" + }, + { + "bbox": [ + 304, + 522, + 547, + 570 + ], + "type": "text", + "content": " and a scaling vector " + }, + { + "bbox": [ + 304, + 522, + 547, + 570 + ], + "type": "inline_equation", + "content": "\\boldsymbol{s} \\in \\mathcal{R}^3" + }, + { + "bbox": [ + 304, + 522, + 547, + 570 + ], + "type": "text", + "content": " respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 570, + 548, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 548, + 666 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 548, + 666 + ], + "type": "text", + "content": "To splot 3D Gaussians, 3D Gaussian Splatting [135, 136] first transforms them from world space to ray coordinate and then projects them to a 2D screen plane. As shown in [135], utilizing a world-to-camera transformation matrix " + }, + { + "bbox": [ + 304, + 570, + 548, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 304, + 570, + 548, + 666 + ], + "type": "text", + "content": ", translation vector " + }, + { + "bbox": [ + 304, + 570, + 548, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 304, + 570, + 548, + 666 + ], + "type": "text", + "content": ", and the Jacobian matrix " + }, + { + "bbox": [ + 304, + 570, + 548, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{J}" + }, + { + "bbox": [ + 304, + 570, + 548, + 666 + ], + "type": "text", + "content": " representing an approximated projective transformation from camera coordinate to ray coordinate, 3D Gaussians in ray coordinate assume the following form:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 334, + 669, + 547, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 669, + 547, + 696 + ], + "spans": [ + { + "bbox": [ + 334, + 669, + 547, + 696 + ], + "type": "interline_equation", + "content": "G (\\boldsymbol {x}) = \\frac {1}{(2 \\pi) ^ {\\frac {3}{2}} | \\boldsymbol {\\Sigma} ^ {\\prime} | ^ {\\frac {1}{2}}} e ^ {- \\frac {1}{2} (\\boldsymbol {x} - \\boldsymbol {p} ^ {\\prime}) ^ {T} \\boldsymbol {\\Sigma} ^ {\\prime} ^ {- 1} (\\boldsymbol {x} - \\boldsymbol {p} ^ {\\prime})} \\tag {2}", + "image_path": "2705ff21e6dff5b2ac95b70ed9479836b3bda7c2371822c88681adef448fc8af.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 700, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 700, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 700, + 547, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 700, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{p}^{\\prime} = m(\\pmb {W}\\pmb {p} + \\pmb {d})" + }, + { + "bbox": [ + 305, + 700, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 700, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\boldsymbol {\\Sigma}' = \\boldsymbol {J}\\boldsymbol {W}\\boldsymbol {\\Sigma}\\boldsymbol{W}^{T}\\boldsymbol{J}^{T}" + }, + { + "bbox": [ + 305, + 700, + 547, + 715 + ], + "type": "text", + "content": " are" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20420" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 64, + 519, + 255 + ], + "blocks": [ + { + "bbox": [ + 67, + 64, + 519, + 255 + ], + "lines": [ + { + "bbox": [ + 67, + 64, + 519, + 255 + ], + "spans": [ + { + "bbox": [ + 67, + 64, + 519, + 255 + ], + "type": "image", + "image_path": "a8ec7681a1cda701a7a20c550df198e9af56e459ce98838a1af63b1818cd95f5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 255, + 548, + 312 + ], + "lines": [ + { + "bbox": [ + 46, + 255, + 548, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 548, + 312 + ], + "type": "text", + "content": "Figure 2. GauHuman Framework. We first initialize 3D Gaussians' positions " + }, + { + "bbox": [ + 46, + 255, + 548, + 312 + ], + "type": "inline_equation", + "content": "\\pmb{p}^{c}" + }, + { + "bbox": [ + 46, + 255, + 548, + 312 + ], + "type": "text", + "content": " from SMPL vertex points. Then we incorporate pose refinement module " + }, + { + "bbox": [ + 46, + 255, + 548, + 312 + ], + "type": "inline_equation", + "content": "\\mathrm{MLP}_{\\Phi_{\\mathrm{pose}}}(\\cdot)" + }, + { + "bbox": [ + 46, + 255, + 548, + 312 + ], + "type": "text", + "content": " and LBS weight field " + }, + { + "bbox": [ + 46, + 255, + 548, + 312 + ], + "type": "inline_equation", + "content": "\\mathrm{MLP}_{\\Phi_{\\mathrm{lbs}}}(\\cdot)" + }, + { + "bbox": [ + 46, + 255, + 548, + 312 + ], + "type": "text", + "content": " module to learn the LBS transformation to transform 3D Gaussians from canonical space to posed space. During optimization, a tile-based differentiable rasterizer is applied to enable fast rendering. To adaptively control the number of 3D Gaussians, we further propose to use human prior (e.g., SMPL) and KL divergence measure to regulate the split, clone, merge, and prune process." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 320, + 288, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 320, + 288, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 320, + 288, + 392 + ], + "type": "text", + "content": "the mean vector and covariance matrix of 3D Gaussians in ray coordinate, " + }, + { + "bbox": [ + 46, + 320, + 288, + 392 + ], + "type": "inline_equation", + "content": "m(\\cdot)" + }, + { + "bbox": [ + 46, + 320, + 288, + 392 + ], + "type": "text", + "content": " denotes the mapping from camera coordinate to ray coordinate. The projected 2D Gaussians in the 2D screen plane can then be obtained by integrating Eqn. (2) along the projection direction or utilizing a property mentioned by [135]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 393, + 288, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 393, + 288, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 393, + 288, + 441 + ], + "type": "text", + "content": "After mapping 3D Gaussians to a 2D screen space, we count 2D Gaussians that overlap with each pixel and calculate their color " + }, + { + "bbox": [ + 46, + 393, + 288, + 441 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 46, + 393, + 288, + 441 + ], + "type": "text", + "content": " and density " + }, + { + "bbox": [ + 46, + 393, + 288, + 441 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 46, + 393, + 288, + 441 + ], + "type": "text", + "content": " contribution. The colors of each pixel can be obtained by blending " + }, + { + "bbox": [ + 46, + 393, + 288, + 441 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 393, + 288, + 441 + ], + "type": "text", + "content": " ordered Gaussians:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 110, + 453, + 288, + 485 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 453, + 288, + 485 + ], + "spans": [ + { + "bbox": [ + 110, + 453, + 288, + 485 + ], + "type": "interline_equation", + "content": "\\hat {C} = \\sum_ {i \\in N} c _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {i}). \\tag {3}", + "image_path": "8f0887a7d9a2da6a77fe7b64305701f868d0d790518c7296ccc9ff3372d27e91.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 493, + 289, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 493, + 289, + 650 + ], + "spans": [ + { + "bbox": [ + 46, + 493, + 289, + 650 + ], + "type": "text", + "content": "During optimization, 3D Gaussians are first initialized from either Structure-from-Motion (SfM) or random sampling. Then adaptive control of Gaussians is employed to improve the rendering quality, which mainly includes three operations, i.e., split, clone, and prune. The split and clone operations are performed on 3D Gaussians with large position gradients. These 3D Gaussians would be 1) split into smaller Gaussians if the magnitude of the scaling matrix is larger than a threshold or 2) cloned if the magnitude of the scaling matrix is smaller than a threshold. While splitting and cloning augments the number of Gaussians, the prune operation selectively eliminates Gaussians with either excessively small opacity " + }, + { + "bbox": [ + 46, + 493, + 289, + 650 + ], + "type": "inline_equation", + "content": "(\\alpha)" + }, + { + "bbox": [ + 46, + 493, + 289, + 650 + ], + "type": "text", + "content": " or overly large scaling magnitudes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 658, + 233, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 233, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 233, + 672 + ], + "type": "text", + "content": "3.2. Articulated 3D Gaussian Splatting" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "Recent research can fast render static scenes [56] or dynamic scenes [72, 119, 125, 126] by splatting a restricted set of 3D points. Motivated by these great successes, it is desirable to" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 320, + 547, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 320, + 547, + 404 + ], + "spans": [ + { + "bbox": [ + 304, + 320, + 547, + 404 + ], + "type": "text", + "content": "explore Gaussian Splatting for real-time rendering of articulated objects such as 3D humans with monocular videos. When employing Gaussian Splatting to articulated 3D humans, there are two main challenges, i.e., 1) How to properly integrate articulated human information from monocular videos in the Gaussian Splatting framework? ii). how to efficiently optimize Gaussian Splatting for fast training?" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 405, + 547, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 405, + 547, + 511 + ], + "spans": [ + { + "bbox": [ + 304, + 405, + 547, + 511 + ], + "type": "text", + "content": "Inspired by previous Human NeRF research, as shown in Fig. 2, we represent 3D humans with 3D Gaussians in a canonical space and then wrap 3D Gaussians from the canonical space to posed space through forward LBS transformation. Thanks to the Gaussian property, we can transform 3D Gaussians from a canonical space to a posed space via linear blend skinning (LBS). Specifically, we rotate and translate the 3D position and covariance of each 3D Gaussian with the estimated LBS transformation matrix." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 354, + 517, + 497, + 530 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 517, + 497, + 530 + ], + "spans": [ + { + "bbox": [ + 354, + 517, + 497, + 530 + ], + "type": "interline_equation", + "content": "\\boldsymbol {p} ^ {t} = \\boldsymbol {G} (\\boldsymbol {J} ^ {t}, \\boldsymbol {\\theta} ^ {t}) \\boldsymbol {p} ^ {c} + \\boldsymbol {b} (\\boldsymbol {J} ^ {t}, \\boldsymbol {\\theta} ^ {t}, \\beta^ {t})", + "image_path": "9814da5f490a71af487cd49369fc73285ebec7ae6509f90ad5bbb5c29b8d1d97.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 355, + 530, + 545, + 547 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 530, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 355, + 530, + 545, + 547 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Sigma} ^ {t} = \\boldsymbol {G} (\\boldsymbol {J} ^ {t}, \\boldsymbol {\\theta} ^ {t}) \\boldsymbol {\\Sigma} ^ {c} \\boldsymbol {G} (\\boldsymbol {J} ^ {t}, \\boldsymbol {\\theta} ^ {t}) ^ {T}, \\tag {4}", + "image_path": "f10898987fd0f5ac9714b59c521f7c01a1ba0a4b3d2c277978f42ca13b1060df.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 555, + 547, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 547, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 547, + 653 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 555, + 547, + 653 + ], + "type": "inline_equation", + "content": "\\pmb{p}^t, \\pmb{\\Sigma}^t, \\pmb{p}^c" + }, + { + "bbox": [ + 304, + 555, + 547, + 653 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 555, + 547, + 653 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}^c" + }, + { + "bbox": [ + 304, + 555, + 547, + 653 + ], + "type": "text", + "content": " are the 3D position vector and covariance matrix in the canonical and posed space respectively. " + }, + { + "bbox": [ + 304, + 555, + 547, + 653 + ], + "type": "inline_equation", + "content": "\\pmb{G}(\\pmb{J}^t, \\pmb{\\theta}^t) = \\sum_{k=1}^{K} w_k \\pmb{G}_k(\\pmb{J}^t, \\pmb{\\theta}^t)" + }, + { + "bbox": [ + 304, + 555, + 547, + 653 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 555, + 547, + 653 + ], + "type": "inline_equation", + "content": "\\pmb{b}(\\pmb{J}^t, \\pmb{\\theta}^t, \\pmb{\\beta}^t) = \\sum_{k=1}^{K} w_k \\pmb{b}_k(\\pmb{J}^t, \\pmb{\\theta}^t, \\pmb{\\beta}^t)" + }, + { + "bbox": [ + 304, + 555, + 547, + 653 + ], + "type": "text", + "content": " are the rotation matrix and translation vector. " + }, + { + "bbox": [ + 304, + 555, + 547, + 653 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 555, + 547, + 653 + ], + "type": "text", + "content": " is the joint number, " + }, + { + "bbox": [ + 304, + 555, + 547, + 653 + ], + "type": "inline_equation", + "content": "\\pmb{G}_k(\\pmb{J}^t, \\pmb{\\theta}^t)" + }, + { + "bbox": [ + 304, + 555, + 547, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 555, + 547, + 653 + ], + "type": "inline_equation", + "content": "\\pmb{b}_k(\\pmb{J}^t, \\pmb{\\theta}^t, \\pmb{\\beta}^t)" + }, + { + "bbox": [ + 304, + 555, + 547, + 653 + ], + "type": "text", + "content": " are the transformation matrix and translation vector of joint " + }, + { + "bbox": [ + 304, + 555, + 547, + 653 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 555, + 547, + 653 + ], + "type": "text", + "content": " respectively, " + }, + { + "bbox": [ + 304, + 555, + 547, + 653 + ], + "type": "inline_equation", + "content": "w_k" + }, + { + "bbox": [ + 304, + 555, + 547, + 653 + ], + "type": "text", + "content": " is the linear blend skinning (LBS) weight we need to estimate." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": "LBS Weight Field Multi-layer perceptrons (MLPs) can be used to predict the LBS weight coefficients " + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "inline_equation", + "content": "w_{k}" + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": " for each 3D Gaussian in the canonical space. However, it would be time-consuming and generate low-quality rendering results. To address this, we start with the LBS weights from SMPL and" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "20421" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 122 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 122 + ], + "type": "text", + "content": "employ an " + }, + { + "bbox": [ + 46, + 72, + 287, + 122 + ], + "type": "inline_equation", + "content": "\\mathrm{MLP}_{\\Phi_{\\mathrm{Ibs}}}(\\cdot)" + }, + { + "bbox": [ + 46, + 72, + 287, + 122 + ], + "type": "text", + "content": " to predict corresponding LBS weight offsets. Specifically, for each 3D Gaussian, we add the LBS weight " + }, + { + "bbox": [ + 46, + 72, + 287, + 122 + ], + "type": "inline_equation", + "content": "w_{k}^{\\mathrm{SMPL}}" + }, + { + "bbox": [ + 46, + 72, + 287, + 122 + ], + "type": "text", + "content": " of nearest SMPL vertex with the predicted offsets " + }, + { + "bbox": [ + 46, + 72, + 287, + 122 + ], + "type": "inline_equation", + "content": "\\mathrm{MLP}_{\\Phi_{\\mathrm{Ibs}}}(\\gamma (\\pmb{p}^{c}))" + }, + { + "bbox": [ + 46, + 72, + 287, + 122 + ], + "type": "text", + "content": ", i.e.," + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 70, + 125, + 287, + 158 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 125, + 287, + 158 + ], + "spans": [ + { + "bbox": [ + 70, + 125, + 287, + 158 + ], + "type": "interline_equation", + "content": "w _ {k} = \\frac {e ^ {\\log \\left(w _ {k} ^ {\\mathrm {S M P L}} + 1 0 ^ {- 8}\\right) + \\mathrm {M L P} _ {\\Phi_ {\\mathrm {l b s}}} \\left(\\gamma \\left(\\boldsymbol {p} ^ {c}\\right)\\right) [ k ]}}{\\sum_ {k = 1} ^ {K} e ^ {\\log \\left(w _ {k} ^ {\\mathrm {S M P L}} + 1 0 ^ {- 8}\\right) + \\mathrm {M L P} _ {\\Phi_ {\\mathrm {l b s}}} \\left(\\gamma \\left(\\boldsymbol {p} ^ {c}\\right)\\right) [ k ]}}, \\tag {5}", + "image_path": "929b6264774c14d778fcf0821e51363ca9b00f1c69b317dcddfec646f73363f3.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 162, + 287, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 162, + 287, + 209 + ], + "spans": [ + { + "bbox": [ + 46, + 162, + 287, + 209 + ], + "type": "text", + "content": "where we apply sinusoidal positional encoding " + }, + { + "bbox": [ + 46, + 162, + 287, + 209 + ], + "type": "inline_equation", + "content": "\\gamma (\\cdot)" + }, + { + "bbox": [ + 46, + 162, + 287, + 209 + ], + "type": "text", + "content": " [76] to the input (3D position " + }, + { + "bbox": [ + 46, + 162, + 287, + 209 + ], + "type": "inline_equation", + "content": "p^c" + }, + { + "bbox": [ + 46, + 162, + 287, + 209 + ], + "type": "text", + "content": ") of " + }, + { + "bbox": [ + 46, + 162, + 287, + 209 + ], + "type": "inline_equation", + "content": "\\mathrm{MLP}_{\\Phi_{\\mathrm{lbs}}(\\cdot)}" + }, + { + "bbox": [ + 46, + 162, + 287, + 209 + ], + "type": "text", + "content": ". The idea of learning LBS weight field is also applied in previous human modelling works [5, 13, 20, 44, 75, 81, 96, 109, 115, 124]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 210, + 288, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 210, + 288, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 210, + 288, + 258 + ], + "type": "text", + "content": "Pose Refinement The body poses " + }, + { + "bbox": [ + 46, + 210, + 288, + 258 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 210, + 288, + 258 + ], + "type": "text", + "content": " estimated from images is often inaccurate, which degrades the rendering quality. To address this, we incorporate a pose correction module " + }, + { + "bbox": [ + 46, + 210, + 288, + 258 + ], + "type": "inline_equation", + "content": "\\mathrm{MLP}_{\\Phi_{\\mathrm{pose}}}(\\cdot)" + }, + { + "bbox": [ + 46, + 210, + 288, + 258 + ], + "type": "text", + "content": " to learn relative updates to joint angles:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 110, + 262, + 287, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 262, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 110, + 262, + 287, + 277 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\theta} = \\boldsymbol {\\theta} ^ {\\mathrm {S M P L}} \\otimes \\mathrm {M L P} _ {\\Phi_ {\\text {p o s e}}} (\\boldsymbol {\\theta}), \\tag {6}", + "image_path": "b03aabca9f19098059b7ff5943b5671661c01135e78c7123b4d83aa3d7584d76.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 281, + 287, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 287, + 317 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 287, + 317 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 281, + 287, + 317 + ], + "type": "inline_equation", + "content": "\\theta^{\\mathrm{SMPL}}" + }, + { + "bbox": [ + 46, + 281, + 287, + 317 + ], + "type": "text", + "content": " is the SMPL body pose parameters estimated from images. The body pose refinement is also investigated in [12, 49, 116]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 318, + 287, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 318, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 46, + 318, + 287, + 414 + ], + "type": "text", + "content": "During inference, thanks to the 3D Gaussian-based representation defined in a canonical space, the learned LBS weight for each 3D Gaussian can be derived from the LBS weight field module in advance, which saves computation time. Meanwhile, the rectified pose parameters can also be saved from the pose refinement module at the end of training. Considering the above, our GauHuman can model high-quality 3D humans with high rendering speed." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 420, + 224, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 420, + 224, + 432 + ], + "spans": [ + { + "bbox": [ + 47, + 420, + 224, + 432 + ], + "type": "text", + "content": "3.3. Fast Optimization of GauHuman" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 437, + 287, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 486 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 486 + ], + "type": "text", + "content": "The current optimization scheme of Gaussian Splitting [56] is specifically designed for static scenes. To facilitate fast optimization for articulated 3D humans, we incorporate 3D human prior and KL divergence measures to speed up." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 487, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 487, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 487, + 287, + 581 + ], + "type": "text", + "content": "3D Gaussian Initialization Current Gaussian Splatting methods utilize either Structure-from-Motion (SfM) [98, 101] or random point cloud to initialize 3D Gaussians, which overlooks 3D human structures. In GauHuman, we initialize 3D Gaussians with human prior like SMPL vertex points. The above initialization encodes human structure, thus accelerating optimization. Note that for 3D Gaussian covariance initialization, we follow the default setting in [56]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "content": "Split & Clone & Merge & Prune As mentioned in Sec. 3.1, the optimization of Gaussian Splatting relies on the adaptive control of 3D Gaussians to ensure superior rendering quality. To adaptively control the number of Gaussians, three operations, i.e., split, clone, and prune, are performed based on the position gradients and magnitude of the scaling matrix. However, although split and clone operations substantially improve the performance, they trivially generate a substantial quantity of 3D Gaussians. This results in a slowdown of the optimization and a significant consumption of storage memory, reaching up to " + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "inline_equation", + "content": "734\\mathrm{MB}" + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "content": " in static scenes. Our" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "type": "text", + "content": "analysis indicates that previous research overlooks a crucial metric, namely, the distance between 3D Gaussians, during the split and clone process. For example, if two Gaussians are close to each other, even if the position gradients are larger than a threshold, they should not be split or cloned as 1) these Gaussians are updating their positions to improve their performance, 2) splitting and cloning these Gaussians have negligible influence on the final performance as they are too close to each other. To regulate the split and clone process, we propose to use Kullback-Leibler (KL) divergence as a measure of the distance of 3D Gaussians, i.e.," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 311, + 217, + 545, + 258 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 217, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 311, + 217, + 545, + 258 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} K L \\left(G \\left(\\boldsymbol {x} _ {0}\\right) \\mid G \\left(\\boldsymbol {x} _ {1}\\right)\\right) = \\frac {1}{2} \\left(t r \\left(\\boldsymbol {\\Sigma} _ {1} ^ {- 1} \\boldsymbol {\\Sigma} _ {0}\\right) + \\ln \\frac {\\det \\boldsymbol {\\Sigma} _ {1}}{\\det \\boldsymbol {\\Sigma} _ {2}} \\right. \\tag {7} \\\\ + \\left(\\boldsymbol {p} _ {1} - \\boldsymbol {p} _ {0}\\right) ^ {T} \\boldsymbol {\\Sigma} _ {1} ^ {- 1} \\left(\\boldsymbol {p} _ {1} - \\boldsymbol {p} _ {0}\\right) - 3), \\\\ \\end{array}", + "image_path": "7558e28a53901ce5295f69634f7b82e80325cc12f19e149735d66c8bd6ca098f.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 261, + 545, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 261, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 304, + 261, + 545, + 285 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 261, + 545, + 285 + ], + "type": "inline_equation", + "content": "\\pmb{p}_0, \\pmb{\\Sigma}_0, \\pmb{p}_1, \\pmb{\\Sigma}_1" + }, + { + "bbox": [ + 304, + 261, + 545, + 285 + ], + "type": "text", + "content": " are the position and covariance matrix of two 3D Gaussians " + }, + { + "bbox": [ + 304, + 261, + 545, + 285 + ], + "type": "inline_equation", + "content": "G(\\pmb{x}_0)" + }, + { + "bbox": [ + 304, + 261, + 545, + 285 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 261, + 545, + 285 + ], + "type": "inline_equation", + "content": "G(\\pmb{x}_1)" + }, + { + "bbox": [ + 304, + 261, + 545, + 285 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 289, + 546, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 546, + 516 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 546, + 516 + ], + "type": "text", + "content": "After obtaining the KL divergence of nearby 3D Gaussian pairs, we select 3D Gaussians with large KL divergence and positional gradients to perform the split and clone. However, the above requires calculating the KL divergence for each pair of 3D Gaussians, which consumes lots of time during the optimization. To solve this problem, we further adopt two approaches to simplify the whole calculation. 1) For each 3D Gaussian, we first identify its closest 3D Gaussian through the widely adopted k-nearest neighbor (k-NN) algorithm, which assesses the distance between their respective centers. Then we calculate a KL divergence for each pair of nearby 3D Gaussians. The aforementioned simplification effectively reduces the time complexity from " + }, + { + "bbox": [ + 304, + 289, + 546, + 516 + ], + "type": "inline_equation", + "content": "O(U^2)" + }, + { + "bbox": [ + 304, + 289, + 546, + 516 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 289, + 546, + 516 + ], + "type": "inline_equation", + "content": "O(U)" + }, + { + "bbox": [ + 304, + 289, + 546, + 516 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 289, + 546, + 516 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 304, + 289, + 546, + 516 + ], + "type": "text", + "content": " represents the total number of 3D Gaussians. 2) As the covariance matrix is decomposed into the product of rotation and scaling matrices, we further simplify the computation of matrix inverse and determinant operations in Eqn. (7) with the diagonal and orthogonal property of rotation and scaling matrices. Details are shown in the appendix." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 520, + 547, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 520, + 547, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 547, + 640 + ], + "type": "text", + "content": "In addition to split and clone operations, we further propose a novel merge operation, which merges redundant 3D Gaussians with 1) large position gradients, 2) small scaling magnitude, and 3) KL divergence less than 0.1. Specifically, we merge two Gaussians by averaging their positions, opacity, and SH coefficients. Additionally, we scale the covariance (scaling) matrix of the first Gaussian by a factor of 1.25 as the initialization for the new Gaussian. For prune operation, we incorporate 3D human prior like SMPL to prune 3D Gaussians far away from SMPL vertex points." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 643, + 536, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 643, + 536, + 655 + ], + "spans": [ + { + "bbox": [ + 317, + 643, + 536, + 655 + ], + "type": "text", + "content": "During optimization, four loss functions are used, i.e.," + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 672, + 545, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 672, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 315, + 672, + 545, + 685 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {c o l o r}} + \\lambda_ {1} \\mathcal {L} _ {\\text {m a s k}} + \\lambda_ {2} \\mathcal {L} _ {\\text {S S I M}} + \\lambda_ {3} \\mathcal {L} _ {\\text {L P I P S}}, \\tag {8}", + "image_path": "34b9598e53118697889eb0687b0f38ccd457aa8a9566251cf233c6eb2def30c7.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "content": "'s are loss weights. Empirically, we set " + }, + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_1 = 0.5" + }, + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_2 = \\lambda_3 = 0.01" + }, + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "content": " to ensure the same magnitude for each loss." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20422" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 122, + 535, + 278 + ], + "blocks": [ + { + "bbox": [ + 46, + 56, + 547, + 121 + ], + "lines": [ + { + "bbox": [ + 46, + 56, + 547, + 121 + ], + "spans": [ + { + "bbox": [ + 46, + 56, + 547, + 121 + ], + "type": "text", + "content": "Table 1. Quantitative comparison of our GauHuman and baseline methods on the ZJU_MoCap and MonoCap data sets. We use bold and underlined text to denote the best and second-best results of each metric respectively. Our GauHuman achieves the state-of-the-art PSNR and LPIPS with fast training (1~2 minutes) and real-time rendering (up to 189 FPS) in both data sets. Note that PixelNeRF [129] and NHP [60] are pre-trained for an additional 10 hours. " + }, + { + "bbox": [ + 46, + 56, + 547, + 121 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 46, + 56, + 547, + 121 + ], + "type": "text", + "content": " denotes the fine-tuning time for generalizable methods. LPIPS\\* = 1000 × LPIPS. Frames per second (FPS) is measured on an RTX 3090. For a fair comparison, we do not conduct test-time optimization of SMPL parameters with images from the test set on InstantAvatar [49]." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 122, + 535, + 278 + ], + "lines": [ + { + "bbox": [ + 56, + 122, + 535, + 278 + ], + "spans": [ + { + "bbox": [ + 56, + 122, + 535, + 278 + ], + "type": "table", + "html": "
MethodZJU_MoCapMonoCap
PSNR↑SSIM↑LPIPS*↓TrainFPSPSNR↑SSIM↑LPIPS*↓TrainFPS
PixelNeRF [129]24.710.892121.86\\( 1h^† \\)1.2026.430.96043.98\\( 1h^† \\)0.75
NHP [60]28.250.95564.77\\( 1h^† \\)0.1530.510.98027.14\\( 1h^† \\)0.05
NB [82]29.030.96442.4710h1.4832.360.98616.7010h0.98
AN [81]29.770.96546.8910h1.1131.070.98519.4710h0.31
AS [83]30.380.97537.2310h0.4032.480.98813.1810h0.29
HumanNeRF [116]30.660.96933.3810h0.3032.680.98715.5210h0.08
DVA [88]29.450.95637.741.5h16.532.990.98315.831.5h10.5
InstantNVR [30]31.010.97138.455m2.2032.610.98816.6810m0.75
InstantAvatar [49]29.730.93868.413m4.1530.790.96439.756m2.54
GauHuman(Ours)31.340.96530.511m18933.450.98513.352m154
", + "image_path": "463984b8ca0818d64cb8aaf80a325b64b357a877f45cfd96b08d83303b4f59a4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 286, + 128, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 286, + 128, + 300 + ], + "spans": [ + { + "bbox": [ + 47, + 286, + 128, + 300 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 306, + 163, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 306, + 163, + 319 + ], + "spans": [ + { + "bbox": [ + 47, + 306, + 163, + 319 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 325, + 288, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 325, + 288, + 539 + ], + "spans": [ + { + "bbox": [ + 46, + 325, + 288, + 539 + ], + "type": "text", + "content": "Datasets. Experiments are conducted on two widely used human modelling data sets, i.e., ZJU_MoCap [82] and MonoCap [36, 37, 83]. For ZJU_MoCap, same as previous work [30, 116], we select 6 human subjects (377, 386, 387, 392, 393, 394) to conduct experiments. We also adopt the same training and testing setting as [30], i.e., one camera is used for training, while the remaining cameras are used for evaluation. For each subject, we sample 1 frame every 5 frames and collect 100 frames for training. The foreground masks, camera, and SMPL parameters provided by the data set are used for evaluation purposes. MonoCap data set contains four multi-view videos collected by [83] from the DeepCap data set [36] and the DynaCap data set [37]. Similar to ZJU_MoCap, it provides human masks, cameras, and SMPL parameters. We adopt the same camera views as [83] for training and testing. For each subject, we collect 100 frames for training by sampling 1 frame every 5 frames. More details are shown in the appendix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 541, + 288, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 541, + 288, + 649 + ], + "spans": [ + { + "bbox": [ + 46, + 541, + 288, + 649 + ], + "type": "text", + "content": "Comparison Methods. We compare our GauHuman with two categories of state-of-the-art human modelling methods, i.e., subject-specific optimization based methods [30, 49, 81-83, 88, 116] and generalizable methods [60, 129]. Following [30], we report the average metric values of all selected human subjects. For a fair comparison, we do not conduct test-time optimization of SMPL parameters with images from the test set on InstantAvatar [49]. More details are shown in the appendix." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 658, + 166, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 166, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 166, + 671 + ], + "type": "text", + "content": "4.2. Quantitative Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "Tab. 1 summarizes the overall novel view synthesis results of our GauHuman, NB [82], AN [81], AS [83], HumanNeRF [116], DVA [88], InstantNVR [30], InstantAvatar [49]," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 287, + 547, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 287, + 547, + 515 + ], + "spans": [ + { + "bbox": [ + 304, + 287, + 547, + 515 + ], + "type": "text", + "content": "PixelNeRF [129], and NHP [60]. Thanks to our proposed articulated 3D human Gaussian-based representation, we can fast optimize it for around 1 or 2 minutes to achieve photorealistic rendering results, while baseline methods (i.e., NB, AN, AS, and HumanNeRF) take around 10 hours to finish the training and generalizable methods takes about 1 hour for fine-tuning after 10 hours' pre-training. To accelerate the training of 3D human modelling, InstantNVR [30] and InstantAvatar [49] reduce the training to several minutes with multi hashing encoding representation. However, their rendering speed still restricts their applications in real-world applications. When compared with these two state-of-the-art methods for fasting training, our GauHuman achieves slightly better performance with 3 times faster training and 45 times faster rendering speed. Among baseline methods, DVA achieves a fast rendering speed (up to 16.5 FPS) with efficient mixtures of volumetric primitives. However, in contrast to our GauHuman, DVA takes about " + }, + { + "bbox": [ + 304, + 287, + 547, + 515 + ], + "type": "inline_equation", + "content": "90\\mathrm{x}" + }, + { + "bbox": [ + 304, + 287, + 547, + 515 + ], + "type": "text", + "content": " more time to converge and " + }, + { + "bbox": [ + 304, + 287, + 547, + 515 + ], + "type": "inline_equation", + "content": "11.45\\mathrm{x}" + }, + { + "bbox": [ + 304, + 287, + 547, + 515 + ], + "type": "text", + "content": " more time to render one image." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 526, + 418, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 526, + 418, + 539 + ], + "spans": [ + { + "bbox": [ + 306, + 526, + 418, + 539 + ], + "type": "text", + "content": "4.3. Qualitative Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "content": "As shown in Fig. 3, our GauHuman shows better appearance and geometry details than DVA [88], NB [82] and NHP [60]. Although NB [82] and NHP [60] show impressive results with 4-view videos, they do not perform well with monocular videos. NB [82] encodes human structure information with latent codes positioned in SMPL vertex's locations, which fails to capture complex human motions. NHP [60] does not generalize well in monocular videos due to the information missing from input images. Thanks to the LBS weight field and deformation field learned in HumanNeRF [116] and AS [83], they achieve comparable visualization results as our GauHuman for " + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "inline_equation", + "content": "600\\mathrm{x}" + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "content": " more time for training and " + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "inline_equation", + "content": "472\\mathrm{x}" + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "content": " more time for rendering. InstantNVR [30] and InstantAvatar [49] reduce the training time to several minutes with a multi" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20423" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 52, + 541, + 390 + ], + "blocks": [ + { + "bbox": [ + 53, + 52, + 541, + 390 + ], + "lines": [ + { + "bbox": [ + 53, + 52, + 541, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 52, + 541, + 390 + ], + "type": "image", + "image_path": "eed4e7ea3430f589fe608b28c4a70010fb51cad4eab29e1ffd345e9e177286f7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 392, + 546, + 415 + ], + "lines": [ + { + "bbox": [ + 46, + 392, + 546, + 415 + ], + "spans": [ + { + "bbox": [ + 46, + 392, + 546, + 415 + ], + "type": "text", + "content": "Figure 3. Novel view synthesis results produced by our GauHuman and baseline methods on ZJU_MoCap and MonoCap. The bottom lines show the training time and rendering speed of each method on ZJU_Mocap. Zoom in for the best view." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 427, + 289, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 427, + 289, + 523 + ], + "spans": [ + { + "bbox": [ + 46, + 427, + 289, + 523 + ], + "type": "text", + "content": "hashing encoding representation, but they produce slightly worse visualization results than ours and require " + }, + { + "bbox": [ + 46, + 427, + 289, + 523 + ], + "type": "inline_equation", + "content": "45\\mathrm{x}" + }, + { + "bbox": [ + 46, + 427, + 289, + 523 + ], + "type": "text", + "content": " more rendering time. DVA [] improves the rendering speed to 16.5 FPS with efficient mixtures of volumetric primitives. However, it fails to produce high-quality results and produces obvious artifacts on garments and faces, due to incomplete information within conditional input images and a lack of mechanisms to adaptively control the number of voxels." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 536, + 141, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 141, + 548 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 141, + 548 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 555, + 288, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 555, + 288, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 555, + 288, + 616 + ], + "type": "text", + "content": "We perform ablation studies with the sequence of 386 from the ZJU_Mocap data set. To highlight the effectiveness of each proposed component, we project the 3D human bounding box to each camera plane to obtain the bounding box mask and report these metrics based on the masked area." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "content": "Pose Refinement and LBS Weight Field. Tab. 2 shows the results of ablating pose refinement and LBS weight field modules. All models are trained in around 1 minute. As shown in Tab. 2, pose refinement and LBS weight field modules significantly improve all metrics with negligible computational cost. The visualization results in Fig. 4 show that these two modules can help erase artifacts and learn fine details like cloth wrinkles. For example, we observe obvious" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 427, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 427, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 305, + 427, + 545, + 449 + ], + "type": "text", + "content": "artifacts in hands and fewer cloth wrinkles without using either one of these two modules." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 309, + 472, + 541, + 529 + ], + "blocks": [ + { + "bbox": [ + 305, + 450, + 545, + 472 + ], + "lines": [ + { + "bbox": [ + 305, + 450, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 305, + 450, + 545, + 472 + ], + "type": "text", + "content": "Table 2. Quantitative Results of ablating pose refinement and LBS weight field modules. " + }, + { + "bbox": [ + 305, + 450, + 545, + 472 + ], + "type": "inline_equation", + "content": "\\mathrm{LPIPS^{*}} = 1000\\times" + }, + { + "bbox": [ + 305, + 450, + 545, + 472 + ], + "type": "text", + "content": " LPIPS." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 472, + 541, + 529 + ], + "lines": [ + { + "bbox": [ + 309, + 472, + 541, + 529 + ], + "spans": [ + { + "bbox": [ + 309, + 472, + 541, + 529 + ], + "type": "table", + "html": "
PSNR↑SSIM↑LPIPS*↓TrainFPS
Ours (full model)28.080.886103.355s189
Ours w/o pose refine27.940.884108.845s192
Ours w/o LBS field27.870.882108.952s190
", + "image_path": "5b05c9a560aff5d2eb1add5161fcb17d3abd9947f8a8dbb574420a89b75a7ff1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 533, + 547, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 547, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 547, + 677 + ], + "type": "text", + "content": "3D Gaussian Initialization. In Gaussian Splitting, previous research [56] indicates that initialization plays an important role in the final results. For static scenes, there are two commonly adopted initialization approaches, i.e., random initialization or initialization with sparse points from Structure-from-Motion (SfM). To incorporate human articulated information into the initialization, we propose to use articulated vertex points from SMPL to initialize our 3D Gaussians in the canonical space. We compare our proposed articulated initialization with random initialization1 in Tab. 3. With the same number of 3D Gaussians at initialization, our proposed articulated initialization produces both better" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 305, + 683, + 547, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 683, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 683, + 547, + 713 + ], + "type": "text", + "content": "Estimating sparse point cloud from Structure-from-Motion (SfM) requires multi-view inputs, which is infeasible with monocular videos. So we do not compare it with our articulated initialization." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20424" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 58, + 544, + 223 + ], + "blocks": [ + { + "bbox": [ + 49, + 58, + 544, + 223 + ], + "lines": [ + { + "bbox": [ + 49, + 58, + 544, + 223 + ], + "spans": [ + { + "bbox": [ + 49, + 58, + 544, + 223 + ], + "type": "image", + "image_path": "7b42fc5011e011a098790b971ec59fa731997588929a8c3cf3a387886253f1a6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 230, + 548, + 265 + ], + "lines": [ + { + "bbox": [ + 46, + 230, + 548, + 265 + ], + "spans": [ + { + "bbox": [ + 46, + 230, + 548, + 265 + ], + "type": "text", + "content": "Figure 4. Qualitative results of ablation study on the 386 sequence of ZJU_Mocap data set. \"w/o pose refinement\" and \"w/o LBS weight field\" mean that we do not adopt the pose refinement or LBS weight field module respectively. \"w/ random Init (#Gau 6890)\" and \"w/ random Init (#Gau 30000)\" mean that we use random initialization to initialize 6890 or 30000 3D Gaussians respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 273, + 289, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 273, + 289, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 273, + 289, + 392 + ], + "type": "text", + "content": "quantitative and qualitative results than random initialization as shown in Tab. 3 and Fig. 4. Notably, our articulated initialization converges x10 times much faster than the random initialization, which illustrates the strength of encoding human articulated information into the initialization. For random initialization, we need to initialize 30k Gaussians to achieve comparable performance as our proposed articulated initialization, while at cost of 30x times more training time. Table 3. Quantitative Results of ablating 3D Gaussian initialization. #Gau denotes the number of 3D Gaussians at initialization." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 48, + 393, + 284, + 449 + ], + "blocks": [ + { + "bbox": [ + 48, + 393, + 284, + 449 + ], + "lines": [ + { + "bbox": [ + 48, + 393, + 284, + 449 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 284, + 449 + ], + "type": "table", + "html": "
PSNR↑SSIM↑LPIPS*↓Train
Articulated Init (w/ #Gau 6890)28.080.886103.355s
Random Init (w/ #Gau 6890)27.980.882108.95m
Random Init (w/ #Gau 30k)28.080.884104.330m
", + "image_path": "2705ac9f144e3ae48cb47e838d780cf6776a8cfad1da18e94bc53eb2b6446ef2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 453, + 287, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 453, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 46, + 453, + 287, + 620 + ], + "type": "text", + "content": "Split/Clone/Merge/Prune. Adaptive control of 3D Gaussians with split, clone, and prune operations substantially improves the rendering quality, but it trivially leads to a large number of 3D Gaussians (up to " + }, + { + "bbox": [ + 46, + 453, + 287, + 620 + ], + "type": "inline_equation", + "content": "100\\mathrm{k}" + }, + { + "bbox": [ + 46, + 453, + 287, + 620 + ], + "type": "text", + "content": "), which slows down the whole optimization process. As mentioned in Sec. 3.3, we propose to use the KL divergence measure to regulate the split and clone process. As shown in Tab. 4, our KL-based split and clone operations lead to " + }, + { + "bbox": [ + 46, + 453, + 287, + 620 + ], + "type": "inline_equation", + "content": "60\\mathrm{x}" + }, + { + "bbox": [ + 46, + 453, + 287, + 620 + ], + "type": "text", + "content": " times faster convergence speed when compared with the original split and clone operation proposed in [56]. Furthermore, our novel merge operation merges about " + }, + { + "bbox": [ + 46, + 453, + 287, + 620 + ], + "type": "inline_equation", + "content": "8.5\\%" + }, + { + "bbox": [ + 46, + 453, + 287, + 620 + ], + "type": "text", + "content": " of 3D Gaussians without degrading the performance. Articulated prune operation with SMPL prior introduced in Sec. 3.3 also helps prune about " + }, + { + "bbox": [ + 46, + 453, + 287, + 620 + ], + "type": "inline_equation", + "content": "0.5\\mathrm{k}" + }, + { + "bbox": [ + 46, + 453, + 287, + 620 + ], + "type": "text", + "content": " 3D Gaussians far away from the 3D human body." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 622, + 287, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 622, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 46, + 622, + 287, + 654 + ], + "type": "text", + "content": "Table 4. Quantitative Results of ablating our proposed KL-based split/clone operations and a novel merge operation. #Gau denotes the final number of 3D Gaussians during optimization." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 50, + 655, + 282, + 722 + ], + "blocks": [ + { + "bbox": [ + 50, + 655, + 282, + 722 + ], + "lines": [ + { + "bbox": [ + 50, + 655, + 282, + 722 + ], + "spans": [ + { + "bbox": [ + 50, + 655, + 282, + 722 + ], + "type": "table", + "html": "
PSNR↑SSIM↑LPIPS*↓#GauTrain
Ours (full model)28.080.886103.31316255s
w/o KL split/clone28.010.885103.1100k1h
w/o KL merge28.080.886103.61437670s
w/o articulated prune28.080.886103.41368757s
", + "image_path": "652903be0e8dd30354576f20fec32e0182801c9b4d19611ac581677338171a23.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 272, + 457, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 272, + 457, + 285 + ], + "spans": [ + { + "bbox": [ + 306, + 272, + 457, + 285 + ], + "type": "text", + "content": "5. Discussion and Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 300, + 547, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 547, + 493 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 547, + 493 + ], + "type": "text", + "content": "To conclude, we propose GauHuman, a 3D human model with Gaussian Splating representation for fast training " + }, + { + "bbox": [ + 304, + 300, + 547, + 493 + ], + "type": "inline_equation", + "content": "(1\\sim 2" + }, + { + "bbox": [ + 304, + 300, + 547, + 493 + ], + "type": "text", + "content": " minutes) and real-time rendering (166 FPS) of 3D humans. Specifically, we encode this representation in a canonical space and transform 3D Gaussian from canonical space to posed space with LBS transformation, in which effective pose refinement and LBS weight field modules are also designed to learn fine details of 3D humans. To enable fast optimization, we initialize and prune 3D Gaussians with 3D human prior, while splitting/cloning 3D Gaussians via KI divergence guidance, accompanied by a novel merge operation for further speeding up. Experiments on monocular videos suggest that our GauHuman achieves state-of-the-art performance while maintaining both fast training and real-time rendering speed. In addition, GauHuman can fast learn 3D human performers with " + }, + { + "bbox": [ + 304, + 300, + 547, + 493 + ], + "type": "inline_equation", + "content": "\\sim 13k" + }, + { + "bbox": [ + 304, + 300, + 547, + 493 + ], + "type": "text", + "content": " Gaussians." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 501, + 547, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 501, + 547, + 586 + ], + "spans": [ + { + "bbox": [ + 304, + 501, + 547, + 586 + ], + "type": "text", + "content": "Limitations. 1). GauHuman is composed of 3D Gaussians and current pipeline does not support extracting 3D meshes. How to extract meshes from 3D Gaussians remains a future work to be investigated. 2). Recovering 3D humans with loose/complex clothing from monocular videos is still a challenging problem to tackle. One direction is to incorporate physical simulation of garments into current pipelines." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "content": "Acknowledgment. This study is supported by the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOET2EP20221-0012), NTU NAP, and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). We would like to thank Zhaoxi Chen, Wei Li, Guangcong Wang from NTU, Xurong Xie from ISCAS, Zhiyuan Yu from ZJU, and Zihao Huang from HUST for the discussions on the paper draft." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20425" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 58, + 91, + 288, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 91, + 288, + 145 + ], + "spans": [ + { + "bbox": [ + 58, + 91, + 288, + 145 + ], + "type": "text", + "content": "[1] Badour AlBahar, Jingwan Lu, Jimei Yang, Zhixin Shu, Eli Shechtman, and Jia-Bin Huang. Pose with style: Detail-preserving pose-guided image synthesis with conditional stylegan. ACM Transactions on Graphics (TOG), 40(6): 1-11, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 147, + 288, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 147, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 58, + 147, + 288, + 201 + ], + "type": "text", + "content": "[2] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 696-712. Springer, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 202, + 288, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 202, + 288, + 256 + ], + "spans": [ + { + "bbox": [ + 58, + 202, + 288, + 256 + ], + "type": "text", + "content": "[3] Thiemo Alldieck, Mihai Zanfir, and Cristian Sminchisescu. Photorealistic monocular 3d reconstruction of humans wearing clothing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1506-1515, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 258, + 288, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 258, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 58, + 258, + 288, + 312 + ], + "type": "text", + "content": "[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 314, + 288, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 314, + 288, + 368 + ], + "spans": [ + { + "bbox": [ + 58, + 314, + 288, + 368 + ], + "type": "text", + "content": "[5] Bharat Lal Bhatnagar, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. Loopreg: Self-supervised learning of implicit surface correspondences, pose and shape for 3d human mesh registration. Advances in Neural Information Processing Systems, 33:12909-12922, 2020. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 369, + 288, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 369, + 288, + 413 + ], + "spans": [ + { + "bbox": [ + 58, + 369, + 288, + 413 + ], + "type": "text", + "content": "[6] Mario Botsch, Alexander Hornung, Matthias Zwicker, and Leif Kobbelt. High-quality surface splatting on today's gpus. In Proceedings Eurographics/IEEE VGTC Symposium Point-Based Graphics, 2005., pages 17-141. IEEE, 2005. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 414, + 288, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 414, + 288, + 457 + ], + "spans": [ + { + "bbox": [ + 58, + 414, + 288, + 457 + ], + "type": "text", + "content": "[7] Chris Buehler, Michael Bosse, Leonard McMillan, Steven Gortler, and Michael Cohen. Unstructured lumigraph rendering. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 497-504. 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 58, + 458, + 288, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 458, + 288, + 513 + ], + "spans": [ + { + "bbox": [ + 58, + 458, + 288, + 513 + ], + "type": "text", + "content": "[8] Sihun Cha, Kwanggyoon Seo, Amirsaman Ashtari, and Jun-yong Noh. Generating texture for 3d human avatar from a single image using sampling and refinement networks. In Computer Graphics Forum, pages 385-396. Wiley Online Library, 2023. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 58, + 514, + 288, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 514, + 288, + 556 + ], + "spans": [ + { + "bbox": [ + 58, + 514, + 288, + 556 + ], + "type": "text", + "content": "[9] Gaurav Chaurasia, Sylvain Duchene, Olga Sorkine-Hornung, and George Drettakis. Depth synthesis and local warps for plausible image-based navigation. ACM Transactions on Graphics (TOG), 32(3):1-12, 2013. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 558, + 288, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 558, + 288, + 602 + ], + "spans": [ + { + "bbox": [ + 53, + 558, + 288, + 602 + ], + "type": "text", + "content": "[10] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. arXiv preprint arXiv:2103.15595, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 602, + 288, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 602, + 288, + 645 + ], + "spans": [ + { + "bbox": [ + 53, + 602, + 288, + 645 + ], + "type": "text", + "content": "[11] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 647, + 288, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 647, + 288, + 690 + ], + "spans": [ + { + "bbox": [ + 53, + 647, + 288, + 690 + ], + "type": "text", + "content": "[12] Jianchuan Chen, Ying Zhang, Di Kang, Xuefei Zhe, Linchao Bao, Xu Jia, and Huchuan Lu. Animatable neural radiance fields from monocular rgb videos. arXiv preprint arXiv:2106.13629, 2021. 1, 3, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 691, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 691, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 691, + 288, + 713 + ], + "type": "text", + "content": "[13] Xu Chen, Yufeng Zheng, Michael J Black, Otmar Hilliges, and Andreas Geiger. Snarf: Differentiable forward skin" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 74, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 331, + 74, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 74, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 331, + 74, + 545, + 106 + ], + "type": "text", + "content": "ning for animating non-rigid neural implicit shapes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11594-11604, 2021. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 108, + 547, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 108, + 547, + 162 + ], + "spans": [ + { + "bbox": [ + 312, + 108, + 547, + 162 + ], + "type": "text", + "content": "[14] Zhaoxi Chen, Fangzhou Hong, Haiyi Mei, Guangcong Wang, Lei Yang, and Ziwei Liu. Primdiffusion: Volumetric primitives diffusion for 3d human generation. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 164, + 547, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 164, + 547, + 229 + ], + "spans": [ + { + "bbox": [ + 312, + 164, + 547, + 229 + ], + "type": "text", + "content": "[15] Wei Cheng, Ruixiang Chen, Siming Fan, Wanqi Yin, Keyu Chen, Zhongang Cai, Jingbo Wang, Yang Gao, Zhengming Yu, Zhengyu Lin, et al. Dna-rendering: A diverse neural actor repository for high-fidelity human-centric rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19982-19993, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 231, + 547, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 231, + 547, + 285 + ], + "spans": [ + { + "bbox": [ + 312, + 231, + 547, + 285 + ], + "type": "text", + "content": "[16] Julian Chibane, Aayush Bansal, Verica Lazova, and Gerard Pons-Moll. Stereo radiance fields (srf): Learning view synthesis for sparse views of novel scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7911-7920, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 287, + 547, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 287, + 547, + 341 + ], + "spans": [ + { + "bbox": [ + 312, + 287, + 547, + 341 + ], + "type": "text", + "content": "[17] Hongsuk Choi, Gyeongsik Moon, Matthieu Armando, Vincent Leroy, Kyoung Mu Lee, and Grégory Rogez. Mononhr: Monocular neural human renderer. In 2022 International Conference on 3D Vision (3DV), pages 242-251. IEEE, 2022. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 343, + 547, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 343, + 547, + 397 + ], + "spans": [ + { + "bbox": [ + 312, + 343, + 547, + 397 + ], + "type": "text", + "content": "[18] Alvaro Collet, Ming Chuang, Pat Sweeney, Don Gillett, Dennis Evseev, David Calabrese, Hugues Hoppe, Adam Kirk, and Steve Sullivan. High-quality streamable free-viewpoint video. ACM Transactions on Graphics (ToG), 34 (4):1-13, 2015. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 399, + 547, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 399, + 547, + 454 + ], + "spans": [ + { + "bbox": [ + 312, + 399, + 547, + 454 + ], + "type": "text", + "content": "[19] François Darmon, Bénédicte Bascle, Jean-Clement Devaux, Pascal Monasse, and Mathieu Aubry. Improving neural implicit surfaces geometry with patch warping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6260-6269, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 456, + 547, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 456, + 547, + 521 + ], + "spans": [ + { + "bbox": [ + 312, + 456, + 547, + 521 + ], + "type": "text", + "content": "[20] Boyang Deng, John P Lewis, Timothy Jeruzalski, Gerard Pons-Moll, Geoffrey Hinton, Mohammad Norouzi, and Andrea Tagliasacchi. Nasa neural articulated shape approximation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VII 16, pages 612-628. Springer, 2020. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 523, + 547, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 523, + 547, + 555 + ], + "spans": [ + { + "bbox": [ + 312, + 523, + 547, + 555 + ], + "type": "text", + "content": "[21] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nef: Fewer views and faster training for free. arXiv preprint arXiv:2107.02791, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 557, + 547, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 557, + 547, + 622 + ], + "spans": [ + { + "bbox": [ + 312, + 557, + 547, + 622 + ], + "type": "text", + "content": "[22] Mingsong Dou, Sameh Khamis, Yury Degtyarev, Philip Davidson, Sean Ryan Fanello, Adarsh Kowdle, Sergio Orts Escolano, Christoph Rhemann, David Kim, Jonathan Taylor, et al. Fusion4d: Real-time performance capture of challenging scenes. ACM Transactions on Graphics (ToG), 35(4): 1-13, 2016. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 624, + 547, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 624, + 547, + 678 + ], + "spans": [ + { + "bbox": [ + 312, + 624, + 547, + 678 + ], + "type": "text", + "content": "[23] Martin Eisemann, Bert De Decker, Marcus Magnor, Philippe Bekaert, Edilson De Aguiar, Naveed Ahmed, Christian Theobalt, and Anita Sellent. Floating textures. In Computer graphics forum, pages 409-418. Wiley Online Library, 2008. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 681, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 681, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 681, + 547, + 713 + ], + "type": "text", + "content": "[24] John Flynn, Ivan Neulander, James Philbin, and Noah Snavely. Deepstereo: Learning to predict new views from the world's imagery. In Proceedings of the IEEE conference" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "20426" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 72, + 72, + 288, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 72, + 288, + 95 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 288, + 95 + ], + "type": "text", + "content": "on computer vision and pattern recognition, pages 5515-5524, 2016.3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 96, + 288, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 96, + 288, + 151 + ], + "spans": [ + { + "bbox": [ + 53, + 96, + 288, + 151 + ], + "type": "text", + "content": "[25] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 152, + 288, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 152, + 288, + 207 + ], + "spans": [ + { + "bbox": [ + 53, + 152, + 288, + 207 + ], + "type": "text", + "content": "[26] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Chen Qian, Chen Change Loy, Wayne Wu, and Ziwei Liu. Stylegan-human: A data-centric odyssey of human generation. In European Conference on Computer Vision, pages 1-19. Springer, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 209, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 209, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 53, + 209, + 288, + 262 + ], + "type": "text", + "content": "[27] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Wayne Wu, and Ziwei Liu. Unitedhuman: Harnessing multi-source data for high-resolution human generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7301-7311, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 264, + 288, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 264, + 288, + 308 + ], + "spans": [ + { + "bbox": [ + 53, + 264, + 288, + 308 + ], + "type": "text", + "content": "[28] Xiangjun Gao, Jiaolong Yang, Jongyoo Kim, Sida Peng, Zicheng Liu, and Xin Tong. Mps-nerf: Generalizable 3d human rendering from multiview images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 310, + 288, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 310, + 288, + 363 + ], + "spans": [ + { + "bbox": [ + 53, + 310, + 288, + 363 + ], + "type": "text", + "content": "[29] Stephan J Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14346-14355, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 365, + 288, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 365, + 288, + 421 + ], + "spans": [ + { + "bbox": [ + 53, + 365, + 288, + 421 + ], + "type": "text", + "content": "[30] Chen Geng, Sida Peng, Zhen Xu, Hujun Bao, and Xiaowei Zhou. Learning neural volumetric representations of dynamic humans in minutes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8759-8770, 2023. 1, 2, 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 422, + 288, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 422, + 288, + 466 + ], + "spans": [ + { + "bbox": [ + 53, + 422, + 288, + 466 + ], + "type": "text", + "content": "[31] Michael Goesele, Noah Snavely, Brian Curless, Hugues Hoppe, and Steven M Seitz. Multi-view stereo for community photo collections. In 2007 IEEE 11th International Conference on Computer Vision, pages 1-8. IEEE, 2007. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 468, + 288, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 468, + 288, + 510 + ], + "spans": [ + { + "bbox": [ + 53, + 468, + 288, + 510 + ], + "type": "text", + "content": "[32] Steven J Gortler, Radek Grzeszcuk, Richard Szeliski, and Michael F Cohen. The lumigraph. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 453-464. 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 512, + 288, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 512, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 53, + 512, + 288, + 533 + ], + "type": "text", + "content": "[33] Markus Gross and Hanspeter Pfister. Point-based graphics. Elsevier, 2011. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 535, + 288, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 535, + 288, + 579 + ], + "spans": [ + { + "bbox": [ + 53, + 535, + 288, + 579 + ], + "type": "text", + "content": "[34] Jeffrey P Grossman and William J Dally. Point sample rendering. In Rendering Techniques' 98: Proceedings of the Eurographics Workshop in Vienna, Austria, June 29—July 1, 1998 9, pages 181–192. Springer, 1998. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 581, + 288, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 581, + 288, + 645 + ], + "spans": [ + { + "bbox": [ + 53, + 581, + 288, + 645 + ], + "type": "text", + "content": "[35] Kaiwen Guo, Peter Lincoln, Philip Davidson, Jay Busch, Xueming Yu, Matt Whalen, Geoff Harvey, Sergio Orts-Escolano, Rohit Pandey, Jason Dourgarian, et al. The relightables: Volumetric performance capture of humans with realistic relighting. ACM Transactions on Graphics (ToG), 38(6):1-19, 2019. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 647, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 647, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 647, + 288, + 712 + ], + "type": "text", + "content": "[36] Marc Habermann, Weipeng Xu, Michael Zollhofer, Gerard Pons-Moll, and Christian Theobalt. Deepcap: Monocular human performance capture using weak supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5052-5063, 2020. 1, 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 312, + 73, + 547, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 547, + 116 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 547, + 116 + ], + "type": "text", + "content": "[37] Marc Habermann, Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Real-time deep dynamic characters. ACM Transactions on Graphics (ToG), 40(4):1-16, 2021. 1, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 118, + 547, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 118, + 547, + 162 + ], + "spans": [ + { + "bbox": [ + 312, + 118, + 547, + 162 + ], + "type": "text", + "content": "[38] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering. ACM Transactions on Graphics (ToG), 37(6):1-15, 2018. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 163, + 547, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 163, + 547, + 207 + ], + "spans": [ + { + "bbox": [ + 312, + 163, + 547, + 207 + ], + "type": "text", + "content": "[39] Philipp Henzler, Niloy J Mitra, and Tobias Ritschel. Escaping plato's cave: 3d shape from adversarial rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9984-9993, 2019. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 209, + 547, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 209, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 312, + 209, + 547, + 251 + ], + "type": "text", + "content": "[40] Fangzhou Hong, Zhaoxi Chen, Yushi Lan, Liang Pan, and Ziwei Liu. Eva3d: Compositional 3d human generation from 2d image collections. arXiv preprint arXiv:2210.04888, 2022.3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 254, + 547, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 254, + 547, + 296 + ], + "spans": [ + { + "bbox": [ + 312, + 254, + 547, + 296 + ], + "type": "text", + "content": "[41] Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. arXiv preprint arXiv:2205.08535, 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 298, + 547, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 298, + 547, + 331 + ], + "spans": [ + { + "bbox": [ + 312, + 298, + 547, + 331 + ], + "type": "text", + "content": "[42] Shoukang Hu, Fangzhou Hong, Liang Pan, Haiyi Mei, Lei Yang, and Ziwei Liu. Sherf: Generalizable human nerf from a single image. arXiv preprint arXiv:2303.12791, 2023. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 332, + 547, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 332, + 547, + 376 + ], + "spans": [ + { + "bbox": [ + 312, + 332, + 547, + 376 + ], + "type": "text", + "content": "[43] Yangyi Huang, Hongwei Yi, Yuliang Xiu, Tingting Liao, Jiaxiang Tang, Deng Cai, and Justus Thies. Tech: Text-guided reconstruction of lifelike clothed humans. arXiv preprint arXiv:2308.08545, 2023. 1, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 377, + 547, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 377, + 547, + 431 + ], + "spans": [ + { + "bbox": [ + 312, + 377, + 547, + 431 + ], + "type": "text", + "content": "[44] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. Arch: Animatable reconstruction of clothed humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3093-3102, 2020. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 434, + 547, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 434, + 547, + 477 + ], + "spans": [ + { + "bbox": [ + 312, + 434, + 547, + 477 + ], + "type": "text", + "content": "[45] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5885-5894, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 479, + 547, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 479, + 547, + 521 + ], + "spans": [ + { + "bbox": [ + 312, + 479, + 547, + 521 + ], + "type": "text", + "content": "[46] Wonbong Jang and Lourdes Agapito. Codenerf: Disentangled neural radiance fields for object categories. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12949-12958, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 523, + 547, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 523, + 547, + 578 + ], + "spans": [ + { + "bbox": [ + 312, + 523, + 547, + 578 + ], + "type": "text", + "content": "[47] Rohit Jena, Ganesh Subramanian Iyer, Siddharth Choudhary, Brandon Smith, Pratik Chaudhari, and James Gee. Splatarmor: Articulated gaussian splatting for animatable humans from monocular rgb videos. arXiv preprint arXiv:2311.10812, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 580, + 547, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 580, + 547, + 633 + ], + "spans": [ + { + "bbox": [ + 312, + 580, + 547, + 633 + ], + "type": "text", + "content": "[48] Boyi Jiang, Yang Hong, Hujun Bao, and Juyong Zhang. Self-con: Self reconstruction your digital avatar from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5605-5615, 2022. 1, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 635, + 547, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 635, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 312, + 635, + 547, + 689 + ], + "type": "text", + "content": "[49] Tianjian Jiang, Xu Chen, Jie Song, and Otmar Hilliges. Instantavatar: Learning avatars from monocular video in 60 seconds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16922-16932, 2023. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 691, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 691, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 691, + 547, + 713 + ], + "type": "text", + "content": "[50] Wei Jiang, Kwang Moo Yi, Golnoosh Samei, Oncel Tuzel, and Anurag Ranjan. Neuman: Neural human radiance field" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20427" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 72, + 72, + 288, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 72, + 288, + 94 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 288, + 94 + ], + "type": "text", + "content": "from a single video. arXiv preprint arXiv:2203.12575, 2022. 1, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 96, + 288, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 96, + 288, + 140 + ], + "spans": [ + { + "bbox": [ + 53, + 96, + 288, + 140 + ], + "type": "text", + "content": "[51] Yuming Jiang, Shuai Yang, Haonan Qiu, Wayne Wu, Chen Change Loy, and Ziwei Liu. Text2human: Text-driven controllable human image generation. ACM Transactions on Graphics (TOG), 41(4):1-11, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 141, + 288, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 141, + 288, + 184 + ], + "spans": [ + { + "bbox": [ + 53, + 141, + 288, + 184 + ], + "type": "text", + "content": "[52] Yuming Jiang, Shuai Yang, Tong Liang Koh, Wayne Wu, Chen Change Loy, and Ziwei Liu. Text2performer: Text-driven human video generation. arXiv preprint arXiv:2304.08483, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 186, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 186, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 53, + 186, + 288, + 239 + ], + "type": "text", + "content": "[53] Mohammad Mahdi Johari, Yann Lepoittevin, and François Fleuret. Geonerf: Generalizing nerf with geometry priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18365-18375, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 241, + 288, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 241, + 288, + 295 + ], + "spans": [ + { + "bbox": [ + 53, + 241, + 288, + 295 + ], + "type": "text", + "content": "[54] HyunJun Jung, Nikolas Brasch, Jifei Song, Eduardo Perez-Pellitero, Yiren Zhou, Zhihao Li, Nassir Navab, and Benjamin Busam. Deformable 3d gaussian splatting for animatable human avatars. arXiv preprint arXiv:2312.15059, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 297, + 288, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 297, + 288, + 330 + ], + "spans": [ + { + "bbox": [ + 53, + 297, + 288, + 330 + ], + "type": "text", + "content": "[55] James T Kajiya and Brian P Von Herzen. Ray tracing volume densities. ACM SIGGRAPH computer graphics, 18(3):165-174, 1984. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 332, + 288, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 332, + 288, + 374 + ], + "spans": [ + { + "bbox": [ + 53, + 332, + 288, + 374 + ], + "type": "text", + "content": "[56] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 42(4):1-14, 2023. 2, 3, 4, 5, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 376, + 288, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 376, + 288, + 409 + ], + "spans": [ + { + "bbox": [ + 53, + 376, + 288, + 409 + ], + "type": "text", + "content": "[57] Muhammed Kocabas, Jen-Hao Rick Chang, James Gabriel, Oncel Tuzel, and Anurag Ranjan. Hugs: Human gaussian splats. arXiv preprint arXiv:2311.17910, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 411, + 288, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 411, + 288, + 454 + ], + "spans": [ + { + "bbox": [ + 53, + 411, + 288, + 454 + ], + "type": "text", + "content": "[58] Georgios Kopanas, Julien Philip, Thomas Leimkuhler, and George Drettakis. Point-based neural rendering with perview optimization. In Computer Graphics Forum, pages 29-43. Wiley Online Library, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 456, + 288, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 456, + 288, + 498 + ], + "spans": [ + { + "bbox": [ + 53, + 456, + 288, + 498 + ], + "type": "text", + "content": "[59] Georgios Kopanas, Thomas Leimkuhler, Gilles Rainer, Clément Jambon, and George Drettakis. Neural point catauustics for novel-view synthesis of reflections. ACM Transactions on Graphics (TOG), 41(6):1-15, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 501, + 288, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 501, + 288, + 554 + ], + "spans": [ + { + "bbox": [ + 53, + 501, + 288, + 554 + ], + "type": "text", + "content": "[60] Youngjoong Kwon, Dahun Kim, Duygu Ceylan, and Henry Fuchs. Neural human performer: Learning generalizable radiance fields for human performance rendering. Advances in Neural Information Processing Systems, 34:24741-24752, 2021. 1, 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 556, + 288, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 556, + 288, + 600 + ], + "spans": [ + { + "bbox": [ + 53, + 556, + 288, + 600 + ], + "type": "text", + "content": "[61] Christoph Lassner and Michael Zollhofer. Pulsar: Efficient sphere-based neural rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1440-1449, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 601, + 288, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 601, + 288, + 634 + ], + "spans": [ + { + "bbox": [ + 53, + 601, + 288, + 634 + ], + "type": "text", + "content": "[62] Marc Levoy and Pat Hanrahan. Light field rendering. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 441-452. 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 635, + 288, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 635, + 288, + 678 + ], + "spans": [ + { + "bbox": [ + 53, + 635, + 288, + 678 + ], + "type": "text", + "content": "[63] Kathleen M Lewis, Srivatsan Varadharajan, and Ira Kemelmacher-Shlizerman. Tryongan: Body-aware try-on via layered interpolation. ACM Transactions on Graphics (TOG), 40(4):1-10, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 681, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 681, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 681, + 288, + 712 + ], + "type": "text", + "content": "[64] Ruilong Li, Matthew Tancik, and Angjoo Kanazawa. Nerve: A general nef acceleration toolbox. arXiv preprint arXiv:2210.04847, 2022. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 312, + 73, + 547, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 547, + 139 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 547, + 139 + ], + "type": "text", + "content": "[65] Tingting Liao, Xiaomei Zhang, Yuliang Xiu, Hongwei Yi, Xudong Liu, Guo-Jun Qi, Yong Zhang, Xuan Wang, Xi-angyu Zhu, and Zhen Lei. High-fidelity clothed avatar reconstruction from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8662–8672, 2023. 1, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 141, + 547, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 141, + 547, + 184 + ], + "spans": [ + { + "bbox": [ + 312, + 141, + 547, + 184 + ], + "type": "text", + "content": "[66] Haotong Lin, Sida Peng, Zhen Xu, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Efficient neural radiance fields for interactive free-viewpoint video. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 186, + 547, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 186, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 312, + 186, + 547, + 228 + ], + "type": "text", + "content": "[67] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 231, + 547, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 231, + 547, + 284 + ], + "spans": [ + { + "bbox": [ + 312, + 231, + 547, + 284 + ], + "type": "text", + "content": "[68] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. ACM transactions on graphics (TOG), 40(6):1-16, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 286, + 547, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 286, + 547, + 351 + ], + "spans": [ + { + "bbox": [ + 312, + 286, + 547, + 351 + ], + "type": "text", + "content": "[69] Yuan Liu, Sida Peng, Lingjie Liu, Qianqian Wang, Peng Wang, Christian Theobalt, Xiaowei Zhou, and Wenping Wang. Neural rays for occlusion-aware image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7824-7833, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 354, + 547, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 354, + 547, + 397 + ], + "spans": [ + { + "bbox": [ + 312, + 354, + 547, + 397 + ], + "type": "text", + "content": "[70] Stephen Lombardi, Tomas Simon, Gabriel Schwartz, Michael Zollhoefer, Yaser Sheikh, and Jason Saragih. Mixture of volumetric primitives for efficient neural rendering. ACM Transactions on Graphics (ToG), 40(4):1-13, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 399, + 547, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 399, + 547, + 442 + ], + "spans": [ + { + "bbox": [ + 312, + 399, + 547, + 442 + ], + "type": "text", + "content": "[71] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi-person linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia), 34(6):248:1-248:16, 2015. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 445, + 547, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 445, + 547, + 487 + ], + "spans": [ + { + "bbox": [ + 312, + 445, + 547, + 487 + ], + "type": "text", + "content": "[72] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 2, 3, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 489, + 547, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 489, + 547, + 520 + ], + "spans": [ + { + "bbox": [ + 312, + 489, + 547, + 520 + ], + "type": "text", + "content": "[73] Nelson Max. Optical models for direct volume rendering. IEEE Transactions on Visualization and Computer Graphics, 1(2):99-108, 1995. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 523, + 547, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 523, + 547, + 578 + ], + "spans": [ + { + "bbox": [ + 312, + 523, + 547, + 578 + ], + "type": "text", + "content": "[74] Yifang Men, Yiming Mao, Yuning Jiang, Wei-Ying Ma, and Zhouhui Lian. Controllable person image synthesis with attribute-decomposed gan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5084-5093, 2020. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 579, + 547, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 579, + 547, + 632 + ], + "spans": [ + { + "bbox": [ + 312, + 579, + 547, + 632 + ], + "type": "text", + "content": "[75] Marko Mihajlovic, Yan Zhang, Michael J Black, and Siyu Tang. Leap: Learning articulated occupancy of people. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10461-10471, 2021. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 635, + 547, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 635, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 312, + 635, + 547, + 689 + ], + "type": "text", + "content": "[76] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 3, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 692, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 692, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 692, + 547, + 713 + ], + "type": "text", + "content": "[77] Arthur Moreau, Jifei Song, Helisa Dhamo, Richard Shaw, Yiren Zhou, and Eduardo Pérez-Pellitero. Human gaussian" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20428" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 73, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 72, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 72, + 73, + 287, + 95 + ], + "type": "text", + "content": "splatting: Real-time rendering of animatable avatars. arXiv preprint arXiv:2311.17113, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 96, + 288, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 96, + 288, + 140 + ], + "spans": [ + { + "bbox": [ + 53, + 96, + 288, + 140 + ], + "type": "text", + "content": "[78] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 141, + 288, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 141, + 288, + 194 + ], + "spans": [ + { + "bbox": [ + 53, + 141, + 288, + 194 + ], + "type": "text", + "content": "[79] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. arXiv preprint arXiv:2112.00724, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 197, + 288, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 197, + 288, + 241 + ], + "spans": [ + { + "bbox": [ + 53, + 197, + 288, + 241 + ], + "type": "text", + "content": "[80] Atsuhiro Noguchi, Xiao Sun, Stephen Lin, and Tatsuya Harada. Neural articulated radiance field. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5762-5772, 2021. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 243, + 288, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 243, + 288, + 285 + ], + "spans": [ + { + "bbox": [ + 53, + 243, + 288, + 285 + ], + "type": "text", + "content": "[81] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Animatable neural radiance fields for modeling dynamic human bodies. In ICCV, 2021. 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 287, + 288, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 287, + 288, + 353 + ], + "spans": [ + { + "bbox": [ + 53, + 287, + 288, + 353 + ], + "type": "text", + "content": "[82] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9054-9063, 2021. 1, 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 354, + 288, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 354, + 288, + 399 + ], + "spans": [ + { + "bbox": [ + 53, + 354, + 288, + 399 + ], + "type": "text", + "content": "[83] Sida Peng, Shangzhan Zhang, Zhen Xu, Chen Geng, Boyi Jiang, Hujun Bao, and Xiaowei Zhou. Animatable neural implicit surfaces for creating avatars from videos. arXiv preprint arXiv:2203.08133, 4(5), 2022. 1, 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 399, + 288, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 399, + 288, + 432 + ], + "spans": [ + { + "bbox": [ + 53, + 399, + 288, + 432 + ], + "type": "text", + "content": "[84] Eric Penner and Li Zhang. Soft 3d reconstruction for view synthesis. ACM Transactions on Graphics (TOG), 36(6): 1-11, 2017. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 434, + 288, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 434, + 288, + 487 + ], + "spans": [ + { + "bbox": [ + 53, + 434, + 288, + 487 + ], + "type": "text", + "content": "[85] Hanspeter Pfister, Matthias Zwicker, Jeroen Van Baar, and Markus Gross. Surfels: Surface elements as rendering primitives. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 335-342, 2000. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 489, + 288, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 489, + 288, + 544 + ], + "spans": [ + { + "bbox": [ + 53, + 489, + 288, + 544 + ], + "type": "text", + "content": "[86] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14335-14345, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 545, + 288, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 545, + 288, + 578 + ], + "spans": [ + { + "bbox": [ + 53, + 545, + 288, + 578 + ], + "type": "text", + "content": "[87] Konstantinos Rematas, Ricardo Martin-Brualla, and Vittorio Ferrari. Sharf: Shape-conditioned radiance fields from a single view. arXiv preprint arXiv:2102.08860, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 580, + 288, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 580, + 288, + 643 + ], + "spans": [ + { + "bbox": [ + 53, + 580, + 288, + 643 + ], + "type": "text", + "content": "[88] Edoardo Remelli, Timur Bagautdinov, Shunsuke Saito, Chenglei Wu, Tomas Simon, Shih-En Wei, Kaiwen Guo, Zhe Cao, Fabian Prada, Jason Saragih, et al. Drivable volumetric avatars using texel-aligned features. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1–9, 2022. 3, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 647, + 288, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 647, + 288, + 690 + ], + "spans": [ + { + "bbox": [ + 53, + 647, + 288, + 690 + ], + "type": "text", + "content": "[89] Liu Ren, Hanspeter Pfister, and Matthias Zwicker. Object space ewa surface splatting: A hardware accelerated approach to high quality point rendering. In Computer Graphics Forum, pages 461-470. Wiley Online Library, 2002. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 692, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 692, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 692, + 288, + 712 + ], + "type": "text", + "content": "[90] Helge Rhodin, Nadia Robertini, Christian Richardt, Hans-Peter Seidel, and Christian Theobalt. A versatile scene" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 331, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 331, + 73, + 545, + 106 + ], + "type": "text", + "content": "model with differentiable visibility applied to generative pose estimation. In Proceedings of the IEEE International Conference on Computer Vision, pages 765-773, 2015. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 107, + 547, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 107, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 312, + 107, + 547, + 150 + ], + "type": "text", + "content": "[91] Gernot Riegler and Vladlen Koltun. Free view synthesis. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIX 16, pages 623-640. Springer, 2020. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 152, + 547, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 152, + 547, + 184 + ], + "spans": [ + { + "bbox": [ + 312, + 152, + 547, + 184 + ], + "type": "text", + "content": "[92] Darius Rückert, Linus Franke, and Marc Stamminger. Adop: Approximate differentiable one-pixel point rendering. ACM Transactions on Graphics (ToG), 41(4):1-14, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 185, + 547, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 185, + 547, + 215 + ], + "spans": [ + { + "bbox": [ + 312, + 185, + 547, + 215 + ], + "type": "text", + "content": "[93] Miguel Sainz and Renato Pajarola. Point-based rendering techniques. Computers & Graphics, 28(6):869-879, 2004. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 217, + 547, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 217, + 547, + 272 + ], + "spans": [ + { + "bbox": [ + 312, + 217, + 547, + 272 + ], + "type": "text", + "content": "[94] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2304-2314, 2019. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 273, + 547, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 273, + 547, + 327 + ], + "spans": [ + { + "bbox": [ + 312, + 273, + 547, + 327 + ], + "type": "text", + "content": "[95] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 84-93, 2020. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 327, + 547, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 327, + 547, + 382 + ], + "spans": [ + { + "bbox": [ + 312, + 327, + 547, + 382 + ], + "type": "text", + "content": "[96] Shunsuke Saito, Jinlong Yang, Qianli Ma, and Michael J Black. Scanimate: Weakly supervised learning of skinned clothed avatar networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2886-2897, 2021. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 383, + 547, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 383, + 547, + 426 + ], + "spans": [ + { + "bbox": [ + 312, + 383, + 547, + 426 + ], + "type": "text", + "content": "[97] Kripasindhu Sarkar, Lingjie Liu, Vladislav Golyanik, and Christian Theobalt. Humangan: A generative model of human images. In 2021 International Conference on 3D Vision (3DV), pages 258-267. IEEE, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 427, + 547, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 427, + 547, + 470 + ], + "spans": [ + { + "bbox": [ + 312, + 427, + 547, + 470 + ], + "type": "text", + "content": "[98] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 2, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 472, + 547, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 472, + 547, + 536 + ], + "spans": [ + { + "bbox": [ + 312, + 472, + 547, + 536 + ], + "type": "text", + "content": "[99] Johannes L Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 501-518. Springer, 2016. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 537, + 547, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 537, + 547, + 592 + ], + "spans": [ + { + "bbox": [ + 307, + 537, + 547, + 592 + ], + "type": "text", + "content": "[100] Vincent Sitzmann, Justus Thies, Felix Heide, Matthias Nießner, Gordon Wetzstein, and Michael Zollhofer. Deepvoxels: Learning persistent 3d feature embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2437-2446, 2019. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 593, + 547, + 625 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 593, + 547, + 625 + ], + "spans": [ + { + "bbox": [ + 307, + 593, + 547, + 625 + ], + "type": "text", + "content": "[101] Noah Snavely, Steven M Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. In ACM siggraph 2006 papers, pages 835-846. 2006. 2, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 626, + 547, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 626, + 547, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 626, + 547, + 679 + ], + "type": "text", + "content": "[102] Carsten Stoll, Nils Hasler, Juergen Gall, Hans-Peter Seidel, and Christian Theobalt. Fast articulated motion tracking using a sums of gaussians body model. In 2011 International Conference on Computer Vision, pages 951-958. IEEE, 2011. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 681, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 547, + 713 + ], + "type": "text", + "content": "[103] Shih-Yang Su, Frank Yu, Michael Zollhöfer, and Helge Rhodin. A-nerf: Articulated neural radiance fields for learning human shape, appearance, and pose. Advances in Neural" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20429" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 72, + 72, + 288, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 72, + 288, + 94 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 288, + 94 + ], + "type": "text", + "content": "Information Processing Systems, 34:12278-12291, 2021. 1, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 288, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 288, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 288, + 163 + ], + "type": "text", + "content": "[104] Zhuo Su, Lan Xu, Zerong Zheng, Tao Yu, Yebin Liu, and Lu Fang. Robustfusion: Human volumetric capture with data-driven visual cues using a rgbd camera. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16, pages 246-264. Springer, 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 164, + 288, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 164, + 288, + 218 + ], + "spans": [ + { + "bbox": [ + 48, + 164, + 288, + 218 + ], + "type": "text", + "content": "[105] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 220, + 288, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 288, + 296 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 288, + 296 + ], + "type": "text", + "content": "[106] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3d shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11358-11367, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 298, + 288, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 288, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 288, + 353 + ], + "type": "text", + "content": "[107] Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srinivasan, Edgar Tretschk, Wang Yifan, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lombardi, et al. Advances in neural rendering. In Computer Graphics Forum, pages 703-735. Wiley Online Library, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 354, + 288, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 288, + 397 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 288, + 397 + ], + "type": "text", + "content": "[108] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: Image synthesis using neural textures. Acm Transactions on Graphics (TOG), 38(4):1-12, 2019. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 399, + 288, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 288, + 454 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 288, + 454 + ], + "type": "text", + "content": "[109] Garvita Tiwari, Nikolaos Sarafianos, Tony Tung, and Gerard Pons-Moll. Neural-gif: Neural generalized implicit functions for animating people in clothing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11708-11718, 2021. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 456, + 288, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 288, + 499 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 288, + 499 + ], + "type": "text", + "content": "[110] Alex Trevithick and Bo Yang. Grf: Learning a general radiance field for 3d representation and rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15182-15192, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 501, + 288, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 288, + 555 + ], + "type": "text", + "content": "[111] Angtian Wang, Peng Wang, Jian Sun, Adam Kortylewski, and Alan Yuille. Voge: a differentiable volume renderer using gaussian ellipsoids for analysis-by-synthesis. In The Eleventh International Conference on Learning Representations, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 557, + 288, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 288, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 288, + 601 + ], + "type": "text", + "content": "[112] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 602, + 288, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 288, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 288, + 668 + ], + "type": "text", + "content": "[113] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snively, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690-4699, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "text", + "content": "[114] Shaofei Wang, Katja Schwarz, Andreas Geiger, and Siyu Tang. Arah: Animatable volume rendering of articulated human sdfs. In European conference on computer vision, 2022. 1, 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "type": "text", + "content": "[115] Chung-Yi Weng, Brian Curless, and Ira Kemelmacher-Shlizerman. Vid2actor: Free-viewpoint animatable person synthesis from video in the wild. arXiv preprint arXiv:2012.12884, 2020. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 118, + 547, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 183 + ], + "type": "text", + "content": "[116] Chung-Yi Weng, Brian Curless, Pratul P. Srinivasan, Jonathan T. Barron, and Ira Kemelmacher-Shlizerman. HumanNeRF: Free-viewpoint rendering of moving people from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16210-16220, 2022. 1, 3, 5, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 184, + 547, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 184, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 184, + 547, + 217 + ], + "type": "text", + "content": "[117] Zhenzhen Weng, Zeyu Wang, and Serena Yeung. Zero-)rator: Zero-shot 3d avatar generation from a single image. arXiv preprint arXiv:2305.16411, 2023. 1, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 217, + 547, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 217, + 547, + 270 + ], + "spans": [ + { + "bbox": [ + 307, + 217, + 547, + 270 + ], + "type": "text", + "content": "[118] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. Synsin: End-to-end view synthesis from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7467-7477, 2020. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 272, + 547, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 272, + 547, + 316 + ], + "spans": [ + { + "bbox": [ + 307, + 272, + 547, + 316 + ], + "type": "text", + "content": "[119] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528, 2023. 2, 3, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 317, + 547, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 317, + 547, + 360 + ], + "spans": [ + { + "bbox": [ + 307, + 317, + 547, + 360 + ], + "type": "text", + "content": "[120] Hongyi Xu, Thiemo Alldieck, and Cristian Sminchisescu. H-nerf: Neural radiance fields for rendering and temporal reconstruction of humans in motion. Advances in Neural Information Processing Systems, 34, 2021. 1, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 361, + 547, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 361, + 547, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 361, + 547, + 415 + ], + "type": "text", + "content": "[121] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5438–5448, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 416, + 547, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 416, + 547, + 459 + ], + "spans": [ + { + "bbox": [ + 307, + 416, + 547, + 459 + ], + "type": "text", + "content": "[122] Zhen Xu, Sida Peng, Haotong Lin, Guangzhao He, Jiaming Sun, Yujun Shen, Hujun Bao, and Xiaowei Zhou. 4k4d: Real-time 4d view synthesis at 4k resolution. arXiv preprint arXiv:2310.11448, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 460, + 547, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 460, + 547, + 514 + ], + "spans": [ + { + "bbox": [ + 307, + 460, + 547, + 514 + ], + "type": "text", + "content": "[123] Jiayu Yang, Wei Mao, Jose M Alvarez, and Miaomiao Liu. Cost volume pyramid based depth inference for multi-view stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4877-4886, 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 514, + 547, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 514, + 547, + 581 + ], + "spans": [ + { + "bbox": [ + 307, + 514, + 547, + 581 + ], + "type": "text", + "content": "[124] Ze Yang, Shenlong Wang, Sivabalan Manivasagam, Zeng Huang, Wei-Chiu Ma, Xinchen Yan, Ersin Yumer, and Raquel Urtasun. S3: Neural shape, skeleton, and skinn-ning fields for 3d human modeling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13284–13293, 2021. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 582, + 547, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 582, + 547, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 582, + 547, + 624 + ], + "type": "text", + "content": "[125] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101, 2023. 2, 3, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 625, + 547, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 625, + 547, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 547, + 668 + ], + "type": "text", + "content": "[126] Zeyu Yang, Hongye Yang, Zijie Pan, Xiatian Zhu, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. arXiv preprint arXiv:2310.10642, 2023. 2, 3, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 670, + 547, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 547, + 712 + ], + "type": "text", + "content": "[127] Wang Yifan, Felice Serena, Shihao Wu, Cengiz Öztireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM Transactions on Graphics (TOG), 38(6):1-14, 2019. 3" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20430" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 476 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "text", + "content": "[128] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoptrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5752-5761, 2021. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 129, + 288, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 129, + 288, + 161 + ], + "spans": [ + { + "bbox": [ + 49, + 129, + 288, + 161 + ], + "type": "text", + "content": "[129] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelNeRF: Neural radiance fields from one or few images. In CVPR, 2021. 3, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 288, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 288, + 196 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 288, + 196 + ], + "type": "text", + "content": "[130] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 198, + 288, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 198, + 288, + 241 + ], + "spans": [ + { + "bbox": [ + 49, + 198, + 288, + 241 + ], + "type": "text", + "content": "[131] Qiang Zhang, Seung-Hwan Baek, Szymon Rusinkiewicz, and Felix Heide. Differentiable point-based radiance fields for efficient view synthesis. In SIGGRAPH Asia 2022 Conference Papers, pages 1-12, 2022. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 242, + 288, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 242, + 288, + 285 + ], + "spans": [ + { + "bbox": [ + 49, + 242, + 288, + 285 + ], + "type": "text", + "content": "[132] Fuqiang Zhao, Wei Yang, Jiakai Zhang, Pei Lin, Yingliang Zhang, Jingyi Yu, and Lan Xu. Humannerf: Generalizable neural human radiance field from sparse inputs. arXiv preprint arXiv:2112.02789, 3, 2021. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 287, + 288, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 287, + 288, + 342 + ], + "spans": [ + { + "bbox": [ + 49, + 287, + 288, + 342 + ], + "type": "text", + "content": "[133] Tinghui Zhou, Shubham Tulsiani, Weilun Sun, Jitendra Malik, and Alexei A Efros. View synthesis by appearance flow. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 286-301. Springer, 2016. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 343, + 288, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 343, + 288, + 385 + ], + "spans": [ + { + "bbox": [ + 49, + 343, + 288, + 385 + ], + "type": "text", + "content": "[134] Wojciech Zielonka, Timur Bagautdinov, Shunsuke Saito, Michael Zollhöfer, Justus Thies, and Javier Romero. Drivable 3d gaussian avatars. arXiv preprint arXiv:2311.08581, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 387, + 288, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 387, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 49, + 387, + 288, + 430 + ], + "type": "text", + "content": "[135] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Ewa volume splatting. In Proceedings Visualization, 2001. VIS'01., pages 29-538. IEEE, 2001. 2, 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 432, + 288, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 432, + 288, + 476 + ], + "spans": [ + { + "bbox": [ + 49, + 432, + 288, + 476 + ], + "type": "text", + "content": "[136] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Surface splatting. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 371-378, 2001. 3" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "20431" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Gaussian Head Avatar_ Ultra High-fidelity Head Avatar via Dynamic Gaussians/24a06516-83b5-41f4-8bb5-a4da984c01e1_content_list.json b/2024/Gaussian Head Avatar_ Ultra High-fidelity Head Avatar via Dynamic Gaussians/24a06516-83b5-41f4-8bb5-a4da984c01e1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0ab2397da3f50fd43cb98906989388ec7f31ecc4 --- /dev/null +++ b/2024/Gaussian Head Avatar_ Ultra High-fidelity Head Avatar via Dynamic Gaussians/24a06516-83b5-41f4-8bb5-a4da984c01e1_content_list.json @@ -0,0 +1,1703 @@ +[ + { + "type": "text", + "text": "Gaussian Head Avatar: Ultra High-fidelity Head Avatar via Dynamic Gaussians", + "text_level": 1, + "bbox": [ + 81, + 130, + 887, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuelang Xu $^{1}$ , Bengwang Chen $^{1}$ , Zhe Li $^{1}$ , Hongwen Zhang $^{2}$ , Lizhen Wang $^{1}$ , Zerong Zheng $^{3}$ , Yebin Liu $^{1\\dagger}$ $^{1}$ Department of Automation, Tsinghua University $^{2}$ Beijing Normal University $^{3}$ NNKosmos Technology", + "bbox": [ + 86, + 179, + 915, + 218 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/566be093fd8aedb6b1d056c06b0149e2f2c2f1bfa9da039746530b81ca4983e3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 76, + 224, + 893, + 375 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b2f5a3e47a3df591cc0811e9b5414e04ffbb9182328730acab361413101d2e75.jpg", + "image_caption": [ + "Figure 1. Gaussian head avatar achieves ultra high-fidelity image synthesis with controllable expressions at 2K resolution. The above shows different views of the synthesized avatar, and the bottom shows different identities animated by the same expression. 16 views are used during the training." + ], + "image_footnote": [], + "bbox": [ + 76, + 376, + 890, + 523 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 585, + 313, + 599 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Creating high-fidelity 3D head avatars has always been a research hotspot, but there remains a great challenge under lightweight sparse view setups. In this paper, we propose Gaussian Head Avatar represented by controllable 3D Gaussians for high-fidelity head avatar modeling. We optimize the neutral 3D Gaussians and a fully learned MLP-based deformation field to capture complex expressions. The two parts benefit each other, thereby our method can model fine-grained dynamic details while ensuring expression accuracy. Furthermore, we devise a well-designed geometry-guided initialization strategy based on implicit SDF and Deep Marching Tetrahedra for the stability and convergence of the training procedure. Experiments show our approach outperforms other state-of-the-art sparse-view methods, achieving ultra high-fidelity rendering quality at $2K$ resolution even under exaggerated expressions. Project page: https://yuelangx.github.io/gaussianheadavatar.", + "bbox": [ + 75, + 603, + 473, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 585, + 630, + 599 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "High-fidelity 3D human head avatar modeling is of great significance in many fields, such as VR/AR, telepresence, digital human and film production. Automatically creating high-fidelity avatars has been a research hotspot in computer vision for decades. Although some traditional head avatars [37, 39, 41, 54] can realize high-fidelity animation, they typically require accurate geometries reconstructed and tracked from dense multi-view videos, thus limiting their applications in lightweight settings.", + "bbox": [ + 496, + 612, + 890, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Thanks to the Neural Radiance Fields (NeRF) [43] which show great capability of novel view synthesis in the absence of accurate geometry, recent methods [38, 57] skip the geometry reconstruction and tracking steps but directly learn high-quality NeRF-based head avatars. Other works [42, 48, 71] have verified that NeRF can be applied to either dense or sparse views, which greatly lowers the threshold for head avatar reconstruction. However, it still remains challenging for these NeRF-based approaches to synthesize high-fidelity images at 2K resolutions with pixel", + "bbox": [ + 496, + 750, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding author.", + "bbox": [ + 101, + 887, + 228, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1931", + "bbox": [ + 483, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "level details, including wrinkles and eyes.", + "bbox": [ + 76, + 90, + 354, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome this bottleneck and further improve the avatar quality, we introduce 3D Gaussian splatting [26] for 3D head avatar modeling. This is an explicit discrete representation that can be well adapted to the rasterization-based rendering pipeline. It has been verified that the 3D Gaussian representation is capable of rendering complex scenes with low computational consumption. Compared to NeRF, the reconstruction quality of static and dynamic scenes [40, 59, 67] is much better while rendering time cost has been significantly reduced. Motivated by this progress, we propose Gaussian Head Avatar, a novel representation that utilizes 3D Gaussian splatting for ultra high-fidelity head avatar modeling. Although recent 4D Gaussian works [40, 59, 66, 67] have been proposed to reconstruct dynamic scenes, all of them cannot be animated. For modeling the animatable head avatar, it is crucial but still unexplored how to effectively control the deformation of 3D Gaussians and model the dynamic appearances through expression coefficients.", + "bbox": [ + 75, + 109, + 470, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Previous explicit [74] and implicit [2, 73, 76] head avatars usually formulate the facial deformation via linear blend skinning (LBS) using the skinning weights and blendshapes like the FLAME model [32]. However, such an LBS-based formulation fails to represent exaggerated and fine-grained expressions by simple linear operations, limiting the representation ability of the head avatars. Inspired by NeRSemble [28], we propose a fully learnable expression-conditioned deformation field for the 3D head Gaussians, avoiding the limited capability of the LBS-based formulation. Specifically, we input the positions of the 3D Gaussians with expression coefficients into an MLP to directly predict the displacements from the neutral expression to the target one. Similarly, we control the motion of nonface areas, such as the neck, using the head pose as the condition. 3D Gaussian-based representation has the powerful ability to reconstruct high-frequency details, enabling our method to learn accurate deformation fields. In turn, the learned accurate deformation field facilitates the dynamic Gaussian head model to fit more dynamic details. As a result, our method is able to reconstruct finer-grained dynamic details of expressive human heads.", + "bbox": [ + 75, + 398, + 470, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unfortunately, as a discretized representation, the gradients back-propagated to the 3D Gaussians cannot spread through the whole space. Thus the convergence of training heavily relies on a plausible initialization for both the geometry and the deformation field. However, simply initializing the 3D head Gaussians with a morphable template like FLAME [32] fails to model the long hairstyle and the shoulders. Hence, we further propose an efficient and well-designed geometry-guided initialization strategy. Specifically, instead of starting from stochastic Gaussians or a FLAME model, we initially optimize an implicit signed dis", + "bbox": [ + 75, + 734, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tance function (SDF) field along with a color field and a deformation MLP for modeling the basic geometry, color, and the expression-conditioned deformations of the head avatar respectively. The SDF field is converted to a mesh through Deep Marching Tetrahedra (DMTet) [52], with the color and deformation of the vertices predicted by the MLPs. Then we render the mesh and optimize them jointly under the supervision of multi-view RGB images. Finally, we use the mesh with per-vertex features from the SDF field to initialize the 3D Gaussians to lie on the basic head surface while the color and deformation MLPs are carried over to the next stage, ensuring stable training for convergence. The entire initialization process takes only around 10 minutes.", + "bbox": [ + 496, + 90, + 890, + 286 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contributions of our method can be summarized as:", + "bbox": [ + 517, + 287, + 887, + 301 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose Gaussian Head Avatar, a new head avatar representation that employs controllable dynamic 3D Gaussians to model expressive human head avatars, producing ultra high-fidelity synthesized images at 2K resolutions.", + "- For modeling high-frequency dynamic details, we employ a fully learned deformation field upon the 3D head Gaussians, which accurately model extremely complex and exaggerated facial expressions.", + "- We carefully design an efficient initialization strategy that leverages implicit representations to initialize the geometry and deformation, leading to efficient and robust convergence when training the Gaussian Head Avatar." + ], + "bbox": [ + 500, + 303, + 890, + 483 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Benefiting from these contributions, our method surpasses recent state-of-the-art methods under lightweight sparse-view setups on the avatar quality by a large margin.", + "bbox": [ + 498, + 484, + 890, + 529 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 542, + 650, + 558 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Head Avatar Reconstruction. Due to the wide application value in the film and digital human industry, 3D head avatar reconstruction from multi-view images has always been a research hotspot. Traditional works [4, 8, 19, 31] reconstruct the scan geometry through multi-view stereo and then register a face mesh template to it. However, such methods usually require heavy computation. With the utilization of deep neural networks, current methods [7, 33, 61, 65] achieve very fast reconstruction, producing even more accurate geometry. Lombardi et al. [37], Bi et al. [5] and Ma et al. [41] represent the full head mesh through a deep neural network and train it with multi-view videos as supervision. However, due to the errors in geometric estimation, mesh-based head avatars typically suffer from texture blur. Therefore, some recent methods [38, 57] utilize NeRF representation [43] to synthesize novel view images without geometry reconstruction, or build NeRF on the head mesh template [39]. Furthermore, the NeRF-based methods are extended to sparse view reconstruction tasks [28, 42, 48, 71] and achieve impressive performance.", + "bbox": [ + 496, + 566, + 890, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Methods which focus on generative model [6, 9, 11, 32, 46, 56, 60] are dedicated to learning general mesh", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1932", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "face templates from large-scale multi-view face images or 3D scans. Recently, implicit SDF-based [68] or NeRF-based [14, 23, 53, 55, 75] methods can learn full-head templates without the limitations of fixed topology, thereby better modeling complex hairstyles and glasses. Cao et al. [14] adopts a hybrid representation of local NeRF built on the mesh surface, which enables high-fidelity rendering and flexible expression control.", + "bbox": [ + 75, + 90, + 468, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D head avatars reconstruction from monocular videos is also a popular yet challenging research topic. Early methods [12, 13, 15, 24, 25, 45] optimize a morphable mesh to fit the training video. Recent methods [20, 27] leverage neural networks to learn non-rigid deformation upon 3DMM face templates [18, 32], thus can recover more dynamic details. Such methods are not flexible enough to handle complex topologies. Therefore, the latest methods explore to construct head avatar models based on implicit SDF [73], point clouds [74] or NeRF [2, 3, 16, 17, 21, 36, 47, 63, 64, 76].", + "bbox": [ + 75, + 212, + 470, + 362 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Point-based Rendering. Point elements as a discrete and unstructured representation can fit geometry with arbitrary topology [69] efficiently. Recent methods [29, 30, 58] open up a differentiable rasterization pipeline, such that the point-based representation is widely used in multi-view reconstruction tasks. Aliev et al. [1] and Ruckert et al. [49] propose to first render the feature map, which is transferred to the images through a convolutional renderer. Xu et al.[62] use neural point cloud associated with neural features to model a NeRF. Recently, 3D Gaussian splatting [26] shows its superior performance, beating NeRF in both novel view synthesis quality and rendering speed. Some approaches [34, 40, 50, 59, 66, 67, 72, 72] extend Gaussian representation to dynamic scene reconstruction. However, these methods can not be migrated to the head avatar reconstruction tasks.", + "bbox": [ + 75, + 363, + 468, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Overview", + "text_level": 1, + "bbox": [ + 76, + 618, + 181, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The pipeline of the reconstruction of Gaussian Head Avatar is illustrated in Fig. 2, including the initialization stage and the training stage of Gaussian Head Avatar. Before the beginning of the pipeline, we remove the background [35] of each image and jointly estimate the 3DMM model [18], 3D facial landmarks and the expression coefficients for each frame. In the initialization stage (Sec. 4.3), we reconstruct an SDF-based neutral geometry, and optimize a deformation MLP and a color MLP from the training data as the guidance model. Next, we extract the neutral mesh through DMTet to initialize the neutral Gaussians while the deformation and color MLPs are also inherited from the initialization stage. In the training stage of Gaussian Head Avatar (Sec. 4.2), we deform the neutral Gaussians to the target expression through the dynamic generator given the driving expression coefficients as the condition. Finally, given a camera view, the expressive Gaussians are rendered to a", + "bbox": [ + 75, + 643, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "feature map, which is then fed into the convolutional super resolution network to generate high-resolution avatar images. The whole model is optimized under the supervision of multi-view RGB videos.", + "bbox": [ + 496, + 90, + 890, + 150 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 500, + 167, + 591, + 184 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Avatar Representation", + "text_level": 1, + "bbox": [ + 500, + 194, + 709, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generally, the static 3D Gaussians [26] with $N$ points are represented by their positions $X$ , the multi-channel color $C$ , the rotation $Q$ , scale $S$ and opacity $A$ . The rotation $Q$ is represented in the form of quaternion. Subsequently, the Gaussians can be rasterized and rendered to the multi-channel image $I$ given the camera parameters $\\mu$ . This process can be formulated as:", + "bbox": [ + 496, + 218, + 890, + 323 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI = \\mathcal {R} (X, C, Q, S, A; \\mu). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 339, + 890, + 354 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our task is to reconstruct a dynamic head avatar controlled by expression coefficients. Therefore, we formulate the head avatar as dynamic 3D Gaussians conditioned on expressions. To handle the dynamic changes, we input the expression coefficients with head pose to the head avatar model and output the position and other attributes of the Gaussians as above.", + "bbox": [ + 496, + 369, + 890, + 473 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, we first construct a canonical neutral Gaussian model with expression-independent attributes: $\\{X_0, F_0, Q_0, S_0, A_0\\}$ , which are fully estimizable. $X_0 \\in \\mathbb{R}^{N \\times 3}$ denotes the positions of the Gaussians with a neutral expression in the canonical space. $F_0 \\in \\mathbb{R}^{N \\times 128}$ denotes the point-wise feature vectors as their intrinsic properties. $Q_0 \\in \\mathbb{R}^{N \\times 4}$ , $S_0 \\in \\mathbb{R}^{N \\times 3}$ and $A_0 \\in \\mathbb{R}^{N \\times 1}$ denotes the neutral rotation, scale and opacity respectively. Note that we do not define the neutral color, but directly predict expression-dependent dynamic color from the point-wise feature vectors $F_0$ . Then, we construct an MLP-based expression conditioned dynamic generator $\\Phi$ to generate all the extra dynamic changes to the neutral model. Overall, the whole Gaussian head avatar can be formulated as:", + "bbox": [ + 496, + 476, + 890, + 686 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\{X, C, Q, S, A \\} = \\Phi \\left(\\boldsymbol {X} _ {0}, \\boldsymbol {F} _ {0}, \\boldsymbol {Q} _ {0}, \\boldsymbol {S} _ {0}, \\boldsymbol {A} _ {0}; \\theta , \\beta\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 702, + 890, + 719 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with $\\theta$ denoting expression coefficients and $\\beta$ denoting the head pose. During the training, we optimize all the parameters of the dynamic generator $\\Phi$ and the neutral Gaussian model $\\{X_0, F_0, Q_0, S_0, A_0\\}$ , which are highlighted in bold in the following.", + "bbox": [ + 496, + 731, + 890, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Next, we explain the process of adding expression-related changes to the neutral Gaussian model through the dynamic generator $\\Phi$ as described in Eqn. 2 in detail.", + "bbox": [ + 496, + 809, + 890, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Positions $X^{\\prime}$ of the Gaussians. Expressions bring about the geometric deformation of the neutral model, which is modeled as the displacements of the Gaussian points.", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1933", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0532ff0b2d16e0d0cf8e507f5e09c0ad3a3cda6c445667a10def86811cf9672f.jpg", + "image_caption": [ + "Figure 2. The overview of the Gaussian Head Avatar rendering and reconstruction. We first optimize the guidance model including a neutral mesh, a deformation MLP and a color MLP in the Initialization stage. Then we use them to initialize the neutral Gaussians and the dynamic generator. Finally, 2K RGB images are synthesized through differentiable rendering and the super-resolution network. The Gaussian Head Avatar are trained under the supervision of multi-view RGB videos." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 895, + 246 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, we predict the displacements respectively controlled by the expression and the head pose in the canonical space through two different MLPs: $\\pmb{f}_{\\text{def}}^{exp} \\in \\Phi$ and $\\pmb{f}_{\\text{def}}^{\\text{pose}} \\in \\Phi$ . Then, we add them to the neutral positions.", + "bbox": [ + 76, + 324, + 470, + 388 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} X ^ {\\prime} = \\boldsymbol {X} _ {0} + \\lambda_ {e x p} (\\boldsymbol {X} _ {0}) \\boldsymbol {f} _ {d e f} ^ {e x p} (\\boldsymbol {X} _ {0}, \\theta) \\tag {3} \\\\ + \\lambda_ {p o s e} (\\boldsymbol {X} _ {0}) \\boldsymbol {f} _ {d e f} ^ {p o s e} (\\boldsymbol {X} _ {0}, \\beta). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 402, + 468, + 443 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\lambda_{exp}(\\cdot)$ and $\\lambda_{pose}(\\cdot)$ represent the extent to which the points are affected by the expression or the head pose respectively. Without decoupling but just using the expression coefficients as the global condition [16] which also controls the shoulders and upper body, will produce jittering results during the animation. Here, we assume that the Gaussian points closer to 3D landmarks are more affected by the expression coefficients and less affected by the head pose, while the opposite is true for the Gaussian points far away. Specifically, The 3D landmarks $P_0$ of the canonical model are first estimated through the 3DMM model in the data preprocessing and then optimized in the initialization stage 4.3. Then for each Gaussian point, we calculate the above weight $\\lambda_{exp}(\\cdot)$ and $\\lambda_{pose}(\\cdot)$ as follows:", + "bbox": [ + 75, + 455, + 468, + 670 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\lambda_ {e x p} (x) = \\left\\{ \\begin{array}{l l} 1, & d i s t (x, \\boldsymbol {P} _ {0}) < t _ {1} \\\\ \\frac {t _ {2} - d i s t (x , \\boldsymbol {P} _ {0})}{t _ {2} - t _ {1}}, & d i s t (x, \\boldsymbol {P} _ {0}) \\in [ t _ {1}, t _ {2} ] \\\\ 0, & d i s t (x, \\boldsymbol {P} _ {0}) > t _ {2} \\end{array} \\right.\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 683, + 442, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with $\\lambda_{\\text{pose}}(x) = 1 - \\lambda_{\\text{exp}}(x)$ . And $x \\in \\mathbf{X}_0$ denotes the position of one neutral Gaussian. $\\text{dist}(x, P_0)$ denotes the minimum distance from the point $x$ to the 3D landmarks $P_0$ . $t_1 = 0.15$ and $t_2 = 0.25$ are predefined hyperparameters when the length of the head is set to approximately 1.", + "bbox": [ + 76, + 747, + 468, + 837 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Color $C'$ of the Gaussians. Modeling the dynamic details typically requires dynamic color that changes with expressions. As we do not pre-define the neutral value in Eqn. 2, the color are directly predict by two color MLPs:", + "bbox": [ + 76, + 839, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\pmb{f}_{col}^{exp}\\in \\Phi$ and $\\pmb{f}_{col}^{pose}\\in \\Phi$", + "bbox": [ + 500, + 323, + 676, + 340 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} C ^ {\\prime} = \\lambda_ {e x p} \\left(\\boldsymbol {X} _ {0}\\right) \\boldsymbol {f} _ {c o l} ^ {e x p} \\left(\\boldsymbol {F} _ {0}, \\theta\\right) \\tag {4} \\\\ + \\lambda_ {p o s e} (\\boldsymbol {X} _ {0}) \\boldsymbol {f} _ {c o l} ^ {p o s e} (\\boldsymbol {F} _ {0}, \\beta). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 349, + 890, + 386 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Rotation, Scale and Opacity $\\{Q', S', A'\\}$ of the Gaussians. These three attributes also dynamic, thereby modeling some detailed expressions-related appearance changes. Here, we just use another two attribute MLPs $f_{att}^{exp} \\in \\Phi$ and $f_{att}^{pose} \\in \\Phi$ to predict their shift from the neutral value.", + "bbox": [ + 498, + 393, + 890, + 472 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\{Q ^ {\\prime}, S ^ {\\prime}, A ^ {\\prime} \\} = \\left\\{\\mathbf {Q} _ {0}, \\mathbf {S} _ {0}, \\mathbf {A} _ {0} \\right\\} \\\\ + \\lambda_ {e x p} (\\boldsymbol {X} _ {0}) \\boldsymbol {f} _ {a t t} ^ {e x p} \\left(\\boldsymbol {F} _ {0}, \\theta\\right) \\tag {5} \\\\ + \\lambda_ {p o s e} (\\boldsymbol {X} _ {0}) \\boldsymbol {f} _ {a t t} ^ {p o s e} (\\boldsymbol {F} _ {0}, \\beta). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 479, + 890, + 536 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, we apply rigid rotations and translations $T(\\cdot)$ to the Gaussians, transforming them from the canonical space to the world space. Note, the transformation is only implemented for directional variables: $\\{X', Q'\\}$ , while the multichannel color, the scale and the opacity $\\{C', S', A'\\}$ are not directional thus remain unchanged.", + "bbox": [ + 496, + 545, + 890, + 635 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\{X, Q \\} = T \\left(\\left\\{X ^ {\\prime}, Q ^ {\\prime} \\right\\}, \\beta\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 645, + 890, + 662 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\{C, S, A \\} = \\left\\{C ^ {\\prime}, S ^ {\\prime}, A ^ {\\prime} \\right\\}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 671, + 890, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Training", + "text_level": 1, + "bbox": [ + 500, + 696, + 604, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this part, we explain the training pipeline of the Gaussian head avatar 4.1 and the loss function. In each iteration, we first generate the expression conditioned 3D Gaussians as Eqn. 2. Then we render a 32-channel image with 512 resolution $I_C \\in \\mathbb{R}^{512 \\times 512 \\times 32}$ referring to Eqn. 1. After that we feed the image to a super resolution network $\\Psi$ to generate a 2048 resolution RGB image $I_{hr} \\in \\mathbb{R}^{2048 \\times 2048 \\times 3}$ , such that more details are recovered and noise caused by uneven ambient light or camera chromatic aberration in the training data will be filtered out [49, 64].", + "bbox": [ + 496, + 719, + 890, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During training, we jointly optimize all the learnable parameters mentioned above in bold, including the neutral", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1934", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Gaussians: $\\{X_0, F_0, Q_0, S_0, A_0\\}$ , the dynamic generator: $\\{f_{col}^{exp}, f_{col}^{pose}, f_{def}^{exp}, f_{def}^{pose}, f_{att}^{exp}, f_{att}^{pose}\\}$ , and the super resolution network $\\Psi$ . For the loss function, we only use the foreground RGB images $I_{gt}$ as supervision to construct an L1 loss and a VGG perceptual loss [70] between the generated images $I_{hr}$ and the ground truth $I_{gt}$ . Besides, we encourage the first three channels of the 32-channel feature image $I_C$ to be RGB channels, which is ensured by a L1 loss term. The total loss is:", + "bbox": [ + 76, + 90, + 468, + 224 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} = \\left\\| I _ {h r} - I _ {g t} \\right\\| _ {1} + \\lambda_ {v g g} V G G \\left(I _ {h r}, I _ {g t}\\right) \\tag {8} \\\\ + \\lambda_ {l r} | | I _ {l r} - I _ {g t} | | _ {1}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 233, + 467, + 268 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with $I_{lr}$ denoting the first three channels of the 32-channel image $I_C$ . We set the weights $\\lambda_{vgg} = 0.1$ and $\\lambda_{lr} = 1$ .", + "bbox": [ + 76, + 277, + 468, + 309 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Geometry-guided Initialization", + "text_level": 1, + "bbox": [ + 76, + 316, + 352, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unlike neural networks, the Gaussians act as an unordered and unstructured representation. Random initialization leads to failure to converge while naively using the FLAME model to initialize will significantly reduce the reconstruction quality. In this section, we describe in detail how to optimize a mesh guidance model to provide reliable initialization for the Gaussians in Sec. 4.1.", + "bbox": [ + 76, + 339, + 468, + 444 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Mesh Guidance Model. Specifically, we first construct a MLP $f_{sdf}$ to represent a signed distance field. In addition, this network will also output the corresponding feature vector of each point, which is used for predicting the point color. It can be formulated as:", + "bbox": [ + 76, + 445, + 468, + 518 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns, \\eta = \\boldsymbol {f} _ {s d f} (x), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 531, + 468, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with $s$ denotes the SDF value, $\\eta$ denotes the feature vector and $x$ denotes the point position. Then through (DMTet) [51], we can differentially extract a mesh with vertices $X$ , per-vertex feature vectors $F$ and its faces. We also predict the per-vertex 32-channel color as Eqn. 4 by the two color MLPs $f_{col}^{exp}$ and $f_{col}^{pose}$ . In parallel, we construct the two deformation MLPs: $f_{def}^{exp}$ and $f_{def}^{pose}$ as described in Sec. 4.1 to predict the displacements and add them to the vertex positions. This process is similar to Eqn 3 above, with the Gaussian positions $X_0$ replaced by the vertex positions $X$ . Finally we also apply rigid rotations and translations to the deformed mesh, transforming it to the world space and render the deformed mesh into an image $I$ and a mask $M$ through differentiable rasterization [44] according to the camera parameters $\\mu$ .", + "bbox": [ + 76, + 556, + 468, + 782 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Loss Function and Training. Next, we can construct the RGB loss and the silhouette loss to train the guidance model:", + "bbox": [ + 76, + 782, + 468, + 825 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {R G B} = \\left| \\left| I _ {r, g, b} - I _ {g t} \\right| \\right| _ {1}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 186, + 827, + 468, + 844 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s i l} = I O U (M, M _ {g t}), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 848, + 467, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with $I_{gt}$ and $M_{gt}$ denote the ground truth RGB image and mask, respectively. $IOU(\\cdot)$ denotes Intersection over", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Union metrics. Note that only the first three channels $R, G, B$ of the 32-channel image $I$ are supervised by the ground truth RGB images.", + "bbox": [ + 500, + 90, + 890, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We also use the estimated 3D facial landmarks $P_{gt}$ as described in Sec. 3 to provide rough guidance for the expression deformation MLP. Specifically, we input the neutral 3D landmarks $P_0$ into the expression deformation MLP to predict the expression conditioned landmarks $P$ :", + "bbox": [ + 498, + 136, + 890, + 210 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP = \\boldsymbol {P} _ {0} + \\boldsymbol {f} _ {\\text {d e f}} ^ {\\exp} (\\boldsymbol {P} _ {0}, \\theta). \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 222, + 890, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then we construct the loss function with 3D facial landmarks $P_{gt}$ as the supervision:", + "bbox": [ + 498, + 250, + 890, + 280 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d e f}} = \\left| \\left| P - P _ {g t} \\right| \\right| _ {2}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 291, + 890, + 308 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Besides, we introduce three constraints: (1) a regular term $\\mathcal{L}_{offset}$ to punish all non-zero displacements to prevent the two deformation MLPs from learning a global constant offset [63], (2) a regular term $\\mathcal{L}_{lmk}$ to limit the SDF value at the 3D landmarks to be close to zero, such that the landmarks are located on the surface of the mesh, (3) a Laplacian term $\\mathcal{L}_{lap}$ for maintaining the extracted mesh smooth to a certain extent. Overall, the total loss function is formulated as:", + "bbox": [ + 498, + 318, + 890, + 452 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} = \\mathcal {L} _ {R G B} + \\lambda_ {s i l} \\mathcal {L} _ {s i l} + \\lambda_ {d e f} \\mathcal {L} _ {d e f} + \\tag {14} \\\\ \\lambda_ {o f f s e t} \\mathcal {L} _ {o f f s e t} + \\lambda_ {l m k} \\mathcal {L} _ {l m k} + \\lambda_ {l a p} \\mathcal {L} _ {l a p}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 462, + 890, + 498 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with $\\lambda$ denoting the weights of each term, which are set as follows: $\\lambda_{sil} = 0.1$ , $\\lambda_{def} = 1$ , $\\lambda_{offset} = 0.01$ , $\\lambda_{lmk} = 0.1$ and $\\lambda_{lap} = 100$ . We jointly optimize the MLPs mentioned above with the neutral 3D landmarks $P_0$ jointly until all MLPs are converged.", + "bbox": [ + 498, + 507, + 890, + 582 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Parameters Transfer. Finally, we use the roughly trained mesh guidance model to initialize the Gaussian model. Specifically, we extract the neutral mesh with vertices $X$ and per-vertex features $F$ through DMTet, and directly assign their values to the neutral positions $\\mathbf{X}_0 = X$ and the per-vertex feature vectors $\\pmb{F}_{0} = F$ of the neutral Gaussians respectively. Then, we retain all the four optimized MLPs: $\\{\\pmb{f}_{col}^{exp},\\pmb{f}_{col}^{pose},\\pmb{f}_{def}^{exp},\\pmb{f}_{def}^{pose}\\}$ for the Gaussian model. For the other neutral attributes: rotation, scale and opacity, we adopt the original initialization strategy in Gaussian Splatting [26]. And the parameters of the two attribute MLPs: $\\{\\pmb{f}_{att}^{exp},\\pmb{f}_{att}^{pose}\\}$ and the super resolution network $\\Psi$ are just randomly initialized.", + "bbox": [ + 498, + 583, + 890, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 792, + 632, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 816, + 715, + 832 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the experiment we use 12 sets of data, 10 of which are from NeRSemble [28], and the other 2 are multi-view video data from HAvatar [71]. For the 10 identities from NeRSemble, each set contains 2500 to 3000 frames, 16", + "bbox": [ + 498, + 839, + 890, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1935", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3fd16cc2aa69a95b28977d998402bb72b458d96f03b5396d6aa3fb0b88048e03.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 76, + 88, + 236, + 276 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d39d1453b17767fac298dc00c9fad722b9368a291dac9f348ede15a1e7f2152b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 241, + 88, + 400, + 275 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7e5bd01e6ccb0454dd029612d2095b570267073250ef5ee364e2fe47f4247048.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 405, + 88, + 563, + 275 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5d0caf3d13100366507cdd69ab838a29ee14101dcdd5b7434cc6576462b47711.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 570, + 88, + 728, + 275 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4914b8a2be00e894b7950b90646037287f0cc87996450f4c7795723bc8bdbd91.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 735, + 89, + 893, + 275 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/68181e56aaf028fcb2eb3d7412505588663d0aef29deda2ec62fd1d9bb1ca8f1.jpg", + "image_caption": [ + "NeRFBlendShape" + ], + "image_footnote": [], + "bbox": [ + 78, + 280, + 235, + 467 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9031d2608924608e6f8a59c32319ffa9b920451b6a89aa6c8b80b83bbe80c704.jpg", + "image_caption": [ + "NeRFace" + ], + "image_footnote": [], + "bbox": [ + 243, + 281, + 400, + 467 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/217d97f9e2a3cd03421d2449f28c231699851035f5b7b16b193efb76a9adcfed.jpg", + "image_caption": [ + "HAvatar" + ], + "image_footnote": [], + "bbox": [ + 406, + 281, + 563, + 467 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/275d43f84f8d006fb2ae6bdcd7430565282943ed3ee96c0a253a7dc20035ac9e.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 571, + 281, + 727, + 467 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/23edde6edb66fc22bb344622df640093796f068abfda61707cac7b45056c0c41.jpg", + "image_caption": [ + "GT", + "Figure 3. Qualitative comparisons of different methods on self reenactment task. From left to right: NeRFBlendShape [17], NeRFace [16], HAvatar [71] and Ours. Our method can reconstruct details like beards, teeth, etc. with high quality." + ], + "image_footnote": [], + "bbox": [ + 735, + 277, + 892, + 468 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "cameras are distributed about 120 degrees in front, and simultaneously capture 2K resolution video. For each identity, We use the sequences marked with \"FREE\" as the evaluation data, and the rest as the training data. For the 2 identities from HAvatar, each set contains 3000 frames, 8 cameras are distributed about 120 degrees in front, and 4K resolution videos are collected simultaneously. Later, we crop the face area and resize to 2K resolution.", + "bbox": [ + 75, + 523, + 468, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For data preprocessing, we first remove the background [35] and extract 68 2D facial landmarks [10] for all the images. Then, for each frame, we use multi-view images to estimate the corresponding 3D landmarks, expression coefficients and head pose by fitting the Basel Face Model (BFM) [18] to the extracted 2D landmarks. Note that we define the 3D landmarks as the usual 68 landmarks with vertices indexed as multiples of 100 in the BFM vertices.", + "bbox": [ + 75, + 647, + 470, + 768 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Results and Comparisons", + "text_level": 1, + "bbox": [ + 76, + 784, + 308, + 800 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Self Reenactment. In this section, we first compare our method with existing SOTA methods in qualitative experiments on self reenactment task. Specifically, NeRFace [16] uses a deep MLP to fit an expression conditioned dynamic NeRF. The current SOTA method HAvatar [71] introduces 3DMM template prior and uses a deep convolutional net", + "bbox": [ + 75, + 809, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "work to generate a human head NeRF represented by three planes from a mesh template with expression. Note, HAvatar leverages the GAN framework using the adversarial loss function to force the network to generate details that are not view-consistent. For a fair comparison, we remove this part and use VGG perceptual loss as Sec. 4.2 instead.", + "bbox": [ + 496, + 523, + 892, + 614 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative results on self reenactment task are shown in the Fig. 3. Our method can accurately reconstruct pixel-level high-frequency details such as beards, teeth, and hair. Besides, our method can achieve expression transfer more accurately, such as eye movements in the figure.", + "bbox": [ + 496, + 614, + 892, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Next, we conduct a quantitative evaluation for the four methods on 5 identities and 6 cameras using the evaluation split. The evaluation metrics include: Peak Signal-to-Noise Ratio (PSNR), Structure Similarity Index (SSIM), Learned Perceptual Image Patch Similarity (LPIPS) [70] and Fréchet Inception Distance (FID) [22]. Note, we calculate FID by comparing the distribution of all the training images and all the rendered images. As the task mainly focuses on the reconstruction of the head, we use face parsing to remove the body parts in the image to eliminate their impact in the experiment. As shown in Tab. 1, our method demonstrates a slight improvement in PSNR and SSIM compared with", + "bbox": [ + 496, + 691, + 893, + 875 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "$^{1}$ https://github.com/zllrunning/faceParsing.PyTorch", + "bbox": [ + 517, + 886, + 795, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1936", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/94e0c8c1d19751671903d66d027f4d7b836d3515e6ecf40ac51c6d7e30256eaa.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPSNR ↑SSIM ↑LPIPS (512) ↓LPIPS (2K) ↓FID (2K) ↓
NeRFBlendShape25.910.8360.1230.22954.80
NeRFace27.140.8490.1470.23465.11
HAVatar27.190.8830.0640.20931.06
Ours (w/o SR)27.820.8870.0800.20245.50
Ours27.700.8830.0560.09818.50
", + "bbox": [ + 114, + 88, + 851, + 185 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4f216f90bc2dd0933614856803541ac8434a5d55705005552af1bd5e5231b262.jpg", + "image_caption": [ + "Figure 4. Qualitative comparisons of different methods on cross-identity reenactment task. From left to right: NeRFBlendShape [17], NeRFace [16], HAvatar [71] and Ours. Our method synthesizes high-fidelity images while ensuring the accuracy of expression transfer." + ], + "image_footnote": [], + "bbox": [ + 76, + 227, + 890, + 503 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4bb0f82960493835c93524465f4158c84387f872cf985e4f9f417f54e7ce4430.jpg", + "table_caption": [ + "Table 1. Quantitative evaluation results of NeRFBlendShape [17], NeRFace [16], HAvatar [71], our method without super resolution and our full method on self reenactment task." + ], + "table_footnote": [], + "table_body": "
MethodPSNR ↑SSIM ↑LPIPS ↓
NeRFBlendShape25.430.8120.148
NeRFace26.650.8250.151
HAVatar27.130.8800.65
Ours27.580.8820.059
", + "bbox": [ + 91, + 558, + 450, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "previous methods, and a significant improvement in LPIPS and FID, which means that our method can generate more high-frequency details.", + "bbox": [ + 75, + 686, + 468, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Cross-Identity Reenactment. We also qualitatively compare our method with the above SOTA methods on cross-identity reenactment task. As shown in the results, our method is able to synthesize higher-fidelity images with more accurate expression transfer and richer emotions.", + "bbox": [ + 75, + 734, + 468, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Novel View Synthesis. In this section, we show the results of novel view synthesis as shown in Fig. 5 shown. In this case, we use video data from 8 views for training and render the image at a new viewpoint. Next, we quantitatively evaluate the 3D consistency of our method and compare it with other SOTA methods mentioned above. Specif", + "bbox": [ + 75, + 810, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ically, we select the 5 identities as above and use the video data from 8 cameras for training while the other 8 holdout cameras for evaluation. The evaluation metrics include PSNR, SSIM and LPIPS (512 resolution). The results are shown in Tab. 2. Our method outperforms other methods in 3D consistency.", + "bbox": [ + 496, + 555, + 892, + 647 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 657, + 653, + 674 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2ceb035ef74c945f55d4aa0174bbac279c66e67a394b98325049f7fcf839f205.jpg", + "table_caption": [ + "Table 2. Quantitative evaluation results of the other SOTA methods and our method on 3D consistency." + ], + "table_footnote": [], + "table_body": "
MethodPSNR ↑SSIM ↑LPIPS ↓
FLAME-Init28.730.8750.123
Mesh-Deform28.830.8740.116
Ours28.940.8760.108
", + "bbox": [ + 527, + 680, + 861, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Quantitative evaluation results of the other two ablation baselines and ours on self reenactment task.", + "bbox": [ + 496, + 752, + 890, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation on Initialization Strategies. In order to verify the effectiveness of our geometry-guided initialization strategy 4.3, we compare it with the strategy to use the FLAME model for initialization (FLAME-Init). Specifically, after fitting a FLAME model through multi-view data, we first subdivide the FLAME mesh 4 times and use the neutral vertices as the positions of the neutral Gaussians. Then, the", + "bbox": [ + 496, + 794, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1937", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2047bfc484c3da4a7cbc4535dea093d79bfb21ca553d99480cca68bdf49ceab6.jpg", + "image_caption": [ + "Figure 5. Novel view synthesis results of our method. We use 8-view synchronized videos for training the avatar." + ], + "image_footnote": [], + "bbox": [ + 76, + 88, + 893, + 212 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4c63f457e03a24d4ef3f2400d188ade0f7fa21b59cd4d716638e1f98f19bb6f2.jpg", + "image_caption": [ + "Figure 6. Ablation study on the initialization strategies: FLAME-initialization and our geometry-guided initialization. Our strategy ensures the hair strands away from the head are well reconstructed." + ], + "image_footnote": [], + "bbox": [ + 76, + 243, + 472, + 477 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "expression deformation MLP is optimized to learn the displacement of FLAME vertices. We set the per-vertex feature to zeros, while randomly initialize the parameters of the expression color MLP. The initialization of other variables is the same as our strategy. Qualitative results are shown in the Fig. 6. Due to the lack of initialization for the hair and shoulders in FLAME-Init, the points to model these parts are offset from nearby vertices, which leads to sparseness of the Gaussians, resulting in blurring.", + "bbox": [ + 75, + 546, + 468, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation on Deformation Modeling Approaches. We compare our fully learned deformation field with the previous mesh-based deformation (Mesh-Deform). Specifically, we migrate the method in INSTA [76] for controlling the NeRF deformation to our Gaussians. First we fit a 3DMM mesh template. Then, for each Gaussian point, find the closest face on the mesh, and calculate the deformation gradient to estimate the displacement. Qualitative results are shown in the Fig. 7. For some expressions that cannot be captured well by the 3DMM mesh template, our method can learn accurate deformation, thereby achieving the modeling of complex expressions. Quantitative results are shown in the Fig. 3. Our method outperforms both the two ablation baselines on PSNR, SSIM and LPIPS metrics.", + "bbox": [ + 75, + 689, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0f815ad2a74e0582c53a4a406698b5868da703d611678876dd1ca739d76d5c22.jpg", + "image_caption": [ + "Figure 7. Ablation study on the deformation modeling: mesh LBS-based deformation and our fully learned deformation. Our approach can learn complex and exaggerated expressions." + ], + "image_footnote": [], + "bbox": [ + 500, + 244, + 893, + 476 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Discussion and Conclusion", + "text_level": 1, + "bbox": [ + 500, + 539, + 746, + 554 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ethical Considerations. Our method is capable of creating artificial portrait videos, which have the potential to disseminate misinformation, influence public perceptions, and erode confidence in media sources.", + "bbox": [ + 498, + 569, + 890, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitation. For the tongue and teeth inside the mouth or long hair, blurring is sometimes produced in our method due to the lack of tracking methods.", + "bbox": [ + 498, + 633, + 890, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Conclusion. In this paper, we propose Gaussian Head Avatar, a novel representation for head avatar reconstruction, which leverages dynamic 3D Gaussians controlled by a fully learned expression deformation. Experiments demonstrate our method can synthesize ultra high-fidelity images while modeling exaggerated expressions. In addition, we propose a well-designed minute-level initialization strategy to ensure the training convergence. We believe our Gaussian Head Avatar will become the mainstream direction for head avatar reconstruction in the future.", + "bbox": [ + 496, + 684, + 890, + 834 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This paper is supported by National Key R&D Program of China (2022YFF0902200), the NSFC project No.62125107, the Beijing Municipal Science & Technology Z231100005923030.", + "bbox": [ + 498, + 839, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1938", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In European Conference on Computer Vision, pages 696-712, 2020. 3", + "[2] ShahRukh Athar, Zexiang Xu, Kalyan Sunkavalli, Eli Shechtman, and Zhixin Shu. Rignerf: Fully controllable neural 3d portraits. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3", + "[3] ShahRukh Athar, Zhixin Shu, and Dimitris Samaras. Flame-in-nerf: Neural control of radiance fields for free view face animation. In IEEE 17th International Conference on Automatic Face and Gesture Recognition (FG), pages 1–8, 2023. 3", + "[4] Thabo Beeler, Bernd Bickel, Paul Beardsley, Bob Sumner, and Markus Gross. High-quality single-shot capture of facial geometry. ACM Trans. Graph., 29(4), 2010. 2", + "[5] Sai Bi, Stephen Lombardi, Shunsuke Saito, Tomas Simon, Shih-En Wei, Kevyn Mcphail, Ravi Ramamoorthi, Yaser Sheikh, and Jason Saragih. Deep reeligible appearance models for animatable faces. ACM Trans. Graph., 40(4), 2021. 2", + "[6] V Blanz and T Vetter. A morphable model for the synthesis of 3d faces. In 26th Annual Conference on Computer Graphics and Interactive Techniques (SIGGRAPH 1999), pages 187-194. ACM Press, 1999. 2", + "[7] Timo Bolkart, Tianye Li, and Michael J. Black. Instant multi-view head capture through learnable registration. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 768-779, 2023. 2", + "[8] Derek Bradley, Wolfgang Heidrich, Tiberiu Popa, and Alla Sheffer. High resolution passive facial performance capture. 29(4), 2010. 2", + "[9] Alan Brunton, Timo Bolkart, and Stefanie Wuhrer. Multilinear wavelets: A statistical shape space for human faces. In Proceedings of the Proceedings of the European Conference on Computer Vision (ECCV), 2014. 2", + "[10] Adrian Bulat and Georgios Tzimiropoulos. How far are we from solving the 2d & 3d face alignment problem? (and a dataset of 230,000 3d facial landmarks). In International Conference on Computer Vision, 2017. 6", + "[11] Chen Cao, Yanlin Weng, Shun Zhou, Y. Tong, and Kun Zhou. Facewarehouse: A 3d facial expression database for visual computing. In IEEE Transactions on Visualization and Computer Graphics, pages 413-425, 2014. 2", + "[12] Chen Cao, Derek Bradley, Kun Zhou, and Thabo Beeler. Real-time high-fidelity facial performance capture. ACM Trans. Graph., 34(4), 2015. 3", + "[13] Chen Cao, Hongzhi Wu, Yanlin Weng, Tianjia Shao, and Kun Zhou. Real-time facial animation with image-based dynamic avatars. ACM Trans. Graph., 35(4), 2016. 3", + "[14] Chen Cao, Tomas Simon, Jin Kyu Kim, Gabe Schwartz, Michael Zollhoefer, Shun-Suke Saito, Stephen Lombardi, Shih-En Wei, Danielle Belko, Shouu-I Yu, Yaser Sheikh, and Jason Saragih. Authentic volumetric avatars from a phone scan. ACM Trans. Graph., 41(4), 2022. 3" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2019. 3", + "[16] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Niessner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8645-8654, 2021. 3, 4, 6, 7", + "[17] Xuan Gao, Chenglai Zhong, Jun Xiang, Yang Hong, Yudong Guo, and Juyong Zhang. Reconstructing personalized semantic facial nerf models from monocular video. ACM Transactions on Graphics (Proceedings of SIGGRAPH Asia), 41(6), 2022. 3, 6, 7", + "[18] Thomas Gerig, Andreas Forster, Clemens Blumer, Bernhard Egger, Marcel Lüthi, Sandro Schonborn, and Thomas Vetter. Morphable face models - an open framework. pages 75-82, 2017. 3, 6", + "[19] Abhijeet Ghosh, Graham Fyffe, Borom Tunwattanapong, Jay Busch, Xueming Yu, and Paul Debevec. Multiview face capture using polarized spherical gradient illumination. ACM Trans. Graph., 30(6):1-10, 2011. 2", + "[20] Philip-William Grassal, Malte Prinzler, Titus Leistner, Carsten Rother, Matthias Nießner, and Justus Thies. Neural head avatars from monocular rgb videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18632–18643, 2022. 3", + "[21] Yudong Guo, Keyu Chen, Sen Liang, Yong-Jin Liu, Hujun Bao, and Juyong Zhang. Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5764–5774, 2021. 3", + "[22] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Neural Information Processing Systems, 2017. 6", + "[23] Yang Hong, Bo Peng, Haiyao Xiao, Ligang Liu, and Juyong Zhang. Headnerf: A real-time nerf-based parametric head model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20374-20384, 2022. 3", + "[24] Liwen Hu, Shunsuke Saito, Lingyu Wei, Koki Nagano, Jaewoo Seo, Jens Fursund, Iman Sadeghi, Carrie Sun, YenChun Chen, and Hao Li. Avatar digitization from a single image for real-time rendering. ACM Trans. Graph., 36(6), 2017. 3", + "[25] Alexandru Eugen Ichim, Sofien Bouaziz, and Mark Pauly. Dynamic 3d avatar creation from hand-held video input. ACM Trans. Graph., 34(4), 2015. 3", + "[26] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 2, 3, 5", + "[27] Taras Khakhulin, Vanessa Sklyarova, Victor Lempitsky, and Egor Zakharov. Realistic one-shot mesh-based head avatars. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 3" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1939", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Tobias Kirschstein, Shenhan Qian, Simon Giebenhain, Tim Walter, and Matthias Nießner. Nersemble: Multi-view radiance field reconstruction of human heads. ACM Trans. Graph., 42(4), 2023. 2, 5", + "[29] Georgios Kopanas, Thomas Leimkuhler, Gilles Rainer, Clément Jambon, and George Drettakis. Neural point cata-caustics for novel-view synthesis of reflections. ACM Transactions on Graphics (TOG), 41(6):1-15, 2022. 3", + "[30] Christoph Lassner and Michael Zollhofer. Pulsar: Efficient sphere-based neural rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1440-1449, 2021. 3", + "[31] Marc Levoy, Kari Pulli, Brian Curless, Szymon Rusinkiewicz, David Koller, Lucas Pereira, Matt Ginzon, Sean Anderson, James Davis, Jeremy Ginsberg, Jonathan Shade, and Duane Fulk. The digital michelangelo project: 3d scanning of large statues. In Proceedings of the 27th Annual Conference on Computer Graphics and Interactive Techniques, page 131-144, USA, 2000. ACM Press/Addison-Wesley Publishing Co. 2", + "[32] Tianye Li, Timo Bolkart, Michael J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6), 2017. 2, 3", + "[33] Tianye Li, Shichen Liu, Timo Bolkart, Jiayi Liu, Hao Li, and Yajie Zhao. Topologically consistent multi-view face inference using volumetric sampling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3824-3834, 2021. 2", + "[34] Zhe Li, Zerong Zheng, Lizhen Wang, and Yebin Liu. Animatable gaussians: Learning pose-dependent gaussian maps for high-fidelity human avatar modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3", + "[35] Shanchuan Lin, Andrey Ryabtsev, Soumyadip Sengupta, Brian Curless, Steve Seitz, and Ira Kemelmacher-Shlizerman. Real-time high-resolution background matting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3, 6", + "[36] Xian Liu, Yinghao Xu, Qianyi Wu, Hang Zhou, Wayne Wu, and Bolei Zhou. Semantic-aware implicit neural auditory-driven video portrait generation. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 3", + "[37] Stephen Lombardi, Jason Saragih, Tomas Simon, and Yaser Sheikh. Deep appearance models for face rendering. ACM Trans. Graph., 37(4):68:1-68:13, 2018. 1, 2", + "[38] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural volumes: Learning dynamic renderable volumes from images. ACM Trans. Graph., 38(4):65:1-65:14, 2019. 1, 2", + "[39] Stephen Lombardi, Tomas Simon, Gabriel Schwartz, Michael Zollhoefer, Yaser Sheikh, and Jason Saragih. Mixture of volumetric primitives for efficient neural rendering. ACM Trans. Graph., 40(4), 2021. 1, 2", + "[40] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis, 2023. 2, 3" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Shugao Ma, Tomas Simon, Jason Saragih, Dawei Wang, Yuecheng Li, Fernando De La Torre, and Yaser Sheikh. Pixel codec avatars. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 64-73, 2021. 1, 2", + "[42] Marko Mihajlovic, Aayush Bansal, Michael Zollhoefer, Siyu Tang, and Shunsuke Saito. Keypointnerf: Generalizing image-based volumetric avatars using relative spatial encoding of keypoints. In European conference on computer vision, 2022. 1, 2", + "[43] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 1, 2", + "[44] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting Triangular 3D Models, Materials, and Lighting From Images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8280-8290, 2022. 5", + "[45] Koki Nagano, Jaewoo Seo, Jun Xing, Lingyu Wei, Zimo Li, Shunsuke Saito, Aviral Agarwal, Jens Fursund, and Hao Li. Pagan: Real-time avatars using dynamic textures. ACM Trans. Graph., 37(6), 2018. 3", + "[46] Pascal Paysan, Reinhard Knothe, Brian Amberg, Sami Romdhani, and Thomas Vetter. A 3d face model for pose and illumination invariant face recognition. In 2009 Sixth IEEE International Conference on Advanced Video and Signal Based Surveillance, pages 296-301, 2009. 2", + "[47] Minghan Qin, Yifan Liu, Yuelang Xu, Xiaochen Zhao, Yebin Liu, and Haoqian Wang. High-fidelity 3d head avatars reconstruction through spatially-varying expression conditioned neural radiance field. In AAAI Conference on Artificial Intelligence, 2023. 3", + "[48] Amit Raj, Michael Zollhoefer, Tomas Simon, Jason Saragih, Shunsuke Saito, James Hays, and Stephen Lombardi. Pva: Pixel-aligned volumetric avatars. In arXiv:2101.02697, 2020. 1, 2", + "[49] Darius Rückert, Linus Franke, and Marc Stamminger. Adop: Approximate differentiable one-pixel point rendering. ACM Trans. Graph., 41(4), 2022. 3, 4", + "[50] Ruizhi Shao, Jingxiang Sun, Cheng Peng, Zerong Zheng, Boyao Zhou, Hongwen Zhang, and Yebin Liu. Control4d: Efficient 4d portrait editing with text. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 3", + "[51] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 5", + "[52] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2", + "[53] Jingxiang Sun, Xuan Wang, Lizhen Wang, Xiaoyu Li, Yong Zhang, Hongwen Zhang, and Yebin Liu. Next3d: Generative neural texture rasterization for 3d-aware head avatars." + ], + "bbox": [ + 501, + 92, + 892, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1940", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3", + "[54] Cong Wang, Di Kang, Yanpei Cao, Linchao Bao, Ying Shan, and Song-Hai Zhang. Neural point-based volumetric avatar: Surface-guided neural points for efficient and photorealistic volumetric head avatar. In ACM SIGGRAPH Asia 2023 Conference Proceedings, 2023. 1", + "[55] Daoye Wang, Prashanth Chandran, Gaspard Zoss, Derek Bradley, and Paulo Gotardo. Morf: Morphable radiance fields for multiview neural head modeling. In ACM SIGGRAPH 2022 Conference Proceedings, New York, NY, USA, 2022. Association for Computing Machinery. 3", + "[56] Lizhen Wang, Zhiyuan Chen, Tao Yu, Chenguang Ma, Liang Li, and Yebin Liu. Faceverse: a fine-grained and detail-controllable 3d face morphable model from a hybrid dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[57] Ziyan Wang, Timur Bagautdinov, Stephen Lombardi, Tomas Simon, Jason Saragih, Jessica Hodgins, and Michael Zollhofer. Learning compositional radiance fields of dynamic human heads. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5704-5713, 2021. 1, 2", + "[58] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. SynSin: End-to-end view synthesis from a single image. In CVPR, 2020. 3", + "[59] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering, 2023. 2, 3", + "[60] Sijing Wu, Yichao Yan, Yunhao Li, Yuhao Cheng, Wenhan Zhu, Ke Gao, Xiaobo Li, and Guangtao Zhai. Ganhead: Towards generative animatable neural head avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 437-447, 2023. 2", + "[61] Yunze Xiao, Hao Zhu, Haotian Yang, Zhengyu Diao, Xiangju Lu, and Xun Cao. Detailed facial geometry recovery from multi-view images by learning an implicit function. In Proceedings of the AAAI Conference on Artificial Intelligence, 2022. 2", + "[62] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5438–5448, 2022. 3", + "[63] Yuelang Xu, Lizhen Wang, Xiaochen Zhao, Hongwen Zhang, and Yebin Liu. Avatarmav: Fast 3d head avatar reconstruction using motion-aware neural voxels. In ACM SIGGRAPH 2023 Conference Proceedings, 2023. 3, 5", + "[64] Yuelang Xu, Hongwen Zhang, Lizhen Wang, Xiaochen Zhao, Huang Han, Qi Guojun, and Yebin Liu. Latentavatar: Learning latent expression code for expressive neural head avatar. In ACM SIGGRAPH 2023 Conference Proceedings, 2023. 3, 4", + "[65] Kai Yang, Hong Shang, Tianyang Shi, Xinghan Chen, Jingkai Zhou, Zhongqian Sun, and Wei Yang. Asm: Adaptive skinning model for high-quality 3d face modeling supplementary material. 2021. 2" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[66] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction, 2023. 2, 3", + "[67] Zeyu Yang, Hongye Yang, Zijie Pan, Xiatian Zhu, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting, 2023. 2, 3", + "[68] T Yenamandra, A Tewari, F Bernard, HP Seidel, M Elgharib, D Cremers, and C Theobalt. i3dmm: Deep implicit 3d morphable model of human heads. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3", + "[69] Wang Yifan, Felice Serena, Shihao Wu, Cengiz Öztireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM Transactions on Graphics (proceedings of ACM SIGGRAPH ASIA), 38(6), 2019. 3", + "[70] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 5, 6", + "[71] Xiaochen Zhao, Lizhen Wang, Jingxiang Sun, Hongwen Zhang, Jinli Suo, and Yebin Liu. Havatar: High-fidelity head avatar via facial model conditioned neural radiance field. ACM Trans. Graph., 2023. Just Accepted. 1, 2, 5, 6, 7", + "[72] Shunyuan Zheng, Boyao Zhou, Ruizhi Shao, Boning Liu, Shengping Zhang, Liqiang Nie, and Yebin Liu. Gpsgaussian: Generalizable pixel-wise 3d gaussian splatting for real-time human novel view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3", + "[73] Yufeng Zheng, Victoria Fernández Abrevaya, Marcel C. Bühler, Xu Chen, Michael J. Black, and Otmar Hilliges. I m avatar: Implicit morphable head avatars from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13535-13545, 2022. 2, 3", + "[74] Yufeng Zheng, Wang Yifan, Gordon Wetzstein, Michael J. Black, and Otmar Hilliges. Pointavatar: Deformable point-based head avatars from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3", + "[75] Yiyu Zhuang, Hao Zhu, Xusen Sun, and Xun Cao. Mofanerf: Morphable facial neural radiance field. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 3", + "[76] Wojciech Zielonka, Timo Bolkart, and Justus Thies. Instant volumetric head avatars, 2022. 2, 3, 8" + ], + "bbox": [ + 501, + 92, + 892, + 753 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1941", + "bbox": [ + 483, + 944, + 513, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Gaussian Head Avatar_ Ultra High-fidelity Head Avatar via Dynamic Gaussians/24a06516-83b5-41f4-8bb5-a4da984c01e1_model.json b/2024/Gaussian Head Avatar_ Ultra High-fidelity Head Avatar via Dynamic Gaussians/24a06516-83b5-41f4-8bb5-a4da984c01e1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..523a04ffcfb7d76cbdff4a9e020349ddaea756eb --- /dev/null +++ b/2024/Gaussian Head Avatar_ Ultra High-fidelity Head Avatar via Dynamic Gaussians/24a06516-83b5-41f4-8bb5-a4da984c01e1_model.json @@ -0,0 +1,2565 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.083, + 0.131, + 0.888, + 0.153 + ], + "angle": 0, + "content": "Gaussian Head Avatar: Ultra High-fidelity Head Avatar via Dynamic Gaussians" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.18, + 0.916, + 0.219 + ], + "angle": 0, + "content": "Yuelang Xu\\(^{1}\\), Bengwang Chen\\(^{1}\\), Zhe Li\\(^{1}\\), Hongwen Zhang\\(^{2}\\), Lizhen Wang\\(^{1}\\), Zerong Zheng\\(^{3}\\), Yebin Liu\\(^{1\\dagger}\\) \n\\(^{1}\\) Department of Automation, Tsinghua University \\(^{2}\\) Beijing Normal University \\(^{3}\\) NNKosmos Technology" + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.226, + 0.895, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.077, + 0.377, + 0.891, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.531, + 0.893, + 0.575 + ], + "angle": 0, + "content": "Figure 1. Gaussian head avatar achieves ultra high-fidelity image synthesis with controllable expressions at 2K resolution. The above shows different views of the synthesized avatar, and the bottom shows different identities animated by the same expression. 16 views are used during the training." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.587, + 0.314, + 0.601 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.604, + 0.474, + 0.876 + ], + "angle": 0, + "content": "Creating high-fidelity 3D head avatars has always been a research hotspot, but there remains a great challenge under lightweight sparse view setups. In this paper, we propose Gaussian Head Avatar represented by controllable 3D Gaussians for high-fidelity head avatar modeling. We optimize the neutral 3D Gaussians and a fully learned MLP-based deformation field to capture complex expressions. The two parts benefit each other, thereby our method can model fine-grained dynamic details while ensuring expression accuracy. Furthermore, we devise a well-designed geometry-guided initialization strategy based on implicit SDF and Deep Marching Tetrahedra for the stability and convergence of the training procedure. Experiments show our approach outperforms other state-of-the-art sparse-view methods, achieving ultra high-fidelity rendering quality at \\(2K\\) resolution even under exaggerated expressions. Project page: https://yuelangx.github.io/gaussianheadavatar." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.587, + 0.631, + 0.601 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.892, + 0.749 + ], + "angle": 0, + "content": "High-fidelity 3D human head avatar modeling is of great significance in many fields, such as VR/AR, telepresence, digital human and film production. Automatically creating high-fidelity avatars has been a research hotspot in computer vision for decades. Although some traditional head avatars [37, 39, 41, 54] can realize high-fidelity animation, they typically require accurate geometries reconstructed and tracked from dense multi-view videos, thus limiting their applications in lightweight settings." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Thanks to the Neural Radiance Fields (NeRF) [43] which show great capability of novel view synthesis in the absence of accurate geometry, recent methods [38, 57] skip the geometry reconstruction and tracking steps but directly learn high-quality NeRF-based head avatars. Other works [42, 48, 71] have verified that NeRF can be applied to either dense or sparse views, which greatly lowers the threshold for head avatar reconstruction. However, it still remains challenging for these NeRF-based approaches to synthesize high-fidelity images at 2K resolutions with pixel" + }, + { + "type": "page_footnote", + "bbox": [ + 0.102, + 0.888, + 0.229, + 0.9 + ], + "angle": 0, + "content": "†Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1931" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.355, + 0.106 + ], + "angle": 0, + "content": "level details, including wrinkles and eyes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.11, + 0.471, + 0.395 + ], + "angle": 0, + "content": "To overcome this bottleneck and further improve the avatar quality, we introduce 3D Gaussian splatting [26] for 3D head avatar modeling. This is an explicit discrete representation that can be well adapted to the rasterization-based rendering pipeline. It has been verified that the 3D Gaussian representation is capable of rendering complex scenes with low computational consumption. Compared to NeRF, the reconstruction quality of static and dynamic scenes [40, 59, 67] is much better while rendering time cost has been significantly reduced. Motivated by this progress, we propose Gaussian Head Avatar, a novel representation that utilizes 3D Gaussian splatting for ultra high-fidelity head avatar modeling. Although recent 4D Gaussian works [40, 59, 66, 67] have been proposed to reconstruct dynamic scenes, all of them cannot be animated. For modeling the animatable head avatar, it is crucial but still unexplored how to effectively control the deformation of 3D Gaussians and model the dynamic appearances through expression coefficients." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.399, + 0.471, + 0.732 + ], + "angle": 0, + "content": "Previous explicit [74] and implicit [2, 73, 76] head avatars usually formulate the facial deformation via linear blend skinning (LBS) using the skinning weights and blendshapes like the FLAME model [32]. However, such an LBS-based formulation fails to represent exaggerated and fine-grained expressions by simple linear operations, limiting the representation ability of the head avatars. Inspired by NeRSemble [28], we propose a fully learnable expression-conditioned deformation field for the 3D head Gaussians, avoiding the limited capability of the LBS-based formulation. Specifically, we input the positions of the 3D Gaussians with expression coefficients into an MLP to directly predict the displacements from the neutral expression to the target one. Similarly, we control the motion of nonface areas, such as the neck, using the head pose as the condition. 3D Gaussian-based representation has the powerful ability to reconstruct high-frequency details, enabling our method to learn accurate deformation fields. In turn, the learned accurate deformation field facilitates the dynamic Gaussian head model to fit more dynamic details. As a result, our method is able to reconstruct finer-grained dynamic details of expressive human heads." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Unfortunately, as a discretized representation, the gradients back-propagated to the 3D Gaussians cannot spread through the whole space. Thus the convergence of training heavily relies on a plausible initialization for both the geometry and the deformation field. However, simply initializing the 3D head Gaussians with a morphable template like FLAME [32] fails to model the long hairstyle and the shoulders. Hence, we further propose an efficient and well-designed geometry-guided initialization strategy. Specifically, instead of starting from stochastic Gaussians or a FLAME model, we initially optimize an implicit signed dis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.287 + ], + "angle": 0, + "content": "tance function (SDF) field along with a color field and a deformation MLP for modeling the basic geometry, color, and the expression-conditioned deformations of the head avatar respectively. The SDF field is converted to a mesh through Deep Marching Tetrahedra (DMTet) [52], with the color and deformation of the vertices predicted by the MLPs. Then we render the mesh and optimize them jointly under the supervision of multi-view RGB images. Finally, we use the mesh with per-vertex features from the SDF field to initialize the 3D Gaussians to lie on the basic head surface while the color and deformation MLPs are carried over to the next stage, ensuring stable training for convergence. The entire initialization process takes only around 10 minutes." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.289, + 0.888, + 0.302 + ], + "angle": 0, + "content": "The contributions of our method can be summarized as:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.304, + 0.891, + 0.363 + ], + "angle": 0, + "content": "- We propose Gaussian Head Avatar, a new head avatar representation that employs controllable dynamic 3D Gaussians to model expressive human head avatars, producing ultra high-fidelity synthesized images at 2K resolutions." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.364, + 0.891, + 0.424 + ], + "angle": 0, + "content": "- For modeling high-frequency dynamic details, we employ a fully learned deformation field upon the 3D head Gaussians, which accurately model extremely complex and exaggerated facial expressions." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.424, + 0.891, + 0.484 + ], + "angle": 0, + "content": "- We carefully design an efficient initialization strategy that leverages implicit representations to initialize the geometry and deformation, leading to efficient and robust convergence when training the Gaussian Head Avatar." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.304, + 0.891, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.485, + 0.891, + 0.53 + ], + "angle": 0, + "content": "Benefiting from these contributions, our method surpasses recent state-of-the-art methods under lightweight sparse-view setups on the avatar quality by a large margin." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.544, + 0.651, + 0.559 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.568, + 0.892, + 0.87 + ], + "angle": 0, + "content": "3D Head Avatar Reconstruction. Due to the wide application value in the film and digital human industry, 3D head avatar reconstruction from multi-view images has always been a research hotspot. Traditional works [4, 8, 19, 31] reconstruct the scan geometry through multi-view stereo and then register a face mesh template to it. However, such methods usually require heavy computation. With the utilization of deep neural networks, current methods [7, 33, 61, 65] achieve very fast reconstruction, producing even more accurate geometry. Lombardi et al. [37], Bi et al. [5] and Ma et al. [41] represent the full head mesh through a deep neural network and train it with multi-view videos as supervision. However, due to the errors in geometric estimation, mesh-based head avatars typically suffer from texture blur. Therefore, some recent methods [38, 57] utilize NeRF representation [43] to synthesize novel view images without geometry reconstruction, or build NeRF on the head mesh template [39]. Furthermore, the NeRF-based methods are extended to sparse view reconstruction tasks [28, 42, 48, 71] and achieve impressive performance." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Methods which focus on generative model [6, 9, 11, 32, 46, 56, 60] are dedicated to learning general mesh" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1932" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.212 + ], + "angle": 0, + "content": "face templates from large-scale multi-view face images or 3D scans. Recently, implicit SDF-based [68] or NeRF-based [14, 23, 53, 55, 75] methods can learn full-head templates without the limitations of fixed topology, thereby better modeling complex hairstyles and glasses. Cao et al. [14] adopts a hybrid representation of local NeRF built on the mesh surface, which enables high-fidelity rendering and flexible expression control." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.213, + 0.471, + 0.363 + ], + "angle": 0, + "content": "3D head avatars reconstruction from monocular videos is also a popular yet challenging research topic. Early methods [12, 13, 15, 24, 25, 45] optimize a morphable mesh to fit the training video. Recent methods [20, 27] leverage neural networks to learn non-rigid deformation upon 3DMM face templates [18, 32], thus can recover more dynamic details. Such methods are not flexible enough to handle complex topologies. Therefore, the latest methods explore to construct head avatar models based on implicit SDF [73], point clouds [74] or NeRF [2, 3, 16, 17, 21, 36, 47, 63, 64, 76]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.364, + 0.47, + 0.605 + ], + "angle": 0, + "content": "Point-based Rendering. Point elements as a discrete and unstructured representation can fit geometry with arbitrary topology [69] efficiently. Recent methods [29, 30, 58] open up a differentiable rasterization pipeline, such that the point-based representation is widely used in multi-view reconstruction tasks. Aliev et al. [1] and Ruckert et al. [49] propose to first render the feature map, which is transferred to the images through a convolutional renderer. Xu et al.[62] use neural point cloud associated with neural features to model a NeRF. Recently, 3D Gaussian splatting [26] shows its superior performance, beating NeRF in both novel view synthesis quality and rendering speed. Some approaches [34, 40, 50, 59, 66, 67, 72, 72] extend Gaussian representation to dynamic scene reconstruction. However, these methods can not be migrated to the head avatar reconstruction tasks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.619, + 0.182, + 0.635 + ], + "angle": 0, + "content": "3. Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.471, + 0.901 + ], + "angle": 0, + "content": "The pipeline of the reconstruction of Gaussian Head Avatar is illustrated in Fig. 2, including the initialization stage and the training stage of Gaussian Head Avatar. Before the beginning of the pipeline, we remove the background [35] of each image and jointly estimate the 3DMM model [18], 3D facial landmarks and the expression coefficients for each frame. In the initialization stage (Sec. 4.3), we reconstruct an SDF-based neutral geometry, and optimize a deformation MLP and a color MLP from the training data as the guidance model. Next, we extract the neutral mesh through DMTet to initialize the neutral Gaussians while the deformation and color MLPs are also inherited from the initialization stage. In the training stage of Gaussian Head Avatar (Sec. 4.2), we deform the neutral Gaussians to the target expression through the dynamic generator given the driving expression coefficients as the condition. Finally, given a camera view, the expressive Gaussians are rendered to a" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.151 + ], + "angle": 0, + "content": "feature map, which is then fed into the convolutional super resolution network to generate high-resolution avatar images. The whole model is optimized under the supervision of multi-view RGB videos." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.169, + 0.593, + 0.185 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.195, + 0.71, + 0.211 + ], + "angle": 0, + "content": "4.1. Avatar Representation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.219, + 0.892, + 0.324 + ], + "angle": 0, + "content": "Generally, the static 3D Gaussians [26] with \\(N\\) points are represented by their positions \\(X\\), the multi-channel color \\(C\\), the rotation \\(Q\\), scale \\(S\\) and opacity \\(A\\). The rotation \\(Q\\) is represented in the form of quaternion. Subsequently, the Gaussians can be rasterized and rendered to the multi-channel image \\(I\\) given the camera parameters \\(\\mu\\). This process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.609, + 0.34, + 0.892, + 0.356 + ], + "angle": 0, + "content": "\\[\nI = \\mathcal {R} (X, C, Q, S, A; \\mu). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.371, + 0.892, + 0.474 + ], + "angle": 0, + "content": "Our task is to reconstruct a dynamic head avatar controlled by expression coefficients. Therefore, we formulate the head avatar as dynamic 3D Gaussians conditioned on expressions. To handle the dynamic changes, we input the expression coefficients with head pose to the head avatar model and output the position and other attributes of the Gaussians as above." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.477, + 0.892, + 0.688 + ], + "angle": 0, + "content": "Specifically, we first construct a canonical neutral Gaussian model with expression-independent attributes: \\(\\{X_0, F_0, Q_0, S_0, A_0\\}\\), which are fully estimizable. \\(X_0 \\in \\mathbb{R}^{N \\times 3}\\) denotes the positions of the Gaussians with a neutral expression in the canonical space. \\(F_0 \\in \\mathbb{R}^{N \\times 128}\\) denotes the point-wise feature vectors as their intrinsic properties. \\(Q_0 \\in \\mathbb{R}^{N \\times 4}\\), \\(S_0 \\in \\mathbb{R}^{N \\times 3}\\) and \\(A_0 \\in \\mathbb{R}^{N \\times 1}\\) denotes the neutral rotation, scale and opacity respectively. Note that we do not define the neutral color, but directly predict expression-dependent dynamic color from the point-wise feature vectors \\(F_0\\). Then, we construct an MLP-based expression conditioned dynamic generator \\(\\Phi\\) to generate all the extra dynamic changes to the neutral model. Overall, the whole Gaussian head avatar can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.703, + 0.892, + 0.72 + ], + "angle": 0, + "content": "\\[\n\\{X, C, Q, S, A \\} = \\Phi \\left(\\boldsymbol {X} _ {0}, \\boldsymbol {F} _ {0}, \\boldsymbol {Q} _ {0}, \\boldsymbol {S} _ {0}, \\boldsymbol {A} _ {0}; \\theta , \\beta\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.732, + 0.892, + 0.808 + ], + "angle": 0, + "content": "with \\(\\theta\\) denoting expression coefficients and \\(\\beta\\) denoting the head pose. During the training, we optimize all the parameters of the dynamic generator \\(\\Phi\\) and the neutral Gaussian model \\(\\{X_0, F_0, Q_0, S_0, A_0\\}\\), which are highlighted in bold in the following." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.855 + ], + "angle": 0, + "content": "Next, we explain the process of adding expression-related changes to the neutral Gaussian model through the dynamic generator \\(\\Phi\\) as described in Eqn. 2 in detail." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Positions \\(X^{\\prime}\\) of the Gaussians. Expressions bring about the geometric deformation of the neutral model, which is modeled as the displacements of the Gaussian points." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1933" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.896, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.257, + 0.895, + 0.314 + ], + "angle": 0, + "content": "Figure 2. The overview of the Gaussian Head Avatar rendering and reconstruction. We first optimize the guidance model including a neutral mesh, a deformation MLP and a color MLP in the Initialization stage. Then we use them to initialize the neutral Gaussians and the dynamic generator. Finally, 2K RGB images are synthesized through differentiable rendering and the super-resolution network. The Gaussian Head Avatar are trained under the supervision of multi-view RGB videos." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.325, + 0.471, + 0.39 + ], + "angle": 0, + "content": "Specifically, we predict the displacements respectively controlled by the expression and the head pose in the canonical space through two different MLPs: \\( \\pmb{f}_{\\text{def}}^{exp} \\in \\Phi \\) and \\( \\pmb{f}_{\\text{def}}^{\\text{pose}} \\in \\Phi \\). Then, we add them to the neutral positions." + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.403, + 0.469, + 0.444 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} X ^ {\\prime} = \\boldsymbol {X} _ {0} + \\lambda_ {e x p} (\\boldsymbol {X} _ {0}) \\boldsymbol {f} _ {d e f} ^ {e x p} (\\boldsymbol {X} _ {0}, \\theta) \\tag {3} \\\\ + \\lambda_ {p o s e} (\\boldsymbol {X} _ {0}) \\boldsymbol {f} _ {d e f} ^ {p o s e} (\\boldsymbol {X} _ {0}, \\beta). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.457, + 0.47, + 0.671 + ], + "angle": 0, + "content": "\\(\\lambda_{exp}(\\cdot)\\) and \\(\\lambda_{pose}(\\cdot)\\) represent the extent to which the points are affected by the expression or the head pose respectively. Without decoupling but just using the expression coefficients as the global condition [16] which also controls the shoulders and upper body, will produce jittering results during the animation. Here, we assume that the Gaussian points closer to 3D landmarks are more affected by the expression coefficients and less affected by the head pose, while the opposite is true for the Gaussian points far away. Specifically, The 3D landmarks \\(P_0\\) of the canonical model are first estimated through the 3DMM model in the data preprocessing and then optimized in the initialization stage 4.3. Then for each Gaussian point, we calculate the above weight \\(\\lambda_{exp}(\\cdot)\\) and \\(\\lambda_{pose}(\\cdot)\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.684, + 0.443, + 0.734 + ], + "angle": 0, + "content": "\\[\n\\lambda_ {e x p} (x) = \\left\\{ \\begin{array}{l l} 1, & d i s t (x, \\boldsymbol {P} _ {0}) < t _ {1} \\\\ \\frac {t _ {2} - d i s t (x , \\boldsymbol {P} _ {0})}{t _ {2} - t _ {1}}, & d i s t (x, \\boldsymbol {P} _ {0}) \\in [ t _ {1}, t _ {2} ] \\\\ 0, & d i s t (x, \\boldsymbol {P} _ {0}) > t _ {2} \\end{array} \\right.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.748, + 0.469, + 0.838 + ], + "angle": 0, + "content": "with \\(\\lambda_{\\text{pose}}(x) = 1 - \\lambda_{\\text{exp}}(x)\\). And \\(x \\in \\mathbf{X}_0\\) denotes the position of one neutral Gaussian. \\(\\text{dist}(x, P_0)\\) denotes the minimum distance from the point \\(x\\) to the 3D landmarks \\(P_0\\). \\(t_1 = 0.15\\) and \\(t_2 = 0.25\\) are predefined hyperparameters when the length of the head is set to approximately 1." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Color \\(C'\\) of the Gaussians. Modeling the dynamic details typically requires dynamic color that changes with expressions. As we do not pre-define the neutral value in Eqn. 2, the color are directly predict by two color MLPs:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.324, + 0.678, + 0.342 + ], + "angle": 0, + "content": "\\(\\pmb{f}_{col}^{exp}\\in \\Phi\\) and \\(\\pmb{f}_{col}^{pose}\\in \\Phi\\)" + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.35, + 0.892, + 0.387 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} C ^ {\\prime} = \\lambda_ {e x p} \\left(\\boldsymbol {X} _ {0}\\right) \\boldsymbol {f} _ {c o l} ^ {e x p} \\left(\\boldsymbol {F} _ {0}, \\theta\\right) \\tag {4} \\\\ + \\lambda_ {p o s e} (\\boldsymbol {X} _ {0}) \\boldsymbol {f} _ {c o l} ^ {p o s e} (\\boldsymbol {F} _ {0}, \\beta). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.395, + 0.892, + 0.473 + ], + "angle": 0, + "content": "Rotation, Scale and Opacity \\(\\{Q', S', A'\\}\\) of the Gaussians. These three attributes also dynamic, thereby modeling some detailed expressions-related appearance changes. Here, we just use another two attribute MLPs \\(f_{att}^{exp} \\in \\Phi\\) and \\(f_{att}^{pose} \\in \\Phi\\) to predict their shift from the neutral value." + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.481, + 0.891, + 0.537 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\{Q ^ {\\prime}, S ^ {\\prime}, A ^ {\\prime} \\} = \\left\\{\\mathbf {Q} _ {0}, \\mathbf {S} _ {0}, \\mathbf {A} _ {0} \\right\\} \\\\ + \\lambda_ {e x p} (\\boldsymbol {X} _ {0}) \\boldsymbol {f} _ {a t t} ^ {e x p} \\left(\\boldsymbol {F} _ {0}, \\theta\\right) \\tag {5} \\\\ + \\lambda_ {p o s e} (\\boldsymbol {X} _ {0}) \\boldsymbol {f} _ {a t t} ^ {p o s e} (\\boldsymbol {F} _ {0}, \\beta). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.546, + 0.892, + 0.636 + ], + "angle": 0, + "content": "Finally, we apply rigid rotations and translations \\( T(\\cdot) \\) to the Gaussians, transforming them from the canonical space to the world space. Note, the transformation is only implemented for directional variables: \\( \\{X', Q'\\} \\), while the multichannel color, the scale and the opacity \\( \\{C', S', A'\\} \\) are not directional thus remain unchanged." + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.646, + 0.891, + 0.663 + ], + "angle": 0, + "content": "\\[\n\\{X, Q \\} = T \\left(\\left\\{X ^ {\\prime}, Q ^ {\\prime} \\right\\}, \\beta\\right), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.609, + 0.672, + 0.891, + 0.689 + ], + "angle": 0, + "content": "\\[\n\\{C, S, A \\} = \\left\\{C ^ {\\prime}, S ^ {\\prime}, A ^ {\\prime} \\right\\}. \\tag {7}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.697, + 0.605, + 0.713 + ], + "angle": 0, + "content": "4.2. Training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.87 + ], + "angle": 0, + "content": "In this part, we explain the training pipeline of the Gaussian head avatar 4.1 and the loss function. In each iteration, we first generate the expression conditioned 3D Gaussians as Eqn. 2. Then we render a 32-channel image with 512 resolution \\( I_C \\in \\mathbb{R}^{512 \\times 512 \\times 32} \\) referring to Eqn. 1. After that we feed the image to a super resolution network \\( \\Psi \\) to generate a 2048 resolution RGB image \\( I_{hr} \\in \\mathbb{R}^{2048 \\times 2048 \\times 3} \\), such that more details are recovered and noise caused by uneven ambient light or camera chromatic aberration in the training data will be filtered out [49, 64]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "During training, we jointly optimize all the learnable parameters mentioned above in bold, including the neutral" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1934" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.091, + 0.47, + 0.226 + ], + "angle": 0, + "content": "Gaussians: \\(\\{X_0, F_0, Q_0, S_0, A_0\\}\\), the dynamic generator: \\(\\{f_{col}^{exp}, f_{col}^{pose}, f_{def}^{exp}, f_{def}^{pose}, f_{att}^{exp}, f_{att}^{pose}\\}\\), and the super resolution network \\(\\Psi\\). For the loss function, we only use the foreground RGB images \\(I_{gt}\\) as supervision to construct an L1 loss and a VGG perceptual loss [70] between the generated images \\(I_{hr}\\) and the ground truth \\(I_{gt}\\). Besides, we encourage the first three channels of the 32-channel feature image \\(I_C\\) to be RGB channels, which is ensured by a L1 loss term. The total loss is:" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.234, + 0.468, + 0.27 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} = \\left\\| I _ {h r} - I _ {g t} \\right\\| _ {1} + \\lambda_ {v g g} V G G \\left(I _ {h r}, I _ {g t}\\right) \\tag {8} \\\\ + \\lambda_ {l r} | | I _ {l r} - I _ {g t} | | _ {1}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.278, + 0.469, + 0.31 + ], + "angle": 0, + "content": "with \\( I_{lr} \\) denoting the first three channels of the 32-channel image \\( I_C \\). We set the weights \\( \\lambda_{vgg} = 0.1 \\) and \\( \\lambda_{lr} = 1 \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.317, + 0.353, + 0.332 + ], + "angle": 0, + "content": "4.3. Geometry-guided Initialization" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.34, + 0.469, + 0.445 + ], + "angle": 0, + "content": "Unlike neural networks, the Gaussians act as an unordered and unstructured representation. Random initialization leads to failure to converge while naively using the FLAME model to initialize will significantly reduce the reconstruction quality. In this section, we describe in detail how to optimize a mesh guidance model to provide reliable initialization for the Gaussians in Sec. 4.1." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.446, + 0.469, + 0.52 + ], + "angle": 0, + "content": "Mesh Guidance Model. Specifically, we first construct a MLP \\( f_{sdf} \\) to represent a signed distance field. In addition, this network will also output the corresponding feature vector of each point, which is used for predicting the point color. It can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.22, + 0.532, + 0.469, + 0.548 + ], + "angle": 0, + "content": "\\[\ns, \\eta = \\boldsymbol {f} _ {s d f} (x), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.557, + 0.469, + 0.783 + ], + "angle": 0, + "content": "with \\( s \\) denotes the SDF value, \\( \\eta \\) denotes the feature vector and \\( x \\) denotes the point position. Then through (DMTet) [51], we can differentially extract a mesh with vertices \\( X \\), per-vertex feature vectors \\( F \\) and its faces. We also predict the per-vertex 32-channel color as Eqn. 4 by the two color MLPs \\( f_{col}^{exp} \\) and \\( f_{col}^{pose} \\). In parallel, we construct the two deformation MLPs: \\( f_{def}^{exp} \\) and \\( f_{def}^{pose} \\) as described in Sec. 4.1 to predict the displacements and add them to the vertex positions. This process is similar to Eqn 3 above, with the Gaussian positions \\( X_0 \\) replaced by the vertex positions \\( X \\). Finally we also apply rigid rotations and translations to the deformed mesh, transforming it to the world space and render the deformed mesh into an image \\( I \\) and a mask \\( M \\) through differentiable rasterization [44] according to the camera parameters \\( \\mu \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.784, + 0.469, + 0.826 + ], + "angle": 0, + "content": "Loss Function and Training. Next, we can construct the RGB loss and the silhouette loss to train the guidance model:" + }, + { + "type": "equation", + "bbox": [ + 0.187, + 0.828, + 0.469, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {R G B} = \\left| \\left| I _ {r, g, b} - I _ {g t} \\right| \\right| _ {1}, \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.196, + 0.849, + 0.468, + 0.866 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s i l} = I O U (M, M _ {g t}), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "with \\(I_{gt}\\) and \\(M_{gt}\\) denote the ground truth RGB image and mask, respectively. \\(IOU(\\cdot)\\) denotes Intersection over" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "Union metrics. Note that only the first three channels \\( R, G, B \\) of the 32-channel image \\( I \\) are supervised by the ground truth RGB images." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.137, + 0.892, + 0.212 + ], + "angle": 0, + "content": "We also use the estimated 3D facial landmarks \\(P_{gt}\\) as described in Sec. 3 to provide rough guidance for the expression deformation MLP. Specifically, we input the neutral 3D landmarks \\(P_0\\) into the expression deformation MLP to predict the expression conditioned landmarks \\(P\\):" + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.223, + 0.892, + 0.241 + ], + "angle": 0, + "content": "\\[\nP = \\boldsymbol {P} _ {0} + \\boldsymbol {f} _ {\\text {d e f}} ^ {\\exp} (\\boldsymbol {P} _ {0}, \\theta). \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.25, + 0.892, + 0.281 + ], + "angle": 0, + "content": "Then we construct the loss function with 3D facial landmarks \\( P_{gt} \\) as the supervision:" + }, + { + "type": "equation", + "bbox": [ + 0.623, + 0.292, + 0.891, + 0.309 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d e f}} = \\left| \\left| P - P _ {g t} \\right| \\right| _ {2}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.319, + 0.892, + 0.453 + ], + "angle": 0, + "content": "Besides, we introduce three constraints: (1) a regular term \\(\\mathcal{L}_{offset}\\) to punish all non-zero displacements to prevent the two deformation MLPs from learning a global constant offset [63], (2) a regular term \\(\\mathcal{L}_{lmk}\\) to limit the SDF value at the 3D landmarks to be close to zero, such that the landmarks are located on the surface of the mesh, (3) a Laplacian term \\(\\mathcal{L}_{lap}\\) for maintaining the extracted mesh smooth to a certain extent. Overall, the total loss function is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.463, + 0.891, + 0.499 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} = \\mathcal {L} _ {R G B} + \\lambda_ {s i l} \\mathcal {L} _ {s i l} + \\lambda_ {d e f} \\mathcal {L} _ {d e f} + \\tag {14} \\\\ \\lambda_ {o f f s e t} \\mathcal {L} _ {o f f s e t} + \\lambda_ {l m k} \\mathcal {L} _ {l m k} + \\lambda_ {l a p} \\mathcal {L} _ {l a p}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.508, + 0.892, + 0.583 + ], + "angle": 0, + "content": "with \\(\\lambda\\) denoting the weights of each term, which are set as follows: \\(\\lambda_{sil} = 0.1\\), \\(\\lambda_{def} = 1\\), \\(\\lambda_{offset} = 0.01\\), \\(\\lambda_{lmk} = 0.1\\) and \\(\\lambda_{lap} = 100\\). We jointly optimize the MLPs mentioned above with the neutral 3D landmarks \\(P_0\\) jointly until all MLPs are converged." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.584, + 0.892, + 0.78 + ], + "angle": 0, + "content": "Parameters Transfer. Finally, we use the roughly trained mesh guidance model to initialize the Gaussian model. Specifically, we extract the neutral mesh with vertices \\(X\\) and per-vertex features \\(F\\) through DMTet, and directly assign their values to the neutral positions \\(\\mathbf{X}_0 = X\\) and the per-vertex feature vectors \\(\\pmb{F}_{0} = F\\) of the neutral Gaussians respectively. Then, we retain all the four optimized MLPs: \\(\\{\\pmb{f}_{col}^{exp},\\pmb{f}_{col}^{pose},\\pmb{f}_{def}^{exp},\\pmb{f}_{def}^{pose}\\}\\) for the Gaussian model. For the other neutral attributes: rotation, scale and opacity, we adopt the original initialization strategy in Gaussian Splatting [26]. And the parameters of the two attribute MLPs: \\(\\{\\pmb{f}_{att}^{exp},\\pmb{f}_{att}^{pose}\\}\\) and the super resolution network \\(\\Psi\\) are just randomly initialized." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.793, + 0.633, + 0.81 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.818, + 0.717, + 0.833 + ], + "angle": 0, + "content": "5.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.892, + 0.9 + ], + "angle": 0, + "content": "In the experiment we use 12 sets of data, 10 of which are from NeRSemble [28], and the other 2 are multi-view video data from HAvatar [71]. For the 10 identities from NeRSemble, each set contains 2500 to 3000 frames, 16" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1935" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.089, + 0.237, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.243, + 0.089, + 0.401, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.089, + 0.565, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.089, + 0.73, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.736, + 0.09, + 0.894, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.281, + 0.236, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.109, + 0.469, + 0.208, + 0.48 + ], + "angle": 0, + "content": "NeRFBlendShape" + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.282, + 0.401, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.296, + 0.469, + 0.348, + 0.479 + ], + "angle": 0, + "content": "NeRFace" + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.282, + 0.565, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.462, + 0.47, + 0.512, + 0.479 + ], + "angle": 0, + "content": "HAvatar" + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.282, + 0.728, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.636, + 0.47, + 0.666, + 0.479 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.736, + 0.279, + 0.893, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.804, + 0.47, + 0.823, + 0.479 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.488, + 0.893, + 0.517 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparisons of different methods on self reenactment task. From left to right: NeRFBlendShape [17], NeRFace [16], HAvatar [71] and Ours. Our method can reconstruct details like beards, teeth, etc. with high quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.525, + 0.47, + 0.645 + ], + "angle": 0, + "content": "cameras are distributed about 120 degrees in front, and simultaneously capture 2K resolution video. For each identity, We use the sequences marked with \"FREE\" as the evaluation data, and the rest as the training data. For the 2 identities from HAvatar, each set contains 3000 frames, 8 cameras are distributed about 120 degrees in front, and 4K resolution videos are collected simultaneously. Later, we crop the face area and resize to 2K resolution." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.648, + 0.471, + 0.769 + ], + "angle": 0, + "content": "For data preprocessing, we first remove the background [35] and extract 68 2D facial landmarks [10] for all the images. Then, for each frame, we use multi-view images to estimate the corresponding 3D landmarks, expression coefficients and head pose by fitting the Basel Face Model (BFM) [18] to the extracted 2D landmarks. Note that we define the 3D landmarks as the usual 68 landmarks with vertices indexed as multiples of 100 in the BFM vertices." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.785, + 0.31, + 0.801 + ], + "angle": 0, + "content": "5.2. Results and Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Self Reenactment. In this section, we first compare our method with existing SOTA methods in qualitative experiments on self reenactment task. Specifically, NeRFace [16] uses a deep MLP to fit an expression conditioned dynamic NeRF. The current SOTA method HAvatar [71] introduces 3DMM template prior and uses a deep convolutional net" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.525, + 0.893, + 0.616 + ], + "angle": 0, + "content": "work to generate a human head NeRF represented by three planes from a mesh template with expression. Note, HAvatar leverages the GAN framework using the adversarial loss function to force the network to generate details that are not view-consistent. For a fair comparison, we remove this part and use VGG perceptual loss as Sec. 4.2 instead." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.616, + 0.893, + 0.692 + ], + "angle": 0, + "content": "Qualitative results on self reenactment task are shown in the Fig. 3. Our method can accurately reconstruct pixel-level high-frequency details such as beards, teeth, and hair. Besides, our method can achieve expression transfer more accurately, such as eye movements in the figure." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.693, + 0.895, + 0.875 + ], + "angle": 0, + "content": "Next, we conduct a quantitative evaluation for the four methods on 5 identities and 6 cameras using the evaluation split. The evaluation metrics include: Peak Signal-to-Noise Ratio (PSNR), Structure Similarity Index (SSIM), Learned Perceptual Image Patch Similarity (LPIPS) [70] and Fréchet Inception Distance (FID) [22]. Note, we calculate FID by comparing the distribution of all the training images and all the rendered images. As the task mainly focuses on the reconstruction of the head, we use face parsing to remove the body parts in the image to eliminate their impact in the experiment. As shown in Tab. 1, our method demonstrates a slight improvement in PSNR and SSIM compared with" + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.887, + 0.796, + 0.901 + ], + "angle": 0, + "content": "\\(^{1}\\)https://github.com/zllrunning/faceParsing.PyTorch" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1936" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.116, + 0.089, + 0.852, + 0.186 + ], + "angle": 0, + "content": "
MethodPSNR ↑SSIM ↑LPIPS (512) ↓LPIPS (2K) ↓FID (2K) ↓
NeRFBlendShape25.910.8360.1230.22954.80
NeRFace27.140.8490.1470.23465.11
HAVatar27.190.8830.0640.20931.06
Ours (w/o SR)27.820.8870.0800.20245.50
Ours27.700.8830.0560.09818.50
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.189, + 0.893, + 0.218 + ], + "angle": 0, + "content": "Table 1. Quantitative evaluation results of NeRFBlendShape [17], NeRFace [16], HAvatar [71], our method without super resolution and our full method on self reenactment task." + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.228, + 0.892, + 0.504 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.512, + 0.893, + 0.542 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparisons of different methods on cross-identity reenactment task. From left to right: NeRFBlendShape [17], NeRFace [16], HAvatar [71] and Ours. Our method synthesizes high-fidelity images while ensuring the accuracy of expression transfer." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.559, + 0.452, + 0.643 + ], + "angle": 0, + "content": "
MethodPSNR ↑SSIM ↑LPIPS ↓
NeRFBlendShape25.430.8120.148
NeRFace26.650.8250.151
HAVatar27.130.8800.65
Ours27.580.8820.059
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.646, + 0.47, + 0.675 + ], + "angle": 0, + "content": "Table 2. Quantitative evaluation results of the other SOTA methods and our method on 3D consistency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.688, + 0.47, + 0.734 + ], + "angle": 0, + "content": "previous methods, and a significant improvement in LPIPS and FID, which means that our method can generate more high-frequency details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.469, + 0.809 + ], + "angle": 0, + "content": "Cross-Identity Reenactment. We also qualitatively compare our method with the above SOTA methods on cross-identity reenactment task. As shown in the results, our method is able to synthesize higher-fidelity images with more accurate expression transfer and richer emotions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Novel View Synthesis. In this section, we show the results of novel view synthesis as shown in Fig. 5 shown. In this case, we use video data from 8 views for training and render the image at a new viewpoint. Next, we quantitatively evaluate the 3D consistency of our method and compare it with other SOTA methods mentioned above. Specif" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.556, + 0.893, + 0.648 + ], + "angle": 0, + "content": "ically, we select the 5 identities as above and use the video data from 8 cameras for training while the other 8 holdout cameras for evaluation. The evaluation metrics include PSNR, SSIM and LPIPS (512 resolution). The results are shown in Tab. 2. Our method outperforms other methods in 3D consistency." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.658, + 0.655, + 0.675 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "table", + "bbox": [ + 0.528, + 0.681, + 0.862, + 0.749 + ], + "angle": 0, + "content": "
MethodPSNR ↑SSIM ↑LPIPS ↓
FLAME-Init28.730.8750.123
Mesh-Deform28.830.8740.116
Ours28.940.8760.108
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.753, + 0.892, + 0.78 + ], + "angle": 0, + "content": "Table 3. Quantitative evaluation results of the other two ablation baselines and ours on self reenactment task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Ablation on Initialization Strategies. In order to verify the effectiveness of our geometry-guided initialization strategy 4.3, we compare it with the strategy to use the FLAME model for initialization (FLAME-Init). Specifically, after fitting a FLAME model through multi-view data, we first subdivide the FLAME mesh 4 times and use the neutral vertices as the positions of the neutral Gaussians. Then, the" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1937" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.089, + 0.895, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.15, + 0.218, + 0.819, + 0.233 + ], + "angle": 0, + "content": "Figure 5. Novel view synthesis results of our method. We use 8-view synchronized videos for training the avatar." + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.244, + 0.473, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.485, + 0.47, + 0.527 + ], + "angle": 0, + "content": "Figure 6. Ablation study on the initialization strategies: FLAME-initialization and our geometry-guided initialization. Our strategy ensures the hair strands away from the head are well reconstructed." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.547, + 0.47, + 0.683 + ], + "angle": 0, + "content": "expression deformation MLP is optimized to learn the displacement of FLAME vertices. We set the per-vertex feature to zeros, while randomly initialize the parameters of the expression color MLP. The initialization of other variables is the same as our strategy. Qualitative results are shown in the Fig. 6. Due to the lack of initialization for the hair and shoulders in FLAME-Init, the points to model these parts are offset from nearby vertices, which leads to sparseness of the Gaussians, resulting in blurring." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.69, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Ablation on Deformation Modeling Approaches. We compare our fully learned deformation field with the previous mesh-based deformation (Mesh-Deform). Specifically, we migrate the method in INSTA [76] for controlling the NeRF deformation to our Gaussians. First we fit a 3DMM mesh template. Then, for each Gaussian point, find the closest face on the mesh, and calculate the deformation gradient to estimate the displacement. Qualitative results are shown in the Fig. 7. For some expressions that cannot be captured well by the 3DMM mesh template, our method can learn accurate deformation, thereby achieving the modeling of complex expressions. Quantitative results are shown in the Fig. 3. Our method outperforms both the two ablation baselines on PSNR, SSIM and LPIPS metrics." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.245, + 0.895, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.484, + 0.893, + 0.527 + ], + "angle": 0, + "content": "Figure 7. Ablation study on the deformation modeling: mesh LBS-based deformation and our fully learned deformation. Our approach can learn complex and exaggerated expressions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.54, + 0.748, + 0.555 + ], + "angle": 0, + "content": "6. Discussion and Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.57, + 0.892, + 0.63 + ], + "angle": 0, + "content": "Ethical Considerations. Our method is capable of creating artificial portrait videos, which have the potential to disseminate misinformation, influence public perceptions, and erode confidence in media sources." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.635, + 0.892, + 0.68 + ], + "angle": 0, + "content": "Limitation. For the tongue and teeth inside the mouth or long hair, blurring is sometimes produced in our method due to the lack of tracking methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.685, + 0.892, + 0.835 + ], + "angle": 0, + "content": "Conclusion. In this paper, we propose Gaussian Head Avatar, a novel representation for head avatar reconstruction, which leverages dynamic 3D Gaussians controlled by a fully learned expression deformation. Experiments demonstrate our method can synthesize ultra high-fidelity images while modeling exaggerated expressions. In addition, we propose a well-designed minute-level initialization strategy to ensure the training convergence. We believe our Gaussian Head Avatar will become the mainstream direction for head avatar reconstruction in the future." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement. This paper is supported by National Key R&D Program of China (2022YFF0902200), the NSFC project No.62125107, the Beijing Municipal Science & Technology Z231100005923030." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1938" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.169 + ], + "angle": 0, + "content": "[1] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In European Conference on Computer Vision, pages 696-712, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.472, + 0.239 + ], + "angle": 0, + "content": "[2] ShahRukh Athar, Zexiang Xu, Kalyan Sunkavalli, Eli Shechtman, and Zhixin Shu. Rignerf: Fully controllable neural 3d portraits. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.241, + 0.47, + 0.309 + ], + "angle": 0, + "content": "[3] ShahRukh Athar, Zhixin Shu, and Dimitris Samaras. Flame-in-nerf: Neural control of radiance fields for free view face animation. In IEEE 17th International Conference on Automatic Face and Gesture Recognition (FG), pages 1–8, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.311, + 0.47, + 0.354 + ], + "angle": 0, + "content": "[4] Thabo Beeler, Bernd Bickel, Paul Beardsley, Bob Sumner, and Markus Gross. High-quality single-shot capture of facial geometry. ACM Trans. Graph., 29(4), 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.354, + 0.47, + 0.422 + ], + "angle": 0, + "content": "[5] Sai Bi, Stephen Lombardi, Shunsuke Saito, Tomas Simon, Shih-En Wei, Kevyn Mcphail, Ravi Ramamoorthi, Yaser Sheikh, and Jason Saragih. Deep reeligible appearance models for animatable faces. ACM Trans. Graph., 40(4), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.47, + 0.478 + ], + "angle": 0, + "content": "[6] V Blanz and T Vetter. A morphable model for the synthesis of 3d faces. In 26th Annual Conference on Computer Graphics and Interactive Techniques (SIGGRAPH 1999), pages 187-194. ACM Press, 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.48, + 0.47, + 0.535 + ], + "angle": 0, + "content": "[7] Timo Bolkart, Tianye Li, and Michael J. Black. Instant multi-view head capture through learnable registration. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 768-779, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.536, + 0.47, + 0.577 + ], + "angle": 0, + "content": "[8] Derek Bradley, Wolfgang Heidrich, Tiberiu Popa, and Alla Sheffer. High resolution passive facial performance capture. 29(4), 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.578, + 0.47, + 0.634 + ], + "angle": 0, + "content": "[9] Alan Brunton, Timo Bolkart, and Stefanie Wuhrer. Multilinear wavelets: A statistical shape space for human faces. In Proceedings of the Proceedings of the European Conference on Computer Vision (ECCV), 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.635, + 0.47, + 0.689 + ], + "angle": 0, + "content": "[10] Adrian Bulat and Georgios Tzimiropoulos. How far are we from solving the 2d & 3d face alignment problem? (and a dataset of 230,000 3d facial landmarks). In International Conference on Computer Vision, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.69, + 0.47, + 0.746 + ], + "angle": 0, + "content": "[11] Chen Cao, Yanlin Weng, Shun Zhou, Y. Tong, and Kun Zhou. Facewarehouse: A 3d facial expression database for visual computing. In IEEE Transactions on Visualization and Computer Graphics, pages 413-425, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.747, + 0.47, + 0.788 + ], + "angle": 0, + "content": "[12] Chen Cao, Derek Bradley, Kun Zhou, and Thabo Beeler. Real-time high-fidelity facial performance capture. ACM Trans. Graph., 34(4), 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.789, + 0.47, + 0.83 + ], + "angle": 0, + "content": "[13] Chen Cao, Hongzhi Wu, Yanlin Weng, Tianjia Shao, and Kun Zhou. Real-time facial animation with image-based dynamic avatars. ACM Trans. Graph., 35(4), 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.831, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[14] Chen Cao, Tomas Simon, Jin Kyu Kim, Gabe Schwartz, Michael Zollhoefer, Shun-Suke Saito, Stephen Lombardi, Shih-En Wei, Danielle Belko, Shouu-I Yu, Yaser Sheikh, and Jason Saragih. Authentic volumetric avatars from a phone scan. ACM Trans. Graph., 41(4), 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.162 + ], + "angle": 0, + "content": "[15] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.894, + 0.232 + ], + "angle": 0, + "content": "[16] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Niessner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8645-8654, 2021. 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.894, + 0.3 + ], + "angle": 0, + "content": "[17] Xuan Gao, Chenglai Zhong, Jun Xiang, Yang Hong, Yudong Guo, and Juyong Zhang. Reconstructing personalized semantic facial nerf models from monocular video. ACM Transactions on Graphics (Proceedings of SIGGRAPH Asia), 41(6), 2022. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.301, + 0.894, + 0.356 + ], + "angle": 0, + "content": "[18] Thomas Gerig, Andreas Forster, Clemens Blumer, Bernhard Egger, Marcel Lüthi, Sandro Schonborn, and Thomas Vetter. Morphable face models - an open framework. pages 75-82, 2017. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.358, + 0.894, + 0.412 + ], + "angle": 0, + "content": "[19] Abhijeet Ghosh, Graham Fyffe, Borom Tunwattanapong, Jay Busch, Xueming Yu, and Paul Debevec. Multiview face capture using polarized spherical gradient illumination. ACM Trans. Graph., 30(6):1-10, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.413, + 0.894, + 0.482 + ], + "angle": 0, + "content": "[20] Philip-William Grassal, Malte Prinzler, Titus Leistner, Carsten Rother, Matthias Nießner, and Justus Thies. Neural head avatars from monocular rgb videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18632–18643, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.483, + 0.894, + 0.552 + ], + "angle": 0, + "content": "[21] Yudong Guo, Keyu Chen, Sen Liang, Yong-Jin Liu, Hujun Bao, and Juyong Zhang. Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5764–5774, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.553, + 0.894, + 0.608 + ], + "angle": 0, + "content": "[22] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Neural Information Processing Systems, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.894, + 0.676 + ], + "angle": 0, + "content": "[23] Yang Hong, Bo Peng, Haiyao Xiao, Ligang Liu, and Juyong Zhang. Headnerf: A real-time nerf-based parametric head model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20374-20384, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.678, + 0.894, + 0.746 + ], + "angle": 0, + "content": "[24] Liwen Hu, Shunsuke Saito, Lingyu Wei, Koki Nagano, Jaewoo Seo, Jens Fursund, Iman Sadeghi, Carrie Sun, YenChun Chen, and Hao Li. Avatar digitization from a single image for real-time rendering. ACM Trans. Graph., 36(6), 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.748, + 0.894, + 0.788 + ], + "angle": 0, + "content": "[25] Alexandru Eugen Ichim, Sofien Bouaziz, and Mark Pauly. Dynamic 3d avatar creation from hand-held video input. ACM Trans. Graph., 34(4), 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.894, + 0.844 + ], + "angle": 0, + "content": "[26] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.845, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[27] Taras Khakhulin, Vanessa Sklyarova, Victor Lempitsky, and Egor Zakharov. Realistic one-shot mesh-based head avatars. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1939" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[28] Tobias Kirschstein, Shenhan Qian, Simon Giebenhain, Tim Walter, and Matthias Nießner. Nersemble: Multi-view radiance field reconstruction of human heads. ACM Trans. Graph., 42(4), 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[29] Georgios Kopanas, Thomas Leimkuhler, Gilles Rainer, Clément Jambon, and George Drettakis. Neural point cata-caustics for novel-view synthesis of reflections. ACM Transactions on Graphics (TOG), 41(6):1-15, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.472, + 0.263 + ], + "angle": 0, + "content": "[30] Christoph Lassner and Michael Zollhofer. Pulsar: Efficient sphere-based neural rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1440-1449, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.472, + 0.375 + ], + "angle": 0, + "content": "[31] Marc Levoy, Kari Pulli, Brian Curless, Szymon Rusinkiewicz, David Koller, Lucas Pereira, Matt Ginzon, Sean Anderson, James Davis, Jeremy Ginsberg, Jonathan Shade, and Duane Fulk. The digital michelangelo project: 3d scanning of large statues. In Proceedings of the 27th Annual Conference on Computer Graphics and Interactive Techniques, page 131-144, USA, 2000. ACM Press/Addison-Wesley Publishing Co. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.376, + 0.47, + 0.431 + ], + "angle": 0, + "content": "[32] Tianye Li, Timo Bolkart, Michael J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6), 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.433, + 0.47, + 0.503 + ], + "angle": 0, + "content": "[33] Tianye Li, Shichen Liu, Timo Bolkart, Jiayi Liu, Hao Li, and Yajie Zhao. Topologically consistent multi-view face inference using volumetric sampling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3824-3834, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.472, + 0.574 + ], + "angle": 0, + "content": "[34] Zhe Li, Zerong Zheng, Lizhen Wang, and Yebin Liu. Animatable gaussians: Learning pose-dependent gaussian maps for high-fidelity human avatar modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.575, + 0.47, + 0.644 + ], + "angle": 0, + "content": "[35] Shanchuan Lin, Andrey Ryabtsev, Soumyadip Sengupta, Brian Curless, Steve Seitz, and Ira Kemelmacher-Shlizerman. Real-time high-resolution background matting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.645, + 0.47, + 0.702 + ], + "angle": 0, + "content": "[36] Xian Liu, Yinghao Xu, Qianyi Wu, Hang Zhou, Wayne Wu, and Bolei Zhou. Semantic-aware implicit neural auditory-driven video portrait generation. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.702, + 0.47, + 0.744 + ], + "angle": 0, + "content": "[37] Stephen Lombardi, Jason Saragih, Tomas Simon, and Yaser Sheikh. Deep appearance models for face rendering. ACM Trans. Graph., 37(4):68:1-68:13, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.745, + 0.47, + 0.802 + ], + "angle": 0, + "content": "[38] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural volumes: Learning dynamic renderable volumes from images. ACM Trans. Graph., 38(4):65:1-65:14, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.47, + 0.858 + ], + "angle": 0, + "content": "[39] Stephen Lombardi, Tomas Simon, Gabriel Schwartz, Michael Zollhoefer, Yaser Sheikh, and Jason Saragih. Mixture of volumetric primitives for efficient neural rendering. ACM Trans. Graph., 40(4), 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[40] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis, 2023. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.161 + ], + "angle": 0, + "content": "[41] Shugao Ma, Tomas Simon, Jason Saragih, Dawei Wang, Yuecheng Li, Fernando De La Torre, and Yaser Sheikh. Pixel codec avatars. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 64-73, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.162, + 0.893, + 0.232 + ], + "angle": 0, + "content": "[42] Marko Mihajlovic, Aayush Bansal, Michael Zollhoefer, Siyu Tang, and Shunsuke Saito. Keypointnerf: Generalizing image-based volumetric avatars using relative spatial encoding of keypoints. In European conference on computer vision, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.893, + 0.302 + ], + "angle": 0, + "content": "[43] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.302, + 0.893, + 0.385 + ], + "angle": 0, + "content": "[44] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting Triangular 3D Models, Materials, and Lighting From Images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8280-8290, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.386, + 0.893, + 0.442 + ], + "angle": 0, + "content": "[45] Koki Nagano, Jaewoo Seo, Jun Xing, Lingyu Wei, Zimo Li, Shunsuke Saito, Aviral Agarwal, Jens Fursund, and Hao Li. Pagan: Real-time avatars using dynamic textures. ACM Trans. Graph., 37(6), 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.893, + 0.512 + ], + "angle": 0, + "content": "[46] Pascal Paysan, Reinhard Knothe, Brian Amberg, Sami Romdhani, and Thomas Vetter. A 3d face model for pose and illumination invariant face recognition. In 2009 Sixth IEEE International Conference on Advanced Video and Signal Based Surveillance, pages 296-301, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.512, + 0.893, + 0.58 + ], + "angle": 0, + "content": "[47] Minghan Qin, Yifan Liu, Yuelang Xu, Xiaochen Zhao, Yebin Liu, and Haoqian Wang. High-fidelity 3d head avatars reconstruction through spatially-varying expression conditioned neural radiance field. In AAAI Conference on Artificial Intelligence, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.58, + 0.893, + 0.635 + ], + "angle": 0, + "content": "[48] Amit Raj, Michael Zollhoefer, Tomas Simon, Jason Saragih, Shunsuke Saito, James Hays, and Stephen Lombardi. Pva: Pixel-aligned volumetric avatars. In arXiv:2101.02697, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.636, + 0.893, + 0.678 + ], + "angle": 0, + "content": "[49] Darius Rückert, Linus Franke, and Marc Stamminger. Adop: Approximate differentiable one-pixel point rendering. ACM Trans. Graph., 41(4), 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.893, + 0.747 + ], + "angle": 0, + "content": "[50] Ruizhi Shao, Jingxiang Sun, Cheng Peng, Zerong Zheng, Boyao Zhou, Hongwen Zhang, and Yebin Liu. Control4d: Efficient 4d portrait editing with text. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.748, + 0.893, + 0.804 + ], + "angle": 0, + "content": "[51] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.893, + 0.859 + ], + "angle": 0, + "content": "[52] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.893, + 0.902 + ], + "angle": 0, + "content": "[53] Jingxiang Sun, Xuan Wang, Lizhen Wang, Xiaoyu Li, Yong Zhang, Hongwen Zhang, and Yebin Liu. Next3d: Generative neural texture rasterization for 3d-aware head avatars." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1940" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.189 + ], + "angle": 0, + "content": "[54] Cong Wang, Di Kang, Yanpei Cao, Linchao Bao, Ying Shan, and Song-Hai Zhang. Neural point-based volumetric avatar: Surface-guided neural points for efficient and photorealistic volumetric head avatar. In ACM SIGGRAPH Asia 2023 Conference Proceedings, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.191, + 0.469, + 0.26 + ], + "angle": 0, + "content": "[55] Daoye Wang, Prashanth Chandran, Gaspard Zoss, Derek Bradley, and Paulo Gotardo. Morf: Morphable radiance fields for multiview neural head modeling. In ACM SIGGRAPH 2022 Conference Proceedings, New York, NY, USA, 2022. Association for Computing Machinery. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.261, + 0.469, + 0.328 + ], + "angle": 0, + "content": "[56] Lizhen Wang, Zhiyuan Chen, Tao Yu, Chenguang Ma, Liang Li, and Yebin Liu. Faceverse: a fine-grained and detail-controllable 3d face morphable model from a hybrid dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.469, + 0.411 + ], + "angle": 0, + "content": "[57] Ziyan Wang, Timur Bagautdinov, Stephen Lombardi, Tomas Simon, Jason Saragih, Jessica Hodgins, and Michael Zollhofer. Learning compositional radiance fields of dynamic human heads. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5704-5713, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.413, + 0.469, + 0.453 + ], + "angle": 0, + "content": "[58] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. SynSin: End-to-end view synthesis from a single image. In CVPR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.469, + 0.509 + ], + "angle": 0, + "content": "[59] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.469, + 0.58 + ], + "angle": 0, + "content": "[60] Sijing Wu, Yichao Yan, Yunhao Li, Yuhao Cheng, Wenhan Zhu, Ke Gao, Xiaobo Li, and Guangtao Zhai. Ganhead: Towards generative animatable neural head avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 437-447, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.469, + 0.649 + ], + "angle": 0, + "content": "[61] Yunze Xiao, Hao Zhu, Haotian Yang, Zhengyu Diao, Xiangju Lu, and Xun Cao. Detailed facial geometry recovery from multi-view images by learning an implicit function. In Proceedings of the AAAI Conference on Artificial Intelligence, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.469, + 0.719 + ], + "angle": 0, + "content": "[62] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5438–5448, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.721, + 0.469, + 0.775 + ], + "angle": 0, + "content": "[63] Yuelang Xu, Lizhen Wang, Xiaochen Zhao, Hongwen Zhang, and Yebin Liu. Avatarmav: Fast 3d head avatar reconstruction using motion-aware neural voxels. In ACM SIGGRAPH 2023 Conference Proceedings, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[64] Yuelang Xu, Hongwen Zhang, Lizhen Wang, Xiaochen Zhao, Huang Han, Qi Guojun, and Yebin Liu. Latentavatar: Learning latent expression code for expressive neural head avatar. In ACM SIGGRAPH 2023 Conference Proceedings, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[65] Kai Yang, Hong Shang, Tianyang Shi, Xinghan Chen, Jingkai Zhou, Zhongqian Sun, and Wei Yang. Asm: Adaptive skinning model for high-quality 3d face modeling supplementary material. 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[66] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[67] Zeyu Yang, Hongye Yang, Zijie Pan, Xiatian Zhu, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.893, + 0.247 + ], + "angle": 0, + "content": "[68] T Yenamandra, A Tewari, F Bernard, HP Seidel, M Elgharib, D Cremers, and C Theobalt. i3dmm: Deep implicit 3d morphable model of human heads. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.249, + 0.892, + 0.316 + ], + "angle": 0, + "content": "[69] Wang Yifan, Felice Serena, Shihao Wu, Cengiz Öztireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM Transactions on Graphics (proceedings of ACM SIGGRAPH ASIA), 38(6), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.319, + 0.893, + 0.388 + ], + "angle": 0, + "content": "[70] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.39, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[71] Xiaochen Zhao, Lizhen Wang, Jingxiang Sun, Hongwen Zhang, Jinli Suo, and Yebin Liu. Havatar: High-fidelity head avatar via facial model conditioned neural radiance field. ACM Trans. Graph., 2023. Just Accepted. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.446, + 0.893, + 0.529 + ], + "angle": 0, + "content": "[72] Shunyuan Zheng, Boyao Zhou, Ruizhi Shao, Boning Liu, Shengping Zhang, Liqiang Nie, and Yebin Liu. Gpsgaussian: Generalizable pixel-wise 3d gaussian splatting for real-time human novel view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.531, + 0.892, + 0.613 + ], + "angle": 0, + "content": "[73] Yufeng Zheng, Victoria Fernández Abrevaya, Marcel C. Bühler, Xu Chen, Michael J. Black, and Otmar Hilliges. I m avatar: Implicit morphable head avatars from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13535-13545, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.615, + 0.892, + 0.683 + ], + "angle": 0, + "content": "[74] Yufeng Zheng, Wang Yifan, Gordon Wetzstein, Michael J. Black, and Otmar Hilliges. Pointavatar: Deformable point-based head avatars from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.685, + 0.892, + 0.727 + ], + "angle": 0, + "content": "[75] Yiyu Zhuang, Hao Zhu, Xusen Sun, and Xun Cao. Mofanerf: Morphable facial neural radiance field. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.728, + 0.892, + 0.755 + ], + "angle": 0, + "content": "[76] Wojciech Zielonka, Timo Bolkart, and Justus Thies. Instant volumetric head avatars, 2022. 2, 3, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.755 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1941" + } + ] +] \ No newline at end of file diff --git a/2024/Gaussian Head Avatar_ Ultra High-fidelity Head Avatar via Dynamic Gaussians/24a06516-83b5-41f4-8bb5-a4da984c01e1_origin.pdf b/2024/Gaussian Head Avatar_ Ultra High-fidelity Head Avatar via Dynamic Gaussians/24a06516-83b5-41f4-8bb5-a4da984c01e1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6b8a36ecf58ae61f3aa327e6b4c68cb42d691a0c --- /dev/null +++ b/2024/Gaussian Head Avatar_ Ultra High-fidelity Head Avatar via Dynamic Gaussians/24a06516-83b5-41f4-8bb5-a4da984c01e1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b818623dadfde852edace765ccb1b1973f18b32c60788e90b6a53f80afcabc7e +size 5685494 diff --git a/2024/Gaussian Head Avatar_ Ultra High-fidelity Head Avatar via Dynamic Gaussians/full.md b/2024/Gaussian Head Avatar_ Ultra High-fidelity Head Avatar via Dynamic Gaussians/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8de0cb7ef9e5f8219ffb0c97aaeef5668ade9505 --- /dev/null +++ b/2024/Gaussian Head Avatar_ Ultra High-fidelity Head Avatar via Dynamic Gaussians/full.md @@ -0,0 +1,362 @@ +# Gaussian Head Avatar: Ultra High-fidelity Head Avatar via Dynamic Gaussians + +Yuelang Xu $^{1}$ , Bengwang Chen $^{1}$ , Zhe Li $^{1}$ , Hongwen Zhang $^{2}$ , Lizhen Wang $^{1}$ , Zerong Zheng $^{3}$ , Yebin Liu $^{1\dagger}$ $^{1}$ Department of Automation, Tsinghua University $^{2}$ Beijing Normal University $^{3}$ NNKosmos Technology + +![](images/566be093fd8aedb6b1d056c06b0149e2f2c2f1bfa9da039746530b81ca4983e3.jpg) + +![](images/b2f5a3e47a3df591cc0811e9b5414e04ffbb9182328730acab361413101d2e75.jpg) +Figure 1. Gaussian head avatar achieves ultra high-fidelity image synthesis with controllable expressions at 2K resolution. The above shows different views of the synthesized avatar, and the bottom shows different identities animated by the same expression. 16 views are used during the training. + +# Abstract + +Creating high-fidelity 3D head avatars has always been a research hotspot, but there remains a great challenge under lightweight sparse view setups. In this paper, we propose Gaussian Head Avatar represented by controllable 3D Gaussians for high-fidelity head avatar modeling. We optimize the neutral 3D Gaussians and a fully learned MLP-based deformation field to capture complex expressions. The two parts benefit each other, thereby our method can model fine-grained dynamic details while ensuring expression accuracy. Furthermore, we devise a well-designed geometry-guided initialization strategy based on implicit SDF and Deep Marching Tetrahedra for the stability and convergence of the training procedure. Experiments show our approach outperforms other state-of-the-art sparse-view methods, achieving ultra high-fidelity rendering quality at $2K$ resolution even under exaggerated expressions. Project page: https://yuelangx.github.io/gaussianheadavatar. + +# 1. Introduction + +High-fidelity 3D human head avatar modeling is of great significance in many fields, such as VR/AR, telepresence, digital human and film production. Automatically creating high-fidelity avatars has been a research hotspot in computer vision for decades. Although some traditional head avatars [37, 39, 41, 54] can realize high-fidelity animation, they typically require accurate geometries reconstructed and tracked from dense multi-view videos, thus limiting their applications in lightweight settings. + +Thanks to the Neural Radiance Fields (NeRF) [43] which show great capability of novel view synthesis in the absence of accurate geometry, recent methods [38, 57] skip the geometry reconstruction and tracking steps but directly learn high-quality NeRF-based head avatars. Other works [42, 48, 71] have verified that NeRF can be applied to either dense or sparse views, which greatly lowers the threshold for head avatar reconstruction. However, it still remains challenging for these NeRF-based approaches to synthesize high-fidelity images at 2K resolutions with pixel + +level details, including wrinkles and eyes. + +To overcome this bottleneck and further improve the avatar quality, we introduce 3D Gaussian splatting [26] for 3D head avatar modeling. This is an explicit discrete representation that can be well adapted to the rasterization-based rendering pipeline. It has been verified that the 3D Gaussian representation is capable of rendering complex scenes with low computational consumption. Compared to NeRF, the reconstruction quality of static and dynamic scenes [40, 59, 67] is much better while rendering time cost has been significantly reduced. Motivated by this progress, we propose Gaussian Head Avatar, a novel representation that utilizes 3D Gaussian splatting for ultra high-fidelity head avatar modeling. Although recent 4D Gaussian works [40, 59, 66, 67] have been proposed to reconstruct dynamic scenes, all of them cannot be animated. For modeling the animatable head avatar, it is crucial but still unexplored how to effectively control the deformation of 3D Gaussians and model the dynamic appearances through expression coefficients. + +Previous explicit [74] and implicit [2, 73, 76] head avatars usually formulate the facial deformation via linear blend skinning (LBS) using the skinning weights and blendshapes like the FLAME model [32]. However, such an LBS-based formulation fails to represent exaggerated and fine-grained expressions by simple linear operations, limiting the representation ability of the head avatars. Inspired by NeRSemble [28], we propose a fully learnable expression-conditioned deformation field for the 3D head Gaussians, avoiding the limited capability of the LBS-based formulation. Specifically, we input the positions of the 3D Gaussians with expression coefficients into an MLP to directly predict the displacements from the neutral expression to the target one. Similarly, we control the motion of nonface areas, such as the neck, using the head pose as the condition. 3D Gaussian-based representation has the powerful ability to reconstruct high-frequency details, enabling our method to learn accurate deformation fields. In turn, the learned accurate deformation field facilitates the dynamic Gaussian head model to fit more dynamic details. As a result, our method is able to reconstruct finer-grained dynamic details of expressive human heads. + +Unfortunately, as a discretized representation, the gradients back-propagated to the 3D Gaussians cannot spread through the whole space. Thus the convergence of training heavily relies on a plausible initialization for both the geometry and the deformation field. However, simply initializing the 3D head Gaussians with a morphable template like FLAME [32] fails to model the long hairstyle and the shoulders. Hence, we further propose an efficient and well-designed geometry-guided initialization strategy. Specifically, instead of starting from stochastic Gaussians or a FLAME model, we initially optimize an implicit signed dis + +tance function (SDF) field along with a color field and a deformation MLP for modeling the basic geometry, color, and the expression-conditioned deformations of the head avatar respectively. The SDF field is converted to a mesh through Deep Marching Tetrahedra (DMTet) [52], with the color and deformation of the vertices predicted by the MLPs. Then we render the mesh and optimize them jointly under the supervision of multi-view RGB images. Finally, we use the mesh with per-vertex features from the SDF field to initialize the 3D Gaussians to lie on the basic head surface while the color and deformation MLPs are carried over to the next stage, ensuring stable training for convergence. The entire initialization process takes only around 10 minutes. + +The contributions of our method can be summarized as: + +- We propose Gaussian Head Avatar, a new head avatar representation that employs controllable dynamic 3D Gaussians to model expressive human head avatars, producing ultra high-fidelity synthesized images at 2K resolutions. +- For modeling high-frequency dynamic details, we employ a fully learned deformation field upon the 3D head Gaussians, which accurately model extremely complex and exaggerated facial expressions. +- We carefully design an efficient initialization strategy that leverages implicit representations to initialize the geometry and deformation, leading to efficient and robust convergence when training the Gaussian Head Avatar. + +Benefiting from these contributions, our method surpasses recent state-of-the-art methods under lightweight sparse-view setups on the avatar quality by a large margin. + +# 2. Related Works + +3D Head Avatar Reconstruction. Due to the wide application value in the film and digital human industry, 3D head avatar reconstruction from multi-view images has always been a research hotspot. Traditional works [4, 8, 19, 31] reconstruct the scan geometry through multi-view stereo and then register a face mesh template to it. However, such methods usually require heavy computation. With the utilization of deep neural networks, current methods [7, 33, 61, 65] achieve very fast reconstruction, producing even more accurate geometry. Lombardi et al. [37], Bi et al. [5] and Ma et al. [41] represent the full head mesh through a deep neural network and train it with multi-view videos as supervision. However, due to the errors in geometric estimation, mesh-based head avatars typically suffer from texture blur. Therefore, some recent methods [38, 57] utilize NeRF representation [43] to synthesize novel view images without geometry reconstruction, or build NeRF on the head mesh template [39]. Furthermore, the NeRF-based methods are extended to sparse view reconstruction tasks [28, 42, 48, 71] and achieve impressive performance. + +Methods which focus on generative model [6, 9, 11, 32, 46, 56, 60] are dedicated to learning general mesh + +face templates from large-scale multi-view face images or 3D scans. Recently, implicit SDF-based [68] or NeRF-based [14, 23, 53, 55, 75] methods can learn full-head templates without the limitations of fixed topology, thereby better modeling complex hairstyles and glasses. Cao et al. [14] adopts a hybrid representation of local NeRF built on the mesh surface, which enables high-fidelity rendering and flexible expression control. + +3D head avatars reconstruction from monocular videos is also a popular yet challenging research topic. Early methods [12, 13, 15, 24, 25, 45] optimize a morphable mesh to fit the training video. Recent methods [20, 27] leverage neural networks to learn non-rigid deformation upon 3DMM face templates [18, 32], thus can recover more dynamic details. Such methods are not flexible enough to handle complex topologies. Therefore, the latest methods explore to construct head avatar models based on implicit SDF [73], point clouds [74] or NeRF [2, 3, 16, 17, 21, 36, 47, 63, 64, 76]. + +Point-based Rendering. Point elements as a discrete and unstructured representation can fit geometry with arbitrary topology [69] efficiently. Recent methods [29, 30, 58] open up a differentiable rasterization pipeline, such that the point-based representation is widely used in multi-view reconstruction tasks. Aliev et al. [1] and Ruckert et al. [49] propose to first render the feature map, which is transferred to the images through a convolutional renderer. Xu et al.[62] use neural point cloud associated with neural features to model a NeRF. Recently, 3D Gaussian splatting [26] shows its superior performance, beating NeRF in both novel view synthesis quality and rendering speed. Some approaches [34, 40, 50, 59, 66, 67, 72, 72] extend Gaussian representation to dynamic scene reconstruction. However, these methods can not be migrated to the head avatar reconstruction tasks. + +# 3. Overview + +The pipeline of the reconstruction of Gaussian Head Avatar is illustrated in Fig. 2, including the initialization stage and the training stage of Gaussian Head Avatar. Before the beginning of the pipeline, we remove the background [35] of each image and jointly estimate the 3DMM model [18], 3D facial landmarks and the expression coefficients for each frame. In the initialization stage (Sec. 4.3), we reconstruct an SDF-based neutral geometry, and optimize a deformation MLP and a color MLP from the training data as the guidance model. Next, we extract the neutral mesh through DMTet to initialize the neutral Gaussians while the deformation and color MLPs are also inherited from the initialization stage. In the training stage of Gaussian Head Avatar (Sec. 4.2), we deform the neutral Gaussians to the target expression through the dynamic generator given the driving expression coefficients as the condition. Finally, given a camera view, the expressive Gaussians are rendered to a + +feature map, which is then fed into the convolutional super resolution network to generate high-resolution avatar images. The whole model is optimized under the supervision of multi-view RGB videos. + +# 4. Method + +# 4.1. Avatar Representation + +Generally, the static 3D Gaussians [26] with $N$ points are represented by their positions $X$ , the multi-channel color $C$ , the rotation $Q$ , scale $S$ and opacity $A$ . The rotation $Q$ is represented in the form of quaternion. Subsequently, the Gaussians can be rasterized and rendered to the multi-channel image $I$ given the camera parameters $\mu$ . This process can be formulated as: + +$$ +I = \mathcal {R} (X, C, Q, S, A; \mu). \tag {1} +$$ + +Our task is to reconstruct a dynamic head avatar controlled by expression coefficients. Therefore, we formulate the head avatar as dynamic 3D Gaussians conditioned on expressions. To handle the dynamic changes, we input the expression coefficients with head pose to the head avatar model and output the position and other attributes of the Gaussians as above. + +Specifically, we first construct a canonical neutral Gaussian model with expression-independent attributes: $\{X_0, F_0, Q_0, S_0, A_0\}$ , which are fully estimizable. $X_0 \in \mathbb{R}^{N \times 3}$ denotes the positions of the Gaussians with a neutral expression in the canonical space. $F_0 \in \mathbb{R}^{N \times 128}$ denotes the point-wise feature vectors as their intrinsic properties. $Q_0 \in \mathbb{R}^{N \times 4}$ , $S_0 \in \mathbb{R}^{N \times 3}$ and $A_0 \in \mathbb{R}^{N \times 1}$ denotes the neutral rotation, scale and opacity respectively. Note that we do not define the neutral color, but directly predict expression-dependent dynamic color from the point-wise feature vectors $F_0$ . Then, we construct an MLP-based expression conditioned dynamic generator $\Phi$ to generate all the extra dynamic changes to the neutral model. Overall, the whole Gaussian head avatar can be formulated as: + +$$ +\{X, C, Q, S, A \} = \Phi \left(\boldsymbol {X} _ {0}, \boldsymbol {F} _ {0}, \boldsymbol {Q} _ {0}, \boldsymbol {S} _ {0}, \boldsymbol {A} _ {0}; \theta , \beta\right), \tag {2} +$$ + +with $\theta$ denoting expression coefficients and $\beta$ denoting the head pose. During the training, we optimize all the parameters of the dynamic generator $\Phi$ and the neutral Gaussian model $\{X_0, F_0, Q_0, S_0, A_0\}$ , which are highlighted in bold in the following. + +Next, we explain the process of adding expression-related changes to the neutral Gaussian model through the dynamic generator $\Phi$ as described in Eqn. 2 in detail. + +Positions $X^{\prime}$ of the Gaussians. Expressions bring about the geometric deformation of the neutral model, which is modeled as the displacements of the Gaussian points. + +![](images/0532ff0b2d16e0d0cf8e507f5e09c0ad3a3cda6c445667a10def86811cf9672f.jpg) +Figure 2. The overview of the Gaussian Head Avatar rendering and reconstruction. We first optimize the guidance model including a neutral mesh, a deformation MLP and a color MLP in the Initialization stage. Then we use them to initialize the neutral Gaussians and the dynamic generator. Finally, 2K RGB images are synthesized through differentiable rendering and the super-resolution network. The Gaussian Head Avatar are trained under the supervision of multi-view RGB videos. + +Specifically, we predict the displacements respectively controlled by the expression and the head pose in the canonical space through two different MLPs: $\pmb{f}_{\text{def}}^{exp} \in \Phi$ and $\pmb{f}_{\text{def}}^{\text{pose}} \in \Phi$ . Then, we add them to the neutral positions. + +$$ +\begin{array}{l} X ^ {\prime} = \boldsymbol {X} _ {0} + \lambda_ {e x p} (\boldsymbol {X} _ {0}) \boldsymbol {f} _ {d e f} ^ {e x p} (\boldsymbol {X} _ {0}, \theta) \tag {3} \\ + \lambda_ {p o s e} (\boldsymbol {X} _ {0}) \boldsymbol {f} _ {d e f} ^ {p o s e} (\boldsymbol {X} _ {0}, \beta). \\ \end{array} +$$ + +$\lambda_{exp}(\cdot)$ and $\lambda_{pose}(\cdot)$ represent the extent to which the points are affected by the expression or the head pose respectively. Without decoupling but just using the expression coefficients as the global condition [16] which also controls the shoulders and upper body, will produce jittering results during the animation. Here, we assume that the Gaussian points closer to 3D landmarks are more affected by the expression coefficients and less affected by the head pose, while the opposite is true for the Gaussian points far away. Specifically, The 3D landmarks $P_0$ of the canonical model are first estimated through the 3DMM model in the data preprocessing and then optimized in the initialization stage 4.3. Then for each Gaussian point, we calculate the above weight $\lambda_{exp}(\cdot)$ and $\lambda_{pose}(\cdot)$ as follows: + +$$ +\lambda_ {e x p} (x) = \left\{ \begin{array}{l l} 1, & d i s t (x, \boldsymbol {P} _ {0}) < t _ {1} \\ \frac {t _ {2} - d i s t (x , \boldsymbol {P} _ {0})}{t _ {2} - t _ {1}}, & d i s t (x, \boldsymbol {P} _ {0}) \in [ t _ {1}, t _ {2} ] \\ 0, & d i s t (x, \boldsymbol {P} _ {0}) > t _ {2} \end{array} \right. +$$ + +with $\lambda_{\text{pose}}(x) = 1 - \lambda_{\text{exp}}(x)$ . And $x \in \mathbf{X}_0$ denotes the position of one neutral Gaussian. $\text{dist}(x, P_0)$ denotes the minimum distance from the point $x$ to the 3D landmarks $P_0$ . $t_1 = 0.15$ and $t_2 = 0.25$ are predefined hyperparameters when the length of the head is set to approximately 1. + +Color $C'$ of the Gaussians. Modeling the dynamic details typically requires dynamic color that changes with expressions. As we do not pre-define the neutral value in Eqn. 2, the color are directly predict by two color MLPs: + +$\pmb{f}_{col}^{exp}\in \Phi$ and $\pmb{f}_{col}^{pose}\in \Phi$ + +$$ +\begin{array}{l} C ^ {\prime} = \lambda_ {e x p} \left(\boldsymbol {X} _ {0}\right) \boldsymbol {f} _ {c o l} ^ {e x p} \left(\boldsymbol {F} _ {0}, \theta\right) \tag {4} \\ + \lambda_ {p o s e} (\boldsymbol {X} _ {0}) \boldsymbol {f} _ {c o l} ^ {p o s e} (\boldsymbol {F} _ {0}, \beta). \\ \end{array} +$$ + +Rotation, Scale and Opacity $\{Q', S', A'\}$ of the Gaussians. These three attributes also dynamic, thereby modeling some detailed expressions-related appearance changes. Here, we just use another two attribute MLPs $f_{att}^{exp} \in \Phi$ and $f_{att}^{pose} \in \Phi$ to predict their shift from the neutral value. + +$$ +\begin{array}{l} \{Q ^ {\prime}, S ^ {\prime}, A ^ {\prime} \} = \left\{\mathbf {Q} _ {0}, \mathbf {S} _ {0}, \mathbf {A} _ {0} \right\} \\ + \lambda_ {e x p} (\boldsymbol {X} _ {0}) \boldsymbol {f} _ {a t t} ^ {e x p} \left(\boldsymbol {F} _ {0}, \theta\right) \tag {5} \\ + \lambda_ {p o s e} (\boldsymbol {X} _ {0}) \boldsymbol {f} _ {a t t} ^ {p o s e} (\boldsymbol {F} _ {0}, \beta). \\ \end{array} +$$ + +Finally, we apply rigid rotations and translations $T(\cdot)$ to the Gaussians, transforming them from the canonical space to the world space. Note, the transformation is only implemented for directional variables: $\{X', Q'\}$ , while the multichannel color, the scale and the opacity $\{C', S', A'\}$ are not directional thus remain unchanged. + +$$ +\{X, Q \} = T \left(\left\{X ^ {\prime}, Q ^ {\prime} \right\}, \beta\right), \tag {6} +$$ + +$$ +\{C, S, A \} = \left\{C ^ {\prime}, S ^ {\prime}, A ^ {\prime} \right\}. \tag {7} +$$ + +# 4.2. Training + +In this part, we explain the training pipeline of the Gaussian head avatar 4.1 and the loss function. In each iteration, we first generate the expression conditioned 3D Gaussians as Eqn. 2. Then we render a 32-channel image with 512 resolution $I_C \in \mathbb{R}^{512 \times 512 \times 32}$ referring to Eqn. 1. After that we feed the image to a super resolution network $\Psi$ to generate a 2048 resolution RGB image $I_{hr} \in \mathbb{R}^{2048 \times 2048 \times 3}$ , such that more details are recovered and noise caused by uneven ambient light or camera chromatic aberration in the training data will be filtered out [49, 64]. + +During training, we jointly optimize all the learnable parameters mentioned above in bold, including the neutral + +Gaussians: $\{X_0, F_0, Q_0, S_0, A_0\}$ , the dynamic generator: $\{f_{col}^{exp}, f_{col}^{pose}, f_{def}^{exp}, f_{def}^{pose}, f_{att}^{exp}, f_{att}^{pose}\}$ , and the super resolution network $\Psi$ . For the loss function, we only use the foreground RGB images $I_{gt}$ as supervision to construct an L1 loss and a VGG perceptual loss [70] between the generated images $I_{hr}$ and the ground truth $I_{gt}$ . Besides, we encourage the first three channels of the 32-channel feature image $I_C$ to be RGB channels, which is ensured by a L1 loss term. The total loss is: + +$$ +\begin{array}{l} \mathcal {L} = \left\| I _ {h r} - I _ {g t} \right\| _ {1} + \lambda_ {v g g} V G G \left(I _ {h r}, I _ {g t}\right) \tag {8} \\ + \lambda_ {l r} | | I _ {l r} - I _ {g t} | | _ {1}, \\ \end{array} +$$ + +with $I_{lr}$ denoting the first three channels of the 32-channel image $I_C$ . We set the weights $\lambda_{vgg} = 0.1$ and $\lambda_{lr} = 1$ . + +# 4.3. Geometry-guided Initialization + +Unlike neural networks, the Gaussians act as an unordered and unstructured representation. Random initialization leads to failure to converge while naively using the FLAME model to initialize will significantly reduce the reconstruction quality. In this section, we describe in detail how to optimize a mesh guidance model to provide reliable initialization for the Gaussians in Sec. 4.1. + +Mesh Guidance Model. Specifically, we first construct a MLP $f_{sdf}$ to represent a signed distance field. In addition, this network will also output the corresponding feature vector of each point, which is used for predicting the point color. It can be formulated as: + +$$ +s, \eta = \boldsymbol {f} _ {s d f} (x), \tag {9} +$$ + +with $s$ denotes the SDF value, $\eta$ denotes the feature vector and $x$ denotes the point position. Then through (DMTet) [51], we can differentially extract a mesh with vertices $X$ , per-vertex feature vectors $F$ and its faces. We also predict the per-vertex 32-channel color as Eqn. 4 by the two color MLPs $f_{col}^{exp}$ and $f_{col}^{pose}$ . In parallel, we construct the two deformation MLPs: $f_{def}^{exp}$ and $f_{def}^{pose}$ as described in Sec. 4.1 to predict the displacements and add them to the vertex positions. This process is similar to Eqn 3 above, with the Gaussian positions $X_0$ replaced by the vertex positions $X$ . Finally we also apply rigid rotations and translations to the deformed mesh, transforming it to the world space and render the deformed mesh into an image $I$ and a mask $M$ through differentiable rasterization [44] according to the camera parameters $\mu$ . + +Loss Function and Training. Next, we can construct the RGB loss and the silhouette loss to train the guidance model: + +$$ +\mathcal {L} _ {R G B} = \left| \left| I _ {r, g, b} - I _ {g t} \right| \right| _ {1}, \tag {10} +$$ + +$$ +\mathcal {L} _ {s i l} = I O U (M, M _ {g t}), \tag {11} +$$ + +with $I_{gt}$ and $M_{gt}$ denote the ground truth RGB image and mask, respectively. $IOU(\cdot)$ denotes Intersection over + +Union metrics. Note that only the first three channels $R, G, B$ of the 32-channel image $I$ are supervised by the ground truth RGB images. + +We also use the estimated 3D facial landmarks $P_{gt}$ as described in Sec. 3 to provide rough guidance for the expression deformation MLP. Specifically, we input the neutral 3D landmarks $P_0$ into the expression deformation MLP to predict the expression conditioned landmarks $P$ : + +$$ +P = \boldsymbol {P} _ {0} + \boldsymbol {f} _ {\text {d e f}} ^ {\exp} (\boldsymbol {P} _ {0}, \theta). \tag {12} +$$ + +Then we construct the loss function with 3D facial landmarks $P_{gt}$ as the supervision: + +$$ +\mathcal {L} _ {\text {d e f}} = \left| \left| P - P _ {g t} \right| \right| _ {2}, \tag {13} +$$ + +Besides, we introduce three constraints: (1) a regular term $\mathcal{L}_{offset}$ to punish all non-zero displacements to prevent the two deformation MLPs from learning a global constant offset [63], (2) a regular term $\mathcal{L}_{lmk}$ to limit the SDF value at the 3D landmarks to be close to zero, such that the landmarks are located on the surface of the mesh, (3) a Laplacian term $\mathcal{L}_{lap}$ for maintaining the extracted mesh smooth to a certain extent. Overall, the total loss function is formulated as: + +$$ +\begin{array}{l} \mathcal {L} = \mathcal {L} _ {R G B} + \lambda_ {s i l} \mathcal {L} _ {s i l} + \lambda_ {d e f} \mathcal {L} _ {d e f} + \tag {14} \\ \lambda_ {o f f s e t} \mathcal {L} _ {o f f s e t} + \lambda_ {l m k} \mathcal {L} _ {l m k} + \lambda_ {l a p} \mathcal {L} _ {l a p}, \\ \end{array} +$$ + +with $\lambda$ denoting the weights of each term, which are set as follows: $\lambda_{sil} = 0.1$ , $\lambda_{def} = 1$ , $\lambda_{offset} = 0.01$ , $\lambda_{lmk} = 0.1$ and $\lambda_{lap} = 100$ . We jointly optimize the MLPs mentioned above with the neutral 3D landmarks $P_0$ jointly until all MLPs are converged. + +Parameters Transfer. Finally, we use the roughly trained mesh guidance model to initialize the Gaussian model. Specifically, we extract the neutral mesh with vertices $X$ and per-vertex features $F$ through DMTet, and directly assign their values to the neutral positions $\mathbf{X}_0 = X$ and the per-vertex feature vectors $\pmb{F}_{0} = F$ of the neutral Gaussians respectively. Then, we retain all the four optimized MLPs: $\{\pmb{f}_{col}^{exp},\pmb{f}_{col}^{pose},\pmb{f}_{def}^{exp},\pmb{f}_{def}^{pose}\}$ for the Gaussian model. For the other neutral attributes: rotation, scale and opacity, we adopt the original initialization strategy in Gaussian Splatting [26]. And the parameters of the two attribute MLPs: $\{\pmb{f}_{att}^{exp},\pmb{f}_{att}^{pose}\}$ and the super resolution network $\Psi$ are just randomly initialized. + +# 5. Experiments + +# 5.1. Implementation Details + +In the experiment we use 12 sets of data, 10 of which are from NeRSemble [28], and the other 2 are multi-view video data from HAvatar [71]. For the 10 identities from NeRSemble, each set contains 2500 to 3000 frames, 16 + +![](images/3fd16cc2aa69a95b28977d998402bb72b458d96f03b5396d6aa3fb0b88048e03.jpg) + +![](images/d39d1453b17767fac298dc00c9fad722b9368a291dac9f348ede15a1e7f2152b.jpg) + +![](images/7e5bd01e6ccb0454dd029612d2095b570267073250ef5ee364e2fe47f4247048.jpg) + +![](images/5d0caf3d13100366507cdd69ab838a29ee14101dcdd5b7434cc6576462b47711.jpg) + +![](images/4914b8a2be00e894b7950b90646037287f0cc87996450f4c7795723bc8bdbd91.jpg) + +![](images/68181e56aaf028fcb2eb3d7412505588663d0aef29deda2ec62fd1d9bb1ca8f1.jpg) +NeRFBlendShape + +![](images/9031d2608924608e6f8a59c32319ffa9b920451b6a89aa6c8b80b83bbe80c704.jpg) +NeRFace + +![](images/217d97f9e2a3cd03421d2449f28c231699851035f5b7b16b193efb76a9adcfed.jpg) +HAvatar + +![](images/275d43f84f8d006fb2ae6bdcd7430565282943ed3ee96c0a253a7dc20035ac9e.jpg) +Ours + +![](images/23edde6edb66fc22bb344622df640093796f068abfda61707cac7b45056c0c41.jpg) +GT +Figure 3. Qualitative comparisons of different methods on self reenactment task. From left to right: NeRFBlendShape [17], NeRFace [16], HAvatar [71] and Ours. Our method can reconstruct details like beards, teeth, etc. with high quality. + +cameras are distributed about 120 degrees in front, and simultaneously capture 2K resolution video. For each identity, We use the sequences marked with "FREE" as the evaluation data, and the rest as the training data. For the 2 identities from HAvatar, each set contains 3000 frames, 8 cameras are distributed about 120 degrees in front, and 4K resolution videos are collected simultaneously. Later, we crop the face area and resize to 2K resolution. + +For data preprocessing, we first remove the background [35] and extract 68 2D facial landmarks [10] for all the images. Then, for each frame, we use multi-view images to estimate the corresponding 3D landmarks, expression coefficients and head pose by fitting the Basel Face Model (BFM) [18] to the extracted 2D landmarks. Note that we define the 3D landmarks as the usual 68 landmarks with vertices indexed as multiples of 100 in the BFM vertices. + +# 5.2. Results and Comparisons + +Self Reenactment. In this section, we first compare our method with existing SOTA methods in qualitative experiments on self reenactment task. Specifically, NeRFace [16] uses a deep MLP to fit an expression conditioned dynamic NeRF. The current SOTA method HAvatar [71] introduces 3DMM template prior and uses a deep convolutional net + +work to generate a human head NeRF represented by three planes from a mesh template with expression. Note, HAvatar leverages the GAN framework using the adversarial loss function to force the network to generate details that are not view-consistent. For a fair comparison, we remove this part and use VGG perceptual loss as Sec. 4.2 instead. + +Qualitative results on self reenactment task are shown in the Fig. 3. Our method can accurately reconstruct pixel-level high-frequency details such as beards, teeth, and hair. Besides, our method can achieve expression transfer more accurately, such as eye movements in the figure. + +Next, we conduct a quantitative evaluation for the four methods on 5 identities and 6 cameras using the evaluation split. The evaluation metrics include: Peak Signal-to-Noise Ratio (PSNR), Structure Similarity Index (SSIM), Learned Perceptual Image Patch Similarity (LPIPS) [70] and Fréchet Inception Distance (FID) [22]. Note, we calculate FID by comparing the distribution of all the training images and all the rendered images. As the task mainly focuses on the reconstruction of the head, we use face parsing to remove the body parts in the image to eliminate their impact in the experiment. As shown in Tab. 1, our method demonstrates a slight improvement in PSNR and SSIM compared with + +
MethodPSNR ↑SSIM ↑LPIPS (512) ↓LPIPS (2K) ↓FID (2K) ↓
NeRFBlendShape25.910.8360.1230.22954.80
NeRFace27.140.8490.1470.23465.11
HAVatar27.190.8830.0640.20931.06
Ours (w/o SR)27.820.8870.0800.20245.50
Ours27.700.8830.0560.09818.50
+ +![](images/4f216f90bc2dd0933614856803541ac8434a5d55705005552af1bd5e5231b262.jpg) +Figure 4. Qualitative comparisons of different methods on cross-identity reenactment task. From left to right: NeRFBlendShape [17], NeRFace [16], HAvatar [71] and Ours. Our method synthesizes high-fidelity images while ensuring the accuracy of expression transfer. + +Table 1. Quantitative evaluation results of NeRFBlendShape [17], NeRFace [16], HAvatar [71], our method without super resolution and our full method on self reenactment task. + +
MethodPSNR ↑SSIM ↑LPIPS ↓
NeRFBlendShape25.430.8120.148
NeRFace26.650.8250.151
HAVatar27.130.8800.65
Ours27.580.8820.059
+ +previous methods, and a significant improvement in LPIPS and FID, which means that our method can generate more high-frequency details. + +Cross-Identity Reenactment. We also qualitatively compare our method with the above SOTA methods on cross-identity reenactment task. As shown in the results, our method is able to synthesize higher-fidelity images with more accurate expression transfer and richer emotions. + +Novel View Synthesis. In this section, we show the results of novel view synthesis as shown in Fig. 5 shown. In this case, we use video data from 8 views for training and render the image at a new viewpoint. Next, we quantitatively evaluate the 3D consistency of our method and compare it with other SOTA methods mentioned above. Specif + +ically, we select the 5 identities as above and use the video data from 8 cameras for training while the other 8 holdout cameras for evaluation. The evaluation metrics include PSNR, SSIM and LPIPS (512 resolution). The results are shown in Tab. 2. Our method outperforms other methods in 3D consistency. + +# 5.3. Ablation Study + +Table 2. Quantitative evaluation results of the other SOTA methods and our method on 3D consistency. + +
MethodPSNR ↑SSIM ↑LPIPS ↓
FLAME-Init28.730.8750.123
Mesh-Deform28.830.8740.116
Ours28.940.8760.108
+ +Table 3. Quantitative evaluation results of the other two ablation baselines and ours on self reenactment task. + +Ablation on Initialization Strategies. In order to verify the effectiveness of our geometry-guided initialization strategy 4.3, we compare it with the strategy to use the FLAME model for initialization (FLAME-Init). Specifically, after fitting a FLAME model through multi-view data, we first subdivide the FLAME mesh 4 times and use the neutral vertices as the positions of the neutral Gaussians. Then, the + +![](images/2047bfc484c3da4a7cbc4535dea093d79bfb21ca553d99480cca68bdf49ceab6.jpg) +Figure 5. Novel view synthesis results of our method. We use 8-view synchronized videos for training the avatar. + +![](images/4c63f457e03a24d4ef3f2400d188ade0f7fa21b59cd4d716638e1f98f19bb6f2.jpg) +Figure 6. Ablation study on the initialization strategies: FLAME-initialization and our geometry-guided initialization. Our strategy ensures the hair strands away from the head are well reconstructed. + +expression deformation MLP is optimized to learn the displacement of FLAME vertices. We set the per-vertex feature to zeros, while randomly initialize the parameters of the expression color MLP. The initialization of other variables is the same as our strategy. Qualitative results are shown in the Fig. 6. Due to the lack of initialization for the hair and shoulders in FLAME-Init, the points to model these parts are offset from nearby vertices, which leads to sparseness of the Gaussians, resulting in blurring. + +Ablation on Deformation Modeling Approaches. We compare our fully learned deformation field with the previous mesh-based deformation (Mesh-Deform). Specifically, we migrate the method in INSTA [76] for controlling the NeRF deformation to our Gaussians. First we fit a 3DMM mesh template. Then, for each Gaussian point, find the closest face on the mesh, and calculate the deformation gradient to estimate the displacement. Qualitative results are shown in the Fig. 7. For some expressions that cannot be captured well by the 3DMM mesh template, our method can learn accurate deformation, thereby achieving the modeling of complex expressions. Quantitative results are shown in the Fig. 3. Our method outperforms both the two ablation baselines on PSNR, SSIM and LPIPS metrics. + +![](images/0f815ad2a74e0582c53a4a406698b5868da703d611678876dd1ca739d76d5c22.jpg) +Figure 7. Ablation study on the deformation modeling: mesh LBS-based deformation and our fully learned deformation. Our approach can learn complex and exaggerated expressions. + +# 6. Discussion and Conclusion + +Ethical Considerations. Our method is capable of creating artificial portrait videos, which have the potential to disseminate misinformation, influence public perceptions, and erode confidence in media sources. + +Limitation. For the tongue and teeth inside the mouth or long hair, blurring is sometimes produced in our method due to the lack of tracking methods. + +Conclusion. In this paper, we propose Gaussian Head Avatar, a novel representation for head avatar reconstruction, which leverages dynamic 3D Gaussians controlled by a fully learned expression deformation. Experiments demonstrate our method can synthesize ultra high-fidelity images while modeling exaggerated expressions. In addition, we propose a well-designed minute-level initialization strategy to ensure the training convergence. We believe our Gaussian Head Avatar will become the mainstream direction for head avatar reconstruction in the future. + +Acknowledgement. This paper is supported by National Key R&D Program of China (2022YFF0902200), the NSFC project No.62125107, the Beijing Municipal Science & Technology Z231100005923030. + +# References + +[1] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In European Conference on Computer Vision, pages 696-712, 2020. 3 +[2] ShahRukh Athar, Zexiang Xu, Kalyan Sunkavalli, Eli Shechtman, and Zhixin Shu. Rignerf: Fully controllable neural 3d portraits. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3 +[3] ShahRukh Athar, Zhixin Shu, and Dimitris Samaras. Flame-in-nerf: Neural control of radiance fields for free view face animation. In IEEE 17th International Conference on Automatic Face and Gesture Recognition (FG), pages 1–8, 2023. 3 +[4] Thabo Beeler, Bernd Bickel, Paul Beardsley, Bob Sumner, and Markus Gross. High-quality single-shot capture of facial geometry. ACM Trans. Graph., 29(4), 2010. 2 +[5] Sai Bi, Stephen Lombardi, Shunsuke Saito, Tomas Simon, Shih-En Wei, Kevyn Mcphail, Ravi Ramamoorthi, Yaser Sheikh, and Jason Saragih. Deep reeligible appearance models for animatable faces. ACM Trans. Graph., 40(4), 2021. 2 +[6] V Blanz and T Vetter. A morphable model for the synthesis of 3d faces. In 26th Annual Conference on Computer Graphics and Interactive Techniques (SIGGRAPH 1999), pages 187-194. ACM Press, 1999. 2 +[7] Timo Bolkart, Tianye Li, and Michael J. Black. Instant multi-view head capture through learnable registration. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 768-779, 2023. 2 +[8] Derek Bradley, Wolfgang Heidrich, Tiberiu Popa, and Alla Sheffer. High resolution passive facial performance capture. 29(4), 2010. 2 +[9] Alan Brunton, Timo Bolkart, and Stefanie Wuhrer. Multilinear wavelets: A statistical shape space for human faces. In Proceedings of the Proceedings of the European Conference on Computer Vision (ECCV), 2014. 2 +[10] Adrian Bulat and Georgios Tzimiropoulos. How far are we from solving the 2d & 3d face alignment problem? (and a dataset of 230,000 3d facial landmarks). In International Conference on Computer Vision, 2017. 6 +[11] Chen Cao, Yanlin Weng, Shun Zhou, Y. Tong, and Kun Zhou. Facewarehouse: A 3d facial expression database for visual computing. In IEEE Transactions on Visualization and Computer Graphics, pages 413-425, 2014. 2 +[12] Chen Cao, Derek Bradley, Kun Zhou, and Thabo Beeler. Real-time high-fidelity facial performance capture. ACM Trans. Graph., 34(4), 2015. 3 +[13] Chen Cao, Hongzhi Wu, Yanlin Weng, Tianjia Shao, and Kun Zhou. Real-time facial animation with image-based dynamic avatars. ACM Trans. Graph., 35(4), 2016. 3 +[14] Chen Cao, Tomas Simon, Jin Kyu Kim, Gabe Schwartz, Michael Zollhoefer, Shun-Suke Saito, Stephen Lombardi, Shih-En Wei, Danielle Belko, Shouu-I Yu, Yaser Sheikh, and Jason Saragih. Authentic volumetric avatars from a phone scan. ACM Trans. Graph., 41(4), 2022. 3 + +[15] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2019. 3 +[16] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Niessner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8645-8654, 2021. 3, 4, 6, 7 +[17] Xuan Gao, Chenglai Zhong, Jun Xiang, Yang Hong, Yudong Guo, and Juyong Zhang. Reconstructing personalized semantic facial nerf models from monocular video. ACM Transactions on Graphics (Proceedings of SIGGRAPH Asia), 41(6), 2022. 3, 6, 7 +[18] Thomas Gerig, Andreas Forster, Clemens Blumer, Bernhard Egger, Marcel Lüthi, Sandro Schonborn, and Thomas Vetter. Morphable face models - an open framework. pages 75-82, 2017. 3, 6 +[19] Abhijeet Ghosh, Graham Fyffe, Borom Tunwattanapong, Jay Busch, Xueming Yu, and Paul Debevec. Multiview face capture using polarized spherical gradient illumination. ACM Trans. Graph., 30(6):1-10, 2011. 2 +[20] Philip-William Grassal, Malte Prinzler, Titus Leistner, Carsten Rother, Matthias Nießner, and Justus Thies. Neural head avatars from monocular rgb videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18632–18643, 2022. 3 +[21] Yudong Guo, Keyu Chen, Sen Liang, Yong-Jin Liu, Hujun Bao, and Juyong Zhang. Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5764–5774, 2021. 3 +[22] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Neural Information Processing Systems, 2017. 6 +[23] Yang Hong, Bo Peng, Haiyao Xiao, Ligang Liu, and Juyong Zhang. Headnerf: A real-time nerf-based parametric head model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20374-20384, 2022. 3 +[24] Liwen Hu, Shunsuke Saito, Lingyu Wei, Koki Nagano, Jaewoo Seo, Jens Fursund, Iman Sadeghi, Carrie Sun, YenChun Chen, and Hao Li. Avatar digitization from a single image for real-time rendering. ACM Trans. Graph., 36(6), 2017. 3 +[25] Alexandru Eugen Ichim, Sofien Bouaziz, and Mark Pauly. Dynamic 3d avatar creation from hand-held video input. ACM Trans. Graph., 34(4), 2015. 3 +[26] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 2, 3, 5 +[27] Taras Khakhulin, Vanessa Sklyarova, Victor Lempitsky, and Egor Zakharov. Realistic one-shot mesh-based head avatars. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 3 + +[28] Tobias Kirschstein, Shenhan Qian, Simon Giebenhain, Tim Walter, and Matthias Nießner. Nersemble: Multi-view radiance field reconstruction of human heads. ACM Trans. Graph., 42(4), 2023. 2, 5 +[29] Georgios Kopanas, Thomas Leimkuhler, Gilles Rainer, Clément Jambon, and George Drettakis. Neural point cata-caustics for novel-view synthesis of reflections. ACM Transactions on Graphics (TOG), 41(6):1-15, 2022. 3 +[30] Christoph Lassner and Michael Zollhofer. Pulsar: Efficient sphere-based neural rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1440-1449, 2021. 3 +[31] Marc Levoy, Kari Pulli, Brian Curless, Szymon Rusinkiewicz, David Koller, Lucas Pereira, Matt Ginzon, Sean Anderson, James Davis, Jeremy Ginsberg, Jonathan Shade, and Duane Fulk. The digital michelangelo project: 3d scanning of large statues. In Proceedings of the 27th Annual Conference on Computer Graphics and Interactive Techniques, page 131-144, USA, 2000. ACM Press/Addison-Wesley Publishing Co. 2 +[32] Tianye Li, Timo Bolkart, Michael J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6), 2017. 2, 3 +[33] Tianye Li, Shichen Liu, Timo Bolkart, Jiayi Liu, Hao Li, and Yajie Zhao. Topologically consistent multi-view face inference using volumetric sampling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3824-3834, 2021. 2 +[34] Zhe Li, Zerong Zheng, Lizhen Wang, and Yebin Liu. Animatable gaussians: Learning pose-dependent gaussian maps for high-fidelity human avatar modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3 +[35] Shanchuan Lin, Andrey Ryabtsev, Soumyadip Sengupta, Brian Curless, Steve Seitz, and Ira Kemelmacher-Shlizerman. Real-time high-resolution background matting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3, 6 +[36] Xian Liu, Yinghao Xu, Qianyi Wu, Hang Zhou, Wayne Wu, and Bolei Zhou. Semantic-aware implicit neural auditory-driven video portrait generation. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 3 +[37] Stephen Lombardi, Jason Saragih, Tomas Simon, and Yaser Sheikh. Deep appearance models for face rendering. ACM Trans. Graph., 37(4):68:1-68:13, 2018. 1, 2 +[38] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural volumes: Learning dynamic renderable volumes from images. ACM Trans. Graph., 38(4):65:1-65:14, 2019. 1, 2 +[39] Stephen Lombardi, Tomas Simon, Gabriel Schwartz, Michael Zollhoefer, Yaser Sheikh, and Jason Saragih. Mixture of volumetric primitives for efficient neural rendering. ACM Trans. Graph., 40(4), 2021. 1, 2 +[40] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis, 2023. 2, 3 + +[41] Shugao Ma, Tomas Simon, Jason Saragih, Dawei Wang, Yuecheng Li, Fernando De La Torre, and Yaser Sheikh. Pixel codec avatars. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 64-73, 2021. 1, 2 +[42] Marko Mihajlovic, Aayush Bansal, Michael Zollhoefer, Siyu Tang, and Shunsuke Saito. Keypointnerf: Generalizing image-based volumetric avatars using relative spatial encoding of keypoints. In European conference on computer vision, 2022. 1, 2 +[43] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 1, 2 +[44] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting Triangular 3D Models, Materials, and Lighting From Images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8280-8290, 2022. 5 +[45] Koki Nagano, Jaewoo Seo, Jun Xing, Lingyu Wei, Zimo Li, Shunsuke Saito, Aviral Agarwal, Jens Fursund, and Hao Li. Pagan: Real-time avatars using dynamic textures. ACM Trans. Graph., 37(6), 2018. 3 +[46] Pascal Paysan, Reinhard Knothe, Brian Amberg, Sami Romdhani, and Thomas Vetter. A 3d face model for pose and illumination invariant face recognition. In 2009 Sixth IEEE International Conference on Advanced Video and Signal Based Surveillance, pages 296-301, 2009. 2 +[47] Minghan Qin, Yifan Liu, Yuelang Xu, Xiaochen Zhao, Yebin Liu, and Haoqian Wang. High-fidelity 3d head avatars reconstruction through spatially-varying expression conditioned neural radiance field. In AAAI Conference on Artificial Intelligence, 2023. 3 +[48] Amit Raj, Michael Zollhoefer, Tomas Simon, Jason Saragih, Shunsuke Saito, James Hays, and Stephen Lombardi. Pva: Pixel-aligned volumetric avatars. In arXiv:2101.02697, 2020. 1, 2 +[49] Darius Rückert, Linus Franke, and Marc Stamminger. Adop: Approximate differentiable one-pixel point rendering. ACM Trans. Graph., 41(4), 2022. 3, 4 +[50] Ruizhi Shao, Jingxiang Sun, Cheng Peng, Zerong Zheng, Boyao Zhou, Hongwen Zhang, and Yebin Liu. Control4d: Efficient 4d portrait editing with text. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 3 +[51] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 5 +[52] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2 +[53] Jingxiang Sun, Xuan Wang, Lizhen Wang, Xiaoyu Li, Yong Zhang, Hongwen Zhang, and Yebin Liu. Next3d: Generative neural texture rasterization for 3d-aware head avatars. + +In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3 +[54] Cong Wang, Di Kang, Yanpei Cao, Linchao Bao, Ying Shan, and Song-Hai Zhang. Neural point-based volumetric avatar: Surface-guided neural points for efficient and photorealistic volumetric head avatar. In ACM SIGGRAPH Asia 2023 Conference Proceedings, 2023. 1 +[55] Daoye Wang, Prashanth Chandran, Gaspard Zoss, Derek Bradley, and Paulo Gotardo. Morf: Morphable radiance fields for multiview neural head modeling. In ACM SIGGRAPH 2022 Conference Proceedings, New York, NY, USA, 2022. Association for Computing Machinery. 3 +[56] Lizhen Wang, Zhiyuan Chen, Tao Yu, Chenguang Ma, Liang Li, and Yebin Liu. Faceverse: a fine-grained and detail-controllable 3d face morphable model from a hybrid dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[57] Ziyan Wang, Timur Bagautdinov, Stephen Lombardi, Tomas Simon, Jason Saragih, Jessica Hodgins, and Michael Zollhofer. Learning compositional radiance fields of dynamic human heads. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5704-5713, 2021. 1, 2 +[58] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. SynSin: End-to-end view synthesis from a single image. In CVPR, 2020. 3 +[59] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering, 2023. 2, 3 +[60] Sijing Wu, Yichao Yan, Yunhao Li, Yuhao Cheng, Wenhan Zhu, Ke Gao, Xiaobo Li, and Guangtao Zhai. Ganhead: Towards generative animatable neural head avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 437-447, 2023. 2 +[61] Yunze Xiao, Hao Zhu, Haotian Yang, Zhengyu Diao, Xiangju Lu, and Xun Cao. Detailed facial geometry recovery from multi-view images by learning an implicit function. In Proceedings of the AAAI Conference on Artificial Intelligence, 2022. 2 +[62] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5438–5448, 2022. 3 +[63] Yuelang Xu, Lizhen Wang, Xiaochen Zhao, Hongwen Zhang, and Yebin Liu. Avatarmav: Fast 3d head avatar reconstruction using motion-aware neural voxels. In ACM SIGGRAPH 2023 Conference Proceedings, 2023. 3, 5 +[64] Yuelang Xu, Hongwen Zhang, Lizhen Wang, Xiaochen Zhao, Huang Han, Qi Guojun, and Yebin Liu. Latentavatar: Learning latent expression code for expressive neural head avatar. In ACM SIGGRAPH 2023 Conference Proceedings, 2023. 3, 4 +[65] Kai Yang, Hong Shang, Tianyang Shi, Xinghan Chen, Jingkai Zhou, Zhongqian Sun, and Wei Yang. Asm: Adaptive skinning model for high-quality 3d face modeling supplementary material. 2021. 2 + +[66] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction, 2023. 2, 3 +[67] Zeyu Yang, Hongye Yang, Zijie Pan, Xiatian Zhu, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting, 2023. 2, 3 +[68] T Yenamandra, A Tewari, F Bernard, HP Seidel, M Elgharib, D Cremers, and C Theobalt. i3dmm: Deep implicit 3d morphable model of human heads. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3 +[69] Wang Yifan, Felice Serena, Shihao Wu, Cengiz Öztireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM Transactions on Graphics (proceedings of ACM SIGGRAPH ASIA), 38(6), 2019. 3 +[70] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 5, 6 +[71] Xiaochen Zhao, Lizhen Wang, Jingxiang Sun, Hongwen Zhang, Jinli Suo, and Yebin Liu. Havatar: High-fidelity head avatar via facial model conditioned neural radiance field. ACM Trans. Graph., 2023. Just Accepted. 1, 2, 5, 6, 7 +[72] Shunyuan Zheng, Boyao Zhou, Ruizhi Shao, Boning Liu, Shengping Zhang, Liqiang Nie, and Yebin Liu. Gpsgaussian: Generalizable pixel-wise 3d gaussian splatting for real-time human novel view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3 +[73] Yufeng Zheng, Victoria Fernández Abrevaya, Marcel C. Bühler, Xu Chen, Michael J. Black, and Otmar Hilliges. I m avatar: Implicit morphable head avatars from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13535-13545, 2022. 2, 3 +[74] Yufeng Zheng, Wang Yifan, Gordon Wetzstein, Michael J. Black, and Otmar Hilliges. Pointavatar: Deformable point-based head avatars from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3 +[75] Yiyu Zhuang, Hao Zhu, Xusen Sun, and Xun Cao. Mofanerf: Morphable facial neural radiance field. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 3 +[76] Wojciech Zielonka, Timo Bolkart, and Justus Thies. Instant volumetric head avatars, 2022. 2, 3, 8 \ No newline at end of file diff --git a/2024/Gaussian Head Avatar_ Ultra High-fidelity Head Avatar via Dynamic Gaussians/images.zip b/2024/Gaussian Head Avatar_ Ultra High-fidelity Head Avatar via Dynamic Gaussians/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3b913142dfc483304fadb348dcf0ec28a98c39bf --- /dev/null +++ b/2024/Gaussian Head Avatar_ Ultra High-fidelity Head Avatar via Dynamic Gaussians/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e5d62a56c0616b0a3e8ef3c763f1b012206b20a3629a988053b6a4b4bc6050be +size 605422 diff --git a/2024/Gaussian Head Avatar_ Ultra High-fidelity Head Avatar via Dynamic Gaussians/layout.json b/2024/Gaussian Head Avatar_ Ultra High-fidelity Head Avatar via Dynamic Gaussians/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..19bb8ebbdaea1f719f50cc74dc120e7e72079573 --- /dev/null +++ b/2024/Gaussian Head Avatar_ Ultra High-fidelity Head Avatar via Dynamic Gaussians/layout.json @@ -0,0 +1,10085 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 50, + 103, + 543, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 103, + 543, + 121 + ], + "spans": [ + { + "bbox": [ + 50, + 103, + 543, + 121 + ], + "type": "text", + "content": "Gaussian Head Avatar: Ultra High-fidelity Head Avatar via Dynamic Gaussians" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 142, + 560, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 142, + 560, + 173 + ], + "spans": [ + { + "bbox": [ + 53, + 142, + 560, + 173 + ], + "type": "text", + "content": "Yuelang Xu" + }, + { + "bbox": [ + 53, + 142, + 560, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 142, + 560, + 173 + ], + "type": "text", + "content": ", Bengwang Chen" + }, + { + "bbox": [ + 53, + 142, + 560, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 142, + 560, + 173 + ], + "type": "text", + "content": ", Zhe Li" + }, + { + "bbox": [ + 53, + 142, + 560, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 142, + 560, + 173 + ], + "type": "text", + "content": ", Hongwen Zhang" + }, + { + "bbox": [ + 53, + 142, + 560, + 173 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 53, + 142, + 560, + 173 + ], + "type": "text", + "content": ", Lizhen Wang" + }, + { + "bbox": [ + 53, + 142, + 560, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 142, + 560, + 173 + ], + "type": "text", + "content": ", Zerong Zheng" + }, + { + "bbox": [ + 53, + 142, + 560, + 173 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 53, + 142, + 560, + 173 + ], + "type": "text", + "content": ", Yebin Liu" + }, + { + "bbox": [ + 53, + 142, + 560, + 173 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 53, + 142, + 560, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 142, + 560, + 173 + ], + "type": "text", + "content": " Department of Automation, Tsinghua University " + }, + { + "bbox": [ + 53, + 142, + 560, + 173 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 53, + 142, + 560, + 173 + ], + "type": "text", + "content": " Beijing Normal University " + }, + { + "bbox": [ + 53, + 142, + 560, + 173 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 53, + 142, + 560, + 173 + ], + "type": "text", + "content": " NNKosmos Technology" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 47, + 178, + 547, + 297 + ], + "blocks": [ + { + "bbox": [ + 47, + 178, + 547, + 297 + ], + "lines": [ + { + "bbox": [ + 47, + 178, + 547, + 297 + ], + "spans": [ + { + "bbox": [ + 47, + 178, + 547, + 297 + ], + "type": "image", + "image_path": "566be093fd8aedb6b1d056c06b0149e2f2c2f1bfa9da039746530b81ca4983e3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 47, + 298, + 545, + 415 + ], + "blocks": [ + { + "bbox": [ + 47, + 298, + 545, + 415 + ], + "lines": [ + { + "bbox": [ + 47, + 298, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 47, + 298, + 545, + 415 + ], + "type": "image", + "image_path": "b2f5a3e47a3df591cc0811e9b5414e04ffbb9182328730acab361413101d2e75.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 420, + 546, + 455 + ], + "lines": [ + { + "bbox": [ + 45, + 420, + 546, + 455 + ], + "spans": [ + { + "bbox": [ + 45, + 420, + 546, + 455 + ], + "type": "text", + "content": "Figure 1. Gaussian head avatar achieves ultra high-fidelity image synthesis with controllable expressions at 2K resolution. The above shows different views of the synthesized avatar, and the bottom shows different identities animated by the same expression. 16 views are used during the training." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 464, + 192, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 464, + 192, + 475 + ], + "spans": [ + { + "bbox": [ + 143, + 464, + 192, + 475 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 478, + 290, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 478, + 290, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 478, + 290, + 693 + ], + "type": "text", + "content": "Creating high-fidelity 3D head avatars has always been a research hotspot, but there remains a great challenge under lightweight sparse view setups. In this paper, we propose Gaussian Head Avatar represented by controllable 3D Gaussians for high-fidelity head avatar modeling. We optimize the neutral 3D Gaussians and a fully learned MLP-based deformation field to capture complex expressions. The two parts benefit each other, thereby our method can model fine-grained dynamic details while ensuring expression accuracy. Furthermore, we devise a well-designed geometry-guided initialization strategy based on implicit SDF and Deep Marching Tetrahedra for the stability and convergence of the training procedure. Experiments show our approach outperforms other state-of-the-art sparse-view methods, achieving ultra high-fidelity rendering quality at " + }, + { + "bbox": [ + 46, + 478, + 290, + 693 + ], + "type": "inline_equation", + "content": "2K" + }, + { + "bbox": [ + 46, + 478, + 290, + 693 + ], + "type": "text", + "content": " resolution even under exaggerated expressions. Project page: https://yuelangx.github.io/gaussianheadavatar." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 307, + 464, + 386, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 464, + 386, + 475 + ], + "spans": [ + { + "bbox": [ + 307, + 464, + 386, + 475 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 485, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 545, + 593 + ], + "type": "text", + "content": "High-fidelity 3D human head avatar modeling is of great significance in many fields, such as VR/AR, telepresence, digital human and film production. Automatically creating high-fidelity avatars has been a research hotspot in computer vision for decades. Although some traditional head avatars [37, 39, 41, 54] can realize high-fidelity animation, they typically require accurate geometries reconstructed and tracked from dense multi-view videos, thus limiting their applications in lightweight settings." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": "Thanks to the Neural Radiance Fields (NeRF) [43] which show great capability of novel view synthesis in the absence of accurate geometry, recent methods [38, 57] skip the geometry reconstruction and tracking steps but directly learn high-quality NeRF-based head avatars. Other works [42, 48, 71] have verified that NeRF can be applied to either dense or sparse views, which greatly lowers the threshold for head avatar reconstruction. However, it still remains challenging for these NeRF-based approaches to synthesize high-fidelity images at 2K resolutions with pixel" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 703, + 140, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 703, + 140, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 703, + 140, + 712 + ], + "type": "text", + "content": "†Corresponding author." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "text", + "content": "1931" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 217, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 217, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 217, + 83 + ], + "type": "text", + "content": "level details, including wrinkles and eyes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 87, + 288, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 87, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 87, + 288, + 312 + ], + "type": "text", + "content": "To overcome this bottleneck and further improve the avatar quality, we introduce 3D Gaussian splatting [26] for 3D head avatar modeling. This is an explicit discrete representation that can be well adapted to the rasterization-based rendering pipeline. It has been verified that the 3D Gaussian representation is capable of rendering complex scenes with low computational consumption. Compared to NeRF, the reconstruction quality of static and dynamic scenes [40, 59, 67] is much better while rendering time cost has been significantly reduced. Motivated by this progress, we propose Gaussian Head Avatar, a novel representation that utilizes 3D Gaussian splatting for ultra high-fidelity head avatar modeling. Although recent 4D Gaussian works [40, 59, 66, 67] have been proposed to reconstruct dynamic scenes, all of them cannot be animated. For modeling the animatable head avatar, it is crucial but still unexplored how to effectively control the deformation of 3D Gaussians and model the dynamic appearances through expression coefficients." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 316, + 288, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 288, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 288, + 579 + ], + "type": "text", + "content": "Previous explicit [74] and implicit [2, 73, 76] head avatars usually formulate the facial deformation via linear blend skinning (LBS) using the skinning weights and blendshapes like the FLAME model [32]. However, such an LBS-based formulation fails to represent exaggerated and fine-grained expressions by simple linear operations, limiting the representation ability of the head avatars. Inspired by NeRSemble [28], we propose a fully learnable expression-conditioned deformation field for the 3D head Gaussians, avoiding the limited capability of the LBS-based formulation. Specifically, we input the positions of the 3D Gaussians with expression coefficients into an MLP to directly predict the displacements from the neutral expression to the target one. Similarly, we control the motion of nonface areas, such as the neck, using the head pose as the condition. 3D Gaussian-based representation has the powerful ability to reconstruct high-frequency details, enabling our method to learn accurate deformation fields. In turn, the learned accurate deformation field facilitates the dynamic Gaussian head model to fit more dynamic details. As a result, our method is able to reconstruct finer-grained dynamic details of expressive human heads." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "text", + "content": "Unfortunately, as a discretized representation, the gradients back-propagated to the 3D Gaussians cannot spread through the whole space. Thus the convergence of training heavily relies on a plausible initialization for both the geometry and the deformation field. However, simply initializing the 3D head Gaussians with a morphable template like FLAME [32] fails to model the long hairstyle and the shoulders. Hence, we further propose an efficient and well-designed geometry-guided initialization strategy. Specifically, instead of starting from stochastic Gaussians or a FLAME model, we initially optimize an implicit signed dis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "text", + "content": "tance function (SDF) field along with a color field and a deformation MLP for modeling the basic geometry, color, and the expression-conditioned deformations of the head avatar respectively. The SDF field is converted to a mesh through Deep Marching Tetrahedra (DMTet) [52], with the color and deformation of the vertices predicted by the MLPs. Then we render the mesh and optimize them jointly under the supervision of multi-view RGB images. Finally, we use the mesh with per-vertex features from the SDF field to initialize the 3D Gaussians to lie on the basic head surface while the color and deformation MLPs are carried over to the next stage, ensuring stable training for convergence. The entire initialization process takes only around 10 minutes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 228, + 543, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 228, + 543, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 228, + 543, + 239 + ], + "type": "text", + "content": "The contributions of our method can be summarized as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 240, + 545, + 383 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 306, + 240, + 545, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 240, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 306, + 240, + 545, + 287 + ], + "type": "text", + "content": "- We propose Gaussian Head Avatar, a new head avatar representation that employs controllable dynamic 3D Gaussians to model expressive human head avatars, producing ultra high-fidelity synthesized images at 2K resolutions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 288, + 545, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 288, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 306, + 288, + 545, + 335 + ], + "type": "text", + "content": "- For modeling high-frequency dynamic details, we employ a fully learned deformation field upon the 3D head Gaussians, which accurately model extremely complex and exaggerated facial expressions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 335, + 545, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 335, + 545, + 383 + ], + "spans": [ + { + "bbox": [ + 306, + 335, + 545, + 383 + ], + "type": "text", + "content": "- We carefully design an efficient initialization strategy that leverages implicit representations to initialize the geometry and deformation, leading to efficient and robust convergence when training the Gaussian Head Avatar." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 305, + 384, + 545, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 384, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 305, + 384, + 545, + 419 + ], + "type": "text", + "content": "Benefiting from these contributions, our method surpasses recent state-of-the-art methods under lightweight sparse-view setups on the avatar quality by a large margin." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 430, + 398, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 430, + 398, + 442 + ], + "spans": [ + { + "bbox": [ + 306, + 430, + 398, + 442 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 449, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 689 + ], + "type": "text", + "content": "3D Head Avatar Reconstruction. Due to the wide application value in the film and digital human industry, 3D head avatar reconstruction from multi-view images has always been a research hotspot. Traditional works [4, 8, 19, 31] reconstruct the scan geometry through multi-view stereo and then register a face mesh template to it. However, such methods usually require heavy computation. With the utilization of deep neural networks, current methods [7, 33, 61, 65] achieve very fast reconstruction, producing even more accurate geometry. Lombardi et al. [37], Bi et al. [5] and Ma et al. [41] represent the full head mesh through a deep neural network and train it with multi-view videos as supervision. However, due to the errors in geometric estimation, mesh-based head avatars typically suffer from texture blur. Therefore, some recent methods [38, 57] utilize NeRF representation [43] to synthesize novel view images without geometry reconstruction, or build NeRF on the head mesh template [39]. Furthermore, the NeRF-based methods are extended to sparse view reconstruction tasks [28, 42, 48, 71] and achieve impressive performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Methods which focus on generative model [6, 9, 11, 32, 46, 56, 60] are dedicated to learning general mesh" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1932" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "content": "face templates from large-scale multi-view face images or 3D scans. Recently, implicit SDF-based [68] or NeRF-based [14, 23, 53, 55, 75] methods can learn full-head templates without the limitations of fixed topology, thereby better modeling complex hairstyles and glasses. Cao et al. [14] adopts a hybrid representation of local NeRF built on the mesh surface, which enables high-fidelity rendering and flexible expression control." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 168, + 288, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 168, + 288, + 287 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 288, + 287 + ], + "type": "text", + "content": "3D head avatars reconstruction from monocular videos is also a popular yet challenging research topic. Early methods [12, 13, 15, 24, 25, 45] optimize a morphable mesh to fit the training video. Recent methods [20, 27] leverage neural networks to learn non-rigid deformation upon 3DMM face templates [18, 32], thus can recover more dynamic details. Such methods are not flexible enough to handle complex topologies. Therefore, the latest methods explore to construct head avatar models based on implicit SDF [73], point clouds [74] or NeRF [2, 3, 16, 17, 21, 36, 47, 63, 64, 76]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 288, + 287, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 288, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 287, + 479 + ], + "type": "text", + "content": "Point-based Rendering. Point elements as a discrete and unstructured representation can fit geometry with arbitrary topology [69] efficiently. Recent methods [29, 30, 58] open up a differentiable rasterization pipeline, such that the point-based representation is widely used in multi-view reconstruction tasks. Aliev et al. [1] and Ruckert et al. [49] propose to first render the feature map, which is transferred to the images through a convolutional renderer. Xu et al.[62] use neural point cloud associated with neural features to model a NeRF. Recently, 3D Gaussian splatting [26] shows its superior performance, beating NeRF in both novel view synthesis quality and rendering speed. Some approaches [34, 40, 50, 59, 66, 67, 72, 72] extend Gaussian representation to dynamic scene reconstruction. However, these methods can not be migrated to the head avatar reconstruction tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 490, + 111, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 490, + 111, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 490, + 111, + 502 + ], + "type": "text", + "content": "3. Overview" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 510, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 288, + 713 + ], + "type": "text", + "content": "The pipeline of the reconstruction of Gaussian Head Avatar is illustrated in Fig. 2, including the initialization stage and the training stage of Gaussian Head Avatar. Before the beginning of the pipeline, we remove the background [35] of each image and jointly estimate the 3DMM model [18], 3D facial landmarks and the expression coefficients for each frame. In the initialization stage (Sec. 4.3), we reconstruct an SDF-based neutral geometry, and optimize a deformation MLP and a color MLP from the training data as the guidance model. Next, we extract the neutral mesh through DMTet to initialize the neutral Gaussians while the deformation and color MLPs are also inherited from the initialization stage. In the training stage of Gaussian Head Avatar (Sec. 4.2), we deform the neutral Gaussians to the target expression through the dynamic generator given the driving expression coefficients as the condition. Finally, given a camera view, the expressive Gaussians are rendered to a" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "text", + "content": "feature map, which is then fed into the convolutional super resolution network to generate high-resolution avatar images. The whole model is optimized under the supervision of multi-view RGB videos." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 133, + 362, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 133, + 362, + 146 + ], + "spans": [ + { + "bbox": [ + 306, + 133, + 362, + 146 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 154, + 434, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 154, + 434, + 167 + ], + "spans": [ + { + "bbox": [ + 306, + 154, + 434, + 167 + ], + "type": "text", + "content": "4.1. Avatar Representation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 173, + 545, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 173, + 545, + 256 + ], + "spans": [ + { + "bbox": [ + 304, + 173, + 545, + 256 + ], + "type": "text", + "content": "Generally, the static 3D Gaussians [26] with " + }, + { + "bbox": [ + 304, + 173, + 545, + 256 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 173, + 545, + 256 + ], + "type": "text", + "content": " points are represented by their positions " + }, + { + "bbox": [ + 304, + 173, + 545, + 256 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 173, + 545, + 256 + ], + "type": "text", + "content": ", the multi-channel color " + }, + { + "bbox": [ + 304, + 173, + 545, + 256 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 173, + 545, + 256 + ], + "type": "text", + "content": ", the rotation " + }, + { + "bbox": [ + 304, + 173, + 545, + 256 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 173, + 545, + 256 + ], + "type": "text", + "content": ", scale " + }, + { + "bbox": [ + 304, + 173, + 545, + 256 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 173, + 545, + 256 + ], + "type": "text", + "content": " and opacity " + }, + { + "bbox": [ + 304, + 173, + 545, + 256 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 304, + 173, + 545, + 256 + ], + "type": "text", + "content": ". The rotation " + }, + { + "bbox": [ + 304, + 173, + 545, + 256 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 173, + 545, + 256 + ], + "type": "text", + "content": " is represented in the form of quaternion. Subsequently, the Gaussians can be rasterized and rendered to the multi-channel image " + }, + { + "bbox": [ + 304, + 173, + 545, + 256 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 173, + 545, + 256 + ], + "type": "text", + "content": " given the camera parameters " + }, + { + "bbox": [ + 304, + 173, + 545, + 256 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 173, + 545, + 256 + ], + "type": "text", + "content": ". This process can be formulated as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 372, + 269, + 545, + 281 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 269, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 372, + 269, + 545, + 281 + ], + "type": "interline_equation", + "content": "I = \\mathcal {R} (X, C, Q, S, A; \\mu). \\tag {1}", + "image_path": "f04e961947dbfb1e8344eb03c8eada2f40822023df0b6441d2f7121b72533a4b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 293, + 545, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 545, + 375 + ], + "type": "text", + "content": "Our task is to reconstruct a dynamic head avatar controlled by expression coefficients. Therefore, we formulate the head avatar as dynamic 3D Gaussians conditioned on expressions. To handle the dynamic changes, we input the expression coefficients with head pose to the head avatar model and output the position and other attributes of the Gaussians as above." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 377, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 377, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 377, + 545, + 544 + ], + "type": "text", + "content": "Specifically, we first construct a canonical neutral Gaussian model with expression-independent attributes: " + }, + { + "bbox": [ + 304, + 377, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\{X_0, F_0, Q_0, S_0, A_0\\}" + }, + { + "bbox": [ + 304, + 377, + 545, + 544 + ], + "type": "text", + "content": ", which are fully estimizable. " + }, + { + "bbox": [ + 304, + 377, + 545, + 544 + ], + "type": "inline_equation", + "content": "X_0 \\in \\mathbb{R}^{N \\times 3}" + }, + { + "bbox": [ + 304, + 377, + 545, + 544 + ], + "type": "text", + "content": " denotes the positions of the Gaussians with a neutral expression in the canonical space. " + }, + { + "bbox": [ + 304, + 377, + 545, + 544 + ], + "type": "inline_equation", + "content": "F_0 \\in \\mathbb{R}^{N \\times 128}" + }, + { + "bbox": [ + 304, + 377, + 545, + 544 + ], + "type": "text", + "content": " denotes the point-wise feature vectors as their intrinsic properties. " + }, + { + "bbox": [ + 304, + 377, + 545, + 544 + ], + "type": "inline_equation", + "content": "Q_0 \\in \\mathbb{R}^{N \\times 4}" + }, + { + "bbox": [ + 304, + 377, + 545, + 544 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 377, + 545, + 544 + ], + "type": "inline_equation", + "content": "S_0 \\in \\mathbb{R}^{N \\times 3}" + }, + { + "bbox": [ + 304, + 377, + 545, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 377, + 545, + 544 + ], + "type": "inline_equation", + "content": "A_0 \\in \\mathbb{R}^{N \\times 1}" + }, + { + "bbox": [ + 304, + 377, + 545, + 544 + ], + "type": "text", + "content": " denotes the neutral rotation, scale and opacity respectively. Note that we do not define the neutral color, but directly predict expression-dependent dynamic color from the point-wise feature vectors " + }, + { + "bbox": [ + 304, + 377, + 545, + 544 + ], + "type": "inline_equation", + "content": "F_0" + }, + { + "bbox": [ + 304, + 377, + 545, + 544 + ], + "type": "text", + "content": ". Then, we construct an MLP-based expression conditioned dynamic generator " + }, + { + "bbox": [ + 304, + 377, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 304, + 377, + 545, + 544 + ], + "type": "text", + "content": " to generate all the extra dynamic changes to the neutral model. Overall, the whole Gaussian head avatar can be formulated as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 318, + 556, + 545, + 570 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 556, + 545, + 570 + ], + "spans": [ + { + "bbox": [ + 318, + 556, + 545, + 570 + ], + "type": "interline_equation", + "content": "\\{X, C, Q, S, A \\} = \\Phi \\left(\\boldsymbol {X} _ {0}, \\boldsymbol {F} _ {0}, \\boldsymbol {Q} _ {0}, \\boldsymbol {S} _ {0}, \\boldsymbol {A} _ {0}; \\theta , \\beta\\right), \\tag {2}", + "image_path": "5f830250b56200a6755a0b53dd5d29ef8ff6d6fc0c5b2da69722dadf213d8021.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 579, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 579, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 579, + 545, + 639 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 304, + 579, + 545, + 639 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 579, + 545, + 639 + ], + "type": "text", + "content": " denoting expression coefficients and " + }, + { + "bbox": [ + 304, + 579, + 545, + 639 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 579, + 545, + 639 + ], + "type": "text", + "content": " denoting the head pose. During the training, we optimize all the parameters of the dynamic generator " + }, + { + "bbox": [ + 304, + 579, + 545, + 639 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 304, + 579, + 545, + 639 + ], + "type": "text", + "content": " and the neutral Gaussian model " + }, + { + "bbox": [ + 304, + 579, + 545, + 639 + ], + "type": "inline_equation", + "content": "\\{X_0, F_0, Q_0, S_0, A_0\\}" + }, + { + "bbox": [ + 304, + 579, + 545, + 639 + ], + "type": "text", + "content": ", which are highlighted in bold in the following." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 641, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 677 + ], + "type": "text", + "content": "Next, we explain the process of adding expression-related changes to the neutral Gaussian model through the dynamic generator " + }, + { + "bbox": [ + 304, + 641, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 304, + 641, + 545, + 677 + ], + "type": "text", + "content": " as described in Eqn. 2 in detail." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Positions " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "X^{\\prime}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " of the Gaussians. Expressions bring about the geometric deformation of the neutral model, which is modeled as the displacements of the Gaussian points." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1933" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 548, + 195 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 548, + 195 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 548, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 548, + 195 + ], + "type": "image", + "image_path": "0532ff0b2d16e0d0cf8e507f5e09c0ad3a3cda6c445667a10def86811cf9672f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 203, + 547, + 248 + ], + "lines": [ + { + "bbox": [ + 46, + 203, + 547, + 248 + ], + "spans": [ + { + "bbox": [ + 46, + 203, + 547, + 248 + ], + "type": "text", + "content": "Figure 2. The overview of the Gaussian Head Avatar rendering and reconstruction. We first optimize the guidance model including a neutral mesh, a deformation MLP and a color MLP in the Initialization stage. Then we use them to initialize the neutral Gaussians and the dynamic generator. Finally, 2K RGB images are synthesized through differentiable rendering and the super-resolution network. The Gaussian Head Avatar are trained under the supervision of multi-view RGB videos." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 257, + 288, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 257, + 288, + 308 + ], + "spans": [ + { + "bbox": [ + 47, + 257, + 288, + 308 + ], + "type": "text", + "content": "Specifically, we predict the displacements respectively controlled by the expression and the head pose in the canonical space through two different MLPs: " + }, + { + "bbox": [ + 47, + 257, + 288, + 308 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{\\text{def}}^{exp} \\in \\Phi" + }, + { + "bbox": [ + 47, + 257, + 288, + 308 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 257, + 288, + 308 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{\\text{def}}^{\\text{pose}} \\in \\Phi" + }, + { + "bbox": [ + 47, + 257, + 288, + 308 + ], + "type": "text", + "content": ". Then, we add them to the neutral positions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 93, + 319, + 287, + 351 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 319, + 287, + 351 + ], + "spans": [ + { + "bbox": [ + 93, + 319, + 287, + 351 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} X ^ {\\prime} = \\boldsymbol {X} _ {0} + \\lambda_ {e x p} (\\boldsymbol {X} _ {0}) \\boldsymbol {f} _ {d e f} ^ {e x p} (\\boldsymbol {X} _ {0}, \\theta) \\tag {3} \\\\ + \\lambda_ {p o s e} (\\boldsymbol {X} _ {0}) \\boldsymbol {f} _ {d e f} ^ {p o s e} (\\boldsymbol {X} _ {0}, \\beta). \\\\ \\end{array}", + "image_path": "fb9b7db0bfff9caf166c9b9b260b4e9cc7d51d034facdcfd5c2bb00217a253cc.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 361, + 287, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 287, + 531 + ], + "type": "inline_equation", + "content": "\\lambda_{exp}(\\cdot)" + }, + { + "bbox": [ + 46, + 361, + 287, + 531 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 361, + 287, + 531 + ], + "type": "inline_equation", + "content": "\\lambda_{pose}(\\cdot)" + }, + { + "bbox": [ + 46, + 361, + 287, + 531 + ], + "type": "text", + "content": " represent the extent to which the points are affected by the expression or the head pose respectively. Without decoupling but just using the expression coefficients as the global condition [16] which also controls the shoulders and upper body, will produce jittering results during the animation. Here, we assume that the Gaussian points closer to 3D landmarks are more affected by the expression coefficients and less affected by the head pose, while the opposite is true for the Gaussian points far away. Specifically, The 3D landmarks " + }, + { + "bbox": [ + 46, + 361, + 287, + 531 + ], + "type": "inline_equation", + "content": "P_0" + }, + { + "bbox": [ + 46, + 361, + 287, + 531 + ], + "type": "text", + "content": " of the canonical model are first estimated through the 3DMM model in the data preprocessing and then optimized in the initialization stage 4.3. Then for each Gaussian point, we calculate the above weight " + }, + { + "bbox": [ + 46, + 361, + 287, + 531 + ], + "type": "inline_equation", + "content": "\\lambda_{exp}(\\cdot)" + }, + { + "bbox": [ + 46, + 361, + 287, + 531 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 361, + 287, + 531 + ], + "type": "inline_equation", + "content": "\\lambda_{pose}(\\cdot)" + }, + { + "bbox": [ + 46, + 361, + 287, + 531 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 541, + 271, + 581 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 541, + 271, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 541, + 271, + 581 + ], + "type": "interline_equation", + "content": "\\lambda_ {e x p} (x) = \\left\\{ \\begin{array}{l l} 1, & d i s t (x, \\boldsymbol {P} _ {0}) < t _ {1} \\\\ \\frac {t _ {2} - d i s t (x , \\boldsymbol {P} _ {0})}{t _ {2} - t _ {1}}, & d i s t (x, \\boldsymbol {P} _ {0}) \\in [ t _ {1}, t _ {2} ] \\\\ 0, & d i s t (x, \\boldsymbol {P} _ {0}) > t _ {2} \\end{array} \\right.", + "image_path": "870bf1d03ecc685e91b893bf39b90f7858e335a09078ca3baf109ece57c17b6a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 592, + 287, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 592, + 287, + 663 + ], + "spans": [ + { + "bbox": [ + 47, + 592, + 287, + 663 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 47, + 592, + 287, + 663 + ], + "type": "inline_equation", + "content": "\\lambda_{\\text{pose}}(x) = 1 - \\lambda_{\\text{exp}}(x)" + }, + { + "bbox": [ + 47, + 592, + 287, + 663 + ], + "type": "text", + "content": ". And " + }, + { + "bbox": [ + 47, + 592, + 287, + 663 + ], + "type": "inline_equation", + "content": "x \\in \\mathbf{X}_0" + }, + { + "bbox": [ + 47, + 592, + 287, + 663 + ], + "type": "text", + "content": " denotes the position of one neutral Gaussian. " + }, + { + "bbox": [ + 47, + 592, + 287, + 663 + ], + "type": "inline_equation", + "content": "\\text{dist}(x, P_0)" + }, + { + "bbox": [ + 47, + 592, + 287, + 663 + ], + "type": "text", + "content": " denotes the minimum distance from the point " + }, + { + "bbox": [ + 47, + 592, + 287, + 663 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 47, + 592, + 287, + 663 + ], + "type": "text", + "content": " to the 3D landmarks " + }, + { + "bbox": [ + 47, + 592, + 287, + 663 + ], + "type": "inline_equation", + "content": "P_0" + }, + { + "bbox": [ + 47, + 592, + 287, + 663 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 592, + 287, + 663 + ], + "type": "inline_equation", + "content": "t_1 = 0.15" + }, + { + "bbox": [ + 47, + 592, + 287, + 663 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 592, + 287, + 663 + ], + "type": "inline_equation", + "content": "t_2 = 0.25" + }, + { + "bbox": [ + 47, + 592, + 287, + 663 + ], + "type": "text", + "content": " are predefined hyperparameters when the length of the head is set to approximately 1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "content": "Color " + }, + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "C'" + }, + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "content": " of the Gaussians. Modeling the dynamic details typically requires dynamic color that changes with expressions. As we do not pre-define the neutral value in Eqn. 2, the color are directly predict by two color MLPs:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 256, + 414, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 256, + 414, + 270 + ], + "spans": [ + { + "bbox": [ + 306, + 256, + 414, + 270 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{col}^{exp}\\in \\Phi" + }, + { + "bbox": [ + 306, + 256, + 414, + 270 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 256, + 414, + 270 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{col}^{pose}\\in \\Phi" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 367, + 277, + 545, + 306 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 277, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 367, + 277, + 545, + 306 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} C ^ {\\prime} = \\lambda_ {e x p} \\left(\\boldsymbol {X} _ {0}\\right) \\boldsymbol {f} _ {c o l} ^ {e x p} \\left(\\boldsymbol {F} _ {0}, \\theta\\right) \\tag {4} \\\\ + \\lambda_ {p o s e} (\\boldsymbol {X} _ {0}) \\boldsymbol {f} _ {c o l} ^ {p o s e} (\\boldsymbol {F} _ {0}, \\beta). \\\\ \\end{array}", + "image_path": "1374c3e826e8493cd273c33256e1ce6347bae09f9e882b1032cf44b25374bb76.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 312, + 545, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 312, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 305, + 312, + 545, + 374 + ], + "type": "text", + "content": "Rotation, Scale and Opacity " + }, + { + "bbox": [ + 305, + 312, + 545, + 374 + ], + "type": "inline_equation", + "content": "\\{Q', S', A'\\}" + }, + { + "bbox": [ + 305, + 312, + 545, + 374 + ], + "type": "text", + "content": " of the Gaussians. These three attributes also dynamic, thereby modeling some detailed expressions-related appearance changes. Here, we just use another two attribute MLPs " + }, + { + "bbox": [ + 305, + 312, + 545, + 374 + ], + "type": "inline_equation", + "content": "f_{att}^{exp} \\in \\Phi" + }, + { + "bbox": [ + 305, + 312, + 545, + 374 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 312, + 545, + 374 + ], + "type": "inline_equation", + "content": "f_{att}^{pose} \\in \\Phi" + }, + { + "bbox": [ + 305, + 312, + 545, + 374 + ], + "type": "text", + "content": " to predict their shift from the neutral value." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 365, + 380, + 545, + 425 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 380, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 365, + 380, + 545, + 425 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\{Q ^ {\\prime}, S ^ {\\prime}, A ^ {\\prime} \\} = \\left\\{\\mathbf {Q} _ {0}, \\mathbf {S} _ {0}, \\mathbf {A} _ {0} \\right\\} \\\\ + \\lambda_ {e x p} (\\boldsymbol {X} _ {0}) \\boldsymbol {f} _ {a t t} ^ {e x p} \\left(\\boldsymbol {F} _ {0}, \\theta\\right) \\tag {5} \\\\ + \\lambda_ {p o s e} (\\boldsymbol {X} _ {0}) \\boldsymbol {f} _ {a t t} ^ {p o s e} (\\boldsymbol {F} _ {0}, \\beta). \\\\ \\end{array}", + "image_path": "c4108f142f26460aa3233b1e059999fcdcb5534d9cf670845c7075e78fc7a1c4.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 432, + 545, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 432, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 304, + 432, + 545, + 503 + ], + "type": "text", + "content": "Finally, we apply rigid rotations and translations " + }, + { + "bbox": [ + 304, + 432, + 545, + 503 + ], + "type": "inline_equation", + "content": "T(\\cdot)" + }, + { + "bbox": [ + 304, + 432, + 545, + 503 + ], + "type": "text", + "content": " to the Gaussians, transforming them from the canonical space to the world space. Note, the transformation is only implemented for directional variables: " + }, + { + "bbox": [ + 304, + 432, + 545, + 503 + ], + "type": "inline_equation", + "content": "\\{X', Q'\\}" + }, + { + "bbox": [ + 304, + 432, + 545, + 503 + ], + "type": "text", + "content": ", while the multichannel color, the scale and the opacity " + }, + { + "bbox": [ + 304, + 432, + 545, + 503 + ], + "type": "inline_equation", + "content": "\\{C', S', A'\\}" + }, + { + "bbox": [ + 304, + 432, + 545, + 503 + ], + "type": "text", + "content": " are not directional thus remain unchanged." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 370, + 511, + 545, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 511, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 370, + 511, + 545, + 525 + ], + "type": "interline_equation", + "content": "\\{X, Q \\} = T \\left(\\left\\{X ^ {\\prime}, Q ^ {\\prime} \\right\\}, \\beta\\right), \\tag {6}", + "image_path": "d16bbdbfef64ef5a6e8b42985ec6ae382e7f87acd369fe465114b99841305909.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 372, + 532, + 545, + 545 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 532, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 372, + 532, + 545, + 545 + ], + "type": "interline_equation", + "content": "\\{C, S, A \\} = \\left\\{C ^ {\\prime}, S ^ {\\prime}, A ^ {\\prime} \\right\\}. \\tag {7}", + "image_path": "02da77dd910a178111f49a3e7d0662d571e219318c8afe6914f15a34ef470b59.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 552, + 370, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 552, + 370, + 564 + ], + "spans": [ + { + "bbox": [ + 306, + 552, + 370, + 564 + ], + "type": "text", + "content": "4.2. Training" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "text", + "content": "In this part, we explain the training pipeline of the Gaussian head avatar 4.1 and the loss function. In each iteration, we first generate the expression conditioned 3D Gaussians as Eqn. 2. Then we render a 32-channel image with 512 resolution " + }, + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "inline_equation", + "content": "I_C \\in \\mathbb{R}^{512 \\times 512 \\times 32}" + }, + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "text", + "content": " referring to Eqn. 1. After that we feed the image to a super resolution network " + }, + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "text", + "content": " to generate a 2048 resolution RGB image " + }, + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "inline_equation", + "content": "I_{hr} \\in \\mathbb{R}^{2048 \\times 2048 \\times 3}" + }, + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "text", + "content": ", such that more details are recovered and noise caused by uneven ambient light or camera chromatic aberration in the training data will be filtered out [49, 64]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "During training, we jointly optimize all the learnable parameters mentioned above in bold, including the neutral" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1934" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 178 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 178 + ], + "type": "text", + "content": "Gaussians: " + }, + { + "bbox": [ + 47, + 72, + 287, + 178 + ], + "type": "inline_equation", + "content": "\\{X_0, F_0, Q_0, S_0, A_0\\}" + }, + { + "bbox": [ + 47, + 72, + 287, + 178 + ], + "type": "text", + "content": ", the dynamic generator: " + }, + { + "bbox": [ + 47, + 72, + 287, + 178 + ], + "type": "inline_equation", + "content": "\\{f_{col}^{exp}, f_{col}^{pose}, f_{def}^{exp}, f_{def}^{pose}, f_{att}^{exp}, f_{att}^{pose}\\}" + }, + { + "bbox": [ + 47, + 72, + 287, + 178 + ], + "type": "text", + "content": ", and the super resolution network " + }, + { + "bbox": [ + 47, + 72, + 287, + 178 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 47, + 72, + 287, + 178 + ], + "type": "text", + "content": ". For the loss function, we only use the foreground RGB images " + }, + { + "bbox": [ + 47, + 72, + 287, + 178 + ], + "type": "inline_equation", + "content": "I_{gt}" + }, + { + "bbox": [ + 47, + 72, + 287, + 178 + ], + "type": "text", + "content": " as supervision to construct an L1 loss and a VGG perceptual loss [70] between the generated images " + }, + { + "bbox": [ + 47, + 72, + 287, + 178 + ], + "type": "inline_equation", + "content": "I_{hr}" + }, + { + "bbox": [ + 47, + 72, + 287, + 178 + ], + "type": "text", + "content": " and the ground truth " + }, + { + "bbox": [ + 47, + 72, + 287, + 178 + ], + "type": "inline_equation", + "content": "I_{gt}" + }, + { + "bbox": [ + 47, + 72, + 287, + 178 + ], + "type": "text", + "content": ". Besides, we encourage the first three channels of the 32-channel feature image " + }, + { + "bbox": [ + 47, + 72, + 287, + 178 + ], + "type": "inline_equation", + "content": "I_C" + }, + { + "bbox": [ + 47, + 72, + 287, + 178 + ], + "type": "text", + "content": " to be RGB channels, which is ensured by a L1 loss term. The total loss is:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 82, + 185, + 286, + 213 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 185, + 286, + 213 + ], + "spans": [ + { + "bbox": [ + 82, + 185, + 286, + 213 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} = \\left\\| I _ {h r} - I _ {g t} \\right\\| _ {1} + \\lambda_ {v g g} V G G \\left(I _ {h r}, I _ {g t}\\right) \\tag {8} \\\\ + \\lambda_ {l r} | | I _ {l r} - I _ {g t} | | _ {1}, \\\\ \\end{array}", + "image_path": "4bed9fa288f851546758164808caa80984f7d333fdd44e846274abb75e414c13.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 220, + 287, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 220, + 287, + 245 + ], + "spans": [ + { + "bbox": [ + 47, + 220, + 287, + 245 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 47, + 220, + 287, + 245 + ], + "type": "inline_equation", + "content": "I_{lr}" + }, + { + "bbox": [ + 47, + 220, + 287, + 245 + ], + "type": "text", + "content": " denoting the first three channels of the 32-channel image " + }, + { + "bbox": [ + 47, + 220, + 287, + 245 + ], + "type": "inline_equation", + "content": "I_C" + }, + { + "bbox": [ + 47, + 220, + 287, + 245 + ], + "type": "text", + "content": ". We set the weights " + }, + { + "bbox": [ + 47, + 220, + 287, + 245 + ], + "type": "inline_equation", + "content": "\\lambda_{vgg} = 0.1" + }, + { + "bbox": [ + 47, + 220, + 287, + 245 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 220, + 287, + 245 + ], + "type": "inline_equation", + "content": "\\lambda_{lr} = 1" + }, + { + "bbox": [ + 47, + 220, + 287, + 245 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 251, + 216, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 251, + 216, + 262 + ], + "spans": [ + { + "bbox": [ + 47, + 251, + 216, + 262 + ], + "type": "text", + "content": "4.3. Geometry-guided Initialization" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 269, + 287, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 269, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 47, + 269, + 287, + 352 + ], + "type": "text", + "content": "Unlike neural networks, the Gaussians act as an unordered and unstructured representation. Random initialization leads to failure to converge while naively using the FLAME model to initialize will significantly reduce the reconstruction quality. In this section, we describe in detail how to optimize a mesh guidance model to provide reliable initialization for the Gaussians in Sec. 4.1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 353, + 287, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 353, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 47, + 353, + 287, + 411 + ], + "type": "text", + "content": "Mesh Guidance Model. Specifically, we first construct a MLP " + }, + { + "bbox": [ + 47, + 353, + 287, + 411 + ], + "type": "inline_equation", + "content": "f_{sdf}" + }, + { + "bbox": [ + 47, + 353, + 287, + 411 + ], + "type": "text", + "content": " to represent a signed distance field. In addition, this network will also output the corresponding feature vector of each point, which is used for predicting the point color. It can be formulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 421, + 287, + 434 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 421, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 134, + 421, + 287, + 434 + ], + "type": "interline_equation", + "content": "s, \\eta = \\boldsymbol {f} _ {s d f} (x), \\tag {9}", + "image_path": "819cc314b96191ddcfc3dabe656519d14f493a8a54990ac21b65559423fa3820.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 441, + 287, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 441, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 47, + 441, + 287, + 620 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 47, + 441, + 287, + 620 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 47, + 441, + 287, + 620 + ], + "type": "text", + "content": " denotes the SDF value, " + }, + { + "bbox": [ + 47, + 441, + 287, + 620 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 47, + 441, + 287, + 620 + ], + "type": "text", + "content": " denotes the feature vector and " + }, + { + "bbox": [ + 47, + 441, + 287, + 620 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 47, + 441, + 287, + 620 + ], + "type": "text", + "content": " denotes the point position. Then through (DMTet) [51], we can differentially extract a mesh with vertices " + }, + { + "bbox": [ + 47, + 441, + 287, + 620 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 47, + 441, + 287, + 620 + ], + "type": "text", + "content": ", per-vertex feature vectors " + }, + { + "bbox": [ + 47, + 441, + 287, + 620 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 47, + 441, + 287, + 620 + ], + "type": "text", + "content": " and its faces. We also predict the per-vertex 32-channel color as Eqn. 4 by the two color MLPs " + }, + { + "bbox": [ + 47, + 441, + 287, + 620 + ], + "type": "inline_equation", + "content": "f_{col}^{exp}" + }, + { + "bbox": [ + 47, + 441, + 287, + 620 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 441, + 287, + 620 + ], + "type": "inline_equation", + "content": "f_{col}^{pose}" + }, + { + "bbox": [ + 47, + 441, + 287, + 620 + ], + "type": "text", + "content": ". In parallel, we construct the two deformation MLPs: " + }, + { + "bbox": [ + 47, + 441, + 287, + 620 + ], + "type": "inline_equation", + "content": "f_{def}^{exp}" + }, + { + "bbox": [ + 47, + 441, + 287, + 620 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 441, + 287, + 620 + ], + "type": "inline_equation", + "content": "f_{def}^{pose}" + }, + { + "bbox": [ + 47, + 441, + 287, + 620 + ], + "type": "text", + "content": " as described in Sec. 4.1 to predict the displacements and add them to the vertex positions. This process is similar to Eqn 3 above, with the Gaussian positions " + }, + { + "bbox": [ + 47, + 441, + 287, + 620 + ], + "type": "inline_equation", + "content": "X_0" + }, + { + "bbox": [ + 47, + 441, + 287, + 620 + ], + "type": "text", + "content": " replaced by the vertex positions " + }, + { + "bbox": [ + 47, + 441, + 287, + 620 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 47, + 441, + 287, + 620 + ], + "type": "text", + "content": ". Finally we also apply rigid rotations and translations to the deformed mesh, transforming it to the world space and render the deformed mesh into an image " + }, + { + "bbox": [ + 47, + 441, + 287, + 620 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 47, + 441, + 287, + 620 + ], + "type": "text", + "content": " and a mask " + }, + { + "bbox": [ + 47, + 441, + 287, + 620 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 441, + 287, + 620 + ], + "type": "text", + "content": " through differentiable rasterization [44] according to the camera parameters " + }, + { + "bbox": [ + 47, + 441, + 287, + 620 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 47, + 441, + 287, + 620 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 620, + 287, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 620, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 47, + 620, + 287, + 654 + ], + "type": "text", + "content": "Loss Function and Training. Next, we can construct the RGB loss and the silhouette loss to train the guidance model:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 114, + 655, + 287, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 655, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 114, + 655, + 287, + 669 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {R G B} = \\left| \\left| I _ {r, g, b} - I _ {g t} \\right| \\right| _ {1}, \\tag {10}", + "image_path": "61c3893b418818611f5a7cd0f47693f27d7e0a11233630551517465e033f561d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 119, + 672, + 286, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 672, + 286, + 685 + ], + "spans": [ + { + "bbox": [ + 119, + 672, + 286, + 685 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s i l} = I O U (M, M _ {g t}), \\tag {11}", + "image_path": "acede5ff316031b35981af921d34be498abdd1315296d13d4eee875b7a7b9702.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "I_{gt}" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "M_{gt}" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": " denote the ground truth RGB image and mask, respectively. " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "IOU(\\cdot)" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": " denotes Intersection over" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 108 + ], + "type": "text", + "content": "Union metrics. Note that only the first three channels " + }, + { + "bbox": [ + 306, + 72, + 545, + 108 + ], + "type": "inline_equation", + "content": "R, G, B" + }, + { + "bbox": [ + 306, + 72, + 545, + 108 + ], + "type": "text", + "content": " of the 32-channel image " + }, + { + "bbox": [ + 306, + 72, + 545, + 108 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 306, + 72, + 545, + 108 + ], + "type": "text", + "content": " are supervised by the ground truth RGB images." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 108, + 545, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 108, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 305, + 108, + 545, + 167 + ], + "type": "text", + "content": "We also use the estimated 3D facial landmarks " + }, + { + "bbox": [ + 305, + 108, + 545, + 167 + ], + "type": "inline_equation", + "content": "P_{gt}" + }, + { + "bbox": [ + 305, + 108, + 545, + 167 + ], + "type": "text", + "content": " as described in Sec. 3 to provide rough guidance for the expression deformation MLP. Specifically, we input the neutral 3D landmarks " + }, + { + "bbox": [ + 305, + 108, + 545, + 167 + ], + "type": "inline_equation", + "content": "P_0" + }, + { + "bbox": [ + 305, + 108, + 545, + 167 + ], + "type": "text", + "content": " into the expression deformation MLP to predict the expression conditioned landmarks " + }, + { + "bbox": [ + 305, + 108, + 545, + 167 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 305, + 108, + 545, + 167 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 375, + 176, + 545, + 190 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 176, + 545, + 190 + ], + "spans": [ + { + "bbox": [ + 375, + 176, + 545, + 190 + ], + "type": "interline_equation", + "content": "P = \\boldsymbol {P} _ {0} + \\boldsymbol {f} _ {\\text {d e f}} ^ {\\exp} (\\boldsymbol {P} _ {0}, \\theta). \\tag {12}", + "image_path": "fdea1bba759f212b743a9781098f435d77b86fe4fb6dca03e900e15ed5b22ba6.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 198, + 545, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 198, + 545, + 222 + ], + "spans": [ + { + "bbox": [ + 305, + 198, + 545, + 222 + ], + "type": "text", + "content": "Then we construct the loss function with 3D facial landmarks " + }, + { + "bbox": [ + 305, + 198, + 545, + 222 + ], + "type": "inline_equation", + "content": "P_{gt}" + }, + { + "bbox": [ + 305, + 198, + 545, + 222 + ], + "type": "text", + "content": " as the supervision:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 381, + 231, + 545, + 244 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 231, + 545, + 244 + ], + "spans": [ + { + "bbox": [ + 381, + 231, + 545, + 244 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d e f}} = \\left| \\left| P - P _ {g t} \\right| \\right| _ {2}, \\tag {13}", + "image_path": "7124bf560650fa31be5f6f527a2a0dd141e88758b9f50b46836b870b57d5e1b1.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 252, + 545, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 252, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 305, + 252, + 545, + 358 + ], + "type": "text", + "content": "Besides, we introduce three constraints: (1) a regular term " + }, + { + "bbox": [ + 305, + 252, + 545, + 358 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{offset}" + }, + { + "bbox": [ + 305, + 252, + 545, + 358 + ], + "type": "text", + "content": " to punish all non-zero displacements to prevent the two deformation MLPs from learning a global constant offset [63], (2) a regular term " + }, + { + "bbox": [ + 305, + 252, + 545, + 358 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{lmk}" + }, + { + "bbox": [ + 305, + 252, + 545, + 358 + ], + "type": "text", + "content": " to limit the SDF value at the 3D landmarks to be close to zero, such that the landmarks are located on the surface of the mesh, (3) a Laplacian term " + }, + { + "bbox": [ + 305, + 252, + 545, + 358 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{lap}" + }, + { + "bbox": [ + 305, + 252, + 545, + 358 + ], + "type": "text", + "content": " for maintaining the extracted mesh smooth to a certain extent. Overall, the total loss function is formulated as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 342, + 366, + 545, + 395 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 366, + 545, + 395 + ], + "spans": [ + { + "bbox": [ + 342, + 366, + 545, + 395 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} = \\mathcal {L} _ {R G B} + \\lambda_ {s i l} \\mathcal {L} _ {s i l} + \\lambda_ {d e f} \\mathcal {L} _ {d e f} + \\tag {14} \\\\ \\lambda_ {o f f s e t} \\mathcal {L} _ {o f f s e t} + \\lambda_ {l m k} \\mathcal {L} _ {l m k} + \\lambda_ {l a p} \\mathcal {L} _ {l a p}, \\\\ \\end{array}", + "image_path": "e6ce691f8a88689c39f76d52fba0de577dbaf558e16e1055b3a9f1c7046c6519.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 402, + 545, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 402, + 545, + 461 + ], + "spans": [ + { + "bbox": [ + 305, + 402, + 545, + 461 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 305, + 402, + 545, + 461 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 305, + 402, + 545, + 461 + ], + "type": "text", + "content": " denoting the weights of each term, which are set as follows: " + }, + { + "bbox": [ + 305, + 402, + 545, + 461 + ], + "type": "inline_equation", + "content": "\\lambda_{sil} = 0.1" + }, + { + "bbox": [ + 305, + 402, + 545, + 461 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 402, + 545, + 461 + ], + "type": "inline_equation", + "content": "\\lambda_{def} = 1" + }, + { + "bbox": [ + 305, + 402, + 545, + 461 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 402, + 545, + 461 + ], + "type": "inline_equation", + "content": "\\lambda_{offset} = 0.01" + }, + { + "bbox": [ + 305, + 402, + 545, + 461 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 402, + 545, + 461 + ], + "type": "inline_equation", + "content": "\\lambda_{lmk} = 0.1" + }, + { + "bbox": [ + 305, + 402, + 545, + 461 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 402, + 545, + 461 + ], + "type": "inline_equation", + "content": "\\lambda_{lap} = 100" + }, + { + "bbox": [ + 305, + 402, + 545, + 461 + ], + "type": "text", + "content": ". We jointly optimize the MLPs mentioned above with the neutral 3D landmarks " + }, + { + "bbox": [ + 305, + 402, + 545, + 461 + ], + "type": "inline_equation", + "content": "P_0" + }, + { + "bbox": [ + 305, + 402, + 545, + 461 + ], + "type": "text", + "content": " jointly until all MLPs are converged." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 462, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 462, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 305, + 462, + 545, + 617 + ], + "type": "text", + "content": "Parameters Transfer. Finally, we use the roughly trained mesh guidance model to initialize the Gaussian model. Specifically, we extract the neutral mesh with vertices " + }, + { + "bbox": [ + 305, + 462, + 545, + 617 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 305, + 462, + 545, + 617 + ], + "type": "text", + "content": " and per-vertex features " + }, + { + "bbox": [ + 305, + 462, + 545, + 617 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 305, + 462, + 545, + 617 + ], + "type": "text", + "content": " through DMTet, and directly assign their values to the neutral positions " + }, + { + "bbox": [ + 305, + 462, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_0 = X" + }, + { + "bbox": [ + 305, + 462, + 545, + 617 + ], + "type": "text", + "content": " and the per-vertex feature vectors " + }, + { + "bbox": [ + 305, + 462, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\pmb{F}_{0} = F" + }, + { + "bbox": [ + 305, + 462, + 545, + 617 + ], + "type": "text", + "content": " of the neutral Gaussians respectively. Then, we retain all the four optimized MLPs: " + }, + { + "bbox": [ + 305, + 462, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\{\\pmb{f}_{col}^{exp},\\pmb{f}_{col}^{pose},\\pmb{f}_{def}^{exp},\\pmb{f}_{def}^{pose}\\}" + }, + { + "bbox": [ + 305, + 462, + 545, + 617 + ], + "type": "text", + "content": " for the Gaussian model. For the other neutral attributes: rotation, scale and opacity, we adopt the original initialization strategy in Gaussian Splatting [26]. And the parameters of the two attribute MLPs: " + }, + { + "bbox": [ + 305, + 462, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\{\\pmb{f}_{att}^{exp},\\pmb{f}_{att}^{pose}\\}" + }, + { + "bbox": [ + 305, + 462, + 545, + 617 + ], + "type": "text", + "content": " and the super resolution network " + }, + { + "bbox": [ + 305, + 462, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 305, + 462, + 545, + 617 + ], + "type": "text", + "content": " are just randomly initialized." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 628, + 387, + 641 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 628, + 387, + 641 + ], + "spans": [ + { + "bbox": [ + 306, + 628, + 387, + 641 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 647, + 438, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 438, + 659 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 438, + 659 + ], + "type": "text", + "content": "5.1. Implementation Details" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 665, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 545, + 712 + ], + "type": "text", + "content": "In the experiment we use 12 sets of data, 10 of which are from NeRSemble [28], and the other 2 are multi-view video data from HAvatar [71]. For the 10 identities from NeRSemble, each set contains 2500 to 3000 frames, 16" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1935" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 70, + 145, + 219 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 145, + 219 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 145, + 219 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 145, + 219 + ], + "type": "image", + "image_path": "3fd16cc2aa69a95b28977d998402bb72b458d96f03b5396d6aa3fb0b88048e03.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 148, + 70, + 245, + 218 + ], + "blocks": [ + { + "bbox": [ + 148, + 70, + 245, + 218 + ], + "lines": [ + { + "bbox": [ + 148, + 70, + 245, + 218 + ], + "spans": [ + { + "bbox": [ + 148, + 70, + 245, + 218 + ], + "type": "image", + "image_path": "d39d1453b17767fac298dc00c9fad722b9368a291dac9f348ede15a1e7f2152b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 248, + 70, + 345, + 218 + ], + "blocks": [ + { + "bbox": [ + 248, + 70, + 345, + 218 + ], + "lines": [ + { + "bbox": [ + 248, + 70, + 345, + 218 + ], + "spans": [ + { + "bbox": [ + 248, + 70, + 345, + 218 + ], + "type": "image", + "image_path": "7e5bd01e6ccb0454dd029612d2095b570267073250ef5ee364e2fe47f4247048.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 349, + 70, + 446, + 218 + ], + "blocks": [ + { + "bbox": [ + 349, + 70, + 446, + 218 + ], + "lines": [ + { + "bbox": [ + 349, + 70, + 446, + 218 + ], + "spans": [ + { + "bbox": [ + 349, + 70, + 446, + 218 + ], + "type": "image", + "image_path": "5d0caf3d13100366507cdd69ab838a29ee14101dcdd5b7434cc6576462b47711.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 450, + 71, + 547, + 218 + ], + "blocks": [ + { + "bbox": [ + 450, + 71, + 547, + 218 + ], + "lines": [ + { + "bbox": [ + 450, + 71, + 547, + 218 + ], + "spans": [ + { + "bbox": [ + 450, + 71, + 547, + 218 + ], + "type": "image", + "image_path": "4914b8a2be00e894b7950b90646037287f0cc87996450f4c7795723bc8bdbd91.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 48, + 222, + 144, + 370 + ], + "blocks": [ + { + "bbox": [ + 48, + 222, + 144, + 370 + ], + "lines": [ + { + "bbox": [ + 48, + 222, + 144, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 222, + 144, + 370 + ], + "type": "image", + "image_path": "68181e56aaf028fcb2eb3d7412505588663d0aef29deda2ec62fd1d9bb1ca8f1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 371, + 127, + 380 + ], + "lines": [ + { + "bbox": [ + 66, + 371, + 127, + 380 + ], + "spans": [ + { + "bbox": [ + 66, + 371, + 127, + 380 + ], + "type": "text", + "content": "NeRFBlendShape" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 149, + 223, + 245, + 370 + ], + "blocks": [ + { + "bbox": [ + 149, + 223, + 245, + 370 + ], + "lines": [ + { + "bbox": [ + 149, + 223, + 245, + 370 + ], + "spans": [ + { + "bbox": [ + 149, + 223, + 245, + 370 + ], + "type": "image", + "image_path": "9031d2608924608e6f8a59c32319ffa9b920451b6a89aa6c8b80b83bbe80c704.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 371, + 212, + 379 + ], + "lines": [ + { + "bbox": [ + 181, + 371, + 212, + 379 + ], + "spans": [ + { + "bbox": [ + 181, + 371, + 212, + 379 + ], + "type": "text", + "content": "NeRFace" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 249, + 223, + 345, + 370 + ], + "blocks": [ + { + "bbox": [ + 249, + 223, + 345, + 370 + ], + "lines": [ + { + "bbox": [ + 249, + 223, + 345, + 370 + ], + "spans": [ + { + "bbox": [ + 249, + 223, + 345, + 370 + ], + "type": "image", + "image_path": "217d97f9e2a3cd03421d2449f28c231699851035f5b7b16b193efb76a9adcfed.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 282, + 372, + 313, + 379 + ], + "lines": [ + { + "bbox": [ + 282, + 372, + 313, + 379 + ], + "spans": [ + { + "bbox": [ + 282, + 372, + 313, + 379 + ], + "type": "text", + "content": "HAvatar" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 350, + 223, + 445, + 370 + ], + "blocks": [ + { + "bbox": [ + 350, + 223, + 445, + 370 + ], + "lines": [ + { + "bbox": [ + 350, + 223, + 445, + 370 + ], + "spans": [ + { + "bbox": [ + 350, + 223, + 445, + 370 + ], + "type": "image", + "image_path": "275d43f84f8d006fb2ae6bdcd7430565282943ed3ee96c0a253a7dc20035ac9e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 389, + 372, + 407, + 379 + ], + "lines": [ + { + "bbox": [ + 389, + 372, + 407, + 379 + ], + "spans": [ + { + "bbox": [ + 389, + 372, + 407, + 379 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 450, + 220, + 546, + 371 + ], + "blocks": [ + { + "bbox": [ + 450, + 220, + 546, + 371 + ], + "lines": [ + { + "bbox": [ + 450, + 220, + 546, + 371 + ], + "spans": [ + { + "bbox": [ + 450, + 220, + 546, + 371 + ], + "type": "image", + "image_path": "23edde6edb66fc22bb344622df640093796f068abfda61707cac7b45056c0c41.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 492, + 372, + 503, + 379 + ], + "lines": [ + { + "bbox": [ + 492, + 372, + 503, + 379 + ], + "spans": [ + { + "bbox": [ + 492, + 372, + 503, + 379 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 386, + 546, + 409 + ], + "lines": [ + { + "bbox": [ + 46, + 386, + 546, + 409 + ], + "spans": [ + { + "bbox": [ + 46, + 386, + 546, + 409 + ], + "type": "text", + "content": "Figure 3. Qualitative comparisons of different methods on self reenactment task. From left to right: NeRFBlendShape [17], NeRFace [16], HAvatar [71] and Ours. Our method can reconstruct details like beards, teeth, etc. with high quality." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 415, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 287, + 510 + ], + "type": "text", + "content": "cameras are distributed about 120 degrees in front, and simultaneously capture 2K resolution video. For each identity, We use the sequences marked with \"FREE\" as the evaluation data, and the rest as the training data. For the 2 identities from HAvatar, each set contains 3000 frames, 8 cameras are distributed about 120 degrees in front, and 4K resolution videos are collected simultaneously. Later, we crop the face area and resize to 2K resolution." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 513, + 288, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 513, + 288, + 609 + ], + "spans": [ + { + "bbox": [ + 46, + 513, + 288, + 609 + ], + "type": "text", + "content": "For data preprocessing, we first remove the background [35] and extract 68 2D facial landmarks [10] for all the images. Then, for each frame, we use multi-view images to estimate the corresponding 3D landmarks, expression coefficients and head pose by fitting the Basel Face Model (BFM) [18] to the extracted 2D landmarks. Note that we define the 3D landmarks as the usual 68 landmarks with vertices indexed as multiples of 100 in the BFM vertices." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 47, + 621, + 189, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 189, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 189, + 634 + ], + "type": "text", + "content": "5.2. Results and Comparisons" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": "Self Reenactment. In this section, we first compare our method with existing SOTA methods in qualitative experiments on self reenactment task. Specifically, NeRFace [16] uses a deep MLP to fit an expression conditioned dynamic NeRF. The current SOTA method HAvatar [71] introduces 3DMM template prior and uses a deep convolutional net" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 415, + 546, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 546, + 487 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 546, + 487 + ], + "type": "text", + "content": "work to generate a human head NeRF represented by three planes from a mesh template with expression. Note, HAvatar leverages the GAN framework using the adversarial loss function to force the network to generate details that are not view-consistent. For a fair comparison, we remove this part and use VGG perceptual loss as Sec. 4.2 instead." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 487, + 546, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 487, + 546, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 487, + 546, + 548 + ], + "type": "text", + "content": "Qualitative results on self reenactment task are shown in the Fig. 3. Our method can accurately reconstruct pixel-level high-frequency details such as beards, teeth, and hair. Besides, our method can achieve expression transfer more accurately, such as eye movements in the figure." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 548, + 547, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 548, + 547, + 693 + ], + "spans": [ + { + "bbox": [ + 304, + 548, + 547, + 693 + ], + "type": "text", + "content": "Next, we conduct a quantitative evaluation for the four methods on 5 identities and 6 cameras using the evaluation split. The evaluation metrics include: Peak Signal-to-Noise Ratio (PSNR), Structure Similarity Index (SSIM), Learned Perceptual Image Patch Similarity (LPIPS) [70] and Fréchet Inception Distance (FID) [22]. Note, we calculate FID by comparing the distribution of all the training images and all the rendered images. As the task mainly focuses on the reconstruction of the head, we use face parsing to remove the body parts in the image to eliminate their impact in the experiment. As shown in Tab. 1, our method demonstrates a slight improvement in PSNR and SSIM compared with" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 702, + 487, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 702, + 487, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 702, + 487, + 713 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 317, + 702, + 487, + 713 + ], + "type": "text", + "content": "https://github.com/zllrunning/faceParsing.PyTorch" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1936" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 70, + 521, + 147 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 521, + 147 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 521, + 147 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 521, + 147 + ], + "type": "table", + "html": "
MethodPSNR ↑SSIM ↑LPIPS (512) ↓LPIPS (2K) ↓FID (2K) ↓
NeRFBlendShape25.910.8360.1230.22954.80
NeRFace27.140.8490.1470.23465.11
HAVatar27.190.8830.0640.20931.06
Ours (w/o SR)27.820.8870.0800.20245.50
Ours27.700.8830.0560.09818.50
", + "image_path": "94e0c8c1d19751671903d66d027f4d7b836d3515e6ecf40ac51c6d7e30256eaa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 47, + 180, + 545, + 399 + ], + "blocks": [ + { + "bbox": [ + 47, + 180, + 545, + 399 + ], + "lines": [ + { + "bbox": [ + 47, + 180, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 47, + 180, + 545, + 399 + ], + "type": "image", + "image_path": "4f216f90bc2dd0933614856803541ac8434a5d55705005552af1bd5e5231b262.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 405, + 546, + 429 + ], + "lines": [ + { + "bbox": [ + 46, + 405, + 546, + 429 + ], + "spans": [ + { + "bbox": [ + 46, + 405, + 546, + 429 + ], + "type": "text", + "content": "Figure 4. Qualitative comparisons of different methods on cross-identity reenactment task. From left to right: NeRFBlendShape [17], NeRFace [16], HAvatar [71] and Ours. Our method synthesizes high-fidelity images while ensuring the accuracy of expression transfer." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 56, + 442, + 276, + 509 + ], + "blocks": [ + { + "bbox": [ + 46, + 149, + 546, + 172 + ], + "lines": [ + { + "bbox": [ + 46, + 149, + 546, + 172 + ], + "spans": [ + { + "bbox": [ + 46, + 149, + 546, + 172 + ], + "type": "text", + "content": "Table 1. Quantitative evaluation results of NeRFBlendShape [17], NeRFace [16], HAvatar [71], our method without super resolution and our full method on self reenactment task." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 442, + 276, + 509 + ], + "lines": [ + { + "bbox": [ + 56, + 442, + 276, + 509 + ], + "spans": [ + { + "bbox": [ + 56, + 442, + 276, + 509 + ], + "type": "table", + "html": "
MethodPSNR ↑SSIM ↑LPIPS ↓
NeRFBlendShape25.430.8120.148
NeRFace26.650.8250.151
HAVatar27.130.8800.65
Ours27.580.8820.059
", + "image_path": "4bb0f82960493835c93524465f4158c84387f872cf985e4f9f417f54e7ce4430.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 544, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 581 + ], + "type": "text", + "content": "previous methods, and a significant improvement in LPIPS and FID, which means that our method can generate more high-frequency details." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 582, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 640 + ], + "type": "text", + "content": "Cross-Identity Reenactment. We also qualitatively compare our method with the above SOTA methods on cross-identity reenactment task. As shown in the results, our method is able to synthesize higher-fidelity images with more accurate expression transfer and richer emotions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": "Novel View Synthesis. In this section, we show the results of novel view synthesis as shown in Fig. 5 shown. In this case, we use video data from 8 views for training and render the image at a new viewpoint. Next, we quantitatively evaluate the 3D consistency of our method and compare it with other SOTA methods mentioned above. Specif" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 440, + 546, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 440, + 546, + 513 + ], + "spans": [ + { + "bbox": [ + 304, + 440, + 546, + 513 + ], + "type": "text", + "content": "ically, we select the 5 identities as above and use the video data from 8 cameras for training while the other 8 holdout cameras for evaluation. The evaluation metrics include PSNR, SSIM and LPIPS (512 resolution). The results are shown in Tab. 2. Our method outperforms other methods in 3D consistency." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 521, + 400, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 521, + 400, + 534 + ], + "spans": [ + { + "bbox": [ + 306, + 521, + 400, + 534 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 323, + 539, + 527, + 593 + ], + "blocks": [ + { + "bbox": [ + 46, + 511, + 287, + 534 + ], + "lines": [ + { + "bbox": [ + 46, + 511, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 46, + 511, + 287, + 534 + ], + "type": "text", + "content": "Table 2. Quantitative evaluation results of the other SOTA methods and our method on 3D consistency." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 323, + 539, + 527, + 593 + ], + "lines": [ + { + "bbox": [ + 323, + 539, + 527, + 593 + ], + "spans": [ + { + "bbox": [ + 323, + 539, + 527, + 593 + ], + "type": "table", + "html": "
MethodPSNR ↑SSIM ↑LPIPS ↓
FLAME-Init28.730.8750.123
Mesh-Deform28.830.8740.116
Ours28.940.8760.108
", + "image_path": "2ceb035ef74c945f55d4aa0174bbac279c66e67a394b98325049f7fcf839f205.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 596, + 545, + 617 + ], + "lines": [ + { + "bbox": [ + 304, + 596, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 596, + 545, + 617 + ], + "type": "text", + "content": "Table 3. Quantitative evaluation results of the other two ablation baselines and ours on self reenactment task." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 629, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 546, + 713 + ], + "type": "text", + "content": "Ablation on Initialization Strategies. In order to verify the effectiveness of our geometry-guided initialization strategy 4.3, we compare it with the strategy to use the FLAME model for initialization (FLAME-Init). Specifically, after fitting a FLAME model through multi-view data, we first subdivide the FLAME mesh 4 times and use the neutral vertices as the positions of the neutral Gaussians. Then, the" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1937" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 70, + 547, + 168 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 547, + 168 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 547, + 168 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 547, + 168 + ], + "type": "image", + "image_path": "2047bfc484c3da4a7cbc4535dea093d79bfb21ca553d99480cca68bdf49ceab6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 91, + 172, + 501, + 184 + ], + "lines": [ + { + "bbox": [ + 91, + 172, + 501, + 184 + ], + "spans": [ + { + "bbox": [ + 91, + 172, + 501, + 184 + ], + "type": "text", + "content": "Figure 5. Novel view synthesis results of our method. We use 8-view synchronized videos for training the avatar." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 47, + 193, + 289, + 378 + ], + "blocks": [ + { + "bbox": [ + 47, + 193, + 289, + 378 + ], + "lines": [ + { + "bbox": [ + 47, + 193, + 289, + 378 + ], + "spans": [ + { + "bbox": [ + 47, + 193, + 289, + 378 + ], + "type": "image", + "image_path": "4c63f457e03a24d4ef3f2400d188ade0f7fa21b59cd4d716638e1f98f19bb6f2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 384, + 287, + 417 + ], + "lines": [ + { + "bbox": [ + 46, + 384, + 287, + 417 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 287, + 417 + ], + "type": "text", + "content": "Figure 6. Ablation study on the initialization strategies: FLAME-initialization and our geometry-guided initialization. Our strategy ensures the hair strands away from the head are well reconstructed." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 433, + 287, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 433, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 433, + 287, + 540 + ], + "type": "text", + "content": "expression deformation MLP is optimized to learn the displacement of FLAME vertices. We set the per-vertex feature to zeros, while randomly initialize the parameters of the expression color MLP. The initialization of other variables is the same as our strategy. Qualitative results are shown in the Fig. 6. Due to the lack of initialization for the hair and shoulders in FLAME-Init, the points to model these parts are offset from nearby vertices, which leads to sparseness of the Gaussians, resulting in blurring." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "type": "text", + "content": "Ablation on Deformation Modeling Approaches. We compare our fully learned deformation field with the previous mesh-based deformation (Mesh-Deform). Specifically, we migrate the method in INSTA [76] for controlling the NeRF deformation to our Gaussians. First we fit a 3DMM mesh template. Then, for each Gaussian point, find the closest face on the mesh, and calculate the deformation gradient to estimate the displacement. Qualitative results are shown in the Fig. 7. For some expressions that cannot be captured well by the 3DMM mesh template, our method can learn accurate deformation, thereby achieving the modeling of complex expressions. Quantitative results are shown in the Fig. 3. Our method outperforms both the two ablation baselines on PSNR, SSIM and LPIPS metrics." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 306, + 194, + 547, + 377 + ], + "blocks": [ + { + "bbox": [ + 306, + 194, + 547, + 377 + ], + "lines": [ + { + "bbox": [ + 306, + 194, + 547, + 377 + ], + "spans": [ + { + "bbox": [ + 306, + 194, + 547, + 377 + ], + "type": "image", + "image_path": "0f815ad2a74e0582c53a4a406698b5868da703d611678876dd1ca739d76d5c22.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 383, + 546, + 417 + ], + "lines": [ + { + "bbox": [ + 305, + 383, + 546, + 417 + ], + "spans": [ + { + "bbox": [ + 305, + 383, + 546, + 417 + ], + "type": "text", + "content": "Figure 7. Ablation study on the deformation modeling: mesh LBS-based deformation and our fully learned deformation. Our approach can learn complex and exaggerated expressions." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 427, + 457, + 439 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 427, + 457, + 439 + ], + "spans": [ + { + "bbox": [ + 306, + 427, + 457, + 439 + ], + "type": "text", + "content": "6. Discussion and Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 451, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 451, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 305, + 451, + 545, + 498 + ], + "type": "text", + "content": "Ethical Considerations. Our method is capable of creating artificial portrait videos, which have the potential to disseminate misinformation, influence public perceptions, and erode confidence in media sources." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 502, + 545, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 502, + 545, + 538 + ], + "spans": [ + { + "bbox": [ + 305, + 502, + 545, + 538 + ], + "type": "text", + "content": "Limitation. For the tongue and teeth inside the mouth or long hair, blurring is sometimes produced in our method due to the lack of tracking methods." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 542, + 545, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 542, + 545, + 661 + ], + "spans": [ + { + "bbox": [ + 304, + 542, + 545, + 661 + ], + "type": "text", + "content": "Conclusion. In this paper, we propose Gaussian Head Avatar, a novel representation for head avatar reconstruction, which leverages dynamic 3D Gaussians controlled by a fully learned expression deformation. Experiments demonstrate our method can synthesize ultra high-fidelity images while modeling exaggerated expressions. In addition, we propose a well-designed minute-level initialization strategy to ensure the training convergence. We believe our Gaussian Head Avatar will become the mainstream direction for head avatar reconstruction in the future." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgement. This paper is supported by National Key R&D Program of China (2022YFF0902200), the NSFC project No.62125107, the Beijing Municipal Science & Technology Z231100005923030." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1938" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "text", + "content": "[1] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In European Conference on Computer Vision, pages 696-712, 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 288, + 189 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 288, + 189 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 288, + 189 + ], + "type": "text", + "content": "[2] ShahRukh Athar, Zexiang Xu, Kalyan Sunkavalli, Eli Shechtman, and Zhixin Shu. Rignerf: Fully controllable neural 3d portraits. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 190, + 287, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 190, + 287, + 244 + ], + "spans": [ + { + "bbox": [ + 53, + 190, + 287, + 244 + ], + "type": "text", + "content": "[3] ShahRukh Athar, Zhixin Shu, and Dimitris Samaras. Flame-in-nerf: Neural control of radiance fields for free view face animation. In IEEE 17th International Conference on Automatic Face and Gesture Recognition (FG), pages 1–8, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 246, + 287, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 246, + 287, + 280 + ], + "spans": [ + { + "bbox": [ + 53, + 246, + 287, + 280 + ], + "type": "text", + "content": "[4] Thabo Beeler, Bernd Bickel, Paul Beardsley, Bob Sumner, and Markus Gross. High-quality single-shot capture of facial geometry. ACM Trans. Graph., 29(4), 2010. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 280, + 287, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 280, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 53, + 280, + 287, + 334 + ], + "type": "text", + "content": "[5] Sai Bi, Stephen Lombardi, Shunsuke Saito, Tomas Simon, Shih-En Wei, Kevyn Mcphail, Ravi Ramamoorthi, Yaser Sheikh, and Jason Saragih. Deep reeligible appearance models for animatable faces. ACM Trans. Graph., 40(4), 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 335, + 287, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 287, + 378 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 287, + 378 + ], + "type": "text", + "content": "[6] V Blanz and T Vetter. A morphable model for the synthesis of 3d faces. In 26th Annual Conference on Computer Graphics and Interactive Techniques (SIGGRAPH 1999), pages 187-194. ACM Press, 1999. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 380, + 287, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 287, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 287, + 423 + ], + "type": "text", + "content": "[7] Timo Bolkart, Tianye Li, and Michael J. Black. Instant multi-view head capture through learnable registration. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 768-779, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 424, + 287, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 424, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 53, + 424, + 287, + 456 + ], + "type": "text", + "content": "[8] Derek Bradley, Wolfgang Heidrich, Tiberiu Popa, and Alla Sheffer. High resolution passive facial performance capture. 29(4), 2010. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 457, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 457, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 53, + 457, + 287, + 502 + ], + "type": "text", + "content": "[9] Alan Brunton, Timo Bolkart, and Stefanie Wuhrer. Multilinear wavelets: A statistical shape space for human faces. In Proceedings of the Proceedings of the European Conference on Computer Vision (ECCV), 2014. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 502, + 287, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 502, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 48, + 502, + 287, + 545 + ], + "type": "text", + "content": "[10] Adrian Bulat and Georgios Tzimiropoulos. How far are we from solving the 2d & 3d face alignment problem? (and a dataset of 230,000 3d facial landmarks). In International Conference on Computer Vision, 2017. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 546, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 546, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 546, + 287, + 590 + ], + "type": "text", + "content": "[11] Chen Cao, Yanlin Weng, Shun Zhou, Y. Tong, and Kun Zhou. Facewarehouse: A 3d facial expression database for visual computing. In IEEE Transactions on Visualization and Computer Graphics, pages 413-425, 2014. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 591, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 287, + 624 + ], + "type": "text", + "content": "[12] Chen Cao, Derek Bradley, Kun Zhou, and Thabo Beeler. Real-time high-fidelity facial performance capture. ACM Trans. Graph., 34(4), 2015. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 624, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 657 + ], + "type": "text", + "content": "[13] Chen Cao, Hongzhi Wu, Yanlin Weng, Tianjia Shao, and Kun Zhou. Real-time facial animation with image-based dynamic avatars. ACM Trans. Graph., 35(4), 2016. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[14] Chen Cao, Tomas Simon, Jin Kyu Kim, Gabe Schwartz, Michael Zollhoefer, Shun-Suke Saito, Stephen Lombardi, Shih-En Wei, Danielle Belko, Shouu-I Yu, Yaser Sheikh, and Jason Saragih. Authentic volumetric avatars from a phone scan. ACM Trans. Graph., 41(4), 2022. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "type": "text", + "content": "[15] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2019. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 129, + 547, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 183 + ], + "type": "text", + "content": "[16] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Niessner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8645-8654, 2021. 3, 4, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 184, + 547, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 547, + 237 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 547, + 237 + ], + "type": "text", + "content": "[17] Xuan Gao, Chenglai Zhong, Jun Xiang, Yang Hong, Yudong Guo, and Juyong Zhang. Reconstructing personalized semantic facial nerf models from monocular video. ACM Transactions on Graphics (Proceedings of SIGGRAPH Asia), 41(6), 2022. 3, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 238, + 547, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 238, + 547, + 281 + ], + "spans": [ + { + "bbox": [ + 308, + 238, + 547, + 281 + ], + "type": "text", + "content": "[18] Thomas Gerig, Andreas Forster, Clemens Blumer, Bernhard Egger, Marcel Lüthi, Sandro Schonborn, and Thomas Vetter. Morphable face models - an open framework. pages 75-82, 2017. 3, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 283, + 547, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 283, + 547, + 326 + ], + "spans": [ + { + "bbox": [ + 308, + 283, + 547, + 326 + ], + "type": "text", + "content": "[19] Abhijeet Ghosh, Graham Fyffe, Borom Tunwattanapong, Jay Busch, Xueming Yu, and Paul Debevec. Multiview face capture using polarized spherical gradient illumination. ACM Trans. Graph., 30(6):1-10, 2011. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 327, + 547, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 547, + 381 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 547, + 381 + ], + "type": "text", + "content": "[20] Philip-William Grassal, Malte Prinzler, Titus Leistner, Carsten Rother, Matthias Nießner, and Justus Thies. Neural head avatars from monocular rgb videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18632–18643, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 382, + 547, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 382, + 547, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 382, + 547, + 437 + ], + "type": "text", + "content": "[21] Yudong Guo, Keyu Chen, Sen Liang, Yong-Jin Liu, Hujun Bao, and Juyong Zhang. Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5764–5774, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 437, + 547, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 547, + 481 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 547, + 481 + ], + "type": "text", + "content": "[22] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Neural Information Processing Systems, 2017. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 482, + 547, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 547, + 535 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 547, + 535 + ], + "type": "text", + "content": "[23] Yang Hong, Bo Peng, Haiyao Xiao, Ligang Liu, and Juyong Zhang. Headnerf: A real-time nerf-based parametric head model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20374-20384, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 536, + 547, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 536, + 547, + 590 + ], + "spans": [ + { + "bbox": [ + 308, + 536, + 547, + 590 + ], + "type": "text", + "content": "[24] Liwen Hu, Shunsuke Saito, Lingyu Wei, Koki Nagano, Jaewoo Seo, Jens Fursund, Iman Sadeghi, Carrie Sun, YenChun Chen, and Hao Li. Avatar digitization from a single image for real-time rendering. ACM Trans. Graph., 36(6), 2017. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 592, + 547, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 592, + 547, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 592, + 547, + 624 + ], + "type": "text", + "content": "[25] Alexandru Eugen Ichim, Sofien Bouaziz, and Mark Pauly. Dynamic 3d avatar creation from hand-held video input. ACM Trans. Graph., 34(4), 2015. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 625, + 547, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 547, + 668 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 547, + 668 + ], + "type": "text", + "content": "[26] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 2, 3, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 669, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 669, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 669, + 547, + 713 + ], + "type": "text", + "content": "[27] Taras Khakhulin, Vanessa Sklyarova, Victor Lempitsky, and Egor Zakharov. Realistic one-shot mesh-based head avatars. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1939" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[28] Tobias Kirschstein, Shenhan Qian, Simon Giebenhain, Tim Walter, and Matthias Nießner. Nersemble: Multi-view radiance field reconstruction of human heads. ACM Trans. Graph., 42(4), 2023. 2, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "type": "text", + "content": "[29] Georgios Kopanas, Thomas Leimkuhler, Gilles Rainer, Clément Jambon, and George Drettakis. Neural point cata-caustics for novel-view synthesis of reflections. ACM Transactions on Graphics (TOG), 41(6):1-15, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 208 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 208 + ], + "type": "text", + "content": "[30] Christoph Lassner and Michael Zollhofer. Pulsar: Efficient sphere-based neural rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1440-1449, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 288, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 288, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 288, + 297 + ], + "type": "text", + "content": "[31] Marc Levoy, Kari Pulli, Brian Curless, Szymon Rusinkiewicz, David Koller, Lucas Pereira, Matt Ginzon, Sean Anderson, James Davis, Jeremy Ginsberg, Jonathan Shade, and Duane Fulk. The digital michelangelo project: 3d scanning of large statues. In Proceedings of the 27th Annual Conference on Computer Graphics and Interactive Techniques, page 131-144, USA, 2000. ACM Press/Addison-Wesley Publishing Co. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 297, + 287, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 297, + 287, + 341 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 287, + 341 + ], + "type": "text", + "content": "[32] Tianye Li, Timo Bolkart, Michael J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6), 2017. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 342, + 287, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 342, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 48, + 342, + 287, + 398 + ], + "type": "text", + "content": "[33] Tianye Li, Shichen Liu, Timo Bolkart, Jiayi Liu, Hao Li, and Yajie Zhao. Topologically consistent multi-view face inference using volumetric sampling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3824-3834, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 399, + 288, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 288, + 454 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 288, + 454 + ], + "type": "text", + "content": "[34] Zhe Li, Zerong Zheng, Lizhen Wang, and Yebin Liu. Animatable gaussians: Learning pose-dependent gaussian maps for high-fidelity human avatar modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 455, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 455, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 455, + 287, + 510 + ], + "type": "text", + "content": "[35] Shanchuan Lin, Andrey Ryabtsev, Soumyadip Sengupta, Brian Curless, Steve Seitz, and Ira Kemelmacher-Shlizerman. Real-time high-resolution background matting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 510, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 287, + 555 + ], + "type": "text", + "content": "[36] Xian Liu, Yinghao Xu, Qianyi Wu, Hang Zhou, Wayne Wu, and Bolei Zhou. Semantic-aware implicit neural auditory-driven video portrait generation. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 555, + 287, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 589 + ], + "type": "text", + "content": "[37] Stephen Lombardi, Jason Saragih, Tomas Simon, and Yaser Sheikh. Deep appearance models for face rendering. ACM Trans. Graph., 37(4):68:1-68:13, 2018. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 590, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 635 + ], + "type": "text", + "content": "[38] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural volumes: Learning dynamic renderable volumes from images. ACM Trans. Graph., 38(4):65:1-65:14, 2019. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 679 + ], + "type": "text", + "content": "[39] Stephen Lombardi, Tomas Simon, Gabriel Schwartz, Michael Zollhoefer, Yaser Sheikh, and Jason Saragih. Mixture of volumetric primitives for efficient neural rendering. ACM Trans. Graph., 40(4), 2021. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "text", + "content": "[40] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis, 2023. 2, 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 546, + 714 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 127 + ], + "type": "text", + "content": "[41] Shugao Ma, Tomas Simon, Jason Saragih, Dawei Wang, Yuecheng Li, Fernando De La Torre, and Yaser Sheikh. Pixel codec avatars. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 64-73, 2021. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 128, + 546, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 128, + 546, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 128, + 546, + 183 + ], + "type": "text", + "content": "[42] Marko Mihajlovic, Aayush Bansal, Michael Zollhoefer, Siyu Tang, and Shunsuke Saito. Keypointnerf: Generalizing image-based volumetric avatars using relative spatial encoding of keypoints. In European conference on computer vision, 2022. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 184, + 546, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 546, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 546, + 239 + ], + "type": "text", + "content": "[43] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 239, + 546, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 546, + 304 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 546, + 304 + ], + "type": "text", + "content": "[44] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting Triangular 3D Models, Materials, and Lighting From Images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8280-8290, 2022. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 305, + 546, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 305, + 546, + 350 + ], + "spans": [ + { + "bbox": [ + 308, + 305, + 546, + 350 + ], + "type": "text", + "content": "[45] Koki Nagano, Jaewoo Seo, Jun Xing, Lingyu Wei, Zimo Li, Shunsuke Saito, Aviral Agarwal, Jens Fursund, and Hao Li. Pagan: Real-time avatars using dynamic textures. ACM Trans. Graph., 37(6), 2018. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 350, + 546, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 546, + 405 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 546, + 405 + ], + "type": "text", + "content": "[46] Pascal Paysan, Reinhard Knothe, Brian Amberg, Sami Romdhani, and Thomas Vetter. A 3d face model for pose and illumination invariant face recognition. In 2009 Sixth IEEE International Conference on Advanced Video and Signal Based Surveillance, pages 296-301, 2009. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 405, + 546, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 405, + 546, + 459 + ], + "spans": [ + { + "bbox": [ + 308, + 405, + 546, + 459 + ], + "type": "text", + "content": "[47] Minghan Qin, Yifan Liu, Yuelang Xu, Xiaochen Zhao, Yebin Liu, and Haoqian Wang. High-fidelity 3d head avatars reconstruction through spatially-varying expression conditioned neural radiance field. In AAAI Conference on Artificial Intelligence, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 459, + 546, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 459, + 546, + 502 + ], + "spans": [ + { + "bbox": [ + 308, + 459, + 546, + 502 + ], + "type": "text", + "content": "[48] Amit Raj, Michael Zollhoefer, Tomas Simon, Jason Saragih, Shunsuke Saito, James Hays, and Stephen Lombardi. Pva: Pixel-aligned volumetric avatars. In arXiv:2101.02697, 2020. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 503, + 546, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 503, + 546, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 503, + 546, + 536 + ], + "type": "text", + "content": "[49] Darius Rückert, Linus Franke, and Marc Stamminger. Adop: Approximate differentiable one-pixel point rendering. ACM Trans. Graph., 41(4), 2022. 3, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 537, + 546, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 546, + 591 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 546, + 591 + ], + "type": "text", + "content": "[50] Ruizhi Shao, Jingxiang Sun, Cheng Peng, Zerong Zheng, Boyao Zhou, Hongwen Zhang, and Yebin Liu. Control4d: Efficient 4d portrait editing with text. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 592, + 546, + 636 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 592, + 546, + 636 + ], + "spans": [ + { + "bbox": [ + 308, + 592, + 546, + 636 + ], + "type": "text", + "content": "[51] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 636, + 546, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 546, + 680 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 546, + 680 + ], + "type": "text", + "content": "[52] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 681, + 546, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 546, + 714 + ], + "type": "text", + "content": "[53] Jingxiang Sun, Xuan Wang, Lizhen Wang, Xiaoyu Li, Yong Zhang, Hongwen Zhang, and Yebin Liu. Next3d: Generative neural texture rasterization for 3d-aware head avatars." + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1940" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "type": "text", + "content": "[54] Cong Wang, Di Kang, Yanpei Cao, Linchao Bao, Ying Shan, and Song-Hai Zhang. Neural point-based volumetric avatar: Surface-guided neural points for efficient and photorealistic volumetric head avatar. In ACM SIGGRAPH Asia 2023 Conference Proceedings, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 151, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 151, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 151, + 287, + 205 + ], + "type": "text", + "content": "[55] Daoye Wang, Prashanth Chandran, Gaspard Zoss, Derek Bradley, and Paulo Gotardo. Morf: Morphable radiance fields for multiview neural head modeling. In ACM SIGGRAPH 2022 Conference Proceedings, New York, NY, USA, 2022. Association for Computing Machinery. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 206, + 287, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 206, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 206, + 287, + 259 + ], + "type": "text", + "content": "[56] Lizhen Wang, Zhiyuan Chen, Tao Yu, Chenguang Ma, Liang Li, and Yebin Liu. Faceverse: a fine-grained and detail-controllable 3d face morphable model from a hybrid dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 261, + 287, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 287, + 325 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 287, + 325 + ], + "type": "text", + "content": "[57] Ziyan Wang, Timur Bagautdinov, Stephen Lombardi, Tomas Simon, Jason Saragih, Jessica Hodgins, and Michael Zollhofer. Learning compositional radiance fields of dynamic human heads. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5704-5713, 2021. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 327, + 287, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 287, + 358 + ], + "type": "text", + "content": "[58] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. SynSin: End-to-end view synthesis from a single image. In CVPR, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 360, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 287, + 403 + ], + "type": "text", + "content": "[59] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering, 2023. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 404, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 287, + 459 + ], + "type": "text", + "content": "[60] Sijing Wu, Yichao Yan, Yunhao Li, Yuhao Cheng, Wenhan Zhu, Ke Gao, Xiaobo Li, and Guangtao Zhai. Ganhead: Towards generative animatable neural head avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 437-447, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 460, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 287, + 514 + ], + "type": "text", + "content": "[61] Yunze Xiao, Hao Zhu, Haotian Yang, Zhengyu Diao, Xiangju Lu, and Xun Cao. Detailed facial geometry recovery from multi-view images by learning an implicit function. In Proceedings of the AAAI Conference on Artificial Intelligence, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 514, + 287, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 569 + ], + "type": "text", + "content": "[62] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5438–5448, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 571, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 571, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 571, + 287, + 613 + ], + "type": "text", + "content": "[63] Yuelang Xu, Lizhen Wang, Xiaochen Zhao, Hongwen Zhang, and Yebin Liu. Avatarmav: Fast 3d head avatar reconstruction using motion-aware neural voxels. In ACM SIGGRAPH 2023 Conference Proceedings, 2023. 3, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 614, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 667 + ], + "type": "text", + "content": "[64] Yuelang Xu, Hongwen Zhang, Lizhen Wang, Xiaochen Zhao, Huang Han, Qi Guojun, and Yebin Liu. Latentavatar: Learning latent expression code for expressive neural head avatar. In ACM SIGGRAPH 2023 Conference Proceedings, 2023. 3, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "text", + "content": "[65] Kai Yang, Hong Shang, Tianyang Shi, Xinghan Chen, Jingkai Zhou, Zhongqian Sun, and Wei Yang. Asm: Adaptive skinning model for high-quality 3d face modeling supplementary material. 2021. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 546, + 597 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[66] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction, 2023. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "text", + "content": "[67] Zeyu Yang, Hongye Yang, Zijie Pan, Xiatian Zhu, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting, 2023. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 141, + 546, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 546, + 195 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 546, + 195 + ], + "type": "text", + "content": "[68] T Yenamandra, A Tewari, F Bernard, HP Seidel, M Elgharib, D Cremers, and C Theobalt. i3dmm: Deep implicit 3d morphable model of human heads. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 197, + 545, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 197, + 545, + 250 + ], + "spans": [ + { + "bbox": [ + 308, + 197, + 545, + 250 + ], + "type": "text", + "content": "[69] Wang Yifan, Felice Serena, Shihao Wu, Cengiz Öztireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM Transactions on Graphics (proceedings of ACM SIGGRAPH ASIA), 38(6), 2019. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 252, + 546, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 252, + 546, + 307 + ], + "spans": [ + { + "bbox": [ + 308, + 252, + 546, + 307 + ], + "type": "text", + "content": "[70] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 308, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 308, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 308, + 308, + 545, + 352 + ], + "type": "text", + "content": "[71] Xiaochen Zhao, Lizhen Wang, Jingxiang Sun, Hongwen Zhang, Jinli Suo, and Yebin Liu. Havatar: High-fidelity head avatar via facial model conditioned neural radiance field. ACM Trans. Graph., 2023. Just Accepted. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 353, + 546, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 353, + 546, + 418 + ], + "spans": [ + { + "bbox": [ + 308, + 353, + 546, + 418 + ], + "type": "text", + "content": "[72] Shunyuan Zheng, Boyao Zhou, Ruizhi Shao, Boning Liu, Shengping Zhang, Liqiang Nie, and Yebin Liu. Gpsgaussian: Generalizable pixel-wise 3d gaussian splatting for real-time human novel view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 420, + 545, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 420, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 308, + 420, + 545, + 485 + ], + "type": "text", + "content": "[73] Yufeng Zheng, Victoria Fernández Abrevaya, Marcel C. Bühler, Xu Chen, Michael J. Black, and Otmar Hilliges. I m avatar: Implicit morphable head avatars from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13535-13545, 2022. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 487, + 545, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 487, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 308, + 487, + 545, + 540 + ], + "type": "text", + "content": "[74] Yufeng Zheng, Wang Yifan, Gordon Wetzstein, Michael J. Black, and Otmar Hilliges. Pointavatar: Deformable point-based head avatars from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 542, + 545, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 542, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 308, + 542, + 545, + 575 + ], + "type": "text", + "content": "[75] Yiyu Zhuang, Hao Zhu, Xusen Sun, and Xun Cao. Mofanerf: Morphable facial neural radiance field. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 576, + 545, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 576, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 308, + 576, + 545, + 597 + ], + "type": "text", + "content": "[76] Wojciech Zielonka, Timo Bolkart, and Justus Thies. Instant volumetric head avatars, 2022. 2, 3, 8" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "text", + "content": "1941" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Gaussian Shading_ Provable Performance-Lossless Image Watermarking for Diffusion Models/cf093428-bde9-4e71-b54b-2506850cecda_content_list.json b/2024/Gaussian Shading_ Provable Performance-Lossless Image Watermarking for Diffusion Models/cf093428-bde9-4e71-b54b-2506850cecda_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..84088151ac51ecf54994598818de29fd537d26ca --- /dev/null +++ b/2024/Gaussian Shading_ Provable Performance-Lossless Image Watermarking for Diffusion Models/cf093428-bde9-4e71-b54b-2506850cecda_content_list.json @@ -0,0 +1,1994 @@ +[ + { + "type": "text", + "text": "Gaussian Shading: Provable Performance-Lossless Image Watermarking for Diffusion Models", + "text_level": 1, + "bbox": [ + 99, + 128, + 870, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zijin Yang $^{1}$ , Kai Zeng $^{1}$ , Kejiang Chen $^{1,*}$ , Han Fang $^{2}$ , Weiming Zhang $^{1}$ , Nenghai Yu $^{1}$ , $^{1}$ Anhui Province Key Laboratory of Digital Security, University of Science and Technology of China $^{2}$ National University of Singapore", + "bbox": [ + 88, + 202, + 880, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{bsmhmmlf@mail.,zk0128@mail.,chenkj@,zhangwm@,ynh@}ustc.edu.cn fanghan@nus.edu.sg", + "bbox": [ + 101, + 258, + 861, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 308, + 313, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ethical concerns surrounding copyright protection and inappropriate content generation pose challenges for the practical implementation of diffusion models. One effective solution involves watermarking the generated images. However, existing methods often compromise the model performance or require additional training, which is undesirable for operators and users. To address this issue, we propose Gaussian Shading, a diffusion model watermarking technique that is both performance-lossless and training-free, while serving the dual purpose of copyright protection and tracing of offending content. Our watermark embedding is free of model parameter modifications and thus is plug-and-play. We map the watermark to latent representations following a standard Gaussian distribution, which is indistinguishable from latent representations obtained from the non-watermarked diffusion model. Therefore we can achieve watermark embedding with lossless performance, for which we also provide theoretical proof. Furthermore, since the watermark is intricately linked with image semantics, it exhibits resilience to lossy processing and erasure attempts. The watermark can be extracted by Denoising Diffusion Implicit Models (DDIM) inversion and inverse sampling. We evaluate Gaussian Shading on multiple versions of Stable Diffusion, and the results demonstrate that Gaussian Shading not only is performance-lossless but also outperforms existing methods in terms of robustness.", + "bbox": [ + 73, + 340, + 473, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 762, + 209, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion models [16, 31-34] signify a noteworthy leap forward in image generation. These well-trained diffusion models, especially commercial diffusion models like Stable Diffusion (SD) [30], Glide [27], and Muse AI [30], enable individuals with diverse backgrounds to create high-quality images effortlessly. However, this raises concerns about", + "bbox": [ + 75, + 786, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bcf03ab1c92533925c1aa3c8a58d737b9fb74127d1de4e4f035c46fea24683cd.jpg", + "image_caption": [ + "Figure 1. Existing watermarking frameworks can be divided into three categories: post-processing-based, fine-tuning-based, and latent-representation-based Tree-Ring. Our method also relies on latent representations but achieves performance-lossless without altering the distribution." + ], + "image_footnote": [], + "bbox": [ + 522, + 306, + 867, + 551 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "intellectual property and whether diffusion models will be stolen or resold twice.", + "bbox": [ + 496, + 626, + 892, + 655 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "On the other hand, the ease of generating realistic images raises concerns about potentially misleading content generation. For example, on May 23, 2023, a Twitter-verified user named Bloomberg Feed posted a tweet titled \"Large explosion near the Pentagon complex in Washington DC-initial report,\" along with a synthetic image. This tweet led to multiple authoritative media accounts sharing it, even causing a brief impact on the stock market1. On October 30, 2023, White House issued an executive order on AI security, emphasizing the need to protect Americans from AI-enabled fraud and deception by establishing standards and best practices for detecting AI-generated content and authenticating official content2. The urgency of labeling gen", + "bbox": [ + 496, + 656, + 892, + 853 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 94, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1Fake image of Pentagon explosion on Twitter", + "bbox": [ + 517, + 862, + 764, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "2FACT SHEET: President Biden Issues Executive Order on Safe, Secure, and Trustworthy Artificial Intelligence", + "bbox": [ + 501, + 876, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12162", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "erated content for copyright authentication and prevention of misuse is evident.", + "bbox": [ + 75, + 90, + 468, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Watermarking is highlighted as a fundamental method for labeling generated content, as it embeds watermark information within the generated image, allowing for subsequent copyright authentication and the tracking of false content. Existing watermarking methods for the diffusion model can be divided into three categories, as shown in Fig. 1. Post-processing-based watermarks [6, 45] adjust robust image features to embed watermarks, thereby directly altering the image and degrading its quality. To mitigate this concern, recent research endeavors propose fine-tuning-based methods [7, 10, 24, 42, 49], which amalgamate the watermark embedding process with the image generation process. Intuitively, these methods need to modify model parameters, introducing supplementary computational overhead. Recently, Wen et al. [41] proposed the latent-representation-based Tree-Ring watermark, which conveys information by adapting the latent representations to match specific patterns. However, it restricts the randomness of sampling, which impacts generative performance.", + "bbox": [ + 75, + 123, + 470, + 411 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Through the above analysis, we can find that these methods compromise model performance to embed watermarks. In practical applications, model performance is paramount for both business interests and user experience. Substantial resource investment is often necessary to pursue enhanced model performance. This leads to a fundamental question: Can watermarks be embedded without compromising model performance?", + "bbox": [ + 75, + 412, + 468, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We affirmatively address the question presented above. Succinctly, the generation process can be delineated into two key phases: latent representation sampling and decoding. Our goal is to align the distribution of the latent representation in watermarked images with that of the latent representation in normally generated images. By keeping the model unaltered, the distribution of watermarked images is naturally consistent with that of normally generated images, enabling the seamless embedding of watermarks without compromising model performance.", + "bbox": [ + 75, + 536, + 468, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building upon this insight, we propose a watermarking method named Gaussian Shading, designed to ensure no deterioration in model performance. The embedding process encompasses three primary elements: watermark diffuse, randomization, and distribution-preserving sampling. Watermark diffusion spreads the watermark information throughout the latent representation. During the generation process, the watermark information will be diffused to the whole semantics of the image, thus achieving excellent robustness. Watermark randomization and distribution-preserving sampling guarantee the congruity of the latent representation distribution with that of watermark-free latent representations, thereby achieving performance-lossless. In the extraction phase, the latent representations", + "bbox": [ + 75, + 689, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "are acquired through Denoising Diffusion Implicit Model (DDIM) inversion [32], allowing for the retrieval of watermark information. Harnessing the extensive scope of the SD latent space, we can achieve a high-capacity watermark of 256 bits, surpassing prior methods.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To the best of our knowledge, ours is the first technique that tackles this challenging problem of performanceless watermarking for diffusion models, and we provide theoretical proof. Moreover, this technique leaves the architecture and parameters of SD unaltered, necessitating no supplementary training. It can seamlessly integrate as a plug-and-play module within the generation process. Model providers can easily replace watermarked models with non-watermarked ones without affecting usability experiences.", + "bbox": [ + 496, + 167, + 890, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conducted thorough experiments on SD. Under strong noise perturbation, the average true positive rate and bit accuracy can exceed 0.99 and 0.97, respectively, validating the superiority of Gaussian Shading in both detection and traceability tasks compared to prior methods. Additionally, experiments on visual quality and image-text similarity serve as indicators of performance preservation in our approach. Lastly, we deliberated on various watermark erasure attacks, affirming the steadfast performance of our watermark in the face of such adversities.", + "bbox": [ + 496, + 303, + 890, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 469, + 640, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Diffusion Models", + "text_level": 1, + "bbox": [ + 500, + 494, + 666, + 508 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by non-equilibrium thermodynamics, Ho et al. [16] introduced the Denoising Diffusion Probabilistic Model (DDPM). DDPM consists of two Markov chains used for adding and removing noise, and subsequent works [8, 11, 15, 25, 28, 30, 32] have adopted this bidirectional chain framework. To reduce computational complexity and improve efficiency, the Latent Diffusion Model (LDM) [30] was designed, in which the diffusion process occurs in a latent space $\\mathcal{Z}$ . To map an image $x \\in \\mathbb{R}^{H \\times W \\times 3}$ to the latent space, the LDM employs an encoder $\\mathcal{E}$ , such that $z_0 = \\mathcal{E}(x) \\in \\mathbb{R}^{h \\times w \\times c}$ . Similarly, to reconstruct an image from the latent space, a decoder $\\mathcal{D}$ is used, such that $x = \\mathcal{D}(z_0)$ . A pretrained LDM can generate images without the encoder $\\mathcal{E}$ . Specifically, a latent representation $z_T$ is first sampled from a standard Gaussian distribution $\\mathcal{N}(0, I)$ . Subsequently, through iterative denoising using methods like DDIM [32], $z_0$ is obtained, and an image can be generated using the decoder: $x = \\mathcal{D}(z_0)$ .", + "bbox": [ + 496, + 518, + 890, + 791 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Image Watermarking", + "text_level": 1, + "bbox": [ + 500, + 800, + 702, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Digital watermarking [38] is an effective means to address copyright protection and content authentication by embedding copyright or traceable identification information within carrier data. Typically, the functionality of a watermark depends on its capacity. For example, a single-bit", + "bbox": [ + 496, + 824, + 890, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12163", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f1602cb88675e865b5ea5a86ef9698a06b6c8a692e398be731b30b89df7b4ea5.jpg", + "image_caption": [ + "Figure 2. Application scenarios for Gaussian Shading." + ], + "image_footnote": [], + "bbox": [ + 176, + 85, + 790, + 200 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "watermark can determine whether an image was generated by a particular diffusion model, i.e., copyright protection; a multi-bit watermark can further determine which user of the diffusion model generated the image, i.e., traceability.", + "bbox": [ + 75, + 223, + 468, + 282 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image watermarking is a method that employs images as carriers for watermarking. Initially, watermark embedding methods primarily focused on the spatial domain [38], but later, to enhance robustness, transform domain watermarking techniques [1, 12, 13, 20, 22, 35, 37] were developed. In recent years, with the advancement of deep learning, researchers have turned their attention to neural networks [21, 39], harnessing their powerful learning capabilities to develop watermarking techniques [18, 19, 26, 36, 43, 50, 51].", + "bbox": [ + 75, + 284, + 468, + 420 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Image Watermarking for Diffusion Models", + "text_level": 1, + "bbox": [ + 76, + 429, + 441, + 445 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Existing Image watermarking methods for the diffusion model [6, 7, 10, 24, 41, 42, 45, 49] can be divided into three categories, as shown in Fig. 1. The image watermarking methods described in the previous section can be applied directly to the images generated by the diffusion model, which is called post-processing-based watermarks [6, 45]. These methods directly modify the image, thus degrading image quality. Recent research endeavors have amalgamated the watermark embedding process with the image generation process to mitigate this issue. Stable Signature [10] fine-tunes the LDM decoder using a pre-trained watermark extractor, facilitating watermark extraction from images produced by the fine-tuned model. Zhao et al. [49] and Liu et al. [24] suggest fine-tuning the diffusion model to implant a backdoor as a watermark, enabling watermark extraction by triggering. These fine-tuning-based approaches enhance the quality of watermarked images but introduce supplementary computational overhead and modify model parameters. Furthermore, Wen et al. [41] introduced the Tree-Ring Watermark, which conveys copyright information by adapting the frequency domain of latent representations to match specific patterns. This method achieves an imperceptible watermark. However, it directly disrupts the Gaussian distribution of noise, limiting the randomness of sampling and resulting in affecting model performance.", + "bbox": [ + 75, + 452, + 470, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 76, + 844, + 174, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we provide an overview of the application scenarios and functionalities in Fig. 2. We then proceed", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to detail the embedding and extraction processes shown in Fig. 3. Finally, we present a mathematical proof of the performance-lossless characteristic of the watermark.", + "bbox": [ + 496, + 223, + 890, + 268 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Application Scenarios", + "text_level": 1, + "bbox": [ + 498, + 277, + 702, + 294 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Scenarios. See Fig. 2, the scenario involves the operator Alice, the thief Carol, and two types of users Bob and Trudy.", + "bbox": [ + 496, + 301, + 890, + 332 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Alice is responsible for training the model, deploying it on the platform, and providing the corresponding API for users, but she does not open-source the code or model weights. Carol does not use Alice's services but steals images generated by her model, claiming ownership of the copyrights. Bob and Trudy, as community users, can utilize the API to generate and disseminate images. While Bob faithfully adheres to the community guidelines, Trudy aims to generate deep fake, and infringing content. To evade detection and traceability, Trudy can employ various data augmentation to modify illicit images.", + "bbox": [ + 496, + 333, + 890, + 498 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Detection. This scenario satisfies the detection (copyright protection) requirement. Alice embeds a single-bit watermark into each generated image. The successful extraction of the watermark from an image serves as evidence of Alice's rightful ownership of the copyright, while also indicating that the image is artificially generated (as opposed to natural images).", + "bbox": [ + 496, + 500, + 890, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Traceability. This scenario fulfills the traceability requirement. Alice allocates a watermark to each user. By extracting the watermark from the illicit content, it enables tracing Trudy, through comparison with the watermark database. Traceability is a higher pursuit than detection and can also achieve copyright protection for different users.", + "bbox": [ + 496, + 604, + 890, + 695 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Details of the statistical tests in both scenarios are shown in Supplementary Material.", + "bbox": [ + 496, + 695, + 890, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Watermark Embedding", + "text_level": 1, + "bbox": [ + 498, + 734, + 718, + 751 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Watermark diffusion. The dimensions of the latent representations are given by $c \\times h \\times w$ , where each dimension can represent $l$ bits of the watermark. Therefore, the watermark capacity becomes $l \\times c \\times h \\times w$ bits. To enhance the robustness of the watermark, we represent the watermark using $\\frac{1}{f_{hw}}$ of the height and width, and $\\frac{1}{f_c}$ of the channel, and replicate the watermark $f_c \\cdot f_{hw}^2$ times. Thus, the watermark $s$ with dimensions $l \\times \\frac{c}{f_c} \\times \\frac{h}{f_{hw}} \\times \\frac{w}{f_{hw}}$ is expanded into a diffused watermark $s^d$ with dimensions $l \\times c \\times h \\times w$ .", + "bbox": [ + 496, + 758, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12164", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e61d5fab0bfa788917f1acf870e9289c022719614719f736758469ebe946cc65.jpg", + "image_caption": [ + "Figure 3. The framework of Gaussian Shading. We utilize a $k$ -bit binary sequence $s$ to represent the watermark. After diffusion and encryption, the watermark can be utilized to drive distribution-preserving sampling, followed by denoising to generate watermarked images $X^s$ . For extraction, it is sufficient to introduce DDIM inversion and the inverse process of all the operations mentioned above." + ], + "image_footnote": [], + "bbox": [ + 156, + 87, + 807, + 309 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The actual watermark capacity is $k = \\frac{l \\times c \\times h \\times w}{f_c \\cdot f_{hw}^2}$ bits.", + "bbox": [ + 76, + 356, + 428, + 377 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Watermark randomization. If we know the distribution of the diffused watermark $s^d$ , we can directly utilize distribution-preserving sampling to obtain the corresponding latent representations $z_T^s$ . However, in practical scenarios, its distribution is always unknown. Hence, we introduce a stream key $K$ to transform $s^d$ into a distribution-unknown randomized watermark $m$ through encryption. Considering the use of computationally secure stream cipher, such as ChaCha20 [3], $m$ follows a uniform distribution, i.e., $m$ is a random binary bit stream.", + "bbox": [ + 75, + 376, + 468, + 525 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Distribution-preserving sampling driven by randomized watermark. When each dimension represents $l$ -bit randomized watermark $m$ , this $l$ bits can be regarded as an integer $y \\in [0, 2^l - 1]$ . Since $m$ is a ciphertext, $y$ follows a discrete uniform distribution, i.e., $p(y) = \\frac{1}{2^l}$ for $y = 0, 1, 2, \\ldots, 2^l - 1$ . Let $f(x)$ denote the probability density function of the Gaussian distribution $\\mathcal{N}(0, I)$ , and $ppf$ denotes the quantile function. We divide $f(x)$ into $2^l$ equal cumulative probability portions. When $y = i$ , the watermarked latent representation $z_T^s$ falls into the $i$ -th interval, which means $z_T^s$ should follow the conditional distribution:", + "bbox": [ + 76, + 526, + 472, + 694 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np \\left(z _ {T} ^ {s} \\mid y = i\\right) = \\left\\{ \\begin{array}{c c} 2 ^ {l} \\cdot f \\left(z _ {T} ^ {s}\\right) & p p f \\left(\\frac {i}{2 ^ {l}}\\right) < z _ {T} ^ {s} \\leq p p f \\left(\\frac {i + 1}{2 ^ {l}}\\right) \\\\ 0 & o t h e r w i s e \\end{array} . \\right. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 75, + 702, + 477, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The probability distribution of $z_T^s$ is given by:", + "bbox": [ + 76, + 750, + 383, + 766 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np \\left(z _ {T} ^ {s}\\right) = \\sum_ {i = 0} ^ {2 ^ {l} - 1} p \\left(z _ {T} ^ {s} \\mid y = i\\right) p (y = i) = f \\left(z _ {T} ^ {s}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 773, + 468, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Eq. (2) indicates that $z_T^s$ follows the same distribution as the randomly sampled latent representation $z_T \\sim \\mathcal{N}(0, I)$ . Next, we elaborate on how this sampling is implemented.", + "bbox": [ + 75, + 824, + 468, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let the cumulative distribution function of $f(x)$ be denoted as $cdf$ . We can obtain the cumulative distribution", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "function of Eq. (1) as follows,", + "bbox": [ + 498, + 357, + 700, + 373 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} F (z _ {T} ^ {s} | y = i) \\\\ = \\left\\{ \\begin{array}{c c} 0 & z _ {T} ^ {s} < p p f \\left(\\frac {i}{2 ^ {l}}\\right) \\\\ 2 ^ {l} \\cdot c d f \\left(z _ {T} ^ {s}\\right) - i & p p f \\left(\\frac {i}{2 ^ {l}}\\right) \\leq z _ {T} ^ {s} \\leq p p f \\left(\\frac {i + 1}{2 ^ {l}}\\right) \\\\ 1 & z _ {T} ^ {s} > p p f \\left(\\frac {i + 1}{2 ^ {l}}\\right) \\end{array} . \\right. \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 504, + 375, + 890, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given $y = i$ , we aim to perform random sampling of $z_{T}^{s}$ within the interval $[ppf(\\frac{i}{2^t}), ppf(\\frac{i + 1}{2^t})]$ . The commonly used method is rejection sampling [4, 17, 47], which can be time-consuming as it requires repeated sampling until $z_{T}^{s}$ falls into the correct interval. Instead, we can utilize the cumulative probability density. When randomly sampling $F(z_{T}^{s}|y = i)$ , the corresponding $z_{T}^{s}$ is naturally obtained through random sampling. Since $F(z_{T}^{s}|y = i)$ takes values in $[0,1]$ , sampling from it is equivalent to sampling from a standard uniform distribution, denoted as $u = F(z_{T}^{s}|y = i) \\sim \\mathcal{U}(0,1)$ . Shift the terms of Eq. (3), and take into account that $cdf$ and $ppf$ are inverse functions, we have", + "bbox": [ + 498, + 457, + 892, + 637 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz _ {T} ^ {s} = p p f \\left(\\frac {u + i}{2 ^ {l}}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 632, + 640, + 890, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Eq. (4) represents the process of sampling the watermarked latent representation $z_{T}^{s}$ driven by the randomized watermark $m$ . To extract the watermark, its inverse map is", + "bbox": [ + 498, + 669, + 890, + 710 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ni = \\left\\lfloor 2 ^ {l} \\cdot c d f \\left(z _ {T} ^ {s}\\right) \\right\\rfloor . \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 713, + 890, + 732 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Image generation. After the sampling process, the watermark is embedded in the latent representation $z_T^s$ , and the subsequent generation process is no different from the regular generation process of SD. Here, we employ the DPM-Solver [25] for iterative denoising of $z_T^s$ . In addition to DPMSolver [25], other continuous-time samplers based on ordinary differential equation (ODE) solvers [32], such as DDIM [32], DEIS [46], PNDM [23], and UniPC [48], can be used too. After obtaining denoised $z_0^s$ , the watermarked image $X^s$ is generated using the decoder $\\mathcal{D}$ : $X^s = \\mathcal{D}(z_0^s)$ .", + "bbox": [ + 496, + 750, + 892, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12165", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Watermark Extraction", + "text_level": 1, + "bbox": [ + 76, + 90, + 289, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "DDIM Inversion. Using the SD encoder $\\mathcal{E}$ , we first restore $X^{\\prime s}$ to the latent space $z_0^{\\prime s} = \\mathcal{E}(X^{\\prime s})$ . Then, we introduce the DDIM inversion [32] to estimate the additive noise. It can be considered that $z_T^{\\prime s} \\approx z_T^s$ . We also observe that although DDIM inversion is derived from DDIM, it can apply to other continuous-time samplers based on ODE solvers.", + "bbox": [ + 75, + 112, + 467, + 204 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Watermark reduction from latent representations. After obtaining $z_T^{\\prime s}$ , according to the inverse transformation defined in Eq. (5), the tensor can be converted into a bit stream $m'$ . Subsequently, $m'$ is decrypted using $K$ to obtain $s'^d$ . Inverse diffusion of the watermark results in $f_c \\cdot f_{hw}^2$ copies of the watermark. Similar to voting, if the bit is set to 1 in more than half of the copies, the corresponding watermark bit is set to 1; otherwise, it is set to 0. This process restores the true binary watermark sequence $s'$ .", + "bbox": [ + 75, + 205, + 467, + 339 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Proof of Lossless Performance", + "text_level": 1, + "bbox": [ + 76, + 347, + 344, + 361 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In prior works, the incorporation of watermark embedding modules inevitably results in a decline in model performance, as typically evaluated using metrics such as Peak Signal-to-Noise Ratio (PSNR) and Fréchet Inception Distance (FID) [14], which are more suitable for assessing post-processing methods. To assess methods that integrate the watermark embedding and generation processes, we propose a definition for the impact of watermark embedding on model performance, drawing on the complexity-theoretic definition of steganographic security [17]. This definition is based on a probabilistic game between a watermarked image $X^{s}$ and a normally generated image $X$ . The tester $\\mathcal{A}$ can use any watermark to drive the sampling process and generate $X^{s}$ , similar to the chosen hidden text attacks [17], which we refer to as chosen watermark tests. The watermarking method is performance-lossless under chosen watermark tests, if for any polynomial-time tester $\\mathcal{A}$ and key $K \\gets \\text{KeyGen}_{\\mathcal{G}(1^{\\rho})}$ , it holds that", + "bbox": [ + 75, + 369, + 467, + 643 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left| \\Pr \\left[ \\mathcal {A} \\left(X ^ {s}\\right) = 1 \\right] - \\Pr \\left[ \\mathcal {A} (X) = 1 \\right] \\right| < \\operatorname {n e g l} (\\rho). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 645, + 467, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $\\rho$ represents the length of the security parameter, such as the key $K$ , and $\\mathsf{negl}(\\rho)$ is a negligible term relative to $\\rho$ .", + "bbox": [ + 76, + 667, + 467, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We prove the statement using a proof by contradiction. First, assume that the watermarked image $X^s$ and the normally generated image $X$ are distinguishable, meaning", + "bbox": [ + 76, + 695, + 467, + 739 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left. \\right.\\left| \\Pr \\left[ \\mathcal {A} \\left(X ^ {s}\\right) = 1 \\right] - \\Pr \\left[ \\mathcal {A} (X) = 1 \\right]\\right| = \\delta , \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 742, + 467, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\delta$ is non-negligible with respect to the key $K$ . Let the iterative denoising process be denoted as $Q(\\cdot)$ , and substitute the LDM encoder $\\mathcal{E}$ into Eq. (7), we have", + "bbox": [ + 76, + 766, + 467, + 808 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\left| \\Pr \\left[ \\mathcal {A} \\left(\\mathcal {E} \\left(Q \\left(z _ {T} ^ {s}\\right)\\right)\\right) = 1 \\mid m = E (K, s ^ {d}) \\right] \\right. \\\\ - \\Pr \\left[ \\mathcal {A} \\left(\\mathcal {E} \\left(Q \\left(z _ {T}\\right)\\right)\\right) = 1 \\mid z _ {T} \\leftarrow \\mathcal {N} (0, I) \\right] | = \\delta , \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 811, + 433, + 849 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the randomized watermark $m$ is obtained by encrypting the diffused watermark $s^d$ using the encryption algorithm $E$ with key $K$ . Note that Eq. (2) contains the fact that", + "bbox": [ + 76, + 854, + 467, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "distribution-preserving sampling driven by randomized watermark and random sampling are equivalent. Therefore, we denote sequence-driven sampling as $S(\\cdot)$ . $z_{T}^{s}$ can naturally be obtained by sampling driven by $m$ , i.e., $z_{T}^{s} = S(m)$ . On the other hand, $z_{T}$ can be considered as obtained by sampling driven by a truly random sequence $r$ of the same length as $m$ , i.e., $z_{T} = S(r)$ . Eq. (8) can be written as", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\left| \\Pr [ \\mathcal {A} (\\mathcal {E} (Q (S (m)))) = 1 | m = E (K, s ^ {d}) ] \\right. \\tag {9} \\\\ - \\Pr \\left[ \\mathcal {A} \\left(\\mathcal {E} (Q (S (r))) = 1 \\right\\rvert \\right] = \\delta . \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 200, + 890, + 238 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sampling $S(\\cdot)$ , denoising $Q(\\cdot)$ , and encoder $\\mathcal{E}$ can be considered as subroutines that the tester $\\mathcal{A}_{\\mathcal{E}, Q, S}$ can use. Thus, Eq. (9) can be simplified,", + "bbox": [ + 496, + 247, + 890, + 289 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\left| \\Pr \\left[ \\mathcal {A} _ {\\varepsilon , Q, S} (m) = 1 \\mid m = E (K, s ^ {d}) \\right] \\right. \\tag {10} \\\\ - \\Pr \\left[ \\mathcal {A} _ {\\mathcal {E}, Q, S} (r) = 1 \\right] | = \\delta . \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 292, + 890, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that $S(\\cdot), Q(\\cdot)$ , and $\\mathcal{E}$ are all polynomial-time programs, so the time taken by the tester $\\mathcal{A}_{\\mathcal{E},Q,S}$ to make the distinction is also polynomial. Eq. (10) essentially states that it is possible to distinguish between $m$ and $r$ in polynomial time. However, we have used the computationally secure stream cipher ChaCha20 [3] in watermark randomization, which means that $m$ as a pseudorandom sequence cannot be distinguished from a truly random sequence in polynomial time. Eq. (10) contradicts the computational security property of ChaCha20 [3]. Therefore, Eq. (10) is not valid, leading us back to our initial assumption that Eq. (7) is also not valid. This implies that the watermarked image $X^s$ and the normally generated image $X$ are indistinguishable in polynomial time. Hence, Gaussian Shading is performance-lossless under chosen watermark tests.", + "bbox": [ + 496, + 340, + 890, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 580, + 630, + 597 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This section focuses on experimental analysis, including details of the experimental setup, performance evaluation of Gaussian Shading, comparison with baseline methods, ablation experiments, and potential attacks.", + "bbox": [ + 496, + 604, + 890, + 666 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 674, + 715, + 689 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SD models. In this paper, we focus on text-to-image LDM, hence we select SD [30] provided by huggingface. We evaluate Gaussian Shading as well as baseline methods, using three versions of SD: V1.4, V2.0, and V2.1. The size of the generated images is $512 \\times 512$ , and the latent space dimension is $4 \\times 64 \\times 64$ . During inference, we employ the prompt from Stable-Diffusion-Prompt3, with a guidance scale of 7.5. We sample 50 steps using DPMSolver [25]. Considering that users tend to propagate the generated images without retaining the corresponding prompts, we use an empty prompt for inversion, with a scale of 1. We perform 50 steps of inversion using DDIM inversion [32].", + "bbox": [ + 496, + 696, + 890, + 878 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "3Stable-Diffusion-Prompts", + "bbox": [ + 517, + 886, + 661, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12166", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/113f92213137cf7560fc650a7bc87aa7fa6167595fbc53eb9e6c952236ba331c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 84, + 87, + 158, + 143 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/391c5ea77b52a7c6e9ad6089279876a9caf13725122d62529f2361a9e94e4313.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 163, + 88, + 233, + 143 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2bc54b6f2c460407df5ec3c81d88e2f7a9f96762a6c17332b8cce2eca805f13c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 235, + 88, + 310, + 143 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a00616e79178046b18344bdef487b7c122716a955f897da335aa553de66f550e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 312, + 88, + 385, + 143 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7a1e6ed37b87a66ec8e8ddf1676f6702fc25431a7fb153cb16423fa84918c87a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 388, + 88, + 462, + 143 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/78308b2122f914f4fac6cf0abe8e2988f0188e26a62c9f26c1470943e865b0d4.jpg", + "image_caption": [ + "(f)" + ], + "image_footnote": [], + "bbox": [ + 81, + 159, + 156, + 215 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3c04e4d0e837a921c04aabbbd4ebaf3cb8aa7899b3418d1658e7736585bc880e.jpg", + "image_caption": [ + "(b)", + "(g)" + ], + "image_footnote": [], + "bbox": [ + 156, + 159, + 230, + 215 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f4b394ed8fb9227ff91deaab30c954771c7a2118d94f43320d95ddb0bf42850b.jpg", + "image_caption": [ + "(h)" + ], + "image_footnote": [], + "bbox": [ + 233, + 159, + 307, + 215 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/44eee57d50c3eeced27c41c772a82e00c47783d7cea6ed9483617ca300b13e12.jpg", + "image_caption": [ + "(d)", + "(i)", + "Figure 4. Watermarked image is attacked by different noise. (a) Watermarked image. (b) JPEG, $QF = 25$ . (c) $60\\%$ area Random Crop (RandCr). (d) $80\\%$ area Random Drop (RandDr). (e) Gaussian Blur, $r = 4$ (GauBlur). (f) Median Filter, $k = 7$ (MedFilter). (g) Gaussian Noise, $\\mu = 0$ , $\\sigma = 0.05$ (GauNoise). (h) Salt and Pepper Noise, $p = 0.05$ (S&PNoise). (i) $25\\%$ Resize and restore (Resize). (j) Brightness, factor $= 6$ ." + ], + "image_footnote": [], + "bbox": [ + 310, + 159, + 383, + 215 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e724f5fe89325981724032fd706b40f092a41863e3b1de3c6d07a334bb23ccca.jpg", + "image_caption": [ + "(j)" + ], + "image_footnote": [], + "bbox": [ + 398, + 159, + 457, + 215 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Watermarking methods. In the main experiments, the settings for Gaussian Shading are $f_{c} = 1$ , $f_{hw} = 8$ , $l = 1$ , resulting in an actual capacity of 256 bits. We select five baseline methods: three officially used by SD, namely Dwt-Dct [6], DwtDctSvd [6], and RivaGAN [45], a multi-bit watermarking called Stable Signature [10], and a train-free invisible watermarking called Tree-Ring [41].", + "bbox": [ + 75, + 330, + 468, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Robustness evaluation To evaluate the robustness, we select nine representative types of noise shown in Fig. 4. We conduct experiments following the noise strength in Fig. 4.", + "bbox": [ + 75, + 439, + 468, + 483 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics. In the detection scenario, we calculate the true positive rate (TPR) corresponding to a fixed false positive rate (FPR). In the traceability scenario, we calculate the bit accuracy. To measure the bias in model performance, we compute the FID [14] and CLIP-Score [29] for 10 batches of watermarked images and perform a $t$ -test on the mean FID and CLIP-Score compared to that of watermark-free images.", + "bbox": [ + 75, + 484, + 468, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "All experiments are conducted using the PyTorch 1.13.0 framework, running on a single RTX 3090 GPU.", + "bbox": [ + 75, + 606, + 468, + 636 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Performance of Gaussian Shading", + "text_level": 1, + "bbox": [ + 76, + 648, + 375, + 665 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Detection. In the detection scenario, we consider Gaussian Shading as a single-bit watermark, with a fixed watermark $s$ . We approximate the FPR to be controlled at $10^{0}, 10^{-1}, \\ldots, 10^{-13}$ , calculate the corresponding threshold $\\tau$ , and test the TPR on 1,000 watermarked images To mitigate the effects of randomness, we perform 5 trials with different $s$ and compute the average TPR. See Fig. 5a, when the FPR is controlled at $10^{-13}$ , the TPR remains at least 0.99 for eight out of the nine cases. Although the TPR for Brightness is only 0.953, it is still a promising result.", + "bbox": [ + 75, + 672, + 468, + 823 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Traceability. In this scenario, Gaussian Shading serves as a multi-bit watermark. Assuming Alice provides services to $N$ users, Alice needs to allocate one watermark for each user. In our experiments, we assume that $N' = 1,000$ users generate images, with each user generating 10 images, re", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/774162d867b79361428a6cb94c725181280f963f31253856bb88a5940fa1d48c.jpg", + "image_caption": [ + "(a) Detection results.", + "Figure 5. Performance of Gaussian Shading." + ], + "image_footnote": [], + "bbox": [ + 516, + 88, + 671, + 178 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/aa640412f42f561a96912774a9dce03db8c6cc463236ac6429791ff69841e72a.jpg", + "image_caption": [ + "(b) Traceability results." + ], + "image_footnote": [], + "bbox": [ + 712, + 88, + 870, + 178 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "sulting in a dataset of 10,000 watermarked images.", + "bbox": [ + 500, + 212, + 836, + 227 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During testing, we calculate the threshold $\\tau$ to control the FPR at $10^{-6}$ . Note that when computing traceability accuracy, we need to consider two types of errors: false positives, where watermarked images are not detected, and traceability errors, where watermarked images are detected but attributed to the wrong user. Therefore, we first determine whether the image contains a watermark. If it does, we calculate the number of matching bits $Acc$ with all $N$ users on the platform. The user with the highest $Acc$ is considered the one who generated the image. Finally, we verify whether the correct user has been traced. When $N > N'$ , it can be assumed that some users have been assigned a watermark but have not generated any images.", + "bbox": [ + 496, + 229, + 890, + 425 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "See Fig. 5b, when $N = 10^{6}$ , Gaussian Shading exhibits almost perfect traceability in seven cases. Although the traceability accuracy for Brightness is only $95.47\\%$ , if a user generates two images, the probability of successfully tracing him is still no less than $99\\%$ .", + "bbox": [ + 496, + 425, + 890, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparison to Baselines", + "text_level": 1, + "bbox": [ + 500, + 512, + 725, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we compare the performance of Gaussian Shading with baselines on SD V1.4, V2.0, and V2.1. We use our implementations for each method, see details in Supplementary Material.", + "bbox": [ + 496, + 537, + 890, + 597 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct tests on 1,000 generated images for each method respectively. See Tab. 1. Gaussian Shading exhibits strong robustness and significantly outperforms baselines in both scenarios. In terms of bit accuracy, it surpasses the best-performing baseline by approximately $7\\%$ . This can be attributed to the extensive diffusion of the watermark throughout the entire latent space, establishing a profound binding between the watermark and the image semantics.", + "bbox": [ + 496, + 598, + 890, + 718 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To measure the performance bias introduced by the watermark embedding, we apply a $t$ -test to evaluate. The hypotheses are $H_0: \\mu_s = \\mu_0, H_1: \\mu_s \\neq \\mu_0$ , where $\\mu_s$ and $\\mu_0$ represent the average FID [14] or CLIP-Score [29] of multiple sets of watermarked and watermark-free images, respectively. A lower $t$ -value indicates a higher probability that $H_0$ holds. If the $t$ -value is larger than a threshold, $H_0$ is rejected, and model performance is considered to have been affected. See Tab. 1, Gaussian Shading achieves the smallest $t$ -value, which indirectly reflects its performance-lossless characteristic. For a detailed analysis of the $t$ -test, please refer to the Supplementary Material.", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12167", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/236704faa0a707312d5346244dee764e1ac373bce22c3ad9b63c8f8871350458.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsMetrics
TPR (Clean)TPR (Adversarial)Bit Acc. (Clean)Bit Acc. (Adversarial)FID (t-value ↓)CLIP-Score (t-value ↓)
Stable Diffusion----25.23±.180.3629±.0006
DwtDct [6]0.825/0.881/0.8660.172/0.178/0.1730.8030/0.8059/0.80230.5696/0.5671/0.562224.97±.19 (3.026)0.3617±.0007 (3.045)
DwtDctSvd [6]1.000/1.000/1.0000.597/0.594/0.5990.9997/0.9987/0.99870.6920/0.6868/0.690524.45±.22 (8.253)0.3609±.0009 (4.452)
RivaGAN [45]0.920/0.945/0.9630.697/0.697/0.7060.9762/0.9877/0.99210.8986/0.9124/0.901924.24±.16 (12.29)0.3611±.0009 (4.259)
Tree-Ring [41]1.000/1.000/1.0000.894/0.898/0.906--25.43±.13 (2.581)0.3632±.0006 (0.8278)
Stable Signature [10]1.000/1.000/1.0000.502/0.505/0.4960.9987/0.9978/0.99790.7520/0.7472/0.750025.45±.18 (2.477)0.3622±.0027 (0.7066)
Ours1.000/1.000/1.0000.997/0.998/0.9960.9999/0.9999/0.99990.9753/0.9749/0.972425.20±.22 (0.3567)0.3631±.0005 (0.6870)
", + "bbox": [ + 81, + 89, + 893, + 213 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Comparison results. We control the FPR at $10^{-6}$ , and evaluate the TPR and bit accuracy for SD V1.4/V2.0/V2.1. To assess the bias in model performance, we conduct a $t$ -test on SD V2.1. Adversarial here refers to the average performance of a series of noises. Additional results can be found in Supplementary Material.", + "bbox": [ + 75, + 215, + 892, + 257 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/db4c4943a11b71c4f21ace1b07c62f59f4d4daf753525c2cda36cf1450e25b33.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 102, + 258, + 251, + 347 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b84c926be2f4085637d5b5c005f44ffb54fa0ad2895819f40434dfc0a5df6b94.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 254, + 258, + 405, + 347 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/87353aed10bc3304f8afcd28ee3be310f682ffb1cfa196dbb473192af11974be.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 410, + 258, + 558, + 347 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1849f464e6b319de72170104ae614ecc91f64b614bc08e5901c5cb9716282939.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 563, + 258, + 712, + 347 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/acc123a5fbb8846a038efd042a4ce1bbab5e99ba593033d37ec567f62895b3e5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 717, + 258, + 867, + 347 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d14af2497e22ab79cac4adf3bc18bb022c2d364c466615121b42a3e96c298b49.jpg", + "image_caption": [ + "(a) Guidance scales.", + "(f) Median Filter." + ], + "image_footnote": [], + "bbox": [ + 98, + 363, + 248, + 450 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/336e0f66502a768f3fd7a00d31195f7d9a2a7a4135001346c9868fbcbec5e3bb.jpg", + "image_caption": [ + "(b) JPEG", + "(g) Gaussian Noise." + ], + "image_footnote": [], + "bbox": [ + 253, + 363, + 401, + 450 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ad94a8eb28303917d25d49dfcd8b99f9fbf0ec49a6a48c1c88eff66d6f002d36.jpg", + "image_caption": [ + "(c) Random Crop.", + "(h) Salt and Pepper Noise.", + "Figure 6. Ablation studies." + ], + "image_footnote": [], + "bbox": [ + 406, + 363, + 553, + 450 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d5c033138c9afd948058844db3e2be752c786e291bc9636a0f800e456fc3b839.jpg", + "image_caption": [ + "(d) Random Drop.", + "(i) Resize." + ], + "image_footnote": [], + "bbox": [ + 560, + 363, + 707, + 450 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/66a5694f8e5461951c77626d1771d39b23d6d145c3288595c87e27d5bc80a502.jpg", + "image_caption": [ + "(e) Gaussian Blur.", + "(j) Brightness." + ], + "image_footnote": [], + "bbox": [ + 714, + 363, + 862, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 483, + 240, + 497 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we conduct comprehensive ablation experiments on SD V2.1 to demonstrate hyperparameter selection. Unless specified, we generate 1,000 images and test the TPR and the bit accuracy with a theoretical FPR of $10^{-6}$ .", + "bbox": [ + 75, + 506, + 468, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Watermark capacity. The watermark capacity is determined by three parameters: channel diffusion factor $f_{c}$ , height-width diffusion factor $f_{hw}$ , and embedding rate $l$ . See Tab. 2, to balance the capacity and robustness of Gaussian Shading, we chose $f_{c} = 1$ and $f_{hw} = 8$ . After fixing $f_{c}$ and $f_{hw}$ , we vary $l$ to examine if it could enhance the capacity, and additional results can be found in Supplementary Material. Considering all factors, we determine that the optimal solution is $f_{c} = 1$ , $f_{hw} = 8$ , and $l = 1$ , resulting in a watermark capacity of 256 bits.", + "bbox": [ + 75, + 566, + 468, + 717 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Sampling methods. To validate the generalization, we select five commonly used sampling methods, all continuous-time samplers based on ODE solvers [32]. See Tab. 3, all of them exhibit excellent performance with a bit accuracy of approximately $97\\%$ against noises.", + "bbox": [ + 75, + 719, + 468, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of the inversion step. In practice, the inference step is often unknown, which introduces a mismatch with the inversion step. See Tab. 4, such mismatch introduces minimal loss in accuracy. Considering the high efficiency of existing samplers, the inference step generally does not exceed 50. Therefore, we set the inversion step to 50.", + "bbox": [ + 75, + 795, + 468, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Guidance scales. Given diverse user preferences for image", + "bbox": [ + 76, + 886, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "prompt alignment, larger guidance scales ensure faithful adherence to prompts, while smaller scales grant the model greater creative freedom. In SD, the guidance scale is typically selected from the range of [5, 15]. Hence, experiments cover the range of 2 to 18. For the inversion, an empty prompt is used for guidance, and the guidance scale is fixed at 1, assuming unknown information during extraction. In Fig. 6a, the bit accuracy of Gaussian Shading surpasses $99.9\\%$ , showing its reliability in real-world-like scenes.", + "bbox": [ + 496, + 483, + 890, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Noise intensities. To further test the robustness, we conduct experiments using different intensities of noises. See Figs. 6b to 6j, for Random Crop and Gaussian Noise, performance declines significantly with higher intensities. However, for the other seven types of noise, even at high intensities, the bit accuracy remains approximately $80\\%$ .", + "bbox": [ + 496, + 619, + 890, + 710 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Attacks against Gaussian Shading", + "text_level": 1, + "bbox": [ + 500, + 724, + 797, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We consider two malicious attacks: compression attack, where the attacker employs a neural network to compress watermarked images, and inversion attack, assuming the attacker is aware of the watermark embedding method, enabling them to modify the image's latent representations.", + "bbox": [ + 496, + 750, + 890, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Compression attack. We utilize popular auto-encoders [2, 5, 9, 30] to compare Stable Signature (SS) with Gaussian Shading across various compression rates. Additionally, we assess the compression quality through the PSNR between the compressed and watermarked images. See Figs. 7a", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12168", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/842d3e3a041d7f93a6daa14896dd108074cedb783e2deef6d14fa76b70826991.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Noisefc-fhw (k bits)
1-2 (4096)4-1 (4096)1-4 (1024)4-2 (1024)1-8 (256)4-4 (256)1-16 (64)4-8 (64)
None0.94130.93800.99850.99800.99990.99991.00001.0000
Adversarial0.73020.72380.87690.86140.97240.96710.99590.9953
", + "bbox": [ + 166, + 88, + 802, + 157 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2d227749735613639e5af587adfd7bd99ce2aa9f8c52ba03f1f958ba386c37a2.jpg", + "table_caption": [ + "Table 2. Bit accuracy with different factors ${f}_{c}$ and ${f}_{hw}$ ,where $l = 1$ . Additional results can be found in Supplementary Material." + ], + "table_footnote": [], + "table_body": "
NoiseSampling Methods
DDIM [32]UniPC [48]PNDM [23]DEIS [46]DPMSolver [25]
None0.99991.00001.00000.99990.9999
Adversarial0.97060.96280.97210.97150.9724
", + "bbox": [ + 91, + 181, + 455, + 263 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e8ef5d3d10c7dea97f508e187dda37875eec9b1aa7a2804c24e59f57738f74d4.jpg", + "table_caption": [ + "Table 3. Bit accuracy with different sampling methods. Additional results can be found in Supplementary Material. These methods differ only in accuracy and order. DDIM is a first-order estimate of the ODE. Accordingly, DDIM inversion ensures a lower bound on the accuracy of the inversion process. Therefore, it can naturally be applied to higher-order and higher-accuracy methods." + ], + "table_footnote": [], + "table_body": "
Inference StepInversion Step
102550100
100.99990.99990.99990.9999
250.99980.99991.00001.0000
500.99950.99970.99990.9999
1000.99940.99960.99990.9999
", + "bbox": [ + 135, + 351, + 410, + 444 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Bit accuracy with different inference and inversion step. and 7b, Gaussian Shading significantly outperforms Stable Signature. This is because Gaussian Shading diffuses the watermark across the entire semantic space of images, while Stable Signature relies solely on the image texture.", + "bbox": [ + 75, + 445, + 468, + 522 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Inversion attack. Assuming the attacker is aware of the embedding method, a more effective approach to erasing is through inversion to obtain latent representations and subsequently modify them. We validate the robustness against such attacks. Importantly, our experiments assume the strongest attacker capability of using the same model as Alice for precise inversion. In real-world scenarios, where the watermark embedding is not publicly available, the attacker's capabilities would be weaker.", + "bbox": [ + 75, + 523, + 468, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Specifically, we perform inversion to obtain latent representations and randomly flip a certain rate of them. Using the flipped latent representations, we regenerate the images and extract the watermark. See Fig. 7c. the watermark can still be reliably extracted when the flipping rate (FR) is less than 0.4. At high FRs, significant changes in images are observed. Although the watermark cannot be accurately extracted, we consider the image transformed into a different one, resulting in the content not intended to be protected.", + "bbox": [ + 75, + 657, + 468, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "From another perspective, the attacker can launch a forgery attack by performing inversion on an innocuous image from Bob and generating harmful content using a different prompt. See Fig. 7c, when the FR is 0, Alice can accurately trace Bob based on the forgeries, enabling the attacker to successfully frame Bob. Therefore, protecting the model from leakage is crucial for operators.", + "bbox": [ + 75, + 795, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/79029d1b4edd943077dd20adea6a00b115b845037db4d141053738cfbf85dc13.jpg", + "image_caption": [ + "(a) Detection results." + ], + "image_footnote": [], + "bbox": [ + 501, + 181, + 629, + 257 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b1e3e8caa276f4aa3216dcbc155f939b4c6b9469e3389d584c640c634f685772.jpg", + "image_caption": [ + "(b) Traceability results.", + "Figure 7. Performance of Gaussian Shading under Malicious Attack, where (a) and (b) are under compression attack and (c) is under inversion attack." + ], + "image_footnote": [], + "bbox": [ + 633, + 181, + 758, + 258 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9d4b715f4373508fe44f20ea037ed3cd283cdaeef1e62d4c41cf8b7d50e169ba.jpg", + "image_caption": [ + "(c) Inversion attack." + ], + "image_footnote": [], + "bbox": [ + 763, + 181, + 888, + 257 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Limitations", + "text_level": 1, + "bbox": [ + 500, + 323, + 622, + 339 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Despite extensive experimental validation of Gaussian Shading's superior performance, our work still has certain limitations. Firstly, the usage scenarios are restricted due to the reliance on DDIM inversion [32], which necessitates the utilization of continuous-time samplers based on ODE solvers [32] like DPMSolver [25]. Secondly, Gaussian Shading employs stream ciphers, necessitating proper key usage and management on the deployment platform. Additionally, we assume that the model is not publicly accessible, and only operators can verify the watermark, providing a certain level of protection against white-box attacks and ensuring security. However, if a legitimate third party requires watermark verification, cooperation from the operators becomes necessary. Lastly, Gaussian Shading is vulnerable to forgery attacks, emphasizing the importance for operators to safeguard the model parameters.", + "bbox": [ + 496, + 349, + 890, + 592 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion and Future Work", + "text_level": 1, + "bbox": [ + 500, + 604, + 769, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose Gaussian Shading, a provably performance-lossless watermarking applied to diffusion models. Compared to baseline methods, Gaussian Shading offers simplicity and effectiveness by making a simple modification in the sampling process of the initial latent representation. Extensive experiments validate the superior performance in both detection and traceability scenarios. To our knowledge, we are the first to propose and implement a performance-lossless approach in image watermarking.", + "bbox": [ + 496, + 628, + 890, + 765 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Regarding future work, we will introduce more efficient inversion methods [40, 44] and include a wider range of sampling methods. Additionally, careful consideration should be given to counteracting forgery attacks.", + "bbox": [ + 496, + 766, + 890, + 825 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work was supported in part by the Natural Science Foundation of China under Grant U2336206, 62102386, 62072421, 62372423, and 62121002.", + "bbox": [ + 496, + 839, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12169", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ali Al-Haj. Combined dwt-dct digital image watermarking. Journal of computer science, 3(9):740-746, 2007. 3", + "[2] Johannes Ballé, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. arXiv preprint arXiv:1802.01436, 2018. 7", + "[3] Daniel J Bernstein et al. Chacha, a variant of salsa20. In Workshop record of SASC, pages 3-5. Citeseer, 2008. 4, 5", + "[4] Kejiang Chen, Hang Zhou, Hanqing Zhao, Dongdong Chen, Weiming Zhang, and Nenghai Yu. Distribution-preserving steganography based on text-to-speech generative models. IEEE Transactions on Dependable and Secure Computing, 19(5):3343-3356, 2021. 4", + "[5] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned image compression with discretized gaussian mixture likelihoods and attention modules. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7939-7948, 2020. 7", + "[6] Ingemar Cox, Matthew Miller, Jeffrey Bloom, Jessica Fridrich, and Ton Kalker. Digital watermarking and steganography. Morgan Kaufmann, 2007. 2, 3, 6, 7", + "[7] Yingqian Cui, Jie Ren, Han Xu, Pengfei He, Hui Liu, Lichao Sun, and Jiliang Tang. Diffusionshield: A watermark for copyright protection against generative diffusion models. arXiv preprint arXiv:2306.04642, 2023. 2, 3", + "[8] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2", + "[9] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 7", + "[10] Pierre Fernandez, Guillaume Couairon, Hervé Jégou, Matthijs Douze, and Teddy Furon. The stable signature: Rooting watermarks in latent diffusion models. arXiv preprint arXiv:2303.15435, 2023. 2, 3, 6, 7", + "[11] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10696-10706, 2022. 2", + "[12] Huiping Guo and Nicolas D Georganas. Digital image watermarking for joint ownership. In Proceedings of the tenth ACM international conference on Multimedia, pages 362-371, 2002. 3", + "[13] Mohamed Hamidi, Mohamed El Haziti, Hocine Cherifi, and Mohammed El Hassouni. Hybrid blind robust image watermarking technique based on dft-dct and arnold transform. Multimedia Tools and Applications, 77:27181-27214, 2018. 3", + "[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5, 6" + ], + "bbox": [ + 78, + 114, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2", + "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2", + "[17] Nicholas J Hopper, John Langford, and Luis Von Ahn. Provably secure steganography. In Advances in Cryptology—CRYPTO 2002: 22nd Annual International Cryptology Conference Santa Barbara, California, USA, August 18–22, 2002 Proceedings 22, pages 77–92. Springer, 2002. 4, 5", + "[18] Zhaoyang Jia, Han Fang, and Weiming Zhang. Mbrs: Enhancing robustness of dnn-based watermarking by minibatch of real and simulatedJPEG compression. In Proceedings of the 29th ACM international conference on multimedia, pages 41-49, 2021. 3", + "[19] Varsha Kishore, Xiangyu Chen, Yan Wang, Boyi Li, and Kilian Q Weinberger. Fixed neural network steganography: Train the images, not the network. In International Conference on Learning Representations, 2021. 3", + "[20] Deepa Kundur and Dimitrios Hatzinakos. A robust digital image watermarking method using wavelet-based fusion. In Proceedings of International Conference on Image Processing, pages 544-547. IEEE, 1997. 3", + "[21] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 3", + "[22] Sunil Lee, Chang D Yoo, and Ton Kalker. Reversible image watermarking based on integer-to-integer wavelet transform. IEEE Transactions on information forensics and security, 2 (3):321-330, 2007. 3", + "[23] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. arXiv preprint arXiv:2202.09778, 2022. 4, 8", + "[24] Yugeng Liu, Zheng Li, Michael Backes, Yun Shen, and Yang Zhang. Watermarking diffusion model. arXiv preprint arXiv:2305.12502, 2023. 2, 3", + "[25] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 2, 4, 5, 8", + "[26] Xiyang Luo, Ruohan Zhan, Huiwen Chang, Feng Yang, and Peyman Milanfar. Distortion agnostic deep watermarking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13548-13557, 2020. 3", + "[27] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1", + "[28] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 2", + "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry," + ], + "bbox": [ + 501, + 92, + 890, + 902 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12170", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6", + "[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 5, 7", + "[31] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 1", + "[32] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2020. 2, 4, 5, 7, 8", + "[33] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019.", + "[34] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2020. 1", + "[35] Srdjan Stankovic, Irena Orovic, and Nikola Zaric. An application of multidimensional time-frequency analysis as a base for the unified watermarking approach. IEEE Transactions on Image Processing, 19(3):736-745, 2009. 3", + "[36] Matthew Tancik, Ben Mildenhall, and Ren Ng. Stegastamp: Invisible hyperlinks in physical photographs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2117-2126, 2020. 3", + "[37] Min-Jen Tsai, Kuang-Yao Yu, and Yi-Zhang Chen. Joint wavelet and spatial transformation for digital watermarking. IEEE Transactions on Consumer Electronics, 46(1):237, 2000. 3", + "[38] Ron G Van Schyndel, Andrew Z Tirkel, and Charles F Osborne. A digital watermark. In Proceedings of 1st international conference on image processing, pages 86-90. IEEE, 1994. 2, 3", + "[39] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3", + "[40] Bram Wallace, Akash Gokul, and Nikhil Naik. Edict: Exact diffusion inversion via coupled transformations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22532-22541, 2023. 8", + "[41] Yuxin Wen, John Kirchenbauer, Jonas Geiping, and Tom Goldstein. Tree-ring watermarks: Fingerprints for diffusion images that are invisible and robust. arXiv preprint arXiv:2305.20030, 2023. 2, 3, 6, 7", + "[42] Cheng Xiong, Chuan Qin, Guorui Feng, and Xinpeng Zhang. Flexible and secure watermarking for latent diffusion model. In Proceedings of the 31st ACM International Conference on Multimedia, pages 1668-1676, 2023. 2, 3" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Chaoning Zhang, Philipp Benz, Adil Karjauv, Geng Sun, and In So Kweon. Udh: Universal deep hiding for steganography, watermarking, and light field messaging. Advances in Neural Information Processing Systems, 33:10223-10234, 2020. 3", + "[44] Jiaxin Zhang, Kamalika Das, and Sricharan Kumar. On the robustness of diffusion inversion in image manipulation. In ICLR 2023 Workshop on Trustworthy and Reliable Large-Scale Machine Learning Models, 2023. 8", + "[45] Kevin Alex Zhang, Lei Xu, Alfredo Cuesta-Infante, and Kalyan Veeramachaneni. Robust invisible video watermarking with attention. arXiv preprint arXiv:1909.01285, 2019. 2,3,6,7", + "[46] Qinsheng Zhang and Yongxin Chen. Fast sampling of diffusion models with exponential integrator. arXiv preprint arXiv:2204.13902, 2022. 4, 8", + "[47] Weiming Zhang, Kejiang Chen, and Nenghai Yu. Provable secure steganography: Theory, application and prospects. Journal of Cybersecurity, 1:38-46, 2023. 4", + "[48] Wenliang Zhao, Lujia Bai, Yongming Rao, Jie Zhou, and Jiwen Lu. Unipc: A unified predictor-corrector framework for fast sampling of diffusion models. arXiv preprint arXiv:2302.04867, 2023. 4, 8", + "[49] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Ngai-Man Cheung, and Min Lin. A recipe for watermarking diffusion models. arXiv preprint arXiv:2303.10137, 2023. 2, 3", + "[50] Xin Zhong, Pei-Chi Huang, Spyridon Mastorakis, and Frank Y Shih. An automated and robust image watermarking scheme based on deep neural networks. IEEE Transactions on Multimedia, 23:1951-1961, 2020. 3", + "[51] Jiren Zhu, Russell Kaplan, Justin Johnson, and Li Fei-Fei. Hidden: Hiding data with deep networks. In Proceedings of the European conference on computer vision (ECCV), pages 657–672, 2018. 3" + ], + "bbox": [ + 501, + 92, + 893, + 571 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12171", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Gaussian Shading_ Provable Performance-Lossless Image Watermarking for Diffusion Models/cf093428-bde9-4e71-b54b-2506850cecda_model.json b/2024/Gaussian Shading_ Provable Performance-Lossless Image Watermarking for Diffusion Models/cf093428-bde9-4e71-b54b-2506850cecda_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6363df82bf8acf42eebc6796dc7d894530910b81 --- /dev/null +++ b/2024/Gaussian Shading_ Provable Performance-Lossless Image Watermarking for Diffusion Models/cf093428-bde9-4e71-b54b-2506850cecda_model.json @@ -0,0 +1,2728 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.1, + 0.13, + 0.872, + 0.175 + ], + "angle": 0, + "content": "Gaussian Shading: Provable Performance-Lossless Image Watermarking for Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.203, + 0.882, + 0.258 + ], + "angle": 0, + "content": "Zijin Yang\\(^{1}\\), Kai Zeng\\(^{1}\\), Kejiang Chen\\(^{1,*}\\), Han Fang\\(^{2}\\), Weiming Zhang\\(^{1}\\), Nenghai Yu\\(^{1}\\), \\(^{1}\\)Anhui Province Key Laboratory of Digital Security, University of Science and Technology of China \\(^{2}\\)National University of Singapore" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.26, + 0.862, + 0.275 + ], + "angle": 0, + "content": "{bsmhmmlf@mail.,zk0128@mail.,chenkj@,zhangwm@,ynh@}ustc.edu.cn fanghan@nus.edu.sg" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.309, + 0.314, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.342, + 0.474, + 0.735 + ], + "angle": 0, + "content": "Ethical concerns surrounding copyright protection and inappropriate content generation pose challenges for the practical implementation of diffusion models. One effective solution involves watermarking the generated images. However, existing methods often compromise the model performance or require additional training, which is undesirable for operators and users. To address this issue, we propose Gaussian Shading, a diffusion model watermarking technique that is both performance-lossless and training-free, while serving the dual purpose of copyright protection and tracing of offending content. Our watermark embedding is free of model parameter modifications and thus is plug-and-play. We map the watermark to latent representations following a standard Gaussian distribution, which is indistinguishable from latent representations obtained from the non-watermarked diffusion model. Therefore we can achieve watermark embedding with lossless performance, for which we also provide theoretical proof. Furthermore, since the watermark is intricately linked with image semantics, it exhibits resilience to lossy processing and erasure attempts. The watermark can be extracted by Denoising Diffusion Implicit Models (DDIM) inversion and inverse sampling. We evaluate Gaussian Shading on multiple versions of Stable Diffusion, and the results demonstrate that Gaussian Shading not only is performance-lossless but also outperforms existing methods in terms of robustness." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.763, + 0.21, + 0.779 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.787, + 0.47, + 0.879 + ], + "angle": 0, + "content": "Diffusion models [16, 31-34] signify a noteworthy leap forward in image generation. These well-trained diffusion models, especially commercial diffusion models like Stable Diffusion (SD) [30], Glide [27], and Muse AI [30], enable individuals with diverse backgrounds to create high-quality images effortlessly. However, this raises concerns about" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.308, + 0.869, + 0.552 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.553, + 0.895, + 0.621 + ], + "angle": 0, + "content": "Figure 1. Existing watermarking frameworks can be divided into three categories: post-processing-based, fine-tuning-based, and latent-representation-based Tree-Ring. Our method also relies on latent representations but achieves performance-lossless without altering the distribution." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.627, + 0.893, + 0.656 + ], + "angle": 0, + "content": "intellectual property and whether diffusion models will be stolen or resold twice." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.657, + 0.893, + 0.854 + ], + "angle": 0, + "content": "On the other hand, the ease of generating realistic images raises concerns about potentially misleading content generation. For example, on May 23, 2023, a Twitter-verified user named Bloomberg Feed posted a tweet titled \"Large explosion near the Pentagon complex in Washington DC-initial report,\" along with a synthetic image. This tweet led to multiple authoritative media accounts sharing it, even causing a brief impact on the stock market1. On October 30, 2023, White House issued an executive order on AI security, emphasizing the need to protect Americans from AI-enabled fraud and deception by establishing standards and best practices for detecting AI-generated content and authenticating official content2. The urgency of labeling gen" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.221, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.863, + 0.766, + 0.876 + ], + "angle": 0, + "content": "1Fake image of Pentagon explosion on Twitter" + }, + { + "type": "page_footnote", + "bbox": [ + 0.502, + 0.877, + 0.891, + 0.901 + ], + "angle": 0, + "content": "2FACT SHEET: President Biden Issues Executive Order on Safe, Secure, and Trustworthy Artificial Intelligence" + }, + { + "type": "list", + "bbox": [ + 0.502, + 0.863, + 0.891, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12162" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "erated content for copyright authentication and prevention of misuse is evident." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.124, + 0.471, + 0.412 + ], + "angle": 0, + "content": "Watermarking is highlighted as a fundamental method for labeling generated content, as it embeds watermark information within the generated image, allowing for subsequent copyright authentication and the tracking of false content. Existing watermarking methods for the diffusion model can be divided into three categories, as shown in Fig. 1. Post-processing-based watermarks [6, 45] adjust robust image features to embed watermarks, thereby directly altering the image and degrading its quality. To mitigate this concern, recent research endeavors propose fine-tuning-based methods [7, 10, 24, 42, 49], which amalgamate the watermark embedding process with the image generation process. Intuitively, these methods need to modify model parameters, introducing supplementary computational overhead. Recently, Wen et al. [41] proposed the latent-representation-based Tree-Ring watermark, which conveys information by adapting the latent representations to match specific patterns. However, it restricts the randomness of sampling, which impacts generative performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.413, + 0.47, + 0.534 + ], + "angle": 0, + "content": "Through the above analysis, we can find that these methods compromise model performance to embed watermarks. In practical applications, model performance is paramount for both business interests and user experience. Substantial resource investment is often necessary to pursue enhanced model performance. This leads to a fundamental question: Can watermarks be embedded without compromising model performance?" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.537, + 0.47, + 0.688 + ], + "angle": 0, + "content": "We affirmatively address the question presented above. Succinctly, the generation process can be delineated into two key phases: latent representation sampling and decoding. Our goal is to align the distribution of the latent representation in watermarked images with that of the latent representation in normally generated images. By keeping the model unaltered, the distribution of watermarked images is naturally consistent with that of normally generated images, enabling the seamless embedding of watermarks without compromising model performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.69, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Building upon this insight, we propose a watermarking method named Gaussian Shading, designed to ensure no deterioration in model performance. The embedding process encompasses three primary elements: watermark diffuse, randomization, and distribution-preserving sampling. Watermark diffusion spreads the watermark information throughout the latent representation. During the generation process, the watermark information will be diffused to the whole semantics of the image, thus achieving excellent robustness. Watermark randomization and distribution-preserving sampling guarantee the congruity of the latent representation distribution with that of watermark-free latent representations, thereby achieving performance-lossless. In the extraction phase, the latent representations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "are acquired through Denoising Diffusion Implicit Model (DDIM) inversion [32], allowing for the retrieval of watermark information. Harnessing the extensive scope of the SD latent space, we can achieve a high-capacity watermark of 256 bits, surpassing prior methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.168, + 0.892, + 0.304 + ], + "angle": 0, + "content": "To the best of our knowledge, ours is the first technique that tackles this challenging problem of performanceless watermarking for diffusion models, and we provide theoretical proof. Moreover, this technique leaves the architecture and parameters of SD unaltered, necessitating no supplementary training. It can seamlessly integrate as a plug-and-play module within the generation process. Model providers can easily replace watermarked models with non-watermarked ones without affecting usability experiences." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.304, + 0.892, + 0.456 + ], + "angle": 0, + "content": "We conducted thorough experiments on SD. Under strong noise perturbation, the average true positive rate and bit accuracy can exceed 0.99 and 0.97, respectively, validating the superiority of Gaussian Shading in both detection and traceability tasks compared to prior methods. Additionally, experiments on visual quality and image-text similarity serve as indicators of performance preservation in our approach. Lastly, we deliberated on various watermark erasure attacks, affirming the steadfast performance of our watermark in the face of such adversities." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.47, + 0.642, + 0.485 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.495, + 0.668, + 0.51 + ], + "angle": 0, + "content": "2.1. Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.519, + 0.892, + 0.792 + ], + "angle": 0, + "content": "Inspired by non-equilibrium thermodynamics, Ho et al. [16] introduced the Denoising Diffusion Probabilistic Model (DDPM). DDPM consists of two Markov chains used for adding and removing noise, and subsequent works [8, 11, 15, 25, 28, 30, 32] have adopted this bidirectional chain framework. To reduce computational complexity and improve efficiency, the Latent Diffusion Model (LDM) [30] was designed, in which the diffusion process occurs in a latent space \\(\\mathcal{Z}\\). To map an image \\(x \\in \\mathbb{R}^{H \\times W \\times 3}\\) to the latent space, the LDM employs an encoder \\(\\mathcal{E}\\), such that \\(z_0 = \\mathcal{E}(x) \\in \\mathbb{R}^{h \\times w \\times c}\\). Similarly, to reconstruct an image from the latent space, a decoder \\(\\mathcal{D}\\) is used, such that \\(x = \\mathcal{D}(z_0)\\). A pretrained LDM can generate images without the encoder \\(\\mathcal{E}\\). Specifically, a latent representation \\(z_T\\) is first sampled from a standard Gaussian distribution \\(\\mathcal{N}(0, I)\\). Subsequently, through iterative denoising using methods like DDIM [32], \\(z_0\\) is obtained, and an image can be generated using the decoder: \\(x = \\mathcal{D}(z_0)\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.801, + 0.704, + 0.818 + ], + "angle": 0, + "content": "2.2. Image Watermarking" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Digital watermarking [38] is an effective means to address copyright protection and content authentication by embedding copyright or traceable identification information within carrier data. Typically, the functionality of a watermark depends on its capacity. For example, a single-bit" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12163" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.178, + 0.087, + 0.791, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.322, + 0.205, + 0.645, + 0.219 + ], + "angle": 0, + "content": "Figure 2. Application scenarios for Gaussian Shading." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.224, + 0.47, + 0.284 + ], + "angle": 0, + "content": "watermark can determine whether an image was generated by a particular diffusion model, i.e., copyright protection; a multi-bit watermark can further determine which user of the diffusion model generated the image, i.e., traceability." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.285, + 0.47, + 0.421 + ], + "angle": 0, + "content": "Image watermarking is a method that employs images as carriers for watermarking. Initially, watermark embedding methods primarily focused on the spatial domain [38], but later, to enhance robustness, transform domain watermarking techniques [1, 12, 13, 20, 22, 35, 37] were developed. In recent years, with the advancement of deep learning, researchers have turned their attention to neural networks [21, 39], harnessing their powerful learning capabilities to develop watermarking techniques [18, 19, 26, 36, 43, 50, 51]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.43, + 0.442, + 0.446 + ], + "angle": 0, + "content": "2.3. Image Watermarking for Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.453, + 0.471, + 0.832 + ], + "angle": 0, + "content": "Existing Image watermarking methods for the diffusion model [6, 7, 10, 24, 41, 42, 45, 49] can be divided into three categories, as shown in Fig. 1. The image watermarking methods described in the previous section can be applied directly to the images generated by the diffusion model, which is called post-processing-based watermarks [6, 45]. These methods directly modify the image, thus degrading image quality. Recent research endeavors have amalgamated the watermark embedding process with the image generation process to mitigate this issue. Stable Signature [10] fine-tunes the LDM decoder using a pre-trained watermark extractor, facilitating watermark extraction from images produced by the fine-tuned model. Zhao et al. [49] and Liu et al. [24] suggest fine-tuning the diffusion model to implant a backdoor as a watermark, enabling watermark extraction by triggering. These fine-tuning-based approaches enhance the quality of watermarked images but introduce supplementary computational overhead and modify model parameters. Furthermore, Wen et al. [41] introduced the Tree-Ring Watermark, which conveys copyright information by adapting the frequency domain of latent representations to match specific patterns. This method achieves an imperceptible watermark. However, it directly disrupts the Gaussian distribution of noise, limiting the randomness of sampling and resulting in affecting model performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.845, + 0.176, + 0.86 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "In this section, we provide an overview of the application scenarios and functionalities in Fig. 2. We then proceed" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.224, + 0.892, + 0.269 + ], + "angle": 0, + "content": "to detail the embedding and extraction processes shown in Fig. 3. Finally, we present a mathematical proof of the performance-lossless characteristic of the watermark." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.279, + 0.704, + 0.295 + ], + "angle": 0, + "content": "3.1. Application Scenarios" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.303, + 0.892, + 0.333 + ], + "angle": 0, + "content": "Scenarios. See Fig. 2, the scenario involves the operator Alice, the thief Carol, and two types of users Bob and Trudy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.334, + 0.892, + 0.499 + ], + "angle": 0, + "content": "Alice is responsible for training the model, deploying it on the platform, and providing the corresponding API for users, but she does not open-source the code or model weights. Carol does not use Alice's services but steals images generated by her model, claiming ownership of the copyrights. Bob and Trudy, as community users, can utilize the API to generate and disseminate images. While Bob faithfully adheres to the community guidelines, Trudy aims to generate deep fake, and infringing content. To evade detection and traceability, Trudy can employ various data augmentation to modify illicit images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.5, + 0.892, + 0.605 + ], + "angle": 0, + "content": "Detection. This scenario satisfies the detection (copyright protection) requirement. Alice embeds a single-bit watermark into each generated image. The successful extraction of the watermark from an image serves as evidence of Alice's rightful ownership of the copyright, while also indicating that the image is artificially generated (as opposed to natural images)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.606, + 0.892, + 0.696 + ], + "angle": 0, + "content": "Traceability. This scenario fulfills the traceability requirement. Alice allocates a watermark to each user. By extracting the watermark from the illicit content, it enables tracing Trudy, through comparison with the watermark database. Traceability is a higher pursuit than detection and can also achieve copyright protection for different users." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.696, + 0.892, + 0.726 + ], + "angle": 0, + "content": "Details of the statistical tests in both scenarios are shown in Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.736, + 0.72, + 0.752 + ], + "angle": 0, + "content": "3.2. Watermark Embedding" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.76, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Watermark diffusion. The dimensions of the latent representations are given by \\( c \\times h \\times w \\), where each dimension can represent \\( l \\) bits of the watermark. Therefore, the watermark capacity becomes \\( l \\times c \\times h \\times w \\) bits. To enhance the robustness of the watermark, we represent the watermark using \\( \\frac{1}{f_{hw}} \\) of the height and width, and \\( \\frac{1}{f_c} \\) of the channel, and replicate the watermark \\( f_c \\cdot f_{hw}^2 \\) times. Thus, the watermark \\( s \\) with dimensions \\( l \\times \\frac{c}{f_c} \\times \\frac{h}{f_{hw}} \\times \\frac{w}{f_{hw}} \\) is expanded into a diffused watermark \\( s^d \\) with dimensions \\( l \\times c \\times h \\times w \\)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12164" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.157, + 0.088, + 0.808, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.311, + 0.895, + 0.355 + ], + "angle": 0, + "content": "Figure 3. The framework of Gaussian Shading. We utilize a \\(k\\)-bit binary sequence \\(s\\) to represent the watermark. After diffusion and encryption, the watermark can be utilized to drive distribution-preserving sampling, followed by denoising to generate watermarked images \\(X^s\\). For extraction, it is sufficient to introduce DDIM inversion and the inverse process of all the operations mentioned above." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.357, + 0.429, + 0.378 + ], + "angle": 0, + "content": "The actual watermark capacity is \\( k = \\frac{l \\times c \\times h \\times w}{f_c \\cdot f_{hw}^2} \\) bits." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.377, + 0.47, + 0.526 + ], + "angle": 0, + "content": "Watermark randomization. If we know the distribution of the diffused watermark \\( s^d \\), we can directly utilize distribution-preserving sampling to obtain the corresponding latent representations \\( z_T^s \\). However, in practical scenarios, its distribution is always unknown. Hence, we introduce a stream key \\( K \\) to transform \\( s^d \\) into a distribution-unknown randomized watermark \\( m \\) through encryption. Considering the use of computationally secure stream cipher, such as ChaCha20 [3], \\( m \\) follows a uniform distribution, i.e., \\( m \\) is a random binary bit stream." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.527, + 0.473, + 0.695 + ], + "angle": 0, + "content": "Distribution-preserving sampling driven by randomized watermark. When each dimension represents \\(l\\)-bit randomized watermark \\(m\\), this \\(l\\) bits can be regarded as an integer \\(y \\in [0, 2^l - 1]\\). Since \\(m\\) is a ciphertext, \\(y\\) follows a discrete uniform distribution, i.e., \\(p(y) = \\frac{1}{2^l}\\) for \\(y = 0, 1, 2, \\ldots, 2^l - 1\\). Let \\(f(x)\\) denote the probability density function of the Gaussian distribution \\(\\mathcal{N}(0, I)\\), and \\(ppf\\) denotes the quantile function. We divide \\(f(x)\\) into \\(2^l\\) equal cumulative probability portions. When \\(y = i\\), the watermarked latent representation \\(z_T^s\\) falls into the \\(i\\)-th interval, which means \\(z_T^s\\) should follow the conditional distribution:" + }, + { + "type": "equation", + "bbox": [ + 0.076, + 0.703, + 0.478, + 0.749 + ], + "angle": 0, + "content": "\\[\np \\left(z _ {T} ^ {s} \\mid y = i\\right) = \\left\\{ \\begin{array}{c c} 2 ^ {l} \\cdot f \\left(z _ {T} ^ {s}\\right) & p p f \\left(\\frac {i}{2 ^ {l}}\\right) < z _ {T} ^ {s} \\leq p p f \\left(\\frac {i + 1}{2 ^ {l}}\\right) \\\\ 0 & o t h e r w i s e \\end{array} . \\right. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.75, + 0.384, + 0.767 + ], + "angle": 0, + "content": "The probability distribution of \\( z_T^s \\) is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.123, + 0.775, + 0.47, + 0.817 + ], + "angle": 0, + "content": "\\[\np \\left(z _ {T} ^ {s}\\right) = \\sum_ {i = 0} ^ {2 ^ {l} - 1} p \\left(z _ {T} ^ {s} \\mid y = i\\right) p (y = i) = f \\left(z _ {T} ^ {s}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.871 + ], + "angle": 0, + "content": "Eq. (2) indicates that \\( z_T^s \\) follows the same distribution as the randomly sampled latent representation \\( z_T \\sim \\mathcal{N}(0, I) \\). Next, we elaborate on how this sampling is implemented." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Let the cumulative distribution function of \\( f(x) \\) be denoted as \\( cdf \\). We can obtain the cumulative distribution" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.358, + 0.702, + 0.374 + ], + "angle": 0, + "content": "function of Eq. (1) as follows," + }, + { + "type": "equation", + "bbox": [ + 0.505, + 0.376, + 0.892, + 0.457 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} F (z _ {T} ^ {s} | y = i) \\\\ = \\left\\{ \\begin{array}{c c} 0 & z _ {T} ^ {s} < p p f \\left(\\frac {i}{2 ^ {l}}\\right) \\\\ 2 ^ {l} \\cdot c d f \\left(z _ {T} ^ {s}\\right) - i & p p f \\left(\\frac {i}{2 ^ {l}}\\right) \\leq z _ {T} ^ {s} \\leq p p f \\left(\\frac {i + 1}{2 ^ {l}}\\right) \\\\ 1 & z _ {T} ^ {s} > p p f \\left(\\frac {i + 1}{2 ^ {l}}\\right) \\end{array} . \\right. \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.458, + 0.893, + 0.638 + ], + "angle": 0, + "content": "Given \\(y = i\\), we aim to perform random sampling of \\(z_{T}^{s}\\) within the interval \\([ppf(\\frac{i}{2^t}), ppf(\\frac{i + 1}{2^t})]\\). The commonly used method is rejection sampling [4, 17, 47], which can be time-consuming as it requires repeated sampling until \\(z_{T}^{s}\\) falls into the correct interval. Instead, we can utilize the cumulative probability density. When randomly sampling \\(F(z_{T}^{s}|y = i)\\), the corresponding \\(z_{T}^{s}\\) is naturally obtained through random sampling. Since \\(F(z_{T}^{s}|y = i)\\) takes values in \\([0,1]\\), sampling from it is equivalent to sampling from a standard uniform distribution, denoted as \\(u = F(z_{T}^{s}|y = i) \\sim \\mathcal{U}(0,1)\\). Shift the terms of Eq. (3), and take into account that \\(cdf\\) and \\(ppf\\) are inverse functions, we have" + }, + { + "type": "equation", + "bbox": [ + 0.633, + 0.641, + 0.892, + 0.669 + ], + "angle": 0, + "content": "\\[\nz _ {T} ^ {s} = p p f \\left(\\frac {u + i}{2 ^ {l}}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.67, + 0.892, + 0.712 + ], + "angle": 0, + "content": "Eq. (4) represents the process of sampling the watermarked latent representation \\( z_{T}^{s} \\) driven by the randomized watermark \\( m \\). To extract the watermark, its inverse map is" + }, + { + "type": "equation", + "bbox": [ + 0.631, + 0.714, + 0.892, + 0.733 + ], + "angle": 0, + "content": "\\[\ni = \\left\\lfloor 2 ^ {l} \\cdot c d f \\left(z _ {T} ^ {s}\\right) \\right\\rfloor . \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Image generation. After the sampling process, the watermark is embedded in the latent representation \\( z_T^s \\), and the subsequent generation process is no different from the regular generation process of SD. Here, we employ the DPM-Solver [25] for iterative denoising of \\( z_T^s \\). In addition to DPMSolver [25], other continuous-time samplers based on ordinary differential equation (ODE) solvers [32], such as DDIM [32], DEIS [46], PNDM [23], and UniPC [48], can be used too. After obtaining denoised \\( z_0^s \\), the watermarked image \\( X^s \\) is generated using the decoder \\( \\mathcal{D} \\): \\( X^s = \\mathcal{D}(z_0^s) \\)." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12165" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.29, + 0.105 + ], + "angle": 0, + "content": "3.3. Watermark Extraction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.113, + 0.468, + 0.205 + ], + "angle": 0, + "content": "DDIM Inversion. Using the SD encoder \\(\\mathcal{E}\\), we first restore \\(X^{\\prime s}\\) to the latent space \\(z_0^{\\prime s} = \\mathcal{E}(X^{\\prime s})\\). Then, we introduce the DDIM inversion [32] to estimate the additive noise. It can be considered that \\(z_T^{\\prime s} \\approx z_T^s\\). We also observe that although DDIM inversion is derived from DDIM, it can apply to other continuous-time samplers based on ODE solvers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.206, + 0.468, + 0.34 + ], + "angle": 0, + "content": "Watermark reduction from latent representations. After obtaining \\( z_T^{\\prime s} \\), according to the inverse transformation defined in Eq. (5), the tensor can be converted into a bit stream \\( m' \\). Subsequently, \\( m' \\) is decrypted using \\( K \\) to obtain \\( s'^d \\). Inverse diffusion of the watermark results in \\( f_c \\cdot f_{hw}^2 \\) copies of the watermark. Similar to voting, if the bit is set to 1 in more than half of the copies, the corresponding watermark bit is set to 1; otherwise, it is set to 0. This process restores the true binary watermark sequence \\( s' \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.348, + 0.345, + 0.362 + ], + "angle": 0, + "content": "3.4. Proof of Lossless Performance" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.371, + 0.468, + 0.644 + ], + "angle": 0, + "content": "In prior works, the incorporation of watermark embedding modules inevitably results in a decline in model performance, as typically evaluated using metrics such as Peak Signal-to-Noise Ratio (PSNR) and Fréchet Inception Distance (FID) [14], which are more suitable for assessing post-processing methods. To assess methods that integrate the watermark embedding and generation processes, we propose a definition for the impact of watermark embedding on model performance, drawing on the complexity-theoretic definition of steganographic security [17]. This definition is based on a probabilistic game between a watermarked image \\( X^{s} \\) and a normally generated image \\( X \\). The tester \\( \\mathcal{A} \\) can use any watermark to drive the sampling process and generate \\( X^{s} \\), similar to the chosen hidden text attacks [17], which we refer to as chosen watermark tests. The watermarking method is performance-lossless under chosen watermark tests, if for any polynomial-time tester \\( \\mathcal{A} \\) and key \\( K \\gets \\text{KeyGen}_{\\mathcal{G}(1^{\\rho})} \\), it holds that" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.646, + 0.468, + 0.663 + ], + "angle": 0, + "content": "\\[\n\\left| \\Pr \\left[ \\mathcal {A} \\left(X ^ {s}\\right) = 1 \\right] - \\Pr \\left[ \\mathcal {A} (X) = 1 \\right] \\right| < \\operatorname {n e g l} (\\rho). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.669, + 0.468, + 0.695 + ], + "angle": 0, + "content": "Here, \\(\\rho\\) represents the length of the security parameter, such as the key \\(K\\), and \\(\\mathsf{negl}(\\rho)\\) is a negligible term relative to \\(\\rho\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.696, + 0.468, + 0.741 + ], + "angle": 0, + "content": "We prove the statement using a proof by contradiction. First, assume that the watermarked image \\( X^s \\) and the normally generated image \\( X \\) are distinguishable, meaning" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.743, + 0.468, + 0.76 + ], + "angle": 0, + "content": "\\[\n\\left. \\right.\\left| \\Pr \\left[ \\mathcal {A} \\left(X ^ {s}\\right) = 1 \\right] - \\Pr \\left[ \\mathcal {A} (X) = 1 \\right]\\right| = \\delta , \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.767, + 0.468, + 0.809 + ], + "angle": 0, + "content": "where \\(\\delta\\) is non-negligible with respect to the key \\(K\\). Let the iterative denoising process be denoted as \\(Q(\\cdot)\\), and substitute the LDM encoder \\(\\mathcal{E}\\) into Eq. (7), we have" + }, + { + "type": "equation", + "bbox": [ + 0.096, + 0.813, + 0.434, + 0.85 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\left| \\Pr \\left[ \\mathcal {A} \\left(\\mathcal {E} \\left(Q \\left(z _ {T} ^ {s}\\right)\\right)\\right) = 1 \\mid m = E (K, s ^ {d}) \\right] \\right. \\\\ - \\Pr \\left[ \\mathcal {A} \\left(\\mathcal {E} \\left(Q \\left(z _ {T}\\right)\\right)\\right) = 1 \\mid z _ {T} \\leftarrow \\mathcal {N} (0, I) \\right] | = \\delta , \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "where the randomized watermark \\(m\\) is obtained by encrypting the diffused watermark \\(s^d\\) using the encryption algorithm \\(E\\) with key \\(K\\). Note that Eq. (2) contains the fact that" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "distribution-preserving sampling driven by randomized watermark and random sampling are equivalent. Therefore, we denote sequence-driven sampling as \\( S(\\cdot) \\). \\( z_{T}^{s} \\) can naturally be obtained by sampling driven by \\( m \\), i.e., \\( z_{T}^{s} = S(m) \\). On the other hand, \\( z_{T} \\) can be considered as obtained by sampling driven by a truly random sequence \\( r \\) of the same length as \\( m \\), i.e., \\( z_{T} = S(r) \\). Eq. (8) can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.202, + 0.892, + 0.239 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\left| \\Pr [ \\mathcal {A} (\\mathcal {E} (Q (S (m)))) = 1 | m = E (K, s ^ {d}) ] \\right. \\tag {9} \\\\ - \\Pr \\left[ \\mathcal {A} \\left(\\mathcal {E} (Q (S (r))) = 1 \\right\\rvert \\right] = \\delta . \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.248, + 0.892, + 0.29 + ], + "angle": 0, + "content": "Sampling \\(S(\\cdot)\\), denoising \\(Q(\\cdot)\\), and encoder \\(\\mathcal{E}\\) can be considered as subroutines that the tester \\(\\mathcal{A}_{\\mathcal{E}, Q, S}\\) can use. Thus, Eq. (9) can be simplified," + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.294, + 0.892, + 0.331 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\left| \\Pr \\left[ \\mathcal {A} _ {\\varepsilon , Q, S} (m) = 1 \\mid m = E (K, s ^ {d}) \\right] \\right. \\tag {10} \\\\ - \\Pr \\left[ \\mathcal {A} _ {\\mathcal {E}, Q, S} (r) = 1 \\right] | = \\delta . \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.341, + 0.892, + 0.568 + ], + "angle": 0, + "content": "Note that \\( S(\\cdot), Q(\\cdot) \\), and \\( \\mathcal{E} \\) are all polynomial-time programs, so the time taken by the tester \\( \\mathcal{A}_{\\mathcal{E},Q,S} \\) to make the distinction is also polynomial. Eq. (10) essentially states that it is possible to distinguish between \\( m \\) and \\( r \\) in polynomial time. However, we have used the computationally secure stream cipher ChaCha20 [3] in watermark randomization, which means that \\( m \\) as a pseudorandom sequence cannot be distinguished from a truly random sequence in polynomial time. Eq. (10) contradicts the computational security property of ChaCha20 [3]. Therefore, Eq. (10) is not valid, leading us back to our initial assumption that Eq. (7) is also not valid. This implies that the watermarked image \\( X^s \\) and the normally generated image \\( X \\) are indistinguishable in polynomial time. Hence, Gaussian Shading is performance-lossless under chosen watermark tests." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.581, + 0.632, + 0.598 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.606, + 0.892, + 0.667 + ], + "angle": 0, + "content": "This section focuses on experimental analysis, including details of the experimental setup, performance evaluation of Gaussian Shading, comparison with baseline methods, ablation experiments, and potential attacks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.675, + 0.716, + 0.69 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.698, + 0.892, + 0.879 + ], + "angle": 0, + "content": "SD models. In this paper, we focus on text-to-image LDM, hence we select SD [30] provided by huggingface. We evaluate Gaussian Shading as well as baseline methods, using three versions of SD: V1.4, V2.0, and V2.1. The size of the generated images is \\(512 \\times 512\\), and the latent space dimension is \\(4 \\times 64 \\times 64\\). During inference, we employ the prompt from Stable-Diffusion-Prompt3, with a guidance scale of 7.5. We sample 50 steps using DPMSolver [25]. Considering that users tend to propagate the generated images without retaining the corresponding prompts, we use an empty prompt for inversion, with a scale of 1. We perform 50 steps of inversion using DDIM inversion [32]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.887, + 0.663, + 0.9 + ], + "angle": 0, + "content": "3Stable-Diffusion-Prompts" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12166" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.088, + 0.159, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.164, + 0.089, + 0.235, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.188, + 0.148, + 0.206, + 0.158 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.236, + 0.089, + 0.311, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.313, + 0.089, + 0.386, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.39, + 0.089, + 0.463, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.339, + 0.148, + 0.357, + 0.158 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.16, + 0.157, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.111, + 0.219, + 0.127, + 0.23 + ], + "angle": 0, + "content": "(f)" + }, + { + "type": "image", + "bbox": [ + 0.158, + 0.16, + 0.232, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.185, + 0.219, + 0.203, + 0.23 + ], + "angle": 0, + "content": "(g)" + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.16, + 0.308, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.261, + 0.219, + 0.278, + 0.23 + ], + "angle": 0, + "content": "(h)" + }, + { + "type": "image", + "bbox": [ + 0.311, + 0.16, + 0.384, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.219, + 0.352, + 0.23 + ], + "angle": 0, + "content": "(i)" + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.16, + 0.459, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.413, + 0.219, + 0.428, + 0.23 + ], + "angle": 0, + "content": "(j)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.232, + 0.47, + 0.329 + ], + "angle": 0, + "content": "Figure 4. Watermarked image is attacked by different noise. (a) Watermarked image. (b) JPEG, \\( QF = 25 \\). (c) \\( 60\\% \\) area Random Crop (RandCr). (d) \\( 80\\% \\) area Random Drop (RandDr). (e) Gaussian Blur, \\( r = 4 \\) (GauBlur). (f) Median Filter, \\( k = 7 \\) (MedFilter). (g) Gaussian Noise, \\( \\mu = 0 \\), \\( \\sigma = 0.05 \\) (GauNoise). (h) Salt and Pepper Noise, \\( p = 0.05 \\) (S&PNoise). (i) \\( 25\\% \\) Resize and restore (Resize). (j) Brightness, factor \\( = 6 \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.332, + 0.469, + 0.438 + ], + "angle": 0, + "content": "Watermarking methods. In the main experiments, the settings for Gaussian Shading are \\( f_{c} = 1 \\), \\( f_{hw} = 8 \\), \\( l = 1 \\), resulting in an actual capacity of 256 bits. We select five baseline methods: three officially used by SD, namely Dwt-Dct [6], DwtDctSvd [6], and RivaGAN [45], a multi-bit watermarking called Stable Signature [10], and a train-free invisible watermarking called Tree-Ring [41]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.44, + 0.469, + 0.484 + ], + "angle": 0, + "content": "Robustness evaluation To evaluate the robustness, we select nine representative types of noise shown in Fig. 4. We conduct experiments following the noise strength in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.485, + 0.469, + 0.606 + ], + "angle": 0, + "content": "Evaluation metrics. In the detection scenario, we calculate the true positive rate (TPR) corresponding to a fixed false positive rate (FPR). In the traceability scenario, we calculate the bit accuracy. To measure the bias in model performance, we compute the FID [14] and CLIP-Score [29] for 10 batches of watermarked images and perform a \\( t \\)-test on the mean FID and CLIP-Score compared to that of watermark-free images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.607, + 0.469, + 0.637 + ], + "angle": 0, + "content": "All experiments are conducted using the PyTorch 1.13.0 framework, running on a single RTX 3090 GPU." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.649, + 0.376, + 0.666 + ], + "angle": 0, + "content": "4.2. Performance of Gaussian Shading" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.469, + 0.824 + ], + "angle": 0, + "content": "Detection. In the detection scenario, we consider Gaussian Shading as a single-bit watermark, with a fixed watermark \\( s \\). We approximate the FPR to be controlled at \\( 10^{0}, 10^{-1}, \\ldots, 10^{-13} \\), calculate the corresponding threshold \\( \\tau \\), and test the TPR on 1,000 watermarked images To mitigate the effects of randomness, we perform 5 trials with different \\( s \\) and compute the average TPR. See Fig. 5a, when the FPR is controlled at \\( 10^{-13} \\), the TPR remains at least 0.99 for eight out of the nine cases. Although the TPR for Brightness is only 0.953, it is still a promising result." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Traceability. In this scenario, Gaussian Shading serves as a multi-bit watermark. Assuming Alice provides services to \\( N \\) users, Alice needs to allocate one watermark for each user. In our experiments, we assume that \\( N' = 1,000 \\) users generate images, with each user generating 10 images, re" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.089, + 0.673, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.183, + 0.65, + 0.194 + ], + "angle": 0, + "content": "(a) Detection results." + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.089, + 0.871, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.73, + 0.183, + 0.854, + 0.195 + ], + "angle": 0, + "content": "(b) Traceability results." + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.197, + 0.829, + 0.21 + ], + "angle": 0, + "content": "Figure 5. Performance of Gaussian Shading." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.213, + 0.838, + 0.228 + ], + "angle": 0, + "content": "sulting in a dataset of 10,000 watermarked images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.23, + 0.892, + 0.426 + ], + "angle": 0, + "content": "During testing, we calculate the threshold \\(\\tau\\) to control the FPR at \\(10^{-6}\\). Note that when computing traceability accuracy, we need to consider two types of errors: false positives, where watermarked images are not detected, and traceability errors, where watermarked images are detected but attributed to the wrong user. Therefore, we first determine whether the image contains a watermark. If it does, we calculate the number of matching bits \\(Acc\\) with all \\(N\\) users on the platform. The user with the highest \\(Acc\\) is considered the one who generated the image. Finally, we verify whether the correct user has been traced. When \\(N > N'\\), it can be assumed that some users have been assigned a watermark but have not generated any images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.426, + 0.892, + 0.501 + ], + "angle": 0, + "content": "See Fig. 5b, when \\( N = 10^{6} \\), Gaussian Shading exhibits almost perfect traceability in seven cases. Although the traceability accuracy for Brightness is only \\( 95.47\\% \\), if a user generates two images, the probability of successfully tracing him is still no less than \\( 99\\% \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.513, + 0.726, + 0.529 + ], + "angle": 0, + "content": "4.3. Comparison to Baselines" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.892, + 0.598 + ], + "angle": 0, + "content": "In this section, we compare the performance of Gaussian Shading with baselines on SD V1.4, V2.0, and V2.1. We use our implementations for each method, see details in Supplementary Material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.719 + ], + "angle": 0, + "content": "We conduct tests on 1,000 generated images for each method respectively. See Tab. 1. Gaussian Shading exhibits strong robustness and significantly outperforms baselines in both scenarios. In terms of bit accuracy, it surpasses the best-performing baseline by approximately \\(7\\%\\). This can be attributed to the extensive diffusion of the watermark throughout the entire latent space, establishing a profound binding between the watermark and the image semantics." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To measure the performance bias introduced by the watermark embedding, we apply a \\(t\\)-test to evaluate. The hypotheses are \\(H_0: \\mu_s = \\mu_0, H_1: \\mu_s \\neq \\mu_0\\), where \\(\\mu_s\\) and \\(\\mu_0\\) represent the average FID [14] or CLIP-Score [29] of multiple sets of watermarked and watermark-free images, respectively. A lower \\(t\\)-value indicates a higher probability that \\(H_0\\) holds. If the \\(t\\)-value is larger than a threshold, \\(H_0\\) is rejected, and model performance is considered to have been affected. See Tab. 1, Gaussian Shading achieves the smallest \\(t\\)-value, which indirectly reflects its performance-lossless characteristic. For a detailed analysis of the \\(t\\)-test, please refer to the Supplementary Material." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12167" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.09, + 0.894, + 0.214 + ], + "angle": 0, + "content": "
MethodsMetrics
TPR (Clean)TPR (Adversarial)Bit Acc. (Clean)Bit Acc. (Adversarial)FID (t-value ↓)CLIP-Score (t-value ↓)
Stable Diffusion----25.23±.180.3629±.0006
DwtDct [6]0.825/0.881/0.8660.172/0.178/0.1730.8030/0.8059/0.80230.5696/0.5671/0.562224.97±.19 (3.026)0.3617±.0007 (3.045)
DwtDctSvd [6]1.000/1.000/1.0000.597/0.594/0.5990.9997/0.9987/0.99870.6920/0.6868/0.690524.45±.22 (8.253)0.3609±.0009 (4.452)
RivaGAN [45]0.920/0.945/0.9630.697/0.697/0.7060.9762/0.9877/0.99210.8986/0.9124/0.901924.24±.16 (12.29)0.3611±.0009 (4.259)
Tree-Ring [41]1.000/1.000/1.0000.894/0.898/0.906--25.43±.13 (2.581)0.3632±.0006 (0.8278)
Stable Signature [10]1.000/1.000/1.0000.502/0.505/0.4960.9987/0.9978/0.99790.7520/0.7472/0.750025.45±.18 (2.477)0.3622±.0027 (0.7066)
Ours1.000/1.000/1.0000.997/0.998/0.9960.9999/0.9999/0.99990.9753/0.9749/0.972425.20±.22 (0.3567)0.3631±.0005 (0.6870)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.216, + 0.893, + 0.258 + ], + "angle": 0, + "content": "Table 1. Comparison results. We control the FPR at \\(10^{-6}\\), and evaluate the TPR and bit accuracy for SD V1.4/V2.0/V2.1. To assess the bias in model performance, we conduct a \\(t\\)-test on SD V2.1. Adversarial here refers to the average performance of a series of noises. Additional results can be found in Supplementary Material." + }, + { + "type": "image", + "bbox": [ + 0.103, + 0.26, + 0.253, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.352, + 0.232, + 0.363 + ], + "angle": 0, + "content": "(a) Guidance scales." + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.26, + 0.406, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.307, + 0.352, + 0.358, + 0.363 + ], + "angle": 0, + "content": "(b) JPEG" + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.26, + 0.56, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.437, + 0.352, + 0.534, + 0.364 + ], + "angle": 0, + "content": "(c) Random Crop." + }, + { + "type": "image", + "bbox": [ + 0.565, + 0.26, + 0.714, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.591, + 0.352, + 0.688, + 0.364 + ], + "angle": 0, + "content": "(d) Random Drop." + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.26, + 0.868, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.743, + 0.352, + 0.839, + 0.363 + ], + "angle": 0, + "content": "(e) Gaussian Blur." + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.364, + 0.249, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.455, + 0.22, + 0.467 + ], + "angle": 0, + "content": "(f) Median Filter." + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.364, + 0.402, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.275, + 0.455, + 0.381, + 0.467 + ], + "angle": 0, + "content": "(g) Gaussian Noise." + }, + { + "type": "image", + "bbox": [ + 0.407, + 0.364, + 0.555, + 0.451 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.411, + 0.455, + 0.551, + 0.467 + ], + "angle": 0, + "content": "(h) Salt and Pepper Noise." + }, + { + "type": "image", + "bbox": [ + 0.561, + 0.364, + 0.709, + 0.451 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.607, + 0.455, + 0.662, + 0.466 + ], + "angle": 0, + "content": "(i) Resize." + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.364, + 0.863, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.751, + 0.455, + 0.827, + 0.468 + ], + "angle": 0, + "content": "(j) Brightness." + }, + { + "type": "image_caption", + "bbox": [ + 0.404, + 0.47, + 0.565, + 0.482 + ], + "angle": 0, + "content": "Figure 6. Ablation studies." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.484, + 0.241, + 0.498 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.469, + 0.567 + ], + "angle": 0, + "content": "In this section, we conduct comprehensive ablation experiments on SD V2.1 to demonstrate hyperparameter selection. Unless specified, we generate 1,000 images and test the TPR and the bit accuracy with a theoretical FPR of \\(10^{-6}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.568, + 0.469, + 0.718 + ], + "angle": 0, + "content": "Watermark capacity. The watermark capacity is determined by three parameters: channel diffusion factor \\( f_{c} \\), height-width diffusion factor \\( f_{hw} \\), and embedding rate \\( l \\). See Tab. 2, to balance the capacity and robustness of Gaussian Shading, we chose \\( f_{c} = 1 \\) and \\( f_{hw} = 8 \\). After fixing \\( f_{c} \\) and \\( f_{hw} \\), we vary \\( l \\) to examine if it could enhance the capacity, and additional results can be found in Supplementary Material. Considering all factors, we determine that the optimal solution is \\( f_{c} = 1 \\), \\( f_{hw} = 8 \\), and \\( l = 1 \\), resulting in a watermark capacity of 256 bits." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.469, + 0.795 + ], + "angle": 0, + "content": "Sampling methods. To validate the generalization, we select five commonly used sampling methods, all continuous-time samplers based on ODE solvers [32]. See Tab. 3, all of them exhibit excellent performance with a bit accuracy of approximately \\(97\\%\\) against noises." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.469, + 0.885 + ], + "angle": 0, + "content": "Impact of the inversion step. In practice, the inference step is often unknown, which introduces a mismatch with the inversion step. See Tab. 4, such mismatch introduces minimal loss in accuracy. Considering the high efficiency of existing samplers, the inference step generally does not exceed 50. Therefore, we set the inversion step to 50." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.887, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Guidance scales. Given diverse user preferences for image" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.484, + 0.892, + 0.621 + ], + "angle": 0, + "content": "prompt alignment, larger guidance scales ensure faithful adherence to prompts, while smaller scales grant the model greater creative freedom. In SD, the guidance scale is typically selected from the range of [5, 15]. Hence, experiments cover the range of 2 to 18. For the inversion, an empty prompt is used for guidance, and the guidance scale is fixed at 1, assuming unknown information during extraction. In Fig. 6a, the bit accuracy of Gaussian Shading surpasses \\(99.9\\%\\), showing its reliability in real-world-like scenes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.621, + 0.892, + 0.712 + ], + "angle": 0, + "content": "Noise intensities. To further test the robustness, we conduct experiments using different intensities of noises. See Figs. 6b to 6j, for Random Crop and Gaussian Noise, performance declines significantly with higher intensities. However, for the other seven types of noise, even at high intensities, the bit accuracy remains approximately \\(80\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.725, + 0.798, + 0.74 + ], + "angle": 0, + "content": "4.5. Attacks against Gaussian Shading" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.825 + ], + "angle": 0, + "content": "We consider two malicious attacks: compression attack, where the attacker employs a neural network to compress watermarked images, and inversion attack, assuming the attacker is aware of the watermark embedding method, enabling them to modify the image's latent representations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Compression attack. We utilize popular auto-encoders [2, 5, 9, 30] to compare Stable Signature (SS) with Gaussian Shading across various compression rates. Additionally, we assess the compression quality through the PSNR between the compressed and watermarked images. See Figs. 7a" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12168" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.167, + 0.089, + 0.803, + 0.159 + ], + "angle": 0, + "content": "
Noisefc-fhw (k bits)
1-2 (4096)4-1 (4096)1-4 (1024)4-2 (1024)1-8 (256)4-4 (256)1-16 (64)4-8 (64)
None0.94130.93800.99850.99800.99990.99991.00001.0000
Adversarial0.73020.72380.87690.86140.97240.96710.99590.9953
" + }, + { + "type": "table_caption", + "bbox": [ + 0.1, + 0.16, + 0.867, + 0.174 + ], + "angle": 0, + "content": "Table 2. Bit accuracy with different factors \\( {f}_{c} \\) and \\( {f}_{hw} \\) ,where \\( l = 1 \\) . Additional results can be found in Supplementary Material." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.183, + 0.457, + 0.264 + ], + "angle": 0, + "content": "
NoiseSampling Methods
DDIM [32]UniPC [48]PNDM [23]DEIS [46]DPMSolver [25]
None0.99991.00001.00000.99990.9999
Adversarial0.97060.96280.97210.97150.9724
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.265, + 0.47, + 0.349 + ], + "angle": 0, + "content": "Table 3. Bit accuracy with different sampling methods. Additional results can be found in Supplementary Material. These methods differ only in accuracy and order. DDIM is a first-order estimate of the ODE. Accordingly, DDIM inversion ensures a lower bound on the accuracy of the inversion process. Therefore, it can naturally be applied to higher-order and higher-accuracy methods." + }, + { + "type": "table", + "bbox": [ + 0.136, + 0.352, + 0.411, + 0.445 + ], + "angle": 0, + "content": "
Inference StepInversion Step
102550100
100.99990.99990.99990.9999
250.99980.99991.00001.0000
500.99950.99970.99990.9999
1000.99940.99960.99990.9999
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.446, + 0.469, + 0.523 + ], + "angle": 0, + "content": "Table 4. Bit accuracy with different inference and inversion step. and 7b, Gaussian Shading significantly outperforms Stable Signature. This is because Gaussian Shading diffuses the watermark across the entire semantic space of images, while Stable Signature relies solely on the image texture." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.524, + 0.469, + 0.658 + ], + "angle": 0, + "content": "Inversion attack. Assuming the attacker is aware of the embedding method, a more effective approach to erasing is through inversion to obtain latent representations and subsequently modify them. We validate the robustness against such attacks. Importantly, our experiments assume the strongest attacker capability of using the same model as Alice for precise inversion. In real-world scenarios, where the watermark embedding is not publicly available, the attacker's capabilities would be weaker." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.469, + 0.795 + ], + "angle": 0, + "content": "Specifically, we perform inversion to obtain latent representations and randomly flip a certain rate of them. Using the flipped latent representations, we regenerate the images and extract the watermark. See Fig. 7c. the watermark can still be reliably extracted when the flipping rate (FR) is less than 0.4. At high FRs, significant changes in images are observed. Although the watermark cannot be accurately extracted, we consider the image transformed into a different one, resulting in the content not intended to be protected." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.469, + 0.902 + ], + "angle": 0, + "content": "From another perspective, the attacker can launch a forgery attack by performing inversion on an innocuous image from Bob and generating harmful content using a different prompt. See Fig. 7c, when the FR is 0, Alice can accurately trace Bob based on the forgeries, enabling the attacker to successfully frame Bob. Therefore, protecting the model from leakage is crucial for operators." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.183, + 0.63, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.511, + 0.261, + 0.623, + 0.273 + ], + "angle": 0, + "content": "(a) Detection results." + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.183, + 0.759, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.635, + 0.261, + 0.758, + 0.273 + ], + "angle": 0, + "content": "(b) Traceability results." + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.183, + 0.889, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.773, + 0.261, + 0.879, + 0.272 + ], + "angle": 0, + "content": "(c) Inversion attack." + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.276, + 0.892, + 0.317 + ], + "angle": 0, + "content": "Figure 7. Performance of Gaussian Shading under Malicious Attack, where (a) and (b) are under compression attack and (c) is under inversion attack." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.324, + 0.623, + 0.34 + ], + "angle": 0, + "content": "5. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.35, + 0.892, + 0.593 + ], + "angle": 0, + "content": "Despite extensive experimental validation of Gaussian Shading's superior performance, our work still has certain limitations. Firstly, the usage scenarios are restricted due to the reliance on DDIM inversion [32], which necessitates the utilization of continuous-time samplers based on ODE solvers [32] like DPMSolver [25]. Secondly, Gaussian Shading employs stream ciphers, necessitating proper key usage and management on the deployment platform. Additionally, we assume that the model is not publicly accessible, and only operators can verify the watermark, providing a certain level of protection against white-box attacks and ensuring security. However, if a legitimate third party requires watermark verification, cooperation from the operators becomes necessary. Lastly, Gaussian Shading is vulnerable to forgery attacks, emphasizing the importance for operators to safeguard the model parameters." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.605, + 0.771, + 0.621 + ], + "angle": 0, + "content": "6. Conclusion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.892, + 0.766 + ], + "angle": 0, + "content": "We propose Gaussian Shading, a provably performance-lossless watermarking applied to diffusion models. Compared to baseline methods, Gaussian Shading offers simplicity and effectiveness by making a simple modification in the sampling process of the initial latent representation. Extensive experiments validate the superior performance in both detection and traceability scenarios. To our knowledge, we are the first to propose and implement a performance-lossless approach in image watermarking." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.767, + 0.892, + 0.827 + ], + "angle": 0, + "content": "Regarding future work, we will introduce more efficient inversion methods [40, 44] and include a wider range of sampling methods. Additionally, careful consideration should be given to counteracting forgery attacks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgement. This work was supported in part by the Natural Science Foundation of China under Grant U2336206, 62102386, 62072421, 62372423, and 62121002." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12169" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.143 + ], + "angle": 0, + "content": "[1] Ali Al-Haj. Combined dwt-dct digital image watermarking. Journal of computer science, 3(9):740-746, 2007. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.145, + 0.469, + 0.2 + ], + "angle": 0, + "content": "[2] Johannes Ballé, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. arXiv preprint arXiv:1802.01436, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.202, + 0.469, + 0.23 + ], + "angle": 0, + "content": "[3] Daniel J Bernstein et al. Chacha, a variant of salsa20. In Workshop record of SASC, pages 3-5. Citeseer, 2008. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.232, + 0.469, + 0.3 + ], + "angle": 0, + "content": "[4] Kejiang Chen, Hang Zhou, Hanqing Zhao, Dongdong Chen, Weiming Zhang, and Nenghai Yu. Distribution-preserving steganography based on text-to-speech generative models. IEEE Transactions on Dependable and Secure Computing, 19(5):3343-3356, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.303, + 0.469, + 0.373 + ], + "angle": 0, + "content": "[5] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned image compression with discretized gaussian mixture likelihoods and attention modules. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7939-7948, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.374, + 0.469, + 0.416 + ], + "angle": 0, + "content": "[6] Ingemar Cox, Matthew Miller, Jeffrey Bloom, Jessica Fridrich, and Ton Kalker. Digital watermarking and steganography. Morgan Kaufmann, 2007. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.418, + 0.469, + 0.473 + ], + "angle": 0, + "content": "[7] Yingqian Cui, Jie Ren, Han Xu, Pengfei He, Hui Liu, Lichao Sun, and Jiliang Tang. Diffusionshield: A watermark for copyright protection against generative diffusion models. arXiv preprint arXiv:2306.04642, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.475, + 0.469, + 0.516 + ], + "angle": 0, + "content": "[8] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.518, + 0.469, + 0.573 + ], + "angle": 0, + "content": "[9] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.575, + 0.469, + 0.631 + ], + "angle": 0, + "content": "[10] Pierre Fernandez, Guillaume Couairon, Hervé Jégou, Matthijs Douze, and Teddy Furon. The stable signature: Rooting watermarks in latent diffusion models. arXiv preprint arXiv:2303.15435, 2023. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.469, + 0.702 + ], + "angle": 0, + "content": "[11] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10696-10706, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.469, + 0.758 + ], + "angle": 0, + "content": "[12] Huiping Guo and Nicolas D Georganas. Digital image watermarking for joint ownership. In Proceedings of the tenth ACM international conference on Multimedia, pages 362-371, 2002. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.469, + 0.829 + ], + "angle": 0, + "content": "[13] Mohamed Hamidi, Mohamed El Haziti, Hocine Cherifi, and Mohammed El Hassouni. Hybrid blind robust image watermarking technique based on dft-dct and arnold transform. Multimedia Tools and Applications, 77:27181-27214, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "[15] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.123, + 0.892, + 0.164 + ], + "angle": 0, + "content": "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.165, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[17] Nicholas J Hopper, John Langford, and Luis Von Ahn. Provably secure steganography. In Advances in Cryptology—CRYPTO 2002: 22nd Annual International Cryptology Conference Santa Barbara, California, USA, August 18–22, 2002 Proceedings 22, pages 77–92. Springer, 2002. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.236, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[18] Zhaoyang Jia, Han Fang, and Weiming Zhang. Mbrs: Enhancing robustness of dnn-based watermarking by minibatch of real and simulatedJPEG compression. In Proceedings of the 29th ACM international conference on multimedia, pages 41-49, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.361 + ], + "angle": 0, + "content": "[19] Varsha Kishore, Xiangyu Chen, Yan Wang, Boyi Li, and Kilian Q Weinberger. Fixed neural network steganography: Train the images, not the network. In International Conference on Learning Representations, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.892, + 0.418 + ], + "angle": 0, + "content": "[20] Deepa Kundur and Dimitrios Hatzinakos. A robust digital image watermarking method using wavelet-based fusion. In Proceedings of International Conference on Image Processing, pages 544-547. IEEE, 1997. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.42, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[21] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.892, + 0.532 + ], + "angle": 0, + "content": "[22] Sunil Lee, Chang D Yoo, and Ton Kalker. Reversible image watermarking based on integer-to-integer wavelet transform. IEEE Transactions on information forensics and security, 2 (3):321-330, 2007. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.533, + 0.892, + 0.575 + ], + "angle": 0, + "content": "[23] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. arXiv preprint arXiv:2202.09778, 2022. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.576, + 0.892, + 0.617 + ], + "angle": 0, + "content": "[24] Yugeng Liu, Zheng Li, Michael Backes, Yun Shen, and Yang Zhang. Watermarking diffusion model. arXiv preprint arXiv:2305.12502, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.619, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[25] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 2, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.892, + 0.745 + ], + "angle": 0, + "content": "[26] Xiyang Luo, Ruohan Zhan, Huiwen Chang, Feng Yang, and Peyman Milanfar. Distortion agnostic deep watermarking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13548-13557, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.746, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[27] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[28] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.903 + ], + "angle": 0, + "content": "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry," + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12170" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.15, + 0.472, + 0.219 + ], + "angle": 0, + "content": "[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.47, + 0.288 + ], + "angle": 0, + "content": "[31] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.47, + 0.332 + ], + "angle": 0, + "content": "[32] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2020. 2, 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.47, + 0.375 + ], + "angle": 0, + "content": "[33] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.47, + 0.445 + ], + "angle": 0, + "content": "[34] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.448, + 0.47, + 0.503 + ], + "angle": 0, + "content": "[35] Srdjan Stankovic, Irena Orovic, and Nikola Zaric. An application of multidimensional time-frequency analysis as a base for the unified watermarking approach. IEEE Transactions on Image Processing, 19(3):736-745, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.47, + 0.56 + ], + "angle": 0, + "content": "[36] Matthew Tancik, Ben Mildenhall, and Ren Ng. Stegastamp: Invisible hyperlinks in physical photographs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2117-2126, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.562, + 0.47, + 0.615 + ], + "angle": 0, + "content": "[37] Min-Jen Tsai, Kuang-Yao Yu, and Yi-Zhang Chen. Joint wavelet and spatial transformation for digital watermarking. IEEE Transactions on Consumer Electronics, 46(1):237, 2000. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.618, + 0.47, + 0.672 + ], + "angle": 0, + "content": "[38] Ron G Van Schyndel, Andrew Z Tirkel, and Charles F Osborne. A digital watermark. In Proceedings of 1st international conference on image processing, pages 86-90. IEEE, 1994. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.676, + 0.47, + 0.731 + ], + "angle": 0, + "content": "[39] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.47, + 0.787 + ], + "angle": 0, + "content": "[40] Bram Wallace, Akash Gokul, and Nikhil Naik. Edict: Exact diffusion inversion via coupled transformations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22532-22541, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[41] Yuxin Wen, John Kirchenbauer, Jonas Geiping, and Tom Goldstein. Tree-ring watermarks: Fingerprints for diffusion images that are invisible and robust. arXiv preprint arXiv:2305.20030, 2023. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[42] Cheng Xiong, Chuan Qin, Guorui Feng, and Xinpeng Zhang. Flexible and secure watermarking for latent diffusion model. In Proceedings of the 31st ACM International Conference on Multimedia, pages 1668-1676, 2023. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[43] Chaoning Zhang, Philipp Benz, Adil Karjauv, Geng Sun, and In So Kweon. Udh: Universal deep hiding for steganography, watermarking, and light field messaging. Advances in Neural Information Processing Systems, 33:10223-10234, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[44] Jiaxin Zhang, Kamalika Das, and Sricharan Kumar. On the robustness of diffusion inversion in image manipulation. In ICLR 2023 Workshop on Trustworthy and Reliable Large-Scale Machine Learning Models, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.26 + ], + "angle": 0, + "content": "[45] Kevin Alex Zhang, Lei Xu, Alfredo Cuesta-Infante, and Kalyan Veeramachaneni. Robust invisible video watermarking with attention. arXiv preprint arXiv:1909.01285, 2019. 2,3,6,7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[46] Qinsheng Zhang and Yongxin Chen. Fast sampling of diffusion models with exponential integrator. arXiv preprint arXiv:2204.13902, 2022. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[47] Weiming Zhang, Kejiang Chen, and Nenghai Yu. Provable secure steganography: Theory, application and prospects. Journal of Cybersecurity, 1:38-46, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[48] Wenliang Zhao, Lujia Bai, Yongming Rao, Jie Zhou, and Jiwen Lu. Unipc: A unified predictor-corrector framework for fast sampling of diffusion models. arXiv preprint arXiv:2302.04867, 2023. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[49] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Ngai-Man Cheung, and Min Lin. A recipe for watermarking diffusion models. arXiv preprint arXiv:2303.10137, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.461, + 0.892, + 0.516 + ], + "angle": 0, + "content": "[50] Xin Zhong, Pei-Chi Huang, Spyridon Mastorakis, and Frank Y Shih. An automated and robust image watermarking scheme based on deep neural networks. IEEE Transactions on Multimedia, 23:1951-1961, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.518, + 0.894, + 0.572 + ], + "angle": 0, + "content": "[51] Jiren Zhu, Russell Kaplan, Justin Johnson, and Li Fei-Fei. Hidden: Hiding data with deep networks. In Proceedings of the European conference on computer vision (ECCV), pages 657–672, 2018. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12171" + } + ] +] \ No newline at end of file diff --git a/2024/Gaussian Shading_ Provable Performance-Lossless Image Watermarking for Diffusion Models/cf093428-bde9-4e71-b54b-2506850cecda_origin.pdf b/2024/Gaussian Shading_ Provable Performance-Lossless Image Watermarking for Diffusion Models/cf093428-bde9-4e71-b54b-2506850cecda_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..092cf476f73e0530cd59582d78a46ea0887bc61c --- /dev/null +++ b/2024/Gaussian Shading_ Provable Performance-Lossless Image Watermarking for Diffusion Models/cf093428-bde9-4e71-b54b-2506850cecda_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a8ec896b102e89e30004d0f235a62d4835a83e42e2b823ae442e136fd27f9ea7 +size 5439354 diff --git a/2024/Gaussian Shading_ Provable Performance-Lossless Image Watermarking for Diffusion Models/full.md b/2024/Gaussian Shading_ Provable Performance-Lossless Image Watermarking for Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..370c8c034a2c80943d81cee45f7208d67b0fe8bb --- /dev/null +++ b/2024/Gaussian Shading_ Provable Performance-Lossless Image Watermarking for Diffusion Models/full.md @@ -0,0 +1,392 @@ +# Gaussian Shading: Provable Performance-Lossless Image Watermarking for Diffusion Models + +Zijin Yang $^{1}$ , Kai Zeng $^{1}$ , Kejiang Chen $^{1,*}$ , Han Fang $^{2}$ , Weiming Zhang $^{1}$ , Nenghai Yu $^{1}$ , $^{1}$ Anhui Province Key Laboratory of Digital Security, University of Science and Technology of China $^{2}$ National University of Singapore + +{bsmhmmlf@mail.,zk0128@mail.,chenkj@,zhangwm@,ynh@}ustc.edu.cn fanghan@nus.edu.sg + +# Abstract + +Ethical concerns surrounding copyright protection and inappropriate content generation pose challenges for the practical implementation of diffusion models. One effective solution involves watermarking the generated images. However, existing methods often compromise the model performance or require additional training, which is undesirable for operators and users. To address this issue, we propose Gaussian Shading, a diffusion model watermarking technique that is both performance-lossless and training-free, while serving the dual purpose of copyright protection and tracing of offending content. Our watermark embedding is free of model parameter modifications and thus is plug-and-play. We map the watermark to latent representations following a standard Gaussian distribution, which is indistinguishable from latent representations obtained from the non-watermarked diffusion model. Therefore we can achieve watermark embedding with lossless performance, for which we also provide theoretical proof. Furthermore, since the watermark is intricately linked with image semantics, it exhibits resilience to lossy processing and erasure attempts. The watermark can be extracted by Denoising Diffusion Implicit Models (DDIM) inversion and inverse sampling. We evaluate Gaussian Shading on multiple versions of Stable Diffusion, and the results demonstrate that Gaussian Shading not only is performance-lossless but also outperforms existing methods in terms of robustness. + +# 1. Introduction + +Diffusion models [16, 31-34] signify a noteworthy leap forward in image generation. These well-trained diffusion models, especially commercial diffusion models like Stable Diffusion (SD) [30], Glide [27], and Muse AI [30], enable individuals with diverse backgrounds to create high-quality images effortlessly. However, this raises concerns about + +![](images/bcf03ab1c92533925c1aa3c8a58d737b9fb74127d1de4e4f035c46fea24683cd.jpg) +Figure 1. Existing watermarking frameworks can be divided into three categories: post-processing-based, fine-tuning-based, and latent-representation-based Tree-Ring. Our method also relies on latent representations but achieves performance-lossless without altering the distribution. + +intellectual property and whether diffusion models will be stolen or resold twice. + +On the other hand, the ease of generating realistic images raises concerns about potentially misleading content generation. For example, on May 23, 2023, a Twitter-verified user named Bloomberg Feed posted a tweet titled "Large explosion near the Pentagon complex in Washington DC-initial report," along with a synthetic image. This tweet led to multiple authoritative media accounts sharing it, even causing a brief impact on the stock market1. On October 30, 2023, White House issued an executive order on AI security, emphasizing the need to protect Americans from AI-enabled fraud and deception by establishing standards and best practices for detecting AI-generated content and authenticating official content2. The urgency of labeling gen + +erated content for copyright authentication and prevention of misuse is evident. + +Watermarking is highlighted as a fundamental method for labeling generated content, as it embeds watermark information within the generated image, allowing for subsequent copyright authentication and the tracking of false content. Existing watermarking methods for the diffusion model can be divided into three categories, as shown in Fig. 1. Post-processing-based watermarks [6, 45] adjust robust image features to embed watermarks, thereby directly altering the image and degrading its quality. To mitigate this concern, recent research endeavors propose fine-tuning-based methods [7, 10, 24, 42, 49], which amalgamate the watermark embedding process with the image generation process. Intuitively, these methods need to modify model parameters, introducing supplementary computational overhead. Recently, Wen et al. [41] proposed the latent-representation-based Tree-Ring watermark, which conveys information by adapting the latent representations to match specific patterns. However, it restricts the randomness of sampling, which impacts generative performance. + +Through the above analysis, we can find that these methods compromise model performance to embed watermarks. In practical applications, model performance is paramount for both business interests and user experience. Substantial resource investment is often necessary to pursue enhanced model performance. This leads to a fundamental question: Can watermarks be embedded without compromising model performance? + +We affirmatively address the question presented above. Succinctly, the generation process can be delineated into two key phases: latent representation sampling and decoding. Our goal is to align the distribution of the latent representation in watermarked images with that of the latent representation in normally generated images. By keeping the model unaltered, the distribution of watermarked images is naturally consistent with that of normally generated images, enabling the seamless embedding of watermarks without compromising model performance. + +Building upon this insight, we propose a watermarking method named Gaussian Shading, designed to ensure no deterioration in model performance. The embedding process encompasses three primary elements: watermark diffuse, randomization, and distribution-preserving sampling. Watermark diffusion spreads the watermark information throughout the latent representation. During the generation process, the watermark information will be diffused to the whole semantics of the image, thus achieving excellent robustness. Watermark randomization and distribution-preserving sampling guarantee the congruity of the latent representation distribution with that of watermark-free latent representations, thereby achieving performance-lossless. In the extraction phase, the latent representations + +are acquired through Denoising Diffusion Implicit Model (DDIM) inversion [32], allowing for the retrieval of watermark information. Harnessing the extensive scope of the SD latent space, we can achieve a high-capacity watermark of 256 bits, surpassing prior methods. + +To the best of our knowledge, ours is the first technique that tackles this challenging problem of performanceless watermarking for diffusion models, and we provide theoretical proof. Moreover, this technique leaves the architecture and parameters of SD unaltered, necessitating no supplementary training. It can seamlessly integrate as a plug-and-play module within the generation process. Model providers can easily replace watermarked models with non-watermarked ones without affecting usability experiences. + +We conducted thorough experiments on SD. Under strong noise perturbation, the average true positive rate and bit accuracy can exceed 0.99 and 0.97, respectively, validating the superiority of Gaussian Shading in both detection and traceability tasks compared to prior methods. Additionally, experiments on visual quality and image-text similarity serve as indicators of performance preservation in our approach. Lastly, we deliberated on various watermark erasure attacks, affirming the steadfast performance of our watermark in the face of such adversities. + +# 2. Related Work + +# 2.1. Diffusion Models + +Inspired by non-equilibrium thermodynamics, Ho et al. [16] introduced the Denoising Diffusion Probabilistic Model (DDPM). DDPM consists of two Markov chains used for adding and removing noise, and subsequent works [8, 11, 15, 25, 28, 30, 32] have adopted this bidirectional chain framework. To reduce computational complexity and improve efficiency, the Latent Diffusion Model (LDM) [30] was designed, in which the diffusion process occurs in a latent space $\mathcal{Z}$ . To map an image $x \in \mathbb{R}^{H \times W \times 3}$ to the latent space, the LDM employs an encoder $\mathcal{E}$ , such that $z_0 = \mathcal{E}(x) \in \mathbb{R}^{h \times w \times c}$ . Similarly, to reconstruct an image from the latent space, a decoder $\mathcal{D}$ is used, such that $x = \mathcal{D}(z_0)$ . A pretrained LDM can generate images without the encoder $\mathcal{E}$ . Specifically, a latent representation $z_T$ is first sampled from a standard Gaussian distribution $\mathcal{N}(0, I)$ . Subsequently, through iterative denoising using methods like DDIM [32], $z_0$ is obtained, and an image can be generated using the decoder: $x = \mathcal{D}(z_0)$ . + +# 2.2. Image Watermarking + +Digital watermarking [38] is an effective means to address copyright protection and content authentication by embedding copyright or traceable identification information within carrier data. Typically, the functionality of a watermark depends on its capacity. For example, a single-bit + +![](images/f1602cb88675e865b5ea5a86ef9698a06b6c8a692e398be731b30b89df7b4ea5.jpg) +Figure 2. Application scenarios for Gaussian Shading. + +watermark can determine whether an image was generated by a particular diffusion model, i.e., copyright protection; a multi-bit watermark can further determine which user of the diffusion model generated the image, i.e., traceability. + +Image watermarking is a method that employs images as carriers for watermarking. Initially, watermark embedding methods primarily focused on the spatial domain [38], but later, to enhance robustness, transform domain watermarking techniques [1, 12, 13, 20, 22, 35, 37] were developed. In recent years, with the advancement of deep learning, researchers have turned their attention to neural networks [21, 39], harnessing their powerful learning capabilities to develop watermarking techniques [18, 19, 26, 36, 43, 50, 51]. + +# 2.3. Image Watermarking for Diffusion Models + +Existing Image watermarking methods for the diffusion model [6, 7, 10, 24, 41, 42, 45, 49] can be divided into three categories, as shown in Fig. 1. The image watermarking methods described in the previous section can be applied directly to the images generated by the diffusion model, which is called post-processing-based watermarks [6, 45]. These methods directly modify the image, thus degrading image quality. Recent research endeavors have amalgamated the watermark embedding process with the image generation process to mitigate this issue. Stable Signature [10] fine-tunes the LDM decoder using a pre-trained watermark extractor, facilitating watermark extraction from images produced by the fine-tuned model. Zhao et al. [49] and Liu et al. [24] suggest fine-tuning the diffusion model to implant a backdoor as a watermark, enabling watermark extraction by triggering. These fine-tuning-based approaches enhance the quality of watermarked images but introduce supplementary computational overhead and modify model parameters. Furthermore, Wen et al. [41] introduced the Tree-Ring Watermark, which conveys copyright information by adapting the frequency domain of latent representations to match specific patterns. This method achieves an imperceptible watermark. However, it directly disrupts the Gaussian distribution of noise, limiting the randomness of sampling and resulting in affecting model performance. + +# 3. Methods + +In this section, we provide an overview of the application scenarios and functionalities in Fig. 2. We then proceed + +to detail the embedding and extraction processes shown in Fig. 3. Finally, we present a mathematical proof of the performance-lossless characteristic of the watermark. + +# 3.1. Application Scenarios + +Scenarios. See Fig. 2, the scenario involves the operator Alice, the thief Carol, and two types of users Bob and Trudy. + +Alice is responsible for training the model, deploying it on the platform, and providing the corresponding API for users, but she does not open-source the code or model weights. Carol does not use Alice's services but steals images generated by her model, claiming ownership of the copyrights. Bob and Trudy, as community users, can utilize the API to generate and disseminate images. While Bob faithfully adheres to the community guidelines, Trudy aims to generate deep fake, and infringing content. To evade detection and traceability, Trudy can employ various data augmentation to modify illicit images. + +Detection. This scenario satisfies the detection (copyright protection) requirement. Alice embeds a single-bit watermark into each generated image. The successful extraction of the watermark from an image serves as evidence of Alice's rightful ownership of the copyright, while also indicating that the image is artificially generated (as opposed to natural images). + +Traceability. This scenario fulfills the traceability requirement. Alice allocates a watermark to each user. By extracting the watermark from the illicit content, it enables tracing Trudy, through comparison with the watermark database. Traceability is a higher pursuit than detection and can also achieve copyright protection for different users. + +Details of the statistical tests in both scenarios are shown in Supplementary Material. + +# 3.2. Watermark Embedding + +Watermark diffusion. The dimensions of the latent representations are given by $c \times h \times w$ , where each dimension can represent $l$ bits of the watermark. Therefore, the watermark capacity becomes $l \times c \times h \times w$ bits. To enhance the robustness of the watermark, we represent the watermark using $\frac{1}{f_{hw}}$ of the height and width, and $\frac{1}{f_c}$ of the channel, and replicate the watermark $f_c \cdot f_{hw}^2$ times. Thus, the watermark $s$ with dimensions $l \times \frac{c}{f_c} \times \frac{h}{f_{hw}} \times \frac{w}{f_{hw}}$ is expanded into a diffused watermark $s^d$ with dimensions $l \times c \times h \times w$ . + +![](images/e61d5fab0bfa788917f1acf870e9289c022719614719f736758469ebe946cc65.jpg) +Figure 3. The framework of Gaussian Shading. We utilize a $k$ -bit binary sequence $s$ to represent the watermark. After diffusion and encryption, the watermark can be utilized to drive distribution-preserving sampling, followed by denoising to generate watermarked images $X^s$ . For extraction, it is sufficient to introduce DDIM inversion and the inverse process of all the operations mentioned above. + +The actual watermark capacity is $k = \frac{l \times c \times h \times w}{f_c \cdot f_{hw}^2}$ bits. + +Watermark randomization. If we know the distribution of the diffused watermark $s^d$ , we can directly utilize distribution-preserving sampling to obtain the corresponding latent representations $z_T^s$ . However, in practical scenarios, its distribution is always unknown. Hence, we introduce a stream key $K$ to transform $s^d$ into a distribution-unknown randomized watermark $m$ through encryption. Considering the use of computationally secure stream cipher, such as ChaCha20 [3], $m$ follows a uniform distribution, i.e., $m$ is a random binary bit stream. + +Distribution-preserving sampling driven by randomized watermark. When each dimension represents $l$ -bit randomized watermark $m$ , this $l$ bits can be regarded as an integer $y \in [0, 2^l - 1]$ . Since $m$ is a ciphertext, $y$ follows a discrete uniform distribution, i.e., $p(y) = \frac{1}{2^l}$ for $y = 0, 1, 2, \ldots, 2^l - 1$ . Let $f(x)$ denote the probability density function of the Gaussian distribution $\mathcal{N}(0, I)$ , and $ppf$ denotes the quantile function. We divide $f(x)$ into $2^l$ equal cumulative probability portions. When $y = i$ , the watermarked latent representation $z_T^s$ falls into the $i$ -th interval, which means $z_T^s$ should follow the conditional distribution: + +$$ +p \left(z _ {T} ^ {s} \mid y = i\right) = \left\{ \begin{array}{c c} 2 ^ {l} \cdot f \left(z _ {T} ^ {s}\right) & p p f \left(\frac {i}{2 ^ {l}}\right) < z _ {T} ^ {s} \leq p p f \left(\frac {i + 1}{2 ^ {l}}\right) \\ 0 & o t h e r w i s e \end{array} . \right. \tag {1} +$$ + +The probability distribution of $z_T^s$ is given by: + +$$ +p \left(z _ {T} ^ {s}\right) = \sum_ {i = 0} ^ {2 ^ {l} - 1} p \left(z _ {T} ^ {s} \mid y = i\right) p (y = i) = f \left(z _ {T} ^ {s}\right). \tag {2} +$$ + +Eq. (2) indicates that $z_T^s$ follows the same distribution as the randomly sampled latent representation $z_T \sim \mathcal{N}(0, I)$ . Next, we elaborate on how this sampling is implemented. + +Let the cumulative distribution function of $f(x)$ be denoted as $cdf$ . We can obtain the cumulative distribution + +function of Eq. (1) as follows, + +$$ +\begin{array}{l} F (z _ {T} ^ {s} | y = i) \\ = \left\{ \begin{array}{c c} 0 & z _ {T} ^ {s} < p p f \left(\frac {i}{2 ^ {l}}\right) \\ 2 ^ {l} \cdot c d f \left(z _ {T} ^ {s}\right) - i & p p f \left(\frac {i}{2 ^ {l}}\right) \leq z _ {T} ^ {s} \leq p p f \left(\frac {i + 1}{2 ^ {l}}\right) \\ 1 & z _ {T} ^ {s} > p p f \left(\frac {i + 1}{2 ^ {l}}\right) \end{array} . \right. \tag {3} \\ \end{array} +$$ + +Given $y = i$ , we aim to perform random sampling of $z_{T}^{s}$ within the interval $[ppf(\frac{i}{2^t}), ppf(\frac{i + 1}{2^t})]$ . The commonly used method is rejection sampling [4, 17, 47], which can be time-consuming as it requires repeated sampling until $z_{T}^{s}$ falls into the correct interval. Instead, we can utilize the cumulative probability density. When randomly sampling $F(z_{T}^{s}|y = i)$ , the corresponding $z_{T}^{s}$ is naturally obtained through random sampling. Since $F(z_{T}^{s}|y = i)$ takes values in $[0,1]$ , sampling from it is equivalent to sampling from a standard uniform distribution, denoted as $u = F(z_{T}^{s}|y = i) \sim \mathcal{U}(0,1)$ . Shift the terms of Eq. (3), and take into account that $cdf$ and $ppf$ are inverse functions, we have + +$$ +z _ {T} ^ {s} = p p f \left(\frac {u + i}{2 ^ {l}}\right). \tag {4} +$$ + +Eq. (4) represents the process of sampling the watermarked latent representation $z_{T}^{s}$ driven by the randomized watermark $m$ . To extract the watermark, its inverse map is + +$$ +i = \left\lfloor 2 ^ {l} \cdot c d f \left(z _ {T} ^ {s}\right) \right\rfloor . \tag {5} +$$ + +Image generation. After the sampling process, the watermark is embedded in the latent representation $z_T^s$ , and the subsequent generation process is no different from the regular generation process of SD. Here, we employ the DPM-Solver [25] for iterative denoising of $z_T^s$ . In addition to DPMSolver [25], other continuous-time samplers based on ordinary differential equation (ODE) solvers [32], such as DDIM [32], DEIS [46], PNDM [23], and UniPC [48], can be used too. After obtaining denoised $z_0^s$ , the watermarked image $X^s$ is generated using the decoder $\mathcal{D}$ : $X^s = \mathcal{D}(z_0^s)$ . + +# 3.3. Watermark Extraction + +DDIM Inversion. Using the SD encoder $\mathcal{E}$ , we first restore $X^{\prime s}$ to the latent space $z_0^{\prime s} = \mathcal{E}(X^{\prime s})$ . Then, we introduce the DDIM inversion [32] to estimate the additive noise. It can be considered that $z_T^{\prime s} \approx z_T^s$ . We also observe that although DDIM inversion is derived from DDIM, it can apply to other continuous-time samplers based on ODE solvers. + +Watermark reduction from latent representations. After obtaining $z_T^{\prime s}$ , according to the inverse transformation defined in Eq. (5), the tensor can be converted into a bit stream $m'$ . Subsequently, $m'$ is decrypted using $K$ to obtain $s'^d$ . Inverse diffusion of the watermark results in $f_c \cdot f_{hw}^2$ copies of the watermark. Similar to voting, if the bit is set to 1 in more than half of the copies, the corresponding watermark bit is set to 1; otherwise, it is set to 0. This process restores the true binary watermark sequence $s'$ . + +# 3.4. Proof of Lossless Performance + +In prior works, the incorporation of watermark embedding modules inevitably results in a decline in model performance, as typically evaluated using metrics such as Peak Signal-to-Noise Ratio (PSNR) and Fréchet Inception Distance (FID) [14], which are more suitable for assessing post-processing methods. To assess methods that integrate the watermark embedding and generation processes, we propose a definition for the impact of watermark embedding on model performance, drawing on the complexity-theoretic definition of steganographic security [17]. This definition is based on a probabilistic game between a watermarked image $X^{s}$ and a normally generated image $X$ . The tester $\mathcal{A}$ can use any watermark to drive the sampling process and generate $X^{s}$ , similar to the chosen hidden text attacks [17], which we refer to as chosen watermark tests. The watermarking method is performance-lossless under chosen watermark tests, if for any polynomial-time tester $\mathcal{A}$ and key $K \gets \text{KeyGen}_{\mathcal{G}(1^{\rho})}$ , it holds that + +$$ +\left| \Pr \left[ \mathcal {A} \left(X ^ {s}\right) = 1 \right] - \Pr \left[ \mathcal {A} (X) = 1 \right] \right| < \operatorname {n e g l} (\rho). \tag {6} +$$ + +Here, $\rho$ represents the length of the security parameter, such as the key $K$ , and $\mathsf{negl}(\rho)$ is a negligible term relative to $\rho$ . + +We prove the statement using a proof by contradiction. First, assume that the watermarked image $X^s$ and the normally generated image $X$ are distinguishable, meaning + +$$ +\left. \right.\left| \Pr \left[ \mathcal {A} \left(X ^ {s}\right) = 1 \right] - \Pr \left[ \mathcal {A} (X) = 1 \right]\right| = \delta , \tag {7} +$$ + +where $\delta$ is non-negligible with respect to the key $K$ . Let the iterative denoising process be denoted as $Q(\cdot)$ , and substitute the LDM encoder $\mathcal{E}$ into Eq. (7), we have + +$$ +\begin{array}{l} \left| \Pr \left[ \mathcal {A} \left(\mathcal {E} \left(Q \left(z _ {T} ^ {s}\right)\right)\right) = 1 \mid m = E (K, s ^ {d}) \right] \right. \\ - \Pr \left[ \mathcal {A} \left(\mathcal {E} \left(Q \left(z _ {T}\right)\right)\right) = 1 \mid z _ {T} \leftarrow \mathcal {N} (0, I) \right] | = \delta , \\ \end{array} +$$ + +where the randomized watermark $m$ is obtained by encrypting the diffused watermark $s^d$ using the encryption algorithm $E$ with key $K$ . Note that Eq. (2) contains the fact that + +distribution-preserving sampling driven by randomized watermark and random sampling are equivalent. Therefore, we denote sequence-driven sampling as $S(\cdot)$ . $z_{T}^{s}$ can naturally be obtained by sampling driven by $m$ , i.e., $z_{T}^{s} = S(m)$ . On the other hand, $z_{T}$ can be considered as obtained by sampling driven by a truly random sequence $r$ of the same length as $m$ , i.e., $z_{T} = S(r)$ . Eq. (8) can be written as + +$$ +\begin{array}{l} \left| \Pr [ \mathcal {A} (\mathcal {E} (Q (S (m)))) = 1 | m = E (K, s ^ {d}) ] \right. \tag {9} \\ - \Pr \left[ \mathcal {A} \left(\mathcal {E} (Q (S (r))) = 1 \right\rvert \right] = \delta . \\ \end{array} +$$ + +Sampling $S(\cdot)$ , denoising $Q(\cdot)$ , and encoder $\mathcal{E}$ can be considered as subroutines that the tester $\mathcal{A}_{\mathcal{E}, Q, S}$ can use. Thus, Eq. (9) can be simplified, + +$$ +\begin{array}{l} \left| \Pr \left[ \mathcal {A} _ {\varepsilon , Q, S} (m) = 1 \mid m = E (K, s ^ {d}) \right] \right. \tag {10} \\ - \Pr \left[ \mathcal {A} _ {\mathcal {E}, Q, S} (r) = 1 \right] | = \delta . \\ \end{array} +$$ + +Note that $S(\cdot), Q(\cdot)$ , and $\mathcal{E}$ are all polynomial-time programs, so the time taken by the tester $\mathcal{A}_{\mathcal{E},Q,S}$ to make the distinction is also polynomial. Eq. (10) essentially states that it is possible to distinguish between $m$ and $r$ in polynomial time. However, we have used the computationally secure stream cipher ChaCha20 [3] in watermark randomization, which means that $m$ as a pseudorandom sequence cannot be distinguished from a truly random sequence in polynomial time. Eq. (10) contradicts the computational security property of ChaCha20 [3]. Therefore, Eq. (10) is not valid, leading us back to our initial assumption that Eq. (7) is also not valid. This implies that the watermarked image $X^s$ and the normally generated image $X$ are indistinguishable in polynomial time. Hence, Gaussian Shading is performance-lossless under chosen watermark tests. + +# 4. Experiments + +This section focuses on experimental analysis, including details of the experimental setup, performance evaluation of Gaussian Shading, comparison with baseline methods, ablation experiments, and potential attacks. + +# 4.1. Implementation Details + +SD models. In this paper, we focus on text-to-image LDM, hence we select SD [30] provided by huggingface. We evaluate Gaussian Shading as well as baseline methods, using three versions of SD: V1.4, V2.0, and V2.1. The size of the generated images is $512 \times 512$ , and the latent space dimension is $4 \times 64 \times 64$ . During inference, we employ the prompt from Stable-Diffusion-Prompt3, with a guidance scale of 7.5. We sample 50 steps using DPMSolver [25]. Considering that users tend to propagate the generated images without retaining the corresponding prompts, we use an empty prompt for inversion, with a scale of 1. We perform 50 steps of inversion using DDIM inversion [32]. + +![](images/113f92213137cf7560fc650a7bc87aa7fa6167595fbc53eb9e6c952236ba331c.jpg) + +![](images/391c5ea77b52a7c6e9ad6089279876a9caf13725122d62529f2361a9e94e4313.jpg) + +![](images/2bc54b6f2c460407df5ec3c81d88e2f7a9f96762a6c17332b8cce2eca805f13c.jpg) + +![](images/a00616e79178046b18344bdef487b7c122716a955f897da335aa553de66f550e.jpg) + +![](images/7a1e6ed37b87a66ec8e8ddf1676f6702fc25431a7fb153cb16423fa84918c87a.jpg) + +![](images/78308b2122f914f4fac6cf0abe8e2988f0188e26a62c9f26c1470943e865b0d4.jpg) +(f) + +![](images/3c04e4d0e837a921c04aabbbd4ebaf3cb8aa7899b3418d1658e7736585bc880e.jpg) +(b) +(g) + +![](images/f4b394ed8fb9227ff91deaab30c954771c7a2118d94f43320d95ddb0bf42850b.jpg) +(h) + +![](images/44eee57d50c3eeced27c41c772a82e00c47783d7cea6ed9483617ca300b13e12.jpg) +(d) +(i) +Figure 4. Watermarked image is attacked by different noise. (a) Watermarked image. (b) JPEG, $QF = 25$ . (c) $60\%$ area Random Crop (RandCr). (d) $80\%$ area Random Drop (RandDr). (e) Gaussian Blur, $r = 4$ (GauBlur). (f) Median Filter, $k = 7$ (MedFilter). (g) Gaussian Noise, $\mu = 0$ , $\sigma = 0.05$ (GauNoise). (h) Salt and Pepper Noise, $p = 0.05$ (S&PNoise). (i) $25\%$ Resize and restore (Resize). (j) Brightness, factor $= 6$ . + +![](images/e724f5fe89325981724032fd706b40f092a41863e3b1de3c6d07a334bb23ccca.jpg) +(j) + +Watermarking methods. In the main experiments, the settings for Gaussian Shading are $f_{c} = 1$ , $f_{hw} = 8$ , $l = 1$ , resulting in an actual capacity of 256 bits. We select five baseline methods: three officially used by SD, namely Dwt-Dct [6], DwtDctSvd [6], and RivaGAN [45], a multi-bit watermarking called Stable Signature [10], and a train-free invisible watermarking called Tree-Ring [41]. + +Robustness evaluation To evaluate the robustness, we select nine representative types of noise shown in Fig. 4. We conduct experiments following the noise strength in Fig. 4. + +Evaluation metrics. In the detection scenario, we calculate the true positive rate (TPR) corresponding to a fixed false positive rate (FPR). In the traceability scenario, we calculate the bit accuracy. To measure the bias in model performance, we compute the FID [14] and CLIP-Score [29] for 10 batches of watermarked images and perform a $t$ -test on the mean FID and CLIP-Score compared to that of watermark-free images. + +All experiments are conducted using the PyTorch 1.13.0 framework, running on a single RTX 3090 GPU. + +# 4.2. Performance of Gaussian Shading + +Detection. In the detection scenario, we consider Gaussian Shading as a single-bit watermark, with a fixed watermark $s$ . We approximate the FPR to be controlled at $10^{0}, 10^{-1}, \ldots, 10^{-13}$ , calculate the corresponding threshold $\tau$ , and test the TPR on 1,000 watermarked images To mitigate the effects of randomness, we perform 5 trials with different $s$ and compute the average TPR. See Fig. 5a, when the FPR is controlled at $10^{-13}$ , the TPR remains at least 0.99 for eight out of the nine cases. Although the TPR for Brightness is only 0.953, it is still a promising result. + +Traceability. In this scenario, Gaussian Shading serves as a multi-bit watermark. Assuming Alice provides services to $N$ users, Alice needs to allocate one watermark for each user. In our experiments, we assume that $N' = 1,000$ users generate images, with each user generating 10 images, re + +![](images/774162d867b79361428a6cb94c725181280f963f31253856bb88a5940fa1d48c.jpg) +(a) Detection results. +Figure 5. Performance of Gaussian Shading. + +![](images/aa640412f42f561a96912774a9dce03db8c6cc463236ac6429791ff69841e72a.jpg) +(b) Traceability results. + +sulting in a dataset of 10,000 watermarked images. + +During testing, we calculate the threshold $\tau$ to control the FPR at $10^{-6}$ . Note that when computing traceability accuracy, we need to consider two types of errors: false positives, where watermarked images are not detected, and traceability errors, where watermarked images are detected but attributed to the wrong user. Therefore, we first determine whether the image contains a watermark. If it does, we calculate the number of matching bits $Acc$ with all $N$ users on the platform. The user with the highest $Acc$ is considered the one who generated the image. Finally, we verify whether the correct user has been traced. When $N > N'$ , it can be assumed that some users have been assigned a watermark but have not generated any images. + +See Fig. 5b, when $N = 10^{6}$ , Gaussian Shading exhibits almost perfect traceability in seven cases. Although the traceability accuracy for Brightness is only $95.47\%$ , if a user generates two images, the probability of successfully tracing him is still no less than $99\%$ . + +# 4.3. Comparison to Baselines + +In this section, we compare the performance of Gaussian Shading with baselines on SD V1.4, V2.0, and V2.1. We use our implementations for each method, see details in Supplementary Material. + +We conduct tests on 1,000 generated images for each method respectively. See Tab. 1. Gaussian Shading exhibits strong robustness and significantly outperforms baselines in both scenarios. In terms of bit accuracy, it surpasses the best-performing baseline by approximately $7\%$ . This can be attributed to the extensive diffusion of the watermark throughout the entire latent space, establishing a profound binding between the watermark and the image semantics. + +To measure the performance bias introduced by the watermark embedding, we apply a $t$ -test to evaluate. The hypotheses are $H_0: \mu_s = \mu_0, H_1: \mu_s \neq \mu_0$ , where $\mu_s$ and $\mu_0$ represent the average FID [14] or CLIP-Score [29] of multiple sets of watermarked and watermark-free images, respectively. A lower $t$ -value indicates a higher probability that $H_0$ holds. If the $t$ -value is larger than a threshold, $H_0$ is rejected, and model performance is considered to have been affected. See Tab. 1, Gaussian Shading achieves the smallest $t$ -value, which indirectly reflects its performance-lossless characteristic. For a detailed analysis of the $t$ -test, please refer to the Supplementary Material. + +
MethodsMetrics
TPR (Clean)TPR (Adversarial)Bit Acc. (Clean)Bit Acc. (Adversarial)FID (t-value ↓)CLIP-Score (t-value ↓)
Stable Diffusion----25.23±.180.3629±.0006
DwtDct [6]0.825/0.881/0.8660.172/0.178/0.1730.8030/0.8059/0.80230.5696/0.5671/0.562224.97±.19 (3.026)0.3617±.0007 (3.045)
DwtDctSvd [6]1.000/1.000/1.0000.597/0.594/0.5990.9997/0.9987/0.99870.6920/0.6868/0.690524.45±.22 (8.253)0.3609±.0009 (4.452)
RivaGAN [45]0.920/0.945/0.9630.697/0.697/0.7060.9762/0.9877/0.99210.8986/0.9124/0.901924.24±.16 (12.29)0.3611±.0009 (4.259)
Tree-Ring [41]1.000/1.000/1.0000.894/0.898/0.906--25.43±.13 (2.581)0.3632±.0006 (0.8278)
Stable Signature [10]1.000/1.000/1.0000.502/0.505/0.4960.9987/0.9978/0.99790.7520/0.7472/0.750025.45±.18 (2.477)0.3622±.0027 (0.7066)
Ours1.000/1.000/1.0000.997/0.998/0.9960.9999/0.9999/0.99990.9753/0.9749/0.972425.20±.22 (0.3567)0.3631±.0005 (0.6870)
+ +Table 1. Comparison results. We control the FPR at $10^{-6}$ , and evaluate the TPR and bit accuracy for SD V1.4/V2.0/V2.1. To assess the bias in model performance, we conduct a $t$ -test on SD V2.1. Adversarial here refers to the average performance of a series of noises. Additional results can be found in Supplementary Material. + +![](images/db4c4943a11b71c4f21ace1b07c62f59f4d4daf753525c2cda36cf1450e25b33.jpg) + +![](images/b84c926be2f4085637d5b5c005f44ffb54fa0ad2895819f40434dfc0a5df6b94.jpg) + +![](images/87353aed10bc3304f8afcd28ee3be310f682ffb1cfa196dbb473192af11974be.jpg) + +![](images/1849f464e6b319de72170104ae614ecc91f64b614bc08e5901c5cb9716282939.jpg) + +![](images/acc123a5fbb8846a038efd042a4ce1bbab5e99ba593033d37ec567f62895b3e5.jpg) + +![](images/d14af2497e22ab79cac4adf3bc18bb022c2d364c466615121b42a3e96c298b49.jpg) +(a) Guidance scales. +(f) Median Filter. + +![](images/336e0f66502a768f3fd7a00d31195f7d9a2a7a4135001346c9868fbcbec5e3bb.jpg) +(b) JPEG +(g) Gaussian Noise. + +![](images/ad94a8eb28303917d25d49dfcd8b99f9fbf0ec49a6a48c1c88eff66d6f002d36.jpg) +(c) Random Crop. +(h) Salt and Pepper Noise. +Figure 6. Ablation studies. + +![](images/d5c033138c9afd948058844db3e2be752c786e291bc9636a0f800e456fc3b839.jpg) +(d) Random Drop. +(i) Resize. + +![](images/66a5694f8e5461951c77626d1771d39b23d6d145c3288595c87e27d5bc80a502.jpg) +(e) Gaussian Blur. +(j) Brightness. + +# 4.4. Ablation Studies + +In this section, we conduct comprehensive ablation experiments on SD V2.1 to demonstrate hyperparameter selection. Unless specified, we generate 1,000 images and test the TPR and the bit accuracy with a theoretical FPR of $10^{-6}$ . + +Watermark capacity. The watermark capacity is determined by three parameters: channel diffusion factor $f_{c}$ , height-width diffusion factor $f_{hw}$ , and embedding rate $l$ . See Tab. 2, to balance the capacity and robustness of Gaussian Shading, we chose $f_{c} = 1$ and $f_{hw} = 8$ . After fixing $f_{c}$ and $f_{hw}$ , we vary $l$ to examine if it could enhance the capacity, and additional results can be found in Supplementary Material. Considering all factors, we determine that the optimal solution is $f_{c} = 1$ , $f_{hw} = 8$ , and $l = 1$ , resulting in a watermark capacity of 256 bits. + +Sampling methods. To validate the generalization, we select five commonly used sampling methods, all continuous-time samplers based on ODE solvers [32]. See Tab. 3, all of them exhibit excellent performance with a bit accuracy of approximately $97\%$ against noises. + +Impact of the inversion step. In practice, the inference step is often unknown, which introduces a mismatch with the inversion step. See Tab. 4, such mismatch introduces minimal loss in accuracy. Considering the high efficiency of existing samplers, the inference step generally does not exceed 50. Therefore, we set the inversion step to 50. + +Guidance scales. Given diverse user preferences for image + +prompt alignment, larger guidance scales ensure faithful adherence to prompts, while smaller scales grant the model greater creative freedom. In SD, the guidance scale is typically selected from the range of [5, 15]. Hence, experiments cover the range of 2 to 18. For the inversion, an empty prompt is used for guidance, and the guidance scale is fixed at 1, assuming unknown information during extraction. In Fig. 6a, the bit accuracy of Gaussian Shading surpasses $99.9\%$ , showing its reliability in real-world-like scenes. + +Noise intensities. To further test the robustness, we conduct experiments using different intensities of noises. See Figs. 6b to 6j, for Random Crop and Gaussian Noise, performance declines significantly with higher intensities. However, for the other seven types of noise, even at high intensities, the bit accuracy remains approximately $80\%$ . + +# 4.5. Attacks against Gaussian Shading + +We consider two malicious attacks: compression attack, where the attacker employs a neural network to compress watermarked images, and inversion attack, assuming the attacker is aware of the watermark embedding method, enabling them to modify the image's latent representations. + +Compression attack. We utilize popular auto-encoders [2, 5, 9, 30] to compare Stable Signature (SS) with Gaussian Shading across various compression rates. Additionally, we assess the compression quality through the PSNR between the compressed and watermarked images. See Figs. 7a + +
Noisefc-fhw (k bits)
1-2 (4096)4-1 (4096)1-4 (1024)4-2 (1024)1-8 (256)4-4 (256)1-16 (64)4-8 (64)
None0.94130.93800.99850.99800.99990.99991.00001.0000
Adversarial0.73020.72380.87690.86140.97240.96710.99590.9953
+ +Table 2. Bit accuracy with different factors ${f}_{c}$ and ${f}_{hw}$ ,where $l = 1$ . Additional results can be found in Supplementary Material. + +
NoiseSampling Methods
DDIM [32]UniPC [48]PNDM [23]DEIS [46]DPMSolver [25]
None0.99991.00001.00000.99990.9999
Adversarial0.97060.96280.97210.97150.9724
+ +Table 3. Bit accuracy with different sampling methods. Additional results can be found in Supplementary Material. These methods differ only in accuracy and order. DDIM is a first-order estimate of the ODE. Accordingly, DDIM inversion ensures a lower bound on the accuracy of the inversion process. Therefore, it can naturally be applied to higher-order and higher-accuracy methods. + +
Inference StepInversion Step
102550100
100.99990.99990.99990.9999
250.99980.99991.00001.0000
500.99950.99970.99990.9999
1000.99940.99960.99990.9999
+ +Table 4. Bit accuracy with different inference and inversion step. and 7b, Gaussian Shading significantly outperforms Stable Signature. This is because Gaussian Shading diffuses the watermark across the entire semantic space of images, while Stable Signature relies solely on the image texture. + +Inversion attack. Assuming the attacker is aware of the embedding method, a more effective approach to erasing is through inversion to obtain latent representations and subsequently modify them. We validate the robustness against such attacks. Importantly, our experiments assume the strongest attacker capability of using the same model as Alice for precise inversion. In real-world scenarios, where the watermark embedding is not publicly available, the attacker's capabilities would be weaker. + +Specifically, we perform inversion to obtain latent representations and randomly flip a certain rate of them. Using the flipped latent representations, we regenerate the images and extract the watermark. See Fig. 7c. the watermark can still be reliably extracted when the flipping rate (FR) is less than 0.4. At high FRs, significant changes in images are observed. Although the watermark cannot be accurately extracted, we consider the image transformed into a different one, resulting in the content not intended to be protected. + +From another perspective, the attacker can launch a forgery attack by performing inversion on an innocuous image from Bob and generating harmful content using a different prompt. See Fig. 7c, when the FR is 0, Alice can accurately trace Bob based on the forgeries, enabling the attacker to successfully frame Bob. Therefore, protecting the model from leakage is crucial for operators. + +![](images/79029d1b4edd943077dd20adea6a00b115b845037db4d141053738cfbf85dc13.jpg) +(a) Detection results. + +![](images/b1e3e8caa276f4aa3216dcbc155f939b4c6b9469e3389d584c640c634f685772.jpg) +(b) Traceability results. +Figure 7. Performance of Gaussian Shading under Malicious Attack, where (a) and (b) are under compression attack and (c) is under inversion attack. + +![](images/9d4b715f4373508fe44f20ea037ed3cd283cdaeef1e62d4c41cf8b7d50e169ba.jpg) +(c) Inversion attack. + +# 5. Limitations + +Despite extensive experimental validation of Gaussian Shading's superior performance, our work still has certain limitations. Firstly, the usage scenarios are restricted due to the reliance on DDIM inversion [32], which necessitates the utilization of continuous-time samplers based on ODE solvers [32] like DPMSolver [25]. Secondly, Gaussian Shading employs stream ciphers, necessitating proper key usage and management on the deployment platform. Additionally, we assume that the model is not publicly accessible, and only operators can verify the watermark, providing a certain level of protection against white-box attacks and ensuring security. However, if a legitimate third party requires watermark verification, cooperation from the operators becomes necessary. Lastly, Gaussian Shading is vulnerable to forgery attacks, emphasizing the importance for operators to safeguard the model parameters. + +# 6. Conclusion and Future Work + +We propose Gaussian Shading, a provably performance-lossless watermarking applied to diffusion models. Compared to baseline methods, Gaussian Shading offers simplicity and effectiveness by making a simple modification in the sampling process of the initial latent representation. Extensive experiments validate the superior performance in both detection and traceability scenarios. To our knowledge, we are the first to propose and implement a performance-lossless approach in image watermarking. + +Regarding future work, we will introduce more efficient inversion methods [40, 44] and include a wider range of sampling methods. Additionally, careful consideration should be given to counteracting forgery attacks. + +Acknowledgement. This work was supported in part by the Natural Science Foundation of China under Grant U2336206, 62102386, 62072421, 62372423, and 62121002. + +# References + +[1] Ali Al-Haj. Combined dwt-dct digital image watermarking. Journal of computer science, 3(9):740-746, 2007. 3 +[2] Johannes Ballé, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. arXiv preprint arXiv:1802.01436, 2018. 7 +[3] Daniel J Bernstein et al. Chacha, a variant of salsa20. In Workshop record of SASC, pages 3-5. Citeseer, 2008. 4, 5 +[4] Kejiang Chen, Hang Zhou, Hanqing Zhao, Dongdong Chen, Weiming Zhang, and Nenghai Yu. Distribution-preserving steganography based on text-to-speech generative models. IEEE Transactions on Dependable and Secure Computing, 19(5):3343-3356, 2021. 4 +[5] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned image compression with discretized gaussian mixture likelihoods and attention modules. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7939-7948, 2020. 7 +[6] Ingemar Cox, Matthew Miller, Jeffrey Bloom, Jessica Fridrich, and Ton Kalker. Digital watermarking and steganography. Morgan Kaufmann, 2007. 2, 3, 6, 7 +[7] Yingqian Cui, Jie Ren, Han Xu, Pengfei He, Hui Liu, Lichao Sun, and Jiliang Tang. Diffusionshield: A watermark for copyright protection against generative diffusion models. arXiv preprint arXiv:2306.04642, 2023. 2, 3 +[8] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2 +[9] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 7 +[10] Pierre Fernandez, Guillaume Couairon, Hervé Jégou, Matthijs Douze, and Teddy Furon. The stable signature: Rooting watermarks in latent diffusion models. arXiv preprint arXiv:2303.15435, 2023. 2, 3, 6, 7 +[11] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10696-10706, 2022. 2 +[12] Huiping Guo and Nicolas D Georganas. Digital image watermarking for joint ownership. In Proceedings of the tenth ACM international conference on Multimedia, pages 362-371, 2002. 3 +[13] Mohamed Hamidi, Mohamed El Haziti, Hocine Cherifi, and Mohammed El Hassouni. Hybrid blind robust image watermarking technique based on dft-dct and arnold transform. Multimedia Tools and Applications, 77:27181-27214, 2018. 3 +[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5, 6 + +[15] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2 +[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2 +[17] Nicholas J Hopper, John Langford, and Luis Von Ahn. Provably secure steganography. In Advances in Cryptology—CRYPTO 2002: 22nd Annual International Cryptology Conference Santa Barbara, California, USA, August 18–22, 2002 Proceedings 22, pages 77–92. Springer, 2002. 4, 5 +[18] Zhaoyang Jia, Han Fang, and Weiming Zhang. Mbrs: Enhancing robustness of dnn-based watermarking by minibatch of real and simulatedJPEG compression. In Proceedings of the 29th ACM international conference on multimedia, pages 41-49, 2021. 3 +[19] Varsha Kishore, Xiangyu Chen, Yan Wang, Boyi Li, and Kilian Q Weinberger. Fixed neural network steganography: Train the images, not the network. In International Conference on Learning Representations, 2021. 3 +[20] Deepa Kundur and Dimitrios Hatzinakos. A robust digital image watermarking method using wavelet-based fusion. In Proceedings of International Conference on Image Processing, pages 544-547. IEEE, 1997. 3 +[21] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 3 +[22] Sunil Lee, Chang D Yoo, and Ton Kalker. Reversible image watermarking based on integer-to-integer wavelet transform. IEEE Transactions on information forensics and security, 2 (3):321-330, 2007. 3 +[23] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. arXiv preprint arXiv:2202.09778, 2022. 4, 8 +[24] Yugeng Liu, Zheng Li, Michael Backes, Yun Shen, and Yang Zhang. Watermarking diffusion model. arXiv preprint arXiv:2305.12502, 2023. 2, 3 +[25] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 2, 4, 5, 8 +[26] Xiyang Luo, Ruohan Zhan, Huiwen Chang, Feng Yang, and Peyman Milanfar. Distortion agnostic deep watermarking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13548-13557, 2020. 3 +[27] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1 +[28] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 2 +[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, + +Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6 +[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 5, 7 +[31] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 1 +[32] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2020. 2, 4, 5, 7, 8 +[33] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. +[34] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2020. 1 +[35] Srdjan Stankovic, Irena Orovic, and Nikola Zaric. An application of multidimensional time-frequency analysis as a base for the unified watermarking approach. IEEE Transactions on Image Processing, 19(3):736-745, 2009. 3 +[36] Matthew Tancik, Ben Mildenhall, and Ren Ng. Stegastamp: Invisible hyperlinks in physical photographs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2117-2126, 2020. 3 +[37] Min-Jen Tsai, Kuang-Yao Yu, and Yi-Zhang Chen. Joint wavelet and spatial transformation for digital watermarking. IEEE Transactions on Consumer Electronics, 46(1):237, 2000. 3 +[38] Ron G Van Schyndel, Andrew Z Tirkel, and Charles F Osborne. A digital watermark. In Proceedings of 1st international conference on image processing, pages 86-90. IEEE, 1994. 2, 3 +[39] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3 +[40] Bram Wallace, Akash Gokul, and Nikhil Naik. Edict: Exact diffusion inversion via coupled transformations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22532-22541, 2023. 8 +[41] Yuxin Wen, John Kirchenbauer, Jonas Geiping, and Tom Goldstein. Tree-ring watermarks: Fingerprints for diffusion images that are invisible and robust. arXiv preprint arXiv:2305.20030, 2023. 2, 3, 6, 7 +[42] Cheng Xiong, Chuan Qin, Guorui Feng, and Xinpeng Zhang. Flexible and secure watermarking for latent diffusion model. In Proceedings of the 31st ACM International Conference on Multimedia, pages 1668-1676, 2023. 2, 3 + +[43] Chaoning Zhang, Philipp Benz, Adil Karjauv, Geng Sun, and In So Kweon. Udh: Universal deep hiding for steganography, watermarking, and light field messaging. Advances in Neural Information Processing Systems, 33:10223-10234, 2020. 3 +[44] Jiaxin Zhang, Kamalika Das, and Sricharan Kumar. On the robustness of diffusion inversion in image manipulation. In ICLR 2023 Workshop on Trustworthy and Reliable Large-Scale Machine Learning Models, 2023. 8 +[45] Kevin Alex Zhang, Lei Xu, Alfredo Cuesta-Infante, and Kalyan Veeramachaneni. Robust invisible video watermarking with attention. arXiv preprint arXiv:1909.01285, 2019. 2,3,6,7 +[46] Qinsheng Zhang and Yongxin Chen. Fast sampling of diffusion models with exponential integrator. arXiv preprint arXiv:2204.13902, 2022. 4, 8 +[47] Weiming Zhang, Kejiang Chen, and Nenghai Yu. Provable secure steganography: Theory, application and prospects. Journal of Cybersecurity, 1:38-46, 2023. 4 +[48] Wenliang Zhao, Lujia Bai, Yongming Rao, Jie Zhou, and Jiwen Lu. Unipc: A unified predictor-corrector framework for fast sampling of diffusion models. arXiv preprint arXiv:2302.04867, 2023. 4, 8 +[49] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Ngai-Man Cheung, and Min Lin. A recipe for watermarking diffusion models. arXiv preprint arXiv:2303.10137, 2023. 2, 3 +[50] Xin Zhong, Pei-Chi Huang, Spyridon Mastorakis, and Frank Y Shih. An automated and robust image watermarking scheme based on deep neural networks. IEEE Transactions on Multimedia, 23:1951-1961, 2020. 3 +[51] Jiren Zhu, Russell Kaplan, Justin Johnson, and Li Fei-Fei. Hidden: Hiding data with deep networks. In Proceedings of the European conference on computer vision (ECCV), pages 657–672, 2018. 3 \ No newline at end of file diff --git a/2024/Gaussian Shading_ Provable Performance-Lossless Image Watermarking for Diffusion Models/images.zip b/2024/Gaussian Shading_ Provable Performance-Lossless Image Watermarking for Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..658f9f448a5bbe1cfb3e4628e25e1e94a2f229f4 --- /dev/null +++ b/2024/Gaussian Shading_ Provable Performance-Lossless Image Watermarking for Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ee824ec3cec9eb1be05368eb934b0af301d38a48cbf1c0a2ce69c4f970b952a +size 557404 diff --git a/2024/Gaussian Shading_ Provable Performance-Lossless Image Watermarking for Diffusion Models/layout.json b/2024/Gaussian Shading_ Provable Performance-Lossless Image Watermarking for Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a1bcec8605e1b8066df53e2fcbe8df6a1ac66ea6 --- /dev/null +++ b/2024/Gaussian Shading_ Provable Performance-Lossless Image Watermarking for Diffusion Models/layout.json @@ -0,0 +1,12423 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 61, + 102, + 533, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 102, + 533, + 138 + ], + "spans": [ + { + "bbox": [ + 61, + 102, + 533, + 138 + ], + "type": "text", + "content": "Gaussian Shading: Provable Performance-Lossless Image Watermarking for Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 160, + 539, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 160, + 539, + 204 + ], + "spans": [ + { + "bbox": [ + 54, + 160, + 539, + 204 + ], + "type": "text", + "content": "Zijin Yang" + }, + { + "bbox": [ + 54, + 160, + 539, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 54, + 160, + 539, + 204 + ], + "type": "text", + "content": ", Kai Zeng" + }, + { + "bbox": [ + 54, + 160, + 539, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 54, + 160, + 539, + 204 + ], + "type": "text", + "content": ", Kejiang Chen" + }, + { + "bbox": [ + 54, + 160, + 539, + 204 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 54, + 160, + 539, + 204 + ], + "type": "text", + "content": ", Han Fang" + }, + { + "bbox": [ + 54, + 160, + 539, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 54, + 160, + 539, + 204 + ], + "type": "text", + "content": ", Weiming Zhang" + }, + { + "bbox": [ + 54, + 160, + 539, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 54, + 160, + 539, + 204 + ], + "type": "text", + "content": ", Nenghai Yu" + }, + { + "bbox": [ + 54, + 160, + 539, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 54, + 160, + 539, + 204 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 160, + 539, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 54, + 160, + 539, + 204 + ], + "type": "text", + "content": "Anhui Province Key Laboratory of Digital Security, University of Science and Technology of China " + }, + { + "bbox": [ + 54, + 160, + 539, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 54, + 160, + 539, + 204 + ], + "type": "text", + "content": "National University of Singapore" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 205, + 527, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 205, + 527, + 217 + ], + "spans": [ + { + "bbox": [ + 62, + 205, + 527, + 217 + ], + "type": "text", + "content": "{bsmhmmlf@mail.,zk0128@mail.,chenkj@,zhangwm@,ynh@}ustc.edu.cn fanghan@nus.edu.sg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "spans": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 270, + 290, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 270, + 290, + 582 + ], + "spans": [ + { + "bbox": [ + 45, + 270, + 290, + 582 + ], + "type": "text", + "content": "Ethical concerns surrounding copyright protection and inappropriate content generation pose challenges for the practical implementation of diffusion models. One effective solution involves watermarking the generated images. However, existing methods often compromise the model performance or require additional training, which is undesirable for operators and users. To address this issue, we propose Gaussian Shading, a diffusion model watermarking technique that is both performance-lossless and training-free, while serving the dual purpose of copyright protection and tracing of offending content. Our watermark embedding is free of model parameter modifications and thus is plug-and-play. We map the watermark to latent representations following a standard Gaussian distribution, which is indistinguishable from latent representations obtained from the non-watermarked diffusion model. Therefore we can achieve watermark embedding with lossless performance, for which we also provide theoretical proof. Furthermore, since the watermark is intricately linked with image semantics, it exhibits resilience to lossy processing and erasure attempts. The watermark can be extracted by Denoising Diffusion Implicit Models (DDIM) inversion and inverse sampling. We evaluate Gaussian Shading on multiple versions of Stable Diffusion, and the results demonstrate that Gaussian Shading not only is performance-lossless but also outperforms existing methods in terms of robustness." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 604, + 128, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 604, + 128, + 616 + ], + "spans": [ + { + "bbox": [ + 47, + 604, + 128, + 616 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 623, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 623, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 623, + 287, + 696 + ], + "type": "text", + "content": "Diffusion models [16, 31-34] signify a noteworthy leap forward in image generation. These well-trained diffusion models, especially commercial diffusion models like Stable Diffusion (SD) [30], Glide [27], and Muse AI [30], enable individuals with diverse backgrounds to create high-quality images effortlessly. However, this raises concerns about" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 320, + 243, + 531, + 437 + ], + "blocks": [ + { + "bbox": [ + 320, + 243, + 531, + 437 + ], + "lines": [ + { + "bbox": [ + 320, + 243, + 531, + 437 + ], + "spans": [ + { + "bbox": [ + 320, + 243, + 531, + 437 + ], + "type": "image", + "image_path": "bcf03ab1c92533925c1aa3c8a58d737b9fb74127d1de4e4f035c46fea24683cd.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 437, + 547, + 491 + ], + "lines": [ + { + "bbox": [ + 304, + 437, + 547, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 547, + 491 + ], + "type": "text", + "content": "Figure 1. Existing watermarking frameworks can be divided into three categories: post-processing-based, fine-tuning-based, and latent-representation-based Tree-Ring. Our method also relies on latent representations but achieves performance-lossless without altering the distribution." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 496, + 546, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 496, + 546, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 496, + 546, + 519 + ], + "type": "text", + "content": "intellectual property and whether diffusion models will be stolen or resold twice." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 520, + 546, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 520, + 546, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 546, + 676 + ], + "type": "text", + "content": "On the other hand, the ease of generating realistic images raises concerns about potentially misleading content generation. For example, on May 23, 2023, a Twitter-verified user named Bloomberg Feed posted a tweet titled \"Large explosion near the Pentagon complex in Washington DC-initial report,\" along with a synthetic image. This tweet led to multiple authoritative media accounts sharing it, even causing a brief impact on the stock market1. On October 30, 2023, White House issued an executive order on AI security, emphasizing the need to protect Americans from AI-enabled fraud and deception by establishing standards and best practices for detecting AI-generated content and authenticating official content2. The urgency of labeling gen" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 683, + 468, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 683, + 468, + 693 + ], + "spans": [ + { + "bbox": [ + 317, + 683, + 468, + 693 + ], + "type": "text", + "content": "1Fake image of Pentagon explosion on Twitter" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 694, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 694, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 694, + 545, + 713 + ], + "type": "text", + "content": "2FACT SHEET: President Biden Issues Executive Order on Safe, Secure, and Trustworthy Artificial Intelligence" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12162" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": "erated content for copyright authentication and prevention of misuse is evident." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 98, + 288, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 98, + 288, + 326 + ], + "spans": [ + { + "bbox": [ + 46, + 98, + 288, + 326 + ], + "type": "text", + "content": "Watermarking is highlighted as a fundamental method for labeling generated content, as it embeds watermark information within the generated image, allowing for subsequent copyright authentication and the tracking of false content. Existing watermarking methods for the diffusion model can be divided into three categories, as shown in Fig. 1. Post-processing-based watermarks [6, 45] adjust robust image features to embed watermarks, thereby directly altering the image and degrading its quality. To mitigate this concern, recent research endeavors propose fine-tuning-based methods [7, 10, 24, 42, 49], which amalgamate the watermark embedding process with the image generation process. Intuitively, these methods need to modify model parameters, introducing supplementary computational overhead. Recently, Wen et al. [41] proposed the latent-representation-based Tree-Ring watermark, which conveys information by adapting the latent representations to match specific patterns. However, it restricts the randomness of sampling, which impacts generative performance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 327, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 287, + 422 + ], + "type": "text", + "content": "Through the above analysis, we can find that these methods compromise model performance to embed watermarks. In practical applications, model performance is paramount for both business interests and user experience. Substantial resource investment is often necessary to pursue enhanced model performance. This leads to a fundamental question: Can watermarks be embedded without compromising model performance?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 425, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 425, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 287, + 544 + ], + "type": "text", + "content": "We affirmatively address the question presented above. Succinctly, the generation process can be delineated into two key phases: latent representation sampling and decoding. Our goal is to align the distribution of the latent representation in watermarked images with that of the latent representation in normally generated images. By keeping the model unaltered, the distribution of watermarked images is naturally consistent with that of normally generated images, enabling the seamless embedding of watermarks without compromising model performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 546, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 546, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 546, + 287, + 713 + ], + "type": "text", + "content": "Building upon this insight, we propose a watermarking method named Gaussian Shading, designed to ensure no deterioration in model performance. The embedding process encompasses three primary elements: watermark diffuse, randomization, and distribution-preserving sampling. Watermark diffusion spreads the watermark information throughout the latent representation. During the generation process, the watermark information will be diffused to the whole semantics of the image, thus achieving excellent robustness. Watermark randomization and distribution-preserving sampling guarantee the congruity of the latent representation distribution with that of watermark-free latent representations, thereby achieving performance-lossless. In the extraction phase, the latent representations" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "are acquired through Denoising Diffusion Implicit Model (DDIM) inversion [32], allowing for the retrieval of watermark information. Harnessing the extensive scope of the SD latent space, we can achieve a high-capacity watermark of 256 bits, surpassing prior methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 133, + 545, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 545, + 240 + ], + "type": "text", + "content": "To the best of our knowledge, ours is the first technique that tackles this challenging problem of performanceless watermarking for diffusion models, and we provide theoretical proof. Moreover, this technique leaves the architecture and parameters of SD unaltered, necessitating no supplementary training. It can seamlessly integrate as a plug-and-play module within the generation process. Model providers can easily replace watermarked models with non-watermarked ones without affecting usability experiences." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 240, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 240, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 240, + 545, + 361 + ], + "type": "text", + "content": "We conducted thorough experiments on SD. Under strong noise perturbation, the average true positive rate and bit accuracy can exceed 0.99 and 0.97, respectively, validating the superiority of Gaussian Shading in both detection and traceability tasks compared to prior methods. Additionally, experiments on visual quality and image-text similarity serve as indicators of performance preservation in our approach. Lastly, we deliberated on various watermark erasure attacks, affirming the steadfast performance of our watermark in the face of such adversities." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 372, + 392, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 372, + 392, + 384 + ], + "spans": [ + { + "bbox": [ + 306, + 372, + 392, + 384 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 392, + 408, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 392, + 408, + 403 + ], + "spans": [ + { + "bbox": [ + 306, + 392, + 408, + 403 + ], + "type": "text", + "content": "2.1. Diffusion Models" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 411, + 545, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 545, + 627 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 545, + 627 + ], + "type": "text", + "content": "Inspired by non-equilibrium thermodynamics, Ho et al. [16] introduced the Denoising Diffusion Probabilistic Model (DDPM). DDPM consists of two Markov chains used for adding and removing noise, and subsequent works [8, 11, 15, 25, 28, 30, 32] have adopted this bidirectional chain framework. To reduce computational complexity and improve efficiency, the Latent Diffusion Model (LDM) [30] was designed, in which the diffusion process occurs in a latent space " + }, + { + "bbox": [ + 304, + 411, + 545, + 627 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 304, + 411, + 545, + 627 + ], + "type": "text", + "content": ". To map an image " + }, + { + "bbox": [ + 304, + 411, + 545, + 627 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 304, + 411, + 545, + 627 + ], + "type": "text", + "content": " to the latent space, the LDM employs an encoder " + }, + { + "bbox": [ + 304, + 411, + 545, + 627 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 304, + 411, + 545, + 627 + ], + "type": "text", + "content": ", such that " + }, + { + "bbox": [ + 304, + 411, + 545, + 627 + ], + "type": "inline_equation", + "content": "z_0 = \\mathcal{E}(x) \\in \\mathbb{R}^{h \\times w \\times c}" + }, + { + "bbox": [ + 304, + 411, + 545, + 627 + ], + "type": "text", + "content": ". Similarly, to reconstruct an image from the latent space, a decoder " + }, + { + "bbox": [ + 304, + 411, + 545, + 627 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 411, + 545, + 627 + ], + "type": "text", + "content": " is used, such that " + }, + { + "bbox": [ + 304, + 411, + 545, + 627 + ], + "type": "inline_equation", + "content": "x = \\mathcal{D}(z_0)" + }, + { + "bbox": [ + 304, + 411, + 545, + 627 + ], + "type": "text", + "content": ". A pretrained LDM can generate images without the encoder " + }, + { + "bbox": [ + 304, + 411, + 545, + 627 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 304, + 411, + 545, + 627 + ], + "type": "text", + "content": ". Specifically, a latent representation " + }, + { + "bbox": [ + 304, + 411, + 545, + 627 + ], + "type": "inline_equation", + "content": "z_T" + }, + { + "bbox": [ + 304, + 411, + 545, + 627 + ], + "type": "text", + "content": " is first sampled from a standard Gaussian distribution " + }, + { + "bbox": [ + 304, + 411, + 545, + 627 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0, I)" + }, + { + "bbox": [ + 304, + 411, + 545, + 627 + ], + "type": "text", + "content": ". Subsequently, through iterative denoising using methods like DDIM [32], " + }, + { + "bbox": [ + 304, + 411, + 545, + 627 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 304, + 411, + 545, + 627 + ], + "type": "text", + "content": " is obtained, and an image can be generated using the decoder: " + }, + { + "bbox": [ + 304, + 411, + 545, + 627 + ], + "type": "inline_equation", + "content": "x = \\mathcal{D}(z_0)" + }, + { + "bbox": [ + 304, + 411, + 545, + 627 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 634, + 430, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 634, + 430, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 634, + 430, + 647 + ], + "type": "text", + "content": "2.2. Image Watermarking" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "type": "text", + "content": "Digital watermarking [38] is an effective means to address copyright protection and content authentication by embedding copyright or traceable identification information within carrier data. Typically, the functionality of a watermark depends on its capacity. For example, a single-bit" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12163" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 68, + 484, + 159 + ], + "blocks": [ + { + "bbox": [ + 108, + 68, + 484, + 159 + ], + "lines": [ + { + "bbox": [ + 108, + 68, + 484, + 159 + ], + "spans": [ + { + "bbox": [ + 108, + 68, + 484, + 159 + ], + "type": "image", + "image_path": "f1602cb88675e865b5ea5a86ef9698a06b6c8a692e398be731b30b89df7b4ea5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 197, + 162, + 394, + 173 + ], + "lines": [ + { + "bbox": [ + 197, + 162, + 394, + 173 + ], + "spans": [ + { + "bbox": [ + 197, + 162, + 394, + 173 + ], + "type": "text", + "content": "Figure 2. Application scenarios for Gaussian Shading." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 177, + 287, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 177, + 287, + 224 + ], + "spans": [ + { + "bbox": [ + 46, + 177, + 287, + 224 + ], + "type": "text", + "content": "watermark can determine whether an image was generated by a particular diffusion model, i.e., copyright protection; a multi-bit watermark can further determine which user of the diffusion model generated the image, i.e., traceability." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 225, + 287, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 225, + 287, + 333 + ], + "spans": [ + { + "bbox": [ + 46, + 225, + 287, + 333 + ], + "type": "text", + "content": "Image watermarking is a method that employs images as carriers for watermarking. Initially, watermark embedding methods primarily focused on the spatial domain [38], but later, to enhance robustness, transform domain watermarking techniques [1, 12, 13, 20, 22, 35, 37] were developed. In recent years, with the advancement of deep learning, researchers have turned their attention to neural networks [21, 39], harnessing their powerful learning capabilities to develop watermarking techniques [18, 19, 26, 36, 43, 50, 51]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 340, + 270, + 353 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 340, + 270, + 353 + ], + "spans": [ + { + "bbox": [ + 47, + 340, + 270, + 353 + ], + "type": "text", + "content": "2.3. Image Watermarking for Diffusion Models" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 358, + 288, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 358, + 288, + 658 + ], + "spans": [ + { + "bbox": [ + 46, + 358, + 288, + 658 + ], + "type": "text", + "content": "Existing Image watermarking methods for the diffusion model [6, 7, 10, 24, 41, 42, 45, 49] can be divided into three categories, as shown in Fig. 1. The image watermarking methods described in the previous section can be applied directly to the images generated by the diffusion model, which is called post-processing-based watermarks [6, 45]. These methods directly modify the image, thus degrading image quality. Recent research endeavors have amalgamated the watermark embedding process with the image generation process to mitigate this issue. Stable Signature [10] fine-tunes the LDM decoder using a pre-trained watermark extractor, facilitating watermark extraction from images produced by the fine-tuned model. Zhao et al. [49] and Liu et al. [24] suggest fine-tuning the diffusion model to implant a backdoor as a watermark, enabling watermark extraction by triggering. These fine-tuning-based approaches enhance the quality of watermarked images but introduce supplementary computational overhead and modify model parameters. Furthermore, Wen et al. [41] introduced the Tree-Ring Watermark, which conveys copyright information by adapting the frequency domain of latent representations to match specific patterns. This method achieves an imperceptible watermark. However, it directly disrupts the Gaussian distribution of noise, limiting the randomness of sampling and resulting in affecting model performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 669, + 107, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 107, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 107, + 681 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "In this section, we provide an overview of the application scenarios and functionalities in Fig. 2. We then proceed" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 177, + 545, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 177, + 545, + 213 + ], + "spans": [ + { + "bbox": [ + 304, + 177, + 545, + 213 + ], + "type": "text", + "content": "to detail the embedding and extraction processes shown in Fig. 3. Finally, we present a mathematical proof of the performance-lossless characteristic of the watermark." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 220, + 430, + 233 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 220, + 430, + 233 + ], + "spans": [ + { + "bbox": [ + 305, + 220, + 430, + 233 + ], + "type": "text", + "content": "3.1. Application Scenarios" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 239, + 545, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 239, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 304, + 239, + 545, + 263 + ], + "type": "text", + "content": "Scenarios. See Fig. 2, the scenario involves the operator Alice, the thief Carol, and two types of users Bob and Trudy." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 264, + 545, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 264, + 545, + 395 + ], + "spans": [ + { + "bbox": [ + 304, + 264, + 545, + 395 + ], + "type": "text", + "content": "Alice is responsible for training the model, deploying it on the platform, and providing the corresponding API for users, but she does not open-source the code or model weights. Carol does not use Alice's services but steals images generated by her model, claiming ownership of the copyrights. Bob and Trudy, as community users, can utilize the API to generate and disseminate images. While Bob faithfully adheres to the community guidelines, Trudy aims to generate deep fake, and infringing content. To evade detection and traceability, Trudy can employ various data augmentation to modify illicit images." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 396, + 545, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 545, + 479 + ], + "type": "text", + "content": "Detection. This scenario satisfies the detection (copyright protection) requirement. Alice embeds a single-bit watermark into each generated image. The successful extraction of the watermark from an image serves as evidence of Alice's rightful ownership of the copyright, while also indicating that the image is artificially generated (as opposed to natural images)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 479, + 545, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 545, + 551 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 545, + 551 + ], + "type": "text", + "content": "Traceability. This scenario fulfills the traceability requirement. Alice allocates a watermark to each user. By extracting the watermark from the illicit content, it enables tracing Trudy, through comparison with the watermark database. Traceability is a higher pursuit than detection and can also achieve copyright protection for different users." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 551, + 545, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 551, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 304, + 551, + 545, + 574 + ], + "type": "text", + "content": "Details of the statistical tests in both scenarios are shown in Supplementary Material." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 582, + 440, + 595 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 582, + 440, + 595 + ], + "spans": [ + { + "bbox": [ + 305, + 582, + 440, + 595 + ], + "type": "text", + "content": "3.2. Watermark Embedding" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 601, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 601, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 601, + 545, + 713 + ], + "type": "text", + "content": "Watermark diffusion. The dimensions of the latent representations are given by " + }, + { + "bbox": [ + 304, + 601, + 545, + 713 + ], + "type": "inline_equation", + "content": "c \\times h \\times w" + }, + { + "bbox": [ + 304, + 601, + 545, + 713 + ], + "type": "text", + "content": ", where each dimension can represent " + }, + { + "bbox": [ + 304, + 601, + 545, + 713 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 601, + 545, + 713 + ], + "type": "text", + "content": " bits of the watermark. Therefore, the watermark capacity becomes " + }, + { + "bbox": [ + 304, + 601, + 545, + 713 + ], + "type": "inline_equation", + "content": "l \\times c \\times h \\times w" + }, + { + "bbox": [ + 304, + 601, + 545, + 713 + ], + "type": "text", + "content": " bits. To enhance the robustness of the watermark, we represent the watermark using " + }, + { + "bbox": [ + 304, + 601, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\frac{1}{f_{hw}}" + }, + { + "bbox": [ + 304, + 601, + 545, + 713 + ], + "type": "text", + "content": " of the height and width, and " + }, + { + "bbox": [ + 304, + 601, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\frac{1}{f_c}" + }, + { + "bbox": [ + 304, + 601, + 545, + 713 + ], + "type": "text", + "content": " of the channel, and replicate the watermark " + }, + { + "bbox": [ + 304, + 601, + 545, + 713 + ], + "type": "inline_equation", + "content": "f_c \\cdot f_{hw}^2" + }, + { + "bbox": [ + 304, + 601, + 545, + 713 + ], + "type": "text", + "content": " times. Thus, the watermark " + }, + { + "bbox": [ + 304, + 601, + 545, + 713 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 601, + 545, + 713 + ], + "type": "text", + "content": " with dimensions " + }, + { + "bbox": [ + 304, + 601, + 545, + 713 + ], + "type": "inline_equation", + "content": "l \\times \\frac{c}{f_c} \\times \\frac{h}{f_{hw}} \\times \\frac{w}{f_{hw}}" + }, + { + "bbox": [ + 304, + 601, + 545, + 713 + ], + "type": "text", + "content": " is expanded into a diffused watermark " + }, + { + "bbox": [ + 304, + 601, + 545, + 713 + ], + "type": "inline_equation", + "content": "s^d" + }, + { + "bbox": [ + 304, + 601, + 545, + 713 + ], + "type": "text", + "content": " with dimensions " + }, + { + "bbox": [ + 304, + 601, + 545, + 713 + ], + "type": "inline_equation", + "content": "l \\times c \\times h \\times w" + }, + { + "bbox": [ + 304, + 601, + 545, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12164" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 96, + 69, + 494, + 245 + ], + "blocks": [ + { + "bbox": [ + 96, + 69, + 494, + 245 + ], + "lines": [ + { + "bbox": [ + 96, + 69, + 494, + 245 + ], + "spans": [ + { + "bbox": [ + 96, + 69, + 494, + 245 + ], + "type": "image", + "image_path": "e61d5fab0bfa788917f1acf870e9289c022719614719f736758469ebe946cc65.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 246, + 547, + 281 + ], + "lines": [ + { + "bbox": [ + 46, + 246, + 547, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 246, + 547, + 281 + ], + "type": "text", + "content": "Figure 3. The framework of Gaussian Shading. We utilize a " + }, + { + "bbox": [ + 46, + 246, + 547, + 281 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 246, + 547, + 281 + ], + "type": "text", + "content": "-bit binary sequence " + }, + { + "bbox": [ + 46, + 246, + 547, + 281 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 246, + 547, + 281 + ], + "type": "text", + "content": " to represent the watermark. After diffusion and encryption, the watermark can be utilized to drive distribution-preserving sampling, followed by denoising to generate watermarked images " + }, + { + "bbox": [ + 46, + 246, + 547, + 281 + ], + "type": "inline_equation", + "content": "X^s" + }, + { + "bbox": [ + 46, + 246, + 547, + 281 + ], + "type": "text", + "content": ". For extraction, it is sufficient to introduce DDIM inversion and the inverse process of all the operations mentioned above." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 282, + 262, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 282, + 262, + 299 + ], + "spans": [ + { + "bbox": [ + 47, + 282, + 262, + 299 + ], + "type": "text", + "content": "The actual watermark capacity is " + }, + { + "bbox": [ + 47, + 282, + 262, + 299 + ], + "type": "inline_equation", + "content": "k = \\frac{l \\times c \\times h \\times w}{f_c \\cdot f_{hw}^2}" + }, + { + "bbox": [ + 47, + 282, + 262, + 299 + ], + "type": "text", + "content": " bits." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 298, + 287, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 298, + 287, + 416 + ], + "spans": [ + { + "bbox": [ + 46, + 298, + 287, + 416 + ], + "type": "text", + "content": "Watermark randomization. If we know the distribution of the diffused watermark " + }, + { + "bbox": [ + 46, + 298, + 287, + 416 + ], + "type": "inline_equation", + "content": "s^d" + }, + { + "bbox": [ + 46, + 298, + 287, + 416 + ], + "type": "text", + "content": ", we can directly utilize distribution-preserving sampling to obtain the corresponding latent representations " + }, + { + "bbox": [ + 46, + 298, + 287, + 416 + ], + "type": "inline_equation", + "content": "z_T^s" + }, + { + "bbox": [ + 46, + 298, + 287, + 416 + ], + "type": "text", + "content": ". However, in practical scenarios, its distribution is always unknown. Hence, we introduce a stream key " + }, + { + "bbox": [ + 46, + 298, + 287, + 416 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 298, + 287, + 416 + ], + "type": "text", + "content": " to transform " + }, + { + "bbox": [ + 46, + 298, + 287, + 416 + ], + "type": "inline_equation", + "content": "s^d" + }, + { + "bbox": [ + 46, + 298, + 287, + 416 + ], + "type": "text", + "content": " into a distribution-unknown randomized watermark " + }, + { + "bbox": [ + 46, + 298, + 287, + 416 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 298, + 287, + 416 + ], + "type": "text", + "content": " through encryption. Considering the use of computationally secure stream cipher, such as ChaCha20 [3], " + }, + { + "bbox": [ + 46, + 298, + 287, + 416 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 298, + 287, + 416 + ], + "type": "text", + "content": " follows a uniform distribution, i.e., " + }, + { + "bbox": [ + 46, + 298, + 287, + 416 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 298, + 287, + 416 + ], + "type": "text", + "content": " is a random binary bit stream." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 417, + 289, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 417, + 289, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 417, + 289, + 550 + ], + "type": "text", + "content": "Distribution-preserving sampling driven by randomized watermark. When each dimension represents " + }, + { + "bbox": [ + 47, + 417, + 289, + 550 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 47, + 417, + 289, + 550 + ], + "type": "text", + "content": "-bit randomized watermark " + }, + { + "bbox": [ + 47, + 417, + 289, + 550 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 47, + 417, + 289, + 550 + ], + "type": "text", + "content": ", this " + }, + { + "bbox": [ + 47, + 417, + 289, + 550 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 47, + 417, + 289, + 550 + ], + "type": "text", + "content": " bits can be regarded as an integer " + }, + { + "bbox": [ + 47, + 417, + 289, + 550 + ], + "type": "inline_equation", + "content": "y \\in [0, 2^l - 1]" + }, + { + "bbox": [ + 47, + 417, + 289, + 550 + ], + "type": "text", + "content": ". Since " + }, + { + "bbox": [ + 47, + 417, + 289, + 550 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 47, + 417, + 289, + 550 + ], + "type": "text", + "content": " is a ciphertext, " + }, + { + "bbox": [ + 47, + 417, + 289, + 550 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 47, + 417, + 289, + 550 + ], + "type": "text", + "content": " follows a discrete uniform distribution, i.e., " + }, + { + "bbox": [ + 47, + 417, + 289, + 550 + ], + "type": "inline_equation", + "content": "p(y) = \\frac{1}{2^l}" + }, + { + "bbox": [ + 47, + 417, + 289, + 550 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 47, + 417, + 289, + 550 + ], + "type": "inline_equation", + "content": "y = 0, 1, 2, \\ldots, 2^l - 1" + }, + { + "bbox": [ + 47, + 417, + 289, + 550 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 47, + 417, + 289, + 550 + ], + "type": "inline_equation", + "content": "f(x)" + }, + { + "bbox": [ + 47, + 417, + 289, + 550 + ], + "type": "text", + "content": " denote the probability density function of the Gaussian distribution " + }, + { + "bbox": [ + 47, + 417, + 289, + 550 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0, I)" + }, + { + "bbox": [ + 47, + 417, + 289, + 550 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 417, + 289, + 550 + ], + "type": "inline_equation", + "content": "ppf" + }, + { + "bbox": [ + 47, + 417, + 289, + 550 + ], + "type": "text", + "content": " denotes the quantile function. We divide " + }, + { + "bbox": [ + 47, + 417, + 289, + 550 + ], + "type": "inline_equation", + "content": "f(x)" + }, + { + "bbox": [ + 47, + 417, + 289, + 550 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 47, + 417, + 289, + 550 + ], + "type": "inline_equation", + "content": "2^l" + }, + { + "bbox": [ + 47, + 417, + 289, + 550 + ], + "type": "text", + "content": " equal cumulative probability portions. When " + }, + { + "bbox": [ + 47, + 417, + 289, + 550 + ], + "type": "inline_equation", + "content": "y = i" + }, + { + "bbox": [ + 47, + 417, + 289, + 550 + ], + "type": "text", + "content": ", the watermarked latent representation " + }, + { + "bbox": [ + 47, + 417, + 289, + 550 + ], + "type": "inline_equation", + "content": "z_T^s" + }, + { + "bbox": [ + 47, + 417, + 289, + 550 + ], + "type": "text", + "content": " falls into the " + }, + { + "bbox": [ + 47, + 417, + 289, + 550 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 417, + 289, + 550 + ], + "type": "text", + "content": "-th interval, which means " + }, + { + "bbox": [ + 47, + 417, + 289, + 550 + ], + "type": "inline_equation", + "content": "z_T^s" + }, + { + "bbox": [ + 47, + 417, + 289, + 550 + ], + "type": "text", + "content": " should follow the conditional distribution:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 556, + 292, + 593 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 556, + 292, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 556, + 292, + 593 + ], + "type": "interline_equation", + "content": "p \\left(z _ {T} ^ {s} \\mid y = i\\right) = \\left\\{ \\begin{array}{c c} 2 ^ {l} \\cdot f \\left(z _ {T} ^ {s}\\right) & p p f \\left(\\frac {i}{2 ^ {l}}\\right) < z _ {T} ^ {s} \\leq p p f \\left(\\frac {i + 1}{2 ^ {l}}\\right) \\\\ 0 & o t h e r w i s e \\end{array} . \\right. \\tag {1}", + "image_path": "51b0baddfb987348754987f578bcb4b9fccd3fbbb157cc200b1bcbe1338727b1.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 594, + 235, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 594, + 235, + 607 + ], + "spans": [ + { + "bbox": [ + 47, + 594, + 235, + 607 + ], + "type": "text", + "content": "The probability distribution of " + }, + { + "bbox": [ + 47, + 594, + 235, + 607 + ], + "type": "inline_equation", + "content": "z_T^s" + }, + { + "bbox": [ + 47, + 594, + 235, + 607 + ], + "type": "text", + "content": " is given by:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 75, + 613, + 287, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 613, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 75, + 613, + 287, + 647 + ], + "type": "interline_equation", + "content": "p \\left(z _ {T} ^ {s}\\right) = \\sum_ {i = 0} ^ {2 ^ {l} - 1} p \\left(z _ {T} ^ {s} \\mid y = i\\right) p (y = i) = f \\left(z _ {T} ^ {s}\\right). \\tag {2}", + "image_path": "8b3042b1e0ad67deb7577200f3a5d7eb3d86697b54ff44a914b4895f7f206c95.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "type": "text", + "content": "Eq. (2) indicates that " + }, + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "type": "inline_equation", + "content": "z_T^s" + }, + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "type": "text", + "content": " follows the same distribution as the randomly sampled latent representation " + }, + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "type": "inline_equation", + "content": "z_T \\sim \\mathcal{N}(0, I)" + }, + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "type": "text", + "content": ". Next, we elaborate on how this sampling is implemented." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "Let the cumulative distribution function of " + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "f(x)" + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": " be denoted as " + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "cdf" + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": ". We can obtain the cumulative distribution" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 283, + 429, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 283, + 429, + 296 + ], + "spans": [ + { + "bbox": [ + 305, + 283, + 429, + 296 + ], + "type": "text", + "content": "function of Eq. (1) as follows," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 309, + 297, + 545, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 297, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 309, + 297, + 545, + 361 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} F (z _ {T} ^ {s} | y = i) \\\\ = \\left\\{ \\begin{array}{c c} 0 & z _ {T} ^ {s} < p p f \\left(\\frac {i}{2 ^ {l}}\\right) \\\\ 2 ^ {l} \\cdot c d f \\left(z _ {T} ^ {s}\\right) - i & p p f \\left(\\frac {i}{2 ^ {l}}\\right) \\leq z _ {T} ^ {s} \\leq p p f \\left(\\frac {i + 1}{2 ^ {l}}\\right) \\\\ 1 & z _ {T} ^ {s} > p p f \\left(\\frac {i + 1}{2 ^ {l}}\\right) \\end{array} . \\right. \\tag {3} \\\\ \\end{array}", + "image_path": "0c1ce46a937d07ce0d686b70bfcab61ddfb4635a687e43a99d588f7849149b4d.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 362, + 546, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 362, + 546, + 505 + ], + "spans": [ + { + "bbox": [ + 305, + 362, + 546, + 505 + ], + "type": "text", + "content": "Given " + }, + { + "bbox": [ + 305, + 362, + 546, + 505 + ], + "type": "inline_equation", + "content": "y = i" + }, + { + "bbox": [ + 305, + 362, + 546, + 505 + ], + "type": "text", + "content": ", we aim to perform random sampling of " + }, + { + "bbox": [ + 305, + 362, + 546, + 505 + ], + "type": "inline_equation", + "content": "z_{T}^{s}" + }, + { + "bbox": [ + 305, + 362, + 546, + 505 + ], + "type": "text", + "content": " within the interval " + }, + { + "bbox": [ + 305, + 362, + 546, + 505 + ], + "type": "inline_equation", + "content": "[ppf(\\frac{i}{2^t}), ppf(\\frac{i + 1}{2^t})]" + }, + { + "bbox": [ + 305, + 362, + 546, + 505 + ], + "type": "text", + "content": ". The commonly used method is rejection sampling [4, 17, 47], which can be time-consuming as it requires repeated sampling until " + }, + { + "bbox": [ + 305, + 362, + 546, + 505 + ], + "type": "inline_equation", + "content": "z_{T}^{s}" + }, + { + "bbox": [ + 305, + 362, + 546, + 505 + ], + "type": "text", + "content": " falls into the correct interval. Instead, we can utilize the cumulative probability density. When randomly sampling " + }, + { + "bbox": [ + 305, + 362, + 546, + 505 + ], + "type": "inline_equation", + "content": "F(z_{T}^{s}|y = i)" + }, + { + "bbox": [ + 305, + 362, + 546, + 505 + ], + "type": "text", + "content": ", the corresponding " + }, + { + "bbox": [ + 305, + 362, + 546, + 505 + ], + "type": "inline_equation", + "content": "z_{T}^{s}" + }, + { + "bbox": [ + 305, + 362, + 546, + 505 + ], + "type": "text", + "content": " is naturally obtained through random sampling. Since " + }, + { + "bbox": [ + 305, + 362, + 546, + 505 + ], + "type": "inline_equation", + "content": "F(z_{T}^{s}|y = i)" + }, + { + "bbox": [ + 305, + 362, + 546, + 505 + ], + "type": "text", + "content": " takes values in " + }, + { + "bbox": [ + 305, + 362, + 546, + 505 + ], + "type": "inline_equation", + "content": "[0,1]" + }, + { + "bbox": [ + 305, + 362, + 546, + 505 + ], + "type": "text", + "content": ", sampling from it is equivalent to sampling from a standard uniform distribution, denoted as " + }, + { + "bbox": [ + 305, + 362, + 546, + 505 + ], + "type": "inline_equation", + "content": "u = F(z_{T}^{s}|y = i) \\sim \\mathcal{U}(0,1)" + }, + { + "bbox": [ + 305, + 362, + 546, + 505 + ], + "type": "text", + "content": ". Shift the terms of Eq. (3), and take into account that " + }, + { + "bbox": [ + 305, + 362, + 546, + 505 + ], + "type": "inline_equation", + "content": "cdf" + }, + { + "bbox": [ + 305, + 362, + 546, + 505 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 362, + 546, + 505 + ], + "type": "inline_equation", + "content": "ppf" + }, + { + "bbox": [ + 305, + 362, + 546, + 505 + ], + "type": "text", + "content": " are inverse functions, we have" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 387, + 507, + 545, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 507, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 387, + 507, + 545, + 529 + ], + "type": "interline_equation", + "content": "z _ {T} ^ {s} = p p f \\left(\\frac {u + i}{2 ^ {l}}\\right). \\tag {4}", + "image_path": "d3e735d4d0d9b8c0982f9045c1aacb84d59d29c7213edf3eefda7ece05e2d2af.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 530, + 545, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 530, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 305, + 530, + 545, + 563 + ], + "type": "text", + "content": "Eq. (4) represents the process of sampling the watermarked latent representation " + }, + { + "bbox": [ + 305, + 530, + 545, + 563 + ], + "type": "inline_equation", + "content": "z_{T}^{s}" + }, + { + "bbox": [ + 305, + 530, + 545, + 563 + ], + "type": "text", + "content": " driven by the randomized watermark " + }, + { + "bbox": [ + 305, + 530, + 545, + 563 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 305, + 530, + 545, + 563 + ], + "type": "text", + "content": ". To extract the watermark, its inverse map is" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 386, + 565, + 545, + 580 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 565, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 386, + 565, + 545, + 580 + ], + "type": "interline_equation", + "content": "i = \\left\\lfloor 2 ^ {l} \\cdot c d f \\left(z _ {T} ^ {s}\\right) \\right\\rfloor . \\tag {5}", + "image_path": "057ad42446172f6976603130ee25be18b38fd3740a40bcea00a0c34bc0f3e8cf.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 594, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 715 + ], + "type": "text", + "content": "Image generation. After the sampling process, the watermark is embedded in the latent representation " + }, + { + "bbox": [ + 304, + 594, + 546, + 715 + ], + "type": "inline_equation", + "content": "z_T^s" + }, + { + "bbox": [ + 304, + 594, + 546, + 715 + ], + "type": "text", + "content": ", and the subsequent generation process is no different from the regular generation process of SD. Here, we employ the DPM-Solver [25] for iterative denoising of " + }, + { + "bbox": [ + 304, + 594, + 546, + 715 + ], + "type": "inline_equation", + "content": "z_T^s" + }, + { + "bbox": [ + 304, + 594, + 546, + 715 + ], + "type": "text", + "content": ". In addition to DPMSolver [25], other continuous-time samplers based on ordinary differential equation (ODE) solvers [32], such as DDIM [32], DEIS [46], PNDM [23], and UniPC [48], can be used too. After obtaining denoised " + }, + { + "bbox": [ + 304, + 594, + 546, + 715 + ], + "type": "inline_equation", + "content": "z_0^s" + }, + { + "bbox": [ + 304, + 594, + 546, + 715 + ], + "type": "text", + "content": ", the watermarked image " + }, + { + "bbox": [ + 304, + 594, + 546, + 715 + ], + "type": "inline_equation", + "content": "X^s" + }, + { + "bbox": [ + 304, + 594, + 546, + 715 + ], + "type": "text", + "content": " is generated using the decoder " + }, + { + "bbox": [ + 304, + 594, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 594, + 546, + 715 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 304, + 594, + 546, + 715 + ], + "type": "inline_equation", + "content": "X^s = \\mathcal{D}(z_0^s)" + }, + { + "bbox": [ + 304, + 594, + 546, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "12165" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 177, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 177, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 177, + 83 + ], + "type": "text", + "content": "3.3. Watermark Extraction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 89, + 286, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 89, + 286, + 162 + ], + "spans": [ + { + "bbox": [ + 46, + 89, + 286, + 162 + ], + "type": "text", + "content": "DDIM Inversion. Using the SD encoder " + }, + { + "bbox": [ + 46, + 89, + 286, + 162 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 46, + 89, + 286, + 162 + ], + "type": "text", + "content": ", we first restore " + }, + { + "bbox": [ + 46, + 89, + 286, + 162 + ], + "type": "inline_equation", + "content": "X^{\\prime s}" + }, + { + "bbox": [ + 46, + 89, + 286, + 162 + ], + "type": "text", + "content": " to the latent space " + }, + { + "bbox": [ + 46, + 89, + 286, + 162 + ], + "type": "inline_equation", + "content": "z_0^{\\prime s} = \\mathcal{E}(X^{\\prime s})" + }, + { + "bbox": [ + 46, + 89, + 286, + 162 + ], + "type": "text", + "content": ". Then, we introduce the DDIM inversion [32] to estimate the additive noise. It can be considered that " + }, + { + "bbox": [ + 46, + 89, + 286, + 162 + ], + "type": "inline_equation", + "content": "z_T^{\\prime s} \\approx z_T^s" + }, + { + "bbox": [ + 46, + 89, + 286, + 162 + ], + "type": "text", + "content": ". We also observe that although DDIM inversion is derived from DDIM, it can apply to other continuous-time samplers based on ODE solvers." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 163, + 286, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 163, + 286, + 269 + ], + "spans": [ + { + "bbox": [ + 46, + 163, + 286, + 269 + ], + "type": "text", + "content": "Watermark reduction from latent representations. After obtaining " + }, + { + "bbox": [ + 46, + 163, + 286, + 269 + ], + "type": "inline_equation", + "content": "z_T^{\\prime s}" + }, + { + "bbox": [ + 46, + 163, + 286, + 269 + ], + "type": "text", + "content": ", according to the inverse transformation defined in Eq. (5), the tensor can be converted into a bit stream " + }, + { + "bbox": [ + 46, + 163, + 286, + 269 + ], + "type": "inline_equation", + "content": "m'" + }, + { + "bbox": [ + 46, + 163, + 286, + 269 + ], + "type": "text", + "content": ". Subsequently, " + }, + { + "bbox": [ + 46, + 163, + 286, + 269 + ], + "type": "inline_equation", + "content": "m'" + }, + { + "bbox": [ + 46, + 163, + 286, + 269 + ], + "type": "text", + "content": " is decrypted using " + }, + { + "bbox": [ + 46, + 163, + 286, + 269 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 163, + 286, + 269 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 46, + 163, + 286, + 269 + ], + "type": "inline_equation", + "content": "s'^d" + }, + { + "bbox": [ + 46, + 163, + 286, + 269 + ], + "type": "text", + "content": ". Inverse diffusion of the watermark results in " + }, + { + "bbox": [ + 46, + 163, + 286, + 269 + ], + "type": "inline_equation", + "content": "f_c \\cdot f_{hw}^2" + }, + { + "bbox": [ + 46, + 163, + 286, + 269 + ], + "type": "text", + "content": " copies of the watermark. Similar to voting, if the bit is set to 1 in more than half of the copies, the corresponding watermark bit is set to 1; otherwise, it is set to 0. This process restores the true binary watermark sequence " + }, + { + "bbox": [ + 46, + 163, + 286, + 269 + ], + "type": "inline_equation", + "content": "s'" + }, + { + "bbox": [ + 46, + 163, + 286, + 269 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 275, + 211, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 275, + 211, + 286 + ], + "spans": [ + { + "bbox": [ + 47, + 275, + 211, + 286 + ], + "type": "text", + "content": "3.4. Proof of Lossless Performance" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 293, + 286, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 286, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 286, + 510 + ], + "type": "text", + "content": "In prior works, the incorporation of watermark embedding modules inevitably results in a decline in model performance, as typically evaluated using metrics such as Peak Signal-to-Noise Ratio (PSNR) and Fréchet Inception Distance (FID) [14], which are more suitable for assessing post-processing methods. To assess methods that integrate the watermark embedding and generation processes, we propose a definition for the impact of watermark embedding on model performance, drawing on the complexity-theoretic definition of steganographic security [17]. This definition is based on a probabilistic game between a watermarked image " + }, + { + "bbox": [ + 46, + 293, + 286, + 510 + ], + "type": "inline_equation", + "content": "X^{s}" + }, + { + "bbox": [ + 46, + 293, + 286, + 510 + ], + "type": "text", + "content": " and a normally generated image " + }, + { + "bbox": [ + 46, + 293, + 286, + 510 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 293, + 286, + 510 + ], + "type": "text", + "content": ". The tester " + }, + { + "bbox": [ + 46, + 293, + 286, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 46, + 293, + 286, + 510 + ], + "type": "text", + "content": " can use any watermark to drive the sampling process and generate " + }, + { + "bbox": [ + 46, + 293, + 286, + 510 + ], + "type": "inline_equation", + "content": "X^{s}" + }, + { + "bbox": [ + 46, + 293, + 286, + 510 + ], + "type": "text", + "content": ", similar to the chosen hidden text attacks [17], which we refer to as chosen watermark tests. The watermarking method is performance-lossless under chosen watermark tests, if for any polynomial-time tester " + }, + { + "bbox": [ + 46, + 293, + 286, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 46, + 293, + 286, + 510 + ], + "type": "text", + "content": " and key " + }, + { + "bbox": [ + 46, + 293, + 286, + 510 + ], + "type": "inline_equation", + "content": "K \\gets \\text{KeyGen}_{\\mathcal{G}(1^{\\rho})}" + }, + { + "bbox": [ + 46, + 293, + 286, + 510 + ], + "type": "text", + "content": ", it holds that" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 511, + 286, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 511, + 286, + 525 + ], + "spans": [ + { + "bbox": [ + 61, + 511, + 286, + 525 + ], + "type": "interline_equation", + "content": "\\left| \\Pr \\left[ \\mathcal {A} \\left(X ^ {s}\\right) = 1 \\right] - \\Pr \\left[ \\mathcal {A} (X) = 1 \\right] \\right| < \\operatorname {n e g l} (\\rho). \\tag {6}", + "image_path": "99e13487184837231db6cb7dd601500be3c38627334a6331b3969faf882e6329.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 529, + 286, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 529, + 286, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 529, + 286, + 550 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 47, + 529, + 286, + 550 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 47, + 529, + 286, + 550 + ], + "type": "text", + "content": " represents the length of the security parameter, such as the key " + }, + { + "bbox": [ + 47, + 529, + 286, + 550 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 529, + 286, + 550 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 529, + 286, + 550 + ], + "type": "inline_equation", + "content": "\\mathsf{negl}(\\rho)" + }, + { + "bbox": [ + 47, + 529, + 286, + 550 + ], + "type": "text", + "content": " is a negligible term relative to " + }, + { + "bbox": [ + 47, + 529, + 286, + 550 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 47, + 529, + 286, + 550 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 551, + 286, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 551, + 286, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 551, + 286, + 586 + ], + "type": "text", + "content": "We prove the statement using a proof by contradiction. First, assume that the watermarked image " + }, + { + "bbox": [ + 47, + 551, + 286, + 586 + ], + "type": "inline_equation", + "content": "X^s" + }, + { + "bbox": [ + 47, + 551, + 286, + 586 + ], + "type": "text", + "content": " and the normally generated image " + }, + { + "bbox": [ + 47, + 551, + 286, + 586 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 47, + 551, + 286, + 586 + ], + "type": "text", + "content": " are distinguishable, meaning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 82, + 588, + 286, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 588, + 286, + 601 + ], + "spans": [ + { + "bbox": [ + 82, + 588, + 286, + 601 + ], + "type": "interline_equation", + "content": "\\left. \\right.\\left| \\Pr \\left[ \\mathcal {A} \\left(X ^ {s}\\right) = 1 \\right] - \\Pr \\left[ \\mathcal {A} (X) = 1 \\right]\\right| = \\delta , \\tag {7}", + "image_path": "08029c68df2b75e7c1b180a0885cf1ad558b66d3883028b83d77a5362371be9f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 607, + 286, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 607, + 286, + 640 + ], + "spans": [ + { + "bbox": [ + 47, + 607, + 286, + 640 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 607, + 286, + 640 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 47, + 607, + 286, + 640 + ], + "type": "text", + "content": " is non-negligible with respect to the key " + }, + { + "bbox": [ + 47, + 607, + 286, + 640 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 607, + 286, + 640 + ], + "type": "text", + "content": ". Let the iterative denoising process be denoted as " + }, + { + "bbox": [ + 47, + 607, + 286, + 640 + ], + "type": "inline_equation", + "content": "Q(\\cdot)" + }, + { + "bbox": [ + 47, + 607, + 286, + 640 + ], + "type": "text", + "content": ", and substitute the LDM encoder " + }, + { + "bbox": [ + 47, + 607, + 286, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 47, + 607, + 286, + 640 + ], + "type": "text", + "content": " into Eq. (7), we have" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 58, + 643, + 265, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 643, + 265, + 673 + ], + "spans": [ + { + "bbox": [ + 58, + 643, + 265, + 673 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\left| \\Pr \\left[ \\mathcal {A} \\left(\\mathcal {E} \\left(Q \\left(z _ {T} ^ {s}\\right)\\right)\\right) = 1 \\mid m = E (K, s ^ {d}) \\right] \\right. \\\\ - \\Pr \\left[ \\mathcal {A} \\left(\\mathcal {E} \\left(Q \\left(z _ {T}\\right)\\right)\\right) = 1 \\mid z _ {T} \\leftarrow \\mathcal {N} (0, I) \\right] | = \\delta , \\\\ \\end{array}", + "image_path": "53145f88f6ebd54b0282156f2417fa9cfda688fa7da6637b44ad659a41cc9cd1.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "text", + "content": "where the randomized watermark " + }, + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "text", + "content": " is obtained by encrypting the diffused watermark " + }, + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "inline_equation", + "content": "s^d" + }, + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "text", + "content": " using the encryption algorithm " + }, + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "text", + "content": " with key " + }, + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "text", + "content": ". Note that Eq. (2) contains the fact that" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "distribution-preserving sampling driven by randomized watermark and random sampling are equivalent. Therefore, we denote sequence-driven sampling as " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "S(\\cdot)" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "z_{T}^{s}" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " can naturally be obtained by sampling driven by " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "z_{T}^{s} = S(m)" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": ". On the other hand, " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "z_{T}" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " can be considered as obtained by sampling driven by a truly random sequence " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " of the same length as " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "z_{T} = S(r)" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": ". Eq. (8) can be written as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 318, + 159, + 545, + 189 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 159, + 545, + 189 + ], + "spans": [ + { + "bbox": [ + 318, + 159, + 545, + 189 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\left| \\Pr [ \\mathcal {A} (\\mathcal {E} (Q (S (m)))) = 1 | m = E (K, s ^ {d}) ] \\right. \\tag {9} \\\\ - \\Pr \\left[ \\mathcal {A} \\left(\\mathcal {E} (Q (S (r))) = 1 \\right\\rvert \\right] = \\delta . \\\\ \\end{array}", + "image_path": "463fa752f885547d5f8f6f165492917b07536e32605147f86d8a65bf710c071e.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 196, + 545, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 196, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 304, + 196, + 545, + 229 + ], + "type": "text", + "content": "Sampling " + }, + { + "bbox": [ + 304, + 196, + 545, + 229 + ], + "type": "inline_equation", + "content": "S(\\cdot)" + }, + { + "bbox": [ + 304, + 196, + 545, + 229 + ], + "type": "text", + "content": ", denoising " + }, + { + "bbox": [ + 304, + 196, + 545, + 229 + ], + "type": "inline_equation", + "content": "Q(\\cdot)" + }, + { + "bbox": [ + 304, + 196, + 545, + 229 + ], + "type": "text", + "content": ", and encoder " + }, + { + "bbox": [ + 304, + 196, + 545, + 229 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 304, + 196, + 545, + 229 + ], + "type": "text", + "content": " can be considered as subroutines that the tester " + }, + { + "bbox": [ + 304, + 196, + 545, + 229 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\mathcal{E}, Q, S}" + }, + { + "bbox": [ + 304, + 196, + 545, + 229 + ], + "type": "text", + "content": " can use. Thus, Eq. (9) can be simplified," + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 232, + 545, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 232, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 320, + 232, + 545, + 262 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\left| \\Pr \\left[ \\mathcal {A} _ {\\varepsilon , Q, S} (m) = 1 \\mid m = E (K, s ^ {d}) \\right] \\right. \\tag {10} \\\\ - \\Pr \\left[ \\mathcal {A} _ {\\mathcal {E}, Q, S} (r) = 1 \\right] | = \\delta . \\\\ \\end{array}", + "image_path": "ca80573a8e171971fb53b58479c115e63f2b177f7d726738cc0de80ed579ff20.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 270, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 270, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 270, + 545, + 449 + ], + "type": "text", + "content": "Note that " + }, + { + "bbox": [ + 304, + 270, + 545, + 449 + ], + "type": "inline_equation", + "content": "S(\\cdot), Q(\\cdot)" + }, + { + "bbox": [ + 304, + 270, + 545, + 449 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 270, + 545, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 304, + 270, + 545, + 449 + ], + "type": "text", + "content": " are all polynomial-time programs, so the time taken by the tester " + }, + { + "bbox": [ + 304, + 270, + 545, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\mathcal{E},Q,S}" + }, + { + "bbox": [ + 304, + 270, + 545, + 449 + ], + "type": "text", + "content": " to make the distinction is also polynomial. Eq. (10) essentially states that it is possible to distinguish between " + }, + { + "bbox": [ + 304, + 270, + 545, + 449 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 270, + 545, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 270, + 545, + 449 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 270, + 545, + 449 + ], + "type": "text", + "content": " in polynomial time. However, we have used the computationally secure stream cipher ChaCha20 [3] in watermark randomization, which means that " + }, + { + "bbox": [ + 304, + 270, + 545, + 449 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 270, + 545, + 449 + ], + "type": "text", + "content": " as a pseudorandom sequence cannot be distinguished from a truly random sequence in polynomial time. Eq. (10) contradicts the computational security property of ChaCha20 [3]. Therefore, Eq. (10) is not valid, leading us back to our initial assumption that Eq. (7) is also not valid. This implies that the watermarked image " + }, + { + "bbox": [ + 304, + 270, + 545, + 449 + ], + "type": "inline_equation", + "content": "X^s" + }, + { + "bbox": [ + 304, + 270, + 545, + 449 + ], + "type": "text", + "content": " and the normally generated image " + }, + { + "bbox": [ + 304, + 270, + 545, + 449 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 270, + 545, + 449 + ], + "type": "text", + "content": " are indistinguishable in polynomial time. Hence, Gaussian Shading is performance-lossless under chosen watermark tests." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 460, + 386, + 473 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 460, + 386, + 473 + ], + "spans": [ + { + "bbox": [ + 306, + 460, + 386, + 473 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 479, + 545, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 545, + 528 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 545, + 528 + ], + "type": "text", + "content": "This section focuses on experimental analysis, including details of the experimental setup, performance evaluation of Gaussian Shading, comparison with baseline methods, ablation experiments, and potential attacks." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 534, + 438, + 546 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 534, + 438, + 546 + ], + "spans": [ + { + "bbox": [ + 306, + 534, + 438, + 546 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 552, + 545, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 552, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 552, + 545, + 696 + ], + "type": "text", + "content": "SD models. In this paper, we focus on text-to-image LDM, hence we select SD [30] provided by huggingface. We evaluate Gaussian Shading as well as baseline methods, using three versions of SD: V1.4, V2.0, and V2.1. The size of the generated images is " + }, + { + "bbox": [ + 304, + 552, + 545, + 696 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 552, + 545, + 696 + ], + "type": "text", + "content": ", and the latent space dimension is " + }, + { + "bbox": [ + 304, + 552, + 545, + 696 + ], + "type": "inline_equation", + "content": "4 \\times 64 \\times 64" + }, + { + "bbox": [ + 304, + 552, + 545, + 696 + ], + "type": "text", + "content": ". During inference, we employ the prompt from Stable-Diffusion-Prompt3, with a guidance scale of 7.5. We sample 50 steps using DPMSolver [25]. Considering that users tend to propagate the generated images without retaining the corresponding prompts, we use an empty prompt for inversion, with a scale of 1. We perform 50 steps of inversion using DDIM inversion [32]." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 702, + 405, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 702, + 405, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 702, + 405, + 712 + ], + "type": "text", + "content": "3Stable-Diffusion-Prompts" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12166" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 69, + 97, + 114 + ], + "blocks": [ + { + "bbox": [ + 52, + 69, + 97, + 114 + ], + "lines": [ + { + "bbox": [ + 52, + 69, + 97, + 114 + ], + "spans": [ + { + "bbox": [ + 52, + 69, + 97, + 114 + ], + "type": "image", + "image_path": "113f92213137cf7560fc650a7bc87aa7fa6167595fbc53eb9e6c952236ba331c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 100, + 70, + 143, + 114 + ], + "blocks": [ + { + "bbox": [ + 100, + 70, + 143, + 114 + ], + "lines": [ + { + "bbox": [ + 100, + 70, + 143, + 114 + ], + "spans": [ + { + "bbox": [ + 100, + 70, + 143, + 114 + ], + "type": "image", + "image_path": "391c5ea77b52a7c6e9ad6089279876a9caf13725122d62529f2361a9e94e4313.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 144, + 70, + 190, + 114 + ], + "blocks": [ + { + "bbox": [ + 144, + 70, + 190, + 114 + ], + "lines": [ + { + "bbox": [ + 144, + 70, + 190, + 114 + ], + "spans": [ + { + "bbox": [ + 144, + 70, + 190, + 114 + ], + "type": "image", + "image_path": "2bc54b6f2c460407df5ec3c81d88e2f7a9f96762a6c17332b8cce2eca805f13c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 191, + 70, + 236, + 114 + ], + "blocks": [ + { + "bbox": [ + 191, + 70, + 236, + 114 + ], + "lines": [ + { + "bbox": [ + 191, + 70, + 236, + 114 + ], + "spans": [ + { + "bbox": [ + 191, + 70, + 236, + 114 + ], + "type": "image", + "image_path": "a00616e79178046b18344bdef487b7c122716a955f897da335aa553de66f550e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 238, + 70, + 283, + 114 + ], + "blocks": [ + { + "bbox": [ + 238, + 70, + 283, + 114 + ], + "lines": [ + { + "bbox": [ + 238, + 70, + 283, + 114 + ], + "spans": [ + { + "bbox": [ + 238, + 70, + 283, + 114 + ], + "type": "image", + "image_path": "7a1e6ed37b87a66ec8e8ddf1676f6702fc25431a7fb153cb16423fa84918c87a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 50, + 126, + 96, + 171 + ], + "blocks": [ + { + "bbox": [ + 50, + 126, + 96, + 171 + ], + "lines": [ + { + "bbox": [ + 50, + 126, + 96, + 171 + ], + "spans": [ + { + "bbox": [ + 50, + 126, + 96, + 171 + ], + "type": "image", + "image_path": "78308b2122f914f4fac6cf0abe8e2988f0188e26a62c9f26c1470943e865b0d4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 173, + 77, + 182 + ], + "lines": [ + { + "bbox": [ + 67, + 173, + 77, + 182 + ], + "spans": [ + { + "bbox": [ + 67, + 173, + 77, + 182 + ], + "type": "text", + "content": "(f)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 96, + 126, + 141, + 171 + ], + "blocks": [ + { + "bbox": [ + 115, + 117, + 126, + 125 + ], + "lines": [ + { + "bbox": [ + 115, + 117, + 126, + 125 + ], + "spans": [ + { + "bbox": [ + 115, + 117, + 126, + 125 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 96, + 126, + 141, + 171 + ], + "lines": [ + { + "bbox": [ + 96, + 126, + 141, + 171 + ], + "spans": [ + { + "bbox": [ + 96, + 126, + 141, + 171 + ], + "type": "image", + "image_path": "3c04e4d0e837a921c04aabbbd4ebaf3cb8aa7899b3418d1658e7736585bc880e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 113, + 173, + 124, + 182 + ], + "lines": [ + { + "bbox": [ + 113, + 173, + 124, + 182 + ], + "spans": [ + { + "bbox": [ + 113, + 173, + 124, + 182 + ], + "type": "text", + "content": "(g)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 143, + 126, + 188, + 171 + ], + "blocks": [ + { + "bbox": [ + 143, + 126, + 188, + 171 + ], + "lines": [ + { + "bbox": [ + 143, + 126, + 188, + 171 + ], + "spans": [ + { + "bbox": [ + 143, + 126, + 188, + 171 + ], + "type": "image", + "image_path": "f4b394ed8fb9227ff91deaab30c954771c7a2118d94f43320d95ddb0bf42850b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 159, + 173, + 170, + 182 + ], + "lines": [ + { + "bbox": [ + 159, + 173, + 170, + 182 + ], + "spans": [ + { + "bbox": [ + 159, + 173, + 170, + 182 + ], + "type": "text", + "content": "(h)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 190, + 126, + 235, + 171 + ], + "blocks": [ + { + "bbox": [ + 207, + 117, + 218, + 125 + ], + "lines": [ + { + "bbox": [ + 207, + 117, + 218, + 125 + ], + "spans": [ + { + "bbox": [ + 207, + 117, + 218, + 125 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 190, + 126, + 235, + 171 + ], + "lines": [ + { + "bbox": [ + 190, + 126, + 235, + 171 + ], + "spans": [ + { + "bbox": [ + 190, + 126, + 235, + 171 + ], + "type": "image", + "image_path": "44eee57d50c3eeced27c41c772a82e00c47783d7cea6ed9483617ca300b13e12.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 173, + 215, + 182 + ], + "lines": [ + { + "bbox": [ + 206, + 173, + 215, + 182 + ], + "spans": [ + { + "bbox": [ + 206, + 173, + 215, + 182 + ], + "type": "text", + "content": "(i)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 183, + 287, + 260 + ], + "lines": [ + { + "bbox": [ + 46, + 183, + 287, + 260 + ], + "spans": [ + { + "bbox": [ + 46, + 183, + 287, + 260 + ], + "type": "text", + "content": "Figure 4. Watermarked image is attacked by different noise. (a) Watermarked image. (b) JPEG, " + }, + { + "bbox": [ + 46, + 183, + 287, + 260 + ], + "type": "inline_equation", + "content": "QF = 25" + }, + { + "bbox": [ + 46, + 183, + 287, + 260 + ], + "type": "text", + "content": ". (c) " + }, + { + "bbox": [ + 46, + 183, + 287, + 260 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 46, + 183, + 287, + 260 + ], + "type": "text", + "content": " area Random Crop (RandCr). (d) " + }, + { + "bbox": [ + 46, + 183, + 287, + 260 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 46, + 183, + 287, + 260 + ], + "type": "text", + "content": " area Random Drop (RandDr). (e) Gaussian Blur, " + }, + { + "bbox": [ + 46, + 183, + 287, + 260 + ], + "type": "inline_equation", + "content": "r = 4" + }, + { + "bbox": [ + 46, + 183, + 287, + 260 + ], + "type": "text", + "content": " (GauBlur). (f) Median Filter, " + }, + { + "bbox": [ + 46, + 183, + 287, + 260 + ], + "type": "inline_equation", + "content": "k = 7" + }, + { + "bbox": [ + 46, + 183, + 287, + 260 + ], + "type": "text", + "content": " (MedFilter). (g) Gaussian Noise, " + }, + { + "bbox": [ + 46, + 183, + 287, + 260 + ], + "type": "inline_equation", + "content": "\\mu = 0" + }, + { + "bbox": [ + 46, + 183, + 287, + 260 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 183, + 287, + 260 + ], + "type": "inline_equation", + "content": "\\sigma = 0.05" + }, + { + "bbox": [ + 46, + 183, + 287, + 260 + ], + "type": "text", + "content": " (GauNoise). (h) Salt and Pepper Noise, " + }, + { + "bbox": [ + 46, + 183, + 287, + 260 + ], + "type": "inline_equation", + "content": "p = 0.05" + }, + { + "bbox": [ + 46, + 183, + 287, + 260 + ], + "type": "text", + "content": " (S&PNoise). (i) " + }, + { + "bbox": [ + 46, + 183, + 287, + 260 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 46, + 183, + 287, + 260 + ], + "type": "text", + "content": " Resize and restore (Resize). (j) Brightness, factor " + }, + { + "bbox": [ + 46, + 183, + 287, + 260 + ], + "type": "inline_equation", + "content": "= 6" + }, + { + "bbox": [ + 46, + 183, + 287, + 260 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 244, + 126, + 280, + 171 + ], + "blocks": [ + { + "bbox": [ + 244, + 126, + 280, + 171 + ], + "lines": [ + { + "bbox": [ + 244, + 126, + 280, + 171 + ], + "spans": [ + { + "bbox": [ + 244, + 126, + 280, + 171 + ], + "type": "image", + "image_path": "e724f5fe89325981724032fd706b40f092a41863e3b1de3c6d07a334bb23ccca.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 252, + 173, + 261, + 182 + ], + "lines": [ + { + "bbox": [ + 252, + 173, + 261, + 182 + ], + "spans": [ + { + "bbox": [ + 252, + 173, + 261, + 182 + ], + "type": "text", + "content": "(j)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 262, + 287, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 262, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 287, + 346 + ], + "type": "text", + "content": "Watermarking methods. In the main experiments, the settings for Gaussian Shading are " + }, + { + "bbox": [ + 46, + 262, + 287, + 346 + ], + "type": "inline_equation", + "content": "f_{c} = 1" + }, + { + "bbox": [ + 46, + 262, + 287, + 346 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 262, + 287, + 346 + ], + "type": "inline_equation", + "content": "f_{hw} = 8" + }, + { + "bbox": [ + 46, + 262, + 287, + 346 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 262, + 287, + 346 + ], + "type": "inline_equation", + "content": "l = 1" + }, + { + "bbox": [ + 46, + 262, + 287, + 346 + ], + "type": "text", + "content": ", resulting in an actual capacity of 256 bits. We select five baseline methods: three officially used by SD, namely Dwt-Dct [6], DwtDctSvd [6], and RivaGAN [45], a multi-bit watermarking called Stable Signature [10], and a train-free invisible watermarking called Tree-Ring [41]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 348, + 287, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 348, + 287, + 383 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 287, + 383 + ], + "type": "text", + "content": "Robustness evaluation To evaluate the robustness, we select nine representative types of noise shown in Fig. 4. We conduct experiments following the noise strength in Fig. 4." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 384, + 287, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 287, + 479 + ], + "type": "text", + "content": "Evaluation metrics. In the detection scenario, we calculate the true positive rate (TPR) corresponding to a fixed false positive rate (FPR). In the traceability scenario, we calculate the bit accuracy. To measure the bias in model performance, we compute the FID [14] and CLIP-Score [29] for 10 batches of watermarked images and perform a " + }, + { + "bbox": [ + 46, + 384, + 287, + 479 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 384, + 287, + 479 + ], + "type": "text", + "content": "-test on the mean FID and CLIP-Score compared to that of watermark-free images." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 480, + 287, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 480, + 287, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 480, + 287, + 504 + ], + "type": "text", + "content": "All experiments are conducted using the PyTorch 1.13.0 framework, running on a single RTX 3090 GPU." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 47, + 514, + 230, + 527 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 230, + 527 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 230, + 527 + ], + "type": "text", + "content": "4.2. Performance of Gaussian Shading" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 46, + 533, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 652 + ], + "type": "text", + "content": "Detection. In the detection scenario, we consider Gaussian Shading as a single-bit watermark, with a fixed watermark " + }, + { + "bbox": [ + 46, + 533, + 287, + 652 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 533, + 287, + 652 + ], + "type": "text", + "content": ". We approximate the FPR to be controlled at " + }, + { + "bbox": [ + 46, + 533, + 287, + 652 + ], + "type": "inline_equation", + "content": "10^{0}, 10^{-1}, \\ldots, 10^{-13}" + }, + { + "bbox": [ + 46, + 533, + 287, + 652 + ], + "type": "text", + "content": ", calculate the corresponding threshold " + }, + { + "bbox": [ + 46, + 533, + 287, + 652 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 533, + 287, + 652 + ], + "type": "text", + "content": ", and test the TPR on 1,000 watermarked images To mitigate the effects of randomness, we perform 5 trials with different " + }, + { + "bbox": [ + 46, + 533, + 287, + 652 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 533, + 287, + 652 + ], + "type": "text", + "content": " and compute the average TPR. See Fig. 5a, when the FPR is controlled at " + }, + { + "bbox": [ + 46, + 533, + 287, + 652 + ], + "type": "inline_equation", + "content": "10^{-13}" + }, + { + "bbox": [ + 46, + 533, + 287, + 652 + ], + "type": "text", + "content": ", the TPR remains at least 0.99 for eight out of the nine cases. Although the TPR for Brightness is only 0.953, it is still a promising result." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "Traceability. In this scenario, Gaussian Shading serves as a multi-bit watermark. Assuming Alice provides services to " + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": " users, Alice needs to allocate one watermark for each user. In our experiments, we assume that " + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "N' = 1,000" + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": " users generate images, with each user generating 10 images, re" + } + ] + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 316, + 70, + 411, + 141 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 411, + 141 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 411, + 141 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 411, + 141 + ], + "type": "image", + "image_path": "774162d867b79361428a6cb94c725181280f963f31253856bb88a5940fa1d48c.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 144, + 397, + 153 + ], + "lines": [ + { + "bbox": [ + 329, + 144, + 397, + 153 + ], + "spans": [ + { + "bbox": [ + 329, + 144, + 397, + 153 + ], + "type": "text", + "content": "(a) Detection results." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 343, + 156, + 507, + 166 + ], + "lines": [ + { + "bbox": [ + 343, + 156, + 507, + 166 + ], + "spans": [ + { + "bbox": [ + 343, + 156, + 507, + 166 + ], + "type": "text", + "content": "Figure 5. Performance of Gaussian Shading." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 436, + 70, + 533, + 141 + ], + "blocks": [ + { + "bbox": [ + 436, + 70, + 533, + 141 + ], + "lines": [ + { + "bbox": [ + 436, + 70, + 533, + 141 + ], + "spans": [ + { + "bbox": [ + 436, + 70, + 533, + 141 + ], + "type": "image", + "image_path": "aa640412f42f561a96912774a9dce03db8c6cc463236ac6429791ff69841e72a.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 446, + 144, + 522, + 154 + ], + "lines": [ + { + "bbox": [ + 446, + 144, + 522, + 154 + ], + "spans": [ + { + "bbox": [ + 446, + 144, + 522, + 154 + ], + "type": "text", + "content": "(b) Traceability results." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 168, + 512, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 168, + 512, + 180 + ], + "spans": [ + { + "bbox": [ + 306, + 168, + 512, + 180 + ], + "type": "text", + "content": "sulting in a dataset of 10,000 watermarked images." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 182, + 545, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 182, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 304, + 182, + 545, + 337 + ], + "type": "text", + "content": "During testing, we calculate the threshold " + }, + { + "bbox": [ + 304, + 182, + 545, + 337 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 182, + 545, + 337 + ], + "type": "text", + "content": " to control the FPR at " + }, + { + "bbox": [ + 304, + 182, + 545, + 337 + ], + "type": "inline_equation", + "content": "10^{-6}" + }, + { + "bbox": [ + 304, + 182, + 545, + 337 + ], + "type": "text", + "content": ". Note that when computing traceability accuracy, we need to consider two types of errors: false positives, where watermarked images are not detected, and traceability errors, where watermarked images are detected but attributed to the wrong user. Therefore, we first determine whether the image contains a watermark. If it does, we calculate the number of matching bits " + }, + { + "bbox": [ + 304, + 182, + 545, + 337 + ], + "type": "inline_equation", + "content": "Acc" + }, + { + "bbox": [ + 304, + 182, + 545, + 337 + ], + "type": "text", + "content": " with all " + }, + { + "bbox": [ + 304, + 182, + 545, + 337 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 182, + 545, + 337 + ], + "type": "text", + "content": " users on the platform. The user with the highest " + }, + { + "bbox": [ + 304, + 182, + 545, + 337 + ], + "type": "inline_equation", + "content": "Acc" + }, + { + "bbox": [ + 304, + 182, + 545, + 337 + ], + "type": "text", + "content": " is considered the one who generated the image. Finally, we verify whether the correct user has been traced. When " + }, + { + "bbox": [ + 304, + 182, + 545, + 337 + ], + "type": "inline_equation", + "content": "N > N'" + }, + { + "bbox": [ + 304, + 182, + 545, + 337 + ], + "type": "text", + "content": ", it can be assumed that some users have been assigned a watermark but have not generated any images." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 304, + 337, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 337, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 337, + 545, + 396 + ], + "type": "text", + "content": "See Fig. 5b, when " + }, + { + "bbox": [ + 304, + 337, + 545, + 396 + ], + "type": "inline_equation", + "content": "N = 10^{6}" + }, + { + "bbox": [ + 304, + 337, + 545, + 396 + ], + "type": "text", + "content": ", Gaussian Shading exhibits almost perfect traceability in seven cases. Although the traceability accuracy for Brightness is only " + }, + { + "bbox": [ + 304, + 337, + 545, + 396 + ], + "type": "inline_equation", + "content": "95.47\\%" + }, + { + "bbox": [ + 304, + 337, + 545, + 396 + ], + "type": "text", + "content": ", if a user generates two images, the probability of successfully tracing him is still no less than " + }, + { + "bbox": [ + 304, + 337, + 545, + 396 + ], + "type": "inline_equation", + "content": "99\\%" + }, + { + "bbox": [ + 304, + 337, + 545, + 396 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 306, + 406, + 444, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 406, + 444, + 418 + ], + "spans": [ + { + "bbox": [ + 306, + 406, + 444, + 418 + ], + "type": "text", + "content": "4.3. Comparison to Baselines" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 304, + 426, + 545, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 473 + ], + "type": "text", + "content": "In this section, we compare the performance of Gaussian Shading with baselines on SD V1.4, V2.0, and V2.1. We use our implementations for each method, see details in Supplementary Material." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 304, + 474, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 569 + ], + "type": "text", + "content": "We conduct tests on 1,000 generated images for each method respectively. See Tab. 1. Gaussian Shading exhibits strong robustness and significantly outperforms baselines in both scenarios. In terms of bit accuracy, it surpasses the best-performing baseline by approximately " + }, + { + "bbox": [ + 304, + 474, + 545, + 569 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 304, + 474, + 545, + 569 + ], + "type": "text", + "content": ". This can be attributed to the extensive diffusion of the watermark throughout the entire latent space, establishing a profound binding between the watermark and the image semantics." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "To measure the performance bias introduced by the watermark embedding, we apply a " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "-test to evaluate. The hypotheses are " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "H_0: \\mu_s = \\mu_0, H_1: \\mu_s \\neq \\mu_0" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mu_s" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mu_0" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " represent the average FID [14] or CLIP-Score [29] of multiple sets of watermarked and watermark-free images, respectively. A lower " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "-value indicates a higher probability that " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "H_0" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " holds. If the " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "-value is larger than a threshold, " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "H_0" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " is rejected, and model performance is considered to have been affected. See Tab. 1, Gaussian Shading achieves the smallest " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "-value, which indirectly reflects its performance-lossless characteristic. For a detailed analysis of the " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "-test, please refer to the Supplementary Material." + } + ] + } + ], + "index": 36 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12167" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 71, + 547, + 169 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 547, + 169 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 547, + 169 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 547, + 169 + ], + "type": "table", + "html": "
MethodsMetrics
TPR (Clean)TPR (Adversarial)Bit Acc. (Clean)Bit Acc. (Adversarial)FID (t-value ↓)CLIP-Score (t-value ↓)
Stable Diffusion----25.23±.180.3629±.0006
DwtDct [6]0.825/0.881/0.8660.172/0.178/0.1730.8030/0.8059/0.80230.5696/0.5671/0.562224.97±.19 (3.026)0.3617±.0007 (3.045)
DwtDctSvd [6]1.000/1.000/1.0000.597/0.594/0.5990.9997/0.9987/0.99870.6920/0.6868/0.690524.45±.22 (8.253)0.3609±.0009 (4.452)
RivaGAN [45]0.920/0.945/0.9630.697/0.697/0.7060.9762/0.9877/0.99210.8986/0.9124/0.901924.24±.16 (12.29)0.3611±.0009 (4.259)
Tree-Ring [41]1.000/1.000/1.0000.894/0.898/0.906--25.43±.13 (2.581)0.3632±.0006 (0.8278)
Stable Signature [10]1.000/1.000/1.0000.502/0.505/0.4960.9987/0.9978/0.99790.7520/0.7472/0.750025.45±.18 (2.477)0.3622±.0027 (0.7066)
Ours1.000/1.000/1.0000.997/0.998/0.9960.9999/0.9999/0.99990.9753/0.9749/0.972425.20±.22 (0.3567)0.3631±.0005 (0.6870)
", + "image_path": "236704faa0a707312d5346244dee764e1ac373bce22c3ad9b63c8f8871350458.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 171, + 546, + 204 + ], + "lines": [ + { + "bbox": [ + 46, + 171, + 546, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 546, + 204 + ], + "type": "text", + "content": "Table 1. Comparison results. We control the FPR at " + }, + { + "bbox": [ + 46, + 171, + 546, + 204 + ], + "type": "inline_equation", + "content": "10^{-6}" + }, + { + "bbox": [ + 46, + 171, + 546, + 204 + ], + "type": "text", + "content": ", and evaluate the TPR and bit accuracy for SD V1.4/V2.0/V2.1. To assess the bias in model performance, we conduct a " + }, + { + "bbox": [ + 46, + 171, + 546, + 204 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 171, + 546, + 204 + ], + "type": "text", + "content": "-test on SD V2.1. Adversarial here refers to the average performance of a series of noises. Additional results can be found in Supplementary Material." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 63, + 205, + 154, + 275 + ], + "blocks": [ + { + "bbox": [ + 63, + 205, + 154, + 275 + ], + "lines": [ + { + "bbox": [ + 63, + 205, + 154, + 275 + ], + "spans": [ + { + "bbox": [ + 63, + 205, + 154, + 275 + ], + "type": "image", + "image_path": "db4c4943a11b71c4f21ace1b07c62f59f4d4daf753525c2cda36cf1450e25b33.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 156, + 205, + 248, + 275 + ], + "blocks": [ + { + "bbox": [ + 156, + 205, + 248, + 275 + ], + "lines": [ + { + "bbox": [ + 156, + 205, + 248, + 275 + ], + "spans": [ + { + "bbox": [ + 156, + 205, + 248, + 275 + ], + "type": "image", + "image_path": "b84c926be2f4085637d5b5c005f44ffb54fa0ad2895819f40434dfc0a5df6b94.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 251, + 205, + 342, + 275 + ], + "blocks": [ + { + "bbox": [ + 251, + 205, + 342, + 275 + ], + "lines": [ + { + "bbox": [ + 251, + 205, + 342, + 275 + ], + "spans": [ + { + "bbox": [ + 251, + 205, + 342, + 275 + ], + "type": "image", + "image_path": "87353aed10bc3304f8afcd28ee3be310f682ffb1cfa196dbb473192af11974be.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 345, + 205, + 436, + 275 + ], + "blocks": [ + { + "bbox": [ + 345, + 205, + 436, + 275 + ], + "lines": [ + { + "bbox": [ + 345, + 205, + 436, + 275 + ], + "spans": [ + { + "bbox": [ + 345, + 205, + 436, + 275 + ], + "type": "image", + "image_path": "1849f464e6b319de72170104ae614ecc91f64b614bc08e5901c5cb9716282939.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 439, + 205, + 531, + 275 + ], + "blocks": [ + { + "bbox": [ + 439, + 205, + 531, + 275 + ], + "lines": [ + { + "bbox": [ + 439, + 205, + 531, + 275 + ], + "spans": [ + { + "bbox": [ + 439, + 205, + 531, + 275 + ], + "type": "image", + "image_path": "acc123a5fbb8846a038efd042a4ce1bbab5e99ba593033d37ec567f62895b3e5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 60, + 288, + 152, + 357 + ], + "blocks": [ + { + "bbox": [ + 76, + 278, + 141, + 287 + ], + "lines": [ + { + "bbox": [ + 76, + 278, + 141, + 287 + ], + "spans": [ + { + "bbox": [ + 76, + 278, + 141, + 287 + ], + "type": "text", + "content": "(a) Guidance scales." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 60, + 288, + 152, + 357 + ], + "lines": [ + { + "bbox": [ + 60, + 288, + 152, + 357 + ], + "spans": [ + { + "bbox": [ + 60, + 288, + 152, + 357 + ], + "type": "image", + "image_path": "d14af2497e22ab79cac4adf3bc18bb022c2d364c466615121b42a3e96c298b49.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 360, + 134, + 369 + ], + "lines": [ + { + "bbox": [ + 78, + 360, + 134, + 369 + ], + "spans": [ + { + "bbox": [ + 78, + 360, + 134, + 369 + ], + "type": "text", + "content": "(f) Median Filter." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 155, + 288, + 246, + 357 + ], + "blocks": [ + { + "bbox": [ + 187, + 278, + 219, + 287 + ], + "lines": [ + { + "bbox": [ + 187, + 278, + 219, + 287 + ], + "spans": [ + { + "bbox": [ + 187, + 278, + 219, + 287 + ], + "type": "text", + "content": "(b) JPEG" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 155, + 288, + 246, + 357 + ], + "lines": [ + { + "bbox": [ + 155, + 288, + 246, + 357 + ], + "spans": [ + { + "bbox": [ + 155, + 288, + 246, + 357 + ], + "type": "image", + "image_path": "336e0f66502a768f3fd7a00d31195f7d9a2a7a4135001346c9868fbcbec5e3bb.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 168, + 360, + 233, + 369 + ], + "lines": [ + { + "bbox": [ + 168, + 360, + 233, + 369 + ], + "spans": [ + { + "bbox": [ + 168, + 360, + 233, + 369 + ], + "type": "text", + "content": "(g) Gaussian Noise." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 249, + 288, + 339, + 357 + ], + "blocks": [ + { + "bbox": [ + 267, + 278, + 326, + 288 + ], + "lines": [ + { + "bbox": [ + 267, + 278, + 326, + 288 + ], + "spans": [ + { + "bbox": [ + 267, + 278, + 326, + 288 + ], + "type": "text", + "content": "(c) Random Crop." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 249, + 288, + 339, + 357 + ], + "lines": [ + { + "bbox": [ + 249, + 288, + 339, + 357 + ], + "spans": [ + { + "bbox": [ + 249, + 288, + 339, + 357 + ], + "type": "image", + "image_path": "ad94a8eb28303917d25d49dfcd8b99f9fbf0ec49a6a48c1c88eff66d6f002d36.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 251, + 360, + 337, + 369 + ], + "lines": [ + { + "bbox": [ + 251, + 360, + 337, + 369 + ], + "spans": [ + { + "bbox": [ + 251, + 360, + 337, + 369 + ], + "type": "text", + "content": "(h) Salt and Pepper Noise." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 247, + 372, + 345, + 381 + ], + "lines": [ + { + "bbox": [ + 247, + 372, + 345, + 381 + ], + "spans": [ + { + "bbox": [ + 247, + 372, + 345, + 381 + ], + "type": "text", + "content": "Figure 6. Ablation studies." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 343, + 288, + 433, + 357 + ], + "blocks": [ + { + "bbox": [ + 361, + 278, + 421, + 288 + ], + "lines": [ + { + "bbox": [ + 361, + 278, + 421, + 288 + ], + "spans": [ + { + "bbox": [ + 361, + 278, + 421, + 288 + ], + "type": "text", + "content": "(d) Random Drop." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 343, + 288, + 433, + 357 + ], + "lines": [ + { + "bbox": [ + 343, + 288, + 433, + 357 + ], + "spans": [ + { + "bbox": [ + 343, + 288, + 433, + 357 + ], + "type": "image", + "image_path": "d5c033138c9afd948058844db3e2be752c786e291bc9636a0f800e456fc3b839.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 371, + 360, + 405, + 369 + ], + "lines": [ + { + "bbox": [ + 371, + 360, + 405, + 369 + ], + "spans": [ + { + "bbox": [ + 371, + 360, + 405, + 369 + ], + "type": "text", + "content": "(i) Resize." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 437, + 288, + 528, + 357 + ], + "blocks": [ + { + "bbox": [ + 454, + 278, + 513, + 287 + ], + "lines": [ + { + "bbox": [ + 454, + 278, + 513, + 287 + ], + "spans": [ + { + "bbox": [ + 454, + 278, + 513, + 287 + ], + "type": "text", + "content": "(e) Gaussian Blur." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 437, + 288, + 528, + 357 + ], + "lines": [ + { + "bbox": [ + 437, + 288, + 528, + 357 + ], + "spans": [ + { + "bbox": [ + 437, + 288, + 528, + 357 + ], + "type": "image", + "image_path": "66a5694f8e5461951c77626d1771d39b23d6d145c3288595c87e27d5bc80a502.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 459, + 360, + 506, + 370 + ], + "lines": [ + { + "bbox": [ + 459, + 360, + 506, + 370 + ], + "spans": [ + { + "bbox": [ + 459, + 360, + 506, + 370 + ], + "type": "text", + "content": "(j) Brightness." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 47, + 383, + 147, + 394 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 383, + 147, + 394 + ], + "spans": [ + { + "bbox": [ + 47, + 383, + 147, + 394 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 46, + 401, + 287, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 287, + 449 + ], + "type": "text", + "content": "In this section, we conduct comprehensive ablation experiments on SD V2.1 to demonstrate hyperparameter selection. Unless specified, we generate 1,000 images and test the TPR and the bit accuracy with a theoretical FPR of " + }, + { + "bbox": [ + 46, + 401, + 287, + 449 + ], + "type": "inline_equation", + "content": "10^{-6}" + }, + { + "bbox": [ + 46, + 401, + 287, + 449 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "text", + "content": "Watermark capacity. The watermark capacity is determined by three parameters: channel diffusion factor " + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "inline_equation", + "content": "f_{c}" + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "text", + "content": ", height-width diffusion factor " + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "inline_equation", + "content": "f_{hw}" + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "text", + "content": ", and embedding rate " + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "text", + "content": ". See Tab. 2, to balance the capacity and robustness of Gaussian Shading, we chose " + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "inline_equation", + "content": "f_{c} = 1" + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "inline_equation", + "content": "f_{hw} = 8" + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "text", + "content": ". After fixing " + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "inline_equation", + "content": "f_{c}" + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "inline_equation", + "content": "f_{hw}" + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "text", + "content": ", we vary " + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "text", + "content": " to examine if it could enhance the capacity, and additional results can be found in Supplementary Material. Considering all factors, we determine that the optimal solution is " + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "inline_equation", + "content": "f_{c} = 1" + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "inline_equation", + "content": "f_{hw} = 8" + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "inline_equation", + "content": "l = 1" + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "text", + "content": ", resulting in a watermark capacity of 256 bits." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 46, + 570, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 629 + ], + "type": "text", + "content": "Sampling methods. To validate the generalization, we select five commonly used sampling methods, all continuous-time samplers based on ODE solvers [32]. See Tab. 3, all of them exhibit excellent performance with a bit accuracy of approximately " + }, + { + "bbox": [ + 46, + 570, + 287, + 629 + ], + "type": "inline_equation", + "content": "97\\%" + }, + { + "bbox": [ + 46, + 570, + 287, + 629 + ], + "type": "text", + "content": " against noises." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 46, + 630, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 700 + ], + "type": "text", + "content": "Impact of the inversion step. In practice, the inference step is often unknown, which introduces a mismatch with the inversion step. See Tab. 4, such mismatch introduces minimal loss in accuracy. Considering the high efficiency of existing samplers, the inference step generally does not exceed 50. Therefore, we set the inversion step to 50." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 47, + 702, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 702, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 702, + 287, + 713 + ], + "type": "text", + "content": "Guidance scales. Given diverse user preferences for image" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 383, + 545, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 383, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 383, + 545, + 491 + ], + "type": "text", + "content": "prompt alignment, larger guidance scales ensure faithful adherence to prompts, while smaller scales grant the model greater creative freedom. In SD, the guidance scale is typically selected from the range of [5, 15]. Hence, experiments cover the range of 2 to 18. For the inversion, an empty prompt is used for guidance, and the guidance scale is fixed at 1, assuming unknown information during extraction. In Fig. 6a, the bit accuracy of Gaussian Shading surpasses " + }, + { + "bbox": [ + 304, + 383, + 545, + 491 + ], + "type": "inline_equation", + "content": "99.9\\%" + }, + { + "bbox": [ + 304, + 383, + 545, + 491 + ], + "type": "text", + "content": ", showing its reliability in real-world-like scenes." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 491, + 545, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 491, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 304, + 491, + 545, + 563 + ], + "type": "text", + "content": "Noise intensities. To further test the robustness, we conduct experiments using different intensities of noises. See Figs. 6b to 6j, for Random Crop and Gaussian Noise, performance declines significantly with higher intensities. However, for the other seven types of noise, even at high intensities, the bit accuracy remains approximately " + }, + { + "bbox": [ + 304, + 491, + 545, + 563 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 304, + 491, + 545, + 563 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 306, + 574, + 488, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 488, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 488, + 586 + ], + "type": "text", + "content": "4.5. Attacks against Gaussian Shading" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 304, + 594, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 653 + ], + "type": "text", + "content": "We consider two malicious attacks: compression attack, where the attacker employs a neural network to compress watermarked images, and inversion attack, assuming the attacker is aware of the watermark embedding method, enabling them to modify the image's latent representations." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "Compression attack. We utilize popular auto-encoders [2, 5, 9, 30] to compare Stable Signature (SS) with Gaussian Shading across various compression rates. Additionally, we assess the compression quality through the PSNR between the compressed and watermarked images. See Figs. 7a" + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12168" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 102, + 70, + 491, + 125 + ], + "blocks": [ + { + "bbox": [ + 102, + 70, + 491, + 125 + ], + "lines": [ + { + "bbox": [ + 102, + 70, + 491, + 125 + ], + "spans": [ + { + "bbox": [ + 102, + 70, + 491, + 125 + ], + "type": "table", + "html": "
Noisefc-fhw (k bits)
1-2 (4096)4-1 (4096)1-4 (1024)4-2 (1024)1-8 (256)4-4 (256)1-16 (64)4-8 (64)
None0.94130.93800.99850.99800.99990.99991.00001.0000
Adversarial0.73020.72380.87690.86140.97240.96710.99590.9953
", + "image_path": "842d3e3a041d7f93a6daa14896dd108074cedb783e2deef6d14fa76b70826991.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 144, + 279, + 209 + ], + "blocks": [ + { + "bbox": [ + 61, + 126, + 530, + 137 + ], + "lines": [ + { + "bbox": [ + 61, + 126, + 530, + 137 + ], + "spans": [ + { + "bbox": [ + 61, + 126, + 530, + 137 + ], + "type": "text", + "content": "Table 2. Bit accuracy with different factors " + }, + { + "bbox": [ + 61, + 126, + 530, + 137 + ], + "type": "inline_equation", + "content": "{f}_{c}" + }, + { + "bbox": [ + 61, + 126, + 530, + 137 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 61, + 126, + 530, + 137 + ], + "type": "inline_equation", + "content": "{f}_{hw}" + }, + { + "bbox": [ + 61, + 126, + 530, + 137 + ], + "type": "text", + "content": " ,where " + }, + { + "bbox": [ + 61, + 126, + 530, + 137 + ], + "type": "inline_equation", + "content": "l = 1" + }, + { + "bbox": [ + 61, + 126, + 530, + 137 + ], + "type": "text", + "content": " . Additional results can be found in Supplementary Material." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 144, + 279, + 209 + ], + "lines": [ + { + "bbox": [ + 56, + 144, + 279, + 209 + ], + "spans": [ + { + "bbox": [ + 56, + 144, + 279, + 209 + ], + "type": "table", + "html": "
NoiseSampling Methods
DDIM [32]UniPC [48]PNDM [23]DEIS [46]DPMSolver [25]
None0.99991.00001.00000.99990.9999
Adversarial0.97060.96280.97210.97150.9724
", + "image_path": "2d227749735613639e5af587adfd7bd99ce2aa9f8c52ba03f1f958ba386c37a2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 83, + 278, + 251, + 352 + ], + "blocks": [ + { + "bbox": [ + 46, + 209, + 287, + 276 + ], + "lines": [ + { + "bbox": [ + 46, + 209, + 287, + 276 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 287, + 276 + ], + "type": "text", + "content": "Table 3. Bit accuracy with different sampling methods. Additional results can be found in Supplementary Material. These methods differ only in accuracy and order. DDIM is a first-order estimate of the ODE. Accordingly, DDIM inversion ensures a lower bound on the accuracy of the inversion process. Therefore, it can naturally be applied to higher-order and higher-accuracy methods." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 83, + 278, + 251, + 352 + ], + "lines": [ + { + "bbox": [ + 83, + 278, + 251, + 352 + ], + "spans": [ + { + "bbox": [ + 83, + 278, + 251, + 352 + ], + "type": "table", + "html": "
Inference StepInversion Step
102550100
100.99990.99990.99990.9999
250.99980.99991.00001.0000
500.99950.99970.99990.9999
1000.99940.99960.99990.9999
", + "image_path": "e8ef5d3d10c7dea97f508e187dda37875eec9b1aa7a2804c24e59f57738f74d4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 353, + 287, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 353, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 46, + 353, + 287, + 414 + ], + "type": "text", + "content": "Table 4. Bit accuracy with different inference and inversion step. and 7b, Gaussian Shading significantly outperforms Stable Signature. This is because Gaussian Shading diffuses the watermark across the entire semantic space of images, while Stable Signature relies solely on the image texture." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 415, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 287, + 521 + ], + "type": "text", + "content": "Inversion attack. Assuming the attacker is aware of the embedding method, a more effective approach to erasing is through inversion to obtain latent representations and subsequently modify them. We validate the robustness against such attacks. Importantly, our experiments assume the strongest attacker capability of using the same model as Alice for precise inversion. In real-world scenarios, where the watermark embedding is not publicly available, the attacker's capabilities would be weaker." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 521, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 629 + ], + "type": "text", + "content": "Specifically, we perform inversion to obtain latent representations and randomly flip a certain rate of them. Using the flipped latent representations, we regenerate the images and extract the watermark. See Fig. 7c. the watermark can still be reliably extracted when the flipping rate (FR) is less than 0.4. At high FRs, significant changes in images are observed. Although the watermark cannot be accurately extracted, we consider the image transformed into a different one, resulting in the content not intended to be protected." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "content": "From another perspective, the attacker can launch a forgery attack by performing inversion on an innocuous image from Bob and generating harmful content using a different prompt. See Fig. 7c, when the FR is 0, Alice can accurately trace Bob based on the forgeries, enabling the attacker to successfully frame Bob. Therefore, protecting the model from leakage is crucial for operators." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 144, + 385, + 204 + ], + "blocks": [ + { + "bbox": [ + 307, + 144, + 385, + 204 + ], + "lines": [ + { + "bbox": [ + 307, + 144, + 385, + 204 + ], + "spans": [ + { + "bbox": [ + 307, + 144, + 385, + 204 + ], + "type": "image", + "image_path": "79029d1b4edd943077dd20adea6a00b115b845037db4d141053738cfbf85dc13.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 206, + 381, + 216 + ], + "lines": [ + { + "bbox": [ + 312, + 206, + 381, + 216 + ], + "spans": [ + { + "bbox": [ + 312, + 206, + 381, + 216 + ], + "type": "text", + "content": "(a) Detection results." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 388, + 144, + 464, + 205 + ], + "blocks": [ + { + "bbox": [ + 388, + 144, + 464, + 205 + ], + "lines": [ + { + "bbox": [ + 388, + 144, + 464, + 205 + ], + "spans": [ + { + "bbox": [ + 388, + 144, + 464, + 205 + ], + "type": "image", + "image_path": "b1e3e8caa276f4aa3216dcbc155f939b4c6b9469e3389d584c640c634f685772.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 388, + 206, + 463, + 216 + ], + "lines": [ + { + "bbox": [ + 388, + 206, + 463, + 216 + ], + "spans": [ + { + "bbox": [ + 388, + 206, + 463, + 216 + ], + "type": "text", + "content": "(b) Traceability results." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 218, + 545, + 251 + ], + "lines": [ + { + "bbox": [ + 305, + 218, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 305, + 218, + 545, + 251 + ], + "type": "text", + "content": "Figure 7. Performance of Gaussian Shading under Malicious Attack, where (a) and (b) are under compression attack and (c) is under inversion attack." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 467, + 144, + 544, + 204 + ], + "blocks": [ + { + "bbox": [ + 467, + 144, + 544, + 204 + ], + "lines": [ + { + "bbox": [ + 467, + 144, + 544, + 204 + ], + "spans": [ + { + "bbox": [ + 467, + 144, + 544, + 204 + ], + "type": "image", + "image_path": "9d4b715f4373508fe44f20ea037ed3cd283cdaeef1e62d4c41cf8b7d50e169ba.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 473, + 206, + 537, + 215 + ], + "lines": [ + { + "bbox": [ + 473, + 206, + 537, + 215 + ], + "spans": [ + { + "bbox": [ + 473, + 206, + 537, + 215 + ], + "type": "text", + "content": "(c) Inversion attack." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 256, + 381, + 269 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 256, + 381, + 269 + ], + "spans": [ + { + "bbox": [ + 306, + 256, + 381, + 269 + ], + "type": "text", + "content": "5. Limitations" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 277, + 545, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 545, + 469 + ], + "type": "text", + "content": "Despite extensive experimental validation of Gaussian Shading's superior performance, our work still has certain limitations. Firstly, the usage scenarios are restricted due to the reliance on DDIM inversion [32], which necessitates the utilization of continuous-time samplers based on ODE solvers [32] like DPMSolver [25]. Secondly, Gaussian Shading employs stream ciphers, necessitating proper key usage and management on the deployment platform. Additionally, we assume that the model is not publicly accessible, and only operators can verify the watermark, providing a certain level of protection against white-box attacks and ensuring security. However, if a legitimate third party requires watermark verification, cooperation from the operators becomes necessary. Lastly, Gaussian Shading is vulnerable to forgery attacks, emphasizing the importance for operators to safeguard the model parameters." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 479, + 471, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 471, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 471, + 491 + ], + "type": "text", + "content": "6. Conclusion and Future Work" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 498, + 545, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 606 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 606 + ], + "type": "text", + "content": "We propose Gaussian Shading, a provably performance-lossless watermarking applied to diffusion models. Compared to baseline methods, Gaussian Shading offers simplicity and effectiveness by making a simple modification in the sampling process of the initial latent representation. Extensive experiments validate the superior performance in both detection and traceability scenarios. To our knowledge, we are the first to propose and implement a performance-lossless approach in image watermarking." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 607, + 545, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 607, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 304, + 607, + 545, + 654 + ], + "type": "text", + "content": "Regarding future work, we will introduce more efficient inversion methods [40, 44] and include a wider range of sampling methods. Additionally, careful consideration should be given to counteracting forgery attacks." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgement. This work was supported in part by the Natural Science Foundation of China under Grant U2336206, 62102386, 62072421, 62372423, and 62121002." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12169" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "type": "text", + "content": "[1] Ali Al-Haj. Combined dwt-dct digital image watermarking. Journal of computer science, 3(9):740-746, 2007. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 114, + 287, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 287, + 158 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 287, + 158 + ], + "type": "text", + "content": "[2] Johannes Ballé, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. arXiv preprint arXiv:1802.01436, 2018. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 159, + 287, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 159, + 287, + 182 + ], + "spans": [ + { + "bbox": [ + 53, + 159, + 287, + 182 + ], + "type": "text", + "content": "[3] Daniel J Bernstein et al. Chacha, a variant of salsa20. In Workshop record of SASC, pages 3-5. Citeseer, 2008. 4, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 183, + 287, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 183, + 287, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 183, + 287, + 237 + ], + "type": "text", + "content": "[4] Kejiang Chen, Hang Zhou, Hanqing Zhao, Dongdong Chen, Weiming Zhang, and Nenghai Yu. Distribution-preserving steganography based on text-to-speech generative models. IEEE Transactions on Dependable and Secure Computing, 19(5):3343-3356, 2021. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 239, + 287, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 287, + 295 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 287, + 295 + ], + "type": "text", + "content": "[5] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned image compression with discretized gaussian mixture likelihoods and attention modules. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7939-7948, 2020. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 296, + 287, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 296, + 287, + 329 + ], + "spans": [ + { + "bbox": [ + 53, + 296, + 287, + 329 + ], + "type": "text", + "content": "[6] Ingemar Cox, Matthew Miller, Jeffrey Bloom, Jessica Fridrich, and Ton Kalker. Digital watermarking and steganography. Morgan Kaufmann, 2007. 2, 3, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 331, + 287, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 331, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 53, + 331, + 287, + 374 + ], + "type": "text", + "content": "[7] Yingqian Cui, Jie Ren, Han Xu, Pengfei He, Hui Liu, Lichao Sun, and Jiliang Tang. Diffusionshield: A watermark for copyright protection against generative diffusion models. arXiv preprint arXiv:2306.04642, 2023. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 376, + 287, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 376, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 53, + 376, + 287, + 408 + ], + "type": "text", + "content": "[8] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 410, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 410, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 53, + 410, + 287, + 453 + ], + "type": "text", + "content": "[9] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 455, + 287, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 455, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 48, + 455, + 287, + 499 + ], + "type": "text", + "content": "[10] Pierre Fernandez, Guillaume Couairon, Hervé Jégou, Matthijs Douze, and Teddy Furon. The stable signature: Rooting watermarks in latent diffusion models. arXiv preprint arXiv:2303.15435, 2023. 2, 3, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 501, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 287, + 555 + ], + "type": "text", + "content": "[11] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10696-10706, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 557, + 287, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 287, + 600 + ], + "type": "text", + "content": "[12] Huiping Guo and Nicolas D Georganas. Digital image watermarking for joint ownership. In Proceedings of the tenth ACM international conference on Multimedia, pages 362-371, 2002. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 602, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 656 + ], + "type": "text", + "content": "[13] Mohamed Hamidi, Mohamed El Haziti, Hocine Cherifi, and Mohammed El Hassouni. Hybrid blind robust image watermarking technique based on dft-dct and arnold transform. Multimedia Tools and Applications, 77:27181-27214, 2018. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5, 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 715 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "text", + "content": "[15] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 97, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 97, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 307, + 97, + 545, + 129 + ], + "type": "text", + "content": "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 130, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 130, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 130, + 545, + 185 + ], + "type": "text", + "content": "[17] Nicholas J Hopper, John Langford, and Luis Von Ahn. Provably secure steganography. In Advances in Cryptology—CRYPTO 2002: 22nd Annual International Cryptology Conference Santa Barbara, California, USA, August 18–22, 2002 Proceedings 22, pages 77–92. Springer, 2002. 4, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 186, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 240 + ], + "type": "text", + "content": "[18] Zhaoyang Jia, Han Fang, and Weiming Zhang. Mbrs: Enhancing robustness of dnn-based watermarking by minibatch of real and simulatedJPEG compression. In Proceedings of the 29th ACM international conference on multimedia, pages 41-49, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "type": "text", + "content": "[19] Varsha Kishore, Xiangyu Chen, Yan Wang, Boyi Li, and Kilian Q Weinberger. Fixed neural network steganography: Train the images, not the network. In International Conference on Learning Representations, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 286, + 545, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 545, + 331 + ], + "type": "text", + "content": "[20] Deepa Kundur and Dimitrios Hatzinakos. A robust digital image watermarking method using wavelet-based fusion. In Proceedings of International Conference on Image Processing, pages 544-547. IEEE, 1997. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 332, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 332, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 307, + 332, + 545, + 374 + ], + "type": "text", + "content": "[21] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 376, + 545, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 421 + ], + "type": "text", + "content": "[22] Sunil Lee, Chang D Yoo, and Ton Kalker. Reversible image watermarking based on integer-to-integer wavelet transform. IEEE Transactions on information forensics and security, 2 (3):321-330, 2007. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 422, + 545, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 545, + 455 + ], + "type": "text", + "content": "[23] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. arXiv preprint arXiv:2202.09778, 2022. 4, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 456, + 545, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 488 + ], + "type": "text", + "content": "[24] Yugeng Liu, Zheng Li, Michael Backes, Yun Shen, and Yang Zhang. Watermarking diffusion model. arXiv preprint arXiv:2305.12502, 2023. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 490, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 490, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 490, + 545, + 544 + ], + "type": "text", + "content": "[25] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 2, 4, 5, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 545, + 545, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 545, + 590 + ], + "type": "text", + "content": "[26] Xiyang Luo, Ruohan Zhan, Huiwen Chang, Feng Yang, and Peyman Milanfar. Distortion agnostic deep watermarking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13548-13557, 2020. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 590, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 644 + ], + "type": "text", + "content": "[27] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "text", + "content": "[28] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 691, + 545, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 715 + ], + "type": "text", + "content": "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry," + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12170" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 116 + ], + "type": "text", + "content": "Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 288, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 288, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 288, + 173 + ], + "type": "text", + "content": "[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 5, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 228 + ], + "type": "text", + "content": "[31] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 230, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 287, + 262 + ], + "type": "text", + "content": "[32] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2020. 2, 4, 5, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 297 + ], + "type": "text", + "content": "[33] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 298, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 287, + 352 + ], + "type": "text", + "content": "[34] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2020. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 354, + 287, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 398 + ], + "type": "text", + "content": "[35] Srdjan Stankovic, Irena Orovic, and Nikola Zaric. An application of multidimensional time-frequency analysis as a base for the unified watermarking approach. IEEE Transactions on Image Processing, 19(3):736-745, 2009. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 399, + 287, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 443 + ], + "type": "text", + "content": "[36] Matthew Tancik, Ben Mildenhall, and Ren Ng. Stegastamp: Invisible hyperlinks in physical photographs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2117-2126, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 445, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 287, + 487 + ], + "type": "text", + "content": "[37] Min-Jen Tsai, Kuang-Yao Yu, and Yi-Zhang Chen. Joint wavelet and spatial transformation for digital watermarking. IEEE Transactions on Consumer Electronics, 46(1):237, 2000. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 489, + 287, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 489, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 48, + 489, + 287, + 532 + ], + "type": "text", + "content": "[38] Ron G Van Schyndel, Andrew Z Tirkel, and Charles F Osborne. A digital watermark. In Proceedings of 1st international conference on image processing, pages 86-90. IEEE, 1994. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 535, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 535, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 535, + 287, + 578 + ], + "type": "text", + "content": "[39] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 580, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 287, + 623 + ], + "type": "text", + "content": "[40] Bram Wallace, Akash Gokul, and Nikhil Naik. Edict: Exact diffusion inversion via coupled transformations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22532-22541, 2023. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "text", + "content": "[41] Yuxin Wen, John Kirchenbauer, Jonas Geiping, and Tom Goldstein. Tree-ring watermarks: Fingerprints for diffusion images that are invisible and robust. arXiv preprint arXiv:2305.20030, 2023. 2, 3, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[42] Cheng Xiong, Chuan Qin, Guorui Feng, and Xinpeng Zhang. Flexible and secure watermarking for latent diffusion model. In Proceedings of the 31st ACM International Conference on Multimedia, pages 1668-1676, 2023. 2, 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 453 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[43] Chaoning Zhang, Philipp Benz, Adil Karjauv, Geng Sun, and In So Kweon. Udh: Universal deep hiding for steganography, watermarking, and light field messaging. Advances in Neural Information Processing Systems, 33:10223-10234, 2020. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "text", + "content": "[44] Jiaxin Zhang, Kamalika Das, and Sricharan Kumar. On the robustness of diffusion inversion in image manipulation. In ICLR 2023 Workshop on Trustworthy and Reliable Large-Scale Machine Learning Models, 2023. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 163, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 205 + ], + "type": "text", + "content": "[45] Kevin Alex Zhang, Lei Xu, Alfredo Cuesta-Infante, and Kalyan Veeramachaneni. Robust invisible video watermarking with attention. arXiv preprint arXiv:1909.01285, 2019. 2,3,6,7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 208, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 545, + 239 + ], + "type": "text", + "content": "[46] Qinsheng Zhang and Yongxin Chen. Fast sampling of diffusion models with exponential integrator. arXiv preprint arXiv:2204.13902, 2022. 4, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 241, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 274 + ], + "type": "text", + "content": "[47] Weiming Zhang, Kejiang Chen, and Nenghai Yu. Provable secure steganography: Theory, application and prospects. Journal of Cybersecurity, 1:38-46, 2023. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 276, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 319 + ], + "type": "text", + "content": "[48] Wenliang Zhao, Lujia Bai, Yongming Rao, Jie Zhou, and Jiwen Lu. Unipc: A unified predictor-corrector framework for fast sampling of diffusion models. arXiv preprint arXiv:2302.04867, 2023. 4, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 320, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 545, + 363 + ], + "type": "text", + "content": "[49] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Ngai-Man Cheung, and Min Lin. A recipe for watermarking diffusion models. arXiv preprint arXiv:2303.10137, 2023. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 365, + 545, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 408 + ], + "type": "text", + "content": "[50] Xin Zhong, Pei-Chi Huang, Spyridon Mastorakis, and Frank Y Shih. An automated and robust image watermarking scheme based on deep neural networks. IEEE Transactions on Multimedia, 23:1951-1961, 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 410, + 547, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 410, + 547, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 410, + 547, + 453 + ], + "type": "text", + "content": "[51] Jiren Zhu, Russell Kaplan, Justin Johnson, and Li Fei-Fei. Hidden: Hiding data with deep networks. In Proceedings of the European conference on computer vision (ECCV), pages 657–672, 2018. 3" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12171" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Gaussian Shadow Casting for Neural Characters/e6fedb1c-caf9-484c-be44-7dbe074386f9_content_list.json b/2024/Gaussian Shadow Casting for Neural Characters/e6fedb1c-caf9-484c-be44-7dbe074386f9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6be8d47b32756d934d16fdeaa96507f3d1c72cb9 --- /dev/null +++ b/2024/Gaussian Shadow Casting for Neural Characters/e6fedb1c-caf9-484c-be44-7dbe074386f9_content_list.json @@ -0,0 +1,1666 @@ +[ + { + "type": "text", + "text": "Gaussian Shadow Casting for Neural Characters", + "text_level": 1, + "bbox": [ + 238, + 130, + 732, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Luis Bolanos1 Shih-Yang Su1 Helge Rhodin1,2 \n1The University of British Columbia 2Bielefeld University", + "bbox": [ + 243, + 179, + 723, + 218 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/42becdbf56c8af7d828f3e01f48ee9b2d7dc0f848629c4e27af18b2fd12753e8.jpg", + "image_caption": [ + "Figure 1. Gaussian Shadow Casting (GSC): Our method is able to reconstruct 3D neural characters from a sparse set of videos in settings with strong directional illumination. GSC uses a sum of Gaussians density model to cast secondary shadow rays efficiently with an analytic formula. Our method learns to remove shadows from the neural color field, allowing us to relight in novel illuminations. All faces are blurred for anonymity." + ], + "image_footnote": [], + "bbox": [ + 127, + 256, + 228, + 407 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/05b584c53cf36fcb7953e597a75f4ee52831125e10766c03ce275d0a02a298fc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 232, + 257, + 264, + 409 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/10eb7278dd5c34c0c3c1687853f86c26fa0bd19aee45a6becbdf9923e4351e99.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 267, + 257, + 357, + 409 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/64ec36f7e9cf948dc8ce3126a554a8a3b72806968bcc52775fb50f8d628f516f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 257, + 423, + 407 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1e4121215bc8cdefe69fffbce202c5aace9093792f72cdf5e4a5c1d44bb56129.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 433, + 265, + 573, + 405 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4af957060dbe8eb456b75f5a43264a2d8bfef292361a076dab9275e40e70557f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 583, + 265, + 663, + 405 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1dd102a236a19906632678a8071a98b96fadb035081f1a26acf5fb6b95c1478d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 678, + 257, + 839, + 409 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 494, + 313, + 510 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural character models can now reconstruct detailed geometry and texture from video, but they lack explicit shadows and shading, leading to artifacts when generating novel views and poses or during relighting. It is particularly difficult to include shadows as they are a global effect and the required casting of secondary rays is costly. We propose a new shadow model using a Gaussian density proxy that replaces sampling with a simple analytic formula. It supports dynamic motion and is tailored for shadow computation, thereby avoiding the affine projection approximation and sorting required by the closely related Gaussian splattering. Combined with a deferred neural rendering model, our Gaussian shadows enable Lambertian shading and shadow casting with minimal overhead. We demonstrate improved reconstructions, with better separation of albedo, shading, and shadows in challenging outdoor scenes with direct sun light and hard shadows. Our method is able to optimize the light direction without any input from the user. As a result, novel poses have fewer shadow artifacts, and relighting in novel scenes is more realistic compared to the state-of-the-art methods, providing new ways to pose neural characters in novel environments, increasing their applicability. Code available at: https://github.com/LuisBolanos17/GaussianShadowCasting", + "bbox": [ + 75, + 527, + 473, + 891 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 527, + 632, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "It is now possible to reconstruct animatable 3D neural avatars from video but methods do not account for accurate lighting and shadows. They have to rely on recordings that have soft uniform lighting, which precludes recording outdoors in direct sun light and on film sets with spotlights, and most are unable to relight characters in novel environments, limiting their applicability in content creation.", + "bbox": [ + 496, + 551, + 890, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The most recent body models [17,20,28,36,37,40] which are based on neural radiance fields (NeRFs) [26], approximate the light transport by casting primary rays between the camera and the scene, sampling the underlying neural network dozens of times along each ray to obtain the density and color. As they do not include an illumination model, the color that the NeRF learns includes lighting, shadow, and view-dependent effects. Learning a body model in a challenging scene with a strong directional light source, such as the sun, leads to the neural field overfitting to the observed shadows. It does not generalize to novel poses, as the cast shadows are global effects where movement of a joint could affect the appearance of other distant areas of the body. Figure 1 shows such setting. This is in contrast to local shading effects such as wrinkles in clothing which current body models can successfully reconstruct.", + "bbox": [ + 496, + 659, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20997", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To cast shadows within NeRF, secondary ray tracing from the reconstructed body model to the light source is an option. Although the predominant NeRF formulation enables casting secondary rays without change, it comes with a massive computational cost. For each sample along the primary ray, an equal number of secondary rays would have to be computed, each with multiple samples, leading to a quadratic, instead of linear, complexity in the number of samples per pixel. As a result, current re-lighting models only support diffuse reflection [16], hard shadows that do not generalize to novel poses [7], and soft dynamic shadow maps by approximate sphere tracing [43].", + "bbox": [ + 75, + 90, + 472, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our core contribution is introducing an additional volumetric density field that is approximate but significantly speeds up dynamic shadow casting while still maintaining differentiability and smoothness for gradient-based optimization. We introduce an anisotropic Gaussian density model and associated rendering functions that approximate the fine-grained density of the NeRF. The Gaussians have the beneficial property that we can integrate their density along a ray in closed form, thereby avoiding any sampling steps. Our derivation and implementation differs significantly from existing work using Gaussians for rendering. Compared to Gaussian Splatting [18, 31, 33], we neither require an affine approximation nor back-to-front ordering. Compared to Gaussian density models we alleviate their sampling [31] with an analytic integration and extend the existing analytic integration [30] to apply to anisotropic Gaussians. Notably, the Gaussian density is optimized alongside the NeRF without requiring a reference mesh such as SMPL [25]; it is template-free.", + "bbox": [ + 75, + 275, + 472, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To further reduce runtime, we use a deferred shading approach [7] in which the first rendering pass computes the albedo, depth, and normal for each pixel. The second pass casts only one secondary ray per pixel from the estimated surface point to the light source. This makes shadow computations independent of the number of samples in the NeRF, avoiding the mentioned quadratic complexity.", + "bbox": [ + 75, + 565, + 472, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our experiments with strong directional light and cast shadows demonstrate that our explicit lighting reduces the occurrence of artifacts in novel-view and novel-pose synthesis tasks. Figure 1 shows how our method is able to disentangle lighting and shadows from the avatar's albedo given sparse-view data from only a single illumination. We take advantage of the dynamic aspect of the data where we can observe the same body part in multiple illuminations as the subject moves. We further demonstrate the ability to optimize the unknown light directions without any user input or careful initialization. Moreover, relighting of the neural character enables us to composite recorded motions into novel scenes realistically, making them directly applicable in computer graphics and entertainment industries, as demonstrated by the HDRi re-lighting in Figure 1-right.", + "bbox": [ + 75, + 674, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 498, + 89, + 642, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We build on neural body models using NeRF [26], which we introduce briefly. The subsequent discussion focuses on relighting methods for 3D scenes and body models as well as how Gaussians are used in rendering and reconstruction. Neural avatars model dynamic performances by conditioning the neural rendering model on a template mesh driven by skeleton motion [4, 16, 20, 23, 24, 40, 49, 50] or template-free by linking neural fields directly to a skeleton [22, 27, 36-38]. Our implementation uses the more flexible template-free approach but it is general enough to extend to any NeRF-based model.", + "bbox": [ + 496, + 114, + 893, + 280 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Static NeRF scene relighting approaches can be categorized by either implicit [6, 9, 32, 34, 48] or explicit [13, 42] implementations. In implicit methods, the NeRF's MLP is extended to further output illumination data such as shadow, direct and indirect illumination or occlusion maps [9, 32, 34], or decompose the scene into material properties such as metallicity and roughness which can be used in a Bidirectional Reflectance Distribution Function (BRDF) lighting model [6, 48]. These extended MLPs are conditioned at training and test time on lighting information such as spherical harmonics coefficients [32], or light direction [9]. Implicit methods require large amounts of data in both multi-view and multiple illuminations with lighting information known or estimated [9, 32]. Explicit methods simulate how real light interacts with the environment which improves the generalizability to novel illuminations but are difficult to extend to dynamic scenes or objects. These methods either utilize a secondary data structure such as proxy geometries where lighting computations can be done using established methods [42], or attempt to cast the necessary secondary rays within the neural field's volume which comes with a significant computational burden [13].", + "bbox": [ + 496, + 281, + 893, + 613 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dynamic neural character relighting has been built on top of volume rendering methods [5, 7, 21, 29, 43, 44, 51] as well as 2D CNN based models [16]. Implicit methods again require large amounts of data which can only be captured using light stages with known illumination [5, 21, 44, 51] or, across multiple subjects for faces that are self-similar, each captured in a different setting with in-the-wild illumination [29]. Our model provides dynamic and explicit shading and is most closely related to the following three methods.", + "bbox": [ + 496, + 613, + 893, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "RANA [16] uses SMPL+D [3] to estimate the coarse geometry of a person and extract an albedo texture map using TextureNet [15]. Given a target pose, they render person-specific neural features alongside coarse albedo and normals from the SMPL-D model. These are passed through two CNNs to refine the albedo and normal maps. Finally, they generate a light map using spherical harmonics and the normal map which is multiplied by the albedo map to obtain the final lit image. While spherical harmonics allow a wide", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "footer", + "text": "20998", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "array of lighting conditions to be simulated, cast shadows are not present, e.g., an arm casting a shadow on the body. Our work implements a Gaussian density model [30, 31] to facilitate fast and efficient secondary ray tracing to compute these cast shadows.", + "bbox": [ + 75, + 90, + 468, + 165 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Likewise, Relighting4D [7] uses SMPL [25] to condition a 4D neural field of latent features which are trilinearly interpolated based on the nearby vertices to the query location. The latent features are passed through an MLP to obtain geometry, occlusion, and reflectance properties which are fed through a BRDF to get the final lit image. It is able to estimate the light probe, and at inference time, be able to switch the light probe to a new illumination. However, Relighting4D was not designed to work with hard shadows in novel poses, which is the focus point of our work.", + "bbox": [ + 75, + 166, + 467, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Finally, Xu et al. [43] utilize a signed distance field (SDF) based approach to learn a neural human avatar which utilizes SMPL-based inverse Linear Blender Skinning (LBS) and a displacement field to obtain canonical features. They utilize Hierarchical Distance Queries (HDQ) to compute minimum distances from world space to surface locations and perform sphere tracing to obtain material and surface properties for each camera ray. They further take advantage of HDQ through the SDF by computing soft visibility maps towards a learned light probe. While HDQ allows for fast occlusion checks, their solution focuses on soft approximate shadows whereas our work enables hard shadow casting.", + "bbox": [ + 75, + 316, + 467, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Gaussians have been used in rendering applications as differentiable methods for computing visibility and occlusions [30, 31, 35], as components of environment maps [46], or as a means to improve rendering efficiency for neural scenes [18]. Most methods are limited to spherical Gaussians [30, 31], while Gaussian Splatting uses an affine approximation that is only accurate when many small Gaussians are used [18], and Sridhar et al. use an approximation by perspective projection of ellipsoids [33]. Our work extends Rhodin et al. [30, 31] to use anisotropic Gaussians, without introducing any approximation, and tailors the analytic formulas and implementation towards shadow casting.", + "bbox": [ + 75, + 513, + 467, + 695 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 723, + 166, + 738 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our method reconstructs a neural character from a set of $N$ images of width $W$ and height $H$ , $\\{\\mathbf{I}_t \\in \\mathbb{R}^{H \\times W \\times 3}\\}_{t=1}^N$ , and corresponding character poses $\\theta_t \\in \\mathbb{R}^{J \\times 4 \\times 4}$ . The pose is represented as a skeleton with one $4 \\times 4$ local-to-world transformation matrix for each of the $J$ joints. Figure 2 gives an overview of our method. A key element of our design is a deferred illumination model [39] that separates the rendering into computing albedo, $\\mathbf{a} \\in \\mathbb{R}^3$ , surface normal, $\\hat{\\mathbf{n}} \\in \\mathbb{R}^3$ , and depth, $\\mathbf{d} \\in \\mathbb{R}$ , in a first pass and subsequently adding shading and shadow, $s \\in [0,1]$ , in a second", + "bbox": [ + 75, + 750, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "pass. Our key contribution is the closed form formula for the shadow $s$ .", + "bbox": [ + 498, + 90, + 890, + 119 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Deferred Neural Illumination", + "text_level": 1, + "bbox": [ + 500, + 130, + 761, + 145 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our volumetric body model is optimized on a reconstruction objective, $\\mathcal{L}_{\\mathrm{RGB}}$ that minimizes the squared difference between the input images $\\mathbf{I}_t$ and the rendering of the model. We test our method using DANBO [36]. It outputs a color and density for samples $\\mathbf{x}$ along the primary view rays. These are subsequently integrated to compute a color, which we interpret as the albedo, $\\mathbf{a}$ . The illuminated color for a given pixel of the reconstructed image, $\\hat{\\mathbf{c}}$ , is computed by a Lambertian reflectance model,", + "bbox": [ + 496, + 154, + 890, + 289 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {c}} = \\mathbf {a} \\left(\\theta_ {t}\\right) \\left(\\hat {\\mathbf {L}} _ {\\mathrm {a m b}} + s \\left(\\theta_ {t}\\right) \\mathbf {L} _ {\\mathrm {c o l}} \\left(\\hat {\\mathbf {L}} _ {\\mathrm {d i r}} \\cdot \\hat {\\mathbf {n}} \\left(\\theta_ {t}\\right)\\right)\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 300, + 890, + 325 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This diffuse shading model illuminates the entire body with an ambient light $\\hat{\\mathbf{L}}_{\\mathrm{amb}}$ and a directional light with color $\\mathbf{L}_{\\mathrm{col}}$ . The directional light intensity is attenuated by the cast shadows $s$ and the cosine angle between the surface normal $\\hat{\\mathbf{n}}$ and light direction $\\hat{\\mathbf{L}}_{\\mathrm{dir}}$ .", + "bbox": [ + 496, + 334, + 890, + 410 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Shading extensions. The benefit of the deferred rendering approach is that it lets us compute lighting information only once for each pixel, as opposed to at every sample location of the volumetric ray tracing leading to significantly faster computation. To be applicable, we extend DANBO to yield surface normals $\\hat{\\mathbf{n}}$ and depth $\\mathbf{d}$ for a given view ray. The former we attain by switching the density formulation to a signed distance function with an Eikonal loss. The normal is then readily estimated by differentiating the distance with respect to the original query location $\\mathbf{x}$ as in [45]. We compute $\\mathbf{d}$ likewise to albedo $\\mathbf{a}$ by integrating the sample's $\\mathbf{x}$ positions along the ray, weighted by their density and transmittance. Furthermore, we fix the intensity of the directional light to white with a magnitude of 1.5. Without fixing the directional light intensity, the equation would be over parametrized and lead to ambiguities. Our model is invariant to light intensities between 1.0 and 2.0, as shown by Table 1.", + "bbox": [ + 496, + 429, + 890, + 699 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/056e61285a12ebaa1aae4c951dab5e31b5d9956b4a2a3a61277ef7f269c4eff1.jpg", + "table_caption": [ + "Table 1. Light intensity modulation: Our method is invariant to moderate light intensity values." + ], + "table_footnote": [], + "table_body": "
Novel View & Pose\nLight Intensity:1.01.251.502.05.0
PSNR ↑24.4624.5224.3724.0018.98
LPIPS ↓0.1650.1640.1650.1630.185
", + "bbox": [ + 500, + 750, + 888, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Gaussian Shadow Casting", + "text_level": 1, + "bbox": [ + 500, + 832, + 735, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For modeling shadows more efficiently, we represent the body shape with a set of Gaussians rigidly attached to the skeleton model. The relative positions, orientation, and size", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20999", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4027d4362b1e9c7c3a80e2f7efc1ba7b34c31e537761b54f5a459d085e205ee9.jpg", + "image_caption": [ + "Figure 2. Method Overview. Our method takes as input images and poses of a person. Using a neural radiance field as a backbone $[36]^1$ , density, normals, and albedo values are volumetrically reconstructed and rendered. We fit a sum of 3D anisotropic Gaussian density model to approximate the neural density field and compute shadow maps using our novel anisotropic Gaussian ray occlusion equations. The shadow map is combined with a diffuse shading pass to produce the lit image. The whole model is optimized with a photometric loss against the training images. Our method is able to optimize the light direction and ambient intensity without any initialization. It also separates albedo from shading and shadow, allowing us to relight the model." + ], + "image_footnote": [], + "bbox": [ + 163, + 97, + 803, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of the Gaussians are optimized to approximate the density of the neural field and to allow for a fast, efficient, and closed-form solution for integration along a ray (occlusion checking). Our model extends previous work [30,31] by using anisotropic Gaussians (variable scale and rotation along each axis) and avoids the need for sampling during integration as in NeRF.", + "bbox": [ + 75, + 478, + 472, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Anisotropic Gaussian body model. We define the anisotropic Gaussian density model as the matrix $\\mathbf{G} \\in \\mathbb{R}^{J \\times K \\times 13}$ , with $K$ being the number of Gaussians per joint, typically $\\sim 8$ , and the columns representing the 3D mean $(\\mu_x, \\mu_y, \\mu_z)$ , the axis aligned standard deviations $(\\sigma_x, \\sigma_y, \\sigma_z)$ , the rotation defined using the 6 DOF representation $(R_{0,0}, R_{0,1}, R_{0,2}, R_{1,0}, R_{1,1}, R_{1,2})$ [52], and density $(\\mathcal{C})$ . Figure 3 gives examples with varying numbers of Gaussians.", + "bbox": [ + 75, + 603, + 468, + 736 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The 3D density function, $\\mathbf{G}(\\mathbf{x})$ , defines the density of the Gaussian model at the query location $\\mathbf{x}$ in world space. We define the density function of a single 3D anisotropic Gaussian as", + "bbox": [ + 75, + 738, + 470, + 799 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {G} _ {i} (\\mathbf {x}) = \\mathcal {C} \\exp \\left[ - 0. 5 \\left((\\mu - \\mathbf {x}) ^ {T}\\right) \\Sigma^ {- 1} (\\mu - \\mathbf {x}) \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 810, + 468, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the precision matrix $\\Sigma^{-1} = \\mathbf{R}^T\\mathbf{D}\\mathbf{R}$ and $\\mathbf{R}$ is the rotation matrix computed from the 6 DOF representation and $\\mathbf{D} = \\mathrm{diagonal}(1 / \\sigma_x^2, 1 / \\sigma_y^2, 1 / \\sigma_z^2)$ .", + "bbox": [ + 75, + 838, + 468, + 886 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The density of the entire Gaussian model is the sum of", + "bbox": [ + 96, + 886, + 470, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ec00476944913dcd817d00f3d41bc2cd0145cd64498b0fb6a23acc5bcd2a10c4.jpg", + "image_caption": [ + "Figure 3. Gaussian Density Model. The approximation to the NeRF's density field using a sum of 3D anisotropic Gaussians using: a) 2 Gaussians per bone, b) 4 Gaussian per bone, and c) 8 Gaussian per bone; d) is the groundtruth mesh. Note: ellipses are scaled to 2.5 STD of the Gaussians ( $99^{th}$ percentile)" + ], + "image_footnote": [], + "bbox": [ + 506, + 477, + 589, + 575 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5a3d4933d5d1871320d25c4edbbc4f4867e9c3b859c730b2cb978dc4b2b22a6e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 602, + 478, + 684, + 575 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/aeb845d270390441afd941d99b6903e97e9558800b639219a6caafb5cbbae736.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 479, + 782, + 575 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/dc24cc99a507b56267afb010ba396393b97ff4687ae1a4e044d2c5864c0ccee6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 794, + 478, + 875, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the density of each. The query location is transformed to the local space of the given Gaussian's joint $j$ at time-step $t$ using the world-to-local transformation matrix $\\theta_{t,j}^{-1}$ , rigidly attaching the Gaussians to the underlying skeleton and facilitating animation,", + "bbox": [ + 496, + 685, + 893, + 762 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {G} (\\mathbf {x}) = \\sum_ {i = 0} ^ {J \\times K} \\mathbf {G} _ {i} \\left(\\theta_ {t, j} ^ {- 1} \\mathbf {x}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 772, + 890, + 815 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We jointly fit the parameters of the Gaussian density model to the neural field by minimizing $\\mathcal{L}_{\\mathrm{gDensity}}$ , the L2 error between the density function $\\mathbf{G}(\\mathbf{x})$ and the target neural density field at query location $\\mathbf{x}$ . We detach the gradients of the neural density field to optimize the Gaussians, fitting the", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21000", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7f25f62062fb066ffb4a0b34fd80565f928d512efda55cac0198afe4b0e9dff6.jpg", + "image_caption": [ + "Figure 4. 3D Anisotropic Gaussian Raytracing. a) A cross-section of a 3D anisotropic Gaussian with rays passing through the Gaussian. b) The computed 1D Gaussians resulting from our derivation in Section 3.2 (colored solid), compared to sampling the 3D Gaussian directly (dashed), with their exact match validating the correctness. c) The transmittance along each ray which is used as the shadow map value." + ], + "image_footnote": [], + "bbox": [ + 83, + 89, + 460, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Gaussians to the neural field and not the other way around.", + "bbox": [ + 76, + 327, + 465, + 342 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Gaussian Ray Tracing. Figure 4 shows how casting a ray, $r$ , with ray origin $\\mathbf{r}_o \\in \\mathbb{R}^{3 \\times 1}$ and direction $\\mathbf{r}_d \\in \\mathbb{R}^{3 \\times 1}$ , through a 3D anisotropic Gaussian results in a 1D Gaussian density along the ray. Through the Gaussian body model, this equates to a sum of 1D Gaussians for which analytic integrals can be computed. The amount of occlusion these rays experience is equal to the sum of the integrals of each of the 1D Gaussians across the rays. The transmittance value, $\\mathcal{T}$ , used as the shadow map value, $s$ , is the exponential of the negative integral from the start of the ray, $t = 0$ to the length of the ray, $t = l$ ,", + "bbox": [ + 75, + 362, + 468, + 529 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns = \\mathcal {T} _ {r} = \\exp \\left[ - \\sum_ {i = 0} ^ {J \\times K} \\int_ {0} ^ {l} \\mathbf {G} _ {i} ^ {r} \\right]. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 541, + 468, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$G_{i}^{r}$ is the 1D Gaussian created by the ray, $r$ , going through the 3D anisotropic Gaussian, $G_{i}$ with mean $\\mu \\in \\mathbb{R}^{3\\times 1}$ and precision matrix $\\Sigma^{-1}\\in \\mathbb{R}^{3\\times 3}$ . We derive in the supplemental how the 1D Gaussian's density function takes the form", + "bbox": [ + 75, + 593, + 468, + 654 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nG _ {i} ^ {r ^ {s}} = \\bar {\\mathcal {C}} \\cdot \\exp \\left(- \\frac {(\\bar {\\mu} - x) ^ {2}}{2 \\bar {\\sigma} ^ {2}}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 666, + 468, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 76, + 705, + 122, + 719 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\bar {\\mathcal {C}} = \\mathcal {C} \\exp \\left(- 0. 5 \\left(\\left(\\mu - \\mathbf {r} _ {o}\\right) ^ {T} \\Sigma^ {- 1} (\\mu - \\mathbf {r} _ {o}) - \\frac {\\bar {\\mu} ^ {2}}{\\bar {\\sigma} ^ {2}}\\right)\\right), \\\\ \\bar {\\mu} = \\frac {\\mathbf {r} _ {d} ^ {T} \\Sigma^ {- 1} (\\boldsymbol {\\mu} - \\mathbf {r} _ {o})}{\\mathbf {r} _ {d} ^ {T} \\Sigma^ {- 1} \\mathbf {r} _ {d}}, \\text {a n d} \\\\ \\bar {\\sigma} = \\sqrt {\\frac {1}{\\mathbf {r} _ {d} ^ {T} \\Sigma^ {- 1} \\mathbf {r} _ {d}}}. \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 729, + 468, + 844 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This formula is more complex than in [30], as it now accounts for anisotropic Gaussians with an arbitrary covariance instead of isotropic Gaussians. The comparison to", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "sampling the 3D Gaussian in Figure 4 validates their correctness. It also lets us compute the cumulative density function (CDF) analytically, thereby avoiding the sampling in classical NeRFs,", + "bbox": [ + 496, + 90, + 890, + 150 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\int_ {0} ^ {x} \\mathbf {G} _ {i} ^ {r ^ {s}} = \\bar {\\mathcal {C}} \\cdot 0. 5 \\cdot \\left(1 + \\operatorname {e r f} \\left(\\frac {x - \\bar {\\mu}}{\\bar {\\sigma} \\sqrt {2}}\\right)\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 156, + 890, + 190 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Together with Equation 4, this integral computes the shadow $s$ when applied to the secondary ray with origin $r_o^s$ , as the point on the subject's surface computed from the depth map $\\mathbf{d}$ , and direction $r_d^s$ towards the light.", + "bbox": [ + 496, + 195, + 890, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Optimization", + "text_level": 1, + "bbox": [ + 500, + 263, + 638, + 279 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition to the introduced reconstruction loss $\\mathcal{L}_{\\mathrm{RGB}}$ , $\\mathcal{L}_{\\mathrm{Eikonal}}$ for SDF regularization as in [12], and Gaussian fitting $\\mathcal{L}_{\\mathrm{gDensity}}$ , we regularize the training with i) a $\\mathcal{L}_{\\mathrm{mask}} = |\\hat{\\rho} - \\rho|$ that regularizes density by minimizing the difference between integrated accumulation, $\\hat{\\rho}$ , and the foreground mask, $\\rho$ , ii) $\\mathcal{L}_{\\mathrm{amb}} = ||\\hat{\\mathbf{L}}_{\\mathrm{amb}} - 0.1||^2$ preferring small ambient light values, iii) $\\mathcal{L}_{\\mathrm{gSigma}}$ that prevent too large or small Gaussians, and iv) $\\mathcal{L}_{\\mathrm{gMean}}$ that pulls Gaussians closer to the center of the bones.", + "bbox": [ + 496, + 287, + 890, + 421 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training proceeds in three stages. In stage I, the reconstruction loss is replaced with one that encourages predicting gray inside the silhouette, to learn a rough body shape without illumination effects. In stage II, $\\mathcal{L}_{\\mathrm{gDensity}}$ and its regularizers are introduced, allowing the Gaussian density model to fit. Finally, in stage III, the $\\mathcal{L}_{\\mathrm{RGB}}$ takes over to optimize the light direction and learn the albedo. Additional training details are provided in the supplemental.", + "bbox": [ + 496, + 422, + 890, + 544 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Results", + "text_level": 1, + "bbox": [ + 500, + 556, + 586, + 570 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our method on synthetic sequences, as done in prior work [16]. However, this does not test performance in real world conditions. Hence, we captured a new dataset in direct sunlight and compare to the most closely related baselines, showing significantly improved relightable body models. The supplemental video and supplemental document provide additional qualitative comparisons, including relighting with HDRi environment maps.", + "bbox": [ + 496, + 580, + 890, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Synthetic datasets. We test our model on the RANA dataset [16], and further create our own synthetic sequence by obtaining a textured mesh of a subject with a 3D full body scanner (VITUS 3D Body Scanner). A Blender [8] cloth simulation was applied to a shirt over the scan and the character was automatically rigged and animated using Mixamo [2]. We use the 'swing-dance' animation as the driving motion as it contains a variety of poses from all body angles. Four (3 train, 1 test) cameras are placed around the subject at 90 degrees from each other. A directional light source illuminates the scene with a slight ambient contribution such that the shadowed areas were not fully black. The", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21001", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "annimations are rendered using the Cycles render engine. In addition, ground-truth pose and segmentation masks are exported. We split the dataset into 600/114 images for the train/test sets, with the test set including 57 novel view images and 57 novel poses, out of which 15 have a strong hard cast shadows that we test separately.", + "bbox": [ + 75, + 90, + 472, + 183 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Outdoor sunlight dataset. We recorded two sequences of real human motion in an outdoor scene during a sunny day. This setting has largely been unexplored in neural body models and few publicly available datasets are available. We capture the data using 3 cameras (Canon EOS R8, Canon EOS 70D, iPhone12) and obtain SMPL estimates using EasyMocap [1,10,11]. Segmentation masks are obtained using the Segment Anything Model (SAM) [19]. We divide the frames into 600/200 images for the train/test splits, using all three cameras for training.", + "bbox": [ + 75, + 203, + 472, + 354 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines We evaluate our method using the hard illumination dataset against Relighting4D [7]. Due to code being unavailable, we were not able to compare against RANA [16] and Xu et al. [43] using our datasets. We instead quantitatively compare albedo estimates between our model and RANA on their dataset as these results were kindly made available. We further compare our work with other template-less neural body models [36, 37], highlighting the drawback when not explicitly modeling lighting. DANBO [36] is our neural field backbone. NPC [37] forms the current state-of-the-art template-less neural character model.", + "bbox": [ + 75, + 376, + 472, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Albedo Estimates (training poses)", + "text_level": 1, + "bbox": [ + 76, + 554, + 372, + 570 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We ran the official implementation of Relighting4D (R4D) [7] on our outdoor dataset, providing the same segmentation masks and SMPL body model as to our method (our method only uses the skeleton, not the surface). As the first stage of R4D is NeuralBody [28] which does not take shadowing into account, it produces dark floaters in the space to approximate the hard shadow, hindering their subsequent relighting module from estimating shadow and shading correctly as seen in Figure 5a.", + "bbox": [ + 75, + 577, + 468, + 714 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To compare against RANA [16], we run our method on subject 1 of their synthetic dataset. Even with the dataset being monocular (light comes from the same direction relative to the camera), our method was able to accurately estimate the light direction (error of 9.9 degrees) and obtain albedo estimates with fewer lighting artifacts compared to RANA, see the back of the left leg in Figure 5b.", + "bbox": [ + 75, + 714, + 470, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Novel-Pose Rendering with Shadows", + "text_level": 1, + "bbox": [ + 76, + 830, + 395, + 847 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this setting, the camera and illumination are unchanged and only novel-poses are tested. These poses created new shadow casts that resulted in large appearance", + "bbox": [ + 75, + 854, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/75c81702be88068639ccb1ec8aeee8d3f81014ec416e6bc25774cda6896999e4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 89, + 893, + 189 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/90d273132a8b3b9cd7032421bcf11795e193ea41b3191fb353453e890f1a07b0.jpg", + "image_caption": [ + "Figure 5. Albedo Estimation. Our method can better separate shadows and lighting from training images to obtain better albedo estimates without lighting artifacts compared to a) Relighting4D [7] and b) RANA [16]." + ], + "image_footnote": [], + "bbox": [ + 504, + 196, + 893, + 297 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "changes distant to the changed body part. As expected, existing methods (NPC, DANBO) overfit the training poses and the shadows created in the novel poses are highly inaccurate. Figure 6 shows how for frames that had body parts casting shadows on other regions, our method produced more accurate shadows. Table 2 quantifies the gains across novel poses and Table 3 across the subset of the novel poses that has a shadow cast across the body.", + "bbox": [ + 496, + 388, + 890, + 511 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Novel-pose synthesis (all test frames). Our Gaussian Shadow Casting model achieves consistently better PSNR scores for novel pose renderings as it properly models the hard shadows cast by the limbs in novel positions. Existing methods only shine in perceptual metrics (SSIM and LPIPS) as these normalize contrast and hence lessen the impact of proper shadows and shading.", + "bbox": [ + 496, + 522, + 890, + 606 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6c161d61c7ca150f7aaaf5b48272356794af5200c132d67c8452506f7fcb8bfd.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Synthetic (N=57)Real S1 (N=200)Real S2 (N=200)Average
PSNR↑SSIM↑LPIPS↓PSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPS
DANBO [36]17.520.7560.19516.570.5990.32817.690.5880.32517.260.6480.283
NPC [37]17.570.7580.18816.330.5900.33417.470.5750.32817.120.6410.283
Ours22.040.8290.16617.570.5920.35618.290.5770.35119.300.6660.291
", + "bbox": [ + 504, + 617, + 887, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Novel-pose synthesis (subset of test set with observed self-casting shadows). Our Gaussian Shadow Casting renders novel poses with strong hard shadows well. Our scores drop marginally on these hard frames compared to all frames in Tab. 2, while the baselines drop significantly.", + "bbox": [ + 496, + 694, + 890, + 763 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7a0a511a089fb45a76deca6b5e7130cfbf162e642cb396c9288b9b6c9a308b76.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Synthetic (N=15)Real S1 (N=41)Real S2 (N=36)Average
PSNR↑SSIM↑LPIPS↓PSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPS
DANBO [36]17.780.7400.20915.110.5590.35416.550.5470.35316.480.6150.305
NPC [37]17.810.7410.20114.880.5530.35716.510.5380.35516.400.6110.304
Ours22.130.8210.17516.880.5720.36517.400.5440.37118.810.6460.303
", + "bbox": [ + 503, + 775, + 887, + 825 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the synthetic sequence, improvements were consistent across all three metrics. In the real outdoor sequence, all methods attain a lower quality as the cameras are spaced further apart and the segmentation masks and 3D input", + "bbox": [ + 496, + 840, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21002", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/24849be7be555d52da5bd2c38741ac76a5b83e6c3124912506f4f23e0509bdbb.jpg", + "image_caption": [ + "Figure 6. Novel Pose Rendering. Our method can more accurately reproduce the shadow in novel poses compared to the baselines." + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 245, + 289 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/88056b478e35e01ec23792f69602136773c892245f7f28bd4614a5f193e433b3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 261, + 90, + 411, + 289 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b2490e5bcf56804ccf59a682b76ed502ff65daea80d2e837381bec0ecf3bac95.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 426, + 90, + 575, + 289 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bdf72406ac60669a6f155216bbdaa5f2e019d213e5055dfc26b4d0ddb52dcf99.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 576, + 90, + 723, + 289 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/de89f37aa0141bdba367edb70c5ac3df779793a01ad914e77ea3db8f1b11ac46.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 741, + 90, + 890, + 289 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "pose, estimated with off-the-shelf 2D pose detection and lifting methods, are less reliable. Nevertheless, Figure 6 shows how our model can accurately optimize the light direction and predict realistic shadows, including on the ground. To map shadows to the ground, we estimate the ground plane from the reconstructed foot positions and cast the Gaussian shadow on it by modulating the static background with the ground shadow map.", + "bbox": [ + 75, + 340, + 468, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2 and Table 3 show that our method consistently improves the PSNR while the perceptual metrics SSIM [41] improves only in one and the baselines perform better for LPIPS [47]. This lower performance in perceptual metrics is expected because these metrics normalize for brightness and contrast differences, thereby lessening the importance of producing proper shading and shadowing. In addition, the texture and geometry detail of our method is slightly lower, which we attribute to the separation into shading and albedo imposing additional constraints, thereby leading to slightly less detailed reconstructions.", + "bbox": [ + 75, + 462, + 468, + 628 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Render Time Comparison", + "text_level": 1, + "bbox": [ + 76, + 635, + 313, + 651 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4 lists the render time of our baseline compared to our full model. Casting shadows with GSC has minimal overhead (0.3s for one ray, only $2\\%$ of the entire render time), enabling efficient training alongside NeRF optimization. Casting a shadow with the NeRF baseline requires processing twice the number of samples by the NeRF. The deferred shading model creates one occlusion ray and each of these secondary rays requires a similar number of samples as for the primary ray. Already with a single light source, this increases runtime by $25\\%$ , a ten-fold difference to GSC.", + "bbox": [ + 75, + 657, + 468, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Relighting with Environment Maps", + "text_level": 1, + "bbox": [ + 76, + 832, + 383, + 848 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The shadow computation not only benefits training time but also enables computing irradiance maps for environment maps. Figure 1 shows relighting with two different", + "bbox": [ + 76, + 854, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c87845d60654d43ca76b222d41600c9ddf5aab3903e3804c5dcd404cd238bd75.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Methodrender time [s]
DANBO + DS17.13
DANBO + DS + GSC17.47
DANBO + DS + NeRFSC21.4
DANBO + DS + GSC-HDRi-820.70
DANBO + DS + GSC-HDRi-1623.57
DANBO + DS + GSC-HDRi-3229.49
DANBO + DS + GSC-HDRi-6441.22
", + "bbox": [ + 553, + 338, + 839, + 441 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Render time. The overhead of Gaussian Shadow Casting (GSC) is minimal on DANBO with diffuse shading (DANBO + DS) and enables casting many rays (64 for GSC-HDRi-64). By contrast, NeRF shadow casting (NeRFSC) doubles the runtime with every light source, making training prohibitively slow and HDRi relighting impractical.", + "bbox": [ + 498, + 453, + 890, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "HDRi maps (obtained from Poly Haven [14]) by casting 64 secondary light rays towards the environment map for each pixel through importance sampling. In both cases, the bright sun casts a strong shadow while the colored light from the environment leads to natural shading that matches the character with the environment. This enables placing the reconstructed characters into new environments and giving them a natural and consistent look with respect to the rest of the scene while still containing cast shadows.", + "bbox": [ + 496, + 564, + 890, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our method is plug-and-play with other neural body models due to the deferred rendering approach. It can be used with higher quality volumetric neural models without degraded quality when training on uniformly lit data and using GSC for relighting as shown in Figure 7.", + "bbox": [ + 496, + 700, + 890, + 776 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 786, + 653, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We test a variety of implementation details in our model, including using only diffuse shading on top of DANBO (DANBO + Diff. Shading), only using the Gaussians to cast shadows (Ours w/o Diff. Shading), providing ground truth lighting (Ours-GT Light), and detaching the normals prior to the diffuse shading (Ours-Detached Normals). The", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21003", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1a86ed186dc1e803ed89d0289d0915b6d8875edc7602814e767993a323d6905b.jpg", + "image_caption": [ + "Figure 7. HDRi Relighting on MonoPerfCap using NPC [37] as the backbone. Our Gaussian relighting method can be utilized at inference with higher quality models on data that is uniformly lit." + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 472, + 194 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "results of which can be seen in Table 5, which shows that each of our contributions improves reconstruction quality at test time.", + "bbox": [ + 75, + 289, + 470, + 333 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/933e2e4ba7842ac96de6765ee76daec877f152d87c6c649709612a9750d20590.jpg", + "table_caption": [ + "Table 5. Ablation on Synthetic Sequence." + ], + "table_footnote": [], + "table_body": "
TrainingNovel PoseNovel View
PSNR↑SSIM↑LPIPS↓PSNRSSIMLPIPSPSNRSSIMLPIPS
DANBO27.660.9130.13317.520.7560.19518.850.7730.179
DANBO + Diff. Shading24.820.8600.18718.600.7850.20824.400.8750.167
Ours w/o Diff. Shading27.170.8790.16121.220.8020.17825.080.8670.163
Ours-Detached Normals25.900.8630.18220.430.8140.19226.180.8930.158
Ours-GT Light25.220.8610.18021.230.8300.18426.680.8950.155
Ours26.600.8760.16522.300.8270.17627.320.8820.154
", + "bbox": [ + 81, + 369, + 464, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Diffuse Shading. Shadowing alone does not account for accurate shading based on how incident the light hits the surface. Moreover, the Gaussians cast long-range shadows, but their smooth and low-resolution approximation to the NeRF's density prevents them from representing finer details such as small extremities (nose, fingers). As a result, finer shading details are missed as seen in Figure 8c. On the other hand, using only shading (DANBO + Diff. Shading) already reduces texture detail quality as seen by the training scores, but achieves improved performance in novel pose and novel view conditions. However, the missing cast shadows play the most significant role in improving test scores.", + "bbox": [ + 75, + 478, + 468, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Detached Normals. We compare results between a model where network gradients could, Figure 8d, and could not, Figure 8b, backpropagate through the surface normals used in the diffuse shading to see whether or not artifacts in the shading would smoothen out the geometry. We find that the surface is indeed affected by the gradients backpropagating through the diffuse computation and observe a smoother geometry reconstruction.", + "bbox": [ + 75, + 659, + 468, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "GT Light. Our method is able to fit the direction of the light source and the ambient intensity with little user input. We observe accurate light recovery when the light is initialized randomly, e.g. when coming from the back the angle error is only 1.36 degrees on our synthetic sequence. We found providing the ground truth light direction did not improve results and may have hindered the model due to the added constraints, Figure 8a.", + "bbox": [ + 75, + 779, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/426c13ca53715a663c15447c85486e59497cd5ded30a493cddc06fd8fa6c54a5.jpg", + "image_caption": [ + "Figure 8. Ablation Comparisons (novel-pose). a) The model trained with the groundtruth light direction. b) The model trained while detaching the gradients from the surface normals during diffuse shading. c) The model trained without diffuse shading. d) Our full model." + ], + "image_footnote": [], + "bbox": [ + 516, + 89, + 656, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/23dd5afffd1a5736a3f73440504cf635069e0fda9f66e42e5f7ed190d0d36f30.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 676, + 88, + 815, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e95aa5d9484d0a0707f1b75fec64992648d5ec513cedc55683e10772d1f0e3d2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 823, + 88, + 892, + 191 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Limitations", + "text_level": 1, + "bbox": [ + 500, + 309, + 625, + 323 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The Gaussian cast shadows model long-range effects, such as the arm casting a shadow on the leg but the smooth Gaussians lack high frequency details. This is a minor drawback since the diffuse shading already faithfully reproduces the light intensity fall-off as the light direction becomes more incident with the surface and therefore shades the back side of small extremities (i.e. nose and fingers) well. A future extension could be to integrate mid-scale effects with screen-space ambient occlusion and shading.", + "bbox": [ + 496, + 334, + 890, + 469 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Moreover, we noticed that disentangling color into shading and albedo, compared to the original DANBO backbone, leads to slightly lower image reconstruction metrics when shading effects are minimal. We attribute this to the additional constraints that are imposed on the model. However, the overall performance in novel light conditions is still improved significantly by our model.", + "bbox": [ + 496, + 472, + 890, + 579 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 598, + 617, + 614 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We enabled the 3D reconstruction of human motions in uncontrolled environments with a Gaussian-based shadow model that applies to dynamic scenes and is differentiable for iterative refinement. The reconstructed characters support reposing and relighting in novel environments. They are equipped with global shadow computation, diffuse shading, geometric reconstruction, and a consistent albedo, much like hand-crafted computer graphics models would provide. The deferred lighting approach allows our method to be combined with other neural body models with efficient shadow computations.", + "bbox": [ + 496, + 626, + 890, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 813, + 686, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by an NSERC Discovery Grant and the computational resources and services provided by Advanced Research Computing at The University of British Columbia.", + "bbox": [ + 496, + 839, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21004", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Easymocap - make human motion capture easier. https://github.com/zju3dv/EasyMocap, 2021.6", + "[2] Adobe. Mixamo. www MIXamo.com, 2023.5", + "[3] Thiemo Alldieck, Marcus Magnor, Bharat Lal Bhatnagar, Christian Theobalt, and Gerard Pons-Moll. Learning to reconstruct people in clothing from a single RGB camera. In CVPR, 2019. 2", + "[4] Timur Bagautdinov, Chenglei Wu, Tomas Simon, Fabian Prada, Takaaki Shiratori, Shih-En Wei, Weipeng Xu, Yaser Sheikh, and Jason Saragih. Driving-signal aware full-body avatars. ACM TOG (Proc. SIGGRAPH), 2021. 2", + "[5] Sai Bi, Stephen Lombardi, Shunsuke Saito, Tomas Simon, Shih-En Wei, Kevyn Mcphail, Ravi Ramamoorthi, Yaser Sheikh, and Jason Saragih. Deep relightable appearance models for animatable faces. ACM TOG (Proc. SIGGRAPH), 2021. 2", + "[6] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P.A. Lensch. Nerd: Neural reflectance decomposition from image collections. In ICCV, 2021. 2", + "[7] Zhaoxi Chen and Ziwei Liu. Relighting4d: Neural relightable human from videos. In ECCV, 2022. 2, 3, 6", + "[8] Blender Online Community. Blender - a 3d modelling and rendering package. http://www.blender.org, 2023.5", + "[9] Dawa Derksen and Dario Izzo. Shadow neural radiance fields for multi-view satellite photogrammetry. In CVPR, 2021. 2", + "[10] Junting Dong, Qi Fang, Wen Jiang, Yurou Yang, Hujun Bao, and Xiaowei Zhou. Fast and robust multi-person 3d pose estimation and tracking from multiple views. In TPAMI, 2021. 6", + "[11] Junting Dong, Qing Shuai, Yuanqing Zhang, Xian Liu, Xiaowei Zhou, and Hujun Bao. Motion capture from internet videos. In ECCV, 2020. 6", + "[12] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In ICML, 2020. 5", + "[13] Michelle Guo, Alireza Fathi, Jiajun Wu, and Thomas Funkhouser. Object-centric neural scene rendering, 2020. 2", + "[14] Poly Haven. Poly haven. www.polyhaven.com/HDris, 2023.7", + "[15] Jingwei Huang, Haotian Zhang, Li Yi, Thomas Funkhouser, Matthias Nießner, and Leonidas J Guibas. Texturenet: Consistent local parametrizations for learning from high-resolution signals on meshes. In CVPR, 2019. 2", + "[16] Umar Iqbal, Akin Caliskan, Koki Nagano, Sameh Khamis, Pavlo Molchanov, and Jan Kautz. Rana: Relightable articulated neural avatars. arXiv preprint arXiv:2212.03237, 2022. 2, 5, 6", + "[17] Wei Jiang, Kwang Moo Yi, Golnoosh Samei, Oncel Tuzel, and Anurag Ranjan. Neuman: Neural human radiance field from a single video. In ECCV, 2022. 1" + ], + "bbox": [ + 78, + 116, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM TOG (Proc. SIGGRAPH), 2023. 2, 3", + "[19] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 6", + "[20] Youngjoong Kwon, Dahun Kim, Duygu Ceylan, and Henry Fuchs. Neural human performer: Learning generalizable radiance fields for human performance rendering. NeurIPS, 2021. 1, 2", + "[21] Junxuan Li, Shunsuke Saito, Tomas Simon, Stephen Lombardi, Hongdong Li, and Jason Saragih. Megane: Morphable eyeglass and avatar network. In CVPR, 2023. 2", + "[22] Ruilong Li, Julian Tanke, Minh Vo, Michael Zollhofer, Jurgen Gall, Angjoo Kanazawa, and Christoph Lassner. Tava: Template-free animatable volumetric actors. In ECCV, 2022. 2", + "[23] Zhe Li, Zerong Zheng, Yuxiao Liu, Boyao Zhou, and Yebin Liu. Posevocab: Learning joint-structured pose embeddings for human avatar modeling. In ACM TOG (Proc. SIGGRAPH), 2023. 2", + "[24] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. ACM TOG (Proc. SIGGRAPH Asia), 2021. 2", + "[25] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. ACM TOG (Proc. SIGGRAPH Asia), 2015. 2, 3", + "[26] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2", + "[27] Atsuhiro Noguchi, Xiao Sun, Stephen Lin, and Tatsuya Harada. Neural articulated radiance field. In ICCV, 2021. 2", + "[28] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In CVPR, 2021. 1, 6", + "[29] Anurag Ranjan, Kwang Moo Yi, Jen-Hao Rick Chang, and Oncel Tuzel. Facelit: Neural 3d relightable faces. In CVPR, 2023. 2", + "[30] Helge Rhodin, Nadia Robertini, Dan Casas, Christian Richardt, Hans-Peter Seidel, and Christian Theobalt. General automatic human shape and motion capture using volumetric contour cues. In Bastian Leibe, Jiri Matas, Nicu Sebe, and Max Welling, editors, ECCV, 2016. 2, 3, 4, 5", + "[31] H. Rhodin, N. Robertini, C. Richardt, H. Seidel, and C. Theobalt. A versatile scene model with differentiable visibility applied to generative pose estimation. In ICCV, 2015. 2, 3, 4" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21005", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Viktor Rudnev, Mohamed Elgharib, William Smith, Lingjie Liu, Vladislav Golyanik, and Christian Theobalt. Nerf for outdoor scene relighting. In ECCV, 2022. 2", + "[33] Srinath Sridhar, Helge Rhodin, Hans-Peter Seidel, Antti Oulasvirta, and Christian Theobalt. Real-time hand tracking using a sum of anisotropic gaussians model. In 3DV, 2014. 2, 3", + "[34] Pratul P. Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T. Barron. Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In CVPR, 2021. 2", + "[35] Carsten Stoll, Nils Hasler, Juergen Gall, Hans-Peter Seidel, and Christian Theobalt. Fast articulated motion tracking using a sums of gaussians body model. ICCV, 2011. 3", + "[36] Shih-Yang Su, Timur Bagautdinov, and Helge Rhodin. Danbo: Disentangled articulated neural body representations via graph neural networks. In ECCV, 2022. 1, 2, 3, 4, 6", + "[37] Shih-Yang Su, Timur Bagautdinov, and Helge Rhodin. Npc: Neural point characters from video. In ICCV, 2023. 1, 2, 6, 8", + "[38] Shih-Yang Su, Frank Yu, Michael Zollhöfer, and Helge Rhodin. A-nerf: Articulated neural radiance fields for learning human shape, appearance, and pose. In NeurIPS, 2021. 2", + "[39] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: Image synthesis using neural textures. ACM TOG (Proc. SIGGRAPH), 2019. 3", + "[40] Shaofei Wang, Katja Schwarz, Andreas Geiger, and Siyu Tang. Arah: Animatable volume rendering of articulated human sdfs. In ECCV, 2022. 1, 2", + "[41] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 7", + "[42] Zian Wang, Tianchang Shen, Jun Gao, Shengyu Huang, Jacob Munkberg, Jon Hasselgren, Zan Gojcic, Wenzheng Chen, and Sanja Fidler. Neural fields meet explicit geometric representations for inverse rendering of urban scenes. In CVPR, 2023. 2", + "[43] Zhen Xu, Sida Peng, Chen Geng, Linzhan Mou, Zihan Yan, Jiaming Sun, Hujun Bao, and Xiaowei Zhou. Relightable and animatable neural avatar from sparse-view video, 2023. 2, 3, 6", + "[44] Haotian Yang, Mingwu Zheng, Wanquan Feng, Haibin Huang, Yu-Kun Lai, Pengfei Wan, Zhongyuan Wang, and Chongyang Ma. Towards practical capture of high-fidelity reconfigurable avatars. In ACM TOG (Proc. SIGGRAPH Asia), 2023. 2", + "[45] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. In NeurIPS, 2021. 3", + "[46] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snavely. PhySG: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In CVPR, 2021. 3", + "[47] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 7" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[48] Xiuming Zhang, Pratul P. Srinivasan, Boyang Deng, Paul Debevec, William T. Freeman, and Jonathan T. Barron. Nefactor: Neural factorization of shape and reflectance under an unknown illumination. ACM TOG (Proc. SIGGRAPH), 2021. 2", + "[49] Zerong Zheng, Han Huang, Tao Yu, Hongwen Zhang, Yandong Guo, and Yebin Liu. Structured local radiance fields for human avatar modeling. In CVPR, 2022. 2", + "[50] Zerong Zheng, Xiaochen Zhao, Hongwen Zhang, Boning Liu, and Yebin Liu. Avatarrex: Real-time expressive full-body avatars. ACM TOG (Proc. SIGGRAPH), 2023. 2", + "[51] Taotao Zhou, Kai He, Di Wu, Teng Xu, Qixuan Zhang, Kuixiang Shao, Wenzheng Chen, Lan Xu, and Jingyi Yu. Relightable neural human assets from multi-view gradient illuminations. In CVPR, 2023. 2", + "[52] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In CVPR, 2019. 4" + ], + "bbox": [ + 501, + 90, + 890, + 345 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21006", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Gaussian Shadow Casting for Neural Characters/e6fedb1c-caf9-484c-be44-7dbe074386f9_model.json b/2024/Gaussian Shadow Casting for Neural Characters/e6fedb1c-caf9-484c-be44-7dbe074386f9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ca76f94d1a6420b46f3bbda51c3d81565cce8c58 --- /dev/null +++ b/2024/Gaussian Shadow Casting for Neural Characters/e6fedb1c-caf9-484c-be44-7dbe074386f9_model.json @@ -0,0 +1,2200 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.239, + 0.131, + 0.733, + 0.154 + ], + "angle": 0, + "content": "Gaussian Shadow Casting for Neural Characters" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.18, + 0.725, + 0.219 + ], + "angle": 0, + "content": "Luis Bolanos1 Shih-Yang Su1 Helge Rhodin1,2 \n1The University of British Columbia 2Bielefeld University" + }, + { + "type": "image", + "bbox": [ + 0.129, + 0.257, + 0.23, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.258, + 0.265, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.268, + 0.258, + 0.359, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.258, + 0.424, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.434, + 0.266, + 0.575, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.584, + 0.266, + 0.665, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.679, + 0.258, + 0.841, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.429, + 0.895, + 0.486 + ], + "angle": 0, + "content": "Figure 1. Gaussian Shadow Casting (GSC): Our method is able to reconstruct 3D neural characters from a sparse set of videos in settings with strong directional illumination. GSC uses a sum of Gaussians density model to cast secondary shadow rays efficiently with an analytic formula. Our method learns to remove shadows from the neural color field, allowing us to relight in novel illuminations. All faces are blurred for anonymity." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.495, + 0.314, + 0.511 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.528, + 0.474, + 0.892 + ], + "angle": 0, + "content": "Neural character models can now reconstruct detailed geometry and texture from video, but they lack explicit shadows and shading, leading to artifacts when generating novel views and poses or during relighting. It is particularly difficult to include shadows as they are a global effect and the required casting of secondary rays is costly. We propose a new shadow model using a Gaussian density proxy that replaces sampling with a simple analytic formula. It supports dynamic motion and is tailored for shadow computation, thereby avoiding the affine projection approximation and sorting required by the closely related Gaussian splattering. Combined with a deferred neural rendering model, our Gaussian shadows enable Lambertian shading and shadow casting with minimal overhead. We demonstrate improved reconstructions, with better separation of albedo, shading, and shadows in challenging outdoor scenes with direct sun light and hard shadows. Our method is able to optimize the light direction without any input from the user. As a result, novel poses have fewer shadow artifacts, and relighting in novel scenes is more realistic compared to the state-of-the-art methods, providing new ways to pose neural characters in novel environments, increasing their applicability. Code available at: https://github.com/LuisBolanos17/GaussianShadowCasting" + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.528, + 0.633, + 0.544 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.659 + ], + "angle": 0, + "content": "It is now possible to reconstruct animatable 3D neural avatars from video but methods do not account for accurate lighting and shadows. They have to rely on recordings that have soft uniform lighting, which precludes recording outdoors in direct sun light and on film sets with spotlights, and most are unable to relight characters in novel environments, limiting their applicability in content creation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.893, + 0.901 + ], + "angle": 0, + "content": "The most recent body models [17,20,28,36,37,40] which are based on neural radiance fields (NeRFs) [26], approximate the light transport by casting primary rays between the camera and the scene, sampling the underlying neural network dozens of times along each ray to obtain the density and color. As they do not include an illumination model, the color that the NeRF learns includes lighting, shadow, and view-dependent effects. Learning a body model in a challenging scene with a strong directional light source, such as the sun, leads to the neural field overfitting to the observed shadows. It does not generalize to novel poses, as the cast shadows are global effects where movement of a joint could affect the appearance of other distant areas of the body. Figure 1 shows such setting. This is in contrast to local shading effects such as wrinkles in clothing which current body models can successfully reconstruct." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20997" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.273 + ], + "angle": 0, + "content": "To cast shadows within NeRF, secondary ray tracing from the reconstructed body model to the light source is an option. Although the predominant NeRF formulation enables casting secondary rays without change, it comes with a massive computational cost. For each sample along the primary ray, an equal number of secondary rays would have to be computed, each with multiple samples, leading to a quadratic, instead of linear, complexity in the number of samples per pixel. As a result, current re-lighting models only support diffuse reflection [16], hard shadows that do not generalize to novel poses [7], and soft dynamic shadow maps by approximate sphere tracing [43]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.276, + 0.473, + 0.563 + ], + "angle": 0, + "content": "Our core contribution is introducing an additional volumetric density field that is approximate but significantly speeds up dynamic shadow casting while still maintaining differentiability and smoothness for gradient-based optimization. We introduce an anisotropic Gaussian density model and associated rendering functions that approximate the fine-grained density of the NeRF. The Gaussians have the beneficial property that we can integrate their density along a ray in closed form, thereby avoiding any sampling steps. Our derivation and implementation differs significantly from existing work using Gaussians for rendering. Compared to Gaussian Splatting [18, 31, 33], we neither require an affine approximation nor back-to-front ordering. Compared to Gaussian density models we alleviate their sampling [31] with an analytic integration and extend the existing analytic integration [30] to apply to anisotropic Gaussians. Notably, the Gaussian density is optimized alongside the NeRF without requiring a reference mesh such as SMPL [25]; it is template-free." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.566, + 0.473, + 0.673 + ], + "angle": 0, + "content": "To further reduce runtime, we use a deferred shading approach [7] in which the first rendering pass computes the albedo, depth, and normal for each pixel. The second pass casts only one secondary ray per pixel from the estimated surface point to the light source. This makes shadow computations independent of the number of samples in the NeRF, avoiding the mentioned quadratic complexity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Our experiments with strong directional light and cast shadows demonstrate that our explicit lighting reduces the occurrence of artifacts in novel-view and novel-pose synthesis tasks. Figure 1 shows how our method is able to disentangle lighting and shadows from the avatar's albedo given sparse-view data from only a single illumination. We take advantage of the dynamic aspect of the data where we can observe the same body part in multiple illuminations as the subject moves. We further demonstrate the ability to optimize the unknown light directions without any user input or careful initialization. Moreover, relighting of the neural character enables us to composite recorded motions into novel scenes realistically, making them directly applicable in computer graphics and entertainment industries, as demonstrated by the HDRi re-lighting in Figure 1-right." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.09, + 0.643, + 0.107 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.895, + 0.281 + ], + "angle": 0, + "content": "We build on neural body models using NeRF [26], which we introduce briefly. The subsequent discussion focuses on relighting methods for 3D scenes and body models as well as how Gaussians are used in rendering and reconstruction. Neural avatars model dynamic performances by conditioning the neural rendering model on a template mesh driven by skeleton motion [4, 16, 20, 23, 24, 40, 49, 50] or template-free by linking neural fields directly to a skeleton [22, 27, 36-38]. Our implementation uses the more flexible template-free approach but it is general enough to extend to any NeRF-based model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.282, + 0.895, + 0.614 + ], + "angle": 0, + "content": "Static NeRF scene relighting approaches can be categorized by either implicit [6, 9, 32, 34, 48] or explicit [13, 42] implementations. In implicit methods, the NeRF's MLP is extended to further output illumination data such as shadow, direct and indirect illumination or occlusion maps [9, 32, 34], or decompose the scene into material properties such as metallicity and roughness which can be used in a Bidirectional Reflectance Distribution Function (BRDF) lighting model [6, 48]. These extended MLPs are conditioned at training and test time on lighting information such as spherical harmonics coefficients [32], or light direction [9]. Implicit methods require large amounts of data in both multi-view and multiple illuminations with lighting information known or estimated [9, 32]. Explicit methods simulate how real light interacts with the environment which improves the generalizability to novel illuminations but are difficult to extend to dynamic scenes or objects. These methods either utilize a secondary data structure such as proxy geometries where lighting computations can be done using established methods [42], or attempt to cast the necessary secondary rays within the neural field's volume which comes with a significant computational burden [13]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.895, + 0.763 + ], + "angle": 0, + "content": "Dynamic neural character relighting has been built on top of volume rendering methods [5, 7, 21, 29, 43, 44, 51] as well as 2D CNN based models [16]. Implicit methods again require large amounts of data which can only be captured using light stages with known illumination [5, 21, 44, 51] or, across multiple subjects for faces that are self-similar, each captured in a different setting with in-the-wild illumination [29]. Our model provides dynamic and explicit shading and is most closely related to the following three methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "RANA [16] uses SMPL+D [3] to estimate the coarse geometry of a person and extract an albedo texture map using TextureNet [15]. Given a target pose, they render person-specific neural features alongside coarse albedo and normals from the SMPL-D model. These are passed through two CNNs to refine the albedo and normal maps. Finally, they generate a light map using spherical harmonics and the normal map which is multiplied by the albedo map to obtain the final lit image. While spherical harmonics allow a wide" + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20998" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.166 + ], + "angle": 0, + "content": "array of lighting conditions to be simulated, cast shadows are not present, e.g., an arm casting a shadow on the body. Our work implements a Gaussian density model [30, 31] to facilitate fast and efficient secondary ray tracing to compute these cast shadows." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.167, + 0.468, + 0.317 + ], + "angle": 0, + "content": "Likewise, Relighting4D [7] uses SMPL [25] to condition a 4D neural field of latent features which are trilinearly interpolated based on the nearby vertices to the query location. The latent features are passed through an MLP to obtain geometry, occlusion, and reflectance properties which are fed through a BRDF to get the final lit image. It is able to estimate the light probe, and at inference time, be able to switch the light probe to a new illumination. However, Relighting4D was not designed to work with hard shadows in novel poses, which is the focus point of our work." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.318, + 0.468, + 0.514 + ], + "angle": 0, + "content": "Finally, Xu et al. [43] utilize a signed distance field (SDF) based approach to learn a neural human avatar which utilizes SMPL-based inverse Linear Blender Skinning (LBS) and a displacement field to obtain canonical features. They utilize Hierarchical Distance Queries (HDQ) to compute minimum distances from world space to surface locations and perform sphere tracing to obtain material and surface properties for each camera ray. They further take advantage of HDQ through the SDF by computing soft visibility maps towards a learned light probe. While HDQ allows for fast occlusion checks, their solution focuses on soft approximate shadows whereas our work enables hard shadow casting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.515, + 0.468, + 0.696 + ], + "angle": 0, + "content": "Gaussians have been used in rendering applications as differentiable methods for computing visibility and occlusions [30, 31, 35], as components of environment maps [46], or as a means to improve rendering efficiency for neural scenes [18]. Most methods are limited to spherical Gaussians [30, 31], while Gaussian Splatting uses an affine approximation that is only accurate when many small Gaussians are used [18], and Sridhar et al. use an approximation by perspective projection of ellipsoids [33]. Our work extends Rhodin et al. [30, 31] to use anisotropic Gaussians, without introducing any approximation, and tailors the analytic formulas and implementation towards shadow casting." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.724, + 0.167, + 0.739 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Our method reconstructs a neural character from a set of \\(N\\) images of width \\(W\\) and height \\(H\\), \\(\\{\\mathbf{I}_t \\in \\mathbb{R}^{H \\times W \\times 3}\\}_{t=1}^N\\), and corresponding character poses \\(\\theta_t \\in \\mathbb{R}^{J \\times 4 \\times 4}\\). The pose is represented as a skeleton with one \\(4 \\times 4\\) local-to-world transformation matrix for each of the \\(J\\) joints. Figure 2 gives an overview of our method. A key element of our design is a deferred illumination model [39] that separates the rendering into computing albedo, \\(\\mathbf{a} \\in \\mathbb{R}^3\\), surface normal, \\(\\hat{\\mathbf{n}} \\in \\mathbb{R}^3\\), and depth, \\(\\mathbf{d} \\in \\mathbb{R}\\), in a first pass and subsequently adding shading and shadow, \\(s \\in [0,1]\\), in a second" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "pass. Our key contribution is the closed form formula for the shadow \\(s\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.131, + 0.762, + 0.146 + ], + "angle": 0, + "content": "3.1. Deferred Neural Illumination" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.155, + 0.892, + 0.29 + ], + "angle": 0, + "content": "Our volumetric body model is optimized on a reconstruction objective, \\(\\mathcal{L}_{\\mathrm{RGB}}\\) that minimizes the squared difference between the input images \\(\\mathbf{I}_t\\) and the rendering of the model. We test our method using DANBO [36]. It outputs a color and density for samples \\(\\mathbf{x}\\) along the primary view rays. These are subsequently integrated to compute a color, which we interpret as the albedo, \\(\\mathbf{a}\\). The illuminated color for a given pixel of the reconstructed image, \\(\\hat{\\mathbf{c}}\\), is computed by a Lambertian reflectance model," + }, + { + "type": "equation", + "bbox": [ + 0.546, + 0.301, + 0.892, + 0.327 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {c}} = \\mathbf {a} \\left(\\theta_ {t}\\right) \\left(\\hat {\\mathbf {L}} _ {\\mathrm {a m b}} + s \\left(\\theta_ {t}\\right) \\mathbf {L} _ {\\mathrm {c o l}} \\left(\\hat {\\mathbf {L}} _ {\\mathrm {d i r}} \\cdot \\hat {\\mathbf {n}} \\left(\\theta_ {t}\\right)\\right)\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.335, + 0.892, + 0.411 + ], + "angle": 0, + "content": "This diffuse shading model illuminates the entire body with an ambient light \\(\\hat{\\mathbf{L}}_{\\mathrm{amb}}\\) and a directional light with color \\(\\mathbf{L}_{\\mathrm{col}}\\). The directional light intensity is attenuated by the cast shadows \\(s\\) and the cosine angle between the surface normal \\(\\hat{\\mathbf{n}}\\) and light direction \\(\\hat{\\mathbf{L}}_{\\mathrm{dir}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.43, + 0.892, + 0.7 + ], + "angle": 0, + "content": "Shading extensions. The benefit of the deferred rendering approach is that it lets us compute lighting information only once for each pixel, as opposed to at every sample location of the volumetric ray tracing leading to significantly faster computation. To be applicable, we extend DANBO to yield surface normals \\(\\hat{\\mathbf{n}}\\) and depth \\(\\mathbf{d}\\) for a given view ray. The former we attain by switching the density formulation to a signed distance function with an Eikonal loss. The normal is then readily estimated by differentiating the distance with respect to the original query location \\(\\mathbf{x}\\) as in [45]. We compute \\(\\mathbf{d}\\) likewise to albedo \\(\\mathbf{a}\\) by integrating the sample's \\(\\mathbf{x}\\) positions along the ray, weighted by their density and transmittance. Furthermore, we fix the intensity of the directional light to white with a magnitude of 1.5. Without fixing the directional light intensity, the equation would be over parametrized and lead to ambiguities. Our model is invariant to light intensities between 1.0 and 2.0, as shown by Table 1." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.711, + 0.892, + 0.74 + ], + "angle": 0, + "content": "Table 1. Light intensity modulation: Our method is invariant to moderate light intensity values." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.75, + 0.89, + 0.807 + ], + "angle": 0, + "content": "
Novel View & Pose\nLight Intensity:1.01.251.502.05.0
PSNR ↑24.4624.5224.3724.0018.98
LPIPS ↓0.1650.1640.1650.1630.185
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.833, + 0.736, + 0.849 + ], + "angle": 0, + "content": "3.2. Gaussian Shadow Casting" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "For modeling shadows more efficiently, we represent the body shape with a set of Gaussians rigidly attached to the skeleton model. The relative positions, orientation, and size" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20999" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.164, + 0.098, + 0.805, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.368, + 0.895, + 0.456 + ], + "angle": 0, + "content": "Figure 2. Method Overview. Our method takes as input images and poses of a person. Using a neural radiance field as a backbone \\([36]^1\\), density, normals, and albedo values are volumetrically reconstructed and rendered. We fit a sum of 3D anisotropic Gaussian density model to approximate the neural density field and compute shadow maps using our novel anisotropic Gaussian ray occlusion equations. The shadow map is combined with a diffuse shading pass to produce the lit image. The whole model is optimized with a photometric loss against the training images. Our method is able to optimize the light direction and ambient intensity without any initialization. It also separates albedo from shading and shadow, allowing us to relight the model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.479, + 0.473, + 0.587 + ], + "angle": 0, + "content": "of the Gaussians are optimized to approximate the density of the neural field and to allow for a fast, efficient, and closed-form solution for integration along a ray (occlusion checking). Our model extends previous work [30,31] by using anisotropic Gaussians (variable scale and rotation along each axis) and avoids the need for sampling during integration as in NeRF." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.604, + 0.47, + 0.737 + ], + "angle": 0, + "content": "Anisotropic Gaussian body model. We define the anisotropic Gaussian density model as the matrix \\(\\mathbf{G} \\in \\mathbb{R}^{J \\times K \\times 13}\\), with \\(K\\) being the number of Gaussians per joint, typically \\(\\sim 8\\), and the columns representing the 3D mean \\((\\mu_x, \\mu_y, \\mu_z)\\), the axis aligned standard deviations \\((\\sigma_x, \\sigma_y, \\sigma_z)\\), the rotation defined using the 6 DOF representation \\((R_{0,0}, R_{0,1}, R_{0,2}, R_{1,0}, R_{1,1}, R_{1,2})\\) [52], and density \\((\\mathcal{C})\\). Figure 3 gives examples with varying numbers of Gaussians." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.739, + 0.471, + 0.8 + ], + "angle": 0, + "content": "The 3D density function, \\(\\mathbf{G}(\\mathbf{x})\\), defines the density of the Gaussian model at the query location \\(\\mathbf{x}\\) in world space. We define the density function of a single 3D anisotropic Gaussian as" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.811, + 0.47, + 0.831 + ], + "angle": 0, + "content": "\\[\n\\mathbf {G} _ {i} (\\mathbf {x}) = \\mathcal {C} \\exp \\left[ - 0. 5 \\left((\\mu - \\mathbf {x}) ^ {T}\\right) \\Sigma^ {- 1} (\\mu - \\mathbf {x}) \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.839, + 0.47, + 0.887 + ], + "angle": 0, + "content": "where the precision matrix \\(\\Sigma^{-1} = \\mathbf{R}^T\\mathbf{D}\\mathbf{R}\\) and \\(\\mathbf{R}\\) is the rotation matrix computed from the 6 DOF representation and \\(\\mathbf{D} = \\mathrm{diagonal}(1 / \\sigma_x^2, 1 / \\sigma_y^2, 1 / \\sigma_z^2)\\)." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.887, + 0.471, + 0.901 + ], + "angle": 0, + "content": "The density of the entire Gaussian model is the sum of" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.478, + 0.59, + 0.577 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.604, + 0.479, + 0.685, + 0.576 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.48, + 0.783, + 0.576 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.795, + 0.479, + 0.877, + 0.576 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.589, + 0.895, + 0.66 + ], + "angle": 0, + "content": "Figure 3. Gaussian Density Model. The approximation to the NeRF's density field using a sum of 3D anisotropic Gaussians using: a) 2 Gaussians per bone, b) 4 Gaussian per bone, and c) 8 Gaussian per bone; d) is the groundtruth mesh. Note: ellipses are scaled to 2.5 STD of the Gaussians (\\(99^{th}\\) percentile)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.686, + 0.894, + 0.763 + ], + "angle": 0, + "content": "the density of each. The query location is transformed to the local space of the given Gaussian's joint \\(j\\) at time-step \\(t\\) using the world-to-local transformation matrix \\(\\theta_{t,j}^{-1}\\), rigidly attaching the Gaussians to the underlying skeleton and facilitating animation," + }, + { + "type": "equation", + "bbox": [ + 0.611, + 0.773, + 0.892, + 0.816 + ], + "angle": 0, + "content": "\\[\n\\mathbf {G} (\\mathbf {x}) = \\sum_ {i = 0} ^ {J \\times K} \\mathbf {G} _ {i} \\left(\\theta_ {t, j} ^ {- 1} \\mathbf {x}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We jointly fit the parameters of the Gaussian density model to the neural field by minimizing \\(\\mathcal{L}_{\\mathrm{gDensity}}\\), the L2 error between the density function \\(\\mathbf{G}(\\mathbf{x})\\) and the target neural density field at query location \\(\\mathbf{x}\\). We detach the gradients of the neural density field to optimize the Gaussians, fitting the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21000" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.09, + 0.462, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.202, + 0.473, + 0.3 + ], + "angle": 0, + "content": "Figure 4. 3D Anisotropic Gaussian Raytracing. a) A cross-section of a 3D anisotropic Gaussian with rays passing through the Gaussian. b) The computed 1D Gaussians resulting from our derivation in Section 3.2 (colored solid), compared to sampling the 3D Gaussian directly (dashed), with their exact match validating the correctness. c) The transmittance along each ray which is used as the shadow map value." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.328, + 0.466, + 0.343 + ], + "angle": 0, + "content": "Gaussians to the neural field and not the other way around." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.363, + 0.47, + 0.53 + ], + "angle": 0, + "content": "Gaussian Ray Tracing. Figure 4 shows how casting a ray, \\( r \\), with ray origin \\( \\mathbf{r}_o \\in \\mathbb{R}^{3 \\times 1} \\) and direction \\( \\mathbf{r}_d \\in \\mathbb{R}^{3 \\times 1} \\), through a 3D anisotropic Gaussian results in a 1D Gaussian density along the ray. Through the Gaussian body model, this equates to a sum of 1D Gaussians for which analytic integrals can be computed. The amount of occlusion these rays experience is equal to the sum of the integrals of each of the 1D Gaussians across the rays. The transmittance value, \\( \\mathcal{T} \\), used as the shadow map value, \\( s \\), is the exponential of the negative integral from the start of the ray, \\( t = 0 \\) to the length of the ray, \\( t = l \\)," + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.542, + 0.469, + 0.584 + ], + "angle": 0, + "content": "\\[\ns = \\mathcal {T} _ {r} = \\exp \\left[ - \\sum_ {i = 0} ^ {J \\times K} \\int_ {0} ^ {l} \\mathbf {G} _ {i} ^ {r} \\right]. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.594, + 0.47, + 0.655 + ], + "angle": 0, + "content": "\\(G_{i}^{r}\\) is the 1D Gaussian created by the ray, \\(r\\), going through the 3D anisotropic Gaussian, \\(G_{i}\\) with mean \\(\\mu \\in \\mathbb{R}^{3\\times 1}\\) and precision matrix \\(\\Sigma^{-1}\\in \\mathbb{R}^{3\\times 3}\\). We derive in the supplemental how the 1D Gaussian's density function takes the form" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.667, + 0.469, + 0.701 + ], + "angle": 0, + "content": "\\[\nG _ {i} ^ {r ^ {s}} = \\bar {\\mathcal {C}} \\cdot \\exp \\left(- \\frac {(\\bar {\\mu} - x) ^ {2}}{2 \\bar {\\sigma} ^ {2}}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.707, + 0.124, + 0.72 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.73, + 0.469, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\bar {\\mathcal {C}} = \\mathcal {C} \\exp \\left(- 0. 5 \\left(\\left(\\mu - \\mathbf {r} _ {o}\\right) ^ {T} \\Sigma^ {- 1} (\\mu - \\mathbf {r} _ {o}) - \\frac {\\bar {\\mu} ^ {2}}{\\bar {\\sigma} ^ {2}}\\right)\\right), \\\\ \\bar {\\mu} = \\frac {\\mathbf {r} _ {d} ^ {T} \\Sigma^ {- 1} (\\boldsymbol {\\mu} - \\mathbf {r} _ {o})}{\\mathbf {r} _ {d} ^ {T} \\Sigma^ {- 1} \\mathbf {r} _ {d}}, \\text {a n d} \\\\ \\bar {\\sigma} = \\sqrt {\\frac {1}{\\mathbf {r} _ {d} ^ {T} \\Sigma^ {- 1} \\mathbf {r} _ {d}}}. \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "This formula is more complex than in [30], as it now accounts for anisotropic Gaussians with an arbitrary covariance instead of isotropic Gaussians. The comparison to" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.151 + ], + "angle": 0, + "content": "sampling the 3D Gaussian in Figure 4 validates their correctness. It also lets us compute the cumulative density function (CDF) analytically, thereby avoiding the sampling in classical NeRFs," + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.157, + 0.892, + 0.191 + ], + "angle": 0, + "content": "\\[\n\\int_ {0} ^ {x} \\mathbf {G} _ {i} ^ {r ^ {s}} = \\bar {\\mathcal {C}} \\cdot 0. 5 \\cdot \\left(1 + \\operatorname {e r f} \\left(\\frac {x - \\bar {\\mu}}{\\bar {\\sigma} \\sqrt {2}}\\right)\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.196, + 0.892, + 0.258 + ], + "angle": 0, + "content": "Together with Equation 4, this integral computes the shadow \\( s \\) when applied to the secondary ray with origin \\( r_o^s \\), as the point on the subject's surface computed from the depth map \\( \\mathbf{d} \\), and direction \\( r_d^s \\) towards the light." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.265, + 0.64, + 0.28 + ], + "angle": 0, + "content": "3.3. Optimization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.288, + 0.892, + 0.422 + ], + "angle": 0, + "content": "In addition to the introduced reconstruction loss \\(\\mathcal{L}_{\\mathrm{RGB}}\\), \\(\\mathcal{L}_{\\mathrm{Eikonal}}\\) for SDF regularization as in [12], and Gaussian fitting \\(\\mathcal{L}_{\\mathrm{gDensity}}\\), we regularize the training with i) a \\(\\mathcal{L}_{\\mathrm{mask}} = |\\hat{\\rho} - \\rho|\\) that regularizes density by minimizing the difference between integrated accumulation, \\(\\hat{\\rho}\\), and the foreground mask, \\(\\rho\\), ii) \\(\\mathcal{L}_{\\mathrm{amb}} = ||\\hat{\\mathbf{L}}_{\\mathrm{amb}} - 0.1||^2\\) preferring small ambient light values, iii) \\(\\mathcal{L}_{\\mathrm{gSigma}}\\) that prevent too large or small Gaussians, and iv) \\(\\mathcal{L}_{\\mathrm{gMean}}\\) that pulls Gaussians closer to the center of the bones." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.424, + 0.892, + 0.545 + ], + "angle": 0, + "content": "Training proceeds in three stages. In stage I, the reconstruction loss is replaced with one that encourages predicting gray inside the silhouette, to learn a rough body shape without illumination effects. In stage II, \\(\\mathcal{L}_{\\mathrm{gDensity}}\\) and its regularizers are introduced, allowing the Gaussian density model to fit. Finally, in stage III, the \\(\\mathcal{L}_{\\mathrm{RGB}}\\) takes over to optimize the light direction and learn the albedo. Additional training details are provided in the supplemental." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.557, + 0.587, + 0.571 + ], + "angle": 0, + "content": "4. Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.892, + 0.703 + ], + "angle": 0, + "content": "We evaluate our method on synthetic sequences, as done in prior work [16]. However, this does not test performance in real world conditions. Hence, we captured a new dataset in direct sunlight and compare to the most closely related baselines, showing significantly improved relightable body models. The supplemental video and supplemental document provide additional qualitative comparisons, including relighting with HDRi environment maps." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Synthetic datasets. We test our model on the RANA dataset [16], and further create our own synthetic sequence by obtaining a textured mesh of a subject with a 3D full body scanner (VITUS 3D Body Scanner). A Blender [8] cloth simulation was applied to a shirt over the scan and the character was automatically rigged and animated using Mixamo [2]. We use the 'swing-dance' animation as the driving motion as it contains a variety of poses from all body angles. Four (3 train, 1 test) cameras are placed around the subject at 90 degrees from each other. A directional light source illuminates the scene with a slight ambient contribution such that the shadowed areas were not fully black. The" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21001" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.184 + ], + "angle": 0, + "content": "annimations are rendered using the Cycles render engine. In addition, ground-truth pose and segmentation masks are exported. We split the dataset into 600/114 images for the train/test sets, with the test set including 57 novel view images and 57 novel poses, out of which 15 have a strong hard cast shadows that we test separately." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.204, + 0.473, + 0.356 + ], + "angle": 0, + "content": "Outdoor sunlight dataset. We recorded two sequences of real human motion in an outdoor scene during a sunny day. This setting has largely been unexplored in neural body models and few publicly available datasets are available. We capture the data using 3 cameras (Canon EOS R8, Canon EOS 70D, iPhone12) and obtain SMPL estimates using EasyMocap [1,10,11]. Segmentation masks are obtained using the Segment Anything Model (SAM) [19]. We divide the frames into 600/200 images for the train/test splits, using all three cameras for training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.377, + 0.473, + 0.543 + ], + "angle": 0, + "content": "Baselines We evaluate our method using the hard illumination dataset against Relighting4D [7]. Due to code being unavailable, we were not able to compare against RANA [16] and Xu et al. [43] using our datasets. We instead quantitatively compare albedo estimates between our model and RANA on their dataset as these results were kindly made available. We further compare our work with other template-less neural body models [36, 37], highlighting the drawback when not explicitly modeling lighting. DANBO [36] is our neural field backbone. NPC [37] forms the current state-of-the-art template-less neural character model." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.555, + 0.373, + 0.571 + ], + "angle": 0, + "content": "4.1. Albedo Estimates (training poses)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.578, + 0.47, + 0.715 + ], + "angle": 0, + "content": "We ran the official implementation of Relighting4D (R4D) [7] on our outdoor dataset, providing the same segmentation masks and SMPL body model as to our method (our method only uses the skeleton, not the surface). As the first stage of R4D is NeuralBody [28] which does not take shadowing into account, it produces dark floaters in the space to approximate the hard shadow, hindering their subsequent relighting module from estimating shadow and shading correctly as seen in Figure 5a." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.715, + 0.471, + 0.822 + ], + "angle": 0, + "content": "To compare against RANA [16], we run our method on subject 1 of their synthetic dataset. Even with the dataset being monocular (light comes from the same direction relative to the camera), our method was able to accurately estimate the light direction (error of 9.9 degrees) and obtain albedo estimates with fewer lighting artifacts compared to RANA, see the back of the left leg in Figure 5b." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.396, + 0.848 + ], + "angle": 0, + "content": "4.2. Novel-Pose Rendering with Shadows" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In this setting, the camera and illumination are unchanged and only novel-poses are tested. These poses created new shadow casts that resulted in large appearance" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.09, + 0.895, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.197, + 0.895, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.308, + 0.895, + 0.366 + ], + "angle": 0, + "content": "Figure 5. Albedo Estimation. Our method can better separate shadows and lighting from training images to obtain better albedo estimates without lighting artifacts compared to a) Relighting4D [7] and b) RANA [16]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.39, + 0.892, + 0.512 + ], + "angle": 0, + "content": "changes distant to the changed body part. As expected, existing methods (NPC, DANBO) overfit the training poses and the shadows created in the novel poses are highly inaccurate. Figure 6 shows how for frames that had body parts casting shadows on other regions, our method produced more accurate shadows. Table 2 quantifies the gains across novel poses and Table 3 across the subset of the novel poses that has a shadow cast across the body." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.523, + 0.892, + 0.607 + ], + "angle": 0, + "content": "Table 2. Novel-pose synthesis (all test frames). Our Gaussian Shadow Casting model achieves consistently better PSNR scores for novel pose renderings as it properly models the hard shadows cast by the limbs in novel positions. Existing methods only shine in perceptual metrics (SSIM and LPIPS) as these normalize contrast and hence lessen the impact of proper shadows and shading." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.618, + 0.888, + 0.669 + ], + "angle": 0, + "content": "
Synthetic (N=57)Real S1 (N=200)Real S2 (N=200)Average
PSNR↑SSIM↑LPIPS↓PSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPS
DANBO [36]17.520.7560.19516.570.5990.32817.690.5880.32517.260.6480.283
NPC [37]17.570.7580.18816.330.5900.33417.470.5750.32817.120.6410.283
Ours22.040.8290.16617.570.5920.35618.290.5770.35119.300.6660.291
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.695, + 0.892, + 0.765 + ], + "angle": 0, + "content": "Table 3. Novel-pose synthesis (subset of test set with observed self-casting shadows). Our Gaussian Shadow Casting renders novel poses with strong hard shadows well. Our scores drop marginally on these hard frames compared to all frames in Tab. 2, while the baselines drop significantly." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.776, + 0.888, + 0.827 + ], + "angle": 0, + "content": "
Synthetic (N=15)Real S1 (N=41)Real S2 (N=36)Average
PSNR↑SSIM↑LPIPS↓PSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPS
DANBO [36]17.780.7400.20915.110.5590.35416.550.5470.35316.480.6150.305
NPC [37]17.810.7410.20114.880.5530.35716.510.5380.35516.400.6110.304
Ours22.130.8210.17516.880.5720.36517.400.5440.37118.810.6460.303
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.895, + 0.903 + ], + "angle": 0, + "content": "For the synthetic sequence, improvements were consistent across all three metrics. In the real outdoor sequence, all methods attain a lower quality as the cameras are spaced further apart and the segmentation masks and 3D input" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21002" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.092, + 0.246, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.262, + 0.092, + 0.412, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.427, + 0.092, + 0.576, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.578, + 0.092, + 0.725, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.742, + 0.092, + 0.892, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.095, + 0.3, + 0.873, + 0.315 + ], + "angle": 0, + "content": "Figure 6. Novel Pose Rendering. Our method can more accurately reproduce the shadow in novel poses compared to the baselines." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.341, + 0.47, + 0.462 + ], + "angle": 0, + "content": "pose, estimated with off-the-shelf 2D pose detection and lifting methods, are less reliable. Nevertheless, Figure 6 shows how our model can accurately optimize the light direction and predict realistic shadows, including on the ground. To map shadows to the ground, we estimate the ground plane from the reconstructed foot positions and cast the Gaussian shadow on it by modulating the static background with the ground shadow map." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.463, + 0.47, + 0.629 + ], + "angle": 0, + "content": "Table 2 and Table 3 show that our method consistently improves the PSNR while the perceptual metrics SSIM [41] improves only in one and the baselines perform better for LPIPS [47]. This lower performance in perceptual metrics is expected because these metrics normalize for brightness and contrast differences, thereby lessening the importance of producing proper shading and shadowing. In addition, the texture and geometry detail of our method is slightly lower, which we attribute to the separation into shading and albedo imposing additional constraints, thereby leading to slightly less detailed reconstructions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.636, + 0.315, + 0.652 + ], + "angle": 0, + "content": "4.3. Render Time Comparison" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.47, + 0.81 + ], + "angle": 0, + "content": "Table 4 lists the render time of our baseline compared to our full model. Casting shadows with GSC has minimal overhead (0.3s for one ray, only \\(2\\%\\) of the entire render time), enabling efficient training alongside NeRF optimization. Casting a shadow with the NeRF baseline requires processing twice the number of samples by the NeRF. The deferred shading model creates one occlusion ray and each of these secondary rays requires a similar number of samples as for the primary ray. Already with a single light source, this increases runtime by \\(25\\%\\), a ten-fold difference to GSC." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.833, + 0.385, + 0.849 + ], + "angle": 0, + "content": "4.4. Relighting with Environment Maps" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "The shadow computation not only benefits training time but also enables computing irradiance maps for environment maps. Figure 1 shows relighting with two different" + }, + { + "type": "table", + "bbox": [ + 0.554, + 0.339, + 0.84, + 0.443 + ], + "angle": 0, + "content": "
Methodrender time [s]
DANBO + DS17.13
DANBO + DS + GSC17.47
DANBO + DS + NeRFSC21.4
DANBO + DS + GSC-HDRi-820.70
DANBO + DS + GSC-HDRi-1623.57
DANBO + DS + GSC-HDRi-3229.49
DANBO + DS + GSC-HDRi-6441.22
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.454, + 0.892, + 0.538 + ], + "angle": 0, + "content": "Table 4. Render time. The overhead of Gaussian Shadow Casting (GSC) is minimal on DANBO with diffuse shading (DANBO + DS) and enables casting many rays (64 for GSC-HDRi-64). By contrast, NeRF shadow casting (NeRFSC) doubles the runtime with every light source, making training prohibitively slow and HDRi relighting impractical." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.565, + 0.892, + 0.701 + ], + "angle": 0, + "content": "HDRi maps (obtained from Poly Haven [14]) by casting 64 secondary light rays towards the environment map for each pixel through importance sampling. In both cases, the bright sun casts a strong shadow while the colored light from the environment leads to natural shading that matches the character with the environment. This enables placing the reconstructed characters into new environments and giving them a natural and consistent look with respect to the rest of the scene while still containing cast shadows." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.702, + 0.892, + 0.777 + ], + "angle": 0, + "content": "Our method is plug-and-play with other neural body models due to the deferred rendering approach. It can be used with higher quality volumetric neural models without degraded quality when training on uniformly lit data and using GSC for relighting as shown in Figure 7." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.654, + 0.803 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We test a variety of implementation details in our model, including using only diffuse shading on top of DANBO (DANBO + Diff. Shading), only using the Gaussians to cast shadows (Ours w/o Diff. Shading), providing ground truth lighting (Ours-GT Light), and detaching the normals prior to the diffuse shading (Ours-Detached Normals). The" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21003" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.09, + 0.473, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.206, + 0.471, + 0.261 + ], + "angle": 0, + "content": "Figure 7. HDRi Relighting on MonoPerfCap using NPC [37] as the backbone. Our Gaussian relighting method can be utilized at inference with higher quality models on data that is uniformly lit." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.29, + 0.471, + 0.334 + ], + "angle": 0, + "content": "results of which can be seen in Table 5, which shows that each of our contributions improves reconstruction quality at test time." + }, + { + "type": "table_caption", + "bbox": [ + 0.144, + 0.346, + 0.402, + 0.36 + ], + "angle": 0, + "content": "Table 5. Ablation on Synthetic Sequence." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.37, + 0.465, + 0.462 + ], + "angle": 0, + "content": "
TrainingNovel PoseNovel View
PSNR↑SSIM↑LPIPS↓PSNRSSIMLPIPSPSNRSSIMLPIPS
DANBO27.660.9130.13317.520.7560.19518.850.7730.179
DANBO + Diff. Shading24.820.8600.18718.600.7850.20824.400.8750.167
Ours w/o Diff. Shading27.170.8790.16121.220.8020.17825.080.8670.163
Ours-Detached Normals25.900.8630.18220.430.8140.19226.180.8930.158
Ours-GT Light25.220.8610.18021.230.8300.18426.680.8950.155
Ours26.600.8760.16522.300.8270.17627.320.8820.154
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.479, + 0.469, + 0.659 + ], + "angle": 0, + "content": "Diffuse Shading. Shadowing alone does not account for accurate shading based on how incident the light hits the surface. Moreover, the Gaussians cast long-range shadows, but their smooth and low-resolution approximation to the NeRF's density prevents them from representing finer details such as small extremities (nose, fingers). As a result, finer shading details are missed as seen in Figure 8c. On the other hand, using only shading (DANBO + Diff. Shading) already reduces texture detail quality as seen by the training scores, but achieves improved performance in novel pose and novel view conditions. However, the missing cast shadows play the most significant role in improving test scores." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.469, + 0.78 + ], + "angle": 0, + "content": "Detached Normals. We compare results between a model where network gradients could, Figure 8d, and could not, Figure 8b, backpropagate through the surface normals used in the diffuse shading to see whether or not artifacts in the shading would smoothen out the geometry. We find that the surface is indeed affected by the gradients backpropagating through the diffuse computation and observe a smoother geometry reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.469, + 0.902 + ], + "angle": 0, + "content": "GT Light. Our method is able to fit the direction of the light source and the ambient intensity with little user input. We observe accurate light recovery when the light is initialized randomly, e.g. when coming from the back the angle error is only 1.36 degrees on our synthetic sequence. We found providing the ground truth light direction did not improve results and may have hindered the model due to the added constraints, Figure 8a." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.09, + 0.657, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.677, + 0.089, + 0.816, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.824, + 0.089, + 0.893, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.209, + 0.893, + 0.278 + ], + "angle": 0, + "content": "Figure 8. Ablation Comparisons (novel-pose). a) The model trained with the groundtruth light direction. b) The model trained while detaching the gradients from the surface normals during diffuse shading. c) The model trained without diffuse shading. d) Our full model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.31, + 0.627, + 0.324 + ], + "angle": 0, + "content": "4.6. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.335, + 0.892, + 0.47 + ], + "angle": 0, + "content": "The Gaussian cast shadows model long-range effects, such as the arm casting a shadow on the leg but the smooth Gaussians lack high frequency details. This is a minor drawback since the diffuse shading already faithfully reproduces the light intensity fall-off as the light direction becomes more incident with the surface and therefore shades the back side of small extremities (i.e. nose and fingers) well. A future extension could be to integrate mid-scale effects with screen-space ambient occlusion and shading." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.473, + 0.892, + 0.58 + ], + "angle": 0, + "content": "Moreover, we noticed that disentangling color into shading and albedo, compared to the original DANBO backbone, leads to slightly lower image reconstruction metrics when shading effects are minimal. We attribute this to the additional constraints that are imposed on the model. However, the overall performance in novel light conditions is still improved significantly by our model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.599, + 0.619, + 0.615 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.627, + 0.892, + 0.793 + ], + "angle": 0, + "content": "We enabled the 3D reconstruction of human motions in uncontrolled environments with a Gaussian-based shadow model that applies to dynamic scenes and is differentiable for iterative refinement. The reconstructed characters support reposing and relighting in novel environments. They are equipped with global shadow computation, diffuse shading, geometric reconstruction, and a consistent albedo, much like hand-crafted computer graphics models would provide. The deferred lighting approach allows our method to be combined with other neural body models with efficient shadow computations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.814, + 0.687, + 0.831 + ], + "angle": 0, + "content": "6. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.9 + ], + "angle": 0, + "content": "This work was supported in part by an NSERC Discovery Grant and the computational resources and services provided by Advanced Research Computing at The University of British Columbia." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21004" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.117, + 0.466, + 0.144 + ], + "angle": 0, + "content": "[1] Easymocap - make human motion capture easier. https://github.com/zju3dv/EasyMocap, 2021.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.147, + 0.397, + 0.16 + ], + "angle": 0, + "content": "[2] Adobe. Mixamo. www MIXamo.com, 2023.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.162, + 0.469, + 0.217 + ], + "angle": 0, + "content": "[3] Thiemo Alldieck, Marcus Magnor, Bharat Lal Bhatnagar, Christian Theobalt, and Gerard Pons-Moll. Learning to reconstruct people in clothing from a single RGB camera. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.22, + 0.469, + 0.275 + ], + "angle": 0, + "content": "[4] Timur Bagautdinov, Chenglei Wu, Tomas Simon, Fabian Prada, Takaaki Shiratori, Shih-En Wei, Weipeng Xu, Yaser Sheikh, and Jason Saragih. Driving-signal aware full-body avatars. ACM TOG (Proc. SIGGRAPH), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.278, + 0.469, + 0.346 + ], + "angle": 0, + "content": "[5] Sai Bi, Stephen Lombardi, Shunsuke Saito, Tomas Simon, Shih-En Wei, Kevyn Mcphail, Ravi Ramamoorthi, Yaser Sheikh, and Jason Saragih. Deep relightable appearance models for animatable faces. ACM TOG (Proc. SIGGRAPH), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.349, + 0.469, + 0.404 + ], + "angle": 0, + "content": "[6] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P.A. Lensch. Nerd: Neural reflectance decomposition from image collections. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.407, + 0.469, + 0.435 + ], + "angle": 0, + "content": "[7] Zhaoxi Chen and Ziwei Liu. Relighting4d: Neural relightable human from videos. In ECCV, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.437, + 0.469, + 0.477 + ], + "angle": 0, + "content": "[8] Blender Online Community. Blender - a 3d modelling and rendering package. http://www.blender.org, 2023.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.469, + 0.521 + ], + "angle": 0, + "content": "[9] Dawa Derksen and Dario Izzo. Shadow neural radiance fields for multi-view satellite photogrammetry. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.469, + 0.579 + ], + "angle": 0, + "content": "[10] Junting Dong, Qi Fang, Wen Jiang, Yurou Yang, Hujun Bao, and Xiaowei Zhou. Fast and robust multi-person 3d pose estimation and tracking from multiple views. In TPAMI, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.582, + 0.469, + 0.623 + ], + "angle": 0, + "content": "[11] Junting Dong, Qing Shuai, Yuanqing Zhang, Xian Liu, Xiaowei Zhou, and Hujun Bao. Motion capture from internet videos. In ECCV, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.626, + 0.469, + 0.667 + ], + "angle": 0, + "content": "[12] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In ICML, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.67, + 0.469, + 0.71 + ], + "angle": 0, + "content": "[13] Michelle Guo, Alireza Fathi, Jiajun Wu, and Thomas Funkhouser. Object-centric neural scene rendering, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.714, + 0.469, + 0.74 + ], + "angle": 0, + "content": "[14] Poly Haven. Poly haven. www.polyhaven.com/HDris, 2023.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.744, + 0.469, + 0.799 + ], + "angle": 0, + "content": "[15] Jingwei Huang, Haotian Zhang, Li Yi, Thomas Funkhouser, Matthias Nießner, and Leonidas J Guibas. Texturenet: Consistent local parametrizations for learning from high-resolution signals on meshes. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.469, + 0.856 + ], + "angle": 0, + "content": "[16] Umar Iqbal, Akin Caliskan, Koki Nagano, Sameh Khamis, Pavlo Molchanov, and Jan Kautz. Rana: Relightable articulated neural avatars. arXiv preprint arXiv:2212.03237, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[17] Wei Jiang, Kwang Moo Yi, Golnoosh Samei, Oncel Tuzel, and Anurag Ranjan. Neuman: Neural human radiance field from a single video. In ECCV, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.117, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[18] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM TOG (Proc. SIGGRAPH), 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.151, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[19] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.222, + 0.892, + 0.276 + ], + "angle": 0, + "content": "[20] Youngjoong Kwon, Dahun Kim, Duygu Ceylan, and Henry Fuchs. Neural human performer: Learning generalizable radiance fields for human performance rendering. NeurIPS, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.28, + 0.892, + 0.322 + ], + "angle": 0, + "content": "[21] Junxuan Li, Shunsuke Saito, Tomas Simon, Stephen Lombardi, Hongdong Li, and Jason Saragih. Megane: Morphable eyeglass and avatar network. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.324, + 0.892, + 0.378 + ], + "angle": 0, + "content": "[22] Ruilong Li, Julian Tanke, Minh Vo, Michael Zollhofer, Jurgen Gall, Angjoo Kanazawa, and Christoph Lassner. Tava: Template-free animatable volumetric actors. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.382, + 0.892, + 0.437 + ], + "angle": 0, + "content": "[23] Zhe Li, Zerong Zheng, Yuxiao Liu, Boyao Zhou, and Yebin Liu. Posevocab: Learning joint-structured pose embeddings for human avatar modeling. In ACM TOG (Proc. SIGGRAPH), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.44, + 0.892, + 0.495 + ], + "angle": 0, + "content": "[24] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. ACM TOG (Proc. SIGGRAPH Asia), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.498, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[25] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. ACM TOG (Proc. SIGGRAPH Asia), 2015. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.556, + 0.892, + 0.61 + ], + "angle": 0, + "content": "[26] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.614, + 0.892, + 0.653 + ], + "angle": 0, + "content": "[27] Atsuhiro Noguchi, Xiao Sun, Stephen Lin, and Tatsuya Harada. Neural articulated radiance field. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.657, + 0.892, + 0.726 + ], + "angle": 0, + "content": "[28] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In CVPR, 2021. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.73, + 0.892, + 0.77 + ], + "angle": 0, + "content": "[29] Anurag Ranjan, Kwang Moo Yi, Jen-Hao Rick Chang, and Oncel Tuzel. Facelit: Neural 3d relightable faces. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[30] Helge Rhodin, Nadia Robertini, Dan Casas, Christian Richardt, Hans-Peter Seidel, and Christian Theobalt. General automatic human shape and motion capture using volumetric contour cues. In Bastian Leibe, Jiri Matas, Nicu Sebe, and Max Welling, editors, ECCV, 2016. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[31] H. Rhodin, N. Robertini, C. Richardt, H. Seidel, and C. Theobalt. A versatile scene model with differentiable visibility applied to generative pose estimation. In ICCV, 2015. 2, 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21005" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[32] Viktor Rudnev, Mohamed Elgharib, William Smith, Lingjie Liu, Vladislav Golyanik, and Christian Theobalt. Nerf for outdoor scene relighting. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.47, + 0.189 + ], + "angle": 0, + "content": "[33] Srinath Sridhar, Helge Rhodin, Hans-Peter Seidel, Antti Oulasvirta, and Christian Theobalt. Real-time hand tracking using a sum of anisotropic gaussians model. In 3DV, 2014. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.19, + 0.469, + 0.245 + ], + "angle": 0, + "content": "[34] Pratul P. Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T. Barron. Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.246, + 0.469, + 0.287 + ], + "angle": 0, + "content": "[35] Carsten Stoll, Nils Hasler, Juergen Gall, Hans-Peter Seidel, and Christian Theobalt. Fast articulated motion tracking using a sums of gaussians body model. ICCV, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.469, + 0.329 + ], + "angle": 0, + "content": "[36] Shih-Yang Su, Timur Bagautdinov, and Helge Rhodin. Danbo: Disentangled articulated neural body representations via graph neural networks. In ECCV, 2022. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.469, + 0.37 + ], + "angle": 0, + "content": "[37] Shih-Yang Su, Timur Bagautdinov, and Helge Rhodin. Npc: Neural point characters from video. In ICCV, 2023. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.371, + 0.469, + 0.425 + ], + "angle": 0, + "content": "[38] Shih-Yang Su, Frank Yu, Michael Zollhöfer, and Helge Rhodin. A-nerf: Articulated neural radiance fields for learning human shape, appearance, and pose. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.427, + 0.469, + 0.468 + ], + "angle": 0, + "content": "[39] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: Image synthesis using neural textures. ACM TOG (Proc. SIGGRAPH), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.469, + 0.51 + ], + "angle": 0, + "content": "[40] Shaofei Wang, Katja Schwarz, Andreas Geiger, and Siyu Tang. Arah: Animatable volume rendering of articulated human sdfs. In ECCV, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.469, + 0.566 + ], + "angle": 0, + "content": "[41] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.469, + 0.635 + ], + "angle": 0, + "content": "[42] Zian Wang, Tianchang Shen, Jun Gao, Shengyu Huang, Jacob Munkberg, Jon Hasselgren, Zan Gojcic, Wenzheng Chen, and Sanja Fidler. Neural fields meet explicit geometric representations for inverse rendering of urban scenes. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.636, + 0.469, + 0.69 + ], + "angle": 0, + "content": "[43] Zhen Xu, Sida Peng, Chen Geng, Linzhan Mou, Zihan Yan, Jiaming Sun, Hujun Bao, and Xiaowei Zhou. Relightable and animatable neural avatar from sparse-view video, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.469, + 0.759 + ], + "angle": 0, + "content": "[44] Haotian Yang, Mingwu Zheng, Wanquan Feng, Haibin Huang, Yu-Kun Lai, Pengfei Wan, Zhongyuan Wang, and Chongyang Ma. Towards practical capture of high-fidelity reconfigurable avatars. In ACM TOG (Proc. SIGGRAPH Asia), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.469, + 0.801 + ], + "angle": 0, + "content": "[45] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. In NeurIPS, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[46] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snavely. PhySG: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[47] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.16 + ], + "angle": 0, + "content": "[48] Xiuming Zhang, Pratul P. Srinivasan, Boyang Deng, Paul Debevec, William T. Freeman, and Jonathan T. Barron. Nefactor: Neural factorization of shape and reflectance under an unknown illumination. ACM TOG (Proc. SIGGRAPH), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[49] Zerong Zheng, Han Huang, Tao Yu, Hongwen Zhang, Yandong Guo, and Yebin Liu. Structured local radiance fields for human avatar modeling. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[50] Zerong Zheng, Xiaochen Zhao, Hongwen Zhang, Boning Liu, and Yebin Liu. Avatarrex: Real-time expressive full-body avatars. ACM TOG (Proc. SIGGRAPH), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[51] Taotao Zhou, Kai He, Di Wu, Teng Xu, Qixuan Zhang, Kuixiang Shao, Wenzheng Chen, Lan Xu, and Jingyi Yu. Relightable neural human assets from multi-view gradient illuminations. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[52] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In CVPR, 2019. 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21006" + } + ] +] \ No newline at end of file diff --git a/2024/Gaussian Shadow Casting for Neural Characters/e6fedb1c-caf9-484c-be44-7dbe074386f9_origin.pdf b/2024/Gaussian Shadow Casting for Neural Characters/e6fedb1c-caf9-484c-be44-7dbe074386f9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..32f00ddcfc5563d5b9d6c7204f5b998e752b5edb --- /dev/null +++ b/2024/Gaussian Shadow Casting for Neural Characters/e6fedb1c-caf9-484c-be44-7dbe074386f9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:21a77fad31c3673357c96494a0104143592b33468288d821a7de85ba91e54e01 +size 7063980 diff --git a/2024/Gaussian Shadow Casting for Neural Characters/full.md b/2024/Gaussian Shadow Casting for Neural Characters/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d5fe618f4e803b1d53c9a76523cb55ef224cc088 --- /dev/null +++ b/2024/Gaussian Shadow Casting for Neural Characters/full.md @@ -0,0 +1,323 @@ +# Gaussian Shadow Casting for Neural Characters + +Luis Bolanos1 Shih-Yang Su1 Helge Rhodin1,2 +1The University of British Columbia 2Bielefeld University + +![](images/42becdbf56c8af7d828f3e01f48ee9b2d7dc0f848629c4e27af18b2fd12753e8.jpg) +Figure 1. Gaussian Shadow Casting (GSC): Our method is able to reconstruct 3D neural characters from a sparse set of videos in settings with strong directional illumination. GSC uses a sum of Gaussians density model to cast secondary shadow rays efficiently with an analytic formula. Our method learns to remove shadows from the neural color field, allowing us to relight in novel illuminations. All faces are blurred for anonymity. + +![](images/05b584c53cf36fcb7953e597a75f4ee52831125e10766c03ce275d0a02a298fc.jpg) + +![](images/10eb7278dd5c34c0c3c1687853f86c26fa0bd19aee45a6becbdf9923e4351e99.jpg) + +![](images/64ec36f7e9cf948dc8ce3126a554a8a3b72806968bcc52775fb50f8d628f516f.jpg) + +![](images/1e4121215bc8cdefe69fffbce202c5aace9093792f72cdf5e4a5c1d44bb56129.jpg) + +![](images/4af957060dbe8eb456b75f5a43264a2d8bfef292361a076dab9275e40e70557f.jpg) + +![](images/1dd102a236a19906632678a8071a98b96fadb035081f1a26acf5fb6b95c1478d.jpg) + +# Abstract + +Neural character models can now reconstruct detailed geometry and texture from video, but they lack explicit shadows and shading, leading to artifacts when generating novel views and poses or during relighting. It is particularly difficult to include shadows as they are a global effect and the required casting of secondary rays is costly. We propose a new shadow model using a Gaussian density proxy that replaces sampling with a simple analytic formula. It supports dynamic motion and is tailored for shadow computation, thereby avoiding the affine projection approximation and sorting required by the closely related Gaussian splattering. Combined with a deferred neural rendering model, our Gaussian shadows enable Lambertian shading and shadow casting with minimal overhead. We demonstrate improved reconstructions, with better separation of albedo, shading, and shadows in challenging outdoor scenes with direct sun light and hard shadows. Our method is able to optimize the light direction without any input from the user. As a result, novel poses have fewer shadow artifacts, and relighting in novel scenes is more realistic compared to the state-of-the-art methods, providing new ways to pose neural characters in novel environments, increasing their applicability. Code available at: https://github.com/LuisBolanos17/GaussianShadowCasting + +# 1. Introduction + +It is now possible to reconstruct animatable 3D neural avatars from video but methods do not account for accurate lighting and shadows. They have to rely on recordings that have soft uniform lighting, which precludes recording outdoors in direct sun light and on film sets with spotlights, and most are unable to relight characters in novel environments, limiting their applicability in content creation. + +The most recent body models [17,20,28,36,37,40] which are based on neural radiance fields (NeRFs) [26], approximate the light transport by casting primary rays between the camera and the scene, sampling the underlying neural network dozens of times along each ray to obtain the density and color. As they do not include an illumination model, the color that the NeRF learns includes lighting, shadow, and view-dependent effects. Learning a body model in a challenging scene with a strong directional light source, such as the sun, leads to the neural field overfitting to the observed shadows. It does not generalize to novel poses, as the cast shadows are global effects where movement of a joint could affect the appearance of other distant areas of the body. Figure 1 shows such setting. This is in contrast to local shading effects such as wrinkles in clothing which current body models can successfully reconstruct. + +To cast shadows within NeRF, secondary ray tracing from the reconstructed body model to the light source is an option. Although the predominant NeRF formulation enables casting secondary rays without change, it comes with a massive computational cost. For each sample along the primary ray, an equal number of secondary rays would have to be computed, each with multiple samples, leading to a quadratic, instead of linear, complexity in the number of samples per pixel. As a result, current re-lighting models only support diffuse reflection [16], hard shadows that do not generalize to novel poses [7], and soft dynamic shadow maps by approximate sphere tracing [43]. + +Our core contribution is introducing an additional volumetric density field that is approximate but significantly speeds up dynamic shadow casting while still maintaining differentiability and smoothness for gradient-based optimization. We introduce an anisotropic Gaussian density model and associated rendering functions that approximate the fine-grained density of the NeRF. The Gaussians have the beneficial property that we can integrate their density along a ray in closed form, thereby avoiding any sampling steps. Our derivation and implementation differs significantly from existing work using Gaussians for rendering. Compared to Gaussian Splatting [18, 31, 33], we neither require an affine approximation nor back-to-front ordering. Compared to Gaussian density models we alleviate their sampling [31] with an analytic integration and extend the existing analytic integration [30] to apply to anisotropic Gaussians. Notably, the Gaussian density is optimized alongside the NeRF without requiring a reference mesh such as SMPL [25]; it is template-free. + +To further reduce runtime, we use a deferred shading approach [7] in which the first rendering pass computes the albedo, depth, and normal for each pixel. The second pass casts only one secondary ray per pixel from the estimated surface point to the light source. This makes shadow computations independent of the number of samples in the NeRF, avoiding the mentioned quadratic complexity. + +Our experiments with strong directional light and cast shadows demonstrate that our explicit lighting reduces the occurrence of artifacts in novel-view and novel-pose synthesis tasks. Figure 1 shows how our method is able to disentangle lighting and shadows from the avatar's albedo given sparse-view data from only a single illumination. We take advantage of the dynamic aspect of the data where we can observe the same body part in multiple illuminations as the subject moves. We further demonstrate the ability to optimize the unknown light directions without any user input or careful initialization. Moreover, relighting of the neural character enables us to composite recorded motions into novel scenes realistically, making them directly applicable in computer graphics and entertainment industries, as demonstrated by the HDRi re-lighting in Figure 1-right. + +# 2. Related Work + +We build on neural body models using NeRF [26], which we introduce briefly. The subsequent discussion focuses on relighting methods for 3D scenes and body models as well as how Gaussians are used in rendering and reconstruction. Neural avatars model dynamic performances by conditioning the neural rendering model on a template mesh driven by skeleton motion [4, 16, 20, 23, 24, 40, 49, 50] or template-free by linking neural fields directly to a skeleton [22, 27, 36-38]. Our implementation uses the more flexible template-free approach but it is general enough to extend to any NeRF-based model. + +Static NeRF scene relighting approaches can be categorized by either implicit [6, 9, 32, 34, 48] or explicit [13, 42] implementations. In implicit methods, the NeRF's MLP is extended to further output illumination data such as shadow, direct and indirect illumination or occlusion maps [9, 32, 34], or decompose the scene into material properties such as metallicity and roughness which can be used in a Bidirectional Reflectance Distribution Function (BRDF) lighting model [6, 48]. These extended MLPs are conditioned at training and test time on lighting information such as spherical harmonics coefficients [32], or light direction [9]. Implicit methods require large amounts of data in both multi-view and multiple illuminations with lighting information known or estimated [9, 32]. Explicit methods simulate how real light interacts with the environment which improves the generalizability to novel illuminations but are difficult to extend to dynamic scenes or objects. These methods either utilize a secondary data structure such as proxy geometries where lighting computations can be done using established methods [42], or attempt to cast the necessary secondary rays within the neural field's volume which comes with a significant computational burden [13]. + +Dynamic neural character relighting has been built on top of volume rendering methods [5, 7, 21, 29, 43, 44, 51] as well as 2D CNN based models [16]. Implicit methods again require large amounts of data which can only be captured using light stages with known illumination [5, 21, 44, 51] or, across multiple subjects for faces that are self-similar, each captured in a different setting with in-the-wild illumination [29]. Our model provides dynamic and explicit shading and is most closely related to the following three methods. + +RANA [16] uses SMPL+D [3] to estimate the coarse geometry of a person and extract an albedo texture map using TextureNet [15]. Given a target pose, they render person-specific neural features alongside coarse albedo and normals from the SMPL-D model. These are passed through two CNNs to refine the albedo and normal maps. Finally, they generate a light map using spherical harmonics and the normal map which is multiplied by the albedo map to obtain the final lit image. While spherical harmonics allow a wide + +array of lighting conditions to be simulated, cast shadows are not present, e.g., an arm casting a shadow on the body. Our work implements a Gaussian density model [30, 31] to facilitate fast and efficient secondary ray tracing to compute these cast shadows. + +Likewise, Relighting4D [7] uses SMPL [25] to condition a 4D neural field of latent features which are trilinearly interpolated based on the nearby vertices to the query location. The latent features are passed through an MLP to obtain geometry, occlusion, and reflectance properties which are fed through a BRDF to get the final lit image. It is able to estimate the light probe, and at inference time, be able to switch the light probe to a new illumination. However, Relighting4D was not designed to work with hard shadows in novel poses, which is the focus point of our work. + +Finally, Xu et al. [43] utilize a signed distance field (SDF) based approach to learn a neural human avatar which utilizes SMPL-based inverse Linear Blender Skinning (LBS) and a displacement field to obtain canonical features. They utilize Hierarchical Distance Queries (HDQ) to compute minimum distances from world space to surface locations and perform sphere tracing to obtain material and surface properties for each camera ray. They further take advantage of HDQ through the SDF by computing soft visibility maps towards a learned light probe. While HDQ allows for fast occlusion checks, their solution focuses on soft approximate shadows whereas our work enables hard shadow casting. + +Gaussians have been used in rendering applications as differentiable methods for computing visibility and occlusions [30, 31, 35], as components of environment maps [46], or as a means to improve rendering efficiency for neural scenes [18]. Most methods are limited to spherical Gaussians [30, 31], while Gaussian Splatting uses an affine approximation that is only accurate when many small Gaussians are used [18], and Sridhar et al. use an approximation by perspective projection of ellipsoids [33]. Our work extends Rhodin et al. [30, 31] to use anisotropic Gaussians, without introducing any approximation, and tailors the analytic formulas and implementation towards shadow casting. + +# 3. Method + +Our method reconstructs a neural character from a set of $N$ images of width $W$ and height $H$ , $\{\mathbf{I}_t \in \mathbb{R}^{H \times W \times 3}\}_{t=1}^N$ , and corresponding character poses $\theta_t \in \mathbb{R}^{J \times 4 \times 4}$ . The pose is represented as a skeleton with one $4 \times 4$ local-to-world transformation matrix for each of the $J$ joints. Figure 2 gives an overview of our method. A key element of our design is a deferred illumination model [39] that separates the rendering into computing albedo, $\mathbf{a} \in \mathbb{R}^3$ , surface normal, $\hat{\mathbf{n}} \in \mathbb{R}^3$ , and depth, $\mathbf{d} \in \mathbb{R}$ , in a first pass and subsequently adding shading and shadow, $s \in [0,1]$ , in a second + +pass. Our key contribution is the closed form formula for the shadow $s$ . + +# 3.1. Deferred Neural Illumination + +Our volumetric body model is optimized on a reconstruction objective, $\mathcal{L}_{\mathrm{RGB}}$ that minimizes the squared difference between the input images $\mathbf{I}_t$ and the rendering of the model. We test our method using DANBO [36]. It outputs a color and density for samples $\mathbf{x}$ along the primary view rays. These are subsequently integrated to compute a color, which we interpret as the albedo, $\mathbf{a}$ . The illuminated color for a given pixel of the reconstructed image, $\hat{\mathbf{c}}$ , is computed by a Lambertian reflectance model, + +$$ +\hat {\mathbf {c}} = \mathbf {a} \left(\theta_ {t}\right) \left(\hat {\mathbf {L}} _ {\mathrm {a m b}} + s \left(\theta_ {t}\right) \mathbf {L} _ {\mathrm {c o l}} \left(\hat {\mathbf {L}} _ {\mathrm {d i r}} \cdot \hat {\mathbf {n}} \left(\theta_ {t}\right)\right)\right). \tag {1} +$$ + +This diffuse shading model illuminates the entire body with an ambient light $\hat{\mathbf{L}}_{\mathrm{amb}}$ and a directional light with color $\mathbf{L}_{\mathrm{col}}$ . The directional light intensity is attenuated by the cast shadows $s$ and the cosine angle between the surface normal $\hat{\mathbf{n}}$ and light direction $\hat{\mathbf{L}}_{\mathrm{dir}}$ . + +Shading extensions. The benefit of the deferred rendering approach is that it lets us compute lighting information only once for each pixel, as opposed to at every sample location of the volumetric ray tracing leading to significantly faster computation. To be applicable, we extend DANBO to yield surface normals $\hat{\mathbf{n}}$ and depth $\mathbf{d}$ for a given view ray. The former we attain by switching the density formulation to a signed distance function with an Eikonal loss. The normal is then readily estimated by differentiating the distance with respect to the original query location $\mathbf{x}$ as in [45]. We compute $\mathbf{d}$ likewise to albedo $\mathbf{a}$ by integrating the sample's $\mathbf{x}$ positions along the ray, weighted by their density and transmittance. Furthermore, we fix the intensity of the directional light to white with a magnitude of 1.5. Without fixing the directional light intensity, the equation would be over parametrized and lead to ambiguities. Our model is invariant to light intensities between 1.0 and 2.0, as shown by Table 1. + +Table 1. Light intensity modulation: Our method is invariant to moderate light intensity values. + +
Novel View & Pose +Light Intensity:1.01.251.502.05.0
PSNR ↑24.4624.5224.3724.0018.98
LPIPS ↓0.1650.1640.1650.1630.185
+ +# 3.2. Gaussian Shadow Casting + +For modeling shadows more efficiently, we represent the body shape with a set of Gaussians rigidly attached to the skeleton model. The relative positions, orientation, and size + +![](images/4027d4362b1e9c7c3a80e2f7efc1ba7b34c31e537761b54f5a459d085e205ee9.jpg) +Figure 2. Method Overview. Our method takes as input images and poses of a person. Using a neural radiance field as a backbone $[36]^1$ , density, normals, and albedo values are volumetrically reconstructed and rendered. We fit a sum of 3D anisotropic Gaussian density model to approximate the neural density field and compute shadow maps using our novel anisotropic Gaussian ray occlusion equations. The shadow map is combined with a diffuse shading pass to produce the lit image. The whole model is optimized with a photometric loss against the training images. Our method is able to optimize the light direction and ambient intensity without any initialization. It also separates albedo from shading and shadow, allowing us to relight the model. + +of the Gaussians are optimized to approximate the density of the neural field and to allow for a fast, efficient, and closed-form solution for integration along a ray (occlusion checking). Our model extends previous work [30,31] by using anisotropic Gaussians (variable scale and rotation along each axis) and avoids the need for sampling during integration as in NeRF. + +Anisotropic Gaussian body model. We define the anisotropic Gaussian density model as the matrix $\mathbf{G} \in \mathbb{R}^{J \times K \times 13}$ , with $K$ being the number of Gaussians per joint, typically $\sim 8$ , and the columns representing the 3D mean $(\mu_x, \mu_y, \mu_z)$ , the axis aligned standard deviations $(\sigma_x, \sigma_y, \sigma_z)$ , the rotation defined using the 6 DOF representation $(R_{0,0}, R_{0,1}, R_{0,2}, R_{1,0}, R_{1,1}, R_{1,2})$ [52], and density $(\mathcal{C})$ . Figure 3 gives examples with varying numbers of Gaussians. + +The 3D density function, $\mathbf{G}(\mathbf{x})$ , defines the density of the Gaussian model at the query location $\mathbf{x}$ in world space. We define the density function of a single 3D anisotropic Gaussian as + +$$ +\mathbf {G} _ {i} (\mathbf {x}) = \mathcal {C} \exp \left[ - 0. 5 \left((\mu - \mathbf {x}) ^ {T}\right) \Sigma^ {- 1} (\mu - \mathbf {x}) \right], \tag {2} +$$ + +where the precision matrix $\Sigma^{-1} = \mathbf{R}^T\mathbf{D}\mathbf{R}$ and $\mathbf{R}$ is the rotation matrix computed from the 6 DOF representation and $\mathbf{D} = \mathrm{diagonal}(1 / \sigma_x^2, 1 / \sigma_y^2, 1 / \sigma_z^2)$ . + +The density of the entire Gaussian model is the sum of + +![](images/ec00476944913dcd817d00f3d41bc2cd0145cd64498b0fb6a23acc5bcd2a10c4.jpg) +Figure 3. Gaussian Density Model. The approximation to the NeRF's density field using a sum of 3D anisotropic Gaussians using: a) 2 Gaussians per bone, b) 4 Gaussian per bone, and c) 8 Gaussian per bone; d) is the groundtruth mesh. Note: ellipses are scaled to 2.5 STD of the Gaussians ( $99^{th}$ percentile) + +![](images/5a3d4933d5d1871320d25c4edbbc4f4867e9c3b859c730b2cb978dc4b2b22a6e.jpg) + +![](images/aeb845d270390441afd941d99b6903e97e9558800b639219a6caafb5cbbae736.jpg) + +![](images/dc24cc99a507b56267afb010ba396393b97ff4687ae1a4e044d2c5864c0ccee6.jpg) + +the density of each. The query location is transformed to the local space of the given Gaussian's joint $j$ at time-step $t$ using the world-to-local transformation matrix $\theta_{t,j}^{-1}$ , rigidly attaching the Gaussians to the underlying skeleton and facilitating animation, + +$$ +\mathbf {G} (\mathbf {x}) = \sum_ {i = 0} ^ {J \times K} \mathbf {G} _ {i} \left(\theta_ {t, j} ^ {- 1} \mathbf {x}\right). \tag {3} +$$ + +We jointly fit the parameters of the Gaussian density model to the neural field by minimizing $\mathcal{L}_{\mathrm{gDensity}}$ , the L2 error between the density function $\mathbf{G}(\mathbf{x})$ and the target neural density field at query location $\mathbf{x}$ . We detach the gradients of the neural density field to optimize the Gaussians, fitting the + +![](images/7f25f62062fb066ffb4a0b34fd80565f928d512efda55cac0198afe4b0e9dff6.jpg) +Figure 4. 3D Anisotropic Gaussian Raytracing. a) A cross-section of a 3D anisotropic Gaussian with rays passing through the Gaussian. b) The computed 1D Gaussians resulting from our derivation in Section 3.2 (colored solid), compared to sampling the 3D Gaussian directly (dashed), with their exact match validating the correctness. c) The transmittance along each ray which is used as the shadow map value. + +Gaussians to the neural field and not the other way around. + +Gaussian Ray Tracing. Figure 4 shows how casting a ray, $r$ , with ray origin $\mathbf{r}_o \in \mathbb{R}^{3 \times 1}$ and direction $\mathbf{r}_d \in \mathbb{R}^{3 \times 1}$ , through a 3D anisotropic Gaussian results in a 1D Gaussian density along the ray. Through the Gaussian body model, this equates to a sum of 1D Gaussians for which analytic integrals can be computed. The amount of occlusion these rays experience is equal to the sum of the integrals of each of the 1D Gaussians across the rays. The transmittance value, $\mathcal{T}$ , used as the shadow map value, $s$ , is the exponential of the negative integral from the start of the ray, $t = 0$ to the length of the ray, $t = l$ , + +$$ +s = \mathcal {T} _ {r} = \exp \left[ - \sum_ {i = 0} ^ {J \times K} \int_ {0} ^ {l} \mathbf {G} _ {i} ^ {r} \right]. \tag {4} +$$ + +$G_{i}^{r}$ is the 1D Gaussian created by the ray, $r$ , going through the 3D anisotropic Gaussian, $G_{i}$ with mean $\mu \in \mathbb{R}^{3\times 1}$ and precision matrix $\Sigma^{-1}\in \mathbb{R}^{3\times 3}$ . We derive in the supplemental how the 1D Gaussian's density function takes the form + +$$ +G _ {i} ^ {r ^ {s}} = \bar {\mathcal {C}} \cdot \exp \left(- \frac {(\bar {\mu} - x) ^ {2}}{2 \bar {\sigma} ^ {2}}\right), \tag {5} +$$ + +where + +$$ +\begin{array}{l} \bar {\mathcal {C}} = \mathcal {C} \exp \left(- 0. 5 \left(\left(\mu - \mathbf {r} _ {o}\right) ^ {T} \Sigma^ {- 1} (\mu - \mathbf {r} _ {o}) - \frac {\bar {\mu} ^ {2}}{\bar {\sigma} ^ {2}}\right)\right), \\ \bar {\mu} = \frac {\mathbf {r} _ {d} ^ {T} \Sigma^ {- 1} (\boldsymbol {\mu} - \mathbf {r} _ {o})}{\mathbf {r} _ {d} ^ {T} \Sigma^ {- 1} \mathbf {r} _ {d}}, \text {a n d} \\ \bar {\sigma} = \sqrt {\frac {1}{\mathbf {r} _ {d} ^ {T} \Sigma^ {- 1} \mathbf {r} _ {d}}}. \tag {6} \\ \end{array} +$$ + +This formula is more complex than in [30], as it now accounts for anisotropic Gaussians with an arbitrary covariance instead of isotropic Gaussians. The comparison to + +sampling the 3D Gaussian in Figure 4 validates their correctness. It also lets us compute the cumulative density function (CDF) analytically, thereby avoiding the sampling in classical NeRFs, + +$$ +\int_ {0} ^ {x} \mathbf {G} _ {i} ^ {r ^ {s}} = \bar {\mathcal {C}} \cdot 0. 5 \cdot \left(1 + \operatorname {e r f} \left(\frac {x - \bar {\mu}}{\bar {\sigma} \sqrt {2}}\right)\right). \tag {7} +$$ + +Together with Equation 4, this integral computes the shadow $s$ when applied to the secondary ray with origin $r_o^s$ , as the point on the subject's surface computed from the depth map $\mathbf{d}$ , and direction $r_d^s$ towards the light. + +# 3.3. Optimization + +In addition to the introduced reconstruction loss $\mathcal{L}_{\mathrm{RGB}}$ , $\mathcal{L}_{\mathrm{Eikonal}}$ for SDF regularization as in [12], and Gaussian fitting $\mathcal{L}_{\mathrm{gDensity}}$ , we regularize the training with i) a $\mathcal{L}_{\mathrm{mask}} = |\hat{\rho} - \rho|$ that regularizes density by minimizing the difference between integrated accumulation, $\hat{\rho}$ , and the foreground mask, $\rho$ , ii) $\mathcal{L}_{\mathrm{amb}} = ||\hat{\mathbf{L}}_{\mathrm{amb}} - 0.1||^2$ preferring small ambient light values, iii) $\mathcal{L}_{\mathrm{gSigma}}$ that prevent too large or small Gaussians, and iv) $\mathcal{L}_{\mathrm{gMean}}$ that pulls Gaussians closer to the center of the bones. + +Training proceeds in three stages. In stage I, the reconstruction loss is replaced with one that encourages predicting gray inside the silhouette, to learn a rough body shape without illumination effects. In stage II, $\mathcal{L}_{\mathrm{gDensity}}$ and its regularizers are introduced, allowing the Gaussian density model to fit. Finally, in stage III, the $\mathcal{L}_{\mathrm{RGB}}$ takes over to optimize the light direction and learn the albedo. Additional training details are provided in the supplemental. + +# 4. Results + +We evaluate our method on synthetic sequences, as done in prior work [16]. However, this does not test performance in real world conditions. Hence, we captured a new dataset in direct sunlight and compare to the most closely related baselines, showing significantly improved relightable body models. The supplemental video and supplemental document provide additional qualitative comparisons, including relighting with HDRi environment maps. + +Synthetic datasets. We test our model on the RANA dataset [16], and further create our own synthetic sequence by obtaining a textured mesh of a subject with a 3D full body scanner (VITUS 3D Body Scanner). A Blender [8] cloth simulation was applied to a shirt over the scan and the character was automatically rigged and animated using Mixamo [2]. We use the 'swing-dance' animation as the driving motion as it contains a variety of poses from all body angles. Four (3 train, 1 test) cameras are placed around the subject at 90 degrees from each other. A directional light source illuminates the scene with a slight ambient contribution such that the shadowed areas were not fully black. The + +annimations are rendered using the Cycles render engine. In addition, ground-truth pose and segmentation masks are exported. We split the dataset into 600/114 images for the train/test sets, with the test set including 57 novel view images and 57 novel poses, out of which 15 have a strong hard cast shadows that we test separately. + +Outdoor sunlight dataset. We recorded two sequences of real human motion in an outdoor scene during a sunny day. This setting has largely been unexplored in neural body models and few publicly available datasets are available. We capture the data using 3 cameras (Canon EOS R8, Canon EOS 70D, iPhone12) and obtain SMPL estimates using EasyMocap [1,10,11]. Segmentation masks are obtained using the Segment Anything Model (SAM) [19]. We divide the frames into 600/200 images for the train/test splits, using all three cameras for training. + +Baselines We evaluate our method using the hard illumination dataset against Relighting4D [7]. Due to code being unavailable, we were not able to compare against RANA [16] and Xu et al. [43] using our datasets. We instead quantitatively compare albedo estimates between our model and RANA on their dataset as these results were kindly made available. We further compare our work with other template-less neural body models [36, 37], highlighting the drawback when not explicitly modeling lighting. DANBO [36] is our neural field backbone. NPC [37] forms the current state-of-the-art template-less neural character model. + +# 4.1. Albedo Estimates (training poses) + +We ran the official implementation of Relighting4D (R4D) [7] on our outdoor dataset, providing the same segmentation masks and SMPL body model as to our method (our method only uses the skeleton, not the surface). As the first stage of R4D is NeuralBody [28] which does not take shadowing into account, it produces dark floaters in the space to approximate the hard shadow, hindering their subsequent relighting module from estimating shadow and shading correctly as seen in Figure 5a. + +To compare against RANA [16], we run our method on subject 1 of their synthetic dataset. Even with the dataset being monocular (light comes from the same direction relative to the camera), our method was able to accurately estimate the light direction (error of 9.9 degrees) and obtain albedo estimates with fewer lighting artifacts compared to RANA, see the back of the left leg in Figure 5b. + +# 4.2. Novel-Pose Rendering with Shadows + +In this setting, the camera and illumination are unchanged and only novel-poses are tested. These poses created new shadow casts that resulted in large appearance + +![](images/75c81702be88068639ccb1ec8aeee8d3f81014ec416e6bc25774cda6896999e4.jpg) + +![](images/90d273132a8b3b9cd7032421bcf11795e193ea41b3191fb353453e890f1a07b0.jpg) +Figure 5. Albedo Estimation. Our method can better separate shadows and lighting from training images to obtain better albedo estimates without lighting artifacts compared to a) Relighting4D [7] and b) RANA [16]. + +changes distant to the changed body part. As expected, existing methods (NPC, DANBO) overfit the training poses and the shadows created in the novel poses are highly inaccurate. Figure 6 shows how for frames that had body parts casting shadows on other regions, our method produced more accurate shadows. Table 2 quantifies the gains across novel poses and Table 3 across the subset of the novel poses that has a shadow cast across the body. + +Table 2. Novel-pose synthesis (all test frames). Our Gaussian Shadow Casting model achieves consistently better PSNR scores for novel pose renderings as it properly models the hard shadows cast by the limbs in novel positions. Existing methods only shine in perceptual metrics (SSIM and LPIPS) as these normalize contrast and hence lessen the impact of proper shadows and shading. + +
Synthetic (N=57)Real S1 (N=200)Real S2 (N=200)Average
PSNR↑SSIM↑LPIPS↓PSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPS
DANBO [36]17.520.7560.19516.570.5990.32817.690.5880.32517.260.6480.283
NPC [37]17.570.7580.18816.330.5900.33417.470.5750.32817.120.6410.283
Ours22.040.8290.16617.570.5920.35618.290.5770.35119.300.6660.291
+ +Table 3. Novel-pose synthesis (subset of test set with observed self-casting shadows). Our Gaussian Shadow Casting renders novel poses with strong hard shadows well. Our scores drop marginally on these hard frames compared to all frames in Tab. 2, while the baselines drop significantly. + +
Synthetic (N=15)Real S1 (N=41)Real S2 (N=36)Average
PSNR↑SSIM↑LPIPS↓PSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPS
DANBO [36]17.780.7400.20915.110.5590.35416.550.5470.35316.480.6150.305
NPC [37]17.810.7410.20114.880.5530.35716.510.5380.35516.400.6110.304
Ours22.130.8210.17516.880.5720.36517.400.5440.37118.810.6460.303
+ +For the synthetic sequence, improvements were consistent across all three metrics. In the real outdoor sequence, all methods attain a lower quality as the cameras are spaced further apart and the segmentation masks and 3D input + +![](images/24849be7be555d52da5bd2c38741ac76a5b83e6c3124912506f4f23e0509bdbb.jpg) +Figure 6. Novel Pose Rendering. Our method can more accurately reproduce the shadow in novel poses compared to the baselines. + +![](images/88056b478e35e01ec23792f69602136773c892245f7f28bd4614a5f193e433b3.jpg) + +![](images/b2490e5bcf56804ccf59a682b76ed502ff65daea80d2e837381bec0ecf3bac95.jpg) + +![](images/bdf72406ac60669a6f155216bbdaa5f2e019d213e5055dfc26b4d0ddb52dcf99.jpg) + +![](images/de89f37aa0141bdba367edb70c5ac3df779793a01ad914e77ea3db8f1b11ac46.jpg) + +pose, estimated with off-the-shelf 2D pose detection and lifting methods, are less reliable. Nevertheless, Figure 6 shows how our model can accurately optimize the light direction and predict realistic shadows, including on the ground. To map shadows to the ground, we estimate the ground plane from the reconstructed foot positions and cast the Gaussian shadow on it by modulating the static background with the ground shadow map. + +Table 2 and Table 3 show that our method consistently improves the PSNR while the perceptual metrics SSIM [41] improves only in one and the baselines perform better for LPIPS [47]. This lower performance in perceptual metrics is expected because these metrics normalize for brightness and contrast differences, thereby lessening the importance of producing proper shading and shadowing. In addition, the texture and geometry detail of our method is slightly lower, which we attribute to the separation into shading and albedo imposing additional constraints, thereby leading to slightly less detailed reconstructions. + +# 4.3. Render Time Comparison + +Table 4 lists the render time of our baseline compared to our full model. Casting shadows with GSC has minimal overhead (0.3s for one ray, only $2\%$ of the entire render time), enabling efficient training alongside NeRF optimization. Casting a shadow with the NeRF baseline requires processing twice the number of samples by the NeRF. The deferred shading model creates one occlusion ray and each of these secondary rays requires a similar number of samples as for the primary ray. Already with a single light source, this increases runtime by $25\%$ , a ten-fold difference to GSC. + +# 4.4. Relighting with Environment Maps + +The shadow computation not only benefits training time but also enables computing irradiance maps for environment maps. Figure 1 shows relighting with two different + +
Methodrender time [s]
DANBO + DS17.13
DANBO + DS + GSC17.47
DANBO + DS + NeRFSC21.4
DANBO + DS + GSC-HDRi-820.70
DANBO + DS + GSC-HDRi-1623.57
DANBO + DS + GSC-HDRi-3229.49
DANBO + DS + GSC-HDRi-6441.22
+ +Table 4. Render time. The overhead of Gaussian Shadow Casting (GSC) is minimal on DANBO with diffuse shading (DANBO + DS) and enables casting many rays (64 for GSC-HDRi-64). By contrast, NeRF shadow casting (NeRFSC) doubles the runtime with every light source, making training prohibitively slow and HDRi relighting impractical. + +HDRi maps (obtained from Poly Haven [14]) by casting 64 secondary light rays towards the environment map for each pixel through importance sampling. In both cases, the bright sun casts a strong shadow while the colored light from the environment leads to natural shading that matches the character with the environment. This enables placing the reconstructed characters into new environments and giving them a natural and consistent look with respect to the rest of the scene while still containing cast shadows. + +Our method is plug-and-play with other neural body models due to the deferred rendering approach. It can be used with higher quality volumetric neural models without degraded quality when training on uniformly lit data and using GSC for relighting as shown in Figure 7. + +# 4.5. Ablation Study + +We test a variety of implementation details in our model, including using only diffuse shading on top of DANBO (DANBO + Diff. Shading), only using the Gaussians to cast shadows (Ours w/o Diff. Shading), providing ground truth lighting (Ours-GT Light), and detaching the normals prior to the diffuse shading (Ours-Detached Normals). The + +![](images/1a86ed186dc1e803ed89d0289d0915b6d8875edc7602814e767993a323d6905b.jpg) +Figure 7. HDRi Relighting on MonoPerfCap using NPC [37] as the backbone. Our Gaussian relighting method can be utilized at inference with higher quality models on data that is uniformly lit. + +results of which can be seen in Table 5, which shows that each of our contributions improves reconstruction quality at test time. + +Table 5. Ablation on Synthetic Sequence. + +
TrainingNovel PoseNovel View
PSNR↑SSIM↑LPIPS↓PSNRSSIMLPIPSPSNRSSIMLPIPS
DANBO27.660.9130.13317.520.7560.19518.850.7730.179
DANBO + Diff. Shading24.820.8600.18718.600.7850.20824.400.8750.167
Ours w/o Diff. Shading27.170.8790.16121.220.8020.17825.080.8670.163
Ours-Detached Normals25.900.8630.18220.430.8140.19226.180.8930.158
Ours-GT Light25.220.8610.18021.230.8300.18426.680.8950.155
Ours26.600.8760.16522.300.8270.17627.320.8820.154
+ +Diffuse Shading. Shadowing alone does not account for accurate shading based on how incident the light hits the surface. Moreover, the Gaussians cast long-range shadows, but their smooth and low-resolution approximation to the NeRF's density prevents them from representing finer details such as small extremities (nose, fingers). As a result, finer shading details are missed as seen in Figure 8c. On the other hand, using only shading (DANBO + Diff. Shading) already reduces texture detail quality as seen by the training scores, but achieves improved performance in novel pose and novel view conditions. However, the missing cast shadows play the most significant role in improving test scores. + +Detached Normals. We compare results between a model where network gradients could, Figure 8d, and could not, Figure 8b, backpropagate through the surface normals used in the diffuse shading to see whether or not artifacts in the shading would smoothen out the geometry. We find that the surface is indeed affected by the gradients backpropagating through the diffuse computation and observe a smoother geometry reconstruction. + +GT Light. Our method is able to fit the direction of the light source and the ambient intensity with little user input. We observe accurate light recovery when the light is initialized randomly, e.g. when coming from the back the angle error is only 1.36 degrees on our synthetic sequence. We found providing the ground truth light direction did not improve results and may have hindered the model due to the added constraints, Figure 8a. + +![](images/426c13ca53715a663c15447c85486e59497cd5ded30a493cddc06fd8fa6c54a5.jpg) +Figure 8. Ablation Comparisons (novel-pose). a) The model trained with the groundtruth light direction. b) The model trained while detaching the gradients from the surface normals during diffuse shading. c) The model trained without diffuse shading. d) Our full model. + +![](images/23dd5afffd1a5736a3f73440504cf635069e0fda9f66e42e5f7ed190d0d36f30.jpg) + +![](images/e95aa5d9484d0a0707f1b75fec64992648d5ec513cedc55683e10772d1f0e3d2.jpg) + +# 4.6. Limitations + +The Gaussian cast shadows model long-range effects, such as the arm casting a shadow on the leg but the smooth Gaussians lack high frequency details. This is a minor drawback since the diffuse shading already faithfully reproduces the light intensity fall-off as the light direction becomes more incident with the surface and therefore shades the back side of small extremities (i.e. nose and fingers) well. A future extension could be to integrate mid-scale effects with screen-space ambient occlusion and shading. + +Moreover, we noticed that disentangling color into shading and albedo, compared to the original DANBO backbone, leads to slightly lower image reconstruction metrics when shading effects are minimal. We attribute this to the additional constraints that are imposed on the model. However, the overall performance in novel light conditions is still improved significantly by our model. + +# 5. Conclusion + +We enabled the 3D reconstruction of human motions in uncontrolled environments with a Gaussian-based shadow model that applies to dynamic scenes and is differentiable for iterative refinement. The reconstructed characters support reposing and relighting in novel environments. They are equipped with global shadow computation, diffuse shading, geometric reconstruction, and a consistent albedo, much like hand-crafted computer graphics models would provide. The deferred lighting approach allows our method to be combined with other neural body models with efficient shadow computations. + +# 6. Acknowledgements + +This work was supported in part by an NSERC Discovery Grant and the computational resources and services provided by Advanced Research Computing at The University of British Columbia. + +# References + +[1] Easymocap - make human motion capture easier. https://github.com/zju3dv/EasyMocap, 2021.6 +[2] Adobe. Mixamo. www MIXamo.com, 2023.5 +[3] Thiemo Alldieck, Marcus Magnor, Bharat Lal Bhatnagar, Christian Theobalt, and Gerard Pons-Moll. Learning to reconstruct people in clothing from a single RGB camera. In CVPR, 2019. 2 +[4] Timur Bagautdinov, Chenglei Wu, Tomas Simon, Fabian Prada, Takaaki Shiratori, Shih-En Wei, Weipeng Xu, Yaser Sheikh, and Jason Saragih. Driving-signal aware full-body avatars. ACM TOG (Proc. SIGGRAPH), 2021. 2 +[5] Sai Bi, Stephen Lombardi, Shunsuke Saito, Tomas Simon, Shih-En Wei, Kevyn Mcphail, Ravi Ramamoorthi, Yaser Sheikh, and Jason Saragih. Deep relightable appearance models for animatable faces. ACM TOG (Proc. SIGGRAPH), 2021. 2 +[6] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P.A. Lensch. Nerd: Neural reflectance decomposition from image collections. In ICCV, 2021. 2 +[7] Zhaoxi Chen and Ziwei Liu. Relighting4d: Neural relightable human from videos. In ECCV, 2022. 2, 3, 6 +[8] Blender Online Community. Blender - a 3d modelling and rendering package. http://www.blender.org, 2023.5 +[9] Dawa Derksen and Dario Izzo. Shadow neural radiance fields for multi-view satellite photogrammetry. In CVPR, 2021. 2 +[10] Junting Dong, Qi Fang, Wen Jiang, Yurou Yang, Hujun Bao, and Xiaowei Zhou. Fast and robust multi-person 3d pose estimation and tracking from multiple views. In TPAMI, 2021. 6 +[11] Junting Dong, Qing Shuai, Yuanqing Zhang, Xian Liu, Xiaowei Zhou, and Hujun Bao. Motion capture from internet videos. In ECCV, 2020. 6 +[12] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In ICML, 2020. 5 +[13] Michelle Guo, Alireza Fathi, Jiajun Wu, and Thomas Funkhouser. Object-centric neural scene rendering, 2020. 2 +[14] Poly Haven. Poly haven. www.polyhaven.com/HDris, 2023.7 +[15] Jingwei Huang, Haotian Zhang, Li Yi, Thomas Funkhouser, Matthias Nießner, and Leonidas J Guibas. Texturenet: Consistent local parametrizations for learning from high-resolution signals on meshes. In CVPR, 2019. 2 +[16] Umar Iqbal, Akin Caliskan, Koki Nagano, Sameh Khamis, Pavlo Molchanov, and Jan Kautz. Rana: Relightable articulated neural avatars. arXiv preprint arXiv:2212.03237, 2022. 2, 5, 6 +[17] Wei Jiang, Kwang Moo Yi, Golnoosh Samei, Oncel Tuzel, and Anurag Ranjan. Neuman: Neural human radiance field from a single video. In ECCV, 2022. 1 + +[18] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM TOG (Proc. SIGGRAPH), 2023. 2, 3 +[19] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 6 +[20] Youngjoong Kwon, Dahun Kim, Duygu Ceylan, and Henry Fuchs. Neural human performer: Learning generalizable radiance fields for human performance rendering. NeurIPS, 2021. 1, 2 +[21] Junxuan Li, Shunsuke Saito, Tomas Simon, Stephen Lombardi, Hongdong Li, and Jason Saragih. Megane: Morphable eyeglass and avatar network. In CVPR, 2023. 2 +[22] Ruilong Li, Julian Tanke, Minh Vo, Michael Zollhofer, Jurgen Gall, Angjoo Kanazawa, and Christoph Lassner. Tava: Template-free animatable volumetric actors. In ECCV, 2022. 2 +[23] Zhe Li, Zerong Zheng, Yuxiao Liu, Boyao Zhou, and Yebin Liu. Posevocab: Learning joint-structured pose embeddings for human avatar modeling. In ACM TOG (Proc. SIGGRAPH), 2023. 2 +[24] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. ACM TOG (Proc. SIGGRAPH Asia), 2021. 2 +[25] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. ACM TOG (Proc. SIGGRAPH Asia), 2015. 2, 3 +[26] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2 +[27] Atsuhiro Noguchi, Xiao Sun, Stephen Lin, and Tatsuya Harada. Neural articulated radiance field. In ICCV, 2021. 2 +[28] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In CVPR, 2021. 1, 6 +[29] Anurag Ranjan, Kwang Moo Yi, Jen-Hao Rick Chang, and Oncel Tuzel. Facelit: Neural 3d relightable faces. In CVPR, 2023. 2 +[30] Helge Rhodin, Nadia Robertini, Dan Casas, Christian Richardt, Hans-Peter Seidel, and Christian Theobalt. General automatic human shape and motion capture using volumetric contour cues. In Bastian Leibe, Jiri Matas, Nicu Sebe, and Max Welling, editors, ECCV, 2016. 2, 3, 4, 5 +[31] H. Rhodin, N. Robertini, C. Richardt, H. Seidel, and C. Theobalt. A versatile scene model with differentiable visibility applied to generative pose estimation. In ICCV, 2015. 2, 3, 4 + +[32] Viktor Rudnev, Mohamed Elgharib, William Smith, Lingjie Liu, Vladislav Golyanik, and Christian Theobalt. Nerf for outdoor scene relighting. In ECCV, 2022. 2 +[33] Srinath Sridhar, Helge Rhodin, Hans-Peter Seidel, Antti Oulasvirta, and Christian Theobalt. Real-time hand tracking using a sum of anisotropic gaussians model. In 3DV, 2014. 2, 3 +[34] Pratul P. Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T. Barron. Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In CVPR, 2021. 2 +[35] Carsten Stoll, Nils Hasler, Juergen Gall, Hans-Peter Seidel, and Christian Theobalt. Fast articulated motion tracking using a sums of gaussians body model. ICCV, 2011. 3 +[36] Shih-Yang Su, Timur Bagautdinov, and Helge Rhodin. Danbo: Disentangled articulated neural body representations via graph neural networks. In ECCV, 2022. 1, 2, 3, 4, 6 +[37] Shih-Yang Su, Timur Bagautdinov, and Helge Rhodin. Npc: Neural point characters from video. In ICCV, 2023. 1, 2, 6, 8 +[38] Shih-Yang Su, Frank Yu, Michael Zollhöfer, and Helge Rhodin. A-nerf: Articulated neural radiance fields for learning human shape, appearance, and pose. In NeurIPS, 2021. 2 +[39] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: Image synthesis using neural textures. ACM TOG (Proc. SIGGRAPH), 2019. 3 +[40] Shaofei Wang, Katja Schwarz, Andreas Geiger, and Siyu Tang. Arah: Animatable volume rendering of articulated human sdfs. In ECCV, 2022. 1, 2 +[41] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 7 +[42] Zian Wang, Tianchang Shen, Jun Gao, Shengyu Huang, Jacob Munkberg, Jon Hasselgren, Zan Gojcic, Wenzheng Chen, and Sanja Fidler. Neural fields meet explicit geometric representations for inverse rendering of urban scenes. In CVPR, 2023. 2 +[43] Zhen Xu, Sida Peng, Chen Geng, Linzhan Mou, Zihan Yan, Jiaming Sun, Hujun Bao, and Xiaowei Zhou. Relightable and animatable neural avatar from sparse-view video, 2023. 2, 3, 6 +[44] Haotian Yang, Mingwu Zheng, Wanquan Feng, Haibin Huang, Yu-Kun Lai, Pengfei Wan, Zhongyuan Wang, and Chongyang Ma. Towards practical capture of high-fidelity reconfigurable avatars. In ACM TOG (Proc. SIGGRAPH Asia), 2023. 2 +[45] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. In NeurIPS, 2021. 3 +[46] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snavely. PhySG: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In CVPR, 2021. 3 +[47] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 7 + +[48] Xiuming Zhang, Pratul P. Srinivasan, Boyang Deng, Paul Debevec, William T. Freeman, and Jonathan T. Barron. Nefactor: Neural factorization of shape and reflectance under an unknown illumination. ACM TOG (Proc. SIGGRAPH), 2021. 2 +[49] Zerong Zheng, Han Huang, Tao Yu, Hongwen Zhang, Yandong Guo, and Yebin Liu. Structured local radiance fields for human avatar modeling. In CVPR, 2022. 2 +[50] Zerong Zheng, Xiaochen Zhao, Hongwen Zhang, Boning Liu, and Yebin Liu. Avatarrex: Real-time expressive full-body avatars. ACM TOG (Proc. SIGGRAPH), 2023. 2 +[51] Taotao Zhou, Kai He, Di Wu, Teng Xu, Qixuan Zhang, Kuixiang Shao, Wenzheng Chen, Lan Xu, and Jingyi Yu. Relightable neural human assets from multi-view gradient illuminations. In CVPR, 2023. 2 +[52] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In CVPR, 2019. 4 \ No newline at end of file diff --git a/2024/Gaussian Shadow Casting for Neural Characters/images.zip b/2024/Gaussian Shadow Casting for Neural Characters/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..dde584bfad171b89aa6d93cb11eae62fdbadb402 --- /dev/null +++ b/2024/Gaussian Shadow Casting for Neural Characters/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6c1b6ad3c3683054c3c4e92e18ab69b20fa093ca7e305cc52d5654aba62a5b87 +size 524580 diff --git a/2024/Gaussian Shadow Casting for Neural Characters/layout.json b/2024/Gaussian Shadow Casting for Neural Characters/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5317a56675d8db49f5875caffc9c93255d058816 --- /dev/null +++ b/2024/Gaussian Shadow Casting for Neural Characters/layout.json @@ -0,0 +1,8464 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 146, + 103, + 448, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 103, + 448, + 121 + ], + "spans": [ + { + "bbox": [ + 146, + 103, + 448, + 121 + ], + "type": "text", + "content": "Gaussian Shadow Casting for Neural Characters" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 149, + 142, + 443, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 142, + 443, + 173 + ], + "spans": [ + { + "bbox": [ + 149, + 142, + 443, + 173 + ], + "type": "text", + "content": "Luis Bolanos1 Shih-Yang Su1 Helge Rhodin1,2 \n1The University of British Columbia 2Bielefeld University" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 78, + 203, + 140, + 323 + ], + "blocks": [ + { + "bbox": [ + 78, + 203, + 140, + 323 + ], + "lines": [ + { + "bbox": [ + 78, + 203, + 140, + 323 + ], + "spans": [ + { + "bbox": [ + 78, + 203, + 140, + 323 + ], + "type": "image", + "image_path": "42becdbf56c8af7d828f3e01f48ee9b2d7dc0f848629c4e27af18b2fd12753e8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 339, + 547, + 384 + ], + "lines": [ + { + "bbox": [ + 46, + 339, + 547, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 339, + 547, + 384 + ], + "type": "text", + "content": "Figure 1. Gaussian Shadow Casting (GSC): Our method is able to reconstruct 3D neural characters from a sparse set of videos in settings with strong directional illumination. GSC uses a sum of Gaussians density model to cast secondary shadow rays efficiently with an analytic formula. Our method learns to remove shadows from the neural color field, allowing us to relight in novel illuminations. All faces are blurred for anonymity." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 142, + 204, + 162, + 324 + ], + "blocks": [ + { + "bbox": [ + 142, + 204, + 162, + 324 + ], + "lines": [ + { + "bbox": [ + 142, + 204, + 162, + 324 + ], + "spans": [ + { + "bbox": [ + 142, + 204, + 162, + 324 + ], + "type": "image", + "image_path": "05b584c53cf36fcb7953e597a75f4ee52831125e10766c03ce275d0a02a298fc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 164, + 204, + 219, + 324 + ], + "blocks": [ + { + "bbox": [ + 164, + 204, + 219, + 324 + ], + "lines": [ + { + "bbox": [ + 164, + 204, + 219, + 324 + ], + "spans": [ + { + "bbox": [ + 164, + 204, + 219, + 324 + ], + "type": "image", + "image_path": "10eb7278dd5c34c0c3c1687853f86c26fa0bd19aee45a6becbdf9923e4351e99.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 220, + 204, + 259, + 323 + ], + "blocks": [ + { + "bbox": [ + 220, + 204, + 259, + 323 + ], + "lines": [ + { + "bbox": [ + 220, + 204, + 259, + 323 + ], + "spans": [ + { + "bbox": [ + 220, + 204, + 259, + 323 + ], + "type": "image", + "image_path": "64ec36f7e9cf948dc8ce3126a554a8a3b72806968bcc52775fb50f8d628f516f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 265, + 210, + 351, + 321 + ], + "blocks": [ + { + "bbox": [ + 265, + 210, + 351, + 321 + ], + "lines": [ + { + "bbox": [ + 265, + 210, + 351, + 321 + ], + "spans": [ + { + "bbox": [ + 265, + 210, + 351, + 321 + ], + "type": "image", + "image_path": "1e4121215bc8cdefe69fffbce202c5aace9093792f72cdf5e4a5c1d44bb56129.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 357, + 210, + 406, + 321 + ], + "blocks": [ + { + "bbox": [ + 357, + 210, + 406, + 321 + ], + "lines": [ + { + "bbox": [ + 357, + 210, + 406, + 321 + ], + "spans": [ + { + "bbox": [ + 357, + 210, + 406, + 321 + ], + "type": "image", + "image_path": "4af957060dbe8eb456b75f5a43264a2d8bfef292361a076dab9275e40e70557f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 415, + 204, + 514, + 324 + ], + "blocks": [ + { + "bbox": [ + 415, + 204, + 514, + 324 + ], + "lines": [ + { + "bbox": [ + 415, + 204, + 514, + 324 + ], + "spans": [ + { + "bbox": [ + 415, + 204, + 514, + 324 + ], + "type": "image", + "image_path": "1dd102a236a19906632678a8071a98b96fadb035081f1a26acf5fb6b95c1478d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 143, + 392, + 192, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 392, + 192, + 404 + ], + "spans": [ + { + "bbox": [ + 143, + 392, + 192, + 404 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 418, + 290, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 418, + 290, + 706 + ], + "spans": [ + { + "bbox": [ + 46, + 418, + 290, + 706 + ], + "type": "text", + "content": "Neural character models can now reconstruct detailed geometry and texture from video, but they lack explicit shadows and shading, leading to artifacts when generating novel views and poses or during relighting. It is particularly difficult to include shadows as they are a global effect and the required casting of secondary rays is costly. We propose a new shadow model using a Gaussian density proxy that replaces sampling with a simple analytic formula. It supports dynamic motion and is tailored for shadow computation, thereby avoiding the affine projection approximation and sorting required by the closely related Gaussian splattering. Combined with a deferred neural rendering model, our Gaussian shadows enable Lambertian shading and shadow casting with minimal overhead. We demonstrate improved reconstructions, with better separation of albedo, shading, and shadows in challenging outdoor scenes with direct sun light and hard shadows. Our method is able to optimize the light direction without any input from the user. As a result, novel poses have fewer shadow artifacts, and relighting in novel scenes is more realistic compared to the state-of-the-art methods, providing new ways to pose neural characters in novel environments, increasing their applicability. Code available at: https://github.com/LuisBolanos17/GaussianShadowCasting" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 418, + 387, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 418, + 387, + 430 + ], + "spans": [ + { + "bbox": [ + 306, + 418, + 387, + 430 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 437, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 521 + ], + "type": "text", + "content": "It is now possible to reconstruct animatable 3D neural avatars from video but methods do not account for accurate lighting and shadows. They have to rely on recordings that have soft uniform lighting, which precludes recording outdoors in direct sun light and on film sets with spotlights, and most are unable to relight characters in novel environments, limiting their applicability in content creation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "text", + "content": "The most recent body models [17,20,28,36,37,40] which are based on neural radiance fields (NeRFs) [26], approximate the light transport by casting primary rays between the camera and the scene, sampling the underlying neural network dozens of times along each ray to obtain the density and color. As they do not include an illumination model, the color that the NeRF learns includes lighting, shadow, and view-dependent effects. Learning a body model in a challenging scene with a strong directional light source, such as the sun, leads to the neural field overfitting to the observed shadows. It does not generalize to novel poses, as the cast shadows are global effects where movement of a joint could affect the appearance of other distant areas of the body. Figure 1 shows such setting. This is in contrast to local shading effects such as wrinkles in clothing which current body models can successfully reconstruct." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20997" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": "To cast shadows within NeRF, secondary ray tracing from the reconstructed body model to the light source is an option. Although the predominant NeRF formulation enables casting secondary rays without change, it comes with a massive computational cost. For each sample along the primary ray, an equal number of secondary rays would have to be computed, each with multiple samples, leading to a quadratic, instead of linear, complexity in the number of samples per pixel. As a result, current re-lighting models only support diffuse reflection [16], hard shadows that do not generalize to novel poses [7], and soft dynamic shadow maps by approximate sphere tracing [43]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 218, + 289, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 218, + 289, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 218, + 289, + 445 + ], + "type": "text", + "content": "Our core contribution is introducing an additional volumetric density field that is approximate but significantly speeds up dynamic shadow casting while still maintaining differentiability and smoothness for gradient-based optimization. We introduce an anisotropic Gaussian density model and associated rendering functions that approximate the fine-grained density of the NeRF. The Gaussians have the beneficial property that we can integrate their density along a ray in closed form, thereby avoiding any sampling steps. Our derivation and implementation differs significantly from existing work using Gaussians for rendering. Compared to Gaussian Splatting [18, 31, 33], we neither require an affine approximation nor back-to-front ordering. Compared to Gaussian density models we alleviate their sampling [31] with an analytic integration and extend the existing analytic integration [30] to apply to anisotropic Gaussians. Notably, the Gaussian density is optimized alongside the NeRF without requiring a reference mesh such as SMPL [25]; it is template-free." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 448, + 289, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 448, + 289, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 448, + 289, + 533 + ], + "type": "text", + "content": "To further reduce runtime, we use a deferred shading approach [7] in which the first rendering pass computes the albedo, depth, and normal for each pixel. The second pass casts only one secondary ray per pixel from the estimated surface point to the light source. This makes shadow computations independent of the number of samples in the NeRF, avoiding the mentioned quadratic complexity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "content": "Our experiments with strong directional light and cast shadows demonstrate that our explicit lighting reduces the occurrence of artifacts in novel-view and novel-pose synthesis tasks. Figure 1 shows how our method is able to disentangle lighting and shadows from the avatar's albedo given sparse-view data from only a single illumination. We take advantage of the dynamic aspect of the data where we can observe the same body part in multiple illuminations as the subject moves. We further demonstrate the ability to optimize the unknown light directions without any user input or careful initialization. Moreover, relighting of the neural character enables us to composite recorded motions into novel scenes realistically, making them directly applicable in computer graphics and entertainment industries, as demonstrated by the HDRi re-lighting in Figure 1-right." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 71, + 393, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 71, + 393, + 84 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 393, + 84 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 91, + 547, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 547, + 222 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 547, + 222 + ], + "type": "text", + "content": "We build on neural body models using NeRF [26], which we introduce briefly. The subsequent discussion focuses on relighting methods for 3D scenes and body models as well as how Gaussians are used in rendering and reconstruction. Neural avatars model dynamic performances by conditioning the neural rendering model on a template mesh driven by skeleton motion [4, 16, 20, 23, 24, 40, 49, 50] or template-free by linking neural fields directly to a skeleton [22, 27, 36-38]. Our implementation uses the more flexible template-free approach but it is general enough to extend to any NeRF-based model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 223, + 547, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 223, + 547, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 223, + 547, + 486 + ], + "type": "text", + "content": "Static NeRF scene relighting approaches can be categorized by either implicit [6, 9, 32, 34, 48] or explicit [13, 42] implementations. In implicit methods, the NeRF's MLP is extended to further output illumination data such as shadow, direct and indirect illumination or occlusion maps [9, 32, 34], or decompose the scene into material properties such as metallicity and roughness which can be used in a Bidirectional Reflectance Distribution Function (BRDF) lighting model [6, 48]. These extended MLPs are conditioned at training and test time on lighting information such as spherical harmonics coefficients [32], or light direction [9]. Implicit methods require large amounts of data in both multi-view and multiple illuminations with lighting information known or estimated [9, 32]. Explicit methods simulate how real light interacts with the environment which improves the generalizability to novel illuminations but are difficult to extend to dynamic scenes or objects. These methods either utilize a secondary data structure such as proxy geometries where lighting computations can be done using established methods [42], or attempt to cast the necessary secondary rays within the neural field's volume which comes with a significant computational burden [13]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 486, + 547, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 547, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 547, + 604 + ], + "type": "text", + "content": "Dynamic neural character relighting has been built on top of volume rendering methods [5, 7, 21, 29, 43, 44, 51] as well as 2D CNN based models [16]. Implicit methods again require large amounts of data which can only be captured using light stages with known illumination [5, 21, 44, 51] or, across multiple subjects for faces that are self-similar, each captured in a different setting with in-the-wild illumination [29]. Our model provides dynamic and explicit shading and is most closely related to the following three methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "RANA [16] uses SMPL+D [3] to estimate the coarse geometry of a person and extract an albedo texture map using TextureNet [15]. Given a target pose, they render person-specific neural features alongside coarse albedo and normals from the SMPL-D model. These are passed through two CNNs to refine the albedo and normal maps. Finally, they generate a light map using spherical harmonics and the normal map which is multiplied by the albedo map to obtain the final lit image. While spherical harmonics allow a wide" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20998" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "content": "array of lighting conditions to be simulated, cast shadows are not present, e.g., an arm casting a shadow on the body. Our work implements a Gaussian density model [30, 31] to facilitate fast and efficient secondary ray tracing to compute these cast shadows." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 132, + 286, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 132, + 286, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 132, + 286, + 251 + ], + "type": "text", + "content": "Likewise, Relighting4D [7] uses SMPL [25] to condition a 4D neural field of latent features which are trilinearly interpolated based on the nearby vertices to the query location. The latent features are passed through an MLP to obtain geometry, occlusion, and reflectance properties which are fed through a BRDF to get the final lit image. It is able to estimate the light probe, and at inference time, be able to switch the light probe to a new illumination. However, Relighting4D was not designed to work with hard shadows in novel poses, which is the focus point of our work." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 251, + 286, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 286, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 286, + 407 + ], + "type": "text", + "content": "Finally, Xu et al. [43] utilize a signed distance field (SDF) based approach to learn a neural human avatar which utilizes SMPL-based inverse Linear Blender Skinning (LBS) and a displacement field to obtain canonical features. They utilize Hierarchical Distance Queries (HDQ) to compute minimum distances from world space to surface locations and perform sphere tracing to obtain material and surface properties for each camera ray. They further take advantage of HDQ through the SDF by computing soft visibility maps towards a learned light probe. While HDQ allows for fast occlusion checks, their solution focuses on soft approximate shadows whereas our work enables hard shadow casting." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 407, + 286, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 407, + 286, + 551 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 286, + 551 + ], + "type": "text", + "content": "Gaussians have been used in rendering applications as differentiable methods for computing visibility and occlusions [30, 31, 35], as components of environment maps [46], or as a means to improve rendering efficiency for neural scenes [18]. Most methods are limited to spherical Gaussians [30, 31], while Gaussian Splatting uses an affine approximation that is only accurate when many small Gaussians are used [18], and Sridhar et al. use an approximation by perspective projection of ellipsoids [33]. Our work extends Rhodin et al. [30, 31] to use anisotropic Gaussians, without introducing any approximation, and tailors the analytic formulas and implementation towards shadow casting." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 573, + 102, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 573, + 102, + 585 + ], + "spans": [ + { + "bbox": [ + 47, + 573, + 102, + 585 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": "Our method reconstructs a neural character from a set of " + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": " images of width " + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": " and height " + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{I}_t \\in \\mathbb{R}^{H \\times W \\times 3}\\}_{t=1}^N" + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": ", and corresponding character poses " + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\theta_t \\in \\mathbb{R}^{J \\times 4 \\times 4}" + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": ". The pose is represented as a skeleton with one " + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": " local-to-world transformation matrix for each of the " + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": " joints. Figure 2 gives an overview of our method. A key element of our design is a deferred illumination model [39] that separates the rendering into computing albedo, " + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{a} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": ", surface normal, " + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{n}} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": ", and depth, " + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{d} \\in \\mathbb{R}" + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": ", in a first pass and subsequently adding shading and shadow, " + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "s \\in [0,1]" + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": ", in a second" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": "pass. Our key contribution is the closed form formula for the shadow " + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 103, + 466, + 115 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 103, + 466, + 115 + ], + "spans": [ + { + "bbox": [ + 306, + 103, + 466, + 115 + ], + "type": "text", + "content": "3.1. Deferred Neural Illumination" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 122, + 545, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 122, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 304, + 122, + 545, + 229 + ], + "type": "text", + "content": "Our volumetric body model is optimized on a reconstruction objective, " + }, + { + "bbox": [ + 304, + 122, + 545, + 229 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{RGB}}" + }, + { + "bbox": [ + 304, + 122, + 545, + 229 + ], + "type": "text", + "content": " that minimizes the squared difference between the input images " + }, + { + "bbox": [ + 304, + 122, + 545, + 229 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_t" + }, + { + "bbox": [ + 304, + 122, + 545, + 229 + ], + "type": "text", + "content": " and the rendering of the model. We test our method using DANBO [36]. It outputs a color and density for samples " + }, + { + "bbox": [ + 304, + 122, + 545, + 229 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 122, + 545, + 229 + ], + "type": "text", + "content": " along the primary view rays. These are subsequently integrated to compute a color, which we interpret as the albedo, " + }, + { + "bbox": [ + 304, + 122, + 545, + 229 + ], + "type": "inline_equation", + "content": "\\mathbf{a}" + }, + { + "bbox": [ + 304, + 122, + 545, + 229 + ], + "type": "text", + "content": ". The illuminated color for a given pixel of the reconstructed image, " + }, + { + "bbox": [ + 304, + 122, + 545, + 229 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{c}}" + }, + { + "bbox": [ + 304, + 122, + 545, + 229 + ], + "type": "text", + "content": ", is computed by a Lambertian reflectance model," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 334, + 238, + 545, + 258 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 238, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 334, + 238, + 545, + 258 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {c}} = \\mathbf {a} \\left(\\theta_ {t}\\right) \\left(\\hat {\\mathbf {L}} _ {\\mathrm {a m b}} + s \\left(\\theta_ {t}\\right) \\mathbf {L} _ {\\mathrm {c o l}} \\left(\\hat {\\mathbf {L}} _ {\\mathrm {d i r}} \\cdot \\hat {\\mathbf {n}} \\left(\\theta_ {t}\\right)\\right)\\right). \\tag {1}", + "image_path": "3207da17df3f8eab57867434cdc1490210620685aa845a7d03f0abe60a3fec59.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 265, + 545, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 265, + 545, + 325 + ], + "spans": [ + { + "bbox": [ + 304, + 265, + 545, + 325 + ], + "type": "text", + "content": "This diffuse shading model illuminates the entire body with an ambient light " + }, + { + "bbox": [ + 304, + 265, + 545, + 325 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{L}}_{\\mathrm{amb}}" + }, + { + "bbox": [ + 304, + 265, + 545, + 325 + ], + "type": "text", + "content": " and a directional light with color " + }, + { + "bbox": [ + 304, + 265, + 545, + 325 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{\\mathrm{col}}" + }, + { + "bbox": [ + 304, + 265, + 545, + 325 + ], + "type": "text", + "content": ". The directional light intensity is attenuated by the cast shadows " + }, + { + "bbox": [ + 304, + 265, + 545, + 325 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 265, + 545, + 325 + ], + "type": "text", + "content": " and the cosine angle between the surface normal " + }, + { + "bbox": [ + 304, + 265, + 545, + 325 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{n}}" + }, + { + "bbox": [ + 304, + 265, + 545, + 325 + ], + "type": "text", + "content": " and light direction " + }, + { + "bbox": [ + 304, + 265, + 545, + 325 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{L}}_{\\mathrm{dir}}" + }, + { + "bbox": [ + 304, + 265, + 545, + 325 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 340, + 545, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 340, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 304, + 340, + 545, + 554 + ], + "type": "text", + "content": "Shading extensions. The benefit of the deferred rendering approach is that it lets us compute lighting information only once for each pixel, as opposed to at every sample location of the volumetric ray tracing leading to significantly faster computation. To be applicable, we extend DANBO to yield surface normals " + }, + { + "bbox": [ + 304, + 340, + 545, + 554 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{n}}" + }, + { + "bbox": [ + 304, + 340, + 545, + 554 + ], + "type": "text", + "content": " and depth " + }, + { + "bbox": [ + 304, + 340, + 545, + 554 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 304, + 340, + 545, + 554 + ], + "type": "text", + "content": " for a given view ray. The former we attain by switching the density formulation to a signed distance function with an Eikonal loss. The normal is then readily estimated by differentiating the distance with respect to the original query location " + }, + { + "bbox": [ + 304, + 340, + 545, + 554 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 340, + 545, + 554 + ], + "type": "text", + "content": " as in [45]. We compute " + }, + { + "bbox": [ + 304, + 340, + 545, + 554 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 304, + 340, + 545, + 554 + ], + "type": "text", + "content": " likewise to albedo " + }, + { + "bbox": [ + 304, + 340, + 545, + 554 + ], + "type": "inline_equation", + "content": "\\mathbf{a}" + }, + { + "bbox": [ + 304, + 340, + 545, + 554 + ], + "type": "text", + "content": " by integrating the sample's " + }, + { + "bbox": [ + 304, + 340, + 545, + 554 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 340, + 545, + 554 + ], + "type": "text", + "content": " positions along the ray, weighted by their density and transmittance. Furthermore, we fix the intensity of the directional light to white with a magnitude of 1.5. Without fixing the directional light intensity, the equation would be over parametrized and lead to ambiguities. Our model is invariant to light intensities between 1.0 and 2.0, as shown by Table 1." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 306, + 594, + 544, + 639 + ], + "blocks": [ + { + "bbox": [ + 305, + 563, + 545, + 586 + ], + "lines": [ + { + "bbox": [ + 305, + 563, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 305, + 563, + 545, + 586 + ], + "type": "text", + "content": "Table 1. Light intensity modulation: Our method is invariant to moderate light intensity values." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 594, + 544, + 639 + ], + "lines": [ + { + "bbox": [ + 306, + 594, + 544, + 639 + ], + "spans": [ + { + "bbox": [ + 306, + 594, + 544, + 639 + ], + "type": "table", + "html": "
Novel View & Pose\nLight Intensity:1.01.251.502.05.0
PSNR ↑24.4624.5224.3724.0018.98
LPIPS ↓0.1650.1640.1650.1630.185
", + "image_path": "056e61285a12ebaa1aae4c951dab5e31b5d9956b4a2a3a61277ef7f269c4eff1.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 659, + 450, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 659, + 450, + 672 + ], + "spans": [ + { + "bbox": [ + 306, + 659, + 450, + 672 + ], + "type": "text", + "content": "3.2. Gaussian Shadow Casting" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "For modeling shadows more efficiently, we represent the body shape with a set of Gaussians rigidly attached to the skeleton model. The relative positions, orientation, and size" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20999" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 100, + 77, + 492, + 281 + ], + "blocks": [ + { + "bbox": [ + 100, + 77, + 492, + 281 + ], + "lines": [ + { + "bbox": [ + 100, + 77, + 492, + 281 + ], + "spans": [ + { + "bbox": [ + 100, + 77, + 492, + 281 + ], + "type": "image", + "image_path": "4027d4362b1e9c7c3a80e2f7efc1ba7b34c31e537761b54f5a459d085e205ee9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 291, + 547, + 361 + ], + "lines": [ + { + "bbox": [ + 46, + 291, + 547, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 291, + 547, + 361 + ], + "type": "text", + "content": "Figure 2. Method Overview. Our method takes as input images and poses of a person. Using a neural radiance field as a backbone " + }, + { + "bbox": [ + 46, + 291, + 547, + 361 + ], + "type": "inline_equation", + "content": "[36]^1" + }, + { + "bbox": [ + 46, + 291, + 547, + 361 + ], + "type": "text", + "content": ", density, normals, and albedo values are volumetrically reconstructed and rendered. We fit a sum of 3D anisotropic Gaussian density model to approximate the neural density field and compute shadow maps using our novel anisotropic Gaussian ray occlusion equations. The shadow map is combined with a diffuse shading pass to produce the lit image. The whole model is optimized with a photometric loss against the training images. Our method is able to optimize the light direction and ambient intensity without any initialization. It also separates albedo from shading and shadow, allowing us to relight the model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 379, + 289, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 379, + 289, + 464 + ], + "spans": [ + { + "bbox": [ + 46, + 379, + 289, + 464 + ], + "type": "text", + "content": "of the Gaussians are optimized to approximate the density of the neural field and to allow for a fast, efficient, and closed-form solution for integration along a ray (occlusion checking). Our model extends previous work [30,31] by using anisotropic Gaussians (variable scale and rotation along each axis) and avoids the need for sampling during integration as in NeRF." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 478, + 287, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 478, + 287, + 583 + ], + "spans": [ + { + "bbox": [ + 46, + 478, + 287, + 583 + ], + "type": "text", + "content": "Anisotropic Gaussian body model. We define the anisotropic Gaussian density model as the matrix " + }, + { + "bbox": [ + 46, + 478, + 287, + 583 + ], + "type": "inline_equation", + "content": "\\mathbf{G} \\in \\mathbb{R}^{J \\times K \\times 13}" + }, + { + "bbox": [ + 46, + 478, + 287, + 583 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 46, + 478, + 287, + 583 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 478, + 287, + 583 + ], + "type": "text", + "content": " being the number of Gaussians per joint, typically " + }, + { + "bbox": [ + 46, + 478, + 287, + 583 + ], + "type": "inline_equation", + "content": "\\sim 8" + }, + { + "bbox": [ + 46, + 478, + 287, + 583 + ], + "type": "text", + "content": ", and the columns representing the 3D mean " + }, + { + "bbox": [ + 46, + 478, + 287, + 583 + ], + "type": "inline_equation", + "content": "(\\mu_x, \\mu_y, \\mu_z)" + }, + { + "bbox": [ + 46, + 478, + 287, + 583 + ], + "type": "text", + "content": ", the axis aligned standard deviations " + }, + { + "bbox": [ + 46, + 478, + 287, + 583 + ], + "type": "inline_equation", + "content": "(\\sigma_x, \\sigma_y, \\sigma_z)" + }, + { + "bbox": [ + 46, + 478, + 287, + 583 + ], + "type": "text", + "content": ", the rotation defined using the 6 DOF representation " + }, + { + "bbox": [ + 46, + 478, + 287, + 583 + ], + "type": "inline_equation", + "content": "(R_{0,0}, R_{0,1}, R_{0,2}, R_{1,0}, R_{1,1}, R_{1,2})" + }, + { + "bbox": [ + 46, + 478, + 287, + 583 + ], + "type": "text", + "content": " [52], and density " + }, + { + "bbox": [ + 46, + 478, + 287, + 583 + ], + "type": "inline_equation", + "content": "(\\mathcal{C})" + }, + { + "bbox": [ + 46, + 478, + 287, + 583 + ], + "type": "text", + "content": ". Figure 3 gives examples with varying numbers of Gaussians." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 585, + 288, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 585, + 288, + 633 + ], + "spans": [ + { + "bbox": [ + 46, + 585, + 288, + 633 + ], + "type": "text", + "content": "The 3D density function, " + }, + { + "bbox": [ + 46, + 585, + 288, + 633 + ], + "type": "inline_equation", + "content": "\\mathbf{G}(\\mathbf{x})" + }, + { + "bbox": [ + 46, + 585, + 288, + 633 + ], + "type": "text", + "content": ", defines the density of the Gaussian model at the query location " + }, + { + "bbox": [ + 46, + 585, + 288, + 633 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 585, + 288, + 633 + ], + "type": "text", + "content": " in world space. We define the density function of a single 3D anisotropic Gaussian as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 60, + 642, + 287, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 642, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 60, + 642, + 287, + 658 + ], + "type": "interline_equation", + "content": "\\mathbf {G} _ {i} (\\mathbf {x}) = \\mathcal {C} \\exp \\left[ - 0. 5 \\left((\\mu - \\mathbf {x}) ^ {T}\\right) \\Sigma^ {- 1} (\\mu - \\mathbf {x}) \\right], \\tag {2}", + "image_path": "714839e8a72686e95e7d79a78797ac1c7ab00892d63635aac2faeb3aec7f6995.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 664, + 287, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 664, + 287, + 702 + ], + "spans": [ + { + "bbox": [ + 46, + 664, + 287, + 702 + ], + "type": "text", + "content": "where the precision matrix " + }, + { + "bbox": [ + 46, + 664, + 287, + 702 + ], + "type": "inline_equation", + "content": "\\Sigma^{-1} = \\mathbf{R}^T\\mathbf{D}\\mathbf{R}" + }, + { + "bbox": [ + 46, + 664, + 287, + 702 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 664, + 287, + 702 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 46, + 664, + 287, + 702 + ], + "type": "text", + "content": " is the rotation matrix computed from the 6 DOF representation and " + }, + { + "bbox": [ + 46, + 664, + 287, + 702 + ], + "type": "inline_equation", + "content": "\\mathbf{D} = \\mathrm{diagonal}(1 / \\sigma_x^2, 1 / \\sigma_y^2, 1 / \\sigma_z^2)" + }, + { + "bbox": [ + 46, + 664, + 287, + 702 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 59, + 702, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 702, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 702, + 288, + 713 + ], + "type": "text", + "content": "The density of the entire Gaussian model is the sum of" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 310, + 378, + 361, + 456 + ], + "blocks": [ + { + "bbox": [ + 310, + 378, + 361, + 456 + ], + "lines": [ + { + "bbox": [ + 310, + 378, + 361, + 456 + ], + "spans": [ + { + "bbox": [ + 310, + 378, + 361, + 456 + ], + "type": "image", + "image_path": "ec00476944913dcd817d00f3d41bc2cd0145cd64498b0fb6a23acc5bcd2a10c4.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 466, + 547, + 522 + ], + "lines": [ + { + "bbox": [ + 305, + 466, + 547, + 522 + ], + "spans": [ + { + "bbox": [ + 305, + 466, + 547, + 522 + ], + "type": "text", + "content": "Figure 3. Gaussian Density Model. The approximation to the NeRF's density field using a sum of 3D anisotropic Gaussians using: a) 2 Gaussians per bone, b) 4 Gaussian per bone, and c) 8 Gaussian per bone; d) is the groundtruth mesh. Note: ellipses are scaled to 2.5 STD of the Gaussians (" + }, + { + "bbox": [ + 305, + 466, + 547, + 522 + ], + "type": "inline_equation", + "content": "99^{th}" + }, + { + "bbox": [ + 305, + 466, + 547, + 522 + ], + "type": "text", + "content": " percentile)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 369, + 379, + 419, + 456 + ], + "blocks": [ + { + "bbox": [ + 369, + 379, + 419, + 456 + ], + "lines": [ + { + "bbox": [ + 369, + 379, + 419, + 456 + ], + "spans": [ + { + "bbox": [ + 369, + 379, + 419, + 456 + ], + "type": "image", + "image_path": "5a3d4933d5d1871320d25c4edbbc4f4867e9c3b859c730b2cb978dc4b2b22a6e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 428, + 380, + 479, + 456 + ], + "blocks": [ + { + "bbox": [ + 428, + 380, + 479, + 456 + ], + "lines": [ + { + "bbox": [ + 428, + 380, + 479, + 456 + ], + "spans": [ + { + "bbox": [ + 428, + 380, + 479, + 456 + ], + "type": "image", + "image_path": "aeb845d270390441afd941d99b6903e97e9558800b639219a6caafb5cbbae736.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 486, + 379, + 536, + 456 + ], + "blocks": [ + { + "bbox": [ + 486, + 379, + 536, + 456 + ], + "lines": [ + { + "bbox": [ + 486, + 379, + 536, + 456 + ], + "spans": [ + { + "bbox": [ + 486, + 379, + 536, + 456 + ], + "type": "image", + "image_path": "dc24cc99a507b56267afb010ba396393b97ff4687ae1a4e044d2c5864c0ccee6.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 543, + 547, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 543, + 547, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 543, + 547, + 604 + ], + "type": "text", + "content": "the density of each. The query location is transformed to the local space of the given Gaussian's joint " + }, + { + "bbox": [ + 304, + 543, + 547, + 604 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 543, + 547, + 604 + ], + "type": "text", + "content": " at time-step " + }, + { + "bbox": [ + 304, + 543, + 547, + 604 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 543, + 547, + 604 + ], + "type": "text", + "content": " using the world-to-local transformation matrix " + }, + { + "bbox": [ + 304, + 543, + 547, + 604 + ], + "type": "inline_equation", + "content": "\\theta_{t,j}^{-1}" + }, + { + "bbox": [ + 304, + 543, + 547, + 604 + ], + "type": "text", + "content": ", rigidly attaching the Gaussians to the underlying skeleton and facilitating animation," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 373, + 612, + 545, + 646 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 612, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 373, + 612, + 545, + 646 + ], + "type": "interline_equation", + "content": "\\mathbf {G} (\\mathbf {x}) = \\sum_ {i = 0} ^ {J \\times K} \\mathbf {G} _ {i} \\left(\\theta_ {t, j} ^ {- 1} \\mathbf {x}\\right). \\tag {3}", + "image_path": "1f10739a13c6b5144daccfc981f3e50cdf8790413e2cc6f0d82b46eac9199ad4.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "We jointly fit the parameters of the Gaussian density model to the neural field by minimizing " + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{gDensity}}" + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": ", the L2 error between the density function " + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{G}(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": " and the target neural density field at query location " + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": ". We detach the gradients of the neural density field to optimize the Gaussians, fitting the" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21000" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 71, + 282, + 152 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 282, + 152 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 282, + 152 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 282, + 152 + ], + "type": "image", + "image_path": "7f25f62062fb066ffb4a0b34fd80565f928d512efda55cac0198afe4b0e9dff6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 159, + 289, + 237 + ], + "lines": [ + { + "bbox": [ + 46, + 159, + 289, + 237 + ], + "spans": [ + { + "bbox": [ + 46, + 159, + 289, + 237 + ], + "type": "text", + "content": "Figure 4. 3D Anisotropic Gaussian Raytracing. a) A cross-section of a 3D anisotropic Gaussian with rays passing through the Gaussian. b) The computed 1D Gaussians resulting from our derivation in Section 3.2 (colored solid), compared to sampling the 3D Gaussian directly (dashed), with their exact match validating the correctness. c) The transmittance along each ray which is used as the shadow map value." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 259, + 285, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 259, + 285, + 271 + ], + "spans": [ + { + "bbox": [ + 47, + 259, + 285, + 271 + ], + "type": "text", + "content": "Gaussians to the neural field and not the other way around." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 287, + 287, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 287, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 46, + 287, + 287, + 419 + ], + "type": "text", + "content": "Gaussian Ray Tracing. Figure 4 shows how casting a ray, " + }, + { + "bbox": [ + 46, + 287, + 287, + 419 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 287, + 287, + 419 + ], + "type": "text", + "content": ", with ray origin " + }, + { + "bbox": [ + 46, + 287, + 287, + 419 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_o \\in \\mathbb{R}^{3 \\times 1}" + }, + { + "bbox": [ + 46, + 287, + 287, + 419 + ], + "type": "text", + "content": " and direction " + }, + { + "bbox": [ + 46, + 287, + 287, + 419 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_d \\in \\mathbb{R}^{3 \\times 1}" + }, + { + "bbox": [ + 46, + 287, + 287, + 419 + ], + "type": "text", + "content": ", through a 3D anisotropic Gaussian results in a 1D Gaussian density along the ray. Through the Gaussian body model, this equates to a sum of 1D Gaussians for which analytic integrals can be computed. The amount of occlusion these rays experience is equal to the sum of the integrals of each of the 1D Gaussians across the rays. The transmittance value, " + }, + { + "bbox": [ + 46, + 287, + 287, + 419 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 46, + 287, + 287, + 419 + ], + "type": "text", + "content": ", used as the shadow map value, " + }, + { + "bbox": [ + 46, + 287, + 287, + 419 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 287, + 287, + 419 + ], + "type": "text", + "content": ", is the exponential of the negative integral from the start of the ray, " + }, + { + "bbox": [ + 46, + 287, + 287, + 419 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 46, + 287, + 287, + 419 + ], + "type": "text", + "content": " to the length of the ray, " + }, + { + "bbox": [ + 46, + 287, + 287, + 419 + ], + "type": "inline_equation", + "content": "t = l" + }, + { + "bbox": [ + 46, + 287, + 287, + 419 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 99, + 429, + 287, + 462 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 429, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 99, + 429, + 287, + 462 + ], + "type": "interline_equation", + "content": "s = \\mathcal {T} _ {r} = \\exp \\left[ - \\sum_ {i = 0} ^ {J \\times K} \\int_ {0} ^ {l} \\mathbf {G} _ {i} ^ {r} \\right]. \\tag {4}", + "image_path": "cf38c6b2f416dc8af0ec755ec0aa1b1ef2b1c528f7ba1c95897fb26a756b1d08.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 470, + 287, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 470, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 470, + 287, + 518 + ], + "type": "inline_equation", + "content": "G_{i}^{r}" + }, + { + "bbox": [ + 46, + 470, + 287, + 518 + ], + "type": "text", + "content": " is the 1D Gaussian created by the ray, " + }, + { + "bbox": [ + 46, + 470, + 287, + 518 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 470, + 287, + 518 + ], + "type": "text", + "content": ", going through the 3D anisotropic Gaussian, " + }, + { + "bbox": [ + 46, + 470, + 287, + 518 + ], + "type": "inline_equation", + "content": "G_{i}" + }, + { + "bbox": [ + 46, + 470, + 287, + 518 + ], + "type": "text", + "content": " with mean " + }, + { + "bbox": [ + 46, + 470, + 287, + 518 + ], + "type": "inline_equation", + "content": "\\mu \\in \\mathbb{R}^{3\\times 1}" + }, + { + "bbox": [ + 46, + 470, + 287, + 518 + ], + "type": "text", + "content": " and precision matrix " + }, + { + "bbox": [ + 46, + 470, + 287, + 518 + ], + "type": "inline_equation", + "content": "\\Sigma^{-1}\\in \\mathbb{R}^{3\\times 3}" + }, + { + "bbox": [ + 46, + 470, + 287, + 518 + ], + "type": "text", + "content": ". We derive in the supplemental how the 1D Gaussian's density function takes the form" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 528, + 287, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 528, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 104, + 528, + 287, + 555 + ], + "type": "interline_equation", + "content": "G _ {i} ^ {r ^ {s}} = \\bar {\\mathcal {C}} \\cdot \\exp \\left(- \\frac {(\\bar {\\mu} - x) ^ {2}}{2 \\bar {\\sigma} ^ {2}}\\right), \\tag {5}", + "image_path": "fdb99e9008d6a1954329549298177e78d1666be38cad66a9cfb408a4d0542ce4.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 559, + 75, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 559, + 75, + 570 + ], + "spans": [ + { + "bbox": [ + 47, + 559, + 75, + 570 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 578, + 287, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 578, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 578, + 287, + 669 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\bar {\\mathcal {C}} = \\mathcal {C} \\exp \\left(- 0. 5 \\left(\\left(\\mu - \\mathbf {r} _ {o}\\right) ^ {T} \\Sigma^ {- 1} (\\mu - \\mathbf {r} _ {o}) - \\frac {\\bar {\\mu} ^ {2}}{\\bar {\\sigma} ^ {2}}\\right)\\right), \\\\ \\bar {\\mu} = \\frac {\\mathbf {r} _ {d} ^ {T} \\Sigma^ {- 1} (\\boldsymbol {\\mu} - \\mathbf {r} _ {o})}{\\mathbf {r} _ {d} ^ {T} \\Sigma^ {- 1} \\mathbf {r} _ {d}}, \\text {a n d} \\\\ \\bar {\\sigma} = \\sqrt {\\frac {1}{\\mathbf {r} _ {d} ^ {T} \\Sigma^ {- 1} \\mathbf {r} _ {d}}}. \\tag {6} \\\\ \\end{array}", + "image_path": "c1138b0ab1ba304c8daf4fcbfd6626cc33cdf589630728d65527046bb3c65ce7.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "This formula is more complex than in [30], as it now accounts for anisotropic Gaussians with an arbitrary covariance instead of isotropic Gaussians. The comparison to" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "text", + "content": "sampling the 3D Gaussian in Figure 4 validates their correctness. It also lets us compute the cumulative density function (CDF) analytically, thereby avoiding the sampling in classical NeRFs," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 339, + 124, + 545, + 151 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 124, + 545, + 151 + ], + "spans": [ + { + "bbox": [ + 339, + 124, + 545, + 151 + ], + "type": "interline_equation", + "content": "\\int_ {0} ^ {x} \\mathbf {G} _ {i} ^ {r ^ {s}} = \\bar {\\mathcal {C}} \\cdot 0. 5 \\cdot \\left(1 + \\operatorname {e r f} \\left(\\frac {x - \\bar {\\mu}}{\\bar {\\sigma} \\sqrt {2}}\\right)\\right). \\tag {7}", + "image_path": "8a5cafe5bf057363ddc057e62f88ca859a287622bad3c19fa4016ec4c7788664.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 155, + 545, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 155, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 155, + 545, + 204 + ], + "type": "text", + "content": "Together with Equation 4, this integral computes the shadow " + }, + { + "bbox": [ + 304, + 155, + 545, + 204 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 155, + 545, + 204 + ], + "type": "text", + "content": " when applied to the secondary ray with origin " + }, + { + "bbox": [ + 304, + 155, + 545, + 204 + ], + "type": "inline_equation", + "content": "r_o^s" + }, + { + "bbox": [ + 304, + 155, + 545, + 204 + ], + "type": "text", + "content": ", as the point on the subject's surface computed from the depth map " + }, + { + "bbox": [ + 304, + 155, + 545, + 204 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 304, + 155, + 545, + 204 + ], + "type": "text", + "content": ", and direction " + }, + { + "bbox": [ + 304, + 155, + 545, + 204 + ], + "type": "inline_equation", + "content": "r_d^s" + }, + { + "bbox": [ + 304, + 155, + 545, + 204 + ], + "type": "text", + "content": " towards the light." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 209, + 391, + 221 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 209, + 391, + 221 + ], + "spans": [ + { + "bbox": [ + 306, + 209, + 391, + 221 + ], + "type": "text", + "content": "3.3. Optimization" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 228, + 545, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 545, + 334 + ], + "type": "text", + "content": "In addition to the introduced reconstruction loss " + }, + { + "bbox": [ + 304, + 228, + 545, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{RGB}}" + }, + { + "bbox": [ + 304, + 228, + 545, + 334 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 228, + 545, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Eikonal}}" + }, + { + "bbox": [ + 304, + 228, + 545, + 334 + ], + "type": "text", + "content": " for SDF regularization as in [12], and Gaussian fitting " + }, + { + "bbox": [ + 304, + 228, + 545, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{gDensity}}" + }, + { + "bbox": [ + 304, + 228, + 545, + 334 + ], + "type": "text", + "content": ", we regularize the training with i) a " + }, + { + "bbox": [ + 304, + 228, + 545, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{mask}} = |\\hat{\\rho} - \\rho|" + }, + { + "bbox": [ + 304, + 228, + 545, + 334 + ], + "type": "text", + "content": " that regularizes density by minimizing the difference between integrated accumulation, " + }, + { + "bbox": [ + 304, + 228, + 545, + 334 + ], + "type": "inline_equation", + "content": "\\hat{\\rho}" + }, + { + "bbox": [ + 304, + 228, + 545, + 334 + ], + "type": "text", + "content": ", and the foreground mask, " + }, + { + "bbox": [ + 304, + 228, + 545, + 334 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 304, + 228, + 545, + 334 + ], + "type": "text", + "content": ", ii) " + }, + { + "bbox": [ + 304, + 228, + 545, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{amb}} = ||\\hat{\\mathbf{L}}_{\\mathrm{amb}} - 0.1||^2" + }, + { + "bbox": [ + 304, + 228, + 545, + 334 + ], + "type": "text", + "content": " preferring small ambient light values, iii) " + }, + { + "bbox": [ + 304, + 228, + 545, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{gSigma}}" + }, + { + "bbox": [ + 304, + 228, + 545, + 334 + ], + "type": "text", + "content": " that prevent too large or small Gaussians, and iv) " + }, + { + "bbox": [ + 304, + 228, + 545, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{gMean}}" + }, + { + "bbox": [ + 304, + 228, + 545, + 334 + ], + "type": "text", + "content": " that pulls Gaussians closer to the center of the bones." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 335, + 545, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 545, + 431 + ], + "type": "text", + "content": "Training proceeds in three stages. In stage I, the reconstruction loss is replaced with one that encourages predicting gray inside the silhouette, to learn a rough body shape without illumination effects. In stage II, " + }, + { + "bbox": [ + 304, + 335, + 545, + 431 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{gDensity}}" + }, + { + "bbox": [ + 304, + 335, + 545, + 431 + ], + "type": "text", + "content": " and its regularizers are introduced, allowing the Gaussian density model to fit. Finally, in stage III, the " + }, + { + "bbox": [ + 304, + 335, + 545, + 431 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{RGB}}" + }, + { + "bbox": [ + 304, + 335, + 545, + 431 + ], + "type": "text", + "content": " takes over to optimize the light direction and learn the albedo. Additional training details are provided in the supplemental." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 441, + 359, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 441, + 359, + 452 + ], + "spans": [ + { + "bbox": [ + 306, + 441, + 359, + 452 + ], + "type": "text", + "content": "4. Results" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 460, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 556 + ], + "type": "text", + "content": "We evaluate our method on synthetic sequences, as done in prior work [16]. However, this does not test performance in real world conditions. Hence, we captured a new dataset in direct sunlight and compare to the most closely related baselines, showing significantly improved relightable body models. The supplemental video and supplemental document provide additional qualitative comparisons, including relighting with HDRi environment maps." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "Synthetic datasets. We test our model on the RANA dataset [16], and further create our own synthetic sequence by obtaining a textured mesh of a subject with a 3D full body scanner (VITUS 3D Body Scanner). A Blender [8] cloth simulation was applied to a shirt over the scan and the character was automatically rigged and animated using Mixamo [2]. We use the 'swing-dance' animation as the driving motion as it contains a variety of poses from all body angles. Four (3 train, 1 test) cameras are placed around the subject at 90 degrees from each other. A directional light source illuminates the scene with a slight ambient contribution such that the shadowed areas were not fully black. The" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21001" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": "annimations are rendered using the Cycles render engine. In addition, ground-truth pose and segmentation masks are exported. We split the dataset into 600/114 images for the train/test sets, with the test set including 57 novel view images and 57 novel poses, out of which 15 have a strong hard cast shadows that we test separately." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 161, + 289, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 161, + 289, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 161, + 289, + 281 + ], + "type": "text", + "content": "Outdoor sunlight dataset. We recorded two sequences of real human motion in an outdoor scene during a sunny day. This setting has largely been unexplored in neural body models and few publicly available datasets are available. We capture the data using 3 cameras (Canon EOS R8, Canon EOS 70D, iPhone12) and obtain SMPL estimates using EasyMocap [1,10,11]. Segmentation masks are obtained using the Segment Anything Model (SAM) [19]. We divide the frames into 600/200 images for the train/test splits, using all three cameras for training." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 298, + 289, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 298, + 289, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 298, + 289, + 430 + ], + "type": "text", + "content": "Baselines We evaluate our method using the hard illumination dataset against Relighting4D [7]. Due to code being unavailable, we were not able to compare against RANA [16] and Xu et al. [43] using our datasets. We instead quantitatively compare albedo estimates between our model and RANA on their dataset as these results were kindly made available. We further compare our work with other template-less neural body models [36, 37], highlighting the drawback when not explicitly modeling lighting. DANBO [36] is our neural field backbone. NPC [37] forms the current state-of-the-art template-less neural character model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 439, + 228, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 439, + 228, + 452 + ], + "spans": [ + { + "bbox": [ + 47, + 439, + 228, + 452 + ], + "type": "text", + "content": "4.1. Albedo Estimates (training poses)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 457, + 287, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 457, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 46, + 457, + 287, + 566 + ], + "type": "text", + "content": "We ran the official implementation of Relighting4D (R4D) [7] on our outdoor dataset, providing the same segmentation masks and SMPL body model as to our method (our method only uses the skeleton, not the surface). As the first stage of R4D is NeuralBody [28] which does not take shadowing into account, it produces dark floaters in the space to approximate the hard shadow, hindering their subsequent relighting module from estimating shadow and shading correctly as seen in Figure 5a." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 566, + 288, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 566, + 288, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 566, + 288, + 651 + ], + "type": "text", + "content": "To compare against RANA [16], we run our method on subject 1 of their synthetic dataset. Even with the dataset being monocular (light comes from the same direction relative to the camera), our method was able to accurately estimate the light direction (error of 9.9 degrees) and obtain albedo estimates with fewer lighting artifacts compared to RANA, see the back of the left leg in Figure 5b." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 658, + 242, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 242, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 242, + 671 + ], + "type": "text", + "content": "4.2. Novel-Pose Rendering with Shadows" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "content": "In this setting, the camera and illumination are unchanged and only novel-poses are tested. These poses created new shadow casts that resulted in large appearance" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 310, + 71, + 547, + 150 + ], + "blocks": [ + { + "bbox": [ + 310, + 71, + 547, + 150 + ], + "lines": [ + { + "bbox": [ + 310, + 71, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 310, + 71, + 547, + 150 + ], + "type": "image", + "image_path": "75c81702be88068639ccb1ec8aeee8d3f81014ec416e6bc25774cda6896999e4.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 156, + 547, + 236 + ], + "blocks": [ + { + "bbox": [ + 309, + 156, + 547, + 236 + ], + "lines": [ + { + "bbox": [ + 309, + 156, + 547, + 236 + ], + "spans": [ + { + "bbox": [ + 309, + 156, + 547, + 236 + ], + "type": "image", + "image_path": "90d273132a8b3b9cd7032421bcf11795e193ea41b3191fb353453e890f1a07b0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 243, + 547, + 289 + ], + "lines": [ + { + "bbox": [ + 304, + 243, + 547, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 547, + 289 + ], + "type": "text", + "content": "Figure 5. Albedo Estimation. Our method can better separate shadows and lighting from training images to obtain better albedo estimates without lighting artifacts compared to a) Relighting4D [7] and b) RANA [16]." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 308, + 545, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 308, + 545, + 405 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 545, + 405 + ], + "type": "text", + "content": "changes distant to the changed body part. As expected, existing methods (NPC, DANBO) overfit the training poses and the shadows created in the novel poses are highly inaccurate. Figure 6 shows how for frames that had body parts casting shadows on other regions, our method produced more accurate shadows. Table 2 quantifies the gains across novel poses and Table 3 across the subset of the novel poses that has a shadow cast across the body." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 414, + 545, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 414, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 304, + 414, + 545, + 480 + ], + "type": "text", + "content": "Table 2. Novel-pose synthesis (all test frames). Our Gaussian Shadow Casting model achieves consistently better PSNR scores for novel pose renderings as it properly models the hard shadows cast by the limbs in novel positions. Existing methods only shine in perceptual metrics (SSIM and LPIPS) as these normalize contrast and hence lessen the impact of proper shadows and shading." + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 309, + 489, + 543, + 529 + ], + "blocks": [ + { + "bbox": [ + 309, + 489, + 543, + 529 + ], + "lines": [ + { + "bbox": [ + 309, + 489, + 543, + 529 + ], + "spans": [ + { + "bbox": [ + 309, + 489, + 543, + 529 + ], + "type": "table", + "html": "
Synthetic (N=57)Real S1 (N=200)Real S2 (N=200)Average
PSNR↑SSIM↑LPIPS↓PSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPS
DANBO [36]17.520.7560.19516.570.5990.32817.690.5880.32517.260.6480.283
NPC [37]17.570.7580.18816.330.5900.33417.470.5750.32817.120.6410.283
Ours22.040.8290.16617.570.5920.35618.290.5770.35119.300.6660.291
", + "image_path": "6c161d61c7ca150f7aaaf5b48272356794af5200c132d67c8452506f7fcb8bfd.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 550, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 550, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 550, + 545, + 605 + ], + "type": "text", + "content": "Table 3. Novel-pose synthesis (subset of test set with observed self-casting shadows). Our Gaussian Shadow Casting renders novel poses with strong hard shadows well. Our scores drop marginally on these hard frames compared to all frames in Tab. 2, while the baselines drop significantly." + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 308, + 614, + 543, + 654 + ], + "blocks": [ + { + "bbox": [ + 308, + 614, + 543, + 654 + ], + "lines": [ + { + "bbox": [ + 308, + 614, + 543, + 654 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 543, + 654 + ], + "type": "table", + "html": "
Synthetic (N=15)Real S1 (N=41)Real S2 (N=36)Average
PSNR↑SSIM↑LPIPS↓PSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPS
DANBO [36]17.780.7400.20915.110.5590.35416.550.5470.35316.480.6150.305
NPC [37]17.810.7410.20114.880.5530.35716.510.5380.35516.400.6110.304
Ours22.130.8210.17516.880.5720.36517.400.5440.37118.810.6460.303
", + "image_path": "7a0a511a089fb45a76deca6b5e7130cfbf162e642cb396c9288b9b6c9a308b76.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 666, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 547, + 715 + ], + "type": "text", + "content": "For the synthetic sequence, improvements were consistent across all three metrics. In the real outdoor sequence, all methods attain a lower quality as the cameras are spaced further apart and the segmentation masks and 3D input" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21002" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 150, + 229 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 150, + 229 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 150, + 229 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 150, + 229 + ], + "type": "image", + "image_path": "24849be7be555d52da5bd2c38741ac76a5b83e6c3124912506f4f23e0509bdbb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 58, + 237, + 534, + 249 + ], + "lines": [ + { + "bbox": [ + 58, + 237, + 534, + 249 + ], + "spans": [ + { + "bbox": [ + 58, + 237, + 534, + 249 + ], + "type": "text", + "content": "Figure 6. Novel Pose Rendering. Our method can more accurately reproduce the shadow in novel poses compared to the baselines." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 160, + 72, + 252, + 229 + ], + "blocks": [ + { + "bbox": [ + 160, + 72, + 252, + 229 + ], + "lines": [ + { + "bbox": [ + 160, + 72, + 252, + 229 + ], + "spans": [ + { + "bbox": [ + 160, + 72, + 252, + 229 + ], + "type": "image", + "image_path": "88056b478e35e01ec23792f69602136773c892245f7f28bd4614a5f193e433b3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 261, + 72, + 352, + 229 + ], + "blocks": [ + { + "bbox": [ + 261, + 72, + 352, + 229 + ], + "lines": [ + { + "bbox": [ + 261, + 72, + 352, + 229 + ], + "spans": [ + { + "bbox": [ + 261, + 72, + 352, + 229 + ], + "type": "image", + "image_path": "b2490e5bcf56804ccf59a682b76ed502ff65daea80d2e837381bec0ecf3bac95.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 353, + 72, + 443, + 229 + ], + "blocks": [ + { + "bbox": [ + 353, + 72, + 443, + 229 + ], + "lines": [ + { + "bbox": [ + 353, + 72, + 443, + 229 + ], + "spans": [ + { + "bbox": [ + 353, + 72, + 443, + 229 + ], + "type": "image", + "image_path": "bdf72406ac60669a6f155216bbdaa5f2e019d213e5055dfc26b4d0ddb52dcf99.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 454, + 72, + 545, + 229 + ], + "blocks": [ + { + "bbox": [ + 454, + 72, + 545, + 229 + ], + "lines": [ + { + "bbox": [ + 454, + 72, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 454, + 72, + 545, + 229 + ], + "type": "image", + "image_path": "de89f37aa0141bdba367edb70c5ac3df779793a01ad914e77ea3db8f1b11ac46.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 270, + 287, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 287, + 365 + ], + "type": "text", + "content": "pose, estimated with off-the-shelf 2D pose detection and lifting methods, are less reliable. Nevertheless, Figure 6 shows how our model can accurately optimize the light direction and predict realistic shadows, including on the ground. To map shadows to the ground, we estimate the ground plane from the reconstructed foot positions and cast the Gaussian shadow on it by modulating the static background with the ground shadow map." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 366, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 366, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 366, + 287, + 498 + ], + "type": "text", + "content": "Table 2 and Table 3 show that our method consistently improves the PSNR while the perceptual metrics SSIM [41] improves only in one and the baselines perform better for LPIPS [47]. This lower performance in perceptual metrics is expected because these metrics normalize for brightness and contrast differences, thereby lessening the importance of producing proper shading and shadowing. In addition, the texture and geometry detail of our method is slightly lower, which we attribute to the separation into shading and albedo imposing additional constraints, thereby leading to slightly less detailed reconstructions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 503, + 192, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 503, + 192, + 516 + ], + "spans": [ + { + "bbox": [ + 47, + 503, + 192, + 516 + ], + "type": "text", + "content": "4.3. Render Time Comparison" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 521, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 641 + ], + "type": "text", + "content": "Table 4 lists the render time of our baseline compared to our full model. Casting shadows with GSC has minimal overhead (0.3s for one ray, only " + }, + { + "bbox": [ + 46, + 521, + 287, + 641 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 46, + 521, + 287, + 641 + ], + "type": "text", + "content": " of the entire render time), enabling efficient training alongside NeRF optimization. Casting a shadow with the NeRF baseline requires processing twice the number of samples by the NeRF. The deferred shading model creates one occlusion ray and each of these secondary rays requires a similar number of samples as for the primary ray. Already with a single light source, this increases runtime by " + }, + { + "bbox": [ + 46, + 521, + 287, + 641 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 46, + 521, + 287, + 641 + ], + "type": "text", + "content": ", a ten-fold difference to GSC." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 659, + 235, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 659, + 235, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 659, + 235, + 672 + ], + "type": "text", + "content": "4.4. Relighting with Environment Maps" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "The shadow computation not only benefits training time but also enables computing irradiance maps for environment maps. Figure 1 shows relighting with two different" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 339, + 268, + 514, + 350 + ], + "blocks": [ + { + "bbox": [ + 339, + 268, + 514, + 350 + ], + "lines": [ + { + "bbox": [ + 339, + 268, + 514, + 350 + ], + "spans": [ + { + "bbox": [ + 339, + 268, + 514, + 350 + ], + "type": "table", + "html": "
Methodrender time [s]
DANBO + DS17.13
DANBO + DS + GSC17.47
DANBO + DS + NeRFSC21.4
DANBO + DS + GSC-HDRi-820.70
DANBO + DS + GSC-HDRi-1623.57
DANBO + DS + GSC-HDRi-3229.49
DANBO + DS + GSC-HDRi-6441.22
", + "image_path": "c87845d60654d43ca76b222d41600c9ddf5aab3903e3804c5dcd404cd238bd75.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 359, + 545, + 426 + ], + "lines": [ + { + "bbox": [ + 305, + 359, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 305, + 359, + 545, + 426 + ], + "type": "text", + "content": "Table 4. Render time. The overhead of Gaussian Shadow Casting (GSC) is minimal on DANBO with diffuse shading (DANBO + DS) and enables casting many rays (64 for GSC-HDRi-64). By contrast, NeRF shadow casting (NeRFSC) doubles the runtime with every light source, making training prohibitively slow and HDRi relighting impractical." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 447, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 447, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 447, + 545, + 555 + ], + "type": "text", + "content": "HDRi maps (obtained from Poly Haven [14]) by casting 64 secondary light rays towards the environment map for each pixel through importance sampling. In both cases, the bright sun casts a strong shadow while the colored light from the environment leads to natural shading that matches the character with the environment. This enables placing the reconstructed characters into new environments and giving them a natural and consistent look with respect to the rest of the scene while still containing cast shadows." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 555, + 545, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 545, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 545, + 615 + ], + "type": "text", + "content": "Our method is plug-and-play with other neural body models due to the deferred rendering approach. It can be used with higher quality volumetric neural models without degraded quality when training on uniformly lit data and using GSC for relighting as shown in Figure 7." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 623, + 400, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 400, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 400, + 635 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "We test a variety of implementation details in our model, including using only diffuse shading on top of DANBO (DANBO + Diff. Shading), only using the Gaussians to cast shadows (Ours w/o Diff. Shading), providing ground truth lighting (Ours-GT Light), and detaching the normals prior to the diffuse shading (Ours-Detached Normals). The" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21003" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 289, + 154 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 289, + 154 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 289, + 154 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 289, + 154 + ], + "type": "image", + "image_path": "1a86ed186dc1e803ed89d0289d0915b6d8875edc7602814e767993a323d6905b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 163, + 288, + 206 + ], + "lines": [ + { + "bbox": [ + 46, + 163, + 288, + 206 + ], + "spans": [ + { + "bbox": [ + 46, + 163, + 288, + 206 + ], + "type": "text", + "content": "Figure 7. HDRi Relighting on MonoPerfCap using NPC [37] as the backbone. Our Gaussian relighting method can be utilized at inference with higher quality models on data that is uniformly lit." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 229, + 288, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 229, + 288, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 229, + 288, + 264 + ], + "type": "text", + "content": "results of which can be seen in Table 5, which shows that each of our contributions improves reconstruction quality at test time." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 50, + 293, + 284, + 365 + ], + "blocks": [ + { + "bbox": [ + 88, + 274, + 246, + 285 + ], + "lines": [ + { + "bbox": [ + 88, + 274, + 246, + 285 + ], + "spans": [ + { + "bbox": [ + 88, + 274, + 246, + 285 + ], + "type": "text", + "content": "Table 5. Ablation on Synthetic Sequence." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 293, + 284, + 365 + ], + "lines": [ + { + "bbox": [ + 50, + 293, + 284, + 365 + ], + "spans": [ + { + "bbox": [ + 50, + 293, + 284, + 365 + ], + "type": "table", + "html": "
TrainingNovel PoseNovel View
PSNR↑SSIM↑LPIPS↓PSNRSSIMLPIPSPSNRSSIMLPIPS
DANBO27.660.9130.13317.520.7560.19518.850.7730.179
DANBO + Diff. Shading24.820.8600.18718.600.7850.20824.400.8750.167
Ours w/o Diff. Shading27.170.8790.16121.220.8020.17825.080.8670.163
Ours-Detached Normals25.900.8630.18220.430.8140.19226.180.8930.158
Ours-GT Light25.220.8610.18021.230.8300.18426.680.8950.155
Ours26.600.8760.16522.300.8270.17627.320.8820.154
", + "image_path": "933e2e4ba7842ac96de6765ee76daec877f152d87c6c649709612a9750d20590.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 379, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 379, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 379, + 287, + 521 + ], + "type": "text", + "content": "Diffuse Shading. Shadowing alone does not account for accurate shading based on how incident the light hits the surface. Moreover, the Gaussians cast long-range shadows, but their smooth and low-resolution approximation to the NeRF's density prevents them from representing finer details such as small extremities (nose, fingers). As a result, finer shading details are missed as seen in Figure 8c. On the other hand, using only shading (DANBO + Diff. Shading) already reduces texture detail quality as seen by the training scores, but achieves improved performance in novel pose and novel view conditions. However, the missing cast shadows play the most significant role in improving test scores." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 522, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 617 + ], + "type": "text", + "content": "Detached Normals. We compare results between a model where network gradients could, Figure 8d, and could not, Figure 8b, backpropagate through the surface normals used in the diffuse shading to see whether or not artifacts in the shading would smoothen out the geometry. We find that the surface is indeed affected by the gradients backpropagating through the diffuse computation and observe a smoother geometry reconstruction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "type": "text", + "content": "GT Light. Our method is able to fit the direction of the light source and the ambient intensity with little user input. We observe accurate light recovery when the light is initialized randomly, e.g. when coming from the back the angle error is only 1.36 degrees on our synthetic sequence. We found providing the ground truth light direction did not improve results and may have hindered the model due to the added constraints, Figure 8a." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 71, + 402, + 156 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 402, + 156 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 402, + 156 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 402, + 156 + ], + "type": "image", + "image_path": "426c13ca53715a663c15447c85486e59497cd5ded30a493cddc06fd8fa6c54a5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 165, + 546, + 220 + ], + "lines": [ + { + "bbox": [ + 305, + 165, + 546, + 220 + ], + "spans": [ + { + "bbox": [ + 305, + 165, + 546, + 220 + ], + "type": "text", + "content": "Figure 8. Ablation Comparisons (novel-pose). a) The model trained with the groundtruth light direction. b) The model trained while detaching the gradients from the surface normals during diffuse shading. c) The model trained without diffuse shading. d) Our full model." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 414, + 70, + 499, + 156 + ], + "blocks": [ + { + "bbox": [ + 414, + 70, + 499, + 156 + ], + "lines": [ + { + "bbox": [ + 414, + 70, + 499, + 156 + ], + "spans": [ + { + "bbox": [ + 414, + 70, + 499, + 156 + ], + "type": "image", + "image_path": "23dd5afffd1a5736a3f73440504cf635069e0fda9f66e42e5f7ed190d0d36f30.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 504, + 70, + 546, + 152 + ], + "blocks": [ + { + "bbox": [ + 504, + 70, + 546, + 152 + ], + "lines": [ + { + "bbox": [ + 504, + 70, + 546, + 152 + ], + "spans": [ + { + "bbox": [ + 504, + 70, + 546, + 152 + ], + "type": "image", + "image_path": "e95aa5d9484d0a0707f1b75fec64992648d5ec513cedc55683e10772d1f0e3d2.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 245, + 383, + 256 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 245, + 383, + 256 + ], + "spans": [ + { + "bbox": [ + 306, + 245, + 383, + 256 + ], + "type": "text", + "content": "4.6. Limitations" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 265, + 545, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 265, + 545, + 372 + ], + "spans": [ + { + "bbox": [ + 304, + 265, + 545, + 372 + ], + "type": "text", + "content": "The Gaussian cast shadows model long-range effects, such as the arm casting a shadow on the leg but the smooth Gaussians lack high frequency details. This is a minor drawback since the diffuse shading already faithfully reproduces the light intensity fall-off as the light direction becomes more incident with the surface and therefore shades the back side of small extremities (i.e. nose and fingers) well. A future extension could be to integrate mid-scale effects with screen-space ambient occlusion and shading." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 374, + 545, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 374, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 304, + 374, + 545, + 459 + ], + "type": "text", + "content": "Moreover, we noticed that disentangling color into shading and albedo, compared to the original DANBO backbone, leads to slightly lower image reconstruction metrics when shading effects are minimal. We attribute this to the additional constraints that are imposed on the model. However, the overall performance in novel light conditions is still improved significantly by our model." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 474, + 378, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 474, + 378, + 487 + ], + "spans": [ + { + "bbox": [ + 306, + 474, + 378, + 487 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 496, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 496, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 496, + 545, + 628 + ], + "type": "text", + "content": "We enabled the 3D reconstruction of human motions in uncontrolled environments with a Gaussian-based shadow model that applies to dynamic scenes and is differentiable for iterative refinement. The reconstructed characters support reposing and relighting in novel environments. They are equipped with global shadow computation, diffuse shading, geometric reconstruction, and a consistent albedo, much like hand-crafted computer graphics models would provide. The deferred lighting approach allows our method to be combined with other neural body models with efficient shadow computations." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 644, + 420, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 644, + 420, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 644, + 420, + 658 + ], + "type": "text", + "content": "6. Acknowledgements" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "content": "This work was supported in part by an NSERC Discovery Grant and the computational resources and services provided by Advanced Research Computing at The University of British Columbia." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21004" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 92, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 53, + 92, + 285, + 114 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 92, + 285, + 114 + ], + "spans": [ + { + "bbox": [ + 53, + 92, + 285, + 114 + ], + "type": "text", + "content": "[1] Easymocap - make human motion capture easier. https://github.com/zju3dv/EasyMocap, 2021.6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 116, + 242, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 116, + 242, + 126 + ], + "spans": [ + { + "bbox": [ + 53, + 116, + 242, + 126 + ], + "type": "text", + "content": "[2] Adobe. Mixamo. www MIXamo.com, 2023.5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 128, + 287, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 128, + 287, + 171 + ], + "spans": [ + { + "bbox": [ + 53, + 128, + 287, + 171 + ], + "type": "text", + "content": "[3] Thiemo Alldieck, Marcus Magnor, Bharat Lal Bhatnagar, Christian Theobalt, and Gerard Pons-Moll. Learning to reconstruct people in clothing from a single RGB camera. In CVPR, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 174, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 174, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 53, + 174, + 287, + 217 + ], + "type": "text", + "content": "[4] Timur Bagautdinov, Chenglei Wu, Tomas Simon, Fabian Prada, Takaaki Shiratori, Shih-En Wei, Weipeng Xu, Yaser Sheikh, and Jason Saragih. Driving-signal aware full-body avatars. ACM TOG (Proc. SIGGRAPH), 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 220, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 220, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 53, + 220, + 287, + 274 + ], + "type": "text", + "content": "[5] Sai Bi, Stephen Lombardi, Shunsuke Saito, Tomas Simon, Shih-En Wei, Kevyn Mcphail, Ravi Ramamoorthi, Yaser Sheikh, and Jason Saragih. Deep relightable appearance models for animatable faces. ACM TOG (Proc. SIGGRAPH), 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 276, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 276, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 53, + 276, + 287, + 319 + ], + "type": "text", + "content": "[6] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P.A. Lensch. Nerd: Neural reflectance decomposition from image collections. In ICCV, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 322, + 287, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 322, + 287, + 344 + ], + "spans": [ + { + "bbox": [ + 53, + 322, + 287, + 344 + ], + "type": "text", + "content": "[7] Zhaoxi Chen and Ziwei Liu. Relighting4d: Neural relightable human from videos. In ECCV, 2022. 2, 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 346, + 287, + 377 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 287, + 377 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 287, + 377 + ], + "type": "text", + "content": "[8] Blender Online Community. Blender - a 3d modelling and rendering package. http://www.blender.org, 2023.5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 380, + 287, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 287, + 412 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 287, + 412 + ], + "type": "text", + "content": "[9] Dawa Derksen and Dario Izzo. Shadow neural radiance fields for multi-view satellite photogrammetry. In CVPR, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 415, + 287, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 458 + ], + "type": "text", + "content": "[10] Junting Dong, Qi Fang, Wen Jiang, Yurou Yang, Hujun Bao, and Xiaowei Zhou. Fast and robust multi-person 3d pose estimation and tracking from multiple views. In TPAMI, 2021. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 460, + 287, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 287, + 493 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 287, + 493 + ], + "type": "text", + "content": "[11] Junting Dong, Qing Shuai, Yuanqing Zhang, Xian Liu, Xiaowei Zhou, and Hujun Bao. Motion capture from internet videos. In ECCV, 2020. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 495, + 287, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 495, + 287, + 528 + ], + "spans": [ + { + "bbox": [ + 48, + 495, + 287, + 528 + ], + "type": "text", + "content": "[12] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In ICML, 2020. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 530, + 287, + 562 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 530, + 287, + 562 + ], + "spans": [ + { + "bbox": [ + 48, + 530, + 287, + 562 + ], + "type": "text", + "content": "[13] Michelle Guo, Alireza Fathi, Jiajun Wu, and Thomas Funkhouser. Object-centric neural scene rendering, 2020. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 565, + 287, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 565, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 565, + 287, + 586 + ], + "type": "text", + "content": "[14] Poly Haven. Poly haven. www.polyhaven.com/HDris, 2023.7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 589, + 287, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 589, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 589, + 287, + 632 + ], + "type": "text", + "content": "[15] Jingwei Huang, Haotian Zhang, Li Yi, Thomas Funkhouser, Matthias Nießner, and Leonidas J Guibas. Texturenet: Consistent local parametrizations for learning from high-resolution signals on meshes. In CVPR, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 635, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 677 + ], + "type": "text", + "content": "[16] Umar Iqbal, Akin Caliskan, Koki Nagano, Sameh Khamis, Pavlo Molchanov, and Jan Kautz. Rana: Relightable articulated neural avatars. arXiv preprint arXiv:2212.03237, 2022. 2, 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[17] Wei Jiang, Kwang Moo Yi, Golnoosh Samei, Oncel Tuzel, and Anurag Ranjan. Neuman: Neural human radiance field from a single video. In ECCV, 2022. 1" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[18] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM TOG (Proc. SIGGRAPH), 2023. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 119, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 119, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 119, + 545, + 172 + ], + "type": "text", + "content": "[19] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 175, + 545, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 218 + ], + "type": "text", + "content": "[20] Youngjoong Kwon, Dahun Kim, Duygu Ceylan, and Henry Fuchs. Neural human performer: Learning generalizable radiance fields for human performance rendering. NeurIPS, 2021. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 221, + 545, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 221, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 307, + 221, + 545, + 255 + ], + "type": "text", + "content": "[21] Junxuan Li, Shunsuke Saito, Tomas Simon, Stephen Lombardi, Hongdong Li, and Jason Saragih. Megane: Morphable eyeglass and avatar network. In CVPR, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 256, + 545, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 256, + 545, + 299 + ], + "spans": [ + { + "bbox": [ + 307, + 256, + 545, + 299 + ], + "type": "text", + "content": "[22] Ruilong Li, Julian Tanke, Minh Vo, Michael Zollhofer, Jurgen Gall, Angjoo Kanazawa, and Christoph Lassner. Tava: Template-free animatable volumetric actors. In ECCV, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 302, + 545, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 302, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 307, + 302, + 545, + 346 + ], + "type": "text", + "content": "[23] Zhe Li, Zerong Zheng, Yuxiao Liu, Boyao Zhou, and Yebin Liu. Posevocab: Learning joint-structured pose embeddings for human avatar modeling. In ACM TOG (Proc. SIGGRAPH), 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 348, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 348, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 348, + 545, + 392 + ], + "type": "text", + "content": "[24] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. ACM TOG (Proc. SIGGRAPH Asia), 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 394, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 394, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 394, + 545, + 437 + ], + "type": "text", + "content": "[25] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. ACM TOG (Proc. SIGGRAPH Asia), 2015. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 440, + 545, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 440, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 307, + 440, + 545, + 483 + ], + "type": "text", + "content": "[26] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 486, + 545, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 486, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 307, + 486, + 545, + 517 + ], + "type": "text", + "content": "[27] Atsuhiro Noguchi, Xiao Sun, Stephen Lin, and Tatsuya Harada. Neural articulated radiance field. In ICCV, 2021. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 520, + 545, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 520, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 307, + 520, + 545, + 574 + ], + "type": "text", + "content": "[28] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In CVPR, 2021. 1, 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 578, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 578, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 307, + 578, + 545, + 609 + ], + "type": "text", + "content": "[29] Anurag Ranjan, Kwang Moo Yi, Jen-Hao Rick Chang, and Oncel Tuzel. Facelit: Neural 3d relightable faces. In CVPR, 2023. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 613, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 667 + ], + "type": "text", + "content": "[30] Helge Rhodin, Nadia Robertini, Dan Casas, Christian Richardt, Hans-Peter Seidel, and Christian Theobalt. General automatic human shape and motion capture using volumetric contour cues. In Bastian Leibe, Jiri Matas, Nicu Sebe, and Max Welling, editors, ECCV, 2016. 2, 3, 4, 5" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[31] H. Rhodin, N. Robertini, C. Richardt, H. Seidel, and C. Theobalt. A versatile scene model with differentiable visibility applied to generative pose estimation. In ICCV, 2015. 2, 3, 4" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21005" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[32] Viktor Rudnev, Mohamed Elgharib, William Smith, Lingjie Liu, Vladislav Golyanik, and Christian Theobalt. Nerf for outdoor scene relighting. In ECCV, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 149 + ], + "type": "text", + "content": "[33] Srinath Sridhar, Helge Rhodin, Hans-Peter Seidel, Antti Oulasvirta, and Christian Theobalt. Real-time hand tracking using a sum of anisotropic gaussians model. In 3DV, 2014. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 150, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 150, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 150, + 287, + 194 + ], + "type": "text", + "content": "[34] Pratul P. Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T. Barron. Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In CVPR, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 194, + 287, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 194, + 287, + 227 + ], + "spans": [ + { + "bbox": [ + 48, + 194, + 287, + 227 + ], + "type": "text", + "content": "[35] Carsten Stoll, Nils Hasler, Juergen Gall, Hans-Peter Seidel, and Christian Theobalt. Fast articulated motion tracking using a sums of gaussians body model. ICCV, 2011. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 228, + 287, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 287, + 260 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 287, + 260 + ], + "type": "text", + "content": "[36] Shih-Yang Su, Timur Bagautdinov, and Helge Rhodin. Danbo: Disentangled articulated neural body representations via graph neural networks. In ECCV, 2022. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 261, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 287, + 293 + ], + "type": "text", + "content": "[37] Shih-Yang Su, Timur Bagautdinov, and Helge Rhodin. Npc: Neural point characters from video. In ICCV, 2023. 1, 2, 6, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 293, + 287, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 293, + 287, + 336 + ], + "spans": [ + { + "bbox": [ + 48, + 293, + 287, + 336 + ], + "type": "text", + "content": "[38] Shih-Yang Su, Frank Yu, Michael Zollhöfer, and Helge Rhodin. A-nerf: Articulated neural radiance fields for learning human shape, appearance, and pose. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 338, + 287, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 287, + 370 + ], + "type": "text", + "content": "[39] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: Image synthesis using neural textures. ACM TOG (Proc. SIGGRAPH), 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 371, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 403 + ], + "type": "text", + "content": "[40] Shaofei Wang, Katja Schwarz, Andreas Geiger, and Siyu Tang. Arah: Animatable volume rendering of articulated human sdfs. In ECCV, 2022. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 404, + 287, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 287, + 448 + ], + "type": "text", + "content": "[41] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 449, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 502 + ], + "type": "text", + "content": "[42] Zian Wang, Tianchang Shen, Jun Gao, Shengyu Huang, Jacob Munkberg, Jon Hasselgren, Zan Gojcic, Wenzheng Chen, and Sanja Fidler. Neural fields meet explicit geometric representations for inverse rendering of urban scenes. In CVPR, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 503, + 287, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 503, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 48, + 503, + 287, + 546 + ], + "type": "text", + "content": "[43] Zhen Xu, Sida Peng, Chen Geng, Linzhan Mou, Zihan Yan, Jiaming Sun, Hujun Bao, and Xiaowei Zhou. Relightable and animatable neural avatar from sparse-view video, 2023. 2, 3, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 548, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 601 + ], + "type": "text", + "content": "[44] Haotian Yang, Mingwu Zheng, Wanquan Feng, Haibin Huang, Yu-Kun Lai, Pengfei Wan, Zhongyuan Wang, and Chongyang Ma. Towards practical capture of high-fidelity reconfigurable avatars. In ACM TOG (Proc. SIGGRAPH Asia), 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 602, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 634 + ], + "type": "text", + "content": "[45] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. In NeurIPS, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 635, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 679 + ], + "type": "text", + "content": "[46] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snavely. PhySG: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In CVPR, 2021. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[47] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 7" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 274 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 126 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 126 + ], + "type": "text", + "content": "[48] Xiuming Zhang, Pratul P. Srinivasan, Boyang Deng, Paul Debevec, William T. Freeman, and Jonathan T. Barron. Nefactor: Neural factorization of shape and reflectance under an unknown illumination. ACM TOG (Proc. SIGGRAPH), 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 129, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 161 + ], + "type": "text", + "content": "[49] Zerong Zheng, Han Huang, Tao Yu, Hongwen Zhang, Yandong Guo, and Yebin Liu. Structured local radiance fields for human avatar modeling. In CVPR, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 163, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 195 + ], + "type": "text", + "content": "[50] Zerong Zheng, Xiaochen Zhao, Hongwen Zhang, Boning Liu, and Yebin Liu. Avatarrex: Real-time expressive full-body avatars. ACM TOG (Proc. SIGGRAPH), 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "type": "text", + "content": "[51] Taotao Zhou, Kai He, Di Wu, Teng Xu, Qixuan Zhang, Kuixiang Shao, Wenzheng Chen, Lan Xu, and Jingyi Yu. Relightable neural human assets from multi-view gradient illuminations. In CVPR, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 241, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 274 + ], + "type": "text", + "content": "[52] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In CVPR, 2019. 4" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "21006" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Gaussian Shell Maps for Efficient 3D Human Generation/c6318955-2d79-4389-9712-ec90ab24e73b_content_list.json b/2024/Gaussian Shell Maps for Efficient 3D Human Generation/c6318955-2d79-4389-9712-ec90ab24e73b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ef8481b6917e4d5067ae364882baa4d5538878ff --- /dev/null +++ b/2024/Gaussian Shell Maps for Efficient 3D Human Generation/c6318955-2d79-4389-9712-ec90ab24e73b_content_list.json @@ -0,0 +1,1685 @@ +[ + { + "type": "text", + "text": "Gaussian Shell Maps for Efficient 3D Human Generation", + "text_level": 1, + "bbox": [ + 196, + 77, + 774, + 98 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rameen Abdal\\*1 Wang Yifan\\*1 Zifan Shi\\*1,2 Yinghao Xu $^{1}$ Ryan Po $^{1}$ Zhengfei Kuang $^{1}$ Qifeng Chen $^{2}$ Dit-Yan Yeung $^{2}$ Gordon Wetzstein $^{1}$ $^{1}$ Stanford University, USA ${}^{2}$ HKUST, Hong Kong", + "bbox": [ + 94, + 125, + 875, + 219 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/33b9ecb905bebd4160c3f5af610ac3b1bca30e445a3e1437885d0330a2a73f02.jpg", + "image_caption": [ + "Figure 1. Gaussian Shell Maps. Gaussian Shell Maps is an efficient framework for 3D human generation connecting 3D Gaussians with CNN-based generators. 3D Gaussians are anchored to \"shells\" derived from the SMPL template [37] (only two shells are visualized for clarity), and the appearance is modeled in texture space. Trained only on 2D images, we show that our method can generate diverse articulable humans in real-time with state-of-the-art quality directly in high resolution without the need for upsampling and hence avoiding aliasing artifacts." + ], + "image_footnote": [], + "bbox": [ + 91, + 253, + 875, + 409 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 494, + 313, + 510 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Efficient generation of 3D digital humans is important in several industries, including virtual reality, social media, and cinematic production. 3D generative adversarial networks (GANs) have demonstrated state-of-the-art (SOTA) quality and diversity for generated assets. Current 3D GAN architectures, however, typically rely on volume representations, which are slow to render, thereby hampering the GAN training and requiring multi-view-inconsistent 2D upsamplers. Here, we introduce Gaussian Shell Maps (GSMs) as a framework that connects SOTA generator network architectures with emerging 3D Gaussian rendering primitives using an articulable multi shell-based scaffold. In this setting, a CNN generates a 3D texture stack with features that are mapped to the shells. The latter represent inflated and deflated versions of a template surface of a digital human in a canonical body pose. Instead of rasterizing the shells directly, we sample 3D Gaussians on the shells whose attributes are encoded in the texture features. These Gaussians are efficiently and differentiably rendered. The ability to articulate the shells is important during GAN training and, at inference time, to deform a body into arbitrary user-defined poses. Our efficient rendering scheme bypasses the need for view-inconsistent upsamplers and achieves high-quality multi-view consistent renderings at a native resolu", + "bbox": [ + 75, + 537, + 472, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tion of $512 \\times 512$ pixels. We demonstrate that GSMs successfully generate 3D humans when trained on single-view datasets, including SHHQ and DeepFashion.", + "bbox": [ + 498, + 496, + 890, + 541 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Project Page: rameenabdal.github.io/GaussianShellMaps", + "bbox": [ + 500, + 542, + 880, + 556 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 584, + 632, + 599 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The ability to generate articulable three-dimensional digital humans augments traditional asset creation and animation workflows, which are laborious and costly. Such generative artificial intelligence-fueled workflows are crucial in several applications, including communication, cinematic production, and interactive gaming, among others.", + "bbox": [ + 496, + 609, + 890, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D Generative Adversarial Networks (GANs) have emerged as the state-of-the-art (SOTA) platform in this domain, enabling the generation of diverse 3D assets at interactive framereates [6, 8, 15, 23, 41, 71, 76, 76]. Most existing 3D GANs build on variants of volumetric scene representations combined with neural volume rendering [65]. However, volume rendering is relatively slow and compromises the training of a GAN, which requires tens of millions of forward rendering passes to converge [11]. Mesh-based representations building on fast differentiable rasterization have been proposed to overcome this limitation [19, 71],", + "bbox": [ + 496, + 700, + 893, + 867 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal Contribution", + "bbox": [ + 522, + 875, + 638, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Work done as a visiting student researcher at Stanford University", + "bbox": [ + 522, + 887, + 877, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9441", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "but these are not expressive enough to realistically model features like hair, clothing, or accessories, which deviate significantly from the template mesh. These limitations, which are largely imposed by a tradeoff between efficient or expressive scene representations, have been constraining the quality and resolution of existing 3D GANs. While partially compensated for by using 2D convolutional neural network (CNN)-based upsamplers [11, 22], upsampling leads to multi-view inconsistency in the form of aliasing.", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Very recently, 3D Gaussians have been introduced as a promising neural scene representation offering fast rendering speed and high expressivity [28]. While 3D Gaussians have been explored in the context of single-scene overfitting, their full potential in generative settings has yet to be unlocked. This is challenging because it is not obvious how to combine SOTA CNN-based generators [26, 27] with 3D Gaussian primitives that inherently do not exist on a regular Cartesian grid and that may vary in numbers.", + "bbox": [ + 75, + 234, + 472, + 371 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We introduce Gaussian Shell Maps (GSMs), a 3D GAN framework that intuitively connects CNN generators with 3D Gaussians used as efficient rendering primitives. Inspired by the traditional computer graphics work on shell maps [47], GSMs use the CNN generator to produce texture maps for a set of \"shell\" meshes appropriately inflated and deflated from the popular SMPL mesh template for human bodies [37]. The textures on the individual shells directly encode the properties of 3D Gaussians, which are sampled on the shell surfaces at fixed locations. The generated images are rendered using highly efficient Gaussian splatting, and articulation of these Gaussians can be naturally enabled through deforming the scaffolding shells with the SMPL model. Since 3D Gaussians have spatial extent, they represent details on, in between, and outside the discrete shells. GSMs are trained exclusively on datasets containing single-view images of human bodies, such as SHHQ [17].", + "bbox": [ + 75, + 377, + 472, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our experiments demonstrate that GSM can generate highly diverse appearances, including loose clothing and accessories, at high resolution, without an upsampler, at a state-of-the-art rendering speed of 125 FPS (or 35FPS including generation). Among various architecture design choices, multiple shells with fixed relative locations of the 3D Gaussian achieve the best results in our experiments.", + "bbox": [ + 75, + 642, + 470, + 748 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, our contributions include", + "bbox": [ + 96, + 753, + 351, + 770 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We propose a novel 3D GAN framework combining a CNN-based generator and 3D Gaussian rendering primitives using shell maps.", + "2. We demonstrate the fastest 3D GAN architecture to date, achieving real-time rendering of $512^{2}$ px without convolutional upsamplers, with image quality and diversity matching the state of the art on challenging single-view datasets of human bodies." + ], + "bbox": [ + 89, + 777, + 467, + 897 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 89, + 640, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D-Aware Generative Models. GANs and diffusion models are two very powerful generative models emerging as a result of efficient architectures and high-quality datasets [25, 48]. With the quality of image generation reaching photo-realism, the community started leveraging these SOTA generative models for 3D generation by combining neural rendering methods [64, 65] to produce high-quality multi-view consistent 3D objects from image collections. Several 3D GAN architectures have explored implicit or explicit neural volume representations for modeling 3D objects, including [2, 10, 11, 14, 22, 39, 40, 42, 57, 59, 62, 69, 70]. 3D diffusion models, on the other hand, typically use the priors encoded by pre-trained text-to-image 2D diffusion models, e.g. [4, 7-9, 12, 30, 34, 35, 44, 46, 51, 53, 66] (see this survey for more details [45]). Due to the lack of high-quality, large-scale multi-view datasets curated for specific categories like humans, the choice of a suitable generative model becomes critical. Using diffusion models to generate high-resolution multi-view consistent 3D objects is still an unsolved problem [45]. 3D GANs, on the other hand, exhibit better quality and multi-view consistency at higher resolutions that do not assume multi-view data [11, 15, 62, 71]. This motivates our choice of building an efficient representation using a 3D GAN framework.", + "bbox": [ + 496, + 114, + 893, + 479 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generative Articulated 3D Digital Humans. 3D-GAN frameworks have been proposed to generate the appearance, geometry, and identity-preserving novel views of digital humans [6, 13, 15, 23, 71, 72, 74, 76]. Most of these GANs are trained on single-view image collections [16, 33]. A popular approach is to use a neural radiance representation [6, 23, 24, 41, 74] where a canonically posed human can be articulated via deformation. Other approaches are based on meshes, where a template can be fixed or learned during the training [3, 18, 19, 63, 72]. Related to this line of work, a concurrent paper, LSV-GAN [71], offsets the SMPL template meshes into layered surfaces and composites the per-layer rasterization result to form the final rendering. While it provides a faster alternative to the volume rendering-based approaches, it can only accommodate a small offset from the SMPL mesh, which hampers diversity. Our GSM method differs from LSV-GAN as we employ 3D Gaussians [28] as primitives on the layered surfaces, which, by having a learnable spatial span, allows for larger deviation from the template mesh and can thus generate more intricate details.", + "bbox": [ + 496, + 484, + 893, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Point-Based Rendering. Earlier point-based methods efficiently render point clouds and rasterize them by fixing the size [20, 21, 54]. While efficient in terms of speed and parallel rasterization on graphics processing units [32, 56], they are not differentiable [28, 49]. To combine these methods with the neural networks and perform view synthesis,", + "bbox": [ + 496, + 809, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9442", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d56a544036a45f6dab10e514f5c5177b741721beb19d76630aaf0a4f39533bdb.jpg", + "image_caption": [ + "Figure 2. Method Overview. We propose an expressive yet highly efficient representation, Gaussian Shell Map (GSM), for 3D human generation. Combining the idea of 3D Gaussians and Shell Maps (Sec. 3), we sample 3D Gaussians on \"shells\", which are mesh layers offsetted from the SMPL template, forming a shell volume to model complex and diverse geometry and appearance; the Gaussian parameters are learned in the texture space, allowing us to leverage existing CNN-based generative architecture (Sec. 4.1). Articulation is straightforward by interpolating the deformation of the shell (Sec. 4.2). The generation is supervised by single-view 2D images using several discriminator critics, including part-specific face, hands, and feet discriminators (Sec. 4.3)." + ], + "image_footnote": [], + "bbox": [ + 89, + 89, + 883, + 296 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "recent works have developed differentiable point-based rendering techniques [1, 52, 67, 68, 73, 75, 77]. More recently, 3D Gaussian-based point splatting gained traction due to the flexibility of anisotropic covariance and density control with efficient depth sorting [28, 29]. This allows 3D Gaussian splats to handle complex scenes composed of high and low-frequency features. Relevant to human bodies, 3D Gaussians have also been used in pose estimation and tracking [38, 50, 61]. While the Gaussian primitives have been used for efficient scene reconstruction and novel view synthesis, it is not trivial to deploy Gaussians in a generative setup. To the best of our knowledge, our method is the first to propose a combination of 3D Gaussians and 3D GANs.", + "bbox": [ + 75, + 412, + 472, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Background", + "text_level": 1, + "bbox": [ + 76, + 627, + 205, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Gaussians. These point-based primitives can be differentiably and efficiently rendered using EWA (elliptical weighted average) volume splatting [78]. 3D Gaussians have recently demonstrated outstanding expressivity for 3D scene reconstruction [28], in which the Gaussian parameters, position $\\mathbf{x} \\in \\mathbb{R}^3$ , opacity $o \\in \\mathbb{R}$ , color $\\mathbf{c} \\in \\mathbb{R}^{sh}$ ( $sh$ representing the spherical harmonic coefficients), scaling $\\mathbf{s} \\in \\mathbb{R}^3$ , and rotation $\\mathbf{q} \\in \\mathbb{R}^4$ parameterized as quaternions are jointly optimized to minimize the photometric errors of the rendered images in a set of known camera views. The optimization is accompanied by adaptive control of the density, where the points are added or removed based on the density, size, and gradient of the Gaussians.", + "bbox": [ + 75, + 652, + 468, + 849 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, each Gaussian is defined as", + "bbox": [ + 96, + 851, + 364, + 867 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nG \\left(\\mathbf {x} ^ {\\prime}; \\mathbf {x}, \\Sigma\\right) = \\exp^ {- \\frac {1}{2} \\left(\\mathbf {x} ^ {\\prime} - \\mathbf {x}\\right) ^ {\\intercal} \\Sigma^ {- 1} \\left(\\mathbf {x} ^ {\\prime} - \\mathbf {x}\\right)}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 880, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\Sigma = R S S^T R^T$ is the covariance matrix parameterized by the rotation and scaling matrices $R$ and $S$ given by the quaternions $\\mathbf{q}$ and scaling $\\mathbf{s}$ .", + "bbox": [ + 498, + 411, + 890, + 457 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The image formation is governed by classic point-based $\\alpha$ -blending of overlapping Gaussians ordered from closest to farthest [31]:", + "bbox": [ + 498, + 457, + 890, + 502 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {C} = \\sum_ {i \\in \\mathcal {N}} \\mathbf {c} _ {i} ^ {\\prime} o _ {i} ^ {\\prime} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {\\mathbf {j}}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 508, + 890, + 550 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where color $\\mathbf{c}_i^{\\prime}$ and opacity $\\sigma_{i}^{\\prime}$ is computed by", + "bbox": [ + 500, + 555, + 803, + 571 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} _ {i} ^ {\\prime} = G \\left(\\mathbf {x} ^ {\\prime}; \\mathbf {x} _ {i}\\right) \\mathbf {c} _ {i} \\quad \\text {a n d} \\quad o _ {i} ^ {\\prime} = G \\left(\\mathbf {x} ^ {\\prime}; \\mathbf {x} _ {i}\\right) o _ {i}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 575, + 890, + 594 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Thanks to their ability to fit complex geometry and appearance, 3D Gaussians are gaining popularity for 3D scene reconstruction. However, deploying 3D Gaussians for generative tasks remains an unexplored topic and is challenging for several reasons. First, Gaussians are of \"Lagrangian\" nature - their sparse number and learnable positions are challenging to combine with SOTA \"Eulerian\" (i.e., grid-based) CNN generators. Second, the parameters of Gaussians are highly correlated. The same radiance field can be equally well explained by many different configurations of Gaussians, varying their locations, sizes, scales, rotations, colors, and opacities. This ambiguity makes it challenging to generalize over a distribution of objects or scenes, which generative methods do.", + "bbox": [ + 496, + 599, + 892, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Shell Maps. Our representation is related to Shell Maps [14, 47, 58], a technique in computer graphics designed to model near-surface details. Shell maps use 3D texture maps to store the fine-scale features in a shell-like volume close to a given base surface, typically represented as a triangle mesh. In essence, they extend UV", + "bbox": [ + 496, + 809, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9443", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/96ad5d1d30cf15b47c61c8e54a497541a9817998667630ae96b6559201797931.jpg", + "image_caption": [ + "Figure 3. Novel Views and Animation. We can render a generated identity in novel views and articulate it for animations." + ], + "image_footnote": [], + "bbox": [ + 84, + 85, + 883, + 202 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4ea0818e6b5c2737d2ca2baf51115b51c91de71387ae230c263c36fa628fdfa6.jpg", + "image_caption": [ + "Figure 4. Latent Code Interpolation. Latent code interpolation of our model trained on DeepFashion Dataset." + ], + "image_footnote": [], + "bbox": [ + 81, + 244, + 467, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "maps such that every point in the shell volume can be bijectively mapped to the 3D texture map for efficient modeling and rendering in texture space. The shell volume is constructed by offsetting the base mesh along the normal direction while maintaining the topology and avoiding self-intersection. The volume is discretized into tetrahedra that connect the vertices of the base and offset meshes. A unique mapping between the shell space and the texture space can be established by identifying the tetrahedron and subsequently querying the barycentric coordinates inside it. Formally, the mapping from $\\mathbf{x}$ in the shell volume to its position in the 3D texture map is defined as", + "bbox": [ + 75, + 464, + 468, + 645 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {t} = \\phi (\\mathbf {T} _ {t}, B (\\mathbf {x}, \\mathbf {T})) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 655, + 468, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{T},\\mathbf{T}_t$ refer to corresponding tetrahedra in the shell and texture space, $B(\\mathbf{x},\\mathbf{T})$ is the barycentric coordinates of $\\mathbf{x}$ in $\\mathbf{T}$ , and $\\phi$ denotes barycentric interpolation.", + "bbox": [ + 75, + 681, + 468, + 727 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Gaussian Shell Maps (GSM)", + "text_level": 1, + "bbox": [ + 76, + 739, + 339, + 757 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "GSM is a framework that connects 3D Gaussians with SOTA CNN-based generator networks [11] in a GAN setting. The key idea is to anchor the 3D Gaussians at fixed locations on a set of \"shells\" derived from the SMPL [37] human body template mesh. These shells span a volume to model surface details that deviate from the unclothed template mesh. We learn the relevant Gaussian parameters in the texture space, allowing us to leverage established CNN-based generative backbones while seamlessly utilizing the", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "parametric deformation model for articulation. The overall pipeline is shown in Figure 2.", + "bbox": [ + 496, + 248, + 890, + 279 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Representation", + "text_level": 1, + "bbox": [ + 498, + 287, + 653, + 304 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our method utilizes the concept of shell maps to leverage the inherent planar structure of texture space for seamless integration with CNN-based generative architectures and, at the same time, encapsulate diverse and complex surface details without directly modifying the template mesh. Specifically, the shell volume is defined by the boundary mesh layers, created by inflating and deflating the T-pose SMPL mesh using Laplacian Smoothing [60] with the smoothing factor set to negative for inflation and positive for deflation. We then represent this shell volume using $N$ mesh layers, i.e. \"shells\", by linearly spacing the vertices between the aforementioned boundary shells. In parallel, we apply a similar discretization strategy to the 3D texture space, creating $N$ 2D texture maps storing neural features that can be referenced for each shell using the UV mapping inherited from the SMPL template.", + "bbox": [ + 496, + 311, + 890, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Figure 2, we use the shell maps to generate all Gaussian parameters except the positions, as those are sampled and anchored w.r.t. to the shells at every iteration (explained below in detail). This results in a feature volume of $\\mathsf{T}^{N\\times H\\times W\\times 11}$ , comprised of 3D color $\\mathsf{T}^{\\mathrm{c}}$ , 1D opacity $\\mathsf{T}^o$ , 3D scaling $\\mathsf{T}^s$ , and 4D rotation $\\mathsf{T}^q$ features.", + "bbox": [ + 496, + 553, + 890, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We create Gaussians in our shell volume. This is done by sampling a fixed number of Gaussians quasi-uniformly on the shells based on the triangle areas. Once sampled, the Gaussians are anchored on the shells using barycentric coordinates so that every Gaussian center $\\mathbf{x}$ can be mapped to a point $\\mathbf{x}_t$ on the corresponding shell map using Eq. (4), except that the tetrahedra are replaced with the triangle in which the Gaussian resides. This anchoring is a crucial design choice, as it enables straightforward feature retrieval from the shell maps. More importantly, it simplifies the learning by fixing the Gaussian positions and offloading geometry modeling to opacity $o$ and sigma $\\Sigma$ .", + "bbox": [ + 496, + 643, + 892, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The spatial span of the Gaussians plays a significant role in reconstructing the features defined on the discrete shells into a continuous signal within the 3D space. It enables every point—whether on the shells, between them, or outside them—to receive a valid opacity and color value by evaluat-", + "bbox": [ + 496, + 825, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9444", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9d4ccb1dc66ec68769e236587cba67738311be5f642f9de02a0f41f4e1467f3e.jpg", + "image_caption": [ + "GNNARF" + ], + "image_footnote": [], + "bbox": [ + 125, + 99, + 454, + 167 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/60a80f93c25fa38b46384e0cab92c9bc0bd3aebc0c784de7de2c61a1989ecd07.jpg", + "image_caption": [ + "EVA3D" + ], + "image_footnote": [], + "bbox": [ + 133, + 176, + 256, + 247 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2abfb83f66150462ae1694569f8f057f0d47f061b55f19face5e827336def53b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 176, + 455, + 246 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ac2e726bb7e7dde5d0327859e176d331caf26c3066813aca6df3d2e65bf09cd1.jpg", + "image_caption": [ + "s", + "DeepFashion" + ], + "image_footnote": [], + "bbox": [ + 135, + 252, + 256, + 325 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0362bac430c50fc9a2689dd7de7fed6c1ad9e13dcf4ccb96cdcc77eb234b55fc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 250, + 455, + 324 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9f63fa811cb63af8ff2daaa0e6bd8a2caf56aeb0298bb39fa81d3bd8bb2bacd8.jpg", + "image_caption": [ + "Figure 5. Qualitative Comparison. We compare our results with GNARF and EVA3D baselines on DeepFashion and SHHQ datasets. In each case, we show the deformed body poses of the identities generated by the methods. The competing methods exhibit artifacts marked in red. Notice that our approach generates high-quality textures, like facial details and more realistic deformations." + ], + "image_footnote": [], + "bbox": [ + 526, + 97, + 851, + 344 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ing Eq. (3). This process allows us to model diverse appearances and body shapes, excelling mesh-based representation while at the same time maintaining efficient rendering, which is critical for GAN training.", + "bbox": [ + 75, + 420, + 468, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Formally, Gaussian opacity, color, scale, and rotation can be interpolated from the shell maps $\\mathsf{T}$", + "bbox": [ + 75, + 481, + 468, + 512 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf = \\mathrm {T} ^ {f} \\left(\\mathbf {x} _ {t}\\right), \\text {w h e r e} f = \\{o, \\mathbf {c}, \\mathbf {s}, \\mathbf {q} \\}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 523, + 468, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Deformation", + "text_level": 1, + "bbox": [ + 76, + 553, + 210, + 568 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The deformation step updates the Gaussians' locations and orientations based on the SMPL template mesh $\\mathbf{M}$ , pose $\\theta$ , and shape $\\beta$ . Note that, different from SMPL, our template mesh could be any of the shells. Since each Gaussian is anchored on the shells using barycentric coordinates, we can query its new location and orientation simply from the associated vertices. In particular, given the barycentric coordinates $B(\\mathbf{x},\\mathbf{T})$ of a Gaussian inside triangle $\\mathbf{T}$ , and the deformed position $\\mathbf{T}_{\\mathrm{new}}$ and the rotation quaternions $\\mathbf{P}$ on the vertices, we can obtain the new location and orientation of the Gaussian as", + "bbox": [ + 75, + 577, + 468, + 742 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {\\text {n e w}} = \\phi (\\mathbf {T} _ {\\text {n e w}}, B (\\mathbf {x}, \\mathbf {T})) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 756, + 468, + 772 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {q} _ {\\text {n e w}} = \\frac {\\hat {\\mathbf {p}}}{\\| \\hat {\\mathbf {p}} \\|} \\mathbf {q}, \\text {w h e r e} \\hat {\\mathbf {p}} = \\phi (\\mathbf {P}, B (\\mathbf {x}, \\mathbf {T})). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 776, + 468, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The deformed mesh is given by the SMPL deformation model SMPL, and the quaternions $\\mathbf{P}$ are a result of the linear blend skinning (LBS) from regressed joints and skinning weights, $\\{w_{j}\\}_{1}^{N_{J}}$ , i.e., $\\mathbf{M}_{\\mathrm{new}} = \\mathrm{SMPL}(\\pmb {\\theta},\\pmb {\\beta},\\mathbf{M})$ and $\\mathbf{p} = \\mathrm{rot2quat}\\left(\\sum_{j = 1}^{N_J}w_jR_j(\\pmb {\\theta},J(\\pmb {\\beta}))\\right)$ , where $R_{j}$ is the", + "bbox": [ + 75, + 816, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "rotation matrix of the $j$ -th joint, and $J$ is the regressor function that maps the shape parameters to the joint locations.", + "bbox": [ + 498, + 420, + 890, + 452 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. GAN training", + "text_level": 1, + "bbox": [ + 500, + 460, + 643, + 478 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Generator. Similar to prior work [11], we adopt StyleGAN2 without camera and pose conditioning [6, 71] for the generator. We use separate MLPs with different activations for $\\mathbf{c}, o, \\mathbf{q}$ , and $\\mathbf{s}$ : for $\\mathbf{c}$ , we use shifted sigmoid following the practice proposed in Mip-NeRF [5]; for $o$ we use sigmoid to constrain the range to $(0, 1)$ ; for $\\mathbf{q}$ we normalize the raw MLP output to ensure they are quaternions (see Eq. (7)); for $\\mathbf{s}$ we use clamped exponential activation to limit the size of the Gaussians, which is critical for convergence based on our empirical study.", + "bbox": [ + 496, + 484, + 890, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Discriminator. Our discriminator closely follows that of [6]. Since it does not have an upsampler, the input to the discriminator is the RGB image concatenated with the alpha channel (foreground mask), which is rendered using Gaussian rasterization with the Gaussian color set to 1 for fake samples and precomputed using off-the-shelf segmentation network [55] for the real samples. We refer to the discriminator using foreground mask \"Mask Discriminator\". Including the alpha channel helps prevent the white background from bleeding into the appearance, which causes artifacts during articulation.", + "bbox": [ + 496, + 640, + 892, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Face, Hand, and Feet Discriminator. As the human body and clothes are diverse, the discriminator may choose to focus on these features and provide a weak signal to the facial, hand, and foot areas, which are crucial for visually appealing results. To avoid this problem, we adopt a dedicated Face Discriminator $\\mathcal{D}_{\\mathrm{face}}$ , Feet-and-Hands Discrimi", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9445", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/7dfe76cbdc86d1d3c6bf1fe9686dbd5d16c9be12c208c4214ccfbbffb8df06d7.jpg", + "table_caption": [ + "Table 1. Quantitative Evaluation. We compare our method with 3D-GAN baselines using DeepFashion and SHHQ datasets. We compute the FID score to evaluate the quality and diversity of the generated samples. Notice that our scores are comparable to state-of-the-art methods. To evaluate deformation consistency, we compute the PCK metric, where our approach consistently outperforms the baselines. INF represents the inference speed measured in ms/img on an A6000 GPU at $512^2$ resolution. Note our method is the fastest across all competing methods. * numbers are adopted from EVA3D [23]; NA represents Not Available; — represents Not Applicable, and + numbers are provided by the authors." + ], + "table_footnote": [], + "table_body": "
ModelDeep FashionSHHQComp.
FID ↓PCK ↑FID ↓PCK ↑INF. ↓
EG3D*26.38—32.96—38
StyleSDF*92.40—14.12—32
ENARF*77.0343.7480.5440.17104
GNARF33.8597.8314.8498.9672
EVA3D*15.9187.5011.9988.95200
StylePeople17.7298.3114.6798.5828
GetAvatar19.00+NANANA44
AG3D10.93NANANA105
Ours15.7899.4813.3099.2728
", + "bbox": [ + 76, + 244, + 478, + 434 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "nator $\\mathcal{D}_{\\mathrm{feet/hands}}$ in addition to the main Discriminator $\\mathcal{D}_{\\mathrm{main}}$ . All of these part-focused discriminators share the same base architecture as the main discriminator, except that we no longer use any conditioning since these cropped images do not contain distinct pose information. The inputs are crops of the corresponding parts, whose spatial span is determined from the SMPL pose and camera parameters.", + "bbox": [ + 75, + 450, + 468, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Scaling Regularization. In our empirical study, we observe when unconstrained, the network tends to learn overly large or extremely small Gaussians early on during the training and rapidly leads to divergence or model collapse. We experimented with multiple regularization strategies and different activation functions and found the following scaling regularization most effective:", + "bbox": [ + 75, + 561, + 468, + 667 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s c a l e}} = \\mathrm {M} \\circ \\| \\mathrm {T} ^ {\\mathrm {s}} - s _ {\\text {r e f}} \\| ^ {2} \\text {w i t h} s _ {\\text {r e f}} = \\frac {1}{P} \\sum_ {i = 1} ^ {P} \\ln \\left(\\delta_ {i}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 88, + 680, + 468, + 720 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $M$ is the binary mask indicating UV-mapped region on the shell maps, and $s_{\\mathrm{ref}}$ is a reference scale determined by the closest-neighbor distance $\\delta$ averaged among all the Gaussians.", + "bbox": [ + 75, + 731, + 468, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiment Settings", + "text_level": 1, + "bbox": [ + 76, + 806, + 272, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Datasets", + "text_level": 1, + "bbox": [ + 76, + 830, + 181, + 845 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our method using the two most common human datasets, DeepFashion [36] and SHHQ [17]. DeepFashion and SHHQ do not provide SMPL parameters, so we use", + "bbox": [ + 75, + 854, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8817a128926a103fe0b05dbf07aaf837b325a5b9d84b5e1b2dd4c62dd0900132.jpg", + "table_caption": [ + "Table 2. Ablation: Number of Shells. Ablation on the number of shells performed on $128^{2}$ resolution trained for $1600k$ images." + ], + "table_footnote": [], + "table_body": "
# of Shells1281015
FID ↓40.325.3118.7021.5721.10
", + "bbox": [ + 509, + 118, + 883, + 166 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SMPLify-X [43] to obtain the SMPL parameters and camera poses.", + "bbox": [ + 498, + 199, + 890, + 229 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Training Details", + "text_level": 1, + "bbox": [ + 500, + 251, + 661, + 268 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "By default, we generate the shell maps at $512 \\times 512$ resolution for 8 shells with a total of $100k$ Gaussians equally distributed across the shells. The offset between adjacent shells is 0.08. The last fully-connected layer of the scaling prediction is adjusted to ensure the initial scaling is within a reasonable range, which is determined similar to $s_{\\mathrm{ref}}$ in Eq. (8). The loss weights for all discriminators are set to 1, whereas the scaling regularization is set to 0.1. $R1$ regularization is used for all the discriminators with a weight of 10. We apply progressive training starting at $256 \\times 256$ rendering resolution for $6000k$ training images and progressively grow the resolution to $512 \\times 512$ for $1000k$ images, then continue training at the fixed $512 \\times 512$ resolution for $3000k$ images. We train our method on 8 A100 GPUs with a batch size of 32. The learning rates are initialized at 0.002 for both the generator and discriminator.", + "bbox": [ + 496, + 279, + 893, + 522 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Evaluation Details", + "text_level": 1, + "bbox": [ + 500, + 542, + 678, + 558 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare with the following volume rendering and rasterization-based methods: EG3D [11], StyleSDF [42], GNARF [6], ENARF [41], StylePeople [19], EVA3D [23], GetAvatar [76], and AG3D [15]. Among these methods, EG3D and StyleSDF do not support articulation; except StylePeople, all methods use neural field and volumetric rendering. Regardless of representation, all these baselines use 2D convolution for postrendering upsampling and/or postprocessing, causing flickering artifacts (see Supplementary Materials).", + "bbox": [ + 496, + 571, + 890, + 723 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics Following prior work, We use two metrics to evaluate the quality, diversity, and quality of animations: Fréchet Inception Distance (FID) and Percentage of Correct Keypoint (PCK). The former measures the visual quality and diversity of the generated samples (using $50k$ samples and the full dataset). The latter measures how close the generated image aligns with the pose control. We compute PCK on $5k$ samples using the implementation provided by GNARF [6].", + "bbox": [ + 496, + 763, + 890, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9446", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5061a58a11f903bee1677c837befcd5079d65d02ff72910bbba64c2fee91c068.jpg", + "image_caption": [ + "Figure 6. Multi-view Consistency. The figure shows a closeup of the garment in two different views. The green dashed line follows the body shape and indicates the same 3D position. In AG3D [15], the pattern significantly changes, while ours, utilizing texture maps, has built-in view consistency." + ], + "image_footnote": [], + "bbox": [ + 83, + 93, + 472, + 325 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6. Experiment Results", + "text_level": 1, + "bbox": [ + 76, + 422, + 266, + 438 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.1. Qualitative", + "text_level": 1, + "bbox": [ + 76, + 446, + 199, + 463 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 1 showcases some of our generated results. Notice that our model is able to generate diverse attributes, including loose clothing and accessories like hats. Further, in Figure 3, we visualize the generated results in varying camera views and body poses. To demonstrate that the latent space of the proposed model is smooth, we show visual results of latent code interpolation in Figure 4 computed on the DeepFashion trained model.", + "bbox": [ + 75, + 470, + 468, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 5 shows the visual comparison of our method with the baselines on DeepFashion and SHHQ datasets, with more examples to be found in Supplementary Materials. Volumetric methods, such as GNARF, operate in a canonical pose and rely on accurate correspondence matching between the observed space and the canonical space, which is ill-defined; thus, these methods tend to produce artifacts for limbs, which undergo large deformation and occlusion. Our method is able to model the complex geometry and appearance of the human body, generating characters with loose clothing and accessories (see Figure 1) and producing convincing results under various articulations.", + "bbox": [ + 75, + 592, + 468, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2.Quantitative", + "text_level": 1, + "bbox": [ + 76, + 785, + 210, + 800 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Table 1, we show the results of our method compared to the competing methods. Many of the methods have not released training or data processing code; we have collected these numbers with our best effort from the authors directly.", + "bbox": [ + 75, + 809, + 468, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our method performs on par with the SOTA methods in terms of visual quality and diversity measured by FID", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1b9fef6c9aecdd64f01c8197f09cc9ae1df82de6a0858c0233bf5db73146f6d5.jpg", + "image_caption": [ + "Figure 7. Per-shell Visualization. We visualize the per-shell contributions. The shells decompose the human body into different parts, where the inner shells capture the torso, and the outer shells capture the clothing and hair." + ], + "image_footnote": [], + "bbox": [ + 511, + 90, + 875, + 339 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "while ranking top in terms of inference speed. By utilizing the texture space, our model can control the articulation in a straightforward manner, contributing to consistent alignment with the pose control input, as reflected by the PCK score. Note that GetAvatar is trained on multiview data, and AG3D uses normal maps for supervision.", + "bbox": [ + 496, + 425, + 890, + 516 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "All other baseline methods adopt post-rendering convolution layers to upsample and add texture details, leading to severe aliasing artifacts not reflected in the quantitative evaluation. We demonstrate this in Figure 6, where we render a generated identity in two different views using our method and AG3D, which achieves the lowest FID on the DeepFashion dataset. Notice that with view change, the features produced by AG3D change significantly, leading to unnatural flickering in animation even though the FID is low.", + "bbox": [ + 496, + 516, + 892, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.3. Ablation", + "text_level": 1, + "bbox": [ + 500, + 662, + 604, + 676 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct ablation studies on different configurations and components of our method to understand their effect on the final model's performance. To identify the trends, we test these configurations at $128^{2}$ resolution on the SHHQ dataset, trained for $1600k$ images without progressive training. $200k$ Gaussians are sampled for all the experiments.", + "bbox": [ + 496, + 685, + 890, + 776 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Number of Shells. The number of shells determines the granularity of variation we can model inside the shell volume, thus impacting the expressivity of our representation. We investigate this effect in Table 2. With only one shell, although the Gaussians can vary in size and cover regions off the shell surface, they still lack the capacity to model diverse human body shapes and appearances. Using two shells es", + "bbox": [ + 496, + 794, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9447", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a2d90ac876d564264aef258571024d9c30e2e5ca71ec75dfd587305158ddf2a7.jpg", + "image_caption": [ + "w/o Mask and Part Discriminators" + ], + "image_footnote": [], + "bbox": [ + 83, + 94, + 184, + 271 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5f52023679c00c6c1b75fdaabbb3e5d45548ec28397e33ed70575d213272f151.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 192, + 95, + 264, + 148 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8996ccb612221ede0468c0437a66707607969c6c94a29d79ca92ab2a135ecccf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 200, + 152, + 254, + 184 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8315f15269a8e6b8ba9f361a68925ff7b4921c7f808c9e732a83baa547dbeee0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 202, + 186, + 254, + 233 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/11a8e5b5a6db3dba21b7956cb9ab3b401cbfd52381df0056c5d1d25c74ec94a5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 187, + 234, + 271, + 268 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/53baec545b2efe830faa9332178e31fff5ba1e3a7b3fd43315ab2dbffc91c83e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 284, + 95, + 354, + 147 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9432951984b7d118844e22980b54eb950da36c6eeba6f8439f61f833db16351f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 152, + 344, + 183 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b1ba2b9fb7f5563cd7191c0e1af4e7fe8a5fda9939ed3073f7365caf57d1017a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 294, + 186, + 344, + 233 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c60e8edd65da92bb61cd635ccf89f5d793b5eb36ddb6ceb64f642744034c0f08.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 279, + 234, + 361, + 268 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4cb8897cca021a8a05c2bde68570c87fcaf91ffb6f259d296917695fa1ed378a.jpg", + "image_caption": [ + "w/ Mask and Part Discriminators", + "Figure 8. Mask + Part Discriminators. The blue box highlights the improved silhouette after concatenating the alpha channels in the input to the discriminator. The other boxes showcase better details using the face, feet, and hands discriminator." + ], + "image_footnote": [], + "bbox": [ + 364, + 95, + 465, + 271 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "mentally defines the boundary for the shell volume. This variation significantly outperforms the single-shell variation, demonstrating the importance of shell volume. Adding 8 shells (our default setting) further improves the visual quality, as the finer discretization enhances the capability to model more complex geometry and appearance. The performance gain stagnates with even more shells, likely because the model struggles to converge when the feature map becomes too large. Figure 7 shows an example identity of what each Gaussian shell has learned: progressing from the innermost to the outermost shell, every shell captures some texture details, and the outer shell models details that deviate from the base template, such as hat and hair.", + "bbox": [ + 75, + 371, + 468, + 568 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Mask and Part Discriminator The mask discriminator involves the inclusion of the alpha channel as part of the input to the RGB images and is very beneficial for disambiguating background by learning plausible human silhouettes. As demonstrated in Figure 8, the model trained without mask discriminator generates disturbing white artifacts outside the human shape. Figure 8 also shows the improvement in small features in the facial area and better structure for hands and feet introduced by facial and feet/hands discriminator. We also quantify the joint improvement of the mask and part discriminator and observed an FID boost from 20.63 to 18.7 at $128^{2}$ resolution.", + "bbox": [ + 75, + 590, + 467, + 771 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Types of Gaussian Anchors. In our representation, we opt to sample the Gaussians on shell meshes. There are other alternatives to sample the Gaussians, including constructing tetrahedra akin to the original Shell Map proposed by Porumbescu et al. [47]; or completely discarding the template by sampling uniformly inside a bounding box. We compared with these alternatives and found the shell repre", + "bbox": [ + 75, + 794, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "sensation achieves the best visual quality and convergence speed. A more detailed discussion can be found in Supplementary Materials.", + "bbox": [ + 498, + 90, + 890, + 136 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Discussion", + "text_level": 1, + "bbox": [ + 500, + 150, + 612, + 165 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Our method is not without limitations. Like almost all existing 3D human generation approaches, our method also relies on a parametric deformation model to enable articulation, which falls short in handling the dynamics of hair and loose clothing. Due to the irregularity and sparsity of Gaussians, it is not obvious how to extract the accurate geometry and normals. Surface splatting, with a stronger surface prior may be a potential solution, which we plan to investigate in the future. Lastly, while ranking among the best methods, our current results still cannot achieve photorealism. The challenge stems from the complexity of human appearance. A promising direction is to combine a small amount of multi-view studio captures with in-the-wild datasets to leverage priors discovered in controlled and richly annotated datasets.", + "bbox": [ + 496, + 171, + 890, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ethical Concerns. GANs, like the one we developed, carry the risk of being utilized for creating altered images of real individuals. This inappropriate application of image generation methods represents a danger to society, and we firmly oppose the use of our research for disseminating false information or damaging reputations. Additionally, we acknowledge that there might be a deficit in diversity in our outcomes, which could be a consequence of inherent biases in the datasets we utilize.", + "bbox": [ + 496, + 412, + 890, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Conclusions. In this work, we present Gaussian Shell Maps (GSMs), a novel framework that effectively combines CNN-based generators with 3D Gaussian rendering primitives for the generation of digital humans. Our experiments demonstrate the potential of GSMs in generating diverse and detailed human figures, including complex features like clothing and accessories. The framework shows promise in enhancing the efficiency of rendering processes, a crucial factor for real-time applications in industries like virtual reality and cinematic production. The potential of GSMs in practical scenarios invites further exploration, and continued research could enhance digital human modeling.", + "bbox": [ + 496, + 564, + 890, + 746 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. We thank the authors of 3D Gaussian Splitting, Bernhard Kerbl, Georgios Kopanas, and George Drettakis for the discussions on the code. We thank Thabo Beeler and Guandao Yang for fruitful discussions and Alexander Bergman for helping with the baseline comparisons. We also thank the authors of GetAvartar for providing additional evaluation results. This work was in part supported by Google, Samsung, Stanford HAI, and the Swiss Postdoc Mobility Fund.", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9448", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In ECCV, 2020. 3", + "[2] Sizhe An, Hongyi Xu, Yichun Shi, Guoxian Song, Umit Y Ogras, and Linjie Luo. Panohead: Geometry-aware 3d full-head synthesis in 360deg. In CVPR, 2023. 2", + "[3] Shivangi Aneja, Justus Thies, Angela Dai, and Matthias Nießner. ClipFace: Text-guided Editing of Textured 3D Morphable Models. In ArXiv preprint arXiv:2212.01406, 2022. 2", + "[4] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. ArXiv, abs/2211.01324, 2022.2", + "[5] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, 2021. 5", + "[6] Alexander Bergman, Petr Kellnhofer, Wang Yifan, Eric Chan, David Lindell, and Gordon Wetzstein. Generative neural articulated radiance fields. NeurIPS, 2022. 1, 2, 5, 6", + "[7] Alexander W. Bergman, Wang Yifan, and Gordon Wetzstein. Articulated 3d head avatar generation using text-to-image diffusion models. 2023. 2", + "[8] Yukang Cao, Yan-Pei Cao, Kai Han, Ying Shan, and KwanYee K. Wong. Dreamavatar: Text-and-shape guided 3d human avatar generation via diffusion models, 2023. 1", + "[9] Eric Chan, Koki Nagano, Matthew A. Chan, Alexander W. Bergman, Jeong Joon Park, Axel Levy, Miika Aittala, Shalini De Mello, Tero Karras, and Gordon Wetzstein. Generative novel view synthesis with 3d-aware diffusion models. ArXiv, abs/2304.02602, 2023. 2", + "[10] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In CVPR, 2021. 2", + "[11] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In CVPR, 2022. 1, 2, 4, 5, 6", + "[12] Rui Chen, Y. Chen, Ningxin Jiao, and Kui Jia. *Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation.* ArXiv, abs/2303.13873, 2023. 2", + "[13] Xinya Chen, Jiaxin Huang, Yanrui Bin, Lu Yu, and Yiyi Liao. Veri3d: Generative vertex-based radiance fields for 3d controllable human image synthesis, 2023. 2", + "[14] Stavros Diolatzis, Jan Novak, Fabrice Rousselle, Jonathan Granskog, Miika Aittala, Ravi Ramamoorthi, and George Drettakis. Mesogan: Generative neural reflectance shells. Comput. Graph. Forum, 2023. 2, 3" + ], + "bbox": [ + 78, + 114, + 467, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Zijian Dong, Xu Chen, Jinlong Yang, Michael J.Black, Otmar Hilliges, and Andreas Geiger. AG3D: Learning to generate 3D avatars from 2D image collections. In ICCV, 2023. 1, 2, 6, 7", + "[16] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Chen Qian, Chen Change Loy, Wayne Wu, and Ziwei Liu. Stylegan-human: A data-centric odyssey of human generation. In ECCV, 2022. 2", + "[17] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Chen Qian, Chen Change Loy, Wayne Wu, and Ziwei Liu. Stylegan-human: A data-centric odyssey of human generation. In ECCV, 2022. 2, 6", + "[18] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. NeurIPS, 2022. 2", + "[19] Artur Grigorev, Karim Iskakov, Anastasia Iinina, Renat Bashirov, Ilya Zakharkin, Alexander Vakhitov, and Victor Lempitsky. Stylepeople: A generative model of fullbody human avatars. In CVPR, 2021. 1, 2, 6", + "[20] Markus Gross and Hanspeter Pfister. Point-based graphics. Elsevier, 2011. 2", + "[21] Jeffrey P Grossman and William J Dally. Point sample rendering. In Proceedings of the Eurographics Workshop, 1998. 2", + "[22] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d aware generator for high-resolution image synthesis. In ICLR, 2022. 2", + "[23] Fangzhou Hong, Zhaoxi Chen, Yushi LAN, Liang Pan, and Ziwei Liu. EVA3d: Compositional 3d human generation from 2d image collections. In ICLR, 2023. 1, 2, 6", + "[24] Suyi Jiang, Haoran Jiang, Ziyu Wang, Haimin Luo, Wenzheng Chen, and Lan Xu. Humangen: Generating human radiance fields with explicit priors. In CVPR, 2023. 2", + "[25] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 2", + "[26] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In Proc. CVPR, 2020. 2", + "[27] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In NeurIPS, 2021. 2", + "[28] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM TOG, 2023. 2, 3", + "[29] Leonid Keselman and Martial Hebert. Approximate differentiable rendering with algebraic surfaces. In European Conference on Computer Vision (ECCV), 2022. 3", + "[30] Nikos Kolotouros, Thiemo Alldieck, Andrei Zanfir, Eduard Gabriel Bazavan, Mihai Fieraru, and Cristian Sminchisescu. Dreamhuman: Animatable 3d avatars from text. ArXiv, abs/2306.09329, 2023. 2", + "[31] Georgios Kopanas, Julien Philip, Thomas Leimkuhler, and George Drettakis. Point-based neural rendering with perview optimization. In Comput. Graph. Forum, 2021. 3" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9449", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Samuli Laine and Tero Karras. High-performance software rasterization on gpus. In Proceedings of the ACM SIGGRAPH Symposium on High Performance Graphics, 2011. 2", + "[33] Ruilong Li, Shan Yang, David A. Ross, and Angjoo Kanazawa. Learn to dance with aist++: Music conditioned 3d dance generation, 2021. 2", + "[34] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. ArXiv, abs/2211.10440, 2022. 2", + "[35] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. ArXiv, abs/2303.11328, 2023. 2", + "[36] Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In CVPR, 2016. 6", + "[37] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. ACM TOG, 2015. 1, 2, 4", + "[38] Youssef A Mejjati, Isa Milefchik, Aaron Gokaslan, Oliver Wang, Kwang In Kim, and James Tompkin. Gaussian: Controllable image synthesis with 3d gaussians from unposed silhouettes. arXiv preprint arXiv:2106.13215, 2021. 3", + "[39] Michael Niemeyer and Andreas Geiger. Campari: Camera-aware decomposed generative neural radiance fields. In 3DV, 2021. 2", + "[40] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In CVPR, 2021. 2", + "[41] Atsuhiro Noguchi, Xiao Sun, Stephen Lin, and Tatsuya Harada. Unsupervised learning of efficient geometry-aware neural articulated representations. In ECCV, 2022. 1, 2, 6", + "[42] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. StyleSDF: High-Resolution 3D-Consistent Image and Geometry Generation. arXiv preprint arXiv:2112.11427, 2021. 2, 6", + "[43] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In CVPR, 2019. 6", + "[44] Ryan Po and Gordon Wetzstein. Compositional 3d scene generation using locally conditioned diffusion. arXiv preprint arXiv:2303.12218, 2023. 2", + "[45] Ryan Po, Wang Yifan, Vladislav Golyanik, Kfir Aberman, Jonathan T. Barron, Amit H. Bermano, Eric Ryan Chan, Tali Dekel, Aleksander Holynski, Angjoo Kanazawa, C. Karen Liu, Lingjie Liu, Ben Mildenhall, Matthias Nießner, Björn Ommer, Christian Theobalt, Peter Wonka, and Gordon Wetzstein. State of the art on diffusion models for visual computing, 2023. 2", + "[46] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2" + ], + "bbox": [ + 78, + 90, + 468, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Serban D. Porumbescu, Brian Budge, Louis Feng, and Kenneth I. Joy. Shell maps. In ACM SIGGRAPH, 2005. 2, 3, 8", + "[48] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2", + "[49] Liu Ren, Hanspeter Pfister, and Matthias Zwicker. Object space ewa surface splatting: A hardware accelerated approach to high quality point rendering. In Comput. Graph. Forum, 2002. 2", + "[50] Helge Rhodin, Nadia Robertini, Christian Richardt, Hans-Peter Seidel, and Christian Theobalt. A versatile scene model with differentiable visibility applied to generative pose estimation. In ICCV, 2015. 3", + "[51] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. CVPR, 2021. 2", + "[52] Darius Rückert, Linus Franke, and Marc Stamminger. Adop: Approximate differentiable one-pixel point rendering. ACM TOG, 2022. 3", + "[53] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Seyedeh Sara Mahdavi, Raphael Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. ArXiv, abs/2205.11487, 2022. 2", + "[54] Miguel Sainz and Renato Pajarola. Point-based rendering techniques. Computers & Graphics, 28(6):869-879, 2004. 2", + "[55] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In CVPR, 2018. 5", + "[56] Markus Schütz, Bernhard Kerbl, and Michael Wimmer. Software rasterization of 2 billion points in real time. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 2022. 2", + "[57] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. In NeurIPS, 2020. 2", + "[58] Zackary P. T. Sin, Peter H. F. Ng, and Hong Va Leong. Ner-fahedron: A primitive for animatable neural rendering with interactive speed. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 2023. 3", + "[59] Ivan Skorokhodov, Aliaksandr Siarohin, Yinghao Xu, Jian Ren, Hsin-Ying Lee, Peter Wonka, and Sergey Tulyakov. 3d generation on imagenet. In ICLR, 2023. 2", + "[60] Olga Sorkine, Daniel Cohen-Or, Yaron Lipman, Marc Alexa, Christian Rössl, and H-P Seidel. Laplacian surface editing. In Eurographics, 2004. 4", + "[61] Carsten Stoll, Nils Hasler, Juergen Gall, Hans-Peter Seidel, and Christian Theobalt. Fast articulated motion tracking using a sums of gaussians body model. In ICCV, 2011. 3", + "[62] Jingxiang Sun, Xuan Wang, Yichun Shi, Lizhen Wang, Jue Wang, and Yebin Liu. Ide-3d: Interactive disentangled editing for high-resolution 3d-aware portrait synthesis. ACM TOG, 2022. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9450", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[63] Jingxiang Sun, Xuan Wang, Lizhen Wang, Xiaoyu Li, Yong Zhang, Hongwen Zhang, and Yebin Liu. Next3d: Generative neural texture rasterization for 3d-aware head avatars. In CVPR, 2023. 2", + "[64] Ayush Tewari, Ohad Fried, Justus Thies, Vincent Sitzmann, Stephen Lombardi, Kalyan Sunkavalli, Ricardo MartinBrualla, Tomas Simon, Jason Saragih, Matthias Nießner, et al. State of the art on neural rendering. In Comput. Graph. Forum, 2020. 2", + "[65] Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srinivasan, Edgar Tretschk, Wang Yifan, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lombardi, et al. Advances in neural rendering. In Comput. Graph. Forum, 2022. 1, 2", + "[66] Christina Tsalicoglou, Fabian Manhardt, Alessio Tonioni, Michael Niemeyer, and Federico Tombari. Textmesh: Generation of realistic 3d meshes from text prompts. ArXiv, abs/2304.12439, 2023. 2", + "[67] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. Synsin: End-to-end view synthesis from a single image. In CVPR, 2020. 3", + "[68] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In CVPR, 2022. 3", + "[69] Yinghao Xu, Sida Peng, Ceyuan Yang, Yujun Shen, and Bolei Zhou. 3d-aware image synthesis via learning structural and textural representations. arXiv preprint arXiv:2112.10759, 2021. 2", + "[70] Yinghao Xu, Menglei Chai, Zifan Shi, Sida Peng, Ivan Skorokhodov, Aliaksandr Siarohin, Ceyuan Yang, Yujun Shen, Hsin-Ying Lee, Bolei Zhou, et al. Discoscope: Spatially disentangled generative radiance fields for controllable 3d-aware scene synthesis. In CVPR, 2023. 2", + "[71] Yinghao Xu, Wang Yifan, Alexander W. Bergman, Mengei Chai, Bolei Zhou, and Gordon Wetzstein. Efficient 3d articulated human generation with layered surface volumes. In 3DV, 2024. 1, 2, 5", + "[72] Zhuoqian Yang, Shikai Li, Wayne Wu, and Bo Dai. 3dhuman: 3d-aware human image generation with 3d pose mapping. In ICCV, 2023. 2", + "[73] Wang Yifan, Felice Serena, Shihao Wu, Cengiz Oztireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM TOG, 2019. 3", + "[74] Jianfeng Zhang, Zihang Jiang, Dingdong Yang, Hongyi Xu, Yichun Shi, Guoxian Song, Zhongcong Xu, Xinchao Wang, and Jiashi Feng. Avatargen: A 3d generative model for animatable human avatars. ArXiv, 2023. 2", + "[75] Qiang Zhang, Seung-Hwan Baek, Szymon Rusinkiewicz, and Felix Heide. Differentiable point-based radiance fields for efficient view synthesis. In SIGGRAPH Asia, 2022. 3", + "[76] Xuanmeng Zhang, Jianfeng Zhang, Chacko Rohan, Hongyi Xu, Guoxian Song, Yi Yang, and Jiashi Feng. Getavatar: Generative textured meshes for animatable human avatars. In ICCV, 2023. 1, 2, 6", + "[77] Yufeng Zheng, Wang Yifan, Gordon Wetzstein, Michael J Black, and Otmar Hilliges. Pointavatar: Deformable point-based head avatars from videos. In CVPR, 2023. 3" + ], + "bbox": [ + 78, + 90, + 468, + 901 + ], + "page_idx": 10 + }, + { + "type": "ref_text", + "text": "[78] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Ewa volume splatting. In Proceedings Visualization, 2001. 3", + "bbox": [ + 501, + 90, + 892, + 133 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "9451", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Gaussian Shell Maps for Efficient 3D Human Generation/c6318955-2d79-4389-9712-ec90ab24e73b_model.json b/2024/Gaussian Shell Maps for Efficient 3D Human Generation/c6318955-2d79-4389-9712-ec90ab24e73b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b3806570874e5e88bff0232a6ed2c8d256a5a454 --- /dev/null +++ b/2024/Gaussian Shell Maps for Efficient 3D Human Generation/c6318955-2d79-4389-9712-ec90ab24e73b_model.json @@ -0,0 +1,2565 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.197, + 0.078, + 0.775, + 0.099 + ], + "angle": 0, + "content": "Gaussian Shell Maps for Efficient 3D Human Generation" + }, + { + "type": "text", + "bbox": [ + 0.095, + 0.126, + 0.877, + 0.22 + ], + "angle": 0, + "content": "Rameen Abdal\\*1 Wang Yifan\\*1 Zifan Shi\\*1,2 Yinghao Xu\\(^{1}\\) Ryan Po\\(^{1}\\) Zhengfei Kuang\\(^{1}\\) Qifeng Chen\\(^{2}\\) Dit-Yan Yeung\\(^{2}\\) Gordon Wetzstein\\(^{1}\\) \n\\(^{1}\\)Stanford University, USA \\({}^{2}\\)HKUST, Hong Kong" + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.254, + 0.877, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.413, + 0.895, + 0.483 + ], + "angle": 0, + "content": "Figure 1. Gaussian Shell Maps. Gaussian Shell Maps is an efficient framework for 3D human generation connecting 3D Gaussians with CNN-based generators. 3D Gaussians are anchored to \"shells\" derived from the SMPL template [37] (only two shells are visualized for clarity), and the appearance is modeled in texture space. Trained only on 2D images, we show that our method can generate diverse articulable humans in real-time with state-of-the-art quality directly in high resolution without the need for upsampling and hence avoiding aliasing artifacts." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.496, + 0.314, + 0.511 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Efficient generation of 3D digital humans is important in several industries, including virtual reality, social media, and cinematic production. 3D generative adversarial networks (GANs) have demonstrated state-of-the-art (SOTA) quality and diversity for generated assets. Current 3D GAN architectures, however, typically rely on volume representations, which are slow to render, thereby hampering the GAN training and requiring multi-view-inconsistent 2D upsamplers. Here, we introduce Gaussian Shell Maps (GSMs) as a framework that connects SOTA generator network architectures with emerging 3D Gaussian rendering primitives using an articulable multi shell-based scaffold. In this setting, a CNN generates a 3D texture stack with features that are mapped to the shells. The latter represent inflated and deflated versions of a template surface of a digital human in a canonical body pose. Instead of rasterizing the shells directly, we sample 3D Gaussians on the shells whose attributes are encoded in the texture features. These Gaussians are efficiently and differentiably rendered. The ability to articulate the shells is important during GAN training and, at inference time, to deform a body into arbitrary user-defined poses. Our efficient rendering scheme bypasses the need for view-inconsistent upsamplers and achieves high-quality multi-view consistent renderings at a native resolu" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.497, + 0.892, + 0.542 + ], + "angle": 0, + "content": "tion of \\(512 \\times 512\\) pixels. We demonstrate that GSMs successfully generate 3D humans when trained on single-view datasets, including SHHQ and DeepFashion." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.543, + 0.882, + 0.558 + ], + "angle": 0, + "content": "Project Page: rameenabdal.github.io/GaussianShellMaps" + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.585, + 0.633, + 0.601 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.61, + 0.892, + 0.701 + ], + "angle": 0, + "content": "The ability to generate articulable three-dimensional digital humans augments traditional asset creation and animation workflows, which are laborious and costly. Such generative artificial intelligence-fueled workflows are crucial in several applications, including communication, cinematic production, and interactive gaming, among others." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.701, + 0.895, + 0.868 + ], + "angle": 0, + "content": "3D Generative Adversarial Networks (GANs) have emerged as the state-of-the-art (SOTA) platform in this domain, enabling the generation of diverse 3D assets at interactive framereates [6, 8, 15, 23, 41, 71, 76, 76]. Most existing 3D GANs build on variants of volumetric scene representations combined with neural volume rendering [65]. However, volume rendering is relatively slow and compromises the training of a GAN, which requires tens of millions of forward rendering passes to converge [11]. Mesh-based representations building on fast differentiable rasterization have been proposed to overcome this limitation [19, 71]," + }, + { + "type": "page_footnote", + "bbox": [ + 0.524, + 0.875, + 0.64, + 0.888 + ], + "angle": 0, + "content": "* Equal Contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.524, + 0.888, + 0.878, + 0.901 + ], + "angle": 0, + "content": "† Work done as a visiting student researcher at Stanford University" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.875, + 0.878, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "9441" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "but these are not expressive enough to realistically model features like hair, clothing, or accessories, which deviate significantly from the template mesh. These limitations, which are largely imposed by a tradeoff between efficient or expressive scene representations, have been constraining the quality and resolution of existing 3D GANs. While partially compensated for by using 2D convolutional neural network (CNN)-based upsamplers [11, 22], upsampling leads to multi-view inconsistency in the form of aliasing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.235, + 0.473, + 0.372 + ], + "angle": 0, + "content": "Very recently, 3D Gaussians have been introduced as a promising neural scene representation offering fast rendering speed and high expressivity [28]. While 3D Gaussians have been explored in the context of single-scene overfitting, their full potential in generative settings has yet to be unlocked. This is challenging because it is not obvious how to combine SOTA CNN-based generators [26, 27] with 3D Gaussian primitives that inherently do not exist on a regular Cartesian grid and that may vary in numbers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.378, + 0.473, + 0.635 + ], + "angle": 0, + "content": "We introduce Gaussian Shell Maps (GSMs), a 3D GAN framework that intuitively connects CNN generators with 3D Gaussians used as efficient rendering primitives. Inspired by the traditional computer graphics work on shell maps [47], GSMs use the CNN generator to produce texture maps for a set of \"shell\" meshes appropriately inflated and deflated from the popular SMPL mesh template for human bodies [37]. The textures on the individual shells directly encode the properties of 3D Gaussians, which are sampled on the shell surfaces at fixed locations. The generated images are rendered using highly efficient Gaussian splatting, and articulation of these Gaussians can be naturally enabled through deforming the scaffolding shells with the SMPL model. Since 3D Gaussians have spatial extent, they represent details on, in between, and outside the discrete shells. GSMs are trained exclusively on datasets containing single-view images of human bodies, such as SHHQ [17]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.643, + 0.472, + 0.749 + ], + "angle": 0, + "content": "Our experiments demonstrate that GSM can generate highly diverse appearances, including loose clothing and accessories, at high resolution, without an upsampler, at a state-of-the-art rendering speed of 125 FPS (or 35FPS including generation). Among various architecture design choices, multiple shells with fixed relative locations of the 3D Gaussian achieve the best results in our experiments." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.755, + 0.352, + 0.771 + ], + "angle": 0, + "content": "Specifically, our contributions include" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.778, + 0.468, + 0.822 + ], + "angle": 0, + "content": "1. We propose a novel 3D GAN framework combining a CNN-based generator and 3D Gaussian rendering primitives using shell maps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.822, + 0.468, + 0.898 + ], + "angle": 0, + "content": "2. We demonstrate the fastest 3D GAN architecture to date, achieving real-time rendering of \\(512^{2}\\) px without convolutional upsamplers, with image quality and diversity matching the state of the art on challenging single-view datasets of human bodies." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.778, + 0.468, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.09, + 0.642, + 0.107 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.116, + 0.895, + 0.48 + ], + "angle": 0, + "content": "3D-Aware Generative Models. GANs and diffusion models are two very powerful generative models emerging as a result of efficient architectures and high-quality datasets [25, 48]. With the quality of image generation reaching photo-realism, the community started leveraging these SOTA generative models for 3D generation by combining neural rendering methods [64, 65] to produce high-quality multi-view consistent 3D objects from image collections. Several 3D GAN architectures have explored implicit or explicit neural volume representations for modeling 3D objects, including [2, 10, 11, 14, 22, 39, 40, 42, 57, 59, 62, 69, 70]. 3D diffusion models, on the other hand, typically use the priors encoded by pre-trained text-to-image 2D diffusion models, e.g. [4, 7-9, 12, 30, 34, 35, 44, 46, 51, 53, 66] (see this survey for more details [45]). Due to the lack of high-quality, large-scale multi-view datasets curated for specific categories like humans, the choice of a suitable generative model becomes critical. Using diffusion models to generate high-resolution multi-view consistent 3D objects is still an unsolved problem [45]. 3D GANs, on the other hand, exhibit better quality and multi-view consistency at higher resolutions that do not assume multi-view data [11, 15, 62, 71]. This motivates our choice of building an efficient representation using a 3D GAN framework." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.486, + 0.895, + 0.803 + ], + "angle": 0, + "content": "Generative Articulated 3D Digital Humans. 3D-GAN frameworks have been proposed to generate the appearance, geometry, and identity-preserving novel views of digital humans [6, 13, 15, 23, 71, 72, 74, 76]. Most of these GANs are trained on single-view image collections [16, 33]. A popular approach is to use a neural radiance representation [6, 23, 24, 41, 74] where a canonically posed human can be articulated via deformation. Other approaches are based on meshes, where a template can be fixed or learned during the training [3, 18, 19, 63, 72]. Related to this line of work, a concurrent paper, LSV-GAN [71], offsets the SMPL template meshes into layered surfaces and composites the per-layer rasterization result to form the final rendering. While it provides a faster alternative to the volume rendering-based approaches, it can only accommodate a small offset from the SMPL mesh, which hampers diversity. Our GSM method differs from LSV-GAN as we employ 3D Gaussians [28] as primitives on the layered surfaces, which, by having a learnable spatial span, allows for larger deviation from the template mesh and can thus generate more intricate details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Point-Based Rendering. Earlier point-based methods efficiently render point clouds and rasterize them by fixing the size [20, 21, 54]. While efficient in terms of speed and parallel rasterization on graphics processing units [32, 56], they are not differentiable [28, 49]. To combine these methods with the neural networks and perform view synthesis," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "9442" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.09, + 0.885, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.302, + 0.895, + 0.388 + ], + "angle": 0, + "content": "Figure 2. Method Overview. We propose an expressive yet highly efficient representation, Gaussian Shell Map (GSM), for 3D human generation. Combining the idea of 3D Gaussians and Shell Maps (Sec. 3), we sample 3D Gaussians on \"shells\", which are mesh layers offsetted from the SMPL template, forming a shell volume to model complex and diverse geometry and appearance; the Gaussian parameters are learned in the texture space, allowing us to leverage existing CNN-based generative architecture (Sec. 4.1). Articulation is straightforward by interpolating the deformation of the shell (Sec. 4.2). The generation is supervised by single-view 2D images using several discriminator critics, including part-specific face, hands, and feet discriminators (Sec. 4.3)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.413, + 0.473, + 0.611 + ], + "angle": 0, + "content": "recent works have developed differentiable point-based rendering techniques [1, 52, 67, 68, 73, 75, 77]. More recently, 3D Gaussian-based point splatting gained traction due to the flexibility of anisotropic covariance and density control with efficient depth sorting [28, 29]. This allows 3D Gaussian splats to handle complex scenes composed of high and low-frequency features. Relevant to human bodies, 3D Gaussians have also been used in pose estimation and tracking [38, 50, 61]. While the Gaussian primitives have been used for efficient scene reconstruction and novel view synthesis, it is not trivial to deploy Gaussians in a generative setup. To the best of our knowledge, our method is the first to propose a combination of 3D Gaussians and 3D GANs." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.628, + 0.206, + 0.645 + ], + "angle": 0, + "content": "3. Background" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.654, + 0.47, + 0.851 + ], + "angle": 0, + "content": "3D Gaussians. These point-based primitives can be differentiably and efficiently rendered using EWA (elliptical weighted average) volume splatting [78]. 3D Gaussians have recently demonstrated outstanding expressivity for 3D scene reconstruction [28], in which the Gaussian parameters, position \\(\\mathbf{x} \\in \\mathbb{R}^3\\), opacity \\(o \\in \\mathbb{R}\\), color \\(\\mathbf{c} \\in \\mathbb{R}^{sh}\\) (\\(sh\\) representing the spherical harmonic coefficients), scaling \\(\\mathbf{s} \\in \\mathbb{R}^3\\), and rotation \\(\\mathbf{q} \\in \\mathbb{R}^4\\) parameterized as quaternions are jointly optimized to minimize the photometric errors of the rendered images in a set of known camera views. The optimization is accompanied by adaptive control of the density, where the points are added or removed based on the density, size, and gradient of the Gaussians." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.852, + 0.366, + 0.868 + ], + "angle": 0, + "content": "Specifically, each Gaussian is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.137, + 0.881, + 0.47, + 0.903 + ], + "angle": 0, + "content": "\\[\nG \\left(\\mathbf {x} ^ {\\prime}; \\mathbf {x}, \\Sigma\\right) = \\exp^ {- \\frac {1}{2} \\left(\\mathbf {x} ^ {\\prime} - \\mathbf {x}\\right) ^ {\\intercal} \\Sigma^ {- 1} \\left(\\mathbf {x} ^ {\\prime} - \\mathbf {x}\\right)}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.412, + 0.892, + 0.458 + ], + "angle": 0, + "content": "where \\(\\Sigma = R S S^T R^T\\) is the covariance matrix parameterized by the rotation and scaling matrices \\(R\\) and \\(S\\) given by the quaternions \\(\\mathbf{q}\\) and scaling \\(\\mathbf{s}\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.458, + 0.892, + 0.503 + ], + "angle": 0, + "content": "The image formation is governed by classic point-based \\(\\alpha\\)-blending of overlapping Gaussians ordered from closest to farthest [31]:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.509, + 0.892, + 0.551 + ], + "angle": 0, + "content": "\\[\n\\mathbf {C} = \\sum_ {i \\in \\mathcal {N}} \\mathbf {c} _ {i} ^ {\\prime} o _ {i} ^ {\\prime} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {\\mathbf {j}}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.556, + 0.804, + 0.573 + ], + "angle": 0, + "content": "where color \\(\\mathbf{c}_i^{\\prime}\\) and opacity \\(\\sigma_{i}^{\\prime}\\) is computed by" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.577, + 0.892, + 0.595 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} _ {i} ^ {\\prime} = G \\left(\\mathbf {x} ^ {\\prime}; \\mathbf {x} _ {i}\\right) \\mathbf {c} _ {i} \\quad \\text {a n d} \\quad o _ {i} ^ {\\prime} = G \\left(\\mathbf {x} ^ {\\prime}; \\mathbf {x} _ {i}\\right) o _ {i}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.6, + 0.893, + 0.81 + ], + "angle": 0, + "content": "Thanks to their ability to fit complex geometry and appearance, 3D Gaussians are gaining popularity for 3D scene reconstruction. However, deploying 3D Gaussians for generative tasks remains an unexplored topic and is challenging for several reasons. First, Gaussians are of \"Lagrangian\" nature - their sparse number and learnable positions are challenging to combine with SOTA \"Eulerian\" (i.e., grid-based) CNN generators. Second, the parameters of Gaussians are highly correlated. The same radiance field can be equally well explained by many different configurations of Gaussians, varying their locations, sizes, scales, rotations, colors, and opacities. This ambiguity makes it challenging to generalize over a distribution of objects or scenes, which generative methods do." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Shell Maps. Our representation is related to Shell Maps [14, 47, 58], a technique in computer graphics designed to model near-surface details. Shell maps use 3D texture maps to store the fine-scale features in a shell-like volume close to a given base surface, typically represented as a triangle mesh. In essence, they extend UV" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9443" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.087, + 0.885, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.208, + 0.846, + 0.223 + ], + "angle": 0, + "content": "Figure 3. Novel Views and Animation. We can render a generated identity in novel views and articulate it for animations." + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.246, + 0.468, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.412, + 0.471, + 0.441 + ], + "angle": 0, + "content": "Figure 4. Latent Code Interpolation. Latent code interpolation of our model trained on DeepFashion Dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.465, + 0.47, + 0.646 + ], + "angle": 0, + "content": "maps such that every point in the shell volume can be bijectively mapped to the 3D texture map for efficient modeling and rendering in texture space. The shell volume is constructed by offsetting the base mesh along the normal direction while maintaining the topology and avoiding self-intersection. The volume is discretized into tetrahedra that connect the vertices of the base and offset meshes. A unique mapping between the shell space and the texture space can be established by identifying the tetrahedron and subsequently querying the barycentric coordinates inside it. Formally, the mapping from \\(\\mathbf{x}\\) in the shell volume to its position in the 3D texture map is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.656, + 0.47, + 0.673 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {t} = \\phi (\\mathbf {T} _ {t}, B (\\mathbf {x}, \\mathbf {T})) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.682, + 0.47, + 0.728 + ], + "angle": 0, + "content": "where \\(\\mathbf{T},\\mathbf{T}_t\\) refer to corresponding tetrahedra in the shell and texture space, \\(B(\\mathbf{x},\\mathbf{T})\\) is the barycentric coordinates of \\(\\mathbf{x}\\) in \\(\\mathbf{T}\\), and \\(\\phi\\) denotes barycentric interpolation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.74, + 0.341, + 0.758 + ], + "angle": 0, + "content": "4. Gaussian Shell Maps (GSM)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "GSM is a framework that connects 3D Gaussians with SOTA CNN-based generator networks [11] in a GAN setting. The key idea is to anchor the 3D Gaussians at fixed locations on a set of \"shells\" derived from the SMPL [37] human body template mesh. These shells span a volume to model surface details that deviate from the unclothed template mesh. We learn the relevant Gaussian parameters in the texture space, allowing us to leverage established CNN-based generative backbones while seamlessly utilizing the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.249, + 0.892, + 0.28 + ], + "angle": 0, + "content": "parametric deformation model for articulation. The overall pipeline is shown in Figure 2." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.289, + 0.654, + 0.305 + ], + "angle": 0, + "content": "4.1. Representation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.312, + 0.892, + 0.553 + ], + "angle": 0, + "content": "Our method utilizes the concept of shell maps to leverage the inherent planar structure of texture space for seamless integration with CNN-based generative architectures and, at the same time, encapsulate diverse and complex surface details without directly modifying the template mesh. Specifically, the shell volume is defined by the boundary mesh layers, created by inflating and deflating the T-pose SMPL mesh using Laplacian Smoothing [60] with the smoothing factor set to negative for inflation and positive for deflation. We then represent this shell volume using \\( N \\) mesh layers, i.e. \"shells\", by linearly spacing the vertices between the aforementioned boundary shells. In parallel, we apply a similar discretization strategy to the 3D texture space, creating \\( N \\) 2D texture maps storing neural features that can be referenced for each shell using the UV mapping inherited from the SMPL template." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.892, + 0.644 + ], + "angle": 0, + "content": "As shown in Figure 2, we use the shell maps to generate all Gaussian parameters except the positions, as those are sampled and anchored w.r.t. to the shells at every iteration (explained below in detail). This results in a feature volume of \\(\\mathsf{T}^{N\\times H\\times W\\times 11}\\), comprised of 3D color \\(\\mathsf{T}^{\\mathrm{c}}\\), 1D opacity \\(\\mathsf{T}^o\\), 3D scaling \\(\\mathsf{T}^s\\), and 4D rotation \\(\\mathsf{T}^q\\) features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.893, + 0.825 + ], + "angle": 0, + "content": "We create Gaussians in our shell volume. This is done by sampling a fixed number of Gaussians quasi-uniformly on the shells based on the triangle areas. Once sampled, the Gaussians are anchored on the shells using barycentric coordinates so that every Gaussian center \\(\\mathbf{x}\\) can be mapped to a point \\(\\mathbf{x}_t\\) on the corresponding shell map using Eq. (4), except that the tetrahedra are replaced with the triangle in which the Gaussian resides. This anchoring is a crucial design choice, as it enables straightforward feature retrieval from the shell maps. More importantly, it simplifies the learning by fixing the Gaussian positions and offloading geometry modeling to opacity \\(o\\) and sigma \\(\\Sigma\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.901 + ], + "angle": 0, + "content": "The spatial span of the Gaussians plays a significant role in reconstructing the features defined on the discrete shells into a continuous signal within the 3D space. It enables every point—whether on the shells, between them, or outside them—to receive a valid opacity and color value by evaluat-" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9444" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.086, + 0.119, + 0.102, + 0.155 + ], + "angle": 0, + "content": "GNNARF" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.193, + 0.105, + 0.226 + ], + "angle": 0, + "content": "EVA3D" + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.279, + 0.105, + 0.301 + ], + "angle": 0, + "content": "s" + }, + { + "type": "image", + "bbox": [ + 0.126, + 0.101, + 0.455, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.135, + 0.178, + 0.258, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.178, + 0.457, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.137, + 0.253, + 0.258, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.251, + 0.457, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.264, + 0.332, + 0.34, + 0.345 + ], + "angle": 0, + "content": "DeepFashion" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.098, + 0.852, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.352, + 0.894, + 0.396 + ], + "angle": 0, + "content": "Figure 5. Qualitative Comparison. We compare our results with GNARF and EVA3D baselines on DeepFashion and SHHQ datasets. In each case, we show the deformed body poses of the identities generated by the methods. The competing methods exhibit artifacts marked in red. Notice that our approach generates high-quality textures, like facial details and more realistic deformations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.421, + 0.469, + 0.481 + ], + "angle": 0, + "content": "ing Eq. (3). This process allows us to model diverse appearances and body shapes, excelling mesh-based representation while at the same time maintaining efficient rendering, which is critical for GAN training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.482, + 0.469, + 0.513 + ], + "angle": 0, + "content": "Formally, Gaussian opacity, color, scale, and rotation can be interpolated from the shell maps \\( \\mathsf{T} \\)" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.524, + 0.469, + 0.543 + ], + "angle": 0, + "content": "\\[\nf = \\mathrm {T} ^ {f} \\left(\\mathbf {x} _ {t}\\right), \\text {w h e r e} f = \\{o, \\mathbf {c}, \\mathbf {s}, \\mathbf {q} \\}. \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.554, + 0.212, + 0.569 + ], + "angle": 0, + "content": "4.2. Deformation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.578, + 0.47, + 0.743 + ], + "angle": 0, + "content": "The deformation step updates the Gaussians' locations and orientations based on the SMPL template mesh \\(\\mathbf{M}\\), pose \\(\\theta\\), and shape \\(\\beta\\). Note that, different from SMPL, our template mesh could be any of the shells. Since each Gaussian is anchored on the shells using barycentric coordinates, we can query its new location and orientation simply from the associated vertices. In particular, given the barycentric coordinates \\(B(\\mathbf{x},\\mathbf{T})\\) of a Gaussian inside triangle \\(\\mathbf{T}\\), and the deformed position \\(\\mathbf{T}_{\\mathrm{new}}\\) and the rotation quaternions \\(\\mathbf{P}\\) on the vertices, we can obtain the new location and orientation of the Gaussian as" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.757, + 0.469, + 0.773 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {\\text {n e w}} = \\phi (\\mathbf {T} _ {\\text {n e w}}, B (\\mathbf {x}, \\mathbf {T})) \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.128, + 0.777, + 0.469, + 0.809 + ], + "angle": 0, + "content": "\\[\n\\mathbf {q} _ {\\text {n e w}} = \\frac {\\hat {\\mathbf {p}}}{\\| \\hat {\\mathbf {p}} \\|} \\mathbf {q}, \\text {w h e r e} \\hat {\\mathbf {p}} = \\phi (\\mathbf {P}, B (\\mathbf {x}, \\mathbf {T})). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.818, + 0.47, + 0.904 + ], + "angle": 0, + "content": "The deformed mesh is given by the SMPL deformation model SMPL, and the quaternions \\(\\mathbf{P}\\) are a result of the linear blend skinning (LBS) from regressed joints and skinning weights, \\(\\{w_{j}\\}_{1}^{N_{J}}\\), i.e., \\(\\mathbf{M}_{\\mathrm{new}} = \\mathrm{SMPL}(\\pmb {\\theta},\\pmb {\\beta},\\mathbf{M})\\) and \\(\\mathbf{p} = \\mathrm{rot2quat}\\left(\\sum_{j = 1}^{N_J}w_jR_j(\\pmb {\\theta},J(\\pmb {\\beta}))\\right)\\), where \\(R_{j}\\) is the" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.421, + 0.892, + 0.453 + ], + "angle": 0, + "content": "rotation matrix of the \\(j\\)-th joint, and \\(J\\) is the regressor function that maps the shape parameters to the joint locations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.462, + 0.644, + 0.479 + ], + "angle": 0, + "content": "4.3. GAN training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.486, + 0.892, + 0.637 + ], + "angle": 0, + "content": "Generator. Similar to prior work [11], we adopt StyleGAN2 without camera and pose conditioning [6, 71] for the generator. We use separate MLPs with different activations for \\(\\mathbf{c}, o, \\mathbf{q}\\), and \\(\\mathbf{s}\\): for \\(\\mathbf{c}\\), we use shifted sigmoid following the practice proposed in Mip-NeRF [5]; for \\(o\\) we use sigmoid to constrain the range to \\((0, 1)\\); for \\(\\mathbf{q}\\) we normalize the raw MLP output to ensure they are quaternions (see Eq. (7)); for \\(\\mathbf{s}\\) we use clamped exponential activation to limit the size of the Gaussians, which is critical for convergence based on our empirical study." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.641, + 0.893, + 0.807 + ], + "angle": 0, + "content": "Discriminator. Our discriminator closely follows that of [6]. Since it does not have an upsampler, the input to the discriminator is the RGB image concatenated with the alpha channel (foreground mask), which is rendered using Gaussian rasterization with the Gaussian color set to 1 for fake samples and precomputed using off-the-shelf segmentation network [55] for the real samples. We refer to the discriminator using foreground mask \"Mask Discriminator\". Including the alpha channel helps prevent the white background from bleeding into the appearance, which causes artifacts during articulation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Face, Hand, and Feet Discriminator. As the human body and clothes are diverse, the discriminator may choose to focus on these features and provide a weak signal to the facial, hand, and foot areas, which are crucial for visually appealing results. To avoid this problem, we adopt a dedicated Face Discriminator \\(\\mathcal{D}_{\\mathrm{face}}\\), Feet-and-Hands Discrimi" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9445" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.472, + 0.243 + ], + "angle": 0, + "content": "Table 1. Quantitative Evaluation. We compare our method with 3D-GAN baselines using DeepFashion and SHHQ datasets. We compute the FID score to evaluate the quality and diversity of the generated samples. Notice that our scores are comparable to state-of-the-art methods. To evaluate deformation consistency, we compute the PCK metric, where our approach consistently outperforms the baselines. INF represents the inference speed measured in ms/img on an A6000 GPU at \\(512^2\\) resolution. Note our method is the fastest across all competing methods. * numbers are adopted from EVA3D [23]; NA represents Not Available; — represents Not Applicable, and + numbers are provided by the authors." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.246, + 0.48, + 0.435 + ], + "angle": 0, + "content": "
ModelDeep FashionSHHQComp.
FID ↓PCK ↑FID ↓PCK ↑INF. ↓
EG3D*26.38—32.96—38
StyleSDF*92.40—14.12—32
ENARF*77.0343.7480.5440.17104
GNARF33.8597.8314.8498.9672
EVA3D*15.9187.5011.9988.95200
StylePeople17.7298.3114.6798.5828
GetAvatar19.00+NANANA44
AG3D10.93NANANA105
Ours15.7899.4813.3099.2728
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.452, + 0.47, + 0.558 + ], + "angle": 0, + "content": "nator \\(\\mathcal{D}_{\\mathrm{feet/hands}}\\) in addition to the main Discriminator \\(\\mathcal{D}_{\\mathrm{main}}\\). All of these part-focused discriminators share the same base architecture as the main discriminator, except that we no longer use any conditioning since these cropped images do not contain distinct pose information. The inputs are crops of the corresponding parts, whose spatial span is determined from the SMPL pose and camera parameters." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.563, + 0.47, + 0.669 + ], + "angle": 0, + "content": "Scaling Regularization. In our empirical study, we observe when unconstrained, the network tends to learn overly large or extremely small Gaussians early on during the training and rapidly leads to divergence or model collapse. We experimented with multiple regularization strategies and different activation functions and found the following scaling regularization most effective:" + }, + { + "type": "equation", + "bbox": [ + 0.089, + 0.681, + 0.47, + 0.722 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s c a l e}} = \\mathrm {M} \\circ \\| \\mathrm {T} ^ {\\mathrm {s}} - s _ {\\text {r e f}} \\| ^ {2} \\text {w i t h} s _ {\\text {r e f}} = \\frac {1}{P} \\sum_ {i = 1} ^ {P} \\ln \\left(\\delta_ {i}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.732, + 0.47, + 0.793 + ], + "angle": 0, + "content": "where \\(M\\) is the binary mask indicating UV-mapped region on the shell maps, and \\(s_{\\mathrm{ref}}\\) is a reference scale determined by the closest-neighbor distance \\(\\delta\\) averaged among all the Gaussians." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.807, + 0.273, + 0.825 + ], + "angle": 0, + "content": "5. Experiment Settings" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.182, + 0.847 + ], + "angle": 0, + "content": "5.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.903 + ], + "angle": 0, + "content": "We evaluate our method using the two most common human datasets, DeepFashion [36] and SHHQ [17]. DeepFashion and SHHQ do not provide SMPL parameters, so we use" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.894, + 0.118 + ], + "angle": 0, + "content": "Table 2. Ablation: Number of Shells. Ablation on the number of shells performed on \\(128^{2}\\) resolution trained for \\(1600k\\) images." + }, + { + "type": "table", + "bbox": [ + 0.511, + 0.119, + 0.885, + 0.167 + ], + "angle": 0, + "content": "
# of Shells1281015
FID ↓40.325.3118.7021.5721.10
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.2, + 0.892, + 0.231 + ], + "angle": 0, + "content": "SMPLify-X [43] to obtain the SMPL parameters and camera poses." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.252, + 0.663, + 0.269 + ], + "angle": 0, + "content": "5.2. Training Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.28, + 0.895, + 0.523 + ], + "angle": 0, + "content": "By default, we generate the shell maps at \\(512 \\times 512\\) resolution for 8 shells with a total of \\(100k\\) Gaussians equally distributed across the shells. The offset between adjacent shells is 0.08. The last fully-connected layer of the scaling prediction is adjusted to ensure the initial scaling is within a reasonable range, which is determined similar to \\(s_{\\mathrm{ref}}\\) in Eq. (8). The loss weights for all discriminators are set to 1, whereas the scaling regularization is set to 0.1. \\(R1\\) regularization is used for all the discriminators with a weight of 10. We apply progressive training starting at \\(256 \\times 256\\) rendering resolution for \\(6000k\\) training images and progressively grow the resolution to \\(512 \\times 512\\) for \\(1000k\\) images, then continue training at the fixed \\(512 \\times 512\\) resolution for \\(3000k\\) images. We train our method on 8 A100 GPUs with a batch size of 32. The learning rates are initialized at 0.002 for both the generator and discriminator." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.544, + 0.679, + 0.559 + ], + "angle": 0, + "content": "5.3. Evaluation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.572, + 0.892, + 0.724 + ], + "angle": 0, + "content": "Baselines. We compare with the following volume rendering and rasterization-based methods: EG3D [11], StyleSDF [42], GNARF [6], ENARF [41], StylePeople [19], EVA3D [23], GetAvatar [76], and AG3D [15]. Among these methods, EG3D and StyleSDF do not support articulation; except StylePeople, all methods use neural field and volumetric rendering. Regardless of representation, all these baselines use 2D convolution for postrendering upsampling and/or postprocessing, causing flickering artifacts (see Supplementary Materials)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Metrics Following prior work, We use two metrics to evaluate the quality, diversity, and quality of animations: Fréchet Inception Distance (FID) and Percentage of Correct Keypoint (PCK). The former measures the visual quality and diversity of the generated samples (using \\(50k\\) samples and the full dataset). The latter measures how close the generated image aligns with the pose control. We compute PCK on \\(5k\\) samples using the implementation provided by GNARF [6]." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9446" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.094, + 0.473, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.327, + 0.47, + 0.397 + ], + "angle": 0, + "content": "Figure 6. Multi-view Consistency. The figure shows a closeup of the garment in two different views. The green dashed line follows the body shape and indicates the same 3D position. In AG3D [15], the pattern significantly changes, while ours, utilizing texture maps, has built-in view consistency." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.423, + 0.267, + 0.439 + ], + "angle": 0, + "content": "6. Experiment Results" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.448, + 0.2, + 0.464 + ], + "angle": 0, + "content": "6.1. Qualitative" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.472, + 0.47, + 0.592 + ], + "angle": 0, + "content": "Figure 1 showcases some of our generated results. Notice that our model is able to generate diverse attributes, including loose clothing and accessories like hats. Further, in Figure 3, we visualize the generated results in varying camera views and body poses. To demonstrate that the latent space of the proposed model is smooth, we show visual results of latent code interpolation in Figure 4 computed on the DeepFashion trained model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.593, + 0.47, + 0.775 + ], + "angle": 0, + "content": "Figure 5 shows the visual comparison of our method with the baselines on DeepFashion and SHHQ datasets, with more examples to be found in Supplementary Materials. Volumetric methods, such as GNARF, operate in a canonical pose and rely on accurate correspondence matching between the observed space and the canonical space, which is ill-defined; thus, these methods tend to produce artifacts for limbs, which undergo large deformation and occlusion. Our method is able to model the complex geometry and appearance of the human body, generating characters with loose clothing and accessories (see Figure 1) and producing convincing results under various articulations." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.786, + 0.212, + 0.801 + ], + "angle": 0, + "content": "6.2.Quantitative" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.469, + 0.87 + ], + "angle": 0, + "content": "In Table 1, we show the results of our method compared to the competing methods. Many of the methods have not released training or data processing code; we have collected these numbers with our best effort from the authors directly." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Our method performs on par with the SOTA methods in terms of visual quality and diversity measured by FID" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.092, + 0.877, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.343, + 0.892, + 0.4 + ], + "angle": 0, + "content": "Figure 7. Per-shell Visualization. We visualize the per-shell contributions. The shells decompose the human body into different parts, where the inner shells capture the torso, and the outer shells capture the clothing and hair." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.426, + 0.892, + 0.517 + ], + "angle": 0, + "content": "while ranking top in terms of inference speed. By utilizing the texture space, our model can control the articulation in a straightforward manner, contributing to consistent alignment with the pose control input, as reflected by the PCK score. Note that GetAvatar is trained on multiview data, and AG3D uses normal maps for supervision." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.517, + 0.893, + 0.653 + ], + "angle": 0, + "content": "All other baseline methods adopt post-rendering convolution layers to upsample and add texture details, leading to severe aliasing artifacts not reflected in the quantitative evaluation. We demonstrate this in Figure 6, where we render a generated identity in two different views using our method and AG3D, which achieves the lowest FID on the DeepFashion dataset. Notice that with view change, the features produced by AG3D change significantly, leading to unnatural flickering in animation even though the FID is low." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.663, + 0.605, + 0.677 + ], + "angle": 0, + "content": "6.3. Ablation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.686, + 0.892, + 0.777 + ], + "angle": 0, + "content": "We conduct ablation studies on different configurations and components of our method to understand their effect on the final model's performance. To identify the trends, we test these configurations at \\(128^{2}\\) resolution on the SHHQ dataset, trained for \\(1600k\\) images without progressive training. \\(200k\\) Gaussians are sampled for all the experiments." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Number of Shells. The number of shells determines the granularity of variation we can model inside the shell volume, thus impacting the expressivity of our representation. We investigate this effect in Table 2. With only one shell, although the Gaussians can vary in size and cover regions off the shell surface, they still lack the capacity to model diverse human body shapes and appearances. Using two shells es" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9447" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.095, + 0.185, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.096, + 0.273, + 0.262, + 0.283 + ], + "angle": 0, + "content": "w/o Mask and Part Discriminators" + }, + { + "type": "image", + "bbox": [ + 0.194, + 0.096, + 0.266, + 0.149 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.202, + 0.153, + 0.255, + 0.185 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.204, + 0.188, + 0.255, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.188, + 0.236, + 0.272, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.285, + 0.096, + 0.356, + 0.148 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.153, + 0.345, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.295, + 0.188, + 0.346, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.236, + 0.362, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.288, + 0.273, + 0.448, + 0.283 + ], + "angle": 0, + "content": "w/ Mask and Part Discriminators" + }, + { + "type": "image", + "bbox": [ + 0.366, + 0.096, + 0.466, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.289, + 0.47, + 0.344 + ], + "angle": 0, + "content": "Figure 8. Mask + Part Discriminators. The blue box highlights the improved silhouette after concatenating the alpha channels in the input to the discriminator. The other boxes showcase better details using the face, feet, and hands discriminator." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.372, + 0.47, + 0.569 + ], + "angle": 0, + "content": "mentally defines the boundary for the shell volume. This variation significantly outperforms the single-shell variation, demonstrating the importance of shell volume. Adding 8 shells (our default setting) further improves the visual quality, as the finer discretization enhances the capability to model more complex geometry and appearance. The performance gain stagnates with even more shells, likely because the model struggles to converge when the feature map becomes too large. Figure 7 shows an example identity of what each Gaussian shell has learned: progressing from the innermost to the outermost shell, every shell captures some texture details, and the outer shell models details that deviate from the base template, such as hat and hair." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.592, + 0.468, + 0.772 + ], + "angle": 0, + "content": "Mask and Part Discriminator The mask discriminator involves the inclusion of the alpha channel as part of the input to the RGB images and is very beneficial for disambiguating background by learning plausible human silhouettes. As demonstrated in Figure 8, the model trained without mask discriminator generates disturbing white artifacts outside the human shape. Figure 8 also shows the improvement in small features in the facial area and better structure for hands and feet introduced by facial and feet/hands discriminator. We also quantify the joint improvement of the mask and part discriminator and observed an FID boost from 20.63 to 18.7 at \\(128^{2}\\) resolution." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Types of Gaussian Anchors. In our representation, we opt to sample the Gaussians on shell meshes. There are other alternatives to sample the Gaussians, including constructing tetrahedra akin to the original Shell Map proposed by Porumbescu et al. [47]; or completely discarding the template by sampling uniformly inside a bounding box. We compared with these alternatives and found the shell repre" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "sensation achieves the best visual quality and convergence speed. A more detailed discussion can be found in Supplementary Materials." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.151, + 0.613, + 0.166 + ], + "angle": 0, + "content": "7. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.172, + 0.892, + 0.399 + ], + "angle": 0, + "content": "Limitations. Our method is not without limitations. Like almost all existing 3D human generation approaches, our method also relies on a parametric deformation model to enable articulation, which falls short in handling the dynamics of hair and loose clothing. Due to the irregularity and sparsity of Gaussians, it is not obvious how to extract the accurate geometry and normals. Surface splatting, with a stronger surface prior may be a potential solution, which we plan to investigate in the future. Lastly, while ranking among the best methods, our current results still cannot achieve photorealism. The challenge stems from the complexity of human appearance. A promising direction is to combine a small amount of multi-view studio captures with in-the-wild datasets to leverage priors discovered in controlled and richly annotated datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.414, + 0.892, + 0.549 + ], + "angle": 0, + "content": "Ethical Concerns. GANs, like the one we developed, carry the risk of being utilized for creating altered images of real individuals. This inappropriate application of image generation methods represents a danger to society, and we firmly oppose the use of our research for disseminating false information or damaging reputations. Additionally, we acknowledge that there might be a deficit in diversity in our outcomes, which could be a consequence of inherent biases in the datasets we utilize." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.565, + 0.892, + 0.747 + ], + "angle": 0, + "content": "Conclusions. In this work, we present Gaussian Shell Maps (GSMs), a novel framework that effectively combines CNN-based generators with 3D Gaussian rendering primitives for the generation of digital humans. Our experiments demonstrate the potential of GSMs in generating diverse and detailed human figures, including complex features like clothing and accessories. The framework shows promise in enhancing the efficiency of rendering processes, a crucial factor for real-time applications in industries like virtual reality and cinematic production. The potential of GSMs in practical scenarios invites further exploration, and continued research could enhance digital human modeling." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. We thank the authors of 3D Gaussian Splitting, Bernhard Kerbl, Georgios Kopanas, and George Drettakis for the discussions on the code. We thank Thabo Beeler and Guandao Yang for fruitful discussions and Alexander Bergman for helping with the baseline comparisons. We also thank the authors of GetAvartar for providing additional evaluation results. This work was in part supported by Google, Samsung, Stanford HAI, and the Swiss Postdoc Mobility Fund." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9448" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.156 + ], + "angle": 0, + "content": "[1] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.468, + 0.2 + ], + "angle": 0, + "content": "[2] Sizhe An, Hongyi Xu, Yichun Shi, Guoxian Song, Umit Y Ogras, and Linjie Luo. Panohead: Geometry-aware 3d full-head synthesis in 360deg. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.202, + 0.468, + 0.256 + ], + "angle": 0, + "content": "[3] Shivangi Aneja, Justus Thies, Angela Dai, and Matthias Nießner. ClipFace: Text-guided Editing of Textured 3D Morphable Models. In ArXiv preprint arXiv:2212.01406, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.259, + 0.468, + 0.34 + ], + "angle": 0, + "content": "[4] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. ArXiv, abs/2211.01324, 2022.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.344, + 0.468, + 0.399 + ], + "angle": 0, + "content": "[5] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.402, + 0.468, + 0.442 + ], + "angle": 0, + "content": "[6] Alexander Bergman, Petr Kellnhofer, Wang Yifan, Eric Chan, David Lindell, and Gordon Wetzstein. Generative neural articulated radiance fields. NeurIPS, 2022. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.445, + 0.468, + 0.485 + ], + "angle": 0, + "content": "[7] Alexander W. Bergman, Wang Yifan, and Gordon Wetzstein. Articulated 3d head avatar generation using text-to-image diffusion models. 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.489, + 0.468, + 0.529 + ], + "angle": 0, + "content": "[8] Yukang Cao, Yan-Pei Cao, Kai Han, Ying Shan, and KwanYee K. Wong. Dreamavatar: Text-and-shape guided 3d human avatar generation via diffusion models, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.532, + 0.468, + 0.6 + ], + "angle": 0, + "content": "[9] Eric Chan, Koki Nagano, Matthew A. Chan, Alexander W. Bergman, Jeong Joon Park, Axel Levy, Miika Aittala, Shalini De Mello, Tero Karras, and Gordon Wetzstein. Generative novel view synthesis with 3d-aware diffusion models. ArXiv, abs/2304.02602, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.603, + 0.468, + 0.657 + ], + "angle": 0, + "content": "[10] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.468, + 0.741 + ], + "angle": 0, + "content": "[11] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In CVPR, 2022. 1, 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.745, + 0.468, + 0.799 + ], + "angle": 0, + "content": "[12] Rui Chen, Y. Chen, Ningxin Jiao, and Kui Jia. *Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation.* ArXiv, abs/2303.13873, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.468, + 0.843 + ], + "angle": 0, + "content": "[13] Xinya Chen, Jiaxin Huang, Yanrui Bin, Lu Yu, and Yiyi Liao. Veri3d: Generative vertex-based radiance fields for 3d controllable human image synthesis, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[14] Stavros Diolatzis, Jan Novak, Fabrice Rousselle, Jonathan Granskog, Miika Aittala, Ravi Ramamoorthi, and George Drettakis. Mesogan: Generative neural reflectance shells. Comput. Graph. Forum, 2023. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[15] Zijian Dong, Xu Chen, Jinlong Yang, Michael J.Black, Otmar Hilliges, and Andreas Geiger. AG3D: Learning to generate 3D avatars from 2D image collections. In ICCV, 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.151, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[16] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Chen Qian, Chen Change Loy, Wayne Wu, and Ziwei Liu. Stylegan-human: A data-centric odyssey of human generation. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.208, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[17] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Chen Qian, Chen Change Loy, Wayne Wu, and Ziwei Liu. Stylegan-human: A data-centric odyssey of human generation. In ECCV, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.265, + 0.892, + 0.32 + ], + "angle": 0, + "content": "[18] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.323, + 0.892, + 0.377 + ], + "angle": 0, + "content": "[19] Artur Grigorev, Karim Iskakov, Anastasia Iinina, Renat Bashirov, Ilya Zakharkin, Alexander Vakhitov, and Victor Lempitsky. Stylepeople: A generative model of fullbody human avatars. In CVPR, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.38, + 0.891, + 0.406 + ], + "angle": 0, + "content": "[20] Markus Gross and Hanspeter Pfister. Point-based graphics. Elsevier, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.41, + 0.891, + 0.449 + ], + "angle": 0, + "content": "[21] Jeffrey P Grossman and William J Dally. Point sample rendering. In Proceedings of the Eurographics Workshop, 1998. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.453, + 0.891, + 0.495 + ], + "angle": 0, + "content": "[22] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d aware generator for high-resolution image synthesis. In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.498, + 0.891, + 0.537 + ], + "angle": 0, + "content": "[23] Fangzhou Hong, Zhaoxi Chen, Yushi LAN, Liang Pan, and Ziwei Liu. EVA3d: Compositional 3d human generation from 2d image collections. In ICLR, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.541, + 0.891, + 0.581 + ], + "angle": 0, + "content": "[24] Suyi Jiang, Haoran Jiang, Ziyu Wang, Haimin Luo, Wenzheng Chen, and Lan Xu. Humangen: Generating human radiance fields with explicit priors. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.584, + 0.891, + 0.624 + ], + "angle": 0, + "content": "[25] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.628, + 0.891, + 0.668 + ], + "angle": 0, + "content": "[26] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In Proc. CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.672, + 0.891, + 0.712 + ], + "angle": 0, + "content": "[27] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.716, + 0.891, + 0.756 + ], + "angle": 0, + "content": "[28] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM TOG, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.759, + 0.891, + 0.8 + ], + "angle": 0, + "content": "[29] Leonid Keselman and Martial Hebert. Approximate differentiable rendering with algebraic surfaces. In European Conference on Computer Vision (ECCV), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.802, + 0.891, + 0.857 + ], + "angle": 0, + "content": "[30] Nikos Kolotouros, Thiemo Alldieck, Andrei Zanfir, Eduard Gabriel Bazavan, Mihai Fieraru, and Cristian Sminchisescu. Dreamhuman: Animatable 3d avatars from text. ArXiv, abs/2306.09329, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.891, + 0.901 + ], + "angle": 0, + "content": "[31] Georgios Kopanas, Julien Philip, Thomas Leimkuhler, and George Drettakis. Point-based neural rendering with perview optimization. In Comput. Graph. Forum, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9449" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[32] Samuli Laine and Tero Karras. High-performance software rasterization on gpus. In Proceedings of the ACM SIGGRAPH Symposium on High Performance Graphics, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[33] Ruilong Li, Shan Yang, David A. Ross, and Angjoo Kanazawa. Learn to dance with aist++: Music conditioned 3d dance generation, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.247 + ], + "angle": 0, + "content": "[34] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. ArXiv, abs/2211.10440, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.469, + 0.303 + ], + "angle": 0, + "content": "[35] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. ArXiv, abs/2303.11328, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.469, + 0.347 + ], + "angle": 0, + "content": "[36] Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In CVPR, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.469, + 0.39 + ], + "angle": 0, + "content": "[37] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. ACM TOG, 2015. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.392, + 0.469, + 0.459 + ], + "angle": 0, + "content": "[38] Youssef A Mejjati, Isa Milefchik, Aaron Gokaslan, Oliver Wang, Kwang In Kim, and James Tompkin. Gaussian: Controllable image synthesis with 3d gaussians from unposed silhouettes. arXiv preprint arXiv:2106.13215, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.462, + 0.469, + 0.503 + ], + "angle": 0, + "content": "[39] Michael Niemeyer and Andreas Geiger. Campari: Camera-aware decomposed generative neural radiance fields. In 3DV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.469, + 0.546 + ], + "angle": 0, + "content": "[40] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.469, + 0.589 + ], + "angle": 0, + "content": "[41] Atsuhiro Noguchi, Xiao Sun, Stephen Lin, and Tatsuya Harada. Unsupervised learning of efficient geometry-aware neural articulated representations. In ECCV, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.591, + 0.469, + 0.659 + ], + "angle": 0, + "content": "[42] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. StyleSDF: High-Resolution 3D-Consistent Image and Geometry Generation. arXiv preprint arXiv:2112.11427, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.469, + 0.717 + ], + "angle": 0, + "content": "[43] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In CVPR, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.469, + 0.76 + ], + "angle": 0, + "content": "[44] Ryan Po and Gordon Wetzstein. Compositional 3d scene generation using locally conditioned diffusion. arXiv preprint arXiv:2303.12218, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[45] Ryan Po, Wang Yifan, Vladislav Golyanik, Kfir Aberman, Jonathan T. Barron, Amit H. Bermano, Eric Ryan Chan, Tali Dekel, Aleksander Holynski, Angjoo Kanazawa, C. Karen Liu, Lingjie Liu, Ben Mildenhall, Matthias Nießner, Björn Ommer, Christian Theobalt, Peter Wonka, and Gordon Wetzstein. State of the art on diffusion models for visual computing, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[46] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "[47] Serban D. Porumbescu, Brian Budge, Louis Feng, and Kenneth I. Joy. Shell maps. In ACM SIGGRAPH, 2005. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[48] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[49] Liu Ren, Hanspeter Pfister, and Matthias Zwicker. Object space ewa surface splatting: A hardware accelerated approach to high quality point rendering. In Comput. Graph. Forum, 2002. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[50] Helge Rhodin, Nadia Robertini, Christian Richardt, Hans-Peter Seidel, and Christian Theobalt. A versatile scene model with differentiable visibility applied to generative pose estimation. In ICCV, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[51] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[52] Darius Rückert, Linus Franke, and Marc Stamminger. Adop: Approximate differentiable one-pixel point rendering. ACM TOG, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.488 + ], + "angle": 0, + "content": "[53] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Seyedeh Sara Mahdavi, Raphael Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. ArXiv, abs/2205.11487, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.49, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[54] Miguel Sainz and Renato Pajarola. Point-based rendering techniques. Computers & Graphics, 28(6):869-879, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.519, + 0.892, + 0.559 + ], + "angle": 0, + "content": "[55] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In CVPR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.561, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[56] Markus Schütz, Bernhard Kerbl, and Michael Wimmer. Software rasterization of 2 billion points in real time. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.892, + 0.658 + ], + "angle": 0, + "content": "[57] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. In NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[58] Zackary P. T. Sin, Peter H. F. Ng, and Hong Va Leong. Ner-fahedron: A primitive for animatable neural rendering with interactive speed. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.718, + 0.892, + 0.759 + ], + "angle": 0, + "content": "[59] Ivan Skorokhodov, Aliaksandr Siarohin, Yinghao Xu, Jian Ren, Hsin-Ying Lee, Peter Wonka, and Sergey Tulyakov. 3d generation on imagenet. In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.761, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[60] Olga Sorkine, Daniel Cohen-Or, Yaron Lipman, Marc Alexa, Christian Rössl, and H-P Seidel. Laplacian surface editing. In Eurographics, 2004. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.804, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[61] Carsten Stoll, Nils Hasler, Juergen Gall, Hans-Peter Seidel, and Christian Theobalt. Fast articulated motion tracking using a sums of gaussians body model. In ICCV, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[62] Jingxiang Sun, Xuan Wang, Yichun Shi, Lizhen Wang, Jue Wang, and Yebin Liu. Ide-3d: Interactive disentangled editing for high-resolution 3d-aware portrait synthesis. ACM TOG, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9450" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[63] Jingxiang Sun, Xuan Wang, Lizhen Wang, Xiaoyu Li, Yong Zhang, Hongwen Zhang, and Yebin Liu. Next3d: Generative neural texture rasterization for 3d-aware head avatars. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.219 + ], + "angle": 0, + "content": "[64] Ayush Tewari, Ohad Fried, Justus Thies, Vincent Sitzmann, Stephen Lombardi, Kalyan Sunkavalli, Ricardo MartinBrualla, Tomas Simon, Jason Saragih, Matthias Nießner, et al. State of the art on neural rendering. In Comput. Graph. Forum, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.47, + 0.289 + ], + "angle": 0, + "content": "[65] Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srinivasan, Edgar Tretschk, Wang Yifan, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lombardi, et al. Advances in neural rendering. In Comput. Graph. Forum, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.47, + 0.345 + ], + "angle": 0, + "content": "[66] Christina Tsalicoglou, Fabian Manhardt, Alessio Tonioni, Michael Niemeyer, and Federico Tombari. Textmesh: Generation of realistic 3d meshes from text prompts. ArXiv, abs/2304.12439, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.347, + 0.47, + 0.389 + ], + "angle": 0, + "content": "[67] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. Synsin: End-to-end view synthesis from a single image. In CVPR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.391, + 0.47, + 0.432 + ], + "angle": 0, + "content": "[68] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.47, + 0.489 + ], + "angle": 0, + "content": "[69] Yinghao Xu, Sida Peng, Ceyuan Yang, Yujun Shen, and Bolei Zhou. 3d-aware image synthesis via learning structural and textural representations. arXiv preprint arXiv:2112.10759, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.47, + 0.559 + ], + "angle": 0, + "content": "[70] Yinghao Xu, Menglei Chai, Zifan Shi, Sida Peng, Ivan Skorokhodov, Aliaksandr Siarohin, Ceyuan Yang, Yujun Shen, Hsin-Ying Lee, Bolei Zhou, et al. Discoscope: Spatially disentangled generative radiance fields for controllable 3d-aware scene synthesis. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.561, + 0.47, + 0.615 + ], + "angle": 0, + "content": "[71] Yinghao Xu, Wang Yifan, Alexander W. Bergman, Mengei Chai, Bolei Zhou, and Gordon Wetzstein. Efficient 3d articulated human generation with layered surface volumes. In 3DV, 2024. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.617, + 0.47, + 0.659 + ], + "angle": 0, + "content": "[72] Zhuoqian Yang, Shikai Li, Wayne Wu, and Bo Dai. 3dhuman: 3d-aware human image generation with 3d pose mapping. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.47, + 0.702 + ], + "angle": 0, + "content": "[73] Wang Yifan, Felice Serena, Shihao Wu, Cengiz Oztireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM TOG, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.47, + 0.758 + ], + "angle": 0, + "content": "[74] Jianfeng Zhang, Zihang Jiang, Dingdong Yang, Hongyi Xu, Yichun Shi, Guoxian Song, Zhongcong Xu, Xinchao Wang, and Jiashi Feng. Avatargen: A 3d generative model for animatable human avatars. ArXiv, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.47, + 0.802 + ], + "angle": 0, + "content": "[75] Qiang Zhang, Seung-Hwan Baek, Szymon Rusinkiewicz, and Felix Heide. Differentiable point-based radiance fields for efficient view synthesis. In SIGGRAPH Asia, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.47, + 0.857 + ], + "angle": 0, + "content": "[76] Xuanmeng Zhang, Jianfeng Zhang, Chacko Rohan, Hongyi Xu, Guoxian Song, Yi Yang, and Jiashi Feng. Getavatar: Generative textured meshes for animatable human avatars. In ICCV, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[77] Yufeng Zheng, Wang Yifan, Gordon Wetzstein, Michael J Black, and Otmar Hilliges. Pointavatar: Deformable point-based head avatars from videos. In CVPR, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.135 + ], + "angle": 0, + "content": "[78] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Ewa volume splatting. In Proceedings Visualization, 2001. 3" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "9451" + } + ] +] \ No newline at end of file diff --git a/2024/Gaussian Shell Maps for Efficient 3D Human Generation/c6318955-2d79-4389-9712-ec90ab24e73b_origin.pdf b/2024/Gaussian Shell Maps for Efficient 3D Human Generation/c6318955-2d79-4389-9712-ec90ab24e73b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d8c316a6131d0bbeb7a338ed00a5b9cd10b7dbfe --- /dev/null +++ b/2024/Gaussian Shell Maps for Efficient 3D Human Generation/c6318955-2d79-4389-9712-ec90ab24e73b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:40ab61dd2313da16cadc29fe90288c3e6f880c98ea9fc90d6a1f9f9e8e75ce86 +size 8332861 diff --git a/2024/Gaussian Shell Maps for Efficient 3D Human Generation/full.md b/2024/Gaussian Shell Maps for Efficient 3D Human Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..82c4394de550b3dec060db251ecef08cb9cd11f0 --- /dev/null +++ b/2024/Gaussian Shell Maps for Efficient 3D Human Generation/full.md @@ -0,0 +1,346 @@ +# Gaussian Shell Maps for Efficient 3D Human Generation + +Rameen Abdal\*1 Wang Yifan\*1 Zifan Shi\*1,2 Yinghao Xu $^{1}$ Ryan Po $^{1}$ Zhengfei Kuang $^{1}$ Qifeng Chen $^{2}$ Dit-Yan Yeung $^{2}$ Gordon Wetzstein $^{1}$ $^{1}$ Stanford University, USA ${}^{2}$ HKUST, Hong Kong + +![](images/33b9ecb905bebd4160c3f5af610ac3b1bca30e445a3e1437885d0330a2a73f02.jpg) +Figure 1. Gaussian Shell Maps. Gaussian Shell Maps is an efficient framework for 3D human generation connecting 3D Gaussians with CNN-based generators. 3D Gaussians are anchored to "shells" derived from the SMPL template [37] (only two shells are visualized for clarity), and the appearance is modeled in texture space. Trained only on 2D images, we show that our method can generate diverse articulable humans in real-time with state-of-the-art quality directly in high resolution without the need for upsampling and hence avoiding aliasing artifacts. + +# Abstract + +Efficient generation of 3D digital humans is important in several industries, including virtual reality, social media, and cinematic production. 3D generative adversarial networks (GANs) have demonstrated state-of-the-art (SOTA) quality and diversity for generated assets. Current 3D GAN architectures, however, typically rely on volume representations, which are slow to render, thereby hampering the GAN training and requiring multi-view-inconsistent 2D upsamplers. Here, we introduce Gaussian Shell Maps (GSMs) as a framework that connects SOTA generator network architectures with emerging 3D Gaussian rendering primitives using an articulable multi shell-based scaffold. In this setting, a CNN generates a 3D texture stack with features that are mapped to the shells. The latter represent inflated and deflated versions of a template surface of a digital human in a canonical body pose. Instead of rasterizing the shells directly, we sample 3D Gaussians on the shells whose attributes are encoded in the texture features. These Gaussians are efficiently and differentiably rendered. The ability to articulate the shells is important during GAN training and, at inference time, to deform a body into arbitrary user-defined poses. Our efficient rendering scheme bypasses the need for view-inconsistent upsamplers and achieves high-quality multi-view consistent renderings at a native resolu + +tion of $512 \times 512$ pixels. We demonstrate that GSMs successfully generate 3D humans when trained on single-view datasets, including SHHQ and DeepFashion. + +Project Page: rameenabdal.github.io/GaussianShellMaps + +# 1. Introduction + +The ability to generate articulable three-dimensional digital humans augments traditional asset creation and animation workflows, which are laborious and costly. Such generative artificial intelligence-fueled workflows are crucial in several applications, including communication, cinematic production, and interactive gaming, among others. + +3D Generative Adversarial Networks (GANs) have emerged as the state-of-the-art (SOTA) platform in this domain, enabling the generation of diverse 3D assets at interactive framereates [6, 8, 15, 23, 41, 71, 76, 76]. Most existing 3D GANs build on variants of volumetric scene representations combined with neural volume rendering [65]. However, volume rendering is relatively slow and compromises the training of a GAN, which requires tens of millions of forward rendering passes to converge [11]. Mesh-based representations building on fast differentiable rasterization have been proposed to overcome this limitation [19, 71], + +but these are not expressive enough to realistically model features like hair, clothing, or accessories, which deviate significantly from the template mesh. These limitations, which are largely imposed by a tradeoff between efficient or expressive scene representations, have been constraining the quality and resolution of existing 3D GANs. While partially compensated for by using 2D convolutional neural network (CNN)-based upsamplers [11, 22], upsampling leads to multi-view inconsistency in the form of aliasing. + +Very recently, 3D Gaussians have been introduced as a promising neural scene representation offering fast rendering speed and high expressivity [28]. While 3D Gaussians have been explored in the context of single-scene overfitting, their full potential in generative settings has yet to be unlocked. This is challenging because it is not obvious how to combine SOTA CNN-based generators [26, 27] with 3D Gaussian primitives that inherently do not exist on a regular Cartesian grid and that may vary in numbers. + +We introduce Gaussian Shell Maps (GSMs), a 3D GAN framework that intuitively connects CNN generators with 3D Gaussians used as efficient rendering primitives. Inspired by the traditional computer graphics work on shell maps [47], GSMs use the CNN generator to produce texture maps for a set of "shell" meshes appropriately inflated and deflated from the popular SMPL mesh template for human bodies [37]. The textures on the individual shells directly encode the properties of 3D Gaussians, which are sampled on the shell surfaces at fixed locations. The generated images are rendered using highly efficient Gaussian splatting, and articulation of these Gaussians can be naturally enabled through deforming the scaffolding shells with the SMPL model. Since 3D Gaussians have spatial extent, they represent details on, in between, and outside the discrete shells. GSMs are trained exclusively on datasets containing single-view images of human bodies, such as SHHQ [17]. + +Our experiments demonstrate that GSM can generate highly diverse appearances, including loose clothing and accessories, at high resolution, without an upsampler, at a state-of-the-art rendering speed of 125 FPS (or 35FPS including generation). Among various architecture design choices, multiple shells with fixed relative locations of the 3D Gaussian achieve the best results in our experiments. + +Specifically, our contributions include + +1. We propose a novel 3D GAN framework combining a CNN-based generator and 3D Gaussian rendering primitives using shell maps. +2. We demonstrate the fastest 3D GAN architecture to date, achieving real-time rendering of $512^{2}$ px without convolutional upsamplers, with image quality and diversity matching the state of the art on challenging single-view datasets of human bodies. + +# 2. Related Work + +3D-Aware Generative Models. GANs and diffusion models are two very powerful generative models emerging as a result of efficient architectures and high-quality datasets [25, 48]. With the quality of image generation reaching photo-realism, the community started leveraging these SOTA generative models for 3D generation by combining neural rendering methods [64, 65] to produce high-quality multi-view consistent 3D objects from image collections. Several 3D GAN architectures have explored implicit or explicit neural volume representations for modeling 3D objects, including [2, 10, 11, 14, 22, 39, 40, 42, 57, 59, 62, 69, 70]. 3D diffusion models, on the other hand, typically use the priors encoded by pre-trained text-to-image 2D diffusion models, e.g. [4, 7-9, 12, 30, 34, 35, 44, 46, 51, 53, 66] (see this survey for more details [45]). Due to the lack of high-quality, large-scale multi-view datasets curated for specific categories like humans, the choice of a suitable generative model becomes critical. Using diffusion models to generate high-resolution multi-view consistent 3D objects is still an unsolved problem [45]. 3D GANs, on the other hand, exhibit better quality and multi-view consistency at higher resolutions that do not assume multi-view data [11, 15, 62, 71]. This motivates our choice of building an efficient representation using a 3D GAN framework. + +Generative Articulated 3D Digital Humans. 3D-GAN frameworks have been proposed to generate the appearance, geometry, and identity-preserving novel views of digital humans [6, 13, 15, 23, 71, 72, 74, 76]. Most of these GANs are trained on single-view image collections [16, 33]. A popular approach is to use a neural radiance representation [6, 23, 24, 41, 74] where a canonically posed human can be articulated via deformation. Other approaches are based on meshes, where a template can be fixed or learned during the training [3, 18, 19, 63, 72]. Related to this line of work, a concurrent paper, LSV-GAN [71], offsets the SMPL template meshes into layered surfaces and composites the per-layer rasterization result to form the final rendering. While it provides a faster alternative to the volume rendering-based approaches, it can only accommodate a small offset from the SMPL mesh, which hampers diversity. Our GSM method differs from LSV-GAN as we employ 3D Gaussians [28] as primitives on the layered surfaces, which, by having a learnable spatial span, allows for larger deviation from the template mesh and can thus generate more intricate details. + +Point-Based Rendering. Earlier point-based methods efficiently render point clouds and rasterize them by fixing the size [20, 21, 54]. While efficient in terms of speed and parallel rasterization on graphics processing units [32, 56], they are not differentiable [28, 49]. To combine these methods with the neural networks and perform view synthesis, + +![](images/d56a544036a45f6dab10e514f5c5177b741721beb19d76630aaf0a4f39533bdb.jpg) +Figure 2. Method Overview. We propose an expressive yet highly efficient representation, Gaussian Shell Map (GSM), for 3D human generation. Combining the idea of 3D Gaussians and Shell Maps (Sec. 3), we sample 3D Gaussians on "shells", which are mesh layers offsetted from the SMPL template, forming a shell volume to model complex and diverse geometry and appearance; the Gaussian parameters are learned in the texture space, allowing us to leverage existing CNN-based generative architecture (Sec. 4.1). Articulation is straightforward by interpolating the deformation of the shell (Sec. 4.2). The generation is supervised by single-view 2D images using several discriminator critics, including part-specific face, hands, and feet discriminators (Sec. 4.3). + +recent works have developed differentiable point-based rendering techniques [1, 52, 67, 68, 73, 75, 77]. More recently, 3D Gaussian-based point splatting gained traction due to the flexibility of anisotropic covariance and density control with efficient depth sorting [28, 29]. This allows 3D Gaussian splats to handle complex scenes composed of high and low-frequency features. Relevant to human bodies, 3D Gaussians have also been used in pose estimation and tracking [38, 50, 61]. While the Gaussian primitives have been used for efficient scene reconstruction and novel view synthesis, it is not trivial to deploy Gaussians in a generative setup. To the best of our knowledge, our method is the first to propose a combination of 3D Gaussians and 3D GANs. + +# 3. Background + +3D Gaussians. These point-based primitives can be differentiably and efficiently rendered using EWA (elliptical weighted average) volume splatting [78]. 3D Gaussians have recently demonstrated outstanding expressivity for 3D scene reconstruction [28], in which the Gaussian parameters, position $\mathbf{x} \in \mathbb{R}^3$ , opacity $o \in \mathbb{R}$ , color $\mathbf{c} \in \mathbb{R}^{sh}$ ( $sh$ representing the spherical harmonic coefficients), scaling $\mathbf{s} \in \mathbb{R}^3$ , and rotation $\mathbf{q} \in \mathbb{R}^4$ parameterized as quaternions are jointly optimized to minimize the photometric errors of the rendered images in a set of known camera views. The optimization is accompanied by adaptive control of the density, where the points are added or removed based on the density, size, and gradient of the Gaussians. + +Specifically, each Gaussian is defined as + +$$ +G \left(\mathbf {x} ^ {\prime}; \mathbf {x}, \Sigma\right) = \exp^ {- \frac {1}{2} \left(\mathbf {x} ^ {\prime} - \mathbf {x}\right) ^ {\intercal} \Sigma^ {- 1} \left(\mathbf {x} ^ {\prime} - \mathbf {x}\right)}, \tag {1} +$$ + +where $\Sigma = R S S^T R^T$ is the covariance matrix parameterized by the rotation and scaling matrices $R$ and $S$ given by the quaternions $\mathbf{q}$ and scaling $\mathbf{s}$ . + +The image formation is governed by classic point-based $\alpha$ -blending of overlapping Gaussians ordered from closest to farthest [31]: + +$$ +\mathbf {C} = \sum_ {i \in \mathcal {N}} \mathbf {c} _ {i} ^ {\prime} o _ {i} ^ {\prime} \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {\mathbf {j}}), \tag {2} +$$ + +where color $\mathbf{c}_i^{\prime}$ and opacity $\sigma_{i}^{\prime}$ is computed by + +$$ +\mathbf {c} _ {i} ^ {\prime} = G \left(\mathbf {x} ^ {\prime}; \mathbf {x} _ {i}\right) \mathbf {c} _ {i} \quad \text {a n d} \quad o _ {i} ^ {\prime} = G \left(\mathbf {x} ^ {\prime}; \mathbf {x} _ {i}\right) o _ {i}. \tag {3} +$$ + +Thanks to their ability to fit complex geometry and appearance, 3D Gaussians are gaining popularity for 3D scene reconstruction. However, deploying 3D Gaussians for generative tasks remains an unexplored topic and is challenging for several reasons. First, Gaussians are of "Lagrangian" nature - their sparse number and learnable positions are challenging to combine with SOTA "Eulerian" (i.e., grid-based) CNN generators. Second, the parameters of Gaussians are highly correlated. The same radiance field can be equally well explained by many different configurations of Gaussians, varying their locations, sizes, scales, rotations, colors, and opacities. This ambiguity makes it challenging to generalize over a distribution of objects or scenes, which generative methods do. + +Shell Maps. Our representation is related to Shell Maps [14, 47, 58], a technique in computer graphics designed to model near-surface details. Shell maps use 3D texture maps to store the fine-scale features in a shell-like volume close to a given base surface, typically represented as a triangle mesh. In essence, they extend UV + +![](images/96ad5d1d30cf15b47c61c8e54a497541a9817998667630ae96b6559201797931.jpg) +Figure 3. Novel Views and Animation. We can render a generated identity in novel views and articulate it for animations. + +![](images/4ea0818e6b5c2737d2ca2baf51115b51c91de71387ae230c263c36fa628fdfa6.jpg) +Figure 4. Latent Code Interpolation. Latent code interpolation of our model trained on DeepFashion Dataset. + +maps such that every point in the shell volume can be bijectively mapped to the 3D texture map for efficient modeling and rendering in texture space. The shell volume is constructed by offsetting the base mesh along the normal direction while maintaining the topology and avoiding self-intersection. The volume is discretized into tetrahedra that connect the vertices of the base and offset meshes. A unique mapping between the shell space and the texture space can be established by identifying the tetrahedron and subsequently querying the barycentric coordinates inside it. Formally, the mapping from $\mathbf{x}$ in the shell volume to its position in the 3D texture map is defined as + +$$ +\mathbf {x} _ {t} = \phi (\mathbf {T} _ {t}, B (\mathbf {x}, \mathbf {T})) \tag {4} +$$ + +where $\mathbf{T},\mathbf{T}_t$ refer to corresponding tetrahedra in the shell and texture space, $B(\mathbf{x},\mathbf{T})$ is the barycentric coordinates of $\mathbf{x}$ in $\mathbf{T}$ , and $\phi$ denotes barycentric interpolation. + +# 4. Gaussian Shell Maps (GSM) + +GSM is a framework that connects 3D Gaussians with SOTA CNN-based generator networks [11] in a GAN setting. The key idea is to anchor the 3D Gaussians at fixed locations on a set of "shells" derived from the SMPL [37] human body template mesh. These shells span a volume to model surface details that deviate from the unclothed template mesh. We learn the relevant Gaussian parameters in the texture space, allowing us to leverage established CNN-based generative backbones while seamlessly utilizing the + +parametric deformation model for articulation. The overall pipeline is shown in Figure 2. + +# 4.1. Representation + +Our method utilizes the concept of shell maps to leverage the inherent planar structure of texture space for seamless integration with CNN-based generative architectures and, at the same time, encapsulate diverse and complex surface details without directly modifying the template mesh. Specifically, the shell volume is defined by the boundary mesh layers, created by inflating and deflating the T-pose SMPL mesh using Laplacian Smoothing [60] with the smoothing factor set to negative for inflation and positive for deflation. We then represent this shell volume using $N$ mesh layers, i.e. "shells", by linearly spacing the vertices between the aforementioned boundary shells. In parallel, we apply a similar discretization strategy to the 3D texture space, creating $N$ 2D texture maps storing neural features that can be referenced for each shell using the UV mapping inherited from the SMPL template. + +As shown in Figure 2, we use the shell maps to generate all Gaussian parameters except the positions, as those are sampled and anchored w.r.t. to the shells at every iteration (explained below in detail). This results in a feature volume of $\mathsf{T}^{N\times H\times W\times 11}$ , comprised of 3D color $\mathsf{T}^{\mathrm{c}}$ , 1D opacity $\mathsf{T}^o$ , 3D scaling $\mathsf{T}^s$ , and 4D rotation $\mathsf{T}^q$ features. + +We create Gaussians in our shell volume. This is done by sampling a fixed number of Gaussians quasi-uniformly on the shells based on the triangle areas. Once sampled, the Gaussians are anchored on the shells using barycentric coordinates so that every Gaussian center $\mathbf{x}$ can be mapped to a point $\mathbf{x}_t$ on the corresponding shell map using Eq. (4), except that the tetrahedra are replaced with the triangle in which the Gaussian resides. This anchoring is a crucial design choice, as it enables straightforward feature retrieval from the shell maps. More importantly, it simplifies the learning by fixing the Gaussian positions and offloading geometry modeling to opacity $o$ and sigma $\Sigma$ . + +The spatial span of the Gaussians plays a significant role in reconstructing the features defined on the discrete shells into a continuous signal within the 3D space. It enables every point—whether on the shells, between them, or outside them—to receive a valid opacity and color value by evaluat- + +![](images/9d4ccb1dc66ec68769e236587cba67738311be5f642f9de02a0f41f4e1467f3e.jpg) +GNNARF + +![](images/60a80f93c25fa38b46384e0cab92c9bc0bd3aebc0c784de7de2c61a1989ecd07.jpg) +EVA3D + +![](images/2abfb83f66150462ae1694569f8f057f0d47f061b55f19face5e827336def53b.jpg) + +![](images/ac2e726bb7e7dde5d0327859e176d331caf26c3066813aca6df3d2e65bf09cd1.jpg) +s +DeepFashion + +![](images/0362bac430c50fc9a2689dd7de7fed6c1ad9e13dcf4ccb96cdcc77eb234b55fc.jpg) + +![](images/9f63fa811cb63af8ff2daaa0e6bd8a2caf56aeb0298bb39fa81d3bd8bb2bacd8.jpg) +Figure 5. Qualitative Comparison. We compare our results with GNARF and EVA3D baselines on DeepFashion and SHHQ datasets. In each case, we show the deformed body poses of the identities generated by the methods. The competing methods exhibit artifacts marked in red. Notice that our approach generates high-quality textures, like facial details and more realistic deformations. + +ing Eq. (3). This process allows us to model diverse appearances and body shapes, excelling mesh-based representation while at the same time maintaining efficient rendering, which is critical for GAN training. + +Formally, Gaussian opacity, color, scale, and rotation can be interpolated from the shell maps $\mathsf{T}$ + +$$ +f = \mathrm {T} ^ {f} \left(\mathbf {x} _ {t}\right), \text {w h e r e} f = \{o, \mathbf {c}, \mathbf {s}, \mathbf {q} \}. \tag {5} +$$ + +# 4.2. Deformation + +The deformation step updates the Gaussians' locations and orientations based on the SMPL template mesh $\mathbf{M}$ , pose $\theta$ , and shape $\beta$ . Note that, different from SMPL, our template mesh could be any of the shells. Since each Gaussian is anchored on the shells using barycentric coordinates, we can query its new location and orientation simply from the associated vertices. In particular, given the barycentric coordinates $B(\mathbf{x},\mathbf{T})$ of a Gaussian inside triangle $\mathbf{T}$ , and the deformed position $\mathbf{T}_{\mathrm{new}}$ and the rotation quaternions $\mathbf{P}$ on the vertices, we can obtain the new location and orientation of the Gaussian as + +$$ +\mathbf {x} _ {\text {n e w}} = \phi (\mathbf {T} _ {\text {n e w}}, B (\mathbf {x}, \mathbf {T})) \tag {6} +$$ + +$$ +\mathbf {q} _ {\text {n e w}} = \frac {\hat {\mathbf {p}}}{\| \hat {\mathbf {p}} \|} \mathbf {q}, \text {w h e r e} \hat {\mathbf {p}} = \phi (\mathbf {P}, B (\mathbf {x}, \mathbf {T})). \tag {7} +$$ + +The deformed mesh is given by the SMPL deformation model SMPL, and the quaternions $\mathbf{P}$ are a result of the linear blend skinning (LBS) from regressed joints and skinning weights, $\{w_{j}\}_{1}^{N_{J}}$ , i.e., $\mathbf{M}_{\mathrm{new}} = \mathrm{SMPL}(\pmb {\theta},\pmb {\beta},\mathbf{M})$ and $\mathbf{p} = \mathrm{rot2quat}\left(\sum_{j = 1}^{N_J}w_jR_j(\pmb {\theta},J(\pmb {\beta}))\right)$ , where $R_{j}$ is the + +rotation matrix of the $j$ -th joint, and $J$ is the regressor function that maps the shape parameters to the joint locations. + +# 4.3. GAN training + +Generator. Similar to prior work [11], we adopt StyleGAN2 without camera and pose conditioning [6, 71] for the generator. We use separate MLPs with different activations for $\mathbf{c}, o, \mathbf{q}$ , and $\mathbf{s}$ : for $\mathbf{c}$ , we use shifted sigmoid following the practice proposed in Mip-NeRF [5]; for $o$ we use sigmoid to constrain the range to $(0, 1)$ ; for $\mathbf{q}$ we normalize the raw MLP output to ensure they are quaternions (see Eq. (7)); for $\mathbf{s}$ we use clamped exponential activation to limit the size of the Gaussians, which is critical for convergence based on our empirical study. + +Discriminator. Our discriminator closely follows that of [6]. Since it does not have an upsampler, the input to the discriminator is the RGB image concatenated with the alpha channel (foreground mask), which is rendered using Gaussian rasterization with the Gaussian color set to 1 for fake samples and precomputed using off-the-shelf segmentation network [55] for the real samples. We refer to the discriminator using foreground mask "Mask Discriminator". Including the alpha channel helps prevent the white background from bleeding into the appearance, which causes artifacts during articulation. + +Face, Hand, and Feet Discriminator. As the human body and clothes are diverse, the discriminator may choose to focus on these features and provide a weak signal to the facial, hand, and foot areas, which are crucial for visually appealing results. To avoid this problem, we adopt a dedicated Face Discriminator $\mathcal{D}_{\mathrm{face}}$ , Feet-and-Hands Discrimi + +Table 1. Quantitative Evaluation. We compare our method with 3D-GAN baselines using DeepFashion and SHHQ datasets. We compute the FID score to evaluate the quality and diversity of the generated samples. Notice that our scores are comparable to state-of-the-art methods. To evaluate deformation consistency, we compute the PCK metric, where our approach consistently outperforms the baselines. INF represents the inference speed measured in ms/img on an A6000 GPU at $512^2$ resolution. Note our method is the fastest across all competing methods. * numbers are adopted from EVA3D [23]; NA represents Not Available; — represents Not Applicable, and + numbers are provided by the authors. + +
ModelDeep FashionSHHQComp.
FID ↓PCK ↑FID ↓PCK ↑INF. ↓
EG3D*26.38—32.96—38
StyleSDF*92.40—14.12—32
ENARF*77.0343.7480.5440.17104
GNARF33.8597.8314.8498.9672
EVA3D*15.9187.5011.9988.95200
StylePeople17.7298.3114.6798.5828
GetAvatar19.00+NANANA44
AG3D10.93NANANA105
Ours15.7899.4813.3099.2728
+ +nator $\mathcal{D}_{\mathrm{feet/hands}}$ in addition to the main Discriminator $\mathcal{D}_{\mathrm{main}}$ . All of these part-focused discriminators share the same base architecture as the main discriminator, except that we no longer use any conditioning since these cropped images do not contain distinct pose information. The inputs are crops of the corresponding parts, whose spatial span is determined from the SMPL pose and camera parameters. + +Scaling Regularization. In our empirical study, we observe when unconstrained, the network tends to learn overly large or extremely small Gaussians early on during the training and rapidly leads to divergence or model collapse. We experimented with multiple regularization strategies and different activation functions and found the following scaling regularization most effective: + +$$ +\mathcal {L} _ {\text {s c a l e}} = \mathrm {M} \circ \| \mathrm {T} ^ {\mathrm {s}} - s _ {\text {r e f}} \| ^ {2} \text {w i t h} s _ {\text {r e f}} = \frac {1}{P} \sum_ {i = 1} ^ {P} \ln \left(\delta_ {i}\right), \tag {8} +$$ + +where $M$ is the binary mask indicating UV-mapped region on the shell maps, and $s_{\mathrm{ref}}$ is a reference scale determined by the closest-neighbor distance $\delta$ averaged among all the Gaussians. + +# 5. Experiment Settings + +# 5.1. Datasets + +We evaluate our method using the two most common human datasets, DeepFashion [36] and SHHQ [17]. DeepFashion and SHHQ do not provide SMPL parameters, so we use + +Table 2. Ablation: Number of Shells. Ablation on the number of shells performed on $128^{2}$ resolution trained for $1600k$ images. + +
# of Shells1281015
FID ↓40.325.3118.7021.5721.10
+ +SMPLify-X [43] to obtain the SMPL parameters and camera poses. + +# 5.2. Training Details + +By default, we generate the shell maps at $512 \times 512$ resolution for 8 shells with a total of $100k$ Gaussians equally distributed across the shells. The offset between adjacent shells is 0.08. The last fully-connected layer of the scaling prediction is adjusted to ensure the initial scaling is within a reasonable range, which is determined similar to $s_{\mathrm{ref}}$ in Eq. (8). The loss weights for all discriminators are set to 1, whereas the scaling regularization is set to 0.1. $R1$ regularization is used for all the discriminators with a weight of 10. We apply progressive training starting at $256 \times 256$ rendering resolution for $6000k$ training images and progressively grow the resolution to $512 \times 512$ for $1000k$ images, then continue training at the fixed $512 \times 512$ resolution for $3000k$ images. We train our method on 8 A100 GPUs with a batch size of 32. The learning rates are initialized at 0.002 for both the generator and discriminator. + +# 5.3. Evaluation Details + +Baselines. We compare with the following volume rendering and rasterization-based methods: EG3D [11], StyleSDF [42], GNARF [6], ENARF [41], StylePeople [19], EVA3D [23], GetAvatar [76], and AG3D [15]. Among these methods, EG3D and StyleSDF do not support articulation; except StylePeople, all methods use neural field and volumetric rendering. Regardless of representation, all these baselines use 2D convolution for postrendering upsampling and/or postprocessing, causing flickering artifacts (see Supplementary Materials). + +Metrics Following prior work, We use two metrics to evaluate the quality, diversity, and quality of animations: Fréchet Inception Distance (FID) and Percentage of Correct Keypoint (PCK). The former measures the visual quality and diversity of the generated samples (using $50k$ samples and the full dataset). The latter measures how close the generated image aligns with the pose control. We compute PCK on $5k$ samples using the implementation provided by GNARF [6]. + +![](images/5061a58a11f903bee1677c837befcd5079d65d02ff72910bbba64c2fee91c068.jpg) +Figure 6. Multi-view Consistency. The figure shows a closeup of the garment in two different views. The green dashed line follows the body shape and indicates the same 3D position. In AG3D [15], the pattern significantly changes, while ours, utilizing texture maps, has built-in view consistency. + +# 6. Experiment Results + +# 6.1. Qualitative + +Figure 1 showcases some of our generated results. Notice that our model is able to generate diverse attributes, including loose clothing and accessories like hats. Further, in Figure 3, we visualize the generated results in varying camera views and body poses. To demonstrate that the latent space of the proposed model is smooth, we show visual results of latent code interpolation in Figure 4 computed on the DeepFashion trained model. + +Figure 5 shows the visual comparison of our method with the baselines on DeepFashion and SHHQ datasets, with more examples to be found in Supplementary Materials. Volumetric methods, such as GNARF, operate in a canonical pose and rely on accurate correspondence matching between the observed space and the canonical space, which is ill-defined; thus, these methods tend to produce artifacts for limbs, which undergo large deformation and occlusion. Our method is able to model the complex geometry and appearance of the human body, generating characters with loose clothing and accessories (see Figure 1) and producing convincing results under various articulations. + +# 6.2.Quantitative + +In Table 1, we show the results of our method compared to the competing methods. Many of the methods have not released training or data processing code; we have collected these numbers with our best effort from the authors directly. + +Our method performs on par with the SOTA methods in terms of visual quality and diversity measured by FID + +![](images/1b9fef6c9aecdd64f01c8197f09cc9ae1df82de6a0858c0233bf5db73146f6d5.jpg) +Figure 7. Per-shell Visualization. We visualize the per-shell contributions. The shells decompose the human body into different parts, where the inner shells capture the torso, and the outer shells capture the clothing and hair. + +while ranking top in terms of inference speed. By utilizing the texture space, our model can control the articulation in a straightforward manner, contributing to consistent alignment with the pose control input, as reflected by the PCK score. Note that GetAvatar is trained on multiview data, and AG3D uses normal maps for supervision. + +All other baseline methods adopt post-rendering convolution layers to upsample and add texture details, leading to severe aliasing artifacts not reflected in the quantitative evaluation. We demonstrate this in Figure 6, where we render a generated identity in two different views using our method and AG3D, which achieves the lowest FID on the DeepFashion dataset. Notice that with view change, the features produced by AG3D change significantly, leading to unnatural flickering in animation even though the FID is low. + +# 6.3. Ablation + +We conduct ablation studies on different configurations and components of our method to understand their effect on the final model's performance. To identify the trends, we test these configurations at $128^{2}$ resolution on the SHHQ dataset, trained for $1600k$ images without progressive training. $200k$ Gaussians are sampled for all the experiments. + +Number of Shells. The number of shells determines the granularity of variation we can model inside the shell volume, thus impacting the expressivity of our representation. We investigate this effect in Table 2. With only one shell, although the Gaussians can vary in size and cover regions off the shell surface, they still lack the capacity to model diverse human body shapes and appearances. Using two shells es + +![](images/a2d90ac876d564264aef258571024d9c30e2e5ca71ec75dfd587305158ddf2a7.jpg) +w/o Mask and Part Discriminators + +![](images/5f52023679c00c6c1b75fdaabbb3e5d45548ec28397e33ed70575d213272f151.jpg) + +![](images/8996ccb612221ede0468c0437a66707607969c6c94a29d79ca92ab2a135ecccf.jpg) + +![](images/8315f15269a8e6b8ba9f361a68925ff7b4921c7f808c9e732a83baa547dbeee0.jpg) + +![](images/11a8e5b5a6db3dba21b7956cb9ab3b401cbfd52381df0056c5d1d25c74ec94a5.jpg) + +![](images/53baec545b2efe830faa9332178e31fff5ba1e3a7b3fd43315ab2dbffc91c83e.jpg) + +![](images/9432951984b7d118844e22980b54eb950da36c6eeba6f8439f61f833db16351f.jpg) + +![](images/b1ba2b9fb7f5563cd7191c0e1af4e7fe8a5fda9939ed3073f7365caf57d1017a.jpg) + +![](images/c60e8edd65da92bb61cd635ccf89f5d793b5eb36ddb6ceb64f642744034c0f08.jpg) + +![](images/4cb8897cca021a8a05c2bde68570c87fcaf91ffb6f259d296917695fa1ed378a.jpg) +w/ Mask and Part Discriminators +Figure 8. Mask + Part Discriminators. The blue box highlights the improved silhouette after concatenating the alpha channels in the input to the discriminator. The other boxes showcase better details using the face, feet, and hands discriminator. + +mentally defines the boundary for the shell volume. This variation significantly outperforms the single-shell variation, demonstrating the importance of shell volume. Adding 8 shells (our default setting) further improves the visual quality, as the finer discretization enhances the capability to model more complex geometry and appearance. The performance gain stagnates with even more shells, likely because the model struggles to converge when the feature map becomes too large. Figure 7 shows an example identity of what each Gaussian shell has learned: progressing from the innermost to the outermost shell, every shell captures some texture details, and the outer shell models details that deviate from the base template, such as hat and hair. + +Mask and Part Discriminator The mask discriminator involves the inclusion of the alpha channel as part of the input to the RGB images and is very beneficial for disambiguating background by learning plausible human silhouettes. As demonstrated in Figure 8, the model trained without mask discriminator generates disturbing white artifacts outside the human shape. Figure 8 also shows the improvement in small features in the facial area and better structure for hands and feet introduced by facial and feet/hands discriminator. We also quantify the joint improvement of the mask and part discriminator and observed an FID boost from 20.63 to 18.7 at $128^{2}$ resolution. + +Types of Gaussian Anchors. In our representation, we opt to sample the Gaussians on shell meshes. There are other alternatives to sample the Gaussians, including constructing tetrahedra akin to the original Shell Map proposed by Porumbescu et al. [47]; or completely discarding the template by sampling uniformly inside a bounding box. We compared with these alternatives and found the shell repre + +sensation achieves the best visual quality and convergence speed. A more detailed discussion can be found in Supplementary Materials. + +# 7. Discussion + +Limitations. Our method is not without limitations. Like almost all existing 3D human generation approaches, our method also relies on a parametric deformation model to enable articulation, which falls short in handling the dynamics of hair and loose clothing. Due to the irregularity and sparsity of Gaussians, it is not obvious how to extract the accurate geometry and normals. Surface splatting, with a stronger surface prior may be a potential solution, which we plan to investigate in the future. Lastly, while ranking among the best methods, our current results still cannot achieve photorealism. The challenge stems from the complexity of human appearance. A promising direction is to combine a small amount of multi-view studio captures with in-the-wild datasets to leverage priors discovered in controlled and richly annotated datasets. + +Ethical Concerns. GANs, like the one we developed, carry the risk of being utilized for creating altered images of real individuals. This inappropriate application of image generation methods represents a danger to society, and we firmly oppose the use of our research for disseminating false information or damaging reputations. Additionally, we acknowledge that there might be a deficit in diversity in our outcomes, which could be a consequence of inherent biases in the datasets we utilize. + +Conclusions. In this work, we present Gaussian Shell Maps (GSMs), a novel framework that effectively combines CNN-based generators with 3D Gaussian rendering primitives for the generation of digital humans. Our experiments demonstrate the potential of GSMs in generating diverse and detailed human figures, including complex features like clothing and accessories. The framework shows promise in enhancing the efficiency of rendering processes, a crucial factor for real-time applications in industries like virtual reality and cinematic production. The potential of GSMs in practical scenarios invites further exploration, and continued research could enhance digital human modeling. + +Acknowledgements. We thank the authors of 3D Gaussian Splitting, Bernhard Kerbl, Georgios Kopanas, and George Drettakis for the discussions on the code. We thank Thabo Beeler and Guandao Yang for fruitful discussions and Alexander Bergman for helping with the baseline comparisons. We also thank the authors of GetAvartar for providing additional evaluation results. This work was in part supported by Google, Samsung, Stanford HAI, and the Swiss Postdoc Mobility Fund. + +# References + +[1] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In ECCV, 2020. 3 +[2] Sizhe An, Hongyi Xu, Yichun Shi, Guoxian Song, Umit Y Ogras, and Linjie Luo. Panohead: Geometry-aware 3d full-head synthesis in 360deg. In CVPR, 2023. 2 +[3] Shivangi Aneja, Justus Thies, Angela Dai, and Matthias Nießner. ClipFace: Text-guided Editing of Textured 3D Morphable Models. In ArXiv preprint arXiv:2212.01406, 2022. 2 +[4] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. ArXiv, abs/2211.01324, 2022.2 +[5] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, 2021. 5 +[6] Alexander Bergman, Petr Kellnhofer, Wang Yifan, Eric Chan, David Lindell, and Gordon Wetzstein. Generative neural articulated radiance fields. NeurIPS, 2022. 1, 2, 5, 6 +[7] Alexander W. Bergman, Wang Yifan, and Gordon Wetzstein. Articulated 3d head avatar generation using text-to-image diffusion models. 2023. 2 +[8] Yukang Cao, Yan-Pei Cao, Kai Han, Ying Shan, and KwanYee K. Wong. Dreamavatar: Text-and-shape guided 3d human avatar generation via diffusion models, 2023. 1 +[9] Eric Chan, Koki Nagano, Matthew A. Chan, Alexander W. Bergman, Jeong Joon Park, Axel Levy, Miika Aittala, Shalini De Mello, Tero Karras, and Gordon Wetzstein. Generative novel view synthesis with 3d-aware diffusion models. ArXiv, abs/2304.02602, 2023. 2 +[10] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In CVPR, 2021. 2 +[11] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In CVPR, 2022. 1, 2, 4, 5, 6 +[12] Rui Chen, Y. Chen, Ningxin Jiao, and Kui Jia. *Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation.* ArXiv, abs/2303.13873, 2023. 2 +[13] Xinya Chen, Jiaxin Huang, Yanrui Bin, Lu Yu, and Yiyi Liao. Veri3d: Generative vertex-based radiance fields for 3d controllable human image synthesis, 2023. 2 +[14] Stavros Diolatzis, Jan Novak, Fabrice Rousselle, Jonathan Granskog, Miika Aittala, Ravi Ramamoorthi, and George Drettakis. Mesogan: Generative neural reflectance shells. Comput. Graph. Forum, 2023. 2, 3 + +[15] Zijian Dong, Xu Chen, Jinlong Yang, Michael J.Black, Otmar Hilliges, and Andreas Geiger. AG3D: Learning to generate 3D avatars from 2D image collections. In ICCV, 2023. 1, 2, 6, 7 +[16] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Chen Qian, Chen Change Loy, Wayne Wu, and Ziwei Liu. Stylegan-human: A data-centric odyssey of human generation. In ECCV, 2022. 2 +[17] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Chen Qian, Chen Change Loy, Wayne Wu, and Ziwei Liu. Stylegan-human: A data-centric odyssey of human generation. In ECCV, 2022. 2, 6 +[18] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. NeurIPS, 2022. 2 +[19] Artur Grigorev, Karim Iskakov, Anastasia Iinina, Renat Bashirov, Ilya Zakharkin, Alexander Vakhitov, and Victor Lempitsky. Stylepeople: A generative model of fullbody human avatars. In CVPR, 2021. 1, 2, 6 +[20] Markus Gross and Hanspeter Pfister. Point-based graphics. Elsevier, 2011. 2 +[21] Jeffrey P Grossman and William J Dally. Point sample rendering. In Proceedings of the Eurographics Workshop, 1998. 2 +[22] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d aware generator for high-resolution image synthesis. In ICLR, 2022. 2 +[23] Fangzhou Hong, Zhaoxi Chen, Yushi LAN, Liang Pan, and Ziwei Liu. EVA3d: Compositional 3d human generation from 2d image collections. In ICLR, 2023. 1, 2, 6 +[24] Suyi Jiang, Haoran Jiang, Ziyu Wang, Haimin Luo, Wenzheng Chen, and Lan Xu. Humangen: Generating human radiance fields with explicit priors. In CVPR, 2023. 2 +[25] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 2 +[26] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In Proc. CVPR, 2020. 2 +[27] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In NeurIPS, 2021. 2 +[28] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM TOG, 2023. 2, 3 +[29] Leonid Keselman and Martial Hebert. Approximate differentiable rendering with algebraic surfaces. In European Conference on Computer Vision (ECCV), 2022. 3 +[30] Nikos Kolotouros, Thiemo Alldieck, Andrei Zanfir, Eduard Gabriel Bazavan, Mihai Fieraru, and Cristian Sminchisescu. Dreamhuman: Animatable 3d avatars from text. ArXiv, abs/2306.09329, 2023. 2 +[31] Georgios Kopanas, Julien Philip, Thomas Leimkuhler, and George Drettakis. Point-based neural rendering with perview optimization. In Comput. Graph. Forum, 2021. 3 + +[32] Samuli Laine and Tero Karras. High-performance software rasterization on gpus. In Proceedings of the ACM SIGGRAPH Symposium on High Performance Graphics, 2011. 2 +[33] Ruilong Li, Shan Yang, David A. Ross, and Angjoo Kanazawa. Learn to dance with aist++: Music conditioned 3d dance generation, 2021. 2 +[34] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. ArXiv, abs/2211.10440, 2022. 2 +[35] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. ArXiv, abs/2303.11328, 2023. 2 +[36] Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In CVPR, 2016. 6 +[37] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. ACM TOG, 2015. 1, 2, 4 +[38] Youssef A Mejjati, Isa Milefchik, Aaron Gokaslan, Oliver Wang, Kwang In Kim, and James Tompkin. Gaussian: Controllable image synthesis with 3d gaussians from unposed silhouettes. arXiv preprint arXiv:2106.13215, 2021. 3 +[39] Michael Niemeyer and Andreas Geiger. Campari: Camera-aware decomposed generative neural radiance fields. In 3DV, 2021. 2 +[40] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In CVPR, 2021. 2 +[41] Atsuhiro Noguchi, Xiao Sun, Stephen Lin, and Tatsuya Harada. Unsupervised learning of efficient geometry-aware neural articulated representations. In ECCV, 2022. 1, 2, 6 +[42] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. StyleSDF: High-Resolution 3D-Consistent Image and Geometry Generation. arXiv preprint arXiv:2112.11427, 2021. 2, 6 +[43] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In CVPR, 2019. 6 +[44] Ryan Po and Gordon Wetzstein. Compositional 3d scene generation using locally conditioned diffusion. arXiv preprint arXiv:2303.12218, 2023. 2 +[45] Ryan Po, Wang Yifan, Vladislav Golyanik, Kfir Aberman, Jonathan T. Barron, Amit H. Bermano, Eric Ryan Chan, Tali Dekel, Aleksander Holynski, Angjoo Kanazawa, C. Karen Liu, Lingjie Liu, Ben Mildenhall, Matthias Nießner, Björn Ommer, Christian Theobalt, Peter Wonka, and Gordon Wetzstein. State of the art on diffusion models for visual computing, 2023. 2 +[46] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2 + +[47] Serban D. Porumbescu, Brian Budge, Louis Feng, and Kenneth I. Joy. Shell maps. In ACM SIGGRAPH, 2005. 2, 3, 8 +[48] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2 +[49] Liu Ren, Hanspeter Pfister, and Matthias Zwicker. Object space ewa surface splatting: A hardware accelerated approach to high quality point rendering. In Comput. Graph. Forum, 2002. 2 +[50] Helge Rhodin, Nadia Robertini, Christian Richardt, Hans-Peter Seidel, and Christian Theobalt. A versatile scene model with differentiable visibility applied to generative pose estimation. In ICCV, 2015. 3 +[51] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. CVPR, 2021. 2 +[52] Darius Rückert, Linus Franke, and Marc Stamminger. Adop: Approximate differentiable one-pixel point rendering. ACM TOG, 2022. 3 +[53] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Seyedeh Sara Mahdavi, Raphael Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. ArXiv, abs/2205.11487, 2022. 2 +[54] Miguel Sainz and Renato Pajarola. Point-based rendering techniques. Computers & Graphics, 28(6):869-879, 2004. 2 +[55] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In CVPR, 2018. 5 +[56] Markus Schütz, Bernhard Kerbl, and Michael Wimmer. Software rasterization of 2 billion points in real time. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 2022. 2 +[57] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. In NeurIPS, 2020. 2 +[58] Zackary P. T. Sin, Peter H. F. Ng, and Hong Va Leong. Ner-fahedron: A primitive for animatable neural rendering with interactive speed. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 2023. 3 +[59] Ivan Skorokhodov, Aliaksandr Siarohin, Yinghao Xu, Jian Ren, Hsin-Ying Lee, Peter Wonka, and Sergey Tulyakov. 3d generation on imagenet. In ICLR, 2023. 2 +[60] Olga Sorkine, Daniel Cohen-Or, Yaron Lipman, Marc Alexa, Christian Rössl, and H-P Seidel. Laplacian surface editing. In Eurographics, 2004. 4 +[61] Carsten Stoll, Nils Hasler, Juergen Gall, Hans-Peter Seidel, and Christian Theobalt. Fast articulated motion tracking using a sums of gaussians body model. In ICCV, 2011. 3 +[62] Jingxiang Sun, Xuan Wang, Yichun Shi, Lizhen Wang, Jue Wang, and Yebin Liu. Ide-3d: Interactive disentangled editing for high-resolution 3d-aware portrait synthesis. ACM TOG, 2022. 2 + +[63] Jingxiang Sun, Xuan Wang, Lizhen Wang, Xiaoyu Li, Yong Zhang, Hongwen Zhang, and Yebin Liu. Next3d: Generative neural texture rasterization for 3d-aware head avatars. In CVPR, 2023. 2 +[64] Ayush Tewari, Ohad Fried, Justus Thies, Vincent Sitzmann, Stephen Lombardi, Kalyan Sunkavalli, Ricardo MartinBrualla, Tomas Simon, Jason Saragih, Matthias Nießner, et al. State of the art on neural rendering. In Comput. Graph. Forum, 2020. 2 +[65] Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srinivasan, Edgar Tretschk, Wang Yifan, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lombardi, et al. Advances in neural rendering. In Comput. Graph. Forum, 2022. 1, 2 +[66] Christina Tsalicoglou, Fabian Manhardt, Alessio Tonioni, Michael Niemeyer, and Federico Tombari. Textmesh: Generation of realistic 3d meshes from text prompts. ArXiv, abs/2304.12439, 2023. 2 +[67] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. Synsin: End-to-end view synthesis from a single image. In CVPR, 2020. 3 +[68] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In CVPR, 2022. 3 +[69] Yinghao Xu, Sida Peng, Ceyuan Yang, Yujun Shen, and Bolei Zhou. 3d-aware image synthesis via learning structural and textural representations. arXiv preprint arXiv:2112.10759, 2021. 2 +[70] Yinghao Xu, Menglei Chai, Zifan Shi, Sida Peng, Ivan Skorokhodov, Aliaksandr Siarohin, Ceyuan Yang, Yujun Shen, Hsin-Ying Lee, Bolei Zhou, et al. Discoscope: Spatially disentangled generative radiance fields for controllable 3d-aware scene synthesis. In CVPR, 2023. 2 +[71] Yinghao Xu, Wang Yifan, Alexander W. Bergman, Mengei Chai, Bolei Zhou, and Gordon Wetzstein. Efficient 3d articulated human generation with layered surface volumes. In 3DV, 2024. 1, 2, 5 +[72] Zhuoqian Yang, Shikai Li, Wayne Wu, and Bo Dai. 3dhuman: 3d-aware human image generation with 3d pose mapping. In ICCV, 2023. 2 +[73] Wang Yifan, Felice Serena, Shihao Wu, Cengiz Oztireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM TOG, 2019. 3 +[74] Jianfeng Zhang, Zihang Jiang, Dingdong Yang, Hongyi Xu, Yichun Shi, Guoxian Song, Zhongcong Xu, Xinchao Wang, and Jiashi Feng. Avatargen: A 3d generative model for animatable human avatars. ArXiv, 2023. 2 +[75] Qiang Zhang, Seung-Hwan Baek, Szymon Rusinkiewicz, and Felix Heide. Differentiable point-based radiance fields for efficient view synthesis. In SIGGRAPH Asia, 2022. 3 +[76] Xuanmeng Zhang, Jianfeng Zhang, Chacko Rohan, Hongyi Xu, Guoxian Song, Yi Yang, and Jiashi Feng. Getavatar: Generative textured meshes for animatable human avatars. In ICCV, 2023. 1, 2, 6 +[77] Yufeng Zheng, Wang Yifan, Gordon Wetzstein, Michael J Black, and Otmar Hilliges. Pointavatar: Deformable point-based head avatars from videos. In CVPR, 2023. 3 + +[78] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Ewa volume splatting. In Proceedings Visualization, 2001. 3 \ No newline at end of file diff --git a/2024/Gaussian Shell Maps for Efficient 3D Human Generation/images.zip b/2024/Gaussian Shell Maps for Efficient 3D Human Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..849f2eab78474ab8a2059cc73edc3a0dad1ac522 --- /dev/null +++ b/2024/Gaussian Shell Maps for Efficient 3D Human Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1e16c2bbf85eeca1f5a521b43016bdbe32365ca4be232a22e87b6f2120e5e787 +size 472569 diff --git a/2024/Gaussian Shell Maps for Efficient 3D Human Generation/layout.json b/2024/Gaussian Shell Maps for Efficient 3D Human Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0ac8ec573cc43e8f7498884b5444f6b7653e6a52 --- /dev/null +++ b/2024/Gaussian Shell Maps for Efficient 3D Human Generation/layout.json @@ -0,0 +1,9641 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 120, + 61, + 474, + 78 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 61, + 474, + 78 + ], + "spans": [ + { + "bbox": [ + 120, + 61, + 474, + 78 + ], + "type": "text", + "content": "Gaussian Shell Maps for Efficient 3D Human Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 99, + 536, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 99, + 536, + 174 + ], + "spans": [ + { + "bbox": [ + 58, + 99, + 536, + 174 + ], + "type": "text", + "content": "Rameen Abdal\\*1 Wang Yifan\\*1 Zifan Shi\\*1,2 Yinghao Xu" + }, + { + "bbox": [ + 58, + 99, + 536, + 174 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 58, + 99, + 536, + 174 + ], + "type": "text", + "content": " Ryan Po" + }, + { + "bbox": [ + 58, + 99, + 536, + 174 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 58, + 99, + 536, + 174 + ], + "type": "text", + "content": " Zhengfei Kuang" + }, + { + "bbox": [ + 58, + 99, + 536, + 174 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 58, + 99, + 536, + 174 + ], + "type": "text", + "content": " Qifeng Chen" + }, + { + "bbox": [ + 58, + 99, + 536, + 174 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 58, + 99, + 536, + 174 + ], + "type": "text", + "content": " Dit-Yan Yeung" + }, + { + "bbox": [ + 58, + 99, + 536, + 174 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 58, + 99, + 536, + 174 + ], + "type": "text", + "content": " Gordon Wetzstein" + }, + { + "bbox": [ + 58, + 99, + 536, + 174 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 58, + 99, + 536, + 174 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 58, + 99, + 536, + 174 + ], + "type": "text", + "content": "Stanford University, USA " + }, + { + "bbox": [ + 58, + 99, + 536, + 174 + ], + "type": "inline_equation", + "content": "{}^{2}" + }, + { + "bbox": [ + 58, + 99, + 536, + 174 + ], + "type": "text", + "content": "HKUST, Hong Kong" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 56, + 201, + 536, + 324 + ], + "blocks": [ + { + "bbox": [ + 56, + 201, + 536, + 324 + ], + "lines": [ + { + "bbox": [ + 56, + 201, + 536, + 324 + ], + "spans": [ + { + "bbox": [ + 56, + 201, + 536, + 324 + ], + "type": "image", + "image_path": "33b9ecb905bebd4160c3f5af610ac3b1bca30e445a3e1437885d0330a2a73f02.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 327, + 547, + 382 + ], + "lines": [ + { + "bbox": [ + 46, + 327, + 547, + 382 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 547, + 382 + ], + "type": "text", + "content": "Figure 1. Gaussian Shell Maps. Gaussian Shell Maps is an efficient framework for 3D human generation connecting 3D Gaussians with CNN-based generators. 3D Gaussians are anchored to \"shells\" derived from the SMPL template [37] (only two shells are visualized for clarity), and the appearance is modeled in texture space. Trained only on 2D images, we show that our method can generate diverse articulable humans in real-time with state-of-the-art quality directly in high resolution without the need for upsampling and hence avoiding aliasing artifacts." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 392, + 192, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 392, + 192, + 404 + ], + "spans": [ + { + "bbox": [ + 143, + 392, + 192, + 404 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "type": "text", + "content": "Efficient generation of 3D digital humans is important in several industries, including virtual reality, social media, and cinematic production. 3D generative adversarial networks (GANs) have demonstrated state-of-the-art (SOTA) quality and diversity for generated assets. Current 3D GAN architectures, however, typically rely on volume representations, which are slow to render, thereby hampering the GAN training and requiring multi-view-inconsistent 2D upsamplers. Here, we introduce Gaussian Shell Maps (GSMs) as a framework that connects SOTA generator network architectures with emerging 3D Gaussian rendering primitives using an articulable multi shell-based scaffold. In this setting, a CNN generates a 3D texture stack with features that are mapped to the shells. The latter represent inflated and deflated versions of a template surface of a digital human in a canonical body pose. Instead of rasterizing the shells directly, we sample 3D Gaussians on the shells whose attributes are encoded in the texture features. These Gaussians are efficiently and differentiably rendered. The ability to articulate the shells is important during GAN training and, at inference time, to deform a body into arbitrary user-defined poses. Our efficient rendering scheme bypasses the need for view-inconsistent upsamplers and achieves high-quality multi-view consistent renderings at a native resolu" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 393, + 545, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 393, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 305, + 393, + 545, + 429 + ], + "type": "text", + "content": "tion of " + }, + { + "bbox": [ + 305, + 393, + 545, + 429 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 305, + 393, + 545, + 429 + ], + "type": "text", + "content": " pixels. We demonstrate that GSMs successfully generate 3D humans when trained on single-view datasets, including SHHQ and DeepFashion." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 430, + 539, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 430, + 539, + 441 + ], + "spans": [ + { + "bbox": [ + 306, + 430, + 539, + 441 + ], + "type": "text", + "content": "Project Page: rameenabdal.github.io/GaussianShellMaps" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 463, + 387, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 463, + 387, + 475 + ], + "spans": [ + { + "bbox": [ + 306, + 463, + 387, + 475 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 483, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 545, + 555 + ], + "type": "text", + "content": "The ability to generate articulable three-dimensional digital humans augments traditional asset creation and animation workflows, which are laborious and costly. Such generative artificial intelligence-fueled workflows are crucial in several applications, including communication, cinematic production, and interactive gaming, among others." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 555, + 547, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 547, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 547, + 687 + ], + "type": "text", + "content": "3D Generative Adversarial Networks (GANs) have emerged as the state-of-the-art (SOTA) platform in this domain, enabling the generation of diverse 3D assets at interactive framereates [6, 8, 15, 23, 41, 71, 76, 76]. Most existing 3D GANs build on variants of volumetric scene representations combined with neural volume rendering [65]. However, volume rendering is relatively slow and compromises the training of a GAN, which requires tens of millions of forward rendering passes to converge [11]. Mesh-based representations building on fast differentiable rasterization have been proposed to overcome this limitation [19, 71]," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 320, + 693, + 391, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 693, + 391, + 703 + ], + "spans": [ + { + "bbox": [ + 320, + 693, + 391, + 703 + ], + "type": "text", + "content": "* Equal Contribution" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 320, + 703, + 537, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 703, + 537, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 703, + 537, + 713 + ], + "type": "text", + "content": "† Work done as a visiting student researcher at Stanford University" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "9441" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "but these are not expressive enough to realistically model features like hair, clothing, or accessories, which deviate significantly from the template mesh. These limitations, which are largely imposed by a tradeoff between efficient or expressive scene representations, have been constraining the quality and resolution of existing 3D GANs. While partially compensated for by using 2D convolutional neural network (CNN)-based upsamplers [11, 22], upsampling leads to multi-view inconsistency in the form of aliasing." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 186, + 289, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 186, + 289, + 294 + ], + "spans": [ + { + "bbox": [ + 46, + 186, + 289, + 294 + ], + "type": "text", + "content": "Very recently, 3D Gaussians have been introduced as a promising neural scene representation offering fast rendering speed and high expressivity [28]. While 3D Gaussians have been explored in the context of single-scene overfitting, their full potential in generative settings has yet to be unlocked. This is challenging because it is not obvious how to combine SOTA CNN-based generators [26, 27] with 3D Gaussian primitives that inherently do not exist on a regular Cartesian grid and that may vary in numbers." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 299, + 289, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 299, + 289, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 299, + 289, + 502 + ], + "type": "text", + "content": "We introduce Gaussian Shell Maps (GSMs), a 3D GAN framework that intuitively connects CNN generators with 3D Gaussians used as efficient rendering primitives. Inspired by the traditional computer graphics work on shell maps [47], GSMs use the CNN generator to produce texture maps for a set of \"shell\" meshes appropriately inflated and deflated from the popular SMPL mesh template for human bodies [37]. The textures on the individual shells directly encode the properties of 3D Gaussians, which are sampled on the shell surfaces at fixed locations. The generated images are rendered using highly efficient Gaussian splatting, and articulation of these Gaussians can be naturally enabled through deforming the scaffolding shells with the SMPL model. Since 3D Gaussians have spatial extent, they represent details on, in between, and outside the discrete shells. GSMs are trained exclusively on datasets containing single-view images of human bodies, such as SHHQ [17]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 509, + 288, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 509, + 288, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 509, + 288, + 593 + ], + "type": "text", + "content": "Our experiments demonstrate that GSM can generate highly diverse appearances, including loose clothing and accessories, at high resolution, without an upsampler, at a state-of-the-art rendering speed of 125 FPS (or 35FPS including generation). Among various architecture design choices, multiple shells with fixed relative locations of the 3D Gaussian achieve the best results in our experiments." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 597, + 215, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 597, + 215, + 610 + ], + "spans": [ + { + "bbox": [ + 59, + 597, + 215, + 610 + ], + "type": "text", + "content": "Specifically, our contributions include" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 616, + 286, + 711 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 616, + 286, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 286, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 286, + 651 + ], + "type": "text", + "content": "1. We propose a novel 3D GAN framework combining a CNN-based generator and 3D Gaussian rendering primitives using shell maps." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 651, + 286, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 286, + 711 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 286, + 711 + ], + "type": "text", + "content": "2. We demonstrate the fastest 3D GAN architecture to date, achieving real-time rendering of " + }, + { + "bbox": [ + 55, + 651, + 286, + 711 + ], + "type": "inline_equation", + "content": "512^{2}" + }, + { + "bbox": [ + 55, + 651, + 286, + 711 + ], + "type": "text", + "content": " px without convolutional upsamplers, with image quality and diversity matching the state of the art on challenging single-view datasets of human bodies." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 91, + 547, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 547, + 380 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 547, + 380 + ], + "type": "text", + "content": "3D-Aware Generative Models. GANs and diffusion models are two very powerful generative models emerging as a result of efficient architectures and high-quality datasets [25, 48]. With the quality of image generation reaching photo-realism, the community started leveraging these SOTA generative models for 3D generation by combining neural rendering methods [64, 65] to produce high-quality multi-view consistent 3D objects from image collections. Several 3D GAN architectures have explored implicit or explicit neural volume representations for modeling 3D objects, including [2, 10, 11, 14, 22, 39, 40, 42, 57, 59, 62, 69, 70]. 3D diffusion models, on the other hand, typically use the priors encoded by pre-trained text-to-image 2D diffusion models, e.g. [4, 7-9, 12, 30, 34, 35, 44, 46, 51, 53, 66] (see this survey for more details [45]). Due to the lack of high-quality, large-scale multi-view datasets curated for specific categories like humans, the choice of a suitable generative model becomes critical. Using diffusion models to generate high-resolution multi-view consistent 3D objects is still an unsolved problem [45]. 3D GANs, on the other hand, exhibit better quality and multi-view consistency at higher resolutions that do not assume multi-view data [11, 15, 62, 71]. This motivates our choice of building an efficient representation using a 3D GAN framework." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 384, + 547, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 547, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 547, + 635 + ], + "type": "text", + "content": "Generative Articulated 3D Digital Humans. 3D-GAN frameworks have been proposed to generate the appearance, geometry, and identity-preserving novel views of digital humans [6, 13, 15, 23, 71, 72, 74, 76]. Most of these GANs are trained on single-view image collections [16, 33]. A popular approach is to use a neural radiance representation [6, 23, 24, 41, 74] where a canonically posed human can be articulated via deformation. Other approaches are based on meshes, where a template can be fixed or learned during the training [3, 18, 19, 63, 72]. Related to this line of work, a concurrent paper, LSV-GAN [71], offsets the SMPL template meshes into layered surfaces and composites the per-layer rasterization result to form the final rendering. While it provides a faster alternative to the volume rendering-based approaches, it can only accommodate a small offset from the SMPL mesh, which hampers diversity. Our GSM method differs from LSV-GAN as we employ 3D Gaussians [28] as primitives on the layered surfaces, which, by having a learnable spatial span, allows for larger deviation from the template mesh and can thus generate more intricate details." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": "Point-Based Rendering. Earlier point-based methods efficiently render point clouds and rasterize them by fixing the size [20, 21, 54]. While efficient in terms of speed and parallel rasterization on graphics processing units [32, 56], they are not differentiable [28, 49]. To combine these methods with the neural networks and perform view synthesis," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "9442" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 71, + 541, + 235 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 541, + 235 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 541, + 235 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 541, + 235 + ], + "type": "image", + "image_path": "d56a544036a45f6dab10e514f5c5177b741721beb19d76630aaf0a4f39533bdb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 239, + 547, + 307 + ], + "lines": [ + { + "bbox": [ + 46, + 239, + 547, + 307 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 547, + 307 + ], + "type": "text", + "content": "Figure 2. Method Overview. We propose an expressive yet highly efficient representation, Gaussian Shell Map (GSM), for 3D human generation. Combining the idea of 3D Gaussians and Shell Maps (Sec. 3), we sample 3D Gaussians on \"shells\", which are mesh layers offsetted from the SMPL template, forming a shell volume to model complex and diverse geometry and appearance; the Gaussian parameters are learned in the texture space, allowing us to leverage existing CNN-based generative architecture (Sec. 4.1). Articulation is straightforward by interpolating the deformation of the shell (Sec. 4.2). The generation is supervised by single-view 2D images using several discriminator critics, including part-specific face, hands, and feet discriminators (Sec. 4.3)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 327, + 289, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 289, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 289, + 483 + ], + "type": "text", + "content": "recent works have developed differentiable point-based rendering techniques [1, 52, 67, 68, 73, 75, 77]. More recently, 3D Gaussian-based point splatting gained traction due to the flexibility of anisotropic covariance and density control with efficient depth sorting [28, 29]. This allows 3D Gaussian splats to handle complex scenes composed of high and low-frequency features. Relevant to human bodies, 3D Gaussians have also been used in pose estimation and tracking [38, 50, 61]. While the Gaussian primitives have been used for efficient scene reconstruction and novel view synthesis, it is not trivial to deploy Gaussians in a generative setup. To the best of our knowledge, our method is the first to propose a combination of 3D Gaussians and 3D GANs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 497, + 126, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 497, + 126, + 510 + ], + "spans": [ + { + "bbox": [ + 47, + 497, + 126, + 510 + ], + "type": "text", + "content": "3. Background" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 517, + 287, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 517, + 287, + 673 + ], + "spans": [ + { + "bbox": [ + 46, + 517, + 287, + 673 + ], + "type": "text", + "content": "3D Gaussians. These point-based primitives can be differentiably and efficiently rendered using EWA (elliptical weighted average) volume splatting [78]. 3D Gaussians have recently demonstrated outstanding expressivity for 3D scene reconstruction [28], in which the Gaussian parameters, position " + }, + { + "bbox": [ + 46, + 517, + 287, + 673 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 517, + 287, + 673 + ], + "type": "text", + "content": ", opacity " + }, + { + "bbox": [ + 46, + 517, + 287, + 673 + ], + "type": "inline_equation", + "content": "o \\in \\mathbb{R}" + }, + { + "bbox": [ + 46, + 517, + 287, + 673 + ], + "type": "text", + "content": ", color " + }, + { + "bbox": [ + 46, + 517, + 287, + 673 + ], + "type": "inline_equation", + "content": "\\mathbf{c} \\in \\mathbb{R}^{sh}" + }, + { + "bbox": [ + 46, + 517, + 287, + 673 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 517, + 287, + 673 + ], + "type": "inline_equation", + "content": "sh" + }, + { + "bbox": [ + 46, + 517, + 287, + 673 + ], + "type": "text", + "content": " representing the spherical harmonic coefficients), scaling " + }, + { + "bbox": [ + 46, + 517, + 287, + 673 + ], + "type": "inline_equation", + "content": "\\mathbf{s} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 517, + 287, + 673 + ], + "type": "text", + "content": ", and rotation " + }, + { + "bbox": [ + 46, + 517, + 287, + 673 + ], + "type": "inline_equation", + "content": "\\mathbf{q} \\in \\mathbb{R}^4" + }, + { + "bbox": [ + 46, + 517, + 287, + 673 + ], + "type": "text", + "content": " parameterized as quaternions are jointly optimized to minimize the photometric errors of the rendered images in a set of known camera views. The optimization is accompanied by adaptive control of the density, where the points are added or removed based on the density, size, and gradient of the Gaussians." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 674, + 223, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 674, + 223, + 687 + ], + "spans": [ + { + "bbox": [ + 59, + 674, + 223, + 687 + ], + "type": "text", + "content": "Specifically, each Gaussian is defined as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 83, + 697, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 697, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 83, + 697, + 287, + 715 + ], + "type": "interline_equation", + "content": "G \\left(\\mathbf {x} ^ {\\prime}; \\mathbf {x}, \\Sigma\\right) = \\exp^ {- \\frac {1}{2} \\left(\\mathbf {x} ^ {\\prime} - \\mathbf {x}\\right) ^ {\\intercal} \\Sigma^ {- 1} \\left(\\mathbf {x} ^ {\\prime} - \\mathbf {x}\\right)}, \\tag {1}", + "image_path": "483b114cb8bf986753ded51c2377a07b1af606d9c86072927e336e0199ba6393.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 326, + 545, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 326, + 545, + 362 + ], + "spans": [ + { + "bbox": [ + 305, + 326, + 545, + 362 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 326, + 545, + 362 + ], + "type": "inline_equation", + "content": "\\Sigma = R S S^T R^T" + }, + { + "bbox": [ + 305, + 326, + 545, + 362 + ], + "type": "text", + "content": " is the covariance matrix parameterized by the rotation and scaling matrices " + }, + { + "bbox": [ + 305, + 326, + 545, + 362 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 305, + 326, + 545, + 362 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 326, + 545, + 362 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 305, + 326, + 545, + 362 + ], + "type": "text", + "content": " given by the quaternions " + }, + { + "bbox": [ + 305, + 326, + 545, + 362 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 305, + 326, + 545, + 362 + ], + "type": "text", + "content": " and scaling " + }, + { + "bbox": [ + 305, + 326, + 545, + 362 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 305, + 326, + 545, + 362 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 362, + 545, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 362, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 305, + 362, + 545, + 398 + ], + "type": "text", + "content": "The image formation is governed by classic point-based " + }, + { + "bbox": [ + 305, + 362, + 545, + 398 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 305, + 362, + 545, + 398 + ], + "type": "text", + "content": "-blending of overlapping Gaussians ordered from closest to farthest [31]:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 367, + 403, + 545, + 436 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 403, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 367, + 403, + 545, + 436 + ], + "type": "interline_equation", + "content": "\\mathbf {C} = \\sum_ {i \\in \\mathcal {N}} \\mathbf {c} _ {i} ^ {\\prime} o _ {i} ^ {\\prime} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {\\mathbf {j}}), \\tag {2}", + "image_path": "c61ae62c713e8eab6dfffcda5e0ecd3de04fd43abae687b51c0bdbd2273ef18b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 440, + 492, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 440, + 492, + 453 + ], + "spans": [ + { + "bbox": [ + 306, + 440, + 492, + 453 + ], + "type": "text", + "content": "where color " + }, + { + "bbox": [ + 306, + 440, + 492, + 453 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i^{\\prime}" + }, + { + "bbox": [ + 306, + 440, + 492, + 453 + ], + "type": "text", + "content": " and opacity " + }, + { + "bbox": [ + 306, + 440, + 492, + 453 + ], + "type": "inline_equation", + "content": "\\sigma_{i}^{\\prime}" + }, + { + "bbox": [ + 306, + 440, + 492, + 453 + ], + "type": "text", + "content": " is computed by" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 336, + 456, + 545, + 471 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 456, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 336, + 456, + 545, + 471 + ], + "type": "interline_equation", + "content": "\\mathbf {c} _ {i} ^ {\\prime} = G \\left(\\mathbf {x} ^ {\\prime}; \\mathbf {x} _ {i}\\right) \\mathbf {c} _ {i} \\quad \\text {a n d} \\quad o _ {i} ^ {\\prime} = G \\left(\\mathbf {x} ^ {\\prime}; \\mathbf {x} _ {i}\\right) o _ {i}. \\tag {3}", + "image_path": "87f33cd4b543eef314bb2135a50017964f8da973fd65e1e354d0070e3f439d79.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 475, + 546, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 475, + 546, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 546, + 641 + ], + "type": "text", + "content": "Thanks to their ability to fit complex geometry and appearance, 3D Gaussians are gaining popularity for 3D scene reconstruction. However, deploying 3D Gaussians for generative tasks remains an unexplored topic and is challenging for several reasons. First, Gaussians are of \"Lagrangian\" nature - their sparse number and learnable positions are challenging to combine with SOTA \"Eulerian\" (i.e., grid-based) CNN generators. Second, the parameters of Gaussians are highly correlated. The same radiance field can be equally well explained by many different configurations of Gaussians, varying their locations, sizes, scales, rotations, colors, and opacities. This ambiguity makes it challenging to generalize over a distribution of objects or scenes, which generative methods do." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "content": "Shell Maps. Our representation is related to Shell Maps [14, 47, 58], a technique in computer graphics designed to model near-surface details. Shell maps use 3D texture maps to store the fine-scale features in a shell-like volume close to a given base surface, typically represented as a triangle mesh. In essence, they extend UV" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9443" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 68, + 541, + 160 + ], + "blocks": [ + { + "bbox": [ + 52, + 68, + 541, + 160 + ], + "lines": [ + { + "bbox": [ + 52, + 68, + 541, + 160 + ], + "spans": [ + { + "bbox": [ + 52, + 68, + 541, + 160 + ], + "type": "image", + "image_path": "96ad5d1d30cf15b47c61c8e54a497541a9817998667630ae96b6559201797931.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 164, + 517, + 176 + ], + "lines": [ + { + "bbox": [ + 74, + 164, + 517, + 176 + ], + "spans": [ + { + "bbox": [ + 74, + 164, + 517, + 176 + ], + "type": "text", + "content": "Figure 3. Novel Views and Animation. We can render a generated identity in novel views and articulate it for animations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 194, + 286, + 319 + ], + "blocks": [ + { + "bbox": [ + 50, + 194, + 286, + 319 + ], + "lines": [ + { + "bbox": [ + 50, + 194, + 286, + 319 + ], + "spans": [ + { + "bbox": [ + 50, + 194, + 286, + 319 + ], + "type": "image", + "image_path": "4ea0818e6b5c2737d2ca2baf51115b51c91de71387ae230c263c36fa628fdfa6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 326, + 288, + 349 + ], + "lines": [ + { + "bbox": [ + 46, + 326, + 288, + 349 + ], + "spans": [ + { + "bbox": [ + 46, + 326, + 288, + 349 + ], + "type": "text", + "content": "Figure 4. Latent Code Interpolation. Latent code interpolation of our model trained on DeepFashion Dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 368, + 287, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 368, + 287, + 511 + ], + "spans": [ + { + "bbox": [ + 46, + 368, + 287, + 511 + ], + "type": "text", + "content": "maps such that every point in the shell volume can be bijectively mapped to the 3D texture map for efficient modeling and rendering in texture space. The shell volume is constructed by offsetting the base mesh along the normal direction while maintaining the topology and avoiding self-intersection. The volume is discretized into tetrahedra that connect the vertices of the base and offset meshes. A unique mapping between the shell space and the texture space can be established by identifying the tetrahedron and subsequently querying the barycentric coordinates inside it. Formally, the mapping from " + }, + { + "bbox": [ + 46, + 368, + 287, + 511 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 368, + 287, + 511 + ], + "type": "text", + "content": " in the shell volume to its position in the 3D texture map is defined as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 117, + 519, + 287, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 519, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 117, + 519, + 287, + 533 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {t} = \\phi (\\mathbf {T} _ {t}, B (\\mathbf {x}, \\mathbf {T})) \\tag {4}", + "image_path": "b8b2af5dc376ac1719f4e2d80e6dd3d8ad0c2940d1f9167f9606d8c2b68bc20e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 540, + 287, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 287, + 576 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 540, + 287, + 576 + ], + "type": "inline_equation", + "content": "\\mathbf{T},\\mathbf{T}_t" + }, + { + "bbox": [ + 46, + 540, + 287, + 576 + ], + "type": "text", + "content": " refer to corresponding tetrahedra in the shell and texture space, " + }, + { + "bbox": [ + 46, + 540, + 287, + 576 + ], + "type": "inline_equation", + "content": "B(\\mathbf{x},\\mathbf{T})" + }, + { + "bbox": [ + 46, + 540, + 287, + 576 + ], + "type": "text", + "content": " is the barycentric coordinates of " + }, + { + "bbox": [ + 46, + 540, + 287, + 576 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 540, + 287, + 576 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 46, + 540, + 287, + 576 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 46, + 540, + 287, + 576 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 540, + 287, + 576 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 540, + 287, + 576 + ], + "type": "text", + "content": " denotes barycentric interpolation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 586, + 208, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 208, + 600 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 208, + 600 + ], + "type": "text", + "content": "4. Gaussian Shell Maps (GSM)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "GSM is a framework that connects 3D Gaussians with SOTA CNN-based generator networks [11] in a GAN setting. The key idea is to anchor the 3D Gaussians at fixed locations on a set of \"shells\" derived from the SMPL [37] human body template mesh. These shells span a volume to model surface details that deviate from the unclothed template mesh. We learn the relevant Gaussian parameters in the texture space, allowing us to leverage established CNN-based generative backbones while seamlessly utilizing the" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 197, + 545, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 197, + 545, + 221 + ], + "spans": [ + { + "bbox": [ + 304, + 197, + 545, + 221 + ], + "type": "text", + "content": "parametric deformation model for articulation. The overall pipeline is shown in Figure 2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 228, + 400, + 241 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 228, + 400, + 241 + ], + "spans": [ + { + "bbox": [ + 305, + 228, + 400, + 241 + ], + "type": "text", + "content": "4.1. Representation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 247, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 247, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 247, + 545, + 437 + ], + "type": "text", + "content": "Our method utilizes the concept of shell maps to leverage the inherent planar structure of texture space for seamless integration with CNN-based generative architectures and, at the same time, encapsulate diverse and complex surface details without directly modifying the template mesh. Specifically, the shell volume is defined by the boundary mesh layers, created by inflating and deflating the T-pose SMPL mesh using Laplacian Smoothing [60] with the smoothing factor set to negative for inflation and positive for deflation. We then represent this shell volume using " + }, + { + "bbox": [ + 304, + 247, + 545, + 437 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 247, + 545, + 437 + ], + "type": "text", + "content": " mesh layers, i.e. \"shells\", by linearly spacing the vertices between the aforementioned boundary shells. In parallel, we apply a similar discretization strategy to the 3D texture space, creating " + }, + { + "bbox": [ + 304, + 247, + 545, + 437 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 247, + 545, + 437 + ], + "type": "text", + "content": " 2D texture maps storing neural features that can be referenced for each shell using the UV mapping inherited from the SMPL template." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 438, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 545, + 510 + ], + "type": "text", + "content": "As shown in Figure 2, we use the shell maps to generate all Gaussian parameters except the positions, as those are sampled and anchored w.r.t. to the shells at every iteration (explained below in detail). This results in a feature volume of " + }, + { + "bbox": [ + 304, + 438, + 545, + 510 + ], + "type": "inline_equation", + "content": "\\mathsf{T}^{N\\times H\\times W\\times 11}" + }, + { + "bbox": [ + 304, + 438, + 545, + 510 + ], + "type": "text", + "content": ", comprised of 3D color " + }, + { + "bbox": [ + 304, + 438, + 545, + 510 + ], + "type": "inline_equation", + "content": "\\mathsf{T}^{\\mathrm{c}}" + }, + { + "bbox": [ + 304, + 438, + 545, + 510 + ], + "type": "text", + "content": ", 1D opacity " + }, + { + "bbox": [ + 304, + 438, + 545, + 510 + ], + "type": "inline_equation", + "content": "\\mathsf{T}^o" + }, + { + "bbox": [ + 304, + 438, + 545, + 510 + ], + "type": "text", + "content": ", 3D scaling " + }, + { + "bbox": [ + 304, + 438, + 545, + 510 + ], + "type": "inline_equation", + "content": "\\mathsf{T}^s" + }, + { + "bbox": [ + 304, + 438, + 545, + 510 + ], + "type": "text", + "content": ", and 4D rotation " + }, + { + "bbox": [ + 304, + 438, + 545, + 510 + ], + "type": "inline_equation", + "content": "\\mathsf{T}^q" + }, + { + "bbox": [ + 304, + 438, + 545, + 510 + ], + "type": "text", + "content": " features." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 510, + 546, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 546, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 546, + 653 + ], + "type": "text", + "content": "We create Gaussians in our shell volume. This is done by sampling a fixed number of Gaussians quasi-uniformly on the shells based on the triangle areas. Once sampled, the Gaussians are anchored on the shells using barycentric coordinates so that every Gaussian center " + }, + { + "bbox": [ + 304, + 510, + 546, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 510, + 546, + 653 + ], + "type": "text", + "content": " can be mapped to a point " + }, + { + "bbox": [ + 304, + 510, + 546, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 304, + 510, + 546, + 653 + ], + "type": "text", + "content": " on the corresponding shell map using Eq. (4), except that the tetrahedra are replaced with the triangle in which the Gaussian resides. This anchoring is a crucial design choice, as it enables straightforward feature retrieval from the shell maps. More importantly, it simplifies the learning by fixing the Gaussian positions and offloading geometry modeling to opacity " + }, + { + "bbox": [ + 304, + 510, + 546, + 653 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 304, + 510, + 546, + 653 + ], + "type": "text", + "content": " and sigma " + }, + { + "bbox": [ + 304, + 510, + 546, + 653 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 304, + 510, + 546, + 653 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": "The spatial span of the Gaussians plays a significant role in reconstructing the features defined on the discrete shells into a continuous signal within the 3D space. It enables every point—whether on the shells, between them, or outside them—to receive a valid opacity and color value by evaluat-" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9444" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 79, + 278, + 133 + ], + "blocks": [ + { + "bbox": [ + 52, + 94, + 62, + 122 + ], + "lines": [ + { + "bbox": [ + 52, + 94, + 62, + 122 + ], + "spans": [ + { + "bbox": [ + 52, + 94, + 62, + 122 + ], + "type": "text", + "content": "GNNARF" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 77, + 79, + 278, + 133 + ], + "lines": [ + { + "bbox": [ + 77, + 79, + 278, + 133 + ], + "spans": [ + { + "bbox": [ + 77, + 79, + 278, + 133 + ], + "type": "image", + "image_path": "9d4ccb1dc66ec68769e236587cba67738311be5f642f9de02a0f41f4e1467f3e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 82, + 140, + 157, + 196 + ], + "blocks": [ + { + "bbox": [ + 54, + 152, + 64, + 178 + ], + "lines": [ + { + "bbox": [ + 54, + 152, + 64, + 178 + ], + "spans": [ + { + "bbox": [ + 54, + 152, + 64, + 178 + ], + "type": "text", + "content": "EVA3D" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 82, + 140, + 157, + 196 + ], + "lines": [ + { + "bbox": [ + 82, + 140, + 157, + 196 + ], + "spans": [ + { + "bbox": [ + 82, + 140, + 157, + 196 + ], + "type": "image", + "image_path": "60a80f93c25fa38b46384e0cab92c9bc0bd3aebc0c784de7de2c61a1989ecd07.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 168, + 140, + 279, + 195 + ], + "blocks": [ + { + "bbox": [ + 168, + 140, + 279, + 195 + ], + "lines": [ + { + "bbox": [ + 168, + 140, + 279, + 195 + ], + "spans": [ + { + "bbox": [ + 168, + 140, + 279, + 195 + ], + "type": "image", + "image_path": "2abfb83f66150462ae1694569f8f057f0d47f061b55f19face5e827336def53b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 83, + 200, + 157, + 258 + ], + "blocks": [ + { + "bbox": [ + 55, + 220, + 64, + 238 + ], + "lines": [ + { + "bbox": [ + 55, + 220, + 64, + 238 + ], + "spans": [ + { + "bbox": [ + 55, + 220, + 64, + 238 + ], + "type": "text", + "content": "s" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 83, + 200, + 157, + 258 + ], + "lines": [ + { + "bbox": [ + 83, + 200, + 157, + 258 + ], + "spans": [ + { + "bbox": [ + 83, + 200, + 157, + 258 + ], + "type": "image", + "image_path": "ac2e726bb7e7dde5d0327859e176d331caf26c3066813aca6df3d2e65bf09cd1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 161, + 262, + 208, + 273 + ], + "lines": [ + { + "bbox": [ + 161, + 262, + 208, + 273 + ], + "spans": [ + { + "bbox": [ + 161, + 262, + 208, + 273 + ], + "type": "text", + "content": "DeepFashion" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 173, + 198, + 279, + 257 + ], + "blocks": [ + { + "bbox": [ + 173, + 198, + 279, + 257 + ], + "lines": [ + { + "bbox": [ + 173, + 198, + 279, + 257 + ], + "spans": [ + { + "bbox": [ + 173, + 198, + 279, + 257 + ], + "type": "image", + "image_path": "0362bac430c50fc9a2689dd7de7fed6c1ad9e13dcf4ccb96cdcc77eb234b55fc.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 322, + 77, + 521, + 273 + ], + "blocks": [ + { + "bbox": [ + 322, + 77, + 521, + 273 + ], + "lines": [ + { + "bbox": [ + 322, + 77, + 521, + 273 + ], + "spans": [ + { + "bbox": [ + 322, + 77, + 521, + 273 + ], + "type": "image", + "image_path": "9f63fa811cb63af8ff2daaa0e6bd8a2caf56aeb0298bb39fa81d3bd8bb2bacd8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 278, + 547, + 313 + ], + "lines": [ + { + "bbox": [ + 46, + 278, + 547, + 313 + ], + "spans": [ + { + "bbox": [ + 46, + 278, + 547, + 313 + ], + "type": "text", + "content": "Figure 5. Qualitative Comparison. We compare our results with GNARF and EVA3D baselines on DeepFashion and SHHQ datasets. In each case, we show the deformed body poses of the identities generated by the methods. The competing methods exhibit artifacts marked in red. Notice that our approach generates high-quality textures, like facial details and more realistic deformations." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 333, + 287, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 333, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 46, + 333, + 287, + 380 + ], + "type": "text", + "content": "ing Eq. (3). This process allows us to model diverse appearances and body shapes, excelling mesh-based representation while at the same time maintaining efficient rendering, which is critical for GAN training." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 381, + 287, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 381, + 287, + 406 + ], + "spans": [ + { + "bbox": [ + 46, + 381, + 287, + 406 + ], + "type": "text", + "content": "Formally, Gaussian opacity, color, scale, and rotation can be interpolated from the shell maps " + }, + { + "bbox": [ + 46, + 381, + 287, + 406 + ], + "type": "inline_equation", + "content": "\\mathsf{T}" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 88, + 415, + 287, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 415, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 88, + 415, + 287, + 430 + ], + "type": "interline_equation", + "content": "f = \\mathrm {T} ^ {f} \\left(\\mathbf {x} _ {t}\\right), \\text {w h e r e} f = \\{o, \\mathbf {c}, \\mathbf {s}, \\mathbf {q} \\}. \\tag {5}", + "image_path": "702a5fb369be73798428d2d2541ad26c2967dd867849a0131c1fe6db9055a997.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 438, + 129, + 450 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 438, + 129, + 450 + ], + "spans": [ + { + "bbox": [ + 47, + 438, + 129, + 450 + ], + "type": "text", + "content": "4.2. Deformation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 457, + 287, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 457, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 46, + 457, + 287, + 588 + ], + "type": "text", + "content": "The deformation step updates the Gaussians' locations and orientations based on the SMPL template mesh " + }, + { + "bbox": [ + 46, + 457, + 287, + 588 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 46, + 457, + 287, + 588 + ], + "type": "text", + "content": ", pose " + }, + { + "bbox": [ + 46, + 457, + 287, + 588 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 457, + 287, + 588 + ], + "type": "text", + "content": ", and shape " + }, + { + "bbox": [ + 46, + 457, + 287, + 588 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 457, + 287, + 588 + ], + "type": "text", + "content": ". Note that, different from SMPL, our template mesh could be any of the shells. Since each Gaussian is anchored on the shells using barycentric coordinates, we can query its new location and orientation simply from the associated vertices. In particular, given the barycentric coordinates " + }, + { + "bbox": [ + 46, + 457, + 287, + 588 + ], + "type": "inline_equation", + "content": "B(\\mathbf{x},\\mathbf{T})" + }, + { + "bbox": [ + 46, + 457, + 287, + 588 + ], + "type": "text", + "content": " of a Gaussian inside triangle " + }, + { + "bbox": [ + 46, + 457, + 287, + 588 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 46, + 457, + 287, + 588 + ], + "type": "text", + "content": ", and the deformed position " + }, + { + "bbox": [ + 46, + 457, + 287, + 588 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathrm{new}}" + }, + { + "bbox": [ + 46, + 457, + 287, + 588 + ], + "type": "text", + "content": " and the rotation quaternions " + }, + { + "bbox": [ + 46, + 457, + 287, + 588 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 46, + 457, + 287, + 588 + ], + "type": "text", + "content": " on the vertices, we can obtain the new location and orientation of the Gaussian as" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 110, + 599, + 287, + 612 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 599, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 110, + 599, + 287, + 612 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {\\text {n e w}} = \\phi (\\mathbf {T} _ {\\text {n e w}}, B (\\mathbf {x}, \\mathbf {T})) \\tag {6}", + "image_path": "34a6490045cdc5e2e73b6ae7f21482ea218218dbf97662fc1af74cee39bded6f.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 78, + 615, + 287, + 640 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 615, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 78, + 615, + 287, + 640 + ], + "type": "interline_equation", + "content": "\\mathbf {q} _ {\\text {n e w}} = \\frac {\\hat {\\mathbf {p}}}{\\| \\hat {\\mathbf {p}} \\|} \\mathbf {q}, \\text {w h e r e} \\hat {\\mathbf {p}} = \\phi (\\mathbf {P}, B (\\mathbf {x}, \\mathbf {T})). \\tag {7}", + "image_path": "53af08dcd3c596847cccea8280f4d1c1532f374f487d37637710dffeb3cdf999.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 647, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 647, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 647, + 287, + 715 + ], + "type": "text", + "content": "The deformed mesh is given by the SMPL deformation model SMPL, and the quaternions " + }, + { + "bbox": [ + 46, + 647, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 46, + 647, + 287, + 715 + ], + "type": "text", + "content": " are a result of the linear blend skinning (LBS) from regressed joints and skinning weights, " + }, + { + "bbox": [ + 46, + 647, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\{w_{j}\\}_{1}^{N_{J}}" + }, + { + "bbox": [ + 46, + 647, + 287, + 715 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 46, + 647, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{new}} = \\mathrm{SMPL}(\\pmb {\\theta},\\pmb {\\beta},\\mathbf{M})" + }, + { + "bbox": [ + 46, + 647, + 287, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 647, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{p} = \\mathrm{rot2quat}\\left(\\sum_{j = 1}^{N_J}w_jR_j(\\pmb {\\theta},J(\\pmb {\\beta}))\\right)" + }, + { + "bbox": [ + 46, + 647, + 287, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 647, + 287, + 715 + ], + "type": "inline_equation", + "content": "R_{j}" + }, + { + "bbox": [ + 46, + 647, + 287, + 715 + ], + "type": "text", + "content": " is the" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 333, + 545, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 333, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 305, + 333, + 545, + 358 + ], + "type": "text", + "content": "rotation matrix of the " + }, + { + "bbox": [ + 305, + 333, + 545, + 358 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 305, + 333, + 545, + 358 + ], + "type": "text", + "content": "-th joint, and " + }, + { + "bbox": [ + 305, + 333, + 545, + 358 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 305, + 333, + 545, + 358 + ], + "type": "text", + "content": " is the regressor function that maps the shape parameters to the joint locations." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 365, + 394, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 365, + 394, + 379 + ], + "spans": [ + { + "bbox": [ + 306, + 365, + 394, + 379 + ], + "type": "text", + "content": "4.3. GAN training" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 384, + 545, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 545, + 504 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 545, + 504 + ], + "type": "text", + "content": "Generator. Similar to prior work [11], we adopt StyleGAN2 without camera and pose conditioning [6, 71] for the generator. We use separate MLPs with different activations for " + }, + { + "bbox": [ + 304, + 384, + 545, + 504 + ], + "type": "inline_equation", + "content": "\\mathbf{c}, o, \\mathbf{q}" + }, + { + "bbox": [ + 304, + 384, + 545, + 504 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 384, + 545, + 504 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 304, + 384, + 545, + 504 + ], + "type": "text", + "content": ": for " + }, + { + "bbox": [ + 304, + 384, + 545, + 504 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 304, + 384, + 545, + 504 + ], + "type": "text", + "content": ", we use shifted sigmoid following the practice proposed in Mip-NeRF [5]; for " + }, + { + "bbox": [ + 304, + 384, + 545, + 504 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 304, + 384, + 545, + 504 + ], + "type": "text", + "content": " we use sigmoid to constrain the range to " + }, + { + "bbox": [ + 304, + 384, + 545, + 504 + ], + "type": "inline_equation", + "content": "(0, 1)" + }, + { + "bbox": [ + 304, + 384, + 545, + 504 + ], + "type": "text", + "content": "; for " + }, + { + "bbox": [ + 304, + 384, + 545, + 504 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 304, + 384, + 545, + 504 + ], + "type": "text", + "content": " we normalize the raw MLP output to ensure they are quaternions (see Eq. (7)); for " + }, + { + "bbox": [ + 304, + 384, + 545, + 504 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 304, + 384, + 545, + 504 + ], + "type": "text", + "content": " we use clamped exponential activation to limit the size of the Gaussians, which is critical for convergence based on our empirical study." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 507, + 546, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 507, + 546, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 507, + 546, + 639 + ], + "type": "text", + "content": "Discriminator. Our discriminator closely follows that of [6]. Since it does not have an upsampler, the input to the discriminator is the RGB image concatenated with the alpha channel (foreground mask), which is rendered using Gaussian rasterization with the Gaussian color set to 1 for fake samples and precomputed using off-the-shelf segmentation network [55] for the real samples. We refer to the discriminator using foreground mask \"Mask Discriminator\". Including the alpha channel helps prevent the white background from bleeding into the appearance, which causes artifacts during articulation." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "Face, Hand, and Feet Discriminator. As the human body and clothes are diverse, the discriminator may choose to focus on these features and provide a weak signal to the facial, hand, and foot areas, which are crucial for visually appealing results. To avoid this problem, we adopt a dedicated Face Discriminator " + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{face}}" + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": ", Feet-and-Hands Discrimi" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9445" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 194, + 293, + 344 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 288, + 192 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 288, + 192 + ], + "type": "text", + "content": "Table 1. Quantitative Evaluation. We compare our method with 3D-GAN baselines using DeepFashion and SHHQ datasets. We compute the FID score to evaluate the quality and diversity of the generated samples. Notice that our scores are comparable to state-of-the-art methods. To evaluate deformation consistency, we compute the PCK metric, where our approach consistently outperforms the baselines. INF represents the inference speed measured in ms/img on an A6000 GPU at " + }, + { + "bbox": [ + 46, + 70, + 288, + 192 + ], + "type": "inline_equation", + "content": "512^2" + }, + { + "bbox": [ + 46, + 70, + 288, + 192 + ], + "type": "text", + "content": " resolution. Note our method is the fastest across all competing methods. * numbers are adopted from EVA3D [23]; NA represents Not Available; — represents Not Applicable, and + numbers are provided by the authors." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 194, + 293, + 344 + ], + "lines": [ + { + "bbox": [ + 47, + 194, + 293, + 344 + ], + "spans": [ + { + "bbox": [ + 47, + 194, + 293, + 344 + ], + "type": "table", + "html": "
ModelDeep FashionSHHQComp.
FID ↓PCK ↑FID ↓PCK ↑INF. ↓
EG3D*26.38—32.96—38
StyleSDF*92.40—14.12—32
ENARF*77.0343.7480.5440.17104
GNARF33.8597.8314.8498.9672
EVA3D*15.9187.5011.9988.95200
StylePeople17.7298.3114.6798.5828
GetAvatar19.00+NANANA44
AG3D10.93NANANA105
Ours15.7899.4813.3099.2728
", + "image_path": "7dfe76cbdc86d1d3c6bf1fe9686dbd5d16c9be12c208c4214ccfbbffb8df06d7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 357, + 287, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 357, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 287, + 441 + ], + "type": "text", + "content": "nator " + }, + { + "bbox": [ + 46, + 357, + 287, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{feet/hands}}" + }, + { + "bbox": [ + 46, + 357, + 287, + 441 + ], + "type": "text", + "content": " in addition to the main Discriminator " + }, + { + "bbox": [ + 46, + 357, + 287, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{main}}" + }, + { + "bbox": [ + 46, + 357, + 287, + 441 + ], + "type": "text", + "content": ". All of these part-focused discriminators share the same base architecture as the main discriminator, except that we no longer use any conditioning since these cropped images do not contain distinct pose information. The inputs are crops of the corresponding parts, whose spatial span is determined from the SMPL pose and camera parameters." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 445, + 287, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 287, + 529 + ], + "type": "text", + "content": "Scaling Regularization. In our empirical study, we observe when unconstrained, the network tends to learn overly large or extremely small Gaussians early on during the training and rapidly leads to divergence or model collapse. We experimented with multiple regularization strategies and different activation functions and found the following scaling regularization most effective:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 539, + 287, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 539, + 287, + 571 + ], + "spans": [ + { + "bbox": [ + 54, + 539, + 287, + 571 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s c a l e}} = \\mathrm {M} \\circ \\| \\mathrm {T} ^ {\\mathrm {s}} - s _ {\\text {r e f}} \\| ^ {2} \\text {w i t h} s _ {\\text {r e f}} = \\frac {1}{P} \\sum_ {i = 1} ^ {P} \\ln \\left(\\delta_ {i}\\right), \\tag {8}", + "image_path": "78a55833d99e4528ee5cdd337829b8c9eaec800420fc0fc40ca21184e41ae7d8.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 579, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 579, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 579, + 287, + 628 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 579, + 287, + 628 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 579, + 287, + 628 + ], + "type": "text", + "content": " is the binary mask indicating UV-mapped region on the shell maps, and " + }, + { + "bbox": [ + 46, + 579, + 287, + 628 + ], + "type": "inline_equation", + "content": "s_{\\mathrm{ref}}" + }, + { + "bbox": [ + 46, + 579, + 287, + 628 + ], + "type": "text", + "content": " is a reference scale determined by the closest-neighbor distance " + }, + { + "bbox": [ + 46, + 579, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 46, + 579, + 287, + 628 + ], + "type": "text", + "content": " averaged among all the Gaussians." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 639, + 167, + 653 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 639, + 167, + 653 + ], + "spans": [ + { + "bbox": [ + 47, + 639, + 167, + 653 + ], + "type": "text", + "content": "5. Experiment Settings" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 658, + 111, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 111, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 111, + 670 + ], + "type": "text", + "content": "5.1. Datasets" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": "We evaluate our method using the two most common human datasets, DeepFashion [36] and SHHQ [17]. DeepFashion and SHHQ do not provide SMPL parameters, so we use" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 312, + 94, + 541, + 132 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 547, + 93 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 547, + 93 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 547, + 93 + ], + "type": "text", + "content": "Table 2. Ablation: Number of Shells. Ablation on the number of shells performed on " + }, + { + "bbox": [ + 305, + 70, + 547, + 93 + ], + "type": "inline_equation", + "content": "128^{2}" + }, + { + "bbox": [ + 305, + 70, + 547, + 93 + ], + "type": "text", + "content": " resolution trained for " + }, + { + "bbox": [ + 305, + 70, + 547, + 93 + ], + "type": "inline_equation", + "content": "1600k" + }, + { + "bbox": [ + 305, + 70, + 547, + 93 + ], + "type": "text", + "content": " images." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 312, + 94, + 541, + 132 + ], + "lines": [ + { + "bbox": [ + 312, + 94, + 541, + 132 + ], + "spans": [ + { + "bbox": [ + 312, + 94, + 541, + 132 + ], + "type": "table", + "html": "
# of Shells1281015
FID ↓40.325.3118.7021.5721.10
", + "image_path": "8817a128926a103fe0b05dbf07aaf837b325a5b9d84b5e1b2dd4c62dd0900132.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 158, + 545, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 158, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 305, + 158, + 545, + 182 + ], + "type": "text", + "content": "SMPLify-X [43] to obtain the SMPL parameters and camera poses." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 199, + 405, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 199, + 405, + 213 + ], + "spans": [ + { + "bbox": [ + 306, + 199, + 405, + 213 + ], + "type": "text", + "content": "5.2. Training Details" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 221, + 547, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 221, + 547, + 414 + ], + "spans": [ + { + "bbox": [ + 304, + 221, + 547, + 414 + ], + "type": "text", + "content": "By default, we generate the shell maps at " + }, + { + "bbox": [ + 304, + 221, + 547, + 414 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 221, + 547, + 414 + ], + "type": "text", + "content": " resolution for 8 shells with a total of " + }, + { + "bbox": [ + 304, + 221, + 547, + 414 + ], + "type": "inline_equation", + "content": "100k" + }, + { + "bbox": [ + 304, + 221, + 547, + 414 + ], + "type": "text", + "content": " Gaussians equally distributed across the shells. The offset between adjacent shells is 0.08. The last fully-connected layer of the scaling prediction is adjusted to ensure the initial scaling is within a reasonable range, which is determined similar to " + }, + { + "bbox": [ + 304, + 221, + 547, + 414 + ], + "type": "inline_equation", + "content": "s_{\\mathrm{ref}}" + }, + { + "bbox": [ + 304, + 221, + 547, + 414 + ], + "type": "text", + "content": " in Eq. (8). The loss weights for all discriminators are set to 1, whereas the scaling regularization is set to 0.1. " + }, + { + "bbox": [ + 304, + 221, + 547, + 414 + ], + "type": "inline_equation", + "content": "R1" + }, + { + "bbox": [ + 304, + 221, + 547, + 414 + ], + "type": "text", + "content": " regularization is used for all the discriminators with a weight of 10. We apply progressive training starting at " + }, + { + "bbox": [ + 304, + 221, + 547, + 414 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 221, + 547, + 414 + ], + "type": "text", + "content": " rendering resolution for " + }, + { + "bbox": [ + 304, + 221, + 547, + 414 + ], + "type": "inline_equation", + "content": "6000k" + }, + { + "bbox": [ + 304, + 221, + 547, + 414 + ], + "type": "text", + "content": " training images and progressively grow the resolution to " + }, + { + "bbox": [ + 304, + 221, + 547, + 414 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 221, + 547, + 414 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 304, + 221, + 547, + 414 + ], + "type": "inline_equation", + "content": "1000k" + }, + { + "bbox": [ + 304, + 221, + 547, + 414 + ], + "type": "text", + "content": " images, then continue training at the fixed " + }, + { + "bbox": [ + 304, + 221, + 547, + 414 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 221, + 547, + 414 + ], + "type": "text", + "content": " resolution for " + }, + { + "bbox": [ + 304, + 221, + 547, + 414 + ], + "type": "inline_equation", + "content": "3000k" + }, + { + "bbox": [ + 304, + 221, + 547, + 414 + ], + "type": "text", + "content": " images. We train our method on 8 A100 GPUs with a batch size of 32. The learning rates are initialized at 0.002 for both the generator and discriminator." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 430, + 415, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 430, + 415, + 442 + ], + "spans": [ + { + "bbox": [ + 306, + 430, + 415, + 442 + ], + "type": "text", + "content": "5.3. Evaluation Details" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 453, + 545, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 453, + 545, + 573 + ], + "spans": [ + { + "bbox": [ + 304, + 453, + 545, + 573 + ], + "type": "text", + "content": "Baselines. We compare with the following volume rendering and rasterization-based methods: EG3D [11], StyleSDF [42], GNARF [6], ENARF [41], StylePeople [19], EVA3D [23], GetAvatar [76], and AG3D [15]. Among these methods, EG3D and StyleSDF do not support articulation; except StylePeople, all methods use neural field and volumetric rendering. Regardless of representation, all these baselines use 2D convolution for postrendering upsampling and/or postprocessing, causing flickering artifacts (see Supplementary Materials)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": "Metrics Following prior work, We use two metrics to evaluate the quality, diversity, and quality of animations: Fréchet Inception Distance (FID) and Percentage of Correct Keypoint (PCK). The former measures the visual quality and diversity of the generated samples (using " + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "inline_equation", + "content": "50k" + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": " samples and the full dataset). The latter measures how close the generated image aligns with the pose control. We compute PCK on " + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "inline_equation", + "content": "5k" + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": " samples using the implementation provided by GNARF [6]." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9446" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 74, + 289, + 258 + ], + "blocks": [ + { + "bbox": [ + 51, + 74, + 289, + 258 + ], + "lines": [ + { + "bbox": [ + 51, + 74, + 289, + 258 + ], + "spans": [ + { + "bbox": [ + 51, + 74, + 289, + 258 + ], + "type": "image", + "image_path": "5061a58a11f903bee1677c837befcd5079d65d02ff72910bbba64c2fee91c068.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 258, + 287, + 314 + ], + "lines": [ + { + "bbox": [ + 46, + 258, + 287, + 314 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 287, + 314 + ], + "type": "text", + "content": "Figure 6. Multi-view Consistency. The figure shows a closeup of the garment in two different views. The green dashed line follows the body shape and indicates the same 3D position. In AG3D [15], the pattern significantly changes, while ours, utilizing texture maps, has built-in view consistency." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 335, + 163, + 347 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 335, + 163, + 347 + ], + "spans": [ + { + "bbox": [ + 47, + 335, + 163, + 347 + ], + "type": "text", + "content": "6. Experiment Results" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 354, + 122, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 354, + 122, + 367 + ], + "spans": [ + { + "bbox": [ + 47, + 354, + 122, + 367 + ], + "type": "text", + "content": "6.1. Qualitative" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 373, + 287, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 287, + 468 + ], + "type": "text", + "content": "Figure 1 showcases some of our generated results. Notice that our model is able to generate diverse attributes, including loose clothing and accessories like hats. Further, in Figure 3, we visualize the generated results in varying camera views and body poses. To demonstrate that the latent space of the proposed model is smooth, we show visual results of latent code interpolation in Figure 4 computed on the DeepFashion trained model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 469, + 287, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 469, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 469, + 287, + 613 + ], + "type": "text", + "content": "Figure 5 shows the visual comparison of our method with the baselines on DeepFashion and SHHQ datasets, with more examples to be found in Supplementary Materials. Volumetric methods, such as GNARF, operate in a canonical pose and rely on accurate correspondence matching between the observed space and the canonical space, which is ill-defined; thus, these methods tend to produce artifacts for limbs, which undergo large deformation and occlusion. Our method is able to model the complex geometry and appearance of the human body, generating characters with loose clothing and accessories (see Figure 1) and producing convincing results under various articulations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 622, + 129, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 622, + 129, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 622, + 129, + 634 + ], + "type": "text", + "content": "6.2.Quantitative" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 641, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 689 + ], + "type": "text", + "content": "In Table 1, we show the results of our method compared to the competing methods. Many of the methods have not released training or data processing code; we have collected these numbers with our best effort from the authors directly." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Our method performs on par with the SOTA methods in terms of visual quality and diversity measured by FID" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 313, + 72, + 536, + 269 + ], + "blocks": [ + { + "bbox": [ + 313, + 72, + 536, + 269 + ], + "lines": [ + { + "bbox": [ + 313, + 72, + 536, + 269 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 536, + 269 + ], + "type": "image", + "image_path": "1b9fef6c9aecdd64f01c8197f09cc9ae1df82de6a0858c0233bf5db73146f6d5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 271, + 545, + 316 + ], + "lines": [ + { + "bbox": [ + 305, + 271, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 305, + 271, + 545, + 316 + ], + "type": "text", + "content": "Figure 7. Per-shell Visualization. We visualize the per-shell contributions. The shells decompose the human body into different parts, where the inner shells capture the torso, and the outer shells capture the clothing and hair." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 337, + 545, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 337, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 304, + 337, + 545, + 409 + ], + "type": "text", + "content": "while ranking top in terms of inference speed. By utilizing the texture space, our model can control the articulation in a straightforward manner, contributing to consistent alignment with the pose control input, as reflected by the PCK score. Note that GetAvatar is trained on multiview data, and AG3D uses normal maps for supervision." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 409, + 546, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 409, + 546, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 409, + 546, + 517 + ], + "type": "text", + "content": "All other baseline methods adopt post-rendering convolution layers to upsample and add texture details, leading to severe aliasing artifacts not reflected in the quantitative evaluation. We demonstrate this in Figure 6, where we render a generated identity in two different views using our method and AG3D, which achieves the lowest FID on the DeepFashion dataset. Notice that with view change, the features produced by AG3D change significantly, leading to unnatural flickering in animation even though the FID is low." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 525, + 370, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 525, + 370, + 536 + ], + "spans": [ + { + "bbox": [ + 306, + 525, + 370, + 536 + ], + "type": "text", + "content": "6.3. Ablation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 543, + 545, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 543, + 545, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 543, + 545, + 615 + ], + "type": "text", + "content": "We conduct ablation studies on different configurations and components of our method to understand their effect on the final model's performance. To identify the trends, we test these configurations at " + }, + { + "bbox": [ + 304, + 543, + 545, + 615 + ], + "type": "inline_equation", + "content": "128^{2}" + }, + { + "bbox": [ + 304, + 543, + 545, + 615 + ], + "type": "text", + "content": " resolution on the SHHQ dataset, trained for " + }, + { + "bbox": [ + 304, + 543, + 545, + 615 + ], + "type": "inline_equation", + "content": "1600k" + }, + { + "bbox": [ + 304, + 543, + 545, + 615 + ], + "type": "text", + "content": " images without progressive training. " + }, + { + "bbox": [ + 304, + 543, + 545, + 615 + ], + "type": "inline_equation", + "content": "200k" + }, + { + "bbox": [ + 304, + 543, + 545, + 615 + ], + "type": "text", + "content": " Gaussians are sampled for all the experiments." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 629, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 546, + 713 + ], + "type": "text", + "content": "Number of Shells. The number of shells determines the granularity of variation we can model inside the shell volume, thus impacting the expressivity of our representation. We investigate this effect in Table 2. With only one shell, although the Gaussians can vary in size and cover regions off the shell surface, they still lack the capacity to model diverse human body shapes and appearances. Using two shells es" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9447" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 75, + 113, + 215 + ], + "blocks": [ + { + "bbox": [ + 51, + 75, + 113, + 215 + ], + "lines": [ + { + "bbox": [ + 51, + 75, + 113, + 215 + ], + "spans": [ + { + "bbox": [ + 51, + 75, + 113, + 215 + ], + "type": "image", + "image_path": "a2d90ac876d564264aef258571024d9c30e2e5ca71ec75dfd587305158ddf2a7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 58, + 216, + 160, + 224 + ], + "lines": [ + { + "bbox": [ + 58, + 216, + 160, + 224 + ], + "spans": [ + { + "bbox": [ + 58, + 216, + 160, + 224 + ], + "type": "text", + "content": "w/o Mask and Part Discriminators" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 118, + 76, + 162, + 118 + ], + "blocks": [ + { + "bbox": [ + 118, + 76, + 162, + 118 + ], + "lines": [ + { + "bbox": [ + 118, + 76, + 162, + 118 + ], + "spans": [ + { + "bbox": [ + 118, + 76, + 162, + 118 + ], + "type": "image", + "image_path": "5f52023679c00c6c1b75fdaabbb3e5d45548ec28397e33ed70575d213272f151.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 123, + 121, + 156, + 146 + ], + "blocks": [ + { + "bbox": [ + 123, + 121, + 156, + 146 + ], + "lines": [ + { + "bbox": [ + 123, + 121, + 156, + 146 + ], + "spans": [ + { + "bbox": [ + 123, + 121, + 156, + 146 + ], + "type": "image", + "image_path": "8996ccb612221ede0468c0437a66707607969c6c94a29d79ca92ab2a135ecccf.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 124, + 148, + 156, + 185 + ], + "blocks": [ + { + "bbox": [ + 124, + 148, + 156, + 185 + ], + "lines": [ + { + "bbox": [ + 124, + 148, + 156, + 185 + ], + "spans": [ + { + "bbox": [ + 124, + 148, + 156, + 185 + ], + "type": "image", + "image_path": "8315f15269a8e6b8ba9f361a68925ff7b4921c7f808c9e732a83baa547dbeee0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 115, + 186, + 166, + 213 + ], + "blocks": [ + { + "bbox": [ + 115, + 186, + 166, + 213 + ], + "lines": [ + { + "bbox": [ + 115, + 186, + 166, + 213 + ], + "spans": [ + { + "bbox": [ + 115, + 186, + 166, + 213 + ], + "type": "image", + "image_path": "11a8e5b5a6db3dba21b7956cb9ab3b401cbfd52381df0056c5d1d25c74ec94a5.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 174, + 76, + 217, + 117 + ], + "blocks": [ + { + "bbox": [ + 174, + 76, + 217, + 117 + ], + "lines": [ + { + "bbox": [ + 174, + 76, + 217, + 117 + ], + "spans": [ + { + "bbox": [ + 174, + 76, + 217, + 117 + ], + "type": "image", + "image_path": "53baec545b2efe830faa9332178e31fff5ba1e3a7b3fd43315ab2dbffc91c83e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 181, + 121, + 211, + 145 + ], + "blocks": [ + { + "bbox": [ + 181, + 121, + 211, + 145 + ], + "lines": [ + { + "bbox": [ + 181, + 121, + 211, + 145 + ], + "spans": [ + { + "bbox": [ + 181, + 121, + 211, + 145 + ], + "type": "image", + "image_path": "9432951984b7d118844e22980b54eb950da36c6eeba6f8439f61f833db16351f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 180, + 148, + 211, + 185 + ], + "blocks": [ + { + "bbox": [ + 180, + 148, + 211, + 185 + ], + "lines": [ + { + "bbox": [ + 180, + 148, + 211, + 185 + ], + "spans": [ + { + "bbox": [ + 180, + 148, + 211, + 185 + ], + "type": "image", + "image_path": "b1ba2b9fb7f5563cd7191c0e1af4e7fe8a5fda9939ed3073f7365caf57d1017a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 171, + 186, + 221, + 213 + ], + "blocks": [ + { + "bbox": [ + 171, + 186, + 221, + 213 + ], + "lines": [ + { + "bbox": [ + 171, + 186, + 221, + 213 + ], + "spans": [ + { + "bbox": [ + 171, + 186, + 221, + 213 + ], + "type": "image", + "image_path": "c60e8edd65da92bb61cd635ccf89f5d793b5eb36ddb6ceb64f642744034c0f08.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 223, + 76, + 285, + 215 + ], + "blocks": [ + { + "bbox": [ + 176, + 216, + 274, + 224 + ], + "lines": [ + { + "bbox": [ + 176, + 216, + 274, + 224 + ], + "spans": [ + { + "bbox": [ + 176, + 216, + 274, + 224 + ], + "type": "text", + "content": "w/ Mask and Part Discriminators" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 223, + 76, + 285, + 215 + ], + "lines": [ + { + "bbox": [ + 223, + 76, + 285, + 215 + ], + "spans": [ + { + "bbox": [ + 223, + 76, + 285, + 215 + ], + "type": "image", + "image_path": "4cb8897cca021a8a05c2bde68570c87fcaf91ffb6f259d296917695fa1ed378a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 228, + 287, + 272 + ], + "lines": [ + { + "bbox": [ + 47, + 228, + 287, + 272 + ], + "spans": [ + { + "bbox": [ + 47, + 228, + 287, + 272 + ], + "type": "text", + "content": "Figure 8. Mask + Part Discriminators. The blue box highlights the improved silhouette after concatenating the alpha channels in the input to the discriminator. The other boxes showcase better details using the face, feet, and hands discriminator." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 294, + 287, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 294, + 287, + 450 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 287, + 450 + ], + "type": "text", + "content": "mentally defines the boundary for the shell volume. This variation significantly outperforms the single-shell variation, demonstrating the importance of shell volume. Adding 8 shells (our default setting) further improves the visual quality, as the finer discretization enhances the capability to model more complex geometry and appearance. The performance gain stagnates with even more shells, likely because the model struggles to converge when the feature map becomes too large. Figure 7 shows an example identity of what each Gaussian shell has learned: progressing from the innermost to the outermost shell, every shell captures some texture details, and the outer shell models details that deviate from the base template, such as hat and hair." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 468, + 286, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 286, + 611 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 286, + 611 + ], + "type": "text", + "content": "Mask and Part Discriminator The mask discriminator involves the inclusion of the alpha channel as part of the input to the RGB images and is very beneficial for disambiguating background by learning plausible human silhouettes. As demonstrated in Figure 8, the model trained without mask discriminator generates disturbing white artifacts outside the human shape. Figure 8 also shows the improvement in small features in the facial area and better structure for hands and feet introduced by facial and feet/hands discriminator. We also quantify the joint improvement of the mask and part discriminator and observed an FID boost from 20.63 to 18.7 at " + }, + { + "bbox": [ + 46, + 468, + 286, + 611 + ], + "type": "inline_equation", + "content": "128^{2}" + }, + { + "bbox": [ + 46, + 468, + 286, + 611 + ], + "type": "text", + "content": " resolution." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "content": "Types of Gaussian Anchors. In our representation, we opt to sample the Gaussians on shell meshes. There are other alternatives to sample the Gaussians, including constructing tetrahedra akin to the original Shell Map proposed by Porumbescu et al. [47]; or completely discarding the template by sampling uniformly inside a bounding box. We compared with these alternatives and found the shell repre" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "content": "sensation achieves the best visual quality and convergence speed. A more detailed discussion can be found in Supplementary Materials." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 119, + 375, + 131 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 119, + 375, + 131 + ], + "spans": [ + { + "bbox": [ + 306, + 119, + 375, + 131 + ], + "type": "text", + "content": "7. Discussion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 136, + 545, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 136, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 304, + 136, + 545, + 316 + ], + "type": "text", + "content": "Limitations. Our method is not without limitations. Like almost all existing 3D human generation approaches, our method also relies on a parametric deformation model to enable articulation, which falls short in handling the dynamics of hair and loose clothing. Due to the irregularity and sparsity of Gaussians, it is not obvious how to extract the accurate geometry and normals. Surface splatting, with a stronger surface prior may be a potential solution, which we plan to investigate in the future. Lastly, while ranking among the best methods, our current results still cannot achieve photorealism. The challenge stems from the complexity of human appearance. A promising direction is to combine a small amount of multi-view studio captures with in-the-wild datasets to leverage priors discovered in controlled and richly annotated datasets." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 327, + 545, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 545, + 434 + ], + "type": "text", + "content": "Ethical Concerns. GANs, like the one we developed, carry the risk of being utilized for creating altered images of real individuals. This inappropriate application of image generation methods represents a danger to society, and we firmly oppose the use of our research for disseminating false information or damaging reputations. Additionally, we acknowledge that there might be a deficit in diversity in our outcomes, which could be a consequence of inherent biases in the datasets we utilize." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 447, + 545, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 447, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 304, + 447, + 545, + 591 + ], + "type": "text", + "content": "Conclusions. In this work, we present Gaussian Shell Maps (GSMs), a novel framework that effectively combines CNN-based generators with 3D Gaussian rendering primitives for the generation of digital humans. Our experiments demonstrate the potential of GSMs in generating diverse and detailed human figures, including complex features like clothing and accessories. The framework shows promise in enhancing the efficiency of rendering processes, a crucial factor for real-time applications in industries like virtual reality and cinematic production. The potential of GSMs in practical scenarios invites further exploration, and continued research could enhance digital human modeling." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgements. We thank the authors of 3D Gaussian Splitting, Bernhard Kerbl, Georgios Kopanas, and George Drettakis for the discussions on the code. We thank Thabo Beeler and Guandao Yang for fruitful discussions and Alexander Bergman for helping with the baseline comparisons. We also thank the authors of GetAvartar for providing additional evaluation results. This work was in part supported by Google, Samsung, Stanford HAI, and the Swiss Postdoc Mobility Fund." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9448" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "text", + "content": "[1] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In ECCV, 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 286, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 286, + 158 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 286, + 158 + ], + "type": "text", + "content": "[2] Sizhe An, Hongyi Xu, Yichun Shi, Guoxian Song, Umit Y Ogras, and Linjie Luo. Panohead: Geometry-aware 3d full-head synthesis in 360deg. In CVPR, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 159, + 286, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 159, + 286, + 202 + ], + "spans": [ + { + "bbox": [ + 54, + 159, + 286, + 202 + ], + "type": "text", + "content": "[3] Shivangi Aneja, Justus Thies, Angela Dai, and Matthias Nießner. ClipFace: Text-guided Editing of Textured 3D Morphable Models. In ArXiv preprint arXiv:2212.01406, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 205, + 286, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 205, + 286, + 269 + ], + "spans": [ + { + "bbox": [ + 54, + 205, + 286, + 269 + ], + "type": "text", + "content": "[4] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. ArXiv, abs/2211.01324, 2022.2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 272, + 286, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 272, + 286, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 272, + 286, + 316 + ], + "type": "text", + "content": "[5] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, 2021. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 318, + 286, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 318, + 286, + 350 + ], + "spans": [ + { + "bbox": [ + 53, + 318, + 286, + 350 + ], + "type": "text", + "content": "[6] Alexander Bergman, Petr Kellnhofer, Wang Yifan, Eric Chan, David Lindell, and Gordon Wetzstein. Generative neural articulated radiance fields. NeurIPS, 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 352, + 286, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 352, + 286, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 352, + 286, + 384 + ], + "type": "text", + "content": "[7] Alexander W. Bergman, Wang Yifan, and Gordon Wetzstein. Articulated 3d head avatar generation using text-to-image diffusion models. 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 387, + 286, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 387, + 286, + 418 + ], + "spans": [ + { + "bbox": [ + 53, + 387, + 286, + 418 + ], + "type": "text", + "content": "[8] Yukang Cao, Yan-Pei Cao, Kai Han, Ying Shan, and KwanYee K. Wong. Dreamavatar: Text-and-shape guided 3d human avatar generation via diffusion models, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 421, + 286, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 421, + 286, + 475 + ], + "spans": [ + { + "bbox": [ + 53, + 421, + 286, + 475 + ], + "type": "text", + "content": "[9] Eric Chan, Koki Nagano, Matthew A. Chan, Alexander W. Bergman, Jeong Joon Park, Axel Levy, Miika Aittala, Shalini De Mello, Tero Karras, and Gordon Wetzstein. Generative novel view synthesis with 3d-aware diffusion models. ArXiv, abs/2304.02602, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 477, + 286, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 477, + 286, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 477, + 286, + 520 + ], + "type": "text", + "content": "[10] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In CVPR, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 522, + 286, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 286, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 286, + 586 + ], + "type": "text", + "content": "[11] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In CVPR, 2022. 1, 2, 4, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 590, + 286, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 286, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 286, + 632 + ], + "type": "text", + "content": "[12] Rui Chen, Y. Chen, Ningxin Jiao, and Kui Jia. *Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation.* ArXiv, abs/2303.13873, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 286, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 286, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 286, + 667 + ], + "type": "text", + "content": "[13] Xinya Chen, Jiaxin Huang, Yanrui Bin, Lu Yu, and Yiyi Liao. Veri3d: Generative vertex-based radiance fields for 3d controllable human image synthesis, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 286, + 713 + ], + "type": "text", + "content": "[14] Stavros Diolatzis, Jan Novak, Fabrice Rousselle, Jonathan Granskog, Miika Aittala, Ravi Ramamoorthi, and George Drettakis. Mesogan: Generative neural reflectance shells. Comput. Graph. Forum, 2023. 2, 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[15] Zijian Dong, Xu Chen, Jinlong Yang, Michael J.Black, Otmar Hilliges, and Andreas Geiger. AG3D: Learning to generate 3D avatars from 2D image collections. In ICCV, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 119, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 119, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 119, + 545, + 162 + ], + "type": "text", + "content": "[16] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Chen Qian, Chen Change Loy, Wayne Wu, and Ziwei Liu. Stylegan-human: A data-centric odyssey of human generation. In ECCV, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 164, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 164, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 308, + 164, + 545, + 207 + ], + "type": "text", + "content": "[17] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Chen Qian, Chen Change Loy, Wayne Wu, and Ziwei Liu. Stylegan-human: A data-centric odyssey of human generation. In ECCV, 2022. 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 209, + 545, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 253 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 253 + ], + "type": "text", + "content": "[18] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. NeurIPS, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 255, + 545, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 255, + 545, + 298 + ], + "spans": [ + { + "bbox": [ + 308, + 255, + 545, + 298 + ], + "type": "text", + "content": "[19] Artur Grigorev, Karim Iskakov, Anastasia Iinina, Renat Bashirov, Ilya Zakharkin, Alexander Vakhitov, and Victor Lempitsky. Stylepeople: A generative model of fullbody human avatars. In CVPR, 2021. 1, 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 300, + 545, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 300, + 545, + 321 + ], + "spans": [ + { + "bbox": [ + 308, + 300, + 545, + 321 + ], + "type": "text", + "content": "[20] Markus Gross and Hanspeter Pfister. Point-based graphics. Elsevier, 2011. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 324, + 545, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 324, + 545, + 355 + ], + "spans": [ + { + "bbox": [ + 308, + 324, + 545, + 355 + ], + "type": "text", + "content": "[21] Jeffrey P Grossman and William J Dally. Point sample rendering. In Proceedings of the Eurographics Workshop, 1998. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 358, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 358, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 358, + 545, + 392 + ], + "type": "text", + "content": "[22] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d aware generator for high-resolution image synthesis. In ICLR, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 394, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 394, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 308, + 394, + 545, + 425 + ], + "type": "text", + "content": "[23] Fangzhou Hong, Zhaoxi Chen, Yushi LAN, Liang Pan, and Ziwei Liu. EVA3d: Compositional 3d human generation from 2d image collections. In ICLR, 2023. 1, 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 428, + 545, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 428, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 308, + 428, + 545, + 460 + ], + "type": "text", + "content": "[24] Suyi Jiang, Haoran Jiang, Ziyu Wang, Haimin Luo, Wenzheng Chen, and Lan Xu. Humangen: Generating human radiance fields with explicit priors. In CVPR, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 462, + 545, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 462, + 545, + 494 + ], + "spans": [ + { + "bbox": [ + 308, + 462, + 545, + 494 + ], + "type": "text", + "content": "[25] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 497, + 545, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 497, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 308, + 497, + 545, + 529 + ], + "type": "text", + "content": "[26] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In Proc. CVPR, 2020. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 532, + 545, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 532, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 308, + 532, + 545, + 563 + ], + "type": "text", + "content": "[27] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 567, + 545, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 567, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 308, + 567, + 545, + 598 + ], + "type": "text", + "content": "[28] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM TOG, 2023. 2, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 601, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 601, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 308, + 601, + 545, + 633 + ], + "type": "text", + "content": "[29] Leonid Keselman and Martial Hebert. Approximate differentiable rendering with algebraic surfaces. In European Conference on Computer Vision (ECCV), 2022. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 635, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 678 + ], + "type": "text", + "content": "[30] Nikos Kolotouros, Thiemo Alldieck, Andrei Zanfir, Eduard Gabriel Bazavan, Mihai Fieraru, and Cristian Sminchisescu. Dreamhuman: Animatable 3d avatars from text. ArXiv, abs/2306.09329, 2023. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "text", + "content": "[31] Georgios Kopanas, Julien Philip, Thomas Leimkuhler, and George Drettakis. Point-based neural rendering with perview optimization. In Comput. Graph. Forum, 2021. 3" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9449" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[32] Samuli Laine and Tero Karras. High-performance software rasterization on gpus. In Proceedings of the ACM SIGGRAPH Symposium on High Performance Graphics, 2011. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "type": "text", + "content": "[33] Ruilong Li, Shan Yang, David A. Ross, and Angjoo Kanazawa. Learn to dance with aist++: Music conditioned 3d dance generation, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 195 + ], + "type": "text", + "content": "[34] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. ArXiv, abs/2211.10440, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 197, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 287, + 239 + ], + "type": "text", + "content": "[35] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. ArXiv, abs/2303.11328, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 241, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 287, + 274 + ], + "type": "text", + "content": "[36] Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In CVPR, 2016. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 276, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 287, + 308 + ], + "type": "text", + "content": "[37] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. ACM TOG, 2015. 1, 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 310, + 287, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 310, + 287, + 363 + ], + "spans": [ + { + "bbox": [ + 48, + 310, + 287, + 363 + ], + "type": "text", + "content": "[38] Youssef A Mejjati, Isa Milefchik, Aaron Gokaslan, Oliver Wang, Kwang In Kim, and James Tompkin. Gaussian: Controllable image synthesis with 3d gaussians from unposed silhouettes. arXiv preprint arXiv:2106.13215, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 365, + 287, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 287, + 398 + ], + "type": "text", + "content": "[39] Michael Niemeyer and Andreas Geiger. Campari: Camera-aware decomposed generative neural radiance fields. In 3DV, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 399, + 287, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 432 + ], + "type": "text", + "content": "[40] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In CVPR, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 434, + 287, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 287, + 466 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 287, + 466 + ], + "type": "text", + "content": "[41] Atsuhiro Noguchi, Xiao Sun, Stephen Lin, and Tatsuya Harada. Unsupervised learning of efficient geometry-aware neural articulated representations. In ECCV, 2022. 1, 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 468, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 468, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 287, + 521 + ], + "type": "text", + "content": "[42] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. StyleSDF: High-Resolution 3D-Consistent Image and Geometry Generation. arXiv preprint arXiv:2112.11427, 2021. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 523, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 567 + ], + "type": "text", + "content": "[43] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In CVPR, 2019. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 568, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 601 + ], + "type": "text", + "content": "[44] Ryan Po and Gordon Wetzstein. Compositional 3d scene generation using locally conditioned diffusion. arXiv preprint arXiv:2303.12218, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 602, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 679 + ], + "type": "text", + "content": "[45] Ryan Po, Wang Yifan, Vladislav Golyanik, Kfir Aberman, Jonathan T. Barron, Amit H. Bermano, Eric Ryan Chan, Tali Dekel, Aleksander Holynski, Angjoo Kanazawa, C. Karen Liu, Lingjie Liu, Ben Mildenhall, Matthias Nießner, Björn Ommer, Christian Theobalt, Peter Wonka, and Gordon Wetzstein. State of the art on diffusion models for visual computing, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "text", + "content": "[46] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "text", + "content": "[47] Serban D. Porumbescu, Brian Budge, Louis Feng, and Kenneth I. Joy. Shell maps. In ACM SIGGRAPH, 2005. 2, 3, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 107, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 149 + ], + "type": "text", + "content": "[48] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "text", + "content": "[49] Liu Ren, Hanspeter Pfister, and Matthias Zwicker. Object space ewa surface splatting: A hardware accelerated approach to high quality point rendering. In Comput. Graph. Forum, 2002. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "type": "text", + "content": "[50] Helge Rhodin, Nadia Robertini, Christian Richardt, Hans-Peter Seidel, and Christian Theobalt. A versatile scene model with differentiable visibility applied to generative pose estimation. In ICCV, 2015. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 241, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 274 + ], + "type": "text", + "content": "[51] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. CVPR, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 276, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 308 + ], + "type": "text", + "content": "[52] Darius Rückert, Linus Franke, and Marc Stamminger. Adop: Approximate differentiable one-pixel point rendering. ACM TOG, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 309, + 545, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 386 + ], + "type": "text", + "content": "[53] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Seyedeh Sara Mahdavi, Raphael Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. ArXiv, abs/2205.11487, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 388, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 409 + ], + "type": "text", + "content": "[54] Miguel Sainz and Renato Pajarola. Point-based rendering techniques. Computers & Graphics, 28(6):869-879, 2004. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 411, + 545, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 442 + ], + "type": "text", + "content": "[55] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In CVPR, 2018. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 444, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 444, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 444, + 545, + 487 + ], + "type": "text", + "content": "[56] Markus Schütz, Bernhard Kerbl, and Michael Wimmer. Software rasterization of 2 billion points in real time. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 489, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 545, + 521 + ], + "type": "text", + "content": "[57] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. In NeurIPS, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 523, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 545, + 567 + ], + "type": "text", + "content": "[58] Zackary P. T. Sin, Peter H. F. Ng, and Hong Va Leong. Ner-fahedron: A primitive for animatable neural rendering with interactive speed. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 568, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 568, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 307, + 568, + 545, + 601 + ], + "type": "text", + "content": "[59] Ivan Skorokhodov, Aliaksandr Siarohin, Yinghao Xu, Jian Ren, Hsin-Ying Lee, Peter Wonka, and Sergey Tulyakov. 3d generation on imagenet. In ICLR, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 602, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 602, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 602, + 545, + 635 + ], + "type": "text", + "content": "[60] Olga Sorkine, Daniel Cohen-Or, Yaron Lipman, Marc Alexa, Christian Rössl, and H-P Seidel. Laplacian surface editing. In Eurographics, 2004. 4" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 636, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 636, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 636, + 545, + 668 + ], + "type": "text", + "content": "[61] Carsten Stoll, Nils Hasler, Juergen Gall, Hans-Peter Seidel, and Christian Theobalt. Fast articulated motion tracking using a sums of gaussians body model. In ICCV, 2011. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[62] Jingxiang Sun, Xuan Wang, Yichun Shi, Lizhen Wang, Jue Wang, and Yebin Liu. Ide-3d: Interactive disentangled editing for high-resolution 3d-aware portrait synthesis. ACM TOG, 2022. 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9450" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[63] Jingxiang Sun, Xuan Wang, Lizhen Wang, Xiaoyu Li, Yong Zhang, Hongwen Zhang, and Yebin Liu. Next3d: Generative neural texture rasterization for 3d-aware head avatars. In CVPR, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 173 + ], + "type": "text", + "content": "[64] Ayush Tewari, Ohad Fried, Justus Thies, Vincent Sitzmann, Stephen Lombardi, Kalyan Sunkavalli, Ricardo MartinBrualla, Tomas Simon, Jason Saragih, Matthias Nießner, et al. State of the art on neural rendering. In Comput. Graph. Forum, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 228 + ], + "type": "text", + "content": "[65] Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srinivasan, Edgar Tretschk, Wang Yifan, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lombardi, et al. Advances in neural rendering. In Comput. Graph. Forum, 2022. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 230, + 287, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 287, + 273 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 287, + 273 + ], + "type": "text", + "content": "[66] Christina Tsalicoglou, Fabian Manhardt, Alessio Tonioni, Michael Niemeyer, and Federico Tombari. Textmesh: Generation of realistic 3d meshes from text prompts. ArXiv, abs/2304.12439, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 274, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 274, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 274, + 287, + 308 + ], + "type": "text", + "content": "[67] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. Synsin: End-to-end view synthesis from a single image. In CVPR, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 309, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 309, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 287, + 342 + ], + "type": "text", + "content": "[68] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In CVPR, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 343, + 287, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 287, + 387 + ], + "type": "text", + "content": "[69] Yinghao Xu, Sida Peng, Ceyuan Yang, Yujun Shen, and Bolei Zhou. 3d-aware image synthesis via learning structural and textural representations. arXiv preprint arXiv:2112.10759, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 388, + 287, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 442 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 442 + ], + "type": "text", + "content": "[70] Yinghao Xu, Menglei Chai, Zifan Shi, Sida Peng, Ivan Skorokhodov, Aliaksandr Siarohin, Ceyuan Yang, Yujun Shen, Hsin-Ying Lee, Bolei Zhou, et al. Discoscope: Spatially disentangled generative radiance fields for controllable 3d-aware scene synthesis. In CVPR, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 444, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 444, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 444, + 287, + 487 + ], + "type": "text", + "content": "[71] Yinghao Xu, Wang Yifan, Alexander W. Bergman, Mengei Chai, Bolei Zhou, and Gordon Wetzstein. Efficient 3d articulated human generation with layered surface volumes. In 3DV, 2024. 1, 2, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 488, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 488, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 488, + 287, + 521 + ], + "type": "text", + "content": "[72] Zhuoqian Yang, Shikai Li, Wayne Wu, and Bo Dai. 3dhuman: 3d-aware human image generation with 3d pose mapping. In ICCV, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 522, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 287, + 555 + ], + "type": "text", + "content": "[73] Wang Yifan, Felice Serena, Shihao Wu, Cengiz Oztireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM TOG, 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 557, + 287, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 287, + 600 + ], + "type": "text", + "content": "[74] Jianfeng Zhang, Zihang Jiang, Dingdong Yang, Hongyi Xu, Yichun Shi, Guoxian Song, Zhongcong Xu, Xinchao Wang, and Jiashi Feng. Avatargen: A 3d generative model for animatable human avatars. ArXiv, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 601, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 635 + ], + "type": "text", + "content": "[75] Qiang Zhang, Seung-Hwan Baek, Szymon Rusinkiewicz, and Felix Heide. Differentiable point-based radiance fields for efficient view synthesis. In SIGGRAPH Asia, 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "text", + "content": "[76] Xuanmeng Zhang, Jianfeng Zhang, Chacko Rohan, Hongyi Xu, Guoxian Song, Yi Yang, and Jiashi Feng. Getavatar: Generative textured meshes for animatable human avatars. In ICCV, 2023. 1, 2, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "text", + "content": "[77] Yufeng Zheng, Wang Yifan, Gordon Wetzstein, Michael J Black, and Otmar Hilliges. Pointavatar: Deformable point-based head avatars from videos. In CVPR, 2023. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 546, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 106 + ], + "type": "text", + "content": "[78] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Ewa volume splatting. In Proceedings Visualization, 2001. 3" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "9451" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Gaussian Splatting SLAM/32db8396-043e-45f8-8764-d01308949b32_content_list.json b/2024/Gaussian Splatting SLAM/32db8396-043e-45f8-8764-d01308949b32_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d24e850ff809f476f5885f50d058acb1ace0d0f4 --- /dev/null +++ b/2024/Gaussian Splatting SLAM/32db8396-043e-45f8-8764-d01308949b32_content_list.json @@ -0,0 +1,1706 @@ +[ + { + "type": "text", + "text": "Gaussian Splatting SLAM", + "text_level": 1, + "bbox": [ + 349, + 130, + 620, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hidenobu Matsuki $^{1*}$", + "bbox": [ + 135, + 179, + 302, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Riku Murai $^{2*}$", + "bbox": [ + 341, + 180, + 450, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Paul H. J. Kelly", + "bbox": [ + 491, + 180, + 625, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Andrew J. Davison1", + "bbox": [ + 666, + 180, + 825, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Dyson Robotics Laboratory, Imperial College London $^{2}$ Software Performance Optimisation Group, Imperial College London", + "bbox": [ + 205, + 205, + 763, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{h.matsuki20, riku.murai15, p.kelly, a.davison}@imperial.ac.uk", + "bbox": [ + 209, + 244, + 754, + 258 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c2f83a8fbdb7d189748a19aee544271d2f9f30f06468e5043b49446e3ad4f1e9.jpg", + "image_caption": [ + "Figure 1. From a single monocular camera, we reconstruct a high fidelity 3D scene live at 3fps. For every incoming RGB frame, 3D Gaussians are incrementally formed and optimised together with the camera poses. We show both the rasterised Gaussians (left) and Gaussians shaded to highlight the geometry (right). Notice the details and the complex material properties (e.g. transparency) captured. Thin structures such as wires are accurately represented by numerous small, elongated Gaussians, and transparent objects are effectively represented by placing the Gaussians along the rim. Our system significantly advances the fidelity a live monocular SLAM system can capture." + ], + "image_footnote": [], + "bbox": [ + 81, + 294, + 196, + 459 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0cf846d57eab8d874d336abf74b7dbdc1f6ca25d48c9d112804de3a177a69eb4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 197, + 295, + 485, + 459 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/de58f0539762e5fe3ede81e0b04853165ffba86077c8ef7af1cf3f36874bb722.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 295, + 774, + 459 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/929b63922c0a7c841cf1939d4afee510145b830db8bcce80ff9e2c27b03447d4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 295, + 888, + 459 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 583, + 313, + 599 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present the first application of 3D Gaussian Splating in monocular SLAM, the most fundamental but the hardest setup for Visual SLAM. Our method, which runs live at 3fps, utilises Gaussians as the only 3D representation, unifying the required representation for accurate, efficient tracking, mapping, and high-quality rendering. Designed for challenging monocular settings, our approach is seamlessly extendable to RGB-D SLAM when an external depth sensor is available. Several innovations are required to continuously reconstruct 3D scenes with high fidelity from a live camera. First, to move beyond the original 3DGS algorithm, which requires accurate poses from an offline Structure from Motion (SfM) system, we formulate camera tracking for 3DGS using direct optimisation against the 3D Gaussians, and show that this enables fast and robust tracking with a wide basin of convergence. Second, by utilising the explicit nature of the Gaussians, we introduce geometric verification", + "bbox": [ + 75, + 616, + 472, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and regularisation to handle the ambiguities occurring in incremental 3D dense reconstruction. Finally, we introduce a full SLAM system which not only achieves state-of-the-art results in novel view synthesis and trajectory estimation but also reconstruction of tiny and even transparent objects.", + "bbox": [ + 500, + 570, + 892, + 645 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 676, + 632, + 693 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A long-term goal of online reconstruction with a single moving camera is near-photorealistic fidelity, which will surely allow new levels of performance in many areas of Spatial AI and robotics as well as opening up a whole range of new applications. While we increasingly see the benefit of applying powerful pre-trained priors to 3D reconstruction, a key avenue for progress is still the invention and development of core 3D representations with advantageous properties. Many \"layered\" SLAM methods exist which tackle the SLAM problem by integrating multiple different 3D representations or existing SLAM components; however, the most interesting advances are when a new unified dense representation can be used for all aspects", + "bbox": [ + 496, + 703, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Authors contributed equally to this work.", + "bbox": [ + 99, + 886, + 326, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18039", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of a system's operation: local representation of detail, large-scale geometric mapping and also camera tracking by direct alignment.", + "bbox": [ + 78, + 90, + 467, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we present the first online visual SLAM system based solely on the 3D Gaussian Splatting (3DGS) representation [10] recently making a big impact in offline scene reconstruction. In 3DGS a scene is represented by a large number of Gaussian blobs with orientation, elongation, colour and opacity. Other previous world/map-centric scene representations used for visual SLAM include occupancy or Signed Distance Function (SDF) voxel grids [23]; meshes [28]; point or surfel clouds [9, 29]; and recently neural fields [33]. Each of these has disadvantages: grids use significant memory and have bounded resolution, and even if octrees or hashing allow more efficiency they cannot be flexibly warped for large corrections [25, 37]; meshes require difficult, irregular topology to fuse new information; surfel clouds are discontinuous and difficult to fuse and optimise; and neural fields require expensive per-pixel raycasting to render. We show that 3DGS has none of these weaknesses. As a SLAM representation, it is most similar to point and surfel clouds, and inherits their efficiency, locality and ability to be easily warped or modified. However, it also represents geometry in a smooth, continuously differentiable way: a dense cloud of Gaussians merge together and jointly define a continuous volumetric function. And crucially, the design of modern graphics cards means that a large number of Gaussians can be efficiently rendered via \"splatting\" rasterisation, up to 200fps at $1080\\mathrm{p}$ . This rapid, differentiable rendering is integral to the tracking and map optimisation loops in our system.", + "bbox": [ + 80, + 142, + 467, + 563 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The 3DGS representation has up until now only been used in offline systems for 3D reconstruction with known camera poses, and we present several innovations to enable online SLAM. We first derive the analytic Jacobian on Lie group of camera pose with respect to a 3D Gaussians map, and show that this can be seamlessly integrated into the existing differentiable rasterisation pipeline to enable camera poses to be optimised alongside scene geometry. Second, we introduce a novel Gaussian isotropic shape regularisation, to ensure geometric consistency, which we have found is important for incremental reconstruction. Third, we propose a novel Gaussian resource allocation and pruning method to keep the geometry clean and enable accurate camera tracking. Our experimental results demonstrate photorealistic online local scene reconstruction, as well as state-of-the-art camera trajectory estimation and mapping for larger scenes compared to other rendering-based SLAM methods. We further show the uniqueness of the Gaussian-based SLAM method such as an extremely large camera pose convergence basin, which can also be useful for map-based camera localisation. Our method works with only monocular input, one of the most challenging scenarios in", + "bbox": [ + 80, + 569, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SLAM. To highlight the intrinsic capability of 3D Gaussian for camera localisation, our method does not use any pretrained monocular depth predictor or other existing tracking modules, but relies solely on RGB image inputs in line with the original 3DGS. Since this is one of the most challenging SLAM scenario, we also show our method can easily be extended to RGB-D SLAM when depth measurements are available.", + "bbox": [ + 503, + 90, + 890, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are as follows:", + "bbox": [ + 521, + 213, + 818, + 226 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- The first near real-time SLAM system which works with a 3DGS as the only underlying scene representation, which can handle monocular only inputs.", + "- Novel techniques within the SLAM framework, including the analytic Jacobian on Lie group for direct camera pose estimation, isotropic regularisation of the Gaussian shape, and geometric verification.", + "- Extensive evaluations on a variety of datasets both for monocular and RGB-D settings, demonstrating competitive performance, particularly in real-world scenarios." + ], + "bbox": [ + 503, + 228, + 890, + 376 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 503, + 392, + 638, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dense SLAM: Dense visual SLAM focuses on reconstructing detailed 3D maps, unlike sparse SLAM methods which excel in pose estimation [4, 5, 21] but typically yield maps useful mainly for localisation. In contrast, dense SLAM creates interactive maps beneficial for broader applications, including AR and robotics. Dense SLAM methods are generally divided into two primary categories: Frame-centric and Map-centric. Frame-centric SLAM minimises photometric error across consecutive frames, jointly estimating per-frame depth and frame-to-frame camera motion. Frame-centric approaches [1, 36] are efficient, as individual frames host local rather than global geometry (e.g. depth maps), and are attractive for long-session SLAM, but if a dense global map is needed, it must be constructed on demand by assembling all of these parts which are not necessarily fully consistent. In contrast, Map-centric SLAM uses a unified 3D representation across the SLAM pipeline, enabling a compact and streamlined system. Compared to purely local frame-to-frame tracking, a map-centric approach leverages global information by tracking against the reconstructed 3D consistent map. Classical map-centric approaches often use voxel grids [2, 23, 26, 40] or points [9, 29, 41] as the underlying 3D representation. While voxels enable a fast look-up of features in 3D, the representation is expensive, and the fixed voxel resolution and distribution are problematic when the spatial characteristics of the environment are not known in advance. On the other hand, a point-based map representation, such as surfel clouds, enables adaptive changes in resolution and spatial distribution by dynamic allocation of point primitives in the 3D space. Such flexibility benefits online applications such as SLAM with deformation-based loop closure [29, 41]. However, op", + "bbox": [ + 503, + 417, + 890, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18040", + "bbox": [ + 482, + 946, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "timising the representation to capture high fidelity is challenging due to the lack of correlation among the primitives. Recently, in addition to classical graphic primitives, neural network-based map representations are a promising alternative. iMAP [33] demonstrated the interesting properties of neural representation, such as sensible hole filling of unobserved geometry. Many recent approaches combine the classical and neural representations to capture finer details [8, 27, 46, 47]; however, the large amount of computation required for neural rendering makes the live operation of such systems challenging.", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Differentiable Rendering: The classical method for creating a 3D representation was to unproject 2D observations into 3D space and to fuse them via weighted averaging [16, 23]. Such an averaging scheme suffers from over-smooth representation and lacks the expressiveness to capture high-quality details. To capture a scene with photorealistic quality, differentiable volumetric rendering [24] has recently been popularised with Neural Radiance Fields (NeRF) [17]. Using a single Multi-Layer Perceptron (MLP) as a scene representation, NeRF performs volume rendering by marching along pixel rays, querying the MLP for opacity and colour. Since volume rendering is naturally differentiable, the MLP representation is optimised to minimise the rendering loss using multiview information to achieve high-quality novel view synthesis. The main weakness of NeRF is its training speed. Recent developments have introduced explicit volume structures such as multi-resolution voxel grids [6, 14, 34] or hash functions [19] to improve performance. Interestingly, these projects demonstrate that the main contributor to high-quality novel view synthesis is not the neural network but rather differentiable volumetric rendering, and that it is possible to avoid the use of an MLP and yet achieve comparable rendering quality to NeRF [6]. However, even in these systems, per-pixel ray marching remains a significant bottleneck for rendering speed. This issue is particularly critical in SLAM, where immediate interaction with the map is essential for tracking. In contrast to NeRF, 3DGS performs differentiable rasterisation. Similar to regular graphics rasterisations, by iterating over the primitives to be rasterised rather than marching along rays, 3DGS leverages the natural sparsity of a 3D scene and achieves a representation which is expressive to capture high-fidelity 3D scenes while offering significantly faster rendering. Several works have applied 3D Gaussians and differentiable rendering to static scene capture [11, 38], and in particular more recent works utilise 3DGS and demonstrate superior results in vision tasks such as dynamic scene capture [15, 42, 44] and 3D generation [35, 45]. Our method adopts a Map-centric approach, utilising 3D Gaussians as the only SLAM representation. Similar to surfel-based SLAM, we dynamically allocate the 3D Gaussians, enabling us to model an arbitrary spatial dis", + "bbox": [ + 75, + 265, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tribution in the scene. Unlike other methods such as Elastic-Fusion [41] and PointFusion [9], however, by using differentiable rasterisation, our SLAM system can capture high-fidelity scene details and represent challenging object properties by direct optimisation against information from every pixel.", + "bbox": [ + 496, + 90, + 893, + 183 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 196, + 591, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Gaussian Splatting", + "text_level": 1, + "bbox": [ + 500, + 220, + 683, + 238 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our SLAM representation is 3DGS, mapping the scene with a set of anisotropic Gaussians $\\mathcal{G}$ . Each Gaussian $\\mathcal{G}^i$ contains optical properties: colour $c^i$ and opacity $\\alpha^i$ . For continuous 3D representation, the mean $\\mu_W^i$ and covariance $\\Sigma_W^i$ , defined in the world coordinate, represent the Gaussian's position and its ellipsoidal shape. For simplicity and speed, in our work we omit the spherical harmonics representing view-dependent radiance. Since 3DGS uses volume rendering, explicit extraction of the surface is not required. Instead, by splatting and blending $\\mathcal{N}$ Gaussians, a pixel colour $\\mathcal{C}_p$ is synthesised:", + "bbox": [ + 496, + 244, + 893, + 412 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} _ {p} = \\sum_ {i \\in \\mathcal {N}} c _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 424, + 890, + 467 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3DGS performs rasterisation, iterating over the Gaussians rather than marching along the camera rays, and hence, free spaces are ignored during rendering. During rasterisation, the contributions of $\\alpha$ are decayed via a Gaussian function, based on the 2D Gaussian formed by splatting a 3D Gaussian. The 3D Gaussians $\\mathcal{N}(\\pmb{\\mu}_W, \\pmb{\\Sigma}_W)$ in world coordinates are related to the 2D Gaussians $\\mathcal{N}(\\pmb{\\mu}_I, \\pmb{\\Sigma}_I)$ on the image plane through a projective transformation:", + "bbox": [ + 496, + 477, + 893, + 599 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\mu} _ {I} = \\pi (\\boldsymbol {T} _ {C W} \\cdot \\boldsymbol {\\mu} _ {W}), \\boldsymbol {\\Sigma} _ {I} = \\mathbf {J} \\mathbf {W} \\boldsymbol {\\Sigma} _ {W} \\mathbf {W} ^ {T} \\mathbf {J} ^ {T}, (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 609, + 890, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\pi$ is the projection operation and $T_{CW} \\in SE(3)$ is the camera pose of the viewpoint. $\\mathbf{J}$ is the Jacobian of the linear approximation of the projective transformation and $\\mathbf{W}$ is the rotational component of $T_{CW}$ . This formulation enables the 3D Gaussians to be differentiable and the blending operation provides gradient flow to the Gaussians. Using first-order gradient descent [12], Gaussians gradually refines both their optic and geometric parameters to represent the captured scene with high fidelity.", + "bbox": [ + 496, + 638, + 893, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Camera Pose Optimisation", + "text_level": 1, + "bbox": [ + 500, + 786, + 743, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To achieve accurate tracking, we typically require at least 50 iterations of gradient descent per frame. This requirement emphasises the necessity of a representation with computationally efficient view synthesis and gradient computation, making the choice of 3D representation a crucial part of designing a SLAM system.", + "bbox": [ + 496, + 809, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18041", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/97ceacd1e19adb68856325d2133821f4017de23d7ca01b34d044a0e689e68a9c.jpg", + "image_caption": [ + "Figure 2. SLAM System Overview: Our SLAM system uses 3D Gaussians as the only representation, unifying all components of SLAM, including tracking, mapping, keyframe management, and novel view synthesis." + ], + "image_footnote": [], + "bbox": [ + 76, + 88, + 895, + 325 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In order to avoid the overhead of automatic differentiation, 3DGS implements rasterisation with CUDA with derivatives for all parameters calculated explicitly. Since rasterisation is performance critical, we similarly derive the camera Jacobians explicitly.", + "bbox": [ + 75, + 381, + 468, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To the best of our knowledge, we provide the first analytical Jacobian of $\\pmb{SE}(3)$ camera pose with respect to the 3D Gaussians used in EWA splatting [48] and 3DGS. This opens up new applications of 3DGS beyond SLAM.", + "bbox": [ + 75, + 455, + 468, + 517 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use Lie algebra to derive the minimal Jacobians, ensuring that the dimensionality of the Jacobians matches the degrees of freedom, eliminating any redundant computations. The terms of Eq. (2) are differentiable with respect to the camera pose $\\pmb{T}_{CW}$ ; using the chain rule:", + "bbox": [ + 75, + 517, + 468, + 594 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial \\boldsymbol {\\mu} _ {I}}{\\partial \\boldsymbol {T} _ {C W}} = \\frac {\\partial \\boldsymbol {\\mu} _ {I}}{\\partial \\boldsymbol {\\mu} _ {C}} \\frac {\\mathcal {D} \\boldsymbol {\\mu} _ {C}}{\\mathcal {D} \\boldsymbol {T} _ {C W}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 597, + 468, + 628 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial \\boldsymbol {\\Sigma} _ {I}}{\\partial \\boldsymbol {T} _ {C W}} = \\frac {\\partial \\boldsymbol {\\Sigma} _ {I}}{\\partial \\mathbf {J}} \\frac {\\partial \\mathbf {J}}{\\partial \\boldsymbol {\\mu} _ {C}} \\frac {\\mathcal {D} \\boldsymbol {\\mu} _ {C}}{\\mathcal {D} \\boldsymbol {T} _ {C W}} + \\frac {\\partial \\boldsymbol {\\Sigma} _ {I}}{\\partial \\mathbf {W}} \\frac {\\mathcal {D} \\mathbf {W}}{\\mathcal {D} \\boldsymbol {T} _ {C W}}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 631, + 468, + 664 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $T_{CW}$ represents the 3D position of Gaussian in the camera coordinate. We take the derivatives on the manifold to derive minimal parameterisation. Borrowing the notation from [30], let $T \\in SE(3)$ and $\\tau \\in \\mathfrak{se}(3)$ . We define the partial derivative on the manifold as:", + "bbox": [ + 75, + 667, + 468, + 744 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\mathcal {D} f (\\boldsymbol {T})}{\\mathcal {D} \\boldsymbol {T}} \\triangleq \\lim _ {\\tau \\rightarrow 0} \\frac {\\operatorname {L o g} (f (\\operatorname {E x p} (\\tau) \\circ \\boldsymbol {T}) \\circ f (\\boldsymbol {T}) ^ {- 1})}{\\tau}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 748, + 468, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\circ$ is a group composition, and Exp, Log are the exponential and logarithmic mappings between Lie algebra and Lie Group. With this, we derive the following:", + "bbox": [ + 75, + 784, + 468, + 829 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\mathcal {D} \\boldsymbol {\\mu} _ {C}}{\\mathcal {D} \\boldsymbol {T} _ {C W}} = \\left[ \\begin{array}{l l} \\boldsymbol {I} & - \\boldsymbol {\\mu} _ {C} ^ {\\times} \\end{array} \\right], \\frac {\\mathcal {D} \\mathbf {W}}{\\mathcal {D} \\boldsymbol {T} _ {C W}} = \\left[ \\begin{array}{l l} \\mathbf {0} & - \\mathbf {W} _ {:, 1} ^ {\\times} \\\\ \\mathbf {0} & - \\mathbf {W} _ {:, 2} ^ {\\times} \\\\ \\mathbf {0} & - \\mathbf {W} _ {:, 3} ^ {\\times} \\end{array} \\right] ， \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 847, + 468, + 897 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\times$ denotes the skew symmetric matrix of a 3D vector, and $\\mathbf{W}_{:,i}$ refers to the $i$ th column of the matrix.", + "bbox": [ + 498, + 381, + 890, + 412 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. SLAM", + "text_level": 1, + "bbox": [ + 500, + 421, + 589, + 435 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we present details of full SLAM framework. The overview of the system is summarised in Fig. 2. Please refer to the supplementary material for the further parameter details.", + "bbox": [ + 498, + 445, + 890, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1 Tracking", + "text_level": 1, + "bbox": [ + 500, + 525, + 617, + 541 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In tracking only the current camera pose is optimised, without updates to the map representation. In the monocular case, we minimise the following photometric residual:", + "bbox": [ + 498, + 550, + 890, + 595 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nE _ {p h o} = \\left\\| I (\\mathcal {G}, \\boldsymbol {T} _ {C W}) - \\bar {I} \\right\\| _ {1}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 606, + 890, + 626 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $I(\\mathcal{G},\\pmb{T}_{CW})$ renders the Gaussians $\\mathcal{G}$ from $T_{CW}$ , and $\\bar{I}$ is an observed image.", + "bbox": [ + 500, + 636, + 890, + 666 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We further optimise affine brightness parameters for varying exposure. When depth observations are available, we define the geometric residual as:", + "bbox": [ + 500, + 667, + 890, + 712 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nE _ {g e o} = \\left\\| D (\\mathcal {G}, \\boldsymbol {T} _ {C W}) - \\bar {D} \\right\\| _ {1}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 723, + 890, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $D(\\mathcal{G},\\mathbf{T}_{CW})$ is depth rasterisation and $\\bar{D}$ is the observed depth. Rather than simply using the depth measurements to initialise the Gaussians, we minimise both photometric and geometric residuals: $\\lambda_{pho}E_{pho} + (1 - \\lambda_{pho})E_{geo}$ , where $\\lambda_{pho}$ is a hyperparameter.", + "bbox": [ + 498, + 753, + 890, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As in Eq. (1), per-pixel depth is rasterised by alphabetending:", + "bbox": [ + 498, + 830, + 890, + 861 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} _ {p} = \\sum_ {i \\in \\mathcal {N}} z _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 861, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18042", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $z_{i}$ is the distance to the mean $\\pmb{\\mu}_{W}$ of Gaussian $i$ along the camera ray. We derive analytical Jacobians for the camera pose optimisation in a similar manner to Eq. (3), (4).", + "bbox": [ + 76, + 90, + 468, + 137 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.2 Keyframing", + "text_level": 1, + "bbox": [ + 76, + 155, + 217, + 171 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since using all the images from a video stream to jointly optimise the Gaussians and camera poses online is infeasible, we maintain a small window $\\mathcal{W}_k$ consisting of carefully selected keyframes based on inter-frame covisibility. Ideal keyframe management will select non-redundant keyframes observing the same area, spanning a wide baseline to provide better multiview constraints. The parameters are detailed in supplementary.", + "bbox": [ + 75, + 179, + 470, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Selection and Management Every tracked frame is checked for keyframe registration based on our simple yet effective criteria. We measure the可视ibility by measuring the intersection over union of the observed Gaussians between the current frame $i$ and the last keyframe $j$ . If the可视ibility drops below a threshold, or if the relative translation $t_{ij}$ is large with respect to the median depth, frame $i$ is registered as a keyframe. For efficiency, we maintain only a small number of keyframes in the current window $\\mathcal{W}_k$ following the keyframe management heuristics of DSO [4]. The main difference is that a keyframe is removed from the current window if the overlap coefficient with the latest keyframe drops below a threshold.", + "bbox": [ + 75, + 319, + 472, + 515 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Gaussian Covisibility An accurate estimate of covisibility simplifies keyframe selection and management. 3DGS respects visibility ordering since the 3D Gaussians are sorted along the camera ray. This property is desirable for covisibility estimation as occlusions are handled by design. A Gaussian is marked to be visible from a view if used in the rasterisation and if the ray's accumulated $\\alpha$ has not yet reached 0.5. This enables our estimated covisibility to handle occlusions without requiring additional heuristics.", + "bbox": [ + 75, + 534, + 468, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Gaussian Insertion and Pruning At every keyframe, new Gaussians are inserted into the scene to capture newly visible scene elements and to refine the fine details. When depth measurements are available, Gaussian means $\\mu_W$ are initialised by back-projecting the depth. In the monocular case, we render the depth at the current frame. For pixels with depth estimates, $\\mu_W$ are initialised around those depths with low variance; for pixels without the depth estimates, we initialise $\\mu_W$ around the median depth of the rendered image with high variance.", + "bbox": [ + 75, + 689, + 468, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the monocular case, the positions of many newly inserted Gaussians are incorrect. While the majority will quickly vanish during optimisation as they violate multiview consistency, we further prune the excess Gaussians by", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "checking the visibility amongst the current window $\\mathcal{W}_k$ . If the Gaussians inserted within the last 3 keyframes are unobserved by at least 3 other frames, we prune them out as they are geometrically unstable.", + "bbox": [ + 496, + 90, + 893, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.3 Mapping", + "text_level": 1, + "bbox": [ + 500, + 170, + 619, + 186 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The purpose of mapping is to maintain a coherent 3D structure and to optimise the newly inserted Gaussians. During mapping, the keyframes in $\\mathcal{W}_k$ are used to reconstruct currently visible regions. Additionally, two random past keyframes $\\mathcal{W}_r$ are selected per iteration to avoid forgetting the global map. Rasterisation of 3DGS imposes no constraint on the Gaussians along the viewing ray direction, even with a depth observation. This is not a problem when sufficient carefully selected viewpoints are provided (e.g. in the novel view synthesis case); however, in continuous SLAM this causes many artefacts, making tracking challenging. We therefore introduce an isotropic regularisation:", + "bbox": [ + 496, + 193, + 893, + 375 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nE _ {i s o} = \\sum_ {i = 1} ^ {| \\mathcal {G} |} \\| \\mathbf {s} _ {i} - \\tilde {\\mathbf {s}} _ {i} \\cdot \\mathbf {1} \\| _ {1} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 385, + 890, + 429 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "to penalise the scaling parameters $\\mathbf{s}_i$ (i.e. stretch of the ellipsoid) by its difference to the mean $\\tilde{\\mathbf{s}}_i$ . As shown in Fig 3, this encourages sphericality, and avoids the problem of Gaussians which are highly elongated along the viewing direction creating artefacts. Let the union of the keyframes in the current window and the randomly selected one be $\\mathcal{W} = \\mathcal{W}_k \\cup \\mathcal{W}_r$ . For mapping, we solve the following problem:", + "bbox": [ + 496, + 438, + 893, + 559 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\boldsymbol {T} _ {C W} ^ {k} \\in \\boldsymbol {S E} (3), \\mathcal {G}, \\forall k \\in \\mathcal {W}} \\sum_ {\\forall k \\in \\mathcal {W}} E _ {p h o} ^ {k} + \\lambda_ {i s o} E _ {i s o}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 570, + 890, + 608 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "If depth observations are available, as in tracking, geometric residuals Eq. (8) are added to the optimisation problem.", + "bbox": [ + 496, + 617, + 890, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Evaluation", + "text_level": 1, + "bbox": [ + 500, + 661, + 617, + 676 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conduct a comprehensive evaluation of our system across a range of both real and synthetic datasets. Additionally, we perform an ablation study to justify our design choices. Finally, we present qualitative results of our system operating live using a monocular camera, illustrating its practicality and high fidelity reconstruction.", + "bbox": [ + 496, + 686, + 890, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 500, + 786, + 689, + 803 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets For our quantitative analysis, we evaluate our method on the TUM RGB-D dataset [32] (3 sequences) and the Replica dataset [31] (8 sequences), following the evaluation in [33]. For qualitative results, we use self-captured real-world sequences recorded by Intel Realsense d455. Since the Replica dataset is designed for RGB-D SLAM", + "bbox": [ + 496, + 810, + 893, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18043", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d3b400550f268e328fb72f93546cffb3b56e822362318d4823f1d64a4bd74d87.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 89, + 465, + 208 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cfb4ab71fd891c902d7dffe34369c688ecaeb250f6f3e0253f77581c79108702.jpg", + "image_caption": [ + "Figure 3. Effect of isotropic regularisation: Top: Rendering close to a training view (looking at the keyboard). Bottom: Rendering 3D Gaussians far from the training views (view from a side of the keyboard) without (left) and with (right) the isotropic loss. When the photometric constraints are insufficient, the Gaussians tend to elongate along the viewing direction, creating artefacts in the novel views, and affecting the camera tracking." + ], + "image_footnote": [], + "bbox": [ + 80, + 210, + 279, + 300 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d4232cbfd10cd50adaf9060720a2b68ce9c4820b6d9dea29265c927bb26fe56b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 210, + 467, + 300 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "evaluation, it contains challenging purely rotational camera motions. We hence use the Replica dataset for RGB-D evaluation only. The TUM RGB-D dataset is used for both monocular and RGB-D evaluation.", + "bbox": [ + 76, + 409, + 467, + 468 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details We run our SLAM on a desktop with Intel Core i9 12900K 3.50GHz and a single NVIDIA GeForce RTX 4090. We present results from our multiprocessing implementation aimed at real-time applications. For a fair comparison with other methods on Replica, we additionally report result for single-process implementation which performs more mapping iterations. As with 3DGS, time-critical rasterisation and gradient computation are implemented using CUDA. The rest of the SLAM pipeline is developed with PyTorch. Details of hyperparameters are provided in the supplementary material.", + "bbox": [ + 76, + 492, + 467, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics For camera tracking accuracy, we report the Root Mean Square Error (RMSE) of the Absolute Trajectory Error (ATE) of the keyframes. To evaluate map quality, we report standard photometric rendering quality metrics (PSNR, SSIM and LPIPS) following the evaluation protocol used in [27]. To evaluate the map quality, on every fifth frame, rendering metrics are computed. We exclude the keyframes (training views). We report the average across three runs for all our evaluations. In the tables, the best result is in bold, and the second best is underlined.", + "bbox": [ + 76, + 681, + 467, + 830 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baseline Methods We primarily benchmark our SLAM method against other approaches that, like ours, do not have explicit loop closure. In monocular settings, we compare", + "bbox": [ + 76, + 854, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "with state-of-the-art classical and learning-based direct visual odometry (VO) methods. Specifically, we compare DSO [4], DepthCov [3], and DROID-SLAM [36] in VO configurations. These methods are selected based on their public reporting of results on the benchmark (TUM dataset) or the availability of their source code for getting the benchmark result. Since one of our focuses is the online scale estimation under monocular scale ambiguity, the method which uses ground truth poses for the system initialisation such as [13] is not considered for the comparison. In the RGB-D case, we compare against neural-implicit SLAM methods [7, 8, 27, 33, 39, 43, 46] which are also map-centric, rendering-based and do not perform loop closure.", + "bbox": [ + 496, + 90, + 890, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 500, + 297, + 720, + 313 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Camera Tracking Accuracy Table 1 shows the tracking results on the TUM RGB-D dataset. In the monocular setting, our method surpasses other baselines without requiring any deep priors. Furthermore, our performance is comparable to systems which perform explicit loop closure. This clearly highlights that there still remains potential for enhancing the tracking of monocular SLAM by exploring fundamental SLAM representations.", + "bbox": [ + 496, + 321, + 890, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our RGB-D method shows better performance than any other baseline method. Notably, our system surpasses ORB-SLAM in the fr1 sequences, narrowing the gap between Map-centric SLAM and the state-of-the-art sparse frame-centric methods. Table 2 reports results on the synthetic Replica dataset. Our single-process implementation shows competitive performance and achieves the best result in 4 out of 8 sequences. Our multi-process implementation which performs fewer mapping iterations still performs comparably. In contrast to other methods, our system demonstrates higher performance on real-world data, by optimising the Gaussian positions to compensate for the sensor noise.", + "bbox": [ + 496, + 441, + 890, + 637 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Novel View Rendering Table 5 summarises the novel view rendering performance of our method with RGB-D input. We consistently show the best performance across most sequences and is least second best. Our rendering FPS is hundreds of times faster than other methods, offering a significant advantage for applications which require real-time map interaction. While Point-SLAM is competitive, that method focuses on view synthesis rather than novel-view synthesis. Their view synthesis is conditional on the availability of depth due to the depth-guided ray-sampling, making novel-view synthesis challenging. On the other hand, our rasterisation-based approach does not require depth guidance and achieves efficient, high-quality, novel view synthesis. Fig. 4 provides a qualitative comparison of the rendering of ours and Point-SLAM (with depth guidance).", + "bbox": [ + 496, + 657, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18044", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a9996696ef891870be6d242ff078e8edc16684df8dc79aaa15611698e0ff80d0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
InputLoop-closureMethodfr1/deskfr2/xyzfr3/officeAvg.
Monocularw/oDSO [4]22.41.109.5011.0
DROID-VO [36]5.2010.77.307.73
DepthCov-VO [3]5.601.2068.825.2
Ours3.784.603.503.96
w/DROID-SLAM [36]1.800.502.801.70
ORB-SLAM2 [20]1.900.602.401.60
RGB-Dw/oiMAP [33]4.902.005.804.23
NICE-SLAM [46]4.266.193.874.77
DI-Fusion [7]4.402.005.804.07
Vox-Fusion [43]3.521.4926.0110.34
ESLAM [8]2.471.112.422.00
Co-SLAM [39]2.401.702.402.17
Point-SLAM [27]4.341.313.483.04
Ours1.501.441.491.47
w/BAD-SLAM [29]1.701.101.701.50
Kintinous [40]3.702.903.003.20
ORB-SLAM2 [20]1.600.401.001.00
", + "bbox": [ + 81, + 88, + 467, + 321 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b649b474e53fb76a34b8433f71bb29fb6db2398a96fe38c4b5ba8401a56f8e50.jpg", + "table_caption": [ + "Table 1. Camera tracking result on TUM for monocular and RGB-D. ATE RMSE in cm is reported. In both monocular and RGB-D cases, we achieve state-of-the-art performance. In particular, in the monocular case, not only do we outperform systems which use deep prior, but we achieve comparable performance with many of the RGB-D systems." + ], + "table_footnote": [], + "table_body": "
Methodr0r1r2o0o1o2o3o4Avg.
iMAP [33]3.122.542.311.691.033.994.051.932.58
NICE-SLAM0.971.311.070.881.001.061.101.131.07
Vox-Fusion [43]1.374.701.478.482.042.581.112.943.09
ESLAM [8]0.710.700.520.570.550.580.720.630.63
Point-SLAM [27]0.610.410.370.380.480.540.690.720.53
Ours0.440.320.310.440.520.230.172.250.58
Ours (sp)0.330.220.290.360.190.250.120.810.32
", + "bbox": [ + 81, + 412, + 465, + 497 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/02faa4e7e5d5616fe9bdd110731759aa7590ebab795b0191b4136c805e81fe44.jpg", + "table_caption": [ + "Table 2. Camera tracking result on Replica for RGB-D SLAM. ATE RMSE in cm is reported. We achieve best performance across most sequences. Here, Ours is our multi-process implementation and Ours (sp) is the single-process implementation which ensures a certain amount of mapping iteration similar to other works." + ], + "table_footnote": [], + "table_body": "
InputMethodfr1/deskfr2/xyzfr3/officeAvg.
Monow/o \\( E_{iso} \\)4.164.665.734.83
w/o kf selection13.24.368.658.73
Ours3.784.603.503.96
RGB-Dw/o \\( E_{geo} \\)2.390.624.982.66
w/o kf selection1.641.492.601.90
Ours1.501.441.491.47
", + "bbox": [ + 109, + 574, + 434, + 652 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9f9d2a943d2f1cba74ad69cf8453bbee9526ebe635285116340c5a7e66847976.jpg", + "table_caption": [ + "Table 3. Ablation Study on TUM RGB-D dataset. We analyse the usefulness of isotropic regularisation, geometric residual, and keyframe selection to our SLAM system. Further isotropic regularisation ablation is available in supplementary." + ], + "table_footnote": [], + "table_body": "
Memory Usage [MB]
iMAP [33]NICE-SLAM [46]Co-SLAM [39]Ours (Mono)Ours (RGB-D)
0.8MB40.3.4MB6.4MB2.6MB3.97MB
", + "bbox": [ + 80, + 715, + 465, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablative Analysis In Table 3, we perform ablation to confirm our design choices. Isotropic regularisation and geometric residual improve the tracking of monocular and RGB-D SLAM respectively, as they aid in constraining the geometry when photometric signals are weak. For both cases, keyframe selection significantly improves systems performance, as it automatically chooses suitable keyframes", + "bbox": [ + 75, + 787, + 468, + 893 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7573fad4bc6edfc0a0c23e59f6984cac66f8b7ab0f9ddc04ec6efceae58797bf.jpg", + "table_caption": [ + "Table 4. Memory Analysis on TUM RGB-D dataset. The baseline numbers are computed from the parameter numbers in [39]" + ], + "table_footnote": [], + "table_body": "
MethodPSNR[db]↑SSIM↑LPIPS↓Rendering FPS
NICE-SLAM[46]24.420.8090.2330.54
Vox-Fusion[43]24.410.8010.2362.17
Point-SLAM [27]35.170.9750.1241.33
ours38.940.9680.070769
", + "bbox": [ + 504, + 88, + 888, + 147 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/22e58c9c0ffa4fd7aec784b64abf8ac6b021bddd9689fd072931ef9cc292d6ff.jpg", + "image_caption": [ + "Figure 4. Rendering examples on Replica. Point-SLAM struggle with rendering fine details due to the stochastic ray sampling." + ], + "image_footnote": [], + "bbox": [ + 503, + 227, + 888, + 477 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "based on our occlusion-aware keyframe selection and management. We further compare the memory usage of different 3D representations in Table 4. MLP-based iMAP is clearly more memory efficient, but it struggles to express high-fidelity 3D scenes due to the limited capacity of small MLP. Compared with a voxel grid of features used in NICE-SLAM, our method uses significantly less memory.", + "bbox": [ + 498, + 517, + 890, + 623 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4d957e0ab619b8b97d7cd5b1b7b0d691a60424ff86934ad6b629c2c474e6fa34.jpg", + "table_caption": [ + "Table 5. Average rendering performance on Replica (RGB-D). Our method outperforms most of the rendering metrics compared to existing methods. Note that Point-SLAM uses ground-truth depth to guide sampling along rays. The full detail is available in supplementary." + ], + "table_footnote": [], + "table_body": "
Methodseq1seq2seq3Avg.
Neural SDF (Hash Grid)0.130.150.160.14
Neural SDF (MLP)0.400.380.220.33
Ours w/o depth0.820.910.650.79
Ours w/ depth0.831.00.650.82
", + "bbox": [ + 522, + 647, + 869, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Camera convergence analysis. We report the ratio of successful camera convergence for the different sequences, across different differentiable 3D representations.", + "bbox": [ + 498, + 720, + 890, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Convergence Basin Analysis In our SLAM experiments, we discovered that 3D Gaussian maps have a notably large convergence basin for camera localisation. To investigate further, we conducted a convergence funnel analysis, an evaluation methodology proposed in [18] and used in [22]. Here, we train a 3D representation (e.g. 3DGS) using 9 fixed views arranged in a square. We set the viewpoint in the middle of the square to be the target view. As shown in Fig 5, we uniformly sample a position, creating a funnel.", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18045", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6479bae5c85977d9e87231cc8aa653b8995ced9101bdf0ea466cbe0fd5dcc0cd.jpg", + "image_caption": [ + "Camera Layout", + "Figure 5. Convergence basin analysis: Left: 3D Gaussian map from training views (Yellow) and visualisation of the test poses (Red) and target pose (Blue). Right: Convergence basin of our method. The green marks success, and the red marks failure." + ], + "image_footnote": [], + "bbox": [ + 80, + 111, + 243, + 297 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3c6933a24620117c00152aa1c63b4984b51f73bfee7a703aeee21bf173b14185.jpg", + "image_caption": [ + "Ours w/ depth" + ], + "image_footnote": [], + "bbox": [ + 251, + 111, + 356, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/398ea8a505dfd31b526338e5c8b68a567f418bb02301e9565b80a773c5ca9f35.jpg", + "image_caption": [ + "Ours w/o depth" + ], + "image_footnote": [], + "bbox": [ + 361, + 111, + 467, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f3a2e348ba530ca0faa18294db78c0359b0718b93ba9cb2d9224e85ba7f00766.jpg", + "image_caption": [ + "Hash Grid SDF" + ], + "image_footnote": [], + "bbox": [ + 251, + 214, + 356, + 295 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/499fec0da5cbf54cc205e566a99a11ddeb17e1db06d1464acbdb4ab2844c6289.jpg", + "image_caption": [ + "MLP SDF" + ], + "image_footnote": [], + "bbox": [ + 361, + 214, + 467, + 294 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/723f9e4c7b39e327f5ec2b8349c052b9a0887637fc93f1ec910a8a91e9c45ae4.jpg", + "image_caption": [ + "Figure 6. Monocular SLAM result on fr1/desk sequence: We show the reconstructed 3D Gaussian maps (Left) and novel view synthesis result (Right)." + ], + "image_footnote": [], + "bbox": [ + 83, + 364, + 356, + 563 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6b963a76d8f3a9f1202632b6815ae6c22899a26fa92e4b325fe42a2dedf021aa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 357, + 364, + 464, + 430 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9e01286b1746790e8df6077f6deeab8ece2b6bc4745c733801623e2e1bb2fde2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 357, + 431, + 464, + 500 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bd4c1275abc36e95390b0c0249c6720ceaa06f0c8a8feb95b1c5ff54438a3fca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 357, + 500, + 464, + 563 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "From the sampled position, given the RGB image of the target view, we perform camera pose optimisation for 1000 iterations. The optimisation is successful if it converges to within $1\\mathrm{cm}$ of the target view within the fixed iterations. We compare our Gaussian approach with Co-SLAM [39]'s network (Hash Grid SDF) and iMAP's [33] network with Co-SLAM's SDF loss for further geometric accuracy (MLP Neural SDF). We render the training views using a synthetic Replica dataset and create three sequences for testing (seq1, seq2 and seq3). The width of the square formed by the training view is $0.5\\mathrm{m}$ , and the test cameras are distributed with radii ranging from $0.2\\mathrm{m}$ to $1.2\\mathrm{m}$ , covering a larger area than the training view. When training the map, the three methods—Ours w/depth, Hash Grid SDF, and MLP SDF—use RGB-D images, whereas Ours w/o depth utilises only colour images. Fig. 5 shows the qualitative results and Table 6 reports the success rate. For both with and without depth for training, our method shows better convergence.", + "bbox": [ + 76, + 618, + 468, + 890 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/47e9dec8d1434bc8a0c86b8dfa8b7be232b9d911067a469086644d77097f97c5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 89, + 694, + 198 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/21a2ca9113fec668cf62842a26c986443f45def672bb057fc41d56349540e139.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 89, + 888, + 198 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b0cccaaa883c0586fd89440b919de4caf748659e1d7776d32578ecfe0d28b27f.jpg", + "image_caption": [ + "Figure 7. Self-captured Scenes: Challenging scenes and objects, for example, transparent glasses and crinkled texture of salad are captured by our monocular SLAM running live." + ], + "image_footnote": [], + "bbox": [ + 504, + 199, + 694, + 306 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6e64f8174f79873aef6837b1b9eed17c58347638bc2c7bbafcb157b8dad47274.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 199, + 888, + 306 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Unlike hashing and positional encoding which can lead to signal conflict, anisotropic Gaussians form a smooth gradient in 3D space, increasing the convergence basin. Further experimental details are available in supplementary ??", + "bbox": [ + 498, + 361, + 890, + 421 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Qualitative Results", + "text_level": 1, + "bbox": [ + 500, + 430, + 681, + 445 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We report both the 3D reconstruction of the SLAM dataset and self-captured sequences. In Fig. 6, we visualise the monocular SLAM reconstruction of fr1/desk. The placements of the Gaussians are geometrically sensible and are 3D coherent, and our rendering from the different viewpoints highlights the quality of our systems' novel view synthesis. In Fig. 7, we self-capture challenging scenes for monocular SLAM. By not explicitly modelling a surface, our system naturally handles transparent objects which is challenging for many other SLAM systems.", + "bbox": [ + 498, + 453, + 890, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 617, + 617, + 632 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have proposed the first SLAM method using 3D Gaussians as a SLAM representation. Via efficient volume rendering, our system significantly advances the fidelity and diversity of object materials a live SLAM system can capture. Our system achieves state-of-the-art performance across benchmarks for both monocular and RGB-D cases. Interesting directions for future research are the integration of loop closure for handling large-scale scenes and extraction of geometry such as surface normal as Gaussians do not explicitly represent the surface.", + "bbox": [ + 496, + 642, + 890, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 800, + 666, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Research presented in this paper has been supported by Dyson Technology Ltd. We are very grateful to Eric Dexheimer, Kirill Mazur, Xin Kong, Marwan Taher, Ignacio Alzugaray, Gwangbin Bae, Aalok Patwardhan, and members of the Dyson Robotics Lab for insightful discussions.", + "bbox": [ + 498, + 825, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18046", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] J. Czarnowski, T. Laidlow, R. Clark, and A. J. Davison. Deepfactors: Real-time probabilistic dense monocular SLAM. IEEE Robotics and Automation Letters (RAL), 5(2): 721-728, 2020.", + "[2] Angela Dai, Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Christian Theobalt. BundleFusion: Real-time Globally Consistent 3D Reconstruction using On-the-fly Surface Re-integration. ACM Transactions on Graphics (TOG), 36(3):24:1-24:18, 2017.", + "[3] Eric Dexheimer and Andrew J. Davison. Learning a Depth Covariance Function. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023.", + "[4] J. Engel, V. Koltun, and D. Cremers. Direct sparse odometry. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 2017.", + "[5] C. Forster, M. Pizzoli, and D. Scaramuzza. SVO: Fast Semi-Direct Monocular Visual Odometry. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2014.", + "[6] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenumxels: Radiance fields without neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022.", + "[7] Jiahui Huang, Shi-Sheng Huang, Haoxuan Song, and ShiMin Hu. Di-fusion: Online implicit 3d reconstruction with deep priors. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021.", + "[8] M. M. Johari, C. Carta, and F. Fleuret. ESLAM: Efficient dense slam system based on hybrid representation of signed distance fields. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023.", + "[9] M. Keller, D. Lefloch, M. Lambers, S. Izadi, T. Weyrich, and A. Kolb. Real-time 3D Reconstruction in Dynamic Scenes using Point-based Fusion. In Proc. of Joint 3DIM/3DPVT Conference (3DV), 2013.", + "[10] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (TOG), 2023.", + "[11] Leonid Keselman and Martial Hebert. Approximate differentiable rendering with algebraic surfaces. In Proceedings of the European Conference on Computer Vision (ECCV), 2022.", + "[12] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proceedings of the International Conference on Learning Representations (ICLR), 2015.", + "[13] Heng Li, Xiaodong Gu, Weihao Yuan, Luwei Yang, Zilong Dong, and Ping Tan. Dense rgb slam with neural implicit maps. In Proceedings of the International Conference on Learning Representations (ICLR), 2023.", + "[14] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. NeurIPS, 2020." + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. 3DV, 2024.", + "[16] J. McCormac, A. Handa, A. J. Davison, and S. Leutenegger. SemanticFusion: Dense 3D semantic mapping with convolutional neural networks. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2017.", + "[17] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Proceedings of the European Conference on Computer Vision (ECCV), 2020.", + "[18] N. J. Mitra, N. Gelfand, H. Pottmann, and L. J. Guibas. Registration of Point Cloud Data from a Geometric Optimization Perspective. In Proceedings of the Symposium on Geometry Processing, 2004.", + "[19] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (TOG), 2022.", + "[20] R. Mur-Artal and J. D. Tardós. ORB-SLAM2: An Open-Source SLAM System for Monocular, Stereo, and RGB-D Cameras. IEEE Transactions on Robotics (T-RO), 33(5): 1255–1262, 2017.", + "[21] R. Mur-Artal, J. M. M Montiel, and J. D. Tardós. ORBSLAM: a Versatile and Accurate Monocular SLAM System. IEEE Transactions on Robotics (T-RO), 31(5):1147-1163, 2015.", + "[22] R. A. Newcombe. Dense Visual SLAM. PhD thesis, Imperial College London, 2012.", + "[23] R. A. Newcombe, S. Izadi, O. Hilliges, D. Molyneaux, D. Kim, A. J. Davison, P. Kohli, J. Shotton, S. Hodges, and A. Fitzgibbon. KinectFusion: Real-Time Dense Surface Mapping and Tracking. In Proceedings of the International Symposium on Mixed and Augmented Reality (ISMAR), 2011.", + "[24] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020.", + "[25] M. Nießner, M. Zollhöfer, S. Izadi, and M. Stamminger. Real-time 3D Reconstruction at Scale using Voxel Hashing. In Proceedings of SIGGRAPH, 2013.", + "[26] Victor Adrian Prisacariu, Olaf Kähler, Ming-Ming Cheng, Carl Yuheng Ren, Julien P. C. Valentin, Philip H. S. Torr, Ian D. Reid, and David W. Murray. A framework for the volumetric integration of depth images. CoRR, abs/1410.0925, 2014.", + "[27] Erik Sandström, Yue Li, Luc Van Gool, and Martin R. Oswald. Point-slam: Dense neural point cloud-based slam. In Proceedings of the International Conference on Computer Vision (ICCV), 2023.", + "[28] Thomas Schöps, Torsten Sattler, and Marc Pollefeys. Surfelmeshing: Online surfel-based mesh reconstruction. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 2020." + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18047", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Thomas Schöps, Torsten Sattler, and Marc Pollefeys. Bad slam: Bundle adjusted direct rgb-d slam. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019.", + "[30] J. Solà, J. Deray, and D. Atchuthan. A micro Lie theory for state estimation in robotics. arXiv:1812.01537, 2018.", + "[31] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, Anton Clarkson, Mingfei Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Braies, Tyler Gillingham, Elias Mueggler, Luis Pesqueira, Manolis Savva, Dhruv Batra, Hauke M. Strasdat, Renzo De Nardi, Michael Goesele, Steven Lovegrove, and Richard Newcombe. The Replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019.", + "[32] J. Sturm, N. Engelhard, F. Endres, W. Burgard, and D. Cremers. A Benchmark for the Evaluation of RGB-D SLAM Systems. In Proceedings of the IEEE/RSJ Conference on Intelligent Robots and Systems (IROS), 2012.", + "[33] E. Sucar, S. Liu, J. Ortiz, and A. J. Davison. iMAP: Implicit mapping and positioning in real-time. In Proceedings of the International Conference on Computer Vision (ICCV), 2021.", + "[34] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022.", + "[35] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. Proceedings of the International Conference on Learning Representations (ICLR), 2024.", + "[36] Zachary Teed and Jia Deng. DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras. In Neural Information Processing Systems (NIPS), 2021.", + "[37] Emanuele Vespa, Nikolay Nikolov, Marius Grimm, Luigi Nardi, Paul HJ Kelly, and Stefan Leutenegger. Efficient octree-based volumetric SLAM supporting signed-distance and occupancy mapping. IEEE Robotics and Automation Letters (RAL), 2018.", + "[38] Angtian Wang, Peng Wang, Jian Sun, Adam Kortylewski, and Alan Yuille. Voge: a differentiable volume renderer using gaussian ellipsoids for analysis-by-synthesis. 2022.", + "[39] Hengyi Wang, Jingwen Wang, and Lourdes Agapito. Coslam: Joint coordinate and sparse parametric encodings for neural real-time slam. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023.", + "[40] T. Whelan, M. Kaess, H. Johannsson, M. F. Fallon, J. J. Leonard, and J. B. McDonald. Real-time large scale dense RGB-D SLAM with volumetric fusion. International Journal of Robotics Research (IJRR), 34(4-5):598-626, 2015.", + "[41] T. Whelan, S. Leutenegger, R. F. Salas-Moreno, B. Glocker, and A. J. Davison. ElasticFusion: Dense SLAM without a pose graph. In Proceedings of Robotics: Science and Systems (RSS), 2015.", + "[42] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering." + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024.", + "[43] Xingrui Yang, Hai Li, Hongjia Zhai, Yuhang Ming, Yuqian Liu, and Guofeng Zhang. Vox-fusion: Dense tracking and mapping with voxel-based neural implicit representation. In Proceedings of the International Symposium on Mixed and Augmented Reality (ISMAR), 2022.", + "[44] Zeyu Yang, Hongye Yang, Zijie Pan, Xiatian Zhu, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. Proceedings of the International Conference on Learning Representations (ICLR), 2024.", + "[45] Taoran Yi, Jiemin Fang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussian-dreamer: Fast generation from text to 3d gaussian splatting with point cloud priors. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024.", + "[46] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R. Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022.", + "[47] Zihan Zhu, Songyou Peng, Viktor Larsson, Zhaopeng Cui, Martin R Oswald, Andreas Geiger, and Marc Pollefeys. Nicer-slam: Neural implicit scene encoding for rgb slam. International Conference on 3D Vision (3DV), 2024.", + "[48] M. Zwicker, H. Pfister, J. van Baar, and M. Gross. Ewa splatting. IEEE Transactions on Visualization and Computer Graphics, 8(3):223-238, 2002." + ], + "bbox": [ + 501, + 92, + 890, + 501 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18048", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Gaussian Splatting SLAM/32db8396-043e-45f8-8764-d01308949b32_model.json b/2024/Gaussian Splatting SLAM/32db8396-043e-45f8-8764-d01308949b32_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2f313eb4513da913ae841d2a8d27109f0659c81f --- /dev/null +++ b/2024/Gaussian Splatting SLAM/32db8396-043e-45f8-8764-d01308949b32_model.json @@ -0,0 +1,2299 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.351, + 0.131, + 0.621, + 0.154 + ], + "angle": 0, + "content": "Gaussian Splatting SLAM" + }, + { + "type": "text", + "bbox": [ + 0.136, + 0.18, + 0.303, + 0.198 + ], + "angle": 0, + "content": "Hidenobu Matsuki\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.342, + 0.181, + 0.452, + 0.198 + ], + "angle": 0, + "content": "Riku Murai\\(^{2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.492, + 0.181, + 0.627, + 0.199 + ], + "angle": 0, + "content": "Paul H. J. Kelly" + }, + { + "type": "text", + "bbox": [ + 0.668, + 0.181, + 0.826, + 0.198 + ], + "angle": 0, + "content": "Andrew J. Davison1" + }, + { + "type": "text", + "bbox": [ + 0.206, + 0.206, + 0.764, + 0.242 + ], + "angle": 0, + "content": "\\(^{1}\\)Dyson Robotics Laboratory, Imperial College London \\(^{2}\\)Software Performance Optimisation Group, Imperial College London" + }, + { + "type": "text", + "bbox": [ + 0.21, + 0.245, + 0.756, + 0.26 + ], + "angle": 0, + "content": "{h.matsuki20, riku.murai15, p.kelly, a.davison}@imperial.ac.uk" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.295, + 0.197, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.198, + 0.296, + 0.486, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.296, + 0.775, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.776, + 0.296, + 0.89, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.463, + 0.893, + 0.546 + ], + "angle": 0, + "content": "Figure 1. From a single monocular camera, we reconstruct a high fidelity 3D scene live at 3fps. For every incoming RGB frame, 3D Gaussians are incrementally formed and optimised together with the camera poses. We show both the rasterised Gaussians (left) and Gaussians shaded to highlight the geometry (right). Notice the details and the complex material properties (e.g. transparency) captured. Thin structures such as wires are accurately represented by numerous small, elongated Gaussians, and transparent objects are effectively represented by placing the Gaussians along the rim. Our system significantly advances the fidelity a live monocular SLAM system can capture." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.584, + 0.314, + 0.6 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.617, + 0.473, + 0.875 + ], + "angle": 0, + "content": "We present the first application of 3D Gaussian Splating in monocular SLAM, the most fundamental but the hardest setup for Visual SLAM. Our method, which runs live at 3fps, utilises Gaussians as the only 3D representation, unifying the required representation for accurate, efficient tracking, mapping, and high-quality rendering. Designed for challenging monocular settings, our approach is seamlessly extendable to RGB-D SLAM when an external depth sensor is available. Several innovations are required to continuously reconstruct 3D scenes with high fidelity from a live camera. First, to move beyond the original 3DGS algorithm, which requires accurate poses from an offline Structure from Motion (SfM) system, we formulate camera tracking for 3DGS using direct optimisation against the 3D Gaussians, and show that this enables fast and robust tracking with a wide basin of convergence. Second, by utilising the explicit nature of the Gaussians, we introduce geometric verification" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.571, + 0.893, + 0.646 + ], + "angle": 0, + "content": "and regularisation to handle the ambiguities occurring in incremental 3D dense reconstruction. Finally, we introduce a full SLAM system which not only achieves state-of-the-art results in novel view synthesis and trajectory estimation but also reconstruction of tiny and even transparent objects." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.678, + 0.633, + 0.694 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.895, + 0.903 + ], + "angle": 0, + "content": "A long-term goal of online reconstruction with a single moving camera is near-photorealistic fidelity, which will surely allow new levels of performance in many areas of Spatial AI and robotics as well as opening up a whole range of new applications. While we increasingly see the benefit of applying powerful pre-trained priors to 3D reconstruction, a key avenue for progress is still the invention and development of core 3D representations with advantageous properties. Many \"layered\" SLAM methods exist which tackle the SLAM problem by integrating multiple different 3D representations or existing SLAM components; however, the most interesting advances are when a new unified dense representation can be used for all aspects" + }, + { + "type": "page_footnote", + "bbox": [ + 0.101, + 0.887, + 0.327, + 0.9 + ], + "angle": 0, + "content": "*Authors contributed equally to this work." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18039" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.137 + ], + "angle": 0, + "content": "of a system's operation: local representation of detail, large-scale geometric mapping and also camera tracking by direct alignment." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.143, + 0.468, + 0.564 + ], + "angle": 0, + "content": "In this paper, we present the first online visual SLAM system based solely on the 3D Gaussian Splatting (3DGS) representation [10] recently making a big impact in offline scene reconstruction. In 3DGS a scene is represented by a large number of Gaussian blobs with orientation, elongation, colour and opacity. Other previous world/map-centric scene representations used for visual SLAM include occupancy or Signed Distance Function (SDF) voxel grids [23]; meshes [28]; point or surfel clouds [9, 29]; and recently neural fields [33]. Each of these has disadvantages: grids use significant memory and have bounded resolution, and even if octrees or hashing allow more efficiency they cannot be flexibly warped for large corrections [25, 37]; meshes require difficult, irregular topology to fuse new information; surfel clouds are discontinuous and difficult to fuse and optimise; and neural fields require expensive per-pixel raycasting to render. We show that 3DGS has none of these weaknesses. As a SLAM representation, it is most similar to point and surfel clouds, and inherits their efficiency, locality and ability to be easily warped or modified. However, it also represents geometry in a smooth, continuously differentiable way: a dense cloud of Gaussians merge together and jointly define a continuous volumetric function. And crucially, the design of modern graphics cards means that a large number of Gaussians can be efficiently rendered via \"splatting\" rasterisation, up to 200fps at \\(1080\\mathrm{p}\\). This rapid, differentiable rendering is integral to the tracking and map optimisation loops in our system." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.57, + 0.468, + 0.9 + ], + "angle": 0, + "content": "The 3DGS representation has up until now only been used in offline systems for 3D reconstruction with known camera poses, and we present several innovations to enable online SLAM. We first derive the analytic Jacobian on Lie group of camera pose with respect to a 3D Gaussians map, and show that this can be seamlessly integrated into the existing differentiable rasterisation pipeline to enable camera poses to be optimised alongside scene geometry. Second, we introduce a novel Gaussian isotropic shape regularisation, to ensure geometric consistency, which we have found is important for incremental reconstruction. Third, we propose a novel Gaussian resource allocation and pruning method to keep the geometry clean and enable accurate camera tracking. Our experimental results demonstrate photorealistic online local scene reconstruction, as well as state-of-the-art camera trajectory estimation and mapping for larger scenes compared to other rendering-based SLAM methods. We further show the uniqueness of the Gaussian-based SLAM method such as an extremely large camera pose convergence basin, which can also be useful for map-based camera localisation. Our method works with only monocular input, one of the most challenging scenarios in" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.092, + 0.891, + 0.211 + ], + "angle": 0, + "content": "SLAM. To highlight the intrinsic capability of 3D Gaussian for camera localisation, our method does not use any pretrained monocular depth predictor or other existing tracking modules, but relies solely on RGB image inputs in line with the original 3DGS. Since this is one of the most challenging SLAM scenario, we also show our method can easily be extended to RGB-D SLAM when depth measurements are available." + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.214, + 0.819, + 0.227 + ], + "angle": 0, + "content": "In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.229, + 0.891, + 0.272 + ], + "angle": 0, + "content": "- The first near real-time SLAM system which works with a 3DGS as the only underlying scene representation, which can handle monocular only inputs." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.274, + 0.89, + 0.331 + ], + "angle": 0, + "content": "- Novel techniques within the SLAM framework, including the analytic Jacobian on Lie group for direct camera pose estimation, isotropic regularisation of the Gaussian shape, and geometric verification." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.334, + 0.89, + 0.377 + ], + "angle": 0, + "content": "- Extensive evaluations on a variety of datasets both for monocular and RGB-D settings, demonstrating competitive performance, particularly in real-world scenarios." + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.229, + 0.891, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.393, + 0.64, + 0.408 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.419, + 0.891, + 0.9 + ], + "angle": 0, + "content": "Dense SLAM: Dense visual SLAM focuses on reconstructing detailed 3D maps, unlike sparse SLAM methods which excel in pose estimation [4, 5, 21] but typically yield maps useful mainly for localisation. In contrast, dense SLAM creates interactive maps beneficial for broader applications, including AR and robotics. Dense SLAM methods are generally divided into two primary categories: Frame-centric and Map-centric. Frame-centric SLAM minimises photometric error across consecutive frames, jointly estimating per-frame depth and frame-to-frame camera motion. Frame-centric approaches [1, 36] are efficient, as individual frames host local rather than global geometry (e.g. depth maps), and are attractive for long-session SLAM, but if a dense global map is needed, it must be constructed on demand by assembling all of these parts which are not necessarily fully consistent. In contrast, Map-centric SLAM uses a unified 3D representation across the SLAM pipeline, enabling a compact and streamlined system. Compared to purely local frame-to-frame tracking, a map-centric approach leverages global information by tracking against the reconstructed 3D consistent map. Classical map-centric approaches often use voxel grids [2, 23, 26, 40] or points [9, 29, 41] as the underlying 3D representation. While voxels enable a fast look-up of features in 3D, the representation is expensive, and the fixed voxel resolution and distribution are problematic when the spatial characteristics of the environment are not known in advance. On the other hand, a point-based map representation, such as surfel clouds, enables adaptive changes in resolution and spatial distribution by dynamic allocation of point primitives in the 3D space. Such flexibility benefits online applications such as SLAM with deformation-based loop closure [29, 41]. However, op" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.947, + 0.518, + 0.956 + ], + "angle": 0, + "content": "18040" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "timising the representation to capture high fidelity is challenging due to the lack of correlation among the primitives. Recently, in addition to classical graphic primitives, neural network-based map representations are a promising alternative. iMAP [33] demonstrated the interesting properties of neural representation, such as sensible hole filling of unobserved geometry. Many recent approaches combine the classical and neural representations to capture finer details [8, 27, 46, 47]; however, the large amount of computation required for neural rendering makes the live operation of such systems challenging." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.266, + 0.473, + 0.904 + ], + "angle": 0, + "content": "Differentiable Rendering: The classical method for creating a 3D representation was to unproject 2D observations into 3D space and to fuse them via weighted averaging [16, 23]. Such an averaging scheme suffers from over-smooth representation and lacks the expressiveness to capture high-quality details. To capture a scene with photorealistic quality, differentiable volumetric rendering [24] has recently been popularised with Neural Radiance Fields (NeRF) [17]. Using a single Multi-Layer Perceptron (MLP) as a scene representation, NeRF performs volume rendering by marching along pixel rays, querying the MLP for opacity and colour. Since volume rendering is naturally differentiable, the MLP representation is optimised to minimise the rendering loss using multiview information to achieve high-quality novel view synthesis. The main weakness of NeRF is its training speed. Recent developments have introduced explicit volume structures such as multi-resolution voxel grids [6, 14, 34] or hash functions [19] to improve performance. Interestingly, these projects demonstrate that the main contributor to high-quality novel view synthesis is not the neural network but rather differentiable volumetric rendering, and that it is possible to avoid the use of an MLP and yet achieve comparable rendering quality to NeRF [6]. However, even in these systems, per-pixel ray marching remains a significant bottleneck for rendering speed. This issue is particularly critical in SLAM, where immediate interaction with the map is essential for tracking. In contrast to NeRF, 3DGS performs differentiable rasterisation. Similar to regular graphics rasterisations, by iterating over the primitives to be rasterised rather than marching along rays, 3DGS leverages the natural sparsity of a 3D scene and achieves a representation which is expressive to capture high-fidelity 3D scenes while offering significantly faster rendering. Several works have applied 3D Gaussians and differentiable rendering to static scene capture [11, 38], and in particular more recent works utilise 3DGS and demonstrate superior results in vision tasks such as dynamic scene capture [15, 42, 44] and 3D generation [35, 45]. Our method adopts a Map-centric approach, utilising 3D Gaussians as the only SLAM representation. Similar to surfel-based SLAM, we dynamically allocate the 3D Gaussians, enabling us to model an arbitrary spatial dis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.184 + ], + "angle": 0, + "content": "tribution in the scene. Unlike other methods such as Elastic-Fusion [41] and PointFusion [9], however, by using differentiable rasterisation, our SLAM system can capture high-fidelity scene details and represent challenging object properties by direct optimisation against information from every pixel." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.197, + 0.593, + 0.212 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.222, + 0.684, + 0.239 + ], + "angle": 0, + "content": "3.1. Gaussian Splatting" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.245, + 0.895, + 0.413 + ], + "angle": 0, + "content": "Our SLAM representation is 3DGS, mapping the scene with a set of anisotropic Gaussians \\(\\mathcal{G}\\). Each Gaussian \\(\\mathcal{G}^i\\) contains optical properties: colour \\(c^i\\) and opacity \\(\\alpha^i\\). For continuous 3D representation, the mean \\(\\mu_W^i\\) and covariance \\(\\Sigma_W^i\\), defined in the world coordinate, represent the Gaussian's position and its ellipsoidal shape. For simplicity and speed, in our work we omit the spherical harmonics representing view-dependent radiance. Since 3DGS uses volume rendering, explicit extraction of the surface is not required. Instead, by splatting and blending \\(\\mathcal{N}\\) Gaussians, a pixel colour \\(\\mathcal{C}_p\\) is synthesised:" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.425, + 0.892, + 0.468 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} _ {p} = \\sum_ {i \\in \\mathcal {N}} c _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.478, + 0.895, + 0.6 + ], + "angle": 0, + "content": "3DGS performs rasterisation, iterating over the Gaussians rather than marching along the camera rays, and hence, free spaces are ignored during rendering. During rasterisation, the contributions of \\(\\alpha\\) are decayed via a Gaussian function, based on the 2D Gaussian formed by splatting a 3D Gaussian. The 3D Gaussians \\(\\mathcal{N}(\\pmb{\\mu}_W, \\pmb{\\Sigma}_W)\\) in world coordinates are related to the 2D Gaussians \\(\\mathcal{N}(\\pmb{\\mu}_I, \\pmb{\\Sigma}_I)\\) on the image plane through a projective transformation:" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.61, + 0.892, + 0.63 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\mu} _ {I} = \\pi (\\boldsymbol {T} _ {C W} \\cdot \\boldsymbol {\\mu} _ {W}), \\boldsymbol {\\Sigma} _ {I} = \\mathbf {J} \\mathbf {W} \\boldsymbol {\\Sigma} _ {W} \\mathbf {W} ^ {T} \\mathbf {J} ^ {T}, (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.64, + 0.895, + 0.779 + ], + "angle": 0, + "content": "where \\(\\pi\\) is the projection operation and \\(T_{CW} \\in SE(3)\\) is the camera pose of the viewpoint. \\(\\mathbf{J}\\) is the Jacobian of the linear approximation of the projective transformation and \\(\\mathbf{W}\\) is the rotational component of \\(T_{CW}\\). This formulation enables the 3D Gaussians to be differentiable and the blending operation provides gradient flow to the Gaussians. Using first-order gradient descent [12], Gaussians gradually refines both their optic and geometric parameters to represent the captured scene with high fidelity." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.744, + 0.803 + ], + "angle": 0, + "content": "3.2. Camera Pose Optimisation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.894, + 0.903 + ], + "angle": 0, + "content": "To achieve accurate tracking, we typically require at least 50 iterations of gradient descent per frame. This requirement emphasises the necessity of a representation with computationally efficient view synthesis and gradient computation, making the choice of 3D representation a crucial part of designing a SLAM system." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18041" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.089, + 0.896, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.327, + 0.895, + 0.357 + ], + "angle": 0, + "content": "Figure 2. SLAM System Overview: Our SLAM system uses 3D Gaussians as the only representation, unifying all components of SLAM, including tracking, mapping, keyframe management, and novel view synthesis." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.382, + 0.47, + 0.457 + ], + "angle": 0, + "content": "In order to avoid the overhead of automatic differentiation, 3DGS implements rasterisation with CUDA with derivatives for all parameters calculated explicitly. Since rasterisation is performance critical, we similarly derive the camera Jacobians explicitly." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.457, + 0.47, + 0.518 + ], + "angle": 0, + "content": "To the best of our knowledge, we provide the first analytical Jacobian of \\( \\pmb{SE}(3) \\) camera pose with respect to the 3D Gaussians used in EWA splatting [48] and 3DGS. This opens up new applications of 3DGS beyond SLAM." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.518, + 0.47, + 0.595 + ], + "angle": 0, + "content": "We use Lie algebra to derive the minimal Jacobians, ensuring that the dimensionality of the Jacobians matches the degrees of freedom, eliminating any redundant computations. The terms of Eq. (2) are differentiable with respect to the camera pose \\( \\pmb{T}_{CW} \\); using the chain rule:" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.598, + 0.47, + 0.63 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial \\boldsymbol {\\mu} _ {I}}{\\partial \\boldsymbol {T} _ {C W}} = \\frac {\\partial \\boldsymbol {\\mu} _ {I}}{\\partial \\boldsymbol {\\mu} _ {C}} \\frac {\\mathcal {D} \\boldsymbol {\\mu} _ {C}}{\\mathcal {D} \\boldsymbol {T} _ {C W}}, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.632, + 0.469, + 0.665 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial \\boldsymbol {\\Sigma} _ {I}}{\\partial \\boldsymbol {T} _ {C W}} = \\frac {\\partial \\boldsymbol {\\Sigma} _ {I}}{\\partial \\mathbf {J}} \\frac {\\partial \\mathbf {J}}{\\partial \\boldsymbol {\\mu} _ {C}} \\frac {\\mathcal {D} \\boldsymbol {\\mu} _ {C}}{\\mathcal {D} \\boldsymbol {T} _ {C W}} + \\frac {\\partial \\boldsymbol {\\Sigma} _ {I}}{\\partial \\mathbf {W}} \\frac {\\mathcal {D} \\mathbf {W}}{\\mathcal {D} \\boldsymbol {T} _ {C W}}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.669, + 0.47, + 0.745 + ], + "angle": 0, + "content": "where \\( T_{CW} \\) represents the 3D position of Gaussian in the camera coordinate. We take the derivatives on the manifold to derive minimal parameterisation. Borrowing the notation from [30], let \\( T \\in SE(3) \\) and \\( \\tau \\in \\mathfrak{se}(3) \\). We define the partial derivative on the manifold as:" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.749, + 0.469, + 0.78 + ], + "angle": 0, + "content": "\\[\n\\frac {\\mathcal {D} f (\\boldsymbol {T})}{\\mathcal {D} \\boldsymbol {T}} \\triangleq \\lim _ {\\tau \\rightarrow 0} \\frac {\\operatorname {L o g} (f (\\operatorname {E x p} (\\tau) \\circ \\boldsymbol {T}) \\circ f (\\boldsymbol {T}) ^ {- 1})}{\\tau}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.785, + 0.47, + 0.83 + ], + "angle": 0, + "content": "where \\( \\circ \\) is a group composition, and Exp, Log are the exponential and logarithmic mappings between Lie algebra and Lie Group. With this, we derive the following:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.848, + 0.47, + 0.898 + ], + "angle": 0, + "content": "\\[\n\\frac {\\mathcal {D} \\boldsymbol {\\mu} _ {C}}{\\mathcal {D} \\boldsymbol {T} _ {C W}} = \\left[ \\begin{array}{l l} \\boldsymbol {I} & - \\boldsymbol {\\mu} _ {C} ^ {\\times} \\end{array} \\right], \\frac {\\mathcal {D} \\mathbf {W}}{\\mathcal {D} \\boldsymbol {T} _ {C W}} = \\left[ \\begin{array}{l l} \\mathbf {0} & - \\mathbf {W} _ {:, 1} ^ {\\times} \\\\ \\mathbf {0} & - \\mathbf {W} _ {:, 2} ^ {\\times} \\\\ \\mathbf {0} & - \\mathbf {W} _ {:, 3} ^ {\\times} \\end{array} \\right] ， \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.382, + 0.892, + 0.413 + ], + "angle": 0, + "content": "where \\(\\times\\) denotes the skew symmetric matrix of a 3D vector, and \\(\\mathbf{W}_{:,i}\\) refers to the \\(i\\)th column of the matrix." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.422, + 0.59, + 0.436 + ], + "angle": 0, + "content": "3.3. SLAM" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.446, + 0.892, + 0.505 + ], + "angle": 0, + "content": "In this section, we present details of full SLAM framework. The overview of the system is summarised in Fig. 2. Please refer to the supplementary material for the further parameter details." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.526, + 0.619, + 0.542 + ], + "angle": 0, + "content": "3.3.1 Tracking" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.551, + 0.892, + 0.596 + ], + "angle": 0, + "content": "In tracking only the current camera pose is optimised, without updates to the map representation. In the monocular case, we minimise the following photometric residual:" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.607, + 0.892, + 0.627 + ], + "angle": 0, + "content": "\\[\nE _ {p h o} = \\left\\| I (\\mathcal {G}, \\boldsymbol {T} _ {C W}) - \\bar {I} \\right\\| _ {1}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.637, + 0.892, + 0.667 + ], + "angle": 0, + "content": "where \\(I(\\mathcal{G},\\pmb{T}_{CW})\\) renders the Gaussians \\(\\mathcal{G}\\) from \\(T_{CW}\\), and \\(\\bar{I}\\) is an observed image." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.668, + 0.892, + 0.713 + ], + "angle": 0, + "content": "We further optimise affine brightness parameters for varying exposure. When depth observations are available, we define the geometric residual as:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.724, + 0.892, + 0.745 + ], + "angle": 0, + "content": "\\[\nE _ {g e o} = \\left\\| D (\\mathcal {G}, \\boldsymbol {T} _ {C W}) - \\bar {D} \\right\\| _ {1}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.755, + 0.892, + 0.832 + ], + "angle": 0, + "content": "where \\(D(\\mathcal{G},\\mathbf{T}_{CW})\\) is depth rasterisation and \\(\\bar{D}\\) is the observed depth. Rather than simply using the depth measurements to initialise the Gaussians, we minimise both photometric and geometric residuals: \\(\\lambda_{pho}E_{pho} + (1 - \\lambda_{pho})E_{geo}\\), where \\(\\lambda_{pho}\\) is a hyperparameter." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.832, + 0.892, + 0.862 + ], + "angle": 0, + "content": "As in Eq. (1), per-pixel depth is rasterised by alphabetending:" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.862, + 0.892, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} _ {p} = \\sum_ {i \\in \\mathcal {N}} z _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {9}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18042" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.138 + ], + "angle": 0, + "content": "where \\(z_{i}\\) is the distance to the mean \\(\\pmb{\\mu}_{W}\\) of Gaussian \\(i\\) along the camera ray. We derive analytical Jacobians for the camera pose optimisation in a similar manner to Eq. (3), (4)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.156, + 0.218, + 0.172 + ], + "angle": 0, + "content": "3.3.2 Keyframing" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.18, + 0.471, + 0.302 + ], + "angle": 0, + "content": "Since using all the images from a video stream to jointly optimise the Gaussians and camera poses online is infeasible, we maintain a small window \\(\\mathcal{W}_k\\) consisting of carefully selected keyframes based on inter-frame covisibility. Ideal keyframe management will select non-redundant keyframes observing the same area, spanning a wide baseline to provide better multiview constraints. The parameters are detailed in supplementary." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.32, + 0.473, + 0.516 + ], + "angle": 0, + "content": "Selection and Management Every tracked frame is checked for keyframe registration based on our simple yet effective criteria. We measure the可视ibility by measuring the intersection over union of the observed Gaussians between the current frame \\( i \\) and the last keyframe \\( j \\). If the可视ibility drops below a threshold, or if the relative translation \\( t_{ij} \\) is large with respect to the median depth, frame \\( i \\) is registered as a keyframe. For efficiency, we maintain only a small number of keyframes in the current window \\( \\mathcal{W}_k \\) following the keyframe management heuristics of DSO [4]. The main difference is that a keyframe is removed from the current window if the overlap coefficient with the latest keyframe drops below a threshold." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.535, + 0.47, + 0.673 + ], + "angle": 0, + "content": "Gaussian Covisibility An accurate estimate of covisibility simplifies keyframe selection and management. 3DGS respects visibility ordering since the 3D Gaussians are sorted along the camera ray. This property is desirable for covisibility estimation as occlusions are handled by design. A Gaussian is marked to be visible from a view if used in the rasterisation and if the ray's accumulated \\(\\alpha\\) has not yet reached 0.5. This enables our estimated covisibility to handle occlusions without requiring additional heuristics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.69, + 0.47, + 0.84 + ], + "angle": 0, + "content": "Gaussian Insertion and Pruning At every keyframe, new Gaussians are inserted into the scene to capture newly visible scene elements and to refine the fine details. When depth measurements are available, Gaussian means \\(\\mu_W\\) are initialised by back-projecting the depth. In the monocular case, we render the depth at the current frame. For pixels with depth estimates, \\(\\mu_W\\) are initialised around those depths with low variance; for pixels without the depth estimates, we initialise \\(\\mu_W\\) around the median depth of the rendered image with high variance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "In the monocular case, the positions of many newly inserted Gaussians are incorrect. While the majority will quickly vanish during optimisation as they violate multiview consistency, we further prune the excess Gaussians by" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.894, + 0.152 + ], + "angle": 0, + "content": "checking the visibility amongst the current window \\(\\mathcal{W}_k\\). If the Gaussians inserted within the last 3 keyframes are unobserved by at least 3 other frames, we prune them out as they are geometrically unstable." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.171, + 0.62, + 0.187 + ], + "angle": 0, + "content": "3.3.3 Mapping" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.194, + 0.895, + 0.376 + ], + "angle": 0, + "content": "The purpose of mapping is to maintain a coherent 3D structure and to optimise the newly inserted Gaussians. During mapping, the keyframes in \\(\\mathcal{W}_k\\) are used to reconstruct currently visible regions. Additionally, two random past keyframes \\(\\mathcal{W}_r\\) are selected per iteration to avoid forgetting the global map. Rasterisation of 3DGS imposes no constraint on the Gaussians along the viewing ray direction, even with a depth observation. This is not a problem when sufficient carefully selected viewpoints are provided (e.g. in the novel view synthesis case); however, in continuous SLAM this causes many artefacts, making tracking challenging. We therefore introduce an isotropic regularisation:" + }, + { + "type": "equation", + "bbox": [ + 0.609, + 0.386, + 0.892, + 0.43 + ], + "angle": 0, + "content": "\\[\nE _ {i s o} = \\sum_ {i = 1} ^ {| \\mathcal {G} |} \\| \\mathbf {s} _ {i} - \\tilde {\\mathbf {s}} _ {i} \\cdot \\mathbf {1} \\| _ {1} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.439, + 0.894, + 0.56 + ], + "angle": 0, + "content": "to penalise the scaling parameters \\(\\mathbf{s}_i\\) (i.e. stretch of the ellipsoid) by its difference to the mean \\(\\tilde{\\mathbf{s}}_i\\). As shown in Fig 3, this encourages sphericality, and avoids the problem of Gaussians which are highly elongated along the viewing direction creating artefacts. Let the union of the keyframes in the current window and the randomly selected one be \\(\\mathcal{W} = \\mathcal{W}_k \\cup \\mathcal{W}_r\\). For mapping, we solve the following problem:" + }, + { + "type": "equation", + "bbox": [ + 0.564, + 0.571, + 0.892, + 0.609 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\boldsymbol {T} _ {C W} ^ {k} \\in \\boldsymbol {S E} (3), \\mathcal {G}, \\forall k \\in \\mathcal {W}} \\sum_ {\\forall k \\in \\mathcal {W}} E _ {p h o} ^ {k} + \\lambda_ {i s o} E _ {i s o}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.618, + 0.892, + 0.649 + ], + "angle": 0, + "content": "If depth observations are available, as in tracking, geometric residuals Eq. (8) are added to the optimisation problem." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.662, + 0.618, + 0.677 + ], + "angle": 0, + "content": "4. Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.687, + 0.892, + 0.779 + ], + "angle": 0, + "content": "We conduct a comprehensive evaluation of our system across a range of both real and synthetic datasets. Additionally, we perform an ablation study to justify our design choices. Finally, we present qualitative results of our system operating live using a monocular camera, illustrating its practicality and high fidelity reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.691, + 0.804 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Datasets For our quantitative analysis, we evaluate our method on the TUM RGB-D dataset [32] (3 sequences) and the Replica dataset [31] (8 sequences), following the evaluation in [33]. For qualitative results, we use self-captured real-world sequences recorded by Intel Realsense d455. Since the Replica dataset is designed for RGB-D SLAM" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18043" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.467, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.211, + 0.28, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.211, + 0.468, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.303, + 0.468, + 0.401 + ], + "angle": 0, + "content": "Figure 3. Effect of isotropic regularisation: Top: Rendering close to a training view (looking at the keyboard). Bottom: Rendering 3D Gaussians far from the training views (view from a side of the keyboard) without (left) and with (right) the isotropic loss. When the photometric constraints are insufficient, the Gaussians tend to elongate along the viewing direction, creating artefacts in the novel views, and affecting the camera tracking." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.41, + 0.468, + 0.469 + ], + "angle": 0, + "content": "evaluation, it contains challenging purely rotational camera motions. We hence use the Replica dataset for RGB-D evaluation only. The TUM RGB-D dataset is used for both monocular and RGB-D evaluation." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.493, + 0.468, + 0.659 + ], + "angle": 0, + "content": "Implementation Details We run our SLAM on a desktop with Intel Core i9 12900K 3.50GHz and a single NVIDIA GeForce RTX 4090. We present results from our multiprocessing implementation aimed at real-time applications. For a fair comparison with other methods on Replica, we additionally report result for single-process implementation which performs more mapping iterations. As with 3DGS, time-critical rasterisation and gradient computation are implemented using CUDA. The rest of the SLAM pipeline is developed with PyTorch. Details of hyperparameters are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.682, + 0.468, + 0.831 + ], + "angle": 0, + "content": "Metrics For camera tracking accuracy, we report the Root Mean Square Error (RMSE) of the Absolute Trajectory Error (ATE) of the keyframes. To evaluate map quality, we report standard photometric rendering quality metrics (PSNR, SSIM and LPIPS) following the evaluation protocol used in [27]. To evaluate the map quality, on every fifth frame, rendering metrics are computed. We exclude the keyframes (training views). We report the average across three runs for all our evaluations. In the tables, the best result is in bold, and the second best is underlined." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Baseline Methods We primarily benchmark our SLAM method against other approaches that, like ours, do not have explicit loop closure. In monocular settings, we compare" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.289 + ], + "angle": 0, + "content": "with state-of-the-art classical and learning-based direct visual odometry (VO) methods. Specifically, we compare DSO [4], DepthCov [3], and DROID-SLAM [36] in VO configurations. These methods are selected based on their public reporting of results on the benchmark (TUM dataset) or the availability of their source code for getting the benchmark result. Since one of our focuses is the online scale estimation under monocular scale ambiguity, the method which uses ground truth poses for the system initialisation such as [13] is not considered for the comparison. In the RGB-D case, we compare against neural-implicit SLAM methods [7, 8, 27, 33, 39, 43, 46] which are also map-centric, rendering-based and do not perform loop closure." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.299, + 0.722, + 0.314 + ], + "angle": 0, + "content": "4.2. Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.322, + 0.892, + 0.442 + ], + "angle": 0, + "content": "Camera Tracking Accuracy Table 1 shows the tracking results on the TUM RGB-D dataset. In the monocular setting, our method surpasses other baselines without requiring any deep priors. Furthermore, our performance is comparable to systems which perform explicit loop closure. This clearly highlights that there still remains potential for enhancing the tracking of monocular SLAM by exploring fundamental SLAM representations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.443, + 0.892, + 0.638 + ], + "angle": 0, + "content": "Our RGB-D method shows better performance than any other baseline method. Notably, our system surpasses ORB-SLAM in the fr1 sequences, narrowing the gap between Map-centric SLAM and the state-of-the-art sparse frame-centric methods. Table 2 reports results on the synthetic Replica dataset. Our single-process implementation shows competitive performance and achieves the best result in 4 out of 8 sequences. Our multi-process implementation which performs fewer mapping iterations still performs comparably. In contrast to other methods, our system demonstrates higher performance on real-world data, by optimising the Gaussian positions to compensate for the sensor noise." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Novel View Rendering Table 5 summarises the novel view rendering performance of our method with RGB-D input. We consistently show the best performance across most sequences and is least second best. Our rendering FPS is hundreds of times faster than other methods, offering a significant advantage for applications which require real-time map interaction. While Point-SLAM is competitive, that method focuses on view synthesis rather than novel-view synthesis. Their view synthesis is conditional on the availability of depth due to the depth-guided ray-sampling, making novel-view synthesis challenging. On the other hand, our rasterisation-based approach does not require depth guidance and achieves efficient, high-quality, novel view synthesis. Fig. 4 provides a qualitative comparison of the rendering of ours and Point-SLAM (with depth guidance)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18044" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.468, + 0.322 + ], + "angle": 0, + "content": "
InputLoop-closureMethodfr1/deskfr2/xyzfr3/officeAvg.
Monocularw/oDSO [4]22.41.109.5011.0
DROID-VO [36]5.2010.77.307.73
DepthCov-VO [3]5.601.2068.825.2
Ours3.784.603.503.96
w/DROID-SLAM [36]1.800.502.801.70
ORB-SLAM2 [20]1.900.602.401.60
RGB-Dw/oiMAP [33]4.902.005.804.23
NICE-SLAM [46]4.266.193.874.77
DI-Fusion [7]4.402.005.804.07
Vox-Fusion [43]3.521.4926.0110.34
ESLAM [8]2.471.112.422.00
Co-SLAM [39]2.401.702.402.17
Point-SLAM [27]4.341.313.483.04
Ours1.501.441.491.47
w/BAD-SLAM [29]1.701.101.701.50
Kintinous [40]3.702.903.003.20
ORB-SLAM2 [20]1.600.401.001.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.324, + 0.47, + 0.407 + ], + "angle": 0, + "content": "Table 1. Camera tracking result on TUM for monocular and RGB-D. ATE RMSE in cm is reported. In both monocular and RGB-D cases, we achieve state-of-the-art performance. In particular, in the monocular case, not only do we outperform systems which use deep prior, but we achieve comparable performance with many of the RGB-D systems." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.413, + 0.466, + 0.498 + ], + "angle": 0, + "content": "
Methodr0r1r2o0o1o2o3o4Avg.
iMAP [33]3.122.542.311.691.033.994.051.932.58
NICE-SLAM0.971.311.070.881.001.061.101.131.07
Vox-Fusion [43]1.374.701.478.482.042.581.112.943.09
ESLAM [8]0.710.700.520.570.550.580.720.630.63
Point-SLAM [27]0.610.410.370.380.480.540.690.720.53
Ours0.440.320.310.440.520.230.172.250.58
Ours (sp)0.330.220.290.360.190.250.120.810.32
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.5, + 0.469, + 0.571 + ], + "angle": 0, + "content": "Table 2. Camera tracking result on Replica for RGB-D SLAM. ATE RMSE in cm is reported. We achieve best performance across most sequences. Here, Ours is our multi-process implementation and Ours (sp) is the single-process implementation which ensures a certain amount of mapping iteration similar to other works." + }, + { + "type": "table", + "bbox": [ + 0.111, + 0.575, + 0.436, + 0.653 + ], + "angle": 0, + "content": "
InputMethodfr1/deskfr2/xyzfr3/officeAvg.
Monow/o \\( E_{iso} \\)4.164.665.734.83
w/o kf selection13.24.368.658.73
Ours3.784.603.503.96
RGB-Dw/o \\( E_{geo} \\)2.390.624.982.66
w/o kf selection1.641.492.601.90
Ours1.501.441.491.47
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.655, + 0.469, + 0.711 + ], + "angle": 0, + "content": "Table 3. Ablation Study on TUM RGB-D dataset. We analyse the usefulness of isotropic regularisation, geometric residual, and keyframe selection to our SLAM system. Further isotropic regularisation ablation is available in supplementary." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.717, + 0.466, + 0.751 + ], + "angle": 0, + "content": "
Memory Usage [MB]
iMAP [33]NICE-SLAM [46]Co-SLAM [39]Ours (Mono)Ours (RGB-D)
0.8MB40.3.4MB6.4MB2.6MB3.97MB
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.752, + 0.468, + 0.781 + ], + "angle": 0, + "content": "Table 4. Memory Analysis on TUM RGB-D dataset. The baseline numbers are computed from the parameter numbers in [39]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.789, + 0.469, + 0.895 + ], + "angle": 0, + "content": "Ablative Analysis In Table 3, we perform ablation to confirm our design choices. Isotropic regularisation and geometric residual improve the tracking of monocular and RGB-D SLAM respectively, as they aid in constraining the geometry when photometric signals are weak. For both cases, keyframe selection significantly improves systems performance, as it automatically chooses suitable keyframes" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.089, + 0.889, + 0.148 + ], + "angle": 0, + "content": "
MethodPSNR[db]↑SSIM↑LPIPS↓Rendering FPS
NICE-SLAM[46]24.420.8090.2330.54
Vox-Fusion[43]24.410.8010.2362.17
Point-SLAM [27]35.170.9750.1241.33
ours38.940.9680.070769
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.15, + 0.892, + 0.219 + ], + "angle": 0, + "content": "Table 5. Average rendering performance on Replica (RGB-D). Our method outperforms most of the rendering metrics compared to existing methods. Note that Point-SLAM uses ground-truth depth to guide sampling along rays. The full detail is available in supplementary." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.228, + 0.89, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.481, + 0.891, + 0.509 + ], + "angle": 0, + "content": "Figure 4. Rendering examples on Replica. Point-SLAM struggle with rendering fine details due to the stochastic ray sampling." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.518, + 0.892, + 0.624 + ], + "angle": 0, + "content": "based on our occlusion-aware keyframe selection and management. We further compare the memory usage of different 3D representations in Table 4. MLP-based iMAP is clearly more memory efficient, but it struggles to express high-fidelity 3D scenes due to the limited capacity of small MLP. Compared with a voxel grid of features used in NICE-SLAM, our method uses significantly less memory." + }, + { + "type": "table", + "bbox": [ + 0.524, + 0.648, + 0.87, + 0.719 + ], + "angle": 0, + "content": "
Methodseq1seq2seq3Avg.
Neural SDF (Hash Grid)0.130.150.160.14
Neural SDF (MLP)0.400.380.220.33
Ours w/o depth0.820.910.650.79
Ours w/ depth0.831.00.650.82
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.722, + 0.892, + 0.763 + ], + "angle": 0, + "content": "Table 6. Camera convergence analysis. We report the ratio of successful camera convergence for the different sequences, across different differentiable 3D representations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Convergence Basin Analysis In our SLAM experiments, we discovered that 3D Gaussian maps have a notably large convergence basin for camera localisation. To investigate further, we conducted a convergence funnel analysis, an evaluation methodology proposed in [18] and used in [22]. Here, we train a 3D representation (e.g. 3DGS) using 9 fixed views arranged in a square. We set the viewpoint in the middle of the square to be the target view. As shown in Fig 5, we uniformly sample a position, creating a funnel." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "18045" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.115, + 0.094, + 0.208, + 0.107 + ], + "angle": 0, + "content": "Camera Layout" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.112, + 0.245, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.259, + 0.094, + 0.346, + 0.107 + ], + "angle": 0, + "content": "Ours w/ depth" + }, + { + "type": "image", + "bbox": [ + 0.252, + 0.112, + 0.357, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.365, + 0.094, + 0.459, + 0.107 + ], + "angle": 0, + "content": "Ours w/o depth" + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.112, + 0.468, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.259, + 0.199, + 0.35, + 0.21 + ], + "angle": 0, + "content": "Hash Grid SDF" + }, + { + "type": "image", + "bbox": [ + 0.252, + 0.215, + 0.357, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.383, + 0.199, + 0.442, + 0.21 + ], + "angle": 0, + "content": "MLP SDF" + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.215, + 0.468, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.301, + 0.47, + 0.357 + ], + "angle": 0, + "content": "Figure 5. Convergence basin analysis: Left: 3D Gaussian map from training views (Yellow) and visualisation of the test poses (Red) and target pose (Blue). Right: Convergence basin of our method. The green marks success, and the red marks failure." + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.366, + 0.357, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.366, + 0.465, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.432, + 0.465, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.501, + 0.465, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.568, + 0.469, + 0.61 + ], + "angle": 0, + "content": "Figure 6. Monocular SLAM result on fr1/desk sequence: We show the reconstructed 3D Gaussian maps (Left) and novel view synthesis result (Right)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.619, + 0.47, + 0.891 + ], + "angle": 0, + "content": "From the sampled position, given the RGB image of the target view, we perform camera pose optimisation for 1000 iterations. The optimisation is successful if it converges to within \\(1\\mathrm{cm}\\) of the target view within the fixed iterations. We compare our Gaussian approach with Co-SLAM [39]'s network (Hash Grid SDF) and iMAP's [33] network with Co-SLAM's SDF loss for further geometric accuracy (MLP Neural SDF). We render the training views using a synthetic Replica dataset and create three sequences for testing (seq1, seq2 and seq3). The width of the square formed by the training view is \\(0.5\\mathrm{m}\\), and the test cameras are distributed with radii ranging from \\(0.2\\mathrm{m}\\) to \\(1.2\\mathrm{m}\\), covering a larger area than the training view. When training the map, the three methods—Ours w/depth, Hash Grid SDF, and MLP SDF—use RGB-D images, whereas Ours w/o depth utilises only colour images. Fig. 5 shows the qualitative results and Table 6 reports the success rate. For both with and without depth for training, our method shows better convergence." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.09, + 0.696, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.09, + 0.889, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.2, + 0.696, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.2, + 0.889, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.311, + 0.892, + 0.353 + ], + "angle": 0, + "content": "Figure 7. Self-captured Scenes: Challenging scenes and objects, for example, transparent glasses and crinkled texture of salad are captured by our monocular SLAM running live." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.362, + 0.892, + 0.422 + ], + "angle": 0, + "content": "Unlike hashing and positional encoding which can lead to signal conflict, anisotropic Gaussians form a smooth gradient in 3D space, increasing the convergence basin. Further experimental details are available in supplementary ??" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.431, + 0.682, + 0.446 + ], + "angle": 0, + "content": "4.3. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.454, + 0.892, + 0.605 + ], + "angle": 0, + "content": "We report both the 3D reconstruction of the SLAM dataset and self-captured sequences. In Fig. 6, we visualise the monocular SLAM reconstruction of fr1/desk. The placements of the Gaussians are geometrically sensible and are 3D coherent, and our rendering from the different viewpoints highlights the quality of our systems' novel view synthesis. In Fig. 7, we self-capture challenging scenes for monocular SLAM. By not explicitly modelling a surface, our system naturally handles transparent objects which is challenging for many other SLAM systems." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.618, + 0.618, + 0.633 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.892, + 0.794 + ], + "angle": 0, + "content": "We have proposed the first SLAM method using 3D Gaussians as a SLAM representation. Via efficient volume rendering, our system significantly advances the fidelity and diversity of object materials a live SLAM system can capture. Our system achieves state-of-the-art performance across benchmarks for both monocular and RGB-D cases. Interesting directions for future research are the integration of loop closure for handling large-scale scenes and extraction of geometry such as surface normal as Gaussians do not explicitly represent the surface." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.801, + 0.667, + 0.817 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Research presented in this paper has been supported by Dyson Technology Ltd. We are very grateful to Eric Dexheimer, Kirill Mazur, Xin Kong, Marwan Taher, Ignacio Alzugaray, Gwangbin Bae, Aalok Patwardhan, and members of the Dyson Robotics Lab for insightful discussions." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18046" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] J. Czarnowski, T. Laidlow, R. Clark, and A. J. Davison. Deepfactors: Real-time probabilistic dense monocular SLAM. IEEE Robotics and Automation Letters (RAL), 5(2): 721-728, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.472, + 0.243 + ], + "angle": 0, + "content": "[2] Angela Dai, Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Christian Theobalt. BundleFusion: Real-time Globally Consistent 3D Reconstruction using On-the-fly Surface Re-integration. ACM Transactions on Graphics (TOG), 36(3):24:1-24:18, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.47, + 0.298 + ], + "angle": 0, + "content": "[3] Eric Dexheimer and Andrew J. Davison. Learning a Depth Covariance Function. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.301, + 0.47, + 0.342 + ], + "angle": 0, + "content": "[4] J. Engel, V. Koltun, and D. Cremers. Direct sparse odometry. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.344, + 0.47, + 0.399 + ], + "angle": 0, + "content": "[5] C. Forster, M. Pizzoli, and D. Scaramuzza. SVO: Fast Semi-Direct Monocular Visual Odometry. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.402, + 0.47, + 0.47 + ], + "angle": 0, + "content": "[6] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenumxels: Radiance fields without neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.472, + 0.469, + 0.528 + ], + "angle": 0, + "content": "[7] Jiahui Huang, Shi-Sheng Huang, Haoxuan Song, and ShiMin Hu. Di-fusion: Online implicit 3d reconstruction with deep priors. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.53, + 0.47, + 0.586 + ], + "angle": 0, + "content": "[8] M. M. Johari, C. Carta, and F. Fleuret. ESLAM: Efficient dense slam system based on hybrid representation of signed distance fields. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.588, + 0.47, + 0.642 + ], + "angle": 0, + "content": "[9] M. Keller, D. Lefloch, M. Lambers, S. Izadi, T. Weyrich, and A. Kolb. Real-time 3D Reconstruction in Dynamic Scenes using Point-based Fusion. In Proc. of Joint 3DIM/3DPVT Conference (3DV), 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.645, + 0.469, + 0.699 + ], + "angle": 0, + "content": "[10] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (TOG), 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.702, + 0.469, + 0.755 + ], + "angle": 0, + "content": "[11] Leonid Keselman and Martial Hebert. Approximate differentiable rendering with algebraic surfaces. In Proceedings of the European Conference on Computer Vision (ECCV), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.758, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[12] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proceedings of the International Conference on Learning Representations (ICLR), 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[13] Heng Li, Xiaodong Gu, Weihao Yuan, Luwei Yang, Zilong Dong, and Ping Tan. Dense rgb slam with neural implicit maps. In Proceedings of the International Conference on Learning Representations (ICLR), 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[14] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. NeurIPS, 2020." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[15] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. 3DV, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[16] J. McCormac, A. Handa, A. J. Davison, and S. Leutenegger. SemanticFusion: Dense 3D semantic mapping with convolutional neural networks. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[17] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Proceedings of the European Conference on Computer Vision (ECCV), 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[18] N. J. Mitra, N. Gelfand, H. Pottmann, and L. J. Guibas. Registration of Point Cloud Data from a Geometric Optimization Perspective. In Proceedings of the Symposium on Geometry Processing, 2004." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[19] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (TOG), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[20] R. Mur-Artal and J. D. Tardós. ORB-SLAM2: An Open-Source SLAM System for Monocular, Stereo, and RGB-D Cameras. IEEE Transactions on Robotics (T-RO), 33(5): 1255–1262, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.448, + 0.892, + 0.502 + ], + "angle": 0, + "content": "[21] R. Mur-Artal, J. M. M Montiel, and J. D. Tardós. ORBSLAM: a Versatile and Accurate Monocular SLAM System. IEEE Transactions on Robotics (T-RO), 31(5):1147-1163, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.892, + 0.532 + ], + "angle": 0, + "content": "[22] R. A. Newcombe. Dense Visual SLAM. PhD thesis, Imperial College London, 2012." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.534, + 0.892, + 0.603 + ], + "angle": 0, + "content": "[23] R. A. Newcombe, S. Izadi, O. Hilliges, D. Molyneaux, D. Kim, A. J. Davison, P. Kohli, J. Shotton, S. Hodges, and A. Fitzgibbon. KinectFusion: Real-Time Dense Surface Mapping and Tracking. In Proceedings of the International Symposium on Mixed and Augmented Reality (ISMAR), 2011." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.605, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[24] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.676, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[25] M. Nießner, M. Zollhöfer, S. Izadi, and M. Stamminger. Real-time 3D Reconstruction at Scale using Voxel Hashing. In Proceedings of SIGGRAPH, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.718, + 0.892, + 0.786 + ], + "angle": 0, + "content": "[26] Victor Adrian Prisacariu, Olaf Kähler, Ming-Ming Cheng, Carl Yuheng Ren, Julien P. C. Valentin, Philip H. S. Torr, Ian D. Reid, and David W. Murray. A framework for the volumetric integration of depth images. CoRR, abs/1410.0925, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[27] Erik Sandström, Yue Li, Luc Van Gool, and Martin R. Oswald. Point-slam: Dense neural point cloud-based slam. In Proceedings of the International Conference on Computer Vision (ICCV), 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[28] Thomas Schöps, Torsten Sattler, and Marc Pollefeys. Surfelmeshing: Online surfel-based mesh reconstruction. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 2020." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18047" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[29] Thomas Schöps, Torsten Sattler, and Marc Pollefeys. Bad slam: Bundle adjusted direct rgb-d slam. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[30] J. Solà, J. Deray, and D. Atchuthan. A micro Lie theory for state estimation in robotics. arXiv:1812.01537, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.468, + 0.317 + ], + "angle": 0, + "content": "[31] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, Anton Clarkson, Mingfei Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Braies, Tyler Gillingham, Elias Mueggler, Luis Pesqueira, Manolis Savva, Dhruv Batra, Hauke M. Strasdat, Renzo De Nardi, Michael Goesele, Steven Lovegrove, and Richard Newcombe. The Replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.319, + 0.468, + 0.374 + ], + "angle": 0, + "content": "[32] J. Sturm, N. Engelhard, F. Endres, W. Burgard, and D. Cremers. A Benchmark for the Evaluation of RGB-D SLAM Systems. In Proceedings of the IEEE/RSJ Conference on Intelligent Robots and Systems (IROS), 2012." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.375, + 0.468, + 0.417 + ], + "angle": 0, + "content": "[33] E. Sucar, S. Liu, J. Ortiz, and A. J. Davison. iMAP: Implicit mapping and positioning in real-time. In Proceedings of the International Conference on Computer Vision (ICCV), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.418, + 0.468, + 0.473 + ], + "angle": 0, + "content": "[34] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.476, + 0.468, + 0.531 + ], + "angle": 0, + "content": "[35] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. Proceedings of the International Conference on Learning Representations (ICLR), 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.532, + 0.468, + 0.573 + ], + "angle": 0, + "content": "[36] Zachary Teed and Jia Deng. DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras. In Neural Information Processing Systems (NIPS), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.575, + 0.468, + 0.644 + ], + "angle": 0, + "content": "[37] Emanuele Vespa, Nikolay Nikolov, Marius Grimm, Luigi Nardi, Paul HJ Kelly, and Stefan Leutenegger. Efficient octree-based volumetric SLAM supporting signed-distance and occupancy mapping. IEEE Robotics and Automation Letters (RAL), 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.646, + 0.468, + 0.687 + ], + "angle": 0, + "content": "[38] Angtian Wang, Peng Wang, Jian Sun, Adam Kortylewski, and Alan Yuille. Voge: a differentiable volume renderer using gaussian ellipsoids for analysis-by-synthesis. 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.468, + 0.743 + ], + "angle": 0, + "content": "[39] Hengyi Wang, Jingwen Wang, and Lourdes Agapito. Coslam: Joint coordinate and sparse parametric encodings for neural real-time slam. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.745, + 0.468, + 0.8 + ], + "angle": 0, + "content": "[40] T. Whelan, M. Kaess, H. Johannsson, M. F. Fallon, J. J. Leonard, and J. B. McDonald. Real-time large scale dense RGB-D SLAM with volumetric fusion. International Journal of Robotics Research (IJRR), 34(4-5):598-626, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.468, + 0.857 + ], + "angle": 0, + "content": "[41] T. Whelan, S. Leutenegger, R. F. Salas-Moreno, B. Glocker, and A. J. Davison. ElasticFusion: Dense SLAM without a pose graph. In Proceedings of Robotics: Science and Systems (RSS), 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[42] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[43] Xingrui Yang, Hai Li, Hongjia Zhai, Yuhang Ming, Yuqian Liu, and Guofeng Zhang. Vox-fusion: Dense tracking and mapping with voxel-based neural implicit representation. In Proceedings of the International Symposium on Mixed and Augmented Reality (ISMAR), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[44] Zeyu Yang, Hongye Yang, Zijie Pan, Xiatian Zhu, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. Proceedings of the International Conference on Learning Representations (ICLR), 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.262, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[45] Taoran Yi, Jiemin Fang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussian-dreamer: Fast generation from text to 3d gaussian splatting with point cloud priors. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.333, + 0.892, + 0.402 + ], + "angle": 0, + "content": "[46] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R. Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.403, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[47] Zihan Zhu, Songyou Peng, Viktor Larsson, Zhaopeng Cui, Martin R Oswald, Andreas Geiger, and Marc Pollefeys. Nicer-slam: Neural implicit scene encoding for rgb slam. International Conference on 3D Vision (3DV), 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.46, + 0.892, + 0.502 + ], + "angle": 0, + "content": "[48] M. Zwicker, H. Pfister, J. van Baar, and M. Gross. Ewa splatting. IEEE Transactions on Visualization and Computer Graphics, 8(3):223-238, 2002." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18048" + } + ] +] \ No newline at end of file diff --git a/2024/Gaussian Splatting SLAM/32db8396-043e-45f8-8764-d01308949b32_origin.pdf b/2024/Gaussian Splatting SLAM/32db8396-043e-45f8-8764-d01308949b32_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ac6a3d81a7a6fe4c13ba977d33ae039efc452c34 --- /dev/null +++ b/2024/Gaussian Splatting SLAM/32db8396-043e-45f8-8764-d01308949b32_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8a41c6975857d80aa3b58993524ecd3614e18d4b35e56fdf44f29edbd00deee2 +size 9924048 diff --git a/2024/Gaussian Splatting SLAM/full.md b/2024/Gaussian Splatting SLAM/full.md new file mode 100644 index 0000000000000000000000000000000000000000..996cc5d67bbe3ca9a81dc7a2d501a8bf8defe1bc --- /dev/null +++ b/2024/Gaussian Splatting SLAM/full.md @@ -0,0 +1,340 @@ +# Gaussian Splatting SLAM + +Hidenobu Matsuki $^{1*}$ + +Riku Murai $^{2*}$ + +Paul H. J. Kelly + +Andrew J. Davison1 + +$^{1}$ Dyson Robotics Laboratory, Imperial College London $^{2}$ Software Performance Optimisation Group, Imperial College London + +{h.matsuki20, riku.murai15, p.kelly, a.davison}@imperial.ac.uk + +![](images/c2f83a8fbdb7d189748a19aee544271d2f9f30f06468e5043b49446e3ad4f1e9.jpg) +Figure 1. From a single monocular camera, we reconstruct a high fidelity 3D scene live at 3fps. For every incoming RGB frame, 3D Gaussians are incrementally formed and optimised together with the camera poses. We show both the rasterised Gaussians (left) and Gaussians shaded to highlight the geometry (right). Notice the details and the complex material properties (e.g. transparency) captured. Thin structures such as wires are accurately represented by numerous small, elongated Gaussians, and transparent objects are effectively represented by placing the Gaussians along the rim. Our system significantly advances the fidelity a live monocular SLAM system can capture. + +![](images/0cf846d57eab8d874d336abf74b7dbdc1f6ca25d48c9d112804de3a177a69eb4.jpg) + +![](images/de58f0539762e5fe3ede81e0b04853165ffba86077c8ef7af1cf3f36874bb722.jpg) + +![](images/929b63922c0a7c841cf1939d4afee510145b830db8bcce80ff9e2c27b03447d4.jpg) + +# Abstract + +We present the first application of 3D Gaussian Splating in monocular SLAM, the most fundamental but the hardest setup for Visual SLAM. Our method, which runs live at 3fps, utilises Gaussians as the only 3D representation, unifying the required representation for accurate, efficient tracking, mapping, and high-quality rendering. Designed for challenging monocular settings, our approach is seamlessly extendable to RGB-D SLAM when an external depth sensor is available. Several innovations are required to continuously reconstruct 3D scenes with high fidelity from a live camera. First, to move beyond the original 3DGS algorithm, which requires accurate poses from an offline Structure from Motion (SfM) system, we formulate camera tracking for 3DGS using direct optimisation against the 3D Gaussians, and show that this enables fast and robust tracking with a wide basin of convergence. Second, by utilising the explicit nature of the Gaussians, we introduce geometric verification + +and regularisation to handle the ambiguities occurring in incremental 3D dense reconstruction. Finally, we introduce a full SLAM system which not only achieves state-of-the-art results in novel view synthesis and trajectory estimation but also reconstruction of tiny and even transparent objects. + +# 1. Introduction + +A long-term goal of online reconstruction with a single moving camera is near-photorealistic fidelity, which will surely allow new levels of performance in many areas of Spatial AI and robotics as well as opening up a whole range of new applications. While we increasingly see the benefit of applying powerful pre-trained priors to 3D reconstruction, a key avenue for progress is still the invention and development of core 3D representations with advantageous properties. Many "layered" SLAM methods exist which tackle the SLAM problem by integrating multiple different 3D representations or existing SLAM components; however, the most interesting advances are when a new unified dense representation can be used for all aspects + +of a system's operation: local representation of detail, large-scale geometric mapping and also camera tracking by direct alignment. + +In this paper, we present the first online visual SLAM system based solely on the 3D Gaussian Splatting (3DGS) representation [10] recently making a big impact in offline scene reconstruction. In 3DGS a scene is represented by a large number of Gaussian blobs with orientation, elongation, colour and opacity. Other previous world/map-centric scene representations used for visual SLAM include occupancy or Signed Distance Function (SDF) voxel grids [23]; meshes [28]; point or surfel clouds [9, 29]; and recently neural fields [33]. Each of these has disadvantages: grids use significant memory and have bounded resolution, and even if octrees or hashing allow more efficiency they cannot be flexibly warped for large corrections [25, 37]; meshes require difficult, irregular topology to fuse new information; surfel clouds are discontinuous and difficult to fuse and optimise; and neural fields require expensive per-pixel raycasting to render. We show that 3DGS has none of these weaknesses. As a SLAM representation, it is most similar to point and surfel clouds, and inherits their efficiency, locality and ability to be easily warped or modified. However, it also represents geometry in a smooth, continuously differentiable way: a dense cloud of Gaussians merge together and jointly define a continuous volumetric function. And crucially, the design of modern graphics cards means that a large number of Gaussians can be efficiently rendered via "splatting" rasterisation, up to 200fps at $1080\mathrm{p}$ . This rapid, differentiable rendering is integral to the tracking and map optimisation loops in our system. + +The 3DGS representation has up until now only been used in offline systems for 3D reconstruction with known camera poses, and we present several innovations to enable online SLAM. We first derive the analytic Jacobian on Lie group of camera pose with respect to a 3D Gaussians map, and show that this can be seamlessly integrated into the existing differentiable rasterisation pipeline to enable camera poses to be optimised alongside scene geometry. Second, we introduce a novel Gaussian isotropic shape regularisation, to ensure geometric consistency, which we have found is important for incremental reconstruction. Third, we propose a novel Gaussian resource allocation and pruning method to keep the geometry clean and enable accurate camera tracking. Our experimental results demonstrate photorealistic online local scene reconstruction, as well as state-of-the-art camera trajectory estimation and mapping for larger scenes compared to other rendering-based SLAM methods. We further show the uniqueness of the Gaussian-based SLAM method such as an extremely large camera pose convergence basin, which can also be useful for map-based camera localisation. Our method works with only monocular input, one of the most challenging scenarios in + +SLAM. To highlight the intrinsic capability of 3D Gaussian for camera localisation, our method does not use any pretrained monocular depth predictor or other existing tracking modules, but relies solely on RGB image inputs in line with the original 3DGS. Since this is one of the most challenging SLAM scenario, we also show our method can easily be extended to RGB-D SLAM when depth measurements are available. + +In summary, our contributions are as follows: + +- The first near real-time SLAM system which works with a 3DGS as the only underlying scene representation, which can handle monocular only inputs. +- Novel techniques within the SLAM framework, including the analytic Jacobian on Lie group for direct camera pose estimation, isotropic regularisation of the Gaussian shape, and geometric verification. +- Extensive evaluations on a variety of datasets both for monocular and RGB-D settings, demonstrating competitive performance, particularly in real-world scenarios. + +# 2. Related Work + +Dense SLAM: Dense visual SLAM focuses on reconstructing detailed 3D maps, unlike sparse SLAM methods which excel in pose estimation [4, 5, 21] but typically yield maps useful mainly for localisation. In contrast, dense SLAM creates interactive maps beneficial for broader applications, including AR and robotics. Dense SLAM methods are generally divided into two primary categories: Frame-centric and Map-centric. Frame-centric SLAM minimises photometric error across consecutive frames, jointly estimating per-frame depth and frame-to-frame camera motion. Frame-centric approaches [1, 36] are efficient, as individual frames host local rather than global geometry (e.g. depth maps), and are attractive for long-session SLAM, but if a dense global map is needed, it must be constructed on demand by assembling all of these parts which are not necessarily fully consistent. In contrast, Map-centric SLAM uses a unified 3D representation across the SLAM pipeline, enabling a compact and streamlined system. Compared to purely local frame-to-frame tracking, a map-centric approach leverages global information by tracking against the reconstructed 3D consistent map. Classical map-centric approaches often use voxel grids [2, 23, 26, 40] or points [9, 29, 41] as the underlying 3D representation. While voxels enable a fast look-up of features in 3D, the representation is expensive, and the fixed voxel resolution and distribution are problematic when the spatial characteristics of the environment are not known in advance. On the other hand, a point-based map representation, such as surfel clouds, enables adaptive changes in resolution and spatial distribution by dynamic allocation of point primitives in the 3D space. Such flexibility benefits online applications such as SLAM with deformation-based loop closure [29, 41]. However, op + +timising the representation to capture high fidelity is challenging due to the lack of correlation among the primitives. Recently, in addition to classical graphic primitives, neural network-based map representations are a promising alternative. iMAP [33] demonstrated the interesting properties of neural representation, such as sensible hole filling of unobserved geometry. Many recent approaches combine the classical and neural representations to capture finer details [8, 27, 46, 47]; however, the large amount of computation required for neural rendering makes the live operation of such systems challenging. + +Differentiable Rendering: The classical method for creating a 3D representation was to unproject 2D observations into 3D space and to fuse them via weighted averaging [16, 23]. Such an averaging scheme suffers from over-smooth representation and lacks the expressiveness to capture high-quality details. To capture a scene with photorealistic quality, differentiable volumetric rendering [24] has recently been popularised with Neural Radiance Fields (NeRF) [17]. Using a single Multi-Layer Perceptron (MLP) as a scene representation, NeRF performs volume rendering by marching along pixel rays, querying the MLP for opacity and colour. Since volume rendering is naturally differentiable, the MLP representation is optimised to minimise the rendering loss using multiview information to achieve high-quality novel view synthesis. The main weakness of NeRF is its training speed. Recent developments have introduced explicit volume structures such as multi-resolution voxel grids [6, 14, 34] or hash functions [19] to improve performance. Interestingly, these projects demonstrate that the main contributor to high-quality novel view synthesis is not the neural network but rather differentiable volumetric rendering, and that it is possible to avoid the use of an MLP and yet achieve comparable rendering quality to NeRF [6]. However, even in these systems, per-pixel ray marching remains a significant bottleneck for rendering speed. This issue is particularly critical in SLAM, where immediate interaction with the map is essential for tracking. In contrast to NeRF, 3DGS performs differentiable rasterisation. Similar to regular graphics rasterisations, by iterating over the primitives to be rasterised rather than marching along rays, 3DGS leverages the natural sparsity of a 3D scene and achieves a representation which is expressive to capture high-fidelity 3D scenes while offering significantly faster rendering. Several works have applied 3D Gaussians and differentiable rendering to static scene capture [11, 38], and in particular more recent works utilise 3DGS and demonstrate superior results in vision tasks such as dynamic scene capture [15, 42, 44] and 3D generation [35, 45]. Our method adopts a Map-centric approach, utilising 3D Gaussians as the only SLAM representation. Similar to surfel-based SLAM, we dynamically allocate the 3D Gaussians, enabling us to model an arbitrary spatial dis + +tribution in the scene. Unlike other methods such as Elastic-Fusion [41] and PointFusion [9], however, by using differentiable rasterisation, our SLAM system can capture high-fidelity scene details and represent challenging object properties by direct optimisation against information from every pixel. + +# 3. Method + +# 3.1. Gaussian Splatting + +Our SLAM representation is 3DGS, mapping the scene with a set of anisotropic Gaussians $\mathcal{G}$ . Each Gaussian $\mathcal{G}^i$ contains optical properties: colour $c^i$ and opacity $\alpha^i$ . For continuous 3D representation, the mean $\mu_W^i$ and covariance $\Sigma_W^i$ , defined in the world coordinate, represent the Gaussian's position and its ellipsoidal shape. For simplicity and speed, in our work we omit the spherical harmonics representing view-dependent radiance. Since 3DGS uses volume rendering, explicit extraction of the surface is not required. Instead, by splatting and blending $\mathcal{N}$ Gaussians, a pixel colour $\mathcal{C}_p$ is synthesised: + +$$ +\mathcal {C} _ {p} = \sum_ {i \in \mathcal {N}} c _ {i} \alpha_ {i} \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {j}). \tag {1} +$$ + +3DGS performs rasterisation, iterating over the Gaussians rather than marching along the camera rays, and hence, free spaces are ignored during rendering. During rasterisation, the contributions of $\alpha$ are decayed via a Gaussian function, based on the 2D Gaussian formed by splatting a 3D Gaussian. The 3D Gaussians $\mathcal{N}(\pmb{\mu}_W, \pmb{\Sigma}_W)$ in world coordinates are related to the 2D Gaussians $\mathcal{N}(\pmb{\mu}_I, \pmb{\Sigma}_I)$ on the image plane through a projective transformation: + +$$ +\boldsymbol {\mu} _ {I} = \pi (\boldsymbol {T} _ {C W} \cdot \boldsymbol {\mu} _ {W}), \boldsymbol {\Sigma} _ {I} = \mathbf {J} \mathbf {W} \boldsymbol {\Sigma} _ {W} \mathbf {W} ^ {T} \mathbf {J} ^ {T}, (2) +$$ + +where $\pi$ is the projection operation and $T_{CW} \in SE(3)$ is the camera pose of the viewpoint. $\mathbf{J}$ is the Jacobian of the linear approximation of the projective transformation and $\mathbf{W}$ is the rotational component of $T_{CW}$ . This formulation enables the 3D Gaussians to be differentiable and the blending operation provides gradient flow to the Gaussians. Using first-order gradient descent [12], Gaussians gradually refines both their optic and geometric parameters to represent the captured scene with high fidelity. + +# 3.2. Camera Pose Optimisation + +To achieve accurate tracking, we typically require at least 50 iterations of gradient descent per frame. This requirement emphasises the necessity of a representation with computationally efficient view synthesis and gradient computation, making the choice of 3D representation a crucial part of designing a SLAM system. + +![](images/97ceacd1e19adb68856325d2133821f4017de23d7ca01b34d044a0e689e68a9c.jpg) +Figure 2. SLAM System Overview: Our SLAM system uses 3D Gaussians as the only representation, unifying all components of SLAM, including tracking, mapping, keyframe management, and novel view synthesis. + +In order to avoid the overhead of automatic differentiation, 3DGS implements rasterisation with CUDA with derivatives for all parameters calculated explicitly. Since rasterisation is performance critical, we similarly derive the camera Jacobians explicitly. + +To the best of our knowledge, we provide the first analytical Jacobian of $\pmb{SE}(3)$ camera pose with respect to the 3D Gaussians used in EWA splatting [48] and 3DGS. This opens up new applications of 3DGS beyond SLAM. + +We use Lie algebra to derive the minimal Jacobians, ensuring that the dimensionality of the Jacobians matches the degrees of freedom, eliminating any redundant computations. The terms of Eq. (2) are differentiable with respect to the camera pose $\pmb{T}_{CW}$ ; using the chain rule: + +$$ +\frac {\partial \boldsymbol {\mu} _ {I}}{\partial \boldsymbol {T} _ {C W}} = \frac {\partial \boldsymbol {\mu} _ {I}}{\partial \boldsymbol {\mu} _ {C}} \frac {\mathcal {D} \boldsymbol {\mu} _ {C}}{\mathcal {D} \boldsymbol {T} _ {C W}}, \tag {3} +$$ + +$$ +\frac {\partial \boldsymbol {\Sigma} _ {I}}{\partial \boldsymbol {T} _ {C W}} = \frac {\partial \boldsymbol {\Sigma} _ {I}}{\partial \mathbf {J}} \frac {\partial \mathbf {J}}{\partial \boldsymbol {\mu} _ {C}} \frac {\mathcal {D} \boldsymbol {\mu} _ {C}}{\mathcal {D} \boldsymbol {T} _ {C W}} + \frac {\partial \boldsymbol {\Sigma} _ {I}}{\partial \mathbf {W}} \frac {\mathcal {D} \mathbf {W}}{\mathcal {D} \boldsymbol {T} _ {C W}}. \tag {4} +$$ + +where $T_{CW}$ represents the 3D position of Gaussian in the camera coordinate. We take the derivatives on the manifold to derive minimal parameterisation. Borrowing the notation from [30], let $T \in SE(3)$ and $\tau \in \mathfrak{se}(3)$ . We define the partial derivative on the manifold as: + +$$ +\frac {\mathcal {D} f (\boldsymbol {T})}{\mathcal {D} \boldsymbol {T}} \triangleq \lim _ {\tau \rightarrow 0} \frac {\operatorname {L o g} (f (\operatorname {E x p} (\tau) \circ \boldsymbol {T}) \circ f (\boldsymbol {T}) ^ {- 1})}{\tau}, \tag {5} +$$ + +where $\circ$ is a group composition, and Exp, Log are the exponential and logarithmic mappings between Lie algebra and Lie Group. With this, we derive the following: + +$$ +\frac {\mathcal {D} \boldsymbol {\mu} _ {C}}{\mathcal {D} \boldsymbol {T} _ {C W}} = \left[ \begin{array}{l l} \boldsymbol {I} & - \boldsymbol {\mu} _ {C} ^ {\times} \end{array} \right], \frac {\mathcal {D} \mathbf {W}}{\mathcal {D} \boldsymbol {T} _ {C W}} = \left[ \begin{array}{l l} \mathbf {0} & - \mathbf {W} _ {:, 1} ^ {\times} \\ \mathbf {0} & - \mathbf {W} _ {:, 2} ^ {\times} \\ \mathbf {0} & - \mathbf {W} _ {:, 3} ^ {\times} \end{array} \right] ， \tag {6} +$$ + +where $\times$ denotes the skew symmetric matrix of a 3D vector, and $\mathbf{W}_{:,i}$ refers to the $i$ th column of the matrix. + +# 3.3. SLAM + +In this section, we present details of full SLAM framework. The overview of the system is summarised in Fig. 2. Please refer to the supplementary material for the further parameter details. + +# 3.3.1 Tracking + +In tracking only the current camera pose is optimised, without updates to the map representation. In the monocular case, we minimise the following photometric residual: + +$$ +E _ {p h o} = \left\| I (\mathcal {G}, \boldsymbol {T} _ {C W}) - \bar {I} \right\| _ {1}, \tag {7} +$$ + +where $I(\mathcal{G},\pmb{T}_{CW})$ renders the Gaussians $\mathcal{G}$ from $T_{CW}$ , and $\bar{I}$ is an observed image. + +We further optimise affine brightness parameters for varying exposure. When depth observations are available, we define the geometric residual as: + +$$ +E _ {g e o} = \left\| D (\mathcal {G}, \boldsymbol {T} _ {C W}) - \bar {D} \right\| _ {1}, \tag {8} +$$ + +where $D(\mathcal{G},\mathbf{T}_{CW})$ is depth rasterisation and $\bar{D}$ is the observed depth. Rather than simply using the depth measurements to initialise the Gaussians, we minimise both photometric and geometric residuals: $\lambda_{pho}E_{pho} + (1 - \lambda_{pho})E_{geo}$ , where $\lambda_{pho}$ is a hyperparameter. + +As in Eq. (1), per-pixel depth is rasterised by alphabetending: + +$$ +\mathcal {D} _ {p} = \sum_ {i \in \mathcal {N}} z _ {i} \alpha_ {i} \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {j}), \tag {9} +$$ + +where $z_{i}$ is the distance to the mean $\pmb{\mu}_{W}$ of Gaussian $i$ along the camera ray. We derive analytical Jacobians for the camera pose optimisation in a similar manner to Eq. (3), (4). + +# 3.3.2 Keyframing + +Since using all the images from a video stream to jointly optimise the Gaussians and camera poses online is infeasible, we maintain a small window $\mathcal{W}_k$ consisting of carefully selected keyframes based on inter-frame covisibility. Ideal keyframe management will select non-redundant keyframes observing the same area, spanning a wide baseline to provide better multiview constraints. The parameters are detailed in supplementary. + +Selection and Management Every tracked frame is checked for keyframe registration based on our simple yet effective criteria. We measure the可视ibility by measuring the intersection over union of the observed Gaussians between the current frame $i$ and the last keyframe $j$ . If the可视ibility drops below a threshold, or if the relative translation $t_{ij}$ is large with respect to the median depth, frame $i$ is registered as a keyframe. For efficiency, we maintain only a small number of keyframes in the current window $\mathcal{W}_k$ following the keyframe management heuristics of DSO [4]. The main difference is that a keyframe is removed from the current window if the overlap coefficient with the latest keyframe drops below a threshold. + +Gaussian Covisibility An accurate estimate of covisibility simplifies keyframe selection and management. 3DGS respects visibility ordering since the 3D Gaussians are sorted along the camera ray. This property is desirable for covisibility estimation as occlusions are handled by design. A Gaussian is marked to be visible from a view if used in the rasterisation and if the ray's accumulated $\alpha$ has not yet reached 0.5. This enables our estimated covisibility to handle occlusions without requiring additional heuristics. + +Gaussian Insertion and Pruning At every keyframe, new Gaussians are inserted into the scene to capture newly visible scene elements and to refine the fine details. When depth measurements are available, Gaussian means $\mu_W$ are initialised by back-projecting the depth. In the monocular case, we render the depth at the current frame. For pixels with depth estimates, $\mu_W$ are initialised around those depths with low variance; for pixels without the depth estimates, we initialise $\mu_W$ around the median depth of the rendered image with high variance. + +In the monocular case, the positions of many newly inserted Gaussians are incorrect. While the majority will quickly vanish during optimisation as they violate multiview consistency, we further prune the excess Gaussians by + +checking the visibility amongst the current window $\mathcal{W}_k$ . If the Gaussians inserted within the last 3 keyframes are unobserved by at least 3 other frames, we prune them out as they are geometrically unstable. + +# 3.3.3 Mapping + +The purpose of mapping is to maintain a coherent 3D structure and to optimise the newly inserted Gaussians. During mapping, the keyframes in $\mathcal{W}_k$ are used to reconstruct currently visible regions. Additionally, two random past keyframes $\mathcal{W}_r$ are selected per iteration to avoid forgetting the global map. Rasterisation of 3DGS imposes no constraint on the Gaussians along the viewing ray direction, even with a depth observation. This is not a problem when sufficient carefully selected viewpoints are provided (e.g. in the novel view synthesis case); however, in continuous SLAM this causes many artefacts, making tracking challenging. We therefore introduce an isotropic regularisation: + +$$ +E _ {i s o} = \sum_ {i = 1} ^ {| \mathcal {G} |} \| \mathbf {s} _ {i} - \tilde {\mathbf {s}} _ {i} \cdot \mathbf {1} \| _ {1} \tag {10} +$$ + +to penalise the scaling parameters $\mathbf{s}_i$ (i.e. stretch of the ellipsoid) by its difference to the mean $\tilde{\mathbf{s}}_i$ . As shown in Fig 3, this encourages sphericality, and avoids the problem of Gaussians which are highly elongated along the viewing direction creating artefacts. Let the union of the keyframes in the current window and the randomly selected one be $\mathcal{W} = \mathcal{W}_k \cup \mathcal{W}_r$ . For mapping, we solve the following problem: + +$$ +\min _ {\boldsymbol {T} _ {C W} ^ {k} \in \boldsymbol {S E} (3), \mathcal {G}, \forall k \in \mathcal {W}} \sum_ {\forall k \in \mathcal {W}} E _ {p h o} ^ {k} + \lambda_ {i s o} E _ {i s o}. \tag {11} +$$ + +If depth observations are available, as in tracking, geometric residuals Eq. (8) are added to the optimisation problem. + +# 4. Evaluation + +We conduct a comprehensive evaluation of our system across a range of both real and synthetic datasets. Additionally, we perform an ablation study to justify our design choices. Finally, we present qualitative results of our system operating live using a monocular camera, illustrating its practicality and high fidelity reconstruction. + +# 4.1. Experimental Setup + +Datasets For our quantitative analysis, we evaluate our method on the TUM RGB-D dataset [32] (3 sequences) and the Replica dataset [31] (8 sequences), following the evaluation in [33]. For qualitative results, we use self-captured real-world sequences recorded by Intel Realsense d455. Since the Replica dataset is designed for RGB-D SLAM + +![](images/d3b400550f268e328fb72f93546cffb3b56e822362318d4823f1d64a4bd74d87.jpg) + +![](images/cfb4ab71fd891c902d7dffe34369c688ecaeb250f6f3e0253f77581c79108702.jpg) +Figure 3. Effect of isotropic regularisation: Top: Rendering close to a training view (looking at the keyboard). Bottom: Rendering 3D Gaussians far from the training views (view from a side of the keyboard) without (left) and with (right) the isotropic loss. When the photometric constraints are insufficient, the Gaussians tend to elongate along the viewing direction, creating artefacts in the novel views, and affecting the camera tracking. + +![](images/d4232cbfd10cd50adaf9060720a2b68ce9c4820b6d9dea29265c927bb26fe56b.jpg) + +evaluation, it contains challenging purely rotational camera motions. We hence use the Replica dataset for RGB-D evaluation only. The TUM RGB-D dataset is used for both monocular and RGB-D evaluation. + +Implementation Details We run our SLAM on a desktop with Intel Core i9 12900K 3.50GHz and a single NVIDIA GeForce RTX 4090. We present results from our multiprocessing implementation aimed at real-time applications. For a fair comparison with other methods on Replica, we additionally report result for single-process implementation which performs more mapping iterations. As with 3DGS, time-critical rasterisation and gradient computation are implemented using CUDA. The rest of the SLAM pipeline is developed with PyTorch. Details of hyperparameters are provided in the supplementary material. + +Metrics For camera tracking accuracy, we report the Root Mean Square Error (RMSE) of the Absolute Trajectory Error (ATE) of the keyframes. To evaluate map quality, we report standard photometric rendering quality metrics (PSNR, SSIM and LPIPS) following the evaluation protocol used in [27]. To evaluate the map quality, on every fifth frame, rendering metrics are computed. We exclude the keyframes (training views). We report the average across three runs for all our evaluations. In the tables, the best result is in bold, and the second best is underlined. + +Baseline Methods We primarily benchmark our SLAM method against other approaches that, like ours, do not have explicit loop closure. In monocular settings, we compare + +with state-of-the-art classical and learning-based direct visual odometry (VO) methods. Specifically, we compare DSO [4], DepthCov [3], and DROID-SLAM [36] in VO configurations. These methods are selected based on their public reporting of results on the benchmark (TUM dataset) or the availability of their source code for getting the benchmark result. Since one of our focuses is the online scale estimation under monocular scale ambiguity, the method which uses ground truth poses for the system initialisation such as [13] is not considered for the comparison. In the RGB-D case, we compare against neural-implicit SLAM methods [7, 8, 27, 33, 39, 43, 46] which are also map-centric, rendering-based and do not perform loop closure. + +# 4.2. Quantitative Evaluation + +Camera Tracking Accuracy Table 1 shows the tracking results on the TUM RGB-D dataset. In the monocular setting, our method surpasses other baselines without requiring any deep priors. Furthermore, our performance is comparable to systems which perform explicit loop closure. This clearly highlights that there still remains potential for enhancing the tracking of monocular SLAM by exploring fundamental SLAM representations. + +Our RGB-D method shows better performance than any other baseline method. Notably, our system surpasses ORB-SLAM in the fr1 sequences, narrowing the gap between Map-centric SLAM and the state-of-the-art sparse frame-centric methods. Table 2 reports results on the synthetic Replica dataset. Our single-process implementation shows competitive performance and achieves the best result in 4 out of 8 sequences. Our multi-process implementation which performs fewer mapping iterations still performs comparably. In contrast to other methods, our system demonstrates higher performance on real-world data, by optimising the Gaussian positions to compensate for the sensor noise. + +Novel View Rendering Table 5 summarises the novel view rendering performance of our method with RGB-D input. We consistently show the best performance across most sequences and is least second best. Our rendering FPS is hundreds of times faster than other methods, offering a significant advantage for applications which require real-time map interaction. While Point-SLAM is competitive, that method focuses on view synthesis rather than novel-view synthesis. Their view synthesis is conditional on the availability of depth due to the depth-guided ray-sampling, making novel-view synthesis challenging. On the other hand, our rasterisation-based approach does not require depth guidance and achieves efficient, high-quality, novel view synthesis. Fig. 4 provides a qualitative comparison of the rendering of ours and Point-SLAM (with depth guidance). + +
InputLoop-closureMethodfr1/deskfr2/xyzfr3/officeAvg.
Monocularw/oDSO [4]22.41.109.5011.0
DROID-VO [36]5.2010.77.307.73
DepthCov-VO [3]5.601.2068.825.2
Ours3.784.603.503.96
w/DROID-SLAM [36]1.800.502.801.70
ORB-SLAM2 [20]1.900.602.401.60
RGB-Dw/oiMAP [33]4.902.005.804.23
NICE-SLAM [46]4.266.193.874.77
DI-Fusion [7]4.402.005.804.07
Vox-Fusion [43]3.521.4926.0110.34
ESLAM [8]2.471.112.422.00
Co-SLAM [39]2.401.702.402.17
Point-SLAM [27]4.341.313.483.04
Ours1.501.441.491.47
w/BAD-SLAM [29]1.701.101.701.50
Kintinous [40]3.702.903.003.20
ORB-SLAM2 [20]1.600.401.001.00
+ +Table 1. Camera tracking result on TUM for monocular and RGB-D. ATE RMSE in cm is reported. In both monocular and RGB-D cases, we achieve state-of-the-art performance. In particular, in the monocular case, not only do we outperform systems which use deep prior, but we achieve comparable performance with many of the RGB-D systems. + +
Methodr0r1r2o0o1o2o3o4Avg.
iMAP [33]3.122.542.311.691.033.994.051.932.58
NICE-SLAM0.971.311.070.881.001.061.101.131.07
Vox-Fusion [43]1.374.701.478.482.042.581.112.943.09
ESLAM [8]0.710.700.520.570.550.580.720.630.63
Point-SLAM [27]0.610.410.370.380.480.540.690.720.53
Ours0.440.320.310.440.520.230.172.250.58
Ours (sp)0.330.220.290.360.190.250.120.810.32
+ +Table 2. Camera tracking result on Replica for RGB-D SLAM. ATE RMSE in cm is reported. We achieve best performance across most sequences. Here, Ours is our multi-process implementation and Ours (sp) is the single-process implementation which ensures a certain amount of mapping iteration similar to other works. + +
InputMethodfr1/deskfr2/xyzfr3/officeAvg.
Monow/o \( E_{iso} \)4.164.665.734.83
w/o kf selection13.24.368.658.73
Ours3.784.603.503.96
RGB-Dw/o \( E_{geo} \)2.390.624.982.66
w/o kf selection1.641.492.601.90
Ours1.501.441.491.47
+ +Table 3. Ablation Study on TUM RGB-D dataset. We analyse the usefulness of isotropic regularisation, geometric residual, and keyframe selection to our SLAM system. Further isotropic regularisation ablation is available in supplementary. + +
Memory Usage [MB]
iMAP [33]NICE-SLAM [46]Co-SLAM [39]Ours (Mono)Ours (RGB-D)
0.8MB40.3.4MB6.4MB2.6MB3.97MB
+ +Ablative Analysis In Table 3, we perform ablation to confirm our design choices. Isotropic regularisation and geometric residual improve the tracking of monocular and RGB-D SLAM respectively, as they aid in constraining the geometry when photometric signals are weak. For both cases, keyframe selection significantly improves systems performance, as it automatically chooses suitable keyframes + +Table 4. Memory Analysis on TUM RGB-D dataset. The baseline numbers are computed from the parameter numbers in [39] + +
MethodPSNR[db]↑SSIM↑LPIPS↓Rendering FPS
NICE-SLAM[46]24.420.8090.2330.54
Vox-Fusion[43]24.410.8010.2362.17
Point-SLAM [27]35.170.9750.1241.33
ours38.940.9680.070769
+ +![](images/22e58c9c0ffa4fd7aec784b64abf8ac6b021bddd9689fd072931ef9cc292d6ff.jpg) +Figure 4. Rendering examples on Replica. Point-SLAM struggle with rendering fine details due to the stochastic ray sampling. + +based on our occlusion-aware keyframe selection and management. We further compare the memory usage of different 3D representations in Table 4. MLP-based iMAP is clearly more memory efficient, but it struggles to express high-fidelity 3D scenes due to the limited capacity of small MLP. Compared with a voxel grid of features used in NICE-SLAM, our method uses significantly less memory. + +Table 5. Average rendering performance on Replica (RGB-D). Our method outperforms most of the rendering metrics compared to existing methods. Note that Point-SLAM uses ground-truth depth to guide sampling along rays. The full detail is available in supplementary. + +
Methodseq1seq2seq3Avg.
Neural SDF (Hash Grid)0.130.150.160.14
Neural SDF (MLP)0.400.380.220.33
Ours w/o depth0.820.910.650.79
Ours w/ depth0.831.00.650.82
+ +Table 6. Camera convergence analysis. We report the ratio of successful camera convergence for the different sequences, across different differentiable 3D representations. + +Convergence Basin Analysis In our SLAM experiments, we discovered that 3D Gaussian maps have a notably large convergence basin for camera localisation. To investigate further, we conducted a convergence funnel analysis, an evaluation methodology proposed in [18] and used in [22]. Here, we train a 3D representation (e.g. 3DGS) using 9 fixed views arranged in a square. We set the viewpoint in the middle of the square to be the target view. As shown in Fig 5, we uniformly sample a position, creating a funnel. + +![](images/6479bae5c85977d9e87231cc8aa653b8995ced9101bdf0ea466cbe0fd5dcc0cd.jpg) +Camera Layout +Figure 5. Convergence basin analysis: Left: 3D Gaussian map from training views (Yellow) and visualisation of the test poses (Red) and target pose (Blue). Right: Convergence basin of our method. The green marks success, and the red marks failure. + +![](images/3c6933a24620117c00152aa1c63b4984b51f73bfee7a703aeee21bf173b14185.jpg) +Ours w/ depth + +![](images/398ea8a505dfd31b526338e5c8b68a567f418bb02301e9565b80a773c5ca9f35.jpg) +Ours w/o depth + +![](images/f3a2e348ba530ca0faa18294db78c0359b0718b93ba9cb2d9224e85ba7f00766.jpg) +Hash Grid SDF + +![](images/499fec0da5cbf54cc205e566a99a11ddeb17e1db06d1464acbdb4ab2844c6289.jpg) +MLP SDF + +![](images/723f9e4c7b39e327f5ec2b8349c052b9a0887637fc93f1ec910a8a91e9c45ae4.jpg) +Figure 6. Monocular SLAM result on fr1/desk sequence: We show the reconstructed 3D Gaussian maps (Left) and novel view synthesis result (Right). + +![](images/6b963a76d8f3a9f1202632b6815ae6c22899a26fa92e4b325fe42a2dedf021aa.jpg) + +![](images/9e01286b1746790e8df6077f6deeab8ece2b6bc4745c733801623e2e1bb2fde2.jpg) + +![](images/bd4c1275abc36e95390b0c0249c6720ceaa06f0c8a8feb95b1c5ff54438a3fca.jpg) + +From the sampled position, given the RGB image of the target view, we perform camera pose optimisation for 1000 iterations. The optimisation is successful if it converges to within $1\mathrm{cm}$ of the target view within the fixed iterations. We compare our Gaussian approach with Co-SLAM [39]'s network (Hash Grid SDF) and iMAP's [33] network with Co-SLAM's SDF loss for further geometric accuracy (MLP Neural SDF). We render the training views using a synthetic Replica dataset and create three sequences for testing (seq1, seq2 and seq3). The width of the square formed by the training view is $0.5\mathrm{m}$ , and the test cameras are distributed with radii ranging from $0.2\mathrm{m}$ to $1.2\mathrm{m}$ , covering a larger area than the training view. When training the map, the three methods—Ours w/depth, Hash Grid SDF, and MLP SDF—use RGB-D images, whereas Ours w/o depth utilises only colour images. Fig. 5 shows the qualitative results and Table 6 reports the success rate. For both with and without depth for training, our method shows better convergence. + +![](images/47e9dec8d1434bc8a0c86b8dfa8b7be232b9d911067a469086644d77097f97c5.jpg) + +![](images/21a2ca9113fec668cf62842a26c986443f45def672bb057fc41d56349540e139.jpg) + +![](images/b0cccaaa883c0586fd89440b919de4caf748659e1d7776d32578ecfe0d28b27f.jpg) +Figure 7. Self-captured Scenes: Challenging scenes and objects, for example, transparent glasses and crinkled texture of salad are captured by our monocular SLAM running live. + +![](images/6e64f8174f79873aef6837b1b9eed17c58347638bc2c7bbafcb157b8dad47274.jpg) + +Unlike hashing and positional encoding which can lead to signal conflict, anisotropic Gaussians form a smooth gradient in 3D space, increasing the convergence basin. Further experimental details are available in supplementary ?? + +# 4.3. Qualitative Results + +We report both the 3D reconstruction of the SLAM dataset and self-captured sequences. In Fig. 6, we visualise the monocular SLAM reconstruction of fr1/desk. The placements of the Gaussians are geometrically sensible and are 3D coherent, and our rendering from the different viewpoints highlights the quality of our systems' novel view synthesis. In Fig. 7, we self-capture challenging scenes for monocular SLAM. By not explicitly modelling a surface, our system naturally handles transparent objects which is challenging for many other SLAM systems. + +# 5. Conclusion + +We have proposed the first SLAM method using 3D Gaussians as a SLAM representation. Via efficient volume rendering, our system significantly advances the fidelity and diversity of object materials a live SLAM system can capture. Our system achieves state-of-the-art performance across benchmarks for both monocular and RGB-D cases. Interesting directions for future research are the integration of loop closure for handling large-scale scenes and extraction of geometry such as surface normal as Gaussians do not explicitly represent the surface. + +# Acknowledgements + +Research presented in this paper has been supported by Dyson Technology Ltd. We are very grateful to Eric Dexheimer, Kirill Mazur, Xin Kong, Marwan Taher, Ignacio Alzugaray, Gwangbin Bae, Aalok Patwardhan, and members of the Dyson Robotics Lab for insightful discussions. + +# References + +[1] J. Czarnowski, T. Laidlow, R. Clark, and A. J. Davison. Deepfactors: Real-time probabilistic dense monocular SLAM. IEEE Robotics and Automation Letters (RAL), 5(2): 721-728, 2020. +[2] Angela Dai, Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Christian Theobalt. BundleFusion: Real-time Globally Consistent 3D Reconstruction using On-the-fly Surface Re-integration. ACM Transactions on Graphics (TOG), 36(3):24:1-24:18, 2017. +[3] Eric Dexheimer and Andrew J. Davison. Learning a Depth Covariance Function. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. +[4] J. Engel, V. Koltun, and D. Cremers. Direct sparse odometry. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 2017. +[5] C. Forster, M. Pizzoli, and D. Scaramuzza. SVO: Fast Semi-Direct Monocular Visual Odometry. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2014. +[6] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenumxels: Radiance fields without neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. +[7] Jiahui Huang, Shi-Sheng Huang, Haoxuan Song, and ShiMin Hu. Di-fusion: Online implicit 3d reconstruction with deep priors. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. +[8] M. M. Johari, C. Carta, and F. Fleuret. ESLAM: Efficient dense slam system based on hybrid representation of signed distance fields. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. +[9] M. Keller, D. Lefloch, M. Lambers, S. Izadi, T. Weyrich, and A. Kolb. Real-time 3D Reconstruction in Dynamic Scenes using Point-based Fusion. In Proc. of Joint 3DIM/3DPVT Conference (3DV), 2013. +[10] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (TOG), 2023. +[11] Leonid Keselman and Martial Hebert. Approximate differentiable rendering with algebraic surfaces. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. +[12] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proceedings of the International Conference on Learning Representations (ICLR), 2015. +[13] Heng Li, Xiaodong Gu, Weihao Yuan, Luwei Yang, Zilong Dong, and Ping Tan. Dense rgb slam with neural implicit maps. In Proceedings of the International Conference on Learning Representations (ICLR), 2023. +[14] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. NeurIPS, 2020. + +[15] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. 3DV, 2024. +[16] J. McCormac, A. Handa, A. J. Davison, and S. Leutenegger. SemanticFusion: Dense 3D semantic mapping with convolutional neural networks. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2017. +[17] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. +[18] N. J. Mitra, N. Gelfand, H. Pottmann, and L. J. Guibas. Registration of Point Cloud Data from a Geometric Optimization Perspective. In Proceedings of the Symposium on Geometry Processing, 2004. +[19] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (TOG), 2022. +[20] R. Mur-Artal and J. D. Tardós. ORB-SLAM2: An Open-Source SLAM System for Monocular, Stereo, and RGB-D Cameras. IEEE Transactions on Robotics (T-RO), 33(5): 1255–1262, 2017. +[21] R. Mur-Artal, J. M. M Montiel, and J. D. Tardós. ORBSLAM: a Versatile and Accurate Monocular SLAM System. IEEE Transactions on Robotics (T-RO), 31(5):1147-1163, 2015. +[22] R. A. Newcombe. Dense Visual SLAM. PhD thesis, Imperial College London, 2012. +[23] R. A. Newcombe, S. Izadi, O. Hilliges, D. Molyneaux, D. Kim, A. J. Davison, P. Kohli, J. Shotton, S. Hodges, and A. Fitzgibbon. KinectFusion: Real-Time Dense Surface Mapping and Tracking. In Proceedings of the International Symposium on Mixed and Augmented Reality (ISMAR), 2011. +[24] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. +[25] M. Nießner, M. Zollhöfer, S. Izadi, and M. Stamminger. Real-time 3D Reconstruction at Scale using Voxel Hashing. In Proceedings of SIGGRAPH, 2013. +[26] Victor Adrian Prisacariu, Olaf Kähler, Ming-Ming Cheng, Carl Yuheng Ren, Julien P. C. Valentin, Philip H. S. Torr, Ian D. Reid, and David W. Murray. A framework for the volumetric integration of depth images. CoRR, abs/1410.0925, 2014. +[27] Erik Sandström, Yue Li, Luc Van Gool, and Martin R. Oswald. Point-slam: Dense neural point cloud-based slam. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. +[28] Thomas Schöps, Torsten Sattler, and Marc Pollefeys. Surfelmeshing: Online surfel-based mesh reconstruction. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 2020. + +[29] Thomas Schöps, Torsten Sattler, and Marc Pollefeys. Bad slam: Bundle adjusted direct rgb-d slam. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. +[30] J. Solà, J. Deray, and D. Atchuthan. A micro Lie theory for state estimation in robotics. arXiv:1812.01537, 2018. +[31] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, Anton Clarkson, Mingfei Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Braies, Tyler Gillingham, Elias Mueggler, Luis Pesqueira, Manolis Savva, Dhruv Batra, Hauke M. Strasdat, Renzo De Nardi, Michael Goesele, Steven Lovegrove, and Richard Newcombe. The Replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. +[32] J. Sturm, N. Engelhard, F. Endres, W. Burgard, and D. Cremers. A Benchmark for the Evaluation of RGB-D SLAM Systems. In Proceedings of the IEEE/RSJ Conference on Intelligent Robots and Systems (IROS), 2012. +[33] E. Sucar, S. Liu, J. Ortiz, and A. J. Davison. iMAP: Implicit mapping and positioning in real-time. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. +[34] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. +[35] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. Proceedings of the International Conference on Learning Representations (ICLR), 2024. +[36] Zachary Teed and Jia Deng. DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras. In Neural Information Processing Systems (NIPS), 2021. +[37] Emanuele Vespa, Nikolay Nikolov, Marius Grimm, Luigi Nardi, Paul HJ Kelly, and Stefan Leutenegger. Efficient octree-based volumetric SLAM supporting signed-distance and occupancy mapping. IEEE Robotics and Automation Letters (RAL), 2018. +[38] Angtian Wang, Peng Wang, Jian Sun, Adam Kortylewski, and Alan Yuille. Voge: a differentiable volume renderer using gaussian ellipsoids for analysis-by-synthesis. 2022. +[39] Hengyi Wang, Jingwen Wang, and Lourdes Agapito. Coslam: Joint coordinate and sparse parametric encodings for neural real-time slam. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. +[40] T. Whelan, M. Kaess, H. Johannsson, M. F. Fallon, J. J. Leonard, and J. B. McDonald. Real-time large scale dense RGB-D SLAM with volumetric fusion. International Journal of Robotics Research (IJRR), 34(4-5):598-626, 2015. +[41] T. Whelan, S. Leutenegger, R. F. Salas-Moreno, B. Glocker, and A. J. Davison. ElasticFusion: Dense SLAM without a pose graph. In Proceedings of Robotics: Science and Systems (RSS), 2015. +[42] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. + +Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. +[43] Xingrui Yang, Hai Li, Hongjia Zhai, Yuhang Ming, Yuqian Liu, and Guofeng Zhang. Vox-fusion: Dense tracking and mapping with voxel-based neural implicit representation. In Proceedings of the International Symposium on Mixed and Augmented Reality (ISMAR), 2022. +[44] Zeyu Yang, Hongye Yang, Zijie Pan, Xiatian Zhu, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. Proceedings of the International Conference on Learning Representations (ICLR), 2024. +[45] Taoran Yi, Jiemin Fang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussian-dreamer: Fast generation from text to 3d gaussian splatting with point cloud priors. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. +[46] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R. Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. +[47] Zihan Zhu, Songyou Peng, Viktor Larsson, Zhaopeng Cui, Martin R Oswald, Andreas Geiger, and Marc Pollefeys. Nicer-slam: Neural implicit scene encoding for rgb slam. International Conference on 3D Vision (3DV), 2024. +[48] M. Zwicker, H. Pfister, J. van Baar, and M. Gross. Ewa splatting. IEEE Transactions on Visualization and Computer Graphics, 8(3):223-238, 2002. \ No newline at end of file diff --git a/2024/Gaussian Splatting SLAM/images.zip b/2024/Gaussian Splatting SLAM/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..765db0143019f2108107fb32e9e70e54f31623ae --- /dev/null +++ b/2024/Gaussian Splatting SLAM/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a9d8f7bd3cb6069323b7a18a48eee37ba31cf7b8a7e835efd8c93bd750d315d4 +size 691537 diff --git a/2024/Gaussian Splatting SLAM/layout.json b/2024/Gaussian Splatting SLAM/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4a62fc6b0db51a679fed8b79b287dbcb1d2b8421 --- /dev/null +++ b/2024/Gaussian Splatting SLAM/layout.json @@ -0,0 +1,8395 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 214, + 103, + 380, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 103, + 380, + 121 + ], + "spans": [ + { + "bbox": [ + 214, + 103, + 380, + 121 + ], + "type": "text", + "content": "Gaussian Splatting SLAM" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 83, + 142, + 185, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 142, + 185, + 156 + ], + "spans": [ + { + "bbox": [ + 83, + 142, + 185, + 156 + ], + "type": "text", + "content": "Hidenobu Matsuki" + }, + { + "bbox": [ + 83, + 142, + 185, + 156 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 209, + 143, + 276, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 143, + 276, + 156 + ], + "spans": [ + { + "bbox": [ + 209, + 143, + 276, + 156 + ], + "type": "text", + "content": "Riku Murai" + }, + { + "bbox": [ + 209, + 143, + 276, + 156 + ], + "type": "inline_equation", + "content": "^{2*}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 301, + 143, + 383, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 143, + 383, + 157 + ], + "spans": [ + { + "bbox": [ + 301, + 143, + 383, + 157 + ], + "type": "text", + "content": "Paul H. J. Kelly" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 408, + 143, + 505, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 143, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 408, + 143, + 505, + 156 + ], + "type": "text", + "content": "Andrew J. Davison1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 126, + 163, + 467, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 163, + 467, + 191 + ], + "spans": [ + { + "bbox": [ + 126, + 163, + 467, + 191 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 126, + 163, + 467, + 191 + ], + "type": "text", + "content": "Dyson Robotics Laboratory, Imperial College London " + }, + { + "bbox": [ + 126, + 163, + 467, + 191 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 126, + 163, + 467, + 191 + ], + "type": "text", + "content": "Software Performance Optimisation Group, Imperial College London" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 128, + 194, + 462, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 194, + 462, + 205 + ], + "spans": [ + { + "bbox": [ + 128, + 194, + 462, + 205 + ], + "type": "text", + "content": "{h.matsuki20, riku.murai15, p.kelly, a.davison}@imperial.ac.uk" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 50, + 233, + 120, + 364 + ], + "blocks": [ + { + "bbox": [ + 50, + 233, + 120, + 364 + ], + "lines": [ + { + "bbox": [ + 50, + 233, + 120, + 364 + ], + "spans": [ + { + "bbox": [ + 50, + 233, + 120, + 364 + ], + "type": "image", + "image_path": "c2f83a8fbdb7d189748a19aee544271d2f9f30f06468e5043b49446e3ad4f1e9.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 366, + 546, + 432 + ], + "lines": [ + { + "bbox": [ + 45, + 366, + 546, + 432 + ], + "spans": [ + { + "bbox": [ + 45, + 366, + 546, + 432 + ], + "type": "text", + "content": "Figure 1. From a single monocular camera, we reconstruct a high fidelity 3D scene live at 3fps. For every incoming RGB frame, 3D Gaussians are incrementally formed and optimised together with the camera poses. We show both the rasterised Gaussians (left) and Gaussians shaded to highlight the geometry (right). Notice the details and the complex material properties (e.g. transparency) captured. Thin structures such as wires are accurately represented by numerous small, elongated Gaussians, and transparent objects are effectively represented by placing the Gaussians along the rim. Our system significantly advances the fidelity a live monocular SLAM system can capture." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 121, + 234, + 297, + 364 + ], + "blocks": [ + { + "bbox": [ + 121, + 234, + 297, + 364 + ], + "lines": [ + { + "bbox": [ + 121, + 234, + 297, + 364 + ], + "spans": [ + { + "bbox": [ + 121, + 234, + 297, + 364 + ], + "type": "image", + "image_path": "0cf846d57eab8d874d336abf74b7dbdc1f6ca25d48c9d112804de3a177a69eb4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 298, + 234, + 474, + 364 + ], + "blocks": [ + { + "bbox": [ + 298, + 234, + 474, + 364 + ], + "lines": [ + { + "bbox": [ + 298, + 234, + 474, + 364 + ], + "spans": [ + { + "bbox": [ + 298, + 234, + 474, + 364 + ], + "type": "image", + "image_path": "de58f0539762e5fe3ede81e0b04853165ffba86077c8ef7af1cf3f36874bb722.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 474, + 234, + 544, + 364 + ], + "blocks": [ + { + "bbox": [ + 474, + 234, + 544, + 364 + ], + "lines": [ + { + "bbox": [ + 474, + 234, + 544, + 364 + ], + "spans": [ + { + "bbox": [ + 474, + 234, + 544, + 364 + ], + "type": "image", + "image_path": "929b63922c0a7c841cf1939d4afee510145b830db8bcce80ff9e2c27b03447d4.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 143, + 462, + 192, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 462, + 192, + 475 + ], + "spans": [ + { + "bbox": [ + 143, + 462, + 192, + 475 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 488, + 289, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 488, + 289, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 488, + 289, + 693 + ], + "type": "text", + "content": "We present the first application of 3D Gaussian Splating in monocular SLAM, the most fundamental but the hardest setup for Visual SLAM. Our method, which runs live at 3fps, utilises Gaussians as the only 3D representation, unifying the required representation for accurate, efficient tracking, mapping, and high-quality rendering. Designed for challenging monocular settings, our approach is seamlessly extendable to RGB-D SLAM when an external depth sensor is available. Several innovations are required to continuously reconstruct 3D scenes with high fidelity from a live camera. First, to move beyond the original 3DGS algorithm, which requires accurate poses from an offline Structure from Motion (SfM) system, we formulate camera tracking for 3DGS using direct optimisation against the 3D Gaussians, and show that this enables fast and robust tracking with a wide basin of convergence. Second, by utilising the explicit nature of the Gaussians, we introduce geometric verification" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 452, + 546, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 452, + 546, + 511 + ], + "spans": [ + { + "bbox": [ + 306, + 452, + 546, + 511 + ], + "type": "text", + "content": "and regularisation to handle the ambiguities occurring in incremental 3D dense reconstruction. Finally, we introduce a full SLAM system which not only achieves state-of-the-art results in novel view synthesis and trajectory estimation but also reconstruction of tiny and even transparent objects." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 536, + 387, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 536, + 387, + 549 + ], + "spans": [ + { + "bbox": [ + 306, + 536, + 387, + 549 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "content": "A long-term goal of online reconstruction with a single moving camera is near-photorealistic fidelity, which will surely allow new levels of performance in many areas of Spatial AI and robotics as well as opening up a whole range of new applications. While we increasingly see the benefit of applying powerful pre-trained priors to 3D reconstruction, a key avenue for progress is still the invention and development of core 3D representations with advantageous properties. Many \"layered\" SLAM methods exist which tackle the SLAM problem by integrating multiple different 3D representations or existing SLAM components; however, the most interesting advances are when a new unified dense representation can be used for all aspects" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 702, + 200, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 702, + 200, + 712 + ], + "spans": [ + { + "bbox": [ + 61, + 702, + 200, + 712 + ], + "type": "text", + "content": "*Authors contributed equally to this work." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18039" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 108 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 108 + ], + "type": "text", + "content": "of a system's operation: local representation of detail, large-scale geometric mapping and also camera tracking by direct alignment." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 113, + 286, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 113, + 286, + 446 + ], + "spans": [ + { + "bbox": [ + 49, + 113, + 286, + 446 + ], + "type": "text", + "content": "In this paper, we present the first online visual SLAM system based solely on the 3D Gaussian Splatting (3DGS) representation [10] recently making a big impact in offline scene reconstruction. In 3DGS a scene is represented by a large number of Gaussian blobs with orientation, elongation, colour and opacity. Other previous world/map-centric scene representations used for visual SLAM include occupancy or Signed Distance Function (SDF) voxel grids [23]; meshes [28]; point or surfel clouds [9, 29]; and recently neural fields [33]. Each of these has disadvantages: grids use significant memory and have bounded resolution, and even if octrees or hashing allow more efficiency they cannot be flexibly warped for large corrections [25, 37]; meshes require difficult, irregular topology to fuse new information; surfel clouds are discontinuous and difficult to fuse and optimise; and neural fields require expensive per-pixel raycasting to render. We show that 3DGS has none of these weaknesses. As a SLAM representation, it is most similar to point and surfel clouds, and inherits their efficiency, locality and ability to be easily warped or modified. However, it also represents geometry in a smooth, continuously differentiable way: a dense cloud of Gaussians merge together and jointly define a continuous volumetric function. And crucially, the design of modern graphics cards means that a large number of Gaussians can be efficiently rendered via \"splatting\" rasterisation, up to 200fps at " + }, + { + "bbox": [ + 49, + 113, + 286, + 446 + ], + "type": "inline_equation", + "content": "1080\\mathrm{p}" + }, + { + "bbox": [ + 49, + 113, + 286, + 446 + ], + "type": "text", + "content": ". This rapid, differentiable rendering is integral to the tracking and map optimisation loops in our system." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 451, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 451, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 451, + 286, + 712 + ], + "type": "text", + "content": "The 3DGS representation has up until now only been used in offline systems for 3D reconstruction with known camera poses, and we present several innovations to enable online SLAM. We first derive the analytic Jacobian on Lie group of camera pose with respect to a 3D Gaussians map, and show that this can be seamlessly integrated into the existing differentiable rasterisation pipeline to enable camera poses to be optimised alongside scene geometry. Second, we introduce a novel Gaussian isotropic shape regularisation, to ensure geometric consistency, which we have found is important for incremental reconstruction. Third, we propose a novel Gaussian resource allocation and pruning method to keep the geometry clean and enable accurate camera tracking. Our experimental results demonstrate photorealistic online local scene reconstruction, as well as state-of-the-art camera trajectory estimation and mapping for larger scenes compared to other rendering-based SLAM methods. We further show the uniqueness of the Gaussian-based SLAM method such as an extremely large camera pose convergence basin, which can also be useful for map-based camera localisation. Our method works with only monocular input, one of the most challenging scenarios in" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 308, + 72, + 545, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 72, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 308, + 72, + 545, + 167 + ], + "type": "text", + "content": "SLAM. To highlight the intrinsic capability of 3D Gaussian for camera localisation, our method does not use any pretrained monocular depth predictor or other existing tracking modules, but relies solely on RGB image inputs in line with the original 3DGS. Since this is one of the most challenging SLAM scenario, we also show our method can easily be extended to RGB-D SLAM when depth measurements are available." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 319, + 169, + 501, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 169, + 501, + 179 + ], + "spans": [ + { + "bbox": [ + 319, + 169, + 501, + 179 + ], + "type": "text", + "content": "In summary, our contributions are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 308, + 181, + 545, + 298 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 308, + 181, + 545, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 181, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 308, + 181, + 545, + 215 + ], + "type": "text", + "content": "- The first near real-time SLAM system which works with a 3DGS as the only underlying scene representation, which can handle monocular only inputs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 308, + 217, + 544, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 544, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 544, + 262 + ], + "type": "text", + "content": "- Novel techniques within the SLAM framework, including the analytic Jacobian on Lie group for direct camera pose estimation, isotropic regularisation of the Gaussian shape, and geometric verification." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 308, + 264, + 544, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 544, + 298 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 544, + 298 + ], + "type": "text", + "content": "- Extensive evaluations on a variety of datasets both for monocular and RGB-D settings, demonstrating competitive performance, particularly in real-world scenarios." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 308, + 311, + 391, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 311, + 391, + 323 + ], + "spans": [ + { + "bbox": [ + 308, + 311, + 391, + 323 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 308, + 331, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 545, + 712 + ], + "type": "text", + "content": "Dense SLAM: Dense visual SLAM focuses on reconstructing detailed 3D maps, unlike sparse SLAM methods which excel in pose estimation [4, 5, 21] but typically yield maps useful mainly for localisation. In contrast, dense SLAM creates interactive maps beneficial for broader applications, including AR and robotics. Dense SLAM methods are generally divided into two primary categories: Frame-centric and Map-centric. Frame-centric SLAM minimises photometric error across consecutive frames, jointly estimating per-frame depth and frame-to-frame camera motion. Frame-centric approaches [1, 36] are efficient, as individual frames host local rather than global geometry (e.g. depth maps), and are attractive for long-session SLAM, but if a dense global map is needed, it must be constructed on demand by assembling all of these parts which are not necessarily fully consistent. In contrast, Map-centric SLAM uses a unified 3D representation across the SLAM pipeline, enabling a compact and streamlined system. Compared to purely local frame-to-frame tracking, a map-centric approach leverages global information by tracking against the reconstructed 3D consistent map. Classical map-centric approaches often use voxel grids [2, 23, 26, 40] or points [9, 29, 41] as the underlying 3D representation. While voxels enable a fast look-up of features in 3D, the representation is expensive, and the fixed voxel resolution and distribution are problematic when the spatial characteristics of the environment are not known in advance. On the other hand, a point-based map representation, such as surfel clouds, enables adaptive changes in resolution and spatial distribution by dynamic allocation of point primitives in the 3D space. Such flexibility benefits online applications such as SLAM with deformation-based loop closure [29, 41]. However, op" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 750, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 750, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 750, + 317, + 757 + ], + "type": "text", + "content": "18040" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "timising the representation to capture high fidelity is challenging due to the lack of correlation among the primitives. Recently, in addition to classical graphic primitives, neural network-based map representations are a promising alternative. iMAP [33] demonstrated the interesting properties of neural representation, such as sensible hole filling of unobserved geometry. Many recent approaches combine the classical and neural representations to capture finer details [8, 27, 46, 47]; however, the large amount of computation required for neural rendering makes the live operation of such systems challenging." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 210, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 210, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 210, + 289, + 715 + ], + "type": "text", + "content": "Differentiable Rendering: The classical method for creating a 3D representation was to unproject 2D observations into 3D space and to fuse them via weighted averaging [16, 23]. Such an averaging scheme suffers from over-smooth representation and lacks the expressiveness to capture high-quality details. To capture a scene with photorealistic quality, differentiable volumetric rendering [24] has recently been popularised with Neural Radiance Fields (NeRF) [17]. Using a single Multi-Layer Perceptron (MLP) as a scene representation, NeRF performs volume rendering by marching along pixel rays, querying the MLP for opacity and colour. Since volume rendering is naturally differentiable, the MLP representation is optimised to minimise the rendering loss using multiview information to achieve high-quality novel view synthesis. The main weakness of NeRF is its training speed. Recent developments have introduced explicit volume structures such as multi-resolution voxel grids [6, 14, 34] or hash functions [19] to improve performance. Interestingly, these projects demonstrate that the main contributor to high-quality novel view synthesis is not the neural network but rather differentiable volumetric rendering, and that it is possible to avoid the use of an MLP and yet achieve comparable rendering quality to NeRF [6]. However, even in these systems, per-pixel ray marching remains a significant bottleneck for rendering speed. This issue is particularly critical in SLAM, where immediate interaction with the map is essential for tracking. In contrast to NeRF, 3DGS performs differentiable rasterisation. Similar to regular graphics rasterisations, by iterating over the primitives to be rasterised rather than marching along rays, 3DGS leverages the natural sparsity of a 3D scene and achieves a representation which is expressive to capture high-fidelity 3D scenes while offering significantly faster rendering. Several works have applied 3D Gaussians and differentiable rendering to static scene capture [11, 38], and in particular more recent works utilise 3DGS and demonstrate superior results in vision tasks such as dynamic scene capture [15, 42, 44] and 3D generation [35, 45]. Our method adopts a Map-centric approach, utilising 3D Gaussians as the only SLAM representation. Similar to surfel-based SLAM, we dynamically allocate the 3D Gaussians, enabling us to model an arbitrary spatial dis" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 304, + 72, + 547, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 145 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 145 + ], + "type": "text", + "content": "tribution in the scene. Unlike other methods such as Elastic-Fusion [41] and PointFusion [9], however, by using differentiable rasterisation, our SLAM system can capture high-fidelity scene details and represent challenging object properties by direct optimisation against information from every pixel." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 306, + 156, + 362, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 156, + 362, + 167 + ], + "spans": [ + { + "bbox": [ + 306, + 156, + 362, + 167 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 175, + 418, + 189 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 175, + 418, + 189 + ], + "spans": [ + { + "bbox": [ + 306, + 175, + 418, + 189 + ], + "type": "text", + "content": "3.1. Gaussian Splatting" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 194, + 547, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 194, + 547, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 194, + 547, + 327 + ], + "type": "text", + "content": "Our SLAM representation is 3DGS, mapping the scene with a set of anisotropic Gaussians " + }, + { + "bbox": [ + 304, + 194, + 547, + 327 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 304, + 194, + 547, + 327 + ], + "type": "text", + "content": ". Each Gaussian " + }, + { + "bbox": [ + 304, + 194, + 547, + 327 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^i" + }, + { + "bbox": [ + 304, + 194, + 547, + 327 + ], + "type": "text", + "content": " contains optical properties: colour " + }, + { + "bbox": [ + 304, + 194, + 547, + 327 + ], + "type": "inline_equation", + "content": "c^i" + }, + { + "bbox": [ + 304, + 194, + 547, + 327 + ], + "type": "text", + "content": " and opacity " + }, + { + "bbox": [ + 304, + 194, + 547, + 327 + ], + "type": "inline_equation", + "content": "\\alpha^i" + }, + { + "bbox": [ + 304, + 194, + 547, + 327 + ], + "type": "text", + "content": ". For continuous 3D representation, the mean " + }, + { + "bbox": [ + 304, + 194, + 547, + 327 + ], + "type": "inline_equation", + "content": "\\mu_W^i" + }, + { + "bbox": [ + 304, + 194, + 547, + 327 + ], + "type": "text", + "content": " and covariance " + }, + { + "bbox": [ + 304, + 194, + 547, + 327 + ], + "type": "inline_equation", + "content": "\\Sigma_W^i" + }, + { + "bbox": [ + 304, + 194, + 547, + 327 + ], + "type": "text", + "content": ", defined in the world coordinate, represent the Gaussian's position and its ellipsoidal shape. For simplicity and speed, in our work we omit the spherical harmonics representing view-dependent radiance. Since 3DGS uses volume rendering, explicit extraction of the surface is not required. Instead, by splatting and blending " + }, + { + "bbox": [ + 304, + 194, + 547, + 327 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 304, + 194, + 547, + 327 + ], + "type": "text", + "content": " Gaussians, a pixel colour " + }, + { + "bbox": [ + 304, + 194, + 547, + 327 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_p" + }, + { + "bbox": [ + 304, + 194, + 547, + 327 + ], + "type": "text", + "content": " is synthesised:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 365, + 336, + 545, + 370 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 336, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 365, + 336, + 545, + 370 + ], + "type": "interline_equation", + "content": "\\mathcal {C} _ {p} = \\sum_ {i \\in \\mathcal {N}} c _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}). \\tag {1}", + "image_path": "a793a0975bf97aa67ee15150800010e9a8ab4622f157434ab8a25ba3bc26e6c9.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 378, + 547, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 378, + 547, + 475 + ], + "spans": [ + { + "bbox": [ + 304, + 378, + 547, + 475 + ], + "type": "text", + "content": "3DGS performs rasterisation, iterating over the Gaussians rather than marching along the camera rays, and hence, free spaces are ignored during rendering. During rasterisation, the contributions of " + }, + { + "bbox": [ + 304, + 378, + 547, + 475 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 378, + 547, + 475 + ], + "type": "text", + "content": " are decayed via a Gaussian function, based on the 2D Gaussian formed by splatting a 3D Gaussian. The 3D Gaussians " + }, + { + "bbox": [ + 304, + 378, + 547, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\pmb{\\mu}_W, \\pmb{\\Sigma}_W)" + }, + { + "bbox": [ + 304, + 378, + 547, + 475 + ], + "type": "text", + "content": " in world coordinates are related to the 2D Gaussians " + }, + { + "bbox": [ + 304, + 378, + 547, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\pmb{\\mu}_I, \\pmb{\\Sigma}_I)" + }, + { + "bbox": [ + 304, + 378, + 547, + 475 + ], + "type": "text", + "content": " on the image plane through a projective transformation:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 329, + 483, + 545, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 483, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 329, + 483, + 545, + 498 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\mu} _ {I} = \\pi (\\boldsymbol {T} _ {C W} \\cdot \\boldsymbol {\\mu} _ {W}), \\boldsymbol {\\Sigma} _ {I} = \\mathbf {J} \\mathbf {W} \\boldsymbol {\\Sigma} _ {W} \\mathbf {W} ^ {T} \\mathbf {J} ^ {T}, (2)", + "image_path": "cd72aac8117dee6877ecf93887e85d6584ad8ae7aa4ffcf50559fea22fc4bc54.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 506, + 547, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 547, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 547, + 616 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 506, + 547, + 616 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 304, + 506, + 547, + 616 + ], + "type": "text", + "content": " is the projection operation and " + }, + { + "bbox": [ + 304, + 506, + 547, + 616 + ], + "type": "inline_equation", + "content": "T_{CW} \\in SE(3)" + }, + { + "bbox": [ + 304, + 506, + 547, + 616 + ], + "type": "text", + "content": " is the camera pose of the viewpoint. " + }, + { + "bbox": [ + 304, + 506, + 547, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{J}" + }, + { + "bbox": [ + 304, + 506, + 547, + 616 + ], + "type": "text", + "content": " is the Jacobian of the linear approximation of the projective transformation and " + }, + { + "bbox": [ + 304, + 506, + 547, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 304, + 506, + 547, + 616 + ], + "type": "text", + "content": " is the rotational component of " + }, + { + "bbox": [ + 304, + 506, + 547, + 616 + ], + "type": "inline_equation", + "content": "T_{CW}" + }, + { + "bbox": [ + 304, + 506, + 547, + 616 + ], + "type": "text", + "content": ". This formulation enables the 3D Gaussians to be differentiable and the blending operation provides gradient flow to the Gaussians. Using first-order gradient descent [12], Gaussians gradually refines both their optic and geometric parameters to represent the captured scene with high fidelity." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 623, + 455, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 455, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 455, + 635 + ], + "type": "text", + "content": "3.2. Camera Pose Optimisation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": "To achieve accurate tracking, we typically require at least 50 iterations of gradient descent per frame. This requirement emphasises the necessity of a representation with computationally efficient view synthesis and gradient computation, making the choice of 3D representation a crucial part of designing a SLAM system." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18041" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 70, + 548, + 258 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 548, + 258 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 548, + 258 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 548, + 258 + ], + "type": "image", + "image_path": "97ceacd1e19adb68856325d2133821f4017de23d7ca01b34d044a0e689e68a9c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 258, + 547, + 282 + ], + "lines": [ + { + "bbox": [ + 46, + 258, + 547, + 282 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 547, + 282 + ], + "type": "text", + "content": "Figure 2. SLAM System Overview: Our SLAM system uses 3D Gaussians as the only representation, unifying all components of SLAM, including tracking, mapping, keyframe management, and novel view synthesis." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 302, + 287, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 302, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 302, + 287, + 361 + ], + "type": "text", + "content": "In order to avoid the overhead of automatic differentiation, 3DGS implements rasterisation with CUDA with derivatives for all parameters calculated explicitly. Since rasterisation is performance critical, we similarly derive the camera Jacobians explicitly." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 361, + 287, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 287, + 410 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 287, + 410 + ], + "type": "text", + "content": "To the best of our knowledge, we provide the first analytical Jacobian of " + }, + { + "bbox": [ + 46, + 361, + 287, + 410 + ], + "type": "inline_equation", + "content": "\\pmb{SE}(3)" + }, + { + "bbox": [ + 46, + 361, + 287, + 410 + ], + "type": "text", + "content": " camera pose with respect to the 3D Gaussians used in EWA splatting [48] and 3DGS. This opens up new applications of 3DGS beyond SLAM." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 410, + 287, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 410, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 410, + 287, + 471 + ], + "type": "text", + "content": "We use Lie algebra to derive the minimal Jacobians, ensuring that the dimensionality of the Jacobians matches the degrees of freedom, eliminating any redundant computations. The terms of Eq. (2) are differentiable with respect to the camera pose " + }, + { + "bbox": [ + 46, + 410, + 287, + 471 + ], + "type": "inline_equation", + "content": "\\pmb{T}_{CW}" + }, + { + "bbox": [ + 46, + 410, + 287, + 471 + ], + "type": "text", + "content": "; using the chain rule:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 63, + 473, + 287, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 473, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 63, + 473, + 287, + 498 + ], + "type": "interline_equation", + "content": "\\frac {\\partial \\boldsymbol {\\mu} _ {I}}{\\partial \\boldsymbol {T} _ {C W}} = \\frac {\\partial \\boldsymbol {\\mu} _ {I}}{\\partial \\boldsymbol {\\mu} _ {C}} \\frac {\\mathcal {D} \\boldsymbol {\\mu} _ {C}}{\\mathcal {D} \\boldsymbol {T} _ {C W}}, \\tag {3}", + "image_path": "522dd74e821d5758cfc0c1f085129ad69f99bc8ebb59b7af9c5a4d09dd4939eb.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 63, + 500, + 287, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 500, + 287, + 526 + ], + "spans": [ + { + "bbox": [ + 63, + 500, + 287, + 526 + ], + "type": "interline_equation", + "content": "\\frac {\\partial \\boldsymbol {\\Sigma} _ {I}}{\\partial \\boldsymbol {T} _ {C W}} = \\frac {\\partial \\boldsymbol {\\Sigma} _ {I}}{\\partial \\mathbf {J}} \\frac {\\partial \\mathbf {J}}{\\partial \\boldsymbol {\\mu} _ {C}} \\frac {\\mathcal {D} \\boldsymbol {\\mu} _ {C}}{\\mathcal {D} \\boldsymbol {T} _ {C W}} + \\frac {\\partial \\boldsymbol {\\Sigma} _ {I}}{\\partial \\mathbf {W}} \\frac {\\mathcal {D} \\mathbf {W}}{\\mathcal {D} \\boldsymbol {T} _ {C W}}. \\tag {4}", + "image_path": "6b69331017d095e61d9a7bea0144376d6019b53a149945486ad9687cd06ab06b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 529, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 529, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 529, + 287, + 590 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 529, + 287, + 590 + ], + "type": "inline_equation", + "content": "T_{CW}" + }, + { + "bbox": [ + 46, + 529, + 287, + 590 + ], + "type": "text", + "content": " represents the 3D position of Gaussian in the camera coordinate. We take the derivatives on the manifold to derive minimal parameterisation. Borrowing the notation from [30], let " + }, + { + "bbox": [ + 46, + 529, + 287, + 590 + ], + "type": "inline_equation", + "content": "T \\in SE(3)" + }, + { + "bbox": [ + 46, + 529, + 287, + 590 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 529, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\tau \\in \\mathfrak{se}(3)" + }, + { + "bbox": [ + 46, + 529, + 287, + 590 + ], + "type": "text", + "content": ". We define the partial derivative on the manifold as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 593, + 287, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 593, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 61, + 593, + 287, + 617 + ], + "type": "interline_equation", + "content": "\\frac {\\mathcal {D} f (\\boldsymbol {T})}{\\mathcal {D} \\boldsymbol {T}} \\triangleq \\lim _ {\\tau \\rightarrow 0} \\frac {\\operatorname {L o g} (f (\\operatorname {E x p} (\\tau) \\circ \\boldsymbol {T}) \\circ f (\\boldsymbol {T}) ^ {- 1})}{\\tau}, \\tag {5}", + "image_path": "a53bb2d73f23bef441fab2c219fc6c09e4917be684aec5e22c28d5f1cca6a04d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 621, + 287, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 621, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 46, + 621, + 287, + 657 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 621, + 287, + 657 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 46, + 621, + 287, + 657 + ], + "type": "text", + "content": " is a group composition, and Exp, Log are the exponential and logarithmic mappings between Lie algebra and Lie Group. With this, we derive the following:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 671, + 287, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 287, + 711 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 287, + 711 + ], + "type": "interline_equation", + "content": "\\frac {\\mathcal {D} \\boldsymbol {\\mu} _ {C}}{\\mathcal {D} \\boldsymbol {T} _ {C W}} = \\left[ \\begin{array}{l l} \\boldsymbol {I} & - \\boldsymbol {\\mu} _ {C} ^ {\\times} \\end{array} \\right], \\frac {\\mathcal {D} \\mathbf {W}}{\\mathcal {D} \\boldsymbol {T} _ {C W}} = \\left[ \\begin{array}{l l} \\mathbf {0} & - \\mathbf {W} _ {:, 1} ^ {\\times} \\\\ \\mathbf {0} & - \\mathbf {W} _ {:, 2} ^ {\\times} \\\\ \\mathbf {0} & - \\mathbf {W} _ {:, 3} ^ {\\times} \\end{array} \\right] ， \\tag {6}", + "image_path": "59024cc4841e4834d88347310f56703ccf1b61e14d2f20fb3e55e3819e5bf09f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 302, + 545, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 302, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 305, + 302, + 545, + 327 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 302, + 545, + 327 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 305, + 302, + 545, + 327 + ], + "type": "text", + "content": " denotes the skew symmetric matrix of a 3D vector, and " + }, + { + "bbox": [ + 305, + 302, + 545, + 327 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{:,i}" + }, + { + "bbox": [ + 305, + 302, + 545, + 327 + ], + "type": "text", + "content": " refers to the " + }, + { + "bbox": [ + 305, + 302, + 545, + 327 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 302, + 545, + 327 + ], + "type": "text", + "content": "th column of the matrix." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 334, + 361, + 345 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 334, + 361, + 345 + ], + "spans": [ + { + "bbox": [ + 306, + 334, + 361, + 345 + ], + "type": "text", + "content": "3.3. SLAM" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 353, + 545, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 353, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 305, + 353, + 545, + 399 + ], + "type": "text", + "content": "In this section, we present details of full SLAM framework. The overview of the system is summarised in Fig. 2. Please refer to the supplementary material for the further parameter details." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 416, + 378, + 429 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 416, + 378, + 429 + ], + "spans": [ + { + "bbox": [ + 306, + 416, + 378, + 429 + ], + "type": "text", + "content": "3.3.1 Tracking" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 436, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 436, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 305, + 436, + 545, + 472 + ], + "type": "text", + "content": "In tracking only the current camera pose is optimised, without updates to the map representation. In the monocular case, we minimise the following photometric residual:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 364, + 480, + 545, + 496 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 480, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 364, + 480, + 545, + 496 + ], + "type": "interline_equation", + "content": "E _ {p h o} = \\left\\| I (\\mathcal {G}, \\boldsymbol {T} _ {C W}) - \\bar {I} \\right\\| _ {1}, \\tag {7}", + "image_path": "27bf125bfc1118307b2612a39059ab9712d3740fb436478f86702e5c62e7a2b6.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 504, + 545, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 504, + 545, + 528 + ], + "spans": [ + { + "bbox": [ + 306, + 504, + 545, + 528 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 504, + 545, + 528 + ], + "type": "inline_equation", + "content": "I(\\mathcal{G},\\pmb{T}_{CW})" + }, + { + "bbox": [ + 306, + 504, + 545, + 528 + ], + "type": "text", + "content": " renders the Gaussians " + }, + { + "bbox": [ + 306, + 504, + 545, + 528 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 306, + 504, + 545, + 528 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 306, + 504, + 545, + 528 + ], + "type": "inline_equation", + "content": "T_{CW}" + }, + { + "bbox": [ + 306, + 504, + 545, + 528 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 306, + 504, + 545, + 528 + ], + "type": "inline_equation", + "content": "\\bar{I}" + }, + { + "bbox": [ + 306, + 504, + 545, + 528 + ], + "type": "text", + "content": " is an observed image." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 529, + 545, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 529, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 306, + 529, + 545, + 564 + ], + "type": "text", + "content": "We further optimise affine brightness parameters for varying exposure. When depth observations are available, we define the geometric residual as:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 361, + 573, + 545, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 573, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 361, + 573, + 545, + 590 + ], + "type": "interline_equation", + "content": "E _ {g e o} = \\left\\| D (\\mathcal {G}, \\boldsymbol {T} _ {C W}) - \\bar {D} \\right\\| _ {1}, \\tag {8}", + "image_path": "20d96b32f0d3be973ac450fdf2f11a0585a7c967bdaae77394688859f6bc35a7.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 597, + 545, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 597, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 305, + 597, + 545, + 658 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 597, + 545, + 658 + ], + "type": "inline_equation", + "content": "D(\\mathcal{G},\\mathbf{T}_{CW})" + }, + { + "bbox": [ + 305, + 597, + 545, + 658 + ], + "type": "text", + "content": " is depth rasterisation and " + }, + { + "bbox": [ + 305, + 597, + 545, + 658 + ], + "type": "inline_equation", + "content": "\\bar{D}" + }, + { + "bbox": [ + 305, + 597, + 545, + 658 + ], + "type": "text", + "content": " is the observed depth. Rather than simply using the depth measurements to initialise the Gaussians, we minimise both photometric and geometric residuals: " + }, + { + "bbox": [ + 305, + 597, + 545, + 658 + ], + "type": "inline_equation", + "content": "\\lambda_{pho}E_{pho} + (1 - \\lambda_{pho})E_{geo}" + }, + { + "bbox": [ + 305, + 597, + 545, + 658 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 597, + 545, + 658 + ], + "type": "inline_equation", + "content": "\\lambda_{pho}" + }, + { + "bbox": [ + 305, + 597, + 545, + 658 + ], + "type": "text", + "content": " is a hyperparameter." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 658, + 545, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 658, + 545, + 682 + ], + "spans": [ + { + "bbox": [ + 305, + 658, + 545, + 682 + ], + "type": "text", + "content": "As in Eq. (1), per-pixel depth is rasterised by alphabetending:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 365, + 682, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 682, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 365, + 682, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {D} _ {p} = \\sum_ {i \\in \\mathcal {N}} z _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {9}", + "image_path": "0955313f1f39f06416da87d9e41d81a6c6b642198a606049736c5dc394e6d8f8.jpg" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18042" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": " is the distance to the mean " + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_{W}" + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": " of Gaussian " + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": " along the camera ray. We derive analytical Jacobians for the camera pose optimisation in a similar manner to Eq. (3), (4)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 123, + 133, + 136 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 123, + 133, + 136 + ], + "spans": [ + { + "bbox": [ + 47, + 123, + 133, + 136 + ], + "type": "text", + "content": "3.3.2 Keyframing" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 142, + 288, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 142, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 142, + 288, + 239 + ], + "type": "text", + "content": "Since using all the images from a video stream to jointly optimise the Gaussians and camera poses online is infeasible, we maintain a small window " + }, + { + "bbox": [ + 46, + 142, + 288, + 239 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_k" + }, + { + "bbox": [ + 46, + 142, + 288, + 239 + ], + "type": "text", + "content": " consisting of carefully selected keyframes based on inter-frame covisibility. Ideal keyframe management will select non-redundant keyframes observing the same area, spanning a wide baseline to provide better multiview constraints. The parameters are detailed in supplementary." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 253, + 289, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 253, + 289, + 408 + ], + "spans": [ + { + "bbox": [ + 46, + 253, + 289, + 408 + ], + "type": "text", + "content": "Selection and Management Every tracked frame is checked for keyframe registration based on our simple yet effective criteria. We measure the可视ibility by measuring the intersection over union of the observed Gaussians between the current frame " + }, + { + "bbox": [ + 46, + 253, + 289, + 408 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 253, + 289, + 408 + ], + "type": "text", + "content": " and the last keyframe " + }, + { + "bbox": [ + 46, + 253, + 289, + 408 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 253, + 289, + 408 + ], + "type": "text", + "content": ". If the可视ibility drops below a threshold, or if the relative translation " + }, + { + "bbox": [ + 46, + 253, + 289, + 408 + ], + "type": "inline_equation", + "content": "t_{ij}" + }, + { + "bbox": [ + 46, + 253, + 289, + 408 + ], + "type": "text", + "content": " is large with respect to the median depth, frame " + }, + { + "bbox": [ + 46, + 253, + 289, + 408 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 253, + 289, + 408 + ], + "type": "text", + "content": " is registered as a keyframe. For efficiency, we maintain only a small number of keyframes in the current window " + }, + { + "bbox": [ + 46, + 253, + 289, + 408 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_k" + }, + { + "bbox": [ + 46, + 253, + 289, + 408 + ], + "type": "text", + "content": " following the keyframe management heuristics of DSO [4]. The main difference is that a keyframe is removed from the current window if the overlap coefficient with the latest keyframe drops below a threshold." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 423, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 423, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 287, + 533 + ], + "type": "text", + "content": "Gaussian Covisibility An accurate estimate of covisibility simplifies keyframe selection and management. 3DGS respects visibility ordering since the 3D Gaussians are sorted along the camera ray. This property is desirable for covisibility estimation as occlusions are handled by design. A Gaussian is marked to be visible from a view if used in the rasterisation and if the ray's accumulated " + }, + { + "bbox": [ + 46, + 423, + 287, + 533 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 423, + 287, + 533 + ], + "type": "text", + "content": " has not yet reached 0.5. This enables our estimated covisibility to handle occlusions without requiring additional heuristics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 546, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 546, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 546, + 287, + 665 + ], + "type": "text", + "content": "Gaussian Insertion and Pruning At every keyframe, new Gaussians are inserted into the scene to capture newly visible scene elements and to refine the fine details. When depth measurements are available, Gaussian means " + }, + { + "bbox": [ + 46, + 546, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\mu_W" + }, + { + "bbox": [ + 46, + 546, + 287, + 665 + ], + "type": "text", + "content": " are initialised by back-projecting the depth. In the monocular case, we render the depth at the current frame. For pixels with depth estimates, " + }, + { + "bbox": [ + 46, + 546, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\mu_W" + }, + { + "bbox": [ + 46, + 546, + 287, + 665 + ], + "type": "text", + "content": " are initialised around those depths with low variance; for pixels without the depth estimates, we initialise " + }, + { + "bbox": [ + 46, + 546, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\mu_W" + }, + { + "bbox": [ + 46, + 546, + 287, + 665 + ], + "type": "text", + "content": " around the median depth of the rendered image with high variance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "In the monocular case, the positions of many newly inserted Gaussians are incorrect. While the majority will quickly vanish during optimisation as they violate multiview consistency, we further prune the excess Gaussians by" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "type": "text", + "content": "checking the visibility amongst the current window " + }, + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_k" + }, + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "type": "text", + "content": ". If the Gaussians inserted within the last 3 keyframes are unobserved by at least 3 other frames, we prune them out as they are geometrically unstable." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 135, + 379, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 135, + 379, + 148 + ], + "spans": [ + { + "bbox": [ + 306, + 135, + 379, + 148 + ], + "type": "text", + "content": "3.3.3 Mapping" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 153, + 547, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 153, + 547, + 297 + ], + "spans": [ + { + "bbox": [ + 304, + 153, + 547, + 297 + ], + "type": "text", + "content": "The purpose of mapping is to maintain a coherent 3D structure and to optimise the newly inserted Gaussians. During mapping, the keyframes in " + }, + { + "bbox": [ + 304, + 153, + 547, + 297 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_k" + }, + { + "bbox": [ + 304, + 153, + 547, + 297 + ], + "type": "text", + "content": " are used to reconstruct currently visible regions. Additionally, two random past keyframes " + }, + { + "bbox": [ + 304, + 153, + 547, + 297 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_r" + }, + { + "bbox": [ + 304, + 153, + 547, + 297 + ], + "type": "text", + "content": " are selected per iteration to avoid forgetting the global map. Rasterisation of 3DGS imposes no constraint on the Gaussians along the viewing ray direction, even with a depth observation. This is not a problem when sufficient carefully selected viewpoints are provided (e.g. in the novel view synthesis case); however, in continuous SLAM this causes many artefacts, making tracking challenging. We therefore introduce an isotropic regularisation:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 372, + 305, + 545, + 340 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 305, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 372, + 305, + 545, + 340 + ], + "type": "interline_equation", + "content": "E _ {i s o} = \\sum_ {i = 1} ^ {| \\mathcal {G} |} \\| \\mathbf {s} _ {i} - \\tilde {\\mathbf {s}} _ {i} \\cdot \\mathbf {1} \\| _ {1} \\tag {10}", + "image_path": "5f51f86295fb616695a663364d14309310b0f5f023b6456ee1ec73ed5535f984.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 347, + 547, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 347, + 547, + 443 + ], + "spans": [ + { + "bbox": [ + 304, + 347, + 547, + 443 + ], + "type": "text", + "content": "to penalise the scaling parameters " + }, + { + "bbox": [ + 304, + 347, + 547, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_i" + }, + { + "bbox": [ + 304, + 347, + 547, + 443 + ], + "type": "text", + "content": " (i.e. stretch of the ellipsoid) by its difference to the mean " + }, + { + "bbox": [ + 304, + 347, + 547, + 443 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{s}}_i" + }, + { + "bbox": [ + 304, + 347, + 547, + 443 + ], + "type": "text", + "content": ". As shown in Fig 3, this encourages sphericality, and avoids the problem of Gaussians which are highly elongated along the viewing direction creating artefacts. Let the union of the keyframes in the current window and the randomly selected one be " + }, + { + "bbox": [ + 304, + 347, + 547, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{W} = \\mathcal{W}_k \\cup \\mathcal{W}_r" + }, + { + "bbox": [ + 304, + 347, + 547, + 443 + ], + "type": "text", + "content": ". For mapping, we solve the following problem:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 345, + 452, + 545, + 482 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 452, + 545, + 482 + ], + "spans": [ + { + "bbox": [ + 345, + 452, + 545, + 482 + ], + "type": "interline_equation", + "content": "\\min _ {\\boldsymbol {T} _ {C W} ^ {k} \\in \\boldsymbol {S E} (3), \\mathcal {G}, \\forall k \\in \\mathcal {W}} \\sum_ {\\forall k \\in \\mathcal {W}} E _ {p h o} ^ {k} + \\lambda_ {i s o} E _ {i s o}. \\tag {11}", + "image_path": "aa7c0b92a3821a919e804c6cf8e929770fabbd7415d654a5ef37aa58d289ce12.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 489, + 545, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 489, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 304, + 489, + 545, + 514 + ], + "type": "text", + "content": "If depth observations are available, as in tracking, geometric residuals Eq. (8) are added to the optimisation problem." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 524, + 378, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 524, + 378, + 536 + ], + "spans": [ + { + "bbox": [ + 306, + 524, + 378, + 536 + ], + "type": "text", + "content": "4. Evaluation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 544, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 545, + 616 + ], + "type": "text", + "content": "We conduct a comprehensive evaluation of our system across a range of both real and synthetic datasets. Additionally, we perform an ablation study to justify our design choices. Finally, we present qualitative results of our system operating live using a monocular camera, illustrating its practicality and high fidelity reconstruction." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 623, + 422, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 422, + 636 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 422, + 636 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "type": "text", + "content": "Datasets For our quantitative analysis, we evaluate our method on the TUM RGB-D dataset [32] (3 sequences) and the Replica dataset [31] (8 sequences), following the evaluation in [33]. For qualitative results, we use self-captured real-world sequences recorded by Intel Realsense d455. Since the Replica dataset is designed for RGB-D SLAM" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18043" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 285, + 165 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 285, + 165 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 285, + 165 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 285, + 165 + ], + "type": "image", + "image_path": "d3b400550f268e328fb72f93546cffb3b56e822362318d4823f1d64a4bd74d87.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 49, + 167, + 171, + 238 + ], + "blocks": [ + { + "bbox": [ + 49, + 167, + 171, + 238 + ], + "lines": [ + { + "bbox": [ + 49, + 167, + 171, + 238 + ], + "spans": [ + { + "bbox": [ + 49, + 167, + 171, + 238 + ], + "type": "image", + "image_path": "cfb4ab71fd891c902d7dffe34369c688ecaeb250f6f3e0253f77581c79108702.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 239, + 286, + 317 + ], + "lines": [ + { + "bbox": [ + 47, + 239, + 286, + 317 + ], + "spans": [ + { + "bbox": [ + 47, + 239, + 286, + 317 + ], + "type": "text", + "content": "Figure 3. Effect of isotropic regularisation: Top: Rendering close to a training view (looking at the keyboard). Bottom: Rendering 3D Gaussians far from the training views (view from a side of the keyboard) without (left) and with (right) the isotropic loss. When the photometric constraints are insufficient, the Gaussians tend to elongate along the viewing direction, creating artefacts in the novel views, and affecting the camera tracking." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 173, + 167, + 286, + 238 + ], + "blocks": [ + { + "bbox": [ + 173, + 167, + 286, + 238 + ], + "lines": [ + { + "bbox": [ + 173, + 167, + 286, + 238 + ], + "spans": [ + { + "bbox": [ + 173, + 167, + 286, + 238 + ], + "type": "image", + "image_path": "d4232cbfd10cd50adaf9060720a2b68ce9c4820b6d9dea29265c927bb26fe56b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 324, + 286, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 324, + 286, + 371 + ], + "spans": [ + { + "bbox": [ + 47, + 324, + 286, + 371 + ], + "type": "text", + "content": "evaluation, it contains challenging purely rotational camera motions. We hence use the Replica dataset for RGB-D evaluation only. The TUM RGB-D dataset is used for both monocular and RGB-D evaluation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 390, + 286, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 390, + 286, + 521 + ], + "spans": [ + { + "bbox": [ + 47, + 390, + 286, + 521 + ], + "type": "text", + "content": "Implementation Details We run our SLAM on a desktop with Intel Core i9 12900K 3.50GHz and a single NVIDIA GeForce RTX 4090. We present results from our multiprocessing implementation aimed at real-time applications. For a fair comparison with other methods on Replica, we additionally report result for single-process implementation which performs more mapping iterations. As with 3DGS, time-critical rasterisation and gradient computation are implemented using CUDA. The rest of the SLAM pipeline is developed with PyTorch. Details of hyperparameters are provided in the supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 540, + 286, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 540, + 286, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 540, + 286, + 658 + ], + "type": "text", + "content": "Metrics For camera tracking accuracy, we report the Root Mean Square Error (RMSE) of the Absolute Trajectory Error (ATE) of the keyframes. To evaluate map quality, we report standard photometric rendering quality metrics (PSNR, SSIM and LPIPS) following the evaluation protocol used in [27]. To evaluate the map quality, on every fifth frame, rendering metrics are computed. We exclude the keyframes (training views). We report the average across three runs for all our evaluations. In the tables, the best result is in bold, and the second best is underlined." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "text", + "content": "Baseline Methods We primarily benchmark our SLAM method against other approaches that, like ours, do not have explicit loop closure. In monocular settings, we compare" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": "with state-of-the-art classical and learning-based direct visual odometry (VO) methods. Specifically, we compare DSO [4], DepthCov [3], and DROID-SLAM [36] in VO configurations. These methods are selected based on their public reporting of results on the benchmark (TUM dataset) or the availability of their source code for getting the benchmark result. Since one of our focuses is the online scale estimation under monocular scale ambiguity, the method which uses ground truth poses for the system initialisation such as [13] is not considered for the comparison. In the RGB-D case, we compare against neural-implicit SLAM methods [7, 8, 27, 33, 39, 43, 46] which are also map-centric, rendering-based and do not perform loop closure." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 236, + 441, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 236, + 441, + 248 + ], + "spans": [ + { + "bbox": [ + 306, + 236, + 441, + 248 + ], + "type": "text", + "content": "4.2. Quantitative Evaluation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 255, + 545, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 545, + 350 + ], + "type": "text", + "content": "Camera Tracking Accuracy Table 1 shows the tracking results on the TUM RGB-D dataset. In the monocular setting, our method surpasses other baselines without requiring any deep priors. Furthermore, our performance is comparable to systems which perform explicit loop closure. This clearly highlights that there still remains potential for enhancing the tracking of monocular SLAM by exploring fundamental SLAM representations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 350, + 545, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 350, + 545, + 505 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 545, + 505 + ], + "type": "text", + "content": "Our RGB-D method shows better performance than any other baseline method. Notably, our system surpasses ORB-SLAM in the fr1 sequences, narrowing the gap between Map-centric SLAM and the state-of-the-art sparse frame-centric methods. Table 2 reports results on the synthetic Replica dataset. Our single-process implementation shows competitive performance and achieves the best result in 4 out of 8 sequences. Our multi-process implementation which performs fewer mapping iterations still performs comparably. In contrast to other methods, our system demonstrates higher performance on real-world data, by optimising the Gaussian positions to compensate for the sensor noise." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": "Novel View Rendering Table 5 summarises the novel view rendering performance of our method with RGB-D input. We consistently show the best performance across most sequences and is least second best. Our rendering FPS is hundreds of times faster than other methods, offering a significant advantage for applications which require real-time map interaction. While Point-SLAM is competitive, that method focuses on view synthesis rather than novel-view synthesis. Their view synthesis is conditional on the availability of depth due to the depth-guided ray-sampling, making novel-view synthesis challenging. On the other hand, our rasterisation-based approach does not require depth guidance and achieves efficient, high-quality, novel view synthesis. Fig. 4 provides a qualitative comparison of the rendering of ours and Point-SLAM (with depth guidance)." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18044" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 286, + 255 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 286, + 255 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 286, + 255 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 286, + 255 + ], + "type": "table", + "html": "
InputLoop-closureMethodfr1/deskfr2/xyzfr3/officeAvg.
Monocularw/oDSO [4]22.41.109.5011.0
DROID-VO [36]5.2010.77.307.73
DepthCov-VO [3]5.601.2068.825.2
Ours3.784.603.503.96
w/DROID-SLAM [36]1.800.502.801.70
ORB-SLAM2 [20]1.900.602.401.60
RGB-Dw/oiMAP [33]4.902.005.804.23
NICE-SLAM [46]4.266.193.874.77
DI-Fusion [7]4.402.005.804.07
Vox-Fusion [43]3.521.4926.0110.34
ESLAM [8]2.471.112.422.00
Co-SLAM [39]2.401.702.402.17
Point-SLAM [27]4.341.313.483.04
Ours1.501.441.491.47
w/BAD-SLAM [29]1.701.101.701.50
Kintinous [40]3.702.903.003.20
ORB-SLAM2 [20]1.600.401.001.00
", + "image_path": "a9996696ef891870be6d242ff078e8edc16684df8dc79aaa15611698e0ff80d0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 327, + 285, + 394 + ], + "blocks": [ + { + "bbox": [ + 47, + 256, + 287, + 322 + ], + "lines": [ + { + "bbox": [ + 47, + 256, + 287, + 322 + ], + "spans": [ + { + "bbox": [ + 47, + 256, + 287, + 322 + ], + "type": "text", + "content": "Table 1. Camera tracking result on TUM for monocular and RGB-D. ATE RMSE in cm is reported. In both monocular and RGB-D cases, we achieve state-of-the-art performance. In particular, in the monocular case, not only do we outperform systems which use deep prior, but we achieve comparable performance with many of the RGB-D systems." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 327, + 285, + 394 + ], + "lines": [ + { + "bbox": [ + 50, + 327, + 285, + 394 + ], + "spans": [ + { + "bbox": [ + 50, + 327, + 285, + 394 + ], + "type": "table", + "html": "
Methodr0r1r2o0o1o2o3o4Avg.
iMAP [33]3.122.542.311.691.033.994.051.932.58
NICE-SLAM0.971.311.070.881.001.061.101.131.07
Vox-Fusion [43]1.374.701.478.482.042.581.112.943.09
ESLAM [8]0.710.700.520.570.550.580.720.630.63
Point-SLAM [27]0.610.410.370.380.480.540.690.720.53
Ours0.440.320.310.440.520.230.172.250.58
Ours (sp)0.330.220.290.360.190.250.120.810.32
", + "image_path": "b649b474e53fb76a34b8433f71bb29fb6db2398a96fe38c4b5ba8401a56f8e50.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 67, + 455, + 266, + 517 + ], + "blocks": [ + { + "bbox": [ + 47, + 396, + 287, + 452 + ], + "lines": [ + { + "bbox": [ + 47, + 396, + 287, + 452 + ], + "spans": [ + { + "bbox": [ + 47, + 396, + 287, + 452 + ], + "type": "text", + "content": "Table 2. Camera tracking result on Replica for RGB-D SLAM. ATE RMSE in cm is reported. We achieve best performance across most sequences. Here, Ours is our multi-process implementation and Ours (sp) is the single-process implementation which ensures a certain amount of mapping iteration similar to other works." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 455, + 266, + 517 + ], + "lines": [ + { + "bbox": [ + 67, + 455, + 266, + 517 + ], + "spans": [ + { + "bbox": [ + 67, + 455, + 266, + 517 + ], + "type": "table", + "html": "
InputMethodfr1/deskfr2/xyzfr3/officeAvg.
Monow/o \\( E_{iso} \\)4.164.665.734.83
w/o kf selection13.24.368.658.73
Ours3.784.603.503.96
RGB-Dw/o \\( E_{geo} \\)2.390.624.982.66
w/o kf selection1.641.492.601.90
Ours1.501.441.491.47
", + "image_path": "02faa4e7e5d5616fe9bdd110731759aa7590ebab795b0191b4136c805e81fe44.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 49, + 567, + 285, + 594 + ], + "blocks": [ + { + "bbox": [ + 47, + 518, + 287, + 563 + ], + "lines": [ + { + "bbox": [ + 47, + 518, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 518, + 287, + 563 + ], + "type": "text", + "content": "Table 3. Ablation Study on TUM RGB-D dataset. We analyse the usefulness of isotropic regularisation, geometric residual, and keyframe selection to our SLAM system. Further isotropic regularisation ablation is available in supplementary." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 567, + 285, + 594 + ], + "lines": [ + { + "bbox": [ + 49, + 567, + 285, + 594 + ], + "spans": [ + { + "bbox": [ + 49, + 567, + 285, + 594 + ], + "type": "table", + "html": "
Memory Usage [MB]
iMAP [33]NICE-SLAM [46]Co-SLAM [39]Ours (Mono)Ours (RGB-D)
0.8MB40.3.4MB6.4MB2.6MB3.97MB
", + "image_path": "9f9d2a943d2f1cba74ad69cf8453bbee9526ebe635285116340c5a7e66847976.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 624, + 287, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 624, + 287, + 708 + ], + "spans": [ + { + "bbox": [ + 46, + 624, + 287, + 708 + ], + "type": "text", + "content": "Ablative Analysis In Table 3, we perform ablation to confirm our design choices. Isotropic regularisation and geometric residual improve the tracking of monocular and RGB-D SLAM respectively, as they aid in constraining the geometry when photometric signals are weak. For both cases, keyframe selection significantly improves systems performance, as it automatically chooses suitable keyframes" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 309, + 70, + 544, + 117 + ], + "blocks": [ + { + "bbox": [ + 47, + 595, + 286, + 618 + ], + "lines": [ + { + "bbox": [ + 47, + 595, + 286, + 618 + ], + "spans": [ + { + "bbox": [ + 47, + 595, + 286, + 618 + ], + "type": "text", + "content": "Table 4. Memory Analysis on TUM RGB-D dataset. The baseline numbers are computed from the parameter numbers in [39]" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 70, + 544, + 117 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 544, + 117 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 544, + 117 + ], + "type": "table", + "html": "
MethodPSNR[db]↑SSIM↑LPIPS↓Rendering FPS
NICE-SLAM[46]24.420.8090.2330.54
Vox-Fusion[43]24.410.8010.2362.17
Point-SLAM [27]35.170.9750.1241.33
ours38.940.9680.070769
", + "image_path": "7573fad4bc6edfc0a0c23e59f6984cac66f8b7ab0f9ddc04ec6efceae58797bf.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 308, + 180, + 544, + 378 + ], + "blocks": [ + { + "bbox": [ + 308, + 180, + 544, + 378 + ], + "lines": [ + { + "bbox": [ + 308, + 180, + 544, + 378 + ], + "spans": [ + { + "bbox": [ + 308, + 180, + 544, + 378 + ], + "type": "image", + "image_path": "22e58c9c0ffa4fd7aec784b64abf8ac6b021bddd9689fd072931ef9cc292d6ff.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 380, + 545, + 403 + ], + "lines": [ + { + "bbox": [ + 306, + 380, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 306, + 380, + 545, + 403 + ], + "type": "text", + "content": "Figure 4. Rendering examples on Replica. Point-SLAM struggle with rendering fine details due to the stochastic ray sampling." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 410, + 545, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 410, + 545, + 494 + ], + "spans": [ + { + "bbox": [ + 305, + 410, + 545, + 494 + ], + "type": "text", + "content": "based on our occlusion-aware keyframe selection and management. We further compare the memory usage of different 3D representations in Table 4. MLP-based iMAP is clearly more memory efficient, but it struggles to express high-fidelity 3D scenes due to the limited capacity of small MLP. Compared with a voxel grid of features used in NICE-SLAM, our method uses significantly less memory." + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 320, + 513, + 532, + 569 + ], + "blocks": [ + { + "bbox": [ + 305, + 118, + 545, + 173 + ], + "lines": [ + { + "bbox": [ + 305, + 118, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 305, + 118, + 545, + 173 + ], + "type": "text", + "content": "Table 5. Average rendering performance on Replica (RGB-D). Our method outperforms most of the rendering metrics compared to existing methods. Note that Point-SLAM uses ground-truth depth to guide sampling along rays. The full detail is available in supplementary." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 320, + 513, + 532, + 569 + ], + "lines": [ + { + "bbox": [ + 320, + 513, + 532, + 569 + ], + "spans": [ + { + "bbox": [ + 320, + 513, + 532, + 569 + ], + "type": "table", + "html": "
Methodseq1seq2seq3Avg.
Neural SDF (Hash Grid)0.130.150.160.14
Neural SDF (MLP)0.400.380.220.33
Ours w/o depth0.820.910.650.79
Ours w/ depth0.831.00.650.82
", + "image_path": "4d957e0ab619b8b97d7cd5b1b7b0d691a60424ff86934ad6b629c2c474e6fa34.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 571, + 545, + 604 + ], + "lines": [ + { + "bbox": [ + 305, + 571, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 305, + 571, + 545, + 604 + ], + "type": "text", + "content": "Table 6. Camera convergence analysis. We report the ratio of successful camera convergence for the different sequences, across different differentiable 3D representations." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Convergence Basin Analysis In our SLAM experiments, we discovered that 3D Gaussian maps have a notably large convergence basin for camera localisation. To investigate further, we conducted a convergence funnel analysis, an evaluation methodology proposed in [18] and used in [22]. Here, we train a 3D representation (e.g. 3DGS) using 9 fixed views arranged in a square. We set the viewpoint in the middle of the square to be the target view. As shown in Fig 5, we uniformly sample a position, creating a funnel." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18045" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 88, + 149, + 236 + ], + "blocks": [ + { + "bbox": [ + 70, + 74, + 127, + 84 + ], + "lines": [ + { + "bbox": [ + 70, + 74, + 127, + 84 + ], + "spans": [ + { + "bbox": [ + 70, + 74, + 127, + 84 + ], + "type": "text", + "content": "Camera Layout" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 88, + 149, + 236 + ], + "lines": [ + { + "bbox": [ + 49, + 88, + 149, + 236 + ], + "spans": [ + { + "bbox": [ + 49, + 88, + 149, + 236 + ], + "type": "image", + "image_path": "6479bae5c85977d9e87231cc8aa653b8995ced9101bdf0ea466cbe0fd5dcc0cd.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 238, + 287, + 282 + ], + "lines": [ + { + "bbox": [ + 47, + 238, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 47, + 238, + 287, + 282 + ], + "type": "text", + "content": "Figure 5. Convergence basin analysis: Left: 3D Gaussian map from training views (Yellow) and visualisation of the test poses (Red) and target pose (Blue). Right: Convergence basin of our method. The green marks success, and the red marks failure." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 154, + 88, + 218, + 152 + ], + "blocks": [ + { + "bbox": [ + 158, + 74, + 211, + 84 + ], + "lines": [ + { + "bbox": [ + 158, + 74, + 211, + 84 + ], + "spans": [ + { + "bbox": [ + 158, + 74, + 211, + 84 + ], + "type": "text", + "content": "Ours w/ depth" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 154, + 88, + 218, + 152 + ], + "lines": [ + { + "bbox": [ + 154, + 88, + 218, + 152 + ], + "spans": [ + { + "bbox": [ + 154, + 88, + 218, + 152 + ], + "type": "image", + "image_path": "3c6933a24620117c00152aa1c63b4984b51f73bfee7a703aeee21bf173b14185.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 221, + 88, + 286, + 152 + ], + "blocks": [ + { + "bbox": [ + 223, + 74, + 280, + 84 + ], + "lines": [ + { + "bbox": [ + 223, + 74, + 280, + 84 + ], + "spans": [ + { + "bbox": [ + 223, + 74, + 280, + 84 + ], + "type": "text", + "content": "Ours w/o depth" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 221, + 88, + 286, + 152 + ], + "lines": [ + { + "bbox": [ + 221, + 88, + 286, + 152 + ], + "spans": [ + { + "bbox": [ + 221, + 88, + 286, + 152 + ], + "type": "image", + "image_path": "398ea8a505dfd31b526338e5c8b68a567f418bb02301e9565b80a773c5ca9f35.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 154, + 170, + 218, + 234 + ], + "blocks": [ + { + "bbox": [ + 158, + 157, + 214, + 166 + ], + "lines": [ + { + "bbox": [ + 158, + 157, + 214, + 166 + ], + "spans": [ + { + "bbox": [ + 158, + 157, + 214, + 166 + ], + "type": "text", + "content": "Hash Grid SDF" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 154, + 170, + 218, + 234 + ], + "lines": [ + { + "bbox": [ + 154, + 170, + 218, + 234 + ], + "spans": [ + { + "bbox": [ + 154, + 170, + 218, + 234 + ], + "type": "image", + "image_path": "f3a2e348ba530ca0faa18294db78c0359b0718b93ba9cb2d9224e85ba7f00766.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 221, + 170, + 286, + 233 + ], + "blocks": [ + { + "bbox": [ + 234, + 157, + 270, + 166 + ], + "lines": [ + { + "bbox": [ + 234, + 157, + 270, + 166 + ], + "spans": [ + { + "bbox": [ + 234, + 157, + 270, + 166 + ], + "type": "text", + "content": "MLP SDF" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 221, + 170, + 286, + 233 + ], + "lines": [ + { + "bbox": [ + 221, + 170, + 286, + 233 + ], + "spans": [ + { + "bbox": [ + 221, + 170, + 286, + 233 + ], + "type": "image", + "image_path": "499fec0da5cbf54cc205e566a99a11ddeb17e1db06d1464acbdb4ab2844c6289.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 51, + 289, + 218, + 446 + ], + "blocks": [ + { + "bbox": [ + 51, + 289, + 218, + 446 + ], + "lines": [ + { + "bbox": [ + 51, + 289, + 218, + 446 + ], + "spans": [ + { + "bbox": [ + 51, + 289, + 218, + 446 + ], + "type": "image", + "image_path": "723f9e4c7b39e327f5ec2b8349c052b9a0887637fc93f1ec910a8a91e9c45ae4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 449, + 287, + 483 + ], + "lines": [ + { + "bbox": [ + 47, + 449, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 47, + 449, + 287, + 483 + ], + "type": "text", + "content": "Figure 6. Monocular SLAM result on fr1/desk sequence: We show the reconstructed 3D Gaussian maps (Left) and novel view synthesis result (Right)." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 219, + 289, + 284, + 341 + ], + "blocks": [ + { + "bbox": [ + 219, + 289, + 284, + 341 + ], + "lines": [ + { + "bbox": [ + 219, + 289, + 284, + 341 + ], + "spans": [ + { + "bbox": [ + 219, + 289, + 284, + 341 + ], + "type": "image", + "image_path": "6b963a76d8f3a9f1202632b6815ae6c22899a26fa92e4b325fe42a2dedf021aa.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 219, + 342, + 284, + 396 + ], + "blocks": [ + { + "bbox": [ + 219, + 342, + 284, + 396 + ], + "lines": [ + { + "bbox": [ + 219, + 342, + 284, + 396 + ], + "spans": [ + { + "bbox": [ + 219, + 342, + 284, + 396 + ], + "type": "image", + "image_path": "9e01286b1746790e8df6077f6deeab8ece2b6bc4745c733801623e2e1bb2fde2.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 219, + 396, + 284, + 446 + ], + "blocks": [ + { + "bbox": [ + 219, + 396, + 284, + 446 + ], + "lines": [ + { + "bbox": [ + 219, + 396, + 284, + 446 + ], + "spans": [ + { + "bbox": [ + 219, + 396, + 284, + 446 + ], + "type": "image", + "image_path": "bd4c1275abc36e95390b0c0249c6720ceaa06f0c8a8feb95b1c5ff54438a3fca.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 490, + 287, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 490, + 287, + 705 + ], + "spans": [ + { + "bbox": [ + 47, + 490, + 287, + 705 + ], + "type": "text", + "content": "From the sampled position, given the RGB image of the target view, we perform camera pose optimisation for 1000 iterations. The optimisation is successful if it converges to within " + }, + { + "bbox": [ + 47, + 490, + 287, + 705 + ], + "type": "inline_equation", + "content": "1\\mathrm{cm}" + }, + { + "bbox": [ + 47, + 490, + 287, + 705 + ], + "type": "text", + "content": " of the target view within the fixed iterations. We compare our Gaussian approach with Co-SLAM [39]'s network (Hash Grid SDF) and iMAP's [33] network with Co-SLAM's SDF loss for further geometric accuracy (MLP Neural SDF). We render the training views using a synthetic Replica dataset and create three sequences for testing (seq1, seq2 and seq3). The width of the square formed by the training view is " + }, + { + "bbox": [ + 47, + 490, + 287, + 705 + ], + "type": "inline_equation", + "content": "0.5\\mathrm{m}" + }, + { + "bbox": [ + 47, + 490, + 287, + 705 + ], + "type": "text", + "content": ", and the test cameras are distributed with radii ranging from " + }, + { + "bbox": [ + 47, + 490, + 287, + 705 + ], + "type": "inline_equation", + "content": "0.2\\mathrm{m}" + }, + { + "bbox": [ + 47, + 490, + 287, + 705 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 47, + 490, + 287, + 705 + ], + "type": "inline_equation", + "content": "1.2\\mathrm{m}" + }, + { + "bbox": [ + 47, + 490, + 287, + 705 + ], + "type": "text", + "content": ", covering a larger area than the training view. When training the map, the three methods—Ours w/depth, Hash Grid SDF, and MLP SDF—use RGB-D images, whereas Ours w/o depth utilises only colour images. Fig. 5 shows the qualitative results and Table 6 reports the success rate. For both with and without depth for training, our method shows better convergence." + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 309, + 71, + 425, + 157 + ], + "blocks": [ + { + "bbox": [ + 309, + 71, + 425, + 157 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 425, + 157 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 425, + 157 + ], + "type": "image", + "image_path": "47e9dec8d1434bc8a0c86b8dfa8b7be232b9d911067a469086644d77097f97c5.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 427, + 71, + 544, + 157 + ], + "blocks": [ + { + "bbox": [ + 427, + 71, + 544, + 157 + ], + "lines": [ + { + "bbox": [ + 427, + 71, + 544, + 157 + ], + "spans": [ + { + "bbox": [ + 427, + 71, + 544, + 157 + ], + "type": "image", + "image_path": "21a2ca9113fec668cf62842a26c986443f45def672bb057fc41d56349540e139.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 309, + 158, + 425, + 243 + ], + "blocks": [ + { + "bbox": [ + 309, + 158, + 425, + 243 + ], + "lines": [ + { + "bbox": [ + 309, + 158, + 425, + 243 + ], + "spans": [ + { + "bbox": [ + 309, + 158, + 425, + 243 + ], + "type": "image", + "image_path": "b0cccaaa883c0586fd89440b919de4caf748659e1d7776d32578ecfe0d28b27f.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 246, + 545, + 279 + ], + "lines": [ + { + "bbox": [ + 306, + 246, + 545, + 279 + ], + "spans": [ + { + "bbox": [ + 306, + 246, + 545, + 279 + ], + "type": "text", + "content": "Figure 7. Self-captured Scenes: Challenging scenes and objects, for example, transparent glasses and crinkled texture of salad are captured by our monocular SLAM running live." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 427, + 158, + 544, + 243 + ], + "blocks": [ + { + "bbox": [ + 427, + 158, + 544, + 243 + ], + "lines": [ + { + "bbox": [ + 427, + 158, + 544, + 243 + ], + "spans": [ + { + "bbox": [ + 427, + 158, + 544, + 243 + ], + "type": "image", + "image_path": "6e64f8174f79873aef6837b1b9eed17c58347638bc2c7bbafcb157b8dad47274.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 286, + 545, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 286, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 305, + 286, + 545, + 334 + ], + "type": "text", + "content": "Unlike hashing and positional encoding which can lead to signal conflict, anisotropic Gaussians form a smooth gradient in 3D space, increasing the convergence basin. Further experimental details are available in supplementary ??" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 341, + 417, + 353 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 341, + 417, + 353 + ], + "spans": [ + { + "bbox": [ + 306, + 341, + 417, + 353 + ], + "type": "text", + "content": "4.3. Qualitative Results" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 359, + 545, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 359, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 305, + 359, + 545, + 479 + ], + "type": "text", + "content": "We report both the 3D reconstruction of the SLAM dataset and self-captured sequences. In Fig. 6, we visualise the monocular SLAM reconstruction of fr1/desk. The placements of the Gaussians are geometrically sensible and are 3D coherent, and our rendering from the different viewpoints highlights the quality of our systems' novel view synthesis. In Fig. 7, we self-capture challenging scenes for monocular SLAM. By not explicitly modelling a surface, our system naturally handles transparent objects which is challenging for many other SLAM systems." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 489, + 378, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 489, + 378, + 501 + ], + "spans": [ + { + "bbox": [ + 306, + 489, + 378, + 501 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 509, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 545, + 628 + ], + "type": "text", + "content": "We have proposed the first SLAM method using 3D Gaussians as a SLAM representation. Via efficient volume rendering, our system significantly advances the fidelity and diversity of object materials a live SLAM system can capture. Our system achieves state-of-the-art performance across benchmarks for both monocular and RGB-D cases. Interesting directions for future research are the integration of loop closure for handling large-scale scenes and extraction of geometry such as surface normal as Gaussians do not explicitly represent the surface." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 634, + 408, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 634, + 408, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 634, + 408, + 647 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "content": "Research presented in this paper has been supported by Dyson Technology Ltd. We are very grateful to Eric Dexheimer, Kirill Mazur, Xin Kong, Marwan Taher, Ignacio Alzugaray, Gwangbin Bae, Aalok Patwardhan, and members of the Dyson Robotics Lab for insightful discussions." + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18046" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] J. Czarnowski, T. Laidlow, R. Clark, and A. J. Davison. Deepfactors: Real-time probabilistic dense monocular SLAM. IEEE Robotics and Automation Letters (RAL), 5(2): 721-728, 2020." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 288, + 192 + ], + "type": "text", + "content": "[2] Angela Dai, Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Christian Theobalt. BundleFusion: Real-time Globally Consistent 3D Reconstruction using On-the-fly Surface Re-integration. ACM Transactions on Graphics (TOG), 36(3):24:1-24:18, 2017." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 287, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 287, + 236 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 287, + 236 + ], + "type": "text", + "content": "[3] Eric Dexheimer and Andrew J. Davison. Learning a Depth Covariance Function. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 238, + 287, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 238, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 53, + 238, + 287, + 270 + ], + "type": "text", + "content": "[4] J. Engel, V. Koltun, and D. Cremers. Direct sparse odometry. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 2017." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 272, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 272, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 272, + 287, + 316 + ], + "type": "text", + "content": "[5] C. Forster, M. Pizzoli, and D. Scaramuzza. SVO: Fast Semi-Direct Monocular Visual Odometry. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2014." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 318, + 287, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 318, + 287, + 372 + ], + "spans": [ + { + "bbox": [ + 53, + 318, + 287, + 372 + ], + "type": "text", + "content": "[6] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenumxels: Radiance fields without neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 373, + 287, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 373, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 53, + 373, + 287, + 418 + ], + "type": "text", + "content": "[7] Jiahui Huang, Shi-Sheng Huang, Haoxuan Song, and ShiMin Hu. Di-fusion: Online implicit 3d reconstruction with deep priors. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 419, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 419, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 53, + 419, + 287, + 464 + ], + "type": "text", + "content": "[8] M. M. Johari, C. Carta, and F. Fleuret. ESLAM: Efficient dense slam system based on hybrid representation of signed distance fields. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 465, + 287, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 465, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 53, + 465, + 287, + 508 + ], + "type": "text", + "content": "[9] M. Keller, D. Lefloch, M. Lambers, S. Izadi, T. Weyrich, and A. Kolb. Real-time 3D Reconstruction in Dynamic Scenes using Point-based Fusion. In Proc. of Joint 3DIM/3DPVT Conference (3DV), 2013." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 510, + 287, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 287, + 553 + ], + "type": "text", + "content": "[10] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (TOG), 2023." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 555, + 287, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 597 + ], + "type": "text", + "content": "[11] Leonid Keselman and Martial Hebert. Approximate differentiable rendering with algebraic surfaces. In Proceedings of the European Conference on Computer Vision (ECCV), 2022." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 600, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 600, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 600, + 287, + 633 + ], + "type": "text", + "content": "[12] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proceedings of the International Conference on Learning Representations (ICLR), 2015." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "text", + "content": "[13] Heng Li, Xiaodong Gu, Weihao Yuan, Luwei Yang, Zilong Dong, and Ping Tan. Dense rgb slam with neural implicit maps. In Proceedings of the International Conference on Learning Representations (ICLR), 2023." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "text", + "content": "[14] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. NeurIPS, 2020." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[15] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. 3DV, 2024." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "type": "text", + "content": "[16] J. McCormac, A. Handa, A. J. Davison, and S. Leutenegger. SemanticFusion: Dense 3D semantic mapping with convolutional neural networks. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2017." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "type": "text", + "content": "[17] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Proceedings of the European Conference on Computer Vision (ECCV), 2020." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 220, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 262 + ], + "type": "text", + "content": "[18] N. J. Mitra, N. Gelfand, H. Pottmann, and L. J. Guibas. Registration of Point Cloud Data from a Geometric Optimization Perspective. In Proceedings of the Symposium on Geometry Processing, 2004." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "type": "text", + "content": "[19] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (TOG), 2022." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 309, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 352 + ], + "type": "text", + "content": "[20] R. Mur-Artal and J. D. Tardós. ORB-SLAM2: An Open-Source SLAM System for Monocular, Stereo, and RGB-D Cameras. IEEE Transactions on Robotics (T-RO), 33(5): 1255–1262, 2017." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 354, + 545, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 397 + ], + "type": "text", + "content": "[21] R. Mur-Artal, J. M. M Montiel, and J. D. Tardós. ORBSLAM: a Versatile and Accurate Monocular SLAM System. IEEE Transactions on Robotics (T-RO), 31(5):1147-1163, 2015." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 399, + 545, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 421 + ], + "type": "text", + "content": "[22] R. A. Newcombe. Dense Visual SLAM. PhD thesis, Imperial College London, 2012." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 422, + 545, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 545, + 477 + ], + "type": "text", + "content": "[23] R. A. Newcombe, S. Izadi, O. Hilliges, D. Molyneaux, D. Kim, A. J. Davison, P. Kohli, J. Shotton, S. Hodges, and A. Fitzgibbon. KinectFusion: Real-Time Dense Surface Mapping and Tracking. In Proceedings of the International Symposium on Mixed and Augmented Reality (ISMAR), 2011." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 479, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 479, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 479, + 545, + 533 + ], + "type": "text", + "content": "[24] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 535, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 535, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 535, + 545, + 567 + ], + "type": "text", + "content": "[25] M. Nießner, M. Zollhöfer, S. Izadi, and M. Stamminger. Real-time 3D Reconstruction at Scale using Voxel Hashing. In Proceedings of SIGGRAPH, 2013." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 568, + 545, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 568, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 307, + 568, + 545, + 622 + ], + "type": "text", + "content": "[26] Victor Adrian Prisacariu, Olaf Kähler, Ming-Ming Cheng, Carl Yuheng Ren, Julien P. C. Valentin, Philip H. S. Torr, Ian D. Reid, and David W. Murray. A framework for the volumetric integration of depth images. CoRR, abs/1410.0925, 2014." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 624, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 667 + ], + "type": "text", + "content": "[27] Erik Sandström, Yue Li, Luc Van Gool, and Martin R. Oswald. Point-slam: Dense neural point cloud-based slam. In Proceedings of the International Conference on Computer Vision (ICCV), 2023." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[28] Thomas Schöps, Torsten Sattler, and Marc Pollefeys. Surfelmeshing: Online surfel-based mesh reconstruction. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 2020." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18047" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[29] Thomas Schöps, Torsten Sattler, and Marc Pollefeys. Bad slam: Bundle adjusted direct rgb-d slam. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 139 + ], + "type": "text", + "content": "[30] J. Solà, J. Deray, and D. Atchuthan. A micro Lie theory for state estimation in robotics. arXiv:1812.01537, 2018." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 286, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 286, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 286, + 251 + ], + "type": "text", + "content": "[31] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, Anton Clarkson, Mingfei Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Braies, Tyler Gillingham, Elias Mueggler, Luis Pesqueira, Manolis Savva, Dhruv Batra, Hauke M. Strasdat, Renzo De Nardi, Michael Goesele, Steven Lovegrove, and Richard Newcombe. The Replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 252, + 286, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 252, + 286, + 296 + ], + "spans": [ + { + "bbox": [ + 48, + 252, + 286, + 296 + ], + "type": "text", + "content": "[32] J. Sturm, N. Engelhard, F. Endres, W. Burgard, and D. Cremers. A Benchmark for the Evaluation of RGB-D SLAM Systems. In Proceedings of the IEEE/RSJ Conference on Intelligent Robots and Systems (IROS), 2012." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 297, + 286, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 297, + 286, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 286, + 330 + ], + "type": "text", + "content": "[33] E. Sucar, S. Liu, J. Ortiz, and A. J. Davison. iMAP: Implicit mapping and positioning in real-time. In Proceedings of the International Conference on Computer Vision (ICCV), 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 331, + 286, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 286, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 286, + 374 + ], + "type": "text", + "content": "[34] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 376, + 286, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 286, + 420 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 286, + 420 + ], + "type": "text", + "content": "[35] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. Proceedings of the International Conference on Learning Representations (ICLR), 2024." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 421, + 286, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 421, + 286, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 421, + 286, + 453 + ], + "type": "text", + "content": "[36] Zachary Teed and Jia Deng. DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras. In Neural Information Processing Systems (NIPS), 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 455, + 286, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 455, + 286, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 455, + 286, + 510 + ], + "type": "text", + "content": "[37] Emanuele Vespa, Nikolay Nikolov, Marius Grimm, Luigi Nardi, Paul HJ Kelly, and Stefan Leutenegger. Efficient octree-based volumetric SLAM supporting signed-distance and occupancy mapping. IEEE Robotics and Automation Letters (RAL), 2018." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 511, + 286, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 511, + 286, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 511, + 286, + 544 + ], + "type": "text", + "content": "[38] Angtian Wang, Peng Wang, Jian Sun, Adam Kortylewski, and Alan Yuille. Voge: a differentiable volume renderer using gaussian ellipsoids for analysis-by-synthesis. 2022." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 545, + 286, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 286, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 286, + 588 + ], + "type": "text", + "content": "[39] Hengyi Wang, Jingwen Wang, and Lourdes Agapito. Coslam: Joint coordinate and sparse parametric encodings for neural real-time slam. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 590, + 286, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 286, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 286, + 633 + ], + "type": "text", + "content": "[40] T. Whelan, M. Kaess, H. Johannsson, M. F. Fallon, J. J. Leonard, and J. B. McDonald. Real-time large scale dense RGB-D SLAM with volumetric fusion. International Journal of Robotics Research (IJRR), 34(4-5):598-626, 2015." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 286, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 286, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 286, + 678 + ], + "type": "text", + "content": "[41] T. Whelan, S. Leutenegger, R. F. Salas-Moreno, B. Glocker, and A. J. Davison. ElasticFusion: Dense SLAM without a pose graph. In Proceedings of Robotics: Science and Systems (RSS), 2015." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 286, + 713 + ], + "type": "text", + "content": "[42] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 397 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 96, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 150 + ], + "type": "text", + "content": "[43] Xingrui Yang, Hai Li, Hongjia Zhai, Yuhang Ming, Yuqian Liu, and Guofeng Zhang. Vox-fusion: Dense tracking and mapping with voxel-based neural implicit representation. In Proceedings of the International Symposium on Mixed and Augmented Reality (ISMAR), 2022." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "type": "text", + "content": "[44] Zeyu Yang, Hongye Yang, Zijie Pan, Xiatian Zhu, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. Proceedings of the International Conference on Learning Representations (ICLR), 2024." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 207, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 207, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 207, + 545, + 262 + ], + "type": "text", + "content": "[45] Taoran Yi, Jiemin Fang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussian-dreamer: Fast generation from text to 3d gaussian splatting with point cloud priors. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 263, + 545, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 263, + 545, + 318 + ], + "spans": [ + { + "bbox": [ + 308, + 263, + 545, + 318 + ], + "type": "text", + "content": "[46] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R. Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 319, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 319, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 308, + 319, + 545, + 363 + ], + "type": "text", + "content": "[47] Zihan Zhu, Songyou Peng, Viktor Larsson, Zhaopeng Cui, Martin R Oswald, Andreas Geiger, and Marc Pollefeys. Nicer-slam: Neural implicit scene encoding for rgb slam. International Conference on 3D Vision (3DV), 2024." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 364, + 545, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 364, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 308, + 364, + 545, + 397 + ], + "type": "text", + "content": "[48] M. Zwicker, H. Pfister, J. van Baar, and M. Gross. Ewa splatting. IEEE Transactions on Visualization and Computer Graphics, 8(3):223-238, 2002." + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18048" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Gaussian-Flow_ 4D Reconstruction with Dynamic 3D Gaussian Particle/3e2469b7-4042-4a31-bb56-8ba156724067_content_list.json b/2024/Gaussian-Flow_ 4D Reconstruction with Dynamic 3D Gaussian Particle/3e2469b7-4042-4a31-bb56-8ba156724067_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e507a4a0fec9585e8cb9f585157927de235c1d5d --- /dev/null +++ b/2024/Gaussian-Flow_ 4D Reconstruction with Dynamic 3D Gaussian Particle/3e2469b7-4042-4a31-bb56-8ba156724067_content_list.json @@ -0,0 +1,1639 @@ +[ + { + "type": "text", + "text": "Gaussian-Flow: 4D Reconstruction with Dynamic 3D Gaussian Particle", + "text_level": 1, + "bbox": [ + 124, + 130, + 844, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Youtian Lin1,2", + "bbox": [ + 217, + 179, + 334, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zuozhuo Dai3", + "bbox": [ + 372, + 180, + 486, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Siyu Zhu", + "bbox": [ + 526, + 180, + 609, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yao $\\mathrm{Yao^{1,2\\boxtimes}}$", + "bbox": [ + 651, + 180, + 751, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ State Key Laboratory for Novel Software Technology, Nanjing University, Nanjing, China \n $^{2}$ School of Intelligence Science and Technology, Nanjing University, Suzhou, China \n $^{3}$ Alibaba Group \n $^{4}$ Fudan University", + "bbox": [ + 125, + 198, + 843, + 268 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/011177e7de7e581e87375234ad14b83ca5958e7b27a20ec981f875dcdaaf5e85.jpg", + "image_caption": [ + "Figure 1. Dynamic reconstruction results of the proposed Gaussian-Flow on the monocular D-NeRF [30] and HyperNeRF Dataset [28] (left) and the multi-view Plenoptic Dataset [18] (right). Our method achieves a $5 \\times$ faster training and rendering speed compared with the per-frame 3DGS modeling and significantly outperforms previous methods in novel view rendering quality." + ], + "image_footnote": [], + "bbox": [ + 81, + 305, + 349, + 542 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8409baae1f126fa7fc775f1a146696fd00a48429075bdd1afce2509f1eda6930.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 306, + 581, + 541 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c203f09eb7b8ed5c101a2eedb8f7c227f8bd12d421e8ad62a461fdb66ceb115f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 586, + 305, + 735, + 544 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cee1196eb8f70fdda808aa8d5768a28dcbd2522dbda0fcde4d193b64ac301275.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 735, + 306, + 888, + 544 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 609, + 313, + 626 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce Gaussian-Flow, a novel point-based approach for fast dynamic scene reconstruction and real-time rendering from both multi-view and monocular videos. In contrast to the prevalent NeRF-based approaches hampered by slow training and rendering speeds, our approach harnesses recent advancements in point-based 3D Gaussian Splatting (3DGS). Specifically, a novel Dual-Domain Deformation Model (DDDM) is proposed to explicitly model attribute deformations of each Gaussian point, where the time-dependent residual of each attribute is captured by a polynomial fitting in the time domain, and a Fourier series fitting in the frequency domain. The proposed DDDM is capable of modeling complex scene deformations across long video footage, eliminating the need for training separate 3DGS for each frame or introducing an additional implicit neural field to model 3D dynamics. Moreover, the explicit deformation modeling for discretized Gaussian points en", + "bbox": [ + 73, + 643, + 472, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sures ultra-fast training and rendering of a 4D scene, which is comparable to the original 3DGS designed for static 3D reconstruction. Our proposed approach showcases a substantial efficiency improvement, achieving a $5 \\times$ faster training speed compared to the per-frame 3DGS modeling. In addition, quantitative results demonstrate that the proposed Gaussian-Flow significantly outperforms previous leading methods in novel view rendering quality. Project page: https://nju-3dv.github.io/projects/Gaussian-Flow.", + "bbox": [ + 496, + 612, + 893, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 797, + 632, + 813 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the realm of digital scene synthesis, achieving a balance between high-quality reconstructions and real-time rendering is paramount, especially for applications like virtual reality (VR) playback, where immediate feedback and immersive experiences are essential. Neural Radiance Fields", + "bbox": [ + 496, + 824, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21136", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(NeRFs) [25] has risen as a promising method for synthesizing intricate scenes. However, despite their ability to produce visually stunning results, NeRFs require costly sampling and evaluation of the neural radiance field at multiple points along each ray. Consequently, the substantial computational demands impede the real-time rendering capabilities of NeRFs. These are attempts to accelerate the rendering process of NeRFs, such as direct volume representation [21, 33, 40], neural hasing [26], and tri-plane structures [2, 4, 9]. However, it still remains a challenge for high-fidelity real-time rendering. Nevertheless, the issue becomes even more severe when turning to dynamic scene reconstruction and rendering.", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent progress on 3D Gaussian Splatting (3DGS) [14] has drawn attention from the 3D computer vision community. With tile-based rasterization instead of plain volume rendering, 3DGS can render images two orders of magnitude faster than the vanilla NeRF. The technique has also been quickly applied to 4D scene reconstruction by extending to the separate per-frame 3DGS optimization [23]. However, such direct extension is storage-intensive and is not applicable to monocular video input. Some other concurrent works [35, 38] try to mix the explicit point-based 3DGS and an implicit neural field for dynamic information modeling, however, require computationally expensive forward passes of the neural network, which significantly lowers the rendering speed of the original 3DGS.", + "bbox": [ + 75, + 291, + 472, + 503 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose Gaussian-Flow, an explicit particle-based deformation model designed specifically for 3DGS to model the dynamic scene without using any neural network. Gaussian-Flow can recover a high-fidelity 4D scene from captured videos, while still preserving the ultrafast training and rendering speed of the original 3DGS. In particular, we formulate a 4D scene as a set of deformable 3D Gaussian points. A novel Dual-Domain Deformation Model (DDDM) is proposed to explicitly model deformations of each Gaussian point's attributes, including position, rotation, and radiance. The time-dependent deformation residual is modeled simultaneously in time and frequency domains: we apply joint polynomial and Fourier series fitting for each deformable attribute. This compact dynamic representation greatly reduces the computation cost of the deformation model, which is a key factor in preserving the rendering speed of 3DGS. Moreover, an adaptive times-tamp scaling technique is introduced to avoid over-fitting the scene to only frames with violent motions. For robust estimation, we also regularize the motion trajectory by the KNN-based rigid and the time smooth constraints. It is also noteworthy that our discretized point-based 4D representation naturally supports the edition of both static and dynamic 3D scenes, showing the potential for unlocking a variety of downstream applications related to dynamic 3D reconstruction and rendering.", + "bbox": [ + 75, + 508, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We have conducted extensive experiments to demonstrate the effectiveness of the proposed method on several multi-view and monocular datasets. The proposed Gaussian-Flow achieves a $5 \\times$ faster training speed compared with the separate per-frame 3DGS modeling, and significantly outperforms prior leading methods in novel view rendering quality. Our major contributions can be summarized as follows:", + "bbox": [ + 496, + 90, + 890, + 210 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce Gaussian-Flow, a novel point-based differentiable rendering approach for dynamic 3D scene reconstruction, setting a new state-of-the-art for training speed, rendering FPS, and novel view synthesis quality for 4D scene reconstruction.", + "- We propose a Dual-Domain Deformation Model for efficient 4D scene training and rendering, eliminating the need for per-frame 3DGS optimization and sampling on implicit neural fields. This preserves a running speed on par with the original 3DGS with minimum overhead.", + "- We demonstrate that our discretized point-based representation supports the segmentation, edition, and composition of both static and dynamic 3D scenes." + ], + "bbox": [ + 500, + 212, + 890, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 422, + 650, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Dynamic Neural Radiance Field", + "text_level": 1, + "bbox": [ + 500, + 449, + 782, + 465 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dynamic NeRF modeling has become a heated research topic in recent years due to the development of the neural radiance field and differentiable rendering. By treating time as an extended input dimension to NeRF, researchers successfully achieve qualified image-based 4D scene rendering [5, 10, 19, 22, 37]. To further improve the reconstruction quality and incorporate prior knowledge of motions and structures, dynamic neural scene flow methods have been proposed [27, 30], where a canonical space is constructed and then transferred to each frame with scene flow or motion fields. HyperNeRF [28] models the deformation of the object topologies by using higher-dimensional inputs, while DyNeRF [18] utilizes time-conditioned NeRF to represent a 4D scene. However, the aforementioned approaches are all based on the vanilla NeRF, which requires a long training time and does not meet the requirement of real-time rendering.", + "bbox": [ + 496, + 472, + 890, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Accelerated Neural Radiance Field", + "text_level": 1, + "bbox": [ + 500, + 739, + 803, + 755 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To expedite NeRF training and rendering, numerous approaches have been suggested, employing more streamlined strategies [6, 11, 12, 16, 20, 29, 36]. Other methods propose the integration of neural implicit functions with explicit 3D structures, forming a hybrid representation for faster radiance field sampling [4, 26, 33, 34]. These approaches establish strong foundations for enhancing dynamic NeRF, concurrently decreasing required training and inference time. Apart from implicit neural representations, explicit NeRF", + "bbox": [ + 496, + 763, + 890, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21137", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "modeling has show promising results for real-time rendering: NSVF [21] employs a neural sparse voxel field for efficient NeRF sampling, which stands for the earliest attempt on the explicit NeRF modeling; PlenOctrees [40] utilizes the explicit octree structure for rendering acceleration.", + "bbox": [ + 75, + 90, + 467, + 166 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent efforts have also emerged to accelerate the intricate dynamic neural radiance field. TensorRF [4] employs multiple planes as explicit representations for direct dynamic scene modeling. More recent recent approaches of this kind include K-Planes [8], Tensor4D [31], and HexPlane [3]. Alternatively, NeRFPlayer[32] introduces a unified streaming representation for both grid-based [26] and plane-based methods, utilizing separate models to distinguish static and dynamic scene components, however, leading to slow rendering times. HyperReel [1] further suggests a flexible sampling network coupled with two planes for dynamic scene representation. While these methods improve the rendering speed of a dynamic scene to some extent, it is still hard to achieve real-time rendering, let alone a good balance between the running speed and rendering quality. In contrast, we resort to the recent 3D Gaussians splatting, which applies an explicit soft point cloud representation for real-time image-based rendering.", + "bbox": [ + 75, + 167, + 468, + 438 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Differentiable Point-based Rendering", + "text_level": 1, + "bbox": [ + 76, + 446, + 401, + 462 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The original idea of using 3D points as rendering primitives was first introduced in [17]. By incorporating differentiable rendering, recent approaches have made remarkable progress in image-based rendering, representative methods include PointRF [41], DSS [39], and 3D Gaussians splatting (3DGS) [13]. Specifically, 3DGS [13] has demonstrated extraordinary performance in novel-view synthesis, achieving real-time rendering speed and state-of-the-art rendering quality. The method adopts a soft point representation with attributes of position, rotation, density, and radiance, and applies differentiable point-based rendering for scene optimization. 3DGS has quickly been extended to dynamic scene modeling by direct separate per-frame optimization [23], however, requires a long optimization time and a large amount of storage for long video footage. Other works [35, 38] apply an implicit motion field to model scene dynamics, but the introduction of the implicit neural network significantly slows down the sampling and rendering speed. In this work, we adopt the approach of representing 4D scenes through a purely discretized point cloud model, ensuring a fast training and rendering speed comparable with the original 3DGS.", + "bbox": [ + 75, + 469, + 468, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Gaussian-Flow", + "text_level": 1, + "bbox": [ + 76, + 815, + 225, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce the proposed Gaussian-Flow for dynamic scene modeling. We first review the 3DGS in Sec. 3.1. Then, we introduce our explicit motion modeling of each Gaussian point by using a novel Dual-Domain", + "bbox": [ + 75, + 839, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Deformation Model (DDDM), as outlined in Sec. 3.2. An adaptive timestamp scaling technique is described in Sec. 3.3 for balanced training of each frame. To ensure the continuity of the motion in both spatial and temporal dimensions, we incorporate appropriate regularizations on each point during the optimization, as detailed in Sec. 3.4.", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Recap on 3D Gaussian Splitting", + "text_level": 1, + "bbox": [ + 498, + 191, + 784, + 208 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Gaussian Splatting [14] is designed to efficiently optimize a 3D scene for real-time and high-quality novel view synthesis. The 3DGS framework has garnered significant attention within the community due to its remarkable enhancements in both training and rendering times, concurrently achieving state-of-the-art rendering quality. In contrast to the volume rendering in the vanilla NeRF which relies on ray marching, 3DGS adopts a tile-based rasterization on a distinctive soft point cloud representation to achieve fast rendering. Specifically, 3DGS models a 3D scene as a large amount of 3D Gaussian points in the world space, where each point is represented by:", + "bbox": [ + 496, + 213, + 890, + 396 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nG (\\boldsymbol {x}) = \\exp \\left(- \\frac {1}{2} \\left(\\boldsymbol {x} - \\boldsymbol {\\mu}\\right) ^ {T} \\boldsymbol {\\Sigma} ^ {- 1} \\left(\\boldsymbol {x} - \\boldsymbol {\\mu}\\right)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 405, + 890, + 434 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mu$ and $\\Sigma$ are the mean position and covariance matrix of a 3D Gaussian particle. The 3DGS takes sparse Structure-from-Motion (SfM) points or even random points as input, and initializes each point to a 3D Gaussian based on its neighbors. Besides, each 3D Gaussian is associated with a learnable view-dependent radiance $c$ and a learnable opacity $\\alpha$ for rendering. Subsequently, an efficient 3D to 2D Gaussian mapping [42] is employed to project the point onto the image plane:", + "bbox": [ + 496, + 443, + 890, + 579 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\mu} ^ {\\prime} = \\boldsymbol {P} \\boldsymbol {W} \\boldsymbol {\\mu}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 648, + 590, + 890, + 607 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Sigma} ^ {\\prime} = \\boldsymbol {J} \\boldsymbol {W} \\boldsymbol {\\Sigma} \\boldsymbol {W} ^ {T} \\boldsymbol {J} ^ {T}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 616, + 890, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mu^{\\prime}$ and $\\Sigma^{\\prime}$ represent the 2D mean position and 2D covariance of the projected 3D Gaussian. $P, W$ and $J$ donate the projective transformation, viewing transformation, and Jacobian of the affine approximation of $P$ . After that, $\\alpha$ -blending is executed to merge the overlapping Gaussians for each pixel, yielding a final color:", + "bbox": [ + 496, + 640, + 890, + 731 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nC = \\sum_ {i = 1} ^ {n} \\boldsymbol {c} _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {i}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 742, + 890, + 785 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $c_{i}$ and $\\alpha_{i}$ are the color and opacity of the $i$ -th Gaussian, and $n$ is the number of overlapping Gaussians.", + "bbox": [ + 496, + 795, + 888, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The attributes of the 3D Gaussian, including the mean $\\mu$ , covariance matrix $\\Sigma$ , opacity $\\alpha$ , and color $c$ , are optimized through backward propagation of the gradient flow. In particular, to ensure positive definiteness during the optimization process, the covariance matrix is parameterized", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21138", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "as a scaling vector $s$ and a rotation matrix $R$ , i.e., $\\Sigma = R\\Lambda(s)\\Lambda(s)^T R^T$ , where $\\Lambda(s)$ is the diagonal matrix of $s$ . To facilitate optimization, the rotation matrix is further parameterized using a quaternion $q$ . Leveraging the inherent flexibility of discrete points, 3DGS incorporates adaptive density control for points. This mechanism utilizes the gradient flow to identify where geometric reconstruction is suboptimal, and employs cloning and splitting to augment the density of points for a higher rendering quality.", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Dual-Domain Deformation Model", + "text_level": 1, + "bbox": [ + 76, + 242, + 372, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We target to directly model the dynamics of each 3D Gaussian point by fitting each of its attributes into a time-dependent curve. Among different approaches, Polynomials fitting in the time domain and Fourier series fitting in the frequency domain are the two most widely used approaches [5, 10, 19, 22, 37], due to their simplicity and effectiveness. However, each method comes with its own advantages and drawbacks: describing the motion of a Gaussian particle in terms of polynomials yields a good fit with smooth motion with a small order of polynomials, however, can easily overfit to a violent motion if assuming a larger order of polynomials, resulting in unreasonable oscillations in the fitted trajectory. Whereas, the Fourier series excels at capturing the variations associated with violent motion, however, requires a manually reduced order when dealing with smooth motion.", + "bbox": [ + 75, + 268, + 472, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this work, our key insight is to use a Dual-Domain Deformation Model (DDDM) for fitting the scene dynamics, which integrates both the time domain polynomials and the frequency domain Fourier series into a unified fitting model. We assume that only the rotation $\\pmb{q}$ , radiance $c$ , and position $\\pmb{\\mu}$ of a 3D Gaussian particle change over time, while the scaling $s$ and opacity $\\alpha$ remain constant. Specifically, we conceptualize the change in each particle's attributes as its base attributes $S_0 \\in \\{\\pmb{\\mu}_0, c_0, q_0\\}$ at the reference time frame $t_0$ (usually set to the first frame), superimposed on a time-dependent attribute residual $D(t)$ . For simplicity, we use lowercase characters to represent a single attribute in $S$ . The time-dependent residual of each attribute is modeled through polynomial fitting in the time domain and Fourier series fitting in the frequency domain, expressed as:", + "bbox": [ + 75, + 511, + 472, + 739 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS (t) = S _ {0} + D (t), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 753, + 468, + 771 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $D(t) = P_N(t) + F_L(t)$ is combined by a polynomial $P_N(t)$ with coefficients $\\pmb{a} = \\{a\\}_{n=0}^{N}$ and a Fourier series $F_L(t)$ with coefficients $\\pmb{f} = \\{f_{sin}^l, f_{cos}^l\\}_{l=0}^L$ . These are respectively denoted as:", + "bbox": [ + 76, + 786, + 470, + 848 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP _ {N} (t) = \\sum_ {n = 0} ^ {N} a _ {n} t ^ {n}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 205, + 862, + 468, + 904 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {L} (t) = \\sum_ {l = 1} ^ {L} \\left(f _ {\\sin} ^ {l} \\cos (l t) + f _ {\\cos} ^ {l} \\sin (l t)\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 89, + 890, + 130 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It is important to note that we assume different dimensions of an attribute are independently changed over time. Therefore, we assign a different $D(t)$ for each dimension of an attribute. For instance, we utilize $\\{D_{\\mu_i}(t)\\}_{i=0}^3$ to describe the motion of a 3D position $\\pmb{\\mu}$ .", + "bbox": [ + 496, + 137, + 890, + 212 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 3 illustrates a comparative analysis of trajectory fitting using polynomial, Fourier series, and the proposed joint DDDM functions. The figure highlights the superior fitting capabilities of the DDDM approach in capturing complex motion trajectories as represented by the sampled data points.", + "bbox": [ + 496, + 213, + 890, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Adaptive Timestemp Scaling", + "text_level": 1, + "bbox": [ + 498, + 315, + 756, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In a typical scenario, a normalized frame index $t$ ranging from 0 to 1 will be used as the temporal input of $D(t)$ . However, this poses a challenge when endeavoring to model substantial motions within a very short time using polynomials and Fourier series. Adhering to the standard temporal division would necessitate an exceedingly large coefficient to accommodate highly intense movements within a very short time frame. This circumstance has the potential to induce instability or even a breakdown in the optimization process. To address this issue, we introduce a time dilation factor $\\lambda$ to scale the temporal input for each Gaussian point, which is formulated as:", + "bbox": [ + 496, + 338, + 890, + 518 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nt _ {s} = \\lambda_ {s} \\cdot t + \\lambda_ {b} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 534, + 890, + 549 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $t_s$ represents the scaled time input, serving as the input of $D(t)$ , $t \\in [0,1]$ denotes the normalized frame index, and $\\lambda$ and $\\lambda_b$ stand for the dilation factor and base factor of a Gaussian, respectively. In all our experiments, $\\lambda_s$ and $\\lambda_b$ are initialized to 1 and 0, respectively.", + "bbox": [ + 496, + 561, + 890, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To summarize, in our dynamic scene setting, a Gaussian particle contains multiple attributes to be optimized, including base attributes $\\{\\mu_0,q_0,s_0,c_0,\\alpha_0\\}$ at reference frame $t_0$ , polynomial coefficients and Fourier coefficients in $\\{D_{\\mu}(t),D_q(t),D_c(t)\\}$ . Since the proposed DDDM optimizes the time-dependent residuals without the need for an intricate neural field structure, our Gaussian-Flow inherits the benefits of extreme-fast training and rendering speed from the vanilla 3DGS.", + "bbox": [ + 496, + 638, + 890, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Regularizations", + "text_level": 1, + "bbox": [ + 500, + 786, + 658, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While the utilization of discrete points as scene representations accelerates rendering, several challenges remain. First, points are optimized individually, losing connections with their spatial neighbors, which do not align with the real-world scenario. Optimizing these Gaussian points without considering continuity will inevitably lead to a", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21139", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0048a63e88b776549374029bf25e4d539c550b3f8d35cd19bab3b5eae4b7e4dd.jpg", + "image_caption": [ + "Adaptive Timestamp Scaling" + ], + "image_footnote": [], + "bbox": [ + 86, + 94, + 112, + 111 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1f113e0850392e0cc881978378908c8e851b6024c72e500a1dfa9cce415dab62.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 292, + 95, + 339, + 109 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/507b87fe0e2e244937d05cd068f4ffc7e90ef1ac0a50cc6db51452aadebf093b.jpg", + "image_caption": [ + "Radiance" + ], + "image_footnote": [], + "bbox": [ + 547, + 95, + 593, + 109 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/af0a463f6e073f6ef74f5ab11dc8d0514d151ac159c84ae9c0d86c84c47b5c64.jpg", + "image_caption": [ + "Forward" + ], + "image_footnote": [], + "bbox": [ + 663, + 95, + 705, + 109 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/413f4d271d9b9c08183d9d2c203e05f07a2cb0011adc7c627d3af178101cb14b.jpg", + "image_caption": [ + "Backward" + ], + "image_footnote": [], + "bbox": [ + 769, + 95, + 815, + 109 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5736ef64b98c15315404595634f02571a3b576feeae6e06e7711771c647a60d7.jpg", + "image_caption": [ + "Dual-Domain Deformation Model" + ], + "image_footnote": [], + "bbox": [ + 76, + 126, + 893, + 297 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/67dee80a23b0eb50805662b4088b8f741b9a27a8df8f1a32c08642bc80092308.jpg", + "image_caption": [ + "Figure 2. Overview of the Gaussian-Flow pipeline. We model the deformation of attributes of each 3D Gaussian point independently by using the Dual-Domain Deformation Model (DDDM), which preserves the discretized nature of the 3D Gaussian points, and thus achieves ultra-fast training and rendering speed comparable with the original 3DGS." + ], + "image_footnote": [], + "bbox": [ + 91, + 393, + 264, + 523 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/943b01e627be6bb69f0a616bfcc9579a1416ad2ecaee302a292f5c25ab2d855b.jpg", + "image_caption": [ + "Figure 3. Two exemplar motion fittings using polynomial, Fourier Series, and the joint DDDM functions. Our DDDM is able to accurately fit complex trajectories denoted by the sampling points." + ], + "image_footnote": [], + "bbox": [ + 282, + 397, + 452, + 522 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "degradation in reconstruction quality and spatial coherence. Additionally, motions should be smoothed over time. Based on these observations, we employ two regularizations, namely time smoothness and a KNN rigid regularization, for robust optimization of Gaussian points and their motions.", + "bbox": [ + 75, + 609, + 468, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Time Smoothness Loss To ensure temporal smoothness over time, we apply a perturbing $\\epsilon$ on the input timestamp $t$ and encourage the time-dependent attributes (i.e., position $\\mu$ , rotation $q$ and radiance $c$ ) at time $t + \\epsilon$ to be consistent with those at time $t$ . The time smoothness term is defined as:", + "bbox": [ + 75, + 729, + 468, + 820 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t} = \\| D (t) - D (t + \\epsilon) \\| _ {2}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 825, + 468, + 843 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It is noteworthy that the magnitude of the perturbing value is adaptively set according to the number of total frames, i.e., $\\epsilon = 0.1/$ frames.", + "bbox": [ + 75, + 854, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "KNN Rigid Loss During the optimization of 3D Gaussians with adaptive density control, points will be dynamically added or removed. This dynamic nature implies that the neighbors of points within a local space are subject to constant changes, posing a challenge for directly enforcing a spatial local consistency constraint. To sidestep this problem, we propose to divide the optimization part into two alternating stages: in the former stage, we optimize all variables with adaptive density control; while in the latter stage, we optimize the attributes without adding or removing points. The local rigid constraint is incorporated in every latter stage, and it is defined as:", + "bbox": [ + 496, + 383, + 893, + 566 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s} = \\sum_ {j \\in \\mathcal {N} _ {i}} \\| D (t) _ {i} - D (t) _ {j} \\| _ {2} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 578, + 890, + 612 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where, $\\mathcal{N}_i$ represents the $K$ nearest neighbor (KNN) of $i$ -th Gaussian.", + "bbox": [ + 496, + 623, + 893, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 669, + 633, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 695, + 718, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We train our model using the Adam [15] optimizer with separate learning rates for different attributes of the Gaussian point. We set a learning rate of $4 \\times 10^{-4}$ for the point position with an exponential decay of $8 \\times 10^{-7}$ . The learning rates for point rotation and all DDDM parameters are set to 0.002 and $4 \\times 10^{-4}$ respectively. We apply a weight decay of $8 \\times 10^{-7}$ to all parameters. The rest of the learning rate follows the 3DGS setting. We train the model for 30K steps and 60K steps for all scenes. All experiments are conducted on a single NVIDIA RTX 4090 GPU with 24GB memory. In addition, we use Taichi to implement our DDDM model, which can parallelize the computation of the DDDM (i.e.", + "bbox": [ + 496, + 719, + 893, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21140", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "polynomial and Fourier series computation) of each Gaussian point.", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Datasets", + "text_level": 1, + "bbox": [ + 76, + 132, + 179, + 146 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our method on both multi-view and monocular datasets, to demonstrate the effectiveness of our method in both settings.", + "bbox": [ + 76, + 156, + 468, + 202 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Plenoptic Video dataset [18] The dataset was captured using 21 cameras at a resolution of $2704 \\times 2028$ , with each camera recording a 10-second video. Six scenes from this dataset are publicly available. For a fair comparison, we downsampled the images to $1352 \\times 1014$ resolution in our experiments to keep the same setting from the concurrent work of 4D Gaussian [35].", + "bbox": [ + 75, + 223, + 470, + 329 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "HyperNeRF dataset [28] This dataset uses a monocular camera (e.g., iPhone) to record real-world motions, which includes real rigid and non-rigidly deforming scenes, such as a person splitting a cookie. The dataset is rather challenging due to large motions, complex lighting conditions, and thin object structures. To ensure a fair comparison, we downsampled images to $540 \\times 960$ in our experiments and followed the training and validation camera split provided by [28]. We conducted experiments on the four \"vrig\" scenes, and we also provided results of the \"interp\" scenes in the supplementary material.", + "bbox": [ + 75, + 351, + 470, + 517 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 529, + 230, + 544 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Deformation Models The deformation model is the core component of the proposed Gaussian-Flow. We conduct ablation studies to validate the effectiveness of the particular choice of DDDM. As DDDM consists of a Fourier series and a polynomial function, hence we first study the Fourier series and the polynomial functions separately, in which we only use the Fourier series or the polynomial function as the deformation model. As shown in Figure. 4, the Fourier series contains more high-frequency components than the polynomial function, thus the Fourier series has sharper image details but results in more artifacts. The polynomial function is smoother than the Fourier series, leading to fewer artifacts but burry scene renderings. Finally, the hybrid DDDM function is able to generate sharper details with fewer artifacts.", + "bbox": [ + 75, + 551, + 468, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Furthermore, we study the orders of the polynomial and Fourier series functions in our DDDM, which are related to the complexity of the scene and are crucial to the final performance. As shown in Figure. 5, the performance of our method increases with the number of Fourier series orders but starts to drop after the order over 32, which should be related to the over-parameterization of the deformation model.", + "bbox": [ + 75, + 779, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3bd3706bbd2d1872071390dc232a03b4643e2ff896e4d6c41132b5b9c09b8a62.jpg", + "image_caption": [ + "Figure 4. Ablation study on different deformation models. From left to right are deformation fitting with polynomial function only, Fourier series only, and our dual-domain deformation fitting. The proposed DDDM achieves the best rendering quality qualitatively." + ], + "image_footnote": [], + "bbox": [ + 516, + 88, + 880, + 383 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8818a3012ee2e26b2c1c54429a121634cf9494ac293c2df8b6b1f0b84e4e1706.jpg", + "image_caption": [ + "Figure 5. Ablation study on different orders of the DDDM. We find an order number of 16 leads to the best novel view rendering results in the HyperNeRF dataset." + ], + "image_footnote": [], + "bbox": [ + 504, + 468, + 888, + 627 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Regularizations Next, we study the effectiveness of the two proposed regularizations in Gaussian-Flow optimization. As shown in Table. 1, adding separate KNN rigid regularization or the time smooth regularization can both improve the novel view rendering quality, and the full model that contains both regularizations can achieve the best performance.", + "bbox": [ + 498, + 710, + 890, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Quantitative Comparisons", + "text_level": 1, + "bbox": [ + 500, + 830, + 740, + 847 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our method with SOTA on D-NeRF [30] synthetic scenes. As presented in Table 2, our method is benchmarked against TiNeuVox, K-Planes, and 4D-GS.", + "bbox": [ + 498, + 854, + 892, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21141", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c297ff6e7f4f0f37e6482425de21161bb484ea3dbd01006f2b04cb22a4dc1300.jpg", + "table_caption": [ + "Table 1. Ablation study on the proposed KNN rigid and time smooth regularizations. The quantitative results demonstrate the effectiveness of both regularizations." + ], + "table_footnote": [], + "table_body": "
reg,w/o KNN rigidw/o Time smoothfull model
PSNR28.4829.1229.92
SSIM0.920.930.94
", + "bbox": [ + 84, + 142, + 460, + 200 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our method consistently outperforms other methods across all metrics.", + "bbox": [ + 76, + 208, + 468, + 239 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c2e7449e7efe6089bb98050adbc5b803221b593b0d6ffcf74f70eb2a72fbc4b2.jpg", + "table_caption": [ + "Table 2. Quantitative results on D-NeRF synthetic blender scenes." + ], + "table_footnote": [], + "table_body": "
MethodPSNR ↑SSIM ↑LPIPS ↓
TiNeuVox [7]32.670.970.04
K-Planes [8]31.610.97/
4D-GS [35]33.300.980.03
Ours34.270.980.03
", + "bbox": [ + 78, + 281, + 468, + 375 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also compare our method against previous SOTA NeRF-based methods, including NeRF [25], Nerfies [27], HyperNeRF [28], NeRFPlayer [32], and TiNeuVox [7] on HyperNeRF [28]. We also provide comparisons with other 3DGS-based approaches concurrently proposed with our Gaussian-Flow. The training time, rendering FPS, and novel view synthesis PSNR of different methods can be found in Table. 3. Previous NeRF-based methods require at least 30 minutes to train the scene, and fail to achieve real-time rendering of the dynamic scene. Our method only requires 7 minutes of training time and can achieve real-time rendering speed, which is much faster than previous methods. Moreover, our method can achieve better performance than previous SOTA methods in terms of PSNR.", + "bbox": [ + 75, + 386, + 468, + 597 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluated various methods on the Plenoptic Video dataset, as summarized in Table. 4. The comparison focuses on training time efficiency and image quality, assessed through PSNR and SSIM. Our approach demonstrated a significant advancement in training efficiency, requiring only 22.5 minutes, a drastic reduction compared to the hours needed by methods like DyNeRF [18] and K-Planes [8]. This efficiency is paramount for practical applications, where reduced training time can be a critical factor. In terms of image quality, our method achieved a PSNR of 30.5 with 30K steps, which, while not the highest, is competitive with the leading methods. However, our method scored 0.97 in SSIM, higher than K-Planes' leading score of 0.96. This indicates a potential trade-off between training efficiency and the ability to preserve structural details in images.", + "bbox": [ + 75, + 598, + 468, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition, we extend our method to 60K steps on both datasets, which can further improve the performance on the HyperNeRF dataset [28], and achieve the highest performance on the Plenoptic Video dataset [18]. However, the", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "training time is also increased by approximately $2 \\times$ for HyperNeRF dataset [28] and about $1.5 \\times$ for Plenoptic Video dataset [18], which is still much faster than previous methods.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Qualitative Comparisons", + "text_level": 1, + "bbox": [ + 500, + 157, + 728, + 174 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we show qualitative comparisons of our method and previous SOTA methods on the HyperNeRF [28] dataset and the Plenoptic Video dataset [18]. Figure. 6 shows the qualitative comparisons of our method and TiNeuVox [7], 4D-GS [35] on HyperNeRF dataset. Notice that our method can produce comparably clear and sharp images than previous SOTA methods, which highlights our method's superior performance in monocular conditions. Despite its overall efficacy, our method does encounter limitations in extremely thin structure, as shown in the 3D Printer scene, the thread of the 3D Printer is not clear, while other methods can produce a clear thread.", + "bbox": [ + 496, + 181, + 890, + 363 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7dbcb503f35e71708120636ecac26a1288ab19c55523199170bbbc4fd77c0d3b.jpg", + "image_caption": [ + "Figure 6. Qualitative comparisons of our method and TiNeVox [7] and 4D-GS [35] on the HyperNeRF [28] vrig. dataset" + ], + "image_footnote": [], + "bbox": [ + 522, + 375, + 870, + 584 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also compare our method with previous SOTA methods on the Plenoptic Video dataset [18], as shown in Figure. 7. Compared with previous SOTA methods, our method can produce more accurate color and correct structure. Moreover, our method successfully reconstructs the flame in the scene, while NeRFPlayer [8] fails to reconstruct the flame. These results suggest that our method can achieve comparable image quality with previous SOTA methods, which demonstrates the effectiveness of our method in the multiview conditions.", + "bbox": [ + 496, + 637, + 890, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 800, + 617, + 814 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this paper, we introduced Gaussian-Flow, a novel framework for dynamic 3D scene reconstruction using a point-based differentiable rendering approach. The core of our innovation lies in the DDDM, which efficiently models deformations of each 3D Gaussian point in both the time and", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21142", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c6323032c553b14af0ff01a30d42ed7430389d73942c93be8cd8e6e482ffb66b.jpg", + "table_caption": [ + "Table 3. Per-scene quantitative comparisons on HyperNeRF [28] dataset. Results are gathered from papers of the corresponding methods. Our method achieves the fastest training time, the highest rendering FPS, and the highest PSNR score for novel view synthesis, setting a new state-of-the-art for image-based dynamic scene rendering." + ], + "table_footnote": [], + "table_body": "
MethodTrain Time↓FPS↑Broom3D PrinterChickenPeel BananaMean
PSNR↑SSIM ↑PSNR↑SSIM ↑PSNR↑SSIM ↑PSNR↑SSIM ↑PSNR↑SSIM ↑
NeRF [25]16 hours0.01319.90.65320.70.78019.90.77720.00.76920.10.745
Nerfies [27]16 hours0.01119.20.56720.60.83026.70.94322.40.87222.20.803
HyperNeRF [28]32 hours0.01119.30.59120.00.82126.90.94823.30.89622.40.814
NeRFPlayer [32]6 hours0.20821.70.63522.90.81026.30.90524.00.86323.70.803
TiNeuVox [7]30 min0.521.50.68622.80.84128.30.94724.40.87324.30.837
Ours (30K)7 min12522.50.69024.30.85729.40.93426.30.90625.60.847
Ours (60K)12 min12522.80.70925.00.87730.40.94527.00.91726.30.862
", + "bbox": [ + 86, + 147, + 883, + 268 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7690f4ca9fd93d0e8eff58e78067e1f36103f3de8ee0121aee4262fc12b102cc.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 176, + 282, + 380, + 401 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/aedc63b4ada16e09833bcb298d59c959d6b1988819d96b588e365e414eaf475e.jpg", + "image_caption": [ + "Ours (30K)" + ], + "image_footnote": [], + "bbox": [ + 383, + 282, + 586, + 401 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cfb3fff214c6e1c7b64af5abb8d4b1dcdffec73cebb80aac0aa7baeddfe6cb88.jpg", + "image_caption": [ + "K-Planes" + ], + "image_footnote": [], + "bbox": [ + 588, + 282, + 792, + 401 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e292cdab9c5421f6ae4efcbbcc426272c97b3e59b836184ba6ab4c56cb5908c1.jpg", + "image_caption": [ + "NeRFPlayer" + ], + "image_footnote": [], + "bbox": [ + 176, + 416, + 380, + 535 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b2ff2c05d169d71bf100718937329c423ffd8aaa19469c16f4028896263a2a16.jpg", + "image_caption": [ + "DyNeRF" + ], + "image_footnote": [], + "bbox": [ + 383, + 416, + 586, + 535 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f852bacbf4bfd3a0d1bf37c7783b525190d48406a149af4e2721c9020d50c2eb.jpg", + "image_caption": [ + "LLFF", + "Figure 7. Qualitative comparisons of our method and K-Planes [8], NeRFPlayer [8], DyNeRF [18], and LLFF on Plenoptic Video dataset." + ], + "image_footnote": [], + "bbox": [ + 589, + 416, + 792, + 535 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5c650e2f1d4e7980bfebe30a77a6f326912607fdfc4540f02406459b3fd05d5f.jpg", + "table_caption": [ + "Table 4. Quantitative comparison on Plenoptic Video dataset [18]. Our training speed is $5 \\times$ faster of magnitude faster than previous leading approaches. Also, we achieved the highest PSNR score among all methods." + ], + "table_footnote": [], + "table_body": "
MethodTrain Time ↓PSNR ↑SSIM ↑
LLFF [24]-23.2-
DyNeRF [18]1344 hours29.60.96
NeRFPlayer [32]5.5 hours30.7-
K-Planes [8]1.8 hours31.60.96
4D-GS [35]2 hours31.00.94
Ours (30K)22.5 min30.50.97
Ours (60K)41.8 min32.00.97
", + "bbox": [ + 78, + 672, + 468, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "frequency domains. This approach has enabled us to set a new state-of-the-art for 4D scene reconstruction in terms of training speed, rendering frames per second, and novel view synthesis quality. Our extensive experiments and ab", + "bbox": [ + 75, + 839, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "lation studies have demonstrated the efficacy of proposed Gaussian-Flow across various datasets. We achieved significant improvements over existing methods, particularly in training speed and rendering performance. The ability to efficiently handle dynamic scenes without the computational overhead of neural networks marks a substantial leap forward in this domain.", + "bbox": [ + 496, + 602, + 890, + 708 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Limitations", + "text_level": 1, + "bbox": [ + 500, + 723, + 622, + 738 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While our method excels in rendering speed and training efficiency, there is room for improvement in maintaining high-fidelity thin structures in the final rendering, as mentioned in the Section. 4.5. The current deformation model may not be able to capture the fine details of thin structures, leading to some loss of image quality. This limitation is particularly evident in scenes with complex motion patterns. Future work could focus on enhancing the balance between speed and image detail preservation, potentially through more refined deformation models or advanced regularization techniques.", + "bbox": [ + 496, + 747, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21143", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 119, + 173, + 136 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Benjamin Attal, Jia-Bin Huang, Christian Richardt, Michael Zollhoefer, Johannes Kopf, Matthew O'Toole, and Changil Kim. HyperReel: High-fidelity 6-DoF video with ray-conditioned sampling. arXiv preprint arXiv:2301.02238, 2023. 3", + "[2] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. arXiv:2301.09632, 2023. 2", + "[3] Ang Cao and Justin Johnson. Hexplane: a fast representation for dynamic scenes. arXiv preprint arXiv:2301.09632, 2023. 3", + "[4] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In Proceedings of the European Conference on Computer Vision, 2022. 2, 3", + "[5] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 14304-14314. IEEE Computer Society, 2021. 2, 4", + "[6] Jiemin Fang, Lingxi Xie, Xinggang Wang, Xiaopeng Zhang, Wenyu Liu, and Qi Tian. Neusample: Neural sample field for efficient view synthesis. arXiv:2111.15552, 2021. 2", + "[7] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. arXiv preprint arXiv:2205.15285, 2022. 7, 8", + "[8] Sara Fridovich-Keil, Giacomo Meanti, Frederik Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. arXiv preprint arXiv:2301.10241, 2023. 3, 7, 8", + "[9] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance, 2023. 2", + "[10] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5712-5721, 2021. 2, 4", + "[11] Stephan J. Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14346-14355, 2021. 2", + "[12] Peter Hedman, Pratul P. Srinivasan, Ben Mildenhall, Jonathan T. Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5875-5884, 2021. 2", + "[13] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 42(4):1-14, 2023. 3", + "[14] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time" + ], + "bbox": [ + 78, + 145, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 2, 3", + "[15] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations, 2015. 5", + "[16] Andreas Kurz, Thomas Neff, Zhaoyang Lv, Michael Zollhofer, and Markus Steinberger. Adanerf: Adaptive sampling for real-time rendering of neural radiance fields. In European Conference on Computer Vision, 2022. 2", + "[17] Marc Levoy and Turner Whitted. The use of points as a display primitive. 1985. 3", + "[18] Tianye Li, Mira Slavcheva, Michael Zollhöfer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, and Zhaoyang Lv. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5521-5531, 2022. 1, 2, 6, 7, 8", + "[19] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2, 4", + "[20] D. B.* Lindell, J. N. P.* Martel, and G. Wetzstein. Autoint: Automatic integration for fast neural volume rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 2", + "[21] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. In Advances in Neural Information Processing Systems, 2020. 2, 3", + "[22] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural volumes: learning dynamic renderable volumes from images. ACM Transactions on Graphics, 38(4):1-14, 2019. 2, 4", + "[23] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 2, 3", + "[24] Ben Mildenhall, Pratul P. Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (Proceedings of SIGGRAPH), 38(4), 2019. 8", + "[25] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision, pages 405-421. Springer, 2020. 2, 7, 8", + "[26] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 2, 3", + "[27] Keunhong Park, Utkarsh Sinha, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Steven M. Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 2, 7, 8" + ], + "bbox": [ + 503, + 92, + 890, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21144", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo Martin-Brulla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Trans. Graph., 40(6), 2021. 1, 2, 6, 7, 8", + "[29] Martin Piala and Ronald Clark. Terminerf: Ray termination prediction for efficient neural rendering. International Conference on 3D Vision, pages 1106-1114, 2021. 2", + "[30] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 1, 2, 6", + "[31] Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boning Liu, Hongwen Zhang, and Yebin Liu. Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. arXiv preprint arXiv:2211.11610, 2022. 3", + "[32] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerfplayer: A streamable dynamic scene representation with decomposed neural radiance fields. arXiv preprint arXiv:2210.15947, 2022. 3, 7, 8", + "[33] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 2", + "[34] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Improved direct voxel grid optimization for radiance fields reconstruction. arXiv preprint arXiv:2206.05085, 2022. 2", + "[35] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528, 2023. 2, 3, 6, 7, 8", + "[36] Liwen Wu, Jae Yong Lee, Anand Bhattachad, Yu-Xiong Wang, and David Forsyth. Diver: Real-time and accurate neural radiance fields with deterministic integration for volume rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16200-16209, 2022. 2", + "[37] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9421-9431, 2021. 2, 4", + "[38] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101, 2023. 2, 3", + "[39] Wang Yifan, Felice Serena, Shihao Wu, Cengiz Öz Tireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM Transactions on Graphics (TOG), 38(6):1-14, 2019. 3", + "[40] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5752-5761, 2021. 2, 3", + "[41] Qiang Zhang, Seung-Hwan Baek, Szymon Rusinkiewicz, and Felix Heide. Differentiable point-based radiance fields for efficient view synthesis. In SIGGRAPH Asia 2022 Conference Papers, 2022. 3", + "[42] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Ewa volume splatting. In Proceedings Visualization, 2001. VIS'01., pages 29-538. IEEE, 2001. 3" + ], + "bbox": [ + 501, + 92, + 890, + 233 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21145", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Gaussian-Flow_ 4D Reconstruction with Dynamic 3D Gaussian Particle/3e2469b7-4042-4a31-bb56-8ba156724067_model.json b/2024/Gaussian-Flow_ 4D Reconstruction with Dynamic 3D Gaussian Particle/3e2469b7-4042-4a31-bb56-8ba156724067_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a3d22124a6754dd0e20272997dc074308d2738e7 --- /dev/null +++ b/2024/Gaussian-Flow_ 4D Reconstruction with Dynamic 3D Gaussian Particle/3e2469b7-4042-4a31-bb56-8ba156724067_model.json @@ -0,0 +1,2233 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.125, + 0.131, + 0.845, + 0.152 + ], + "angle": 0, + "content": "Gaussian-Flow: 4D Reconstruction with Dynamic 3D Gaussian Particle" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.18, + 0.335, + 0.198 + ], + "angle": 0, + "content": "Youtian Lin1,2" + }, + { + "type": "text", + "bbox": [ + 0.373, + 0.181, + 0.487, + 0.198 + ], + "angle": 0, + "content": "Zuozhuo Dai3" + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.181, + 0.611, + 0.198 + ], + "angle": 0, + "content": "Siyu Zhu" + }, + { + "type": "text", + "bbox": [ + 0.653, + 0.181, + 0.752, + 0.198 + ], + "angle": 0, + "content": "Yao \\(\\mathrm{Yao^{1,2\\boxtimes}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.199, + 0.844, + 0.27 + ], + "angle": 0, + "content": "\\(^{1}\\)State Key Laboratory for Novel Software Technology, Nanjing University, Nanjing, China \n\\(^{2}\\)School of Intelligence Science and Technology, Nanjing University, Suzhou, China \n\\(^{3}\\)Alibaba Group \n\\(^{4}\\)Fudan University" + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.306, + 0.35, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.308, + 0.583, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.306, + 0.736, + 0.545 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.736, + 0.307, + 0.89, + 0.545 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.556, + 0.893, + 0.6 + ], + "angle": 0, + "content": "Figure 1. Dynamic reconstruction results of the proposed Gaussian-Flow on the monocular D-NeRF [30] and HyperNeRF Dataset [28] (left) and the multi-view Plenoptic Dataset [18] (right). Our method achieves a \\(5 \\times\\) faster training and rendering speed compared with the per-frame 3DGS modeling and significantly outperforms previous methods in novel view rendering quality." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.611, + 0.314, + 0.627 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.644, + 0.473, + 0.903 + ], + "angle": 0, + "content": "We introduce Gaussian-Flow, a novel point-based approach for fast dynamic scene reconstruction and real-time rendering from both multi-view and monocular videos. In contrast to the prevalent NeRF-based approaches hampered by slow training and rendering speeds, our approach harnesses recent advancements in point-based 3D Gaussian Splatting (3DGS). Specifically, a novel Dual-Domain Deformation Model (DDDM) is proposed to explicitly model attribute deformations of each Gaussian point, where the time-dependent residual of each attribute is captured by a polynomial fitting in the time domain, and a Fourier series fitting in the frequency domain. The proposed DDDM is capable of modeling complex scene deformations across long video footage, eliminating the need for training separate 3DGS for each frame or introducing an additional implicit neural field to model 3D dynamics. Moreover, the explicit deformation modeling for discretized Gaussian points en" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.895, + 0.763 + ], + "angle": 0, + "content": "sures ultra-fast training and rendering of a 4D scene, which is comparable to the original 3DGS designed for static 3D reconstruction. Our proposed approach showcases a substantial efficiency improvement, achieving a \\(5 \\times\\) faster training speed compared to the per-frame 3DGS modeling. In addition, quantitative results demonstrate that the proposed Gaussian-Flow significantly outperforms previous leading methods in novel view rendering quality. Project page: https://nju-3dv.github.io/projects/Gaussian-Flow." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.798, + 0.633, + 0.814 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.901 + ], + "angle": 0, + "content": "In the realm of digital scene synthesis, achieving a balance between high-quality reconstructions and real-time rendering is paramount, especially for applications like virtual reality (VR) playback, where immediate feedback and immersive experiences are essential. Neural Radiance Fields" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21136" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.288 + ], + "angle": 0, + "content": "(NeRFs) [25] has risen as a promising method for synthesizing intricate scenes. However, despite their ability to produce visually stunning results, NeRFs require costly sampling and evaluation of the neural radiance field at multiple points along each ray. Consequently, the substantial computational demands impede the real-time rendering capabilities of NeRFs. These are attempts to accelerate the rendering process of NeRFs, such as direct volume representation [21, 33, 40], neural hasing [26], and tri-plane structures [2, 4, 9]. However, it still remains a challenge for high-fidelity real-time rendering. Nevertheless, the issue becomes even more severe when turning to dynamic scene reconstruction and rendering." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.292, + 0.473, + 0.504 + ], + "angle": 0, + "content": "Recent progress on 3D Gaussian Splatting (3DGS) [14] has drawn attention from the 3D computer vision community. With tile-based rasterization instead of plain volume rendering, 3DGS can render images two orders of magnitude faster than the vanilla NeRF. The technique has also been quickly applied to 4D scene reconstruction by extending to the separate per-frame 3DGS optimization [23]. However, such direct extension is storage-intensive and is not applicable to monocular video input. Some other concurrent works [35, 38] try to mix the explicit point-based 3DGS and an implicit neural field for dynamic information modeling, however, require computationally expensive forward passes of the neural network, which significantly lowers the rendering speed of the original 3DGS." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.509, + 0.473, + 0.903 + ], + "angle": 0, + "content": "In this work, we propose Gaussian-Flow, an explicit particle-based deformation model designed specifically for 3DGS to model the dynamic scene without using any neural network. Gaussian-Flow can recover a high-fidelity 4D scene from captured videos, while still preserving the ultrafast training and rendering speed of the original 3DGS. In particular, we formulate a 4D scene as a set of deformable 3D Gaussian points. A novel Dual-Domain Deformation Model (DDDM) is proposed to explicitly model deformations of each Gaussian point's attributes, including position, rotation, and radiance. The time-dependent deformation residual is modeled simultaneously in time and frequency domains: we apply joint polynomial and Fourier series fitting for each deformable attribute. This compact dynamic representation greatly reduces the computation cost of the deformation model, which is a key factor in preserving the rendering speed of 3DGS. Moreover, an adaptive times-tamp scaling technique is introduced to avoid over-fitting the scene to only frames with violent motions. For robust estimation, we also regularize the motion trajectory by the KNN-based rigid and the time smooth constraints. It is also noteworthy that our discretized point-based 4D representation naturally supports the edition of both static and dynamic 3D scenes, showing the potential for unlocking a variety of downstream applications related to dynamic 3D reconstruction and rendering." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.212 + ], + "angle": 0, + "content": "We have conducted extensive experiments to demonstrate the effectiveness of the proposed method on several multi-view and monocular datasets. The proposed Gaussian-Flow achieves a \\(5 \\times\\) faster training speed compared with the separate per-frame 3DGS modeling, and significantly outperforms prior leading methods in novel view rendering quality. Our major contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.213, + 0.892, + 0.287 + ], + "angle": 0, + "content": "- We introduce Gaussian-Flow, a novel point-based differentiable rendering approach for dynamic 3D scene reconstruction, setting a new state-of-the-art for training speed, rendering FPS, and novel view synthesis quality for 4D scene reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.289, + 0.892, + 0.363 + ], + "angle": 0, + "content": "- We propose a Dual-Domain Deformation Model for efficient 4D scene training and rendering, eliminating the need for per-frame 3DGS optimization and sampling on implicit neural fields. This preserves a running speed on par with the original 3DGS with minimum overhead." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.365, + 0.892, + 0.409 + ], + "angle": 0, + "content": "- We demonstrate that our discretized point-based representation supports the segmentation, edition, and composition of both static and dynamic 3D scenes." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.213, + 0.892, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.424, + 0.651, + 0.44 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.45, + 0.783, + 0.466 + ], + "angle": 0, + "content": "2.1. Dynamic Neural Radiance Field" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.473, + 0.892, + 0.732 + ], + "angle": 0, + "content": "Dynamic NeRF modeling has become a heated research topic in recent years due to the development of the neural radiance field and differentiable rendering. By treating time as an extended input dimension to NeRF, researchers successfully achieve qualified image-based 4D scene rendering [5, 10, 19, 22, 37]. To further improve the reconstruction quality and incorporate prior knowledge of motions and structures, dynamic neural scene flow methods have been proposed [27, 30], where a canonical space is constructed and then transferred to each frame with scene flow or motion fields. HyperNeRF [28] models the deformation of the object topologies by using higher-dimensional inputs, while DyNeRF [18] utilizes time-conditioned NeRF to represent a 4D scene. However, the aforementioned approaches are all based on the vanilla NeRF, which requires a long training time and does not meet the requirement of real-time rendering." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.741, + 0.805, + 0.756 + ], + "angle": 0, + "content": "2.2. Accelerated Neural Radiance Field" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.903 + ], + "angle": 0, + "content": "To expedite NeRF training and rendering, numerous approaches have been suggested, employing more streamlined strategies [6, 11, 12, 16, 20, 29, 36]. Other methods propose the integration of neural implicit functions with explicit 3D structures, forming a hybrid representation for faster radiance field sampling [4, 26, 33, 34]. These approaches establish strong foundations for enhancing dynamic NeRF, concurrently decreasing required training and inference time. Apart from implicit neural representations, explicit NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21137" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.167 + ], + "angle": 0, + "content": "modeling has show promising results for real-time rendering: NSVF [21] employs a neural sparse voxel field for efficient NeRF sampling, which stands for the earliest attempt on the explicit NeRF modeling; PlenOctrees [40] utilizes the explicit octree structure for rendering acceleration." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.47, + 0.439 + ], + "angle": 0, + "content": "Recent efforts have also emerged to accelerate the intricate dynamic neural radiance field. TensorRF [4] employs multiple planes as explicit representations for direct dynamic scene modeling. More recent recent approaches of this kind include K-Planes [8], Tensor4D [31], and HexPlane [3]. Alternatively, NeRFPlayer[32] introduces a unified streaming representation for both grid-based [26] and plane-based methods, utilizing separate models to distinguish static and dynamic scene components, however, leading to slow rendering times. HyperReel [1] further suggests a flexible sampling network coupled with two planes for dynamic scene representation. While these methods improve the rendering speed of a dynamic scene to some extent, it is still hard to achieve real-time rendering, let alone a good balance between the running speed and rendering quality. In contrast, we resort to the recent 3D Gaussians splatting, which applies an explicit soft point cloud representation for real-time image-based rendering." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.447, + 0.402, + 0.463 + ], + "angle": 0, + "content": "2.3. Differentiable Point-based Rendering" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.47, + 0.47, + 0.803 + ], + "angle": 0, + "content": "The original idea of using 3D points as rendering primitives was first introduced in [17]. By incorporating differentiable rendering, recent approaches have made remarkable progress in image-based rendering, representative methods include PointRF [41], DSS [39], and 3D Gaussians splatting (3DGS) [13]. Specifically, 3DGS [13] has demonstrated extraordinary performance in novel-view synthesis, achieving real-time rendering speed and state-of-the-art rendering quality. The method adopts a soft point representation with attributes of position, rotation, density, and radiance, and applies differentiable point-based rendering for scene optimization. 3DGS has quickly been extended to dynamic scene modeling by direct separate per-frame optimization [23], however, requires a long optimization time and a large amount of storage for long video footage. Other works [35, 38] apply an implicit motion field to model scene dynamics, but the introduction of the implicit neural network significantly slows down the sampling and rendering speed. In this work, we adopt the approach of representing 4D scenes through a purely discretized point cloud model, ensuring a fast training and rendering speed comparable with the original 3DGS." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.816, + 0.227, + 0.831 + ], + "angle": 0, + "content": "3. Gaussian-Flow" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.468, + 0.901 + ], + "angle": 0, + "content": "In this section, we introduce the proposed Gaussian-Flow for dynamic scene modeling. We first review the 3DGS in Sec. 3.1. Then, we introduce our explicit motion modeling of each Gaussian point by using a novel Dual-Domain" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.183 + ], + "angle": 0, + "content": "Deformation Model (DDDM), as outlined in Sec. 3.2. An adaptive timestamp scaling technique is described in Sec. 3.3 for balanced training of each frame. To ensure the continuity of the motion in both spatial and temporal dimensions, we incorporate appropriate regularizations on each point during the optimization, as detailed in Sec. 3.4." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.192, + 0.785, + 0.209 + ], + "angle": 0, + "content": "3.1. Recap on 3D Gaussian Splitting" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.214, + 0.892, + 0.397 + ], + "angle": 0, + "content": "3D Gaussian Splatting [14] is designed to efficiently optimize a 3D scene for real-time and high-quality novel view synthesis. The 3DGS framework has garnered significant attention within the community due to its remarkable enhancements in both training and rendering times, concurrently achieving state-of-the-art rendering quality. In contrast to the volume rendering in the vanilla NeRF which relies on ray marching, 3DGS adopts a tile-based rasterization on a distinctive soft point cloud representation to achieve fast rendering. Specifically, 3DGS models a 3D scene as a large amount of 3D Gaussian points in the world space, where each point is represented by:" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.406, + 0.892, + 0.435 + ], + "angle": 0, + "content": "\\[\nG (\\boldsymbol {x}) = \\exp \\left(- \\frac {1}{2} \\left(\\boldsymbol {x} - \\boldsymbol {\\mu}\\right) ^ {T} \\boldsymbol {\\Sigma} ^ {- 1} \\left(\\boldsymbol {x} - \\boldsymbol {\\mu}\\right)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.444, + 0.892, + 0.58 + ], + "angle": 0, + "content": "where \\(\\mu\\) and \\(\\Sigma\\) are the mean position and covariance matrix of a 3D Gaussian particle. The 3DGS takes sparse Structure-from-Motion (SfM) points or even random points as input, and initializes each point to a 3D Gaussian based on its neighbors. Besides, each 3D Gaussian is associated with a learnable view-dependent radiance \\(c\\) and a learnable opacity \\(\\alpha\\) for rendering. Subsequently, an efficient 3D to 2D Gaussian mapping [42] is employed to project the point onto the image plane:" + }, + { + "type": "equation", + "bbox": [ + 0.649, + 0.592, + 0.891, + 0.608 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\mu} ^ {\\prime} = \\boldsymbol {P} \\boldsymbol {W} \\boldsymbol {\\mu}, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.622, + 0.617, + 0.891, + 0.634 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Sigma} ^ {\\prime} = \\boldsymbol {J} \\boldsymbol {W} \\boldsymbol {\\Sigma} \\boldsymbol {W} ^ {T} \\boldsymbol {J} ^ {T}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.641, + 0.892, + 0.732 + ], + "angle": 0, + "content": "where \\(\\mu^{\\prime}\\) and \\(\\Sigma^{\\prime}\\) represent the 2D mean position and 2D covariance of the projected 3D Gaussian. \\(P, W\\) and \\(J\\) donate the projective transformation, viewing transformation, and Jacobian of the affine approximation of \\(P\\). After that, \\(\\alpha\\)-blending is executed to merge the overlapping Gaussians for each pixel, yielding a final color:" + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.743, + 0.892, + 0.786 + ], + "angle": 0, + "content": "\\[\nC = \\sum_ {i = 1} ^ {n} \\boldsymbol {c} _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {i}), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.89, + 0.825 + ], + "angle": 0, + "content": "where \\( c_{i} \\) and \\( \\alpha_{i} \\) are the color and opacity of the \\( i \\)-th Gaussian, and \\( n \\) is the number of overlapping Gaussians." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The attributes of the 3D Gaussian, including the mean \\(\\mu\\), covariance matrix \\(\\Sigma\\), opacity \\(\\alpha\\), and color \\(c\\), are optimized through backward propagation of the gradient flow. In particular, to ensure positive definiteness during the optimization process, the covariance matrix is parameterized" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21138" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "as a scaling vector \\( s \\) and a rotation matrix \\( R \\), i.e., \\( \\Sigma = R\\Lambda(s)\\Lambda(s)^T R^T \\), where \\( \\Lambda(s) \\) is the diagonal matrix of \\( s \\). To facilitate optimization, the rotation matrix is further parameterized using a quaternion \\( q \\). Leveraging the inherent flexibility of discrete points, 3DGS incorporates adaptive density control for points. This mechanism utilizes the gradient flow to identify where geometric reconstruction is suboptimal, and employs cloning and splitting to augment the density of points for a higher rendering quality." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.243, + 0.373, + 0.259 + ], + "angle": 0, + "content": "3.2. Dual-Domain Deformation Model" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.269, + 0.473, + 0.509 + ], + "angle": 0, + "content": "We target to directly model the dynamics of each 3D Gaussian point by fitting each of its attributes into a time-dependent curve. Among different approaches, Polynomials fitting in the time domain and Fourier series fitting in the frequency domain are the two most widely used approaches [5, 10, 19, 22, 37], due to their simplicity and effectiveness. However, each method comes with its own advantages and drawbacks: describing the motion of a Gaussian particle in terms of polynomials yields a good fit with smooth motion with a small order of polynomials, however, can easily overfit to a violent motion if assuming a larger order of polynomials, resulting in unreasonable oscillations in the fitted trajectory. Whereas, the Fourier series excels at capturing the variations associated with violent motion, however, requires a manually reduced order when dealing with smooth motion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.512, + 0.473, + 0.74 + ], + "angle": 0, + "content": "In this work, our key insight is to use a Dual-Domain Deformation Model (DDDM) for fitting the scene dynamics, which integrates both the time domain polynomials and the frequency domain Fourier series into a unified fitting model. We assume that only the rotation \\( \\pmb{q} \\), radiance \\( c \\), and position \\( \\pmb{\\mu} \\) of a 3D Gaussian particle change over time, while the scaling \\( s \\) and opacity \\( \\alpha \\) remain constant. Specifically, we conceptualize the change in each particle's attributes as its base attributes \\( S_0 \\in \\{\\pmb{\\mu}_0, c_0, q_0\\} \\) at the reference time frame \\( t_0 \\) (usually set to the first frame), superimposed on a time-dependent attribute residual \\( D(t) \\). For simplicity, we use lowercase characters to represent a single attribute in \\( S \\). The time-dependent residual of each attribute is modeled through polynomial fitting in the time domain and Fourier series fitting in the frequency domain, expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.208, + 0.755, + 0.47, + 0.772 + ], + "angle": 0, + "content": "\\[\nS (t) = S _ {0} + D (t), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.787, + 0.472, + 0.849 + ], + "angle": 0, + "content": "where \\(D(t) = P_N(t) + F_L(t)\\) is combined by a polynomial \\(P_N(t)\\) with coefficients \\(\\pmb{a} = \\{a\\}_{n=0}^{N}\\) and a Fourier series \\(F_L(t)\\) with coefficients \\(\\pmb{f} = \\{f_{sin}^l, f_{cos}^l\\}_{l=0}^L\\). These are respectively denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.207, + 0.863, + 0.47, + 0.905 + ], + "angle": 0, + "content": "\\[\nP _ {N} (t) = \\sum_ {n = 0} ^ {N} a _ {n} t ^ {n}, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.09, + 0.892, + 0.131 + ], + "angle": 0, + "content": "\\[\nF _ {L} (t) = \\sum_ {l = 1} ^ {L} \\left(f _ {\\sin} ^ {l} \\cos (l t) + f _ {\\cos} ^ {l} \\sin (l t)\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.138, + 0.892, + 0.213 + ], + "angle": 0, + "content": "It is important to note that we assume different dimensions of an attribute are independently changed over time. Therefore, we assign a different \\( D(t) \\) for each dimension of an attribute. For instance, we utilize \\( \\{D_{\\mu_i}(t)\\}_{i=0}^3 \\) to describe the motion of a 3D position \\( \\pmb{\\mu} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.214, + 0.892, + 0.304 + ], + "angle": 0, + "content": "Figure 3 illustrates a comparative analysis of trajectory fitting using polynomial, Fourier series, and the proposed joint DDDM functions. The figure highlights the superior fitting capabilities of the DDDM approach in capturing complex motion trajectories as represented by the sampled data points." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.316, + 0.757, + 0.332 + ], + "angle": 0, + "content": "3.3. Adaptive Timestemp Scaling" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.339, + 0.892, + 0.52 + ], + "angle": 0, + "content": "In a typical scenario, a normalized frame index \\( t \\) ranging from 0 to 1 will be used as the temporal input of \\( D(t) \\). However, this poses a challenge when endeavoring to model substantial motions within a very short time using polynomials and Fourier series. Adhering to the standard temporal division would necessitate an exceedingly large coefficient to accommodate highly intense movements within a very short time frame. This circumstance has the potential to induce instability or even a breakdown in the optimization process. To address this issue, we introduce a time dilation factor \\( \\lambda \\) to scale the temporal input for each Gaussian point, which is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.641, + 0.535, + 0.892, + 0.55 + ], + "angle": 0, + "content": "\\[\nt _ {s} = \\lambda_ {s} \\cdot t + \\lambda_ {b} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.563, + 0.892, + 0.639 + ], + "angle": 0, + "content": "where \\( t_s \\) represents the scaled time input, serving as the input of \\( D(t) \\), \\( t \\in [0,1] \\) denotes the normalized frame index, and \\( \\lambda \\) and \\( \\lambda_b \\) stand for the dilation factor and base factor of a Gaussian, respectively. In all our experiments, \\( \\lambda_s \\) and \\( \\lambda_b \\) are initialized to 1 and 0, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.64, + 0.892, + 0.775 + ], + "angle": 0, + "content": "To summarize, in our dynamic scene setting, a Gaussian particle contains multiple attributes to be optimized, including base attributes \\(\\{\\mu_0,q_0,s_0,c_0,\\alpha_0\\}\\) at reference frame \\(t_0\\), polynomial coefficients and Fourier coefficients in \\(\\{D_{\\mu}(t),D_q(t),D_c(t)\\}\\). Since the proposed DDDM optimizes the time-dependent residuals without the need for an intricate neural field structure, our Gaussian-Flow inherits the benefits of extreme-fast training and rendering speed from the vanilla 3DGS." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.659, + 0.803 + ], + "angle": 0, + "content": "3.4. Regularizations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "While the utilization of discrete points as scene representations accelerates rendering, several challenges remain. First, points are optimized individually, losing connections with their spatial neighbors, which do not align with the real-world scenario. Optimizing these Gaussian points without considering continuity will inevitably lead to a" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21139" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.095, + 0.113, + 0.112 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.097, + 0.277, + 0.111 + ], + "angle": 0, + "content": "Adaptive Timestamp Scaling" + }, + { + "type": "image", + "bbox": [ + 0.294, + 0.097, + 0.341, + 0.11 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.35, + 0.097, + 0.534, + 0.11 + ], + "angle": 0, + "content": "Dual-Domain Deformation Model" + }, + { + "type": "image", + "bbox": [ + 0.548, + 0.097, + 0.594, + 0.11 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.6, + 0.097, + 0.652, + 0.11 + ], + "angle": 0, + "content": "Radiance" + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.097, + 0.707, + 0.11 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.715, + 0.097, + 0.762, + 0.11 + ], + "angle": 0, + "content": "Forward" + }, + { + "type": "image", + "bbox": [ + 0.77, + 0.097, + 0.816, + 0.11 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.824, + 0.097, + 0.882, + 0.11 + ], + "angle": 0, + "content": "Backward" + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.127, + 0.895, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.316, + 0.895, + 0.36 + ], + "angle": 0, + "content": "Figure 2. Overview of the Gaussian-Flow pipeline. We model the deformation of attributes of each 3D Gaussian point independently by using the Dual-Domain Deformation Model (DDDM), which preserves the discretized nature of the 3D Gaussian points, and thus achieves ultra-fast training and rendering speed comparable with the original 3DGS." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.395, + 0.266, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.398, + 0.454, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.538, + 0.471, + 0.582 + ], + "angle": 0, + "content": "Figure 3. Two exemplar motion fittings using polynomial, Fourier Series, and the joint DDDM functions. Our DDDM is able to accurately fit complex trajectories denoted by the sampling points." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.611, + 0.47, + 0.702 + ], + "angle": 0, + "content": "degradation in reconstruction quality and spatial coherence. Additionally, motions should be smoothed over time. Based on these observations, we employ two regularizations, namely time smoothness and a KNN rigid regularization, for robust optimization of Gaussian points and their motions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.731, + 0.47, + 0.821 + ], + "angle": 0, + "content": "Time Smoothness Loss To ensure temporal smoothness over time, we apply a perturbing \\(\\epsilon\\) on the input timestamp \\(t\\) and encourage the time-dependent attributes (i.e., position \\(\\mu\\), rotation \\(q\\) and radiance \\(c\\)) at time \\(t + \\epsilon\\) to be consistent with those at time \\(t\\). The time smoothness term is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.181, + 0.827, + 0.47, + 0.844 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {t} = \\| D (t) - D (t + \\epsilon) \\| _ {2}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.902 + ], + "angle": 0, + "content": "It is noteworthy that the magnitude of the perturbing value is adaptively set according to the number of total frames, i.e., \\(\\epsilon = 0.1/\\) frames." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.384, + 0.895, + 0.567 + ], + "angle": 0, + "content": "KNN Rigid Loss During the optimization of 3D Gaussians with adaptive density control, points will be dynamically added or removed. This dynamic nature implies that the neighbors of points within a local space are subject to constant changes, posing a challenge for directly enforcing a spatial local consistency constraint. To sidestep this problem, we propose to divide the optimization part into two alternating stages: in the former stage, we optimize all variables with adaptive density control; while in the latter stage, we optimize the attributes without adding or removing points. The local rigid constraint is incorporated in every latter stage, and it is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.579, + 0.892, + 0.613 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s} = \\sum_ {j \\in \\mathcal {N} _ {i}} \\| D (t) _ {i} - D (t) _ {j} \\| _ {2} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.624, + 0.894, + 0.654 + ], + "angle": 0, + "content": "where, \\(\\mathcal{N}_i\\) represents the \\(K\\) nearest neighbor (KNN) of \\(i\\)-th Gaussian." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.67, + 0.634, + 0.687 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.696, + 0.719, + 0.712 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.902 + ], + "angle": 0, + "content": "We train our model using the Adam [15] optimizer with separate learning rates for different attributes of the Gaussian point. We set a learning rate of \\(4 \\times 10^{-4}\\) for the point position with an exponential decay of \\(8 \\times 10^{-7}\\). The learning rates for point rotation and all DDDM parameters are set to 0.002 and \\(4 \\times 10^{-4}\\) respectively. We apply a weight decay of \\(8 \\times 10^{-7}\\) to all parameters. The rest of the learning rate follows the 3DGS setting. We train the model for 30K steps and 60K steps for all scenes. All experiments are conducted on a single NVIDIA RTX 4090 GPU with 24GB memory. In addition, we use Taichi to implement our DDDM model, which can parallelize the computation of the DDDM (i.e." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21140" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "polynomial and Fourier series computation) of each Gaussian point." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.133, + 0.181, + 0.147 + ], + "angle": 0, + "content": "4.2. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.157, + 0.47, + 0.203 + ], + "angle": 0, + "content": "We evaluate our method on both multi-view and monocular datasets, to demonstrate the effectiveness of our method in both settings." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.224, + 0.471, + 0.33 + ], + "angle": 0, + "content": "Plenoptic Video dataset [18] The dataset was captured using 21 cameras at a resolution of \\(2704 \\times 2028\\), with each camera recording a 10-second video. Six scenes from this dataset are publicly available. For a fair comparison, we downsampled the images to \\(1352 \\times 1014\\) resolution in our experiments to keep the same setting from the concurrent work of 4D Gaussian [35]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.352, + 0.471, + 0.518 + ], + "angle": 0, + "content": "HyperNeRF dataset [28] This dataset uses a monocular camera (e.g., iPhone) to record real-world motions, which includes real rigid and non-rigidly deforming scenes, such as a person splitting a cookie. The dataset is rather challenging due to large motions, complex lighting conditions, and thin object structures. To ensure a fair comparison, we downsampled images to \\(540 \\times 960\\) in our experiments and followed the training and validation camera split provided by [28]. We conducted experiments on the four \"vrig\" scenes, and we also provided results of the \"interp\" scenes in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.53, + 0.232, + 0.545 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.47, + 0.779 + ], + "angle": 0, + "content": "Deformation Models The deformation model is the core component of the proposed Gaussian-Flow. We conduct ablation studies to validate the effectiveness of the particular choice of DDDM. As DDDM consists of a Fourier series and a polynomial function, hence we first study the Fourier series and the polynomial functions separately, in which we only use the Fourier series or the polynomial function as the deformation model. As shown in Figure. 4, the Fourier series contains more high-frequency components than the polynomial function, thus the Fourier series has sharper image details but results in more artifacts. The polynomial function is smoother than the Fourier series, leading to fewer artifacts but burry scene renderings. Finally, the hybrid DDDM function is able to generate sharper details with fewer artifacts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Furthermore, we study the orders of the polynomial and Fourier series functions in our DDDM, which are related to the complexity of the scene and are crucial to the final performance. As shown in Figure. 5, the performance of our method increases with the number of Fourier series orders but starts to drop after the order over 32, which should be related to the over-parameterization of the deformation model." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.089, + 0.882, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.395, + 0.893, + 0.453 + ], + "angle": 0, + "content": "Figure 4. Ablation study on different deformation models. From left to right are deformation fitting with polynomial function only, Fourier series only, and our dual-domain deformation fitting. The proposed DDDM achieves the best rendering quality qualitatively." + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.469, + 0.89, + 0.628 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.64, + 0.893, + 0.682 + ], + "angle": 0, + "content": "Figure 5. Ablation study on different orders of the DDDM. We find an order number of 16 leads to the best novel view rendering results in the HyperNeRF dataset." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.712, + 0.892, + 0.817 + ], + "angle": 0, + "content": "Regularizations Next, we study the effectiveness of the two proposed regularizations in Gaussian-Flow optimization. As shown in Table. 1, adding separate KNN rigid regularization or the time smooth regularization can both improve the novel view rendering quality, and the full model that contains both regularizations can achieve the best performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.831, + 0.741, + 0.848 + ], + "angle": 0, + "content": "4.4. Quantitative Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.893, + 0.901 + ], + "angle": 0, + "content": "We compare our method with SOTA on D-NeRF [30] synthetic scenes. As presented in Table 2, our method is benchmarked against TiNeuVox, K-Planes, and 4D-GS." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21141" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.472, + 0.132 + ], + "angle": 0, + "content": "Table 1. Ablation study on the proposed KNN rigid and time smooth regularizations. The quantitative results demonstrate the effectiveness of both regularizations." + }, + { + "type": "table", + "bbox": [ + 0.085, + 0.143, + 0.462, + 0.202 + ], + "angle": 0, + "content": "
reg,w/o KNN rigidw/o Time smoothfull model
PSNR28.4829.1229.92
SSIM0.920.930.94
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.209, + 0.47, + 0.24 + ], + "angle": 0, + "content": "Our method consistently outperforms other methods across all metrics." + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.251, + 0.468, + 0.266 + ], + "angle": 0, + "content": "Table 2. Quantitative results on D-NeRF synthetic blender scenes." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.282, + 0.47, + 0.376 + ], + "angle": 0, + "content": "
MethodPSNR ↑SSIM ↑LPIPS ↓
TiNeuVox [7]32.670.970.04
K-Planes [8]31.610.97/
4D-GS [35]33.300.980.03
Ours34.270.980.03
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.387, + 0.47, + 0.598 + ], + "angle": 0, + "content": "We also compare our method against previous SOTA NeRF-based methods, including NeRF [25], Nerfies [27], HyperNeRF [28], NeRFPlayer [32], and TiNeuVox [7] on HyperNeRF [28]. We also provide comparisons with other 3DGS-based approaches concurrently proposed with our Gaussian-Flow. The training time, rendering FPS, and novel view synthesis PSNR of different methods can be found in Table. 3. Previous NeRF-based methods require at least 30 minutes to train the scene, and fail to achieve real-time rendering of the dynamic scene. Our method only requires 7 minutes of training time and can achieve real-time rendering speed, which is much faster than previous methods. Moreover, our method can achieve better performance than previous SOTA methods in terms of PSNR." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.47, + 0.84 + ], + "angle": 0, + "content": "We evaluated various methods on the Plenoptic Video dataset, as summarized in Table. 4. The comparison focuses on training time efficiency and image quality, assessed through PSNR and SSIM. Our approach demonstrated a significant advancement in training efficiency, requiring only 22.5 minutes, a drastic reduction compared to the hours needed by methods like DyNeRF [18] and K-Planes [8]. This efficiency is paramount for practical applications, where reduced training time can be a critical factor. In terms of image quality, our method achieved a PSNR of 30.5 with 30K steps, which, while not the highest, is competitive with the leading methods. However, our method scored 0.97 in SSIM, higher than K-Planes' leading score of 0.96. This indicates a potential trade-off between training efficiency and the ability to preserve structural details in images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "In addition, we extend our method to 60K steps on both datasets, which can further improve the performance on the HyperNeRF dataset [28], and achieve the highest performance on the Plenoptic Video dataset [18]. However, the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "training time is also increased by approximately \\(2 \\times\\) for HyperNeRF dataset [28] and about \\(1.5 \\times\\) for Plenoptic Video dataset [18], which is still much faster than previous methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.159, + 0.729, + 0.175 + ], + "angle": 0, + "content": "4.5. Qualitative Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.182, + 0.892, + 0.364 + ], + "angle": 0, + "content": "In this section, we show qualitative comparisons of our method and previous SOTA methods on the HyperNeRF [28] dataset and the Plenoptic Video dataset [18]. Figure. 6 shows the qualitative comparisons of our method and TiNeuVox [7], 4D-GS [35] on HyperNeRF dataset. Notice that our method can produce comparably clear and sharp images than previous SOTA methods, which highlights our method's superior performance in monocular conditions. Despite its overall efficacy, our method does encounter limitations in extremely thin structure, as shown in the 3D Printer scene, the thread of the 3D Printer is not clear, while other methods can produce a clear thread." + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.375, + 0.872, + 0.585 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.596, + 0.892, + 0.625 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparisons of our method and TiNeVox [7] and 4D-GS [35] on the HyperNeRF [28] vrig. dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.638, + 0.892, + 0.789 + ], + "angle": 0, + "content": "We also compare our method with previous SOTA methods on the Plenoptic Video dataset [18], as shown in Figure. 7. Compared with previous SOTA methods, our method can produce more accurate color and correct structure. Moreover, our method successfully reconstructs the flame in the scene, while NeRFPlayer [8] fails to reconstruct the flame. These results suggest that our method can achieve comparable image quality with previous SOTA methods, which demonstrates the effectiveness of our method in the multiview conditions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.801, + 0.619, + 0.815 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this paper, we introduced Gaussian-Flow, a novel framework for dynamic 3D scene reconstruction using a point-based differentiable rendering approach. The core of our innovation lies in the DDDM, which efficiently models deformations of each 3D Gaussian point in both the time and" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21142" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.892, + 0.132 + ], + "angle": 0, + "content": "Table 3. Per-scene quantitative comparisons on HyperNeRF [28] dataset. Results are gathered from papers of the corresponding methods. Our method achieves the fastest training time, the highest rendering FPS, and the highest PSNR score for novel view synthesis, setting a new state-of-the-art for image-based dynamic scene rendering." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.148, + 0.885, + 0.27 + ], + "angle": 0, + "content": "
MethodTrain Time↓FPS↑Broom3D PrinterChickenPeel BananaMean
PSNR↑SSIM ↑PSNR↑SSIM ↑PSNR↑SSIM ↑PSNR↑SSIM ↑PSNR↑SSIM ↑
NeRF [25]16 hours0.01319.90.65320.70.78019.90.77720.00.76920.10.745
Nerfies [27]16 hours0.01119.20.56720.60.83026.70.94322.40.87222.20.803
HyperNeRF [28]32 hours0.01119.30.59120.00.82126.90.94823.30.89622.40.814
NeRFPlayer [32]6 hours0.20821.70.63522.90.81026.30.90524.00.86323.70.803
TiNeuVox [7]30 min0.521.50.68622.80.84128.30.94724.40.87324.30.837
Ours (30K)7 min12522.50.69024.30.85729.40.93426.30.90625.60.847
Ours (60K)12 min12522.80.70925.00.87730.40.94527.00.91726.30.862
" + }, + { + "type": "image", + "bbox": [ + 0.177, + 0.283, + 0.382, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.269, + 0.404, + 0.289, + 0.415 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.283, + 0.588, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.454, + 0.404, + 0.517, + 0.415 + ], + "angle": 0, + "content": "Ours (30K)" + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.283, + 0.794, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.667, + 0.404, + 0.716, + 0.415 + ], + "angle": 0, + "content": "K-Planes" + }, + { + "type": "image", + "bbox": [ + 0.177, + 0.417, + 0.382, + 0.536 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.245, + 0.538, + 0.314, + 0.551 + ], + "angle": 0, + "content": "NeRFPlayer" + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.417, + 0.588, + 0.536 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.46, + 0.538, + 0.511, + 0.55 + ], + "angle": 0, + "content": "DyNeRF" + }, + { + "type": "image", + "bbox": [ + 0.59, + 0.417, + 0.794, + 0.536 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.675, + 0.538, + 0.708, + 0.55 + ], + "angle": 0, + "content": "LLFF" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.562, + 0.89, + 0.578 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparisons of our method and K-Planes [8], NeRFPlayer [8], DyNeRF [18], and LLFF on Plenoptic Video dataset." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.602, + 0.47, + 0.658 + ], + "angle": 0, + "content": "Table 4. Quantitative comparison on Plenoptic Video dataset [18]. Our training speed is \\(5 \\times\\) faster of magnitude faster than previous leading approaches. Also, we achieved the highest PSNR score among all methods." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.674, + 0.47, + 0.809 + ], + "angle": 0, + "content": "
MethodTrain Time ↓PSNR ↑SSIM ↑
LLFF [24]-23.2-
DyNeRF [18]1344 hours29.60.96
NeRFPlayer [32]5.5 hours30.7-
K-Planes [8]1.8 hours31.60.96
4D-GS [35]2 hours31.00.94
Ours (30K)22.5 min30.50.97
Ours (60K)41.8 min32.00.97
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.903 + ], + "angle": 0, + "content": "frequency domains. This approach has enabled us to set a new state-of-the-art for 4D scene reconstruction in terms of training speed, rendering frames per second, and novel view synthesis quality. Our extensive experiments and ab" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.603, + 0.892, + 0.709 + ], + "angle": 0, + "content": "lation studies have demonstrated the efficacy of proposed Gaussian-Flow across various datasets. We achieved significant improvements over existing methods, particularly in training speed and rendering performance. The ability to efficiently handle dynamic scenes without the computational overhead of neural networks marks a substantial leap forward in this domain." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.724, + 0.623, + 0.739 + ], + "angle": 0, + "content": "6. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.748, + 0.892, + 0.901 + ], + "angle": 0, + "content": "While our method excels in rendering speed and training efficiency, there is room for improvement in maintaining high-fidelity thin structures in the final rendering, as mentioned in the Section. 4.5. The current deformation model may not be able to capture the fine details of thin structures, leading to some loss of image quality. This limitation is particularly evident in scenes with complex motion patterns. Future work could focus on enhancing the balance between speed and image detail preservation, potentially through more refined deformation models or advanced regularization techniques." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21143" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.121, + 0.174, + 0.137 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.146, + 0.47, + 0.215 + ], + "angle": 0, + "content": "[1] Benjamin Attal, Jia-Bin Huang, Christian Richardt, Michael Zollhoefer, Johannes Kopf, Matthew O'Toole, and Changil Kim. HyperReel: High-fidelity 6-DoF video with ray-conditioned sampling. arXiv preprint arXiv:2301.02238, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.218, + 0.47, + 0.245 + ], + "angle": 0, + "content": "[2] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. arXiv:2301.09632, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.247, + 0.47, + 0.288 + ], + "angle": 0, + "content": "[3] Ang Cao and Justin Johnson. Hexplane: a fast representation for dynamic scenes. arXiv preprint arXiv:2301.09632, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.29, + 0.471, + 0.332 + ], + "angle": 0, + "content": "[4] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In Proceedings of the European Conference on Computer Vision, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.334, + 0.47, + 0.403 + ], + "angle": 0, + "content": "[5] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 14304-14314. IEEE Computer Society, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.405, + 0.47, + 0.445 + ], + "angle": 0, + "content": "[6] Jiemin Fang, Lingxi Xie, Xinggang Wang, Xiaopeng Zhang, Wenyu Liu, and Qi Tian. Neusample: Neural sample field for efficient view synthesis. arXiv:2111.15552, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.447, + 0.469, + 0.502 + ], + "angle": 0, + "content": "[7] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. arXiv preprint arXiv:2205.15285, 2022. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.504, + 0.47, + 0.559 + ], + "angle": 0, + "content": "[8] Sara Fridovich-Keil, Giacomo Meanti, Frederik Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. arXiv preprint arXiv:2301.10241, 2023. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.561, + 0.469, + 0.615 + ], + "angle": 0, + "content": "[9] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.618, + 0.469, + 0.673 + ], + "angle": 0, + "content": "[10] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5712-5721, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.469, + 0.743 + ], + "angle": 0, + "content": "[11] Stephan J. Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14346-14355, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[12] Peter Hedman, Pratul P. Srinivasan, Ben Mildenhall, Jonathan T. Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5875-5884, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[13] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 42(4):1-14, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[14] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.146, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.164 + ], + "angle": 0, + "content": "[15] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.165, + 0.892, + 0.22 + ], + "angle": 0, + "content": "[16] Andreas Kurz, Thomas Neff, Zhaoyang Lv, Michael Zollhofer, and Markus Steinberger. Adanerf: Adaptive sampling for real-time rendering of neural radiance fields. In European Conference on Computer Vision, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.222, + 0.892, + 0.25 + ], + "angle": 0, + "content": "[17] Marc Levoy and Turner Whitted. The use of points as a display primitive. 1985. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.252, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[18] Tianye Li, Mira Slavcheva, Michael Zollhöfer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, and Zhaoyang Lv. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5521-5531, 2022. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.892, + 0.405 + ], + "angle": 0, + "content": "[19] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.407, + 0.892, + 0.461 + ], + "angle": 0, + "content": "[20] D. B.* Lindell, J. N. P.* Martel, and G. Wetzstein. Autoint: Automatic integration for fast neural volume rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.463, + 0.892, + 0.505 + ], + "angle": 0, + "content": "[21] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. In Advances in Neural Information Processing Systems, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.507, + 0.892, + 0.561 + ], + "angle": 0, + "content": "[22] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural volumes: learning dynamic renderable volumes from images. ACM Transactions on Graphics, 38(4):1-14, 2019. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.563, + 0.892, + 0.618 + ], + "angle": 0, + "content": "[23] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.62, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[24] Ben Mildenhall, Pratul P. Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (Proceedings of SIGGRAPH), 38(4), 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.704, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[25] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision, pages 405-421. Springer, 2020. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.775, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[26] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[27] Keunhong Park, Utkarsh Sinha, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Steven M. Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 2, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21144" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.174 + ], + "angle": 0, + "content": "[28] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo Martin-Brulla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Trans. Graph., 40(6), 2021. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.178, + 0.469, + 0.219 + ], + "angle": 0, + "content": "[29] Martin Piala and Ronald Clark. Terminerf: Ray termination prediction for efficient neural rendering. International Conference on 3D Vision, pages 1106-1114, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.221, + 0.469, + 0.289 + ], + "angle": 0, + "content": "[30] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.291, + 0.469, + 0.358 + ], + "angle": 0, + "content": "[31] Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boning Liu, Hongwen Zhang, and Yebin Liu. Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. arXiv preprint arXiv:2211.11610, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.362, + 0.469, + 0.431 + ], + "angle": 0, + "content": "[32] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerfplayer: A streamable dynamic scene representation with decomposed neural radiance fields. arXiv preprint arXiv:2210.15947, 2022. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.433, + 0.469, + 0.5 + ], + "angle": 0, + "content": "[33] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.504, + 0.469, + 0.545 + ], + "angle": 0, + "content": "[34] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Improved direct voxel grid optimization for radiance fields reconstruction. arXiv preprint arXiv:2206.05085, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.548, + 0.469, + 0.602 + ], + "angle": 0, + "content": "[35] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528, 2023. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.604, + 0.469, + 0.685 + ], + "angle": 0, + "content": "[36] Liwen Wu, Jae Yong Lee, Anand Bhattachad, Yu-Xiong Wang, and David Forsyth. Diver: Real-time and accurate neural radiance fields with deterministic integration for volume rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16200-16209, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.688, + 0.469, + 0.757 + ], + "angle": 0, + "content": "[37] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9421-9431, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.76, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[38] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.817, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[39] Wang Yifan, Felice Serena, Shihao Wu, Cengiz Öz Tireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM Transactions on Graphics (TOG), 38(6):1-14, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.874, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[40] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5752-5761, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[41] Qiang Zhang, Seung-Hwan Baek, Szymon Rusinkiewicz, and Felix Heide. Differentiable point-based radiance fields for efficient view synthesis. In SIGGRAPH Asia 2022 Conference Papers, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[42] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Ewa volume splatting. In Proceedings Visualization, 2001. VIS'01., pages 29-538. IEEE, 2001. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21145" + } + ] +] \ No newline at end of file diff --git a/2024/Gaussian-Flow_ 4D Reconstruction with Dynamic 3D Gaussian Particle/3e2469b7-4042-4a31-bb56-8ba156724067_origin.pdf b/2024/Gaussian-Flow_ 4D Reconstruction with Dynamic 3D Gaussian Particle/3e2469b7-4042-4a31-bb56-8ba156724067_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..82bf11d5ae590508ca2711b756c867eafbb7d364 --- /dev/null +++ b/2024/Gaussian-Flow_ 4D Reconstruction with Dynamic 3D Gaussian Particle/3e2469b7-4042-4a31-bb56-8ba156724067_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0d5de186eafc4bb042f0eff02bc90816438a32e541b41ebb37e083f3924e3705 +size 4972015 diff --git a/2024/Gaussian-Flow_ 4D Reconstruction with Dynamic 3D Gaussian Particle/full.md b/2024/Gaussian-Flow_ 4D Reconstruction with Dynamic 3D Gaussian Particle/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2f00c1094b2b3434c690cc434df4b05c41986e74 --- /dev/null +++ b/2024/Gaussian-Flow_ 4D Reconstruction with Dynamic 3D Gaussian Particle/full.md @@ -0,0 +1,330 @@ +# Gaussian-Flow: 4D Reconstruction with Dynamic 3D Gaussian Particle + +Youtian Lin1,2 + +Zuozhuo Dai3 + +Siyu Zhu + +Yao $\mathrm{Yao^{1,2\boxtimes}}$ + +$^{1}$ State Key Laboratory for Novel Software Technology, Nanjing University, Nanjing, China + $^{2}$ School of Intelligence Science and Technology, Nanjing University, Suzhou, China + $^{3}$ Alibaba Group + $^{4}$ Fudan University + +![](images/011177e7de7e581e87375234ad14b83ca5958e7b27a20ec981f875dcdaaf5e85.jpg) +Figure 1. Dynamic reconstruction results of the proposed Gaussian-Flow on the monocular D-NeRF [30] and HyperNeRF Dataset [28] (left) and the multi-view Plenoptic Dataset [18] (right). Our method achieves a $5 \times$ faster training and rendering speed compared with the per-frame 3DGS modeling and significantly outperforms previous methods in novel view rendering quality. + +![](images/8409baae1f126fa7fc775f1a146696fd00a48429075bdd1afce2509f1eda6930.jpg) + +![](images/c203f09eb7b8ed5c101a2eedb8f7c227f8bd12d421e8ad62a461fdb66ceb115f.jpg) + +![](images/cee1196eb8f70fdda808aa8d5768a28dcbd2522dbda0fcde4d193b64ac301275.jpg) + +# Abstract + +We introduce Gaussian-Flow, a novel point-based approach for fast dynamic scene reconstruction and real-time rendering from both multi-view and monocular videos. In contrast to the prevalent NeRF-based approaches hampered by slow training and rendering speeds, our approach harnesses recent advancements in point-based 3D Gaussian Splatting (3DGS). Specifically, a novel Dual-Domain Deformation Model (DDDM) is proposed to explicitly model attribute deformations of each Gaussian point, where the time-dependent residual of each attribute is captured by a polynomial fitting in the time domain, and a Fourier series fitting in the frequency domain. The proposed DDDM is capable of modeling complex scene deformations across long video footage, eliminating the need for training separate 3DGS for each frame or introducing an additional implicit neural field to model 3D dynamics. Moreover, the explicit deformation modeling for discretized Gaussian points en + +sures ultra-fast training and rendering of a 4D scene, which is comparable to the original 3DGS designed for static 3D reconstruction. Our proposed approach showcases a substantial efficiency improvement, achieving a $5 \times$ faster training speed compared to the per-frame 3DGS modeling. In addition, quantitative results demonstrate that the proposed Gaussian-Flow significantly outperforms previous leading methods in novel view rendering quality. Project page: https://nju-3dv.github.io/projects/Gaussian-Flow. + +# 1. Introduction + +In the realm of digital scene synthesis, achieving a balance between high-quality reconstructions and real-time rendering is paramount, especially for applications like virtual reality (VR) playback, where immediate feedback and immersive experiences are essential. Neural Radiance Fields + +(NeRFs) [25] has risen as a promising method for synthesizing intricate scenes. However, despite their ability to produce visually stunning results, NeRFs require costly sampling and evaluation of the neural radiance field at multiple points along each ray. Consequently, the substantial computational demands impede the real-time rendering capabilities of NeRFs. These are attempts to accelerate the rendering process of NeRFs, such as direct volume representation [21, 33, 40], neural hasing [26], and tri-plane structures [2, 4, 9]. However, it still remains a challenge for high-fidelity real-time rendering. Nevertheless, the issue becomes even more severe when turning to dynamic scene reconstruction and rendering. + +Recent progress on 3D Gaussian Splatting (3DGS) [14] has drawn attention from the 3D computer vision community. With tile-based rasterization instead of plain volume rendering, 3DGS can render images two orders of magnitude faster than the vanilla NeRF. The technique has also been quickly applied to 4D scene reconstruction by extending to the separate per-frame 3DGS optimization [23]. However, such direct extension is storage-intensive and is not applicable to monocular video input. Some other concurrent works [35, 38] try to mix the explicit point-based 3DGS and an implicit neural field for dynamic information modeling, however, require computationally expensive forward passes of the neural network, which significantly lowers the rendering speed of the original 3DGS. + +In this work, we propose Gaussian-Flow, an explicit particle-based deformation model designed specifically for 3DGS to model the dynamic scene without using any neural network. Gaussian-Flow can recover a high-fidelity 4D scene from captured videos, while still preserving the ultrafast training and rendering speed of the original 3DGS. In particular, we formulate a 4D scene as a set of deformable 3D Gaussian points. A novel Dual-Domain Deformation Model (DDDM) is proposed to explicitly model deformations of each Gaussian point's attributes, including position, rotation, and radiance. The time-dependent deformation residual is modeled simultaneously in time and frequency domains: we apply joint polynomial and Fourier series fitting for each deformable attribute. This compact dynamic representation greatly reduces the computation cost of the deformation model, which is a key factor in preserving the rendering speed of 3DGS. Moreover, an adaptive times-tamp scaling technique is introduced to avoid over-fitting the scene to only frames with violent motions. For robust estimation, we also regularize the motion trajectory by the KNN-based rigid and the time smooth constraints. It is also noteworthy that our discretized point-based 4D representation naturally supports the edition of both static and dynamic 3D scenes, showing the potential for unlocking a variety of downstream applications related to dynamic 3D reconstruction and rendering. + +We have conducted extensive experiments to demonstrate the effectiveness of the proposed method on several multi-view and monocular datasets. The proposed Gaussian-Flow achieves a $5 \times$ faster training speed compared with the separate per-frame 3DGS modeling, and significantly outperforms prior leading methods in novel view rendering quality. Our major contributions can be summarized as follows: + +- We introduce Gaussian-Flow, a novel point-based differentiable rendering approach for dynamic 3D scene reconstruction, setting a new state-of-the-art for training speed, rendering FPS, and novel view synthesis quality for 4D scene reconstruction. +- We propose a Dual-Domain Deformation Model for efficient 4D scene training and rendering, eliminating the need for per-frame 3DGS optimization and sampling on implicit neural fields. This preserves a running speed on par with the original 3DGS with minimum overhead. +- We demonstrate that our discretized point-based representation supports the segmentation, edition, and composition of both static and dynamic 3D scenes. + +# 2. Related Works + +# 2.1. Dynamic Neural Radiance Field + +Dynamic NeRF modeling has become a heated research topic in recent years due to the development of the neural radiance field and differentiable rendering. By treating time as an extended input dimension to NeRF, researchers successfully achieve qualified image-based 4D scene rendering [5, 10, 19, 22, 37]. To further improve the reconstruction quality and incorporate prior knowledge of motions and structures, dynamic neural scene flow methods have been proposed [27, 30], where a canonical space is constructed and then transferred to each frame with scene flow or motion fields. HyperNeRF [28] models the deformation of the object topologies by using higher-dimensional inputs, while DyNeRF [18] utilizes time-conditioned NeRF to represent a 4D scene. However, the aforementioned approaches are all based on the vanilla NeRF, which requires a long training time and does not meet the requirement of real-time rendering. + +# 2.2. Accelerated Neural Radiance Field + +To expedite NeRF training and rendering, numerous approaches have been suggested, employing more streamlined strategies [6, 11, 12, 16, 20, 29, 36]. Other methods propose the integration of neural implicit functions with explicit 3D structures, forming a hybrid representation for faster radiance field sampling [4, 26, 33, 34]. These approaches establish strong foundations for enhancing dynamic NeRF, concurrently decreasing required training and inference time. Apart from implicit neural representations, explicit NeRF + +modeling has show promising results for real-time rendering: NSVF [21] employs a neural sparse voxel field for efficient NeRF sampling, which stands for the earliest attempt on the explicit NeRF modeling; PlenOctrees [40] utilizes the explicit octree structure for rendering acceleration. + +Recent efforts have also emerged to accelerate the intricate dynamic neural radiance field. TensorRF [4] employs multiple planes as explicit representations for direct dynamic scene modeling. More recent recent approaches of this kind include K-Planes [8], Tensor4D [31], and HexPlane [3]. Alternatively, NeRFPlayer[32] introduces a unified streaming representation for both grid-based [26] and plane-based methods, utilizing separate models to distinguish static and dynamic scene components, however, leading to slow rendering times. HyperReel [1] further suggests a flexible sampling network coupled with two planes for dynamic scene representation. While these methods improve the rendering speed of a dynamic scene to some extent, it is still hard to achieve real-time rendering, let alone a good balance between the running speed and rendering quality. In contrast, we resort to the recent 3D Gaussians splatting, which applies an explicit soft point cloud representation for real-time image-based rendering. + +# 2.3. Differentiable Point-based Rendering + +The original idea of using 3D points as rendering primitives was first introduced in [17]. By incorporating differentiable rendering, recent approaches have made remarkable progress in image-based rendering, representative methods include PointRF [41], DSS [39], and 3D Gaussians splatting (3DGS) [13]. Specifically, 3DGS [13] has demonstrated extraordinary performance in novel-view synthesis, achieving real-time rendering speed and state-of-the-art rendering quality. The method adopts a soft point representation with attributes of position, rotation, density, and radiance, and applies differentiable point-based rendering for scene optimization. 3DGS has quickly been extended to dynamic scene modeling by direct separate per-frame optimization [23], however, requires a long optimization time and a large amount of storage for long video footage. Other works [35, 38] apply an implicit motion field to model scene dynamics, but the introduction of the implicit neural network significantly slows down the sampling and rendering speed. In this work, we adopt the approach of representing 4D scenes through a purely discretized point cloud model, ensuring a fast training and rendering speed comparable with the original 3DGS. + +# 3. Gaussian-Flow + +In this section, we introduce the proposed Gaussian-Flow for dynamic scene modeling. We first review the 3DGS in Sec. 3.1. Then, we introduce our explicit motion modeling of each Gaussian point by using a novel Dual-Domain + +Deformation Model (DDDM), as outlined in Sec. 3.2. An adaptive timestamp scaling technique is described in Sec. 3.3 for balanced training of each frame. To ensure the continuity of the motion in both spatial and temporal dimensions, we incorporate appropriate regularizations on each point during the optimization, as detailed in Sec. 3.4. + +# 3.1. Recap on 3D Gaussian Splitting + +3D Gaussian Splatting [14] is designed to efficiently optimize a 3D scene for real-time and high-quality novel view synthesis. The 3DGS framework has garnered significant attention within the community due to its remarkable enhancements in both training and rendering times, concurrently achieving state-of-the-art rendering quality. In contrast to the volume rendering in the vanilla NeRF which relies on ray marching, 3DGS adopts a tile-based rasterization on a distinctive soft point cloud representation to achieve fast rendering. Specifically, 3DGS models a 3D scene as a large amount of 3D Gaussian points in the world space, where each point is represented by: + +$$ +G (\boldsymbol {x}) = \exp \left(- \frac {1}{2} \left(\boldsymbol {x} - \boldsymbol {\mu}\right) ^ {T} \boldsymbol {\Sigma} ^ {- 1} \left(\boldsymbol {x} - \boldsymbol {\mu}\right)\right), \tag {1} +$$ + +where $\mu$ and $\Sigma$ are the mean position and covariance matrix of a 3D Gaussian particle. The 3DGS takes sparse Structure-from-Motion (SfM) points or even random points as input, and initializes each point to a 3D Gaussian based on its neighbors. Besides, each 3D Gaussian is associated with a learnable view-dependent radiance $c$ and a learnable opacity $\alpha$ for rendering. Subsequently, an efficient 3D to 2D Gaussian mapping [42] is employed to project the point onto the image plane: + +$$ +\boldsymbol {\mu} ^ {\prime} = \boldsymbol {P} \boldsymbol {W} \boldsymbol {\mu}, \tag {2} +$$ + +$$ +\boldsymbol {\Sigma} ^ {\prime} = \boldsymbol {J} \boldsymbol {W} \boldsymbol {\Sigma} \boldsymbol {W} ^ {T} \boldsymbol {J} ^ {T}, \tag {3} +$$ + +where $\mu^{\prime}$ and $\Sigma^{\prime}$ represent the 2D mean position and 2D covariance of the projected 3D Gaussian. $P, W$ and $J$ donate the projective transformation, viewing transformation, and Jacobian of the affine approximation of $P$ . After that, $\alpha$ -blending is executed to merge the overlapping Gaussians for each pixel, yielding a final color: + +$$ +C = \sum_ {i = 1} ^ {n} \boldsymbol {c} _ {i} \alpha_ {i} \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {i}), \tag {4} +$$ + +where $c_{i}$ and $\alpha_{i}$ are the color and opacity of the $i$ -th Gaussian, and $n$ is the number of overlapping Gaussians. + +The attributes of the 3D Gaussian, including the mean $\mu$ , covariance matrix $\Sigma$ , opacity $\alpha$ , and color $c$ , are optimized through backward propagation of the gradient flow. In particular, to ensure positive definiteness during the optimization process, the covariance matrix is parameterized + +as a scaling vector $s$ and a rotation matrix $R$ , i.e., $\Sigma = R\Lambda(s)\Lambda(s)^T R^T$ , where $\Lambda(s)$ is the diagonal matrix of $s$ . To facilitate optimization, the rotation matrix is further parameterized using a quaternion $q$ . Leveraging the inherent flexibility of discrete points, 3DGS incorporates adaptive density control for points. This mechanism utilizes the gradient flow to identify where geometric reconstruction is suboptimal, and employs cloning and splitting to augment the density of points for a higher rendering quality. + +# 3.2. Dual-Domain Deformation Model + +We target to directly model the dynamics of each 3D Gaussian point by fitting each of its attributes into a time-dependent curve. Among different approaches, Polynomials fitting in the time domain and Fourier series fitting in the frequency domain are the two most widely used approaches [5, 10, 19, 22, 37], due to their simplicity and effectiveness. However, each method comes with its own advantages and drawbacks: describing the motion of a Gaussian particle in terms of polynomials yields a good fit with smooth motion with a small order of polynomials, however, can easily overfit to a violent motion if assuming a larger order of polynomials, resulting in unreasonable oscillations in the fitted trajectory. Whereas, the Fourier series excels at capturing the variations associated with violent motion, however, requires a manually reduced order when dealing with smooth motion. + +In this work, our key insight is to use a Dual-Domain Deformation Model (DDDM) for fitting the scene dynamics, which integrates both the time domain polynomials and the frequency domain Fourier series into a unified fitting model. We assume that only the rotation $\pmb{q}$ , radiance $c$ , and position $\pmb{\mu}$ of a 3D Gaussian particle change over time, while the scaling $s$ and opacity $\alpha$ remain constant. Specifically, we conceptualize the change in each particle's attributes as its base attributes $S_0 \in \{\pmb{\mu}_0, c_0, q_0\}$ at the reference time frame $t_0$ (usually set to the first frame), superimposed on a time-dependent attribute residual $D(t)$ . For simplicity, we use lowercase characters to represent a single attribute in $S$ . The time-dependent residual of each attribute is modeled through polynomial fitting in the time domain and Fourier series fitting in the frequency domain, expressed as: + +$$ +S (t) = S _ {0} + D (t), \tag {5} +$$ + +where $D(t) = P_N(t) + F_L(t)$ is combined by a polynomial $P_N(t)$ with coefficients $\pmb{a} = \{a\}_{n=0}^{N}$ and a Fourier series $F_L(t)$ with coefficients $\pmb{f} = \{f_{sin}^l, f_{cos}^l\}_{l=0}^L$ . These are respectively denoted as: + +$$ +P _ {N} (t) = \sum_ {n = 0} ^ {N} a _ {n} t ^ {n}, \tag {6} +$$ + +$$ +F _ {L} (t) = \sum_ {l = 1} ^ {L} \left(f _ {\sin} ^ {l} \cos (l t) + f _ {\cos} ^ {l} \sin (l t)\right). \tag {7} +$$ + +It is important to note that we assume different dimensions of an attribute are independently changed over time. Therefore, we assign a different $D(t)$ for each dimension of an attribute. For instance, we utilize $\{D_{\mu_i}(t)\}_{i=0}^3$ to describe the motion of a 3D position $\pmb{\mu}$ . + +Figure 3 illustrates a comparative analysis of trajectory fitting using polynomial, Fourier series, and the proposed joint DDDM functions. The figure highlights the superior fitting capabilities of the DDDM approach in capturing complex motion trajectories as represented by the sampled data points. + +# 3.3. Adaptive Timestemp Scaling + +In a typical scenario, a normalized frame index $t$ ranging from 0 to 1 will be used as the temporal input of $D(t)$ . However, this poses a challenge when endeavoring to model substantial motions within a very short time using polynomials and Fourier series. Adhering to the standard temporal division would necessitate an exceedingly large coefficient to accommodate highly intense movements within a very short time frame. This circumstance has the potential to induce instability or even a breakdown in the optimization process. To address this issue, we introduce a time dilation factor $\lambda$ to scale the temporal input for each Gaussian point, which is formulated as: + +$$ +t _ {s} = \lambda_ {s} \cdot t + \lambda_ {b} \tag {8} +$$ + +where $t_s$ represents the scaled time input, serving as the input of $D(t)$ , $t \in [0,1]$ denotes the normalized frame index, and $\lambda$ and $\lambda_b$ stand for the dilation factor and base factor of a Gaussian, respectively. In all our experiments, $\lambda_s$ and $\lambda_b$ are initialized to 1 and 0, respectively. + +To summarize, in our dynamic scene setting, a Gaussian particle contains multiple attributes to be optimized, including base attributes $\{\mu_0,q_0,s_0,c_0,\alpha_0\}$ at reference frame $t_0$ , polynomial coefficients and Fourier coefficients in $\{D_{\mu}(t),D_q(t),D_c(t)\}$ . Since the proposed DDDM optimizes the time-dependent residuals without the need for an intricate neural field structure, our Gaussian-Flow inherits the benefits of extreme-fast training and rendering speed from the vanilla 3DGS. + +# 3.4. Regularizations + +While the utilization of discrete points as scene representations accelerates rendering, several challenges remain. First, points are optimized individually, losing connections with their spatial neighbors, which do not align with the real-world scenario. Optimizing these Gaussian points without considering continuity will inevitably lead to a + +![](images/0048a63e88b776549374029bf25e4d539c550b3f8d35cd19bab3b5eae4b7e4dd.jpg) +Adaptive Timestamp Scaling + +![](images/1f113e0850392e0cc881978378908c8e851b6024c72e500a1dfa9cce415dab62.jpg) + +![](images/507b87fe0e2e244937d05cd068f4ffc7e90ef1ac0a50cc6db51452aadebf093b.jpg) +Radiance + +![](images/af0a463f6e073f6ef74f5ab11dc8d0514d151ac159c84ae9c0d86c84c47b5c64.jpg) +Forward + +![](images/413f4d271d9b9c08183d9d2c203e05f07a2cb0011adc7c627d3af178101cb14b.jpg) +Backward + +![](images/5736ef64b98c15315404595634f02571a3b576feeae6e06e7711771c647a60d7.jpg) +Dual-Domain Deformation Model + +![](images/67dee80a23b0eb50805662b4088b8f741b9a27a8df8f1a32c08642bc80092308.jpg) +Figure 2. Overview of the Gaussian-Flow pipeline. We model the deformation of attributes of each 3D Gaussian point independently by using the Dual-Domain Deformation Model (DDDM), which preserves the discretized nature of the 3D Gaussian points, and thus achieves ultra-fast training and rendering speed comparable with the original 3DGS. + +![](images/943b01e627be6bb69f0a616bfcc9579a1416ad2ecaee302a292f5c25ab2d855b.jpg) +Figure 3. Two exemplar motion fittings using polynomial, Fourier Series, and the joint DDDM functions. Our DDDM is able to accurately fit complex trajectories denoted by the sampling points. + +degradation in reconstruction quality and spatial coherence. Additionally, motions should be smoothed over time. Based on these observations, we employ two regularizations, namely time smoothness and a KNN rigid regularization, for robust optimization of Gaussian points and their motions. + +Time Smoothness Loss To ensure temporal smoothness over time, we apply a perturbing $\epsilon$ on the input timestamp $t$ and encourage the time-dependent attributes (i.e., position $\mu$ , rotation $q$ and radiance $c$ ) at time $t + \epsilon$ to be consistent with those at time $t$ . The time smoothness term is defined as: + +$$ +\mathcal {L} _ {t} = \| D (t) - D (t + \epsilon) \| _ {2}. \tag {9} +$$ + +It is noteworthy that the magnitude of the perturbing value is adaptively set according to the number of total frames, i.e., $\epsilon = 0.1/$ frames. + +KNN Rigid Loss During the optimization of 3D Gaussians with adaptive density control, points will be dynamically added or removed. This dynamic nature implies that the neighbors of points within a local space are subject to constant changes, posing a challenge for directly enforcing a spatial local consistency constraint. To sidestep this problem, we propose to divide the optimization part into two alternating stages: in the former stage, we optimize all variables with adaptive density control; while in the latter stage, we optimize the attributes without adding or removing points. The local rigid constraint is incorporated in every latter stage, and it is defined as: + +$$ +\mathcal {L} _ {s} = \sum_ {j \in \mathcal {N} _ {i}} \| D (t) _ {i} - D (t) _ {j} \| _ {2} \tag {10} +$$ + +where, $\mathcal{N}_i$ represents the $K$ nearest neighbor (KNN) of $i$ -th Gaussian. + +# 4. Experiments + +# 4.1. Implementation Details + +We train our model using the Adam [15] optimizer with separate learning rates for different attributes of the Gaussian point. We set a learning rate of $4 \times 10^{-4}$ for the point position with an exponential decay of $8 \times 10^{-7}$ . The learning rates for point rotation and all DDDM parameters are set to 0.002 and $4 \times 10^{-4}$ respectively. We apply a weight decay of $8 \times 10^{-7}$ to all parameters. The rest of the learning rate follows the 3DGS setting. We train the model for 30K steps and 60K steps for all scenes. All experiments are conducted on a single NVIDIA RTX 4090 GPU with 24GB memory. In addition, we use Taichi to implement our DDDM model, which can parallelize the computation of the DDDM (i.e. + +polynomial and Fourier series computation) of each Gaussian point. + +# 4.2. Datasets + +We evaluate our method on both multi-view and monocular datasets, to demonstrate the effectiveness of our method in both settings. + +Plenoptic Video dataset [18] The dataset was captured using 21 cameras at a resolution of $2704 \times 2028$ , with each camera recording a 10-second video. Six scenes from this dataset are publicly available. For a fair comparison, we downsampled the images to $1352 \times 1014$ resolution in our experiments to keep the same setting from the concurrent work of 4D Gaussian [35]. + +HyperNeRF dataset [28] This dataset uses a monocular camera (e.g., iPhone) to record real-world motions, which includes real rigid and non-rigidly deforming scenes, such as a person splitting a cookie. The dataset is rather challenging due to large motions, complex lighting conditions, and thin object structures. To ensure a fair comparison, we downsampled images to $540 \times 960$ in our experiments and followed the training and validation camera split provided by [28]. We conducted experiments on the four "vrig" scenes, and we also provided results of the "interp" scenes in the supplementary material. + +# 4.3. Ablation Study + +Deformation Models The deformation model is the core component of the proposed Gaussian-Flow. We conduct ablation studies to validate the effectiveness of the particular choice of DDDM. As DDDM consists of a Fourier series and a polynomial function, hence we first study the Fourier series and the polynomial functions separately, in which we only use the Fourier series or the polynomial function as the deformation model. As shown in Figure. 4, the Fourier series contains more high-frequency components than the polynomial function, thus the Fourier series has sharper image details but results in more artifacts. The polynomial function is smoother than the Fourier series, leading to fewer artifacts but burry scene renderings. Finally, the hybrid DDDM function is able to generate sharper details with fewer artifacts. + +Furthermore, we study the orders of the polynomial and Fourier series functions in our DDDM, which are related to the complexity of the scene and are crucial to the final performance. As shown in Figure. 5, the performance of our method increases with the number of Fourier series orders but starts to drop after the order over 32, which should be related to the over-parameterization of the deformation model. + +![](images/3bd3706bbd2d1872071390dc232a03b4643e2ff896e4d6c41132b5b9c09b8a62.jpg) +Figure 4. Ablation study on different deformation models. From left to right are deformation fitting with polynomial function only, Fourier series only, and our dual-domain deformation fitting. The proposed DDDM achieves the best rendering quality qualitatively. + +![](images/8818a3012ee2e26b2c1c54429a121634cf9494ac293c2df8b6b1f0b84e4e1706.jpg) +Figure 5. Ablation study on different orders of the DDDM. We find an order number of 16 leads to the best novel view rendering results in the HyperNeRF dataset. + +Regularizations Next, we study the effectiveness of the two proposed regularizations in Gaussian-Flow optimization. As shown in Table. 1, adding separate KNN rigid regularization or the time smooth regularization can both improve the novel view rendering quality, and the full model that contains both regularizations can achieve the best performance. + +# 4.4. Quantitative Comparisons + +We compare our method with SOTA on D-NeRF [30] synthetic scenes. As presented in Table 2, our method is benchmarked against TiNeuVox, K-Planes, and 4D-GS. + +Table 1. Ablation study on the proposed KNN rigid and time smooth regularizations. The quantitative results demonstrate the effectiveness of both regularizations. + +
reg,w/o KNN rigidw/o Time smoothfull model
PSNR28.4829.1229.92
SSIM0.920.930.94
+ +Our method consistently outperforms other methods across all metrics. + +Table 2. Quantitative results on D-NeRF synthetic blender scenes. + +
MethodPSNR ↑SSIM ↑LPIPS ↓
TiNeuVox [7]32.670.970.04
K-Planes [8]31.610.97/
4D-GS [35]33.300.980.03
Ours34.270.980.03
+ +We also compare our method against previous SOTA NeRF-based methods, including NeRF [25], Nerfies [27], HyperNeRF [28], NeRFPlayer [32], and TiNeuVox [7] on HyperNeRF [28]. We also provide comparisons with other 3DGS-based approaches concurrently proposed with our Gaussian-Flow. The training time, rendering FPS, and novel view synthesis PSNR of different methods can be found in Table. 3. Previous NeRF-based methods require at least 30 minutes to train the scene, and fail to achieve real-time rendering of the dynamic scene. Our method only requires 7 minutes of training time and can achieve real-time rendering speed, which is much faster than previous methods. Moreover, our method can achieve better performance than previous SOTA methods in terms of PSNR. + +We evaluated various methods on the Plenoptic Video dataset, as summarized in Table. 4. The comparison focuses on training time efficiency and image quality, assessed through PSNR and SSIM. Our approach demonstrated a significant advancement in training efficiency, requiring only 22.5 minutes, a drastic reduction compared to the hours needed by methods like DyNeRF [18] and K-Planes [8]. This efficiency is paramount for practical applications, where reduced training time can be a critical factor. In terms of image quality, our method achieved a PSNR of 30.5 with 30K steps, which, while not the highest, is competitive with the leading methods. However, our method scored 0.97 in SSIM, higher than K-Planes' leading score of 0.96. This indicates a potential trade-off between training efficiency and the ability to preserve structural details in images. + +In addition, we extend our method to 60K steps on both datasets, which can further improve the performance on the HyperNeRF dataset [28], and achieve the highest performance on the Plenoptic Video dataset [18]. However, the + +training time is also increased by approximately $2 \times$ for HyperNeRF dataset [28] and about $1.5 \times$ for Plenoptic Video dataset [18], which is still much faster than previous methods. + +# 4.5. Qualitative Comparisons + +In this section, we show qualitative comparisons of our method and previous SOTA methods on the HyperNeRF [28] dataset and the Plenoptic Video dataset [18]. Figure. 6 shows the qualitative comparisons of our method and TiNeuVox [7], 4D-GS [35] on HyperNeRF dataset. Notice that our method can produce comparably clear and sharp images than previous SOTA methods, which highlights our method's superior performance in monocular conditions. Despite its overall efficacy, our method does encounter limitations in extremely thin structure, as shown in the 3D Printer scene, the thread of the 3D Printer is not clear, while other methods can produce a clear thread. + +![](images/7dbcb503f35e71708120636ecac26a1288ab19c55523199170bbbc4fd77c0d3b.jpg) +Figure 6. Qualitative comparisons of our method and TiNeVox [7] and 4D-GS [35] on the HyperNeRF [28] vrig. dataset + +We also compare our method with previous SOTA methods on the Plenoptic Video dataset [18], as shown in Figure. 7. Compared with previous SOTA methods, our method can produce more accurate color and correct structure. Moreover, our method successfully reconstructs the flame in the scene, while NeRFPlayer [8] fails to reconstruct the flame. These results suggest that our method can achieve comparable image quality with previous SOTA methods, which demonstrates the effectiveness of our method in the multiview conditions. + +# 5. Conclusion + +In this paper, we introduced Gaussian-Flow, a novel framework for dynamic 3D scene reconstruction using a point-based differentiable rendering approach. The core of our innovation lies in the DDDM, which efficiently models deformations of each 3D Gaussian point in both the time and + +Table 3. Per-scene quantitative comparisons on HyperNeRF [28] dataset. Results are gathered from papers of the corresponding methods. Our method achieves the fastest training time, the highest rendering FPS, and the highest PSNR score for novel view synthesis, setting a new state-of-the-art for image-based dynamic scene rendering. + +
MethodTrain Time↓FPS↑Broom3D PrinterChickenPeel BananaMean
PSNR↑SSIM ↑PSNR↑SSIM ↑PSNR↑SSIM ↑PSNR↑SSIM ↑PSNR↑SSIM ↑
NeRF [25]16 hours0.01319.90.65320.70.78019.90.77720.00.76920.10.745
Nerfies [27]16 hours0.01119.20.56720.60.83026.70.94322.40.87222.20.803
HyperNeRF [28]32 hours0.01119.30.59120.00.82126.90.94823.30.89622.40.814
NeRFPlayer [32]6 hours0.20821.70.63522.90.81026.30.90524.00.86323.70.803
TiNeuVox [7]30 min0.521.50.68622.80.84128.30.94724.40.87324.30.837
Ours (30K)7 min12522.50.69024.30.85729.40.93426.30.90625.60.847
Ours (60K)12 min12522.80.70925.00.87730.40.94527.00.91726.30.862
+ +![](images/7690f4ca9fd93d0e8eff58e78067e1f36103f3de8ee0121aee4262fc12b102cc.jpg) +GT + +![](images/aedc63b4ada16e09833bcb298d59c959d6b1988819d96b588e365e414eaf475e.jpg) +Ours (30K) + +![](images/cfb3fff214c6e1c7b64af5abb8d4b1dcdffec73cebb80aac0aa7baeddfe6cb88.jpg) +K-Planes + +![](images/e292cdab9c5421f6ae4efcbbcc426272c97b3e59b836184ba6ab4c56cb5908c1.jpg) +NeRFPlayer + +![](images/b2ff2c05d169d71bf100718937329c423ffd8aaa19469c16f4028896263a2a16.jpg) +DyNeRF + +![](images/f852bacbf4bfd3a0d1bf37c7783b525190d48406a149af4e2721c9020d50c2eb.jpg) +LLFF +Figure 7. Qualitative comparisons of our method and K-Planes [8], NeRFPlayer [8], DyNeRF [18], and LLFF on Plenoptic Video dataset. + +Table 4. Quantitative comparison on Plenoptic Video dataset [18]. Our training speed is $5 \times$ faster of magnitude faster than previous leading approaches. Also, we achieved the highest PSNR score among all methods. + +
MethodTrain Time ↓PSNR ↑SSIM ↑
LLFF [24]-23.2-
DyNeRF [18]1344 hours29.60.96
NeRFPlayer [32]5.5 hours30.7-
K-Planes [8]1.8 hours31.60.96
4D-GS [35]2 hours31.00.94
Ours (30K)22.5 min30.50.97
Ours (60K)41.8 min32.00.97
+ +frequency domains. This approach has enabled us to set a new state-of-the-art for 4D scene reconstruction in terms of training speed, rendering frames per second, and novel view synthesis quality. Our extensive experiments and ab + +lation studies have demonstrated the efficacy of proposed Gaussian-Flow across various datasets. We achieved significant improvements over existing methods, particularly in training speed and rendering performance. The ability to efficiently handle dynamic scenes without the computational overhead of neural networks marks a substantial leap forward in this domain. + +# 6. Limitations + +While our method excels in rendering speed and training efficiency, there is room for improvement in maintaining high-fidelity thin structures in the final rendering, as mentioned in the Section. 4.5. The current deformation model may not be able to capture the fine details of thin structures, leading to some loss of image quality. This limitation is particularly evident in scenes with complex motion patterns. Future work could focus on enhancing the balance between speed and image detail preservation, potentially through more refined deformation models or advanced regularization techniques. + +# References + +[1] Benjamin Attal, Jia-Bin Huang, Christian Richardt, Michael Zollhoefer, Johannes Kopf, Matthew O'Toole, and Changil Kim. HyperReel: High-fidelity 6-DoF video with ray-conditioned sampling. arXiv preprint arXiv:2301.02238, 2023. 3 +[2] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. arXiv:2301.09632, 2023. 2 +[3] Ang Cao and Justin Johnson. Hexplane: a fast representation for dynamic scenes. arXiv preprint arXiv:2301.09632, 2023. 3 +[4] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In Proceedings of the European Conference on Computer Vision, 2022. 2, 3 +[5] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 14304-14314. IEEE Computer Society, 2021. 2, 4 +[6] Jiemin Fang, Lingxi Xie, Xinggang Wang, Xiaopeng Zhang, Wenyu Liu, and Qi Tian. Neusample: Neural sample field for efficient view synthesis. arXiv:2111.15552, 2021. 2 +[7] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. arXiv preprint arXiv:2205.15285, 2022. 7, 8 +[8] Sara Fridovich-Keil, Giacomo Meanti, Frederik Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. arXiv preprint arXiv:2301.10241, 2023. 3, 7, 8 +[9] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance, 2023. 2 +[10] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5712-5721, 2021. 2, 4 +[11] Stephan J. Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14346-14355, 2021. 2 +[12] Peter Hedman, Pratul P. Srinivasan, Ben Mildenhall, Jonathan T. Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5875-5884, 2021. 2 +[13] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 42(4):1-14, 2023. 3 +[14] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time + +radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 2, 3 +[15] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations, 2015. 5 +[16] Andreas Kurz, Thomas Neff, Zhaoyang Lv, Michael Zollhofer, and Markus Steinberger. Adanerf: Adaptive sampling for real-time rendering of neural radiance fields. In European Conference on Computer Vision, 2022. 2 +[17] Marc Levoy and Turner Whitted. The use of points as a display primitive. 1985. 3 +[18] Tianye Li, Mira Slavcheva, Michael Zollhöfer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, and Zhaoyang Lv. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5521-5531, 2022. 1, 2, 6, 7, 8 +[19] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2, 4 +[20] D. B.* Lindell, J. N. P.* Martel, and G. Wetzstein. Autoint: Automatic integration for fast neural volume rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 2 +[21] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. In Advances in Neural Information Processing Systems, 2020. 2, 3 +[22] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural volumes: learning dynamic renderable volumes from images. ACM Transactions on Graphics, 38(4):1-14, 2019. 2, 4 +[23] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 2, 3 +[24] Ben Mildenhall, Pratul P. Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (Proceedings of SIGGRAPH), 38(4), 2019. 8 +[25] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision, pages 405-421. Springer, 2020. 2, 7, 8 +[26] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 2, 3 +[27] Keunhong Park, Utkarsh Sinha, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Steven M. Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 2, 7, 8 + +[28] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo Martin-Brulla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Trans. Graph., 40(6), 2021. 1, 2, 6, 7, 8 +[29] Martin Piala and Ronald Clark. Terminerf: Ray termination prediction for efficient neural rendering. International Conference on 3D Vision, pages 1106-1114, 2021. 2 +[30] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 1, 2, 6 +[31] Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boning Liu, Hongwen Zhang, and Yebin Liu. Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. arXiv preprint arXiv:2211.11610, 2022. 3 +[32] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerfplayer: A streamable dynamic scene representation with decomposed neural radiance fields. arXiv preprint arXiv:2210.15947, 2022. 3, 7, 8 +[33] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 2 +[34] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Improved direct voxel grid optimization for radiance fields reconstruction. arXiv preprint arXiv:2206.05085, 2022. 2 +[35] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528, 2023. 2, 3, 6, 7, 8 +[36] Liwen Wu, Jae Yong Lee, Anand Bhattachad, Yu-Xiong Wang, and David Forsyth. Diver: Real-time and accurate neural radiance fields with deterministic integration for volume rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16200-16209, 2022. 2 +[37] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9421-9431, 2021. 2, 4 +[38] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101, 2023. 2, 3 +[39] Wang Yifan, Felice Serena, Shihao Wu, Cengiz Öz Tireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM Transactions on Graphics (TOG), 38(6):1-14, 2019. 3 +[40] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering + +of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5752-5761, 2021. 2, 3 +[41] Qiang Zhang, Seung-Hwan Baek, Szymon Rusinkiewicz, and Felix Heide. Differentiable point-based radiance fields for efficient view synthesis. In SIGGRAPH Asia 2022 Conference Papers, 2022. 3 +[42] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Ewa volume splatting. In Proceedings Visualization, 2001. VIS'01., pages 29-538. IEEE, 2001. 3 \ No newline at end of file diff --git a/2024/Gaussian-Flow_ 4D Reconstruction with Dynamic 3D Gaussian Particle/images.zip b/2024/Gaussian-Flow_ 4D Reconstruction with Dynamic 3D Gaussian Particle/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..95d37d2c44b293f6ee7cc08d19d86d8aa4dbf669 --- /dev/null +++ b/2024/Gaussian-Flow_ 4D Reconstruction with Dynamic 3D Gaussian Particle/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a1e03d43da399ebd57ba1f9c49d86e07de814437824b0600b75df28198249b3e +size 604220 diff --git a/2024/Gaussian-Flow_ 4D Reconstruction with Dynamic 3D Gaussian Particle/layout.json b/2024/Gaussian-Flow_ 4D Reconstruction with Dynamic 3D Gaussian Particle/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fe3516c6a17435c6c44135681fce6534c0ca832d --- /dev/null +++ b/2024/Gaussian-Flow_ 4D Reconstruction with Dynamic 3D Gaussian Particle/layout.json @@ -0,0 +1,8608 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 76, + 103, + 517, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 103, + 517, + 120 + ], + "spans": [ + { + "bbox": [ + 76, + 103, + 517, + 120 + ], + "type": "text", + "content": "Gaussian-Flow: 4D Reconstruction with Dynamic 3D Gaussian Particle" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 142, + 205, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 142, + 205, + 156 + ], + "spans": [ + { + "bbox": [ + 133, + 142, + 205, + 156 + ], + "type": "text", + "content": "Youtian Lin1,2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 228, + 143, + 298, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 143, + 298, + 156 + ], + "spans": [ + { + "bbox": [ + 228, + 143, + 298, + 156 + ], + "type": "text", + "content": "Zuozhuo Dai3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 322, + 143, + 373, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 143, + 373, + 156 + ], + "spans": [ + { + "bbox": [ + 322, + 143, + 373, + 156 + ], + "type": "text", + "content": "Siyu Zhu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 399, + 143, + 460, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 143, + 460, + 156 + ], + "spans": [ + { + "bbox": [ + 399, + 143, + 460, + 156 + ], + "type": "text", + "content": "Yao " + }, + { + "bbox": [ + 399, + 143, + 460, + 156 + ], + "type": "inline_equation", + "content": "\\mathrm{Yao^{1,2\\boxtimes}}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 77, + 157, + 516, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 157, + 516, + 213 + ], + "spans": [ + { + "bbox": [ + 77, + 157, + 516, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 157, + 516, + 213 + ], + "type": "text", + "content": "State Key Laboratory for Novel Software Technology, Nanjing University, Nanjing, China \n" + }, + { + "bbox": [ + 77, + 157, + 516, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 77, + 157, + 516, + 213 + ], + "type": "text", + "content": "School of Intelligence Science and Technology, Nanjing University, Suzhou, China \n" + }, + { + "bbox": [ + 77, + 157, + 516, + 213 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 77, + 157, + 516, + 213 + ], + "type": "text", + "content": "Alibaba Group \n" + }, + { + "bbox": [ + 77, + 157, + 516, + 213 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 77, + 157, + 516, + 213 + ], + "type": "text", + "content": "Fudan University" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 50, + 242, + 214, + 430 + ], + "blocks": [ + { + "bbox": [ + 50, + 242, + 214, + 430 + ], + "lines": [ + { + "bbox": [ + 50, + 242, + 214, + 430 + ], + "spans": [ + { + "bbox": [ + 50, + 242, + 214, + 430 + ], + "type": "image", + "image_path": "011177e7de7e581e87375234ad14b83ca5958e7b27a20ec981f875dcdaaf5e85.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 440, + 546, + 475 + ], + "lines": [ + { + "bbox": [ + 45, + 440, + 546, + 475 + ], + "spans": [ + { + "bbox": [ + 45, + 440, + 546, + 475 + ], + "type": "text", + "content": "Figure 1. Dynamic reconstruction results of the proposed Gaussian-Flow on the monocular D-NeRF [30] and HyperNeRF Dataset [28] (left) and the multi-view Plenoptic Dataset [18] (right). Our method achieves a " + }, + { + "bbox": [ + 45, + 440, + 546, + 475 + ], + "type": "inline_equation", + "content": "5 \\times" + }, + { + "bbox": [ + 45, + 440, + 546, + 475 + ], + "type": "text", + "content": " faster training and rendering speed compared with the per-frame 3DGS modeling and significantly outperforms previous methods in novel view rendering quality." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 220, + 243, + 356, + 429 + ], + "blocks": [ + { + "bbox": [ + 220, + 243, + 356, + 429 + ], + "lines": [ + { + "bbox": [ + 220, + 243, + 356, + 429 + ], + "spans": [ + { + "bbox": [ + 220, + 243, + 356, + 429 + ], + "type": "image", + "image_path": "8409baae1f126fa7fc775f1a146696fd00a48429075bdd1afce2509f1eda6930.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 359, + 242, + 450, + 431 + ], + "blocks": [ + { + "bbox": [ + 359, + 242, + 450, + 431 + ], + "lines": [ + { + "bbox": [ + 359, + 242, + 450, + 431 + ], + "spans": [ + { + "bbox": [ + 359, + 242, + 450, + 431 + ], + "type": "image", + "image_path": "c203f09eb7b8ed5c101a2eedb8f7c227f8bd12d421e8ad62a461fdb66ceb115f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 450, + 243, + 544, + 431 + ], + "blocks": [ + { + "bbox": [ + 450, + 243, + 544, + 431 + ], + "lines": [ + { + "bbox": [ + 450, + 243, + 544, + 431 + ], + "spans": [ + { + "bbox": [ + 450, + 243, + 544, + 431 + ], + "type": "image", + "image_path": "cee1196eb8f70fdda808aa8d5768a28dcbd2522dbda0fcde4d193b64ac301275.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 143, + 483, + 192, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 483, + 192, + 496 + ], + "spans": [ + { + "bbox": [ + 143, + 483, + 192, + 496 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 45, + 510, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 510, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 45, + 510, + 289, + 715 + ], + "type": "text", + "content": "We introduce Gaussian-Flow, a novel point-based approach for fast dynamic scene reconstruction and real-time rendering from both multi-view and monocular videos. In contrast to the prevalent NeRF-based approaches hampered by slow training and rendering speeds, our approach harnesses recent advancements in point-based 3D Gaussian Splatting (3DGS). Specifically, a novel Dual-Domain Deformation Model (DDDM) is proposed to explicitly model attribute deformations of each Gaussian point, where the time-dependent residual of each attribute is captured by a polynomial fitting in the time domain, and a Fourier series fitting in the frequency domain. The proposed DDDM is capable of modeling complex scene deformations across long video footage, eliminating the need for training separate 3DGS for each frame or introducing an additional implicit neural field to model 3D dynamics. Moreover, the explicit deformation modeling for discretized Gaussian points en" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 485, + 547, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 547, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 547, + 604 + ], + "type": "text", + "content": "sures ultra-fast training and rendering of a 4D scene, which is comparable to the original 3DGS designed for static 3D reconstruction. Our proposed approach showcases a substantial efficiency improvement, achieving a " + }, + { + "bbox": [ + 304, + 485, + 547, + 604 + ], + "type": "inline_equation", + "content": "5 \\times" + }, + { + "bbox": [ + 304, + 485, + 547, + 604 + ], + "type": "text", + "content": " faster training speed compared to the per-frame 3DGS modeling. In addition, quantitative results demonstrate that the proposed Gaussian-Flow significantly outperforms previous leading methods in novel view rendering quality. Project page: https://nju-3dv.github.io/projects/Gaussian-Flow." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 632, + 387, + 644 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 632, + 387, + 644 + ], + "spans": [ + { + "bbox": [ + 306, + 632, + 387, + 644 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "content": "In the realm of digital scene synthesis, achieving a balance between high-quality reconstructions and real-time rendering is paramount, especially for applications like virtual reality (VR) playback, where immediate feedback and immersive experiences are essential. Neural Radiance Fields" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21136" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "(NeRFs) [25] has risen as a promising method for synthesizing intricate scenes. However, despite their ability to produce visually stunning results, NeRFs require costly sampling and evaluation of the neural radiance field at multiple points along each ray. Consequently, the substantial computational demands impede the real-time rendering capabilities of NeRFs. These are attempts to accelerate the rendering process of NeRFs, such as direct volume representation [21, 33, 40], neural hasing [26], and tri-plane structures [2, 4, 9]. However, it still remains a challenge for high-fidelity real-time rendering. Nevertheless, the issue becomes even more severe when turning to dynamic scene reconstruction and rendering." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 231, + 289, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 231, + 289, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 231, + 289, + 399 + ], + "type": "text", + "content": "Recent progress on 3D Gaussian Splatting (3DGS) [14] has drawn attention from the 3D computer vision community. With tile-based rasterization instead of plain volume rendering, 3DGS can render images two orders of magnitude faster than the vanilla NeRF. The technique has also been quickly applied to 4D scene reconstruction by extending to the separate per-frame 3DGS optimization [23]. However, such direct extension is storage-intensive and is not applicable to monocular video input. Some other concurrent works [35, 38] try to mix the explicit point-based 3DGS and an implicit neural field for dynamic information modeling, however, require computationally expensive forward passes of the neural network, which significantly lowers the rendering speed of the original 3DGS." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 403, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 289, + 715 + ], + "type": "text", + "content": "In this work, we propose Gaussian-Flow, an explicit particle-based deformation model designed specifically for 3DGS to model the dynamic scene without using any neural network. Gaussian-Flow can recover a high-fidelity 4D scene from captured videos, while still preserving the ultrafast training and rendering speed of the original 3DGS. In particular, we formulate a 4D scene as a set of deformable 3D Gaussian points. A novel Dual-Domain Deformation Model (DDDM) is proposed to explicitly model deformations of each Gaussian point's attributes, including position, rotation, and radiance. The time-dependent deformation residual is modeled simultaneously in time and frequency domains: we apply joint polynomial and Fourier series fitting for each deformable attribute. This compact dynamic representation greatly reduces the computation cost of the deformation model, which is a key factor in preserving the rendering speed of 3DGS. Moreover, an adaptive times-tamp scaling technique is introduced to avoid over-fitting the scene to only frames with violent motions. For robust estimation, we also regularize the motion trajectory by the KNN-based rigid and the time smooth constraints. It is also noteworthy that our discretized point-based 4D representation naturally supports the edition of both static and dynamic 3D scenes, showing the potential for unlocking a variety of downstream applications related to dynamic 3D reconstruction and rendering." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": "We have conducted extensive experiments to demonstrate the effectiveness of the proposed method on several multi-view and monocular datasets. The proposed Gaussian-Flow achieves a " + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "inline_equation", + "content": "5 \\times" + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": " faster training speed compared with the separate per-frame 3DGS modeling, and significantly outperforms prior leading methods in novel view rendering quality. Our major contributions can be summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 168, + 545, + 323 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 306, + 168, + 545, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 168, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 306, + 168, + 545, + 227 + ], + "type": "text", + "content": "- We introduce Gaussian-Flow, a novel point-based differentiable rendering approach for dynamic 3D scene reconstruction, setting a new state-of-the-art for training speed, rendering FPS, and novel view synthesis quality for 4D scene reconstruction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 228, + 545, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 228, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 306, + 228, + 545, + 287 + ], + "type": "text", + "content": "- We propose a Dual-Domain Deformation Model for efficient 4D scene training and rendering, eliminating the need for per-frame 3DGS optimization and sampling on implicit neural fields. This preserves a running speed on par with the original 3DGS with minimum overhead." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 289, + 545, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 289, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 306, + 289, + 545, + 323 + ], + "type": "text", + "content": "- We demonstrate that our discretized point-based representation supports the segmentation, edition, and composition of both static and dynamic 3D scenes." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 335, + 398, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 335, + 398, + 348 + ], + "spans": [ + { + "bbox": [ + 306, + 335, + 398, + 348 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 356, + 479, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 356, + 479, + 369 + ], + "spans": [ + { + "bbox": [ + 306, + 356, + 479, + 369 + ], + "type": "text", + "content": "2.1. Dynamic Neural Radiance Field" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 374, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 374, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 374, + 545, + 579 + ], + "type": "text", + "content": "Dynamic NeRF modeling has become a heated research topic in recent years due to the development of the neural radiance field and differentiable rendering. By treating time as an extended input dimension to NeRF, researchers successfully achieve qualified image-based 4D scene rendering [5, 10, 19, 22, 37]. To further improve the reconstruction quality and incorporate prior knowledge of motions and structures, dynamic neural scene flow methods have been proposed [27, 30], where a canonical space is constructed and then transferred to each frame with scene flow or motion fields. HyperNeRF [28] models the deformation of the object topologies by using higher-dimensional inputs, while DyNeRF [18] utilizes time-conditioned NeRF to represent a 4D scene. However, the aforementioned approaches are all based on the vanilla NeRF, which requires a long training time and does not meet the requirement of real-time rendering." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 586, + 492, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 492, + 598 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 492, + 598 + ], + "type": "text", + "content": "2.2. Accelerated Neural Radiance Field" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": "To expedite NeRF training and rendering, numerous approaches have been suggested, employing more streamlined strategies [6, 11, 12, 16, 20, 29, 36]. Other methods propose the integration of neural implicit functions with explicit 3D structures, forming a hybrid representation for faster radiance field sampling [4, 26, 33, 34]. These approaches establish strong foundations for enhancing dynamic NeRF, concurrently decreasing required training and inference time. Apart from implicit neural representations, explicit NeRF" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21137" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 132 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 132 + ], + "type": "text", + "content": "modeling has show promising results for real-time rendering: NSVF [21] employs a neural sparse voxel field for efficient NeRF sampling, which stands for the earliest attempt on the explicit NeRF modeling; PlenOctrees [40] utilizes the explicit octree structure for rendering acceleration." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 287, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 287, + 347 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 287, + 347 + ], + "type": "text", + "content": "Recent efforts have also emerged to accelerate the intricate dynamic neural radiance field. TensorRF [4] employs multiple planes as explicit representations for direct dynamic scene modeling. More recent recent approaches of this kind include K-Planes [8], Tensor4D [31], and HexPlane [3]. Alternatively, NeRFPlayer[32] introduces a unified streaming representation for both grid-based [26] and plane-based methods, utilizing separate models to distinguish static and dynamic scene components, however, leading to slow rendering times. HyperReel [1] further suggests a flexible sampling network coupled with two planes for dynamic scene representation. While these methods improve the rendering speed of a dynamic scene to some extent, it is still hard to achieve real-time rendering, let alone a good balance between the running speed and rendering quality. In contrast, we resort to the recent 3D Gaussians splatting, which applies an explicit soft point cloud representation for real-time image-based rendering." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 354, + 246, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 354, + 246, + 366 + ], + "spans": [ + { + "bbox": [ + 47, + 354, + 246, + 366 + ], + "type": "text", + "content": "2.3. Differentiable Point-based Rendering" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 372, + 287, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 372, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 372, + 287, + 635 + ], + "type": "text", + "content": "The original idea of using 3D points as rendering primitives was first introduced in [17]. By incorporating differentiable rendering, recent approaches have made remarkable progress in image-based rendering, representative methods include PointRF [41], DSS [39], and 3D Gaussians splatting (3DGS) [13]. Specifically, 3DGS [13] has demonstrated extraordinary performance in novel-view synthesis, achieving real-time rendering speed and state-of-the-art rendering quality. The method adopts a soft point representation with attributes of position, rotation, density, and radiance, and applies differentiable point-based rendering for scene optimization. 3DGS has quickly been extended to dynamic scene modeling by direct separate per-frame optimization [23], however, requires a long optimization time and a large amount of storage for long video footage. Other works [35, 38] apply an implicit motion field to model scene dynamics, but the introduction of the implicit neural network significantly slows down the sampling and rendering speed. In this work, we adopt the approach of representing 4D scenes through a purely discretized point cloud model, ensuring a fast training and rendering speed comparable with the original 3DGS." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 646, + 138, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 646, + 138, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 646, + 138, + 658 + ], + "type": "text", + "content": "3. Gaussian-Flow" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "content": "In this section, we introduce the proposed Gaussian-Flow for dynamic scene modeling. We first review the 3DGS in Sec. 3.1. Then, we introduce our explicit motion modeling of each Gaussian point by using a novel Dual-Domain" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "Deformation Model (DDDM), as outlined in Sec. 3.2. An adaptive timestamp scaling technique is described in Sec. 3.3 for balanced training of each frame. To ensure the continuity of the motion in both spatial and temporal dimensions, we incorporate appropriate regularizations on each point during the optimization, as detailed in Sec. 3.4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 152, + 480, + 165 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 152, + 480, + 165 + ], + "spans": [ + { + "bbox": [ + 305, + 152, + 480, + 165 + ], + "type": "text", + "content": "3.1. Recap on 3D Gaussian Splitting" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 169, + 545, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 169, + 545, + 314 + ], + "spans": [ + { + "bbox": [ + 304, + 169, + 545, + 314 + ], + "type": "text", + "content": "3D Gaussian Splatting [14] is designed to efficiently optimize a 3D scene for real-time and high-quality novel view synthesis. The 3DGS framework has garnered significant attention within the community due to its remarkable enhancements in both training and rendering times, concurrently achieving state-of-the-art rendering quality. In contrast to the volume rendering in the vanilla NeRF which relies on ray marching, 3DGS adopts a tile-based rasterization on a distinctive soft point cloud representation to achieve fast rendering. Specifically, 3DGS models a 3D scene as a large amount of 3D Gaussian points in the world space, where each point is represented by:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 339, + 321, + 545, + 344 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 321, + 545, + 344 + ], + "spans": [ + { + "bbox": [ + 339, + 321, + 545, + 344 + ], + "type": "interline_equation", + "content": "G (\\boldsymbol {x}) = \\exp \\left(- \\frac {1}{2} \\left(\\boldsymbol {x} - \\boldsymbol {\\mu}\\right) ^ {T} \\boldsymbol {\\Sigma} ^ {- 1} \\left(\\boldsymbol {x} - \\boldsymbol {\\mu}\\right)\\right), \\tag {1}", + "image_path": "e755cc5ddb4745282edf36ab0bc2b721e984c01b75fbfbeaa775f6db0a0bbeed.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 351, + 545, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 351, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 304, + 351, + 545, + 459 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 351, + 545, + 459 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 351, + 545, + 459 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 351, + 545, + 459 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 304, + 351, + 545, + 459 + ], + "type": "text", + "content": " are the mean position and covariance matrix of a 3D Gaussian particle. The 3DGS takes sparse Structure-from-Motion (SfM) points or even random points as input, and initializes each point to a 3D Gaussian based on its neighbors. Besides, each 3D Gaussian is associated with a learnable view-dependent radiance " + }, + { + "bbox": [ + 304, + 351, + 545, + 459 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 351, + 545, + 459 + ], + "type": "text", + "content": " and a learnable opacity " + }, + { + "bbox": [ + 304, + 351, + 545, + 459 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 351, + 545, + 459 + ], + "type": "text", + "content": " for rendering. Subsequently, an efficient 3D to 2D Gaussian mapping [42] is employed to project the point onto the image plane:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 397, + 468, + 545, + 481 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 468, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 397, + 468, + 545, + 481 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\mu} ^ {\\prime} = \\boldsymbol {P} \\boldsymbol {W} \\boldsymbol {\\mu}, \\tag {2}", + "image_path": "3ddd7ba81209b70d4020ab464cac0f2f00c90d01c5b4aa109706a3cd22bfea37.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 380, + 488, + 545, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 488, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 380, + 488, + 545, + 502 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Sigma} ^ {\\prime} = \\boldsymbol {J} \\boldsymbol {W} \\boldsymbol {\\Sigma} \\boldsymbol {W} ^ {T} \\boldsymbol {J} ^ {T}, \\tag {3}", + "image_path": "e054525e8ccba9a7e6836af8b9ccfd446d1c66e4c7412c676e51c0bacc012fc7.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 507, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 507, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 507, + 545, + 579 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 507, + 545, + 579 + ], + "type": "inline_equation", + "content": "\\mu^{\\prime}" + }, + { + "bbox": [ + 304, + 507, + 545, + 579 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 507, + 545, + 579 + ], + "type": "inline_equation", + "content": "\\Sigma^{\\prime}" + }, + { + "bbox": [ + 304, + 507, + 545, + 579 + ], + "type": "text", + "content": " represent the 2D mean position and 2D covariance of the projected 3D Gaussian. " + }, + { + "bbox": [ + 304, + 507, + 545, + 579 + ], + "type": "inline_equation", + "content": "P, W" + }, + { + "bbox": [ + 304, + 507, + 545, + 579 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 507, + 545, + 579 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 304, + 507, + 545, + 579 + ], + "type": "text", + "content": " donate the projective transformation, viewing transformation, and Jacobian of the affine approximation of " + }, + { + "bbox": [ + 304, + 507, + 545, + 579 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 507, + 545, + 579 + ], + "type": "text", + "content": ". After that, " + }, + { + "bbox": [ + 304, + 507, + 545, + 579 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 507, + 545, + 579 + ], + "type": "text", + "content": "-blending is executed to merge the overlapping Gaussians for each pixel, yielding a final color:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 364, + 588, + 545, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 588, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 364, + 588, + 545, + 622 + ], + "type": "interline_equation", + "content": "C = \\sum_ {i = 1} ^ {n} \\boldsymbol {c} _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {i}), \\tag {4}", + "image_path": "5c44193c6f81070adb74601dae19723501ba8a53d320a34ae978db7147ab4a29.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 630, + 544, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 544, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 544, + 653 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 630, + 544, + 653 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 304, + 630, + 544, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 630, + 544, + 653 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 304, + 630, + 544, + 653 + ], + "type": "text", + "content": " are the color and opacity of the " + }, + { + "bbox": [ + 304, + 630, + 544, + 653 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 630, + 544, + 653 + ], + "type": "text", + "content": "-th Gaussian, and " + }, + { + "bbox": [ + 304, + 630, + 544, + 653 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 630, + 544, + 653 + ], + "type": "text", + "content": " is the number of overlapping Gaussians." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "The attributes of the 3D Gaussian, including the mean " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": ", covariance matrix " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": ", opacity " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": ", and color " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": ", are optimized through backward propagation of the gradient flow. In particular, to ensure positive definiteness during the optimization process, the covariance matrix is parameterized" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21138" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "as a scaling vector " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": " and a rotation matrix " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\Sigma = R\\Lambda(s)\\Lambda(s)^T R^T" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\Lambda(s)" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": " is the diagonal matrix of " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": ". To facilitate optimization, the rotation matrix is further parameterized using a quaternion " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": ". Leveraging the inherent flexibility of discrete points, 3DGS incorporates adaptive density control for points. This mechanism utilizes the gradient flow to identify where geometric reconstruction is suboptimal, and employs cloning and splitting to augment the density of points for a higher rendering quality." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 192, + 228, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 192, + 228, + 205 + ], + "spans": [ + { + "bbox": [ + 47, + 192, + 228, + 205 + ], + "type": "text", + "content": "3.2. Dual-Domain Deformation Model" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 213, + 289, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 213, + 289, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 289, + 403 + ], + "type": "text", + "content": "We target to directly model the dynamics of each 3D Gaussian point by fitting each of its attributes into a time-dependent curve. Among different approaches, Polynomials fitting in the time domain and Fourier series fitting in the frequency domain are the two most widely used approaches [5, 10, 19, 22, 37], due to their simplicity and effectiveness. However, each method comes with its own advantages and drawbacks: describing the motion of a Gaussian particle in terms of polynomials yields a good fit with smooth motion with a small order of polynomials, however, can easily overfit to a violent motion if assuming a larger order of polynomials, resulting in unreasonable oscillations in the fitted trajectory. Whereas, the Fourier series excels at capturing the variations associated with violent motion, however, requires a manually reduced order when dealing with smooth motion." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 405, + 289, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 405, + 289, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 405, + 289, + 586 + ], + "type": "text", + "content": "In this work, our key insight is to use a Dual-Domain Deformation Model (DDDM) for fitting the scene dynamics, which integrates both the time domain polynomials and the frequency domain Fourier series into a unified fitting model. We assume that only the rotation " + }, + { + "bbox": [ + 46, + 405, + 289, + 586 + ], + "type": "inline_equation", + "content": "\\pmb{q}" + }, + { + "bbox": [ + 46, + 405, + 289, + 586 + ], + "type": "text", + "content": ", radiance " + }, + { + "bbox": [ + 46, + 405, + 289, + 586 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 405, + 289, + 586 + ], + "type": "text", + "content": ", and position " + }, + { + "bbox": [ + 46, + 405, + 289, + 586 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}" + }, + { + "bbox": [ + 46, + 405, + 289, + 586 + ], + "type": "text", + "content": " of a 3D Gaussian particle change over time, while the scaling " + }, + { + "bbox": [ + 46, + 405, + 289, + 586 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 405, + 289, + 586 + ], + "type": "text", + "content": " and opacity " + }, + { + "bbox": [ + 46, + 405, + 289, + 586 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 405, + 289, + 586 + ], + "type": "text", + "content": " remain constant. Specifically, we conceptualize the change in each particle's attributes as its base attributes " + }, + { + "bbox": [ + 46, + 405, + 289, + 586 + ], + "type": "inline_equation", + "content": "S_0 \\in \\{\\pmb{\\mu}_0, c_0, q_0\\}" + }, + { + "bbox": [ + 46, + 405, + 289, + 586 + ], + "type": "text", + "content": " at the reference time frame " + }, + { + "bbox": [ + 46, + 405, + 289, + 586 + ], + "type": "inline_equation", + "content": "t_0" + }, + { + "bbox": [ + 46, + 405, + 289, + 586 + ], + "type": "text", + "content": " (usually set to the first frame), superimposed on a time-dependent attribute residual " + }, + { + "bbox": [ + 46, + 405, + 289, + 586 + ], + "type": "inline_equation", + "content": "D(t)" + }, + { + "bbox": [ + 46, + 405, + 289, + 586 + ], + "type": "text", + "content": ". For simplicity, we use lowercase characters to represent a single attribute in " + }, + { + "bbox": [ + 46, + 405, + 289, + 586 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 405, + 289, + 586 + ], + "type": "text", + "content": ". The time-dependent residual of each attribute is modeled through polynomial fitting in the time domain and Fourier series fitting in the frequency domain, expressed as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 127, + 597, + 287, + 611 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 597, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 127, + 597, + 287, + 611 + ], + "type": "interline_equation", + "content": "S (t) = S _ {0} + D (t), \\tag {5}", + "image_path": "f182f88e72da0495d72d2edf00a9a88849b31014b675a9cf01d4b987405c4009.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 623, + 288, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 288, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 288, + 672 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 623, + 288, + 672 + ], + "type": "inline_equation", + "content": "D(t) = P_N(t) + F_L(t)" + }, + { + "bbox": [ + 47, + 623, + 288, + 672 + ], + "type": "text", + "content": " is combined by a polynomial " + }, + { + "bbox": [ + 47, + 623, + 288, + 672 + ], + "type": "inline_equation", + "content": "P_N(t)" + }, + { + "bbox": [ + 47, + 623, + 288, + 672 + ], + "type": "text", + "content": " with coefficients " + }, + { + "bbox": [ + 47, + 623, + 288, + 672 + ], + "type": "inline_equation", + "content": "\\pmb{a} = \\{a\\}_{n=0}^{N}" + }, + { + "bbox": [ + 47, + 623, + 288, + 672 + ], + "type": "text", + "content": " and a Fourier series " + }, + { + "bbox": [ + 47, + 623, + 288, + 672 + ], + "type": "inline_equation", + "content": "F_L(t)" + }, + { + "bbox": [ + 47, + 623, + 288, + 672 + ], + "type": "text", + "content": " with coefficients " + }, + { + "bbox": [ + 47, + 623, + 288, + 672 + ], + "type": "inline_equation", + "content": "\\pmb{f} = \\{f_{sin}^l, f_{cos}^l\\}_{l=0}^L" + }, + { + "bbox": [ + 47, + 623, + 288, + 672 + ], + "type": "text", + "content": ". These are respectively denoted as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 126, + 683, + 287, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 683, + 287, + 716 + ], + "spans": [ + { + "bbox": [ + 126, + 683, + 287, + 716 + ], + "type": "interline_equation", + "content": "P _ {N} (t) = \\sum_ {n = 0} ^ {N} a _ {n} t ^ {n}, \\tag {6}", + "image_path": "a2f6b42ebe495fcc492bb0a884be4fa460936ec7fdc18058ebea0f54debd197d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 339, + 71, + 545, + 103 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 71, + 545, + 103 + ], + "spans": [ + { + "bbox": [ + 339, + 71, + 545, + 103 + ], + "type": "interline_equation", + "content": "F _ {L} (t) = \\sum_ {l = 1} ^ {L} \\left(f _ {\\sin} ^ {l} \\cos (l t) + f _ {\\cos} ^ {l} \\sin (l t)\\right). \\tag {7}", + "image_path": "7825b9a7ee4ac065419ce74c2f132ad89f946a698248f5a32ebab4c0e31cb1f5.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 109, + 545, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 109, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 109, + 545, + 168 + ], + "type": "text", + "content": "It is important to note that we assume different dimensions of an attribute are independently changed over time. Therefore, we assign a different " + }, + { + "bbox": [ + 304, + 109, + 545, + 168 + ], + "type": "inline_equation", + "content": "D(t)" + }, + { + "bbox": [ + 304, + 109, + 545, + 168 + ], + "type": "text", + "content": " for each dimension of an attribute. For instance, we utilize " + }, + { + "bbox": [ + 304, + 109, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\{D_{\\mu_i}(t)\\}_{i=0}^3" + }, + { + "bbox": [ + 304, + 109, + 545, + 168 + ], + "type": "text", + "content": " to describe the motion of a 3D position " + }, + { + "bbox": [ + 304, + 109, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}" + }, + { + "bbox": [ + 304, + 109, + 545, + 168 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 169, + 545, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 169, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 169, + 545, + 240 + ], + "type": "text", + "content": "Figure 3 illustrates a comparative analysis of trajectory fitting using polynomial, Fourier series, and the proposed joint DDDM functions. The figure highlights the superior fitting capabilities of the DDDM approach in capturing complex motion trajectories as represented by the sampled data points." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 250, + 463, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 250, + 463, + 262 + ], + "spans": [ + { + "bbox": [ + 305, + 250, + 463, + 262 + ], + "type": "text", + "content": "3.3. Adaptive Timestemp Scaling" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 268, + 545, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 268, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 268, + 545, + 411 + ], + "type": "text", + "content": "In a typical scenario, a normalized frame index " + }, + { + "bbox": [ + 304, + 268, + 545, + 411 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 268, + 545, + 411 + ], + "type": "text", + "content": " ranging from 0 to 1 will be used as the temporal input of " + }, + { + "bbox": [ + 304, + 268, + 545, + 411 + ], + "type": "inline_equation", + "content": "D(t)" + }, + { + "bbox": [ + 304, + 268, + 545, + 411 + ], + "type": "text", + "content": ". However, this poses a challenge when endeavoring to model substantial motions within a very short time using polynomials and Fourier series. Adhering to the standard temporal division would necessitate an exceedingly large coefficient to accommodate highly intense movements within a very short time frame. This circumstance has the potential to induce instability or even a breakdown in the optimization process. To address this issue, we introduce a time dilation factor " + }, + { + "bbox": [ + 304, + 268, + 545, + 411 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 268, + 545, + 411 + ], + "type": "text", + "content": " to scale the temporal input for each Gaussian point, which is formulated as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 392, + 423, + 545, + 435 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 423, + 545, + 435 + ], + "spans": [ + { + "bbox": [ + 392, + 423, + 545, + 435 + ], + "type": "interline_equation", + "content": "t _ {s} = \\lambda_ {s} \\cdot t + \\lambda_ {b} \\tag {8}", + "image_path": "12cbf58f6e3040392438967aea767558ec0d5801b66cf53f9809d30684574c80.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 445, + 545, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 545, + 506 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 445, + 545, + 506 + ], + "type": "inline_equation", + "content": "t_s" + }, + { + "bbox": [ + 304, + 445, + 545, + 506 + ], + "type": "text", + "content": " represents the scaled time input, serving as the input of " + }, + { + "bbox": [ + 304, + 445, + 545, + 506 + ], + "type": "inline_equation", + "content": "D(t)" + }, + { + "bbox": [ + 304, + 445, + 545, + 506 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 445, + 545, + 506 + ], + "type": "inline_equation", + "content": "t \\in [0,1]" + }, + { + "bbox": [ + 304, + 445, + 545, + 506 + ], + "type": "text", + "content": " denotes the normalized frame index, and " + }, + { + "bbox": [ + 304, + 445, + 545, + 506 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 445, + 545, + 506 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 445, + 545, + 506 + ], + "type": "inline_equation", + "content": "\\lambda_b" + }, + { + "bbox": [ + 304, + 445, + 545, + 506 + ], + "type": "text", + "content": " stand for the dilation factor and base factor of a Gaussian, respectively. In all our experiments, " + }, + { + "bbox": [ + 304, + 445, + 545, + 506 + ], + "type": "inline_equation", + "content": "\\lambda_s" + }, + { + "bbox": [ + 304, + 445, + 545, + 506 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 445, + 545, + 506 + ], + "type": "inline_equation", + "content": "\\lambda_b" + }, + { + "bbox": [ + 304, + 445, + 545, + 506 + ], + "type": "text", + "content": " are initialized to 1 and 0, respectively." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 506, + 545, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 545, + 613 + ], + "type": "text", + "content": "To summarize, in our dynamic scene setting, a Gaussian particle contains multiple attributes to be optimized, including base attributes " + }, + { + "bbox": [ + 304, + 506, + 545, + 613 + ], + "type": "inline_equation", + "content": "\\{\\mu_0,q_0,s_0,c_0,\\alpha_0\\}" + }, + { + "bbox": [ + 304, + 506, + 545, + 613 + ], + "type": "text", + "content": " at reference frame " + }, + { + "bbox": [ + 304, + 506, + 545, + 613 + ], + "type": "inline_equation", + "content": "t_0" + }, + { + "bbox": [ + 304, + 506, + 545, + 613 + ], + "type": "text", + "content": ", polynomial coefficients and Fourier coefficients in " + }, + { + "bbox": [ + 304, + 506, + 545, + 613 + ], + "type": "inline_equation", + "content": "\\{D_{\\mu}(t),D_q(t),D_c(t)\\}" + }, + { + "bbox": [ + 304, + 506, + 545, + 613 + ], + "type": "text", + "content": ". Since the proposed DDDM optimizes the time-dependent residuals without the need for an intricate neural field structure, our Gaussian-Flow inherits the benefits of extreme-fast training and rendering speed from the vanilla 3DGS." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 623, + 403, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 403, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 403, + 635 + ], + "type": "text", + "content": "3.4. Regularizations" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "While the utilization of discrete points as scene representations accelerates rendering, several challenges remain. First, points are optimized individually, losing connections with their spatial neighbors, which do not align with the real-world scenario. Optimizing these Gaussian points without considering continuity will inevitably lead to a" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21139" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 75, + 69, + 88 + ], + "blocks": [ + { + "bbox": [ + 53, + 75, + 69, + 88 + ], + "lines": [ + { + "bbox": [ + 53, + 75, + 69, + 88 + ], + "spans": [ + { + "bbox": [ + 53, + 75, + 69, + 88 + ], + "type": "image", + "image_path": "0048a63e88b776549374029bf25e4d539c550b3f8d35cd19bab3b5eae4b7e4dd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 76, + 169, + 87 + ], + "lines": [ + { + "bbox": [ + 72, + 76, + 169, + 87 + ], + "spans": [ + { + "bbox": [ + 72, + 76, + 169, + 87 + ], + "type": "text", + "content": "Adaptive Timestamp Scaling" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 179, + 76, + 208, + 87 + ], + "blocks": [ + { + "bbox": [ + 179, + 76, + 208, + 87 + ], + "lines": [ + { + "bbox": [ + 179, + 76, + 208, + 87 + ], + "spans": [ + { + "bbox": [ + 179, + 76, + 208, + 87 + ], + "type": "image", + "image_path": "1f113e0850392e0cc881978378908c8e851b6024c72e500a1dfa9cce415dab62.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 335, + 76, + 363, + 87 + ], + "blocks": [ + { + "bbox": [ + 335, + 76, + 363, + 87 + ], + "lines": [ + { + "bbox": [ + 335, + 76, + 363, + 87 + ], + "spans": [ + { + "bbox": [ + 335, + 76, + 363, + 87 + ], + "type": "image", + "image_path": "507b87fe0e2e244937d05cd068f4ffc7e90ef1ac0a50cc6db51452aadebf093b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 367, + 76, + 399, + 87 + ], + "lines": [ + { + "bbox": [ + 367, + 76, + 399, + 87 + ], + "spans": [ + { + "bbox": [ + 367, + 76, + 399, + 87 + ], + "type": "text", + "content": "Radiance" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 406, + 76, + 432, + 87 + ], + "blocks": [ + { + "bbox": [ + 406, + 76, + 432, + 87 + ], + "lines": [ + { + "bbox": [ + 406, + 76, + 432, + 87 + ], + "spans": [ + { + "bbox": [ + 406, + 76, + 432, + 87 + ], + "type": "image", + "image_path": "af0a463f6e073f6ef74f5ab11dc8d0514d151ac159c84ae9c0d86c84c47b5c64.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 437, + 76, + 466, + 87 + ], + "lines": [ + { + "bbox": [ + 437, + 76, + 466, + 87 + ], + "spans": [ + { + "bbox": [ + 437, + 76, + 466, + 87 + ], + "type": "text", + "content": "Forward" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 471, + 76, + 499, + 87 + ], + "blocks": [ + { + "bbox": [ + 471, + 76, + 499, + 87 + ], + "lines": [ + { + "bbox": [ + 471, + 76, + 499, + 87 + ], + "spans": [ + { + "bbox": [ + 471, + 76, + 499, + 87 + ], + "type": "image", + "image_path": "413f4d271d9b9c08183d9d2c203e05f07a2cb0011adc7c627d3af178101cb14b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 504, + 76, + 539, + 87 + ], + "lines": [ + { + "bbox": [ + 504, + 76, + 539, + 87 + ], + "spans": [ + { + "bbox": [ + 504, + 76, + 539, + 87 + ], + "type": "text", + "content": "Backward" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 47, + 100, + 547, + 236 + ], + "blocks": [ + { + "bbox": [ + 214, + 76, + 326, + 87 + ], + "lines": [ + { + "bbox": [ + 214, + 76, + 326, + 87 + ], + "spans": [ + { + "bbox": [ + 214, + 76, + 326, + 87 + ], + "type": "text", + "content": "Dual-Domain Deformation Model" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 100, + 547, + 236 + ], + "lines": [ + { + "bbox": [ + 47, + 100, + 547, + 236 + ], + "spans": [ + { + "bbox": [ + 47, + 100, + 547, + 236 + ], + "type": "image", + "image_path": "5736ef64b98c15315404595634f02571a3b576feeae6e06e7711771c647a60d7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 56, + 312, + 162, + 415 + ], + "blocks": [ + { + "bbox": [ + 46, + 250, + 547, + 285 + ], + "lines": [ + { + "bbox": [ + 46, + 250, + 547, + 285 + ], + "spans": [ + { + "bbox": [ + 46, + 250, + 547, + 285 + ], + "type": "text", + "content": "Figure 2. Overview of the Gaussian-Flow pipeline. We model the deformation of attributes of each 3D Gaussian point independently by using the Dual-Domain Deformation Model (DDDM), which preserves the discretized nature of the 3D Gaussian points, and thus achieves ultra-fast training and rendering speed comparable with the original 3DGS." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 312, + 162, + 415 + ], + "lines": [ + { + "bbox": [ + 56, + 312, + 162, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 312, + 162, + 415 + ], + "type": "image", + "image_path": "67dee80a23b0eb50805662b4088b8f741b9a27a8df8f1a32c08642bc80092308.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 173, + 315, + 277, + 414 + ], + "blocks": [ + { + "bbox": [ + 173, + 315, + 277, + 414 + ], + "lines": [ + { + "bbox": [ + 173, + 315, + 277, + 414 + ], + "spans": [ + { + "bbox": [ + 173, + 315, + 277, + 414 + ], + "type": "image", + "image_path": "943b01e627be6bb69f0a616bfcc9579a1416ad2ecaee302a292f5c25ab2d855b.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 426, + 288, + 460 + ], + "lines": [ + { + "bbox": [ + 46, + 426, + 288, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 288, + 460 + ], + "type": "text", + "content": "Figure 3. Two exemplar motion fittings using polynomial, Fourier Series, and the joint DDDM functions. Our DDDM is able to accurately fit complex trajectories denoted by the sampling points." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 483, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 287, + 555 + ], + "type": "text", + "content": "degradation in reconstruction quality and spatial coherence. Additionally, motions should be smoothed over time. Based on these observations, we employ two regularizations, namely time smoothness and a KNN rigid regularization, for robust optimization of Gaussian points and their motions." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 578, + 287, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 578, + 287, + 650 + ], + "spans": [ + { + "bbox": [ + 46, + 578, + 287, + 650 + ], + "type": "text", + "content": "Time Smoothness Loss To ensure temporal smoothness over time, we apply a perturbing " + }, + { + "bbox": [ + 46, + 578, + 287, + 650 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 578, + 287, + 650 + ], + "type": "text", + "content": " on the input timestamp " + }, + { + "bbox": [ + 46, + 578, + 287, + 650 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 578, + 287, + 650 + ], + "type": "text", + "content": " and encourage the time-dependent attributes (i.e., position " + }, + { + "bbox": [ + 46, + 578, + 287, + 650 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 578, + 287, + 650 + ], + "type": "text", + "content": ", rotation " + }, + { + "bbox": [ + 46, + 578, + 287, + 650 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 46, + 578, + 287, + 650 + ], + "type": "text", + "content": " and radiance " + }, + { + "bbox": [ + 46, + 578, + 287, + 650 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 578, + 287, + 650 + ], + "type": "text", + "content": ") at time " + }, + { + "bbox": [ + 46, + 578, + 287, + 650 + ], + "type": "inline_equation", + "content": "t + \\epsilon" + }, + { + "bbox": [ + 46, + 578, + 287, + 650 + ], + "type": "text", + "content": " to be consistent with those at time " + }, + { + "bbox": [ + 46, + 578, + 287, + 650 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 578, + 287, + 650 + ], + "type": "text", + "content": ". The time smoothness term is defined as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 110, + 654, + 287, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 654, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 110, + 654, + 287, + 668 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {t} = \\| D (t) - D (t + \\epsilon) \\| _ {2}. \\tag {9}", + "image_path": "c1050336f4588b5c2518098dd61a130afd6fe7fbdb3bde04d44791859aee9761.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": "It is noteworthy that the magnitude of the perturbing value is adaptively set according to the number of total frames, i.e., " + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon = 0.1/" + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": " frames." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 304, + 547, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 547, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 547, + 449 + ], + "type": "text", + "content": "KNN Rigid Loss During the optimization of 3D Gaussians with adaptive density control, points will be dynamically added or removed. This dynamic nature implies that the neighbors of points within a local space are subject to constant changes, posing a challenge for directly enforcing a spatial local consistency constraint. To sidestep this problem, we propose to divide the optimization part into two alternating stages: in the former stage, we optimize all variables with adaptive density control; while in the latter stage, we optimize the attributes without adding or removing points. The local rigid constraint is incorporated in every latter stage, and it is defined as:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 364, + 458, + 545, + 485 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 458, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 364, + 458, + 545, + 485 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s} = \\sum_ {j \\in \\mathcal {N} _ {i}} \\| D (t) _ {i} - D (t) _ {j} \\| _ {2} \\tag {10}", + "image_path": "79a110c3f285be0881c72ea9933f99107afbd692f077cd00053b59e9159423bd.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 494, + 547, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 494, + 547, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 494, + 547, + 517 + ], + "type": "text", + "content": "where, " + }, + { + "bbox": [ + 304, + 494, + 547, + 517 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_i" + }, + { + "bbox": [ + 304, + 494, + 547, + 517 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 304, + 494, + 547, + 517 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 494, + 547, + 517 + ], + "type": "text", + "content": " nearest neighbor (KNN) of " + }, + { + "bbox": [ + 304, + 494, + 547, + 517 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 494, + 547, + 517 + ], + "type": "text", + "content": "-th Gaussian." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 530, + 388, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 530, + 388, + 544 + ], + "spans": [ + { + "bbox": [ + 306, + 530, + 388, + 544 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 551, + 440, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 551, + 440, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 551, + 440, + 563 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": "We train our model using the Adam [15] optimizer with separate learning rates for different attributes of the Gaussian point. We set a learning rate of " + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "inline_equation", + "content": "4 \\times 10^{-4}" + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": " for the point position with an exponential decay of " + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "inline_equation", + "content": "8 \\times 10^{-7}" + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": ". The learning rates for point rotation and all DDDM parameters are set to 0.002 and " + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "inline_equation", + "content": "4 \\times 10^{-4}" + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": " respectively. We apply a weight decay of " + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "inline_equation", + "content": "8 \\times 10^{-7}" + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": " to all parameters. The rest of the learning rate follows the 3DGS setting. We train the model for 30K steps and 60K steps for all scenes. All experiments are conducted on a single NVIDIA RTX 4090 GPU with 24GB memory. In addition, we use Taichi to implement our DDDM model, which can parallelize the computation of the DDDM (i.e." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21140" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "polynomial and Fourier series computation) of each Gaussian point." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 105, + 110, + 116 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 105, + 110, + 116 + ], + "spans": [ + { + "bbox": [ + 47, + 105, + 110, + 116 + ], + "type": "text", + "content": "4.2. Datasets" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 124, + 287, + 160 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 124, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 47, + 124, + 287, + 160 + ], + "type": "text", + "content": "We evaluate our method on both multi-view and monocular datasets, to demonstrate the effectiveness of our method in both settings." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 177, + 288, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 177, + 288, + 261 + ], + "spans": [ + { + "bbox": [ + 46, + 177, + 288, + 261 + ], + "type": "text", + "content": "Plenoptic Video dataset [18] The dataset was captured using 21 cameras at a resolution of " + }, + { + "bbox": [ + 46, + 177, + 288, + 261 + ], + "type": "inline_equation", + "content": "2704 \\times 2028" + }, + { + "bbox": [ + 46, + 177, + 288, + 261 + ], + "type": "text", + "content": ", with each camera recording a 10-second video. Six scenes from this dataset are publicly available. For a fair comparison, we downsampled the images to " + }, + { + "bbox": [ + 46, + 177, + 288, + 261 + ], + "type": "inline_equation", + "content": "1352 \\times 1014" + }, + { + "bbox": [ + 46, + 177, + 288, + 261 + ], + "type": "text", + "content": " resolution in our experiments to keep the same setting from the concurrent work of 4D Gaussian [35]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 278, + 288, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 278, + 288, + 410 + ], + "spans": [ + { + "bbox": [ + 46, + 278, + 288, + 410 + ], + "type": "text", + "content": "HyperNeRF dataset [28] This dataset uses a monocular camera (e.g., iPhone) to record real-world motions, which includes real rigid and non-rigidly deforming scenes, such as a person splitting a cookie. The dataset is rather challenging due to large motions, complex lighting conditions, and thin object structures. To ensure a fair comparison, we downsampled images to " + }, + { + "bbox": [ + 46, + 278, + 288, + 410 + ], + "type": "inline_equation", + "content": "540 \\times 960" + }, + { + "bbox": [ + 46, + 278, + 288, + 410 + ], + "type": "text", + "content": " in our experiments and followed the training and validation camera split provided by [28]. We conducted experiments on the four \"vrig\" scenes, and we also provided results of the \"interp\" scenes in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 419, + 141, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 419, + 141, + 431 + ], + "spans": [ + { + "bbox": [ + 47, + 419, + 141, + 431 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 437, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 616 + ], + "type": "text", + "content": "Deformation Models The deformation model is the core component of the proposed Gaussian-Flow. We conduct ablation studies to validate the effectiveness of the particular choice of DDDM. As DDDM consists of a Fourier series and a polynomial function, hence we first study the Fourier series and the polynomial functions separately, in which we only use the Fourier series or the polynomial function as the deformation model. As shown in Figure. 4, the Fourier series contains more high-frequency components than the polynomial function, thus the Fourier series has sharper image details but results in more artifacts. The polynomial function is smoother than the Fourier series, leading to fewer artifacts but burry scene renderings. Finally, the hybrid DDDM function is able to generate sharper details with fewer artifacts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "content": "Furthermore, we study the orders of the polynomial and Fourier series functions in our DDDM, which are related to the complexity of the scene and are crucial to the final performance. As shown in Figure. 5, the performance of our method increases with the number of Fourier series orders but starts to drop after the order over 32, which should be related to the over-parameterization of the deformation model." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 70, + 539, + 304 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 539, + 304 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 539, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 539, + 304 + ], + "type": "image", + "image_path": "3bd3706bbd2d1872071390dc232a03b4643e2ff896e4d6c41132b5b9c09b8a62.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 312, + 546, + 358 + ], + "lines": [ + { + "bbox": [ + 305, + 312, + 546, + 358 + ], + "spans": [ + { + "bbox": [ + 305, + 312, + 546, + 358 + ], + "type": "text", + "content": "Figure 4. Ablation study on different deformation models. From left to right are deformation fitting with polynomial function only, Fourier series only, and our dual-domain deformation fitting. The proposed DDDM achieves the best rendering quality qualitatively." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 371, + 544, + 497 + ], + "blocks": [ + { + "bbox": [ + 309, + 371, + 544, + 497 + ], + "lines": [ + { + "bbox": [ + 309, + 371, + 544, + 497 + ], + "spans": [ + { + "bbox": [ + 309, + 371, + 544, + 497 + ], + "type": "image", + "image_path": "8818a3012ee2e26b2c1c54429a121634cf9494ac293c2df8b6b1f0b84e4e1706.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 506, + 546, + 540 + ], + "lines": [ + { + "bbox": [ + 305, + 506, + 546, + 540 + ], + "spans": [ + { + "bbox": [ + 305, + 506, + 546, + 540 + ], + "type": "text", + "content": "Figure 5. Ablation study on different orders of the DDDM. We find an order number of 16 leads to the best novel view rendering results in the HyperNeRF dataset." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 563, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 563, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 305, + 563, + 545, + 647 + ], + "type": "text", + "content": "Regularizations Next, we study the effectiveness of the two proposed regularizations in Gaussian-Flow optimization. As shown in Table. 1, adding separate KNN rigid regularization or the time smooth regularization can both improve the novel view rendering quality, and the full model that contains both regularizations can achieve the best performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 658, + 453, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 453, + 671 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 453, + 671 + ], + "type": "text", + "content": "4.4. Quantitative Comparisons" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 677, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 546, + 713 + ], + "type": "text", + "content": "We compare our method with SOTA on D-NeRF [30] synthetic scenes. As presented in Table 2, our method is benchmarked against TiNeuVox, K-Planes, and 4D-GS." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21141" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 113, + 282, + 159 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 288, + 104 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 288, + 104 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 288, + 104 + ], + "type": "text", + "content": "Table 1. Ablation study on the proposed KNN rigid and time smooth regularizations. The quantitative results demonstrate the effectiveness of both regularizations." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 113, + 282, + 159 + ], + "lines": [ + { + "bbox": [ + 52, + 113, + 282, + 159 + ], + "spans": [ + { + "bbox": [ + 52, + 113, + 282, + 159 + ], + "type": "table", + "html": "
reg,w/o KNN rigidw/o Time smoothfull model
PSNR28.4829.1229.92
SSIM0.920.930.94
", + "image_path": "c297ff6e7f4f0f37e6482425de21161bb484ea3dbd01006f2b04cb22a4dc1300.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 165, + 287, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 165, + 287, + 190 + ], + "spans": [ + { + "bbox": [ + 47, + 165, + 287, + 190 + ], + "type": "text", + "content": "Our method consistently outperforms other methods across all metrics." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 48, + 223, + 287, + 297 + ], + "blocks": [ + { + "bbox": [ + 47, + 198, + 286, + 210 + ], + "lines": [ + { + "bbox": [ + 47, + 198, + 286, + 210 + ], + "spans": [ + { + "bbox": [ + 47, + 198, + 286, + 210 + ], + "type": "text", + "content": "Table 2. Quantitative results on D-NeRF synthetic blender scenes." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 223, + 287, + 297 + ], + "lines": [ + { + "bbox": [ + 48, + 223, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 223, + 287, + 297 + ], + "type": "table", + "html": "
MethodPSNR ↑SSIM ↑LPIPS ↓
TiNeuVox [7]32.670.970.04
K-Planes [8]31.610.97/
4D-GS [35]33.300.980.03
Ours34.270.980.03
", + "image_path": "c2e7449e7efe6089bb98050adbc5b803221b593b0d6ffcf74f70eb2a72fbc4b2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 306, + 287, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 306, + 287, + 473 + ], + "spans": [ + { + "bbox": [ + 46, + 306, + 287, + 473 + ], + "type": "text", + "content": "We also compare our method against previous SOTA NeRF-based methods, including NeRF [25], Nerfies [27], HyperNeRF [28], NeRFPlayer [32], and TiNeuVox [7] on HyperNeRF [28]. We also provide comparisons with other 3DGS-based approaches concurrently proposed with our Gaussian-Flow. The training time, rendering FPS, and novel view synthesis PSNR of different methods can be found in Table. 3. Previous NeRF-based methods require at least 30 minutes to train the scene, and fail to achieve real-time rendering of the dynamic scene. Our method only requires 7 minutes of training time and can achieve real-time rendering speed, which is much faster than previous methods. Moreover, our method can achieve better performance than previous SOTA methods in terms of PSNR." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 474, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 287, + 665 + ], + "type": "text", + "content": "We evaluated various methods on the Plenoptic Video dataset, as summarized in Table. 4. The comparison focuses on training time efficiency and image quality, assessed through PSNR and SSIM. Our approach demonstrated a significant advancement in training efficiency, requiring only 22.5 minutes, a drastic reduction compared to the hours needed by methods like DyNeRF [18] and K-Planes [8]. This efficiency is paramount for practical applications, where reduced training time can be a critical factor. In terms of image quality, our method achieved a PSNR of 30.5 with 30K steps, which, while not the highest, is competitive with the leading methods. However, our method scored 0.97 in SSIM, higher than K-Planes' leading score of 0.96. This indicates a potential trade-off between training efficiency and the ability to preserve structural details in images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "In addition, we extend our method to 60K steps on both datasets, which can further improve the performance on the HyperNeRF dataset [28], and achieve the highest performance on the Plenoptic Video dataset [18]. However, the" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "training time is also increased by approximately " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": " for HyperNeRF dataset [28] and about " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "1.5 \\times" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": " for Plenoptic Video dataset [18], which is still much faster than previous methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 125, + 446, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 125, + 446, + 138 + ], + "spans": [ + { + "bbox": [ + 306, + 125, + 446, + 138 + ], + "type": "text", + "content": "4.5. Qualitative Comparisons" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 144, + 545, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 144, + 545, + 288 + ], + "spans": [ + { + "bbox": [ + 304, + 144, + 545, + 288 + ], + "type": "text", + "content": "In this section, we show qualitative comparisons of our method and previous SOTA methods on the HyperNeRF [28] dataset and the Plenoptic Video dataset [18]. Figure. 6 shows the qualitative comparisons of our method and TiNeuVox [7], 4D-GS [35] on HyperNeRF dataset. Notice that our method can produce comparably clear and sharp images than previous SOTA methods, which highlights our method's superior performance in monocular conditions. Despite its overall efficacy, our method does encounter limitations in extremely thin structure, as shown in the 3D Printer scene, the thread of the 3D Printer is not clear, while other methods can produce a clear thread." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 320, + 297, + 533, + 463 + ], + "blocks": [ + { + "bbox": [ + 320, + 297, + 533, + 463 + ], + "lines": [ + { + "bbox": [ + 320, + 297, + 533, + 463 + ], + "spans": [ + { + "bbox": [ + 320, + 297, + 533, + 463 + ], + "type": "image", + "image_path": "7dbcb503f35e71708120636ecac26a1288ab19c55523199170bbbc4fd77c0d3b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 472, + 545, + 495 + ], + "lines": [ + { + "bbox": [ + 305, + 472, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 305, + 472, + 545, + 495 + ], + "type": "text", + "content": "Figure 6. Qualitative comparisons of our method and TiNeVox [7] and 4D-GS [35] on the HyperNeRF [28] vrig. dataset" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 505, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 505, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 505, + 545, + 624 + ], + "type": "text", + "content": "We also compare our method with previous SOTA methods on the Plenoptic Video dataset [18], as shown in Figure. 7. Compared with previous SOTA methods, our method can produce more accurate color and correct structure. Moreover, our method successfully reconstructs the flame in the scene, while NeRFPlayer [8] fails to reconstruct the flame. These results suggest that our method can achieve comparable image quality with previous SOTA methods, which demonstrates the effectiveness of our method in the multiview conditions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 634, + 378, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 634, + 378, + 645 + ], + "spans": [ + { + "bbox": [ + 306, + 634, + 378, + 645 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "In this paper, we introduced Gaussian-Flow, a novel framework for dynamic 3D scene reconstruction using a point-based differentiable rendering approach. The core of our innovation lies in the DDDM, which efficiently models deformations of each 3D Gaussian point in both the time and" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21142" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 53, + 117, + 541, + 213 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 545, + 104 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 545, + 104 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 545, + 104 + ], + "type": "text", + "content": "Table 3. Per-scene quantitative comparisons on HyperNeRF [28] dataset. Results are gathered from papers of the corresponding methods. Our method achieves the fastest training time, the highest rendering FPS, and the highest PSNR score for novel view synthesis, setting a new state-of-the-art for image-based dynamic scene rendering." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 53, + 117, + 541, + 213 + ], + "lines": [ + { + "bbox": [ + 53, + 117, + 541, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 117, + 541, + 213 + ], + "type": "table", + "html": "
MethodTrain Time↓FPS↑Broom3D PrinterChickenPeel BananaMean
PSNR↑SSIM ↑PSNR↑SSIM ↑PSNR↑SSIM ↑PSNR↑SSIM ↑PSNR↑SSIM ↑
NeRF [25]16 hours0.01319.90.65320.70.78019.90.77720.00.76920.10.745
Nerfies [27]16 hours0.01119.20.56720.60.83026.70.94322.40.87222.20.803
HyperNeRF [28]32 hours0.01119.30.59120.00.82126.90.94823.30.89622.40.814
NeRFPlayer [32]6 hours0.20821.70.63522.90.81026.30.90524.00.86323.70.803
TiNeuVox [7]30 min0.521.50.68622.80.84128.30.94724.40.87324.30.837
Ours (30K)7 min12522.50.69024.30.85729.40.93426.30.90625.60.847
Ours (60K)12 min12522.80.70925.00.87730.40.94527.00.91726.30.862
", + "image_path": "c6323032c553b14af0ff01a30d42ed7430389d73942c93be8cd8e6e482ffb66b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 108, + 224, + 233, + 318 + ], + "blocks": [ + { + "bbox": [ + 108, + 224, + 233, + 318 + ], + "lines": [ + { + "bbox": [ + 108, + 224, + 233, + 318 + ], + "spans": [ + { + "bbox": [ + 108, + 224, + 233, + 318 + ], + "type": "image", + "image_path": "7690f4ca9fd93d0e8eff58e78067e1f36103f3de8ee0121aee4262fc12b102cc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 164, + 319, + 176, + 328 + ], + "lines": [ + { + "bbox": [ + 164, + 319, + 176, + 328 + ], + "spans": [ + { + "bbox": [ + 164, + 319, + 176, + 328 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 235, + 224, + 359, + 318 + ], + "blocks": [ + { + "bbox": [ + 235, + 224, + 359, + 318 + ], + "lines": [ + { + "bbox": [ + 235, + 224, + 359, + 318 + ], + "spans": [ + { + "bbox": [ + 235, + 224, + 359, + 318 + ], + "type": "image", + "image_path": "aedc63b4ada16e09833bcb298d59c959d6b1988819d96b588e365e414eaf475e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 277, + 319, + 316, + 328 + ], + "lines": [ + { + "bbox": [ + 277, + 319, + 316, + 328 + ], + "spans": [ + { + "bbox": [ + 277, + 319, + 316, + 328 + ], + "type": "text", + "content": "Ours (30K)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 360, + 224, + 485, + 318 + ], + "blocks": [ + { + "bbox": [ + 360, + 224, + 485, + 318 + ], + "lines": [ + { + "bbox": [ + 360, + 224, + 485, + 318 + ], + "spans": [ + { + "bbox": [ + 360, + 224, + 485, + 318 + ], + "type": "image", + "image_path": "cfb3fff214c6e1c7b64af5abb8d4b1dcdffec73cebb80aac0aa7baeddfe6cb88.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 408, + 319, + 438, + 328 + ], + "lines": [ + { + "bbox": [ + 408, + 319, + 438, + 328 + ], + "spans": [ + { + "bbox": [ + 408, + 319, + 438, + 328 + ], + "type": "text", + "content": "K-Planes" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 108, + 330, + 233, + 424 + ], + "blocks": [ + { + "bbox": [ + 108, + 330, + 233, + 424 + ], + "lines": [ + { + "bbox": [ + 108, + 330, + 233, + 424 + ], + "spans": [ + { + "bbox": [ + 108, + 330, + 233, + 424 + ], + "type": "image", + "image_path": "e292cdab9c5421f6ae4efcbbcc426272c97b3e59b836184ba6ab4c56cb5908c1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 149, + 426, + 192, + 436 + ], + "lines": [ + { + "bbox": [ + 149, + 426, + 192, + 436 + ], + "spans": [ + { + "bbox": [ + 149, + 426, + 192, + 436 + ], + "type": "text", + "content": "NeRFPlayer" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 235, + 330, + 359, + 424 + ], + "blocks": [ + { + "bbox": [ + 235, + 330, + 359, + 424 + ], + "lines": [ + { + "bbox": [ + 235, + 330, + 359, + 424 + ], + "spans": [ + { + "bbox": [ + 235, + 330, + 359, + 424 + ], + "type": "image", + "image_path": "b2ff2c05d169d71bf100718937329c423ffd8aaa19469c16f4028896263a2a16.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 281, + 426, + 312, + 435 + ], + "lines": [ + { + "bbox": [ + 281, + 426, + 312, + 435 + ], + "spans": [ + { + "bbox": [ + 281, + 426, + 312, + 435 + ], + "type": "text", + "content": "DyNeRF" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 361, + 330, + 485, + 424 + ], + "blocks": [ + { + "bbox": [ + 361, + 330, + 485, + 424 + ], + "lines": [ + { + "bbox": [ + 361, + 330, + 485, + 424 + ], + "spans": [ + { + "bbox": [ + 361, + 330, + 485, + 424 + ], + "type": "image", + "image_path": "f852bacbf4bfd3a0d1bf37c7783b525190d48406a149af4e2721c9020d50c2eb.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 413, + 426, + 433, + 435 + ], + "lines": [ + { + "bbox": [ + 413, + 426, + 433, + 435 + ], + "spans": [ + { + "bbox": [ + 413, + 426, + 433, + 435 + ], + "type": "text", + "content": "LLFF" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 445, + 544, + 457 + ], + "lines": [ + { + "bbox": [ + 46, + 445, + 544, + 457 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 544, + 457 + ], + "type": "text", + "content": "Figure 7. Qualitative comparisons of our method and K-Planes [8], NeRFPlayer [8], DyNeRF [18], and LLFF on Plenoptic Video dataset." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 48, + 533, + 287, + 640 + ], + "blocks": [ + { + "bbox": [ + 46, + 476, + 287, + 521 + ], + "lines": [ + { + "bbox": [ + 46, + 476, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 476, + 287, + 521 + ], + "type": "text", + "content": "Table 4. Quantitative comparison on Plenoptic Video dataset [18]. Our training speed is " + }, + { + "bbox": [ + 46, + 476, + 287, + 521 + ], + "type": "inline_equation", + "content": "5 \\times" + }, + { + "bbox": [ + 46, + 476, + 287, + 521 + ], + "type": "text", + "content": " faster of magnitude faster than previous leading approaches. Also, we achieved the highest PSNR score among all methods." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 533, + 287, + 640 + ], + "lines": [ + { + "bbox": [ + 48, + 533, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 287, + 640 + ], + "type": "table", + "html": "
MethodTrain Time ↓PSNR ↑SSIM ↑
LLFF [24]-23.2-
DyNeRF [18]1344 hours29.60.96
NeRFPlayer [32]5.5 hours30.7-
K-Planes [8]1.8 hours31.60.96
4D-GS [35]2 hours31.00.94
Ours (30K)22.5 min30.50.97
Ours (60K)41.8 min32.00.97
", + "image_path": "5c650e2f1d4e7980bfebe30a77a6f326912607fdfc4540f02406459b3fd05d5f.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "content": "frequency domains. This approach has enabled us to set a new state-of-the-art for 4D scene reconstruction in terms of training speed, rendering frames per second, and novel view synthesis quality. Our extensive experiments and ab" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 477, + 545, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 477, + 545, + 561 + ], + "spans": [ + { + "bbox": [ + 304, + 477, + 545, + 561 + ], + "type": "text", + "content": "lation studies have demonstrated the efficacy of proposed Gaussian-Flow across various datasets. We achieved significant improvements over existing methods, particularly in training speed and rendering performance. The ability to efficiently handle dynamic scenes without the computational overhead of neural networks marks a substantial leap forward in this domain." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 573, + 381, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 573, + 381, + 585 + ], + "spans": [ + { + "bbox": [ + 306, + 573, + 381, + 585 + ], + "type": "text", + "content": "6. Limitations" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 592, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 592, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 592, + 545, + 713 + ], + "type": "text", + "content": "While our method excels in rendering speed and training efficiency, there is room for improvement in maintaining high-fidelity thin structures in the final rendering, as mentioned in the Section. 4.5. The current deformation model may not be able to capture the fine details of thin structures, leading to some loss of image quality. This limitation is particularly evident in scenes with complex motion patterns. Future work could focus on enhancing the balance between speed and image detail preservation, potentially through more refined deformation models or advanced regularization techniques." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21143" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 95, + 106, + 108 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 106, + 108 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 106, + 108 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 115, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 115, + 287, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 115, + 287, + 170 + ], + "spans": [ + { + "bbox": [ + 53, + 115, + 287, + 170 + ], + "type": "text", + "content": "[1] Benjamin Attal, Jia-Bin Huang, Christian Richardt, Michael Zollhoefer, Johannes Kopf, Matthew O'Toole, and Changil Kim. HyperReel: High-fidelity 6-DoF video with ray-conditioned sampling. arXiv preprint arXiv:2301.02238, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 172, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 172, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 53, + 172, + 287, + 194 + ], + "type": "text", + "content": "[2] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. arXiv:2301.09632, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 195, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 195, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 53, + 195, + 287, + 228 + ], + "type": "text", + "content": "[3] Ang Cao and Justin Johnson. Hexplane: a fast representation for dynamic scenes. arXiv preprint arXiv:2301.09632, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 229, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 229, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 53, + 229, + 288, + 262 + ], + "type": "text", + "content": "[4] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In Proceedings of the European Conference on Computer Vision, 2022. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 264, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 264, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 53, + 264, + 287, + 319 + ], + "type": "text", + "content": "[5] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 14304-14314. IEEE Computer Society, 2021. 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 320, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 320, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 53, + 320, + 287, + 352 + ], + "type": "text", + "content": "[6] Jiemin Fang, Lingxi Xie, Xinggang Wang, Xiaopeng Zhang, Wenyu Liu, and Qi Tian. Neusample: Neural sample field for efficient view synthesis. arXiv:2111.15552, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 354, + 287, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 354, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 53, + 354, + 287, + 397 + ], + "type": "text", + "content": "[7] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. arXiv preprint arXiv:2205.15285, 2022. 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 399, + 287, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 399, + 287, + 442 + ], + "spans": [ + { + "bbox": [ + 53, + 399, + 287, + 442 + ], + "type": "text", + "content": "[8] Sara Fridovich-Keil, Giacomo Meanti, Frederik Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. arXiv preprint arXiv:2301.10241, 2023. 3, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 444, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 444, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 53, + 444, + 287, + 487 + ], + "type": "text", + "content": "[9] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 489, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 489, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 489, + 287, + 533 + ], + "type": "text", + "content": "[10] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5712-5721, 2021. 2, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 534, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 287, + 588 + ], + "type": "text", + "content": "[11] Stephan J. Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14346-14355, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 590, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 645 + ], + "type": "text", + "content": "[12] Peter Hedman, Pratul P. Srinivasan, Ben Mildenhall, Jonathan T. Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5875-5884, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[13] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 42(4):1-14, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "text", + "content": "[14] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "type": "text", + "content": "[15] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations, 2015. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 130, + 545, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 130, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 308, + 130, + 545, + 174 + ], + "type": "text", + "content": "[16] Andreas Kurz, Thomas Neff, Zhaoyang Lv, Michael Zollhofer, and Markus Steinberger. Adanerf: Adaptive sampling for real-time rendering of neural radiance fields. In European Conference on Computer Vision, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 175, + 545, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 175, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 308, + 175, + 545, + 198 + ], + "type": "text", + "content": "[17] Marc Levoy and Turner Whitted. The use of points as a display primitive. 1985. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 199, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 199, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 199, + 545, + 274 + ], + "type": "text", + "content": "[18] Tianye Li, Mira Slavcheva, Michael Zollhöfer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, and Zhaoyang Lv. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5521-5531, 2022. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 276, + 545, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 545, + 320 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 545, + 320 + ], + "type": "text", + "content": "[19] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 322, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 322, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 308, + 322, + 545, + 365 + ], + "type": "text", + "content": "[20] D. B.* Lindell, J. N. P.* Martel, and G. Wetzstein. Autoint: Automatic integration for fast neural volume rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 366, + 545, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 366, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 308, + 366, + 545, + 399 + ], + "type": "text", + "content": "[21] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. In Advances in Neural Information Processing Systems, 2020. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 401, + 545, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 401, + 545, + 444 + ], + "spans": [ + { + "bbox": [ + 308, + 401, + 545, + 444 + ], + "type": "text", + "content": "[22] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural volumes: learning dynamic renderable volumes from images. ACM Transactions on Graphics, 38(4):1-14, 2019. 2, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 445, + 545, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 445, + 545, + 489 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 545, + 489 + ], + "type": "text", + "content": "[23] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 491, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 491, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 491, + 545, + 555 + ], + "type": "text", + "content": "[24] Ben Mildenhall, Pratul P. Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (Proceedings of SIGGRAPH), 38(4), 2019. 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 557, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 557, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 557, + 545, + 612 + ], + "type": "text", + "content": "[25] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision, pages 405-421. Springer, 2020. 2, 7, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 613, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 657 + ], + "type": "text", + "content": "[26] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 2, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 658, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 714 + ], + "type": "text", + "content": "[27] Keunhong Park, Utkarsh Sinha, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Steven M. Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 2, 7, 8" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "21144" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "type": "text", + "content": "[28] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo Martin-Brulla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Trans. Graph., 40(6), 2021. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 140, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 140, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 49, + 140, + 287, + 173 + ], + "type": "text", + "content": "[29] Martin Piala and Ronald Clark. Terminerf: Ray termination prediction for efficient neural rendering. International Conference on 3D Vision, pages 1106-1114, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 175, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 175, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 49, + 175, + 287, + 228 + ], + "type": "text", + "content": "[30] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 1, 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 230, + 287, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 230, + 287, + 283 + ], + "spans": [ + { + "bbox": [ + 49, + 230, + 287, + 283 + ], + "type": "text", + "content": "[31] Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boning Liu, Hongwen Zhang, and Yebin Liu. Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. arXiv preprint arXiv:2211.11610, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 286, + 287, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 286, + 287, + 341 + ], + "spans": [ + { + "bbox": [ + 49, + 286, + 287, + 341 + ], + "type": "text", + "content": "[32] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerfplayer: A streamable dynamic scene representation with decomposed neural radiance fields. arXiv preprint arXiv:2210.15947, 2022. 3, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 342, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 342, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 49, + 342, + 287, + 396 + ], + "type": "text", + "content": "[33] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 399, + 287, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 399, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 49, + 399, + 287, + 431 + ], + "type": "text", + "content": "[34] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Improved direct voxel grid optimization for radiance fields reconstruction. arXiv preprint arXiv:2206.05085, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 434, + 287, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 434, + 287, + 476 + ], + "spans": [ + { + "bbox": [ + 49, + 434, + 287, + 476 + ], + "type": "text", + "content": "[35] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528, 2023. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 478, + 287, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 478, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 49, + 478, + 287, + 542 + ], + "type": "text", + "content": "[36] Liwen Wu, Jae Yong Lee, Anand Bhattachad, Yu-Xiong Wang, and David Forsyth. Diver: Real-time and accurate neural radiance fields with deterministic integration for volume rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16200-16209, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 544, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 544, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 49, + 544, + 287, + 599 + ], + "type": "text", + "content": "[37] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9421-9431, 2021. 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 601, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 601, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 49, + 601, + 287, + 644 + ], + "type": "text", + "content": "[38] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101, 2023. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 49, + 647, + 287, + 689 + ], + "type": "text", + "content": "[39] Wang Yifan, Felice Serena, Shihao Wu, Cengiz Öz Tireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM Transactions on Graphics (TOG), 38(6):1-14, 2019. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 692, + 287, + 713 + ], + "type": "text", + "content": "[40] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 185 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "text", + "content": "of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5752-5761, 2021. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "text", + "content": "[41] Qiang Zhang, Seung-Hwan Baek, Szymon Rusinkiewicz, and Felix Heide. Differentiable point-based radiance fields for efficient view synthesis. In SIGGRAPH Asia 2022 Conference Papers, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "type": "text", + "content": "[42] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Ewa volume splatting. In Proceedings Visualization, 2001. VIS'01., pages 29-538. IEEE, 2001. 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21145" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/GaussianAvatar_ Towards Realistic Human Avatar Modeling from a Single Video via Animatable 3D Gaussians/30b3a92a-861e-47e6-8246-2e90d1765e38_content_list.json b/2024/GaussianAvatar_ Towards Realistic Human Avatar Modeling from a Single Video via Animatable 3D Gaussians/30b3a92a-861e-47e6-8246-2e90d1765e38_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a2b74f4b518e15b15bbfffd18f5db8a46282fb7c --- /dev/null +++ b/2024/GaussianAvatar_ Towards Realistic Human Avatar Modeling from a Single Video via Animatable 3D Gaussians/30b3a92a-861e-47e6-8246-2e90d1765e38_content_list.json @@ -0,0 +1,1603 @@ +[ + { + "type": "text", + "text": "GaussianAvatar: Towards Realistic Human Avatar Modeling from a Single Video via Animatable 3D Gaussians", + "text_level": 1, + "bbox": [ + 176, + 130, + 794, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Liangxiao $\\mathrm{Hu}^{\\dagger,1}$ , Hongwen Zhang $^{2}$ , Yuxiang Zhang $^{3}$ , Boyao Zhou $^{3}$ , Boning Liu $^{3}$ , Shengping Zhang $^{*,1,4}$ , Liqiang Nie $^{1}$", + "bbox": [ + 161, + 202, + 805, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Harbin Institute of Technology 2Beijing Normal University", + "bbox": [ + 248, + 239, + 720, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Tsinghua University $^{4}$ Peng Cheng Laboratory", + "bbox": [ + 300, + 256, + 669, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{lx.hu, s.zhang}@hit.edu.cn, zhanghongwen@bnu.edu.cn, yx-z19@mails.tsinghua.edu.cn", + "bbox": [ + 122, + 276, + 844, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{bzhou22, liuboning}@mail.tsinghua.edu.cn, nieliqiang@gmail.com", + "bbox": [ + 202, + 295, + 758, + 309 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e78040bbddb2f4b799711aeaf5810b13030766248e31885acd533c979589b050.jpg", + "image_caption": [ + "Figure 1. We propose GaussianAvatar, which learns animatable 3D Gaussians to represent detailed human avatars from a single video. Our method maintains a 3D consistent appearance even when animated by out-of-distribution motions." + ], + "image_footnote": [], + "bbox": [ + 102, + 359, + 851, + 589 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 646, + 313, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present GaussianAvatar, an efficient approach to creating realistic human avatars with dynamic 3D appearances from a single video. We start by introducing animatable 3D Gaussians to explicitly represent humans in various poses and clothing styles. Such an explicit and animatable representation can fuse 3D appearances more efficiently and consistently from 2D observations. Our representation is further augmented with dynamic properties to support pose-dependent appearance modeling, where a dynamic appearance network along with an estimizable feature tensor is designed to learn the motion-to-appearance mapping. Moreover, by leveraging the differentiable motion", + "bbox": [ + 75, + 679, + 470, + 861 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "condition, our method enables a joint optimization of motions and appearances during avatar modeling, which helps to tackle the long-standing issue of inaccurate motion estimation in monocular settings. The efficacy of GaussianAvatar is validated on both the public dataset and our collected dataset, demonstrating its superior performances in terms of appearance quality and rendering efficiency. The code and dataset are available at https://github.com/aipixel/GaussianAvatar.", + "bbox": [ + 500, + 648, + 890, + 782 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 814, + 632, + 830 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Creating a customized human avatar from a single video has great potential for many applications including virtual and augmented reality, the Metaverse, gaming, and movie industries. This task is appealing yet challenging, as the", + "bbox": [ + 498, + 839, + 892, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Work done during an internship at Tsinghua University.", + "bbox": [ + 99, + 875, + 405, + 888 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author.", + "bbox": [ + 99, + 888, + 233, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "634", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "monocular observations are highly underdetermined for the modeling of a 3D animatable avatar. Meanwhile, the inaccurate body motion estimations and the complex wrinkle deformations also make it extremely difficult to create a realistic avatar from a single video.", + "bbox": [ + 75, + 90, + 470, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The modeling of 3D human avatars from monocular videos involves a fusion process of 2D observations to a 3D consistent human model. For this purpose, existing solutions have leveraged both implicit and explicit representations to create the base model of avatars. Among them, implicit methods [6, 15, 16, 49, 57] define a deformable human NeRF to fuse the image observation from current motion space to a canonical space by inverse skinning. However, the implicit 3D volume is inefficient in representing human surfaces and the inverse skinning also introduces ambiguous correspondences during the fusion process. These issues make it hard for implicit solutions to capture fine-grained details of moving people. As the avatar appearances are concentrated around human surfaces, explicit representations are much more efficient in modeling avatars. Previous attempts [12, 63] have employed differentiable mesh rendering to reconstruct the human surface, but these methods struggle to preserve wrinkle details due to a fixed mesh topology. On the other hand, point-based representations [18, 37, 50, 56, 65] are more effective to handle flexible topology but require millions of points to capture detailed appearances. How to represent humans remains one of the fundamental problems for avatar modeling.", + "bbox": [ + 75, + 169, + 472, + 517 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "When creating an animatable avatar from a video of moving people, the algorithm is required to learn the relationships between body motions and corresponding appearances. However, the motions estimated from monocular videos are typically faulty, leading to large artifacts in the modeling of dynamic cloth deformations. To address this issue, previous works [1, 15, 16, 42, 45, 49, 57] have attempted to optimize body motions along with the learning of animatable avatar volumes. As the body motions are explicitly represented as parametric meshes, the implicit 3D volume of previous methods makes the optimization indirect and less effective. This issue has even become the main obstacle to achieving high-quality avatar modeling from monocular videos.", + "bbox": [ + 75, + 518, + 472, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle the above issues, we introduce new representations and solutions to achieve high-quality avatar modeling from a single video. The key insight of our solution is to model dynamic human surfaces explicitly and optimize both the motion and appearances jointly in an end-to-end manner. To this end, we propose GaussianAvatar, a method to reconstruct human avatars with dynamic appearances using the 3D Gaussian representation [17]. As an explicit representation, 3D Gaussian can be easily posed from the canonical space to the motion space via a forward skinning process. Such an animatable 3D Gaus", + "bbox": [ + 75, + 734, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "sian representation bypasses the inverse skinning process used in the aforementioned NeRF-based methods and overcomes the previous one-to-many issue [4] during canonicalization. Based on the animatable 3D Gaussian, our method can fuse 3D appearances more consistently from 2D observations to a canonical 3D space. To model dynamic human appearances under different poses, we additionally add pose-dependent properties to 3D Gaussians and incorporate them with the canonical human surfaces. Inspired by previous work [28, 46], we learn a dynamic appearance network on the 2D manifolds of the underlying human mesh to predict dynamic properties of 3D Gaussians. However, due to the strong bias of limited training poses, modeling dynamic appearance solely conditioned on pose information struggles to generalize to novel views and poses. To address this issue, we introduce an estimizable feature tensor to capture a coarse global appearance of human avatars. Subsequently, we incorporate pose-dependent effects on the feature tensor to decode fine-grained details such as wrinkles.", + "bbox": [ + 496, + 90, + 893, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As the proposed animatable 3D Gaussians are differentiable with respect to the motion conditions, it enables a joint optimization of motion and appearances. This merit allows our network to refine the motion along with the avatar modeling process, which helps to tackle the long-standing issue of inaccurate motion estimation in monocular settings. Besides, the refined motion further enhances the accuracy of avatar modeling since the 3D appearance fusion process relies on motion-based skinning. As shown in our experiments, our method is quite robust to initial motion estimation and has the ability to correct the misalignment of motion capture results.", + "bbox": [ + 496, + 378, + 893, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our main contributions are as follows:", + "bbox": [ + 517, + 569, + 872, + 583 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce animatable 3D Gaussians for realistic human avatar modeling from a single video. By representing human surfaces explicitly, our method can fuse 3D appearances more consistently and efficiently from 2D observations.", + "- We augment the animatable 3D Gaussians with dynamic properties to support pose-dependent appearance modeling, where a dynamic appearance network along with an estimizable feature tensor is designed to learn the motion-to-appearance mapping.", + "- We propose to jointly optimize the motion and appearance during the avatar modeling, enabling our method to correct the misalignment of initial motion and improve the final appearance quality." + ], + "bbox": [ + 500, + 587, + 890, + 797 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 814, + 640, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural Rendering for Human Reconstruction. Without the need to define a template mesh for avatar modeling [2, 9, 53], neural rendering has emerged as a potent technique that enables learning avatars directly from im", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "635", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ages. Here we briefly review precious work that aims to reconstruct humans using neural rendering.", + "bbox": [ + 75, + 90, + 468, + 119 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Due to the high-quality rendering of neural radiance field [29], various efforts [20, 21, 24, 32, 33, 47, 67, 68] have been made to reconstruct the dynamic appearance of moving people. Neural Body [33] associates a latent code to each SMPL [26] vertex to encode the appearance, which is transformed into observation space based on the human pose. Neural Actor [24] learns a deformable radiance field with SMPL as guidance and utilizes a texture map to improve its final rendering quality. TAVA [20] proposes to jointly model the non-rigid warping field and shading effects directly conditioned on the pose vectors. Posevocab [21] designs joint-structured pose embeddings to encode the dynamic appearances under different key poses, such embeddings can better learn joint-related appearance. NeRF-based methods have demonstrated appealing rendering results on human avatar reconstruction, but still struggle to represent human surfaces with the implicit 3D volume. Explicit modeling of human surfaces is a more straightforward way for this task, as the dynamic appearance of humans is mostly reflected on the human surface.", + "bbox": [ + 75, + 122, + 468, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Explicit representations have great potential for human reconstruction. In HF-Avatar [63], meshes are used as the base representation in a coarse-to-fine framework that combines neural texture with dynamic surface deformation for avatar creation. EMA [12] proposes Meshy neural fields to reconstruct human avatars by optimizing the canonical mesh, material, and motion dynamics through inverse rendering in an end-to-end process. PointAvatar [65] employs a deformable point-based representation to separate source color into intrinsic albedo and normal-dependent shading. DVA [35] extends mixtures of volumetric primitives [25] for human avatar modeling. All these attempts have demonstrated the significant potential of explicit representations and their under-exploration. However, meshes are constrained by fixed topologies, whereas point clouds demand a multitude of points to encompass intricate details. 3D Gaussians [17] have showcased their capability in various human tasks [22, 40, 55, 64]. In essence, 3D Gaussians hold promise for human avatar reconstruction and are currently a subject of active research.", + "bbox": [ + 75, + 425, + 468, + 727 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Avatar Modeling from Monocular Videos. Numerous methods investigate reconstructing humans from single images or monocular videos. Regression-based methods[10, 11, 13, 38, 39, 51, 52, 66] directly recover clothed 3D humans just from a single image. While these methods produce attractive results, they can not recover the dynamic appearance in the reconstruction across the entire sequence. Traditional methods aim to capture human dynamics by tracking individuals in videos using pre-scanned rigged templates [7, 8, 54]. However, the pre-scanning", + "bbox": [ + 75, + 750, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and manual rigging processes prevent their real-life applications. [1, 5, 30] try to bypass the requirement for predefined human models but face challenges in preserving fine details because of the fixed mesh resolution.", + "bbox": [ + 496, + 90, + 890, + 150 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The emergence of neural radiance fields [29] has facilitated the creation of various techniques for reconstructing animatable avatars from monocular videos [3, 14, 16, 41-43, 49]. However, inaccuracies in estimating human motions from monocular videos result in pronounced artifacts. To address the problem, HumanNeRF [49] solves for an update to the inaccurate poses. NeuMan [16] introduces an error-correction network to enable training with erroneous estimates. Vid2Avatar[6] and InstantAvatar [15] also jointly optimize motions by back-propagating the gradient of the image reconstruction loss to the pose parameters. MonoHuman [57] introduces bi-directional constraints to alleviate ambiguous correspondence on novel poses. Attempting to address the inaccurate pose estimation issue with implicit human NeRF models has proven to be inefficient and imprecise. Consequently, we endeavor to address this issue by leveraging an explicit 3D Gaussian representation.", + "bbox": [ + 496, + 152, + 892, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 420, + 591, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to create a human avatar that enables free-viewpoint rendering and realistic animation using a single video. We first introduce an expressive representation, namely animatable 3D Gaussians, to represent human avatars in Sec. 3.1. With this representation, modeling the dynamic appearance of humans can be regarded as dynamic 3D Gaussian property estimation. Then we build a dynamic appearance network along with an estimable feature tensor to learn the motion-to-appearance mapping in Sec. 3.2. To mitigate the artifacts caused by inaccurate pose estimation, we adopt a joint motion and appearance optimization approach for refining human poses during training (Sec. 3.3). In the following, we provide a comprehensive explanation of the technical details.", + "bbox": [ + 496, + 446, + 890, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Animatable 3D Gaussians", + "text_level": 1, + "bbox": [ + 500, + 666, + 735, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Point-based representation [23, 28, 34, 59, 69] has proven its topological flexibility in generating realistic human avatars from scans. Extending this explicit representation to create human avatars from images is a significant endeavor. With this goal in mind, we introduce a novel representation, termed animatable 3D Gaussians, which effectively reconstructs human surfaces with a 3D consistent appearance.", + "bbox": [ + 496, + 688, + 890, + 795 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Gaussian Splitting [17] is a point-based scene representation that allows high-quality real-time rendering. The scene representation is parameterized by a set of static 3D Gaussians, each of which has the following parameters: 3D center position $\\mathbf{x} \\in \\mathbb{R}^3$ , color $\\mathbf{c} \\in \\mathbb{R}^3$ , opacity $\\alpha \\in \\mathbb{R}$ , 3D rotation in form of quaternion $\\mathbf{q} \\in \\mathbb{R}^4$ and 3D scaling factor $\\mathbf{s} \\in \\mathbb{R}^3$ . With these properties, we can generate rendered", + "bbox": [ + 496, + 796, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "636", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/26330d876fda00891ab4b8a45025625ad03c508ee6cef6ff63f263f78a64a74b.jpg", + "image_caption": [ + "Figure 2. Overview of GaussianAvatar. Given a fitted SMPL or SMPL-X model on the current frame, we sample the points on its surface and record their positions on a UV positional map $I$ , which is then passed to a pose encoder to obtain the pose feature. An estimizable feature tensor is pixel-aligned with the pose feature and learned to capture the coarse appearance of humans. Then the two aligned feature tensors are input into the Gaussian parameter decoder, which predicts each point's offset $\\Delta \\hat{\\mathbf{x}}$ , color $\\hat{\\mathbf{c}}$ , and scale $\\hat{s}$ . These predictions, along with the fixed rotations $\\mathbf{q}$ and opacity $\\alpha$ , collectively constitute the animatable 3D Gaussians in canonical space." + ], + "image_footnote": [], + "bbox": [ + 83, + 90, + 890, + 262 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "images from any viewpoint. We refer readers to [17, 27] for the rendering details of 3D Gaussians.", + "bbox": [ + 75, + 358, + 470, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To extend this representation for human avatar modeling, we integrate it with either the SMPL [26] or SMPL-X [31] model as follows:", + "bbox": [ + 75, + 388, + 470, + 434 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nG (\\pmb {\\beta}, \\pmb {\\theta}, \\mathbf {D}, \\mathbf {P}) = S p l a t t i n g (W (\\mathbf {D}, J (\\pmb {\\beta}), \\pmb {\\theta}, \\omega), \\mathbf {P}), (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 446, + 468, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $G(\\cdot)$ represents a rendered image, and $Splatting(\\cdot)$ denotes the rendering process of 3D Gaussians from any viewpoint, $W(\\cdot)$ is a standard linear blend skinning function employed for reposing 3D Gaussians, $\\mathbf{D} = T(\\boldsymbol{\\beta}) + dT$ represents the locations of 3D Gaussians in canonical space, formed by adding corrective point displacements $dT$ on the template mesh surface $T(\\beta)$ , $\\mathbf{P}$ denotes the remaining properties of 3D Gaussians, excluding the positions. $\\beta$ and $\\theta$ are the shape and pose parameters, $J(\\beta)$ outputs 3D joint locations. Note that we propagate the skinning weight $\\omega$ from the vertices of the SMPL or SMPL-X model to the nearest 3D Gaussians. With the proposed representation, we can now repose these canonical 3D Gaussians to the motion space for free-view rendering.", + "bbox": [ + 75, + 474, + 472, + 690 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Dynamic 3D Gaussian Property Estimation", + "text_level": 1, + "bbox": [ + 76, + 696, + 447, + 713 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following the proposed animatable Gaussians, human appearances are determined by the point displacements $dT$ and properties $\\mathbf{P}$ . Modeling dynamic human appearances can be regarded as estimating these dynamic properties. To model dynamic human appearances under various poses, we introduce a dynamic appearance network along with an estimizable feature tensor to predict these pose-dependent properties of 3D Gaussians. Despite sharing similarities in network structure with [28], we present this framework for a distinct purpose. In [28], the feature tensor serves to decouple the pose-independent human shape from the decoder, and we repurpose it to capture a coarse global appearance", + "bbox": [ + 75, + 719, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/460762fcaac45792567d669584836768ed1f08403826bda9b989f98bccf7a450.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 506, + 358, + 586, + 496 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/504c07309e5eb4d0857085dea30ea95c4095228d8fbbe4ac05908ea4f34a72f1.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 589, + 358, + 650, + 497 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/19a40eee079f096e44a7e5d285e4cf533fd72d2157fea90a5023c365cc358594.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 656, + 358, + 733, + 497 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6155fdef5b9e0184ee5365daa6468f4f474dfdfa0965aedaf128dd841b127a6a.jpg", + "image_caption": [ + "(d)", + "Figure 3. Effect of iostropy of 3D Gaussians. (a) Input image, (b)(d) front and back views trained with isotropic 3D Gaussians, (c)(e) front and back views trained with anisotropic 3D Gaussians." + ], + "image_footnote": [], + "bbox": [ + 736, + 358, + 805, + 497 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e2785d0a2f35a1401d0267206f2370467c2bb39c30f6d4534495c815d5808177.jpg", + "image_caption": [ + "(e)" + ], + "image_footnote": [], + "bbox": [ + 813, + 358, + 885, + 497 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of human avatars. The motivation behind this modification is that directly learning a mapping from human poses to dynamic properties is susceptible to overfitting on the limited training poses. To integrate the global appearance into the feature tensor, we introduce a two-stage training strategy, as discussed in Sec. 3.4.", + "bbox": [ + 496, + 584, + 893, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The dynamic appearance network is designed to learn a mapping from a 2D manifold representing the underlying human shape to the dynamic properties of 3D Gaussians as follows:", + "bbox": [ + 496, + 676, + 893, + 734 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\phi}: \\mathcal {S} ^ {2} \\in \\mathbb {R} ^ {3} \\rightarrow \\mathbb {R} ^ {7}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 737, + 890, + 755 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 2, the 2D human manifold $S^2$ is depicted by a UV positional map $I \\in R^{H \\times W \\times 3}$ , where each valid pixel stores the position $(x, y, z)$ of one point on the posed body surface. The final predictions consist of per point offset $\\Delta \\hat{\\mathbf{x}} \\in \\mathbb{R}^3$ , color $\\hat{\\mathbf{c}} \\in \\mathbb{R}^3$ , and scale $\\hat{s} \\in \\mathbb{R}$ on the canonical surface.", + "bbox": [ + 496, + 763, + 893, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Instead of predicting all properties, we make these slight adjustments to our task based on experimental results. Due to the unbalanced viewpoints in monocular videos,", + "bbox": [ + 496, + 854, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "637", + "bbox": [ + 486, + 944, + 511, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "anisotropic 3D Gaussians are prone to learning an inaccurate 3D shape to fit the most frequently seen view, resulting in poorer performance in side views. As shown in Fig. 3, training our network with isotropic 3D Gaussians yields superior results. We ensure isotropy among all 3D Gaussians by maintaining uniform size across dimensions. Therefore, we introduce a scalar $\\hat{s}$ to represent the scales of 3D Gaussians and set rotations $\\mathbf{q}$ as $[1,0,0,0]$ . Through experimental observations, we notice that the network tends to learn an opacity value $(\\alpha)$ of zero on the boundary to correct the human shape. To address this, we fix opacity $\\alpha = 1$ to keep all 3D Gaussians visible, enforcing the network to predict accurate positions of the 3D Gaussians.", + "bbox": [ + 75, + 90, + 472, + 286 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The dynamic appearance network consists of two parts: a pose encoder and a Gaussian parameter decoder. The pose encoder takes the UV positional map of posed body points as input to generate a pose-conditioned feature tensor $O \\in \\mathbb{R}^{H \\times W \\times C}$ . We then integrate the pixel-aligned optimized feature tensor $F \\in \\mathbb{R}^{H \\times W \\times C}$ with the pose features before feeding it into the Gaussian parameter decoder to generate final predictions. Following this, we add the predicted offsets to the canonical 3D Gaussians and associate the predicted properties with the corresponding 3D Gaussians. With the estimated poses, we can then repose the canonical 3D Gaussians to the motion space for rendering.", + "bbox": [ + 75, + 287, + 472, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Joint Motion and Appearance Optimization", + "text_level": 1, + "bbox": [ + 76, + 477, + 449, + 493 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, owing to the imprecise estimation of human poses $\\pmb{\\theta} = (\\theta_{1},\\theta_{2},\\dots,\\theta_{J})$ and translations $\\mathbf{t}$ from monocular videos, the reposed 3D Gaussians in motion space are inaccurately represented and may lead to unsatisfactory rendering outcomes. To address this, we propose to jointly optimize human motions and appearances. To optimize human motions with image loss, we solve for an update $(\\Delta \\pmb {\\theta},\\Delta \\mathbf{t})$ to the estimated body poses and translations as follows:", + "bbox": [ + 75, + 500, + 468, + 621 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {\\Theta}} = (\\boldsymbol {\\theta} + \\Delta \\boldsymbol {\\theta}, \\mathbf {t} + \\Delta \\mathbf {t}). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 628, + 468, + 647 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We modify $\\theta$ in Eq. 1 using $\\Theta$ to render the proposed animatable 3D Gaussians differentiable with respect to the motion conditions. Different from previous work [6, 16, 49] that jointly optimizes human poses via inverse skinning, we optimize the updates in a forward skinning process, which benefits both motion and appearance optimization.", + "bbox": [ + 75, + 657, + 468, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Training Strategy", + "text_level": 1, + "bbox": [ + 76, + 756, + 250, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we outline our approach to training the network with inaccurate human motions. We conduct a two-stage optimization process using different loss functions. In the first stage, we aim to fuse the sequential appearances to the estimizable feature tensor and conduct motion optimization to get accurate poses for the dynamic appearance network. In this stage, we optimize the framework without incorporating any pose-dependent information by excluding", + "bbox": [ + 75, + 780, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the training of the pose encoder. Specifically, we utilize the following loss functions to train our network:", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {s t a g e} _ {1}} = \\lambda_ {r b g} \\mathcal {L} _ {r b g} + \\lambda_ {s s i m} \\mathcal {L} _ {s s i m} + \\lambda_ {l p i p s} \\mathcal {L} _ {l p i p s} \\tag {4} \\\\ + \\lambda_ {f} \\mathcal {L} _ {f} + \\lambda_ {\\text {o f f s e t}} \\mathcal {L} _ {\\text {o f f s e t}} + \\lambda_ {\\text {s c a l e}} \\mathcal {L} _ {\\text {s c a l e}}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 137, + 890, + 172 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{rbg}$ , $\\mathcal{L}_{ssim}$ , and $\\mathcal{L}_{lpips}$ are the L1 loss, SSIM loss [48], and LPIPS loss [61], respectively. $\\mathcal{L}_f$ , $\\mathcal{L}_{offset}$ , $\\mathcal{L}_{scale}$ calculate the L2-norm of the feature map, predicted offsets and scales, respectively. We set $\\lambda_{rbg} = 0.8$ , $\\lambda_{ssim} = 0.2$ , $\\lambda_{lpips} = 0.2$ , $\\lambda_f = 1$ , $\\lambda_{offset} = 10$ , $\\lambda_{scale} = 1$ .", + "bbox": [ + 498, + 186, + 890, + 263 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After the first stage of training, we obtain more accurate human motions and an optimized feature tensor $F$ . The optimized feature tensor $F$ captures a coarse appearance of human avatars. In the second stage, we incorporate the pose features encoded by the pose encoder with the trained feature tensor $F$ . We replace $\\mathcal{L}_f$ with the L2-norm loss $\\mathcal{L}_p$ , which plays the same role as $\\mathcal{L}_f$ in regularizing the limited pose space. By penalizing the pose-dependent features, we can eliminate the strong bias of limited training poses and thus generalize to unseen viewpoints and poses.", + "bbox": [ + 496, + 265, + 892, + 416 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 434, + 633, + 452 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and Metrics", + "text_level": 1, + "bbox": [ + 500, + 460, + 697, + 477 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "People-Snapshot Dataset. This dataset [1] comprises videos of individuals rotating in front of a stationary camera. To ensure a fair quantitative comparison, we follow the same evaluation protocol outlined in InstantAvatar [15].", + "bbox": [ + 498, + 487, + 890, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "NeuMan Dataset. To assess even more challenging scenarios, we employ outdoor collections from the NeuMan dataset [16]. These videos are recorded using a mobile phone for moving individuals. Specifically, we select four sequences (bike, citizen, Seattle, jogging) that exhibit most body regions and contain minimal blurry images. We initialize the estimated poses with an off-the-shelf method [44], which is also utilized in [15].", + "bbox": [ + 496, + 549, + 890, + 670 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "DynVideo Dataset. With the limited cloth deformation presented in the above two datasets, we propose the Dynvideo dataset to capture dynamic human appearance. To this end, we utilize a mobile phone to record videos of a character performing various movements, especially rotation, in front of the device. Each of these videos takes about one minute and provides a comprehensive and detailed representation of human movement. We also provide the corresponding SMPL parameter sequences for all videos, obtained by our proposed method. This dataset serves as a valuable resource for evaluating reconstruction quality, with a particular emphasis on dynamic appearances.", + "bbox": [ + 496, + 671, + 890, + 853 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. We consider three metrics: PSNR, SSIM [48], and LPIPS [61] to access the reconstruction quality on three datasets.", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "638", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/a475419c29254a8096fff7388fb327eaa7b75d1bb1dd933f60539cd2fd89fb25.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Methodmale-3-casualmale-4-casualfemale-3-casualfemale-4-casual
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
HumanNeRF [49]26.900.96050.018125.500.93970.035724.460.95160.026927.070.96150.0152
InstantAvatar [15]29.530.97160.015527.670.96260.030727.660.97090.021029.110.96830.0167
Baseline27.710.97130.021825.090.96140.030624.800.96090.031026.160.96100.0227
Baseline + Opt.30.980.97910.014328.770.97530.023029.590.97680.022030.830.97680.0145
Baseline + Opt. + Dyn.30.980.97900.014528.780.97550.022829.550.97620.022530.840.97710.0140
", + "bbox": [ + 81, + 88, + 893, + 200 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5280fc7045287bd9fc45e2542bbd0594fb9279e03b59180fd49a3a1ef7ca8740.jpg", + "table_caption": [ + "Table 1. Quantitative evaluation on the People-Snapshot [1] dataset. Opt. denotes the motion optimization and Dyn. refers to dynamic appearance modeling." + ], + "table_footnote": [], + "table_body": "
MethodPSNR ↑SSIM ↑LPIPS ↓
HumanNeRF [49]27.060.96690.0192
InstantAvatar [15]28.470.97150.0277
Baseline27.060.96920.0183
Baseline + Opt.29.930.97940.0126
Baseline + Opt. + Dyn.29.940.97950.0124
", + "bbox": [ + 91, + 252, + 455, + 343 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8c9f7f6ead0584698e210df7c047eec3735a59536e3de3e27119d92bb43b4ff5.jpg", + "table_caption": [ + "Table 2. Quantitative evaluation on NeuMan [16] dataset." + ], + "table_footnote": [], + "table_body": "
MethodPSNR ↑SSIM ↑LPIPS ↓
HumanNeRF [49]24.390.93490.0349
InstantAvatar [15]20.310.91830.1046
Baseline23.550.93400.0354
Baseline + Opt.26.160.95220.0237
Baseline + Opt. + Dyn.28.580.96160.0169
", + "bbox": [ + 91, + 382, + 455, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Quantitative evaluation on DynVideo dataset.", + "bbox": [ + 102, + 481, + 442, + 496 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/53a69b099812bce6361a1a404aedc0d370eee4bf278ca822a7380654c7ab654e.jpg", + "image_caption": [ + "(a)", + "Figure 4. Motion optimization results. (a)(d) Original image, (b)(e) our optimized SMPL, (c)(f) ROMP [44] estimates." + ], + "image_footnote": [], + "bbox": [ + 84, + 513, + 148, + 669 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/efd5c94998dee1d0fe52f5c704e4f0e6d22feb6dd7cb275b820ab8da6000209b.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 148, + 513, + 209, + 667 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e913b401eb4b088d37fc059217f7dde737b89f15cc2605a2ba4dfe51097849ac.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 210, + 513, + 272, + 667 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6c5cfc78e55dbcebbdfbbfe3ca57a466ea619cd89213370d6501cd7527fcd977.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 274, + 513, + 336, + 667 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7df6d01934acbddd2669f26b3ed4834657cf5e4d869545cc0f80a139848dd10a.jpg", + "image_caption": [ + "(e)" + ], + "image_footnote": [], + "bbox": [ + 338, + 513, + 398, + 667 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f387d007031e81449e82fc9f40b230bd1e8b8aea88bf0342dfdb45261aaef90b.jpg", + "image_caption": [ + "(f)" + ], + "image_footnote": [], + "bbox": [ + 400, + 513, + 460, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Experimental Settings", + "text_level": 1, + "bbox": [ + 76, + 753, + 284, + 768 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. We employ a U-Net [36] for extracting pose-dependent features, and the Gaussian parameter decoder is implemented as an 8-layer multilayer perceptron (MLP). We sample approximately 200,000 points on the SMPL mesh surface. The entire framework is trained on a single NVIDIA RTX 3090 GPU, with training times ranging from 0.5 to 6 hours.", + "bbox": [ + 75, + 777, + 468, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baseline. To showcase the efficacy of the proposed mod", + "bbox": [ + 76, + 885, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6a03ee5740839210e69ab2df20dfc0bc36461068aaa8f2ca98e4ad2a5f5b77fb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 253, + 694, + 368 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e617c07bc6d3c71539c702e09c6052177451ec63859c9d93a4c8934fdef87f95.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 509, + 371, + 602, + 479 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/474aa8c9485bc9fdb7e8fcd4d3eb1c0ccd7f3f84f9d607e0108bb13af96db098.jpg", + "image_caption": [ + "(b)", + "Figure 5. Qualitative ablation studies. (a) Ground truth, (b) baseline + Opt. + Dyn., (c) baseline + Opt., (d) baseline." + ], + "image_footnote": [], + "bbox": [ + 604, + 371, + 694, + 479 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ae64d062660dc7f762a9673ed9e20ea2eef4bd48db0accf99e0bf0fb46699caa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 255, + 785, + 368 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fb82564be95791f788df7d5b2ee4d9ff2d7f16ac0b1d81352ea14146fea9b511.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 697, + 371, + 787, + 479 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f9cf3cdef3710d38af1acc9f23a9387c7c7f50a73eb54546d54a68313878d137.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 789, + 255, + 880, + 368 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fe8ad7e17cf8f04981501fd9a8f6dd6bff1fab79f27b28aece4fefd9a1757966.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 790, + 371, + 880, + 479 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ules, our method can be partitioned into three components: baseline, motion optimization, and dynamic appearance modeling. As our baseline, we suspend the motion optimization process and exclude the pose-dependent information from the pose encoder.", + "bbox": [ + 496, + 551, + 890, + 626 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Methods for Comparison. We compare our method against (1) HumanNeRF [49], which implicitly represents a human avatar with a canonical appearance neural field and a motion field; (2) InstantAvatar [15], which achieves fast avatar modeling by using several acceleration strategies for neural fields.", + "bbox": [ + 496, + 626, + 890, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparisons with the State of the Art", + "text_level": 1, + "bbox": [ + 500, + 726, + 828, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We report the numeric evaluation results on three datasets. As shown in Table 1, 2, and 3, our proposed method outperforms all baselines on all metrics for recovering more details of dynamic appearance and correcting the artifacts caused by initialized poses. To demonstrate the qualitative evaluation of these datasets, we also visualize the novel view synthesis results on the test splits. As implicit representations, HumanNeRF and InstantAvatar are prone to generate ghosting effects at boundary areas, as shown in Fig. 6. InstantAvatar lacks the capability to model pose", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "639", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3307401eced9c5d32e29fda4fbbfa79b0db791157a873df09435bf499faec07c.jpg", + "image_caption": [ + "Figure 6. Qualitative comparison of novel view synthesis. We compare the novel view synthesis quality on the People-Snapshot dataset (first row), NeuMan dataset (second row), and DynVideo dataset (last two rows)." + ], + "image_footnote": [], + "bbox": [ + 84, + 95, + 519, + 738 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8e2a9f2db1de5d5c275a29e7874b1600f18cd2110e5e9ecb06c8aa7306e48897.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 531, + 95, + 874, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "dependent deformations, leading to challenges in effectively handling the dynamic appearance of moving people.", + "bbox": [ + 75, + 797, + 470, + 829 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To showcase the robustness of our avatar modeling, we collect more challenging poses [19] captured by the monocular camera and evaluate our method and other methods for avatar animation and novel view synthesis. As ground truth", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "is unavailable for these out-of-distribution poses, we illustrate qualitative results of novel pose and view synthesis in Fig. 1 and Fig. 7. Our method generates realistic animation results with these challenging poses, demonstrating a consistent 3D appearance in novel views with respect to other methods.", + "bbox": [ + 496, + 797, + 893, + 888 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "640", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/56a851efbb927e2ca51d324b297af18c839938f557224fde54d862a77573b7f5.jpg", + "image_caption": [ + "Figure 7. Animation results on out-of-distribution motions. We compare the animation results on the People-Snapshot dataset (first row), NeuMan dataset (second row), and DynVideo dataset (last row)." + ], + "image_footnote": [], + "bbox": [ + 89, + 95, + 883, + 440 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 508, + 240, + 522 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we conduct ablation studies to validate each component of our methods. As shown in Table 1, 2, and 3, our proposed motion optimization module dramatically improves over the baseline on all metrics, demonstrating its effectiveness in modeling human avatars. To highlight the effectiveness of our method in human motion optimization, we illustrate the initialized poses obtained by ROMP [44] and the optimized ones by our method on NeuMan and DynVideo datasets in Fig. 4. Experiments demonstrate that joint motion optimization is capable of correcting inaccurate motion estimation, even for the side and back views. Furthermore, our approach readily extends to enhance the accuracy of existing motion capture methods [58, 60, 62].", + "bbox": [ + 75, + 534, + 468, + 731 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also notice that the dynamic appearance modeling achieves superior results on NeuMan and DynVideo datasets in Table 2 and 3 and slightly favorable outcomes in Table 1, considering that People-Snapshot has limited pose variation with respect to the other two datasets. In Fig. 5, we further illustrate the visible improvements of each component of our method. The motion optimization scheme improves the global quality of the baseline for better pose estimation. The dynamic appearance modeling additionally preserves pose-dependent details such as cloth wrinkles on the human surface.", + "bbox": [ + 75, + 734, + 470, + 898 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Discussion", + "text_level": 1, + "bbox": [ + 500, + 507, + 746, + 522 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce GaussianAvatar, a human avatar reconstruction method based on the proposed animatable 3D Gaussians from monocular videos. For dynamic human appearance modeling, we leverage a dynamic appearance network along with an estimizable feature tensor to enhance the representation with dynamic properties. Besides, we implement a joint motion and appearance optimization scheme to rectify estimated motion and enhance the overall reconstruction quality. Our method shows the capability to reconstruct avatars with dynamic appearances, enabling realistic animation while maintaining real-time rendering speed.", + "bbox": [ + 496, + 532, + 890, + 698 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitation. Similar to [15, 49, 57], our method may generate artifacts due to inaccurate foreground segmentations in videos and encounter challenges in modeling loose outfits such as dresses.", + "bbox": [ + 496, + 699, + 890, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Potential Social Impact. Given our method's capability to reconstruct a realistic personalized character from a monocular video, it is imperative to exercise caution and consider the potential for technology misuse.", + "bbox": [ + 496, + 758, + 890, + 819 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work was supported by the NSFC project (Nos. 62272134, 62236003, 62072141, 62301298, and 62125107), Shenzhen College Stability Support Plan (Grant No. GXWD20220817144428005), and the Major Key Project of PCL (PCL2023A10-2).", + "bbox": [ + 496, + 819, + 890, + 895 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "641", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Thiemo Alldieck, Marcus Magnor, Weipeng Xu, Christian Theobalt, and Gerard Pons-Moll. Video based reconstruction of 3d people models. In CVPR, 2018. 2, 3, 5, 6", + "[2] Timur Bagautdinov, Chenglei Wu, Tomas Simon, Fabian Prada, Takaaki Shiratori, Shih-En Wei, Weipeng Xu, Yaser Sheikh, and Jason Saragih. Driving-signal aware full-body avatars. ACM TOG, 2021. 2", + "[3] Jianchuan Chen, Ying Zhang, Di Kang, Xuefei Zhe, Linchao Bao, Xu Jia, and Huchuan Lu. Animatable neural radiance fields from monocular rgb videos. arXiv preprint arXiv:2106.13629, 2021. 3", + "[4] Xu Chen, Yufeng Zheng, Michael J Black, Otmar Hilliges, and Andreas Geiger. Snarf: Differentiable forward skinning for animating non-rigid neural implicit shapes. In ICCV, 2021. 2", + "[5] Chen Guo, Xu Chen, Jie Song, and Otmar Hilliges. Human performance capture from monocular video in the wild. In 3DV, 2021. 3", + "[6] Chen Guo, Tianjian Jiang, Xu Chen, Jie Song, and Otmar Hilliges. Vid2 avatar: 3d avatar reconstruction from videos in the wild via self-supervised scene decomposition. In CVPR, 2023. 2, 3, 5", + "[7] Marc Habermann, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Livecap: Real-time human performance capture from monocular video. ACM TOG, 2019. 3", + "[8] Marc Habermann, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Deepcap: Monocular human performance capture using weak supervision. In CVPR, 2020. 3", + "[9] Marc Habermann, Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Real-time deep dynamic characters. ACM TOG, 2021. 2", + "[10] Tong He, John Collomosse, Hailin Jin, and Stefano Soatto. Geo-pifu: Geometry and pixel aligned implicit functions for single-view human reconstruction. In NeurIPS, 2020. 3", + "[11] Tong He, Yuanlu Xu, Shunsuke Saito, Stefano Soatto, and Tony Tung. Arch++: Animation-ready clothed human reconstruction revisited. In ICCV, 2021. 3", + "[12] Xiaoke Huang, Yiji Cheng, Yansong Tang, Xiu Li, Jie Zhou, and Jiwen Lu. Efficient meshy neural fields for animatable human avatars. arXiv preprint arXiv:2303.12965, 2023. 2, 3", + "[13] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. Arch: Animatable reconstruction of clothed humans. In CVPR, 2020. 3", + "[14] Boyi Jiang, Yang Hong, Hujun Bao, and Juyong Zhang. Self-con: Self reconstruction your digital avatar from monocular video. In CVPR, 2022. 3", + "[15] Tianjian Jiang, Xu Chen, Jie Song, and Otmar Hilliges. Instantavatar: Learning avatars from monocular video in 60 seconds. In CVPR, 2023. 2, 3, 5, 6, 8", + "[16] Wei Jiang, Kwang Moo Yi, Golnoosh Samei, Oncel Tuzel, and Anurag Ranjan. Neuman: Neural human radiance field from a single video. In ECCV, 2022. 2, 3, 5, 6" + ], + "bbox": [ + 78, + 114, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM TOG, 2023. 2, 3, 4", + "[18] Christoph Lassner and Michael Zollhofer. Pulsar: Efficient sphere-based neural rendering. In CVPR, 2021. 2", + "[19] Ruilong Li, Shan Yang, David A. Ross, and Angjoo Kanazawa. Learn to dance with aist++: Music conditioned 3d dance generation. In ICCV, 2021. 7", + "[20] Ruilong Li, Julian Tanke, Minh Vo, Michael Zollhöfer, Jürgen Gall, Angjoo Kanazawa, and Christoph Lassner. Tava: Template-free animatable volumetric actors. In ECCV, 2022. 3", + "[21] Zhe Li, Zerong Zheng, Yuxiao Liu, Boyao Zhou, and Yebin Liu. Posevocab: Learning joint-structured pose embeddings for human avatar modeling. In SIGGRAPH, 2023. 3", + "[22] Zhe Li, Zerong Zheng, Lizhen Wang, and Yebin Liu. Animatable gaussians: Learning pose-dependent gaussian maps for high-fidelity human avatar modeling. In CVPR, 2024. 3", + "[23] Siyou Lin, Hongwen Zhang, Zerong Zheng, Ruizhi Shao, and Yebin Liu. Learning implicit templates for point-based clothed human modeling. In ECCV, 2022. 3", + "[24] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. ACM TOG, 2021. 3", + "[25] Stephen Lombardi, Tomas Simon, Gabriel Schwartz, Michael Zollhoefer, Yaser Sheikh, and Jason Saragih. Mixture of volumetric primitives for efficient neural rendering. ACM Trans. Graph., 2021. 3", + "[26] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. ACM TOG, 2015. 3, 4", + "[27] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. 3DV, 2024. 4", + "[28] Qianli Ma, Jinlong Yang, Siyu Tang, and Michael J Black. The power of points for modeling humans in clothing. In ICCV, 2021. 2, 3, 4", + "[29] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 3", + "[30] Gyeongsik Moon, Hyeongjin Nam, Takaaki Shiratori, and Young Mu Lee. 3d clothed human reconstruction in the wild. In ECCV, 2022. 3", + "[31] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3d hands, face, and body from a single image. In CVPR, 2019. 4", + "[32] Sida Peng, Shangzhan Zhang, Zhen Xu, Chen Geng, Boyi Jiang, Hujun Bao, and Xiaowei Zhou. Animatable neural implicit surfaces for creating avatars from videos. In ICCV, 2021. 3", + "[33] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In CVPR, 2021. 3" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "642", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Sergey Prokudin, Qianli Ma, Maxime Raafat, Julien Valentin, and Siyu Tang. Dynamic point fields. In ICCV, 2023. 3", + "[35] Edoardo Remelli, Timur Bagautdinov, Shunsuke Saito, Chenglei Wu, Tomas Simon, Shih-En Wei, Kaiwen Guo, Zhe Cao, Fabian Prada, Jason Saragih, et al. Drivable volumetric avatars using texel-aligned features. In SIGGRAPH, 2022. 3", + "[36] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 6", + "[37] Darius Rückert, Linus Franke, and Marc Stamminger. Adop: Approximate differentiable one-pixel point rendering. ACM TOG, 2022. 2", + "[38] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In ICCV, 2019. 3", + "[39] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In CVPR, 2020. 3", + "[40] Ruizhi Shao, Jingxiang Sun, Cheng Peng, Zerong Zheng, Boyao Zhou, Hongwen Zhang, and Yebin Liu. Control4d: Efficient 4d portrait editing with text. In CVPR, 2024. 3", + "[41] Shih-Yang Su, Frank Yu, Michael Zollhöfer, and Helge Rhodin. A-nerf: Articulated neural radiance fields for learning human shape, appearance, and pose. In NeurIPS, 2021. 3", + "[42] Shih-Yang Su, Timur Bagautdinov, and Helge Rhodin. Danbo: Disentangled articulated neural body representations via graph neural networks. In ECCV, 2022. 2", + "[43] Shih-Yang Su, Timur Bagautdinov, and Helge Rhodin. Npc: Neural point characters from video. In ICCV, 2023. 3", + "[44] Yu Sun, Qian Bao, Wu Liu, Yili Fu, Michael J Black, and Tao Mei. Monocular, one-stage, regression of multiple 3d people. In ICCV, 2021. 5, 6, 8", + "[45] Gusi Te, Xiu Li, Xiao Li, Jinglu Wang, Wei Hu, and Yan Lu. Neural capture of animatable 3d human from monocular video. In ECCV, 2022. 2", + "[46] Maria Vakalopoulou, Guillaume Chassagnon, Norbert Bus, Rafael Marini, Evangelia I Zacharaki, M-P Revel, and Nikos Paragios. Atlasnet: Multi-atlas non-linear deep networks for medical image segmentation. In MICCAI, 2018. 2", + "[47] Shaofei Wang, Katja Schwarz, Andreas Geiger, and Siyu Tang. Arah: Animatable volume rendering of articulated human sdfs. In ECCV, 2022. 3", + "[48] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 2004. 5", + "[49] Chung-Yi Weng, Brian Curless, Pratul P Srinivasan, Jonathan T Barron, and Ira Kemelmacher-Shlizerman. Humaner: Free-viewpoint rendering of moving people from monocular video. In CVPR, 2022. 2, 3, 5, 6, 8", + "[50] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. Synsin: End-to-end view synthesis from a single image. In CVPR, 2020. 2", + "[51] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J Black. Icon: Implicit clothed humans obtained from normals. In CVPR, 2022. 3" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Yuliang Xiu, Jinlong Yang, Xu Cao, Dimitrios Tzionas, and Michael J Black. Econ: Explicit clothed humans optimized via normal integration. In CVPR, 2023. 3", + "[53] Feng Xu, Yebin Liu, Carsten Stoll, James Tompkin, Gaurav Bharaj, Qionghai Dai, Hans-Peter Seidel, Jan Kautz, and Christian Theobalt. Video-based characters: creating new human performances from a multi-view video database. ACM TOG, 2011. 2", + "[54] Weipeng Xu, Avishek Chatterjee, Michael Zollhöfer, Helge Rhodin, Dushyant Mehta, Hans-Peter Seidel, and Christian Theobalt. Monoperfcap: Human performance capture from monocular video. In SIGGRAPH, 2018. 3", + "[55] Yuelang Xu, Benwang Chen, Zhe Li, Hongwen Zhang, Lizhen Wang, Zerong Zheng, and Yebin Liu. Gaussian head avatar: Ultra high-fidelity head avatar via dynamic gaussians. In CVPR, 2024. 3", + "[56] Wang Yifan, Felice Serena, Shihao Wu, Cengiz Oztireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM TOG, 2019. 2", + "[57] Zhengming Yu, Wei Cheng, Xian Liu, Wayne Wu, and Kwan-Yee Lin. Monohuman: Animatable human neural field from monocular video. In CVPR, 2023. 2, 3, 8", + "[58] Hongwen Zhang, Yating Tian, Xinchi Zhou, Wanli Ouyang, Yebin Liu, Limin Wang, and Zhenan Sun. Pymaf: 3d human pose and shape regression with pyramidal mesh alignment feedback loop. In ICCV, 2021. 8", + "[59] Hongwen Zhang, Siyou Lin, Ruizhi Shao, Yuxiang Zhang, Zerong Zheng, Han Huang, Yandong Guo, and Yebin Liu. Closet: Modeling clothed humans on continuous surface with explicit template decomposition. In CVPR, 2023. 3", + "[60] Hongwen Zhang, Yating Tian, Yuxiang Zhang, Mengcheng Li, Liang An, Zhenan Sun, and Yebin Liu. Pymaf-x: Towards well-aligned full-body model regression from monocular images. IEEE T-PAMI, 2023. 8", + "[61] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5", + "[62] Yuxiang Zhang, Hongwen Zhang, Liangxiao Hu, Jiajun Zhang, Hongwei Yi, Shengping Zhang, and Yebin Liu. Proxycap: Real-time monocular full-body capture in world space via human-centric proxy-to-motion learning. In CVPR, 2024. 8", + "[63] Hao Zhao, Jinsong Zhang, Yu-Kun Lai, Zerong Zheng, Yingdi Xie, Yebin Liu, and Kun Li. High-fidelity human avatars from a single rgb camera. In CVPR, 2022. 2, 3", + "[64] Shunyuan Zheng, Boyao Zhou, Ruizhi Shao, Boring Liu, Shengping Zhang, Liqiang Nie, and Yebin Liu. Gps-gaussian: Generalizable pixel-wise 3d gaussian splattering for real-time human novel view synthesis. In CVPR, 2024. 3", + "[65] Yufeng Zheng, Wang Yifan, Gordon Wetzstein, Michael J Black, and Otmar Hilliges. Pointavatar: Deformable point-based head avatars from videos. In CVPR, 2023. 2, 3", + "[66] Zerong Zheng, Tao Yu, Yebin Liu, and Qionghai Dai. Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction. IEEE T-PAMI, 2021. 3" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "643", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[67] Zerong Zheng, Han Huang, Tao Yu, Hongwen Zhang, Yandong Guo, and Yebin Liu. Structured local radiance fields for human avatar modeling. In CVPR, 2022. 3", + "[68] Zerong Zheng, Xiaochen Zhao, Hongwen Zhang, Boning Liu, and Yebin Liu. Avatarrex: Real-time expressive full-body avatars. ACM TOG, 2023. 3", + "[69] Boyao Zhou, Jean-Sébastien Franco, Federica Bogo, Bugra Tekin, and Edmond Boyer. Reconstructing human body mesh from point clouds by adversarial gp network. In ACCV, 2020. 3" + ], + "bbox": [ + 78, + 90, + 468, + 231 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "644", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/GaussianAvatar_ Towards Realistic Human Avatar Modeling from a Single Video via Animatable 3D Gaussians/30b3a92a-861e-47e6-8246-2e90d1765e38_model.json b/2024/GaussianAvatar_ Towards Realistic Human Avatar Modeling from a Single Video via Animatable 3D Gaussians/30b3a92a-861e-47e6-8246-2e90d1765e38_model.json new file mode 100644 index 0000000000000000000000000000000000000000..91d90bb9c234e4e4e511ab8fa55aed6a47fff913 --- /dev/null +++ b/2024/GaussianAvatar_ Towards Realistic Human Avatar Modeling from a Single Video via Animatable 3D Gaussians/30b3a92a-861e-47e6-8246-2e90d1765e38_model.json @@ -0,0 +1,2488 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.178, + 0.131, + 0.795, + 0.177 + ], + "angle": 0, + "content": "GaussianAvatar: Towards Realistic Human Avatar Modeling from a Single Video via Animatable 3D Gaussians" + }, + { + "type": "text", + "bbox": [ + 0.162, + 0.203, + 0.806, + 0.239 + ], + "angle": 0, + "content": "Liangxiao \\(\\mathrm{Hu}^{\\dagger,1}\\), Hongwen Zhang\\(^{2}\\), Yuxiang Zhang\\(^{3}\\), Boyao Zhou\\(^{3}\\), Boning Liu\\(^{3}\\), Shengping Zhang\\(^{*,1,4}\\), Liqiang Nie\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.249, + 0.24, + 0.721, + 0.258 + ], + "angle": 0, + "content": "1Harbin Institute of Technology 2Beijing Normal University" + }, + { + "type": "text", + "bbox": [ + 0.301, + 0.257, + 0.671, + 0.275 + ], + "angle": 0, + "content": "\\(^{3}\\)Tsinghua University \\(^{4}\\)Peng Cheng Laboratory" + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.277, + 0.846, + 0.293 + ], + "angle": 0, + "content": "{lx.hu, s.zhang}@hit.edu.cn, zhanghongwen@bnu.edu.cn, yx-z19@mails.tsinghua.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.204, + 0.296, + 0.759, + 0.31 + ], + "angle": 0, + "content": "{bzhou22, liuboning}@mail.tsinghua.edu.cn, nieliqiang@gmail.com" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.361, + 0.852, + 0.59 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.604, + 0.893, + 0.633 + ], + "angle": 0, + "content": "Figure 1. We propose GaussianAvatar, which learns animatable 3D Gaussians to represent detailed human avatars from a single video. Our method maintains a 3D consistent appearance even when animated by out-of-distribution motions." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.647, + 0.314, + 0.663 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.68, + 0.471, + 0.862 + ], + "angle": 0, + "content": "We present GaussianAvatar, an efficient approach to creating realistic human avatars with dynamic 3D appearances from a single video. We start by introducing animatable 3D Gaussians to explicitly represent humans in various poses and clothing styles. Such an explicit and animatable representation can fuse 3D appearances more efficiently and consistently from 2D observations. Our representation is further augmented with dynamic properties to support pose-dependent appearance modeling, where a dynamic appearance network along with an estimizable feature tensor is designed to learn the motion-to-appearance mapping. Moreover, by leveraging the differentiable motion" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.649, + 0.892, + 0.784 + ], + "angle": 0, + "content": "condition, our method enables a joint optimization of motions and appearances during avatar modeling, which helps to tackle the long-standing issue of inaccurate motion estimation in monocular settings. The efficacy of GaussianAvatar is validated on both the public dataset and our collected dataset, demonstrating its superior performances in terms of appearance quality and rendering efficiency. The code and dataset are available at https://github.com/aipixel/GaussianAvatar." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.815, + 0.633, + 0.831 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Creating a customized human avatar from a single video has great potential for many applications including virtual and augmented reality, the Metaverse, gaming, and movie industries. This task is appealing yet challenging, as the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.1, + 0.875, + 0.406, + 0.889 + ], + "angle": 0, + "content": "† Work done during an internship at Tsinghua University." + }, + { + "type": "page_footnote", + "bbox": [ + 0.1, + 0.889, + 0.234, + 0.901 + ], + "angle": 0, + "content": "* Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.875, + 0.406, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "634" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.168 + ], + "angle": 0, + "content": "monocular observations are highly underdetermined for the modeling of a 3D animatable avatar. Meanwhile, the inaccurate body motion estimations and the complex wrinkle deformations also make it extremely difficult to create a realistic avatar from a single video." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.17, + 0.473, + 0.518 + ], + "angle": 0, + "content": "The modeling of 3D human avatars from monocular videos involves a fusion process of 2D observations to a 3D consistent human model. For this purpose, existing solutions have leveraged both implicit and explicit representations to create the base model of avatars. Among them, implicit methods [6, 15, 16, 49, 57] define a deformable human NeRF to fuse the image observation from current motion space to a canonical space by inverse skinning. However, the implicit 3D volume is inefficient in representing human surfaces and the inverse skinning also introduces ambiguous correspondences during the fusion process. These issues make it hard for implicit solutions to capture fine-grained details of moving people. As the avatar appearances are concentrated around human surfaces, explicit representations are much more efficient in modeling avatars. Previous attempts [12, 63] have employed differentiable mesh rendering to reconstruct the human surface, but these methods struggle to preserve wrinkle details due to a fixed mesh topology. On the other hand, point-based representations [18, 37, 50, 56, 65] are more effective to handle flexible topology but require millions of points to capture detailed appearances. How to represent humans remains one of the fundamental problems for avatar modeling." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.52, + 0.473, + 0.732 + ], + "angle": 0, + "content": "When creating an animatable avatar from a video of moving people, the algorithm is required to learn the relationships between body motions and corresponding appearances. However, the motions estimated from monocular videos are typically faulty, leading to large artifacts in the modeling of dynamic cloth deformations. To address this issue, previous works [1, 15, 16, 42, 45, 49, 57] have attempted to optimize body motions along with the learning of animatable avatar volumes. As the body motions are explicitly represented as parametric meshes, the implicit 3D volume of previous methods makes the optimization indirect and less effective. This issue has even become the main obstacle to achieving high-quality avatar modeling from monocular videos." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.473, + 0.903 + ], + "angle": 0, + "content": "To tackle the above issues, we introduce new representations and solutions to achieve high-quality avatar modeling from a single video. The key insight of our solution is to model dynamic human surfaces explicitly and optimize both the motion and appearances jointly in an end-to-end manner. To this end, we propose GaussianAvatar, a method to reconstruct human avatars with dynamic appearances using the 3D Gaussian representation [17]. As an explicit representation, 3D Gaussian can be easily posed from the canonical space to the motion space via a forward skinning process. Such an animatable 3D Gaus" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.379 + ], + "angle": 0, + "content": "sian representation bypasses the inverse skinning process used in the aforementioned NeRF-based methods and overcomes the previous one-to-many issue [4] during canonicalization. Based on the animatable 3D Gaussian, our method can fuse 3D appearances more consistently from 2D observations to a canonical 3D space. To model dynamic human appearances under different poses, we additionally add pose-dependent properties to 3D Gaussians and incorporate them with the canonical human surfaces. Inspired by previous work [28, 46], we learn a dynamic appearance network on the 2D manifolds of the underlying human mesh to predict dynamic properties of 3D Gaussians. However, due to the strong bias of limited training poses, modeling dynamic appearance solely conditioned on pose information struggles to generalize to novel views and poses. To address this issue, we introduce an estimizable feature tensor to capture a coarse global appearance of human avatars. Subsequently, we incorporate pose-dependent effects on the feature tensor to decode fine-grained details such as wrinkles." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.379, + 0.895, + 0.56 + ], + "angle": 0, + "content": "As the proposed animatable 3D Gaussians are differentiable with respect to the motion conditions, it enables a joint optimization of motion and appearances. This merit allows our network to refine the motion along with the avatar modeling process, which helps to tackle the long-standing issue of inaccurate motion estimation in monocular settings. Besides, the refined motion further enhances the accuracy of avatar modeling since the 3D appearance fusion process relies on motion-based skinning. As shown in our experiments, our method is quite robust to initial motion estimation and has the ability to correct the misalignment of motion capture results." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.57, + 0.874, + 0.584 + ], + "angle": 0, + "content": "To summarize, our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.588, + 0.892, + 0.661 + ], + "angle": 0, + "content": "- We introduce animatable 3D Gaussians for realistic human avatar modeling from a single video. By representing human surfaces explicitly, our method can fuse 3D appearances more consistently and efficiently from 2D observations." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.664, + 0.892, + 0.738 + ], + "angle": 0, + "content": "- We augment the animatable 3D Gaussians with dynamic properties to support pose-dependent appearance modeling, where a dynamic appearance network along with an estimizable feature tensor is designed to learn the motion-to-appearance mapping." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.739, + 0.892, + 0.799 + ], + "angle": 0, + "content": "- We propose to jointly optimize the motion and appearance during the avatar modeling, enabling our method to correct the misalignment of initial motion and improve the final appearance quality." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.588, + 0.892, + 0.799 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.815, + 0.642, + 0.831 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Neural Rendering for Human Reconstruction. Without the need to define a template mesh for avatar modeling [2, 9, 53], neural rendering has emerged as a potent technique that enables learning avatars directly from im" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "635" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "ages. Here we briefly review precious work that aims to reconstruct humans using neural rendering." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.123, + 0.47, + 0.424 + ], + "angle": 0, + "content": "Due to the high-quality rendering of neural radiance field [29], various efforts [20, 21, 24, 32, 33, 47, 67, 68] have been made to reconstruct the dynamic appearance of moving people. Neural Body [33] associates a latent code to each SMPL [26] vertex to encode the appearance, which is transformed into observation space based on the human pose. Neural Actor [24] learns a deformable radiance field with SMPL as guidance and utilizes a texture map to improve its final rendering quality. TAVA [20] proposes to jointly model the non-rigid warping field and shading effects directly conditioned on the pose vectors. Posevocab [21] designs joint-structured pose embeddings to encode the dynamic appearances under different key poses, such embeddings can better learn joint-related appearance. NeRF-based methods have demonstrated appealing rendering results on human avatar reconstruction, but still struggle to represent human surfaces with the implicit 3D volume. Explicit modeling of human surfaces is a more straightforward way for this task, as the dynamic appearance of humans is mostly reflected on the human surface." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.426, + 0.47, + 0.728 + ], + "angle": 0, + "content": "Explicit representations have great potential for human reconstruction. In HF-Avatar [63], meshes are used as the base representation in a coarse-to-fine framework that combines neural texture with dynamic surface deformation for avatar creation. EMA [12] proposes Meshy neural fields to reconstruct human avatars by optimizing the canonical mesh, material, and motion dynamics through inverse rendering in an end-to-end process. PointAvatar [65] employs a deformable point-based representation to separate source color into intrinsic albedo and normal-dependent shading. DVA [35] extends mixtures of volumetric primitives [25] for human avatar modeling. All these attempts have demonstrated the significant potential of explicit representations and their under-exploration. However, meshes are constrained by fixed topologies, whereas point clouds demand a multitude of points to encompass intricate details. 3D Gaussians [17] have showcased their capability in various human tasks [22, 40, 55, 64]. In essence, 3D Gaussians hold promise for human avatar reconstruction and are currently a subject of active research." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Avatar Modeling from Monocular Videos. Numerous methods investigate reconstructing humans from single images or monocular videos. Regression-based methods[10, 11, 13, 38, 39, 51, 52, 66] directly recover clothed 3D humans just from a single image. While these methods produce attractive results, they can not recover the dynamic appearance in the reconstruction across the entire sequence. Traditional methods aim to capture human dynamics by tracking individuals in videos using pre-scanned rigged templates [7, 8, 54]. However, the pre-scanning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.151 + ], + "angle": 0, + "content": "and manual rigging processes prevent their real-life applications. [1, 5, 30] try to bypass the requirement for predefined human models but face challenges in preserving fine details because of the fixed mesh resolution." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.153, + 0.893, + 0.41 + ], + "angle": 0, + "content": "The emergence of neural radiance fields [29] has facilitated the creation of various techniques for reconstructing animatable avatars from monocular videos [3, 14, 16, 41-43, 49]. However, inaccuracies in estimating human motions from monocular videos result in pronounced artifacts. To address the problem, HumanNeRF [49] solves for an update to the inaccurate poses. NeuMan [16] introduces an error-correction network to enable training with erroneous estimates. Vid2Avatar[6] and InstantAvatar [15] also jointly optimize motions by back-propagating the gradient of the image reconstruction loss to the pose parameters. MonoHuman [57] introduces bi-directional constraints to alleviate ambiguous correspondence on novel poses. Attempting to address the inaccurate pose estimation issue with implicit human NeRF models has proven to be inefficient and imprecise. Consequently, we endeavor to address this issue by leveraging an explicit 3D Gaussian representation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.421, + 0.593, + 0.437 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.447, + 0.892, + 0.658 + ], + "angle": 0, + "content": "Our goal is to create a human avatar that enables free-viewpoint rendering and realistic animation using a single video. We first introduce an expressive representation, namely animatable 3D Gaussians, to represent human avatars in Sec. 3.1. With this representation, modeling the dynamic appearance of humans can be regarded as dynamic 3D Gaussian property estimation. Then we build a dynamic appearance network along with an estimable feature tensor to learn the motion-to-appearance mapping in Sec. 3.2. To mitigate the artifacts caused by inaccurate pose estimation, we adopt a joint motion and appearance optimization approach for refining human poses during training (Sec. 3.3). In the following, we provide a comprehensive explanation of the technical details." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.667, + 0.736, + 0.682 + ], + "angle": 0, + "content": "3.1. Animatable 3D Gaussians" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.796 + ], + "angle": 0, + "content": "Point-based representation [23, 28, 34, 59, 69] has proven its topological flexibility in generating realistic human avatars from scans. Extending this explicit representation to create human avatars from images is a significant endeavor. With this goal in mind, we introduce a novel representation, termed animatable 3D Gaussians, which effectively reconstructs human surfaces with a 3D consistent appearance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.797, + 0.892, + 0.901 + ], + "angle": 0, + "content": "3D Gaussian Splitting [17] is a point-based scene representation that allows high-quality real-time rendering. The scene representation is parameterized by a set of static 3D Gaussians, each of which has the following parameters: 3D center position \\(\\mathbf{x} \\in \\mathbb{R}^3\\), color \\(\\mathbf{c} \\in \\mathbb{R}^3\\), opacity \\(\\alpha \\in \\mathbb{R}\\), 3D rotation in form of quaternion \\(\\mathbf{q} \\in \\mathbb{R}^4\\) and 3D scaling factor \\(\\mathbf{s} \\in \\mathbb{R}^3\\). With these properties, we can generate rendered" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "636" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.092, + 0.891, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.274, + 0.897, + 0.348 + ], + "angle": 0, + "content": "Figure 2. Overview of GaussianAvatar. Given a fitted SMPL or SMPL-X model on the current frame, we sample the points on its surface and record their positions on a UV positional map \\(I\\), which is then passed to a pose encoder to obtain the pose feature. An estimizable feature tensor is pixel-aligned with the pose feature and learned to capture the coarse appearance of humans. Then the two aligned feature tensors are input into the Gaussian parameter decoder, which predicts each point's offset \\(\\Delta \\hat{\\mathbf{x}}\\), color \\(\\hat{\\mathbf{c}}\\), and scale \\(\\hat{s}\\). These predictions, along with the fixed rotations \\(\\mathbf{q}\\) and opacity \\(\\alpha\\), collectively constitute the animatable 3D Gaussians in canonical space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.359, + 0.471, + 0.39 + ], + "angle": 0, + "content": "images from any viewpoint. We refer readers to [17, 27] for the rendering details of 3D Gaussians." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.39, + 0.471, + 0.435 + ], + "angle": 0, + "content": "To extend this representation for human avatar modeling, we integrate it with either the SMPL [26] or SMPL-X [31] model as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.448, + 0.47, + 0.467 + ], + "angle": 0, + "content": "\\[\nG (\\pmb {\\beta}, \\pmb {\\theta}, \\mathbf {D}, \\mathbf {P}) = S p l a t t i n g (W (\\mathbf {D}, J (\\pmb {\\beta}), \\pmb {\\theta}, \\omega), \\mathbf {P}), (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.476, + 0.473, + 0.691 + ], + "angle": 0, + "content": "where \\( G(\\cdot) \\) represents a rendered image, and \\( Splatting(\\cdot) \\) denotes the rendering process of 3D Gaussians from any viewpoint, \\( W(\\cdot) \\) is a standard linear blend skinning function employed for reposing 3D Gaussians, \\( \\mathbf{D} = T(\\boldsymbol{\\beta}) + dT \\) represents the locations of 3D Gaussians in canonical space, formed by adding corrective point displacements \\( dT \\) on the template mesh surface \\( T(\\beta) \\), \\( \\mathbf{P} \\) denotes the remaining properties of 3D Gaussians, excluding the positions. \\( \\beta \\) and \\( \\theta \\) are the shape and pose parameters, \\( J(\\beta) \\) outputs 3D joint locations. Note that we propagate the skinning weight \\( \\omega \\) from the vertices of the SMPL or SMPL-X model to the nearest 3D Gaussians. With the proposed representation, we can now repose these canonical 3D Gaussians to the motion space for free-view rendering." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.697, + 0.449, + 0.714 + ], + "angle": 0, + "content": "3.2. Dynamic 3D Gaussian Property Estimation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Following the proposed animatable Gaussians, human appearances are determined by the point displacements \\(dT\\) and properties \\(\\mathbf{P}\\). Modeling dynamic human appearances can be regarded as estimating these dynamic properties. To model dynamic human appearances under various poses, we introduce a dynamic appearance network along with an estimizable feature tensor to predict these pose-dependent properties of 3D Gaussians. Despite sharing similarities in network structure with [28], we present this framework for a distinct purpose. In [28], the feature tensor serves to decouple the pose-independent human shape from the decoder, and we repurpose it to capture a coarse global appearance" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.359, + 0.588, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.501, + 0.553, + 0.514 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.359, + 0.651, + 0.498 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.609, + 0.5, + 0.629, + 0.514 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.359, + 0.735, + 0.498 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.686, + 0.5, + 0.704, + 0.513 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.738, + 0.359, + 0.807, + 0.498 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.762, + 0.5, + 0.782, + 0.513 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.814, + 0.359, + 0.887, + 0.498 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.84, + 0.499, + 0.858, + 0.511 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.525, + 0.895, + 0.568 + ], + "angle": 0, + "content": "Figure 3. Effect of iostropy of 3D Gaussians. (a) Input image, (b)(d) front and back views trained with isotropic 3D Gaussians, (c)(e) front and back views trained with anisotropic 3D Gaussians." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.585, + 0.895, + 0.675 + ], + "angle": 0, + "content": "of human avatars. The motivation behind this modification is that directly learning a mapping from human poses to dynamic properties is susceptible to overfitting on the limited training poses. To integrate the global appearance into the feature tensor, we introduce a two-stage training strategy, as discussed in Sec. 3.4." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.677, + 0.895, + 0.735 + ], + "angle": 0, + "content": "The dynamic appearance network is designed to learn a mapping from a 2D manifold representing the underlying human shape to the dynamic properties of 3D Gaussians as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.738, + 0.892, + 0.756 + ], + "angle": 0, + "content": "\\[\nf _ {\\phi}: \\mathcal {S} ^ {2} \\in \\mathbb {R} ^ {3} \\rightarrow \\mathbb {R} ^ {7}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.764, + 0.895, + 0.854 + ], + "angle": 0, + "content": "As shown in Fig. 2, the 2D human manifold \\(S^2\\) is depicted by a UV positional map \\(I \\in R^{H \\times W \\times 3}\\), where each valid pixel stores the position \\((x, y, z)\\) of one point on the posed body surface. The final predictions consist of per point offset \\(\\Delta \\hat{\\mathbf{x}} \\in \\mathbb{R}^3\\), color \\(\\hat{\\mathbf{c}} \\in \\mathbb{R}^3\\), and scale \\(\\hat{s} \\in \\mathbb{R}\\) on the canonical surface." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Instead of predicting all properties, we make these slight adjustments to our task based on experimental results. Due to the unbalanced viewpoints in monocular videos," + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.945, + 0.512, + 0.957 + ], + "angle": 0, + "content": "637" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.287 + ], + "angle": 0, + "content": "anisotropic 3D Gaussians are prone to learning an inaccurate 3D shape to fit the most frequently seen view, resulting in poorer performance in side views. As shown in Fig. 3, training our network with isotropic 3D Gaussians yields superior results. We ensure isotropy among all 3D Gaussians by maintaining uniform size across dimensions. Therefore, we introduce a scalar \\(\\hat{s}\\) to represent the scales of 3D Gaussians and set rotations \\(\\mathbf{q}\\) as \\([1,0,0,0]\\). Through experimental observations, we notice that the network tends to learn an opacity value \\((\\alpha)\\) of zero on the boundary to correct the human shape. To address this, we fix opacity \\(\\alpha = 1\\) to keep all 3D Gaussians visible, enforcing the network to predict accurate positions of the 3D Gaussians." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.288, + 0.473, + 0.471 + ], + "angle": 0, + "content": "The dynamic appearance network consists of two parts: a pose encoder and a Gaussian parameter decoder. The pose encoder takes the UV positional map of posed body points as input to generate a pose-conditioned feature tensor \\( O \\in \\mathbb{R}^{H \\times W \\times C} \\). We then integrate the pixel-aligned optimized feature tensor \\( F \\in \\mathbb{R}^{H \\times W \\times C} \\) with the pose features before feeding it into the Gaussian parameter decoder to generate final predictions. Following this, we add the predicted offsets to the canonical 3D Gaussians and associate the predicted properties with the corresponding 3D Gaussians. With the estimated poses, we can then repose the canonical 3D Gaussians to the motion space for rendering." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.478, + 0.45, + 0.494 + ], + "angle": 0, + "content": "3.3. Joint Motion and Appearance Optimization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.501, + 0.47, + 0.622 + ], + "angle": 0, + "content": "However, owing to the imprecise estimation of human poses \\(\\pmb{\\theta} = (\\theta_{1},\\theta_{2},\\dots,\\theta_{J})\\) and translations \\(\\mathbf{t}\\) from monocular videos, the reposed 3D Gaussians in motion space are inaccurately represented and may lead to unsatisfactory rendering outcomes. To address this, we propose to jointly optimize human motions and appearances. To optimize human motions with image loss, we solve for an update \\((\\Delta \\pmb {\\theta},\\Delta \\mathbf{t})\\) to the estimated body poses and translations as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.63, + 0.47, + 0.648 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {\\Theta}} = (\\boldsymbol {\\theta} + \\Delta \\boldsymbol {\\theta}, \\mathbf {t} + \\Delta \\mathbf {t}). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.658, + 0.47, + 0.75 + ], + "angle": 0, + "content": "We modify \\(\\theta\\) in Eq. 1 using \\(\\Theta\\) to render the proposed animatable 3D Gaussians differentiable with respect to the motion conditions. Different from previous work [6, 16, 49] that jointly optimizes human poses via inverse skinning, we optimize the updates in a forward skinning process, which benefits both motion and appearance optimization." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.757, + 0.251, + 0.773 + ], + "angle": 0, + "content": "3.4. Training Strategy" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In this section, we outline our approach to training the network with inaccurate human motions. We conduct a two-stage optimization process using different loss functions. In the first stage, we aim to fuse the sequential appearances to the estimizable feature tensor and conduct motion optimization to get accurate poses for the dynamic appearance network. In this stage, we optimize the framework without incorporating any pose-dependent information by excluding" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "the training of the pose encoder. Specifically, we utilize the following loss functions to train our network:" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.138, + 0.892, + 0.173 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {s t a g e} _ {1}} = \\lambda_ {r b g} \\mathcal {L} _ {r b g} + \\lambda_ {s s i m} \\mathcal {L} _ {s s i m} + \\lambda_ {l p i p s} \\mathcal {L} _ {l p i p s} \\tag {4} \\\\ + \\lambda_ {f} \\mathcal {L} _ {f} + \\lambda_ {\\text {o f f s e t}} \\mathcal {L} _ {\\text {o f f s e t}} + \\lambda_ {\\text {s c a l e}} \\mathcal {L} _ {\\text {s c a l e}}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.187, + 0.892, + 0.265 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{rbg}\\), \\(\\mathcal{L}_{ssim}\\), and \\(\\mathcal{L}_{lpips}\\) are the L1 loss, SSIM loss [48], and LPIPS loss [61], respectively. \\(\\mathcal{L}_f\\), \\(\\mathcal{L}_{offset}\\), \\(\\mathcal{L}_{scale}\\) calculate the L2-norm of the feature map, predicted offsets and scales, respectively. We set \\(\\lambda_{rbg} = 0.8\\), \\(\\lambda_{ssim} = 0.2\\), \\(\\lambda_{lpips} = 0.2\\), \\(\\lambda_f = 1\\), \\(\\lambda_{offset} = 10\\), \\(\\lambda_{scale} = 1\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.266, + 0.893, + 0.417 + ], + "angle": 0, + "content": "After the first stage of training, we obtain more accurate human motions and an optimized feature tensor \\( F \\). The optimized feature tensor \\( F \\) captures a coarse appearance of human avatars. In the second stage, we incorporate the pose features encoded by the pose encoder with the trained feature tensor \\( F \\). We replace \\( \\mathcal{L}_f \\) with the L2-norm loss \\( \\mathcal{L}_p \\), which plays the same role as \\( \\mathcal{L}_f \\) in regularizing the limited pose space. By penalizing the pose-dependent features, we can eliminate the strong bias of limited training poses and thus generalize to unseen viewpoints and poses." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.435, + 0.634, + 0.453 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.462, + 0.698, + 0.478 + ], + "angle": 0, + "content": "4.1. Datasets and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.488, + 0.892, + 0.548 + ], + "angle": 0, + "content": "People-Snapshot Dataset. This dataset [1] comprises videos of individuals rotating in front of a stationary camera. To ensure a fair quantitative comparison, we follow the same evaluation protocol outlined in InstantAvatar [15]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.55, + 0.892, + 0.671 + ], + "angle": 0, + "content": "NeuMan Dataset. To assess even more challenging scenarios, we employ outdoor collections from the NeuMan dataset [16]. These videos are recorded using a mobile phone for moving individuals. Specifically, we select four sequences (bike, citizen, Seattle, jogging) that exhibit most body regions and contain minimal blurry images. We initialize the estimated poses with an off-the-shelf method [44], which is also utilized in [15]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.672, + 0.892, + 0.854 + ], + "angle": 0, + "content": "DynVideo Dataset. With the limited cloth deformation presented in the above two datasets, we propose the Dynvideo dataset to capture dynamic human appearance. To this end, we utilize a mobile phone to record videos of a character performing various movements, especially rotation, in front of the device. Each of these videos takes about one minute and provides a comprehensive and detailed representation of human movement. We also provide the corresponding SMPL parameter sequences for all videos, obtained by our proposed method. This dataset serves as a valuable resource for evaluating reconstruction quality, with a particular emphasis on dynamic appearances." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Evaluation Metrics. We consider three metrics: PSNR, SSIM [48], and LPIPS [61] to access the reconstruction quality on three datasets." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "638" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.895, + 0.201 + ], + "angle": 0, + "content": "
Methodmale-3-casualmale-4-casualfemale-3-casualfemale-4-casual
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
HumanNeRF [49]26.900.96050.018125.500.93970.035724.460.95160.026927.070.96150.0152
InstantAvatar [15]29.530.97160.015527.670.96260.030727.660.97090.021029.110.96830.0167
Baseline27.710.97130.021825.090.96140.030624.800.96090.031026.160.96100.0227
Baseline + Opt.30.980.97910.014328.770.97530.023029.590.97680.022030.830.97680.0145
Baseline + Opt. + Dyn.30.980.97900.014528.780.97550.022829.550.97620.022530.840.97710.0140
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.209, + 0.895, + 0.239 + ], + "angle": 0, + "content": "Table 1. Quantitative evaluation on the People-Snapshot [1] dataset. Opt. denotes the motion optimization and Dyn. refers to dynamic appearance modeling." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.253, + 0.457, + 0.344 + ], + "angle": 0, + "content": "
MethodPSNR ↑SSIM ↑LPIPS ↓
HumanNeRF [49]27.060.96690.0192
InstantAvatar [15]28.470.97150.0277
Baseline27.060.96920.0183
Baseline + Opt.29.930.97940.0126
Baseline + Opt. + Dyn.29.940.97950.0124
" + }, + { + "type": "table_caption", + "bbox": [ + 0.093, + 0.353, + 0.454, + 0.368 + ], + "angle": 0, + "content": "Table 2. Quantitative evaluation on NeuMan [16] dataset." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.383, + 0.456, + 0.472 + ], + "angle": 0, + "content": "
MethodPSNR ↑SSIM ↑LPIPS ↓
HumanNeRF [49]24.390.93490.0349
InstantAvatar [15]20.310.91830.1046
Baseline23.550.93400.0354
Baseline + Opt.26.160.95220.0237
Baseline + Opt. + Dyn.28.580.96160.0169
" + }, + { + "type": "table_caption", + "bbox": [ + 0.103, + 0.482, + 0.444, + 0.497 + ], + "angle": 0, + "content": "Table 3. Quantitative evaluation on DynVideo dataset." + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.515, + 0.149, + 0.67 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.111, + 0.672, + 0.129, + 0.683 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.15, + 0.515, + 0.21, + 0.669 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.171, + 0.672, + 0.187, + 0.683 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.212, + 0.515, + 0.273, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.233, + 0.672, + 0.249, + 0.683 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.515, + 0.338, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.301, + 0.672, + 0.317, + 0.682 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.339, + 0.515, + 0.399, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.361, + 0.672, + 0.376, + 0.683 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image", + "bbox": [ + 0.401, + 0.515, + 0.462, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.423, + 0.672, + 0.438, + 0.683 + ], + "angle": 0, + "content": "(f)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.697, + 0.47, + 0.725 + ], + "angle": 0, + "content": "Figure 4. Motion optimization results. (a)(d) Original image, (b)(e) our optimized SMPL, (c)(f) ROMP [44] estimates." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.754, + 0.285, + 0.77 + ], + "angle": 0, + "content": "4.2. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.779, + 0.469, + 0.884 + ], + "angle": 0, + "content": "Implementation details. We employ a U-Net [36] for extracting pose-dependent features, and the Gaussian parameter decoder is implemented as an 8-layer multilayer perceptron (MLP). We sample approximately 200,000 points on the SMPL mesh surface. The entire framework is trained on a single NVIDIA RTX 3090 GPU, with training times ranging from 0.5 to 6 hours." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Baseline. To showcase the efficacy of the proposed mod" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.255, + 0.695, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.372, + 0.604, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.554, + 0.482, + 0.569, + 0.493 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.606, + 0.372, + 0.696, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.647, + 0.482, + 0.662, + 0.493 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.256, + 0.787, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.372, + 0.789, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.739, + 0.482, + 0.753, + 0.492 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.79, + 0.256, + 0.882, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.791, + 0.372, + 0.881, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.83, + 0.482, + 0.846, + 0.492 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.507, + 0.892, + 0.537 + ], + "angle": 0, + "content": "Figure 5. Qualitative ablation studies. (a) Ground truth, (b) baseline + Opt. + Dyn., (c) baseline + Opt., (d) baseline." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.552, + 0.892, + 0.627 + ], + "angle": 0, + "content": "ules, our method can be partitioned into three components: baseline, motion optimization, and dynamic appearance modeling. As our baseline, we suspend the motion optimization process and exclude the pose-dependent information from the pose encoder." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.627, + 0.892, + 0.716 + ], + "angle": 0, + "content": "Methods for Comparison. We compare our method against (1) HumanNeRF [49], which implicitly represents a human avatar with a canonical appearance neural field and a motion field; (2) InstantAvatar [15], which achieves fast avatar modeling by using several acceleration strategies for neural fields." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.727, + 0.829, + 0.743 + ], + "angle": 0, + "content": "4.3. Comparisons with the State of the Art" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We report the numeric evaluation results on three datasets. As shown in Table 1, 2, and 3, our proposed method outperforms all baselines on all metrics for recovering more details of dynamic appearance and correcting the artifacts caused by initialized poses. To demonstrate the qualitative evaluation of these datasets, we also visualize the novel view synthesis results on the test splits. As implicit representations, HumanNeRF and InstantAvatar are prone to generate ghosting effects at boundary areas, as shown in Fig. 6. InstantAvatar lacks the capability to model pose" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "639" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.097, + 0.52, + 0.739 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.097, + 0.875, + 0.735 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.747, + 0.895, + 0.777 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison of novel view synthesis. We compare the novel view synthesis quality on the People-Snapshot dataset (first row), NeuMan dataset (second row), and DynVideo dataset (last two rows)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.799, + 0.471, + 0.83 + ], + "angle": 0, + "content": "dependent deformations, leading to challenges in effectively handling the dynamic appearance of moving people." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.472, + 0.903 + ], + "angle": 0, + "content": "To showcase the robustness of our avatar modeling, we collect more challenging poses [19] captured by the monocular camera and evaluate our method and other methods for avatar animation and novel view synthesis. As ground truth" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.799, + 0.895, + 0.889 + ], + "angle": 0, + "content": "is unavailable for these out-of-distribution poses, we illustrate qualitative results of novel pose and view synthesis in Fig. 1 and Fig. 7. Our method generates realistic animation results with these challenging poses, demonstrating a consistent 3D appearance in novel views with respect to other methods." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.513, + 0.957 + ], + "angle": 0, + "content": "640" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.096, + 0.885, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.458, + 0.893, + 0.486 + ], + "angle": 0, + "content": "Figure 7. Animation results on out-of-distribution motions. We compare the animation results on the People-Snapshot dataset (first row), NeuMan dataset (second row), and DynVideo dataset (last row)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.509, + 0.241, + 0.523 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.535, + 0.47, + 0.732 + ], + "angle": 0, + "content": "In this section, we conduct ablation studies to validate each component of our methods. As shown in Table 1, 2, and 3, our proposed motion optimization module dramatically improves over the baseline on all metrics, demonstrating its effectiveness in modeling human avatars. To highlight the effectiveness of our method in human motion optimization, we illustrate the initialized poses obtained by ROMP [44] and the optimized ones by our method on NeuMan and DynVideo datasets in Fig. 4. Experiments demonstrate that joint motion optimization is capable of correcting inaccurate motion estimation, even for the side and back views. Furthermore, our approach readily extends to enhance the accuracy of existing motion capture methods [58, 60, 62]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.9 + ], + "angle": 0, + "content": "We also notice that the dynamic appearance modeling achieves superior results on NeuMan and DynVideo datasets in Table 2 and 3 and slightly favorable outcomes in Table 1, considering that People-Snapshot has limited pose variation with respect to the other two datasets. In Fig. 5, we further illustrate the visible improvements of each component of our method. The motion optimization scheme improves the global quality of the baseline for better pose estimation. The dynamic appearance modeling additionally preserves pose-dependent details such as cloth wrinkles on the human surface." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.508, + 0.747, + 0.523 + ], + "angle": 0, + "content": "5. Conclusion and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.534, + 0.892, + 0.699 + ], + "angle": 0, + "content": "We introduce GaussianAvatar, a human avatar reconstruction method based on the proposed animatable 3D Gaussians from monocular videos. For dynamic human appearance modeling, we leverage a dynamic appearance network along with an estimizable feature tensor to enhance the representation with dynamic properties. Besides, we implement a joint motion and appearance optimization scheme to rectify estimated motion and enhance the overall reconstruction quality. Our method shows the capability to reconstruct avatars with dynamic appearances, enabling realistic animation while maintaining real-time rendering speed." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.7, + 0.892, + 0.759 + ], + "angle": 0, + "content": "Limitation. Similar to [15, 49, 57], our method may generate artifacts due to inaccurate foreground segmentations in videos and encounter challenges in modeling loose outfits such as dresses." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.76, + 0.892, + 0.82 + ], + "angle": 0, + "content": "Potential Social Impact. Given our method's capability to reconstruct a realistic personalized character from a monocular video, it is imperative to exercise caution and consider the potential for technology misuse." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.82, + 0.892, + 0.896 + ], + "angle": 0, + "content": "Acknowledgement. This work was supported by the NSFC project (Nos. 62272134, 62236003, 62072141, 62301298, and 62125107), Shenzhen College Stability Support Plan (Grant No. GXWD20220817144428005), and the Major Key Project of PCL (PCL2023A10-2)." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.51, + 0.957 + ], + "angle": 0, + "content": "641" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.469, + 0.156 + ], + "angle": 0, + "content": "[1] Thiemo Alldieck, Marcus Magnor, Weipeng Xu, Christian Theobalt, and Gerard Pons-Moll. Video based reconstruction of 3d people models. In CVPR, 2018. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.469, + 0.214 + ], + "angle": 0, + "content": "[2] Timur Bagautdinov, Chenglei Wu, Tomas Simon, Fabian Prada, Takaaki Shiratori, Shih-En Wei, Weipeng Xu, Yaser Sheikh, and Jason Saragih. Driving-signal aware full-body avatars. ACM TOG, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.217, + 0.469, + 0.272 + ], + "angle": 0, + "content": "[3] Jianchuan Chen, Ying Zhang, Di Kang, Xuefei Zhe, Linchao Bao, Xu Jia, and Huchuan Lu. Animatable neural radiance fields from monocular rgb videos. arXiv preprint arXiv:2106.13629, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.275, + 0.469, + 0.328 + ], + "angle": 0, + "content": "[4] Xu Chen, Yufeng Zheng, Michael J Black, Otmar Hilliges, and Andreas Geiger. Snarf: Differentiable forward skinning for animating non-rigid neural implicit shapes. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.333, + 0.469, + 0.374 + ], + "angle": 0, + "content": "[5] Chen Guo, Xu Chen, Jie Song, and Otmar Hilliges. Human performance capture from monocular video in the wild. In 3DV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.377, + 0.469, + 0.432 + ], + "angle": 0, + "content": "[6] Chen Guo, Tianjian Jiang, Xu Chen, Jie Song, and Otmar Hilliges. Vid2 avatar: 3d avatar reconstruction from videos in the wild via self-supervised scene decomposition. In CVPR, 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.435, + 0.469, + 0.489 + ], + "angle": 0, + "content": "[7] Marc Habermann, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Livecap: Real-time human performance capture from monocular video. ACM TOG, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.493, + 0.469, + 0.547 + ], + "angle": 0, + "content": "[8] Marc Habermann, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Deepcap: Monocular human performance capture using weak supervision. In CVPR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.551, + 0.469, + 0.592 + ], + "angle": 0, + "content": "[9] Marc Habermann, Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Real-time deep dynamic characters. ACM TOG, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.469, + 0.636 + ], + "angle": 0, + "content": "[10] Tong He, John Collomosse, Hailin Jin, and Stefano Soatto. Geo-pifu: Geometry and pixel aligned implicit functions for single-view human reconstruction. In NeurIPS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.639, + 0.469, + 0.68 + ], + "angle": 0, + "content": "[11] Tong He, Yuanlu Xu, Shunsuke Saito, Stefano Soatto, and Tony Tung. Arch++: Animation-ready clothed human reconstruction revisited. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.683, + 0.469, + 0.724 + ], + "angle": 0, + "content": "[12] Xiaoke Huang, Yiji Cheng, Yansong Tang, Xiu Li, Jie Zhou, and Jiwen Lu. Efficient meshy neural fields for animatable human avatars. arXiv preprint arXiv:2303.12965, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.727, + 0.469, + 0.768 + ], + "angle": 0, + "content": "[13] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. Arch: Animatable reconstruction of clothed humans. In CVPR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.771, + 0.469, + 0.812 + ], + "angle": 0, + "content": "[14] Boyi Jiang, Yang Hong, Hujun Bao, and Juyong Zhang. Self-con: Self reconstruction your digital avatar from monocular video. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.815, + 0.469, + 0.856 + ], + "angle": 0, + "content": "[15] Tianjian Jiang, Xu Chen, Jie Song, and Otmar Hilliges. Instantavatar: Learning avatars from monocular video in 60 seconds. In CVPR, 2023. 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[16] Wei Jiang, Kwang Moo Yi, Golnoosh Samei, Oncel Tuzel, and Anurag Ranjan. Neuman: Neural human radiance field from a single video. In ECCV, 2022. 2, 3, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM TOG, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[18] Christoph Lassner and Michael Zollhofer. Pulsar: Efficient sphere-based neural rendering. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.164, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[19] Ruilong Li, Shan Yang, David A. Ross, and Angjoo Kanazawa. Learn to dance with aist++: Music conditioned 3d dance generation. In ICCV, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.258 + ], + "angle": 0, + "content": "[20] Ruilong Li, Julian Tanke, Minh Vo, Michael Zollhöfer, Jürgen Gall, Angjoo Kanazawa, and Christoph Lassner. Tava: Template-free animatable volumetric actors. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.302 + ], + "angle": 0, + "content": "[21] Zhe Li, Zerong Zheng, Yuxiao Liu, Boyao Zhou, and Yebin Liu. Posevocab: Learning joint-structured pose embeddings for human avatar modeling. In SIGGRAPH, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.304, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[22] Zhe Li, Zerong Zheng, Lizhen Wang, and Yebin Liu. Animatable gaussians: Learning pose-dependent gaussian maps for high-fidelity human avatar modeling. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.345, + 0.892, + 0.385 + ], + "angle": 0, + "content": "[23] Siyou Lin, Hongwen Zhang, Zerong Zheng, Ruizhi Shao, and Yebin Liu. Learning implicit templates for point-based clothed human modeling. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.387, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[24] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. ACM TOG, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.443, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[25] Stephen Lombardi, Tomas Simon, Gabriel Schwartz, Michael Zollhoefer, Yaser Sheikh, and Jason Saragih. Mixture of volumetric primitives for efficient neural rendering. ACM Trans. Graph., 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.498, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[26] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. ACM TOG, 2015. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.54, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[27] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. 3DV, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.582, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[28] Qianli Ma, Jinlong Yang, Siyu Tang, and Michael J Black. The power of points for modeling humans in clothing. In ICCV, 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.624, + 0.892, + 0.677 + ], + "angle": 0, + "content": "[29] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[30] Gyeongsik Moon, Hyeongjin Nam, Takaaki Shiratori, and Young Mu Lee. 3d clothed human reconstruction in the wild. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.721, + 0.892, + 0.775 + ], + "angle": 0, + "content": "[31] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3d hands, face, and body from a single image. In CVPR, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.777, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[32] Sida Peng, Shangzhan Zhang, Zhen Xu, Chen Geng, Boyi Jiang, Hujun Bao, and Xiaowei Zhou. Animatable neural implicit surfaces for creating avatars from videos. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[33] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In CVPR, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "642" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "[34] Sergey Prokudin, Qianli Ma, Maxime Raafat, Julien Valentin, and Siyu Tang. Dynamic point fields. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[35] Edoardo Remelli, Timur Bagautdinov, Shunsuke Saito, Chenglei Wu, Tomas Simon, Shih-En Wei, Kaiwen Guo, Zhe Cao, Fabian Prada, Jason Saragih, et al. Drivable volumetric avatars using texel-aligned features. In SIGGRAPH, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.19, + 0.469, + 0.232 + ], + "angle": 0, + "content": "[36] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.232, + 0.47, + 0.273 + ], + "angle": 0, + "content": "[37] Darius Rückert, Linus Franke, and Marc Stamminger. Adop: Approximate differentiable one-pixel point rendering. ACM TOG, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.469, + 0.328 + ], + "angle": 0, + "content": "[38] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In ICCV, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.469, + 0.372 + ], + "angle": 0, + "content": "[39] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In CVPR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.372, + 0.469, + 0.413 + ], + "angle": 0, + "content": "[40] Ruizhi Shao, Jingxiang Sun, Cheng Peng, Zerong Zheng, Boyao Zhou, Hongwen Zhang, and Yebin Liu. Control4d: Efficient 4d portrait editing with text. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.414, + 0.469, + 0.467 + ], + "angle": 0, + "content": "[41] Shih-Yang Su, Frank Yu, Michael Zollhöfer, and Helge Rhodin. A-nerf: Articulated neural radiance fields for learning human shape, appearance, and pose. In NeurIPS, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.469, + 0.51 + ], + "angle": 0, + "content": "[42] Shih-Yang Su, Timur Bagautdinov, and Helge Rhodin. Danbo: Disentangled articulated neural body representations via graph neural networks. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.469, + 0.539 + ], + "angle": 0, + "content": "[43] Shih-Yang Su, Timur Bagautdinov, and Helge Rhodin. Npc: Neural point characters from video. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.469, + 0.58 + ], + "angle": 0, + "content": "[44] Yu Sun, Qian Bao, Wu Liu, Yili Fu, Michael J Black, and Tao Mei. Monocular, one-stage, regression of multiple 3d people. In ICCV, 2021. 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.469, + 0.622 + ], + "angle": 0, + "content": "[45] Gusi Te, Xiu Li, Xiao Li, Jinglu Wang, Wei Hu, and Yan Lu. Neural capture of animatable 3d human from monocular video. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.469, + 0.678 + ], + "angle": 0, + "content": "[46] Maria Vakalopoulou, Guillaume Chassagnon, Norbert Bus, Rafael Marini, Evangelia I Zacharaki, M-P Revel, and Nikos Paragios. Atlasnet: Multi-atlas non-linear deep networks for medical image segmentation. In MICCAI, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.469, + 0.718 + ], + "angle": 0, + "content": "[47] Shaofei Wang, Katja Schwarz, Andreas Geiger, and Siyu Tang. Arah: Animatable volume rendering of articulated human sdfs. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.469, + 0.761 + ], + "angle": 0, + "content": "[48] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.469, + 0.817 + ], + "angle": 0, + "content": "[49] Chung-Yi Weng, Brian Curless, Pratul P Srinivasan, Jonathan T Barron, and Ira Kemelmacher-Shlizerman. Humaner: Free-viewpoint rendering of moving people from monocular video. In CVPR, 2022. 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.859 + ], + "angle": 0, + "content": "[50] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. Synsin: End-to-end view synthesis from a single image. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[51] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J Black. Icon: Implicit clothed humans obtained from normals. In CVPR, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[52] Yuliang Xiu, Jinlong Yang, Xu Cao, Dimitrios Tzionas, and Michael J Black. Econ: Explicit clothed humans optimized via normal integration. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[53] Feng Xu, Yebin Liu, Carsten Stoll, James Tompkin, Gaurav Bharaj, Qionghai Dai, Hans-Peter Seidel, Jan Kautz, and Christian Theobalt. Video-based characters: creating new human performances from a multi-view video database. ACM TOG, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.208, + 0.892, + 0.263 + ], + "angle": 0, + "content": "[54] Weipeng Xu, Avishek Chatterjee, Michael Zollhöfer, Helge Rhodin, Dushyant Mehta, Hans-Peter Seidel, and Christian Theobalt. Monoperfcap: Human performance capture from monocular video. In SIGGRAPH, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.265, + 0.892, + 0.32 + ], + "angle": 0, + "content": "[55] Yuelang Xu, Benwang Chen, Zhe Li, Hongwen Zhang, Lizhen Wang, Zerong Zheng, and Yebin Liu. Gaussian head avatar: Ultra high-fidelity head avatar via dynamic gaussians. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.323, + 0.892, + 0.366 + ], + "angle": 0, + "content": "[56] Wang Yifan, Felice Serena, Shihao Wu, Cengiz Oztireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM TOG, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.367, + 0.892, + 0.409 + ], + "angle": 0, + "content": "[57] Zhengming Yu, Wei Cheng, Xian Liu, Wayne Wu, and Kwan-Yee Lin. Monohuman: Animatable human neural field from monocular video. In CVPR, 2023. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.411, + 0.892, + 0.467 + ], + "angle": 0, + "content": "[58] Hongwen Zhang, Yating Tian, Xinchi Zhou, Wanli Ouyang, Yebin Liu, Limin Wang, and Zhenan Sun. Pymaf: 3d human pose and shape regression with pyramidal mesh alignment feedback loop. In ICCV, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.469, + 0.892, + 0.525 + ], + "angle": 0, + "content": "[59] Hongwen Zhang, Siyou Lin, Ruizhi Shao, Yuxiang Zhang, Zerong Zheng, Han Huang, Yandong Guo, and Yebin Liu. Closet: Modeling clothed humans on continuous surface with explicit template decomposition. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.527, + 0.892, + 0.582 + ], + "angle": 0, + "content": "[60] Hongwen Zhang, Yating Tian, Yuxiang Zhang, Mengcheng Li, Liang An, Zhenan Sun, and Yebin Liu. Pymaf-x: Towards well-aligned full-body model regression from monocular images. IEEE T-PAMI, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.584, + 0.892, + 0.627 + ], + "angle": 0, + "content": "[61] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.628, + 0.892, + 0.696 + ], + "angle": 0, + "content": "[62] Yuxiang Zhang, Hongwen Zhang, Liangxiao Hu, Jiajun Zhang, Hongwei Yi, Shengping Zhang, and Yebin Liu. Proxycap: Real-time monocular full-body capture in world space via human-centric proxy-to-motion learning. In CVPR, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.7, + 0.892, + 0.742 + ], + "angle": 0, + "content": "[63] Hao Zhao, Jinsong Zhang, Yu-Kun Lai, Zerong Zheng, Yingdi Xie, Yebin Liu, and Kun Li. High-fidelity human avatars from a single rgb camera. In CVPR, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.743, + 0.892, + 0.799 + ], + "angle": 0, + "content": "[64] Shunyuan Zheng, Boyao Zhou, Ruizhi Shao, Boring Liu, Shengping Zhang, Liqiang Nie, and Yebin Liu. Gps-gaussian: Generalizable pixel-wise 3d gaussian splattering for real-time human novel view synthesis. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.801, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[65] Yufeng Zheng, Wang Yifan, Gordon Wetzstein, Michael J Black, and Otmar Hilliges. Pointavatar: Deformable point-based head avatars from videos. In CVPR, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[66] Zerong Zheng, Tao Yu, Yebin Liu, and Qionghai Dai. Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction. IEEE T-PAMI, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "643" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[67] Zerong Zheng, Han Huang, Tao Yu, Hongwen Zhang, Yandong Guo, and Yebin Liu. Structured local radiance fields for human avatar modeling. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.178 + ], + "angle": 0, + "content": "[68] Zerong Zheng, Xiaochen Zhao, Hongwen Zhang, Boning Liu, and Yebin Liu. Avatarrex: Real-time expressive full-body avatars. ACM TOG, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.232 + ], + "angle": 0, + "content": "[69] Boyao Zhou, Jean-Sébastien Franco, Federica Bogo, Bugra Tekin, and Edmond Boyer. Reconstructing human body mesh from point clouds by adversarial gp network. In ACCV, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "644" + } + ] +] \ No newline at end of file diff --git a/2024/GaussianAvatar_ Towards Realistic Human Avatar Modeling from a Single Video via Animatable 3D Gaussians/30b3a92a-861e-47e6-8246-2e90d1765e38_origin.pdf b/2024/GaussianAvatar_ Towards Realistic Human Avatar Modeling from a Single Video via Animatable 3D Gaussians/30b3a92a-861e-47e6-8246-2e90d1765e38_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ba0964d0e068e627c867407c3421ddca406ec3b3 --- /dev/null +++ b/2024/GaussianAvatar_ Towards Realistic Human Avatar Modeling from a Single Video via Animatable 3D Gaussians/30b3a92a-861e-47e6-8246-2e90d1765e38_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:751d1523ba65983f35749e3d076fb9cbb7a7ddb2560cd2f7c738552d1ce9755d +size 3351494 diff --git a/2024/GaussianAvatar_ Towards Realistic Human Avatar Modeling from a Single Video via Animatable 3D Gaussians/full.md b/2024/GaussianAvatar_ Towards Realistic Human Avatar Modeling from a Single Video via Animatable 3D Gaussians/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e7271ca8d40c34e4f0b03ee28f4f84c0f82c5de1 --- /dev/null +++ b/2024/GaussianAvatar_ Towards Realistic Human Avatar Modeling from a Single Video via Animatable 3D Gaussians/full.md @@ -0,0 +1,323 @@ +# GaussianAvatar: Towards Realistic Human Avatar Modeling from a Single Video via Animatable 3D Gaussians + +Liangxiao $\mathrm{Hu}^{\dagger,1}$ , Hongwen Zhang $^{2}$ , Yuxiang Zhang $^{3}$ , Boyao Zhou $^{3}$ , Boning Liu $^{3}$ , Shengping Zhang $^{*,1,4}$ , Liqiang Nie $^{1}$ + +1Harbin Institute of Technology 2Beijing Normal University + +$^{3}$ Tsinghua University $^{4}$ Peng Cheng Laboratory + +{lx.hu, s.zhang}@hit.edu.cn, zhanghongwen@bnu.edu.cn, yx-z19@mails.tsinghua.edu.cn + +{bzhou22, liuboning}@mail.tsinghua.edu.cn, nieliqiang@gmail.com + +![](images/e78040bbddb2f4b799711aeaf5810b13030766248e31885acd533c979589b050.jpg) +Figure 1. We propose GaussianAvatar, which learns animatable 3D Gaussians to represent detailed human avatars from a single video. Our method maintains a 3D consistent appearance even when animated by out-of-distribution motions. + +# Abstract + +We present GaussianAvatar, an efficient approach to creating realistic human avatars with dynamic 3D appearances from a single video. We start by introducing animatable 3D Gaussians to explicitly represent humans in various poses and clothing styles. Such an explicit and animatable representation can fuse 3D appearances more efficiently and consistently from 2D observations. Our representation is further augmented with dynamic properties to support pose-dependent appearance modeling, where a dynamic appearance network along with an estimizable feature tensor is designed to learn the motion-to-appearance mapping. Moreover, by leveraging the differentiable motion + +condition, our method enables a joint optimization of motions and appearances during avatar modeling, which helps to tackle the long-standing issue of inaccurate motion estimation in monocular settings. The efficacy of GaussianAvatar is validated on both the public dataset and our collected dataset, demonstrating its superior performances in terms of appearance quality and rendering efficiency. The code and dataset are available at https://github.com/aipixel/GaussianAvatar. + +# 1. Introduction + +Creating a customized human avatar from a single video has great potential for many applications including virtual and augmented reality, the Metaverse, gaming, and movie industries. This task is appealing yet challenging, as the + +monocular observations are highly underdetermined for the modeling of a 3D animatable avatar. Meanwhile, the inaccurate body motion estimations and the complex wrinkle deformations also make it extremely difficult to create a realistic avatar from a single video. + +The modeling of 3D human avatars from monocular videos involves a fusion process of 2D observations to a 3D consistent human model. For this purpose, existing solutions have leveraged both implicit and explicit representations to create the base model of avatars. Among them, implicit methods [6, 15, 16, 49, 57] define a deformable human NeRF to fuse the image observation from current motion space to a canonical space by inverse skinning. However, the implicit 3D volume is inefficient in representing human surfaces and the inverse skinning also introduces ambiguous correspondences during the fusion process. These issues make it hard for implicit solutions to capture fine-grained details of moving people. As the avatar appearances are concentrated around human surfaces, explicit representations are much more efficient in modeling avatars. Previous attempts [12, 63] have employed differentiable mesh rendering to reconstruct the human surface, but these methods struggle to preserve wrinkle details due to a fixed mesh topology. On the other hand, point-based representations [18, 37, 50, 56, 65] are more effective to handle flexible topology but require millions of points to capture detailed appearances. How to represent humans remains one of the fundamental problems for avatar modeling. + +When creating an animatable avatar from a video of moving people, the algorithm is required to learn the relationships between body motions and corresponding appearances. However, the motions estimated from monocular videos are typically faulty, leading to large artifacts in the modeling of dynamic cloth deformations. To address this issue, previous works [1, 15, 16, 42, 45, 49, 57] have attempted to optimize body motions along with the learning of animatable avatar volumes. As the body motions are explicitly represented as parametric meshes, the implicit 3D volume of previous methods makes the optimization indirect and less effective. This issue has even become the main obstacle to achieving high-quality avatar modeling from monocular videos. + +To tackle the above issues, we introduce new representations and solutions to achieve high-quality avatar modeling from a single video. The key insight of our solution is to model dynamic human surfaces explicitly and optimize both the motion and appearances jointly in an end-to-end manner. To this end, we propose GaussianAvatar, a method to reconstruct human avatars with dynamic appearances using the 3D Gaussian representation [17]. As an explicit representation, 3D Gaussian can be easily posed from the canonical space to the motion space via a forward skinning process. Such an animatable 3D Gaus + +sian representation bypasses the inverse skinning process used in the aforementioned NeRF-based methods and overcomes the previous one-to-many issue [4] during canonicalization. Based on the animatable 3D Gaussian, our method can fuse 3D appearances more consistently from 2D observations to a canonical 3D space. To model dynamic human appearances under different poses, we additionally add pose-dependent properties to 3D Gaussians and incorporate them with the canonical human surfaces. Inspired by previous work [28, 46], we learn a dynamic appearance network on the 2D manifolds of the underlying human mesh to predict dynamic properties of 3D Gaussians. However, due to the strong bias of limited training poses, modeling dynamic appearance solely conditioned on pose information struggles to generalize to novel views and poses. To address this issue, we introduce an estimizable feature tensor to capture a coarse global appearance of human avatars. Subsequently, we incorporate pose-dependent effects on the feature tensor to decode fine-grained details such as wrinkles. + +As the proposed animatable 3D Gaussians are differentiable with respect to the motion conditions, it enables a joint optimization of motion and appearances. This merit allows our network to refine the motion along with the avatar modeling process, which helps to tackle the long-standing issue of inaccurate motion estimation in monocular settings. Besides, the refined motion further enhances the accuracy of avatar modeling since the 3D appearance fusion process relies on motion-based skinning. As shown in our experiments, our method is quite robust to initial motion estimation and has the ability to correct the misalignment of motion capture results. + +To summarize, our main contributions are as follows: + +- We introduce animatable 3D Gaussians for realistic human avatar modeling from a single video. By representing human surfaces explicitly, our method can fuse 3D appearances more consistently and efficiently from 2D observations. +- We augment the animatable 3D Gaussians with dynamic properties to support pose-dependent appearance modeling, where a dynamic appearance network along with an estimizable feature tensor is designed to learn the motion-to-appearance mapping. +- We propose to jointly optimize the motion and appearance during the avatar modeling, enabling our method to correct the misalignment of initial motion and improve the final appearance quality. + +# 2. Related Work + +Neural Rendering for Human Reconstruction. Without the need to define a template mesh for avatar modeling [2, 9, 53], neural rendering has emerged as a potent technique that enables learning avatars directly from im + +ages. Here we briefly review precious work that aims to reconstruct humans using neural rendering. + +Due to the high-quality rendering of neural radiance field [29], various efforts [20, 21, 24, 32, 33, 47, 67, 68] have been made to reconstruct the dynamic appearance of moving people. Neural Body [33] associates a latent code to each SMPL [26] vertex to encode the appearance, which is transformed into observation space based on the human pose. Neural Actor [24] learns a deformable radiance field with SMPL as guidance and utilizes a texture map to improve its final rendering quality. TAVA [20] proposes to jointly model the non-rigid warping field and shading effects directly conditioned on the pose vectors. Posevocab [21] designs joint-structured pose embeddings to encode the dynamic appearances under different key poses, such embeddings can better learn joint-related appearance. NeRF-based methods have demonstrated appealing rendering results on human avatar reconstruction, but still struggle to represent human surfaces with the implicit 3D volume. Explicit modeling of human surfaces is a more straightforward way for this task, as the dynamic appearance of humans is mostly reflected on the human surface. + +Explicit representations have great potential for human reconstruction. In HF-Avatar [63], meshes are used as the base representation in a coarse-to-fine framework that combines neural texture with dynamic surface deformation for avatar creation. EMA [12] proposes Meshy neural fields to reconstruct human avatars by optimizing the canonical mesh, material, and motion dynamics through inverse rendering in an end-to-end process. PointAvatar [65] employs a deformable point-based representation to separate source color into intrinsic albedo and normal-dependent shading. DVA [35] extends mixtures of volumetric primitives [25] for human avatar modeling. All these attempts have demonstrated the significant potential of explicit representations and their under-exploration. However, meshes are constrained by fixed topologies, whereas point clouds demand a multitude of points to encompass intricate details. 3D Gaussians [17] have showcased their capability in various human tasks [22, 40, 55, 64]. In essence, 3D Gaussians hold promise for human avatar reconstruction and are currently a subject of active research. + +Avatar Modeling from Monocular Videos. Numerous methods investigate reconstructing humans from single images or monocular videos. Regression-based methods[10, 11, 13, 38, 39, 51, 52, 66] directly recover clothed 3D humans just from a single image. While these methods produce attractive results, they can not recover the dynamic appearance in the reconstruction across the entire sequence. Traditional methods aim to capture human dynamics by tracking individuals in videos using pre-scanned rigged templates [7, 8, 54]. However, the pre-scanning + +and manual rigging processes prevent their real-life applications. [1, 5, 30] try to bypass the requirement for predefined human models but face challenges in preserving fine details because of the fixed mesh resolution. + +The emergence of neural radiance fields [29] has facilitated the creation of various techniques for reconstructing animatable avatars from monocular videos [3, 14, 16, 41-43, 49]. However, inaccuracies in estimating human motions from monocular videos result in pronounced artifacts. To address the problem, HumanNeRF [49] solves for an update to the inaccurate poses. NeuMan [16] introduces an error-correction network to enable training with erroneous estimates. Vid2Avatar[6] and InstantAvatar [15] also jointly optimize motions by back-propagating the gradient of the image reconstruction loss to the pose parameters. MonoHuman [57] introduces bi-directional constraints to alleviate ambiguous correspondence on novel poses. Attempting to address the inaccurate pose estimation issue with implicit human NeRF models has proven to be inefficient and imprecise. Consequently, we endeavor to address this issue by leveraging an explicit 3D Gaussian representation. + +# 3. Method + +Our goal is to create a human avatar that enables free-viewpoint rendering and realistic animation using a single video. We first introduce an expressive representation, namely animatable 3D Gaussians, to represent human avatars in Sec. 3.1. With this representation, modeling the dynamic appearance of humans can be regarded as dynamic 3D Gaussian property estimation. Then we build a dynamic appearance network along with an estimable feature tensor to learn the motion-to-appearance mapping in Sec. 3.2. To mitigate the artifacts caused by inaccurate pose estimation, we adopt a joint motion and appearance optimization approach for refining human poses during training (Sec. 3.3). In the following, we provide a comprehensive explanation of the technical details. + +# 3.1. Animatable 3D Gaussians + +Point-based representation [23, 28, 34, 59, 69] has proven its topological flexibility in generating realistic human avatars from scans. Extending this explicit representation to create human avatars from images is a significant endeavor. With this goal in mind, we introduce a novel representation, termed animatable 3D Gaussians, which effectively reconstructs human surfaces with a 3D consistent appearance. + +3D Gaussian Splitting [17] is a point-based scene representation that allows high-quality real-time rendering. The scene representation is parameterized by a set of static 3D Gaussians, each of which has the following parameters: 3D center position $\mathbf{x} \in \mathbb{R}^3$ , color $\mathbf{c} \in \mathbb{R}^3$ , opacity $\alpha \in \mathbb{R}$ , 3D rotation in form of quaternion $\mathbf{q} \in \mathbb{R}^4$ and 3D scaling factor $\mathbf{s} \in \mathbb{R}^3$ . With these properties, we can generate rendered + +![](images/26330d876fda00891ab4b8a45025625ad03c508ee6cef6ff63f263f78a64a74b.jpg) +Figure 2. Overview of GaussianAvatar. Given a fitted SMPL or SMPL-X model on the current frame, we sample the points on its surface and record their positions on a UV positional map $I$ , which is then passed to a pose encoder to obtain the pose feature. An estimizable feature tensor is pixel-aligned with the pose feature and learned to capture the coarse appearance of humans. Then the two aligned feature tensors are input into the Gaussian parameter decoder, which predicts each point's offset $\Delta \hat{\mathbf{x}}$ , color $\hat{\mathbf{c}}$ , and scale $\hat{s}$ . These predictions, along with the fixed rotations $\mathbf{q}$ and opacity $\alpha$ , collectively constitute the animatable 3D Gaussians in canonical space. + +images from any viewpoint. We refer readers to [17, 27] for the rendering details of 3D Gaussians. + +To extend this representation for human avatar modeling, we integrate it with either the SMPL [26] or SMPL-X [31] model as follows: + +$$ +G (\pmb {\beta}, \pmb {\theta}, \mathbf {D}, \mathbf {P}) = S p l a t t i n g (W (\mathbf {D}, J (\pmb {\beta}), \pmb {\theta}, \omega), \mathbf {P}), (1) +$$ + +where $G(\cdot)$ represents a rendered image, and $Splatting(\cdot)$ denotes the rendering process of 3D Gaussians from any viewpoint, $W(\cdot)$ is a standard linear blend skinning function employed for reposing 3D Gaussians, $\mathbf{D} = T(\boldsymbol{\beta}) + dT$ represents the locations of 3D Gaussians in canonical space, formed by adding corrective point displacements $dT$ on the template mesh surface $T(\beta)$ , $\mathbf{P}$ denotes the remaining properties of 3D Gaussians, excluding the positions. $\beta$ and $\theta$ are the shape and pose parameters, $J(\beta)$ outputs 3D joint locations. Note that we propagate the skinning weight $\omega$ from the vertices of the SMPL or SMPL-X model to the nearest 3D Gaussians. With the proposed representation, we can now repose these canonical 3D Gaussians to the motion space for free-view rendering. + +# 3.2. Dynamic 3D Gaussian Property Estimation + +Following the proposed animatable Gaussians, human appearances are determined by the point displacements $dT$ and properties $\mathbf{P}$ . Modeling dynamic human appearances can be regarded as estimating these dynamic properties. To model dynamic human appearances under various poses, we introduce a dynamic appearance network along with an estimizable feature tensor to predict these pose-dependent properties of 3D Gaussians. Despite sharing similarities in network structure with [28], we present this framework for a distinct purpose. In [28], the feature tensor serves to decouple the pose-independent human shape from the decoder, and we repurpose it to capture a coarse global appearance + +![](images/460762fcaac45792567d669584836768ed1f08403826bda9b989f98bccf7a450.jpg) +(a) + +![](images/504c07309e5eb4d0857085dea30ea95c4095228d8fbbe4ac05908ea4f34a72f1.jpg) +(b) + +![](images/19a40eee079f096e44a7e5d285e4cf533fd72d2157fea90a5023c365cc358594.jpg) +(c) + +![](images/6155fdef5b9e0184ee5365daa6468f4f474dfdfa0965aedaf128dd841b127a6a.jpg) +(d) +Figure 3. Effect of iostropy of 3D Gaussians. (a) Input image, (b)(d) front and back views trained with isotropic 3D Gaussians, (c)(e) front and back views trained with anisotropic 3D Gaussians. + +![](images/e2785d0a2f35a1401d0267206f2370467c2bb39c30f6d4534495c815d5808177.jpg) +(e) + +of human avatars. The motivation behind this modification is that directly learning a mapping from human poses to dynamic properties is susceptible to overfitting on the limited training poses. To integrate the global appearance into the feature tensor, we introduce a two-stage training strategy, as discussed in Sec. 3.4. + +The dynamic appearance network is designed to learn a mapping from a 2D manifold representing the underlying human shape to the dynamic properties of 3D Gaussians as follows: + +$$ +f _ {\phi}: \mathcal {S} ^ {2} \in \mathbb {R} ^ {3} \rightarrow \mathbb {R} ^ {7}. \tag {2} +$$ + +As shown in Fig. 2, the 2D human manifold $S^2$ is depicted by a UV positional map $I \in R^{H \times W \times 3}$ , where each valid pixel stores the position $(x, y, z)$ of one point on the posed body surface. The final predictions consist of per point offset $\Delta \hat{\mathbf{x}} \in \mathbb{R}^3$ , color $\hat{\mathbf{c}} \in \mathbb{R}^3$ , and scale $\hat{s} \in \mathbb{R}$ on the canonical surface. + +Instead of predicting all properties, we make these slight adjustments to our task based on experimental results. Due to the unbalanced viewpoints in monocular videos, + +anisotropic 3D Gaussians are prone to learning an inaccurate 3D shape to fit the most frequently seen view, resulting in poorer performance in side views. As shown in Fig. 3, training our network with isotropic 3D Gaussians yields superior results. We ensure isotropy among all 3D Gaussians by maintaining uniform size across dimensions. Therefore, we introduce a scalar $\hat{s}$ to represent the scales of 3D Gaussians and set rotations $\mathbf{q}$ as $[1,0,0,0]$ . Through experimental observations, we notice that the network tends to learn an opacity value $(\alpha)$ of zero on the boundary to correct the human shape. To address this, we fix opacity $\alpha = 1$ to keep all 3D Gaussians visible, enforcing the network to predict accurate positions of the 3D Gaussians. + +The dynamic appearance network consists of two parts: a pose encoder and a Gaussian parameter decoder. The pose encoder takes the UV positional map of posed body points as input to generate a pose-conditioned feature tensor $O \in \mathbb{R}^{H \times W \times C}$ . We then integrate the pixel-aligned optimized feature tensor $F \in \mathbb{R}^{H \times W \times C}$ with the pose features before feeding it into the Gaussian parameter decoder to generate final predictions. Following this, we add the predicted offsets to the canonical 3D Gaussians and associate the predicted properties with the corresponding 3D Gaussians. With the estimated poses, we can then repose the canonical 3D Gaussians to the motion space for rendering. + +# 3.3. Joint Motion and Appearance Optimization + +However, owing to the imprecise estimation of human poses $\pmb{\theta} = (\theta_{1},\theta_{2},\dots,\theta_{J})$ and translations $\mathbf{t}$ from monocular videos, the reposed 3D Gaussians in motion space are inaccurately represented and may lead to unsatisfactory rendering outcomes. To address this, we propose to jointly optimize human motions and appearances. To optimize human motions with image loss, we solve for an update $(\Delta \pmb {\theta},\Delta \mathbf{t})$ to the estimated body poses and translations as follows: + +$$ +\hat {\boldsymbol {\Theta}} = (\boldsymbol {\theta} + \Delta \boldsymbol {\theta}, \mathbf {t} + \Delta \mathbf {t}). \tag {3} +$$ + +We modify $\theta$ in Eq. 1 using $\Theta$ to render the proposed animatable 3D Gaussians differentiable with respect to the motion conditions. Different from previous work [6, 16, 49] that jointly optimizes human poses via inverse skinning, we optimize the updates in a forward skinning process, which benefits both motion and appearance optimization. + +# 3.4. Training Strategy + +In this section, we outline our approach to training the network with inaccurate human motions. We conduct a two-stage optimization process using different loss functions. In the first stage, we aim to fuse the sequential appearances to the estimizable feature tensor and conduct motion optimization to get accurate poses for the dynamic appearance network. In this stage, we optimize the framework without incorporating any pose-dependent information by excluding + +the training of the pose encoder. Specifically, we utilize the following loss functions to train our network: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {s t a g e} _ {1}} = \lambda_ {r b g} \mathcal {L} _ {r b g} + \lambda_ {s s i m} \mathcal {L} _ {s s i m} + \lambda_ {l p i p s} \mathcal {L} _ {l p i p s} \tag {4} \\ + \lambda_ {f} \mathcal {L} _ {f} + \lambda_ {\text {o f f s e t}} \mathcal {L} _ {\text {o f f s e t}} + \lambda_ {\text {s c a l e}} \mathcal {L} _ {\text {s c a l e}}, \\ \end{array} +$$ + +where $\mathcal{L}_{rbg}$ , $\mathcal{L}_{ssim}$ , and $\mathcal{L}_{lpips}$ are the L1 loss, SSIM loss [48], and LPIPS loss [61], respectively. $\mathcal{L}_f$ , $\mathcal{L}_{offset}$ , $\mathcal{L}_{scale}$ calculate the L2-norm of the feature map, predicted offsets and scales, respectively. We set $\lambda_{rbg} = 0.8$ , $\lambda_{ssim} = 0.2$ , $\lambda_{lpips} = 0.2$ , $\lambda_f = 1$ , $\lambda_{offset} = 10$ , $\lambda_{scale} = 1$ . + +After the first stage of training, we obtain more accurate human motions and an optimized feature tensor $F$ . The optimized feature tensor $F$ captures a coarse appearance of human avatars. In the second stage, we incorporate the pose features encoded by the pose encoder with the trained feature tensor $F$ . We replace $\mathcal{L}_f$ with the L2-norm loss $\mathcal{L}_p$ , which plays the same role as $\mathcal{L}_f$ in regularizing the limited pose space. By penalizing the pose-dependent features, we can eliminate the strong bias of limited training poses and thus generalize to unseen viewpoints and poses. + +# 4. Experiments + +# 4.1. Datasets and Metrics + +People-Snapshot Dataset. This dataset [1] comprises videos of individuals rotating in front of a stationary camera. To ensure a fair quantitative comparison, we follow the same evaluation protocol outlined in InstantAvatar [15]. + +NeuMan Dataset. To assess even more challenging scenarios, we employ outdoor collections from the NeuMan dataset [16]. These videos are recorded using a mobile phone for moving individuals. Specifically, we select four sequences (bike, citizen, Seattle, jogging) that exhibit most body regions and contain minimal blurry images. We initialize the estimated poses with an off-the-shelf method [44], which is also utilized in [15]. + +DynVideo Dataset. With the limited cloth deformation presented in the above two datasets, we propose the Dynvideo dataset to capture dynamic human appearance. To this end, we utilize a mobile phone to record videos of a character performing various movements, especially rotation, in front of the device. Each of these videos takes about one minute and provides a comprehensive and detailed representation of human movement. We also provide the corresponding SMPL parameter sequences for all videos, obtained by our proposed method. This dataset serves as a valuable resource for evaluating reconstruction quality, with a particular emphasis on dynamic appearances. + +Evaluation Metrics. We consider three metrics: PSNR, SSIM [48], and LPIPS [61] to access the reconstruction quality on three datasets. + +
Methodmale-3-casualmale-4-casualfemale-3-casualfemale-4-casual
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
HumanNeRF [49]26.900.96050.018125.500.93970.035724.460.95160.026927.070.96150.0152
InstantAvatar [15]29.530.97160.015527.670.96260.030727.660.97090.021029.110.96830.0167
Baseline27.710.97130.021825.090.96140.030624.800.96090.031026.160.96100.0227
Baseline + Opt.30.980.97910.014328.770.97530.023029.590.97680.022030.830.97680.0145
Baseline + Opt. + Dyn.30.980.97900.014528.780.97550.022829.550.97620.022530.840.97710.0140
+ +Table 1. Quantitative evaluation on the People-Snapshot [1] dataset. Opt. denotes the motion optimization and Dyn. refers to dynamic appearance modeling. + +
MethodPSNR ↑SSIM ↑LPIPS ↓
HumanNeRF [49]27.060.96690.0192
InstantAvatar [15]28.470.97150.0277
Baseline27.060.96920.0183
Baseline + Opt.29.930.97940.0126
Baseline + Opt. + Dyn.29.940.97950.0124
+ +Table 2. Quantitative evaluation on NeuMan [16] dataset. + +
MethodPSNR ↑SSIM ↑LPIPS ↓
HumanNeRF [49]24.390.93490.0349
InstantAvatar [15]20.310.91830.1046
Baseline23.550.93400.0354
Baseline + Opt.26.160.95220.0237
Baseline + Opt. + Dyn.28.580.96160.0169
+ +Table 3. Quantitative evaluation on DynVideo dataset. + +![](images/53a69b099812bce6361a1a404aedc0d370eee4bf278ca822a7380654c7ab654e.jpg) +(a) +Figure 4. Motion optimization results. (a)(d) Original image, (b)(e) our optimized SMPL, (c)(f) ROMP [44] estimates. + +![](images/efd5c94998dee1d0fe52f5c704e4f0e6d22feb6dd7cb275b820ab8da6000209b.jpg) +(b) + +![](images/e913b401eb4b088d37fc059217f7dde737b89f15cc2605a2ba4dfe51097849ac.jpg) +(c) + +![](images/6c5cfc78e55dbcebbdfbbfe3ca57a466ea619cd89213370d6501cd7527fcd977.jpg) +(d) + +![](images/7df6d01934acbddd2669f26b3ed4834657cf5e4d869545cc0f80a139848dd10a.jpg) +(e) + +![](images/f387d007031e81449e82fc9f40b230bd1e8b8aea88bf0342dfdb45261aaef90b.jpg) +(f) + +# 4.2. Experimental Settings + +Implementation details. We employ a U-Net [36] for extracting pose-dependent features, and the Gaussian parameter decoder is implemented as an 8-layer multilayer perceptron (MLP). We sample approximately 200,000 points on the SMPL mesh surface. The entire framework is trained on a single NVIDIA RTX 3090 GPU, with training times ranging from 0.5 to 6 hours. + +Baseline. To showcase the efficacy of the proposed mod + +![](images/6a03ee5740839210e69ab2df20dfc0bc36461068aaa8f2ca98e4ad2a5f5b77fb.jpg) + +![](images/e617c07bc6d3c71539c702e09c6052177451ec63859c9d93a4c8934fdef87f95.jpg) +(a) + +![](images/474aa8c9485bc9fdb7e8fcd4d3eb1c0ccd7f3f84f9d607e0108bb13af96db098.jpg) +(b) +Figure 5. Qualitative ablation studies. (a) Ground truth, (b) baseline + Opt. + Dyn., (c) baseline + Opt., (d) baseline. + +![](images/ae64d062660dc7f762a9673ed9e20ea2eef4bd48db0accf99e0bf0fb46699caa.jpg) + +![](images/fb82564be95791f788df7d5b2ee4d9ff2d7f16ac0b1d81352ea14146fea9b511.jpg) +(c) + +![](images/f9cf3cdef3710d38af1acc9f23a9387c7c7f50a73eb54546d54a68313878d137.jpg) + +![](images/fe8ad7e17cf8f04981501fd9a8f6dd6bff1fab79f27b28aece4fefd9a1757966.jpg) +(d) + +ules, our method can be partitioned into three components: baseline, motion optimization, and dynamic appearance modeling. As our baseline, we suspend the motion optimization process and exclude the pose-dependent information from the pose encoder. + +Methods for Comparison. We compare our method against (1) HumanNeRF [49], which implicitly represents a human avatar with a canonical appearance neural field and a motion field; (2) InstantAvatar [15], which achieves fast avatar modeling by using several acceleration strategies for neural fields. + +# 4.3. Comparisons with the State of the Art + +We report the numeric evaluation results on three datasets. As shown in Table 1, 2, and 3, our proposed method outperforms all baselines on all metrics for recovering more details of dynamic appearance and correcting the artifacts caused by initialized poses. To demonstrate the qualitative evaluation of these datasets, we also visualize the novel view synthesis results on the test splits. As implicit representations, HumanNeRF and InstantAvatar are prone to generate ghosting effects at boundary areas, as shown in Fig. 6. InstantAvatar lacks the capability to model pose + +![](images/3307401eced9c5d32e29fda4fbbfa79b0db791157a873df09435bf499faec07c.jpg) +Figure 6. Qualitative comparison of novel view synthesis. We compare the novel view synthesis quality on the People-Snapshot dataset (first row), NeuMan dataset (second row), and DynVideo dataset (last two rows). + +![](images/8e2a9f2db1de5d5c275a29e7874b1600f18cd2110e5e9ecb06c8aa7306e48897.jpg) + +dependent deformations, leading to challenges in effectively handling the dynamic appearance of moving people. + +To showcase the robustness of our avatar modeling, we collect more challenging poses [19] captured by the monocular camera and evaluate our method and other methods for avatar animation and novel view synthesis. As ground truth + +is unavailable for these out-of-distribution poses, we illustrate qualitative results of novel pose and view synthesis in Fig. 1 and Fig. 7. Our method generates realistic animation results with these challenging poses, demonstrating a consistent 3D appearance in novel views with respect to other methods. + +![](images/56a851efbb927e2ca51d324b297af18c839938f557224fde54d862a77573b7f5.jpg) +Figure 7. Animation results on out-of-distribution motions. We compare the animation results on the People-Snapshot dataset (first row), NeuMan dataset (second row), and DynVideo dataset (last row). + +# 4.4. Ablation Studies + +In this section, we conduct ablation studies to validate each component of our methods. As shown in Table 1, 2, and 3, our proposed motion optimization module dramatically improves over the baseline on all metrics, demonstrating its effectiveness in modeling human avatars. To highlight the effectiveness of our method in human motion optimization, we illustrate the initialized poses obtained by ROMP [44] and the optimized ones by our method on NeuMan and DynVideo datasets in Fig. 4. Experiments demonstrate that joint motion optimization is capable of correcting inaccurate motion estimation, even for the side and back views. Furthermore, our approach readily extends to enhance the accuracy of existing motion capture methods [58, 60, 62]. + +We also notice that the dynamic appearance modeling achieves superior results on NeuMan and DynVideo datasets in Table 2 and 3 and slightly favorable outcomes in Table 1, considering that People-Snapshot has limited pose variation with respect to the other two datasets. In Fig. 5, we further illustrate the visible improvements of each component of our method. The motion optimization scheme improves the global quality of the baseline for better pose estimation. The dynamic appearance modeling additionally preserves pose-dependent details such as cloth wrinkles on the human surface. + +# 5. Conclusion and Discussion + +We introduce GaussianAvatar, a human avatar reconstruction method based on the proposed animatable 3D Gaussians from monocular videos. For dynamic human appearance modeling, we leverage a dynamic appearance network along with an estimizable feature tensor to enhance the representation with dynamic properties. Besides, we implement a joint motion and appearance optimization scheme to rectify estimated motion and enhance the overall reconstruction quality. Our method shows the capability to reconstruct avatars with dynamic appearances, enabling realistic animation while maintaining real-time rendering speed. + +Limitation. Similar to [15, 49, 57], our method may generate artifacts due to inaccurate foreground segmentations in videos and encounter challenges in modeling loose outfits such as dresses. + +Potential Social Impact. Given our method's capability to reconstruct a realistic personalized character from a monocular video, it is imperative to exercise caution and consider the potential for technology misuse. + +Acknowledgement. This work was supported by the NSFC project (Nos. 62272134, 62236003, 62072141, 62301298, and 62125107), Shenzhen College Stability Support Plan (Grant No. GXWD20220817144428005), and the Major Key Project of PCL (PCL2023A10-2). + +# References + +[1] Thiemo Alldieck, Marcus Magnor, Weipeng Xu, Christian Theobalt, and Gerard Pons-Moll. Video based reconstruction of 3d people models. In CVPR, 2018. 2, 3, 5, 6 +[2] Timur Bagautdinov, Chenglei Wu, Tomas Simon, Fabian Prada, Takaaki Shiratori, Shih-En Wei, Weipeng Xu, Yaser Sheikh, and Jason Saragih. Driving-signal aware full-body avatars. ACM TOG, 2021. 2 +[3] Jianchuan Chen, Ying Zhang, Di Kang, Xuefei Zhe, Linchao Bao, Xu Jia, and Huchuan Lu. Animatable neural radiance fields from monocular rgb videos. arXiv preprint arXiv:2106.13629, 2021. 3 +[4] Xu Chen, Yufeng Zheng, Michael J Black, Otmar Hilliges, and Andreas Geiger. Snarf: Differentiable forward skinning for animating non-rigid neural implicit shapes. In ICCV, 2021. 2 +[5] Chen Guo, Xu Chen, Jie Song, and Otmar Hilliges. Human performance capture from monocular video in the wild. In 3DV, 2021. 3 +[6] Chen Guo, Tianjian Jiang, Xu Chen, Jie Song, and Otmar Hilliges. Vid2 avatar: 3d avatar reconstruction from videos in the wild via self-supervised scene decomposition. In CVPR, 2023. 2, 3, 5 +[7] Marc Habermann, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Livecap: Real-time human performance capture from monocular video. ACM TOG, 2019. 3 +[8] Marc Habermann, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Deepcap: Monocular human performance capture using weak supervision. In CVPR, 2020. 3 +[9] Marc Habermann, Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Real-time deep dynamic characters. ACM TOG, 2021. 2 +[10] Tong He, John Collomosse, Hailin Jin, and Stefano Soatto. Geo-pifu: Geometry and pixel aligned implicit functions for single-view human reconstruction. In NeurIPS, 2020. 3 +[11] Tong He, Yuanlu Xu, Shunsuke Saito, Stefano Soatto, and Tony Tung. Arch++: Animation-ready clothed human reconstruction revisited. In ICCV, 2021. 3 +[12] Xiaoke Huang, Yiji Cheng, Yansong Tang, Xiu Li, Jie Zhou, and Jiwen Lu. Efficient meshy neural fields for animatable human avatars. arXiv preprint arXiv:2303.12965, 2023. 2, 3 +[13] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. Arch: Animatable reconstruction of clothed humans. In CVPR, 2020. 3 +[14] Boyi Jiang, Yang Hong, Hujun Bao, and Juyong Zhang. Self-con: Self reconstruction your digital avatar from monocular video. In CVPR, 2022. 3 +[15] Tianjian Jiang, Xu Chen, Jie Song, and Otmar Hilliges. Instantavatar: Learning avatars from monocular video in 60 seconds. In CVPR, 2023. 2, 3, 5, 6, 8 +[16] Wei Jiang, Kwang Moo Yi, Golnoosh Samei, Oncel Tuzel, and Anurag Ranjan. Neuman: Neural human radiance field from a single video. In ECCV, 2022. 2, 3, 5, 6 + +[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM TOG, 2023. 2, 3, 4 +[18] Christoph Lassner and Michael Zollhofer. Pulsar: Efficient sphere-based neural rendering. In CVPR, 2021. 2 +[19] Ruilong Li, Shan Yang, David A. Ross, and Angjoo Kanazawa. Learn to dance with aist++: Music conditioned 3d dance generation. In ICCV, 2021. 7 +[20] Ruilong Li, Julian Tanke, Minh Vo, Michael Zollhöfer, Jürgen Gall, Angjoo Kanazawa, and Christoph Lassner. Tava: Template-free animatable volumetric actors. In ECCV, 2022. 3 +[21] Zhe Li, Zerong Zheng, Yuxiao Liu, Boyao Zhou, and Yebin Liu. Posevocab: Learning joint-structured pose embeddings for human avatar modeling. In SIGGRAPH, 2023. 3 +[22] Zhe Li, Zerong Zheng, Lizhen Wang, and Yebin Liu. Animatable gaussians: Learning pose-dependent gaussian maps for high-fidelity human avatar modeling. In CVPR, 2024. 3 +[23] Siyou Lin, Hongwen Zhang, Zerong Zheng, Ruizhi Shao, and Yebin Liu. Learning implicit templates for point-based clothed human modeling. In ECCV, 2022. 3 +[24] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. ACM TOG, 2021. 3 +[25] Stephen Lombardi, Tomas Simon, Gabriel Schwartz, Michael Zollhoefer, Yaser Sheikh, and Jason Saragih. Mixture of volumetric primitives for efficient neural rendering. ACM Trans. Graph., 2021. 3 +[26] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. ACM TOG, 2015. 3, 4 +[27] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. 3DV, 2024. 4 +[28] Qianli Ma, Jinlong Yang, Siyu Tang, and Michael J Black. The power of points for modeling humans in clothing. In ICCV, 2021. 2, 3, 4 +[29] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 3 +[30] Gyeongsik Moon, Hyeongjin Nam, Takaaki Shiratori, and Young Mu Lee. 3d clothed human reconstruction in the wild. In ECCV, 2022. 3 +[31] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3d hands, face, and body from a single image. In CVPR, 2019. 4 +[32] Sida Peng, Shangzhan Zhang, Zhen Xu, Chen Geng, Boyi Jiang, Hujun Bao, and Xiaowei Zhou. Animatable neural implicit surfaces for creating avatars from videos. In ICCV, 2021. 3 +[33] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In CVPR, 2021. 3 + +[34] Sergey Prokudin, Qianli Ma, Maxime Raafat, Julien Valentin, and Siyu Tang. Dynamic point fields. In ICCV, 2023. 3 +[35] Edoardo Remelli, Timur Bagautdinov, Shunsuke Saito, Chenglei Wu, Tomas Simon, Shih-En Wei, Kaiwen Guo, Zhe Cao, Fabian Prada, Jason Saragih, et al. Drivable volumetric avatars using texel-aligned features. In SIGGRAPH, 2022. 3 +[36] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 6 +[37] Darius Rückert, Linus Franke, and Marc Stamminger. Adop: Approximate differentiable one-pixel point rendering. ACM TOG, 2022. 2 +[38] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In ICCV, 2019. 3 +[39] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In CVPR, 2020. 3 +[40] Ruizhi Shao, Jingxiang Sun, Cheng Peng, Zerong Zheng, Boyao Zhou, Hongwen Zhang, and Yebin Liu. Control4d: Efficient 4d portrait editing with text. In CVPR, 2024. 3 +[41] Shih-Yang Su, Frank Yu, Michael Zollhöfer, and Helge Rhodin. A-nerf: Articulated neural radiance fields for learning human shape, appearance, and pose. In NeurIPS, 2021. 3 +[42] Shih-Yang Su, Timur Bagautdinov, and Helge Rhodin. Danbo: Disentangled articulated neural body representations via graph neural networks. In ECCV, 2022. 2 +[43] Shih-Yang Su, Timur Bagautdinov, and Helge Rhodin. Npc: Neural point characters from video. In ICCV, 2023. 3 +[44] Yu Sun, Qian Bao, Wu Liu, Yili Fu, Michael J Black, and Tao Mei. Monocular, one-stage, regression of multiple 3d people. In ICCV, 2021. 5, 6, 8 +[45] Gusi Te, Xiu Li, Xiao Li, Jinglu Wang, Wei Hu, and Yan Lu. Neural capture of animatable 3d human from monocular video. In ECCV, 2022. 2 +[46] Maria Vakalopoulou, Guillaume Chassagnon, Norbert Bus, Rafael Marini, Evangelia I Zacharaki, M-P Revel, and Nikos Paragios. Atlasnet: Multi-atlas non-linear deep networks for medical image segmentation. In MICCAI, 2018. 2 +[47] Shaofei Wang, Katja Schwarz, Andreas Geiger, and Siyu Tang. Arah: Animatable volume rendering of articulated human sdfs. In ECCV, 2022. 3 +[48] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 2004. 5 +[49] Chung-Yi Weng, Brian Curless, Pratul P Srinivasan, Jonathan T Barron, and Ira Kemelmacher-Shlizerman. Humaner: Free-viewpoint rendering of moving people from monocular video. In CVPR, 2022. 2, 3, 5, 6, 8 +[50] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. Synsin: End-to-end view synthesis from a single image. In CVPR, 2020. 2 +[51] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J Black. Icon: Implicit clothed humans obtained from normals. In CVPR, 2022. 3 + +[52] Yuliang Xiu, Jinlong Yang, Xu Cao, Dimitrios Tzionas, and Michael J Black. Econ: Explicit clothed humans optimized via normal integration. In CVPR, 2023. 3 +[53] Feng Xu, Yebin Liu, Carsten Stoll, James Tompkin, Gaurav Bharaj, Qionghai Dai, Hans-Peter Seidel, Jan Kautz, and Christian Theobalt. Video-based characters: creating new human performances from a multi-view video database. ACM TOG, 2011. 2 +[54] Weipeng Xu, Avishek Chatterjee, Michael Zollhöfer, Helge Rhodin, Dushyant Mehta, Hans-Peter Seidel, and Christian Theobalt. Monoperfcap: Human performance capture from monocular video. In SIGGRAPH, 2018. 3 +[55] Yuelang Xu, Benwang Chen, Zhe Li, Hongwen Zhang, Lizhen Wang, Zerong Zheng, and Yebin Liu. Gaussian head avatar: Ultra high-fidelity head avatar via dynamic gaussians. In CVPR, 2024. 3 +[56] Wang Yifan, Felice Serena, Shihao Wu, Cengiz Oztireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM TOG, 2019. 2 +[57] Zhengming Yu, Wei Cheng, Xian Liu, Wayne Wu, and Kwan-Yee Lin. Monohuman: Animatable human neural field from monocular video. In CVPR, 2023. 2, 3, 8 +[58] Hongwen Zhang, Yating Tian, Xinchi Zhou, Wanli Ouyang, Yebin Liu, Limin Wang, and Zhenan Sun. Pymaf: 3d human pose and shape regression with pyramidal mesh alignment feedback loop. In ICCV, 2021. 8 +[59] Hongwen Zhang, Siyou Lin, Ruizhi Shao, Yuxiang Zhang, Zerong Zheng, Han Huang, Yandong Guo, and Yebin Liu. Closet: Modeling clothed humans on continuous surface with explicit template decomposition. In CVPR, 2023. 3 +[60] Hongwen Zhang, Yating Tian, Yuxiang Zhang, Mengcheng Li, Liang An, Zhenan Sun, and Yebin Liu. Pymaf-x: Towards well-aligned full-body model regression from monocular images. IEEE T-PAMI, 2023. 8 +[61] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5 +[62] Yuxiang Zhang, Hongwen Zhang, Liangxiao Hu, Jiajun Zhang, Hongwei Yi, Shengping Zhang, and Yebin Liu. Proxycap: Real-time monocular full-body capture in world space via human-centric proxy-to-motion learning. In CVPR, 2024. 8 +[63] Hao Zhao, Jinsong Zhang, Yu-Kun Lai, Zerong Zheng, Yingdi Xie, Yebin Liu, and Kun Li. High-fidelity human avatars from a single rgb camera. In CVPR, 2022. 2, 3 +[64] Shunyuan Zheng, Boyao Zhou, Ruizhi Shao, Boring Liu, Shengping Zhang, Liqiang Nie, and Yebin Liu. Gps-gaussian: Generalizable pixel-wise 3d gaussian splattering for real-time human novel view synthesis. In CVPR, 2024. 3 +[65] Yufeng Zheng, Wang Yifan, Gordon Wetzstein, Michael J Black, and Otmar Hilliges. Pointavatar: Deformable point-based head avatars from videos. In CVPR, 2023. 2, 3 +[66] Zerong Zheng, Tao Yu, Yebin Liu, and Qionghai Dai. Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction. IEEE T-PAMI, 2021. 3 + +[67] Zerong Zheng, Han Huang, Tao Yu, Hongwen Zhang, Yandong Guo, and Yebin Liu. Structured local radiance fields for human avatar modeling. In CVPR, 2022. 3 +[68] Zerong Zheng, Xiaochen Zhao, Hongwen Zhang, Boning Liu, and Yebin Liu. Avatarrex: Real-time expressive full-body avatars. ACM TOG, 2023. 3 +[69] Boyao Zhou, Jean-Sébastien Franco, Federica Bogo, Bugra Tekin, and Edmond Boyer. Reconstructing human body mesh from point clouds by adversarial gp network. In ACCV, 2020. 3 \ No newline at end of file diff --git a/2024/GaussianAvatar_ Towards Realistic Human Avatar Modeling from a Single Video via Animatable 3D Gaussians/images.zip b/2024/GaussianAvatar_ Towards Realistic Human Avatar Modeling from a Single Video via Animatable 3D Gaussians/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1e8561c17e4bbe9ad4ecee100481f8facad87a3b --- /dev/null +++ b/2024/GaussianAvatar_ Towards Realistic Human Avatar Modeling from a Single Video via Animatable 3D Gaussians/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e9914406165f7169b63260f9f3f9cfb3dfb6970eea47309e3c696b1752cf60aa +size 669840 diff --git a/2024/GaussianAvatar_ Towards Realistic Human Avatar Modeling from a Single Video via Animatable 3D Gaussians/layout.json b/2024/GaussianAvatar_ Towards Realistic Human Avatar Modeling from a Single Video via Animatable 3D Gaussians/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..37b955762dc513a8efeb7131ecf2d4421b652192 --- /dev/null +++ b/2024/GaussianAvatar_ Towards Realistic Human Avatar Modeling from a Single Video via Animatable 3D Gaussians/layout.json @@ -0,0 +1,9051 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 108, + 103, + 486, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 103, + 486, + 140 + ], + "spans": [ + { + "bbox": [ + 108, + 103, + 486, + 140 + ], + "type": "text", + "content": "GaussianAvatar: Towards Realistic Human Avatar Modeling from a Single Video via Animatable 3D Gaussians" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 99, + 160, + 493, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 160, + 493, + 189 + ], + "spans": [ + { + "bbox": [ + 99, + 160, + 493, + 189 + ], + "type": "text", + "content": "Liangxiao " + }, + { + "bbox": [ + 99, + 160, + 493, + 189 + ], + "type": "inline_equation", + "content": "\\mathrm{Hu}^{\\dagger,1}" + }, + { + "bbox": [ + 99, + 160, + 493, + 189 + ], + "type": "text", + "content": ", Hongwen Zhang" + }, + { + "bbox": [ + 99, + 160, + 493, + 189 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 99, + 160, + 493, + 189 + ], + "type": "text", + "content": ", Yuxiang Zhang" + }, + { + "bbox": [ + 99, + 160, + 493, + 189 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 99, + 160, + 493, + 189 + ], + "type": "text", + "content": ", Boyao Zhou" + }, + { + "bbox": [ + 99, + 160, + 493, + 189 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 99, + 160, + 493, + 189 + ], + "type": "text", + "content": ", Boning Liu" + }, + { + "bbox": [ + 99, + 160, + 493, + 189 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 99, + 160, + 493, + 189 + ], + "type": "text", + "content": ", Shengping Zhang" + }, + { + "bbox": [ + 99, + 160, + 493, + 189 + ], + "type": "inline_equation", + "content": "^{*,1,4}" + }, + { + "bbox": [ + 99, + 160, + 493, + 189 + ], + "type": "text", + "content": ", Liqiang Nie" + }, + { + "bbox": [ + 99, + 160, + 493, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 152, + 190, + 441, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 190, + 441, + 204 + ], + "spans": [ + { + "bbox": [ + 152, + 190, + 441, + 204 + ], + "type": "text", + "content": "1Harbin Institute of Technology 2Beijing Normal University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 184, + 203, + 410, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 203, + 410, + 217 + ], + "spans": [ + { + "bbox": [ + 184, + 203, + 410, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 184, + 203, + 410, + 217 + ], + "type": "text", + "content": "Tsinghua University " + }, + { + "bbox": [ + 184, + 203, + 410, + 217 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 184, + 203, + 410, + 217 + ], + "type": "text", + "content": "Peng Cheng Laboratory" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 75, + 219, + 517, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 219, + 517, + 232 + ], + "spans": [ + { + "bbox": [ + 75, + 219, + 517, + 232 + ], + "type": "text", + "content": "{lx.hu, s.zhang}@hit.edu.cn, zhanghongwen@bnu.edu.cn, yx-z19@mails.tsinghua.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 124, + 234, + 464, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 234, + 464, + 245 + ], + "spans": [ + { + "bbox": [ + 124, + 234, + 464, + 245 + ], + "type": "text", + "content": "{bzhou22, liuboning}@mail.tsinghua.edu.cn, nieliqiang@gmail.com" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 63, + 285, + 521, + 467 + ], + "blocks": [ + { + "bbox": [ + 63, + 285, + 521, + 467 + ], + "lines": [ + { + "bbox": [ + 63, + 285, + 521, + 467 + ], + "spans": [ + { + "bbox": [ + 63, + 285, + 521, + 467 + ], + "type": "image", + "image_path": "e78040bbddb2f4b799711aeaf5810b13030766248e31885acd533c979589b050.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 478, + 546, + 501 + ], + "lines": [ + { + "bbox": [ + 46, + 478, + 546, + 501 + ], + "spans": [ + { + "bbox": [ + 46, + 478, + 546, + 501 + ], + "type": "text", + "content": "Figure 1. We propose GaussianAvatar, which learns animatable 3D Gaussians to represent detailed human avatars from a single video. Our method maintains a 3D consistent appearance even when animated by out-of-distribution motions." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 512, + 192, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 512, + 192, + 525 + ], + "spans": [ + { + "bbox": [ + 143, + 512, + 192, + 525 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 538, + 288, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 538, + 288, + 682 + ], + "spans": [ + { + "bbox": [ + 46, + 538, + 288, + 682 + ], + "type": "text", + "content": "We present GaussianAvatar, an efficient approach to creating realistic human avatars with dynamic 3D appearances from a single video. We start by introducing animatable 3D Gaussians to explicitly represent humans in various poses and clothing styles. Such an explicit and animatable representation can fuse 3D appearances more efficiently and consistently from 2D observations. Our representation is further augmented with dynamic properties to support pose-dependent appearance modeling, where a dynamic appearance network along with an estimizable feature tensor is designed to learn the motion-to-appearance mapping. Moreover, by leveraging the differentiable motion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 514, + 545, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 514, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 306, + 514, + 545, + 620 + ], + "type": "text", + "content": "condition, our method enables a joint optimization of motions and appearances during avatar modeling, which helps to tackle the long-standing issue of inaccurate motion estimation in monocular settings. The efficacy of GaussianAvatar is validated on both the public dataset and our collected dataset, demonstrating its superior performances in terms of appearance quality and rendering efficiency. The code and dataset are available at https://github.com/aipixel/GaussianAvatar." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 645, + 387, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 645, + 387, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 645, + 387, + 658 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 665, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 546, + 715 + ], + "type": "text", + "content": "Creating a customized human avatar from a single video has great potential for many applications including virtual and augmented reality, the Metaverse, gaming, and movie industries. This task is appealing yet challenging, as the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 693, + 248, + 704 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 693, + 248, + 704 + ], + "spans": [ + { + "bbox": [ + 61, + 693, + 248, + 704 + ], + "type": "text", + "content": "† Work done during an internship at Tsinghua University." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 61, + 704, + 143, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 704, + 143, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 704, + 143, + 713 + ], + "type": "text", + "content": "* Corresponding author." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "634" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": "monocular observations are highly underdetermined for the modeling of a 3D animatable avatar. Meanwhile, the inaccurate body motion estimations and the complex wrinkle deformations also make it extremely difficult to create a realistic avatar from a single video." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 134, + 289, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 134, + 289, + 410 + ], + "spans": [ + { + "bbox": [ + 46, + 134, + 289, + 410 + ], + "type": "text", + "content": "The modeling of 3D human avatars from monocular videos involves a fusion process of 2D observations to a 3D consistent human model. For this purpose, existing solutions have leveraged both implicit and explicit representations to create the base model of avatars. Among them, implicit methods [6, 15, 16, 49, 57] define a deformable human NeRF to fuse the image observation from current motion space to a canonical space by inverse skinning. However, the implicit 3D volume is inefficient in representing human surfaces and the inverse skinning also introduces ambiguous correspondences during the fusion process. These issues make it hard for implicit solutions to capture fine-grained details of moving people. As the avatar appearances are concentrated around human surfaces, explicit representations are much more efficient in modeling avatars. Previous attempts [12, 63] have employed differentiable mesh rendering to reconstruct the human surface, but these methods struggle to preserve wrinkle details due to a fixed mesh topology. On the other hand, point-based representations [18, 37, 50, 56, 65] are more effective to handle flexible topology but require millions of points to capture detailed appearances. How to represent humans remains one of the fundamental problems for avatar modeling." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 411, + 289, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 289, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 289, + 579 + ], + "type": "text", + "content": "When creating an animatable avatar from a video of moving people, the algorithm is required to learn the relationships between body motions and corresponding appearances. However, the motions estimated from monocular videos are typically faulty, leading to large artifacts in the modeling of dynamic cloth deformations. To address this issue, previous works [1, 15, 16, 42, 45, 49, 57] have attempted to optimize body motions along with the learning of animatable avatar volumes. As the body motions are explicitly represented as parametric meshes, the implicit 3D volume of previous methods makes the optimization indirect and less effective. This issue has even become the main obstacle to achieving high-quality avatar modeling from monocular videos." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": "To tackle the above issues, we introduce new representations and solutions to achieve high-quality avatar modeling from a single video. The key insight of our solution is to model dynamic human surfaces explicitly and optimize both the motion and appearances jointly in an end-to-end manner. To this end, we propose GaussianAvatar, a method to reconstruct human avatars with dynamic appearances using the 3D Gaussian representation [17]. As an explicit representation, 3D Gaussian can be easily posed from the canonical space to the motion space via a forward skinning process. Such an animatable 3D Gaus" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 300 + ], + "type": "text", + "content": "sian representation bypasses the inverse skinning process used in the aforementioned NeRF-based methods and overcomes the previous one-to-many issue [4] during canonicalization. Based on the animatable 3D Gaussian, our method can fuse 3D appearances more consistently from 2D observations to a canonical 3D space. To model dynamic human appearances under different poses, we additionally add pose-dependent properties to 3D Gaussians and incorporate them with the canonical human surfaces. Inspired by previous work [28, 46], we learn a dynamic appearance network on the 2D manifolds of the underlying human mesh to predict dynamic properties of 3D Gaussians. However, due to the strong bias of limited training poses, modeling dynamic appearance solely conditioned on pose information struggles to generalize to novel views and poses. To address this issue, we introduce an estimizable feature tensor to capture a coarse global appearance of human avatars. Subsequently, we incorporate pose-dependent effects on the feature tensor to decode fine-grained details such as wrinkles." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 300, + 547, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 547, + 443 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 547, + 443 + ], + "type": "text", + "content": "As the proposed animatable 3D Gaussians are differentiable with respect to the motion conditions, it enables a joint optimization of motion and appearances. This merit allows our network to refine the motion along with the avatar modeling process, which helps to tackle the long-standing issue of inaccurate motion estimation in monocular settings. Besides, the refined motion further enhances the accuracy of avatar modeling since the 3D appearance fusion process relies on motion-based skinning. As shown in our experiments, our method is quite robust to initial motion estimation and has the ability to correct the misalignment of motion capture results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 451, + 534, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 451, + 534, + 462 + ], + "spans": [ + { + "bbox": [ + 317, + 451, + 534, + 462 + ], + "type": "text", + "content": "To summarize, our main contributions are as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 465, + 545, + 632 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 306, + 465, + 545, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 465, + 545, + 523 + ], + "spans": [ + { + "bbox": [ + 306, + 465, + 545, + 523 + ], + "type": "text", + "content": "- We introduce animatable 3D Gaussians for realistic human avatar modeling from a single video. By representing human surfaces explicitly, our method can fuse 3D appearances more consistently and efficiently from 2D observations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 525, + 545, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 525, + 545, + 584 + ], + "spans": [ + { + "bbox": [ + 306, + 525, + 545, + 584 + ], + "type": "text", + "content": "- We augment the animatable 3D Gaussians with dynamic properties to support pose-dependent appearance modeling, where a dynamic appearance network along with an estimizable feature tensor is designed to learn the motion-to-appearance mapping." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 585, + 545, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 585, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 306, + 585, + 545, + 632 + ], + "type": "text", + "content": "- We propose to jointly optimize the motion and appearance during the avatar modeling, enabling our method to correct the misalignment of initial motion and improve the final appearance quality." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 645, + 392, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 645, + 392, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 645, + 392, + 658 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "Neural Rendering for Human Reconstruction. Without the need to define a template mesh for avatar modeling [2, 9, 53], neural rendering has emerged as a potent technique that enables learning avatars directly from im" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "635" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": "ages. Here we briefly review precious work that aims to reconstruct humans using neural rendering." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 97, + 287, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 97, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 97, + 287, + 335 + ], + "type": "text", + "content": "Due to the high-quality rendering of neural radiance field [29], various efforts [20, 21, 24, 32, 33, 47, 67, 68] have been made to reconstruct the dynamic appearance of moving people. Neural Body [33] associates a latent code to each SMPL [26] vertex to encode the appearance, which is transformed into observation space based on the human pose. Neural Actor [24] learns a deformable radiance field with SMPL as guidance and utilizes a texture map to improve its final rendering quality. TAVA [20] proposes to jointly model the non-rigid warping field and shading effects directly conditioned on the pose vectors. Posevocab [21] designs joint-structured pose embeddings to encode the dynamic appearances under different key poses, such embeddings can better learn joint-related appearance. NeRF-based methods have demonstrated appealing rendering results on human avatar reconstruction, but still struggle to represent human surfaces with the implicit 3D volume. Explicit modeling of human surfaces is a more straightforward way for this task, as the dynamic appearance of humans is mostly reflected on the human surface." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 337, + 287, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 337, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 46, + 337, + 287, + 576 + ], + "type": "text", + "content": "Explicit representations have great potential for human reconstruction. In HF-Avatar [63], meshes are used as the base representation in a coarse-to-fine framework that combines neural texture with dynamic surface deformation for avatar creation. EMA [12] proposes Meshy neural fields to reconstruct human avatars by optimizing the canonical mesh, material, and motion dynamics through inverse rendering in an end-to-end process. PointAvatar [65] employs a deformable point-based representation to separate source color into intrinsic albedo and normal-dependent shading. DVA [35] extends mixtures of volumetric primitives [25] for human avatar modeling. All these attempts have demonstrated the significant potential of explicit representations and their under-exploration. However, meshes are constrained by fixed topologies, whereas point clouds demand a multitude of points to encompass intricate details. 3D Gaussians [17] have showcased their capability in various human tasks [22, 40, 55, 64]. In essence, 3D Gaussians hold promise for human avatar reconstruction and are currently a subject of active research." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 594, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 715 + ], + "type": "text", + "content": "Avatar Modeling from Monocular Videos. Numerous methods investigate reconstructing humans from single images or monocular videos. Regression-based methods[10, 11, 13, 38, 39, 51, 52, 66] directly recover clothed 3D humans just from a single image. While these methods produce attractive results, they can not recover the dynamic appearance in the reconstruction across the entire sequence. Traditional methods aim to capture human dynamics by tracking individuals in videos using pre-scanned rigged templates [7, 8, 54]. However, the pre-scanning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "text", + "content": "and manual rigging processes prevent their real-life applications. [1, 5, 30] try to bypass the requirement for predefined human models but face challenges in preserving fine details because of the fixed mesh resolution." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 121, + 546, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 121, + 546, + 324 + ], + "spans": [ + { + "bbox": [ + 304, + 121, + 546, + 324 + ], + "type": "text", + "content": "The emergence of neural radiance fields [29] has facilitated the creation of various techniques for reconstructing animatable avatars from monocular videos [3, 14, 16, 41-43, 49]. However, inaccuracies in estimating human motions from monocular videos result in pronounced artifacts. To address the problem, HumanNeRF [49] solves for an update to the inaccurate poses. NeuMan [16] introduces an error-correction network to enable training with erroneous estimates. Vid2Avatar[6] and InstantAvatar [15] also jointly optimize motions by back-propagating the gradient of the image reconstruction loss to the pose parameters. MonoHuman [57] introduces bi-directional constraints to alleviate ambiguous correspondence on novel poses. Attempting to address the inaccurate pose estimation issue with implicit human NeRF models has proven to be inefficient and imprecise. Consequently, we endeavor to address this issue by leveraging an explicit 3D Gaussian representation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 333, + 362, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 333, + 362, + 346 + ], + "spans": [ + { + "bbox": [ + 306, + 333, + 362, + 346 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 354, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 545, + 521 + ], + "type": "text", + "content": "Our goal is to create a human avatar that enables free-viewpoint rendering and realistic animation using a single video. We first introduce an expressive representation, namely animatable 3D Gaussians, to represent human avatars in Sec. 3.1. With this representation, modeling the dynamic appearance of humans can be regarded as dynamic 3D Gaussian property estimation. Then we build a dynamic appearance network along with an estimable feature tensor to learn the motion-to-appearance mapping in Sec. 3.2. To mitigate the artifacts caused by inaccurate pose estimation, we adopt a joint motion and appearance optimization approach for refining human poses during training (Sec. 3.3). In the following, we provide a comprehensive explanation of the technical details." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 528, + 450, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 528, + 450, + 540 + ], + "spans": [ + { + "bbox": [ + 306, + 528, + 450, + 540 + ], + "type": "text", + "content": "3.1. Animatable 3D Gaussians" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 545, + 545, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 630 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 630 + ], + "type": "text", + "content": "Point-based representation [23, 28, 34, 59, 69] has proven its topological flexibility in generating realistic human avatars from scans. Extending this explicit representation to create human avatars from images is a significant endeavor. With this goal in mind, we introduce a novel representation, termed animatable 3D Gaussians, which effectively reconstructs human surfaces with a 3D consistent appearance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 631, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 631, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 631, + 545, + 713 + ], + "type": "text", + "content": "3D Gaussian Splitting [17] is a point-based scene representation that allows high-quality real-time rendering. The scene representation is parameterized by a set of static 3D Gaussians, each of which has the following parameters: 3D center position " + }, + { + "bbox": [ + 304, + 631, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 631, + 545, + 713 + ], + "type": "text", + "content": ", color " + }, + { + "bbox": [ + 304, + 631, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{c} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 631, + 545, + 713 + ], + "type": "text", + "content": ", opacity " + }, + { + "bbox": [ + 304, + 631, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\alpha \\in \\mathbb{R}" + }, + { + "bbox": [ + 304, + 631, + 545, + 713 + ], + "type": "text", + "content": ", 3D rotation in form of quaternion " + }, + { + "bbox": [ + 304, + 631, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{q} \\in \\mathbb{R}^4" + }, + { + "bbox": [ + 304, + 631, + 545, + 713 + ], + "type": "text", + "content": " and 3D scaling factor " + }, + { + "bbox": [ + 304, + 631, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{s} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 631, + 545, + 713 + ], + "type": "text", + "content": ". With these properties, we can generate rendered" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "636" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 72, + 545, + 208 + ], + "blocks": [ + { + "bbox": [ + 51, + 72, + 545, + 208 + ], + "lines": [ + { + "bbox": [ + 51, + 72, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 51, + 72, + 545, + 208 + ], + "type": "image", + "image_path": "26330d876fda00891ab4b8a45025625ad03c508ee6cef6ff63f263f78a64a74b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 217, + 548, + 275 + ], + "lines": [ + { + "bbox": [ + 46, + 217, + 548, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 548, + 275 + ], + "type": "text", + "content": "Figure 2. Overview of GaussianAvatar. Given a fitted SMPL or SMPL-X model on the current frame, we sample the points on its surface and record their positions on a UV positional map " + }, + { + "bbox": [ + 46, + 217, + 548, + 275 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 217, + 548, + 275 + ], + "type": "text", + "content": ", which is then passed to a pose encoder to obtain the pose feature. An estimizable feature tensor is pixel-aligned with the pose feature and learned to capture the coarse appearance of humans. Then the two aligned feature tensors are input into the Gaussian parameter decoder, which predicts each point's offset " + }, + { + "bbox": [ + 46, + 217, + 548, + 275 + ], + "type": "inline_equation", + "content": "\\Delta \\hat{\\mathbf{x}}" + }, + { + "bbox": [ + 46, + 217, + 548, + 275 + ], + "type": "text", + "content": ", color " + }, + { + "bbox": [ + 46, + 217, + 548, + 275 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{c}}" + }, + { + "bbox": [ + 46, + 217, + 548, + 275 + ], + "type": "text", + "content": ", and scale " + }, + { + "bbox": [ + 46, + 217, + 548, + 275 + ], + "type": "inline_equation", + "content": "\\hat{s}" + }, + { + "bbox": [ + 46, + 217, + 548, + 275 + ], + "type": "text", + "content": ". These predictions, along with the fixed rotations " + }, + { + "bbox": [ + 46, + 217, + 548, + 275 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 46, + 217, + 548, + 275 + ], + "type": "text", + "content": " and opacity " + }, + { + "bbox": [ + 46, + 217, + 548, + 275 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 217, + 548, + 275 + ], + "type": "text", + "content": ", collectively constitute the animatable 3D Gaussians in canonical space." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 284, + 288, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 284, + 288, + 308 + ], + "spans": [ + { + "bbox": [ + 46, + 284, + 288, + 308 + ], + "type": "text", + "content": "images from any viewpoint. We refer readers to [17, 27] for the rendering details of 3D Gaussians." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 308, + 288, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 288, + 344 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 288, + 344 + ], + "type": "text", + "content": "To extend this representation for human avatar modeling, we integrate it with either the SMPL [26] or SMPL-X [31] model as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 354, + 287, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 354, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 53, + 354, + 287, + 369 + ], + "type": "interline_equation", + "content": "G (\\pmb {\\beta}, \\pmb {\\theta}, \\mathbf {D}, \\mathbf {P}) = S p l a t t i n g (W (\\mathbf {D}, J (\\pmb {\\beta}), \\pmb {\\theta}, \\omega), \\mathbf {P}), (1)", + "image_path": "7771b0fcf1f03f8bb712ed0dcf8f5157a8a87f945171b4033c9ec0f16179c963.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 376, + 289, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 289, + 547 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 289, + 547 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 376, + 289, + 547 + ], + "type": "inline_equation", + "content": "G(\\cdot)" + }, + { + "bbox": [ + 46, + 376, + 289, + 547 + ], + "type": "text", + "content": " represents a rendered image, and " + }, + { + "bbox": [ + 46, + 376, + 289, + 547 + ], + "type": "inline_equation", + "content": "Splatting(\\cdot)" + }, + { + "bbox": [ + 46, + 376, + 289, + 547 + ], + "type": "text", + "content": " denotes the rendering process of 3D Gaussians from any viewpoint, " + }, + { + "bbox": [ + 46, + 376, + 289, + 547 + ], + "type": "inline_equation", + "content": "W(\\cdot)" + }, + { + "bbox": [ + 46, + 376, + 289, + 547 + ], + "type": "text", + "content": " is a standard linear blend skinning function employed for reposing 3D Gaussians, " + }, + { + "bbox": [ + 46, + 376, + 289, + 547 + ], + "type": "inline_equation", + "content": "\\mathbf{D} = T(\\boldsymbol{\\beta}) + dT" + }, + { + "bbox": [ + 46, + 376, + 289, + 547 + ], + "type": "text", + "content": " represents the locations of 3D Gaussians in canonical space, formed by adding corrective point displacements " + }, + { + "bbox": [ + 46, + 376, + 289, + 547 + ], + "type": "inline_equation", + "content": "dT" + }, + { + "bbox": [ + 46, + 376, + 289, + 547 + ], + "type": "text", + "content": " on the template mesh surface " + }, + { + "bbox": [ + 46, + 376, + 289, + 547 + ], + "type": "inline_equation", + "content": "T(\\beta)" + }, + { + "bbox": [ + 46, + 376, + 289, + 547 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 376, + 289, + 547 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 46, + 376, + 289, + 547 + ], + "type": "text", + "content": " denotes the remaining properties of 3D Gaussians, excluding the positions. " + }, + { + "bbox": [ + 46, + 376, + 289, + 547 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 376, + 289, + 547 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 376, + 289, + 547 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 376, + 289, + 547 + ], + "type": "text", + "content": " are the shape and pose parameters, " + }, + { + "bbox": [ + 46, + 376, + 289, + 547 + ], + "type": "inline_equation", + "content": "J(\\beta)" + }, + { + "bbox": [ + 46, + 376, + 289, + 547 + ], + "type": "text", + "content": " outputs 3D joint locations. Note that we propagate the skinning weight " + }, + { + "bbox": [ + 46, + 376, + 289, + 547 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 46, + 376, + 289, + 547 + ], + "type": "text", + "content": " from the vertices of the SMPL or SMPL-X model to the nearest 3D Gaussians. With the proposed representation, we can now repose these canonical 3D Gaussians to the motion space for free-view rendering." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 552, + 274, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 274, + 565 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 274, + 565 + ], + "type": "text", + "content": "3.2. Dynamic 3D Gaussian Property Estimation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": "Following the proposed animatable Gaussians, human appearances are determined by the point displacements " + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "inline_equation", + "content": "dT" + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": " and properties " + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": ". Modeling dynamic human appearances can be regarded as estimating these dynamic properties. To model dynamic human appearances under various poses, we introduce a dynamic appearance network along with an estimizable feature tensor to predict these pose-dependent properties of 3D Gaussians. Despite sharing similarities in network structure with [28], we present this framework for a distinct purpose. In [28], the feature tensor serves to decouple the pose-independent human shape from the decoder, and we repurpose it to capture a coarse global appearance" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 310, + 284, + 359, + 393 + ], + "blocks": [ + { + "bbox": [ + 310, + 284, + 359, + 393 + ], + "lines": [ + { + "bbox": [ + 310, + 284, + 359, + 393 + ], + "spans": [ + { + "bbox": [ + 310, + 284, + 359, + 393 + ], + "type": "image", + "image_path": "460762fcaac45792567d669584836768ed1f08403826bda9b989f98bccf7a450.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 325, + 396, + 338, + 407 + ], + "lines": [ + { + "bbox": [ + 325, + 396, + 338, + 407 + ], + "spans": [ + { + "bbox": [ + 325, + 396, + 338, + 407 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 361, + 284, + 398, + 394 + ], + "blocks": [ + { + "bbox": [ + 361, + 284, + 398, + 394 + ], + "lines": [ + { + "bbox": [ + 361, + 284, + 398, + 394 + ], + "spans": [ + { + "bbox": [ + 361, + 284, + 398, + 394 + ], + "type": "image", + "image_path": "504c07309e5eb4d0857085dea30ea95c4095228d8fbbe4ac05908ea4f34a72f1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 372, + 396, + 384, + 407 + ], + "lines": [ + { + "bbox": [ + 372, + 396, + 384, + 407 + ], + "spans": [ + { + "bbox": [ + 372, + 396, + 384, + 407 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 402, + 284, + 449, + 394 + ], + "blocks": [ + { + "bbox": [ + 402, + 284, + 449, + 394 + ], + "lines": [ + { + "bbox": [ + 402, + 284, + 449, + 394 + ], + "spans": [ + { + "bbox": [ + 402, + 284, + 449, + 394 + ], + "type": "image", + "image_path": "19a40eee079f096e44a7e5d285e4cf533fd72d2157fea90a5023c365cc358594.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 419, + 396, + 430, + 406 + ], + "lines": [ + { + "bbox": [ + 419, + 396, + 430, + 406 + ], + "spans": [ + { + "bbox": [ + 419, + 396, + 430, + 406 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 451, + 284, + 493, + 394 + ], + "blocks": [ + { + "bbox": [ + 451, + 284, + 493, + 394 + ], + "lines": [ + { + "bbox": [ + 451, + 284, + 493, + 394 + ], + "spans": [ + { + "bbox": [ + 451, + 284, + 493, + 394 + ], + "type": "image", + "image_path": "6155fdef5b9e0184ee5365daa6468f4f474dfdfa0965aedaf128dd841b127a6a.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 466, + 396, + 478, + 406 + ], + "lines": [ + { + "bbox": [ + 466, + 396, + 478, + 406 + ], + "spans": [ + { + "bbox": [ + 466, + 396, + 478, + 406 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 415, + 547, + 449 + ], + "lines": [ + { + "bbox": [ + 305, + 415, + 547, + 449 + ], + "spans": [ + { + "bbox": [ + 305, + 415, + 547, + 449 + ], + "type": "text", + "content": "Figure 3. Effect of iostropy of 3D Gaussians. (a) Input image, (b)(d) front and back views trained with isotropic 3D Gaussians, (c)(e) front and back views trained with anisotropic 3D Gaussians." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 498, + 284, + 542, + 394 + ], + "blocks": [ + { + "bbox": [ + 498, + 284, + 542, + 394 + ], + "lines": [ + { + "bbox": [ + 498, + 284, + 542, + 394 + ], + "spans": [ + { + "bbox": [ + 498, + 284, + 542, + 394 + ], + "type": "image", + "image_path": "e2785d0a2f35a1401d0267206f2370467c2bb39c30f6d4534495c815d5808177.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 514, + 395, + 525, + 404 + ], + "lines": [ + { + "bbox": [ + 514, + 395, + 525, + 404 + ], + "spans": [ + { + "bbox": [ + 514, + 395, + 525, + 404 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 463, + 547, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 463, + 547, + 534 + ], + "spans": [ + { + "bbox": [ + 304, + 463, + 547, + 534 + ], + "type": "text", + "content": "of human avatars. The motivation behind this modification is that directly learning a mapping from human poses to dynamic properties is susceptible to overfitting on the limited training poses. To integrate the global appearance into the feature tensor, we introduce a two-stage training strategy, as discussed in Sec. 3.4." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 536, + 547, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 536, + 547, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 536, + 547, + 582 + ], + "type": "text", + "content": "The dynamic appearance network is designed to learn a mapping from a 2D manifold representing the underlying human shape to the dynamic properties of 3D Gaussians as follows:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 382, + 584, + 545, + 598 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 584, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 382, + 584, + 545, + 598 + ], + "type": "interline_equation", + "content": "f _ {\\phi}: \\mathcal {S} ^ {2} \\in \\mathbb {R} ^ {3} \\rightarrow \\mathbb {R} ^ {7}. \\tag {2}", + "image_path": "9179c52a5cfa10f5e8f8780cac7124222315d37d79cf87d75a55ef7e18bf5839.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 605, + 547, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 676 + ], + "type": "text", + "content": "As shown in Fig. 2, the 2D human manifold " + }, + { + "bbox": [ + 304, + 605, + 547, + 676 + ], + "type": "inline_equation", + "content": "S^2" + }, + { + "bbox": [ + 304, + 605, + 547, + 676 + ], + "type": "text", + "content": " is depicted by a UV positional map " + }, + { + "bbox": [ + 304, + 605, + 547, + 676 + ], + "type": "inline_equation", + "content": "I \\in R^{H \\times W \\times 3}" + }, + { + "bbox": [ + 304, + 605, + 547, + 676 + ], + "type": "text", + "content": ", where each valid pixel stores the position " + }, + { + "bbox": [ + 304, + 605, + 547, + 676 + ], + "type": "inline_equation", + "content": "(x, y, z)" + }, + { + "bbox": [ + 304, + 605, + 547, + 676 + ], + "type": "text", + "content": " of one point on the posed body surface. The final predictions consist of per point offset " + }, + { + "bbox": [ + 304, + 605, + 547, + 676 + ], + "type": "inline_equation", + "content": "\\Delta \\hat{\\mathbf{x}} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 605, + 547, + 676 + ], + "type": "text", + "content": ", color " + }, + { + "bbox": [ + 304, + 605, + 547, + 676 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{c}} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 605, + 547, + 676 + ], + "type": "text", + "content": ", and scale " + }, + { + "bbox": [ + 304, + 605, + 547, + 676 + ], + "type": "inline_equation", + "content": "\\hat{s} \\in \\mathbb{R}" + }, + { + "bbox": [ + 304, + 605, + 547, + 676 + ], + "type": "text", + "content": " on the canonical surface." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "Instead of predicting all properties, we make these slight adjustments to our task based on experimental results. Due to the unbalanced viewpoints in monocular videos," + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 748, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 748, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 748, + 313, + 757 + ], + "type": "text", + "content": "637" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 227 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 227 + ], + "type": "text", + "content": "anisotropic 3D Gaussians are prone to learning an inaccurate 3D shape to fit the most frequently seen view, resulting in poorer performance in side views. As shown in Fig. 3, training our network with isotropic 3D Gaussians yields superior results. We ensure isotropy among all 3D Gaussians by maintaining uniform size across dimensions. Therefore, we introduce a scalar " + }, + { + "bbox": [ + 46, + 72, + 289, + 227 + ], + "type": "inline_equation", + "content": "\\hat{s}" + }, + { + "bbox": [ + 46, + 72, + 289, + 227 + ], + "type": "text", + "content": " to represent the scales of 3D Gaussians and set rotations " + }, + { + "bbox": [ + 46, + 72, + 289, + 227 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 46, + 72, + 289, + 227 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 72, + 289, + 227 + ], + "type": "inline_equation", + "content": "[1,0,0,0]" + }, + { + "bbox": [ + 46, + 72, + 289, + 227 + ], + "type": "text", + "content": ". Through experimental observations, we notice that the network tends to learn an opacity value " + }, + { + "bbox": [ + 46, + 72, + 289, + 227 + ], + "type": "inline_equation", + "content": "(\\alpha)" + }, + { + "bbox": [ + 46, + 72, + 289, + 227 + ], + "type": "text", + "content": " of zero on the boundary to correct the human shape. To address this, we fix opacity " + }, + { + "bbox": [ + 46, + 72, + 289, + 227 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 46, + 72, + 289, + 227 + ], + "type": "text", + "content": " to keep all 3D Gaussians visible, enforcing the network to predict accurate positions of the 3D Gaussians." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 228, + 289, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 289, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 289, + 373 + ], + "type": "text", + "content": "The dynamic appearance network consists of two parts: a pose encoder and a Gaussian parameter decoder. The pose encoder takes the UV positional map of posed body points as input to generate a pose-conditioned feature tensor " + }, + { + "bbox": [ + 46, + 228, + 289, + 373 + ], + "type": "inline_equation", + "content": "O \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 46, + 228, + 289, + 373 + ], + "type": "text", + "content": ". We then integrate the pixel-aligned optimized feature tensor " + }, + { + "bbox": [ + 46, + 228, + 289, + 373 + ], + "type": "inline_equation", + "content": "F \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 46, + 228, + 289, + 373 + ], + "type": "text", + "content": " with the pose features before feeding it into the Gaussian parameter decoder to generate final predictions. Following this, we add the predicted offsets to the canonical 3D Gaussians and associate the predicted properties with the corresponding 3D Gaussians. With the estimated poses, we can then repose the canonical 3D Gaussians to the motion space for rendering." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 378, + 275, + 391 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 378, + 275, + 391 + ], + "spans": [ + { + "bbox": [ + 47, + 378, + 275, + 391 + ], + "type": "text", + "content": "3.3. Joint Motion and Appearance Optimization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 396, + 287, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 287, + 492 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 287, + 492 + ], + "type": "text", + "content": "However, owing to the imprecise estimation of human poses " + }, + { + "bbox": [ + 46, + 396, + 287, + 492 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta} = (\\theta_{1},\\theta_{2},\\dots,\\theta_{J})" + }, + { + "bbox": [ + 46, + 396, + 287, + 492 + ], + "type": "text", + "content": " and translations " + }, + { + "bbox": [ + 46, + 396, + 287, + 492 + ], + "type": "inline_equation", + "content": "\\mathbf{t}" + }, + { + "bbox": [ + 46, + 396, + 287, + 492 + ], + "type": "text", + "content": " from monocular videos, the reposed 3D Gaussians in motion space are inaccurately represented and may lead to unsatisfactory rendering outcomes. To address this, we propose to jointly optimize human motions and appearances. To optimize human motions with image loss, we solve for an update " + }, + { + "bbox": [ + 46, + 396, + 287, + 492 + ], + "type": "inline_equation", + "content": "(\\Delta \\pmb {\\theta},\\Delta \\mathbf{t})" + }, + { + "bbox": [ + 46, + 396, + 287, + 492 + ], + "type": "text", + "content": " to the estimated body poses and translations as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 498, + 287, + 513 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 498, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 116, + 498, + 287, + 513 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {\\Theta}} = (\\boldsymbol {\\theta} + \\Delta \\boldsymbol {\\theta}, \\mathbf {t} + \\Delta \\mathbf {t}). \\tag {3}", + "image_path": "2b6e20acf6e6a3754cbf23308ffe14466505757723a93e9d8eb1238c9770b589.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 521, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 594 + ], + "type": "text", + "content": "We modify " + }, + { + "bbox": [ + 46, + 521, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 521, + 287, + 594 + ], + "type": "text", + "content": " in Eq. 1 using " + }, + { + "bbox": [ + 46, + 521, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 46, + 521, + 287, + 594 + ], + "type": "text", + "content": " to render the proposed animatable 3D Gaussians differentiable with respect to the motion conditions. Different from previous work [6, 16, 49] that jointly optimizes human poses via inverse skinning, we optimize the updates in a forward skinning process, which benefits both motion and appearance optimization." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 599, + 153, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 153, + 612 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 153, + 612 + ], + "type": "text", + "content": "3.4. Training Strategy" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "content": "In this section, we outline our approach to training the network with inaccurate human motions. We conduct a two-stage optimization process using different loss functions. In the first stage, we aim to fuse the sequential appearances to the estimizable feature tensor and conduct motion optimization to get accurate poses for the dynamic appearance network. In this stage, we optimize the framework without incorporating any pose-dependent information by excluding" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "the training of the pose encoder. Specifically, we utilize the following loss functions to train our network:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 109, + 545, + 137 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 109, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 317, + 109, + 545, + 137 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {s t a g e} _ {1}} = \\lambda_ {r b g} \\mathcal {L} _ {r b g} + \\lambda_ {s s i m} \\mathcal {L} _ {s s i m} + \\lambda_ {l p i p s} \\mathcal {L} _ {l p i p s} \\tag {4} \\\\ + \\lambda_ {f} \\mathcal {L} _ {f} + \\lambda_ {\\text {o f f s e t}} \\mathcal {L} _ {\\text {o f f s e t}} + \\lambda_ {\\text {s c a l e}} \\mathcal {L} _ {\\text {s c a l e}}, \\\\ \\end{array}", + "image_path": "a45f262c8785edaf8019e26f4ae0d637bf10fa57c1e7903a3b68428f0f397018.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 148, + 545, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 148, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 305, + 148, + 545, + 209 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 148, + 545, + 209 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rbg}" + }, + { + "bbox": [ + 305, + 148, + 545, + 209 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 148, + 545, + 209 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ssim}" + }, + { + "bbox": [ + 305, + 148, + 545, + 209 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 148, + 545, + 209 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{lpips}" + }, + { + "bbox": [ + 305, + 148, + 545, + 209 + ], + "type": "text", + "content": " are the L1 loss, SSIM loss [48], and LPIPS loss [61], respectively. " + }, + { + "bbox": [ + 305, + 148, + 545, + 209 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_f" + }, + { + "bbox": [ + 305, + 148, + 545, + 209 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 148, + 545, + 209 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{offset}" + }, + { + "bbox": [ + 305, + 148, + 545, + 209 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 148, + 545, + 209 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{scale}" + }, + { + "bbox": [ + 305, + 148, + 545, + 209 + ], + "type": "text", + "content": " calculate the L2-norm of the feature map, predicted offsets and scales, respectively. We set " + }, + { + "bbox": [ + 305, + 148, + 545, + 209 + ], + "type": "inline_equation", + "content": "\\lambda_{rbg} = 0.8" + }, + { + "bbox": [ + 305, + 148, + 545, + 209 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 148, + 545, + 209 + ], + "type": "inline_equation", + "content": "\\lambda_{ssim} = 0.2" + }, + { + "bbox": [ + 305, + 148, + 545, + 209 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 148, + 545, + 209 + ], + "type": "inline_equation", + "content": "\\lambda_{lpips} = 0.2" + }, + { + "bbox": [ + 305, + 148, + 545, + 209 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 148, + 545, + 209 + ], + "type": "inline_equation", + "content": "\\lambda_f = 1" + }, + { + "bbox": [ + 305, + 148, + 545, + 209 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 148, + 545, + 209 + ], + "type": "inline_equation", + "content": "\\lambda_{offset} = 10" + }, + { + "bbox": [ + 305, + 148, + 545, + 209 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 148, + 545, + 209 + ], + "type": "inline_equation", + "content": "\\lambda_{scale} = 1" + }, + { + "bbox": [ + 305, + 148, + 545, + 209 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 210, + 546, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 210, + 546, + 330 + ], + "spans": [ + { + "bbox": [ + 304, + 210, + 546, + 330 + ], + "type": "text", + "content": "After the first stage of training, we obtain more accurate human motions and an optimized feature tensor " + }, + { + "bbox": [ + 304, + 210, + 546, + 330 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 210, + 546, + 330 + ], + "type": "text", + "content": ". The optimized feature tensor " + }, + { + "bbox": [ + 304, + 210, + 546, + 330 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 210, + 546, + 330 + ], + "type": "text", + "content": " captures a coarse appearance of human avatars. In the second stage, we incorporate the pose features encoded by the pose encoder with the trained feature tensor " + }, + { + "bbox": [ + 304, + 210, + 546, + 330 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 210, + 546, + 330 + ], + "type": "text", + "content": ". We replace " + }, + { + "bbox": [ + 304, + 210, + 546, + 330 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_f" + }, + { + "bbox": [ + 304, + 210, + 546, + 330 + ], + "type": "text", + "content": " with the L2-norm loss " + }, + { + "bbox": [ + 304, + 210, + 546, + 330 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_p" + }, + { + "bbox": [ + 304, + 210, + 546, + 330 + ], + "type": "text", + "content": ", which plays the same role as " + }, + { + "bbox": [ + 304, + 210, + 546, + 330 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_f" + }, + { + "bbox": [ + 304, + 210, + 546, + 330 + ], + "type": "text", + "content": " in regularizing the limited pose space. By penalizing the pose-dependent features, we can eliminate the strong bias of limited training poses and thus generalize to unseen viewpoints and poses." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 344, + 388, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 344, + 388, + 358 + ], + "spans": [ + { + "bbox": [ + 306, + 344, + 388, + 358 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 365, + 427, + 378 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 365, + 427, + 378 + ], + "spans": [ + { + "bbox": [ + 306, + 365, + 427, + 378 + ], + "type": "text", + "content": "4.1. Datasets and Metrics" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 386, + 545, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 386, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 305, + 386, + 545, + 434 + ], + "type": "text", + "content": "People-Snapshot Dataset. This dataset [1] comprises videos of individuals rotating in front of a stationary camera. To ensure a fair quantitative comparison, we follow the same evaluation protocol outlined in InstantAvatar [15]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 435, + 545, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 435, + 545, + 531 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 545, + 531 + ], + "type": "text", + "content": "NeuMan Dataset. To assess even more challenging scenarios, we employ outdoor collections from the NeuMan dataset [16]. These videos are recorded using a mobile phone for moving individuals. Specifically, we select four sequences (bike, citizen, Seattle, jogging) that exhibit most body regions and contain minimal blurry images. We initialize the estimated poses with an off-the-shelf method [44], which is also utilized in [15]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 532, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 532, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 532, + 545, + 676 + ], + "type": "text", + "content": "DynVideo Dataset. With the limited cloth deformation presented in the above two datasets, we propose the Dynvideo dataset to capture dynamic human appearance. To this end, we utilize a mobile phone to record videos of a character performing various movements, especially rotation, in front of the device. Each of these videos takes about one minute and provides a comprehensive and detailed representation of human movement. We also provide the corresponding SMPL parameter sequences for all videos, obtained by our proposed method. This dataset serves as a valuable resource for evaluating reconstruction quality, with a particular emphasis on dynamic appearances." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "Evaluation Metrics. We consider three metrics: PSNR, SSIM [48], and LPIPS [61] to access the reconstruction quality on three datasets." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "638" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 547, + 159 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 547, + 159 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 547, + 159 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 547, + 159 + ], + "type": "table", + "html": "
Methodmale-3-casualmale-4-casualfemale-3-casualfemale-4-casual
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
HumanNeRF [49]26.900.96050.018125.500.93970.035724.460.95160.026927.070.96150.0152
InstantAvatar [15]29.530.97160.015527.670.96260.030727.660.97090.021029.110.96830.0167
Baseline27.710.97130.021825.090.96140.030624.800.96090.031026.160.96100.0227
Baseline + Opt.30.980.97910.014328.770.97530.023029.590.97680.022030.830.97680.0145
Baseline + Opt. + Dyn.30.980.97900.014528.780.97550.022829.550.97620.022530.840.97710.0140
", + "image_path": "a475419c29254a8096fff7388fb327eaa7b75d1bb1dd933f60539cd2fd89fb25.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 200, + 279, + 272 + ], + "blocks": [ + { + "bbox": [ + 46, + 165, + 547, + 189 + ], + "lines": [ + { + "bbox": [ + 46, + 165, + 547, + 189 + ], + "spans": [ + { + "bbox": [ + 46, + 165, + 547, + 189 + ], + "type": "text", + "content": "Table 1. Quantitative evaluation on the People-Snapshot [1] dataset. Opt. denotes the motion optimization and Dyn. refers to dynamic appearance modeling." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 200, + 279, + 272 + ], + "lines": [ + { + "bbox": [ + 56, + 200, + 279, + 272 + ], + "spans": [ + { + "bbox": [ + 56, + 200, + 279, + 272 + ], + "type": "table", + "html": "
MethodPSNR ↑SSIM ↑LPIPS ↓
HumanNeRF [49]27.060.96690.0192
InstantAvatar [15]28.470.97150.0277
Baseline27.060.96920.0183
Baseline + Opt.29.930.97940.0126
Baseline + Opt. + Dyn.29.940.97950.0124
", + "image_path": "5280fc7045287bd9fc45e2542bbd0594fb9279e03b59180fd49a3a1ef7ca8740.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 56, + 303, + 279, + 373 + ], + "blocks": [ + { + "bbox": [ + 56, + 279, + 277, + 291 + ], + "lines": [ + { + "bbox": [ + 56, + 279, + 277, + 291 + ], + "spans": [ + { + "bbox": [ + 56, + 279, + 277, + 291 + ], + "type": "text", + "content": "Table 2. Quantitative evaluation on NeuMan [16] dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 303, + 279, + 373 + ], + "lines": [ + { + "bbox": [ + 56, + 303, + 279, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 303, + 279, + 373 + ], + "type": "table", + "html": "
MethodPSNR ↑SSIM ↑LPIPS ↓
HumanNeRF [49]24.390.93490.0349
InstantAvatar [15]20.310.91830.1046
Baseline23.550.93400.0354
Baseline + Opt.26.160.95220.0237
Baseline + Opt. + Dyn.28.580.96160.0169
", + "image_path": "8c9f7f6ead0584698e210df7c047eec3735a59536e3de3e27119d92bb43b4ff5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 63, + 381, + 271, + 393 + ], + "lines": [ + { + "bbox": [ + 63, + 381, + 271, + 393 + ], + "spans": [ + { + "bbox": [ + 63, + 381, + 271, + 393 + ], + "type": "text", + "content": "Table 3. Quantitative evaluation on DynVideo dataset." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 52, + 407, + 91, + 530 + ], + "blocks": [ + { + "bbox": [ + 52, + 407, + 91, + 530 + ], + "lines": [ + { + "bbox": [ + 52, + 407, + 91, + 530 + ], + "spans": [ + { + "bbox": [ + 52, + 407, + 91, + 530 + ], + "type": "image", + "image_path": "53a69b099812bce6361a1a404aedc0d370eee4bf278ca822a7380654c7ab654e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 532, + 78, + 540 + ], + "lines": [ + { + "bbox": [ + 67, + 532, + 78, + 540 + ], + "spans": [ + { + "bbox": [ + 67, + 532, + 78, + 540 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 552, + 287, + 574 + ], + "lines": [ + { + "bbox": [ + 46, + 552, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 552, + 287, + 574 + ], + "type": "text", + "content": "Figure 4. Motion optimization results. (a)(d) Original image, (b)(e) our optimized SMPL, (c)(f) ROMP [44] estimates." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 91, + 407, + 128, + 529 + ], + "blocks": [ + { + "bbox": [ + 91, + 407, + 128, + 529 + ], + "lines": [ + { + "bbox": [ + 91, + 407, + 128, + 529 + ], + "spans": [ + { + "bbox": [ + 91, + 407, + 128, + 529 + ], + "type": "image", + "image_path": "efd5c94998dee1d0fe52f5c704e4f0e6d22feb6dd7cb275b820ab8da6000209b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 104, + 532, + 114, + 540 + ], + "lines": [ + { + "bbox": [ + 104, + 532, + 114, + 540 + ], + "spans": [ + { + "bbox": [ + 104, + 532, + 114, + 540 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 129, + 407, + 167, + 529 + ], + "blocks": [ + { + "bbox": [ + 129, + 407, + 167, + 529 + ], + "lines": [ + { + "bbox": [ + 129, + 407, + 167, + 529 + ], + "spans": [ + { + "bbox": [ + 129, + 407, + 167, + 529 + ], + "type": "image", + "image_path": "e913b401eb4b088d37fc059217f7dde737b89f15cc2605a2ba4dfe51097849ac.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 142, + 532, + 152, + 540 + ], + "lines": [ + { + "bbox": [ + 142, + 532, + 152, + 540 + ], + "spans": [ + { + "bbox": [ + 142, + 532, + 152, + 540 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 168, + 407, + 206, + 529 + ], + "blocks": [ + { + "bbox": [ + 168, + 407, + 206, + 529 + ], + "lines": [ + { + "bbox": [ + 168, + 407, + 206, + 529 + ], + "spans": [ + { + "bbox": [ + 168, + 407, + 206, + 529 + ], + "type": "image", + "image_path": "6c5cfc78e55dbcebbdfbbfe3ca57a466ea619cd89213370d6501cd7527fcd977.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 532, + 194, + 540 + ], + "lines": [ + { + "bbox": [ + 184, + 532, + 194, + 540 + ], + "spans": [ + { + "bbox": [ + 184, + 532, + 194, + 540 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 207, + 407, + 244, + 529 + ], + "blocks": [ + { + "bbox": [ + 207, + 407, + 244, + 529 + ], + "lines": [ + { + "bbox": [ + 207, + 407, + 244, + 529 + ], + "spans": [ + { + "bbox": [ + 207, + 407, + 244, + 529 + ], + "type": "image", + "image_path": "7df6d01934acbddd2669f26b3ed4834657cf5e4d869545cc0f80a139848dd10a.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 532, + 230, + 540 + ], + "lines": [ + { + "bbox": [ + 220, + 532, + 230, + 540 + ], + "spans": [ + { + "bbox": [ + 220, + 532, + 230, + 540 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 245, + 407, + 282, + 529 + ], + "blocks": [ + { + "bbox": [ + 245, + 407, + 282, + 529 + ], + "lines": [ + { + "bbox": [ + 245, + 407, + 282, + 529 + ], + "spans": [ + { + "bbox": [ + 245, + 407, + 282, + 529 + ], + "type": "image", + "image_path": "f387d007031e81449e82fc9f40b230bd1e8b8aea88bf0342dfdb45261aaef90b.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 258, + 532, + 268, + 540 + ], + "lines": [ + { + "bbox": [ + 258, + 532, + 268, + 540 + ], + "spans": [ + { + "bbox": [ + 258, + 532, + 268, + 540 + ], + "type": "text", + "content": "(f)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 47, + 597, + 174, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 174, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 174, + 609 + ], + "type": "text", + "content": "4.2. Experimental Settings" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 616, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 616, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 616, + 287, + 700 + ], + "type": "text", + "content": "Implementation details. We employ a U-Net [36] for extracting pose-dependent features, and the Gaussian parameter decoder is implemented as an 8-layer multilayer perceptron (MLP). We sample approximately 200,000 points on the SMPL mesh surface. The entire framework is trained on a single NVIDIA RTX 3090 GPU, with training times ranging from 0.5 to 6 hours." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "type": "text", + "content": "Baseline. To showcase the efficacy of the proposed mod" + } + ] + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 310, + 201, + 425, + 292 + ], + "blocks": [ + { + "bbox": [ + 310, + 201, + 425, + 292 + ], + "lines": [ + { + "bbox": [ + 310, + 201, + 425, + 292 + ], + "spans": [ + { + "bbox": [ + 310, + 201, + 425, + 292 + ], + "type": "image", + "image_path": "6a03ee5740839210e69ab2df20dfc0bc36461068aaa8f2ca98e4ad2a5f5b77fb.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 312, + 294, + 369, + 380 + ], + "blocks": [ + { + "bbox": [ + 312, + 294, + 369, + 380 + ], + "lines": [ + { + "bbox": [ + 312, + 294, + 369, + 380 + ], + "spans": [ + { + "bbox": [ + 312, + 294, + 369, + 380 + ], + "type": "image", + "image_path": "e617c07bc6d3c71539c702e09c6052177451ec63859c9d93a4c8934fdef87f95.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 381, + 348, + 390 + ], + "lines": [ + { + "bbox": [ + 339, + 381, + 348, + 390 + ], + "spans": [ + { + "bbox": [ + 339, + 381, + 348, + 390 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 370, + 294, + 425, + 380 + ], + "blocks": [ + { + "bbox": [ + 370, + 294, + 425, + 380 + ], + "lines": [ + { + "bbox": [ + 370, + 294, + 425, + 380 + ], + "spans": [ + { + "bbox": [ + 370, + 294, + 425, + 380 + ], + "type": "image", + "image_path": "474aa8c9485bc9fdb7e8fcd4d3eb1c0ccd7f3f84f9d607e0108bb13af96db098.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 395, + 381, + 405, + 390 + ], + "lines": [ + { + "bbox": [ + 395, + 381, + 405, + 390 + ], + "spans": [ + { + "bbox": [ + 395, + 381, + 405, + 390 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 401, + 545, + 425 + ], + "lines": [ + { + "bbox": [ + 306, + 401, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 306, + 401, + 545, + 425 + ], + "type": "text", + "content": "Figure 5. Qualitative ablation studies. (a) Ground truth, (b) baseline + Opt. + Dyn., (c) baseline + Opt., (d) baseline." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 425, + 202, + 481, + 292 + ], + "blocks": [ + { + "bbox": [ + 425, + 202, + 481, + 292 + ], + "lines": [ + { + "bbox": [ + 425, + 202, + 481, + 292 + ], + "spans": [ + { + "bbox": [ + 425, + 202, + 481, + 292 + ], + "type": "image", + "image_path": "ae64d062660dc7f762a9673ed9e20ea2eef4bd48db0accf99e0bf0fb46699caa.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 427, + 294, + 482, + 380 + ], + "blocks": [ + { + "bbox": [ + 427, + 294, + 482, + 380 + ], + "lines": [ + { + "bbox": [ + 427, + 294, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 427, + 294, + 482, + 380 + ], + "type": "image", + "image_path": "fb82564be95791f788df7d5b2ee4d9ff2d7f16ac0b1d81352ea14146fea9b511.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 452, + 381, + 460, + 389 + ], + "lines": [ + { + "bbox": [ + 452, + 381, + 460, + 389 + ], + "spans": [ + { + "bbox": [ + 452, + 381, + 460, + 389 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 483, + 202, + 539, + 292 + ], + "blocks": [ + { + "bbox": [ + 483, + 202, + 539, + 292 + ], + "lines": [ + { + "bbox": [ + 483, + 202, + 539, + 292 + ], + "spans": [ + { + "bbox": [ + 483, + 202, + 539, + 292 + ], + "type": "image", + "image_path": "f9cf3cdef3710d38af1acc9f23a9387c7c7f50a73eb54546d54a68313878d137.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 484, + 294, + 539, + 380 + ], + "blocks": [ + { + "bbox": [ + 484, + 294, + 539, + 380 + ], + "lines": [ + { + "bbox": [ + 484, + 294, + 539, + 380 + ], + "spans": [ + { + "bbox": [ + 484, + 294, + 539, + 380 + ], + "type": "image", + "image_path": "fe8ad7e17cf8f04981501fd9a8f6dd6bff1fab79f27b28aece4fefd9a1757966.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 507, + 381, + 517, + 389 + ], + "lines": [ + { + "bbox": [ + 507, + 381, + 517, + 389 + ], + "spans": [ + { + "bbox": [ + 507, + 381, + 517, + 389 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "bbox": [ + 304, + 437, + 545, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 496 + ], + "type": "text", + "content": "ules, our method can be partitioned into three components: baseline, motion optimization, and dynamic appearance modeling. As our baseline, we suspend the motion optimization process and exclude the pose-dependent information from the pose encoder." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 304, + 496, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 496, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 496, + 545, + 567 + ], + "type": "text", + "content": "Methods for Comparison. We compare our method against (1) HumanNeRF [49], which implicitly represents a human avatar with a canonical appearance neural field and a motion field; (2) InstantAvatar [15], which achieves fast avatar modeling by using several acceleration strategies for neural fields." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 306, + 575, + 507, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 575, + 507, + 588 + ], + "spans": [ + { + "bbox": [ + 306, + 575, + 507, + 588 + ], + "type": "text", + "content": "4.3. Comparisons with the State of the Art" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "We report the numeric evaluation results on three datasets. As shown in Table 1, 2, and 3, our proposed method outperforms all baselines on all metrics for recovering more details of dynamic appearance and correcting the artifacts caused by initialized poses. To demonstrate the qualitative evaluation of these datasets, we also visualize the novel view synthesis results on the test splits. As implicit representations, HumanNeRF and InstantAvatar are prone to generate ghosting effects at boundary areas, as shown in Fig. 6. InstantAvatar lacks the capability to model pose" + } + ] + } + ], + "index": 37 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "639" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 76, + 318, + 585 + ], + "blocks": [ + { + "bbox": [ + 52, + 76, + 318, + 585 + ], + "lines": [ + { + "bbox": [ + 52, + 76, + 318, + 585 + ], + "spans": [ + { + "bbox": [ + 52, + 76, + 318, + 585 + ], + "type": "image", + "image_path": "3307401eced9c5d32e29fda4fbbfa79b0db791157a873df09435bf499faec07c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 591, + 547, + 615 + ], + "lines": [ + { + "bbox": [ + 46, + 591, + 547, + 615 + ], + "spans": [ + { + "bbox": [ + 46, + 591, + 547, + 615 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison of novel view synthesis. We compare the novel view synthesis quality on the People-Snapshot dataset (first row), NeuMan dataset (second row), and DynVideo dataset (last two rows)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 325, + 76, + 535, + 582 + ], + "blocks": [ + { + "bbox": [ + 325, + 76, + 535, + 582 + ], + "lines": [ + { + "bbox": [ + 325, + 76, + 535, + 582 + ], + "spans": [ + { + "bbox": [ + 325, + 76, + 535, + 582 + ], + "type": "image", + "image_path": "8e2a9f2db1de5d5c275a29e7874b1600f18cd2110e5e9ecb06c8aa7306e48897.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 632, + 288, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 632, + 288, + 657 + ], + "spans": [ + { + "bbox": [ + 46, + 632, + 288, + 657 + ], + "type": "text", + "content": "dependent deformations, leading to challenges in effectively handling the dynamic appearance of moving people." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "To showcase the robustness of our avatar modeling, we collect more challenging poses [19] captured by the monocular camera and evaluate our method and other methods for avatar animation and novel view synthesis. As ground truth" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 632, + 547, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 632, + 547, + 704 + ], + "spans": [ + { + "bbox": [ + 304, + 632, + 547, + 704 + ], + "type": "text", + "content": "is unavailable for these out-of-distribution poses, we illustrate qualitative results of novel pose and view synthesis in Fig. 1 and Fig. 7. Our method generates realistic animation results with these challenging poses, demonstrating a consistent 3D appearance in novel views with respect to other methods." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "640" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 76, + 541, + 349 + ], + "blocks": [ + { + "bbox": [ + 55, + 76, + 541, + 349 + ], + "lines": [ + { + "bbox": [ + 55, + 76, + 541, + 349 + ], + "spans": [ + { + "bbox": [ + 55, + 76, + 541, + 349 + ], + "type": "image", + "image_path": "56a851efbb927e2ca51d324b297af18c839938f557224fde54d862a77573b7f5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 362, + 546, + 384 + ], + "lines": [ + { + "bbox": [ + 46, + 362, + 546, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 362, + 546, + 384 + ], + "type": "text", + "content": "Figure 7. Animation results on out-of-distribution motions. We compare the animation results on the People-Snapshot dataset (first row), NeuMan dataset (second row), and DynVideo dataset (last row)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 403, + 147, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 403, + 147, + 414 + ], + "spans": [ + { + "bbox": [ + 47, + 403, + 147, + 414 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 423, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 423, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 287, + 579 + ], + "type": "text", + "content": "In this section, we conduct ablation studies to validate each component of our methods. As shown in Table 1, 2, and 3, our proposed motion optimization module dramatically improves over the baseline on all metrics, demonstrating its effectiveness in modeling human avatars. To highlight the effectiveness of our method in human motion optimization, we illustrate the initialized poses obtained by ROMP [44] and the optimized ones by our method on NeuMan and DynVideo datasets in Fig. 4. Experiments demonstrate that joint motion optimization is capable of correcting inaccurate motion estimation, even for the side and back views. Furthermore, our approach readily extends to enhance the accuracy of existing motion capture methods [58, 60, 62]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 582, + 288, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 712 + ], + "type": "text", + "content": "We also notice that the dynamic appearance modeling achieves superior results on NeuMan and DynVideo datasets in Table 2 and 3 and slightly favorable outcomes in Table 1, considering that People-Snapshot has limited pose variation with respect to the other two datasets. In Fig. 5, we further illustrate the visible improvements of each component of our method. The motion optimization scheme improves the global quality of the baseline for better pose estimation. The dynamic appearance modeling additionally preserves pose-dependent details such as cloth wrinkles on the human surface." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 402, + 457, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 402, + 457, + 414 + ], + "spans": [ + { + "bbox": [ + 306, + 402, + 457, + 414 + ], + "type": "text", + "content": "5. Conclusion and Discussion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 422, + 545, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 545, + 553 + ], + "type": "text", + "content": "We introduce GaussianAvatar, a human avatar reconstruction method based on the proposed animatable 3D Gaussians from monocular videos. For dynamic human appearance modeling, we leverage a dynamic appearance network along with an estimizable feature tensor to enhance the representation with dynamic properties. Besides, we implement a joint motion and appearance optimization scheme to rectify estimated motion and enhance the overall reconstruction quality. Our method shows the capability to reconstruct avatars with dynamic appearances, enabling realistic animation while maintaining real-time rendering speed." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 554, + 545, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 554, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 554, + 545, + 601 + ], + "type": "text", + "content": "Limitation. Similar to [15, 49, 57], our method may generate artifacts due to inaccurate foreground segmentations in videos and encounter challenges in modeling loose outfits such as dresses." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 601, + 545, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 601, + 545, + 649 + ], + "spans": [ + { + "bbox": [ + 304, + 601, + 545, + 649 + ], + "type": "text", + "content": "Potential Social Impact. Given our method's capability to reconstruct a realistic personalized character from a monocular video, it is imperative to exercise caution and consider the potential for technology misuse." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 649, + 545, + 709 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 649, + 545, + 709 + ], + "spans": [ + { + "bbox": [ + 304, + 649, + 545, + 709 + ], + "type": "text", + "content": "Acknowledgement. This work was supported by the NSFC project (Nos. 62272134, 62236003, 62072141, 62301298, and 62125107), Shenzhen College Stability Support Plan (Grant No. GXWD20220817144428005), and the Major Key Project of PCL (PCL2023A10-2)." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "641" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Thiemo Alldieck, Marcus Magnor, Weipeng Xu, Christian Theobalt, and Gerard Pons-Moll. Video based reconstruction of 3d people models. In CVPR, 2018. 2, 3, 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 287, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 287, + 169 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 287, + 169 + ], + "type": "text", + "content": "[2] Timur Bagautdinov, Chenglei Wu, Tomas Simon, Fabian Prada, Takaaki Shiratori, Shih-En Wei, Weipeng Xu, Yaser Sheikh, and Jason Saragih. Driving-signal aware full-body avatars. ACM TOG, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 171, + 287, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 287, + 215 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 287, + 215 + ], + "type": "text", + "content": "[3] Jianchuan Chen, Ying Zhang, Di Kang, Xuefei Zhe, Linchao Bao, Xu Jia, and Huchuan Lu. Animatable neural radiance fields from monocular rgb videos. arXiv preprint arXiv:2106.13629, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 217, + 287, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 217, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 53, + 217, + 287, + 259 + ], + "type": "text", + "content": "[4] Xu Chen, Yufeng Zheng, Michael J Black, Otmar Hilliges, and Andreas Geiger. Snarf: Differentiable forward skinning for animating non-rigid neural implicit shapes. In ICCV, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 263, + 287, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 263, + 287, + 296 + ], + "spans": [ + { + "bbox": [ + 53, + 263, + 287, + 296 + ], + "type": "text", + "content": "[5] Chen Guo, Xu Chen, Jie Song, and Otmar Hilliges. Human performance capture from monocular video in the wild. In 3DV, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 298, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 298, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 53, + 298, + 287, + 342 + ], + "type": "text", + "content": "[6] Chen Guo, Tianjian Jiang, Xu Chen, Jie Song, and Otmar Hilliges. Vid2 avatar: 3d avatar reconstruction from videos in the wild via self-supervised scene decomposition. In CVPR, 2023. 2, 3, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 344, + 287, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 344, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 53, + 344, + 287, + 387 + ], + "type": "text", + "content": "[7] Marc Habermann, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Livecap: Real-time human performance capture from monocular video. ACM TOG, 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 390, + 287, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 390, + 287, + 433 + ], + "spans": [ + { + "bbox": [ + 53, + 390, + 287, + 433 + ], + "type": "text", + "content": "[8] Marc Habermann, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Deepcap: Monocular human performance capture using weak supervision. In CVPR, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 436, + 287, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 436, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 436, + 287, + 468 + ], + "type": "text", + "content": "[9] Marc Habermann, Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Real-time deep dynamic characters. ACM TOG, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 471, + 287, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 287, + 503 + ], + "type": "text", + "content": "[10] Tong He, John Collomosse, Hailin Jin, and Stefano Soatto. Geo-pifu: Geometry and pixel aligned implicit functions for single-view human reconstruction. In NeurIPS, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 506, + 287, + 538 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 506, + 287, + 538 + ], + "spans": [ + { + "bbox": [ + 48, + 506, + 287, + 538 + ], + "type": "text", + "content": "[11] Tong He, Yuanlu Xu, Shunsuke Saito, Stefano Soatto, and Tony Tung. Arch++: Animation-ready clothed human reconstruction revisited. In ICCV, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 540, + 287, + 573 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 540, + 287, + 573 + ], + "spans": [ + { + "bbox": [ + 48, + 540, + 287, + 573 + ], + "type": "text", + "content": "[12] Xiaoke Huang, Yiji Cheng, Yansong Tang, Xiu Li, Jie Zhou, and Jiwen Lu. Efficient meshy neural fields for animatable human avatars. arXiv preprint arXiv:2303.12965, 2023. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 575, + 287, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 575, + 287, + 608 + ], + "spans": [ + { + "bbox": [ + 48, + 575, + 287, + 608 + ], + "type": "text", + "content": "[13] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. Arch: Animatable reconstruction of clothed humans. In CVPR, 2020. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 610, + 287, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 610, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 610, + 287, + 643 + ], + "type": "text", + "content": "[14] Boyi Jiang, Yang Hong, Hujun Bao, and Juyong Zhang. Self-con: Self reconstruction your digital avatar from monocular video. In CVPR, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 645, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 645, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 645, + 287, + 677 + ], + "type": "text", + "content": "[15] Tianjian Jiang, Xu Chen, Jie Song, and Otmar Hilliges. Instantavatar: Learning avatars from monocular video in 60 seconds. In CVPR, 2023. 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "text", + "content": "[16] Wei Jiang, Kwang Moo Yi, Golnoosh Samei, Oncel Tuzel, and Anurag Ranjan. Neuman: Neural human radiance field from a single video. In ECCV, 2022. 2, 3, 5, 6" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM TOG, 2023. 2, 3, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 107, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 128 + ], + "type": "text", + "content": "[18] Christoph Lassner and Michael Zollhofer. Pulsar: Efficient sphere-based neural rendering. In CVPR, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 129, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 129, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 308, + 129, + 545, + 160 + ], + "type": "text", + "content": "[19] Ruilong Li, Shan Yang, David A. Ross, and Angjoo Kanazawa. Learn to dance with aist++: Music conditioned 3d dance generation. In ICCV, 2021. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 162, + 545, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 204 + ], + "type": "text", + "content": "[20] Ruilong Li, Julian Tanke, Minh Vo, Michael Zollhöfer, Jürgen Gall, Angjoo Kanazawa, and Christoph Lassner. Tava: Template-free animatable volumetric actors. In ECCV, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 206, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 239 + ], + "type": "text", + "content": "[21] Zhe Li, Zerong Zheng, Yuxiao Liu, Boyao Zhou, and Yebin Liu. Posevocab: Learning joint-structured pose embeddings for human avatar modeling. In SIGGRAPH, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 240, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 240, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 308, + 240, + 545, + 271 + ], + "type": "text", + "content": "[22] Zhe Li, Zerong Zheng, Lizhen Wang, and Yebin Liu. Animatable gaussians: Learning pose-dependent gaussian maps for high-fidelity human avatar modeling. In CVPR, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 273, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 273, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 308, + 273, + 545, + 304 + ], + "type": "text", + "content": "[23] Siyou Lin, Hongwen Zhang, Zerong Zheng, Ruizhi Shao, and Yebin Liu. Learning implicit templates for point-based clothed human modeling. In ECCV, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 306, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 306, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 308, + 306, + 545, + 348 + ], + "type": "text", + "content": "[24] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. ACM TOG, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "type": "text", + "content": "[25] Stephen Lombardi, Tomas Simon, Gabriel Schwartz, Michael Zollhoefer, Yaser Sheikh, and Jason Saragih. Mixture of volumetric primitives for efficient neural rendering. ACM Trans. Graph., 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 394, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 394, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 308, + 394, + 545, + 425 + ], + "type": "text", + "content": "[26] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. ACM TOG, 2015. 3, 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 427, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 427, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 308, + 427, + 545, + 459 + ], + "type": "text", + "content": "[27] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. 3DV, 2024. 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 460, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 491 + ], + "type": "text", + "content": "[28] Qianli Ma, Jinlong Yang, Siyu Tang, and Michael J Black. The power of points for modeling humans in clothing. In ICCV, 2021. 2, 3, 4" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 494, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 494, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 494, + 545, + 536 + ], + "type": "text", + "content": "[29] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 537, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 545, + 568 + ], + "type": "text", + "content": "[30] Gyeongsik Moon, Hyeongjin Nam, Takaaki Shiratori, and Young Mu Lee. 3d clothed human reconstruction in the wild. In ECCV, 2022. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 571, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 571, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 571, + 545, + 613 + ], + "type": "text", + "content": "[31] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3d hands, face, and body from a single image. In CVPR, 2019. 4" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 615, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 615, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 308, + 615, + 545, + 656 + ], + "type": "text", + "content": "[32] Sida Peng, Shangzhan Zhang, Zhen Xu, Chen Geng, Boyi Jiang, Hujun Bao, and Xiaowei Zhou. Animatable neural implicit surfaces for creating avatars from videos. In ICCV, 2021. 3" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "text", + "content": "[33] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In CVPR, 2021. 3" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "642" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "text", + "content": "[34] Sergey Prokudin, Qianli Ma, Maxime Raafat, Julien Valentin, and Siyu Tang. Dynamic point fields. In ICCV, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 150 + ], + "type": "text", + "content": "[35] Edoardo Remelli, Timur Bagautdinov, Shunsuke Saito, Chenglei Wu, Tomas Simon, Shih-En Wei, Kaiwen Guo, Zhe Cao, Fabian Prada, Jason Saragih, et al. Drivable volumetric avatars using texel-aligned features. In SIGGRAPH, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 150, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 150, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 150, + 287, + 183 + ], + "type": "text", + "content": "[36] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 183, + 287, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 183, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 183, + 287, + 216 + ], + "type": "text", + "content": "[37] Darius Rückert, Linus Franke, and Marc Stamminger. Adop: Approximate differentiable one-pixel point rendering. ACM TOG, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 217, + 287, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 259 + ], + "type": "text", + "content": "[38] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In ICCV, 2019. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 261, + 287, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 287, + 294 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 287, + 294 + ], + "type": "text", + "content": "[39] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In CVPR, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 294, + 287, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 294, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 48, + 294, + 287, + 327 + ], + "type": "text", + "content": "[40] Ruizhi Shao, Jingxiang Sun, Cheng Peng, Zerong Zheng, Boyao Zhou, Hongwen Zhang, and Yebin Liu. Control4d: Efficient 4d portrait editing with text. In CVPR, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 327, + 287, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 287, + 369 + ], + "type": "text", + "content": "[41] Shih-Yang Su, Frank Yu, Michael Zollhöfer, and Helge Rhodin. A-nerf: Articulated neural radiance fields for learning human shape, appearance, and pose. In NeurIPS, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 371, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 403 + ], + "type": "text", + "content": "[42] Shih-Yang Su, Timur Bagautdinov, and Helge Rhodin. Danbo: Disentangled articulated neural body representations via graph neural networks. In ECCV, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 404, + 287, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 287, + 426 + ], + "type": "text", + "content": "[43] Shih-Yang Su, Timur Bagautdinov, and Helge Rhodin. Npc: Neural point characters from video. In ICCV, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 426, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 287, + 459 + ], + "type": "text", + "content": "[44] Yu Sun, Qian Bao, Wu Liu, Yili Fu, Michael J Black, and Tao Mei. Monocular, one-stage, regression of multiple 3d people. In ICCV, 2021. 5, 6, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 460, + 287, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 287, + 492 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 287, + 492 + ], + "type": "text", + "content": "[45] Gusi Te, Xiu Li, Xiao Li, Jinglu Wang, Wei Hu, and Yan Lu. Neural capture of animatable 3d human from monocular video. In ECCV, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 492, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 492, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 492, + 287, + 536 + ], + "type": "text", + "content": "[46] Maria Vakalopoulou, Guillaume Chassagnon, Norbert Bus, Rafael Marini, Evangelia I Zacharaki, M-P Revel, and Nikos Paragios. Atlasnet: Multi-atlas non-linear deep networks for medical image segmentation. In MICCAI, 2018. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 537, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 287, + 568 + ], + "type": "text", + "content": "[47] Shaofei Wang, Katja Schwarz, Andreas Geiger, and Siyu Tang. Arah: Animatable volume rendering of articulated human sdfs. In ECCV, 2022. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 570, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 602 + ], + "type": "text", + "content": "[48] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 2004. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 603, + 287, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 287, + 647 + ], + "type": "text", + "content": "[49] Chung-Yi Weng, Brian Curless, Pratul P Srinivasan, Jonathan T Barron, and Ira Kemelmacher-Shlizerman. Humaner: Free-viewpoint rendering of moving people from monocular video. In CVPR, 2022. 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 647, + 287, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 680 + ], + "type": "text", + "content": "[50] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. Synsin: End-to-end view synthesis from a single image. In CVPR, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "text", + "content": "[51] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J Black. Icon: Implicit clothed humans obtained from normals. In CVPR, 2022. 3" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[52] Yuliang Xiu, Jinlong Yang, Xu Cao, Dimitrios Tzionas, and Michael J Black. Econ: Explicit clothed humans optimized via normal integration. In CVPR, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 107, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 162 + ], + "type": "text", + "content": "[53] Feng Xu, Yebin Liu, Carsten Stoll, James Tompkin, Gaurav Bharaj, Qionghai Dai, Hans-Peter Seidel, Jan Kautz, and Christian Theobalt. Video-based characters: creating new human performances from a multi-view video database. ACM TOG, 2011. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 164, + 545, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 164, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 307, + 164, + 545, + 208 + ], + "type": "text", + "content": "[54] Weipeng Xu, Avishek Chatterjee, Michael Zollhöfer, Helge Rhodin, Dushyant Mehta, Hans-Peter Seidel, and Christian Theobalt. Monoperfcap: Human performance capture from monocular video. In SIGGRAPH, 2018. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 209, + 545, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 253 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 253 + ], + "type": "text", + "content": "[55] Yuelang Xu, Benwang Chen, Zhe Li, Hongwen Zhang, Lizhen Wang, Zerong Zheng, and Yebin Liu. Gaussian head avatar: Ultra high-fidelity head avatar via dynamic gaussians. In CVPR, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 255, + 545, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 255, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 307, + 255, + 545, + 289 + ], + "type": "text", + "content": "[56] Wang Yifan, Felice Serena, Shihao Wu, Cengiz Oztireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM TOG, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 290, + 545, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 290, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 307, + 290, + 545, + 323 + ], + "type": "text", + "content": "[57] Zhengming Yu, Wei Cheng, Xian Liu, Wayne Wu, and Kwan-Yee Lin. Monohuman: Animatable human neural field from monocular video. In CVPR, 2023. 2, 3, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 325, + 545, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 325, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 307, + 325, + 545, + 369 + ], + "type": "text", + "content": "[58] Hongwen Zhang, Yating Tian, Xinchi Zhou, Wanli Ouyang, Yebin Liu, Limin Wang, and Zhenan Sun. Pymaf: 3d human pose and shape regression with pyramidal mesh alignment feedback loop. In ICCV, 2021. 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 371, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 371, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 371, + 545, + 415 + ], + "type": "text", + "content": "[59] Hongwen Zhang, Siyou Lin, Ruizhi Shao, Yuxiang Zhang, Zerong Zheng, Han Huang, Yandong Guo, and Yebin Liu. Closet: Modeling clothed humans on continuous surface with explicit template decomposition. In CVPR, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 417, + 545, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 417, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 307, + 417, + 545, + 460 + ], + "type": "text", + "content": "[60] Hongwen Zhang, Yating Tian, Yuxiang Zhang, Mengcheng Li, Liang An, Zhenan Sun, and Yebin Liu. Pymaf-x: Towards well-aligned full-body model regression from monocular images. IEEE T-PAMI, 2023. 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 462, + 545, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 462, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 307, + 462, + 545, + 496 + ], + "type": "text", + "content": "[61] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 497, + 545, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 497, + 545, + 551 + ], + "spans": [ + { + "bbox": [ + 307, + 497, + 545, + 551 + ], + "type": "text", + "content": "[62] Yuxiang Zhang, Hongwen Zhang, Liangxiao Hu, Jiajun Zhang, Hongwei Yi, Shengping Zhang, and Yebin Liu. Proxycap: Real-time monocular full-body capture in world space via human-centric proxy-to-motion learning. In CVPR, 2024. 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 554, + 545, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 554, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 307, + 554, + 545, + 587 + ], + "type": "text", + "content": "[63] Hao Zhao, Jinsong Zhang, Yu-Kun Lai, Zerong Zheng, Yingdi Xie, Yebin Liu, and Kun Li. High-fidelity human avatars from a single rgb camera. In CVPR, 2022. 2, 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 588, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 588, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 307, + 588, + 545, + 632 + ], + "type": "text", + "content": "[64] Shunyuan Zheng, Boyao Zhou, Ruizhi Shao, Boring Liu, Shengping Zhang, Liqiang Nie, and Yebin Liu. Gps-gaussian: Generalizable pixel-wise 3d gaussian splattering for real-time human novel view synthesis. In CVPR, 2024. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 634, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 634, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 634, + 545, + 667 + ], + "type": "text", + "content": "[65] Yufeng Zheng, Wang Yifan, Gordon Wetzstein, Michael J Black, and Otmar Hilliges. Pointavatar: Deformable point-based head avatars from videos. In CVPR, 2023. 2, 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "text", + "content": "[66] Zerong Zheng, Tao Yu, Yebin Liu, and Qionghai Dai. Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction. IEEE T-PAMI, 2021. 3" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "643" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 183 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[67] Zerong Zheng, Han Huang, Tao Yu, Hongwen Zhang, Yandong Guo, and Yebin Liu. Structured local radiance fields for human avatar modeling. In CVPR, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "type": "text", + "content": "[68] Zerong Zheng, Xiaochen Zhao, Hongwen Zhang, Boning Liu, and Yebin Liu. Avatarrex: Real-time expressive full-body avatars. ACM TOG, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 183 + ], + "type": "text", + "content": "[69] Boyao Zhou, Jean-Sébastien Franco, Federica Bogo, Bugra Tekin, and Edmond Boyer. Reconstructing human body mesh from point clouds by adversarial gp network. In ACCV, 2020. 3" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "644" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/GaussianAvatars_ Photorealistic Head Avatars with Rigged 3D Gaussians/41411f0e-3b8e-40c3-bbbd-d16b793f1be0_content_list.json b/2024/GaussianAvatars_ Photorealistic Head Avatars with Rigged 3D Gaussians/41411f0e-3b8e-40c3-bbbd-d16b793f1be0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0affebc52743c04a3eb29d26bec55ac47dcc3360 --- /dev/null +++ b/2024/GaussianAvatars_ Photorealistic Head Avatars with Rigged 3D Gaussians/41411f0e-3b8e-40c3-bbbd-d16b793f1be0_content_list.json @@ -0,0 +1,1619 @@ +[ + { + "type": "text", + "text": "GaussianAvatars: Photorealistic Head Avatars with Rigged 3D Gaussians", + "text_level": 1, + "bbox": [ + 114, + 130, + 854, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shenhan Qian", + "bbox": [ + 158, + 179, + 281, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tobias Kirschstein", + "bbox": [ + 321, + 180, + 477, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Simon Giebenhain", + "bbox": [ + 321, + 199, + 475, + 214 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Liam Schoneveld²", + "bbox": [ + 519, + 180, + 666, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Matthias Nießner1", + "bbox": [ + 519, + 199, + 663, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Davide Davoli3", + "bbox": [ + 707, + 180, + 831, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Technical University of Munich", + "bbox": [ + 127, + 222, + 388, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Woven by Toyota", + "bbox": [ + 431, + 222, + 581, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Toyota Motor Europe NV/SA", + "bbox": [ + 625, + 223, + 870, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "associated partner by contracted service", + "bbox": [ + 643, + 238, + 854, + 250 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f59dab12bb1ea77179eb856be965b62c7c983a18f3ffe57ab32782be22c1bb9f.jpg", + "image_caption": [ + "Figure 1. We introduce GaussianAvatars, a new method to create photorealistic head avatars from multi-view videos. Our avatar is represented by 3D Gaussian splats rigged to a parametric face model. We can fully control and animate our avatars in terms of pose, expression, and view point as shown in the example renderings above." + ], + "image_footnote": [], + "bbox": [ + 76, + 282, + 220, + 406 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3e6fbc01b4fa6c1c8d37c709694d69dc3c88fdc2f2ecee4018946d716093785c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 225, + 282, + 357, + 405 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/918425896ae7b016e99a03843bfd1e49dd6bfdbb52cf760dd0fae1634a466322.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 281, + 483, + 405 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/47968548cd067be4da20614ef275e3f13d4b28c33c946a6f875ee623b2e04447.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 495, + 281, + 616, + 405 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8f621531ff7e4ad335a8b7ff4740587045c33681115361931c9dcb852e5c0550.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 290, + 750, + 405 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b178e2f8cd547acafcf9c25dd607f9c79419a78a92102eddb821de93fc45938d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 753, + 290, + 890, + 405 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 470, + 313, + 487 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce GaussianAvatars1, a new method to create photorealistic head avatars that are fully controllable in terms of expression, pose, and viewpoint. The core idea is a dynamic 3D representation based on 3D Gaussian splats that are rigged to a parametric morphable face model. This combination facilitates photorealistic rendering while allowing for precise animation control via the underlying parametric model, e.g., through expression transfer from a driving sequence or by manually changing the morphable model parameters. We parameterize each splat by a local coordinate frame of a triangle and optimize for explicit displacement offset to obtain a more accurate geometric representation. During avatar reconstruction, we jointly optimize for the morphable model parameters and Gaussian splat parameters in an end-to-end fashion. We demonstrate the animation capabilities of our photorealistic avatar in several challenging scenarios. For instance, we show reenactments from a driving video, where our method outperforms existing works by a significant margin.", + "bbox": [ + 73, + 503, + 473, + 791 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 820, + 209, + 835 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Creating animatable avatars of human heads has been a longstanding problem in computer vision and graphics.", + "bbox": [ + 75, + 845, + 468, + 876 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In particular, the ability to render photorealistic dynamic avatars from arbitrary viewpoints enables numerous applications in gaming, movie production, immersive telepresence, and augmented or virtual reality. For such applications, it is also crucial to be able to control the avatar, and for it to generalize well to novel poses and expressions.", + "bbox": [ + 500, + 472, + 890, + 563 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Reconstructing a 3D representation able to jointly capture the appearance, geometry and dynamics of human heads represents a major challenge for high-fidelity avatar generation. The under-constrained nature of this reconstruction problem significantly complicates the task of achieving a representation that combines novel-view rendering photorealism with expression controllability. Moreover, extreme expressions and facial details, like wrinkles, the mouth interior, and hair, are difficult to capture, and can produce visual artifacts easily noticed by humans.", + "bbox": [ + 496, + 565, + 892, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural Radiance Fields (NeRF [28]) and its variants [2, 6, 29] have shown impressive results in reconstructing static scenes from multi-view observations. Follow-up works have extended NeRF to model dynamic scenes for both arbitrary [22, 30, 31] and human-tailored scenarios [13, 18]. These works achieve impressive results for novel view rendering; however, they lack controllability and as such do not generalize well to novel poses and expressions.", + "bbox": [ + 496, + 717, + 892, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The recent 3D Gaussian Splitting method [15] achieves even higher rendering quality than NeRF for novel-view synthesis with real-time performance, optimizing for discrete geometric primitives (3D Gaussians) throughout 3D", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1Project page: https://shenhanqian.github.io/gaussian-avatars", + "bbox": [ + 94, + 886, + 418, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20299", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "space. This method has been extended to capture dynamic scenes by building explicit correspondences across time steps [27]; however, they do not allow for animations of the reconstructed outputs.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we propose GaussianAvatars, a method for dynamic 3D representation of human head based on 3D Gaussian splats that are rigged to a parametric morphable face model. Given a FLAME [21] mesh, we initialize a 3D Gaussian at the center of each triangle. When the FLAME mesh is animated, each Gaussian then translates, rotates, and scales according to its parent triangle. The 3D Gaussians then form a radiance field on top of the mesh, compensating for regions where the mesh is either not accurately aligned, or is incapable of reproducing certain visual elements. To achieve high fidelity of the reconstructed avatar, we introduce a binding inheritance strategy to support Gaussian splats without losing controllability. We also explore balancing fidelity and robustness to animate the avatars with novel expressions and poses. Our method outperforms existing works by a significant margin on both novel-view rendering and reenactment from a driving video.", + "bbox": [ + 75, + 152, + 470, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are as follows:", + "bbox": [ + 76, + 412, + 295, + 426 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose GaussianAvatars, a method to create animatable head avatars by rigging 3D Gaussians to parametric mesh models.", + "- We design a binding inheritance strategy to support adding and removing 3D Gaussians without losing controllability." + ], + "bbox": [ + 76, + 429, + 467, + 518 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 534, + 218, + 549 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Radiance Field Reconstruction", + "text_level": 1, + "bbox": [ + 76, + 559, + 349, + 573 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "NeRF [28] stores the radiance field of a scene in a neural network and provides photorealistic renderings of novel views with volumetric rendering. Later work [37, 47] represents the scene as voxel grids, achieving comparable rendering quality in a shorter time. Efficiency can be further improved by employing compression methods such as voxel hashing [29] or tensor decomposition [6]. PointNeRF [43] uses point clouds as a scene representation, whereas 3D Gaussian Splatting [15] uses anisotropic 3D Gaussians that are rendered by sorting and rasterization, achieving superior visual quality with real-time rendering. Mixture of Volumetric Primitives [25] uses surface-aligned volumes to achieve fast rendering with high visual fidelity. It also applies residual transformations to primitives. Our method follows 3D Gaussian Splatting [15], benefiting from the expressiveness of anisotropic Gaussians.", + "bbox": [ + 75, + 583, + 467, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A common paradigm to model a dynamic scene is directly storing it in an MLP with 4D coordinates as the input [9, 42] or a 4D tensor with the time dimension or spatial dimensions compressed [1, 4, 7, 36]. These methods can faithfully replay a dynamic scene and produce realistic", + "bbox": [ + 75, + 825, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "novel-view rendering but lack an explicit handle to manipulate the content. Another paradigm learns a static canonical space and the maps time steps back to the canonical space via separate MLPs [30, 31, 34]. Instead of using a deformation MLP, a proxy geometry provides more direct controllability. Liu et al. [24] warp points from an observed space to the canonical space based on the movement of the nearest triangle on an SMPL [26] mesh. Peng et al. [33] deform points with the skeleton of SMPL and neural blending weights. Concurrent works create human body avatars with forward deformation [14, 19, 20, 23, 46] and cage-based deformation [51]. Unlike these methods, we directly attach 3D Gaussians to triangles and explicitly move them, obviating the need for a canonical space and enabling effective mesh finetuning.", + "bbox": [ + 496, + 90, + 890, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Human Head Reconstruction and Animation", + "text_level": 1, + "bbox": [ + 498, + 325, + 879, + 340 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Head avatar creation advanced through the uptake of differentiable rendering and scene representations. Thies et al. [39] instigated a shift toward digital avatars with real-time face tracking and authentic face reenactment. Advancements in image synthesis with neural networks [5, 40] boosted the controllability of head avatars from lip-syncing to expression transfer and head motions [16, 38, 45]. Gafni et al. [8] learn a NeRF conditioned on an expression vector from monocular videos. Grassal et al. [11] subdivide and add offsets to FLAME [21] to enhance its geometry and enable a dynamic texture via an expression-dependent texture field. IMavatar [49] learns a 3D morphable head avatar with neural implicit functions, solving for a map from observed to canonical space via iterative root-finding. HeadNeRF [12] learns a NeRF-based parametric head model with 2D neural rendering for efficiency.", + "bbox": [ + 496, + 349, + 890, + 590 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "INSTA [52] deforms query points to a canonical space by finding the nearest triangle on a FLAME [21] mesh, and combines this with InstantNGP [29] to achieve fast rendering. Like INSTA, we also use triangles to warp the scene, but we build a consistent correspondence between a 3D Gaussian and a triangle instead of querying the nearest one for each timestep. Zheng et al. [50] explore point-based representations with differential point splatting. They define a point set in canonical space and learn a deformation field conditioned on FLAME's expression vectors to animate the head avatar. While the scale of a point has to be manually set for their method, it is an estimizable parameter for 3D Gaussians. NeRFBlendShape [10] models a dynamic scene by blending hash tables with 3DMM parameters [3, 32]. AvatarMAV [44] decouples motion and appearance, blending voxel grids only for the motion field.", + "bbox": [ + 496, + 590, + 890, + 832 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 844, + 589, + 859 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As shown in Fig. 2, the input to our method is a multi-view video recording of a human head. For each time step, we", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20300", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a1c9c7ee575ff82d5eaa16685d863275bd6674ba0703d139b8583552bbe1f504.jpg", + "image_caption": [ + "Figure 2. Overview. Our method binds 3D Gaussian splats to a FLAME [21] mesh locally. We take the tracked mesh for each frame and transform the splats from local to global space before rendering them with 3D Gaussian Splatting [15]. We optimize the splats in the local space by minimizing color loss on the rendering. We add and remove splats adaptively with their binding relation to triangles inherited so that all splats remain rigged throughout the optimization procedure. Further, we regularize the position and scaling of 3D Gaussian splats to suppress artifacts during animation." + ], + "image_footnote": [], + "bbox": [ + 76, + 85, + 893, + 429 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "use a photometric head tracker based on [39] to fit FLAME [21] parameters with multi-view observations and known camera parameters. The FLAME meshes have vertices at varied positions but share the same topology. Therefore, we can build a consistent connection between triangles of the mesh and 3D Gaussian splats (Sec. 3.2). The splats are rendered into images via a differentiable tile rasterizer [15]. These images are then supervised by the groundtruth images towards learning a photorealistic human head avatar. As per static scenes, it is also necessary to densify and prune Gaussian splats for optimal quality, with a set of adaptive density control operations [15]. To achieve this without breaking the connection between triangles and splats, we design a binding inheritance strategy (Sec. 3.3) so that new Gaussian points remain rigged to the FLAME mesh. In addition to the color loss, we also find it crucial to regularize the local position and scaling of Gaussian splats to avoid quality degradation under novel expressions and poses (Sec. 3.4).", + "bbox": [ + 75, + 535, + 472, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 76, + 830, + 207, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Gaussian Splitting [15] provides a solution to reconstruct a static scene with anisotropic 3D Gaussians given images and camera parameters. A scene is represented by a", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "set of Gaussian splats, each defined by a covariance matrix $\\Sigma$ centered at point (mean) $\\mu$ :", + "bbox": [ + 498, + 535, + 890, + 566 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nG (\\boldsymbol {x}) = e ^ {- \\frac {1}{2} (\\boldsymbol {x} - \\boldsymbol {\\mu}) ^ {T} \\Sigma^ {- 1} (\\boldsymbol {x} - \\boldsymbol {\\mu})} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 571, + 890, + 592 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Note that covariance matrices have physical meaning only when they are semi-definite, which cannot be guaranteed for an optimization process with gradient descent. Therefore, Kerbl et al. [15] first define a parametric ellipse with a scaling matrix $S$ and a rotation matrix $R$ , then construct the covariance matrix by:", + "bbox": [ + 496, + 599, + 890, + 690 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Sigma = R S S ^ {T} R ^ {T}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 638, + 696, + 890, + 714 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Practically, an ellipse is stored as a position vector $\\pmb{\\mu} \\in \\mathbb{R}^3$ , a scaling vector $\\pmb{s} \\in \\mathbb{R}^3$ , and a quaternion $\\pmb{q} \\in \\mathbb{R}^4$ . In our paper, we use $\\pmb{r} \\in \\mathbb{R}^{3 \\times 3}$ to denote the corresponding rotation matrix to $\\pmb{q}$ .", + "bbox": [ + 496, + 722, + 890, + 782 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For rendering, the color $C$ of a pixel is computed by blending all 3D Gaussians overlapping the pixel:", + "bbox": [ + 496, + 784, + 890, + 814 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {C} = \\sum_ {i = 1} \\boldsymbol {c} _ {i} \\alpha_ {i} ^ {\\prime} \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\alpha_ {j} ^ {\\prime}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 821, + 890, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $c_{i}$ is the color of each point modeled by 3-degree spherical harmonics. The blending weight $\\alpha^{\\prime}$ is given by", + "bbox": [ + 496, + 869, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20301", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "evaluating the 2D projection of the 3D Gaussian multiplied by a per-point opacity $\\alpha$ . The Gaussian splats are sorted by depth before blending to respect visibility order.", + "bbox": [ + 75, + 90, + 468, + 137 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. 3D Gaussian Rigging", + "text_level": 1, + "bbox": [ + 76, + 146, + 274, + 162 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The key component of our method is how we build connections between the FLAME [21] mesh and 3D Gaussian splats. Initially, we pair each triangle of the mesh with a 3D Gaussian and let the 3D Gaussian move with the triangle across time steps. In other words, the 3D Gaussian is static in the local space of its parent triangle but dynamic in the global (metric) space as the triangle moves. Given the vertices and edges of a triangle, we take the mean position $T$ of the vertices as the origin of the local space. Then, we concatenate the direction vector of one of the edges, the normal vector of the triangle, and their cross product as column vectors to form a rotation matrix $R$ , which describes the orientation of the triangle in the global space. We also compute a scalar $k$ by the mean length of one of the edges and its perpendicular to describe the triangle scaling.", + "bbox": [ + 75, + 169, + 468, + 396 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the paired 3D Gaussian of a triangle, we define its location $\\mu$ , rotation $r$ , and anisotropic scaling $s$ all in the local space. We initialize the location $\\mu$ at the local origin, the rotation $r$ as an identity rotation matrix, and the scaling $s$ as a unit vector. At rendering time, we convert these properties into the global space by:", + "bbox": [ + 75, + 396, + 468, + 487 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {r} ^ {\\prime} = \\boldsymbol {R} \\boldsymbol {r}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 500, + 468, + 513 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\mu} ^ {\\prime} = k \\boldsymbol {R} \\boldsymbol {\\mu} + \\boldsymbol {T}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 517, + 468, + 534 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {s} ^ {\\prime} = k \\boldsymbol {s}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 537, + 468, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We incorporate triangle scaling in Eqs. (5) and (6) so that the local position and scaling of a 3D Gaussian are defined relative to the absolute scale of a triangle. This enables an adaptive step size in the metric space with a constant learning rate for parameters defined in the local space. For example, a 3D Gaussian paired with a smaller triangle will move slower in an iteration step than those paired with large triangles. This also makes interpreting the parameters regarding the distance from the triangle center easier.", + "bbox": [ + 75, + 565, + 468, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Binding Inheritance", + "text_level": 1, + "bbox": [ + 76, + 710, + 267, + 726 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Only having the same numbers of Gaussian splats as the triangles is insufficient to capture details. For instance, representing a curved hair strand requires multiple splats, while a triangle on the scalp may intersect with several strands. Therefore, we also need the adaptive density control strategy [15], which adds and removes splats based on the view-space positional gradient and the opacity of each Gaussian.", + "bbox": [ + 75, + 734, + 468, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each 3D Gaussian with a large view-space positional gradient, we split it into two smaller ones if it is large or clone it if it is small. We conduct this in the local space and ensure a newly created Gaussian is close to the old one that", + "bbox": [ + 75, + 840, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "triggers this densification operation. Then, it is reasonable to bind a new 3D Gaussian to the same triangle as the old one because it was created to enhance the fidelity of the local region. Therefore, each 3D Gaussian must carry one more parameter, the index of its parent triangle, to enable binding inheritance during densification.", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Besides densification, we also use the pruning operation as a part of the adaptive density control strategy [15]. It periodically resets the opacity of all splats close to zero and removes points with opacity below a threshold. This technique is effective in suppressing floating artifacts, however, such pruning can also cause problems in a dynamic scene. For instance, regions of the face that are often occluded (such as eyeball triangles), can be overly sensitive to this pruning strategy, and often end up with few or no attached Gaussians. To prevent this, we keep track of the number of splats attached to each triangle, and ensure that every triangle always has at least one splat attached.", + "bbox": [ + 496, + 181, + 890, + 363 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Optimization and Regularization", + "text_level": 1, + "bbox": [ + 500, + 373, + 790, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We supervise the rendered images with a combination of $\\mathcal{L}_1$ term and a D-SSIM term following [15]:", + "bbox": [ + 498, + 396, + 888, + 426 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {r g b}} = (1 - \\lambda) \\mathcal {L} _ {1} + \\lambda \\mathcal {L} _ {\\mathrm {D} - \\mathrm {S S I M}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 439, + 890, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with $\\lambda = 0.2$ . This already results in good re-rendering quality without additional supervision, such as depth or silhouette supervision, thanks to the powerful tile rasterizer [15]. We found however, that if we try to animate these splats via FLAME [21] to novel expressions and poses, large spike- and blob-like artifacts appear wildly throughout the scene. This is due to a poor alignment between the Gaussian splats and the triangles.", + "bbox": [ + 496, + 467, + 890, + 587 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Position loss with threshold. A basic hypothesis behind 3D Gaussian rigging is that the Gaussian splats should roughly match the underlying mesh. They should also match their locations; for instance a Gaussian representing a spot on the nose should not be rigged to a triangle on the cheek. Although our splats are initialized at triangle centers, and new splats are added nearby these existing ones, it is not guaranteed that the primitives remain close to their parent triangle after the optimization. To address this, we regularize the local position of each Gaussian by:", + "bbox": [ + 496, + 588, + 890, + 739 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {p o s i t i o n}} = \\| \\max (\\boldsymbol {\\mu}, \\epsilon_ {\\text {p o s i t i o n}}) \\| _ {2}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 751, + 890, + 768 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\epsilon_{\\mathrm{position}} = 1$ is a threshold that tolerates small errors within the scaling of its parent triangle.", + "bbox": [ + 496, + 779, + 888, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Scaling loss with threshold. Aside from position, the scaling of 3D Gaussians is even more essential for the visual quality during animation. Specifically, if a 3D Gaussian is large in comparison to its parent triangle, small rotations of the triangle – barely noticeable at the scale of the triangle – will be magnified by the scale of the 3D Gaussian, resulting", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "footer", + "text": "20302", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "in unpleasant jittering artifacts. To mitigate this, we also regularize the local scale of each 3D Gaussian by:", + "bbox": [ + 75, + 90, + 468, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s c a l i n g}} = \\left\\| \\max \\left(\\boldsymbol {s}, \\epsilon_ {\\text {s c a l i n g}}\\right) \\right\\| _ {2}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 131, + 468, + 148 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\epsilon_{\\mathrm{scaling}} = 0.6$ is a threshold that disables this loss term when the scale of a Gaussian less than $0.6\\times$ the scale of its parent triangle. This $\\epsilon_{\\mathrm{scaling}}$ -tolerance is indispensable; without it, the Gaussian splats shrink excessively, causing rendering speeds to deteriorate, as camera rays need to hit more splats before zero transmittance is reached.", + "bbox": [ + 75, + 156, + 468, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our final loss function is thus:", + "bbox": [ + 96, + 247, + 297, + 260 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {r g b}} + \\lambda_ {\\text {p o s i t i o n}} \\mathcal {L} _ {\\text {p o s i t i o n}} + \\lambda_ {\\text {s c a l i n g}} \\mathcal {L} _ {\\text {s c a l i n g}}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 271, + 468, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{\\mathrm{position}} = 0.01$ and $\\lambda_{\\mathrm{scaling}} = 1$ . Note that we only apply $\\mathcal{L}_{\\mathrm{position}}$ and $\\mathcal{L}_{\\mathrm{scaling}}$ to visible splats. Thereby, we only regularize points when the color loss $\\mathcal{L}_{\\mathrm{rgb}}$ is present. This helps to maintain the learned structure of often-occluded regions such as teeth and eyeballs.", + "bbox": [ + 75, + 296, + 468, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. We use Adam [17] for parameter optimization (the same hyperparameter values are used across all subjects). We set the learning rate to 5e-3 for the position and 1.7e-2 for the scaling of 3D Gaussians and keep the same learning rates as 3D Gaussian Splatting [15] for the rest of the parameters. Alongside the Gaussian splat parameters, we also finetune the translation, joint rotation, and expression parameters of FLAME [21] for each timestep, using learning rates 1e-6, 1e-5, and 1e-3, respectively. We train for 600,000 iterations, and exponentially decay the learning rate for the splat positions until the final iteration, where it reaches $0.01 \\times$ the initial value. We enable adaptive density control with binding inheritance every 2,000 iterations, from iteration 10,000 until the end. Every 60,000 iterations, we reset the Gaussians' opacities. We use a photo-metric head tracker to obtain the FLAME parameters, including shape $\\beta$ , translation $t$ , pose $\\theta$ , expression $\\psi$ , and vertex offset $\\Delta v$ in the canonical space.", + "bbox": [ + 75, + 372, + 468, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 656, + 207, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Setup", + "text_level": 1, + "bbox": [ + 76, + 681, + 158, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Settings. We evaluate the quality of head avatars with three settings: 1) novel-view synthesis: driving an avatar with head poses and expressions from training sequences and rendering from a held-out viewpoint. 2) self-reenactment: driving an avatar with unseen poses and expressions from a held-out sequence of the same subject and rendering all 16 camera views. 3) cross-identity reenactment: animating an avatar with poses and expressions from another subject.", + "bbox": [ + 75, + 703, + 468, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. We conduct experiments on video recordings of 9 subjects from the NeRSample [18] dataset. All recordings contain 16 views covering the front and sides of a subject. For each subject, we take 11 video sequences and down-sample images to a resolution of $802 \\times 550$ . Participants", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c3e9ad8c45420afd947ff7949c58cdbaee4b09780c0ba91e78b72a51bd64c5fa.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Novel-View SynthesisSelf-Reenactment
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
AvatarMAV [44]29.50.9130.15224.30.8870.168
PointAvatar [50]25.80.8930.09723.40.8840.102
INSTA [52]26.70.8990.12226.30.9060.110
Ours31.60.9380.06526.00.9100.076
", + "bbox": [ + 500, + 89, + 890, + 193 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparison with state-of-the-art methods. Green indicates the best and yellow indicates the second.", + "bbox": [ + 500, + 204, + 888, + 234 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "were instructed to perform specific expressions or emotions in 10 sequences and freely perform in the last one. To conduct quantitative evaluation for each subject, we train our method and three baselines (INSTA [52], PointAvatar [50], and AvatarMAV [44]) using 9 out of the 10 prescribed sequences and 15 out of 16 available cameras. We use the $11^{\\mathrm{th}}$ (free-performance) sequence to visually assess cross-identity reenactment capabilities. Please refer to the supplementary material for details of the dataset and baselines.", + "bbox": [ + 498, + 263, + 890, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Head Avatar Reconstruction and Animation", + "text_level": 1, + "bbox": [ + 498, + 409, + 874, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate the reconstruction quality of avatars by novel-view synthesis and animation fidelity by self-reenactment. Fig. 3 shows qualitative comparisons. For novel-view synthesis, all methods produce plausible rendering results. A close inspection of PointAvatar's [50] results show dotted artifacts, owing its fixed point size. In our case, the anisotropic scaling of 3D Gaussians alleviates this issue.", + "bbox": [ + 496, + 431, + 890, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "INSTA [52] shows clean results for the face, however regions around the neck and shoulder can be noisy, as the tracked FLAME meshes are often misaligned in those regions. This causes issues for INSTA, as its warping process is based on the nearest triangle. In the case of our method, each Gaussian splat is rigged to a consistent triangle, regardless of the pose or expression. When the tracked mesh is inaccurate, the positional gradient to a Gaussian splat can consistently back-propagate to the same triangle. This enables misalignments due to incorrect FLAME tracking to be corrected while the 3D Gaussians are being optimized.", + "bbox": [ + 496, + 537, + 890, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "AvatarMAV [44] shows comparable quality to other methods in novel-view synthesis but struggles with novel-expression synthesis. This is because it only uses the expression vector of a 3DMM as conditioning. Since the control from the expression vectors to the deformation bases must be learned, it struggles to reproduce expressions that are far from the training distribution. Similar conclusions can be drawn from the quantitative comparison shown in Tab. 1. Our approach outperforms others by a large margin regarding metrics for novel-view synthesis. Our method also stands out in self-reenactment, with significantly lower perceptual differences in terms of LPIPS [48]. Note that self-reenactment is based on tracked FLAME meshes that", + "bbox": [ + 496, + 704, + 892, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20303", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/63b0543c00372643cd9905b20c1c1dc1ce384518bff6d979432871d41c22c233.jpg", + "image_caption": [ + "Figure 3. Qualitative comparison on novel-view synthesis and self-reenactment of head avatars. Our method outperforms state-of-the-art methods by producing significantly sharper rendering outputs. We obtain precise reconstruction of details such as reflective light on eyes, hair strands, teeth, etc. Our results for self-reenactment show more accurate expressions compare to baselines." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 893, + 801 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20304", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8d168d18932ce5a214ea0236828bf75ca52660c73dfc0e91ae22a35af2241f43.jpg", + "image_caption": [ + "Figure 4. Cross-identity reenactment of head avatars. We use the tracked FLAME expression and pose parameters of source actors to drive the reconstructed avatars. Our method produces high-quality rendering and transfers expressions vividly, while baseline methods suffer from artifacts and generalize poorly to novel expressions." + ], + "image_footnote": [], + "bbox": [ + 84, + 92, + 890, + 549 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "may not perfectly align with the target images, thus bringing disadvantages to our results with more visual details regarding pixel-wise metrics such as PSNR.", + "bbox": [ + 75, + 627, + 468, + 671 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For a real-world test of avatar animation, we conduct experiments on cross-identity reenactment in Fig. 4. Our avatars accurately reproduce eye blinks and mouth movements from source actors showing lively, complex dynamics such as wrinkles. INSTA [52] suffers from aliasing artifacts when the avatars move beyond the occupancy grid of I-NGP [29] optimized for training sequences. The movement of results from PointAvatar [50] is not precise because its deformation space is not guaranteed to be consistent with FLAME. AvatarMAV [44] exhibits large degradations in reenactment due to a lack of deformation priors.", + "bbox": [ + 75, + 672, + 472, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 847, + 230, + 863 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To validate the effectiveness of our method components, we deactivate each of them and report results in Tab. 2.", + "bbox": [ + 75, + 869, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d66860fcbeff6c9a795e3ddf45bc43e2c4b25c44f93ae5d501d4a4f498d9040d.jpg", + "image_caption": [ + "Figure 5. Fine-tuning FLAME parameters leads to better alignment of the mesh to the input image. In the example above, the movements of cheeks and lips are better captured after fine-tuning." + ], + "image_footnote": [], + "bbox": [ + 500, + 625, + 619, + 763 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c7e03eab471cd10db96f4e471599724263803167295ac80cb4664fa1966b2009.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 625, + 754, + 765 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/062dfa4389a1cdd500091b7c605f119e83a99bd3c1bc38b645d913a485d5074b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 771, + 625, + 890, + 765 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Adaptive density control with binding inheritance. Without binding inheritance, we lose the ability to add and remove Gaussian splats besides the initial ones. With a lim", + "bbox": [ + 498, + 854, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20305", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ba7ca60a662282e1923c158487d90e14d05c283010b1bfc568485a789c2fbbd5.jpg", + "table_caption": [], + "table_footnote": [ + "Table 2. Ablation study on subject #304. Green indicates the best and yellow indicates the second. “ADC” refers to Adaptive Density Control with binding inheritance. ‘FLAME ft.’ refers to FLAME parameter fine-tuning." + ], + "table_body": "
Novel-ViewSelf-Reenactment
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Ours28.80.8830.09825.10.8530.101
w/o ADC26.80.8540.20625.10.8600.183
w/o Lscaling28.00.8770.11424.90.8520.109
w/o εscaling25.00.8330.19524.10.8430.176
w/o Lposition29.70.8940.09124.90.8510.096
w/o εposition28.70.8820.10525.00.8550.106
w/o FLAME ft.26.10.8550.13125.50.8620.124
", + "bbox": [ + 76, + 88, + 460, + 234 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ited number of splats, the scaling of each splat increases to occupy the same space, causing blurry renderings, leading to huge fidelity loss, as shown in the second row of Tab. 2.", + "bbox": [ + 75, + 333, + 468, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Regularization on the scaling of Gaussian splats. Without the scaling loss, artifacts show up as spikes (the middle column in Fig. 6), and slightly worsen the image quality (the third row in Tab. 2). But when we use the scaling loss without an error tolerance, all metrics drastically deteriorate (the fourth row in Tab. 2) because all the splats are regularized to be infinitely small, making it hard to construct opaque surfaces.", + "bbox": [ + 75, + 378, + 468, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Regularization on the position of Gaussian splats. When disabling the threshold for the position loss, we get slightly worse metrics (the sixth row in Tab. 2). But when turning off the position loss, the metrics on novel-view synthesis become the best in the table (the fifth row in Tab. 2). This is reasonable because without constraints to the mesh Gaussian splats can freely move in the space to achieve minimal re-rendering error. However, since the distribution of Gaussian splats are overfitted to training frames, the avatar shows artifacts such as cracks and fly-around blobs (the left column in Fig. 6) in novel expressions and poses. Therefore, the position loss is necessary to enforce a conservative deviation of splats from the mesh surface so that we can animate the reconstructed avatar with unseen facial motion.", + "bbox": [ + 75, + 500, + 468, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "FLAME fine-tuning. Thanks to our consistent binding between Gaussian splats and triangles, we can effectively fine-tune FLAME parameters while optimizing Gaussian splats (Fig. 5). As shown in the last row in Tab. 2, turning off the fine-tuning negatively affects image quality for novel-view synthesis. For self-reenactment, fine-tuning FLAME parameters also leads to lower perceptual error.", + "bbox": [ + 75, + 710, + 468, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Limitations and Potential Negative Impacts", + "text_level": 1, + "bbox": [ + 76, + 829, + 465, + 847 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our approach, based on 3D Gaussian splats, directly captures the radiance field without decoupling material and lighting. Hence, relighting the avatar is not feasible within", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3ae5dfe4b1dbc53e16309a6ca9b831392ec6b77995354bd7e0b1b12f65ccc3ef.jpg", + "image_caption": [ + "Figure 6. The position loss and the scaling loss helps prevent artifacts during animation with novel expressions and poses." + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 890, + 246 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "our current approach. Additionally, we lack control over areas that are not modeled by FLAME such as hair and other accessories. Here, we believe that modeling these parts directly would provide an interesting future research avenue; for instance, recent methods show promising result in reconstructing hair [35, 41].", + "bbox": [ + 496, + 309, + 890, + 400 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The creation of photorealistic avatars may introduce a wide spectrum of malevolent uses. Privacy violations are a significant concern, with potential for unauthorized manipulation of people's likeness. Moreover the creation of deceptive content, like deepfake videos, can lead to misinformation, defamation and harm to reputations. Identity theft is an additional risk, as avatars may be exploited for impersonation and fraudulent activities. We strongly condemn unauthorized and malevolent uses of this technology, emphasizing the need for ethical considerations in all applications utilizing our method.", + "bbox": [ + 496, + 401, + 892, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 580, + 617, + 594 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "GaussianAvatars is a novel approach that creates photorealistic avatars from video sequences. It features a dynamic 3D representation based on 3D Gaussian splats that are rigged to a parametric morphable face model. This enables flexible control and precise attribute transfer of a dynamic, photorealistic avatar. The set of Gaussian splats can deviate from the mesh surface to compensate the absence or inaccuracy of the morphable model, exhibiting astonishing ability to model fine details on human heads. Our approach outperforms state-of-the-art methods on image quality and expression accuracy for a large margin, indicating a large potential for more applications in related domains.", + "bbox": [ + 496, + 604, + 890, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 800, + 666, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by Toyota Motor Europe and Woven by Toyota. This work was also supported by the ERC Starting Grant Scan2CAD (804724). We thank Justus Thies, Andrei Burov, and Lei Li for constructive discussions and Angela Dai for the video voice-over.", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20306", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Benjamin Attal, Jia-Bin Huang, Christian Richardt, Michael Zollhoefer, Johannes Kopf, Matthew O'Toole, and Changil Kim. Hyperreel: High-fidelity 6-dof video with ray-conditioned sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16610-16620, 2023. 2", + "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 1", + "[3] Volker Blanz and Thomas Vetter. Face recognition based on fitting a 3d morphable model. IEEE Transactions on pattern analysis and machine intelligence, 25(9):1063-1074, 2003. 2", + "[4] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 130-141, 2023. 2", + "[5] Caroline Chan, Shiry Ginosar, Tinghui Zhou, and Alexei A Efros. Everybody dance now. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5933-5942, 2019. 2", + "[6] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 1, 2", + "[7] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbaek Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12479-12488, 2023. 2", + "[8] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 2", + "[9] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5712-5721, 2021. 2", + "[10] Xuan Gao, Chenglai Zhong, Jun Xiang, Yang Hong, Yudong Guo, and Juyong Zhang. Reconstructing personalized semantic facial nerf models from monocular video. ACM Transactions on Graphics (Proceedings of SIGGRAPH Asia), 41(6), 2022. 2", + "[11] Philip-William Grassal, Malte Prinzler, Titus Leistner, Carsten Rother, Matthias Nießner, and Justus Thies. Neural head avatars from monocular rgb videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18653–18664, 2022. 2", + "[12] Yang Hong, Bo Peng, Haiyao Xiao, Ligang Liu, and Juyong Zhang. Headnerf: A real-time nerf-based parametric head model. In Proceedings of the IEEE/CVF Conference" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Computer Vision and Pattern Recognition, pages 20374-20384, 2022. 2", + "[13] Mustafa Işik, Martin Rünz, Markos Georgopoulos, Taras Khakhulin, Jonathan Starck, Lourdes Agapito, and Matthias Nießner. Humanrf: High-fidelity neural radiance fields for humans in motion. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 1", + "[14] Rohit Jena, Ganesh Subramanian Iyer, Siddharth Choudhary, Brandon Smith, Pratik Chaudhari, and James Gee. Splatarmor: Articulated gaussian splattering for animatable humans from monocular rgb videos. arXiv preprint arXiv:2311.10812, 2023. 2", + "[15] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 42(4):1-14, 2023. 1, 2, 3, 4, 5", + "[16] Hyeongwoo Kim, Pablo Garrido, Ayush Tewari, Weipeng Xu, Justus Thies, Matthias Niessner, Patrick Pérez, Christian Richardt, Michael Zollhöfer, and Christian Theobalt. Deep video portraits. ACM transactions on graphics (TOG), 37(4): 1-14, 2018. 2", + "[17] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5", + "[18] Tobias Kirschstein, Shenhan Qian, Simon Giebenhain, Tim Walter, and Matthias Nießner. Nersemble: Multi-view radiance field reconstruction of human heads. ACM Trans. Graph., 42(4), 2023. 1, 5", + "[19] Muhammed Kocabas, Jen-Hao Rick Chang, James Gabriel, Oncel Tuzel, and Anurag Ranjan. Hugs: Human gaussian splats. arXiv preprint arXiv:2311.17910, 2023. 2", + "[20] Jiahui Lei, Yufu Wang, Georgios Pavlakos, Lingjie Liu, and Kostas Daniilidis. Gart: Gaussian articulated template models. arXiv preprint arXiv:2311.16099, 2023. 2", + "[21] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6):194:1-194:17, 2017. 2, 3, 4, 5", + "[22] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 1", + "[23] Zhe Li, Zerong Zheng, Lizhen Wang, and Yebin Liu. Animatable gaussians: Learning pose-dependent gaussian maps for high-fidelity human avatar modeling. arXiv, 2023. 2", + "[24] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. ACM transactions on graphics (TOG), 40(6):1-16, 2021. 2", + "[25] Stephen Lombardi, Tomas Simon, Gabriel Schwartz, Michael Zollhoefer, Yaser Sheikh, and Jason Saragih. Mixture of volumetric primitives for efficient neural rendering. ACM Transactions on Graphics (ToG), 40(4):1-13, 2021. 2" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20307", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. In *Seminal Graphics Papers: Pushing the Boundaries*, Volume 2, pages 851–866. 2023. 2", + "[27] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 2", + "[28] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2", + "[29] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 1, 2, 7", + "[30] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 1, 2", + "[31] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. 1, 2", + "[32] Pascal Paysan, Reinhard Knothe, Brian Amberg, Sami Romdhani, and Thomas Vetter. A 3d face model for pose and illumination invariant face recognition. In 2009 sixth IEEE international conference on advanced video and signal based surveillance, pages 296-301. IEEE, 2009. 2", + "[33] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Animatable neural radiance fields for modeling dynamic human bodies. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14314-14323, 2021. 2", + "[34] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 2", + "[35] Radu Alexandru Rosu, Shunsuke Saito, Ziyan Wang, Chenglei Wu, Sven Behnke, and Giljoo Nam. Neural strands: Learning hair geometry and appearance from multi-view images. In European Conference on Computer Vision, pages 73-89. Springer, 2022. 8", + "[36] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerfplayer: A streamable dynamic scene representation with decomposed neural radiance fields. IEEE Transactions on Visualization and Computer Graphics, 29(5):2732-2742, 2023. 2", + "[37] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 2" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Supasorn Suwajanakorn, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Synthesizing obama: learning lip sync from audio. ACM Transactions on Graphics (ToG), 36(4):1-13, 2017. 2", + "[39] Justus Thies, Michael Zollhofer, Marc Stamminger, Christian Theobalt, and Matthias Nießner. Face2face: Real-time face capture and reenactment of rgb videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2387-2395, 2016. 2, 3", + "[40] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: Image synthesis using neural textures. Acm Transactions on Graphics (TOG), 38(4):1-12, 2019. 2", + "[41] Ziyan Wang, Giljoo Nam, Tuur Stuyck, Stephen Lombardi, Chen Cao, Jason Saragih, Michael Zollhöfer, Jessica Hodgins, and Christoph Lassner. Neuwigs: A neural dynamic model for volumetric hair capture and animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8641-8651, 2023. 8", + "[42] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9421-9431, 2021. 2", + "[43] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5438–5448, 2022. 2", + "[44] Yuelang Xu, Lizhen Wang, Xiaochen Zhao, Hongwen Zhang, and Yebin Liu. Avatarmav: Fast 3d head avatar reconstruction using motion-aware neural voxels. In ACM SIGGRAPH 2023 Conference Proceedings, 2023. 2, 5, 7", + "[45] Yuelang Xu, Hongwen Zhang, Lizhen Wang, Xiaochen Zhao, Huang Han, Qi Guojun, and Yebin Liu. Latentavatar: Learning latent expression code for expressive neural head avatar. In ACM SIGGRAPH 2023 Conference Proceedings, 2023. 2", + "[46] Keyang Ye, Tianjia Shao, and Kun Zhou. Animatable 3d gaussians for high-fidelity synthesis of human motions. arXiv preprint arXiv:2311.13404, 2023. 2", + "[47] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4578-4587, 2021. 2", + "[48] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5", + "[49] Yufeng Zheng, Victoria Fernández Abrevaya, Marcel C Bühler, Xu Chen, Michael J Black, and Otmar Hilliges. Im avatar: Implicit morphable head avatars from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13545-13555, 2022. 2", + "[50] Yufeng Zheng, Wang Yifan, Gordon Wetzstein, Michael J Black, and Otmar Hilliges. Pointavatar: Deformable point" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20308", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "based head avatars from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21057-21067, 2023. 2, 5, 7", + "[51] Wojciech Zielonka, Timur Bagautdinov, Shunsuke Saito, Michael Zollhöfer, Justus Thies, and Javier Romero. Drivable 3d gaussian avatars. 2023. 2", + "[52] Wojciech Zielonka, Timo Bolkart, and Justus Thies. Instant volumetric head avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4574-4584, 2023. 2, 5, 7" + ], + "bbox": [ + 78, + 90, + 468, + 233 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "20309", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/GaussianAvatars_ Photorealistic Head Avatars with Rigged 3D Gaussians/41411f0e-3b8e-40c3-bbbd-d16b793f1be0_model.json b/2024/GaussianAvatars_ Photorealistic Head Avatars with Rigged 3D Gaussians/41411f0e-3b8e-40c3-bbbd-d16b793f1be0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d8697db77a6de4d47f0263dd402d44d692ee8f4b --- /dev/null +++ b/2024/GaussianAvatars_ Photorealistic Head Avatars with Rigged 3D Gaussians/41411f0e-3b8e-40c3-bbbd-d16b793f1be0_model.json @@ -0,0 +1,2191 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.116, + 0.131, + 0.856, + 0.154 + ], + "angle": 0, + "content": "GaussianAvatars: Photorealistic Head Avatars with Rigged 3D Gaussians" + }, + { + "type": "text", + "bbox": [ + 0.16, + 0.18, + 0.282, + 0.2 + ], + "angle": 0, + "content": "Shenhan Qian" + }, + { + "type": "text", + "bbox": [ + 0.323, + 0.181, + 0.478, + 0.198 + ], + "angle": 0, + "content": "Tobias Kirschstein" + }, + { + "type": "text", + "bbox": [ + 0.323, + 0.2, + 0.477, + 0.215 + ], + "angle": 0, + "content": "Simon Giebenhain" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.181, + 0.667, + 0.198 + ], + "angle": 0, + "content": "Liam Schoneveld²" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.2, + 0.665, + 0.216 + ], + "angle": 0, + "content": "Matthias Nießner1" + }, + { + "type": "text", + "bbox": [ + 0.708, + 0.181, + 0.833, + 0.198 + ], + "angle": 0, + "content": "Davide Davoli3" + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.223, + 0.389, + 0.242 + ], + "angle": 0, + "content": "1Technical University of Munich" + }, + { + "type": "text", + "bbox": [ + 0.433, + 0.223, + 0.583, + 0.242 + ], + "angle": 0, + "content": "\\(^{2}\\)Woven by Toyota" + }, + { + "type": "text", + "bbox": [ + 0.627, + 0.224, + 0.871, + 0.241 + ], + "angle": 0, + "content": "\\(^{3}\\)Toyota Motor Europe NV/SA" + }, + { + "type": "text", + "bbox": [ + 0.644, + 0.239, + 0.855, + 0.251 + ], + "angle": 0, + "content": "associated partner by contracted service" + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.283, + 0.221, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.283, + 0.359, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.282, + 0.484, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.282, + 0.617, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.291, + 0.751, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.754, + 0.291, + 0.892, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.417, + 0.895, + 0.46 + ], + "angle": 0, + "content": "Figure 1. We introduce GaussianAvatars, a new method to create photorealistic head avatars from multi-view videos. Our avatar is represented by 3D Gaussian splats rigged to a parametric face model. We can fully control and animate our avatars in terms of pose, expression, and view point as shown in the example renderings above." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.471, + 0.314, + 0.488 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.504, + 0.474, + 0.792 + ], + "angle": 0, + "content": "We introduce GaussianAvatars1, a new method to create photorealistic head avatars that are fully controllable in terms of expression, pose, and viewpoint. The core idea is a dynamic 3D representation based on 3D Gaussian splats that are rigged to a parametric morphable face model. This combination facilitates photorealistic rendering while allowing for precise animation control via the underlying parametric model, e.g., through expression transfer from a driving sequence or by manually changing the morphable model parameters. We parameterize each splat by a local coordinate frame of a triangle and optimize for explicit displacement offset to obtain a more accurate geometric representation. During avatar reconstruction, we jointly optimize for the morphable model parameters and Gaussian splat parameters in an end-to-end fashion. We demonstrate the animation capabilities of our photorealistic avatar in several challenging scenarios. For instance, we show reenactments from a driving video, where our method outperforms existing works by a significant margin." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.821, + 0.21, + 0.836 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.846, + 0.47, + 0.877 + ], + "angle": 0, + "content": "Creating animatable avatars of human heads has been a longstanding problem in computer vision and graphics." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.473, + 0.892, + 0.564 + ], + "angle": 0, + "content": "In particular, the ability to render photorealistic dynamic avatars from arbitrary viewpoints enables numerous applications in gaming, movie production, immersive telepresence, and augmented or virtual reality. For such applications, it is also crucial to be able to control the avatar, and for it to generalize well to novel poses and expressions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.566, + 0.893, + 0.717 + ], + "angle": 0, + "content": "Reconstructing a 3D representation able to jointly capture the appearance, geometry and dynamics of human heads represents a major challenge for high-fidelity avatar generation. The under-constrained nature of this reconstruction problem significantly complicates the task of achieving a representation that combines novel-view rendering photorealism with expression controllability. Moreover, extreme expressions and facial details, like wrinkles, the mouth interior, and hair, are difficult to capture, and can produce visual artifacts easily noticed by humans." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.718, + 0.893, + 0.839 + ], + "angle": 0, + "content": "Neural Radiance Fields (NeRF [28]) and its variants [2, 6, 29] have shown impressive results in reconstructing static scenes from multi-view observations. Follow-up works have extended NeRF to model dynamic scenes for both arbitrary [22, 30, 31] and human-tailored scenarios [13, 18]. These works achieve impressive results for novel view rendering; however, they lack controllability and as such do not generalize well to novel poses and expressions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "The recent 3D Gaussian Splitting method [15] achieves even higher rendering quality than NeRF for novel-view synthesis with real-time performance, optimizing for discrete geometric primitives (3D Gaussians) throughout 3D" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.419, + 0.901 + ], + "angle": 0, + "content": "1Project page: https://shenhanqian.github.io/gaussian-avatars" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20299" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "space. This method has been extended to capture dynamic scenes by building explicit correspondences across time steps [27]; however, they do not allow for animations of the reconstructed outputs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.153, + 0.471, + 0.41 + ], + "angle": 0, + "content": "To this end, we propose GaussianAvatars, a method for dynamic 3D representation of human head based on 3D Gaussian splats that are rigged to a parametric morphable face model. Given a FLAME [21] mesh, we initialize a 3D Gaussian at the center of each triangle. When the FLAME mesh is animated, each Gaussian then translates, rotates, and scales according to its parent triangle. The 3D Gaussians then form a radiance field on top of the mesh, compensating for regions where the mesh is either not accurately aligned, or is incapable of reproducing certain visual elements. To achieve high fidelity of the reconstructed avatar, we introduce a binding inheritance strategy to support Gaussian splats without losing controllability. We also explore balancing fidelity and robustness to animate the avatars with novel expressions and poses. Our method outperforms existing works by a significant margin on both novel-view rendering and reenactment from a driving video." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.413, + 0.297, + 0.427 + ], + "angle": 0, + "content": "Our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.43, + 0.468, + 0.473 + ], + "angle": 0, + "content": "- We propose GaussianAvatars, a method to create animatable head avatars by rigging 3D Gaussians to parametric mesh models." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.475, + 0.468, + 0.519 + ], + "angle": 0, + "content": "- We design a binding inheritance strategy to support adding and removing 3D Gaussians without losing controllability." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.43, + 0.468, + 0.519 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.535, + 0.219, + 0.55 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.56, + 0.35, + 0.574 + ], + "angle": 0, + "content": "2.1. Radiance Field Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.468, + 0.825 + ], + "angle": 0, + "content": "NeRF [28] stores the radiance field of a scene in a neural network and provides photorealistic renderings of novel views with volumetric rendering. Later work [37, 47] represents the scene as voxel grids, achieving comparable rendering quality in a shorter time. Efficiency can be further improved by employing compression methods such as voxel hashing [29] or tensor decomposition [6]. PointNeRF [43] uses point clouds as a scene representation, whereas 3D Gaussian Splatting [15] uses anisotropic 3D Gaussians that are rendered by sorting and rasterization, achieving superior visual quality with real-time rendering. Mixture of Volumetric Primitives [25] uses surface-aligned volumes to achieve fast rendering with high visual fidelity. It also applies residual transformations to primitives. Our method follows 3D Gaussian Splatting [15], benefiting from the expressiveness of anisotropic Gaussians." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.468, + 0.901 + ], + "angle": 0, + "content": "A common paradigm to model a dynamic scene is directly storing it in an MLP with 4D coordinates as the input [9, 42] or a 4D tensor with the time dimension or spatial dimensions compressed [1, 4, 7, 36]. These methods can faithfully replay a dynamic scene and produce realistic" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.318 + ], + "angle": 0, + "content": "novel-view rendering but lack an explicit handle to manipulate the content. Another paradigm learns a static canonical space and the maps time steps back to the canonical space via separate MLPs [30, 31, 34]. Instead of using a deformation MLP, a proxy geometry provides more direct controllability. Liu et al. [24] warp points from an observed space to the canonical space based on the movement of the nearest triangle on an SMPL [26] mesh. Peng et al. [33] deform points with the skeleton of SMPL and neural blending weights. Concurrent works create human body avatars with forward deformation [14, 19, 20, 23, 46] and cage-based deformation [51]. Unlike these methods, we directly attach 3D Gaussians to triangles and explicitly move them, obviating the need for a canonical space and enabling effective mesh finetuning." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.327, + 0.88, + 0.342 + ], + "angle": 0, + "content": "2.2. Human Head Reconstruction and Animation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.35, + 0.892, + 0.591 + ], + "angle": 0, + "content": "Head avatar creation advanced through the uptake of differentiable rendering and scene representations. Thies et al. [39] instigated a shift toward digital avatars with real-time face tracking and authentic face reenactment. Advancements in image synthesis with neural networks [5, 40] boosted the controllability of head avatars from lip-syncing to expression transfer and head motions [16, 38, 45]. Gafni et al. [8] learn a NeRF conditioned on an expression vector from monocular videos. Grassal et al. [11] subdivide and add offsets to FLAME [21] to enhance its geometry and enable a dynamic texture via an expression-dependent texture field. IMavatar [49] learns a 3D morphable head avatar with neural implicit functions, solving for a map from observed to canonical space via iterative root-finding. HeadNeRF [12] learns a NeRF-based parametric head model with 2D neural rendering for efficiency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.592, + 0.892, + 0.833 + ], + "angle": 0, + "content": "INSTA [52] deforms query points to a canonical space by finding the nearest triangle on a FLAME [21] mesh, and combines this with InstantNGP [29] to achieve fast rendering. Like INSTA, we also use triangles to warp the scene, but we build a consistent correspondence between a 3D Gaussian and a triangle instead of querying the nearest one for each timestep. Zheng et al. [50] explore point-based representations with differential point splatting. They define a point set in canonical space and learn a deformation field conditioned on FLAME's expression vectors to animate the head avatar. While the scale of a point has to be manually set for their method, it is an estimizable parameter for 3D Gaussians. NeRFBlendShape [10] models a dynamic scene by blending hash tables with 3DMM parameters [3, 32]. AvatarMAV [44] decouples motion and appearance, blending voxel grids only for the motion field." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.845, + 0.591, + 0.86 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "As shown in Fig. 2, the input to our method is a multi-view video recording of a human head. For each time step, we" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20300" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.086, + 0.895, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.44, + 0.893, + 0.51 + ], + "angle": 0, + "content": "Figure 2. Overview. Our method binds 3D Gaussian splats to a FLAME [21] mesh locally. We take the tracked mesh for each frame and transform the splats from local to global space before rendering them with 3D Gaussian Splatting [15]. We optimize the splats in the local space by minimizing color loss on the rendering. We add and remove splats adaptively with their binding relation to triangles inherited so that all splats remain rigged throughout the optimization procedure. Further, we regularize the position and scaling of 3D Gaussian splats to suppress artifacts during animation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.536, + 0.473, + 0.822 + ], + "angle": 0, + "content": "use a photometric head tracker based on [39] to fit FLAME [21] parameters with multi-view observations and known camera parameters. The FLAME meshes have vertices at varied positions but share the same topology. Therefore, we can build a consistent connection between triangles of the mesh and 3D Gaussian splats (Sec. 3.2). The splats are rendered into images via a differentiable tile rasterizer [15]. These images are then supervised by the groundtruth images towards learning a photorealistic human head avatar. As per static scenes, it is also necessary to densify and prune Gaussian splats for optimal quality, with a set of adaptive density control operations [15]. To achieve this without breaking the connection between triangles and splats, we design a binding inheritance strategy (Sec. 3.3) so that new Gaussian points remain rigged to the FLAME mesh. In addition to the color loss, we also find it crucial to regularize the local position and scaling of Gaussian splats to avoid quality degradation under novel expressions and poses (Sec. 3.4)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.208, + 0.849 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "3D Gaussian Splitting [15] provides a solution to reconstruct a static scene with anisotropic 3D Gaussians given images and camera parameters. A scene is represented by a" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.536, + 0.892, + 0.567 + ], + "angle": 0, + "content": "set of Gaussian splats, each defined by a covariance matrix \\(\\Sigma\\) centered at point (mean) \\(\\mu\\):" + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.573, + 0.892, + 0.593 + ], + "angle": 0, + "content": "\\[\nG (\\boldsymbol {x}) = e ^ {- \\frac {1}{2} (\\boldsymbol {x} - \\boldsymbol {\\mu}) ^ {T} \\Sigma^ {- 1} (\\boldsymbol {x} - \\boldsymbol {\\mu})} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.601, + 0.892, + 0.691 + ], + "angle": 0, + "content": "Note that covariance matrices have physical meaning only when they are semi-definite, which cannot be guaranteed for an optimization process with gradient descent. Therefore, Kerbl et al. [15] first define a parametric ellipse with a scaling matrix \\( S \\) and a rotation matrix \\( R \\), then construct the covariance matrix by:" + }, + { + "type": "equation", + "bbox": [ + 0.64, + 0.698, + 0.891, + 0.715 + ], + "angle": 0, + "content": "\\[\n\\Sigma = R S S ^ {T} R ^ {T}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.723, + 0.892, + 0.784 + ], + "angle": 0, + "content": "Practically, an ellipse is stored as a position vector \\(\\pmb{\\mu} \\in \\mathbb{R}^3\\), a scaling vector \\(\\pmb{s} \\in \\mathbb{R}^3\\), and a quaternion \\(\\pmb{q} \\in \\mathbb{R}^4\\). In our paper, we use \\(\\pmb{r} \\in \\mathbb{R}^{3 \\times 3}\\) to denote the corresponding rotation matrix to \\(\\pmb{q}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.785, + 0.892, + 0.815 + ], + "angle": 0, + "content": "For rendering, the color \\( C \\) of a pixel is computed by blending all 3D Gaussians overlapping the pixel:" + }, + { + "type": "equation", + "bbox": [ + 0.603, + 0.822, + 0.892, + 0.864 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {C} = \\sum_ {i = 1} \\boldsymbol {c} _ {i} \\alpha_ {i} ^ {\\prime} \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\alpha_ {j} ^ {\\prime}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\( c_{i} \\) is the color of each point modeled by 3-degree spherical harmonics. The blending weight \\( \\alpha^{\\prime} \\) is given by" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20301" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.138 + ], + "angle": 0, + "content": "evaluating the 2D projection of the 3D Gaussian multiplied by a per-point opacity \\(\\alpha\\). The Gaussian splats are sorted by depth before blending to respect visibility order." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.147, + 0.276, + 0.163 + ], + "angle": 0, + "content": "3.2. 3D Gaussian Rigging" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.17, + 0.469, + 0.397 + ], + "angle": 0, + "content": "The key component of our method is how we build connections between the FLAME [21] mesh and 3D Gaussian splats. Initially, we pair each triangle of the mesh with a 3D Gaussian and let the 3D Gaussian move with the triangle across time steps. In other words, the 3D Gaussian is static in the local space of its parent triangle but dynamic in the global (metric) space as the triangle moves. Given the vertices and edges of a triangle, we take the mean position \\( T \\) of the vertices as the origin of the local space. Then, we concatenate the direction vector of one of the edges, the normal vector of the triangle, and their cross product as column vectors to form a rotation matrix \\( R \\), which describes the orientation of the triangle in the global space. We also compute a scalar \\( k \\) by the mean length of one of the edges and its perpendicular to describe the triangle scaling." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.397, + 0.469, + 0.488 + ], + "angle": 0, + "content": "For the paired 3D Gaussian of a triangle, we define its location \\(\\mu\\), rotation \\(r\\), and anisotropic scaling \\(s\\) all in the local space. We initialize the location \\(\\mu\\) at the local origin, the rotation \\(r\\) as an identity rotation matrix, and the scaling \\(s\\) as a unit vector. At rendering time, we convert these properties into the global space by:" + }, + { + "type": "equation", + "bbox": [ + 0.217, + 0.5, + 0.469, + 0.515 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {r} ^ {\\prime} = \\boldsymbol {R} \\boldsymbol {r}, \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.217, + 0.518, + 0.469, + 0.535 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\mu} ^ {\\prime} = k \\boldsymbol {R} \\boldsymbol {\\mu} + \\boldsymbol {T}, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.22, + 0.538, + 0.469, + 0.552 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {s} ^ {\\prime} = k \\boldsymbol {s}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.566, + 0.469, + 0.702 + ], + "angle": 0, + "content": "We incorporate triangle scaling in Eqs. (5) and (6) so that the local position and scaling of a 3D Gaussian are defined relative to the absolute scale of a triangle. This enables an adaptive step size in the metric space with a constant learning rate for parameters defined in the local space. For example, a 3D Gaussian paired with a smaller triangle will move slower in an iteration step than those paired with large triangles. This also makes interpreting the parameters regarding the distance from the triangle center easier." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.712, + 0.268, + 0.727 + ], + "angle": 0, + "content": "3.3. Binding Inheritance" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.469, + 0.841 + ], + "angle": 0, + "content": "Only having the same numbers of Gaussian splats as the triangles is insufficient to capture details. For instance, representing a curved hair strand requires multiple splats, while a triangle on the scalp may intersect with several strands. Therefore, we also need the adaptive density control strategy [15], which adds and removes splats based on the view-space positional gradient and the opacity of each Gaussian." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.841, + 0.469, + 0.901 + ], + "angle": 0, + "content": "For each 3D Gaussian with a large view-space positional gradient, we split it into two smaller ones if it is large or clone it if it is small. We conduct this in the local space and ensure a newly created Gaussian is close to the old one that" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.182 + ], + "angle": 0, + "content": "triggers this densification operation. Then, it is reasonable to bind a new 3D Gaussian to the same triangle as the old one because it was created to enhance the fidelity of the local region. Therefore, each 3D Gaussian must carry one more parameter, the index of its parent triangle, to enable binding inheritance during densification." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.183, + 0.892, + 0.364 + ], + "angle": 0, + "content": "Besides densification, we also use the pruning operation as a part of the adaptive density control strategy [15]. It periodically resets the opacity of all splats close to zero and removes points with opacity below a threshold. This technique is effective in suppressing floating artifacts, however, such pruning can also cause problems in a dynamic scene. For instance, regions of the face that are often occluded (such as eyeball triangles), can be overly sensitive to this pruning strategy, and often end up with few or no attached Gaussians. To prevent this, we keep track of the number of splats attached to each triangle, and ensure that every triangle always has at least one splat attached." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.374, + 0.791, + 0.389 + ], + "angle": 0, + "content": "3.4. Optimization and Regularization" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.397, + 0.89, + 0.427 + ], + "angle": 0, + "content": "We supervise the rendered images with a combination of \\(\\mathcal{L}_1\\) term and a D-SSIM term following [15]:" + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.44, + 0.891, + 0.457 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {r g b}} = (1 - \\lambda) \\mathcal {L} _ {1} + \\lambda \\mathcal {L} _ {\\mathrm {D} - \\mathrm {S S I M}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.468, + 0.892, + 0.588 + ], + "angle": 0, + "content": "with \\(\\lambda = 0.2\\). This already results in good re-rendering quality without additional supervision, such as depth or silhouette supervision, thanks to the powerful tile rasterizer [15]. We found however, that if we try to animate these splats via FLAME [21] to novel expressions and poses, large spike- and blob-like artifacts appear wildly throughout the scene. This is due to a poor alignment between the Gaussian splats and the triangles." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.589, + 0.892, + 0.74 + ], + "angle": 0, + "content": "Position loss with threshold. A basic hypothesis behind 3D Gaussian rigging is that the Gaussian splats should roughly match the underlying mesh. They should also match their locations; for instance a Gaussian representing a spot on the nose should not be rigged to a triangle on the cheek. Although our splats are initialized at triangle centers, and new splats are added nearby these existing ones, it is not guaranteed that the primitives remain close to their parent triangle after the optimization. To address this, we regularize the local position of each Gaussian by:" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.752, + 0.891, + 0.769 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {p o s i t i o n}} = \\| \\max (\\boldsymbol {\\mu}, \\epsilon_ {\\text {p o s i t i o n}}) \\| _ {2}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.89, + 0.81 + ], + "angle": 0, + "content": "where \\(\\epsilon_{\\mathrm{position}} = 1\\) is a threshold that tolerates small errors within the scaling of its parent triangle." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Scaling loss with threshold. Aside from position, the scaling of 3D Gaussians is even more essential for the visual quality during animation. Specifically, if a 3D Gaussian is large in comparison to its parent triangle, small rotations of the triangle – barely noticeable at the scale of the triangle – will be magnified by the scale of the 3D Gaussian, resulting" + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20302" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "in unpleasant jittering artifacts. To mitigate this, we also regularize the local scale of each 3D Gaussian by:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.132, + 0.469, + 0.149 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s c a l i n g}} = \\left\\| \\max \\left(\\boldsymbol {s}, \\epsilon_ {\\text {s c a l i n g}}\\right) \\right\\| _ {2}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.157, + 0.47, + 0.246 + ], + "angle": 0, + "content": "where \\(\\epsilon_{\\mathrm{scaling}} = 0.6\\) is a threshold that disables this loss term when the scale of a Gaussian less than \\(0.6\\times\\) the scale of its parent triangle. This \\(\\epsilon_{\\mathrm{scaling}}\\)-tolerance is indispensable; without it, the Gaussian splats shrink excessively, causing rendering speeds to deteriorate, as camera rays need to hit more splats before zero transmittance is reached." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.248, + 0.299, + 0.261 + ], + "angle": 0, + "content": "Our final loss function is thus:" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.272, + 0.469, + 0.289 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {r g b}} + \\lambda_ {\\text {p o s i t i o n}} \\mathcal {L} _ {\\text {p o s i t i o n}} + \\lambda_ {\\text {s c a l i n g}} \\mathcal {L} _ {\\text {s c a l i n g}}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.297, + 0.469, + 0.371 + ], + "angle": 0, + "content": "where \\(\\lambda_{\\mathrm{position}} = 0.01\\) and \\(\\lambda_{\\mathrm{scaling}} = 1\\). Note that we only apply \\(\\mathcal{L}_{\\mathrm{position}}\\) and \\(\\mathcal{L}_{\\mathrm{scaling}}\\) to visible splats. Thereby, we only regularize points when the color loss \\(\\mathcal{L}_{\\mathrm{rgb}}\\) is present. This helps to maintain the learned structure of often-occluded regions such as teeth and eyeballs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.373, + 0.47, + 0.645 + ], + "angle": 0, + "content": "Implementation details. We use Adam [17] for parameter optimization (the same hyperparameter values are used across all subjects). We set the learning rate to 5e-3 for the position and 1.7e-2 for the scaling of 3D Gaussians and keep the same learning rates as 3D Gaussian Splatting [15] for the rest of the parameters. Alongside the Gaussian splat parameters, we also finetune the translation, joint rotation, and expression parameters of FLAME [21] for each timestep, using learning rates 1e-6, 1e-5, and 1e-3, respectively. We train for 600,000 iterations, and exponentially decay the learning rate for the splat positions until the final iteration, where it reaches \\(0.01 \\times\\) the initial value. We enable adaptive density control with binding inheritance every 2,000 iterations, from iteration 10,000 until the end. Every 60,000 iterations, we reset the Gaussians' opacities. We use a photo-metric head tracker to obtain the FLAME parameters, including shape \\(\\beta\\), translation \\(t\\), pose \\(\\theta\\), expression \\(\\psi\\), and vertex offset \\(\\Delta v\\) in the canonical space." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.657, + 0.209, + 0.673 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.682, + 0.159, + 0.697 + ], + "angle": 0, + "content": "4.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.469, + 0.825 + ], + "angle": 0, + "content": "Settings. We evaluate the quality of head avatars with three settings: 1) novel-view synthesis: driving an avatar with head poses and expressions from training sequences and rendering from a held-out viewpoint. 2) self-reenactment: driving an avatar with unseen poses and expressions from a held-out sequence of the same subject and rendering all 16 camera views. 3) cross-identity reenactment: animating an avatar with poses and expressions from another subject." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Dataset. We conduct experiments on video recordings of 9 subjects from the NeRSample [18] dataset. All recordings contain 16 views covering the front and sides of a subject. For each subject, we take 11 video sequences and down-sample images to a resolution of \\(802 \\times 550\\). Participants" + }, + { + "type": "table", + "bbox": [ + 0.5, + 0.09, + 0.892, + 0.194 + ], + "angle": 0, + "content": "
Novel-View SynthesisSelf-Reenactment
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
AvatarMAV [44]29.50.9130.15224.30.8870.168
PointAvatar [50]25.80.8930.09723.40.8840.102
INSTA [52]26.70.8990.12226.30.9060.110
Ours31.60.9380.06526.00.9100.076
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.205, + 0.89, + 0.235 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison with state-of-the-art methods. Green indicates the best and yellow indicates the second." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.264, + 0.892, + 0.4 + ], + "angle": 0, + "content": "were instructed to perform specific expressions or emotions in 10 sequences and freely perform in the last one. To conduct quantitative evaluation for each subject, we train our method and three baselines (INSTA [52], PointAvatar [50], and AvatarMAV [44]) using 9 out of the 10 prescribed sequences and 15 out of 16 available cameras. We use the \\(11^{\\mathrm{th}}\\) (free-performance) sequence to visually assess cross-identity reenactment capabilities. Please refer to the supplementary material for details of the dataset and baselines." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.41, + 0.875, + 0.424 + ], + "angle": 0, + "content": "4.2. Head Avatar Reconstruction and Animation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.432, + 0.892, + 0.538 + ], + "angle": 0, + "content": "We evaluate the reconstruction quality of avatars by novel-view synthesis and animation fidelity by self-reenactment. Fig. 3 shows qualitative comparisons. For novel-view synthesis, all methods produce plausible rendering results. A close inspection of PointAvatar's [50] results show dotted artifacts, owing its fixed point size. In our case, the anisotropic scaling of 3D Gaussians alleviates this issue." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.892, + 0.704 + ], + "angle": 0, + "content": "INSTA [52] shows clean results for the face, however regions around the neck and shoulder can be noisy, as the tracked FLAME meshes are often misaligned in those regions. This causes issues for INSTA, as its warping process is based on the nearest triangle. In the case of our method, each Gaussian splat is rigged to a consistent triangle, regardless of the pose or expression. When the tracked mesh is inaccurate, the positional gradient to a Gaussian splat can consistently back-propagate to the same triangle. This enables misalignments due to incorrect FLAME tracking to be corrected while the 3D Gaussians are being optimized." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.893, + 0.9 + ], + "angle": 0, + "content": "AvatarMAV [44] shows comparable quality to other methods in novel-view synthesis but struggles with novel-expression synthesis. This is because it only uses the expression vector of a 3DMM as conditioning. Since the control from the expression vectors to the deformation bases must be learned, it struggles to reproduce expressions that are far from the training distribution. Similar conclusions can be drawn from the quantitative comparison shown in Tab. 1. Our approach outperforms others by a large margin regarding metrics for novel-view synthesis. Our method also stands out in self-reenactment, with significantly lower perceptual differences in terms of LPIPS [48]. Note that self-reenactment is based on tracked FLAME meshes that" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20303" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.895, + 0.803 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.81, + 0.895, + 0.855 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparison on novel-view synthesis and self-reenactment of head avatars. Our method outperforms state-of-the-art methods by producing significantly sharper rendering outputs. We obtain precise reconstruction of details such as reflective light on eyes, hair strands, teeth, etc. Our results for self-reenactment show more accurate expressions compare to baselines." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "20304" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.093, + 0.891, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.559, + 0.895, + 0.603 + ], + "angle": 0, + "content": "Figure 4. Cross-identity reenactment of head avatars. We use the tracked FLAME expression and pose parameters of source actors to drive the reconstructed avatars. Our method produces high-quality rendering and transfers expressions vividly, while baseline methods suffer from artifacts and generalize poorly to novel expressions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.628, + 0.47, + 0.672 + ], + "angle": 0, + "content": "may not perfectly align with the target images, thus bringing disadvantages to our results with more visual details regarding pixel-wise metrics such as PSNR." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.473, + 0.84 + ], + "angle": 0, + "content": "For a real-world test of avatar animation, we conduct experiments on cross-identity reenactment in Fig. 4. Our avatars accurately reproduce eye blinks and mouth movements from source actors showing lively, complex dynamics such as wrinkles. INSTA [52] suffers from aliasing artifacts when the avatars move beyond the occupancy grid of I-NGP [29] optimized for training sequences. The movement of results from PointAvatar [50] is not precise because its deformation space is not guaranteed to be consistent with FLAME. AvatarMAV [44] exhibits large degradations in reenactment due to a lack of deformation priors." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.231, + 0.864 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "To validate the effectiveness of our method components, we deactivate each of them and report results in Tab. 2." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.626, + 0.62, + 0.765 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.626, + 0.756, + 0.766 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.772, + 0.626, + 0.892, + 0.766 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.776, + 0.894, + 0.819 + ], + "angle": 0, + "content": "Figure 5. Fine-tuning FLAME parameters leads to better alignment of the mesh to the input image. In the example above, the movements of cheeks and lips are better captured after fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Adaptive density control with binding inheritance. Without binding inheritance, we lose the ability to add and remove Gaussian splats besides the initial ones. With a lim" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "20305" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.078, + 0.089, + 0.462, + 0.235 + ], + "angle": 0, + "content": "
Novel-ViewSelf-Reenactment
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Ours28.80.8830.09825.10.8530.101
w/o ADC26.80.8540.20625.10.8600.183
w/o Lscaling28.00.8770.11424.90.8520.109
w/o εscaling25.00.8330.19524.10.8430.176
w/o Lposition29.70.8940.09124.90.8510.096
w/o εposition28.70.8820.10525.00.8550.106
w/o FLAME ft.26.10.8550.13125.50.8620.124
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.248, + 0.47, + 0.31 + ], + "angle": 0, + "content": "Table 2. Ablation study on subject #304. Green indicates the best and yellow indicates the second. “ADC” refers to Adaptive Density Control with binding inheritance. ‘FLAME ft.’ refers to FLAME parameter fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.334, + 0.469, + 0.379 + ], + "angle": 0, + "content": "ited number of splats, the scaling of each splat increases to occupy the same space, causing blurry renderings, leading to huge fidelity loss, as shown in the second row of Tab. 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.38, + 0.469, + 0.5 + ], + "angle": 0, + "content": "Regularization on the scaling of Gaussian splats. Without the scaling loss, artifacts show up as spikes (the middle column in Fig. 6), and slightly worsen the image quality (the third row in Tab. 2). But when we use the scaling loss without an error tolerance, all metrics drastically deteriorate (the fourth row in Tab. 2) because all the splats are regularized to be infinitely small, making it hard to construct opaque surfaces." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.501, + 0.469, + 0.711 + ], + "angle": 0, + "content": "Regularization on the position of Gaussian splats. When disabling the threshold for the position loss, we get slightly worse metrics (the sixth row in Tab. 2). But when turning off the position loss, the metrics on novel-view synthesis become the best in the table (the fifth row in Tab. 2). This is reasonable because without constraints to the mesh Gaussian splats can freely move in the space to achieve minimal re-rendering error. However, since the distribution of Gaussian splats are overfitted to training frames, the avatar shows artifacts such as cracks and fly-around blobs (the left column in Fig. 6) in novel expressions and poses. Therefore, the position loss is necessary to enforce a conservative deviation of splats from the mesh surface so that we can animate the reconstructed avatar with unseen facial motion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.712, + 0.469, + 0.818 + ], + "angle": 0, + "content": "FLAME fine-tuning. Thanks to our consistent binding between Gaussian splats and triangles, we can effectively fine-tune FLAME parameters while optimizing Gaussian splats (Fig. 5). As shown in the last row in Tab. 2, turning off the fine-tuning negatively affects image quality for novel-view synthesis. For self-reenactment, fine-tuning FLAME parameters also leads to lower perceptual error." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.83, + 0.466, + 0.848 + ], + "angle": 0, + "content": "5. Limitations and Potential Negative Impacts" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Our approach, based on 3D Gaussian splats, directly captures the radiance field without decoupling material and lighting. Hence, relighting the avatar is not feasible within" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.891, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.256, + 0.892, + 0.285 + ], + "angle": 0, + "content": "Figure 6. The position loss and the scaling loss helps prevent artifacts during animation with novel expressions and poses." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.31, + 0.892, + 0.401 + ], + "angle": 0, + "content": "our current approach. Additionally, we lack control over areas that are not modeled by FLAME such as hair and other accessories. Here, we believe that modeling these parts directly would provide an interesting future research avenue; for instance, recent methods show promising result in reconstructing hair [35, 41]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.402, + 0.893, + 0.567 + ], + "angle": 0, + "content": "The creation of photorealistic avatars may introduce a wide spectrum of malevolent uses. Privacy violations are a significant concern, with potential for unauthorized manipulation of people's likeness. Moreover the creation of deceptive content, like deepfake videos, can lead to misinformation, defamation and harm to reputations. Identity theft is an additional risk, as avatars may be exploited for impersonation and fraudulent activities. We strongly condemn unauthorized and malevolent uses of this technology, emphasizing the need for ethical considerations in all applications utilizing our method." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.581, + 0.619, + 0.595 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.606, + 0.892, + 0.788 + ], + "angle": 0, + "content": "GaussianAvatars is a novel approach that creates photorealistic avatars from video sequences. It features a dynamic 3D representation based on 3D Gaussian splats that are rigged to a parametric morphable face model. This enables flexible control and precise attribute transfer of a dynamic, photorealistic avatar. The set of Gaussian splats can deviate from the mesh surface to compensate the absence or inaccuracy of the morphable model, exhibiting astonishing ability to model fine details on human heads. Our approach outperforms state-of-the-art methods on image quality and expression accuracy for a large margin, indicating a large potential for more applications in related domains." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.801, + 0.667, + 0.817 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This work was supported by Toyota Motor Europe and Woven by Toyota. This work was also supported by the ERC Starting Grant Scan2CAD (804724). We thank Justus Thies, Andrei Burov, and Lei Li for constructive discussions and Angela Dai for the video voice-over." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20306" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.198 + ], + "angle": 0, + "content": "[1] Benjamin Attal, Jia-Bin Huang, Christian Richardt, Michael Zollhoefer, Johannes Kopf, Matthew O'Toole, and Changil Kim. Hyperreel: High-fidelity 6-dof video with ray-conditioned sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16610-16620, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.201, + 0.472, + 0.284 + ], + "angle": 0, + "content": "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.286, + 0.472, + 0.342 + ], + "angle": 0, + "content": "[3] Volker Blanz and Thomas Vetter. Face recognition based on fitting a 3d morphable model. IEEE Transactions on pattern analysis and machine intelligence, 25(9):1063-1074, 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.344, + 0.472, + 0.399 + ], + "angle": 0, + "content": "[4] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 130-141, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.402, + 0.472, + 0.457 + ], + "angle": 0, + "content": "[5] Caroline Chan, Shiry Ginosar, Tinghui Zhou, and Alexei A Efros. Everybody dance now. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5933-5942, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.459, + 0.472, + 0.514 + ], + "angle": 0, + "content": "[6] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.517, + 0.472, + 0.586 + ], + "angle": 0, + "content": "[7] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbaek Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12479-12488, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.588, + 0.472, + 0.657 + ], + "angle": 0, + "content": "[8] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.66, + 0.472, + 0.715 + ], + "angle": 0, + "content": "[9] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5712-5721, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.717, + 0.472, + 0.786 + ], + "angle": 0, + "content": "[10] Xuan Gao, Chenglai Zhong, Jun Xiang, Yang Hong, Yudong Guo, and Juyong Zhang. Reconstructing personalized semantic facial nerf models from monocular video. ACM Transactions on Graphics (Proceedings of SIGGRAPH Asia), 41(6), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.788, + 0.472, + 0.857 + ], + "angle": 0, + "content": "[11] Philip-William Grassal, Malte Prinzler, Titus Leistner, Carsten Rother, Matthias Nießner, and Justus Thies. Neural head avatars from monocular rgb videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18653–18664, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.859, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[12] Yang Hong, Bo Peng, Haiyao Xiao, Ligang Liu, and Juyong Zhang. Headnerf: A real-time nerf-based parametric head model. In Proceedings of the IEEE/CVF Conference" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "on Computer Vision and Pattern Recognition, pages 20374-20384, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.19 + ], + "angle": 0, + "content": "[13] Mustafa Işik, Martin Rünz, Markos Georgopoulos, Taras Khakhulin, Jonathan Starck, Lourdes Agapito, and Matthias Nießner. Humanrf: High-fidelity neural radiance fields for humans in motion. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.894, + 0.261 + ], + "angle": 0, + "content": "[14] Rohit Jena, Ganesh Subramanian Iyer, Siddharth Choudhary, Brandon Smith, Pratik Chaudhari, and James Gee. Splatarmor: Articulated gaussian splattering for animatable humans from monocular rgb videos. arXiv preprint arXiv:2311.10812, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.264, + 0.894, + 0.319 + ], + "angle": 0, + "content": "[15] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 42(4):1-14, 2023. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.321, + 0.894, + 0.389 + ], + "angle": 0, + "content": "[16] Hyeongwoo Kim, Pablo Garrido, Ayush Tewari, Weipeng Xu, Justus Thies, Matthias Niessner, Patrick Pérez, Christian Richardt, Michael Zollhöfer, and Christian Theobalt. Deep video portraits. ACM transactions on graphics (TOG), 37(4): 1-14, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.894, + 0.432 + ], + "angle": 0, + "content": "[17] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.894, + 0.489 + ], + "angle": 0, + "content": "[18] Tobias Kirschstein, Shenhan Qian, Simon Giebenhain, Tim Walter, and Matthias Nießner. Nersemble: Multi-view radiance field reconstruction of human heads. ACM Trans. Graph., 42(4), 2023. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.491, + 0.894, + 0.533 + ], + "angle": 0, + "content": "[19] Muhammed Kocabas, Jen-Hao Rick Chang, James Gabriel, Oncel Tuzel, and Anurag Ranjan. Hugs: Human gaussian splats. arXiv preprint arXiv:2311.17910, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.535, + 0.894, + 0.575 + ], + "angle": 0, + "content": "[20] Jiahui Lei, Yufu Wang, Georgios Pavlakos, Lingjie Liu, and Kostas Daniilidis. Gart: Gaussian articulated template models. arXiv preprint arXiv:2311.16099, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.577, + 0.894, + 0.645 + ], + "angle": 0, + "content": "[21] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6):194:1-194:17, 2017. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.648, + 0.894, + 0.731 + ], + "angle": 0, + "content": "[22] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.733, + 0.894, + 0.774 + ], + "angle": 0, + "content": "[23] Zhe Li, Zerong Zheng, Lizhen Wang, and Yebin Liu. Animatable gaussians: Learning pose-dependent gaussian maps for high-fidelity human avatar modeling. arXiv, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.776, + 0.894, + 0.843 + ], + "angle": 0, + "content": "[24] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. ACM transactions on graphics (TOG), 40(6):1-16, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.894, + 0.902 + ], + "angle": 0, + "content": "[25] Stephen Lombardi, Tomas Simon, Gabriel Schwartz, Michael Zollhoefer, Yaser Sheikh, and Jason Saragih. Mixture of volumetric primitives for efficient neural rendering. ACM Transactions on Graphics (ToG), 40(4):1-13, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20307" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[26] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. In *Seminal Graphics Papers: Pushing the Boundaries*, Volume 2, pages 851–866. 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.203 + ], + "angle": 0, + "content": "[27] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.205, + 0.469, + 0.272 + ], + "angle": 0, + "content": "[28] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.274, + 0.469, + 0.328 + ], + "angle": 0, + "content": "[29] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.331, + 0.469, + 0.399 + ], + "angle": 0, + "content": "[30] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.401, + 0.469, + 0.468 + ], + "angle": 0, + "content": "[31] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.47, + 0.469, + 0.538 + ], + "angle": 0, + "content": "[32] Pascal Paysan, Reinhard Knothe, Brian Amberg, Sami Romdhani, and Thomas Vetter. A 3d face model for pose and illumination invariant face recognition. In 2009 sixth IEEE international conference on advanced video and signal based surveillance, pages 296-301. IEEE, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.54, + 0.469, + 0.607 + ], + "angle": 0, + "content": "[33] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Animatable neural radiance fields for modeling dynamic human bodies. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14314-14323, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.609, + 0.469, + 0.676 + ], + "angle": 0, + "content": "[34] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.679, + 0.469, + 0.747 + ], + "angle": 0, + "content": "[35] Radu Alexandru Rosu, Shunsuke Saito, Ziyan Wang, Chenglei Wu, Sven Behnke, and Giljoo Nam. Neural strands: Learning hair geometry and appearance from multi-view images. In European Conference on Computer Vision, pages 73-89. Springer, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.749, + 0.469, + 0.829 + ], + "angle": 0, + "content": "[36] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerfplayer: A streamable dynamic scene representation with decomposed neural radiance fields. IEEE Transactions on Visualization and Computer Graphics, 29(5):2732-2742, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.832, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[37] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[38] Supasorn Suwajanakorn, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Synthesizing obama: learning lip sync from audio. ACM Transactions on Graphics (ToG), 36(4):1-13, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.149, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[39] Justus Thies, Michael Zollhofer, Marc Stamminger, Christian Theobalt, and Matthias Nießner. Face2face: Real-time face capture and reenactment of rgb videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2387-2395, 2016. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.219, + 0.892, + 0.272 + ], + "angle": 0, + "content": "[40] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: Image synthesis using neural textures. Acm Transactions on Graphics (TOG), 38(4):1-12, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.274, + 0.892, + 0.356 + ], + "angle": 0, + "content": "[41] Ziyan Wang, Giljoo Nam, Tuur Stuyck, Stephen Lombardi, Chen Cao, Jason Saragih, Michael Zollhöfer, Jessica Hodgins, and Christoph Lassner. Neuwigs: A neural dynamic model for volumetric hair capture and animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8641-8651, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.358, + 0.892, + 0.425 + ], + "angle": 0, + "content": "[42] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9421-9431, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.427, + 0.892, + 0.495 + ], + "angle": 0, + "content": "[43] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5438–5448, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.498, + 0.892, + 0.551 + ], + "angle": 0, + "content": "[44] Yuelang Xu, Lizhen Wang, Xiaochen Zhao, Hongwen Zhang, and Yebin Liu. Avatarmav: Fast 3d head avatar reconstruction using motion-aware neural voxels. In ACM SIGGRAPH 2023 Conference Proceedings, 2023. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.553, + 0.892, + 0.62 + ], + "angle": 0, + "content": "[45] Yuelang Xu, Hongwen Zhang, Lizhen Wang, Xiaochen Zhao, Huang Han, Qi Guojun, and Yebin Liu. Latentavatar: Learning latent expression code for expressive neural head avatar. In ACM SIGGRAPH 2023 Conference Proceedings, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.623, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[46] Keyang Ye, Tianjia Shao, and Kun Zhou. Animatable 3d gaussians for high-fidelity synthesis of human motions. arXiv preprint arXiv:2311.13404, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.665, + 0.892, + 0.732 + ], + "angle": 0, + "content": "[47] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4578-4587, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.734, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[48] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[49] Yufeng Zheng, Victoria Fernández Abrevaya, Marcel C Bühler, Xu Chen, Michael J Black, and Otmar Hilliges. Im avatar: Implicit morphable head avatars from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13545-13555, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[50] Yufeng Zheng, Wang Yifan, Gordon Wetzstein, Michael J Black, and Otmar Hilliges. Pointavatar: Deformable point" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20308" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.135 + ], + "angle": 0, + "content": "based head avatars from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21057-21067, 2023. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.178 + ], + "angle": 0, + "content": "[51] Wojciech Zielonka, Timur Bagautdinov, Shunsuke Saito, Michael Zollhöfer, Justus Thies, and Javier Romero. Drivable 3d gaussian avatars. 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.47, + 0.234 + ], + "angle": 0, + "content": "[52] Wojciech Zielonka, Timo Bolkart, and Justus Thies. Instant volumetric head avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4574-4584, 2023. 2, 5, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20309" + } + ] +] \ No newline at end of file diff --git a/2024/GaussianAvatars_ Photorealistic Head Avatars with Rigged 3D Gaussians/41411f0e-3b8e-40c3-bbbd-d16b793f1be0_origin.pdf b/2024/GaussianAvatars_ Photorealistic Head Avatars with Rigged 3D Gaussians/41411f0e-3b8e-40c3-bbbd-d16b793f1be0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..86f416a835afa0de41a098f685c8301144a89380 --- /dev/null +++ b/2024/GaussianAvatars_ Photorealistic Head Avatars with Rigged 3D Gaussians/41411f0e-3b8e-40c3-bbbd-d16b793f1be0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ffd8223618614a089f8b8252b8b193c338cddb4d7f391cce9d9d797e148a9484 +size 3154937 diff --git a/2024/GaussianAvatars_ Photorealistic Head Avatars with Rigged 3D Gaussians/full.md b/2024/GaussianAvatars_ Photorealistic Head Avatars with Rigged 3D Gaussians/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d3f30994cf43566f1626b6dde9e072d59a740877 --- /dev/null +++ b/2024/GaussianAvatars_ Photorealistic Head Avatars with Rigged 3D Gaussians/full.md @@ -0,0 +1,317 @@ +# GaussianAvatars: Photorealistic Head Avatars with Rigged 3D Gaussians + +Shenhan Qian + +Tobias Kirschstein + +Simon Giebenhain + +Liam Schoneveld² + +Matthias Nießner1 + +Davide Davoli3 + +1Technical University of Munich + +$^{2}$ Woven by Toyota + +$^{3}$ Toyota Motor Europe NV/SA + +associated partner by contracted service + +![](images/f59dab12bb1ea77179eb856be965b62c7c983a18f3ffe57ab32782be22c1bb9f.jpg) +Figure 1. We introduce GaussianAvatars, a new method to create photorealistic head avatars from multi-view videos. Our avatar is represented by 3D Gaussian splats rigged to a parametric face model. We can fully control and animate our avatars in terms of pose, expression, and view point as shown in the example renderings above. + +![](images/3e6fbc01b4fa6c1c8d37c709694d69dc3c88fdc2f2ecee4018946d716093785c.jpg) + +![](images/918425896ae7b016e99a03843bfd1e49dd6bfdbb52cf760dd0fae1634a466322.jpg) + +![](images/47968548cd067be4da20614ef275e3f13d4b28c33c946a6f875ee623b2e04447.jpg) + +![](images/8f621531ff7e4ad335a8b7ff4740587045c33681115361931c9dcb852e5c0550.jpg) + +![](images/b178e2f8cd547acafcf9c25dd607f9c79419a78a92102eddb821de93fc45938d.jpg) + +# Abstract + +We introduce GaussianAvatars1, a new method to create photorealistic head avatars that are fully controllable in terms of expression, pose, and viewpoint. The core idea is a dynamic 3D representation based on 3D Gaussian splats that are rigged to a parametric morphable face model. This combination facilitates photorealistic rendering while allowing for precise animation control via the underlying parametric model, e.g., through expression transfer from a driving sequence or by manually changing the morphable model parameters. We parameterize each splat by a local coordinate frame of a triangle and optimize for explicit displacement offset to obtain a more accurate geometric representation. During avatar reconstruction, we jointly optimize for the morphable model parameters and Gaussian splat parameters in an end-to-end fashion. We demonstrate the animation capabilities of our photorealistic avatar in several challenging scenarios. For instance, we show reenactments from a driving video, where our method outperforms existing works by a significant margin. + +# 1. Introduction + +Creating animatable avatars of human heads has been a longstanding problem in computer vision and graphics. + +In particular, the ability to render photorealistic dynamic avatars from arbitrary viewpoints enables numerous applications in gaming, movie production, immersive telepresence, and augmented or virtual reality. For such applications, it is also crucial to be able to control the avatar, and for it to generalize well to novel poses and expressions. + +Reconstructing a 3D representation able to jointly capture the appearance, geometry and dynamics of human heads represents a major challenge for high-fidelity avatar generation. The under-constrained nature of this reconstruction problem significantly complicates the task of achieving a representation that combines novel-view rendering photorealism with expression controllability. Moreover, extreme expressions and facial details, like wrinkles, the mouth interior, and hair, are difficult to capture, and can produce visual artifacts easily noticed by humans. + +Neural Radiance Fields (NeRF [28]) and its variants [2, 6, 29] have shown impressive results in reconstructing static scenes from multi-view observations. Follow-up works have extended NeRF to model dynamic scenes for both arbitrary [22, 30, 31] and human-tailored scenarios [13, 18]. These works achieve impressive results for novel view rendering; however, they lack controllability and as such do not generalize well to novel poses and expressions. + +The recent 3D Gaussian Splitting method [15] achieves even higher rendering quality than NeRF for novel-view synthesis with real-time performance, optimizing for discrete geometric primitives (3D Gaussians) throughout 3D + +space. This method has been extended to capture dynamic scenes by building explicit correspondences across time steps [27]; however, they do not allow for animations of the reconstructed outputs. + +To this end, we propose GaussianAvatars, a method for dynamic 3D representation of human head based on 3D Gaussian splats that are rigged to a parametric morphable face model. Given a FLAME [21] mesh, we initialize a 3D Gaussian at the center of each triangle. When the FLAME mesh is animated, each Gaussian then translates, rotates, and scales according to its parent triangle. The 3D Gaussians then form a radiance field on top of the mesh, compensating for regions where the mesh is either not accurately aligned, or is incapable of reproducing certain visual elements. To achieve high fidelity of the reconstructed avatar, we introduce a binding inheritance strategy to support Gaussian splats without losing controllability. We also explore balancing fidelity and robustness to animate the avatars with novel expressions and poses. Our method outperforms existing works by a significant margin on both novel-view rendering and reenactment from a driving video. + +Our contributions are as follows: + +- We propose GaussianAvatars, a method to create animatable head avatars by rigging 3D Gaussians to parametric mesh models. +- We design a binding inheritance strategy to support adding and removing 3D Gaussians without losing controllability. + +# 2. Related Work + +# 2.1. Radiance Field Reconstruction + +NeRF [28] stores the radiance field of a scene in a neural network and provides photorealistic renderings of novel views with volumetric rendering. Later work [37, 47] represents the scene as voxel grids, achieving comparable rendering quality in a shorter time. Efficiency can be further improved by employing compression methods such as voxel hashing [29] or tensor decomposition [6]. PointNeRF [43] uses point clouds as a scene representation, whereas 3D Gaussian Splatting [15] uses anisotropic 3D Gaussians that are rendered by sorting and rasterization, achieving superior visual quality with real-time rendering. Mixture of Volumetric Primitives [25] uses surface-aligned volumes to achieve fast rendering with high visual fidelity. It also applies residual transformations to primitives. Our method follows 3D Gaussian Splatting [15], benefiting from the expressiveness of anisotropic Gaussians. + +A common paradigm to model a dynamic scene is directly storing it in an MLP with 4D coordinates as the input [9, 42] or a 4D tensor with the time dimension or spatial dimensions compressed [1, 4, 7, 36]. These methods can faithfully replay a dynamic scene and produce realistic + +novel-view rendering but lack an explicit handle to manipulate the content. Another paradigm learns a static canonical space and the maps time steps back to the canonical space via separate MLPs [30, 31, 34]. Instead of using a deformation MLP, a proxy geometry provides more direct controllability. Liu et al. [24] warp points from an observed space to the canonical space based on the movement of the nearest triangle on an SMPL [26] mesh. Peng et al. [33] deform points with the skeleton of SMPL and neural blending weights. Concurrent works create human body avatars with forward deformation [14, 19, 20, 23, 46] and cage-based deformation [51]. Unlike these methods, we directly attach 3D Gaussians to triangles and explicitly move them, obviating the need for a canonical space and enabling effective mesh finetuning. + +# 2.2. Human Head Reconstruction and Animation + +Head avatar creation advanced through the uptake of differentiable rendering and scene representations. Thies et al. [39] instigated a shift toward digital avatars with real-time face tracking and authentic face reenactment. Advancements in image synthesis with neural networks [5, 40] boosted the controllability of head avatars from lip-syncing to expression transfer and head motions [16, 38, 45]. Gafni et al. [8] learn a NeRF conditioned on an expression vector from monocular videos. Grassal et al. [11] subdivide and add offsets to FLAME [21] to enhance its geometry and enable a dynamic texture via an expression-dependent texture field. IMavatar [49] learns a 3D morphable head avatar with neural implicit functions, solving for a map from observed to canonical space via iterative root-finding. HeadNeRF [12] learns a NeRF-based parametric head model with 2D neural rendering for efficiency. + +INSTA [52] deforms query points to a canonical space by finding the nearest triangle on a FLAME [21] mesh, and combines this with InstantNGP [29] to achieve fast rendering. Like INSTA, we also use triangles to warp the scene, but we build a consistent correspondence between a 3D Gaussian and a triangle instead of querying the nearest one for each timestep. Zheng et al. [50] explore point-based representations with differential point splatting. They define a point set in canonical space and learn a deformation field conditioned on FLAME's expression vectors to animate the head avatar. While the scale of a point has to be manually set for their method, it is an estimizable parameter for 3D Gaussians. NeRFBlendShape [10] models a dynamic scene by blending hash tables with 3DMM parameters [3, 32]. AvatarMAV [44] decouples motion and appearance, blending voxel grids only for the motion field. + +# 3. Method + +As shown in Fig. 2, the input to our method is a multi-view video recording of a human head. For each time step, we + +![](images/a1c9c7ee575ff82d5eaa16685d863275bd6674ba0703d139b8583552bbe1f504.jpg) +Figure 2. Overview. Our method binds 3D Gaussian splats to a FLAME [21] mesh locally. We take the tracked mesh for each frame and transform the splats from local to global space before rendering them with 3D Gaussian Splatting [15]. We optimize the splats in the local space by minimizing color loss on the rendering. We add and remove splats adaptively with their binding relation to triangles inherited so that all splats remain rigged throughout the optimization procedure. Further, we regularize the position and scaling of 3D Gaussian splats to suppress artifacts during animation. + +use a photometric head tracker based on [39] to fit FLAME [21] parameters with multi-view observations and known camera parameters. The FLAME meshes have vertices at varied positions but share the same topology. Therefore, we can build a consistent connection between triangles of the mesh and 3D Gaussian splats (Sec. 3.2). The splats are rendered into images via a differentiable tile rasterizer [15]. These images are then supervised by the groundtruth images towards learning a photorealistic human head avatar. As per static scenes, it is also necessary to densify and prune Gaussian splats for optimal quality, with a set of adaptive density control operations [15]. To achieve this without breaking the connection between triangles and splats, we design a binding inheritance strategy (Sec. 3.3) so that new Gaussian points remain rigged to the FLAME mesh. In addition to the color loss, we also find it crucial to regularize the local position and scaling of Gaussian splats to avoid quality degradation under novel expressions and poses (Sec. 3.4). + +# 3.1. Preliminary + +3D Gaussian Splitting [15] provides a solution to reconstruct a static scene with anisotropic 3D Gaussians given images and camera parameters. A scene is represented by a + +set of Gaussian splats, each defined by a covariance matrix $\Sigma$ centered at point (mean) $\mu$ : + +$$ +G (\boldsymbol {x}) = e ^ {- \frac {1}{2} (\boldsymbol {x} - \boldsymbol {\mu}) ^ {T} \Sigma^ {- 1} (\boldsymbol {x} - \boldsymbol {\mu})} \tag {1} +$$ + +Note that covariance matrices have physical meaning only when they are semi-definite, which cannot be guaranteed for an optimization process with gradient descent. Therefore, Kerbl et al. [15] first define a parametric ellipse with a scaling matrix $S$ and a rotation matrix $R$ , then construct the covariance matrix by: + +$$ +\Sigma = R S S ^ {T} R ^ {T}. \tag {2} +$$ + +Practically, an ellipse is stored as a position vector $\pmb{\mu} \in \mathbb{R}^3$ , a scaling vector $\pmb{s} \in \mathbb{R}^3$ , and a quaternion $\pmb{q} \in \mathbb{R}^4$ . In our paper, we use $\pmb{r} \in \mathbb{R}^{3 \times 3}$ to denote the corresponding rotation matrix to $\pmb{q}$ . + +For rendering, the color $C$ of a pixel is computed by blending all 3D Gaussians overlapping the pixel: + +$$ +\boldsymbol {C} = \sum_ {i = 1} \boldsymbol {c} _ {i} \alpha_ {i} ^ {\prime} \prod_ {j = 1} ^ {i - 1} \left(1 - \alpha_ {j} ^ {\prime}\right), \tag {3} +$$ + +where $c_{i}$ is the color of each point modeled by 3-degree spherical harmonics. The blending weight $\alpha^{\prime}$ is given by + +evaluating the 2D projection of the 3D Gaussian multiplied by a per-point opacity $\alpha$ . The Gaussian splats are sorted by depth before blending to respect visibility order. + +# 3.2. 3D Gaussian Rigging + +The key component of our method is how we build connections between the FLAME [21] mesh and 3D Gaussian splats. Initially, we pair each triangle of the mesh with a 3D Gaussian and let the 3D Gaussian move with the triangle across time steps. In other words, the 3D Gaussian is static in the local space of its parent triangle but dynamic in the global (metric) space as the triangle moves. Given the vertices and edges of a triangle, we take the mean position $T$ of the vertices as the origin of the local space. Then, we concatenate the direction vector of one of the edges, the normal vector of the triangle, and their cross product as column vectors to form a rotation matrix $R$ , which describes the orientation of the triangle in the global space. We also compute a scalar $k$ by the mean length of one of the edges and its perpendicular to describe the triangle scaling. + +For the paired 3D Gaussian of a triangle, we define its location $\mu$ , rotation $r$ , and anisotropic scaling $s$ all in the local space. We initialize the location $\mu$ at the local origin, the rotation $r$ as an identity rotation matrix, and the scaling $s$ as a unit vector. At rendering time, we convert these properties into the global space by: + +$$ +\boldsymbol {r} ^ {\prime} = \boldsymbol {R} \boldsymbol {r}, \tag {4} +$$ + +$$ +\boldsymbol {\mu} ^ {\prime} = k \boldsymbol {R} \boldsymbol {\mu} + \boldsymbol {T}, \tag {5} +$$ + +$$ +\boldsymbol {s} ^ {\prime} = k \boldsymbol {s}. \tag {6} +$$ + +We incorporate triangle scaling in Eqs. (5) and (6) so that the local position and scaling of a 3D Gaussian are defined relative to the absolute scale of a triangle. This enables an adaptive step size in the metric space with a constant learning rate for parameters defined in the local space. For example, a 3D Gaussian paired with a smaller triangle will move slower in an iteration step than those paired with large triangles. This also makes interpreting the parameters regarding the distance from the triangle center easier. + +# 3.3. Binding Inheritance + +Only having the same numbers of Gaussian splats as the triangles is insufficient to capture details. For instance, representing a curved hair strand requires multiple splats, while a triangle on the scalp may intersect with several strands. Therefore, we also need the adaptive density control strategy [15], which adds and removes splats based on the view-space positional gradient and the opacity of each Gaussian. + +For each 3D Gaussian with a large view-space positional gradient, we split it into two smaller ones if it is large or clone it if it is small. We conduct this in the local space and ensure a newly created Gaussian is close to the old one that + +triggers this densification operation. Then, it is reasonable to bind a new 3D Gaussian to the same triangle as the old one because it was created to enhance the fidelity of the local region. Therefore, each 3D Gaussian must carry one more parameter, the index of its parent triangle, to enable binding inheritance during densification. + +Besides densification, we also use the pruning operation as a part of the adaptive density control strategy [15]. It periodically resets the opacity of all splats close to zero and removes points with opacity below a threshold. This technique is effective in suppressing floating artifacts, however, such pruning can also cause problems in a dynamic scene. For instance, regions of the face that are often occluded (such as eyeball triangles), can be overly sensitive to this pruning strategy, and often end up with few or no attached Gaussians. To prevent this, we keep track of the number of splats attached to each triangle, and ensure that every triangle always has at least one splat attached. + +# 3.4. Optimization and Regularization + +We supervise the rendered images with a combination of $\mathcal{L}_1$ term and a D-SSIM term following [15]: + +$$ +\mathcal {L} _ {\mathrm {r g b}} = (1 - \lambda) \mathcal {L} _ {1} + \lambda \mathcal {L} _ {\mathrm {D} - \mathrm {S S I M}}, \tag {7} +$$ + +with $\lambda = 0.2$ . This already results in good re-rendering quality without additional supervision, such as depth or silhouette supervision, thanks to the powerful tile rasterizer [15]. We found however, that if we try to animate these splats via FLAME [21] to novel expressions and poses, large spike- and blob-like artifacts appear wildly throughout the scene. This is due to a poor alignment between the Gaussian splats and the triangles. + +Position loss with threshold. A basic hypothesis behind 3D Gaussian rigging is that the Gaussian splats should roughly match the underlying mesh. They should also match their locations; for instance a Gaussian representing a spot on the nose should not be rigged to a triangle on the cheek. Although our splats are initialized at triangle centers, and new splats are added nearby these existing ones, it is not guaranteed that the primitives remain close to their parent triangle after the optimization. To address this, we regularize the local position of each Gaussian by: + +$$ +\mathcal {L} _ {\text {p o s i t i o n}} = \| \max (\boldsymbol {\mu}, \epsilon_ {\text {p o s i t i o n}}) \| _ {2}, \tag {8} +$$ + +where $\epsilon_{\mathrm{position}} = 1$ is a threshold that tolerates small errors within the scaling of its parent triangle. + +Scaling loss with threshold. Aside from position, the scaling of 3D Gaussians is even more essential for the visual quality during animation. Specifically, if a 3D Gaussian is large in comparison to its parent triangle, small rotations of the triangle – barely noticeable at the scale of the triangle – will be magnified by the scale of the 3D Gaussian, resulting + +in unpleasant jittering artifacts. To mitigate this, we also regularize the local scale of each 3D Gaussian by: + +$$ +\mathcal {L} _ {\text {s c a l i n g}} = \left\| \max \left(\boldsymbol {s}, \epsilon_ {\text {s c a l i n g}}\right) \right\| _ {2}, \tag {9} +$$ + +where $\epsilon_{\mathrm{scaling}} = 0.6$ is a threshold that disables this loss term when the scale of a Gaussian less than $0.6\times$ the scale of its parent triangle. This $\epsilon_{\mathrm{scaling}}$ -tolerance is indispensable; without it, the Gaussian splats shrink excessively, causing rendering speeds to deteriorate, as camera rays need to hit more splats before zero transmittance is reached. + +Our final loss function is thus: + +$$ +\mathcal {L} = \mathcal {L} _ {\mathrm {r g b}} + \lambda_ {\text {p o s i t i o n}} \mathcal {L} _ {\text {p o s i t i o n}} + \lambda_ {\text {s c a l i n g}} \mathcal {L} _ {\text {s c a l i n g}}, \tag {10} +$$ + +where $\lambda_{\mathrm{position}} = 0.01$ and $\lambda_{\mathrm{scaling}} = 1$ . Note that we only apply $\mathcal{L}_{\mathrm{position}}$ and $\mathcal{L}_{\mathrm{scaling}}$ to visible splats. Thereby, we only regularize points when the color loss $\mathcal{L}_{\mathrm{rgb}}$ is present. This helps to maintain the learned structure of often-occluded regions such as teeth and eyeballs. + +Implementation details. We use Adam [17] for parameter optimization (the same hyperparameter values are used across all subjects). We set the learning rate to 5e-3 for the position and 1.7e-2 for the scaling of 3D Gaussians and keep the same learning rates as 3D Gaussian Splatting [15] for the rest of the parameters. Alongside the Gaussian splat parameters, we also finetune the translation, joint rotation, and expression parameters of FLAME [21] for each timestep, using learning rates 1e-6, 1e-5, and 1e-3, respectively. We train for 600,000 iterations, and exponentially decay the learning rate for the splat positions until the final iteration, where it reaches $0.01 \times$ the initial value. We enable adaptive density control with binding inheritance every 2,000 iterations, from iteration 10,000 until the end. Every 60,000 iterations, we reset the Gaussians' opacities. We use a photo-metric head tracker to obtain the FLAME parameters, including shape $\beta$ , translation $t$ , pose $\theta$ , expression $\psi$ , and vertex offset $\Delta v$ in the canonical space. + +# 4. Experiments + +# 4.1. Setup + +Settings. We evaluate the quality of head avatars with three settings: 1) novel-view synthesis: driving an avatar with head poses and expressions from training sequences and rendering from a held-out viewpoint. 2) self-reenactment: driving an avatar with unseen poses and expressions from a held-out sequence of the same subject and rendering all 16 camera views. 3) cross-identity reenactment: animating an avatar with poses and expressions from another subject. + +Dataset. We conduct experiments on video recordings of 9 subjects from the NeRSample [18] dataset. All recordings contain 16 views covering the front and sides of a subject. For each subject, we take 11 video sequences and down-sample images to a resolution of $802 \times 550$ . Participants + +
Novel-View SynthesisSelf-Reenactment
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
AvatarMAV [44]29.50.9130.15224.30.8870.168
PointAvatar [50]25.80.8930.09723.40.8840.102
INSTA [52]26.70.8990.12226.30.9060.110
Ours31.60.9380.06526.00.9100.076
+ +Table 1. Quantitative comparison with state-of-the-art methods. Green indicates the best and yellow indicates the second. + +were instructed to perform specific expressions or emotions in 10 sequences and freely perform in the last one. To conduct quantitative evaluation for each subject, we train our method and three baselines (INSTA [52], PointAvatar [50], and AvatarMAV [44]) using 9 out of the 10 prescribed sequences and 15 out of 16 available cameras. We use the $11^{\mathrm{th}}$ (free-performance) sequence to visually assess cross-identity reenactment capabilities. Please refer to the supplementary material for details of the dataset and baselines. + +# 4.2. Head Avatar Reconstruction and Animation + +We evaluate the reconstruction quality of avatars by novel-view synthesis and animation fidelity by self-reenactment. Fig. 3 shows qualitative comparisons. For novel-view synthesis, all methods produce plausible rendering results. A close inspection of PointAvatar's [50] results show dotted artifacts, owing its fixed point size. In our case, the anisotropic scaling of 3D Gaussians alleviates this issue. + +INSTA [52] shows clean results for the face, however regions around the neck and shoulder can be noisy, as the tracked FLAME meshes are often misaligned in those regions. This causes issues for INSTA, as its warping process is based on the nearest triangle. In the case of our method, each Gaussian splat is rigged to a consistent triangle, regardless of the pose or expression. When the tracked mesh is inaccurate, the positional gradient to a Gaussian splat can consistently back-propagate to the same triangle. This enables misalignments due to incorrect FLAME tracking to be corrected while the 3D Gaussians are being optimized. + +AvatarMAV [44] shows comparable quality to other methods in novel-view synthesis but struggles with novel-expression synthesis. This is because it only uses the expression vector of a 3DMM as conditioning. Since the control from the expression vectors to the deformation bases must be learned, it struggles to reproduce expressions that are far from the training distribution. Similar conclusions can be drawn from the quantitative comparison shown in Tab. 1. Our approach outperforms others by a large margin regarding metrics for novel-view synthesis. Our method also stands out in self-reenactment, with significantly lower perceptual differences in terms of LPIPS [48]. Note that self-reenactment is based on tracked FLAME meshes that + +![](images/63b0543c00372643cd9905b20c1c1dc1ce384518bff6d979432871d41c22c233.jpg) +Figure 3. Qualitative comparison on novel-view synthesis and self-reenactment of head avatars. Our method outperforms state-of-the-art methods by producing significantly sharper rendering outputs. We obtain precise reconstruction of details such as reflective light on eyes, hair strands, teeth, etc. Our results for self-reenactment show more accurate expressions compare to baselines. + +![](images/8d168d18932ce5a214ea0236828bf75ca52660c73dfc0e91ae22a35af2241f43.jpg) +Figure 4. Cross-identity reenactment of head avatars. We use the tracked FLAME expression and pose parameters of source actors to drive the reconstructed avatars. Our method produces high-quality rendering and transfers expressions vividly, while baseline methods suffer from artifacts and generalize poorly to novel expressions. + +may not perfectly align with the target images, thus bringing disadvantages to our results with more visual details regarding pixel-wise metrics such as PSNR. + +For a real-world test of avatar animation, we conduct experiments on cross-identity reenactment in Fig. 4. Our avatars accurately reproduce eye blinks and mouth movements from source actors showing lively, complex dynamics such as wrinkles. INSTA [52] suffers from aliasing artifacts when the avatars move beyond the occupancy grid of I-NGP [29] optimized for training sequences. The movement of results from PointAvatar [50] is not precise because its deformation space is not guaranteed to be consistent with FLAME. AvatarMAV [44] exhibits large degradations in reenactment due to a lack of deformation priors. + +# 4.3. Ablation Study + +To validate the effectiveness of our method components, we deactivate each of them and report results in Tab. 2. + +![](images/d66860fcbeff6c9a795e3ddf45bc43e2c4b25c44f93ae5d501d4a4f498d9040d.jpg) +Figure 5. Fine-tuning FLAME parameters leads to better alignment of the mesh to the input image. In the example above, the movements of cheeks and lips are better captured after fine-tuning. + +![](images/c7e03eab471cd10db96f4e471599724263803167295ac80cb4664fa1966b2009.jpg) + +![](images/062dfa4389a1cdd500091b7c605f119e83a99bd3c1bc38b645d913a485d5074b.jpg) + +Adaptive density control with binding inheritance. Without binding inheritance, we lose the ability to add and remove Gaussian splats besides the initial ones. With a lim + +
Novel-ViewSelf-Reenactment
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Ours28.80.8830.09825.10.8530.101
w/o ADC26.80.8540.20625.10.8600.183
w/o Lscaling28.00.8770.11424.90.8520.109
w/o εscaling25.00.8330.19524.10.8430.176
w/o Lposition29.70.8940.09124.90.8510.096
w/o εposition28.70.8820.10525.00.8550.106
w/o FLAME ft.26.10.8550.13125.50.8620.124
+ +Table 2. Ablation study on subject #304. Green indicates the best and yellow indicates the second. “ADC” refers to Adaptive Density Control with binding inheritance. ‘FLAME ft.’ refers to FLAME parameter fine-tuning. + +ited number of splats, the scaling of each splat increases to occupy the same space, causing blurry renderings, leading to huge fidelity loss, as shown in the second row of Tab. 2. + +Regularization on the scaling of Gaussian splats. Without the scaling loss, artifacts show up as spikes (the middle column in Fig. 6), and slightly worsen the image quality (the third row in Tab. 2). But when we use the scaling loss without an error tolerance, all metrics drastically deteriorate (the fourth row in Tab. 2) because all the splats are regularized to be infinitely small, making it hard to construct opaque surfaces. + +Regularization on the position of Gaussian splats. When disabling the threshold for the position loss, we get slightly worse metrics (the sixth row in Tab. 2). But when turning off the position loss, the metrics on novel-view synthesis become the best in the table (the fifth row in Tab. 2). This is reasonable because without constraints to the mesh Gaussian splats can freely move in the space to achieve minimal re-rendering error. However, since the distribution of Gaussian splats are overfitted to training frames, the avatar shows artifacts such as cracks and fly-around blobs (the left column in Fig. 6) in novel expressions and poses. Therefore, the position loss is necessary to enforce a conservative deviation of splats from the mesh surface so that we can animate the reconstructed avatar with unseen facial motion. + +FLAME fine-tuning. Thanks to our consistent binding between Gaussian splats and triangles, we can effectively fine-tune FLAME parameters while optimizing Gaussian splats (Fig. 5). As shown in the last row in Tab. 2, turning off the fine-tuning negatively affects image quality for novel-view synthesis. For self-reenactment, fine-tuning FLAME parameters also leads to lower perceptual error. + +# 5. Limitations and Potential Negative Impacts + +Our approach, based on 3D Gaussian splats, directly captures the radiance field without decoupling material and lighting. Hence, relighting the avatar is not feasible within + +![](images/3ae5dfe4b1dbc53e16309a6ca9b831392ec6b77995354bd7e0b1b12f65ccc3ef.jpg) +Figure 6. The position loss and the scaling loss helps prevent artifacts during animation with novel expressions and poses. + +our current approach. Additionally, we lack control over areas that are not modeled by FLAME such as hair and other accessories. Here, we believe that modeling these parts directly would provide an interesting future research avenue; for instance, recent methods show promising result in reconstructing hair [35, 41]. + +The creation of photorealistic avatars may introduce a wide spectrum of malevolent uses. Privacy violations are a significant concern, with potential for unauthorized manipulation of people's likeness. Moreover the creation of deceptive content, like deepfake videos, can lead to misinformation, defamation and harm to reputations. Identity theft is an additional risk, as avatars may be exploited for impersonation and fraudulent activities. We strongly condemn unauthorized and malevolent uses of this technology, emphasizing the need for ethical considerations in all applications utilizing our method. + +# 6. Conclusion + +GaussianAvatars is a novel approach that creates photorealistic avatars from video sequences. It features a dynamic 3D representation based on 3D Gaussian splats that are rigged to a parametric morphable face model. This enables flexible control and precise attribute transfer of a dynamic, photorealistic avatar. The set of Gaussian splats can deviate from the mesh surface to compensate the absence or inaccuracy of the morphable model, exhibiting astonishing ability to model fine details on human heads. Our approach outperforms state-of-the-art methods on image quality and expression accuracy for a large margin, indicating a large potential for more applications in related domains. + +# Acknowledgements + +This work was supported by Toyota Motor Europe and Woven by Toyota. This work was also supported by the ERC Starting Grant Scan2CAD (804724). We thank Justus Thies, Andrei Burov, and Lei Li for constructive discussions and Angela Dai for the video voice-over. + +# References + +[1] Benjamin Attal, Jia-Bin Huang, Christian Richardt, Michael Zollhoefer, Johannes Kopf, Matthew O'Toole, and Changil Kim. Hyperreel: High-fidelity 6-dof video with ray-conditioned sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16610-16620, 2023. 2 +[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 1 +[3] Volker Blanz and Thomas Vetter. Face recognition based on fitting a 3d morphable model. IEEE Transactions on pattern analysis and machine intelligence, 25(9):1063-1074, 2003. 2 +[4] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 130-141, 2023. 2 +[5] Caroline Chan, Shiry Ginosar, Tinghui Zhou, and Alexei A Efros. Everybody dance now. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5933-5942, 2019. 2 +[6] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 1, 2 +[7] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbaek Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12479-12488, 2023. 2 +[8] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 2 +[9] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5712-5721, 2021. 2 +[10] Xuan Gao, Chenglai Zhong, Jun Xiang, Yang Hong, Yudong Guo, and Juyong Zhang. Reconstructing personalized semantic facial nerf models from monocular video. ACM Transactions on Graphics (Proceedings of SIGGRAPH Asia), 41(6), 2022. 2 +[11] Philip-William Grassal, Malte Prinzler, Titus Leistner, Carsten Rother, Matthias Nießner, and Justus Thies. Neural head avatars from monocular rgb videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18653–18664, 2022. 2 +[12] Yang Hong, Bo Peng, Haiyao Xiao, Ligang Liu, and Juyong Zhang. Headnerf: A real-time nerf-based parametric head model. In Proceedings of the IEEE/CVF Conference + +on Computer Vision and Pattern Recognition, pages 20374-20384, 2022. 2 +[13] Mustafa Işik, Martin Rünz, Markos Georgopoulos, Taras Khakhulin, Jonathan Starck, Lourdes Agapito, and Matthias Nießner. Humanrf: High-fidelity neural radiance fields for humans in motion. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 1 +[14] Rohit Jena, Ganesh Subramanian Iyer, Siddharth Choudhary, Brandon Smith, Pratik Chaudhari, and James Gee. Splatarmor: Articulated gaussian splattering for animatable humans from monocular rgb videos. arXiv preprint arXiv:2311.10812, 2023. 2 +[15] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 42(4):1-14, 2023. 1, 2, 3, 4, 5 +[16] Hyeongwoo Kim, Pablo Garrido, Ayush Tewari, Weipeng Xu, Justus Thies, Matthias Niessner, Patrick Pérez, Christian Richardt, Michael Zollhöfer, and Christian Theobalt. Deep video portraits. ACM transactions on graphics (TOG), 37(4): 1-14, 2018. 2 +[17] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5 +[18] Tobias Kirschstein, Shenhan Qian, Simon Giebenhain, Tim Walter, and Matthias Nießner. Nersemble: Multi-view radiance field reconstruction of human heads. ACM Trans. Graph., 42(4), 2023. 1, 5 +[19] Muhammed Kocabas, Jen-Hao Rick Chang, James Gabriel, Oncel Tuzel, and Anurag Ranjan. Hugs: Human gaussian splats. arXiv preprint arXiv:2311.17910, 2023. 2 +[20] Jiahui Lei, Yufu Wang, Georgios Pavlakos, Lingjie Liu, and Kostas Daniilidis. Gart: Gaussian articulated template models. arXiv preprint arXiv:2311.16099, 2023. 2 +[21] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6):194:1-194:17, 2017. 2, 3, 4, 5 +[22] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 1 +[23] Zhe Li, Zerong Zheng, Lizhen Wang, and Yebin Liu. Animatable gaussians: Learning pose-dependent gaussian maps for high-fidelity human avatar modeling. arXiv, 2023. 2 +[24] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. ACM transactions on graphics (TOG), 40(6):1-16, 2021. 2 +[25] Stephen Lombardi, Tomas Simon, Gabriel Schwartz, Michael Zollhoefer, Yaser Sheikh, and Jason Saragih. Mixture of volumetric primitives for efficient neural rendering. ACM Transactions on Graphics (ToG), 40(4):1-13, 2021. 2 + +[26] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. In *Seminal Graphics Papers: Pushing the Boundaries*, Volume 2, pages 851–866. 2023. 2 +[27] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 2 +[28] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2 +[29] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 1, 2, 7 +[30] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 1, 2 +[31] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. 1, 2 +[32] Pascal Paysan, Reinhard Knothe, Brian Amberg, Sami Romdhani, and Thomas Vetter. A 3d face model for pose and illumination invariant face recognition. In 2009 sixth IEEE international conference on advanced video and signal based surveillance, pages 296-301. IEEE, 2009. 2 +[33] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Animatable neural radiance fields for modeling dynamic human bodies. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14314-14323, 2021. 2 +[34] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 2 +[35] Radu Alexandru Rosu, Shunsuke Saito, Ziyan Wang, Chenglei Wu, Sven Behnke, and Giljoo Nam. Neural strands: Learning hair geometry and appearance from multi-view images. In European Conference on Computer Vision, pages 73-89. Springer, 2022. 8 +[36] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerfplayer: A streamable dynamic scene representation with decomposed neural radiance fields. IEEE Transactions on Visualization and Computer Graphics, 29(5):2732-2742, 2023. 2 +[37] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 2 + +[38] Supasorn Suwajanakorn, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Synthesizing obama: learning lip sync from audio. ACM Transactions on Graphics (ToG), 36(4):1-13, 2017. 2 +[39] Justus Thies, Michael Zollhofer, Marc Stamminger, Christian Theobalt, and Matthias Nießner. Face2face: Real-time face capture and reenactment of rgb videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2387-2395, 2016. 2, 3 +[40] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: Image synthesis using neural textures. Acm Transactions on Graphics (TOG), 38(4):1-12, 2019. 2 +[41] Ziyan Wang, Giljoo Nam, Tuur Stuyck, Stephen Lombardi, Chen Cao, Jason Saragih, Michael Zollhöfer, Jessica Hodgins, and Christoph Lassner. Neuwigs: A neural dynamic model for volumetric hair capture and animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8641-8651, 2023. 8 +[42] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9421-9431, 2021. 2 +[43] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5438–5448, 2022. 2 +[44] Yuelang Xu, Lizhen Wang, Xiaochen Zhao, Hongwen Zhang, and Yebin Liu. Avatarmav: Fast 3d head avatar reconstruction using motion-aware neural voxels. In ACM SIGGRAPH 2023 Conference Proceedings, 2023. 2, 5, 7 +[45] Yuelang Xu, Hongwen Zhang, Lizhen Wang, Xiaochen Zhao, Huang Han, Qi Guojun, and Yebin Liu. Latentavatar: Learning latent expression code for expressive neural head avatar. In ACM SIGGRAPH 2023 Conference Proceedings, 2023. 2 +[46] Keyang Ye, Tianjia Shao, and Kun Zhou. Animatable 3d gaussians for high-fidelity synthesis of human motions. arXiv preprint arXiv:2311.13404, 2023. 2 +[47] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4578-4587, 2021. 2 +[48] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5 +[49] Yufeng Zheng, Victoria Fernández Abrevaya, Marcel C Bühler, Xu Chen, Michael J Black, and Otmar Hilliges. Im avatar: Implicit morphable head avatars from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13545-13555, 2022. 2 +[50] Yufeng Zheng, Wang Yifan, Gordon Wetzstein, Michael J Black, and Otmar Hilliges. Pointavatar: Deformable point + +based head avatars from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21057-21067, 2023. 2, 5, 7 +[51] Wojciech Zielonka, Timur Bagautdinov, Shunsuke Saito, Michael Zollhöfer, Justus Thies, and Javier Romero. Drivable 3d gaussian avatars. 2023. 2 +[52] Wojciech Zielonka, Timo Bolkart, and Justus Thies. Instant volumetric head avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4574-4584, 2023. 2, 5, 7 \ No newline at end of file diff --git a/2024/GaussianAvatars_ Photorealistic Head Avatars with Rigged 3D Gaussians/images.zip b/2024/GaussianAvatars_ Photorealistic Head Avatars with Rigged 3D Gaussians/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..82fdc11c5cb47c94481ea01985824af04c3b5496 --- /dev/null +++ b/2024/GaussianAvatars_ Photorealistic Head Avatars with Rigged 3D Gaussians/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9e4590eaaf198d5ad9e4411d83236042ce5922108ddcc1038f338a08e7c88363 +size 632035 diff --git a/2024/GaussianAvatars_ Photorealistic Head Avatars with Rigged 3D Gaussians/layout.json b/2024/GaussianAvatars_ Photorealistic Head Avatars with Rigged 3D Gaussians/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..42061e0493988f13a686a9e525ad3568474170bc --- /dev/null +++ b/2024/GaussianAvatars_ Photorealistic Head Avatars with Rigged 3D Gaussians/layout.json @@ -0,0 +1,7546 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 70, + 103, + 523, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 103, + 523, + 121 + ], + "spans": [ + { + "bbox": [ + 70, + 103, + 523, + 121 + ], + "type": "text", + "content": "GaussianAvatars: Photorealistic Head Avatars with Rigged 3D Gaussians" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 97, + 142, + 172, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 142, + 172, + 158 + ], + "spans": [ + { + "bbox": [ + 97, + 142, + 172, + 158 + ], + "type": "text", + "content": "Shenhan Qian" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 197, + 143, + 292, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 143, + 292, + 156 + ], + "spans": [ + { + "bbox": [ + 197, + 143, + 292, + 156 + ], + "type": "text", + "content": "Tobias Kirschstein" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 197, + 158, + 291, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 158, + 291, + 170 + ], + "spans": [ + { + "bbox": [ + 197, + 158, + 291, + 170 + ], + "type": "text", + "content": "Simon Giebenhain" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 318, + 143, + 408, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 143, + 408, + 156 + ], + "spans": [ + { + "bbox": [ + 318, + 143, + 408, + 156 + ], + "type": "text", + "content": "Liam Schoneveld²" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 318, + 158, + 406, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 158, + 406, + 171 + ], + "spans": [ + { + "bbox": [ + 318, + 158, + 406, + 171 + ], + "type": "text", + "content": "Matthias Nießner1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 433, + 143, + 509, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 433, + 143, + 509, + 156 + ], + "spans": [ + { + "bbox": [ + 433, + 143, + 509, + 156 + ], + "type": "text", + "content": "Davide Davoli3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 78, + 176, + 238, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 176, + 238, + 191 + ], + "spans": [ + { + "bbox": [ + 78, + 176, + 238, + 191 + ], + "type": "text", + "content": "1Technical University of Munich" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 264, + 176, + 356, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 176, + 356, + 191 + ], + "spans": [ + { + "bbox": [ + 264, + 176, + 356, + 191 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 264, + 176, + 356, + 191 + ], + "type": "text", + "content": "Woven by Toyota" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 383, + 177, + 533, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 177, + 533, + 190 + ], + "spans": [ + { + "bbox": [ + 383, + 177, + 533, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 383, + 177, + 533, + 190 + ], + "type": "text", + "content": "Toyota Motor Europe NV/SA" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 394, + 189, + 523, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 189, + 523, + 198 + ], + "spans": [ + { + "bbox": [ + 394, + 189, + 523, + 198 + ], + "type": "text", + "content": "associated partner by contracted service" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 47, + 224, + 135, + 322 + ], + "blocks": [ + { + "bbox": [ + 47, + 224, + 135, + 322 + ], + "lines": [ + { + "bbox": [ + 47, + 224, + 135, + 322 + ], + "spans": [ + { + "bbox": [ + 47, + 224, + 135, + 322 + ], + "type": "image", + "image_path": "f59dab12bb1ea77179eb856be965b62c7c983a18f3ffe57ab32782be22c1bb9f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 330, + 547, + 364 + ], + "lines": [ + { + "bbox": [ + 46, + 330, + 547, + 364 + ], + "spans": [ + { + "bbox": [ + 46, + 330, + 547, + 364 + ], + "type": "text", + "content": "Figure 1. We introduce GaussianAvatars, a new method to create photorealistic head avatars from multi-view videos. Our avatar is represented by 3D Gaussian splats rigged to a parametric face model. We can fully control and animate our avatars in terms of pose, expression, and view point as shown in the example renderings above." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 138, + 224, + 219, + 321 + ], + "blocks": [ + { + "bbox": [ + 138, + 224, + 219, + 321 + ], + "lines": [ + { + "bbox": [ + 138, + 224, + 219, + 321 + ], + "spans": [ + { + "bbox": [ + 138, + 224, + 219, + 321 + ], + "type": "image", + "image_path": "3e6fbc01b4fa6c1c8d37c709694d69dc3c88fdc2f2ecee4018946d716093785c.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 223, + 223, + 296, + 321 + ], + "blocks": [ + { + "bbox": [ + 223, + 223, + 296, + 321 + ], + "lines": [ + { + "bbox": [ + 223, + 223, + 296, + 321 + ], + "spans": [ + { + "bbox": [ + 223, + 223, + 296, + 321 + ], + "type": "image", + "image_path": "918425896ae7b016e99a03843bfd1e49dd6bfdbb52cf760dd0fae1634a466322.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 303, + 223, + 377, + 321 + ], + "blocks": [ + { + "bbox": [ + 303, + 223, + 377, + 321 + ], + "lines": [ + { + "bbox": [ + 303, + 223, + 377, + 321 + ], + "spans": [ + { + "bbox": [ + 303, + 223, + 377, + 321 + ], + "type": "image", + "image_path": "47968548cd067be4da20614ef275e3f13d4b28c33c946a6f875ee623b2e04447.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 378, + 230, + 459, + 321 + ], + "blocks": [ + { + "bbox": [ + 378, + 230, + 459, + 321 + ], + "lines": [ + { + "bbox": [ + 378, + 230, + 459, + 321 + ], + "spans": [ + { + "bbox": [ + 378, + 230, + 459, + 321 + ], + "type": "image", + "image_path": "8f621531ff7e4ad335a8b7ff4740587045c33681115361931c9dcb852e5c0550.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 461, + 230, + 545, + 321 + ], + "blocks": [ + { + "bbox": [ + 461, + 230, + 545, + 321 + ], + "lines": [ + { + "bbox": [ + 461, + 230, + 545, + 321 + ], + "spans": [ + { + "bbox": [ + 461, + 230, + 545, + 321 + ], + "type": "image", + "image_path": "b178e2f8cd547acafcf9c25dd607f9c79419a78a92102eddb821de93fc45938d.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 143, + 373, + 192, + 386 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 373, + 192, + 386 + ], + "spans": [ + { + "bbox": [ + 143, + 373, + 192, + 386 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 45, + 399, + 290, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 399, + 290, + 627 + ], + "spans": [ + { + "bbox": [ + 45, + 399, + 290, + 627 + ], + "type": "text", + "content": "We introduce GaussianAvatars1, a new method to create photorealistic head avatars that are fully controllable in terms of expression, pose, and viewpoint. The core idea is a dynamic 3D representation based on 3D Gaussian splats that are rigged to a parametric morphable face model. This combination facilitates photorealistic rendering while allowing for precise animation control via the underlying parametric model, e.g., through expression transfer from a driving sequence or by manually changing the morphable model parameters. We parameterize each splat by a local coordinate frame of a triangle and optimize for explicit displacement offset to obtain a more accurate geometric representation. During avatar reconstruction, we jointly optimize for the morphable model parameters and Gaussian splat parameters in an end-to-end fashion. We demonstrate the animation capabilities of our photorealistic avatar in several challenging scenarios. For instance, we show reenactments from a driving video, where our method outperforms existing works by a significant margin." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 47, + 650, + 128, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 650, + 128, + 662 + ], + "spans": [ + { + "bbox": [ + 47, + 650, + 128, + 662 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 46, + 670, + 287, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 670, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 46, + 670, + 287, + 694 + ], + "type": "text", + "content": "Creating animatable avatars of human heads has been a longstanding problem in computer vision and graphics." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 374, + 545, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 374, + 545, + 446 + ], + "spans": [ + { + "bbox": [ + 306, + 374, + 545, + 446 + ], + "type": "text", + "content": "In particular, the ability to render photorealistic dynamic avatars from arbitrary viewpoints enables numerous applications in gaming, movie production, immersive telepresence, and augmented or virtual reality. For such applications, it is also crucial to be able to control the avatar, and for it to generalize well to novel poses and expressions." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 448, + 546, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 448, + 546, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 448, + 546, + 567 + ], + "type": "text", + "content": "Reconstructing a 3D representation able to jointly capture the appearance, geometry and dynamics of human heads represents a major challenge for high-fidelity avatar generation. The under-constrained nature of this reconstruction problem significantly complicates the task of achieving a representation that combines novel-view rendering photorealism with expression controllability. Moreover, extreme expressions and facial details, like wrinkles, the mouth interior, and hair, are difficult to capture, and can produce visual artifacts easily noticed by humans." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 568, + 546, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 568, + 546, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 568, + 546, + 664 + ], + "type": "text", + "content": "Neural Radiance Fields (NeRF [28]) and its variants [2, 6, 29] have shown impressive results in reconstructing static scenes from multi-view observations. Follow-up works have extended NeRF to model dynamic scenes for both arbitrary [22, 30, 31] and human-tailored scenarios [13, 18]. These works achieve impressive results for novel view rendering; however, they lack controllability and as such do not generalize well to novel poses and expressions." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "The recent 3D Gaussian Splitting method [15] achieves even higher rendering quality than NeRF for novel-view synthesis with real-time performance, optimizing for discrete geometric primitives (3D Gaussians) throughout 3D" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 256, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 256, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 256, + 713 + ], + "type": "text", + "content": "1Project page: https://shenhanqian.github.io/gaussian-avatars" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20299" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "space. This method has been extended to capture dynamic scenes by building explicit correspondences across time steps [27]; however, they do not allow for animations of the reconstructed outputs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 121, + 288, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 288, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 288, + 324 + ], + "type": "text", + "content": "To this end, we propose GaussianAvatars, a method for dynamic 3D representation of human head based on 3D Gaussian splats that are rigged to a parametric morphable face model. Given a FLAME [21] mesh, we initialize a 3D Gaussian at the center of each triangle. When the FLAME mesh is animated, each Gaussian then translates, rotates, and scales according to its parent triangle. The 3D Gaussians then form a radiance field on top of the mesh, compensating for regions where the mesh is either not accurately aligned, or is incapable of reproducing certain visual elements. To achieve high fidelity of the reconstructed avatar, we introduce a binding inheritance strategy to support Gaussian splats without losing controllability. We also explore balancing fidelity and robustness to animate the avatars with novel expressions and poses. Our method outperforms existing works by a significant margin on both novel-view rendering and reenactment from a driving video." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 327, + 181, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 327, + 181, + 338 + ], + "spans": [ + { + "bbox": [ + 47, + 327, + 181, + 338 + ], + "type": "text", + "content": "Our contributions are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 340, + 286, + 411 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 47, + 340, + 286, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 340, + 286, + 374 + ], + "spans": [ + { + "bbox": [ + 47, + 340, + 286, + 374 + ], + "type": "text", + "content": "- We propose GaussianAvatars, a method to create animatable head avatars by rigging 3D Gaussians to parametric mesh models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 376, + 286, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 376, + 286, + 411 + ], + "spans": [ + { + "bbox": [ + 47, + 376, + 286, + 411 + ], + "type": "text", + "content": "- We design a binding inheritance strategy to support adding and removing 3D Gaussians without losing controllability." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 423, + 134, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 423, + 134, + 435 + ], + "spans": [ + { + "bbox": [ + 47, + 423, + 134, + 435 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 443, + 214, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 443, + 214, + 454 + ], + "spans": [ + { + "bbox": [ + 47, + 443, + 214, + 454 + ], + "type": "text", + "content": "2.1. Radiance Field Reconstruction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 462, + 286, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 286, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 286, + 653 + ], + "type": "text", + "content": "NeRF [28] stores the radiance field of a scene in a neural network and provides photorealistic renderings of novel views with volumetric rendering. Later work [37, 47] represents the scene as voxel grids, achieving comparable rendering quality in a shorter time. Efficiency can be further improved by employing compression methods such as voxel hashing [29] or tensor decomposition [6]. PointNeRF [43] uses point clouds as a scene representation, whereas 3D Gaussian Splatting [15] uses anisotropic 3D Gaussians that are rendered by sorting and rasterization, achieving superior visual quality with real-time rendering. Mixture of Volumetric Primitives [25] uses surface-aligned volumes to achieve fast rendering with high visual fidelity. It also applies residual transformations to primitives. Our method follows 3D Gaussian Splatting [15], benefiting from the expressiveness of anisotropic Gaussians." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 654, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 286, + 713 + ], + "type": "text", + "content": "A common paradigm to model a dynamic scene is directly storing it in an MLP with 4D coordinates as the input [9, 42] or a 4D tensor with the time dimension or spatial dimensions compressed [1, 4, 7, 36]. These methods can faithfully replay a dynamic scene and produce realistic" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "type": "text", + "content": "novel-view rendering but lack an explicit handle to manipulate the content. Another paradigm learns a static canonical space and the maps time steps back to the canonical space via separate MLPs [30, 31, 34]. Instead of using a deformation MLP, a proxy geometry provides more direct controllability. Liu et al. [24] warp points from an observed space to the canonical space based on the movement of the nearest triangle on an SMPL [26] mesh. Peng et al. [33] deform points with the skeleton of SMPL and neural blending weights. Concurrent works create human body avatars with forward deformation [14, 19, 20, 23, 46] and cage-based deformation [51]. Unlike these methods, we directly attach 3D Gaussians to triangles and explicitly move them, obviating the need for a canonical space and enabling effective mesh finetuning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 258, + 538, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 258, + 538, + 270 + ], + "spans": [ + { + "bbox": [ + 305, + 258, + 538, + 270 + ], + "type": "text", + "content": "2.2. Human Head Reconstruction and Animation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 277, + 545, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 545, + 468 + ], + "type": "text", + "content": "Head avatar creation advanced through the uptake of differentiable rendering and scene representations. Thies et al. [39] instigated a shift toward digital avatars with real-time face tracking and authentic face reenactment. Advancements in image synthesis with neural networks [5, 40] boosted the controllability of head avatars from lip-syncing to expression transfer and head motions [16, 38, 45]. Gafni et al. [8] learn a NeRF conditioned on an expression vector from monocular videos. Grassal et al. [11] subdivide and add offsets to FLAME [21] to enhance its geometry and enable a dynamic texture via an expression-dependent texture field. IMavatar [49] learns a 3D morphable head avatar with neural implicit functions, solving for a map from observed to canonical space via iterative root-finding. HeadNeRF [12] learns a NeRF-based parametric head model with 2D neural rendering for efficiency." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 468, + 545, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 468, + 545, + 659 + ], + "spans": [ + { + "bbox": [ + 304, + 468, + 545, + 659 + ], + "type": "text", + "content": "INSTA [52] deforms query points to a canonical space by finding the nearest triangle on a FLAME [21] mesh, and combines this with InstantNGP [29] to achieve fast rendering. Like INSTA, we also use triangles to warp the scene, but we build a consistent correspondence between a 3D Gaussian and a triangle instead of querying the nearest one for each timestep. Zheng et al. [50] explore point-based representations with differential point splatting. They define a point set in canonical space and learn a deformation field conditioned on FLAME's expression vectors to animate the head avatar. While the scale of a point has to be manually set for their method, it is an estimizable parameter for 3D Gaussians. NeRFBlendShape [10] models a dynamic scene by blending hash tables with 3DMM parameters [3, 32]. AvatarMAV [44] decouples motion and appearance, blending voxel grids only for the motion field." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 669, + 361, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 669, + 361, + 681 + ], + "spans": [ + { + "bbox": [ + 306, + 669, + 361, + 681 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "As shown in Fig. 2, the input to our method is a multi-view video recording of a human head. For each time step, we" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20300" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 547, + 340 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 547, + 340 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 547, + 340 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 547, + 340 + ], + "type": "image", + "image_path": "a1c9c7ee575ff82d5eaa16685d863275bd6674ba0703d139b8583552bbe1f504.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 348, + 546, + 403 + ], + "lines": [ + { + "bbox": [ + 46, + 348, + 546, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 546, + 403 + ], + "type": "text", + "content": "Figure 2. Overview. Our method binds 3D Gaussian splats to a FLAME [21] mesh locally. We take the tracked mesh for each frame and transform the splats from local to global space before rendering them with 3D Gaussian Splatting [15]. We optimize the splats in the local space by minimizing color loss on the rendering. We add and remove splats adaptively with their binding relation to triangles inherited so that all splats remain rigged throughout the optimization procedure. Further, we regularize the position and scaling of 3D Gaussian splats to suppress artifacts during animation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 424, + 289, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 424, + 289, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 424, + 289, + 651 + ], + "type": "text", + "content": "use a photometric head tracker based on [39] to fit FLAME [21] parameters with multi-view observations and known camera parameters. The FLAME meshes have vertices at varied positions but share the same topology. Therefore, we can build a consistent connection between triangles of the mesh and 3D Gaussian splats (Sec. 3.2). The splats are rendered into images via a differentiable tile rasterizer [15]. These images are then supervised by the groundtruth images towards learning a photorealistic human head avatar. As per static scenes, it is also necessary to densify and prune Gaussian splats for optimal quality, with a set of adaptive density control operations [15]. To achieve this without breaking the connection between triangles and splats, we design a binding inheritance strategy (Sec. 3.3) so that new Gaussian points remain rigged to the FLAME mesh. In addition to the color loss, we also find it crucial to regularize the local position and scaling of Gaussian splats to avoid quality degradation under novel expressions and poses (Sec. 3.4)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 658, + 127, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 127, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 127, + 672 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "3D Gaussian Splitting [15] provides a solution to reconstruct a static scene with anisotropic 3D Gaussians given images and camera parameters. A scene is represented by a" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 424, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 424, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 305, + 424, + 545, + 449 + ], + "type": "text", + "content": "set of Gaussian splats, each defined by a covariance matrix " + }, + { + "bbox": [ + 305, + 424, + 545, + 449 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 305, + 424, + 545, + 449 + ], + "type": "text", + "content": " centered at point (mean) " + }, + { + "bbox": [ + 305, + 424, + 545, + 449 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 305, + 424, + 545, + 449 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 364, + 453, + 545, + 469 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 453, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 364, + 453, + 545, + 469 + ], + "type": "interline_equation", + "content": "G (\\boldsymbol {x}) = e ^ {- \\frac {1}{2} (\\boldsymbol {x} - \\boldsymbol {\\mu}) ^ {T} \\Sigma^ {- 1} (\\boldsymbol {x} - \\boldsymbol {\\mu})} \\tag {1}", + "image_path": "0bbbd0073218c488aeec1ccefd1f3d09b037527c34efb74e40fd2f04fab9e1c1.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 475, + 545, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 475, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 545, + 547 + ], + "type": "text", + "content": "Note that covariance matrices have physical meaning only when they are semi-definite, which cannot be guaranteed for an optimization process with gradient descent. Therefore, Kerbl et al. [15] first define a parametric ellipse with a scaling matrix " + }, + { + "bbox": [ + 304, + 475, + 545, + 547 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 475, + 545, + 547 + ], + "type": "text", + "content": " and a rotation matrix " + }, + { + "bbox": [ + 304, + 475, + 545, + 547 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 475, + 545, + 547 + ], + "type": "text", + "content": ", then construct the covariance matrix by:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 391, + 552, + 545, + 566 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 552, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 391, + 552, + 545, + 566 + ], + "type": "interline_equation", + "content": "\\Sigma = R S S ^ {T} R ^ {T}. \\tag {2}", + "image_path": "8c732a60306c92adaa554df275455af88fe2ec0f75f2dfd0a0a22ac5c2413a75.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 572, + 545, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 572, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 304, + 572, + 545, + 620 + ], + "type": "text", + "content": "Practically, an ellipse is stored as a position vector " + }, + { + "bbox": [ + 304, + 572, + 545, + 620 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 572, + 545, + 620 + ], + "type": "text", + "content": ", a scaling vector " + }, + { + "bbox": [ + 304, + 572, + 545, + 620 + ], + "type": "inline_equation", + "content": "\\pmb{s} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 572, + 545, + 620 + ], + "type": "text", + "content": ", and a quaternion " + }, + { + "bbox": [ + 304, + 572, + 545, + 620 + ], + "type": "inline_equation", + "content": "\\pmb{q} \\in \\mathbb{R}^4" + }, + { + "bbox": [ + 304, + 572, + 545, + 620 + ], + "type": "text", + "content": ". In our paper, we use " + }, + { + "bbox": [ + 304, + 572, + 545, + 620 + ], + "type": "inline_equation", + "content": "\\pmb{r} \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 304, + 572, + 545, + 620 + ], + "type": "text", + "content": " to denote the corresponding rotation matrix to " + }, + { + "bbox": [ + 304, + 572, + 545, + 620 + ], + "type": "inline_equation", + "content": "\\pmb{q}" + }, + { + "bbox": [ + 304, + 572, + 545, + 620 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 621, + 545, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 621, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 304, + 621, + 545, + 645 + ], + "type": "text", + "content": "For rendering, the color " + }, + { + "bbox": [ + 304, + 621, + 545, + 645 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 621, + 545, + 645 + ], + "type": "text", + "content": " of a pixel is computed by blending all 3D Gaussians overlapping the pixel:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 369, + 651, + 545, + 684 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 651, + 545, + 684 + ], + "spans": [ + { + "bbox": [ + 369, + 651, + 545, + 684 + ], + "type": "interline_equation", + "content": "\\boldsymbol {C} = \\sum_ {i = 1} \\boldsymbol {c} _ {i} \\alpha_ {i} ^ {\\prime} \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\alpha_ {j} ^ {\\prime}\\right), \\tag {3}", + "image_path": "f74babba6f7180964487fbb8bfdb3292e2d9be7d707136dfdb40c7002ead4fc3.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": " is the color of each point modeled by 3-degree spherical harmonics. The blending weight " + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\alpha^{\\prime}" + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": " is given by" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20301" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "content": "evaluating the 2D projection of the 3D Gaussian multiplied by a per-point opacity " + }, + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "content": ". The Gaussian splats are sorted by depth before blending to respect visibility order." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 116, + 168, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 116, + 168, + 129 + ], + "spans": [ + { + "bbox": [ + 47, + 116, + 168, + 129 + ], + "type": "text", + "content": "3.2. 3D Gaussian Rigging" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 134, + 287, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 134, + 287, + 314 + ], + "spans": [ + { + "bbox": [ + 46, + 134, + 287, + 314 + ], + "type": "text", + "content": "The key component of our method is how we build connections between the FLAME [21] mesh and 3D Gaussian splats. Initially, we pair each triangle of the mesh with a 3D Gaussian and let the 3D Gaussian move with the triangle across time steps. In other words, the 3D Gaussian is static in the local space of its parent triangle but dynamic in the global (metric) space as the triangle moves. Given the vertices and edges of a triangle, we take the mean position " + }, + { + "bbox": [ + 46, + 134, + 287, + 314 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 134, + 287, + 314 + ], + "type": "text", + "content": " of the vertices as the origin of the local space. Then, we concatenate the direction vector of one of the edges, the normal vector of the triangle, and their cross product as column vectors to form a rotation matrix " + }, + { + "bbox": [ + 46, + 134, + 287, + 314 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 134, + 287, + 314 + ], + "type": "text", + "content": ", which describes the orientation of the triangle in the global space. We also compute a scalar " + }, + { + "bbox": [ + 46, + 134, + 287, + 314 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 134, + 287, + 314 + ], + "type": "text", + "content": " by the mean length of one of the edges and its perpendicular to describe the triangle scaling." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 314, + 287, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 314, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 46, + 314, + 287, + 386 + ], + "type": "text", + "content": "For the paired 3D Gaussian of a triangle, we define its location " + }, + { + "bbox": [ + 46, + 314, + 287, + 386 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 314, + 287, + 386 + ], + "type": "text", + "content": ", rotation " + }, + { + "bbox": [ + 46, + 314, + 287, + 386 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 314, + 287, + 386 + ], + "type": "text", + "content": ", and anisotropic scaling " + }, + { + "bbox": [ + 46, + 314, + 287, + 386 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 314, + 287, + 386 + ], + "type": "text", + "content": " all in the local space. We initialize the location " + }, + { + "bbox": [ + 46, + 314, + 287, + 386 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 314, + 287, + 386 + ], + "type": "text", + "content": " at the local origin, the rotation " + }, + { + "bbox": [ + 46, + 314, + 287, + 386 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 314, + 287, + 386 + ], + "type": "text", + "content": " as an identity rotation matrix, and the scaling " + }, + { + "bbox": [ + 46, + 314, + 287, + 386 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 314, + 287, + 386 + ], + "type": "text", + "content": " as a unit vector. At rendering time, we convert these properties into the global space by:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 396, + 287, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 287, + 407 + ], + "type": "interline_equation", + "content": "\\boldsymbol {r} ^ {\\prime} = \\boldsymbol {R} \\boldsymbol {r}, \\tag {4}", + "image_path": "2e18df2a6922feda132978a1b4f6e846a1f1b7381a955fcb0edd04cf8697ea5f.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 410, + 287, + 423 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 410, + 287, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 410, + 287, + 423 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\mu} ^ {\\prime} = k \\boldsymbol {R} \\boldsymbol {\\mu} + \\boldsymbol {T}, \\tag {5}", + "image_path": "8ce07886d57d26423f054ec575b773f5400d15d3b4fd6b2a8c5c84d71824d2e9.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 426, + 287, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 426, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 134, + 426, + 287, + 437 + ], + "type": "interline_equation", + "content": "\\boldsymbol {s} ^ {\\prime} = k \\boldsymbol {s}. \\tag {6}", + "image_path": "20e630466b65a0267b9fcf0a2bdf08e53c78771dc7945180cff0ceb201d5ea11.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 448, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 448, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 448, + 287, + 555 + ], + "type": "text", + "content": "We incorporate triangle scaling in Eqs. (5) and (6) so that the local position and scaling of a 3D Gaussian are defined relative to the absolute scale of a triangle. This enables an adaptive step size in the metric space with a constant learning rate for parameters defined in the local space. For example, a 3D Gaussian paired with a smaller triangle will move slower in an iteration step than those paired with large triangles. This also makes interpreting the parameters regarding the distance from the triangle center easier." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 563, + 164, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 164, + 575 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 164, + 575 + ], + "type": "text", + "content": "3.3. Binding Inheritance" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 582, + 287, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 666 + ], + "type": "text", + "content": "Only having the same numbers of Gaussian splats as the triangles is insufficient to capture details. For instance, representing a curved hair strand requires multiple splats, while a triangle on the scalp may intersect with several strands. Therefore, we also need the adaptive density control strategy [15], which adds and removes splats based on the view-space positional gradient and the opacity of each Gaussian." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "type": "text", + "content": "For each 3D Gaussian with a large view-space positional gradient, we split it into two smaller ones if it is large or clone it if it is small. We conduct this in the local space and ensure a newly created Gaussian is close to the old one that" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "triggers this densification operation. Then, it is reasonable to bind a new 3D Gaussian to the same triangle as the old one because it was created to enhance the fidelity of the local region. Therefore, each 3D Gaussian must carry one more parameter, the index of its parent triangle, to enable binding inheritance during densification." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 144, + 545, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 144, + 545, + 288 + ], + "spans": [ + { + "bbox": [ + 304, + 144, + 545, + 288 + ], + "type": "text", + "content": "Besides densification, we also use the pruning operation as a part of the adaptive density control strategy [15]. It periodically resets the opacity of all splats close to zero and removes points with opacity below a threshold. This technique is effective in suppressing floating artifacts, however, such pruning can also cause problems in a dynamic scene. For instance, regions of the face that are often occluded (such as eyeball triangles), can be overly sensitive to this pruning strategy, and often end up with few or no attached Gaussians. To prevent this, we keep track of the number of splats attached to each triangle, and ensure that every triangle always has at least one splat attached." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 296, + 484, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 296, + 484, + 308 + ], + "spans": [ + { + "bbox": [ + 306, + 296, + 484, + 308 + ], + "type": "text", + "content": "3.4. Optimization and Regularization" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 314, + 544, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 314, + 544, + 338 + ], + "spans": [ + { + "bbox": [ + 305, + 314, + 544, + 338 + ], + "type": "text", + "content": "We supervise the rendered images with a combination of " + }, + { + "bbox": [ + 305, + 314, + 544, + 338 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 305, + 314, + 544, + 338 + ], + "type": "text", + "content": " term and a D-SSIM term following [15]:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 362, + 348, + 545, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 348, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 362, + 348, + 545, + 361 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {r g b}} = (1 - \\lambda) \\mathcal {L} _ {1} + \\lambda \\mathcal {L} _ {\\mathrm {D} - \\mathrm {S S I M}}, \\tag {7}", + "image_path": "1e9784c53fcef62201a0a8d9cc7dc4d0557a550028dbca3675dc01288999b193.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 370, + 545, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 370, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 304, + 370, + 545, + 465 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 304, + 370, + 545, + 465 + ], + "type": "inline_equation", + "content": "\\lambda = 0.2" + }, + { + "bbox": [ + 304, + 370, + 545, + 465 + ], + "type": "text", + "content": ". This already results in good re-rendering quality without additional supervision, such as depth or silhouette supervision, thanks to the powerful tile rasterizer [15]. We found however, that if we try to animate these splats via FLAME [21] to novel expressions and poses, large spike- and blob-like artifacts appear wildly throughout the scene. This is due to a poor alignment between the Gaussian splats and the triangles." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 466, + 545, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 466, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 304, + 466, + 545, + 586 + ], + "type": "text", + "content": "Position loss with threshold. A basic hypothesis behind 3D Gaussian rigging is that the Gaussian splats should roughly match the underlying mesh. They should also match their locations; for instance a Gaussian representing a spot on the nose should not be rigged to a triangle on the cheek. Although our splats are initialized at triangle centers, and new splats are added nearby these existing ones, it is not guaranteed that the primitives remain close to their parent triangle after the optimization. To address this, we regularize the local position of each Gaussian by:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 359, + 595, + 545, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 595, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 359, + 595, + 545, + 609 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {p o s i t i o n}} = \\| \\max (\\boldsymbol {\\mu}, \\epsilon_ {\\text {p o s i t i o n}}) \\| _ {2}, \\tag {8}", + "image_path": "ff3c9e7c7af2ce82aed077415e5a612a8299be53415dc46e0ba4944feed576f8.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 617, + 544, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 544, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 544, + 641 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 617, + 544, + 641 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\mathrm{position}} = 1" + }, + { + "bbox": [ + 304, + 617, + 544, + 641 + ], + "type": "text", + "content": " is a threshold that tolerates small errors within the scaling of its parent triangle." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "Scaling loss with threshold. Aside from position, the scaling of 3D Gaussians is even more essential for the visual quality during animation. Specifically, if a 3D Gaussian is large in comparison to its parent triangle, small rotations of the triangle – barely noticeable at the scale of the triangle – will be magnified by the scale of the 3D Gaussian, resulting" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20302" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "text", + "content": "in unpleasant jittering artifacts. To mitigate this, we also regularize the local scale of each 3D Gaussian by:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 104, + 287, + 118 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 104, + 287, + 118 + ], + "spans": [ + { + "bbox": [ + 104, + 104, + 287, + 118 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s c a l i n g}} = \\left\\| \\max \\left(\\boldsymbol {s}, \\epsilon_ {\\text {s c a l i n g}}\\right) \\right\\| _ {2}, \\tag {9}", + "image_path": "1b409b0932777c366dac3377ee8c8bca5da7e3445738d3a630c8e6d2581f794b.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 124, + 287, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 124, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 46, + 124, + 287, + 194 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 124, + 287, + 194 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\mathrm{scaling}} = 0.6" + }, + { + "bbox": [ + 46, + 124, + 287, + 194 + ], + "type": "text", + "content": " is a threshold that disables this loss term when the scale of a Gaussian less than " + }, + { + "bbox": [ + 46, + 124, + 287, + 194 + ], + "type": "inline_equation", + "content": "0.6\\times" + }, + { + "bbox": [ + 46, + 124, + 287, + 194 + ], + "type": "text", + "content": " the scale of its parent triangle. This " + }, + { + "bbox": [ + 46, + 124, + 287, + 194 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\mathrm{scaling}}" + }, + { + "bbox": [ + 46, + 124, + 287, + 194 + ], + "type": "text", + "content": "-tolerance is indispensable; without it, the Gaussian splats shrink excessively, causing rendering speeds to deteriorate, as camera rays need to hit more splats before zero transmittance is reached." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 196, + 182, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 196, + 182, + 206 + ], + "spans": [ + { + "bbox": [ + 59, + 196, + 182, + 206 + ], + "type": "text", + "content": "Our final loss function is thus:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 215, + 287, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 215, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 69, + 215, + 287, + 228 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\mathrm {r g b}} + \\lambda_ {\\text {p o s i t i o n}} \\mathcal {L} _ {\\text {p o s i t i o n}} + \\lambda_ {\\text {s c a l i n g}} \\mathcal {L} _ {\\text {s c a l i n g}}, \\tag {10}", + "image_path": "eb403189780576f8f5c9c4d27a2943f0533a16da9f03b459b68833d14d8e1d6d.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 235, + 287, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 235, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 235, + 287, + 293 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 235, + 287, + 293 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{position}} = 0.01" + }, + { + "bbox": [ + 46, + 235, + 287, + 293 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 235, + 287, + 293 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{scaling}} = 1" + }, + { + "bbox": [ + 46, + 235, + 287, + 293 + ], + "type": "text", + "content": ". Note that we only apply " + }, + { + "bbox": [ + 46, + 235, + 287, + 293 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{position}}" + }, + { + "bbox": [ + 46, + 235, + 287, + 293 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 235, + 287, + 293 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{scaling}}" + }, + { + "bbox": [ + 46, + 235, + 287, + 293 + ], + "type": "text", + "content": " to visible splats. Thereby, we only regularize points when the color loss " + }, + { + "bbox": [ + 46, + 235, + 287, + 293 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rgb}}" + }, + { + "bbox": [ + 46, + 235, + 287, + 293 + ], + "type": "text", + "content": " is present. This helps to maintain the learned structure of often-occluded regions such as teeth and eyeballs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 295, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 295, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 295, + 287, + 510 + ], + "type": "text", + "content": "Implementation details. We use Adam [17] for parameter optimization (the same hyperparameter values are used across all subjects). We set the learning rate to 5e-3 for the position and 1.7e-2 for the scaling of 3D Gaussians and keep the same learning rates as 3D Gaussian Splatting [15] for the rest of the parameters. Alongside the Gaussian splat parameters, we also finetune the translation, joint rotation, and expression parameters of FLAME [21] for each timestep, using learning rates 1e-6, 1e-5, and 1e-3, respectively. We train for 600,000 iterations, and exponentially decay the learning rate for the splat positions until the final iteration, where it reaches " + }, + { + "bbox": [ + 46, + 295, + 287, + 510 + ], + "type": "inline_equation", + "content": "0.01 \\times" + }, + { + "bbox": [ + 46, + 295, + 287, + 510 + ], + "type": "text", + "content": " the initial value. We enable adaptive density control with binding inheritance every 2,000 iterations, from iteration 10,000 until the end. Every 60,000 iterations, we reset the Gaussians' opacities. We use a photo-metric head tracker to obtain the FLAME parameters, including shape " + }, + { + "bbox": [ + 46, + 295, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 295, + 287, + 510 + ], + "type": "text", + "content": ", translation " + }, + { + "bbox": [ + 46, + 295, + 287, + 510 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 295, + 287, + 510 + ], + "type": "text", + "content": ", pose " + }, + { + "bbox": [ + 46, + 295, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 295, + 287, + 510 + ], + "type": "text", + "content": ", expression " + }, + { + "bbox": [ + 46, + 295, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 46, + 295, + 287, + 510 + ], + "type": "text", + "content": ", and vertex offset " + }, + { + "bbox": [ + 46, + 295, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\Delta v" + }, + { + "bbox": [ + 46, + 295, + 287, + 510 + ], + "type": "text", + "content": " in the canonical space." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 520, + 127, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 520, + 127, + 533 + ], + "spans": [ + { + "bbox": [ + 47, + 520, + 127, + 533 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 540, + 97, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 540, + 97, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 540, + 97, + 552 + ], + "type": "text", + "content": "4.1. Setup" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 557, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 287, + 653 + ], + "type": "text", + "content": "Settings. We evaluate the quality of head avatars with three settings: 1) novel-view synthesis: driving an avatar with head poses and expressions from training sequences and rendering from a held-out viewpoint. 2) self-reenactment: driving an avatar with unseen poses and expressions from a held-out sequence of the same subject and rendering all 16 camera views. 3) cross-identity reenactment: animating an avatar with poses and expressions from another subject." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "Dataset. We conduct experiments on video recordings of 9 subjects from the NeRSample [18] dataset. All recordings contain 16 views covering the front and sides of a subject. For each subject, we take 11 video sequences and down-sample images to a resolution of " + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "802 \\times 550" + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": ". Participants" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 306, + 71, + 545, + 153 + ], + "blocks": [ + { + "bbox": [ + 306, + 71, + 545, + 153 + ], + "lines": [ + { + "bbox": [ + 306, + 71, + 545, + 153 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 545, + 153 + ], + "type": "table", + "html": "
Novel-View SynthesisSelf-Reenactment
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
AvatarMAV [44]29.50.9130.15224.30.8870.168
PointAvatar [50]25.80.8930.09723.40.8840.102
INSTA [52]26.70.8990.12226.30.9060.110
Ours31.60.9380.06526.00.9100.076
", + "image_path": "c3e9ad8c45420afd947ff7949c58cdbaee4b09780c0ba91e78b72a51bd64c5fa.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 162, + 544, + 186 + ], + "lines": [ + { + "bbox": [ + 306, + 162, + 544, + 186 + ], + "spans": [ + { + "bbox": [ + 306, + 162, + 544, + 186 + ], + "type": "text", + "content": "Table 1. Quantitative comparison with state-of-the-art methods. Green indicates the best and yellow indicates the second." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 209, + 545, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 209, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 305, + 209, + 545, + 316 + ], + "type": "text", + "content": "were instructed to perform specific expressions or emotions in 10 sequences and freely perform in the last one. To conduct quantitative evaluation for each subject, we train our method and three baselines (INSTA [52], PointAvatar [50], and AvatarMAV [44]) using 9 out of the 10 prescribed sequences and 15 out of 16 available cameras. We use the " + }, + { + "bbox": [ + 305, + 209, + 545, + 316 + ], + "type": "inline_equation", + "content": "11^{\\mathrm{th}}" + }, + { + "bbox": [ + 305, + 209, + 545, + 316 + ], + "type": "text", + "content": " (free-performance) sequence to visually assess cross-identity reenactment capabilities. Please refer to the supplementary material for details of the dataset and baselines." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 324, + 535, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 324, + 535, + 335 + ], + "spans": [ + { + "bbox": [ + 305, + 324, + 535, + 335 + ], + "type": "text", + "content": "4.2. Head Avatar Reconstruction and Animation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 342, + 545, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 545, + 426 + ], + "type": "text", + "content": "We evaluate the reconstruction quality of avatars by novel-view synthesis and animation fidelity by self-reenactment. Fig. 3 shows qualitative comparisons. For novel-view synthesis, all methods produce plausible rendering results. A close inspection of PointAvatar's [50] results show dotted artifacts, owing its fixed point size. In our case, the anisotropic scaling of 3D Gaussians alleviates this issue." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 426, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 557 + ], + "type": "text", + "content": "INSTA [52] shows clean results for the face, however regions around the neck and shoulder can be noisy, as the tracked FLAME meshes are often misaligned in those regions. This causes issues for INSTA, as its warping process is based on the nearest triangle. In the case of our method, each Gaussian splat is rigged to a consistent triangle, regardless of the pose or expression. When the tracked mesh is inaccurate, the positional gradient to a Gaussian splat can consistently back-propagate to the same triangle. This enables misalignments due to incorrect FLAME tracking to be corrected while the 3D Gaussians are being optimized." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 558, + 546, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 546, + 712 + ], + "type": "text", + "content": "AvatarMAV [44] shows comparable quality to other methods in novel-view synthesis but struggles with novel-expression synthesis. This is because it only uses the expression vector of a 3DMM as conditioning. Since the control from the expression vectors to the deformation bases must be learned, it struggles to reproduce expressions that are far from the training distribution. Similar conclusions can be drawn from the quantitative comparison shown in Tab. 1. Our approach outperforms others by a large margin regarding metrics for novel-view synthesis. Our method also stands out in self-reenactment, with significantly lower perceptual differences in terms of LPIPS [48]. Note that self-reenactment is based on tracked FLAME meshes that" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20303" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 547, + 635 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 547, + 635 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 547, + 635 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 547, + 635 + ], + "type": "image", + "image_path": "63b0543c00372643cd9905b20c1c1dc1ce384518bff6d979432871d41c22c233.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 641, + 547, + 677 + ], + "lines": [ + { + "bbox": [ + 46, + 641, + 547, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 547, + 677 + ], + "type": "text", + "content": "Figure 3. Qualitative comparison on novel-view synthesis and self-reenactment of head avatars. Our method outperforms state-of-the-art methods by producing significantly sharper rendering outputs. We obtain precise reconstruction of details such as reflective light on eyes, hair strands, teeth, etc. Our results for self-reenactment show more accurate expressions compare to baselines." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20304" + } + ] + } + ], + "index": 2 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 73, + 545, + 435 + ], + "blocks": [ + { + "bbox": [ + 52, + 73, + 545, + 435 + ], + "lines": [ + { + "bbox": [ + 52, + 73, + 545, + 435 + ], + "spans": [ + { + "bbox": [ + 52, + 73, + 545, + 435 + ], + "type": "image", + "image_path": "8d168d18932ce5a214ea0236828bf75ca52660c73dfc0e91ae22a35af2241f43.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 442, + 547, + 477 + ], + "lines": [ + { + "bbox": [ + 46, + 442, + 547, + 477 + ], + "spans": [ + { + "bbox": [ + 46, + 442, + 547, + 477 + ], + "type": "text", + "content": "Figure 4. Cross-identity reenactment of head avatars. We use the tracked FLAME expression and pose parameters of source actors to drive the reconstructed avatars. Our method produces high-quality rendering and transfers expressions vividly, while baseline methods suffer from artifacts and generalize poorly to novel expressions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 497, + 287, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 497, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 46, + 497, + 287, + 532 + ], + "type": "text", + "content": "may not perfectly align with the target images, thus bringing disadvantages to our results with more visual details regarding pixel-wise metrics such as PSNR." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 533, + 289, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 289, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 289, + 665 + ], + "type": "text", + "content": "For a real-world test of avatar animation, we conduct experiments on cross-identity reenactment in Fig. 4. Our avatars accurately reproduce eye blinks and mouth movements from source actors showing lively, complex dynamics such as wrinkles. INSTA [52] suffers from aliasing artifacts when the avatars move beyond the occupancy grid of I-NGP [29] optimized for training sequences. The movement of results from PointAvatar [50] is not precise because its deformation space is not guaranteed to be consistent with FLAME. AvatarMAV [44] exhibits large degradations in reenactment due to a lack of deformation priors." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 671, + 141, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 141, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 141, + 684 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": "To validate the effectiveness of our method components, we deactivate each of them and report results in Tab. 2." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 306, + 495, + 379, + 605 + ], + "blocks": [ + { + "bbox": [ + 306, + 495, + 379, + 605 + ], + "lines": [ + { + "bbox": [ + 306, + 495, + 379, + 605 + ], + "spans": [ + { + "bbox": [ + 306, + 495, + 379, + 605 + ], + "type": "image", + "image_path": "d66860fcbeff6c9a795e3ddf45bc43e2c4b25c44f93ae5d501d4a4f498d9040d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 614, + 547, + 648 + ], + "lines": [ + { + "bbox": [ + 305, + 614, + 547, + 648 + ], + "spans": [ + { + "bbox": [ + 305, + 614, + 547, + 648 + ], + "type": "text", + "content": "Figure 5. Fine-tuning FLAME parameters leads to better alignment of the mesh to the input image. In the example above, the movements of cheeks and lips are better captured after fine-tuning." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 389, + 495, + 462, + 606 + ], + "blocks": [ + { + "bbox": [ + 389, + 495, + 462, + 606 + ], + "lines": [ + { + "bbox": [ + 389, + 495, + 462, + 606 + ], + "spans": [ + { + "bbox": [ + 389, + 495, + 462, + 606 + ], + "type": "image", + "image_path": "c7e03eab471cd10db96f4e471599724263803167295ac80cb4664fa1966b2009.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 472, + 495, + 545, + 606 + ], + "blocks": [ + { + "bbox": [ + 472, + 495, + 545, + 606 + ], + "lines": [ + { + "bbox": [ + 472, + 495, + 545, + 606 + ], + "spans": [ + { + "bbox": [ + 472, + 495, + 545, + 606 + ], + "type": "image", + "image_path": "062dfa4389a1cdd500091b7c605f119e83a99bd3c1bc38b645d913a485d5074b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "content": "Adaptive density control with binding inheritance. Without binding inheritance, we lose the ability to add and remove Gaussian splats besides the initial ones. With a lim" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20305" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 70, + 282, + 186 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 282, + 186 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 282, + 186 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 282, + 186 + ], + "type": "table", + "html": "
Novel-ViewSelf-Reenactment
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Ours28.80.8830.09825.10.8530.101
w/o ADC26.80.8540.20625.10.8600.183
w/o Lscaling28.00.8770.11424.90.8520.109
w/o εscaling25.00.8330.19524.10.8430.176
w/o Lposition29.70.8940.09124.90.8510.096
w/o εposition28.70.8820.10525.00.8550.106
w/o FLAME ft.26.10.8550.13125.50.8620.124
", + "image_path": "ba7ca60a662282e1923c158487d90e14d05c283010b1bfc568485a789c2fbbd5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 196, + 287, + 245 + ], + "lines": [ + { + "bbox": [ + 46, + 196, + 287, + 245 + ], + "spans": [ + { + "bbox": [ + 46, + 196, + 287, + 245 + ], + "type": "text", + "content": "Table 2. Ablation study on subject #304. Green indicates the best and yellow indicates the second. “ADC” refers to Adaptive Density Control with binding inheritance. ‘FLAME ft.’ refers to FLAME parameter fine-tuning." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 264, + 287, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 264, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 264, + 287, + 300 + ], + "type": "text", + "content": "ited number of splats, the scaling of each splat increases to occupy the same space, causing blurry renderings, leading to huge fidelity loss, as shown in the second row of Tab. 2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 300, + 287, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 287, + 396 + ], + "type": "text", + "content": "Regularization on the scaling of Gaussian splats. Without the scaling loss, artifacts show up as spikes (the middle column in Fig. 6), and slightly worsen the image quality (the third row in Tab. 2). But when we use the scaling loss without an error tolerance, all metrics drastically deteriorate (the fourth row in Tab. 2) because all the splats are regularized to be infinitely small, making it hard to construct opaque surfaces." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 396, + 287, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 287, + 563 + ], + "type": "text", + "content": "Regularization on the position of Gaussian splats. When disabling the threshold for the position loss, we get slightly worse metrics (the sixth row in Tab. 2). But when turning off the position loss, the metrics on novel-view synthesis become the best in the table (the fifth row in Tab. 2). This is reasonable because without constraints to the mesh Gaussian splats can freely move in the space to achieve minimal re-rendering error. However, since the distribution of Gaussian splats are overfitted to training frames, the avatar shows artifacts such as cracks and fly-around blobs (the left column in Fig. 6) in novel expressions and poses. Therefore, the position loss is necessary to enforce a conservative deviation of splats from the mesh surface so that we can animate the reconstructed avatar with unseen facial motion." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 563, + 287, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 563, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 563, + 287, + 647 + ], + "type": "text", + "content": "FLAME fine-tuning. Thanks to our consistent binding between Gaussian splats and triangles, we can effectively fine-tune FLAME parameters while optimizing Gaussian splats (Fig. 5). As shown in the last row in Tab. 2, turning off the fine-tuning negatively affects image quality for novel-view synthesis. For self-reenactment, fine-tuning FLAME parameters also leads to lower perceptual error." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 657, + 285, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 657, + 285, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 657, + 285, + 671 + ], + "type": "text", + "content": "5. Limitations and Potential Negative Impacts" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Our approach, based on 3D Gaussian splats, directly captures the radiance field without decoupling material and lighting. Hence, relighting the avatar is not feasible within" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 545, + 195 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 545, + 195 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 545, + 195 + ], + "type": "image", + "image_path": "3ae5dfe4b1dbc53e16309a6ca9b831392ec6b77995354bd7e0b1b12f65ccc3ef.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 202, + 545, + 225 + ], + "lines": [ + { + "bbox": [ + 305, + 202, + 545, + 225 + ], + "spans": [ + { + "bbox": [ + 305, + 202, + 545, + 225 + ], + "type": "text", + "content": "Figure 6. The position loss and the scaling loss helps prevent artifacts during animation with novel expressions and poses." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 245, + 545, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 245, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 304, + 245, + 545, + 317 + ], + "type": "text", + "content": "our current approach. Additionally, we lack control over areas that are not modeled by FLAME such as hair and other accessories. Here, we believe that modeling these parts directly would provide an interesting future research avenue; for instance, recent methods show promising result in reconstructing hair [35, 41]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 318, + 546, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 318, + 546, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 318, + 546, + 449 + ], + "type": "text", + "content": "The creation of photorealistic avatars may introduce a wide spectrum of malevolent uses. Privacy violations are a significant concern, with potential for unauthorized manipulation of people's likeness. Moreover the creation of deceptive content, like deepfake videos, can lead to misinformation, defamation and harm to reputations. Identity theft is an additional risk, as avatars may be exploited for impersonation and fraudulent activities. We strongly condemn unauthorized and malevolent uses of this technology, emphasizing the need for ethical considerations in all applications utilizing our method." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 460, + 378, + 471 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 460, + 378, + 471 + ], + "spans": [ + { + "bbox": [ + 306, + 460, + 378, + 471 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 479, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 545, + 624 + ], + "type": "text", + "content": "GaussianAvatars is a novel approach that creates photorealistic avatars from video sequences. It features a dynamic 3D representation based on 3D Gaussian splats that are rigged to a parametric morphable face model. This enables flexible control and precise attribute transfer of a dynamic, photorealistic avatar. The set of Gaussian splats can deviate from the mesh surface to compensate the absence or inaccuracy of the morphable model, exhibiting astonishing ability to model fine details on human heads. Our approach outperforms state-of-the-art methods on image quality and expression accuracy for a large margin, indicating a large potential for more applications in related domains." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 634, + 408, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 634, + 408, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 634, + 408, + 647 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "This work was supported by Toyota Motor Europe and Woven by Toyota. This work was also supported by the ERC Starting Grant Scan2CAD (804724). We thank Justus Thies, Andrei Burov, and Lei Li for constructive discussions and Angela Dai for the video voice-over." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20306" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "text", + "content": "[1] Benjamin Attal, Jia-Bin Huang, Christian Richardt, Michael Zollhoefer, Johannes Kopf, Matthew O'Toole, and Changil Kim. Hyperreel: High-fidelity 6-dof video with ray-conditioned sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16610-16620, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 159, + 288, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 159, + 288, + 224 + ], + "spans": [ + { + "bbox": [ + 53, + 159, + 288, + 224 + ], + "type": "text", + "content": "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 226, + 288, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 226, + 288, + 270 + ], + "spans": [ + { + "bbox": [ + 53, + 226, + 288, + 270 + ], + "type": "text", + "content": "[3] Volker Blanz and Thomas Vetter. Face recognition based on fitting a 3d morphable model. IEEE Transactions on pattern analysis and machine intelligence, 25(9):1063-1074, 2003. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 272, + 288, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 272, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 272, + 288, + 316 + ], + "type": "text", + "content": "[4] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 130-141, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 318, + 288, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 318, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 53, + 318, + 288, + 361 + ], + "type": "text", + "content": "[5] Caroline Chan, Shiry Ginosar, Tinghui Zhou, and Alexei A Efros. Everybody dance now. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5933-5942, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 363, + 288, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 363, + 288, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 363, + 288, + 407 + ], + "type": "text", + "content": "[6] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 409, + 288, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 409, + 288, + 464 + ], + "spans": [ + { + "bbox": [ + 53, + 409, + 288, + 464 + ], + "type": "text", + "content": "[7] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbaek Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12479-12488, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 465, + 288, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 465, + 288, + 520 + ], + "spans": [ + { + "bbox": [ + 53, + 465, + 288, + 520 + ], + "type": "text", + "content": "[8] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 522, + 288, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 522, + 288, + 566 + ], + "spans": [ + { + "bbox": [ + 53, + 522, + 288, + 566 + ], + "type": "text", + "content": "[9] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5712-5721, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 567, + 288, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 288, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 288, + 622 + ], + "type": "text", + "content": "[10] Xuan Gao, Chenglai Zhong, Jun Xiang, Yang Hong, Yudong Guo, and Juyong Zhang. Reconstructing personalized semantic facial nerf models from monocular video. ACM Transactions on Graphics (Proceedings of SIGGRAPH Asia), 41(6), 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 624, + 288, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 288, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 288, + 678 + ], + "type": "text", + "content": "[11] Philip-William Grassal, Malte Prinzler, Titus Leistner, Carsten Rother, Matthias Nießner, and Justus Thies. Neural head avatars from monocular rgb videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18653–18664, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "text", + "content": "[12] Yang Hong, Bo Peng, Haiyao Xiao, Ligang Liu, and Juyong Zhang. Headnerf: A real-time nerf-based parametric head model. In Proceedings of the IEEE/CVF Conference" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "on Computer Vision and Pattern Recognition, pages 20374-20384, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 547, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 150 + ], + "type": "text", + "content": "[13] Mustafa Işik, Martin Rünz, Markos Georgopoulos, Taras Khakhulin, Jonathan Starck, Lourdes Agapito, and Matthias Nießner. Humanrf: High-fidelity neural radiance fields for humans in motion. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 152, + 547, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 547, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 547, + 206 + ], + "type": "text", + "content": "[14] Rohit Jena, Ganesh Subramanian Iyer, Siddharth Choudhary, Brandon Smith, Pratik Chaudhari, and James Gee. Splatarmor: Articulated gaussian splattering for animatable humans from monocular rgb videos. arXiv preprint arXiv:2311.10812, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 209, + 547, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 547, + 252 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 547, + 252 + ], + "type": "text", + "content": "[15] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 42(4):1-14, 2023. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 254, + 547, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 254, + 547, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 254, + 547, + 308 + ], + "type": "text", + "content": "[16] Hyeongwoo Kim, Pablo Garrido, Ayush Tewari, Weipeng Xu, Justus Thies, Matthias Niessner, Patrick Pérez, Christian Richardt, Michael Zollhöfer, and Christian Theobalt. Deep video portraits. ACM transactions on graphics (TOG), 37(4): 1-14, 2018. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 309, + 547, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 547, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 547, + 342 + ], + "type": "text", + "content": "[17] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 343, + 547, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 547, + 387 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 547, + 387 + ], + "type": "text", + "content": "[18] Tobias Kirschstein, Shenhan Qian, Simon Giebenhain, Tim Walter, and Matthias Nießner. Nersemble: Multi-view radiance field reconstruction of human heads. ACM Trans. Graph., 42(4), 2023. 1, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 388, + 547, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 547, + 422 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 547, + 422 + ], + "type": "text", + "content": "[19] Muhammed Kocabas, Jen-Hao Rick Chang, James Gabriel, Oncel Tuzel, and Anurag Ranjan. Hugs: Human gaussian splats. arXiv preprint arXiv:2311.17910, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 423, + 547, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 423, + 547, + 455 + ], + "spans": [ + { + "bbox": [ + 307, + 423, + 547, + 455 + ], + "type": "text", + "content": "[20] Jiahui Lei, Yufu Wang, Georgios Pavlakos, Lingjie Liu, and Kostas Daniilidis. Gart: Gaussian articulated template models. arXiv preprint arXiv:2311.16099, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 456, + 547, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 547, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 547, + 510 + ], + "type": "text", + "content": "[21] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6):194:1-194:17, 2017. 2, 3, 4, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 513, + 547, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 513, + 547, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 513, + 547, + 578 + ], + "type": "text", + "content": "[22] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 580, + 547, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 580, + 547, + 613 + ], + "spans": [ + { + "bbox": [ + 307, + 580, + 547, + 613 + ], + "type": "text", + "content": "[23] Zhe Li, Zerong Zheng, Lizhen Wang, and Yebin Liu. Animatable gaussians: Learning pose-dependent gaussian maps for high-fidelity human avatar modeling. arXiv, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 614, + 547, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 614, + 547, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 614, + 547, + 667 + ], + "type": "text", + "content": "[24] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. ACM transactions on graphics (TOG), 40(6):1-16, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 670, + 547, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 547, + 714 + ], + "type": "text", + "content": "[25] Stephen Lombardi, Tomas Simon, Gabriel Schwartz, Michael Zollhoefer, Yaser Sheikh, and Jason Saragih. Mixture of volumetric primitives for efficient neural rendering. ACM Transactions on Graphics (ToG), 40(4):1-13, 2021. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20307" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[26] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. In *Seminal Graphics Papers: Pushing the Boundaries*, Volume 2, pages 851–866. 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 160 + ], + "type": "text", + "content": "[27] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 162, + 287, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 162, + 287, + 215 + ], + "spans": [ + { + "bbox": [ + 49, + 162, + 287, + 215 + ], + "type": "text", + "content": "[28] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 217, + 287, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 217, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 49, + 217, + 287, + 259 + ], + "type": "text", + "content": "[29] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 1, 2, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 262, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 262, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 49, + 262, + 287, + 316 + ], + "type": "text", + "content": "[30] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 317, + 287, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 317, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 49, + 317, + 287, + 370 + ], + "type": "text", + "content": "[31] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 372, + 287, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 372, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 49, + 372, + 287, + 426 + ], + "type": "text", + "content": "[32] Pascal Paysan, Reinhard Knothe, Brian Amberg, Sami Romdhani, and Thomas Vetter. A 3d face model for pose and illumination invariant face recognition. In 2009 sixth IEEE international conference on advanced video and signal based surveillance, pages 296-301. IEEE, 2009. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 427, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 427, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 49, + 427, + 287, + 480 + ], + "type": "text", + "content": "[33] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Animatable neural radiance fields for modeling dynamic human bodies. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14314-14323, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 482, + 287, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 482, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 49, + 482, + 287, + 535 + ], + "type": "text", + "content": "[34] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 537, + 287, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 537, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 49, + 537, + 287, + 591 + ], + "type": "text", + "content": "[35] Radu Alexandru Rosu, Shunsuke Saito, Ziyan Wang, Chenglei Wu, Sven Behnke, and Giljoo Nam. Neural strands: Learning hair geometry and appearance from multi-view images. In European Conference on Computer Vision, pages 73-89. Springer, 2022. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 593, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 593, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 49, + 593, + 287, + 656 + ], + "type": "text", + "content": "[36] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerfplayer: A streamable dynamic scene representation with decomposed neural radiance fields. IEEE Transactions on Visualization and Computer Graphics, 29(5):2732-2742, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 658, + 287, + 712 + ], + "type": "text", + "content": "[37] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[38] Supasorn Suwajanakorn, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Synthesizing obama: learning lip sync from audio. ACM Transactions on Graphics (ToG), 36(4):1-13, 2017. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 118, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 545, + 171 + ], + "type": "text", + "content": "[39] Justus Thies, Michael Zollhofer, Marc Stamminger, Christian Theobalt, and Matthias Nießner. Face2face: Real-time face capture and reenactment of rgb videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2387-2395, 2016. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 173, + 545, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 173, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 308, + 173, + 545, + 215 + ], + "type": "text", + "content": "[40] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: Image synthesis using neural textures. Acm Transactions on Graphics (TOG), 38(4):1-12, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 217, + 545, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 281 + ], + "type": "text", + "content": "[41] Ziyan Wang, Giljoo Nam, Tuur Stuyck, Stephen Lombardi, Chen Cao, Jason Saragih, Michael Zollhöfer, Jessica Hodgins, and Christoph Lassner. Neuwigs: A neural dynamic model for volumetric hair capture and animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8641-8651, 2023. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 283, + 545, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 283, + 545, + 336 + ], + "spans": [ + { + "bbox": [ + 308, + 283, + 545, + 336 + ], + "type": "text", + "content": "[42] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9421-9431, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 338, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 545, + 392 + ], + "type": "text", + "content": "[43] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5438–5448, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 394, + 545, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 394, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 308, + 394, + 545, + 436 + ], + "type": "text", + "content": "[44] Yuelang Xu, Lizhen Wang, Xiaochen Zhao, Hongwen Zhang, and Yebin Liu. Avatarmav: Fast 3d head avatar reconstruction using motion-aware neural voxels. In ACM SIGGRAPH 2023 Conference Proceedings, 2023. 2, 5, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 437, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 545, + 491 + ], + "type": "text", + "content": "[45] Yuelang Xu, Hongwen Zhang, Lizhen Wang, Xiaochen Zhao, Huang Han, Qi Guojun, and Yebin Liu. Latentavatar: Learning latent expression code for expressive neural head avatar. In ACM SIGGRAPH 2023 Conference Proceedings, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 493, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 493, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 493, + 545, + 525 + ], + "type": "text", + "content": "[46] Keyang Ye, Tianjia Shao, and Kun Zhou. Animatable 3d gaussians for high-fidelity synthesis of human motions. arXiv preprint arXiv:2311.13404, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 526, + 545, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 526, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 308, + 526, + 545, + 579 + ], + "type": "text", + "content": "[47] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4578-4587, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 581, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 581, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 581, + 545, + 635 + ], + "type": "text", + "content": "[48] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 636, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 690 + ], + "type": "text", + "content": "[49] Yufeng Zheng, Victoria Fernández Abrevaya, Marcel C Bühler, Xu Chen, Michael J Black, and Otmar Hilliges. Im avatar: Implicit morphable head avatars from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13545-13555, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "text", + "content": "[50] Yufeng Zheng, Wang Yifan, Gordon Wetzstein, Michael J Black, and Otmar Hilliges. Pointavatar: Deformable point" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "20308" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 185 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "based head avatars from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21057-21067, 2023. 2, 5, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "type": "text", + "content": "[51] Wojciech Zielonka, Timur Bagautdinov, Shunsuke Saito, Michael Zollhöfer, Justus Thies, and Javier Romero. Drivable 3d gaussian avatars. 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 185 + ], + "type": "text", + "content": "[52] Wojciech Zielonka, Timo Bolkart, and Justus Thies. Instant volumetric head avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4574-4584, 2023. 2, 5, 7" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20309" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/GaussianDreamer_ Fast Generation from Text to 3D Gaussians by Bridging 2D and 3D Diffusion Models/68c13cd6-fa8d-4333-a8a0-12d60a7c6cac_content_list.json b/2024/GaussianDreamer_ Fast Generation from Text to 3D Gaussians by Bridging 2D and 3D Diffusion Models/68c13cd6-fa8d-4333-a8a0-12d60a7c6cac_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bdfdf8fbf4d564ed8eeafaefbd04e2d4b2c6f1a8 --- /dev/null +++ b/2024/GaussianDreamer_ Fast Generation from Text to 3D Gaussians by Bridging 2D and 3D Diffusion Models/68c13cd6-fa8d-4333-a8a0-12d60a7c6cac_content_list.json @@ -0,0 +1,2206 @@ +[ + { + "type": "text", + "text": "GaussianDreamer: Fast Generation from Text to 3D Gaussians by Bridging 2D and 3D Diffusion Models", + "text_level": 1, + "bbox": [ + 166, + 130, + 803, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Taoran Yi $^{1}$ , Jiemin Fang $^{2\\dagger}$ , Junjie Wang $^{2}$ , Guanjun Wu $^{3}$ , Lingxi Xie $^{2}$ , Xiaopeng Zhang $^{2}$ , Wenyu Liu $^{1}$ , Qi Tian $^{2}$ , Xinggang Wang $^{1\\dagger}$", + "bbox": [ + 196, + 202, + 772, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of EIC, Huazhong University of Science and Technology $^{2}$ Huawei Inc. $^{3}$ School of CS, Huazhong University of Science and Technology", + "bbox": [ + 163, + 238, + 803, + 275 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{taoranyi, guajuwu, liuwy, xgwang}@hust.edu.cn", + "bbox": [ + 281, + 277, + 687, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{jaminfong, is.wangjunjie, 198808xc, zxphistory}@gmail.com tian.qil@huawei.com", + "bbox": [ + 133, + 294, + 831, + 309 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b43f1cb133f73d9bb4060ed0df294a63043f74bfdc32fe9e00b27168bc0e64fe.jpg", + "image_caption": [ + "Figure 1. We propose a simple yet efficient framework called GaussianDreamer. It bridges the 3D and 2D diffusion models via Gaussian splatting, having both 3D consistency and rich generation details. Our method can complete training within 15 minutes on a single GPU and achieve real-time rendering." + ], + "image_footnote": [], + "bbox": [ + 76, + 311, + 248, + 494 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5f20021c96402161f3769d9d4f9e5b399c37d00b7500c4bfbfbbb693e9b49e4e.jpg", + "image_caption": [ + "3D Gaussians with Temporal Variation" + ], + "image_footnote": [], + "bbox": [ + 253, + 311, + 892, + 484 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 544, + 313, + 559 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent times, the generation of 3D assets from text prompts has shown impressive results. Both 2D and 3D diffusion models can help generate decent 3D objects based on prompts. 3D diffusion models have good 3D consistency, but their quality and generalization are limited as trainable 3D data is expensive and hard to obtain. 2D diffusion models enjoy strong abilities of generalization and fine generation, but 3D consistency is hard to guarantee. This paper attempts to bridge the power from the two types of diffusion models via the recent explicit and efficient 3D Gaussian splatting representation. A fast 3D object generation framework, named as GaussianDreamer, is proposed, where the 3D diffusion model provides priors for initialization and the 2D diffusion model enriches the geometry and appearance. Operations of noisy point growing and color perturbation are introduced to enhance the initialized Gaussians. Our GaussianDreamer can generate a high-quality 3D instance or 3D avatar within 15 minutes on one GPU, much faster than previous methods, while the gener", + "bbox": [ + 73, + 575, + 472, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ated instances can be directly rendered in real time. Demos and code are available at https://taoranyi.com/gaussiandreamer/.", + "bbox": [ + 500, + 545, + 892, + 590 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 618, + 632, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D asset generation has been an expensive and professional work in conventional pipelines. Recently, diffusion models [58] have achieved great success in creating high-quality and realistic 2D images. Many research works [2, 6, 9, 12, 15, 24, 32, 48, 52, 67, 70, 76, 77, 81, 90] try to transfer the power of 2D diffusion models to the 3D field for easing and assisting the process of 3D assets creation, e.g. the most common text-to-3D task.", + "bbox": [ + 496, + 643, + 890, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Here come two main streams for achieving this goal: (i) training a new diffusion model with 3D data [12, 15, 24, 48] (namely the 3D diffusion model) and (ii) lifting the 2D diffusion model to 3D [2, 6, 9, 32, 52, 58, 67, 70, 76, 77, 81, 90]. The former one is direct to implement and holds strong 3D consistency, but struggles to extend into a large generation domain as 3D data is usually hard and expensive to obtain. The scale of current 3D datasets is far smaller than 2D datasets. This results in the generated 3D assets falling", + "bbox": [ + 496, + 763, + 892, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ Project lead.", + "bbox": [ + 94, + 875, + 171, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\ddagger$ Corresponding author.", + "bbox": [ + 94, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6796", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "short in dealing with complex text prompts and producing complex/fine geometry and appearance. The latter benefits from the large data domain of the 2D diffusion models, which can handle various text prompts and produce highly detailed and complex geometry and appearance. However, as 2D diffusion models are unaware of the camera view, the generated 3D assets are hard to form geometry consistency, especially for structure-complicated instances.", + "bbox": [ + 75, + 90, + 472, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This paper proposes to use recent 3D Gaussian Splating [25] to bridge the two aforementioned approaches, simultaneously having the geometry consistency from 3D diffusion models and rich details from 2D diffusion models. 3D Gaussians are one type of efficient and explicit representation, which intrinsically enjoys geometry priors due to the point-cloud-like structure. Specifically, we use one of two types of 3D diffusion models: text-to-3D and text-to-motion diffusion models, e.g. Shap-E [24] and MDM [72] in our implementation, to generate a coarse 3D instance. Based on the coarse 3D instance, a group of 3D Gaussians are initialized. We introduce two operations of noisy point growing and color perturbation to supplement the initialized Gaussians for follow-up enriching the 3D instance. Then the 3D Gaussians can be improved and optimized by interacting with the 2D diffusion model via the Score Distillation Sampling [52] (SDS) loss. Due to the geometry priors from both the 3D diffusion model and 3D Gaussian Splating itself, the training process can be finished in a very short time. The generated 3D asset can be rendered in real time without transformation into structures like mesh via the splatting process.", + "bbox": [ + 75, + 212, + 472, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows.", + "bbox": [ + 96, + 544, + 419, + 558 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a text-to-3D method, named as Gaussian-Dreamer which bridges the 3D and 2D diffusion models via Gaussian splitting, enjoying both 3D consistency and rich generation details.", + "- Noisy point growing and color perturbation are introduced to supplement the initialized 3D Gaussians for further content enrichment.", + "- The overall method is simple and quite effective. A 3D instance can be generated within 15 minutes on one GPU, much faster than previous methods, and can be directly rendered in real time." + ], + "bbox": [ + 76, + 560, + 468, + 724 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 76, + 739, + 225, + 753 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Pretrained Diffusion Models. Recently, text-to-3D asset generation using diffusion models has achieved great success. Currently, it is mainly divided into lifting 2D diffusion models to 3D and 3D pretrained diffusion models, the difference lies in whether the training data used is 2D or 3D. 3D pretrained diffusion models [12, 15, 24, 48], referred to as 3D diffusion models in our paper, are models pretrained on text-3D pairs. After pretraining, they can generate 3D assets through only inference, and mod", + "bbox": [ + 75, + 763, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "els such as Point-E [48] and Shape-E [24] can generate 3D assets in minutes. In addition to generating 3D assets from text, there are methods where 3D diffusion models [1, 7, 10, 26, 55, 63, 72, 88, 89, 91] generate motion sequences based on text-motion data. By pretraining on text-motion pairs, these models can generate reasonable motion sequences for different texts. The generated motion sequences can be transformed into the SMPL (Skinned Multi-Person Linear) model [39] based on mesh representation, but texture information is not included. In our method, we can paint transformed SMPL by using different text prompts.", + "bbox": [ + 496, + 90, + 893, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Lifting 2D Diffusion Models to 3D. In text-to-3D asset generation methods [8, 16, 21, 23, 40, 49, 50, 53, 57, 61, 62, 68, 73, 80, 85, 87, 92], in addition to using 3D pretrained diffusion models, lifting a 2D diffusion model to 3D is a training-free approach. Moreover, due to the abundance of 2D image data, this method produces assets with higher diversity and fidelity. Some single-image-to-3D methods [13, 19, 30, 33–36, 38, 54, 60, 65, 66, 71, 78, 79, 82, 83] also employ similar ideas. DreamFusion [52] first proposes SDS (Score Distillation Sampling) method, which is to update the 3D representation model using the 2D diffusion model. [76] proposes the method of SJC (Score Jacobian Chaining) to lift the 2D diffusion model to 3D. Later methods [6, 31, 32, 69, 77] build on DreamFusion and further improve the quality of 3D generation. Among them, generated 3D assets may suffer from multi-face problems. To address this issue, some methods strengthen the semantics of different views [2] and use multi-view information [67, 90] to alleviate such problems. There are also models [18, 22, 29, 44, 46, 59, 75, 81] which adopt CLIP [56] to align each view of the 3D representation model with the text.", + "bbox": [ + 496, + 291, + 893, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Representation Methods. In recent times, neural radiance fields (NeRF) [45] have achieved impressive results in 3D representation, and many methods in text-to-3D asset generation have also adopted NeRF or its variants [3, 47] as the representation method. Some methods [6, 12, 31, 32] use explicit optimization mesh representation methods like DMTET [64] to reduce rendering costs and further improve resolution. In addition to that, there are also generation methods that utilize point clouds [41, 48, 53, 74] and meshes [37] as 3D representations.", + "bbox": [ + 496, + 643, + 893, + 794 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, 3D Gaussian Splatting [25] has been introduced as a representation method for 3D scenes, which can achieve rendering effects comparable to NeRF-based methods and enable real-time rendering. Two concurrent works [9, 70] also construct the 3D representation with 3D Gaussian Splatting [25]. DreamGaussian [70] uses a single image as a condition to generate 3D assets, while GS-", + "bbox": [ + 496, + 795, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6797", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/79c713f12bc171b2a6e360462360b1ffac48208e70c3cc7d2d1e66999abcfca3.jpg", + "image_caption": [ + "Figure 2. Overall framework of GaussianDreamer. Firstly, we utilize a 3D diffusion model to generate the initialized point clouds. After executing noisy point growing and color perturbation on the point clouds, we use them to initialize the 3D Gaussians. The initialized 3D Gaussians are further optimized using the SDS method [52] with a 2D diffusion model. Finally, we render the image using the 3D Gaussians by employing 3D Gaussian Splatting [25]. We can use one of various 3D diffusion models to generate the initialized point clouds. In this case, we take text-to-3D and text-to-motion diffusion models as examples." + ], + "image_footnote": [], + "bbox": [ + 78, + 85, + 893, + 265 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "GEN [9] implements high-quality generation from text to 3D. Our method shares a similar idea of using 3D Gaussian Splating as the representation method, which significantly reduces the cost of improving resolution and achieving a much faster optimization speed compared to estimizable mesh representation methods. And we can generate detailed 3D assets based on prompt texts in a very short time.", + "bbox": [ + 75, + 347, + 472, + 454 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 473, + 169, + 488 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first review 2D and 3D diffusion models and the 3D representation method - 3D Gaussian Splating [25]. We give an overview of the whole framework in Sec. 3.2. Then, in Sec. 3.3, we describe the process of initializing the 3D Gaussians with the assistance of 3D diffusion models. The further optimization of 3D Gaussians using the 2D diffusion model is described in Sec. 3.4.", + "bbox": [ + 75, + 502, + 470, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 631, + 218, + 645 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "DreamFusion. DreamFusion [52] is one of the most representative methods to lift 2D diffusion models to 3D, which proposes to optimize the 3D representation with the score distillation sampling (SDS) loss via a pre-trained 2D diffusion model $\\phi$ . Specifically, it takes MipNeRF [3] as the 3D representation method, whose parameters $\\theta$ are optimized. Taking the rendering method as $g$ , the rendered image results in $\\mathbf{x} = g(\\theta)$ . To make the rendered image $\\mathbf{x}$ similar to the samples obtained from the diffusion model $\\phi$ , DreamFusion uses a scoring estimation function: $\\hat{\\epsilon}_{\\phi}(\\mathbf{z}_t; y, t)$ , which predicts the sampled noise $\\hat{\\epsilon}_{\\phi}$ given the noisy image $\\mathbf{z}_t$ , text embedding $y$ , and noise level $t$ . By measuring the difference between the Gaussian noise $\\epsilon$ added to the rendered image $\\mathbf{x}$ and the predicted noise $\\hat{\\epsilon}_{\\phi}$ , this scoring estimation function can provide the direction for updating the parameter $\\theta$ . The formula for computing", + "bbox": [ + 75, + 657, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the gradient is as", + "bbox": [ + 500, + 348, + 617, + 363 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {S D S}} (\\phi , \\mathbf {x} = g (\\theta)) \\triangleq \\mathbb {E} _ {t, \\epsilon} \\left[ w (t) \\left(\\hat {\\epsilon} _ {\\phi} \\left(\\mathbf {z} _ {t}; y, t\\right) - \\epsilon\\right) \\frac {\\partial \\mathbf {x}}{\\partial \\theta} \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 376, + 893, + 425 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $w(t)$ is a weighting function.", + "bbox": [ + 500, + 440, + 736, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Gaussian Splatting. 3D Gaussian Splatting [25] (3D-GS) is a recent groundbreaking method for novel-view synthesis. Unlike implicit representation methods such as NeRF [45], which renders images based on volume rendering, 3D-GS renders images through splatting [84], achieving real-time speed. Specifically, 3D-GS represents the scene through a set of anisotropic Gaussians, defined with its center position $\\mu \\in \\mathbb{R}^3$ , covariance $\\Sigma \\in \\mathbb{R}^7$ , color $c \\in \\mathbb{R}^3$ , and opacity $\\alpha \\in \\mathbb{R}^1$ . And the 3D Gaussians can be queried as follows:", + "bbox": [ + 498, + 477, + 890, + 628 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nG (x) = e ^ {- \\frac {1}{2} (x) ^ {T} \\Sigma^ {- 1} (x)}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 643, + 890, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $x$ represents the distance between $\\mu$ and the query point. For computing the color of each pixel, it uses a typical neural point-based rendering [27, 28]. A ray $r$ is cast from the center of the camera, and the color and density of the 3D Gaussians that the ray intersects are computed along the ray. The rendering process is as follows:", + "bbox": [ + 496, + 678, + 893, + 768 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nC (r) = \\sum_ {i \\in \\mathcal {N}} c _ {i} \\sigma_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\sigma_ {j}), \\quad \\sigma_ {i} = \\alpha_ {i} G (x _ {i}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 784, + 890, + 825 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{N}$ represents the number of sample points on the ray $r$ , $c_{i}$ and $\\alpha_{i}$ denote the color and opacity of the i-th Gaussian, and $x_{i}$ is the distance between the point and the i-th Gaussian.", + "bbox": [ + 496, + 839, + 893, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6798", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Overall Framework", + "text_level": 1, + "bbox": [ + 76, + 90, + 266, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our overall framework consists of two parts, initialization with 3D diffusion model priors and optimization with the 2D diffusion model, as shown in Fig. 2. For initialization with 3D diffusion model priors, we use the 3D diffusion models $F_{3D}$ , instantiated with the text-to-3D and text-to-motion diffusion models, to generate the triangle mesh $m$ based on the text prompt $y$ , which can be denoted as $m = F_{3D}(y)$ . One set of generated point clouds is transformed from the mesh $m$ . Then the 3D Gaussians $\\theta_b$ are initialized via the generated point clouds after noisy point growing and color perturbation. For better quality, we utilize the 2D diffusion model $F_{2D}$ to further optimize the initialized 3D Gaussians $\\theta_b$ via SDS [52] with prompts $y$ , resulting in the final 3D Gaussians $\\theta_f$ . The target instance can be rendered in real time by splatting the generated Gaussians.", + "bbox": [ + 75, + 113, + 472, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Gaussian Initialization with 3D Diffusion Model Priors", + "text_level": 1, + "bbox": [ + 76, + 362, + 468, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we mainly discuss how to initialize the 3D Gaussians with 3D diffusion model priors. First, we use the 3D diffusion model $F_{3D}$ to generate 3D assets based on the prompts $y$ . Then we convert the 3D assets into point clouds and use the transformed point clouds to initialize the 3D Gaussians. We have employed two types of 3D diffusion models to generate 3D assets. Below, we explain how to initialize the 3D Gaussians using each of these models.", + "bbox": [ + 75, + 401, + 468, + 521 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1 Text-to-3D Diffusion Model", + "text_level": 1, + "bbox": [ + 76, + 532, + 323, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When using a text-based 3D generation model, generated 3D assets employ multi-layer perceptrons (MLPs) to predict SDF values and texture colors. To construct the triangle mesh $m$ , we query the SDF values at vertices along a regular grid with the size of $128^3$ within the MLPs. Then we query the texture colors at each vertex of $m$ . We convert the vertices and colors of $m$ into point clouds, denoted as $\\pmb{p}\\pmb{t}_m(\\pmb{p}_m,\\pmb{c}_m)$ . $\\pmb{p}_m\\in \\mathbb{R}^3$ refers to the position of the point clouds, which equals to the vertex coordinated of $m$ . $c_{m}\\in \\mathbb{R}^{3}$ refers to the color of the point clouds, which is the same as the color of $m$ . However, the obtained colors $c_{m}$ are relatively simple, and the positions $p_m$ are sparse.", + "bbox": [ + 75, + 556, + 468, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Noisy Point Growing and Color Perturbation. We do not use the generated point clouds $\\pmb{pt}_m$ to directly initialize the 3D Gaussians. To improve the quality of initialization, we perform noisy point growing and color perturbation around point clouds $\\pmb{pt}_m$ . First, we compute the bounding box (BBox) of the surface on $\\pmb{pt}_m$ and then uniformly grow point clouds $\\pmb{pt}_r(\\pmb{p}_r, \\pmb{c}_r)$ within the BBox. $\\pmb{p}_r$ and $\\pmb{c}_r$ represent the positions and colors of $\\pmb{pt}_r$ . To enable fast searching, we construct a KDTree [4] $K_m$ using the positions $\\pmb{p}_m$ . Based on the distances between the positions $\\pmb{p}_r$", + "bbox": [ + 75, + 750, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/bb582a35d7e07fd94b09d153ed9d7211e1ceb6f1e72bd01be21944fcde6c8680.jpg", + "image_caption": [ + "Figure 3. The process of noisy point growing and color perturbation. \"Grow&Pertb.\" denotes noisy point growing and color perturbation." + ], + "image_footnote": [], + "bbox": [ + 509, + 85, + 883, + 234 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and the nearest points found in the KDTree $K_{m}$ , we determine which points to keep. In this process, we select points within the (normalized) distance of 0.01. For the noisy point clouds, we make their colors $\\pmb{c}_r$ similar to $\\pmb{c}_m$ , and also add some perturbations:", + "bbox": [ + 496, + 296, + 890, + 372 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {c} _ {r} = \\boldsymbol {c} _ {m} + \\boldsymbol {a}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 645, + 386, + 890, + 400 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the values of $\\mathbf{a}$ are randomly sampled between 0 and 0.2. We merge the positions and colors of $\\mathbf{p}\\mathbf{t}_{m}$ and $\\mathbf{p}\\mathbf{t}_{r}$ to obtain the final point clouds.", + "bbox": [ + 496, + 412, + 890, + 458 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {p} \\boldsymbol {t} \\left(\\boldsymbol {p} _ {f}, \\boldsymbol {c} _ {f}\\right) = \\left(\\boldsymbol {p} _ {m} \\oplus \\boldsymbol {p} _ {r}, \\boldsymbol {c} _ {m} \\oplus \\boldsymbol {c} _ {r}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 470, + 890, + 486 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\oplus$ is the concatenation operation. Fig. 3 illustrates the process of noisy point growing and color perturbation. Finally, we initialize the positions $\\mu_{b}$ and colors $c_{b}$ of the 3D Gaussians $\\theta_{b}(\\mu_{b}, c_{b}, \\Sigma_{b}, \\alpha_{b})$ using both the positions $\\pmb{p}_{f}$ and colors $\\pmb{c}_{f}$ of final point clouds $\\pmb{pt}$ . The opacity $\\alpha_{b}$ of the 3D Gaussians is initialized to 0.1, and the covariance $\\Sigma_{b}$ is calculated as the distance between the nearest two points. Algorithm 1 shows the specific algorithm flowchart.", + "bbox": [ + 496, + 498, + 890, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.2 Text-to-Motion Diffusion Model", + "text_level": 1, + "bbox": [ + 500, + 625, + 774, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We generate a sequence of human body motions using text and select a human pose that best matches the given text. We then convert the keypoints of this human pose into the SMPL model [39], which is represented by a triangle mesh $m$ . We then convert the mesh $m$ into point clouds $\\pmb{pt}_m(\\pmb{p}_m, \\pmb{c}_m)$ , where the position $\\pmb{p}_m$ of each point in the point clouds corresponds to the vertices of $m$ . As for the color $\\pmb{c}_m$ of $\\pmb{pt}_m$ , since the SMPL model used here does not have textures, we randomly initialize $\\pmb{c}_m$ . To move $\\pmb{pt}_m$ near the origin, we calculate the center point $\\pmb{p}_c \\in \\mathbb{R}^3$ of $\\pmb{p}_m$ and subtract the position of point clouds $\\pmb{pt}_m$ from the center point $\\pmb{p}_c$ .", + "bbox": [ + 496, + 648, + 890, + 830 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {p} \\boldsymbol {t} \\left(\\boldsymbol {p} _ {f}, \\boldsymbol {c} _ {f}\\right) = \\boldsymbol {p} \\boldsymbol {t} _ {m} \\left(\\boldsymbol {p} _ {m} - \\boldsymbol {p} _ {c}, \\boldsymbol {c} _ {m}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 843, + 890, + 859 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, we use the point clouds $pt$ to initialize the 3D Gaussians, similar to what is described in Sec. 3.3.1. To improve", + "bbox": [ + 496, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6799", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [], + "code_body": "Algorithm 1 The 3D Gaussian Initialization. \n $\\pmb{pt}_m(\\pmb{p}_m, \\pmb{c}_m)$ : Point clouds generated from $F_{3D}$ . \n $\\pmb{pt}_r(\\pmb{p}_r, \\pmb{c}_r)$ : Growing point clouds within the BBox. \n $\\pmb{pt}(\\pmb{p}_f, \\pmb{c}_f)$ : Point clouds used for initializing the 3D Gaussians. \n $\\theta_b(\\mu_b, c_b, \\Sigma_b, \\alpha_b)$ : Initialized 3D Gaussians.", + "bbox": [ + 76, + 90, + 468, + 181 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [], + "code_body": "Stage 1: Generate points in the BBox. \n $K_{m} \\leftarrow$ BuildKDTree( $p_{m}$ ) \n $\\text{BBox} \\leftarrow p_{m}$ $\\text{BBox} \\leftarrow$ Positions bounding box. \nLow, High $\\leftarrow$ BBox.MinBound, BBox.MaxBound \n $\\text{BBox} \\leftarrow$ Boundary of bounding box. \n $\\boldsymbol{ps}_{u} \\leftarrow$ Uniform(Low, High, size = (NumPoints, 3)) \n $\\triangleright$ Points in the BBox.", + "bbox": [ + 93, + 186, + 468, + 291 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Stage 2: Keep the points that meet the distance requirement.", + "bbox": [ + 93, + 292, + 468, + 320 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [], + "code_body": "$\\pmb{p}_r,\\pmb {c}_r = \\left[ \\begin{array}{ll} \\end{array} \\right],\\left[ \\begin{array}{ll} \\end{array} \\right]$ \nfor all $\\pmb{p_u}$ in $p s_{u}$ do $\\pmb{p}_{un},i\\gets K_m.\\mathrm{SearchNearest}(\\pmb {p}_u)$ Nearest point and its index in $K_{m}$ if $|\\pmb {p}_{un} - \\pmb {p}_u| < 0.01$ then $\\pmb{p}_r$ .append $(\\pmb {p}_u)$ $c_{r}$ .append $(c_{m}[i] + 0.2\\times \\mathrm{Random}(\\mathrm{size} = 3))$ $\\triangleright$ Color of the nearest point plus perturbation.", + "bbox": [ + 93, + 321, + 468, + 441 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [], + "code_body": "end if \nend for \n $\\pmb {p}_f\\gets \\pmb {p}_m\\oplus \\pmb {p}_r$ $c_{f}\\gets c_{m}\\oplus c_{r}$ \nStage 3: Initialize the 3D Gaussians. \n $\\mu_b,c_b\\gets p_f,c_f$ $\\triangleright$ Positions and colors of the 3D Gaussians. \n $D\\leftarrow \\mu_{b}\\quad \\triangleright$ Distance between the nearest two positions. \n $\\Sigma_{b},\\alpha_{b}\\gets D,0.1$ $\\triangleright$ Covariance and opacity of the 3D Gaussians.", + "bbox": [ + 93, + 443, + 468, + 593 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the generation of motion sequences, we simplify the text by retaining only the relevant parts related to the motion and add a subject. For example, if the text prompt is \"Iron man kicks with his left leg\", we transform it into \"Someone kicks with the left leg\" when generating the motion sequences.", + "bbox": [ + 75, + 621, + 468, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Optimization with the 2D Diffusion Model", + "text_level": 1, + "bbox": [ + 76, + 709, + 437, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To enrich details and improve the quality of the 3D asset, we optimize the 3D Gaussians $\\theta_{b}$ with a 2D diffusion model $F_{2D}$ after initializing them with 3D diffusion model priors. We employ the SDS (Score Distillation Sampling) loss to optimize the 3D Gaussians. First, we use the method of 3D Gaussian Splatting [25] to obtain the rendered image $\\mathbf{x} = g(\\theta_{i})$ . Here, $g$ represents the splatting rendering method as in Eq. 3. Then, we use Eq. 1 to calculate the gradients for updating the Gaussian parameters $\\theta_{i}$ with the 2D diffusion model $F_{2D}$ . After a short optimization period using the 2D diffusion model $F_{2D}$ , the final generated 3D", + "bbox": [ + 75, + 734, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c3b93ff047c7d6bb6208570425dcd25786d3ff784530b45cbd46db7da22cd83c.jpg", + "table_caption": [ + "Table 1. Quantitative comparisons on ${\\mathrm{T}}^{3}$ Bench [17]." + ], + "table_footnote": [ + "Scores above are average of two metrics (quality and alignment).", + "$\\dagger$ GPU time counted in their papers." + ], + "table_body": "
MethodTime†Single Obj.Single w/ Surr.Multi Obj.Average
SJC [76]-24.719.811.718.7
DreamFusion [52]6 hours24.424.616.121.7
Fantasia3D [6]6 hours26.427.018.524.0
LatentNeRF [43]15 minutes33.130.620.628.1
Magic3D [32]5.3 hours37.035.425.732.7
ProlificDreamer [77]~10 hours49.444.835.843.3
Ours15 minutes54.048.634.545.7
", + "bbox": [ + 501, + 114, + 908, + 215 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "instance $\\theta_{f}$ achieves high quality and fidelity on top of the 3D consistency provided by the 3D diffusion model $F_{3D}$ .", + "bbox": [ + 498, + 253, + 890, + 284 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 297, + 632, + 314 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we first present the implementation details in Sec. 4.1. In Sec. 4.2, we show the quantitative comparison results. Then, in Sec. 4.3, we showcase the visualization results of our method and compare them with other methods. In Sec. 4.4, we conduct a series of ablation experiments to validate the effectiveness of our method. Finally, we discuss the limitations of our method.", + "bbox": [ + 496, + 321, + 890, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 436, + 718, + 452 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our method is implemented in PyTorch [51], based on ThreeStudio [14]. The 3D diffusion models used in our method are Shap-E [24] and MDM [72], and we load the Shap-E model finetuned on Objaverse [11] in Cap3D [42]. For the 2D diffusion model, we use stabilityai/stable-diffusion-2-1-base [58], with a guidance scale of 100. The timestamps we use are uniformly sampled from 0.02 to 0.98 before 500 iterations, and change to 0.02 to 0.55 after 500 iterations. For the 3D Gaussians, the learning rates of opacity $\\alpha$ and position $\\mu$ are $10^{-2}$ and $5 \\times 10^{-5}$ . The color $c$ of the 3d Gaussians is represented by the sh coefficient, with the degree set to 0 and the learning rate set to $1.25 \\times 10^{-2}$ . The covariance of the 3D Gaussians is converted into scaling and rotation for optimization, with learning rates of $10^{-3}$ and $10^{-2}$ , respectively. The radius of the camera we use for rendering is from 1.5 to 4.0, with the azimuth in the range of -180 to 180 degrees and the elevation in the range of -10 to 60 degrees. The total training iterations are 1200. All our experiments can be completed within 15 minutes on a single RTX 3090 with a batch size of 4. The resolution we use for rendering is $1024 \\times 1024$ , which is scaled to 512 $\\times$ 512 when optimizing using the 2D diffusion model. We can render in real time at $512 \\times 512$ resolution. And all our code will be released.", + "bbox": [ + 496, + 460, + 892, + 821 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 500, + 832, + 720, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate quality and consistency following T3Bench [17], which provides a comprehensive benchmark for text-to-3D generation. Three text categories are", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6800", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c4b46be0ed11935322d0c31b2d3bb349e22df589648bbded554dea0a93925251.jpg", + "image_caption": [ + "DreamFusion 6.0 Hours" + ], + "image_footnote": [], + "bbox": [ + 101, + 118, + 230, + 170 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/79922dd0a52a901b81385541823270f804f0894692275bd8d0b8e4a1ab9bf33c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 259, + 126, + 393, + 170 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/45155875c3469a406d73196d631c4d425f392d5197d3d1336d4eab0671160888.jpg", + "image_caption": [ + "Magic3D 5.3 Hours", + "Fantasia3D 6.0 Hours" + ], + "image_footnote": [], + "bbox": [ + 408, + 119, + 549, + 170 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4fc682d9acda770c6b6876118a58e26afdfb7968429e468e00b1398a9d180c0d.jpg", + "image_caption": [ + "ProlificDreamer ~Hours", + "GaussianDreamer (Ours) 15 Minutes" + ], + "image_footnote": [], + "bbox": [ + 575, + 114, + 722, + 172 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/20649408e8da75ccdd629cbb303ee9b982d1d0ebbd7b4a807d972d194c73d22a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 730, + 125, + 805, + 167 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3df616b454f46e13d39cb16915bb76b907e4b74a245f4c6b5b805df783fe0d7a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 810, + 125, + 885, + 167 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7290185a6994106d0a5a8516f71e33486740dd15a357030589b6eae517aff470.jpg", + "image_caption": [ + "Figure 4. Qualitative comparisons between our method and DreamFusion [52], Magic3D [32], Fantasy3D [6] and ProlificDreamer [77]. Here we count the GPU time in their papers. The time for DreamFusion is measured on TPUv4, Magic3D is measured on A100, Fantasy3D is measured on RTX 3090, and our method is measured on RTX 3090." + ], + "image_footnote": [], + "bbox": [ + 81, + 194, + 158, + 244 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/56c02fa1558038816f48335b17cb2aba7f56ed36ef33a19395d676333b3d92e9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 161, + 195, + 230, + 244 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/753289c373588837d7220f4a0d5e9e8da89bc317da0dbaa0bd95d1c961b26d9e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 251, + 186, + 318, + 243 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b4448fb8506ebdf152b2561d19cafa572bd2e5bc7e483177892fb3e199e53101.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 188, + 395, + 244 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/169a62acecb4392a588a8975c36490e2b888819033ac8c73554bda5b07700fc9.jpg", + "image_caption": [ + "A plate piled high with chocolate chip cookies." + ], + "image_footnote": [], + "bbox": [ + 411, + 186, + 480, + 242 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f2692df22d77547f9ebc2ebdd6f765eea1ac84c701030bbdd06d0877f3b20697.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 493, + 186, + 566, + 242 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ae5e65b857a6db3701ba16e0eeabc65a30c56496df91840937db887fe2be934f.jpg", + "image_caption": [ + "A 3D model of an adorable cottage with a thatched roof." + ], + "image_footnote": [], + "bbox": [ + 573, + 189, + 718, + 247 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5579c54f6d7b9f8b045568f74bf15d7f0203d1344e19ba30a5ffd7c66e1a609d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 725, + 190, + 805, + 244 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9247cd7677cc732004fdd032f1084a7da218c12b9ca42b6d12fa72d53a225881.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 812, + 193, + 879, + 244 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ada454a8eb06f50484c9c7782bb9e057256733ac589ebd69888de2aa944d7a7b.jpg", + "image_caption": [ + "A fox" + ], + "image_footnote": [], + "bbox": [ + 84, + 315, + 241, + 386 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/feeec47e65c5f8eff091d6cfcdcfa442960f01a2934b13c9f5fe3bf016d2784b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 245, + 316, + 294, + 387 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8c14e8aec249158efe551260eb11e4ec9dc1bfe39ecfcb526272efa55b10c608.jpg", + "image_caption": [ + "An amigurumi motorcycle..." + ], + "image_footnote": [], + "bbox": [ + 81, + 417, + 207, + 484 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/404b6a7aada9388f28bdba8b8b99e37db7883c3ccda42b411c647b6a89c179d6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 222, + 422, + 321, + 484 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2710ada4be39e44be8d6923f5a8dd5eda21f943faea69e74d4711f08e3ebdffe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 344, + 318, + 446, + 387 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ea366c3716a1473a6be7cb841be0be38d34eae3309221c1e4a85f9f26d991bee.jpg", + "image_caption": [ + "An opulent couch from the palace of Versailles" + ], + "image_footnote": [], + "bbox": [ + 450, + 321, + 584, + 387 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b2d0ead48d1c61ef323021840b5d61c2a6496cef6e99c7e080e020fe26ba17ff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 606, + 332, + 730, + 385 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/59806878a7759f0f8438656b24f462d6ac9c206c2bb48edb0efc8753c970ae7a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 750, + 325, + 833, + 387 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7df71c74c5a0c7f6f31995bfb3097faefbdd82a80d09358763f36f8094a5ccd3.jpg", + "image_caption": [ + "Magic gun, game asset, mistery...", + "Magic dagger, mistery, ancient...", + "Figure 5. More generated samples by our GaussianDreamer. Two views of each sample are shown." + ], + "image_footnote": [], + "bbox": [ + 398, + 401, + 509, + 496 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7e657b51919a4558f204dec6904f0da93dd2720ac6c261de0835cb160652020f.jpg", + "image_caption": [ + "A zoomed out DSLR photo of a lion's mane jellyfish" + ], + "image_footnote": [], + "bbox": [ + 612, + 412, + 705, + 491 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f890a4999f962286b63d61c7709edc53b66c63eb54a13a65ca0a94445bff3c97.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 414, + 821, + 491 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "designed for 3D generation with increasing complexity – single objects, single objects with surroundings, and multi objects. In Tab. 1, our method outperforms the compared methods while enjoying a short generation time.", + "bbox": [ + 75, + 532, + 468, + 594 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Visualization Results", + "text_level": 1, + "bbox": [ + 76, + 606, + 274, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we present the results of initializing the 3D Gaussians with two different 3D diffusion models: text-to-3D and text-to-motion diffusion models.", + "bbox": [ + 75, + 630, + 468, + 675 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Initialization with Text-to-3D Diffusion Model. We show the comparison results with DreamFusion [52], Magic3D [32], Fantasy3D [6] and ProlificDreamer [77] in Fig. 4. In addition to our method, the figures of other methods are downloaded from the paper of ProlificDreamer. When encountering prompts that involve the combination of multiple objects, such as the prompt \"A plate piled high with chocolate chip cookies\", the generated results from Magic3D, Fantasy3D, and ProlificDreamer do not include a plate. In contrast, our generated result can effectively combine a plate and chocolate chip cookies. Furthermore, compared to DreamFusion, the plate we generate has better patterns. Our method shows comparable quality while saving 21 - 24 times the generation time compared to their", + "bbox": [ + 75, + 688, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "methods. Moreover, the 3D Gaussians generated by our method can directly achieve real-time rendering without further transformation into mesh-like structures. Fig. 5 visualizes more samples generated by our GaussianDreamer from various prompts, which show good 3D consistency while having high-quality details.", + "bbox": [ + 496, + 532, + 893, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Initialization with Text-to-Motion Diffusion Model. In Fig. 6, we present the comparison results with DreamFusion [52], DreamAvatar [5], DreamWaltz [20], and AvatarVerse [86]. In addition to our method, the figures of other methods are downloaded from the paper of AvatarVerse. It is worth noting that our prompt is \"Spiderman/Stormtrooper stands with open arms\", while the prompts for other methods are \"Spiderman/Stormtrooper\". This is because when generating motion using text-to-motion diffusion model as initialization, we require more specific action descriptions. Our method achieves a speedup of 4-24 times compared to other methods, while maintaining comparable quality. Additionally, our method allows for generating 3D avatars with specified body poses. In Fig. 7, we provide more results generated with different human body poses. We first generate a sequence of motions that match the text prompt using a text-to-motion 3d diffusion model, and then initialize the", + "bbox": [ + 496, + 643, + 892, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6801", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0cfc1257a36d9ad19d9bf38570bddbbf78df0b6ce4206c80e7c433b10c1cd8e0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 78, + 87, + 251, + 231 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d4b09dcf5d6995715bc51a8ea393ea2c634c32f67eaf132566c4f2b25e040929.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 254, + 88, + 452, + 231 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f8832450bbe09a68ea5aa1fed5a5c62c2bc15309ae6bff4a546bf4be874239db.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 460, + 88, + 633, + 231 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3399a3158964f6a65d26795c956329a332dd2a19cf2efcf69223255a2e436d7a.jpg", + "image_caption": [ + "GaussianDreamer (Ours) 15 Minutes" + ], + "image_footnote": [], + "bbox": [ + 638, + 88, + 756, + 222 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/84996c3396cf61b95e47c8f381ef314d2432e42526d5dbaddba30ba6f1041219.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 241, + 245, + 366 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9f086da7d7c61a81113c6d0f895fe1f9581a4c06fc5fd0a81367a8823db5fb89.jpg", + "image_caption": [ + "Stormtrooper" + ], + "image_footnote": [], + "bbox": [ + 259, + 243, + 447, + 367 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/44cef7937051be096397c49f171c7f367570f6b49fe250aa88146fd258700661.jpg", + "image_caption": [ + "Spiderman" + ], + "image_footnote": [], + "bbox": [ + 460, + 243, + 630, + 367 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fdb541959261afc19ef371e5c0033a236a3baca59a6e849bc55d847b8a1bb003.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 255, + 754, + 361 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f38748c4101231bfb0f09515c973d036bd80949127338995681d58505b84b99f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 260, + 870, + 361 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9b44a10d6ece8b727f465d7903ef5f0dffa3c1ae84cc4b9b98dc405816f1a392.jpg", + "image_caption": [ + "Figure 6. Qualitative comparisons between our method and DreamFusion [52], DreamAvatar [5], DreamWaltz [20], and AvatarVerse [86].", + "Figure 7. More generated 3D avatars by our GaussianDreamer initialized with the different poses of SMPL [39]. Here, the different poses of SMPL are generated using a text-to-motion diffusion model." + ], + "image_footnote": [], + "bbox": [ + 76, + 400, + 470, + 671 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3d Gaussians with the SMPL of one selected pose in motions. Our method can generate 3D avatars in any desired pose.", + "bbox": [ + 75, + 736, + 468, + 782 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study and Analysis", + "text_level": 1, + "bbox": [ + 76, + 786, + 333, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The Role of Initialization. As shown in Fig. 8, we first conduct an ablation experiment on the initialization of the 3D Gaussians to validate that initialization can improve 3D consistency. The first column is the rendering result of Shap-E [24] with NeRF [45] as the 3D representation. The second column is the result of optimizing the 3D Gaussians", + "bbox": [ + 75, + 809, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "randomly initialized within a cube using the SDS loss, and the third column is our method. We show the initialization effects on 3 samples. In the first and second rows, ShapE has good generation results while our method provides more complex geometries and more realistic appearances. Compared with random initialization, in the first row, details of our method are better. In the second row, the 3D assets generated by random initialization have the multi-head problem, which does not occur in our method. The initialization of the 3D diffusion model can avoid unreasonable geometry. In the third row, the generation result of Shap-E is far different from the given text prompt while our method makes the 3D assets closer to the prompt through the 2D diffusion model. Our method can expand the domain of Shap-E prompts, allowing for the generation of 3D assets based on a wider range of prompts.", + "bbox": [ + 496, + 407, + 893, + 650 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Noisy Point Growing and Color Perturbation. Fig. 9 illustrates the ablation results of noisy point growing and color perturbation. With noisy point growing and color perturbation applied, the first row showcases improved details in the sniper rifle. Additionally, the second column generates an amigurumi motorcycle that better aligns with the style characteristics of amigurumi mentioned in the prompt, compared to the case without noisy point growing and color perturbation.", + "bbox": [ + 496, + 661, + 893, + 797 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Initialization with Different Text-to-3D Diffusion Models. We select two text-to-3D generation models, ShapeE [24] and Point-E [48], to validate the effectiveness of our framework. We load the Point-E model finetuned on Objaverse [11] in Cap3D[42]. Fig. 10, we showcase the generated results after initializing the 3D Gaussians using", + "bbox": [ + 496, + 809, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6802", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1807572c3b4a4dcc7a809670101e10432bb4ac3afede8ee74eb3462a2422a0af.jpg", + "image_caption": [ + "Shap-E" + ], + "image_footnote": [], + "bbox": [ + 93, + 102, + 349, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5e25c51298fd20ea10da18d92f718411b9e43e6904456d22a7e0259068530b09.jpg", + "image_caption": [ + "Ours w/o Point Cloud Prior" + ], + "image_footnote": [], + "bbox": [ + 357, + 111, + 480, + 181 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c80410bf9f9e49ef5cc21b67a47c063c53106cb9a3fc5b5d847bd910e504c75c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 112, + 627, + 180 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7975de07c8bec90db2745111a489b16e732d4d9ab3c20ad1e2fe11e8f7e87573.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 109, + 754, + 184 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d50db4694a3713138bec67aeb5b66782e706c844f78ce7e4016efa8d0b902549.jpg", + "image_caption": [ + "GaussianDreamer (Ours)" + ], + "image_footnote": [], + "bbox": [ + 764, + 108, + 888, + 185 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e85175ebd52e7829a9b9d709beb8c5f2965ca828c112b7bf8a3844822a6cdab8.jpg", + "image_caption": [ + "A stack of pancakes covered in maple syrup" + ], + "image_footnote": [], + "bbox": [ + 93, + 205, + 352, + 289 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1cd727f291e3666492d0fe66f68a9419dbbbc61990802c14b84110925e41eaf1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 207, + 493, + 287 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6bfd90bfc09b8e436553ff957360f6965f5880169e40baed95a09c1d296c3991.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 212, + 620, + 289 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e637eae71eeee2c858b591bfef01b6e77dbf0c01105092f25c7c58d8e73fe4f5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 215, + 754, + 301 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/97ec9a04ccf38a249c6d9f60b6e1be68ca6f14350fd4007241457038e7df0598.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 758, + 217, + 870, + 301 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/95f05e8ebe08383af6510ec9ed620522020bd6904899b22d9c4cb26f6e7ccdfb.jpg", + "image_caption": [ + "Figure 8. Ablation studies of the initialization of the 3D Gaussians. The Shap-E [24] rendering resolution here is $256 \\times 256$ ." + ], + "image_footnote": [], + "bbox": [ + 78, + 309, + 372, + 378 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c82c6c8330d5b8c021d620571379e98034ff0b5518f26bffe2ad6b67dc03e14d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 387, + 306, + 493, + 382 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/03947ce9b4662d0af060f95acc3bfa577f3e427e417b69f36b454c3e0f2a13a3.jpg", + "image_caption": [ + "A lionfish" + ], + "image_footnote": [], + "bbox": [ + 503, + 306, + 620, + 383 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/379e57fae2221050da7d14800c0d41221bc64214b31cb4e2fbd7f5f61cc362bd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 622, + 311, + 750, + 381 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/29084092880b0720a0a55b982efda37476f45335831fe66cd772233e870f3c25.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 753, + 311, + 883, + 381 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/003ebaa33cdd9170969b8dacc3eacea37d58121fba4fa36839e61f1813ad19e1.jpg", + "image_caption": [ + "Ours w/o Grow&Pertb.", + "GaussianDreamer (Ours)", + "Sniper rifle, asset, scifi, cyberpunk..." + ], + "image_footnote": [], + "bbox": [ + 81, + 441, + 468, + 468 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9e5fe0ab69529996c392ef7eb29bfc293fd18c2cd212503dd2cba6ecd8d231d6.jpg", + "image_caption": [ + "An amigurumi motorcycle...", + "Figure 9. Ablation studies of noisy point growing and color perturbation. \"Grow&Pertb.\" denotes noisy point growing and color perturbation." + ], + "image_footnote": [], + "bbox": [ + 84, + 479, + 455, + 523 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/33f58febb62d7925bb0334415925dabb5be1e55409b01b8859917efbd0dcf946.jpg", + "image_caption": [ + "Initialization with Point-E", + "Initialization with Shap-E", + "An airplane made out of wood" + ], + "image_footnote": [], + "bbox": [ + 89, + 593, + 439, + 628 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/75c33677a6349772832ee11dd09544e8901798823d9fe45a8a16f3a422e1b31e.jpg", + "image_caption": [ + "Flamethrower, with fire, scif, cyberpunk...", + "Figure 10. Ablation studies of initialization with different text-to-3D diffusion models: Point-E [48] and Shap-E [24]." + ], + "image_footnote": [], + "bbox": [ + 81, + 638, + 467, + 676 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "one of two text-to-3D generation models. It can be seen that both initializations yield good generation results. However, considering that Shap-E generates 3D assets based on NeRF and SDF, which provide higher fidelity compared to the point cloud representation used by Point-E, the geometry of the airplane in the first row of Fig. 10 appears better when initialized using Shap-E.", + "bbox": [ + 75, + 729, + 468, + 837 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Limitations", + "text_level": 1, + "bbox": [ + 76, + 845, + 202, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The edges of the 3D assets generated by our method are not always sharp, and there may be unnecessary 3D Gaussians", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "around the object surface. How to filter these point clouds will be a possible direction for improvement. Our approach utilizes 3D diffusion model priors, which greatly alleviates the issue of multi-face problems. However, there is still a small chance of encountering the multi-face problems in scenes where there is minimal geometric difference but significant appearance discrepancy between the front and back of objects such as a backpack. Utilizing 3D-aware diffusion models [35, 67, 90] may be able to solve this problem. Additionally, our method has limited effectiveness in generating large-scale scenes, such as indoor scenes.", + "bbox": [ + 496, + 431, + 893, + 598 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 627, + 619, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose a fast text-to-3D method GaussianDreamer by bridging the abilities of 3D and 2D diffusion models via the Gaussian splatting representation. GaussianDreamer can generate detailed and realistic geometry and appearance while maintaining 3D consistency. The 3D diffusion model priors and geometry priors from the 3D Gaussians effectively promote the convergence speed. Each sample can be generated within 15 minutes on one GPU. We believe the approach of bridging 3D and 2D diffusion models could be a promising direction to generate 3D assets efficiently.", + "bbox": [ + 496, + 657, + 893, + 811 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 500, + 839, + 658, + 857 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by the National Natural Science Foundation of China (No. 62376102).", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6803", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Tenglong Ao, Zeyi Zhang, and Libin Liu. Gesture diffuclip: Gesture diffusion model with clip latents. arXiv preprint arXiv:2303.14613, 2023. 2", + "[2] Mohammadreza Armandpour, Huangjie Zheng, Ali Sadeghian, Amir Sadeghian, and Mingyuan Zhou. Re-imagine the negative prompt algorithm: Transform 2d diffusion into 3d, alleviate janus problem and beyond. arXiv preprint arXiv:2304.04968, 2023.1, 2", + "[3] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, pages 5855-5864, 2021. 2, 3", + "[4] Jon Louis Bentley. Multidimensional binary search trees used for associative searching. Communications of the ACM, 18(9):509-517, 1975. 4", + "[5] Yukang Cao, Yan-Pei Cao, Kai Han, Ying Shan, and KwanYee K Wong. Dreamavatar: Text-and-shape guided 3d human avatar generation via diffusion models. arXiv preprint arXiv:2304.00916, 2023. 6, 7", + "[6] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. *Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation.* arXiv preprint arXiv:2303.13873, 2023. 1, 2, 5, 6", + "[7] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In CVPR, pages 18000-18010, 2023. 2", + "[8] Yiwen Chen, Chi Zhang, Xiaofeng Yang, Zhongang Cai, Gang Yu, Lei Yang, and Guosheng Lin. It3d: Improved text-to-3d generation with explicit view synthesis. arXiv preprint arXiv:2308.11473, 2023. 2", + "[9] Zilong Chen, Feng Wang, and Huaping Liu. Text-to-3d using gaussian splatting. arXiv preprint arXiv:2309.16585, 2023. 1, 2, 3", + "[10] Rishabh Dabral, Muhammad Hamza Mughal, Vladislav Golyanik, and Christian Theobalt. Mofusion: A framework for denoising-diffusion-based motion synthesis. In CVPR, pages 9760-9770, 2023. 2", + "[11] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In CVPR, pages 13142-13153, 2023. 5, 7", + "[12] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. NeurIPS, 35:31841-31854, 2022. 1, 2", + "[13] Jiatao Gu, Alex Trevithick, Kai-En Lin, Joshua M Susskind, Christian Theobalt, Lingjie Liu, and Ravi Ramamoorthi. Nerfdiff: Single-image view synthesis with nerf-guided distillation from 3d-aware diffusion. In ICML, pages 11808–11826. PMLR, 2023. 2", + "[14] Yuan-Chen Guo, Ying-Tian Liu, Ruizhi Shao, Christian Laforte, Vikram Voleti, Guan Luo, Chia-Hao Chen, Zi-Xin Zou, Chen Wang, Yan-Pei Cao, and Song-Hai Zhang." + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "threestudio: A unified framework for 3d content generation. https://github.com/threestudio-project/threestudio, 2023.5", + "[15] Anchit Gupta, Wenhan Xiong, Yixin Nie, Ian Jones, and Barlas Oğuz. 3dgen: Triplane latent diffusion for textured mesh generation. arXiv preprint arXiv:2303.05371, 2023. 1, 2", + "[16] Xiao Han, Yukang Cao, Kai Han, Xiatian Zhu, Jiankang Deng, Yi-Zhe Song, Tao Xiang, and Kwan-Yee K Wong. Headsculpt: Crafting 3d head avatars with text. In NeurIPS, 2023. 2", + "[17] Yuze He, Yushi Bai, Matthieu Lin, Wang Zhao, Yubin Hu, Jenny Sheng, Ran Yi, Juanzi Li, and Yong-Jin Liu. T 3 bench: Benchmarking current progress in text-to-3d generation. arXiv preprint arXiv:2310.02977, 2023. 5", + "[18] Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. arXiv preprint arXiv:2205.08535, 2022. 2", + "[19] Shoukang Hu, Fangzhou Hong, Tao Hu, Liang Pan, Haiyi Mei, Weiye Xiao, Lei Yang, and Ziwei Liu. Humanliff: Layer-wise 3d human generation with diffusion model. arXiv preprint arXiv:2308.09712, 2023. 2", + "[20] Yukun Huang, Jianan Wang, Ailing Zeng, He Cao, Xianbiao Qi, Yukai Shi, Zheng-Jun Zha, and Lei Zhang. Dreamwaltz: Make a scene with complex 3d animatable avatars. arXiv preprint arXiv:2305.12529, 2023. 6, 7", + "[21] Yangyi Huang, Hongwei Yi, Yuliang Xiu, Tingting Liao, Ji-axiang Tang, Deng Cai, and Justus Thies. Tech: Text-guided reconstruction of lifelike clothed humans. arXiv preprint arXiv:2308.08545, 2023. 2", + "[22] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In CVPR, pages 867-876, 2022. 2", + "[23] Ruixiang Jiang, Can Wang, Jingbo Zhang, Mengei Chai, Mingming He, Dongdong Chen, and Jing Liao. Avatarcraft: Transforming text into neural human avatars with parameterized shape and pose control. arXiv preprint arXiv:2303.17606, 2023. 2", + "[24] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 1, 2, 5, 7, 8", + "[25] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 2, 3, 5", + "[26] Jihoon Kim, Jiseob Kim, and Sungjoon Choi. Flame: Freeform language-based motion synthesis & editing. In AAAI, pages 8255-8263, 2023. 2", + "[27] Georgios Kopanas, Julien Philip, Thomas Leimkuhler, and George Drettakis. Point-based neural rendering with perview optimization. In Computer Graphics Forum, pages 29-43. Wiley Online Library, 2021. 3", + "[28] Georgios Kopanas, Thomas Leimkuhler, Gilles Rainer, Clément Jambon, and George Drettakis. Neural point cata-caustics for novel-view synthesis of reflections. ACM Transactions on Graphics (TOG), 41(6):1-15, 2022. 3" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6804", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Jiabao Lei, Yabin Zhang, Kui Jia, et al. Tango: Text-driven photorealistic and robust 3d stylization via lighting decomposition. NeurIPS, 35:30923-30936, 2022. 2", + "[30] Gang Li, Heliang Zheng, Chaoyue Wang, Chang Li, Changwen Zheng, and Dacheng Tao. 3ddesigner: Towards photorealistic 3d object generation and editing with text-guided diffusion models. arXiv preprint arXiv:2211.14108, 2022. 2", + "[31] Weiyu Li, Rui Chen, Xuelin Chen, and Ping Tan. Sweetdreamer: Aligning geometric priors in 2d diffusion for consistent text-to-3d. arXiv preprint arXiv:2310.02596, 2023. 2", + "[32] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, pages 300–309, 2023. 1, 2, 5, 6", + "[33] Yukang Lin, Haonan Han, Chaoqun Gong, Zunnan Xu, Yachao Zhang, and Xiu Li. Consistent123: One image to highly consistent 3d asset using case-aware diffusion priors. arXiv preprint arXiv:2309.17261, 2023. 2", + "[34] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Zexiang Xu, Hao Su, et al. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. arXiv preprint arXiv:2306.16928, 2023.", + "[35] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. arXiv preprint arXiv:2303.11328, 2023. 8", + "[36] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 2", + "[37] Zhen Liu, Yao Feng, Michael J Black, Derek Nowrouzezahrai, Liam Paull, and Weiyang Liu. Meshdiffusion: Score-based generative 3d mesh modeling. arXiv preprint arXiv:2303.08133, 2023. 2", + "[38] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. arXiv preprint arXiv:2310.15008, 2023. 2", + "[39] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 851-866. 2023. 2, 4, 7", + "[40] Jonathan Lorraine, Kevin Xie, Xiaohui Zeng, Chen-Hsuan Lin, Towaki Takikawa, Nicholas Sharp, Tsung-Yi Lin, Ming-Yu Liu, Sanja Fidler, and James Lucas. Att3d: Amortized text-to-3d object synthesis. arXiv preprint arXiv:2306.07349, 2023. 2", + "[41] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In CVPR, pages 2837-2845, 2021. 2", + "[42] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. arXiv preprint arXiv:2306.07279, 2023. 5, 7", + "[43] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of" + ], + "bbox": [ + 78, + 90, + 470, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "3d shapes and textures. In $CVPR$ , pages 12663-12673, 2023. 5", + "[44] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In CVPR, pages 13492-13502, 2022. 2", + "[45] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, pages 405-421, 2020. 2, 3, 7", + "[46] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 conference papers, pages 1-8, 2022. 2", + "[47] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 2", + "[48] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 1, 2, 7, 8", + "[49] Yichen Ouyang, Wenhao Chai, Jiayi Ye, Dapeng Tao, Yibing Zhan, and Gaoang Wang. Chasing consistency in text-to-3d generation from a single image. arXiv preprint arXiv:2309.03599, 2023. 2", + "[50] Jangho Park, Gihyun Kwon, and Jong Chul Ye. Ed-nerf: Efficient text-guided editing of 3d scene using latent space nerf. arXiv preprint arXiv:2310.02712, 2023. 2", + "[51] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 32, 2019. 5", + "[52] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 1, 2, 3, 4, 5, 6, 7", + "[53] Zekun Qi, Muzhou Yu, Runpei Dong, and Kaisheng Ma. Vpp: Efficient conditional 3d generation via voxel-point progressive representation. arXiv preprint arXiv:2307.16605, 2023. 2", + "[54] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, et al. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. arXiv preprint arXiv:2306.17843, 2023. 2", + "[55] Sigal Raab, Inbal Leibovitch, Guy Tevet, Moab Arar, Amit H Bermano, and Daniel Cohen-Or. Single motion diffusion. arXiv preprint arXiv:2302.05905, 2023. 2", + "[56] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 2", + "[57] Amit Raj, Srinivas Kaza, Ben Poole, Michael Niemeyer, Nataniel Ruiz, Ben Mildenhall, Shiran Zada, Kfir Aberman, Michael Rubinstein, Jonathan Barron, et al. Dream" + ], + "bbox": [ + 501, + 90, + 890, + 902 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6805", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "booth3d: Subject-driven text-to-3d generation. arXiv preprint arXiv:2303.13508, 2023. 2", + "[58] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1, 5", + "[59] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, Marco Fumero, and Kamal Rahimi Malekshan. Clip-forge: Towards zero-shot text-to-shape generation. In CVPR, pages 18603-18613, 2022. 2", + "[60] Kyle Sargent, Zizhang Li, Tanmay Shah, Charles Herrmann, Hong-Xing Yu, Yunzhi Zhang, Eric Ryan Chan, Dmitry Lagun, Li Fei-Fei, Deqing Sun, et al. Zeronvs: Zero-shot 360-degree view synthesis from a single real image. arXiv preprint arXiv:2310.17994, 2023. 2", + "[61] Hoigi Seo, Hayeon Kim, Gwanghyun Kim, and Se Young Chun. Ditto-nerf: Diffusion-based iterative text to omnidirectional 3d model. arXiv preprint arXiv:2304.02827, 2023. 2", + "[62] Junyoung Seo, Wooseok Jang, Min-Seop Kwak, Jaehoon Ko, Hyeonsu Kim, Junho Kim, Jin-Hwa Kim, Jiyoung Lee, and Seungryong Kim. Let 2d diffusion model know 3d-consistency for robust text-to-3d generation. arXiv preprint arXiv:2303.07937, 2023. 2", + "[63] Yonatan Shafir, Guy Tevet, Roy Kapon, and Amit H Bermano. Human motion diffusion as a generative prior. arXiv preprint arXiv:2303.01418, 2023. 2", + "[64] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. NeurIPS, 34: 6087-6101, 2021. 2", + "[65] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. 2", + "[66] Yukai Shi, Jianan Wang, He Cao, Boshi Tang, Xianbiao Qi, Tianyu Yang, Yukun Huang, Shilong Liu, Lei Zhang, and Heung-Yeung Shum. Toss: High-quality text-guided novel view synthesis from a single image. arXiv preprint arXiv:2310.10644, 2023. 2", + "[67] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 1, 2, 8", + "[68] Liangchen Song, Liangliang Cao, Hongyu Xu, Kai Kang, Feng Tang, Junsong Yuan, and Yang Zhao. Roomdreamer: Text-driven 3d indoor scene synthesis with coherent geometry and texture. arXiv preprint arXiv:2305.11337, 2023. 2", + "[69] Jingxiang Sun, Bo Zhang, Ruizhi Shao, Lizhen Wang, Wen Liu, Zhenda Xie, and Yebin Liu. Dreamcraft3d: Hierarchical 3d generation with bootstrapped diffusion prior. arXiv preprint arXiv:2310.16818, 2023. 2", + "[70] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. 1, 2", + "[71] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity" + ], + "bbox": [ + 78, + 90, + 468, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "3d creation from a single image with diffusion prior. arXiv preprint arXiv:2303.14184, 2023. 2", + "[72] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 2, 5", + "[73] Christina Tsalicoglou, Fabian Manhardt, Alessio Tonioni, Michael Niemeyer, and Federico Tombari. Textmesh: Generation of realistic 3d meshes from text prompts. arXiv preprint arXiv:2304.12439, 2023. 2", + "[74] Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, Karsten Kreis, et al. Lion: Latent point diffusion models for 3d shape generation. NeurIPS, 35:10021-10039, 2022. 2", + "[75] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In CVPR, pages 3835-3844, 2022. 2", + "[76] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In CVPR, pages 12619-12629, 2023. 1, 2, 5", + "[77] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213, 2023. 1, 2, 5, 6", + "[78] Haohan Weng, Tianyu Yang, Jianan Wang, Yu Li, Tong Zhang, CL Chen, and Lei Zhang. Consistent123: Improve consistency for one image to 3d object synthesis. arXiv preprint arXiv:2310.08092, 2023. 2", + "[79] Zhenzhen Weng, Zeyu Wang, and Serena Yeung. Zero-)rator: Zero-shot 3d avatar generation from a single image. arXiv preprint arXiv:2305.16411, 2023. 2", + "[80] Jinbo Wu, Xiaobo Gao, Xing Liu, Zhengyang Shen, Chen Zhao, Haocheng Feng, Jingtuo Liu, and Errui Ding. Hdfusion: Detailed text-to-3d generation leveraging multiple noise estimation. arXiv preprint arXiv:2307.16183, 2023. 2", + "[81] Jiale Xu, Xintao Wang, Weihao Cheng, Yan-Pei Cao, Ying Shan, Xiaohu Qie, and Shenghua Gao. Dream3d: Zero-shot text-to-3d synthesis using 3d shape prior and text-to-image diffusion models. In CVPR, pages 20908–20918, 2023. 1, 2", + "[82] Jiayu Yang, Ziang Cheng, Yunfei Duan, Pan Ji, and Hongdong Li. Consistnet: Enforcing 3d consistency for multiview images diffusion. arXiv preprint arXiv:2310.10343, 2023. 2", + "[83] Jianglong Ye, Peng Wang, Kejie Li, Yichun Shi, and Heng Wang. Consistent-1-to-3: Consistent image to 3d view synthesis via geometry-aware diffusion models. arXiv preprint arXiv:2310.03020, 2023. 2", + "[84] Wang Yifan, Felice Serena, Shihao Wu, Cengiz Öztireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM Transactions on Graphics (TOG), 38(6):1-14, 2019. 3", + "[85] Chaohui Yu, Qiang Zhou, Jingliang Li, Zhe Zhang, Zhibin Wang, and Fan Wang. Points-to-3d: Bridging the gap be" + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "6806", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tween sparse points and shape-controllable text-to-3d generation. In ACM MM, pages 6841-6850, 2023. 2", + "[86] Huichao Zhang, Bowen Chen, Hao Yang, Liao Qu, Xu Wang, Li Chen, Chao Long, Feida Zhu, Kang Du, and Min Zheng. Avatarverse: High-quality & stable 3d avatar creation from text and pose. arXiv preprint arXiv:2308.03610, 2023. 6, 7", + "[87] Longwen Zhang, Qiwei Qiu, Hongyang Lin, Qixuan Zhang, Cheng Shi, Wei Yang, Ye Shi, Sibei Yang, Lan Xu, and Jingyi Yu. Dreamface: Progressive generation of animatable 3d faces under text guidance. arXiv preprint arXiv:2304.03117, 2023. 2", + "[88] Mingyuan Zhang, Xinying Guo, Liang Pan, Zhongang Cai, Fangzhou Hong, Huirong Li, Lei Yang, and Ziwei Liu. Remodiffuse: Retrieval-augmented motion diffusion model. arXiv preprint arXiv:2304.01116, 2023. 2", + "[89] Mengyi Zhao, Mengyuan Liu, Bin Ren, Shuling Dai, and Nicu Sebe. Modiff: Action-conditioned 3d motion generation with denoising diffusion probabilistic models. arXiv preprint arXiv:2301.03949, 2023. 2", + "[90] Minda Zhao, Chaoyi Zhao, Xinyue Liang, Lincheng Li, Zeng Zhao, Zhipeng Hu, Changjie Fan, and Xin Yu. Efficientdreamer: High-fidelity and robust 3d creation via orthogonal-view diffusion prior. arXiv preprint arXiv:2308.13223, 2023. 1, 2, 8", + "[91] Zixiang Zhou and Baoyuan Wang. Ude: A unified driving engine for human motion generation. In CVPR, pages 5632-5641, 2023. 2", + "[92] Joseph Zhu and Peiye Zhuang. Hifa: High-fidelity text-to-3d with advanced diffusion guidance. arXiv preprint arXiv:2305.18766, 2023. 2" + ], + "bbox": [ + 78, + 92, + 468, + 527 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "6807", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2024/GaussianDreamer_ Fast Generation from Text to 3D Gaussians by Bridging 2D and 3D Diffusion Models/68c13cd6-fa8d-4333-a8a0-12d60a7c6cac_model.json b/2024/GaussianDreamer_ Fast Generation from Text to 3D Gaussians by Bridging 2D and 3D Diffusion Models/68c13cd6-fa8d-4333-a8a0-12d60a7c6cac_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7d37c410ee7b5ef104c491f54b5da36dfbf58c84 --- /dev/null +++ b/2024/GaussianDreamer_ Fast Generation from Text to 3D Gaussians by Bridging 2D and 3D Diffusion Models/68c13cd6-fa8d-4333-a8a0-12d60a7c6cac_model.json @@ -0,0 +1,3480 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.168, + 0.131, + 0.804, + 0.177 + ], + "angle": 0, + "content": "GaussianDreamer: Fast Generation from Text to 3D Gaussians by Bridging 2D and 3D Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.197, + 0.203, + 0.773, + 0.24 + ], + "angle": 0, + "content": "Taoran Yi\\(^{1}\\), Jiemin Fang\\(^{2\\dagger}\\), Junjie Wang\\(^{2}\\), Guanjun Wu\\(^{3}\\), Lingxi Xie\\(^{2}\\), Xiaopeng Zhang\\(^{2}\\), Wenyu Liu\\(^{1}\\), Qi Tian\\(^{2}\\), Xinggang Wang\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.165, + 0.239, + 0.804, + 0.276 + ], + "angle": 0, + "content": "\\(^{1}\\)School of EIC, Huazhong University of Science and Technology \\(^{2}\\)Huawei Inc. \\(^{3}\\)School of CS, Huazhong University of Science and Technology" + }, + { + "type": "text", + "bbox": [ + 0.282, + 0.278, + 0.688, + 0.293 + ], + "angle": 0, + "content": "{taoranyi, guajuwu, liuwy, xgwang}@hust.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.134, + 0.295, + 0.832, + 0.31 + ], + "angle": 0, + "content": "{jaminfong, is.wangjunjie, 198808xc, zxphistory}@gmail.com tian.qil@huawei.com" + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.313, + 0.249, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.313, + 0.893, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.408, + 0.487, + 0.621, + 0.499 + ], + "angle": 0, + "content": "3D Gaussians with Temporal Variation" + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.5, + 0.893, + 0.542 + ], + "angle": 0, + "content": "Figure 1. We propose a simple yet efficient framework called GaussianDreamer. It bridges the 3D and 2D diffusion models via Gaussian splatting, having both 3D consistency and rich generation details. Our method can complete training within 15 minutes on a single GPU and achieve real-time rendering." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.545, + 0.314, + 0.56 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.577, + 0.473, + 0.865 + ], + "angle": 0, + "content": "In recent times, the generation of 3D assets from text prompts has shown impressive results. Both 2D and 3D diffusion models can help generate decent 3D objects based on prompts. 3D diffusion models have good 3D consistency, but their quality and generalization are limited as trainable 3D data is expensive and hard to obtain. 2D diffusion models enjoy strong abilities of generalization and fine generation, but 3D consistency is hard to guarantee. This paper attempts to bridge the power from the two types of diffusion models via the recent explicit and efficient 3D Gaussian splatting representation. A fast 3D object generation framework, named as GaussianDreamer, is proposed, where the 3D diffusion model provides priors for initialization and the 2D diffusion model enriches the geometry and appearance. Operations of noisy point growing and color perturbation are introduced to enhance the initialized Gaussians. Our GaussianDreamer can generate a high-quality 3D instance or 3D avatar within 15 minutes on one GPU, much faster than previous methods, while the gener" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.546, + 0.893, + 0.591 + ], + "angle": 0, + "content": "ated instances can be directly rendered in real time. Demos and code are available at https://taoranyi.com/gaussiandreamer/." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.619, + 0.633, + 0.635 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.764 + ], + "angle": 0, + "content": "3D asset generation has been an expensive and professional work in conventional pipelines. Recently, diffusion models [58] have achieved great success in creating high-quality and realistic 2D images. Many research works [2, 6, 9, 12, 15, 24, 32, 48, 52, 67, 70, 76, 77, 81, 90] try to transfer the power of 2D diffusion models to the 3D field for easing and assisting the process of 3D assets creation, e.g. the most common text-to-3D task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Here come two main streams for achieving this goal: (i) training a new diffusion model with 3D data [12, 15, 24, 48] (namely the 3D diffusion model) and (ii) lifting the 2D diffusion model to 3D [2, 6, 9, 32, 52, 58, 67, 70, 76, 77, 81, 90]. The former one is direct to implement and holds strong 3D consistency, but struggles to extend into a large generation domain as 3D data is usually hard and expensive to obtain. The scale of current 3D datasets is far smaller than 2D datasets. This results in the generated 3D assets falling" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.173, + 0.888 + ], + "angle": 0, + "content": "\\(\\dagger\\) Project lead." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.222, + 0.9 + ], + "angle": 0, + "content": "\\(\\ddagger\\) Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.222, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6796" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.212 + ], + "angle": 0, + "content": "short in dealing with complex text prompts and producing complex/fine geometry and appearance. The latter benefits from the large data domain of the 2D diffusion models, which can handle various text prompts and produce highly detailed and complex geometry and appearance. However, as 2D diffusion models are unaware of the camera view, the generated 3D assets are hard to form geometry consistency, especially for structure-complicated instances." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.213, + 0.473, + 0.544 + ], + "angle": 0, + "content": "This paper proposes to use recent 3D Gaussian Splating [25] to bridge the two aforementioned approaches, simultaneously having the geometry consistency from 3D diffusion models and rich details from 2D diffusion models. 3D Gaussians are one type of efficient and explicit representation, which intrinsically enjoys geometry priors due to the point-cloud-like structure. Specifically, we use one of two types of 3D diffusion models: text-to-3D and text-to-motion diffusion models, e.g. Shap-E [24] and MDM [72] in our implementation, to generate a coarse 3D instance. Based on the coarse 3D instance, a group of 3D Gaussians are initialized. We introduce two operations of noisy point growing and color perturbation to supplement the initialized Gaussians for follow-up enriching the 3D instance. Then the 3D Gaussians can be improved and optimized by interacting with the 2D diffusion model via the Score Distillation Sampling [52] (SDS) loss. Due to the geometry priors from both the 3D diffusion model and 3D Gaussian Splating itself, the training process can be finished in a very short time. The generated 3D asset can be rendered in real time without transformation into structures like mesh via the splatting process." + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.545, + 0.421, + 0.559 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.561, + 0.47, + 0.619 + ], + "angle": 0, + "content": "- We propose a text-to-3D method, named as Gaussian-Dreamer which bridges the 3D and 2D diffusion models via Gaussian splitting, enjoying both 3D consistency and rich generation details." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.62, + 0.469, + 0.664 + ], + "angle": 0, + "content": "- Noisy point growing and color perturbation are introduced to supplement the initialized 3D Gaussians for further content enrichment." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.665, + 0.469, + 0.725 + ], + "angle": 0, + "content": "- The overall method is simple and quite effective. A 3D instance can be generated within 15 minutes on one GPU, much faster than previous methods, and can be directly rendered in real time." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.561, + 0.47, + 0.725 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.74, + 0.227, + 0.755 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.901 + ], + "angle": 0, + "content": "3D Pretrained Diffusion Models. Recently, text-to-3D asset generation using diffusion models has achieved great success. Currently, it is mainly divided into lifting 2D diffusion models to 3D and 3D pretrained diffusion models, the difference lies in whether the training data used is 2D or 3D. 3D pretrained diffusion models [12, 15, 24, 48], referred to as 3D diffusion models in our paper, are models pretrained on text-3D pairs. After pretraining, they can generate 3D assets through only inference, and mod" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.273 + ], + "angle": 0, + "content": "els such as Point-E [48] and Shape-E [24] can generate 3D assets in minutes. In addition to generating 3D assets from text, there are methods where 3D diffusion models [1, 7, 10, 26, 55, 63, 72, 88, 89, 91] generate motion sequences based on text-motion data. By pretraining on text-motion pairs, these models can generate reasonable motion sequences for different texts. The generated motion sequences can be transformed into the SMPL (Skinned Multi-Person Linear) model [39] based on mesh representation, but texture information is not included. In our method, we can paint transformed SMPL by using different text prompts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.292, + 0.895, + 0.625 + ], + "angle": 0, + "content": "Lifting 2D Diffusion Models to 3D. In text-to-3D asset generation methods [8, 16, 21, 23, 40, 49, 50, 53, 57, 61, 62, 68, 73, 80, 85, 87, 92], in addition to using 3D pretrained diffusion models, lifting a 2D diffusion model to 3D is a training-free approach. Moreover, due to the abundance of 2D image data, this method produces assets with higher diversity and fidelity. Some single-image-to-3D methods [13, 19, 30, 33–36, 38, 54, 60, 65, 66, 71, 78, 79, 82, 83] also employ similar ideas. DreamFusion [52] first proposes SDS (Score Distillation Sampling) method, which is to update the 3D representation model using the 2D diffusion model. [76] proposes the method of SJC (Score Jacobian Chaining) to lift the 2D diffusion model to 3D. Later methods [6, 31, 32, 69, 77] build on DreamFusion and further improve the quality of 3D generation. Among them, generated 3D assets may suffer from multi-face problems. To address this issue, some methods strengthen the semantics of different views [2] and use multi-view information [67, 90] to alleviate such problems. There are also models [18, 22, 29, 44, 46, 59, 75, 81] which adopt CLIP [56] to align each view of the 3D representation model with the text." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.894, + 0.795 + ], + "angle": 0, + "content": "3D Representation Methods. In recent times, neural radiance fields (NeRF) [45] have achieved impressive results in 3D representation, and many methods in text-to-3D asset generation have also adopted NeRF or its variants [3, 47] as the representation method. Some methods [6, 12, 31, 32] use explicit optimization mesh representation methods like DMTET [64] to reduce rendering costs and further improve resolution. In addition to that, there are also generation methods that utilize point clouds [41, 48, 53, 74] and meshes [37] as 3D representations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Recently, 3D Gaussian Splatting [25] has been introduced as a representation method for 3D scenes, which can achieve rendering effects comparable to NeRF-based methods and enable real-time rendering. Two concurrent works [9, 70] also construct the 3D representation with 3D Gaussian Splatting [25]. DreamGaussian [70] uses a single image as a condition to generate 3D assets, while GS-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6797" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.087, + 0.895, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.27, + 0.895, + 0.342 + ], + "angle": 0, + "content": "Figure 2. Overall framework of GaussianDreamer. Firstly, we utilize a 3D diffusion model to generate the initialized point clouds. After executing noisy point growing and color perturbation on the point clouds, we use them to initialize the 3D Gaussians. The initialized 3D Gaussians are further optimized using the SDS method [52] with a 2D diffusion model. Finally, we render the image using the 3D Gaussians by employing 3D Gaussian Splatting [25]. We can use one of various 3D diffusion models to generate the initialized point clouds. In this case, we take text-to-3D and text-to-motion diffusion models as examples." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.348, + 0.473, + 0.455 + ], + "angle": 0, + "content": "GEN [9] implements high-quality generation from text to 3D. Our method shares a similar idea of using 3D Gaussian Splating as the representation method, which significantly reduces the cost of improving resolution and achieving a much faster optimization speed compared to estimizable mesh representation methods. And we can generate detailed 3D assets based on prompt texts in a very short time." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.474, + 0.17, + 0.489 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.503, + 0.471, + 0.61 + ], + "angle": 0, + "content": "In this section, we first review 2D and 3D diffusion models and the 3D representation method - 3D Gaussian Splating [25]. We give an overview of the whole framework in Sec. 3.2. Then, in Sec. 3.3, we describe the process of initializing the 3D Gaussians with the assistance of 3D diffusion models. The further optimization of 3D Gaussians using the 2D diffusion model is described in Sec. 3.4." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.632, + 0.219, + 0.646 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.473, + 0.903 + ], + "angle": 0, + "content": "DreamFusion. DreamFusion [52] is one of the most representative methods to lift 2D diffusion models to 3D, which proposes to optimize the 3D representation with the score distillation sampling (SDS) loss via a pre-trained 2D diffusion model \\(\\phi\\). Specifically, it takes MipNeRF [3] as the 3D representation method, whose parameters \\(\\theta\\) are optimized. Taking the rendering method as \\(g\\), the rendered image results in \\(\\mathbf{x} = g(\\theta)\\). To make the rendered image \\(\\mathbf{x}\\) similar to the samples obtained from the diffusion model \\(\\phi\\), DreamFusion uses a scoring estimation function: \\(\\hat{\\epsilon}_{\\phi}(\\mathbf{z}_t; y, t)\\), which predicts the sampled noise \\(\\hat{\\epsilon}_{\\phi}\\) given the noisy image \\(\\mathbf{z}_t\\), text embedding \\(y\\), and noise level \\(t\\). By measuring the difference between the Gaussian noise \\(\\epsilon\\) added to the rendered image \\(\\mathbf{x}\\) and the predicted noise \\(\\hat{\\epsilon}_{\\phi}\\), this scoring estimation function can provide the direction for updating the parameter \\(\\theta\\). The formula for computing" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.349, + 0.618, + 0.364 + ], + "angle": 0, + "content": "the gradient is as" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.377, + 0.895, + 0.426 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {S D S}} (\\phi , \\mathbf {x} = g (\\theta)) \\triangleq \\mathbb {E} _ {t, \\epsilon} \\left[ w (t) \\left(\\hat {\\epsilon} _ {\\phi} \\left(\\mathbf {z} _ {t}; y, t\\right) - \\epsilon\\right) \\frac {\\partial \\mathbf {x}}{\\partial \\theta} \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.441, + 0.738, + 0.457 + ], + "angle": 0, + "content": "where \\( w(t) \\) is a weighting function." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.478, + 0.892, + 0.629 + ], + "angle": 0, + "content": "3D Gaussian Splatting. 3D Gaussian Splatting [25] (3D-GS) is a recent groundbreaking method for novel-view synthesis. Unlike implicit representation methods such as NeRF [45], which renders images based on volume rendering, 3D-GS renders images through splatting [84], achieving real-time speed. Specifically, 3D-GS represents the scene through a set of anisotropic Gaussians, defined with its center position \\(\\mu \\in \\mathbb{R}^3\\), covariance \\(\\Sigma \\in \\mathbb{R}^7\\), color \\(c \\in \\mathbb{R}^3\\), and opacity \\(\\alpha \\in \\mathbb{R}^1\\). And the 3D Gaussians can be queried as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.644, + 0.892, + 0.663 + ], + "angle": 0, + "content": "\\[\nG (x) = e ^ {- \\frac {1}{2} (x) ^ {T} \\Sigma^ {- 1} (x)}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.679, + 0.894, + 0.77 + ], + "angle": 0, + "content": "where \\( x \\) represents the distance between \\( \\mu \\) and the query point. For computing the color of each pixel, it uses a typical neural point-based rendering [27, 28]. A ray \\( r \\) is cast from the center of the camera, and the color and density of the 3D Gaussians that the ray intersects are computed along the ray. The rendering process is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.785, + 0.892, + 0.827 + ], + "angle": 0, + "content": "\\[\nC (r) = \\sum_ {i \\in \\mathcal {N}} c _ {i} \\sigma_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\sigma_ {j}), \\quad \\sigma_ {i} = \\alpha_ {i} G (x _ {i}), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.9 + ], + "angle": 0, + "content": "where \\(\\mathcal{N}\\) represents the number of sample points on the ray \\(r\\), \\(c_{i}\\) and \\(\\alpha_{i}\\) denote the color and opacity of the i-th Gaussian, and \\(x_{i}\\) is the distance between the point and the i-th Gaussian." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6798" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.091, + 0.267, + 0.105 + ], + "angle": 0, + "content": "3.2. Overall Framework" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.114, + 0.473, + 0.355 + ], + "angle": 0, + "content": "Our overall framework consists of two parts, initialization with 3D diffusion model priors and optimization with the 2D diffusion model, as shown in Fig. 2. For initialization with 3D diffusion model priors, we use the 3D diffusion models \\( F_{3D} \\), instantiated with the text-to-3D and text-to-motion diffusion models, to generate the triangle mesh \\( m \\) based on the text prompt \\( y \\), which can be denoted as \\( m = F_{3D}(y) \\). One set of generated point clouds is transformed from the mesh \\( m \\). Then the 3D Gaussians \\( \\theta_b \\) are initialized via the generated point clouds after noisy point growing and color perturbation. For better quality, we utilize the 2D diffusion model \\( F_{2D} \\) to further optimize the initialized 3D Gaussians \\( \\theta_b \\) via SDS [52] with prompts \\( y \\), resulting in the final 3D Gaussians \\( \\theta_f \\). The target instance can be rendered in real time by splatting the generated Gaussians." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.363, + 0.47, + 0.394 + ], + "angle": 0, + "content": "3.3. Gaussian Initialization with 3D Diffusion Model Priors" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.402, + 0.47, + 0.522 + ], + "angle": 0, + "content": "In this section, we mainly discuss how to initialize the 3D Gaussians with 3D diffusion model priors. First, we use the 3D diffusion model \\( F_{3D} \\) to generate 3D assets based on the prompts \\( y \\). Then we convert the 3D assets into point clouds and use the transformed point clouds to initialize the 3D Gaussians. We have employed two types of 3D diffusion models to generate 3D assets. Below, we explain how to initialize the 3D Gaussians using each of these models." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.534, + 0.325, + 0.548 + ], + "angle": 0, + "content": "3.3.1 Text-to-3D Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.558, + 0.47, + 0.74 + ], + "angle": 0, + "content": "When using a text-based 3D generation model, generated 3D assets employ multi-layer perceptrons (MLPs) to predict SDF values and texture colors. To construct the triangle mesh \\( m \\), we query the SDF values at vertices along a regular grid with the size of \\( 128^3 \\) within the MLPs. Then we query the texture colors at each vertex of \\( m \\). We convert the vertices and colors of \\( m \\) into point clouds, denoted as \\( \\pmb{p}\\pmb{t}_m(\\pmb{p}_m,\\pmb{c}_m) \\). \\( \\pmb{p}_m\\in \\mathbb{R}^3 \\) refers to the position of the point clouds, which equals to the vertex coordinated of \\( m \\). \\( c_{m}\\in \\mathbb{R}^{3} \\) refers to the color of the point clouds, which is the same as the color of \\( m \\). However, the obtained colors \\( c_{m} \\) are relatively simple, and the positions \\( p_m \\) are sparse." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Noisy Point Growing and Color Perturbation. We do not use the generated point clouds \\( \\pmb{pt}_m \\) to directly initialize the 3D Gaussians. To improve the quality of initialization, we perform noisy point growing and color perturbation around point clouds \\( \\pmb{pt}_m \\). First, we compute the bounding box (BBox) of the surface on \\( \\pmb{pt}_m \\) and then uniformly grow point clouds \\( \\pmb{pt}_r(\\pmb{p}_r, \\pmb{c}_r) \\) within the BBox. \\( \\pmb{p}_r \\) and \\( \\pmb{c}_r \\) represent the positions and colors of \\( \\pmb{pt}_r \\). To enable fast searching, we construct a KDTree [4] \\( K_m \\) using the positions \\( \\pmb{p}_m \\). Based on the distances between the positions \\( \\pmb{p}_r \\)" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.087, + 0.885, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.244, + 0.892, + 0.286 + ], + "angle": 0, + "content": "Figure 3. The process of noisy point growing and color perturbation. \"Grow&Pertb.\" denotes noisy point growing and color perturbation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.297, + 0.892, + 0.373 + ], + "angle": 0, + "content": "and the nearest points found in the KDTree \\(K_{m}\\), we determine which points to keep. In this process, we select points within the (normalized) distance of 0.01. For the noisy point clouds, we make their colors \\(\\pmb{c}_r\\) similar to \\(\\pmb{c}_m\\), and also add some perturbations:" + }, + { + "type": "equation", + "bbox": [ + 0.647, + 0.387, + 0.892, + 0.401 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {c} _ {r} = \\boldsymbol {c} _ {m} + \\boldsymbol {a}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.413, + 0.892, + 0.459 + ], + "angle": 0, + "content": "where the values of \\( \\mathbf{a} \\) are randomly sampled between 0 and 0.2. We merge the positions and colors of \\( \\mathbf{p}\\mathbf{t}_{m} \\) and \\( \\mathbf{p}\\mathbf{t}_{r} \\) to obtain the final point clouds." + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.472, + 0.892, + 0.487 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {p} \\boldsymbol {t} \\left(\\boldsymbol {p} _ {f}, \\boldsymbol {c} _ {f}\\right) = \\left(\\boldsymbol {p} _ {m} \\oplus \\boldsymbol {p} _ {r}, \\boldsymbol {c} _ {m} \\oplus \\boldsymbol {c} _ {r}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.499, + 0.892, + 0.62 + ], + "angle": 0, + "content": "where \\(\\oplus\\) is the concatenation operation. Fig. 3 illustrates the process of noisy point growing and color perturbation. Finally, we initialize the positions \\(\\mu_{b}\\) and colors \\(c_{b}\\) of the 3D Gaussians \\(\\theta_{b}(\\mu_{b}, c_{b}, \\Sigma_{b}, \\alpha_{b})\\) using both the positions \\(\\pmb{p}_{f}\\) and colors \\(\\pmb{c}_{f}\\) of final point clouds \\(\\pmb{pt}\\). The opacity \\(\\alpha_{b}\\) of the 3D Gaussians is initialized to 0.1, and the covariance \\(\\Sigma_{b}\\) is calculated as the distance between the nearest two points. Algorithm 1 shows the specific algorithm flowchart." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.626, + 0.776, + 0.639 + ], + "angle": 0, + "content": "3.3.2 Text-to-Motion Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.649, + 0.892, + 0.831 + ], + "angle": 0, + "content": "We generate a sequence of human body motions using text and select a human pose that best matches the given text. We then convert the keypoints of this human pose into the SMPL model [39], which is represented by a triangle mesh \\( m \\). We then convert the mesh \\( m \\) into point clouds \\( \\pmb{pt}_m(\\pmb{p}_m, \\pmb{c}_m) \\), where the position \\( \\pmb{p}_m \\) of each point in the point clouds corresponds to the vertices of \\( m \\). As for the color \\( \\pmb{c}_m \\) of \\( \\pmb{pt}_m \\), since the SMPL model used here does not have textures, we randomly initialize \\( \\pmb{c}_m \\). To move \\( \\pmb{pt}_m \\) near the origin, we calculate the center point \\( \\pmb{p}_c \\in \\mathbb{R}^3 \\) of \\( \\pmb{p}_m \\) and subtract the position of point clouds \\( \\pmb{pt}_m \\) from the center point \\( \\pmb{p}_c \\)." + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.844, + 0.892, + 0.86 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {p} \\boldsymbol {t} \\left(\\boldsymbol {p} _ {f}, \\boldsymbol {c} _ {f}\\right) = \\boldsymbol {p} \\boldsymbol {t} _ {m} \\left(\\boldsymbol {p} _ {m} - \\boldsymbol {p} _ {c}, \\boldsymbol {c} _ {m}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Finally, we use the point clouds \\( pt \\) to initialize the 3D Gaussians, similar to what is described in Sec. 3.3.1. To improve" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6799" + } + ], + [ + { + "type": "algorithm", + "bbox": [ + 0.078, + 0.091, + 0.47, + 0.183 + ], + "angle": 0, + "content": "Algorithm 1 The 3D Gaussian Initialization. \n\\(\\pmb{pt}_m(\\pmb{p}_m, \\pmb{c}_m)\\): Point clouds generated from \\(F_{3D}\\). \n\\(\\pmb{pt}_r(\\pmb{p}_r, \\pmb{c}_r)\\): Growing point clouds within the BBox. \n\\(\\pmb{pt}(\\pmb{p}_f, \\pmb{c}_f)\\): Point clouds used for initializing the 3D Gaussians. \n\\(\\theta_b(\\mu_b, c_b, \\Sigma_b, \\alpha_b)\\): Initialized 3D Gaussians." + }, + { + "type": "algorithm", + "bbox": [ + 0.094, + 0.187, + 0.47, + 0.292 + ], + "angle": 0, + "content": "Stage 1: Generate points in the BBox. \n\\(K_{m} \\leftarrow\\) BuildKDTree(\\(p_{m}\\)) \n\\(\\text{BBox} \\leftarrow p_{m}\\) \n\\(\\text{BBox} \\leftarrow\\) Positions bounding box. \nLow, High \\(\\leftarrow\\) BBox.MinBound, BBox.MaxBound \n\\(\\text{BBox} \\leftarrow\\) Boundary of bounding box. \n\\(\\boldsymbol{ps}_{u} \\leftarrow\\) Uniform(Low, High, size = (NumPoints, 3)) \n\\(\\triangleright\\) Points in the BBox." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.293, + 0.469, + 0.321 + ], + "angle": 0, + "content": "Stage 2: Keep the points that meet the distance requirement." + }, + { + "type": "algorithm", + "bbox": [ + 0.094, + 0.323, + 0.469, + 0.443 + ], + "angle": 0, + "content": "\\(\\pmb{p}_r,\\pmb {c}_r = \\left[ \\begin{array}{ll} \\end{array} \\right],\\left[ \\begin{array}{ll} \\end{array} \\right]\\) \nfor all \\(\\pmb{p_u}\\) in \\(p s_{u}\\) do \\(\\pmb{p}_{un},i\\gets K_m.\\mathrm{SearchNearest}(\\pmb {p}_u)\\) Nearest point and its index in \\(K_{m}\\) if \\(|\\pmb {p}_{un} - \\pmb {p}_u| < 0.01\\) then \\(\\pmb{p}_r\\) .append \\((\\pmb {p}_u)\\) \\(c_{r}\\) .append \\((c_{m}[i] + 0.2\\times \\mathrm{Random}(\\mathrm{size} = 3))\\) \\(\\triangleright\\) Color of the nearest point plus perturbation." + }, + { + "type": "algorithm", + "bbox": [ + 0.094, + 0.444, + 0.469, + 0.594 + ], + "angle": 0, + "content": "end if \nend for \n\\(\\pmb {p}_f\\gets \\pmb {p}_m\\oplus \\pmb {p}_r\\) \n\\(c_{f}\\gets c_{m}\\oplus c_{r}\\) \nStage 3: Initialize the 3D Gaussians. \n\\(\\mu_b,c_b\\gets p_f,c_f\\) \n\\(\\triangleright\\) Positions and colors of the 3D Gaussians. \n\\(D\\leftarrow \\mu_{b}\\quad \\triangleright\\) Distance between the nearest two positions. \n\\(\\Sigma_{b},\\alpha_{b}\\gets D,0.1\\) \n\\(\\triangleright\\) Covariance and opacity of the 3D Gaussians." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.622, + 0.47, + 0.699 + ], + "angle": 0, + "content": "the generation of motion sequences, we simplify the text by retaining only the relevant parts related to the motion and add a subject. For example, if the text prompt is \"Iron man kicks with his left leg\", we transform it into \"Someone kicks with the left leg\" when generating the motion sequences." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.71, + 0.438, + 0.727 + ], + "angle": 0, + "content": "3.4. Optimization with the 2D Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.902 + ], + "angle": 0, + "content": "To enrich details and improve the quality of the 3D asset, we optimize the 3D Gaussians \\(\\theta_{b}\\) with a 2D diffusion model \\(F_{2D}\\) after initializing them with 3D diffusion model priors. We employ the SDS (Score Distillation Sampling) loss to optimize the 3D Gaussians. First, we use the method of 3D Gaussian Splatting [25] to obtain the rendered image \\(\\mathbf{x} = g(\\theta_{i})\\). Here, \\(g\\) represents the splatting rendering method as in Eq. 3. Then, we use Eq. 1 to calculate the gradients for updating the Gaussian parameters \\(\\theta_{i}\\) with the 2D diffusion model \\(F_{2D}\\). After a short optimization period using the 2D diffusion model \\(F_{2D}\\), the final generated 3D" + }, + { + "type": "table_caption", + "bbox": [ + 0.54, + 0.089, + 0.853, + 0.104 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons on \\( {\\mathrm{T}}^{3} \\) Bench [17]." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.115, + 0.91, + 0.216 + ], + "angle": 0, + "content": "
MethodTime†Single Obj.Single w/ Surr.Multi Obj.Average
SJC [76]-24.719.811.718.7
DreamFusion [52]6 hours24.424.616.121.7
Fantasia3D [6]6 hours26.427.018.524.0
LatentNeRF [43]15 minutes33.130.620.628.1
Magic3D [32]5.3 hours37.035.425.732.7
ProlificDreamer [77]~10 hours49.444.835.843.3
Ours15 minutes54.048.634.545.7
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.525, + 0.217, + 0.868, + 0.229 + ], + "angle": 0, + "content": "Scores above are average of two metrics (quality and alignment)." + }, + { + "type": "table_footnote", + "bbox": [ + 0.527, + 0.229, + 0.714, + 0.242 + ], + "angle": 0, + "content": "\\(\\dagger\\) GPU time counted in their papers." + }, + { + "type": "list", + "bbox": [ + 0.525, + 0.217, + 0.868, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.254, + 0.892, + 0.285 + ], + "angle": 0, + "content": "instance \\(\\theta_{f}\\) achieves high quality and fidelity on top of the 3D consistency provided by the 3D diffusion model \\(F_{3D}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.298, + 0.633, + 0.315 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.323, + 0.892, + 0.428 + ], + "angle": 0, + "content": "In this section, we first present the implementation details in Sec. 4.1. In Sec. 4.2, we show the quantitative comparison results. Then, in Sec. 4.3, we showcase the visualization results of our method and compare them with other methods. In Sec. 4.4, we conduct a series of ablation experiments to validate the effectiveness of our method. Finally, we discuss the limitations of our method." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.438, + 0.719, + 0.453 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.461, + 0.893, + 0.823 + ], + "angle": 0, + "content": "Our method is implemented in PyTorch [51], based on ThreeStudio [14]. The 3D diffusion models used in our method are Shap-E [24] and MDM [72], and we load the Shap-E model finetuned on Objaverse [11] in Cap3D [42]. For the 2D diffusion model, we use stabilityai/stable-diffusion-2-1-base [58], with a guidance scale of 100. The timestamps we use are uniformly sampled from 0.02 to 0.98 before 500 iterations, and change to 0.02 to 0.55 after 500 iterations. For the 3D Gaussians, the learning rates of opacity \\(\\alpha\\) and position \\(\\mu\\) are \\(10^{-2}\\) and \\(5 \\times 10^{-5}\\). The color \\(c\\) of the 3d Gaussians is represented by the sh coefficient, with the degree set to 0 and the learning rate set to \\(1.25 \\times 10^{-2}\\). The covariance of the 3D Gaussians is converted into scaling and rotation for optimization, with learning rates of \\(10^{-3}\\) and \\(10^{-2}\\), respectively. The radius of the camera we use for rendering is from 1.5 to 4.0, with the azimuth in the range of -180 to 180 degrees and the elevation in the range of -10 to 60 degrees. The total training iterations are 1200. All our experiments can be completed within 15 minutes on a single RTX 3090 with a batch size of 4. The resolution we use for rendering is \\(1024 \\times 1024\\), which is scaled to 512 \\(\\times\\) 512 when optimizing using the 2D diffusion model. We can render in real time at \\(512 \\times 512\\) resolution. And all our code will be released." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.833, + 0.722, + 0.849 + ], + "angle": 0, + "content": "4.2. Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We evaluate quality and consistency following T3Bench [17], which provides a comprehensive benchmark for text-to-3D generation. Three text categories are" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6800" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.127, + 0.089, + 0.204, + 0.111 + ], + "angle": 0, + "content": "DreamFusion 6.0 Hours" + }, + { + "type": "image_caption", + "bbox": [ + 0.304, + 0.089, + 0.356, + 0.11 + ], + "angle": 0, + "content": "Magic3D 5.3 Hours" + }, + { + "type": "image_caption", + "bbox": [ + 0.452, + 0.089, + 0.518, + 0.111 + ], + "angle": 0, + "content": "Fantasia3D 6.0 Hours" + }, + { + "type": "image_caption", + "bbox": [ + 0.609, + 0.089, + 0.701, + 0.111 + ], + "angle": 0, + "content": "ProlificDreamer ~Hours" + }, + { + "type": "image_caption", + "bbox": [ + 0.731, + 0.09, + 0.871, + 0.111 + ], + "angle": 0, + "content": "GaussianDreamer (Ours) 15 Minutes" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.119, + 0.231, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.261, + 0.127, + 0.395, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.409, + 0.121, + 0.55, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.576, + 0.115, + 0.723, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.731, + 0.125, + 0.807, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.812, + 0.126, + 0.886, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.375, + 0.174, + 0.632, + 0.187 + ], + "angle": 0, + "content": "A plate piled high with chocolate chip cookies." + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.195, + 0.159, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.162, + 0.196, + 0.232, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.188, + 0.319, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.332, + 0.189, + 0.397, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.413, + 0.188, + 0.481, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.494, + 0.188, + 0.567, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.574, + 0.19, + 0.719, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.727, + 0.191, + 0.807, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.813, + 0.194, + 0.88, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.346, + 0.25, + 0.66, + 0.262 + ], + "angle": 0, + "content": "A 3D model of an adorable cottage with a thatched roof." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.265, + 0.893, + 0.307 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparisons between our method and DreamFusion [52], Magic3D [32], Fantasy3D [6] and ProlificDreamer [77]. Here we count the GPU time in their papers. The time for DreamFusion is measured on TPUv4, Magic3D is measured on A100, Fantasy3D is measured on RTX 3090, and our method is measured on RTX 3090." + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.316, + 0.243, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.176, + 0.389, + 0.212, + 0.403 + ], + "angle": 0, + "content": "A fox" + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.317, + 0.295, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.419, + 0.209, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.423, + 0.323, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.495, + 0.281, + 0.508 + ], + "angle": 0, + "content": "An amigurumi motorcycle..." + }, + { + "type": "image", + "bbox": [ + 0.346, + 0.319, + 0.447, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.452, + 0.323, + 0.586, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.607, + 0.333, + 0.732, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.751, + 0.326, + 0.834, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.352, + 0.39, + 0.562, + 0.402 + ], + "angle": 0, + "content": "Magic gun, game asset, mistery..." + }, + { + "type": "image_caption", + "bbox": [ + 0.582, + 0.389, + 0.876, + 0.403 + ], + "angle": 0, + "content": "An opulent couch from the palace of Versailles" + }, + { + "type": "image", + "bbox": [ + 0.4, + 0.402, + 0.51, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.614, + 0.413, + 0.706, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.415, + 0.822, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.339, + 0.495, + 0.545, + 0.508 + ], + "angle": 0, + "content": "Magic dagger, mistery, ancient..." + }, + { + "type": "image_caption", + "bbox": [ + 0.566, + 0.495, + 0.892, + 0.509 + ], + "angle": 0, + "content": "A zoomed out DSLR photo of a lion's mane jellyfish" + }, + { + "type": "image_caption", + "bbox": [ + 0.191, + 0.513, + 0.778, + 0.527 + ], + "angle": 0, + "content": "Figure 5. More generated samples by our GaussianDreamer. Two views of each sample are shown." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.534, + 0.47, + 0.595 + ], + "angle": 0, + "content": "designed for 3D generation with increasing complexity – single objects, single objects with surroundings, and multi objects. In Tab. 1, our method outperforms the compared methods while enjoying a short generation time." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.607, + 0.275, + 0.621 + ], + "angle": 0, + "content": "4.3. Visualization Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.631, + 0.47, + 0.676 + ], + "angle": 0, + "content": "In this section, we present the results of initializing the 3D Gaussians with two different 3D diffusion models: text-to-3D and text-to-motion diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Initialization with Text-to-3D Diffusion Model. We show the comparison results with DreamFusion [52], Magic3D [32], Fantasy3D [6] and ProlificDreamer [77] in Fig. 4. In addition to our method, the figures of other methods are downloaded from the paper of ProlificDreamer. When encountering prompts that involve the combination of multiple objects, such as the prompt \"A plate piled high with chocolate chip cookies\", the generated results from Magic3D, Fantasy3D, and ProlificDreamer do not include a plate. In contrast, our generated result can effectively combine a plate and chocolate chip cookies. Furthermore, compared to DreamFusion, the plate we generate has better patterns. Our method shows comparable quality while saving 21 - 24 times the generation time compared to their" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.534, + 0.894, + 0.626 + ], + "angle": 0, + "content": "methods. Moreover, the 3D Gaussians generated by our method can directly achieve real-time rendering without further transformation into mesh-like structures. Fig. 5 visualizes more samples generated by our GaussianDreamer from various prompts, which show good 3D consistency while having high-quality details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Initialization with Text-to-Motion Diffusion Model. In Fig. 6, we present the comparison results with DreamFusion [52], DreamAvatar [5], DreamWaltz [20], and AvatarVerse [86]. In addition to our method, the figures of other methods are downloaded from the paper of AvatarVerse. It is worth noting that our prompt is \"Spiderman/Stormtrooper stands with open arms\", while the prompts for other methods are \"Spiderman/Stormtrooper\". This is because when generating motion using text-to-motion diffusion model as initialization, we require more specific action descriptions. Our method achieves a speedup of 4-24 times compared to other methods, while maintaining comparable quality. Additionally, our method allows for generating 3D avatars with specified body poses. In Fig. 7, we provide more results generated with different human body poses. We first generate a sequence of motions that match the text prompt using a text-to-motion 3d diffusion model, and then initialize the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "6801" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.088, + 0.252, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.089, + 0.453, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.461, + 0.089, + 0.635, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.089, + 0.758, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.762, + 0.089, + 0.892, + 0.11 + ], + "angle": 0, + "content": "GaussianDreamer (Ours) 15 Minutes" + }, + { + "type": "image_caption", + "bbox": [ + 0.451, + 0.233, + 0.512, + 0.244 + ], + "angle": 0, + "content": "Spiderman" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.242, + 0.246, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.26, + 0.244, + 0.449, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.462, + 0.244, + 0.632, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.256, + 0.755, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.261, + 0.871, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.447, + 0.371, + 0.521, + 0.383 + ], + "angle": 0, + "content": "Stormtrooper" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.384, + 0.891, + 0.398 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparisons between our method and DreamFusion [52], DreamAvatar [5], DreamWaltz [20], and AvatarVerse [86]." + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.401, + 0.472, + 0.672 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.673, + 0.47, + 0.728 + ], + "angle": 0, + "content": "Figure 7. More generated 3D avatars by our GaussianDreamer initialized with the different poses of SMPL [39]. Here, the different poses of SMPL are generated using a text-to-motion diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.737, + 0.47, + 0.784 + ], + "angle": 0, + "content": "3d Gaussians with the SMPL of one selected pose in motions. Our method can generate 3D avatars in any desired pose." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.334, + 0.803 + ], + "angle": 0, + "content": "4.4. Ablation Study and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.903 + ], + "angle": 0, + "content": "The Role of Initialization. As shown in Fig. 8, we first conduct an ablation experiment on the initialization of the 3D Gaussians to validate that initialization can improve 3D consistency. The first column is the rendering result of Shap-E [24] with NeRF [45] as the 3D representation. The second column is the result of optimizing the 3D Gaussians" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.408, + 0.895, + 0.651 + ], + "angle": 0, + "content": "randomly initialized within a cube using the SDS loss, and the third column is our method. We show the initialization effects on 3 samples. In the first and second rows, ShapE has good generation results while our method provides more complex geometries and more realistic appearances. Compared with random initialization, in the first row, details of our method are better. In the second row, the 3D assets generated by random initialization have the multi-head problem, which does not occur in our method. The initialization of the 3D diffusion model can avoid unreasonable geometry. In the third row, the generation result of Shap-E is far different from the given text prompt while our method makes the 3D assets closer to the prompt through the 2D diffusion model. Our method can expand the domain of Shap-E prompts, allowing for the generation of 3D assets based on a wider range of prompts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.662, + 0.894, + 0.799 + ], + "angle": 0, + "content": "Noisy Point Growing and Color Perturbation. Fig. 9 illustrates the ablation results of noisy point growing and color perturbation. With noisy point growing and color perturbation applied, the first row showcases improved details in the sniper rifle. Additionally, the second column generates an amigurumi motorcycle that better aligns with the style characteristics of amigurumi mentioned in the prompt, compared to the case without noisy point growing and color perturbation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Initialization with Different Text-to-3D Diffusion Models. We select two text-to-3D generation models, ShapeE [24] and Point-E [48], to validate the effectiveness of our framework. We load the Point-E model finetuned on Objaverse [11] in Cap3D[42]. Fig. 10, we showcase the generated results after initializing the 3D Gaussians using" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6802" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.194, + 0.088, + 0.241, + 0.101 + ], + "angle": 0, + "content": "Shap-E" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.103, + 0.351, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.417, + 0.089, + 0.585, + 0.102 + ], + "angle": 0, + "content": "Ours w/o Point Cloud Prior" + }, + { + "type": "image", + "bbox": [ + 0.359, + 0.112, + 0.481, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.113, + 0.629, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.679, + 0.088, + 0.832, + 0.102 + ], + "angle": 0, + "content": "GaussianDreamer (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.11, + 0.756, + 0.185 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.109, + 0.89, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.366, + 0.189, + 0.63, + 0.203 + ], + "angle": 0, + "content": "A stack of pancakes covered in maple syrup" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.207, + 0.353, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.385, + 0.208, + 0.495, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.213, + 0.621, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.217, + 0.755, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.759, + 0.218, + 0.871, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.31, + 0.374, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.388, + 0.307, + 0.495, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.307, + 0.622, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.623, + 0.312, + 0.75, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.754, + 0.312, + 0.885, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.467, + 0.384, + 0.528, + 0.397 + ], + "angle": 0, + "content": "A lionfish" + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.41, + 0.846, + 0.424 + ], + "angle": 0, + "content": "Figure 8. Ablation studies of the initialization of the 3D Gaussians. The Shap-E [24] rendering resolution here is \\(256 \\times 256\\)." + }, + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.43, + 0.264, + 0.441 + ], + "angle": 0, + "content": "Ours w/o Grow&Pertb." + }, + { + "type": "image_caption", + "bbox": [ + 0.323, + 0.43, + 0.461, + 0.44 + ], + "angle": 0, + "content": "GaussianDreamer (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.443, + 0.469, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.197, + 0.469, + 0.375, + 0.478 + ], + "angle": 0, + "content": "Sniper rifle, asset, scifi, cyberpunk..." + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.48, + 0.457, + 0.524 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.219, + 0.524, + 0.354, + 0.532 + ], + "angle": 0, + "content": "An amigurumi motorcycle..." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.535, + 0.47, + 0.576 + ], + "angle": 0, + "content": "Figure 9. Ablation studies of noisy point growing and color perturbation. \"Grow&Pertb.\" denotes noisy point growing and color perturbation." + }, + { + "type": "image_caption", + "bbox": [ + 0.124, + 0.583, + 0.247, + 0.593 + ], + "angle": 0, + "content": "Initialization with Point-E" + }, + { + "type": "image_caption", + "bbox": [ + 0.291, + 0.583, + 0.411, + 0.593 + ], + "angle": 0, + "content": "Initialization with Shap-E" + }, + { + "type": "image", + "bbox": [ + 0.09, + 0.594, + 0.44, + 0.63 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.2, + 0.629, + 0.345, + 0.639 + ], + "angle": 0, + "content": "An airplane made out of wood" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.64, + 0.468, + 0.677 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.176, + 0.679, + 0.374, + 0.69 + ], + "angle": 0, + "content": "Flamethrower, with fire, scif, cyberpunk..." + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.692, + 0.469, + 0.72 + ], + "angle": 0, + "content": "Figure 10. Ablation studies of initialization with different text-to-3D diffusion models: Point-E [48] and Shap-E [24]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.731, + 0.47, + 0.838 + ], + "angle": 0, + "content": "one of two text-to-3D generation models. It can be seen that both initializations yield good generation results. However, considering that Shap-E generates 3D assets based on NeRF and SDF, which provide higher fidelity compared to the point cloud representation used by Point-E, the geometry of the airplane in the first row of Fig. 10 appears better when initialized using Shap-E." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.847, + 0.204, + 0.861 + ], + "angle": 0, + "content": "4.5. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "The edges of the 3D assets generated by our method are not always sharp, and there may be unnecessary 3D Gaussians" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.432, + 0.895, + 0.599 + ], + "angle": 0, + "content": "around the object surface. How to filter these point clouds will be a possible direction for improvement. Our approach utilizes 3D diffusion model priors, which greatly alleviates the issue of multi-face problems. However, there is still a small chance of encountering the multi-face problems in scenes where there is minimal geometric difference but significant appearance discrepancy between the front and back of objects such as a backpack. Utilizing 3D-aware diffusion models [35, 67, 90] may be able to solve this problem. Additionally, our method has limited effectiveness in generating large-scale scenes, such as indoor scenes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.628, + 0.62, + 0.644 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.894, + 0.812 + ], + "angle": 0, + "content": "We propose a fast text-to-3D method GaussianDreamer by bridging the abilities of 3D and 2D diffusion models via the Gaussian splatting representation. GaussianDreamer can generate detailed and realistic geometry and appearance while maintaining 3D consistency. The 3D diffusion model priors and geometry priors from the 3D Gaussians effectively promote the convergence speed. Each sample can be generated within 15 minutes on one GPU. We believe the approach of bridging 3D and 2D diffusion models could be a promising direction to generate 3D assets efficiently." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.84, + 0.66, + 0.858 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This work was supported by the National Natural Science Foundation of China (No. 62376102)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6803" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Tenglong Ao, Zeyi Zhang, and Libin Liu. Gesture diffuclip: Gesture diffusion model with clip latents. arXiv preprint arXiv:2303.14613, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.158, + 0.472, + 0.227 + ], + "angle": 0, + "content": "[2] Mohammadreza Armandpour, Huangjie Zheng, Ali Sadeghian, Amir Sadeghian, and Mingyuan Zhou. Re-imagine the negative prompt algorithm: Transform 2d diffusion into 3d, alleviate janus problem and beyond. arXiv preprint arXiv:2304.04968, 2023.1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.471, + 0.284 + ], + "angle": 0, + "content": "[3] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, pages 5855-5864, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.285, + 0.471, + 0.325 + ], + "angle": 0, + "content": "[4] Jon Louis Bentley. Multidimensional binary search trees used for associative searching. Communications of the ACM, 18(9):509-517, 1975. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.327, + 0.471, + 0.381 + ], + "angle": 0, + "content": "[5] Yukang Cao, Yan-Pei Cao, Kai Han, Ying Shan, and KwanYee K Wong. Dreamavatar: Text-and-shape guided 3d human avatar generation via diffusion models. arXiv preprint arXiv:2304.00916, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.383, + 0.471, + 0.437 + ], + "angle": 0, + "content": "[6] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. *Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation.* arXiv preprint arXiv:2303.13873, 2023. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.439, + 0.471, + 0.493 + ], + "angle": 0, + "content": "[7] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In CVPR, pages 18000-18010, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.495, + 0.471, + 0.549 + ], + "angle": 0, + "content": "[8] Yiwen Chen, Chi Zhang, Xiaofeng Yang, Zhongang Cai, Gang Yu, Lei Yang, and Guosheng Lin. It3d: Improved text-to-3d generation with explicit view synthesis. arXiv preprint arXiv:2308.11473, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.551, + 0.471, + 0.592 + ], + "angle": 0, + "content": "[9] Zilong Chen, Feng Wang, and Huaping Liu. Text-to-3d using gaussian splatting. arXiv preprint arXiv:2309.16585, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.594, + 0.47, + 0.649 + ], + "angle": 0, + "content": "[10] Rishabh Dabral, Muhammad Hamza Mughal, Vladislav Golyanik, and Christian Theobalt. Mofusion: A framework for denoising-diffusion-based motion synthesis. In CVPR, pages 9760-9770, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.47, + 0.718 + ], + "angle": 0, + "content": "[11] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In CVPR, pages 13142-13153, 2023. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.47, + 0.788 + ], + "angle": 0, + "content": "[12] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. NeurIPS, 35:31841-31854, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.47, + 0.857 + ], + "angle": 0, + "content": "[13] Jiatao Gu, Alex Trevithick, Kai-En Lin, Joshua M Susskind, Christian Theobalt, Lingjie Liu, and Ravi Ramamoorthi. Nerfdiff: Single-image view synthesis with nerf-guided distillation from 3d-aware diffusion. In ICML, pages 11808–11826. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[14] Yuan-Chen Guo, Ying-Tian Liu, Ruizhi Shao, Christian Laforte, Vikram Voleti, Guan Luo, Chia-Hao Chen, Zi-Xin Zou, Chen Wang, Yan-Pei Cao, and Song-Hai Zhang." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "threestudio: A unified framework for 3d content generation. https://github.com/threestudio-project/threestudio, 2023.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[15] Anchit Gupta, Wenhan Xiong, Yixin Nie, Ian Jones, and Barlas Oğuz. 3dgen: Triplane latent diffusion for textured mesh generation. arXiv preprint arXiv:2303.05371, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[16] Xiao Han, Yukang Cao, Kai Han, Xiatian Zhu, Jiankang Deng, Yi-Zhe Song, Tao Xiang, and Kwan-Yee K Wong. Headsculpt: Crafting 3d head avatars with text. In NeurIPS, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.235, + 0.892, + 0.291 + ], + "angle": 0, + "content": "[17] Yuze He, Yushi Bai, Matthieu Lin, Wang Zhao, Yubin Hu, Jenny Sheng, Ran Yi, Juanzi Li, and Yong-Jin Liu. T 3 bench: Benchmarking current progress in text-to-3d generation. arXiv preprint arXiv:2310.02977, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.293, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[18] Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. arXiv preprint arXiv:2205.08535, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[19] Shoukang Hu, Fangzhou Hong, Tao Hu, Liang Pan, Haiyi Mei, Weiye Xiao, Lei Yang, and Ziwei Liu. Humanliff: Layer-wise 3d human generation with diffusion model. arXiv preprint arXiv:2308.09712, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.406, + 0.892, + 0.462 + ], + "angle": 0, + "content": "[20] Yukun Huang, Jianan Wang, Ailing Zeng, He Cao, Xianbiao Qi, Yukai Shi, Zheng-Jun Zha, and Lei Zhang. Dreamwaltz: Make a scene with complex 3d animatable avatars. arXiv preprint arXiv:2305.12529, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.463, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[21] Yangyi Huang, Hongwei Yi, Yuliang Xiu, Tingting Liao, Ji-axiang Tang, Deng Cai, and Justus Thies. Tech: Text-guided reconstruction of lifelike clothed humans. arXiv preprint arXiv:2308.08545, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.519, + 0.892, + 0.572 + ], + "angle": 0, + "content": "[22] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In CVPR, pages 867-876, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.576, + 0.892, + 0.644 + ], + "angle": 0, + "content": "[23] Ruixiang Jiang, Can Wang, Jingbo Zhang, Mengei Chai, Mingming He, Dongdong Chen, and Jing Liao. Avatarcraft: Transforming text into neural human avatars with parameterized shape and pose control. arXiv preprint arXiv:2303.17606, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.647, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[24] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 1, 2, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.689, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[25] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.746, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[26] Jihoon Kim, Jiseob Kim, and Sungjoon Choi. Flame: Freeform language-based motion synthesis & editing. In AAAI, pages 8255-8263, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[27] Georgios Kopanas, Julien Philip, Thomas Leimkuhler, and George Drettakis. Point-based neural rendering with perview optimization. In Computer Graphics Forum, pages 29-43. Wiley Online Library, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[28] Georgios Kopanas, Thomas Leimkuhler, Gilles Rainer, Clément Jambon, and George Drettakis. Neural point cata-caustics for novel-view synthesis of reflections. ACM Transactions on Graphics (TOG), 41(6):1-15, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6804" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[29] Jiabao Lei, Yabin Zhang, Kui Jia, et al. Tango: Text-driven photorealistic and robust 3d stylization via lighting decomposition. NeurIPS, 35:30923-30936, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.471, + 0.19 + ], + "angle": 0, + "content": "[30] Gang Li, Heliang Zheng, Chaoyue Wang, Chang Li, Changwen Zheng, and Dacheng Tao. 3ddesigner: Towards photorealistic 3d object generation and editing with text-guided diffusion models. arXiv preprint arXiv:2211.14108, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.19, + 0.471, + 0.245 + ], + "angle": 0, + "content": "[31] Weiyu Li, Rui Chen, Xuelin Chen, and Ping Tan. Sweetdreamer: Aligning geometric priors in 2d diffusion for consistent text-to-3d. arXiv preprint arXiv:2310.02596, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.246, + 0.471, + 0.314 + ], + "angle": 0, + "content": "[32] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, pages 300–309, 2023. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.316, + 0.471, + 0.372 + ], + "angle": 0, + "content": "[33] Yukang Lin, Haonan Han, Chaoqun Gong, Zunnan Xu, Yachao Zhang, and Xiu Li. Consistent123: One image to highly consistent 3d asset using case-aware diffusion priors. arXiv preprint arXiv:2309.17261, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.372, + 0.471, + 0.426 + ], + "angle": 0, + "content": "[34] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Zexiang Xu, Hao Su, et al. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. arXiv preprint arXiv:2306.16928, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.426, + 0.471, + 0.482 + ], + "angle": 0, + "content": "[35] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. arXiv preprint arXiv:2303.11328, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.482, + 0.471, + 0.538 + ], + "angle": 0, + "content": "[36] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.471, + 0.594 + ], + "angle": 0, + "content": "[37] Zhen Liu, Yao Feng, Michael J Black, Derek Nowrouzezahrai, Liam Paull, and Weiyang Liu. Meshdiffusion: Score-based generative 3d mesh modeling. arXiv preprint arXiv:2303.08133, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.471, + 0.663 + ], + "angle": 0, + "content": "[38] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. arXiv preprint arXiv:2310.15008, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.471, + 0.72 + ], + "angle": 0, + "content": "[39] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 851-866. 2023. 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.471, + 0.788 + ], + "angle": 0, + "content": "[40] Jonathan Lorraine, Kevin Xie, Xiaohui Zeng, Chen-Hsuan Lin, Towaki Takikawa, Nicholas Sharp, Tsung-Yi Lin, Ming-Yu Liu, Sanja Fidler, and James Lucas. Att3d: Amortized text-to-3d object synthesis. arXiv preprint arXiv:2306.07349, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.471, + 0.83 + ], + "angle": 0, + "content": "[41] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In CVPR, pages 2837-2845, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.471, + 0.873 + ], + "angle": 0, + "content": "[42] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. arXiv preprint arXiv:2306.07279, 2023. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.903 + ], + "angle": 0, + "content": "[43] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.092, + 0.892, + 0.12 + ], + "angle": 0, + "content": "3d shapes and textures. In \\(CVPR\\), pages 12663-12673, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.164 + ], + "angle": 0, + "content": "[44] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In CVPR, pages 13492-13502, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.221 + ], + "angle": 0, + "content": "[45] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, pages 405-421, 2020. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.277 + ], + "angle": 0, + "content": "[46] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 conference papers, pages 1-8, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.333 + ], + "angle": 0, + "content": "[47] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.39 + ], + "angle": 0, + "content": "[48] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.447 + ], + "angle": 0, + "content": "[49] Yichen Ouyang, Wenhao Chai, Jiayi Ye, Dapeng Tao, Yibing Zhan, and Gaoang Wang. Chasing consistency in text-to-3d generation from a single image. arXiv preprint arXiv:2309.03599, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.448, + 0.892, + 0.49 + ], + "angle": 0, + "content": "[50] Jangho Park, Gihyun Kwon, and Jong Chul Ye. Ed-nerf: Efficient text-guided editing of 3d scene using latent space nerf. arXiv preprint arXiv:2310.02712, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.491, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[51] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 32, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.561, + 0.892, + 0.604 + ], + "angle": 0, + "content": "[52] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.605, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[53] Zekun Qi, Muzhou Yu, Runpei Dong, and Kaisheng Ma. Vpp: Efficient conditional 3d generation via voxel-point progressive representation. arXiv preprint arXiv:2307.16605, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[54] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, et al. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. arXiv preprint arXiv:2306.17843, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.745, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[55] Sigal Raab, Inbal Leibovitch, Guy Tevet, Moab Arar, Amit H Bermano, and Daniel Cohen-Or. Single motion diffusion. arXiv preprint arXiv:2302.05905, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[56] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.903 + ], + "angle": 0, + "content": "[57] Amit Raj, Srinivas Kaza, Ben Poole, Michael Niemeyer, Nataniel Ruiz, Ben Mildenhall, Shiran Zada, Kfir Aberman, Michael Rubinstein, Jonathan Barron, et al. Dream" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6805" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "booth3d: Subject-driven text-to-3d generation. arXiv preprint arXiv:2303.13508, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.47, + 0.175 + ], + "angle": 0, + "content": "[58] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.177, + 0.469, + 0.232 + ], + "angle": 0, + "content": "[59] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, Marco Fumero, and Kamal Rahimi Malekshan. Clip-forge: Towards zero-shot text-to-shape generation. In CVPR, pages 18603-18613, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.469, + 0.3 + ], + "angle": 0, + "content": "[60] Kyle Sargent, Zizhang Li, Tanmay Shah, Charles Herrmann, Hong-Xing Yu, Yunzhi Zhang, Eric Ryan Chan, Dmitry Lagun, Li Fei-Fei, Deqing Sun, et al. Zeronvs: Zero-shot 360-degree view synthesis from a single real image. arXiv preprint arXiv:2310.17994, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.469, + 0.356 + ], + "angle": 0, + "content": "[61] Hoigi Seo, Hayeon Kim, Gwanghyun Kim, and Se Young Chun. Ditto-nerf: Diffusion-based iterative text to omnidirectional 3d model. arXiv preprint arXiv:2304.02827, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.358, + 0.469, + 0.425 + ], + "angle": 0, + "content": "[62] Junyoung Seo, Wooseok Jang, Min-Seop Kwak, Jaehoon Ko, Hyeonsu Kim, Junho Kim, Jin-Hwa Kim, Jiyoung Lee, and Seungryong Kim. Let 2d diffusion model know 3d-consistency for robust text-to-3d generation. arXiv preprint arXiv:2303.07937, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.427, + 0.469, + 0.468 + ], + "angle": 0, + "content": "[63] Yonatan Shafir, Guy Tevet, Roy Kapon, and Amit H Bermano. Human motion diffusion as a generative prior. arXiv preprint arXiv:2303.01418, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.469, + 0.523 + ], + "angle": 0, + "content": "[64] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. NeurIPS, 34: 6087-6101, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.469, + 0.591 + ], + "angle": 0, + "content": "[65] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.469, + 0.662 + ], + "angle": 0, + "content": "[66] Yukai Shi, Jianan Wang, He Cao, Boshi Tang, Xianbiao Qi, Tianyu Yang, Yukun Huang, Shilong Liu, Lei Zhang, and Heung-Yeung Shum. Toss: High-quality text-guided novel view synthesis from a single image. arXiv preprint arXiv:2310.10644, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.469, + 0.705 + ], + "angle": 0, + "content": "[67] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.707, + 0.469, + 0.761 + ], + "angle": 0, + "content": "[68] Liangchen Song, Liangliang Cao, Hongyu Xu, Kai Kang, Feng Tang, Junsong Yuan, and Yang Zhao. Roomdreamer: Text-driven 3d indoor scene synthesis with coherent geometry and texture. arXiv preprint arXiv:2305.11337, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.469, + 0.817 + ], + "angle": 0, + "content": "[69] Jingxiang Sun, Bo Zhang, Ruizhi Shao, Lizhen Wang, Wen Liu, Zhenda Xie, and Yebin Liu. Dreamcraft3d: Hierarchical 3d generation with bootstrapped diffusion prior. arXiv preprint arXiv:2310.16818, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[70] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[71] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "3d creation from a single image with diffusion prior. arXiv preprint arXiv:2303.14184, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[72] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[73] Christina Tsalicoglou, Fabian Manhardt, Alessio Tonioni, Michael Niemeyer, and Federico Tombari. Textmesh: Generation of realistic 3d meshes from text prompts. arXiv preprint arXiv:2304.12439, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[74] Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, Karsten Kreis, et al. Lion: Latent point diffusion models for 3d shape generation. NeurIPS, 35:10021-10039, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.292, + 0.892, + 0.345 + ], + "angle": 0, + "content": "[75] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In CVPR, pages 3835-3844, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[76] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In CVPR, pages 12619-12629, 2023. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.406, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[77] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213, 2023. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.892, + 0.532 + ], + "angle": 0, + "content": "[78] Haohan Weng, Tianyu Yang, Jianan Wang, Yu Li, Tong Zhang, CL Chen, and Lei Zhang. Consistent123: Improve consistency for one image to 3d object synthesis. arXiv preprint arXiv:2310.08092, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.533, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[79] Zhenzhen Weng, Zeyu Wang, and Serena Yeung. Zero-)rator: Zero-shot 3d avatar generation from a single image. arXiv preprint arXiv:2305.16411, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.576, + 0.892, + 0.643 + ], + "angle": 0, + "content": "[80] Jinbo Wu, Xiaobo Gao, Xing Liu, Zhengyang Shen, Chen Zhao, Haocheng Feng, Jingtuo Liu, and Errui Ding. Hdfusion: Detailed text-to-3d generation leveraging multiple noise estimation. arXiv preprint arXiv:2307.16183, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.646, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[81] Jiale Xu, Xintao Wang, Weihao Cheng, Yan-Pei Cao, Ying Shan, Xiaohu Qie, and Shenghua Gao. Dream3d: Zero-shot text-to-3d synthesis using 3d shape prior and text-to-image diffusion models. In CVPR, pages 20908–20918, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.703, + 0.892, + 0.757 + ], + "angle": 0, + "content": "[82] Jiayu Yang, Ziang Cheng, Yunfei Duan, Pan Ji, and Hongdong Li. Consistnet: Enforcing 3d consistency for multiview images diffusion. arXiv preprint arXiv:2310.10343, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.76, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[83] Jianglong Ye, Peng Wang, Kejie Li, Yichun Shi, and Heng Wang. Consistent-1-to-3: Consistent image to 3d view synthesis via geometry-aware diffusion models. arXiv preprint arXiv:2310.03020, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[84] Wang Yifan, Felice Serena, Shihao Wu, Cengiz Öztireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM Transactions on Graphics (TOG), 38(6):1-14, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[85] Chaohui Yu, Qiang Zhou, Jingliang Li, Zhe Zhang, Zhibin Wang, and Fan Wang. Points-to-3d: Bridging the gap be" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6806" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.093, + 0.468, + 0.12 + ], + "angle": 0, + "content": "tween sparse points and shape-controllable text-to-3d generation. In ACM MM, pages 6841-6850, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.189 + ], + "angle": 0, + "content": "[86] Huichao Zhang, Bowen Chen, Hao Yang, Liao Qu, Xu Wang, Li Chen, Chao Long, Feida Zhu, Kang Du, and Min Zheng. Avatarverse: High-quality & stable 3d avatar creation from text and pose. arXiv preprint arXiv:2308.03610, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.261 + ], + "angle": 0, + "content": "[87] Longwen Zhang, Qiwei Qiu, Hongyang Lin, Qixuan Zhang, Cheng Shi, Wei Yang, Ye Shi, Sibei Yang, Lan Xu, and Jingyi Yu. Dreamface: Progressive generation of animatable 3d faces under text guidance. arXiv preprint arXiv:2304.03117, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.468, + 0.318 + ], + "angle": 0, + "content": "[88] Mingyuan Zhang, Xinying Guo, Liang Pan, Zhongang Cai, Fangzhou Hong, Huirong Li, Lei Yang, and Ziwei Liu. Remodiffuse: Retrieval-augmented motion diffusion model. arXiv preprint arXiv:2304.01116, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.32, + 0.469, + 0.374 + ], + "angle": 0, + "content": "[89] Mengyi Zhao, Mengyuan Liu, Bin Ren, Shuling Dai, and Nicu Sebe. Modiff: Action-conditioned 3d motion generation with denoising diffusion probabilistic models. arXiv preprint arXiv:2301.03949, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.376, + 0.469, + 0.444 + ], + "angle": 0, + "content": "[90] Minda Zhao, Chaoyi Zhao, Xinyue Liang, Lincheng Li, Zeng Zhao, Zhipeng Hu, Changjie Fan, and Xin Yu. Efficientdreamer: High-fidelity and robust 3d creation via orthogonal-view diffusion prior. arXiv preprint arXiv:2308.13223, 2023. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.446, + 0.469, + 0.487 + ], + "angle": 0, + "content": "[91] Zixiang Zhou and Baoyuan Wang. Ude: A unified driving engine for human motion generation. In CVPR, pages 5632-5641, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.489, + 0.469, + 0.529 + ], + "angle": 0, + "content": "[92] Joseph Zhu and Peiye Zhuang. Hifa: High-fidelity text-to-3d with advanced diffusion guidance. arXiv preprint arXiv:2305.18766, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.469, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6807" + } + ] +] \ No newline at end of file diff --git a/2024/GaussianDreamer_ Fast Generation from Text to 3D Gaussians by Bridging 2D and 3D Diffusion Models/68c13cd6-fa8d-4333-a8a0-12d60a7c6cac_origin.pdf b/2024/GaussianDreamer_ Fast Generation from Text to 3D Gaussians by Bridging 2D and 3D Diffusion Models/68c13cd6-fa8d-4333-a8a0-12d60a7c6cac_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cfcacac12d1c3ef58c6d3fb5fcf1a5c4332af602 --- /dev/null +++ b/2024/GaussianDreamer_ Fast Generation from Text to 3D Gaussians by Bridging 2D and 3D Diffusion Models/68c13cd6-fa8d-4333-a8a0-12d60a7c6cac_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d56010d7afe0aa854bceaa6415b4bb0d39a826cc8971b13d5c4f2de44df6d01c +size 6631104 diff --git a/2024/GaussianDreamer_ Fast Generation from Text to 3D Gaussians by Bridging 2D and 3D Diffusion Models/full.md b/2024/GaussianDreamer_ Fast Generation from Text to 3D Gaussians by Bridging 2D and 3D Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c715fef445324d88423593b6664028ef807b865c --- /dev/null +++ b/2024/GaussianDreamer_ Fast Generation from Text to 3D Gaussians by Bridging 2D and 3D Diffusion Models/full.md @@ -0,0 +1,471 @@ +# GaussianDreamer: Fast Generation from Text to 3D Gaussians by Bridging 2D and 3D Diffusion Models + +Taoran Yi $^{1}$ , Jiemin Fang $^{2\dagger}$ , Junjie Wang $^{2}$ , Guanjun Wu $^{3}$ , Lingxi Xie $^{2}$ , Xiaopeng Zhang $^{2}$ , Wenyu Liu $^{1}$ , Qi Tian $^{2}$ , Xinggang Wang $^{1\dagger}$ + +$^{1}$ School of EIC, Huazhong University of Science and Technology $^{2}$ Huawei Inc. $^{3}$ School of CS, Huazhong University of Science and Technology + +{taoranyi, guajuwu, liuwy, xgwang}@hust.edu.cn + +{jaminfong, is.wangjunjie, 198808xc, zxphistory}@gmail.com tian.qil@huawei.com + +![](images/b43f1cb133f73d9bb4060ed0df294a63043f74bfdc32fe9e00b27168bc0e64fe.jpg) +Figure 1. We propose a simple yet efficient framework called GaussianDreamer. It bridges the 3D and 2D diffusion models via Gaussian splatting, having both 3D consistency and rich generation details. Our method can complete training within 15 minutes on a single GPU and achieve real-time rendering. + +![](images/5f20021c96402161f3769d9d4f9e5b399c37d00b7500c4bfbfbbb693e9b49e4e.jpg) +3D Gaussians with Temporal Variation + +# Abstract + +In recent times, the generation of 3D assets from text prompts has shown impressive results. Both 2D and 3D diffusion models can help generate decent 3D objects based on prompts. 3D diffusion models have good 3D consistency, but their quality and generalization are limited as trainable 3D data is expensive and hard to obtain. 2D diffusion models enjoy strong abilities of generalization and fine generation, but 3D consistency is hard to guarantee. This paper attempts to bridge the power from the two types of diffusion models via the recent explicit and efficient 3D Gaussian splatting representation. A fast 3D object generation framework, named as GaussianDreamer, is proposed, where the 3D diffusion model provides priors for initialization and the 2D diffusion model enriches the geometry and appearance. Operations of noisy point growing and color perturbation are introduced to enhance the initialized Gaussians. Our GaussianDreamer can generate a high-quality 3D instance or 3D avatar within 15 minutes on one GPU, much faster than previous methods, while the gener + +ated instances can be directly rendered in real time. Demos and code are available at https://taoranyi.com/gaussiandreamer/. + +# 1. Introduction + +3D asset generation has been an expensive and professional work in conventional pipelines. Recently, diffusion models [58] have achieved great success in creating high-quality and realistic 2D images. Many research works [2, 6, 9, 12, 15, 24, 32, 48, 52, 67, 70, 76, 77, 81, 90] try to transfer the power of 2D diffusion models to the 3D field for easing and assisting the process of 3D assets creation, e.g. the most common text-to-3D task. + +Here come two main streams for achieving this goal: (i) training a new diffusion model with 3D data [12, 15, 24, 48] (namely the 3D diffusion model) and (ii) lifting the 2D diffusion model to 3D [2, 6, 9, 32, 52, 58, 67, 70, 76, 77, 81, 90]. The former one is direct to implement and holds strong 3D consistency, but struggles to extend into a large generation domain as 3D data is usually hard and expensive to obtain. The scale of current 3D datasets is far smaller than 2D datasets. This results in the generated 3D assets falling + +short in dealing with complex text prompts and producing complex/fine geometry and appearance. The latter benefits from the large data domain of the 2D diffusion models, which can handle various text prompts and produce highly detailed and complex geometry and appearance. However, as 2D diffusion models are unaware of the camera view, the generated 3D assets are hard to form geometry consistency, especially for structure-complicated instances. + +This paper proposes to use recent 3D Gaussian Splating [25] to bridge the two aforementioned approaches, simultaneously having the geometry consistency from 3D diffusion models and rich details from 2D diffusion models. 3D Gaussians are one type of efficient and explicit representation, which intrinsically enjoys geometry priors due to the point-cloud-like structure. Specifically, we use one of two types of 3D diffusion models: text-to-3D and text-to-motion diffusion models, e.g. Shap-E [24] and MDM [72] in our implementation, to generate a coarse 3D instance. Based on the coarse 3D instance, a group of 3D Gaussians are initialized. We introduce two operations of noisy point growing and color perturbation to supplement the initialized Gaussians for follow-up enriching the 3D instance. Then the 3D Gaussians can be improved and optimized by interacting with the 2D diffusion model via the Score Distillation Sampling [52] (SDS) loss. Due to the geometry priors from both the 3D diffusion model and 3D Gaussian Splating itself, the training process can be finished in a very short time. The generated 3D asset can be rendered in real time without transformation into structures like mesh via the splatting process. + +Our contributions can be summarized as follows. + +- We propose a text-to-3D method, named as Gaussian-Dreamer which bridges the 3D and 2D diffusion models via Gaussian splitting, enjoying both 3D consistency and rich generation details. +- Noisy point growing and color perturbation are introduced to supplement the initialized 3D Gaussians for further content enrichment. +- The overall method is simple and quite effective. A 3D instance can be generated within 15 minutes on one GPU, much faster than previous methods, and can be directly rendered in real time. + +# 2. Related Works + +3D Pretrained Diffusion Models. Recently, text-to-3D asset generation using diffusion models has achieved great success. Currently, it is mainly divided into lifting 2D diffusion models to 3D and 3D pretrained diffusion models, the difference lies in whether the training data used is 2D or 3D. 3D pretrained diffusion models [12, 15, 24, 48], referred to as 3D diffusion models in our paper, are models pretrained on text-3D pairs. After pretraining, they can generate 3D assets through only inference, and mod + +els such as Point-E [48] and Shape-E [24] can generate 3D assets in minutes. In addition to generating 3D assets from text, there are methods where 3D diffusion models [1, 7, 10, 26, 55, 63, 72, 88, 89, 91] generate motion sequences based on text-motion data. By pretraining on text-motion pairs, these models can generate reasonable motion sequences for different texts. The generated motion sequences can be transformed into the SMPL (Skinned Multi-Person Linear) model [39] based on mesh representation, but texture information is not included. In our method, we can paint transformed SMPL by using different text prompts. + +Lifting 2D Diffusion Models to 3D. In text-to-3D asset generation methods [8, 16, 21, 23, 40, 49, 50, 53, 57, 61, 62, 68, 73, 80, 85, 87, 92], in addition to using 3D pretrained diffusion models, lifting a 2D diffusion model to 3D is a training-free approach. Moreover, due to the abundance of 2D image data, this method produces assets with higher diversity and fidelity. Some single-image-to-3D methods [13, 19, 30, 33–36, 38, 54, 60, 65, 66, 71, 78, 79, 82, 83] also employ similar ideas. DreamFusion [52] first proposes SDS (Score Distillation Sampling) method, which is to update the 3D representation model using the 2D diffusion model. [76] proposes the method of SJC (Score Jacobian Chaining) to lift the 2D diffusion model to 3D. Later methods [6, 31, 32, 69, 77] build on DreamFusion and further improve the quality of 3D generation. Among them, generated 3D assets may suffer from multi-face problems. To address this issue, some methods strengthen the semantics of different views [2] and use multi-view information [67, 90] to alleviate such problems. There are also models [18, 22, 29, 44, 46, 59, 75, 81] which adopt CLIP [56] to align each view of the 3D representation model with the text. + +3D Representation Methods. In recent times, neural radiance fields (NeRF) [45] have achieved impressive results in 3D representation, and many methods in text-to-3D asset generation have also adopted NeRF or its variants [3, 47] as the representation method. Some methods [6, 12, 31, 32] use explicit optimization mesh representation methods like DMTET [64] to reduce rendering costs and further improve resolution. In addition to that, there are also generation methods that utilize point clouds [41, 48, 53, 74] and meshes [37] as 3D representations. + +Recently, 3D Gaussian Splatting [25] has been introduced as a representation method for 3D scenes, which can achieve rendering effects comparable to NeRF-based methods and enable real-time rendering. Two concurrent works [9, 70] also construct the 3D representation with 3D Gaussian Splatting [25]. DreamGaussian [70] uses a single image as a condition to generate 3D assets, while GS- + +![](images/79c713f12bc171b2a6e360462360b1ffac48208e70c3cc7d2d1e66999abcfca3.jpg) +Figure 2. Overall framework of GaussianDreamer. Firstly, we utilize a 3D diffusion model to generate the initialized point clouds. After executing noisy point growing and color perturbation on the point clouds, we use them to initialize the 3D Gaussians. The initialized 3D Gaussians are further optimized using the SDS method [52] with a 2D diffusion model. Finally, we render the image using the 3D Gaussians by employing 3D Gaussian Splatting [25]. We can use one of various 3D diffusion models to generate the initialized point clouds. In this case, we take text-to-3D and text-to-motion diffusion models as examples. + +GEN [9] implements high-quality generation from text to 3D. Our method shares a similar idea of using 3D Gaussian Splating as the representation method, which significantly reduces the cost of improving resolution and achieving a much faster optimization speed compared to estimizable mesh representation methods. And we can generate detailed 3D assets based on prompt texts in a very short time. + +# 3. Method + +In this section, we first review 2D and 3D diffusion models and the 3D representation method - 3D Gaussian Splating [25]. We give an overview of the whole framework in Sec. 3.2. Then, in Sec. 3.3, we describe the process of initializing the 3D Gaussians with the assistance of 3D diffusion models. The further optimization of 3D Gaussians using the 2D diffusion model is described in Sec. 3.4. + +# 3.1. Preliminaries + +DreamFusion. DreamFusion [52] is one of the most representative methods to lift 2D diffusion models to 3D, which proposes to optimize the 3D representation with the score distillation sampling (SDS) loss via a pre-trained 2D diffusion model $\phi$ . Specifically, it takes MipNeRF [3] as the 3D representation method, whose parameters $\theta$ are optimized. Taking the rendering method as $g$ , the rendered image results in $\mathbf{x} = g(\theta)$ . To make the rendered image $\mathbf{x}$ similar to the samples obtained from the diffusion model $\phi$ , DreamFusion uses a scoring estimation function: $\hat{\epsilon}_{\phi}(\mathbf{z}_t; y, t)$ , which predicts the sampled noise $\hat{\epsilon}_{\phi}$ given the noisy image $\mathbf{z}_t$ , text embedding $y$ , and noise level $t$ . By measuring the difference between the Gaussian noise $\epsilon$ added to the rendered image $\mathbf{x}$ and the predicted noise $\hat{\epsilon}_{\phi}$ , this scoring estimation function can provide the direction for updating the parameter $\theta$ . The formula for computing + +the gradient is as + +$$ +\nabla_ {\theta} \mathcal {L} _ {\mathrm {S D S}} (\phi , \mathbf {x} = g (\theta)) \triangleq \mathbb {E} _ {t, \epsilon} \left[ w (t) \left(\hat {\epsilon} _ {\phi} \left(\mathbf {z} _ {t}; y, t\right) - \epsilon\right) \frac {\partial \mathbf {x}}{\partial \theta} \right], \tag {1} +$$ + +where $w(t)$ is a weighting function. + +3D Gaussian Splatting. 3D Gaussian Splatting [25] (3D-GS) is a recent groundbreaking method for novel-view synthesis. Unlike implicit representation methods such as NeRF [45], which renders images based on volume rendering, 3D-GS renders images through splatting [84], achieving real-time speed. Specifically, 3D-GS represents the scene through a set of anisotropic Gaussians, defined with its center position $\mu \in \mathbb{R}^3$ , covariance $\Sigma \in \mathbb{R}^7$ , color $c \in \mathbb{R}^3$ , and opacity $\alpha \in \mathbb{R}^1$ . And the 3D Gaussians can be queried as follows: + +$$ +G (x) = e ^ {- \frac {1}{2} (x) ^ {T} \Sigma^ {- 1} (x)}, \tag {2} +$$ + +where $x$ represents the distance between $\mu$ and the query point. For computing the color of each pixel, it uses a typical neural point-based rendering [27, 28]. A ray $r$ is cast from the center of the camera, and the color and density of the 3D Gaussians that the ray intersects are computed along the ray. The rendering process is as follows: + +$$ +C (r) = \sum_ {i \in \mathcal {N}} c _ {i} \sigma_ {i} \prod_ {j = 1} ^ {i - 1} (1 - \sigma_ {j}), \quad \sigma_ {i} = \alpha_ {i} G (x _ {i}), \tag {3} +$$ + +where $\mathcal{N}$ represents the number of sample points on the ray $r$ , $c_{i}$ and $\alpha_{i}$ denote the color and opacity of the i-th Gaussian, and $x_{i}$ is the distance between the point and the i-th Gaussian. + +# 3.2. Overall Framework + +Our overall framework consists of two parts, initialization with 3D diffusion model priors and optimization with the 2D diffusion model, as shown in Fig. 2. For initialization with 3D diffusion model priors, we use the 3D diffusion models $F_{3D}$ , instantiated with the text-to-3D and text-to-motion diffusion models, to generate the triangle mesh $m$ based on the text prompt $y$ , which can be denoted as $m = F_{3D}(y)$ . One set of generated point clouds is transformed from the mesh $m$ . Then the 3D Gaussians $\theta_b$ are initialized via the generated point clouds after noisy point growing and color perturbation. For better quality, we utilize the 2D diffusion model $F_{2D}$ to further optimize the initialized 3D Gaussians $\theta_b$ via SDS [52] with prompts $y$ , resulting in the final 3D Gaussians $\theta_f$ . The target instance can be rendered in real time by splatting the generated Gaussians. + +# 3.3. Gaussian Initialization with 3D Diffusion Model Priors + +In this section, we mainly discuss how to initialize the 3D Gaussians with 3D diffusion model priors. First, we use the 3D diffusion model $F_{3D}$ to generate 3D assets based on the prompts $y$ . Then we convert the 3D assets into point clouds and use the transformed point clouds to initialize the 3D Gaussians. We have employed two types of 3D diffusion models to generate 3D assets. Below, we explain how to initialize the 3D Gaussians using each of these models. + +# 3.3.1 Text-to-3D Diffusion Model + +When using a text-based 3D generation model, generated 3D assets employ multi-layer perceptrons (MLPs) to predict SDF values and texture colors. To construct the triangle mesh $m$ , we query the SDF values at vertices along a regular grid with the size of $128^3$ within the MLPs. Then we query the texture colors at each vertex of $m$ . We convert the vertices and colors of $m$ into point clouds, denoted as $\pmb{p}\pmb{t}_m(\pmb{p}_m,\pmb{c}_m)$ . $\pmb{p}_m\in \mathbb{R}^3$ refers to the position of the point clouds, which equals to the vertex coordinated of $m$ . $c_{m}\in \mathbb{R}^{3}$ refers to the color of the point clouds, which is the same as the color of $m$ . However, the obtained colors $c_{m}$ are relatively simple, and the positions $p_m$ are sparse. + +Noisy Point Growing and Color Perturbation. We do not use the generated point clouds $\pmb{pt}_m$ to directly initialize the 3D Gaussians. To improve the quality of initialization, we perform noisy point growing and color perturbation around point clouds $\pmb{pt}_m$ . First, we compute the bounding box (BBox) of the surface on $\pmb{pt}_m$ and then uniformly grow point clouds $\pmb{pt}_r(\pmb{p}_r, \pmb{c}_r)$ within the BBox. $\pmb{p}_r$ and $\pmb{c}_r$ represent the positions and colors of $\pmb{pt}_r$ . To enable fast searching, we construct a KDTree [4] $K_m$ using the positions $\pmb{p}_m$ . Based on the distances between the positions $\pmb{p}_r$ + +![](images/bb582a35d7e07fd94b09d153ed9d7211e1ceb6f1e72bd01be21944fcde6c8680.jpg) +Figure 3. The process of noisy point growing and color perturbation. "Grow&Pertb." denotes noisy point growing and color perturbation. + +and the nearest points found in the KDTree $K_{m}$ , we determine which points to keep. In this process, we select points within the (normalized) distance of 0.01. For the noisy point clouds, we make their colors $\pmb{c}_r$ similar to $\pmb{c}_m$ , and also add some perturbations: + +$$ +\boldsymbol {c} _ {r} = \boldsymbol {c} _ {m} + \boldsymbol {a}, \tag {4} +$$ + +where the values of $\mathbf{a}$ are randomly sampled between 0 and 0.2. We merge the positions and colors of $\mathbf{p}\mathbf{t}_{m}$ and $\mathbf{p}\mathbf{t}_{r}$ to obtain the final point clouds. + +$$ +\boldsymbol {p} \boldsymbol {t} \left(\boldsymbol {p} _ {f}, \boldsymbol {c} _ {f}\right) = \left(\boldsymbol {p} _ {m} \oplus \boldsymbol {p} _ {r}, \boldsymbol {c} _ {m} \oplus \boldsymbol {c} _ {r}\right), \tag {5} +$$ + +where $\oplus$ is the concatenation operation. Fig. 3 illustrates the process of noisy point growing and color perturbation. Finally, we initialize the positions $\mu_{b}$ and colors $c_{b}$ of the 3D Gaussians $\theta_{b}(\mu_{b}, c_{b}, \Sigma_{b}, \alpha_{b})$ using both the positions $\pmb{p}_{f}$ and colors $\pmb{c}_{f}$ of final point clouds $\pmb{pt}$ . The opacity $\alpha_{b}$ of the 3D Gaussians is initialized to 0.1, and the covariance $\Sigma_{b}$ is calculated as the distance between the nearest two points. Algorithm 1 shows the specific algorithm flowchart. + +# 3.3.2 Text-to-Motion Diffusion Model + +We generate a sequence of human body motions using text and select a human pose that best matches the given text. We then convert the keypoints of this human pose into the SMPL model [39], which is represented by a triangle mesh $m$ . We then convert the mesh $m$ into point clouds $\pmb{pt}_m(\pmb{p}_m, \pmb{c}_m)$ , where the position $\pmb{p}_m$ of each point in the point clouds corresponds to the vertices of $m$ . As for the color $\pmb{c}_m$ of $\pmb{pt}_m$ , since the SMPL model used here does not have textures, we randomly initialize $\pmb{c}_m$ . To move $\pmb{pt}_m$ near the origin, we calculate the center point $\pmb{p}_c \in \mathbb{R}^3$ of $\pmb{p}_m$ and subtract the position of point clouds $\pmb{pt}_m$ from the center point $\pmb{p}_c$ . + +$$ +\boldsymbol {p} \boldsymbol {t} \left(\boldsymbol {p} _ {f}, \boldsymbol {c} _ {f}\right) = \boldsymbol {p} \boldsymbol {t} _ {m} \left(\boldsymbol {p} _ {m} - \boldsymbol {p} _ {c}, \boldsymbol {c} _ {m}\right), \tag {6} +$$ + +Finally, we use the point clouds $pt$ to initialize the 3D Gaussians, similar to what is described in Sec. 3.3.1. To improve + +Algorithm 1 The 3D Gaussian Initialization. + $\pmb{pt}_m(\pmb{p}_m, \pmb{c}_m)$ : Point clouds generated from $F_{3D}$ . + $\pmb{pt}_r(\pmb{p}_r, \pmb{c}_r)$ : Growing point clouds within the BBox. + $\pmb{pt}(\pmb{p}_f, \pmb{c}_f)$ : Point clouds used for initializing the 3D Gaussians. + $\theta_b(\mu_b, c_b, \Sigma_b, \alpha_b)$ : Initialized 3D Gaussians. + +Stage 1: Generate points in the BBox. + $K_{m} \leftarrow$ BuildKDTree( $p_{m}$ ) + $\text{BBox} \leftarrow p_{m}$ $\text{BBox} \leftarrow$ Positions bounding box. +Low, High $\leftarrow$ BBox.MinBound, BBox.MaxBound + $\text{BBox} \leftarrow$ Boundary of bounding box. + $\boldsymbol{ps}_{u} \leftarrow$ Uniform(Low, High, size = (NumPoints, 3)) + $\triangleright$ Points in the BBox. + +Stage 2: Keep the points that meet the distance requirement. + +$\pmb{p}_r,\pmb {c}_r = \left[ \begin{array}{ll} \end{array} \right],\left[ \begin{array}{ll} \end{array} \right]$ +for all $\pmb{p_u}$ in $p s_{u}$ do $\pmb{p}_{un},i\gets K_m.\mathrm{SearchNearest}(\pmb {p}_u)$ Nearest point and its index in $K_{m}$ if $|\pmb {p}_{un} - \pmb {p}_u| < 0.01$ then $\pmb{p}_r$ .append $(\pmb {p}_u)$ $c_{r}$ .append $(c_{m}[i] + 0.2\times \mathrm{Random}(\mathrm{size} = 3))$ $\triangleright$ Color of the nearest point plus perturbation. + +end if +end for + $\pmb {p}_f\gets \pmb {p}_m\oplus \pmb {p}_r$ $c_{f}\gets c_{m}\oplus c_{r}$ +Stage 3: Initialize the 3D Gaussians. + $\mu_b,c_b\gets p_f,c_f$ $\triangleright$ Positions and colors of the 3D Gaussians. + $D\leftarrow \mu_{b}\quad \triangleright$ Distance between the nearest two positions. + $\Sigma_{b},\alpha_{b}\gets D,0.1$ $\triangleright$ Covariance and opacity of the 3D Gaussians. + +the generation of motion sequences, we simplify the text by retaining only the relevant parts related to the motion and add a subject. For example, if the text prompt is "Iron man kicks with his left leg", we transform it into "Someone kicks with the left leg" when generating the motion sequences. + +# 3.4. Optimization with the 2D Diffusion Model + +To enrich details and improve the quality of the 3D asset, we optimize the 3D Gaussians $\theta_{b}$ with a 2D diffusion model $F_{2D}$ after initializing them with 3D diffusion model priors. We employ the SDS (Score Distillation Sampling) loss to optimize the 3D Gaussians. First, we use the method of 3D Gaussian Splatting [25] to obtain the rendered image $\mathbf{x} = g(\theta_{i})$ . Here, $g$ represents the splatting rendering method as in Eq. 3. Then, we use Eq. 1 to calculate the gradients for updating the Gaussian parameters $\theta_{i}$ with the 2D diffusion model $F_{2D}$ . After a short optimization period using the 2D diffusion model $F_{2D}$ , the final generated 3D + +Table 1. Quantitative comparisons on ${\mathrm{T}}^{3}$ Bench [17]. + +
MethodTime†Single Obj.Single w/ Surr.Multi Obj.Average
SJC [76]-24.719.811.718.7
DreamFusion [52]6 hours24.424.616.121.7
Fantasia3D [6]6 hours26.427.018.524.0
LatentNeRF [43]15 minutes33.130.620.628.1
Magic3D [32]5.3 hours37.035.425.732.7
ProlificDreamer [77]~10 hours49.444.835.843.3
Ours15 minutes54.048.634.545.7
+ +Scores above are average of two metrics (quality and alignment). +$\dagger$ GPU time counted in their papers. + +instance $\theta_{f}$ achieves high quality and fidelity on top of the 3D consistency provided by the 3D diffusion model $F_{3D}$ . + +# 4. Experiments + +In this section, we first present the implementation details in Sec. 4.1. In Sec. 4.2, we show the quantitative comparison results. Then, in Sec. 4.3, we showcase the visualization results of our method and compare them with other methods. In Sec. 4.4, we conduct a series of ablation experiments to validate the effectiveness of our method. Finally, we discuss the limitations of our method. + +# 4.1. Implementation Details + +Our method is implemented in PyTorch [51], based on ThreeStudio [14]. The 3D diffusion models used in our method are Shap-E [24] and MDM [72], and we load the Shap-E model finetuned on Objaverse [11] in Cap3D [42]. For the 2D diffusion model, we use stabilityai/stable-diffusion-2-1-base [58], with a guidance scale of 100. The timestamps we use are uniformly sampled from 0.02 to 0.98 before 500 iterations, and change to 0.02 to 0.55 after 500 iterations. For the 3D Gaussians, the learning rates of opacity $\alpha$ and position $\mu$ are $10^{-2}$ and $5 \times 10^{-5}$ . The color $c$ of the 3d Gaussians is represented by the sh coefficient, with the degree set to 0 and the learning rate set to $1.25 \times 10^{-2}$ . The covariance of the 3D Gaussians is converted into scaling and rotation for optimization, with learning rates of $10^{-3}$ and $10^{-2}$ , respectively. The radius of the camera we use for rendering is from 1.5 to 4.0, with the azimuth in the range of -180 to 180 degrees and the elevation in the range of -10 to 60 degrees. The total training iterations are 1200. All our experiments can be completed within 15 minutes on a single RTX 3090 with a batch size of 4. The resolution we use for rendering is $1024 \times 1024$ , which is scaled to 512 $\times$ 512 when optimizing using the 2D diffusion model. We can render in real time at $512 \times 512$ resolution. And all our code will be released. + +# 4.2. Quantitative Evaluation + +We evaluate quality and consistency following T3Bench [17], which provides a comprehensive benchmark for text-to-3D generation. Three text categories are + +![](images/c4b46be0ed11935322d0c31b2d3bb349e22df589648bbded554dea0a93925251.jpg) +DreamFusion 6.0 Hours + +![](images/79922dd0a52a901b81385541823270f804f0894692275bd8d0b8e4a1ab9bf33c.jpg) + +![](images/45155875c3469a406d73196d631c4d425f392d5197d3d1336d4eab0671160888.jpg) +Magic3D 5.3 Hours +Fantasia3D 6.0 Hours + +![](images/4fc682d9acda770c6b6876118a58e26afdfb7968429e468e00b1398a9d180c0d.jpg) +ProlificDreamer ~Hours +GaussianDreamer (Ours) 15 Minutes + +![](images/20649408e8da75ccdd629cbb303ee9b982d1d0ebbd7b4a807d972d194c73d22a.jpg) + +![](images/3df616b454f46e13d39cb16915bb76b907e4b74a245f4c6b5b805df783fe0d7a.jpg) + +![](images/7290185a6994106d0a5a8516f71e33486740dd15a357030589b6eae517aff470.jpg) +Figure 4. Qualitative comparisons between our method and DreamFusion [52], Magic3D [32], Fantasy3D [6] and ProlificDreamer [77]. Here we count the GPU time in their papers. The time for DreamFusion is measured on TPUv4, Magic3D is measured on A100, Fantasy3D is measured on RTX 3090, and our method is measured on RTX 3090. + +![](images/56c02fa1558038816f48335b17cb2aba7f56ed36ef33a19395d676333b3d92e9.jpg) + +![](images/753289c373588837d7220f4a0d5e9e8da89bc317da0dbaa0bd95d1c961b26d9e.jpg) + +![](images/b4448fb8506ebdf152b2561d19cafa572bd2e5bc7e483177892fb3e199e53101.jpg) + +![](images/169a62acecb4392a588a8975c36490e2b888819033ac8c73554bda5b07700fc9.jpg) +A plate piled high with chocolate chip cookies. + +![](images/f2692df22d77547f9ebc2ebdd6f765eea1ac84c701030bbdd06d0877f3b20697.jpg) + +![](images/ae5e65b857a6db3701ba16e0eeabc65a30c56496df91840937db887fe2be934f.jpg) +A 3D model of an adorable cottage with a thatched roof. + +![](images/5579c54f6d7b9f8b045568f74bf15d7f0203d1344e19ba30a5ffd7c66e1a609d.jpg) + +![](images/9247cd7677cc732004fdd032f1084a7da218c12b9ca42b6d12fa72d53a225881.jpg) + +![](images/ada454a8eb06f50484c9c7782bb9e057256733ac589ebd69888de2aa944d7a7b.jpg) +A fox + +![](images/feeec47e65c5f8eff091d6cfcdcfa442960f01a2934b13c9f5fe3bf016d2784b.jpg) + +![](images/8c14e8aec249158efe551260eb11e4ec9dc1bfe39ecfcb526272efa55b10c608.jpg) +An amigurumi motorcycle... + +![](images/404b6a7aada9388f28bdba8b8b99e37db7883c3ccda42b411c647b6a89c179d6.jpg) + +![](images/2710ada4be39e44be8d6923f5a8dd5eda21f943faea69e74d4711f08e3ebdffe.jpg) + +![](images/ea366c3716a1473a6be7cb841be0be38d34eae3309221c1e4a85f9f26d991bee.jpg) +An opulent couch from the palace of Versailles + +![](images/b2d0ead48d1c61ef323021840b5d61c2a6496cef6e99c7e080e020fe26ba17ff.jpg) + +![](images/59806878a7759f0f8438656b24f462d6ac9c206c2bb48edb0efc8753c970ae7a.jpg) + +![](images/7df71c74c5a0c7f6f31995bfb3097faefbdd82a80d09358763f36f8094a5ccd3.jpg) +Magic gun, game asset, mistery... +Magic dagger, mistery, ancient... +Figure 5. More generated samples by our GaussianDreamer. Two views of each sample are shown. + +![](images/7e657b51919a4558f204dec6904f0da93dd2720ac6c261de0835cb160652020f.jpg) +A zoomed out DSLR photo of a lion's mane jellyfish + +![](images/f890a4999f962286b63d61c7709edc53b66c63eb54a13a65ca0a94445bff3c97.jpg) + +designed for 3D generation with increasing complexity – single objects, single objects with surroundings, and multi objects. In Tab. 1, our method outperforms the compared methods while enjoying a short generation time. + +# 4.3. Visualization Results + +In this section, we present the results of initializing the 3D Gaussians with two different 3D diffusion models: text-to-3D and text-to-motion diffusion models. + +Initialization with Text-to-3D Diffusion Model. We show the comparison results with DreamFusion [52], Magic3D [32], Fantasy3D [6] and ProlificDreamer [77] in Fig. 4. In addition to our method, the figures of other methods are downloaded from the paper of ProlificDreamer. When encountering prompts that involve the combination of multiple objects, such as the prompt "A plate piled high with chocolate chip cookies", the generated results from Magic3D, Fantasy3D, and ProlificDreamer do not include a plate. In contrast, our generated result can effectively combine a plate and chocolate chip cookies. Furthermore, compared to DreamFusion, the plate we generate has better patterns. Our method shows comparable quality while saving 21 - 24 times the generation time compared to their + +methods. Moreover, the 3D Gaussians generated by our method can directly achieve real-time rendering without further transformation into mesh-like structures. Fig. 5 visualizes more samples generated by our GaussianDreamer from various prompts, which show good 3D consistency while having high-quality details. + +Initialization with Text-to-Motion Diffusion Model. In Fig. 6, we present the comparison results with DreamFusion [52], DreamAvatar [5], DreamWaltz [20], and AvatarVerse [86]. In addition to our method, the figures of other methods are downloaded from the paper of AvatarVerse. It is worth noting that our prompt is "Spiderman/Stormtrooper stands with open arms", while the prompts for other methods are "Spiderman/Stormtrooper". This is because when generating motion using text-to-motion diffusion model as initialization, we require more specific action descriptions. Our method achieves a speedup of 4-24 times compared to other methods, while maintaining comparable quality. Additionally, our method allows for generating 3D avatars with specified body poses. In Fig. 7, we provide more results generated with different human body poses. We first generate a sequence of motions that match the text prompt using a text-to-motion 3d diffusion model, and then initialize the + +![](images/0cfc1257a36d9ad19d9bf38570bddbbf78df0b6ce4206c80e7c433b10c1cd8e0.jpg) + +![](images/d4b09dcf5d6995715bc51a8ea393ea2c634c32f67eaf132566c4f2b25e040929.jpg) + +![](images/f8832450bbe09a68ea5aa1fed5a5c62c2bc15309ae6bff4a546bf4be874239db.jpg) + +![](images/3399a3158964f6a65d26795c956329a332dd2a19cf2efcf69223255a2e436d7a.jpg) +GaussianDreamer (Ours) 15 Minutes + +![](images/84996c3396cf61b95e47c8f381ef314d2432e42526d5dbaddba30ba6f1041219.jpg) + +![](images/9f086da7d7c61a81113c6d0f895fe1f9581a4c06fc5fd0a81367a8823db5fb89.jpg) +Stormtrooper + +![](images/44cef7937051be096397c49f171c7f367570f6b49fe250aa88146fd258700661.jpg) +Spiderman + +![](images/fdb541959261afc19ef371e5c0033a236a3baca59a6e849bc55d847b8a1bb003.jpg) + +![](images/f38748c4101231bfb0f09515c973d036bd80949127338995681d58505b84b99f.jpg) + +![](images/9b44a10d6ece8b727f465d7903ef5f0dffa3c1ae84cc4b9b98dc405816f1a392.jpg) +Figure 6. Qualitative comparisons between our method and DreamFusion [52], DreamAvatar [5], DreamWaltz [20], and AvatarVerse [86]. +Figure 7. More generated 3D avatars by our GaussianDreamer initialized with the different poses of SMPL [39]. Here, the different poses of SMPL are generated using a text-to-motion diffusion model. + +3d Gaussians with the SMPL of one selected pose in motions. Our method can generate 3D avatars in any desired pose. + +# 4.4. Ablation Study and Analysis + +The Role of Initialization. As shown in Fig. 8, we first conduct an ablation experiment on the initialization of the 3D Gaussians to validate that initialization can improve 3D consistency. The first column is the rendering result of Shap-E [24] with NeRF [45] as the 3D representation. The second column is the result of optimizing the 3D Gaussians + +randomly initialized within a cube using the SDS loss, and the third column is our method. We show the initialization effects on 3 samples. In the first and second rows, ShapE has good generation results while our method provides more complex geometries and more realistic appearances. Compared with random initialization, in the first row, details of our method are better. In the second row, the 3D assets generated by random initialization have the multi-head problem, which does not occur in our method. The initialization of the 3D diffusion model can avoid unreasonable geometry. In the third row, the generation result of Shap-E is far different from the given text prompt while our method makes the 3D assets closer to the prompt through the 2D diffusion model. Our method can expand the domain of Shap-E prompts, allowing for the generation of 3D assets based on a wider range of prompts. + +Noisy Point Growing and Color Perturbation. Fig. 9 illustrates the ablation results of noisy point growing and color perturbation. With noisy point growing and color perturbation applied, the first row showcases improved details in the sniper rifle. Additionally, the second column generates an amigurumi motorcycle that better aligns with the style characteristics of amigurumi mentioned in the prompt, compared to the case without noisy point growing and color perturbation. + +Initialization with Different Text-to-3D Diffusion Models. We select two text-to-3D generation models, ShapeE [24] and Point-E [48], to validate the effectiveness of our framework. We load the Point-E model finetuned on Objaverse [11] in Cap3D[42]. Fig. 10, we showcase the generated results after initializing the 3D Gaussians using + +![](images/1807572c3b4a4dcc7a809670101e10432bb4ac3afede8ee74eb3462a2422a0af.jpg) +Shap-E + +![](images/5e25c51298fd20ea10da18d92f718411b9e43e6904456d22a7e0259068530b09.jpg) +Ours w/o Point Cloud Prior + +![](images/c80410bf9f9e49ef5cc21b67a47c063c53106cb9a3fc5b5d847bd910e504c75c.jpg) + +![](images/7975de07c8bec90db2745111a489b16e732d4d9ab3c20ad1e2fe11e8f7e87573.jpg) + +![](images/d50db4694a3713138bec67aeb5b66782e706c844f78ce7e4016efa8d0b902549.jpg) +GaussianDreamer (Ours) + +![](images/e85175ebd52e7829a9b9d709beb8c5f2965ca828c112b7bf8a3844822a6cdab8.jpg) +A stack of pancakes covered in maple syrup + +![](images/1cd727f291e3666492d0fe66f68a9419dbbbc61990802c14b84110925e41eaf1.jpg) + +![](images/6bfd90bfc09b8e436553ff957360f6965f5880169e40baed95a09c1d296c3991.jpg) + +![](images/e637eae71eeee2c858b591bfef01b6e77dbf0c01105092f25c7c58d8e73fe4f5.jpg) + +![](images/97ec9a04ccf38a249c6d9f60b6e1be68ca6f14350fd4007241457038e7df0598.jpg) + +![](images/95f05e8ebe08383af6510ec9ed620522020bd6904899b22d9c4cb26f6e7ccdfb.jpg) +Figure 8. Ablation studies of the initialization of the 3D Gaussians. The Shap-E [24] rendering resolution here is $256 \times 256$ . + +![](images/c82c6c8330d5b8c021d620571379e98034ff0b5518f26bffe2ad6b67dc03e14d.jpg) + +![](images/03947ce9b4662d0af060f95acc3bfa577f3e427e417b69f36b454c3e0f2a13a3.jpg) +A lionfish + +![](images/379e57fae2221050da7d14800c0d41221bc64214b31cb4e2fbd7f5f61cc362bd.jpg) + +![](images/29084092880b0720a0a55b982efda37476f45335831fe66cd772233e870f3c25.jpg) + +![](images/003ebaa33cdd9170969b8dacc3eacea37d58121fba4fa36839e61f1813ad19e1.jpg) +Ours w/o Grow&Pertb. +GaussianDreamer (Ours) +Sniper rifle, asset, scifi, cyberpunk... + +![](images/9e5fe0ab69529996c392ef7eb29bfc293fd18c2cd212503dd2cba6ecd8d231d6.jpg) +An amigurumi motorcycle... +Figure 9. Ablation studies of noisy point growing and color perturbation. "Grow&Pertb." denotes noisy point growing and color perturbation. + +![](images/33f58febb62d7925bb0334415925dabb5be1e55409b01b8859917efbd0dcf946.jpg) +Initialization with Point-E +Initialization with Shap-E +An airplane made out of wood + +![](images/75c33677a6349772832ee11dd09544e8901798823d9fe45a8a16f3a422e1b31e.jpg) +Flamethrower, with fire, scif, cyberpunk... +Figure 10. Ablation studies of initialization with different text-to-3D diffusion models: Point-E [48] and Shap-E [24]. + +one of two text-to-3D generation models. It can be seen that both initializations yield good generation results. However, considering that Shap-E generates 3D assets based on NeRF and SDF, which provide higher fidelity compared to the point cloud representation used by Point-E, the geometry of the airplane in the first row of Fig. 10 appears better when initialized using Shap-E. + +# 4.5. Limitations + +The edges of the 3D assets generated by our method are not always sharp, and there may be unnecessary 3D Gaussians + +around the object surface. How to filter these point clouds will be a possible direction for improvement. Our approach utilizes 3D diffusion model priors, which greatly alleviates the issue of multi-face problems. However, there is still a small chance of encountering the multi-face problems in scenes where there is minimal geometric difference but significant appearance discrepancy between the front and back of objects such as a backpack. Utilizing 3D-aware diffusion models [35, 67, 90] may be able to solve this problem. Additionally, our method has limited effectiveness in generating large-scale scenes, such as indoor scenes. + +# 5. Conclusion + +We propose a fast text-to-3D method GaussianDreamer by bridging the abilities of 3D and 2D diffusion models via the Gaussian splatting representation. GaussianDreamer can generate detailed and realistic geometry and appearance while maintaining 3D consistency. The 3D diffusion model priors and geometry priors from the 3D Gaussians effectively promote the convergence speed. Each sample can be generated within 15 minutes on one GPU. We believe the approach of bridging 3D and 2D diffusion models could be a promising direction to generate 3D assets efficiently. + +# Acknowledgments + +This work was supported by the National Natural Science Foundation of China (No. 62376102). + +# References + +[1] Tenglong Ao, Zeyi Zhang, and Libin Liu. Gesture diffuclip: Gesture diffusion model with clip latents. arXiv preprint arXiv:2303.14613, 2023. 2 +[2] Mohammadreza Armandpour, Huangjie Zheng, Ali Sadeghian, Amir Sadeghian, and Mingyuan Zhou. Re-imagine the negative prompt algorithm: Transform 2d diffusion into 3d, alleviate janus problem and beyond. arXiv preprint arXiv:2304.04968, 2023.1, 2 +[3] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, pages 5855-5864, 2021. 2, 3 +[4] Jon Louis Bentley. Multidimensional binary search trees used for associative searching. Communications of the ACM, 18(9):509-517, 1975. 4 +[5] Yukang Cao, Yan-Pei Cao, Kai Han, Ying Shan, and KwanYee K Wong. Dreamavatar: Text-and-shape guided 3d human avatar generation via diffusion models. arXiv preprint arXiv:2304.00916, 2023. 6, 7 +[6] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. *Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation.* arXiv preprint arXiv:2303.13873, 2023. 1, 2, 5, 6 +[7] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In CVPR, pages 18000-18010, 2023. 2 +[8] Yiwen Chen, Chi Zhang, Xiaofeng Yang, Zhongang Cai, Gang Yu, Lei Yang, and Guosheng Lin. It3d: Improved text-to-3d generation with explicit view synthesis. arXiv preprint arXiv:2308.11473, 2023. 2 +[9] Zilong Chen, Feng Wang, and Huaping Liu. Text-to-3d using gaussian splatting. arXiv preprint arXiv:2309.16585, 2023. 1, 2, 3 +[10] Rishabh Dabral, Muhammad Hamza Mughal, Vladislav Golyanik, and Christian Theobalt. Mofusion: A framework for denoising-diffusion-based motion synthesis. In CVPR, pages 9760-9770, 2023. 2 +[11] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In CVPR, pages 13142-13153, 2023. 5, 7 +[12] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. NeurIPS, 35:31841-31854, 2022. 1, 2 +[13] Jiatao Gu, Alex Trevithick, Kai-En Lin, Joshua M Susskind, Christian Theobalt, Lingjie Liu, and Ravi Ramamoorthi. Nerfdiff: Single-image view synthesis with nerf-guided distillation from 3d-aware diffusion. In ICML, pages 11808–11826. PMLR, 2023. 2 +[14] Yuan-Chen Guo, Ying-Tian Liu, Ruizhi Shao, Christian Laforte, Vikram Voleti, Guan Luo, Chia-Hao Chen, Zi-Xin Zou, Chen Wang, Yan-Pei Cao, and Song-Hai Zhang. + +threestudio: A unified framework for 3d content generation. https://github.com/threestudio-project/threestudio, 2023.5 +[15] Anchit Gupta, Wenhan Xiong, Yixin Nie, Ian Jones, and Barlas Oğuz. 3dgen: Triplane latent diffusion for textured mesh generation. arXiv preprint arXiv:2303.05371, 2023. 1, 2 +[16] Xiao Han, Yukang Cao, Kai Han, Xiatian Zhu, Jiankang Deng, Yi-Zhe Song, Tao Xiang, and Kwan-Yee K Wong. Headsculpt: Crafting 3d head avatars with text. In NeurIPS, 2023. 2 +[17] Yuze He, Yushi Bai, Matthieu Lin, Wang Zhao, Yubin Hu, Jenny Sheng, Ran Yi, Juanzi Li, and Yong-Jin Liu. T 3 bench: Benchmarking current progress in text-to-3d generation. arXiv preprint arXiv:2310.02977, 2023. 5 +[18] Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. arXiv preprint arXiv:2205.08535, 2022. 2 +[19] Shoukang Hu, Fangzhou Hong, Tao Hu, Liang Pan, Haiyi Mei, Weiye Xiao, Lei Yang, and Ziwei Liu. Humanliff: Layer-wise 3d human generation with diffusion model. arXiv preprint arXiv:2308.09712, 2023. 2 +[20] Yukun Huang, Jianan Wang, Ailing Zeng, He Cao, Xianbiao Qi, Yukai Shi, Zheng-Jun Zha, and Lei Zhang. Dreamwaltz: Make a scene with complex 3d animatable avatars. arXiv preprint arXiv:2305.12529, 2023. 6, 7 +[21] Yangyi Huang, Hongwei Yi, Yuliang Xiu, Tingting Liao, Ji-axiang Tang, Deng Cai, and Justus Thies. Tech: Text-guided reconstruction of lifelike clothed humans. arXiv preprint arXiv:2308.08545, 2023. 2 +[22] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In CVPR, pages 867-876, 2022. 2 +[23] Ruixiang Jiang, Can Wang, Jingbo Zhang, Mengei Chai, Mingming He, Dongdong Chen, and Jing Liao. Avatarcraft: Transforming text into neural human avatars with parameterized shape and pose control. arXiv preprint arXiv:2303.17606, 2023. 2 +[24] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 1, 2, 5, 7, 8 +[25] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 2, 3, 5 +[26] Jihoon Kim, Jiseob Kim, and Sungjoon Choi. Flame: Freeform language-based motion synthesis & editing. In AAAI, pages 8255-8263, 2023. 2 +[27] Georgios Kopanas, Julien Philip, Thomas Leimkuhler, and George Drettakis. Point-based neural rendering with perview optimization. In Computer Graphics Forum, pages 29-43. Wiley Online Library, 2021. 3 +[28] Georgios Kopanas, Thomas Leimkuhler, Gilles Rainer, Clément Jambon, and George Drettakis. Neural point cata-caustics for novel-view synthesis of reflections. ACM Transactions on Graphics (TOG), 41(6):1-15, 2022. 3 + +[29] Jiabao Lei, Yabin Zhang, Kui Jia, et al. Tango: Text-driven photorealistic and robust 3d stylization via lighting decomposition. NeurIPS, 35:30923-30936, 2022. 2 +[30] Gang Li, Heliang Zheng, Chaoyue Wang, Chang Li, Changwen Zheng, and Dacheng Tao. 3ddesigner: Towards photorealistic 3d object generation and editing with text-guided diffusion models. arXiv preprint arXiv:2211.14108, 2022. 2 +[31] Weiyu Li, Rui Chen, Xuelin Chen, and Ping Tan. Sweetdreamer: Aligning geometric priors in 2d diffusion for consistent text-to-3d. arXiv preprint arXiv:2310.02596, 2023. 2 +[32] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, pages 300–309, 2023. 1, 2, 5, 6 +[33] Yukang Lin, Haonan Han, Chaoqun Gong, Zunnan Xu, Yachao Zhang, and Xiu Li. Consistent123: One image to highly consistent 3d asset using case-aware diffusion priors. arXiv preprint arXiv:2309.17261, 2023. 2 +[34] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Zexiang Xu, Hao Su, et al. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. arXiv preprint arXiv:2306.16928, 2023. +[35] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. arXiv preprint arXiv:2303.11328, 2023. 8 +[36] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 2 +[37] Zhen Liu, Yao Feng, Michael J Black, Derek Nowrouzezahrai, Liam Paull, and Weiyang Liu. Meshdiffusion: Score-based generative 3d mesh modeling. arXiv preprint arXiv:2303.08133, 2023. 2 +[38] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. arXiv preprint arXiv:2310.15008, 2023. 2 +[39] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 851-866. 2023. 2, 4, 7 +[40] Jonathan Lorraine, Kevin Xie, Xiaohui Zeng, Chen-Hsuan Lin, Towaki Takikawa, Nicholas Sharp, Tsung-Yi Lin, Ming-Yu Liu, Sanja Fidler, and James Lucas. Att3d: Amortized text-to-3d object synthesis. arXiv preprint arXiv:2306.07349, 2023. 2 +[41] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In CVPR, pages 2837-2845, 2021. 2 +[42] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. arXiv preprint arXiv:2306.07279, 2023. 5, 7 +[43] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of + +3d shapes and textures. In $CVPR$ , pages 12663-12673, 2023. 5 +[44] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In CVPR, pages 13492-13502, 2022. 2 +[45] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, pages 405-421, 2020. 2, 3, 7 +[46] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 conference papers, pages 1-8, 2022. 2 +[47] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 2 +[48] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 1, 2, 7, 8 +[49] Yichen Ouyang, Wenhao Chai, Jiayi Ye, Dapeng Tao, Yibing Zhan, and Gaoang Wang. Chasing consistency in text-to-3d generation from a single image. arXiv preprint arXiv:2309.03599, 2023. 2 +[50] Jangho Park, Gihyun Kwon, and Jong Chul Ye. Ed-nerf: Efficient text-guided editing of 3d scene using latent space nerf. arXiv preprint arXiv:2310.02712, 2023. 2 +[51] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 32, 2019. 5 +[52] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 1, 2, 3, 4, 5, 6, 7 +[53] Zekun Qi, Muzhou Yu, Runpei Dong, and Kaisheng Ma. Vpp: Efficient conditional 3d generation via voxel-point progressive representation. arXiv preprint arXiv:2307.16605, 2023. 2 +[54] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, et al. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. arXiv preprint arXiv:2306.17843, 2023. 2 +[55] Sigal Raab, Inbal Leibovitch, Guy Tevet, Moab Arar, Amit H Bermano, and Daniel Cohen-Or. Single motion diffusion. arXiv preprint arXiv:2302.05905, 2023. 2 +[56] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 2 +[57] Amit Raj, Srinivas Kaza, Ben Poole, Michael Niemeyer, Nataniel Ruiz, Ben Mildenhall, Shiran Zada, Kfir Aberman, Michael Rubinstein, Jonathan Barron, et al. Dream + +booth3d: Subject-driven text-to-3d generation. arXiv preprint arXiv:2303.13508, 2023. 2 +[58] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1, 5 +[59] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, Marco Fumero, and Kamal Rahimi Malekshan. Clip-forge: Towards zero-shot text-to-shape generation. In CVPR, pages 18603-18613, 2022. 2 +[60] Kyle Sargent, Zizhang Li, Tanmay Shah, Charles Herrmann, Hong-Xing Yu, Yunzhi Zhang, Eric Ryan Chan, Dmitry Lagun, Li Fei-Fei, Deqing Sun, et al. Zeronvs: Zero-shot 360-degree view synthesis from a single real image. arXiv preprint arXiv:2310.17994, 2023. 2 +[61] Hoigi Seo, Hayeon Kim, Gwanghyun Kim, and Se Young Chun. Ditto-nerf: Diffusion-based iterative text to omnidirectional 3d model. arXiv preprint arXiv:2304.02827, 2023. 2 +[62] Junyoung Seo, Wooseok Jang, Min-Seop Kwak, Jaehoon Ko, Hyeonsu Kim, Junho Kim, Jin-Hwa Kim, Jiyoung Lee, and Seungryong Kim. Let 2d diffusion model know 3d-consistency for robust text-to-3d generation. arXiv preprint arXiv:2303.07937, 2023. 2 +[63] Yonatan Shafir, Guy Tevet, Roy Kapon, and Amit H Bermano. Human motion diffusion as a generative prior. arXiv preprint arXiv:2303.01418, 2023. 2 +[64] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. NeurIPS, 34: 6087-6101, 2021. 2 +[65] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. 2 +[66] Yukai Shi, Jianan Wang, He Cao, Boshi Tang, Xianbiao Qi, Tianyu Yang, Yukun Huang, Shilong Liu, Lei Zhang, and Heung-Yeung Shum. Toss: High-quality text-guided novel view synthesis from a single image. arXiv preprint arXiv:2310.10644, 2023. 2 +[67] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 1, 2, 8 +[68] Liangchen Song, Liangliang Cao, Hongyu Xu, Kai Kang, Feng Tang, Junsong Yuan, and Yang Zhao. Roomdreamer: Text-driven 3d indoor scene synthesis with coherent geometry and texture. arXiv preprint arXiv:2305.11337, 2023. 2 +[69] Jingxiang Sun, Bo Zhang, Ruizhi Shao, Lizhen Wang, Wen Liu, Zhenda Xie, and Yebin Liu. Dreamcraft3d: Hierarchical 3d generation with bootstrapped diffusion prior. arXiv preprint arXiv:2310.16818, 2023. 2 +[70] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. 1, 2 +[71] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity + +3d creation from a single image with diffusion prior. arXiv preprint arXiv:2303.14184, 2023. 2 +[72] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 2, 5 +[73] Christina Tsalicoglou, Fabian Manhardt, Alessio Tonioni, Michael Niemeyer, and Federico Tombari. Textmesh: Generation of realistic 3d meshes from text prompts. arXiv preprint arXiv:2304.12439, 2023. 2 +[74] Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, Karsten Kreis, et al. Lion: Latent point diffusion models for 3d shape generation. NeurIPS, 35:10021-10039, 2022. 2 +[75] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In CVPR, pages 3835-3844, 2022. 2 +[76] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In CVPR, pages 12619-12629, 2023. 1, 2, 5 +[77] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213, 2023. 1, 2, 5, 6 +[78] Haohan Weng, Tianyu Yang, Jianan Wang, Yu Li, Tong Zhang, CL Chen, and Lei Zhang. Consistent123: Improve consistency for one image to 3d object synthesis. arXiv preprint arXiv:2310.08092, 2023. 2 +[79] Zhenzhen Weng, Zeyu Wang, and Serena Yeung. Zero-)rator: Zero-shot 3d avatar generation from a single image. arXiv preprint arXiv:2305.16411, 2023. 2 +[80] Jinbo Wu, Xiaobo Gao, Xing Liu, Zhengyang Shen, Chen Zhao, Haocheng Feng, Jingtuo Liu, and Errui Ding. Hdfusion: Detailed text-to-3d generation leveraging multiple noise estimation. arXiv preprint arXiv:2307.16183, 2023. 2 +[81] Jiale Xu, Xintao Wang, Weihao Cheng, Yan-Pei Cao, Ying Shan, Xiaohu Qie, and Shenghua Gao. Dream3d: Zero-shot text-to-3d synthesis using 3d shape prior and text-to-image diffusion models. In CVPR, pages 20908–20918, 2023. 1, 2 +[82] Jiayu Yang, Ziang Cheng, Yunfei Duan, Pan Ji, and Hongdong Li. Consistnet: Enforcing 3d consistency for multiview images diffusion. arXiv preprint arXiv:2310.10343, 2023. 2 +[83] Jianglong Ye, Peng Wang, Kejie Li, Yichun Shi, and Heng Wang. Consistent-1-to-3: Consistent image to 3d view synthesis via geometry-aware diffusion models. arXiv preprint arXiv:2310.03020, 2023. 2 +[84] Wang Yifan, Felice Serena, Shihao Wu, Cengiz Öztireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM Transactions on Graphics (TOG), 38(6):1-14, 2019. 3 +[85] Chaohui Yu, Qiang Zhou, Jingliang Li, Zhe Zhang, Zhibin Wang, and Fan Wang. Points-to-3d: Bridging the gap be + +tween sparse points and shape-controllable text-to-3d generation. In ACM MM, pages 6841-6850, 2023. 2 +[86] Huichao Zhang, Bowen Chen, Hao Yang, Liao Qu, Xu Wang, Li Chen, Chao Long, Feida Zhu, Kang Du, and Min Zheng. Avatarverse: High-quality & stable 3d avatar creation from text and pose. arXiv preprint arXiv:2308.03610, 2023. 6, 7 +[87] Longwen Zhang, Qiwei Qiu, Hongyang Lin, Qixuan Zhang, Cheng Shi, Wei Yang, Ye Shi, Sibei Yang, Lan Xu, and Jingyi Yu. Dreamface: Progressive generation of animatable 3d faces under text guidance. arXiv preprint arXiv:2304.03117, 2023. 2 +[88] Mingyuan Zhang, Xinying Guo, Liang Pan, Zhongang Cai, Fangzhou Hong, Huirong Li, Lei Yang, and Ziwei Liu. Remodiffuse: Retrieval-augmented motion diffusion model. arXiv preprint arXiv:2304.01116, 2023. 2 +[89] Mengyi Zhao, Mengyuan Liu, Bin Ren, Shuling Dai, and Nicu Sebe. Modiff: Action-conditioned 3d motion generation with denoising diffusion probabilistic models. arXiv preprint arXiv:2301.03949, 2023. 2 +[90] Minda Zhao, Chaoyi Zhao, Xinyue Liang, Lincheng Li, Zeng Zhao, Zhipeng Hu, Changjie Fan, and Xin Yu. Efficientdreamer: High-fidelity and robust 3d creation via orthogonal-view diffusion prior. arXiv preprint arXiv:2308.13223, 2023. 1, 2, 8 +[91] Zixiang Zhou and Baoyuan Wang. Ude: A unified driving engine for human motion generation. In CVPR, pages 5632-5641, 2023. 2 +[92] Joseph Zhu and Peiye Zhuang. Hifa: High-fidelity text-to-3d with advanced diffusion guidance. arXiv preprint arXiv:2305.18766, 2023. 2 \ No newline at end of file diff --git a/2024/GaussianDreamer_ Fast Generation from Text to 3D Gaussians by Bridging 2D and 3D Diffusion Models/images.zip b/2024/GaussianDreamer_ Fast Generation from Text to 3D Gaussians by Bridging 2D and 3D Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3b72111b717e1360b9dd80067a993437a6e6020d --- /dev/null +++ b/2024/GaussianDreamer_ Fast Generation from Text to 3D Gaussians by Bridging 2D and 3D Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:436c54e7dbb1defdd5a3861e82b7a8fdfd41b11d01daa1d5a95b58abeb16282f +size 664493 diff --git a/2024/GaussianDreamer_ Fast Generation from Text to 3D Gaussians by Bridging 2D and 3D Diffusion Models/layout.json b/2024/GaussianDreamer_ Fast Generation from Text to 3D Gaussians by Bridging 2D and 3D Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..610a8528f3c1d80be4b0782dc45a4ea7330da9e6 --- /dev/null +++ b/2024/GaussianDreamer_ Fast Generation from Text to 3D Gaussians by Bridging 2D and 3D Diffusion Models/layout.json @@ -0,0 +1,14111 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 102, + 103, + 492, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 103, + 492, + 140 + ], + "spans": [ + { + "bbox": [ + 102, + 103, + 492, + 140 + ], + "type": "text", + "content": "GaussianDreamer: Fast Generation from Text to 3D Gaussians by Bridging 2D and 3D Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 120, + 160, + 473, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 160, + 473, + 190 + ], + "spans": [ + { + "bbox": [ + 120, + 160, + 473, + 190 + ], + "type": "text", + "content": "Taoran Yi" + }, + { + "bbox": [ + 120, + 160, + 473, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 120, + 160, + 473, + 190 + ], + "type": "text", + "content": ", Jiemin Fang" + }, + { + "bbox": [ + 120, + 160, + 473, + 190 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 120, + 160, + 473, + 190 + ], + "type": "text", + "content": ", Junjie Wang" + }, + { + "bbox": [ + 120, + 160, + 473, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 120, + 160, + 473, + 190 + ], + "type": "text", + "content": ", Guanjun Wu" + }, + { + "bbox": [ + 120, + 160, + 473, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 120, + 160, + 473, + 190 + ], + "type": "text", + "content": ", Lingxi Xie" + }, + { + "bbox": [ + 120, + 160, + 473, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 120, + 160, + 473, + 190 + ], + "type": "text", + "content": ", Xiaopeng Zhang" + }, + { + "bbox": [ + 120, + 160, + 473, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 120, + 160, + 473, + 190 + ], + "type": "text", + "content": ", Wenyu Liu" + }, + { + "bbox": [ + 120, + 160, + 473, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 120, + 160, + 473, + 190 + ], + "type": "text", + "content": ", Qi Tian" + }, + { + "bbox": [ + 120, + 160, + 473, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 120, + 160, + 473, + 190 + ], + "type": "text", + "content": ", Xinggang Wang" + }, + { + "bbox": [ + 120, + 160, + 473, + 190 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 100, + 189, + 492, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 189, + 492, + 218 + ], + "spans": [ + { + "bbox": [ + 100, + 189, + 492, + 218 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 100, + 189, + 492, + 218 + ], + "type": "text", + "content": "School of EIC, Huazhong University of Science and Technology " + }, + { + "bbox": [ + 100, + 189, + 492, + 218 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 100, + 189, + 492, + 218 + ], + "type": "text", + "content": "Huawei Inc. " + }, + { + "bbox": [ + 100, + 189, + 492, + 218 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 100, + 189, + 492, + 218 + ], + "type": "text", + "content": "School of CS, Huazhong University of Science and Technology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 172, + 220, + 421, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 220, + 421, + 232 + ], + "spans": [ + { + "bbox": [ + 172, + 220, + 421, + 232 + ], + "type": "text", + "content": "{taoranyi, guajuwu, liuwy, xgwang}@hust.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 82, + 233, + 509, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 233, + 509, + 245 + ], + "spans": [ + { + "bbox": [ + 82, + 233, + 509, + 245 + ], + "type": "text", + "content": "{jaminfong, is.wangjunjie, 198808xc, zxphistory}@gmail.com tian.qil@huawei.com" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 47, + 247, + 152, + 392 + ], + "blocks": [ + { + "bbox": [ + 47, + 247, + 152, + 392 + ], + "lines": [ + { + "bbox": [ + 47, + 247, + 152, + 392 + ], + "spans": [ + { + "bbox": [ + 47, + 247, + 152, + 392 + ], + "type": "image", + "image_path": "b43f1cb133f73d9bb4060ed0df294a63043f74bfdc32fe9e00b27168bc0e64fe.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 396, + 546, + 429 + ], + "lines": [ + { + "bbox": [ + 45, + 396, + 546, + 429 + ], + "spans": [ + { + "bbox": [ + 45, + 396, + 546, + 429 + ], + "type": "text", + "content": "Figure 1. We propose a simple yet efficient framework called GaussianDreamer. It bridges the 3D and 2D diffusion models via Gaussian splatting, having both 3D consistency and rich generation details. Our method can complete training within 15 minutes on a single GPU and achieve real-time rendering." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 155, + 247, + 546, + 384 + ], + "blocks": [ + { + "bbox": [ + 155, + 247, + 546, + 384 + ], + "lines": [ + { + "bbox": [ + 155, + 247, + 546, + 384 + ], + "spans": [ + { + "bbox": [ + 155, + 247, + 546, + 384 + ], + "type": "image", + "image_path": "5f20021c96402161f3769d9d4f9e5b399c37d00b7500c4bfbfbbb693e9b49e4e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 249, + 385, + 380, + 395 + ], + "lines": [ + { + "bbox": [ + 249, + 385, + 380, + 395 + ], + "spans": [ + { + "bbox": [ + 249, + 385, + 380, + 395 + ], + "type": "text", + "content": "3D Gaussians with Temporal Variation" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 431, + 192, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 431, + 192, + 443 + ], + "spans": [ + { + "bbox": [ + 143, + 431, + 192, + 443 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 45, + 456, + 289, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 456, + 289, + 685 + ], + "spans": [ + { + "bbox": [ + 45, + 456, + 289, + 685 + ], + "type": "text", + "content": "In recent times, the generation of 3D assets from text prompts has shown impressive results. Both 2D and 3D diffusion models can help generate decent 3D objects based on prompts. 3D diffusion models have good 3D consistency, but their quality and generalization are limited as trainable 3D data is expensive and hard to obtain. 2D diffusion models enjoy strong abilities of generalization and fine generation, but 3D consistency is hard to guarantee. This paper attempts to bridge the power from the two types of diffusion models via the recent explicit and efficient 3D Gaussian splatting representation. A fast 3D object generation framework, named as GaussianDreamer, is proposed, where the 3D diffusion model provides priors for initialization and the 2D diffusion model enriches the geometry and appearance. Operations of noisy point growing and color perturbation are introduced to enhance the initialized Gaussians. Our GaussianDreamer can generate a high-quality 3D instance or 3D avatar within 15 minutes on one GPU, much faster than previous methods, while the gener" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 432, + 546, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 432, + 546, + 468 + ], + "spans": [ + { + "bbox": [ + 306, + 432, + 546, + 468 + ], + "type": "text", + "content": "ated instances can be directly rendered in real time. Demos and code are available at https://taoranyi.com/gaussiandreamer/." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 490, + 387, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 490, + 387, + 502 + ], + "spans": [ + { + "bbox": [ + 306, + 490, + 387, + 502 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 510, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 605 + ], + "type": "text", + "content": "3D asset generation has been an expensive and professional work in conventional pipelines. Recently, diffusion models [58] have achieved great success in creating high-quality and realistic 2D images. Many research works [2, 6, 9, 12, 15, 24, 32, 48, 52, 67, 70, 76, 77, 81, 90] try to transfer the power of 2D diffusion models to the 3D field for easing and assisting the process of 3D assets creation, e.g. the most common text-to-3D task." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "type": "text", + "content": "Here come two main streams for achieving this goal: (i) training a new diffusion model with 3D data [12, 15, 24, 48] (namely the 3D diffusion model) and (ii) lifting the 2D diffusion model to 3D [2, 6, 9, 32, 52, 58, 67, 70, 76, 77, 81, 90]. The former one is direct to implement and holds strong 3D consistency, but struggles to extend into a large generation domain as 3D data is usually hard and expensive to obtain. The scale of current 3D datasets is far smaller than 2D datasets. This results in the generated 3D assets falling" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 105, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 105, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 105, + 703 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 58, + 693, + 105, + 703 + ], + "type": "text", + "content": " Project lead." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "text", + "content": " Corresponding author." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6796" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": "short in dealing with complex text prompts and producing complex/fine geometry and appearance. The latter benefits from the large data domain of the 2D diffusion models, which can handle various text prompts and produce highly detailed and complex geometry and appearance. However, as 2D diffusion models are unaware of the camera view, the generated 3D assets are hard to form geometry consistency, especially for structure-complicated instances." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 168, + 289, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 168, + 289, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 289, + 430 + ], + "type": "text", + "content": "This paper proposes to use recent 3D Gaussian Splating [25] to bridge the two aforementioned approaches, simultaneously having the geometry consistency from 3D diffusion models and rich details from 2D diffusion models. 3D Gaussians are one type of efficient and explicit representation, which intrinsically enjoys geometry priors due to the point-cloud-like structure. Specifically, we use one of two types of 3D diffusion models: text-to-3D and text-to-motion diffusion models, e.g. Shap-E [24] and MDM [72] in our implementation, to generate a coarse 3D instance. Based on the coarse 3D instance, a group of 3D Gaussians are initialized. We introduce two operations of noisy point growing and color perturbation to supplement the initialized Gaussians for follow-up enriching the 3D instance. Then the 3D Gaussians can be improved and optimized by interacting with the 2D diffusion model via the Score Distillation Sampling [52] (SDS) loss. Due to the geometry priors from both the 3D diffusion model and 3D Gaussian Splating itself, the training process can be finished in a very short time. The generated 3D asset can be rendered in real time without transformation into structures like mesh via the splatting process." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 431, + 257, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 431, + 257, + 442 + ], + "spans": [ + { + "bbox": [ + 59, + 431, + 257, + 442 + ], + "type": "text", + "content": "Our contributions can be summarized as follows." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 444, + 287, + 574 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 444, + 287, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 444, + 287, + 490 + ], + "spans": [ + { + "bbox": [ + 47, + 444, + 287, + 490 + ], + "type": "text", + "content": "- We propose a text-to-3D method, named as Gaussian-Dreamer which bridges the 3D and 2D diffusion models via Gaussian splitting, enjoying both 3D consistency and rich generation details." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 491, + 287, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 491, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 491, + 287, + 525 + ], + "type": "text", + "content": "- Noisy point growing and color perturbation are introduced to supplement the initialized 3D Gaussians for further content enrichment." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 526, + 287, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 526, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 526, + 287, + 574 + ], + "type": "text", + "content": "- The overall method is simple and quite effective. A 3D instance can be generated within 15 minutes on one GPU, much faster than previous methods, and can be directly rendered in real time." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 586, + 138, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 138, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 138, + 597 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": "3D Pretrained Diffusion Models. Recently, text-to-3D asset generation using diffusion models has achieved great success. Currently, it is mainly divided into lifting 2D diffusion models to 3D and 3D pretrained diffusion models, the difference lies in whether the training data used is 2D or 3D. 3D pretrained diffusion models [12, 15, 24, 48], referred to as 3D diffusion models in our paper, are models pretrained on text-3D pairs. After pretraining, they can generate 3D assets through only inference, and mod" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 547, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 216 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 216 + ], + "type": "text", + "content": "els such as Point-E [48] and Shape-E [24] can generate 3D assets in minutes. In addition to generating 3D assets from text, there are methods where 3D diffusion models [1, 7, 10, 26, 55, 63, 72, 88, 89, 91] generate motion sequences based on text-motion data. By pretraining on text-motion pairs, these models can generate reasonable motion sequences for different texts. The generated motion sequences can be transformed into the SMPL (Skinned Multi-Person Linear) model [39] based on mesh representation, but texture information is not included. In our method, we can paint transformed SMPL by using different text prompts." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 231, + 547, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 231, + 547, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 231, + 547, + 495 + ], + "type": "text", + "content": "Lifting 2D Diffusion Models to 3D. In text-to-3D asset generation methods [8, 16, 21, 23, 40, 49, 50, 53, 57, 61, 62, 68, 73, 80, 85, 87, 92], in addition to using 3D pretrained diffusion models, lifting a 2D diffusion model to 3D is a training-free approach. Moreover, due to the abundance of 2D image data, this method produces assets with higher diversity and fidelity. Some single-image-to-3D methods [13, 19, 30, 33–36, 38, 54, 60, 65, 66, 71, 78, 79, 82, 83] also employ similar ideas. DreamFusion [52] first proposes SDS (Score Distillation Sampling) method, which is to update the 3D representation model using the 2D diffusion model. [76] proposes the method of SJC (Score Jacobian Chaining) to lift the 2D diffusion model to 3D. Later methods [6, 31, 32, 69, 77] build on DreamFusion and further improve the quality of 3D generation. Among them, generated 3D assets may suffer from multi-face problems. To address this issue, some methods strengthen the semantics of different views [2] and use multi-view information [67, 90] to alleviate such problems. There are also models [18, 22, 29, 44, 46, 59, 75, 81] which adopt CLIP [56] to align each view of the 3D representation model with the text." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 510, + 547, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 547, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 547, + 629 + ], + "type": "text", + "content": "3D Representation Methods. In recent times, neural radiance fields (NeRF) [45] have achieved impressive results in 3D representation, and many methods in text-to-3D asset generation have also adopted NeRF or its variants [3, 47] as the representation method. Some methods [6, 12, 31, 32] use explicit optimization mesh representation methods like DMTET [64] to reduce rendering costs and further improve resolution. In addition to that, there are also generation methods that utilize point clouds [41, 48, 53, 74] and meshes [37] as 3D representations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "content": "Recently, 3D Gaussian Splatting [25] has been introduced as a representation method for 3D scenes, which can achieve rendering effects comparable to NeRF-based methods and enable real-time rendering. Two concurrent works [9, 70] also construct the 3D representation with 3D Gaussian Splatting [25]. DreamGaussian [70] uses a single image as a condition to generate 3D assets, while GS-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6797" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 68, + 547, + 210 + ], + "blocks": [ + { + "bbox": [ + 48, + 68, + 547, + 210 + ], + "lines": [ + { + "bbox": [ + 48, + 68, + 547, + 210 + ], + "spans": [ + { + "bbox": [ + 48, + 68, + 547, + 210 + ], + "type": "image", + "image_path": "79c713f12bc171b2a6e360462360b1ffac48208e70c3cc7d2d1e66999abcfca3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 213, + 547, + 270 + ], + "lines": [ + { + "bbox": [ + 46, + 213, + 547, + 270 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 547, + 270 + ], + "type": "text", + "content": "Figure 2. Overall framework of GaussianDreamer. Firstly, we utilize a 3D diffusion model to generate the initialized point clouds. After executing noisy point growing and color perturbation on the point clouds, we use them to initialize the 3D Gaussians. The initialized 3D Gaussians are further optimized using the SDS method [52] with a 2D diffusion model. Finally, we render the image using the 3D Gaussians by employing 3D Gaussian Splatting [25]. We can use one of various 3D diffusion models to generate the initialized point clouds. In this case, we take text-to-3D and text-to-motion diffusion models as examples." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 275, + 289, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 275, + 289, + 360 + ], + "spans": [ + { + "bbox": [ + 46, + 275, + 289, + 360 + ], + "type": "text", + "content": "GEN [9] implements high-quality generation from text to 3D. Our method shares a similar idea of using 3D Gaussian Splating as the representation method, which significantly reduces the cost of improving resolution and achieving a much faster optimization speed compared to estimizable mesh representation methods. And we can generate detailed 3D assets based on prompt texts in a very short time." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 375, + 104, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 375, + 104, + 387 + ], + "spans": [ + { + "bbox": [ + 47, + 375, + 104, + 387 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 398, + 288, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 398, + 288, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 398, + 288, + 483 + ], + "type": "text", + "content": "In this section, we first review 2D and 3D diffusion models and the 3D representation method - 3D Gaussian Splating [25]. We give an overview of the whole framework in Sec. 3.2. Then, in Sec. 3.3, we describe the process of initializing the 3D Gaussians with the assistance of 3D diffusion models. The further optimization of 3D Gaussians using the 2D diffusion model is described in Sec. 3.4." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 500, + 134, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 500, + 134, + 511 + ], + "spans": [ + { + "bbox": [ + 47, + 500, + 134, + 511 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "text", + "content": "DreamFusion. DreamFusion [52] is one of the most representative methods to lift 2D diffusion models to 3D, which proposes to optimize the 3D representation with the score distillation sampling (SDS) loss via a pre-trained 2D diffusion model " + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "text", + "content": ". Specifically, it takes MipNeRF [3] as the 3D representation method, whose parameters " + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "text", + "content": " are optimized. Taking the rendering method as " + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "text", + "content": ", the rendered image results in " + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x} = g(\\theta)" + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "text", + "content": ". To make the rendered image " + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "text", + "content": " similar to the samples obtained from the diffusion model " + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "text", + "content": ", DreamFusion uses a scoring estimation function: " + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_{\\phi}(\\mathbf{z}_t; y, t)" + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "text", + "content": ", which predicts the sampled noise " + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_{\\phi}" + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "text", + "content": " given the noisy image " + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "text", + "content": ", text embedding " + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "text", + "content": ", and noise level " + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "text", + "content": ". By measuring the difference between the Gaussian noise " + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "text", + "content": " added to the rendered image " + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "text", + "content": " and the predicted noise " + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_{\\phi}" + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "text", + "content": ", this scoring estimation function can provide the direction for updating the parameter " + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "text", + "content": ". The formula for computing" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 276, + 378, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 276, + 378, + 288 + ], + "spans": [ + { + "bbox": [ + 306, + 276, + 378, + 288 + ], + "type": "text", + "content": "the gradient is as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 298, + 547, + 337 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 298, + 547, + 337 + ], + "spans": [ + { + "bbox": [ + 306, + 298, + 547, + 337 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {S D S}} (\\phi , \\mathbf {x} = g (\\theta)) \\triangleq \\mathbb {E} _ {t, \\epsilon} \\left[ w (t) \\left(\\hat {\\epsilon} _ {\\phi} \\left(\\mathbf {z} _ {t}; y, t\\right) - \\epsilon\\right) \\frac {\\partial \\mathbf {x}}{\\partial \\theta} \\right], \\tag {1}", + "image_path": "30b88a89dce6e109938301b9e6e2844048198e58343b0a626f2903587027bc61.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 349, + 451, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 349, + 451, + 361 + ], + "spans": [ + { + "bbox": [ + 306, + 349, + 451, + 361 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 349, + 451, + 361 + ], + "type": "inline_equation", + "content": "w(t)" + }, + { + "bbox": [ + 306, + 349, + 451, + 361 + ], + "type": "text", + "content": " is a weighting function." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 378, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 378, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 305, + 378, + 545, + 498 + ], + "type": "text", + "content": "3D Gaussian Splatting. 3D Gaussian Splatting [25] (3D-GS) is a recent groundbreaking method for novel-view synthesis. Unlike implicit representation methods such as NeRF [45], which renders images based on volume rendering, 3D-GS renders images through splatting [84], achieving real-time speed. Specifically, 3D-GS represents the scene through a set of anisotropic Gaussians, defined with its center position " + }, + { + "bbox": [ + 305, + 378, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\mu \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 305, + 378, + 545, + 498 + ], + "type": "text", + "content": ", covariance " + }, + { + "bbox": [ + 305, + 378, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\Sigma \\in \\mathbb{R}^7" + }, + { + "bbox": [ + 305, + 378, + 545, + 498 + ], + "type": "text", + "content": ", color " + }, + { + "bbox": [ + 305, + 378, + 545, + 498 + ], + "type": "inline_equation", + "content": "c \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 305, + 378, + 545, + 498 + ], + "type": "text", + "content": ", and opacity " + }, + { + "bbox": [ + 305, + 378, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\alpha \\in \\mathbb{R}^1" + }, + { + "bbox": [ + 305, + 378, + 545, + 498 + ], + "type": "text", + "content": ". And the 3D Gaussians can be queried as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 375, + 510, + 545, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 510, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 375, + 510, + 545, + 525 + ], + "type": "interline_equation", + "content": "G (x) = e ^ {- \\frac {1}{2} (x) ^ {T} \\Sigma^ {- 1} (x)}, \\tag {2}", + "image_path": "28ae95ea8c83e93777ea4cc6565c8f1906f4bff1f50b341afd0e8c1f80867ae3.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 537, + 547, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 537, + 547, + 609 + ], + "spans": [ + { + "bbox": [ + 304, + 537, + 547, + 609 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 537, + 547, + 609 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 537, + 547, + 609 + ], + "type": "text", + "content": " represents the distance between " + }, + { + "bbox": [ + 304, + 537, + 547, + 609 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 537, + 547, + 609 + ], + "type": "text", + "content": " and the query point. For computing the color of each pixel, it uses a typical neural point-based rendering [27, 28]. A ray " + }, + { + "bbox": [ + 304, + 537, + 547, + 609 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 537, + 547, + 609 + ], + "type": "text", + "content": " is cast from the center of the camera, and the color and density of the 3D Gaussians that the ray intersects are computed along the ray. The rendering process is as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 321, + 621, + 545, + 654 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 621, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 321, + 621, + 545, + 654 + ], + "type": "interline_equation", + "content": "C (r) = \\sum_ {i \\in \\mathcal {N}} c _ {i} \\sigma_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\sigma_ {j}), \\quad \\sigma_ {i} = \\alpha_ {i} G (x _ {i}), \\tag {3}", + "image_path": "0a1eaf21ab377f83cac50c97eae0d80f7c15eca21781b53e440d0f55a240be32.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 665, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 712 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 665, + 547, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 304, + 665, + 547, + 712 + ], + "type": "text", + "content": " represents the number of sample points on the ray " + }, + { + "bbox": [ + 304, + 665, + 547, + 712 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 665, + 547, + 712 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 665, + 547, + 712 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 304, + 665, + 547, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 665, + 547, + 712 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 304, + 665, + 547, + 712 + ], + "type": "text", + "content": " denote the color and opacity of the i-th Gaussian, and " + }, + { + "bbox": [ + 304, + 665, + 547, + 712 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 304, + 665, + 547, + 712 + ], + "type": "text", + "content": " is the distance between the point and the i-th Gaussian." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6798" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 163, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 163, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 163, + 83 + ], + "type": "text", + "content": "3.2. Overall Framework" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 90, + 289, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 90, + 289, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 90, + 289, + 281 + ], + "type": "text", + "content": "Our overall framework consists of two parts, initialization with 3D diffusion model priors and optimization with the 2D diffusion model, as shown in Fig. 2. For initialization with 3D diffusion model priors, we use the 3D diffusion models " + }, + { + "bbox": [ + 46, + 90, + 289, + 281 + ], + "type": "inline_equation", + "content": "F_{3D}" + }, + { + "bbox": [ + 46, + 90, + 289, + 281 + ], + "type": "text", + "content": ", instantiated with the text-to-3D and text-to-motion diffusion models, to generate the triangle mesh " + }, + { + "bbox": [ + 46, + 90, + 289, + 281 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 90, + 289, + 281 + ], + "type": "text", + "content": " based on the text prompt " + }, + { + "bbox": [ + 46, + 90, + 289, + 281 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 90, + 289, + 281 + ], + "type": "text", + "content": ", which can be denoted as " + }, + { + "bbox": [ + 46, + 90, + 289, + 281 + ], + "type": "inline_equation", + "content": "m = F_{3D}(y)" + }, + { + "bbox": [ + 46, + 90, + 289, + 281 + ], + "type": "text", + "content": ". One set of generated point clouds is transformed from the mesh " + }, + { + "bbox": [ + 46, + 90, + 289, + 281 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 90, + 289, + 281 + ], + "type": "text", + "content": ". Then the 3D Gaussians " + }, + { + "bbox": [ + 46, + 90, + 289, + 281 + ], + "type": "inline_equation", + "content": "\\theta_b" + }, + { + "bbox": [ + 46, + 90, + 289, + 281 + ], + "type": "text", + "content": " are initialized via the generated point clouds after noisy point growing and color perturbation. For better quality, we utilize the 2D diffusion model " + }, + { + "bbox": [ + 46, + 90, + 289, + 281 + ], + "type": "inline_equation", + "content": "F_{2D}" + }, + { + "bbox": [ + 46, + 90, + 289, + 281 + ], + "type": "text", + "content": " to further optimize the initialized 3D Gaussians " + }, + { + "bbox": [ + 46, + 90, + 289, + 281 + ], + "type": "inline_equation", + "content": "\\theta_b" + }, + { + "bbox": [ + 46, + 90, + 289, + 281 + ], + "type": "text", + "content": " via SDS [52] with prompts " + }, + { + "bbox": [ + 46, + 90, + 289, + 281 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 90, + 289, + 281 + ], + "type": "text", + "content": ", resulting in the final 3D Gaussians " + }, + { + "bbox": [ + 46, + 90, + 289, + 281 + ], + "type": "inline_equation", + "content": "\\theta_f" + }, + { + "bbox": [ + 46, + 90, + 289, + 281 + ], + "type": "text", + "content": ". The target instance can be rendered in real time by splatting the generated Gaussians." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 287, + 287, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 287, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 47, + 287, + 287, + 312 + ], + "type": "text", + "content": "3.3. Gaussian Initialization with 3D Diffusion Model Priors" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 318, + 287, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 318, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 318, + 287, + 413 + ], + "type": "text", + "content": "In this section, we mainly discuss how to initialize the 3D Gaussians with 3D diffusion model priors. First, we use the 3D diffusion model " + }, + { + "bbox": [ + 46, + 318, + 287, + 413 + ], + "type": "inline_equation", + "content": "F_{3D}" + }, + { + "bbox": [ + 46, + 318, + 287, + 413 + ], + "type": "text", + "content": " to generate 3D assets based on the prompts " + }, + { + "bbox": [ + 46, + 318, + 287, + 413 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 318, + 287, + 413 + ], + "type": "text", + "content": ". Then we convert the 3D assets into point clouds and use the transformed point clouds to initialize the 3D Gaussians. We have employed two types of 3D diffusion models to generate 3D assets. Below, we explain how to initialize the 3D Gaussians using each of these models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 422, + 198, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 422, + 198, + 434 + ], + "spans": [ + { + "bbox": [ + 47, + 422, + 198, + 434 + ], + "type": "text", + "content": "3.3.1 Text-to-3D Diffusion Model" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 441, + 287, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 441, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 441, + 287, + 586 + ], + "type": "text", + "content": "When using a text-based 3D generation model, generated 3D assets employ multi-layer perceptrons (MLPs) to predict SDF values and texture colors. To construct the triangle mesh " + }, + { + "bbox": [ + 46, + 441, + 287, + 586 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 441, + 287, + 586 + ], + "type": "text", + "content": ", we query the SDF values at vertices along a regular grid with the size of " + }, + { + "bbox": [ + 46, + 441, + 287, + 586 + ], + "type": "inline_equation", + "content": "128^3" + }, + { + "bbox": [ + 46, + 441, + 287, + 586 + ], + "type": "text", + "content": " within the MLPs. Then we query the texture colors at each vertex of " + }, + { + "bbox": [ + 46, + 441, + 287, + 586 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 441, + 287, + 586 + ], + "type": "text", + "content": ". We convert the vertices and colors of " + }, + { + "bbox": [ + 46, + 441, + 287, + 586 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 441, + 287, + 586 + ], + "type": "text", + "content": " into point clouds, denoted as " + }, + { + "bbox": [ + 46, + 441, + 287, + 586 + ], + "type": "inline_equation", + "content": "\\pmb{p}\\pmb{t}_m(\\pmb{p}_m,\\pmb{c}_m)" + }, + { + "bbox": [ + 46, + 441, + 287, + 586 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 441, + 287, + 586 + ], + "type": "inline_equation", + "content": "\\pmb{p}_m\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 441, + 287, + 586 + ], + "type": "text", + "content": " refers to the position of the point clouds, which equals to the vertex coordinated of " + }, + { + "bbox": [ + 46, + 441, + 287, + 586 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 441, + 287, + 586 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 441, + 287, + 586 + ], + "type": "inline_equation", + "content": "c_{m}\\in \\mathbb{R}^{3}" + }, + { + "bbox": [ + 46, + 441, + 287, + 586 + ], + "type": "text", + "content": " refers to the color of the point clouds, which is the same as the color of " + }, + { + "bbox": [ + 46, + 441, + 287, + 586 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 441, + 287, + 586 + ], + "type": "text", + "content": ". However, the obtained colors " + }, + { + "bbox": [ + 46, + 441, + 287, + 586 + ], + "type": "inline_equation", + "content": "c_{m}" + }, + { + "bbox": [ + 46, + 441, + 287, + 586 + ], + "type": "text", + "content": " are relatively simple, and the positions " + }, + { + "bbox": [ + 46, + 441, + 287, + 586 + ], + "type": "inline_equation", + "content": "p_m" + }, + { + "bbox": [ + 46, + 441, + 287, + 586 + ], + "type": "text", + "content": " are sparse." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": "Noisy Point Growing and Color Perturbation. We do not use the generated point clouds " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{pt}_m" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": " to directly initialize the 3D Gaussians. To improve the quality of initialization, we perform noisy point growing and color perturbation around point clouds " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{pt}_m" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": ". First, we compute the bounding box (BBox) of the surface on " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{pt}_m" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": " and then uniformly grow point clouds " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{pt}_r(\\pmb{p}_r, \\pmb{c}_r)" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": " within the BBox. " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{p}_r" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{c}_r" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": " represent the positions and colors of " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{pt}_r" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": ". To enable fast searching, we construct a KDTree [4] " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "K_m" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": " using the positions " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{p}_m" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": ". Based on the distances between the positions " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{p}_r" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 312, + 68, + 541, + 186 + ], + "blocks": [ + { + "bbox": [ + 312, + 68, + 541, + 186 + ], + "lines": [ + { + "bbox": [ + 312, + 68, + 541, + 186 + ], + "spans": [ + { + "bbox": [ + 312, + 68, + 541, + 186 + ], + "type": "image", + "image_path": "bb582a35d7e07fd94b09d153ed9d7211e1ceb6f1e72bd01be21944fcde6c8680.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 193, + 545, + 226 + ], + "lines": [ + { + "bbox": [ + 304, + 193, + 545, + 226 + ], + "spans": [ + { + "bbox": [ + 304, + 193, + 545, + 226 + ], + "type": "text", + "content": "Figure 3. The process of noisy point growing and color perturbation. \"Grow&Pertb.\" denotes noisy point growing and color perturbation." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 235, + 545, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 235, + 545, + 295 + ], + "spans": [ + { + "bbox": [ + 304, + 235, + 545, + 295 + ], + "type": "text", + "content": "and the nearest points found in the KDTree " + }, + { + "bbox": [ + 304, + 235, + 545, + 295 + ], + "type": "inline_equation", + "content": "K_{m}" + }, + { + "bbox": [ + 304, + 235, + 545, + 295 + ], + "type": "text", + "content": ", we determine which points to keep. In this process, we select points within the (normalized) distance of 0.01. For the noisy point clouds, we make their colors " + }, + { + "bbox": [ + 304, + 235, + 545, + 295 + ], + "type": "inline_equation", + "content": "\\pmb{c}_r" + }, + { + "bbox": [ + 304, + 235, + 545, + 295 + ], + "type": "text", + "content": " similar to " + }, + { + "bbox": [ + 304, + 235, + 545, + 295 + ], + "type": "inline_equation", + "content": "\\pmb{c}_m" + }, + { + "bbox": [ + 304, + 235, + 545, + 295 + ], + "type": "text", + "content": ", and also add some perturbations:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 395, + 306, + 545, + 317 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 306, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 395, + 306, + 545, + 317 + ], + "type": "interline_equation", + "content": "\\boldsymbol {c} _ {r} = \\boldsymbol {c} _ {m} + \\boldsymbol {a}, \\tag {4}", + "image_path": "7706d51e3b13670d0a120a724572d74d456ff76165ea15377bacfc2763491304.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 327, + 545, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 545, + 363 + ], + "type": "text", + "content": "where the values of " + }, + { + "bbox": [ + 304, + 327, + 545, + 363 + ], + "type": "inline_equation", + "content": "\\mathbf{a}" + }, + { + "bbox": [ + 304, + 327, + 545, + 363 + ], + "type": "text", + "content": " are randomly sampled between 0 and 0.2. We merge the positions and colors of " + }, + { + "bbox": [ + 304, + 327, + 545, + 363 + ], + "type": "inline_equation", + "content": "\\mathbf{p}\\mathbf{t}_{m}" + }, + { + "bbox": [ + 304, + 327, + 545, + 363 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 327, + 545, + 363 + ], + "type": "inline_equation", + "content": "\\mathbf{p}\\mathbf{t}_{r}" + }, + { + "bbox": [ + 304, + 327, + 545, + 363 + ], + "type": "text", + "content": " to obtain the final point clouds." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 352, + 373, + 545, + 385 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 373, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 352, + 373, + 545, + 385 + ], + "type": "interline_equation", + "content": "\\boldsymbol {p} \\boldsymbol {t} \\left(\\boldsymbol {p} _ {f}, \\boldsymbol {c} _ {f}\\right) = \\left(\\boldsymbol {p} _ {m} \\oplus \\boldsymbol {p} _ {r}, \\boldsymbol {c} _ {m} \\oplus \\boldsymbol {c} _ {r}\\right), \\tag {5}", + "image_path": "b39df4822ba52ec1907655bf4dd7b50f75f33eb7d6a1a799aed74a0b4da25e0d.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 395, + 545, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 395, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 395, + 545, + 491 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 395, + 545, + 491 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 304, + 395, + 545, + 491 + ], + "type": "text", + "content": " is the concatenation operation. Fig. 3 illustrates the process of noisy point growing and color perturbation. Finally, we initialize the positions " + }, + { + "bbox": [ + 304, + 395, + 545, + 491 + ], + "type": "inline_equation", + "content": "\\mu_{b}" + }, + { + "bbox": [ + 304, + 395, + 545, + 491 + ], + "type": "text", + "content": " and colors " + }, + { + "bbox": [ + 304, + 395, + 545, + 491 + ], + "type": "inline_equation", + "content": "c_{b}" + }, + { + "bbox": [ + 304, + 395, + 545, + 491 + ], + "type": "text", + "content": " of the 3D Gaussians " + }, + { + "bbox": [ + 304, + 395, + 545, + 491 + ], + "type": "inline_equation", + "content": "\\theta_{b}(\\mu_{b}, c_{b}, \\Sigma_{b}, \\alpha_{b})" + }, + { + "bbox": [ + 304, + 395, + 545, + 491 + ], + "type": "text", + "content": " using both the positions " + }, + { + "bbox": [ + 304, + 395, + 545, + 491 + ], + "type": "inline_equation", + "content": "\\pmb{p}_{f}" + }, + { + "bbox": [ + 304, + 395, + 545, + 491 + ], + "type": "text", + "content": " and colors " + }, + { + "bbox": [ + 304, + 395, + 545, + 491 + ], + "type": "inline_equation", + "content": "\\pmb{c}_{f}" + }, + { + "bbox": [ + 304, + 395, + 545, + 491 + ], + "type": "text", + "content": " of final point clouds " + }, + { + "bbox": [ + 304, + 395, + 545, + 491 + ], + "type": "inline_equation", + "content": "\\pmb{pt}" + }, + { + "bbox": [ + 304, + 395, + 545, + 491 + ], + "type": "text", + "content": ". The opacity " + }, + { + "bbox": [ + 304, + 395, + 545, + 491 + ], + "type": "inline_equation", + "content": "\\alpha_{b}" + }, + { + "bbox": [ + 304, + 395, + 545, + 491 + ], + "type": "text", + "content": " of the 3D Gaussians is initialized to 0.1, and the covariance " + }, + { + "bbox": [ + 304, + 395, + 545, + 491 + ], + "type": "inline_equation", + "content": "\\Sigma_{b}" + }, + { + "bbox": [ + 304, + 395, + 545, + 491 + ], + "type": "text", + "content": " is calculated as the distance between the nearest two points. Algorithm 1 shows the specific algorithm flowchart." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 495, + 474, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 495, + 474, + 506 + ], + "spans": [ + { + "bbox": [ + 306, + 495, + 474, + 506 + ], + "type": "text", + "content": "3.3.2 Text-to-Motion Diffusion Model" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "text", + "content": "We generate a sequence of human body motions using text and select a human pose that best matches the given text. We then convert the keypoints of this human pose into the SMPL model [39], which is represented by a triangle mesh " + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "text", + "content": ". We then convert the mesh " + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "text", + "content": " into point clouds " + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "inline_equation", + "content": "\\pmb{pt}_m(\\pmb{p}_m, \\pmb{c}_m)" + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "text", + "content": ", where the position " + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "inline_equation", + "content": "\\pmb{p}_m" + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "text", + "content": " of each point in the point clouds corresponds to the vertices of " + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "text", + "content": ". As for the color " + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "inline_equation", + "content": "\\pmb{c}_m" + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "inline_equation", + "content": "\\pmb{pt}_m" + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "text", + "content": ", since the SMPL model used here does not have textures, we randomly initialize " + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "inline_equation", + "content": "\\pmb{c}_m" + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "text", + "content": ". To move " + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "inline_equation", + "content": "\\pmb{pt}_m" + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "text", + "content": " near the origin, we calculate the center point " + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "inline_equation", + "content": "\\pmb{p}_c \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "inline_equation", + "content": "\\pmb{p}_m" + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "text", + "content": " and subtract the position of point clouds " + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "inline_equation", + "content": "\\pmb{pt}_m" + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "text", + "content": " from the center point " + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "inline_equation", + "content": "\\pmb{p}_c" + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 354, + 668, + 545, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 668, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 354, + 668, + 545, + 681 + ], + "type": "interline_equation", + "content": "\\boldsymbol {p} \\boldsymbol {t} \\left(\\boldsymbol {p} _ {f}, \\boldsymbol {c} _ {f}\\right) = \\boldsymbol {p} \\boldsymbol {t} _ {m} \\left(\\boldsymbol {p} _ {m} - \\boldsymbol {p} _ {c}, \\boldsymbol {c} _ {m}\\right), \\tag {6}", + "image_path": "9fc52ea7a5ffede09a3385d596e61da4a24c77fbff2a2f833ee3b03aa287630e.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": "Finally, we use the point clouds " + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "pt" + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": " to initialize the 3D Gaussians, similar to what is described in Sec. 3.3.1. To improve" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6799" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 47, + 72, + 287, + 144 + ], + "blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 144 + ], + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 144 + ], + "type": "text", + "content": "Algorithm 1 The 3D Gaussian Initialization. \n" + }, + { + "bbox": [ + 47, + 72, + 287, + 144 + ], + "type": "inline_equation", + "content": "\\pmb{pt}_m(\\pmb{p}_m, \\pmb{c}_m)" + }, + { + "bbox": [ + 47, + 72, + 287, + 144 + ], + "type": "text", + "content": ": Point clouds generated from " + }, + { + "bbox": [ + 47, + 72, + 287, + 144 + ], + "type": "inline_equation", + "content": "F_{3D}" + }, + { + "bbox": [ + 47, + 72, + 287, + 144 + ], + "type": "text", + "content": ". \n" + }, + { + "bbox": [ + 47, + 72, + 287, + 144 + ], + "type": "inline_equation", + "content": "\\pmb{pt}_r(\\pmb{p}_r, \\pmb{c}_r)" + }, + { + "bbox": [ + 47, + 72, + 287, + 144 + ], + "type": "text", + "content": ": Growing point clouds within the BBox. \n" + }, + { + "bbox": [ + 47, + 72, + 287, + 144 + ], + "type": "inline_equation", + "content": "\\pmb{pt}(\\pmb{p}_f, \\pmb{c}_f)" + }, + { + "bbox": [ + 47, + 72, + 287, + 144 + ], + "type": "text", + "content": ": Point clouds used for initializing the 3D Gaussians. \n" + }, + { + "bbox": [ + 47, + 72, + 287, + 144 + ], + "type": "inline_equation", + "content": "\\theta_b(\\mu_b, c_b, \\Sigma_b, \\alpha_b)" + }, + { + "bbox": [ + 47, + 72, + 287, + 144 + ], + "type": "text", + "content": ": Initialized 3D Gaussians." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_body" + } + ], + "index": 0, + "sub_type": "algorithm" + }, + { + "type": "code", + "bbox": [ + 57, + 148, + 287, + 231 + ], + "blocks": [ + { + "bbox": [ + 57, + 148, + 287, + 231 + ], + "lines": [ + { + "bbox": [ + 57, + 148, + 287, + 231 + ], + "spans": [ + { + "bbox": [ + 57, + 148, + 287, + 231 + ], + "type": "text", + "content": "Stage 1: Generate points in the BBox. \n" + }, + { + "bbox": [ + 57, + 148, + 287, + 231 + ], + "type": "inline_equation", + "content": "K_{m} \\leftarrow" + }, + { + "bbox": [ + 57, + 148, + 287, + 231 + ], + "type": "text", + "content": " BuildKDTree(" + }, + { + "bbox": [ + 57, + 148, + 287, + 231 + ], + "type": "inline_equation", + "content": "p_{m}" + }, + { + "bbox": [ + 57, + 148, + 287, + 231 + ], + "type": "text", + "content": ") \n" + }, + { + "bbox": [ + 57, + 148, + 287, + 231 + ], + "type": "inline_equation", + "content": "\\text{BBox} \\leftarrow p_{m}" + }, + { + "bbox": [ + 57, + 148, + 287, + 231 + ], + "type": "inline_equation", + "content": "\\text{BBox} \\leftarrow" + }, + { + "bbox": [ + 57, + 148, + 287, + 231 + ], + "type": "text", + "content": " Positions bounding box. \nLow, High " + }, + { + "bbox": [ + 57, + 148, + 287, + 231 + ], + "type": "inline_equation", + "content": "\\leftarrow" + }, + { + "bbox": [ + 57, + 148, + 287, + 231 + ], + "type": "text", + "content": " BBox.MinBound, BBox.MaxBound \n" + }, + { + "bbox": [ + 57, + 148, + 287, + 231 + ], + "type": "inline_equation", + "content": "\\text{BBox} \\leftarrow" + }, + { + "bbox": [ + 57, + 148, + 287, + 231 + ], + "type": "text", + "content": " Boundary of bounding box. \n" + }, + { + "bbox": [ + 57, + 148, + 287, + 231 + ], + "type": "inline_equation", + "content": "\\boldsymbol{ps}_{u} \\leftarrow" + }, + { + "bbox": [ + 57, + 148, + 287, + 231 + ], + "type": "text", + "content": " Uniform(Low, High, size = (NumPoints, 3)) \n" + }, + { + "bbox": [ + 57, + 148, + 287, + 231 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 57, + 148, + 287, + 231 + ], + "type": "text", + "content": " Points in the BBox." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "algorithm" + }, + { + "bbox": [ + 57, + 232, + 287, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 232, + 287, + 254 + ], + "spans": [ + { + "bbox": [ + 57, + 232, + 287, + 254 + ], + "type": "text", + "content": "Stage 2: Keep the points that meet the distance requirement." + } + ] + } + ], + "index": 2 + }, + { + "type": "code", + "bbox": [ + 57, + 255, + 287, + 350 + ], + "blocks": [ + { + "bbox": [ + 57, + 255, + 287, + 350 + ], + "lines": [ + { + "bbox": [ + 57, + 255, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 57, + 255, + 287, + 350 + ], + "type": "inline_equation", + "content": "\\pmb{p}_r,\\pmb {c}_r = \\left[ \\begin{array}{ll} \\end{array} \\right],\\left[ \\begin{array}{ll} \\end{array} \\right]" + }, + { + "bbox": [ + 57, + 255, + 287, + 350 + ], + "type": "text", + "content": " \nfor all " + }, + { + "bbox": [ + 57, + 255, + 287, + 350 + ], + "type": "inline_equation", + "content": "\\pmb{p_u}" + }, + { + "bbox": [ + 57, + 255, + 287, + 350 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 57, + 255, + 287, + 350 + ], + "type": "inline_equation", + "content": "p s_{u}" + }, + { + "bbox": [ + 57, + 255, + 287, + 350 + ], + "type": "text", + "content": " do " + }, + { + "bbox": [ + 57, + 255, + 287, + 350 + ], + "type": "inline_equation", + "content": "\\pmb{p}_{un},i\\gets K_m.\\mathrm{SearchNearest}(\\pmb {p}_u)" + }, + { + "bbox": [ + 57, + 255, + 287, + 350 + ], + "type": "text", + "content": " Nearest point and its index in " + }, + { + "bbox": [ + 57, + 255, + 287, + 350 + ], + "type": "inline_equation", + "content": "K_{m}" + }, + { + "bbox": [ + 57, + 255, + 287, + 350 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 57, + 255, + 287, + 350 + ], + "type": "inline_equation", + "content": "|\\pmb {p}_{un} - \\pmb {p}_u| < 0.01" + }, + { + "bbox": [ + 57, + 255, + 287, + 350 + ], + "type": "text", + "content": " then " + }, + { + "bbox": [ + 57, + 255, + 287, + 350 + ], + "type": "inline_equation", + "content": "\\pmb{p}_r" + }, + { + "bbox": [ + 57, + 255, + 287, + 350 + ], + "type": "text", + "content": " .append " + }, + { + "bbox": [ + 57, + 255, + 287, + 350 + ], + "type": "inline_equation", + "content": "(\\pmb {p}_u)" + }, + { + "bbox": [ + 57, + 255, + 287, + 350 + ], + "type": "inline_equation", + "content": "c_{r}" + }, + { + "bbox": [ + 57, + 255, + 287, + 350 + ], + "type": "text", + "content": " .append " + }, + { + "bbox": [ + 57, + 255, + 287, + 350 + ], + "type": "inline_equation", + "content": "(c_{m}[i] + 0.2\\times \\mathrm{Random}(\\mathrm{size} = 3))" + }, + { + "bbox": [ + 57, + 255, + 287, + 350 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 57, + 255, + 287, + 350 + ], + "type": "text", + "content": " Color of the nearest point plus perturbation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "type": "code", + "bbox": [ + 57, + 351, + 287, + 470 + ], + "blocks": [ + { + "bbox": [ + 57, + 351, + 287, + 470 + ], + "lines": [ + { + "bbox": [ + 57, + 351, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 57, + 351, + 287, + 470 + ], + "type": "text", + "content": "end if \nend for \n" + }, + { + "bbox": [ + 57, + 351, + 287, + 470 + ], + "type": "inline_equation", + "content": "\\pmb {p}_f\\gets \\pmb {p}_m\\oplus \\pmb {p}_r" + }, + { + "bbox": [ + 57, + 351, + 287, + 470 + ], + "type": "inline_equation", + "content": "c_{f}\\gets c_{m}\\oplus c_{r}" + }, + { + "bbox": [ + 57, + 351, + 287, + 470 + ], + "type": "text", + "content": " \nStage 3: Initialize the 3D Gaussians. \n" + }, + { + "bbox": [ + 57, + 351, + 287, + 470 + ], + "type": "inline_equation", + "content": "\\mu_b,c_b\\gets p_f,c_f" + }, + { + "bbox": [ + 57, + 351, + 287, + 470 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 57, + 351, + 287, + 470 + ], + "type": "text", + "content": " Positions and colors of the 3D Gaussians. \n" + }, + { + "bbox": [ + 57, + 351, + 287, + 470 + ], + "type": "inline_equation", + "content": "D\\leftarrow \\mu_{b}\\quad \\triangleright" + }, + { + "bbox": [ + 57, + 351, + 287, + 470 + ], + "type": "text", + "content": " Distance between the nearest two positions. \n" + }, + { + "bbox": [ + 57, + 351, + 287, + 470 + ], + "type": "inline_equation", + "content": "\\Sigma_{b},\\alpha_{b}\\gets D,0.1" + }, + { + "bbox": [ + 57, + 351, + 287, + 470 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 57, + 351, + 287, + 470 + ], + "type": "text", + "content": " Covariance and opacity of the 3D Gaussians." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "code_body" + } + ], + "index": 4, + "sub_type": "algorithm" + }, + { + "bbox": [ + 46, + 492, + 287, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 492, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 46, + 492, + 287, + 553 + ], + "type": "text", + "content": "the generation of motion sequences, we simplify the text by retaining only the relevant parts related to the motion and add a subject. For example, if the text prompt is \"Iron man kicks with his left leg\", we transform it into \"Someone kicks with the left leg\" when generating the motion sequences." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 562, + 268, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 562, + 268, + 575 + ], + "spans": [ + { + "bbox": [ + 47, + 562, + 268, + 575 + ], + "type": "text", + "content": "3.4. Optimization with the 2D Diffusion Model" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": "To enrich details and improve the quality of the 3D asset, we optimize the 3D Gaussians " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\theta_{b}" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": " with a 2D diffusion model " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "F_{2D}" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": " after initializing them with 3D diffusion model priors. We employ the SDS (Score Distillation Sampling) loss to optimize the 3D Gaussians. First, we use the method of 3D Gaussian Splatting [25] to obtain the rendered image " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x} = g(\\theta_{i})" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": " represents the splatting rendering method as in Eq. 3. Then, we use Eq. 1 to calculate the gradients for updating the Gaussian parameters " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\theta_{i}" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": " with the 2D diffusion model " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "F_{2D}" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": ". After a short optimization period using the 2D diffusion model " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "F_{2D}" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": ", the final generated 3D" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 307, + 91, + 556, + 171 + ], + "blocks": [ + { + "bbox": [ + 330, + 70, + 522, + 82 + ], + "lines": [ + { + "bbox": [ + 330, + 70, + 522, + 82 + ], + "spans": [ + { + "bbox": [ + 330, + 70, + 522, + 82 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons on " + }, + { + "bbox": [ + 330, + 70, + 522, + 82 + ], + "type": "inline_equation", + "content": "{\\mathrm{T}}^{3}" + }, + { + "bbox": [ + 330, + 70, + 522, + 82 + ], + "type": "text", + "content": " Bench [17]." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 91, + 556, + 171 + ], + "lines": [ + { + "bbox": [ + 307, + 91, + 556, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 91, + 556, + 171 + ], + "type": "table", + "html": "
MethodTime†Single Obj.Single w/ Surr.Multi Obj.Average
SJC [76]-24.719.811.718.7
DreamFusion [52]6 hours24.424.616.121.7
Fantasia3D [6]6 hours26.427.018.524.0
LatentNeRF [43]15 minutes33.130.620.628.1
Magic3D [32]5.3 hours37.035.425.732.7
ProlificDreamer [77]~10 hours49.444.835.843.3
Ours15 minutes54.048.634.545.7
", + "image_path": "c3b93ff047c7d6bb6208570425dcd25786d3ff784530b45cbd46db7da22cd83c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 321, + 171, + 531, + 181 + ], + "lines": [ + { + "bbox": [ + 321, + 171, + 531, + 181 + ], + "spans": [ + { + "bbox": [ + 321, + 171, + 531, + 181 + ], + "type": "text", + "content": "Scores above are average of two metrics (quality and alignment)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 322, + 181, + 436, + 191 + ], + "lines": [ + { + "bbox": [ + 322, + 181, + 436, + 191 + ], + "spans": [ + { + "bbox": [ + 322, + 181, + 436, + 191 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 322, + 181, + 436, + 191 + ], + "type": "text", + "content": " GPU time counted in their papers." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 201, + 545, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 201, + 545, + 225 + ], + "spans": [ + { + "bbox": [ + 305, + 201, + 545, + 225 + ], + "type": "text", + "content": "instance " + }, + { + "bbox": [ + 305, + 201, + 545, + 225 + ], + "type": "inline_equation", + "content": "\\theta_{f}" + }, + { + "bbox": [ + 305, + 201, + 545, + 225 + ], + "type": "text", + "content": " achieves high quality and fidelity on top of the 3D consistency provided by the 3D diffusion model " + }, + { + "bbox": [ + 305, + 201, + 545, + 225 + ], + "type": "inline_equation", + "content": "F_{3D}" + }, + { + "bbox": [ + 305, + 201, + 545, + 225 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 236, + 387, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 236, + 387, + 249 + ], + "spans": [ + { + "bbox": [ + 306, + 236, + 387, + 249 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 255, + 545, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 545, + 338 + ], + "type": "text", + "content": "In this section, we first present the implementation details in Sec. 4.1. In Sec. 4.2, we show the quantitative comparison results. Then, in Sec. 4.3, we showcase the visualization results of our method and compare them with other methods. In Sec. 4.4, we conduct a series of ablation experiments to validate the effectiveness of our method. Finally, we discuss the limitations of our method." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 346, + 440, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 346, + 440, + 358 + ], + "spans": [ + { + "bbox": [ + 306, + 346, + 440, + 358 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 365, + 546, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 546, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 546, + 651 + ], + "type": "text", + "content": "Our method is implemented in PyTorch [51], based on ThreeStudio [14]. The 3D diffusion models used in our method are Shap-E [24] and MDM [72], and we load the Shap-E model finetuned on Objaverse [11] in Cap3D [42]. For the 2D diffusion model, we use stabilityai/stable-diffusion-2-1-base [58], with a guidance scale of 100. The timestamps we use are uniformly sampled from 0.02 to 0.98 before 500 iterations, and change to 0.02 to 0.55 after 500 iterations. For the 3D Gaussians, the learning rates of opacity " + }, + { + "bbox": [ + 304, + 365, + 546, + 651 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 365, + 546, + 651 + ], + "type": "text", + "content": " and position " + }, + { + "bbox": [ + 304, + 365, + 546, + 651 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 365, + 546, + 651 + ], + "type": "text", + "content": " are " + }, + { + "bbox": [ + 304, + 365, + 546, + 651 + ], + "type": "inline_equation", + "content": "10^{-2}" + }, + { + "bbox": [ + 304, + 365, + 546, + 651 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 365, + 546, + 651 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-5}" + }, + { + "bbox": [ + 304, + 365, + 546, + 651 + ], + "type": "text", + "content": ". The color " + }, + { + "bbox": [ + 304, + 365, + 546, + 651 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 365, + 546, + 651 + ], + "type": "text", + "content": " of the 3d Gaussians is represented by the sh coefficient, with the degree set to 0 and the learning rate set to " + }, + { + "bbox": [ + 304, + 365, + 546, + 651 + ], + "type": "inline_equation", + "content": "1.25 \\times 10^{-2}" + }, + { + "bbox": [ + 304, + 365, + 546, + 651 + ], + "type": "text", + "content": ". The covariance of the 3D Gaussians is converted into scaling and rotation for optimization, with learning rates of " + }, + { + "bbox": [ + 304, + 365, + 546, + 651 + ], + "type": "inline_equation", + "content": "10^{-3}" + }, + { + "bbox": [ + 304, + 365, + 546, + 651 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 365, + 546, + 651 + ], + "type": "inline_equation", + "content": "10^{-2}" + }, + { + "bbox": [ + 304, + 365, + 546, + 651 + ], + "type": "text", + "content": ", respectively. The radius of the camera we use for rendering is from 1.5 to 4.0, with the azimuth in the range of -180 to 180 degrees and the elevation in the range of -10 to 60 degrees. The total training iterations are 1200. All our experiments can be completed within 15 minutes on a single RTX 3090 with a batch size of 4. The resolution we use for rendering is " + }, + { + "bbox": [ + 304, + 365, + 546, + 651 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 304, + 365, + 546, + 651 + ], + "type": "text", + "content": ", which is scaled to 512 " + }, + { + "bbox": [ + 304, + 365, + 546, + 651 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 304, + 365, + 546, + 651 + ], + "type": "text", + "content": " 512 when optimizing using the 2D diffusion model. We can render in real time at " + }, + { + "bbox": [ + 304, + 365, + 546, + 651 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 365, + 546, + 651 + ], + "type": "text", + "content": " resolution. And all our code will be released." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 659, + 441, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 659, + 441, + 672 + ], + "spans": [ + { + "bbox": [ + 306, + 659, + 441, + 672 + ], + "type": "text", + "content": "4.2. Quantitative Evaluation" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "We evaluate quality and consistency following T3Bench [17], which provides a comprehensive benchmark for text-to-3D generation. Three text categories are" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6800" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 94, + 141, + 135 + ], + "blocks": [ + { + "bbox": [ + 77, + 70, + 124, + 87 + ], + "lines": [ + { + "bbox": [ + 77, + 70, + 124, + 87 + ], + "spans": [ + { + "bbox": [ + 77, + 70, + 124, + 87 + ], + "type": "text", + "content": "DreamFusion 6.0 Hours" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 62, + 94, + 141, + 135 + ], + "lines": [ + { + "bbox": [ + 62, + 94, + 141, + 135 + ], + "spans": [ + { + "bbox": [ + 62, + 94, + 141, + 135 + ], + "type": "image", + "image_path": "c4b46be0ed11935322d0c31b2d3bb349e22df589648bbded554dea0a93925251.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 159, + 100, + 241, + 135 + ], + "blocks": [ + { + "bbox": [ + 159, + 100, + 241, + 135 + ], + "lines": [ + { + "bbox": [ + 159, + 100, + 241, + 135 + ], + "spans": [ + { + "bbox": [ + 159, + 100, + 241, + 135 + ], + "type": "image", + "image_path": "79922dd0a52a901b81385541823270f804f0894692275bd8d0b8e4a1ab9bf33c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 250, + 95, + 336, + 135 + ], + "blocks": [ + { + "bbox": [ + 186, + 70, + 217, + 87 + ], + "lines": [ + { + "bbox": [ + 186, + 70, + 217, + 87 + ], + "spans": [ + { + "bbox": [ + 186, + 70, + 217, + 87 + ], + "type": "text", + "content": "Magic3D 5.3 Hours" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 276, + 70, + 317, + 87 + ], + "lines": [ + { + "bbox": [ + 276, + 70, + 317, + 87 + ], + "spans": [ + { + "bbox": [ + 276, + 70, + 317, + 87 + ], + "type": "text", + "content": "Fantasia3D 6.0 Hours" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 250, + 95, + 336, + 135 + ], + "lines": [ + { + "bbox": [ + 250, + 95, + 336, + 135 + ], + "spans": [ + { + "bbox": [ + 250, + 95, + 336, + 135 + ], + "type": "image", + "image_path": "45155875c3469a406d73196d631c4d425f392d5197d3d1336d4eab0671160888.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 352, + 91, + 442, + 137 + ], + "blocks": [ + { + "bbox": [ + 372, + 70, + 429, + 87 + ], + "lines": [ + { + "bbox": [ + 372, + 70, + 429, + 87 + ], + "spans": [ + { + "bbox": [ + 372, + 70, + 429, + 87 + ], + "type": "text", + "content": "ProlificDreamer ~Hours" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 447, + 71, + 533, + 87 + ], + "lines": [ + { + "bbox": [ + 447, + 71, + 533, + 87 + ], + "spans": [ + { + "bbox": [ + 447, + 71, + 533, + 87 + ], + "type": "text", + "content": "GaussianDreamer (Ours) 15 Minutes" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 352, + 91, + 442, + 137 + ], + "lines": [ + { + "bbox": [ + 352, + 91, + 442, + 137 + ], + "spans": [ + { + "bbox": [ + 352, + 91, + 442, + 137 + ], + "type": "image", + "image_path": "4fc682d9acda770c6b6876118a58e26afdfb7968429e468e00b1398a9d180c0d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 447, + 99, + 493, + 133 + ], + "blocks": [ + { + "bbox": [ + 447, + 99, + 493, + 133 + ], + "lines": [ + { + "bbox": [ + 447, + 99, + 493, + 133 + ], + "spans": [ + { + "bbox": [ + 447, + 99, + 493, + 133 + ], + "type": "image", + "image_path": "20649408e8da75ccdd629cbb303ee9b982d1d0ebbd7b4a807d972d194c73d22a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 496, + 99, + 542, + 133 + ], + "blocks": [ + { + "bbox": [ + 496, + 99, + 542, + 133 + ], + "lines": [ + { + "bbox": [ + 496, + 99, + 542, + 133 + ], + "spans": [ + { + "bbox": [ + 496, + 99, + 542, + 133 + ], + "type": "image", + "image_path": "3df616b454f46e13d39cb16915bb76b907e4b74a245f4c6b5b805df783fe0d7a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 50, + 154, + 97, + 194 + ], + "blocks": [ + { + "bbox": [ + 50, + 154, + 97, + 194 + ], + "lines": [ + { + "bbox": [ + 50, + 154, + 97, + 194 + ], + "spans": [ + { + "bbox": [ + 50, + 154, + 97, + 194 + ], + "type": "image", + "image_path": "7290185a6994106d0a5a8516f71e33486740dd15a357030589b6eae517aff470.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 209, + 546, + 243 + ], + "lines": [ + { + "bbox": [ + 46, + 209, + 546, + 243 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 546, + 243 + ], + "type": "text", + "content": "Figure 4. Qualitative comparisons between our method and DreamFusion [52], Magic3D [32], Fantasy3D [6] and ProlificDreamer [77]. Here we count the GPU time in their papers. The time for DreamFusion is measured on TPUv4, Magic3D is measured on A100, Fantasy3D is measured on RTX 3090, and our method is measured on RTX 3090." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 99, + 155, + 141, + 194 + ], + "blocks": [ + { + "bbox": [ + 99, + 155, + 141, + 194 + ], + "lines": [ + { + "bbox": [ + 99, + 155, + 141, + 194 + ], + "spans": [ + { + "bbox": [ + 99, + 155, + 141, + 194 + ], + "type": "image", + "image_path": "56c02fa1558038816f48335b17cb2aba7f56ed36ef33a19395d676333b3d92e9.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 154, + 148, + 195, + 193 + ], + "blocks": [ + { + "bbox": [ + 154, + 148, + 195, + 193 + ], + "lines": [ + { + "bbox": [ + 154, + 148, + 195, + 193 + ], + "spans": [ + { + "bbox": [ + 154, + 148, + 195, + 193 + ], + "type": "image", + "image_path": "753289c373588837d7220f4a0d5e9e8da89bc317da0dbaa0bd95d1c961b26d9e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 203, + 149, + 242, + 194 + ], + "blocks": [ + { + "bbox": [ + 203, + 149, + 242, + 194 + ], + "lines": [ + { + "bbox": [ + 203, + 149, + 242, + 194 + ], + "spans": [ + { + "bbox": [ + 203, + 149, + 242, + 194 + ], + "type": "image", + "image_path": "b4448fb8506ebdf152b2561d19cafa572bd2e5bc7e483177892fb3e199e53101.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 252, + 148, + 294, + 192 + ], + "blocks": [ + { + "bbox": [ + 229, + 137, + 386, + 148 + ], + "lines": [ + { + "bbox": [ + 229, + 137, + 386, + 148 + ], + "spans": [ + { + "bbox": [ + 229, + 137, + 386, + 148 + ], + "type": "text", + "content": "A plate piled high with chocolate chip cookies." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 252, + 148, + 294, + 192 + ], + "lines": [ + { + "bbox": [ + 252, + 148, + 294, + 192 + ], + "spans": [ + { + "bbox": [ + 252, + 148, + 294, + 192 + ], + "type": "image", + "image_path": "169a62acecb4392a588a8975c36490e2b888819033ac8c73554bda5b07700fc9.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 302, + 148, + 347, + 192 + ], + "blocks": [ + { + "bbox": [ + 302, + 148, + 347, + 192 + ], + "lines": [ + { + "bbox": [ + 302, + 148, + 347, + 192 + ], + "spans": [ + { + "bbox": [ + 302, + 148, + 347, + 192 + ], + "type": "image", + "image_path": "f2692df22d77547f9ebc2ebdd6f765eea1ac84c701030bbdd06d0877f3b20697.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 351, + 150, + 440, + 196 + ], + "blocks": [ + { + "bbox": [ + 351, + 150, + 440, + 196 + ], + "lines": [ + { + "bbox": [ + 351, + 150, + 440, + 196 + ], + "spans": [ + { + "bbox": [ + 351, + 150, + 440, + 196 + ], + "type": "image", + "image_path": "ae5e65b857a6db3701ba16e0eeabc65a30c56496df91840937db887fe2be934f.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 211, + 198, + 403, + 207 + ], + "lines": [ + { + "bbox": [ + 211, + 198, + 403, + 207 + ], + "spans": [ + { + "bbox": [ + 211, + 198, + 403, + 207 + ], + "type": "text", + "content": "A 3D model of an adorable cottage with a thatched roof." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 444, + 151, + 493, + 194 + ], + "blocks": [ + { + "bbox": [ + 444, + 151, + 493, + 194 + ], + "lines": [ + { + "bbox": [ + 444, + 151, + 493, + 194 + ], + "spans": [ + { + "bbox": [ + 444, + 151, + 493, + 194 + ], + "type": "image", + "image_path": "5579c54f6d7b9f8b045568f74bf15d7f0203d1344e19ba30a5ffd7c66e1a609d.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 497, + 153, + 538, + 194 + ], + "blocks": [ + { + "bbox": [ + 497, + 153, + 538, + 194 + ], + "lines": [ + { + "bbox": [ + 497, + 153, + 538, + 194 + ], + "spans": [ + { + "bbox": [ + 497, + 153, + 538, + 194 + ], + "type": "image", + "image_path": "9247cd7677cc732004fdd032f1084a7da218c12b9ca42b6d12fa72d53a225881.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 52, + 250, + 148, + 306 + ], + "blocks": [ + { + "bbox": [ + 52, + 250, + 148, + 306 + ], + "lines": [ + { + "bbox": [ + 52, + 250, + 148, + 306 + ], + "spans": [ + { + "bbox": [ + 52, + 250, + 148, + 306 + ], + "type": "image", + "image_path": "ada454a8eb06f50484c9c7782bb9e057256733ac589ebd69888de2aa944d7a7b.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 107, + 308, + 129, + 319 + ], + "lines": [ + { + "bbox": [ + 107, + 308, + 129, + 319 + ], + "spans": [ + { + "bbox": [ + 107, + 308, + 129, + 319 + ], + "type": "text", + "content": "A fox" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 150, + 251, + 180, + 307 + ], + "blocks": [ + { + "bbox": [ + 150, + 251, + 180, + 307 + ], + "lines": [ + { + "bbox": [ + 150, + 251, + 180, + 307 + ], + "spans": [ + { + "bbox": [ + 150, + 251, + 180, + 307 + ], + "type": "image", + "image_path": "feeec47e65c5f8eff091d6cfcdcfa442960f01a2934b13c9f5fe3bf016d2784b.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 50, + 331, + 127, + 384 + ], + "blocks": [ + { + "bbox": [ + 50, + 331, + 127, + 384 + ], + "lines": [ + { + "bbox": [ + 50, + 331, + 127, + 384 + ], + "spans": [ + { + "bbox": [ + 50, + 331, + 127, + 384 + ], + "type": "image", + "image_path": "8c14e8aec249158efe551260eb11e4ec9dc1bfe39ecfcb526272efa55b10c608.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 392, + 171, + 402 + ], + "lines": [ + { + "bbox": [ + 63, + 392, + 171, + 402 + ], + "spans": [ + { + "bbox": [ + 63, + 392, + 171, + 402 + ], + "type": "text", + "content": "An amigurumi motorcycle..." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 136, + 335, + 197, + 384 + ], + "blocks": [ + { + "bbox": [ + 136, + 335, + 197, + 384 + ], + "lines": [ + { + "bbox": [ + 136, + 335, + 197, + 384 + ], + "spans": [ + { + "bbox": [ + 136, + 335, + 197, + 384 + ], + "type": "image", + "image_path": "404b6a7aada9388f28bdba8b8b99e37db7883c3ccda42b411c647b6a89c179d6.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 211, + 252, + 273, + 307 + ], + "blocks": [ + { + "bbox": [ + 211, + 252, + 273, + 307 + ], + "lines": [ + { + "bbox": [ + 211, + 252, + 273, + 307 + ], + "spans": [ + { + "bbox": [ + 211, + 252, + 273, + 307 + ], + "type": "image", + "image_path": "2710ada4be39e44be8d6923f5a8dd5eda21f943faea69e74d4711f08e3ebdffe.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 276, + 255, + 358, + 307 + ], + "blocks": [ + { + "bbox": [ + 276, + 255, + 358, + 307 + ], + "lines": [ + { + "bbox": [ + 276, + 255, + 358, + 307 + ], + "spans": [ + { + "bbox": [ + 276, + 255, + 358, + 307 + ], + "type": "image", + "image_path": "ea366c3716a1473a6be7cb841be0be38d34eae3309221c1e4a85f9f26d991bee.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 356, + 308, + 536, + 319 + ], + "lines": [ + { + "bbox": [ + 356, + 308, + 536, + 319 + ], + "spans": [ + { + "bbox": [ + 356, + 308, + 536, + 319 + ], + "type": "text", + "content": "An opulent couch from the palace of Versailles" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 371, + 263, + 447, + 305 + ], + "blocks": [ + { + "bbox": [ + 371, + 263, + 447, + 305 + ], + "lines": [ + { + "bbox": [ + 371, + 263, + 447, + 305 + ], + "spans": [ + { + "bbox": [ + 371, + 263, + 447, + 305 + ], + "type": "image", + "image_path": "b2d0ead48d1c61ef323021840b5d61c2a6496cef6e99c7e080e020fe26ba17ff.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 459, + 258, + 510, + 307 + ], + "blocks": [ + { + "bbox": [ + 459, + 258, + 510, + 307 + ], + "lines": [ + { + "bbox": [ + 459, + 258, + 510, + 307 + ], + "spans": [ + { + "bbox": [ + 459, + 258, + 510, + 307 + ], + "type": "image", + "image_path": "59806878a7759f0f8438656b24f462d6ac9c206c2bb48edb0efc8753c970ae7a.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 244, + 318, + 312, + 393 + ], + "blocks": [ + { + "bbox": [ + 215, + 308, + 343, + 318 + ], + "lines": [ + { + "bbox": [ + 215, + 308, + 343, + 318 + ], + "spans": [ + { + "bbox": [ + 215, + 308, + 343, + 318 + ], + "type": "text", + "content": "Magic gun, game asset, mistery..." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 244, + 318, + 312, + 393 + ], + "lines": [ + { + "bbox": [ + 244, + 318, + 312, + 393 + ], + "spans": [ + { + "bbox": [ + 244, + 318, + 312, + 393 + ], + "type": "image", + "image_path": "7df71c74c5a0c7f6f31995bfb3097faefbdd82a80d09358763f36f8094a5ccd3.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 207, + 392, + 333, + 402 + ], + "lines": [ + { + "bbox": [ + 207, + 392, + 333, + 402 + ], + "spans": [ + { + "bbox": [ + 207, + 392, + 333, + 402 + ], + "type": "text", + "content": "Magic dagger, mistery, ancient..." + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 116, + 406, + 476, + 417 + ], + "lines": [ + { + "bbox": [ + 116, + 406, + 476, + 417 + ], + "spans": [ + { + "bbox": [ + 116, + 406, + 476, + 417 + ], + "type": "text", + "content": "Figure 5. More generated samples by our GaussianDreamer. Two views of each sample are shown." + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 375, + 327, + 432, + 389 + ], + "blocks": [ + { + "bbox": [ + 375, + 327, + 432, + 389 + ], + "lines": [ + { + "bbox": [ + 375, + 327, + 432, + 389 + ], + "spans": [ + { + "bbox": [ + 375, + 327, + 432, + 389 + ], + "type": "image", + "image_path": "7e657b51919a4558f204dec6904f0da93dd2720ac6c261de0835cb160652020f.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 346, + 392, + 545, + 403 + ], + "lines": [ + { + "bbox": [ + 346, + 392, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 346, + 392, + 545, + 403 + ], + "type": "text", + "content": "A zoomed out DSLR photo of a lion's mane jellyfish" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 440, + 328, + 503, + 389 + ], + "blocks": [ + { + "bbox": [ + 440, + 328, + 503, + 389 + ], + "lines": [ + { + "bbox": [ + 440, + 328, + 503, + 389 + ], + "spans": [ + { + "bbox": [ + 440, + 328, + 503, + 389 + ], + "type": "image", + "image_path": "f890a4999f962286b63d61c7709edc53b66c63eb54a13a65ca0a94445bff3c97.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "bbox": [ + 46, + 422, + 287, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 287, + 471 + ], + "type": "text", + "content": "designed for 3D generation with increasing complexity – single objects, single objects with surroundings, and multi objects. In Tab. 1, our method outperforms the compared methods while enjoying a short generation time." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 47, + 480, + 168, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 480, + 168, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 480, + 168, + 491 + ], + "type": "text", + "content": "4.3. Visualization Results" + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 46, + 499, + 287, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 499, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 46, + 499, + 287, + 535 + ], + "type": "text", + "content": "In this section, we present the results of initializing the 3D Gaussians with two different 3D diffusion models: text-to-3D and text-to-motion diffusion models." + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": "Initialization with Text-to-3D Diffusion Model. We show the comparison results with DreamFusion [52], Magic3D [32], Fantasy3D [6] and ProlificDreamer [77] in Fig. 4. In addition to our method, the figures of other methods are downloaded from the paper of ProlificDreamer. When encountering prompts that involve the combination of multiple objects, such as the prompt \"A plate piled high with chocolate chip cookies\", the generated results from Magic3D, Fantasy3D, and ProlificDreamer do not include a plate. In contrast, our generated result can effectively combine a plate and chocolate chip cookies. Furthermore, compared to DreamFusion, the plate we generate has better patterns. Our method shows comparable quality while saving 21 - 24 times the generation time compared to their" + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 304, + 422, + 547, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 547, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 547, + 495 + ], + "type": "text", + "content": "methods. Moreover, the 3D Gaussians generated by our method can directly achieve real-time rendering without further transformation into mesh-like structures. Fig. 5 visualizes more samples generated by our GaussianDreamer from various prompts, which show good 3D consistency while having high-quality details." + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "type": "text", + "content": "Initialization with Text-to-Motion Diffusion Model. In Fig. 6, we present the comparison results with DreamFusion [52], DreamAvatar [5], DreamWaltz [20], and AvatarVerse [86]. In addition to our method, the figures of other methods are downloaded from the paper of AvatarVerse. It is worth noting that our prompt is \"Spiderman/Stormtrooper stands with open arms\", while the prompts for other methods are \"Spiderman/Stormtrooper\". This is because when generating motion using text-to-motion diffusion model as initialization, we require more specific action descriptions. Our method achieves a speedup of 4-24 times compared to other methods, while maintaining comparable quality. Additionally, our method allows for generating 3D avatars with specified body poses. In Fig. 7, we provide more results generated with different human body poses. We first generate a sequence of motions that match the text prompt using a text-to-motion 3d diffusion model, and then initialize the" + } + ] + } + ], + "index": 46 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "6801" + } + ] + } + ], + "index": 47 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 69, + 154, + 183 + ], + "blocks": [ + { + "bbox": [ + 48, + 69, + 154, + 183 + ], + "lines": [ + { + "bbox": [ + 48, + 69, + 154, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 69, + 154, + 183 + ], + "type": "image", + "image_path": "0cfc1257a36d9ad19d9bf38570bddbbf78df0b6ce4206c80e7c433b10c1cd8e0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 156, + 70, + 277, + 183 + ], + "blocks": [ + { + "bbox": [ + 156, + 70, + 277, + 183 + ], + "lines": [ + { + "bbox": [ + 156, + 70, + 277, + 183 + ], + "spans": [ + { + "bbox": [ + 156, + 70, + 277, + 183 + ], + "type": "image", + "image_path": "d4b09dcf5d6995715bc51a8ea393ea2c634c32f67eaf132566c4f2b25e040929.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 282, + 70, + 388, + 183 + ], + "blocks": [ + { + "bbox": [ + 282, + 70, + 388, + 183 + ], + "lines": [ + { + "bbox": [ + 282, + 70, + 388, + 183 + ], + "spans": [ + { + "bbox": [ + 282, + 70, + 388, + 183 + ], + "type": "image", + "image_path": "f8832450bbe09a68ea5aa1fed5a5c62c2bc15309ae6bff4a546bf4be874239db.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 391, + 70, + 463, + 176 + ], + "blocks": [ + { + "bbox": [ + 391, + 70, + 463, + 176 + ], + "lines": [ + { + "bbox": [ + 391, + 70, + 463, + 176 + ], + "spans": [ + { + "bbox": [ + 391, + 70, + 463, + 176 + ], + "type": "image", + "image_path": "3399a3158964f6a65d26795c956329a332dd2a19cf2efcf69223255a2e436d7a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 466, + 70, + 545, + 87 + ], + "lines": [ + { + "bbox": [ + 466, + 70, + 545, + 87 + ], + "spans": [ + { + "bbox": [ + 466, + 70, + 545, + 87 + ], + "type": "text", + "content": "GaussianDreamer (Ours) 15 Minutes" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 49, + 191, + 150, + 290 + ], + "blocks": [ + { + "bbox": [ + 49, + 191, + 150, + 290 + ], + "lines": [ + { + "bbox": [ + 49, + 191, + 150, + 290 + ], + "spans": [ + { + "bbox": [ + 49, + 191, + 150, + 290 + ], + "type": "image", + "image_path": "84996c3396cf61b95e47c8f381ef314d2432e42526d5dbaddba30ba6f1041219.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 159, + 193, + 274, + 291 + ], + "blocks": [ + { + "bbox": [ + 159, + 193, + 274, + 291 + ], + "lines": [ + { + "bbox": [ + 159, + 193, + 274, + 291 + ], + "spans": [ + { + "bbox": [ + 159, + 193, + 274, + 291 + ], + "type": "image", + "image_path": "9f086da7d7c61a81113c6d0f895fe1f9581a4c06fc5fd0a81367a8823db5fb89.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 273, + 293, + 318, + 303 + ], + "lines": [ + { + "bbox": [ + 273, + 293, + 318, + 303 + ], + "spans": [ + { + "bbox": [ + 273, + 293, + 318, + 303 + ], + "type": "text", + "content": "Stormtrooper" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 282, + 193, + 386, + 291 + ], + "blocks": [ + { + "bbox": [ + 276, + 184, + 313, + 193 + ], + "lines": [ + { + "bbox": [ + 276, + 184, + 313, + 193 + ], + "spans": [ + { + "bbox": [ + 276, + 184, + 313, + 193 + ], + "type": "text", + "content": "Spiderman" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 282, + 193, + 386, + 291 + ], + "lines": [ + { + "bbox": [ + 282, + 193, + 386, + 291 + ], + "spans": [ + { + "bbox": [ + 282, + 193, + 386, + 291 + ], + "type": "image", + "image_path": "44cef7937051be096397c49f171c7f367570f6b49fe250aa88146fd258700661.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 391, + 202, + 462, + 286 + ], + "blocks": [ + { + "bbox": [ + 391, + 202, + 462, + 286 + ], + "lines": [ + { + "bbox": [ + 391, + 202, + 462, + 286 + ], + "spans": [ + { + "bbox": [ + 391, + 202, + 462, + 286 + ], + "type": "image", + "image_path": "fdb541959261afc19ef371e5c0033a236a3baca59a6e849bc55d847b8a1bb003.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 474, + 206, + 533, + 286 + ], + "blocks": [ + { + "bbox": [ + 474, + 206, + 533, + 286 + ], + "lines": [ + { + "bbox": [ + 474, + 206, + 533, + 286 + ], + "spans": [ + { + "bbox": [ + 474, + 206, + 533, + 286 + ], + "type": "image", + "image_path": "f38748c4101231bfb0f09515c973d036bd80949127338995681d58505b84b99f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 47, + 317, + 288, + 532 + ], + "blocks": [ + { + "bbox": [ + 47, + 304, + 545, + 315 + ], + "lines": [ + { + "bbox": [ + 47, + 304, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 47, + 304, + 545, + 315 + ], + "type": "text", + "content": "Figure 6. Qualitative comparisons between our method and DreamFusion [52], DreamAvatar [5], DreamWaltz [20], and AvatarVerse [86]." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 317, + 288, + 532 + ], + "lines": [ + { + "bbox": [ + 47, + 317, + 288, + 532 + ], + "spans": [ + { + "bbox": [ + 47, + 317, + 288, + 532 + ], + "type": "image", + "image_path": "9b44a10d6ece8b727f465d7903ef5f0dffa3c1ae84cc4b9b98dc405816f1a392.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 533, + 287, + 576 + ], + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 576 + ], + "type": "text", + "content": "Figure 7. More generated 3D avatars by our GaussianDreamer initialized with the different poses of SMPL [39]. Here, the different poses of SMPL are generated using a text-to-motion diffusion model." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 583, + 287, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 583, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 46, + 583, + 287, + 620 + ], + "type": "text", + "content": "3d Gaussians with the SMPL of one selected pose in motions. Our method can generate 3D avatars in any desired pose." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 623, + 204, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 204, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 204, + 635 + ], + "type": "text", + "content": "4.4. Ablation Study and Analysis" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": "The Role of Initialization. As shown in Fig. 8, we first conduct an ablation experiment on the initialization of the 3D Gaussians to validate that initialization can improve 3D consistency. The first column is the rendering result of Shap-E [24] with NeRF [45] as the 3D representation. The second column is the result of optimizing the 3D Gaussians" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 323, + 547, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 323, + 547, + 515 + ], + "spans": [ + { + "bbox": [ + 304, + 323, + 547, + 515 + ], + "type": "text", + "content": "randomly initialized within a cube using the SDS loss, and the third column is our method. We show the initialization effects on 3 samples. In the first and second rows, ShapE has good generation results while our method provides more complex geometries and more realistic appearances. Compared with random initialization, in the first row, details of our method are better. In the second row, the 3D assets generated by random initialization have the multi-head problem, which does not occur in our method. The initialization of the 3D diffusion model can avoid unreasonable geometry. In the third row, the generation result of Shap-E is far different from the given text prompt while our method makes the 3D assets closer to the prompt through the 2D diffusion model. Our method can expand the domain of Shap-E prompts, allowing for the generation of 3D assets based on a wider range of prompts." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 524, + 547, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 524, + 547, + 632 + ], + "spans": [ + { + "bbox": [ + 304, + 524, + 547, + 632 + ], + "type": "text", + "content": "Noisy Point Growing and Color Perturbation. Fig. 9 illustrates the ablation results of noisy point growing and color perturbation. With noisy point growing and color perturbation applied, the first row showcases improved details in the sniper rifle. Additionally, the second column generates an amigurumi motorcycle that better aligns with the style characteristics of amigurumi mentioned in the prompt, compared to the case without noisy point growing and color perturbation." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": "Initialization with Different Text-to-3D Diffusion Models. We select two text-to-3D generation models, ShapeE [24] and Point-E [48], to validate the effectiveness of our framework. We load the Point-E model finetuned on Objaverse [11] in Cap3D[42]. Fig. 10, we showcase the generated results after initializing the 3D Gaussians using" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6802" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 81, + 214, + 148 + ], + "blocks": [ + { + "bbox": [ + 118, + 69, + 147, + 79 + ], + "lines": [ + { + "bbox": [ + 118, + 69, + 147, + 79 + ], + "spans": [ + { + "bbox": [ + 118, + 69, + 147, + 79 + ], + "type": "text", + "content": "Shap-E" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 81, + 214, + 148 + ], + "lines": [ + { + "bbox": [ + 57, + 81, + 214, + 148 + ], + "spans": [ + { + "bbox": [ + 57, + 81, + 214, + 148 + ], + "type": "image", + "image_path": "1807572c3b4a4dcc7a809670101e10432bb4ac3afede8ee74eb3462a2422a0af.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 219, + 88, + 294, + 144 + ], + "blocks": [ + { + "bbox": [ + 255, + 70, + 358, + 80 + ], + "lines": [ + { + "bbox": [ + 255, + 70, + 358, + 80 + ], + "spans": [ + { + "bbox": [ + 255, + 70, + 358, + 80 + ], + "type": "text", + "content": "Ours w/o Point Cloud Prior" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 219, + 88, + 294, + 144 + ], + "lines": [ + { + "bbox": [ + 219, + 88, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 219, + 88, + 294, + 144 + ], + "type": "image", + "image_path": "5e25c51298fd20ea10da18d92f718411b9e43e6904456d22a7e0259068530b09.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 304, + 89, + 384, + 143 + ], + "blocks": [ + { + "bbox": [ + 304, + 89, + 384, + 143 + ], + "lines": [ + { + "bbox": [ + 304, + 89, + 384, + 143 + ], + "spans": [ + { + "bbox": [ + 304, + 89, + 384, + 143 + ], + "type": "image", + "image_path": "c80410bf9f9e49ef5cc21b67a47c063c53106cb9a3fc5b5d847bd910e504c75c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 391, + 87, + 462, + 146 + ], + "blocks": [ + { + "bbox": [ + 391, + 87, + 462, + 146 + ], + "lines": [ + { + "bbox": [ + 391, + 87, + 462, + 146 + ], + "spans": [ + { + "bbox": [ + 391, + 87, + 462, + 146 + ], + "type": "image", + "image_path": "7975de07c8bec90db2745111a489b16e732d4d9ab3c20ad1e2fe11e8f7e87573.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 468, + 86, + 544, + 147 + ], + "blocks": [ + { + "bbox": [ + 415, + 69, + 509, + 80 + ], + "lines": [ + { + "bbox": [ + 415, + 69, + 509, + 80 + ], + "spans": [ + { + "bbox": [ + 415, + 69, + 509, + 80 + ], + "type": "text", + "content": "GaussianDreamer (Ours)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 468, + 86, + 544, + 147 + ], + "lines": [ + { + "bbox": [ + 468, + 86, + 544, + 147 + ], + "spans": [ + { + "bbox": [ + 468, + 86, + 544, + 147 + ], + "type": "image", + "image_path": "d50db4694a3713138bec67aeb5b66782e706c844f78ce7e4016efa8d0b902549.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 57, + 163, + 216, + 229 + ], + "blocks": [ + { + "bbox": [ + 223, + 149, + 385, + 160 + ], + "lines": [ + { + "bbox": [ + 223, + 149, + 385, + 160 + ], + "spans": [ + { + "bbox": [ + 223, + 149, + 385, + 160 + ], + "type": "text", + "content": "A stack of pancakes covered in maple syrup" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 163, + 216, + 229 + ], + "lines": [ + { + "bbox": [ + 57, + 163, + 216, + 229 + ], + "spans": [ + { + "bbox": [ + 57, + 163, + 216, + 229 + ], + "type": "image", + "image_path": "e85175ebd52e7829a9b9d709beb8c5f2965ca828c112b7bf8a3844822a6cdab8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 235, + 164, + 302, + 228 + ], + "blocks": [ + { + "bbox": [ + 235, + 164, + 302, + 228 + ], + "lines": [ + { + "bbox": [ + 235, + 164, + 302, + 228 + ], + "spans": [ + { + "bbox": [ + 235, + 164, + 302, + 228 + ], + "type": "image", + "image_path": "1cd727f291e3666492d0fe66f68a9419dbbbc61990802c14b84110925e41eaf1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 309, + 168, + 380, + 229 + ], + "blocks": [ + { + "bbox": [ + 309, + 168, + 380, + 229 + ], + "lines": [ + { + "bbox": [ + 309, + 168, + 380, + 229 + ], + "spans": [ + { + "bbox": [ + 309, + 168, + 380, + 229 + ], + "type": "image", + "image_path": "6bfd90bfc09b8e436553ff957360f6965f5880169e40baed95a09c1d296c3991.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 392, + 171, + 462, + 239 + ], + "blocks": [ + { + "bbox": [ + 392, + 171, + 462, + 239 + ], + "lines": [ + { + "bbox": [ + 392, + 171, + 462, + 239 + ], + "spans": [ + { + "bbox": [ + 392, + 171, + 462, + 239 + ], + "type": "image", + "image_path": "e637eae71eeee2c858b591bfef01b6e77dbf0c01105092f25c7c58d8e73fe4f5.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 464, + 172, + 533, + 239 + ], + "blocks": [ + { + "bbox": [ + 464, + 172, + 533, + 239 + ], + "lines": [ + { + "bbox": [ + 464, + 172, + 533, + 239 + ], + "spans": [ + { + "bbox": [ + 464, + 172, + 533, + 239 + ], + "type": "image", + "image_path": "97ec9a04ccf38a249c6d9f60b6e1be68ca6f14350fd4007241457038e7df0598.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 48, + 245, + 228, + 300 + ], + "blocks": [ + { + "bbox": [ + 48, + 245, + 228, + 300 + ], + "lines": [ + { + "bbox": [ + 48, + 245, + 228, + 300 + ], + "spans": [ + { + "bbox": [ + 48, + 245, + 228, + 300 + ], + "type": "image", + "image_path": "95f05e8ebe08383af6510ec9ed620522020bd6904899b22d9c4cb26f6e7ccdfb.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 324, + 517, + 335 + ], + "lines": [ + { + "bbox": [ + 74, + 324, + 517, + 335 + ], + "spans": [ + { + "bbox": [ + 74, + 324, + 517, + 335 + ], + "type": "text", + "content": "Figure 8. Ablation studies of the initialization of the 3D Gaussians. The Shap-E [24] rendering resolution here is " + }, + { + "bbox": [ + 74, + 324, + 517, + 335 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 74, + 324, + 517, + 335 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 237, + 243, + 302, + 303 + ], + "blocks": [ + { + "bbox": [ + 237, + 243, + 302, + 303 + ], + "lines": [ + { + "bbox": [ + 237, + 243, + 302, + 303 + ], + "spans": [ + { + "bbox": [ + 237, + 243, + 302, + 303 + ], + "type": "image", + "image_path": "c82c6c8330d5b8c021d620571379e98034ff0b5518f26bffe2ad6b67dc03e14d.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 308, + 243, + 380, + 304 + ], + "blocks": [ + { + "bbox": [ + 308, + 243, + 380, + 304 + ], + "lines": [ + { + "bbox": [ + 308, + 243, + 380, + 304 + ], + "spans": [ + { + "bbox": [ + 308, + 243, + 380, + 304 + ], + "type": "image", + "image_path": "03947ce9b4662d0af060f95acc3bfa577f3e427e417b69f36b454c3e0f2a13a3.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 285, + 304, + 323, + 314 + ], + "lines": [ + { + "bbox": [ + 285, + 304, + 323, + 314 + ], + "spans": [ + { + "bbox": [ + 285, + 304, + 323, + 314 + ], + "type": "text", + "content": "A lionfish" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 381, + 247, + 459, + 302 + ], + "blocks": [ + { + "bbox": [ + 381, + 247, + 459, + 302 + ], + "lines": [ + { + "bbox": [ + 381, + 247, + 459, + 302 + ], + "spans": [ + { + "bbox": [ + 381, + 247, + 459, + 302 + ], + "type": "image", + "image_path": "379e57fae2221050da7d14800c0d41221bc64214b31cb4e2fbd7f5f61cc362bd.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 461, + 247, + 541, + 302 + ], + "blocks": [ + { + "bbox": [ + 461, + 247, + 541, + 302 + ], + "lines": [ + { + "bbox": [ + 461, + 247, + 541, + 302 + ], + "spans": [ + { + "bbox": [ + 461, + 247, + 541, + 302 + ], + "type": "image", + "image_path": "29084092880b0720a0a55b982efda37476f45335831fe66cd772233e870f3c25.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 50, + 350, + 287, + 371 + ], + "blocks": [ + { + "bbox": [ + 83, + 340, + 161, + 349 + ], + "lines": [ + { + "bbox": [ + 83, + 340, + 161, + 349 + ], + "spans": [ + { + "bbox": [ + 83, + 340, + 161, + 349 + ], + "type": "text", + "content": "Ours w/o Grow&Pertb." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 197, + 340, + 282, + 348 + ], + "lines": [ + { + "bbox": [ + 197, + 340, + 282, + 348 + ], + "spans": [ + { + "bbox": [ + 197, + 340, + 282, + 348 + ], + "type": "text", + "content": "GaussianDreamer (Ours)" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 350, + 287, + 371 + ], + "lines": [ + { + "bbox": [ + 50, + 350, + 287, + 371 + ], + "spans": [ + { + "bbox": [ + 50, + 350, + 287, + 371 + ], + "type": "image", + "image_path": "003ebaa33cdd9170969b8dacc3eacea37d58121fba4fa36839e61f1813ad19e1.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 120, + 371, + 229, + 378 + ], + "lines": [ + { + "bbox": [ + 120, + 371, + 229, + 378 + ], + "spans": [ + { + "bbox": [ + 120, + 371, + 229, + 378 + ], + "type": "text", + "content": "Sniper rifle, asset, scifi, cyberpunk..." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 52, + 380, + 279, + 415 + ], + "blocks": [ + { + "bbox": [ + 52, + 380, + 279, + 415 + ], + "lines": [ + { + "bbox": [ + 52, + 380, + 279, + 415 + ], + "spans": [ + { + "bbox": [ + 52, + 380, + 279, + 415 + ], + "type": "image", + "image_path": "9e5fe0ab69529996c392ef7eb29bfc293fd18c2cd212503dd2cba6ecd8d231d6.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 134, + 415, + 216, + 421 + ], + "lines": [ + { + "bbox": [ + 134, + 415, + 216, + 421 + ], + "spans": [ + { + "bbox": [ + 134, + 415, + 216, + 421 + ], + "type": "text", + "content": "An amigurumi motorcycle..." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 423, + 287, + 456 + ], + "lines": [ + { + "bbox": [ + 46, + 423, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 287, + 456 + ], + "type": "text", + "content": "Figure 9. Ablation studies of noisy point growing and color perturbation. \"Grow&Pertb.\" denotes noisy point growing and color perturbation." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 55, + 470, + 269, + 498 + ], + "blocks": [ + { + "bbox": [ + 75, + 461, + 151, + 469 + ], + "lines": [ + { + "bbox": [ + 75, + 461, + 151, + 469 + ], + "spans": [ + { + "bbox": [ + 75, + 461, + 151, + 469 + ], + "type": "text", + "content": "Initialization with Point-E" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 178, + 461, + 251, + 469 + ], + "lines": [ + { + "bbox": [ + 178, + 461, + 251, + 469 + ], + "spans": [ + { + "bbox": [ + 178, + 461, + 251, + 469 + ], + "type": "text", + "content": "Initialization with Shap-E" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 470, + 269, + 498 + ], + "lines": [ + { + "bbox": [ + 55, + 470, + 269, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 470, + 269, + 498 + ], + "type": "image", + "image_path": "33f58febb62d7925bb0334415925dabb5be1e55409b01b8859917efbd0dcf946.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 122, + 498, + 211, + 506 + ], + "lines": [ + { + "bbox": [ + 122, + 498, + 211, + 506 + ], + "spans": [ + { + "bbox": [ + 122, + 498, + 211, + 506 + ], + "type": "text", + "content": "An airplane made out of wood" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 50, + 506, + 286, + 536 + ], + "blocks": [ + { + "bbox": [ + 50, + 506, + 286, + 536 + ], + "lines": [ + { + "bbox": [ + 50, + 506, + 286, + 536 + ], + "spans": [ + { + "bbox": [ + 50, + 506, + 286, + 536 + ], + "type": "image", + "image_path": "75c33677a6349772832ee11dd09544e8901798823d9fe45a8a16f3a422e1b31e.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 107, + 537, + 228, + 546 + ], + "lines": [ + { + "bbox": [ + 107, + 537, + 228, + 546 + ], + "spans": [ + { + "bbox": [ + 107, + 537, + 228, + 546 + ], + "type": "text", + "content": "Flamethrower, with fire, scif, cyberpunk..." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 548, + 287, + 570 + ], + "lines": [ + { + "bbox": [ + 47, + 548, + 287, + 570 + ], + "spans": [ + { + "bbox": [ + 47, + 548, + 287, + 570 + ], + "type": "text", + "content": "Figure 10. Ablation studies of initialization with different text-to-3D diffusion models: Point-E [48] and Shap-E [24]." + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "bbox": [ + 46, + 578, + 287, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 578, + 287, + 663 + ], + "spans": [ + { + "bbox": [ + 46, + 578, + 287, + 663 + ], + "type": "text", + "content": "one of two text-to-3D generation models. It can be seen that both initializations yield good generation results. However, considering that Shap-E generates 3D assets based on NeRF and SDF, which provide higher fidelity compared to the point cloud representation used by Point-E, the geometry of the airplane in the first row of Fig. 10 appears better when initialized using Shap-E." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 47, + 670, + 124, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 124, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 124, + 681 + ], + "type": "text", + "content": "4.5. Limitations" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "The edges of the 3D assets generated by our method are not always sharp, and there may be unnecessary 3D Gaussians" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 304, + 342, + 547, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 547, + 474 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 547, + 474 + ], + "type": "text", + "content": "around the object surface. How to filter these point clouds will be a possible direction for improvement. Our approach utilizes 3D diffusion model priors, which greatly alleviates the issue of multi-face problems. However, there is still a small chance of encountering the multi-face problems in scenes where there is minimal geometric difference but significant appearance discrepancy between the front and back of objects such as a backpack. Utilizing 3D-aware diffusion models [35, 67, 90] may be able to solve this problem. Additionally, our method has limited effectiveness in generating large-scale scenes, such as indoor scenes." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 306, + 497, + 379, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 497, + 379, + 510 + ], + "spans": [ + { + "bbox": [ + 306, + 497, + 379, + 510 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 304, + 521, + 547, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 547, + 643 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 547, + 643 + ], + "type": "text", + "content": "We propose a fast text-to-3D method GaussianDreamer by bridging the abilities of 3D and 2D diffusion models via the Gaussian splatting representation. GaussianDreamer can generate detailed and realistic geometry and appearance while maintaining 3D consistency. The 3D diffusion model priors and geometry priors from the 3D Gaussians effectively promote the convergence speed. Each sample can be generated within 15 minutes on one GPU. We believe the approach of bridging 3D and 2D diffusion models could be a promising direction to generate 3D assets efficiently." + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 306, + 665, + 403, + 679 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 665, + 403, + 679 + ], + "spans": [ + { + "bbox": [ + 306, + 665, + 403, + 679 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "This work was supported by the National Natural Science Foundation of China (No. 62376102)." + } + ] + } + ], + "index": 42 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6803" + } + ] + } + ], + "index": 43 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Tenglong Ao, Zeyi Zhang, and Libin Liu. Gesture diffuclip: Gesture diffusion model with clip latents. arXiv preprint arXiv:2303.14613, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 288, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 288, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 288, + 179 + ], + "type": "text", + "content": "[2] Mohammadreza Armandpour, Huangjie Zheng, Ali Sadeghian, Amir Sadeghian, and Mingyuan Zhou. Re-imagine the negative prompt algorithm: Transform 2d diffusion into 3d, alleviate janus problem and beyond. arXiv preprint arXiv:2304.04968, 2023.1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 288, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 288, + 224 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 288, + 224 + ], + "type": "text", + "content": "[3] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, pages 5855-5864, 2021. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 225, + 288, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 225, + 288, + 257 + ], + "spans": [ + { + "bbox": [ + 53, + 225, + 288, + 257 + ], + "type": "text", + "content": "[4] Jon Louis Bentley. Multidimensional binary search trees used for associative searching. Communications of the ACM, 18(9):509-517, 1975. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 258, + 288, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 288, + 301 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 288, + 301 + ], + "type": "text", + "content": "[5] Yukang Cao, Yan-Pei Cao, Kai Han, Ying Shan, and KwanYee K Wong. Dreamavatar: Text-and-shape guided 3d human avatar generation via diffusion models. arXiv preprint arXiv:2304.00916, 2023. 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 303, + 288, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 303, + 288, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 303, + 288, + 346 + ], + "type": "text", + "content": "[6] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. *Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation.* arXiv preprint arXiv:2303.13873, 2023. 1, 2, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 347, + 288, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 347, + 288, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 347, + 288, + 390 + ], + "type": "text", + "content": "[7] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In CVPR, pages 18000-18010, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 392, + 288, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 392, + 288, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 392, + 288, + 434 + ], + "type": "text", + "content": "[8] Yiwen Chen, Chi Zhang, Xiaofeng Yang, Zhongang Cai, Gang Yu, Lei Yang, and Guosheng Lin. It3d: Improved text-to-3d generation with explicit view synthesis. arXiv preprint arXiv:2308.11473, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 436, + 288, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 436, + 288, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 436, + 288, + 468 + ], + "type": "text", + "content": "[9] Zilong Chen, Feng Wang, and Huaping Liu. Text-to-3d using gaussian splatting. arXiv preprint arXiv:2309.16585, 2023. 1, 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 470, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 470, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 470, + 287, + 514 + ], + "type": "text", + "content": "[10] Rishabh Dabral, Muhammad Hamza Mughal, Vladislav Golyanik, and Christian Theobalt. Mofusion: A framework for denoising-diffusion-based motion synthesis. In CVPR, pages 9760-9770, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 514, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 568 + ], + "type": "text", + "content": "[11] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In CVPR, pages 13142-13153, 2023. 5, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 570, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 624 + ], + "type": "text", + "content": "[12] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. NeurIPS, 35:31841-31854, 2022. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 625, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 678 + ], + "type": "text", + "content": "[13] Jiatao Gu, Alex Trevithick, Kai-En Lin, Joshua M Susskind, Christian Theobalt, Lingjie Liu, and Ravi Ramamoorthi. Nerfdiff: Single-image view synthesis with nerf-guided distillation from 3d-aware diffusion. In ICML, pages 11808–11826. PMLR, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "text", + "content": "[14] Yuan-Chen Guo, Ying-Tian Liu, Ruizhi Shao, Christian Laforte, Vikram Voleti, Guan Luo, Chia-Hao Chen, Zi-Xin Zou, Chen Wang, Yan-Pei Cao, and Song-Hai Zhang." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "text", + "content": "threestudio: A unified framework for 3d content generation. https://github.com/threestudio-project/threestudio, 2023.5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 107, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 140 + ], + "type": "text", + "content": "[15] Anchit Gupta, Wenhan Xiong, Yixin Nie, Ian Jones, and Barlas Oğuz. 3dgen: Triplane latent diffusion for textured mesh generation. arXiv preprint arXiv:2303.05371, 2023. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 141, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 545, + 184 + ], + "type": "text", + "content": "[16] Xiao Han, Yukang Cao, Kai Han, Xiatian Zhu, Jiankang Deng, Yi-Zhe Song, Tao Xiang, and Kwan-Yee K Wong. Headsculpt: Crafting 3d head avatars with text. In NeurIPS, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 186, + 545, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 545, + 230 + ], + "type": "text", + "content": "[17] Yuze He, Yushi Bai, Matthieu Lin, Wang Zhao, Yubin Hu, Jenny Sheng, Ran Yi, Juanzi Li, and Yong-Jin Liu. T 3 bench: Benchmarking current progress in text-to-3d generation. arXiv preprint arXiv:2310.02977, 2023. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 232, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 232, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 232, + 545, + 274 + ], + "type": "text", + "content": "[18] Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. arXiv preprint arXiv:2205.08535, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 276, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 545, + 319 + ], + "type": "text", + "content": "[19] Shoukang Hu, Fangzhou Hong, Tao Hu, Liang Pan, Haiyi Mei, Weiye Xiao, Lei Yang, and Ziwei Liu. Humanliff: Layer-wise 3d human generation with diffusion model. arXiv preprint arXiv:2308.09712, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 321, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 321, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 308, + 321, + 545, + 365 + ], + "type": "text", + "content": "[20] Yukun Huang, Jianan Wang, Ailing Zeng, He Cao, Xianbiao Qi, Yukai Shi, Zheng-Jun Zha, and Lei Zhang. Dreamwaltz: Make a scene with complex 3d animatable avatars. arXiv preprint arXiv:2305.12529, 2023. 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 366, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 366, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 308, + 366, + 545, + 409 + ], + "type": "text", + "content": "[21] Yangyi Huang, Hongwei Yi, Yuliang Xiu, Tingting Liao, Ji-axiang Tang, Deng Cai, and Justus Thies. Tech: Text-guided reconstruction of lifelike clothed humans. arXiv preprint arXiv:2308.08545, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 411, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 453 + ], + "type": "text", + "content": "[22] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In CVPR, pages 867-876, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 456, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 456, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 456, + 545, + 510 + ], + "type": "text", + "content": "[23] Ruixiang Jiang, Can Wang, Jingbo Zhang, Mengei Chai, Mingming He, Dongdong Chen, and Jing Liao. Avatarcraft: Transforming text into neural human avatars with parameterized shape and pose control. arXiv preprint arXiv:2303.17606, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 512, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 512, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 512, + 545, + 544 + ], + "type": "text", + "content": "[24] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 1, 2, 5, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 545, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 545, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 308, + 545, + 545, + 589 + ], + "type": "text", + "content": "[25] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 2, 3, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 590, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 545, + 624 + ], + "type": "text", + "content": "[26] Jihoon Kim, Jiseob Kim, and Sungjoon Choi. Flame: Freeform language-based motion synthesis & editing. In AAAI, pages 8255-8263, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 624, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 668 + ], + "type": "text", + "content": "[27] Georgios Kopanas, Julien Philip, Thomas Leimkuhler, and George Drettakis. Point-based neural rendering with perview optimization. In Computer Graphics Forum, pages 29-43. Wiley Online Library, 2021. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "text", + "content": "[28] Georgios Kopanas, Thomas Leimkuhler, Gilles Rainer, Clément Jambon, and George Drettakis. Neural point cata-caustics for novel-view synthesis of reflections. ACM Transactions on Graphics (TOG), 41(6):1-15, 2022. 3" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6804" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[29] Jiabao Lei, Yabin Zhang, Kui Jia, et al. Tango: Text-driven photorealistic and robust 3d stylization via lighting decomposition. NeurIPS, 35:30923-30936, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 288, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 288, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 288, + 150 + ], + "type": "text", + "content": "[30] Gang Li, Heliang Zheng, Chaoyue Wang, Chang Li, Changwen Zheng, and Dacheng Tao. 3ddesigner: Towards photorealistic 3d object generation and editing with text-guided diffusion models. arXiv preprint arXiv:2211.14108, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 150, + 288, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 150, + 288, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 150, + 288, + 194 + ], + "type": "text", + "content": "[31] Weiyu Li, Rui Chen, Xuelin Chen, and Ping Tan. Sweetdreamer: Aligning geometric priors in 2d diffusion for consistent text-to-3d. arXiv preprint arXiv:2310.02596, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 194, + 288, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 194, + 288, + 248 + ], + "spans": [ + { + "bbox": [ + 48, + 194, + 288, + 248 + ], + "type": "text", + "content": "[32] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, pages 300–309, 2023. 1, 2, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 250, + 288, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 288, + 294 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 288, + 294 + ], + "type": "text", + "content": "[33] Yukang Lin, Haonan Han, Chaoqun Gong, Zunnan Xu, Yachao Zhang, and Xiu Li. Consistent123: One image to highly consistent 3d asset using case-aware diffusion priors. arXiv preprint arXiv:2309.17261, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 294, + 288, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 294, + 288, + 337 + ], + "spans": [ + { + "bbox": [ + 48, + 294, + 288, + 337 + ], + "type": "text", + "content": "[34] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Zexiang Xu, Hao Su, et al. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. arXiv preprint arXiv:2306.16928, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 337, + 288, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 337, + 288, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 337, + 288, + 381 + ], + "type": "text", + "content": "[35] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. arXiv preprint arXiv:2303.11328, 2023. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 381, + 288, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 381, + 288, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 381, + 288, + 426 + ], + "type": "text", + "content": "[36] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 426, + 288, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 288, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 288, + 470 + ], + "type": "text", + "content": "[37] Zhen Liu, Yao Feng, Michael J Black, Derek Nowrouzezahrai, Liam Paull, and Weiyang Liu. Meshdiffusion: Score-based generative 3d mesh modeling. arXiv preprint arXiv:2303.08133, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 471, + 288, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 288, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 288, + 525 + ], + "type": "text", + "content": "[38] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. arXiv preprint arXiv:2310.15008, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 525, + 288, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 288, + 570 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 288, + 570 + ], + "type": "text", + "content": "[39] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 851-866. 2023. 2, 4, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 570, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 288, + 624 + ], + "type": "text", + "content": "[40] Jonathan Lorraine, Kevin Xie, Xiaohui Zeng, Chen-Hsuan Lin, Towaki Takikawa, Nicholas Sharp, Tsung-Yi Lin, Ming-Yu Liu, Sanja Fidler, and James Lucas. Att3d: Amortized text-to-3d object synthesis. arXiv preprint arXiv:2306.07349, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 288, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 288, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 288, + 657 + ], + "type": "text", + "content": "[41] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In CVPR, pages 2837-2845, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 288, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 691 + ], + "type": "text", + "content": "[42] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. arXiv preprint arXiv:2306.07279, 2023. 5, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "text", + "content": "[43] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 715 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 327, + 72, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 72, + 545, + 95 + ], + "type": "text", + "content": "3d shapes and textures. In " + }, + { + "bbox": [ + 327, + 72, + 545, + 95 + ], + "type": "inline_equation", + "content": "CVPR" + }, + { + "bbox": [ + 327, + 72, + 545, + 95 + ], + "type": "text", + "content": ", pages 12663-12673, 2023. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "type": "text", + "content": "[44] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In CVPR, pages 13492-13502, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 129, + 545, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 175 + ], + "type": "text", + "content": "[45] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, pages 405-421, 2020. 2, 3, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 175, + 545, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 219 + ], + "type": "text", + "content": "[46] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 conference papers, pages 1-8, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 220, + 545, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 263 + ], + "type": "text", + "content": "[47] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "type": "text", + "content": "[48] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 1, 2, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 309, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 354 + ], + "type": "text", + "content": "[49] Yichen Ouyang, Wenhao Chai, Jiayi Ye, Dapeng Tao, Yibing Zhan, and Gaoang Wang. Chasing consistency in text-to-3d generation from a single image. arXiv preprint arXiv:2309.03599, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 354, + 545, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 388 + ], + "type": "text", + "content": "[50] Jangho Park, Gihyun Kwon, and Jong Chul Ye. Ed-nerf: Efficient text-guided editing of 3d scene using latent space nerf. arXiv preprint arXiv:2310.02712, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 388, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 443 + ], + "type": "text", + "content": "[51] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 32, 2019. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 444, + 545, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 444, + 545, + 478 + ], + "spans": [ + { + "bbox": [ + 307, + 444, + 545, + 478 + ], + "type": "text", + "content": "[52] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 479, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 479, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 479, + 545, + 521 + ], + "type": "text", + "content": "[53] Zekun Qi, Muzhou Yu, Runpei Dong, and Kaisheng Ma. Vpp: Efficient conditional 3d generation via voxel-point progressive representation. arXiv preprint arXiv:2307.16605, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 523, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 545, + 589 + ], + "type": "text", + "content": "[54] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, et al. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. arXiv preprint arXiv:2306.17843, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 590, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 624 + ], + "type": "text", + "content": "[55] Sigal Raab, Inbal Leibovitch, Guy Tevet, Moab Arar, Amit H Bermano, and Daniel Cohen-Or. Single motion diffusion. arXiv preprint arXiv:2302.05905, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 624, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 679 + ], + "type": "text", + "content": "[56] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 680, + 545, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 715 + ], + "type": "text", + "content": "[57] Amit Raj, Srinivas Kaza, Ben Poole, Michael Niemeyer, Nataniel Ruiz, Ben Mildenhall, Shiran Zada, Kfir Aberman, Michael Rubinstein, Jonathan Barron, et al. Dream" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6805" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "booth3d: Subject-driven text-to-3d generation. arXiv preprint arXiv:2303.13508, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 138 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 138 + ], + "type": "text", + "content": "[58] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "type": "text", + "content": "[59] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, Marco Fumero, and Kamal Rahimi Malekshan. Clip-forge: Towards zero-shot text-to-shape generation. In CVPR, pages 18603-18613, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 184, + 287, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 287, + 237 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 287, + 237 + ], + "type": "text", + "content": "[60] Kyle Sargent, Zizhang Li, Tanmay Shah, Charles Herrmann, Hong-Xing Yu, Yunzhi Zhang, Eric Ryan Chan, Dmitry Lagun, Li Fei-Fei, Deqing Sun, et al. Zeronvs: Zero-shot 360-degree view synthesis from a single real image. arXiv preprint arXiv:2310.17994, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 239, + 287, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 287, + 281 + ], + "type": "text", + "content": "[61] Hoigi Seo, Hayeon Kim, Gwanghyun Kim, and Se Young Chun. Ditto-nerf: Diffusion-based iterative text to omnidirectional 3d model. arXiv preprint arXiv:2304.02827, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 283, + 287, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 283, + 287, + 336 + ], + "spans": [ + { + "bbox": [ + 48, + 283, + 287, + 336 + ], + "type": "text", + "content": "[62] Junyoung Seo, Wooseok Jang, Min-Seop Kwak, Jaehoon Ko, Hyeonsu Kim, Junho Kim, Jin-Hwa Kim, Jiyoung Lee, and Seungryong Kim. Let 2d diffusion model know 3d-consistency for robust text-to-3d generation. arXiv preprint arXiv:2303.07937, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 338, + 287, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 287, + 370 + ], + "type": "text", + "content": "[63] Yonatan Shafir, Guy Tevet, Roy Kapon, and Amit H Bermano. Human motion diffusion as a generative prior. arXiv preprint arXiv:2303.01418, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 371, + 287, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 414 + ], + "type": "text", + "content": "[64] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. NeurIPS, 34: 6087-6101, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 415, + 287, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 468 + ], + "type": "text", + "content": "[65] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 471, + 287, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 287, + 524 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 287, + 524 + ], + "type": "text", + "content": "[66] Yukai Shi, Jianan Wang, He Cao, Boshi Tang, Xianbiao Qi, Tianyu Yang, Yukun Huang, Shilong Liu, Lei Zhang, and Heung-Yeung Shum. Toss: High-quality text-guided novel view synthesis from a single image. arXiv preprint arXiv:2310.10644, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 525, + 287, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 558 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 558 + ], + "type": "text", + "content": "[67] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 1, 2, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 559, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 602 + ], + "type": "text", + "content": "[68] Liangchen Song, Liangliang Cao, Hongyu Xu, Kai Kang, Feng Tang, Junsong Yuan, and Yang Zhao. Roomdreamer: Text-driven 3d indoor scene synthesis with coherent geometry and texture. arXiv preprint arXiv:2305.11337, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 603, + 287, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 287, + 647 + ], + "type": "text", + "content": "[69] Jingxiang Sun, Bo Zhang, Ruizhi Shao, Lizhen Wang, Wen Liu, Zhenda Xie, and Yebin Liu. Dreamcraft3d: Hierarchical 3d generation with bootstrapped diffusion prior. arXiv preprint arXiv:2310.16818, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[70] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "text", + "content": "[71] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "3d creation from a single image with diffusion prior. arXiv preprint arXiv:2303.14184, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 96, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 139 + ], + "type": "text", + "content": "[72] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 2, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 141, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 185 + ], + "type": "text", + "content": "[73] Christina Tsalicoglou, Fabian Manhardt, Alessio Tonioni, Michael Niemeyer, and Federico Tombari. Textmesh: Generation of realistic 3d meshes from text prompts. arXiv preprint arXiv:2304.12439, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 186, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 228 + ], + "type": "text", + "content": "[74] Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, Karsten Kreis, et al. Lion: Latent point diffusion models for 3d shape generation. NeurIPS, 35:10021-10039, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 231, + 545, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 231, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 307, + 231, + 545, + 273 + ], + "type": "text", + "content": "[75] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In CVPR, pages 3835-3844, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 276, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 319 + ], + "type": "text", + "content": "[76] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In CVPR, pages 12619-12629, 2023. 1, 2, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 321, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 321, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 307, + 321, + 545, + 374 + ], + "type": "text", + "content": "[77] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213, 2023. 1, 2, 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 376, + 545, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 421 + ], + "type": "text", + "content": "[78] Haohan Weng, Tianyu Yang, Jianan Wang, Yu Li, Tong Zhang, CL Chen, and Lei Zhang. Consistent123: Improve consistency for one image to 3d object synthesis. arXiv preprint arXiv:2310.08092, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 422, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 545, + 454 + ], + "type": "text", + "content": "[79] Zhenzhen Weng, Zeyu Wang, and Serena Yeung. Zero-)rator: Zero-shot 3d avatar generation from a single image. arXiv preprint arXiv:2305.16411, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 456, + 545, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 509 + ], + "type": "text", + "content": "[80] Jinbo Wu, Xiaobo Gao, Xing Liu, Zhengyang Shen, Chen Zhao, Haocheng Feng, Jingtuo Liu, and Errui Ding. Hdfusion: Detailed text-to-3d generation leveraging multiple noise estimation. arXiv preprint arXiv:2307.16183, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 511, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 511, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 511, + 545, + 555 + ], + "type": "text", + "content": "[81] Jiale Xu, Xintao Wang, Weihao Cheng, Yan-Pei Cao, Ying Shan, Xiaohu Qie, and Shenghua Gao. Dream3d: Zero-shot text-to-3d synthesis using 3d shape prior and text-to-image diffusion models. In CVPR, pages 20908–20918, 2023. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 556, + 545, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 556, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 307, + 556, + 545, + 599 + ], + "type": "text", + "content": "[82] Jiayu Yang, Ziang Cheng, Yunfei Duan, Pan Ji, and Hongdong Li. Consistnet: Enforcing 3d consistency for multiview images diffusion. arXiv preprint arXiv:2310.10343, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 601, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 644 + ], + "type": "text", + "content": "[83] Jianglong Ye, Peng Wang, Kejie Li, Yichun Shi, and Heng Wang. Consistent-1-to-3: Consistent image to 3d view synthesis via geometry-aware diffusion models. arXiv preprint arXiv:2310.03020, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "text", + "content": "[84] Wang Yifan, Felice Serena, Shihao Wu, Cengiz Öztireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM Transactions on Graphics (TOG), 38(6):1-14, 2019. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "type": "text", + "content": "[85] Chaohui Yu, Qiang Zhou, Jingliang Li, Zhe Zhang, Zhibin Wang, and Fan Wang. Points-to-3d: Bridging the gap be" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6806" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 418 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 66, + 73, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 73, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 73, + 286, + 95 + ], + "type": "text", + "content": "tween sparse points and shape-controllable text-to-3d generation. In ACM MM, pages 6841-6850, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "type": "text", + "content": "[86] Huichao Zhang, Bowen Chen, Hao Yang, Liao Qu, Xu Wang, Li Chen, Chao Long, Feida Zhu, Kang Du, and Min Zheng. Avatarverse: High-quality & stable 3d avatar creation from text and pose. arXiv preprint arXiv:2308.03610, 2023. 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 206 + ], + "type": "text", + "content": "[87] Longwen Zhang, Qiwei Qiu, Hongyang Lin, Qixuan Zhang, Cheng Shi, Wei Yang, Ye Shi, Sibei Yang, Lan Xu, and Jingyi Yu. Dreamface: Progressive generation of animatable 3d faces under text guidance. arXiv preprint arXiv:2304.03117, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 286, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 286, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 286, + 251 + ], + "type": "text", + "content": "[88] Mingyuan Zhang, Xinying Guo, Liang Pan, Zhongang Cai, Fangzhou Hong, Huirong Li, Lei Yang, and Ziwei Liu. Remodiffuse: Retrieval-augmented motion diffusion model. arXiv preprint arXiv:2304.01116, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 253, + 287, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 253, + 287, + 296 + ], + "spans": [ + { + "bbox": [ + 48, + 253, + 287, + 296 + ], + "type": "text", + "content": "[89] Mengyi Zhao, Mengyuan Liu, Bin Ren, Shuling Dai, and Nicu Sebe. Modiff: Action-conditioned 3d motion generation with denoising diffusion probabilistic models. arXiv preprint arXiv:2301.03949, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 297, + 287, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 297, + 287, + 351 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 287, + 351 + ], + "type": "text", + "content": "[90] Minda Zhao, Chaoyi Zhao, Xinyue Liang, Lincheng Li, Zeng Zhao, Zhipeng Hu, Changjie Fan, and Xin Yu. Efficientdreamer: High-fidelity and robust 3d creation via orthogonal-view diffusion prior. arXiv preprint arXiv:2308.13223, 2023. 1, 2, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 353, + 287, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 353, + 287, + 385 + ], + "spans": [ + { + "bbox": [ + 48, + 353, + 287, + 385 + ], + "type": "text", + "content": "[91] Zixiang Zhou and Baoyuan Wang. Ude: A unified driving engine for human motion generation. In CVPR, pages 5632-5641, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 387, + 287, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 387, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 48, + 387, + 287, + 418 + ], + "type": "text", + "content": "[92] Joseph Zhu and Peiye Zhuang. Hifa: High-fidelity text-to-3d with advanced diffusion guidance. arXiv preprint arXiv:2305.18766, 2023. 2" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6807" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/GaussianEditor_ Editing 3D Gaussians Delicately with Text Instructions/22c39c85-c3be-4622-b7a4-f891bcbdf14e_content_list.json b/2024/GaussianEditor_ Editing 3D Gaussians Delicately with Text Instructions/22c39c85-c3be-4622-b7a4-f891bcbdf14e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1394e0ef106f3ed617457f64d88dc3a707cc436e --- /dev/null +++ b/2024/GaussianEditor_ Editing 3D Gaussians Delicately with Text Instructions/22c39c85-c3be-4622-b7a4-f891bcbdf14e_content_list.json @@ -0,0 +1,1749 @@ +[ + { + "type": "text", + "text": "GaussianEditor: Editing 3D Gaussians Delicately with Text Instructions", + "text_level": 1, + "bbox": [ + 120, + 130, + 849, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Junjie Wang, $^{\\star}$ Jiemin Fang $^{\\star}$ , Xiaopeng Zhang, Lingxi Xie, Qi Tian Huawei Inc.", + "bbox": [ + 192, + 179, + 776, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{is.wangjunjie, jaminfong, zxphistory, 198808xc}@gmail.com tian.qil@huawei.com", + "bbox": [ + 133, + 218, + 834, + 234 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/82c2d2fb93de361a4093dd459c31c3ea3797437513a565caafe64e3ee9fd1981.jpg", + "image_caption": [ + "Figure 1. We propose GaussianEditor, an interactive framework to achieve delicate 3D scene editing following text instructions. As shown in this figure, our method can precisely control the editing region and achieve multi-round editing." + ], + "image_footnote": [], + "bbox": [ + 156, + 247, + 789, + 578 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 626, + 313, + 641 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, impressive results have been achieved in 3D scene editing with text instructions based on a 2D diffusion model. However, current diffusion models primarily generate images by predicting noise in the latent space, and the editing is usually applied to the whole image, which makes it challenging to perform delicate, especially localized, editing for 3D scenes. Inspired by recent 3D Gaussian splatting, we propose a systematic framework, named GaussianEditor, to edit 3D scenes delicately via 3D Gaussians with text instructions. Benefiting from the explicit property of 3D Gaussians, we design a series of techniques to achieve delicate editing. Specifically, we first extract the region of interest (RoI) corresponding to the text instruction, aligning it to 3D Gaussians. The Gaussian RoI is fur", + "bbox": [ + 75, + 646, + 473, + 859 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ther used to control the editing process. Our framework can achieve more delicate and precise editing of 3D scenes than previous methods while enjoying much faster training speed, i.e. within 20 minutes on a single V100 GPU, more than twice as fast as Instruct-NeRF2NeRF (45 minutes - 2 hours)1. The project page is at GaussianEditor.github.io.", + "bbox": [ + 500, + 627, + 893, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 762, + 633, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Creating 3D assets has played a critical role in many applications and industries, e.g. movie/game production, artistic creation, AR, VR etc. However, this process is usually expensive and cumbersome, especially for traditional pipelines. Designers need to take a lot of labor and time to", + "bbox": [ + 496, + 787, + 893, + 866 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contributions.", + "bbox": [ + 94, + 875, + 212, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 96, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1The editing time varies in different scenes according to the scene structure complexity.", + "bbox": [ + 500, + 875, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20902", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "finish each step, e.g. sketching, building structures, creating textures etc. One cheap and effective way of creating high-quality 3D assets is to start from an existing scene, capturing, modeling, and editing the scene and obtaining the wanted one. This approach can be also used for user-interactive entertainment applications.", + "bbox": [ + 75, + 90, + 468, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural radiance field methods [2, 3, 6, 29, 31, 42, 45] have shown great power in representing 3D scenes and synthesizing novel-view images. Past years have witnessed the rapid development of NeRF and its variants, from both quality and efficiency perspectives. Editing a pre-trained NeRF model becomes a promising way to edit 3D scenes. Represented by Instruct-NeRF2NeRF [11], researchers propose to use the image-conditioned 2D diffusion model, e.g. InstructPix2Pix [4], to edit 3D scenes simply with text instructions. Notable results have been achieved as real scenes can be changed following the text instruction. However, current 2D diffusion models face challenges in accurately localizing editing regions, which hinders the generation of finely edited scenes due to the change of unintended regions. Even though some works [30] propose to constrain the editing region on edited 2D images, the editing region is not accurately localized and hard to apply to the 3D representation. Besides, NeRF-based methods [9, 48] bear coupling effects between different spatial positions, e.g. different points are queried from the same MLP field (for implicit representations) or voxel vertices (for explicit representations).", + "bbox": [ + 75, + 184, + 472, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent 3D Gaussian Splatting [18] (3D-GS) has been a groundbreaking work in the radiance field, which is the first to achieve a real sense of real-time rendering while enjoying high rendering quality and training speed. Besides its efficiency, we further notice its natural explicit property. 3D-GS has a great advantage for editing tasks as each 3D Gaussian exists individually. Editing 3D scenes by directly manipulating 3D Gaussians with desired constraints is easy.", + "bbox": [ + 75, + 505, + 468, + 626 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Aiming at editing 3D scenes delicately, we propose to represent the scene with 3D Gaussians, which can be edited with text instructions, and name our method as GaussianEditor. GaussianEditor is divided into three main parts to achieve precise control for editing regions. The first is the region of interest (RoI) extraction from the given text instruction. The instruction may be complex or indirect while this module helps extract the keywords matching the RoI for editing. The second part aligns the extracted text RoI to the 3D Gaussian space through the image space, where a grounding segmentation module is applied. The last part is to edit the original 3D Gaussians delicately with constraints in the obtained 3D Gaussian RoI. With the above processes, the region for editing can be precisely localized simply from text instructions, which constrains the 3D Gaussian updating to obtain a delicately edited new 3D scene. Besides, we enable interfaces for users to introduce more exact instructions for more delicate editing, e.g. Gaussian point selecting", + "bbox": [ + 75, + 628, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and 3D boxes for modifying the editing regions2.", + "bbox": [ + 498, + 90, + 823, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows.", + "bbox": [ + 517, + 107, + 841, + 119 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- As far as we know, our GaussianEditor is one of the first systematic methods to achieve delicate 3D scene editing based on 3D Gaussian splatting.", + "- A series of techniques are designed and proposed to precisely localize the editing region of interest, which are aligned and applied to 3D Gaussians. Though some submodules are from existing works, we believe integrating these awesome techniques to work effectively is a valuable topic, which is what we focus on in this paper.", + "- Our method achieves a series of more delicate editing results compared with the previous representative work Instruct-NeRF2NeRF [11] with much shorter training time (within 20 minutes v.s. 45 minutes - 2 hours)." + ], + "bbox": [ + 500, + 122, + 890, + 318 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 333, + 640, + 348 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2D Image Editing with Diffusion Models. By advancements in diffusion model technology [13, 43], numerous generative models [40] have attained remarkably impressive outcomes in image synthesis. Recent progressions in diffusion models have showcased their adeptness in fabricating lifelike images from arbitrary textual inputs [8, 14, 39, 41, 44]. Harnessing the robust semantic comprehension and image generation capabilities of foundational diffusion models, an escalating number of research explorations are currently employing diffusion models as a fundamental framework for implementing text-based image editing functionalities [33, 36, 37, 40]. Some of these methodologies necessitate the manual provision of captions for both the original and edited images [12], while others mandate specific scenario-based training for optimization [38]. These requisites have rendered it arduous for ordinary users to avail themselves of such techniques. Expanding upon this foundation, iP2P [4] introduces instruction-based capabilities to image editing, enabling users to simply input an image and appraise the model of the desired alterations. This user-friendly approach facilitates the democratization of image editing in a more accessible manner.", + "bbox": [ + 496, + 358, + 890, + 690 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Scene Editing of Radiance Fields. 3D Scene Editing of Radiance Fields has gained significant popularity as a recent research direction [1, 10, 15, 20, 22-25, 28, 34, 46, 47, 51-53]. The main objective of such methods is to enable the manipulation of both the geometry and appearance of 3D scene representations. However, editing such scenes poses inherent challenges due to the implicit nature of traditional NeRF representations, which lack precise localization capabilities. As a result, previous works have primarily focused on achieving global style transformations of 3D scenes [7, 16, 17, 32, 48, 50, 54]. While some efforts", + "bbox": [ + 496, + 698, + 890, + 864 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "2These additional instructions are applied to generate the man with two different edited half faces in Fig. 1.", + "bbox": [ + 500, + 875, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20903", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/45ed546d67486b7683a68eef9a6eb990c955081fec28ed5a0c96e682cea5dcfa.jpg", + "image_caption": [ + "Figure 2. Our framework, named GaussianEditor, consists of three key steps. First, a module $\\mathcal{M}_{Desc}$ is used to get the description of the input scene, which is put to an LLM assistant $\\mathcal{M}_{LLM}$ with the text instruction $\\mathcal{T}$ provided by the user to obtain the text RoI $\\mathcal{T}_{RoI}$ . Second, a grounding segmentation module $\\mathcal{M}_{Seg}$ is used to convert $\\mathcal{T}_{RoI}$ to image RoI $\\mathcal{I}_{RoI}$ , which is then lifted to 3D Gaussians RoI $\\mathcal{G}_{RoI}$ by RoI lifting $\\mathcal{M}_{Lift}$ , where additional user instructions $\\mathcal{O}$ can be incorporated. Third, following the user instruction $\\mathcal{T}$ , rendered image $\\mathcal{I}_{render}$ from randomly chosen views is edited by a diffusion model $\\mathcal{M}_{DM}$ . The loss between $\\mathcal{I}_{render}$ and edited one $\\mathcal{I}_{edit}$ is calculated. Finally, gradient backpropagation and optimization are performed within the Gaussian RoI $\\mathcal{G}_{RoI}$ to get the edited scene $\\mathcal{G}_{edit}$ ." + ], + "image_footnote": [], + "bbox": [ + 83, + 89, + 883, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "have been made towards object-centric scene editing[55], keeping the background unchanged has been a persistent challenge. For example, a recently proposed work, Instruct-NeRF2NeRF [11], implements text instruction-controlled 3D scene editing capabilities, achieving excellent editing effects while maintaining user-friendliness. However, it relies on the editing effect of 2D images, which may cause global changes to the 3D scene. A subsequent work [30] attempts to compute the relevance map between edited and unedited images to localize the editing area. The relevance map may be unreliable when the 2D IP2P [4] model fails. Other efforts [23] attempt to rely on the 3D coordinates entered by the user to decide the editing area. The introduction of 3D Gaussians [18] has provided an opportunity to address this limitation. Its explicit 3D representation methodology enables accurate selection and manipulation of editing areas. By incorporating LLMs, the whole process can be more automated.", + "bbox": [ + 75, + 628, + 472, + 898 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 627, + 589, + 642 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first review 3D representation methods in Sec. 3.1. Subsequently, in Sec. 3.2, we overview our proposed approach, which mainly includes three modules. Sec. 3.3 delves into the precise Region of Interest (RoI) extraction of text instructions, using scene description generation module $\\mathcal{M}_{Desc}$ and LLM assistant $\\mathcal{M}_{LLM}$ . Sec. 3.4 introduces how to align the instruction RoI with 3D Gaussians, using grounding segmentation module $\\mathcal{M}_{Seg}$ and RoI lifting module $\\mathcal{M}_{Lift}$ . Finally, Sec. 3.5 describes the delicate editing process within the obtained Gaussian RoI, using text to image diffusion model $\\mathcal{M}_{DM}$ .", + "bbox": [ + 496, + 652, + 890, + 819 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 500, + 830, + 638, + 845 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Gaussian Splatting. 3D Gaussian splatting [18] is a recent powerful 3D representation method. It represents the 3D scene with point-like 3D Gaussians $\\mathcal{G} = \\{g_1, g_2 \\dots g_N\\}$ ,", + "bbox": [ + 498, + 854, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20904", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $g_{i} = \\{\\mu, \\Sigma, c, \\alpha\\}$ and $i \\in \\{1, \\dots, N\\}$ . Among them, $\\mu \\in \\mathbb{R}^3$ is the position where the Gaussian centers, $\\Sigma \\in \\mathbb{R}^7$ denotes the 3D covariance matrix, $c \\in \\mathbb{R}^3$ is the RGB color and $\\alpha \\in \\mathbb{R}^1$ is the opacity. Benefitting from the compact representation of Gaussians and efficient differentiable rendering approach, 3D Gaussian splatting achieves real-time rendering with high quality. The splatting rendering process can be formulated as", + "bbox": [ + 75, + 90, + 472, + 210 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC = \\sum_ {i \\in \\mathcal {N}} c _ {i} \\sigma_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\sigma_ {j}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 215, + 468, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\sigma_{i} = \\alpha_{i}e^{-\\frac{1}{2}(x_{i})^{T}\\Sigma^{-1}(x_{i})}$ represents the influence of the Gaussian to the image pixel and $x_{i}$ is the distance between the 3D point and the center of the i-th Gaussian.", + "bbox": [ + 76, + 263, + 470, + 313 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Overall Framework", + "text_level": 1, + "bbox": [ + 76, + 319, + 266, + 334 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a group of 3D Gaussians $\\mathcal{G}_{input}$ for an input scene and a text instruction $\\mathcal{T}$ for editing, our Gaussian editor $\\mathcal{E}$ can edit the 3D Gaussians delicately into a new one, denoted as $\\mathcal{G}_{edit}$ , with the guidance of the instruction. The whole process can be formulated as", + "bbox": [ + 75, + 342, + 468, + 417 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} _ {\\text {e d i t}} = \\mathcal {E} \\left(\\mathcal {G} _ {\\text {i n p u t}}, \\mathcal {T}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 425, + 468, + 441 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fig. 2 illustrates the overall framework of our approach, which consists of three main steps. First, the Region of Interest (RoI) is extracted from the text instruction. In this step, we employ a module named scene description generation $\\mathcal{M}_{\\text{Description}}$ to get the description of the input scene. We then input the scene description $\\mathcal{T}_{\\text{scene}}$ and text instruction $\\mathcal{T}$ into a large language model assistant $\\mathcal{M}_{\\text{LLM}}$ to determine where we should make edits in the scene. The output of this step is referred to as the instruction RoI $\\mathcal{T}_{\\text{RoI}}$ .", + "bbox": [ + 75, + 446, + 468, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The next step is the 3D Gaussian RoI alignment. We use a grounding segmentation module $\\mathcal{M}_{Seg}$ to convert the RoI from text space, i.e. $\\mathcal{T}_{RoI}$ , to the image space, i.e. $\\mathcal{I}_{RoI}$ . Then the image RoI $\\mathcal{I}_{RoI}$ is lifted to the RoI of 3D Gaussians $\\mathcal{G}_{RoI}$ through RoI lifting module $\\mathcal{M}_{Lift}$ . The Gaussian RoI allows us to control the regions where edits will be applied precisely.", + "bbox": [ + 75, + 584, + 468, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The last step is delicate editing within the Gaussian RoI. In this step, we randomly sample the view to obtain the rendered image $\\mathcal{I}_{\\text {render }}$ . A 2D diffusion model $\\mathcal{M}_{DM}$ is used to perform the editing process on the rendered image $\\mathcal{I}_{\\text {render }}$ , with the user instruction $\\mathcal{T}$ and the image $\\mathcal{I}_{\\text {input }}$ of input scene as conditions. The resulting edited image is denoted as $\\mathcal{I}_{\\text {edit }}$ . Subsequently, we calculate the loss between $\\mathcal{I}_{\\text {edit }}$ and $\\mathcal{I}_{\\text {render }}$ and make gradient backpropagation within $\\mathcal{G}_{\\text {RoI }}$ . This implies that only the regions specified by the RoI can receive corresponding gradients during the back-propagation process. Finally, optimization is executed based on these gradients. The final optimized scene representation $\\mathcal{G}_{\\text {edit }}$ is obtained through several rounds of iterative optimization.", + "bbox": [ + 75, + 688, + 470, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0fd46d91fccb7c66678e177ba3b61ae5b740bbdcddb305bf1811e5bdbf7f46f2.jpg", + "image_caption": [ + "Figure 3. The process of obtaining scene description." + ], + "image_footnote": [], + "bbox": [ + 511, + 90, + 875, + 415 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. RoI Extraction of Text Instruction", + "text_level": 1, + "bbox": [ + 500, + 444, + 799, + 459 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The instruction RoI is extracted for the editing regions from both the input 3D scene $\\mathcal{G}_{input}$ and the text instruction $\\mathcal{T}$ provided by the user. To achieve this, we employ a multimodal model $\\mathcal{M}_{MM}$ in conjunction with the large language model assistant $\\mathcal{M}_{LLM}$ . The first step is the scene description generation $\\mathcal{M}_{Desc}$ , which aims to get the scene description $\\mathcal{T}_{scene}$ from 3D Gaussians $\\mathcal{G}_{input}$ :", + "bbox": [ + 496, + 467, + 890, + 575 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {T} _ {\\text {s c e n e}} = \\mathcal {M} _ {\\text {D e s c}} \\left(\\mathcal {G} _ {\\text {i n p u t}}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 582, + 890, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The process of the scene description generation $\\mathcal{M}_{Desc}$ is shown in Fig. 3. By leveraging the technique of differentiable splatting as shown in Eq. 1, a set of 2D image samples $\\{\\mathcal{I}_{sample}\\}$ are generated and then inputted into a multimodal model $\\mathcal{M}_{MM}$ to generate corresponding text descriptions $\\{\\mathcal{T}_{sample}\\}$ :", + "bbox": [ + 496, + 607, + 890, + 699 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {T} _ {\\text {s a m p l e}} = \\mathcal {M} _ {M M} \\left(\\mathcal {P} _ {M M}, \\mathcal {I} _ {\\text {s a m p l e}}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 705, + 890, + 723 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{P}_{MM}$ is a prompt, such as \"What is the content of the image\", for multimodal model $\\mathcal{M}_{MM}$ to get precise description. Subsequently, these descriptions $\\{\\mathcal{T}_{\\text {sample }}\\}$ are fed into a large language model $\\mathcal{M}_{LLM}$ , which is specifically instructed by a prompt $\\mathcal{P}_{\\text {merge }}$ to merge descriptions of diverse views into one detailed scene description $\\mathcal{T}_{\\text {scene }}$ :", + "bbox": [ + 496, + 731, + 890, + 837 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {T} _ {\\text {s c e n e}} = \\mathcal {M} _ {L L M} \\left(\\mathcal {P} _ {\\text {m e r g e}}, \\left\\{\\mathcal {T} _ {\\text {s a m p l e}} \\right\\}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 845, + 890, + 863 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After that, the scene description $\\mathcal{T}_{\\text{scene}}$ and the user instruction $\\mathcal{T}$ are combined with a predefined tem", + "bbox": [ + 496, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "footer", + "text": "20905", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "plate $\\mathcal{T}_{\\text{template}}$ : \"Text description: $\\mathcal{T}_{\\text{scene}}$ Edit Instruction: $\\mathcal{T}$ Answer:\" to form the user message $\\mathcal{T}_{\\text{user}} = \\mathcal{T}_{\\text{template}}(\\mathcal{T}_{\\text{scene}}, \\mathcal{T})$ . The LLM model $\\mathcal{M}_{\\text{LLM}}$ is used to extract the instruction RoI $\\mathcal{T}_{\\text{RoI}}$ from user message $\\mathcal{T}_{\\text{user}}$ with a new prompt $\\mathcal{P}_{\\text{extract}}$ :", + "bbox": [ + 76, + 90, + 468, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {T} _ {R o I} = \\mathcal {M} _ {L L M} \\left(\\mathcal {P} _ {\\text {e x t r a c t}}, \\mathcal {T} _ {\\text {u s e r}}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 178, + 468, + 194 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. 3D Gaussian RoI Alignment", + "text_level": 1, + "bbox": [ + 76, + 204, + 330, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To confine the 3D editing region within the instruction RoI, 3D Gaussian RoI $\\mathcal{G}_{RoI}$ is aligned with the text RoI $\\mathcal{T}_{RoI}$ . First, The RoI in the text space is transformed into the image space via a grounding segmentation module $\\mathcal{M}_{Seg}$ :", + "bbox": [ + 76, + 227, + 468, + 289 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {I} _ {R o I} = \\mathcal {M} _ {S e g} \\left(\\mathcal {I} _ {\\text {i n p u t}}, \\mathcal {T} _ {R o I}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 300, + 468, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{I}_{input}$ is rendered image of the input scene $\\mathcal{G}_{input}$ .", + "bbox": [ + 76, + 327, + 455, + 343 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then we lift the RoI $\\mathcal{I}_{RoI}$ in the image space to 3D Gaussian $\\mathcal{G}_{RoI}$ through training. To achieve this, an additional RoI attribute $r\\in \\mathbb{R}^1$ was added to 3D Gaussian $g_{i} = \\{\\mu_{i},\\Sigma_{i},c_{i},\\alpha_{i},r_{i}\\}$ . $r$ is initialized to 0, which means it is not in the Gaussians RoI, and 1 means it is inside the RoI. The set of $r$ is denoted as $\\mathcal{R}\\in \\mathbb{R}^{\\mathcal{N},1}$ , where the $\\mathcal{N}$ is the number of 3D Gaussians $\\mathcal{G}_{input}$ .", + "bbox": [ + 75, + 343, + 468, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then the color $c_{i}$ in Eq. 1 was rewritten with $r_{i}$ to get the rendered RoI $\\mathcal{T}_{RoI}^{render}$ :", + "bbox": [ + 76, + 449, + 468, + 479 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {I} _ {R o I} ^ {\\text {r e n d e r}} = \\sum_ {i \\in \\mathcal {N}} r _ {i} \\sigma_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\sigma_ {j}). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 489, + 468, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Taking inspiration from SA3D [5], to get the trained Gaussians RoI $\\mathcal{G}_{RoI}^{train}$ , we adopt a similar loss function to supervise the training process:", + "bbox": [ + 76, + 542, + 468, + 587 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p r o j} = \\lambda_ {1} \\sum \\left(\\mathcal {I} _ {R o I} ^ {\\text {r e n d e r}} \\cdot \\mathcal {I} _ {R o I}\\right) + \\lambda_ {2} \\sum \\left(\\left(1 - \\mathcal {I} _ {R o I} ^ {\\text {r e n d e r}}\\right) \\cdot \\mathcal {I} _ {R o I}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 598, + 478, + 632 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{1}$ and $\\lambda_{2}$ are hyperparameters. The $r$ in Eq. 8 is updated via $r\\gets r - \\eta \\frac{\\partial\\mathcal{L}_{proj}}{\\partial r}$ with gradient descent, where $\\eta$ denotes the learning rate. Eq. 9 encourages rendered RoI to cover the Image RoI and not exceed it. Additionally, the user can modify the trained Gaussian RoI $\\mathcal{G}_{RoI}^{train}$ by giving added Gaussian RoI $\\mathcal{G}_{RoI}^{add}$ , deleted Gaussian RoI $\\mathcal{G}_{RoI}^{del}$ and 3D box $\\mathcal{B}_{3D}$ :", + "bbox": [ + 76, + 633, + 468, + 739 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} _ {R o I} = \\left(\\mathcal {G} _ {R o I} ^ {\\text {t r a i n}} \\cup \\mathcal {G} _ {R o I} ^ {\\text {a d d}} - \\mathcal {G} _ {R o I} ^ {\\text {d e l}}\\right) \\cap \\mathcal {B} _ {3 D}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 750, + 468, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathcal{G}_{RoI}^{add}$ represents the 3D Gaussians user wants to edit, $\\mathcal{G}_{RoI}^{del}$ means 3D Gaussians user wants to keep from editing, $\\mathcal{B}_{3D}$ is the coordinates of 3D cuboid it limits RoI to inside the box. $\\mathcal{G}_{RoI}$ is the aligned RoI with the text RoI. For example, when editing the left face of the man in Fig. 1, grounding segmentation failed to ground \"left face\", instead, it grounded the whole face. In this scenario, the user can use the interactive interface to set the right face as $\\mathcal{G}_{RoI}^{del}$ or enter", + "bbox": [ + 76, + 777, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the rectangular box where the left face is located as $\\mathcal{B}_{3D}$ . The lifting process $\\mathcal{M}_{\\text{Lift}}$ can be represented as:", + "bbox": [ + 500, + 90, + 890, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} _ {R o I} = \\mathcal {M} _ {\\text {L i f t}} \\left(\\mathcal {I} _ {R o I}, \\mathcal {O}\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 606, + 132, + 890, + 148 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{O} = \\{\\mathcal{G}_{RoI}^{add},\\mathcal{G}_{RoI}^{del},\\mathcal{B}_{3D}\\}$ is optional instructions.", + "bbox": [ + 500, + 157, + 866, + 175 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Delicate Editing within Gaussian RoI", + "text_level": 1, + "bbox": [ + 500, + 181, + 821, + 198 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To achieve delicate editing in 3D scenes, we use the Gaussian RoI to constrain the editing area. In particular, we randomly sample viewpoints from the 3D scene and render 2D image $\\mathcal{I}_{\\text {render }}$ . After that, $\\mathcal{I}_{\\text {render }}$ and noise level $t$ are put into 2D diffusion model $\\mathcal{M}_{DM}$ , with the user instruction $\\mathcal{T}$ and image $\\mathcal{I}_{\\text {input }}$ of input scene as conditions, to get edited image $\\mathcal{I}_{\\text {edit }}$ :", + "bbox": [ + 498, + 205, + 890, + 311 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {I} _ {\\text {e d i t}} = \\mathcal {D} \\left(\\mathcal {I} _ {\\text {r e n d e r}}, t; \\mathcal {T}, \\mathcal {I} _ {\\text {i n p u t}}\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 323, + 890, + 339 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $t$ is a randomly chosen noise level from $[t_{min}, t_{max}]$ .", + "bbox": [ + 500, + 349, + 890, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similar to 3D-GS [18], we apply the $\\mathcal{L}_{\\perp}$ and D-SSIM loss functions during editing.", + "bbox": [ + 500, + 364, + 890, + 393 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = (1 - \\beta) \\mathcal {L} _ {\\mathbb {1}} + \\beta \\mathcal {L} _ {D - S S I M}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 406, + 890, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the two losses are calculated between the 2D edited image $\\mathcal{I}_{ed}$ and the rendered image $\\mathcal{I}_{rd}$ . Then, gradient backpropagation is performed within Gaussian RoI $\\mathcal{G}_{RoI}$ :", + "bbox": [ + 498, + 431, + 890, + 478 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla \\mathcal {G} = \\frac {\\partial \\mathcal {L}}{\\partial \\mathcal {G}} \\cdot \\mathcal {R}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 642, + 487, + 890, + 517 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{R}$ is the set of RoI attributes. That means only Gaussians in RoI can receive gradients. Finally, we utilize the Adam algorithm to optimize the 3D Gaussians. After many rounds of training, the edited 3D scene $\\mathcal{G}_{edit}$ is obtained.", + "bbox": [ + 498, + 526, + 890, + 587 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 599, + 632, + 617 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 625, + 715, + 640 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our method is implemented in PyTorch [35] and CUDA, based on 3D Gaussian splatting. The multimodal model we used in our method is BLIP2 [21], and we use GPT-3.5 Turbo to ground the text ROI. For grounding segmentation, We use the cascade strategy, i.e. first using Grounding DINO [26] to get the box on the image corresponding to the text, and then using SAM [19] to get the corresponding image RoI. The 2D diffusion model used in our method is Instruct Pix2Pix [4]. We leave more details in the Appendix.", + "bbox": [ + 498, + 647, + 890, + 784 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Qualitative Evaluation", + "text_level": 1, + "bbox": [ + 500, + 792, + 709, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Visualization Results. In Fig. 1 and Fig. 4, we present the visual results of GaussianEditor, demonstrating the precise editing effects while ensuring 3D consistency. Fig. 1 shows the editing capabilities for characters. The first column displays the original scenes. In the second column, the first row \"Give him a red nose\" illustrates", + "bbox": [ + 498, + 810, + 890, + 898 + ], + "page_idx": 4 + }, + { + "type": "footer", + "text": "20906", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5f22c4a02c7b65486ef4fb653e956cefd86e95e24b3b25abc09a7a5e82bd0d59.jpg", + "image_caption": [ + "Original" + ], + "image_footnote": [], + "bbox": [ + 84, + 89, + 357, + 224 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dde0ef5cc26e127ed167a336f9adbd5ad7ecb5e0379614f90378c982d759ecbe.jpg", + "image_caption": [ + "\"Make the road look like the meadow\"" + ], + "image_footnote": [], + "bbox": [ + 359, + 92, + 616, + 224 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/16a796c5d293519271f3400bda7ee90a8574d5271b81cd14066c9fdb97a43998.jpg", + "image_caption": [ + "\"Turn the road into river\"" + ], + "image_footnote": [], + "bbox": [ + 617, + 92, + 874, + 224 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/010f77f08f1bd6536a5983de22cb13cc1864a0994c00b63f478036ce138fa462.jpg", + "image_caption": [ + "\"Turn the bench red\"" + ], + "image_footnote": [], + "bbox": [ + 86, + 237, + 357, + 372 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3d66bb5ecd3784d4ae29acb40369765b4acae0537e1aa5ded42e9096f2c030ef.jpg", + "image_caption": [ + "\"Turn the bench into wood\"" + ], + "image_footnote": [], + "bbox": [ + 359, + 238, + 616, + 372 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/26112cada5024cf6b88eef66b41d5c8f9558a08ebfc7e4de524289e5afeeeee3.jpg", + "image_caption": [ + "\"Turn the bench into Turquoise\"" + ], + "image_footnote": [], + "bbox": [ + 617, + 238, + 874, + 371 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5b5009c5cf6088020e9801714120bad0c1c87f59087fdd087e1b95e2ec78238e.jpg", + "image_caption": [ + "Original", + "Figure 4. Qualitative results on outdoor scenes. Our method supports separate foreground and background editing in real-world scenes." + ], + "image_footnote": [], + "bbox": [ + 96, + 387, + 290, + 500 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b1058324351dc5258469e00cc9b25eef929993a41e0a07440c813b625881bc48.jpg", + "image_caption": [ + "\"Turn the bear into a polar bear\"" + ], + "image_footnote": [], + "bbox": [ + 292, + 388, + 483, + 500 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9cfc433aa299af5c76cfa9c52f2973827c84038a5ba4ce6b63a0ded4f66cfe2f.jpg", + "image_caption": [ + "\"Turn the bear into a panda\"" + ], + "image_footnote": [], + "bbox": [ + 485, + 388, + 676, + 500 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c78ef437ca4fbccac2f8d4d19a9efbe17cf57011f3eca3f4e9ea752b8f25adfc.jpg", + "image_caption": [ + "\"Turn the bear into a grizzly bear\"" + ], + "image_footnote": [], + "bbox": [ + 679, + 388, + 870, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "color-changing ability, while the third row, \"Make him completely bald\", showcases capabilities of retexturing and slight geometry editing. The second row in the second column demonstrates precise editing ability by exclusively editing the left side of the face. Based on that, we achieve editing in the third column, focusing on the right side of the face, showcasing the ability of multi-round edits, and accurately fulfilling user instructions. Fig. 4 showcases the precise editing capabilities in open 3D scenes. In the upper portion, the bicycle scene allows us to accurately locate the position of the road and edit its texture, transforming it into the grass, a river. In the experiment where we change the texture to a river, our method accurately constructs the reflection, making it appear realistic. Based on editing the road into a river, we further edited the bench, proving that our method can achieve multiple rounds of editing. The lower portion demonstrates the results of editing the bear, which fully preserves the original appearance of the background area and focuses the edits on the bear.", + "bbox": [ + 76, + 534, + 468, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparisons with Instruct-NeRF2NeRF. Fig. 5 compares the results of our method with those of IN2N [11], on the scenes presented in IN2N. From the figure, it is evident that our method changes the texture of the pants without affecting the clothes, and vice versa, demonstrating", + "bbox": [ + 76, + 824, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the effectiveness of our method in distinguishing different objects within the foreground. Additionally, when editing the clothes and pants, the background remains unaffected, indicating our method's effective separation of foreground and background. Furthermore, the last column reveals that IN2N, limited by 2D diffusion, distorts the face, while our method maintains a superior rendering quality of faces.", + "bbox": [ + 496, + 534, + 890, + 641 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Complex Multi-Object Scenes. Furthermore, we present the results of our editing in a complex scene featuring multiple objects, as depicted in Fig 6. We specifically selected three distinct object types for editing purposes. The first object type is the background. In a particular sub-scene, the desktop shares similarities with the background. Employing a caption-based approach, we successfully transformed the desktop into a wooden material. As evident in the image, the edited result exhibits a distinct and pronounced wood texture. The second object type is a foreground object, the flowerpot. We opted to change the color of the flowerpot to red, and the outcome was highly successful. Lastly, the most intricate editing task involved the rolling pin, which was occluded by multiple objects from various perspectives. As illustrated in the lower right corner of the picture, we managed to edit it into a cucumber without any adverse impact on the other objects.", + "bbox": [ + 496, + 643, + 892, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20907", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/708aec49c9cafdd3fe8a679bbe324ff268d481f8fb6ad070ec7bf7239ee84e84.jpg", + "image_caption": [ + "IN2N" + ], + "image_footnote": [], + "bbox": [ + 80, + 90, + 151, + 239 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/034e98c8f141d54b5efd4f1d503b90feb7866de98e04acbcdeeda0ba4dd087cc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 176, + 90, + 321, + 239 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7a72df2a250f22c3c24084e86a09ea16191a5a3aab175c256eb9a0f3db5d0367.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 325, + 90, + 470, + 239 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/06bbde5350ffd895dbcc5ee14c1d5b5fd62bf26eaf71128c623ad62979a4033f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 473, + 90, + 617, + 239 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/66740c2206392f85d313bdc12fbe63c287529d8cd856b75f4b37a2002a4be59c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 622, + 90, + 888, + 239 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0c6c750adaf9a8ef3d823ff004ac4cbf87ddd3cf348b876acb87c895ec71422b.jpg", + "image_caption": [ + "Original", + "Ours" + ], + "image_footnote": [], + "bbox": [ + 80, + 241, + 151, + 386 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2b4c37b005bca97a16961b72f46b18b0e4f787d3ff5863caf159d08208a9040b.jpg", + "image_caption": [ + "\"Turn his pants into yellow.\"" + ], + "image_footnote": [], + "bbox": [ + 176, + 241, + 320, + 387 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/db5a6d9d9b23e1a617def74bbd5044e16fad1d6ac3ef06bcdc460ab648d50e26.jpg", + "image_caption": [ + "Turn his pants into a bronze statue." + ], + "image_footnote": [], + "bbox": [ + 323, + 241, + 470, + 387 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/89f0d60fa75b90e235ec080c6df3f97077da368e655a292c816379632ec0975f.jpg", + "image_caption": [ + "\"Turn his t-shirt into white.\"" + ], + "image_footnote": [], + "bbox": [ + 473, + 241, + 617, + 387 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9452d962b82a94bf81a3ddd21526532d1f6813bcb7e144e52a288018fbb825e8.jpg", + "image_caption": [ + "'Make his t-shirt out of leather" + ], + "image_footnote": [], + "bbox": [ + 624, + 241, + 888, + 387 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2b5812dd0e21bdfb7a1158c73e58d03e6c059d64c65171c905f4568749e01f5e.jpg", + "image_caption": [ + "Figure 5. Comparisons with Instruct-NeRF2NeRF (IN2N) [11] on the scene presented in their paper.", + "Original", + "Figure 6. Qualitative results on complex multi-object scenes. The background \"desk\", the foreground \"flower pot\", and the multiview blocked foreground \"rolling pin\" are edited separately." + ], + "image_footnote": [], + "bbox": [ + 81, + 424, + 284, + 527 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7d785d265676c476a31937017422b796380e02da12b72a22bef1e420abbe5ab6.jpg", + "image_caption": [ + "\"Turn the desk into wood\"" + ], + "image_footnote": [], + "bbox": [ + 285, + 424, + 478, + 527 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d203d78c3c3f4986abb31e1dced7ad8764bb2146549371e508739125abc1f1f7.jpg", + "image_caption": [ + "\"Turn the flowerpot red\"", + "\"Turn the rolling pin into cucumber\"" + ], + "image_footnote": [], + "bbox": [ + 480, + 424, + 679, + 527 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b7703d6d142d19a1d51628333f931145036b637db4645123ac1bfbae28d90ade.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 424, + 880, + 527 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/cc3a22842e7d83d77e4e98d5b2fb30be29fa8385e559d7193d303bfebc4109f0.jpg", + "table_caption": [ + "Table 1. Quantitative evaluation on the bicycle scene of the MipNeRF360 dataset [3]." + ], + "table_footnote": [], + "table_body": "
MethodsCTIDS↑IIS↑FID↓Time↓
IN2N [11]0.220.8510351 min
Ours-DVGO [45]0.110.8214840 min
Ours-3DGS0.280.955120 min
", + "bbox": [ + 78, + 614, + 464, + 689 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 76, + 702, + 297, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Metric Comparisons. Table 1 shows quantitative results on the bicycle scene of the Mip-NeRF360 dataset [3], comparing with IN2N [11] and Direct Voxel Grid Optimization (DVGO) [45] as the representation. The metrics include CLIP Text-Image Direction Similarity (CTIDS), Image-Image Similarity (IIS), FID, and training time. GaussianEditor achieves the best results in all metrics. The test data is shown in the supplementary material.", + "bbox": [ + 75, + 726, + 467, + 847 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "User Study. We perform a user study comparing with IN2N on the bear scene in Fig. 4 and the human scene in Fig. 5, involving 21 participants. GaussianEditor gets a", + "bbox": [ + 75, + 854, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "87.07% voting percentage, while IN2N gets 12.93%.", + "bbox": [ + 500, + 580, + 846, + 597 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study and Analysis", + "text_level": 1, + "bbox": [ + 500, + 604, + 754, + 621 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation of Gaussian RoI, Text RoI, RoI Lifting. To validate the effectiveness of each module in our framework, we design three variant approaches: (1) w/o Gaussian RoI: We discontinued the use of Gaussian RoI $\\mathcal{G}_{RoI}$ to control the gradients of Gaussian points, as mentioned in Eq. 14. (2) w/o Text ROI: In this scenario, we ceased the selection of text ROI $\\mathcal{T}_{RoI}$ using LLM assistant $\\mathcal{M}_{LLM}$ . Instead, all the words in the user's instruction are put to $\\mathcal{M}_{Seg}$ to get $\\mathcal{I}_{RoI}$ . (3) w/o RoI lifting: Instead of lifting the image RoI $\\mathcal{T}_{RoI}$ to 3D Gaussians, the image RoI $\\mathcal{T}_{RoI}$ is used to govern the calculation of the loss. That is, only the pixels within the image RoI $\\mathcal{T}_{RoI}$ are taken into account for the loss computation. Fig. 7 showcases the outcomes of our ablation experiment, which aimed to edit the doll based on the instruction \"Turn its mouth into red.\" The results reveal the following findings. (1) When the Gaussian RoI is not used, the 3D scene is all turned red because the 2D diffusion fails to control the editing area. (2) In cases where", + "bbox": [ + 496, + 628, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20908", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2f3b4558d05a720cdccebd35f1b2fa8dc23fc0ace4f66a8d90c517a280250ea9.jpg", + "image_caption": [ + "Figure 7. Ablation experiment of RoI." + ], + "image_footnote": [], + "bbox": [ + 89, + 88, + 452, + 205 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "text ROI $\\mathcal{T}_{Roi}$ is not utilized, the grounding segmentation model tends to segment the entire foreground object, leading to the doll being entirely edited to red. (3) When RoI lifting $\\mathcal{M}_{Lift}$ is not employed, the doll's mouth is successfully turned red, but other facial areas are also affected. Because the grounding segmentation model may fail to parse specific views, noise exists on the image RoI $\\mathcal{I}_{Roi}$ . Consequently, leakage occurs during the editing process. Our proposed RoI lifting module effectively addresses this issue during training. In conclusion, our ablation experiment demonstrates the effectiveness of several RoI-related modules in our method.", + "bbox": [ + 75, + 226, + 472, + 406 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation of Scene Description Generation. We further conduct experiments to evaluate the role of scene description generation, employing three distinct experimental setups. The first one composes the user message $\\mathcal{T}_{user}$ , without employing scene description. The second method randomly samples a view and extracts the corresponding image's text description as the scene description. The third one represents the complete version of our approach. The test scene involves a park where a bike and a bench are positioned closely together. The editing instruction is \"Turn the thing next to the bike orange\". The obtained results are presented in Fig. 8. As shown in the image, when scene description is not employed, the LLM fails to acquire the Text ROI according to the user instructions, resulting in editing failure. The second one randomly samples images to obtain scene descriptions, resulting in incomplete descriptions and leading to an incorrect text ROI prediction by the LLM. Consequently, the final editing result turns the road into an orange color. In contrast, our method flawlessly executes the editing task. This success can be attributed to scene description generation, which obtains an accurate text description encompassing the relative positional relationship between the bicycle and the bench. This enables the LLM to analyze and determine the user's intention to edit the bench. Consequently, the desired color change of the bench is successfully implemented.", + "bbox": [ + 75, + 412, + 472, + 819 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Limitations", + "text_level": 1, + "bbox": [ + 76, + 830, + 204, + 845 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Although our framework has solved some problems inherited from the integrated sub-modules, e.g. noise in the results of grounding segmentation, there are still some prob", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4120a7d7f36014858efa3395d672fbe704f351f111af9ab379803e57e7f07d4e.jpg", + "image_caption": [ + "Figure 8. Ablation results about the scene description generation." + ], + "image_footnote": [], + "bbox": [ + 511, + 88, + 877, + 467 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "lems that the current system cannot completely avoid. In scene description generation, the descriptions from different views of the same object may differ from each other. When the differences are large enough, the LLM may misunderstand these descriptions as those from multiple objects. This issue does not affect the results in the current experiment, but we would like to optimize this in the future. In addition, our system cannot achieve good editing results in scenes where the grounding segmentation or diffusion model completely fails, such as drastic geometric editing.", + "bbox": [ + 496, + 493, + 890, + 645 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 662, + 617, + 678 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper proposes a systematic framework, named GaussianEditor, for text-guided delicate 3D scene editing. As we know, GaussianEditor is one of the first works to edit 3D Gaussians, taking advantage of the explicit property of 3D Gaussians and making it easy to control the editing area precisely. Several techniques are proposed to achieve delicate editing, including extracting instruction RoI from texts, aligning the RoI to 3D Gaussians, and editing the scene with the Gaussian RoI. GaussianEditor achieves notably more delicate editing results than IN2N [11] with much shorter training time (within 20 minutes v.s. 45 minutes - 2 hours). Noticing recent works [27, 49] have extended Gaussian splatting to dynamic scenes, we leave the delicate editing in dynamic scenes as future work.", + "bbox": [ + 496, + 689, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20909", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Chong Bao, Yinda Zhang, Bangbang Yang, Tianxing Fan, Zesong Yang, Hujun Bao, Guofeng Zhang, and Zhaopeng Cui. Sine: Semantic-driven image-based nerf editing with prior-guided editing field. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2", + "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, 2021. 2", + "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022. 2, 7", + "[4] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 2, 3, 5", + "[5] Jiazhong Cen, Zanwei Zhou, Jiemin Fang, Chen Yang, Wei Shen, Lingxi Xie, Dongsheng Jiang, Xiaopeng Zhang, and Qi Tian. Segment anything in 3d with nerfs. In NeurIPS, 2023. 5", + "[6] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In ECCV, 2022. 2", + "[7] Pei-Ze Chiang, Meng-Shiun Tsai, Hung-Yu Tseng, Wei-Sheng Lai, and Wei-Chen Chiu. Stylizing 3d scene via implicit representation and hypernetwork. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2022. 2", + "[8] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 2021. 2", + "[9] Shuangkang Fang, Yufeng Wang, Yi Yang, Yi Yang, Yi-Hsuan Tsai, Wenrui Ding, Ming-Hsuan Yang, and Shuchang Zhou. Text-driven editing of 3d scenes without retraining. Arxiv preprint arXiv:2309.04917, 2023. 2", + "[10] William Gao, Noam Aigerman, Thibault Groueix, Vova Kim, and Rana Hanocka. Textdeformer: Geometry manipulation using text guidance. In ACM SIGGRAPH 2023 Conference Proceedings, 2023. 2", + "[11] Ayaan Haque, Matthew Tancik, Alexei Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. In CVPR, 2023. 2, 3, 6, 7, 8", + "[12] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2", + "[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 2", + "[14] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. The Journal of Machine Learning Research, 2022. 2", + "[15] Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. Avatarclip: Zero-shot text" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "driven generation and animation of 3d avatars. arXiv preprint arXiv:2205.08535, 2022. 2", + "[16] Hsin-Ping Huang, Hung-Yu Tseng, Saurabh Saini, Maneesh Singh, and Ming-Hsuan Yang. Learning to stylize novel views. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 2", + "[17] Yi-Hua Huang, Yue He, Yu-Jie Yuan, Yu-Kun Lai, and Lin Gao. Stylizednerf: consistent 3d scene stylization as stylized nerf via 2d-3d mutual learning. In CVPR, 2022. 2", + "[18] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 2023. 2, 3, 5", + "[19] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 5", + "[20] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nef for editing via feature field distillation. Advances in Neural Information Processing Systems, 2022. 2", + "[21] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 5", + "[22] Yuan Li, Zhi-Hao Lin, David Forsyth, Jia-Bin Huang, and Shenlong Wang. Climatenerf: Physically-based neural rendering for extreme climate synthesis. arXiv e-prints, 2022. 2", + "[23] Yuhan Li, Yishun Dou, Yue Shi, Yu Lei, Xuanhong Chen, Yi Zhang, Peng Zhou, and Bingbing Ni. Focaldreamer: Text-driven 3d editing via focal-fusion assembly. arXiv preprint arXiv:2308.10608, 2023. 3", + "[24] Hao-Kang Liu, I Shen, Bing-Yu Chen, et al. Nerf-in: Free-form nerf inpainting with rgb-d priors. arXiv preprint arXiv:2206.04901, 2022.", + "[25] Steven Liu, Xiuming Zhang, Zhoutong Zhang, Richard Zhang, Jun-Yan Zhu, and Bryan Russell. Editing conditional radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, 2021. 2", + "[26] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 5", + "[27] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 8", + "[28] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2", + "[29] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 2021. 2" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20910", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Ashkan Mirzaei, Tristan Aumentado-Armstrong, Marcus A. Brubaker, Jonathan Kelly, Alex Levinshtein, Konstantinos G. Derpanis, and Igor Gilitschenski. Watch your steps: Local image and scene editing by text instructions. In arXiv preprint arXiv:2308.08947, 2023. 2, 3", + "[31] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 2022. 2", + "[32] Thu Nguyen-Phuoc, Feng Liu, and Lei Xiao. Snerf: stylized neural implicit representations for 3d scenes. arXiv preprint arXiv:2207.02363, 2022. 2", + "[33] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2", + "[34] Atsuhiro Noguchi, Xiao Sun, Stephen Lin, and Tatsuya Harada. Neural articulated radiance field. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 2", + "[35] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 2019. 5", + "[36] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2", + "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022. 2", + "[38] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 2", + "[39] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, 2022. 2", + "[40] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 2022. 2", + "[41] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2", + "[42] Sara Fridovich-Keil and Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 2", + "[43] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using" + ], + "bbox": [ + 78, + 90, + 470, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "nonequilibrium thermodynamics. In International conference on machine learning, 2015. 2", + "[44] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 2019. 2", + "[45] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, 2022. 2, 7", + "[46] Vadim Tschernezki, Iro Laina, Diane Larlus, and Andrea Vedaldi. Neural feature fusion fields: 3d distillation of self-supervised 2d image representations. In 2022 International Conference on 3D Vision (3DV), 2022. 2", + "[47] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2", + "[48] Can Wang, Ruixiang Jiang, Mengei Chai, Mingming He, Dongdong Chen, and Jing Liao. Nerf-art: Text-driven neural radiance fields stylization. TVCG, 2023. 2", + "[49] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Wang Xinggang. 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528, 2023. 8", + "[50] Qiling Wu, Jianchao Tan, and Kun Xu. Palettenerf: Palette-based color editing for nerfs. arXiv preprint arXiv:2212.12871, 2022. 2", + "[51] Jiale Xu, Xintao Wang, Yan-Pei Cao, Weihao Cheng, Ying Shan, and Shenghua Gao. Instructp2p: Learning to edit 3d point clouds with text instructions. arXiv preprint arXiv:2306.07154, 2023. 2", + "[52] Tianhan Xu and Tatsuya Harada. Deforming radiance fields with cages. In European Conference on Computer Vision, 2022.", + "[53] Bangbang Yang, Chong Bao, Junyi Zeng, Hujun Bao, Yinda Zhang, Zhaopeng Cui, and Guofeng Zhang. Neumesh: Learning disentangled neural mesh-based implicit field for geometry and texture editing. In European Conference on Computer Vision, 2022. 2", + "[54] Kai Zhang, Nick Kolkin, Sai Bi, Fujun Luan, Zexiang Xu, Eli Shechtman, and Noah Snavely. Arf: Artistic radiance fields. In European Conference on Computer Vision, 2022. 2", + "[55] Jingyu Zhuang, Chen Wang, Lingjie Liu, Liang Lin, and Guanbin Li. Dreameditor: Text-driven 3d scene editing with neural fields. arXiv preprint arXiv:2306.13455, 2023. 3" + ], + "bbox": [ + 501, + 92, + 890, + 743 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20911", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/GaussianEditor_ Editing 3D Gaussians Delicately with Text Instructions/22c39c85-c3be-4622-b7a4-f891bcbdf14e_model.json b/2024/GaussianEditor_ Editing 3D Gaussians Delicately with Text Instructions/22c39c85-c3be-4622-b7a4-f891bcbdf14e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ab54b5ef478cc077e718f9e8c10d0df64f0dc069 --- /dev/null +++ b/2024/GaussianEditor_ Editing 3D Gaussians Delicately with Text Instructions/22c39c85-c3be-4622-b7a4-f891bcbdf14e_model.json @@ -0,0 +1,2552 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.122, + 0.131, + 0.85, + 0.154 + ], + "angle": 0, + "content": "GaussianEditor: Editing 3D Gaussians Delicately with Text Instructions" + }, + { + "type": "text", + "bbox": [ + 0.194, + 0.18, + 0.777, + 0.216 + ], + "angle": 0, + "content": "Junjie Wang,\\(^{\\star}\\) Jiemin Fang\\(^{\\star}\\), Xiaopeng Zhang, Lingxi Xie, Qi Tian Huawei Inc." + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.219, + 0.836, + 0.235 + ], + "angle": 0, + "content": "{is.wangjunjie, jaminfong, zxphistory, 198808xc}@gmail.com tian.qil@huawei.com" + }, + { + "type": "image", + "bbox": [ + 0.158, + 0.248, + 0.79, + 0.579 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.585, + 0.895, + 0.614 + ], + "angle": 0, + "content": "Figure 1. We propose GaussianEditor, an interactive framework to achieve delicate 3D scene editing following text instructions. As shown in this figure, our method can precisely control the editing region and achieve multi-round editing." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.627, + 0.314, + 0.642 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.647, + 0.474, + 0.86 + ], + "angle": 0, + "content": "Recently, impressive results have been achieved in 3D scene editing with text instructions based on a 2D diffusion model. However, current diffusion models primarily generate images by predicting noise in the latent space, and the editing is usually applied to the whole image, which makes it challenging to perform delicate, especially localized, editing for 3D scenes. Inspired by recent 3D Gaussian splatting, we propose a systematic framework, named GaussianEditor, to edit 3D scenes delicately via 3D Gaussians with text instructions. Benefiting from the explicit property of 3D Gaussians, we design a series of techniques to achieve delicate editing. Specifically, we first extract the region of interest (RoI) corresponding to the text instruction, aligning it to 3D Gaussians. The Gaussian RoI is fur" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.628, + 0.895, + 0.734 + ], + "angle": 0, + "content": "ther used to control the editing process. Our framework can achieve more delicate and precise editing of 3D scenes than previous methods while enjoying much faster training speed, i.e. within 20 minutes on a single V100 GPU, more than twice as fast as Instruct-NeRF2NeRF (45 minutes - 2 hours)1. The project page is at GaussianEditor.github.io." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.763, + 0.634, + 0.78 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.789, + 0.895, + 0.867 + ], + "angle": 0, + "content": "Creating 3D assets has played a critical role in many applications and industries, e.g. movie/game production, artistic creation, AR, VR etc. However, this process is usually expensive and cumbersome, especially for traditional pipelines. Designers need to take a lot of labor and time to" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.213, + 0.888 + ], + "angle": 0, + "content": "*Equal contributions." + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.888, + 0.222, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.222, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.892, + 0.901 + ], + "angle": 0, + "content": "1The editing time varies in different scenes according to the scene structure complexity." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20902" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.182 + ], + "angle": 0, + "content": "finish each step, e.g. sketching, building structures, creating textures etc. One cheap and effective way of creating high-quality 3D assets is to start from an existing scene, capturing, modeling, and editing the scene and obtaining the wanted one. This approach can be also used for user-interactive entertainment applications." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.185, + 0.473, + 0.502 + ], + "angle": 0, + "content": "Neural radiance field methods [2, 3, 6, 29, 31, 42, 45] have shown great power in representing 3D scenes and synthesizing novel-view images. Past years have witnessed the rapid development of NeRF and its variants, from both quality and efficiency perspectives. Editing a pre-trained NeRF model becomes a promising way to edit 3D scenes. Represented by Instruct-NeRF2NeRF [11], researchers propose to use the image-conditioned 2D diffusion model, e.g. InstructPix2Pix [4], to edit 3D scenes simply with text instructions. Notable results have been achieved as real scenes can be changed following the text instruction. However, current 2D diffusion models face challenges in accurately localizing editing regions, which hinders the generation of finely edited scenes due to the change of unintended regions. Even though some works [30] propose to constrain the editing region on edited 2D images, the editing region is not accurately localized and hard to apply to the 3D representation. Besides, NeRF-based methods [9, 48] bear coupling effects between different spatial positions, e.g. different points are queried from the same MLP field (for implicit representations) or voxel vertices (for explicit representations)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.506, + 0.47, + 0.627 + ], + "angle": 0, + "content": "Recent 3D Gaussian Splatting [18] (3D-GS) has been a groundbreaking work in the radiance field, which is the first to achieve a real sense of real-time rendering while enjoying high rendering quality and training speed. Besides its efficiency, we further notice its natural explicit property. 3D-GS has a great advantage for editing tasks as each 3D Gaussian exists individually. Editing 3D scenes by directly manipulating 3D Gaussians with desired constraints is easy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Aiming at editing 3D scenes delicately, we propose to represent the scene with 3D Gaussians, which can be edited with text instructions, and name our method as GaussianEditor. GaussianEditor is divided into three main parts to achieve precise control for editing regions. The first is the region of interest (RoI) extraction from the given text instruction. The instruction may be complex or indirect while this module helps extract the keywords matching the RoI for editing. The second part aligns the extracted text RoI to the 3D Gaussian space through the image space, where a grounding segmentation module is applied. The last part is to edit the original 3D Gaussians delicately with constraints in the obtained 3D Gaussian RoI. With the above processes, the region for editing can be precisely localized simply from text instructions, which constrains the 3D Gaussian updating to obtain a delicately edited new 3D scene. Besides, we enable interfaces for users to introduce more exact instructions for more delicate editing, e.g. Gaussian point selecting" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.825, + 0.107 + ], + "angle": 0, + "content": "and 3D boxes for modifying the editing regions2." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.108, + 0.843, + 0.121 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.123, + 0.892, + 0.168 + ], + "angle": 0, + "content": "- As far as we know, our GaussianEditor is one of the first systematic methods to achieve delicate 3D scene editing based on 3D Gaussian splatting." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.169, + 0.892, + 0.258 + ], + "angle": 0, + "content": "- A series of techniques are designed and proposed to precisely localize the editing region of interest, which are aligned and applied to 3D Gaussians. Though some submodules are from existing works, we believe integrating these awesome techniques to work effectively is a valuable topic, which is what we focus on in this paper." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.259, + 0.892, + 0.319 + ], + "angle": 0, + "content": "- Our method achieves a series of more delicate editing results compared with the previous representative work Instruct-NeRF2NeRF [11] with much shorter training time (within 20 minutes v.s. 45 minutes - 2 hours)." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.123, + 0.892, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.334, + 0.642, + 0.349 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.359, + 0.892, + 0.691 + ], + "angle": 0, + "content": "2D Image Editing with Diffusion Models. By advancements in diffusion model technology [13, 43], numerous generative models [40] have attained remarkably impressive outcomes in image synthesis. Recent progressions in diffusion models have showcased their adeptness in fabricating lifelike images from arbitrary textual inputs [8, 14, 39, 41, 44]. Harnessing the robust semantic comprehension and image generation capabilities of foundational diffusion models, an escalating number of research explorations are currently employing diffusion models as a fundamental framework for implementing text-based image editing functionalities [33, 36, 37, 40]. Some of these methodologies necessitate the manual provision of captions for both the original and edited images [12], while others mandate specific scenario-based training for optimization [38]. These requisites have rendered it arduous for ordinary users to avail themselves of such techniques. Expanding upon this foundation, iP2P [4] introduces instruction-based capabilities to image editing, enabling users to simply input an image and appraise the model of the desired alterations. This user-friendly approach facilitates the democratization of image editing in a more accessible manner." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.699, + 0.892, + 0.866 + ], + "angle": 0, + "content": "3D Scene Editing of Radiance Fields. 3D Scene Editing of Radiance Fields has gained significant popularity as a recent research direction [1, 10, 15, 20, 22-25, 28, 34, 46, 47, 51-53]. The main objective of such methods is to enable the manipulation of both the geometry and appearance of 3D scene representations. However, editing such scenes poses inherent challenges due to the implicit nature of traditional NeRF representations, which lack precise localization capabilities. As a result, previous works have primarily focused on achieving global style transformations of 3D scenes [7, 16, 17, 32, 48, 50, 54]. While some efforts" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.892, + 0.901 + ], + "angle": 0, + "content": "2These additional instructions are applied to generate the man with two different edited half faces in Fig. 1." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20903" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.09, + 0.884, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.531, + 0.895, + 0.616 + ], + "angle": 0, + "content": "Figure 2. Our framework, named GaussianEditor, consists of three key steps. First, a module \\(\\mathcal{M}_{Desc}\\) is used to get the description of the input scene, which is put to an LLM assistant \\(\\mathcal{M}_{LLM}\\) with the text instruction \\(\\mathcal{T}\\) provided by the user to obtain the text RoI \\(\\mathcal{T}_{RoI}\\). Second, a grounding segmentation module \\(\\mathcal{M}_{Seg}\\) is used to convert \\(\\mathcal{T}_{RoI}\\) to image RoI \\(\\mathcal{I}_{RoI}\\), which is then lifted to 3D Gaussians RoI \\(\\mathcal{G}_{RoI}\\) by RoI lifting \\(\\mathcal{M}_{Lift}\\), where additional user instructions \\(\\mathcal{O}\\) can be incorporated. Third, following the user instruction \\(\\mathcal{T}\\), rendered image \\(\\mathcal{I}_{render}\\) from randomly chosen views is edited by a diffusion model \\(\\mathcal{M}_{DM}\\). The loss between \\(\\mathcal{I}_{render}\\) and edited one \\(\\mathcal{I}_{edit}\\) is calculated. Finally, gradient backpropagation and optimization are performed within the Gaussian RoI \\(\\mathcal{G}_{RoI}\\) to get the edited scene \\(\\mathcal{G}_{edit}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.473, + 0.9 + ], + "angle": 0, + "content": "have been made towards object-centric scene editing[55], keeping the background unchanged has been a persistent challenge. For example, a recently proposed work, Instruct-NeRF2NeRF [11], implements text instruction-controlled 3D scene editing capabilities, achieving excellent editing effects while maintaining user-friendliness. However, it relies on the editing effect of 2D images, which may cause global changes to the 3D scene. A subsequent work [30] attempts to compute the relevance map between edited and unedited images to localize the editing area. The relevance map may be unreliable when the 2D IP2P [4] model fails. Other efforts [23] attempt to rely on the 3D coordinates entered by the user to decide the editing area. The introduction of 3D Gaussians [18] has provided an opportunity to address this limitation. Its explicit 3D representation methodology enables accurate selection and manipulation of editing areas. By incorporating LLMs, the whole process can be more automated." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.628, + 0.591, + 0.643 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.654, + 0.892, + 0.82 + ], + "angle": 0, + "content": "In this section, we first review 3D representation methods in Sec. 3.1. Subsequently, in Sec. 3.2, we overview our proposed approach, which mainly includes three modules. Sec. 3.3 delves into the precise Region of Interest (RoI) extraction of text instructions, using scene description generation module \\( \\mathcal{M}_{Desc} \\) and LLM assistant \\( \\mathcal{M}_{LLM} \\). Sec. 3.4 introduces how to align the instruction RoI with 3D Gaussians, using grounding segmentation module \\( \\mathcal{M}_{Seg} \\) and RoI lifting module \\( \\mathcal{M}_{Lift} \\). Finally, Sec. 3.5 describes the delicate editing process within the obtained Gaussian RoI, using text to image diffusion model \\( \\mathcal{M}_{DM} \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.832, + 0.64, + 0.846 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.903 + ], + "angle": 0, + "content": "3D Gaussian Splatting. 3D Gaussian splatting [18] is a recent powerful 3D representation method. It represents the 3D scene with point-like 3D Gaussians \\(\\mathcal{G} = \\{g_1, g_2 \\dots g_N\\}\\)," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20904" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.211 + ], + "angle": 0, + "content": "where \\( g_{i} = \\{\\mu, \\Sigma, c, \\alpha\\} \\) and \\( i \\in \\{1, \\dots, N\\} \\). Among them, \\( \\mu \\in \\mathbb{R}^3 \\) is the position where the Gaussian centers, \\( \\Sigma \\in \\mathbb{R}^7 \\) denotes the 3D covariance matrix, \\( c \\in \\mathbb{R}^3 \\) is the RGB color and \\( \\alpha \\in \\mathbb{R}^1 \\) is the opacity. Benefitting from the compact representation of Gaussians and efficient differentiable rendering approach, 3D Gaussian splatting achieves real-time rendering with high quality. The splatting rendering process can be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.216, + 0.47, + 0.258 + ], + "angle": 0, + "content": "\\[\nC = \\sum_ {i \\in \\mathcal {N}} c _ {i} \\sigma_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\sigma_ {j}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.265, + 0.472, + 0.314 + ], + "angle": 0, + "content": "where \\(\\sigma_{i} = \\alpha_{i}e^{-\\frac{1}{2}(x_{i})^{T}\\Sigma^{-1}(x_{i})}\\) represents the influence of the Gaussian to the image pixel and \\(x_{i}\\) is the distance between the 3D point and the center of the i-th Gaussian." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.32, + 0.267, + 0.335 + ], + "angle": 0, + "content": "3.2. Overall Framework" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.343, + 0.47, + 0.419 + ], + "angle": 0, + "content": "Given a group of 3D Gaussians \\(\\mathcal{G}_{input}\\) for an input scene and a text instruction \\(\\mathcal{T}\\) for editing, our Gaussian editor \\(\\mathcal{E}\\) can edit the 3D Gaussians delicately into a new one, denoted as \\(\\mathcal{G}_{edit}\\), with the guidance of the instruction. The whole process can be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.198, + 0.426, + 0.469, + 0.442 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} _ {\\text {e d i t}} = \\mathcal {E} \\left(\\mathcal {G} _ {\\text {i n p u t}}, \\mathcal {T}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.448, + 0.47, + 0.584 + ], + "angle": 0, + "content": "Fig. 2 illustrates the overall framework of our approach, which consists of three main steps. First, the Region of Interest (RoI) is extracted from the text instruction. In this step, we employ a module named scene description generation \\(\\mathcal{M}_{\\text{Description}}\\) to get the description of the input scene. We then input the scene description \\(\\mathcal{T}_{\\text{scene}}\\) and text instruction \\(\\mathcal{T}\\) into a large language model assistant \\(\\mathcal{M}_{\\text{LLM}}\\) to determine where we should make edits in the scene. The output of this step is referred to as the instruction RoI \\(\\mathcal{T}_{\\text{RoI}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.585, + 0.47, + 0.689 + ], + "angle": 0, + "content": "The next step is the 3D Gaussian RoI alignment. We use a grounding segmentation module \\(\\mathcal{M}_{Seg}\\) to convert the RoI from text space, i.e. \\(\\mathcal{T}_{RoI}\\), to the image space, i.e. \\(\\mathcal{I}_{RoI}\\). Then the image RoI \\(\\mathcal{I}_{RoI}\\) is lifted to the RoI of 3D Gaussians \\(\\mathcal{G}_{RoI}\\) through RoI lifting module \\(\\mathcal{M}_{Lift}\\). The Gaussian RoI allows us to control the regions where edits will be applied precisely." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.901 + ], + "angle": 0, + "content": "The last step is delicate editing within the Gaussian RoI. In this step, we randomly sample the view to obtain the rendered image \\(\\mathcal{I}_{\\text {render }}\\). A 2D diffusion model \\(\\mathcal{M}_{DM}\\) is used to perform the editing process on the rendered image \\(\\mathcal{I}_{\\text {render }}\\), with the user instruction \\(\\mathcal{T}\\) and the image \\(\\mathcal{I}_{\\text {input }}\\) of input scene as conditions. The resulting edited image is denoted as \\(\\mathcal{I}_{\\text {edit }}\\). Subsequently, we calculate the loss between \\(\\mathcal{I}_{\\text {edit }}\\) and \\(\\mathcal{I}_{\\text {render }}\\) and make gradient backpropagation within \\(\\mathcal{G}_{\\text {RoI }}\\). This implies that only the regions specified by the RoI can receive corresponding gradients during the back-propagation process. Finally, optimization is executed based on these gradients. The final optimized scene representation \\(\\mathcal{G}_{\\text {edit }}\\) is obtained through several rounds of iterative optimization." + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.091, + 0.877, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.537, + 0.423, + 0.855, + 0.437 + ], + "angle": 0, + "content": "Figure 3. The process of obtaining scene description." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.445, + 0.8, + 0.46 + ], + "angle": 0, + "content": "3.3. RoI Extraction of Text Instruction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.468, + 0.892, + 0.576 + ], + "angle": 0, + "content": "The instruction RoI is extracted for the editing regions from both the input 3D scene \\(\\mathcal{G}_{input}\\) and the text instruction \\(\\mathcal{T}\\) provided by the user. To achieve this, we employ a multimodal model \\(\\mathcal{M}_{MM}\\) in conjunction with the large language model assistant \\(\\mathcal{M}_{LLM}\\). The first step is the scene description generation \\(\\mathcal{M}_{Desc}\\), which aims to get the scene description \\(\\mathcal{T}_{scene}\\) from 3D Gaussians \\(\\mathcal{G}_{input}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.583, + 0.891, + 0.6 + ], + "angle": 0, + "content": "\\[\n\\mathcal {T} _ {\\text {s c e n e}} = \\mathcal {M} _ {\\text {D e s c}} \\left(\\mathcal {G} _ {\\text {i n p u t}}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.608, + 0.892, + 0.7 + ], + "angle": 0, + "content": "The process of the scene description generation \\(\\mathcal{M}_{Desc}\\) is shown in Fig. 3. By leveraging the technique of differentiable splatting as shown in Eq. 1, a set of 2D image samples \\(\\{\\mathcal{I}_{sample}\\}\\) are generated and then inputted into a multimodal model \\(\\mathcal{M}_{MM}\\) to generate corresponding text descriptions \\(\\{\\mathcal{T}_{sample}\\}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.707, + 0.891, + 0.724 + ], + "angle": 0, + "content": "\\[\n\\mathcal {T} _ {\\text {s a m p l e}} = \\mathcal {M} _ {M M} \\left(\\mathcal {P} _ {M M}, \\mathcal {I} _ {\\text {s a m p l e}}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.732, + 0.892, + 0.838 + ], + "angle": 0, + "content": "where \\(\\mathcal{P}_{MM}\\) is a prompt, such as \"What is the content of the image\", for multimodal model \\(\\mathcal{M}_{MM}\\) to get precise description. Subsequently, these descriptions \\(\\{\\mathcal{T}_{\\text {sample }}\\}\\) are fed into a large language model \\(\\mathcal{M}_{LLM}\\), which is specifically instructed by a prompt \\(\\mathcal{P}_{\\text {merge }}\\) to merge descriptions of diverse views into one detailed scene description \\(\\mathcal{T}_{\\text {scene }}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.564, + 0.846, + 0.891, + 0.864 + ], + "angle": 0, + "content": "\\[\n\\mathcal {T} _ {\\text {s c e n e}} = \\mathcal {M} _ {L L M} \\left(\\mathcal {P} _ {\\text {m e r g e}}, \\left\\{\\mathcal {T} _ {\\text {s a m p l e}} \\right\\}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "After that, the scene description \\(\\mathcal{T}_{\\text{scene}}\\) and the user instruction \\(\\mathcal{T}\\) are combined with a predefined tem" + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20905" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.168 + ], + "angle": 0, + "content": "plate \\(\\mathcal{T}_{\\text{template}}\\): \"Text description: \\(\\mathcal{T}_{\\text{scene}}\\) Edit Instruction: \\(\\mathcal{T}\\) Answer:\" to form the user message \\(\\mathcal{T}_{\\text{user}} = \\mathcal{T}_{\\text{template}}(\\mathcal{T}_{\\text{scene}}, \\mathcal{T})\\). The LLM model \\(\\mathcal{M}_{\\text{LLM}}\\) is used to extract the instruction RoI \\(\\mathcal{T}_{\\text{RoI}}\\) from user message \\(\\mathcal{T}_{\\text{user}}\\) with a new prompt \\(\\mathcal{P}_{\\text{extract}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.179, + 0.469, + 0.195 + ], + "angle": 0, + "content": "\\[\n\\mathcal {T} _ {R o I} = \\mathcal {M} _ {L L M} \\left(\\mathcal {P} _ {\\text {e x t r a c t}}, \\mathcal {T} _ {\\text {u s e r}}\\right). \\tag {6}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.205, + 0.331, + 0.221 + ], + "angle": 0, + "content": "3.4. 3D Gaussian RoI Alignment" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.228, + 0.469, + 0.29 + ], + "angle": 0, + "content": "To confine the 3D editing region within the instruction RoI, 3D Gaussian RoI \\(\\mathcal{G}_{RoI}\\) is aligned with the text RoI \\(\\mathcal{T}_{RoI}\\). First, The RoI in the text space is transformed into the image space via a grounding segmentation module \\(\\mathcal{M}_{Seg}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.301, + 0.469, + 0.318 + ], + "angle": 0, + "content": "\\[\n\\mathcal {I} _ {R o I} = \\mathcal {M} _ {S e g} \\left(\\mathcal {I} _ {\\text {i n p u t}}, \\mathcal {T} _ {R o I}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.328, + 0.456, + 0.344 + ], + "angle": 0, + "content": "where \\(\\mathcal{I}_{input}\\) is rendered image of the input scene \\(\\mathcal{G}_{input}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.344, + 0.469, + 0.448 + ], + "angle": 0, + "content": "Then we lift the RoI \\(\\mathcal{I}_{RoI}\\) in the image space to 3D Gaussian \\(\\mathcal{G}_{RoI}\\) through training. To achieve this, an additional RoI attribute \\(r\\in \\mathbb{R}^1\\) was added to 3D Gaussian \\(g_{i} = \\{\\mu_{i},\\Sigma_{i},c_{i},\\alpha_{i},r_{i}\\}\\). \\(r\\) is initialized to 0, which means it is not in the Gaussians RoI, and 1 means it is inside the RoI. The set of \\(r\\) is denoted as \\(\\mathcal{R}\\in \\mathbb{R}^{\\mathcal{N},1}\\), where the \\(\\mathcal{N}\\) is the number of 3D Gaussians \\(\\mathcal{G}_{input}\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.45, + 0.469, + 0.48 + ], + "angle": 0, + "content": "Then the color \\( c_{i} \\) in Eq. 1 was rewritten with \\( r_{i} \\) to get the rendered RoI \\( \\mathcal{T}_{RoI}^{render} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.491, + 0.469, + 0.533 + ], + "angle": 0, + "content": "\\[\n\\mathcal {I} _ {R o I} ^ {\\text {r e n d e r}} = \\sum_ {i \\in \\mathcal {N}} r _ {i} \\sigma_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\sigma_ {j}). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.543, + 0.47, + 0.588 + ], + "angle": 0, + "content": "Taking inspiration from SA3D [5], to get the trained Gaussians RoI \\(\\mathcal{G}_{RoI}^{train}\\), we adopt a similar loss function to supervise the training process:" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.599, + 0.48, + 0.633 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p r o j} = \\lambda_ {1} \\sum \\left(\\mathcal {I} _ {R o I} ^ {\\text {r e n d e r}} \\cdot \\mathcal {I} _ {R o I}\\right) + \\lambda_ {2} \\sum \\left(\\left(1 - \\mathcal {I} _ {R o I} ^ {\\text {r e n d e r}}\\right) \\cdot \\mathcal {I} _ {R o I}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.634, + 0.469, + 0.741 + ], + "angle": 0, + "content": "where \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) are hyperparameters. The \\(r\\) in Eq. 8 is updated via \\(r\\gets r - \\eta \\frac{\\partial\\mathcal{L}_{proj}}{\\partial r}\\) with gradient descent, where \\(\\eta\\) denotes the learning rate. Eq. 9 encourages rendered RoI to cover the Image RoI and not exceed it. Additionally, the user can modify the trained Gaussian RoI \\(\\mathcal{G}_{RoI}^{train}\\) by giving added Gaussian RoI \\(\\mathcal{G}_{RoI}^{add}\\), deleted Gaussian RoI \\(\\mathcal{G}_{RoI}^{del}\\) and 3D box \\(\\mathcal{B}_{3D}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.751, + 0.469, + 0.77 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} _ {R o I} = \\left(\\mathcal {G} _ {R o I} ^ {\\text {t r a i n}} \\cup \\mathcal {G} _ {R o I} ^ {\\text {a d d}} - \\mathcal {G} _ {R o I} ^ {\\text {d e l}}\\right) \\cap \\mathcal {B} _ {3 D}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.779, + 0.469, + 0.902 + ], + "angle": 0, + "content": "\\(\\mathcal{G}_{RoI}^{add}\\) represents the 3D Gaussians user wants to edit, \\(\\mathcal{G}_{RoI}^{del}\\) means 3D Gaussians user wants to keep from editing, \\(\\mathcal{B}_{3D}\\) is the coordinates of 3D cuboid it limits RoI to inside the box. \\(\\mathcal{G}_{RoI}\\) is the aligned RoI with the text RoI. For example, when editing the left face of the man in Fig. 1, grounding segmentation failed to ground \"left face\", instead, it grounded the whole face. In this scenario, the user can use the interactive interface to set the right face as \\(\\mathcal{G}_{RoI}^{del}\\) or enter" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "the rectangular box where the left face is located as \\(\\mathcal{B}_{3D}\\). The lifting process \\(\\mathcal{M}_{\\text{Lift}}\\) can be represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.607, + 0.133, + 0.892, + 0.15 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} _ {R o I} = \\mathcal {M} _ {\\text {L i f t}} \\left(\\mathcal {I} _ {R o I}, \\mathcal {O}\\right), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.158, + 0.867, + 0.176 + ], + "angle": 0, + "content": "where \\(\\mathcal{O} = \\{\\mathcal{G}_{RoI}^{add},\\mathcal{G}_{RoI}^{del},\\mathcal{B}_{3D}\\}\\) is optional instructions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.183, + 0.822, + 0.199 + ], + "angle": 0, + "content": "3.5. Delicate Editing within Gaussian RoI" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.207, + 0.892, + 0.313 + ], + "angle": 0, + "content": "To achieve delicate editing in 3D scenes, we use the Gaussian RoI to constrain the editing area. In particular, we randomly sample viewpoints from the 3D scene and render 2D image \\(\\mathcal{I}_{\\text {render }}\\). After that, \\(\\mathcal{I}_{\\text {render }}\\) and noise level \\(t\\) are put into 2D diffusion model \\(\\mathcal{M}_{DM}\\), with the user instruction \\(\\mathcal{T}\\) and image \\(\\mathcal{I}_{\\text {input }}\\) of input scene as conditions, to get edited image \\(\\mathcal{I}_{\\text {edit }}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.324, + 0.892, + 0.34 + ], + "angle": 0, + "content": "\\[\n\\mathcal {I} _ {\\text {e d i t}} = \\mathcal {D} \\left(\\mathcal {I} _ {\\text {r e n d e r}}, t; \\mathcal {T}, \\mathcal {I} _ {\\text {i n p u t}}\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.351, + 0.891, + 0.366 + ], + "angle": 0, + "content": "where \\( t \\) is a randomly chosen noise level from \\( [t_{min}, t_{max}] \\)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.366, + 0.891, + 0.395 + ], + "angle": 0, + "content": "Similar to 3D-GS [18], we apply the \\(\\mathcal{L}_{\\perp}\\) and D-SSIM loss functions during editing." + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.407, + 0.891, + 0.423 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = (1 - \\beta) \\mathcal {L} _ {\\mathbb {1}} + \\beta \\mathcal {L} _ {D - S S I M}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.433, + 0.892, + 0.479 + ], + "angle": 0, + "content": "the two losses are calculated between the 2D edited image \\(\\mathcal{I}_{ed}\\) and the rendered image \\(\\mathcal{I}_{rd}\\). Then, gradient backpropagation is performed within Gaussian RoI \\(\\mathcal{G}_{RoI}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.643, + 0.488, + 0.891, + 0.518 + ], + "angle": 0, + "content": "\\[\n\\nabla \\mathcal {G} = \\frac {\\partial \\mathcal {L}}{\\partial \\mathcal {G}} \\cdot \\mathcal {R}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.527, + 0.892, + 0.588 + ], + "angle": 0, + "content": "where \\(\\mathcal{R}\\) is the set of RoI attributes. That means only Gaussians in RoI can receive gradients. Finally, we utilize the Adam algorithm to optimize the 3D Gaussians. After many rounds of training, the edited 3D scene \\(\\mathcal{G}_{edit}\\) is obtained." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.601, + 0.633, + 0.618 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.625, + 0.717, + 0.641 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.648, + 0.892, + 0.785 + ], + "angle": 0, + "content": "Our method is implemented in PyTorch [35] and CUDA, based on 3D Gaussian splatting. The multimodal model we used in our method is BLIP2 [21], and we use GPT-3.5 Turbo to ground the text ROI. For grounding segmentation, We use the cascade strategy, i.e. first using Grounding DINO [26] to get the box on the image corresponding to the text, and then using SAM [19] to get the corresponding image RoI. The 2D diffusion model used in our method is Instruct Pix2Pix [4]. We leave more details in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.793, + 0.71, + 0.809 + ], + "angle": 0, + "content": "4.2. Qualitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.811, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Visualization Results. In Fig. 1 and Fig. 4, we present the visual results of GaussianEditor, demonstrating the precise editing effects while ensuring 3D consistency. Fig. 1 shows the editing capabilities for characters. The first column displays the original scenes. In the second column, the first row \"Give him a red nose\" illustrates" + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20906" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.09, + 0.359, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.207, + 0.226, + 0.257, + 0.237 + ], + "angle": 0, + "content": "Original" + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.093, + 0.617, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.388, + 0.227, + 0.593, + 0.237 + ], + "angle": 0, + "content": "\"Make the road look like the meadow\"" + }, + { + "type": "image", + "bbox": [ + 0.618, + 0.093, + 0.875, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.678, + 0.227, + 0.818, + 0.237 + ], + "angle": 0, + "content": "\"Turn the road into river\"" + }, + { + "type": "image", + "bbox": [ + 0.087, + 0.238, + 0.359, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.176, + 0.374, + 0.288, + 0.385 + ], + "angle": 0, + "content": "\"Turn the bench red\"" + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.239, + 0.617, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.424, + 0.374, + 0.573, + 0.385 + ], + "angle": 0, + "content": "\"Turn the bench into wood\"" + }, + { + "type": "image", + "bbox": [ + 0.618, + 0.239, + 0.875, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.662, + 0.373, + 0.835, + 0.384 + ], + "angle": 0, + "content": "\"Turn the bench into Turquoise\"" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.388, + 0.291, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.169, + 0.5, + 0.218, + 0.512 + ], + "angle": 0, + "content": "Original" + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.389, + 0.484, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.296, + 0.5, + 0.475, + 0.512 + ], + "angle": 0, + "content": "\"Turn the bear into a polar bear\"" + }, + { + "type": "image", + "bbox": [ + 0.486, + 0.389, + 0.678, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.504, + 0.5, + 0.661, + 0.512 + ], + "angle": 0, + "content": "\"Turn the bear into a panda\"" + }, + { + "type": "image", + "bbox": [ + 0.681, + 0.389, + 0.872, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.683, + 0.5, + 0.869, + 0.512 + ], + "angle": 0, + "content": "\"Turn the bear into a grizzly bear\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.084, + 0.514, + 0.882, + 0.528 + ], + "angle": 0, + "content": "Figure 4. Qualitative results on outdoor scenes. Our method supports separate foreground and background editing in real-world scenes." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.535, + 0.47, + 0.822 + ], + "angle": 0, + "content": "color-changing ability, while the third row, \"Make him completely bald\", showcases capabilities of retexturing and slight geometry editing. The second row in the second column demonstrates precise editing ability by exclusively editing the left side of the face. Based on that, we achieve editing in the third column, focusing on the right side of the face, showcasing the ability of multi-round edits, and accurately fulfilling user instructions. Fig. 4 showcases the precise editing capabilities in open 3D scenes. In the upper portion, the bicycle scene allows us to accurately locate the position of the road and edit its texture, transforming it into the grass, a river. In the experiment where we change the texture to a river, our method accurately constructs the reflection, making it appear realistic. Based on editing the road into a river, we further edited the bench, proving that our method can achieve multiple rounds of editing. The lower portion demonstrates the results of editing the bear, which fully preserves the original appearance of the background area and focuses the edits on the bear." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Comparisons with Instruct-NeRF2NeRF. Fig. 5 compares the results of our method with those of IN2N [11], on the scenes presented in IN2N. From the figure, it is evident that our method changes the texture of the pants without affecting the clothes, and vice versa, demonstrating" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.892, + 0.642 + ], + "angle": 0, + "content": "the effectiveness of our method in distinguishing different objects within the foreground. Additionally, when editing the clothes and pants, the background remains unaffected, indicating our method's effective separation of foreground and background. Furthermore, the last column reveals that IN2N, limited by 2D diffusion, distorts the face, while our method maintains a superior rendering quality of faces." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Complex Multi-Object Scenes. Furthermore, we present the results of our editing in a complex scene featuring multiple objects, as depicted in Fig 6. We specifically selected three distinct object types for editing purposes. The first object type is the background. In a particular sub-scene, the desktop shares similarities with the background. Employing a caption-based approach, we successfully transformed the desktop into a wooden material. As evident in the image, the edited result exhibits a distinct and pronounced wood texture. The second object type is a foreground object, the flowerpot. We opted to change the color of the flowerpot to red, and the outcome was highly successful. Lastly, the most intricate editing task involved the rolling pin, which was occluded by multiple objects from various perspectives. As illustrated in the lower right corner of the picture, we managed to edit it into a cucumber without any adverse impact on the other objects." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20907" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.092, + 0.152, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.153, + 0.162, + 0.177, + 0.171 + ], + "angle": 0, + "content": "IN2N" + }, + { + "type": "image", + "bbox": [ + 0.178, + 0.092, + 0.323, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.326, + 0.092, + 0.471, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.475, + 0.092, + 0.619, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.624, + 0.092, + 0.89, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.242, + 0.152, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.388, + 0.136, + 0.397 + ], + "angle": 0, + "content": "Original" + }, + { + "type": "image_caption", + "bbox": [ + 0.153, + 0.31, + 0.175, + 0.318 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.178, + 0.242, + 0.321, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.191, + 0.389, + 0.303, + 0.397 + ], + "angle": 0, + "content": "\"Turn his pants into yellow.\"" + }, + { + "type": "image", + "bbox": [ + 0.325, + 0.242, + 0.471, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.325, + 0.389, + 0.473, + 0.397 + ], + "angle": 0, + "content": "Turn his pants into a bronze statue." + }, + { + "type": "image", + "bbox": [ + 0.475, + 0.242, + 0.619, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.492, + 0.389, + 0.605, + 0.397 + ], + "angle": 0, + "content": "\"Turn his t-shirt into white.\"" + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.242, + 0.89, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.633, + 0.389, + 0.763, + 0.397 + ], + "angle": 0, + "content": "'Make his t-shirt out of leather" + }, + { + "type": "image_caption", + "bbox": [ + 0.186, + 0.401, + 0.782, + 0.416 + ], + "angle": 0, + "content": "Figure 5. Comparisons with Instruct-NeRF2NeRF (IN2N) [11] on the scene presented in their paper." + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.425, + 0.285, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.158, + 0.531, + 0.211, + 0.543 + ], + "angle": 0, + "content": "Original" + }, + { + "type": "image", + "bbox": [ + 0.286, + 0.425, + 0.48, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.308, + 0.531, + 0.459, + 0.542 + ], + "angle": 0, + "content": "\"Turn the desk into wood\"" + }, + { + "type": "image", + "bbox": [ + 0.481, + 0.425, + 0.681, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.531, + 0.653, + 0.543 + ], + "angle": 0, + "content": "\"Turn the flowerpot red\"" + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.425, + 0.882, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.678, + 0.531, + 0.887, + 0.543 + ], + "angle": 0, + "content": "\"Turn the rolling pin into cucumber\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.547, + 0.89, + 0.574 + ], + "angle": 0, + "content": "Figure 6. Qualitative results on complex multi-object scenes. The background \"desk\", the foreground \"flower pot\", and the multiview blocked foreground \"rolling pin\" are edited separately." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.58, + 0.468, + 0.608 + ], + "angle": 0, + "content": "Table 1. Quantitative evaluation on the bicycle scene of the MipNeRF360 dataset [3]." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.615, + 0.465, + 0.69 + ], + "angle": 0, + "content": "
MethodsCTIDS↑IIS↑FID↓Time↓
IN2N [11]0.220.8510351 min
Ours-DVGO [45]0.110.8214840 min
Ours-3DGS0.280.955120 min
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.703, + 0.298, + 0.719 + ], + "angle": 0, + "content": "4.3. Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.727, + 0.468, + 0.848 + ], + "angle": 0, + "content": "Metric Comparisons. Table 1 shows quantitative results on the bicycle scene of the Mip-NeRF360 dataset [3], comparing with IN2N [11] and Direct Voxel Grid Optimization (DVGO) [45] as the representation. The metrics include CLIP Text-Image Direction Similarity (CTIDS), Image-Image Similarity (IIS), FID, and training time. GaussianEditor achieves the best results in all metrics. The test data is shown in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.902 + ], + "angle": 0, + "content": "User Study. We perform a user study comparing with IN2N on the bear scene in Fig. 4 and the human scene in Fig. 5, involving 21 participants. GaussianEditor gets a" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.582, + 0.848, + 0.598 + ], + "angle": 0, + "content": "87.07% voting percentage, while IN2N gets 12.93%." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.606, + 0.756, + 0.622 + ], + "angle": 0, + "content": "4.4. Ablation Study and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Ablation of Gaussian RoI, Text RoI, RoI Lifting. To validate the effectiveness of each module in our framework, we design three variant approaches: (1) w/o Gaussian RoI: We discontinued the use of Gaussian RoI \\(\\mathcal{G}_{RoI}\\) to control the gradients of Gaussian points, as mentioned in Eq. 14. (2) w/o Text ROI: In this scenario, we ceased the selection of text ROI \\(\\mathcal{T}_{RoI}\\) using LLM assistant \\(\\mathcal{M}_{LLM}\\). Instead, all the words in the user's instruction are put to \\(\\mathcal{M}_{Seg}\\) to get \\(\\mathcal{I}_{RoI}\\). (3) w/o RoI lifting: Instead of lifting the image RoI \\(\\mathcal{T}_{RoI}\\) to 3D Gaussians, the image RoI \\(\\mathcal{T}_{RoI}\\) is used to govern the calculation of the loss. That is, only the pixels within the image RoI \\(\\mathcal{T}_{RoI}\\) are taken into account for the loss computation. Fig. 7 showcases the outcomes of our ablation experiment, which aimed to edit the doll based on the instruction \"Turn its mouth into red.\" The results reveal the following findings. (1) When the Gaussian RoI is not used, the 3D scene is all turned red because the 2D diffusion fails to control the editing area. (2) In cases where" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20908" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.09, + 0.089, + 0.454, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.159, + 0.21, + 0.386, + 0.223 + ], + "angle": 0, + "content": "Figure 7. Ablation experiment of RoI." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.227, + 0.473, + 0.407 + ], + "angle": 0, + "content": "text ROI \\(\\mathcal{T}_{Roi}\\) is not utilized, the grounding segmentation model tends to segment the entire foreground object, leading to the doll being entirely edited to red. (3) When RoI lifting \\(\\mathcal{M}_{Lift}\\) is not employed, the doll's mouth is successfully turned red, but other facial areas are also affected. Because the grounding segmentation model may fail to parse specific views, noise exists on the image RoI \\(\\mathcal{I}_{Roi}\\). Consequently, leakage occurs during the editing process. Our proposed RoI lifting module effectively addresses this issue during training. In conclusion, our ablation experiment demonstrates the effectiveness of several RoI-related modules in our method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.413, + 0.473, + 0.82 + ], + "angle": 0, + "content": "Ablation of Scene Description Generation. We further conduct experiments to evaluate the role of scene description generation, employing three distinct experimental setups. The first one composes the user message \\( \\mathcal{T}_{user} \\), without employing scene description. The second method randomly samples a view and extracts the corresponding image's text description as the scene description. The third one represents the complete version of our approach. The test scene involves a park where a bike and a bench are positioned closely together. The editing instruction is \"Turn the thing next to the bike orange\". The obtained results are presented in Fig. 8. As shown in the image, when scene description is not employed, the LLM fails to acquire the Text ROI according to the user instructions, resulting in editing failure. The second one randomly samples images to obtain scene descriptions, resulting in incomplete descriptions and leading to an incorrect text ROI prediction by the LLM. Consequently, the final editing result turns the road into an orange color. In contrast, our method flawlessly executes the editing task. This success can be attributed to scene description generation, which obtains an accurate text description encompassing the relative positional relationship between the bicycle and the bench. This enables the LLM to analyze and determine the user's intention to edit the bench. Consequently, the desired color change of the bench is successfully implemented." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.205, + 0.846 + ], + "angle": 0, + "content": "4.5. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Although our framework has solved some problems inherited from the integrated sub-modules, e.g. noise in the results of grounding segmentation, there are still some prob" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.089, + 0.878, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.501, + 0.473, + 0.89, + 0.487 + ], + "angle": 0, + "content": "Figure 8. Ablation results about the scene description generation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.494, + 0.892, + 0.646 + ], + "angle": 0, + "content": "lems that the current system cannot completely avoid. In scene description generation, the descriptions from different views of the same object may differ from each other. When the differences are large enough, the LLM may misunderstand these descriptions as those from multiple objects. This issue does not affect the results in the current experiment, but we would like to optimize this in the future. In addition, our system cannot achieve good editing results in scenes where the grounding segmentation or diffusion model completely fails, such as drastic geometric editing." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.663, + 0.619, + 0.679 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This paper proposes a systematic framework, named GaussianEditor, for text-guided delicate 3D scene editing. As we know, GaussianEditor is one of the first works to edit 3D Gaussians, taking advantage of the explicit property of 3D Gaussians and making it easy to control the editing area precisely. Several techniques are proposed to achieve delicate editing, including extracting instruction RoI from texts, aligning the RoI to 3D Gaussians, and editing the scene with the Gaussian RoI. GaussianEditor achieves notably more delicate editing results than IN2N [11] with much shorter training time (within 20 minutes v.s. 45 minutes - 2 hours). Noticing recent works [27, 49] have extended Gaussian splatting to dynamic scenes, we leave the delicate editing in dynamic scenes as future work." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20909" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.197 + ], + "angle": 0, + "content": "[1] Chong Bao, Yinda Zhang, Bangbang Yang, Tianxing Fan, Zesong Yang, Hujun Bao, Guofeng Zhang, and Zhaopeng Cui. Sine: Semantic-driven image-based nerf editing with prior-guided editing field. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.472, + 0.256 + ], + "angle": 0, + "content": "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.258, + 0.471, + 0.299 + ], + "angle": 0, + "content": "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.3, + 0.47, + 0.342 + ], + "angle": 0, + "content": "[4] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.344, + 0.47, + 0.398 + ], + "angle": 0, + "content": "[5] Jiazhong Cen, Zanwei Zhou, Jiemin Fang, Chen Yang, Wei Shen, Lingxi Xie, Dongsheng Jiang, Xiaopeng Zhang, and Qi Tian. Segment anything in 3d with nerfs. In NeurIPS, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.401, + 0.47, + 0.441 + ], + "angle": 0, + "content": "[6] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.444, + 0.47, + 0.513 + ], + "angle": 0, + "content": "[7] Pei-Ze Chiang, Meng-Shiun Tsai, Hung-Yu Tseng, Wei-Sheng Lai, and Wei-Chen Chiu. Stylizing 3d scene via implicit representation and hypernetwork. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.515, + 0.47, + 0.557 + ], + "angle": 0, + "content": "[8] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.559, + 0.47, + 0.613 + ], + "angle": 0, + "content": "[9] Shuangkang Fang, Yufeng Wang, Yi Yang, Yi Yang, Yi-Hsuan Tsai, Wenrui Ding, Ming-Hsuan Yang, and Shuchang Zhou. Text-driven editing of 3d scenes without retraining. Arxiv preprint arXiv:2309.04917, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.615, + 0.47, + 0.671 + ], + "angle": 0, + "content": "[10] William Gao, Noam Aigerman, Thibault Groueix, Vova Kim, and Rana Hanocka. Textdeformer: Geometry manipulation using text guidance. In ACM SIGGRAPH 2023 Conference Proceedings, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.673, + 0.47, + 0.727 + ], + "angle": 0, + "content": "[11] Ayaan Haque, Matthew Tancik, Alexei Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. In CVPR, 2023. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.729, + 0.47, + 0.784 + ], + "angle": 0, + "content": "[12] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.786, + 0.47, + 0.814 + ], + "angle": 0, + "content": "[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.471, + 0.871 + ], + "angle": 0, + "content": "[14] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. The Journal of Machine Learning Research, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[15] Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. Avatarclip: Zero-shot text" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "driven generation and animation of 3d avatars. arXiv preprint arXiv:2205.08535, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[16] Hsin-Ping Huang, Hung-Yu Tseng, Saurabh Saini, Maneesh Singh, and Ming-Hsuan Yang. Learning to stylize novel views. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.892, + 0.22 + ], + "angle": 0, + "content": "[17] Yi-Hua Huang, Yue He, Yu-Jie Yuan, Yu-Kun Lai, and Lin Gao. Stylizednerf: consistent 3d scene stylization as stylized nerf via 2d-3d mutual learning. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.222, + 0.892, + 0.276 + ], + "angle": 0, + "content": "[18] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.278, + 0.892, + 0.333 + ], + "angle": 0, + "content": "[19] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.335, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[20] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nef for editing via feature field distillation. Advances in Neural Information Processing Systems, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.391, + 0.892, + 0.447 + ], + "angle": 0, + "content": "[21] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.449, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[22] Yuan Li, Zhi-Hao Lin, David Forsyth, Jia-Bin Huang, and Shenlong Wang. Climatenerf: Physically-based neural rendering for extreme climate synthesis. arXiv e-prints, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.505, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[23] Yuhan Li, Yishun Dou, Yue Shi, Yu Lei, Xuanhong Chen, Yi Zhang, Peng Zhou, and Bingbing Ni. Focaldreamer: Text-driven 3d editing via focal-fusion assembly. arXiv preprint arXiv:2308.10608, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.562, + 0.892, + 0.603 + ], + "angle": 0, + "content": "[24] Hao-Kang Liu, I Shen, Bing-Yu Chen, et al. Nerf-in: Free-form nerf inpainting with rgb-d priors. arXiv preprint arXiv:2206.04901, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.605, + 0.892, + 0.66 + ], + "angle": 0, + "content": "[25] Steven Liu, Xiuming Zhang, Zhoutong Zhang, Richard Zhang, Jun-Yan Zhu, and Bryan Russell. Editing conditional radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.661, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[26] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.732, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[27] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[28] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.845, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[29] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20910" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[30] Ashkan Mirzaei, Tristan Aumentado-Armstrong, Marcus A. Brubaker, Jonathan Kelly, Alex Levinshtein, Konstantinos G. Derpanis, and Igor Gilitschenski. Watch your steps: Local image and scene editing by text instructions. In arXiv preprint arXiv:2308.08947, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.47, + 0.218 + ], + "angle": 0, + "content": "[31] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.47, + 0.259 + ], + "angle": 0, + "content": "[32] Thu Nguyen-Phuoc, Feng Liu, and Lei Xiao. Snerf: stylized neural implicit representations for 3d scenes. arXiv preprint arXiv:2207.02363, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.26, + 0.47, + 0.328 + ], + "angle": 0, + "content": "[33] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.47, + 0.384 + ], + "angle": 0, + "content": "[34] Atsuhiro Noguchi, Xiao Sun, Stephen Lin, and Tatsuya Harada. Neural articulated radiance field. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.385, + 0.47, + 0.453 + ], + "angle": 0, + "content": "[35] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.47, + 0.509 + ], + "angle": 0, + "content": "[36] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.471, + 0.58 + ], + "angle": 0, + "content": "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.58, + 0.47, + 0.635 + ], + "angle": 0, + "content": "[38] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.636, + 0.47, + 0.692 + ], + "angle": 0, + "content": "[39] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.47, + 0.775 + ], + "angle": 0, + "content": "[40] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.47, + 0.831 + ], + "angle": 0, + "content": "[41] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.47, + 0.872 + ], + "angle": 0, + "content": "[42] Sara Fridovich-Keil and Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.872, + 0.47, + 0.903 + ], + "angle": 0, + "content": "[43] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "nonequilibrium thermodynamics. In International conference on machine learning, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.892, + 0.163 + ], + "angle": 0, + "content": "[44] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[45] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[46] Vadim Tschernezki, Iro Laina, Diane Larlus, and Andrea Vedaldi. Neural feature fusion fields: 3d distillation of self-supervised 2d image representations. In 2022 International Conference on 3D Vision (3DV), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[47] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[48] Can Wang, Ruixiang Jiang, Mengei Chai, Mingming He, Dongdong Chen, and Jing Liao. Nerf-art: Text-driven neural radiance fields stylization. TVCG, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.376, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[49] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Wang Xinggang. 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.433, + 0.892, + 0.474 + ], + "angle": 0, + "content": "[50] Qiling Wu, Jianchao Tan, and Kun Xu. Palettenerf: Palette-based color editing for nerfs. arXiv preprint arXiv:2212.12871, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[51] Jiale Xu, Xintao Wang, Yan-Pei Cao, Weihao Cheng, Ying Shan, and Shenghua Gao. Instructp2p: Learning to edit 3d point clouds with text instructions. arXiv preprint arXiv:2306.07154, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.532, + 0.892, + 0.573 + ], + "angle": 0, + "content": "[52] Tianhan Xu and Tatsuya Harada. Deforming radiance fields with cages. In European Conference on Computer Vision, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.575, + 0.892, + 0.645 + ], + "angle": 0, + "content": "[53] Bangbang Yang, Chong Bao, Junyi Zeng, Hujun Bao, Yinda Zhang, Zhaopeng Cui, and Guofeng Zhang. Neumesh: Learning disentangled neural mesh-based implicit field for geometry and texture editing. In European Conference on Computer Vision, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.645, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[54] Kai Zhang, Nick Kolkin, Sai Bi, Fujun Luan, Zexiang Xu, Eli Shechtman, and Noah Snavely. Arf: Artistic radiance fields. In European Conference on Computer Vision, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.702, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[55] Jingyu Zhuang, Chen Wang, Lingjie Liu, Liang Lin, and Guanbin Li. Dreameditor: Text-driven 3d scene editing with neural fields. arXiv preprint arXiv:2306.13455, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.744 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20911" + } + ] +] \ No newline at end of file diff --git a/2024/GaussianEditor_ Editing 3D Gaussians Delicately with Text Instructions/22c39c85-c3be-4622-b7a4-f891bcbdf14e_origin.pdf b/2024/GaussianEditor_ Editing 3D Gaussians Delicately with Text Instructions/22c39c85-c3be-4622-b7a4-f891bcbdf14e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9c65aa1c1e4453c24dbb48a81df3bd5f2ab637f4 --- /dev/null +++ b/2024/GaussianEditor_ Editing 3D Gaussians Delicately with Text Instructions/22c39c85-c3be-4622-b7a4-f891bcbdf14e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ccfbe96a3bcab3d8bcfb36aa939cad7ee4c77caddf12b7542b8d616791b7d770 +size 5819255 diff --git a/2024/GaussianEditor_ Editing 3D Gaussians Delicately with Text Instructions/full.md b/2024/GaussianEditor_ Editing 3D Gaussians Delicately with Text Instructions/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3ee195e646ce68a3ba2759c594ece16ef5d7d69e --- /dev/null +++ b/2024/GaussianEditor_ Editing 3D Gaussians Delicately with Text Instructions/full.md @@ -0,0 +1,359 @@ +# GaussianEditor: Editing 3D Gaussians Delicately with Text Instructions + +Junjie Wang, $^{\star}$ Jiemin Fang $^{\star}$ , Xiaopeng Zhang, Lingxi Xie, Qi Tian Huawei Inc. + +{is.wangjunjie, jaminfong, zxphistory, 198808xc}@gmail.com tian.qil@huawei.com + +![](images/82c2d2fb93de361a4093dd459c31c3ea3797437513a565caafe64e3ee9fd1981.jpg) +Figure 1. We propose GaussianEditor, an interactive framework to achieve delicate 3D scene editing following text instructions. As shown in this figure, our method can precisely control the editing region and achieve multi-round editing. + +# Abstract + +Recently, impressive results have been achieved in 3D scene editing with text instructions based on a 2D diffusion model. However, current diffusion models primarily generate images by predicting noise in the latent space, and the editing is usually applied to the whole image, which makes it challenging to perform delicate, especially localized, editing for 3D scenes. Inspired by recent 3D Gaussian splatting, we propose a systematic framework, named GaussianEditor, to edit 3D scenes delicately via 3D Gaussians with text instructions. Benefiting from the explicit property of 3D Gaussians, we design a series of techniques to achieve delicate editing. Specifically, we first extract the region of interest (RoI) corresponding to the text instruction, aligning it to 3D Gaussians. The Gaussian RoI is fur + +ther used to control the editing process. Our framework can achieve more delicate and precise editing of 3D scenes than previous methods while enjoying much faster training speed, i.e. within 20 minutes on a single V100 GPU, more than twice as fast as Instruct-NeRF2NeRF (45 minutes - 2 hours)1. The project page is at GaussianEditor.github.io. + +# 1. Introduction + +Creating 3D assets has played a critical role in many applications and industries, e.g. movie/game production, artistic creation, AR, VR etc. However, this process is usually expensive and cumbersome, especially for traditional pipelines. Designers need to take a lot of labor and time to + +finish each step, e.g. sketching, building structures, creating textures etc. One cheap and effective way of creating high-quality 3D assets is to start from an existing scene, capturing, modeling, and editing the scene and obtaining the wanted one. This approach can be also used for user-interactive entertainment applications. + +Neural radiance field methods [2, 3, 6, 29, 31, 42, 45] have shown great power in representing 3D scenes and synthesizing novel-view images. Past years have witnessed the rapid development of NeRF and its variants, from both quality and efficiency perspectives. Editing a pre-trained NeRF model becomes a promising way to edit 3D scenes. Represented by Instruct-NeRF2NeRF [11], researchers propose to use the image-conditioned 2D diffusion model, e.g. InstructPix2Pix [4], to edit 3D scenes simply with text instructions. Notable results have been achieved as real scenes can be changed following the text instruction. However, current 2D diffusion models face challenges in accurately localizing editing regions, which hinders the generation of finely edited scenes due to the change of unintended regions. Even though some works [30] propose to constrain the editing region on edited 2D images, the editing region is not accurately localized and hard to apply to the 3D representation. Besides, NeRF-based methods [9, 48] bear coupling effects between different spatial positions, e.g. different points are queried from the same MLP field (for implicit representations) or voxel vertices (for explicit representations). + +Recent 3D Gaussian Splatting [18] (3D-GS) has been a groundbreaking work in the radiance field, which is the first to achieve a real sense of real-time rendering while enjoying high rendering quality and training speed. Besides its efficiency, we further notice its natural explicit property. 3D-GS has a great advantage for editing tasks as each 3D Gaussian exists individually. Editing 3D scenes by directly manipulating 3D Gaussians with desired constraints is easy. + +Aiming at editing 3D scenes delicately, we propose to represent the scene with 3D Gaussians, which can be edited with text instructions, and name our method as GaussianEditor. GaussianEditor is divided into three main parts to achieve precise control for editing regions. The first is the region of interest (RoI) extraction from the given text instruction. The instruction may be complex or indirect while this module helps extract the keywords matching the RoI for editing. The second part aligns the extracted text RoI to the 3D Gaussian space through the image space, where a grounding segmentation module is applied. The last part is to edit the original 3D Gaussians delicately with constraints in the obtained 3D Gaussian RoI. With the above processes, the region for editing can be precisely localized simply from text instructions, which constrains the 3D Gaussian updating to obtain a delicately edited new 3D scene. Besides, we enable interfaces for users to introduce more exact instructions for more delicate editing, e.g. Gaussian point selecting + +and 3D boxes for modifying the editing regions2. + +Our contributions can be summarized as follows. + +- As far as we know, our GaussianEditor is one of the first systematic methods to achieve delicate 3D scene editing based on 3D Gaussian splatting. +- A series of techniques are designed and proposed to precisely localize the editing region of interest, which are aligned and applied to 3D Gaussians. Though some submodules are from existing works, we believe integrating these awesome techniques to work effectively is a valuable topic, which is what we focus on in this paper. +- Our method achieves a series of more delicate editing results compared with the previous representative work Instruct-NeRF2NeRF [11] with much shorter training time (within 20 minutes v.s. 45 minutes - 2 hours). + +# 2. Related Work + +2D Image Editing with Diffusion Models. By advancements in diffusion model technology [13, 43], numerous generative models [40] have attained remarkably impressive outcomes in image synthesis. Recent progressions in diffusion models have showcased their adeptness in fabricating lifelike images from arbitrary textual inputs [8, 14, 39, 41, 44]. Harnessing the robust semantic comprehension and image generation capabilities of foundational diffusion models, an escalating number of research explorations are currently employing diffusion models as a fundamental framework for implementing text-based image editing functionalities [33, 36, 37, 40]. Some of these methodologies necessitate the manual provision of captions for both the original and edited images [12], while others mandate specific scenario-based training for optimization [38]. These requisites have rendered it arduous for ordinary users to avail themselves of such techniques. Expanding upon this foundation, iP2P [4] introduces instruction-based capabilities to image editing, enabling users to simply input an image and appraise the model of the desired alterations. This user-friendly approach facilitates the democratization of image editing in a more accessible manner. + +3D Scene Editing of Radiance Fields. 3D Scene Editing of Radiance Fields has gained significant popularity as a recent research direction [1, 10, 15, 20, 22-25, 28, 34, 46, 47, 51-53]. The main objective of such methods is to enable the manipulation of both the geometry and appearance of 3D scene representations. However, editing such scenes poses inherent challenges due to the implicit nature of traditional NeRF representations, which lack precise localization capabilities. As a result, previous works have primarily focused on achieving global style transformations of 3D scenes [7, 16, 17, 32, 48, 50, 54]. While some efforts + +![](images/45ed546d67486b7683a68eef9a6eb990c955081fec28ed5a0c96e682cea5dcfa.jpg) +Figure 2. Our framework, named GaussianEditor, consists of three key steps. First, a module $\mathcal{M}_{Desc}$ is used to get the description of the input scene, which is put to an LLM assistant $\mathcal{M}_{LLM}$ with the text instruction $\mathcal{T}$ provided by the user to obtain the text RoI $\mathcal{T}_{RoI}$ . Second, a grounding segmentation module $\mathcal{M}_{Seg}$ is used to convert $\mathcal{T}_{RoI}$ to image RoI $\mathcal{I}_{RoI}$ , which is then lifted to 3D Gaussians RoI $\mathcal{G}_{RoI}$ by RoI lifting $\mathcal{M}_{Lift}$ , where additional user instructions $\mathcal{O}$ can be incorporated. Third, following the user instruction $\mathcal{T}$ , rendered image $\mathcal{I}_{render}$ from randomly chosen views is edited by a diffusion model $\mathcal{M}_{DM}$ . The loss between $\mathcal{I}_{render}$ and edited one $\mathcal{I}_{edit}$ is calculated. Finally, gradient backpropagation and optimization are performed within the Gaussian RoI $\mathcal{G}_{RoI}$ to get the edited scene $\mathcal{G}_{edit}$ . + +have been made towards object-centric scene editing[55], keeping the background unchanged has been a persistent challenge. For example, a recently proposed work, Instruct-NeRF2NeRF [11], implements text instruction-controlled 3D scene editing capabilities, achieving excellent editing effects while maintaining user-friendliness. However, it relies on the editing effect of 2D images, which may cause global changes to the 3D scene. A subsequent work [30] attempts to compute the relevance map between edited and unedited images to localize the editing area. The relevance map may be unreliable when the 2D IP2P [4] model fails. Other efforts [23] attempt to rely on the 3D coordinates entered by the user to decide the editing area. The introduction of 3D Gaussians [18] has provided an opportunity to address this limitation. Its explicit 3D representation methodology enables accurate selection and manipulation of editing areas. By incorporating LLMs, the whole process can be more automated. + +# 3. Method + +In this section, we first review 3D representation methods in Sec. 3.1. Subsequently, in Sec. 3.2, we overview our proposed approach, which mainly includes three modules. Sec. 3.3 delves into the precise Region of Interest (RoI) extraction of text instructions, using scene description generation module $\mathcal{M}_{Desc}$ and LLM assistant $\mathcal{M}_{LLM}$ . Sec. 3.4 introduces how to align the instruction RoI with 3D Gaussians, using grounding segmentation module $\mathcal{M}_{Seg}$ and RoI lifting module $\mathcal{M}_{Lift}$ . Finally, Sec. 3.5 describes the delicate editing process within the obtained Gaussian RoI, using text to image diffusion model $\mathcal{M}_{DM}$ . + +# 3.1. Preliminaries + +3D Gaussian Splatting. 3D Gaussian splatting [18] is a recent powerful 3D representation method. It represents the 3D scene with point-like 3D Gaussians $\mathcal{G} = \{g_1, g_2 \dots g_N\}$ , + +where $g_{i} = \{\mu, \Sigma, c, \alpha\}$ and $i \in \{1, \dots, N\}$ . Among them, $\mu \in \mathbb{R}^3$ is the position where the Gaussian centers, $\Sigma \in \mathbb{R}^7$ denotes the 3D covariance matrix, $c \in \mathbb{R}^3$ is the RGB color and $\alpha \in \mathbb{R}^1$ is the opacity. Benefitting from the compact representation of Gaussians and efficient differentiable rendering approach, 3D Gaussian splatting achieves real-time rendering with high quality. The splatting rendering process can be formulated as + +$$ +C = \sum_ {i \in \mathcal {N}} c _ {i} \sigma_ {i} \prod_ {j = 1} ^ {i - 1} (1 - \sigma_ {j}), \tag {1} +$$ + +where $\sigma_{i} = \alpha_{i}e^{-\frac{1}{2}(x_{i})^{T}\Sigma^{-1}(x_{i})}$ represents the influence of the Gaussian to the image pixel and $x_{i}$ is the distance between the 3D point and the center of the i-th Gaussian. + +# 3.2. Overall Framework + +Given a group of 3D Gaussians $\mathcal{G}_{input}$ for an input scene and a text instruction $\mathcal{T}$ for editing, our Gaussian editor $\mathcal{E}$ can edit the 3D Gaussians delicately into a new one, denoted as $\mathcal{G}_{edit}$ , with the guidance of the instruction. The whole process can be formulated as + +$$ +\mathcal {G} _ {\text {e d i t}} = \mathcal {E} \left(\mathcal {G} _ {\text {i n p u t}}, \mathcal {T}\right). \tag {2} +$$ + +Fig. 2 illustrates the overall framework of our approach, which consists of three main steps. First, the Region of Interest (RoI) is extracted from the text instruction. In this step, we employ a module named scene description generation $\mathcal{M}_{\text{Description}}$ to get the description of the input scene. We then input the scene description $\mathcal{T}_{\text{scene}}$ and text instruction $\mathcal{T}$ into a large language model assistant $\mathcal{M}_{\text{LLM}}$ to determine where we should make edits in the scene. The output of this step is referred to as the instruction RoI $\mathcal{T}_{\text{RoI}}$ . + +The next step is the 3D Gaussian RoI alignment. We use a grounding segmentation module $\mathcal{M}_{Seg}$ to convert the RoI from text space, i.e. $\mathcal{T}_{RoI}$ , to the image space, i.e. $\mathcal{I}_{RoI}$ . Then the image RoI $\mathcal{I}_{RoI}$ is lifted to the RoI of 3D Gaussians $\mathcal{G}_{RoI}$ through RoI lifting module $\mathcal{M}_{Lift}$ . The Gaussian RoI allows us to control the regions where edits will be applied precisely. + +The last step is delicate editing within the Gaussian RoI. In this step, we randomly sample the view to obtain the rendered image $\mathcal{I}_{\text {render }}$ . A 2D diffusion model $\mathcal{M}_{DM}$ is used to perform the editing process on the rendered image $\mathcal{I}_{\text {render }}$ , with the user instruction $\mathcal{T}$ and the image $\mathcal{I}_{\text {input }}$ of input scene as conditions. The resulting edited image is denoted as $\mathcal{I}_{\text {edit }}$ . Subsequently, we calculate the loss between $\mathcal{I}_{\text {edit }}$ and $\mathcal{I}_{\text {render }}$ and make gradient backpropagation within $\mathcal{G}_{\text {RoI }}$ . This implies that only the regions specified by the RoI can receive corresponding gradients during the back-propagation process. Finally, optimization is executed based on these gradients. The final optimized scene representation $\mathcal{G}_{\text {edit }}$ is obtained through several rounds of iterative optimization. + +![](images/0fd46d91fccb7c66678e177ba3b61ae5b740bbdcddb305bf1811e5bdbf7f46f2.jpg) +Figure 3. The process of obtaining scene description. + +# 3.3. RoI Extraction of Text Instruction + +The instruction RoI is extracted for the editing regions from both the input 3D scene $\mathcal{G}_{input}$ and the text instruction $\mathcal{T}$ provided by the user. To achieve this, we employ a multimodal model $\mathcal{M}_{MM}$ in conjunction with the large language model assistant $\mathcal{M}_{LLM}$ . The first step is the scene description generation $\mathcal{M}_{Desc}$ , which aims to get the scene description $\mathcal{T}_{scene}$ from 3D Gaussians $\mathcal{G}_{input}$ : + +$$ +\mathcal {T} _ {\text {s c e n e}} = \mathcal {M} _ {\text {D e s c}} \left(\mathcal {G} _ {\text {i n p u t}}\right). \tag {3} +$$ + +The process of the scene description generation $\mathcal{M}_{Desc}$ is shown in Fig. 3. By leveraging the technique of differentiable splatting as shown in Eq. 1, a set of 2D image samples $\{\mathcal{I}_{sample}\}$ are generated and then inputted into a multimodal model $\mathcal{M}_{MM}$ to generate corresponding text descriptions $\{\mathcal{T}_{sample}\}$ : + +$$ +\mathcal {T} _ {\text {s a m p l e}} = \mathcal {M} _ {M M} \left(\mathcal {P} _ {M M}, \mathcal {I} _ {\text {s a m p l e}}\right), \tag {4} +$$ + +where $\mathcal{P}_{MM}$ is a prompt, such as "What is the content of the image", for multimodal model $\mathcal{M}_{MM}$ to get precise description. Subsequently, these descriptions $\{\mathcal{T}_{\text {sample }}\}$ are fed into a large language model $\mathcal{M}_{LLM}$ , which is specifically instructed by a prompt $\mathcal{P}_{\text {merge }}$ to merge descriptions of diverse views into one detailed scene description $\mathcal{T}_{\text {scene }}$ : + +$$ +\mathcal {T} _ {\text {s c e n e}} = \mathcal {M} _ {L L M} \left(\mathcal {P} _ {\text {m e r g e}}, \left\{\mathcal {T} _ {\text {s a m p l e}} \right\}\right). \tag {5} +$$ + +After that, the scene description $\mathcal{T}_{\text{scene}}$ and the user instruction $\mathcal{T}$ are combined with a predefined tem + +plate $\mathcal{T}_{\text{template}}$ : "Text description: $\mathcal{T}_{\text{scene}}$ Edit Instruction: $\mathcal{T}$ Answer:" to form the user message $\mathcal{T}_{\text{user}} = \mathcal{T}_{\text{template}}(\mathcal{T}_{\text{scene}}, \mathcal{T})$ . The LLM model $\mathcal{M}_{\text{LLM}}$ is used to extract the instruction RoI $\mathcal{T}_{\text{RoI}}$ from user message $\mathcal{T}_{\text{user}}$ with a new prompt $\mathcal{P}_{\text{extract}}$ : + +$$ +\mathcal {T} _ {R o I} = \mathcal {M} _ {L L M} \left(\mathcal {P} _ {\text {e x t r a c t}}, \mathcal {T} _ {\text {u s e r}}\right). \tag {6} +$$ + +# 3.4. 3D Gaussian RoI Alignment + +To confine the 3D editing region within the instruction RoI, 3D Gaussian RoI $\mathcal{G}_{RoI}$ is aligned with the text RoI $\mathcal{T}_{RoI}$ . First, The RoI in the text space is transformed into the image space via a grounding segmentation module $\mathcal{M}_{Seg}$ : + +$$ +\mathcal {I} _ {R o I} = \mathcal {M} _ {S e g} \left(\mathcal {I} _ {\text {i n p u t}}, \mathcal {T} _ {R o I}\right), \tag {7} +$$ + +where $\mathcal{I}_{input}$ is rendered image of the input scene $\mathcal{G}_{input}$ . + +Then we lift the RoI $\mathcal{I}_{RoI}$ in the image space to 3D Gaussian $\mathcal{G}_{RoI}$ through training. To achieve this, an additional RoI attribute $r\in \mathbb{R}^1$ was added to 3D Gaussian $g_{i} = \{\mu_{i},\Sigma_{i},c_{i},\alpha_{i},r_{i}\}$ . $r$ is initialized to 0, which means it is not in the Gaussians RoI, and 1 means it is inside the RoI. The set of $r$ is denoted as $\mathcal{R}\in \mathbb{R}^{\mathcal{N},1}$ , where the $\mathcal{N}$ is the number of 3D Gaussians $\mathcal{G}_{input}$ . + +Then the color $c_{i}$ in Eq. 1 was rewritten with $r_{i}$ to get the rendered RoI $\mathcal{T}_{RoI}^{render}$ : + +$$ +\mathcal {I} _ {R o I} ^ {\text {r e n d e r}} = \sum_ {i \in \mathcal {N}} r _ {i} \sigma_ {i} \prod_ {j = 1} ^ {i - 1} (1 - \sigma_ {j}). \tag {8} +$$ + +Taking inspiration from SA3D [5], to get the trained Gaussians RoI $\mathcal{G}_{RoI}^{train}$ , we adopt a similar loss function to supervise the training process: + +$$ +\mathcal {L} _ {p r o j} = \lambda_ {1} \sum \left(\mathcal {I} _ {R o I} ^ {\text {r e n d e r}} \cdot \mathcal {I} _ {R o I}\right) + \lambda_ {2} \sum \left(\left(1 - \mathcal {I} _ {R o I} ^ {\text {r e n d e r}}\right) \cdot \mathcal {I} _ {R o I}\right), \tag {9} +$$ + +where $\lambda_{1}$ and $\lambda_{2}$ are hyperparameters. The $r$ in Eq. 8 is updated via $r\gets r - \eta \frac{\partial\mathcal{L}_{proj}}{\partial r}$ with gradient descent, where $\eta$ denotes the learning rate. Eq. 9 encourages rendered RoI to cover the Image RoI and not exceed it. Additionally, the user can modify the trained Gaussian RoI $\mathcal{G}_{RoI}^{train}$ by giving added Gaussian RoI $\mathcal{G}_{RoI}^{add}$ , deleted Gaussian RoI $\mathcal{G}_{RoI}^{del}$ and 3D box $\mathcal{B}_{3D}$ : + +$$ +\mathcal {G} _ {R o I} = \left(\mathcal {G} _ {R o I} ^ {\text {t r a i n}} \cup \mathcal {G} _ {R o I} ^ {\text {a d d}} - \mathcal {G} _ {R o I} ^ {\text {d e l}}\right) \cap \mathcal {B} _ {3 D}, \tag {10} +$$ + +$\mathcal{G}_{RoI}^{add}$ represents the 3D Gaussians user wants to edit, $\mathcal{G}_{RoI}^{del}$ means 3D Gaussians user wants to keep from editing, $\mathcal{B}_{3D}$ is the coordinates of 3D cuboid it limits RoI to inside the box. $\mathcal{G}_{RoI}$ is the aligned RoI with the text RoI. For example, when editing the left face of the man in Fig. 1, grounding segmentation failed to ground "left face", instead, it grounded the whole face. In this scenario, the user can use the interactive interface to set the right face as $\mathcal{G}_{RoI}^{del}$ or enter + +the rectangular box where the left face is located as $\mathcal{B}_{3D}$ . The lifting process $\mathcal{M}_{\text{Lift}}$ can be represented as: + +$$ +\mathcal {G} _ {R o I} = \mathcal {M} _ {\text {L i f t}} \left(\mathcal {I} _ {R o I}, \mathcal {O}\right), \tag {11} +$$ + +where $\mathcal{O} = \{\mathcal{G}_{RoI}^{add},\mathcal{G}_{RoI}^{del},\mathcal{B}_{3D}\}$ is optional instructions. + +# 3.5. Delicate Editing within Gaussian RoI + +To achieve delicate editing in 3D scenes, we use the Gaussian RoI to constrain the editing area. In particular, we randomly sample viewpoints from the 3D scene and render 2D image $\mathcal{I}_{\text {render }}$ . After that, $\mathcal{I}_{\text {render }}$ and noise level $t$ are put into 2D diffusion model $\mathcal{M}_{DM}$ , with the user instruction $\mathcal{T}$ and image $\mathcal{I}_{\text {input }}$ of input scene as conditions, to get edited image $\mathcal{I}_{\text {edit }}$ : + +$$ +\mathcal {I} _ {\text {e d i t}} = \mathcal {D} \left(\mathcal {I} _ {\text {r e n d e r}}, t; \mathcal {T}, \mathcal {I} _ {\text {i n p u t}}\right), \tag {12} +$$ + +where $t$ is a randomly chosen noise level from $[t_{min}, t_{max}]$ . + +Similar to 3D-GS [18], we apply the $\mathcal{L}_{\perp}$ and D-SSIM loss functions during editing. + +$$ +\mathcal {L} = (1 - \beta) \mathcal {L} _ {\mathbb {1}} + \beta \mathcal {L} _ {D - S S I M}. \tag {13} +$$ + +the two losses are calculated between the 2D edited image $\mathcal{I}_{ed}$ and the rendered image $\mathcal{I}_{rd}$ . Then, gradient backpropagation is performed within Gaussian RoI $\mathcal{G}_{RoI}$ : + +$$ +\nabla \mathcal {G} = \frac {\partial \mathcal {L}}{\partial \mathcal {G}} \cdot \mathcal {R}, \tag {14} +$$ + +where $\mathcal{R}$ is the set of RoI attributes. That means only Gaussians in RoI can receive gradients. Finally, we utilize the Adam algorithm to optimize the 3D Gaussians. After many rounds of training, the edited 3D scene $\mathcal{G}_{edit}$ is obtained. + +# 4. Experiments + +# 4.1. Implementation Details + +Our method is implemented in PyTorch [35] and CUDA, based on 3D Gaussian splatting. The multimodal model we used in our method is BLIP2 [21], and we use GPT-3.5 Turbo to ground the text ROI. For grounding segmentation, We use the cascade strategy, i.e. first using Grounding DINO [26] to get the box on the image corresponding to the text, and then using SAM [19] to get the corresponding image RoI. The 2D diffusion model used in our method is Instruct Pix2Pix [4]. We leave more details in the Appendix. + +# 4.2. Qualitative Evaluation + +Visualization Results. In Fig. 1 and Fig. 4, we present the visual results of GaussianEditor, demonstrating the precise editing effects while ensuring 3D consistency. Fig. 1 shows the editing capabilities for characters. The first column displays the original scenes. In the second column, the first row "Give him a red nose" illustrates + +![](images/5f22c4a02c7b65486ef4fb653e956cefd86e95e24b3b25abc09a7a5e82bd0d59.jpg) +Original + +![](images/dde0ef5cc26e127ed167a336f9adbd5ad7ecb5e0379614f90378c982d759ecbe.jpg) +"Make the road look like the meadow" + +![](images/16a796c5d293519271f3400bda7ee90a8574d5271b81cd14066c9fdb97a43998.jpg) +"Turn the road into river" + +![](images/010f77f08f1bd6536a5983de22cb13cc1864a0994c00b63f478036ce138fa462.jpg) +"Turn the bench red" + +![](images/3d66bb5ecd3784d4ae29acb40369765b4acae0537e1aa5ded42e9096f2c030ef.jpg) +"Turn the bench into wood" + +![](images/26112cada5024cf6b88eef66b41d5c8f9558a08ebfc7e4de524289e5afeeeee3.jpg) +"Turn the bench into Turquoise" + +![](images/5b5009c5cf6088020e9801714120bad0c1c87f59087fdd087e1b95e2ec78238e.jpg) +Original +Figure 4. Qualitative results on outdoor scenes. Our method supports separate foreground and background editing in real-world scenes. + +![](images/b1058324351dc5258469e00cc9b25eef929993a41e0a07440c813b625881bc48.jpg) +"Turn the bear into a polar bear" + +![](images/9cfc433aa299af5c76cfa9c52f2973827c84038a5ba4ce6b63a0ded4f66cfe2f.jpg) +"Turn the bear into a panda" + +![](images/c78ef437ca4fbccac2f8d4d19a9efbe17cf57011f3eca3f4e9ea752b8f25adfc.jpg) +"Turn the bear into a grizzly bear" + +color-changing ability, while the third row, "Make him completely bald", showcases capabilities of retexturing and slight geometry editing. The second row in the second column demonstrates precise editing ability by exclusively editing the left side of the face. Based on that, we achieve editing in the third column, focusing on the right side of the face, showcasing the ability of multi-round edits, and accurately fulfilling user instructions. Fig. 4 showcases the precise editing capabilities in open 3D scenes. In the upper portion, the bicycle scene allows us to accurately locate the position of the road and edit its texture, transforming it into the grass, a river. In the experiment where we change the texture to a river, our method accurately constructs the reflection, making it appear realistic. Based on editing the road into a river, we further edited the bench, proving that our method can achieve multiple rounds of editing. The lower portion demonstrates the results of editing the bear, which fully preserves the original appearance of the background area and focuses the edits on the bear. + +Comparisons with Instruct-NeRF2NeRF. Fig. 5 compares the results of our method with those of IN2N [11], on the scenes presented in IN2N. From the figure, it is evident that our method changes the texture of the pants without affecting the clothes, and vice versa, demonstrating + +the effectiveness of our method in distinguishing different objects within the foreground. Additionally, when editing the clothes and pants, the background remains unaffected, indicating our method's effective separation of foreground and background. Furthermore, the last column reveals that IN2N, limited by 2D diffusion, distorts the face, while our method maintains a superior rendering quality of faces. + +Complex Multi-Object Scenes. Furthermore, we present the results of our editing in a complex scene featuring multiple objects, as depicted in Fig 6. We specifically selected three distinct object types for editing purposes. The first object type is the background. In a particular sub-scene, the desktop shares similarities with the background. Employing a caption-based approach, we successfully transformed the desktop into a wooden material. As evident in the image, the edited result exhibits a distinct and pronounced wood texture. The second object type is a foreground object, the flowerpot. We opted to change the color of the flowerpot to red, and the outcome was highly successful. Lastly, the most intricate editing task involved the rolling pin, which was occluded by multiple objects from various perspectives. As illustrated in the lower right corner of the picture, we managed to edit it into a cucumber without any adverse impact on the other objects. + +![](images/708aec49c9cafdd3fe8a679bbe324ff268d481f8fb6ad070ec7bf7239ee84e84.jpg) +IN2N + +![](images/034e98c8f141d54b5efd4f1d503b90feb7866de98e04acbcdeeda0ba4dd087cc.jpg) + +![](images/7a72df2a250f22c3c24084e86a09ea16191a5a3aab175c256eb9a0f3db5d0367.jpg) + +![](images/06bbde5350ffd895dbcc5ee14c1d5b5fd62bf26eaf71128c623ad62979a4033f.jpg) + +![](images/66740c2206392f85d313bdc12fbe63c287529d8cd856b75f4b37a2002a4be59c.jpg) + +![](images/0c6c750adaf9a8ef3d823ff004ac4cbf87ddd3cf348b876acb87c895ec71422b.jpg) +Original +Ours + +![](images/2b4c37b005bca97a16961b72f46b18b0e4f787d3ff5863caf159d08208a9040b.jpg) +"Turn his pants into yellow." + +![](images/db5a6d9d9b23e1a617def74bbd5044e16fad1d6ac3ef06bcdc460ab648d50e26.jpg) +Turn his pants into a bronze statue. + +![](images/89f0d60fa75b90e235ec080c6df3f97077da368e655a292c816379632ec0975f.jpg) +"Turn his t-shirt into white." + +![](images/9452d962b82a94bf81a3ddd21526532d1f6813bcb7e144e52a288018fbb825e8.jpg) +'Make his t-shirt out of leather + +![](images/2b5812dd0e21bdfb7a1158c73e58d03e6c059d64c65171c905f4568749e01f5e.jpg) +Figure 5. Comparisons with Instruct-NeRF2NeRF (IN2N) [11] on the scene presented in their paper. +Original +Figure 6. Qualitative results on complex multi-object scenes. The background "desk", the foreground "flower pot", and the multiview blocked foreground "rolling pin" are edited separately. + +![](images/7d785d265676c476a31937017422b796380e02da12b72a22bef1e420abbe5ab6.jpg) +"Turn the desk into wood" + +![](images/d203d78c3c3f4986abb31e1dced7ad8764bb2146549371e508739125abc1f1f7.jpg) +"Turn the flowerpot red" +"Turn the rolling pin into cucumber" + +![](images/b7703d6d142d19a1d51628333f931145036b637db4645123ac1bfbae28d90ade.jpg) + +Table 1. Quantitative evaluation on the bicycle scene of the MipNeRF360 dataset [3]. + +
MethodsCTIDS↑IIS↑FID↓Time↓
IN2N [11]0.220.8510351 min
Ours-DVGO [45]0.110.8214840 min
Ours-3DGS0.280.955120 min
+ +# 4.3. Quantitative Evaluation + +Metric Comparisons. Table 1 shows quantitative results on the bicycle scene of the Mip-NeRF360 dataset [3], comparing with IN2N [11] and Direct Voxel Grid Optimization (DVGO) [45] as the representation. The metrics include CLIP Text-Image Direction Similarity (CTIDS), Image-Image Similarity (IIS), FID, and training time. GaussianEditor achieves the best results in all metrics. The test data is shown in the supplementary material. + +User Study. We perform a user study comparing with IN2N on the bear scene in Fig. 4 and the human scene in Fig. 5, involving 21 participants. GaussianEditor gets a + +87.07% voting percentage, while IN2N gets 12.93%. + +# 4.4. Ablation Study and Analysis + +Ablation of Gaussian RoI, Text RoI, RoI Lifting. To validate the effectiveness of each module in our framework, we design three variant approaches: (1) w/o Gaussian RoI: We discontinued the use of Gaussian RoI $\mathcal{G}_{RoI}$ to control the gradients of Gaussian points, as mentioned in Eq. 14. (2) w/o Text ROI: In this scenario, we ceased the selection of text ROI $\mathcal{T}_{RoI}$ using LLM assistant $\mathcal{M}_{LLM}$ . Instead, all the words in the user's instruction are put to $\mathcal{M}_{Seg}$ to get $\mathcal{I}_{RoI}$ . (3) w/o RoI lifting: Instead of lifting the image RoI $\mathcal{T}_{RoI}$ to 3D Gaussians, the image RoI $\mathcal{T}_{RoI}$ is used to govern the calculation of the loss. That is, only the pixels within the image RoI $\mathcal{T}_{RoI}$ are taken into account for the loss computation. Fig. 7 showcases the outcomes of our ablation experiment, which aimed to edit the doll based on the instruction "Turn its mouth into red." The results reveal the following findings. (1) When the Gaussian RoI is not used, the 3D scene is all turned red because the 2D diffusion fails to control the editing area. (2) In cases where + +![](images/2f3b4558d05a720cdccebd35f1b2fa8dc23fc0ace4f66a8d90c517a280250ea9.jpg) +Figure 7. Ablation experiment of RoI. + +text ROI $\mathcal{T}_{Roi}$ is not utilized, the grounding segmentation model tends to segment the entire foreground object, leading to the doll being entirely edited to red. (3) When RoI lifting $\mathcal{M}_{Lift}$ is not employed, the doll's mouth is successfully turned red, but other facial areas are also affected. Because the grounding segmentation model may fail to parse specific views, noise exists on the image RoI $\mathcal{I}_{Roi}$ . Consequently, leakage occurs during the editing process. Our proposed RoI lifting module effectively addresses this issue during training. In conclusion, our ablation experiment demonstrates the effectiveness of several RoI-related modules in our method. + +Ablation of Scene Description Generation. We further conduct experiments to evaluate the role of scene description generation, employing three distinct experimental setups. The first one composes the user message $\mathcal{T}_{user}$ , without employing scene description. The second method randomly samples a view and extracts the corresponding image's text description as the scene description. The third one represents the complete version of our approach. The test scene involves a park where a bike and a bench are positioned closely together. The editing instruction is "Turn the thing next to the bike orange". The obtained results are presented in Fig. 8. As shown in the image, when scene description is not employed, the LLM fails to acquire the Text ROI according to the user instructions, resulting in editing failure. The second one randomly samples images to obtain scene descriptions, resulting in incomplete descriptions and leading to an incorrect text ROI prediction by the LLM. Consequently, the final editing result turns the road into an orange color. In contrast, our method flawlessly executes the editing task. This success can be attributed to scene description generation, which obtains an accurate text description encompassing the relative positional relationship between the bicycle and the bench. This enables the LLM to analyze and determine the user's intention to edit the bench. Consequently, the desired color change of the bench is successfully implemented. + +# 4.5. Limitations + +Although our framework has solved some problems inherited from the integrated sub-modules, e.g. noise in the results of grounding segmentation, there are still some prob + +![](images/4120a7d7f36014858efa3395d672fbe704f351f111af9ab379803e57e7f07d4e.jpg) +Figure 8. Ablation results about the scene description generation. + +lems that the current system cannot completely avoid. In scene description generation, the descriptions from different views of the same object may differ from each other. When the differences are large enough, the LLM may misunderstand these descriptions as those from multiple objects. This issue does not affect the results in the current experiment, but we would like to optimize this in the future. In addition, our system cannot achieve good editing results in scenes where the grounding segmentation or diffusion model completely fails, such as drastic geometric editing. + +# 5. Conclusion + +This paper proposes a systematic framework, named GaussianEditor, for text-guided delicate 3D scene editing. As we know, GaussianEditor is one of the first works to edit 3D Gaussians, taking advantage of the explicit property of 3D Gaussians and making it easy to control the editing area precisely. Several techniques are proposed to achieve delicate editing, including extracting instruction RoI from texts, aligning the RoI to 3D Gaussians, and editing the scene with the Gaussian RoI. GaussianEditor achieves notably more delicate editing results than IN2N [11] with much shorter training time (within 20 minutes v.s. 45 minutes - 2 hours). Noticing recent works [27, 49] have extended Gaussian splatting to dynamic scenes, we leave the delicate editing in dynamic scenes as future work. + +# References + +[1] Chong Bao, Yinda Zhang, Bangbang Yang, Tianxing Fan, Zesong Yang, Hujun Bao, Guofeng Zhang, and Zhaopeng Cui. Sine: Semantic-driven image-based nerf editing with prior-guided editing field. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2 +[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, 2021. 2 +[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022. 2, 7 +[4] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 2, 3, 5 +[5] Jiazhong Cen, Zanwei Zhou, Jiemin Fang, Chen Yang, Wei Shen, Lingxi Xie, Dongsheng Jiang, Xiaopeng Zhang, and Qi Tian. Segment anything in 3d with nerfs. In NeurIPS, 2023. 5 +[6] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In ECCV, 2022. 2 +[7] Pei-Ze Chiang, Meng-Shiun Tsai, Hung-Yu Tseng, Wei-Sheng Lai, and Wei-Chen Chiu. Stylizing 3d scene via implicit representation and hypernetwork. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2022. 2 +[8] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 2021. 2 +[9] Shuangkang Fang, Yufeng Wang, Yi Yang, Yi Yang, Yi-Hsuan Tsai, Wenrui Ding, Ming-Hsuan Yang, and Shuchang Zhou. Text-driven editing of 3d scenes without retraining. Arxiv preprint arXiv:2309.04917, 2023. 2 +[10] William Gao, Noam Aigerman, Thibault Groueix, Vova Kim, and Rana Hanocka. Textdeformer: Geometry manipulation using text guidance. In ACM SIGGRAPH 2023 Conference Proceedings, 2023. 2 +[11] Ayaan Haque, Matthew Tancik, Alexei Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. In CVPR, 2023. 2, 3, 6, 7, 8 +[12] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2 +[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 2 +[14] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. The Journal of Machine Learning Research, 2022. 2 +[15] Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. Avatarclip: Zero-shot text + +driven generation and animation of 3d avatars. arXiv preprint arXiv:2205.08535, 2022. 2 +[16] Hsin-Ping Huang, Hung-Yu Tseng, Saurabh Saini, Maneesh Singh, and Ming-Hsuan Yang. Learning to stylize novel views. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 2 +[17] Yi-Hua Huang, Yue He, Yu-Jie Yuan, Yu-Kun Lai, and Lin Gao. Stylizednerf: consistent 3d scene stylization as stylized nerf via 2d-3d mutual learning. In CVPR, 2022. 2 +[18] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 2023. 2, 3, 5 +[19] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 5 +[20] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nef for editing via feature field distillation. Advances in Neural Information Processing Systems, 2022. 2 +[21] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 5 +[22] Yuan Li, Zhi-Hao Lin, David Forsyth, Jia-Bin Huang, and Shenlong Wang. Climatenerf: Physically-based neural rendering for extreme climate synthesis. arXiv e-prints, 2022. 2 +[23] Yuhan Li, Yishun Dou, Yue Shi, Yu Lei, Xuanhong Chen, Yi Zhang, Peng Zhou, and Bingbing Ni. Focaldreamer: Text-driven 3d editing via focal-fusion assembly. arXiv preprint arXiv:2308.10608, 2023. 3 +[24] Hao-Kang Liu, I Shen, Bing-Yu Chen, et al. Nerf-in: Free-form nerf inpainting with rgb-d priors. arXiv preprint arXiv:2206.04901, 2022. +[25] Steven Liu, Xiuming Zhang, Zhoutong Zhang, Richard Zhang, Jun-Yan Zhu, and Bryan Russell. Editing conditional radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, 2021. 2 +[26] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 5 +[27] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 8 +[28] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2 +[29] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 2021. 2 + +[30] Ashkan Mirzaei, Tristan Aumentado-Armstrong, Marcus A. Brubaker, Jonathan Kelly, Alex Levinshtein, Konstantinos G. Derpanis, and Igor Gilitschenski. Watch your steps: Local image and scene editing by text instructions. In arXiv preprint arXiv:2308.08947, 2023. 2, 3 +[31] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 2022. 2 +[32] Thu Nguyen-Phuoc, Feng Liu, and Lei Xiao. Snerf: stylized neural implicit representations for 3d scenes. arXiv preprint arXiv:2207.02363, 2022. 2 +[33] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2 +[34] Atsuhiro Noguchi, Xiao Sun, Stephen Lin, and Tatsuya Harada. Neural articulated radiance field. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 2 +[35] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 2019. 5 +[36] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2 +[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022. 2 +[38] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 2 +[39] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, 2022. 2 +[40] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 2022. 2 +[41] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2 +[42] Sara Fridovich-Keil and Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 2 +[43] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using + +nonequilibrium thermodynamics. In International conference on machine learning, 2015. 2 +[44] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 2019. 2 +[45] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, 2022. 2, 7 +[46] Vadim Tschernezki, Iro Laina, Diane Larlus, and Andrea Vedaldi. Neural feature fusion fields: 3d distillation of self-supervised 2d image representations. In 2022 International Conference on 3D Vision (3DV), 2022. 2 +[47] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2 +[48] Can Wang, Ruixiang Jiang, Mengei Chai, Mingming He, Dongdong Chen, and Jing Liao. Nerf-art: Text-driven neural radiance fields stylization. TVCG, 2023. 2 +[49] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Wang Xinggang. 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528, 2023. 8 +[50] Qiling Wu, Jianchao Tan, and Kun Xu. Palettenerf: Palette-based color editing for nerfs. arXiv preprint arXiv:2212.12871, 2022. 2 +[51] Jiale Xu, Xintao Wang, Yan-Pei Cao, Weihao Cheng, Ying Shan, and Shenghua Gao. Instructp2p: Learning to edit 3d point clouds with text instructions. arXiv preprint arXiv:2306.07154, 2023. 2 +[52] Tianhan Xu and Tatsuya Harada. Deforming radiance fields with cages. In European Conference on Computer Vision, 2022. +[53] Bangbang Yang, Chong Bao, Junyi Zeng, Hujun Bao, Yinda Zhang, Zhaopeng Cui, and Guofeng Zhang. Neumesh: Learning disentangled neural mesh-based implicit field for geometry and texture editing. In European Conference on Computer Vision, 2022. 2 +[54] Kai Zhang, Nick Kolkin, Sai Bi, Fujun Luan, Zexiang Xu, Eli Shechtman, and Noah Snavely. Arf: Artistic radiance fields. In European Conference on Computer Vision, 2022. 2 +[55] Jingyu Zhuang, Chen Wang, Lingjie Liu, Liang Lin, and Guanbin Li. Dreameditor: Text-driven 3d scene editing with neural fields. arXiv preprint arXiv:2306.13455, 2023. 3 \ No newline at end of file diff --git a/2024/GaussianEditor_ Editing 3D Gaussians Delicately with Text Instructions/images.zip b/2024/GaussianEditor_ Editing 3D Gaussians Delicately with Text Instructions/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9fb9750acda22d53448e67da80ca1fbb28523a1c --- /dev/null +++ b/2024/GaussianEditor_ Editing 3D Gaussians Delicately with Text Instructions/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:782c029170eb7c3b29b920244d23cf085bb6e74fc467d4cd927a102bb705ee0f +size 940044 diff --git a/2024/GaussianEditor_ Editing 3D Gaussians Delicately with Text Instructions/layout.json b/2024/GaussianEditor_ Editing 3D Gaussians Delicately with Text Instructions/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..38f09a32ad8603941260f0b080ea18978d3b1d2a --- /dev/null +++ b/2024/GaussianEditor_ Editing 3D Gaussians Delicately with Text Instructions/layout.json @@ -0,0 +1,10883 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 74, + 103, + 520, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 103, + 520, + 121 + ], + "spans": [ + { + "bbox": [ + 74, + 103, + 520, + 121 + ], + "type": "text", + "content": "GaussianEditor: Editing 3D Gaussians Delicately with Text Instructions" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 118, + 142, + 475, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 142, + 475, + 171 + ], + "spans": [ + { + "bbox": [ + 118, + 142, + 475, + 171 + ], + "type": "text", + "content": "Junjie Wang," + }, + { + "bbox": [ + 118, + 142, + 475, + 171 + ], + "type": "inline_equation", + "content": "^{\\star}" + }, + { + "bbox": [ + 118, + 142, + 475, + 171 + ], + "type": "text", + "content": " Jiemin Fang" + }, + { + "bbox": [ + 118, + 142, + 475, + 171 + ], + "type": "inline_equation", + "content": "^{\\star}" + }, + { + "bbox": [ + 118, + 142, + 475, + 171 + ], + "type": "text", + "content": ", Xiaopeng Zhang, Lingxi Xie, Qi Tian Huawei Inc." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 82, + 173, + 511, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 173, + 511, + 186 + ], + "spans": [ + { + "bbox": [ + 82, + 173, + 511, + 186 + ], + "type": "text", + "content": "{is.wangjunjie, jaminfong, zxphistory, 198808xc}@gmail.com tian.qil@huawei.com" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 96, + 196, + 483, + 458 + ], + "blocks": [ + { + "bbox": [ + 96, + 196, + 483, + 458 + ], + "lines": [ + { + "bbox": [ + 96, + 196, + 483, + 458 + ], + "spans": [ + { + "bbox": [ + 96, + 196, + 483, + 458 + ], + "type": "image", + "image_path": "82c2d2fb93de361a4093dd459c31c3ea3797437513a565caafe64e3ee9fd1981.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 463, + 547, + 486 + ], + "lines": [ + { + "bbox": [ + 45, + 463, + 547, + 486 + ], + "spans": [ + { + "bbox": [ + 45, + 463, + 547, + 486 + ], + "type": "text", + "content": "Figure 1. We propose GaussianEditor, an interactive framework to achieve delicate 3D scene editing following text instructions. As shown in this figure, our method can precisely control the editing region and achieve multi-round editing." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 496, + 192, + 508 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 496, + 192, + 508 + ], + "spans": [ + { + "bbox": [ + 143, + 496, + 192, + 508 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 512, + 290, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 512, + 290, + 681 + ], + "spans": [ + { + "bbox": [ + 46, + 512, + 290, + 681 + ], + "type": "text", + "content": "Recently, impressive results have been achieved in 3D scene editing with text instructions based on a 2D diffusion model. However, current diffusion models primarily generate images by predicting noise in the latent space, and the editing is usually applied to the whole image, which makes it challenging to perform delicate, especially localized, editing for 3D scenes. Inspired by recent 3D Gaussian splatting, we propose a systematic framework, named GaussianEditor, to edit 3D scenes delicately via 3D Gaussians with text instructions. Benefiting from the explicit property of 3D Gaussians, we design a series of techniques to achieve delicate editing. Specifically, we first extract the region of interest (RoI) corresponding to the text instruction, aligning it to 3D Gaussians. The Gaussian RoI is fur" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 497, + 547, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 497, + 547, + 581 + ], + "spans": [ + { + "bbox": [ + 306, + 497, + 547, + 581 + ], + "type": "text", + "content": "ther used to control the editing process. Our framework can achieve more delicate and precise editing of 3D scenes than previous methods while enjoying much faster training speed, i.e. within 20 minutes on a single V100 GPU, more than twice as fast as Instruct-NeRF2NeRF (45 minutes - 2 hours)1. The project page is at GaussianEditor.github.io." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 604, + 388, + 617 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 604, + 388, + 617 + ], + "spans": [ + { + "bbox": [ + 306, + 604, + 388, + 617 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 624, + 547, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 624, + 547, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 624, + 547, + 686 + ], + "type": "text", + "content": "Creating 3D assets has played a critical role in many applications and industries, e.g. movie/game production, artistic creation, AR, VR etc. However, this process is usually expensive and cumbersome, especially for traditional pipelines. Designers need to take a lot of labor and time to" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 130, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 130, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 130, + 703 + ], + "type": "text", + "content": "*Equal contributions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "text", + "content": "1The editing time varies in different scenes according to the scene structure complexity." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20902" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": "finish each step, e.g. sketching, building structures, creating textures etc. One cheap and effective way of creating high-quality 3D assets is to start from an existing scene, capturing, modeling, and editing the scene and obtaining the wanted one. This approach can be also used for user-interactive entertainment applications." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 146, + 289, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 146, + 289, + 397 + ], + "spans": [ + { + "bbox": [ + 46, + 146, + 289, + 397 + ], + "type": "text", + "content": "Neural radiance field methods [2, 3, 6, 29, 31, 42, 45] have shown great power in representing 3D scenes and synthesizing novel-view images. Past years have witnessed the rapid development of NeRF and its variants, from both quality and efficiency perspectives. Editing a pre-trained NeRF model becomes a promising way to edit 3D scenes. Represented by Instruct-NeRF2NeRF [11], researchers propose to use the image-conditioned 2D diffusion model, e.g. InstructPix2Pix [4], to edit 3D scenes simply with text instructions. Notable results have been achieved as real scenes can be changed following the text instruction. However, current 2D diffusion models face challenges in accurately localizing editing regions, which hinders the generation of finely edited scenes due to the change of unintended regions. Even though some works [30] propose to constrain the editing region on edited 2D images, the editing region is not accurately localized and hard to apply to the 3D representation. Besides, NeRF-based methods [9, 48] bear coupling effects between different spatial positions, e.g. different points are queried from the same MLP field (for implicit representations) or voxel vertices (for explicit representations)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 400, + 287, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 400, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 287, + 496 + ], + "type": "text", + "content": "Recent 3D Gaussian Splatting [18] (3D-GS) has been a groundbreaking work in the radiance field, which is the first to achieve a real sense of real-time rendering while enjoying high rendering quality and training speed. Besides its efficiency, we further notice its natural explicit property. 3D-GS has a great advantage for editing tasks as each 3D Gaussian exists individually. Editing 3D scenes by directly manipulating 3D Gaussians with desired constraints is easy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "text", + "content": "Aiming at editing 3D scenes delicately, we propose to represent the scene with 3D Gaussians, which can be edited with text instructions, and name our method as GaussianEditor. GaussianEditor is divided into three main parts to achieve precise control for editing regions. The first is the region of interest (RoI) extraction from the given text instruction. The instruction may be complex or indirect while this module helps extract the keywords matching the RoI for editing. The second part aligns the extracted text RoI to the 3D Gaussian space through the image space, where a grounding segmentation module is applied. The last part is to edit the original 3D Gaussians delicately with constraints in the obtained 3D Gaussian RoI. With the above processes, the region for editing can be precisely localized simply from text instructions, which constrains the 3D Gaussian updating to obtain a delicately edited new 3D scene. Besides, we enable interfaces for users to introduce more exact instructions for more delicate editing, e.g. Gaussian point selecting" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 72, + 504, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 504, + 84 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 504, + 84 + ], + "type": "text", + "content": "and 3D boxes for modifying the editing regions2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 85, + 515, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 85, + 515, + 95 + ], + "spans": [ + { + "bbox": [ + 317, + 85, + 515, + 95 + ], + "type": "text", + "content": "Our contributions can be summarized as follows." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 97, + 545, + 252 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 306, + 97, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 97, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 306, + 97, + 545, + 133 + ], + "type": "text", + "content": "- As far as we know, our GaussianEditor is one of the first systematic methods to achieve delicate 3D scene editing based on 3D Gaussian splatting." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 133, + 545, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 133, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 306, + 133, + 545, + 204 + ], + "type": "text", + "content": "- A series of techniques are designed and proposed to precisely localize the editing region of interest, which are aligned and applied to 3D Gaussians. Though some submodules are from existing works, we believe integrating these awesome techniques to work effectively is a valuable topic, which is what we focus on in this paper." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 205, + 545, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 205, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 306, + 205, + 545, + 252 + ], + "type": "text", + "content": "- Our method achieves a series of more delicate editing results compared with the previous representative work Instruct-NeRF2NeRF [11] with much shorter training time (within 20 minutes v.s. 45 minutes - 2 hours)." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 264, + 392, + 276 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 264, + 392, + 276 + ], + "spans": [ + { + "bbox": [ + 306, + 264, + 392, + 276 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 284, + 545, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 284, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 304, + 284, + 545, + 547 + ], + "type": "text", + "content": "2D Image Editing with Diffusion Models. By advancements in diffusion model technology [13, 43], numerous generative models [40] have attained remarkably impressive outcomes in image synthesis. Recent progressions in diffusion models have showcased their adeptness in fabricating lifelike images from arbitrary textual inputs [8, 14, 39, 41, 44]. Harnessing the robust semantic comprehension and image generation capabilities of foundational diffusion models, an escalating number of research explorations are currently employing diffusion models as a fundamental framework for implementing text-based image editing functionalities [33, 36, 37, 40]. Some of these methodologies necessitate the manual provision of captions for both the original and edited images [12], while others mandate specific scenario-based training for optimization [38]. These requisites have rendered it arduous for ordinary users to avail themselves of such techniques. Expanding upon this foundation, iP2P [4] introduces instruction-based capabilities to image editing, enabling users to simply input an image and appraise the model of the desired alterations. This user-friendly approach facilitates the democratization of image editing in a more accessible manner." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 553, + 545, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 553, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 304, + 553, + 545, + 685 + ], + "type": "text", + "content": "3D Scene Editing of Radiance Fields. 3D Scene Editing of Radiance Fields has gained significant popularity as a recent research direction [1, 10, 15, 20, 22-25, 28, 34, 46, 47, 51-53]. The main objective of such methods is to enable the manipulation of both the geometry and appearance of 3D scene representations. However, editing such scenes poses inherent challenges due to the implicit nature of traditional NeRF representations, which lack precise localization capabilities. As a result, previous works have primarily focused on achieving global style transformations of 3D scenes [7, 16, 17, 32, 48, 50, 54]. While some efforts" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "text", + "content": "2These additional instructions are applied to generate the man with two different edited half faces in Fig. 1." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20903" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 71, + 541, + 413 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 541, + 413 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 541, + 413 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 541, + 413 + ], + "type": "image", + "image_path": "45ed546d67486b7683a68eef9a6eb990c955081fec28ed5a0c96e682cea5dcfa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 420, + 547, + 487 + ], + "lines": [ + { + "bbox": [ + 46, + 420, + 547, + 487 + ], + "spans": [ + { + "bbox": [ + 46, + 420, + 547, + 487 + ], + "type": "text", + "content": "Figure 2. Our framework, named GaussianEditor, consists of three key steps. First, a module " + }, + { + "bbox": [ + 46, + 420, + 547, + 487 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{Desc}" + }, + { + "bbox": [ + 46, + 420, + 547, + 487 + ], + "type": "text", + "content": " is used to get the description of the input scene, which is put to an LLM assistant " + }, + { + "bbox": [ + 46, + 420, + 547, + 487 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{LLM}" + }, + { + "bbox": [ + 46, + 420, + 547, + 487 + ], + "type": "text", + "content": " with the text instruction " + }, + { + "bbox": [ + 46, + 420, + 547, + 487 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 46, + 420, + 547, + 487 + ], + "type": "text", + "content": " provided by the user to obtain the text RoI " + }, + { + "bbox": [ + 46, + 420, + 547, + 487 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{RoI}" + }, + { + "bbox": [ + 46, + 420, + 547, + 487 + ], + "type": "text", + "content": ". Second, a grounding segmentation module " + }, + { + "bbox": [ + 46, + 420, + 547, + 487 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{Seg}" + }, + { + "bbox": [ + 46, + 420, + 547, + 487 + ], + "type": "text", + "content": " is used to convert " + }, + { + "bbox": [ + 46, + 420, + 547, + 487 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{RoI}" + }, + { + "bbox": [ + 46, + 420, + 547, + 487 + ], + "type": "text", + "content": " to image RoI " + }, + { + "bbox": [ + 46, + 420, + 547, + 487 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{RoI}" + }, + { + "bbox": [ + 46, + 420, + 547, + 487 + ], + "type": "text", + "content": ", which is then lifted to 3D Gaussians RoI " + }, + { + "bbox": [ + 46, + 420, + 547, + 487 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{RoI}" + }, + { + "bbox": [ + 46, + 420, + 547, + 487 + ], + "type": "text", + "content": " by RoI lifting " + }, + { + "bbox": [ + 46, + 420, + 547, + 487 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{Lift}" + }, + { + "bbox": [ + 46, + 420, + 547, + 487 + ], + "type": "text", + "content": ", where additional user instructions " + }, + { + "bbox": [ + 46, + 420, + 547, + 487 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 46, + 420, + 547, + 487 + ], + "type": "text", + "content": " can be incorporated. Third, following the user instruction " + }, + { + "bbox": [ + 46, + 420, + 547, + 487 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 46, + 420, + 547, + 487 + ], + "type": "text", + "content": ", rendered image " + }, + { + "bbox": [ + 46, + 420, + 547, + 487 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{render}" + }, + { + "bbox": [ + 46, + 420, + 547, + 487 + ], + "type": "text", + "content": " from randomly chosen views is edited by a diffusion model " + }, + { + "bbox": [ + 46, + 420, + 547, + 487 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{DM}" + }, + { + "bbox": [ + 46, + 420, + 547, + 487 + ], + "type": "text", + "content": ". The loss between " + }, + { + "bbox": [ + 46, + 420, + 547, + 487 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{render}" + }, + { + "bbox": [ + 46, + 420, + 547, + 487 + ], + "type": "text", + "content": " and edited one " + }, + { + "bbox": [ + 46, + 420, + 547, + 487 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{edit}" + }, + { + "bbox": [ + 46, + 420, + 547, + 487 + ], + "type": "text", + "content": " is calculated. Finally, gradient backpropagation and optimization are performed within the Gaussian RoI " + }, + { + "bbox": [ + 46, + 420, + 547, + 487 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{RoI}" + }, + { + "bbox": [ + 46, + 420, + 547, + 487 + ], + "type": "text", + "content": " to get the edited scene " + }, + { + "bbox": [ + 46, + 420, + 547, + 487 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{edit}" + }, + { + "bbox": [ + 46, + 420, + 547, + 487 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 498, + 289, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 289, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 289, + 712 + ], + "type": "text", + "content": "have been made towards object-centric scene editing[55], keeping the background unchanged has been a persistent challenge. For example, a recently proposed work, Instruct-NeRF2NeRF [11], implements text instruction-controlled 3D scene editing capabilities, achieving excellent editing effects while maintaining user-friendliness. However, it relies on the editing effect of 2D images, which may cause global changes to the 3D scene. A subsequent work [30] attempts to compute the relevance map between edited and unedited images to localize the editing area. The relevance map may be unreliable when the 2D IP2P [4] model fails. Other efforts [23] attempt to rely on the 3D coordinates entered by the user to decide the editing area. The introduction of 3D Gaussians [18] has provided an opportunity to address this limitation. Its explicit 3D representation methodology enables accurate selection and manipulation of editing areas. By incorporating LLMs, the whole process can be more automated." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 306, + 497, + 361, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 497, + 361, + 509 + ], + "spans": [ + { + "bbox": [ + 306, + 497, + 361, + 509 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 517, + 545, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 517, + 545, + 649 + ], + "spans": [ + { + "bbox": [ + 304, + 517, + 545, + 649 + ], + "type": "text", + "content": "In this section, we first review 3D representation methods in Sec. 3.1. Subsequently, in Sec. 3.2, we overview our proposed approach, which mainly includes three modules. Sec. 3.3 delves into the precise Region of Interest (RoI) extraction of text instructions, using scene description generation module " + }, + { + "bbox": [ + 304, + 517, + 545, + 649 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{Desc}" + }, + { + "bbox": [ + 304, + 517, + 545, + 649 + ], + "type": "text", + "content": " and LLM assistant " + }, + { + "bbox": [ + 304, + 517, + 545, + 649 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{LLM}" + }, + { + "bbox": [ + 304, + 517, + 545, + 649 + ], + "type": "text", + "content": ". Sec. 3.4 introduces how to align the instruction RoI with 3D Gaussians, using grounding segmentation module " + }, + { + "bbox": [ + 304, + 517, + 545, + 649 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{Seg}" + }, + { + "bbox": [ + 304, + 517, + 545, + 649 + ], + "type": "text", + "content": " and RoI lifting module " + }, + { + "bbox": [ + 304, + 517, + 545, + 649 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{Lift}" + }, + { + "bbox": [ + 304, + 517, + 545, + 649 + ], + "type": "text", + "content": ". Finally, Sec. 3.5 describes the delicate editing process within the obtained Gaussian RoI, using text to image diffusion model " + }, + { + "bbox": [ + 304, + 517, + 545, + 649 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{DM}" + }, + { + "bbox": [ + 304, + 517, + 545, + 649 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 658, + 391, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 391, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 391, + 670 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 677, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 715 + ], + "type": "text", + "content": "3D Gaussian Splatting. 3D Gaussian splatting [18] is a recent powerful 3D representation method. It represents the 3D scene with point-like 3D Gaussians " + }, + { + "bbox": [ + 305, + 677, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{G} = \\{g_1, g_2 \\dots g_N\\}" + }, + { + "bbox": [ + 305, + 677, + 545, + 715 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20904" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "inline_equation", + "content": "g_{i} = \\{\\mu, \\Sigma, c, \\alpha\\}" + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "inline_equation", + "content": "i \\in \\{1, \\dots, N\\}" + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": ". Among them, " + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "inline_equation", + "content": "\\mu \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": " is the position where the Gaussian centers, " + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "inline_equation", + "content": "\\Sigma \\in \\mathbb{R}^7" + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": " denotes the 3D covariance matrix, " + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "inline_equation", + "content": "c \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": " is the RGB color and " + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "inline_equation", + "content": "\\alpha \\in \\mathbb{R}^1" + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": " is the opacity. Benefitting from the compact representation of Gaussians and efficient differentiable rendering approach, 3D Gaussian splatting achieves real-time rendering with high quality. The splatting rendering process can be formulated as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 111, + 171, + 287, + 204 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 171, + 287, + 204 + ], + "spans": [ + { + "bbox": [ + 111, + 171, + 287, + 204 + ], + "type": "interline_equation", + "content": "C = \\sum_ {i \\in \\mathcal {N}} c _ {i} \\sigma_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\sigma_ {j}), \\tag {1}", + "image_path": "9541c5a560690b4fdd07cb29923a2d91e2d5e236915cbd5b53cb6df9a998df68.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 209, + 288, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 209, + 288, + 248 + ], + "spans": [ + { + "bbox": [ + 47, + 209, + 288, + 248 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 209, + 288, + 248 + ], + "type": "inline_equation", + "content": "\\sigma_{i} = \\alpha_{i}e^{-\\frac{1}{2}(x_{i})^{T}\\Sigma^{-1}(x_{i})}" + }, + { + "bbox": [ + 47, + 209, + 288, + 248 + ], + "type": "text", + "content": " represents the influence of the Gaussian to the image pixel and " + }, + { + "bbox": [ + 47, + 209, + 288, + 248 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 47, + 209, + 288, + 248 + ], + "type": "text", + "content": " is the distance between the 3D point and the center of the i-th Gaussian." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 253, + 163, + 265 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 253, + 163, + 265 + ], + "spans": [ + { + "bbox": [ + 47, + 253, + 163, + 265 + ], + "type": "text", + "content": "3.2. Overall Framework" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 271, + 287, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 271, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 271, + 287, + 331 + ], + "type": "text", + "content": "Given a group of 3D Gaussians " + }, + { + "bbox": [ + 46, + 271, + 287, + 331 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{input}" + }, + { + "bbox": [ + 46, + 271, + 287, + 331 + ], + "type": "text", + "content": " for an input scene and a text instruction " + }, + { + "bbox": [ + 46, + 271, + 287, + 331 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 46, + 271, + 287, + 331 + ], + "type": "text", + "content": " for editing, our Gaussian editor " + }, + { + "bbox": [ + 46, + 271, + 287, + 331 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 46, + 271, + 287, + 331 + ], + "type": "text", + "content": " can edit the 3D Gaussians delicately into a new one, denoted as " + }, + { + "bbox": [ + 46, + 271, + 287, + 331 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{edit}" + }, + { + "bbox": [ + 46, + 271, + 287, + 331 + ], + "type": "text", + "content": ", with the guidance of the instruction. The whole process can be formulated as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 121, + 337, + 287, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 337, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 121, + 337, + 287, + 350 + ], + "type": "interline_equation", + "content": "\\mathcal {G} _ {\\text {e d i t}} = \\mathcal {E} \\left(\\mathcal {G} _ {\\text {i n p u t}}, \\mathcal {T}\\right). \\tag {2}", + "image_path": "9190cdf7bea62b3cc56fc898c9c8d5e413084b29ef0cdc187931eb4af7e7e10f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 354, + 287, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 287, + 462 + ], + "type": "text", + "content": "Fig. 2 illustrates the overall framework of our approach, which consists of three main steps. First, the Region of Interest (RoI) is extracted from the text instruction. In this step, we employ a module named scene description generation " + }, + { + "bbox": [ + 46, + 354, + 287, + 462 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\text{Description}}" + }, + { + "bbox": [ + 46, + 354, + 287, + 462 + ], + "type": "text", + "content": " to get the description of the input scene. We then input the scene description " + }, + { + "bbox": [ + 46, + 354, + 287, + 462 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\text{scene}}" + }, + { + "bbox": [ + 46, + 354, + 287, + 462 + ], + "type": "text", + "content": " and text instruction " + }, + { + "bbox": [ + 46, + 354, + 287, + 462 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 46, + 354, + 287, + 462 + ], + "type": "text", + "content": " into a large language model assistant " + }, + { + "bbox": [ + 46, + 354, + 287, + 462 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\text{LLM}}" + }, + { + "bbox": [ + 46, + 354, + 287, + 462 + ], + "type": "text", + "content": " to determine where we should make edits in the scene. The output of this step is referred to as the instruction RoI " + }, + { + "bbox": [ + 46, + 354, + 287, + 462 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\text{RoI}}" + }, + { + "bbox": [ + 46, + 354, + 287, + 462 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 463, + 287, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 463, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 463, + 287, + 545 + ], + "type": "text", + "content": "The next step is the 3D Gaussian RoI alignment. We use a grounding segmentation module " + }, + { + "bbox": [ + 46, + 463, + 287, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{Seg}" + }, + { + "bbox": [ + 46, + 463, + 287, + 545 + ], + "type": "text", + "content": " to convert the RoI from text space, i.e. " + }, + { + "bbox": [ + 46, + 463, + 287, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{RoI}" + }, + { + "bbox": [ + 46, + 463, + 287, + 545 + ], + "type": "text", + "content": ", to the image space, i.e. " + }, + { + "bbox": [ + 46, + 463, + 287, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{RoI}" + }, + { + "bbox": [ + 46, + 463, + 287, + 545 + ], + "type": "text", + "content": ". Then the image RoI " + }, + { + "bbox": [ + 46, + 463, + 287, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{RoI}" + }, + { + "bbox": [ + 46, + 463, + 287, + 545 + ], + "type": "text", + "content": " is lifted to the RoI of 3D Gaussians " + }, + { + "bbox": [ + 46, + 463, + 287, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{RoI}" + }, + { + "bbox": [ + 46, + 463, + 287, + 545 + ], + "type": "text", + "content": " through RoI lifting module " + }, + { + "bbox": [ + 46, + 463, + 287, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{Lift}" + }, + { + "bbox": [ + 46, + 463, + 287, + 545 + ], + "type": "text", + "content": ". The Gaussian RoI allows us to control the regions where edits will be applied precisely." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": "The last step is delicate editing within the Gaussian RoI. In this step, we randomly sample the view to obtain the rendered image " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\text {render }}" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": ". A 2D diffusion model " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{DM}" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": " is used to perform the editing process on the rendered image " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\text {render }}" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": ", with the user instruction " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": " and the image " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\text {input }}" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": " of input scene as conditions. The resulting edited image is denoted as " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\text {edit }}" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": ". Subsequently, we calculate the loss between " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\text {edit }}" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\text {render }}" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": " and make gradient backpropagation within " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\text {RoI }}" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": ". This implies that only the regions specified by the RoI can receive corresponding gradients during the back-propagation process. Finally, optimization is executed based on these gradients. The final optimized scene representation " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\text {edit }}" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": " is obtained through several rounds of iterative optimization." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 313, + 72, + 536, + 329 + ], + "blocks": [ + { + "bbox": [ + 313, + 72, + 536, + 329 + ], + "lines": [ + { + "bbox": [ + 313, + 72, + 536, + 329 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 536, + 329 + ], + "type": "image", + "image_path": "0fd46d91fccb7c66678e177ba3b61ae5b740bbdcddb305bf1811e5bdbf7f46f2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 335, + 523, + 346 + ], + "lines": [ + { + "bbox": [ + 328, + 335, + 523, + 346 + ], + "spans": [ + { + "bbox": [ + 328, + 335, + 523, + 346 + ], + "type": "text", + "content": "Figure 3. The process of obtaining scene description." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 352, + 489, + 364 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 352, + 489, + 364 + ], + "spans": [ + { + "bbox": [ + 306, + 352, + 489, + 364 + ], + "type": "text", + "content": "3.3. RoI Extraction of Text Instruction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 370, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 370, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 370, + 545, + 456 + ], + "type": "text", + "content": "The instruction RoI is extracted for the editing regions from both the input 3D scene " + }, + { + "bbox": [ + 304, + 370, + 545, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{input}" + }, + { + "bbox": [ + 304, + 370, + 545, + 456 + ], + "type": "text", + "content": " and the text instruction " + }, + { + "bbox": [ + 304, + 370, + 545, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 304, + 370, + 545, + 456 + ], + "type": "text", + "content": " provided by the user. To achieve this, we employ a multimodal model " + }, + { + "bbox": [ + 304, + 370, + 545, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{MM}" + }, + { + "bbox": [ + 304, + 370, + 545, + 456 + ], + "type": "text", + "content": " in conjunction with the large language model assistant " + }, + { + "bbox": [ + 304, + 370, + 545, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{LLM}" + }, + { + "bbox": [ + 304, + 370, + 545, + 456 + ], + "type": "text", + "content": ". The first step is the scene description generation " + }, + { + "bbox": [ + 304, + 370, + 545, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{Desc}" + }, + { + "bbox": [ + 304, + 370, + 545, + 456 + ], + "type": "text", + "content": ", which aims to get the scene description " + }, + { + "bbox": [ + 304, + 370, + 545, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{scene}" + }, + { + "bbox": [ + 304, + 370, + 545, + 456 + ], + "type": "text", + "content": " from 3D Gaussians " + }, + { + "bbox": [ + 304, + 370, + 545, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{input}" + }, + { + "bbox": [ + 304, + 370, + 545, + 456 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 370, + 461, + 545, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 461, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 370, + 461, + 545, + 475 + ], + "type": "interline_equation", + "content": "\\mathcal {T} _ {\\text {s c e n e}} = \\mathcal {M} _ {\\text {D e s c}} \\left(\\mathcal {G} _ {\\text {i n p u t}}\\right). \\tag {3}", + "image_path": "4492be7e7926a46ad8b6417a86c35c9cf3232087fa2a00f6a8a0156053283b02.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 481, + 545, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 481, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 304, + 481, + 545, + 554 + ], + "type": "text", + "content": "The process of the scene description generation " + }, + { + "bbox": [ + 304, + 481, + 545, + 554 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{Desc}" + }, + { + "bbox": [ + 304, + 481, + 545, + 554 + ], + "type": "text", + "content": " is shown in Fig. 3. By leveraging the technique of differentiable splatting as shown in Eq. 1, a set of 2D image samples " + }, + { + "bbox": [ + 304, + 481, + 545, + 554 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{I}_{sample}\\}" + }, + { + "bbox": [ + 304, + 481, + 545, + 554 + ], + "type": "text", + "content": " are generated and then inputted into a multimodal model " + }, + { + "bbox": [ + 304, + 481, + 545, + 554 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{MM}" + }, + { + "bbox": [ + 304, + 481, + 545, + 554 + ], + "type": "text", + "content": " to generate corresponding text descriptions " + }, + { + "bbox": [ + 304, + 481, + 545, + 554 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{T}_{sample}\\}" + }, + { + "bbox": [ + 304, + 481, + 545, + 554 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 351, + 559, + 545, + 573 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 559, + 545, + 573 + ], + "spans": [ + { + "bbox": [ + 351, + 559, + 545, + 573 + ], + "type": "interline_equation", + "content": "\\mathcal {T} _ {\\text {s a m p l e}} = \\mathcal {M} _ {M M} \\left(\\mathcal {P} _ {M M}, \\mathcal {I} _ {\\text {s a m p l e}}\\right), \\tag {4}", + "image_path": "610e5ff2cbebd126fd5bd1ecfb06827dda3f15fa469b981e840743d69d740d54.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 579, + 545, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 579, + 545, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 579, + 545, + 663 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 579, + 545, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{MM}" + }, + { + "bbox": [ + 304, + 579, + 545, + 663 + ], + "type": "text", + "content": " is a prompt, such as \"What is the content of the image\", for multimodal model " + }, + { + "bbox": [ + 304, + 579, + 545, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{MM}" + }, + { + "bbox": [ + 304, + 579, + 545, + 663 + ], + "type": "text", + "content": " to get precise description. Subsequently, these descriptions " + }, + { + "bbox": [ + 304, + 579, + 545, + 663 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{T}_{\\text {sample }}\\}" + }, + { + "bbox": [ + 304, + 579, + 545, + 663 + ], + "type": "text", + "content": " are fed into a large language model " + }, + { + "bbox": [ + 304, + 579, + 545, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{LLM}" + }, + { + "bbox": [ + 304, + 579, + 545, + 663 + ], + "type": "text", + "content": ", which is specifically instructed by a prompt " + }, + { + "bbox": [ + 304, + 579, + 545, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\text {merge }}" + }, + { + "bbox": [ + 304, + 579, + 545, + 663 + ], + "type": "text", + "content": " to merge descriptions of diverse views into one detailed scene description " + }, + { + "bbox": [ + 304, + 579, + 545, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\text {scene }}" + }, + { + "bbox": [ + 304, + 579, + 545, + 663 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 345, + 670, + 545, + 684 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 670, + 545, + 684 + ], + "spans": [ + { + "bbox": [ + 345, + 670, + 545, + 684 + ], + "type": "interline_equation", + "content": "\\mathcal {T} _ {\\text {s c e n e}} = \\mathcal {M} _ {L L M} \\left(\\mathcal {P} _ {\\text {m e r g e}}, \\left\\{\\mathcal {T} _ {\\text {s a m p l e}} \\right\\}\\right). \\tag {5}", + "image_path": "a6529d5ef28623b9c48418f48367493bee71f5462923eab3d421306880374859.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": "After that, the scene description " + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\text{scene}}" + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": " and the user instruction " + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": " are combined with a predefined tem" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20905" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": "plate " + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\text{template}}" + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": ": \"Text description: " + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\text{scene}}" + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": " Edit Instruction: " + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": " Answer:\" to form the user message " + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\text{user}} = \\mathcal{T}_{\\text{template}}(\\mathcal{T}_{\\text{scene}}, \\mathcal{T})" + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": ". The LLM model " + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\text{LLM}}" + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": " is used to extract the instruction RoI " + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\text{RoI}}" + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": " from user message " + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\text{user}}" + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": " with a new prompt " + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\text{extract}}" + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 97, + 141, + 287, + 154 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 141, + 287, + 154 + ], + "spans": [ + { + "bbox": [ + 97, + 141, + 287, + 154 + ], + "type": "interline_equation", + "content": "\\mathcal {T} _ {R o I} = \\mathcal {M} _ {L L M} \\left(\\mathcal {P} _ {\\text {e x t r a c t}}, \\mathcal {T} _ {\\text {u s e r}}\\right). \\tag {6}", + "image_path": "8a1c4dee41d19562822970717264b42ce115819b4334d9ed7ae54fcae2dedda4.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 162, + 202, + 175 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 162, + 202, + 175 + ], + "spans": [ + { + "bbox": [ + 47, + 162, + 202, + 175 + ], + "type": "text", + "content": "3.4. 3D Gaussian RoI Alignment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 180, + 287, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 180, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 47, + 180, + 287, + 229 + ], + "type": "text", + "content": "To confine the 3D editing region within the instruction RoI, 3D Gaussian RoI " + }, + { + "bbox": [ + 47, + 180, + 287, + 229 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{RoI}" + }, + { + "bbox": [ + 47, + 180, + 287, + 229 + ], + "type": "text", + "content": " is aligned with the text RoI " + }, + { + "bbox": [ + 47, + 180, + 287, + 229 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{RoI}" + }, + { + "bbox": [ + 47, + 180, + 287, + 229 + ], + "type": "text", + "content": ". First, The RoI in the text space is transformed into the image space via a grounding segmentation module " + }, + { + "bbox": [ + 47, + 180, + 287, + 229 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{Seg}" + }, + { + "bbox": [ + 47, + 180, + 287, + 229 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 238, + 287, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 238, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 287, + 251 + ], + "type": "interline_equation", + "content": "\\mathcal {I} _ {R o I} = \\mathcal {M} _ {S e g} \\left(\\mathcal {I} _ {\\text {i n p u t}}, \\mathcal {T} _ {R o I}\\right), \\tag {7}", + "image_path": "ed3755bba2f6ad1ef413ce888ce626992d9374363793e984697e79ea22055323.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 259, + 279, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 259, + 279, + 272 + ], + "spans": [ + { + "bbox": [ + 47, + 259, + 279, + 272 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 259, + 279, + 272 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{input}" + }, + { + "bbox": [ + 47, + 259, + 279, + 272 + ], + "type": "text", + "content": " is rendered image of the input scene " + }, + { + "bbox": [ + 47, + 259, + 279, + 272 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{input}" + }, + { + "bbox": [ + 47, + 259, + 279, + 272 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 272, + 287, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 272, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 272, + 287, + 354 + ], + "type": "text", + "content": "Then we lift the RoI " + }, + { + "bbox": [ + 46, + 272, + 287, + 354 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{RoI}" + }, + { + "bbox": [ + 46, + 272, + 287, + 354 + ], + "type": "text", + "content": " in the image space to 3D Gaussian " + }, + { + "bbox": [ + 46, + 272, + 287, + 354 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{RoI}" + }, + { + "bbox": [ + 46, + 272, + 287, + 354 + ], + "type": "text", + "content": " through training. To achieve this, an additional RoI attribute " + }, + { + "bbox": [ + 46, + 272, + 287, + 354 + ], + "type": "inline_equation", + "content": "r\\in \\mathbb{R}^1" + }, + { + "bbox": [ + 46, + 272, + 287, + 354 + ], + "type": "text", + "content": " was added to 3D Gaussian " + }, + { + "bbox": [ + 46, + 272, + 287, + 354 + ], + "type": "inline_equation", + "content": "g_{i} = \\{\\mu_{i},\\Sigma_{i},c_{i},\\alpha_{i},r_{i}\\}" + }, + { + "bbox": [ + 46, + 272, + 287, + 354 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 272, + 287, + 354 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 272, + 287, + 354 + ], + "type": "text", + "content": " is initialized to 0, which means it is not in the Gaussians RoI, and 1 means it is inside the RoI. The set of " + }, + { + "bbox": [ + 46, + 272, + 287, + 354 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 272, + 287, + 354 + ], + "type": "text", + "content": " is denoted as " + }, + { + "bbox": [ + 46, + 272, + 287, + 354 + ], + "type": "inline_equation", + "content": "\\mathcal{R}\\in \\mathbb{R}^{\\mathcal{N},1}" + }, + { + "bbox": [ + 46, + 272, + 287, + 354 + ], + "type": "text", + "content": ", where the " + }, + { + "bbox": [ + 46, + 272, + 287, + 354 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 46, + 272, + 287, + 354 + ], + "type": "text", + "content": " is the number of 3D Gaussians " + }, + { + "bbox": [ + 46, + 272, + 287, + 354 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{input}" + }, + { + "bbox": [ + 46, + 272, + 287, + 354 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 356, + 287, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 356, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 47, + 356, + 287, + 380 + ], + "type": "text", + "content": "Then the color " + }, + { + "bbox": [ + 47, + 356, + 287, + 380 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 47, + 356, + 287, + 380 + ], + "type": "text", + "content": " in Eq. 1 was rewritten with " + }, + { + "bbox": [ + 47, + 356, + 287, + 380 + ], + "type": "inline_equation", + "content": "r_{i}" + }, + { + "bbox": [ + 47, + 356, + 287, + 380 + ], + "type": "text", + "content": " to get the rendered RoI " + }, + { + "bbox": [ + 47, + 356, + 287, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{RoI}^{render}" + }, + { + "bbox": [ + 47, + 356, + 287, + 380 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 99, + 388, + 287, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 388, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 99, + 388, + 287, + 422 + ], + "type": "interline_equation", + "content": "\\mathcal {I} _ {R o I} ^ {\\text {r e n d e r}} = \\sum_ {i \\in \\mathcal {N}} r _ {i} \\sigma_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\sigma_ {j}). \\tag {8}", + "image_path": "20251cda4bd5d863e1e7c2a3a53f5c8de007b2b73d432a255fc10e2f03cebdb6.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 430, + 287, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 430, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 47, + 430, + 287, + 465 + ], + "type": "text", + "content": "Taking inspiration from SA3D [5], to get the trained Gaussians RoI " + }, + { + "bbox": [ + 47, + 430, + 287, + 465 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{RoI}^{train}" + }, + { + "bbox": [ + 47, + 430, + 287, + 465 + ], + "type": "text", + "content": ", we adopt a similar loss function to supervise the training process:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 474, + 293, + 501 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 474, + 293, + 501 + ], + "spans": [ + { + "bbox": [ + 47, + 474, + 293, + 501 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p r o j} = \\lambda_ {1} \\sum \\left(\\mathcal {I} _ {R o I} ^ {\\text {r e n d e r}} \\cdot \\mathcal {I} _ {R o I}\\right) + \\lambda_ {2} \\sum \\left(\\left(1 - \\mathcal {I} _ {R o I} ^ {\\text {r e n d e r}}\\right) \\cdot \\mathcal {I} _ {R o I}\\right), \\tag {9}", + "image_path": "6744f3e05691fd9f656d66fdfd8492d764f8512addf6b606e193e577e289935a.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 502, + 287, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 502, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 502, + 287, + 586 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 502, + 287, + 586 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 47, + 502, + 287, + 586 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 502, + 287, + 586 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 47, + 502, + 287, + 586 + ], + "type": "text", + "content": " are hyperparameters. The " + }, + { + "bbox": [ + 47, + 502, + 287, + 586 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 47, + 502, + 287, + 586 + ], + "type": "text", + "content": " in Eq. 8 is updated via " + }, + { + "bbox": [ + 47, + 502, + 287, + 586 + ], + "type": "inline_equation", + "content": "r\\gets r - \\eta \\frac{\\partial\\mathcal{L}_{proj}}{\\partial r}" + }, + { + "bbox": [ + 47, + 502, + 287, + 586 + ], + "type": "text", + "content": " with gradient descent, where " + }, + { + "bbox": [ + 47, + 502, + 287, + 586 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 47, + 502, + 287, + 586 + ], + "type": "text", + "content": " denotes the learning rate. Eq. 9 encourages rendered RoI to cover the Image RoI and not exceed it. Additionally, the user can modify the trained Gaussian RoI " + }, + { + "bbox": [ + 47, + 502, + 287, + 586 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{RoI}^{train}" + }, + { + "bbox": [ + 47, + 502, + 287, + 586 + ], + "type": "text", + "content": " by giving added Gaussian RoI " + }, + { + "bbox": [ + 47, + 502, + 287, + 586 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{RoI}^{add}" + }, + { + "bbox": [ + 47, + 502, + 287, + 586 + ], + "type": "text", + "content": ", deleted Gaussian RoI " + }, + { + "bbox": [ + 47, + 502, + 287, + 586 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{RoI}^{del}" + }, + { + "bbox": [ + 47, + 502, + 287, + 586 + ], + "type": "text", + "content": " and 3D box " + }, + { + "bbox": [ + 47, + 502, + 287, + 586 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{3D}" + }, + { + "bbox": [ + 47, + 502, + 287, + 586 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 81, + 594, + 287, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 594, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 81, + 594, + 287, + 609 + ], + "type": "interline_equation", + "content": "\\mathcal {G} _ {R o I} = \\left(\\mathcal {G} _ {R o I} ^ {\\text {t r a i n}} \\cup \\mathcal {G} _ {R o I} ^ {\\text {a d d}} - \\mathcal {G} _ {R o I} ^ {\\text {d e l}}\\right) \\cap \\mathcal {B} _ {3 D}, \\tag {10}", + "image_path": "26597d45e709229235ed7671bc639aa915d2bb174e411aa25b1957f6e7a6608a.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 616, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 616, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 616, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{RoI}^{add}" + }, + { + "bbox": [ + 47, + 616, + 287, + 714 + ], + "type": "text", + "content": " represents the 3D Gaussians user wants to edit, " + }, + { + "bbox": [ + 47, + 616, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{RoI}^{del}" + }, + { + "bbox": [ + 47, + 616, + 287, + 714 + ], + "type": "text", + "content": " means 3D Gaussians user wants to keep from editing, " + }, + { + "bbox": [ + 47, + 616, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{3D}" + }, + { + "bbox": [ + 47, + 616, + 287, + 714 + ], + "type": "text", + "content": " is the coordinates of 3D cuboid it limits RoI to inside the box. " + }, + { + "bbox": [ + 47, + 616, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{RoI}" + }, + { + "bbox": [ + 47, + 616, + 287, + 714 + ], + "type": "text", + "content": " is the aligned RoI with the text RoI. For example, when editing the left face of the man in Fig. 1, grounding segmentation failed to ground \"left face\", instead, it grounded the whole face. In this scenario, the user can use the interactive interface to set the right face as " + }, + { + "bbox": [ + 47, + 616, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{RoI}^{del}" + }, + { + "bbox": [ + 47, + 616, + 287, + 714 + ], + "type": "text", + "content": " or enter" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 96 + ], + "type": "text", + "content": "the rectangular box where the left face is located as " + }, + { + "bbox": [ + 306, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{3D}" + }, + { + "bbox": [ + 306, + 72, + 545, + 96 + ], + "type": "text", + "content": ". The lifting process " + }, + { + "bbox": [ + 306, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\text{Lift}}" + }, + { + "bbox": [ + 306, + 72, + 545, + 96 + ], + "type": "text", + "content": " can be represented as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 371, + 105, + 545, + 118 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 105, + 545, + 118 + ], + "spans": [ + { + "bbox": [ + 371, + 105, + 545, + 118 + ], + "type": "interline_equation", + "content": "\\mathcal {G} _ {R o I} = \\mathcal {M} _ {\\text {L i f t}} \\left(\\mathcal {I} _ {R o I}, \\mathcal {O}\\right), \\tag {11}", + "image_path": "e978c4f9f758b7bd0ca447b51c719acd97335ff875f5acd1e6a1d48aa73e0b6b.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 125, + 530, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 125, + 530, + 139 + ], + "spans": [ + { + "bbox": [ + 306, + 125, + 530, + 139 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 125, + 530, + 139 + ], + "type": "inline_equation", + "content": "\\mathcal{O} = \\{\\mathcal{G}_{RoI}^{add},\\mathcal{G}_{RoI}^{del},\\mathcal{B}_{3D}\\}" + }, + { + "bbox": [ + 306, + 125, + 530, + 139 + ], + "type": "text", + "content": " is optional instructions." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 144, + 503, + 157 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 144, + 503, + 157 + ], + "spans": [ + { + "bbox": [ + 306, + 144, + 503, + 157 + ], + "type": "text", + "content": "3.5. Delicate Editing within Gaussian RoI" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 163, + 545, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 163, + 545, + 247 + ], + "spans": [ + { + "bbox": [ + 305, + 163, + 545, + 247 + ], + "type": "text", + "content": "To achieve delicate editing in 3D scenes, we use the Gaussian RoI to constrain the editing area. In particular, we randomly sample viewpoints from the 3D scene and render 2D image " + }, + { + "bbox": [ + 305, + 163, + 545, + 247 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\text {render }}" + }, + { + "bbox": [ + 305, + 163, + 545, + 247 + ], + "type": "text", + "content": ". After that, " + }, + { + "bbox": [ + 305, + 163, + 545, + 247 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\text {render }}" + }, + { + "bbox": [ + 305, + 163, + 545, + 247 + ], + "type": "text", + "content": " and noise level " + }, + { + "bbox": [ + 305, + 163, + 545, + 247 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 163, + 545, + 247 + ], + "type": "text", + "content": " are put into 2D diffusion model " + }, + { + "bbox": [ + 305, + 163, + 545, + 247 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{DM}" + }, + { + "bbox": [ + 305, + 163, + 545, + 247 + ], + "type": "text", + "content": ", with the user instruction " + }, + { + "bbox": [ + 305, + 163, + 545, + 247 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 305, + 163, + 545, + 247 + ], + "type": "text", + "content": " and image " + }, + { + "bbox": [ + 305, + 163, + 545, + 247 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\text {input }}" + }, + { + "bbox": [ + 305, + 163, + 545, + 247 + ], + "type": "text", + "content": " of input scene as conditions, to get edited image " + }, + { + "bbox": [ + 305, + 163, + 545, + 247 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\text {edit }}" + }, + { + "bbox": [ + 305, + 163, + 545, + 247 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 358, + 256, + 545, + 269 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 256, + 545, + 269 + ], + "spans": [ + { + "bbox": [ + 358, + 256, + 545, + 269 + ], + "type": "interline_equation", + "content": "\\mathcal {I} _ {\\text {e d i t}} = \\mathcal {D} \\left(\\mathcal {I} _ {\\text {r e n d e r}}, t; \\mathcal {T}, \\mathcal {I} _ {\\text {i n p u t}}\\right), \\tag {12}", + "image_path": "ea1d3b039b5d1d4e117db908d77bcd01b4d6a44360a49e7535cb016f1a402377.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 277, + 545, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 277, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 306, + 277, + 545, + 289 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 277, + 545, + 289 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 306, + 277, + 545, + 289 + ], + "type": "text", + "content": " is a randomly chosen noise level from " + }, + { + "bbox": [ + 306, + 277, + 545, + 289 + ], + "type": "inline_equation", + "content": "[t_{min}, t_{max}]" + }, + { + "bbox": [ + 306, + 277, + 545, + 289 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 289, + 545, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 289, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 306, + 289, + 545, + 312 + ], + "type": "text", + "content": "Similar to 3D-GS [18], we apply the " + }, + { + "bbox": [ + 306, + 289, + 545, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\perp}" + }, + { + "bbox": [ + 306, + 289, + 545, + 312 + ], + "type": "text", + "content": " and D-SSIM loss functions during editing." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 359, + 322, + 545, + 335 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 322, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 359, + 322, + 545, + 335 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = (1 - \\beta) \\mathcal {L} _ {\\mathbb {1}} + \\beta \\mathcal {L} _ {D - S S I M}. \\tag {13}", + "image_path": "8bc70025b386b7cd357419c4400d41ff837219955204fcfc53277405b15751c3.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 342, + 545, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 342, + 545, + 379 + ], + "spans": [ + { + "bbox": [ + 305, + 342, + 545, + 379 + ], + "type": "text", + "content": "the two losses are calculated between the 2D edited image " + }, + { + "bbox": [ + 305, + 342, + 545, + 379 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{ed}" + }, + { + "bbox": [ + 305, + 342, + 545, + 379 + ], + "type": "text", + "content": " and the rendered image " + }, + { + "bbox": [ + 305, + 342, + 545, + 379 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{rd}" + }, + { + "bbox": [ + 305, + 342, + 545, + 379 + ], + "type": "text", + "content": ". Then, gradient backpropagation is performed within Gaussian RoI " + }, + { + "bbox": [ + 305, + 342, + 545, + 379 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{RoI}" + }, + { + "bbox": [ + 305, + 342, + 545, + 379 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 393, + 386, + 545, + 410 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 386, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 393, + 386, + 545, + 410 + ], + "type": "interline_equation", + "content": "\\nabla \\mathcal {G} = \\frac {\\partial \\mathcal {L}}{\\partial \\mathcal {G}} \\cdot \\mathcal {R}, \\tag {14}", + "image_path": "b24047f0f2df980d14eb2d846aa39ffa6e9970da7a139748aa26d7f269c637ba.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 305, + 417, + 545, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 417, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 305, + 417, + 545, + 465 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 417, + 545, + 465 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 305, + 417, + 545, + 465 + ], + "type": "text", + "content": " is the set of RoI attributes. That means only Gaussians in RoI can receive gradients. Finally, we utilize the Adam algorithm to optimize the 3D Gaussians. After many rounds of training, the edited 3D scene " + }, + { + "bbox": [ + 305, + 417, + 545, + 465 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{edit}" + }, + { + "bbox": [ + 305, + 417, + 545, + 465 + ], + "type": "text", + "content": " is obtained." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 475, + 387, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 475, + 387, + 489 + ], + "spans": [ + { + "bbox": [ + 306, + 475, + 387, + 489 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 495, + 438, + 507 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 495, + 438, + 507 + ], + "spans": [ + { + "bbox": [ + 306, + 495, + 438, + 507 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 305, + 513, + 545, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 513, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 305, + 513, + 545, + 621 + ], + "type": "text", + "content": "Our method is implemented in PyTorch [35] and CUDA, based on 3D Gaussian splatting. The multimodal model we used in our method is BLIP2 [21], and we use GPT-3.5 Turbo to ground the text ROI. For grounding segmentation, We use the cascade strategy, i.e. first using Grounding DINO [26] to get the box on the image corresponding to the text, and then using SAM [19] to get the corresponding image RoI. The 2D diffusion model used in our method is Instruct Pix2Pix [4]. We leave more details in the Appendix." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 306, + 628, + 434, + 640 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 628, + 434, + 640 + ], + "spans": [ + { + "bbox": [ + 306, + 628, + 434, + 640 + ], + "type": "text", + "content": "4.2. Qualitative Evaluation" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 305, + 642, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 642, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 642, + 545, + 712 + ], + "type": "text", + "content": "Visualization Results. In Fig. 1 and Fig. 4, we present the visual results of GaussianEditor, demonstrating the precise editing effects while ensuring 3D consistency. Fig. 1 shows the editing capabilities for characters. The first column displays the original scenes. In the second column, the first row \"Give him a red nose\" illustrates" + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20906" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 71, + 219, + 178 + ], + "blocks": [ + { + "bbox": [ + 52, + 71, + 219, + 178 + ], + "lines": [ + { + "bbox": [ + 52, + 71, + 219, + 178 + ], + "spans": [ + { + "bbox": [ + 52, + 71, + 219, + 178 + ], + "type": "image", + "image_path": "5f22c4a02c7b65486ef4fb653e956cefd86e95e24b3b25abc09a7a5e82bd0d59.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 126, + 178, + 157, + 187 + ], + "lines": [ + { + "bbox": [ + 126, + 178, + 157, + 187 + ], + "spans": [ + { + "bbox": [ + 126, + 178, + 157, + 187 + ], + "type": "text", + "content": "Original" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 220, + 73, + 377, + 178 + ], + "blocks": [ + { + "bbox": [ + 220, + 73, + 377, + 178 + ], + "lines": [ + { + "bbox": [ + 220, + 73, + 377, + 178 + ], + "spans": [ + { + "bbox": [ + 220, + 73, + 377, + 178 + ], + "type": "image", + "image_path": "dde0ef5cc26e127ed167a336f9adbd5ad7ecb5e0379614f90378c982d759ecbe.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 237, + 179, + 362, + 187 + ], + "lines": [ + { + "bbox": [ + 237, + 179, + 362, + 187 + ], + "spans": [ + { + "bbox": [ + 237, + 179, + 362, + 187 + ], + "type": "text", + "content": "\"Make the road look like the meadow\"" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 378, + 73, + 535, + 178 + ], + "blocks": [ + { + "bbox": [ + 378, + 73, + 535, + 178 + ], + "lines": [ + { + "bbox": [ + 378, + 73, + 535, + 178 + ], + "spans": [ + { + "bbox": [ + 378, + 73, + 535, + 178 + ], + "type": "image", + "image_path": "16a796c5d293519271f3400bda7ee90a8574d5271b81cd14066c9fdb97a43998.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 414, + 179, + 500, + 187 + ], + "lines": [ + { + "bbox": [ + 414, + 179, + 500, + 187 + ], + "spans": [ + { + "bbox": [ + 414, + 179, + 500, + 187 + ], + "type": "text", + "content": "\"Turn the road into river\"" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 53, + 188, + 219, + 295 + ], + "blocks": [ + { + "bbox": [ + 53, + 188, + 219, + 295 + ], + "lines": [ + { + "bbox": [ + 53, + 188, + 219, + 295 + ], + "spans": [ + { + "bbox": [ + 53, + 188, + 219, + 295 + ], + "type": "image", + "image_path": "010f77f08f1bd6536a5983de22cb13cc1864a0994c00b63f478036ce138fa462.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 107, + 296, + 176, + 304 + ], + "lines": [ + { + "bbox": [ + 107, + 296, + 176, + 304 + ], + "spans": [ + { + "bbox": [ + 107, + 296, + 176, + 304 + ], + "type": "text", + "content": "\"Turn the bench red\"" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 220, + 189, + 377, + 295 + ], + "blocks": [ + { + "bbox": [ + 220, + 189, + 377, + 295 + ], + "lines": [ + { + "bbox": [ + 220, + 189, + 377, + 295 + ], + "spans": [ + { + "bbox": [ + 220, + 189, + 377, + 295 + ], + "type": "image", + "image_path": "3d66bb5ecd3784d4ae29acb40369765b4acae0537e1aa5ded42e9096f2c030ef.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 259, + 296, + 350, + 304 + ], + "lines": [ + { + "bbox": [ + 259, + 296, + 350, + 304 + ], + "spans": [ + { + "bbox": [ + 259, + 296, + 350, + 304 + ], + "type": "text", + "content": "\"Turn the bench into wood\"" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 378, + 189, + 535, + 294 + ], + "blocks": [ + { + "bbox": [ + 378, + 189, + 535, + 294 + ], + "lines": [ + { + "bbox": [ + 378, + 189, + 535, + 294 + ], + "spans": [ + { + "bbox": [ + 378, + 189, + 535, + 294 + ], + "type": "image", + "image_path": "26112cada5024cf6b88eef66b41d5c8f9558a08ebfc7e4de524289e5afeeeee3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 405, + 295, + 511, + 304 + ], + "lines": [ + { + "bbox": [ + 405, + 295, + 511, + 304 + ], + "spans": [ + { + "bbox": [ + 405, + 295, + 511, + 304 + ], + "type": "text", + "content": "\"Turn the bench into Turquoise\"" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 59, + 307, + 178, + 396 + ], + "blocks": [ + { + "bbox": [ + 59, + 307, + 178, + 396 + ], + "lines": [ + { + "bbox": [ + 59, + 307, + 178, + 396 + ], + "spans": [ + { + "bbox": [ + 59, + 307, + 178, + 396 + ], + "type": "image", + "image_path": "5b5009c5cf6088020e9801714120bad0c1c87f59087fdd087e1b95e2ec78238e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 103, + 396, + 133, + 405 + ], + "lines": [ + { + "bbox": [ + 103, + 396, + 133, + 405 + ], + "spans": [ + { + "bbox": [ + 103, + 396, + 133, + 405 + ], + "type": "text", + "content": "Original" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 51, + 407, + 539, + 418 + ], + "lines": [ + { + "bbox": [ + 51, + 407, + 539, + 418 + ], + "spans": [ + { + "bbox": [ + 51, + 407, + 539, + 418 + ], + "type": "text", + "content": "Figure 4. Qualitative results on outdoor scenes. Our method supports separate foreground and background editing in real-world scenes." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 179, + 308, + 296, + 396 + ], + "blocks": [ + { + "bbox": [ + 179, + 308, + 296, + 396 + ], + "lines": [ + { + "bbox": [ + 179, + 308, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 179, + 308, + 296, + 396 + ], + "type": "image", + "image_path": "b1058324351dc5258469e00cc9b25eef929993a41e0a07440c813b625881bc48.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 396, + 290, + 405 + ], + "lines": [ + { + "bbox": [ + 181, + 396, + 290, + 405 + ], + "spans": [ + { + "bbox": [ + 181, + 396, + 290, + 405 + ], + "type": "text", + "content": "\"Turn the bear into a polar bear\"" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 297, + 308, + 414, + 396 + ], + "blocks": [ + { + "bbox": [ + 297, + 308, + 414, + 396 + ], + "lines": [ + { + "bbox": [ + 297, + 308, + 414, + 396 + ], + "spans": [ + { + "bbox": [ + 297, + 308, + 414, + 396 + ], + "type": "image", + "image_path": "9cfc433aa299af5c76cfa9c52f2973827c84038a5ba4ce6b63a0ded4f66cfe2f.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 308, + 396, + 404, + 405 + ], + "lines": [ + { + "bbox": [ + 308, + 396, + 404, + 405 + ], + "spans": [ + { + "bbox": [ + 308, + 396, + 404, + 405 + ], + "type": "text", + "content": "\"Turn the bear into a panda\"" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 416, + 308, + 533, + 396 + ], + "blocks": [ + { + "bbox": [ + 416, + 308, + 533, + 396 + ], + "lines": [ + { + "bbox": [ + 416, + 308, + 533, + 396 + ], + "spans": [ + { + "bbox": [ + 416, + 308, + 533, + 396 + ], + "type": "image", + "image_path": "c78ef437ca4fbccac2f8d4d19a9efbe17cf57011f3eca3f4e9ea752b8f25adfc.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 417, + 396, + 531, + 405 + ], + "lines": [ + { + "bbox": [ + 417, + 396, + 531, + 405 + ], + "spans": [ + { + "bbox": [ + 417, + 396, + 531, + 405 + ], + "type": "text", + "content": "\"Turn the bear into a grizzly bear\"" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 47, + 423, + 287, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 423, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 47, + 423, + 287, + 651 + ], + "type": "text", + "content": "color-changing ability, while the third row, \"Make him completely bald\", showcases capabilities of retexturing and slight geometry editing. The second row in the second column demonstrates precise editing ability by exclusively editing the left side of the face. Based on that, we achieve editing in the third column, focusing on the right side of the face, showcasing the ability of multi-round edits, and accurately fulfilling user instructions. Fig. 4 showcases the precise editing capabilities in open 3D scenes. In the upper portion, the bicycle scene allows us to accurately locate the position of the road and edit its texture, transforming it into the grass, a river. In the experiment where we change the texture to a river, our method accurately constructs the reflection, making it appear realistic. Based on editing the road into a river, we further edited the bench, proving that our method can achieve multiple rounds of editing. The lower portion demonstrates the results of editing the bear, which fully preserves the original appearance of the background area and focuses the edits on the bear." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 47, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 653, + 287, + 715 + ], + "type": "text", + "content": "Comparisons with Instruct-NeRF2NeRF. Fig. 5 compares the results of our method with those of IN2N [11], on the scenes presented in IN2N. From the figure, it is evident that our method changes the texture of the pants without affecting the clothes, and vice versa, demonstrating" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 423, + 545, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 545, + 508 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 545, + 508 + ], + "type": "text", + "content": "the effectiveness of our method in distinguishing different objects within the foreground. Additionally, when editing the clothes and pants, the background remains unaffected, indicating our method's effective separation of foreground and background. Furthermore, the last column reveals that IN2N, limited by 2D diffusion, distorts the face, while our method maintains a superior rendering quality of faces." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 510, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 546, + 714 + ], + "type": "text", + "content": "Complex Multi-Object Scenes. Furthermore, we present the results of our editing in a complex scene featuring multiple objects, as depicted in Fig 6. We specifically selected three distinct object types for editing purposes. The first object type is the background. In a particular sub-scene, the desktop shares similarities with the background. Employing a caption-based approach, we successfully transformed the desktop into a wooden material. As evident in the image, the edited result exhibits a distinct and pronounced wood texture. The second object type is a foreground object, the flowerpot. We opted to change the color of the flowerpot to red, and the outcome was highly successful. Lastly, the most intricate editing task involved the rolling pin, which was occluded by multiple objects from various perspectives. As illustrated in the lower right corner of the picture, we managed to edit it into a cucumber without any adverse impact on the other objects." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20907" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 72, + 93, + 190 + ], + "blocks": [ + { + "bbox": [ + 49, + 72, + 93, + 190 + ], + "lines": [ + { + "bbox": [ + 49, + 72, + 93, + 190 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 93, + 190 + ], + "type": "image", + "image_path": "708aec49c9cafdd3fe8a679bbe324ff268d481f8fb6ad070ec7bf7239ee84e84.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 93, + 128, + 108, + 135 + ], + "lines": [ + { + "bbox": [ + 93, + 128, + 108, + 135 + ], + "spans": [ + { + "bbox": [ + 93, + 128, + 108, + 135 + ], + "type": "text", + "content": "IN2N" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 108, + 72, + 197, + 190 + ], + "blocks": [ + { + "bbox": [ + 108, + 72, + 197, + 190 + ], + "lines": [ + { + "bbox": [ + 108, + 72, + 197, + 190 + ], + "spans": [ + { + "bbox": [ + 108, + 72, + 197, + 190 + ], + "type": "image", + "image_path": "034e98c8f141d54b5efd4f1d503b90feb7866de98e04acbcdeeda0ba4dd087cc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 199, + 72, + 288, + 190 + ], + "blocks": [ + { + "bbox": [ + 199, + 72, + 288, + 190 + ], + "lines": [ + { + "bbox": [ + 199, + 72, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 199, + 72, + 288, + 190 + ], + "type": "image", + "image_path": "7a72df2a250f22c3c24084e86a09ea16191a5a3aab175c256eb9a0f3db5d0367.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 290, + 72, + 378, + 190 + ], + "blocks": [ + { + "bbox": [ + 290, + 72, + 378, + 190 + ], + "lines": [ + { + "bbox": [ + 290, + 72, + 378, + 190 + ], + "spans": [ + { + "bbox": [ + 290, + 72, + 378, + 190 + ], + "type": "image", + "image_path": "06bbde5350ffd895dbcc5ee14c1d5b5fd62bf26eaf71128c623ad62979a4033f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 381, + 72, + 544, + 190 + ], + "blocks": [ + { + "bbox": [ + 381, + 72, + 544, + 190 + ], + "lines": [ + { + "bbox": [ + 381, + 72, + 544, + 190 + ], + "spans": [ + { + "bbox": [ + 381, + 72, + 544, + 190 + ], + "type": "image", + "image_path": "66740c2206392f85d313bdc12fbe63c287529d8cd856b75f4b37a2002a4be59c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 49, + 191, + 93, + 306 + ], + "blocks": [ + { + "bbox": [ + 49, + 191, + 93, + 306 + ], + "lines": [ + { + "bbox": [ + 49, + 191, + 93, + 306 + ], + "spans": [ + { + "bbox": [ + 49, + 191, + 93, + 306 + ], + "type": "image", + "image_path": "0c6c750adaf9a8ef3d823ff004ac4cbf87ddd3cf348b876acb87c895ec71422b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 307, + 83, + 314 + ], + "lines": [ + { + "bbox": [ + 59, + 307, + 83, + 314 + ], + "spans": [ + { + "bbox": [ + 59, + 307, + 83, + 314 + ], + "type": "text", + "content": "Original" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 93, + 245, + 107, + 251 + ], + "lines": [ + { + "bbox": [ + 93, + 245, + 107, + 251 + ], + "spans": [ + { + "bbox": [ + 93, + 245, + 107, + 251 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 108, + 191, + 196, + 307 + ], + "blocks": [ + { + "bbox": [ + 108, + 191, + 196, + 307 + ], + "lines": [ + { + "bbox": [ + 108, + 191, + 196, + 307 + ], + "spans": [ + { + "bbox": [ + 108, + 191, + 196, + 307 + ], + "type": "image", + "image_path": "2b4c37b005bca97a16961b72f46b18b0e4f787d3ff5863caf159d08208a9040b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 116, + 308, + 185, + 314 + ], + "lines": [ + { + "bbox": [ + 116, + 308, + 185, + 314 + ], + "spans": [ + { + "bbox": [ + 116, + 308, + 185, + 314 + ], + "type": "text", + "content": "\"Turn his pants into yellow.\"" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 198, + 191, + 288, + 307 + ], + "blocks": [ + { + "bbox": [ + 198, + 191, + 288, + 307 + ], + "lines": [ + { + "bbox": [ + 198, + 191, + 288, + 307 + ], + "spans": [ + { + "bbox": [ + 198, + 191, + 288, + 307 + ], + "type": "image", + "image_path": "db5a6d9d9b23e1a617def74bbd5044e16fad1d6ac3ef06bcdc460ab648d50e26.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 198, + 308, + 289, + 314 + ], + "lines": [ + { + "bbox": [ + 198, + 308, + 289, + 314 + ], + "spans": [ + { + "bbox": [ + 198, + 308, + 289, + 314 + ], + "type": "text", + "content": "Turn his pants into a bronze statue." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 290, + 191, + 378, + 307 + ], + "blocks": [ + { + "bbox": [ + 290, + 191, + 378, + 307 + ], + "lines": [ + { + "bbox": [ + 290, + 191, + 378, + 307 + ], + "spans": [ + { + "bbox": [ + 290, + 191, + 378, + 307 + ], + "type": "image", + "image_path": "89f0d60fa75b90e235ec080c6df3f97077da368e655a292c816379632ec0975f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 301, + 308, + 370, + 314 + ], + "lines": [ + { + "bbox": [ + 301, + 308, + 370, + 314 + ], + "spans": [ + { + "bbox": [ + 301, + 308, + 370, + 314 + ], + "type": "text", + "content": "\"Turn his t-shirt into white.\"" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 382, + 191, + 544, + 307 + ], + "blocks": [ + { + "bbox": [ + 382, + 191, + 544, + 307 + ], + "lines": [ + { + "bbox": [ + 382, + 191, + 544, + 307 + ], + "spans": [ + { + "bbox": [ + 382, + 191, + 544, + 307 + ], + "type": "image", + "image_path": "9452d962b82a94bf81a3ddd21526532d1f6813bcb7e144e52a288018fbb825e8.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 387, + 308, + 466, + 314 + ], + "lines": [ + { + "bbox": [ + 387, + 308, + 466, + 314 + ], + "spans": [ + { + "bbox": [ + 387, + 308, + 466, + 314 + ], + "type": "text", + "content": "'Make his t-shirt out of leather" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 50, + 336, + 174, + 418 + ], + "blocks": [ + { + "bbox": [ + 113, + 317, + 478, + 329 + ], + "lines": [ + { + "bbox": [ + 113, + 317, + 478, + 329 + ], + "spans": [ + { + "bbox": [ + 113, + 317, + 478, + 329 + ], + "type": "text", + "content": "Figure 5. Comparisons with Instruct-NeRF2NeRF (IN2N) [11] on the scene presented in their paper." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 336, + 174, + 418 + ], + "lines": [ + { + "bbox": [ + 50, + 336, + 174, + 418 + ], + "spans": [ + { + "bbox": [ + 50, + 336, + 174, + 418 + ], + "type": "image", + "image_path": "2b5812dd0e21bdfb7a1158c73e58d03e6c059d64c65171c905f4568749e01f5e.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 96, + 420, + 129, + 430 + ], + "lines": [ + { + "bbox": [ + 96, + 420, + 129, + 430 + ], + "spans": [ + { + "bbox": [ + 96, + 420, + 129, + 430 + ], + "type": "text", + "content": "Original" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 433, + 544, + 454 + ], + "lines": [ + { + "bbox": [ + 46, + 433, + 544, + 454 + ], + "spans": [ + { + "bbox": [ + 46, + 433, + 544, + 454 + ], + "type": "text", + "content": "Figure 6. Qualitative results on complex multi-object scenes. The background \"desk\", the foreground \"flower pot\", and the multiview blocked foreground \"rolling pin\" are edited separately." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 175, + 336, + 293, + 418 + ], + "blocks": [ + { + "bbox": [ + 175, + 336, + 293, + 418 + ], + "lines": [ + { + "bbox": [ + 175, + 336, + 293, + 418 + ], + "spans": [ + { + "bbox": [ + 175, + 336, + 293, + 418 + ], + "type": "image", + "image_path": "7d785d265676c476a31937017422b796380e02da12b72a22bef1e420abbe5ab6.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 188, + 420, + 280, + 429 + ], + "lines": [ + { + "bbox": [ + 188, + 420, + 280, + 429 + ], + "spans": [ + { + "bbox": [ + 188, + 420, + 280, + 429 + ], + "type": "text", + "content": "\"Turn the desk into wood\"" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 294, + 336, + 416, + 418 + ], + "blocks": [ + { + "bbox": [ + 294, + 336, + 416, + 418 + ], + "lines": [ + { + "bbox": [ + 294, + 336, + 416, + 418 + ], + "spans": [ + { + "bbox": [ + 294, + 336, + 416, + 418 + ], + "type": "image", + "image_path": "d203d78c3c3f4986abb31e1dced7ad8764bb2146549371e508739125abc1f1f7.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 420, + 399, + 430 + ], + "lines": [ + { + "bbox": [ + 313, + 420, + 399, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 420, + 399, + 430 + ], + "type": "text", + "content": "\"Turn the flowerpot red\"" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 414, + 420, + 542, + 430 + ], + "lines": [ + { + "bbox": [ + 414, + 420, + 542, + 430 + ], + "spans": [ + { + "bbox": [ + 414, + 420, + 542, + 430 + ], + "type": "text", + "content": "\"Turn the rolling pin into cucumber\"" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 417, + 336, + 539, + 418 + ], + "blocks": [ + { + "bbox": [ + 417, + 336, + 539, + 418 + ], + "lines": [ + { + "bbox": [ + 417, + 336, + 539, + 418 + ], + "spans": [ + { + "bbox": [ + 417, + 336, + 539, + 418 + ], + "type": "image", + "image_path": "b7703d6d142d19a1d51628333f931145036b637db4645123ac1bfbae28d90ade.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "table", + "bbox": [ + 48, + 487, + 284, + 546 + ], + "blocks": [ + { + "bbox": [ + 47, + 459, + 286, + 481 + ], + "lines": [ + { + "bbox": [ + 47, + 459, + 286, + 481 + ], + "spans": [ + { + "bbox": [ + 47, + 459, + 286, + 481 + ], + "type": "text", + "content": "Table 1. Quantitative evaluation on the bicycle scene of the MipNeRF360 dataset [3]." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 487, + 284, + 546 + ], + "lines": [ + { + "bbox": [ + 48, + 487, + 284, + 546 + ], + "spans": [ + { + "bbox": [ + 48, + 487, + 284, + 546 + ], + "type": "table", + "html": "
MethodsCTIDS↑IIS↑FID↓Time↓
IN2N [11]0.220.8510351 min
Ours-DVGO [45]0.110.8214840 min
Ours-3DGS0.280.955120 min
", + "image_path": "cc3a22842e7d83d77e4e98d5b2fb30be29fa8385e559d7193d303bfebc4109f0.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "table_body" + } + ], + "index": 28 + }, + { + "bbox": [ + 47, + 556, + 182, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 556, + 182, + 569 + ], + "spans": [ + { + "bbox": [ + 47, + 556, + 182, + 569 + ], + "type": "text", + "content": "4.3. Quantitative Evaluation" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 46, + 575, + 286, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 575, + 286, + 671 + ], + "spans": [ + { + "bbox": [ + 46, + 575, + 286, + 671 + ], + "type": "text", + "content": "Metric Comparisons. Table 1 shows quantitative results on the bicycle scene of the Mip-NeRF360 dataset [3], comparing with IN2N [11] and Direct Voxel Grid Optimization (DVGO) [45] as the representation. The metrics include CLIP Text-Image Direction Similarity (CTIDS), Image-Image Similarity (IIS), FID, and training time. GaussianEditor achieves the best results in all metrics. The test data is shown in the supplementary material." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": "User Study. We perform a user study comparing with IN2N on the bear scene in Fig. 4 and the human scene in Fig. 5, involving 21 participants. GaussianEditor gets a" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 306, + 460, + 518, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 460, + 518, + 473 + ], + "spans": [ + { + "bbox": [ + 306, + 460, + 518, + 473 + ], + "type": "text", + "content": "87.07% voting percentage, while IN2N gets 12.93%." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 306, + 479, + 462, + 492 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 462, + 492 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 462, + 492 + ], + "type": "text", + "content": "4.4. Ablation Study and Analysis" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": "Ablation of Gaussian RoI, Text RoI, RoI Lifting. To validate the effectiveness of each module in our framework, we design three variant approaches: (1) w/o Gaussian RoI: We discontinued the use of Gaussian RoI " + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{RoI}" + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": " to control the gradients of Gaussian points, as mentioned in Eq. 14. (2) w/o Text ROI: In this scenario, we ceased the selection of text ROI " + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{RoI}" + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": " using LLM assistant " + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{LLM}" + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": ". Instead, all the words in the user's instruction are put to " + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{Seg}" + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": " to get " + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{RoI}" + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": ". (3) w/o RoI lifting: Instead of lifting the image RoI " + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{RoI}" + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": " to 3D Gaussians, the image RoI " + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{RoI}" + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": " is used to govern the calculation of the loss. That is, only the pixels within the image RoI " + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{RoI}" + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": " are taken into account for the loss computation. Fig. 7 showcases the outcomes of our ablation experiment, which aimed to edit the doll based on the instruction \"Turn its mouth into red.\" The results reveal the following findings. (1) When the Gaussian RoI is not used, the 3D scene is all turned red because the 2D diffusion fails to control the editing area. (2) In cases where" + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20908" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 70, + 277, + 163 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 277, + 163 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 277, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 277, + 163 + ], + "type": "image", + "image_path": "2f3b4558d05a720cdccebd35f1b2fa8dc23fc0ace4f66a8d90c517a280250ea9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 97, + 166, + 236, + 176 + ], + "lines": [ + { + "bbox": [ + 97, + 166, + 236, + 176 + ], + "spans": [ + { + "bbox": [ + 97, + 166, + 236, + 176 + ], + "type": "text", + "content": "Figure 7. Ablation experiment of RoI." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 179, + 289, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 179, + 289, + 322 + ], + "spans": [ + { + "bbox": [ + 46, + 179, + 289, + 322 + ], + "type": "text", + "content": "text ROI " + }, + { + "bbox": [ + 46, + 179, + 289, + 322 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{Roi}" + }, + { + "bbox": [ + 46, + 179, + 289, + 322 + ], + "type": "text", + "content": " is not utilized, the grounding segmentation model tends to segment the entire foreground object, leading to the doll being entirely edited to red. (3) When RoI lifting " + }, + { + "bbox": [ + 46, + 179, + 289, + 322 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{Lift}" + }, + { + "bbox": [ + 46, + 179, + 289, + 322 + ], + "type": "text", + "content": " is not employed, the doll's mouth is successfully turned red, but other facial areas are also affected. Because the grounding segmentation model may fail to parse specific views, noise exists on the image RoI " + }, + { + "bbox": [ + 46, + 179, + 289, + 322 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{Roi}" + }, + { + "bbox": [ + 46, + 179, + 289, + 322 + ], + "type": "text", + "content": ". Consequently, leakage occurs during the editing process. Our proposed RoI lifting module effectively addresses this issue during training. In conclusion, our ablation experiment demonstrates the effectiveness of several RoI-related modules in our method." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 327, + 289, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 289, + 649 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 289, + 649 + ], + "type": "text", + "content": "Ablation of Scene Description Generation. We further conduct experiments to evaluate the role of scene description generation, employing three distinct experimental setups. The first one composes the user message " + }, + { + "bbox": [ + 46, + 327, + 289, + 649 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{user}" + }, + { + "bbox": [ + 46, + 327, + 289, + 649 + ], + "type": "text", + "content": ", without employing scene description. The second method randomly samples a view and extracts the corresponding image's text description as the scene description. The third one represents the complete version of our approach. The test scene involves a park where a bike and a bench are positioned closely together. The editing instruction is \"Turn the thing next to the bike orange\". The obtained results are presented in Fig. 8. As shown in the image, when scene description is not employed, the LLM fails to acquire the Text ROI according to the user instructions, resulting in editing failure. The second one randomly samples images to obtain scene descriptions, resulting in incomplete descriptions and leading to an incorrect text ROI prediction by the LLM. Consequently, the final editing result turns the road into an orange color. In contrast, our method flawlessly executes the editing task. This success can be attributed to scene description generation, which obtains an accurate text description encompassing the relative positional relationship between the bicycle and the bench. This enables the LLM to analyze and determine the user's intention to edit the bench. Consequently, the desired color change of the bench is successfully implemented." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 658, + 125, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 125, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 125, + 670 + ], + "type": "text", + "content": "4.5. Limitations" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "Although our framework has solved some problems inherited from the integrated sub-modules, e.g. noise in the results of grounding segmentation, there are still some prob" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 313, + 70, + 537, + 370 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 537, + 370 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 537, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 537, + 370 + ], + "type": "image", + "image_path": "4120a7d7f36014858efa3395d672fbe704f351f111af9ab379803e57e7f07d4e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 374, + 544, + 385 + ], + "lines": [ + { + "bbox": [ + 306, + 374, + 544, + 385 + ], + "spans": [ + { + "bbox": [ + 306, + 374, + 544, + 385 + ], + "type": "text", + "content": "Figure 8. Ablation results about the scene description generation." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 391, + 545, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 391, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 304, + 391, + 545, + 511 + ], + "type": "text", + "content": "lems that the current system cannot completely avoid. In scene description generation, the descriptions from different views of the same object may differ from each other. When the differences are large enough, the LLM may misunderstand these descriptions as those from multiple objects. This issue does not affect the results in the current experiment, but we would like to optimize this in the future. In addition, our system cannot achieve good editing results in scenes where the grounding segmentation or diffusion model completely fails, such as drastic geometric editing." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 525, + 378, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 525, + 378, + 537 + ], + "spans": [ + { + "bbox": [ + 306, + 525, + 378, + 537 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "content": "This paper proposes a systematic framework, named GaussianEditor, for text-guided delicate 3D scene editing. As we know, GaussianEditor is one of the first works to edit 3D Gaussians, taking advantage of the explicit property of 3D Gaussians and making it easy to control the editing area precisely. Several techniques are proposed to achieve delicate editing, including extracting instruction RoI from texts, aligning the RoI to 3D Gaussians, and editing the scene with the Gaussian RoI. GaussianEditor achieves notably more delicate editing results than IN2N [11] with much shorter training time (within 20 minutes v.s. 45 minutes - 2 hours). Noticing recent works [27, 49] have extended Gaussian splatting to dynamic scenes, we leave the delicate editing in dynamic scenes as future work." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20909" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "text", + "content": "[1] Chong Bao, Yinda Zhang, Bangbang Yang, Tianxing Fan, Zesong Yang, Hujun Bao, Guofeng Zhang, and Zhaopeng Cui. Sine: Semantic-driven image-based nerf editing with prior-guided editing field. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 158, + 288, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 288, + 202 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 288, + 202 + ], + "type": "text", + "content": "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 204, + 288, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 204, + 288, + 236 + ], + "spans": [ + { + "bbox": [ + 53, + 204, + 288, + 236 + ], + "type": "text", + "content": "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022. 2, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 237, + 287, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 237, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 53, + 237, + 287, + 270 + ], + "type": "text", + "content": "[4] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 2, 3, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 272, + 287, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 272, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 53, + 272, + 287, + 315 + ], + "type": "text", + "content": "[5] Jiazhong Cen, Zanwei Zhou, Jiemin Fang, Chen Yang, Wei Shen, Lingxi Xie, Dongsheng Jiang, Xiaopeng Zhang, and Qi Tian. Segment anything in 3d with nerfs. In NeurIPS, 2023. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 317, + 287, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 317, + 287, + 349 + ], + "spans": [ + { + "bbox": [ + 53, + 317, + 287, + 349 + ], + "type": "text", + "content": "[6] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In ECCV, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 351, + 287, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 351, + 287, + 406 + ], + "spans": [ + { + "bbox": [ + 53, + 351, + 287, + 406 + ], + "type": "text", + "content": "[7] Pei-Ze Chiang, Meng-Shiun Tsai, Hung-Yu Tseng, Wei-Sheng Lai, and Wei-Chen Chiu. Stylizing 3d scene via implicit representation and hypernetwork. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 407, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 407, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 407, + 287, + 441 + ], + "type": "text", + "content": "[8] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 442, + 287, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 442, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 53, + 442, + 287, + 485 + ], + "type": "text", + "content": "[9] Shuangkang Fang, Yufeng Wang, Yi Yang, Yi Yang, Yi-Hsuan Tsai, Wenrui Ding, Ming-Hsuan Yang, and Shuchang Zhou. Text-driven editing of 3d scenes without retraining. Arxiv preprint arXiv:2309.04917, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 487, + 287, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 487, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 48, + 487, + 287, + 531 + ], + "type": "text", + "content": "[10] William Gao, Noam Aigerman, Thibault Groueix, Vova Kim, and Rana Hanocka. Textdeformer: Geometry manipulation using text guidance. In ACM SIGGRAPH 2023 Conference Proceedings, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 533, + 287, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 287, + 575 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 287, + 575 + ], + "type": "text", + "content": "[11] Ayaan Haque, Matthew Tancik, Alexei Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. In CVPR, 2023. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 577, + 287, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 577, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 48, + 577, + 287, + 620 + ], + "type": "text", + "content": "[12] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 622, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 622, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 622, + 287, + 644 + ], + "type": "text", + "content": "[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "type": "text", + "content": "[14] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. The Journal of Machine Learning Research, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[15] Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. Avatarclip: Zero-shot text" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "driven generation and animation of 3d avatars. arXiv preprint arXiv:2205.08535, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "type": "text", + "content": "[16] Hsin-Ping Huang, Hung-Yu Tseng, Saurabh Saini, Maneesh Singh, and Ming-Hsuan Yang. Learning to stylize novel views. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 141, + 545, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 545, + 174 + ], + "type": "text", + "content": "[17] Yi-Hua Huang, Yue He, Yu-Jie Yuan, Yu-Kun Lai, and Lin Gao. Stylizednerf: consistent 3d scene stylization as stylized nerf via 2d-3d mutual learning. In CVPR, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 175, + 545, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 175, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 308, + 175, + 545, + 218 + ], + "type": "text", + "content": "[18] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 2023. 2, 3, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 220, + 545, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 545, + 263 + ], + "type": "text", + "content": "[19] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 265, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 265, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 265, + 545, + 308 + ], + "type": "text", + "content": "[20] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nef for editing via feature field distillation. Advances in Neural Information Processing Systems, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 309, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 309, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 308, + 309, + 545, + 354 + ], + "type": "text", + "content": "[21] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 355, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 355, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 308, + 355, + 545, + 398 + ], + "type": "text", + "content": "[22] Yuan Li, Zhi-Hao Lin, David Forsyth, Jia-Bin Huang, and Shenlong Wang. Climatenerf: Physically-based neural rendering for extreme climate synthesis. arXiv e-prints, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 399, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 443 + ], + "type": "text", + "content": "[23] Yuhan Li, Yishun Dou, Yue Shi, Yu Lei, Xuanhong Chen, Yi Zhang, Peng Zhou, and Bingbing Ni. Focaldreamer: Text-driven 3d editing via focal-fusion assembly. arXiv preprint arXiv:2308.10608, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 445, + 545, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 445, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 545, + 477 + ], + "type": "text", + "content": "[24] Hao-Kang Liu, I Shen, Bing-Yu Chen, et al. Nerf-in: Free-form nerf inpainting with rgb-d priors. arXiv preprint arXiv:2206.04901, 2022." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 479, + 545, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 479, + 545, + 522 + ], + "spans": [ + { + "bbox": [ + 308, + 479, + 545, + 522 + ], + "type": "text", + "content": "[25] Steven Liu, Xiuming Zhang, Zhoutong Zhang, Richard Zhang, Jun-Yan Zhu, and Bryan Russell. Editing conditional radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 523, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 523, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 523, + 545, + 578 + ], + "type": "text", + "content": "[26] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 579, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 579, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 308, + 579, + 545, + 623 + ], + "type": "text", + "content": "[27] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 624, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 668 + ], + "type": "text", + "content": "[28] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "type": "text", + "content": "[29] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 2021. 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20910" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[30] Ashkan Mirzaei, Tristan Aumentado-Armstrong, Marcus A. Brubaker, Jonathan Kelly, Alex Levinshtein, Konstantinos G. Derpanis, and Igor Gilitschenski. Watch your steps: Local image and scene editing by text instructions. In arXiv preprint arXiv:2308.08947, 2023. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 287, + 172 + ], + "type": "text", + "content": "[31] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 287, + 205 + ], + "type": "text", + "content": "[32] Thu Nguyen-Phuoc, Feng Liu, and Lei Xiao. Snerf: stylized neural implicit representations for 3d scenes. arXiv preprint arXiv:2207.02363, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 205, + 287, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 205, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 205, + 287, + 259 + ], + "type": "text", + "content": "[33] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "type": "text", + "content": "[34] Atsuhiro Noguchi, Xiao Sun, Stephen Lin, and Tatsuya Harada. Neural articulated radiance field. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 304, + 287, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 304, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 287, + 358 + ], + "type": "text", + "content": "[35] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 2019. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 360, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 287, + 403 + ], + "type": "text", + "content": "[36] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 404, + 288, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 288, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 288, + 459 + ], + "type": "text", + "content": "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 459, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 459, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 459, + 287, + 502 + ], + "type": "text", + "content": "[38] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 503, + 287, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 503, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 48, + 503, + 287, + 548 + ], + "type": "text", + "content": "[39] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 548, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 613 + ], + "type": "text", + "content": "[40] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 614, + 287, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 658 + ], + "type": "text", + "content": "[41] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 690 + ], + "type": "text", + "content": "[42] Sara Fridovich-Keil and Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 690, + 287, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 690, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 690, + 287, + 715 + ], + "type": "text", + "content": "[43] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 589 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "nonequilibrium thermodynamics. In International conference on machine learning, 2015. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 95, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 545, + 129 + ], + "type": "text", + "content": "[44] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 129, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 162 + ], + "type": "text", + "content": "[45] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, 2022. 2, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 163, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 207 + ], + "type": "text", + "content": "[46] Vadim Tschernezki, Iro Laina, Diane Larlus, and Andrea Vedaldi. Neural feature fusion fields: 3d distillation of self-supervised 2d image representations. In 2022 International Conference on 3D Vision (3DV), 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 208, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 545, + 262 + ], + "type": "text", + "content": "[47] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 264, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 297 + ], + "type": "text", + "content": "[48] Can Wang, Ruixiang Jiang, Mengei Chai, Mingming He, Dongdong Chen, and Jing Liao. Nerf-art: Text-driven neural radiance fields stylization. TVCG, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 297, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 545, + 342 + ], + "type": "text", + "content": "[49] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Wang Xinggang. 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528, 2023. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 342, + 545, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 545, + 375 + ], + "type": "text", + "content": "[50] Qiling Wu, Jianchao Tan, and Kun Xu. Palettenerf: Palette-based color editing for nerfs. arXiv preprint arXiv:2212.12871, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 376, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 420 + ], + "type": "text", + "content": "[51] Jiale Xu, Xintao Wang, Yan-Pei Cao, Weihao Cheng, Ying Shan, and Shenghua Gao. Instructp2p: Learning to edit 3d point clouds with text instructions. arXiv preprint arXiv:2306.07154, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 421, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 421, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 421, + 545, + 453 + ], + "type": "text", + "content": "[52] Tianhan Xu and Tatsuya Harada. Deforming radiance fields with cages. In European Conference on Computer Vision, 2022." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 455, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 455, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 455, + 545, + 510 + ], + "type": "text", + "content": "[53] Bangbang Yang, Chong Bao, Junyi Zeng, Hujun Bao, Yinda Zhang, Zhaopeng Cui, and Guofeng Zhang. Neumesh: Learning disentangled neural mesh-based implicit field for geometry and texture editing. In European Conference on Computer Vision, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 510, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 510, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 510, + 545, + 555 + ], + "type": "text", + "content": "[54] Kai Zhang, Nick Kolkin, Sai Bi, Fujun Luan, Zexiang Xu, Eli Shechtman, and Noah Snavely. Arf: Artistic radiance fields. In European Conference on Computer Vision, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 555, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 555, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 307, + 555, + 545, + 589 + ], + "type": "text", + "content": "[55] Jingyu Zhuang, Chen Wang, Lingjie Liu, Liang Lin, and Guanbin Li. Dreameditor: Text-driven 3d scene editing with neural fields. arXiv preprint arXiv:2306.13455, 2023. 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20911" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/GaussianEditor_ Swift and Controllable 3D Editing with Gaussian Splatting/f458e425-2ba9-468a-94e0-95be47f99560_content_list.json b/2024/GaussianEditor_ Swift and Controllable 3D Editing with Gaussian Splatting/f458e425-2ba9-468a-94e0-95be47f99560_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..668b174b44b14c325e5b4edd232728828d5e07c3 --- /dev/null +++ b/2024/GaussianEditor_ Swift and Controllable 3D Editing with Gaussian Splatting/f458e425-2ba9-468a-94e0-95be47f99560_content_list.json @@ -0,0 +1,2277 @@ +[ + { + "type": "text", + "text": "GaussianEditor: Swift and Controllable 3D Editing with Gaussian Splitting", + "text_level": 1, + "bbox": [ + 101, + 130, + 869, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yiwen Chen\\*1,2 Yikai Wang3", + "bbox": [ + 145, + 178, + 277, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zilong Chen\\*3 Zhongang Cai4", + "bbox": [ + 281, + 180, + 418, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chi Zhang $^{2}$ Feng Wang $^{3}$ Lei Yang $^{4}$ Huaping Liu $^{3}$", + "bbox": [ + 433, + 180, + 666, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiaofeng Yang $^{2}$ Guosheng Lin $^{**1,2}$", + "bbox": [ + 679, + 180, + 831, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ S-Lab, Nanyang Technological University", + "bbox": [ + 313, + 223, + 655, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ School of Computer Science and Engineering, Nanyang Technological University", + "bbox": [ + 160, + 241, + 808, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Department of Computer Science and Technology, Tsinghua University", + "bbox": [ + 200, + 258, + 769, + 277 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 SenseTime Research", + "bbox": [ + 398, + 277, + 570, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://buaacyw.github.io/gaussian-editor", + "bbox": [ + 279, + 296, + 683, + 310 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b1755dd51b2004f2668c564909ff8df1419c4f5ec82753cd6286fdac4e74143c.jpg", + "image_caption": [ + "Original view" + ], + "image_footnote": [], + "bbox": [ + 83, + 327, + 212, + 395 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1b317d9448d69eddea2abdff283e407cd469970daf35dd808042852c4e1a8385.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 215, + 327, + 344, + 395 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/87f5389d17dc044a89bf047023a9ee8265a427318ac2923cf56450f7914413c6.jpg", + "image_caption": [ + "Original view" + ], + "image_footnote": [], + "bbox": [ + 81, + 412, + 212, + 489 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4d01b8eb2a4c0777f8d5bf2daab7928aba67db5ecc08b91fe83e9b49a67c7ad6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 214, + 412, + 343, + 489 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fd0245785b004ba71a77391a241d951b1ce1e46f9d3e08894430bb300626f5a4.jpg", + "image_caption": [ + "\"Delete the bike. Add a dog on the bench.\"" + ], + "image_footnote": [], + "bbox": [ + 352, + 327, + 483, + 395 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0f9f38f043f6dd091e6966a5d2cc8d1f9658026ec88df06a0c39d6b3352af481.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 485, + 327, + 617, + 396 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c980c3a12f5374acd59b22c36dc54b6cab85718f61652d0c78e4ee60cd0c0829.jpg", + "image_caption": [ + "\"Make the grass on fire\"" + ], + "image_footnote": [], + "bbox": [ + 622, + 327, + 754, + 395 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a1275c47cdc9ddf269efd347734e12ff216b4c517ebe38ac7b5f27711707913a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 756, + 327, + 888, + 396 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c947c3270de861600e875163c95f6d05763b917f5a067cfa60bea57c2bbd282d.jpg", + "image_caption": [ + "\"Turn him into Hulk\"" + ], + "image_footnote": [], + "bbox": [ + 351, + 412, + 483, + 489 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d184213e68d0eb12d53708a6ed46c1632bb9d7ffee55bc7e3d003ab23bddd464.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 412, + 614, + 489 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2383e2c6b65793deef4a5cebd936f0f0b1aaeb237f7a35c247ee43349c135486.jpg", + "image_caption": [ + "\"Turn him into a clown\"" + ], + "image_footnote": [], + "bbox": [ + 622, + 412, + 754, + 489 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c4f73d96cc8bd4d4cc9b9a8a6a2897d066242e93be424638cbc3109d787e8f4e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 756, + 412, + 887, + 489 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a8e74e43b88a4ca5fc855f2a74b589bc1ef7ff20e4fd83a6557ae8a58c5073d8.jpg", + "image_caption": [ + "Figure 1. Results of GaussianEditor. GaussianEditor offers swift, controllable, and versatile 3D editing. A single editing session only takes 5-10 minutes. Please note our precise editing control, where only the desired parts are modified. Taking the \"Make the grass on fire\" example from the first row of the figure, other objects in the scene such as the bench and tree remain unaffected." + ], + "image_footnote": [], + "bbox": [ + 81, + 511, + 212, + 577 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d95b7a9658e0fb1f0851be7d14f34bab34d7ab501329ce41fb7097cfbe1997fc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 214, + 511, + 344, + 577 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c214d612d42c940b93cc8e5d41358a7c0e0d7e084601fa11a1e48020919373ec.jpg", + "image_caption": [ + "\"Make it autumn\"" + ], + "image_footnote": [], + "bbox": [ + 352, + 511, + 483, + 577 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7ec76c98117228ac50a0ade37bf90207fff56e2df91b8a31ae089fa4ebdce523.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 485, + 511, + 616, + 577 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/111cc0f0d3f08c038633ad405e330215f71c3097dc84e5b5cdccda89ce852873.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 622, + 511, + 753, + 577 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/92e1a2207ff24cae03a4bfc2f9cf8c8b4337b92aef53f4c770be15d0eb17d716.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 756, + 511, + 887, + 577 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 684, + 312, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D editing plays a crucial role in many areas such as gaming and virtual reality. Traditional 3D editing methods, which rely on representations like meshes and point clouds, often fall short in realistically depicting complex scenes. On the other hand, methods based on implicit 3D representations, like Neural Radiance Field (NeRF), render complex scenes effectively but suffer from slow processing speeds and limited control over specific scene areas. In response to these challenges, our paper presents GaussianEditor, the first 3D editing algorithm based on Gaussian Splitting (GS), a novel", + "bbox": [ + 75, + 715, + 470, + 866 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D representation. GaussianEditor enhances precision and control in editing through our proposed Gaussian semantic tracing, which traces the editing target throughout the training process. Additionally, we propose Hierarchical Gaussian splatting (HGS) to achieve stabilized and fine results under stochastic generative guidance from 2D diffusion models. We also develop editing strategies for efficient object removal and integration, a challenging task for existing methods. Our comprehensive experiments demonstrate GaussianEditor's superior control, effective, and efficient performance, marking a significant advancement in 3D editing.", + "bbox": [ + 496, + 670, + 893, + 837 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "** Corresponding author.", + "bbox": [ + 99, + 875, + 238, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* The first two authors contributed equally to this work.", + "bbox": [ + 99, + 887, + 393, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21476", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 80, + 90, + 207, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the evolving field of computer vision, the development of user-friendly 3D representations and editing algorithms is a key objective. Such technologies are vital in various applications, ranging from digital gaming to the growing MetaVerse. Traditional 3D representations like meshes and point clouds have been preferred due to their interactive editing capabilities. However, these methods face challenges in accurately rendering complex 3D scenes.", + "bbox": [ + 80, + 114, + 467, + 234 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The recent rise of implicit 3D representations, exemplified by the Neural Radiance Field (NeRF) [28], represents a paradigm shift in 3D scene rendering. NeRF's capacity for high-fidelity rendering, coupled with its implicit nature that offers significant expansibility, marks a substantial improvement over conventional approaches[2, 32, 56]. This dual advantage has placed a significant focus on the NeRF framework in 3D editing [12, 31, 46, 47, 57], establishing it as a foundational approach for a considerable duration. However, NeRF's reliance on high-dimensional multilayer perception (MLP) networks for scene data encoding presents limitations. It restricts direct modification of specific scene parts and complicates tasks, like inpainting and scene composition. This complexity extends to the training and rendering processes, hindering practical applications.", + "bbox": [ + 80, + 236, + 467, + 460 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In light of these challenges, our research is focused on developing an advanced 3D editing algorithm. This algorithm aims for flexible and rapid editing of 3D scenes, integrating both implicit editing, like text-based editing, and explicit control, such as bounding box usage for specific area modifications. To achieve these goals, we choose Gaussian Splatting (GS) [15] for its real-time rendering and explicit point cloud-like representations.", + "bbox": [ + 80, + 462, + 467, + 582 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, editing Gaussian Splatting (GS) [15] faces distinct challenges. A primary issue is the absence of efficient methods to accurately identify target Gaussians, which is crucial for precise controllable editing. Moreover, it has been observed in [7, 45, 53] that optimizing Gaussian Splatting (GS) using highly random generative guidance like Score Distillation Sampling [36] poses significant challenges. One possible explanation is that, unlike implicit representations buffered by neural networks, GS is directly affected by the randomness in loss. Such direct exposure results in unstable updates, as the properties of Gaussians are directly changed during training. Besides, each training step of GS may involve updates to a vast number of Gaussian points. This process occurs without the moderating influence of neural network-style buffering mechanisms. As a result, the excessive fluidity of the 3D GS scene hinders its ability to converge to finely detailed results like implicit representations when trained with generative guidance.", + "bbox": [ + 80, + 583, + 467, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To counter these issues, in this work, we propose GaussianEditor , a novel, swift, and highly controllable 3D editing algorithm for Gaussian Splatting. GaussianEditor can fulfill", + "bbox": [ + 80, + 854, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "various high-quality editing needs within minutes. A key feature of our method is the introduction of Gaussian semantic tracing, which enables precise control over Gaussian Splitting (GS). Gaussian semantic tracing consistently identifies the Gaussians requiring editing at every moment during training. This contrasts with traditional 3D editing methods that often depend on static 2D or 3D masks. Such masks become less effective as the geometries and appearances of 3D models evolve during training. Gaussian semantic tracing is achieved by unprojecting 2D segmentation masks into 3D Gaussians and assigning each Gaussian a semantic tag. As the Gaussians evolve during training, these semantic tags enable the tracking of the specific Gaussians targeted for editing. Our Gaussian tracing algorithm ensures that only the targeted areas are modified, enabling precise and controllable editing.", + "bbox": [ + 503, + 90, + 890, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Additionally, to tackle the significant challenge of Gaussian Splatting (GS) struggling to fit fine results under highly random generative guidance, we propose a novel GS representation: hierarchical Gaussian splatting (HGS). In HGS, Gaussians are organized into generations based on their sequence in multiple densification processes during training. Gaussians formed in earlier densification stages are deemed older generations and are subject to stricter constraints, aimed at preserving their original state and thus reducing their mobility. Conversely, those formed in later stages are considered younger generations and are subjected to fewer or no constraints, allowing for more adaptability. HGS's design effectively moderates the fluidity of GS by imposing restrictions on older generations while preserving the flexibility of newer generations. This approach enables continuous optimization towards better outcomes, thereby simulating the buffering function achieved in implicit representations through neural networks. Our experiments also demonstrate that HGS is more adept at adapting to highly random generative guidance.", + "bbox": [ + 503, + 335, + 890, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, we have specifically designed a 3D inpainting algorithm for Gaussian Splatting (GS). As demonstrated in Fig. 1, we have successfully removed specific objects from scenes and seamlessly integrated new objects into designated areas. For object removal, we developed a specialized local repair algorithm that efficiently eliminates artifacts at the intersection of the object and the scene. For adding objects, we first request users to provide a prompt and a 2D inpainting mask for a particular view of the GS. Subsequently, we employ a 2D inpainting method to generate a single-view image of the object to be added. This image is then transformed into a coarse 3D mesh using image-to-3D conversion techniques. The 3D mesh is subsequently converted into the HGS representation and refined. Finally, this refined representation is concatenated into the original GS. The entire inpainting process described above is completed within 5 minutes.", + "bbox": [ + 503, + 643, + 890, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21477", + "bbox": [ + 478, + 946, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "GaussianEditor offers swift, controllable, and versatile 3D editing. A single editing session typically only takes 5-10 minutes, significantly faster than previous editing processes. Our contributions can be summarized in four aspects:", + "bbox": [ + 76, + 90, + 468, + 150 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We have introduced Gaussian semantic tracing, enabling more detailed and effective editing control.", + "2. We propose Hierarchical Gaussian Splitting (HGS), a novel GS representation capable of converging more stably to refined results under highly random generative guidance.", + "3. We have specifically designed a 3D inpainting algorithm for Gaussian Splatting, which allows swift removal and addition objects.", + "4. Extensive experiments demonstrate that our method surpasses previous 3D editing methods in terms of effectiveness, speed, and controllability." + ], + "bbox": [ + 76, + 151, + 468, + 332 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 76, + 347, + 225, + 362 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. 3D Representations", + "text_level": 1, + "bbox": [ + 76, + 372, + 263, + 388 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Various 3D representations have been proposed to address diverse 3D tasks. The groundbreaking work, Neural Radiance Fields (NeRF) [28], employs volumetric rendering and has gained popularity for enabling 3D optimization with only 2D supervision. However, optimizing NeRF can be time-consuming, despite its wide usage in 3D reconstruction [3, 6, 13, 21, 44] and generation [22, 35] tasks.", + "bbox": [ + 75, + 396, + 468, + 501 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While efforts have been made to accelerate NeRF training [29, 40], these approaches primarily focus on the reconstruction setting, leaving the generation setting less optimized. The common technique of spatial pruning does not effectively speed up the generation setting.", + "bbox": [ + 75, + 503, + 468, + 578 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, 3D Gaussian splatting [15] has emerged as an alternative 3D representation to NeRF, showcasing impressive quality and speed in 3D and 4D reconstruction tasks [15, 26, 48, 51, 52]. It has also attracted considerable research interest in the field of generation [7, 45, 53]. Its efficient differentiable rendering implementation and model design facilitate fast training without the need for spatial pruning.", + "bbox": [ + 75, + 579, + 468, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we pioneer the adaptation of 3D Gaussian splating to 3D editing tasks, aiming to achieve swift and controllable 3D editing, harnessing the advantages of this representation for the first time in this context.", + "bbox": [ + 75, + 700, + 468, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2.3D Editing", + "text_level": 1, + "bbox": [ + 76, + 770, + 196, + 786 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Editing neural fields is inherently challenging due to the intricate interplay between their shape and appearance. EditNeRF [24] stands as a pioneering work in this domain, as they edit both the shape and color of neural fields by conditioning them on latent codes. Additionally, some works [1, 10, 46, 47] leverage CLIP models to facilitate editing through the use of text prompts or reference images.", + "bbox": [ + 75, + 794, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Another line of research focuses on predefined template models or skeletons to support actions like re-positioning or re-rendering within specific categories [30, 33]. Geometry-based methods [20, 49, 50, 55] translate neural fields into meshes and synchronize mesh deformation with implicit fields. Additionally, 3D editing techniques involve combining 2D image manipulation, such as inpainting, with neural fields training [19, 23].", + "bbox": [ + 496, + 90, + 893, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Concurrent works [31, 57] leverage static 2D and 3D masks to constrain the edit area of NeRF. However, these approaches have their limitations because the training of 3D models is a dynamic process, and static masks cannot effectively constrain it. In contrast, our research employs Gaussian semantic tracing to track the target Gaussian throughout the entire training process.", + "bbox": [ + 496, + 212, + 893, + 319 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminary", + "text_level": 1, + "bbox": [ + 500, + 335, + 624, + 352 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. 3D Gaussian Splitting", + "text_level": 1, + "bbox": [ + 500, + 361, + 707, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "GS (Gaussian Splitting) [15] represents an explicit 3D scene using point clouds, where Gaussians are employed to depict the scene's structure. In this representation, every Gaussian is defined by a center point, denoted as $x$ , and a covariance matrix $\\Sigma$ . The center point $x$ is commonly known as the Gaussian's mean value:", + "bbox": [ + 496, + 385, + 890, + 474 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nG (x) = e ^ {- \\frac {1}{2} x ^ {T} \\Sigma^ {- 1} x}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 491, + 890, + 511 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The covariance matrix $\\Sigma$ can be decomposed into a rotation matrix $\\mathbf{R}$ and a scaling matrix $\\mathbf{S}$ for differentiable optimization:", + "bbox": [ + 496, + 522, + 893, + 565 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Sigma = \\mathbf {R} \\mathbf {S} \\mathbf {S} ^ {T} \\mathbf {R} ^ {T}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 638, + 569, + 890, + 585 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the calculation of gradient flow is detailed in [15].", + "bbox": [ + 500, + 597, + 830, + 612 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For rendering new viewpoints, the method of splatting, as described in [54], is utilized for positioning the Gaussians on the camera planes. This technique, originally presented in [58], involves a viewing transformation denoted by $W$ and the Jacobian $J$ of the affine approximation of the projective transformation. Using these, the covariance matrix $\\Sigma^{\\prime}$ in camera coordinates is determined as follows:", + "bbox": [ + 496, + 613, + 890, + 717 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Sigma^ {\\prime} = J W \\Sigma W ^ {T} J ^ {T}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 736, + 890, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To summarize, each Gaussian point in the model is characterized by a set of attributes: its position, denoted as $x \\in \\mathbb{R}^3$ , its color represented by spherical harmonics coefficients $c \\in \\mathbb{R}^k$ (where $k$ indicates the degrees of freedom), its opacity $\\alpha \\in \\mathbb{R}$ , a rotation quaternion $q \\in \\mathbb{R}^4$ , and a scaling factor $s \\in \\mathbb{R}^3$ . Particularly, for every pixel, the color and opacity of all Gaussians are calculated based on the Gaussian's representation as described in Eq. 1. The blending process of $N$ ordered points overlapping a pixel follows a specific", + "bbox": [ + 496, + 763, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21478", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2c145577e90ac4526669128608404b14a2e757be251d48765c68c0149cf376a8.jpg", + "image_caption": [ + "Before Editing" + ], + "image_footnote": [], + "bbox": [ + 80, + 78, + 272, + 152 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d4119158fb0ed9e6ca8f8f7a57ba4eab02650eb22d129120a17d5e657b99f26c.jpg", + "image_caption": [ + "Figure 2. Illustration of Gaussian semantic tracing. Prompt: Turn him into an old lady. The red mask in the images represents the projection of the Gaussians that will be updated and densified. The dynamic change of the masked area during the training process, as driven by the updating of Gaussians, ensures consistent effectiveness throughout the training duration. Despite starting with potentially inaccurate segmentation masks due to 2D segmentation errors, Gaussian semantic tracing still guarantees high-quality editing results." + ], + "image_footnote": [], + "bbox": [ + 274, + 78, + 361, + 152 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/86224287b4b87fd52842f1201a2029a0b510b24b084391d661c70659ba9a41ce.jpg", + "image_caption": [ + "During Editing" + ], + "image_footnote": [], + "bbox": [ + 367, + 78, + 468, + 152 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "formula:", + "bbox": [ + 76, + 319, + 137, + 330 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC = \\sum_ {i \\in N} c _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 334, + 468, + 377 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $c_{i}$ and $\\alpha_{i}$ signify the color and density of a given point respectively. These values are determined by a Gaussian with a covariance matrix $\\Sigma$ , which is then scaled by estimizable per-point opacity and spherical harmonics (SH) color coefficients.", + "bbox": [ + 75, + 390, + 470, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Diffusion-based Editing Guidance", + "text_level": 1, + "bbox": [ + 76, + 479, + 374, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recent advancements have seen numerous works elevating 2D diffusion processes to 3D, applying these processes extensively in the realm of 3D editing. Broadly, these works can be categorized into two types. The first type [8, 27, 31, 36, 42, 57], exemplified by Dreamfusion's [36] introduction of SDS loss, involves feeding the noised rendering of the current 3D model, along with other conditions, into a 2D diffusion model [39]. The scores generated by the diffusion model then guide the direction of model updates. The second type [5, 12, 37, 43] focuses on conducting 2D editing based on given prompts for the multiview rendering of a 3D model. This approach creates a multi-view 2D image dataset, which is then utilized as a training target to provide guidance for the 3D model.", + "bbox": [ + 75, + 505, + 468, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our work centers on leveraging the exemplary properties of Gaussian Splatting's explicit representation to enhance 3D editing. Consequently, we do not design specific editing guidance mechanisms but instead directly employ the guidance methods mentioned above. Both types of guidance can be applied in our method. For simplicity, we denote the guidance universally as $D$ . Given the parameters of a 3D model, $\\Theta$ , along with the rendered camera pose $p$ and prompt $e$ , the editing loss from the 2D diffusion prior can be formulated as follows:", + "bbox": [ + 75, + 719, + 468, + 867 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {E d i t}} = D (\\Theta ; p, e) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 885, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 500, + 89, + 589, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We define the task of 3D editing on Gaussian Splatting (GS) as follows: Given a prompt $y$ and a 3D scene represented by 3D Gaussians, denoted by $\\Theta$ , where each $\\Theta_{i} = \\{x_{i}, s_{i}, q_{i}, \\alpha_{i}, c_{i}\\}$ represents the parameters of the $i$ -th Gaussian as detailed in Sec. 3.1, the objective is to achieve an edited 3D Gaussians, referred to as $\\Theta_{y}$ , that aligns with or adheres to the specifications of the prompt $y$ .", + "bbox": [ + 496, + 114, + 893, + 220 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We then introduce our novel framework for performing editing tasks on GS. We first introduce Gaussian semantic tracing in Sec. 4.1, along with a new representation method known as Hierarchical Gaussian Splatting (HGS) in Sec. 4.2. The GS semantic tracing enables precise segmentation and tracing within GS, facilitating controllable editing operations. Compared to the standard GS, the HGS representation demonstrates greater robustness against randomness in generative guidance and is more adept at accommodating a diverse range of editing scenarios. Additionally, we have specifically designed 3D inpainting for GS, which encompasses object removal and addition (Sec. 4.3).", + "bbox": [ + 496, + 220, + 893, + 401 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Gaussian Semantic Tracing", + "text_level": 1, + "bbox": [ + 498, + 409, + 746, + 425 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Previous works [31, 57] in 3D editing usually utilize static 2D or 3D masks to apply loss only within the masked pixels, thus constraining the editing process to only edit the desired area. However, this method has limitations. As 3D representations dynamically change during training, static segmentation masks would become inaccurate or even ineffective. Furthermore, the use of static masks to control gradients in NeRF editing poses a significant limitation, as it confines the editing strictly within the masked area. This restriction prevents the edited content from naturally extending beyond the mask, thus 'locking' the content within a specified spatial boundary.", + "bbox": [ + 496, + 431, + 893, + 613 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address the aforementioned issue, we have chosen Gaussian Splitting (GS) as our 3D representation due to its explicit nature. This allows us to directly assign semantic labels to each Gaussian point, thereby facilitating semantic tracing in 3D scenes.", + "bbox": [ + 496, + 614, + 890, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, we enhance the 3D Gaussians $\\Theta$ by adding a new attribute $m$ , where $m_{ij}$ represents the semantic Gaussian mask for the $i$ -th Gaussian point and the $j$ -th semantic label. With this attribute, we can precisely control the editing process by selectively updating only the target 3D Gaussians. During the densification process, newly densified points inherit the semantic label of their parent point. This ensures that we have an accurate 3D semantic mask at every moment throughout the training process.", + "bbox": [ + 496, + 689, + 893, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As illustrated in Fig. 2, Gaussian semantic tracing enables continuous tracking of each Gaussian's categories during training, adjusting to their evolving properties and numbers. This feature is vital, as it permits selective application of gradients, densification and pruning of Gaussians linked to the", + "bbox": [ + 496, + 825, + 893, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21479", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/923b04fef09ed931a05144eeb6b48a5d66a42fc1fe063c5f0f498fd5b2f6567a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 88, + 267, + 186 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/062709a121e4bcf93b2b49fbba502684237bea4820849f11dbb177daa08e97d3.jpg", + "image_caption": [ + "A Garfield cat on the bench" + ], + "image_footnote": [], + "bbox": [ + 277, + 88, + 467, + 186 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6d3665f71103db50b794991f419d05985516ac85334e4e420e8f60b6d54e5c41.jpg", + "image_caption": [ + "Original GS", + "Novel view 1", + "Figure 3. 3D inpainting for object incorporation. GaussianEditor is capable of adding objects at specified locations in a scene, given a 2D inpainting mask and a text prompt from a single view. The whole process takes merely five minutes." + ], + "image_footnote": [], + "bbox": [ + 80, + 200, + 267, + 297 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4764f866214cdd17259cb45499430e816685a5d70198c852579fb834905f3fb6.jpg", + "image_caption": [ + "Novel view 2" + ], + "image_footnote": [], + "bbox": [ + 274, + 200, + 462, + 297 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c2a5a62afb95861758c4a3d2d16d260e5669863306bcbdbe923d3ec5aff5b3ad.jpg", + "image_caption": [ + "Original View", + "Figure 4. 3D inpainting for object removal. Typically, removing the target object based on a Gaussian semantic mask generates artifacts at the interface between the target object and the scene. To address this, we generate a repaired image using a 2D inpainting method and employ Mean Squared Error (MSE) loss for supervision. The whole process takes merely two minutes." + ], + "image_footnote": [], + "bbox": [ + 78, + 387, + 205, + 520 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/63539f4d440ad98e394d16f5078a6f2c08547fb24de3e2585946e7b9b466f72d.jpg", + "image_caption": [ + "Removed" + ], + "image_footnote": [], + "bbox": [ + 209, + 387, + 336, + 520 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/506b53d4a3d74a0f6b367692a87d459924debdf1eff497015d50a7c31513f1f1.jpg", + "image_caption": [ + "Inpainted" + ], + "image_footnote": [], + "bbox": [ + 339, + 387, + 468, + 520 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "specified category. Additionally, it facilitates training solely by rendering the target object, significantly speeding up the process in complex scenes. The semantic Gaussian mask $m$ functions as a dynamic 3D segmentation mask, evolving with the training, allowing content to expand freely in space. This contrasts with NeRF, where content is restricted to a fixed spatial area.", + "bbox": [ + 75, + 655, + 468, + 760 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, we discuss Gaussian Splating unprojection, the method we propose to obtain semantic Gaussian mask $m$ . For a set of 3D Gaussians $\\Theta$ , we render them from multiple viewpoints to generate a series of renderings $\\mathcal{I}$ . These renderings are then processed using 2D segmentation techniques [18] to obtain 2D segmentation masks $\\mathcal{M}$ , with each $\\mathcal{M}_j$ , representing the $j$ -th semantic labels.", + "bbox": [ + 75, + 762, + 470, + 868 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To obtain the semantic label for each Gaussian, we unproject the posed 2D semantic label back to the Gaussians", + "bbox": [ + 76, + 869, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with inverse rendering. Concretely, we maintain a weight and a counter for each Gaussian. For pixel $p$ on the semantic maps, we unproject the semantic label back to the Gaussians that affects it by", + "bbox": [ + 496, + 90, + 892, + 151 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw _ {i} ^ {j} = \\sum o _ {i} (\\boldsymbol {p}) * T _ {i} ^ {j} (\\boldsymbol {p}) * \\mathcal {M} ^ {j} (\\boldsymbol {p}), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 164, + 892, + 186 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $w_{i}^{j}$ represents the weight of the $i$ -th Gaussian for the $j$ -th semantic label, while $o_{i}(\\pmb{p})$ , $T_{i}^{j}(\\pmb{p})$ , and $\\mathcal{M}^j (\\pmb{p})$ denote the opacity, transmittance from pixel $\\pmb{p}$ , and semantic mask of pixel $\\pmb{p}$ for the $i$ -th Gaussian, respectively. After updating all the Gaussian weights and counters, we determine whether a Gaussian belongs to the $j$ -th semantic class based on whether its average weight exceeds a manually set threshold.", + "bbox": [ + 496, + 195, + 893, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The entire labeling process is remarkably fast, typically taking less than a second. Once this semantic label assignment is completed, the entire Gaussian scene becomes parsed by us, making a variety of operations possible. Notably, 2D diffusion guidance often struggles to effectively edit small objects in complex scenes. Thanks to Gaussian semantic tracing, we can now render these small objects independently and input them into the 2D diffusion model, thereby achieving more precise supervision.", + "bbox": [ + 496, + 318, + 893, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Hierarchical Gaussian Splitting", + "text_level": 1, + "bbox": [ + 500, + 465, + 782, + 482 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The effectiveness of vanilla GS [17] in reconstruction tasks lies in the high-quality initialization provided by point clouds derived from SFM [41], coupled with stable supervision from ground truth datasets.", + "bbox": [ + 496, + 492, + 890, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, the scenario changes in the field of generation. In previous work involving GS in text-to-3D and image-to-3D [7, 45, 53], GS has shown limitations when facing the randomness of generative guidance due to its nature as a point cloud-like representation. This instability in GS is mainly due to their direct exposure to the randomness of loss functions, unlike implicit representations. GS models, which update a large number of Gaussian points each training step, lack the memorization and moderating ability of neural networks. This leads to erratic updates and prevents GS from achieving the detailed results as GS's excessive fluidity hampers its convergence in generative training.", + "bbox": [ + 496, + 551, + 893, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address these challenges, we introduce Hierarchical Gaussian Splitting (HGS), a structured representation of GS that is more suitable for generative and editing scenarios. HGS categorizes GS into different generations based on the densification round in which a particular Gaussian point is produced. The initial Gaussians $\\Theta$ , are all assigned a generation of 0. During the training process for editing, points generated in the $k$ -th densification round are marked as generation $k$ .", + "bbox": [ + 496, + 734, + 893, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Subsequently, we impose varying constraints on Gaussians from different generations to control their degree of", + "bbox": [ + 500, + 869, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21480", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3ebabeaf66e0af185ce7b5cad3c6bfe929f0724965f0895d0ce4e6abccb73c11.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 98, + 236, + 189 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/904ef5a71e062b30d166abdee9c344f46dee956893cd7f79a9f1aa6d4abeb11d.jpg", + "image_caption": [ + "Original view" + ], + "image_footnote": [], + "bbox": [ + 80, + 200, + 236, + 292 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e5431b158ca0c78d12e3cd5cdfd6ba72dfe7d0ed703ceb8cfe3471ec45adb669.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 243, + 61, + 447, + 146 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e08735f3a506f77a852a134fb40a8b0aff814788bd092f94774788da1a1e55e7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 243, + 147, + 447, + 233 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ad97e0ff4e94d396e2c8a8ae09059fb0865b1b6798ff91162c554b64952a1609.jpg", + "image_caption": [ + "\"Turn him into a clown\"", + "Figure 5. Qualitative comparison. It's important to note the level of control we maintain over the editing area (the whole body of the man). Background and other non-target regions are essentially unaffected, in contrast to Instruct-Nerf2Nerf [12] where the entire scene undergoes changes. GaussianEditor-DDS and GaussianEditor-iN2N indicate that we utilize delta denoising score [14] and Instruct-Nerf2Nerf [12] respectively, as guidance for editing." + ], + "image_footnote": [], + "bbox": [ + 243, + 234, + 447, + 320 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4da21a43eafbefc7a60a8342837cbd66788954ca0ea8f873080cd01f6793bcd9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 454, + 61, + 658, + 145 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6229dbee583e751dcecbd87e0f7766ea3be48d5ac4b60109301d4e21c201c7ec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 454, + 147, + 658, + 233 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2994751f0235f9bfa8e155a4b98f6960fc0a14400c95b094e271b16d13c1adcb.jpg", + "image_caption": [ + "\"Turn him into Albert Einstein\"" + ], + "image_footnote": [], + "bbox": [ + 454, + 234, + 658, + 320 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/20e87eb054ea179d88612b0c8bccc0b2b0dfd7678b5cb0aabf304734d6e8744f.jpg", + "image_caption": [ + "Instruct-N2N" + ], + "image_footnote": [], + "bbox": [ + 666, + 61, + 867, + 145 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0a5fe71e53754e53af6e354f566a2cbefce62b938d84d6d1f152a89e3356f814.jpg", + "image_caption": [ + "GSEeditor-DDS" + ], + "image_footnote": [], + "bbox": [ + 666, + 147, + 869, + 233 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ab3efa8079fd0cd213a6342973ef38407967576490cac8045ecd0f2081f4bd4e.jpg", + "image_caption": [ + "\"Turn him into a Modigliani painting.\"", + "GSEoperator-iN2N" + ], + "image_footnote": [], + "bbox": [ + 666, + 234, + 869, + 320 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "flexibility. The older the generation, the stronger the constraints applied. Anchor loss is utilized to enforce these constraints. At the beginning of training, HGS records the attributes of all Gaussians as anchors. These anchors are then updated to reflect the current state of the Gaussians at each densification process. During training, MSE loss between the anchor state and the current state is employed to ensure that the Gaussians do not deviate too far from their respective anchors:", + "bbox": [ + 75, + 407, + 470, + 544 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a n c h o r}} ^ {P} = \\sum_ {i = 0} ^ {n} \\lambda_ {i} \\left(P _ {i} - \\hat {P} _ {i}\\right) ^ {2} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 551, + 468, + 590 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $n$ represents the total number of Gaussians and $P$ denotes a certain property of the current Gaussian, including elements from the set $x, s, q, \\alpha, c$ . Here, $\\hat{P}$ refers to the same property recorded in the anchor state. The term $\\lambda_{i}$ indicates the strength of the anchor loss applied to the $i$ -th Gaussian, which varies based on its generation. The overall training loss is defined as:", + "bbox": [ + 75, + 597, + 468, + 702 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {E d i t} + \\sum_ {P \\in \\{x, s, q, \\alpha , c \\}} \\lambda_ {P} \\mathcal {L} _ {a n c h o r} ^ {P} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 708, + 468, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this equation, $\\lambda_{P}$ signifies the strength of the anchor loss applied to property $P$ , and $\\mathcal{L}_{Edit}$ is the edit loss defined in Sec. 3.2.", + "bbox": [ + 75, + 750, + 468, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This generational design in HGS prevents the issue of excessive flexibility in GS when faced with stochastic loss. With each densification, the anchor loss weight $\\lambda_{i}$ for all previous generations of Gaussians is increased. As a result, the fluidity of the existing generations gradually decreases until it nearly solidifies. This approach ensures stable geometry formation under stochastic losses, relying on the almost", + "bbox": [ + 75, + 795, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "unconstrained Gaussians from new densifications to carve out details. Furthermore, this method of applying anchor loss can effectively meet various editing needs. For instance, to limit changes in the original GS, one can increase the anchor loss weight for generation 0. Similarly, if there is no desire to alter color or geometry during editing, a stronger anchor loss can be applied to these specific properties.", + "bbox": [ + 496, + 409, + 892, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Additionally, to address the challenge of manually determining a densification threshold, we regulate the densification process based on a percentage criterion. In this method, during each densification step, we selectively densify only those Gaussians whose 3D position gradients are within the top $k\\%$ . This strategy proves to be more manageable and intuitive than directly setting a threshold value in the Hierarchical Gaussian Splatting (HGS) framework.", + "bbox": [ + 496, + 515, + 893, + 636 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. 3D Inpainting", + "text_level": 1, + "bbox": [ + 500, + 648, + 643, + 665 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Object Removal. Simply removing Gaussians identified by a mask can lead to artifacts, especially at the interfaces where the object intersects with other Gaussians. To address this, we employ 2D inpainting techniques to provide guidance for filling these areas. However, effective 2D inpainting requires precise masks to offer better guidance. To generate these masks, after deletion, we use the KNN algorithm to identify Gaussians nearest to the ones removed, which are likely at the interface. These are then projected onto various views. We subsequently dilate the mask and fix any holes to accurately represent the interface area, thus creating a refined mask for the boundary zones. The whole object removal procedure typically takes only two minutes.", + "bbox": [ + 496, + 672, + 892, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Object Incorporation. We define this task as follows: Within the 3D Gaussians $\\theta$ , given a camera pose $p$ and the", + "bbox": [ + 498, + 869, + 893, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21481", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9599e95f325b4aef1c133d57fb1078f364f81f6f3842b900c43845bf2016623b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 61, + 277, + 162 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2c769853edacba0890b0e4c18d8821f36c2c1659794645a0b862b334b3c21099.jpg", + "image_caption": [ + "\"Make it snowy\"" + ], + "image_footnote": [], + "bbox": [ + 282, + 59, + 482, + 162 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c5e470ff4bdb73e8b4696f570da95ad40f25e00aa32c865e2b164f78ba00e1b1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 59, + 686, + 162 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ebb9c7e1701d8bc5025bf005659c02391b3663970ee1092fc12fe7c5c7ee59d1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 692, + 59, + 890, + 162 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/65f3fa13daa9c9311c3f699342e5ca879d8fef3fcb8a74f5378966b3bbe5b147.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 181, + 274, + 296 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c9bb3b3505323a0a8196e6b4f85ab7f411be00d561e91a7c2e9df1fdcfbe2e24.jpg", + "image_caption": [ + "\"Turn him into an old lady\"" + ], + "image_footnote": [], + "bbox": [ + 282, + 181, + 478, + 296 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0e26f8c1349b33571e7a255613d9b3918f26733e21df07532efc56d13c7191f6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 181, + 686, + 296 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2cd1bb935d0d135a14b1a790e7e0da193362d77f605458b276e07b7d53463048.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 692, + 181, + 890, + 296 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fd03a26a7a83e1fe6d04f48da2a46c11c77a3802382dc58c7f7336b2cfb6d928.jpg", + "image_caption": [ + "Figure 6. Extensive Results of GaussianEditor. Our method is capable of various editing tasks, including face and scene editing. In face and bear editing, we restrict the editing area to the face using Gaussian semantic tracing, ensuring that undesired areas remain unchanged. The leftmost column demonstrates the original view, while the right three columns show the images after editing." + ], + "image_footnote": [], + "bbox": [ + 80, + 314, + 277, + 429 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3556919a9a88b24b62f61a436e1f03f5a0346c156db177fec6227da8aefe487a.jpg", + "image_caption": [ + "\"Turn the bear into a grizzly bear\"" + ], + "image_footnote": [], + "bbox": [ + 282, + 314, + 482, + 429 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/232cb283a868f1a6540735f1f505f2b11ad7b58d77bc0526265c02b9dbe2bb40.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 314, + 686, + 429 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d87325b0cfb59cb8265af45d88ef40dd70eced37096277c7dfc4d18004cfbc6b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 692, + 314, + 890, + 429 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "corresponding rendering $I$ from this viewpoint, the user provides a 2D mask $M$ on $I$ indicating the area they wish to inpaint. Additionally, a prompt $y$ is provided to specify the content of the inpainting. We then update $\\theta$ to fulfill the inpainting request.", + "bbox": [ + 75, + 518, + 470, + 594 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Given $I$ , $M$ , and $y$ , the process begins with generating a 2D inpainted image $I^M y$ , utilizing a 2D inpainting diffusion model as per [34]. Subsequently, the foreground object from $I^M y$ , created by [34], is segmented and input into the image-to-3D method referenced in [25] to generate a coarse 3D mesh. This coarse mesh is then transformed into 3D Gaussians $\\theta_y$ , and refined with HGS detailed in Sec. 4.2.", + "bbox": [ + 75, + 599, + 468, + 705 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For aligning the coordinate system of $\\theta_y$ with $\\theta$ , the depth of $I_y^M$ is first estimated using the technique from [38]. This depth is then aligned with the depth map rendered by $\\theta$ at camera pose $p$ , using the least squares method. With this alignment, we can accurately determine the coordinates and scale of the inpainted foreground object in the coordinate system of $\\theta$ . After transforming $\\theta_y$ into the coordinate system of $\\theta$ , we simply concatenate them to produce the final inpainted 3D Gaussians.", + "bbox": [ + 75, + 712, + 470, + 849 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "It is important to note that due to our efficient design, the entire object incorporation procedure can be completed in approximately 5 minutes.", + "bbox": [ + 75, + 854, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 516, + 632, + 534 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 542, + 718, + 558 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We utilize the highly optimized renderer implementation from [16] for Gaussian rendering and base our implementation on Threestudio [11]. All the original 3D Gaussians used in this work are trained using the methods described in [16]. Our experiments are conducted on a single RTX A6000 GPU. As detailed in Sec. 4.1, once we obtain segmentation masks from the 2D segmentation method outlined in [18], segmenting the 3D Gaussians takes only about 1 second.", + "bbox": [ + 496, + 566, + 893, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For editing large scenes, the camera poses employed during the editing process are selected from a subset of the multi-view image dataset initially used for reconstruction. Depending on the complexity of the scene, the number of camera poses used in our experiments varies from 24 to 96. The editing process, influenced by the specified prompt and the complexity of the scene, typically involves optimizing for 500-1000 steps, taking about 5-10 minutes in total.", + "bbox": [ + 496, + 688, + 893, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Regarding 3D inpainting for object incorporation, as detailed in Sec. 4.3, it takes approximately 3 minutes to generate a 3D mesh using the method from [25] and an additional 2 minutes to transfer this mesh into 3D Gaussians and refine it, while the composition process of two Gaussians takes less than 1 second.", + "bbox": [ + 496, + 810, + 893, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21482", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/45031bcd2073026aca19d4de3ba66516ad3104370caf41dcbaea304f05f5fcb8.jpg", + "image_caption": [ + "Original View" + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 207, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b79313259f059997c6a3e93f4268768e763521f2e67d6fc0ec4ee0b5f92213a5.jpg", + "image_caption": [ + "GS" + ], + "image_footnote": [], + "bbox": [ + 210, + 88, + 336, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b3991c369740ac2e2ff04101002d89855f37350aa35ca8bbee066c047d5f0cee.jpg", + "image_caption": [ + "HGS", + "Figure 7. Ablation study on Hierarchical Gaussian Splitting (HGS). Prompt: make the grass on fire. Even when specifying the editing area with prompts, generative methods like InstructPix2Pix [4] tend to edit the entire 2D image. Without HGS, Gaussians tend to conform to this whole-image editing by spreading and densifying across the entire scene, leading to uncontrollable densification and blurring of the image. With HGS, however, this kind of diffusion is effectively restrained." + ], + "image_footnote": [], + "bbox": [ + 341, + 88, + 468, + 220 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Qualitative Comparisons", + "text_level": 1, + "bbox": [ + 76, + 385, + 305, + 401 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As illustrated in Fig. 5, GaussianEditor-iN2N surpasses other methods in both the quality of edits and controllability. Instruct-Nerf2Nerf, while producing edits with insufficient detail, cannot also control the editing area. GaussianEditorDDS, due to the more challenging control of guidance offered by DDS loss [14] compared to Instruct-pix2pix [4], tends to result in oversaturated colors and less precise editing outcomes.", + "bbox": [ + 75, + 407, + 470, + 527 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Additionally, our method exhibits exceptional control over the editing area. This is achieved through Gaussian semantic tracing, which identifies the Gaussians that require editing at each training step, for example, the entire human body in Fig. 5. It's important to note that in Instruct-Nerf2Nerf [12], the use of static 2D or 3D masks restricts the spatial freedom of the edits, as the permissible area for the edited subject is limited by these masks. Furthermore, the effectiveness of static masks diminishes as the geometries and appearances of 3D models evolve during training.", + "bbox": [ + 75, + 529, + 468, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Fig. 6, we demonstrate that GaussianEditor can accommodate a variety of scenarios, such as editing in large-scale scenes and facial swaps. In the case of large scenes, we did not apply Gaussian semantic tracing. However, for facial swaps, we traced the Gaussians corresponding to facial regions, achieving controllable and realistic editing.", + "bbox": [ + 75, + 681, + 470, + 772 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3.Quantitative Comparisons", + "text_level": 1, + "bbox": [ + 76, + 782, + 316, + 796 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Table 1, we conduct quantitative comparisons on user study and CLIP directional similarity (as shown in InstructPix2Pix [4] and StyleGAN-Nada [9]). GaussianEditor-iN2N not only demonstrates superior outcomes in user studies but also excels in CLIP Directional Similarity. Besides, Instruct-Nerf2Nerf [12] typically requires more than 30 minutes to complete the editing of a scene, whereas our method", + "bbox": [ + 75, + 796, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/75f672c4926f39a0ae9d4382b6aff1a73dfdfe1131c107b8b31bee82c0874c8f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
iN2N [12]Ours (DDS)Ours (iN2N)
User study15.45%12.27%72.28%
CLIP Directional Similarity0.16000.18130.2071
", + "bbox": [ + 504, + 89, + 890, + 138 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 1. Quantitative Comparison. GaussianEditor-iN2N outperforms in both user study evaluations and CLIP Directional Similarity [9] metrics.", + "bbox": [ + 498, + 150, + 893, + 191 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "only takes between 5 to 10 minutes.", + "bbox": [ + 500, + 217, + 736, + 232 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 239, + 651, + 257 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As demonstrated in Fig. 7, we conducted ablation experiments on Hierarchical Gaussian Splitting(HGS). Without HGS, Gaussians tend to spread and densify across the scene, leading to uncontrolled densification and image blurring. This is typically caused by the tendency of methods like Instruct-Pix2Pix to edit the entire 2D image when prompts are used to define editing areas. However, HGS effectively circumvents this issue by constraining the mobility of the Gaussian in the old generation, ensuring that the overall scene does not exhibit excessive mobility.", + "bbox": [ + 496, + 263, + 893, + 415 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 426, + 617, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In our research, we introduce GaussianEditor, an innovative 3D editing algorithm based on Gaussian Splatting, designed for enhanced control and efficiency. Our method employs Gaussian semantic tracing for precise identification and targeting of editing areas, followed by Hierarchical Gaussian Splatting (HGS) to balance fluidity and stability in achieving detailed results under stochastic guidance. Additionally, we developed a specialized 3D inpainting algorithm for Gaussian Splatting, streamlining object removal and integration, and greatly reducing editing time.", + "bbox": [ + 496, + 452, + 893, + 603 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitation. Similar to previous 3D editing works based on 2D diffusion models, GaussianEditor relies on these models to provide effective supervision. However, current 2D diffusion models struggle to offer effective guidance for certain complex prompts, leading to limitations in 3D editing.", + "bbox": [ + 496, + 604, + 893, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 691, + 687, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This study is supported under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). This research is also supported by the MoE AcRF Tier 2 grant (MOE-T2EP20220-0007).", + "bbox": [ + 496, + 717, + 893, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 500, + 805, + 596, + 820 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Chong Bao, Yinda Zhang, and Bangbang et al. Yang. Sine: Semantic-driven image-based nef editing with prior-guided editing field. In CVPR 2023, pages 20919-20929, 2023. 3", + "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan." + ], + "bbox": [ + 506, + 829, + 893, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21483", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855-5864, 2021. 2", + "[3] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. CVPR, 2022. 3", + "[4] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. arXiv preprint arXiv:2211.09800, 2022. 8", + "[5] Yiwen Chen, Chi Zhang, Xiaofeng Yang, Zhongang Cai, Gang Yu, Lei Yang, and Guosheng Lin. It3d: Improved text-to-3d generation with explicit view synthesis. arXiv preprint arXiv:2308.11473, 2023. 4", + "[6] Zhiqin Chen, Thomas Funkhouser, Peter Hedman, and Andrea Tagliasacchi. Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. arXiv preprint arXiv:2208.00277, 2022. 3", + "[7] Zilong Chen, Feng Wang, and Huaping Liu. Text-to-3d using gaussian splatting. arXiv preprint arXiv:2309.16585, 2023. 2, 3, 5", + "[8] Xinhua Cheng, Tianyu Yang, Jianan Wang, Yu Li, Lei Zhang, Jian Zhang, and Li Yuan. Progressive3d: Progressively local editing for text-to-3d content creation with complex semantic prompts. arXiv preprint arXiv:2310.11784, 2023. 4", + "[9] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clipguided domain adaptation of image generators. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 8", + "[10] William Gao, Noam Aigerman, Thibault Groueix, Vladimir G Kim, and Rana Hanocka. Textdeformer: Geometry manipulation using text guidance. arXiv preprint arXiv:2304.13348, 2023. 3", + "[11] Yuan-Chen Guo, Ying-Tian Liu, Chen Wang, Zi-Xin Zou, Guan Luo, Chia-Hao Chen, Yan-Pei Cao, and Song-Hai Zhang. threestudio: A unified framework for 3d content generation. https://github.com/threestudio-project/threestudio, 2023.7", + "[12] Ayaan Haque, Matthew Tancik, Alexei A Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. arXiv preprint arXiv:2303.12789, 2023. 2, 4, 6, 8", + "[13] Peter Hedman, Pratul P. Srinivasan, Ben Mildenhall, Jonathan T. Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. ICCV, 2021. 3", + "[14] Amir Hertz, Kfir Aberman, and Daniel Cohen-Or. Delta denoising score. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2328-2337, 2023. 6, 8", + "[15] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 42(4): 1-14, 2023. 2, 3", + "[16] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ToG, 42(4):1-14, 2023. 7", + "[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "field rendering. ACM Transactions on Graphics, 42(4), 2023. 5", + "[18] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 5, 7", + "[19] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nerf for editing via feature field distillation. arXiv preprint arXiv:2205.15585, 2022. 3", + "[20] Yuan Li, Zhi-Hao Lin, David Forsyth, Jia-Bin Huang, and Shenlong Wang. Climatenerf: Physically-based neural rendering for extreme climate synthesis. arXiv e-prints, pages arXiv-2211, 2022. 3", + "[21] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In CVPR, 2023. 3", + "[22] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, pages 300–309, 2023. 3", + "[23] Hao-Kang Liu, I Shen, Bing-Yu Chen, et al. Nerf-in: Free-form nerf inpainting with rgb-d priors. arXiv preprint arXiv:2206.04901, 2022. 3", + "[24] Steven Liu, Xiuming Zhang, Zhoutong Zhang, Richard Zhang, Jun-Yan Zhu, and Bryan Russell. Editing conditional radiance fields. In ICCV 2021, pages 5773-5783, 2021. 3", + "[25] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. arXiv preprint arXiv:2310.15008, 2023. 7", + "[26] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 3", + "[27] Aryan Mikaeili, Or Perel, Mehdi Safaee, Daniel Cohen-Or, and Ali Mahdavi-Amiri. Sked: Sketch-guided text-based 3d editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14607-14619, 2023. 4", + "[28] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2, 3", + "[29] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 3", + "[30] Atsuhiro Noguchi, Xiao Sun, Stephen Lin, and Tatsuya Harada. Neural articulated radiance field. In ICCV 2021, pages 5762-5772, 2021. 3", + "[31] Jangho Park, Gihyun Kwon, and Jong Chul Ye. Ed-nerf: Efficient text-guided editing of 3d scene using latent space nerf. arXiv preprint arXiv:2310.02712, 2023. 2, 3, 4", + "[32] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo Martin-" + ], + "bbox": [ + 503, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21484", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Brualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. 2", + "[33] Sida Peng, Yuanqing Zhang, Yinghao Xu, and et al. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In CVPR 2021, pages 9054–9063, 2021. 3", + "[34] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 7", + "[35] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 3", + "[36] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 2, 4", + "[37] Amit Raj, Srinivas Kaza, Ben Poole, Michael Niemeyer, Nataniel Ruiz, Ben Mildenhall, Shiran Zada, Kfir Aberman, Michael Rubinstein, Jonathan Barron, et al. Dreambooth3d: Subject-driven text-to-3d generation. arXiv preprint arXiv:2303.13508, 2023. 4", + "[38] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12179-12188, 2021. 7", + "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 4", + "[40] Sara Fridovich-Keil and Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 3", + "[41] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 5", + "[42] Etai Sella, Gal Fiebelman, Peter Hedman, and Hadar Averbuch-Elor. Vox-e: Text-guided voxel editing of 3d objects. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 430-440, 2023. 4", + "[43] Ruizhi Shao, Jingxiang Sun, Cheng Peng, Zerong Zheng, Boyao Zhou, Hongwen Zhang, and Yebin Liu. Control4d: Dynamic portrait editing by learning 4d gan from 2d diffusion-based editor. arXiv preprint arXiv:2305.20082, 2023. 4", + "[44] Chaoyue Song, Tianyi Chen, Yiwen Chen, Jiacheng Wei, Chuan Sheng Foo, Fayao Liu, and Guosheng Lin. Moda: Modeling deformable 3d objects from casual videos. arXiv preprint arXiv:2304.08279, 2023. 3", + "[45] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. 2, 3, 5", + "[46] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3835-3844, 2022. 2, 3", + "[47] Can Wang, Ruixiang Jiang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Nerf-art: Text-driven neural radiance fields stylization. IEEE Transactions on Visualization and Computer Graphics, 2023. 2, 3", + "[48] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528, 2023. 3", + "[49] Tianhan Xu and Tatsuya Harada. Deforming radiance fields with cages. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIII, pages 159-175. Springer, 2022. 3", + "[50] Bangbang Yang, Chong Bao, and Junyi et al. Zeng. Neumesh: Learning disentangled neural mesh-based implicit field for geometry and texture editing. In ECCV 2022, pages 597-614. Springer, 2022. 3", + "[51] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101, 2023. 3", + "[52] Zeyu Yang, Hongye Yang, Zijie Pan, Xiatian Zhu, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. arXiv preprint arXiv:2310.10642, 2023. 3", + "[53] Taoran Yi, Jiemin Fang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussian-dreamer: Fast generation from text to 3d gaussian splatting with point cloud priors. arXiv preprint arXiv:2310.08529, 2023. 2, 3, 5", + "[54] Wang Yifan, Felice Serena, Shihao Wu, Cengiz Öztireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM Transactions on Graphics (TOG), 38(6):1-14, 2019. 3", + "[55] Yu-Jie Yuan, Yang-Tian Sun, Yu-Kun Lai, and et al. Nerf-editing: geometry editing of neural radiance fields. In CVPR 2022, pages 18353-18364, 2022. 3", + "[56] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 2", + "[57] Jingyu Zhuang, Chen Wang, Lingjie Liu, Liang Lin, and Guanbin Li. Dreameditor: Text-driven 3d scene editing with neural fields. arXiv preprint arXiv:2306.13455, 2023. 2, 3, 4", + "[58] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Surface splatting. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 371-378, 2001. 3" + ], + "bbox": [ + 501, + 92, + 893, + 785 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21485", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/GaussianEditor_ Swift and Controllable 3D Editing with Gaussian Splatting/f458e425-2ba9-468a-94e0-95be47f99560_model.json b/2024/GaussianEditor_ Swift and Controllable 3D Editing with Gaussian Splatting/f458e425-2ba9-468a-94e0-95be47f99560_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bb47591ff8067f4e2c3a17d68edb7ecfa57a6067 --- /dev/null +++ b/2024/GaussianEditor_ Swift and Controllable 3D Editing with Gaussian Splatting/f458e425-2ba9-468a-94e0-95be47f99560_model.json @@ -0,0 +1,3146 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.131, + 0.87, + 0.154 + ], + "angle": 0, + "content": "GaussianEditor: Swift and Controllable 3D Editing with Gaussian Splitting" + }, + { + "type": "text", + "bbox": [ + 0.147, + 0.179, + 0.278, + 0.217 + ], + "angle": 0, + "content": "Yiwen Chen\\*1,2 Yikai Wang3" + }, + { + "type": "text", + "bbox": [ + 0.282, + 0.181, + 0.419, + 0.217 + ], + "angle": 0, + "content": "Zilong Chen\\*3 Zhongang Cai4" + }, + { + "type": "text", + "bbox": [ + 0.434, + 0.181, + 0.668, + 0.217 + ], + "angle": 0, + "content": "Chi Zhang\\(^{2}\\) Feng Wang\\(^{3}\\) Lei Yang\\(^{4}\\) Huaping Liu\\(^{3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.681, + 0.181, + 0.832, + 0.217 + ], + "angle": 0, + "content": "Xiaofeng Yang\\(^{2}\\) Guosheng Lin\\(^{**1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.315, + 0.224, + 0.656, + 0.242 + ], + "angle": 0, + "content": "\\(^{1}\\)S-Lab, Nanyang Technological University" + }, + { + "type": "text", + "bbox": [ + 0.161, + 0.242, + 0.81, + 0.259 + ], + "angle": 0, + "content": "\\(^{2}\\)School of Computer Science and Engineering, Nanyang Technological University" + }, + { + "type": "text", + "bbox": [ + 0.201, + 0.26, + 0.77, + 0.278 + ], + "angle": 0, + "content": "\\(^{3}\\)Department of Computer Science and Technology, Tsinghua University" + }, + { + "type": "text", + "bbox": [ + 0.399, + 0.278, + 0.571, + 0.293 + ], + "angle": 0, + "content": "4 SenseTime Research" + }, + { + "type": "text", + "bbox": [ + 0.281, + 0.297, + 0.684, + 0.311 + ], + "angle": 0, + "content": "https://buaacyw.github.io/gaussian-editor" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.328, + 0.213, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.175, + 0.4, + 0.243, + 0.41 + ], + "angle": 0, + "content": "Original view" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.328, + 0.345, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.414, + 0.213, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.176, + 0.492, + 0.245, + 0.502 + ], + "angle": 0, + "content": "Original view" + }, + { + "type": "image", + "bbox": [ + 0.215, + 0.414, + 0.344, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.354, + 0.328, + 0.484, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.375, + 0.399, + 0.58, + 0.41 + ], + "angle": 0, + "content": "\"Delete the bike. Add a dog on the bench.\"" + }, + { + "type": "image", + "bbox": [ + 0.486, + 0.328, + 0.619, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.624, + 0.328, + 0.756, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.758, + 0.328, + 0.89, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.693, + 0.399, + 0.814, + 0.41 + ], + "angle": 0, + "content": "\"Make the grass on fire\"" + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.414, + 0.484, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.435, + 0.491, + 0.543, + 0.501 + ], + "angle": 0, + "content": "\"Turn him into Hulk\"" + }, + { + "type": "image", + "bbox": [ + 0.485, + 0.414, + 0.616, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.623, + 0.414, + 0.755, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.757, + 0.414, + 0.888, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.701, + 0.491, + 0.821, + 0.501 + ], + "angle": 0, + "content": "\"Turn him into a clown\"" + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.512, + 0.213, + 0.578 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.215, + 0.512, + 0.346, + 0.578 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.354, + 0.512, + 0.484, + 0.578 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.486, + 0.512, + 0.617, + 0.578 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.623, + 0.512, + 0.754, + 0.578 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.757, + 0.512, + 0.888, + 0.578 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.442, + 0.579, + 0.531, + 0.589 + ], + "angle": 0, + "content": "\"Make it autumn\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.6, + 0.895, + 0.642 + ], + "angle": 0, + "content": "Figure 1. Results of GaussianEditor. GaussianEditor offers swift, controllable, and versatile 3D editing. A single editing session only takes 5-10 minutes. Please note our precise editing control, where only the desired parts are modified. Taking the \"Make the grass on fire\" example from the first row of the figure, other objects in the scene such as the bench and tree remain unaffected." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.685, + 0.313, + 0.701 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.716, + 0.471, + 0.867 + ], + "angle": 0, + "content": "3D editing plays a crucial role in many areas such as gaming and virtual reality. Traditional 3D editing methods, which rely on representations like meshes and point clouds, often fall short in realistically depicting complex scenes. On the other hand, methods based on implicit 3D representations, like Neural Radiance Field (NeRF), render complex scenes effectively but suffer from slow processing speeds and limited control over specific scene areas. In response to these challenges, our paper presents GaussianEditor, the first 3D editing algorithm based on Gaussian Splitting (GS), a novel" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.671, + 0.895, + 0.838 + ], + "angle": 0, + "content": "3D representation. GaussianEditor enhances precision and control in editing through our proposed Gaussian semantic tracing, which traces the editing target throughout the training process. Additionally, we propose Hierarchical Gaussian splatting (HGS) to achieve stabilized and fine results under stochastic generative guidance from 2D diffusion models. We also develop editing strategies for efficient object removal and integration, a challenging task for existing methods. Our comprehensive experiments demonstrate GaussianEditor's superior control, effective, and efficient performance, marking a significant advancement in 3D editing." + }, + { + "type": "page_footnote", + "bbox": [ + 0.101, + 0.876, + 0.239, + 0.888 + ], + "angle": 0, + "content": "** Corresponding author." + }, + { + "type": "page_footnote", + "bbox": [ + 0.101, + 0.888, + 0.395, + 0.9 + ], + "angle": 0, + "content": "* The first two authors contributed equally to this work." + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.876, + 0.395, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21476" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.081, + 0.091, + 0.208, + 0.105 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.115, + 0.468, + 0.235 + ], + "angle": 0, + "content": "In the evolving field of computer vision, the development of user-friendly 3D representations and editing algorithms is a key objective. Such technologies are vital in various applications, ranging from digital gaming to the growing MetaVerse. Traditional 3D representations like meshes and point clouds have been preferred due to their interactive editing capabilities. However, these methods face challenges in accurately rendering complex 3D scenes." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.237, + 0.468, + 0.462 + ], + "angle": 0, + "content": "The recent rise of implicit 3D representations, exemplified by the Neural Radiance Field (NeRF) [28], represents a paradigm shift in 3D scene rendering. NeRF's capacity for high-fidelity rendering, coupled with its implicit nature that offers significant expansibility, marks a substantial improvement over conventional approaches[2, 32, 56]. This dual advantage has placed a significant focus on the NeRF framework in 3D editing [12, 31, 46, 47, 57], establishing it as a foundational approach for a considerable duration. However, NeRF's reliance on high-dimensional multilayer perception (MLP) networks for scene data encoding presents limitations. It restricts direct modification of specific scene parts and complicates tasks, like inpainting and scene composition. This complexity extends to the training and rendering processes, hindering practical applications." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.463, + 0.468, + 0.583 + ], + "angle": 0, + "content": "In light of these challenges, our research is focused on developing an advanced 3D editing algorithm. This algorithm aims for flexible and rapid editing of 3D scenes, integrating both implicit editing, like text-based editing, and explicit control, such as bounding box usage for specific area modifications. To achieve these goals, we choose Gaussian Splatting (GS) [15] for its real-time rendering and explicit point cloud-like representations." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.584, + 0.468, + 0.855 + ], + "angle": 0, + "content": "However, editing Gaussian Splatting (GS) [15] faces distinct challenges. A primary issue is the absence of efficient methods to accurately identify target Gaussians, which is crucial for precise controllable editing. Moreover, it has been observed in [7, 45, 53] that optimizing Gaussian Splatting (GS) using highly random generative guidance like Score Distillation Sampling [36] poses significant challenges. One possible explanation is that, unlike implicit representations buffered by neural networks, GS is directly affected by the randomness in loss. Such direct exposure results in unstable updates, as the properties of Gaussians are directly changed during training. Besides, each training step of GS may involve updates to a vast number of Gaussian points. This process occurs without the moderating influence of neural network-style buffering mechanisms. As a result, the excessive fluidity of the 3D GS scene hinders its ability to converge to finely detailed results like implicit representations when trained with generative guidance." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.856, + 0.468, + 0.9 + ], + "angle": 0, + "content": "To counter these issues, in this work, we propose GaussianEditor , a novel, swift, and highly controllable 3D editing algorithm for Gaussian Splatting. GaussianEditor can fulfill" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.092, + 0.892, + 0.333 + ], + "angle": 0, + "content": "various high-quality editing needs within minutes. A key feature of our method is the introduction of Gaussian semantic tracing, which enables precise control over Gaussian Splitting (GS). Gaussian semantic tracing consistently identifies the Gaussians requiring editing at every moment during training. This contrasts with traditional 3D editing methods that often depend on static 2D or 3D masks. Such masks become less effective as the geometries and appearances of 3D models evolve during training. Gaussian semantic tracing is achieved by unprojecting 2D segmentation masks into 3D Gaussians and assigning each Gaussian a semantic tag. As the Gaussians evolve during training, these semantic tags enable the tracking of the specific Gaussians targeted for editing. Our Gaussian tracing algorithm ensures that only the targeted areas are modified, enabling precise and controllable editing." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.337, + 0.892, + 0.639 + ], + "angle": 0, + "content": "Additionally, to tackle the significant challenge of Gaussian Splatting (GS) struggling to fit fine results under highly random generative guidance, we propose a novel GS representation: hierarchical Gaussian splatting (HGS). In HGS, Gaussians are organized into generations based on their sequence in multiple densification processes during training. Gaussians formed in earlier densification stages are deemed older generations and are subject to stricter constraints, aimed at preserving their original state and thus reducing their mobility. Conversely, those formed in later stages are considered younger generations and are subjected to fewer or no constraints, allowing for more adaptability. HGS's design effectively moderates the fluidity of GS by imposing restrictions on older generations while preserving the flexibility of newer generations. This approach enables continuous optimization towards better outcomes, thereby simulating the buffering function achieved in implicit representations through neural networks. Our experiments also demonstrate that HGS is more adept at adapting to highly random generative guidance." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.644, + 0.892, + 0.899 + ], + "angle": 0, + "content": "Finally, we have specifically designed a 3D inpainting algorithm for Gaussian Splatting (GS). As demonstrated in Fig. 1, we have successfully removed specific objects from scenes and seamlessly integrated new objects into designated areas. For object removal, we developed a specialized local repair algorithm that efficiently eliminates artifacts at the intersection of the object and the scene. For adding objects, we first request users to provide a prompt and a 2D inpainting mask for a particular view of the GS. Subsequently, we employ a 2D inpainting method to generate a single-view image of the object to be added. This image is then transformed into a coarse 3D mesh using image-to-3D conversion techniques. The 3D mesh is subsequently converted into the HGS representation and refined. Finally, this refined representation is concatenated into the original GS. The entire inpainting process described above is completed within 5 minutes." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.947, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21477" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.151 + ], + "angle": 0, + "content": "GaussianEditor offers swift, controllable, and versatile 3D editing. A single editing session typically only takes 5-10 minutes, significantly faster than previous editing processes. Our contributions can be summarized in four aspects:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.152, + 0.469, + 0.181 + ], + "angle": 0, + "content": "1. We have introduced Gaussian semantic tracing, enabling more detailed and effective editing control." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.183, + 0.469, + 0.241 + ], + "angle": 0, + "content": "2. We propose Hierarchical Gaussian Splitting (HGS), a novel GS representation capable of converging more stably to refined results under highly random generative guidance." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.243, + 0.469, + 0.287 + ], + "angle": 0, + "content": "3. We have specifically designed a 3D inpainting algorithm for Gaussian Splatting, which allows swift removal and addition objects." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.288, + 0.47, + 0.333 + ], + "angle": 0, + "content": "4. Extensive experiments demonstrate that our method surpasses previous 3D editing methods in terms of effectiveness, speed, and controllability." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.152, + 0.47, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.348, + 0.227, + 0.363 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.373, + 0.264, + 0.389 + ], + "angle": 0, + "content": "2.1. 3D Representations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.397, + 0.47, + 0.502 + ], + "angle": 0, + "content": "Various 3D representations have been proposed to address diverse 3D tasks. The groundbreaking work, Neural Radiance Fields (NeRF) [28], employs volumetric rendering and has gained popularity for enabling 3D optimization with only 2D supervision. However, optimizing NeRF can be time-consuming, despite its wide usage in 3D reconstruction [3, 6, 13, 21, 44] and generation [22, 35] tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.504, + 0.47, + 0.579 + ], + "angle": 0, + "content": "While efforts have been made to accelerate NeRF training [29, 40], these approaches primarily focus on the reconstruction setting, leaving the generation setting less optimized. The common technique of spatial pruning does not effectively speed up the generation setting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.58, + 0.469, + 0.699 + ], + "angle": 0, + "content": "Recently, 3D Gaussian splatting [15] has emerged as an alternative 3D representation to NeRF, showcasing impressive quality and speed in 3D and 4D reconstruction tasks [15, 26, 48, 51, 52]. It has also attracted considerable research interest in the field of generation [7, 45, 53]. Its efficient differentiable rendering implementation and model design facilitate fast training without the need for spatial pruning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.701, + 0.469, + 0.76 + ], + "angle": 0, + "content": "In this work, we pioneer the adaptation of 3D Gaussian splating to 3D editing tasks, aiming to achieve swift and controllable 3D editing, harnessing the advantages of this representation for the first time in this context." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.771, + 0.197, + 0.787 + ], + "angle": 0, + "content": "2.2.3D Editing" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Editing neural fields is inherently challenging due to the intricate interplay between their shape and appearance. EditNeRF [24] stands as a pioneering work in this domain, as they edit both the shape and color of neural fields by conditioning them on latent codes. Additionally, some works [1, 10, 46, 47] leverage CLIP models to facilitate editing through the use of text prompts or reference images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.894, + 0.212 + ], + "angle": 0, + "content": "Another line of research focuses on predefined template models or skeletons to support actions like re-positioning or re-rendering within specific categories [30, 33]. Geometry-based methods [20, 49, 50, 55] translate neural fields into meshes and synchronize mesh deformation with implicit fields. Additionally, 3D editing techniques involve combining 2D image manipulation, such as inpainting, with neural fields training [19, 23]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.213, + 0.894, + 0.32 + ], + "angle": 0, + "content": "Concurrent works [31, 57] leverage static 2D and 3D masks to constrain the edit area of NeRF. However, these approaches have their limitations because the training of 3D models is a dynamic process, and static masks cannot effectively constrain it. In contrast, our research employs Gaussian semantic tracing to track the target Gaussian throughout the entire training process." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.336, + 0.625, + 0.353 + ], + "angle": 0, + "content": "3. Preliminary" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.362, + 0.708, + 0.379 + ], + "angle": 0, + "content": "3.1. 3D Gaussian Splitting" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.386, + 0.892, + 0.476 + ], + "angle": 0, + "content": "GS (Gaussian Splitting) [15] represents an explicit 3D scene using point clouds, where Gaussians are employed to depict the scene's structure. In this representation, every Gaussian is defined by a center point, denoted as \\( x \\), and a covariance matrix \\( \\Sigma \\). The center point \\( x \\) is commonly known as the Gaussian's mean value:" + }, + { + "type": "equation", + "bbox": [ + 0.624, + 0.492, + 0.892, + 0.512 + ], + "angle": 0, + "content": "\\[\nG (x) = e ^ {- \\frac {1}{2} x ^ {T} \\Sigma^ {- 1} x}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.523, + 0.894, + 0.566 + ], + "angle": 0, + "content": "The covariance matrix \\(\\Sigma\\) can be decomposed into a rotation matrix \\(\\mathbf{R}\\) and a scaling matrix \\(\\mathbf{S}\\) for differentiable optimization:" + }, + { + "type": "equation", + "bbox": [ + 0.64, + 0.57, + 0.892, + 0.587 + ], + "angle": 0, + "content": "\\[\n\\Sigma = \\mathbf {R} \\mathbf {S} \\mathbf {S} ^ {T} \\mathbf {R} ^ {T}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.598, + 0.831, + 0.613 + ], + "angle": 0, + "content": "the calculation of gradient flow is detailed in [15]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.718 + ], + "angle": 0, + "content": "For rendering new viewpoints, the method of splatting, as described in [54], is utilized for positioning the Gaussians on the camera planes. This technique, originally presented in [58], involves a viewing transformation denoted by \\(W\\) and the Jacobian \\(J\\) of the affine approximation of the projective transformation. Using these, the covariance matrix \\(\\Sigma^{\\prime}\\) in camera coordinates is determined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.737, + 0.892, + 0.755 + ], + "angle": 0, + "content": "\\[\n\\Sigma^ {\\prime} = J W \\Sigma W ^ {T} J ^ {T}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.894, + 0.901 + ], + "angle": 0, + "content": "To summarize, each Gaussian point in the model is characterized by a set of attributes: its position, denoted as \\( x \\in \\mathbb{R}^3 \\), its color represented by spherical harmonics coefficients \\( c \\in \\mathbb{R}^k \\) (where \\( k \\) indicates the degrees of freedom), its opacity \\( \\alpha \\in \\mathbb{R} \\), a rotation quaternion \\( q \\in \\mathbb{R}^4 \\), and a scaling factor \\( s \\in \\mathbb{R}^3 \\). Particularly, for every pixel, the color and opacity of all Gaussians are calculated based on the Gaussian's representation as described in Eq. 1. The blending process of \\( N \\) ordered points overlapping a pixel follows a specific" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21478" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.079, + 0.273, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.156, + 0.216, + 0.168 + ], + "angle": 0, + "content": "Before Editing" + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.079, + 0.362, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.369, + 0.079, + 0.47, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.331, + 0.156, + 0.412, + 0.169 + ], + "angle": 0, + "content": "During Editing" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.175, + 0.47, + 0.3 + ], + "angle": 0, + "content": "Figure 2. Illustration of Gaussian semantic tracing. Prompt: Turn him into an old lady. The red mask in the images represents the projection of the Gaussians that will be updated and densified. The dynamic change of the masked area during the training process, as driven by the updating of Gaussians, ensures consistent effectiveness throughout the training duration. Despite starting with potentially inaccurate segmentation masks due to 2D segmentation errors, Gaussian semantic tracing still guarantees high-quality editing results." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.32, + 0.138, + 0.332 + ], + "angle": 0, + "content": "formula:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.335, + 0.47, + 0.378 + ], + "angle": 0, + "content": "\\[\nC = \\sum_ {i \\in N} c _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.391, + 0.471, + 0.466 + ], + "angle": 0, + "content": "where \\( c_{i} \\) and \\( \\alpha_{i} \\) signify the color and density of a given point respectively. These values are determined by a Gaussian with a covariance matrix \\( \\Sigma \\), which is then scaled by estimizable per-point opacity and spherical harmonics (SH) color coefficients." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.481, + 0.375, + 0.496 + ], + "angle": 0, + "content": "3.2. Diffusion-based Editing Guidance" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.506, + 0.47, + 0.717 + ], + "angle": 0, + "content": "Recent advancements have seen numerous works elevating 2D diffusion processes to 3D, applying these processes extensively in the realm of 3D editing. Broadly, these works can be categorized into two types. The first type [8, 27, 31, 36, 42, 57], exemplified by Dreamfusion's [36] introduction of SDS loss, involves feeding the noised rendering of the current 3D model, along with other conditions, into a 2D diffusion model [39]. The scores generated by the diffusion model then guide the direction of model updates. The second type [5, 12, 37, 43] focuses on conducting 2D editing based on given prompts for the multiview rendering of a 3D model. This approach creates a multi-view 2D image dataset, which is then utilized as a training target to provide guidance for the 3D model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.868 + ], + "angle": 0, + "content": "Our work centers on leveraging the exemplary properties of Gaussian Splatting's explicit representation to enhance 3D editing. Consequently, we do not design specific editing guidance mechanisms but instead directly employ the guidance methods mentioned above. Both types of guidance can be applied in our method. For simplicity, we denote the guidance universally as \\( D \\). Given the parameters of a 3D model, \\( \\Theta \\), along with the rendered camera pose \\( p \\) and prompt \\( e \\), the editing loss from the 2D diffusion prior can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.209, + 0.886, + 0.47, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {E d i t}} = D (\\Theta ; p, e) \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.59, + 0.105 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.894, + 0.221 + ], + "angle": 0, + "content": "We define the task of 3D editing on Gaussian Splatting (GS) as follows: Given a prompt \\( y \\) and a 3D scene represented by 3D Gaussians, denoted by \\( \\Theta \\), where each \\( \\Theta_{i} = \\{x_{i}, s_{i}, q_{i}, \\alpha_{i}, c_{i}\\} \\) represents the parameters of the \\( i \\)-th Gaussian as detailed in Sec. 3.1, the objective is to achieve an edited 3D Gaussians, referred to as \\( \\Theta_{y} \\), that aligns with or adheres to the specifications of the prompt \\( y \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.222, + 0.895, + 0.402 + ], + "angle": 0, + "content": "We then introduce our novel framework for performing editing tasks on GS. We first introduce Gaussian semantic tracing in Sec. 4.1, along with a new representation method known as Hierarchical Gaussian Splatting (HGS) in Sec. 4.2. The GS semantic tracing enables precise segmentation and tracing within GS, facilitating controllable editing operations. Compared to the standard GS, the HGS representation demonstrates greater robustness against randomness in generative guidance and is more adept at accommodating a diverse range of editing scenarios. Additionally, we have specifically designed 3D inpainting for GS, which encompasses object removal and addition (Sec. 4.3)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.41, + 0.747, + 0.426 + ], + "angle": 0, + "content": "4.1. Gaussian Semantic Tracing" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.433, + 0.894, + 0.614 + ], + "angle": 0, + "content": "Previous works [31, 57] in 3D editing usually utilize static 2D or 3D masks to apply loss only within the masked pixels, thus constraining the editing process to only edit the desired area. However, this method has limitations. As 3D representations dynamically change during training, static segmentation masks would become inaccurate or even ineffective. Furthermore, the use of static masks to control gradients in NeRF editing poses a significant limitation, as it confines the editing strictly within the masked area. This restriction prevents the edited content from naturally extending beyond the mask, thus 'locking' the content within a specified spatial boundary." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.615, + 0.892, + 0.689 + ], + "angle": 0, + "content": "To address the aforementioned issue, we have chosen Gaussian Splitting (GS) as our 3D representation due to its explicit nature. This allows us to directly assign semantic labels to each Gaussian point, thereby facilitating semantic tracing in 3D scenes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.894, + 0.825 + ], + "angle": 0, + "content": "Specifically, we enhance the 3D Gaussians \\(\\Theta\\) by adding a new attribute \\(m\\), where \\(m_{ij}\\) represents the semantic Gaussian mask for the \\(i\\)-th Gaussian point and the \\(j\\)-th semantic label. With this attribute, we can precisely control the editing process by selectively updating only the target 3D Gaussians. During the densification process, newly densified points inherit the semantic label of their parent point. This ensures that we have an accurate 3D semantic mask at every moment throughout the training process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.9 + ], + "angle": 0, + "content": "As illustrated in Fig. 2, Gaussian semantic tracing enables continuous tracking of each Gaussian's categories during training, adjusting to their evolving properties and numbers. This feature is vital, as it permits selective application of gradients, densification and pruning of Gaussians linked to the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21479" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.269, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.141, + 0.189, + 0.207, + 0.201 + ], + "angle": 0, + "content": "Original GS" + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.089, + 0.468, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.301, + 0.189, + 0.447, + 0.2 + ], + "angle": 0, + "content": "A Garfield cat on the bench" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.202, + 0.268, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.138, + 0.302, + 0.209, + 0.312 + ], + "angle": 0, + "content": "Novel view 1" + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.202, + 0.464, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.334, + 0.302, + 0.407, + 0.312 + ], + "angle": 0, + "content": "Novel view 2" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.315, + 0.471, + 0.371 + ], + "angle": 0, + "content": "Figure 3. 3D inpainting for object incorporation. GaussianEditor is capable of adding objects at specified locations in a scene, given a 2D inpainting mask and a text prompt from a single view. The whole process takes merely five minutes." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.388, + 0.207, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.105, + 0.524, + 0.182, + 0.535 + ], + "angle": 0, + "content": "Original View" + }, + { + "type": "image", + "bbox": [ + 0.21, + 0.388, + 0.338, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.524, + 0.301, + 0.535 + ], + "angle": 0, + "content": "Removed" + }, + { + "type": "image", + "bbox": [ + 0.34, + 0.388, + 0.469, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.379, + 0.524, + 0.431, + 0.535 + ], + "angle": 0, + "content": "Inpainted" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.542, + 0.47, + 0.626 + ], + "angle": 0, + "content": "Figure 4. 3D inpainting for object removal. Typically, removing the target object based on a Gaussian semantic mask generates artifacts at the interface between the target object and the scene. To address this, we generate a repaired image using a 2D inpainting method and employ Mean Squared Error (MSE) loss for supervision. The whole process takes merely two minutes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.656, + 0.47, + 0.761 + ], + "angle": 0, + "content": "specified category. Additionally, it facilitates training solely by rendering the target object, significantly speeding up the process in complex scenes. The semantic Gaussian mask \\( m \\) functions as a dynamic 3D segmentation mask, evolving with the training, allowing content to expand freely in space. This contrasts with NeRF, where content is restricted to a fixed spatial area." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.763, + 0.471, + 0.869 + ], + "angle": 0, + "content": "Next, we discuss Gaussian Splating unprojection, the method we propose to obtain semantic Gaussian mask \\( m \\). For a set of 3D Gaussians \\( \\Theta \\), we render them from multiple viewpoints to generate a series of renderings \\( \\mathcal{I} \\). These renderings are then processed using 2D segmentation techniques [18] to obtain 2D segmentation masks \\( \\mathcal{M} \\), with each \\( \\mathcal{M}_j \\), representing the \\( j \\)-th semantic labels." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.901 + ], + "angle": 0, + "content": "To obtain the semantic label for each Gaussian, we unproject the posed 2D semantic label back to the Gaussians" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.893, + 0.152 + ], + "angle": 0, + "content": "with inverse rendering. Concretely, we maintain a weight and a counter for each Gaussian. For pixel \\( p \\) on the semantic maps, we unproject the semantic label back to the Gaussians that affects it by" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.165, + 0.893, + 0.188 + ], + "angle": 0, + "content": "\\[\nw _ {i} ^ {j} = \\sum o _ {i} (\\boldsymbol {p}) * T _ {i} ^ {j} (\\boldsymbol {p}) * \\mathcal {M} ^ {j} (\\boldsymbol {p}), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.196, + 0.895, + 0.318 + ], + "angle": 0, + "content": "where \\( w_{i}^{j} \\) represents the weight of the \\( i \\)-th Gaussian for the \\( j \\)-th semantic label, while \\( o_{i}(\\pmb{p}) \\), \\( T_{i}^{j}(\\pmb{p}) \\), and \\( \\mathcal{M}^j (\\pmb{p}) \\) denote the opacity, transmittance from pixel \\( \\pmb{p} \\), and semantic mask of pixel \\( \\pmb{p} \\) for the \\( i \\)-th Gaussian, respectively. After updating all the Gaussian weights and counters, we determine whether a Gaussian belongs to the \\( j \\)-th semantic class based on whether its average weight exceeds a manually set threshold." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.319, + 0.895, + 0.456 + ], + "angle": 0, + "content": "The entire labeling process is remarkably fast, typically taking less than a second. Once this semantic label assignment is completed, the entire Gaussian scene becomes parsed by us, making a variety of operations possible. Notably, 2D diffusion guidance often struggles to effectively edit small objects in complex scenes. Thanks to Gaussian semantic tracing, we can now render these small objects independently and input them into the 2D diffusion model, thereby achieving more precise supervision." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.466, + 0.783, + 0.483 + ], + "angle": 0, + "content": "4.2. Hierarchical Gaussian Splitting" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.892, + 0.553 + ], + "angle": 0, + "content": "The effectiveness of vanilla GS [17] in reconstruction tasks lies in the high-quality initialization provided by point clouds derived from SFM [41], coupled with stable supervision from ground truth datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.894, + 0.734 + ], + "angle": 0, + "content": "However, the scenario changes in the field of generation. In previous work involving GS in text-to-3D and image-to-3D [7, 45, 53], GS has shown limitations when facing the randomness of generative guidance due to its nature as a point cloud-like representation. This instability in GS is mainly due to their direct exposure to the randomness of loss functions, unlike implicit representations. GS models, which update a large number of Gaussian points each training step, lack the memorization and moderating ability of neural networks. This leads to erratic updates and prevents GS from achieving the detailed results as GS's excessive fluidity hampers its convergence in generative training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.894, + 0.87 + ], + "angle": 0, + "content": "To address these challenges, we introduce Hierarchical Gaussian Splitting (HGS), a structured representation of GS that is more suitable for generative and editing scenarios. HGS categorizes GS into different generations based on the densification round in which a particular Gaussian point is produced. The initial Gaussians \\(\\Theta\\), are all assigned a generation of 0. During the training process for editing, points generated in the \\(k\\)-th densification round are marked as generation \\(k\\)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Subsequently, we impose varying constraints on Gaussians from different generations to control their degree of" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21480" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.099, + 0.238, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.201, + 0.238, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.294, + 0.192, + 0.305 + ], + "angle": 0, + "content": "Original view" + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.062, + 0.449, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.148, + 0.448, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.236, + 0.448, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.287, + 0.323, + 0.407, + 0.334 + ], + "angle": 0, + "content": "\"Turn him into a clown\"" + }, + { + "type": "image", + "bbox": [ + 0.455, + 0.062, + 0.66, + 0.146 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.455, + 0.148, + 0.66, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.455, + 0.236, + 0.66, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.482, + 0.323, + 0.636, + 0.334 + ], + "angle": 0, + "content": "\"Turn him into Albert Einstein\"" + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.062, + 0.869, + 0.146 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.148, + 0.87, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.236, + 0.87, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.675, + 0.323, + 0.86, + 0.335 + ], + "angle": 0, + "content": "\"Turn him into a Modigliani painting.\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.873, + 0.075, + 0.885, + 0.126 + ], + "angle": 90, + "content": "Instruct-N2N" + }, + { + "type": "image_caption", + "bbox": [ + 0.873, + 0.168, + 0.885, + 0.227 + ], + "angle": 90, + "content": "GSEeditor-DDS" + }, + { + "type": "image_caption", + "bbox": [ + 0.873, + 0.253, + 0.885, + 0.314 + ], + "angle": 90, + "content": "GSEoperator-iN2N" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.341, + 0.894, + 0.397 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison. It's important to note the level of control we maintain over the editing area (the whole body of the man). Background and other non-target regions are essentially unaffected, in contrast to Instruct-Nerf2Nerf [12] where the entire scene undergoes changes. GaussianEditor-DDS and GaussianEditor-iN2N indicate that we utilize delta denoising score [14] and Instruct-Nerf2Nerf [12] respectively, as guidance for editing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.409, + 0.471, + 0.545 + ], + "angle": 0, + "content": "flexibility. The older the generation, the stronger the constraints applied. Anchor loss is utilized to enforce these constraints. At the beginning of training, HGS records the attributes of all Gaussians as anchors. These anchors are then updated to reflect the current state of the Gaussians at each densification process. During training, MSE loss between the anchor state and the current state is employed to ensure that the Gaussians do not deviate too far from their respective anchors:" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.553, + 0.47, + 0.592 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a n c h o r}} ^ {P} = \\sum_ {i = 0} ^ {n} \\lambda_ {i} \\left(P _ {i} - \\hat {P} _ {i}\\right) ^ {2} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.47, + 0.703 + ], + "angle": 0, + "content": "where \\( n \\) represents the total number of Gaussians and \\( P \\) denotes a certain property of the current Gaussian, including elements from the set \\( x, s, q, \\alpha, c \\). Here, \\( \\hat{P} \\) refers to the same property recorded in the anchor state. The term \\( \\lambda_{i} \\) indicates the strength of the anchor loss applied to the \\( i \\)-th Gaussian, which varies based on its generation. The overall training loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.709, + 0.47, + 0.743 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {E d i t} + \\sum_ {P \\in \\{x, s, q, \\alpha , c \\}} \\lambda_ {P} \\mathcal {L} _ {a n c h o r} ^ {P} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.794 + ], + "angle": 0, + "content": "In this equation, \\(\\lambda_{P}\\) signifies the strength of the anchor loss applied to property \\(P\\), and \\(\\mathcal{L}_{Edit}\\) is the edit loss defined in Sec. 3.2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.471, + 0.902 + ], + "angle": 0, + "content": "This generational design in HGS prevents the issue of excessive flexibility in GS when faced with stochastic loss. With each densification, the anchor loss weight \\(\\lambda_{i}\\) for all previous generations of Gaussians is increased. As a result, the fluidity of the existing generations gradually decreases until it nearly solidifies. This approach ensures stable geometry formation under stochastic losses, relying on the almost" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.41, + 0.893, + 0.515 + ], + "angle": 0, + "content": "unconstrained Gaussians from new densifications to carve out details. Furthermore, this method of applying anchor loss can effectively meet various editing needs. For instance, to limit changes in the original GS, one can increase the anchor loss weight for generation 0. Similarly, if there is no desire to alter color or geometry during editing, a stronger anchor loss can be applied to these specific properties." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.516, + 0.894, + 0.637 + ], + "angle": 0, + "content": "Additionally, to address the challenge of manually determining a densification threshold, we regulate the densification process based on a percentage criterion. In this method, during each densification step, we selectively densify only those Gaussians whose 3D position gradients are within the top \\( k\\% \\). This strategy proves to be more manageable and intuitive than directly setting a threshold value in the Hierarchical Gaussian Splatting (HGS) framework." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.65, + 0.644, + 0.666 + ], + "angle": 0, + "content": "4.3. 3D Inpainting" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.673, + 0.893, + 0.87 + ], + "angle": 0, + "content": "Object Removal. Simply removing Gaussians identified by a mask can lead to artifacts, especially at the interfaces where the object intersects with other Gaussians. To address this, we employ 2D inpainting techniques to provide guidance for filling these areas. However, effective 2D inpainting requires precise masks to offer better guidance. To generate these masks, after deletion, we use the KNN algorithm to identify Gaussians nearest to the ones removed, which are likely at the interface. These are then projected onto various views. We subsequently dilate the mask and fix any holes to accurately represent the interface area, thus creating a refined mask for the boundary zones. The whole object removal procedure typically takes only two minutes." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Object Incorporation. We define this task as follows: Within the 3D Gaussians \\(\\theta\\), given a camera pose \\(p\\) and the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21481" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.062, + 0.278, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.06, + 0.483, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.488, + 0.06, + 0.687, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.06, + 0.892, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.425, + 0.166, + 0.537, + 0.179 + ], + "angle": 0, + "content": "\"Make it snowy\"" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.182, + 0.276, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.182, + 0.48, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.488, + 0.182, + 0.687, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.182, + 0.892, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.389, + 0.299, + 0.573, + 0.313 + ], + "angle": 0, + "content": "\"Turn him into an old lady\"" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.315, + 0.279, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.315, + 0.483, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.488, + 0.315, + 0.687, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.315, + 0.892, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.37, + 0.432, + 0.6, + 0.446 + ], + "angle": 0, + "content": "\"Turn the bear into a grizzly bear\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.457, + 0.895, + 0.499 + ], + "angle": 0, + "content": "Figure 6. Extensive Results of GaussianEditor. Our method is capable of various editing tasks, including face and scene editing. In face and bear editing, we restrict the editing area to the face using Gaussian semantic tracing, ensuring that undesired areas remain unchanged. The leftmost column demonstrates the original view, while the right three columns show the images after editing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.519, + 0.471, + 0.595 + ], + "angle": 0, + "content": "corresponding rendering \\( I \\) from this viewpoint, the user provides a 2D mask \\( M \\) on \\( I \\) indicating the area they wish to inpaint. Additionally, a prompt \\( y \\) is provided to specify the content of the inpainting. We then update \\( \\theta \\) to fulfill the inpainting request." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.601, + 0.47, + 0.707 + ], + "angle": 0, + "content": "Given \\( I \\), \\( M \\), and \\( y \\), the process begins with generating a 2D inpainted image \\( I^M y \\), utilizing a 2D inpainting diffusion model as per [34]. Subsequently, the foreground object from \\( I^M y \\), created by [34], is segmented and input into the image-to-3D method referenced in [25] to generate a coarse 3D mesh. This coarse mesh is then transformed into 3D Gaussians \\( \\theta_y \\), and refined with HGS detailed in Sec. 4.2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.713, + 0.471, + 0.85 + ], + "angle": 0, + "content": "For aligning the coordinate system of \\(\\theta_y\\) with \\(\\theta\\), the depth of \\(I_y^M\\) is first estimated using the technique from [38]. This depth is then aligned with the depth map rendered by \\(\\theta\\) at camera pose \\(p\\), using the least squares method. With this alignment, we can accurately determine the coordinates and scale of the inpainted foreground object in the coordinate system of \\(\\theta\\). After transforming \\(\\theta_y\\) into the coordinate system of \\(\\theta\\), we simply concatenate them to produce the final inpainted 3D Gaussians." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.902 + ], + "angle": 0, + "content": "It is important to note that due to our efficient design, the entire object incorporation procedure can be completed in approximately 5 minutes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.517, + 0.633, + 0.535 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.543, + 0.719, + 0.559 + ], + "angle": 0, + "content": "5.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.895, + 0.688 + ], + "angle": 0, + "content": "We utilize the highly optimized renderer implementation from [16] for Gaussian rendering and base our implementation on Threestudio [11]. All the original 3D Gaussians used in this work are trained using the methods described in [16]. Our experiments are conducted on a single RTX A6000 GPU. As detailed in Sec. 4.1, once we obtain segmentation masks from the 2D segmentation method outlined in [18], segmenting the 3D Gaussians takes only about 1 second." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.809 + ], + "angle": 0, + "content": "For editing large scenes, the camera poses employed during the editing process are selected from a subset of the multi-view image dataset initially used for reconstruction. Depending on the complexity of the scene, the number of camera poses used in our experiments varies from 24 to 96. The editing process, influenced by the specified prompt and the complexity of the scene, typically involves optimizing for 500-1000 steps, taking about 5-10 minutes in total." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.895, + 0.9 + ], + "angle": 0, + "content": "Regarding 3D inpainting for object incorporation, as detailed in Sec. 4.3, it takes approximately 3 minutes to generate a 3D mesh using the method from [25] and an additional 2 minutes to transfer this mesh into 3D Gaussians and refine it, while the composition process of two Gaussians takes less than 1 second." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21482" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.209, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.105, + 0.223, + 0.182, + 0.235 + ], + "angle": 0, + "content": "Original View" + }, + { + "type": "image", + "bbox": [ + 0.211, + 0.089, + 0.338, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.265, + 0.224, + 0.283, + 0.234 + ], + "angle": 0, + "content": "GS" + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.089, + 0.469, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.391, + 0.224, + 0.419, + 0.234 + ], + "angle": 0, + "content": "HGS" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.247, + 0.473, + 0.359 + ], + "angle": 0, + "content": "Figure 7. Ablation study on Hierarchical Gaussian Splitting (HGS). Prompt: make the grass on fire. Even when specifying the editing area with prompts, generative methods like InstructPix2Pix [4] tend to edit the entire 2D image. Without HGS, Gaussians tend to conform to this whole-image editing by spreading and densifying across the entire scene, leading to uncontrollable densification and blurring of the image. With HGS, however, this kind of diffusion is effectively restrained." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.386, + 0.307, + 0.402 + ], + "angle": 0, + "content": "5.2. Qualitative Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.409, + 0.471, + 0.529 + ], + "angle": 0, + "content": "As illustrated in Fig. 5, GaussianEditor-iN2N surpasses other methods in both the quality of edits and controllability. Instruct-Nerf2Nerf, while producing edits with insufficient detail, cannot also control the editing area. GaussianEditorDDS, due to the more challenging control of guidance offered by DDS loss [14] compared to Instruct-pix2pix [4], tends to result in oversaturated colors and less precise editing outcomes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.53, + 0.47, + 0.682 + ], + "angle": 0, + "content": "Additionally, our method exhibits exceptional control over the editing area. This is achieved through Gaussian semantic tracing, which identifies the Gaussians that require editing at each training step, for example, the entire human body in Fig. 5. It's important to note that in Instruct-Nerf2Nerf [12], the use of static 2D or 3D masks restricts the spatial freedom of the edits, as the permissible area for the edited subject is limited by these masks. Furthermore, the effectiveness of static masks diminishes as the geometries and appearances of 3D models evolve during training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.682, + 0.472, + 0.773 + ], + "angle": 0, + "content": "In Fig. 6, we demonstrate that GaussianEditor can accommodate a variety of scenarios, such as editing in large-scale scenes and facial swaps. In the case of large scenes, we did not apply Gaussian semantic tracing. However, for facial swaps, we traced the Gaussians corresponding to facial regions, achieving controllable and realistic editing." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.783, + 0.317, + 0.797 + ], + "angle": 0, + "content": "5.3.Quantitative Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.797, + 0.472, + 0.901 + ], + "angle": 0, + "content": "As shown in Table 1, we conduct quantitative comparisons on user study and CLIP directional similarity (as shown in InstructPix2Pix [4] and StyleGAN-Nada [9]). GaussianEditor-iN2N not only demonstrates superior outcomes in user studies but also excels in CLIP Directional Similarity. Besides, Instruct-Nerf2Nerf [12] typically requires more than 30 minutes to complete the editing of a scene, whereas our method" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.09, + 0.892, + 0.14 + ], + "angle": 0, + "content": "
iN2N [12]Ours (DDS)Ours (iN2N)
User study15.45%12.27%72.28%
CLIP Directional Similarity0.16000.18130.2071
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.151, + 0.895, + 0.193 + ], + "angle": 0, + "content": "Table 1. Quantitative Comparison. GaussianEditor-iN2N outperforms in both user study evaluations and CLIP Directional Similarity [9] metrics." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.218, + 0.738, + 0.233 + ], + "angle": 0, + "content": "only takes between 5 to 10 minutes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.241, + 0.653, + 0.258 + ], + "angle": 0, + "content": "5.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.264, + 0.895, + 0.416 + ], + "angle": 0, + "content": "As demonstrated in Fig. 7, we conducted ablation experiments on Hierarchical Gaussian Splitting(HGS). Without HGS, Gaussians tend to spread and densify across the scene, leading to uncontrolled densification and image blurring. This is typically caused by the tendency of methods like Instruct-Pix2Pix to edit the entire 2D image when prompts are used to define editing areas. However, HGS effectively circumvents this issue by constraining the mobility of the Gaussian in the old generation, ensuring that the overall scene does not exhibit excessive mobility." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.428, + 0.619, + 0.443 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.453, + 0.895, + 0.604 + ], + "angle": 0, + "content": "In our research, we introduce GaussianEditor, an innovative 3D editing algorithm based on Gaussian Splatting, designed for enhanced control and efficiency. Our method employs Gaussian semantic tracing for precise identification and targeting of editing areas, followed by Hierarchical Gaussian Splatting (HGS) to balance fluidity and stability in achieving detailed results under stochastic guidance. Additionally, we developed a specialized 3D inpainting algorithm for Gaussian Splatting, streamlining object removal and integration, and greatly reducing editing time." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.605, + 0.895, + 0.68 + ], + "angle": 0, + "content": "Limitation. Similar to previous 3D editing works based on 2D diffusion models, GaussianEditor relies on these models to provide effective supervision. However, current 2D diffusion models struggle to offer effective guidance for certain complex prompts, leading to limitations in 3D editing." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.692, + 0.688, + 0.71 + ], + "angle": 0, + "content": "7. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.718, + 0.894, + 0.793 + ], + "angle": 0, + "content": "This study is supported under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). This research is also supported by the MoE AcRF Tier 2 grant (MOE-T2EP20220-0007)." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.806, + 0.597, + 0.821 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.508, + 0.83, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[1] Chong Bao, Yinda Zhang, and Bangbang et al. Yang. Sine: Semantic-driven image-based nef editing with prior-guided editing field. In CVPR 2023, pages 20919-20929, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.873, + 0.894, + 0.9 + ], + "angle": 0, + "content": "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan." + }, + { + "type": "list", + "bbox": [ + 0.508, + 0.83, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21483" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.135 + ], + "angle": 0, + "content": "Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855-5864, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.136, + 0.472, + 0.177 + ], + "angle": 0, + "content": "[3] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.179, + 0.472, + 0.22 + ], + "angle": 0, + "content": "[4] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. arXiv preprint arXiv:2211.09800, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.222, + 0.471, + 0.276 + ], + "angle": 0, + "content": "[5] Yiwen Chen, Chi Zhang, Xiaofeng Yang, Zhongang Cai, Gang Yu, Lei Yang, and Guosheng Lin. It3d: Improved text-to-3d generation with explicit view synthesis. arXiv preprint arXiv:2308.11473, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.278, + 0.471, + 0.332 + ], + "angle": 0, + "content": "[6] Zhiqin Chen, Thomas Funkhouser, Peter Hedman, and Andrea Tagliasacchi. Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. arXiv preprint arXiv:2208.00277, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.334, + 0.471, + 0.375 + ], + "angle": 0, + "content": "[7] Zilong Chen, Feng Wang, and Huaping Liu. Text-to-3d using gaussian splatting. arXiv preprint arXiv:2309.16585, 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.377, + 0.471, + 0.433 + ], + "angle": 0, + "content": "[8] Xinhua Cheng, Tianyu Yang, Jianan Wang, Yu Li, Lei Zhang, Jian Zhang, and Li Yuan. Progressive3d: Progressively local editing for text-to-3d content creation with complex semantic prompts. arXiv preprint arXiv:2310.11784, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.435, + 0.471, + 0.489 + ], + "angle": 0, + "content": "[9] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clipguided domain adaptation of image generators. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.491, + 0.471, + 0.545 + ], + "angle": 0, + "content": "[10] William Gao, Noam Aigerman, Thibault Groueix, Vladimir G Kim, and Rana Hanocka. Textdeformer: Geometry manipulation using text guidance. arXiv preprint arXiv:2304.13348, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.471, + 0.616 + ], + "angle": 0, + "content": "[11] Yuan-Chen Guo, Ying-Tian Liu, Chen Wang, Zi-Xin Zou, Guan Luo, Chia-Hao Chen, Yan-Pei Cao, and Song-Hai Zhang. threestudio: A unified framework for 3d content generation. https://github.com/threestudio-project/threestudio, 2023.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.618, + 0.471, + 0.672 + ], + "angle": 0, + "content": "[12] Ayaan Haque, Matthew Tancik, Alexei A Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. arXiv preprint arXiv:2303.12789, 2023. 2, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.471, + 0.716 + ], + "angle": 0, + "content": "[13] Peter Hedman, Pratul P. Srinivasan, Ben Mildenhall, Jonathan T. Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.471, + 0.771 + ], + "angle": 0, + "content": "[14] Amir Hertz, Kfir Aberman, and Daniel Cohen-Or. Delta denoising score. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2328-2337, 2023. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.471, + 0.828 + ], + "angle": 0, + "content": "[15] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 42(4): 1-14, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.471, + 0.871 + ], + "angle": 0, + "content": "[16] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ToG, 42(4):1-14, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.894, + 0.119 + ], + "angle": 0, + "content": "field rendering. ACM Transactions on Graphics, 42(4), 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.894, + 0.177 + ], + "angle": 0, + "content": "[18] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.894, + 0.22 + ], + "angle": 0, + "content": "[19] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nerf for editing via feature field distillation. arXiv preprint arXiv:2205.15585, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.222, + 0.894, + 0.275 + ], + "angle": 0, + "content": "[20] Yuan Li, Zhi-Hao Lin, David Forsyth, Jia-Bin Huang, and Shenlong Wang. Climatenerf: Physically-based neural rendering for extreme climate synthesis. arXiv e-prints, pages arXiv-2211, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.278, + 0.894, + 0.332 + ], + "angle": 0, + "content": "[21] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.894, + 0.389 + ], + "angle": 0, + "content": "[22] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, pages 300–309, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.391, + 0.894, + 0.432 + ], + "angle": 0, + "content": "[23] Hao-Kang Liu, I Shen, Bing-Yu Chen, et al. Nerf-in: Free-form nerf inpainting with rgb-d priors. arXiv preprint arXiv:2206.04901, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.434, + 0.894, + 0.475 + ], + "angle": 0, + "content": "[24] Steven Liu, Xiuming Zhang, Zhoutong Zhang, Richard Zhang, Jun-Yan Zhu, and Bryan Russell. Editing conditional radiance fields. In ICCV 2021, pages 5773-5783, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.477, + 0.894, + 0.545 + ], + "angle": 0, + "content": "[25] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. arXiv preprint arXiv:2310.15008, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.547, + 0.894, + 0.601 + ], + "angle": 0, + "content": "[26] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.604, + 0.894, + 0.671 + ], + "angle": 0, + "content": "[27] Aryan Mikaeili, Or Perel, Mehdi Safaee, Daniel Cohen-Or, and Ali Mahdavi-Amiri. Sked: Sketch-guided text-based 3d editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14607-14619, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.674, + 0.894, + 0.729 + ], + "angle": 0, + "content": "[28] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.731, + 0.894, + 0.785 + ], + "angle": 0, + "content": "[29] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.788, + 0.894, + 0.829 + ], + "angle": 0, + "content": "[30] Atsuhiro Noguchi, Xiao Sun, Stephen Lin, and Tatsuya Harada. Neural articulated radiance field. In ICCV 2021, pages 5762-5772, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.831, + 0.894, + 0.871 + ], + "angle": 0, + "content": "[31] Jangho Park, Gihyun Kwon, and Jong Chul Ye. Ed-nerf: Efficient text-guided editing of 3d scene using latent space nerf. arXiv preprint arXiv:2310.02712, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[32] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo Martin-" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21484" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "Brualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[33] Sida Peng, Yuanqing Zhang, Yinghao Xu, and et al. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In CVPR 2021, pages 9054–9063, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.472, + 0.247 + ], + "angle": 0, + "content": "[34] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.472, + 0.289 + ], + "angle": 0, + "content": "[35] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.292, + 0.472, + 0.359 + ], + "angle": 0, + "content": "[36] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.472, + 0.431 + ], + "angle": 0, + "content": "[37] Amit Raj, Srinivas Kaza, Ben Poole, Michael Niemeyer, Nataniel Ruiz, Ben Mildenhall, Shiran Zada, Kfir Aberman, Michael Rubinstein, Jonathan Barron, et al. Dreambooth3d: Subject-driven text-to-3d generation. arXiv preprint arXiv:2303.13508, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.433, + 0.472, + 0.489 + ], + "angle": 0, + "content": "[38] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12179-12188, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.472, + 0.544 + ], + "angle": 0, + "content": "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.547, + 0.472, + 0.587 + ], + "angle": 0, + "content": "[40] Sara Fridovich-Keil and Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.59, + 0.472, + 0.643 + ], + "angle": 0, + "content": "[41] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.646, + 0.472, + 0.701 + ], + "angle": 0, + "content": "[42] Etai Sella, Gal Fiebelman, Peter Hedman, and Hadar Averbuch-Elor. Vox-e: Text-guided voxel editing of 3d objects. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 430-440, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.703, + 0.472, + 0.758 + ], + "angle": 0, + "content": "[43] Ruizhi Shao, Jingxiang Sun, Cheng Peng, Zerong Zheng, Boyao Zhou, Hongwen Zhang, and Yebin Liu. Control4d: Dynamic portrait editing by learning 4d gan from 2d diffusion-based editor. arXiv preprint arXiv:2305.20082, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.472, + 0.815 + ], + "angle": 0, + "content": "[44] Chaoyue Song, Tianyi Chen, Yiwen Chen, Jiacheng Wei, Chuan Sheng Foo, Fayao Liu, and Guosheng Lin. Moda: Modeling deformable 3d objects from casual videos. arXiv preprint arXiv:2304.08279, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.472, + 0.87 + ], + "angle": 0, + "content": "[45] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[46] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.893, + 0.134 + ], + "angle": 0, + "content": "of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3835-3844, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.894, + 0.19 + ], + "angle": 0, + "content": "[47] Can Wang, Ruixiang Jiang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Nerf-art: Text-driven neural radiance fields stylization. IEEE Transactions on Visualization and Computer Graphics, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.894, + 0.247 + ], + "angle": 0, + "content": "[48] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.894, + 0.304 + ], + "angle": 0, + "content": "[49] Tianhan Xu and Tatsuya Harada. Deforming radiance fields with cages. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIII, pages 159-175. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.894, + 0.36 + ], + "angle": 0, + "content": "[50] Bangbang Yang, Chong Bao, and Junyi et al. Zeng. Neumesh: Learning disentangled neural mesh-based implicit field for geometry and texture editing. In ECCV 2022, pages 597-614. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.893, + 0.417 + ], + "angle": 0, + "content": "[51] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.419, + 0.893, + 0.473 + ], + "angle": 0, + "content": "[52] Zeyu Yang, Hongye Yang, Zijie Pan, Xiatian Zhu, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. arXiv preprint arXiv:2310.10642, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.475, + 0.894, + 0.543 + ], + "angle": 0, + "content": "[53] Taoran Yi, Jiemin Fang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussian-dreamer: Fast generation from text to 3d gaussian splatting with point cloud priors. arXiv preprint arXiv:2310.08529, 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.546, + 0.893, + 0.6 + ], + "angle": 0, + "content": "[54] Wang Yifan, Felice Serena, Shihao Wu, Cengiz Öztireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM Transactions on Graphics (TOG), 38(6):1-14, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.602, + 0.893, + 0.643 + ], + "angle": 0, + "content": "[55] Yu-Jie Yuan, Yang-Tian Sun, Yu-Kun Lai, and et al. Nerf-editing: geometry editing of neural radiance fields. In CVPR 2022, pages 18353-18364, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.645, + 0.893, + 0.686 + ], + "angle": 0, + "content": "[56] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.688, + 0.893, + 0.73 + ], + "angle": 0, + "content": "[57] Jingyu Zhuang, Chen Wang, Lingjie Liu, Liang Lin, and Guanbin Li. Dreameditor: Text-driven 3d scene editing with neural fields. arXiv preprint arXiv:2306.13455, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.731, + 0.893, + 0.786 + ], + "angle": 0, + "content": "[58] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Surface splatting. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 371-378, 2001. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.786 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21485" + } + ] +] \ No newline at end of file diff --git a/2024/GaussianEditor_ Swift and Controllable 3D Editing with Gaussian Splatting/f458e425-2ba9-468a-94e0-95be47f99560_origin.pdf b/2024/GaussianEditor_ Swift and Controllable 3D Editing with Gaussian Splatting/f458e425-2ba9-468a-94e0-95be47f99560_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5deab527b26a911d89675795f754e3f9888e332e --- /dev/null +++ b/2024/GaussianEditor_ Swift and Controllable 3D Editing with Gaussian Splatting/f458e425-2ba9-468a-94e0-95be47f99560_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:07005488002d134000aa24b428f47e26582ac1b1fcec77fb64a6376f31b61acc +size 3897752 diff --git a/2024/GaussianEditor_ Swift and Controllable 3D Editing with Gaussian Splatting/full.md b/2024/GaussianEditor_ Swift and Controllable 3D Editing with Gaussian Splatting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..50e287a021a090dbdaaff7ee0a9301d59416eedd --- /dev/null +++ b/2024/GaussianEditor_ Swift and Controllable 3D Editing with Gaussian Splatting/full.md @@ -0,0 +1,445 @@ +# GaussianEditor: Swift and Controllable 3D Editing with Gaussian Splitting + +Yiwen Chen\*1,2 Yikai Wang3 + +Zilong Chen\*3 Zhongang Cai4 + +Chi Zhang $^{2}$ Feng Wang $^{3}$ Lei Yang $^{4}$ Huaping Liu $^{3}$ + +Xiaofeng Yang $^{2}$ Guosheng Lin $^{**1,2}$ + +$^{1}$ S-Lab, Nanyang Technological University + +$^{2}$ School of Computer Science and Engineering, Nanyang Technological University + +$^{3}$ Department of Computer Science and Technology, Tsinghua University + +4 SenseTime Research + +https://buaacyw.github.io/gaussian-editor + +![](images/b1755dd51b2004f2668c564909ff8df1419c4f5ec82753cd6286fdac4e74143c.jpg) +Original view + +![](images/1b317d9448d69eddea2abdff283e407cd469970daf35dd808042852c4e1a8385.jpg) + +![](images/87f5389d17dc044a89bf047023a9ee8265a427318ac2923cf56450f7914413c6.jpg) +Original view + +![](images/4d01b8eb2a4c0777f8d5bf2daab7928aba67db5ecc08b91fe83e9b49a67c7ad6.jpg) + +![](images/fd0245785b004ba71a77391a241d951b1ce1e46f9d3e08894430bb300626f5a4.jpg) +"Delete the bike. Add a dog on the bench." + +![](images/0f9f38f043f6dd091e6966a5d2cc8d1f9658026ec88df06a0c39d6b3352af481.jpg) + +![](images/c980c3a12f5374acd59b22c36dc54b6cab85718f61652d0c78e4ee60cd0c0829.jpg) +"Make the grass on fire" + +![](images/a1275c47cdc9ddf269efd347734e12ff216b4c517ebe38ac7b5f27711707913a.jpg) + +![](images/c947c3270de861600e875163c95f6d05763b917f5a067cfa60bea57c2bbd282d.jpg) +"Turn him into Hulk" + +![](images/d184213e68d0eb12d53708a6ed46c1632bb9d7ffee55bc7e3d003ab23bddd464.jpg) + +![](images/2383e2c6b65793deef4a5cebd936f0f0b1aaeb237f7a35c247ee43349c135486.jpg) +"Turn him into a clown" + +![](images/c4f73d96cc8bd4d4cc9b9a8a6a2897d066242e93be424638cbc3109d787e8f4e.jpg) + +![](images/a8e74e43b88a4ca5fc855f2a74b589bc1ef7ff20e4fd83a6557ae8a58c5073d8.jpg) +Figure 1. Results of GaussianEditor. GaussianEditor offers swift, controllable, and versatile 3D editing. A single editing session only takes 5-10 minutes. Please note our precise editing control, where only the desired parts are modified. Taking the "Make the grass on fire" example from the first row of the figure, other objects in the scene such as the bench and tree remain unaffected. + +![](images/d95b7a9658e0fb1f0851be7d14f34bab34d7ab501329ce41fb7097cfbe1997fc.jpg) + +![](images/c214d612d42c940b93cc8e5d41358a7c0e0d7e084601fa11a1e48020919373ec.jpg) +"Make it autumn" + +![](images/7ec76c98117228ac50a0ade37bf90207fff56e2df91b8a31ae089fa4ebdce523.jpg) + +![](images/111cc0f0d3f08c038633ad405e330215f71c3097dc84e5b5cdccda89ce852873.jpg) + +![](images/92e1a2207ff24cae03a4bfc2f9cf8c8b4337b92aef53f4c770be15d0eb17d716.jpg) + +# Abstract + +3D editing plays a crucial role in many areas such as gaming and virtual reality. Traditional 3D editing methods, which rely on representations like meshes and point clouds, often fall short in realistically depicting complex scenes. On the other hand, methods based on implicit 3D representations, like Neural Radiance Field (NeRF), render complex scenes effectively but suffer from slow processing speeds and limited control over specific scene areas. In response to these challenges, our paper presents GaussianEditor, the first 3D editing algorithm based on Gaussian Splitting (GS), a novel + +3D representation. GaussianEditor enhances precision and control in editing through our proposed Gaussian semantic tracing, which traces the editing target throughout the training process. Additionally, we propose Hierarchical Gaussian splatting (HGS) to achieve stabilized and fine results under stochastic generative guidance from 2D diffusion models. We also develop editing strategies for efficient object removal and integration, a challenging task for existing methods. Our comprehensive experiments demonstrate GaussianEditor's superior control, effective, and efficient performance, marking a significant advancement in 3D editing. + +# 1. Introduction + +In the evolving field of computer vision, the development of user-friendly 3D representations and editing algorithms is a key objective. Such technologies are vital in various applications, ranging from digital gaming to the growing MetaVerse. Traditional 3D representations like meshes and point clouds have been preferred due to their interactive editing capabilities. However, these methods face challenges in accurately rendering complex 3D scenes. + +The recent rise of implicit 3D representations, exemplified by the Neural Radiance Field (NeRF) [28], represents a paradigm shift in 3D scene rendering. NeRF's capacity for high-fidelity rendering, coupled with its implicit nature that offers significant expansibility, marks a substantial improvement over conventional approaches[2, 32, 56]. This dual advantage has placed a significant focus on the NeRF framework in 3D editing [12, 31, 46, 47, 57], establishing it as a foundational approach for a considerable duration. However, NeRF's reliance on high-dimensional multilayer perception (MLP) networks for scene data encoding presents limitations. It restricts direct modification of specific scene parts and complicates tasks, like inpainting and scene composition. This complexity extends to the training and rendering processes, hindering practical applications. + +In light of these challenges, our research is focused on developing an advanced 3D editing algorithm. This algorithm aims for flexible and rapid editing of 3D scenes, integrating both implicit editing, like text-based editing, and explicit control, such as bounding box usage for specific area modifications. To achieve these goals, we choose Gaussian Splatting (GS) [15] for its real-time rendering and explicit point cloud-like representations. + +However, editing Gaussian Splatting (GS) [15] faces distinct challenges. A primary issue is the absence of efficient methods to accurately identify target Gaussians, which is crucial for precise controllable editing. Moreover, it has been observed in [7, 45, 53] that optimizing Gaussian Splatting (GS) using highly random generative guidance like Score Distillation Sampling [36] poses significant challenges. One possible explanation is that, unlike implicit representations buffered by neural networks, GS is directly affected by the randomness in loss. Such direct exposure results in unstable updates, as the properties of Gaussians are directly changed during training. Besides, each training step of GS may involve updates to a vast number of Gaussian points. This process occurs without the moderating influence of neural network-style buffering mechanisms. As a result, the excessive fluidity of the 3D GS scene hinders its ability to converge to finely detailed results like implicit representations when trained with generative guidance. + +To counter these issues, in this work, we propose GaussianEditor , a novel, swift, and highly controllable 3D editing algorithm for Gaussian Splatting. GaussianEditor can fulfill + +various high-quality editing needs within minutes. A key feature of our method is the introduction of Gaussian semantic tracing, which enables precise control over Gaussian Splitting (GS). Gaussian semantic tracing consistently identifies the Gaussians requiring editing at every moment during training. This contrasts with traditional 3D editing methods that often depend on static 2D or 3D masks. Such masks become less effective as the geometries and appearances of 3D models evolve during training. Gaussian semantic tracing is achieved by unprojecting 2D segmentation masks into 3D Gaussians and assigning each Gaussian a semantic tag. As the Gaussians evolve during training, these semantic tags enable the tracking of the specific Gaussians targeted for editing. Our Gaussian tracing algorithm ensures that only the targeted areas are modified, enabling precise and controllable editing. + +Additionally, to tackle the significant challenge of Gaussian Splatting (GS) struggling to fit fine results under highly random generative guidance, we propose a novel GS representation: hierarchical Gaussian splatting (HGS). In HGS, Gaussians are organized into generations based on their sequence in multiple densification processes during training. Gaussians formed in earlier densification stages are deemed older generations and are subject to stricter constraints, aimed at preserving their original state and thus reducing their mobility. Conversely, those formed in later stages are considered younger generations and are subjected to fewer or no constraints, allowing for more adaptability. HGS's design effectively moderates the fluidity of GS by imposing restrictions on older generations while preserving the flexibility of newer generations. This approach enables continuous optimization towards better outcomes, thereby simulating the buffering function achieved in implicit representations through neural networks. Our experiments also demonstrate that HGS is more adept at adapting to highly random generative guidance. + +Finally, we have specifically designed a 3D inpainting algorithm for Gaussian Splatting (GS). As demonstrated in Fig. 1, we have successfully removed specific objects from scenes and seamlessly integrated new objects into designated areas. For object removal, we developed a specialized local repair algorithm that efficiently eliminates artifacts at the intersection of the object and the scene. For adding objects, we first request users to provide a prompt and a 2D inpainting mask for a particular view of the GS. Subsequently, we employ a 2D inpainting method to generate a single-view image of the object to be added. This image is then transformed into a coarse 3D mesh using image-to-3D conversion techniques. The 3D mesh is subsequently converted into the HGS representation and refined. Finally, this refined representation is concatenated into the original GS. The entire inpainting process described above is completed within 5 minutes. + +GaussianEditor offers swift, controllable, and versatile 3D editing. A single editing session typically only takes 5-10 minutes, significantly faster than previous editing processes. Our contributions can be summarized in four aspects: + +1. We have introduced Gaussian semantic tracing, enabling more detailed and effective editing control. +2. We propose Hierarchical Gaussian Splitting (HGS), a novel GS representation capable of converging more stably to refined results under highly random generative guidance. +3. We have specifically designed a 3D inpainting algorithm for Gaussian Splatting, which allows swift removal and addition objects. +4. Extensive experiments demonstrate that our method surpasses previous 3D editing methods in terms of effectiveness, speed, and controllability. + +# 2. Related Works + +# 2.1. 3D Representations + +Various 3D representations have been proposed to address diverse 3D tasks. The groundbreaking work, Neural Radiance Fields (NeRF) [28], employs volumetric rendering and has gained popularity for enabling 3D optimization with only 2D supervision. However, optimizing NeRF can be time-consuming, despite its wide usage in 3D reconstruction [3, 6, 13, 21, 44] and generation [22, 35] tasks. + +While efforts have been made to accelerate NeRF training [29, 40], these approaches primarily focus on the reconstruction setting, leaving the generation setting less optimized. The common technique of spatial pruning does not effectively speed up the generation setting. + +Recently, 3D Gaussian splatting [15] has emerged as an alternative 3D representation to NeRF, showcasing impressive quality and speed in 3D and 4D reconstruction tasks [15, 26, 48, 51, 52]. It has also attracted considerable research interest in the field of generation [7, 45, 53]. Its efficient differentiable rendering implementation and model design facilitate fast training without the need for spatial pruning. + +In this work, we pioneer the adaptation of 3D Gaussian splating to 3D editing tasks, aiming to achieve swift and controllable 3D editing, harnessing the advantages of this representation for the first time in this context. + +# 2.2.3D Editing + +Editing neural fields is inherently challenging due to the intricate interplay between their shape and appearance. EditNeRF [24] stands as a pioneering work in this domain, as they edit both the shape and color of neural fields by conditioning them on latent codes. Additionally, some works [1, 10, 46, 47] leverage CLIP models to facilitate editing through the use of text prompts or reference images. + +Another line of research focuses on predefined template models or skeletons to support actions like re-positioning or re-rendering within specific categories [30, 33]. Geometry-based methods [20, 49, 50, 55] translate neural fields into meshes and synchronize mesh deformation with implicit fields. Additionally, 3D editing techniques involve combining 2D image manipulation, such as inpainting, with neural fields training [19, 23]. + +Concurrent works [31, 57] leverage static 2D and 3D masks to constrain the edit area of NeRF. However, these approaches have their limitations because the training of 3D models is a dynamic process, and static masks cannot effectively constrain it. In contrast, our research employs Gaussian semantic tracing to track the target Gaussian throughout the entire training process. + +# 3. Preliminary + +# 3.1. 3D Gaussian Splitting + +GS (Gaussian Splitting) [15] represents an explicit 3D scene using point clouds, where Gaussians are employed to depict the scene's structure. In this representation, every Gaussian is defined by a center point, denoted as $x$ , and a covariance matrix $\Sigma$ . The center point $x$ is commonly known as the Gaussian's mean value: + +$$ +G (x) = e ^ {- \frac {1}{2} x ^ {T} \Sigma^ {- 1} x}. \tag {1} +$$ + +The covariance matrix $\Sigma$ can be decomposed into a rotation matrix $\mathbf{R}$ and a scaling matrix $\mathbf{S}$ for differentiable optimization: + +$$ +\Sigma = \mathbf {R} \mathbf {S} \mathbf {S} ^ {T} \mathbf {R} ^ {T}, \tag {2} +$$ + +the calculation of gradient flow is detailed in [15]. + +For rendering new viewpoints, the method of splatting, as described in [54], is utilized for positioning the Gaussians on the camera planes. This technique, originally presented in [58], involves a viewing transformation denoted by $W$ and the Jacobian $J$ of the affine approximation of the projective transformation. Using these, the covariance matrix $\Sigma^{\prime}$ in camera coordinates is determined as follows: + +$$ +\Sigma^ {\prime} = J W \Sigma W ^ {T} J ^ {T}. \tag {3} +$$ + +To summarize, each Gaussian point in the model is characterized by a set of attributes: its position, denoted as $x \in \mathbb{R}^3$ , its color represented by spherical harmonics coefficients $c \in \mathbb{R}^k$ (where $k$ indicates the degrees of freedom), its opacity $\alpha \in \mathbb{R}$ , a rotation quaternion $q \in \mathbb{R}^4$ , and a scaling factor $s \in \mathbb{R}^3$ . Particularly, for every pixel, the color and opacity of all Gaussians are calculated based on the Gaussian's representation as described in Eq. 1. The blending process of $N$ ordered points overlapping a pixel follows a specific + +![](images/2c145577e90ac4526669128608404b14a2e757be251d48765c68c0149cf376a8.jpg) +Before Editing + +![](images/d4119158fb0ed9e6ca8f8f7a57ba4eab02650eb22d129120a17d5e657b99f26c.jpg) +Figure 2. Illustration of Gaussian semantic tracing. Prompt: Turn him into an old lady. The red mask in the images represents the projection of the Gaussians that will be updated and densified. The dynamic change of the masked area during the training process, as driven by the updating of Gaussians, ensures consistent effectiveness throughout the training duration. Despite starting with potentially inaccurate segmentation masks due to 2D segmentation errors, Gaussian semantic tracing still guarantees high-quality editing results. + +![](images/86224287b4b87fd52842f1201a2029a0b510b24b084391d661c70659ba9a41ce.jpg) +During Editing + +formula: + +$$ +C = \sum_ {i \in N} c _ {i} \alpha_ {i} \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {j}). \tag {4} +$$ + +where $c_{i}$ and $\alpha_{i}$ signify the color and density of a given point respectively. These values are determined by a Gaussian with a covariance matrix $\Sigma$ , which is then scaled by estimizable per-point opacity and spherical harmonics (SH) color coefficients. + +# 3.2. Diffusion-based Editing Guidance + +Recent advancements have seen numerous works elevating 2D diffusion processes to 3D, applying these processes extensively in the realm of 3D editing. Broadly, these works can be categorized into two types. The first type [8, 27, 31, 36, 42, 57], exemplified by Dreamfusion's [36] introduction of SDS loss, involves feeding the noised rendering of the current 3D model, along with other conditions, into a 2D diffusion model [39]. The scores generated by the diffusion model then guide the direction of model updates. The second type [5, 12, 37, 43] focuses on conducting 2D editing based on given prompts for the multiview rendering of a 3D model. This approach creates a multi-view 2D image dataset, which is then utilized as a training target to provide guidance for the 3D model. + +Our work centers on leveraging the exemplary properties of Gaussian Splatting's explicit representation to enhance 3D editing. Consequently, we do not design specific editing guidance mechanisms but instead directly employ the guidance methods mentioned above. Both types of guidance can be applied in our method. For simplicity, we denote the guidance universally as $D$ . Given the parameters of a 3D model, $\Theta$ , along with the rendered camera pose $p$ and prompt $e$ , the editing loss from the 2D diffusion prior can be formulated as follows: + +$$ +\mathcal {L} _ {\text {E d i t}} = D (\Theta ; p, e) \tag {5} +$$ + +# 4. Method + +We define the task of 3D editing on Gaussian Splatting (GS) as follows: Given a prompt $y$ and a 3D scene represented by 3D Gaussians, denoted by $\Theta$ , where each $\Theta_{i} = \{x_{i}, s_{i}, q_{i}, \alpha_{i}, c_{i}\}$ represents the parameters of the $i$ -th Gaussian as detailed in Sec. 3.1, the objective is to achieve an edited 3D Gaussians, referred to as $\Theta_{y}$ , that aligns with or adheres to the specifications of the prompt $y$ . + +We then introduce our novel framework for performing editing tasks on GS. We first introduce Gaussian semantic tracing in Sec. 4.1, along with a new representation method known as Hierarchical Gaussian Splatting (HGS) in Sec. 4.2. The GS semantic tracing enables precise segmentation and tracing within GS, facilitating controllable editing operations. Compared to the standard GS, the HGS representation demonstrates greater robustness against randomness in generative guidance and is more adept at accommodating a diverse range of editing scenarios. Additionally, we have specifically designed 3D inpainting for GS, which encompasses object removal and addition (Sec. 4.3). + +# 4.1. Gaussian Semantic Tracing + +Previous works [31, 57] in 3D editing usually utilize static 2D or 3D masks to apply loss only within the masked pixels, thus constraining the editing process to only edit the desired area. However, this method has limitations. As 3D representations dynamically change during training, static segmentation masks would become inaccurate or even ineffective. Furthermore, the use of static masks to control gradients in NeRF editing poses a significant limitation, as it confines the editing strictly within the masked area. This restriction prevents the edited content from naturally extending beyond the mask, thus 'locking' the content within a specified spatial boundary. + +To address the aforementioned issue, we have chosen Gaussian Splitting (GS) as our 3D representation due to its explicit nature. This allows us to directly assign semantic labels to each Gaussian point, thereby facilitating semantic tracing in 3D scenes. + +Specifically, we enhance the 3D Gaussians $\Theta$ by adding a new attribute $m$ , where $m_{ij}$ represents the semantic Gaussian mask for the $i$ -th Gaussian point and the $j$ -th semantic label. With this attribute, we can precisely control the editing process by selectively updating only the target 3D Gaussians. During the densification process, newly densified points inherit the semantic label of their parent point. This ensures that we have an accurate 3D semantic mask at every moment throughout the training process. + +As illustrated in Fig. 2, Gaussian semantic tracing enables continuous tracking of each Gaussian's categories during training, adjusting to their evolving properties and numbers. This feature is vital, as it permits selective application of gradients, densification and pruning of Gaussians linked to the + +![](images/923b04fef09ed931a05144eeb6b48a5d66a42fc1fe063c5f0f498fd5b2f6567a.jpg) + +![](images/062709a121e4bcf93b2b49fbba502684237bea4820849f11dbb177daa08e97d3.jpg) +A Garfield cat on the bench + +![](images/6d3665f71103db50b794991f419d05985516ac85334e4e420e8f60b6d54e5c41.jpg) +Original GS +Novel view 1 +Figure 3. 3D inpainting for object incorporation. GaussianEditor is capable of adding objects at specified locations in a scene, given a 2D inpainting mask and a text prompt from a single view. The whole process takes merely five minutes. + +![](images/4764f866214cdd17259cb45499430e816685a5d70198c852579fb834905f3fb6.jpg) +Novel view 2 + +![](images/c2a5a62afb95861758c4a3d2d16d260e5669863306bcbdbe923d3ec5aff5b3ad.jpg) +Original View +Figure 4. 3D inpainting for object removal. Typically, removing the target object based on a Gaussian semantic mask generates artifacts at the interface between the target object and the scene. To address this, we generate a repaired image using a 2D inpainting method and employ Mean Squared Error (MSE) loss for supervision. The whole process takes merely two minutes. + +![](images/63539f4d440ad98e394d16f5078a6f2c08547fb24de3e2585946e7b9b466f72d.jpg) +Removed + +![](images/506b53d4a3d74a0f6b367692a87d459924debdf1eff497015d50a7c31513f1f1.jpg) +Inpainted + +specified category. Additionally, it facilitates training solely by rendering the target object, significantly speeding up the process in complex scenes. The semantic Gaussian mask $m$ functions as a dynamic 3D segmentation mask, evolving with the training, allowing content to expand freely in space. This contrasts with NeRF, where content is restricted to a fixed spatial area. + +Next, we discuss Gaussian Splating unprojection, the method we propose to obtain semantic Gaussian mask $m$ . For a set of 3D Gaussians $\Theta$ , we render them from multiple viewpoints to generate a series of renderings $\mathcal{I}$ . These renderings are then processed using 2D segmentation techniques [18] to obtain 2D segmentation masks $\mathcal{M}$ , with each $\mathcal{M}_j$ , representing the $j$ -th semantic labels. + +To obtain the semantic label for each Gaussian, we unproject the posed 2D semantic label back to the Gaussians + +with inverse rendering. Concretely, we maintain a weight and a counter for each Gaussian. For pixel $p$ on the semantic maps, we unproject the semantic label back to the Gaussians that affects it by + +$$ +w _ {i} ^ {j} = \sum o _ {i} (\boldsymbol {p}) * T _ {i} ^ {j} (\boldsymbol {p}) * \mathcal {M} ^ {j} (\boldsymbol {p}), \tag {6} +$$ + +where $w_{i}^{j}$ represents the weight of the $i$ -th Gaussian for the $j$ -th semantic label, while $o_{i}(\pmb{p})$ , $T_{i}^{j}(\pmb{p})$ , and $\mathcal{M}^j (\pmb{p})$ denote the opacity, transmittance from pixel $\pmb{p}$ , and semantic mask of pixel $\pmb{p}$ for the $i$ -th Gaussian, respectively. After updating all the Gaussian weights and counters, we determine whether a Gaussian belongs to the $j$ -th semantic class based on whether its average weight exceeds a manually set threshold. + +The entire labeling process is remarkably fast, typically taking less than a second. Once this semantic label assignment is completed, the entire Gaussian scene becomes parsed by us, making a variety of operations possible. Notably, 2D diffusion guidance often struggles to effectively edit small objects in complex scenes. Thanks to Gaussian semantic tracing, we can now render these small objects independently and input them into the 2D diffusion model, thereby achieving more precise supervision. + +# 4.2. Hierarchical Gaussian Splitting + +The effectiveness of vanilla GS [17] in reconstruction tasks lies in the high-quality initialization provided by point clouds derived from SFM [41], coupled with stable supervision from ground truth datasets. + +However, the scenario changes in the field of generation. In previous work involving GS in text-to-3D and image-to-3D [7, 45, 53], GS has shown limitations when facing the randomness of generative guidance due to its nature as a point cloud-like representation. This instability in GS is mainly due to their direct exposure to the randomness of loss functions, unlike implicit representations. GS models, which update a large number of Gaussian points each training step, lack the memorization and moderating ability of neural networks. This leads to erratic updates and prevents GS from achieving the detailed results as GS's excessive fluidity hampers its convergence in generative training. + +To address these challenges, we introduce Hierarchical Gaussian Splitting (HGS), a structured representation of GS that is more suitable for generative and editing scenarios. HGS categorizes GS into different generations based on the densification round in which a particular Gaussian point is produced. The initial Gaussians $\Theta$ , are all assigned a generation of 0. During the training process for editing, points generated in the $k$ -th densification round are marked as generation $k$ . + +Subsequently, we impose varying constraints on Gaussians from different generations to control their degree of + +![](images/3ebabeaf66e0af185ce7b5cad3c6bfe929f0724965f0895d0ce4e6abccb73c11.jpg) + +![](images/904ef5a71e062b30d166abdee9c344f46dee956893cd7f79a9f1aa6d4abeb11d.jpg) +Original view + +![](images/e5431b158ca0c78d12e3cd5cdfd6ba72dfe7d0ed703ceb8cfe3471ec45adb669.jpg) + +![](images/e08735f3a506f77a852a134fb40a8b0aff814788bd092f94774788da1a1e55e7.jpg) + +![](images/ad97e0ff4e94d396e2c8a8ae09059fb0865b1b6798ff91162c554b64952a1609.jpg) +"Turn him into a clown" +Figure 5. Qualitative comparison. It's important to note the level of control we maintain over the editing area (the whole body of the man). Background and other non-target regions are essentially unaffected, in contrast to Instruct-Nerf2Nerf [12] where the entire scene undergoes changes. GaussianEditor-DDS and GaussianEditor-iN2N indicate that we utilize delta denoising score [14] and Instruct-Nerf2Nerf [12] respectively, as guidance for editing. + +![](images/4da21a43eafbefc7a60a8342837cbd66788954ca0ea8f873080cd01f6793bcd9.jpg) + +![](images/6229dbee583e751dcecbd87e0f7766ea3be48d5ac4b60109301d4e21c201c7ec.jpg) + +![](images/2994751f0235f9bfa8e155a4b98f6960fc0a14400c95b094e271b16d13c1adcb.jpg) +"Turn him into Albert Einstein" + +![](images/20e87eb054ea179d88612b0c8bccc0b2b0dfd7678b5cb0aabf304734d6e8744f.jpg) +Instruct-N2N + +![](images/0a5fe71e53754e53af6e354f566a2cbefce62b938d84d6d1f152a89e3356f814.jpg) +GSEeditor-DDS + +![](images/ab3efa8079fd0cd213a6342973ef38407967576490cac8045ecd0f2081f4bd4e.jpg) +"Turn him into a Modigliani painting." +GSEoperator-iN2N + +flexibility. The older the generation, the stronger the constraints applied. Anchor loss is utilized to enforce these constraints. At the beginning of training, HGS records the attributes of all Gaussians as anchors. These anchors are then updated to reflect the current state of the Gaussians at each densification process. During training, MSE loss between the anchor state and the current state is employed to ensure that the Gaussians do not deviate too far from their respective anchors: + +$$ +\mathcal {L} _ {\text {a n c h o r}} ^ {P} = \sum_ {i = 0} ^ {n} \lambda_ {i} \left(P _ {i} - \hat {P} _ {i}\right) ^ {2} \tag {7} +$$ + +where $n$ represents the total number of Gaussians and $P$ denotes a certain property of the current Gaussian, including elements from the set $x, s, q, \alpha, c$ . Here, $\hat{P}$ refers to the same property recorded in the anchor state. The term $\lambda_{i}$ indicates the strength of the anchor loss applied to the $i$ -th Gaussian, which varies based on its generation. The overall training loss is defined as: + +$$ +\mathcal {L} = \mathcal {L} _ {E d i t} + \sum_ {P \in \{x, s, q, \alpha , c \}} \lambda_ {P} \mathcal {L} _ {a n c h o r} ^ {P} \tag {8} +$$ + +In this equation, $\lambda_{P}$ signifies the strength of the anchor loss applied to property $P$ , and $\mathcal{L}_{Edit}$ is the edit loss defined in Sec. 3.2. + +This generational design in HGS prevents the issue of excessive flexibility in GS when faced with stochastic loss. With each densification, the anchor loss weight $\lambda_{i}$ for all previous generations of Gaussians is increased. As a result, the fluidity of the existing generations gradually decreases until it nearly solidifies. This approach ensures stable geometry formation under stochastic losses, relying on the almost + +unconstrained Gaussians from new densifications to carve out details. Furthermore, this method of applying anchor loss can effectively meet various editing needs. For instance, to limit changes in the original GS, one can increase the anchor loss weight for generation 0. Similarly, if there is no desire to alter color or geometry during editing, a stronger anchor loss can be applied to these specific properties. + +Additionally, to address the challenge of manually determining a densification threshold, we regulate the densification process based on a percentage criterion. In this method, during each densification step, we selectively densify only those Gaussians whose 3D position gradients are within the top $k\%$ . This strategy proves to be more manageable and intuitive than directly setting a threshold value in the Hierarchical Gaussian Splatting (HGS) framework. + +# 4.3. 3D Inpainting + +Object Removal. Simply removing Gaussians identified by a mask can lead to artifacts, especially at the interfaces where the object intersects with other Gaussians. To address this, we employ 2D inpainting techniques to provide guidance for filling these areas. However, effective 2D inpainting requires precise masks to offer better guidance. To generate these masks, after deletion, we use the KNN algorithm to identify Gaussians nearest to the ones removed, which are likely at the interface. These are then projected onto various views. We subsequently dilate the mask and fix any holes to accurately represent the interface area, thus creating a refined mask for the boundary zones. The whole object removal procedure typically takes only two minutes. + +Object Incorporation. We define this task as follows: Within the 3D Gaussians $\theta$ , given a camera pose $p$ and the + +![](images/9599e95f325b4aef1c133d57fb1078f364f81f6f3842b900c43845bf2016623b.jpg) + +![](images/2c769853edacba0890b0e4c18d8821f36c2c1659794645a0b862b334b3c21099.jpg) +"Make it snowy" + +![](images/c5e470ff4bdb73e8b4696f570da95ad40f25e00aa32c865e2b164f78ba00e1b1.jpg) + +![](images/ebb9c7e1701d8bc5025bf005659c02391b3663970ee1092fc12fe7c5c7ee59d1.jpg) + +![](images/65f3fa13daa9c9311c3f699342e5ca879d8fef3fcb8a74f5378966b3bbe5b147.jpg) + +![](images/c9bb3b3505323a0a8196e6b4f85ab7f411be00d561e91a7c2e9df1fdcfbe2e24.jpg) +"Turn him into an old lady" + +![](images/0e26f8c1349b33571e7a255613d9b3918f26733e21df07532efc56d13c7191f6.jpg) + +![](images/2cd1bb935d0d135a14b1a790e7e0da193362d77f605458b276e07b7d53463048.jpg) + +![](images/fd03a26a7a83e1fe6d04f48da2a46c11c77a3802382dc58c7f7336b2cfb6d928.jpg) +Figure 6. Extensive Results of GaussianEditor. Our method is capable of various editing tasks, including face and scene editing. In face and bear editing, we restrict the editing area to the face using Gaussian semantic tracing, ensuring that undesired areas remain unchanged. The leftmost column demonstrates the original view, while the right three columns show the images after editing. + +![](images/3556919a9a88b24b62f61a436e1f03f5a0346c156db177fec6227da8aefe487a.jpg) +"Turn the bear into a grizzly bear" + +![](images/232cb283a868f1a6540735f1f505f2b11ad7b58d77bc0526265c02b9dbe2bb40.jpg) + +![](images/d87325b0cfb59cb8265af45d88ef40dd70eced37096277c7dfc4d18004cfbc6b.jpg) + +corresponding rendering $I$ from this viewpoint, the user provides a 2D mask $M$ on $I$ indicating the area they wish to inpaint. Additionally, a prompt $y$ is provided to specify the content of the inpainting. We then update $\theta$ to fulfill the inpainting request. + +Given $I$ , $M$ , and $y$ , the process begins with generating a 2D inpainted image $I^M y$ , utilizing a 2D inpainting diffusion model as per [34]. Subsequently, the foreground object from $I^M y$ , created by [34], is segmented and input into the image-to-3D method referenced in [25] to generate a coarse 3D mesh. This coarse mesh is then transformed into 3D Gaussians $\theta_y$ , and refined with HGS detailed in Sec. 4.2. + +For aligning the coordinate system of $\theta_y$ with $\theta$ , the depth of $I_y^M$ is first estimated using the technique from [38]. This depth is then aligned with the depth map rendered by $\theta$ at camera pose $p$ , using the least squares method. With this alignment, we can accurately determine the coordinates and scale of the inpainted foreground object in the coordinate system of $\theta$ . After transforming $\theta_y$ into the coordinate system of $\theta$ , we simply concatenate them to produce the final inpainted 3D Gaussians. + +It is important to note that due to our efficient design, the entire object incorporation procedure can be completed in approximately 5 minutes. + +# 5. Experiments + +# 5.1. Implementation Details + +We utilize the highly optimized renderer implementation from [16] for Gaussian rendering and base our implementation on Threestudio [11]. All the original 3D Gaussians used in this work are trained using the methods described in [16]. Our experiments are conducted on a single RTX A6000 GPU. As detailed in Sec. 4.1, once we obtain segmentation masks from the 2D segmentation method outlined in [18], segmenting the 3D Gaussians takes only about 1 second. + +For editing large scenes, the camera poses employed during the editing process are selected from a subset of the multi-view image dataset initially used for reconstruction. Depending on the complexity of the scene, the number of camera poses used in our experiments varies from 24 to 96. The editing process, influenced by the specified prompt and the complexity of the scene, typically involves optimizing for 500-1000 steps, taking about 5-10 minutes in total. + +Regarding 3D inpainting for object incorporation, as detailed in Sec. 4.3, it takes approximately 3 minutes to generate a 3D mesh using the method from [25] and an additional 2 minutes to transfer this mesh into 3D Gaussians and refine it, while the composition process of two Gaussians takes less than 1 second. + +![](images/45031bcd2073026aca19d4de3ba66516ad3104370caf41dcbaea304f05f5fcb8.jpg) +Original View + +![](images/b79313259f059997c6a3e93f4268768e763521f2e67d6fc0ec4ee0b5f92213a5.jpg) +GS + +![](images/b3991c369740ac2e2ff04101002d89855f37350aa35ca8bbee066c047d5f0cee.jpg) +HGS +Figure 7. Ablation study on Hierarchical Gaussian Splitting (HGS). Prompt: make the grass on fire. Even when specifying the editing area with prompts, generative methods like InstructPix2Pix [4] tend to edit the entire 2D image. Without HGS, Gaussians tend to conform to this whole-image editing by spreading and densifying across the entire scene, leading to uncontrollable densification and blurring of the image. With HGS, however, this kind of diffusion is effectively restrained. + +# 5.2. Qualitative Comparisons + +As illustrated in Fig. 5, GaussianEditor-iN2N surpasses other methods in both the quality of edits and controllability. Instruct-Nerf2Nerf, while producing edits with insufficient detail, cannot also control the editing area. GaussianEditorDDS, due to the more challenging control of guidance offered by DDS loss [14] compared to Instruct-pix2pix [4], tends to result in oversaturated colors and less precise editing outcomes. + +Additionally, our method exhibits exceptional control over the editing area. This is achieved through Gaussian semantic tracing, which identifies the Gaussians that require editing at each training step, for example, the entire human body in Fig. 5. It's important to note that in Instruct-Nerf2Nerf [12], the use of static 2D or 3D masks restricts the spatial freedom of the edits, as the permissible area for the edited subject is limited by these masks. Furthermore, the effectiveness of static masks diminishes as the geometries and appearances of 3D models evolve during training. + +In Fig. 6, we demonstrate that GaussianEditor can accommodate a variety of scenarios, such as editing in large-scale scenes and facial swaps. In the case of large scenes, we did not apply Gaussian semantic tracing. However, for facial swaps, we traced the Gaussians corresponding to facial regions, achieving controllable and realistic editing. + +# 5.3.Quantitative Comparisons + +As shown in Table 1, we conduct quantitative comparisons on user study and CLIP directional similarity (as shown in InstructPix2Pix [4] and StyleGAN-Nada [9]). GaussianEditor-iN2N not only demonstrates superior outcomes in user studies but also excels in CLIP Directional Similarity. Besides, Instruct-Nerf2Nerf [12] typically requires more than 30 minutes to complete the editing of a scene, whereas our method + +
iN2N [12]Ours (DDS)Ours (iN2N)
User study15.45%12.27%72.28%
CLIP Directional Similarity0.16000.18130.2071
+ +Table 1. Quantitative Comparison. GaussianEditor-iN2N outperforms in both user study evaluations and CLIP Directional Similarity [9] metrics. + +only takes between 5 to 10 minutes. + +# 5.4. Ablation Study + +As demonstrated in Fig. 7, we conducted ablation experiments on Hierarchical Gaussian Splitting(HGS). Without HGS, Gaussians tend to spread and densify across the scene, leading to uncontrolled densification and image blurring. This is typically caused by the tendency of methods like Instruct-Pix2Pix to edit the entire 2D image when prompts are used to define editing areas. However, HGS effectively circumvents this issue by constraining the mobility of the Gaussian in the old generation, ensuring that the overall scene does not exhibit excessive mobility. + +# 6. Conclusion + +In our research, we introduce GaussianEditor, an innovative 3D editing algorithm based on Gaussian Splatting, designed for enhanced control and efficiency. Our method employs Gaussian semantic tracing for precise identification and targeting of editing areas, followed by Hierarchical Gaussian Splatting (HGS) to balance fluidity and stability in achieving detailed results under stochastic guidance. Additionally, we developed a specialized 3D inpainting algorithm for Gaussian Splatting, streamlining object removal and integration, and greatly reducing editing time. + +Limitation. Similar to previous 3D editing works based on 2D diffusion models, GaussianEditor relies on these models to provide effective supervision. However, current 2D diffusion models struggle to offer effective guidance for certain complex prompts, leading to limitations in 3D editing. + +# 7. Acknowledgements + +This study is supported under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). This research is also supported by the MoE AcRF Tier 2 grant (MOE-T2EP20220-0007). + +# References + +[1] Chong Bao, Yinda Zhang, and Bangbang et al. Yang. Sine: Semantic-driven image-based nef editing with prior-guided editing field. In CVPR 2023, pages 20919-20929, 2023. 3 +[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. + +Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855-5864, 2021. 2 +[3] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. CVPR, 2022. 3 +[4] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. arXiv preprint arXiv:2211.09800, 2022. 8 +[5] Yiwen Chen, Chi Zhang, Xiaofeng Yang, Zhongang Cai, Gang Yu, Lei Yang, and Guosheng Lin. It3d: Improved text-to-3d generation with explicit view synthesis. arXiv preprint arXiv:2308.11473, 2023. 4 +[6] Zhiqin Chen, Thomas Funkhouser, Peter Hedman, and Andrea Tagliasacchi. Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. arXiv preprint arXiv:2208.00277, 2022. 3 +[7] Zilong Chen, Feng Wang, and Huaping Liu. Text-to-3d using gaussian splatting. arXiv preprint arXiv:2309.16585, 2023. 2, 3, 5 +[8] Xinhua Cheng, Tianyu Yang, Jianan Wang, Yu Li, Lei Zhang, Jian Zhang, and Li Yuan. Progressive3d: Progressively local editing for text-to-3d content creation with complex semantic prompts. arXiv preprint arXiv:2310.11784, 2023. 4 +[9] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clipguided domain adaptation of image generators. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 8 +[10] William Gao, Noam Aigerman, Thibault Groueix, Vladimir G Kim, and Rana Hanocka. Textdeformer: Geometry manipulation using text guidance. arXiv preprint arXiv:2304.13348, 2023. 3 +[11] Yuan-Chen Guo, Ying-Tian Liu, Chen Wang, Zi-Xin Zou, Guan Luo, Chia-Hao Chen, Yan-Pei Cao, and Song-Hai Zhang. threestudio: A unified framework for 3d content generation. https://github.com/threestudio-project/threestudio, 2023.7 +[12] Ayaan Haque, Matthew Tancik, Alexei A Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. arXiv preprint arXiv:2303.12789, 2023. 2, 4, 6, 8 +[13] Peter Hedman, Pratul P. Srinivasan, Ben Mildenhall, Jonathan T. Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. ICCV, 2021. 3 +[14] Amir Hertz, Kfir Aberman, and Daniel Cohen-Or. Delta denoising score. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2328-2337, 2023. 6, 8 +[15] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 42(4): 1-14, 2023. 2, 3 +[16] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ToG, 42(4):1-14, 2023. 7 +[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance + +field rendering. ACM Transactions on Graphics, 42(4), 2023. 5 +[18] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 5, 7 +[19] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nerf for editing via feature field distillation. arXiv preprint arXiv:2205.15585, 2022. 3 +[20] Yuan Li, Zhi-Hao Lin, David Forsyth, Jia-Bin Huang, and Shenlong Wang. Climatenerf: Physically-based neural rendering for extreme climate synthesis. arXiv e-prints, pages arXiv-2211, 2022. 3 +[21] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In CVPR, 2023. 3 +[22] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, pages 300–309, 2023. 3 +[23] Hao-Kang Liu, I Shen, Bing-Yu Chen, et al. Nerf-in: Free-form nerf inpainting with rgb-d priors. arXiv preprint arXiv:2206.04901, 2022. 3 +[24] Steven Liu, Xiuming Zhang, Zhoutong Zhang, Richard Zhang, Jun-Yan Zhu, and Bryan Russell. Editing conditional radiance fields. In ICCV 2021, pages 5773-5783, 2021. 3 +[25] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. arXiv preprint arXiv:2310.15008, 2023. 7 +[26] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 3 +[27] Aryan Mikaeili, Or Perel, Mehdi Safaee, Daniel Cohen-Or, and Ali Mahdavi-Amiri. Sked: Sketch-guided text-based 3d editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14607-14619, 2023. 4 +[28] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2, 3 +[29] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 3 +[30] Atsuhiro Noguchi, Xiao Sun, Stephen Lin, and Tatsuya Harada. Neural articulated radiance field. In ICCV 2021, pages 5762-5772, 2021. 3 +[31] Jangho Park, Gihyun Kwon, and Jong Chul Ye. Ed-nerf: Efficient text-guided editing of 3d scene using latent space nerf. arXiv preprint arXiv:2310.02712, 2023. 2, 3, 4 +[32] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo Martin- + +Brualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. 2 +[33] Sida Peng, Yuanqing Zhang, Yinghao Xu, and et al. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In CVPR 2021, pages 9054–9063, 2021. 3 +[34] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 7 +[35] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 3 +[36] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 2, 4 +[37] Amit Raj, Srinivas Kaza, Ben Poole, Michael Niemeyer, Nataniel Ruiz, Ben Mildenhall, Shiran Zada, Kfir Aberman, Michael Rubinstein, Jonathan Barron, et al. Dreambooth3d: Subject-driven text-to-3d generation. arXiv preprint arXiv:2303.13508, 2023. 4 +[38] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12179-12188, 2021. 7 +[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 4 +[40] Sara Fridovich-Keil and Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 3 +[41] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 5 +[42] Etai Sella, Gal Fiebelman, Peter Hedman, and Hadar Averbuch-Elor. Vox-e: Text-guided voxel editing of 3d objects. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 430-440, 2023. 4 +[43] Ruizhi Shao, Jingxiang Sun, Cheng Peng, Zerong Zheng, Boyao Zhou, Hongwen Zhang, and Yebin Liu. Control4d: Dynamic portrait editing by learning 4d gan from 2d diffusion-based editor. arXiv preprint arXiv:2305.20082, 2023. 4 +[44] Chaoyue Song, Tianyi Chen, Yiwen Chen, Jiacheng Wei, Chuan Sheng Foo, Fayao Liu, and Guosheng Lin. Moda: Modeling deformable 3d objects from casual videos. arXiv preprint arXiv:2304.08279, 2023. 3 +[45] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. 2, 3, 5 +[46] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation + +of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3835-3844, 2022. 2, 3 +[47] Can Wang, Ruixiang Jiang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Nerf-art: Text-driven neural radiance fields stylization. IEEE Transactions on Visualization and Computer Graphics, 2023. 2, 3 +[48] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528, 2023. 3 +[49] Tianhan Xu and Tatsuya Harada. Deforming radiance fields with cages. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIII, pages 159-175. Springer, 2022. 3 +[50] Bangbang Yang, Chong Bao, and Junyi et al. Zeng. Neumesh: Learning disentangled neural mesh-based implicit field for geometry and texture editing. In ECCV 2022, pages 597-614. Springer, 2022. 3 +[51] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101, 2023. 3 +[52] Zeyu Yang, Hongye Yang, Zijie Pan, Xiatian Zhu, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. arXiv preprint arXiv:2310.10642, 2023. 3 +[53] Taoran Yi, Jiemin Fang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussian-dreamer: Fast generation from text to 3d gaussian splatting with point cloud priors. arXiv preprint arXiv:2310.08529, 2023. 2, 3, 5 +[54] Wang Yifan, Felice Serena, Shihao Wu, Cengiz Öztireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM Transactions on Graphics (TOG), 38(6):1-14, 2019. 3 +[55] Yu-Jie Yuan, Yang-Tian Sun, Yu-Kun Lai, and et al. Nerf-editing: geometry editing of neural radiance fields. In CVPR 2022, pages 18353-18364, 2022. 3 +[56] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 2 +[57] Jingyu Zhuang, Chen Wang, Lingjie Liu, Liang Lin, and Guanbin Li. Dreameditor: Text-driven 3d scene editing with neural fields. arXiv preprint arXiv:2306.13455, 2023. 2, 3, 4 +[58] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Surface splatting. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 371-378, 2001. 3 \ No newline at end of file diff --git a/2024/GaussianEditor_ Swift and Controllable 3D Editing with Gaussian Splatting/images.zip b/2024/GaussianEditor_ Swift and Controllable 3D Editing with Gaussian Splatting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..42b80a0896cb52b3e470fba4cc2b9d87d482f34d --- /dev/null +++ b/2024/GaussianEditor_ Swift and Controllable 3D Editing with Gaussian Splatting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1f5dc2dfd4b23a18f966dc0ff933a9f020a4e4e16835b22906612deaa4668a15 +size 717578 diff --git a/2024/GaussianEditor_ Swift and Controllable 3D Editing with Gaussian Splatting/layout.json b/2024/GaussianEditor_ Swift and Controllable 3D Editing with Gaussian Splatting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0aa9cb664025dde4abc11f485204fd2de5be304e --- /dev/null +++ b/2024/GaussianEditor_ Swift and Controllable 3D Editing with Gaussian Splatting/layout.json @@ -0,0 +1,11880 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 62, + 103, + 532, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 103, + 532, + 121 + ], + "spans": [ + { + "bbox": [ + 62, + 103, + 532, + 121 + ], + "type": "text", + "content": "GaussianEditor: Swift and Controllable 3D Editing with Gaussian Splitting" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 141, + 170, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 141, + 170, + 171 + ], + "spans": [ + { + "bbox": [ + 89, + 141, + 170, + 171 + ], + "type": "text", + "content": "Yiwen Chen\\*1,2 Yikai Wang3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 172, + 143, + 256, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 143, + 256, + 171 + ], + "spans": [ + { + "bbox": [ + 172, + 143, + 256, + 171 + ], + "type": "text", + "content": "Zilong Chen\\*3 Zhongang Cai4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 265, + 143, + 408, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 265, + 143, + 408, + 171 + ], + "spans": [ + { + "bbox": [ + 265, + 143, + 408, + 171 + ], + "type": "text", + "content": "Chi Zhang" + }, + { + "bbox": [ + 265, + 143, + 408, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 265, + 143, + 408, + 171 + ], + "type": "text", + "content": " Feng Wang" + }, + { + "bbox": [ + 265, + 143, + 408, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 265, + 143, + 408, + 171 + ], + "type": "text", + "content": " Lei Yang" + }, + { + "bbox": [ + 265, + 143, + 408, + 171 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 265, + 143, + 408, + 171 + ], + "type": "text", + "content": " Huaping Liu" + }, + { + "bbox": [ + 265, + 143, + 408, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 416, + 143, + 509, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 416, + 143, + 509, + 171 + ], + "spans": [ + { + "bbox": [ + 416, + 143, + 509, + 171 + ], + "type": "text", + "content": "Xiaofeng Yang" + }, + { + "bbox": [ + 416, + 143, + 509, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 416, + 143, + 509, + 171 + ], + "type": "text", + "content": " Guosheng Lin" + }, + { + "bbox": [ + 416, + 143, + 509, + 171 + ], + "type": "inline_equation", + "content": "^{**1,2}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 192, + 177, + 401, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 177, + 401, + 191 + ], + "spans": [ + { + "bbox": [ + 192, + 177, + 401, + 191 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 192, + 177, + 401, + 191 + ], + "type": "text", + "content": "S-Lab, Nanyang Technological University" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 98, + 191, + 495, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 191, + 495, + 205 + ], + "spans": [ + { + "bbox": [ + 98, + 191, + 495, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 98, + 191, + 495, + 205 + ], + "type": "text", + "content": "School of Computer Science and Engineering, Nanyang Technological University" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 123, + 205, + 471, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 205, + 471, + 220 + ], + "spans": [ + { + "bbox": [ + 123, + 205, + 471, + 220 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 123, + 205, + 471, + 220 + ], + "type": "text", + "content": "Department of Computer Science and Technology, Tsinghua University" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 244, + 220, + 349, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 220, + 349, + 232 + ], + "spans": [ + { + "bbox": [ + 244, + 220, + 349, + 232 + ], + "type": "text", + "content": "4 SenseTime Research" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 171, + 235, + 418, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 235, + 418, + 246 + ], + "spans": [ + { + "bbox": [ + 171, + 235, + 418, + 246 + ], + "type": "text", + "content": "https://buaacyw.github.io/gaussian-editor" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 51, + 259, + 130, + 313 + ], + "blocks": [ + { + "bbox": [ + 51, + 259, + 130, + 313 + ], + "lines": [ + { + "bbox": [ + 51, + 259, + 130, + 313 + ], + "spans": [ + { + "bbox": [ + 51, + 259, + 130, + 313 + ], + "type": "image", + "image_path": "b1755dd51b2004f2668c564909ff8df1419c4f5ec82753cd6286fdac4e74143c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 107, + 316, + 148, + 324 + ], + "lines": [ + { + "bbox": [ + 107, + 316, + 148, + 324 + ], + "spans": [ + { + "bbox": [ + 107, + 316, + 148, + 324 + ], + "type": "text", + "content": "Original view" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 132, + 259, + 211, + 313 + ], + "blocks": [ + { + "bbox": [ + 132, + 259, + 211, + 313 + ], + "lines": [ + { + "bbox": [ + 132, + 259, + 211, + 313 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 211, + 313 + ], + "type": "image", + "image_path": "1b317d9448d69eddea2abdff283e407cd469970daf35dd808042852c4e1a8385.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 50, + 327, + 130, + 388 + ], + "blocks": [ + { + "bbox": [ + 50, + 327, + 130, + 388 + ], + "lines": [ + { + "bbox": [ + 50, + 327, + 130, + 388 + ], + "spans": [ + { + "bbox": [ + 50, + 327, + 130, + 388 + ], + "type": "image", + "image_path": "87f5389d17dc044a89bf047023a9ee8265a427318ac2923cf56450f7914413c6.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 107, + 389, + 149, + 397 + ], + "lines": [ + { + "bbox": [ + 107, + 389, + 149, + 397 + ], + "spans": [ + { + "bbox": [ + 107, + 389, + 149, + 397 + ], + "type": "text", + "content": "Original view" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 131, + 327, + 210, + 388 + ], + "blocks": [ + { + "bbox": [ + 131, + 327, + 210, + 388 + ], + "lines": [ + { + "bbox": [ + 131, + 327, + 210, + 388 + ], + "spans": [ + { + "bbox": [ + 131, + 327, + 210, + 388 + ], + "type": "image", + "image_path": "4d01b8eb2a4c0777f8d5bf2daab7928aba67db5ecc08b91fe83e9b49a67c7ad6.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 216, + 259, + 296, + 313 + ], + "blocks": [ + { + "bbox": [ + 216, + 259, + 296, + 313 + ], + "lines": [ + { + "bbox": [ + 216, + 259, + 296, + 313 + ], + "spans": [ + { + "bbox": [ + 216, + 259, + 296, + 313 + ], + "type": "image", + "image_path": "fd0245785b004ba71a77391a241d951b1ce1e46f9d3e08894430bb300626f5a4.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 229, + 316, + 354, + 324 + ], + "lines": [ + { + "bbox": [ + 229, + 316, + 354, + 324 + ], + "spans": [ + { + "bbox": [ + 229, + 316, + 354, + 324 + ], + "type": "text", + "content": "\"Delete the bike. Add a dog on the bench.\"" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 297, + 259, + 378, + 314 + ], + "blocks": [ + { + "bbox": [ + 297, + 259, + 378, + 314 + ], + "lines": [ + { + "bbox": [ + 297, + 259, + 378, + 314 + ], + "spans": [ + { + "bbox": [ + 297, + 259, + 378, + 314 + ], + "type": "image", + "image_path": "0f9f38f043f6dd091e6966a5d2cc8d1f9658026ec88df06a0c39d6b3352af481.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 381, + 259, + 462, + 313 + ], + "blocks": [ + { + "bbox": [ + 381, + 259, + 462, + 313 + ], + "lines": [ + { + "bbox": [ + 381, + 259, + 462, + 313 + ], + "spans": [ + { + "bbox": [ + 381, + 259, + 462, + 313 + ], + "type": "image", + "image_path": "c980c3a12f5374acd59b22c36dc54b6cab85718f61652d0c78e4ee60cd0c0829.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 424, + 316, + 498, + 324 + ], + "lines": [ + { + "bbox": [ + 424, + 316, + 498, + 324 + ], + "spans": [ + { + "bbox": [ + 424, + 316, + 498, + 324 + ], + "type": "text", + "content": "\"Make the grass on fire\"" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 463, + 259, + 544, + 314 + ], + "blocks": [ + { + "bbox": [ + 463, + 259, + 544, + 314 + ], + "lines": [ + { + "bbox": [ + 463, + 259, + 544, + 314 + ], + "spans": [ + { + "bbox": [ + 463, + 259, + 544, + 314 + ], + "type": "image", + "image_path": "a1275c47cdc9ddf269efd347734e12ff216b4c517ebe38ac7b5f27711707913a.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 215, + 327, + 296, + 388 + ], + "blocks": [ + { + "bbox": [ + 215, + 327, + 296, + 388 + ], + "lines": [ + { + "bbox": [ + 215, + 327, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 215, + 327, + 296, + 388 + ], + "type": "image", + "image_path": "c947c3270de861600e875163c95f6d05763b917f5a067cfa60bea57c2bbd282d.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 266, + 388, + 332, + 396 + ], + "lines": [ + { + "bbox": [ + 266, + 388, + 332, + 396 + ], + "spans": [ + { + "bbox": [ + 266, + 388, + 332, + 396 + ], + "type": "text", + "content": "\"Turn him into Hulk\"" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 296, + 327, + 376, + 388 + ], + "blocks": [ + { + "bbox": [ + 296, + 327, + 376, + 388 + ], + "lines": [ + { + "bbox": [ + 296, + 327, + 376, + 388 + ], + "spans": [ + { + "bbox": [ + 296, + 327, + 376, + 388 + ], + "type": "image", + "image_path": "d184213e68d0eb12d53708a6ed46c1632bb9d7ffee55bc7e3d003ab23bddd464.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 381, + 327, + 462, + 388 + ], + "blocks": [ + { + "bbox": [ + 381, + 327, + 462, + 388 + ], + "lines": [ + { + "bbox": [ + 381, + 327, + 462, + 388 + ], + "spans": [ + { + "bbox": [ + 381, + 327, + 462, + 388 + ], + "type": "image", + "image_path": "2383e2c6b65793deef4a5cebd936f0f0b1aaeb237f7a35c247ee43349c135486.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 429, + 388, + 502, + 396 + ], + "lines": [ + { + "bbox": [ + 429, + 388, + 502, + 396 + ], + "spans": [ + { + "bbox": [ + 429, + 388, + 502, + 396 + ], + "type": "text", + "content": "\"Turn him into a clown\"" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 463, + 327, + 543, + 388 + ], + "blocks": [ + { + "bbox": [ + 463, + 327, + 543, + 388 + ], + "lines": [ + { + "bbox": [ + 463, + 327, + 543, + 388 + ], + "spans": [ + { + "bbox": [ + 463, + 327, + 543, + 388 + ], + "type": "image", + "image_path": "c4f73d96cc8bd4d4cc9b9a8a6a2897d066242e93be424638cbc3109d787e8f4e.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 50, + 405, + 130, + 457 + ], + "blocks": [ + { + "bbox": [ + 50, + 405, + 130, + 457 + ], + "lines": [ + { + "bbox": [ + 50, + 405, + 130, + 457 + ], + "spans": [ + { + "bbox": [ + 50, + 405, + 130, + 457 + ], + "type": "image", + "image_path": "a8e74e43b88a4ca5fc855f2a74b589bc1ef7ff20e4fd83a6557ae8a58c5073d8.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 475, + 547, + 508 + ], + "lines": [ + { + "bbox": [ + 45, + 475, + 547, + 508 + ], + "spans": [ + { + "bbox": [ + 45, + 475, + 547, + 508 + ], + "type": "text", + "content": "Figure 1. Results of GaussianEditor. GaussianEditor offers swift, controllable, and versatile 3D editing. A single editing session only takes 5-10 minutes. Please note our precise editing control, where only the desired parts are modified. Taking the \"Make the grass on fire\" example from the first row of the figure, other objects in the scene such as the bench and tree remain unaffected." + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 131, + 405, + 211, + 457 + ], + "blocks": [ + { + "bbox": [ + 131, + 405, + 211, + 457 + ], + "lines": [ + { + "bbox": [ + 131, + 405, + 211, + 457 + ], + "spans": [ + { + "bbox": [ + 131, + 405, + 211, + 457 + ], + "type": "image", + "image_path": "d95b7a9658e0fb1f0851be7d14f34bab34d7ab501329ce41fb7097cfbe1997fc.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 216, + 405, + 296, + 457 + ], + "blocks": [ + { + "bbox": [ + 216, + 405, + 296, + 457 + ], + "lines": [ + { + "bbox": [ + 216, + 405, + 296, + 457 + ], + "spans": [ + { + "bbox": [ + 216, + 405, + 296, + 457 + ], + "type": "image", + "image_path": "c214d612d42c940b93cc8e5d41358a7c0e0d7e084601fa11a1e48020919373ec.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 270, + 458, + 324, + 466 + ], + "lines": [ + { + "bbox": [ + 270, + 458, + 324, + 466 + ], + "spans": [ + { + "bbox": [ + 270, + 458, + 324, + 466 + ], + "type": "text", + "content": "\"Make it autumn\"" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 297, + 405, + 377, + 457 + ], + "blocks": [ + { + "bbox": [ + 297, + 405, + 377, + 457 + ], + "lines": [ + { + "bbox": [ + 297, + 405, + 377, + 457 + ], + "spans": [ + { + "bbox": [ + 297, + 405, + 377, + 457 + ], + "type": "image", + "image_path": "7ec76c98117228ac50a0ade37bf90207fff56e2df91b8a31ae089fa4ebdce523.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 381, + 405, + 461, + 457 + ], + "blocks": [ + { + "bbox": [ + 381, + 405, + 461, + 457 + ], + "lines": [ + { + "bbox": [ + 381, + 405, + 461, + 457 + ], + "spans": [ + { + "bbox": [ + 381, + 405, + 461, + 457 + ], + "type": "image", + "image_path": "111cc0f0d3f08c038633ad405e330215f71c3097dc84e5b5cdccda89ce852873.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 463, + 405, + 543, + 457 + ], + "blocks": [ + { + "bbox": [ + 463, + 405, + 543, + 457 + ], + "lines": [ + { + "bbox": [ + 463, + 405, + 543, + 457 + ], + "spans": [ + { + "bbox": [ + 463, + 405, + 543, + 457 + ], + "type": "image", + "image_path": "92e1a2207ff24cae03a4bfc2f9cf8c8b4337b92aef53f4c770be15d0eb17d716.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "bbox": [ + 143, + 542, + 191, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 542, + 191, + 555 + ], + "spans": [ + { + "bbox": [ + 143, + 542, + 191, + 555 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 46, + 567, + 288, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 567, + 288, + 686 + ], + "spans": [ + { + "bbox": [ + 46, + 567, + 288, + 686 + ], + "type": "text", + "content": "3D editing plays a crucial role in many areas such as gaming and virtual reality. Traditional 3D editing methods, which rely on representations like meshes and point clouds, often fall short in realistically depicting complex scenes. On the other hand, methods based on implicit 3D representations, like Neural Radiance Field (NeRF), render complex scenes effectively but suffer from slow processing speeds and limited control over specific scene areas. In response to these challenges, our paper presents GaussianEditor, the first 3D editing algorithm based on Gaussian Splitting (GS), a novel" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 304, + 531, + 547, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 531, + 547, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 547, + 663 + ], + "type": "text", + "content": "3D representation. GaussianEditor enhances precision and control in editing through our proposed Gaussian semantic tracing, which traces the editing target throughout the training process. Additionally, we propose Hierarchical Gaussian splatting (HGS) to achieve stabilized and fine results under stochastic generative guidance from 2D diffusion models. We also develop editing strategies for efficient object removal and integration, a challenging task for existing methods. Our comprehensive experiments demonstrate GaussianEditor's superior control, effective, and efficient performance, marking a significant advancement in 3D editing." + } + ] + } + ], + "index": 40 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 693, + 146, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 693, + 146, + 703 + ], + "spans": [ + { + "bbox": [ + 61, + 693, + 146, + 703 + ], + "type": "text", + "content": "** Corresponding author." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 61, + 703, + 241, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 703, + 241, + 712 + ], + "spans": [ + { + "bbox": [ + 61, + 703, + 241, + 712 + ], + "type": "text", + "content": "* The first two authors contributed equally to this work." + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21476" + } + ] + } + ], + "index": 44 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 49, + 72, + 127, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 72, + 127, + 83 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 127, + 83 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 91, + 286, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 91, + 286, + 186 + ], + "spans": [ + { + "bbox": [ + 49, + 91, + 286, + 186 + ], + "type": "text", + "content": "In the evolving field of computer vision, the development of user-friendly 3D representations and editing algorithms is a key objective. Such technologies are vital in various applications, ranging from digital gaming to the growing MetaVerse. Traditional 3D representations like meshes and point clouds have been preferred due to their interactive editing capabilities. However, these methods face challenges in accurately rendering complex 3D scenes." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 187, + 286, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 187, + 286, + 365 + ], + "spans": [ + { + "bbox": [ + 49, + 187, + 286, + 365 + ], + "type": "text", + "content": "The recent rise of implicit 3D representations, exemplified by the Neural Radiance Field (NeRF) [28], represents a paradigm shift in 3D scene rendering. NeRF's capacity for high-fidelity rendering, coupled with its implicit nature that offers significant expansibility, marks a substantial improvement over conventional approaches[2, 32, 56]. This dual advantage has placed a significant focus on the NeRF framework in 3D editing [12, 31, 46, 47, 57], establishing it as a foundational approach for a considerable duration. However, NeRF's reliance on high-dimensional multilayer perception (MLP) networks for scene data encoding presents limitations. It restricts direct modification of specific scene parts and complicates tasks, like inpainting and scene composition. This complexity extends to the training and rendering processes, hindering practical applications." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 366, + 286, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 366, + 286, + 461 + ], + "spans": [ + { + "bbox": [ + 49, + 366, + 286, + 461 + ], + "type": "text", + "content": "In light of these challenges, our research is focused on developing an advanced 3D editing algorithm. This algorithm aims for flexible and rapid editing of 3D scenes, integrating both implicit editing, like text-based editing, and explicit control, such as bounding box usage for specific area modifications. To achieve these goals, we choose Gaussian Splatting (GS) [15] for its real-time rendering and explicit point cloud-like representations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 462, + 286, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 462, + 286, + 677 + ], + "spans": [ + { + "bbox": [ + 49, + 462, + 286, + 677 + ], + "type": "text", + "content": "However, editing Gaussian Splatting (GS) [15] faces distinct challenges. A primary issue is the absence of efficient methods to accurately identify target Gaussians, which is crucial for precise controllable editing. Moreover, it has been observed in [7, 45, 53] that optimizing Gaussian Splatting (GS) using highly random generative guidance like Score Distillation Sampling [36] poses significant challenges. One possible explanation is that, unlike implicit representations buffered by neural networks, GS is directly affected by the randomness in loss. Such direct exposure results in unstable updates, as the properties of Gaussians are directly changed during training. Besides, each training step of GS may involve updates to a vast number of Gaussian points. This process occurs without the moderating influence of neural network-style buffering mechanisms. As a result, the excessive fluidity of the 3D GS scene hinders its ability to converge to finely detailed results like implicit representations when trained with generative guidance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 677, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 677, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 677, + 286, + 712 + ], + "type": "text", + "content": "To counter these issues, in this work, we propose GaussianEditor , a novel, swift, and highly controllable 3D editing algorithm for Gaussian Splatting. GaussianEditor can fulfill" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 308, + 72, + 545, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 72, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 308, + 72, + 545, + 263 + ], + "type": "text", + "content": "various high-quality editing needs within minutes. A key feature of our method is the introduction of Gaussian semantic tracing, which enables precise control over Gaussian Splitting (GS). Gaussian semantic tracing consistently identifies the Gaussians requiring editing at every moment during training. This contrasts with traditional 3D editing methods that often depend on static 2D or 3D masks. Such masks become less effective as the geometries and appearances of 3D models evolve during training. Gaussian semantic tracing is achieved by unprojecting 2D segmentation masks into 3D Gaussians and assigning each Gaussian a semantic tag. As the Gaussians evolve during training, these semantic tags enable the tracking of the specific Gaussians targeted for editing. Our Gaussian tracing algorithm ensures that only the targeted areas are modified, enabling precise and controllable editing." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 308, + 266, + 545, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 266, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 308, + 266, + 545, + 506 + ], + "type": "text", + "content": "Additionally, to tackle the significant challenge of Gaussian Splatting (GS) struggling to fit fine results under highly random generative guidance, we propose a novel GS representation: hierarchical Gaussian splatting (HGS). In HGS, Gaussians are organized into generations based on their sequence in multiple densification processes during training. Gaussians formed in earlier densification stages are deemed older generations and are subject to stricter constraints, aimed at preserving their original state and thus reducing their mobility. Conversely, those formed in later stages are considered younger generations and are subjected to fewer or no constraints, allowing for more adaptability. HGS's design effectively moderates the fluidity of GS by imposing restrictions on older generations while preserving the flexibility of newer generations. This approach enables continuous optimization towards better outcomes, thereby simulating the buffering function achieved in implicit representations through neural networks. Our experiments also demonstrate that HGS is more adept at adapting to highly random generative guidance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 308, + 510, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 510, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 510, + 545, + 712 + ], + "type": "text", + "content": "Finally, we have specifically designed a 3D inpainting algorithm for Gaussian Splatting (GS). As demonstrated in Fig. 1, we have successfully removed specific objects from scenes and seamlessly integrated new objects into designated areas. For object removal, we developed a specialized local repair algorithm that efficiently eliminates artifacts at the intersection of the object and the scene. For adding objects, we first request users to provide a prompt and a 2D inpainting mask for a particular view of the GS. Subsequently, we employ a 2D inpainting method to generate a single-view image of the object to be added. This image is then transformed into a coarse 3D mesh using image-to-3D conversion techniques. The 3D mesh is subsequently converted into the HGS representation and refined. Finally, this refined representation is concatenated into the original GS. The entire inpainting process described above is completed within 5 minutes." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 750, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 750, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 750, + 317, + 757 + ], + "type": "text", + "content": "21477" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 119 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 119 + ], + "type": "text", + "content": "GaussianEditor offers swift, controllable, and versatile 3D editing. A single editing session typically only takes 5-10 minutes, significantly faster than previous editing processes. Our contributions can be summarized in four aspects:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 120, + 287, + 263 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 47, + 120, + 287, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 120, + 287, + 143 + ], + "spans": [ + { + "bbox": [ + 47, + 120, + 287, + 143 + ], + "type": "text", + "content": "1. We have introduced Gaussian semantic tracing, enabling more detailed and effective editing control." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 144, + 287, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 144, + 287, + 190 + ], + "spans": [ + { + "bbox": [ + 47, + 144, + 287, + 190 + ], + "type": "text", + "content": "2. We propose Hierarchical Gaussian Splitting (HGS), a novel GS representation capable of converging more stably to refined results under highly random generative guidance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 192, + 287, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 192, + 287, + 227 + ], + "spans": [ + { + "bbox": [ + 47, + 192, + 287, + 227 + ], + "type": "text", + "content": "3. We have specifically designed a 3D inpainting algorithm for Gaussian Splatting, which allows swift removal and addition objects." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 228, + 287, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 228, + 287, + 263 + ], + "spans": [ + { + "bbox": [ + 47, + 228, + 287, + 263 + ], + "type": "text", + "content": "4. Extensive experiments demonstrate that our method surpasses previous 3D editing methods in terms of effectiveness, speed, and controllability." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 275, + 138, + 287 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 275, + 138, + 287 + ], + "spans": [ + { + "bbox": [ + 47, + 275, + 138, + 287 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 295, + 161, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 295, + 161, + 308 + ], + "spans": [ + { + "bbox": [ + 47, + 295, + 161, + 308 + ], + "type": "text", + "content": "2.1. 3D Representations" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 314, + 287, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 314, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 46, + 314, + 287, + 397 + ], + "type": "text", + "content": "Various 3D representations have been proposed to address diverse 3D tasks. The groundbreaking work, Neural Radiance Fields (NeRF) [28], employs volumetric rendering and has gained popularity for enabling 3D optimization with only 2D supervision. However, optimizing NeRF can be time-consuming, despite its wide usage in 3D reconstruction [3, 6, 13, 21, 44] and generation [22, 35] tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 399, + 287, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 287, + 458 + ], + "type": "text", + "content": "While efforts have been made to accelerate NeRF training [29, 40], these approaches primarily focus on the reconstruction setting, leaving the generation setting less optimized. The common technique of spatial pruning does not effectively speed up the generation setting." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 459, + 287, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 459, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 46, + 459, + 287, + 553 + ], + "type": "text", + "content": "Recently, 3D Gaussian splatting [15] has emerged as an alternative 3D representation to NeRF, showcasing impressive quality and speed in 3D and 4D reconstruction tasks [15, 26, 48, 51, 52]. It has also attracted considerable research interest in the field of generation [7, 45, 53]. Its efficient differentiable rendering implementation and model design facilitate fast training without the need for spatial pruning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 555, + 287, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 555, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 555, + 287, + 601 + ], + "type": "text", + "content": "In this work, we pioneer the adaptation of 3D Gaussian splating to 3D editing tasks, aiming to achieve swift and controllable 3D editing, harnessing the advantages of this representation for the first time in this context." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 610, + 120, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 610, + 120, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 610, + 120, + 623 + ], + "type": "text", + "content": "2.2.3D Editing" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": "Editing neural fields is inherently challenging due to the intricate interplay between their shape and appearance. EditNeRF [24] stands as a pioneering work in this domain, as they edit both the shape and color of neural fields by conditioning them on latent codes. Additionally, some works [1, 10, 46, 47] leverage CLIP models to facilitate editing through the use of text prompts or reference images." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 72, + 547, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 167 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 167 + ], + "type": "text", + "content": "Another line of research focuses on predefined template models or skeletons to support actions like re-positioning or re-rendering within specific categories [30, 33]. Geometry-based methods [20, 49, 50, 55] translate neural fields into meshes and synchronize mesh deformation with implicit fields. Additionally, 3D editing techniques involve combining 2D image manipulation, such as inpainting, with neural fields training [19, 23]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 168, + 547, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 168, + 547, + 253 + ], + "spans": [ + { + "bbox": [ + 304, + 168, + 547, + 253 + ], + "type": "text", + "content": "Concurrent works [31, 57] leverage static 2D and 3D masks to constrain the edit area of NeRF. However, these approaches have their limitations because the training of 3D models is a dynamic process, and static masks cannot effectively constrain it. In contrast, our research employs Gaussian semantic tracing to track the target Gaussian throughout the entire training process." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 266, + 382, + 279 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 266, + 382, + 279 + ], + "spans": [ + { + "bbox": [ + 306, + 266, + 382, + 279 + ], + "type": "text", + "content": "3. Preliminary" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 286, + 433, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 286, + 433, + 300 + ], + "spans": [ + { + "bbox": [ + 306, + 286, + 433, + 300 + ], + "type": "text", + "content": "3.1. 3D Gaussian Splitting" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 305, + 545, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 305, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 304, + 305, + 545, + 376 + ], + "type": "text", + "content": "GS (Gaussian Splitting) [15] represents an explicit 3D scene using point clouds, where Gaussians are employed to depict the scene's structure. In this representation, every Gaussian is defined by a center point, denoted as " + }, + { + "bbox": [ + 304, + 305, + 545, + 376 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 305, + 545, + 376 + ], + "type": "text", + "content": ", and a covariance matrix " + }, + { + "bbox": [ + 304, + 305, + 545, + 376 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 304, + 305, + 545, + 376 + ], + "type": "text", + "content": ". The center point " + }, + { + "bbox": [ + 304, + 305, + 545, + 376 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 305, + 545, + 376 + ], + "type": "text", + "content": " is commonly known as the Gaussian's mean value:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 381, + 389, + 545, + 405 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 389, + 545, + 405 + ], + "spans": [ + { + "bbox": [ + 381, + 389, + 545, + 405 + ], + "type": "interline_equation", + "content": "G (x) = e ^ {- \\frac {1}{2} x ^ {T} \\Sigma^ {- 1} x}. \\tag {1}", + "image_path": "23ac699c25bbd8c2c912c23efe5bab559229276b3de6c70f82d404625f95cf5a.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 414, + 547, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 414, + 547, + 448 + ], + "spans": [ + { + "bbox": [ + 304, + 414, + 547, + 448 + ], + "type": "text", + "content": "The covariance matrix " + }, + { + "bbox": [ + 304, + 414, + 547, + 448 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 304, + 414, + 547, + 448 + ], + "type": "text", + "content": " can be decomposed into a rotation matrix " + }, + { + "bbox": [ + 304, + 414, + 547, + 448 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 304, + 414, + 547, + 448 + ], + "type": "text", + "content": " and a scaling matrix " + }, + { + "bbox": [ + 304, + 414, + 547, + 448 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 304, + 414, + 547, + 448 + ], + "type": "text", + "content": " for differentiable optimization:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 391, + 451, + 545, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 451, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 391, + 451, + 545, + 464 + ], + "type": "interline_equation", + "content": "\\Sigma = \\mathbf {R} \\mathbf {S} \\mathbf {S} ^ {T} \\mathbf {R} ^ {T}, \\tag {2}", + "image_path": "c9b3a5befbe1975da76789b17349daa089b4332315e08ea66c12982a1deebd13.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 473, + 508, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 473, + 508, + 485 + ], + "spans": [ + { + "bbox": [ + 306, + 473, + 508, + 485 + ], + "type": "text", + "content": "the calculation of gradient flow is detailed in [15]." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 486, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 568 + ], + "type": "text", + "content": "For rendering new viewpoints, the method of splatting, as described in [54], is utilized for positioning the Gaussians on the camera planes. This technique, originally presented in [58], involves a viewing transformation denoted by " + }, + { + "bbox": [ + 304, + 486, + 545, + 568 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 486, + 545, + 568 + ], + "type": "text", + "content": " and the Jacobian " + }, + { + "bbox": [ + 304, + 486, + 545, + 568 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 304, + 486, + 545, + 568 + ], + "type": "text", + "content": " of the affine approximation of the projective transformation. Using these, the covariance matrix " + }, + { + "bbox": [ + 304, + 486, + 545, + 568 + ], + "type": "inline_equation", + "content": "\\Sigma^{\\prime}" + }, + { + "bbox": [ + 304, + 486, + 545, + 568 + ], + "type": "text", + "content": " in camera coordinates is determined as follows:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 383, + 583, + 545, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 583, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 383, + 583, + 545, + 597 + ], + "type": "interline_equation", + "content": "\\Sigma^ {\\prime} = J W \\Sigma W ^ {T} J ^ {T}. \\tag {3}", + "image_path": "99d3235585c7dc839816010c092aeed1f52c0303615fd9dd0ac3e93e2f2d28db.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": "To summarize, each Gaussian point in the model is characterized by a set of attributes: its position, denoted as " + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": ", its color represented by spherical harmonics coefficients " + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "inline_equation", + "content": "c \\in \\mathbb{R}^k" + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": " (where " + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": " indicates the degrees of freedom), its opacity " + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\alpha \\in \\mathbb{R}" + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": ", a rotation quaternion " + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "inline_equation", + "content": "q \\in \\mathbb{R}^4" + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": ", and a scaling factor " + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "inline_equation", + "content": "s \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": ". Particularly, for every pixel, the color and opacity of all Gaussians are calculated based on the Gaussian's representation as described in Eq. 1. The blending process of " + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": " ordered points overlapping a pixel follows a specific" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21478" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 62, + 167, + 121 + ], + "blocks": [ + { + "bbox": [ + 49, + 62, + 167, + 121 + ], + "lines": [ + { + "bbox": [ + 49, + 62, + 167, + 121 + ], + "spans": [ + { + "bbox": [ + 49, + 62, + 167, + 121 + ], + "type": "image", + "image_path": "2c145577e90ac4526669128608404b14a2e757be251d48765c68c0149cf376a8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 123, + 132, + 133 + ], + "lines": [ + { + "bbox": [ + 83, + 123, + 132, + 133 + ], + "spans": [ + { + "bbox": [ + 83, + 123, + 132, + 133 + ], + "type": "text", + "content": "Before Editing" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 168, + 62, + 221, + 121 + ], + "blocks": [ + { + "bbox": [ + 168, + 62, + 221, + 121 + ], + "lines": [ + { + "bbox": [ + 168, + 62, + 221, + 121 + ], + "spans": [ + { + "bbox": [ + 168, + 62, + 221, + 121 + ], + "type": "image", + "image_path": "d4119158fb0ed9e6ca8f8f7a57ba4eab02650eb22d129120a17d5e657b99f26c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 138, + 287, + 237 + ], + "lines": [ + { + "bbox": [ + 46, + 138, + 287, + 237 + ], + "spans": [ + { + "bbox": [ + 46, + 138, + 287, + 237 + ], + "type": "text", + "content": "Figure 2. Illustration of Gaussian semantic tracing. Prompt: Turn him into an old lady. The red mask in the images represents the projection of the Gaussians that will be updated and densified. The dynamic change of the masked area during the training process, as driven by the updating of Gaussians, ensures consistent effectiveness throughout the training duration. Despite starting with potentially inaccurate segmentation masks due to 2D segmentation errors, Gaussian semantic tracing still guarantees high-quality editing results." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 225, + 62, + 287, + 121 + ], + "blocks": [ + { + "bbox": [ + 225, + 62, + 287, + 121 + ], + "lines": [ + { + "bbox": [ + 225, + 62, + 287, + 121 + ], + "spans": [ + { + "bbox": [ + 225, + 62, + 287, + 121 + ], + "type": "image", + "image_path": "86224287b4b87fd52842f1201a2029a0b510b24b084391d661c70659ba9a41ce.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 202, + 123, + 252, + 133 + ], + "lines": [ + { + "bbox": [ + 202, + 123, + 252, + 133 + ], + "spans": [ + { + "bbox": [ + 202, + 123, + 252, + 133 + ], + "type": "text", + "content": "During Editing" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 253, + 84, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 253, + 84, + 262 + ], + "spans": [ + { + "bbox": [ + 47, + 253, + 84, + 262 + ], + "type": "text", + "content": "formula:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 110, + 265, + 287, + 299 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 265, + 287, + 299 + ], + "spans": [ + { + "bbox": [ + 110, + 265, + 287, + 299 + ], + "type": "interline_equation", + "content": "C = \\sum_ {i \\in N} c _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}). \\tag {4}", + "image_path": "1548c0aad28ed690f83ed077967449d3c265f2d75f5c9106170d52847593209b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 309, + 288, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 309, + 288, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 309, + 288, + 369 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 309, + 288, + 369 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 46, + 309, + 288, + 369 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 309, + 288, + 369 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 46, + 309, + 288, + 369 + ], + "type": "text", + "content": " signify the color and density of a given point respectively. These values are determined by a Gaussian with a covariance matrix " + }, + { + "bbox": [ + 46, + 309, + 288, + 369 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 46, + 309, + 288, + 369 + ], + "type": "text", + "content": ", which is then scaled by estimizable per-point opacity and spherical harmonics (SH) color coefficients." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 380, + 229, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 380, + 229, + 392 + ], + "spans": [ + { + "bbox": [ + 47, + 380, + 229, + 392 + ], + "type": "text", + "content": "3.2. Diffusion-based Editing Guidance" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 400, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 400, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 287, + 567 + ], + "type": "text", + "content": "Recent advancements have seen numerous works elevating 2D diffusion processes to 3D, applying these processes extensively in the realm of 3D editing. Broadly, these works can be categorized into two types. The first type [8, 27, 31, 36, 42, 57], exemplified by Dreamfusion's [36] introduction of SDS loss, involves feeding the noised rendering of the current 3D model, along with other conditions, into a 2D diffusion model [39]. The scores generated by the diffusion model then guide the direction of model updates. The second type [5, 12, 37, 43] focuses on conducting 2D editing based on given prompts for the multiview rendering of a 3D model. This approach creates a multi-view 2D image dataset, which is then utilized as a training target to provide guidance for the 3D model." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 570, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 687 + ], + "type": "text", + "content": "Our work centers on leveraging the exemplary properties of Gaussian Splatting's explicit representation to enhance 3D editing. Consequently, we do not design specific editing guidance mechanisms but instead directly employ the guidance methods mentioned above. Both types of guidance can be applied in our method. For simplicity, we denote the guidance universally as " + }, + { + "bbox": [ + 46, + 570, + 287, + 687 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 570, + 287, + 687 + ], + "type": "text", + "content": ". Given the parameters of a 3D model, " + }, + { + "bbox": [ + 46, + 570, + 287, + 687 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 46, + 570, + 287, + 687 + ], + "type": "text", + "content": ", along with the rendered camera pose " + }, + { + "bbox": [ + 46, + 570, + 287, + 687 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 570, + 287, + 687 + ], + "type": "text", + "content": " and prompt " + }, + { + "bbox": [ + 46, + 570, + 287, + 687 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 46, + 570, + 287, + 687 + ], + "type": "text", + "content": ", the editing loss from the 2D diffusion prior can be formulated as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 127, + 701, + 287, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 127, + 701, + 287, + 714 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {E d i t}} = D (\\Theta ; p, e) \\tag {5}", + "image_path": "4f02aaa74641c959d86c7c6417310032e49dfc592145be013fe9cf122dee6fdf.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 71, + 361, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 361, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 361, + 83 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 91, + 547, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 547, + 175 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 547, + 175 + ], + "type": "text", + "content": "We define the task of 3D editing on Gaussian Splatting (GS) as follows: Given a prompt " + }, + { + "bbox": [ + 304, + 91, + 547, + 175 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 91, + 547, + 175 + ], + "type": "text", + "content": " and a 3D scene represented by 3D Gaussians, denoted by " + }, + { + "bbox": [ + 304, + 91, + 547, + 175 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 304, + 91, + 547, + 175 + ], + "type": "text", + "content": ", where each " + }, + { + "bbox": [ + 304, + 91, + 547, + 175 + ], + "type": "inline_equation", + "content": "\\Theta_{i} = \\{x_{i}, s_{i}, q_{i}, \\alpha_{i}, c_{i}\\}" + }, + { + "bbox": [ + 304, + 91, + 547, + 175 + ], + "type": "text", + "content": " represents the parameters of the " + }, + { + "bbox": [ + 304, + 91, + 547, + 175 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 91, + 547, + 175 + ], + "type": "text", + "content": "-th Gaussian as detailed in Sec. 3.1, the objective is to achieve an edited 3D Gaussians, referred to as " + }, + { + "bbox": [ + 304, + 91, + 547, + 175 + ], + "type": "inline_equation", + "content": "\\Theta_{y}" + }, + { + "bbox": [ + 304, + 91, + 547, + 175 + ], + "type": "text", + "content": ", that aligns with or adheres to the specifications of the prompt " + }, + { + "bbox": [ + 304, + 91, + 547, + 175 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 91, + 547, + 175 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 175, + 547, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 175, + 547, + 318 + ], + "spans": [ + { + "bbox": [ + 304, + 175, + 547, + 318 + ], + "type": "text", + "content": "We then introduce our novel framework for performing editing tasks on GS. We first introduce Gaussian semantic tracing in Sec. 4.1, along with a new representation method known as Hierarchical Gaussian Splatting (HGS) in Sec. 4.2. The GS semantic tracing enables precise segmentation and tracing within GS, facilitating controllable editing operations. Compared to the standard GS, the HGS representation demonstrates greater robustness against randomness in generative guidance and is more adept at accommodating a diverse range of editing scenarios. Additionally, we have specifically designed 3D inpainting for GS, which encompasses object removal and addition (Sec. 4.3)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 324, + 457, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 324, + 457, + 337 + ], + "spans": [ + { + "bbox": [ + 305, + 324, + 457, + 337 + ], + "type": "text", + "content": "4.1. Gaussian Semantic Tracing" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 342, + 547, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 547, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 547, + 486 + ], + "type": "text", + "content": "Previous works [31, 57] in 3D editing usually utilize static 2D or 3D masks to apply loss only within the masked pixels, thus constraining the editing process to only edit the desired area. However, this method has limitations. As 3D representations dynamically change during training, static segmentation masks would become inaccurate or even ineffective. Furthermore, the use of static masks to control gradients in NeRF editing poses a significant limitation, as it confines the editing strictly within the masked area. This restriction prevents the edited content from naturally extending beyond the mask, thus 'locking' the content within a specified spatial boundary." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 487, + 545, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 487, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 487, + 545, + 545 + ], + "type": "text", + "content": "To address the aforementioned issue, we have chosen Gaussian Splitting (GS) as our 3D representation due to its explicit nature. This allows us to directly assign semantic labels to each Gaussian point, thereby facilitating semantic tracing in 3D scenes." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 546, + 547, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 547, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 547, + 653 + ], + "type": "text", + "content": "Specifically, we enhance the 3D Gaussians " + }, + { + "bbox": [ + 304, + 546, + 547, + 653 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 304, + 546, + 547, + 653 + ], + "type": "text", + "content": " by adding a new attribute " + }, + { + "bbox": [ + 304, + 546, + 547, + 653 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 546, + 547, + 653 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 546, + 547, + 653 + ], + "type": "inline_equation", + "content": "m_{ij}" + }, + { + "bbox": [ + 304, + 546, + 547, + 653 + ], + "type": "text", + "content": " represents the semantic Gaussian mask for the " + }, + { + "bbox": [ + 304, + 546, + 547, + 653 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 546, + 547, + 653 + ], + "type": "text", + "content": "-th Gaussian point and the " + }, + { + "bbox": [ + 304, + 546, + 547, + 653 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 546, + 547, + 653 + ], + "type": "text", + "content": "-th semantic label. With this attribute, we can precisely control the editing process by selectively updating only the target 3D Gaussians. During the densification process, newly densified points inherit the semantic label of their parent point. This ensures that we have an accurate 3D semantic mask at every moment throughout the training process." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 654, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 712 + ], + "type": "text", + "content": "As illustrated in Fig. 2, Gaussian semantic tracing enables continuous tracking of each Gaussian's categories during training, adjusting to their evolving properties and numbers. This feature is vital, as it permits selective application of gradients, densification and pruning of Gaussians linked to the" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21479" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 164, + 148 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 164, + 148 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 164, + 148 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 164, + 148 + ], + "type": "image", + "image_path": "923b04fef09ed931a05144eeb6b48a5d66a42fc1fe063c5f0f498fd5b2f6567a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 170, + 70, + 286, + 148 + ], + "blocks": [ + { + "bbox": [ + 170, + 70, + 286, + 148 + ], + "lines": [ + { + "bbox": [ + 170, + 70, + 286, + 148 + ], + "spans": [ + { + "bbox": [ + 170, + 70, + 286, + 148 + ], + "type": "image", + "image_path": "062709a121e4bcf93b2b49fbba502684237bea4820849f11dbb177daa08e97d3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 149, + 273, + 158 + ], + "lines": [ + { + "bbox": [ + 184, + 149, + 273, + 158 + ], + "spans": [ + { + "bbox": [ + 184, + 149, + 273, + 158 + ], + "type": "text", + "content": "A Garfield cat on the bench" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 49, + 159, + 164, + 236 + ], + "blocks": [ + { + "bbox": [ + 86, + 149, + 126, + 159 + ], + "lines": [ + { + "bbox": [ + 86, + 149, + 126, + 159 + ], + "spans": [ + { + "bbox": [ + 86, + 149, + 126, + 159 + ], + "type": "text", + "content": "Original GS" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 159, + 164, + 236 + ], + "lines": [ + { + "bbox": [ + 49, + 159, + 164, + 236 + ], + "spans": [ + { + "bbox": [ + 49, + 159, + 164, + 236 + ], + "type": "image", + "image_path": "6d3665f71103db50b794991f419d05985516ac85334e4e420e8f60b6d54e5c41.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 84, + 239, + 127, + 247 + ], + "lines": [ + { + "bbox": [ + 84, + 239, + 127, + 247 + ], + "spans": [ + { + "bbox": [ + 84, + 239, + 127, + 247 + ], + "type": "text", + "content": "Novel view 1" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 249, + 288, + 293 + ], + "lines": [ + { + "bbox": [ + 47, + 249, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 47, + 249, + 288, + 293 + ], + "type": "text", + "content": "Figure 3. 3D inpainting for object incorporation. GaussianEditor is capable of adding objects at specified locations in a scene, given a 2D inpainting mask and a text prompt from a single view. The whole process takes merely five minutes." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 168, + 159, + 283, + 236 + ], + "blocks": [ + { + "bbox": [ + 168, + 159, + 283, + 236 + ], + "lines": [ + { + "bbox": [ + 168, + 159, + 283, + 236 + ], + "spans": [ + { + "bbox": [ + 168, + 159, + 283, + 236 + ], + "type": "image", + "image_path": "4764f866214cdd17259cb45499430e816685a5d70198c852579fb834905f3fb6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 204, + 239, + 249, + 247 + ], + "lines": [ + { + "bbox": [ + 204, + 239, + 249, + 247 + ], + "spans": [ + { + "bbox": [ + 204, + 239, + 249, + 247 + ], + "type": "text", + "content": "Novel view 2" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 48, + 307, + 126, + 412 + ], + "blocks": [ + { + "bbox": [ + 48, + 307, + 126, + 412 + ], + "lines": [ + { + "bbox": [ + 48, + 307, + 126, + 412 + ], + "spans": [ + { + "bbox": [ + 48, + 307, + 126, + 412 + ], + "type": "image", + "image_path": "c2a5a62afb95861758c4a3d2d16d260e5669863306bcbdbe923d3ec5aff5b3ad.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 415, + 111, + 423 + ], + "lines": [ + { + "bbox": [ + 64, + 415, + 111, + 423 + ], + "spans": [ + { + "bbox": [ + 64, + 415, + 111, + 423 + ], + "type": "text", + "content": "Original View" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 429, + 287, + 495 + ], + "lines": [ + { + "bbox": [ + 46, + 429, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 429, + 287, + 495 + ], + "type": "text", + "content": "Figure 4. 3D inpainting for object removal. Typically, removing the target object based on a Gaussian semantic mask generates artifacts at the interface between the target object and the scene. To address this, we generate a repaired image using a 2D inpainting method and employ Mean Squared Error (MSE) loss for supervision. The whole process takes merely two minutes." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 128, + 307, + 206, + 412 + ], + "blocks": [ + { + "bbox": [ + 128, + 307, + 206, + 412 + ], + "lines": [ + { + "bbox": [ + 128, + 307, + 206, + 412 + ], + "spans": [ + { + "bbox": [ + 128, + 307, + 206, + 412 + ], + "type": "image", + "image_path": "63539f4d440ad98e394d16f5078a6f2c08547fb24de3e2585946e7b9b466f72d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 415, + 184, + 423 + ], + "lines": [ + { + "bbox": [ + 151, + 415, + 184, + 423 + ], + "spans": [ + { + "bbox": [ + 151, + 415, + 184, + 423 + ], + "type": "text", + "content": "Removed" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 208, + 307, + 287, + 412 + ], + "blocks": [ + { + "bbox": [ + 208, + 307, + 287, + 412 + ], + "lines": [ + { + "bbox": [ + 208, + 307, + 287, + 412 + ], + "spans": [ + { + "bbox": [ + 208, + 307, + 287, + 412 + ], + "type": "image", + "image_path": "506b53d4a3d74a0f6b367692a87d459924debdf1eff497015d50a7c31513f1f1.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 231, + 415, + 263, + 423 + ], + "lines": [ + { + "bbox": [ + 231, + 415, + 263, + 423 + ], + "spans": [ + { + "bbox": [ + 231, + 415, + 263, + 423 + ], + "type": "text", + "content": "Inpainted" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 519, + 287, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 519, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 46, + 519, + 287, + 602 + ], + "type": "text", + "content": "specified category. Additionally, it facilitates training solely by rendering the target object, significantly speeding up the process in complex scenes. The semantic Gaussian mask " + }, + { + "bbox": [ + 46, + 519, + 287, + 602 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 519, + 287, + 602 + ], + "type": "text", + "content": " functions as a dynamic 3D segmentation mask, evolving with the training, allowing content to expand freely in space. This contrasts with NeRF, where content is restricted to a fixed spatial area." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 604, + 288, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 604, + 288, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 604, + 288, + 688 + ], + "type": "text", + "content": "Next, we discuss Gaussian Splating unprojection, the method we propose to obtain semantic Gaussian mask " + }, + { + "bbox": [ + 46, + 604, + 288, + 688 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 604, + 288, + 688 + ], + "type": "text", + "content": ". For a set of 3D Gaussians " + }, + { + "bbox": [ + 46, + 604, + 288, + 688 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 46, + 604, + 288, + 688 + ], + "type": "text", + "content": ", we render them from multiple viewpoints to generate a series of renderings " + }, + { + "bbox": [ + 46, + 604, + 288, + 688 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 46, + 604, + 288, + 688 + ], + "type": "text", + "content": ". These renderings are then processed using 2D segmentation techniques [18] to obtain 2D segmentation masks " + }, + { + "bbox": [ + 46, + 604, + 288, + 688 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 604, + 288, + 688 + ], + "type": "text", + "content": ", with each " + }, + { + "bbox": [ + 46, + 604, + 288, + 688 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_j" + }, + { + "bbox": [ + 46, + 604, + 288, + 688 + ], + "type": "text", + "content": ", representing the " + }, + { + "bbox": [ + 46, + 604, + 288, + 688 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 604, + 288, + 688 + ], + "type": "text", + "content": "-th semantic labels." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "type": "text", + "content": "To obtain the semantic label for each Gaussian, we unproject the posed 2D semantic label back to the Gaussians" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 72, + 546, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 546, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 546, + 120 + ], + "type": "text", + "content": "with inverse rendering. Concretely, we maintain a weight and a counter for each Gaussian. For pixel " + }, + { + "bbox": [ + 304, + 72, + 546, + 120 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 72, + 546, + 120 + ], + "type": "text", + "content": " on the semantic maps, we unproject the semantic label back to the Gaussians that affects it by" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 354, + 130, + 546, + 148 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 130, + 546, + 148 + ], + "spans": [ + { + "bbox": [ + 354, + 130, + 546, + 148 + ], + "type": "interline_equation", + "content": "w _ {i} ^ {j} = \\sum o _ {i} (\\boldsymbol {p}) * T _ {i} ^ {j} (\\boldsymbol {p}) * \\mathcal {M} ^ {j} (\\boldsymbol {p}), \\tag {6}", + "image_path": "ae1bc276d37fd89cc128eb4bc4acca2ea18ebabcdb54a100e4750b68ff0c2e7a.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 155, + 547, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 155, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 155, + 547, + 251 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 155, + 547, + 251 + ], + "type": "inline_equation", + "content": "w_{i}^{j}" + }, + { + "bbox": [ + 304, + 155, + 547, + 251 + ], + "type": "text", + "content": " represents the weight of the " + }, + { + "bbox": [ + 304, + 155, + 547, + 251 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 155, + 547, + 251 + ], + "type": "text", + "content": "-th Gaussian for the " + }, + { + "bbox": [ + 304, + 155, + 547, + 251 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 155, + 547, + 251 + ], + "type": "text", + "content": "-th semantic label, while " + }, + { + "bbox": [ + 304, + 155, + 547, + 251 + ], + "type": "inline_equation", + "content": "o_{i}(\\pmb{p})" + }, + { + "bbox": [ + 304, + 155, + 547, + 251 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 155, + 547, + 251 + ], + "type": "inline_equation", + "content": "T_{i}^{j}(\\pmb{p})" + }, + { + "bbox": [ + 304, + 155, + 547, + 251 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 155, + 547, + 251 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^j (\\pmb{p})" + }, + { + "bbox": [ + 304, + 155, + 547, + 251 + ], + "type": "text", + "content": " denote the opacity, transmittance from pixel " + }, + { + "bbox": [ + 304, + 155, + 547, + 251 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 304, + 155, + 547, + 251 + ], + "type": "text", + "content": ", and semantic mask of pixel " + }, + { + "bbox": [ + 304, + 155, + 547, + 251 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 304, + 155, + 547, + 251 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 304, + 155, + 547, + 251 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 155, + 547, + 251 + ], + "type": "text", + "content": "-th Gaussian, respectively. After updating all the Gaussian weights and counters, we determine whether a Gaussian belongs to the " + }, + { + "bbox": [ + 304, + 155, + 547, + 251 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 155, + 547, + 251 + ], + "type": "text", + "content": "-th semantic class based on whether its average weight exceeds a manually set threshold." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 252, + 547, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 252, + 547, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 252, + 547, + 361 + ], + "type": "text", + "content": "The entire labeling process is remarkably fast, typically taking less than a second. Once this semantic label assignment is completed, the entire Gaussian scene becomes parsed by us, making a variety of operations possible. Notably, 2D diffusion guidance often struggles to effectively edit small objects in complex scenes. Thanks to Gaussian semantic tracing, we can now render these small objects independently and input them into the 2D diffusion model, thereby achieving more precise supervision." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 369, + 479, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 369, + 479, + 382 + ], + "spans": [ + { + "bbox": [ + 306, + 369, + 479, + 382 + ], + "type": "text", + "content": "4.2. Hierarchical Gaussian Splitting" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 390, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 545, + 437 + ], + "type": "text", + "content": "The effectiveness of vanilla GS [17] in reconstruction tasks lies in the high-quality initialization provided by point clouds derived from SFM [41], coupled with stable supervision from ground truth datasets." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 437, + 547, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 547, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 547, + 581 + ], + "type": "text", + "content": "However, the scenario changes in the field of generation. In previous work involving GS in text-to-3D and image-to-3D [7, 45, 53], GS has shown limitations when facing the randomness of generative guidance due to its nature as a point cloud-like representation. This instability in GS is mainly due to their direct exposure to the randomness of loss functions, unlike implicit representations. GS models, which update a large number of Gaussian points each training step, lack the memorization and moderating ability of neural networks. This leads to erratic updates and prevents GS from achieving the detailed results as GS's excessive fluidity hampers its convergence in generative training." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 582, + 547, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 689 + ], + "type": "text", + "content": "To address these challenges, we introduce Hierarchical Gaussian Splitting (HGS), a structured representation of GS that is more suitable for generative and editing scenarios. HGS categorizes GS into different generations based on the densification round in which a particular Gaussian point is produced. The initial Gaussians " + }, + { + "bbox": [ + 304, + 582, + 547, + 689 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 304, + 582, + 547, + 689 + ], + "type": "text", + "content": ", are all assigned a generation of 0. During the training process for editing, points generated in the " + }, + { + "bbox": [ + 304, + 582, + 547, + 689 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 582, + 547, + 689 + ], + "type": "text", + "content": "-th densification round are marked as generation " + }, + { + "bbox": [ + 304, + 582, + 547, + 689 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 582, + 547, + 689 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "content": "Subsequently, we impose varying constraints on Gaussians from different generations to control their degree of" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21480" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 78, + 145, + 150 + ], + "blocks": [ + { + "bbox": [ + 49, + 78, + 145, + 150 + ], + "lines": [ + { + "bbox": [ + 49, + 78, + 145, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 78, + 145, + 150 + ], + "type": "image", + "image_path": "3ebabeaf66e0af185ce7b5cad3c6bfe929f0724965f0895d0ce4e6abccb73c11.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 49, + 159, + 145, + 232 + ], + "blocks": [ + { + "bbox": [ + 49, + 159, + 145, + 232 + ], + "lines": [ + { + "bbox": [ + 49, + 159, + 145, + 232 + ], + "spans": [ + { + "bbox": [ + 49, + 159, + 145, + 232 + ], + "type": "image", + "image_path": "904ef5a71e062b30d166abdee9c344f46dee956893cd7f79a9f1aa6d4abeb11d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 76, + 232, + 117, + 241 + ], + "lines": [ + { + "bbox": [ + 76, + 232, + 117, + 241 + ], + "spans": [ + { + "bbox": [ + 76, + 232, + 117, + 241 + ], + "type": "text", + "content": "Original view" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 149, + 49, + 274, + 116 + ], + "blocks": [ + { + "bbox": [ + 149, + 49, + 274, + 116 + ], + "lines": [ + { + "bbox": [ + 149, + 49, + 274, + 116 + ], + "spans": [ + { + "bbox": [ + 149, + 49, + 274, + 116 + ], + "type": "image", + "image_path": "e5431b158ca0c78d12e3cd5cdfd6ba72dfe7d0ed703ceb8cfe3471ec45adb669.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 149, + 117, + 274, + 185 + ], + "blocks": [ + { + "bbox": [ + 149, + 117, + 274, + 185 + ], + "lines": [ + { + "bbox": [ + 149, + 117, + 274, + 185 + ], + "spans": [ + { + "bbox": [ + 149, + 117, + 274, + 185 + ], + "type": "image", + "image_path": "e08735f3a506f77a852a134fb40a8b0aff814788bd092f94774788da1a1e55e7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 149, + 186, + 274, + 254 + ], + "blocks": [ + { + "bbox": [ + 149, + 186, + 274, + 254 + ], + "lines": [ + { + "bbox": [ + 149, + 186, + 274, + 254 + ], + "spans": [ + { + "bbox": [ + 149, + 186, + 274, + 254 + ], + "type": "image", + "image_path": "ad97e0ff4e94d396e2c8a8ae09059fb0865b1b6798ff91162c554b64952a1609.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 175, + 255, + 249, + 264 + ], + "lines": [ + { + "bbox": [ + 175, + 255, + 249, + 264 + ], + "spans": [ + { + "bbox": [ + 175, + 255, + 249, + 264 + ], + "type": "text", + "content": "\"Turn him into a clown\"" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 270, + 547, + 314 + ], + "lines": [ + { + "bbox": [ + 46, + 270, + 547, + 314 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 547, + 314 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison. It's important to note the level of control we maintain over the editing area (the whole body of the man). Background and other non-target regions are essentially unaffected, in contrast to Instruct-Nerf2Nerf [12] where the entire scene undergoes changes. GaussianEditor-DDS and GaussianEditor-iN2N indicate that we utilize delta denoising score [14] and Instruct-Nerf2Nerf [12] respectively, as guidance for editing." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 278, + 49, + 403, + 115 + ], + "blocks": [ + { + "bbox": [ + 278, + 49, + 403, + 115 + ], + "lines": [ + { + "bbox": [ + 278, + 49, + 403, + 115 + ], + "spans": [ + { + "bbox": [ + 278, + 49, + 403, + 115 + ], + "type": "image", + "image_path": "4da21a43eafbefc7a60a8342837cbd66788954ca0ea8f873080cd01f6793bcd9.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 278, + 117, + 403, + 185 + ], + "blocks": [ + { + "bbox": [ + 278, + 117, + 403, + 185 + ], + "lines": [ + { + "bbox": [ + 278, + 117, + 403, + 185 + ], + "spans": [ + { + "bbox": [ + 278, + 117, + 403, + 185 + ], + "type": "image", + "image_path": "6229dbee583e751dcecbd87e0f7766ea3be48d5ac4b60109301d4e21c201c7ec.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 278, + 186, + 403, + 254 + ], + "blocks": [ + { + "bbox": [ + 278, + 186, + 403, + 254 + ], + "lines": [ + { + "bbox": [ + 278, + 186, + 403, + 254 + ], + "spans": [ + { + "bbox": [ + 278, + 186, + 403, + 254 + ], + "type": "image", + "image_path": "2994751f0235f9bfa8e155a4b98f6960fc0a14400c95b094e271b16d13c1adcb.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 294, + 255, + 389, + 264 + ], + "lines": [ + { + "bbox": [ + 294, + 255, + 389, + 264 + ], + "spans": [ + { + "bbox": [ + 294, + 255, + 389, + 264 + ], + "type": "text", + "content": "\"Turn him into Albert Einstein\"" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 408, + 49, + 531, + 115 + ], + "blocks": [ + { + "bbox": [ + 408, + 49, + 531, + 115 + ], + "lines": [ + { + "bbox": [ + 408, + 49, + 531, + 115 + ], + "spans": [ + { + "bbox": [ + 408, + 49, + 531, + 115 + ], + "type": "image", + "image_path": "20e87eb054ea179d88612b0c8bccc0b2b0dfd7678b5cb0aabf304734d6e8744f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 534, + 59, + 541, + 99 + ], + "lines": [ + { + "bbox": [ + 534, + 59, + 541, + 99 + ], + "spans": [ + { + "bbox": [ + 534, + 59, + 541, + 99 + ], + "type": "text", + "content": "Instruct-N2N" + } + ] + } + ], + "index": 15, + "angle": 90, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 408, + 117, + 532, + 185 + ], + "blocks": [ + { + "bbox": [ + 408, + 117, + 532, + 185 + ], + "lines": [ + { + "bbox": [ + 408, + 117, + 532, + 185 + ], + "spans": [ + { + "bbox": [ + 408, + 117, + 532, + 185 + ], + "type": "image", + "image_path": "0a5fe71e53754e53af6e354f566a2cbefce62b938d84d6d1f152a89e3356f814.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 534, + 133, + 541, + 179 + ], + "lines": [ + { + "bbox": [ + 534, + 133, + 541, + 179 + ], + "spans": [ + { + "bbox": [ + 534, + 133, + 541, + 179 + ], + "type": "text", + "content": "GSEeditor-DDS" + } + ] + } + ], + "index": 16, + "angle": 90, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 408, + 186, + 532, + 254 + ], + "blocks": [ + { + "bbox": [ + 408, + 186, + 532, + 254 + ], + "lines": [ + { + "bbox": [ + 408, + 186, + 532, + 254 + ], + "spans": [ + { + "bbox": [ + 408, + 186, + 532, + 254 + ], + "type": "image", + "image_path": "ab3efa8079fd0cd213a6342973ef38407967576490cac8045ecd0f2081f4bd4e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 413, + 255, + 526, + 265 + ], + "lines": [ + { + "bbox": [ + 413, + 255, + 526, + 265 + ], + "spans": [ + { + "bbox": [ + 413, + 255, + 526, + 265 + ], + "type": "text", + "content": "\"Turn him into a Modigliani painting.\"" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 534, + 200, + 541, + 248 + ], + "lines": [ + { + "bbox": [ + 534, + 200, + 541, + 248 + ], + "spans": [ + { + "bbox": [ + 534, + 200, + 541, + 248 + ], + "type": "text", + "content": "GSEoperator-iN2N" + } + ] + } + ], + "index": 17, + "angle": 90, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 323, + 288, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 323, + 288, + 431 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 288, + 431 + ], + "type": "text", + "content": "flexibility. The older the generation, the stronger the constraints applied. Anchor loss is utilized to enforce these constraints. At the beginning of training, HGS records the attributes of all Gaussians as anchors. These anchors are then updated to reflect the current state of the Gaussians at each densification process. During training, MSE loss between the anchor state and the current state is employed to ensure that the Gaussians do not deviate too far from their respective anchors:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 107, + 437, + 287, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 437, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 107, + 437, + 287, + 468 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a n c h o r}} ^ {P} = \\sum_ {i = 0} ^ {n} \\lambda_ {i} \\left(P _ {i} - \\hat {P} _ {i}\\right) ^ {2} \\tag {7}", + "image_path": "bcd8322d174585b7e0ac0d138b1815cbfd619d9dd16f941b924657396fadd667.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 473, + 287, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 287, + 556 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 473, + 287, + 556 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 473, + 287, + 556 + ], + "type": "text", + "content": " represents the total number of Gaussians and " + }, + { + "bbox": [ + 46, + 473, + 287, + 556 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 473, + 287, + 556 + ], + "type": "text", + "content": " denotes a certain property of the current Gaussian, including elements from the set " + }, + { + "bbox": [ + 46, + 473, + 287, + 556 + ], + "type": "inline_equation", + "content": "x, s, q, \\alpha, c" + }, + { + "bbox": [ + 46, + 473, + 287, + 556 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 46, + 473, + 287, + 556 + ], + "type": "inline_equation", + "content": "\\hat{P}" + }, + { + "bbox": [ + 46, + 473, + 287, + 556 + ], + "type": "text", + "content": " refers to the same property recorded in the anchor state. The term " + }, + { + "bbox": [ + 46, + 473, + 287, + 556 + ], + "type": "inline_equation", + "content": "\\lambda_{i}" + }, + { + "bbox": [ + 46, + 473, + 287, + 556 + ], + "type": "text", + "content": " indicates the strength of the anchor loss applied to the " + }, + { + "bbox": [ + 46, + 473, + 287, + 556 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 473, + 287, + 556 + ], + "type": "text", + "content": "-th Gaussian, which varies based on its generation. The overall training loss is defined as:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 88, + 561, + 287, + 588 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 561, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 88, + 561, + 287, + 588 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {E d i t} + \\sum_ {P \\in \\{x, s, q, \\alpha , c \\}} \\lambda_ {P} \\mathcal {L} _ {a n c h o r} ^ {P} \\tag {8}", + "image_path": "5ac6a29c91c2bb8ce7fb63e101d09d26b96870dafe6d3bbeb954d09a337859bc.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 46, + 594, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 628 + ], + "type": "text", + "content": "In this equation, " + }, + { + "bbox": [ + 46, + 594, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\lambda_{P}" + }, + { + "bbox": [ + 46, + 594, + 287, + 628 + ], + "type": "text", + "content": " signifies the strength of the anchor loss applied to property " + }, + { + "bbox": [ + 46, + 594, + 287, + 628 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 594, + 287, + 628 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 594, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Edit}" + }, + { + "bbox": [ + 46, + 594, + 287, + 628 + ], + "type": "text", + "content": " is the edit loss defined in Sec. 3.2." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "content": "This generational design in HGS prevents the issue of excessive flexibility in GS when faced with stochastic loss. With each densification, the anchor loss weight " + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{i}" + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "content": " for all previous generations of Gaussians is increased. As a result, the fluidity of the existing generations gradually decreases until it nearly solidifies. This approach ensures stable geometry formation under stochastic losses, relying on the almost" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 324, + 546, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 324, + 546, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 324, + 546, + 407 + ], + "type": "text", + "content": "unconstrained Gaussians from new densifications to carve out details. Furthermore, this method of applying anchor loss can effectively meet various editing needs. For instance, to limit changes in the original GS, one can increase the anchor loss weight for generation 0. Similarly, if there is no desire to alter color or geometry during editing, a stronger anchor loss can be applied to these specific properties." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 408, + 547, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 408, + 547, + 504 + ], + "spans": [ + { + "bbox": [ + 304, + 408, + 547, + 504 + ], + "type": "text", + "content": "Additionally, to address the challenge of manually determining a densification threshold, we regulate the densification process based on a percentage criterion. In this method, during each densification step, we selectively densify only those Gaussians whose 3D position gradients are within the top " + }, + { + "bbox": [ + 304, + 408, + 547, + 504 + ], + "type": "inline_equation", + "content": "k\\%" + }, + { + "bbox": [ + 304, + 408, + 547, + 504 + ], + "type": "text", + "content": ". This strategy proves to be more manageable and intuitive than directly setting a threshold value in the Hierarchical Gaussian Splatting (HGS) framework." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 514, + 394, + 527 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 514, + 394, + 527 + ], + "spans": [ + { + "bbox": [ + 306, + 514, + 394, + 527 + ], + "type": "text", + "content": "4.3. 3D Inpainting" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 533, + 546, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 546, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 546, + 689 + ], + "type": "text", + "content": "Object Removal. Simply removing Gaussians identified by a mask can lead to artifacts, especially at the interfaces where the object intersects with other Gaussians. To address this, we employ 2D inpainting techniques to provide guidance for filling these areas. However, effective 2D inpainting requires precise masks to offer better guidance. To generate these masks, after deletion, we use the KNN algorithm to identify Gaussians nearest to the ones removed, which are likely at the interface. These are then projected onto various views. We subsequently dilate the mask and fix any holes to accurately represent the interface area, thus creating a refined mask for the boundary zones. The whole object removal procedure typically takes only two minutes." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "text", + "content": "Object Incorporation. We define this task as follows: Within the 3D Gaussians " + }, + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "text", + "content": ", given a camera pose " + }, + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "text", + "content": " and the" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21481" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 49, + 170, + 129 + ], + "blocks": [ + { + "bbox": [ + 49, + 49, + 170, + 129 + ], + "lines": [ + { + "bbox": [ + 49, + 49, + 170, + 129 + ], + "spans": [ + { + "bbox": [ + 49, + 49, + 170, + 129 + ], + "type": "image", + "image_path": "9599e95f325b4aef1c133d57fb1078f364f81f6f3842b900c43845bf2016623b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 173, + 47, + 295, + 129 + ], + "blocks": [ + { + "bbox": [ + 173, + 47, + 295, + 129 + ], + "lines": [ + { + "bbox": [ + 173, + 47, + 295, + 129 + ], + "spans": [ + { + "bbox": [ + 173, + 47, + 295, + 129 + ], + "type": "image", + "image_path": "2c769853edacba0890b0e4c18d8821f36c2c1659794645a0b862b334b3c21099.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 260, + 131, + 328, + 141 + ], + "lines": [ + { + "bbox": [ + 260, + 131, + 328, + 141 + ], + "spans": [ + { + "bbox": [ + 260, + 131, + 328, + 141 + ], + "type": "text", + "content": "\"Make it snowy\"" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 298, + 47, + 420, + 129 + ], + "blocks": [ + { + "bbox": [ + 298, + 47, + 420, + 129 + ], + "lines": [ + { + "bbox": [ + 298, + 47, + 420, + 129 + ], + "spans": [ + { + "bbox": [ + 298, + 47, + 420, + 129 + ], + "type": "image", + "image_path": "c5e470ff4bdb73e8b4696f570da95ad40f25e00aa32c865e2b164f78ba00e1b1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 424, + 47, + 545, + 129 + ], + "blocks": [ + { + "bbox": [ + 424, + 47, + 545, + 129 + ], + "lines": [ + { + "bbox": [ + 424, + 47, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 424, + 47, + 545, + 129 + ], + "type": "image", + "image_path": "ebb9c7e1701d8bc5025bf005659c02391b3663970ee1092fc12fe7c5c7ee59d1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 49, + 144, + 168, + 235 + ], + "blocks": [ + { + "bbox": [ + 49, + 144, + 168, + 235 + ], + "lines": [ + { + "bbox": [ + 49, + 144, + 168, + 235 + ], + "spans": [ + { + "bbox": [ + 49, + 144, + 168, + 235 + ], + "type": "image", + "image_path": "65f3fa13daa9c9311c3f699342e5ca879d8fef3fcb8a74f5378966b3bbe5b147.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 173, + 144, + 293, + 235 + ], + "blocks": [ + { + "bbox": [ + 173, + 144, + 293, + 235 + ], + "lines": [ + { + "bbox": [ + 173, + 144, + 293, + 235 + ], + "spans": [ + { + "bbox": [ + 173, + 144, + 293, + 235 + ], + "type": "image", + "image_path": "c9bb3b3505323a0a8196e6b4f85ab7f411be00d561e91a7c2e9df1fdcfbe2e24.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 238, + 236, + 350, + 247 + ], + "lines": [ + { + "bbox": [ + 238, + 236, + 350, + 247 + ], + "spans": [ + { + "bbox": [ + 238, + 236, + 350, + 247 + ], + "type": "text", + "content": "\"Turn him into an old lady\"" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 298, + 144, + 420, + 235 + ], + "blocks": [ + { + "bbox": [ + 298, + 144, + 420, + 235 + ], + "lines": [ + { + "bbox": [ + 298, + 144, + 420, + 235 + ], + "spans": [ + { + "bbox": [ + 298, + 144, + 420, + 235 + ], + "type": "image", + "image_path": "0e26f8c1349b33571e7a255613d9b3918f26733e21df07532efc56d13c7191f6.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 424, + 144, + 545, + 235 + ], + "blocks": [ + { + "bbox": [ + 424, + 144, + 545, + 235 + ], + "lines": [ + { + "bbox": [ + 424, + 144, + 545, + 235 + ], + "spans": [ + { + "bbox": [ + 424, + 144, + 545, + 235 + ], + "type": "image", + "image_path": "2cd1bb935d0d135a14b1a790e7e0da193362d77f605458b276e07b7d53463048.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 49, + 249, + 170, + 340 + ], + "blocks": [ + { + "bbox": [ + 49, + 249, + 170, + 340 + ], + "lines": [ + { + "bbox": [ + 49, + 249, + 170, + 340 + ], + "spans": [ + { + "bbox": [ + 49, + 249, + 170, + 340 + ], + "type": "image", + "image_path": "fd03a26a7a83e1fe6d04f48da2a46c11c77a3802382dc58c7f7336b2cfb6d928.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 361, + 547, + 395 + ], + "lines": [ + { + "bbox": [ + 46, + 361, + 547, + 395 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 547, + 395 + ], + "type": "text", + "content": "Figure 6. Extensive Results of GaussianEditor. Our method is capable of various editing tasks, including face and scene editing. In face and bear editing, we restrict the editing area to the face using Gaussian semantic tracing, ensuring that undesired areas remain unchanged. The leftmost column demonstrates the original view, while the right three columns show the images after editing." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 173, + 249, + 295, + 340 + ], + "blocks": [ + { + "bbox": [ + 173, + 249, + 295, + 340 + ], + "lines": [ + { + "bbox": [ + 173, + 249, + 295, + 340 + ], + "spans": [ + { + "bbox": [ + 173, + 249, + 295, + 340 + ], + "type": "image", + "image_path": "3556919a9a88b24b62f61a436e1f03f5a0346c156db177fec6227da8aefe487a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 226, + 342, + 367, + 353 + ], + "lines": [ + { + "bbox": [ + 226, + 342, + 367, + 353 + ], + "spans": [ + { + "bbox": [ + 226, + 342, + 367, + 353 + ], + "type": "text", + "content": "\"Turn the bear into a grizzly bear\"" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 298, + 249, + 420, + 340 + ], + "blocks": [ + { + "bbox": [ + 298, + 249, + 420, + 340 + ], + "lines": [ + { + "bbox": [ + 298, + 249, + 420, + 340 + ], + "spans": [ + { + "bbox": [ + 298, + 249, + 420, + 340 + ], + "type": "image", + "image_path": "232cb283a868f1a6540735f1f505f2b11ad7b58d77bc0526265c02b9dbe2bb40.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 424, + 249, + 545, + 340 + ], + "blocks": [ + { + "bbox": [ + 424, + 249, + 545, + 340 + ], + "lines": [ + { + "bbox": [ + 424, + 249, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 424, + 249, + 545, + 340 + ], + "type": "image", + "image_path": "d87325b0cfb59cb8265af45d88ef40dd70eced37096277c7dfc4d18004cfbc6b.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 411, + 288, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 288, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 288, + 471 + ], + "type": "text", + "content": "corresponding rendering " + }, + { + "bbox": [ + 46, + 411, + 288, + 471 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 411, + 288, + 471 + ], + "type": "text", + "content": " from this viewpoint, the user provides a 2D mask " + }, + { + "bbox": [ + 46, + 411, + 288, + 471 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 411, + 288, + 471 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 46, + 411, + 288, + 471 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 411, + 288, + 471 + ], + "type": "text", + "content": " indicating the area they wish to inpaint. Additionally, a prompt " + }, + { + "bbox": [ + 46, + 411, + 288, + 471 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 411, + 288, + 471 + ], + "type": "text", + "content": " is provided to specify the content of the inpainting. We then update " + }, + { + "bbox": [ + 46, + 411, + 288, + 471 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 411, + 288, + 471 + ], + "type": "text", + "content": " to fulfill the inpainting request." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 475, + 287, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 475, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 475, + 287, + 559 + ], + "type": "text", + "content": "Given " + }, + { + "bbox": [ + 46, + 475, + 287, + 559 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 475, + 287, + 559 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 475, + 287, + 559 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 475, + 287, + 559 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 475, + 287, + 559 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 475, + 287, + 559 + ], + "type": "text", + "content": ", the process begins with generating a 2D inpainted image " + }, + { + "bbox": [ + 46, + 475, + 287, + 559 + ], + "type": "inline_equation", + "content": "I^M y" + }, + { + "bbox": [ + 46, + 475, + 287, + 559 + ], + "type": "text", + "content": ", utilizing a 2D inpainting diffusion model as per [34]. Subsequently, the foreground object from " + }, + { + "bbox": [ + 46, + 475, + 287, + 559 + ], + "type": "inline_equation", + "content": "I^M y" + }, + { + "bbox": [ + 46, + 475, + 287, + 559 + ], + "type": "text", + "content": ", created by [34], is segmented and input into the image-to-3D method referenced in [25] to generate a coarse 3D mesh. This coarse mesh is then transformed into 3D Gaussians " + }, + { + "bbox": [ + 46, + 475, + 287, + 559 + ], + "type": "inline_equation", + "content": "\\theta_y" + }, + { + "bbox": [ + 46, + 475, + 287, + 559 + ], + "type": "text", + "content": ", and refined with HGS detailed in Sec. 4.2." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 564, + 288, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 564, + 288, + 673 + ], + "spans": [ + { + "bbox": [ + 46, + 564, + 288, + 673 + ], + "type": "text", + "content": "For aligning the coordinate system of " + }, + { + "bbox": [ + 46, + 564, + 288, + 673 + ], + "type": "inline_equation", + "content": "\\theta_y" + }, + { + "bbox": [ + 46, + 564, + 288, + 673 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 564, + 288, + 673 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 564, + 288, + 673 + ], + "type": "text", + "content": ", the depth of " + }, + { + "bbox": [ + 46, + 564, + 288, + 673 + ], + "type": "inline_equation", + "content": "I_y^M" + }, + { + "bbox": [ + 46, + 564, + 288, + 673 + ], + "type": "text", + "content": " is first estimated using the technique from [38]. This depth is then aligned with the depth map rendered by " + }, + { + "bbox": [ + 46, + 564, + 288, + 673 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 564, + 288, + 673 + ], + "type": "text", + "content": " at camera pose " + }, + { + "bbox": [ + 46, + 564, + 288, + 673 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 564, + 288, + 673 + ], + "type": "text", + "content": ", using the least squares method. With this alignment, we can accurately determine the coordinates and scale of the inpainted foreground object in the coordinate system of " + }, + { + "bbox": [ + 46, + 564, + 288, + 673 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 564, + 288, + 673 + ], + "type": "text", + "content": ". After transforming " + }, + { + "bbox": [ + 46, + 564, + 288, + 673 + ], + "type": "inline_equation", + "content": "\\theta_y" + }, + { + "bbox": [ + 46, + 564, + 288, + 673 + ], + "type": "text", + "content": " into the coordinate system of " + }, + { + "bbox": [ + 46, + 564, + 288, + 673 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 564, + 288, + 673 + ], + "type": "text", + "content": ", we simply concatenate them to produce the final inpainted 3D Gaussians." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "content": "It is important to note that due to our efficient design, the entire object incorporation procedure can be completed in approximately 5 minutes." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 409, + 387, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 409, + 387, + 423 + ], + "spans": [ + { + "bbox": [ + 306, + 409, + 387, + 423 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 430, + 440, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 430, + 440, + 442 + ], + "spans": [ + { + "bbox": [ + 306, + 430, + 440, + 442 + ], + "type": "text", + "content": "5.1. Implementation Details" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 449, + 547, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 547, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 547, + 544 + ], + "type": "text", + "content": "We utilize the highly optimized renderer implementation from [16] for Gaussian rendering and base our implementation on Threestudio [11]. All the original 3D Gaussians used in this work are trained using the methods described in [16]. Our experiments are conducted on a single RTX A6000 GPU. As detailed in Sec. 4.1, once we obtain segmentation masks from the 2D segmentation method outlined in [18], segmenting the 3D Gaussians takes only about 1 second." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 545, + 547, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 640 + ], + "type": "text", + "content": "For editing large scenes, the camera poses employed during the editing process are selected from a subset of the multi-view image dataset initially used for reconstruction. Depending on the complexity of the scene, the number of camera poses used in our experiments varies from 24 to 96. The editing process, influenced by the specified prompt and the complexity of the scene, typically involves optimizing for 500-1000 steps, taking about 5-10 minutes in total." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 642, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 712 + ], + "type": "text", + "content": "Regarding 3D inpainting for object incorporation, as detailed in Sec. 4.3, it takes approximately 3 minutes to generate a 3D mesh using the method from [25] and an additional 2 minutes to transfer this mesh into 3D Gaussians and refine it, while the composition process of two Gaussians takes less than 1 second." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21482" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 127, + 175 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 127, + 175 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 127, + 175 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 127, + 175 + ], + "type": "image", + "image_path": "45031bcd2073026aca19d4de3ba66516ad3104370caf41dcbaea304f05f5fcb8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 176, + 111, + 186 + ], + "lines": [ + { + "bbox": [ + 64, + 176, + 111, + 186 + ], + "spans": [ + { + "bbox": [ + 64, + 176, + 111, + 186 + ], + "type": "text", + "content": "Original View" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 129, + 70, + 206, + 175 + ], + "blocks": [ + { + "bbox": [ + 129, + 70, + 206, + 175 + ], + "lines": [ + { + "bbox": [ + 129, + 70, + 206, + 175 + ], + "spans": [ + { + "bbox": [ + 129, + 70, + 206, + 175 + ], + "type": "image", + "image_path": "b79313259f059997c6a3e93f4268768e763521f2e67d6fc0ec4ee0b5f92213a5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 177, + 173, + 185 + ], + "lines": [ + { + "bbox": [ + 162, + 177, + 173, + 185 + ], + "spans": [ + { + "bbox": [ + 162, + 177, + 173, + 185 + ], + "type": "text", + "content": "GS" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 209, + 70, + 287, + 175 + ], + "blocks": [ + { + "bbox": [ + 209, + 70, + 287, + 175 + ], + "lines": [ + { + "bbox": [ + 209, + 70, + 287, + 175 + ], + "spans": [ + { + "bbox": [ + 209, + 70, + 287, + 175 + ], + "type": "image", + "image_path": "b3991c369740ac2e2ff04101002d89855f37350aa35ca8bbee066c047d5f0cee.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 239, + 177, + 256, + 185 + ], + "lines": [ + { + "bbox": [ + 239, + 177, + 256, + 185 + ], + "spans": [ + { + "bbox": [ + 239, + 177, + 256, + 185 + ], + "type": "text", + "content": "HGS" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 195, + 289, + 284 + ], + "lines": [ + { + "bbox": [ + 46, + 195, + 289, + 284 + ], + "spans": [ + { + "bbox": [ + 46, + 195, + 289, + 284 + ], + "type": "text", + "content": "Figure 7. Ablation study on Hierarchical Gaussian Splitting (HGS). Prompt: make the grass on fire. Even when specifying the editing area with prompts, generative methods like InstructPix2Pix [4] tend to edit the entire 2D image. Without HGS, Gaussians tend to conform to this whole-image editing by spreading and densifying across the entire scene, leading to uncontrollable densification and blurring of the image. With HGS, however, this kind of diffusion is effectively restrained." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 305, + 187, + 318 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 305, + 187, + 318 + ], + "spans": [ + { + "bbox": [ + 47, + 305, + 187, + 318 + ], + "type": "text", + "content": "5.2. Qualitative Comparisons" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 323, + 288, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 323, + 288, + 418 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 288, + 418 + ], + "type": "text", + "content": "As illustrated in Fig. 5, GaussianEditor-iN2N surpasses other methods in both the quality of edits and controllability. Instruct-Nerf2Nerf, while producing edits with insufficient detail, cannot also control the editing area. GaussianEditorDDS, due to the more challenging control of guidance offered by DDS loss [14] compared to Instruct-pix2pix [4], tends to result in oversaturated colors and less precise editing outcomes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 419, + 287, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 419, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 419, + 287, + 540 + ], + "type": "text", + "content": "Additionally, our method exhibits exceptional control over the editing area. This is achieved through Gaussian semantic tracing, which identifies the Gaussians that require editing at each training step, for example, the entire human body in Fig. 5. It's important to note that in Instruct-Nerf2Nerf [12], the use of static 2D or 3D masks restricts the spatial freedom of the edits, as the permissible area for the edited subject is limited by these masks. Furthermore, the effectiveness of static masks diminishes as the geometries and appearances of 3D models evolve during training." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 540, + 288, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 288, + 612 + ], + "type": "text", + "content": "In Fig. 6, we demonstrate that GaussianEditor can accommodate a variety of scenarios, such as editing in large-scale scenes and facial swaps. In the case of large scenes, we did not apply Gaussian semantic tracing. However, for facial swaps, we traced the Gaussians corresponding to facial regions, achieving controllable and realistic editing." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 620, + 194, + 631 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 620, + 194, + 631 + ], + "spans": [ + { + "bbox": [ + 47, + 620, + 194, + 631 + ], + "type": "text", + "content": "5.3.Quantitative Comparisons" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 631, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 631, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 631, + 288, + 713 + ], + "type": "text", + "content": "As shown in Table 1, we conduct quantitative comparisons on user study and CLIP directional similarity (as shown in InstructPix2Pix [4] and StyleGAN-Nada [9]). GaussianEditor-iN2N not only demonstrates superior outcomes in user studies but also excels in CLIP Directional Similarity. Besides, Instruct-Nerf2Nerf [12] typically requires more than 30 minutes to complete the editing of a scene, whereas our method" + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 309, + 71, + 545, + 110 + ], + "blocks": [ + { + "bbox": [ + 309, + 71, + 545, + 110 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 545, + 110 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 545, + 110 + ], + "type": "table", + "html": "
iN2N [12]Ours (DDS)Ours (iN2N)
User study15.45%12.27%72.28%
CLIP Directional Similarity0.16000.18130.2071
", + "image_path": "75f672c4926f39a0ae9d4382b6aff1a73dfdfe1131c107b8b31bee82c0874c8f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 119, + 547, + 152 + ], + "lines": [ + { + "bbox": [ + 305, + 119, + 547, + 152 + ], + "spans": [ + { + "bbox": [ + 305, + 119, + 547, + 152 + ], + "type": "text", + "content": "Table 1. Quantitative Comparison. GaussianEditor-iN2N outperforms in both user study evaluations and CLIP Directional Similarity [9] metrics." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 172, + 451, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 172, + 451, + 184 + ], + "spans": [ + { + "bbox": [ + 306, + 172, + 451, + 184 + ], + "type": "text", + "content": "only takes between 5 to 10 minutes." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 190, + 399, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 190, + 399, + 204 + ], + "spans": [ + { + "bbox": [ + 306, + 190, + 399, + 204 + ], + "type": "text", + "content": "5.4. Ablation Study" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 209, + 547, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 209, + 547, + 329 + ], + "spans": [ + { + "bbox": [ + 304, + 209, + 547, + 329 + ], + "type": "text", + "content": "As demonstrated in Fig. 7, we conducted ablation experiments on Hierarchical Gaussian Splitting(HGS). Without HGS, Gaussians tend to spread and densify across the scene, leading to uncontrolled densification and image blurring. This is typically caused by the tendency of methods like Instruct-Pix2Pix to edit the entire 2D image when prompts are used to define editing areas. However, HGS effectively circumvents this issue by constraining the mobility of the Gaussian in the old generation, ensuring that the overall scene does not exhibit excessive mobility." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 338, + 378, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 338, + 378, + 350 + ], + "spans": [ + { + "bbox": [ + 306, + 338, + 378, + 350 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 358, + 547, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 358, + 547, + 478 + ], + "spans": [ + { + "bbox": [ + 304, + 358, + 547, + 478 + ], + "type": "text", + "content": "In our research, we introduce GaussianEditor, an innovative 3D editing algorithm based on Gaussian Splatting, designed for enhanced control and efficiency. Our method employs Gaussian semantic tracing for precise identification and targeting of editing areas, followed by Hierarchical Gaussian Splatting (HGS) to balance fluidity and stability in achieving detailed results under stochastic guidance. Additionally, we developed a specialized 3D inpainting algorithm for Gaussian Splatting, streamlining object removal and integration, and greatly reducing editing time." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 479, + 547, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 547, + 538 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 547, + 538 + ], + "type": "text", + "content": "Limitation. Similar to previous 3D editing works based on 2D diffusion models, GaussianEditor relies on these models to provide effective supervision. However, current 2D diffusion models struggle to offer effective guidance for certain complex prompts, leading to limitations in 3D editing." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 548, + 421, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 548, + 421, + 562 + ], + "spans": [ + { + "bbox": [ + 306, + 548, + 421, + 562 + ], + "type": "text", + "content": "7. Acknowledgements" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 568, + 547, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 568, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 568, + 547, + 628 + ], + "type": "text", + "content": "This study is supported under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). This research is also supported by the MoE AcRF Tier 2 grant (MOE-T2EP20220-0007)." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 638, + 365, + 650 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 638, + 365, + 650 + ], + "spans": [ + { + "bbox": [ + 306, + 638, + 365, + 650 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 310, + 657, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 310, + 657, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 657, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 310, + 657, + 545, + 689 + ], + "type": "text", + "content": "[1] Chong Bao, Yinda Zhang, and Bangbang et al. Yang. Sine: Semantic-driven image-based nef editing with prior-guided editing field. In CVPR 2023, pages 20919-20929, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 311, + 691, + 547, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 691, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 311, + 691, + 547, + 712 + ], + "type": "text", + "content": "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan." + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21483" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855-5864, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 107, + 288, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 107, + 288, + 140 + ], + "spans": [ + { + "bbox": [ + 53, + 107, + 288, + 140 + ], + "type": "text", + "content": "[3] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. CVPR, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 141, + 288, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 141, + 288, + 174 + ], + "spans": [ + { + "bbox": [ + 53, + 141, + 288, + 174 + ], + "type": "text", + "content": "[4] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. arXiv preprint arXiv:2211.09800, 2022. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 175, + 288, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 175, + 288, + 218 + ], + "spans": [ + { + "bbox": [ + 53, + 175, + 288, + 218 + ], + "type": "text", + "content": "[5] Yiwen Chen, Chi Zhang, Xiaofeng Yang, Zhongang Cai, Gang Yu, Lei Yang, and Guosheng Lin. It3d: Improved text-to-3d generation with explicit view synthesis. arXiv preprint arXiv:2308.11473, 2023. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 220, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 220, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 53, + 220, + 288, + 262 + ], + "type": "text", + "content": "[6] Zhiqin Chen, Thomas Funkhouser, Peter Hedman, and Andrea Tagliasacchi. Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. arXiv preprint arXiv:2208.00277, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 264, + 288, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 264, + 288, + 297 + ], + "spans": [ + { + "bbox": [ + 53, + 264, + 288, + 297 + ], + "type": "text", + "content": "[7] Zilong Chen, Feng Wang, and Huaping Liu. Text-to-3d using gaussian splatting. arXiv preprint arXiv:2309.16585, 2023. 2, 3, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 298, + 288, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 298, + 288, + 342 + ], + "spans": [ + { + "bbox": [ + 53, + 298, + 288, + 342 + ], + "type": "text", + "content": "[8] Xinhua Cheng, Tianyu Yang, Jianan Wang, Yu Li, Lei Zhang, Jian Zhang, and Li Yuan. Progressive3d: Progressively local editing for text-to-3d content creation with complex semantic prompts. arXiv preprint arXiv:2310.11784, 2023. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 344, + 288, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 344, + 288, + 387 + ], + "spans": [ + { + "bbox": [ + 53, + 344, + 288, + 387 + ], + "type": "text", + "content": "[9] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clipguided domain adaptation of image generators. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 388, + 288, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 288, + 431 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 288, + 431 + ], + "type": "text", + "content": "[10] William Gao, Noam Aigerman, Thibault Groueix, Vladimir G Kim, and Rana Hanocka. Textdeformer: Geometry manipulation using text guidance. arXiv preprint arXiv:2304.13348, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 434, + 288, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 288, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 288, + 487 + ], + "type": "text", + "content": "[11] Yuan-Chen Guo, Ying-Tian Liu, Chen Wang, Zi-Xin Zou, Guan Luo, Chia-Hao Chen, Yan-Pei Cao, and Song-Hai Zhang. threestudio: A unified framework for 3d content generation. https://github.com/threestudio-project/threestudio, 2023.7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 489, + 288, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 489, + 288, + 532 + ], + "spans": [ + { + "bbox": [ + 48, + 489, + 288, + 532 + ], + "type": "text", + "content": "[12] Ayaan Haque, Matthew Tancik, Alexei A Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. arXiv preprint arXiv:2303.12789, 2023. 2, 4, 6, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 534, + 288, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 288, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 288, + 567 + ], + "type": "text", + "content": "[13] Peter Hedman, Pratul P. Srinivasan, Ben Mildenhall, Jonathan T. Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. ICCV, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 568, + 288, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 288, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 288, + 610 + ], + "type": "text", + "content": "[14] Amir Hertz, Kfir Aberman, and Daniel Cohen-Or. Delta denoising score. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2328-2337, 2023. 6, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 613, + 288, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 288, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 288, + 655 + ], + "type": "text", + "content": "[15] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 42(4): 1-14, 2023. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 658, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 689 + ], + "type": "text", + "content": "[16] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ToG, 42(4):1-14, 2023. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "text", + "content": "[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 327, + 73, + 547, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 547, + 94 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 547, + 94 + ], + "type": "text", + "content": "field rendering. ACM Transactions on Graphics, 42(4), 2023. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 96, + 547, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 547, + 140 + ], + "type": "text", + "content": "[18] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 5, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 141, + 547, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 547, + 174 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 547, + 174 + ], + "type": "text", + "content": "[19] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nerf for editing via feature field distillation. arXiv preprint arXiv:2205.15585, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 175, + 547, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 175, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 175, + 547, + 217 + ], + "type": "text", + "content": "[20] Yuan Li, Zhi-Hao Lin, David Forsyth, Jia-Bin Huang, and Shenlong Wang. Climatenerf: Physically-based neural rendering for extreme climate synthesis. arXiv e-prints, pages arXiv-2211, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 220, + 547, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 547, + 262 + ], + "type": "text", + "content": "[21] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In CVPR, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 264, + 547, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 547, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 547, + 308 + ], + "type": "text", + "content": "[22] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, pages 300–309, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 309, + 547, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 309, + 547, + 342 + ], + "spans": [ + { + "bbox": [ + 308, + 309, + 547, + 342 + ], + "type": "text", + "content": "[23] Hao-Kang Liu, I Shen, Bing-Yu Chen, et al. Nerf-in: Free-form nerf inpainting with rgb-d priors. arXiv preprint arXiv:2206.04901, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 343, + 547, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 343, + 547, + 376 + ], + "spans": [ + { + "bbox": [ + 308, + 343, + 547, + 376 + ], + "type": "text", + "content": "[24] Steven Liu, Xiuming Zhang, Zhoutong Zhang, Richard Zhang, Jun-Yan Zhu, and Bryan Russell. Editing conditional radiance fields. In ICCV 2021, pages 5773-5783, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 377, + 547, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 377, + 547, + 431 + ], + "spans": [ + { + "bbox": [ + 308, + 377, + 547, + 431 + ], + "type": "text", + "content": "[25] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. arXiv preprint arXiv:2310.15008, 2023. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 433, + 547, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 433, + 547, + 475 + ], + "spans": [ + { + "bbox": [ + 308, + 433, + 547, + 475 + ], + "type": "text", + "content": "[26] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 478, + 547, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 478, + 547, + 531 + ], + "spans": [ + { + "bbox": [ + 308, + 478, + 547, + 531 + ], + "type": "text", + "content": "[27] Aryan Mikaeili, Or Perel, Mehdi Safaee, Daniel Cohen-Or, and Ali Mahdavi-Amiri. Sked: Sketch-guided text-based 3d editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14607-14619, 2023. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 533, + 547, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 533, + 547, + 577 + ], + "spans": [ + { + "bbox": [ + 308, + 533, + 547, + 577 + ], + "type": "text", + "content": "[28] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 578, + 547, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 578, + 547, + 621 + ], + "spans": [ + { + "bbox": [ + 308, + 578, + 547, + 621 + ], + "type": "text", + "content": "[29] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 624, + 547, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 547, + 656 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 547, + 656 + ], + "type": "text", + "content": "[30] Atsuhiro Noguchi, Xiao Sun, Stephen Lin, and Tatsuya Harada. Neural articulated radiance field. In ICCV 2021, pages 5762-5772, 2021. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 658, + 547, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 547, + 689 + ], + "type": "text", + "content": "[31] Jangho Park, Gihyun Kwon, and Jong Chul Ye. Ed-nerf: Efficient text-guided editing of 3d scene using latent space nerf. arXiv preprint arXiv:2310.02712, 2023. 2, 3, 4" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 691, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 547, + 713 + ], + "type": "text", + "content": "[32] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo Martin-" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "21484" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "Brualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "text", + "content": "[33] Sida Peng, Yuanqing Zhang, Yinghao Xu, and et al. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In CVPR 2021, pages 9054–9063, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 288, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 288, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 288, + 195 + ], + "type": "text", + "content": "[34] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 197, + 288, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 288, + 228 + ], + "type": "text", + "content": "[35] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 231, + 288, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 231, + 288, + 284 + ], + "spans": [ + { + "bbox": [ + 48, + 231, + 288, + 284 + ], + "type": "text", + "content": "[36] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 286, + 288, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 288, + 341 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 288, + 341 + ], + "type": "text", + "content": "[37] Amit Raj, Srinivas Kaza, Ben Poole, Michael Niemeyer, Nataniel Ruiz, Ben Mildenhall, Shiran Zada, Kfir Aberman, Michael Rubinstein, Jonathan Barron, et al. Dreambooth3d: Subject-driven text-to-3d generation. arXiv preprint arXiv:2303.13508, 2023. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 342, + 288, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 342, + 288, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 342, + 288, + 387 + ], + "type": "text", + "content": "[38] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12179-12188, 2021. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 388, + 288, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 288, + 430 + ], + "type": "text", + "content": "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 433, + 288, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 433, + 288, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 433, + 288, + 464 + ], + "type": "text", + "content": "[40] Sara Fridovich-Keil and Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 467, + 288, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 467, + 288, + 509 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 288, + 509 + ], + "type": "text", + "content": "[41] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 511, + 288, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 511, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 511, + 288, + 555 + ], + "type": "text", + "content": "[42] Etai Sella, Gal Fiebelman, Peter Hedman, and Hadar Averbuch-Elor. Vox-e: Text-guided voxel editing of 3d objects. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 430-440, 2023. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 556, + 288, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 556, + 288, + 600 + ], + "spans": [ + { + "bbox": [ + 48, + 556, + 288, + 600 + ], + "type": "text", + "content": "[43] Ruizhi Shao, Jingxiang Sun, Cheng Peng, Zerong Zheng, Boyao Zhou, Hongwen Zhang, and Yebin Liu. Control4d: Dynamic portrait editing by learning 4d gan from 2d diffusion-based editor. arXiv preprint arXiv:2305.20082, 2023. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 601, + 288, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 288, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 288, + 645 + ], + "type": "text", + "content": "[44] Chaoyue Song, Tianyi Chen, Yiwen Chen, Jiacheng Wei, Chuan Sheng Foo, Fayao Liu, and Guosheng Lin. Moda: Modeling deformable 3d objects from casual videos. arXiv preprint arXiv:2304.08279, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 689 + ], + "type": "text", + "content": "[45] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. 2, 3, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "text", + "content": "[46] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 622 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 546, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 546, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 546, + 106 + ], + "type": "text", + "content": "of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3835-3844, 2022. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 107, + 547, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 547, + 150 + ], + "type": "text", + "content": "[47] Can Wang, Ruixiang Jiang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Nerf-art: Text-driven neural radiance fields stylization. IEEE Transactions on Visualization and Computer Graphics, 2023. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 152, + 547, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 547, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 547, + 195 + ], + "type": "text", + "content": "[48] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 197, + 547, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 547, + 240 + ], + "type": "text", + "content": "[49] Tianhan Xu and Tatsuya Harada. Deforming radiance fields with cages. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIII, pages 159-175. Springer, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 241, + 547, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 547, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 547, + 285 + ], + "type": "text", + "content": "[50] Bangbang Yang, Chong Bao, and Junyi et al. Zeng. Neumesh: Learning disentangled neural mesh-based implicit field for geometry and texture editing. In ECCV 2022, pages 597-614. Springer, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 286, + 546, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 546, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 546, + 330 + ], + "type": "text", + "content": "[51] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 331, + 546, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 546, + 374 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 546, + 374 + ], + "type": "text", + "content": "[52] Zeyu Yang, Hongye Yang, Zijie Pan, Xiatian Zhu, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. arXiv preprint arXiv:2310.10642, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 376, + 547, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 547, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 547, + 430 + ], + "type": "text", + "content": "[53] Taoran Yi, Jiemin Fang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussian-dreamer: Fast generation from text to 3d gaussian splatting with point cloud priors. arXiv preprint arXiv:2310.08529, 2023. 2, 3, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 432, + 546, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 432, + 546, + 475 + ], + "spans": [ + { + "bbox": [ + 307, + 432, + 546, + 475 + ], + "type": "text", + "content": "[54] Wang Yifan, Felice Serena, Shihao Wu, Cengiz Öztireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM Transactions on Graphics (TOG), 38(6):1-14, 2019. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 476, + 546, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 476, + 546, + 509 + ], + "spans": [ + { + "bbox": [ + 307, + 476, + 546, + 509 + ], + "type": "text", + "content": "[55] Yu-Jie Yuan, Yang-Tian Sun, Yu-Kun Lai, and et al. Nerf-editing: geometry editing of neural radiance fields. In CVPR 2022, pages 18353-18364, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 510, + 546, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 510, + 546, + 543 + ], + "spans": [ + { + "bbox": [ + 307, + 510, + 546, + 543 + ], + "type": "text", + "content": "[56] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 544, + 546, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 544, + 546, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 544, + 546, + 578 + ], + "type": "text", + "content": "[57] Jingyu Zhuang, Chen Wang, Lingjie Liu, Liang Lin, and Guanbin Li. Dreameditor: Text-driven 3d scene editing with neural fields. arXiv preprint arXiv:2306.13455, 2023. 2, 3, 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 578, + 546, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 578, + 546, + 622 + ], + "spans": [ + { + "bbox": [ + 307, + 578, + 546, + 622 + ], + "type": "text", + "content": "[58] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Surface splatting. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 371-378, 2001. 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21485" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/GaussianShader_ 3D Gaussian Splatting with Shading Functions for Reflective Surfaces/e3ac7488-6672-4bab-abdc-38e188c8adc6_content_list.json b/2024/GaussianShader_ 3D Gaussian Splatting with Shading Functions for Reflective Surfaces/e3ac7488-6672-4bab-abdc-38e188c8adc6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..71327a6965f60ec6354599c156dea3462f1df9d7 --- /dev/null +++ b/2024/GaussianShader_ 3D Gaussian Splatting with Shading Functions for Reflective Surfaces/e3ac7488-6672-4bab-abdc-38e188c8adc6_content_list.json @@ -0,0 +1,1665 @@ +[ + { + "type": "text", + "text": "GaussianShader: 3D Gaussian Splating with Shading Functions for Reflective Surfaces", + "text_level": 1, + "bbox": [ + 158, + 128, + 810, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yingwenqi Jiang $^{1}$ Jiadong Tu $^{1}$ Yuan Liu $^{2}$ Xifeng Gao $^{3}$", + "bbox": [ + 238, + 202, + 730, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiaoxiao Long $^{2,*}$ Wenping Wang $^{4}$ Yuexin Ma $^{1,*}$", + "bbox": [ + 285, + 220, + 705, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ ShanghaiTech University", + "bbox": [ + 107, + 243, + 284, + 261 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2The University of Hong Kong", + "bbox": [ + 303, + 243, + 509, + 261 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3Tencent America", + "bbox": [ + 529, + 243, + 653, + 260 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Texas A&M University", + "bbox": [ + 673, + 243, + 836, + 261 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bd5035014d6f472068fcdb1864689d3d3862a3f53469223ea44f1c5f1c1dbffa.jpg", + "image_caption": [ + "PSNR:21.69" + ], + "image_footnote": [], + "bbox": [ + 76, + 305, + 189, + 373 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c69300f47475fd7e15f3521d175106b04c2c1913e4c6bddd1a3c69ca24d31e28.jpg", + "image_caption": [ + "PSNR: 28.87" + ], + "image_footnote": [], + "bbox": [ + 196, + 306, + 308, + 372 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0081ed2f08744dc3bb8ad153208d7331c3d3203dced177e3e0ac530988bcf6ce.jpg", + "image_caption": [ + "PSNR:32.67" + ], + "image_footnote": [], + "bbox": [ + 313, + 306, + 426, + 372 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8a9006e4b11d153459688da78ef59907630788cc29ed695d4dd11e76566e049b.jpg", + "image_caption": [ + "PSNR:32.46" + ], + "image_footnote": [], + "bbox": [ + 431, + 306, + 544, + 373 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e4c9d9ee7e4c248ab9a75b9e99441f9970e9819a7fb9b399da3cf2a903306579.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 549, + 306, + 661, + 375 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e6a7bb1b1a0b7da5b62ffd5100d2ee0b3ec3784a7b812a98c7385ac4018d7db2.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 558, + 385, + 638, + 472 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/439a08bcd9b263589bd7bc0e52209218cb1f4da2c557044437df4185499cdd5b.jpg", + "image_caption": [ + "(b) Comparisons between FPS, training time and accuracy", + "Figure 1. GaussianShader maintains real-time rendering speed and renders high-fidelity images for both general and reflective surfaces. Ref-NeRF[45] and ENVIDR[27] attempt to handle reflective surfaces, but they suffer from quite time-consuming optimization and slow rendering speed. 3D Gaussian splattering [21] keeps high efficiency but cannot handle such reflective surfaces." + ], + "image_footnote": [], + "bbox": [ + 669, + 316, + 890, + 483 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 561, + 313, + 575 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The advent of neural 3D Gaussians [21] has recently brought about a revolution in the field of neural rendering, facilitating the generation of high-quality renderings at real-time speeds. However, the explicit and discrete representation encounters challenges when applied to scenes featuring reflective surfaces. In this paper, we present GaussianShader, a novel method that applies a simplified shading function on 3D Gaussians to enhance the neural rendering in scenes with reflective surfaces while preserving the training and rendering efficiency. The main challenge in applying the shading function lies in the accurate normal estimation on discrete 3D Gaussians. Specifically, we proposed a novel normal estimation framework based on the shortest axis directions of 3D Gaussians with a delicately designed loss to make the consistency between the normals and the geometries of Gaussian spheres. Exper", + "bbox": [ + 75, + 578, + 468, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "periments show that GaussianShader strikes a commendable balance between efficiency and visual quality. Our method surpasses Gaussian Splatting [21] in PSNR on specular object datasets, exhibiting an improvement of 1.57dB. When compared to prior works handling reflective surfaces, such as Ref-NeRF [45], our optimization time is significantly accelerated (23h vs. 0.58h). Please click on our project website to see more results.", + "bbox": [ + 496, + 561, + 892, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 708, + 632, + 723 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, the field of 3D computer vision has witnessed remarkable advancements in the 3D reconstruction and visualization of 3D scenes. Innovations such as Neural Radiance Fields (NeRF) [32] have achieved substantial breakthroughs in generating novel views of 3D objects and scenes, presenting the potential for high-quality and photorealistic renderings. Despite these advancements, NeRF-related methodologies [6, 27, 45], still contend with challenges such as computationally expensive optimization and slow rendering speed. These limitations restrict their application in real-time interactive scenarios.", + "bbox": [ + 496, + 734, + 892, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author. This work was supported by NSFC (No.62206173), MoE Key Laboratory of Intelligent Perception and Human-Machine Collaboration (ShanghaiTech University), Shanghai Frontiers Science Center of Human-centered Artificial Intelligence, the Innovation and Technology Commission of the HKSAR Government under the InnoHK initiative, and Ref. T45-205/21-N of Hong Kong RGC.", + "bbox": [ + 75, + 828, + 468, + 902 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5322", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "More recently, 3D Gaussian Splatting [21] combines 3D Gaussian representation and tile-based splatting techniques to achieve high-quality 3D scene modeling and real-time rendering, making it possible to employ neural rendering techniques in real applications. However, it suffers from a performance drop on scenes featuring specular and reflective surfaces. This is because 3D Gaussian Splatting [21] does not explicitly model appearance properties, so that fails to capture significant view-dependent changes, particularly specular highlights. This constraint presents a substantial obstacle in the pursuit of achieving photorealistic rendering across a diverse array of materials, particularly those characterized by prominent reflective attributes.", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Accurately modeling reflective surfaces is a challenging task. Ref-NeRF [45] and ENVIDR [27] combine the shading functions in implicit representations and present promising quality on reflective surfaces. However, they suffer from time-consuming optimization (hours) and slow rendering speed. Due to the limited flexibility of SDF, ENVIDR [27] even fails to model complex scenes and presents a significant performance drop on general objects. It is still an unexplored problem how to combine the shading functions in a 3D Gaussian Splatting framework to improve its ability to handle reflections while preserving the efficiency in training and rendering.", + "bbox": [ + 75, + 290, + 472, + 472 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we present GaussianShader, a novel method that enhances the neural rendering of 3D Gaussians within scenes that contain reflective surfaces by incorporating a shading function on 3D Gaussians. To ensure the efficiency of GaussianShader, evaluating the shading function cannot be too expensive while still retaining the ability to model the reflections. In light of this, we propose a novel simplified shading function that considers the diffuse colors and the direct reflections while putting all advanced complex reflections into a residual color term. In comparison with the shading function of Ref-NeRF which can only consider direct reflections, the utilization of this residual color enables GaussianShader to render more complex reflective appearances with efficiency.", + "bbox": [ + 75, + 474, + 472, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another challenge in computing a shading function is how to predict accurate normals on the discrete 3D Gaussian spheres. First, it is hard to get a locally-continuous surface from the 3D Gaussians to compute the surface normals. Second, associating multiple 3D Gaussians for normal computation would be very expensive using neighborhood searching. In GaussianShader, we address this problem by introducing a new normal representation, which is based on the shortest axis direction of a Gaussian sphere and learns a normal residual on this axis direction. Then, to enforce consistency between the estimated normals and the geometry formulated by Gaussian spheres, we introduce an efficient normal-geometry constraint between the predicted normals and the normals derived from the rendered depths.", + "bbox": [ + 75, + 688, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Both the normal representation and the constraint lead to an accurate normal estimation of Gaussian spheres, which helps us compute the shading function.", + "bbox": [ + 496, + 90, + 890, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Built upon 3D Gaussian Splatting, GaussianShader maintains real-time rendering speed while still accommodating various materials like reflective surfaces. Experiments show that, compared to prior works, our method keeps a good balance between efficiency and robust performance on both general scenes and reflective surfaces. In summary, our method offers several significant advantages:", + "bbox": [ + 496, + 136, + 890, + 242 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Our method explicitly approximates the rendering equation by a simplified shading function, significantly enhancing the realism of rendered scenes, particularly for highly specular and reflective surfaces.", + "2. We propose a new normal estimation framework on 3D Gaussians with a new regularization loss that allows precise normal estimation.", + "3. Leveraging the efficiency of Gaussian Splatting, our approach provides real-time rendering capabilities, making it suitable for interactive applications and scenarios that demand efficient rendering." + ], + "bbox": [ + 500, + 243, + 890, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 424, + 640, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Neural Radiance Fields", + "text_level": 1, + "bbox": [ + 500, + 449, + 718, + 464 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural Radiance Fields (NeRF) [32] gains remarkable progress in photo-realistic novel view synthesis using implicit representation and volume rendering. Recently, NeRF has inspired many follow-up works in various directions. [5, 6] improves NeRF in rendering quality by introducing 3D conical frustum, achieving state-of-art performance in NVS. [17, 36, 47, 52, 53] combine implicit surface representations with NeRF for more accurate geometric reconstruction. [43, 44] propose solutions for city-scale scene rendering using NeRF. [41, 45, 51, 57] targets a special kind of scenes with high specularities and reflections. Another important line of works [10, 14, 15, 28, 33, 42] focuses on acceleration due to the low training and rendering speed of NeRF using voxel grid or hash table. Although great progress has been made, NeRF-based methods still suffer from low rendering speed and high training-time memory usage due to their implicit nature [11]. Based on works on point-based neural rendering [2, 39, 50, 55], a recent milestone work [21] introduces anisotropic 3D Gaussian as an effective representation of the scene, and renders the image using a fast tile-based differentiable rasterizer, surpassing existing implicit neural representation methods in both quality and efficiency.", + "bbox": [ + 496, + 472, + 890, + 820 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Reflective Object Rendering", + "text_level": 1, + "bbox": [ + 500, + 830, + 754, + 848 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Rendering views of reflective objects from multi-view images has been a challenging task due to the complex light interactions. Previous approaches rely on simple light field", + "bbox": [ + 496, + 854, + 890, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5323", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/cf80fa45aca6f1450b54c5a979828213ba1f8a07fe344c948dbeff6ed4bb0504.jpg", + "image_caption": [ + "Figure 2. GaussianShader initiates with the neural 3D Gaussian spheres that integrate both conventional attributes and the newly introduced shading attributes to accurately capture view-dependent appearances. We incorporate a differentiable environment lighting map to simulate realistic lighting. The end-to-end training leads to a model that reconstructs both reflective and diffuse surfaces, achieving high material and lighting fidelity." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 893, + 228 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "interpolations to achieve high-fidelity rendering of novel perspectives [16, 25, 49], yet were constrained by the necessity for dense discrete captures. The accurate rendering of reflective surfaces hinges upon the precise estimation of scene illumination (e.g. environment light) and material properties (e.g. BRDF), which is the task of inverse rendering [4, 35]. Previous studies have demonstrated methodologies for predicting BRDF under known lighting conditions [1, 13, 31], as well as techniques for estimating lighting given known geometry [24, 37, 38]. Furthermore, [12, 19, 51, 58] incorporate indirect illumination, thereby enhancing the fidelity of the estimated BRDF.", + "bbox": [ + 75, + 300, + 468, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Some NeRF related works attempt to model reflectance by disentangling the visual appearance into lighting and material properties, such as [7-9, 41, 54, 57], which can jointly predict environmental illumination and surface reflectance properties under unknown or varying lighting conditions. Ref-NeRF [45] introduces a new parameterization and structuring of view-dependent outgoing radiance, as well as a regularizer on normal vectors. Recent works [26, 27, 29] utilize SDF-based representation to learn geometry from high specular surfaces, obtaining more accurate normals for physically based rendering. However, these method suffer from extremely time-consuming optimization and slow rendering speed, which hinders their employments in real applications.", + "bbox": [ + 75, + 482, + 470, + 693 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 702, + 218, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3.1 3D Gaussian Splitting Rasterization", + "text_level": 1, + "bbox": [ + 76, + 724, + 383, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our method builds upon Gaussian Splatting [21], which begins with a collection of images capturing a static scene, their corresponding camera parameters, and a sparse point cloud generated through Structure-from-Motion (SfM) [40]. These points construct a set of Gaussians, each defined by position (mean) $\\mathbf{p}$ and a 3D covariance matrix $\\Sigma$ . While a direct optimization of the covariance matrix $\\Sigma$ might seem intuitive, it presents challenges due to the requirement of positive semi-definiteness. As an alternative, in Gaussian Splatting, a more intuitive yet", + "bbox": [ + 75, + 750, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "equally expressive representation is adopted as an ellipsoid for optimization. The ellipsoid is separated into scaling and rotation components, represented by a scaling matrix $S$ and a rotation matrix $R$ , i.e., $\\Sigma = RSS^T R^T$ .", + "bbox": [ + 498, + 300, + 893, + 361 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3.2 Rendering with 3D Gaussians", + "text_level": 1, + "bbox": [ + 500, + 383, + 761, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "By projecting to 2D frame space according to the camera parameters, 3D Gaussian spheres are both differentiable and amenable to rapid rendering through 2D splatting with $\\alpha$ -blending. For radiance field modeling, the directional appearance aspect (color) $\\mathbf{c}$ is conveyed through spherical harmonics (SH). By tile-based rasterization, we could sum up the pixel color $\\mathbf{C}$ after sorting $\\mathbf{c}_i$ by depth:", + "bbox": [ + 498, + 409, + 893, + 516 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {C} = \\sum_ {i \\in N} \\mathbf {c} _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 527, + 890, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha_{i}$ is obtained by multiplying Gaussian weight with opacity $\\alpha$ associated to Gaussian sphere. After rasterization, the color loss could be applied to the rendered image:", + "bbox": [ + 498, + 583, + 890, + 628 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c o l o r}} = \\left| \\left| \\mathbf {C} - \\mathbf {C} _ {\\mathrm {g t}} \\right| \\right| ^ {2} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 643, + 890, + 664 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 676, + 589, + 693 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overview of our method is depicted in Fig. 2. Our approach begins by adopting 3D Gaussian spheres, encompassing shape attributes including covariance $\\Sigma$ , opacity $\\alpha$ , and position $\\mathbf{p}$ . To enhance the representation ability on reflections, we compute the appearances of these Gaussian spheres with a shading function, which requires a set of shading attributes, including diffuse color, roughness, specular tint, normal, and residual color, as detailed in Sec. 3.1. Subsequently, we employ a differentiable environment light map to model the direct lighting, as elaborated in Sec. 3.2. The shading process relies significantly on accurate normal estimation, as discussed in Sec. 3.3. Finally, we introduce the losses used in the whole training process in Sec. 3.4.", + "bbox": [ + 496, + 703, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5324", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Shading on 3D Gaussians", + "text_level": 1, + "bbox": [ + 76, + 90, + 308, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Gaussian Splatting [21] models the appearances of Gaussians with simple spherical harmonic functions without considering the light-surface interactions. Thus, Gaussian Splatting fails to accurately represent strong specular surfaces. However, accurately considering the light-surface interactions requires an exact evaluation of the Rendering Equation [20], which requires extensive computational time and complex BRDF parameters. We adopt a simplified approximation of the rendering equation which enables us to achieve high-quality rendering results on reflective surfaces in a considerably shorter time.", + "bbox": [ + 75, + 113, + 468, + 279 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, for a Gaussian sphere, its rendered color $\\mathbf{c}$ for the viewing direction $\\omega_{o}$ is computed by", + "bbox": [ + 76, + 280, + 468, + 311 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} \\left(\\omega_ {o}\\right) = \\gamma \\left(\\mathbf {c} _ {d} + \\mathbf {s} \\odot L _ {s} \\left(\\omega_ {o}, \\mathbf {n}, \\rho\\right) + \\mathbf {c} _ {r} \\left(\\omega_ {o}\\right)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 321, + 468, + 339 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\gamma$ is a gamma tone mapping function [3], $\\mathbf{c}_d\\in [0,1]^3$ is the diffuse color of this Gaussian sphere, $\\mathbf{s}\\in [0,1]^3$ is the specular tint defined on this sphere, $L_{s}(\\omega_{o},\\mathbf{n},\\rho)$ is the direct specular light for this sphere in this direction, $\\mathbf{n}$ is the normal of this Gaussian sphere, $\\rho \\in [0,1]$ is the roughness of the sphere, $\\mathbf{c}_r:\\mathbb{R}^3\\to \\mathbb{R}^3$ is so-called residual colors, and $\\odot$ is the element-wise multiplication.", + "bbox": [ + 75, + 348, + 468, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Explanations on Eq. 3. We explain our motivation of this shading model in the following three aspects. a) Diffuse color $\\mathbf{c}_d$ represents the consistent colors of this Gaussian sphere, which do not change with viewing directions. b) $\\mathbf{s} \\odot L_s(\\omega_o, \\mathbf{n}, \\rho)$ describes the interactions between the surface intrinsic color $\\mathbf{s}$ and the direct specular light $L_s$ . This term enables us to represent most of the reflections in rendering. c) Since there are still some reflections that cannot be explained by the above reflections of direct lights, such as scattering and reflection on indirect lights, we add a residual color term $\\mathbf{c}_r(\\omega_o)$ to account for these complex reflections. In comparison, Ref-NeRF [45] adopts a similar shading function without such a residual color term, which makes it struggle to handle advanced complex reflections. $\\mathbf{c}_r(\\omega_o)$ is parameterized by spherical harmonic functions. d) $\\mathbf{c}_d$ , $\\mathbf{s}$ , $\\rho$ , and the coefficients of the spherical harmonics in $\\mathbf{c}_r(\\omega_o)$ all are trainable parameters associated with this Gaussian sphere. In the following, we will introduce how to compute $L_s(\\omega_o, \\mathbf{n}, \\rho)$ , and $\\mathbf{n}$ .", + "bbox": [ + 75, + 454, + 470, + 743 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Specular Light", + "text_level": 1, + "bbox": [ + 76, + 750, + 228, + 767 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We compute the specular light $L_{s}$ by integrating incoming radiance with the specular GGX [46] Normal Distribution Function $D$ visualized in Fig. 3", + "bbox": [ + 75, + 773, + 468, + 820 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {s} (\\omega_ {o}, \\mathbf {n}, \\rho) = \\int_ {\\Omega} L (\\omega_ {i}) D (\\mathbf {r}, \\rho) (\\omega_ {i} \\cdot \\mathbf {n}) d \\omega_ {i}, \\qquad (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 829, + 468, + 861 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\Omega$ represents the whole upper semi-sphere, $\\omega_{i}$ is the direction for the input radiance and $D$ characterizes the", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e9c77fddb150afd8d57161c042d8e52ccd0296a8e4962286773e0228e8ef8f00.jpg", + "image_caption": [ + "Figure 3. Normal Distribution Function D in Eq. 4 is determined by the roughness $\\rho$ and reflective direction $\\mathbf{r}$ . Surface with small $\\rho$ has a smaller specular lobe and that with large $\\rho$ has a larger specular lobe." + ], + "image_footnote": [], + "bbox": [ + 558, + 95, + 823, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "specular lobe (effective integral range). When the surface is rough, the specular lobe will be larger around the reflective direction $\\mathbf{r}$ while if the surface is smooth, the specular lobe will be smaller. The reflection direction $\\mathbf{r}$ is calculated by view direction $\\omega_{o}$ and normal $\\mathbf{n}$ using $\\mathbf{r} = 2(\\omega_{o} \\cdot \\mathbf{n})\\mathbf{n} - \\omega_{o}$ . In our approach, environment light $L(\\omega_{i})$ is represented by a trainable $6 \\times 64 \\times 64$ cube map. All pixel values in our lighting maps are uniformly initialized to 0.8 (within a range of 0 to 1). Note that the lighting map is global, with each scene being optimized using a single lighting map.", + "bbox": [ + 496, + 253, + 890, + 405 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The light modeling is based on NVDiffRec [34]. Specifically, we begin by preprocessing the environment lighting map into multiple mipmap levels, each storing the integrated light across different reflective directions and roughness. When calculating light integration, we interpolate between these mipmaps based on the input of specific reflection directions and roughness, to obtain the required light integrals. This mipmap-based framework, in contrast to Ref-NeRF's [45] approach using integrated directional encoding, enhances training efficiency. The reflection direction $r$ is calculated by combining the observation direction $\\omega_{o}$ with the normal $n$ .", + "bbox": [ + 496, + 407, + 892, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Normal Estimation", + "text_level": 1, + "bbox": [ + 500, + 602, + 684, + 617 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Normal estimation on Gaussian spheres is difficult. Gaussian spheres are a collection of discrete entities, each representing a localized point in space without a continuous surface or defined edge. This discrete structure makes it inherently difficult to directly calculate normals, which typically requires a continuous surface. In the following, we observed that the shortest axis direction of a Gaussian can serve as an approximated normal and we further associate a predicted normal residual on it.", + "bbox": [ + 496, + 627, + 890, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Shortest axis direction. In experiments, we made an interesting observation that the aspect ratio of the 3D Gaussian sphere—specifically, the ratios of their longest, intermediate, and shortest axis—gradually increased during optimization, as shown in Fig. 4. This observation suggests that Gaussian spheres gradually become more flattened and approach a planar shape. This observation inspires us to select the shortest axis as the normal of this \"planarized\" Gaussian sphere, denoted by $\\mathbf{v}$ .", + "bbox": [ + 496, + 763, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5325", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1d0aa677c70ebf18f9b0caca1a4919505c99916197b23b36f0bb357cf77b1bb3.jpg", + "image_caption": [ + "Figure 4. The geometric evolving process of a 3D Gaussian sphere in the optimization, which gradually becomes planar." + ], + "image_footnote": [], + "bbox": [ + 102, + 89, + 452, + 195 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Predicted normal residual. The shortest axis $\\mathbf{v}$ only serves as an approximated normal. To make the normal computation more accurate, we further introduce a trainable normal residual $\\Delta \\mathbf{n}$ on every Gaussian sphere. However, the orientation of the shortest axis $\\mathbf{v}$ has an ambiguity because the direction of the shortest axis could either point outward or inward from the surface. To handle this ambiguity, we optimize two separate normal residuals to accommodate both scenarios. Given a specific viewing direction $\\omega_{o}$ , we first select the direction aligned with the viewing direction $\\omega_{\\mathrm{o}}$ as the active normal direction for this viewing direction and then apply the corresponding normal residual to the active normal. This process is described by", + "bbox": [ + 75, + 239, + 472, + 436 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {n} = \\left\\{ \\begin{array}{l l} \\mathbf {v} + \\Delta \\mathbf {n} _ {1} & \\text {i f} \\omega_ {\\mathbf {o}} \\cdot \\mathbf {v} > 0, \\\\ - (\\mathbf {v} + \\Delta \\mathbf {n} _ {2}) & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 443, + 468, + 483 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To prevent the normal residual from deviating too much from the shortest axis, we add a penalty towards normal residual, making sure it is small enough.", + "bbox": [ + 76, + 489, + 470, + 535 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e g}} = \\left\\| \\Delta \\mathbf {n} \\right\\| ^ {2} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 542, + 468, + 560 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The normal residual is shown on the left of Fig. 5.", + "bbox": [ + 76, + 568, + 406, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Normal-geometry consistency. The above shortest axis direction and normal residuals are defined on each Gaussian sphere separately. However, a noticeable problem is that a normal reveals the gradient of the local geometry, which is supposed to be associated with all the Gaussian spheres in a local region. We find that simply applying a color loss to train the aforementioned normal residuals leads to inconsistency between the local geometry and the estimated normals. The main reason is that every Gaussian sphere learns its normal residuals separately without knowing the local geometry formulated by its neighbor Gaussian spheres. Thus, we have to correlate multiple Gaussian spheres in a local region with their normals to ensure normal-geometry consistency. A straightforward and naive solution is to search for its $K$ neighborhoods in space and estimate a coarse normal from all the neighboring spheres. However, such KNN searching would be extremely expensive during training because all the Gaussian spheres are dynamically moving in the optimization process. Instead, we propose a simple yet effective way to ensure normal-geometry consistency as follows.", + "bbox": [ + 75, + 584, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e13ffe9690f31500f8555ac92c41357f6d5f57dcc1bcf22260ee48d6f49ad889.jpg", + "image_caption": [ + "Figure 5. Visualization of the relationship between shortes axis $\\mathbf{v}$ , normal residual $\\Delta \\mathbf{n}$ , normal $\\mathbf{n}$ and depth-gradient normal $\\hat{\\mathbf{n}}$ . The supervision $\\mathcal{L}_{normal}$ enforces the normal-geometry consistency." + ], + "image_footnote": [], + "bbox": [ + 500, + 87, + 893, + 148 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We associate the local geometry with predicted normals by minimizing the difference between the grad normals derived from the rendered depth map and the rendered normal maps using the predicted normals", + "bbox": [ + 496, + 208, + 893, + 268 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {n o r m a l}} = \\left| \\left| \\bar {\\mathbf {n}} - \\hat {\\mathbf {n}} \\right| \\right| ^ {2}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 277, + 890, + 296 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\bar{\\mathbf{n}}$ is the rendered normal map and $\\hat{\\mathbf{n}}$ is computed by applying the Sobel-like operator on the rendered depth map. $\\hat{\\mathbf{n}}$ reveals the local geometry formulated by multiple Gaussian spheres because it is computed from the rendered depth maps. $\\bar{\\mathbf{n}}$ contains the information from the separately-defined normal on each Gaussian sphere. By minimizing the difference, we enforce the consistency between the local geometry and the estimated normals.", + "bbox": [ + 496, + 306, + 892, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Losses", + "text_level": 1, + "bbox": [ + 500, + 436, + 586, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition to $\\mathcal{L}_{\\mathrm{color}}$ , $\\mathcal{L}_{reg}$ , and $\\mathcal{L}_{\\mathrm{normal}}$ , we make use of sparse loss [30, 50] to encourages Gaussian spheres' opacity values $\\alpha$ to approach either 0 or 1 by", + "bbox": [ + 498, + 459, + 893, + 503 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s p a r s e}} = \\frac {1}{| \\alpha |} \\sum_ {\\alpha_ {i}} \\left[ \\log \\left(\\alpha_ {i}\\right) + \\log \\left(1 - \\alpha_ {i}\\right) \\right]. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 513, + 890, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This sparsity loss helps the geometry of Gaussian spheres converge to a single thin plate and improves the rendering quality. In summary, the total training loss $\\mathcal{L}$ is", + "bbox": [ + 496, + 558, + 890, + 604 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {c o l o r}} + \\lambda_ {\\mathrm {n}} \\mathcal {L} _ {\\text {n o r m a l}} + \\lambda_ {\\mathrm {s}} \\mathcal {L} _ {\\text {s p a r s e}} + \\lambda_ {\\mathrm {r}} \\mathcal {L} _ {\\text {r e g}} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 614, + 890, + 631 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{\\mathrm{n}} = 0.01, \\lambda_{\\mathrm{s}} = 0.001, \\lambda_{\\mathrm{r}} = 0.001$", + "bbox": [ + 500, + 641, + 779, + 656 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 669, + 632, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 500, + 694, + 601, + 709 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To comprehensively validate the effectiveness of GaussianShader, we conduct evaluation on various datasets: a) widely-used NVS dataset: NeRF Synthetic [32]. b) reflective objects datasets: Shiny Blender [45] and Glossy Synthetic [29]. c) real-world large-scale scenes: Tanks and Temples [23].", + "bbox": [ + 496, + 717, + 893, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Baselines and Metrics", + "text_level": 1, + "bbox": [ + 500, + 816, + 704, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compare our method against the following baselines: a) 3D Gaussian Splatting [21]: a real-time radiance field rendering method based on efficient 3D Gaussian representation; b) VolSDF [52]: a classical neural implicit surface", + "bbox": [ + 496, + 840, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5326", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/384b6f476321c46466d38cb76fbb1f655326cdc7c3ea509ac7e2645470fb8c7e.jpg", + "table_caption": [ + "Table 1. The quantitative comparisons (PSNR / SSIM / LPIPS) on NeRF Synthetic dataset [32]." + ], + "table_footnote": [], + "table_body": "
ChairDrumsNeRF Synthetic [32]
LegoMicMaterialsShipHotdogFicusAvg.
PSNR↑
NeRF [32]33.0025.0132.5432.9129.6228.6536.1830.1331.01
VolSDF [52]30.5720.4329.4630.5329.1325.5135.1122.9127.96
Ref-NeRF [45]33.9825.4335.1033.6527.1029.2437.0428.7431.29
ENVIDR [27]31.2222.9929.5532.1729.5221.5731.4426.6028.13
Gaussian Splitting [21]35.8226.1735.6935.3430.0030.8737.6734.8333.30
Ours35.8326.3635.8735.2330.0730.8237.8534.9733.38
SSIM↑
NeRF [32]0.9670.9250.9610.9800.9490.8560.9740.9640.947
VolSDF [52]0.9490.8930.9510.9690.9540.8420.9720.9290.932
Ref-NeRF [45]0.9740.9290.9750.9830.9210.8640.9790.9540.947
ENVIDR [27]0.9760.9300.9610.9840.9680.8550.9630.9870.956
Gaussian Splitting [21]0.9870.9540.9830.9910.9600.9070.9850.9870.969
Ours0.9870.9490.9830.9910.9600.9050.9850.9850.968
LPIPS↓
NeRF [32]0.0460.0910.0500.0280.0630.2060.1210.0440.081
VolSDF [52]0.0560.1190.0540.1910.0480.1910.0430.0680.096
Ref-NeRF [45]0.0290.0730.0250.0180.0780.1580.0280.0560.058
ENVIDR [27]0.0310.0800.0540.0210.0450.2280.0720.0100.067
Gaussian Splitting [21]0.0120.0370.0160.0060.0340.1060.0200.0120.030
Ours0.0120.0400.0140.0060.0330.0980.0190.0130.029
", + "bbox": [ + 181, + 104, + 789, + 406 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e92b80a024832da86c0d9ef310a11b1dae0625e102480f157df6397aea4aa98c.jpg", + "image_caption": [ + "Figure 6. The qualitative comparisons with 3D Gaussian Splitting [21] on Glossy dataset [29]. Our method not only renders the objects with high fidelity but also provides detailed normal and lighting maps. Some areas are zoomed in for better visualization." + ], + "image_footnote": [], + "bbox": [ + 76, + 417, + 477, + 628 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/61a2d2e5f090b2e4d4c12f64b600f566adc3a190f10162b2ecb79458930e0007.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 480, + 417, + 890, + 627 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "reconstruction method based on SDF; c) Ref-NeRF [45]: state-of-the-art method in novel view synthesis of reflective objects; d) NVDiffRec [34], NVDiffRecMC [18]: top-performing neural inverse rendering methods; e) ENVIDR [27], NeRO [29]: top-performing SDF-based neural implicit methods for reconstructing reflective objects. The evaluation metrics to measure rendering quality are reported in PSNR, SSIM [48], and LPIPS [56].", + "bbox": [ + 75, + 685, + 468, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 816, + 294, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "All experiments are conducted on a Nvidia RTX 3090 graphics card. We optimize our models using Adam [22] optimizer for 30,000 iterations. For Ref-NeRF [45], ENVIDR [27], and 3D Gaussian Splitting [21], we retrain", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "them using their official codes keeping all the original settings (training iterations, sampled rays, etc.) and obtain results from the retrained models. Other works' results are imported from their original papers.", + "bbox": [ + 496, + 685, + 890, + 747 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Comparisons", + "text_level": 1, + "bbox": [ + 500, + 756, + 638, + 772 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "NeRF Synthetic dataset [32]. We first evaluate our model on the NeRF Synthetic dataset, which contains objects with complex geometry and realistic non-Lambertian materials. We show the quantitative results in Tab. 1 and visual comparisons in Fig. 8. Our approach achieves numerically and visually comparable results with both Gaussian Splatting and neural rendering methods, demonstrating the effectiveness of our method in rendering general objects.", + "bbox": [ + 496, + 779, + 892, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5327", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/185816575b056867e8c1cca272e0fd3ffa0097de86d7c017fa8034a35e2e3741.jpg", + "image_caption": [ + "Figure 7. Qualitative comparisons on Shiny Blender dataset [45]. Error maps corresponding to each rendering result demonstrate the enhanced fidelity of our approach in highly specular areas." + ], + "image_footnote": [], + "bbox": [ + 76, + 85, + 472, + 223 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Note that SDF-based ENVIDR [27] cannot perform well on shadowed regions of complex objects, as shown in Fig. 8.", + "bbox": [ + 75, + 284, + 468, + 315 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Shiny Blender dataset [45]. Quantitative results on the Shiny Blender dataset are reported in Tab. 2. With the ability to model light-surface interactions, our method outperforms original Gaussian Splatting in all scenes while slightly underperforms ENVIDR and Ref-NeRF. ENVIDR capitalizes on the continuous properties of Signed Distance Fields to naturally create smoother surfaces and normals, resulting in high rendering quality, particularly for reflective objects. Fig. 7 shows our comparisons with Gaussian Splatting. We can see that our method correctly renders surfaces with strong specular appearances. High-quality modeling of reflections is contingent upon the accurate estimation of normals as shown in Fig. 9.", + "bbox": [ + 75, + 316, + 468, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Additionally, we report the average PSNR, training time, and rendering FPS in Tab. 3. The reported results are averaged among all objects of the Shiny Blender and NeRF synthetic datasets. On the same hardware, our method only takes about 0.5 hour for training while MLP-based methods, like ENVIDR and Ref-NeRF, require 6 hours and 23 hours for optimizing. Due to the introduce of extra lighting and shading attributes, our method is a bit slower than 3D Gaussian Splatting [21] but still keeps reasonable efficiency and real-time rendering speed.", + "bbox": [ + 75, + 516, + 468, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tanks and Temples dataset [23]. To explore our method's scalability in larger-scale environments instead of only small objects, we conducted experiments using Tanks and Temples [23]. The qualitative results in Fig. 10 show that our model outperforms Gaussian Splatting [21].", + "bbox": [ + 75, + 670, + 468, + 746 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Glossy Synthetic dataset [29]. We present a comparative visualization between Gaussian Splatting and our method on the Glossy dataset shown in Fig. 6. Our method not only renders objects with exceptional fidelity but also provides more accurate normal and lightings, capturing the nuances of reflective surfaces with greater precision. These maps reveal our method's ability to approximate the rendering function, resulting in a more lifelike portrayal of specular lighting. Please refer to the supplementary materials provided for an extensive showcase of our results.", + "bbox": [ + 75, + 750, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c5759a7d083a879186c7a10cdb0ae320969e29386844b611739142ee8e4aba9d.jpg", + "table_caption": [ + "Table 2. Qualitative comparisons on Shiny Blender dataset [45]. Our method is comparable with both Gaussian Splatting [21] and prior reflective object reconstruction methods." + ], + "table_footnote": [], + "table_body": "
Shiny Blender [45]
CarBallHelmetTeapotToasterCoffeeAvg.
PSNR↑
NVDiffRec [34]27.9821.7726.9740.4424.3130.7428.70
NVDiffMC [18]25.9330.8526.2738.4422.1829.6028.88
Ref-NeRF [45]30.4129.1429.9245.1925.2933.9932.32
NeRO [29]25.5330.2629.2038.7026.4628.8929.84
ENVIDR [27]28.4638.8932.7341.5926.1129.4832.88
Gaussian Splitting [21]27.2427.6928.3245.6820.9932.3230.37
Ours27.9030.9828.3245.8626.2132.3931.94
SSIM↑
NVDiffRec [34]0.9630.8580.9510.9960.9280.9730.945
NVDiffMC [18]0.9400.9400.9400.9950.8860.9650.944
Ref-NeRF [45]0.9490.9560.9550.9950.9100.9720.956
NeRO [29]0.9490.9740.9710.9950.9290.9560.962
ENVIDR [27]0.9610.9910.9800.9960.9390.9490.969
Gaussian Splitting [21]0.9300.9370.9510.9960.8950.9710.947
Ours0.9310.9650.9500.9960.9290.9710.957
LPIPS↓
NVDiffRec [34]0.0450.2970.1180.0110.1690.0760.119
NVDiffMC [18]0.0770.3120.1570.0140.2250.0970.147
Ref-NeRF [45]0.0510.3070.0870.0130.1180.0820.109
NeRO [29]0.0740.0940.0500.0120.0890.1100.072
ENVIDR [27]0.0490.0670.0510.0110.1160.1390.072
Gaussian Splitting [21]0.0470.1610.0790.0070.1260.0780.083
Ours0.0450.1210.0760.0070.0790.0780.068
", + "bbox": [ + 503, + 135, + 890, + 395 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b839f48ab899f9aeebb5bb055629929c54f04f08f7fba0453f383ac31ccc5a1c.jpg", + "table_caption": [ + "Table 3. Comparisons over average PSNR, training time and FPS. Our method achieves the best rendering quality and competitive training/rendering speed." + ], + "table_footnote": [], + "table_body": "
PSNRTraining TimeFPS
Ref-NeRF [45]31.7323h0.03
ENVIDR [27]30.046h1.33
Gaussian Splitting [21]32.050.25h274
Ours32.760.58h97
", + "bbox": [ + 535, + 448, + 856, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 540, + 651, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we conduct ablation studies of our model on Shiny Blender dataset. The numerical results are reported in Tab. 4. All ablation experiments are done on Shiny Blender dataset with $1/2$ image resolution.", + "bbox": [ + 498, + 565, + 890, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Loss functions. We trained our model serperately under the setting \"without sparse loss $\\mathcal{L}_{\\mathrm{sparse}}$ and \"without normal loss $\\mathcal{L}_{\\mathrm{normal}}$ . The incorporation of these losses, especially the direct regularization on normals, proved critical, yielding a marked improvement in rendering quality by effectively guiding the normal optimization process as shown in Tab. 4.", + "bbox": [ + 496, + 626, + 890, + 731 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Residual color. Residual color, comprising 3rd order SH coefficients, is used to compensate for indirect light. As shown in Table 4, the introduce of residual color enhances our model, allowing it to simulate a wide array of real-world intricate lighting situations and significantly improve rendering quality.", + "bbox": [ + 496, + 733, + 890, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Normal formulation. In our full model, we use the shortest axis direction $\\mathbf{v}$ with a learnable residual $\\Delta \\mathbf{n}$ to formulate normal, as described by Eqn. 3.3. In Tab. 4, we compared this novel normal formation with a naive normal prediction manner (indicated by \"w/o axis $\\mathbf{v}$ ). Instead of", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5328", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5f0b70463b9d7846603de94af9f82cc29de25f203370cd7b39ad97fcf7a07858.jpg", + "image_caption": [ + "Figure 8. We present a qualitative comparisons of our method on NeRF Synthetic dataset [32] against previous techniques and corresponding ground truth images from test views. For clarity, we zoom in the images where differences in quality are especially notable." + ], + "image_footnote": [], + "bbox": [ + 109, + 85, + 861, + 290 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/baf25c508b4a5b28b3951f669dea5a80c2968e1692cf22e757ac083b655eda9e.jpg", + "image_caption": [ + "Figure 9. Visualized normal results on Shiny Blender dataset [45], compared with Gaussian Splatting [21]. It illustrates the superior normal estimation achieved by our method. Table 4. Ablation studies on loss function, lighting representation, normal axis and residual color using Shiny Blender dataset." + ], + "image_footnote": [], + "bbox": [ + 76, + 330, + 468, + 521 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b110e4a2b60fd33097c6520f18ea210532f3304f3650092be923fd231026c443.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PSNR↑SSIM↑LPIPS↓
w/o Lsparse31.790.9520.056
w/o Lnormal30.930.9410.060
w/o cr31.490.9480.060
w/o v31.470.9510.058
MLP Lighting29.730.9360.075
Full Model32.090.9530.054
", + "bbox": [ + 125, + 609, + 415, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "correlating with the geometry of 3D gaussians, we directly optimize a single normal attribute. The result shows our full model can achieve higher rendering quality, since our predicted normals are correlated with the true geometry of Gaussian spheres.", + "bbox": [ + 75, + 719, + 468, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Lighting representation. A crucial component of our approach is the modeling of lighting. \"MLP Lighting\" means we use a directional MLP similar to Ref-NeRF to implicitly optimize lighting. In our full model, we use \"Envmap Lighting\" representation as illustrated in Sec. 3.2. The result proves that our explicit lighting representation is better suited for 3D Gaussian rendering framework to cor", + "bbox": [ + 75, + 795, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a60dd242baef56b0f2a97f39aedd39f9bd9988886242d480f46e847e5bbe7046.jpg", + "image_caption": [ + "Figure 10. Qualitative comparisons with Gaussian Splitting [21] on the Tanks and Temples dataset [23]. Our approach yields a smoother and more reasonable normal." + ], + "image_footnote": [], + "bbox": [ + 496, + 330, + 890, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "rectly render reflective objects.", + "bbox": [ + 500, + 671, + 705, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 705, + 625, + 722 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In summary, we propose GaussianShader, advancing the rendering of both reflective and general objects through an extended 3D Gaussian model. Specifically, our method integrates shading functions with 3D Gaussian representation for handling view-dependent appearances, successfully handling reflective surfaces. Moreover, a novel normal prediction method is proposed to achieve high-quality normals, which correlates the shading parameters and true geometry of 3D Gaussians. The results show our approach achieves high-quality renderings, marking a substantial improvement in efficient, realistic 3D object rendering.", + "bbox": [ + 496, + 734, + 893, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5329", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Miika Aittala, Timo Aila, and Jaakko Lehtinen. Reflectance modeling by neural texture synthesis. ACM Transactions on Graphics (ToG), 35(4):1-13, 2016. 3", + "[2] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 696-712. Springer, 2020. 2", + "[3] Matthew Anderson, Ricardo Motta, Srinivasan Chandrasekar, and Michael Stokes. Proposal for a standard default color space for the internet—srgb. In Color and imaging conference, pages 238–245. Society for Imaging Science and Technology, 1996. 4", + "[4] Jonathan T Barron and Jitendra Malik. Shape, illumination, and reflectance from shading. IEEE transactions on pattern analysis and machine intelligence, 37(8):1670-1687, 2014. 3", + "[5] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 2", + "[6] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470-5479, 2022. 1, 2", + "[7] Sai Bi, Zexiang Xu, Pratul Srinivasan, Ben Mildenhall, Kalyan Sunkavalli, Milos Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Neural reflectance fields for appearance acquisition. arXiv preprint arXiv:2008.03824, 2020. 3", + "[8] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T Barron, Ce Liu, and Hendrik Lensch. Nerd: Neural reflectance decomposition from image collections. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12684-12694, 2021.", + "[9] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan Barron, and Hendrik Lensch. Neural-pil: Neural pre-integrated lighting for reflectance decomposition. Advances in Neural Information Processing Systems, 34: 10691-10704, 2021. 3", + "[10] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 2", + "[11] Zilong Chen, Feng Wang, and Huaping Liu. Text-to-3d using gaussian splatting. arXiv preprint arXiv:2309.16585, 2023. 2", + "[12] Youming Deng, Xueting Li, Sifei Liu, and Ming-Hsuan Yang. Dip: Differentiable interreflection-aware physics-based inverse rendering. arXiv preprint arXiv:2212.04705, 2022.3", + "[13] Valentin Deschaintre, Miika Aittala, Fredo Durand, George Drettakis, and Adrien Bousseau. Single-image svbrdf cap" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ture with a rendering-aware deep network. ACM Transactions on Graphics (ToG), 37(4):1-15, 2018. 3", + "[14] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 2", + "[15] Stephan J Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14346-14355, 2021. 2", + "[16] Steven J Gortler, Radek Grzesczuk, Richard Szeliski, and Michael F Cohen. The lumigraph. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 453-464. 2023. 3", + "[17] Haoyu Guo, Sida Peng, Haotong Lin, Qianqian Wang, Guofeng Zhang, Hujun Bao, and Xiaowei Zhou. Neural 3d scene reconstruction with the manhattan-world assumption. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5511-5520, 2022. 2", + "[18] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, Light, and Material Decomposition from Images using Monte Carlo Rendering and Denoising. arXiv:2206.03380, 2022. 6, 7", + "[19] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, light, and material decomposition from images using monte carlo rendering and denoising. Advances in Neural Information Processing Systems, 35:22856-22869, 2022. 3", + "[20] James T Kajiya. The rendering equation. In Proceedings of the 13th annual conference on Computer graphics and interactive techniques, pages 143-150, 1986. 4", + "[21] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 42(4):1-14, 2023. 1, 2, 3, 4, 5, 6, 7, 8", + "[22] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[23] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36 (4):1-13, 2017. 5, 7, 8, 1", + "[24] Chloe LeGendre, Wan-Chun Ma, Graham Fyffe, John Flynn, Laurent Charbonnel, Jay Busch, and Paul Debevec. Deeplight: Learning illumination for unconstrained mobile mixed reality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5918-5928, 2019. 3", + "[25] Marc Levoy and Pat Hanrahan. Light field rendering. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 441-452. 2023. 3", + "[26] Ruofan Liang, Jiahao Zhang, Haoda Li, Chen Yang, Yushi Guan, and Nandita Vijaykumar. Spidr: Sdf-based neural point fields for illumination and deformation. arXiv preprint arXiv:2210.08398, 2022. 3" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5330", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Ruofan Liang, Huiting Chen, Chunlin Li, Fan Chen, Selvakumar Panneer, and Nandita Vijaykumar. Envidr: Implicit differentiable renderer with neural environment lighting. arXiv preprint arXiv:2303.13022, 2023. 1, 2, 3, 6, 7", + "[28] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 2", + "[29] Yuan Liu, Peng Wang, Cheng Lin, Xiaoxiao Long, Jiepeng Wang, Lingjie Liu, Taku Komura, and Wenping Wang. Nano: Neural geometry and brdf reconstruction of reflective objects from multiview images. arXiv preprint arXiv:2305.17398, 2023. 3, 5, 6, 7, 1", + "[30] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural volumes: Learning dynamic renderable volumes from images. arXiv preprint arXiv:1906.07751, 2019. 5", + "[31] Wojciech Matusik. A data-driven reflectance model. PhD thesis, Massachusetts Institute of Technology, 2003. 3", + "[32] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 5, 6, 8", + "[33] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 2", + "[34] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3d models, materials, and lighting from images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8280-8290, 2022. 4, 6, 7", + "[35] Merlin Nimier-David, Delio Vicini, Tizian Zeltner, and Wenzel Jakob. Mitsuba 2: A retargetable forward and inverse renderer. ACM Transactions on Graphics (TOG), 38(6):1-17, 2019. 3", + "[36] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5589-5599, 2021. 2", + "[37] Jeong Joon Park, Aleksander Holynski, and Steven M Seitz. Seeing the world in a bag of chips. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1417-1427, 2020. 3", + "[38] Thomas Richter-Trummer, Denis Kalkofen, Jinwoo Park, and Dieter Schmalstieg. Instant mixed reality lighting from casual scanning. In 2016 IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 27-36. IEEE, 2016. 3", + "[39] Darius Rückert, Linus Franke, and Marc Stamminger. Adop: Approximate differentiable one-pixel point rendering. ACM Transactions on Graphics (ToG), 41(4):1-14, 2022. 2", + "[40] Noah Snavely, Steven M Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. In ACM siggraph 2006 papers, pages 835-846. 2006. 3" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Pratul P Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T Barron. Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7495-7504, 2021. 2, 3", + "[42] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 2", + "[43] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8248–8258, 2022. 2", + "[44] Haithem Turki, Deva Ramanan, and Mahadev Satyanarayanan. Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12922–12931, 2022. 2", + "[45] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481-5490. IEEE, 2022. 1, 2, 3, 4, 5, 6, 7, 8", + "[46] Bruce Walter, Stephen R Marschner, Hongsong Li, and Kenneth E Torrance. Microfacet models for refraction through rough surfaces. In Proceedings of the 18th Eurographics conference on Rendering Techniques, pages 195-206, 2007. 4", + "[47] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 2", + "[48] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6", + "[49] Daniel N Wood, Daniel I Azuma, Ken Aldinger, Brian Curless, Tom Duchamp, David H Salesin, and Werner Stuetzle. Surface light fields for 3d photography. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 487-496. 2023. 3", + "[50] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5438–5448, 2022. 2, 5", + "[51] Yao Yao, Jingyang Zhang, Jingbo Liu, Yihang Qu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Neilf: Neural incident light field for physically-based material estimation. In European Conference on Computer Vision, pages 700-716. Springer, 2022. 2, 3", + "[52] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neu" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5331", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ral Information Processing Systems, 34:4805-4815, 2021. 2, 5, 6", + "[53] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in neural information processing systems, 35:25018-25032, 2022. 2", + "[54] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snavely. Physg: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5453-5462, 2021. 3", + "[55] Qiang Zhang, Seung-Hwan Baek, Szymon Rusinkiewicz, and Felix Heide. Differentiable point-based radiance fields for efficient view synthesis. In SIGGRAPH Asia 2022 Conference Papers, pages 1-12, 2022. 2", + "[56] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6", + "[57] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Ner-factor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (ToG), 40(6):1-18, 2021. 2, 3", + "[58] Yuanqing Zhang, Jiaming Sun, Xingyi He, Huan Fu, Rongfei Jia, and Xiaowei Zhou. Modeling indirect illumination for inverse rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18643-18652, 2022. 3" + ], + "bbox": [ + 78, + 90, + 468, + 527 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "5332", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/GaussianShader_ 3D Gaussian Splatting with Shading Functions for Reflective Surfaces/e3ac7488-6672-4bab-abdc-38e188c8adc6_model.json b/2024/GaussianShader_ 3D Gaussian Splatting with Shading Functions for Reflective Surfaces/e3ac7488-6672-4bab-abdc-38e188c8adc6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9ae33f6409a3c1236ecc1f1bc0ba219b7b837666 --- /dev/null +++ b/2024/GaussianShader_ 3D Gaussian Splatting with Shading Functions for Reflective Surfaces/e3ac7488-6672-4bab-abdc-38e188c8adc6_model.json @@ -0,0 +1,2400 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.16, + 0.13, + 0.812, + 0.175 + ], + "angle": 0, + "content": "GaussianShader: 3D Gaussian Splating with Shading Functions for Reflective Surfaces" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.203, + 0.732, + 0.221 + ], + "angle": 0, + "content": "Yingwenqi Jiang\\(^{1}\\) Jiadong Tu\\(^{1}\\) Yuan Liu\\(^{2}\\) Xifeng Gao\\(^{3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.286, + 0.221, + 0.706, + 0.241 + ], + "angle": 0, + "content": "Xiaoxiao Long\\(^{2,*}\\) Wenping Wang\\(^{4}\\) Yuexin Ma\\(^{1,*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.108, + 0.244, + 0.285, + 0.262 + ], + "angle": 0, + "content": "\\(^{1}\\)ShanghaiTech University" + }, + { + "type": "text", + "bbox": [ + 0.305, + 0.244, + 0.511, + 0.262 + ], + "angle": 0, + "content": "2The University of Hong Kong" + }, + { + "type": "text", + "bbox": [ + 0.53, + 0.244, + 0.655, + 0.261 + ], + "angle": 0, + "content": "3Tencent America" + }, + { + "type": "text", + "bbox": [ + 0.674, + 0.244, + 0.838, + 0.262 + ], + "angle": 0, + "content": "Texas A&M University" + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.306, + 0.191, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.375, + 0.163, + 0.383 + ], + "angle": 0, + "content": "PSNR:21.69" + }, + { + "type": "image", + "bbox": [ + 0.197, + 0.307, + 0.309, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.222, + 0.375, + 0.28, + 0.383 + ], + "angle": 0, + "content": "PSNR: 28.87" + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.308, + 0.427, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.339, + 0.375, + 0.398, + 0.383 + ], + "angle": 0, + "content": "PSNR:32.67" + }, + { + "type": "image", + "bbox": [ + 0.432, + 0.308, + 0.545, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.457, + 0.375, + 0.515, + 0.383 + ], + "angle": 0, + "content": "PSNR:32.46" + }, + { + "type": "image", + "bbox": [ + 0.55, + 0.308, + 0.662, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.386, + 0.639, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.571, + 0.482, + 0.636, + 0.492 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.317, + 0.892, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.712, + 0.488, + 0.868, + 0.509 + ], + "angle": 0, + "content": "(b) Comparisons between FPS, training time and accuracy" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.51, + 0.893, + 0.553 + ], + "angle": 0, + "content": "Figure 1. GaussianShader maintains real-time rendering speed and renders high-fidelity images for both general and reflective surfaces. Ref-NeRF[45] and ENVIDR[27] attempt to handle reflective surfaces, but they suffer from quite time-consuming optimization and slow rendering speed. 3D Gaussian splattering [21] keeps high efficiency but cannot handle such reflective surfaces." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.562, + 0.314, + 0.577 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.579, + 0.47, + 0.822 + ], + "angle": 0, + "content": "The advent of neural 3D Gaussians [21] has recently brought about a revolution in the field of neural rendering, facilitating the generation of high-quality renderings at real-time speeds. However, the explicit and discrete representation encounters challenges when applied to scenes featuring reflective surfaces. In this paper, we present GaussianShader, a novel method that applies a simplified shading function on 3D Gaussians to enhance the neural rendering in scenes with reflective surfaces while preserving the training and rendering efficiency. The main challenge in applying the shading function lies in the accurate normal estimation on discrete 3D Gaussians. Specifically, we proposed a novel normal estimation framework based on the shortest axis directions of 3D Gaussians with a delicately designed loss to make the consistency between the normals and the geometries of Gaussian spheres. Exper" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.829, + 0.47, + 0.903 + ], + "angle": 0, + "content": "*Corresponding author. This work was supported by NSFC (No.62206173), MoE Key Laboratory of Intelligent Perception and Human-Machine Collaboration (ShanghaiTech University), Shanghai Frontiers Science Center of Human-centered Artificial Intelligence, the Innovation and Technology Commission of the HKSAR Government under the InnoHK initiative, and Ref. T45-205/21-N of Hong Kong RGC." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.563, + 0.893, + 0.683 + ], + "angle": 0, + "content": "periments show that GaussianShader strikes a commendable balance between efficiency and visual quality. Our method surpasses Gaussian Splatting [21] in PSNR on specular object datasets, exhibiting an improvement of 1.57dB. When compared to prior works handling reflective surfaces, such as Ref-NeRF [45], our optimization time is significantly accelerated (23h vs. 0.58h). Please click on our project website to see more results." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.709, + 0.633, + 0.724 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.736, + 0.893, + 0.903 + ], + "angle": 0, + "content": "In recent years, the field of 3D computer vision has witnessed remarkable advancements in the 3D reconstruction and visualization of 3D scenes. Innovations such as Neural Radiance Fields (NeRF) [32] have achieved substantial breakthroughs in generating novel views of 3D objects and scenes, presenting the potential for high-quality and photorealistic renderings. Despite these advancements, NeRF-related methodologies [6, 27, 45], still contend with challenges such as computationally expensive optimization and slow rendering speed. These limitations restrict their application in real-time interactive scenarios." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5322" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.288 + ], + "angle": 0, + "content": "More recently, 3D Gaussian Splatting [21] combines 3D Gaussian representation and tile-based splatting techniques to achieve high-quality 3D scene modeling and real-time rendering, making it possible to employ neural rendering techniques in real applications. However, it suffers from a performance drop on scenes featuring specular and reflective surfaces. This is because 3D Gaussian Splatting [21] does not explicitly model appearance properties, so that fails to capture significant view-dependent changes, particularly specular highlights. This constraint presents a substantial obstacle in the pursuit of achieving photorealistic rendering across a diverse array of materials, particularly those characterized by prominent reflective attributes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.291, + 0.473, + 0.473 + ], + "angle": 0, + "content": "Accurately modeling reflective surfaces is a challenging task. Ref-NeRF [45] and ENVIDR [27] combine the shading functions in implicit representations and present promising quality on reflective surfaces. However, they suffer from time-consuming optimization (hours) and slow rendering speed. Due to the limited flexibility of SDF, ENVIDR [27] even fails to model complex scenes and presents a significant performance drop on general objects. It is still an unexplored problem how to combine the shading functions in a 3D Gaussian Splatting framework to improve its ability to handle reflections while preserving the efficiency in training and rendering." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.476, + 0.473, + 0.688 + ], + "angle": 0, + "content": "In this paper, we present GaussianShader, a novel method that enhances the neural rendering of 3D Gaussians within scenes that contain reflective surfaces by incorporating a shading function on 3D Gaussians. To ensure the efficiency of GaussianShader, evaluating the shading function cannot be too expensive while still retaining the ability to model the reflections. In light of this, we propose a novel simplified shading function that considers the diffuse colors and the direct reflections while putting all advanced complex reflections into a residual color term. In comparison with the shading function of Ref-NeRF which can only consider direct reflections, the utilization of this residual color enables GaussianShader to render more complex reflective appearances with efficiency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Another challenge in computing a shading function is how to predict accurate normals on the discrete 3D Gaussian spheres. First, it is hard to get a locally-continuous surface from the 3D Gaussians to compute the surface normals. Second, associating multiple 3D Gaussians for normal computation would be very expensive using neighborhood searching. In GaussianShader, we address this problem by introducing a new normal representation, which is based on the shortest axis direction of a Gaussian sphere and learns a normal residual on this axis direction. Then, to enforce consistency between the estimated normals and the geometry formulated by Gaussian spheres, we introduce an efficient normal-geometry constraint between the predicted normals and the normals derived from the rendered depths." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "Both the normal representation and the constraint lead to an accurate normal estimation of Gaussian spheres, which helps us compute the shading function." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.137, + 0.892, + 0.243 + ], + "angle": 0, + "content": "Built upon 3D Gaussian Splatting, GaussianShader maintains real-time rendering speed while still accommodating various materials like reflective surfaces. Experiments show that, compared to prior works, our method keeps a good balance between efficiency and robust performance on both general scenes and reflective surfaces. In summary, our method offers several significant advantages:" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.244, + 0.892, + 0.304 + ], + "angle": 0, + "content": "1. Our method explicitly approximates the rendering equation by a simplified shading function, significantly enhancing the realism of rendered scenes, particularly for highly specular and reflective surfaces." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.304, + 0.892, + 0.348 + ], + "angle": 0, + "content": "2. We propose a new normal estimation framework on 3D Gaussians with a new regularization loss that allows precise normal estimation." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.349, + 0.892, + 0.41 + ], + "angle": 0, + "content": "3. Leveraging the efficiency of Gaussian Splatting, our approach provides real-time rendering capabilities, making it suitable for interactive applications and scenarios that demand efficient rendering." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.244, + 0.892, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.425, + 0.642, + 0.44 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.45, + 0.719, + 0.465 + ], + "angle": 0, + "content": "2.1. Neural Radiance Fields" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.473, + 0.892, + 0.821 + ], + "angle": 0, + "content": "Neural Radiance Fields (NeRF) [32] gains remarkable progress in photo-realistic novel view synthesis using implicit representation and volume rendering. Recently, NeRF has inspired many follow-up works in various directions. [5, 6] improves NeRF in rendering quality by introducing 3D conical frustum, achieving state-of-art performance in NVS. [17, 36, 47, 52, 53] combine implicit surface representations with NeRF for more accurate geometric reconstruction. [43, 44] propose solutions for city-scale scene rendering using NeRF. [41, 45, 51, 57] targets a special kind of scenes with high specularities and reflections. Another important line of works [10, 14, 15, 28, 33, 42] focuses on acceleration due to the low training and rendering speed of NeRF using voxel grid or hash table. Although great progress has been made, NeRF-based methods still suffer from low rendering speed and high training-time memory usage due to their implicit nature [11]. Based on works on point-based neural rendering [2, 39, 50, 55], a recent milestone work [21] introduces anisotropic 3D Gaussian as an effective representation of the scene, and renders the image using a fast tile-based differentiable rasterizer, surpassing existing implicit neural representation methods in both quality and efficiency." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.832, + 0.755, + 0.849 + ], + "angle": 0, + "content": "2.2. Reflective Object Rendering" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Rendering views of reflective objects from multi-view images has been a challenging task due to the complex light interactions. Previous approaches rely on simple light field" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5323" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.895, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.235, + 0.895, + 0.293 + ], + "angle": 0, + "content": "Figure 2. GaussianShader initiates with the neural 3D Gaussian spheres that integrate both conventional attributes and the newly introduced shading attributes to accurately capture view-dependent appearances. We incorporate a differentiable environment lighting map to simulate realistic lighting. The end-to-end training leads to a model that reconstructs both reflective and diffuse surfaces, achieving high material and lighting fidelity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.47, + 0.481 + ], + "angle": 0, + "content": "interpolations to achieve high-fidelity rendering of novel perspectives [16, 25, 49], yet were constrained by the necessity for dense discrete captures. The accurate rendering of reflective surfaces hinges upon the precise estimation of scene illumination (e.g. environment light) and material properties (e.g. BRDF), which is the task of inverse rendering [4, 35]. Previous studies have demonstrated methodologies for predicting BRDF under known lighting conditions [1, 13, 31], as well as techniques for estimating lighting given known geometry [24, 37, 38]. Furthermore, [12, 19, 51, 58] incorporate indirect illumination, thereby enhancing the fidelity of the estimated BRDF." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.483, + 0.471, + 0.694 + ], + "angle": 0, + "content": "Some NeRF related works attempt to model reflectance by disentangling the visual appearance into lighting and material properties, such as [7-9, 41, 54, 57], which can jointly predict environmental illumination and surface reflectance properties under unknown or varying lighting conditions. Ref-NeRF [45] introduces a new parameterization and structuring of view-dependent outgoing radiance, as well as a regularizer on normal vectors. Recent works [26, 27, 29] utilize SDF-based representation to learn geometry from high specular surfaces, obtaining more accurate normals for physically based rendering. However, these method suffer from extremely time-consuming optimization and slow rendering speed, which hinders their employments in real applications." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.703, + 0.219, + 0.717 + ], + "angle": 0, + "content": "2.3. Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.726, + 0.385, + 0.743 + ], + "angle": 0, + "content": "2.3.1 3D Gaussian Splitting Rasterization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Our method builds upon Gaussian Splatting [21], which begins with a collection of images capturing a static scene, their corresponding camera parameters, and a sparse point cloud generated through Structure-from-Motion (SfM) [40]. These points construct a set of Gaussians, each defined by position (mean) \\(\\mathbf{p}\\) and a 3D covariance matrix \\(\\Sigma\\). While a direct optimization of the covariance matrix \\(\\Sigma\\) might seem intuitive, it presents challenges due to the requirement of positive semi-definiteness. As an alternative, in Gaussian Splatting, a more intuitive yet" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.301, + 0.895, + 0.362 + ], + "angle": 0, + "content": "equally expressive representation is adopted as an ellipsoid for optimization. The ellipsoid is separated into scaling and rotation components, represented by a scaling matrix \\( S \\) and a rotation matrix \\( R \\), i.e., \\( \\Sigma = RSS^T R^T \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.385, + 0.762, + 0.4 + ], + "angle": 0, + "content": "2.3.2 Rendering with 3D Gaussians" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.41, + 0.894, + 0.517 + ], + "angle": 0, + "content": "By projecting to 2D frame space according to the camera parameters, 3D Gaussian spheres are both differentiable and amenable to rapid rendering through 2D splatting with \\(\\alpha\\)-blending. For radiance field modeling, the directional appearance aspect (color) \\(\\mathbf{c}\\) is conveyed through spherical harmonics (SH). By tile-based rasterization, we could sum up the pixel color \\(\\mathbf{C}\\) after sorting \\(\\mathbf{c}_i\\) by depth:" + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.529, + 0.892, + 0.572 + ], + "angle": 0, + "content": "\\[\n\\mathbf {C} = \\sum_ {i \\in N} \\mathbf {c} _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.584, + 0.892, + 0.63 + ], + "angle": 0, + "content": "where \\(\\alpha_{i}\\) is obtained by multiplying Gaussian weight with opacity \\(\\alpha\\) associated to Gaussian sphere. After rasterization, the color loss could be applied to the rendered image:" + }, + { + "type": "equation", + "bbox": [ + 0.624, + 0.645, + 0.892, + 0.665 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c o l o r}} = \\left| \\left| \\mathbf {C} - \\mathbf {C} _ {\\mathrm {g t}} \\right| \\right| ^ {2} \\tag {2}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.678, + 0.591, + 0.694 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.894, + 0.901 + ], + "angle": 0, + "content": "The overview of our method is depicted in Fig. 2. Our approach begins by adopting 3D Gaussian spheres, encompassing shape attributes including covariance \\(\\Sigma\\), opacity \\(\\alpha\\), and position \\(\\mathbf{p}\\). To enhance the representation ability on reflections, we compute the appearances of these Gaussian spheres with a shading function, which requires a set of shading attributes, including diffuse color, roughness, specular tint, normal, and residual color, as detailed in Sec. 3.1. Subsequently, we employ a differentiable environment light map to model the direct lighting, as elaborated in Sec. 3.2. The shading process relies significantly on accurate normal estimation, as discussed in Sec. 3.3. Finally, we introduce the losses used in the whole training process in Sec. 3.4." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5324" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.091, + 0.31, + 0.108 + ], + "angle": 0, + "content": "3.1. Shading on 3D Gaussians" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.114, + 0.47, + 0.28 + ], + "angle": 0, + "content": "Gaussian Splatting [21] models the appearances of Gaussians with simple spherical harmonic functions without considering the light-surface interactions. Thus, Gaussian Splatting fails to accurately represent strong specular surfaces. However, accurately considering the light-surface interactions requires an exact evaluation of the Rendering Equation [20], which requires extensive computational time and complex BRDF parameters. We adopt a simplified approximation of the rendering equation which enables us to achieve high-quality rendering results on reflective surfaces in a considerably shorter time." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.281, + 0.47, + 0.312 + ], + "angle": 0, + "content": "Specifically, for a Gaussian sphere, its rendered color \\(\\mathbf{c}\\) for the viewing direction \\(\\omega_{o}\\) is computed by" + }, + { + "type": "equation", + "bbox": [ + 0.121, + 0.322, + 0.47, + 0.34 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} \\left(\\omega_ {o}\\right) = \\gamma \\left(\\mathbf {c} _ {d} + \\mathbf {s} \\odot L _ {s} \\left(\\omega_ {o}, \\mathbf {n}, \\rho\\right) + \\mathbf {c} _ {r} \\left(\\omega_ {o}\\right)\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.349, + 0.47, + 0.456 + ], + "angle": 0, + "content": "where \\(\\gamma\\) is a gamma tone mapping function [3], \\(\\mathbf{c}_d\\in [0,1]^3\\) is the diffuse color of this Gaussian sphere, \\(\\mathbf{s}\\in [0,1]^3\\) is the specular tint defined on this sphere, \\(L_{s}(\\omega_{o},\\mathbf{n},\\rho)\\) is the direct specular light for this sphere in this direction, \\(\\mathbf{n}\\) is the normal of this Gaussian sphere, \\(\\rho \\in [0,1]\\) is the roughness of the sphere, \\(\\mathbf{c}_r:\\mathbb{R}^3\\to \\mathbb{R}^3\\) is so-called residual colors, and \\(\\odot\\) is the element-wise multiplication." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.455, + 0.471, + 0.744 + ], + "angle": 0, + "content": "Explanations on Eq. 3. We explain our motivation of this shading model in the following three aspects. a) Diffuse color \\(\\mathbf{c}_d\\) represents the consistent colors of this Gaussian sphere, which do not change with viewing directions. b) \\(\\mathbf{s} \\odot L_s(\\omega_o, \\mathbf{n}, \\rho)\\) describes the interactions between the surface intrinsic color \\(\\mathbf{s}\\) and the direct specular light \\(L_s\\). This term enables us to represent most of the reflections in rendering. c) Since there are still some reflections that cannot be explained by the above reflections of direct lights, such as scattering and reflection on indirect lights, we add a residual color term \\(\\mathbf{c}_r(\\omega_o)\\) to account for these complex reflections. In comparison, Ref-NeRF [45] adopts a similar shading function without such a residual color term, which makes it struggle to handle advanced complex reflections. \\(\\mathbf{c}_r(\\omega_o)\\) is parameterized by spherical harmonic functions. d) \\(\\mathbf{c}_d\\), \\(\\mathbf{s}\\), \\(\\rho\\), and the coefficients of the spherical harmonics in \\(\\mathbf{c}_r(\\omega_o)\\) all are trainable parameters associated with this Gaussian sphere. In the following, we will introduce how to compute \\(L_s(\\omega_o, \\mathbf{n}, \\rho)\\), and \\(\\mathbf{n}\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.751, + 0.229, + 0.768 + ], + "angle": 0, + "content": "3.2. Specular Light" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.775, + 0.469, + 0.821 + ], + "angle": 0, + "content": "We compute the specular light \\( L_{s} \\) by integrating incoming radiance with the specular GGX [46] Normal Distribution Function \\( D \\) visualized in Fig. 3" + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.83, + 0.47, + 0.862 + ], + "angle": 0, + "content": "\\[\nL _ {s} (\\omega_ {o}, \\mathbf {n}, \\rho) = \\int_ {\\Omega} L (\\omega_ {i}) D (\\mathbf {r}, \\rho) (\\omega_ {i} \\cdot \\mathbf {n}) d \\omega_ {i}, \\qquad (4)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "where \\(\\Omega\\) represents the whole upper semi-sphere, \\(\\omega_{i}\\) is the direction for the input radiance and \\(D\\) characterizes the" + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.096, + 0.825, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.186, + 0.895, + 0.243 + ], + "angle": 0, + "content": "Figure 3. Normal Distribution Function D in Eq. 4 is determined by the roughness \\(\\rho\\) and reflective direction \\(\\mathbf{r}\\). Surface with small \\(\\rho\\) has a smaller specular lobe and that with large \\(\\rho\\) has a larger specular lobe." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.255, + 0.892, + 0.406 + ], + "angle": 0, + "content": "specular lobe (effective integral range). When the surface is rough, the specular lobe will be larger around the reflective direction \\(\\mathbf{r}\\) while if the surface is smooth, the specular lobe will be smaller. The reflection direction \\(\\mathbf{r}\\) is calculated by view direction \\(\\omega_{o}\\) and normal \\(\\mathbf{n}\\) using \\(\\mathbf{r} = 2(\\omega_{o} \\cdot \\mathbf{n})\\mathbf{n} - \\omega_{o}\\). In our approach, environment light \\(L(\\omega_{i})\\) is represented by a trainable \\(6 \\times 64 \\times 64\\) cube map. All pixel values in our lighting maps are uniformly initialized to 0.8 (within a range of 0 to 1). Note that the lighting map is global, with each scene being optimized using a single lighting map." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.408, + 0.893, + 0.59 + ], + "angle": 0, + "content": "The light modeling is based on NVDiffRec [34]. Specifically, we begin by preprocessing the environment lighting map into multiple mipmap levels, each storing the integrated light across different reflective directions and roughness. When calculating light integration, we interpolate between these mipmaps based on the input of specific reflection directions and roughness, to obtain the required light integrals. This mipmap-based framework, in contrast to Ref-NeRF's [45] approach using integrated directional encoding, enhances training efficiency. The reflection direction \\( r \\) is calculated by combining the observation direction \\( \\omega_{o} \\) with the normal \\( n \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.603, + 0.685, + 0.618 + ], + "angle": 0, + "content": "3.3. Normal Estimation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.892, + 0.764 + ], + "angle": 0, + "content": "Normal estimation on Gaussian spheres is difficult. Gaussian spheres are a collection of discrete entities, each representing a localized point in space without a continuous surface or defined edge. This discrete structure makes it inherently difficult to directly calculate normals, which typically requires a continuous surface. In the following, we observed that the shortest axis direction of a Gaussian can serve as an approximated normal and we further associate a predicted normal residual on it." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Shortest axis direction. In experiments, we made an interesting observation that the aspect ratio of the 3D Gaussian sphere—specifically, the ratios of their longest, intermediate, and shortest axis—gradually increased during optimization, as shown in Fig. 4. This observation suggests that Gaussian spheres gradually become more flattened and approach a planar shape. This observation inspires us to select the shortest axis as the normal of this \"planarized\" Gaussian sphere, denoted by \\(\\mathbf{v}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5325" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.09, + 0.454, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.205, + 0.47, + 0.234 + ], + "angle": 0, + "content": "Figure 4. The geometric evolving process of a 3D Gaussian sphere in the optimization, which gradually becomes planar." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.24, + 0.473, + 0.437 + ], + "angle": 0, + "content": "Predicted normal residual. The shortest axis \\(\\mathbf{v}\\) only serves as an approximated normal. To make the normal computation more accurate, we further introduce a trainable normal residual \\(\\Delta \\mathbf{n}\\) on every Gaussian sphere. However, the orientation of the shortest axis \\(\\mathbf{v}\\) has an ambiguity because the direction of the shortest axis could either point outward or inward from the surface. To handle this ambiguity, we optimize two separate normal residuals to accommodate both scenarios. Given a specific viewing direction \\(\\omega_{o}\\), we first select the direction aligned with the viewing direction \\(\\omega_{\\mathrm{o}}\\) as the active normal direction for this viewing direction and then apply the corresponding normal residual to the active normal. This process is described by" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.444, + 0.47, + 0.484 + ], + "angle": 0, + "content": "\\[\n\\mathbf {n} = \\left\\{ \\begin{array}{l l} \\mathbf {v} + \\Delta \\mathbf {n} _ {1} & \\text {i f} \\omega_ {\\mathbf {o}} \\cdot \\mathbf {v} > 0, \\\\ - (\\mathbf {v} + \\Delta \\mathbf {n} _ {2}) & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.491, + 0.471, + 0.536 + ], + "angle": 0, + "content": "To prevent the normal residual from deviating too much from the shortest axis, we add a penalty towards normal residual, making sure it is small enough." + }, + { + "type": "equation", + "bbox": [ + 0.221, + 0.543, + 0.47, + 0.561 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e g}} = \\left\\| \\Delta \\mathbf {n} \\right\\| ^ {2} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.569, + 0.408, + 0.584 + ], + "angle": 0, + "content": "The normal residual is shown on the left of Fig. 5." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.585, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Normal-geometry consistency. The above shortest axis direction and normal residuals are defined on each Gaussian sphere separately. However, a noticeable problem is that a normal reveals the gradient of the local geometry, which is supposed to be associated with all the Gaussian spheres in a local region. We find that simply applying a color loss to train the aforementioned normal residuals leads to inconsistency between the local geometry and the estimated normals. The main reason is that every Gaussian sphere learns its normal residuals separately without knowing the local geometry formulated by its neighbor Gaussian spheres. Thus, we have to correlate multiple Gaussian spheres in a local region with their normals to ensure normal-geometry consistency. A straightforward and naive solution is to search for its \\( K \\) neighborhoods in space and estimate a coarse normal from all the neighboring spheres. However, such KNN searching would be extremely expensive during training because all the Gaussian spheres are dynamically moving in the optimization process. Instead, we propose a simple yet effective way to ensure normal-geometry consistency as follows." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.088, + 0.895, + 0.149 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.158, + 0.894, + 0.201 + ], + "angle": 0, + "content": "Figure 5. Visualization of the relationship between shortes axis \\(\\mathbf{v}\\), normal residual \\(\\Delta \\mathbf{n}\\), normal \\(\\mathbf{n}\\) and depth-gradient normal \\(\\hat{\\mathbf{n}}\\). The supervision \\(\\mathcal{L}_{normal}\\) enforces the normal-geometry consistency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.209, + 0.894, + 0.269 + ], + "angle": 0, + "content": "We associate the local geometry with predicted normals by minimizing the difference between the grad normals derived from the rendered depth map and the rendered normal maps using the predicted normals" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.279, + 0.892, + 0.297 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {n o r m a l}} = \\left| \\left| \\bar {\\mathbf {n}} - \\hat {\\mathbf {n}} \\right| \\right| ^ {2}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.307, + 0.893, + 0.428 + ], + "angle": 0, + "content": "where \\(\\bar{\\mathbf{n}}\\) is the rendered normal map and \\(\\hat{\\mathbf{n}}\\) is computed by applying the Sobel-like operator on the rendered depth map. \\(\\hat{\\mathbf{n}}\\) reveals the local geometry formulated by multiple Gaussian spheres because it is computed from the rendered depth maps. \\(\\bar{\\mathbf{n}}\\) contains the information from the separately-defined normal on each Gaussian sphere. By minimizing the difference, we enforce the consistency between the local geometry and the estimated normals." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.437, + 0.587, + 0.451 + ], + "angle": 0, + "content": "3.4. Losses" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.46, + 0.894, + 0.505 + ], + "angle": 0, + "content": "In addition to \\(\\mathcal{L}_{\\mathrm{color}}\\), \\(\\mathcal{L}_{reg}\\), and \\(\\mathcal{L}_{\\mathrm{normal}}\\), we make use of sparse loss [30, 50] to encourages Gaussian spheres' opacity values \\(\\alpha\\) to approach either 0 or 1 by" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.514, + 0.892, + 0.549 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s p a r s e}} = \\frac {1}{| \\alpha |} \\sum_ {\\alpha_ {i}} \\left[ \\log \\left(\\alpha_ {i}\\right) + \\log \\left(1 - \\alpha_ {i}\\right) \\right]. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.559, + 0.892, + 0.605 + ], + "angle": 0, + "content": "This sparsity loss helps the geometry of Gaussian spheres converge to a single thin plate and improves the rendering quality. In summary, the total training loss \\(\\mathcal{L}\\) is" + }, + { + "type": "equation", + "bbox": [ + 0.551, + 0.616, + 0.892, + 0.632 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {c o l o r}} + \\lambda_ {\\mathrm {n}} \\mathcal {L} _ {\\text {n o r m a l}} + \\lambda_ {\\mathrm {s}} \\mathcal {L} _ {\\text {s p a r s e}} + \\lambda_ {\\mathrm {r}} \\mathcal {L} _ {\\text {r e g}} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.642, + 0.78, + 0.657 + ], + "angle": 0, + "content": "where \\(\\lambda_{\\mathrm{n}} = 0.01, \\lambda_{\\mathrm{s}} = 0.001, \\lambda_{\\mathrm{r}} = 0.001\\)" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.67, + 0.633, + 0.687 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.695, + 0.602, + 0.71 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.718, + 0.894, + 0.809 + ], + "angle": 0, + "content": "To comprehensively validate the effectiveness of GaussianShader, we conduct evaluation on various datasets: a) widely-used NVS dataset: NeRF Synthetic [32]. b) reflective objects datasets: Shiny Blender [45] and Glossy Synthetic [29]. c) real-world large-scale scenes: Tanks and Temples [23]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.818, + 0.705, + 0.832 + ], + "angle": 0, + "content": "4.2. Baselines and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.894, + 0.901 + ], + "angle": 0, + "content": "We compare our method against the following baselines: a) 3D Gaussian Splatting [21]: a real-time radiance field rendering method based on efficient 3D Gaussian representation; b) VolSDF [52]: a classical neural implicit surface" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5326" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.201, + 0.089, + 0.768, + 0.104 + ], + "angle": 0, + "content": "Table 1. The quantitative comparisons (PSNR / SSIM / LPIPS) on NeRF Synthetic dataset [32]." + }, + { + "type": "table", + "bbox": [ + 0.182, + 0.105, + 0.79, + 0.407 + ], + "angle": 0, + "content": "
ChairDrumsNeRF Synthetic [32]
LegoMicMaterialsShipHotdogFicusAvg.
PSNR↑
NeRF [32]33.0025.0132.5432.9129.6228.6536.1830.1331.01
VolSDF [52]30.5720.4329.4630.5329.1325.5135.1122.9127.96
Ref-NeRF [45]33.9825.4335.1033.6527.1029.2437.0428.7431.29
ENVIDR [27]31.2222.9929.5532.1729.5221.5731.4426.6028.13
Gaussian Splitting [21]35.8226.1735.6935.3430.0030.8737.6734.8333.30
Ours35.8326.3635.8735.2330.0730.8237.8534.9733.38
SSIM↑
NeRF [32]0.9670.9250.9610.9800.9490.8560.9740.9640.947
VolSDF [52]0.9490.8930.9510.9690.9540.8420.9720.9290.932
Ref-NeRF [45]0.9740.9290.9750.9830.9210.8640.9790.9540.947
ENVIDR [27]0.9760.9300.9610.9840.9680.8550.9630.9870.956
Gaussian Splitting [21]0.9870.9540.9830.9910.9600.9070.9850.9870.969
Ours0.9870.9490.9830.9910.9600.9050.9850.9850.968
LPIPS↓
NeRF [32]0.0460.0910.0500.0280.0630.2060.1210.0440.081
VolSDF [52]0.0560.1190.0540.1910.0480.1910.0430.0680.096
Ref-NeRF [45]0.0290.0730.0250.0180.0780.1580.0280.0560.058
ENVIDR [27]0.0310.0800.0540.0210.0450.2280.0720.0100.067
Gaussian Splitting [21]0.0120.0370.0160.0060.0340.1060.0200.0120.030
Ours0.0120.0400.0140.0060.0330.0980.0190.0130.029
" + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.418, + 0.478, + 0.63 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.481, + 0.418, + 0.892, + 0.628 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.642, + 0.893, + 0.673 + ], + "angle": 0, + "content": "Figure 6. The qualitative comparisons with 3D Gaussian Splitting [21] on Glossy dataset [29]. Our method not only renders the objects with high fidelity but also provides detailed normal and lighting maps. Some areas are zoomed in for better visualization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.686, + 0.47, + 0.807 + ], + "angle": 0, + "content": "reconstruction method based on SDF; c) Ref-NeRF [45]: state-of-the-art method in novel view synthesis of reflective objects; d) NVDiffRec [34], NVDiffRecMC [18]: top-performing neural inverse rendering methods; e) ENVIDR [27], NeRO [29]: top-performing SDF-based neural implicit methods for reconstructing reflective objects. The evaluation metrics to measure rendering quality are reported in PSNR, SSIM [48], and LPIPS [56]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.295, + 0.833 + ], + "angle": 0, + "content": "4.3. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "All experiments are conducted on a Nvidia RTX 3090 graphics card. We optimize our models using Adam [22] optimizer for 30,000 iterations. For Ref-NeRF [45], ENVIDR [27], and 3D Gaussian Splitting [21], we retrain" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.686, + 0.892, + 0.748 + ], + "angle": 0, + "content": "them using their official codes keeping all the original settings (training iterations, sampled rays, etc.) and obtain results from the retrained models. Other works' results are imported from their original papers." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.757, + 0.64, + 0.773 + ], + "angle": 0, + "content": "4.4. Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.893, + 0.903 + ], + "angle": 0, + "content": "NeRF Synthetic dataset [32]. We first evaluate our model on the NeRF Synthetic dataset, which contains objects with complex geometry and realistic non-Lambertian materials. We show the quantitative results in Tab. 1 and visual comparisons in Fig. 8. Our approach achieves numerically and visually comparable results with both Gaussian Splatting and neural rendering methods, demonstrating the effectiveness of our method in rendering general objects." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5327" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.087, + 0.473, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.232, + 0.47, + 0.275 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparisons on Shiny Blender dataset [45]. Error maps corresponding to each rendering result demonstrate the enhanced fidelity of our approach in highly specular areas." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.285, + 0.469, + 0.316 + ], + "angle": 0, + "content": "Note that SDF-based ENVIDR [27] cannot perform well on shadowed regions of complex objects, as shown in Fig. 8." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.318, + 0.469, + 0.514 + ], + "angle": 0, + "content": "Shiny Blender dataset [45]. Quantitative results on the Shiny Blender dataset are reported in Tab. 2. With the ability to model light-surface interactions, our method outperforms original Gaussian Splatting in all scenes while slightly underperforms ENVIDR and Ref-NeRF. ENVIDR capitalizes on the continuous properties of Signed Distance Fields to naturally create smoother surfaces and normals, resulting in high rendering quality, particularly for reflective objects. Fig. 7 shows our comparisons with Gaussian Splatting. We can see that our method correctly renders surfaces with strong specular appearances. High-quality modeling of reflections is contingent upon the accurate estimation of normals as shown in Fig. 9." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.517, + 0.469, + 0.669 + ], + "angle": 0, + "content": "Additionally, we report the average PSNR, training time, and rendering FPS in Tab. 3. The reported results are averaged among all objects of the Shiny Blender and NeRF synthetic datasets. On the same hardware, our method only takes about 0.5 hour for training while MLP-based methods, like ENVIDR and Ref-NeRF, require 6 hours and 23 hours for optimizing. Due to the introduce of extra lighting and shading attributes, our method is a bit slower than 3D Gaussian Splatting [21] but still keeps reasonable efficiency and real-time rendering speed." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.671, + 0.47, + 0.747 + ], + "angle": 0, + "content": "Tanks and Temples dataset [23]. To explore our method's scalability in larger-scale environments instead of only small objects, we conducted experiments using Tanks and Temples [23]. The qualitative results in Fig. 10 show that our model outperforms Gaussian Splatting [21]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Glossy Synthetic dataset [29]. We present a comparative visualization between Gaussian Splatting and our method on the Glossy dataset shown in Fig. 6. Our method not only renders objects with exceptional fidelity but also provides more accurate normal and lightings, capturing the nuances of reflective surfaces with greater precision. These maps reveal our method's ability to approximate the rendering function, resulting in a more lifelike portrayal of specular lighting. Please refer to the supplementary materials provided for an extensive showcase of our results." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.132 + ], + "angle": 0, + "content": "Table 2. Qualitative comparisons on Shiny Blender dataset [45]. Our method is comparable with both Gaussian Splatting [21] and prior reflective object reconstruction methods." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.396 + ], + "angle": 0, + "content": "
Shiny Blender [45]
CarBallHelmetTeapotToasterCoffeeAvg.
PSNR↑
NVDiffRec [34]27.9821.7726.9740.4424.3130.7428.70
NVDiffMC [18]25.9330.8526.2738.4422.1829.6028.88
Ref-NeRF [45]30.4129.1429.9245.1925.2933.9932.32
NeRO [29]25.5330.2629.2038.7026.4628.8929.84
ENVIDR [27]28.4638.8932.7341.5926.1129.4832.88
Gaussian Splitting [21]27.2427.6928.3245.6820.9932.3230.37
Ours27.9030.9828.3245.8626.2132.3931.94
SSIM↑
NVDiffRec [34]0.9630.8580.9510.9960.9280.9730.945
NVDiffMC [18]0.9400.9400.9400.9950.8860.9650.944
Ref-NeRF [45]0.9490.9560.9550.9950.9100.9720.956
NeRO [29]0.9490.9740.9710.9950.9290.9560.962
ENVIDR [27]0.9610.9910.9800.9960.9390.9490.969
Gaussian Splitting [21]0.9300.9370.9510.9960.8950.9710.947
Ours0.9310.9650.9500.9960.9290.9710.957
LPIPS↓
NVDiffRec [34]0.0450.2970.1180.0110.1690.0760.119
NVDiffMC [18]0.0770.3120.1570.0140.2250.0970.147
Ref-NeRF [45]0.0510.3070.0870.0130.1180.0820.109
NeRO [29]0.0740.0940.0500.0120.0890.1100.072
ENVIDR [27]0.0490.0670.0510.0110.1160.1390.072
Gaussian Splitting [21]0.0470.1610.0790.0070.1260.0780.083
Ours0.0450.1210.0760.0070.0790.0780.068
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.402, + 0.892, + 0.444 + ], + "angle": 0, + "content": "Table 3. Comparisons over average PSNR, training time and FPS. Our method achieves the best rendering quality and competitive training/rendering speed." + }, + { + "type": "table", + "bbox": [ + 0.536, + 0.449, + 0.857, + 0.514 + ], + "angle": 0, + "content": "
PSNRTraining TimeFPS
Ref-NeRF [45]31.7323h0.03
ENVIDR [27]30.046h1.33
Gaussian Splitting [21]32.050.25h274
Ours32.760.58h97
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.541, + 0.653, + 0.557 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.566, + 0.892, + 0.626 + ], + "angle": 0, + "content": "In this section, we conduct ablation studies of our model on Shiny Blender dataset. The numerical results are reported in Tab. 4. All ablation experiments are done on Shiny Blender dataset with \\(1/2\\) image resolution." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.627, + 0.892, + 0.732 + ], + "angle": 0, + "content": "Loss functions. We trained our model serperately under the setting \"without sparse loss \\(\\mathcal{L}_{\\mathrm{sparse}}\\) and \"without normal loss \\(\\mathcal{L}_{\\mathrm{normal}}\\). The incorporation of these losses, especially the direct regularization on normals, proved critical, yielding a marked improvement in rendering quality by effectively guiding the normal optimization process as shown in Tab. 4." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.734, + 0.892, + 0.825 + ], + "angle": 0, + "content": "Residual color. Residual color, comprising 3rd order SH coefficients, is used to compensate for indirect light. As shown in Table 4, the introduce of residual color enhances our model, allowing it to simulate a wide array of real-world intricate lighting situations and significantly improve rendering quality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Normal formulation. In our full model, we use the shortest axis direction \\(\\mathbf{v}\\) with a learnable residual \\(\\Delta \\mathbf{n}\\) to formulate normal, as described by Eqn. 3.3. In Tab. 4, we compared this novel normal formation with a naive normal prediction manner (indicated by \"w/o axis \\(\\mathbf{v}\\) ). Instead of" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5328" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.11, + 0.087, + 0.862, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.299, + 0.893, + 0.328 + ], + "angle": 0, + "content": "Figure 8. We present a qualitative comparisons of our method on NeRF Synthetic dataset [32] against previous techniques and corresponding ground truth images from test views. For clarity, we zoom in the images where differences in quality are especially notable." + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.332, + 0.47, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.533, + 0.47, + 0.601 + ], + "angle": 0, + "content": "Figure 9. Visualized normal results on Shiny Blender dataset [45], compared with Gaussian Splatting [21]. It illustrates the superior normal estimation achieved by our method. Table 4. Ablation studies on loss function, lighting representation, normal axis and residual color using Shiny Blender dataset." + }, + { + "type": "table", + "bbox": [ + 0.126, + 0.611, + 0.416, + 0.713 + ], + "angle": 0, + "content": "
PSNR↑SSIM↑LPIPS↓
w/o Lsparse31.790.9520.056
w/o Lnormal30.930.9410.060
w/o cr31.490.9480.060
w/o v31.470.9510.058
MLP Lighting29.730.9360.075
Full Model32.090.9530.054
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.795 + ], + "angle": 0, + "content": "correlating with the geometry of 3D gaussians, we directly optimize a single normal attribute. The result shows our full model can achieve higher rendering quality, since our predicted normals are correlated with the true geometry of Gaussian spheres." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Lighting representation. A crucial component of our approach is the modeling of lighting. \"MLP Lighting\" means we use a directional MLP similar to Ref-NeRF to implicitly optimize lighting. In our full model, we use \"Envmap Lighting\" representation as illustrated in Sec. 3.2. The result proves that our explicit lighting representation is better suited for 3D Gaussian rendering framework to cor" + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.332, + 0.892, + 0.605 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.615, + 0.894, + 0.657 + ], + "angle": 0, + "content": "Figure 10. Qualitative comparisons with Gaussian Splitting [21] on the Tanks and Temples dataset [23]. Our approach yields a smoother and more reasonable normal." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.672, + 0.707, + 0.687 + ], + "angle": 0, + "content": "rectly render reflective objects." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.707, + 0.627, + 0.723 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.894, + 0.902 + ], + "angle": 0, + "content": "In summary, we propose GaussianShader, advancing the rendering of both reflective and general objects through an extended 3D Gaussian model. Specifically, our method integrates shading functions with 3D Gaussian representation for handling view-dependent appearances, successfully handling reflective surfaces. Moreover, a novel normal prediction method is proposed to achieve high-quality normals, which correlates the shading parameters and true geometry of 3D Gaussians. The results show our approach achieves high-quality renderings, marking a substantial improvement in efficient, realistic 3D object rendering." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5329" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.157 + ], + "angle": 0, + "content": "[1] Miika Aittala, Timo Aila, and Jaakko Lehtinen. Reflectance modeling by neural texture synthesis. ACM Transactions on Graphics (ToG), 35(4):1-13, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.472, + 0.227 + ], + "angle": 0, + "content": "[2] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 696-712. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.227, + 0.471, + 0.298 + ], + "angle": 0, + "content": "[3] Matthew Anderson, Ricardo Motta, Srinivasan Chandrasekar, and Michael Stokes. Proposal for a standard default color space for the internet—srgb. In Color and imaging conference, pages 238–245. Society for Imaging Science and Technology, 1996. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.471, + 0.352 + ], + "angle": 0, + "content": "[4] Jonathan T Barron and Jitendra Malik. Shape, illumination, and reflectance from shading. IEEE transactions on pattern analysis and machine intelligence, 37(8):1670-1687, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.354, + 0.471, + 0.436 + ], + "angle": 0, + "content": "[5] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.471, + 0.508 + ], + "angle": 0, + "content": "[6] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470-5479, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.508, + 0.471, + 0.577 + ], + "angle": 0, + "content": "[7] Sai Bi, Zexiang Xu, Pratul Srinivasan, Ben Mildenhall, Kalyan Sunkavalli, Milos Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Neural reflectance fields for appearance acquisition. arXiv preprint arXiv:2008.03824, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.578, + 0.472, + 0.648 + ], + "angle": 0, + "content": "[8] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T Barron, Ce Liu, and Hendrik Lensch. Nerd: Neural reflectance decomposition from image collections. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12684-12694, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.649, + 0.471, + 0.717 + ], + "angle": 0, + "content": "[9] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan Barron, and Hendrik Lensch. Neural-pil: Neural pre-integrated lighting for reflectance decomposition. Advances in Neural Information Processing Systems, 34: 10691-10704, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.718, + 0.471, + 0.773 + ], + "angle": 0, + "content": "[10] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.774, + 0.471, + 0.814 + ], + "angle": 0, + "content": "[11] Zilong Chen, Feng Wang, and Huaping Liu. Text-to-3d using gaussian splatting. arXiv preprint arXiv:2309.16585, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.816, + 0.471, + 0.871 + ], + "angle": 0, + "content": "[12] Youming Deng, Xueting Li, Sifei Liu, and Ming-Hsuan Yang. Dip: Differentiable interreflection-aware physics-based inverse rendering. arXiv preprint arXiv:2212.04705, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.872, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[13] Valentin Deschaintre, Miika Aittala, Fredo Durand, George Drettakis, and Adrien Bousseau. Single-image svbrdf cap" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "ture with a rendering-aware deep network. ACM Transactions on Graphics (ToG), 37(4):1-15, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.192 + ], + "angle": 0, + "content": "[14] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.194, + 0.892, + 0.263 + ], + "angle": 0, + "content": "[15] Stephan J Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14346-14355, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.265, + 0.892, + 0.32 + ], + "angle": 0, + "content": "[16] Steven J Gortler, Radek Grzesczuk, Richard Szeliski, and Michael F Cohen. The lumigraph. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 453-464. 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.324, + 0.892, + 0.393 + ], + "angle": 0, + "content": "[17] Haoyu Guo, Sida Peng, Haotong Lin, Qianqian Wang, Guofeng Zhang, Hujun Bao, and Xiaowei Zhou. Neural 3d scene reconstruction with the manhattan-world assumption. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5511-5520, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.395, + 0.892, + 0.451 + ], + "angle": 0, + "content": "[18] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, Light, and Material Decomposition from Images using Monte Carlo Rendering and Denoising. arXiv:2206.03380, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.453, + 0.892, + 0.51 + ], + "angle": 0, + "content": "[19] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, light, and material decomposition from images using monte carlo rendering and denoising. Advances in Neural Information Processing Systems, 35:22856-22869, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.511, + 0.892, + 0.554 + ], + "angle": 0, + "content": "[20] James T Kajiya. The rendering equation. In Proceedings of the 13th annual conference on Computer graphics and interactive techniques, pages 143-150, 1986. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.555, + 0.892, + 0.611 + ], + "angle": 0, + "content": "[21] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 42(4):1-14, 2023. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.612, + 0.892, + 0.654 + ], + "angle": 0, + "content": "[22] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.656, + 0.892, + 0.713 + ], + "angle": 0, + "content": "[23] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36 (4):1-13, 2017. 5, 7, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.714, + 0.892, + 0.799 + ], + "angle": 0, + "content": "[24] Chloe LeGendre, Wan-Chun Ma, Graham Fyffe, John Flynn, Laurent Charbonnel, Jay Busch, and Paul Debevec. Deeplight: Learning illumination for unconstrained mobile mixed reality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5918-5928, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.8, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[25] Marc Levoy and Pat Hanrahan. Light field rendering. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 441-452. 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[26] Ruofan Liang, Jiahao Zhang, Haoda Li, Chen Yang, Yushi Guan, and Nandita Vijaykumar. Spidr: Sdf-based neural point fields for illumination and deformation. arXiv preprint arXiv:2210.08398, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "5330" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[27] Ruofan Liang, Huiting Chen, Chunlin Li, Fan Chen, Selvakumar Panneer, and Nandita Vijaykumar. Envidr: Implicit differentiable renderer with neural environment lighting. arXiv preprint arXiv:2303.13022, 2023. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.203 + ], + "angle": 0, + "content": "[28] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.205, + 0.469, + 0.272 + ], + "angle": 0, + "content": "[29] Yuan Liu, Peng Wang, Cheng Lin, Xiaoxiao Long, Jiepeng Wang, Lingjie Liu, Taku Komura, and Wenping Wang. Nano: Neural geometry and brdf reconstruction of reflective objects from multiview images. arXiv preprint arXiv:2305.17398, 2023. 3, 5, 6, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.274, + 0.469, + 0.328 + ], + "angle": 0, + "content": "[30] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural volumes: Learning dynamic renderable volumes from images. arXiv preprint arXiv:1906.07751, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.33, + 0.469, + 0.357 + ], + "angle": 0, + "content": "[31] Wojciech Matusik. A data-driven reflectance model. PhD thesis, Massachusetts Institute of Technology, 2003. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.359, + 0.469, + 0.426 + ], + "angle": 0, + "content": "[32] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.428, + 0.469, + 0.482 + ], + "angle": 0, + "content": "[33] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.484, + 0.469, + 0.565 + ], + "angle": 0, + "content": "[34] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3d models, materials, and lighting from images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8280-8290, 2022. 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.567, + 0.469, + 0.62 + ], + "angle": 0, + "content": "[35] Merlin Nimier-David, Delio Vicini, Tizian Zeltner, and Wenzel Jakob. Mitsuba 2: A retargetable forward and inverse renderer. ACM Transactions on Graphics (TOG), 38(6):1-17, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.623, + 0.469, + 0.691 + ], + "angle": 0, + "content": "[36] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5589-5599, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.693, + 0.47, + 0.747 + ], + "angle": 0, + "content": "[37] Jeong Joon Park, Aleksander Holynski, and Steven M Seitz. Seeing the world in a bag of chips. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1417-1427, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.749, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[38] Thomas Richter-Trummer, Denis Kalkofen, Jinwoo Park, and Dieter Schmalstieg. Instant mixed reality lighting from casual scanning. In 2016 IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 27-36. IEEE, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.818, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[39] Darius Rückert, Linus Franke, and Marc Stamminger. Adop: Approximate differentiable one-pixel point rendering. ACM Transactions on Graphics (ToG), 41(4):1-14, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.86, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[40] Noah Snavely, Steven M Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. In ACM siggraph 2006 papers, pages 835-846. 2006. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.174 + ], + "angle": 0, + "content": "[41] Pratul P Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T Barron. Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7495-7504, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.176, + 0.892, + 0.243 + ], + "angle": 0, + "content": "[42] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.246, + 0.892, + 0.327 + ], + "angle": 0, + "content": "[43] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8248–8258, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.329, + 0.892, + 0.398 + ], + "angle": 0, + "content": "[44] Haithem Turki, Deva Ramanan, and Mahadev Satyanarayanan. Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12922–12931, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.892, + 0.481 + ], + "angle": 0, + "content": "[45] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481-5490. IEEE, 2022. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.55 + ], + "angle": 0, + "content": "[46] Bruce Walter, Stephen R Marschner, Hongsong Li, and Kenneth E Torrance. Microfacet models for refraction through rough surfaces. In Proceedings of the 18th Eurographics conference on Rendering Techniques, pages 195-206, 2007. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.552, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[47] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[48] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.665, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[49] Daniel N Wood, Daniel I Azuma, Ken Aldinger, Brian Curless, Tom Duchamp, David H Salesin, and Werner Stuetzle. Surface light fields for 3d photography. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 487-496. 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.734, + 0.892, + 0.803 + ], + "angle": 0, + "content": "[50] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5438–5448, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[51] Yao Yao, Jingyang Zhang, Jingbo Liu, Yihang Qu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Neilf: Neural incident light field for physically-based material estimation. In European Conference on Computer Vision, pages 700-716. Springer, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[52] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neu" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "5331" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.468, + 0.119 + ], + "angle": 0, + "content": "ral Information Processing Systems, 34:4805-4815, 2021. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.189 + ], + "angle": 0, + "content": "[53] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in neural information processing systems, 35:25018-25032, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.261 + ], + "angle": 0, + "content": "[54] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snavely. Physg: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5453-5462, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.469, + 0.318 + ], + "angle": 0, + "content": "[55] Qiang Zhang, Seung-Hwan Baek, Szymon Rusinkiewicz, and Felix Heide. Differentiable point-based radiance fields for efficient view synthesis. In SIGGRAPH Asia 2022 Conference Papers, pages 1-12, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.32, + 0.469, + 0.387 + ], + "angle": 0, + "content": "[56] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.39, + 0.469, + 0.458 + ], + "angle": 0, + "content": "[57] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Ner-factor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (ToG), 40(6):1-18, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.46, + 0.469, + 0.528 + ], + "angle": 0, + "content": "[58] Yuanqing Zhang, Jiaming Sun, Xingyi He, Huan Fu, Rongfei Jia, and Xiaowei Zhou. Modeling indirect illumination for inverse rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18643-18652, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5332" + } + ] +] \ No newline at end of file diff --git a/2024/GaussianShader_ 3D Gaussian Splatting with Shading Functions for Reflective Surfaces/e3ac7488-6672-4bab-abdc-38e188c8adc6_origin.pdf b/2024/GaussianShader_ 3D Gaussian Splatting with Shading Functions for Reflective Surfaces/e3ac7488-6672-4bab-abdc-38e188c8adc6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7e1da8f7cd08fe09fabb796f9284adf8f4641590 --- /dev/null +++ b/2024/GaussianShader_ 3D Gaussian Splatting with Shading Functions for Reflective Surfaces/e3ac7488-6672-4bab-abdc-38e188c8adc6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3dddcabe99c25fea7f37737f3eb0ddd0db43513e55dfa548a07b7b77c330a327 +size 5510450 diff --git a/2024/GaussianShader_ 3D Gaussian Splatting with Shading Functions for Reflective Surfaces/full.md b/2024/GaussianShader_ 3D Gaussian Splatting with Shading Functions for Reflective Surfaces/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4c89dbb710f3016cc5f855b9e4281afa184a8f11 --- /dev/null +++ b/2024/GaussianShader_ 3D Gaussian Splatting with Shading Functions for Reflective Surfaces/full.md @@ -0,0 +1,336 @@ +# GaussianShader: 3D Gaussian Splating with Shading Functions for Reflective Surfaces + +Yingwenqi Jiang $^{1}$ Jiadong Tu $^{1}$ Yuan Liu $^{2}$ Xifeng Gao $^{3}$ + +Xiaoxiao Long $^{2,*}$ Wenping Wang $^{4}$ Yuexin Ma $^{1,*}$ + +$^{1}$ ShanghaiTech University + +2The University of Hong Kong + +3Tencent America + +Texas A&M University + +![](images/bd5035014d6f472068fcdb1864689d3d3862a3f53469223ea44f1c5f1c1dbffa.jpg) +PSNR:21.69 + +![](images/c69300f47475fd7e15f3521d175106b04c2c1913e4c6bddd1a3c69ca24d31e28.jpg) +PSNR: 28.87 + +![](images/0081ed2f08744dc3bb8ad153208d7331c3d3203dced177e3e0ac530988bcf6ce.jpg) +PSNR:32.67 + +![](images/8a9006e4b11d153459688da78ef59907630788cc29ed695d4dd11e76566e049b.jpg) +PSNR:32.46 + +![](images/e4c9d9ee7e4c248ab9a75b9e99441f9970e9819a7fb9b399da3cf2a903306579.jpg) + +![](images/e6a7bb1b1a0b7da5b62ffd5100d2ee0b3ec3784a7b812a98c7385ac4018d7db2.jpg) +Ground Truth + +![](images/439a08bcd9b263589bd7bc0e52209218cb1f4da2c557044437df4185499cdd5b.jpg) +(b) Comparisons between FPS, training time and accuracy +Figure 1. GaussianShader maintains real-time rendering speed and renders high-fidelity images for both general and reflective surfaces. Ref-NeRF[45] and ENVIDR[27] attempt to handle reflective surfaces, but they suffer from quite time-consuming optimization and slow rendering speed. 3D Gaussian splattering [21] keeps high efficiency but cannot handle such reflective surfaces. + +# Abstract + +The advent of neural 3D Gaussians [21] has recently brought about a revolution in the field of neural rendering, facilitating the generation of high-quality renderings at real-time speeds. However, the explicit and discrete representation encounters challenges when applied to scenes featuring reflective surfaces. In this paper, we present GaussianShader, a novel method that applies a simplified shading function on 3D Gaussians to enhance the neural rendering in scenes with reflective surfaces while preserving the training and rendering efficiency. The main challenge in applying the shading function lies in the accurate normal estimation on discrete 3D Gaussians. Specifically, we proposed a novel normal estimation framework based on the shortest axis directions of 3D Gaussians with a delicately designed loss to make the consistency between the normals and the geometries of Gaussian spheres. Exper + +periments show that GaussianShader strikes a commendable balance between efficiency and visual quality. Our method surpasses Gaussian Splatting [21] in PSNR on specular object datasets, exhibiting an improvement of 1.57dB. When compared to prior works handling reflective surfaces, such as Ref-NeRF [45], our optimization time is significantly accelerated (23h vs. 0.58h). Please click on our project website to see more results. + +# 1. Introduction + +In recent years, the field of 3D computer vision has witnessed remarkable advancements in the 3D reconstruction and visualization of 3D scenes. Innovations such as Neural Radiance Fields (NeRF) [32] have achieved substantial breakthroughs in generating novel views of 3D objects and scenes, presenting the potential for high-quality and photorealistic renderings. Despite these advancements, NeRF-related methodologies [6, 27, 45], still contend with challenges such as computationally expensive optimization and slow rendering speed. These limitations restrict their application in real-time interactive scenarios. + +More recently, 3D Gaussian Splatting [21] combines 3D Gaussian representation and tile-based splatting techniques to achieve high-quality 3D scene modeling and real-time rendering, making it possible to employ neural rendering techniques in real applications. However, it suffers from a performance drop on scenes featuring specular and reflective surfaces. This is because 3D Gaussian Splatting [21] does not explicitly model appearance properties, so that fails to capture significant view-dependent changes, particularly specular highlights. This constraint presents a substantial obstacle in the pursuit of achieving photorealistic rendering across a diverse array of materials, particularly those characterized by prominent reflective attributes. + +Accurately modeling reflective surfaces is a challenging task. Ref-NeRF [45] and ENVIDR [27] combine the shading functions in implicit representations and present promising quality on reflective surfaces. However, they suffer from time-consuming optimization (hours) and slow rendering speed. Due to the limited flexibility of SDF, ENVIDR [27] even fails to model complex scenes and presents a significant performance drop on general objects. It is still an unexplored problem how to combine the shading functions in a 3D Gaussian Splatting framework to improve its ability to handle reflections while preserving the efficiency in training and rendering. + +In this paper, we present GaussianShader, a novel method that enhances the neural rendering of 3D Gaussians within scenes that contain reflective surfaces by incorporating a shading function on 3D Gaussians. To ensure the efficiency of GaussianShader, evaluating the shading function cannot be too expensive while still retaining the ability to model the reflections. In light of this, we propose a novel simplified shading function that considers the diffuse colors and the direct reflections while putting all advanced complex reflections into a residual color term. In comparison with the shading function of Ref-NeRF which can only consider direct reflections, the utilization of this residual color enables GaussianShader to render more complex reflective appearances with efficiency. + +Another challenge in computing a shading function is how to predict accurate normals on the discrete 3D Gaussian spheres. First, it is hard to get a locally-continuous surface from the 3D Gaussians to compute the surface normals. Second, associating multiple 3D Gaussians for normal computation would be very expensive using neighborhood searching. In GaussianShader, we address this problem by introducing a new normal representation, which is based on the shortest axis direction of a Gaussian sphere and learns a normal residual on this axis direction. Then, to enforce consistency between the estimated normals and the geometry formulated by Gaussian spheres, we introduce an efficient normal-geometry constraint between the predicted normals and the normals derived from the rendered depths. + +Both the normal representation and the constraint lead to an accurate normal estimation of Gaussian spheres, which helps us compute the shading function. + +Built upon 3D Gaussian Splatting, GaussianShader maintains real-time rendering speed while still accommodating various materials like reflective surfaces. Experiments show that, compared to prior works, our method keeps a good balance between efficiency and robust performance on both general scenes and reflective surfaces. In summary, our method offers several significant advantages: + +1. Our method explicitly approximates the rendering equation by a simplified shading function, significantly enhancing the realism of rendered scenes, particularly for highly specular and reflective surfaces. +2. We propose a new normal estimation framework on 3D Gaussians with a new regularization loss that allows precise normal estimation. +3. Leveraging the efficiency of Gaussian Splatting, our approach provides real-time rendering capabilities, making it suitable for interactive applications and scenarios that demand efficient rendering. + +# 2. Related Work + +# 2.1. Neural Radiance Fields + +Neural Radiance Fields (NeRF) [32] gains remarkable progress in photo-realistic novel view synthesis using implicit representation and volume rendering. Recently, NeRF has inspired many follow-up works in various directions. [5, 6] improves NeRF in rendering quality by introducing 3D conical frustum, achieving state-of-art performance in NVS. [17, 36, 47, 52, 53] combine implicit surface representations with NeRF for more accurate geometric reconstruction. [43, 44] propose solutions for city-scale scene rendering using NeRF. [41, 45, 51, 57] targets a special kind of scenes with high specularities and reflections. Another important line of works [10, 14, 15, 28, 33, 42] focuses on acceleration due to the low training and rendering speed of NeRF using voxel grid or hash table. Although great progress has been made, NeRF-based methods still suffer from low rendering speed and high training-time memory usage due to their implicit nature [11]. Based on works on point-based neural rendering [2, 39, 50, 55], a recent milestone work [21] introduces anisotropic 3D Gaussian as an effective representation of the scene, and renders the image using a fast tile-based differentiable rasterizer, surpassing existing implicit neural representation methods in both quality and efficiency. + +# 2.2. Reflective Object Rendering + +Rendering views of reflective objects from multi-view images has been a challenging task due to the complex light interactions. Previous approaches rely on simple light field + +![](images/cf80fa45aca6f1450b54c5a979828213ba1f8a07fe344c948dbeff6ed4bb0504.jpg) +Figure 2. GaussianShader initiates with the neural 3D Gaussian spheres that integrate both conventional attributes and the newly introduced shading attributes to accurately capture view-dependent appearances. We incorporate a differentiable environment lighting map to simulate realistic lighting. The end-to-end training leads to a model that reconstructs both reflective and diffuse surfaces, achieving high material and lighting fidelity. + +interpolations to achieve high-fidelity rendering of novel perspectives [16, 25, 49], yet were constrained by the necessity for dense discrete captures. The accurate rendering of reflective surfaces hinges upon the precise estimation of scene illumination (e.g. environment light) and material properties (e.g. BRDF), which is the task of inverse rendering [4, 35]. Previous studies have demonstrated methodologies for predicting BRDF under known lighting conditions [1, 13, 31], as well as techniques for estimating lighting given known geometry [24, 37, 38]. Furthermore, [12, 19, 51, 58] incorporate indirect illumination, thereby enhancing the fidelity of the estimated BRDF. + +Some NeRF related works attempt to model reflectance by disentangling the visual appearance into lighting and material properties, such as [7-9, 41, 54, 57], which can jointly predict environmental illumination and surface reflectance properties under unknown or varying lighting conditions. Ref-NeRF [45] introduces a new parameterization and structuring of view-dependent outgoing radiance, as well as a regularizer on normal vectors. Recent works [26, 27, 29] utilize SDF-based representation to learn geometry from high specular surfaces, obtaining more accurate normals for physically based rendering. However, these method suffer from extremely time-consuming optimization and slow rendering speed, which hinders their employments in real applications. + +# 2.3. Preliminaries + +# 2.3.1 3D Gaussian Splitting Rasterization + +Our method builds upon Gaussian Splatting [21], which begins with a collection of images capturing a static scene, their corresponding camera parameters, and a sparse point cloud generated through Structure-from-Motion (SfM) [40]. These points construct a set of Gaussians, each defined by position (mean) $\mathbf{p}$ and a 3D covariance matrix $\Sigma$ . While a direct optimization of the covariance matrix $\Sigma$ might seem intuitive, it presents challenges due to the requirement of positive semi-definiteness. As an alternative, in Gaussian Splatting, a more intuitive yet + +equally expressive representation is adopted as an ellipsoid for optimization. The ellipsoid is separated into scaling and rotation components, represented by a scaling matrix $S$ and a rotation matrix $R$ , i.e., $\Sigma = RSS^T R^T$ . + +# 2.3.2 Rendering with 3D Gaussians + +By projecting to 2D frame space according to the camera parameters, 3D Gaussian spheres are both differentiable and amenable to rapid rendering through 2D splatting with $\alpha$ -blending. For radiance field modeling, the directional appearance aspect (color) $\mathbf{c}$ is conveyed through spherical harmonics (SH). By tile-based rasterization, we could sum up the pixel color $\mathbf{C}$ after sorting $\mathbf{c}_i$ by depth: + +$$ +\mathbf {C} = \sum_ {i \in N} \mathbf {c} _ {i} \alpha_ {i} \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {j}) \tag {1} +$$ + +where $\alpha_{i}$ is obtained by multiplying Gaussian weight with opacity $\alpha$ associated to Gaussian sphere. After rasterization, the color loss could be applied to the rendered image: + +$$ +\mathcal {L} _ {\mathrm {c o l o r}} = \left| \left| \mathbf {C} - \mathbf {C} _ {\mathrm {g t}} \right| \right| ^ {2} \tag {2} +$$ + +# 3. Method + +The overview of our method is depicted in Fig. 2. Our approach begins by adopting 3D Gaussian spheres, encompassing shape attributes including covariance $\Sigma$ , opacity $\alpha$ , and position $\mathbf{p}$ . To enhance the representation ability on reflections, we compute the appearances of these Gaussian spheres with a shading function, which requires a set of shading attributes, including diffuse color, roughness, specular tint, normal, and residual color, as detailed in Sec. 3.1. Subsequently, we employ a differentiable environment light map to model the direct lighting, as elaborated in Sec. 3.2. The shading process relies significantly on accurate normal estimation, as discussed in Sec. 3.3. Finally, we introduce the losses used in the whole training process in Sec. 3.4. + +# 3.1. Shading on 3D Gaussians + +Gaussian Splatting [21] models the appearances of Gaussians with simple spherical harmonic functions without considering the light-surface interactions. Thus, Gaussian Splatting fails to accurately represent strong specular surfaces. However, accurately considering the light-surface interactions requires an exact evaluation of the Rendering Equation [20], which requires extensive computational time and complex BRDF parameters. We adopt a simplified approximation of the rendering equation which enables us to achieve high-quality rendering results on reflective surfaces in a considerably shorter time. + +Specifically, for a Gaussian sphere, its rendered color $\mathbf{c}$ for the viewing direction $\omega_{o}$ is computed by + +$$ +\mathbf {c} \left(\omega_ {o}\right) = \gamma \left(\mathbf {c} _ {d} + \mathbf {s} \odot L _ {s} \left(\omega_ {o}, \mathbf {n}, \rho\right) + \mathbf {c} _ {r} \left(\omega_ {o}\right)\right), \tag {3} +$$ + +where $\gamma$ is a gamma tone mapping function [3], $\mathbf{c}_d\in [0,1]^3$ is the diffuse color of this Gaussian sphere, $\mathbf{s}\in [0,1]^3$ is the specular tint defined on this sphere, $L_{s}(\omega_{o},\mathbf{n},\rho)$ is the direct specular light for this sphere in this direction, $\mathbf{n}$ is the normal of this Gaussian sphere, $\rho \in [0,1]$ is the roughness of the sphere, $\mathbf{c}_r:\mathbb{R}^3\to \mathbb{R}^3$ is so-called residual colors, and $\odot$ is the element-wise multiplication. + +Explanations on Eq. 3. We explain our motivation of this shading model in the following three aspects. a) Diffuse color $\mathbf{c}_d$ represents the consistent colors of this Gaussian sphere, which do not change with viewing directions. b) $\mathbf{s} \odot L_s(\omega_o, \mathbf{n}, \rho)$ describes the interactions between the surface intrinsic color $\mathbf{s}$ and the direct specular light $L_s$ . This term enables us to represent most of the reflections in rendering. c) Since there are still some reflections that cannot be explained by the above reflections of direct lights, such as scattering and reflection on indirect lights, we add a residual color term $\mathbf{c}_r(\omega_o)$ to account for these complex reflections. In comparison, Ref-NeRF [45] adopts a similar shading function without such a residual color term, which makes it struggle to handle advanced complex reflections. $\mathbf{c}_r(\omega_o)$ is parameterized by spherical harmonic functions. d) $\mathbf{c}_d$ , $\mathbf{s}$ , $\rho$ , and the coefficients of the spherical harmonics in $\mathbf{c}_r(\omega_o)$ all are trainable parameters associated with this Gaussian sphere. In the following, we will introduce how to compute $L_s(\omega_o, \mathbf{n}, \rho)$ , and $\mathbf{n}$ . + +# 3.2. Specular Light + +We compute the specular light $L_{s}$ by integrating incoming radiance with the specular GGX [46] Normal Distribution Function $D$ visualized in Fig. 3 + +$$ +L _ {s} (\omega_ {o}, \mathbf {n}, \rho) = \int_ {\Omega} L (\omega_ {i}) D (\mathbf {r}, \rho) (\omega_ {i} \cdot \mathbf {n}) d \omega_ {i}, \qquad (4) +$$ + +where $\Omega$ represents the whole upper semi-sphere, $\omega_{i}$ is the direction for the input radiance and $D$ characterizes the + +![](images/e9c77fddb150afd8d57161c042d8e52ccd0296a8e4962286773e0228e8ef8f00.jpg) +Figure 3. Normal Distribution Function D in Eq. 4 is determined by the roughness $\rho$ and reflective direction $\mathbf{r}$ . Surface with small $\rho$ has a smaller specular lobe and that with large $\rho$ has a larger specular lobe. + +specular lobe (effective integral range). When the surface is rough, the specular lobe will be larger around the reflective direction $\mathbf{r}$ while if the surface is smooth, the specular lobe will be smaller. The reflection direction $\mathbf{r}$ is calculated by view direction $\omega_{o}$ and normal $\mathbf{n}$ using $\mathbf{r} = 2(\omega_{o} \cdot \mathbf{n})\mathbf{n} - \omega_{o}$ . In our approach, environment light $L(\omega_{i})$ is represented by a trainable $6 \times 64 \times 64$ cube map. All pixel values in our lighting maps are uniformly initialized to 0.8 (within a range of 0 to 1). Note that the lighting map is global, with each scene being optimized using a single lighting map. + +The light modeling is based on NVDiffRec [34]. Specifically, we begin by preprocessing the environment lighting map into multiple mipmap levels, each storing the integrated light across different reflective directions and roughness. When calculating light integration, we interpolate between these mipmaps based on the input of specific reflection directions and roughness, to obtain the required light integrals. This mipmap-based framework, in contrast to Ref-NeRF's [45] approach using integrated directional encoding, enhances training efficiency. The reflection direction $r$ is calculated by combining the observation direction $\omega_{o}$ with the normal $n$ . + +# 3.3. Normal Estimation + +Normal estimation on Gaussian spheres is difficult. Gaussian spheres are a collection of discrete entities, each representing a localized point in space without a continuous surface or defined edge. This discrete structure makes it inherently difficult to directly calculate normals, which typically requires a continuous surface. In the following, we observed that the shortest axis direction of a Gaussian can serve as an approximated normal and we further associate a predicted normal residual on it. + +Shortest axis direction. In experiments, we made an interesting observation that the aspect ratio of the 3D Gaussian sphere—specifically, the ratios of their longest, intermediate, and shortest axis—gradually increased during optimization, as shown in Fig. 4. This observation suggests that Gaussian spheres gradually become more flattened and approach a planar shape. This observation inspires us to select the shortest axis as the normal of this "planarized" Gaussian sphere, denoted by $\mathbf{v}$ . + +![](images/1d0aa677c70ebf18f9b0caca1a4919505c99916197b23b36f0bb357cf77b1bb3.jpg) +Figure 4. The geometric evolving process of a 3D Gaussian sphere in the optimization, which gradually becomes planar. + +Predicted normal residual. The shortest axis $\mathbf{v}$ only serves as an approximated normal. To make the normal computation more accurate, we further introduce a trainable normal residual $\Delta \mathbf{n}$ on every Gaussian sphere. However, the orientation of the shortest axis $\mathbf{v}$ has an ambiguity because the direction of the shortest axis could either point outward or inward from the surface. To handle this ambiguity, we optimize two separate normal residuals to accommodate both scenarios. Given a specific viewing direction $\omega_{o}$ , we first select the direction aligned with the viewing direction $\omega_{\mathrm{o}}$ as the active normal direction for this viewing direction and then apply the corresponding normal residual to the active normal. This process is described by + +$$ +\mathbf {n} = \left\{ \begin{array}{l l} \mathbf {v} + \Delta \mathbf {n} _ {1} & \text {i f} \omega_ {\mathbf {o}} \cdot \mathbf {v} > 0, \\ - (\mathbf {v} + \Delta \mathbf {n} _ {2}) & \text {o t h e r w i s e .} \end{array} \right. \tag {5} +$$ + +To prevent the normal residual from deviating too much from the shortest axis, we add a penalty towards normal residual, making sure it is small enough. + +$$ +\mathcal {L} _ {\text {r e g}} = \left\| \Delta \mathbf {n} \right\| ^ {2} \tag {6} +$$ + +The normal residual is shown on the left of Fig. 5. + +Normal-geometry consistency. The above shortest axis direction and normal residuals are defined on each Gaussian sphere separately. However, a noticeable problem is that a normal reveals the gradient of the local geometry, which is supposed to be associated with all the Gaussian spheres in a local region. We find that simply applying a color loss to train the aforementioned normal residuals leads to inconsistency between the local geometry and the estimated normals. The main reason is that every Gaussian sphere learns its normal residuals separately without knowing the local geometry formulated by its neighbor Gaussian spheres. Thus, we have to correlate multiple Gaussian spheres in a local region with their normals to ensure normal-geometry consistency. A straightforward and naive solution is to search for its $K$ neighborhoods in space and estimate a coarse normal from all the neighboring spheres. However, such KNN searching would be extremely expensive during training because all the Gaussian spheres are dynamically moving in the optimization process. Instead, we propose a simple yet effective way to ensure normal-geometry consistency as follows. + +![](images/e13ffe9690f31500f8555ac92c41357f6d5f57dcc1bcf22260ee48d6f49ad889.jpg) +Figure 5. Visualization of the relationship between shortes axis $\mathbf{v}$ , normal residual $\Delta \mathbf{n}$ , normal $\mathbf{n}$ and depth-gradient normal $\hat{\mathbf{n}}$ . The supervision $\mathcal{L}_{normal}$ enforces the normal-geometry consistency. + +We associate the local geometry with predicted normals by minimizing the difference between the grad normals derived from the rendered depth map and the rendered normal maps using the predicted normals + +$$ +\mathcal {L} _ {\text {n o r m a l}} = \left| \left| \bar {\mathbf {n}} - \hat {\mathbf {n}} \right| \right| ^ {2}, \tag {7} +$$ + +where $\bar{\mathbf{n}}$ is the rendered normal map and $\hat{\mathbf{n}}$ is computed by applying the Sobel-like operator on the rendered depth map. $\hat{\mathbf{n}}$ reveals the local geometry formulated by multiple Gaussian spheres because it is computed from the rendered depth maps. $\bar{\mathbf{n}}$ contains the information from the separately-defined normal on each Gaussian sphere. By minimizing the difference, we enforce the consistency between the local geometry and the estimated normals. + +# 3.4. Losses + +In addition to $\mathcal{L}_{\mathrm{color}}$ , $\mathcal{L}_{reg}$ , and $\mathcal{L}_{\mathrm{normal}}$ , we make use of sparse loss [30, 50] to encourages Gaussian spheres' opacity values $\alpha$ to approach either 0 or 1 by + +$$ +\mathcal {L} _ {\text {s p a r s e}} = \frac {1}{| \alpha |} \sum_ {\alpha_ {i}} \left[ \log \left(\alpha_ {i}\right) + \log \left(1 - \alpha_ {i}\right) \right]. \tag {8} +$$ + +This sparsity loss helps the geometry of Gaussian spheres converge to a single thin plate and improves the rendering quality. In summary, the total training loss $\mathcal{L}$ is + +$$ +\mathcal {L} = \mathcal {L} _ {\text {c o l o r}} + \lambda_ {\mathrm {n}} \mathcal {L} _ {\text {n o r m a l}} + \lambda_ {\mathrm {s}} \mathcal {L} _ {\text {s p a r s e}} + \lambda_ {\mathrm {r}} \mathcal {L} _ {\text {r e g}} \tag {9} +$$ + +where $\lambda_{\mathrm{n}} = 0.01, \lambda_{\mathrm{s}} = 0.001, \lambda_{\mathrm{r}} = 0.001$ + +# 4. Experiments + +# 4.1. Datasets + +To comprehensively validate the effectiveness of GaussianShader, we conduct evaluation on various datasets: a) widely-used NVS dataset: NeRF Synthetic [32]. b) reflective objects datasets: Shiny Blender [45] and Glossy Synthetic [29]. c) real-world large-scale scenes: Tanks and Temples [23]. + +# 4.2. Baselines and Metrics + +We compare our method against the following baselines: a) 3D Gaussian Splatting [21]: a real-time radiance field rendering method based on efficient 3D Gaussian representation; b) VolSDF [52]: a classical neural implicit surface + +Table 1. The quantitative comparisons (PSNR / SSIM / LPIPS) on NeRF Synthetic dataset [32]. + +
ChairDrumsNeRF Synthetic [32]
LegoMicMaterialsShipHotdogFicusAvg.
PSNR↑
NeRF [32]33.0025.0132.5432.9129.6228.6536.1830.1331.01
VolSDF [52]30.5720.4329.4630.5329.1325.5135.1122.9127.96
Ref-NeRF [45]33.9825.4335.1033.6527.1029.2437.0428.7431.29
ENVIDR [27]31.2222.9929.5532.1729.5221.5731.4426.6028.13
Gaussian Splitting [21]35.8226.1735.6935.3430.0030.8737.6734.8333.30
Ours35.8326.3635.8735.2330.0730.8237.8534.9733.38
SSIM↑
NeRF [32]0.9670.9250.9610.9800.9490.8560.9740.9640.947
VolSDF [52]0.9490.8930.9510.9690.9540.8420.9720.9290.932
Ref-NeRF [45]0.9740.9290.9750.9830.9210.8640.9790.9540.947
ENVIDR [27]0.9760.9300.9610.9840.9680.8550.9630.9870.956
Gaussian Splitting [21]0.9870.9540.9830.9910.9600.9070.9850.9870.969
Ours0.9870.9490.9830.9910.9600.9050.9850.9850.968
LPIPS↓
NeRF [32]0.0460.0910.0500.0280.0630.2060.1210.0440.081
VolSDF [52]0.0560.1190.0540.1910.0480.1910.0430.0680.096
Ref-NeRF [45]0.0290.0730.0250.0180.0780.1580.0280.0560.058
ENVIDR [27]0.0310.0800.0540.0210.0450.2280.0720.0100.067
Gaussian Splitting [21]0.0120.0370.0160.0060.0340.1060.0200.0120.030
Ours0.0120.0400.0140.0060.0330.0980.0190.0130.029
+ +![](images/e92b80a024832da86c0d9ef310a11b1dae0625e102480f157df6397aea4aa98c.jpg) +Figure 6. The qualitative comparisons with 3D Gaussian Splitting [21] on Glossy dataset [29]. Our method not only renders the objects with high fidelity but also provides detailed normal and lighting maps. Some areas are zoomed in for better visualization. + +![](images/61a2d2e5f090b2e4d4c12f64b600f566adc3a190f10162b2ecb79458930e0007.jpg) + +reconstruction method based on SDF; c) Ref-NeRF [45]: state-of-the-art method in novel view synthesis of reflective objects; d) NVDiffRec [34], NVDiffRecMC [18]: top-performing neural inverse rendering methods; e) ENVIDR [27], NeRO [29]: top-performing SDF-based neural implicit methods for reconstructing reflective objects. The evaluation metrics to measure rendering quality are reported in PSNR, SSIM [48], and LPIPS [56]. + +# 4.3. Implementation Details + +All experiments are conducted on a Nvidia RTX 3090 graphics card. We optimize our models using Adam [22] optimizer for 30,000 iterations. For Ref-NeRF [45], ENVIDR [27], and 3D Gaussian Splitting [21], we retrain + +them using their official codes keeping all the original settings (training iterations, sampled rays, etc.) and obtain results from the retrained models. Other works' results are imported from their original papers. + +# 4.4. Comparisons + +NeRF Synthetic dataset [32]. We first evaluate our model on the NeRF Synthetic dataset, which contains objects with complex geometry and realistic non-Lambertian materials. We show the quantitative results in Tab. 1 and visual comparisons in Fig. 8. Our approach achieves numerically and visually comparable results with both Gaussian Splatting and neural rendering methods, demonstrating the effectiveness of our method in rendering general objects. + +![](images/185816575b056867e8c1cca272e0fd3ffa0097de86d7c017fa8034a35e2e3741.jpg) +Figure 7. Qualitative comparisons on Shiny Blender dataset [45]. Error maps corresponding to each rendering result demonstrate the enhanced fidelity of our approach in highly specular areas. + +Note that SDF-based ENVIDR [27] cannot perform well on shadowed regions of complex objects, as shown in Fig. 8. + +Shiny Blender dataset [45]. Quantitative results on the Shiny Blender dataset are reported in Tab. 2. With the ability to model light-surface interactions, our method outperforms original Gaussian Splatting in all scenes while slightly underperforms ENVIDR and Ref-NeRF. ENVIDR capitalizes on the continuous properties of Signed Distance Fields to naturally create smoother surfaces and normals, resulting in high rendering quality, particularly for reflective objects. Fig. 7 shows our comparisons with Gaussian Splatting. We can see that our method correctly renders surfaces with strong specular appearances. High-quality modeling of reflections is contingent upon the accurate estimation of normals as shown in Fig. 9. + +Additionally, we report the average PSNR, training time, and rendering FPS in Tab. 3. The reported results are averaged among all objects of the Shiny Blender and NeRF synthetic datasets. On the same hardware, our method only takes about 0.5 hour for training while MLP-based methods, like ENVIDR and Ref-NeRF, require 6 hours and 23 hours for optimizing. Due to the introduce of extra lighting and shading attributes, our method is a bit slower than 3D Gaussian Splatting [21] but still keeps reasonable efficiency and real-time rendering speed. + +Tanks and Temples dataset [23]. To explore our method's scalability in larger-scale environments instead of only small objects, we conducted experiments using Tanks and Temples [23]. The qualitative results in Fig. 10 show that our model outperforms Gaussian Splatting [21]. + +Glossy Synthetic dataset [29]. We present a comparative visualization between Gaussian Splatting and our method on the Glossy dataset shown in Fig. 6. Our method not only renders objects with exceptional fidelity but also provides more accurate normal and lightings, capturing the nuances of reflective surfaces with greater precision. These maps reveal our method's ability to approximate the rendering function, resulting in a more lifelike portrayal of specular lighting. Please refer to the supplementary materials provided for an extensive showcase of our results. + +Table 2. Qualitative comparisons on Shiny Blender dataset [45]. Our method is comparable with both Gaussian Splatting [21] and prior reflective object reconstruction methods. + +
Shiny Blender [45]
CarBallHelmetTeapotToasterCoffeeAvg.
PSNR↑
NVDiffRec [34]27.9821.7726.9740.4424.3130.7428.70
NVDiffMC [18]25.9330.8526.2738.4422.1829.6028.88
Ref-NeRF [45]30.4129.1429.9245.1925.2933.9932.32
NeRO [29]25.5330.2629.2038.7026.4628.8929.84
ENVIDR [27]28.4638.8932.7341.5926.1129.4832.88
Gaussian Splitting [21]27.2427.6928.3245.6820.9932.3230.37
Ours27.9030.9828.3245.8626.2132.3931.94
SSIM↑
NVDiffRec [34]0.9630.8580.9510.9960.9280.9730.945
NVDiffMC [18]0.9400.9400.9400.9950.8860.9650.944
Ref-NeRF [45]0.9490.9560.9550.9950.9100.9720.956
NeRO [29]0.9490.9740.9710.9950.9290.9560.962
ENVIDR [27]0.9610.9910.9800.9960.9390.9490.969
Gaussian Splitting [21]0.9300.9370.9510.9960.8950.9710.947
Ours0.9310.9650.9500.9960.9290.9710.957
LPIPS↓
NVDiffRec [34]0.0450.2970.1180.0110.1690.0760.119
NVDiffMC [18]0.0770.3120.1570.0140.2250.0970.147
Ref-NeRF [45]0.0510.3070.0870.0130.1180.0820.109
NeRO [29]0.0740.0940.0500.0120.0890.1100.072
ENVIDR [27]0.0490.0670.0510.0110.1160.1390.072
Gaussian Splitting [21]0.0470.1610.0790.0070.1260.0780.083
Ours0.0450.1210.0760.0070.0790.0780.068
+ +Table 3. Comparisons over average PSNR, training time and FPS. Our method achieves the best rendering quality and competitive training/rendering speed. + +
PSNRTraining TimeFPS
Ref-NeRF [45]31.7323h0.03
ENVIDR [27]30.046h1.33
Gaussian Splitting [21]32.050.25h274
Ours32.760.58h97
+ +# 4.5. Ablation Study + +In this section, we conduct ablation studies of our model on Shiny Blender dataset. The numerical results are reported in Tab. 4. All ablation experiments are done on Shiny Blender dataset with $1/2$ image resolution. + +Loss functions. We trained our model serperately under the setting "without sparse loss $\mathcal{L}_{\mathrm{sparse}}$ and "without normal loss $\mathcal{L}_{\mathrm{normal}}$ . The incorporation of these losses, especially the direct regularization on normals, proved critical, yielding a marked improvement in rendering quality by effectively guiding the normal optimization process as shown in Tab. 4. + +Residual color. Residual color, comprising 3rd order SH coefficients, is used to compensate for indirect light. As shown in Table 4, the introduce of residual color enhances our model, allowing it to simulate a wide array of real-world intricate lighting situations and significantly improve rendering quality. + +Normal formulation. In our full model, we use the shortest axis direction $\mathbf{v}$ with a learnable residual $\Delta \mathbf{n}$ to formulate normal, as described by Eqn. 3.3. In Tab. 4, we compared this novel normal formation with a naive normal prediction manner (indicated by "w/o axis $\mathbf{v}$ ). Instead of + +![](images/5f0b70463b9d7846603de94af9f82cc29de25f203370cd7b39ad97fcf7a07858.jpg) +Figure 8. We present a qualitative comparisons of our method on NeRF Synthetic dataset [32] against previous techniques and corresponding ground truth images from test views. For clarity, we zoom in the images where differences in quality are especially notable. + +![](images/baf25c508b4a5b28b3951f669dea5a80c2968e1692cf22e757ac083b655eda9e.jpg) +Figure 9. Visualized normal results on Shiny Blender dataset [45], compared with Gaussian Splatting [21]. It illustrates the superior normal estimation achieved by our method. Table 4. Ablation studies on loss function, lighting representation, normal axis and residual color using Shiny Blender dataset. + +
PSNR↑SSIM↑LPIPS↓
w/o Lsparse31.790.9520.056
w/o Lnormal30.930.9410.060
w/o cr31.490.9480.060
w/o v31.470.9510.058
MLP Lighting29.730.9360.075
Full Model32.090.9530.054
+ +correlating with the geometry of 3D gaussians, we directly optimize a single normal attribute. The result shows our full model can achieve higher rendering quality, since our predicted normals are correlated with the true geometry of Gaussian spheres. + +Lighting representation. A crucial component of our approach is the modeling of lighting. "MLP Lighting" means we use a directional MLP similar to Ref-NeRF to implicitly optimize lighting. In our full model, we use "Envmap Lighting" representation as illustrated in Sec. 3.2. The result proves that our explicit lighting representation is better suited for 3D Gaussian rendering framework to cor + +![](images/a60dd242baef56b0f2a97f39aedd39f9bd9988886242d480f46e847e5bbe7046.jpg) +Figure 10. Qualitative comparisons with Gaussian Splitting [21] on the Tanks and Temples dataset [23]. Our approach yields a smoother and more reasonable normal. + +rectly render reflective objects. + +# 5. Conclusions + +In summary, we propose GaussianShader, advancing the rendering of both reflective and general objects through an extended 3D Gaussian model. Specifically, our method integrates shading functions with 3D Gaussian representation for handling view-dependent appearances, successfully handling reflective surfaces. Moreover, a novel normal prediction method is proposed to achieve high-quality normals, which correlates the shading parameters and true geometry of 3D Gaussians. The results show our approach achieves high-quality renderings, marking a substantial improvement in efficient, realistic 3D object rendering. + +# References + +[1] Miika Aittala, Timo Aila, and Jaakko Lehtinen. Reflectance modeling by neural texture synthesis. ACM Transactions on Graphics (ToG), 35(4):1-13, 2016. 3 +[2] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 696-712. Springer, 2020. 2 +[3] Matthew Anderson, Ricardo Motta, Srinivasan Chandrasekar, and Michael Stokes. Proposal for a standard default color space for the internet—srgb. In Color and imaging conference, pages 238–245. Society for Imaging Science and Technology, 1996. 4 +[4] Jonathan T Barron and Jitendra Malik. Shape, illumination, and reflectance from shading. IEEE transactions on pattern analysis and machine intelligence, 37(8):1670-1687, 2014. 3 +[5] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 2 +[6] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470-5479, 2022. 1, 2 +[7] Sai Bi, Zexiang Xu, Pratul Srinivasan, Ben Mildenhall, Kalyan Sunkavalli, Milos Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Neural reflectance fields for appearance acquisition. arXiv preprint arXiv:2008.03824, 2020. 3 +[8] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T Barron, Ce Liu, and Hendrik Lensch. Nerd: Neural reflectance decomposition from image collections. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12684-12694, 2021. +[9] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan Barron, and Hendrik Lensch. Neural-pil: Neural pre-integrated lighting for reflectance decomposition. Advances in Neural Information Processing Systems, 34: 10691-10704, 2021. 3 +[10] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 2 +[11] Zilong Chen, Feng Wang, and Huaping Liu. Text-to-3d using gaussian splatting. arXiv preprint arXiv:2309.16585, 2023. 2 +[12] Youming Deng, Xueting Li, Sifei Liu, and Ming-Hsuan Yang. Dip: Differentiable interreflection-aware physics-based inverse rendering. arXiv preprint arXiv:2212.04705, 2022.3 +[13] Valentin Deschaintre, Miika Aittala, Fredo Durand, George Drettakis, and Adrien Bousseau. Single-image svbrdf cap + +ture with a rendering-aware deep network. ACM Transactions on Graphics (ToG), 37(4):1-15, 2018. 3 +[14] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 2 +[15] Stephan J Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14346-14355, 2021. 2 +[16] Steven J Gortler, Radek Grzesczuk, Richard Szeliski, and Michael F Cohen. The lumigraph. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 453-464. 2023. 3 +[17] Haoyu Guo, Sida Peng, Haotong Lin, Qianqian Wang, Guofeng Zhang, Hujun Bao, and Xiaowei Zhou. Neural 3d scene reconstruction with the manhattan-world assumption. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5511-5520, 2022. 2 +[18] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, Light, and Material Decomposition from Images using Monte Carlo Rendering and Denoising. arXiv:2206.03380, 2022. 6, 7 +[19] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, light, and material decomposition from images using monte carlo rendering and denoising. Advances in Neural Information Processing Systems, 35:22856-22869, 2022. 3 +[20] James T Kajiya. The rendering equation. In Proceedings of the 13th annual conference on Computer graphics and interactive techniques, pages 143-150, 1986. 4 +[21] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 42(4):1-14, 2023. 1, 2, 3, 4, 5, 6, 7, 8 +[22] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[23] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36 (4):1-13, 2017. 5, 7, 8, 1 +[24] Chloe LeGendre, Wan-Chun Ma, Graham Fyffe, John Flynn, Laurent Charbonnel, Jay Busch, and Paul Debevec. Deeplight: Learning illumination for unconstrained mobile mixed reality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5918-5928, 2019. 3 +[25] Marc Levoy and Pat Hanrahan. Light field rendering. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 441-452. 2023. 3 +[26] Ruofan Liang, Jiahao Zhang, Haoda Li, Chen Yang, Yushi Guan, and Nandita Vijaykumar. Spidr: Sdf-based neural point fields for illumination and deformation. arXiv preprint arXiv:2210.08398, 2022. 3 + +[27] Ruofan Liang, Huiting Chen, Chunlin Li, Fan Chen, Selvakumar Panneer, and Nandita Vijaykumar. Envidr: Implicit differentiable renderer with neural environment lighting. arXiv preprint arXiv:2303.13022, 2023. 1, 2, 3, 6, 7 +[28] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 2 +[29] Yuan Liu, Peng Wang, Cheng Lin, Xiaoxiao Long, Jiepeng Wang, Lingjie Liu, Taku Komura, and Wenping Wang. Nano: Neural geometry and brdf reconstruction of reflective objects from multiview images. arXiv preprint arXiv:2305.17398, 2023. 3, 5, 6, 7, 1 +[30] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural volumes: Learning dynamic renderable volumes from images. arXiv preprint arXiv:1906.07751, 2019. 5 +[31] Wojciech Matusik. A data-driven reflectance model. PhD thesis, Massachusetts Institute of Technology, 2003. 3 +[32] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 5, 6, 8 +[33] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 2 +[34] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3d models, materials, and lighting from images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8280-8290, 2022. 4, 6, 7 +[35] Merlin Nimier-David, Delio Vicini, Tizian Zeltner, and Wenzel Jakob. Mitsuba 2: A retargetable forward and inverse renderer. ACM Transactions on Graphics (TOG), 38(6):1-17, 2019. 3 +[36] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5589-5599, 2021. 2 +[37] Jeong Joon Park, Aleksander Holynski, and Steven M Seitz. Seeing the world in a bag of chips. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1417-1427, 2020. 3 +[38] Thomas Richter-Trummer, Denis Kalkofen, Jinwoo Park, and Dieter Schmalstieg. Instant mixed reality lighting from casual scanning. In 2016 IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 27-36. IEEE, 2016. 3 +[39] Darius Rückert, Linus Franke, and Marc Stamminger. Adop: Approximate differentiable one-pixel point rendering. ACM Transactions on Graphics (ToG), 41(4):1-14, 2022. 2 +[40] Noah Snavely, Steven M Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. In ACM siggraph 2006 papers, pages 835-846. 2006. 3 + +[41] Pratul P Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T Barron. Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7495-7504, 2021. 2, 3 +[42] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 2 +[43] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8248–8258, 2022. 2 +[44] Haithem Turki, Deva Ramanan, and Mahadev Satyanarayanan. Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12922–12931, 2022. 2 +[45] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481-5490. IEEE, 2022. 1, 2, 3, 4, 5, 6, 7, 8 +[46] Bruce Walter, Stephen R Marschner, Hongsong Li, and Kenneth E Torrance. Microfacet models for refraction through rough surfaces. In Proceedings of the 18th Eurographics conference on Rendering Techniques, pages 195-206, 2007. 4 +[47] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 2 +[48] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6 +[49] Daniel N Wood, Daniel I Azuma, Ken Aldinger, Brian Curless, Tom Duchamp, David H Salesin, and Werner Stuetzle. Surface light fields for 3d photography. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 487-496. 2023. 3 +[50] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5438–5448, 2022. 2, 5 +[51] Yao Yao, Jingyang Zhang, Jingbo Liu, Yihang Qu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Neilf: Neural incident light field for physically-based material estimation. In European Conference on Computer Vision, pages 700-716. Springer, 2022. 2, 3 +[52] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neu + +ral Information Processing Systems, 34:4805-4815, 2021. 2, 5, 6 +[53] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in neural information processing systems, 35:25018-25032, 2022. 2 +[54] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snavely. Physg: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5453-5462, 2021. 3 +[55] Qiang Zhang, Seung-Hwan Baek, Szymon Rusinkiewicz, and Felix Heide. Differentiable point-based radiance fields for efficient view synthesis. In SIGGRAPH Asia 2022 Conference Papers, pages 1-12, 2022. 2 +[56] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6 +[57] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Ner-factor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (ToG), 40(6):1-18, 2021. 2, 3 +[58] Yuanqing Zhang, Jiaming Sun, Xingyi He, Huan Fu, Rongfei Jia, and Xiaowei Zhou. Modeling indirect illumination for inverse rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18643-18652, 2022. 3 \ No newline at end of file diff --git a/2024/GaussianShader_ 3D Gaussian Splatting with Shading Functions for Reflective Surfaces/images.zip b/2024/GaussianShader_ 3D Gaussian Splatting with Shading Functions for Reflective Surfaces/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f945e24f9b1a4c8ff92eef295f605a64d24cb93e --- /dev/null +++ b/2024/GaussianShader_ 3D Gaussian Splatting with Shading Functions for Reflective Surfaces/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a8780a277ae10bede26a5c5b1b3208419df9019427d8a01d9c4d10fc2b676ccc +size 729425 diff --git a/2024/GaussianShader_ 3D Gaussian Splatting with Shading Functions for Reflective Surfaces/layout.json b/2024/GaussianShader_ 3D Gaussian Splatting with Shading Functions for Reflective Surfaces/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fc30dc6a547c423789be8b63c99bf4d0f6dc8aad --- /dev/null +++ b/2024/GaussianShader_ 3D Gaussian Splatting with Shading Functions for Reflective Surfaces/layout.json @@ -0,0 +1,9206 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 97, + 102, + 496, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 102, + 496, + 138 + ], + "spans": [ + { + "bbox": [ + 97, + 102, + 496, + 138 + ], + "type": "text", + "content": "GaussianShader: 3D Gaussian Splating with Shading Functions for Reflective Surfaces" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 160, + 447, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 160, + 447, + 175 + ], + "spans": [ + { + "bbox": [ + 146, + 160, + 447, + 175 + ], + "type": "text", + "content": "Yingwenqi Jiang" + }, + { + "bbox": [ + 146, + 160, + 447, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 146, + 160, + 447, + 175 + ], + "type": "text", + "content": " Jiadong Tu" + }, + { + "bbox": [ + 146, + 160, + 447, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 146, + 160, + 447, + 175 + ], + "type": "text", + "content": " Yuan Liu" + }, + { + "bbox": [ + 146, + 160, + 447, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 146, + 160, + 447, + 175 + ], + "type": "text", + "content": " Xifeng Gao" + }, + { + "bbox": [ + 146, + 160, + 447, + 175 + ], + "type": "inline_equation", + "content": "^{3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 175, + 175, + 432, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 175, + 432, + 190 + ], + "spans": [ + { + "bbox": [ + 175, + 175, + 432, + 190 + ], + "type": "text", + "content": "Xiaoxiao Long" + }, + { + "bbox": [ + 175, + 175, + 432, + 190 + ], + "type": "inline_equation", + "content": "^{2,*}" + }, + { + "bbox": [ + 175, + 175, + 432, + 190 + ], + "type": "text", + "content": " Wenping Wang" + }, + { + "bbox": [ + 175, + 175, + 432, + 190 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 175, + 175, + 432, + 190 + ], + "type": "text", + "content": " Yuexin Ma" + }, + { + "bbox": [ + 175, + 175, + 432, + 190 + ], + "type": "inline_equation", + "content": "^{1,*}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 193, + 174, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 193, + 174, + 207 + ], + "spans": [ + { + "bbox": [ + 66, + 193, + 174, + 207 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 66, + 193, + 174, + 207 + ], + "type": "text", + "content": "ShanghaiTech University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 186, + 193, + 312, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 193, + 312, + 207 + ], + "spans": [ + { + "bbox": [ + 186, + 193, + 312, + 207 + ], + "type": "text", + "content": "2The University of Hong Kong" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 324, + 193, + 400, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 193, + 400, + 206 + ], + "spans": [ + { + "bbox": [ + 324, + 193, + 400, + 206 + ], + "type": "text", + "content": "3Tencent America" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 412, + 193, + 512, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 412, + 193, + 512, + 207 + ], + "spans": [ + { + "bbox": [ + 412, + 193, + 512, + 207 + ], + "type": "text", + "content": "Texas A&M University" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 47, + 242, + 116, + 296 + ], + "blocks": [ + { + "bbox": [ + 47, + 242, + 116, + 296 + ], + "lines": [ + { + "bbox": [ + 47, + 242, + 116, + 296 + ], + "spans": [ + { + "bbox": [ + 47, + 242, + 116, + 296 + ], + "type": "image", + "image_path": "bd5035014d6f472068fcdb1864689d3d3862a3f53469223ea44f1c5f1c1dbffa.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 297, + 99, + 303 + ], + "lines": [ + { + "bbox": [ + 63, + 297, + 99, + 303 + ], + "spans": [ + { + "bbox": [ + 63, + 297, + 99, + 303 + ], + "type": "text", + "content": "PSNR:21.69" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 120, + 243, + 189, + 295 + ], + "blocks": [ + { + "bbox": [ + 120, + 243, + 189, + 295 + ], + "lines": [ + { + "bbox": [ + 120, + 243, + 189, + 295 + ], + "spans": [ + { + "bbox": [ + 120, + 243, + 189, + 295 + ], + "type": "image", + "image_path": "c69300f47475fd7e15f3521d175106b04c2c1913e4c6bddd1a3c69ca24d31e28.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 135, + 297, + 171, + 303 + ], + "lines": [ + { + "bbox": [ + 135, + 297, + 171, + 303 + ], + "spans": [ + { + "bbox": [ + 135, + 297, + 171, + 303 + ], + "type": "text", + "content": "PSNR: 28.87" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 192, + 243, + 261, + 295 + ], + "blocks": [ + { + "bbox": [ + 192, + 243, + 261, + 295 + ], + "lines": [ + { + "bbox": [ + 192, + 243, + 261, + 295 + ], + "spans": [ + { + "bbox": [ + 192, + 243, + 261, + 295 + ], + "type": "image", + "image_path": "0081ed2f08744dc3bb8ad153208d7331c3d3203dced177e3e0ac530988bcf6ce.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 207, + 297, + 243, + 303 + ], + "lines": [ + { + "bbox": [ + 207, + 297, + 243, + 303 + ], + "spans": [ + { + "bbox": [ + 207, + 297, + 243, + 303 + ], + "type": "text", + "content": "PSNR:32.67" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 264, + 243, + 333, + 296 + ], + "blocks": [ + { + "bbox": [ + 264, + 243, + 333, + 296 + ], + "lines": [ + { + "bbox": [ + 264, + 243, + 333, + 296 + ], + "spans": [ + { + "bbox": [ + 264, + 243, + 333, + 296 + ], + "type": "image", + "image_path": "8a9006e4b11d153459688da78ef59907630788cc29ed695d4dd11e76566e049b.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 279, + 297, + 315, + 303 + ], + "lines": [ + { + "bbox": [ + 279, + 297, + 315, + 303 + ], + "spans": [ + { + "bbox": [ + 279, + 297, + 315, + 303 + ], + "type": "text", + "content": "PSNR:32.46" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 336, + 243, + 405, + 297 + ], + "blocks": [ + { + "bbox": [ + 336, + 243, + 405, + 297 + ], + "lines": [ + { + "bbox": [ + 336, + 243, + 405, + 297 + ], + "spans": [ + { + "bbox": [ + 336, + 243, + 405, + 297 + ], + "type": "image", + "image_path": "e4c9d9ee7e4c248ab9a75b9e99441f9970e9819a7fb9b399da3cf2a903306579.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 342, + 305, + 391, + 374 + ], + "blocks": [ + { + "bbox": [ + 342, + 305, + 391, + 374 + ], + "lines": [ + { + "bbox": [ + 342, + 305, + 391, + 374 + ], + "spans": [ + { + "bbox": [ + 342, + 305, + 391, + 374 + ], + "type": "image", + "image_path": "e6a7bb1b1a0b7da5b62ffd5100d2ee0b3ec3784a7b812a98c7385ac4018d7db2.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 349, + 381, + 389, + 389 + ], + "lines": [ + { + "bbox": [ + 349, + 381, + 389, + 389 + ], + "spans": [ + { + "bbox": [ + 349, + 381, + 389, + 389 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 410, + 251, + 545, + 383 + ], + "blocks": [ + { + "bbox": [ + 410, + 251, + 545, + 383 + ], + "lines": [ + { + "bbox": [ + 410, + 251, + 545, + 383 + ], + "spans": [ + { + "bbox": [ + 410, + 251, + 545, + 383 + ], + "type": "image", + "image_path": "439a08bcd9b263589bd7bc0e52209218cb1f4da2c557044437df4185499cdd5b.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 435, + 386, + 531, + 403 + ], + "lines": [ + { + "bbox": [ + 435, + 386, + 531, + 403 + ], + "spans": [ + { + "bbox": [ + 435, + 386, + 531, + 403 + ], + "type": "text", + "content": "(b) Comparisons between FPS, training time and accuracy" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 403, + 546, + 437 + ], + "lines": [ + { + "bbox": [ + 46, + 403, + 546, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 546, + 437 + ], + "type": "text", + "content": "Figure 1. GaussianShader maintains real-time rendering speed and renders high-fidelity images for both general and reflective surfaces. Ref-NeRF[45] and ENVIDR[27] attempt to handle reflective surfaces, but they suffer from quite time-consuming optimization and slow rendering speed. 3D Gaussian splattering [21] keeps high efficiency but cannot handle such reflective surfaces." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 143, + 445, + 192, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 445, + 192, + 456 + ], + "spans": [ + { + "bbox": [ + 143, + 445, + 192, + 456 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 46, + 458, + 287, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 458, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 458, + 287, + 651 + ], + "type": "text", + "content": "The advent of neural 3D Gaussians [21] has recently brought about a revolution in the field of neural rendering, facilitating the generation of high-quality renderings at real-time speeds. However, the explicit and discrete representation encounters challenges when applied to scenes featuring reflective surfaces. In this paper, we present GaussianShader, a novel method that applies a simplified shading function on 3D Gaussians to enhance the neural rendering in scenes with reflective surfaces while preserving the training and rendering efficiency. The main challenge in applying the shading function lies in the accurate normal estimation on discrete 3D Gaussians. Specifically, we proposed a novel normal estimation framework based on the shortest axis directions of 3D Gaussians with a delicately designed loss to make the consistency between the normals and the geometries of Gaussian spheres. Exper" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 445, + 546, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 546, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 546, + 540 + ], + "type": "text", + "content": "periments show that GaussianShader strikes a commendable balance between efficiency and visual quality. Our method surpasses Gaussian Splatting [21] in PSNR on specular object datasets, exhibiting an improvement of 1.57dB. When compared to prior works handling reflective surfaces, such as Ref-NeRF [45], our optimization time is significantly accelerated (23h vs. 0.58h). Please click on our project website to see more results." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 561, + 387, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 561, + 387, + 573 + ], + "spans": [ + { + "bbox": [ + 306, + 561, + 387, + 573 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 582, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 546, + 715 + ], + "type": "text", + "content": "In recent years, the field of 3D computer vision has witnessed remarkable advancements in the 3D reconstruction and visualization of 3D scenes. Innovations such as Neural Radiance Fields (NeRF) [32] have achieved substantial breakthroughs in generating novel views of 3D objects and scenes, presenting the potential for high-quality and photorealistic renderings. Despite these advancements, NeRF-related methodologies [6, 27, 45], still contend with challenges such as computationally expensive optimization and slow rendering speed. These limitations restrict their application in real-time interactive scenarios." + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 656, + 287, + 715 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 656, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 656, + 287, + 715 + ], + "type": "text", + "content": "*Corresponding author. This work was supported by NSFC (No.62206173), MoE Key Laboratory of Intelligent Perception and Human-Machine Collaboration (ShanghaiTech University), Shanghai Frontiers Science Center of Human-centered Artificial Intelligence, the Innovation and Technology Commission of the HKSAR Government under the InnoHK initiative, and Ref. T45-205/21-N of Hong Kong RGC." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5322" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "More recently, 3D Gaussian Splatting [21] combines 3D Gaussian representation and tile-based splatting techniques to achieve high-quality 3D scene modeling and real-time rendering, making it possible to employ neural rendering techniques in real applications. However, it suffers from a performance drop on scenes featuring specular and reflective surfaces. This is because 3D Gaussian Splatting [21] does not explicitly model appearance properties, so that fails to capture significant view-dependent changes, particularly specular highlights. This constraint presents a substantial obstacle in the pursuit of achieving photorealistic rendering across a diverse array of materials, particularly those characterized by prominent reflective attributes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 230, + 289, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 230, + 289, + 374 + ], + "spans": [ + { + "bbox": [ + 46, + 230, + 289, + 374 + ], + "type": "text", + "content": "Accurately modeling reflective surfaces is a challenging task. Ref-NeRF [45] and ENVIDR [27] combine the shading functions in implicit representations and present promising quality on reflective surfaces. However, they suffer from time-consuming optimization (hours) and slow rendering speed. Due to the limited flexibility of SDF, ENVIDR [27] even fails to model complex scenes and presents a significant performance drop on general objects. It is still an unexplored problem how to combine the shading functions in a 3D Gaussian Splatting framework to improve its ability to handle reflections while preserving the efficiency in training and rendering." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 376, + 289, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 289, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 289, + 544 + ], + "type": "text", + "content": "In this paper, we present GaussianShader, a novel method that enhances the neural rendering of 3D Gaussians within scenes that contain reflective surfaces by incorporating a shading function on 3D Gaussians. To ensure the efficiency of GaussianShader, evaluating the shading function cannot be too expensive while still retaining the ability to model the reflections. In light of this, we propose a novel simplified shading function that considers the diffuse colors and the direct reflections while putting all advanced complex reflections into a residual color term. In comparison with the shading function of Ref-NeRF which can only consider direct reflections, the utilization of this residual color enables GaussianShader to render more complex reflective appearances with efficiency." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": "Another challenge in computing a shading function is how to predict accurate normals on the discrete 3D Gaussian spheres. First, it is hard to get a locally-continuous surface from the 3D Gaussians to compute the surface normals. Second, associating multiple 3D Gaussians for normal computation would be very expensive using neighborhood searching. In GaussianShader, we address this problem by introducing a new normal representation, which is based on the shortest axis direction of a Gaussian sphere and learns a normal residual on this axis direction. Then, to enforce consistency between the estimated normals and the geometry formulated by Gaussian spheres, we introduce an efficient normal-geometry constraint between the predicted normals and the normals derived from the rendered depths." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": "Both the normal representation and the constraint lead to an accurate normal estimation of Gaussian spheres, which helps us compute the shading function." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 108, + 545, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 108, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 304, + 108, + 545, + 192 + ], + "type": "text", + "content": "Built upon 3D Gaussian Splatting, GaussianShader maintains real-time rendering speed while still accommodating various materials like reflective surfaces. Experiments show that, compared to prior works, our method keeps a good balance between efficiency and robust performance on both general scenes and reflective surfaces. In summary, our method offers several significant advantages:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 193, + 545, + 324 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 306, + 193, + 545, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 193, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 306, + 193, + 545, + 240 + ], + "type": "text", + "content": "1. Our method explicitly approximates the rendering equation by a simplified shading function, significantly enhancing the realism of rendered scenes, particularly for highly specular and reflective surfaces." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 240, + 545, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 240, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 306, + 240, + 545, + 275 + ], + "type": "text", + "content": "2. We propose a new normal estimation framework on 3D Gaussians with a new regularization loss that allows precise normal estimation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 276, + 545, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 276, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 306, + 276, + 545, + 324 + ], + "type": "text", + "content": "3. Leveraging the efficiency of Gaussian Splatting, our approach provides real-time rendering capabilities, making it suitable for interactive applications and scenarios that demand efficient rendering." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 336, + 392, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 336, + 392, + 348 + ], + "spans": [ + { + "bbox": [ + 306, + 336, + 392, + 348 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 356, + 440, + 368 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 356, + 440, + 368 + ], + "spans": [ + { + "bbox": [ + 306, + 356, + 440, + 368 + ], + "type": "text", + "content": "2.1. Neural Radiance Fields" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 374, + 545, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 374, + 545, + 650 + ], + "spans": [ + { + "bbox": [ + 304, + 374, + 545, + 650 + ], + "type": "text", + "content": "Neural Radiance Fields (NeRF) [32] gains remarkable progress in photo-realistic novel view synthesis using implicit representation and volume rendering. Recently, NeRF has inspired many follow-up works in various directions. [5, 6] improves NeRF in rendering quality by introducing 3D conical frustum, achieving state-of-art performance in NVS. [17, 36, 47, 52, 53] combine implicit surface representations with NeRF for more accurate geometric reconstruction. [43, 44] propose solutions for city-scale scene rendering using NeRF. [41, 45, 51, 57] targets a special kind of scenes with high specularities and reflections. Another important line of works [10, 14, 15, 28, 33, 42] focuses on acceleration due to the low training and rendering speed of NeRF using voxel grid or hash table. Although great progress has been made, NeRF-based methods still suffer from low rendering speed and high training-time memory usage due to their implicit nature [11]. Based on works on point-based neural rendering [2, 39, 50, 55], a recent milestone work [21] introduces anisotropic 3D Gaussian as an effective representation of the scene, and renders the image using a fast tile-based differentiable rasterizer, surpassing existing implicit neural representation methods in both quality and efficiency." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 658, + 462, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 462, + 672 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 462, + 672 + ], + "type": "text", + "content": "2.2. Reflective Object Rendering" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": "Rendering views of reflective objects from multi-view images has been a challenging task due to the complex light interactions. Previous approaches rely on simple light field" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5323" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 547, + 181 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 547, + 181 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 547, + 181 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 547, + 181 + ], + "type": "image", + "image_path": "cf80fa45aca6f1450b54c5a979828213ba1f8a07fe344c948dbeff6ed4bb0504.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 186, + 547, + 232 + ], + "lines": [ + { + "bbox": [ + 46, + 186, + 547, + 232 + ], + "spans": [ + { + "bbox": [ + 46, + 186, + 547, + 232 + ], + "type": "text", + "content": "Figure 2. GaussianShader initiates with the neural 3D Gaussian spheres that integrate both conventional attributes and the newly introduced shading attributes to accurately capture view-dependent appearances. We incorporate a differentiable environment lighting map to simulate realistic lighting. The end-to-end training leads to a model that reconstructs both reflective and diffuse surfaces, achieving high material and lighting fidelity." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 238, + 287, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 287, + 380 + ], + "type": "text", + "content": "interpolations to achieve high-fidelity rendering of novel perspectives [16, 25, 49], yet were constrained by the necessity for dense discrete captures. The accurate rendering of reflective surfaces hinges upon the precise estimation of scene illumination (e.g. environment light) and material properties (e.g. BRDF), which is the task of inverse rendering [4, 35]. Previous studies have demonstrated methodologies for predicting BRDF under known lighting conditions [1, 13, 31], as well as techniques for estimating lighting given known geometry [24, 37, 38]. Furthermore, [12, 19, 51, 58] incorporate indirect illumination, thereby enhancing the fidelity of the estimated BRDF." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 382, + 288, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 382, + 288, + 549 + ], + "spans": [ + { + "bbox": [ + 46, + 382, + 288, + 549 + ], + "type": "text", + "content": "Some NeRF related works attempt to model reflectance by disentangling the visual appearance into lighting and material properties, such as [7-9, 41, 54, 57], which can jointly predict environmental illumination and surface reflectance properties under unknown or varying lighting conditions. Ref-NeRF [45] introduces a new parameterization and structuring of view-dependent outgoing radiance, as well as a regularizer on normal vectors. Recent works [26, 27, 29] utilize SDF-based representation to learn geometry from high specular surfaces, obtaining more accurate normals for physically based rendering. However, these method suffer from extremely time-consuming optimization and slow rendering speed, which hinders their employments in real applications." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 556, + 134, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 556, + 134, + 567 + ], + "spans": [ + { + "bbox": [ + 47, + 556, + 134, + 567 + ], + "type": "text", + "content": "2.3. Preliminaries" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 574, + 235, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 235, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 235, + 588 + ], + "type": "text", + "content": "2.3.1 3D Gaussian Splitting Rasterization" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": "Our method builds upon Gaussian Splatting [21], which begins with a collection of images capturing a static scene, their corresponding camera parameters, and a sparse point cloud generated through Structure-from-Motion (SfM) [40]. These points construct a set of Gaussians, each defined by position (mean) " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " and a 3D covariance matrix " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": ". While a direct optimization of the covariance matrix " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " might seem intuitive, it presents challenges due to the requirement of positive semi-definiteness. As an alternative, in Gaussian Splatting, a more intuitive yet" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 238, + 547, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 238, + 547, + 286 + ], + "spans": [ + { + "bbox": [ + 305, + 238, + 547, + 286 + ], + "type": "text", + "content": "equally expressive representation is adopted as an ellipsoid for optimization. The ellipsoid is separated into scaling and rotation components, represented by a scaling matrix " + }, + { + "bbox": [ + 305, + 238, + 547, + 286 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 305, + 238, + 547, + 286 + ], + "type": "text", + "content": " and a rotation matrix " + }, + { + "bbox": [ + 305, + 238, + 547, + 286 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 305, + 238, + 547, + 286 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 305, + 238, + 547, + 286 + ], + "type": "inline_equation", + "content": "\\Sigma = RSS^T R^T" + }, + { + "bbox": [ + 305, + 238, + 547, + 286 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 304, + 466, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 304, + 466, + 316 + ], + "spans": [ + { + "bbox": [ + 306, + 304, + 466, + 316 + ], + "type": "text", + "content": "2.3.2 Rendering with 3D Gaussians" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 324, + 547, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 324, + 547, + 409 + ], + "spans": [ + { + "bbox": [ + 305, + 324, + 547, + 409 + ], + "type": "text", + "content": "By projecting to 2D frame space according to the camera parameters, 3D Gaussian spheres are both differentiable and amenable to rapid rendering through 2D splatting with " + }, + { + "bbox": [ + 305, + 324, + 547, + 409 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 305, + 324, + 547, + 409 + ], + "type": "text", + "content": "-blending. For radiance field modeling, the directional appearance aspect (color) " + }, + { + "bbox": [ + 305, + 324, + 547, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 305, + 324, + 547, + 409 + ], + "type": "text", + "content": " is conveyed through spherical harmonics (SH). By tile-based rasterization, we could sum up the pixel color " + }, + { + "bbox": [ + 305, + 324, + 547, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 305, + 324, + 547, + 409 + ], + "type": "text", + "content": " after sorting " + }, + { + "bbox": [ + 305, + 324, + 547, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i" + }, + { + "bbox": [ + 305, + 324, + 547, + 409 + ], + "type": "text", + "content": " by depth:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 370, + 418, + 545, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 418, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 370, + 418, + 545, + 453 + ], + "type": "interline_equation", + "content": "\\mathbf {C} = \\sum_ {i \\in N} \\mathbf {c} _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}) \\tag {1}", + "image_path": "83cca5c99d9575c0f96df6433fc4d89e559806e2b57813647bb023c53f12fb5d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 462, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 462, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 305, + 462, + 545, + 498 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 462, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 305, + 462, + 545, + 498 + ], + "type": "text", + "content": " is obtained by multiplying Gaussian weight with opacity " + }, + { + "bbox": [ + 305, + 462, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 305, + 462, + 545, + 498 + ], + "type": "text", + "content": " associated to Gaussian sphere. After rasterization, the color loss could be applied to the rendered image:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 381, + 510, + 545, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 510, + 545, + 526 + ], + "spans": [ + { + "bbox": [ + 381, + 510, + 545, + 526 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c o l o r}} = \\left| \\left| \\mathbf {C} - \\mathbf {C} _ {\\mathrm {g t}} \\right| \\right| ^ {2} \\tag {2}", + "image_path": "3177592db43a3248e0df431eff293450a03ec47d4656b4c30fdf186114488301.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 536, + 361, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 536, + 361, + 549 + ], + "spans": [ + { + "bbox": [ + 306, + 536, + 361, + 549 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "content": "The overview of our method is depicted in Fig. 2. Our approach begins by adopting 3D Gaussian spheres, encompassing shape attributes including covariance " + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "content": ", opacity " + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "content": ", and position " + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "content": ". To enhance the representation ability on reflections, we compute the appearances of these Gaussian spheres with a shading function, which requires a set of shading attributes, including diffuse color, roughness, specular tint, normal, and residual color, as detailed in Sec. 3.1. Subsequently, we employ a differentiable environment light map to model the direct lighting, as elaborated in Sec. 3.2. The shading process relies significantly on accurate normal estimation, as discussed in Sec. 3.3. Finally, we introduce the losses used in the whole training process in Sec. 3.4." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5324" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 189, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 189, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 189, + 85 + ], + "type": "text", + "content": "3.1. Shading on 3D Gaussians" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 90, + 287, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 90, + 287, + 221 + ], + "spans": [ + { + "bbox": [ + 46, + 90, + 287, + 221 + ], + "type": "text", + "content": "Gaussian Splatting [21] models the appearances of Gaussians with simple spherical harmonic functions without considering the light-surface interactions. Thus, Gaussian Splatting fails to accurately represent strong specular surfaces. However, accurately considering the light-surface interactions requires an exact evaluation of the Rendering Equation [20], which requires extensive computational time and complex BRDF parameters. We adopt a simplified approximation of the rendering equation which enables us to achieve high-quality rendering results on reflective surfaces in a considerably shorter time." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 222, + 287, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 222, + 287, + 247 + ], + "spans": [ + { + "bbox": [ + 47, + 222, + 287, + 247 + ], + "type": "text", + "content": "Specifically, for a Gaussian sphere, its rendered color " + }, + { + "bbox": [ + 47, + 222, + 287, + 247 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 47, + 222, + 287, + 247 + ], + "type": "text", + "content": " for the viewing direction " + }, + { + "bbox": [ + 47, + 222, + 287, + 247 + ], + "type": "inline_equation", + "content": "\\omega_{o}" + }, + { + "bbox": [ + 47, + 222, + 287, + 247 + ], + "type": "text", + "content": " is computed by" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 74, + 255, + 287, + 269 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 255, + 287, + 269 + ], + "spans": [ + { + "bbox": [ + 74, + 255, + 287, + 269 + ], + "type": "interline_equation", + "content": "\\mathbf {c} \\left(\\omega_ {o}\\right) = \\gamma \\left(\\mathbf {c} _ {d} + \\mathbf {s} \\odot L _ {s} \\left(\\omega_ {o}, \\mathbf {n}, \\rho\\right) + \\mathbf {c} _ {r} \\left(\\omega_ {o}\\right)\\right), \\tag {3}", + "image_path": "ab78a0f1d400708cf90c11caea6b4e64b3008b086443dd407dce394daea74467.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 276, + 287, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 276, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 276, + 287, + 361 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 276, + 287, + 361 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 276, + 287, + 361 + ], + "type": "text", + "content": " is a gamma tone mapping function [3], " + }, + { + "bbox": [ + 46, + 276, + 287, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_d\\in [0,1]^3" + }, + { + "bbox": [ + 46, + 276, + 287, + 361 + ], + "type": "text", + "content": " is the diffuse color of this Gaussian sphere, " + }, + { + "bbox": [ + 46, + 276, + 287, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{s}\\in [0,1]^3" + }, + { + "bbox": [ + 46, + 276, + 287, + 361 + ], + "type": "text", + "content": " is the specular tint defined on this sphere, " + }, + { + "bbox": [ + 46, + 276, + 287, + 361 + ], + "type": "inline_equation", + "content": "L_{s}(\\omega_{o},\\mathbf{n},\\rho)" + }, + { + "bbox": [ + 46, + 276, + 287, + 361 + ], + "type": "text", + "content": " is the direct specular light for this sphere in this direction, " + }, + { + "bbox": [ + 46, + 276, + 287, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 46, + 276, + 287, + 361 + ], + "type": "text", + "content": " is the normal of this Gaussian sphere, " + }, + { + "bbox": [ + 46, + 276, + 287, + 361 + ], + "type": "inline_equation", + "content": "\\rho \\in [0,1]" + }, + { + "bbox": [ + 46, + 276, + 287, + 361 + ], + "type": "text", + "content": " is the roughness of the sphere, " + }, + { + "bbox": [ + 46, + 276, + 287, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_r:\\mathbb{R}^3\\to \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 276, + 287, + 361 + ], + "type": "text", + "content": " is so-called residual colors, and " + }, + { + "bbox": [ + 46, + 276, + 287, + 361 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 46, + 276, + 287, + 361 + ], + "type": "text", + "content": " is the element-wise multiplication." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 360, + 288, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 360, + 288, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 360, + 288, + 589 + ], + "type": "text", + "content": "Explanations on Eq. 3. We explain our motivation of this shading model in the following three aspects. a) Diffuse color " + }, + { + "bbox": [ + 46, + 360, + 288, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_d" + }, + { + "bbox": [ + 46, + 360, + 288, + 589 + ], + "type": "text", + "content": " represents the consistent colors of this Gaussian sphere, which do not change with viewing directions. b) " + }, + { + "bbox": [ + 46, + 360, + 288, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{s} \\odot L_s(\\omega_o, \\mathbf{n}, \\rho)" + }, + { + "bbox": [ + 46, + 360, + 288, + 589 + ], + "type": "text", + "content": " describes the interactions between the surface intrinsic color " + }, + { + "bbox": [ + 46, + 360, + 288, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 46, + 360, + 288, + 589 + ], + "type": "text", + "content": " and the direct specular light " + }, + { + "bbox": [ + 46, + 360, + 288, + 589 + ], + "type": "inline_equation", + "content": "L_s" + }, + { + "bbox": [ + 46, + 360, + 288, + 589 + ], + "type": "text", + "content": ". This term enables us to represent most of the reflections in rendering. c) Since there are still some reflections that cannot be explained by the above reflections of direct lights, such as scattering and reflection on indirect lights, we add a residual color term " + }, + { + "bbox": [ + 46, + 360, + 288, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_r(\\omega_o)" + }, + { + "bbox": [ + 46, + 360, + 288, + 589 + ], + "type": "text", + "content": " to account for these complex reflections. In comparison, Ref-NeRF [45] adopts a similar shading function without such a residual color term, which makes it struggle to handle advanced complex reflections. " + }, + { + "bbox": [ + 46, + 360, + 288, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_r(\\omega_o)" + }, + { + "bbox": [ + 46, + 360, + 288, + 589 + ], + "type": "text", + "content": " is parameterized by spherical harmonic functions. d) " + }, + { + "bbox": [ + 46, + 360, + 288, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_d" + }, + { + "bbox": [ + 46, + 360, + 288, + 589 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 360, + 288, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 46, + 360, + 288, + 589 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 360, + 288, + 589 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 46, + 360, + 288, + 589 + ], + "type": "text", + "content": ", and the coefficients of the spherical harmonics in " + }, + { + "bbox": [ + 46, + 360, + 288, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_r(\\omega_o)" + }, + { + "bbox": [ + 46, + 360, + 288, + 589 + ], + "type": "text", + "content": " all are trainable parameters associated with this Gaussian sphere. In the following, we will introduce how to compute " + }, + { + "bbox": [ + 46, + 360, + 288, + 589 + ], + "type": "inline_equation", + "content": "L_s(\\omega_o, \\mathbf{n}, \\rho)" + }, + { + "bbox": [ + 46, + 360, + 288, + 589 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 360, + 288, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 46, + 360, + 288, + 589 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 594, + 140, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 594, + 140, + 608 + ], + "spans": [ + { + "bbox": [ + 47, + 594, + 140, + 608 + ], + "type": "text", + "content": "3.2. Specular Light" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 613, + 287, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 613, + 287, + 650 + ], + "spans": [ + { + "bbox": [ + 46, + 613, + 287, + 650 + ], + "type": "text", + "content": "We compute the specular light " + }, + { + "bbox": [ + 46, + 613, + 287, + 650 + ], + "type": "inline_equation", + "content": "L_{s}" + }, + { + "bbox": [ + 46, + 613, + 287, + 650 + ], + "type": "text", + "content": " by integrating incoming radiance with the specular GGX [46] Normal Distribution Function " + }, + { + "bbox": [ + 46, + 613, + 287, + 650 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 613, + 287, + 650 + ], + "type": "text", + "content": " visualized in Fig. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 75, + 657, + 287, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 657, + 287, + 682 + ], + "spans": [ + { + "bbox": [ + 75, + 657, + 287, + 682 + ], + "type": "interline_equation", + "content": "L _ {s} (\\omega_ {o}, \\mathbf {n}, \\rho) = \\int_ {\\Omega} L (\\omega_ {i}) D (\\mathbf {r}, \\rho) (\\omega_ {i} \\cdot \\mathbf {n}) d \\omega_ {i}, \\qquad (4)", + "image_path": "d3b483ef782bc11da45bc834f9ff035f8746328391951c95ecc485589069ebc9.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": " represents the whole upper semi-sphere, " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\omega_{i}" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": " is the direction for the input radiance and " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": " characterizes the" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 342, + 76, + 504, + 140 + ], + "blocks": [ + { + "bbox": [ + 342, + 76, + 504, + 140 + ], + "lines": [ + { + "bbox": [ + 342, + 76, + 504, + 140 + ], + "spans": [ + { + "bbox": [ + 342, + 76, + 504, + 140 + ], + "type": "image", + "image_path": "e9c77fddb150afd8d57161c042d8e52ccd0296a8e4962286773e0228e8ef8f00.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 147, + 547, + 192 + ], + "lines": [ + { + "bbox": [ + 304, + 147, + 547, + 192 + ], + "spans": [ + { + "bbox": [ + 304, + 147, + 547, + 192 + ], + "type": "text", + "content": "Figure 3. Normal Distribution Function D in Eq. 4 is determined by the roughness " + }, + { + "bbox": [ + 304, + 147, + 547, + 192 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 304, + 147, + 547, + 192 + ], + "type": "text", + "content": " and reflective direction " + }, + { + "bbox": [ + 304, + 147, + 547, + 192 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 304, + 147, + 547, + 192 + ], + "type": "text", + "content": ". Surface with small " + }, + { + "bbox": [ + 304, + 147, + 547, + 192 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 304, + 147, + 547, + 192 + ], + "type": "text", + "content": " has a smaller specular lobe and that with large " + }, + { + "bbox": [ + 304, + 147, + 547, + 192 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 304, + 147, + 547, + 192 + ], + "type": "text", + "content": " has a larger specular lobe." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 201, + 545, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 201, + 545, + 321 + ], + "spans": [ + { + "bbox": [ + 304, + 201, + 545, + 321 + ], + "type": "text", + "content": "specular lobe (effective integral range). When the surface is rough, the specular lobe will be larger around the reflective direction " + }, + { + "bbox": [ + 304, + 201, + 545, + 321 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 304, + 201, + 545, + 321 + ], + "type": "text", + "content": " while if the surface is smooth, the specular lobe will be smaller. The reflection direction " + }, + { + "bbox": [ + 304, + 201, + 545, + 321 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 304, + 201, + 545, + 321 + ], + "type": "text", + "content": " is calculated by view direction " + }, + { + "bbox": [ + 304, + 201, + 545, + 321 + ], + "type": "inline_equation", + "content": "\\omega_{o}" + }, + { + "bbox": [ + 304, + 201, + 545, + 321 + ], + "type": "text", + "content": " and normal " + }, + { + "bbox": [ + 304, + 201, + 545, + 321 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 304, + 201, + 545, + 321 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 304, + 201, + 545, + 321 + ], + "type": "inline_equation", + "content": "\\mathbf{r} = 2(\\omega_{o} \\cdot \\mathbf{n})\\mathbf{n} - \\omega_{o}" + }, + { + "bbox": [ + 304, + 201, + 545, + 321 + ], + "type": "text", + "content": ". In our approach, environment light " + }, + { + "bbox": [ + 304, + 201, + 545, + 321 + ], + "type": "inline_equation", + "content": "L(\\omega_{i})" + }, + { + "bbox": [ + 304, + 201, + 545, + 321 + ], + "type": "text", + "content": " is represented by a trainable " + }, + { + "bbox": [ + 304, + 201, + 545, + 321 + ], + "type": "inline_equation", + "content": "6 \\times 64 \\times 64" + }, + { + "bbox": [ + 304, + 201, + 545, + 321 + ], + "type": "text", + "content": " cube map. All pixel values in our lighting maps are uniformly initialized to 0.8 (within a range of 0 to 1). Note that the lighting map is global, with each scene being optimized using a single lighting map." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 323, + 546, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 323, + 546, + 467 + ], + "spans": [ + { + "bbox": [ + 304, + 323, + 546, + 467 + ], + "type": "text", + "content": "The light modeling is based on NVDiffRec [34]. Specifically, we begin by preprocessing the environment lighting map into multiple mipmap levels, each storing the integrated light across different reflective directions and roughness. When calculating light integration, we interpolate between these mipmaps based on the input of specific reflection directions and roughness, to obtain the required light integrals. This mipmap-based framework, in contrast to Ref-NeRF's [45] approach using integrated directional encoding, enhances training efficiency. The reflection direction " + }, + { + "bbox": [ + 304, + 323, + 546, + 467 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 323, + 546, + 467 + ], + "type": "text", + "content": " is calculated by combining the observation direction " + }, + { + "bbox": [ + 304, + 323, + 546, + 467 + ], + "type": "inline_equation", + "content": "\\omega_{o}" + }, + { + "bbox": [ + 304, + 323, + 546, + 467 + ], + "type": "text", + "content": " with the normal " + }, + { + "bbox": [ + 304, + 323, + 546, + 467 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 323, + 546, + 467 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 477, + 419, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 477, + 419, + 489 + ], + "spans": [ + { + "bbox": [ + 306, + 477, + 419, + 489 + ], + "type": "text", + "content": "3.3. Normal Estimation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 497, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 545, + 605 + ], + "type": "text", + "content": "Normal estimation on Gaussian spheres is difficult. Gaussian spheres are a collection of discrete entities, each representing a localized point in space without a continuous surface or defined edge. This discrete structure makes it inherently difficult to directly calculate normals, which typically requires a continuous surface. In the following, we observed that the shortest axis direction of a Gaussian can serve as an approximated normal and we further associate a predicted normal residual on it." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": "Shortest axis direction. In experiments, we made an interesting observation that the aspect ratio of the 3D Gaussian sphere—specifically, the ratios of their longest, intermediate, and shortest axis—gradually increased during optimization, as shown in Fig. 4. This observation suggests that Gaussian spheres gradually become more flattened and approach a planar shape. This observation inspires us to select the shortest axis as the normal of this \"planarized\" Gaussian sphere, denoted by " + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5325" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 71, + 277, + 155 + ], + "blocks": [ + { + "bbox": [ + 63, + 71, + 277, + 155 + ], + "lines": [ + { + "bbox": [ + 63, + 71, + 277, + 155 + ], + "spans": [ + { + "bbox": [ + 63, + 71, + 277, + 155 + ], + "type": "image", + "image_path": "1d0aa677c70ebf18f9b0caca1a4919505c99916197b23b36f0bb357cf77b1bb3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 162, + 287, + 185 + ], + "lines": [ + { + "bbox": [ + 46, + 162, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 46, + 162, + 287, + 185 + ], + "type": "text", + "content": "Figure 4. The geometric evolving process of a 3D Gaussian sphere in the optimization, which gradually becomes planar." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 190, + 289, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 190, + 289, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 289, + 346 + ], + "type": "text", + "content": "Predicted normal residual. The shortest axis " + }, + { + "bbox": [ + 46, + 190, + 289, + 346 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 46, + 190, + 289, + 346 + ], + "type": "text", + "content": " only serves as an approximated normal. To make the normal computation more accurate, we further introduce a trainable normal residual " + }, + { + "bbox": [ + 46, + 190, + 289, + 346 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{n}" + }, + { + "bbox": [ + 46, + 190, + 289, + 346 + ], + "type": "text", + "content": " on every Gaussian sphere. However, the orientation of the shortest axis " + }, + { + "bbox": [ + 46, + 190, + 289, + 346 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 46, + 190, + 289, + 346 + ], + "type": "text", + "content": " has an ambiguity because the direction of the shortest axis could either point outward or inward from the surface. To handle this ambiguity, we optimize two separate normal residuals to accommodate both scenarios. Given a specific viewing direction " + }, + { + "bbox": [ + 46, + 190, + 289, + 346 + ], + "type": "inline_equation", + "content": "\\omega_{o}" + }, + { + "bbox": [ + 46, + 190, + 289, + 346 + ], + "type": "text", + "content": ", we first select the direction aligned with the viewing direction " + }, + { + "bbox": [ + 46, + 190, + 289, + 346 + ], + "type": "inline_equation", + "content": "\\omega_{\\mathrm{o}}" + }, + { + "bbox": [ + 46, + 190, + 289, + 346 + ], + "type": "text", + "content": " as the active normal direction for this viewing direction and then apply the corresponding normal residual to the active normal. This process is described by" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 351, + 287, + 383 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 351, + 287, + 383 + ], + "spans": [ + { + "bbox": [ + 91, + 351, + 287, + 383 + ], + "type": "interline_equation", + "content": "\\mathbf {n} = \\left\\{ \\begin{array}{l l} \\mathbf {v} + \\Delta \\mathbf {n} _ {1} & \\text {i f} \\omega_ {\\mathbf {o}} \\cdot \\mathbf {v} > 0, \\\\ - (\\mathbf {v} + \\Delta \\mathbf {n} _ {2}) & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {5}", + "image_path": "31ceb45cc43f0673cd5a7bd177a77c806f40b17ac03a48c78ff213c10a335ff9.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 388, + 288, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 388, + 288, + 424 + ], + "spans": [ + { + "bbox": [ + 47, + 388, + 288, + 424 + ], + "type": "text", + "content": "To prevent the normal residual from deviating too much from the shortest axis, we add a penalty towards normal residual, making sure it is small enough." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 135, + 430, + 287, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 430, + 287, + 444 + ], + "spans": [ + { + "bbox": [ + 135, + 430, + 287, + 444 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e g}} = \\left\\| \\Delta \\mathbf {n} \\right\\| ^ {2} \\tag {6}", + "image_path": "e2dbf225bea9b855ca4e99f4e5271dc4069ca5627a6e4d1a0736aa64b3ca4b58.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 450, + 249, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 450, + 249, + 462 + ], + "spans": [ + { + "bbox": [ + 47, + 450, + 249, + 462 + ], + "type": "text", + "content": "The normal residual is shown on the left of Fig. 5." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 463, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 463, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 463, + 288, + 714 + ], + "type": "text", + "content": "Normal-geometry consistency. The above shortest axis direction and normal residuals are defined on each Gaussian sphere separately. However, a noticeable problem is that a normal reveals the gradient of the local geometry, which is supposed to be associated with all the Gaussian spheres in a local region. We find that simply applying a color loss to train the aforementioned normal residuals leads to inconsistency between the local geometry and the estimated normals. The main reason is that every Gaussian sphere learns its normal residuals separately without knowing the local geometry formulated by its neighbor Gaussian spheres. Thus, we have to correlate multiple Gaussian spheres in a local region with their normals to ensure normal-geometry consistency. A straightforward and naive solution is to search for its " + }, + { + "bbox": [ + 46, + 463, + 288, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 463, + 288, + 714 + ], + "type": "text", + "content": " neighborhoods in space and estimate a coarse normal from all the neighboring spheres. However, such KNN searching would be extremely expensive during training because all the Gaussian spheres are dynamically moving in the optimization process. Instead, we propose a simple yet effective way to ensure normal-geometry consistency as follows." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 306, + 69, + 547, + 118 + ], + "blocks": [ + { + "bbox": [ + 306, + 69, + 547, + 118 + ], + "lines": [ + { + "bbox": [ + 306, + 69, + 547, + 118 + ], + "spans": [ + { + "bbox": [ + 306, + 69, + 547, + 118 + ], + "type": "image", + "image_path": "e13ffe9690f31500f8555ac92c41357f6d5f57dcc1bcf22260ee48d6f49ad889.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 125, + 547, + 159 + ], + "lines": [ + { + "bbox": [ + 305, + 125, + 547, + 159 + ], + "spans": [ + { + "bbox": [ + 305, + 125, + 547, + 159 + ], + "type": "text", + "content": "Figure 5. Visualization of the relationship between shortes axis " + }, + { + "bbox": [ + 305, + 125, + 547, + 159 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 305, + 125, + 547, + 159 + ], + "type": "text", + "content": ", normal residual " + }, + { + "bbox": [ + 305, + 125, + 547, + 159 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{n}" + }, + { + "bbox": [ + 305, + 125, + 547, + 159 + ], + "type": "text", + "content": ", normal " + }, + { + "bbox": [ + 305, + 125, + 547, + 159 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 305, + 125, + 547, + 159 + ], + "type": "text", + "content": " and depth-gradient normal " + }, + { + "bbox": [ + 305, + 125, + 547, + 159 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{n}}" + }, + { + "bbox": [ + 305, + 125, + 547, + 159 + ], + "type": "text", + "content": ". The supervision " + }, + { + "bbox": [ + 305, + 125, + 547, + 159 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{normal}" + }, + { + "bbox": [ + 305, + 125, + 547, + 159 + ], + "type": "text", + "content": " enforces the normal-geometry consistency." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 165, + 547, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 165, + 547, + 213 + ], + "spans": [ + { + "bbox": [ + 304, + 165, + 547, + 213 + ], + "type": "text", + "content": "We associate the local geometry with predicted normals by minimizing the difference between the grad normals derived from the rendered depth map and the rendered normal maps using the predicted normals" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 382, + 220, + 545, + 235 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 220, + 545, + 235 + ], + "spans": [ + { + "bbox": [ + 382, + 220, + 545, + 235 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {n o r m a l}} = \\left| \\left| \\bar {\\mathbf {n}} - \\hat {\\mathbf {n}} \\right| \\right| ^ {2}, \\tag {7}", + "image_path": "da65d3eee3525db2b02e8159fbb61e2ce98bcab6e4020345de7a616fa43fa0a2.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 243, + 546, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 546, + 338 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 546, + 338 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 243, + 546, + 338 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{n}}" + }, + { + "bbox": [ + 304, + 243, + 546, + 338 + ], + "type": "text", + "content": " is the rendered normal map and " + }, + { + "bbox": [ + 304, + 243, + 546, + 338 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{n}}" + }, + { + "bbox": [ + 304, + 243, + 546, + 338 + ], + "type": "text", + "content": " is computed by applying the Sobel-like operator on the rendered depth map. " + }, + { + "bbox": [ + 304, + 243, + 546, + 338 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{n}}" + }, + { + "bbox": [ + 304, + 243, + 546, + 338 + ], + "type": "text", + "content": " reveals the local geometry formulated by multiple Gaussian spheres because it is computed from the rendered depth maps. " + }, + { + "bbox": [ + 304, + 243, + 546, + 338 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{n}}" + }, + { + "bbox": [ + 304, + 243, + 546, + 338 + ], + "type": "text", + "content": " contains the information from the separately-defined normal on each Gaussian sphere. By minimizing the difference, we enforce the consistency between the local geometry and the estimated normals." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 346, + 359, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 346, + 359, + 357 + ], + "spans": [ + { + "bbox": [ + 306, + 346, + 359, + 357 + ], + "type": "text", + "content": "3.4. Losses" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 364, + 547, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 364, + 547, + 399 + ], + "spans": [ + { + "bbox": [ + 305, + 364, + 547, + 399 + ], + "type": "text", + "content": "In addition to " + }, + { + "bbox": [ + 305, + 364, + 547, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{color}}" + }, + { + "bbox": [ + 305, + 364, + 547, + 399 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 364, + 547, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{reg}" + }, + { + "bbox": [ + 305, + 364, + 547, + 399 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 364, + 547, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{normal}}" + }, + { + "bbox": [ + 305, + 364, + 547, + 399 + ], + "type": "text", + "content": ", we make use of sparse loss [30, 50] to encourages Gaussian spheres' opacity values " + }, + { + "bbox": [ + 305, + 364, + 547, + 399 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 305, + 364, + 547, + 399 + ], + "type": "text", + "content": " to approach either 0 or 1 by" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 339, + 407, + 545, + 434 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 407, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 339, + 407, + 545, + 434 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s p a r s e}} = \\frac {1}{| \\alpha |} \\sum_ {\\alpha_ {i}} \\left[ \\log \\left(\\alpha_ {i}\\right) + \\log \\left(1 - \\alpha_ {i}\\right) \\right]. \\tag {8}", + "image_path": "4bbf0efe11a8a1f4366e270c82f345cd9eef623628950224e89ecc34074510da.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 442, + 545, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 442, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 442, + 545, + 479 + ], + "type": "text", + "content": "This sparsity loss helps the geometry of Gaussian spheres converge to a single thin plate and improves the rendering quality. In summary, the total training loss " + }, + { + "bbox": [ + 304, + 442, + 545, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 304, + 442, + 545, + 479 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 337, + 487, + 545, + 500 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 487, + 545, + 500 + ], + "spans": [ + { + "bbox": [ + 337, + 487, + 545, + 500 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {c o l o r}} + \\lambda_ {\\mathrm {n}} \\mathcal {L} _ {\\text {n o r m a l}} + \\lambda_ {\\mathrm {s}} \\mathcal {L} _ {\\text {s p a r s e}} + \\lambda_ {\\mathrm {r}} \\mathcal {L} _ {\\text {r e g}} \\tag {9}", + "image_path": "cfd209a34ccf32e8339cc137af05fadb3a532c92c766f10ab95fda10c8535c2b.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 508, + 477, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 508, + 477, + 520 + ], + "spans": [ + { + "bbox": [ + 306, + 508, + 477, + 520 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 508, + 477, + 520 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{n}} = 0.01, \\lambda_{\\mathrm{s}} = 0.001, \\lambda_{\\mathrm{r}} = 0.001" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 530, + 387, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 530, + 387, + 544 + ], + "spans": [ + { + "bbox": [ + 306, + 530, + 387, + 544 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 550, + 368, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 550, + 368, + 562 + ], + "spans": [ + { + "bbox": [ + 306, + 550, + 368, + 562 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 568, + 547, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 568, + 547, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 568, + 547, + 640 + ], + "type": "text", + "content": "To comprehensively validate the effectiveness of GaussianShader, we conduct evaluation on various datasets: a) widely-used NVS dataset: NeRF Synthetic [32]. b) reflective objects datasets: Shiny Blender [45] and Glossy Synthetic [29]. c) real-world large-scale scenes: Tanks and Temples [23]." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 647, + 431, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 431, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 431, + 658 + ], + "type": "text", + "content": "4.2. Baselines and Metrics" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 666, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 547, + 713 + ], + "type": "text", + "content": "We compare our method against the following baselines: a) 3D Gaussian Splatting [21]: a real-time radiance field rendering method based on efficient 3D Gaussian representation; b) VolSDF [52]: a classical neural implicit surface" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5326" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 111, + 83, + 483, + 322 + ], + "blocks": [ + { + "bbox": [ + 123, + 70, + 470, + 82 + ], + "lines": [ + { + "bbox": [ + 123, + 70, + 470, + 82 + ], + "spans": [ + { + "bbox": [ + 123, + 70, + 470, + 82 + ], + "type": "text", + "content": "Table 1. The quantitative comparisons (PSNR / SSIM / LPIPS) on NeRF Synthetic dataset [32]." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 111, + 83, + 483, + 322 + ], + "lines": [ + { + "bbox": [ + 111, + 83, + 483, + 322 + ], + "spans": [ + { + "bbox": [ + 111, + 83, + 483, + 322 + ], + "type": "table", + "html": "
ChairDrumsNeRF Synthetic [32]
LegoMicMaterialsShipHotdogFicusAvg.
PSNR↑
NeRF [32]33.0025.0132.5432.9129.6228.6536.1830.1331.01
VolSDF [52]30.5720.4329.4630.5329.1325.5135.1122.9127.96
Ref-NeRF [45]33.9825.4335.1033.6527.1029.2437.0428.7431.29
ENVIDR [27]31.2222.9929.5532.1729.5221.5731.4426.6028.13
Gaussian Splitting [21]35.8226.1735.6935.3430.0030.8737.6734.8333.30
Ours35.8326.3635.8735.2330.0730.8237.8534.9733.38
SSIM↑
NeRF [32]0.9670.9250.9610.9800.9490.8560.9740.9640.947
VolSDF [52]0.9490.8930.9510.9690.9540.8420.9720.9290.932
Ref-NeRF [45]0.9740.9290.9750.9830.9210.8640.9790.9540.947
ENVIDR [27]0.9760.9300.9610.9840.9680.8550.9630.9870.956
Gaussian Splitting [21]0.9870.9540.9830.9910.9600.9070.9850.9870.969
Ours0.9870.9490.9830.9910.9600.9050.9850.9850.968
LPIPS↓
NeRF [32]0.0460.0910.0500.0280.0630.2060.1210.0440.081
VolSDF [52]0.0560.1190.0540.1910.0480.1910.0430.0680.096
Ref-NeRF [45]0.0290.0730.0250.0180.0780.1580.0280.0560.058
ENVIDR [27]0.0310.0800.0540.0210.0450.2280.0720.0100.067
Gaussian Splitting [21]0.0120.0370.0160.0060.0340.1060.0200.0120.030
Ours0.0120.0400.0140.0060.0330.0980.0190.0130.029
", + "image_path": "384b6f476321c46466d38cb76fbb1f655326cdc7c3ea509ac7e2645470fb8c7e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 47, + 331, + 292, + 498 + ], + "blocks": [ + { + "bbox": [ + 47, + 331, + 292, + 498 + ], + "lines": [ + { + "bbox": [ + 47, + 331, + 292, + 498 + ], + "spans": [ + { + "bbox": [ + 47, + 331, + 292, + 498 + ], + "type": "image", + "image_path": "e92b80a024832da86c0d9ef310a11b1dae0625e102480f157df6397aea4aa98c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 508, + 546, + 533 + ], + "lines": [ + { + "bbox": [ + 46, + 508, + 546, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 508, + 546, + 533 + ], + "type": "text", + "content": "Figure 6. The qualitative comparisons with 3D Gaussian Splitting [21] on Glossy dataset [29]. Our method not only renders the objects with high fidelity but also provides detailed normal and lighting maps. Some areas are zoomed in for better visualization." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 294, + 331, + 545, + 497 + ], + "blocks": [ + { + "bbox": [ + 294, + 331, + 545, + 497 + ], + "lines": [ + { + "bbox": [ + 294, + 331, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 294, + 331, + 545, + 497 + ], + "type": "image", + "image_path": "61a2d2e5f090b2e4d4c12f64b600f566adc3a190f10162b2ecb79458930e0007.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 543, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 543, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 543, + 287, + 639 + ], + "type": "text", + "content": "reconstruction method based on SDF; c) Ref-NeRF [45]: state-of-the-art method in novel view synthesis of reflective objects; d) NVDiffRec [34], NVDiffRecMC [18]: top-performing neural inverse rendering methods; e) ENVIDR [27], NeRO [29]: top-performing SDF-based neural implicit methods for reconstructing reflective objects. The evaluation metrics to measure rendering quality are reported in PSNR, SSIM [48], and LPIPS [56]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 647, + 180, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 180, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 180, + 659 + ], + "type": "text", + "content": "4.3. Implementation Details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "All experiments are conducted on a Nvidia RTX 3090 graphics card. We optimize our models using Adam [22] optimizer for 30,000 iterations. For Ref-NeRF [45], ENVIDR [27], and 3D Gaussian Splitting [21], we retrain" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 543, + 545, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 543, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 304, + 543, + 545, + 592 + ], + "type": "text", + "content": "them using their official codes keeping all the original settings (training iterations, sampled rays, etc.) and obtain results from the retrained models. Other works' results are imported from their original papers." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 599, + 391, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 599, + 391, + 612 + ], + "spans": [ + { + "bbox": [ + 306, + 599, + 391, + 612 + ], + "type": "text", + "content": "4.4. Comparisons" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 617, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 546, + 715 + ], + "type": "text", + "content": "NeRF Synthetic dataset [32]. We first evaluate our model on the NeRF Synthetic dataset, which contains objects with complex geometry and realistic non-Lambertian materials. We show the quantitative results in Tab. 1 and visual comparisons in Fig. 8. Our approach achieves numerically and visually comparable results with both Gaussian Splatting and neural rendering methods, demonstrating the effectiveness of our method in rendering general objects." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5327" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 289, + 177 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 289, + 177 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 289, + 177 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 289, + 177 + ], + "type": "image", + "image_path": "185816575b056867e8c1cca272e0fd3ffa0097de86d7c017fa8034a35e2e3741.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 183, + 287, + 217 + ], + "lines": [ + { + "bbox": [ + 46, + 183, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 46, + 183, + 287, + 217 + ], + "type": "text", + "content": "Figure 7. Qualitative comparisons on Shiny Blender dataset [45]. Error maps corresponding to each rendering result demonstrate the enhanced fidelity of our approach in highly specular areas." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 225, + 287, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 225, + 287, + 250 + ], + "spans": [ + { + "bbox": [ + 46, + 225, + 287, + 250 + ], + "type": "text", + "content": "Note that SDF-based ENVIDR [27] cannot perform well on shadowed regions of complex objects, as shown in Fig. 8." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 251, + 287, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 287, + 407 + ], + "type": "text", + "content": "Shiny Blender dataset [45]. Quantitative results on the Shiny Blender dataset are reported in Tab. 2. With the ability to model light-surface interactions, our method outperforms original Gaussian Splatting in all scenes while slightly underperforms ENVIDR and Ref-NeRF. ENVIDR capitalizes on the continuous properties of Signed Distance Fields to naturally create smoother surfaces and normals, resulting in high rendering quality, particularly for reflective objects. Fig. 7 shows our comparisons with Gaussian Splatting. We can see that our method correctly renders surfaces with strong specular appearances. High-quality modeling of reflections is contingent upon the accurate estimation of normals as shown in Fig. 9." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 409, + 287, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 409, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 409, + 287, + 529 + ], + "type": "text", + "content": "Additionally, we report the average PSNR, training time, and rendering FPS in Tab. 3. The reported results are averaged among all objects of the Shiny Blender and NeRF synthetic datasets. On the same hardware, our method only takes about 0.5 hour for training while MLP-based methods, like ENVIDR and Ref-NeRF, require 6 hours and 23 hours for optimizing. Due to the introduce of extra lighting and shading attributes, our method is a bit slower than 3D Gaussian Splatting [21] but still keeps reasonable efficiency and real-time rendering speed." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 531, + 287, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 531, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 46, + 531, + 287, + 591 + ], + "type": "text", + "content": "Tanks and Temples dataset [23]. To explore our method's scalability in larger-scale environments instead of only small objects, we conducted experiments using Tanks and Temples [23]. The qualitative results in Fig. 10 show that our model outperforms Gaussian Splatting [21]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": "Glossy Synthetic dataset [29]. We present a comparative visualization between Gaussian Splatting and our method on the Glossy dataset shown in Fig. 6. Our method not only renders objects with exceptional fidelity but also provides more accurate normal and lightings, capturing the nuances of reflective surfaces with greater precision. These maps reveal our method's ability to approximate the rendering function, resulting in a more lifelike portrayal of specular lighting. Please refer to the supplementary materials provided for an extensive showcase of our results." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 308, + 107, + 545, + 313 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 545, + 104 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 104 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 104 + ], + "type": "text", + "content": "Table 2. Qualitative comparisons on Shiny Blender dataset [45]. Our method is comparable with both Gaussian Splatting [21] and prior reflective object reconstruction methods." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 107, + 545, + 313 + ], + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 313 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 313 + ], + "type": "table", + "html": "
Shiny Blender [45]
CarBallHelmetTeapotToasterCoffeeAvg.
PSNR↑
NVDiffRec [34]27.9821.7726.9740.4424.3130.7428.70
NVDiffMC [18]25.9330.8526.2738.4422.1829.6028.88
Ref-NeRF [45]30.4129.1429.9245.1925.2933.9932.32
NeRO [29]25.5330.2629.2038.7026.4628.8929.84
ENVIDR [27]28.4638.8932.7341.5926.1129.4832.88
Gaussian Splitting [21]27.2427.6928.3245.6820.9932.3230.37
Ours27.9030.9828.3245.8626.2132.3931.94
SSIM↑
NVDiffRec [34]0.9630.8580.9510.9960.9280.9730.945
NVDiffMC [18]0.9400.9400.9400.9950.8860.9650.944
Ref-NeRF [45]0.9490.9560.9550.9950.9100.9720.956
NeRO [29]0.9490.9740.9710.9950.9290.9560.962
ENVIDR [27]0.9610.9910.9800.9960.9390.9490.969
Gaussian Splitting [21]0.9300.9370.9510.9960.8950.9710.947
Ours0.9310.9650.9500.9960.9290.9710.957
LPIPS↓
NVDiffRec [34]0.0450.2970.1180.0110.1690.0760.119
NVDiffMC [18]0.0770.3120.1570.0140.2250.0970.147
Ref-NeRF [45]0.0510.3070.0870.0130.1180.0820.109
NeRO [29]0.0740.0940.0500.0120.0890.1100.072
ENVIDR [27]0.0490.0670.0510.0110.1160.1390.072
Gaussian Splitting [21]0.0470.1610.0790.0070.1260.0780.083
Ours0.0450.1210.0760.0070.0790.0780.068
", + "image_path": "c5759a7d083a879186c7a10cdb0ae320969e29386844b611739142ee8e4aba9d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 328, + 355, + 524, + 407 + ], + "blocks": [ + { + "bbox": [ + 305, + 318, + 545, + 351 + ], + "lines": [ + { + "bbox": [ + 305, + 318, + 545, + 351 + ], + "spans": [ + { + "bbox": [ + 305, + 318, + 545, + 351 + ], + "type": "text", + "content": "Table 3. Comparisons over average PSNR, training time and FPS. Our method achieves the best rendering quality and competitive training/rendering speed." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 355, + 524, + 407 + ], + "lines": [ + { + "bbox": [ + 328, + 355, + 524, + 407 + ], + "spans": [ + { + "bbox": [ + 328, + 355, + 524, + 407 + ], + "type": "table", + "html": "
PSNRTraining TimeFPS
Ref-NeRF [45]31.7323h0.03
ENVIDR [27]30.046h1.33
Gaussian Splitting [21]32.050.25h274
Ours32.760.58h97
", + "image_path": "b839f48ab899f9aeebb5bb055629929c54f04f08f7fba0453f383ac31ccc5a1c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 428, + 399, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 428, + 399, + 441 + ], + "spans": [ + { + "bbox": [ + 306, + 428, + 399, + 441 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 448, + 545, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 448, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 305, + 448, + 545, + 495 + ], + "type": "text", + "content": "In this section, we conduct ablation studies of our model on Shiny Blender dataset. The numerical results are reported in Tab. 4. All ablation experiments are done on Shiny Blender dataset with " + }, + { + "bbox": [ + 305, + 448, + 545, + 495 + ], + "type": "inline_equation", + "content": "1/2" + }, + { + "bbox": [ + 305, + 448, + 545, + 495 + ], + "type": "text", + "content": " image resolution." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 496, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 496, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 496, + 545, + 579 + ], + "type": "text", + "content": "Loss functions. We trained our model serperately under the setting \"without sparse loss " + }, + { + "bbox": [ + 304, + 496, + 545, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{sparse}}" + }, + { + "bbox": [ + 304, + 496, + 545, + 579 + ], + "type": "text", + "content": " and \"without normal loss " + }, + { + "bbox": [ + 304, + 496, + 545, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{normal}}" + }, + { + "bbox": [ + 304, + 496, + 545, + 579 + ], + "type": "text", + "content": ". The incorporation of these losses, especially the direct regularization on normals, proved critical, yielding a marked improvement in rendering quality by effectively guiding the normal optimization process as shown in Tab. 4." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 581, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 581, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 581, + 545, + 653 + ], + "type": "text", + "content": "Residual color. Residual color, comprising 3rd order SH coefficients, is used to compensate for indirect light. As shown in Table 4, the introduce of residual color enhances our model, allowing it to simulate a wide array of real-world intricate lighting situations and significantly improve rendering quality." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "Normal formulation. In our full model, we use the shortest axis direction " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " with a learnable residual " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{n}" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " to formulate normal, as described by Eqn. 3.3. In Tab. 4, we compared this novel normal formation with a naive normal prediction manner (indicated by \"w/o axis " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " ). Instead of" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5328" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 68, + 527, + 230 + ], + "blocks": [ + { + "bbox": [ + 67, + 68, + 527, + 230 + ], + "lines": [ + { + "bbox": [ + 67, + 68, + 527, + 230 + ], + "spans": [ + { + "bbox": [ + 67, + 68, + 527, + 230 + ], + "type": "image", + "image_path": "5f0b70463b9d7846603de94af9f82cc29de25f203370cd7b39ad97fcf7a07858.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 236, + 546, + 259 + ], + "lines": [ + { + "bbox": [ + 46, + 236, + 546, + 259 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 546, + 259 + ], + "type": "text", + "content": "Figure 8. We present a qualitative comparisons of our method on NeRF Synthetic dataset [32] against previous techniques and corresponding ground truth images from test views. For clarity, we zoom in the images where differences in quality are especially notable." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 47, + 262, + 287, + 413 + ], + "blocks": [ + { + "bbox": [ + 47, + 262, + 287, + 413 + ], + "lines": [ + { + "bbox": [ + 47, + 262, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 47, + 262, + 287, + 413 + ], + "type": "image", + "image_path": "baf25c508b4a5b28b3951f669dea5a80c2968e1692cf22e757ac083b655eda9e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 422, + 287, + 475 + ], + "lines": [ + { + "bbox": [ + 46, + 422, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 287, + 475 + ], + "type": "text", + "content": "Figure 9. Visualized normal results on Shiny Blender dataset [45], compared with Gaussian Splatting [21]. It illustrates the superior normal estimation achieved by our method. Table 4. Ablation studies on loss function, lighting representation, normal axis and residual color using Shiny Blender dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 77, + 483, + 254, + 564 + ], + "blocks": [ + { + "bbox": [ + 77, + 483, + 254, + 564 + ], + "lines": [ + { + "bbox": [ + 77, + 483, + 254, + 564 + ], + "spans": [ + { + "bbox": [ + 77, + 483, + 254, + 564 + ], + "type": "table", + "html": "
PSNR↑SSIM↑LPIPS↓
w/o Lsparse31.790.9520.056
w/o Lnormal30.930.9410.060
w/o cr31.490.9480.060
w/o v31.470.9510.058
MLP Lighting29.730.9360.075
Full Model32.090.9530.054
", + "image_path": "b110e4a2b60fd33097c6520f18ea210532f3304f3650092be923fd231026c443.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 570, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 629 + ], + "type": "text", + "content": "correlating with the geometry of 3D gaussians, we directly optimize a single normal attribute. The result shows our full model can achieve higher rendering quality, since our predicted normals are correlated with the true geometry of Gaussian spheres." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "content": "Lighting representation. A crucial component of our approach is the modeling of lighting. \"MLP Lighting\" means we use a directional MLP similar to Ref-NeRF to implicitly optimize lighting. In our full model, we use \"Envmap Lighting\" representation as illustrated in Sec. 3.2. The result proves that our explicit lighting representation is better suited for 3D Gaussian rendering framework to cor" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 304, + 262, + 545, + 479 + ], + "blocks": [ + { + "bbox": [ + 304, + 262, + 545, + 479 + ], + "lines": [ + { + "bbox": [ + 304, + 262, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 262, + 545, + 479 + ], + "type": "image", + "image_path": "a60dd242baef56b0f2a97f39aedd39f9bd9988886242d480f46e847e5bbe7046.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 487, + 547, + 520 + ], + "lines": [ + { + "bbox": [ + 304, + 487, + 547, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 487, + 547, + 520 + ], + "type": "text", + "content": "Figure 10. Qualitative comparisons with Gaussian Splitting [21] on the Tanks and Temples dataset [23]. Our approach yields a smoother and more reasonable normal." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 532, + 432, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 532, + 432, + 544 + ], + "spans": [ + { + "bbox": [ + 306, + 532, + 432, + 544 + ], + "type": "text", + "content": "rectly render reflective objects." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 559, + 383, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 559, + 383, + 572 + ], + "spans": [ + { + "bbox": [ + 306, + 559, + 383, + 572 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "text", + "content": "In summary, we propose GaussianShader, advancing the rendering of both reflective and general objects through an extended 3D Gaussian model. Specifically, our method integrates shading functions with 3D Gaussian representation for handling view-dependent appearances, successfully handling reflective surfaces. Moreover, a novel normal prediction method is proposed to achieve high-quality normals, which correlates the shading parameters and true geometry of 3D Gaussians. The results show our approach achieves high-quality renderings, marking a substantial improvement in efficient, realistic 3D object rendering." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5329" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "text", + "content": "[1] Miika Aittala, Timo Aila, and Jaakko Lehtinen. Reflectance modeling by neural texture synthesis. ACM Transactions on Graphics (ToG), 35(4):1-13, 2016. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 288, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 288, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 288, + 179 + ], + "type": "text", + "content": "[2] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 696-712. Springer, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 179, + 288, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 179, + 288, + 236 + ], + "spans": [ + { + "bbox": [ + 53, + 179, + 288, + 236 + ], + "type": "text", + "content": "[3] Matthew Anderson, Ricardo Motta, Srinivasan Chandrasekar, and Michael Stokes. Proposal for a standard default color space for the internet—srgb. In Color and imaging conference, pages 238–245. Society for Imaging Science and Technology, 1996. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 288, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 288, + 278 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 288, + 278 + ], + "type": "text", + "content": "[4] Jonathan T Barron and Jitendra Malik. Shape, illumination, and reflectance from shading. IEEE transactions on pattern analysis and machine intelligence, 37(8):1670-1687, 2014. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 280, + 288, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 280, + 288, + 345 + ], + "spans": [ + { + "bbox": [ + 53, + 280, + 288, + 345 + ], + "type": "text", + "content": "[5] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 346, + 288, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 288, + 402 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 288, + 402 + ], + "type": "text", + "content": "[6] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470-5479, 2022. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 402, + 288, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 402, + 288, + 456 + ], + "spans": [ + { + "bbox": [ + 53, + 402, + 288, + 456 + ], + "type": "text", + "content": "[7] Sai Bi, Zexiang Xu, Pratul Srinivasan, Ben Mildenhall, Kalyan Sunkavalli, Milos Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Neural reflectance fields for appearance acquisition. arXiv preprint arXiv:2008.03824, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 457, + 288, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 457, + 288, + 513 + ], + "spans": [ + { + "bbox": [ + 53, + 457, + 288, + 513 + ], + "type": "text", + "content": "[8] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T Barron, Ce Liu, and Hendrik Lensch. Nerd: Neural reflectance decomposition from image collections. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12684-12694, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 514, + 288, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 514, + 288, + 567 + ], + "spans": [ + { + "bbox": [ + 53, + 514, + 288, + 567 + ], + "type": "text", + "content": "[9] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan Barron, and Hendrik Lensch. Neural-pil: Neural pre-integrated lighting for reflectance decomposition. Advances in Neural Information Processing Systems, 34: 10691-10704, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 568, + 288, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 288, + 612 + ], + "type": "text", + "content": "[10] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 288, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 288, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 288, + 644 + ], + "type": "text", + "content": "[11] Zilong Chen, Feng Wang, and Huaping Liu. Text-to-3d using gaussian splatting. arXiv preprint arXiv:2309.16585, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "type": "text", + "content": "[12] Youming Deng, Xueting Li, Sifei Liu, and Ming-Hsuan Yang. Dip: Differentiable interreflection-aware physics-based inverse rendering. arXiv preprint arXiv:2212.04705, 2022.3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 690, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 690, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 690, + 288, + 714 + ], + "type": "text", + "content": "[13] Valentin Deschaintre, Miika Aittala, Fredo Durand, George Drettakis, and Adrien Bousseau. Single-image svbrdf cap" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "ture with a rendering-aware deep network. ACM Transactions on Graphics (ToG), 37(4):1-15, 2018. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 96, + 547, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 152 + ], + "type": "text", + "content": "[14] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 153, + 545, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 153, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 307, + 153, + 545, + 208 + ], + "type": "text", + "content": "[15] Stephan J Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14346-14355, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 209, + 545, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 253 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 253 + ], + "type": "text", + "content": "[16] Steven J Gortler, Radek Grzesczuk, Richard Szeliski, and Michael F Cohen. The lumigraph. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 453-464. 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 256, + 545, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 256, + 545, + 311 + ], + "spans": [ + { + "bbox": [ + 307, + 256, + 545, + 311 + ], + "type": "text", + "content": "[17] Haoyu Guo, Sida Peng, Haotong Lin, Qianqian Wang, Guofeng Zhang, Hujun Bao, and Xiaowei Zhou. Neural 3d scene reconstruction with the manhattan-world assumption. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5511-5520, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 312, + 545, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 312, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 307, + 312, + 545, + 357 + ], + "type": "text", + "content": "[18] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, Light, and Material Decomposition from Images using Monte Carlo Rendering and Denoising. arXiv:2206.03380, 2022. 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 358, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 358, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 358, + 545, + 403 + ], + "type": "text", + "content": "[19] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, light, and material decomposition from images using monte carlo rendering and denoising. Advances in Neural Information Processing Systems, 35:22856-22869, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 404, + 545, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 404, + 545, + 438 + ], + "spans": [ + { + "bbox": [ + 307, + 404, + 545, + 438 + ], + "type": "text", + "content": "[20] James T Kajiya. The rendering equation. In Proceedings of the 13th annual conference on Computer graphics and interactive techniques, pages 143-150, 1986. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 439, + 545, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 439, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 307, + 439, + 545, + 483 + ], + "type": "text", + "content": "[21] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 42(4):1-14, 2023. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 484, + 545, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 484, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 307, + 484, + 545, + 517 + ], + "type": "text", + "content": "[22] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 519, + 545, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 519, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 307, + 519, + 545, + 564 + ], + "type": "text", + "content": "[23] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36 (4):1-13, 2017. 5, 7, 8, 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 565, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 565, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 307, + 565, + 545, + 632 + ], + "type": "text", + "content": "[24] Chloe LeGendre, Wan-Chun Ma, Graham Fyffe, John Flynn, Laurent Charbonnel, Jay Busch, and Paul Debevec. Deeplight: Learning illumination for unconstrained mobile mixed reality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5918-5928, 2019. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 633, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 633, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 633, + 545, + 667 + ], + "type": "text", + "content": "[25] Marc Levoy and Pat Hanrahan. Light field rendering. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 441-452. 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "text", + "content": "[26] Ruofan Liang, Jiahao Zhang, Haoda Li, Chen Yang, Yushi Guan, and Nandita Vijaykumar. Spidr: Sdf-based neural point fields for illumination and deformation. arXiv preprint arXiv:2210.08398, 2022. 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "5330" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[27] Ruofan Liang, Huiting Chen, Chunlin Li, Fan Chen, Selvakumar Panneer, and Nandita Vijaykumar. Envidr: Implicit differentiable renderer with neural environment lighting. arXiv preprint arXiv:2303.13022, 2023. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 160 + ], + "type": "text", + "content": "[28] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 162, + 287, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 162, + 287, + 215 + ], + "spans": [ + { + "bbox": [ + 49, + 162, + 287, + 215 + ], + "type": "text", + "content": "[29] Yuan Liu, Peng Wang, Cheng Lin, Xiaoxiao Long, Jiepeng Wang, Lingjie Liu, Taku Komura, and Wenping Wang. Nano: Neural geometry and brdf reconstruction of reflective objects from multiview images. arXiv preprint arXiv:2305.17398, 2023. 3, 5, 6, 7, 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 217, + 287, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 217, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 49, + 217, + 287, + 259 + ], + "type": "text", + "content": "[30] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural volumes: Learning dynamic renderable volumes from images. arXiv preprint arXiv:1906.07751, 2019. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 261, + 287, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 261, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 49, + 261, + 287, + 282 + ], + "type": "text", + "content": "[31] Wojciech Matusik. A data-driven reflectance model. PhD thesis, Massachusetts Institute of Technology, 2003. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 284, + 287, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 284, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 49, + 284, + 287, + 337 + ], + "type": "text", + "content": "[32] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 5, 6, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 338, + 287, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 338, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 49, + 338, + 287, + 381 + ], + "type": "text", + "content": "[33] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 383, + 287, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 383, + 287, + 447 + ], + "spans": [ + { + "bbox": [ + 49, + 383, + 287, + 447 + ], + "type": "text", + "content": "[34] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3d models, materials, and lighting from images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8280-8290, 2022. 4, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 449, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 449, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 49, + 449, + 287, + 491 + ], + "type": "text", + "content": "[35] Merlin Nimier-David, Delio Vicini, Tizian Zeltner, and Wenzel Jakob. Mitsuba 2: A retargetable forward and inverse renderer. ACM Transactions on Graphics (TOG), 38(6):1-17, 2019. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 493, + 287, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 493, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 49, + 493, + 287, + 547 + ], + "type": "text", + "content": "[36] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5589-5599, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 548, + 287, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 548, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 49, + 548, + 287, + 591 + ], + "type": "text", + "content": "[37] Jeong Joon Park, Aleksander Holynski, and Steven M Seitz. Seeing the world in a bag of chips. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1417-1427, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 593, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 593, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 49, + 593, + 287, + 645 + ], + "type": "text", + "content": "[38] Thomas Richter-Trummer, Denis Kalkofen, Jinwoo Park, and Dieter Schmalstieg. Instant mixed reality lighting from casual scanning. In 2016 IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 27-36. IEEE, 2016. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 647, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 647, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 49, + 647, + 287, + 679 + ], + "type": "text", + "content": "[39] Darius Rückert, Linus Franke, and Marc Stamminger. Adop: Approximate differentiable one-pixel point rendering. ACM Transactions on Graphics (ToG), 41(4):1-14, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 681, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 681, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 681, + 287, + 712 + ], + "type": "text", + "content": "[40] Noah Snavely, Steven M Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. In ACM siggraph 2006 papers, pages 835-846. 2006. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 137 + ], + "type": "text", + "content": "[41] Pratul P Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T Barron. Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7495-7504, 2021. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 139, + 545, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 139, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 308, + 139, + 545, + 192 + ], + "type": "text", + "content": "[42] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 194, + 545, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 194, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 308, + 194, + 545, + 258 + ], + "type": "text", + "content": "[43] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8248–8258, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 260, + 545, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 260, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 308, + 260, + 545, + 315 + ], + "type": "text", + "content": "[44] Haithem Turki, Deva Ramanan, and Mahadev Satyanarayanan. Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12922–12931, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 316, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 380 + ], + "type": "text", + "content": "[45] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481-5490. IEEE, 2022. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 383, + 545, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 435 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 435 + ], + "type": "text", + "content": "[46] Bruce Walter, Stephen R Marschner, Hongsong Li, and Kenneth E Torrance. Microfacet models for refraction through rough surfaces. In Proceedings of the 18th Eurographics conference on Rendering Techniques, pages 195-206, 2007. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 437, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 545, + 480 + ], + "type": "text", + "content": "[47] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 482, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 545, + 525 + ], + "type": "text", + "content": "[48] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 526, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 526, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 526, + 545, + 578 + ], + "type": "text", + "content": "[49] Daniel N Wood, Daniel I Azuma, Ken Aldinger, Brian Curless, Tom Duchamp, David H Salesin, and Werner Stuetzle. Surface light fields for 3d photography. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 487-496. 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 581, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 581, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 581, + 545, + 635 + ], + "type": "text", + "content": "[50] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5438–5448, 2022. 2, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 636, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 690 + ], + "type": "text", + "content": "[51] Yao Yao, Jingyang Zhang, Jingbo Liu, Yihang Qu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Neilf: Neural incident light field for physically-based material estimation. In European Conference on Computer Vision, pages 700-716. Springer, 2022. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 692, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 545, + 712 + ], + "type": "text", + "content": "[52] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neu" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "5331" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 418 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "type": "text", + "content": "ral Information Processing Systems, 34:4805-4815, 2021. 2, 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "type": "text", + "content": "[53] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in neural information processing systems, 35:25018-25032, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 206 + ], + "type": "text", + "content": "[54] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snavely. Physg: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5453-5462, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 287, + 251 + ], + "type": "text", + "content": "[55] Qiang Zhang, Seung-Hwan Baek, Szymon Rusinkiewicz, and Felix Heide. Differentiable point-based radiance fields for efficient view synthesis. In SIGGRAPH Asia 2022 Conference Papers, pages 1-12, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 253, + 287, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 253, + 287, + 306 + ], + "spans": [ + { + "bbox": [ + 48, + 253, + 287, + 306 + ], + "type": "text", + "content": "[56] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 308, + 287, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 308, + 287, + 362 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 287, + 362 + ], + "type": "text", + "content": "[57] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Ner-factor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (ToG), 40(6):1-18, 2021. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 364, + 287, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 364, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 48, + 364, + 287, + 418 + ], + "type": "text", + "content": "[58] Yuanqing Zhang, Jiaming Sun, Xingyi He, Huan Fu, Rongfei Jia, and Xiaowei Zhou. Modeling indirect illumination for inverse rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18643-18652, 2022. 3" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5332" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Gear-NeRF_ Free-Viewpoint Rendering and Tracking with Motion-aware Spatio-Temporal Sampling/8ba31972-a909-4d60-aef6-d492fd479dbb_content_list.json b/2024/Gear-NeRF_ Free-Viewpoint Rendering and Tracking with Motion-aware Spatio-Temporal Sampling/8ba31972-a909-4d60-aef6-d492fd479dbb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..61156bf306e801b28c0272396ddf87f6b5aff806 --- /dev/null +++ b/2024/Gear-NeRF_ Free-Viewpoint Rendering and Tracking with Motion-aware Spatio-Temporal Sampling/8ba31972-a909-4d60-aef6-d492fd479dbb_content_list.json @@ -0,0 +1,2125 @@ +[ + { + "type": "text", + "text": "Gear-NeRF: Free-Viewpoint Rendering and Tracking with Motion-aware Spatio-Temporal Sampling", + "text_level": 1, + "bbox": [ + 214, + 111, + 756, + 157 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xinhang Liu $^{1\\dagger*}$ Yu-Wing Tai $^{2}$ Chi-Keung Tang $^{1\\dagger}$ Pedro Miraldo $^{3}$ Suhas Lohit $^{3}$ Moitreya Chatterjee $^{3}$ $^{1}$ HKUST $^{2}$ Dartmouth College $^{3}$ Mitsubishi Electric Research Laboratories (MERL) xliufe@connect.ust.hk, yu-wing.tai@dartmouth.edu, cktang@cse.ust.hk, miraldo@merl.com, slohit@merl.com, metro.smiles@gmail.com", + "bbox": [ + 138, + 184, + 831, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 308, + 313, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Extensions of Neural Radiance Fields (NeRFs) to model dynamic scenes have enabled their near photo-realistic, free-viewpoint rendering. Although these methods have shown some potential in creating immersive experiences, two drawbacks limit their ubiquity: (i) a significant reduction in reconstruction quality when the computing budget is limited, and (ii) a lack of semantic understanding of the underlying scenes. To address these issues, we introduce Gear-NeRF, which leverages semantic information from powerful image segmentation models. Our approach presents a principled way for learning a spatio-temporal (4D) semantic embedding, based on which we introduce the concept of gears to allow for stratified modeling of dynamic regions of the scene based on the extent of their motion. Such differentiation allows us to adjust the spatiotemporal sampling resolution for each region in proportion to its motion scale, achieving more photo-realistic dynamic novel view synthesis. At the same time, almost for free, our approach enables free-viewpoint tracking of objects of interest - a functionality not yet achieved by existing NeRF-based methods. Empirical studies validate the effectiveness of our method, where we achieve state-of-the-art rendering and tracking performance on multiple challenging datasets. The project page is available at: https://merl.com/research/highlights/gear-nerf.", + "bbox": [ + 75, + 340, + 473, + 720 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 751, + 209, + 767 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Reconstructing 3D scenes has a broad range of applications, including Virtual Reality/Augmented Reality (VR/AR), 3D animation, game production, and film creation which allow users to observe scenes from any desired viewpoint. While it is crucial to reconstruct static scenes, towards which significant progress has been made, it is even more crucial to", + "bbox": [ + 75, + 777, + 468, + 869 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/388e1c857a088b1e69eef4f1773783926afe85c265cffcd81b0d9b12d5e592d1.jpg", + "image_caption": [ + "Figure 1. (a) Our method takes RGB videos captured from a camera array as input. (b) Trained Gear-NeRF achieves photo-realistic real-time free-viewpoint rendering of a dynamic scene. (c) With users giving a single click at any time and from any viewpoint, our method can perform free-viewpoint tracking of the target object." + ], + "image_footnote": [], + "bbox": [ + 503, + 309, + 890, + 508 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "reconstruct dynamic scenes, as the world around us is characterized by a constant state of flux, with many objects in it - in a state of motion.", + "bbox": [ + 498, + 602, + 892, + 646 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in novel view synthesis, such as Neural Radiance Fields (NeRFs) [56] have inspired numerous studies to extend them to dynamic 3D scenes. Existing approaches either employ a deformation field to map neural fields from a given time to a canonical space [22, 45, 60, 61, 65, 99], or directly model dynamic scenes as a 4D space-time grid [1, 6, 26]. Though these methods offer improved rendering quality by utilizing more accessible inputs compared to previous solutions [21, 42, 43, 77], they still struggle to ensure rendering quality in low-resource settings, requiring carefully engineered efforts. Further, most dynamic radiance field approaches adopt a naive spatiotemporal sampling strategy, without discerning the different scales of motion across different regions in the scene.", + "bbox": [ + 496, + 648, + 893, + 859 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose to fix this issue by leveraging a semantic understanding of dynamic scenes. Intuitively, a reconstruction system aware of the distinction between static and dynamic regions in a scene can perform more focused", + "bbox": [ + 498, + 861, + 893, + 921 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work mainly done when XL was an intern at MERL.", + "bbox": [ + 94, + 883, + 382, + 895 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ XL and CT are supported in part by the Research Grant Council of the Hong Kong SAR grant no. 16201420.", + "bbox": [ + 76, + 896, + 468, + 921 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19667", + "bbox": [ + 480, + 955, + 517, + 965 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/500abea19269eadef36d0ba7557bc09d847c481cb0d319ca0167698cf0cf9a5a.jpg", + "image_caption": [ + "Figure 2. Pipeline of Gear-NeRF: Gear-NeRF takes multi-view videos as input. After optimizing the serial 4D feature volumes (Section 4.1), it maps space-time coordinates to a 4D semantic embedding (Section 4.2), in addition to the volume density and view-dependent radiance color. Regions with larger motion are automatically assigned higher gear levels (Section 4.3) and as a result, receive higher-resolution spatio-temporal sampling (Section 4.4). Furthermore, Gear-NeRF is capable of performing free-viewpoint tracking of a target object with prompts as simple as a user click (Section 4.5)." + ], + "image_footnote": [], + "bbox": [ + 81, + 77, + 893, + 300 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "sampling in the dynamic regions, which inherently require more resources per unit volume than static regions, due to their time-evolving nature. Accordingly, dynamic regions can be further stratified according to their scale of motion. To this end, this paper presents Gear-NeRF, a framework that leverages semantic embedding from powerful image segmentation models [38] for stratified modeling of 4D scenes. Gear-NeRF optimizes for a 4D semantic embedding, based on which we introduce the concept of gear to smartly determine the appropriate region-specific resolution of spatio-temporal sampling in the NeRF. Regions with larger motion scales are assigned higher gears, through our gear determination scheme and we accordingly perform higher-resolution spatio-temporal sampling. Empirical studies reveal that this motion-aware sampling strategy improves the quality of synthesized images, over competing approaches. As a by-product of our semantically embedded representation, we achieve free-viewpoint object tracking, given user prompts. Figure 1 presents an overview of the capabilities of our method.", + "bbox": [ + 75, + 388, + 472, + 691 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Gear-NeRF makes two primary advancements: (i) enhanced dynamic novel view synthesis by resorting to smarter spatio-temporal sampling, and (ii) the ability for free-viewpoint tracking of objects of interest. The latter is a capability not yet realized by existing NeRF methods for dynamic scenes. We perform extensive experiments on multiple datasets to validate the generalizability and robustness of our method, which shows state-of-the-art performances for both tasks across all datasets.", + "bbox": [ + 75, + 693, + 468, + 829 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 848, + 218, + 864 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural Radiance Fields: NeRF [56] is a recent breakthrough among novel view synthesis methods that uses multilayer perceptrons (MLPs) to parameterize the appearance", + "bbox": [ + 75, + 875, + 468, + 922 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and density for each point in 3D space, given any viewing direction of the scene. Researchers have extended NeRF along various dimensions [78], including improving rendering quality [2-4, 12, 18, 31], handling challenging conditions such as large scenes [55, 68, 75], view-dependent appearances [30, 52, 80], and sparse inputs [34, 51, 59, 86, 87, 93]. NeRF-like neural representations have also found applications in semantic segmentation [40, 50, 102] and 3D content generation [9, 10, 49, 64]. Recent work has shown that replacing the deep MLPs with a feature voxel grid can significantly improve training and inference speed [11, 25, 58, 74]. On the other hand, a more recent approach to further improve visual quality, rendering time, and performance entails representing the scene with 3D Gaussians [36]. Our approach, while drawing upon many of these approaches, deals with dynamic scenes which is beyond the scope of these methods.", + "bbox": [ + 496, + 388, + 893, + 645 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural Representations for Dynamic Scenes: NeRF-like representations have recently been extended to model dynamic scenes in high fidelity [17, 28, 33, 39, 46-48, 53, 63, 66, 84, 89, 90, 97]. One straightforward approach to do this is to directly condition the radiance field on time [27, 44, 45, 88]. Alternatively, several methods model a deformation field to map coordinates from different time stamps to a common canonical space [22, 24, 60, 61, 65, 79, 98, 99]. Some recent approaches [6, 26, 71, 81] represent the scene using a 4D space-time grid, which is decomposed into sets of planar representations for training efficiency. Other techniques for improving rendering fidelity and frame rate include Fourier PlenOctrees [82], ray-conditioned sample prediction networks [1], 4D space decomposition (static/dynamic/newly appeared regions) [73], and explicit voxel grids [24]. 3D Gaussians have also been adapted to model dynamic scenes [54, 85, 95, 96]. While these approaches paved the initial path for rendering dy", + "bbox": [ + 496, + 648, + 895, + 921 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19668", + "bbox": [ + 480, + 955, + 517, + 965 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "namic scenes, semantically aware modeling of the scene is absent, a caveat that our proposed method seeks to address. Segment Anything Model (SAM): SAM [38] is a powerful promptable image segmentation model, which showcases remarkable zero-shot generalization abilities and can produce semantically consistent masks, given a single foreground point on the image. HQ-SAM [35] is an improvement on SAM that enhances the quality of masks, especially on objects with intricate boundaries and structures. Recent works [20, 92] have extended SAM to perform interactive video object segmentation. These methods utilize SAM for mask initialization or correction and then employ state-of-the-art mask trackers [19, 94] for mask tracking and prediction [14]. Recent methods have also leveraged SAM for tracking multiple reference objects in a video [13, 101]. This work uses SAM to profile the scene into semantic regions, which are then grouped based on motion scales.", + "bbox": [ + 75, + 73, + 468, + 329 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Semantic Understanding: Existing methods for 3D visual understanding [15, 32, 76, 91] mainly focus on closed set segmentation of point clouds or voxels. NeRF's ability to integrate information across multiple views has led to its applications in 3D semantic segmentation [102], object segmentation [23, 50, 57], panoptic segmentation [72], and interactive segmentation [29, 69]. Kobayashi et al. [40] explored the effectiveness of embedding pixel-aligned features [7, 41] into NeRFs for 3D manipulations. LERF [37] fuses CLIP embeddings [67] and NeRFs to enable language-based localization in 3D NeRF scenes. Recent work has enabled click/text-based 3D segmentation by learning a 3D SAM embedding [16] or inverse rendering of SAM-generated masks [8]. We, on the other hand, seek to utilize the synergy of dynamic NeRFs and SAM segments to derive a semantic understanding of a dynamic 3D scene for tracking objects of interest in novel views – a first of its kind effort.", + "bbox": [ + 75, + 330, + 470, + 603 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 621, + 215, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Neural Radiance Fields (NeRFs): Vanilla NeRFs [56] employ a multi-layer perceptron (MLP) with sinusoidal positional encoding to map a 3D-spatial coordinate $\\mathbf{x} = (x, y, z)$ and a viewing direction $\\mathbf{d} = (\\theta, \\phi)$ to a volume density $\\sigma \\in [0,1]$ and an emitted RGB, $\\mathbf{c} \\in \\mathbb{R}^3$ . Rendering each image pixel involves casting a ray $\\mathbf{r}(t) = \\mathbf{o} + t\\mathbf{d}$ from the camera center $\\mathbf{o}$ through the pixel along direction $\\mathbf{d}$ . The predicted color for the corresponding pixel is computed as:", + "bbox": [ + 75, + 647, + 468, + 768 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {C}} (\\mathbf {r}) = \\sum_ {i = 1} ^ {N} T _ {i} \\alpha_ {i} \\mathbf {c} _ {i}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 782, + 468, + 823 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $T_{i} = \\exp \\left(-\\sum_{j=1}^{i-1} \\sigma_{j} \\delta_{j}\\right)$ , $\\alpha_{i} = 1 - \\exp (-\\sigma_{i} \\delta_{i})$ , and $\\delta_{j} = t_{j+1} - t_{j}$ . A vanilla NeRF is trained by minimizing the mean squared error between the input images and the predicted images, obtained by rendering the scene from the viewpoints from which the input images have been", + "bbox": [ + 75, + 838, + 468, + 921 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "captured, with the training loss given by:", + "bbox": [ + 500, + 73, + 771, + 88 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {p h o}} = \\sum_ {\\mathbf {r} \\in \\mathcal {R}} \\| \\hat {\\mathbf {C}} (\\mathbf {r}) - \\mathbf {C} (\\mathbf {r}) \\| _ {2} ^ {2}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 99, + 890, + 132 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{R}$ is the set of all rays projected from the input image. Planar-Factorized 4D Volumes: A recent emerging trend of handling dynamics using radiance field representations is to directly adapt them to be conditioned on a frame index $t$ (denoting time) in addition to $\\mathbf{x}$ and $\\mathbf{d}$ . This can be accomplished by learning a mapping from $(\\mathbf{x},\\mathbf{d},t)$ to $(\\sigma ,\\mathbf{c})$ using planar-factorized 4D volumes [1, 6, 26, 71]. These methods attempt to learn a 4D feature vector for every $(\\mathbf{x},t)$ , by projecting it to a set of 2D-planes. Embeddings of these projections on these planes can then be integrated to obtain the embedding for the 4D point. This can be mathematically represented as follows:", + "bbox": [ + 496, + 143, + 890, + 325 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {f} (\\mathbf {x}, t) = \\mathbf {B} _ {1} \\left(\\mathbf {h} _ {1} (x, y) \\odot \\mathbf {k} _ {1} (z, t)\\right) \\\\ + \\mathbf {B} _ {2} (\\mathbf {h} _ {2} (x, z) \\odot \\mathbf {k} _ {2} (y, t)) \\\\ + \\mathbf {B} _ {3} \\left(\\mathbf {h} _ {3} (y, z) \\odot \\mathbf {k} _ {3} (x, t)\\right). \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 339, + 890, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{h}_i(\\cdot ,\\cdot)$ and $\\mathbf{k}_i(\\cdot ,\\cdot)$ are functions (evaluated by bilinear interpolation on regularly spaced 2D feature grids) embedding coordinate tuples to features of dimension $M$ \" $\\odot$ denotes an element-wise product, and $\\mathbf{B}_i(\\cdot)$ denotes a linear transform which maps the products to feature vectors. Subsequently, a tiny MLP can map the feature vector $\\mathbf{f}(\\cdot ,\\cdot)$ to the volume density, $\\sigma$ , and the view-dependent emitted color, c, given the viewing direction d.", + "bbox": [ + 496, + 405, + 890, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Proposed Method", + "text_level": 1, + "bbox": [ + 498, + 542, + 671, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a set of $W$ input videos, $\\mathcal{V} = \\{V_1, V_2, \\dots, V_W\\}$ of a dynamic scene, with calibrated camera poses, our approach represents the scene using a series of 4D feature volumes (Section 4.1) along with 4D semantic embeddings (Section 4.2).", + "bbox": [ + 496, + 568, + 890, + 642 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Analogous to multiple gears in motor vehicles for optimizing engine performance, Gear-NeRF stratifies this semantically embedded scene representation into $N_{\\mathrm{gear}}$ levels, based on the motion scales. Each of these levels is called a gear. Through our training scheme, regions with larger motion are assigned higher gear levels (Section 4.3) and as a result, are more densely sampled (Section 4.4) for improved dynamic novel view synthesis. Our 4D semantic embedding also enables a new functionality, almost for free - free-viewpoint tracking of target objects, given simple user prompts like clicks (Section 4.5). Figure 2 shows the overall pipeline of Gear-NeRF.", + "bbox": [ + 496, + 643, + 890, + 825 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Serial 4D Feature Volumes", + "text_level": 1, + "bbox": [ + 498, + 835, + 741, + 851 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Instead of using a unified 4D volume to represent a dynamic scene [1, 6, 26, 71], our representation consists of a series of feature volumes, each corresponding to a gear level, $\\mathcal{G}$ . Specifically, for any space-time coordinate $(\\mathbf{x}, t)$ , its feature", + "bbox": [ + 496, + 859, + 890, + 921 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19669", + "bbox": [ + 480, + 955, + 517, + 965 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "vector corresponding to $\\mathcal{G}$ is computed as follows:", + "bbox": [ + 76, + 73, + 410, + 88 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {f} ^ {\\mathcal {G}} (\\mathbf {x}, t) = \\mathbf {B} _ {1} \\left(\\mathbf {h} _ {1} (x, y) \\odot \\mathbf {k} _ {1} ^ {\\mathcal {G}} (z, t)\\right) \\\\ + \\mathbf {B} _ {2} (\\mathbf {h} _ {2} (x, z) \\odot \\mathbf {k} _ {2} ^ {\\mathcal {G}} (y, t)) \\\\ + \\mathbf {B} _ {3} \\left(\\mathbf {h} _ {3} (y, z) \\odot \\mathbf {k} _ {3} ^ {\\mathcal {G}} (x, t)\\right). \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 95, + 468, + 157 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The vector-valued functions $\\mathbf{h}_j(\\cdot, \\cdot)$ and linear transforms $\\mathbf{B}_j(\\cdot)$ are shared by all gears, while each gear has its own spatio-temporal embedding $\\mathbf{k}_j^G(\\cdot, \\cdot)$ , in $M$ -dimensional space. Therefore, each gear describes regions of a certain scale of motion while the purely spatial features can be shared among all gears.", + "bbox": [ + 76, + 167, + 468, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We obtain the gear level at any spatio-temporal coordinate also from a planar-factorized 4D feature volume. Specifically, the gear level at $(\\mathbf{x},t)$ is computed as:", + "bbox": [ + 76, + 258, + 468, + 305 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} g (\\mathbf {x}, t) = \\mathbf {1} ^ {\\top} \\left(\\mathbf {h} _ {1} ^ {\\prime} (x, y) \\odot \\mathbf {k} _ {1} ^ {\\prime} (z, t)\\right) \\\\ + \\mathbf {1} ^ {\\top} \\left(\\mathbf {h} _ {2} ^ {\\prime} (x, z) \\odot \\mathbf {k} _ {2} ^ {\\prime} (y, t)\\right) \\\\ + \\mathbf {1} ^ {\\top} \\left(\\mathbf {h} _ {3} ^ {\\prime} (y, z) \\odot \\mathbf {k} _ {3} ^ {\\prime} (x, t)\\right), \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 314, + 468, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{1}$ is a vector of ones, $\\mathbf{h}_i^{\\prime}(\\cdot ,\\cdot)$ and $\\mathbf{k}_i^\\prime (\\cdot ,\\cdot)$ are $M$ dimensional embedding functions. This however defines a continuous feature volume. To map it to the gear level integers, we apply the following projection operation:", + "bbox": [ + 76, + 383, + 468, + 446 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np (\\mathbf {x}, t) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} g (\\mathbf {x}, t) < 1, \\\\ N _ {\\text {g e a r}}, & \\text {i f} g (\\mathbf {x}, t) \\geq N _ {\\text {g e a r}}, \\\\ \\lceil g (\\mathbf {x}, t) \\rceil , & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 455, + 468, + 511 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on this gear level volume, we define a 4D mask for a region at gear level $\\mathcal{G}$ as:", + "bbox": [ + 76, + 521, + 468, + 551 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nm _ {\\mathcal {G}} (\\mathbf {x}, t) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} p (\\mathbf {x}, t) = \\mathcal {G}, \\\\ 0, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 561, + 468, + 602 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The final feature vector at $(\\mathbf{x},t)$ is computed as:", + "bbox": [ + 76, + 612, + 395, + 628 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} (\\mathbf {x}, t) = \\sum_ {\\mathcal {G} = 1} ^ {N _ {\\text {g e a r}}} m _ {\\mathcal {G}} (\\mathbf {x}, t) \\mathbf {f} ^ {\\mathcal {G}} (\\mathbf {x}, t). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 638, + 468, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Subsequently, a tiny MLP, $F_{\\theta}$ , maps these feature vectors $\\mathbf{f}(\\cdot ,\\cdot)$ as well as the viewing direction $\\mathbf{d}$ to the volume density $\\sigma$ and radiance color $\\mathbf{c}$ . This allows us to obtain a photometric rendering of the scene.", + "bbox": [ + 76, + 691, + 468, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. 4D Semantic Embedding", + "text_level": 1, + "bbox": [ + 76, + 762, + 303, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Gear-NeRF leverages the strong object priors of the SAM [38] model to acquire a semantic understanding of the scene, for improved photometric rendering (Section 4.3 and Section 4.4) as well as free-viewpoint object tracking (Section 4.5). Toward this end, we utilize the SAM encoder to obtain 2D feature maps from the frames of each video. We then optimize a 4D SAM embedding field by supervising it with these 2D feature maps. In particular, the MLP above, $F_{\\theta}$ , is configured to output a 4D semantic embedding s for", + "bbox": [ + 76, + 784, + 468, + 921 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e3fa72dc36fd506e57f807a0b5f7022d82a22a739737276e39de2b2edcc87568.jpg", + "image_caption": [ + "$1^{\\mathrm{st}}$ Gear Assignment Update" + ], + "image_footnote": [], + "bbox": [ + 511, + 71, + 632, + 141 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/50fa637dd9b5cd1d8ef5c3734619da2a8b18532884ebf3307e56e6012fac32e4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 71, + 754, + 141 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e23f902aa511437326737a562472d7caf86849a36e8cd344b80d94bdfa92a64a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 759, + 71, + 880, + 141 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d87f3f86106845ae486f9020b20c21ed3ab64b99550e216c7f09eea295157c5c.jpg", + "image_caption": [ + "2nd Gear Assignment Update" + ], + "image_footnote": [], + "bbox": [ + 511, + 155, + 632, + 224 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/47af5c1ed79559ca8d3a83c738ee9115e1dc867b9daf89f179fe35cfc79ac20c.jpg", + "image_caption": [ + "$3^{\\mathrm{rd}}$ Gear Assignment Update" + ], + "image_footnote": [], + "bbox": [ + 635, + 155, + 754, + 224 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/440fbf6fbb39d32553f1849c7425139eb95591a5e7cf66188db92618d076026a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 759, + 155, + 880, + 224 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2943450a8edc3e657892f88efead7cc1277c40ff2564f801a6aba17522ec63e1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 511, + 238, + 632, + 308 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/dbea8c9a4761b4dfee689b657b03a59d77398d43bea0d872794a3f642fdd62ea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 238, + 754, + 308 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/07dc913a232a4ca56bdf8c7f81a07d84290cd2bbaf7e7b9422c685494c190bec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 759, + 238, + 880, + 308 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e66b9f16b1e8424443e7fdb89e5a5849e28053e972b233b72ef7e949bedcee11.jpg", + "image_caption": [ + "Rendered Image", + "Figure 3. Illustration of Gear Assignment Update: For each gear assignment update, we calculate the rendering loss map between the rendered RGB-SAM map and the ground truth and identify the centers of the patches with the maximum and minimum losses, marked in red and green (second column). These points are then fed into the SAM decoder as positive and negative prompts to generate an upshift mask representing the areas that need to be shifted to a higher gear (last column). After the first gear assignment update, we see that the next candidate region for upshift is situated where the horse is located, and so on. Upshift mask colors imply the gear levels after the update (blue-2, green-3, red-4)." + ], + "image_footnote": [], + "bbox": [ + 511, + 321, + 632, + 392 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b7d48a34fdb91fa0479f22030faadcd57dedd2e93119424b892303001284c9bd.jpg", + "image_caption": [ + "$4^{\\mathrm{th}}$ Gear Assignment Update", + "Rendering Loss Map" + ], + "image_footnote": [], + "bbox": [ + 635, + 321, + 754, + 392 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e53c1f54b24deda748147abfcf6d354e8558e6bd4ac49d6499b0bdd54d13466f.jpg", + "image_caption": [ + "Upshift Mask" + ], + "image_footnote": [], + "bbox": [ + 759, + 321, + 880, + 392 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "a given space-time coordinate in addition to the density, $\\sigma$ , and color, c. To render 2D semantic feature maps in a given view, we compute the semantic feature of a pixel in the feature map by tracing a ray through it and perform volume rendering analogous to Equation 1, as follows:", + "bbox": [ + 496, + 578, + 890, + 654 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {S}} (\\mathbf {r}) = \\sum_ {i = 1} ^ {N} T _ {i} \\alpha_ {i} \\mathbf {s} _ {i}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 664, + 890, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This SAM embedding is supervised by minimizing the mean squared error between the prediction and the ground truth features $(\\mathbf{S}(\\mathbf{r}))$ from the SAM encoder, as shown:", + "bbox": [ + 496, + 714, + 890, + 760 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {S A M}} = \\sum_ {\\mathbf {r} \\in \\mathcal {R}} \\| \\hat {\\mathbf {S}} (\\mathbf {r}) - \\mathbf {S} (\\mathbf {r}) \\| _ {2} ^ {2}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 770, + 890, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Training Scheme with Gear Assignment", + "text_level": 1, + "bbox": [ + 498, + 821, + 841, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With gear initialization $g(\\mathbf{x}, t) = 1, \\forall \\mathbf{x}, t$ , the (semantically embedded) radiance field optimization and gear assignment updating take place in an alternating fashion.", + "bbox": [ + 496, + 845, + 890, + 890 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Gear Assignment Update: As illustrated in Figure 3, when updating the gear assignment after a period of radiance field", + "bbox": [ + 498, + 891, + 890, + 921 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19670", + "bbox": [ + 480, + 955, + 517, + 965 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "optimization, we find the regions rendered most poorly from the rendering loss maps and increment their gears for denser spatio-temporal sampling. The following steps lay out the process for updating gear assignments to regions:", + "bbox": [ + 75, + 73, + 468, + 133 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We sample a number of viewpoints and time steps and render 2D-images/SAM features for it. For every rendered RGB-SAM map, we compute a rendering loss map. Each pixel of the rendering loss map is computed as: $\\mathcal{L}(\\mathbf{r})\\coloneqq \\mathcal{L}_{\\mathrm{pho}}(\\mathbf{r}) + \\lambda \\mathcal{L}_{\\mathrm{SAM}}(\\mathbf{r})$ . See Figure 3 for example loss maps.", + "- Next, we patchify each rendering loss map to find patches with the top- $k$ largest/smallest average loss. The center coordinate of these patches serve as positive/negative prompts for the next step. See Figure 3 for example positive (red) / negative (green) prompts.", + "- We then feed the ground truth RGB images together with positive and negative prompts into the SAM decoder [38] to estimate an upshift mask. These masks tend to cover regions that have motions and are not satisfactorily rendered with the current sampling resolution. We have multiple upshift masks at different viewpoints and time steps.", + "- For every pixel of an upshift mask, we trace a ray and sample points along it and update the gear assignment by pushing $g(\\mathbf{x}, t)$ towards incremented values." + ], + "bbox": [ + 76, + 133, + 467, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In particular, in the last step, for each pixel within an upshift mask, a corresponding ray is traced that connects it with the camera center o, along direction d. Next, a set of points are sampled along this ray. The collection of sampled points, that lie on the rays emanating from within the masked region constitutes the set $S_{\\mathrm{upshift}} = \\{(x_i^{\\mathrm{upshift}},t_i^{\\mathrm{upshift}})\\}_{i = 1}^{N_{\\mathrm{upshift}}}$ , where $N_{\\mathrm{upshift}}$ represents the total count of points sampled from rays pertaining to the masked region. We then follow a similar procedure to sample a set of points pertaining to the unmasked region. We denote this set as: $\\mathcal{S}_{\\mathrm{stay}} = \\{(x_i^{\\mathrm{stay}},t_i^{\\mathrm{stay}})\\}_{i = 1}^{N_{\\mathrm{stay}}}$ , with $N_{\\mathrm{stay}}$ indicating the total number of points sampled from rays that pertain to the unmasked area. Next, for each sample point in each of $S_{\\mathrm{upshift}}$ and $S_{\\mathrm{stay}}$ , we query its current gear level $p(\\mathbf{x},t)$ . In order to assign new gear levels, we need to update the gear assignment function $g(\\cdot ,\\cdot)$ , which proceeds with the objective function:", + "bbox": [ + 75, + 436, + 467, + 696 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {u p s h i f t}} = \\frac {1}{N _ {\\text {u p s h i f t}}} \\sum_ {(\\mathbf {x}, t) \\in \\mathcal {S} _ {\\text {u p s h i f t}}} \\| g (\\mathbf {x}, t; \\boldsymbol {\\Theta}) - (p (\\mathbf {x}, t) + 1) \\| _ {2} ^ {2} \\\\ + \\frac {\\lambda_ {\\text {s t a y}}}{N _ {\\text {s t a y}}} \\sum_ {(\\mathbf {x}, t) \\in \\mathcal {S} _ {\\text {s t a y}}} \\| g (\\mathbf {x}, t; \\boldsymbol {\\Theta}) - p (\\mathbf {x}, t) \\| _ {2} ^ {2}, \\tag {11} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 707, + 467, + 799 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Theta$ denotes the set of estimizable parameters for $g(\\cdot ,\\cdot)$ . The minimization of the first term encourages sample points within the masked region $S_{\\mathrm{upshift}}$ to have a gear level equal to their current gear incremented by one, resulting in an upshift of gear. Conversely, minimizing the second term encourages the points in $S_{\\mathrm{stay}}$ to maintain their gear values, thereby encouraging the remaining regions to keep their current gear levels and avoiding unwanted up", + "bbox": [ + 75, + 801, + 467, + 921 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "shifts. We update $\\Theta$ via a single step of gradient descent as follows:", + "bbox": [ + 500, + 73, + 888, + 102 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Theta} \\leftarrow \\boldsymbol {\\Theta} - \\alpha \\nabla_ {\\boldsymbol {\\Theta}} \\mathcal {L} _ {\\text {u p s h i f t}}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 111, + 888, + 127 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha$ is the learning rate. Since, $g(\\cdot ,\\cdot)$ is essentially derived from the embedding functions, $\\mathbf{h}_i^{\\prime}(\\cdot ,\\cdot)$ and $\\mathbf{k}_i^\\prime (\\cdot ,\\cdot)$ for $i\\in \\{1,2,3\\}$ , the aforementioned optimization step amounts to updating these embedding functions. Once updated, we proceed with a fresh round of gear assignment to increase the spatio-temporal sampling resolution for the regions that end up at a higher gear level than before.", + "bbox": [ + 496, + 142, + 890, + 248 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Radiance Field Optimization: With the updated gear assignment, we increase the resolution of spatio-temporal sampling for the gear-shifted regions (Section 4.4) and then resume optimizing the radiance field. We alternate between the two processes: radiance field optimization (each time for $L$ epochs), and gear assignment updates until the average variance of each rendering loss map is below a predetermined threshold. After this, we optimize the radiance field for an additional $L'$ epochs without further gear assignment updates.", + "bbox": [ + 496, + 251, + 890, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Motion-aware Spatio-Temporal Sampling", + "text_level": 1, + "bbox": [ + 498, + 421, + 854, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this subsection, we explain our motion-aware spatiotemporal sampling strategy based on assigned gears, permitting differential processing of regions at different gear levels. By temporal sampling, we imply the choice of temporal resolution for planar-factorized 4D feature volumes, and by spatial sampling, we mean the strategy used to choose sampling points along each ray for volume rendering.", + "bbox": [ + 496, + 446, + 890, + 568 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Motion-aware Temporal Sampling: To handle the increasing intensity of object motion, as reflected by their growing gear levels, we increment the temporal resolution for voxel grids. Specifically, $\\mathbf{k}_j^{\\mathcal{G}}$ in Equation 4 has increasing resolution along the time axis, thereby empowering the 4D feature volumes to better model the dynamics along the temporal axis. This ensures fast-moving objects can be more faithfully modeled without unsightly blurring. The temporal resolution for each gear's feature volume is determined by linear interpolation between 1 (for $\\mathcal{G} = 1$ ) and the total number of frames (for $\\mathcal{G} = N_{\\mathrm{gear}}$ ).", + "bbox": [ + 496, + 570, + 890, + 737 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Motion-aware Spatial Sampling: While denser sampling of points can improve reconstruction accuracy, increasing the number of sampling points throughout the scene can lead to prohibitive computational costs. Therefore, we propose a 3D point-splitting strategy as illustrated in Figure 4. We begin by sampling a relatively small number, $n$ , of samples along each ray, assuming it is at the lowest gear level. If a sampled point belongs to a region with a higher gear, as determined by $p(\\mathbf{x}, t)$ , we then sample more densely in that region. For every sampled point in that region, we split it into $2^{p(\\mathbf{x}, t) - 1}$ points, equally spaced within the corresponding ray segment (at that gear level).", + "bbox": [ + 496, + 739, + 890, + 921 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19671", + "bbox": [ + 480, + 955, + 517, + 965 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/67bfcb663407705ef6fd0df36ebff6879367fd9d28578ab337310359da5d1ed8.jpg", + "image_caption": [ + "Figure 4. Motion-aware Spatial Sampling: We split each sampled point into $2^{p(\\mathbf{x},t)}$ points, equally spaced within the corresponding ray segment. The top row shows the vanilla uniformly sampled points, while the bottom one shows the densely sampled points after splitting using our strategy." + ], + "image_footnote": [], + "bbox": [ + 99, + 70, + 433, + 189 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.5. Free-Viewpoint Tracking with User Prompts", + "text_level": 1, + "bbox": [ + 76, + 279, + 455, + 296 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our 4D SAM embedding enables another useful functionality, almost for free - free-viewpoint object tracking, where the user only needs to provide as few as one click to extract the target object based on the 4D embedding. Next, we describe how, given a user-supplied point click at any arbitrary viewpoint and time step, we obtain the corresponding object mask at a novel viewpoint and time step.", + "bbox": [ + 75, + 303, + 467, + 409 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Masks for Novel Viewpoints: The first step for this task entails finding the 3D correspondence of the user click. We trace a ray through the selected pixel, and by utilizing the volume density, we determine the depth at which the ray intersects with the first object surface it encounters. This yields the 3D coordinates of the point of intersection. Subsequently, the 3D coordinates of this intersection can be easily mapped into a 2D coordinate within any novel viewpoint image, using the camera pose of the new viewpoint. Alongside the rendered SAM feature map of the novel view, we feed this coordinate into the SAM decoder to generate the object mask for the novel view.", + "bbox": [ + 75, + 410, + 467, + 589 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Masks for Novel Time Steps: For this task, we propagate an object mask to its neighboring time step. Specifically, with an object mask for a specific frame $t$ , we calculate the bounding box of this mask and use this bounding box as a prompt to SAM for neighboring frames $t' = t + 1$ or $t' = t - 1$ . By inputting this prompt along with the rendered SAM feature map at $t'$ into the SAM decoder, we can obtain the object mask for $t'$ . Combining the above two processes, we can start from a single click and get the object mask in any viewpoint and time step.", + "bbox": [ + 75, + 592, + 467, + 743 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 758, + 209, + 776 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We assess the performance of our proposed Gear-NeRF for dynamic novel view synthesis and free-viewpoint object tracking across a range of challenging datasets, comparing it with state-of-the-art methods. Through ablation studies, we provide empirical evidence of the effectiveness of its fundamental components. We kindly refer the reader to our supplementary material for additional experimental details and results, including videos for free-viewpoint rendering and tracking.", + "bbox": [ + 75, + 784, + 467, + 921 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 500, + 71, + 689, + 88 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details: We implement our method using PyTorch [62] and conduct experiments on an NVIDIA RTX 4090 GPU with 24 GB RAM. We divide each input video into chunks of 100 frames. For every chunk, we train a model for approximately 2.5 hours. Our 4D feature volumes yield embeddings with a dimension of $M = 32$ . We set the gear number $N_{\\mathrm{gear}}$ to 4. We find patches with top- $k = 3$ largest/smallest average loss for gear assignment updates to obtain prompts. The rendering loss map is computed with $\\lambda = 0.01$ . For the optimization of radiance fields, $L = 3$ and $L' = 10$ . In our motion-aware spatial sampling, each ray initially has $n = 64$ sampling points. We use an initial learning rate of 0.02 for all the parameters. For Equation 12, we use $\\alpha = 0.02$ .", + "bbox": [ + 496, + 98, + 890, + 306 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets: (i) The Technicolor light field dataset [70] includes diverse indoor environment videos captured by a $4 \\times 4$ camera rig. We evaluate on 4 sequences (Train, Theater, Painter, Birthday) at the original $2048 \\times 1088$ resolution, holding out the same view as prior work [1] (the second row and second column) for evaluation. (ii) The Neural 3D Video dataset [44] includes indoor multi-view video sequences captured by 20 cameras at a resolution of $2704 \\times 2028$ pixels. We experiment on 6 sequences (Cut Roasted Beef, Flame Steak, Coffee Martini, Cook Spinach, Flame Salmon, Sear Steak), downsampling by a factor of 2 and holding out the central view (akin to prior work [1]) for evaluation. (iii) The Google Immersive dataset[5] contains light field videos of indoor and outdoor scenes captured by a time-synchronized 46-fisheye camera rig, with a resolution of $2560 \\times 1920$ pixels. We experiment with 9 sequences from it (Flames, Truck, Horse, Car, Welder, Exhibit, Face Paint 1, Face Paint 2, Cave). We downsample the video resolution by a factor of 2 and hold out the central view (like prior work [1]) for evaluation. For our experiments, we adopted the same resolution and held-out view selection as prior work [1].", + "bbox": [ + 496, + 311, + 890, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics: For the task of novel view synthesis of dynamic scenes, we evaluate using the following standard metrics: (i) Peak Signal-to-Noise Ratio (PSNR), (ii) Structural Similarity Index Measure (SSIM) [83] and (iii) Learned Perceptual Image Patch Similarity (LPIPS) [100], by comparing the reconstructed frames against ground truth images. These metrics are computed on the held-out view and averaged across all frames. For the free-viewpoint object tracking task, we designate a specific viewpoint and time step for the user to give prompting clicks. Subsequently, we assess the quality of the predicted object masks at novel viewpoints. The quality of the object mask is quantified in terms of the Mean Intersection over Union (mIoU) and accuracy (Acc.), which are calculated against the ground truth mask, manually annotated utilizing Adobe Photoshop. Additionally, we present the same metrics computed for novel time steps, denoted as t-mIoU and t-Acc.", + "bbox": [ + 496, + 646, + 890, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines: We run a comprehensive comparison of our", + "bbox": [ + 500, + 905, + 890, + 920 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19672", + "bbox": [ + 480, + 954, + 517, + 965 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/995f64513dc7dcc780c6471fd2e6af6df9a29b5181a29b9bd39043bebb702214.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 117, + 71, + 299, + 176 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dc52ec33367132fb857c04026dd2a395088ccd663d23f174067dcaabf1cbdca7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 300, + 71, + 478, + 176 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d4db5f3a30ed9868c493d8acb4701fa98355ebc454111ac934fbdec4a31fc069.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 482, + 71, + 661, + 176 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a2cb6d42855b09e182add840584accf719154eb6b74ae6f1764b137bdc28e1d2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 665, + 71, + 844, + 176 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dfe3e86592245362b13d0599a3240e10d819a2960401b7398bb90495e4d0b5b1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 117, + 180, + 299, + 286 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e6fe6e58d2bf1f0e4a8852645168917a64beb9d389006baf35e56d2beef68533.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 300, + 181, + 480, + 286 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f26ffd8f80ec8c5d2c9669fe25d385d4d25418244054b97e1c9eec17915a6003.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 181, + 661, + 285 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a25bb988e9ffcd1291ae92e5a376aff57865a49f26eddafc48e22c8280f74f5d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 665, + 181, + 844, + 286 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f659904e0228a212bd8293989fd4201852d0d9015be72fb69713b180863352b7.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 117, + 290, + 299, + 395 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/466c8a2e5a7934f4152f4690fde1baa9090778ebd4b92c1135daf028aab74129.jpg", + "image_caption": [ + "HyperReel" + ], + "image_footnote": [], + "bbox": [ + 300, + 291, + 480, + 393 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2072749a5c5a3b280a167ec83d172e289318ca9278b7384a4a7c0e2bb1d8e93e.jpg", + "image_caption": [ + "Neural 3D Video", + "Figure 5. Qualitative comparisons for novel view synthesis on the Technicolor dataset [70]: We qualitatively compare our approach against HyperReel [1] and Neural 3D Video [44]. Our approach better recovers fine details like patterns on the toys or stripes on the shirt." + ], + "image_footnote": [], + "bbox": [ + 483, + 291, + 661, + 393 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2a982f4a69cc86b234e01c70e77fc14a9d2325513fb985926b171a73241f5754.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 665, + 291, + 844, + 393 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "method against a range of recent NeRF-based baseline methods: (i) ST-NeRF [99], (ii) HexPlane [6], (iii) HyperReel [1], and (iv) MixVoxels [81]. As the first method to enable promptable free-viewpoint object tracking under the NeRF setting, there is no established direct baseline for this specific task. However, SA3D [8], a recent method for segmenting static scenes, is treated as a baseline for predicting masks at novel viewpoints corresponding to the prompted time step.", + "bbox": [ + 75, + 457, + 472, + 594 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Results", + "text_level": 1, + "bbox": [ + 76, + 607, + 171, + 622 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Dynamic Novel View Synthesis: As shown in Figure 5, Gear-NeRF produces high-quality novel view synthesis of dynamic scenes, accurately modeling real-world scenes with intricate motions and fine details. For example, patterns on the toys or stripes on the shirt, are faithfully rendered, resulting in more photo-realistic images compared to all of the baselines. Table 1, presents quantitative comparisons of our method against the baselines. While Gear-NeRF has longer training (Tr. Time) / inference times (FPS) compared to some baselines, it almost always achieves the best performance in terms of rendering quality.", + "bbox": [ + 75, + 632, + 468, + 799 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Free-Viewpoint Tracking: In Figure 6, we present the object masks obtained by our method based on the user prompts provided at a specified viewpoint and time step. Specifically, the first row displays masked objects at the prompted viewpoint and time step. The second row shows novel view masks at the prompted time step. The third row shows novel view masks at novel time steps. We see that masks obtained from Gear-NeRF show precise boundaries,", + "bbox": [ + 75, + 800, + 470, + 921 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/61c795a2452bec2f26f2300b03a2323a1890dab704922567c5ace877ef5aa187.jpg", + "table_caption": [ + "Table 1. Quantitative comparisons for dynamic novel view synthesis: Our method outperforms all baselines across all datasets on all metrics. We report means over all scenes for each dataset. Best and second best results are highlighted." + ], + "table_footnote": [], + "table_body": "
DatasetMethodPSNR (↑)SSIM (↑)LPIPS (↓)Tr. Time(↓)FPS(↑)
Technicolor [70]ST-NeRF [99]30.860.8830.101344 min0.7
HyperReel [1]31.040.8870.09297 min7.7
MixVoxels [81]28.990.8420.10314 min18.9
Ours32.210.9190.058148 min7.9
Google Immersive [5]HexPlane [6]27.670.8080.188527 min0.9
HyperReel [1]28.320.8620.145117 min7.4
MixVoxels [81]27.140.8350.20919 min16.8
Ours28.740.8760.122189 min7.0
Neural 3D Video [44]ST-NeRF [99]31.030.8900.081367 min0.7
HyperReel [1]31.120.9280.065129 min6.1
MixVoxels [81]30.690.9440.13915 min16.3
Ours31.800.9360.058204 min6.8
", + "bbox": [ + 500, + 512, + 890, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "compared to SA3D. Table 2 presents quantitative assessment of the quality of the masks generated of our method versus SA3D. Our method exceeds $90\\%$ across all metrics and datasets, demonstrating the effectiveness of our approach. Our gains over SA3D can be attributed to the fact that SA3D, as a static scene segmentation method, does not utilize information across all time frames, whereas our approach does. SA3D is incapable of predicting masks for novel time steps, and as such, the corresponding entries are marked as not applicable, a shortcoming which our method does not have.", + "bbox": [ + 496, + 645, + 890, + 810 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Evaluations", + "text_level": 1, + "bbox": [ + 500, + 821, + 627, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To verify the effectiveness of the design choices of GearNeRF, we perform extensive ablation studies on the Truck scene of the Google Immersive dataset [5].", + "bbox": [ + 496, + 845, + 890, + 888 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Motion-aware Temporal Sampling: An intuitive strategy for adjusting the temporal sampling is to directly modify", + "bbox": [ + 498, + 891, + 890, + 921 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19673", + "bbox": [ + 480, + 955, + 517, + 965 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c72389cbbf17f9d3b80ca8280e02a5715b7ac600a62094223b1da648b940c8bf.jpg", + "image_caption": [ + "★ User Prompt" + ], + "image_footnote": [], + "bbox": [ + 240, + 74, + 392, + 204 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/584ffa2ede060c98cbd97fc22a40b6418869315379b5d85f7db39d9701e145c9.jpg", + "image_caption": [ + "Rendering from Novel View", + "Rendering from Novel View at Novel Time", + "Rendered Image", + "Figure 6. Qualitative comparisons of free-viewpoint object tracking on Technicolor [70] and Neural 3D Video [44] datasets: Our method can obtain desirable object masks, with clear edges, from prompting points provided by users at desired time steps and viewpoints." + ], + "image_footnote": [], + "bbox": [ + 240, + 205, + 392, + 268 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4230cacf43138d46dfe8dd3e44d0a838b064007f64f61bfececcf09953f8dbfd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 413, + 79, + 457, + 135 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8d6c4c78db77b3f57a91aacb3d8bbe92fc2654618baf82e02bfdbb80f9ba8d21.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 397, + 143, + 467, + 203 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/94b307177c859840ad0d1b8da6483e1303c8216e5b89e585e9a8d613e74eb3ba.jpg", + "image_caption": [ + "SA3D" + ], + "image_footnote": [], + "bbox": [ + 397, + 204, + 480, + 268 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4c70833af9728c2eaae13facae9812e0ae2642aadc1180731d70446c4038e3f9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 75, + 545, + 135 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1bc3f75de1bf3b7d6e9526f53bd2708669632d330658a88a0eb27d7d6eb45781.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 142, + 550, + 200 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3032b11ac1cc1e51c64f9c1849c499d2a0dec2be2e55d8d0fdf793875c7bf1ea.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 498, + 210, + 550, + 263 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/81483608b12406fd56271932d4a42eb35caf3c997ce397cbfc91db5d0df4d59d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 568, + 74, + 678, + 138 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d7b36c0da5e6aaa56cc01d6895d00a1603c0c597853ebdd657e7b17e615c259e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 568, + 140, + 678, + 203 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/74f70bbb2b3b9c1d85479f1ef17b95d46dea3fd1bac1b504f5471667aec3cf79.jpg", + "image_caption": [ + "Rendered Image" + ], + "image_footnote": [], + "bbox": [ + 568, + 204, + 678, + 268 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e9ce9c240794324dbf67d0d133faf98920106b37155c4d8da9cf7eb2431354d4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 77, + 746, + 137 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/baf9ccfe19d602f96e0583f966b0389b237d013d4b5c1834fcb94e38b5f91ac3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 150, + 763, + 198 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5d20b5754326b29d9840ff2b282124d7e4f00b45bebe71696a0fda8c374283c5.jpg", + "image_caption": [ + "SA3D" + ], + "image_footnote": [], + "bbox": [ + 681, + 204, + 764, + 268 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a08f7e5b49cc3ef1f811bd27167ac01f75380d1e696b8c7b05f80fd7b75354ec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 787, + 75, + 831, + 125 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bfb82b569b7877578e997be0670a95f7e91ac38b0f09965fc3683e683d324a64.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 785, + 148, + 830, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d8079c3abe7497ee3ef2e41c709929aa5bdbc990a7e0582a04be49801357d3d9.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 779, + 215, + 826, + 262 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e7e5c38ccf57c7d2e54616aecff35052013dbeecc13b0669b005ae95d70209f1.jpg", + "table_caption": [ + "Table 2. Quantitative comparisons for free-viewpoint tracking: t-mIoU and t-Acc are metrics used for evaluating novel view masks at novel time steps, not applicable to SA3D. Reported metrics are averages over all scenes for each dataset." + ], + "table_footnote": [], + "table_body": "
DatasetMethodmIoU (↑)Acc. (↑)t-mIoU (↑)t-Acc. (↑)
Technicolor [70]SA3D [8]96.497.1N/AN/A
Ours97.497.692.193.3
Google Immersive [5]SA3D [8]94.194.8N/AN/A
Ours94.395.091.592.8
Neural 3D Video [44]SA3D [8]93.194.0N/AN/A
Ours93.494.390.692.3
", + "bbox": [ + 76, + 388, + 468, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the temporal resolution of the 4D feature volume and make it uniform across all regions of the scene. In Table 3, we demonstrate the performance of this naive temporal sampling strategy using medium (25) or dense (100) temporal resolutions on 100-frame input videos. The results show that increasing the temporal resolution universally does not necessarily yield better performance. In contrast, our proposed method, which utilizes motion-aware temporal sampling, achieves the best results. We surmise that this may be attributed to the model distributing the volume's capacity sparsely across a large number of time steps, which is not the case for our method.", + "bbox": [ + 75, + 492, + 468, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Motion-aware Spatial Sampling: To validate the effectiveness of our motion-aware spatial sampling strategy, we compare it against other sampling strategies. The first variant involves uniform sampling along the ray, and the second uses a sample prediction network (SPN) like Attal et al. [1] for sampling. These variants sample 128 points on each ray. Results in Table 3 show that our full model employing motion-aware spatial sampling outperforms these variants, perhaps by having a better sense of where to sample more from, derived from its semantic-aware embedding.", + "bbox": [ + 75, + 676, + 468, + 828 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "SAM embedding: To verify that introducing SAM embedding can improve rendering quality (rather than just enabling segmentation or tracking), we tested a variant of Gear-NeRF without SAM embedding (it instead thresholds loss maps of rendered RGB frames to obtain upshift masks). As shown in Table 3, the absence of SAM em", + "bbox": [ + 75, + 830, + 468, + 921 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4d826034e68a4163178c44706ad90cb01fb0737a3881d4442434a9489a1dc18d.jpg", + "table_caption": [ + "Table 3. Ablation Study: Ablations on our spatio-temporal sampling strategy and the number of gears (Truck/Google Immersive). Best and second best results are highlighted." + ], + "table_footnote": [], + "table_body": "
MethodPSNR (↑)SSIM (↑)LPIPS (↓)
Naive temporal sampling (medium)26.930.7780.144
Naive temporal sampling (dense)26.850.7600.161
Naive spatial sampling24.800.7340.222
SPN spatial sampling26.540.7870.162
Ours (w/o embedding)27.100.8310.177
Ours (Ngear = 2)26.930.7850.166
Ours (Ngear = 3)27.140.8750.145
Ours (Ngear = 4)27.490.8920.136
Ours (Ngear = 5)27.460.9010.131
", + "bbox": [ + 500, + 375, + 890, + 529 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "bedding for guiding gear assignment reduces the model's rendering quality.", + "bbox": [ + 500, + 532, + 890, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Number of Gears: We ablate on the numbers of gears. Increasing the number of gears allows for more fine-grained motion-aware spatio-temporal sampling, while increasing the computational cost. As shown in Table 3, a choice of up to 4 gear levels seems optimal, further increasing the number of gears does not result in significant improvements.", + "bbox": [ + 500, + 563, + 890, + 654 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 667, + 625, + 684 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we introduced Gear-NeRF, an extension of dynamic NeRFs that leverages semantic information from powerful segmentation models for stratified modeling of dynamic scenes. Our approach learns a 4D (spatiotemporal) semantic embedding and introduces the concept of \"gears\" for differentiated modeling of scene regions based on their motion intensity. With determined gear assignments, Gear-NeRF adaptively adjusts its spatiotemporal sampling resolution to improve the photo-realism of rendered views. At the same time, Gear-NeRF provides the new functionality of free-viewpoint object tracking with user prompts as simple as a click. Our empirical studies underscore the effectiveness of Gear-NeRF, showcasing state-of-the-art performance in both rendering quality and object tracking across multiple challenging datasets.", + "bbox": [ + 496, + 694, + 890, + 921 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19674", + "bbox": [ + 480, + 955, + 517, + 965 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 71, + 171, + 85 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Benjamin Attal, Jia-Bin Huang, Christian Richardt, Michael Zollhoefer, Johannes Kopf, Matthew O'Toole, and Changil Kim. HyperReel: High-fidelity 6-DoF video with ray-conditioned sampling. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 3, 6, 7, 8", + "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 5855–5864, 2021. 2", + "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5470-5479, 2022.", + "[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 19697-19705, 2023. 2", + "[5] Michael Broxton, John Flynn, Ryan Overbeck, Daniel Erickson, Peter Hedman, Matthew Duvall, Jason Dourgarian, Jay Busch, Matt Whalen, and Paul Debevec. Immersive light field video with a layered mesh representation. ACM Transactions on Graphics (TOG), 39(4):86-1, 2020. 6, 7, 8", + "[6] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 130-141, 2023. 1, 2, 3, 7", + "[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers, 2021. 3", + "[8] Jiazhong Cen, Zanwei Zhou, Jiemin Fang, Chen Yang, Wei Shen, Lingxi Xie, Dongsheng Jiang, Xiaopeng Zhang, and Qi Tian. Segment anything in 3d with nerfs, 2023. 3, 7, 8", + "[9] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5799-5809, 2021. 2", + "[10] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 2", + "[11] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision (ECCV), pages 333-350. Springer, 2022. 2", + "[12] Anpei Chen, Zexiang Xu, Xinyue Wei, Siyu Tang, Hao Su, and Andreas Geiger. Factor fields: A unified framework for neural fields and beyond. arXiv preprint arXiv:2302.01226, 2023. 2", + "[13] Jiaben Chen and Huaizu Jiang. Sportsslomo: A new benchmark and baselines for human-centric video frame interpo" + ], + "bbox": [ + 86, + 95, + 470, + 921 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "lation. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3", + "[14] Jiaben Chen, Renrui Zhang, Dongze Lian, Jiaqi Yang, Ziyao Zeng, and Jianbo Shi. iquery: Instruments as queries for audio-visual sound separation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14675–14686, 2023. 3", + "[15] Xiaokang Chen, Kwan-Yee Lin, Chen Qian, Gang Zeng, and Hongsheng Li. 3d sketch-aware semantic scene completion via semi-supervised structure prior. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3", + "[16] Xiaokang Chen, Jiaxiang Tang, Diwen Wan, Jingbo Wang, and Gang Zeng. Interactive segment anything nef with feature imitation. arXiv preprint arXiv:2305.16233, 2023. 3", + "[17] Yue Chen, Xuan Wang, Xingyu Chen, Qi Zhang, Xiaoyu Li, Yu Guo, Jue Wang, and Fei Wang. Uv volumes for real-time rendering of editable free-view human performance. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16621-16631, 2023. 2", + "[18] Zhang Chen, Zhong Li, Liangchen Song, Lele Chen, Jingyi Yu, Junsong Yuan, and Yi Xu. Neurbf: A neural fields representation with adaptive radial basis functions. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 4182-4194, 2023. 2", + "[19] Ho Kei Cheng and Alexander G. Schwing. Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model, 2022. 3", + "[20] Yangming Cheng, Liulei Li, Yuanyou Xu, Xiaodi Li, Zongxin Yang, Wenguan Wang, and Yi Yang. Segment and track anything. arXiv preprint arXiv:2305.06558, 2023. 3", + "[21] Alvaro Collet, Ming Chuang, Pat Sweeney, Don Gillett, Dennis Evseev, David Calabrese, Hugues Hoppe, Adam Kirk, and Steve Sullivan. High-quality streamable free-viewpoint video. ACM Transactions on Graphics (TOG), 34(4):1-13, 2015. 1", + "[22] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 14304-14314. IEEE Computer Society, 2021. 1, 2", + "[23] Zhiwen Fan, Peihao Wang, Xinyu Gong, Yifan Jiang, Dejia Xu, and Zhangyang Wang. Nerf-sos: Any-view self-supervised object segmentation from complex real-world scenes. International Conference on Learning Representations (ICLR), pages arXiv-2209, 2023. 3", + "[24] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. In ACM Transactions on Graphics (SIGGRAPH Asia), pages 1-9, 2022. 2", + "[25] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5501–5510, 2022. 2", + "[26] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In IEEE/CVF Conference on Computer Vision and" + ], + "bbox": [ + 509, + 74, + 890, + 921 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19675", + "bbox": [ + 480, + 955, + 517, + 965 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Pattern Recognition (CVPR), pages 12479-12488, 2023. 1, 2, 3", + "[27] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 5712-5721, 2021. 2", + "[28] Chen Geng, Sida Peng, Zhen Xu, Hujun Bao, and Xiaowei Zhou. Learning neural volumetric representations of dynamic humans in minutes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8759-8770, 2023. 2", + "[29] Rahul Goel, Dhawal Sirikonda, Saurabh Saini, and P.J. Narayanan. Interactive Segmentation of Radiance Fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3", + "[30] Yuan-Chen Guo, Di Kang, Linchao Bao, Yu He, and Song-Hai Zhang. Nerfren: Neural radiance fields with reflections. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18409-18418, 2022. 2", + "[31] Wenbo Hu, Yuling Wang, Lin Ma, Bangbang Yang, Lin Gao, Xiao Liu, and Yuewen Ma. Tri-miprf: Tri-mip representation for efficient anti-aliasing neural radiance fields. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 19774–19783, 2023. 2", + "[32] Qiangui Huang, Weiyue Wang, and Ulrich Neumann. Recurrent slice networks for 3d segmentation on point clouds. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3", + "[33] Mustafa Isik, Martin Rünz, Markos Georgopoulos, Taras Khakhulin, Jonathan Starck, Lourdes Agapito, and Matthias Nießner. Humanrf: High-fidelity neural radiance fields for humans in motion. ACM Transactions on Graphics (SIGGRAPH), 2023. 2", + "[34] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 5885-5894, 2021. 2", + "[35] Lei Ke, Mingqiao Ye, Martin Danelljan, Yifan Liu, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Segment anything in high quality. Advances in Neural Information Processing Systems (NeurIPS), 2023. 3", + "[36] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (TOG), 42(4):1-14, 2023. 2", + "[37] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. In IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 3", + "[38] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2, 3, 4, 5", + "[39] Tobias Kirschstein, Shenhan Qian, Simon Giebenhain, Tim Walter, and Matthias Nießner. Nersemble: Multi-view radiance field reconstruction of human heads. ACM Transactions on Graphics (SIGGRAPH), 2023. 2", + "[40] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nerf for editing via feature field distil" + ], + "bbox": [ + 86, + 74, + 470, + 921 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "lation. Advances in Neural Information Processing Systems (NeurIPS), 35:23311-23330, 2022. 2, 3", + "[41] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and Rene Ranftl. Language-driven semantic segmentation. In International Conference on Learning Representations (ICLR), 2022. 3", + "[42] Hao Li, Bart Adams, Leonidas J Guibas, and Mark Pauly. Robust single-view geometry and motion reconstruction. ACM Transactions on Graphics (TOG), 28(5):1-10, 2009. 1", + "[43] Hao Li, Linjie Luo, Daniel Vlasic, Pieter Peers, Jovan Popovic, Mark Pauly, and Szymon Rusinkiewicz. Temporally coherent completion of dynamic shapes. ACM Transactions on Graphics (TOG), 31(1):1-11, 2012. 1", + "[44] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5521-5531, 2022. 2, 6, 7, 8", + "[45] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021. 1, 2", + "[46] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4273-4284, 2023. 2", + "[47] Haotong Lin, Sida Peng, Zhen Xu, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Efficient neural radiance fields for interactive free-viewpoint video. In ACM Transactions on Graphics (SIGGRAPH Asia), pages 1-9, 2022.", + "[48] Haotong Lin, Sida Peng, Zhen Xu, Tao Xie, Xingyi He, Hujun Bao, and Xiaowei Zhou. Im4d: High-fidelity and real-time novel view synthesis for dynamic scenes. ACM Transactions on Graphics (SIGGRAPH Asia), 2023. 2", + "[49] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 9298–9309, 2023. 2", + "[50] Xinhang Liu, Jiaben Chen, Huai Yu, Yu-Wing Tai, and Chi-Keung Tang. Unsupervised multi-view object segmentation using radiance field propagation. Advances in Neural Information Processing Systems (NeurIPS), 35:17730-17743, 2022. 2, 3", + "[51] Xinhang Liu, Shiu-hong Kao, Jiaben Chen, Yu-Wing Tai, and Chi-Keung Tang. Deceptive-nerf: Enhancing nerf reconstruction using pseudo-observations from diffusion models. arXiv preprint arXiv:2305.15171, 2023. 2", + "[52] Xinhang Liu, Yu-Wing Tai, and Chi-Keung Tang. Clean-nerf: Reformulating nerf to account for view-dependent observations. arXiv preprint arXiv:2303.14707, 2023. 2", + "[53] Yu-Lun Liu, Chen Gao, Andreas Meuleman, Hung-Yu Tseng, Ayush Saraf, Changil Kim, Yung-Yu Chuang, Johannes Kopf, and Jia-Bin Huang. Robust dynamic radiance fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13-23, 2023. 2" + ], + "bbox": [ + 509, + 74, + 890, + 921 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19676", + "bbox": [ + 480, + 955, + 519, + 965 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[54] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 2", + "[55] Andreas Meuleman, Yu-Lun Liu, Chen Gao, Jia-Bin Huang, Changil Kim, Min H Kim, and Johannes Kopf. Progressively optimized local radiance fields for robust view synthesis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16539-16548, 2023. 2", + "[56] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision (ECCV), 2020. 1, 2, 3", + "[57] Ashkan Mirzaei, Tristan Aumentado-Armstrong, Konstantinos G Derpanis, Jonathan Kelly, Marcus A Brubaker, Igor Gilitschenski, and Alex Levinshtein. Spin-nerf: Multiview segmentation and perceptual inpainting with neural radiance fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20669–20679, 2023. 3", + "[58] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (TOG), 41(4):1-15, 2022. 2", + "[59] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5480-5490, 2022. 2", + "[60] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 5865-5874, 2021. 1, 2", + "[61] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Transactions on Graphics (TOG), 2021. 1, 2", + "[62] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in Neural Information Processing Systems (NeurIPS), 32, 2019. 6", + "[63] Sida Peng, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Representing volumetric videos as dynamic mlp maps. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4252-4262, 2023. 2", + "[64] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. International Conference on Learning Representations (ICLR), 2023. 2", + "[65] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields" + ], + "bbox": [ + 86, + 74, + 470, + 920 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for dynamic scenes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10318-10327, 2021. 1, 2", + "[66] Yi-Ling Qiao, Alexander Gao, Yiran Xu, Yue Feng, Jia-Bin Huang, and Ming C Lin. Dynamic mesh-aware radiance fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 385-396, 2023. 2", + "[67] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of Machine Learning Research (PMLR), pages 8748-8763, 2021. 3", + "[68] Konstantinos Rematas, Andrew Liu, Pratul P Srinivasan, Jonathan T Barron, Andrea Tagliasacchi, Thomas Funkhouser, and Vittorio Ferrari. Urban radiance fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12932-12942, 2022. 2", + "[69] Zhongzheng Ren, Aseem Agarwala, Bryan Russell, Alexander G. Schwing, and Oliver Wang. Neural volumetric object selection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3", + "[70] Neus Sabater, Guillaume Boisson, Benoit Vandame, Paul Kerbiriou, Frederic Babon, Matthieu Hog, Remy Gendrot, Tristan Langlois, Olivier Bureller, Arno Schubert, et al. Dataset and pipeline for multi-view light-field video. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 30-40, 2017. 6, 7, 8", + "[71] Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boning Liu, Hongwen Zhang, and Yebin Liu. Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16632-16642, 2023. 2, 3", + "[72] Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulò, Norman Müller, Matthias Nießner, Angela Dai, and Peter Kontschieder. Panoptic lifting for 3d scene understanding with neural fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9043-9052, 2023. 3", + "[73] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerf-player: A streamable dynamic scene representation with decomposed neural radiance fields. IEEE Transactions on Visualization and Computer Graphics (TVCG), 29(5): 2732-2742, 2023. 2", + "[74] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5459-5469, 2022. 2", + "[75] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8248–8258, 2022. 2", + "[76] Jiaxiang Tang, Xiaokang Chen, Jingbo Wang, and Gang Zeng. Point scene understanding via disentangled instance" + ], + "bbox": [ + 509, + 74, + 890, + 920 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "19677", + "bbox": [ + 480, + 955, + 517, + 965 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "mesh reconstruction. European Conference on Computer Vision (ECCV), 2022. 3", + "[77] Jonathan Taylor, Jamie Shotton, Toby Sharp, and Andrew Fitzgibbon. The vitruvian manifold: Inferring dense correspondences for one-shot human pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 103-110. IEEE, 2012. 1", + "[78] Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srinivasan, Edgar Tretschk, Wang Yifan, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lombardi, et al. Advances in neural rendering. In Computer Graphics Forum, pages 703-735. Wiley Online Library, 2022. 2", + "[79] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 12959-12970, 2021. 2", + "[80] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481-5490. IEEE, 2022. 2", + "[81] Feng Wang, Sinan Tan, Xinghang Li, Zeyue Tian, and Huaping Liu. Mixed neural voxels for fast multi-view video synthesis. IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2, 7", + "[82] Liao Wang, Jiakai Zhang, Xinhang Liu, Fuqiang Zhao, Yanshun Zhang, Yingliang Zhang, Minye Wu, Jingyi Yu, and Lan Xu. Fourier plenoctrees for dynamic radiance field rendering in real-time. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13524-13534, 2022. 2", + "[83] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing (TIP), 13(4):600-612, 2004. 6", + "[84] Ziyan Wang, Giljoo Nam, Tuur Stuyck, Stephen Lombardi, Chen Cao, Jason Saragih, Michael Zollhöfer, Jessica Hodgins, and Christoph Lassner. Neuwigs: A neural dynamic model for volumetric hair capture and animation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8641-8651, 2023. 2", + "[85] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528, 2023. 2", + "[86] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P Srinivasan, Dor Verbin, Jonathan T Barron, Ben Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. arXiv preprint arXiv:2312.02981, 2023. 2", + "[87] Jamie Wynn and Daniyar Turmukhambetov. Diffusionerf: Regularizing neural radiance fields with denoising diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4180-4189, 2023. 2", + "[88] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint" + ], + "bbox": [ + 86, + 74, + 470, + 921 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "video. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9421-9431, 2021. 2", + "[89] Zhen Xu, Sida Peng, Haotong Lin, Guangzhao He, Jiaming Sun, Yujun Shen, Hujun Bao, and Xiaowei Zhou. 4k4d: Real-time 4d view synthesis at 4k resolution. arXiv preprint arXiv:2310.11448, 2023. 2", + "[90] Zhiwen Yan, Chen Li, and Gim Hee Lee. Nerf-ds: Neural radiance fields for dynamic specular objects. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8285-8295, 2023. 2", + "[91] Bo Yang, Jianan Wang, Ronald Clark, Qingyong Hu, Sen Wang, Andrew Markham, and Niki Trigoni. Learning object bounding boxes for 3d instance segmentation on point clouds, 2019. 3", + "[92] Jinyu Yang, Mingqi Gao, Zhe Li, Shang Gao, Fangjing Wang, and Feng Zheng. Track anything: Segment anything meets videos, 2023. 3", + "[93] Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Improving few-shot neural rendering with free frequency regularization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8254-8263, 2023. 2", + "[94] Zongxin Yang and Yi Yang. Decoupling features in hierarchical propagation for video object segmentation, 2022. 3", + "[95] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101, 2023. 2", + "[96] Zeyu Yang, Hongye Yang, Zijie Pan, Xiatian Zhu, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. arXiv preprint arXiv:2310.10642, 2023. 2", + "[97] Zhengming Yu, Wei Cheng, Xian Liu, Wayne Wu, and Kwan-Yee Lin. Monohuman: Animatable human neural field from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16943-16953, 2023. 2", + "[98] Wentao Yuan, Zhaoyang Lv, Tanner Schmidt, and Steven Lovegrove. Star: Self-supervised tracking and reconstruction of rigid objects in motion with neural rendering. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13144-13152, 2021. 2", + "[99] Jiakai Zhang, Xinhang Liu, Xinyi Ye, Fuqiang Zhao, Yanshun Zhang, Minye Wu, Yingliang Zhang, Lan Xu, and Jingyi Yu. Editable free-viewpoint video using a layered neural representation. ACM Transactions on Graphics (TOG), 40(4):1-18, 2021. 1, 2, 7", + "[100] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 6", + "[101] Renrui Zhang, Zhengkai Jiang, Ziyu Guo, Shilin Yan, Junting Pan, Hao Dong, Peng Gao, and Hongsheng Li. Personalize segment anything model with one shot. arXiv preprint arXiv:2305.03048, 2023. 3", + "[102] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J Davison. In-place scene labelling and understanding with implicit scene representation. In IEEE/CVF" + ], + "bbox": [ + 501, + 74, + 890, + 920 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "19678", + "bbox": [ + 480, + 955, + 519, + 965 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "International Conference on Computer Vision (ICCV), pages 15838-15847, 2021. 2, 3", + "bbox": [ + 116, + 74, + 468, + 102 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "19679", + "bbox": [ + 480, + 955, + 517, + 965 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/2024/Gear-NeRF_ Free-Viewpoint Rendering and Tracking with Motion-aware Spatio-Temporal Sampling/8ba31972-a909-4d60-aef6-d492fd479dbb_model.json b/2024/Gear-NeRF_ Free-Viewpoint Rendering and Tracking with Motion-aware Spatio-Temporal Sampling/8ba31972-a909-4d60-aef6-d492fd479dbb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0ed2f4bb88a7694994e3dd6fbe1223559f08b11d --- /dev/null +++ b/2024/Gear-NeRF_ Free-Viewpoint Rendering and Tracking with Motion-aware Spatio-Temporal Sampling/8ba31972-a909-4d60-aef6-d492fd479dbb_model.json @@ -0,0 +1,3405 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.112, + 0.758, + 0.159 + ], + "angle": 0, + "content": "Gear-NeRF: Free-Viewpoint Rendering and Tracking with Motion-aware Spatio-Temporal Sampling" + }, + { + "type": "text", + "bbox": [ + 0.139, + 0.185, + 0.832, + 0.275 + ], + "angle": 0, + "content": "Xinhang Liu\\(^{1\\dagger*}\\) Yu-Wing Tai\\(^{2}\\) Chi-Keung Tang\\(^{1\\dagger}\\) Pedro Miraldo\\(^{3}\\) Suhas Lohit\\(^{3}\\) Moitreya Chatterjee\\(^{3}\\) \n\\(^{1}\\)HKUST \\(^{2}\\)Dartmouth College \\(^{3}\\)Mitsubishi Electric Research Laboratories (MERL) xliufe@connect.ust.hk, yu-wing.tai@dartmouth.edu, cktang@cse.ust.hk, miraldo@merl.com, slohit@merl.com, metro.smiles@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.309, + 0.314, + 0.325 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.342, + 0.474, + 0.721 + ], + "angle": 0, + "content": "Extensions of Neural Radiance Fields (NeRFs) to model dynamic scenes have enabled their near photo-realistic, free-viewpoint rendering. Although these methods have shown some potential in creating immersive experiences, two drawbacks limit their ubiquity: (i) a significant reduction in reconstruction quality when the computing budget is limited, and (ii) a lack of semantic understanding of the underlying scenes. To address these issues, we introduce Gear-NeRF, which leverages semantic information from powerful image segmentation models. Our approach presents a principled way for learning a spatio-temporal (4D) semantic embedding, based on which we introduce the concept of gears to allow for stratified modeling of dynamic regions of the scene based on the extent of their motion. Such differentiation allows us to adjust the spatiotemporal sampling resolution for each region in proportion to its motion scale, achieving more photo-realistic dynamic novel view synthesis. At the same time, almost for free, our approach enables free-viewpoint tracking of objects of interest - a functionality not yet achieved by existing NeRF-based methods. Empirical studies validate the effectiveness of our method, where we achieve state-of-the-art rendering and tracking performance on multiple challenging datasets. The project page is available at: https://merl.com/research/highlights/gear-nerf." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.752, + 0.21, + 0.768 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.779, + 0.47, + 0.87 + ], + "angle": 0, + "content": "Reconstructing 3D scenes has a broad range of applications, including Virtual Reality/Augmented Reality (VR/AR), 3D animation, game production, and film creation which allow users to observe scenes from any desired viewpoint. While it is crucial to reconstruct static scenes, towards which significant progress has been made, it is even more crucial to" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.31, + 0.892, + 0.509 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.522, + 0.894, + 0.592 + ], + "angle": 0, + "content": "Figure 1. (a) Our method takes RGB videos captured from a camera array as input. (b) Trained Gear-NeRF achieves photo-realistic real-time free-viewpoint rendering of a dynamic scene. (c) With users giving a single click at any time and from any viewpoint, our method can perform free-viewpoint tracking of the target object." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.603, + 0.893, + 0.647 + ], + "angle": 0, + "content": "reconstruct dynamic scenes, as the world around us is characterized by a constant state of flux, with many objects in it - in a state of motion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.649, + 0.895, + 0.86 + ], + "angle": 0, + "content": "Recent advances in novel view synthesis, such as Neural Radiance Fields (NeRFs) [56] have inspired numerous studies to extend them to dynamic 3D scenes. Existing approaches either employ a deformation field to map neural fields from a given time to a canonical space [22, 45, 60, 61, 65, 99], or directly model dynamic scenes as a 4D space-time grid [1, 6, 26]. Though these methods offer improved rendering quality by utilizing more accessible inputs compared to previous solutions [21, 42, 43, 77], they still struggle to ensure rendering quality in low-resource settings, requiring carefully engineered efforts. Further, most dynamic radiance field approaches adopt a naive spatiotemporal sampling strategy, without discerning the different scales of motion across different regions in the scene." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.862, + 0.894, + 0.922 + ], + "angle": 0, + "content": "We propose to fix this issue by leveraging a semantic understanding of dynamic scenes. Intuitively, a reconstruction system aware of the distinction between static and dynamic regions in a scene can perform more focused" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.884, + 0.383, + 0.896 + ], + "angle": 0, + "content": "*Work mainly done when XL was an intern at MERL." + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.897, + 0.469, + 0.922 + ], + "angle": 0, + "content": "\\(\\dagger\\) XL and CT are supported in part by the Research Grant Council of the Hong Kong SAR grant no. 16201420." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.884, + 0.469, + 0.922 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.956, + 0.519, + 0.967 + ], + "angle": 0, + "content": "19667" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.078, + 0.895, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.311, + 0.897, + 0.382 + ], + "angle": 0, + "content": "Figure 2. Pipeline of Gear-NeRF: Gear-NeRF takes multi-view videos as input. After optimizing the serial 4D feature volumes (Section 4.1), it maps space-time coordinates to a 4D semantic embedding (Section 4.2), in addition to the volume density and view-dependent radiance color. Regions with larger motion are automatically assigned higher gear levels (Section 4.3) and as a result, receive higher-resolution spatio-temporal sampling (Section 4.4). Furthermore, Gear-NeRF is capable of performing free-viewpoint tracking of a target object with prompts as simple as a user click (Section 4.5)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.39, + 0.473, + 0.692 + ], + "angle": 0, + "content": "sampling in the dynamic regions, which inherently require more resources per unit volume than static regions, due to their time-evolving nature. Accordingly, dynamic regions can be further stratified according to their scale of motion. To this end, this paper presents Gear-NeRF, a framework that leverages semantic embedding from powerful image segmentation models [38] for stratified modeling of 4D scenes. Gear-NeRF optimizes for a 4D semantic embedding, based on which we introduce the concept of gear to smartly determine the appropriate region-specific resolution of spatio-temporal sampling in the NeRF. Regions with larger motion scales are assigned higher gears, through our gear determination scheme and we accordingly perform higher-resolution spatio-temporal sampling. Empirical studies reveal that this motion-aware sampling strategy improves the quality of synthesized images, over competing approaches. As a by-product of our semantically embedded representation, we achieve free-viewpoint object tracking, given user prompts. Figure 1 presents an overview of the capabilities of our method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.694, + 0.47, + 0.83 + ], + "angle": 0, + "content": "Gear-NeRF makes two primary advancements: (i) enhanced dynamic novel view synthesis by resorting to smarter spatio-temporal sampling, and (ii) the ability for free-viewpoint tracking of objects of interest. The latter is a capability not yet realized by existing NeRF methods for dynamic scenes. We perform extensive experiments on multiple datasets to validate the generalizability and robustness of our method, which shows state-of-the-art performances for both tasks across all datasets." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.849, + 0.22, + 0.866 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.876, + 0.469, + 0.923 + ], + "angle": 0, + "content": "Neural Radiance Fields: NeRF [56] is a recent breakthrough among novel view synthesis methods that uses multilayer perceptrons (MLPs) to parameterize the appearance" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.39, + 0.895, + 0.646 + ], + "angle": 0, + "content": "and density for each point in 3D space, given any viewing direction of the scene. Researchers have extended NeRF along various dimensions [78], including improving rendering quality [2-4, 12, 18, 31], handling challenging conditions such as large scenes [55, 68, 75], view-dependent appearances [30, 52, 80], and sparse inputs [34, 51, 59, 86, 87, 93]. NeRF-like neural representations have also found applications in semantic segmentation [40, 50, 102] and 3D content generation [9, 10, 49, 64]. Recent work has shown that replacing the deep MLPs with a feature voxel grid can significantly improve training and inference speed [11, 25, 58, 74]. On the other hand, a more recent approach to further improve visual quality, rendering time, and performance entails representing the scene with 3D Gaussians [36]. Our approach, while drawing upon many of these approaches, deals with dynamic scenes which is beyond the scope of these methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.649, + 0.896, + 0.922 + ], + "angle": 0, + "content": "Neural Representations for Dynamic Scenes: NeRF-like representations have recently been extended to model dynamic scenes in high fidelity [17, 28, 33, 39, 46-48, 53, 63, 66, 84, 89, 90, 97]. One straightforward approach to do this is to directly condition the radiance field on time [27, 44, 45, 88]. Alternatively, several methods model a deformation field to map coordinates from different time stamps to a common canonical space [22, 24, 60, 61, 65, 79, 98, 99]. Some recent approaches [6, 26, 71, 81] represent the scene using a 4D space-time grid, which is decomposed into sets of planar representations for training efficiency. Other techniques for improving rendering fidelity and frame rate include Fourier PlenOctrees [82], ray-conditioned sample prediction networks [1], 4D space decomposition (static/dynamic/newly appeared regions) [73], and explicit voxel grids [24]. 3D Gaussians have also been adapted to model dynamic scenes [54, 85, 95, 96]. While these approaches paved the initial path for rendering dy" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.956, + 0.519, + 0.967 + ], + "angle": 0, + "content": "19668" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.074, + 0.47, + 0.33 + ], + "angle": 0, + "content": "namic scenes, semantically aware modeling of the scene is absent, a caveat that our proposed method seeks to address. Segment Anything Model (SAM): SAM [38] is a powerful promptable image segmentation model, which showcases remarkable zero-shot generalization abilities and can produce semantically consistent masks, given a single foreground point on the image. HQ-SAM [35] is an improvement on SAM that enhances the quality of masks, especially on objects with intricate boundaries and structures. Recent works [20, 92] have extended SAM to perform interactive video object segmentation. These methods utilize SAM for mask initialization or correction and then employ state-of-the-art mask trackers [19, 94] for mask tracking and prediction [14]. Recent methods have also leveraged SAM for tracking multiple reference objects in a video [13, 101]. This work uses SAM to profile the scene into semantic regions, which are then grouped based on motion scales." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.332, + 0.471, + 0.604 + ], + "angle": 0, + "content": "3D Semantic Understanding: Existing methods for 3D visual understanding [15, 32, 76, 91] mainly focus on closed set segmentation of point clouds or voxels. NeRF's ability to integrate information across multiple views has led to its applications in 3D semantic segmentation [102], object segmentation [23, 50, 57], panoptic segmentation [72], and interactive segmentation [29, 69]. Kobayashi et al. [40] explored the effectiveness of embedding pixel-aligned features [7, 41] into NeRFs for 3D manipulations. LERF [37] fuses CLIP embeddings [67] and NeRFs to enable language-based localization in 3D NeRF scenes. Recent work has enabled click/text-based 3D segmentation by learning a 3D SAM embedding [16] or inverse rendering of SAM-generated masks [8]. We, on the other hand, seek to utilize the synergy of dynamic NeRFs and SAM segments to derive a semantic understanding of a dynamic 3D scene for tracking objects of interest in novel views – a first of its kind effort." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.622, + 0.216, + 0.637 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.648, + 0.469, + 0.77 + ], + "angle": 0, + "content": "Neural Radiance Fields (NeRFs): Vanilla NeRFs [56] employ a multi-layer perceptron (MLP) with sinusoidal positional encoding to map a 3D-spatial coordinate \\(\\mathbf{x} = (x, y, z)\\) and a viewing direction \\(\\mathbf{d} = (\\theta, \\phi)\\) to a volume density \\(\\sigma \\in [0,1]\\) and an emitted RGB, \\(\\mathbf{c} \\in \\mathbb{R}^3\\). Rendering each image pixel involves casting a ray \\(\\mathbf{r}(t) = \\mathbf{o} + t\\mathbf{d}\\) from the camera center \\(\\mathbf{o}\\) through the pixel along direction \\(\\mathbf{d}\\). The predicted color for the corresponding pixel is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.783, + 0.47, + 0.824 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {C}} (\\mathbf {r}) = \\sum_ {i = 1} ^ {N} T _ {i} \\alpha_ {i} \\mathbf {c} _ {i}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.839, + 0.469, + 0.922 + ], + "angle": 0, + "content": "where \\( T_{i} = \\exp \\left(-\\sum_{j=1}^{i-1} \\sigma_{j} \\delta_{j}\\right) \\), \\( \\alpha_{i} = 1 - \\exp (-\\sigma_{i} \\delta_{i}) \\), and \\( \\delta_{j} = t_{j+1} - t_{j} \\). A vanilla NeRF is trained by minimizing the mean squared error between the input images and the predicted images, obtained by rendering the scene from the viewpoints from which the input images have been" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.074, + 0.772, + 0.089 + ], + "angle": 0, + "content": "captured, with the training loss given by:" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.101, + 0.892, + 0.133 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {p h o}} = \\sum_ {\\mathbf {r} \\in \\mathcal {R}} \\| \\hat {\\mathbf {C}} (\\mathbf {r}) - \\mathbf {C} (\\mathbf {r}) \\| _ {2} ^ {2}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.145, + 0.892, + 0.326 + ], + "angle": 0, + "content": "where \\(\\mathcal{R}\\) is the set of all rays projected from the input image. Planar-Factorized 4D Volumes: A recent emerging trend of handling dynamics using radiance field representations is to directly adapt them to be conditioned on a frame index \\(t\\) (denoting time) in addition to \\(\\mathbf{x}\\) and \\(\\mathbf{d}\\). This can be accomplished by learning a mapping from \\((\\mathbf{x},\\mathbf{d},t)\\) to \\((\\sigma ,\\mathbf{c})\\) using planar-factorized 4D volumes [1, 6, 26, 71]. These methods attempt to learn a 4D feature vector for every \\((\\mathbf{x},t)\\), by projecting it to a set of 2D-planes. Embeddings of these projections on these planes can then be integrated to obtain the embedding for the 4D point. This can be mathematically represented as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.34, + 0.892, + 0.395 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {f} (\\mathbf {x}, t) = \\mathbf {B} _ {1} \\left(\\mathbf {h} _ {1} (x, y) \\odot \\mathbf {k} _ {1} (z, t)\\right) \\\\ + \\mathbf {B} _ {2} (\\mathbf {h} _ {2} (x, z) \\odot \\mathbf {k} _ {2} (y, t)) \\\\ + \\mathbf {B} _ {3} \\left(\\mathbf {h} _ {3} (y, z) \\odot \\mathbf {k} _ {3} (x, t)\\right). \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.406, + 0.892, + 0.528 + ], + "angle": 0, + "content": "where \\(\\mathbf{h}_i(\\cdot ,\\cdot)\\) and \\(\\mathbf{k}_i(\\cdot ,\\cdot)\\) are functions (evaluated by bilinear interpolation on regularly spaced 2D feature grids) embedding coordinate tuples to features of dimension \\(M\\) \" \\(\\odot\\) denotes an element-wise product, and \\(\\mathbf{B}_i(\\cdot)\\) denotes a linear transform which maps the products to feature vectors. Subsequently, a tiny MLP can map the feature vector \\(\\mathbf{f}(\\cdot ,\\cdot)\\) to the volume density, \\(\\sigma\\) , and the view-dependent emitted color, c, given the viewing direction d." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.543, + 0.673, + 0.56 + ], + "angle": 0, + "content": "4. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.892, + 0.643 + ], + "angle": 0, + "content": "Given a set of \\( W \\) input videos, \\( \\mathcal{V} = \\{V_1, V_2, \\dots, V_W\\} \\) of a dynamic scene, with calibrated camera poses, our approach represents the scene using a series of 4D feature volumes (Section 4.1) along with 4D semantic embeddings (Section 4.2)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.892, + 0.826 + ], + "angle": 0, + "content": "Analogous to multiple gears in motor vehicles for optimizing engine performance, Gear-NeRF stratifies this semantically embedded scene representation into \\( N_{\\mathrm{gear}} \\) levels, based on the motion scales. Each of these levels is called a gear. Through our training scheme, regions with larger motion are assigned higher gear levels (Section 4.3) and as a result, are more densely sampled (Section 4.4) for improved dynamic novel view synthesis. Our 4D semantic embedding also enables a new functionality, almost for free - free-viewpoint tracking of target objects, given simple user prompts like clicks (Section 4.5). Figure 2 shows the overall pipeline of Gear-NeRF." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.837, + 0.742, + 0.852 + ], + "angle": 0, + "content": "4.1. Serial 4D Feature Volumes" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.861, + 0.892, + 0.922 + ], + "angle": 0, + "content": "Instead of using a unified 4D volume to represent a dynamic scene [1, 6, 26, 71], our representation consists of a series of feature volumes, each corresponding to a gear level, \\(\\mathcal{G}\\). Specifically, for any space-time coordinate \\((\\mathbf{x}, t)\\), its feature" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.956, + 0.519, + 0.967 + ], + "angle": 0, + "content": "19669" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.074, + 0.411, + 0.089 + ], + "angle": 0, + "content": "vector corresponding to \\(\\mathcal{G}\\) is computed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.097, + 0.469, + 0.159 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {f} ^ {\\mathcal {G}} (\\mathbf {x}, t) = \\mathbf {B} _ {1} \\left(\\mathbf {h} _ {1} (x, y) \\odot \\mathbf {k} _ {1} ^ {\\mathcal {G}} (z, t)\\right) \\\\ + \\mathbf {B} _ {2} (\\mathbf {h} _ {2} (x, z) \\odot \\mathbf {k} _ {2} ^ {\\mathcal {G}} (y, t)) \\\\ + \\mathbf {B} _ {3} \\left(\\mathbf {h} _ {3} (y, z) \\odot \\mathbf {k} _ {3} ^ {\\mathcal {G}} (x, t)\\right). \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.169, + 0.469, + 0.259 + ], + "angle": 0, + "content": "The vector-valued functions \\(\\mathbf{h}_j(\\cdot, \\cdot)\\) and linear transforms \\(\\mathbf{B}_j(\\cdot)\\) are shared by all gears, while each gear has its own spatio-temporal embedding \\(\\mathbf{k}_j^G(\\cdot, \\cdot)\\), in \\(M\\)-dimensional space. Therefore, each gear describes regions of a certain scale of motion while the purely spatial features can be shared among all gears." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.26, + 0.469, + 0.306 + ], + "angle": 0, + "content": "We obtain the gear level at any spatio-temporal coordinate also from a planar-factorized 4D feature volume. Specifically, the gear level at \\((\\mathbf{x},t)\\) is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.315, + 0.469, + 0.375 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} g (\\mathbf {x}, t) = \\mathbf {1} ^ {\\top} \\left(\\mathbf {h} _ {1} ^ {\\prime} (x, y) \\odot \\mathbf {k} _ {1} ^ {\\prime} (z, t)\\right) \\\\ + \\mathbf {1} ^ {\\top} \\left(\\mathbf {h} _ {2} ^ {\\prime} (x, z) \\odot \\mathbf {k} _ {2} ^ {\\prime} (y, t)\\right) \\\\ + \\mathbf {1} ^ {\\top} \\left(\\mathbf {h} _ {3} ^ {\\prime} (y, z) \\odot \\mathbf {k} _ {3} ^ {\\prime} (x, t)\\right), \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.385, + 0.47, + 0.447 + ], + "angle": 0, + "content": "where \\(\\mathbf{1}\\) is a vector of ones, \\(\\mathbf{h}_i^{\\prime}(\\cdot ,\\cdot)\\) and \\(\\mathbf{k}_i^\\prime (\\cdot ,\\cdot)\\) are \\(M\\) dimensional embedding functions. This however defines a continuous feature volume. To map it to the gear level integers, we apply the following projection operation:" + }, + { + "type": "equation", + "bbox": [ + 0.123, + 0.457, + 0.469, + 0.512 + ], + "angle": 0, + "content": "\\[\np (\\mathbf {x}, t) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} g (\\mathbf {x}, t) < 1, \\\\ N _ {\\text {g e a r}}, & \\text {i f} g (\\mathbf {x}, t) \\geq N _ {\\text {g e a r}}, \\\\ \\lceil g (\\mathbf {x}, t) \\rceil , & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.522, + 0.469, + 0.553 + ], + "angle": 0, + "content": "Based on this gear level volume, we define a 4D mask for a region at gear level \\(\\mathcal{G}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.563, + 0.469, + 0.603 + ], + "angle": 0, + "content": "\\[\nm _ {\\mathcal {G}} (\\mathbf {x}, t) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} p (\\mathbf {x}, t) = \\mathcal {G}, \\\\ 0, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.613, + 0.396, + 0.629 + ], + "angle": 0, + "content": "The final feature vector at \\((\\mathbf{x},t)\\) is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.64, + 0.469, + 0.682 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} (\\mathbf {x}, t) = \\sum_ {\\mathcal {G} = 1} ^ {N _ {\\text {g e a r}}} m _ {\\mathcal {G}} (\\mathbf {x}, t) \\mathbf {f} ^ {\\mathcal {G}} (\\mathbf {x}, t). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.693, + 0.469, + 0.754 + ], + "angle": 0, + "content": "Subsequently, a tiny MLP, \\( F_{\\theta} \\), maps these feature vectors \\( \\mathbf{f}(\\cdot ,\\cdot) \\) as well as the viewing direction \\( \\mathbf{d} \\) to the volume density \\( \\sigma \\) and radiance color \\( \\mathbf{c} \\). This allows us to obtain a photometric rendering of the scene." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.763, + 0.305, + 0.779 + ], + "angle": 0, + "content": "4.2. 4D Semantic Embedding" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.785, + 0.469, + 0.922 + ], + "angle": 0, + "content": "Gear-NeRF leverages the strong object priors of the SAM [38] model to acquire a semantic understanding of the scene, for improved photometric rendering (Section 4.3 and Section 4.4) as well as free-viewpoint object tracking (Section 4.5). Toward this end, we utilize the SAM encoder to obtain 2D feature maps from the frames of each video. We then optimize a 4D SAM embedding field by supervising it with these 2D feature maps. In particular, the MLP above, \\( F_{\\theta} \\), is configured to output a 4D semantic embedding s for" + }, + { + "type": "image_caption", + "bbox": [ + 0.637, + 0.061, + 0.757, + 0.071 + ], + "angle": 0, + "content": "\\(1^{\\mathrm{st}}\\) Gear Assignment Update" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.072, + 0.633, + 0.142 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.072, + 0.756, + 0.142 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.072, + 0.882, + 0.142 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.635, + 0.144, + 0.757, + 0.154 + ], + "angle": 0, + "content": "2nd Gear Assignment Update" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.156, + 0.633, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.156, + 0.756, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.156, + 0.881, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.636, + 0.228, + 0.757, + 0.238 + ], + "angle": 0, + "content": "\\(3^{\\mathrm{rd}}\\) Gear Assignment Update" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.239, + 0.633, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.239, + 0.756, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.239, + 0.881, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.636, + 0.311, + 0.757, + 0.322 + ], + "angle": 0, + "content": "\\(4^{\\mathrm{th}}\\) Gear Assignment Update" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.323, + 0.633, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.395, + 0.61, + 0.405 + ], + "angle": 0, + "content": "Rendered Image" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.323, + 0.756, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.651, + 0.395, + 0.742, + 0.405 + ], + "angle": 0, + "content": "Rendering Loss Map" + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.323, + 0.881, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.791, + 0.394, + 0.851, + 0.404 + ], + "angle": 0, + "content": "Upshift Mask" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.418, + 0.892, + 0.571 + ], + "angle": 0, + "content": "Figure 3. Illustration of Gear Assignment Update: For each gear assignment update, we calculate the rendering loss map between the rendered RGB-SAM map and the ground truth and identify the centers of the patches with the maximum and minimum losses, marked in red and green (second column). These points are then fed into the SAM decoder as positive and negative prompts to generate an upshift mask representing the areas that need to be shifted to a higher gear (last column). After the first gear assignment update, we see that the next candidate region for upshift is situated where the horse is located, and so on. Upshift mask colors imply the gear levels after the update (blue-2, green-3, red-4)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.579, + 0.892, + 0.655 + ], + "angle": 0, + "content": "a given space-time coordinate in addition to the density, \\(\\sigma\\), and color, c. To render 2D semantic feature maps in a given view, we compute the semantic feature of a pixel in the feature map by tracing a ray through it and perform volume rendering analogous to Equation 1, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.629, + 0.665, + 0.891, + 0.706 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {S}} (\\mathbf {r}) = \\sum_ {i = 1} ^ {N} T _ {i} \\alpha_ {i} \\mathbf {s} _ {i}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.715, + 0.892, + 0.761 + ], + "angle": 0, + "content": "This SAM embedding is supervised by minimizing the mean squared error between the prediction and the ground truth features \\((\\mathbf{S}(\\mathbf{r}))\\) from the SAM encoder, as shown:" + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.771, + 0.891, + 0.803 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {S A M}} = \\sum_ {\\mathbf {r} \\in \\mathcal {R}} \\| \\hat {\\mathbf {S}} (\\mathbf {r}) - \\mathbf {S} (\\mathbf {r}) \\| _ {2} ^ {2}. \\tag {10}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.822, + 0.843, + 0.839 + ], + "angle": 0, + "content": "4.3. Training Scheme with Gear Assignment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.846, + 0.892, + 0.891 + ], + "angle": 0, + "content": "With gear initialization \\( g(\\mathbf{x}, t) = 1, \\forall \\mathbf{x}, t \\), the (semantically embedded) radiance field optimization and gear assignment updating take place in an alternating fashion." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.892, + 0.892, + 0.922 + ], + "angle": 0, + "content": "Gear Assignment Update: As illustrated in Figure 3, when updating the gear assignment after a period of radiance field" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.956, + 0.519, + 0.967 + ], + "angle": 0, + "content": "19670" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.074, + 0.47, + 0.134 + ], + "angle": 0, + "content": "optimization, we find the regions rendered most poorly from the rendering loss maps and increment their gears for denser spatio-temporal sampling. The following steps lay out the process for updating gear assignments to regions:" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.135, + 0.468, + 0.224 + ], + "angle": 0, + "content": "- We sample a number of viewpoints and time steps and render 2D-images/SAM features for it. For every rendered RGB-SAM map, we compute a rendering loss map. Each pixel of the rendering loss map is computed as: \\(\\mathcal{L}(\\mathbf{r})\\coloneqq \\mathcal{L}_{\\mathrm{pho}}(\\mathbf{r}) + \\lambda \\mathcal{L}_{\\mathrm{SAM}}(\\mathbf{r})\\). See Figure 3 for example loss maps." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.225, + 0.468, + 0.3 + ], + "angle": 0, + "content": "- Next, we patchify each rendering loss map to find patches with the top-\\(k\\) largest/smallest average loss. The center coordinate of these patches serve as positive/negative prompts for the next step. See Figure 3 for example positive (red) / negative (green) prompts." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.3, + 0.468, + 0.39 + ], + "angle": 0, + "content": "- We then feed the ground truth RGB images together with positive and negative prompts into the SAM decoder [38] to estimate an upshift mask. These masks tend to cover regions that have motions and are not satisfactorily rendered with the current sampling resolution. We have multiple upshift masks at different viewpoints and time steps." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.391, + 0.468, + 0.435 + ], + "angle": 0, + "content": "- For every pixel of an upshift mask, we trace a ray and sample points along it and update the gear assignment by pushing \\( g(\\mathbf{x}, t) \\) towards incremented values." + }, + { + "type": "list", + "bbox": [ + 0.077, + 0.135, + 0.468, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.437, + 0.468, + 0.697 + ], + "angle": 0, + "content": "In particular, in the last step, for each pixel within an upshift mask, a corresponding ray is traced that connects it with the camera center o, along direction d. Next, a set of points are sampled along this ray. The collection of sampled points, that lie on the rays emanating from within the masked region constitutes the set \\(S_{\\mathrm{upshift}} = \\{(x_i^{\\mathrm{upshift}},t_i^{\\mathrm{upshift}})\\}_{i = 1}^{N_{\\mathrm{upshift}}}\\), where \\(N_{\\mathrm{upshift}}\\) represents the total count of points sampled from rays pertaining to the masked region. We then follow a similar procedure to sample a set of points pertaining to the unmasked region. We denote this set as: \\(\\mathcal{S}_{\\mathrm{stay}} = \\{(x_i^{\\mathrm{stay}},t_i^{\\mathrm{stay}})\\}_{i = 1}^{N_{\\mathrm{stay}}}\\), with \\(N_{\\mathrm{stay}}\\) indicating the total number of points sampled from rays that pertain to the unmasked area. Next, for each sample point in each of \\(S_{\\mathrm{upshift}}\\) and \\(S_{\\mathrm{stay}}\\), we query its current gear level \\(p(\\mathbf{x},t)\\). In order to assign new gear levels, we need to update the gear assignment function \\(g(\\cdot ,\\cdot)\\), which proceeds with the objective function:" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.708, + 0.468, + 0.8 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {u p s h i f t}} = \\frac {1}{N _ {\\text {u p s h i f t}}} \\sum_ {(\\mathbf {x}, t) \\in \\mathcal {S} _ {\\text {u p s h i f t}}} \\| g (\\mathbf {x}, t; \\boldsymbol {\\Theta}) - (p (\\mathbf {x}, t) + 1) \\| _ {2} ^ {2} \\\\ + \\frac {\\lambda_ {\\text {s t a y}}}{N _ {\\text {s t a y}}} \\sum_ {(\\mathbf {x}, t) \\in \\mathcal {S} _ {\\text {s t a y}}} \\| g (\\mathbf {x}, t; \\boldsymbol {\\Theta}) - p (\\mathbf {x}, t) \\| _ {2} ^ {2}, \\tag {11} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.802, + 0.468, + 0.922 + ], + "angle": 0, + "content": "where \\(\\Theta\\) denotes the set of estimizable parameters for \\(g(\\cdot ,\\cdot)\\). The minimization of the first term encourages sample points within the masked region \\(S_{\\mathrm{upshift}}\\) to have a gear level equal to their current gear incremented by one, resulting in an upshift of gear. Conversely, minimizing the second term encourages the points in \\(S_{\\mathrm{stay}}\\) to maintain their gear values, thereby encouraging the remaining regions to keep their current gear levels and avoiding unwanted up" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.074, + 0.89, + 0.103 + ], + "angle": 0, + "content": "shifts. We update \\(\\Theta\\) via a single step of gradient descent as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.112, + 0.89, + 0.128 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Theta} \\leftarrow \\boldsymbol {\\Theta} - \\alpha \\nabla_ {\\boldsymbol {\\Theta}} \\mathcal {L} _ {\\text {u p s h i f t}}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.143, + 0.892, + 0.249 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is the learning rate. Since, \\(g(\\cdot ,\\cdot)\\) is essentially derived from the embedding functions, \\(\\mathbf{h}_i^{\\prime}(\\cdot ,\\cdot)\\) and \\(\\mathbf{k}_i^\\prime (\\cdot ,\\cdot)\\) for \\(i\\in \\{1,2,3\\}\\), the aforementioned optimization step amounts to updating these embedding functions. Once updated, we proceed with a fresh round of gear assignment to increase the spatio-temporal sampling resolution for the regions that end up at a higher gear level than before." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.252, + 0.892, + 0.403 + ], + "angle": 0, + "content": "Radiance Field Optimization: With the updated gear assignment, we increase the resolution of spatio-temporal sampling for the gear-shifted regions (Section 4.4) and then resume optimizing the radiance field. We alternate between the two processes: radiance field optimization (each time for \\( L \\) epochs), and gear assignment updates until the average variance of each rendering loss map is below a predetermined threshold. After this, we optimize the radiance field for an additional \\( L' \\) epochs without further gear assignment updates." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.422, + 0.855, + 0.438 + ], + "angle": 0, + "content": "4.4. Motion-aware Spatio-Temporal Sampling" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.448, + 0.892, + 0.569 + ], + "angle": 0, + "content": "In this subsection, we explain our motion-aware spatiotemporal sampling strategy based on assigned gears, permitting differential processing of regions at different gear levels. By temporal sampling, we imply the choice of temporal resolution for planar-factorized 4D feature volumes, and by spatial sampling, we mean the strategy used to choose sampling points along each ray for volume rendering." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.571, + 0.892, + 0.738 + ], + "angle": 0, + "content": "Motion-aware Temporal Sampling: To handle the increasing intensity of object motion, as reflected by their growing gear levels, we increment the temporal resolution for voxel grids. Specifically, \\(\\mathbf{k}_j^{\\mathcal{G}}\\) in Equation 4 has increasing resolution along the time axis, thereby empowering the 4D feature volumes to better model the dynamics along the temporal axis. This ensures fast-moving objects can be more faithfully modeled without unsightly blurring. The temporal resolution for each gear's feature volume is determined by linear interpolation between 1 (for \\(\\mathcal{G} = 1\\)) and the total number of frames (for \\(\\mathcal{G} = N_{\\mathrm{gear}}\\))." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.74, + 0.892, + 0.922 + ], + "angle": 0, + "content": "Motion-aware Spatial Sampling: While denser sampling of points can improve reconstruction accuracy, increasing the number of sampling points throughout the scene can lead to prohibitive computational costs. Therefore, we propose a 3D point-splitting strategy as illustrated in Figure 4. We begin by sampling a relatively small number, \\( n \\), of samples along each ray, assuming it is at the lowest gear level. If a sampled point belongs to a region with a higher gear, as determined by \\( p(\\mathbf{x}, t) \\), we then sample more densely in that region. For every sampled point in that region, we split it into \\( 2^{p(\\mathbf{x}, t) - 1} \\) points, equally spaced within the corresponding ray segment (at that gear level)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.956, + 0.518, + 0.967 + ], + "angle": 0, + "content": "19671" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.071, + 0.434, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.2, + 0.47, + 0.272 + ], + "angle": 0, + "content": "Figure 4. Motion-aware Spatial Sampling: We split each sampled point into \\(2^{p(\\mathbf{x},t)}\\) points, equally spaced within the corresponding ray segment. The top row shows the vanilla uniformly sampled points, while the bottom one shows the densely sampled points after splitting using our strategy." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.28, + 0.456, + 0.297 + ], + "angle": 0, + "content": "4.5. Free-Viewpoint Tracking with User Prompts" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.304, + 0.468, + 0.41 + ], + "angle": 0, + "content": "Our 4D SAM embedding enables another useful functionality, almost for free - free-viewpoint object tracking, where the user only needs to provide as few as one click to extract the target object based on the 4D embedding. Next, we describe how, given a user-supplied point click at any arbitrary viewpoint and time step, we obtain the corresponding object mask at a novel viewpoint and time step." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.411, + 0.468, + 0.59 + ], + "angle": 0, + "content": "Masks for Novel Viewpoints: The first step for this task entails finding the 3D correspondence of the user click. We trace a ray through the selected pixel, and by utilizing the volume density, we determine the depth at which the ray intersects with the first object surface it encounters. This yields the 3D coordinates of the point of intersection. Subsequently, the 3D coordinates of this intersection can be easily mapped into a 2D coordinate within any novel viewpoint image, using the camera pose of the new viewpoint. Alongside the rendered SAM feature map of the novel view, we feed this coordinate into the SAM decoder to generate the object mask for the novel view." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.593, + 0.468, + 0.744 + ], + "angle": 0, + "content": "Masks for Novel Time Steps: For this task, we propagate an object mask to its neighboring time step. Specifically, with an object mask for a specific frame \\( t \\), we calculate the bounding box of this mask and use this bounding box as a prompt to SAM for neighboring frames \\( t' = t + 1 \\) or \\( t' = t - 1 \\). By inputting this prompt along with the rendered SAM feature map at \\( t' \\) into the SAM decoder, we can obtain the object mask for \\( t' \\). Combining the above two processes, we can start from a single click and get the object mask in any viewpoint and time step." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.759, + 0.21, + 0.777 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.785, + 0.468, + 0.922 + ], + "angle": 0, + "content": "We assess the performance of our proposed Gear-NeRF for dynamic novel view synthesis and free-viewpoint object tracking across a range of challenging datasets, comparing it with state-of-the-art methods. Through ablation studies, we provide empirical evidence of the effectiveness of its fundamental components. We kindly refer the reader to our supplementary material for additional experimental details and results, including videos for free-viewpoint rendering and tracking." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.073, + 0.691, + 0.089 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.099, + 0.892, + 0.308 + ], + "angle": 0, + "content": "Implementation Details: We implement our method using PyTorch [62] and conduct experiments on an NVIDIA RTX 4090 GPU with 24 GB RAM. We divide each input video into chunks of 100 frames. For every chunk, we train a model for approximately 2.5 hours. Our 4D feature volumes yield embeddings with a dimension of \\( M = 32 \\). We set the gear number \\( N_{\\mathrm{gear}} \\) to 4. We find patches with top- \\( k = 3 \\) largest/smallest average loss for gear assignment updates to obtain prompts. The rendering loss map is computed with \\( \\lambda = 0.01 \\). For the optimization of radiance fields, \\( L = 3 \\) and \\( L' = 10 \\). In our motion-aware spatial sampling, each ray initially has \\( n = 64 \\) sampling points. We use an initial learning rate of 0.02 for all the parameters. For Equation 12, we use \\( \\alpha = 0.02 \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.312, + 0.892, + 0.644 + ], + "angle": 0, + "content": "Datasets: (i) The Technicolor light field dataset [70] includes diverse indoor environment videos captured by a \\(4 \\times 4\\) camera rig. We evaluate on 4 sequences (Train, Theater, Painter, Birthday) at the original \\(2048 \\times 1088\\) resolution, holding out the same view as prior work [1] (the second row and second column) for evaluation. (ii) The Neural 3D Video dataset [44] includes indoor multi-view video sequences captured by 20 cameras at a resolution of \\(2704 \\times 2028\\) pixels. We experiment on 6 sequences (Cut Roasted Beef, Flame Steak, Coffee Martini, Cook Spinach, Flame Salmon, Sear Steak), downsampling by a factor of 2 and holding out the central view (akin to prior work [1]) for evaluation. (iii) The Google Immersive dataset[5] contains light field videos of indoor and outdoor scenes captured by a time-synchronized 46-fisheye camera rig, with a resolution of \\(2560 \\times 1920\\) pixels. We experiment with 9 sequences from it (Flames, Truck, Horse, Car, Welder, Exhibit, Face Paint 1, Face Paint 2, Cave). We downsample the video resolution by a factor of 2 and hold out the central view (like prior work [1]) for evaluation. For our experiments, we adopted the same resolution and held-out view selection as prior work [1]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.647, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Evaluation Metrics: For the task of novel view synthesis of dynamic scenes, we evaluate using the following standard metrics: (i) Peak Signal-to-Noise Ratio (PSNR), (ii) Structural Similarity Index Measure (SSIM) [83] and (iii) Learned Perceptual Image Patch Similarity (LPIPS) [100], by comparing the reconstructed frames against ground truth images. These metrics are computed on the held-out view and averaged across all frames. For the free-viewpoint object tracking task, we designate a specific viewpoint and time step for the user to give prompting clicks. Subsequently, we assess the quality of the predicted object masks at novel viewpoints. The quality of the object mask is quantified in terms of the Mean Intersection over Union (mIoU) and accuracy (Acc.), which are calculated against the ground truth mask, manually annotated utilizing Adobe Photoshop. Additionally, we present the same metrics computed for novel time steps, denoted as t-mIoU and t-Acc." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.906, + 0.892, + 0.921 + ], + "angle": 0, + "content": "Baselines: We run a comprehensive comparison of our" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.955, + 0.519, + 0.967 + ], + "angle": 0, + "content": "19672" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.119, + 0.072, + 0.3, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.073, + 0.48, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.483, + 0.073, + 0.663, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.666, + 0.073, + 0.845, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.181, + 0.3, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.182, + 0.481, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.182, + 0.662, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.666, + 0.182, + 0.845, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.291, + 0.3, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.171, + 0.397, + 0.246, + 0.409 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.292, + 0.481, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.361, + 0.397, + 0.421, + 0.41 + ], + "angle": 0, + "content": "HyperReel" + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.292, + 0.663, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.397, + 0.619, + 0.409 + ], + "angle": 0, + "content": "Neural 3D Video" + }, + { + "type": "image", + "bbox": [ + 0.666, + 0.292, + 0.845, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.741, + 0.398, + 0.77, + 0.408 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.421, + 0.892, + 0.451 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparisons for novel view synthesis on the Technicolor dataset [70]: We qualitatively compare our approach against HyperReel [1] and Neural 3D Video [44]. Our approach better recovers fine details like patterns on the toys or stripes on the shirt." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.458, + 0.473, + 0.595 + ], + "angle": 0, + "content": "method against a range of recent NeRF-based baseline methods: (i) ST-NeRF [99], (ii) HexPlane [6], (iii) HyperReel [1], and (iv) MixVoxels [81]. As the first method to enable promptable free-viewpoint object tracking under the NeRF setting, there is no established direct baseline for this specific task. However, SA3D [8], a recent method for segmenting static scenes, is treated as a baseline for predicting masks at novel viewpoints corresponding to the prompted time step." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.608, + 0.172, + 0.623 + ], + "angle": 0, + "content": "5.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.633, + 0.47, + 0.8 + ], + "angle": 0, + "content": "Dynamic Novel View Synthesis: As shown in Figure 5, Gear-NeRF produces high-quality novel view synthesis of dynamic scenes, accurately modeling real-world scenes with intricate motions and fine details. For example, patterns on the toys or stripes on the shirt, are faithfully rendered, resulting in more photo-realistic images compared to all of the baselines. Table 1, presents quantitative comparisons of our method against the baselines. While Gear-NeRF has longer training (Tr. Time) / inference times (FPS) compared to some baselines, it almost always achieves the best performance in terms of rendering quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.801, + 0.471, + 0.922 + ], + "angle": 0, + "content": "Free-Viewpoint Tracking: In Figure 6, we present the object masks obtained by our method based on the user prompts provided at a specified viewpoint and time step. Specifically, the first row displays masked objects at the prompted viewpoint and time step. The second row shows novel view masks at the prompted time step. The third row shows novel view masks at novel time steps. We see that masks obtained from Gear-NeRF show precise boundaries," + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.457, + 0.892, + 0.512 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons for dynamic novel view synthesis: Our method outperforms all baselines across all datasets on all metrics. We report means over all scenes for each dataset. Best and second best results are highlighted." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.513, + 0.892, + 0.644 + ], + "angle": 0, + "content": "
DatasetMethodPSNR (↑)SSIM (↑)LPIPS (↓)Tr. Time(↓)FPS(↑)
Technicolor [70]ST-NeRF [99]30.860.8830.101344 min0.7
HyperReel [1]31.040.8870.09297 min7.7
MixVoxels [81]28.990.8420.10314 min18.9
Ours32.210.9190.058148 min7.9
Google Immersive [5]HexPlane [6]27.670.8080.188527 min0.9
HyperReel [1]28.320.8620.145117 min7.4
MixVoxels [81]27.140.8350.20919 min16.8
Ours28.740.8760.122189 min7.0
Neural 3D Video [44]ST-NeRF [99]31.030.8900.081367 min0.7
HyperReel [1]31.120.9280.065129 min6.1
MixVoxels [81]30.690.9440.13915 min16.3
Ours31.800.9360.058204 min6.8
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.646, + 0.892, + 0.811 + ], + "angle": 0, + "content": "compared to SA3D. Table 2 presents quantitative assessment of the quality of the masks generated of our method versus SA3D. Our method exceeds \\(90\\%\\) across all metrics and datasets, demonstrating the effectiveness of our approach. Our gains over SA3D can be attributed to the fact that SA3D, as a static scene segmentation method, does not utilize information across all time frames, whereas our approach does. SA3D is incapable of predicting masks for novel time steps, and as such, the corresponding entries are marked as not applicable, a shortcoming which our method does not have." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.822, + 0.628, + 0.837 + ], + "angle": 0, + "content": "5.3. Evaluations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.846, + 0.892, + 0.89 + ], + "angle": 0, + "content": "To verify the effectiveness of the design choices of GearNeRF, we perform extensive ablation studies on the Truck scene of the Google Immersive dataset [5]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.892, + 0.892, + 0.922 + ], + "angle": 0, + "content": "Motion-aware Temporal Sampling: An intuitive strategy for adjusting the temporal sampling is to directly modify" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.956, + 0.519, + 0.967 + ], + "angle": 0, + "content": "19673" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.129, + 0.1, + 0.215, + 0.113 + ], + "angle": 0, + "content": "★ User Prompt" + }, + { + "type": "image_caption", + "bbox": [ + 0.139, + 0.166, + 0.22, + 0.188 + ], + "angle": 0, + "content": "Rendering from Novel View" + }, + { + "type": "image_caption", + "bbox": [ + 0.124, + 0.226, + 0.236, + 0.248 + ], + "angle": 0, + "content": "Rendering from Novel View at Novel Time" + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.075, + 0.393, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.206, + 0.393, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.275, + 0.273, + 0.358, + 0.285 + ], + "angle": 0, + "content": "Rendered Image" + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.08, + 0.458, + 0.136 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.398, + 0.144, + 0.468, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.398, + 0.205, + 0.481, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.422, + 0.274, + 0.455, + 0.284 + ], + "angle": 0, + "content": "SA3D" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.077, + 0.546, + 0.136 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.143, + 0.551, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.212, + 0.551, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.274, + 0.537, + 0.284 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.57, + 0.075, + 0.679, + 0.14 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.57, + 0.141, + 0.679, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.57, + 0.205, + 0.679, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.583, + 0.274, + 0.666, + 0.285 + ], + "angle": 0, + "content": "Rendered Image" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.078, + 0.747, + 0.138 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.683, + 0.151, + 0.764, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.205, + 0.766, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.706, + 0.274, + 0.74, + 0.284 + ], + "angle": 0, + "content": "SA3D" + }, + { + "type": "image", + "bbox": [ + 0.789, + 0.077, + 0.832, + 0.125 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.786, + 0.149, + 0.831, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.781, + 0.216, + 0.828, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.797, + 0.275, + 0.823, + 0.283 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.299, + 0.892, + 0.327 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparisons of free-viewpoint object tracking on Technicolor [70] and Neural 3D Video [44] datasets: Our method can obtain desirable object masks, with clear edges, from prompting points provided by users at desired time steps and viewpoints." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.334, + 0.469, + 0.388 + ], + "angle": 0, + "content": "Table 2. Quantitative comparisons for free-viewpoint tracking: t-mIoU and t-Acc are metrics used for evaluating novel view masks at novel time steps, not applicable to SA3D. Reported metrics are averages over all scenes for each dataset." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.39, + 0.47, + 0.486 + ], + "angle": 0, + "content": "
DatasetMethodmIoU (↑)Acc. (↑)t-mIoU (↑)t-Acc. (↑)
Technicolor [70]SA3D [8]96.497.1N/AN/A
Ours97.497.692.193.3
Google Immersive [5]SA3D [8]94.194.8N/AN/A
Ours94.395.091.592.8
Neural 3D Video [44]SA3D [8]93.194.0N/AN/A
Ours93.494.390.692.3
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.493, + 0.469, + 0.674 + ], + "angle": 0, + "content": "the temporal resolution of the 4D feature volume and make it uniform across all regions of the scene. In Table 3, we demonstrate the performance of this naive temporal sampling strategy using medium (25) or dense (100) temporal resolutions on 100-frame input videos. The results show that increasing the temporal resolution universally does not necessarily yield better performance. In contrast, our proposed method, which utilizes motion-aware temporal sampling, achieves the best results. We surmise that this may be attributed to the model distributing the volume's capacity sparsely across a large number of time steps, which is not the case for our method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.677, + 0.469, + 0.829 + ], + "angle": 0, + "content": "Motion-aware Spatial Sampling: To validate the effectiveness of our motion-aware spatial sampling strategy, we compare it against other sampling strategies. The first variant involves uniform sampling along the ray, and the second uses a sample prediction network (SPN) like Attal et al. [1] for sampling. These variants sample 128 points on each ray. Results in Table 3 show that our full model employing motion-aware spatial sampling outperforms these variants, perhaps by having a better sense of where to sample more from, derived from its semantic-aware embedding." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.831, + 0.469, + 0.922 + ], + "angle": 0, + "content": "SAM embedding: To verify that introducing SAM embedding can improve rendering quality (rather than just enabling segmentation or tracking), we tested a variant of Gear-NeRF without SAM embedding (it instead thresholds loss maps of rendered RGB frames to obtain upshift masks). As shown in Table 3, the absence of SAM em" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.334, + 0.892, + 0.375 + ], + "angle": 0, + "content": "Table 3. Ablation Study: Ablations on our spatio-temporal sampling strategy and the number of gears (Truck/Google Immersive). Best and second best results are highlighted." + }, + { + "type": "table", + "bbox": [ + 0.5, + 0.376, + 0.892, + 0.53 + ], + "angle": 0, + "content": "
MethodPSNR (↑)SSIM (↑)LPIPS (↓)
Naive temporal sampling (medium)26.930.7780.144
Naive temporal sampling (dense)26.850.7600.161
Naive spatial sampling24.800.7340.222
SPN spatial sampling26.540.7870.162
Ours (w/o embedding)27.100.8310.177
Ours (Ngear = 2)26.930.7850.166
Ours (Ngear = 3)27.140.8750.145
Ours (Ngear = 4)27.490.8920.136
Ours (Ngear = 5)27.460.9010.131
" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.533, + 0.892, + 0.563 + ], + "angle": 0, + "content": "bedding for guiding gear assignment reduces the model's rendering quality." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.564, + 0.892, + 0.655 + ], + "angle": 0, + "content": "Number of Gears: We ablate on the numbers of gears. Increasing the number of gears allows for more fine-grained motion-aware spatio-temporal sampling, while increasing the computational cost. As shown in Table 3, a choice of up to 4 gear levels seems optimal, further increasing the number of gears does not result in significant improvements." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.669, + 0.627, + 0.685 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.695, + 0.892, + 0.922 + ], + "angle": 0, + "content": "In this work, we introduced Gear-NeRF, an extension of dynamic NeRFs that leverages semantic information from powerful segmentation models for stratified modeling of dynamic scenes. Our approach learns a 4D (spatiotemporal) semantic embedding and introduces the concept of \"gears\" for differentiated modeling of scene regions based on their motion intensity. With determined gear assignments, Gear-NeRF adaptively adjusts its spatiotemporal sampling resolution to improve the photo-realism of rendered views. At the same time, Gear-NeRF provides the new functionality of free-viewpoint object tracking with user prompts as simple as a click. Our empirical studies underscore the effectiveness of Gear-NeRF, showcasing state-of-the-art performance in both rendering quality and object tracking across multiple challenging datasets." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.956, + 0.519, + 0.967 + ], + "angle": 0, + "content": "19674" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.072, + 0.173, + 0.087 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.097, + 0.47, + 0.18 + ], + "angle": 0, + "content": "[1] Benjamin Attal, Jia-Bin Huang, Christian Richardt, Michael Zollhoefer, Johannes Kopf, Matthew O'Toole, and Changil Kim. HyperReel: High-fidelity 6-DoF video with ray-conditioned sampling. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.182, + 0.472, + 0.25 + ], + "angle": 0, + "content": "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 5855–5864, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.251, + 0.471, + 0.32 + ], + "angle": 0, + "content": "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5470-5479, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.471, + 0.389 + ], + "angle": 0, + "content": "[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 19697-19705, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.471, + 0.46 + ], + "angle": 0, + "content": "[5] Michael Broxton, John Flynn, Ryan Overbeck, Daniel Erickson, Peter Hedman, Matthew Duvall, Jason Dourgarian, Jay Busch, Matt Whalen, and Paul Debevec. Immersive light field video with a layered mesh representation. ACM Transactions on Graphics (TOG), 39(4):86-1, 2020. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.461, + 0.47, + 0.514 + ], + "angle": 0, + "content": "[6] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 130-141, 2023. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.516, + 0.47, + 0.572 + ], + "angle": 0, + "content": "[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.573, + 0.47, + 0.614 + ], + "angle": 0, + "content": "[8] Jiazhong Cen, Zanwei Zhou, Jiemin Fang, Chen Yang, Wei Shen, Lingxi Xie, Dongsheng Jiang, Xiaopeng Zhang, and Qi Tian. Segment anything in 3d with nerfs, 2023. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.615, + 0.47, + 0.683 + ], + "angle": 0, + "content": "[9] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5799-5809, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.685, + 0.47, + 0.78 + ], + "angle": 0, + "content": "[10] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.782, + 0.47, + 0.837 + ], + "angle": 0, + "content": "[11] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision (ECCV), pages 333-350. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.838, + 0.47, + 0.892 + ], + "angle": 0, + "content": "[12] Anpei Chen, Zexiang Xu, Xinyue Wei, Siyu Tang, Hao Su, and Andreas Geiger. Factor fields: A unified framework for neural fields and beyond. arXiv preprint arXiv:2302.01226, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.894, + 0.47, + 0.922 + ], + "angle": 0, + "content": "[13] Jiaben Chen and Huaizu Jiang. Sportsslomo: A new benchmark and baselines for human-centric video frame interpo" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.097, + 0.472, + 0.922 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.075, + 0.892, + 0.102 + ], + "angle": 0, + "content": "lation. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.104, + 0.892, + 0.173 + ], + "angle": 0, + "content": "[14] Jiaben Chen, Renrui Zhang, Dongze Lian, Jiaqi Yang, Ziyao Zeng, and Jianbo Shi. iquery: Instruments as queries for audio-visual sound separation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14675–14686, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.175, + 0.892, + 0.242 + ], + "angle": 0, + "content": "[15] Xiaokang Chen, Kwan-Yee Lin, Chen Qian, Gang Zeng, and Hongsheng Li. 3d sketch-aware semantic scene completion via semi-supervised structure prior. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.245, + 0.892, + 0.285 + ], + "angle": 0, + "content": "[16] Xiaokang Chen, Jiaxiang Tang, Diwen Wan, Jingbo Wang, and Gang Zeng. Interactive segment anything nef with feature imitation. arXiv preprint arXiv:2305.16233, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.287, + 0.892, + 0.356 + ], + "angle": 0, + "content": "[17] Yue Chen, Xuan Wang, Xingyu Chen, Qi Zhang, Xiaoyu Li, Yu Guo, Jue Wang, and Fei Wang. Uv volumes for real-time rendering of editable free-view human performance. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16621-16631, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.358, + 0.892, + 0.427 + ], + "angle": 0, + "content": "[18] Zhang Chen, Zhong Li, Liangchen Song, Lele Chen, Jingyi Yu, Junsong Yuan, and Yi Xu. Neurbf: A neural fields representation with adaptive radial basis functions. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 4182-4194, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.429, + 0.892, + 0.469 + ], + "angle": 0, + "content": "[19] Ho Kei Cheng and Alexander G. Schwing. Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.472, + 0.892, + 0.512 + ], + "angle": 0, + "content": "[20] Yangming Cheng, Liulei Li, Yuanyou Xu, Xiaodi Li, Zongxin Yang, Wenguan Wang, and Yi Yang. Segment and track anything. arXiv preprint arXiv:2305.06558, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.514, + 0.892, + 0.581 + ], + "angle": 0, + "content": "[21] Alvaro Collet, Ming Chuang, Pat Sweeney, Don Gillett, Dennis Evseev, David Calabrese, Hugues Hoppe, Adam Kirk, and Steve Sullivan. High-quality streamable free-viewpoint video. ACM Transactions on Graphics (TOG), 34(4):1-13, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.584, + 0.892, + 0.653 + ], + "angle": 0, + "content": "[22] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 14304-14314. IEEE Computer Society, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.655, + 0.892, + 0.723 + ], + "angle": 0, + "content": "[23] Zhiwen Fan, Peihao Wang, Xinyu Gong, Yifan Jiang, Dejia Xu, and Zhangyang Wang. Nerf-sos: Any-view self-supervised object segmentation from complex real-world scenes. International Conference on Learning Representations (ICLR), pages arXiv-2209, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.726, + 0.892, + 0.794 + ], + "angle": 0, + "content": "[24] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. In ACM Transactions on Graphics (SIGGRAPH Asia), pages 1-9, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.796, + 0.892, + 0.864 + ], + "angle": 0, + "content": "[25] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5501–5510, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.867, + 0.892, + 0.922 + ], + "angle": 0, + "content": "[26] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In IEEE/CVF Conference on Computer Vision and" + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.075, + 0.892, + 0.922 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.956, + 0.519, + 0.967 + ], + "angle": 0, + "content": "19675" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.119, + 0.075, + 0.468, + 0.101 + ], + "angle": 0, + "content": "Pattern Recognition (CVPR), pages 12479-12488, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.103, + 0.47, + 0.159 + ], + "angle": 0, + "content": "[27] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 5712-5721, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.16, + 0.471, + 0.228 + ], + "angle": 0, + "content": "[28] Chen Geng, Sida Peng, Zhen Xu, Hujun Bao, and Xiaowei Zhou. Learning neural volumetric representations of dynamic humans in minutes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8759-8770, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.23, + 0.47, + 0.286 + ], + "angle": 0, + "content": "[29] Rahul Goel, Dhawal Sirikonda, Saurabh Saini, and P.J. Narayanan. Interactive Segmentation of Radiance Fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.288, + 0.471, + 0.342 + ], + "angle": 0, + "content": "[30] Yuan-Chen Guo, Di Kang, Linchao Bao, Yu He, and Song-Hai Zhang. Nerfren: Neural radiance fields with reflections. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18409-18418, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.343, + 0.471, + 0.413 + ], + "angle": 0, + "content": "[31] Wenbo Hu, Yuling Wang, Lin Ma, Bangbang Yang, Lin Gao, Xiao Liu, and Yuewen Ma. Tri-miprf: Tri-mip representation for efficient anti-aliasing neural radiance fields. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 19774–19783, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.414, + 0.471, + 0.469 + ], + "angle": 0, + "content": "[32] Qiangui Huang, Weiyue Wang, and Ulrich Neumann. Recurrent slice networks for 3d segmentation on point clouds. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.47, + 0.471, + 0.539 + ], + "angle": 0, + "content": "[33] Mustafa Isik, Martin Rünz, Markos Georgopoulos, Taras Khakhulin, Jonathan Starck, Lourdes Agapito, and Matthias Nießner. Humanrf: High-fidelity neural radiance fields for humans in motion. ACM Transactions on Graphics (SIGGRAPH), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.541, + 0.471, + 0.596 + ], + "angle": 0, + "content": "[34] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 5885-5894, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.598, + 0.471, + 0.653 + ], + "angle": 0, + "content": "[35] Lei Ke, Mingqiao Ye, Martin Danelljan, Yifan Liu, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Segment anything in high quality. Advances in Neural Information Processing Systems (NeurIPS), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.654, + 0.471, + 0.709 + ], + "angle": 0, + "content": "[36] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (TOG), 42(4):1-14, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.71, + 0.471, + 0.765 + ], + "angle": 0, + "content": "[37] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. In IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.767, + 0.471, + 0.836 + ], + "angle": 0, + "content": "[38] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.837, + 0.471, + 0.893 + ], + "angle": 0, + "content": "[39] Tobias Kirschstein, Shenhan Qian, Simon Giebenhain, Tim Walter, and Matthias Nießner. Nersemble: Multi-view radiance field reconstruction of human heads. ACM Transactions on Graphics (SIGGRAPH), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.894, + 0.471, + 0.922 + ], + "angle": 0, + "content": "[40] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nerf for editing via feature field distil" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.075, + 0.471, + 0.922 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.075, + 0.892, + 0.102 + ], + "angle": 0, + "content": "lation. Advances in Neural Information Processing Systems (NeurIPS), 35:23311-23330, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.103, + 0.892, + 0.159 + ], + "angle": 0, + "content": "[41] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and Rene Ranftl. Language-driven semantic segmentation. In International Conference on Learning Representations (ICLR), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.16, + 0.892, + 0.215 + ], + "angle": 0, + "content": "[42] Hao Li, Bart Adams, Leonidas J Guibas, and Mark Pauly. Robust single-view geometry and motion reconstruction. ACM Transactions on Graphics (TOG), 28(5):1-10, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.217, + 0.892, + 0.272 + ], + "angle": 0, + "content": "[43] Hao Li, Linjie Luo, Daniel Vlasic, Pieter Peers, Jovan Popovic, Mark Pauly, and Szymon Rusinkiewicz. Temporally coherent completion of dynamic shapes. ACM Transactions on Graphics (TOG), 31(1):1-11, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.274, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[44] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5521-5531, 2022. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.358, + 0.892, + 0.427 + ], + "angle": 0, + "content": "[45] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.428, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[46] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4273-4284, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.498, + 0.892, + 0.555 + ], + "angle": 0, + "content": "[47] Haotong Lin, Sida Peng, Zhen Xu, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Efficient neural radiance fields for interactive free-viewpoint video. In ACM Transactions on Graphics (SIGGRAPH Asia), pages 1-9, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.556, + 0.892, + 0.611 + ], + "angle": 0, + "content": "[48] Haotong Lin, Sida Peng, Zhen Xu, Tao Xie, Xingyi He, Hujun Bao, and Xiaowei Zhou. Im4d: High-fidelity and real-time novel view synthesis for dynamic scenes. ACM Transactions on Graphics (SIGGRAPH Asia), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.612, + 0.892, + 0.68 + ], + "angle": 0, + "content": "[49] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 9298–9309, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.682, + 0.892, + 0.751 + ], + "angle": 0, + "content": "[50] Xinhang Liu, Jiaben Chen, Huai Yu, Yu-Wing Tai, and Chi-Keung Tang. Unsupervised multi-view object segmentation using radiance field propagation. Advances in Neural Information Processing Systems (NeurIPS), 35:17730-17743, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.753, + 0.892, + 0.808 + ], + "angle": 0, + "content": "[51] Xinhang Liu, Shiu-hong Kao, Jiaben Chen, Yu-Wing Tai, and Chi-Keung Tang. Deceptive-nerf: Enhancing nerf reconstruction using pseudo-observations from diffusion models. arXiv preprint arXiv:2305.15171, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.81, + 0.892, + 0.851 + ], + "angle": 0, + "content": "[52] Xinhang Liu, Yu-Wing Tai, and Chi-Keung Tang. Clean-nerf: Reformulating nerf to account for view-dependent observations. arXiv preprint arXiv:2303.14707, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.852, + 0.892, + 0.922 + ], + "angle": 0, + "content": "[53] Yu-Lun Liu, Chen Gao, Andreas Meuleman, Hung-Yu Tseng, Ayush Saraf, Changil Kim, Yung-Yu Chuang, Johannes Kopf, and Jia-Bin Huang. Robust dynamic radiance fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13-23, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.075, + 0.892, + 0.922 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.956, + 0.52, + 0.967 + ], + "angle": 0, + "content": "19676" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.075, + 0.47, + 0.129 + ], + "angle": 0, + "content": "[54] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.131, + 0.472, + 0.213 + ], + "angle": 0, + "content": "[55] Andreas Meuleman, Yu-Lun Liu, Chen Gao, Jia-Bin Huang, Changil Kim, Min H Kim, and Johannes Kopf. Progressively optimized local radiance fields for robust view synthesis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16539-16548, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.216, + 0.47, + 0.285 + ], + "angle": 0, + "content": "[56] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision (ECCV), 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.287, + 0.47, + 0.383 + ], + "angle": 0, + "content": "[57] Ashkan Mirzaei, Tristan Aumentado-Armstrong, Konstantinos G Derpanis, Jonathan Kelly, Marcus A Brubaker, Igor Gilitschenski, and Alex Levinshtein. Spin-nerf: Multiview segmentation and perceptual inpainting with neural radiance fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20669–20679, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.385, + 0.47, + 0.44 + ], + "angle": 0, + "content": "[58] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (TOG), 41(4):1-15, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.442, + 0.47, + 0.524 + ], + "angle": 0, + "content": "[59] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5480-5490, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.526, + 0.47, + 0.595 + ], + "angle": 0, + "content": "[60] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 5865-5874, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.597, + 0.47, + 0.679 + ], + "angle": 0, + "content": "[61] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Transactions on Graphics (TOG), 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.681, + 0.47, + 0.764 + ], + "angle": 0, + "content": "[62] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in Neural Information Processing Systems (NeurIPS), 32, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.766, + 0.47, + 0.834 + ], + "angle": 0, + "content": "[63] Sida Peng, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Representing volumetric videos as dynamic mlp maps. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4252-4262, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.837, + 0.47, + 0.891 + ], + "angle": 0, + "content": "[64] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. International Conference on Learning Representations (ICLR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.894, + 0.47, + 0.921 + ], + "angle": 0, + "content": "[65] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.075, + 0.472, + 0.921 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.075, + 0.892, + 0.115 + ], + "angle": 0, + "content": "for dynamic scenes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10318-10327, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.117, + 0.892, + 0.171 + ], + "angle": 0, + "content": "[66] Yi-Ling Qiao, Alexander Gao, Yiran Xu, Yue Feng, Jia-Bin Huang, and Ming C Lin. Dynamic mesh-aware radiance fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 385-396, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.173, + 0.892, + 0.268 + ], + "angle": 0, + "content": "[67] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of Machine Learning Research (PMLR), pages 8748-8763, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.269, + 0.892, + 0.338 + ], + "angle": 0, + "content": "[68] Konstantinos Rematas, Andrew Liu, Pratul P Srinivasan, Jonathan T Barron, Andrea Tagliasacchi, Thomas Funkhouser, and Vittorio Ferrari. Urban radiance fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12932-12942, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.339, + 0.892, + 0.394 + ], + "angle": 0, + "content": "[69] Zhongzheng Ren, Aseem Agarwala, Bryan Russell, Alexander G. Schwing, and Oliver Wang. Neural volumetric object selection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.395, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[70] Neus Sabater, Guillaume Boisson, Benoit Vandame, Paul Kerbiriou, Frederic Babon, Matthieu Hog, Remy Gendrot, Tristan Langlois, Olivier Bureller, Arno Schubert, et al. Dataset and pipeline for multi-view light-field video. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 30-40, 2017. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.491, + 0.892, + 0.572 + ], + "angle": 0, + "content": "[71] Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boning Liu, Hongwen Zhang, and Yebin Liu. Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16632-16642, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.575, + 0.892, + 0.655 + ], + "angle": 0, + "content": "[72] Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulò, Norman Müller, Matthias Nießner, Angela Dai, and Peter Kontschieder. Panoptic lifting for 3d scene understanding with neural fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9043-9052, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.657, + 0.892, + 0.739 + ], + "angle": 0, + "content": "[73] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerf-player: A streamable dynamic scene representation with decomposed neural radiance fields. IEEE Transactions on Visualization and Computer Graphics (TVCG), 29(5): 2732-2742, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.741, + 0.892, + 0.808 + ], + "angle": 0, + "content": "[74] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5459-5469, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.81, + 0.892, + 0.893 + ], + "angle": 0, + "content": "[75] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8248–8258, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.894, + 0.892, + 0.921 + ], + "angle": 0, + "content": "[76] Jiaxiang Tang, Xiaokang Chen, Jingbo Wang, and Gang Zeng. Point scene understanding via disentangled instance" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.075, + 0.892, + 0.921 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.956, + 0.519, + 0.967 + ], + "angle": 0, + "content": "19677" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.119, + 0.075, + 0.47, + 0.101 + ], + "angle": 0, + "content": "mesh reconstruction. European Conference on Computer Vision (ECCV), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.104, + 0.47, + 0.173 + ], + "angle": 0, + "content": "[77] Jonathan Taylor, Jamie Shotton, Toby Sharp, and Andrew Fitzgibbon. The vitruvian manifold: Inferring dense correspondences for one-shot human pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 103-110. IEEE, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.175, + 0.471, + 0.257 + ], + "angle": 0, + "content": "[78] Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srinivasan, Edgar Tretschk, Wang Yifan, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lombardi, et al. Advances in neural rendering. In Computer Graphics Forum, pages 703-735. Wiley Online Library, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.259, + 0.471, + 0.342 + ], + "angle": 0, + "content": "[79] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 12959-12970, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.343, + 0.47, + 0.424 + ], + "angle": 0, + "content": "[80] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481-5490. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.428, + 0.471, + 0.482 + ], + "angle": 0, + "content": "[81] Feng Wang, Sinan Tan, Xinghang Li, Zeyue Tian, and Huaping Liu. Mixed neural voxels for fast multi-view video synthesis. IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.485, + 0.47, + 0.567 + ], + "angle": 0, + "content": "[82] Liao Wang, Jiakai Zhang, Xinhang Liu, Fuqiang Zhao, Yanshun Zhang, Yingliang Zhang, Minye Wu, Jingyi Yu, and Lan Xu. Fourier plenoctrees for dynamic radiance field rendering in real-time. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13524-13534, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.569, + 0.47, + 0.624 + ], + "angle": 0, + "content": "[83] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing (TIP), 13(4):600-612, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.626, + 0.47, + 0.707 + ], + "angle": 0, + "content": "[84] Ziyan Wang, Giljoo Nam, Tuur Stuyck, Stephen Lombardi, Chen Cao, Jason Saragih, Michael Zollhöfer, Jessica Hodgins, and Christoph Lassner. Neuwigs: A neural dynamic model for volumetric hair capture and animation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8641-8651, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.71, + 0.471, + 0.765 + ], + "angle": 0, + "content": "[85] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.767, + 0.47, + 0.835 + ], + "angle": 0, + "content": "[86] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P Srinivasan, Dor Verbin, Jonathan T Barron, Ben Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. arXiv preprint arXiv:2312.02981, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.837, + 0.47, + 0.892 + ], + "angle": 0, + "content": "[87] Jamie Wynn and Daniyar Turmukhambetov. Diffusionerf: Regularizing neural radiance fields with denoising diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4180-4189, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.894, + 0.47, + 0.922 + ], + "angle": 0, + "content": "[88] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.075, + 0.471, + 0.922 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.075, + 0.892, + 0.102 + ], + "angle": 0, + "content": "video. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9421-9431, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.104, + 0.892, + 0.157 + ], + "angle": 0, + "content": "[89] Zhen Xu, Sida Peng, Haotong Lin, Guangzhao He, Jiaming Sun, Yujun Shen, Hujun Bao, and Xiaowei Zhou. 4k4d: Real-time 4d view synthesis at 4k resolution. arXiv preprint arXiv:2310.11448, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.16, + 0.892, + 0.215 + ], + "angle": 0, + "content": "[90] Zhiwen Yan, Chen Li, and Gim Hee Lee. Nerf-ds: Neural radiance fields for dynamic specular objects. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8285-8295, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.217, + 0.892, + 0.271 + ], + "angle": 0, + "content": "[91] Bo Yang, Jianan Wang, Ronald Clark, Qingyong Hu, Sen Wang, Andrew Markham, and Niki Trigoni. Learning object bounding boxes for 3d instance segmentation on point clouds, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.273, + 0.892, + 0.314 + ], + "angle": 0, + "content": "[92] Jinyu Yang, Mingqi Gao, Zhe Li, Shang Gao, Fangjing Wang, and Feng Zheng. Track anything: Segment anything meets videos, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.316, + 0.892, + 0.383 + ], + "angle": 0, + "content": "[93] Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Improving few-shot neural rendering with free frequency regularization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8254-8263, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.386, + 0.892, + 0.425 + ], + "angle": 0, + "content": "[94] Zongxin Yang and Yi Yang. Decoupling features in hierarchical propagation for video object segmentation, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.429, + 0.892, + 0.483 + ], + "angle": 0, + "content": "[95] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.486, + 0.892, + 0.541 + ], + "angle": 0, + "content": "[96] Zeyu Yang, Hongye Yang, Zijie Pan, Xiatian Zhu, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. arXiv preprint arXiv:2310.10642, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.543, + 0.892, + 0.611 + ], + "angle": 0, + "content": "[97] Zhengming Yu, Wei Cheng, Xian Liu, Wayne Wu, and Kwan-Yee Lin. Monohuman: Animatable human neural field from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16943-16953, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.612, + 0.892, + 0.681 + ], + "angle": 0, + "content": "[98] Wentao Yuan, Zhaoyang Lv, Tanner Schmidt, and Steven Lovegrove. Star: Self-supervised tracking and reconstruction of rigid objects in motion with neural rendering. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13144-13152, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.683, + 0.892, + 0.751 + ], + "angle": 0, + "content": "[99] Jiakai Zhang, Xinhang Liu, Xinyi Ye, Fuqiang Zhao, Yanshun Zhang, Minye Wu, Yingliang Zhang, Lan Xu, and Jingyi Yu. Editable free-viewpoint video using a layered neural representation. ACM Transactions on Graphics (TOG), 40(4):1-18, 2021. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.753, + 0.892, + 0.822 + ], + "angle": 0, + "content": "[100] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.824, + 0.892, + 0.878 + ], + "angle": 0, + "content": "[101] Renrui Zhang, Zhengkai Jiang, Ziyu Guo, Shilin Yan, Junting Pan, Hao Dong, Peng Gao, and Hongsheng Li. Personalize segment anything model with one shot. arXiv preprint arXiv:2305.03048, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.88, + 0.892, + 0.921 + ], + "angle": 0, + "content": "[102] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J Davison. In-place scene labelling and understanding with implicit scene representation. In IEEE/CVF" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.075, + 0.892, + 0.921 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.956, + 0.52, + 0.967 + ], + "angle": 0, + "content": "19678" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.117, + 0.075, + 0.47, + 0.103 + ], + "angle": 0, + "content": "International Conference on Computer Vision (ICCV), pages 15838-15847, 2021. 2, 3" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.956, + 0.519, + 0.967 + ], + "angle": 0, + "content": "19679" + } + ] +] \ No newline at end of file diff --git a/2024/Gear-NeRF_ Free-Viewpoint Rendering and Tracking with Motion-aware Spatio-Temporal Sampling/8ba31972-a909-4d60-aef6-d492fd479dbb_origin.pdf b/2024/Gear-NeRF_ Free-Viewpoint Rendering and Tracking with Motion-aware Spatio-Temporal Sampling/8ba31972-a909-4d60-aef6-d492fd479dbb_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..042eafd5d46c9dd40fa4f01882316b0237abf9b0 --- /dev/null +++ b/2024/Gear-NeRF_ Free-Viewpoint Rendering and Tracking with Motion-aware Spatio-Temporal Sampling/8ba31972-a909-4d60-aef6-d492fd479dbb_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0f72d361410a270d6195790e8f5b0067078d99a947cf52b805b0880704a281fd +size 4944574 diff --git a/2024/Gear-NeRF_ Free-Viewpoint Rendering and Tracking with Motion-aware Spatio-Temporal Sampling/full.md b/2024/Gear-NeRF_ Free-Viewpoint Rendering and Tracking with Motion-aware Spatio-Temporal Sampling/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6b3ad1141cbf237940bd5111ffaf32fcbc91840a --- /dev/null +++ b/2024/Gear-NeRF_ Free-Viewpoint Rendering and Tracking with Motion-aware Spatio-Temporal Sampling/full.md @@ -0,0 +1,454 @@ +# Gear-NeRF: Free-Viewpoint Rendering and Tracking with Motion-aware Spatio-Temporal Sampling + +Xinhang Liu $^{1\dagger*}$ Yu-Wing Tai $^{2}$ Chi-Keung Tang $^{1\dagger}$ Pedro Miraldo $^{3}$ Suhas Lohit $^{3}$ Moitreya Chatterjee $^{3}$ $^{1}$ HKUST $^{2}$ Dartmouth College $^{3}$ Mitsubishi Electric Research Laboratories (MERL) xliufe@connect.ust.hk, yu-wing.tai@dartmouth.edu, cktang@cse.ust.hk, miraldo@merl.com, slohit@merl.com, metro.smiles@gmail.com + +# Abstract + +Extensions of Neural Radiance Fields (NeRFs) to model dynamic scenes have enabled their near photo-realistic, free-viewpoint rendering. Although these methods have shown some potential in creating immersive experiences, two drawbacks limit their ubiquity: (i) a significant reduction in reconstruction quality when the computing budget is limited, and (ii) a lack of semantic understanding of the underlying scenes. To address these issues, we introduce Gear-NeRF, which leverages semantic information from powerful image segmentation models. Our approach presents a principled way for learning a spatio-temporal (4D) semantic embedding, based on which we introduce the concept of gears to allow for stratified modeling of dynamic regions of the scene based on the extent of their motion. Such differentiation allows us to adjust the spatiotemporal sampling resolution for each region in proportion to its motion scale, achieving more photo-realistic dynamic novel view synthesis. At the same time, almost for free, our approach enables free-viewpoint tracking of objects of interest - a functionality not yet achieved by existing NeRF-based methods. Empirical studies validate the effectiveness of our method, where we achieve state-of-the-art rendering and tracking performance on multiple challenging datasets. The project page is available at: https://merl.com/research/highlights/gear-nerf. + +# 1. Introduction + +Reconstructing 3D scenes has a broad range of applications, including Virtual Reality/Augmented Reality (VR/AR), 3D animation, game production, and film creation which allow users to observe scenes from any desired viewpoint. While it is crucial to reconstruct static scenes, towards which significant progress has been made, it is even more crucial to + +![](images/388e1c857a088b1e69eef4f1773783926afe85c265cffcd81b0d9b12d5e592d1.jpg) +Figure 1. (a) Our method takes RGB videos captured from a camera array as input. (b) Trained Gear-NeRF achieves photo-realistic real-time free-viewpoint rendering of a dynamic scene. (c) With users giving a single click at any time and from any viewpoint, our method can perform free-viewpoint tracking of the target object. + +reconstruct dynamic scenes, as the world around us is characterized by a constant state of flux, with many objects in it - in a state of motion. + +Recent advances in novel view synthesis, such as Neural Radiance Fields (NeRFs) [56] have inspired numerous studies to extend them to dynamic 3D scenes. Existing approaches either employ a deformation field to map neural fields from a given time to a canonical space [22, 45, 60, 61, 65, 99], or directly model dynamic scenes as a 4D space-time grid [1, 6, 26]. Though these methods offer improved rendering quality by utilizing more accessible inputs compared to previous solutions [21, 42, 43, 77], they still struggle to ensure rendering quality in low-resource settings, requiring carefully engineered efforts. Further, most dynamic radiance field approaches adopt a naive spatiotemporal sampling strategy, without discerning the different scales of motion across different regions in the scene. + +We propose to fix this issue by leveraging a semantic understanding of dynamic scenes. Intuitively, a reconstruction system aware of the distinction between static and dynamic regions in a scene can perform more focused + +![](images/500abea19269eadef36d0ba7557bc09d847c481cb0d319ca0167698cf0cf9a5a.jpg) +Figure 2. Pipeline of Gear-NeRF: Gear-NeRF takes multi-view videos as input. After optimizing the serial 4D feature volumes (Section 4.1), it maps space-time coordinates to a 4D semantic embedding (Section 4.2), in addition to the volume density and view-dependent radiance color. Regions with larger motion are automatically assigned higher gear levels (Section 4.3) and as a result, receive higher-resolution spatio-temporal sampling (Section 4.4). Furthermore, Gear-NeRF is capable of performing free-viewpoint tracking of a target object with prompts as simple as a user click (Section 4.5). + +sampling in the dynamic regions, which inherently require more resources per unit volume than static regions, due to their time-evolving nature. Accordingly, dynamic regions can be further stratified according to their scale of motion. To this end, this paper presents Gear-NeRF, a framework that leverages semantic embedding from powerful image segmentation models [38] for stratified modeling of 4D scenes. Gear-NeRF optimizes for a 4D semantic embedding, based on which we introduce the concept of gear to smartly determine the appropriate region-specific resolution of spatio-temporal sampling in the NeRF. Regions with larger motion scales are assigned higher gears, through our gear determination scheme and we accordingly perform higher-resolution spatio-temporal sampling. Empirical studies reveal that this motion-aware sampling strategy improves the quality of synthesized images, over competing approaches. As a by-product of our semantically embedded representation, we achieve free-viewpoint object tracking, given user prompts. Figure 1 presents an overview of the capabilities of our method. + +Gear-NeRF makes two primary advancements: (i) enhanced dynamic novel view synthesis by resorting to smarter spatio-temporal sampling, and (ii) the ability for free-viewpoint tracking of objects of interest. The latter is a capability not yet realized by existing NeRF methods for dynamic scenes. We perform extensive experiments on multiple datasets to validate the generalizability and robustness of our method, which shows state-of-the-art performances for both tasks across all datasets. + +# 2. Related Work + +Neural Radiance Fields: NeRF [56] is a recent breakthrough among novel view synthesis methods that uses multilayer perceptrons (MLPs) to parameterize the appearance + +and density for each point in 3D space, given any viewing direction of the scene. Researchers have extended NeRF along various dimensions [78], including improving rendering quality [2-4, 12, 18, 31], handling challenging conditions such as large scenes [55, 68, 75], view-dependent appearances [30, 52, 80], and sparse inputs [34, 51, 59, 86, 87, 93]. NeRF-like neural representations have also found applications in semantic segmentation [40, 50, 102] and 3D content generation [9, 10, 49, 64]. Recent work has shown that replacing the deep MLPs with a feature voxel grid can significantly improve training and inference speed [11, 25, 58, 74]. On the other hand, a more recent approach to further improve visual quality, rendering time, and performance entails representing the scene with 3D Gaussians [36]. Our approach, while drawing upon many of these approaches, deals with dynamic scenes which is beyond the scope of these methods. + +Neural Representations for Dynamic Scenes: NeRF-like representations have recently been extended to model dynamic scenes in high fidelity [17, 28, 33, 39, 46-48, 53, 63, 66, 84, 89, 90, 97]. One straightforward approach to do this is to directly condition the radiance field on time [27, 44, 45, 88]. Alternatively, several methods model a deformation field to map coordinates from different time stamps to a common canonical space [22, 24, 60, 61, 65, 79, 98, 99]. Some recent approaches [6, 26, 71, 81] represent the scene using a 4D space-time grid, which is decomposed into sets of planar representations for training efficiency. Other techniques for improving rendering fidelity and frame rate include Fourier PlenOctrees [82], ray-conditioned sample prediction networks [1], 4D space decomposition (static/dynamic/newly appeared regions) [73], and explicit voxel grids [24]. 3D Gaussians have also been adapted to model dynamic scenes [54, 85, 95, 96]. While these approaches paved the initial path for rendering dy + +namic scenes, semantically aware modeling of the scene is absent, a caveat that our proposed method seeks to address. Segment Anything Model (SAM): SAM [38] is a powerful promptable image segmentation model, which showcases remarkable zero-shot generalization abilities and can produce semantically consistent masks, given a single foreground point on the image. HQ-SAM [35] is an improvement on SAM that enhances the quality of masks, especially on objects with intricate boundaries and structures. Recent works [20, 92] have extended SAM to perform interactive video object segmentation. These methods utilize SAM for mask initialization or correction and then employ state-of-the-art mask trackers [19, 94] for mask tracking and prediction [14]. Recent methods have also leveraged SAM for tracking multiple reference objects in a video [13, 101]. This work uses SAM to profile the scene into semantic regions, which are then grouped based on motion scales. + +3D Semantic Understanding: Existing methods for 3D visual understanding [15, 32, 76, 91] mainly focus on closed set segmentation of point clouds or voxels. NeRF's ability to integrate information across multiple views has led to its applications in 3D semantic segmentation [102], object segmentation [23, 50, 57], panoptic segmentation [72], and interactive segmentation [29, 69]. Kobayashi et al. [40] explored the effectiveness of embedding pixel-aligned features [7, 41] into NeRFs for 3D manipulations. LERF [37] fuses CLIP embeddings [67] and NeRFs to enable language-based localization in 3D NeRF scenes. Recent work has enabled click/text-based 3D segmentation by learning a 3D SAM embedding [16] or inverse rendering of SAM-generated masks [8]. We, on the other hand, seek to utilize the synergy of dynamic NeRFs and SAM segments to derive a semantic understanding of a dynamic 3D scene for tracking objects of interest in novel views – a first of its kind effort. + +# 3. Preliminaries + +Neural Radiance Fields (NeRFs): Vanilla NeRFs [56] employ a multi-layer perceptron (MLP) with sinusoidal positional encoding to map a 3D-spatial coordinate $\mathbf{x} = (x, y, z)$ and a viewing direction $\mathbf{d} = (\theta, \phi)$ to a volume density $\sigma \in [0,1]$ and an emitted RGB, $\mathbf{c} \in \mathbb{R}^3$ . Rendering each image pixel involves casting a ray $\mathbf{r}(t) = \mathbf{o} + t\mathbf{d}$ from the camera center $\mathbf{o}$ through the pixel along direction $\mathbf{d}$ . The predicted color for the corresponding pixel is computed as: + +$$ +\hat {\mathbf {C}} (\mathbf {r}) = \sum_ {i = 1} ^ {N} T _ {i} \alpha_ {i} \mathbf {c} _ {i}, \tag {1} +$$ + +where $T_{i} = \exp \left(-\sum_{j=1}^{i-1} \sigma_{j} \delta_{j}\right)$ , $\alpha_{i} = 1 - \exp (-\sigma_{i} \delta_{i})$ , and $\delta_{j} = t_{j+1} - t_{j}$ . A vanilla NeRF is trained by minimizing the mean squared error between the input images and the predicted images, obtained by rendering the scene from the viewpoints from which the input images have been + +captured, with the training loss given by: + +$$ +\mathcal {L} _ {\mathrm {p h o}} = \sum_ {\mathbf {r} \in \mathcal {R}} \| \hat {\mathbf {C}} (\mathbf {r}) - \mathbf {C} (\mathbf {r}) \| _ {2} ^ {2}. \tag {2} +$$ + +where $\mathcal{R}$ is the set of all rays projected from the input image. Planar-Factorized 4D Volumes: A recent emerging trend of handling dynamics using radiance field representations is to directly adapt them to be conditioned on a frame index $t$ (denoting time) in addition to $\mathbf{x}$ and $\mathbf{d}$ . This can be accomplished by learning a mapping from $(\mathbf{x},\mathbf{d},t)$ to $(\sigma ,\mathbf{c})$ using planar-factorized 4D volumes [1, 6, 26, 71]. These methods attempt to learn a 4D feature vector for every $(\mathbf{x},t)$ , by projecting it to a set of 2D-planes. Embeddings of these projections on these planes can then be integrated to obtain the embedding for the 4D point. This can be mathematically represented as follows: + +$$ +\begin{array}{l} \mathbf {f} (\mathbf {x}, t) = \mathbf {B} _ {1} \left(\mathbf {h} _ {1} (x, y) \odot \mathbf {k} _ {1} (z, t)\right) \\ + \mathbf {B} _ {2} (\mathbf {h} _ {2} (x, z) \odot \mathbf {k} _ {2} (y, t)) \\ + \mathbf {B} _ {3} \left(\mathbf {h} _ {3} (y, z) \odot \mathbf {k} _ {3} (x, t)\right). \tag {3} \\ \end{array} +$$ + +where $\mathbf{h}_i(\cdot ,\cdot)$ and $\mathbf{k}_i(\cdot ,\cdot)$ are functions (evaluated by bilinear interpolation on regularly spaced 2D feature grids) embedding coordinate tuples to features of dimension $M$ " $\odot$ denotes an element-wise product, and $\mathbf{B}_i(\cdot)$ denotes a linear transform which maps the products to feature vectors. Subsequently, a tiny MLP can map the feature vector $\mathbf{f}(\cdot ,\cdot)$ to the volume density, $\sigma$ , and the view-dependent emitted color, c, given the viewing direction d. + +# 4. Proposed Method + +Given a set of $W$ input videos, $\mathcal{V} = \{V_1, V_2, \dots, V_W\}$ of a dynamic scene, with calibrated camera poses, our approach represents the scene using a series of 4D feature volumes (Section 4.1) along with 4D semantic embeddings (Section 4.2). + +Analogous to multiple gears in motor vehicles for optimizing engine performance, Gear-NeRF stratifies this semantically embedded scene representation into $N_{\mathrm{gear}}$ levels, based on the motion scales. Each of these levels is called a gear. Through our training scheme, regions with larger motion are assigned higher gear levels (Section 4.3) and as a result, are more densely sampled (Section 4.4) for improved dynamic novel view synthesis. Our 4D semantic embedding also enables a new functionality, almost for free - free-viewpoint tracking of target objects, given simple user prompts like clicks (Section 4.5). Figure 2 shows the overall pipeline of Gear-NeRF. + +# 4.1. Serial 4D Feature Volumes + +Instead of using a unified 4D volume to represent a dynamic scene [1, 6, 26, 71], our representation consists of a series of feature volumes, each corresponding to a gear level, $\mathcal{G}$ . Specifically, for any space-time coordinate $(\mathbf{x}, t)$ , its feature + +vector corresponding to $\mathcal{G}$ is computed as follows: + +$$ +\begin{array}{l} \mathbf {f} ^ {\mathcal {G}} (\mathbf {x}, t) = \mathbf {B} _ {1} \left(\mathbf {h} _ {1} (x, y) \odot \mathbf {k} _ {1} ^ {\mathcal {G}} (z, t)\right) \\ + \mathbf {B} _ {2} (\mathbf {h} _ {2} (x, z) \odot \mathbf {k} _ {2} ^ {\mathcal {G}} (y, t)) \\ + \mathbf {B} _ {3} \left(\mathbf {h} _ {3} (y, z) \odot \mathbf {k} _ {3} ^ {\mathcal {G}} (x, t)\right). \tag {4} \\ \end{array} +$$ + +The vector-valued functions $\mathbf{h}_j(\cdot, \cdot)$ and linear transforms $\mathbf{B}_j(\cdot)$ are shared by all gears, while each gear has its own spatio-temporal embedding $\mathbf{k}_j^G(\cdot, \cdot)$ , in $M$ -dimensional space. Therefore, each gear describes regions of a certain scale of motion while the purely spatial features can be shared among all gears. + +We obtain the gear level at any spatio-temporal coordinate also from a planar-factorized 4D feature volume. Specifically, the gear level at $(\mathbf{x},t)$ is computed as: + +$$ +\begin{array}{l} g (\mathbf {x}, t) = \mathbf {1} ^ {\top} \left(\mathbf {h} _ {1} ^ {\prime} (x, y) \odot \mathbf {k} _ {1} ^ {\prime} (z, t)\right) \\ + \mathbf {1} ^ {\top} \left(\mathbf {h} _ {2} ^ {\prime} (x, z) \odot \mathbf {k} _ {2} ^ {\prime} (y, t)\right) \\ + \mathbf {1} ^ {\top} \left(\mathbf {h} _ {3} ^ {\prime} (y, z) \odot \mathbf {k} _ {3} ^ {\prime} (x, t)\right), \tag {5} \\ \end{array} +$$ + +where $\mathbf{1}$ is a vector of ones, $\mathbf{h}_i^{\prime}(\cdot ,\cdot)$ and $\mathbf{k}_i^\prime (\cdot ,\cdot)$ are $M$ dimensional embedding functions. This however defines a continuous feature volume. To map it to the gear level integers, we apply the following projection operation: + +$$ +p (\mathbf {x}, t) = \left\{ \begin{array}{l l} 1, & \text {i f} g (\mathbf {x}, t) < 1, \\ N _ {\text {g e a r}}, & \text {i f} g (\mathbf {x}, t) \geq N _ {\text {g e a r}}, \\ \lceil g (\mathbf {x}, t) \rceil , & \text {o t h e r w i s e .} \end{array} \right. \tag {6} +$$ + +Based on this gear level volume, we define a 4D mask for a region at gear level $\mathcal{G}$ as: + +$$ +m _ {\mathcal {G}} (\mathbf {x}, t) = \left\{ \begin{array}{l l} 1, & \text {i f} p (\mathbf {x}, t) = \mathcal {G}, \\ 0, & \text {o t h e r w i s e .} \end{array} \right. \tag {7} +$$ + +The final feature vector at $(\mathbf{x},t)$ is computed as: + +$$ +\mathbf {f} (\mathbf {x}, t) = \sum_ {\mathcal {G} = 1} ^ {N _ {\text {g e a r}}} m _ {\mathcal {G}} (\mathbf {x}, t) \mathbf {f} ^ {\mathcal {G}} (\mathbf {x}, t). \tag {8} +$$ + +Subsequently, a tiny MLP, $F_{\theta}$ , maps these feature vectors $\mathbf{f}(\cdot ,\cdot)$ as well as the viewing direction $\mathbf{d}$ to the volume density $\sigma$ and radiance color $\mathbf{c}$ . This allows us to obtain a photometric rendering of the scene. + +# 4.2. 4D Semantic Embedding + +Gear-NeRF leverages the strong object priors of the SAM [38] model to acquire a semantic understanding of the scene, for improved photometric rendering (Section 4.3 and Section 4.4) as well as free-viewpoint object tracking (Section 4.5). Toward this end, we utilize the SAM encoder to obtain 2D feature maps from the frames of each video. We then optimize a 4D SAM embedding field by supervising it with these 2D feature maps. In particular, the MLP above, $F_{\theta}$ , is configured to output a 4D semantic embedding s for + +![](images/e3fa72dc36fd506e57f807a0b5f7022d82a22a739737276e39de2b2edcc87568.jpg) +$1^{\mathrm{st}}$ Gear Assignment Update + +![](images/50fa637dd9b5cd1d8ef5c3734619da2a8b18532884ebf3307e56e6012fac32e4.jpg) + +![](images/e23f902aa511437326737a562472d7caf86849a36e8cd344b80d94bdfa92a64a.jpg) + +![](images/d87f3f86106845ae486f9020b20c21ed3ab64b99550e216c7f09eea295157c5c.jpg) +2nd Gear Assignment Update + +![](images/47af5c1ed79559ca8d3a83c738ee9115e1dc867b9daf89f179fe35cfc79ac20c.jpg) +$3^{\mathrm{rd}}$ Gear Assignment Update + +![](images/440fbf6fbb39d32553f1849c7425139eb95591a5e7cf66188db92618d076026a.jpg) + +![](images/2943450a8edc3e657892f88efead7cc1277c40ff2564f801a6aba17522ec63e1.jpg) + +![](images/dbea8c9a4761b4dfee689b657b03a59d77398d43bea0d872794a3f642fdd62ea.jpg) + +![](images/07dc913a232a4ca56bdf8c7f81a07d84290cd2bbaf7e7b9422c685494c190bec.jpg) + +![](images/e66b9f16b1e8424443e7fdb89e5a5849e28053e972b233b72ef7e949bedcee11.jpg) +Rendered Image +Figure 3. Illustration of Gear Assignment Update: For each gear assignment update, we calculate the rendering loss map between the rendered RGB-SAM map and the ground truth and identify the centers of the patches with the maximum and minimum losses, marked in red and green (second column). These points are then fed into the SAM decoder as positive and negative prompts to generate an upshift mask representing the areas that need to be shifted to a higher gear (last column). After the first gear assignment update, we see that the next candidate region for upshift is situated where the horse is located, and so on. Upshift mask colors imply the gear levels after the update (blue-2, green-3, red-4). + +![](images/b7d48a34fdb91fa0479f22030faadcd57dedd2e93119424b892303001284c9bd.jpg) +$4^{\mathrm{th}}$ Gear Assignment Update +Rendering Loss Map + +![](images/e53c1f54b24deda748147abfcf6d354e8558e6bd4ac49d6499b0bdd54d13466f.jpg) +Upshift Mask + +a given space-time coordinate in addition to the density, $\sigma$ , and color, c. To render 2D semantic feature maps in a given view, we compute the semantic feature of a pixel in the feature map by tracing a ray through it and perform volume rendering analogous to Equation 1, as follows: + +$$ +\hat {\mathbf {S}} (\mathbf {r}) = \sum_ {i = 1} ^ {N} T _ {i} \alpha_ {i} \mathbf {s} _ {i}. \tag {9} +$$ + +This SAM embedding is supervised by minimizing the mean squared error between the prediction and the ground truth features $(\mathbf{S}(\mathbf{r}))$ from the SAM encoder, as shown: + +$$ +\mathcal {L} _ {\mathrm {S A M}} = \sum_ {\mathbf {r} \in \mathcal {R}} \| \hat {\mathbf {S}} (\mathbf {r}) - \mathbf {S} (\mathbf {r}) \| _ {2} ^ {2}. \tag {10} +$$ + +# 4.3. Training Scheme with Gear Assignment + +With gear initialization $g(\mathbf{x}, t) = 1, \forall \mathbf{x}, t$ , the (semantically embedded) radiance field optimization and gear assignment updating take place in an alternating fashion. + +Gear Assignment Update: As illustrated in Figure 3, when updating the gear assignment after a period of radiance field + +optimization, we find the regions rendered most poorly from the rendering loss maps and increment their gears for denser spatio-temporal sampling. The following steps lay out the process for updating gear assignments to regions: + +- We sample a number of viewpoints and time steps and render 2D-images/SAM features for it. For every rendered RGB-SAM map, we compute a rendering loss map. Each pixel of the rendering loss map is computed as: $\mathcal{L}(\mathbf{r})\coloneqq \mathcal{L}_{\mathrm{pho}}(\mathbf{r}) + \lambda \mathcal{L}_{\mathrm{SAM}}(\mathbf{r})$ . See Figure 3 for example loss maps. +- Next, we patchify each rendering loss map to find patches with the top- $k$ largest/smallest average loss. The center coordinate of these patches serve as positive/negative prompts for the next step. See Figure 3 for example positive (red) / negative (green) prompts. +- We then feed the ground truth RGB images together with positive and negative prompts into the SAM decoder [38] to estimate an upshift mask. These masks tend to cover regions that have motions and are not satisfactorily rendered with the current sampling resolution. We have multiple upshift masks at different viewpoints and time steps. +- For every pixel of an upshift mask, we trace a ray and sample points along it and update the gear assignment by pushing $g(\mathbf{x}, t)$ towards incremented values. + +In particular, in the last step, for each pixel within an upshift mask, a corresponding ray is traced that connects it with the camera center o, along direction d. Next, a set of points are sampled along this ray. The collection of sampled points, that lie on the rays emanating from within the masked region constitutes the set $S_{\mathrm{upshift}} = \{(x_i^{\mathrm{upshift}},t_i^{\mathrm{upshift}})\}_{i = 1}^{N_{\mathrm{upshift}}}$ , where $N_{\mathrm{upshift}}$ represents the total count of points sampled from rays pertaining to the masked region. We then follow a similar procedure to sample a set of points pertaining to the unmasked region. We denote this set as: $\mathcal{S}_{\mathrm{stay}} = \{(x_i^{\mathrm{stay}},t_i^{\mathrm{stay}})\}_{i = 1}^{N_{\mathrm{stay}}}$ , with $N_{\mathrm{stay}}$ indicating the total number of points sampled from rays that pertain to the unmasked area. Next, for each sample point in each of $S_{\mathrm{upshift}}$ and $S_{\mathrm{stay}}$ , we query its current gear level $p(\mathbf{x},t)$ . In order to assign new gear levels, we need to update the gear assignment function $g(\cdot ,\cdot)$ , which proceeds with the objective function: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {u p s h i f t}} = \frac {1}{N _ {\text {u p s h i f t}}} \sum_ {(\mathbf {x}, t) \in \mathcal {S} _ {\text {u p s h i f t}}} \| g (\mathbf {x}, t; \boldsymbol {\Theta}) - (p (\mathbf {x}, t) + 1) \| _ {2} ^ {2} \\ + \frac {\lambda_ {\text {s t a y}}}{N _ {\text {s t a y}}} \sum_ {(\mathbf {x}, t) \in \mathcal {S} _ {\text {s t a y}}} \| g (\mathbf {x}, t; \boldsymbol {\Theta}) - p (\mathbf {x}, t) \| _ {2} ^ {2}, \tag {11} \\ \end{array} +$$ + +where $\Theta$ denotes the set of estimizable parameters for $g(\cdot ,\cdot)$ . The minimization of the first term encourages sample points within the masked region $S_{\mathrm{upshift}}$ to have a gear level equal to their current gear incremented by one, resulting in an upshift of gear. Conversely, minimizing the second term encourages the points in $S_{\mathrm{stay}}$ to maintain their gear values, thereby encouraging the remaining regions to keep their current gear levels and avoiding unwanted up + +shifts. We update $\Theta$ via a single step of gradient descent as follows: + +$$ +\boldsymbol {\Theta} \leftarrow \boldsymbol {\Theta} - \alpha \nabla_ {\boldsymbol {\Theta}} \mathcal {L} _ {\text {u p s h i f t}}, \tag {12} +$$ + +where $\alpha$ is the learning rate. Since, $g(\cdot ,\cdot)$ is essentially derived from the embedding functions, $\mathbf{h}_i^{\prime}(\cdot ,\cdot)$ and $\mathbf{k}_i^\prime (\cdot ,\cdot)$ for $i\in \{1,2,3\}$ , the aforementioned optimization step amounts to updating these embedding functions. Once updated, we proceed with a fresh round of gear assignment to increase the spatio-temporal sampling resolution for the regions that end up at a higher gear level than before. + +Radiance Field Optimization: With the updated gear assignment, we increase the resolution of spatio-temporal sampling for the gear-shifted regions (Section 4.4) and then resume optimizing the radiance field. We alternate between the two processes: radiance field optimization (each time for $L$ epochs), and gear assignment updates until the average variance of each rendering loss map is below a predetermined threshold. After this, we optimize the radiance field for an additional $L'$ epochs without further gear assignment updates. + +# 4.4. Motion-aware Spatio-Temporal Sampling + +In this subsection, we explain our motion-aware spatiotemporal sampling strategy based on assigned gears, permitting differential processing of regions at different gear levels. By temporal sampling, we imply the choice of temporal resolution for planar-factorized 4D feature volumes, and by spatial sampling, we mean the strategy used to choose sampling points along each ray for volume rendering. + +Motion-aware Temporal Sampling: To handle the increasing intensity of object motion, as reflected by their growing gear levels, we increment the temporal resolution for voxel grids. Specifically, $\mathbf{k}_j^{\mathcal{G}}$ in Equation 4 has increasing resolution along the time axis, thereby empowering the 4D feature volumes to better model the dynamics along the temporal axis. This ensures fast-moving objects can be more faithfully modeled without unsightly blurring. The temporal resolution for each gear's feature volume is determined by linear interpolation between 1 (for $\mathcal{G} = 1$ ) and the total number of frames (for $\mathcal{G} = N_{\mathrm{gear}}$ ). + +Motion-aware Spatial Sampling: While denser sampling of points can improve reconstruction accuracy, increasing the number of sampling points throughout the scene can lead to prohibitive computational costs. Therefore, we propose a 3D point-splitting strategy as illustrated in Figure 4. We begin by sampling a relatively small number, $n$ , of samples along each ray, assuming it is at the lowest gear level. If a sampled point belongs to a region with a higher gear, as determined by $p(\mathbf{x}, t)$ , we then sample more densely in that region. For every sampled point in that region, we split it into $2^{p(\mathbf{x}, t) - 1}$ points, equally spaced within the corresponding ray segment (at that gear level). + +![](images/67bfcb663407705ef6fd0df36ebff6879367fd9d28578ab337310359da5d1ed8.jpg) +Figure 4. Motion-aware Spatial Sampling: We split each sampled point into $2^{p(\mathbf{x},t)}$ points, equally spaced within the corresponding ray segment. The top row shows the vanilla uniformly sampled points, while the bottom one shows the densely sampled points after splitting using our strategy. + +# 4.5. Free-Viewpoint Tracking with User Prompts + +Our 4D SAM embedding enables another useful functionality, almost for free - free-viewpoint object tracking, where the user only needs to provide as few as one click to extract the target object based on the 4D embedding. Next, we describe how, given a user-supplied point click at any arbitrary viewpoint and time step, we obtain the corresponding object mask at a novel viewpoint and time step. + +Masks for Novel Viewpoints: The first step for this task entails finding the 3D correspondence of the user click. We trace a ray through the selected pixel, and by utilizing the volume density, we determine the depth at which the ray intersects with the first object surface it encounters. This yields the 3D coordinates of the point of intersection. Subsequently, the 3D coordinates of this intersection can be easily mapped into a 2D coordinate within any novel viewpoint image, using the camera pose of the new viewpoint. Alongside the rendered SAM feature map of the novel view, we feed this coordinate into the SAM decoder to generate the object mask for the novel view. + +Masks for Novel Time Steps: For this task, we propagate an object mask to its neighboring time step. Specifically, with an object mask for a specific frame $t$ , we calculate the bounding box of this mask and use this bounding box as a prompt to SAM for neighboring frames $t' = t + 1$ or $t' = t - 1$ . By inputting this prompt along with the rendered SAM feature map at $t'$ into the SAM decoder, we can obtain the object mask for $t'$ . Combining the above two processes, we can start from a single click and get the object mask in any viewpoint and time step. + +# 5. Experiments + +We assess the performance of our proposed Gear-NeRF for dynamic novel view synthesis and free-viewpoint object tracking across a range of challenging datasets, comparing it with state-of-the-art methods. Through ablation studies, we provide empirical evidence of the effectiveness of its fundamental components. We kindly refer the reader to our supplementary material for additional experimental details and results, including videos for free-viewpoint rendering and tracking. + +# 5.1. Experimental Setup + +Implementation Details: We implement our method using PyTorch [62] and conduct experiments on an NVIDIA RTX 4090 GPU with 24 GB RAM. We divide each input video into chunks of 100 frames. For every chunk, we train a model for approximately 2.5 hours. Our 4D feature volumes yield embeddings with a dimension of $M = 32$ . We set the gear number $N_{\mathrm{gear}}$ to 4. We find patches with top- $k = 3$ largest/smallest average loss for gear assignment updates to obtain prompts. The rendering loss map is computed with $\lambda = 0.01$ . For the optimization of radiance fields, $L = 3$ and $L' = 10$ . In our motion-aware spatial sampling, each ray initially has $n = 64$ sampling points. We use an initial learning rate of 0.02 for all the parameters. For Equation 12, we use $\alpha = 0.02$ . + +Datasets: (i) The Technicolor light field dataset [70] includes diverse indoor environment videos captured by a $4 \times 4$ camera rig. We evaluate on 4 sequences (Train, Theater, Painter, Birthday) at the original $2048 \times 1088$ resolution, holding out the same view as prior work [1] (the second row and second column) for evaluation. (ii) The Neural 3D Video dataset [44] includes indoor multi-view video sequences captured by 20 cameras at a resolution of $2704 \times 2028$ pixels. We experiment on 6 sequences (Cut Roasted Beef, Flame Steak, Coffee Martini, Cook Spinach, Flame Salmon, Sear Steak), downsampling by a factor of 2 and holding out the central view (akin to prior work [1]) for evaluation. (iii) The Google Immersive dataset[5] contains light field videos of indoor and outdoor scenes captured by a time-synchronized 46-fisheye camera rig, with a resolution of $2560 \times 1920$ pixels. We experiment with 9 sequences from it (Flames, Truck, Horse, Car, Welder, Exhibit, Face Paint 1, Face Paint 2, Cave). We downsample the video resolution by a factor of 2 and hold out the central view (like prior work [1]) for evaluation. For our experiments, we adopted the same resolution and held-out view selection as prior work [1]. + +Evaluation Metrics: For the task of novel view synthesis of dynamic scenes, we evaluate using the following standard metrics: (i) Peak Signal-to-Noise Ratio (PSNR), (ii) Structural Similarity Index Measure (SSIM) [83] and (iii) Learned Perceptual Image Patch Similarity (LPIPS) [100], by comparing the reconstructed frames against ground truth images. These metrics are computed on the held-out view and averaged across all frames. For the free-viewpoint object tracking task, we designate a specific viewpoint and time step for the user to give prompting clicks. Subsequently, we assess the quality of the predicted object masks at novel viewpoints. The quality of the object mask is quantified in terms of the Mean Intersection over Union (mIoU) and accuracy (Acc.), which are calculated against the ground truth mask, manually annotated utilizing Adobe Photoshop. Additionally, we present the same metrics computed for novel time steps, denoted as t-mIoU and t-Acc. + +Baselines: We run a comprehensive comparison of our + +![](images/995f64513dc7dcc780c6471fd2e6af6df9a29b5181a29b9bd39043bebb702214.jpg) + +![](images/dc52ec33367132fb857c04026dd2a395088ccd663d23f174067dcaabf1cbdca7.jpg) + +![](images/d4db5f3a30ed9868c493d8acb4701fa98355ebc454111ac934fbdec4a31fc069.jpg) + +![](images/a2cb6d42855b09e182add840584accf719154eb6b74ae6f1764b137bdc28e1d2.jpg) + +![](images/dfe3e86592245362b13d0599a3240e10d819a2960401b7398bb90495e4d0b5b1.jpg) + +![](images/e6fe6e58d2bf1f0e4a8852645168917a64beb9d389006baf35e56d2beef68533.jpg) + +![](images/f26ffd8f80ec8c5d2c9669fe25d385d4d25418244054b97e1c9eec17915a6003.jpg) + +![](images/a25bb988e9ffcd1291ae92e5a376aff57865a49f26eddafc48e22c8280f74f5d.jpg) + +![](images/f659904e0228a212bd8293989fd4201852d0d9015be72fb69713b180863352b7.jpg) +Ground Truth + +![](images/466c8a2e5a7934f4152f4690fde1baa9090778ebd4b92c1135daf028aab74129.jpg) +HyperReel + +![](images/2072749a5c5a3b280a167ec83d172e289318ca9278b7384a4a7c0e2bb1d8e93e.jpg) +Neural 3D Video +Figure 5. Qualitative comparisons for novel view synthesis on the Technicolor dataset [70]: We qualitatively compare our approach against HyperReel [1] and Neural 3D Video [44]. Our approach better recovers fine details like patterns on the toys or stripes on the shirt. + +![](images/2a982f4a69cc86b234e01c70e77fc14a9d2325513fb985926b171a73241f5754.jpg) +Ours + +method against a range of recent NeRF-based baseline methods: (i) ST-NeRF [99], (ii) HexPlane [6], (iii) HyperReel [1], and (iv) MixVoxels [81]. As the first method to enable promptable free-viewpoint object tracking under the NeRF setting, there is no established direct baseline for this specific task. However, SA3D [8], a recent method for segmenting static scenes, is treated as a baseline for predicting masks at novel viewpoints corresponding to the prompted time step. + +# 5.2. Results + +Dynamic Novel View Synthesis: As shown in Figure 5, Gear-NeRF produces high-quality novel view synthesis of dynamic scenes, accurately modeling real-world scenes with intricate motions and fine details. For example, patterns on the toys or stripes on the shirt, are faithfully rendered, resulting in more photo-realistic images compared to all of the baselines. Table 1, presents quantitative comparisons of our method against the baselines. While Gear-NeRF has longer training (Tr. Time) / inference times (FPS) compared to some baselines, it almost always achieves the best performance in terms of rendering quality. + +Free-Viewpoint Tracking: In Figure 6, we present the object masks obtained by our method based on the user prompts provided at a specified viewpoint and time step. Specifically, the first row displays masked objects at the prompted viewpoint and time step. The second row shows novel view masks at the prompted time step. The third row shows novel view masks at novel time steps. We see that masks obtained from Gear-NeRF show precise boundaries, + +Table 1. Quantitative comparisons for dynamic novel view synthesis: Our method outperforms all baselines across all datasets on all metrics. We report means over all scenes for each dataset. Best and second best results are highlighted. + +
DatasetMethodPSNR (↑)SSIM (↑)LPIPS (↓)Tr. Time(↓)FPS(↑)
Technicolor [70]ST-NeRF [99]30.860.8830.101344 min0.7
HyperReel [1]31.040.8870.09297 min7.7
MixVoxels [81]28.990.8420.10314 min18.9
Ours32.210.9190.058148 min7.9
Google Immersive [5]HexPlane [6]27.670.8080.188527 min0.9
HyperReel [1]28.320.8620.145117 min7.4
MixVoxels [81]27.140.8350.20919 min16.8
Ours28.740.8760.122189 min7.0
Neural 3D Video [44]ST-NeRF [99]31.030.8900.081367 min0.7
HyperReel [1]31.120.9280.065129 min6.1
MixVoxels [81]30.690.9440.13915 min16.3
Ours31.800.9360.058204 min6.8
+ +compared to SA3D. Table 2 presents quantitative assessment of the quality of the masks generated of our method versus SA3D. Our method exceeds $90\%$ across all metrics and datasets, demonstrating the effectiveness of our approach. Our gains over SA3D can be attributed to the fact that SA3D, as a static scene segmentation method, does not utilize information across all time frames, whereas our approach does. SA3D is incapable of predicting masks for novel time steps, and as such, the corresponding entries are marked as not applicable, a shortcoming which our method does not have. + +# 5.3. Evaluations + +To verify the effectiveness of the design choices of GearNeRF, we perform extensive ablation studies on the Truck scene of the Google Immersive dataset [5]. + +Motion-aware Temporal Sampling: An intuitive strategy for adjusting the temporal sampling is to directly modify + +![](images/c72389cbbf17f9d3b80ca8280e02a5715b7ac600a62094223b1da648b940c8bf.jpg) +★ User Prompt + +![](images/584ffa2ede060c98cbd97fc22a40b6418869315379b5d85f7db39d9701e145c9.jpg) +Rendering from Novel View +Rendering from Novel View at Novel Time +Rendered Image +Figure 6. Qualitative comparisons of free-viewpoint object tracking on Technicolor [70] and Neural 3D Video [44] datasets: Our method can obtain desirable object masks, with clear edges, from prompting points provided by users at desired time steps and viewpoints. + +![](images/4230cacf43138d46dfe8dd3e44d0a838b064007f64f61bfececcf09953f8dbfd.jpg) + +![](images/8d6c4c78db77b3f57a91aacb3d8bbe92fc2654618baf82e02bfdbb80f9ba8d21.jpg) + +![](images/94b307177c859840ad0d1b8da6483e1303c8216e5b89e585e9a8d613e74eb3ba.jpg) +SA3D + +![](images/4c70833af9728c2eaae13facae9812e0ae2642aadc1180731d70446c4038e3f9.jpg) + +![](images/1bc3f75de1bf3b7d6e9526f53bd2708669632d330658a88a0eb27d7d6eb45781.jpg) + +![](images/3032b11ac1cc1e51c64f9c1849c499d2a0dec2be2e55d8d0fdf793875c7bf1ea.jpg) +Ours + +![](images/81483608b12406fd56271932d4a42eb35caf3c997ce397cbfc91db5d0df4d59d.jpg) + +![](images/d7b36c0da5e6aaa56cc01d6895d00a1603c0c597853ebdd657e7b17e615c259e.jpg) + +![](images/74f70bbb2b3b9c1d85479f1ef17b95d46dea3fd1bac1b504f5471667aec3cf79.jpg) +Rendered Image + +![](images/e9ce9c240794324dbf67d0d133faf98920106b37155c4d8da9cf7eb2431354d4.jpg) + +![](images/baf9ccfe19d602f96e0583f966b0389b237d013d4b5c1834fcb94e38b5f91ac3.jpg) + +![](images/5d20b5754326b29d9840ff2b282124d7e4f00b45bebe71696a0fda8c374283c5.jpg) +SA3D + +![](images/a08f7e5b49cc3ef1f811bd27167ac01f75380d1e696b8c7b05f80fd7b75354ec.jpg) + +![](images/bfb82b569b7877578e997be0670a95f7e91ac38b0f09965fc3683e683d324a64.jpg) + +![](images/d8079c3abe7497ee3ef2e41c709929aa5bdbc990a7e0582a04be49801357d3d9.jpg) +Ours + +Table 2. Quantitative comparisons for free-viewpoint tracking: t-mIoU and t-Acc are metrics used for evaluating novel view masks at novel time steps, not applicable to SA3D. Reported metrics are averages over all scenes for each dataset. + +
DatasetMethodmIoU (↑)Acc. (↑)t-mIoU (↑)t-Acc. (↑)
Technicolor [70]SA3D [8]96.497.1N/AN/A
Ours97.497.692.193.3
Google Immersive [5]SA3D [8]94.194.8N/AN/A
Ours94.395.091.592.8
Neural 3D Video [44]SA3D [8]93.194.0N/AN/A
Ours93.494.390.692.3
+ +the temporal resolution of the 4D feature volume and make it uniform across all regions of the scene. In Table 3, we demonstrate the performance of this naive temporal sampling strategy using medium (25) or dense (100) temporal resolutions on 100-frame input videos. The results show that increasing the temporal resolution universally does not necessarily yield better performance. In contrast, our proposed method, which utilizes motion-aware temporal sampling, achieves the best results. We surmise that this may be attributed to the model distributing the volume's capacity sparsely across a large number of time steps, which is not the case for our method. + +Motion-aware Spatial Sampling: To validate the effectiveness of our motion-aware spatial sampling strategy, we compare it against other sampling strategies. The first variant involves uniform sampling along the ray, and the second uses a sample prediction network (SPN) like Attal et al. [1] for sampling. These variants sample 128 points on each ray. Results in Table 3 show that our full model employing motion-aware spatial sampling outperforms these variants, perhaps by having a better sense of where to sample more from, derived from its semantic-aware embedding. + +SAM embedding: To verify that introducing SAM embedding can improve rendering quality (rather than just enabling segmentation or tracking), we tested a variant of Gear-NeRF without SAM embedding (it instead thresholds loss maps of rendered RGB frames to obtain upshift masks). As shown in Table 3, the absence of SAM em + +Table 3. Ablation Study: Ablations on our spatio-temporal sampling strategy and the number of gears (Truck/Google Immersive). Best and second best results are highlighted. + +
MethodPSNR (↑)SSIM (↑)LPIPS (↓)
Naive temporal sampling (medium)26.930.7780.144
Naive temporal sampling (dense)26.850.7600.161
Naive spatial sampling24.800.7340.222
SPN spatial sampling26.540.7870.162
Ours (w/o embedding)27.100.8310.177
Ours (Ngear = 2)26.930.7850.166
Ours (Ngear = 3)27.140.8750.145
Ours (Ngear = 4)27.490.8920.136
Ours (Ngear = 5)27.460.9010.131
+ +bedding for guiding gear assignment reduces the model's rendering quality. + +Number of Gears: We ablate on the numbers of gears. Increasing the number of gears allows for more fine-grained motion-aware spatio-temporal sampling, while increasing the computational cost. As shown in Table 3, a choice of up to 4 gear levels seems optimal, further increasing the number of gears does not result in significant improvements. + +# 6. Conclusions + +In this work, we introduced Gear-NeRF, an extension of dynamic NeRFs that leverages semantic information from powerful segmentation models for stratified modeling of dynamic scenes. Our approach learns a 4D (spatiotemporal) semantic embedding and introduces the concept of "gears" for differentiated modeling of scene regions based on their motion intensity. With determined gear assignments, Gear-NeRF adaptively adjusts its spatiotemporal sampling resolution to improve the photo-realism of rendered views. At the same time, Gear-NeRF provides the new functionality of free-viewpoint object tracking with user prompts as simple as a click. Our empirical studies underscore the effectiveness of Gear-NeRF, showcasing state-of-the-art performance in both rendering quality and object tracking across multiple challenging datasets. + +# References + +[1] Benjamin Attal, Jia-Bin Huang, Christian Richardt, Michael Zollhoefer, Johannes Kopf, Matthew O'Toole, and Changil Kim. HyperReel: High-fidelity 6-DoF video with ray-conditioned sampling. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 3, 6, 7, 8 +[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 5855–5864, 2021. 2 +[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5470-5479, 2022. +[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 19697-19705, 2023. 2 +[5] Michael Broxton, John Flynn, Ryan Overbeck, Daniel Erickson, Peter Hedman, Matthew Duvall, Jason Dourgarian, Jay Busch, Matt Whalen, and Paul Debevec. Immersive light field video with a layered mesh representation. ACM Transactions on Graphics (TOG), 39(4):86-1, 2020. 6, 7, 8 +[6] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 130-141, 2023. 1, 2, 3, 7 +[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers, 2021. 3 +[8] Jiazhong Cen, Zanwei Zhou, Jiemin Fang, Chen Yang, Wei Shen, Lingxi Xie, Dongsheng Jiang, Xiaopeng Zhang, and Qi Tian. Segment anything in 3d with nerfs, 2023. 3, 7, 8 +[9] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5799-5809, 2021. 2 +[10] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 2 +[11] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision (ECCV), pages 333-350. Springer, 2022. 2 +[12] Anpei Chen, Zexiang Xu, Xinyue Wei, Siyu Tang, Hao Su, and Andreas Geiger. Factor fields: A unified framework for neural fields and beyond. arXiv preprint arXiv:2302.01226, 2023. 2 +[13] Jiaben Chen and Huaizu Jiang. Sportsslomo: A new benchmark and baselines for human-centric video frame interpo + +lation. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3 +[14] Jiaben Chen, Renrui Zhang, Dongze Lian, Jiaqi Yang, Ziyao Zeng, and Jianbo Shi. iquery: Instruments as queries for audio-visual sound separation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14675–14686, 2023. 3 +[15] Xiaokang Chen, Kwan-Yee Lin, Chen Qian, Gang Zeng, and Hongsheng Li. 3d sketch-aware semantic scene completion via semi-supervised structure prior. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3 +[16] Xiaokang Chen, Jiaxiang Tang, Diwen Wan, Jingbo Wang, and Gang Zeng. Interactive segment anything nef with feature imitation. arXiv preprint arXiv:2305.16233, 2023. 3 +[17] Yue Chen, Xuan Wang, Xingyu Chen, Qi Zhang, Xiaoyu Li, Yu Guo, Jue Wang, and Fei Wang. Uv volumes for real-time rendering of editable free-view human performance. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16621-16631, 2023. 2 +[18] Zhang Chen, Zhong Li, Liangchen Song, Lele Chen, Jingyi Yu, Junsong Yuan, and Yi Xu. Neurbf: A neural fields representation with adaptive radial basis functions. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 4182-4194, 2023. 2 +[19] Ho Kei Cheng and Alexander G. Schwing. Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model, 2022. 3 +[20] Yangming Cheng, Liulei Li, Yuanyou Xu, Xiaodi Li, Zongxin Yang, Wenguan Wang, and Yi Yang. Segment and track anything. arXiv preprint arXiv:2305.06558, 2023. 3 +[21] Alvaro Collet, Ming Chuang, Pat Sweeney, Don Gillett, Dennis Evseev, David Calabrese, Hugues Hoppe, Adam Kirk, and Steve Sullivan. High-quality streamable free-viewpoint video. ACM Transactions on Graphics (TOG), 34(4):1-13, 2015. 1 +[22] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 14304-14314. IEEE Computer Society, 2021. 1, 2 +[23] Zhiwen Fan, Peihao Wang, Xinyu Gong, Yifan Jiang, Dejia Xu, and Zhangyang Wang. Nerf-sos: Any-view self-supervised object segmentation from complex real-world scenes. International Conference on Learning Representations (ICLR), pages arXiv-2209, 2023. 3 +[24] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. In ACM Transactions on Graphics (SIGGRAPH Asia), pages 1-9, 2022. 2 +[25] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5501–5510, 2022. 2 +[26] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In IEEE/CVF Conference on Computer Vision and + +Pattern Recognition (CVPR), pages 12479-12488, 2023. 1, 2, 3 +[27] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 5712-5721, 2021. 2 +[28] Chen Geng, Sida Peng, Zhen Xu, Hujun Bao, and Xiaowei Zhou. Learning neural volumetric representations of dynamic humans in minutes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8759-8770, 2023. 2 +[29] Rahul Goel, Dhawal Sirikonda, Saurabh Saini, and P.J. Narayanan. Interactive Segmentation of Radiance Fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3 +[30] Yuan-Chen Guo, Di Kang, Linchao Bao, Yu He, and Song-Hai Zhang. Nerfren: Neural radiance fields with reflections. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18409-18418, 2022. 2 +[31] Wenbo Hu, Yuling Wang, Lin Ma, Bangbang Yang, Lin Gao, Xiao Liu, and Yuewen Ma. Tri-miprf: Tri-mip representation for efficient anti-aliasing neural radiance fields. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 19774–19783, 2023. 2 +[32] Qiangui Huang, Weiyue Wang, and Ulrich Neumann. Recurrent slice networks for 3d segmentation on point clouds. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3 +[33] Mustafa Isik, Martin Rünz, Markos Georgopoulos, Taras Khakhulin, Jonathan Starck, Lourdes Agapito, and Matthias Nießner. Humanrf: High-fidelity neural radiance fields for humans in motion. ACM Transactions on Graphics (SIGGRAPH), 2023. 2 +[34] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 5885-5894, 2021. 2 +[35] Lei Ke, Mingqiao Ye, Martin Danelljan, Yifan Liu, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Segment anything in high quality. Advances in Neural Information Processing Systems (NeurIPS), 2023. 3 +[36] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (TOG), 42(4):1-14, 2023. 2 +[37] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. In IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 3 +[38] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2, 3, 4, 5 +[39] Tobias Kirschstein, Shenhan Qian, Simon Giebenhain, Tim Walter, and Matthias Nießner. Nersemble: Multi-view radiance field reconstruction of human heads. ACM Transactions on Graphics (SIGGRAPH), 2023. 2 +[40] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nerf for editing via feature field distil + +lation. Advances in Neural Information Processing Systems (NeurIPS), 35:23311-23330, 2022. 2, 3 +[41] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and Rene Ranftl. Language-driven semantic segmentation. In International Conference on Learning Representations (ICLR), 2022. 3 +[42] Hao Li, Bart Adams, Leonidas J Guibas, and Mark Pauly. Robust single-view geometry and motion reconstruction. ACM Transactions on Graphics (TOG), 28(5):1-10, 2009. 1 +[43] Hao Li, Linjie Luo, Daniel Vlasic, Pieter Peers, Jovan Popovic, Mark Pauly, and Szymon Rusinkiewicz. Temporally coherent completion of dynamic shapes. ACM Transactions on Graphics (TOG), 31(1):1-11, 2012. 1 +[44] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5521-5531, 2022. 2, 6, 7, 8 +[45] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021. 1, 2 +[46] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4273-4284, 2023. 2 +[47] Haotong Lin, Sida Peng, Zhen Xu, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Efficient neural radiance fields for interactive free-viewpoint video. In ACM Transactions on Graphics (SIGGRAPH Asia), pages 1-9, 2022. +[48] Haotong Lin, Sida Peng, Zhen Xu, Tao Xie, Xingyi He, Hujun Bao, and Xiaowei Zhou. Im4d: High-fidelity and real-time novel view synthesis for dynamic scenes. ACM Transactions on Graphics (SIGGRAPH Asia), 2023. 2 +[49] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 9298–9309, 2023. 2 +[50] Xinhang Liu, Jiaben Chen, Huai Yu, Yu-Wing Tai, and Chi-Keung Tang. Unsupervised multi-view object segmentation using radiance field propagation. Advances in Neural Information Processing Systems (NeurIPS), 35:17730-17743, 2022. 2, 3 +[51] Xinhang Liu, Shiu-hong Kao, Jiaben Chen, Yu-Wing Tai, and Chi-Keung Tang. Deceptive-nerf: Enhancing nerf reconstruction using pseudo-observations from diffusion models. arXiv preprint arXiv:2305.15171, 2023. 2 +[52] Xinhang Liu, Yu-Wing Tai, and Chi-Keung Tang. Clean-nerf: Reformulating nerf to account for view-dependent observations. arXiv preprint arXiv:2303.14707, 2023. 2 +[53] Yu-Lun Liu, Chen Gao, Andreas Meuleman, Hung-Yu Tseng, Ayush Saraf, Changil Kim, Yung-Yu Chuang, Johannes Kopf, and Jia-Bin Huang. Robust dynamic radiance fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13-23, 2023. 2 + +[54] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 2 +[55] Andreas Meuleman, Yu-Lun Liu, Chen Gao, Jia-Bin Huang, Changil Kim, Min H Kim, and Johannes Kopf. Progressively optimized local radiance fields for robust view synthesis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16539-16548, 2023. 2 +[56] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision (ECCV), 2020. 1, 2, 3 +[57] Ashkan Mirzaei, Tristan Aumentado-Armstrong, Konstantinos G Derpanis, Jonathan Kelly, Marcus A Brubaker, Igor Gilitschenski, and Alex Levinshtein. Spin-nerf: Multiview segmentation and perceptual inpainting with neural radiance fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20669–20679, 2023. 3 +[58] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (TOG), 41(4):1-15, 2022. 2 +[59] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5480-5490, 2022. 2 +[60] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 5865-5874, 2021. 1, 2 +[61] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Transactions on Graphics (TOG), 2021. 1, 2 +[62] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in Neural Information Processing Systems (NeurIPS), 32, 2019. 6 +[63] Sida Peng, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Representing volumetric videos as dynamic mlp maps. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4252-4262, 2023. 2 +[64] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. International Conference on Learning Representations (ICLR), 2023. 2 +[65] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields + +for dynamic scenes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10318-10327, 2021. 1, 2 +[66] Yi-Ling Qiao, Alexander Gao, Yiran Xu, Yue Feng, Jia-Bin Huang, and Ming C Lin. Dynamic mesh-aware radiance fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 385-396, 2023. 2 +[67] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of Machine Learning Research (PMLR), pages 8748-8763, 2021. 3 +[68] Konstantinos Rematas, Andrew Liu, Pratul P Srinivasan, Jonathan T Barron, Andrea Tagliasacchi, Thomas Funkhouser, and Vittorio Ferrari. Urban radiance fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12932-12942, 2022. 2 +[69] Zhongzheng Ren, Aseem Agarwala, Bryan Russell, Alexander G. Schwing, and Oliver Wang. Neural volumetric object selection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3 +[70] Neus Sabater, Guillaume Boisson, Benoit Vandame, Paul Kerbiriou, Frederic Babon, Matthieu Hog, Remy Gendrot, Tristan Langlois, Olivier Bureller, Arno Schubert, et al. Dataset and pipeline for multi-view light-field video. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 30-40, 2017. 6, 7, 8 +[71] Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boning Liu, Hongwen Zhang, and Yebin Liu. Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16632-16642, 2023. 2, 3 +[72] Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulò, Norman Müller, Matthias Nießner, Angela Dai, and Peter Kontschieder. Panoptic lifting for 3d scene understanding with neural fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9043-9052, 2023. 3 +[73] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerf-player: A streamable dynamic scene representation with decomposed neural radiance fields. IEEE Transactions on Visualization and Computer Graphics (TVCG), 29(5): 2732-2742, 2023. 2 +[74] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5459-5469, 2022. 2 +[75] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8248–8258, 2022. 2 +[76] Jiaxiang Tang, Xiaokang Chen, Jingbo Wang, and Gang Zeng. Point scene understanding via disentangled instance + +mesh reconstruction. European Conference on Computer Vision (ECCV), 2022. 3 +[77] Jonathan Taylor, Jamie Shotton, Toby Sharp, and Andrew Fitzgibbon. The vitruvian manifold: Inferring dense correspondences for one-shot human pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 103-110. IEEE, 2012. 1 +[78] Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srinivasan, Edgar Tretschk, Wang Yifan, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lombardi, et al. Advances in neural rendering. In Computer Graphics Forum, pages 703-735. Wiley Online Library, 2022. 2 +[79] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 12959-12970, 2021. 2 +[80] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481-5490. IEEE, 2022. 2 +[81] Feng Wang, Sinan Tan, Xinghang Li, Zeyue Tian, and Huaping Liu. Mixed neural voxels for fast multi-view video synthesis. IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2, 7 +[82] Liao Wang, Jiakai Zhang, Xinhang Liu, Fuqiang Zhao, Yanshun Zhang, Yingliang Zhang, Minye Wu, Jingyi Yu, and Lan Xu. Fourier plenoctrees for dynamic radiance field rendering in real-time. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13524-13534, 2022. 2 +[83] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing (TIP), 13(4):600-612, 2004. 6 +[84] Ziyan Wang, Giljoo Nam, Tuur Stuyck, Stephen Lombardi, Chen Cao, Jason Saragih, Michael Zollhöfer, Jessica Hodgins, and Christoph Lassner. Neuwigs: A neural dynamic model for volumetric hair capture and animation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8641-8651, 2023. 2 +[85] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528, 2023. 2 +[86] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P Srinivasan, Dor Verbin, Jonathan T Barron, Ben Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. arXiv preprint arXiv:2312.02981, 2023. 2 +[87] Jamie Wynn and Daniyar Turmukhambetov. Diffusionerf: Regularizing neural radiance fields with denoising diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4180-4189, 2023. 2 +[88] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint + +video. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9421-9431, 2021. 2 +[89] Zhen Xu, Sida Peng, Haotong Lin, Guangzhao He, Jiaming Sun, Yujun Shen, Hujun Bao, and Xiaowei Zhou. 4k4d: Real-time 4d view synthesis at 4k resolution. arXiv preprint arXiv:2310.11448, 2023. 2 +[90] Zhiwen Yan, Chen Li, and Gim Hee Lee. Nerf-ds: Neural radiance fields for dynamic specular objects. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8285-8295, 2023. 2 +[91] Bo Yang, Jianan Wang, Ronald Clark, Qingyong Hu, Sen Wang, Andrew Markham, and Niki Trigoni. Learning object bounding boxes for 3d instance segmentation on point clouds, 2019. 3 +[92] Jinyu Yang, Mingqi Gao, Zhe Li, Shang Gao, Fangjing Wang, and Feng Zheng. Track anything: Segment anything meets videos, 2023. 3 +[93] Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Improving few-shot neural rendering with free frequency regularization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8254-8263, 2023. 2 +[94] Zongxin Yang and Yi Yang. Decoupling features in hierarchical propagation for video object segmentation, 2022. 3 +[95] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101, 2023. 2 +[96] Zeyu Yang, Hongye Yang, Zijie Pan, Xiatian Zhu, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. arXiv preprint arXiv:2310.10642, 2023. 2 +[97] Zhengming Yu, Wei Cheng, Xian Liu, Wayne Wu, and Kwan-Yee Lin. Monohuman: Animatable human neural field from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16943-16953, 2023. 2 +[98] Wentao Yuan, Zhaoyang Lv, Tanner Schmidt, and Steven Lovegrove. Star: Self-supervised tracking and reconstruction of rigid objects in motion with neural rendering. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13144-13152, 2021. 2 +[99] Jiakai Zhang, Xinhang Liu, Xinyi Ye, Fuqiang Zhao, Yanshun Zhang, Minye Wu, Yingliang Zhang, Lan Xu, and Jingyi Yu. Editable free-viewpoint video using a layered neural representation. ACM Transactions on Graphics (TOG), 40(4):1-18, 2021. 1, 2, 7 +[100] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 6 +[101] Renrui Zhang, Zhengkai Jiang, Ziyu Guo, Shilin Yan, Junting Pan, Hao Dong, Peng Gao, and Hongsheng Li. Personalize segment anything model with one shot. arXiv preprint arXiv:2305.03048, 2023. 3 +[102] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J Davison. In-place scene labelling and understanding with implicit scene representation. In IEEE/CVF + +International Conference on Computer Vision (ICCV), pages 15838-15847, 2021. 2, 3 \ No newline at end of file diff --git a/2024/Gear-NeRF_ Free-Viewpoint Rendering and Tracking with Motion-aware Spatio-Temporal Sampling/images.zip b/2024/Gear-NeRF_ Free-Viewpoint Rendering and Tracking with Motion-aware Spatio-Temporal Sampling/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6174e42acf11eead54973b9e33d66f2d5aafc058 --- /dev/null +++ b/2024/Gear-NeRF_ Free-Viewpoint Rendering and Tracking with Motion-aware Spatio-Temporal Sampling/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5a3d21affc4bc2549339c8321b43d4154eaf51773785bea64123af95b46f7eef +size 729108 diff --git a/2024/Gear-NeRF_ Free-Viewpoint Rendering and Tracking with Motion-aware Spatio-Temporal Sampling/layout.json b/2024/Gear-NeRF_ Free-Viewpoint Rendering and Tracking with Motion-aware Spatio-Temporal Sampling/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7076ca4010f96f3db7df28f2d89d97fb07e282e3 --- /dev/null +++ b/2024/Gear-NeRF_ Free-Viewpoint Rendering and Tracking with Motion-aware Spatio-Temporal Sampling/layout.json @@ -0,0 +1,12825 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 131, + 88, + 463, + 125 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 88, + 463, + 125 + ], + "spans": [ + { + "bbox": [ + 131, + 88, + 463, + 125 + ], + "type": "text", + "content": "Gear-NeRF: Free-Viewpoint Rendering and Tracking with Motion-aware Spatio-Temporal Sampling" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 85, + 146, + 509, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 146, + 509, + 217 + ], + "spans": [ + { + "bbox": [ + 85, + 146, + 509, + 217 + ], + "type": "text", + "content": "Xinhang Liu" + }, + { + "bbox": [ + 85, + 146, + 509, + 217 + ], + "type": "inline_equation", + "content": "^{1\\dagger*}" + }, + { + "bbox": [ + 85, + 146, + 509, + 217 + ], + "type": "text", + "content": " Yu-Wing Tai" + }, + { + "bbox": [ + 85, + 146, + 509, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 85, + 146, + 509, + 217 + ], + "type": "text", + "content": " Chi-Keung Tang" + }, + { + "bbox": [ + 85, + 146, + 509, + 217 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 85, + 146, + 509, + 217 + ], + "type": "text", + "content": " Pedro Miraldo" + }, + { + "bbox": [ + 85, + 146, + 509, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 85, + 146, + 509, + 217 + ], + "type": "text", + "content": " Suhas Lohit" + }, + { + "bbox": [ + 85, + 146, + 509, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 85, + 146, + 509, + 217 + ], + "type": "text", + "content": " Moitreya Chatterjee" + }, + { + "bbox": [ + 85, + 146, + 509, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 85, + 146, + 509, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 85, + 146, + 509, + 217 + ], + "type": "text", + "content": "HKUST " + }, + { + "bbox": [ + 85, + 146, + 509, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 85, + 146, + 509, + 217 + ], + "type": "text", + "content": "Dartmouth College " + }, + { + "bbox": [ + 85, + 146, + 509, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 85, + 146, + 509, + 217 + ], + "type": "text", + "content": "Mitsubishi Electric Research Laboratories (MERL) xliufe@connect.ust.hk, yu-wing.tai@dartmouth.edu, cktang@cse.ust.hk, miraldo@merl.com, slohit@merl.com, metro.smiles@gmail.com" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 244, + 192, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 244, + 192, + 257 + ], + "spans": [ + { + "bbox": [ + 143, + 244, + 192, + 257 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 270, + 290, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 290, + 571 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 290, + 571 + ], + "type": "text", + "content": "Extensions of Neural Radiance Fields (NeRFs) to model dynamic scenes have enabled their near photo-realistic, free-viewpoint rendering. Although these methods have shown some potential in creating immersive experiences, two drawbacks limit their ubiquity: (i) a significant reduction in reconstruction quality when the computing budget is limited, and (ii) a lack of semantic understanding of the underlying scenes. To address these issues, we introduce Gear-NeRF, which leverages semantic information from powerful image segmentation models. Our approach presents a principled way for learning a spatio-temporal (4D) semantic embedding, based on which we introduce the concept of gears to allow for stratified modeling of dynamic regions of the scene based on the extent of their motion. Such differentiation allows us to adjust the spatiotemporal sampling resolution for each region in proportion to its motion scale, achieving more photo-realistic dynamic novel view synthesis. At the same time, almost for free, our approach enables free-viewpoint tracking of objects of interest - a functionality not yet achieved by existing NeRF-based methods. Empirical studies validate the effectiveness of our method, where we achieve state-of-the-art rendering and tracking performance on multiple challenging datasets. The project page is available at: https://merl.com/research/highlights/gear-nerf." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 595, + 128, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 595, + 128, + 608 + ], + "spans": [ + { + "bbox": [ + 47, + 595, + 128, + 608 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 616, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 616, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 616, + 287, + 689 + ], + "type": "text", + "content": "Reconstructing 3D scenes has a broad range of applications, including Virtual Reality/Augmented Reality (VR/AR), 3D animation, game production, and film creation which allow users to observe scenes from any desired viewpoint. While it is crucial to reconstruct static scenes, towards which significant progress has been made, it is even more crucial to" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 308, + 245, + 545, + 403 + ], + "blocks": [ + { + "bbox": [ + 308, + 245, + 545, + 403 + ], + "lines": [ + { + "bbox": [ + 308, + 245, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 245, + 545, + 403 + ], + "type": "image", + "image_path": "388e1c857a088b1e69eef4f1773783926afe85c265cffcd81b0d9b12d5e592d1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 413, + 547, + 468 + ], + "lines": [ + { + "bbox": [ + 305, + 413, + 547, + 468 + ], + "spans": [ + { + "bbox": [ + 305, + 413, + 547, + 468 + ], + "type": "text", + "content": "Figure 1. (a) Our method takes RGB videos captured from a camera array as input. (b) Trained Gear-NeRF achieves photo-realistic real-time free-viewpoint rendering of a dynamic scene. (c) With users giving a single click at any time and from any viewpoint, our method can perform free-viewpoint tracking of the target object." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 477, + 546, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 477, + 546, + 512 + ], + "spans": [ + { + "bbox": [ + 305, + 477, + 546, + 512 + ], + "type": "text", + "content": "reconstruct dynamic scenes, as the world around us is characterized by a constant state of flux, with many objects in it - in a state of motion." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 514, + 547, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 514, + 547, + 681 + ], + "spans": [ + { + "bbox": [ + 304, + 514, + 547, + 681 + ], + "type": "text", + "content": "Recent advances in novel view synthesis, such as Neural Radiance Fields (NeRFs) [56] have inspired numerous studies to extend them to dynamic 3D scenes. Existing approaches either employ a deformation field to map neural fields from a given time to a canonical space [22, 45, 60, 61, 65, 99], or directly model dynamic scenes as a 4D space-time grid [1, 6, 26]. Though these methods offer improved rendering quality by utilizing more accessible inputs compared to previous solutions [21, 42, 43, 77], they still struggle to ensure rendering quality in low-resource settings, requiring carefully engineered efforts. Further, most dynamic radiance field approaches adopt a naive spatiotemporal sampling strategy, without discerning the different scales of motion across different regions in the scene." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 682, + 547, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 682, + 547, + 730 + ], + "spans": [ + { + "bbox": [ + 305, + 682, + 547, + 730 + ], + "type": "text", + "content": "We propose to fix this issue by leveraging a semantic understanding of dynamic scenes. Intuitively, a reconstruction system aware of the distinction between static and dynamic regions in a scene can perform more focused" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 700, + 234, + 709 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 700, + 234, + 709 + ], + "spans": [ + { + "bbox": [ + 58, + 700, + 234, + 709 + ], + "type": "text", + "content": "*Work mainly done when XL was an intern at MERL." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 710, + 287, + 730 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 710, + 287, + 730 + ], + "spans": [ + { + "bbox": [ + 47, + 710, + 287, + 730 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 47, + 710, + 287, + 730 + ], + "type": "text", + "content": " XL and CT are supported in part by the Research Grant Council of the Hong Kong SAR grant no. 16201420." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 757, + 317, + 765 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 757, + 317, + 765 + ], + "spans": [ + { + "bbox": [ + 294, + 757, + 317, + 765 + ], + "type": "text", + "content": "19667" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 61, + 547, + 238 + ], + "blocks": [ + { + "bbox": [ + 50, + 61, + 547, + 238 + ], + "lines": [ + { + "bbox": [ + 50, + 61, + 547, + 238 + ], + "spans": [ + { + "bbox": [ + 50, + 61, + 547, + 238 + ], + "type": "image", + "image_path": "500abea19269eadef36d0ba7557bc09d847c481cb0d319ca0167698cf0cf9a5a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 246, + 548, + 302 + ], + "lines": [ + { + "bbox": [ + 46, + 246, + 548, + 302 + ], + "spans": [ + { + "bbox": [ + 46, + 246, + 548, + 302 + ], + "type": "text", + "content": "Figure 2. Pipeline of Gear-NeRF: Gear-NeRF takes multi-view videos as input. After optimizing the serial 4D feature volumes (Section 4.1), it maps space-time coordinates to a 4D semantic embedding (Section 4.2), in addition to the volume density and view-dependent radiance color. Regions with larger motion are automatically assigned higher gear levels (Section 4.3) and as a result, receive higher-resolution spatio-temporal sampling (Section 4.4). Furthermore, Gear-NeRF is capable of performing free-viewpoint tracking of a target object with prompts as simple as a user click (Section 4.5)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 308, + 289, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 289, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 289, + 548 + ], + "type": "text", + "content": "sampling in the dynamic regions, which inherently require more resources per unit volume than static regions, due to their time-evolving nature. Accordingly, dynamic regions can be further stratified according to their scale of motion. To this end, this paper presents Gear-NeRF, a framework that leverages semantic embedding from powerful image segmentation models [38] for stratified modeling of 4D scenes. Gear-NeRF optimizes for a 4D semantic embedding, based on which we introduce the concept of gear to smartly determine the appropriate region-specific resolution of spatio-temporal sampling in the NeRF. Regions with larger motion scales are assigned higher gears, through our gear determination scheme and we accordingly perform higher-resolution spatio-temporal sampling. Empirical studies reveal that this motion-aware sampling strategy improves the quality of synthesized images, over competing approaches. As a by-product of our semantically embedded representation, we achieve free-viewpoint object tracking, given user prompts. Figure 1 presents an overview of the capabilities of our method." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 549, + 287, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 549, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 46, + 549, + 287, + 657 + ], + "type": "text", + "content": "Gear-NeRF makes two primary advancements: (i) enhanced dynamic novel view synthesis by resorting to smarter spatio-temporal sampling, and (ii) the ability for free-viewpoint tracking of objects of interest. The latter is a capability not yet realized by existing NeRF methods for dynamic scenes. We perform extensive experiments on multiple datasets to validate the generalizability and robustness of our method, which shows state-of-the-art performances for both tasks across all datasets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 672, + 134, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 672, + 134, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 672, + 134, + 685 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 693, + 287, + 731 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 693, + 287, + 731 + ], + "spans": [ + { + "bbox": [ + 46, + 693, + 287, + 731 + ], + "type": "text", + "content": "Neural Radiance Fields: NeRF [56] is a recent breakthrough among novel view synthesis methods that uses multilayer perceptrons (MLPs) to parameterize the appearance" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 308, + 547, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 308, + 547, + 511 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 547, + 511 + ], + "type": "text", + "content": "and density for each point in 3D space, given any viewing direction of the scene. Researchers have extended NeRF along various dimensions [78], including improving rendering quality [2-4, 12, 18, 31], handling challenging conditions such as large scenes [55, 68, 75], view-dependent appearances [30, 52, 80], and sparse inputs [34, 51, 59, 86, 87, 93]. NeRF-like neural representations have also found applications in semantic segmentation [40, 50, 102] and 3D content generation [9, 10, 49, 64]. Recent work has shown that replacing the deep MLPs with a feature voxel grid can significantly improve training and inference speed [11, 25, 58, 74]. On the other hand, a more recent approach to further improve visual quality, rendering time, and performance entails representing the scene with 3D Gaussians [36]. Our approach, while drawing upon many of these approaches, deals with dynamic scenes which is beyond the scope of these methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 514, + 548, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 514, + 548, + 730 + ], + "spans": [ + { + "bbox": [ + 304, + 514, + 548, + 730 + ], + "type": "text", + "content": "Neural Representations for Dynamic Scenes: NeRF-like representations have recently been extended to model dynamic scenes in high fidelity [17, 28, 33, 39, 46-48, 53, 63, 66, 84, 89, 90, 97]. One straightforward approach to do this is to directly condition the radiance field on time [27, 44, 45, 88]. Alternatively, several methods model a deformation field to map coordinates from different time stamps to a common canonical space [22, 24, 60, 61, 65, 79, 98, 99]. Some recent approaches [6, 26, 71, 81] represent the scene using a 4D space-time grid, which is decomposed into sets of planar representations for training efficiency. Other techniques for improving rendering fidelity and frame rate include Fourier PlenOctrees [82], ray-conditioned sample prediction networks [1], 4D space decomposition (static/dynamic/newly appeared regions) [73], and explicit voxel grids [24]. 3D Gaussians have also been adapted to model dynamic scenes [54, 85, 95, 96]. While these approaches paved the initial path for rendering dy" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 757, + 317, + 765 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 757, + 317, + 765 + ], + "spans": [ + { + "bbox": [ + 294, + 757, + 317, + 765 + ], + "type": "text", + "content": "19668" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 58, + 287, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 58, + 287, + 261 + ], + "spans": [ + { + "bbox": [ + 46, + 58, + 287, + 261 + ], + "type": "text", + "content": "namic scenes, semantically aware modeling of the scene is absent, a caveat that our proposed method seeks to address. Segment Anything Model (SAM): SAM [38] is a powerful promptable image segmentation model, which showcases remarkable zero-shot generalization abilities and can produce semantically consistent masks, given a single foreground point on the image. HQ-SAM [35] is an improvement on SAM that enhances the quality of masks, especially on objects with intricate boundaries and structures. Recent works [20, 92] have extended SAM to perform interactive video object segmentation. These methods utilize SAM for mask initialization or correction and then employ state-of-the-art mask trackers [19, 94] for mask tracking and prediction [14]. Recent methods have also leveraged SAM for tracking multiple reference objects in a video [13, 101]. This work uses SAM to profile the scene into semantic regions, which are then grouped based on motion scales." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 262, + 288, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 262, + 288, + 478 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 288, + 478 + ], + "type": "text", + "content": "3D Semantic Understanding: Existing methods for 3D visual understanding [15, 32, 76, 91] mainly focus on closed set segmentation of point clouds or voxels. NeRF's ability to integrate information across multiple views has led to its applications in 3D semantic segmentation [102], object segmentation [23, 50, 57], panoptic segmentation [72], and interactive segmentation [29, 69]. Kobayashi et al. [40] explored the effectiveness of embedding pixel-aligned features [7, 41] into NeRFs for 3D manipulations. LERF [37] fuses CLIP embeddings [67] and NeRFs to enable language-based localization in 3D NeRF scenes. Recent work has enabled click/text-based 3D segmentation by learning a 3D SAM embedding [16] or inverse rendering of SAM-generated masks [8]. We, on the other hand, seek to utilize the synergy of dynamic NeRFs and SAM segments to derive a semantic understanding of a dynamic 3D scene for tracking objects of interest in novel views – a first of its kind effort." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 492, + 132, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 492, + 132, + 504 + ], + "spans": [ + { + "bbox": [ + 47, + 492, + 132, + 504 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 513, + 287, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 513, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 46, + 513, + 287, + 609 + ], + "type": "text", + "content": "Neural Radiance Fields (NeRFs): Vanilla NeRFs [56] employ a multi-layer perceptron (MLP) with sinusoidal positional encoding to map a 3D-spatial coordinate " + }, + { + "bbox": [ + 46, + 513, + 287, + 609 + ], + "type": "inline_equation", + "content": "\\mathbf{x} = (x, y, z)" + }, + { + "bbox": [ + 46, + 513, + 287, + 609 + ], + "type": "text", + "content": " and a viewing direction " + }, + { + "bbox": [ + 46, + 513, + 287, + 609 + ], + "type": "inline_equation", + "content": "\\mathbf{d} = (\\theta, \\phi)" + }, + { + "bbox": [ + 46, + 513, + 287, + 609 + ], + "type": "text", + "content": " to a volume density " + }, + { + "bbox": [ + 46, + 513, + 287, + 609 + ], + "type": "inline_equation", + "content": "\\sigma \\in [0,1]" + }, + { + "bbox": [ + 46, + 513, + 287, + 609 + ], + "type": "text", + "content": " and an emitted RGB, " + }, + { + "bbox": [ + 46, + 513, + 287, + 609 + ], + "type": "inline_equation", + "content": "\\mathbf{c} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 513, + 287, + 609 + ], + "type": "text", + "content": ". Rendering each image pixel involves casting a ray " + }, + { + "bbox": [ + 46, + 513, + 287, + 609 + ], + "type": "inline_equation", + "content": "\\mathbf{r}(t) = \\mathbf{o} + t\\mathbf{d}" + }, + { + "bbox": [ + 46, + 513, + 287, + 609 + ], + "type": "text", + "content": " from the camera center " + }, + { + "bbox": [ + 46, + 513, + 287, + 609 + ], + "type": "inline_equation", + "content": "\\mathbf{o}" + }, + { + "bbox": [ + 46, + 513, + 287, + 609 + ], + "type": "text", + "content": " through the pixel along direction " + }, + { + "bbox": [ + 46, + 513, + 287, + 609 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 46, + 513, + 287, + 609 + ], + "type": "text", + "content": ". The predicted color for the corresponding pixel is computed as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 125, + 620, + 287, + 652 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 620, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 125, + 620, + 287, + 652 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {C}} (\\mathbf {r}) = \\sum_ {i = 1} ^ {N} T _ {i} \\alpha_ {i} \\mathbf {c} _ {i}, \\tag {1}", + "image_path": "0420f9ce66a35f11f2d04966adaec7f72d17eeb2d610eb23fc2864e681c8d005.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 664, + 287, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 664, + 287, + 730 + ], + "spans": [ + { + "bbox": [ + 46, + 664, + 287, + 730 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 664, + 287, + 730 + ], + "type": "inline_equation", + "content": "T_{i} = \\exp \\left(-\\sum_{j=1}^{i-1} \\sigma_{j} \\delta_{j}\\right)" + }, + { + "bbox": [ + 46, + 664, + 287, + 730 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 664, + 287, + 730 + ], + "type": "inline_equation", + "content": "\\alpha_{i} = 1 - \\exp (-\\sigma_{i} \\delta_{i})" + }, + { + "bbox": [ + 46, + 664, + 287, + 730 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 664, + 287, + 730 + ], + "type": "inline_equation", + "content": "\\delta_{j} = t_{j+1} - t_{j}" + }, + { + "bbox": [ + 46, + 664, + 287, + 730 + ], + "type": "text", + "content": ". A vanilla NeRF is trained by minimizing the mean squared error between the input images and the predicted images, obtained by rendering the scene from the viewpoints from which the input images have been" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 58, + 472, + 70 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 58, + 472, + 70 + ], + "spans": [ + { + "bbox": [ + 306, + 58, + 472, + 70 + ], + "type": "text", + "content": "captured, with the training loss given by:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 365, + 79, + 545, + 105 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 79, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 365, + 79, + 545, + 105 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {p h o}} = \\sum_ {\\mathbf {r} \\in \\mathcal {R}} \\| \\hat {\\mathbf {C}} (\\mathbf {r}) - \\mathbf {C} (\\mathbf {r}) \\| _ {2} ^ {2}. \\tag {2}", + "image_path": "33dcb31fed00dd4cdf6787c672ae3ea314aff70596fb4d1df165313c2e4114d0.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 114, + 545, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 114, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 304, + 114, + 545, + 258 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 114, + 545, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 304, + 114, + 545, + 258 + ], + "type": "text", + "content": " is the set of all rays projected from the input image. Planar-Factorized 4D Volumes: A recent emerging trend of handling dynamics using radiance field representations is to directly adapt them to be conditioned on a frame index " + }, + { + "bbox": [ + 304, + 114, + 545, + 258 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 114, + 545, + 258 + ], + "type": "text", + "content": " (denoting time) in addition to " + }, + { + "bbox": [ + 304, + 114, + 545, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 114, + 545, + 258 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 114, + 545, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 304, + 114, + 545, + 258 + ], + "type": "text", + "content": ". This can be accomplished by learning a mapping from " + }, + { + "bbox": [ + 304, + 114, + 545, + 258 + ], + "type": "inline_equation", + "content": "(\\mathbf{x},\\mathbf{d},t)" + }, + { + "bbox": [ + 304, + 114, + 545, + 258 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 114, + 545, + 258 + ], + "type": "inline_equation", + "content": "(\\sigma ,\\mathbf{c})" + }, + { + "bbox": [ + 304, + 114, + 545, + 258 + ], + "type": "text", + "content": " using planar-factorized 4D volumes [1, 6, 26, 71]. These methods attempt to learn a 4D feature vector for every " + }, + { + "bbox": [ + 304, + 114, + 545, + 258 + ], + "type": "inline_equation", + "content": "(\\mathbf{x},t)" + }, + { + "bbox": [ + 304, + 114, + 545, + 258 + ], + "type": "text", + "content": ", by projecting it to a set of 2D-planes. Embeddings of these projections on these planes can then be integrated to obtain the embedding for the 4D point. This can be mathematically represented as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 353, + 269, + 545, + 312 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 269, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 353, + 269, + 545, + 312 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {f} (\\mathbf {x}, t) = \\mathbf {B} _ {1} \\left(\\mathbf {h} _ {1} (x, y) \\odot \\mathbf {k} _ {1} (z, t)\\right) \\\\ + \\mathbf {B} _ {2} (\\mathbf {h} _ {2} (x, z) \\odot \\mathbf {k} _ {2} (y, t)) \\\\ + \\mathbf {B} _ {3} \\left(\\mathbf {h} _ {3} (y, z) \\odot \\mathbf {k} _ {3} (x, t)\\right). \\tag {3} \\\\ \\end{array}", + "image_path": "60d38196798d1b371666ddd36a494491402fc7087075af86394640c2818c386a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 321, + 545, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 321, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 304, + 321, + 545, + 418 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 321, + 545, + 418 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_i(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 304, + 321, + 545, + 418 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 321, + 545, + 418 + ], + "type": "inline_equation", + "content": "\\mathbf{k}_i(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 304, + 321, + 545, + 418 + ], + "type": "text", + "content": " are functions (evaluated by bilinear interpolation on regularly spaced 2D feature grids) embedding coordinate tuples to features of dimension " + }, + { + "bbox": [ + 304, + 321, + 545, + 418 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 321, + 545, + 418 + ], + "type": "text", + "content": " \" " + }, + { + "bbox": [ + 304, + 321, + 545, + 418 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 304, + 321, + 545, + 418 + ], + "type": "text", + "content": " denotes an element-wise product, and " + }, + { + "bbox": [ + 304, + 321, + 545, + 418 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_i(\\cdot)" + }, + { + "bbox": [ + 304, + 321, + 545, + 418 + ], + "type": "text", + "content": " denotes a linear transform which maps the products to feature vectors. Subsequently, a tiny MLP can map the feature vector " + }, + { + "bbox": [ + 304, + 321, + 545, + 418 + ], + "type": "inline_equation", + "content": "\\mathbf{f}(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 304, + 321, + 545, + 418 + ], + "type": "text", + "content": " to the volume density, " + }, + { + "bbox": [ + 304, + 321, + 545, + 418 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 321, + 545, + 418 + ], + "type": "text", + "content": " , and the view-dependent emitted color, c, given the viewing direction d." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 430, + 411, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 430, + 411, + 443 + ], + "spans": [ + { + "bbox": [ + 305, + 430, + 411, + 443 + ], + "type": "text", + "content": "4. Proposed Method" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 450, + 545, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 545, + 509 + ], + "type": "text", + "content": "Given a set of " + }, + { + "bbox": [ + 304, + 450, + 545, + 509 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 450, + 545, + 509 + ], + "type": "text", + "content": " input videos, " + }, + { + "bbox": [ + 304, + 450, + 545, + 509 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = \\{V_1, V_2, \\dots, V_W\\}" + }, + { + "bbox": [ + 304, + 450, + 545, + 509 + ], + "type": "text", + "content": " of a dynamic scene, with calibrated camera poses, our approach represents the scene using a series of 4D feature volumes (Section 4.1) along with 4D semantic embeddings (Section 4.2)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 510, + 545, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 654 + ], + "type": "text", + "content": "Analogous to multiple gears in motor vehicles for optimizing engine performance, Gear-NeRF stratifies this semantically embedded scene representation into " + }, + { + "bbox": [ + 304, + 510, + 545, + 654 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{gear}}" + }, + { + "bbox": [ + 304, + 510, + 545, + 654 + ], + "type": "text", + "content": " levels, based on the motion scales. Each of these levels is called a gear. Through our training scheme, regions with larger motion are assigned higher gear levels (Section 4.3) and as a result, are more densely sampled (Section 4.4) for improved dynamic novel view synthesis. Our 4D semantic embedding also enables a new functionality, almost for free - free-viewpoint tracking of target objects, given simple user prompts like clicks (Section 4.5). Figure 2 shows the overall pipeline of Gear-NeRF." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 662, + 454, + 674 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 662, + 454, + 674 + ], + "spans": [ + { + "bbox": [ + 305, + 662, + 454, + 674 + ], + "type": "text", + "content": "4.1. Serial 4D Feature Volumes" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 681, + 545, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 681, + 545, + 730 + ], + "spans": [ + { + "bbox": [ + 304, + 681, + 545, + 730 + ], + "type": "text", + "content": "Instead of using a unified 4D volume to represent a dynamic scene [1, 6, 26, 71], our representation consists of a series of feature volumes, each corresponding to a gear level, " + }, + { + "bbox": [ + 304, + 681, + 545, + 730 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 304, + 681, + 545, + 730 + ], + "type": "text", + "content": ". Specifically, for any space-time coordinate " + }, + { + "bbox": [ + 304, + 681, + 545, + 730 + ], + "type": "inline_equation", + "content": "(\\mathbf{x}, t)" + }, + { + "bbox": [ + 304, + 681, + 545, + 730 + ], + "type": "text", + "content": ", its feature" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 757, + 317, + 765 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 757, + 317, + 765 + ], + "spans": [ + { + "bbox": [ + 294, + 757, + 317, + 765 + ], + "type": "text", + "content": "19669" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 58, + 251, + 70 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 58, + 251, + 70 + ], + "spans": [ + { + "bbox": [ + 47, + 58, + 251, + 70 + ], + "type": "text", + "content": "vector corresponding to " + }, + { + "bbox": [ + 47, + 58, + 251, + 70 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 47, + 58, + 251, + 70 + ], + "type": "text", + "content": " is computed as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 92, + 76, + 287, + 125 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 76, + 287, + 125 + ], + "spans": [ + { + "bbox": [ + 92, + 76, + 287, + 125 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {f} ^ {\\mathcal {G}} (\\mathbf {x}, t) = \\mathbf {B} _ {1} \\left(\\mathbf {h} _ {1} (x, y) \\odot \\mathbf {k} _ {1} ^ {\\mathcal {G}} (z, t)\\right) \\\\ + \\mathbf {B} _ {2} (\\mathbf {h} _ {2} (x, z) \\odot \\mathbf {k} _ {2} ^ {\\mathcal {G}} (y, t)) \\\\ + \\mathbf {B} _ {3} \\left(\\mathbf {h} _ {3} (y, z) \\odot \\mathbf {k} _ {3} ^ {\\mathcal {G}} (x, t)\\right). \\tag {4} \\\\ \\end{array}", + "image_path": "db987b861980170bab8bc8f5b48fb38fe05652d2820cab4732026100058e63c7.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 133, + 287, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 133, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 47, + 133, + 287, + 205 + ], + "type": "text", + "content": "The vector-valued functions " + }, + { + "bbox": [ + 47, + 133, + 287, + 205 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_j(\\cdot, \\cdot)" + }, + { + "bbox": [ + 47, + 133, + 287, + 205 + ], + "type": "text", + "content": " and linear transforms " + }, + { + "bbox": [ + 47, + 133, + 287, + 205 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_j(\\cdot)" + }, + { + "bbox": [ + 47, + 133, + 287, + 205 + ], + "type": "text", + "content": " are shared by all gears, while each gear has its own spatio-temporal embedding " + }, + { + "bbox": [ + 47, + 133, + 287, + 205 + ], + "type": "inline_equation", + "content": "\\mathbf{k}_j^G(\\cdot, \\cdot)" + }, + { + "bbox": [ + 47, + 133, + 287, + 205 + ], + "type": "text", + "content": ", in " + }, + { + "bbox": [ + 47, + 133, + 287, + 205 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 133, + 287, + 205 + ], + "type": "text", + "content": "-dimensional space. Therefore, each gear describes regions of a certain scale of motion while the purely spatial features can be shared among all gears." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 205, + 287, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 205, + 287, + 242 + ], + "spans": [ + { + "bbox": [ + 47, + 205, + 287, + 242 + ], + "type": "text", + "content": "We obtain the gear level at any spatio-temporal coordinate also from a planar-factorized 4D feature volume. Specifically, the gear level at " + }, + { + "bbox": [ + 47, + 205, + 287, + 242 + ], + "type": "inline_equation", + "content": "(\\mathbf{x},t)" + }, + { + "bbox": [ + 47, + 205, + 287, + 242 + ], + "type": "text", + "content": " is computed as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 94, + 249, + 287, + 297 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 249, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 94, + 249, + 287, + 297 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} g (\\mathbf {x}, t) = \\mathbf {1} ^ {\\top} \\left(\\mathbf {h} _ {1} ^ {\\prime} (x, y) \\odot \\mathbf {k} _ {1} ^ {\\prime} (z, t)\\right) \\\\ + \\mathbf {1} ^ {\\top} \\left(\\mathbf {h} _ {2} ^ {\\prime} (x, z) \\odot \\mathbf {k} _ {2} ^ {\\prime} (y, t)\\right) \\\\ + \\mathbf {1} ^ {\\top} \\left(\\mathbf {h} _ {3} ^ {\\prime} (y, z) \\odot \\mathbf {k} _ {3} ^ {\\prime} (x, t)\\right), \\tag {5} \\\\ \\end{array}", + "image_path": "09a0a6353bd9a9ecd0c326cd4e245fb82164f4018b08ba26753b603748468942.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 304, + 287, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 304, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 47, + 304, + 287, + 354 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 304, + 287, + 354 + ], + "type": "inline_equation", + "content": "\\mathbf{1}" + }, + { + "bbox": [ + 47, + 304, + 287, + 354 + ], + "type": "text", + "content": " is a vector of ones, " + }, + { + "bbox": [ + 47, + 304, + 287, + 354 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_i^{\\prime}(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 47, + 304, + 287, + 354 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 304, + 287, + 354 + ], + "type": "inline_equation", + "content": "\\mathbf{k}_i^\\prime (\\cdot ,\\cdot)" + }, + { + "bbox": [ + 47, + 304, + 287, + 354 + ], + "type": "text", + "content": " are " + }, + { + "bbox": [ + 47, + 304, + 287, + 354 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 304, + 287, + 354 + ], + "type": "text", + "content": " dimensional embedding functions. This however defines a continuous feature volume. To map it to the gear level integers, we apply the following projection operation:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 75, + 361, + 287, + 405 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 361, + 287, + 405 + ], + "spans": [ + { + "bbox": [ + 75, + 361, + 287, + 405 + ], + "type": "interline_equation", + "content": "p (\\mathbf {x}, t) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} g (\\mathbf {x}, t) < 1, \\\\ N _ {\\text {g e a r}}, & \\text {i f} g (\\mathbf {x}, t) \\geq N _ {\\text {g e a r}}, \\\\ \\lceil g (\\mathbf {x}, t) \\rceil , & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {6}", + "image_path": "efbfa6e53b9a5f9119a4d02c76f5d269ca5210fd9b4ea87ad40d29d9d31f7734.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 413, + 287, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 413, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 47, + 413, + 287, + 437 + ], + "type": "text", + "content": "Based on this gear level volume, we define a 4D mask for a region at gear level " + }, + { + "bbox": [ + 47, + 413, + 287, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 47, + 413, + 287, + 437 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 94, + 445, + 287, + 477 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 445, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 94, + 445, + 287, + 477 + ], + "type": "interline_equation", + "content": "m _ {\\mathcal {G}} (\\mathbf {x}, t) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} p (\\mathbf {x}, t) = \\mathcal {G}, \\\\ 0, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {7}", + "image_path": "8597e358fdecb921a009823b0612604c5a3dcf4243e0e1be5ddb21408a0a2449.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 485, + 242, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 485, + 242, + 498 + ], + "spans": [ + { + "bbox": [ + 47, + 485, + 242, + 498 + ], + "type": "text", + "content": "The final feature vector at " + }, + { + "bbox": [ + 47, + 485, + 242, + 498 + ], + "type": "inline_equation", + "content": "(\\mathbf{x},t)" + }, + { + "bbox": [ + 47, + 485, + 242, + 498 + ], + "type": "text", + "content": " is computed as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 100, + 506, + 287, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 506, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 100, + 506, + 287, + 540 + ], + "type": "interline_equation", + "content": "\\mathbf {f} (\\mathbf {x}, t) = \\sum_ {\\mathcal {G} = 1} ^ {N _ {\\text {g e a r}}} m _ {\\mathcal {G}} (\\mathbf {x}, t) \\mathbf {f} ^ {\\mathcal {G}} (\\mathbf {x}, t). \\tag {8}", + "image_path": "3573646ee1bb928b8df571cb998263c6845cc1212cf8f8e68be61530f48b9de5.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 548, + 287, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 548, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 548, + 287, + 597 + ], + "type": "text", + "content": "Subsequently, a tiny MLP, " + }, + { + "bbox": [ + 47, + 548, + 287, + 597 + ], + "type": "inline_equation", + "content": "F_{\\theta}" + }, + { + "bbox": [ + 47, + 548, + 287, + 597 + ], + "type": "text", + "content": ", maps these feature vectors " + }, + { + "bbox": [ + 47, + 548, + 287, + 597 + ], + "type": "inline_equation", + "content": "\\mathbf{f}(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 47, + 548, + 287, + 597 + ], + "type": "text", + "content": " as well as the viewing direction " + }, + { + "bbox": [ + 47, + 548, + 287, + 597 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 47, + 548, + 287, + 597 + ], + "type": "text", + "content": " to the volume density " + }, + { + "bbox": [ + 47, + 548, + 287, + 597 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 47, + 548, + 287, + 597 + ], + "type": "text", + "content": " and radiance color " + }, + { + "bbox": [ + 47, + 548, + 287, + 597 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 47, + 548, + 287, + 597 + ], + "type": "text", + "content": ". This allows us to obtain a photometric rendering of the scene." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 604, + 186, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 604, + 186, + 616 + ], + "spans": [ + { + "bbox": [ + 47, + 604, + 186, + 616 + ], + "type": "text", + "content": "4.2. 4D Semantic Embedding" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 621, + 287, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 287, + 730 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 287, + 730 + ], + "type": "text", + "content": "Gear-NeRF leverages the strong object priors of the SAM [38] model to acquire a semantic understanding of the scene, for improved photometric rendering (Section 4.3 and Section 4.4) as well as free-viewpoint object tracking (Section 4.5). Toward this end, we utilize the SAM encoder to obtain 2D feature maps from the frames of each video. We then optimize a 4D SAM embedding field by supervising it with these 2D feature maps. In particular, the MLP above, " + }, + { + "bbox": [ + 47, + 621, + 287, + 730 + ], + "type": "inline_equation", + "content": "F_{\\theta}" + }, + { + "bbox": [ + 47, + 621, + 287, + 730 + ], + "type": "text", + "content": ", is configured to output a 4D semantic embedding s for" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 313, + 57, + 387, + 112 + ], + "blocks": [ + { + "bbox": [ + 389, + 48, + 463, + 56 + ], + "lines": [ + { + "bbox": [ + 389, + 48, + 463, + 56 + ], + "spans": [ + { + "bbox": [ + 389, + 48, + 463, + 56 + ], + "type": "inline_equation", + "content": "1^{\\mathrm{st}}" + }, + { + "bbox": [ + 389, + 48, + 463, + 56 + ], + "type": "text", + "content": " Gear Assignment Update" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 57, + 387, + 112 + ], + "lines": [ + { + "bbox": [ + 313, + 57, + 387, + 112 + ], + "spans": [ + { + "bbox": [ + 313, + 57, + 387, + 112 + ], + "type": "image", + "image_path": "e3fa72dc36fd506e57f807a0b5f7022d82a22a739737276e39de2b2edcc87568.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 389, + 57, + 462, + 112 + ], + "blocks": [ + { + "bbox": [ + 389, + 57, + 462, + 112 + ], + "lines": [ + { + "bbox": [ + 389, + 57, + 462, + 112 + ], + "spans": [ + { + "bbox": [ + 389, + 57, + 462, + 112 + ], + "type": "image", + "image_path": "50fa637dd9b5cd1d8ef5c3734619da2a8b18532884ebf3307e56e6012fac32e4.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 465, + 57, + 539, + 112 + ], + "blocks": [ + { + "bbox": [ + 465, + 57, + 539, + 112 + ], + "lines": [ + { + "bbox": [ + 465, + 57, + 539, + 112 + ], + "spans": [ + { + "bbox": [ + 465, + 57, + 539, + 112 + ], + "type": "image", + "image_path": "e23f902aa511437326737a562472d7caf86849a36e8cd344b80d94bdfa92a64a.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 313, + 123, + 387, + 178 + ], + "blocks": [ + { + "bbox": [ + 388, + 114, + 463, + 121 + ], + "lines": [ + { + "bbox": [ + 388, + 114, + 463, + 121 + ], + "spans": [ + { + "bbox": [ + 388, + 114, + 463, + 121 + ], + "type": "text", + "content": "2nd Gear Assignment Update" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 123, + 387, + 178 + ], + "lines": [ + { + "bbox": [ + 313, + 123, + 387, + 178 + ], + "spans": [ + { + "bbox": [ + 313, + 123, + 387, + 178 + ], + "type": "image", + "image_path": "d87f3f86106845ae486f9020b20c21ed3ab64b99550e216c7f09eea295157c5c.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 389, + 123, + 462, + 178 + ], + "blocks": [ + { + "bbox": [ + 389, + 123, + 462, + 178 + ], + "lines": [ + { + "bbox": [ + 389, + 123, + 462, + 178 + ], + "spans": [ + { + "bbox": [ + 389, + 123, + 462, + 178 + ], + "type": "image", + "image_path": "47af5c1ed79559ca8d3a83c738ee9115e1dc867b9daf89f179fe35cfc79ac20c.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 389, + 180, + 463, + 188 + ], + "lines": [ + { + "bbox": [ + 389, + 180, + 463, + 188 + ], + "spans": [ + { + "bbox": [ + 389, + 180, + 463, + 188 + ], + "type": "inline_equation", + "content": "3^{\\mathrm{rd}}" + }, + { + "bbox": [ + 389, + 180, + 463, + 188 + ], + "type": "text", + "content": " Gear Assignment Update" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 465, + 123, + 539, + 178 + ], + "blocks": [ + { + "bbox": [ + 465, + 123, + 539, + 178 + ], + "lines": [ + { + "bbox": [ + 465, + 123, + 539, + 178 + ], + "spans": [ + { + "bbox": [ + 465, + 123, + 539, + 178 + ], + "type": "image", + "image_path": "440fbf6fbb39d32553f1849c7425139eb95591a5e7cf66188db92618d076026a.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 313, + 189, + 387, + 244 + ], + "blocks": [ + { + "bbox": [ + 313, + 189, + 387, + 244 + ], + "lines": [ + { + "bbox": [ + 313, + 189, + 387, + 244 + ], + "spans": [ + { + "bbox": [ + 313, + 189, + 387, + 244 + ], + "type": "image", + "image_path": "2943450a8edc3e657892f88efead7cc1277c40ff2564f801a6aba17522ec63e1.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 389, + 189, + 462, + 244 + ], + "blocks": [ + { + "bbox": [ + 389, + 189, + 462, + 244 + ], + "lines": [ + { + "bbox": [ + 389, + 189, + 462, + 244 + ], + "spans": [ + { + "bbox": [ + 389, + 189, + 462, + 244 + ], + "type": "image", + "image_path": "dbea8c9a4761b4dfee689b657b03a59d77398d43bea0d872794a3f642fdd62ea.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 465, + 189, + 539, + 244 + ], + "blocks": [ + { + "bbox": [ + 465, + 189, + 539, + 244 + ], + "lines": [ + { + "bbox": [ + 465, + 189, + 539, + 244 + ], + "spans": [ + { + "bbox": [ + 465, + 189, + 539, + 244 + ], + "type": "image", + "image_path": "07dc913a232a4ca56bdf8c7f81a07d84290cd2bbaf7e7b9422c685494c190bec.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 313, + 255, + 387, + 311 + ], + "blocks": [ + { + "bbox": [ + 313, + 255, + 387, + 311 + ], + "lines": [ + { + "bbox": [ + 313, + 255, + 387, + 311 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 387, + 311 + ], + "type": "image", + "image_path": "e66b9f16b1e8424443e7fdb89e5a5849e28053e972b233b72ef7e949bedcee11.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 312, + 373, + 320 + ], + "lines": [ + { + "bbox": [ + 329, + 312, + 373, + 320 + ], + "spans": [ + { + "bbox": [ + 329, + 312, + 373, + 320 + ], + "type": "text", + "content": "Rendered Image" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 331, + 545, + 452 + ], + "lines": [ + { + "bbox": [ + 304, + 331, + 545, + 452 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 545, + 452 + ], + "type": "text", + "content": "Figure 3. Illustration of Gear Assignment Update: For each gear assignment update, we calculate the rendering loss map between the rendered RGB-SAM map and the ground truth and identify the centers of the patches with the maximum and minimum losses, marked in red and green (second column). These points are then fed into the SAM decoder as positive and negative prompts to generate an upshift mask representing the areas that need to be shifted to a higher gear (last column). After the first gear assignment update, we see that the next candidate region for upshift is situated where the horse is located, and so on. Upshift mask colors imply the gear levels after the update (blue-2, green-3, red-4)." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 389, + 255, + 462, + 311 + ], + "blocks": [ + { + "bbox": [ + 389, + 246, + 463, + 255 + ], + "lines": [ + { + "bbox": [ + 389, + 246, + 463, + 255 + ], + "spans": [ + { + "bbox": [ + 389, + 246, + 463, + 255 + ], + "type": "inline_equation", + "content": "4^{\\mathrm{th}}" + }, + { + "bbox": [ + 389, + 246, + 463, + 255 + ], + "type": "text", + "content": " Gear Assignment Update" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 389, + 255, + 462, + 311 + ], + "lines": [ + { + "bbox": [ + 389, + 255, + 462, + 311 + ], + "spans": [ + { + "bbox": [ + 389, + 255, + 462, + 311 + ], + "type": "image", + "image_path": "b7d48a34fdb91fa0479f22030faadcd57dedd2e93119424b892303001284c9bd.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 398, + 312, + 454, + 320 + ], + "lines": [ + { + "bbox": [ + 398, + 312, + 454, + 320 + ], + "spans": [ + { + "bbox": [ + 398, + 312, + 454, + 320 + ], + "type": "text", + "content": "Rendering Loss Map" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 465, + 255, + 539, + 311 + ], + "blocks": [ + { + "bbox": [ + 465, + 255, + 539, + 311 + ], + "lines": [ + { + "bbox": [ + 465, + 255, + 539, + 311 + ], + "spans": [ + { + "bbox": [ + 465, + 255, + 539, + 311 + ], + "type": "image", + "image_path": "e53c1f54b24deda748147abfcf6d354e8558e6bd4ac49d6499b0bdd54d13466f.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 484, + 312, + 520, + 319 + ], + "lines": [ + { + "bbox": [ + 484, + 312, + 520, + 319 + ], + "spans": [ + { + "bbox": [ + 484, + 312, + 520, + 319 + ], + "type": "text", + "content": "Upshift Mask" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "bbox": [ + 304, + 458, + 545, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 458, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 458, + 545, + 518 + ], + "type": "text", + "content": "a given space-time coordinate in addition to the density, " + }, + { + "bbox": [ + 304, + 458, + 545, + 518 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 458, + 545, + 518 + ], + "type": "text", + "content": ", and color, c. To render 2D semantic feature maps in a given view, we compute the semantic feature of a pixel in the feature map by tracing a ray through it and perform volume rendering analogous to Equation 1, as follows:" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 384, + 526, + 545, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 526, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 384, + 526, + 545, + 559 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {S}} (\\mathbf {r}) = \\sum_ {i = 1} ^ {N} T _ {i} \\alpha_ {i} \\mathbf {s} _ {i}. \\tag {9}", + "image_path": "669c40a9777b271e217553fad90be8fa41829f2a92188845cfe4b2ea5a1d342e.jpg" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 304, + 566, + 545, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 566, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 566, + 545, + 602 + ], + "type": "text", + "content": "This SAM embedding is supervised by minimizing the mean squared error between the prediction and the ground truth features " + }, + { + "bbox": [ + 304, + 566, + 545, + 602 + ], + "type": "inline_equation", + "content": "(\\mathbf{S}(\\mathbf{r}))" + }, + { + "bbox": [ + 304, + 566, + 545, + 602 + ], + "type": "text", + "content": " from the SAM encoder, as shown:" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 364, + 610, + 545, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 610, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 364, + 610, + 545, + 635 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {S A M}} = \\sum_ {\\mathbf {r} \\in \\mathcal {R}} \\| \\hat {\\mathbf {S}} (\\mathbf {r}) - \\mathbf {S} (\\mathbf {r}) \\| _ {2} ^ {2}. \\tag {10}", + "image_path": "9e75d4d15a612f7c5437349bc8e4aae544465bc252ead285609a53967821b66c.jpg" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 305, + 651, + 515, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 651, + 515, + 664 + ], + "spans": [ + { + "bbox": [ + 305, + 651, + 515, + 664 + ], + "type": "text", + "content": "4.3. Training Scheme with Gear Assignment" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 304, + 670, + 545, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 670, + 545, + 705 + ], + "spans": [ + { + "bbox": [ + 304, + 670, + 545, + 705 + ], + "type": "text", + "content": "With gear initialization " + }, + { + "bbox": [ + 304, + 670, + 545, + 705 + ], + "type": "inline_equation", + "content": "g(\\mathbf{x}, t) = 1, \\forall \\mathbf{x}, t" + }, + { + "bbox": [ + 304, + 670, + 545, + 705 + ], + "type": "text", + "content": ", the (semantically embedded) radiance field optimization and gear assignment updating take place in an alternating fashion." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 305, + 706, + 545, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 706, + 545, + 730 + ], + "spans": [ + { + "bbox": [ + 305, + 706, + 545, + 730 + ], + "type": "text", + "content": "Gear Assignment Update: As illustrated in Figure 3, when updating the gear assignment after a period of radiance field" + } + ] + } + ], + "index": 40 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 757, + 317, + 765 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 757, + 317, + 765 + ], + "spans": [ + { + "bbox": [ + 294, + 757, + 317, + 765 + ], + "type": "text", + "content": "19670" + } + ] + } + ], + "index": 41 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 58, + 287, + 106 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 58, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 46, + 58, + 287, + 106 + ], + "type": "text", + "content": "optimization, we find the regions rendered most poorly from the rendering loss maps and increment their gears for denser spatio-temporal sampling. The following steps lay out the process for updating gear assignments to regions:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 106, + 286, + 344 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 47, + 106, + 286, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 106, + 286, + 177 + ], + "spans": [ + { + "bbox": [ + 47, + 106, + 286, + 177 + ], + "type": "text", + "content": "- We sample a number of viewpoints and time steps and render 2D-images/SAM features for it. For every rendered RGB-SAM map, we compute a rendering loss map. Each pixel of the rendering loss map is computed as: " + }, + { + "bbox": [ + 47, + 106, + 286, + 177 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\mathbf{r})\\coloneqq \\mathcal{L}_{\\mathrm{pho}}(\\mathbf{r}) + \\lambda \\mathcal{L}_{\\mathrm{SAM}}(\\mathbf{r})" + }, + { + "bbox": [ + 47, + 106, + 286, + 177 + ], + "type": "text", + "content": ". See Figure 3 for example loss maps." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 178, + 286, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 178, + 286, + 237 + ], + "spans": [ + { + "bbox": [ + 47, + 178, + 286, + 237 + ], + "type": "text", + "content": "- Next, we patchify each rendering loss map to find patches with the top-" + }, + { + "bbox": [ + 47, + 178, + 286, + 237 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 178, + 286, + 237 + ], + "type": "text", + "content": " largest/smallest average loss. The center coordinate of these patches serve as positive/negative prompts for the next step. See Figure 3 for example positive (red) / negative (green) prompts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 237, + 286, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 237, + 286, + 308 + ], + "spans": [ + { + "bbox": [ + 47, + 237, + 286, + 308 + ], + "type": "text", + "content": "- We then feed the ground truth RGB images together with positive and negative prompts into the SAM decoder [38] to estimate an upshift mask. These masks tend to cover regions that have motions and are not satisfactorily rendered with the current sampling resolution. We have multiple upshift masks at different viewpoints and time steps." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 309, + 286, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 309, + 286, + 344 + ], + "spans": [ + { + "bbox": [ + 47, + 309, + 286, + 344 + ], + "type": "text", + "content": "- For every pixel of an upshift mask, we trace a ray and sample points along it and update the gear assignment by pushing " + }, + { + "bbox": [ + 47, + 309, + 286, + 344 + ], + "type": "inline_equation", + "content": "g(\\mathbf{x}, t)" + }, + { + "bbox": [ + 47, + 309, + 286, + 344 + ], + "type": "text", + "content": " towards incremented values." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 346, + 286, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 346, + 286, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 286, + 552 + ], + "type": "text", + "content": "In particular, in the last step, for each pixel within an upshift mask, a corresponding ray is traced that connects it with the camera center o, along direction d. Next, a set of points are sampled along this ray. The collection of sampled points, that lie on the rays emanating from within the masked region constitutes the set " + }, + { + "bbox": [ + 46, + 346, + 286, + 552 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{upshift}} = \\{(x_i^{\\mathrm{upshift}},t_i^{\\mathrm{upshift}})\\}_{i = 1}^{N_{\\mathrm{upshift}}}" + }, + { + "bbox": [ + 46, + 346, + 286, + 552 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 346, + 286, + 552 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{upshift}}" + }, + { + "bbox": [ + 46, + 346, + 286, + 552 + ], + "type": "text", + "content": " represents the total count of points sampled from rays pertaining to the masked region. We then follow a similar procedure to sample a set of points pertaining to the unmasked region. We denote this set as: " + }, + { + "bbox": [ + 46, + 346, + 286, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{S}_{\\mathrm{stay}} = \\{(x_i^{\\mathrm{stay}},t_i^{\\mathrm{stay}})\\}_{i = 1}^{N_{\\mathrm{stay}}}" + }, + { + "bbox": [ + 46, + 346, + 286, + 552 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 46, + 346, + 286, + 552 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{stay}}" + }, + { + "bbox": [ + 46, + 346, + 286, + 552 + ], + "type": "text", + "content": " indicating the total number of points sampled from rays that pertain to the unmasked area. Next, for each sample point in each of " + }, + { + "bbox": [ + 46, + 346, + 286, + 552 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{upshift}}" + }, + { + "bbox": [ + 46, + 346, + 286, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 346, + 286, + 552 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{stay}}" + }, + { + "bbox": [ + 46, + 346, + 286, + 552 + ], + "type": "text", + "content": ", we query its current gear level " + }, + { + "bbox": [ + 46, + 346, + 286, + 552 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x},t)" + }, + { + "bbox": [ + 46, + 346, + 286, + 552 + ], + "type": "text", + "content": ". In order to assign new gear levels, we need to update the gear assignment function " + }, + { + "bbox": [ + 46, + 346, + 286, + 552 + ], + "type": "inline_equation", + "content": "g(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 46, + 346, + 286, + 552 + ], + "type": "text", + "content": ", which proceeds with the objective function:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 52, + 560, + 286, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 560, + 286, + 633 + ], + "spans": [ + { + "bbox": [ + 52, + 560, + 286, + 633 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {u p s h i f t}} = \\frac {1}{N _ {\\text {u p s h i f t}}} \\sum_ {(\\mathbf {x}, t) \\in \\mathcal {S} _ {\\text {u p s h i f t}}} \\| g (\\mathbf {x}, t; \\boldsymbol {\\Theta}) - (p (\\mathbf {x}, t) + 1) \\| _ {2} ^ {2} \\\\ + \\frac {\\lambda_ {\\text {s t a y}}}{N _ {\\text {s t a y}}} \\sum_ {(\\mathbf {x}, t) \\in \\mathcal {S} _ {\\text {s t a y}}} \\| g (\\mathbf {x}, t; \\boldsymbol {\\Theta}) - p (\\mathbf {x}, t) \\| _ {2} ^ {2}, \\tag {11} \\\\ \\end{array}", + "image_path": "bfa10a13ffb41ddac4790ac05548d6054810ae7b7998a27b2f3d58d67fca96eb.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 635, + 286, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 635, + 286, + 730 + ], + "spans": [ + { + "bbox": [ + 46, + 635, + 286, + 730 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 635, + 286, + 730 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 46, + 635, + 286, + 730 + ], + "type": "text", + "content": " denotes the set of estimizable parameters for " + }, + { + "bbox": [ + 46, + 635, + 286, + 730 + ], + "type": "inline_equation", + "content": "g(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 46, + 635, + 286, + 730 + ], + "type": "text", + "content": ". The minimization of the first term encourages sample points within the masked region " + }, + { + "bbox": [ + 46, + 635, + 286, + 730 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{upshift}}" + }, + { + "bbox": [ + 46, + 635, + 286, + 730 + ], + "type": "text", + "content": " to have a gear level equal to their current gear incremented by one, resulting in an upshift of gear. Conversely, minimizing the second term encourages the points in " + }, + { + "bbox": [ + 46, + 635, + 286, + 730 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{stay}}" + }, + { + "bbox": [ + 46, + 635, + 286, + 730 + ], + "type": "text", + "content": " to maintain their gear values, thereby encouraging the remaining regions to keep their current gear levels and avoiding unwanted up" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 58, + 544, + 81 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 58, + 544, + 81 + ], + "spans": [ + { + "bbox": [ + 306, + 58, + 544, + 81 + ], + "type": "text", + "content": "shifts. We update " + }, + { + "bbox": [ + 306, + 58, + 544, + 81 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 306, + 58, + 544, + 81 + ], + "type": "text", + "content": " via a single step of gradient descent as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 375, + 88, + 544, + 101 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 88, + 544, + 101 + ], + "spans": [ + { + "bbox": [ + 375, + 88, + 544, + 101 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Theta} \\leftarrow \\boldsymbol {\\Theta} - \\alpha \\nabla_ {\\boldsymbol {\\Theta}} \\mathcal {L} _ {\\text {u p s h i f t}}, \\tag {12}", + "image_path": "04191ea3ee82d1b7e5d0de8f89fbe915433ce235a067f305b61fdd7b55b40f07.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 113, + 545, + 197 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 113, + 545, + 197 + ], + "spans": [ + { + "bbox": [ + 304, + 113, + 545, + 197 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 113, + 545, + 197 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 113, + 545, + 197 + ], + "type": "text", + "content": " is the learning rate. Since, " + }, + { + "bbox": [ + 304, + 113, + 545, + 197 + ], + "type": "inline_equation", + "content": "g(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 304, + 113, + 545, + 197 + ], + "type": "text", + "content": " is essentially derived from the embedding functions, " + }, + { + "bbox": [ + 304, + 113, + 545, + 197 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_i^{\\prime}(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 304, + 113, + 545, + 197 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 113, + 545, + 197 + ], + "type": "inline_equation", + "content": "\\mathbf{k}_i^\\prime (\\cdot ,\\cdot)" + }, + { + "bbox": [ + 304, + 113, + 545, + 197 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 304, + 113, + 545, + 197 + ], + "type": "inline_equation", + "content": "i\\in \\{1,2,3\\}" + }, + { + "bbox": [ + 304, + 113, + 545, + 197 + ], + "type": "text", + "content": ", the aforementioned optimization step amounts to updating these embedding functions. Once updated, we proceed with a fresh round of gear assignment to increase the spatio-temporal sampling resolution for the regions that end up at a higher gear level than before." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 199, + 545, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 199, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 304, + 199, + 545, + 319 + ], + "type": "text", + "content": "Radiance Field Optimization: With the updated gear assignment, we increase the resolution of spatio-temporal sampling for the gear-shifted regions (Section 4.4) and then resume optimizing the radiance field. We alternate between the two processes: radiance field optimization (each time for " + }, + { + "bbox": [ + 304, + 199, + 545, + 319 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 199, + 545, + 319 + ], + "type": "text", + "content": " epochs), and gear assignment updates until the average variance of each rendering loss map is below a predetermined threshold. After this, we optimize the radiance field for an additional " + }, + { + "bbox": [ + 304, + 199, + 545, + 319 + ], + "type": "inline_equation", + "content": "L'" + }, + { + "bbox": [ + 304, + 199, + 545, + 319 + ], + "type": "text", + "content": " epochs without further gear assignment updates." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 334, + 523, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 334, + 523, + 346 + ], + "spans": [ + { + "bbox": [ + 305, + 334, + 523, + 346 + ], + "type": "text", + "content": "4.4. Motion-aware Spatio-Temporal Sampling" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 354, + 545, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 545, + 450 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 545, + 450 + ], + "type": "text", + "content": "In this subsection, we explain our motion-aware spatiotemporal sampling strategy based on assigned gears, permitting differential processing of regions at different gear levels. By temporal sampling, we imply the choice of temporal resolution for planar-factorized 4D feature volumes, and by spatial sampling, we mean the strategy used to choose sampling points along each ray for volume rendering." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 452, + 545, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 452, + 545, + 584 + ], + "spans": [ + { + "bbox": [ + 304, + 452, + 545, + 584 + ], + "type": "text", + "content": "Motion-aware Temporal Sampling: To handle the increasing intensity of object motion, as reflected by their growing gear levels, we increment the temporal resolution for voxel grids. Specifically, " + }, + { + "bbox": [ + 304, + 452, + 545, + 584 + ], + "type": "inline_equation", + "content": "\\mathbf{k}_j^{\\mathcal{G}}" + }, + { + "bbox": [ + 304, + 452, + 545, + 584 + ], + "type": "text", + "content": " in Equation 4 has increasing resolution along the time axis, thereby empowering the 4D feature volumes to better model the dynamics along the temporal axis. This ensures fast-moving objects can be more faithfully modeled without unsightly blurring. The temporal resolution for each gear's feature volume is determined by linear interpolation between 1 (for " + }, + { + "bbox": [ + 304, + 452, + 545, + 584 + ], + "type": "inline_equation", + "content": "\\mathcal{G} = 1" + }, + { + "bbox": [ + 304, + 452, + 545, + 584 + ], + "type": "text", + "content": ") and the total number of frames (for " + }, + { + "bbox": [ + 304, + 452, + 545, + 584 + ], + "type": "inline_equation", + "content": "\\mathcal{G} = N_{\\mathrm{gear}}" + }, + { + "bbox": [ + 304, + 452, + 545, + 584 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 586, + 545, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 586, + 545, + 730 + ], + "spans": [ + { + "bbox": [ + 304, + 586, + 545, + 730 + ], + "type": "text", + "content": "Motion-aware Spatial Sampling: While denser sampling of points can improve reconstruction accuracy, increasing the number of sampling points throughout the scene can lead to prohibitive computational costs. Therefore, we propose a 3D point-splitting strategy as illustrated in Figure 4. We begin by sampling a relatively small number, " + }, + { + "bbox": [ + 304, + 586, + 545, + 730 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 586, + 545, + 730 + ], + "type": "text", + "content": ", of samples along each ray, assuming it is at the lowest gear level. If a sampled point belongs to a region with a higher gear, as determined by " + }, + { + "bbox": [ + 304, + 586, + 545, + 730 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}, t)" + }, + { + "bbox": [ + 304, + 586, + 545, + 730 + ], + "type": "text", + "content": ", we then sample more densely in that region. For every sampled point in that region, we split it into " + }, + { + "bbox": [ + 304, + 586, + 545, + 730 + ], + "type": "inline_equation", + "content": "2^{p(\\mathbf{x}, t) - 1}" + }, + { + "bbox": [ + 304, + 586, + 545, + 730 + ], + "type": "text", + "content": " points, equally spaced within the corresponding ray segment (at that gear level)." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 757, + 317, + 765 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 757, + 317, + 765 + ], + "spans": [ + { + "bbox": [ + 294, + 757, + 317, + 765 + ], + "type": "text", + "content": "19671" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 56, + 265, + 150 + ], + "blocks": [ + { + "bbox": [ + 61, + 56, + 265, + 150 + ], + "lines": [ + { + "bbox": [ + 61, + 56, + 265, + 150 + ], + "spans": [ + { + "bbox": [ + 61, + 56, + 265, + 150 + ], + "type": "image", + "image_path": "67bfcb663407705ef6fd0df36ebff6879367fd9d28578ab337310359da5d1ed8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 158, + 287, + 215 + ], + "lines": [ + { + "bbox": [ + 46, + 158, + 287, + 215 + ], + "spans": [ + { + "bbox": [ + 46, + 158, + 287, + 215 + ], + "type": "text", + "content": "Figure 4. Motion-aware Spatial Sampling: We split each sampled point into " + }, + { + "bbox": [ + 46, + 158, + 287, + 215 + ], + "type": "inline_equation", + "content": "2^{p(\\mathbf{x},t)}" + }, + { + "bbox": [ + 46, + 158, + 287, + 215 + ], + "type": "text", + "content": " points, equally spaced within the corresponding ray segment. The top row shows the vanilla uniformly sampled points, while the bottom one shows the densely sampled points after splitting using our strategy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 221, + 279, + 235 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 221, + 279, + 235 + ], + "spans": [ + { + "bbox": [ + 47, + 221, + 279, + 235 + ], + "type": "text", + "content": "4.5. Free-Viewpoint Tracking with User Prompts" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 240, + 286, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 240, + 286, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 286, + 324 + ], + "type": "text", + "content": "Our 4D SAM embedding enables another useful functionality, almost for free - free-viewpoint object tracking, where the user only needs to provide as few as one click to extract the target object based on the 4D embedding. Next, we describe how, given a user-supplied point click at any arbitrary viewpoint and time step, we obtain the corresponding object mask at a novel viewpoint and time step." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 325, + 286, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 325, + 286, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 325, + 286, + 467 + ], + "type": "text", + "content": "Masks for Novel Viewpoints: The first step for this task entails finding the 3D correspondence of the user click. We trace a ray through the selected pixel, and by utilizing the volume density, we determine the depth at which the ray intersects with the first object surface it encounters. This yields the 3D coordinates of the point of intersection. Subsequently, the 3D coordinates of this intersection can be easily mapped into a 2D coordinate within any novel viewpoint image, using the camera pose of the new viewpoint. Alongside the rendered SAM feature map of the novel view, we feed this coordinate into the SAM decoder to generate the object mask for the novel view." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 469, + 286, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 469, + 286, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 469, + 286, + 589 + ], + "type": "text", + "content": "Masks for Novel Time Steps: For this task, we propagate an object mask to its neighboring time step. Specifically, with an object mask for a specific frame " + }, + { + "bbox": [ + 46, + 469, + 286, + 589 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 469, + 286, + 589 + ], + "type": "text", + "content": ", we calculate the bounding box of this mask and use this bounding box as a prompt to SAM for neighboring frames " + }, + { + "bbox": [ + 46, + 469, + 286, + 589 + ], + "type": "inline_equation", + "content": "t' = t + 1" + }, + { + "bbox": [ + 46, + 469, + 286, + 589 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 469, + 286, + 589 + ], + "type": "inline_equation", + "content": "t' = t - 1" + }, + { + "bbox": [ + 46, + 469, + 286, + 589 + ], + "type": "text", + "content": ". By inputting this prompt along with the rendered SAM feature map at " + }, + { + "bbox": [ + 46, + 469, + 286, + 589 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 46, + 469, + 286, + 589 + ], + "type": "text", + "content": " into the SAM decoder, we can obtain the object mask for " + }, + { + "bbox": [ + 46, + 469, + 286, + 589 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 46, + 469, + 286, + 589 + ], + "type": "text", + "content": ". Combining the above two processes, we can start from a single click and get the object mask in any viewpoint and time step." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 601, + 128, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 601, + 128, + 615 + ], + "spans": [ + { + "bbox": [ + 47, + 601, + 128, + 615 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 621, + 286, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 621, + 286, + 730 + ], + "spans": [ + { + "bbox": [ + 46, + 621, + 286, + 730 + ], + "type": "text", + "content": "We assess the performance of our proposed Gear-NeRF for dynamic novel view synthesis and free-viewpoint object tracking across a range of challenging datasets, comparing it with state-of-the-art methods. Through ablation studies, we provide empirical evidence of the effectiveness of its fundamental components. We kindly refer the reader to our supplementary material for additional experimental details and results, including videos for free-viewpoint rendering and tracking." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 57, + 422, + 70 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 57, + 422, + 70 + ], + "spans": [ + { + "bbox": [ + 306, + 57, + 422, + 70 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 78, + 545, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 78, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 304, + 78, + 545, + 243 + ], + "type": "text", + "content": "Implementation Details: We implement our method using PyTorch [62] and conduct experiments on an NVIDIA RTX 4090 GPU with 24 GB RAM. We divide each input video into chunks of 100 frames. For every chunk, we train a model for approximately 2.5 hours. Our 4D feature volumes yield embeddings with a dimension of " + }, + { + "bbox": [ + 304, + 78, + 545, + 243 + ], + "type": "inline_equation", + "content": "M = 32" + }, + { + "bbox": [ + 304, + 78, + 545, + 243 + ], + "type": "text", + "content": ". We set the gear number " + }, + { + "bbox": [ + 304, + 78, + 545, + 243 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{gear}}" + }, + { + "bbox": [ + 304, + 78, + 545, + 243 + ], + "type": "text", + "content": " to 4. We find patches with top- " + }, + { + "bbox": [ + 304, + 78, + 545, + 243 + ], + "type": "inline_equation", + "content": "k = 3" + }, + { + "bbox": [ + 304, + 78, + 545, + 243 + ], + "type": "text", + "content": " largest/smallest average loss for gear assignment updates to obtain prompts. The rendering loss map is computed with " + }, + { + "bbox": [ + 304, + 78, + 545, + 243 + ], + "type": "inline_equation", + "content": "\\lambda = 0.01" + }, + { + "bbox": [ + 304, + 78, + 545, + 243 + ], + "type": "text", + "content": ". For the optimization of radiance fields, " + }, + { + "bbox": [ + 304, + 78, + 545, + 243 + ], + "type": "inline_equation", + "content": "L = 3" + }, + { + "bbox": [ + 304, + 78, + 545, + 243 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 78, + 545, + 243 + ], + "type": "inline_equation", + "content": "L' = 10" + }, + { + "bbox": [ + 304, + 78, + 545, + 243 + ], + "type": "text", + "content": ". In our motion-aware spatial sampling, each ray initially has " + }, + { + "bbox": [ + 304, + 78, + 545, + 243 + ], + "type": "inline_equation", + "content": "n = 64" + }, + { + "bbox": [ + 304, + 78, + 545, + 243 + ], + "type": "text", + "content": " sampling points. We use an initial learning rate of 0.02 for all the parameters. For Equation 12, we use " + }, + { + "bbox": [ + 304, + 78, + 545, + 243 + ], + "type": "inline_equation", + "content": "\\alpha = 0.02" + }, + { + "bbox": [ + 304, + 78, + 545, + 243 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 247, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 247, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 247, + 545, + 510 + ], + "type": "text", + "content": "Datasets: (i) The Technicolor light field dataset [70] includes diverse indoor environment videos captured by a " + }, + { + "bbox": [ + 304, + 247, + 545, + 510 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 304, + 247, + 545, + 510 + ], + "type": "text", + "content": " camera rig. We evaluate on 4 sequences (Train, Theater, Painter, Birthday) at the original " + }, + { + "bbox": [ + 304, + 247, + 545, + 510 + ], + "type": "inline_equation", + "content": "2048 \\times 1088" + }, + { + "bbox": [ + 304, + 247, + 545, + 510 + ], + "type": "text", + "content": " resolution, holding out the same view as prior work [1] (the second row and second column) for evaluation. (ii) The Neural 3D Video dataset [44] includes indoor multi-view video sequences captured by 20 cameras at a resolution of " + }, + { + "bbox": [ + 304, + 247, + 545, + 510 + ], + "type": "inline_equation", + "content": "2704 \\times 2028" + }, + { + "bbox": [ + 304, + 247, + 545, + 510 + ], + "type": "text", + "content": " pixels. We experiment on 6 sequences (Cut Roasted Beef, Flame Steak, Coffee Martini, Cook Spinach, Flame Salmon, Sear Steak), downsampling by a factor of 2 and holding out the central view (akin to prior work [1]) for evaluation. (iii) The Google Immersive dataset[5] contains light field videos of indoor and outdoor scenes captured by a time-synchronized 46-fisheye camera rig, with a resolution of " + }, + { + "bbox": [ + 304, + 247, + 545, + 510 + ], + "type": "inline_equation", + "content": "2560 \\times 1920" + }, + { + "bbox": [ + 304, + 247, + 545, + 510 + ], + "type": "text", + "content": " pixels. We experiment with 9 sequences from it (Flames, Truck, Horse, Car, Welder, Exhibit, Face Paint 1, Face Paint 2, Cave). We downsample the video resolution by a factor of 2 and hold out the central view (like prior work [1]) for evaluation. For our experiments, we adopted the same resolution and held-out view selection as prior work [1]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 512, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 512, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 512, + 545, + 715 + ], + "type": "text", + "content": "Evaluation Metrics: For the task of novel view synthesis of dynamic scenes, we evaluate using the following standard metrics: (i) Peak Signal-to-Noise Ratio (PSNR), (ii) Structural Similarity Index Measure (SSIM) [83] and (iii) Learned Perceptual Image Patch Similarity (LPIPS) [100], by comparing the reconstructed frames against ground truth images. These metrics are computed on the held-out view and averaged across all frames. For the free-viewpoint object tracking task, we designate a specific viewpoint and time step for the user to give prompting clicks. Subsequently, we assess the quality of the predicted object masks at novel viewpoints. The quality of the object mask is quantified in terms of the Mean Intersection over Union (mIoU) and accuracy (Acc.), which are calculated against the ground truth mask, manually annotated utilizing Adobe Photoshop. Additionally, we present the same metrics computed for novel time steps, denoted as t-mIoU and t-Acc." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 717, + 545, + 729 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 717, + 545, + 729 + ], + "spans": [ + { + "bbox": [ + 306, + 717, + 545, + 729 + ], + "type": "text", + "content": "Baselines: We run a comprehensive comparison of our" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 756, + 317, + 765 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 756, + 317, + 765 + ], + "spans": [ + { + "bbox": [ + 294, + 756, + 317, + 765 + ], + "type": "text", + "content": "19672" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 57, + 183, + 140 + ], + "blocks": [ + { + "bbox": [ + 72, + 57, + 183, + 140 + ], + "lines": [ + { + "bbox": [ + 72, + 57, + 183, + 140 + ], + "spans": [ + { + "bbox": [ + 72, + 57, + 183, + 140 + ], + "type": "image", + "image_path": "995f64513dc7dcc780c6471fd2e6af6df9a29b5181a29b9bd39043bebb702214.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 184, + 57, + 293, + 140 + ], + "blocks": [ + { + "bbox": [ + 184, + 57, + 293, + 140 + ], + "lines": [ + { + "bbox": [ + 184, + 57, + 293, + 140 + ], + "spans": [ + { + "bbox": [ + 184, + 57, + 293, + 140 + ], + "type": "image", + "image_path": "dc52ec33367132fb857c04026dd2a395088ccd663d23f174067dcaabf1cbdca7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 295, + 57, + 405, + 140 + ], + "blocks": [ + { + "bbox": [ + 295, + 57, + 405, + 140 + ], + "lines": [ + { + "bbox": [ + 295, + 57, + 405, + 140 + ], + "spans": [ + { + "bbox": [ + 295, + 57, + 405, + 140 + ], + "type": "image", + "image_path": "d4db5f3a30ed9868c493d8acb4701fa98355ebc454111ac934fbdec4a31fc069.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 407, + 57, + 517, + 140 + ], + "blocks": [ + { + "bbox": [ + 407, + 57, + 517, + 140 + ], + "lines": [ + { + "bbox": [ + 407, + 57, + 517, + 140 + ], + "spans": [ + { + "bbox": [ + 407, + 57, + 517, + 140 + ], + "type": "image", + "image_path": "a2cb6d42855b09e182add840584accf719154eb6b74ae6f1764b137bdc28e1d2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 72, + 143, + 183, + 227 + ], + "blocks": [ + { + "bbox": [ + 72, + 143, + 183, + 227 + ], + "lines": [ + { + "bbox": [ + 72, + 143, + 183, + 227 + ], + "spans": [ + { + "bbox": [ + 72, + 143, + 183, + 227 + ], + "type": "image", + "image_path": "dfe3e86592245362b13d0599a3240e10d819a2960401b7398bb90495e4d0b5b1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 184, + 144, + 294, + 227 + ], + "blocks": [ + { + "bbox": [ + 184, + 144, + 294, + 227 + ], + "lines": [ + { + "bbox": [ + 184, + 144, + 294, + 227 + ], + "spans": [ + { + "bbox": [ + 184, + 144, + 294, + 227 + ], + "type": "image", + "image_path": "e6fe6e58d2bf1f0e4a8852645168917a64beb9d389006baf35e56d2beef68533.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 296, + 144, + 405, + 226 + ], + "blocks": [ + { + "bbox": [ + 296, + 144, + 405, + 226 + ], + "lines": [ + { + "bbox": [ + 296, + 144, + 405, + 226 + ], + "spans": [ + { + "bbox": [ + 296, + 144, + 405, + 226 + ], + "type": "image", + "image_path": "f26ffd8f80ec8c5d2c9669fe25d385d4d25418244054b97e1c9eec17915a6003.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 407, + 144, + 517, + 227 + ], + "blocks": [ + { + "bbox": [ + 407, + 144, + 517, + 227 + ], + "lines": [ + { + "bbox": [ + 407, + 144, + 517, + 227 + ], + "spans": [ + { + "bbox": [ + 407, + 144, + 517, + 227 + ], + "type": "image", + "image_path": "a25bb988e9ffcd1291ae92e5a376aff57865a49f26eddafc48e22c8280f74f5d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 72, + 230, + 183, + 313 + ], + "blocks": [ + { + "bbox": [ + 72, + 230, + 183, + 313 + ], + "lines": [ + { + "bbox": [ + 72, + 230, + 183, + 313 + ], + "spans": [ + { + "bbox": [ + 72, + 230, + 183, + 313 + ], + "type": "image", + "image_path": "f659904e0228a212bd8293989fd4201852d0d9015be72fb69713b180863352b7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 104, + 314, + 150, + 323 + ], + "lines": [ + { + "bbox": [ + 104, + 314, + 150, + 323 + ], + "spans": [ + { + "bbox": [ + 104, + 314, + 150, + 323 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 184, + 231, + 294, + 312 + ], + "blocks": [ + { + "bbox": [ + 184, + 231, + 294, + 312 + ], + "lines": [ + { + "bbox": [ + 184, + 231, + 294, + 312 + ], + "spans": [ + { + "bbox": [ + 184, + 231, + 294, + 312 + ], + "type": "image", + "image_path": "466c8a2e5a7934f4152f4690fde1baa9090778ebd4b92c1135daf028aab74129.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 314, + 257, + 324 + ], + "lines": [ + { + "bbox": [ + 220, + 314, + 257, + 324 + ], + "spans": [ + { + "bbox": [ + 220, + 314, + 257, + 324 + ], + "type": "text", + "content": "HyperReel" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 296, + 231, + 405, + 312 + ], + "blocks": [ + { + "bbox": [ + 296, + 231, + 405, + 312 + ], + "lines": [ + { + "bbox": [ + 296, + 231, + 405, + 312 + ], + "spans": [ + { + "bbox": [ + 296, + 231, + 405, + 312 + ], + "type": "image", + "image_path": "2072749a5c5a3b280a167ec83d172e289318ca9278b7384a4a7c0e2bb1d8e93e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 314, + 378, + 323 + ], + "lines": [ + { + "bbox": [ + 321, + 314, + 378, + 323 + ], + "spans": [ + { + "bbox": [ + 321, + 314, + 378, + 323 + ], + "type": "text", + "content": "Neural 3D Video" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 333, + 545, + 357 + ], + "lines": [ + { + "bbox": [ + 46, + 333, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 333, + 545, + 357 + ], + "type": "text", + "content": "Figure 5. Qualitative comparisons for novel view synthesis on the Technicolor dataset [70]: We qualitatively compare our approach against HyperReel [1] and Neural 3D Video [44]. Our approach better recovers fine details like patterns on the toys or stripes on the shirt." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 407, + 231, + 517, + 312 + ], + "blocks": [ + { + "bbox": [ + 407, + 231, + 517, + 312 + ], + "lines": [ + { + "bbox": [ + 407, + 231, + 517, + 312 + ], + "spans": [ + { + "bbox": [ + 407, + 231, + 517, + 312 + ], + "type": "image", + "image_path": "2a982f4a69cc86b234e01c70e77fc14a9d2325513fb985926b171a73241f5754.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 453, + 315, + 471, + 323 + ], + "lines": [ + { + "bbox": [ + 453, + 315, + 471, + 323 + ], + "spans": [ + { + "bbox": [ + 453, + 315, + 471, + 323 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 362, + 289, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 362, + 289, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 362, + 289, + 471 + ], + "type": "text", + "content": "method against a range of recent NeRF-based baseline methods: (i) ST-NeRF [99], (ii) HexPlane [6], (iii) HyperReel [1], and (iv) MixVoxels [81]. As the first method to enable promptable free-viewpoint object tracking under the NeRF setting, there is no established direct baseline for this specific task. However, SA3D [8], a recent method for segmenting static scenes, is treated as a baseline for predicting masks at novel viewpoints corresponding to the prompted time step." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 47, + 481, + 105, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 481, + 105, + 493 + ], + "spans": [ + { + "bbox": [ + 47, + 481, + 105, + 493 + ], + "type": "text", + "content": "5.2. Results" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 501, + 287, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 501, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 46, + 501, + 287, + 633 + ], + "type": "text", + "content": "Dynamic Novel View Synthesis: As shown in Figure 5, Gear-NeRF produces high-quality novel view synthesis of dynamic scenes, accurately modeling real-world scenes with intricate motions and fine details. For example, patterns on the toys or stripes on the shirt, are faithfully rendered, resulting in more photo-realistic images compared to all of the baselines. Table 1, presents quantitative comparisons of our method against the baselines. While Gear-NeRF has longer training (Tr. Time) / inference times (FPS) compared to some baselines, it almost always achieves the best performance in terms of rendering quality." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 634, + 288, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 634, + 288, + 730 + ], + "spans": [ + { + "bbox": [ + 46, + 634, + 288, + 730 + ], + "type": "text", + "content": "Free-Viewpoint Tracking: In Figure 6, we present the object masks obtained by our method based on the user prompts provided at a specified viewpoint and time step. Specifically, the first row displays masked objects at the prompted viewpoint and time step. The second row shows novel view masks at the prompted time step. The third row shows novel view masks at novel time steps. We see that masks obtained from Gear-NeRF show precise boundaries," + } + ] + } + ], + "index": 20 + }, + { + "type": "table", + "bbox": [ + 306, + 406, + 545, + 510 + ], + "blocks": [ + { + "bbox": [ + 305, + 361, + 545, + 405 + ], + "lines": [ + { + "bbox": [ + 305, + 361, + 545, + 405 + ], + "spans": [ + { + "bbox": [ + 305, + 361, + 545, + 405 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons for dynamic novel view synthesis: Our method outperforms all baselines across all datasets on all metrics. We report means over all scenes for each dataset. Best and second best results are highlighted." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 406, + 545, + 510 + ], + "lines": [ + { + "bbox": [ + 306, + 406, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 306, + 406, + 545, + 510 + ], + "type": "table", + "html": "
DatasetMethodPSNR (↑)SSIM (↑)LPIPS (↓)Tr. Time(↓)FPS(↑)
Technicolor [70]ST-NeRF [99]30.860.8830.101344 min0.7
HyperReel [1]31.040.8870.09297 min7.7
MixVoxels [81]28.990.8420.10314 min18.9
Ours32.210.9190.058148 min7.9
Google Immersive [5]HexPlane [6]27.670.8080.188527 min0.9
HyperReel [1]28.320.8620.145117 min7.4
MixVoxels [81]27.140.8350.20919 min16.8
Ours28.740.8760.122189 min7.0
Neural 3D Video [44]ST-NeRF [99]31.030.8900.081367 min0.7
HyperReel [1]31.120.9280.065129 min6.1
MixVoxels [81]30.690.9440.13915 min16.3
Ours31.800.9360.058204 min6.8
", + "image_path": "61c795a2452bec2f26f2300b03a2323a1890dab704922567c5ace877ef5aa187.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "table_body" + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 511, + 545, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 511, + 545, + 642 + ], + "spans": [ + { + "bbox": [ + 304, + 511, + 545, + 642 + ], + "type": "text", + "content": "compared to SA3D. Table 2 presents quantitative assessment of the quality of the masks generated of our method versus SA3D. Our method exceeds " + }, + { + "bbox": [ + 304, + 511, + 545, + 642 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 304, + 511, + 545, + 642 + ], + "type": "text", + "content": " across all metrics and datasets, demonstrating the effectiveness of our approach. Our gains over SA3D can be attributed to the fact that SA3D, as a static scene segmentation method, does not utilize information across all time frames, whereas our approach does. SA3D is incapable of predicting masks for novel time steps, and as such, the corresponding entries are marked as not applicable, a shortcoming which our method does not have." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 651, + 384, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 651, + 384, + 662 + ], + "spans": [ + { + "bbox": [ + 306, + 651, + 384, + 662 + ], + "type": "text", + "content": "5.3. Evaluations" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 670, + 545, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 670, + 545, + 704 + ], + "spans": [ + { + "bbox": [ + 304, + 670, + 545, + 704 + ], + "type": "text", + "content": "To verify the effectiveness of the design choices of GearNeRF, we perform extensive ablation studies on the Truck scene of the Google Immersive dataset [5]." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 305, + 706, + 545, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 706, + 545, + 730 + ], + "spans": [ + { + "bbox": [ + 305, + 706, + 545, + 730 + ], + "type": "text", + "content": "Motion-aware Temporal Sampling: An intuitive strategy for adjusting the temporal sampling is to directly modify" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 757, + 317, + 765 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 757, + 317, + 765 + ], + "spans": [ + { + "bbox": [ + 294, + 757, + 317, + 765 + ], + "type": "text", + "content": "19673" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 147, + 59, + 240, + 162 + ], + "blocks": [ + { + "bbox": [ + 78, + 79, + 131, + 89 + ], + "lines": [ + { + "bbox": [ + 78, + 79, + 131, + 89 + ], + "spans": [ + { + "bbox": [ + 78, + 79, + 131, + 89 + ], + "type": "text", + "content": "★ User Prompt" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 147, + 59, + 240, + 162 + ], + "lines": [ + { + "bbox": [ + 147, + 59, + 240, + 162 + ], + "spans": [ + { + "bbox": [ + 147, + 59, + 240, + 162 + ], + "type": "image", + "image_path": "c72389cbbf17f9d3b80ca8280e02a5715b7ac600a62094223b1da648b940c8bf.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 147, + 163, + 240, + 213 + ], + "blocks": [ + { + "bbox": [ + 85, + 131, + 134, + 148 + ], + "lines": [ + { + "bbox": [ + 85, + 131, + 134, + 148 + ], + "spans": [ + { + "bbox": [ + 85, + 131, + 134, + 148 + ], + "type": "text", + "content": "Rendering from Novel View" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 75, + 178, + 144, + 196 + ], + "lines": [ + { + "bbox": [ + 75, + 178, + 144, + 196 + ], + "spans": [ + { + "bbox": [ + 75, + 178, + 144, + 196 + ], + "type": "text", + "content": "Rendering from Novel View at Novel Time" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 147, + 163, + 240, + 213 + ], + "lines": [ + { + "bbox": [ + 147, + 163, + 240, + 213 + ], + "spans": [ + { + "bbox": [ + 147, + 163, + 240, + 213 + ], + "type": "image", + "image_path": "584ffa2ede060c98cbd97fc22a40b6418869315379b5d85f7db39d9701e145c9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 168, + 216, + 219, + 225 + ], + "lines": [ + { + "bbox": [ + 168, + 216, + 219, + 225 + ], + "spans": [ + { + "bbox": [ + 168, + 216, + 219, + 225 + ], + "type": "text", + "content": "Rendered Image" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 236, + 545, + 258 + ], + "lines": [ + { + "bbox": [ + 46, + 236, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 545, + 258 + ], + "type": "text", + "content": "Figure 6. Qualitative comparisons of free-viewpoint object tracking on Technicolor [70] and Neural 3D Video [44] datasets: Our method can obtain desirable object masks, with clear edges, from prompting points provided by users at desired time steps and viewpoints." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 253, + 63, + 280, + 107 + ], + "blocks": [ + { + "bbox": [ + 253, + 63, + 280, + 107 + ], + "lines": [ + { + "bbox": [ + 253, + 63, + 280, + 107 + ], + "spans": [ + { + "bbox": [ + 253, + 63, + 280, + 107 + ], + "type": "image", + "image_path": "4230cacf43138d46dfe8dd3e44d0a838b064007f64f61bfececcf09953f8dbfd.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 243, + 114, + 286, + 161 + ], + "blocks": [ + { + "bbox": [ + 243, + 114, + 286, + 161 + ], + "lines": [ + { + "bbox": [ + 243, + 114, + 286, + 161 + ], + "spans": [ + { + "bbox": [ + 243, + 114, + 286, + 161 + ], + "type": "image", + "image_path": "8d6c4c78db77b3f57a91aacb3d8bbe92fc2654618baf82e02bfdbb80f9ba8d21.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 243, + 162, + 294, + 213 + ], + "blocks": [ + { + "bbox": [ + 243, + 162, + 294, + 213 + ], + "lines": [ + { + "bbox": [ + 243, + 162, + 294, + 213 + ], + "spans": [ + { + "bbox": [ + 243, + 162, + 294, + 213 + ], + "type": "image", + "image_path": "94b307177c859840ad0d1b8da6483e1303c8216e5b89e585e9a8d613e74eb3ba.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 258, + 217, + 278, + 224 + ], + "lines": [ + { + "bbox": [ + 258, + 217, + 278, + 224 + ], + "spans": [ + { + "bbox": [ + 258, + 217, + 278, + 224 + ], + "type": "text", + "content": "SA3D" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 306, + 60, + 334, + 107 + ], + "blocks": [ + { + "bbox": [ + 306, + 60, + 334, + 107 + ], + "lines": [ + { + "bbox": [ + 306, + 60, + 334, + 107 + ], + "spans": [ + { + "bbox": [ + 306, + 60, + 334, + 107 + ], + "type": "image", + "image_path": "4c70833af9728c2eaae13facae9812e0ae2642aadc1180731d70446c4038e3f9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 309, + 113, + 337, + 159 + ], + "blocks": [ + { + "bbox": [ + 309, + 113, + 337, + 159 + ], + "lines": [ + { + "bbox": [ + 309, + 113, + 337, + 159 + ], + "spans": [ + { + "bbox": [ + 309, + 113, + 337, + 159 + ], + "type": "image", + "image_path": "1bc3f75de1bf3b7d6e9526f53bd2708669632d330658a88a0eb27d7d6eb45781.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 305, + 167, + 337, + 209 + ], + "blocks": [ + { + "bbox": [ + 305, + 167, + 337, + 209 + ], + "lines": [ + { + "bbox": [ + 305, + 167, + 337, + 209 + ], + "spans": [ + { + "bbox": [ + 305, + 167, + 337, + 209 + ], + "type": "image", + "image_path": "3032b11ac1cc1e51c64f9c1849c499d2a0dec2be2e55d8d0fdf793875c7bf1ea.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 217, + 328, + 224 + ], + "lines": [ + { + "bbox": [ + 313, + 217, + 328, + 224 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 328, + 224 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 348, + 59, + 415, + 110 + ], + "blocks": [ + { + "bbox": [ + 348, + 59, + 415, + 110 + ], + "lines": [ + { + "bbox": [ + 348, + 59, + 415, + 110 + ], + "spans": [ + { + "bbox": [ + 348, + 59, + 415, + 110 + ], + "type": "image", + "image_path": "81483608b12406fd56271932d4a42eb35caf3c997ce397cbfc91db5d0df4d59d.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 348, + 111, + 415, + 161 + ], + "blocks": [ + { + "bbox": [ + 348, + 111, + 415, + 161 + ], + "lines": [ + { + "bbox": [ + 348, + 111, + 415, + 161 + ], + "spans": [ + { + "bbox": [ + 348, + 111, + 415, + 161 + ], + "type": "image", + "image_path": "d7b36c0da5e6aaa56cc01d6895d00a1603c0c597853ebdd657e7b17e615c259e.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 348, + 162, + 415, + 213 + ], + "blocks": [ + { + "bbox": [ + 348, + 162, + 415, + 213 + ], + "lines": [ + { + "bbox": [ + 348, + 162, + 415, + 213 + ], + "spans": [ + { + "bbox": [ + 348, + 162, + 415, + 213 + ], + "type": "image", + "image_path": "74f70bbb2b3b9c1d85479f1ef17b95d46dea3fd1bac1b504f5471667aec3cf79.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 356, + 217, + 407, + 225 + ], + "lines": [ + { + "bbox": [ + 356, + 217, + 407, + 225 + ], + "spans": [ + { + "bbox": [ + 356, + 217, + 407, + 225 + ], + "type": "text", + "content": "Rendered Image" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 427, + 61, + 457, + 109 + ], + "blocks": [ + { + "bbox": [ + 427, + 61, + 457, + 109 + ], + "lines": [ + { + "bbox": [ + 427, + 61, + 457, + 109 + ], + "spans": [ + { + "bbox": [ + 427, + 61, + 457, + 109 + ], + "type": "image", + "image_path": "e9ce9c240794324dbf67d0d133faf98920106b37155c4d8da9cf7eb2431354d4.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 417, + 119, + 467, + 157 + ], + "blocks": [ + { + "bbox": [ + 417, + 119, + 467, + 157 + ], + "lines": [ + { + "bbox": [ + 417, + 119, + 467, + 157 + ], + "spans": [ + { + "bbox": [ + 417, + 119, + 467, + 157 + ], + "type": "image", + "image_path": "baf9ccfe19d602f96e0583f966b0389b237d013d4b5c1834fcb94e38b5f91ac3.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 417, + 162, + 468, + 213 + ], + "blocks": [ + { + "bbox": [ + 417, + 162, + 468, + 213 + ], + "lines": [ + { + "bbox": [ + 417, + 162, + 468, + 213 + ], + "spans": [ + { + "bbox": [ + 417, + 162, + 468, + 213 + ], + "type": "image", + "image_path": "5d20b5754326b29d9840ff2b282124d7e4f00b45bebe71696a0fda8c374283c5.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 432, + 217, + 452, + 224 + ], + "lines": [ + { + "bbox": [ + 432, + 217, + 452, + 224 + ], + "spans": [ + { + "bbox": [ + 432, + 217, + 452, + 224 + ], + "type": "text", + "content": "SA3D" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 482, + 60, + 509, + 99 + ], + "blocks": [ + { + "bbox": [ + 482, + 60, + 509, + 99 + ], + "lines": [ + { + "bbox": [ + 482, + 60, + 509, + 99 + ], + "spans": [ + { + "bbox": [ + 482, + 60, + 509, + 99 + ], + "type": "image", + "image_path": "a08f7e5b49cc3ef1f811bd27167ac01f75380d1e696b8c7b05f80fd7b75354ec.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 481, + 118, + 508, + 156 + ], + "blocks": [ + { + "bbox": [ + 481, + 118, + 508, + 156 + ], + "lines": [ + { + "bbox": [ + 481, + 118, + 508, + 156 + ], + "spans": [ + { + "bbox": [ + 481, + 118, + 508, + 156 + ], + "type": "image", + "image_path": "bfb82b569b7877578e997be0670a95f7e91ac38b0f09965fc3683e683d324a64.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 477, + 171, + 506, + 208 + ], + "blocks": [ + { + "bbox": [ + 477, + 171, + 506, + 208 + ], + "lines": [ + { + "bbox": [ + 477, + 171, + 506, + 208 + ], + "spans": [ + { + "bbox": [ + 477, + 171, + 506, + 208 + ], + "type": "image", + "image_path": "d8079c3abe7497ee3ef2e41c709929aa5bdbc990a7e0582a04be49801357d3d9.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 487, + 217, + 503, + 224 + ], + "lines": [ + { + "bbox": [ + 487, + 217, + 503, + 224 + ], + "spans": [ + { + "bbox": [ + 487, + 217, + 503, + 224 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "table", + "bbox": [ + 47, + 308, + 287, + 384 + ], + "blocks": [ + { + "bbox": [ + 46, + 264, + 287, + 307 + ], + "lines": [ + { + "bbox": [ + 46, + 264, + 287, + 307 + ], + "spans": [ + { + "bbox": [ + 46, + 264, + 287, + 307 + ], + "type": "text", + "content": "Table 2. Quantitative comparisons for free-viewpoint tracking: t-mIoU and t-Acc are metrics used for evaluating novel view masks at novel time steps, not applicable to SA3D. Reported metrics are averages over all scenes for each dataset." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 308, + 287, + 384 + ], + "lines": [ + { + "bbox": [ + 47, + 308, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 47, + 308, + 287, + 384 + ], + "type": "table", + "html": "
DatasetMethodmIoU (↑)Acc. (↑)t-mIoU (↑)t-Acc. (↑)
Technicolor [70]SA3D [8]96.497.1N/AN/A
Ours97.497.692.193.3
Google Immersive [5]SA3D [8]94.194.8N/AN/A
Ours94.395.091.592.8
Neural 3D Video [44]SA3D [8]93.194.0N/AN/A
Ours93.494.390.692.3
", + "image_path": "e7e5c38ccf57c7d2e54616aecff35052013dbeecc13b0669b005ae95d70209f1.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "table_body" + } + ], + "index": 28 + }, + { + "bbox": [ + 46, + 390, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 390, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 390, + 287, + 533 + ], + "type": "text", + "content": "the temporal resolution of the 4D feature volume and make it uniform across all regions of the scene. In Table 3, we demonstrate the performance of this naive temporal sampling strategy using medium (25) or dense (100) temporal resolutions on 100-frame input videos. The results show that increasing the temporal resolution universally does not necessarily yield better performance. In contrast, our proposed method, which utilizes motion-aware temporal sampling, achieves the best results. We surmise that this may be attributed to the model distributing the volume's capacity sparsely across a large number of time steps, which is not the case for our method." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 46, + 536, + 287, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 536, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 46, + 536, + 287, + 656 + ], + "type": "text", + "content": "Motion-aware Spatial Sampling: To validate the effectiveness of our motion-aware spatial sampling strategy, we compare it against other sampling strategies. The first variant involves uniform sampling along the ray, and the second uses a sample prediction network (SPN) like Attal et al. [1] for sampling. These variants sample 128 points on each ray. Results in Table 3 show that our full model employing motion-aware spatial sampling outperforms these variants, perhaps by having a better sense of where to sample more from, derived from its semantic-aware embedding." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 46, + 658, + 287, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 658, + 287, + 730 + ], + "spans": [ + { + "bbox": [ + 46, + 658, + 287, + 730 + ], + "type": "text", + "content": "SAM embedding: To verify that introducing SAM embedding can improve rendering quality (rather than just enabling segmentation or tracking), we tested a variant of Gear-NeRF without SAM embedding (it instead thresholds loss maps of rendered RGB frames to obtain upshift masks). As shown in Table 3, the absence of SAM em" + } + ] + } + ], + "index": 31 + }, + { + "type": "table", + "bbox": [ + 306, + 297, + 545, + 419 + ], + "blocks": [ + { + "bbox": [ + 305, + 264, + 545, + 297 + ], + "lines": [ + { + "bbox": [ + 305, + 264, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 305, + 264, + 545, + 297 + ], + "type": "text", + "content": "Table 3. Ablation Study: Ablations on our spatio-temporal sampling strategy and the number of gears (Truck/Google Immersive). Best and second best results are highlighted." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 297, + 545, + 419 + ], + "lines": [ + { + "bbox": [ + 306, + 297, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 306, + 297, + 545, + 419 + ], + "type": "table", + "html": "
MethodPSNR (↑)SSIM (↑)LPIPS (↓)
Naive temporal sampling (medium)26.930.7780.144
Naive temporal sampling (dense)26.850.7600.161
Naive spatial sampling24.800.7340.222
SPN spatial sampling26.540.7870.162
Ours (w/o embedding)27.100.8310.177
Ours (Ngear = 2)26.930.7850.166
Ours (Ngear = 3)27.140.8750.145
Ours (Ngear = 4)27.490.8920.136
Ours (Ngear = 5)27.460.9010.131
", + "image_path": "4d826034e68a4163178c44706ad90cb01fb0737a3881d4442434a9489a1dc18d.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "table_body" + } + ], + "index": 33 + }, + { + "bbox": [ + 306, + 422, + 545, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 422, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 306, + 422, + 545, + 445 + ], + "type": "text", + "content": "bedding for guiding gear assignment reduces the model's rendering quality." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 306, + 446, + 545, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 446, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 306, + 446, + 545, + 518 + ], + "type": "text", + "content": "Number of Gears: We ablate on the numbers of gears. Increasing the number of gears allows for more fine-grained motion-aware spatio-temporal sampling, while increasing the computational cost. As shown in Table 3, a choice of up to 4 gear levels seems optimal, further increasing the number of gears does not result in significant improvements." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 306, + 529, + 383, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 529, + 383, + 542 + ], + "spans": [ + { + "bbox": [ + 306, + 529, + 383, + 542 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 304, + 550, + 545, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 550, + 545, + 730 + ], + "spans": [ + { + "bbox": [ + 304, + 550, + 545, + 730 + ], + "type": "text", + "content": "In this work, we introduced Gear-NeRF, an extension of dynamic NeRFs that leverages semantic information from powerful segmentation models for stratified modeling of dynamic scenes. Our approach learns a 4D (spatiotemporal) semantic embedding and introduces the concept of \"gears\" for differentiated modeling of scene regions based on their motion intensity. With determined gear assignments, Gear-NeRF adaptively adjusts its spatiotemporal sampling resolution to improve the photo-realism of rendered views. At the same time, Gear-NeRF provides the new functionality of free-viewpoint object tracking with user prompts as simple as a click. Our empirical studies underscore the effectiveness of Gear-NeRF, showcasing state-of-the-art performance in both rendering quality and object tracking across multiple challenging datasets." + } + ] + } + ], + "index": 37 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 757, + 317, + 765 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 757, + 317, + 765 + ], + "spans": [ + { + "bbox": [ + 294, + 757, + 317, + 765 + ], + "type": "text", + "content": "19674" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 57, + 105, + 68 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 57, + 105, + 68 + ], + "spans": [ + { + "bbox": [ + 48, + 57, + 105, + 68 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 76, + 288, + 730 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 76, + 287, + 142 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 76, + 287, + 142 + ], + "spans": [ + { + "bbox": [ + 56, + 76, + 287, + 142 + ], + "type": "text", + "content": "[1] Benjamin Attal, Jia-Bin Huang, Christian Richardt, Michael Zollhoefer, Johannes Kopf, Matthew O'Toole, and Changil Kim. HyperReel: High-fidelity 6-DoF video with ray-conditioned sampling. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 144, + 288, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 144, + 288, + 198 + ], + "spans": [ + { + "bbox": [ + 56, + 144, + 288, + 198 + ], + "type": "text", + "content": "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 5855–5864, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 288, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 288, + 253 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 288, + 253 + ], + "type": "text", + "content": "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5470-5479, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 288, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 288, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 288, + 308 + ], + "type": "text", + "content": "[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 19697-19705, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 309, + 288, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 288, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 288, + 364 + ], + "type": "text", + "content": "[5] Michael Broxton, John Flynn, Ryan Overbeck, Daniel Erickson, Peter Hedman, Matthew Duvall, Jason Dourgarian, Jay Busch, Matt Whalen, and Paul Debevec. Immersive light field video with a layered mesh representation. ACM Transactions on Graphics (TOG), 39(4):86-1, 2020. 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 287, + 407 + ], + "type": "text", + "content": "[6] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 130-141, 2023. 1, 2, 3, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 408, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 408, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 408, + 287, + 453 + ], + "type": "text", + "content": "[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 453, + 287, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 453, + 287, + 486 + ], + "spans": [ + { + "bbox": [ + 56, + 453, + 287, + 486 + ], + "type": "text", + "content": "[8] Jiazhong Cen, Zanwei Zhou, Jiemin Fang, Chen Yang, Wei Shen, Lingxi Xie, Dongsheng Jiang, Xiaopeng Zhang, and Qi Tian. Segment anything in 3d with nerfs, 2023. 3, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 487, + 287, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 487, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 56, + 487, + 287, + 540 + ], + "type": "text", + "content": "[9] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5799-5809, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 542, + 287, + 617 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 542, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 53, + 542, + 287, + 617 + ], + "type": "text", + "content": "[10] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 619, + 287, + 662 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 619, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 53, + 619, + 287, + 662 + ], + "type": "text", + "content": "[11] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision (ECCV), pages 333-350. Springer, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 663, + 287, + 706 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 663, + 287, + 706 + ], + "spans": [ + { + "bbox": [ + 53, + 663, + 287, + 706 + ], + "type": "text", + "content": "[12] Anpei Chen, Zexiang Xu, Xinyue Wei, Siyu Tang, Hao Su, and Andreas Geiger. Factor fields: A unified framework for neural fields and beyond. arXiv preprint arXiv:2302.01226, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 708, + 287, + 730 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 708, + 287, + 730 + ], + "spans": [ + { + "bbox": [ + 53, + 708, + 287, + 730 + ], + "type": "text", + "content": "[13] Jiaben Chen and Huaizu Jiang. Sportsslomo: A new benchmark and baselines for human-centric video frame interpo" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 59, + 545, + 730 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 331, + 59, + 545, + 80 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 59, + 545, + 80 + ], + "spans": [ + { + "bbox": [ + 331, + 59, + 545, + 80 + ], + "type": "text", + "content": "lation. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 82, + 545, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 82, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 312, + 82, + 545, + 137 + ], + "type": "text", + "content": "[14] Jiaben Chen, Renrui Zhang, Dongze Lian, Jiaqi Yang, Ziyao Zeng, and Jianbo Shi. iquery: Instruments as queries for audio-visual sound separation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14675–14686, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 138, + 545, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 138, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 312, + 138, + 545, + 191 + ], + "type": "text", + "content": "[15] Xiaokang Chen, Kwan-Yee Lin, Chen Qian, Gang Zeng, and Hongsheng Li. 3d sketch-aware semantic scene completion via semi-supervised structure prior. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 194, + 545, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 194, + 545, + 225 + ], + "spans": [ + { + "bbox": [ + 312, + 194, + 545, + 225 + ], + "type": "text", + "content": "[16] Xiaokang Chen, Jiaxiang Tang, Diwen Wan, Jingbo Wang, and Gang Zeng. Interactive segment anything nef with feature imitation. arXiv preprint arXiv:2305.16233, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 227, + 545, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 227, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 312, + 227, + 545, + 281 + ], + "type": "text", + "content": "[17] Yue Chen, Xuan Wang, Xingyu Chen, Qi Zhang, Xiaoyu Li, Yu Guo, Jue Wang, and Fei Wang. Uv volumes for real-time rendering of editable free-view human performance. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16621-16631, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 283, + 545, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 283, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 312, + 283, + 545, + 338 + ], + "type": "text", + "content": "[18] Zhang Chen, Zhong Li, Liangchen Song, Lele Chen, Jingyi Yu, Junsong Yuan, and Yi Xu. Neurbf: A neural fields representation with adaptive radial basis functions. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 4182-4194, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 339, + 545, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 339, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 312, + 339, + 545, + 371 + ], + "type": "text", + "content": "[19] Ho Kei Cheng and Alexander G. Schwing. Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 373, + 545, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 373, + 545, + 405 + ], + "spans": [ + { + "bbox": [ + 312, + 373, + 545, + 405 + ], + "type": "text", + "content": "[20] Yangming Cheng, Liulei Li, Yuanyou Xu, Xiaodi Li, Zongxin Yang, Wenguan Wang, and Yi Yang. Segment and track anything. arXiv preprint arXiv:2305.06558, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 407, + 545, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 407, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 312, + 407, + 545, + 460 + ], + "type": "text", + "content": "[21] Alvaro Collet, Ming Chuang, Pat Sweeney, Don Gillett, Dennis Evseev, David Calabrese, Hugues Hoppe, Adam Kirk, and Steve Sullivan. High-quality streamable free-viewpoint video. ACM Transactions on Graphics (TOG), 34(4):1-13, 2015. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 462, + 545, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 462, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 312, + 462, + 545, + 517 + ], + "type": "text", + "content": "[22] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 14304-14314. IEEE Computer Society, 2021. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 518, + 545, + 572 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 518, + 545, + 572 + ], + "spans": [ + { + "bbox": [ + 312, + 518, + 545, + 572 + ], + "type": "text", + "content": "[23] Zhiwen Fan, Peihao Wang, Xinyu Gong, Yifan Jiang, Dejia Xu, and Zhangyang Wang. Nerf-sos: Any-view self-supervised object segmentation from complex real-world scenes. International Conference on Learning Representations (ICLR), pages arXiv-2209, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 574, + 545, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 574, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 312, + 574, + 545, + 628 + ], + "type": "text", + "content": "[24] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. In ACM Transactions on Graphics (SIGGRAPH Asia), pages 1-9, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 630, + 545, + 684 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 630, + 545, + 684 + ], + "spans": [ + { + "bbox": [ + 312, + 630, + 545, + 684 + ], + "type": "text", + "content": "[25] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5501–5510, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 686, + 545, + 730 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 686, + 545, + 730 + ], + "spans": [ + { + "bbox": [ + 312, + 686, + 545, + 730 + ], + "type": "text", + "content": "[26] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In IEEE/CVF Conference on Computer Vision and" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 757, + 317, + 765 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 757, + 317, + 765 + ], + "spans": [ + { + "bbox": [ + 294, + 757, + 317, + 765 + ], + "type": "text", + "content": "19675" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 59, + 288, + 730 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 72, + 59, + 286, + 79 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 59, + 286, + 79 + ], + "spans": [ + { + "bbox": [ + 72, + 59, + 286, + 79 + ], + "type": "text", + "content": "Pattern Recognition (CVPR), pages 12479-12488, 2023. 1, 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 81, + 287, + 125 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 81, + 287, + 125 + ], + "spans": [ + { + "bbox": [ + 53, + 81, + 287, + 125 + ], + "type": "text", + "content": "[27] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 5712-5721, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 126, + 288, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 126, + 288, + 180 + ], + "spans": [ + { + "bbox": [ + 53, + 126, + 288, + 180 + ], + "type": "text", + "content": "[28] Chen Geng, Sida Peng, Zhen Xu, Hujun Bao, and Xiaowei Zhou. Learning neural volumetric representations of dynamic humans in minutes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8759-8770, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 287, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 287, + 226 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 287, + 226 + ], + "type": "text", + "content": "[29] Rahul Goel, Dhawal Sirikonda, Saurabh Saini, and P.J. Narayanan. Interactive Segmentation of Radiance Fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 228, + 288, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 228, + 288, + 270 + ], + "spans": [ + { + "bbox": [ + 53, + 228, + 288, + 270 + ], + "type": "text", + "content": "[30] Yuan-Chen Guo, Di Kang, Linchao Bao, Yu He, and Song-Hai Zhang. Nerfren: Neural radiance fields with reflections. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18409-18418, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 271, + 288, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 271, + 288, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 271, + 288, + 327 + ], + "type": "text", + "content": "[31] Wenbo Hu, Yuling Wang, Lin Ma, Bangbang Yang, Lin Gao, Xiao Liu, and Yuewen Ma. Tri-miprf: Tri-mip representation for efficient anti-aliasing neural radiance fields. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 19774–19783, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 327, + 288, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 327, + 288, + 371 + ], + "spans": [ + { + "bbox": [ + 53, + 327, + 288, + 371 + ], + "type": "text", + "content": "[32] Qiangui Huang, Weiyue Wang, and Ulrich Neumann. Recurrent slice networks for 3d segmentation on point clouds. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 372, + 288, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 372, + 288, + 426 + ], + "spans": [ + { + "bbox": [ + 53, + 372, + 288, + 426 + ], + "type": "text", + "content": "[33] Mustafa Isik, Martin Rünz, Markos Georgopoulos, Taras Khakhulin, Jonathan Starck, Lourdes Agapito, and Matthias Nießner. Humanrf: High-fidelity neural radiance fields for humans in motion. ACM Transactions on Graphics (SIGGRAPH), 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 428, + 288, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 428, + 288, + 472 + ], + "spans": [ + { + "bbox": [ + 53, + 428, + 288, + 472 + ], + "type": "text", + "content": "[34] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 5885-5894, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 473, + 288, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 473, + 288, + 517 + ], + "spans": [ + { + "bbox": [ + 53, + 473, + 288, + 517 + ], + "type": "text", + "content": "[35] Lei Ke, Mingqiao Ye, Martin Danelljan, Yifan Liu, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Segment anything in high quality. Advances in Neural Information Processing Systems (NeurIPS), 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 517, + 288, + 561 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 517, + 288, + 561 + ], + "spans": [ + { + "bbox": [ + 53, + 517, + 288, + 561 + ], + "type": "text", + "content": "[36] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (TOG), 42(4):1-14, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 562, + 288, + 605 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 562, + 288, + 605 + ], + "spans": [ + { + "bbox": [ + 53, + 562, + 288, + 605 + ], + "type": "text", + "content": "[37] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. In IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 607, + 288, + 662 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 607, + 288, + 662 + ], + "spans": [ + { + "bbox": [ + 53, + 607, + 288, + 662 + ], + "type": "text", + "content": "[38] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2, 3, 4, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 662, + 288, + 707 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 662, + 288, + 707 + ], + "spans": [ + { + "bbox": [ + 53, + 662, + 288, + 707 + ], + "type": "text", + "content": "[39] Tobias Kirschstein, Shenhan Qian, Simon Giebenhain, Tim Walter, and Matthias Nießner. Nersemble: Multi-view radiance field reconstruction of human heads. ACM Transactions on Graphics (SIGGRAPH), 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 708, + 288, + 730 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 708, + 288, + 730 + ], + "spans": [ + { + "bbox": [ + 53, + 708, + 288, + 730 + ], + "type": "text", + "content": "[40] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nerf for editing via feature field distil" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 59, + 545, + 730 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 331, + 59, + 545, + 80 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 59, + 545, + 80 + ], + "spans": [ + { + "bbox": [ + 331, + 59, + 545, + 80 + ], + "type": "text", + "content": "lation. Advances in Neural Information Processing Systems (NeurIPS), 35:23311-23330, 2022. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 81, + 545, + 125 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 81, + 545, + 125 + ], + "spans": [ + { + "bbox": [ + 312, + 81, + 545, + 125 + ], + "type": "text", + "content": "[41] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and Rene Ranftl. Language-driven semantic segmentation. In International Conference on Learning Representations (ICLR), 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 126, + 545, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 126, + 545, + 170 + ], + "spans": [ + { + "bbox": [ + 312, + 126, + 545, + 170 + ], + "type": "text", + "content": "[42] Hao Li, Bart Adams, Leonidas J Guibas, and Mark Pauly. Robust single-view geometry and motion reconstruction. ACM Transactions on Graphics (TOG), 28(5):1-10, 2009. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 171, + 545, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 171, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 312, + 171, + 545, + 215 + ], + "type": "text", + "content": "[43] Hao Li, Linjie Luo, Daniel Vlasic, Pieter Peers, Jovan Popovic, Mark Pauly, and Szymon Rusinkiewicz. Temporally coherent completion of dynamic shapes. ACM Transactions on Graphics (TOG), 31(1):1-11, 2012. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 217, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 217, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 312, + 217, + 545, + 282 + ], + "type": "text", + "content": "[44] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5521-5531, 2022. 2, 6, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 283, + 545, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 283, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 312, + 283, + 545, + 338 + ], + "type": "text", + "content": "[45] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 338, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 338, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 312, + 338, + 545, + 392 + ], + "type": "text", + "content": "[46] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4273-4284, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 394, + 545, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 394, + 545, + 439 + ], + "spans": [ + { + "bbox": [ + 312, + 394, + 545, + 439 + ], + "type": "text", + "content": "[47] Haotong Lin, Sida Peng, Zhen Xu, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Efficient neural radiance fields for interactive free-viewpoint video. In ACM Transactions on Graphics (SIGGRAPH Asia), pages 1-9, 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 440, + 545, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 440, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 312, + 440, + 545, + 483 + ], + "type": "text", + "content": "[48] Haotong Lin, Sida Peng, Zhen Xu, Tao Xie, Xingyi He, Hujun Bao, and Xiaowei Zhou. Im4d: High-fidelity and real-time novel view synthesis for dynamic scenes. ACM Transactions on Graphics (SIGGRAPH Asia), 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 484, + 545, + 538 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 484, + 545, + 538 + ], + "spans": [ + { + "bbox": [ + 312, + 484, + 545, + 538 + ], + "type": "text", + "content": "[49] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 9298–9309, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 540, + 545, + 594 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 540, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 312, + 540, + 545, + 594 + ], + "type": "text", + "content": "[50] Xinhang Liu, Jiaben Chen, Huai Yu, Yu-Wing Tai, and Chi-Keung Tang. Unsupervised multi-view object segmentation using radiance field propagation. Advances in Neural Information Processing Systems (NeurIPS), 35:17730-17743, 2022. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 596, + 545, + 639 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 596, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 312, + 596, + 545, + 639 + ], + "type": "text", + "content": "[51] Xinhang Liu, Shiu-hong Kao, Jiaben Chen, Yu-Wing Tai, and Chi-Keung Tang. Deceptive-nerf: Enhancing nerf reconstruction using pseudo-observations from diffusion models. arXiv preprint arXiv:2305.15171, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 641, + 545, + 673 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 641, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 312, + 641, + 545, + 673 + ], + "type": "text", + "content": "[52] Xinhang Liu, Yu-Wing Tai, and Chi-Keung Tang. Clean-nerf: Reformulating nerf to account for view-dependent observations. arXiv preprint arXiv:2303.14707, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 674, + 545, + 730 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 674, + 545, + 730 + ], + "spans": [ + { + "bbox": [ + 312, + 674, + 545, + 730 + ], + "type": "text", + "content": "[53] Yu-Lun Liu, Chen Gao, Andreas Meuleman, Hung-Yu Tseng, Ayush Saraf, Changil Kim, Yung-Yu Chuang, Johannes Kopf, and Jia-Bin Huang. Robust dynamic radiance fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13-23, 2023. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 757, + 318, + 765 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 757, + 318, + 765 + ], + "spans": [ + { + "bbox": [ + 294, + 757, + 318, + 765 + ], + "type": "text", + "content": "19676" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 59, + 288, + 729 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 59, + 287, + 102 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 59, + 287, + 102 + ], + "spans": [ + { + "bbox": [ + 53, + 59, + 287, + 102 + ], + "type": "text", + "content": "[54] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 103, + 288, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 103, + 288, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 103, + 288, + 168 + ], + "type": "text", + "content": "[55] Andreas Meuleman, Yu-Lun Liu, Chen Gao, Jia-Bin Huang, Changil Kim, Min H Kim, and Johannes Kopf. Progressively optimized local radiance fields for robust view synthesis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16539-16548, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 171, + 287, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 287, + 225 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 287, + 225 + ], + "type": "text", + "content": "[56] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision (ECCV), 2020. 1, 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 227, + 287, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 227, + 287, + 303 + ], + "spans": [ + { + "bbox": [ + 53, + 227, + 287, + 303 + ], + "type": "text", + "content": "[57] Ashkan Mirzaei, Tristan Aumentado-Armstrong, Konstantinos G Derpanis, Jonathan Kelly, Marcus A Brubaker, Igor Gilitschenski, and Alex Levinshtein. Spin-nerf: Multiview segmentation and perceptual inpainting with neural radiance fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20669–20679, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 304, + 287, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 304, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 53, + 304, + 287, + 348 + ], + "type": "text", + "content": "[58] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (TOG), 41(4):1-15, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 350, + 287, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 350, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 53, + 350, + 287, + 415 + ], + "type": "text", + "content": "[59] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5480-5490, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 416, + 287, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 416, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 53, + 416, + 287, + 471 + ], + "type": "text", + "content": "[60] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 5865-5874, 2021. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 472, + 287, + 537 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 472, + 287, + 537 + ], + "spans": [ + { + "bbox": [ + 53, + 472, + 287, + 537 + ], + "type": "text", + "content": "[61] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Transactions on Graphics (TOG), 2021. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 539, + 287, + 605 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 539, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 53, + 539, + 287, + 605 + ], + "type": "text", + "content": "[62] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in Neural Information Processing Systems (NeurIPS), 32, 2019. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 606, + 287, + 660 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 606, + 287, + 660 + ], + "spans": [ + { + "bbox": [ + 53, + 606, + 287, + 660 + ], + "type": "text", + "content": "[63] Sida Peng, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Representing volumetric videos as dynamic mlp maps. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4252-4262, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 662, + 287, + 705 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 662, + 287, + 705 + ], + "spans": [ + { + "bbox": [ + 53, + 662, + 287, + 705 + ], + "type": "text", + "content": "[64] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. International Conference on Learning Representations (ICLR), 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 708, + 287, + 729 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 708, + 287, + 729 + ], + "spans": [ + { + "bbox": [ + 53, + 708, + 287, + 729 + ], + "type": "text", + "content": "[65] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 59, + 545, + 729 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 331, + 59, + 545, + 91 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 59, + 545, + 91 + ], + "spans": [ + { + "bbox": [ + 331, + 59, + 545, + 91 + ], + "type": "text", + "content": "for dynamic scenes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10318-10327, 2021. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 92, + 545, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 92, + 545, + 135 + ], + "spans": [ + { + "bbox": [ + 312, + 92, + 545, + 135 + ], + "type": "text", + "content": "[66] Yi-Ling Qiao, Alexander Gao, Yiran Xu, Yue Feng, Jia-Bin Huang, and Ming C Lin. Dynamic mesh-aware radiance fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 385-396, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 137, + 545, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 137, + 545, + 212 + ], + "spans": [ + { + "bbox": [ + 312, + 137, + 545, + 212 + ], + "type": "text", + "content": "[67] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of Machine Learning Research (PMLR), pages 8748-8763, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 213, + 545, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 213, + 545, + 267 + ], + "spans": [ + { + "bbox": [ + 312, + 213, + 545, + 267 + ], + "type": "text", + "content": "[68] Konstantinos Rematas, Andrew Liu, Pratul P Srinivasan, Jonathan T Barron, Andrea Tagliasacchi, Thomas Funkhouser, and Vittorio Ferrari. Urban radiance fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12932-12942, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 268, + 545, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 268, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 312, + 268, + 545, + 312 + ], + "type": "text", + "content": "[69] Zhongzheng Ren, Aseem Agarwala, Bryan Russell, Alexander G. Schwing, and Oliver Wang. Neural volumetric object selection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 312, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 312, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 312, + 312, + 545, + 387 + ], + "type": "text", + "content": "[70] Neus Sabater, Guillaume Boisson, Benoit Vandame, Paul Kerbiriou, Frederic Babon, Matthieu Hog, Remy Gendrot, Tristan Langlois, Olivier Bureller, Arno Schubert, et al. Dataset and pipeline for multi-view light-field video. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 30-40, 2017. 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 388, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 388, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 312, + 388, + 545, + 453 + ], + "type": "text", + "content": "[71] Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boning Liu, Hongwen Zhang, and Yebin Liu. Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16632-16642, 2023. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 455, + 545, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 455, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 312, + 455, + 545, + 518 + ], + "type": "text", + "content": "[72] Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulò, Norman Müller, Matthias Nießner, Angela Dai, and Peter Kontschieder. Panoptic lifting for 3d scene understanding with neural fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9043-9052, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 520, + 545, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 520, + 545, + 585 + ], + "spans": [ + { + "bbox": [ + 312, + 520, + 545, + 585 + ], + "type": "text", + "content": "[73] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerf-player: A streamable dynamic scene representation with decomposed neural radiance fields. IEEE Transactions on Visualization and Computer Graphics (TVCG), 29(5): 2732-2742, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 586, + 545, + 639 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 586, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 312, + 586, + 545, + 639 + ], + "type": "text", + "content": "[74] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5459-5469, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 641, + 545, + 707 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 641, + 545, + 707 + ], + "spans": [ + { + "bbox": [ + 312, + 641, + 545, + 707 + ], + "type": "text", + "content": "[75] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8248–8258, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 708, + 545, + 729 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 708, + 545, + 729 + ], + "spans": [ + { + "bbox": [ + 312, + 708, + 545, + 729 + ], + "type": "text", + "content": "[76] Jiaxiang Tang, Xiaokang Chen, Jingbo Wang, and Gang Zeng. Point scene understanding via disentangled instance" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 757, + 317, + 765 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 757, + 317, + 765 + ], + "spans": [ + { + "bbox": [ + 294, + 757, + 317, + 765 + ], + "type": "text", + "content": "19677" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 59, + 288, + 730 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 72, + 59, + 287, + 79 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 59, + 287, + 79 + ], + "spans": [ + { + "bbox": [ + 72, + 59, + 287, + 79 + ], + "type": "text", + "content": "mesh reconstruction. European Conference on Computer Vision (ECCV), 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 82, + 287, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 82, + 287, + 137 + ], + "spans": [ + { + "bbox": [ + 53, + 82, + 287, + 137 + ], + "type": "text", + "content": "[77] Jonathan Taylor, Jamie Shotton, Toby Sharp, and Andrew Fitzgibbon. The vitruvian manifold: Inferring dense correspondences for one-shot human pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 103-110. IEEE, 2012. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 138, + 288, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 138, + 288, + 203 + ], + "spans": [ + { + "bbox": [ + 54, + 138, + 288, + 203 + ], + "type": "text", + "content": "[78] Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srinivasan, Edgar Tretschk, Wang Yifan, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lombardi, et al. Advances in neural rendering. In Computer Graphics Forum, pages 703-735. Wiley Online Library, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 205, + 288, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 205, + 288, + 270 + ], + "spans": [ + { + "bbox": [ + 54, + 205, + 288, + 270 + ], + "type": "text", + "content": "[79] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 12959-12970, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 271, + 287, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 271, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 53, + 271, + 287, + 335 + ], + "type": "text", + "content": "[80] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481-5490. IEEE, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 338, + 288, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 338, + 288, + 381 + ], + "spans": [ + { + "bbox": [ + 53, + 338, + 288, + 381 + ], + "type": "text", + "content": "[81] Feng Wang, Sinan Tan, Xinghang Li, Zeyue Tian, and Huaping Liu. Mixed neural voxels for fast multi-view video synthesis. IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 384, + 287, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 384, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 53, + 384, + 287, + 449 + ], + "type": "text", + "content": "[82] Liao Wang, Jiakai Zhang, Xinhang Liu, Fuqiang Zhao, Yanshun Zhang, Yingliang Zhang, Minye Wu, Jingyi Yu, and Lan Xu. Fourier plenoctrees for dynamic radiance field rendering in real-time. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13524-13534, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 450, + 287, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 450, + 287, + 494 + ], + "spans": [ + { + "bbox": [ + 53, + 450, + 287, + 494 + ], + "type": "text", + "content": "[83] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing (TIP), 13(4):600-612, 2004. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 495, + 287, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 495, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 53, + 495, + 287, + 559 + ], + "type": "text", + "content": "[84] Ziyan Wang, Giljoo Nam, Tuur Stuyck, Stephen Lombardi, Chen Cao, Jason Saragih, Michael Zollhöfer, Jessica Hodgins, and Christoph Lassner. Neuwigs: A neural dynamic model for volumetric hair capture and animation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8641-8651, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 562, + 288, + 605 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 562, + 288, + 605 + ], + "spans": [ + { + "bbox": [ + 53, + 562, + 288, + 605 + ], + "type": "text", + "content": "[85] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 607, + 287, + 661 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 607, + 287, + 661 + ], + "spans": [ + { + "bbox": [ + 53, + 607, + 287, + 661 + ], + "type": "text", + "content": "[86] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P Srinivasan, Dor Verbin, Jonathan T Barron, Ben Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. arXiv preprint arXiv:2312.02981, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 662, + 287, + 706 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 662, + 287, + 706 + ], + "spans": [ + { + "bbox": [ + 53, + 662, + 287, + 706 + ], + "type": "text", + "content": "[87] Jamie Wynn and Daniyar Turmukhambetov. Diffusionerf: Regularizing neural radiance fields with denoising diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4180-4189, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 708, + 287, + 730 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 708, + 287, + 730 + ], + "spans": [ + { + "bbox": [ + 53, + 708, + 287, + 730 + ], + "type": "text", + "content": "[88] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 59, + 545, + 729 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 331, + 59, + 545, + 80 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 59, + 545, + 80 + ], + "spans": [ + { + "bbox": [ + 331, + 59, + 545, + 80 + ], + "type": "text", + "content": "video. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9421-9431, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 82, + 545, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 82, + 545, + 124 + ], + "spans": [ + { + "bbox": [ + 312, + 82, + 545, + 124 + ], + "type": "text", + "content": "[89] Zhen Xu, Sida Peng, Haotong Lin, Guangzhao He, Jiaming Sun, Yujun Shen, Hujun Bao, and Xiaowei Zhou. 4k4d: Real-time 4d view synthesis at 4k resolution. arXiv preprint arXiv:2310.11448, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 126, + 545, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 126, + 545, + 170 + ], + "spans": [ + { + "bbox": [ + 312, + 126, + 545, + 170 + ], + "type": "text", + "content": "[90] Zhiwen Yan, Chen Li, and Gim Hee Lee. Nerf-ds: Neural radiance fields for dynamic specular objects. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8285-8295, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 171, + 545, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 171, + 545, + 214 + ], + "spans": [ + { + "bbox": [ + 312, + 171, + 545, + 214 + ], + "type": "text", + "content": "[91] Bo Yang, Jianan Wang, Ronald Clark, Qingyong Hu, Sen Wang, Andrew Markham, and Niki Trigoni. Learning object bounding boxes for 3d instance segmentation on point clouds, 2019. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 216, + 545, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 216, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 312, + 216, + 545, + 248 + ], + "type": "text", + "content": "[92] Jinyu Yang, Mingqi Gao, Zhe Li, Shang Gao, Fangjing Wang, and Feng Zheng. Track anything: Segment anything meets videos, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 250, + 545, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 250, + 545, + 303 + ], + "spans": [ + { + "bbox": [ + 312, + 250, + 545, + 303 + ], + "type": "text", + "content": "[93] Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Improving few-shot neural rendering with free frequency regularization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8254-8263, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 305, + 545, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 305, + 545, + 336 + ], + "spans": [ + { + "bbox": [ + 312, + 305, + 545, + 336 + ], + "type": "text", + "content": "[94] Zongxin Yang and Yi Yang. Decoupling features in hierarchical propagation for video object segmentation, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 339, + 545, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 339, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 312, + 339, + 545, + 382 + ], + "type": "text", + "content": "[95] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 384, + 545, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 384, + 545, + 428 + ], + "spans": [ + { + "bbox": [ + 312, + 384, + 545, + 428 + ], + "type": "text", + "content": "[96] Zeyu Yang, Hongye Yang, Zijie Pan, Xiatian Zhu, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. arXiv preprint arXiv:2310.10642, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 430, + 545, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 430, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 312, + 430, + 545, + 483 + ], + "type": "text", + "content": "[97] Zhengming Yu, Wei Cheng, Xian Liu, Wayne Wu, and Kwan-Yee Lin. Monohuman: Animatable human neural field from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16943-16953, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 484, + 545, + 539 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 484, + 545, + 539 + ], + "spans": [ + { + "bbox": [ + 312, + 484, + 545, + 539 + ], + "type": "text", + "content": "[98] Wentao Yuan, Zhaoyang Lv, Tanner Schmidt, and Steven Lovegrove. Star: Self-supervised tracking and reconstruction of rigid objects in motion with neural rendering. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13144-13152, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 540, + 545, + 594 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 540, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 312, + 540, + 545, + 594 + ], + "type": "text", + "content": "[99] Jiakai Zhang, Xinhang Liu, Xinyi Ye, Fuqiang Zhao, Yanshun Zhang, Minye Wu, Yingliang Zhang, Lan Xu, and Jingyi Yu. Editable free-viewpoint video using a layered neural representation. ACM Transactions on Graphics (TOG), 40(4):1-18, 2021. 1, 2, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 596, + 545, + 651 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 596, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 307, + 596, + 545, + 651 + ], + "type": "text", + "content": "[100] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 652, + 545, + 695 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 652, + 545, + 695 + ], + "spans": [ + { + "bbox": [ + 307, + 652, + 545, + 695 + ], + "type": "text", + "content": "[101] Renrui Zhang, Zhengkai Jiang, Ziyu Guo, Shilin Yan, Junting Pan, Hao Dong, Peng Gao, and Hongsheng Li. Personalize segment anything model with one shot. arXiv preprint arXiv:2305.03048, 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 696, + 545, + 729 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 696, + 545, + 729 + ], + "spans": [ + { + "bbox": [ + 308, + 696, + 545, + 729 + ], + "type": "text", + "content": "[102] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J Davison. In-place scene labelling and understanding with implicit scene representation. In IEEE/CVF" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 757, + 318, + 765 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 757, + 318, + 765 + ], + "spans": [ + { + "bbox": [ + 294, + 757, + 318, + 765 + ], + "type": "text", + "content": "19678" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 71, + 59, + 287, + 81 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 59, + 287, + 81 + ], + "spans": [ + { + "bbox": [ + 71, + 59, + 287, + 81 + ], + "type": "text", + "content": "International Conference on Computer Vision (ICCV), pages 15838-15847, 2021. 2, 3" + } + ] + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 757, + 317, + 765 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 757, + 317, + 765 + ], + "spans": [ + { + "bbox": [ + 294, + 757, + 317, + 765 + ], + "type": "text", + "content": "19679" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/GenFlow_ Generalizable Recurrent Flow for 6D Pose Refinement of Novel Objects/ff0e87f9-2cf2-45f0-8b79-ad267112b5a6_content_list.json b/2024/GenFlow_ Generalizable Recurrent Flow for 6D Pose Refinement of Novel Objects/ff0e87f9-2cf2-45f0-8b79-ad267112b5a6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..80deea2c187ac5b98c6efe7f3d28813179903cf5 --- /dev/null +++ b/2024/GenFlow_ Generalizable Recurrent Flow for 6D Pose Refinement of Novel Objects/ff0e87f9-2cf2-45f0-8b79-ad267112b5a6_content_list.json @@ -0,0 +1,1588 @@ +[ + { + "type": "text", + "text": "GenFlow: Generalizable Recurrent Flow for 6D Pose Refinement of Novel Objects", + "text_level": 1, + "bbox": [ + 114, + 128, + 854, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sungphill Moon*", + "bbox": [ + 165, + 203, + 308, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hyeontae Son*", + "bbox": [ + 346, + 203, + 470, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dongcheol Hur", + "bbox": [ + 511, + 204, + 637, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sangwook Kim", + "bbox": [ + 676, + 204, + 802, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "NAVER LABS", + "bbox": [ + 421, + 222, + 545, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{sungphill.moon, son.ht, dongcheol.hur, o.s.w.a.l.k}@naverlabs.com", + "bbox": [ + 189, + 241, + 772, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 312, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the progress of learning-based methods for 6D object pose estimation, the trade-off between accuracy and scalability for novel objects still exists. Specifically, previous methods for novel objects do not make good use of the target object's 3D shape information since they focus on generalization by processing the shape indirectly, making them less effective. We present GenFlow, an approach that enables both accuracy and generalization to novel objects with the guidance of the target object's shape. Our method predicts optical flow between the rendered image and the observed image and refines the 6D pose iteratively. It boosts the performance by a constraint of the 3D shape and the generalizable geometric knowledge learned from an end-to-end differentiable system. We further improve our model by designing a cascade network architecture to exploit the multi-scale correlations and coarse-to-fine refinement. GenFlow ranked first on the unseen object pose estimation benchmarks in both the RGB and RGB-D cases. It also achieves performance competitive with existing state-of-the-art methods for the seen object pose estimation without any fine-tuning.", + "bbox": [ + 75, + 323, + 473, + 641 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 670, + 207, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Estimating accurate 6D object pose from an RGB or RGB-D image is crucial for robotic tasks and augmented reality applications. In the standard setup of 6D object pose estimation, 3D models of target objects are given and available in training and testing. Most state-of-the-art algorithms are learning-based and require training the model per object or dataset containing few objects. However, these methods are difficult to handle novel objects since they memorize the 3D shapes of the targets and suffer from the training cost to support them.", + "bbox": [ + 75, + 695, + 468, + 845 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent works [5, 37, 51, 54] have proposed generalizable models that can estimate the 6D pose of novel objects, i.e.,", + "bbox": [ + 76, + 845, + 468, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/606dece7bed8cf402cafcfdde4e308e20119e9e6352721798c50b3c1c9d2d9f8.jpg", + "image_caption": [ + "Figure 1. Examples of 6D pose estimation of novel objects. Our method estimates correspondences between the input image and rendered image, and 6D object pose." + ], + "image_footnote": [], + "bbox": [ + 519, + 295, + 874, + 417 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "objects unseen during training. Specifically, OSOP [54] and MegaPose [37] show remarkable performance with the generalizable pose refinement. Their refinement methods have the scalability to novel objects since the networks learn to compare the rendered image and the observed image without memorizing 3D shapes. However, their refiners avail themselves of 3D shape information indirectly. OSOP refiner is trained by the surrogate 2D-2D matching loss, which makes it suboptimal for 6D pose estimation. MegaPose refiner regresses the relative 6D pose between the rendered image and the input, but the regression makes no use of the projective geometry.", + "bbox": [ + 496, + 505, + 890, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our research question is Is there a better design for generalizable pose refinement that takes advantage of shape information?. Recently, 2D optical flow-based methods [20, 31] show noticeable performance on 6D pose estimation. These methods estimate dense 2D-2D correspondences between the rendered image and the observed image. The estimated correspondences are lifted to 2D-3D correspondences on the 3D shape of the target object, then the 6D pose is recovered using a PnP solver [31] or pose regressor networks [20]. The optical flow-based pose refinement provides a clue for the generalization due to its robustness to the domain shift [31] and shows the potential to utilize 3D information in the model [20].", + "bbox": [ + 496, + 688, + 892, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose GenFlow, the optical flow-based iterative", + "bbox": [ + 519, + 885, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*These authors contributed equally.", + "bbox": [ + 99, + 887, + 290, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10039", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "refinement for the generalizable object pose estimation guided by the 3D shape information of target objects. Given an image and initial 6D pose of an object in the image, GenFlow estimates 2D dense correspondences between the rendered image and the input image and refined 6D pose through a PnP layer. This series of processes is performed iteratively, so it can improve accuracy and robustness to outliers [43].", + "bbox": [ + 75, + 90, + 468, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We adopt the shape-constraint recurrent flow framework inspired by [20] for our model to use 3D information. This framework utilizes the shape information implicitly to enjoy both generalization and performance. Its end-to-end differentiable system learns to estimate the optical flow, confidence, and object pose. These estimations are updated iteratively complying with the 3D shape. We further introduce a cascade network design to make use of the multi-scale correlation volumes. It takes advantage of the coarse-to-fine refinement and helps the model estimate a more accurate pose.", + "bbox": [ + 75, + 212, + 470, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We validate our method on various datasets provided in the BOP challenge [28]. Our model is designed to estimate 6D pose for the RGB input but is also applicable to the RGB-D case with the RANSAC-Kabsch algorithm [18, 34]. For both RGB and RGB-D cases, our method achieves the overall best performance for the unseen object pose estimation task using the default detection [50] provided in the BOP challenge 2023. It is also notable that our method achieves competitive performance with the state-of-the-art methods for seen objects when used with better detection results.", + "bbox": [ + 75, + 378, + 468, + 544 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 558, + 218, + 573 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Iterative Refinement DeepIM [41] pioneered the iterative render-and-compare strategy for 6D pose estimation. It predicts the relative pose that better matches a rendered image against the observed image. By iterative refinement, the rendered and observed images become more similar. CosyPose [38] improved the idea by introducing better parameterization for rotation [69], network architecture [60], loss design [55, 62] and data augmentation. These iterative methods show accurate results and the possibility of being applied to the unseen objects [41]. In contrast to the previous works, CIR [43] introduces a novel differentiable PnP layer to utilize the projective geometry with deep learning. Unlike CIR, which assumes the RGB-D input, ours focuses on the RGB input. For the RGB input case, CIR obtains the rendered depth as the substitution of input depth for every flow update, so the inference suffers from time inefficiency. On the other hand, our refinement method requires fewer renderings.", + "bbox": [ + 75, + 583, + 468, + 853 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Optical Flow Estimation The classical approaches [2, 6, 29] dealt with optical flow estimation as a hand-crafted optimization problem in pursuit of the balance of bright-", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ness consistency and motion plausibility. Deep learning methods [12, 32, 33, 57, 66] emerged as promising alternatives of the classical methods, achieving considerable performance with faster inference times. Recently, RAFT [61] introduced a recurrent deep network architecture which has shown significant improvement over the previous methods regarding accuracy, efficiency, and generalization. PFA [31] applied the optical flow estimation to 6D pose refinement for data-limited cases. Since learning to estimate dense optical flow helps the networks to focus on the lower-level information, it could be robust to domain shift. Following PFA, SCFlow [20] further improved the performance by introducing an end-to-end trainable system using a pose regressor and shape-constraint lookup. Our method capitalizes on a differentiable PnP layer to the framework in order to make use of projective geometry.", + "bbox": [ + 496, + 90, + 890, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unseen Object Pose Estimation Category-level 6D pose estimation [7, 40, 48, 62] tackled the case of unseen objects of which category is given in advance. Since these methods depend on the prior of the 3D shapes belonging to specific category, they struggle with handling the objects of unknown categories. Recently, several works proposed class-agnostic generalizable pose estimators that can handle unseen objects. A few works attempted the CAD model-free pose estimation [23, 24, 45, 53, 58] to cover everyday objects, but the results were only evaluated on datasets without occlusion and cluttered scenes. [49, 59, 64, 65] focused on estimating the orientation of novel objects by comparing the deep image features of the rendered image and the observed image. ZePHyR [51] proposed a generalizable scoring function combined with hand-crafted features [15, 47]. However, it uses the features selectively according to the richness of texture in the datasets. ZeroPose [5] estimates 6D pose by hierarchical feature matching which requires the depth input by its nature and pose refinement for better performance. Like ours, OSOP [54] and MegaPose [37] proposed CAD-based pose estimation methods that can process RGB and RGB-D inputs. Specifically, MegaPose achieved state-of-the-art performance by its iterative pose refiner. The pose refiner leverages the multi-view rendered images containing surface normals for utilizing rich information about the novel object's shape, anchor point, and coordinate system. Our method refines the pose with the guidance of the target object's shape without rendering multiview images per refinement step.", + "bbox": [ + 496, + 333, + 890, + 770 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 784, + 589, + 799 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This section describes our methodology for 6D pose estimation of novel objects. Given an RGB image and 3D models of target objects, which are unseen during training, our goal is to estimate the 6D poses of the objects with respect to the camera. Our pose estimation pipeline consists of 3 stages: object detection, coarse pose estimation, and", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10040", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e0d38f96da9325429b40f9986f7b3827c355feef3851a1f5ea20f6bdaffe6686.jpg", + "image_caption": [ + "(a) Naïve Sampling $(N = 576)$" + ], + "image_footnote": [], + "bbox": [ + 183, + 94, + 367, + 160 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/aa3ec160f48a69d74c34ece7005891027a3e8024f85f54a2ba9c7fab38ebd3ad.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 401, + 148, + 459, + 183 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fe6f8b1333f676f5f84571ce1c3e959148008c987e410a03443a5ff38a2b045d.jpg", + "image_caption": [ + "(b) Sampling using GMM $(M = 104)$" + ], + "image_footnote": [], + "bbox": [ + 86, + 188, + 269, + 255 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a8cca536b3a41d16a71c1c7c5b60f623554edb6dfe5c3b7167e90860d9de4301.jpg", + "image_caption": [ + "Figure 2. Visualization of the scores predicted by the coarse model. Each dot denotes the sampled rotation, and its size represents the relative score of the coarse model for the corresponding rotation. The green circles are boundaries of the area where the refiner is interested. Figures show that the high-scoring poses are clustered in the vicinity of the green circles." + ], + "image_footnote": [], + "bbox": [ + 277, + 188, + 460, + 255 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "pose refinement. Note that the detection of novel objects is out of the scope of this paper. We first provide an overview of the whole process and then focus on the pose refinement method, which is the main contribution of this work.", + "bbox": [ + 75, + 398, + 468, + 458 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 468, + 217, + 483 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a calibrated RGB image $\\mathcal{I}$ and a set of 3D models of the target objects $\\{\\mathcal{M}_i\\}$ , we estimate the 6D poses of the visible objects in the image. A 6D pose is defined by a matrix $\\mathbf{P}$ representing the rigid transformation from the object space to the camera space. Note that we assume the input image is calibrated, i.e., the intrinsic camera matrix $\\mathbf{K}$ is known.", + "bbox": [ + 75, + 491, + 468, + 597 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2D Object Detection Given an RGB image, the object detection method produces a list of object detections consisting of 2D bounding boxes and associated labels. For the unseen object detection, we apply the CNOS [50] detection leveraging the foundation models [36, 52]. To show the extensive results of our work, we also apply the YOLOX detector [19] from GDRNPP [44] which is used for seen object detection. For each detection, the image $\\mathcal{I}$ is cropped and further resized to the specific resolution in the following steps. The intrinsic parameters are adjusted according to the resized resolution. For convenience, the following explanation focuses on the 6D pose estimation of a single detected object of which the 3D model is $\\mathcal{M}$ .", + "bbox": [ + 75, + 598, + 468, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Coarse Pose Estimation Given object detection, the coarse pose estimator predicts rough initial poses. We used the classification-based method proposed in [37] for better generalization. Given an input image and an arbitrary pose $\\mathbf{P}$ for the 3D model $\\mathcal{M}$ , the coarse model classifies whether or not the pose $\\mathbf{P}$ could be improved to the pose in proximity to the ground-truth by the refiner. It uses a pre-defined", + "bbox": [ + 75, + 795, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "set of 3D rotations of which the cardinality is $N$ to generate the various arbitrary pose hypotheses. For each rotation, a rough 3D translation is estimated for the points of the object's 3D model to be approximately inside the 2D bounding box. We extract scores of the pose hypotheses from the model and select the top- $n$ scoring poses. Such a classification-based estimation can handle novel objects, but it suffers from the cost of rendering a large number of images for the pre-defined rotations. Inspired by the property of high-scoring rotations to be sparse and clustered as shown in Fig 2, we introduce sampling using a Gaussian mixture model (GMM) [1] for efficiency. First, we choose the top $k$ -scoring samples $\\{R_i = (\\psi_i, \\theta_i, \\phi_i)\\}_{i=1}^k$ from $M$ pre-defined rotations (Fig 2b left). Then we create a density distribution $p(X)$ for the Euler angle $X$ using a GMM which satisfies that", + "bbox": [ + 496, + 90, + 890, + 332 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np (X) = \\sum_ {i = 1} ^ {k} \\frac {1}{k} \\mathcal {N} (X | \\mu_ {i}, \\Sigma), \\mu_ {i} = R _ {i} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 342, + 890, + 382 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Finally, we sample additional $M$ rotations from the density distribution $p(X)$ (Fig 2b right) and choose the best-scoring $n$ hypotheses among the $2M$ samples. This method renders $2M < N$ images while achieving better performance. We used $M = 104$ and $k = 16$ for our experiments.", + "bbox": [ + 496, + 391, + 890, + 465 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pose Refinement Given a coarse pose estimate, we refine the pose for the rendered image to match against the input image crop. Starting from the coarse estimate $\\mathbf{P}_0$ as an initial pose, we iteratively refine the pose using our refiner. We denote the $k^{th}$ 6D pose estimate of the refiner as $\\mathbf{P}_k$ . For the $k^{th}$ iteration, we obtain an image $\\mathcal{I}_{r,k}$ by rendering the 3D model $\\mathcal{M}$ for the pose $\\mathbf{P}_{k-1}$ . Our refiner then estimates the 2D optical flow $\\mathbf{F}_k$ between the image $\\mathcal{I}_{r,k}$ and the input image crop and the confidence weights $\\mathbf{W}_k$ which are used to solve the pose $\\mathbf{P}_k$ . A detailed description is in the following section.", + "bbox": [ + 496, + 467, + 890, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. GenFlow Refinement", + "text_level": 1, + "bbox": [ + 500, + 642, + 697, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.1 Overview", + "text_level": 1, + "bbox": [ + 500, + 665, + 620, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We describe the specific operation of GenFlow refiner to estimate the refined pose $\\mathbf{P}_k$ from the previous result $\\mathbf{P}_{k - 1}$ but hereafter, we omit the $k$ notation if possible for convenience. First, we obtain a synthetic image by rendering the 3D model $\\mathcal{M}$ for the pose $\\mathbf{P}$ and adjusted intrinsic parameters. Specifically, the results of the rendering consists of the RGB image, depth image and binary rendered mask, denoted by $\\mathcal{I}_r\\in R^{H\\times W\\times 3}$ , $\\mathcal{D}_r\\in R^{H\\times W}$ , and $\\mathcal{M}_r\\in \\{0,1\\}^{H\\times W}$ respectively. Then, we extract features from both images and refine the pose with the GenFlow module.", + "bbox": [ + 496, + 688, + 890, + 853 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Feature Extraction We extract the image features from the input image crop and the rendered image $\\mathcal{I}_r$ with the same feature encoder and the context features from", + "bbox": [ + 496, + 854, + 890, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10041", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/53dcdc771a5c4b98eb9305338aa5bf17832de51b7f7e3bdb5588082b0b881c14.jpg", + "image_caption": [ + "Figure 3. Overview of GenFlow refinement. We visualize the process of feature extraction and GenFlow update in the $k^{th}$ refinement. $T_{C2O}$ refers to a transformation that maps the coordinates from camera space to object space." + ], + "image_footnote": [], + "bbox": [ + 102, + 85, + 890, + 441 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the $\\mathcal{I}_r$ . The feature maps from the $i^{th}$ layer of the feature encoder are used to construct the correlation volumes $\\mathbf{C}_i \\in R^{H_i \\times W_i \\times H_i \\times W_i}$ by the dot product.", + "bbox": [ + 75, + 508, + 468, + 553 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "GenFlow Module For each correlation volume and context feature map, we construct a module named \"GenFlow module\". A GenFlow module consists of a convolution-based Gated Recurrent Unit (GRU) [9], a differentiable PnP solver, and correlation lookup based on the pose-induced flow as shown in the Fig 3. GenFlow module iteratively updates the optical flow, confidence weights, and 6D pose. For the $j^{th}$ update, the output of the module is the refined 6D pose $\\mathbf{P}^j$ . Note that we use subscripts for the outer updates and superscripts for the inner updates, i.e., GenFlow updates. The detailed illustration of updating the 6D pose with a GenFlow module for $i^{th}$ correlation volume is as follows.", + "bbox": [ + 75, + 554, + 470, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2 GenFlow Update", + "text_level": 1, + "bbox": [ + 76, + 768, + 250, + 784 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "GRU Update For the $j^{th}$ update, the GRU of the GenFlow module consumes the previous hidden state $h^{j-1}$ , correlation features $r^{j-1}$ and context features. The output of the GRU $\\{\\mathbf{h}^j, \\mathbf{f}^j, \\mathbf{u}^j, \\mathbf{c}^j, \\mathbf{s}^j\\}$ includes the updated hidden state $h^j$ , optical flow $\\mathbf{f}^j$ , flow upsampling mask $\\mathbf{u}^j$ , certainty $\\mathbf{c}^j$ and pose sensitivity $\\mathbf{s}^j$ . The flow $\\mathbf{f}^j \\in R^{H_i \\times W_i \\times 2}$ is upsampled to $\\mathbf{F}^j$ to match the resolution of previous $(i-1)^{th}$", + "bbox": [ + 75, + 792, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "feature map using the mask $\\mathbf{u}^j$ . We apply the convex combination method of 3x3 neighbors for the optical flow proposed in [61]. The certainty $\\mathbf{c}^j$ and pose sensitivity $\\mathbf{s}^j$ are upsampled by bilinear interpolation, and then we multiply them to obtain the confidence weights $\\mathbf{W}^j$ . The head networks for certainty and pose sensitivity are supervised by different losses, respectively. This explicit factorization of the confidence weights helps to be robust to occlusion with the help of certainty estimation as shown in Fig 4. The upsampled flow $\\mathbf{F}^j$ and confidence weights $\\mathbf{W}^j$ are used to establish weighted 2D-3D correspondences.", + "bbox": [ + 496, + 508, + 892, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pose Update As we know the depth of synthetic image $\\mathcal{D}_r$ for the pose $\\mathbf{P}_k^0 = [\\mathbf{R}_k^0 |\\mathbf{t}_k^0 ]$ and the intrinsic camera matrix $\\mathbf{K}$ , we can compute the object-space 3D coordinate of each pixel by", + "bbox": [ + 496, + 674, + 890, + 736 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left( \\begin{array}{l} x \\\\ y \\\\ z \\end{array} \\right) = \\mathbf {R} _ {k} ^ {\\mathrm {T}} \\left(\\mathbf {K} ^ {- 1} \\mathcal {D} _ {r} (u, v) \\left( \\begin{array}{l} u \\\\ v \\\\ 1 \\end{array} \\right) - \\mathbf {t} _ {k}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 739, + 890, + 789 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $(u,v)$ is the 2D coordinate of a pixel and the $\\mathcal{D}_r(u,v)$ is depth of the pixel. Through this lifting process, we obtain the 2D-3D correspondences for the flow $\\mathbf{F}^j$ . We use $\\mathbf{W}^j$ for the confidence of correspondences and binary rendered mask $\\mathcal{M}_r$ to consider the object region only.", + "bbox": [ + 496, + 794, + 890, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To obtain the 6D pose $\\mathbf{P}^j$ , we utilize the iterative PnP solver based on the Levenberg-Marquardt (LM) algorithm.", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10042", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0931a216aea849aa4cbfa6888d19dcc1afbd0d2381b50b917f04b5efd6ec7be1.jpg", + "image_caption": [ + "Figure 4. Visualization of the outputs of GRU. Our method factorizes the confidence weights into two terms: certainty and pose sensitivity. Certainty estimation helps to the robustness to occlusion. Pose sensitivity highlights the rich texture regions and extremities of the object." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 467, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It optimizes the 6D pose by minimizing the sum of squared weighted reprojection error:", + "bbox": [ + 76, + 387, + 468, + 417 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\underset {R, t} {\\operatorname {a r g m i n}} \\frac {1}{2} \\sum_ {u} \\sum_ {v} \\| \\mathbf {W} ^ {j} (u, v) \\times (\\pi (R \\left( \\begin{array}{l} x \\\\ y \\\\ z \\end{array} \\right) + t) \\tag {3} \\\\ - (\\left( \\begin{array}{c} u \\\\ v \\end{array} \\right) + \\mathbf {F} ^ {j} (u, v))) \\| ^ {2}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 428, + 468, + 512 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pi$ is the projection function with the intrinsic camera parameters and $(x,y,z)^{\\mathrm{T}}$ is the 3D coordinate of the pixel $(u,v)$ computed by the equation 2. Due to its differentiable nature, the PnP layer enables end-to-end training. However, the backpropagation by unrolling the iterative optimization causes numerical instability, making the training process hard. For stable end-to-end training, we introduce the regularization method [4], which backpropagates the gradient through the last step of a differentiable optimization. Specifically, we first obtain a detached pose $\\mathbf{P}^*$ through 3 LM steps and update the pose via another iteration of the Gauss-Newton algorithm to obtain the pose $\\mathbf{P}^j$ . Note that the gradient is not backpropagated through $\\mathbf{P}^*$ .", + "bbox": [ + 75, + 523, + 468, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Correlation Lookup We compute the pose-induced flow, which is the 2D optical flow between the $\\mathcal{I}_r$ and the image with respect to the pose $\\mathbf{P}^j$ . The pose-induced flow is calculated as the displacement of each pixel in the $\\mathcal{I}_r$ by lifting and reprojecting the pixel using the depth $\\mathcal{D}_r$ . To embed the 3D shape of the target object in the lookup operation [20, 43], we use the pose-induced flow rather than the estimated flow $\\mathbf{f}^j$ for indexing the correlation volume $\\mathbf{C}$ . This way of imposing a shape constraint on the model induces the networks to predict dense matching to comply with the target's 3D shape information. From the lookup operation, we obtain the correlation feature $r^j$ , which is the input of", + "bbox": [ + 75, + 719, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/34a61526b8c2ffbc7928a0358b70caef434c7ad8c897d3a40e20098ab15be32f.jpg", + "image_caption": [ + "Figure 5. Cascade design of GenFlow modules. Multiple GenFlow modules are attached to each level on the feature pyramid. The last updated flow $\\mathbf{F}_k^M$ and 6D pose $\\mathbf{P}_k^M$ from the higher-level GenFlow module are used to initialize the flow and 6D pose of the lower-level module. With the cascade architecture, the 6D pose is recovered in a coarse to fine manner." + ], + "image_footnote": [], + "bbox": [ + 504, + 90, + 890, + 237 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "GRU for the next GenFlow update.", + "bbox": [ + 500, + 363, + 732, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.3 Cascade Architecture", + "text_level": 1, + "bbox": [ + 500, + 400, + 705, + 414 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inspired by the previous methods that exploit the multiscale features [8, 42], we design a cascade architecture using multiple GenFlow modules. Given correlation volumes $\\{\\mathbf{C}_i\\}$ , we assign a GenFlow module to each volume. Then, we use the 2D optical flow and refined 6D pose of the last inner updates to initialize those of the next module as shown in Fig 5. This cascade architecture takes advantage of progressive refinement, which makes use of the high-level semantics to the low-level fine details. We found that it improves the accuracy of 6D pose rather than using a single GenFlow module.", + "bbox": [ + 496, + 426, + 890, + 592 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Training", + "text_level": 1, + "bbox": [ + 500, + 604, + 604, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset Both the coarse and refiner models are trained using the large dataset provided by MegaPose [37]. It is a large synthetic dataset comprising 2 million RGB-D images generated by BlenderProc [11], which contains diverse 3D models with ground-truth 6D pose annotations. The 3D shapes are collected from ShapeNet [3] and Google Scanned Object (GSO) [14] datasets, and we trained our model on both of the datasets.", + "bbox": [ + 496, + 628, + 890, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Refiner Model Our refiner model is trained similarly to the previous iterative refinement methods [37, 38]. Given an image with the ground-truth 6D pose $\\mathbf{P}_{gt}$ for the 3D model $\\mathcal{M}$ , we perturb the pose to generate an initial pose $\\mathbf{P}_{init}$ by applying random noises to the translation and rotation of the pose respectively. Translation noise is sampled from a normal distribution with a standard deviation of (0.01, 0.01, 0.05) centimeters, and rotation noise is sampled from another normal distribution of Euler angles with a standard deviation of 15 degrees in each axis. The refiner model", + "bbox": [ + 496, + 750, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10043", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "learns to predict the pose $\\mathbf{P}_{gt}$ from the perturbed initial pose $\\mathbf{P}_{init}$ .", + "bbox": [ + 76, + 90, + 468, + 119 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Coarse Model Given an input RGB image $\\mathcal{I}$ and a pose $\\mathbf{P}$ , a coarse model is trained to classify whether the refiner could estimate the 6D pose in proximity to the ground truth pose from the initial pose $\\mathbf{P}$ . We randomly generate the positive and negative pose samples and train the network to classify them using the binary cross entropy loss. For the positive samples, we use the same method used to generate the perturbed initial pose in the refiner model. Negative samples are generated using the method proposed in [37].", + "bbox": [ + 75, + 122, + 468, + 258 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 273, + 294, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Coarse Estimation We use $224 \\times 224$ resolution for the resized input crop and the synthetic image $\\mathcal{I}_r$ . The coarse model is implemented using the ConvNeXt backbone [46] and fully connected head. We used the ConvNext model pre-trained on ImageNet [10] classification to initialize the networks' weights. For inference, we select top- $n$ hypotheses using our GMM-based sampling strategy with $M = 104$ .", + "bbox": [ + 75, + 297, + 468, + 417 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "GenFlow Refinement First, we obtain a synthetic image $\\mathcal{I}_r$ of which resolution is $256\\times 256$ and an image crop resized to the same resolution as the synthetic image with adjusted intrinsic parameters. Like the coarse model, we use the pre-trained ConvNeXt for the feature and context encoders. The feature maps from the first two layers of the feature encoder are used to construct correlation volumes $\\mathbf{C}_1\\in R^{\\frac{H}{4}\\times \\frac{W}{4}\\times \\frac{H}{4}\\times \\frac{W}{4}}$ and $\\mathbf{C}_2\\in R^{\\frac{H}{8}\\times \\frac{W}{8}\\times \\frac{H}{8}\\times \\frac{W}{8}}$ . Our GRU is implemented with a ConvGRU unit [61] and following four CNN-based heads for the 2D optical flow $\\mathbf{f}$ , flow upsampling mask $\\mathbf{u}$ , certainty $\\mathbf{c}$ and pose sensitivity $\\mathbf{s}$ . Two GenFlow modules are attached to the correlation volumes for the cascade architecture. They are serially connected so that the estimations from higher-level features are utilized as initial values for the next module.", + "bbox": [ + 75, + 420, + 468, + 645 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The optical flow and its upsampling mask produce the upsampled flow, which is supervised by the L1 endpoint error loss with the ground-truth optical flow. We also utilize the disentangled point matching loss for the pose loss, supervising the rotation, 2D object center, and depth of the object individually. Specifically, given an estimated pose $\\mathbf{P} = [\\mathbf{R}][\\mathbf{t}_x,\\mathbf{t}_y,\\mathbf{t}_z]^{\\mathrm{T}}]$ and ground truth pose $\\mathbf{P}_{gt} = [\\mathbf{R}^{\\star}][\\mathbf{t}_x^{\\star},\\mathbf{t}_y^{\\star},\\mathbf{t}_z^{\\star}]^{\\mathrm{T}}]$ , the pose loss is computed as follows:", + "bbox": [ + 76, + 648, + 468, + 771 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {p o s e} = \\mathcal {D} ([ \\mathbf {R} | [ \\mathbf {t} _ {x} ^ {\\star}, \\mathbf {t} _ {y} ^ {\\star}, \\mathbf {t} _ {z} ^ {\\star} ] ^ {\\mathrm {T}} ], \\mathbf {P} _ {g t}) \\\\ + \\mathcal {D} \\left( \\right.\\left[ \\right. \\mathbf {R} ^ {\\star} \\left. \\right|\\left[ \\mathbf {t} _ {x}, \\mathbf {t} _ {y}, \\mathbf {t} _ {z} ^ {\\star} \\right] ^ {\\mathrm {T}}\\left. \\right), \\mathbf {P} _ {g t}\\left. \\right) \\tag {4} \\\\ + \\mathcal {D} ([ \\mathbf {R} ^ {\\star} | [ \\mathbf {t} _ {x} ^ {\\star}, \\mathbf {t} _ {y} ^ {\\star}, \\mathbf {t} _ {z} ] ^ {\\mathrm {T}} ], \\mathbf {P} _ {g t}) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 796, + 468, + 857 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where the distance $\\mathcal{D}$ between two 6D poses $\\mathbf{P}_1$ and $\\mathbf{P}_2$ using the 3D points $\\mathcal{X}_{\\mathcal{M}}$ on the 3D model $\\mathcal{M}$ is defined as", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} \\left(\\mathbf {P} _ {1}, \\mathbf {P} _ {2}\\right) = \\frac {1}{\\left| \\mathcal {X} _ {\\mathcal {M}} \\right|} \\sum_ {x \\in \\mathcal {X} _ {\\mathcal {M}}} \\left| \\mathbf {P} _ {1} x - \\mathbf {P} _ {2} x \\right| \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 122, + 890, + 160 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To supervise the certainty $\\mathbf{c}$ , we introduce the certainty estimation by classifying the depth consistency similarly to [17]. Given 6D poses $\\mathbf{P}_{init}$ and $\\mathbf{P}_{gt}$ for the 3D model $\\mathcal{M}$ and intrinsic camera parameters $\\mathbf{K}$ , we compute the projected depth $\\mathcal{D}_{r\\rightarrow t}$ by warping the rendered image to the target input image using the depth $\\mathcal{D}_r$ . We also obtain the depth map of warped pixels $\\mathcal{D}_t$ given the synthetic depth of the target image. Then we train the networks by minimizing the binary cross entropy loss $\\mathcal{L}_{cert}$ between $\\mathbf{c}$ and $\\mathbf{c}_{r\\rightarrow t}$ which is a binary mask $\\mathbf{c}_{r\\rightarrow t} = |\\mathcal{D}_{r\\rightarrow t} - \\mathcal{D}_t| < d_{th}$ where $d_{th}$ is the distance threshold. It is worth noting that we place a stop-gradient on the certainty mask $\\mathbf{c}^j$ when computing the pose loss $\\mathcal{L}_{pose}^j$ for disentangling the certainty $\\mathbf{c}$ and pose sensitivity $\\mathbf{s}$ .", + "bbox": [ + 496, + 183, + 890, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given the RGB-D images with annotated ground truth 6D poses, the overall loss is defined as", + "bbox": [ + 498, + 398, + 890, + 428 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\sum_ {j = 1} ^ {N} \\gamma^ {j - N} \\left(\\mathcal {L} _ {\\text {f l o w}} ^ {j} + \\alpha \\mathcal {L} _ {\\text {c e r t}} ^ {j} + \\beta \\mathcal {L} _ {\\text {p o s e}} ^ {j}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 448, + 890, + 491 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "following the strategy of exponentially increasing weights [61] where we use the weight $\\gamma = 0.8$ and the iteration of GenFlow updates $N = 8$ in this work. In specific, the GenFlow module assigned to $\\mathbf{C}_2$ outputs $\\{\\mathbf{P}^j\\}_{j=1}^{N/2}$ and the another module outputs $\\{\\mathbf{P}^j\\}_{j=N/2}^N$ from $\\mathbf{C}_1$ .", + "bbox": [ + 496, + 510, + 890, + 593 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Multi-Hypotheses Strategy We select $n$ best-scoring hypotheses in the coarse estimation. The selected hypotheses are refined, and we choose the best estimate by evaluating the refined hypotheses using the coarse model again. This strategy requires more inference time for evaluating $n$ iterations of the rendering and refinement process, but it can boost the accuracy by avoiding the risk of the \"Winner-Takes-All\" strategy.", + "bbox": [ + 496, + 594, + 890, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "RGB-D Input If the observed image consists of depth information, we apply the depth refinement after the GenFlow loop for every refinement step. First, we obtain the optical flow $\\mathbf{F}$ and upsampled certainty from the results of the last GenFlow update. 3D-3D correspondences are established by the optical flow and the input depth. We filter out the correspondences of certainty lower than a specific threshold to remove the outliers. Then we apply the RANSAC-Kabsch [18, 34] for the depth refinement. Using certainty for filtering out the outliers is more effective than using confidence weights since the high-confidence pixels are sparse and the input depth is prone to be noisy.", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10044", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3ce43fc09fe0b6ed08000c39334057d5da74428f1443de099a7db8c1cd8bee73.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
2D LocalizationPose InitializationPose RefinementAverage Recall (↑)
RGB-D InputMethodNovel ObjectsMethodNovel ObjectsMethodNovel ObjectsLM-O T-LESS TUD-L IC-BIN ITODD HB YCB-VMEAN
1XMask-RCNN [22, 38]XCosyPose [38]XCosyPose [38]X63.364.068.558.321.665.657.457.0
2XMask-RCNN [22, 38]XSurfEmb [21]XBFGSX66.373.571.558.841.379.164.765.0
3XYOLOX [19, 44]XZebraPose [56]X-X72.981.175.659.250.492.172.972.0
4XYOLOX [19, 44]XMegaPose [37]✓MegaPose+MH [37]✓64.878.174.156.942.286.370.267.5
5XYOLOX [19, 44]XOurs✓Ours+MH✓68.382.877.859.650.189.770.871.3
6✓YOLOX [19, 44]XWDR-Pose [30]XPFA [31]+KabschX79.284.996.370.652.686.789.980.0
7✓YOLOX [19, 44]XMegaPose [37]✓MegaPose+MH [37]+Teaserpp [67]✓70.471.891.659.255.387.285.574.4
8✓YOLOX [19, 44]XOurs✓Ours+MH+Kabsch✓74.278.392.864.965.292.088.379.4
9XOSOP [54]✓OSOP [54]✓OSOP+PnP+MH [54]✓31.2----49.233.2-
10XCNOS-det. [50]✓MegaPose [37]✓MegaPose+MH [37]✓56.050.868.741.934.670.662.054.9
11XCNOS-det. [50]✓Ours✓Ours+MH✓57.553.069.145.640.874.563.957.8
12✓OSOP [54]✓OSOP [54]✓OSOP+Kabsch+MH [54]+ICP✓48.2----60.557.2-
13✓CNOS-seg. [50]✓ZeroPose [5]✓MegaPose+MH [37]✓53.840.083.539.252.165.365.357.0
14✓CNOS-det. [50]✓MegaPose [37]✓MegaPose+MH [37]+Teaserpp [67]✓62.648.785.146.746.873.076.462.8
15✓CNOS-det. [50]✓Ours✓Ours+Kabsch+MH✓62.951.785.853.355.978.282.567.2
", + "bbox": [ + 76, + 88, + 919, + 319 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 75, + 397, + 209, + 414 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Datasets and Evaluation Metrics", + "text_level": 1, + "bbox": [ + 75, + 421, + 362, + 436 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Datasets We evaluate our method on seven datasets of the BOP challenge [28]: LM-O [25], T-LESS [26], TUD-L [27], IC-BIN [13], ITODD [16], HomebrewdDB [35] and YCB-V [63]. These datasets contain cluttered real-world scenes captured from different lighting and camera sensors. Each scene consists of multiple objects including occlusions between them. The objects vary in terms of texture richness, symmetry, and usage.", + "bbox": [ + 75, + 444, + 468, + 565 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation Metrics Following the evaluation methodology of BOP challenge [28], we report the average recall (AR) considering three pose-error functions: Visible Surface Discrepancy (VSD), Maximum Symmetry-Aware Surface Distance (MSSD), and Maximum Symmetry-Aware Projection Distance (MSPD). VSD measures the misalignment of the visible part by computing the ratio of pixels in which the difference between the estimated distance to the camera center and ground truth is under a specific tolerance. In consideration of the global symmetry information, MSSD computes the maximum distance between the estimated camera-space coordinates and corresponding ground truth. In contrast, MSPD computes the maximum distance between the estimated image-space coordinates and the corresponding ground truth. Concerning each pose-error function, the AR is obtained by averaging recall calculated for multiple settings of the correctness thresholds (and misalignment tolerances for VSD).", + "bbox": [ + 75, + 566, + 470, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. BOP Benchmark Results", + "text_level": 1, + "bbox": [ + 75, + 847, + 305, + 862 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1 shows the results of our method on the BOP datasets. For a fair comparison, all reported pose initializa", + "bbox": [ + 75, + 869, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/166b6174dd27fa33dfc3ca37065a687781515dec4e61ef5eb3c5c09847e7d3b0.jpg", + "table_caption": [ + "Table 1. 6D pose estimation results on the BOP challenge datasets. We report the Average Recall (AR) scores across the datasets for various methods. The higher score the better. The best results among the comparable methods are in bold. We denote the multi-hypotheses strategy as MH for simplicity." + ], + "table_footnote": [], + "table_body": "
Method # of renderings (↓)Average Recall (↑)
LM-OT-LESSTUD-LIC-BINYCB-VMEAN
Naïve57622.525.630.617.516.422.5
Ours14428.835.533.829.728.931.3
20832.940.440.031.134.735.8
", + "bbox": [ + 501, + 395, + 903, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Ablation study of the hypotheses generation strategy for the coarse estimation. The AR scores are reported for the best pose hypothesis from the candidates from different sampling methods.", + "bbox": [ + 498, + 472, + 890, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tion and pose refinement methods are trained with synthetic data containing images rendered with BlenderProc [11]. We reported the AR scores of MegaPose and ours by setting the number of hypotheses $n$ as 10 for the multi-hypotheses strategy and the number of outer updates as 5. For ours, the number of GenFlow updates per single outer update is set to 8.", + "bbox": [ + 496, + 542, + 890, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Unseen Objects For both RGB input (row 9-11) and RGB-D input (row 12-15), the results show that our approach outperforms all other methods for novel objects.", + "bbox": [ + 496, + 650, + 890, + 695 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Seen Objects From row 1 to 8, the pose estimation results on the 2D detection trained for target datasets are reported. Our method still outperforms MegaPose for the same detection results. It is noticeable that our method provides competitive results to the state-of-the-art methods, though it is not the best, even without fine-tuning on the target datasets.", + "bbox": [ + 496, + 696, + 890, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 800, + 653, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct ablation studies to validate the efficacy of our method. The ablation results are reported for the RGB inputs we mainly focus on. All experiments were performed on 5 BOP datasets: LM-O, T-LESS, TUD-L, IC-BIN, and YCB-V. We used the 2D detection results of YOLOX [19,", + "bbox": [ + 496, + 825, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10045", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a24a26797b84c15bcf8cc0e790a057de4a8b8059235839fac8e418bf973863eb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTrainingAverage Recall (↑)
LM-OT-LESSTUD-LIC-BINYCB-VMEAN
1No shape-constraint + RANSAC-PnP [39]Lflow61.777.672.454.265.766.3
2Shape-constraintLflow, Lpose64.981.875.157.070.169.8
3Shape-constraint + CascadeLflow, Lpose65.782.375.058.969.970.4
4Shape-constraint + Confidence factorizationLflow, Lcert, Lpose64.581.875.357.370.569.9
5Shape-constraint + Confidence factorization + CascadeLflow, Lcert, Lpose65.982.076.159.569.870.7
", + "bbox": [ + 78, + 88, + 883, + 200 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Ablation study of GenFlow design. The best result is in bold. Our method accomplishes the best performance (row 5).", + "bbox": [ + 107, + 209, + 857, + 224 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "44] trained on target datasets to focus on the accuracy of 6D pose estimation by preventing the performance bottleneck caused by the detection methods for novel objects.", + "bbox": [ + 75, + 250, + 468, + 297 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3.1 Coarse Pose Estimation", + "text_level": 1, + "bbox": [ + 76, + 335, + 295, + 349 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2 shows the effectiveness of the GMM-based sampling method for generating pose hypotheses. Pre-defined rotations are generated following [68] for the Naive sampling and for creating GMM of our method. We used the $N = 576$ rotations for the Naive sampling, and $M = 72$ and $M = 104$ which conduct $2M$ renderings (144 and 208, respectively) for ours. The results show that our method can generate a better initial pose despite the fewer renderings and network inferences.", + "bbox": [ + 75, + 366, + 470, + 502 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3.2 GenFlow design", + "text_level": 1, + "bbox": [ + 76, + 541, + 243, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3 shows the results on different refiner designs. All experiments are conducted using the same result of coarse estimation, which uses our GMM-based sampling with $M = 104$ without the multi-hypotheses strategy. We run 5 iterations of GenFlow refinement, each containing the 8 iterations of the inner update. We evaluated the final 6D pose from the GenFlow refiner.", + "bbox": [ + 75, + 571, + 468, + 678 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Shape-Constraint Learning to estimate 2D correspondences with flow loss only is suboptimal for 6D pose estimation since it makes no use of the target shape (row 1). Imposing shape-constraint on the model benefits from the guide of 3D information without loss of generalization performance (row 2-5).", + "bbox": [ + 75, + 681, + 468, + 772 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Cascade Architecture The proposed cascade architecture with multiple GenFlow modules contributes to the performance improvement (row 3, 5).", + "bbox": [ + 75, + 775, + 468, + 821 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Confidence Factorization Factorizing the confidence weights to certainty and pose sensitivity helps to increase the accuracy of the output pose. Certainty estimation leverages the training depth data and makes our method robust to occlusion (row 4, 5).", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0da2290d5dcb4220551cdb25a3c6c9e59c5c1292f6be71943e10efbf3983b475.jpg", + "image_caption": [ + "Figure 6. GenFlow vs MegaPose refiner for the RGB inputs." + ], + "image_footnote": [], + "bbox": [ + 506, + 252, + 885, + 411 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3.3 MegaPose Refiner vs Ours", + "text_level": 1, + "bbox": [ + 500, + 468, + 738, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We compare the performance of the MegaPose refiner and ours for the RGB input. The same pose initializations from our coarse estimation were used for a fair evaluation. We reported the changes in AR score according to the iteration of the outer updates for both methods in Fig 6. The reported score is computed as the arithmetic mean of the AR scores for the five datasets. A single outer update takes 66.5 milliseconds for the MegaPose refiner and 98.2 milliseconds for ours, respectively, on an RTX 3090 GPU and Core i9-10920X CPU. Although ours is slower than the MegaPose refiner for a single refinement step, it achieves better performance for the same and even fewer iterations. Note that ours performs effectively by comparing a single-view rendered image for an outer update, while MegaPose refiner uses multiple rendered views as input.", + "bbox": [ + 496, + 493, + 890, + 720 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 736, + 619, + 752 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present GenFlow, an optical flow-based method for 6D pose refinement of novel objects. First, we generate pose hypotheses using a GMM-based sampling strategy for efficient and effective coarse estimation. Then, we refine the high-scoring hypotheses with our pose refiner, our main contribution. GenFlow iteratively refines the flow and confidence with the guidance of 3D shape. Our method achieves state-of-the-art performance of 6D pose estimation for novel objects concerning RGB and RGB-D inputs.", + "bbox": [ + 496, + 760, + 890, + 887 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10046", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Christopher M. Bishop. Pattern recognition and machine learning, 5th Edition. Springer, 2007. 3", + "[2] Michael J. Black and P. Anandan. A framework for the robust estimation of optical flow. In Fourth International Conference on Computer Vision, ICCV 1993, Berlin, Germany, 11-14 May, 1993, Proceedings, pages 231-236. IEEE Computer Society, 1993. 2", + "[3] Angel X. Chang, Thomas A. Funkhouser, Leonidas J. Guibas, Pat Hanrahan, Qi-Xing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. Shapenet: An information-rich 3d model repository. CoRR, abs/1512.03012, 2015. 5", + "[4] Hansheng Chen, Pichao Wang, Fan Wang, Wei Tian, Lu Xiong, and Hao Li. Epro-pnp: Generalized end-to-end probabilistic perspective-n-points for monocular object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2781-2790, 2022. 5", + "[5] Jianqiu Chen, Mingshan Sun, Tianpeng Bao, Rui Zhao, Liwei Wu, and Zhenyu He. Zeropose: Cad-model-based zero-shot pose estimation, 2023. 1, 2, 7", + "[6] Qifeng Chen and Vladlen Koltun. Full flow: Optical flow estimation by global optimization over regular grids. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 4706-4714. IEEE Computer Society, 2016. 2", + "[7] Xu Chen, Zijian Dong, Jie Song, Andreas Geiger, and Otmar Hilliges. Category level object pose estimation via neural analysis-by-synthesis. 2020. 2", + "[8] Yilun Chen, Zhicheng Wang, Yuxiang Peng, Zhiqiang Zhang, Gang Yu, and Jian Sun. Cascaded pyramid network for multi-person pose estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 5", + "[9] Kyunghyun Cho, Bart van Merrienboer, Dzmitry Bahdanau, and Yoshua Bengio. On the properties of neural machine translation: Encoder-decoder approaches. In Proceedings of SSST@EMNLP 2014, Eighth Workshop on Syntax, Semantics and Structure in Statistical Translation, Doha, Qatar, 25 October 2014, pages 103-111. Association for Computational Linguistics, 2014. 4", + "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2009), 20-25 June 2009, Miami, Florida, USA, pages 248-255. IEEE Computer Society, 2009. 6", + "[11] Maximilian Denninger, Dominik Winkelbauer, Martin Sundermeyer, Wout Boerdijk, Markus Knauer, Klaus H. Strobl, Matthias Humt, and Rudolph Triebel. Blenderproc2: A procedural pipeline for photorealistic rendering. Journal of Open Source Software, 8(82):4901, 2023. 5, 7", + "[12] Alexey Dosovitskiy, Philipp Fischer, Eddy Ilg, Philip Hausser, Caner Hazirbas, Vladimir Golkov, Patrick van der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. In Proceedings" + ], + "bbox": [ + 78, + 114, + 467, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "of the IEEE International Conference on Computer Vision (ICCV), 2015. 2", + "[13] Andreas Doumanoglou, Rigas Kouskouridas, Sotiris Malassiotis, and Tae-Kyun Kim. Recovering 6d object pose and predicting next-best-view in the crowd. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 3583-3592. IEEE Computer Society, 2016. 7", + "[14] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas Barlow McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. In 2022 International Conference on Robotics and Automation, ICRA 2022, Philadelphia, PA, USA, May 23-27, 2022, pages 2553-2560. IEEE, 2022. 5", + "[15] Bertram Drost, Markus Ulrich, Nassir Navab, and Slobodan Ilic. Model globally, match locally: Efficient and robust 3d object recognition. In 2010 IEEE computer society conference on computer vision and pattern recognition, pages 998-1005. IEEE, 2010. 2", + "[16] Bertram Drost, Markus Ulrich, Paul Bergmann, Philipp Hartinger, and Carsten Steger. Introducing mvtec ITODD - A dataset for 3d object recognition in industry. In 2017 IEEE International Conference on Computer Vision Workshops, ICCV Workshops 2017, Venice, Italy, October 22-29, 2017, pages 2200-2208. IEEE Computer Society, 2017. 7", + "[17] Johan Edstedt, Ioannis Athanasiadis, Mårten Wadenbäck, and Michael Felsberg. Dkm: Dense kernelized feature matching for geometry estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17765-17775, 2023. 6", + "[18] Martin A. Fischler and Robert C. Bolles. Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM, 24(6):381-395, 1981. 2, 6", + "[19] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. YOLOX: exceeding YOLO series in 2021. CoRR, abs/2107.08430, 2021. 3, 7", + "[20] Yang Hai, Rui Song, Jiaojiao Li, and Yinlin Hu. Shape-constraint recurrent flow for 6d object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4831-4840, 2023. 1, 2, 5", + "[21] Rasmus Laurvig Haugaard and Anders Glent Buch. Surfemb: Dense and continuous correspondence distributions for object pose estimation with learnt surface embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6749-6758, 2022. 7", + "[22] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross B. Girshick. Mask R-CNN. In IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017, pages 2980-2988. IEEE Computer Society, 2017. 7", + "[23] Xingyi He, Jiaming Sun, Yuang Wang, Di Huang, Hujun Bao, and Xiaowei Zhou. Onepose++: Keypoint-free one-shot object pose estimation without CAD models. In Advances in Neural Information Processing Systems, 2022. 2" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10047", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Yisheng He, Yao Wang, Haoqiang Fan, Jian Sun, and Qifeng Chen. Fs6d: Few-shot 6d pose estimation of novel objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6814-6824, 2022. 2", + "[25] Stefan Hinterstoisser, Vincent Lepetit, Slobodan Ilic, Stefan Holzer, Gary R. Bradski, Kurt Konolige, and Nassir Navab. Model based training, detection and pose estimation of texture-less 3d objects in heavily cluttered scenes. In Computer Vision - ACCV 2012 - 11th Asian Conference on Computer Vision, Daejeon, Korea, November 5-9, 2012, Revised Selected Papers, Part I, pages 548-562. Springer, 2012. 7", + "[26] Tomas Hodan, Pavel Haluza, Stepán Obdrzánek, Jiri Matas, Manolis I. A. Lourakis, and Xenophon Zabulis. T-LESS: an RGB-D dataset for 6d pose estimation of texture-less objects. In 2017 IEEE Winter Conference on Applications of Computer Vision, WACV 2017, Santa Rosa, CA, USA, March 24-31, 2017, pages 880-888. IEEE Computer Society, 2017. 7", + "[27] Tomás Hodań, Frank Michel, Eric Brachmann, Wadim Kehl, Anders Glent Buch, Dirk Kraft, Bertram Drost, Joel Vidal, Stephan Ihrke, Xenophon Zabulis, Caner Sahin, Fabian Manhardt, Federico Tombari, Tae-Kyun Kim, Jiri Matas, and Carsten Rother. BOP: Benchmark for 6D object pose estimation. European Conference on Computer Vision (ECCV), 2018. 7", + "[28] Tomás Hodan, Martin Sundermeyer, Bertram Drost, Yann Labbé, Eric Brachmann, Frank Michel, Carsten Rother, and Jiri Matas. BOP challenge 2020 on 6d object localization. In Computer Vision - ECCV 2020 Workshops - Glasgow, UK, August 23-28, 2020, Proceedings, Part II, pages 577-594. Springer, 2020. 2, 7", + "[29] Berthold K. P. Horn and Brian G. Schunck. Determining optical flow. Artif. Intell., 17(1-3):185-203, 1981. 2", + "[30] Yinlin Hu, Sébastien Speierer, Wenzel Jakob, Pascal Fua, and Mathieu Salzmann. Wide-depth-range 6d object pose estimation in space. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 15870-15879. Computer Vision Foundation / IEEE, 2021. 7", + "[31] Yinlin Hu, Pascal Fua, and Mathieu Salzmann. Perspective flow aggregation for data-limited 6d object pose estimation. In ECCV, 2022. 1, 2, 7", + "[32] Tak-Wai Hui, Xiaou Tang, and Chen Change Loy. Liteflownet: A lightweight convolutional neural network for optical flow estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2", + "[33] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2", + "[34] Wolfgang Kabsch. A solution for the best rotation to relate two sets of vectors. Acta Crystallographica Section A: Crystal Physics, Diffraction, Theoretical and General Crystallography, 32(5):922-923, 1976. 2, 6" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Roman Kaskman, Sergey Zakharov, Ivan Shugurov, and Slobodan Ilic. Homebrewdb: RGB-D dataset for 6d pose estimation of 3d objects. In 2019 IEEE/CVF International Conference on Computer Vision Workshops, ICCV Workshops 2019, Seoul, Korea (South), October 27-28, 2019, pages 2767-2776. IEEE, 2019. 7", + "[36] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 3", + "[37] Yann Labbe, Lucas Manuelli, Arsalan Mousavian, Stephen Tyree, Stan Birchfield, Jonathan Tremblay, Justin Carpentier, Mathieu Aubry, Dieter Fox, and Josef Sivic. MegaPose: 6D Pose Estimation of Novel Objects via Render & Compare. In CoRL. 1, 2, 3, 5, 6, 7", + "[38] Y. Labbe, J. Carpentier, M. Aubry, and J. Sivic. Cosyposse: Consistent multi-view multi-object 6d pose estimation. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2, 5, 7", + "[39] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Epnp: An accurate o(n) solution to the pnp problem. International Journal Of Computer Vision, 81:155-166, 2009. 8", + "[40] Fu Li, Ivan Shugurov, Benjamin Busam, Minglong Li, Shaowu Yang, and Slobodan Ilic. Polarmesh: A star-convex 3d shape approximation for object pose estimation. IEEE Robotics Autom. Lett., 7(2):4416-4423, 2022. 2", + "[41] Yi Li, Gu Wang, Xiangyang Ji, Yu Xiang, and Dieter Fox. Deepim: Deep iterative matching for 6d pose estimation. In European Conference on Computer Vision (ECCV), 2018. 2", + "[42] Tsung-Yi Lin, Piotr Dollar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 5", + "[43] Lahav Lipson, Zachary Teed, Ankit Goyal, and Jia Deng. Coupled iterative refinement for 6d multi-object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6728-6737, 2022. 2, 5", + "[44] Xingyu Liu, Ruida Zhang, Chenyangguang Zhang, Bowen Fu, Jiwen Tang, Xiquan Liang, Jingyi Tang, Xiaotian Cheng, Yukang Zhang, Gu Wang, and Xiangyang Ji. Gdrnpp. https://github.com/shanice-1/gdrnpp_bop2022, 2022.3,7,8", + "[45] Yuan Liu, Yilin Wen, Sida Peng, Cheng Lin, Xiaoxiao Long, Taku Komura, and Wenping Wang. Gen6d: Generalizable model-free 6-dof object pose estimation from rgb images. In ECCV, 2022. 2", + "[46] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11976-11986, 2022. 6", + "[47] David G. Lowe. Object recognition from local scale-invariant features. In Proceedings of the International Conference on Computer Vision, Kerkyra, Corfu, Greece," + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10048", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "September 20-25, 1999, pages 1150-1157. IEEE Computer Society, 1999. 2", + "[48] Fabian Manhardt, Gu Wang, Benjamin Busam, Manuel Nickel, Sven Meier, Luca Minciullo, Xiangyang Ji, and Nassir Navab. Cps++: Improving class-level 6d pose and shape estimation from monocular images with self-supervised learning, 2020. 2", + "[49] Van Nguyen Nguyen, Yinlin Hu, Yang Xiao, Mathieu Salzmann, and Vincent Lepetit. Templates for 3d object pose estimation revisited: Generalization to new objects and robustness to occlusions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6771-6780, 2022. 2", + "[50] Van Nguyen Nguyen, Thibault Groueix, Georgy Ponimatkin, Vincent Lepetit, and Tomas Hodan. Cnos: A strong baseline for cad-based novel object segmentation. arXiv preprint arXiv:2307.11067, 2023. 2, 3, 7", + "[51] Brian Okorn, Qiao Gu, Martial Hebert, and David Held. Zephyr: Zero-shot pose hypothesis rating. In IEEE International Conference on Robotics and Automation, ICRA 2021, Xi'an, China, May 30 - June 5, 2021, pages 14141-14148. IEEE, 2021. 1, 2", + "[52] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 3", + "[53] Keunhong Park, Arsalan Mousavian, Yu Xiang, and Dieter Fox. Latentfusion: End-to-end differentiable reconstruction and rendering for unseen object pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2", + "[54] Ivan Shugurov, Fu Li, Benjamin Busam, and Slobodan Ilic. Osop: A multi-stage one shot object pose estimation framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6835-6844, 2022. 1, 2, 7", + "[55] Andrea Simonelli, Samuel Rota Bulo, Lorenzo Porzi, Manuel Lopez-Antequera, and Peter Kontschieder. Disentangling monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 2", + "[56] Yongzhi Su, Mahdi Saleh, Torben Fetzer, Jason Rambach, Nassir Navab, Benjamin Busam, Didier Stricker, and Federico Tombari. Zebrapose: Coarse to fine surface encoding for 6 dof object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6738-6748, 2022. 7", + "[57] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2", + "[58] Jiaming Sun, Zihao Wang, Siyu Zhang, Xingyi He, Hongcheng Zhao, Guofeng Zhang, and Xiaowei Zhou." + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Onepose: One-shot object pose estimation without cad models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6825-6834, 2022. 2", + "[59] Martin Sundermeyer, Maximilian Durner, En Yen Puang, Zoltan-Csaba Marton, Narunas Vaskevicius, Kai O. Arras, and Rudolph Triebel. Multi-path learning for object pose estimation across domains. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2", + "[60] Mingxing Tan and Quoc V. Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In Proceedings of the 36th International Conference on Machine Learning, ICML 2019, 9-15 June 2019, Long Beach, California, USA, pages 6105-6114. PMLR, 2019. 2", + "[61] Zachary Teed and Jia Deng. RAFT: recurrent all-pairs field transforms for optical flow. In Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II, pages 402-419. Springer, 2020. 2, 4, 6", + "[62] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J. Guibas. Normalized object coordinate space for category-level 6d object pose and size estimation. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2", + "[63] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes. 2018. 7", + "[64] Yang Xiao, Xuchong Qiu, Pierre-Alain Langlois, Mathieu Aubry, and Renaud Marlet. Pose from shape: Deep pose estimation for arbitrary 3d objects. In 30th British Machine Vision Conference 2019, BMVC 2019, Cardiff, UK, September 9-12, 2019, page 61. BMVA Press, 2019. 2", + "[65] Yang Xiao, Yuming Du, and Renaud Marlet. Posecontrast: Class-agnostic object viewpoint estimation in the wild with pose-aware contrastive learning. In International Conference on 3D Vision (3DV), 2021. 2", + "[66] Gengshan Yang and Deva Ramanan. Volumetric correspondence networks for optical flow. In Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada, pages 793-803, 2019. 2", + "[67] Heng Yang, Jingnan Shi, and Luca Carlone. TEASER: fast and certifiable point cloud registration. IEEE Trans. Robotics, 37(2):314-333, 2021. 7", + "[68] Anna Yershova, Swati Jain, Steven M. LaValle, and Julie C. Mitchell. Generating uniform incremental grids on $SO(3)$ using the hopf fibration. Int. J. Robotics Res., 29(7):801-812, 2010. 8", + "[69] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + ], + "bbox": [ + 503, + 92, + 890, + 839 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "10049", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/GenFlow_ Generalizable Recurrent Flow for 6D Pose Refinement of Novel Objects/ff0e87f9-2cf2-45f0-8b79-ad267112b5a6_model.json b/2024/GenFlow_ Generalizable Recurrent Flow for 6D Pose Refinement of Novel Objects/ff0e87f9-2cf2-45f0-8b79-ad267112b5a6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d7b6516c9661ed737449189630b2ba357732c6ce --- /dev/null +++ b/2024/GenFlow_ Generalizable Recurrent Flow for 6D Pose Refinement of Novel Objects/ff0e87f9-2cf2-45f0-8b79-ad267112b5a6_model.json @@ -0,0 +1,2345 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.13, + 0.856, + 0.176 + ], + "angle": 0, + "content": "GenFlow: Generalizable Recurrent Flow for 6D Pose Refinement of Novel Objects" + }, + { + "type": "text", + "bbox": [ + 0.166, + 0.204, + 0.309, + 0.222 + ], + "angle": 0, + "content": "Sungphill Moon*" + }, + { + "type": "text", + "bbox": [ + 0.348, + 0.204, + 0.472, + 0.221 + ], + "angle": 0, + "content": "Hyeontae Son*" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.205, + 0.638, + 0.221 + ], + "angle": 0, + "content": "Dongcheol Hur" + }, + { + "type": "text", + "bbox": [ + 0.677, + 0.205, + 0.803, + 0.222 + ], + "angle": 0, + "content": "Sangwook Kim" + }, + { + "type": "text", + "bbox": [ + 0.423, + 0.223, + 0.546, + 0.238 + ], + "angle": 0, + "content": "NAVER LABS" + }, + { + "type": "text", + "bbox": [ + 0.191, + 0.242, + 0.774, + 0.257 + ], + "angle": 0, + "content": "{sungphill.moon, son.ht, dongcheol.hur, o.s.w.a.l.k}@naverlabs.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.313, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.324, + 0.474, + 0.642 + ], + "angle": 0, + "content": "Despite the progress of learning-based methods for 6D object pose estimation, the trade-off between accuracy and scalability for novel objects still exists. Specifically, previous methods for novel objects do not make good use of the target object's 3D shape information since they focus on generalization by processing the shape indirectly, making them less effective. We present GenFlow, an approach that enables both accuracy and generalization to novel objects with the guidance of the target object's shape. Our method predicts optical flow between the rendered image and the observed image and refines the 6D pose iteratively. It boosts the performance by a constraint of the 3D shape and the generalizable geometric knowledge learned from an end-to-end differentiable system. We further improve our model by designing a cascade network architecture to exploit the multi-scale correlations and coarse-to-fine refinement. GenFlow ranked first on the unseen object pose estimation benchmarks in both the RGB and RGB-D cases. It also achieves performance competitive with existing state-of-the-art methods for the seen object pose estimation without any fine-tuning." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.671, + 0.208, + 0.685 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.696, + 0.47, + 0.846 + ], + "angle": 0, + "content": "Estimating accurate 6D object pose from an RGB or RGB-D image is crucial for robotic tasks and augmented reality applications. In the standard setup of 6D object pose estimation, 3D models of target objects are given and available in training and testing. Most state-of-the-art algorithms are learning-based and require training the model per object or dataset containing few objects. However, these methods are difficult to handle novel objects since they memorize the 3D shapes of the targets and suffer from the training cost to support them." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.847, + 0.469, + 0.877 + ], + "angle": 0, + "content": "Recent works [5, 37, 51, 54] have proposed generalizable models that can estimate the 6D pose of novel objects, i.e.," + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.296, + 0.875, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.436, + 0.895, + 0.478 + ], + "angle": 0, + "content": "Figure 1. Examples of 6D pose estimation of novel objects. Our method estimates correspondences between the input image and rendered image, and 6D object pose." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.506, + 0.892, + 0.688 + ], + "angle": 0, + "content": "objects unseen during training. Specifically, OSOP [54] and MegaPose [37] show remarkable performance with the generalizable pose refinement. Their refinement methods have the scalability to novel objects since the networks learn to compare the rendered image and the observed image without memorizing 3D shapes. However, their refiners avail themselves of 3D shape information indirectly. OSOP refiner is trained by the surrogate 2D-2D matching loss, which makes it suboptimal for 6D pose estimation. MegaPose refiner regresses the relative 6D pose between the rendered image and the input, but the regression makes no use of the projective geometry." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.893, + 0.884 + ], + "angle": 0, + "content": "Our research question is Is there a better design for generalizable pose refinement that takes advantage of shape information?. Recently, 2D optical flow-based methods [20, 31] show noticeable performance on 6D pose estimation. These methods estimate dense 2D-2D correspondences between the rendered image and the observed image. The estimated correspondences are lifted to 2D-3D correspondences on the 3D shape of the target object, then the 6D pose is recovered using a PnP solver [31] or pose regressor networks [20]. The optical flow-based pose refinement provides a clue for the generalization due to its robustness to the domain shift [31] and shows the potential to utilize 3D information in the model [20]." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We propose GenFlow, the optical flow-based iterative" + }, + { + "type": "page_footnote", + "bbox": [ + 0.101, + 0.888, + 0.292, + 0.901 + ], + "angle": 0, + "content": "*These authors contributed equally." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10039" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.212 + ], + "angle": 0, + "content": "refinement for the generalizable object pose estimation guided by the 3D shape information of target objects. Given an image and initial 6D pose of an object in the image, GenFlow estimates 2D dense correspondences between the rendered image and the input image and refined 6D pose through a PnP layer. This series of processes is performed iteratively, so it can improve accuracy and robustness to outliers [43]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.213, + 0.471, + 0.379 + ], + "angle": 0, + "content": "We adopt the shape-constraint recurrent flow framework inspired by [20] for our model to use 3D information. This framework utilizes the shape information implicitly to enjoy both generalization and performance. Its end-to-end differentiable system learns to estimate the optical flow, confidence, and object pose. These estimations are updated iteratively complying with the 3D shape. We further introduce a cascade network design to make use of the multi-scale correlation volumes. It takes advantage of the coarse-to-fine refinement and helps the model estimate a more accurate pose." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.379, + 0.47, + 0.545 + ], + "angle": 0, + "content": "We validate our method on various datasets provided in the BOP challenge [28]. Our model is designed to estimate 6D pose for the RGB input but is also applicable to the RGB-D case with the RANSAC-Kabsch algorithm [18, 34]. For both RGB and RGB-D cases, our method achieves the overall best performance for the unseen object pose estimation task using the default detection [50] provided in the BOP challenge 2023. It is also notable that our method achieves competitive performance with the state-of-the-art methods for seen objects when used with better detection results." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.559, + 0.22, + 0.574 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.47, + 0.854 + ], + "angle": 0, + "content": "Iterative Refinement DeepIM [41] pioneered the iterative render-and-compare strategy for 6D pose estimation. It predicts the relative pose that better matches a rendered image against the observed image. By iterative refinement, the rendered and observed images become more similar. CosyPose [38] improved the idea by introducing better parameterization for rotation [69], network architecture [60], loss design [55, 62] and data augmentation. These iterative methods show accurate results and the possibility of being applied to the unseen objects [41]. In contrast to the previous works, CIR [43] introduces a novel differentiable PnP layer to utilize the projective geometry with deep learning. Unlike CIR, which assumes the RGB-D input, ours focuses on the RGB input. For the RGB input case, CIR obtains the rendered depth as the substitution of input depth for every flow update, so the inference suffers from time inefficiency. On the other hand, our refinement method requires fewer renderings." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Optical Flow Estimation The classical approaches [2, 6, 29] dealt with optical flow estimation as a hand-crafted optimization problem in pursuit of the balance of bright-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.333 + ], + "angle": 0, + "content": "ness consistency and motion plausibility. Deep learning methods [12, 32, 33, 57, 66] emerged as promising alternatives of the classical methods, achieving considerable performance with faster inference times. Recently, RAFT [61] introduced a recurrent deep network architecture which has shown significant improvement over the previous methods regarding accuracy, efficiency, and generalization. PFA [31] applied the optical flow estimation to 6D pose refinement for data-limited cases. Since learning to estimate dense optical flow helps the networks to focus on the lower-level information, it could be robust to domain shift. Following PFA, SCFlow [20] further improved the performance by introducing an end-to-end trainable system using a pose regressor and shape-constraint lookup. Our method capitalizes on a differentiable PnP layer to the framework in order to make use of projective geometry." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.334, + 0.892, + 0.771 + ], + "angle": 0, + "content": "Unseen Object Pose Estimation Category-level 6D pose estimation [7, 40, 48, 62] tackled the case of unseen objects of which category is given in advance. Since these methods depend on the prior of the 3D shapes belonging to specific category, they struggle with handling the objects of unknown categories. Recently, several works proposed class-agnostic generalizable pose estimators that can handle unseen objects. A few works attempted the CAD model-free pose estimation [23, 24, 45, 53, 58] to cover everyday objects, but the results were only evaluated on datasets without occlusion and cluttered scenes. [49, 59, 64, 65] focused on estimating the orientation of novel objects by comparing the deep image features of the rendered image and the observed image. ZePHyR [51] proposed a generalizable scoring function combined with hand-crafted features [15, 47]. However, it uses the features selectively according to the richness of texture in the datasets. ZeroPose [5] estimates 6D pose by hierarchical feature matching which requires the depth input by its nature and pose refinement for better performance. Like ours, OSOP [54] and MegaPose [37] proposed CAD-based pose estimation methods that can process RGB and RGB-D inputs. Specifically, MegaPose achieved state-of-the-art performance by its iterative pose refiner. The pose refiner leverages the multi-view rendered images containing surface normals for utilizing rich information about the novel object's shape, anchor point, and coordinate system. Our method refines the pose with the guidance of the target object's shape without rendering multiview images per refinement step." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.785, + 0.591, + 0.8 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This section describes our methodology for 6D pose estimation of novel objects. Given an RGB image and 3D models of target objects, which are unseen during training, our goal is to estimate the 6D poses of the objects with respect to the camera. Our pose estimation pipeline consists of 3 stages: object detection, coarse pose estimation, and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10040" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.184, + 0.095, + 0.368, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.207, + 0.167, + 0.341, + 0.178 + ], + "angle": 0, + "content": "(a) Naïve Sampling \\((N = 576)\\)" + }, + { + "type": "image", + "bbox": [ + 0.403, + 0.149, + 0.46, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.189, + 0.271, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.19, + 0.261, + 0.356, + 0.273 + ], + "angle": 0, + "content": "(b) Sampling using GMM \\((M = 104)\\)" + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.189, + 0.461, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.287, + 0.471, + 0.371 + ], + "angle": 0, + "content": "Figure 2. Visualization of the scores predicted by the coarse model. Each dot denotes the sampled rotation, and its size represents the relative score of the coarse model for the corresponding rotation. The green circles are boundaries of the area where the refiner is interested. Figures show that the high-scoring poses are clustered in the vicinity of the green circles." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.399, + 0.47, + 0.459 + ], + "angle": 0, + "content": "pose refinement. Note that the detection of novel objects is out of the scope of this paper. We first provide an overview of the whole process and then focus on the pose refinement method, which is the main contribution of this work." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.469, + 0.218, + 0.484 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.492, + 0.469, + 0.598 + ], + "angle": 0, + "content": "Given a calibrated RGB image \\(\\mathcal{I}\\) and a set of 3D models of the target objects \\(\\{\\mathcal{M}_i\\}\\), we estimate the 6D poses of the visible objects in the image. A 6D pose is defined by a matrix \\(\\mathbf{P}\\) representing the rigid transformation from the object space to the camera space. Note that we assume the input image is calibrated, i.e., the intrinsic camera matrix \\(\\mathbf{K}\\) is known." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.469, + 0.794 + ], + "angle": 0, + "content": "2D Object Detection Given an RGB image, the object detection method produces a list of object detections consisting of 2D bounding boxes and associated labels. For the unseen object detection, we apply the CNOS [50] detection leveraging the foundation models [36, 52]. To show the extensive results of our work, we also apply the YOLOX detector [19] from GDRNPP [44] which is used for seen object detection. For each detection, the image \\(\\mathcal{I}\\) is cropped and further resized to the specific resolution in the following steps. The intrinsic parameters are adjusted according to the resized resolution. For convenience, the following explanation focuses on the 6D pose estimation of a single detected object of which the 3D model is \\(\\mathcal{M}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Coarse Pose Estimation Given object detection, the coarse pose estimator predicts rough initial poses. We used the classification-based method proposed in [37] for better generalization. Given an input image and an arbitrary pose \\(\\mathbf{P}\\) for the 3D model \\(\\mathcal{M}\\), the coarse model classifies whether or not the pose \\(\\mathbf{P}\\) could be improved to the pose in proximity to the ground-truth by the refiner. It uses a pre-defined" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.333 + ], + "angle": 0, + "content": "set of 3D rotations of which the cardinality is \\( N \\) to generate the various arbitrary pose hypotheses. For each rotation, a rough 3D translation is estimated for the points of the object's 3D model to be approximately inside the 2D bounding box. We extract scores of the pose hypotheses from the model and select the top-\\( n \\) scoring poses. Such a classification-based estimation can handle novel objects, but it suffers from the cost of rendering a large number of images for the pre-defined rotations. Inspired by the property of high-scoring rotations to be sparse and clustered as shown in Fig 2, we introduce sampling using a Gaussian mixture model (GMM) [1] for efficiency. First, we choose the top \\( k \\)-scoring samples \\( \\{R_i = (\\psi_i, \\theta_i, \\phi_i)\\}_{i=1}^k \\) from \\( M \\) pre-defined rotations (Fig 2b left). Then we create a density distribution \\( p(X) \\) for the Euler angle \\( X \\) using a GMM which satisfies that" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.343, + 0.892, + 0.383 + ], + "angle": 0, + "content": "\\[\np (X) = \\sum_ {i = 1} ^ {k} \\frac {1}{k} \\mathcal {N} (X | \\mu_ {i}, \\Sigma), \\mu_ {i} = R _ {i} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.392, + 0.892, + 0.467 + ], + "angle": 0, + "content": "Finally, we sample additional \\(M\\) rotations from the density distribution \\(p(X)\\) (Fig 2b right) and choose the best-scoring \\(n\\) hypotheses among the \\(2M\\) samples. This method renders \\(2M < N\\) images while achieving better performance. We used \\(M = 104\\) and \\(k = 16\\) for our experiments." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.468, + 0.892, + 0.634 + ], + "angle": 0, + "content": "Pose Refinement Given a coarse pose estimate, we refine the pose for the rendered image to match against the input image crop. Starting from the coarse estimate \\(\\mathbf{P}_0\\) as an initial pose, we iteratively refine the pose using our refiner. We denote the \\(k^{th}\\) 6D pose estimate of the refiner as \\(\\mathbf{P}_k\\). For the \\(k^{th}\\) iteration, we obtain an image \\(\\mathcal{I}_{r,k}\\) by rendering the 3D model \\(\\mathcal{M}\\) for the pose \\(\\mathbf{P}_{k-1}\\). Our refiner then estimates the 2D optical flow \\(\\mathbf{F}_k\\) between the image \\(\\mathcal{I}_{r,k}\\) and the input image crop and the confidence weights \\(\\mathbf{W}_k\\) which are used to solve the pose \\(\\mathbf{P}_k\\). A detailed description is in the following section." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.643, + 0.698, + 0.658 + ], + "angle": 0, + "content": "3.2. GenFlow Refinement" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.666, + 0.621, + 0.68 + ], + "angle": 0, + "content": "3.2.1 Overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.854 + ], + "angle": 0, + "content": "We describe the specific operation of GenFlow refiner to estimate the refined pose \\(\\mathbf{P}_k\\) from the previous result \\(\\mathbf{P}_{k - 1}\\) but hereafter, we omit the \\(k\\) notation if possible for convenience. First, we obtain a synthetic image by rendering the 3D model \\(\\mathcal{M}\\) for the pose \\(\\mathbf{P}\\) and adjusted intrinsic parameters. Specifically, the results of the rendering consists of the RGB image, depth image and binary rendered mask, denoted by \\(\\mathcal{I}_r\\in R^{H\\times W\\times 3}\\), \\(\\mathcal{D}_r\\in R^{H\\times W}\\), and \\(\\mathcal{M}_r\\in \\{0,1\\}^{H\\times W}\\) respectively. Then, we extract features from both images and refine the pose with the GenFlow module." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Feature Extraction We extract the image features from the input image crop and the rendered image \\(\\mathcal{I}_r\\) with the same feature encoder and the context features from" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10041" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.087, + 0.891, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.453, + 0.893, + 0.484 + ], + "angle": 0, + "content": "Figure 3. Overview of GenFlow refinement. We visualize the process of feature extraction and GenFlow update in the \\(k^{th}\\) refinement. \\(T_{C2O}\\) refers to a transformation that maps the coordinates from camera space to object space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.509, + 0.47, + 0.554 + ], + "angle": 0, + "content": "the \\(\\mathcal{I}_r\\). The feature maps from the \\(i^{th}\\) layer of the feature encoder are used to construct the correlation volumes \\(\\mathbf{C}_i \\in R^{H_i \\times W_i \\times H_i \\times W_i}\\) by the dot product." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.555, + 0.471, + 0.75 + ], + "angle": 0, + "content": "GenFlow Module For each correlation volume and context feature map, we construct a module named \"GenFlow module\". A GenFlow module consists of a convolution-based Gated Recurrent Unit (GRU) [9], a differentiable PnP solver, and correlation lookup based on the pose-induced flow as shown in the Fig 3. GenFlow module iteratively updates the optical flow, confidence weights, and 6D pose. For the \\( j^{th} \\) update, the output of the module is the refined 6D pose \\( \\mathbf{P}^j \\). Note that we use subscripts for the outer updates and superscripts for the inner updates, i.e., GenFlow updates. The detailed illustration of updating the 6D pose with a GenFlow module for \\( i^{th} \\) correlation volume is as follows." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.77, + 0.25, + 0.785 + ], + "angle": 0, + "content": "3.2.2 GenFlow Update" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.793, + 0.47, + 0.903 + ], + "angle": 0, + "content": "GRU Update For the \\(j^{th}\\) update, the GRU of the GenFlow module consumes the previous hidden state \\(h^{j-1}\\), correlation features \\(r^{j-1}\\) and context features. The output of the GRU \\(\\{\\mathbf{h}^j, \\mathbf{f}^j, \\mathbf{u}^j, \\mathbf{c}^j, \\mathbf{s}^j\\}\\) includes the updated hidden state \\(h^j\\), optical flow \\(\\mathbf{f}^j\\), flow upsampling mask \\(\\mathbf{u}^j\\), certainty \\(\\mathbf{c}^j\\) and pose sensitivity \\(\\mathbf{s}^j\\). The flow \\(\\mathbf{f}^j \\in R^{H_i \\times W_i \\times 2}\\) is upsampled to \\(\\mathbf{F}^j\\) to match the resolution of previous \\((i-1)^{th}\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.509, + 0.893, + 0.675 + ], + "angle": 0, + "content": "feature map using the mask \\(\\mathbf{u}^j\\). We apply the convex combination method of 3x3 neighbors for the optical flow proposed in [61]. The certainty \\(\\mathbf{c}^j\\) and pose sensitivity \\(\\mathbf{s}^j\\) are upsampled by bilinear interpolation, and then we multiply them to obtain the confidence weights \\(\\mathbf{W}^j\\). The head networks for certainty and pose sensitivity are supervised by different losses, respectively. This explicit factorization of the confidence weights helps to be robust to occlusion with the help of certainty estimation as shown in Fig 4. The upsampled flow \\(\\mathbf{F}^j\\) and confidence weights \\(\\mathbf{W}^j\\) are used to establish weighted 2D-3D correspondences." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.737 + ], + "angle": 0, + "content": "Pose Update As we know the depth of synthetic image \\(\\mathcal{D}_r\\) for the pose \\(\\mathbf{P}_k^0 = [\\mathbf{R}_k^0 |\\mathbf{t}_k^0 ]\\) and the intrinsic camera matrix \\(\\mathbf{K}\\), we can compute the object-space 3D coordinate of each pixel by" + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.741, + 0.892, + 0.79 + ], + "angle": 0, + "content": "\\[\n\\left( \\begin{array}{l} x \\\\ y \\\\ z \\end{array} \\right) = \\mathbf {R} _ {k} ^ {\\mathrm {T}} \\left(\\mathbf {K} ^ {- 1} \\mathcal {D} _ {r} (u, v) \\left( \\begin{array}{l} u \\\\ v \\\\ 1 \\end{array} \\right) - \\mathbf {t} _ {k}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.87 + ], + "angle": 0, + "content": "where \\((u,v)\\) is the 2D coordinate of a pixel and the \\(\\mathcal{D}_r(u,v)\\) is depth of the pixel. Through this lifting process, we obtain the 2D-3D correspondences for the flow \\(\\mathbf{F}^j\\). We use \\(\\mathbf{W}^j\\) for the confidence of correspondences and binary rendered mask \\(\\mathcal{M}_r\\) to consider the object region only." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To obtain the 6D pose \\(\\mathbf{P}^j\\), we utilize the iterative PnP solver based on the Levenberg-Marquardt (LM) algorithm." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10042" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.468, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.292, + 0.471, + 0.361 + ], + "angle": 0, + "content": "Figure 4. Visualization of the outputs of GRU. Our method factorizes the confidence weights into two terms: certainty and pose sensitivity. Certainty estimation helps to the robustness to occlusion. Pose sensitivity highlights the rich texture regions and extremities of the object." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.388, + 0.47, + 0.419 + ], + "angle": 0, + "content": "It optimizes the 6D pose by minimizing the sum of squared weighted reprojection error:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.429, + 0.47, + 0.513 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\underset {R, t} {\\operatorname {a r g m i n}} \\frac {1}{2} \\sum_ {u} \\sum_ {v} \\| \\mathbf {W} ^ {j} (u, v) \\times (\\pi (R \\left( \\begin{array}{l} x \\\\ y \\\\ z \\end{array} \\right) + t) \\tag {3} \\\\ - (\\left( \\begin{array}{c} u \\\\ v \\end{array} \\right) + \\mathbf {F} ^ {j} (u, v))) \\| ^ {2}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.524, + 0.47, + 0.719 + ], + "angle": 0, + "content": "where \\(\\pi\\) is the projection function with the intrinsic camera parameters and \\((x,y,z)^{\\mathrm{T}}\\) is the 3D coordinate of the pixel \\((u,v)\\) computed by the equation 2. Due to its differentiable nature, the PnP layer enables end-to-end training. However, the backpropagation by unrolling the iterative optimization causes numerical instability, making the training process hard. For stable end-to-end training, we introduce the regularization method [4], which backpropagates the gradient through the last step of a differentiable optimization. Specifically, we first obtain a detached pose \\(\\mathbf{P}^*\\) through 3 LM steps and update the pose via another iteration of the Gauss-Newton algorithm to obtain the pose \\(\\mathbf{P}^j\\). Note that the gradient is not backpropagated through \\(\\mathbf{P}^*\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Correlation Lookup We compute the pose-induced flow, which is the 2D optical flow between the \\(\\mathcal{I}_r\\) and the image with respect to the pose \\(\\mathbf{P}^j\\). The pose-induced flow is calculated as the displacement of each pixel in the \\(\\mathcal{I}_r\\) by lifting and reprojecting the pixel using the depth \\(\\mathcal{D}_r\\). To embed the 3D shape of the target object in the lookup operation [20, 43], we use the pose-induced flow rather than the estimated flow \\(\\mathbf{f}^j\\) for indexing the correlation volume \\(\\mathbf{C}\\). This way of imposing a shape constraint on the model induces the networks to predict dense matching to comply with the target's 3D shape information. From the lookup operation, we obtain the correlation feature \\(r^j\\), which is the input of" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.092, + 0.891, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.251, + 0.894, + 0.335 + ], + "angle": 0, + "content": "Figure 5. Cascade design of GenFlow modules. Multiple GenFlow modules are attached to each level on the feature pyramid. The last updated flow \\(\\mathbf{F}_k^M\\) and 6D pose \\(\\mathbf{P}_k^M\\) from the higher-level GenFlow module are used to initialize the flow and 6D pose of the lower-level module. With the cascade architecture, the 6D pose is recovered in a coarse to fine manner." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.364, + 0.733, + 0.379 + ], + "angle": 0, + "content": "GRU for the next GenFlow update." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.401, + 0.706, + 0.415 + ], + "angle": 0, + "content": "3.2.3 Cascade Architecture" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.427, + 0.892, + 0.593 + ], + "angle": 0, + "content": "Inspired by the previous methods that exploit the multiscale features [8, 42], we design a cascade architecture using multiple GenFlow modules. Given correlation volumes \\(\\{\\mathbf{C}_i\\}\\), we assign a GenFlow module to each volume. Then, we use the 2D optical flow and refined 6D pose of the last inner updates to initialize those of the next module as shown in Fig 5. This cascade architecture takes advantage of progressive refinement, which makes use of the high-level semantics to the low-level fine details. We found that it improves the accuracy of 6D pose rather than using a single GenFlow module." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.605, + 0.605, + 0.62 + ], + "angle": 0, + "content": "3.3. Training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.748 + ], + "angle": 0, + "content": "Dataset Both the coarse and refiner models are trained using the large dataset provided by MegaPose [37]. It is a large synthetic dataset comprising 2 million RGB-D images generated by BlenderProc [11], which contains diverse 3D models with ground-truth 6D pose annotations. The 3D shapes are collected from ShapeNet [3] and Google Scanned Object (GSO) [14] datasets, and we trained our model on both of the datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Refiner Model Our refiner model is trained similarly to the previous iterative refinement methods [37, 38]. Given an image with the ground-truth 6D pose \\(\\mathbf{P}_{gt}\\) for the 3D model \\(\\mathcal{M}\\), we perturb the pose to generate an initial pose \\(\\mathbf{P}_{init}\\) by applying random noises to the translation and rotation of the pose respectively. Translation noise is sampled from a normal distribution with a standard deviation of (0.01, 0.01, 0.05) centimeters, and rotation noise is sampled from another normal distribution of Euler angles with a standard deviation of 15 degrees in each axis. The refiner model" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10043" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "learns to predict the pose \\(\\mathbf{P}_{gt}\\) from the perturbed initial pose \\(\\mathbf{P}_{init}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.123, + 0.47, + 0.26 + ], + "angle": 0, + "content": "Coarse Model Given an input RGB image \\(\\mathcal{I}\\) and a pose \\(\\mathbf{P}\\), a coarse model is trained to classify whether the refiner could estimate the 6D pose in proximity to the ground truth pose from the initial pose \\(\\mathbf{P}\\). We randomly generate the positive and negative pose samples and train the network to classify them using the binary cross entropy loss. For the positive samples, we use the same method used to generate the perturbed initial pose in the refiner model. Negative samples are generated using the method proposed in [37]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.274, + 0.295, + 0.289 + ], + "angle": 0, + "content": "3.4. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.299, + 0.469, + 0.418 + ], + "angle": 0, + "content": "Coarse Estimation We use \\(224 \\times 224\\) resolution for the resized input crop and the synthetic image \\(\\mathcal{I}_r\\). The coarse model is implemented using the ConvNeXt backbone [46] and fully connected head. We used the ConvNext model pre-trained on ImageNet [10] classification to initialize the networks' weights. For inference, we select top-\\(n\\) hypotheses using our GMM-based sampling strategy with \\(M = 104\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.421, + 0.469, + 0.646 + ], + "angle": 0, + "content": "GenFlow Refinement First, we obtain a synthetic image \\(\\mathcal{I}_r\\) of which resolution is \\(256\\times 256\\) and an image crop resized to the same resolution as the synthetic image with adjusted intrinsic parameters. Like the coarse model, we use the pre-trained ConvNeXt for the feature and context encoders. The feature maps from the first two layers of the feature encoder are used to construct correlation volumes \\(\\mathbf{C}_1\\in R^{\\frac{H}{4}\\times \\frac{W}{4}\\times \\frac{H}{4}\\times \\frac{W}{4}}\\) and \\(\\mathbf{C}_2\\in R^{\\frac{H}{8}\\times \\frac{W}{8}\\times \\frac{H}{8}\\times \\frac{W}{8}}\\). Our GRU is implemented with a ConvGRU unit [61] and following four CNN-based heads for the 2D optical flow \\(\\mathbf{f}\\), flow upsampling mask \\(\\mathbf{u}\\), certainty \\(\\mathbf{c}\\) and pose sensitivity \\(\\mathbf{s}\\). Two GenFlow modules are attached to the correlation volumes for the cascade architecture. They are serially connected so that the estimations from higher-level features are utilized as initial values for the next module." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.649, + 0.47, + 0.772 + ], + "angle": 0, + "content": "The optical flow and its upsampling mask produce the upsampled flow, which is supervised by the L1 endpoint error loss with the ground-truth optical flow. We also utilize the disentangled point matching loss for the pose loss, supervising the rotation, 2D object center, and depth of the object individually. Specifically, given an estimated pose \\(\\mathbf{P} = [\\mathbf{R}][\\mathbf{t}_x,\\mathbf{t}_y,\\mathbf{t}_z]^{\\mathrm{T}}]\\) and ground truth pose \\(\\mathbf{P}_{gt} = [\\mathbf{R}^{\\star}][\\mathbf{t}_x^{\\star},\\mathbf{t}_y^{\\star},\\mathbf{t}_z^{\\star}]^{\\mathrm{T}}]\\), the pose loss is computed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.797, + 0.469, + 0.858 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {p o s e} = \\mathcal {D} ([ \\mathbf {R} | [ \\mathbf {t} _ {x} ^ {\\star}, \\mathbf {t} _ {y} ^ {\\star}, \\mathbf {t} _ {z} ^ {\\star} ] ^ {\\mathrm {T}} ], \\mathbf {P} _ {g t}) \\\\ + \\mathcal {D} \\left( \\right.\\left[ \\right. \\mathbf {R} ^ {\\star} \\left. \\right|\\left[ \\mathbf {t} _ {x}, \\mathbf {t} _ {y}, \\mathbf {t} _ {z} ^ {\\star} \\right] ^ {\\mathrm {T}}\\left. \\right), \\mathbf {P} _ {g t}\\left. \\right) \\tag {4} \\\\ + \\mathcal {D} ([ \\mathbf {R} ^ {\\star} | [ \\mathbf {t} _ {x} ^ {\\star}, \\mathbf {t} _ {y} ^ {\\star}, \\mathbf {t} _ {z} ] ^ {\\mathrm {T}} ], \\mathbf {P} _ {g t}) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "where the distance \\(\\mathcal{D}\\) between two 6D poses \\(\\mathbf{P}_1\\) and \\(\\mathbf{P}_2\\) using the 3D points \\(\\mathcal{X}_{\\mathcal{M}}\\) on the 3D model \\(\\mathcal{M}\\) is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.559, + 0.123, + 0.892, + 0.161 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} \\left(\\mathbf {P} _ {1}, \\mathbf {P} _ {2}\\right) = \\frac {1}{\\left| \\mathcal {X} _ {\\mathcal {M}} \\right|} \\sum_ {x \\in \\mathcal {X} _ {\\mathcal {M}}} \\left| \\mathbf {P} _ {1} x - \\mathbf {P} _ {2} x \\right| \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.184, + 0.892, + 0.395 + ], + "angle": 0, + "content": "To supervise the certainty \\(\\mathbf{c}\\), we introduce the certainty estimation by classifying the depth consistency similarly to [17]. Given 6D poses \\(\\mathbf{P}_{init}\\) and \\(\\mathbf{P}_{gt}\\) for the 3D model \\(\\mathcal{M}\\) and intrinsic camera parameters \\(\\mathbf{K}\\), we compute the projected depth \\(\\mathcal{D}_{r\\rightarrow t}\\) by warping the rendered image to the target input image using the depth \\(\\mathcal{D}_r\\). We also obtain the depth map of warped pixels \\(\\mathcal{D}_t\\) given the synthetic depth of the target image. Then we train the networks by minimizing the binary cross entropy loss \\(\\mathcal{L}_{cert}\\) between \\(\\mathbf{c}\\) and \\(\\mathbf{c}_{r\\rightarrow t}\\) which is a binary mask \\(\\mathbf{c}_{r\\rightarrow t} = |\\mathcal{D}_{r\\rightarrow t} - \\mathcal{D}_t| < d_{th}\\) where \\(d_{th}\\) is the distance threshold. It is worth noting that we place a stop-gradient on the certainty mask \\(\\mathbf{c}^j\\) when computing the pose loss \\(\\mathcal{L}_{pose}^j\\) for disentangling the certainty \\(\\mathbf{c}\\) and pose sensitivity \\(\\mathbf{s}\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.399, + 0.892, + 0.429 + ], + "angle": 0, + "content": "Given the RGB-D images with annotated ground truth 6D poses, the overall loss is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.551, + 0.449, + 0.892, + 0.492 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\sum_ {j = 1} ^ {N} \\gamma^ {j - N} \\left(\\mathcal {L} _ {\\text {f l o w}} ^ {j} + \\alpha \\mathcal {L} _ {\\text {c e r t}} ^ {j} + \\beta \\mathcal {L} _ {\\text {p o s e}} ^ {j}\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.511, + 0.892, + 0.594 + ], + "angle": 0, + "content": "following the strategy of exponentially increasing weights [61] where we use the weight \\(\\gamma = 0.8\\) and the iteration of GenFlow updates \\(N = 8\\) in this work. In specific, the GenFlow module assigned to \\(\\mathbf{C}_2\\) outputs \\(\\{\\mathbf{P}^j\\}_{j=1}^{N/2}\\) and the another module outputs \\(\\{\\mathbf{P}^j\\}_{j=N/2}^N\\) from \\(\\mathbf{C}_1\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.595, + 0.892, + 0.716 + ], + "angle": 0, + "content": "Multi-Hypotheses Strategy We select \\( n \\) best-scoring hypotheses in the coarse estimation. The selected hypotheses are refined, and we choose the best estimate by evaluating the refined hypotheses using the coarse model again. This strategy requires more inference time for evaluating \\( n \\) iterations of the rendering and refinement process, but it can boost the accuracy by avoiding the risk of the \"Winner-Takes-All\" strategy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "RGB-D Input If the observed image consists of depth information, we apply the depth refinement after the GenFlow loop for every refinement step. First, we obtain the optical flow \\(\\mathbf{F}\\) and upsampled certainty from the results of the last GenFlow update. 3D-3D correspondences are established by the optical flow and the input depth. We filter out the correspondences of certainty lower than a specific threshold to remove the outliers. Then we apply the RANSAC-Kabsch [18, 34] for the depth refinement. Using certainty for filtering out the outliers is more effective than using confidence weights since the high-confidence pixels are sparse and the input depth is prone to be noisy." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10044" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.078, + 0.089, + 0.921, + 0.32 + ], + "angle": 0, + "content": "
2D LocalizationPose InitializationPose RefinementAverage Recall (↑)
RGB-D InputMethodNovel ObjectsMethodNovel ObjectsMethodNovel ObjectsLM-O T-LESS TUD-L IC-BIN ITODD HB YCB-VMEAN
1XMask-RCNN [22, 38]XCosyPose [38]XCosyPose [38]X63.364.068.558.321.665.657.457.0
2XMask-RCNN [22, 38]XSurfEmb [21]XBFGSX66.373.571.558.841.379.164.765.0
3XYOLOX [19, 44]XZebraPose [56]X-X72.981.175.659.250.492.172.972.0
4XYOLOX [19, 44]XMegaPose [37]✓MegaPose+MH [37]✓64.878.174.156.942.286.370.267.5
5XYOLOX [19, 44]XOurs✓Ours+MH✓68.382.877.859.650.189.770.871.3
6✓YOLOX [19, 44]XWDR-Pose [30]XPFA [31]+KabschX79.284.996.370.652.686.789.980.0
7✓YOLOX [19, 44]XMegaPose [37]✓MegaPose+MH [37]+Teaserpp [67]✓70.471.891.659.255.387.285.574.4
8✓YOLOX [19, 44]XOurs✓Ours+MH+Kabsch✓74.278.392.864.965.292.088.379.4
9XOSOP [54]✓OSOP [54]✓OSOP+PnP+MH [54]✓31.2----49.233.2-
10XCNOS-det. [50]✓MegaPose [37]✓MegaPose+MH [37]✓56.050.868.741.934.670.662.054.9
11XCNOS-det. [50]✓Ours✓Ours+MH✓57.553.069.145.640.874.563.957.8
12✓OSOP [54]✓OSOP [54]✓OSOP+Kabsch+MH [54]+ICP✓48.2----60.557.2-
13✓CNOS-seg. [50]✓ZeroPose [5]✓MegaPose+MH [37]✓53.840.083.539.252.165.365.357.0
14✓CNOS-det. [50]✓MegaPose [37]✓MegaPose+MH [37]+Teaserpp [67]✓62.648.785.146.746.873.076.462.8
15✓CNOS-det. [50]✓Ours✓Ours+Kabsch+MH✓62.951.785.853.355.978.282.567.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.075, + 0.33, + 0.895, + 0.375 + ], + "angle": 0, + "content": "Table 1. 6D pose estimation results on the BOP challenge datasets. We report the Average Recall (AR) scores across the datasets for various methods. The higher score the better. The best results among the comparable methods are in bold. We denote the multi-hypotheses strategy as MH for simplicity." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.398, + 0.21, + 0.415 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.422, + 0.364, + 0.437 + ], + "angle": 0, + "content": "4.1. Datasets and Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.445, + 0.47, + 0.566 + ], + "angle": 0, + "content": "Datasets We evaluate our method on seven datasets of the BOP challenge [28]: LM-O [25], T-LESS [26], TUD-L [27], IC-BIN [13], ITODD [16], HomebrewdDB [35] and YCB-V [63]. These datasets contain cluttered real-world scenes captured from different lighting and camera sensors. Each scene consists of multiple objects including occlusions between them. The objects vary in terms of texture richness, symmetry, and usage." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.567, + 0.471, + 0.837 + ], + "angle": 0, + "content": "Evaluation Metrics Following the evaluation methodology of BOP challenge [28], we report the average recall (AR) considering three pose-error functions: Visible Surface Discrepancy (VSD), Maximum Symmetry-Aware Surface Distance (MSSD), and Maximum Symmetry-Aware Projection Distance (MSPD). VSD measures the misalignment of the visible part by computing the ratio of pixels in which the difference between the estimated distance to the camera center and ground truth is under a specific tolerance. In consideration of the global symmetry information, MSSD computes the maximum distance between the estimated camera-space coordinates and corresponding ground truth. In contrast, MSPD computes the maximum distance between the estimated image-space coordinates and the corresponding ground truth. Concerning each pose-error function, the AR is obtained by averaging recall calculated for multiple settings of the correctness thresholds (and misalignment tolerances for VSD)." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.848, + 0.307, + 0.863 + ], + "angle": 0, + "content": "4.2. BOP Benchmark Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Table 1 shows the results of our method on the BOP datasets. For a fair comparison, all reported pose initializa" + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.396, + 0.905, + 0.461 + ], + "angle": 0, + "content": "
Method # of renderings (↓)Average Recall (↑)
LM-OT-LESSTUD-LIC-BINYCB-VMEAN
Naïve57622.525.630.617.516.422.5
Ours14428.835.533.829.728.931.3
20832.940.440.031.134.735.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.473, + 0.892, + 0.515 + ], + "angle": 0, + "content": "Table 2. Ablation study of the hypotheses generation strategy for the coarse estimation. The AR scores are reported for the best pose hypothesis from the candidates from different sampling methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.544, + 0.892, + 0.649 + ], + "angle": 0, + "content": "tion and pose refinement methods are trained with synthetic data containing images rendered with BlenderProc [11]. We reported the AR scores of MegaPose and ours by setting the number of hypotheses \\( n \\) as 10 for the multi-hypotheses strategy and the number of outer updates as 5. For ours, the number of GenFlow updates per single outer update is set to 8." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.651, + 0.892, + 0.696 + ], + "angle": 0, + "content": "Unseen Objects For both RGB input (row 9-11) and RGB-D input (row 12-15), the results show that our approach outperforms all other methods for novel objects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.697, + 0.892, + 0.788 + ], + "angle": 0, + "content": "Seen Objects From row 1 to 8, the pose estimation results on the 2D detection trained for target datasets are reported. Our method still outperforms MegaPose for the same detection results. It is noticeable that our method provides competitive results to the state-of-the-art methods, though it is not the best, even without fine-tuning on the target datasets." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.801, + 0.655, + 0.817 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.901 + ], + "angle": 0, + "content": "We conduct ablation studies to validate the efficacy of our method. The ablation results are reported for the RGB inputs we mainly focus on. All experiments were performed on 5 BOP datasets: LM-O, T-LESS, TUD-L, IC-BIN, and YCB-V. We used the 2D detection results of YOLOX [19," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10045" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.089, + 0.885, + 0.201 + ], + "angle": 0, + "content": "
MethodTrainingAverage Recall (↑)
LM-OT-LESSTUD-LIC-BINYCB-VMEAN
1No shape-constraint + RANSAC-PnP [39]Lflow61.777.672.454.265.766.3
2Shape-constraintLflow, Lpose64.981.875.157.070.169.8
3Shape-constraint + CascadeLflow, Lpose65.782.375.058.969.970.4
4Shape-constraint + Confidence factorizationLflow, Lcert, Lpose64.581.875.357.370.569.9
5Shape-constraint + Confidence factorization + CascadeLflow, Lcert, Lpose65.982.076.159.569.870.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.109, + 0.21, + 0.859, + 0.225 + ], + "angle": 0, + "content": "Table 3. Ablation study of GenFlow design. The best result is in bold. Our method accomplishes the best performance (row 5)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.251, + 0.47, + 0.298 + ], + "angle": 0, + "content": "44] trained on target datasets to focus on the accuracy of 6D pose estimation by preventing the performance bottleneck caused by the detection methods for novel objects." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.336, + 0.296, + 0.35 + ], + "angle": 0, + "content": "4.3.1 Coarse Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.367, + 0.471, + 0.503 + ], + "angle": 0, + "content": "Table 2 shows the effectiveness of the GMM-based sampling method for generating pose hypotheses. Pre-defined rotations are generated following [68] for the Naive sampling and for creating GMM of our method. We used the \\( N = 576 \\) rotations for the Naive sampling, and \\( M = 72 \\) and \\( M = 104 \\) which conduct \\( 2M \\) renderings (144 and 208, respectively) for ours. The results show that our method can generate a better initial pose despite the fewer renderings and network inferences." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.542, + 0.245, + 0.558 + ], + "angle": 0, + "content": "4.3.2 GenFlow design" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.573, + 0.47, + 0.679 + ], + "angle": 0, + "content": "Table 3 shows the results on different refiner designs. All experiments are conducted using the same result of coarse estimation, which uses our GMM-based sampling with \\( M = 104 \\) without the multi-hypotheses strategy. We run 5 iterations of GenFlow refinement, each containing the 8 iterations of the inner update. We evaluated the final 6D pose from the GenFlow refiner." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.682, + 0.469, + 0.773 + ], + "angle": 0, + "content": "Shape-Constraint Learning to estimate 2D correspondences with flow loss only is suboptimal for 6D pose estimation since it makes no use of the target shape (row 1). Imposing shape-constraint on the model benefits from the guide of 3D information without loss of generalization performance (row 2-5)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.776, + 0.469, + 0.822 + ], + "angle": 0, + "content": "Cascade Architecture The proposed cascade architecture with multiple GenFlow modules contributes to the performance improvement (row 3, 5)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Confidence Factorization Factorizing the confidence weights to certainty and pose sensitivity helps to increase the accuracy of the output pose. Certainty estimation leverages the training depth data and makes our method robust to occlusion (row 4, 5)." + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.253, + 0.887, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.427, + 0.875, + 0.442 + ], + "angle": 0, + "content": "Figure 6. GenFlow vs MegaPose refiner for the RGB inputs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.469, + 0.74, + 0.485 + ], + "angle": 0, + "content": "4.3.3 MegaPose Refiner vs Ours" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.494, + 0.892, + 0.721 + ], + "angle": 0, + "content": "We compare the performance of the MegaPose refiner and ours for the RGB input. The same pose initializations from our coarse estimation were used for a fair evaluation. We reported the changes in AR score according to the iteration of the outer updates for both methods in Fig 6. The reported score is computed as the arithmetic mean of the AR scores for the five datasets. A single outer update takes 66.5 milliseconds for the MegaPose refiner and 98.2 milliseconds for ours, respectively, on an RTX 3090 GPU and Core i9-10920X CPU. Although ours is slower than the MegaPose refiner for a single refinement step, it achieves better performance for the same and even fewer iterations. Note that ours performs effectively by comparing a single-view rendered image for an outer update, while MegaPose refiner uses multiple rendered views as input." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.737, + 0.62, + 0.753 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.761, + 0.892, + 0.888 + ], + "angle": 0, + "content": "In this paper, we present GenFlow, an optical flow-based method for 6D pose refinement of novel objects. First, we generate pose hypotheses using a GMM-based sampling strategy for efficient and effective coarse estimation. Then, we refine the high-scoring hypotheses with our pose refiner, our main contribution. GenFlow iteratively refines the flow and confidence with the guidance of 3D shape. Our method achieves state-of-the-art performance of 6D pose estimation for novel objects concerning RGB and RGB-D inputs." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10046" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.143 + ], + "angle": 0, + "content": "[1] Christopher M. Bishop. Pattern recognition and machine learning, 5th Edition. Springer, 2007. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.145, + 0.468, + 0.213 + ], + "angle": 0, + "content": "[2] Michael J. Black and P. Anandan. A framework for the robust estimation of optical flow. In Fourth International Conference on Computer Vision, ICCV 1993, Berlin, Germany, 11-14 May, 1993, Proceedings, pages 231-236. IEEE Computer Society, 1993. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.215, + 0.468, + 0.284 + ], + "angle": 0, + "content": "[3] Angel X. Chang, Thomas A. Funkhouser, Leonidas J. Guibas, Pat Hanrahan, Qi-Xing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. Shapenet: An information-rich 3d model repository. CoRR, abs/1512.03012, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.285, + 0.468, + 0.366 + ], + "angle": 0, + "content": "[4] Hansheng Chen, Pichao Wang, Fan Wang, Wei Tian, Lu Xiong, and Hao Li. Epro-pnp: Generalized end-to-end probabilistic perspective-n-points for monocular object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2781-2790, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.369, + 0.468, + 0.41 + ], + "angle": 0, + "content": "[5] Jianqiu Chen, Mingshan Sun, Tianpeng Bao, Rui Zhao, Liwei Wu, and Zhenyu He. Zeropose: Cad-model-based zero-shot pose estimation, 2023. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.411, + 0.468, + 0.48 + ], + "angle": 0, + "content": "[6] Qifeng Chen and Vladlen Koltun. Full flow: Optical flow estimation by global optimization over regular grids. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 4706-4714. IEEE Computer Society, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.482, + 0.468, + 0.523 + ], + "angle": 0, + "content": "[7] Xu Chen, Zijian Dong, Jie Song, Andreas Geiger, and Otmar Hilliges. Category level object pose estimation via neural analysis-by-synthesis. 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.525, + 0.468, + 0.592 + ], + "angle": 0, + "content": "[8] Yilun Chen, Zhicheng Wang, Yuxiang Peng, Zhiqiang Zhang, Gang Yu, and Jian Sun. Cascaded pyramid network for multi-person pose estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.594, + 0.468, + 0.69 + ], + "angle": 0, + "content": "[9] Kyunghyun Cho, Bart van Merrienboer, Dzmitry Bahdanau, and Yoshua Bengio. On the properties of neural machine translation: Encoder-decoder approaches. In Proceedings of SSST@EMNLP 2014, Eighth Workshop on Syntax, Semantics and Structure in Statistical Translation, Doha, Qatar, 25 October 2014, pages 103-111. Association for Computational Linguistics, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.468, + 0.774 + ], + "angle": 0, + "content": "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2009), 20-25 June 2009, Miami, Florida, USA, pages 248-255. IEEE Computer Society, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.468, + 0.844 + ], + "angle": 0, + "content": "[11] Maximilian Denninger, Dominik Winkelbauer, Martin Sundermeyer, Wout Boerdijk, Markus Knauer, Klaus H. Strobl, Matthias Humt, and Rudolph Triebel. Blenderproc2: A procedural pipeline for photorealistic rendering. Journal of Open Source Software, 8(82):4901, 2023. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[12] Alexey Dosovitskiy, Philipp Fischer, Eddy Ilg, Philip Hausser, Caner Hazirbas, Vladimir Golkov, Patrick van der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. In Proceedings" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "of the IEEE International Conference on Computer Vision (ICCV), 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[13] Andreas Doumanoglou, Rigas Kouskouridas, Sotiris Malassiotis, and Tae-Kyun Kim. Recovering 6d object pose and predicting next-best-view in the crowd. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 3583-3592. IEEE Computer Society, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[14] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas Barlow McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. In 2022 International Conference on Robotics and Automation, ICRA 2022, Philadelphia, PA, USA, May 23-27, 2022, pages 2553-2560. IEEE, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.302, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[15] Bertram Drost, Markus Ulrich, Nassir Navab, and Slobodan Ilic. Model globally, match locally: Efficient and robust 3d object recognition. In 2010 IEEE computer society conference on computer vision and pattern recognition, pages 998-1005. IEEE, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[16] Bertram Drost, Markus Ulrich, Paul Bergmann, Philipp Hartinger, and Carsten Steger. Introducing mvtec ITODD - A dataset for 3d object recognition in industry. In 2017 IEEE International Conference on Computer Vision Workshops, ICCV Workshops 2017, Venice, Italy, October 22-29, 2017, pages 2200-2208. IEEE Computer Society, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.456, + 0.892, + 0.524 + ], + "angle": 0, + "content": "[17] Johan Edstedt, Ioannis Athanasiadis, Mårten Wadenbäck, and Michael Felsberg. Dkm: Dense kernelized feature matching for geometry estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17765-17775, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.525, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[18] Martin A. Fischler and Robert C. Bolles. Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM, 24(6):381-395, 1981. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.622 + ], + "angle": 0, + "content": "[19] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. YOLOX: exceeding YOLO series in 2021. CoRR, abs/2107.08430, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.623, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[20] Yang Hai, Rui Song, Jiaojiao Li, and Yinlin Hu. Shape-constraint recurrent flow for 6d object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4831-4840, 2023. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.892, + 0.774 + ], + "angle": 0, + "content": "[21] Rasmus Laurvig Haugaard and Anders Glent Buch. Surfemb: Dense and continuous correspondence distributions for object pose estimation with learnt surface embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6749-6758, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[22] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross B. Girshick. Mask R-CNN. In IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017, pages 2980-2988. IEEE Computer Society, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.845, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[23] Xingyi He, Jiaming Sun, Yuang Wang, Di Huang, Hujun Bao, and Xiaowei Zhou. Onepose++: Keypoint-free one-shot object pose estimation without CAD models. In Advances in Neural Information Processing Systems, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "10047" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.16 + ], + "angle": 0, + "content": "[24] Yisheng He, Yao Wang, Haoqiang Fan, Jian Sun, and Qifeng Chen. Fs6d: Few-shot 6d pose estimation of novel objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6814-6824, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.471, + 0.272 + ], + "angle": 0, + "content": "[25] Stefan Hinterstoisser, Vincent Lepetit, Slobodan Ilic, Stefan Holzer, Gary R. Bradski, Kurt Konolige, and Nassir Navab. Model based training, detection and pose estimation of texture-less 3d objects in heavily cluttered scenes. In Computer Vision - ACCV 2012 - 11th Asian Conference on Computer Vision, Daejeon, Korea, November 5-9, 2012, Revised Selected Papers, Part I, pages 548-562. Springer, 2012. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.471, + 0.369 + ], + "angle": 0, + "content": "[26] Tomas Hodan, Pavel Haluza, Stepán Obdrzánek, Jiri Matas, Manolis I. A. Lourakis, and Xenophon Zabulis. T-LESS: an RGB-D dataset for 6d pose estimation of texture-less objects. In 2017 IEEE Winter Conference on Applications of Computer Vision, WACV 2017, Santa Rosa, CA, USA, March 24-31, 2017, pages 880-888. IEEE Computer Society, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.371, + 0.469, + 0.466 + ], + "angle": 0, + "content": "[27] Tomás Hodań, Frank Michel, Eric Brachmann, Wadim Kehl, Anders Glent Buch, Dirk Kraft, Bertram Drost, Joel Vidal, Stephan Ihrke, Xenophon Zabulis, Caner Sahin, Fabian Manhardt, Federico Tombari, Tae-Kyun Kim, Jiri Matas, and Carsten Rother. BOP: Benchmark for 6D object pose estimation. European Conference on Computer Vision (ECCV), 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.468, + 0.469, + 0.55 + ], + "angle": 0, + "content": "[28] Tomás Hodan, Martin Sundermeyer, Bertram Drost, Yann Labbé, Eric Brachmann, Frank Michel, Carsten Rother, and Jiri Matas. BOP challenge 2020 on 6d object localization. In Computer Vision - ECCV 2020 Workshops - Glasgow, UK, August 23-28, 2020, Proceedings, Part II, pages 577-594. Springer, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.552, + 0.469, + 0.58 + ], + "angle": 0, + "content": "[29] Berthold K. P. Horn and Brian G. Schunck. Determining optical flow. Artif. Intell., 17(1-3):185-203, 1981. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.469, + 0.662 + ], + "angle": 0, + "content": "[30] Yinlin Hu, Sébastien Speierer, Wenzel Jakob, Pascal Fua, and Mathieu Salzmann. Wide-depth-range 6d object pose estimation in space. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 15870-15879. Computer Vision Foundation / IEEE, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.469, + 0.704 + ], + "angle": 0, + "content": "[31] Yinlin Hu, Pascal Fua, and Mathieu Salzmann. Perspective flow aggregation for data-limited 6d object pose estimation. In ECCV, 2022. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.469, + 0.773 + ], + "angle": 0, + "content": "[32] Tak-Wai Hui, Xiaou Tang, and Chen Change Loy. Liteflownet: A lightweight convolutional neural network for optical flow estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[33] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[34] Wolfgang Kabsch. A solution for the best rotation to relate two sets of vectors. Acta Crystallographica Section A: Crystal Physics, Diffraction, Theoretical and General Crystallography, 32(5):922-923, 1976. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.175 + ], + "angle": 0, + "content": "[35] Roman Kaskman, Sergey Zakharov, Ivan Shugurov, and Slobodan Ilic. Homebrewdb: RGB-D dataset for 6d pose estimation of 3d objects. In 2019 IEEE/CVF International Conference on Computer Vision Workshops, ICCV Workshops 2019, Seoul, Korea (South), October 27-28, 2019, pages 2767-2776. IEEE, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[36] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.316 + ], + "angle": 0, + "content": "[37] Yann Labbe, Lucas Manuelli, Arsalan Mousavian, Stephen Tyree, Stan Birchfield, Jonathan Tremblay, Justin Carpentier, Mathieu Aubry, Dieter Fox, and Josef Sivic. MegaPose: 6D Pose Estimation of Novel Objects via Render & Compare. In CoRL. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.319, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[38] Y. Labbe, J. Carpentier, M. Aubry, and J. Sivic. Cosyposse: Consistent multi-view multi-object 6d pose estimation. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.376, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[39] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Epnp: An accurate o(n) solution to the pnp problem. International Journal Of Computer Vision, 81:155-166, 2009. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.42, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[40] Fu Li, Ivan Shugurov, Benjamin Busam, Minglong Li, Shaowu Yang, and Slobodan Ilic. Polarmesh: A star-convex 3d shape approximation for object pose estimation. IEEE Robotics Autom. Lett., 7(2):4416-4423, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[41] Yi Li, Gu Wang, Xiangyang Ji, Yu Xiang, and Dieter Fox. Deepim: Deep iterative matching for 6d pose estimation. In European Conference on Computer Vision (ECCV), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.519, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[42] Tsung-Yi Lin, Piotr Dollar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.591, + 0.892, + 0.657 + ], + "angle": 0, + "content": "[43] Lahav Lipson, Zachary Teed, Ankit Goyal, and Jia Deng. Coupled iterative refinement for 6d multi-object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6728-6737, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[44] Xingyu Liu, Ruida Zhang, Chenyangguang Zhang, Bowen Fu, Jiwen Tang, Xiquan Liang, Jingyi Tang, Xiaotian Cheng, Yukang Zhang, Gu Wang, and Xiangyang Ji. Gdrnpp. https://github.com/shanice-1/gdrnpp_bop2022, 2022.3,7,8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.892, + 0.785 + ], + "angle": 0, + "content": "[45] Yuan Liu, Yilin Wen, Sida Peng, Cheng Lin, Xiaoxiao Long, Taku Komura, and Wenping Wang. Gen6d: Generalizable model-free 6-dof object pose estimation from rgb images. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.856 + ], + "angle": 0, + "content": "[46] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11976-11986, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[47] David G. Lowe. Object recognition from local scale-invariant features. In Proceedings of the International Conference on Computer Vision, Kerkyra, Corfu, Greece," + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10048" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "September 20-25, 1999, pages 1150-1157. IEEE Computer Society, 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.189 + ], + "angle": 0, + "content": "[48] Fabian Manhardt, Gu Wang, Benjamin Busam, Manuel Nickel, Sven Meier, Luca Minciullo, Xiangyang Ji, and Nassir Navab. Cps++: Improving class-level 6d pose and shape estimation from monocular images with self-supervised learning, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.191, + 0.468, + 0.273 + ], + "angle": 0, + "content": "[49] Van Nguyen Nguyen, Yinlin Hu, Yang Xiao, Mathieu Salzmann, and Vincent Lepetit. Templates for 3d object pose estimation revisited: Generalization to new objects and robustness to occlusions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6771-6780, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.275, + 0.468, + 0.328 + ], + "angle": 0, + "content": "[50] Van Nguyen Nguyen, Thibault Groueix, Georgy Ponimatkin, Vincent Lepetit, and Tomas Hodan. Cnos: A strong baseline for cad-based novel object segmentation. arXiv preprint arXiv:2307.11067, 2023. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.468, + 0.398 + ], + "angle": 0, + "content": "[51] Brian Okorn, Qiao Gu, Martial Hebert, and David Held. Zephyr: Zero-shot pose hypothesis rating. In IEEE International Conference on Robotics and Automation, ICRA 2021, Xi'an, China, May 30 - June 5, 2021, pages 14141-14148. IEEE, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.468, + 0.523 + ], + "angle": 0, + "content": "[52] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.468, + 0.594 + ], + "angle": 0, + "content": "[53] Keunhong Park, Arsalan Mousavian, Yu Xiang, and Dieter Fox. Latentfusion: End-to-end differentiable reconstruction and rendering for unseen object pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.468, + 0.663 + ], + "angle": 0, + "content": "[54] Ivan Shugurov, Fu Li, Benjamin Busam, and Slobodan Ilic. Osop: A multi-stage one shot object pose estimation framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6835-6844, 2022. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.468, + 0.732 + ], + "angle": 0, + "content": "[55] Andrea Simonelli, Samuel Rota Bulo, Lorenzo Porzi, Manuel Lopez-Antequera, and Peter Kontschieder. Disentangling monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.468, + 0.816 + ], + "angle": 0, + "content": "[56] Yongzhi Su, Mahdi Saleh, Torben Fetzer, Jason Rambach, Nassir Navab, Benjamin Busam, Didier Stricker, and Federico Tombari. Zebrapose: Coarse to fine surface encoding for 6 dof object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6738-6748, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.468, + 0.872 + ], + "angle": 0, + "content": "[57] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[58] Jiaming Sun, Zihao Wang, Siyu Zhang, Xingyi He, Hongcheng Zhao, Guofeng Zhang, and Xiaowei Zhou." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "Onepose: One-shot object pose estimation without cad models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6825-6834, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.149, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[59] Martin Sundermeyer, Maximilian Durner, En Yen Puang, Zoltan-Csaba Marton, Narunas Vaskevicius, Kai O. Arras, and Rudolph Triebel. Multi-path learning for object pose estimation across domains. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.221, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[60] Mingxing Tan and Quoc V. Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In Proceedings of the 36th International Conference on Machine Learning, ICML 2019, 9-15 June 2019, Long Beach, California, USA, pages 6105-6114. PMLR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.291, + 0.892, + 0.358 + ], + "angle": 0, + "content": "[61] Zachary Teed and Jia Deng. RAFT: recurrent all-pairs field transforms for optical flow. In Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II, pages 402-419. Springer, 2020. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.361, + 0.892, + 0.43 + ], + "angle": 0, + "content": "[62] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J. Guibas. Normalized object coordinate space for category-level 6d object pose and size estimation. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.432, + 0.892, + 0.472 + ], + "angle": 0, + "content": "[63] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes. 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.474, + 0.892, + 0.543 + ], + "angle": 0, + "content": "[64] Yang Xiao, Xuchong Qiu, Pierre-Alain Langlois, Mathieu Aubry, and Renaud Marlet. Pose from shape: Deep pose estimation for arbitrary 3d objects. In 30th British Machine Vision Conference 2019, BMVC 2019, Cardiff, UK, September 9-12, 2019, page 61. BMVA Press, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.545, + 0.892, + 0.599 + ], + "angle": 0, + "content": "[65] Yang Xiao, Yuming Du, and Renaud Marlet. Posecontrast: Class-agnostic object viewpoint estimation in the wild with pose-aware contrastive learning. In International Conference on 3D Vision (3DV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.601, + 0.892, + 0.683 + ], + "angle": 0, + "content": "[66] Gengshan Yang and Deva Ramanan. Volumetric correspondence networks for optical flow. In Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada, pages 793-803, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.686, + 0.892, + 0.726 + ], + "angle": 0, + "content": "[67] Heng Yang, Jingnan Shi, and Luca Carlone. TEASER: fast and certifiable point cloud registration. IEEE Trans. Robotics, 37(2):314-333, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.728, + 0.892, + 0.782 + ], + "angle": 0, + "content": "[68] Anna Yershova, Swati Jain, Steven M. LaValle, and Julie C. Mitchell. Generating uniform incremental grids on \\(SO(3)\\) using the hopf fibration. Int. J. Robotics Res., 29(7):801-812, 2010. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.785, + 0.892, + 0.84 + ], + "angle": 0, + "content": "[69] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.84 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10049" + } + ] +] \ No newline at end of file diff --git a/2024/GenFlow_ Generalizable Recurrent Flow for 6D Pose Refinement of Novel Objects/ff0e87f9-2cf2-45f0-8b79-ad267112b5a6_origin.pdf b/2024/GenFlow_ Generalizable Recurrent Flow for 6D Pose Refinement of Novel Objects/ff0e87f9-2cf2-45f0-8b79-ad267112b5a6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..385d8794ba378e0fe52a8107a0f2991562c93992 --- /dev/null +++ b/2024/GenFlow_ Generalizable Recurrent Flow for 6D Pose Refinement of Novel Objects/ff0e87f9-2cf2-45f0-8b79-ad267112b5a6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a1a3e071080ab564e0b6cc25e9a3ad6a08877d6856e1a21b574b827b615c51f4 +size 4358587 diff --git a/2024/GenFlow_ Generalizable Recurrent Flow for 6D Pose Refinement of Novel Objects/full.md b/2024/GenFlow_ Generalizable Recurrent Flow for 6D Pose Refinement of Novel Objects/full.md new file mode 100644 index 0000000000000000000000000000000000000000..abc277631f96c4995c4fc7417de3d8632d99ea76 --- /dev/null +++ b/2024/GenFlow_ Generalizable Recurrent Flow for 6D Pose Refinement of Novel Objects/full.md @@ -0,0 +1,319 @@ +# GenFlow: Generalizable Recurrent Flow for 6D Pose Refinement of Novel Objects + +Sungphill Moon* + +Hyeontae Son* + +Dongcheol Hur + +Sangwook Kim + +NAVER LABS + +{sungphill.moon, son.ht, dongcheol.hur, o.s.w.a.l.k}@naverlabs.com + +# Abstract + +Despite the progress of learning-based methods for 6D object pose estimation, the trade-off between accuracy and scalability for novel objects still exists. Specifically, previous methods for novel objects do not make good use of the target object's 3D shape information since they focus on generalization by processing the shape indirectly, making them less effective. We present GenFlow, an approach that enables both accuracy and generalization to novel objects with the guidance of the target object's shape. Our method predicts optical flow between the rendered image and the observed image and refines the 6D pose iteratively. It boosts the performance by a constraint of the 3D shape and the generalizable geometric knowledge learned from an end-to-end differentiable system. We further improve our model by designing a cascade network architecture to exploit the multi-scale correlations and coarse-to-fine refinement. GenFlow ranked first on the unseen object pose estimation benchmarks in both the RGB and RGB-D cases. It also achieves performance competitive with existing state-of-the-art methods for the seen object pose estimation without any fine-tuning. + +# 1. Introduction + +Estimating accurate 6D object pose from an RGB or RGB-D image is crucial for robotic tasks and augmented reality applications. In the standard setup of 6D object pose estimation, 3D models of target objects are given and available in training and testing. Most state-of-the-art algorithms are learning-based and require training the model per object or dataset containing few objects. However, these methods are difficult to handle novel objects since they memorize the 3D shapes of the targets and suffer from the training cost to support them. + +Recent works [5, 37, 51, 54] have proposed generalizable models that can estimate the 6D pose of novel objects, i.e., + +![](images/606dece7bed8cf402cafcfdde4e308e20119e9e6352721798c50b3c1c9d2d9f8.jpg) +Figure 1. Examples of 6D pose estimation of novel objects. Our method estimates correspondences between the input image and rendered image, and 6D object pose. + +objects unseen during training. Specifically, OSOP [54] and MegaPose [37] show remarkable performance with the generalizable pose refinement. Their refinement methods have the scalability to novel objects since the networks learn to compare the rendered image and the observed image without memorizing 3D shapes. However, their refiners avail themselves of 3D shape information indirectly. OSOP refiner is trained by the surrogate 2D-2D matching loss, which makes it suboptimal for 6D pose estimation. MegaPose refiner regresses the relative 6D pose between the rendered image and the input, but the regression makes no use of the projective geometry. + +Our research question is Is there a better design for generalizable pose refinement that takes advantage of shape information?. Recently, 2D optical flow-based methods [20, 31] show noticeable performance on 6D pose estimation. These methods estimate dense 2D-2D correspondences between the rendered image and the observed image. The estimated correspondences are lifted to 2D-3D correspondences on the 3D shape of the target object, then the 6D pose is recovered using a PnP solver [31] or pose regressor networks [20]. The optical flow-based pose refinement provides a clue for the generalization due to its robustness to the domain shift [31] and shows the potential to utilize 3D information in the model [20]. + +We propose GenFlow, the optical flow-based iterative + +refinement for the generalizable object pose estimation guided by the 3D shape information of target objects. Given an image and initial 6D pose of an object in the image, GenFlow estimates 2D dense correspondences between the rendered image and the input image and refined 6D pose through a PnP layer. This series of processes is performed iteratively, so it can improve accuracy and robustness to outliers [43]. + +We adopt the shape-constraint recurrent flow framework inspired by [20] for our model to use 3D information. This framework utilizes the shape information implicitly to enjoy both generalization and performance. Its end-to-end differentiable system learns to estimate the optical flow, confidence, and object pose. These estimations are updated iteratively complying with the 3D shape. We further introduce a cascade network design to make use of the multi-scale correlation volumes. It takes advantage of the coarse-to-fine refinement and helps the model estimate a more accurate pose. + +We validate our method on various datasets provided in the BOP challenge [28]. Our model is designed to estimate 6D pose for the RGB input but is also applicable to the RGB-D case with the RANSAC-Kabsch algorithm [18, 34]. For both RGB and RGB-D cases, our method achieves the overall best performance for the unseen object pose estimation task using the default detection [50] provided in the BOP challenge 2023. It is also notable that our method achieves competitive performance with the state-of-the-art methods for seen objects when used with better detection results. + +# 2. Related Work + +Iterative Refinement DeepIM [41] pioneered the iterative render-and-compare strategy for 6D pose estimation. It predicts the relative pose that better matches a rendered image against the observed image. By iterative refinement, the rendered and observed images become more similar. CosyPose [38] improved the idea by introducing better parameterization for rotation [69], network architecture [60], loss design [55, 62] and data augmentation. These iterative methods show accurate results and the possibility of being applied to the unseen objects [41]. In contrast to the previous works, CIR [43] introduces a novel differentiable PnP layer to utilize the projective geometry with deep learning. Unlike CIR, which assumes the RGB-D input, ours focuses on the RGB input. For the RGB input case, CIR obtains the rendered depth as the substitution of input depth for every flow update, so the inference suffers from time inefficiency. On the other hand, our refinement method requires fewer renderings. + +Optical Flow Estimation The classical approaches [2, 6, 29] dealt with optical flow estimation as a hand-crafted optimization problem in pursuit of the balance of bright- + +ness consistency and motion plausibility. Deep learning methods [12, 32, 33, 57, 66] emerged as promising alternatives of the classical methods, achieving considerable performance with faster inference times. Recently, RAFT [61] introduced a recurrent deep network architecture which has shown significant improvement over the previous methods regarding accuracy, efficiency, and generalization. PFA [31] applied the optical flow estimation to 6D pose refinement for data-limited cases. Since learning to estimate dense optical flow helps the networks to focus on the lower-level information, it could be robust to domain shift. Following PFA, SCFlow [20] further improved the performance by introducing an end-to-end trainable system using a pose regressor and shape-constraint lookup. Our method capitalizes on a differentiable PnP layer to the framework in order to make use of projective geometry. + +Unseen Object Pose Estimation Category-level 6D pose estimation [7, 40, 48, 62] tackled the case of unseen objects of which category is given in advance. Since these methods depend on the prior of the 3D shapes belonging to specific category, they struggle with handling the objects of unknown categories. Recently, several works proposed class-agnostic generalizable pose estimators that can handle unseen objects. A few works attempted the CAD model-free pose estimation [23, 24, 45, 53, 58] to cover everyday objects, but the results were only evaluated on datasets without occlusion and cluttered scenes. [49, 59, 64, 65] focused on estimating the orientation of novel objects by comparing the deep image features of the rendered image and the observed image. ZePHyR [51] proposed a generalizable scoring function combined with hand-crafted features [15, 47]. However, it uses the features selectively according to the richness of texture in the datasets. ZeroPose [5] estimates 6D pose by hierarchical feature matching which requires the depth input by its nature and pose refinement for better performance. Like ours, OSOP [54] and MegaPose [37] proposed CAD-based pose estimation methods that can process RGB and RGB-D inputs. Specifically, MegaPose achieved state-of-the-art performance by its iterative pose refiner. The pose refiner leverages the multi-view rendered images containing surface normals for utilizing rich information about the novel object's shape, anchor point, and coordinate system. Our method refines the pose with the guidance of the target object's shape without rendering multiview images per refinement step. + +# 3. Method + +This section describes our methodology for 6D pose estimation of novel objects. Given an RGB image and 3D models of target objects, which are unseen during training, our goal is to estimate the 6D poses of the objects with respect to the camera. Our pose estimation pipeline consists of 3 stages: object detection, coarse pose estimation, and + +![](images/e0d38f96da9325429b40f9986f7b3827c355feef3851a1f5ea20f6bdaffe6686.jpg) +(a) Naïve Sampling $(N = 576)$ + +![](images/aa3ec160f48a69d74c34ece7005891027a3e8024f85f54a2ba9c7fab38ebd3ad.jpg) + +![](images/fe6f8b1333f676f5f84571ce1c3e959148008c987e410a03443a5ff38a2b045d.jpg) +(b) Sampling using GMM $(M = 104)$ + +![](images/a8cca536b3a41d16a71c1c7c5b60f623554edb6dfe5c3b7167e90860d9de4301.jpg) +Figure 2. Visualization of the scores predicted by the coarse model. Each dot denotes the sampled rotation, and its size represents the relative score of the coarse model for the corresponding rotation. The green circles are boundaries of the area where the refiner is interested. Figures show that the high-scoring poses are clustered in the vicinity of the green circles. + +pose refinement. Note that the detection of novel objects is out of the scope of this paper. We first provide an overview of the whole process and then focus on the pose refinement method, which is the main contribution of this work. + +# 3.1. Preliminaries + +Given a calibrated RGB image $\mathcal{I}$ and a set of 3D models of the target objects $\{\mathcal{M}_i\}$ , we estimate the 6D poses of the visible objects in the image. A 6D pose is defined by a matrix $\mathbf{P}$ representing the rigid transformation from the object space to the camera space. Note that we assume the input image is calibrated, i.e., the intrinsic camera matrix $\mathbf{K}$ is known. + +2D Object Detection Given an RGB image, the object detection method produces a list of object detections consisting of 2D bounding boxes and associated labels. For the unseen object detection, we apply the CNOS [50] detection leveraging the foundation models [36, 52]. To show the extensive results of our work, we also apply the YOLOX detector [19] from GDRNPP [44] which is used for seen object detection. For each detection, the image $\mathcal{I}$ is cropped and further resized to the specific resolution in the following steps. The intrinsic parameters are adjusted according to the resized resolution. For convenience, the following explanation focuses on the 6D pose estimation of a single detected object of which the 3D model is $\mathcal{M}$ . + +Coarse Pose Estimation Given object detection, the coarse pose estimator predicts rough initial poses. We used the classification-based method proposed in [37] for better generalization. Given an input image and an arbitrary pose $\mathbf{P}$ for the 3D model $\mathcal{M}$ , the coarse model classifies whether or not the pose $\mathbf{P}$ could be improved to the pose in proximity to the ground-truth by the refiner. It uses a pre-defined + +set of 3D rotations of which the cardinality is $N$ to generate the various arbitrary pose hypotheses. For each rotation, a rough 3D translation is estimated for the points of the object's 3D model to be approximately inside the 2D bounding box. We extract scores of the pose hypotheses from the model and select the top- $n$ scoring poses. Such a classification-based estimation can handle novel objects, but it suffers from the cost of rendering a large number of images for the pre-defined rotations. Inspired by the property of high-scoring rotations to be sparse and clustered as shown in Fig 2, we introduce sampling using a Gaussian mixture model (GMM) [1] for efficiency. First, we choose the top $k$ -scoring samples $\{R_i = (\psi_i, \theta_i, \phi_i)\}_{i=1}^k$ from $M$ pre-defined rotations (Fig 2b left). Then we create a density distribution $p(X)$ for the Euler angle $X$ using a GMM which satisfies that + +$$ +p (X) = \sum_ {i = 1} ^ {k} \frac {1}{k} \mathcal {N} (X | \mu_ {i}, \Sigma), \mu_ {i} = R _ {i} \tag {1} +$$ + +Finally, we sample additional $M$ rotations from the density distribution $p(X)$ (Fig 2b right) and choose the best-scoring $n$ hypotheses among the $2M$ samples. This method renders $2M < N$ images while achieving better performance. We used $M = 104$ and $k = 16$ for our experiments. + +Pose Refinement Given a coarse pose estimate, we refine the pose for the rendered image to match against the input image crop. Starting from the coarse estimate $\mathbf{P}_0$ as an initial pose, we iteratively refine the pose using our refiner. We denote the $k^{th}$ 6D pose estimate of the refiner as $\mathbf{P}_k$ . For the $k^{th}$ iteration, we obtain an image $\mathcal{I}_{r,k}$ by rendering the 3D model $\mathcal{M}$ for the pose $\mathbf{P}_{k-1}$ . Our refiner then estimates the 2D optical flow $\mathbf{F}_k$ between the image $\mathcal{I}_{r,k}$ and the input image crop and the confidence weights $\mathbf{W}_k$ which are used to solve the pose $\mathbf{P}_k$ . A detailed description is in the following section. + +# 3.2. GenFlow Refinement + +# 3.2.1 Overview + +We describe the specific operation of GenFlow refiner to estimate the refined pose $\mathbf{P}_k$ from the previous result $\mathbf{P}_{k - 1}$ but hereafter, we omit the $k$ notation if possible for convenience. First, we obtain a synthetic image by rendering the 3D model $\mathcal{M}$ for the pose $\mathbf{P}$ and adjusted intrinsic parameters. Specifically, the results of the rendering consists of the RGB image, depth image and binary rendered mask, denoted by $\mathcal{I}_r\in R^{H\times W\times 3}$ , $\mathcal{D}_r\in R^{H\times W}$ , and $\mathcal{M}_r\in \{0,1\}^{H\times W}$ respectively. Then, we extract features from both images and refine the pose with the GenFlow module. + +Feature Extraction We extract the image features from the input image crop and the rendered image $\mathcal{I}_r$ with the same feature encoder and the context features from + +![](images/53dcdc771a5c4b98eb9305338aa5bf17832de51b7f7e3bdb5588082b0b881c14.jpg) +Figure 3. Overview of GenFlow refinement. We visualize the process of feature extraction and GenFlow update in the $k^{th}$ refinement. $T_{C2O}$ refers to a transformation that maps the coordinates from camera space to object space. + +the $\mathcal{I}_r$ . The feature maps from the $i^{th}$ layer of the feature encoder are used to construct the correlation volumes $\mathbf{C}_i \in R^{H_i \times W_i \times H_i \times W_i}$ by the dot product. + +GenFlow Module For each correlation volume and context feature map, we construct a module named "GenFlow module". A GenFlow module consists of a convolution-based Gated Recurrent Unit (GRU) [9], a differentiable PnP solver, and correlation lookup based on the pose-induced flow as shown in the Fig 3. GenFlow module iteratively updates the optical flow, confidence weights, and 6D pose. For the $j^{th}$ update, the output of the module is the refined 6D pose $\mathbf{P}^j$ . Note that we use subscripts for the outer updates and superscripts for the inner updates, i.e., GenFlow updates. The detailed illustration of updating the 6D pose with a GenFlow module for $i^{th}$ correlation volume is as follows. + +# 3.2.2 GenFlow Update + +GRU Update For the $j^{th}$ update, the GRU of the GenFlow module consumes the previous hidden state $h^{j-1}$ , correlation features $r^{j-1}$ and context features. The output of the GRU $\{\mathbf{h}^j, \mathbf{f}^j, \mathbf{u}^j, \mathbf{c}^j, \mathbf{s}^j\}$ includes the updated hidden state $h^j$ , optical flow $\mathbf{f}^j$ , flow upsampling mask $\mathbf{u}^j$ , certainty $\mathbf{c}^j$ and pose sensitivity $\mathbf{s}^j$ . The flow $\mathbf{f}^j \in R^{H_i \times W_i \times 2}$ is upsampled to $\mathbf{F}^j$ to match the resolution of previous $(i-1)^{th}$ + +feature map using the mask $\mathbf{u}^j$ . We apply the convex combination method of 3x3 neighbors for the optical flow proposed in [61]. The certainty $\mathbf{c}^j$ and pose sensitivity $\mathbf{s}^j$ are upsampled by bilinear interpolation, and then we multiply them to obtain the confidence weights $\mathbf{W}^j$ . The head networks for certainty and pose sensitivity are supervised by different losses, respectively. This explicit factorization of the confidence weights helps to be robust to occlusion with the help of certainty estimation as shown in Fig 4. The upsampled flow $\mathbf{F}^j$ and confidence weights $\mathbf{W}^j$ are used to establish weighted 2D-3D correspondences. + +Pose Update As we know the depth of synthetic image $\mathcal{D}_r$ for the pose $\mathbf{P}_k^0 = [\mathbf{R}_k^0 |\mathbf{t}_k^0 ]$ and the intrinsic camera matrix $\mathbf{K}$ , we can compute the object-space 3D coordinate of each pixel by + +$$ +\left( \begin{array}{l} x \\ y \\ z \end{array} \right) = \mathbf {R} _ {k} ^ {\mathrm {T}} \left(\mathbf {K} ^ {- 1} \mathcal {D} _ {r} (u, v) \left( \begin{array}{l} u \\ v \\ 1 \end{array} \right) - \mathbf {t} _ {k}\right) \tag {2} +$$ + +where $(u,v)$ is the 2D coordinate of a pixel and the $\mathcal{D}_r(u,v)$ is depth of the pixel. Through this lifting process, we obtain the 2D-3D correspondences for the flow $\mathbf{F}^j$ . We use $\mathbf{W}^j$ for the confidence of correspondences and binary rendered mask $\mathcal{M}_r$ to consider the object region only. + +To obtain the 6D pose $\mathbf{P}^j$ , we utilize the iterative PnP solver based on the Levenberg-Marquardt (LM) algorithm. + +![](images/0931a216aea849aa4cbfa6888d19dcc1afbd0d2381b50b917f04b5efd6ec7be1.jpg) +Figure 4. Visualization of the outputs of GRU. Our method factorizes the confidence weights into two terms: certainty and pose sensitivity. Certainty estimation helps to the robustness to occlusion. Pose sensitivity highlights the rich texture regions and extremities of the object. + +It optimizes the 6D pose by minimizing the sum of squared weighted reprojection error: + +$$ +\begin{array}{l} \underset {R, t} {\operatorname {a r g m i n}} \frac {1}{2} \sum_ {u} \sum_ {v} \| \mathbf {W} ^ {j} (u, v) \times (\pi (R \left( \begin{array}{l} x \\ y \\ z \end{array} \right) + t) \tag {3} \\ - (\left( \begin{array}{c} u \\ v \end{array} \right) + \mathbf {F} ^ {j} (u, v))) \| ^ {2}, \\ \end{array} +$$ + +where $\pi$ is the projection function with the intrinsic camera parameters and $(x,y,z)^{\mathrm{T}}$ is the 3D coordinate of the pixel $(u,v)$ computed by the equation 2. Due to its differentiable nature, the PnP layer enables end-to-end training. However, the backpropagation by unrolling the iterative optimization causes numerical instability, making the training process hard. For stable end-to-end training, we introduce the regularization method [4], which backpropagates the gradient through the last step of a differentiable optimization. Specifically, we first obtain a detached pose $\mathbf{P}^*$ through 3 LM steps and update the pose via another iteration of the Gauss-Newton algorithm to obtain the pose $\mathbf{P}^j$ . Note that the gradient is not backpropagated through $\mathbf{P}^*$ . + +Correlation Lookup We compute the pose-induced flow, which is the 2D optical flow between the $\mathcal{I}_r$ and the image with respect to the pose $\mathbf{P}^j$ . The pose-induced flow is calculated as the displacement of each pixel in the $\mathcal{I}_r$ by lifting and reprojecting the pixel using the depth $\mathcal{D}_r$ . To embed the 3D shape of the target object in the lookup operation [20, 43], we use the pose-induced flow rather than the estimated flow $\mathbf{f}^j$ for indexing the correlation volume $\mathbf{C}$ . This way of imposing a shape constraint on the model induces the networks to predict dense matching to comply with the target's 3D shape information. From the lookup operation, we obtain the correlation feature $r^j$ , which is the input of + +![](images/34a61526b8c2ffbc7928a0358b70caef434c7ad8c897d3a40e20098ab15be32f.jpg) +Figure 5. Cascade design of GenFlow modules. Multiple GenFlow modules are attached to each level on the feature pyramid. The last updated flow $\mathbf{F}_k^M$ and 6D pose $\mathbf{P}_k^M$ from the higher-level GenFlow module are used to initialize the flow and 6D pose of the lower-level module. With the cascade architecture, the 6D pose is recovered in a coarse to fine manner. + +GRU for the next GenFlow update. + +# 3.2.3 Cascade Architecture + +Inspired by the previous methods that exploit the multiscale features [8, 42], we design a cascade architecture using multiple GenFlow modules. Given correlation volumes $\{\mathbf{C}_i\}$ , we assign a GenFlow module to each volume. Then, we use the 2D optical flow and refined 6D pose of the last inner updates to initialize those of the next module as shown in Fig 5. This cascade architecture takes advantage of progressive refinement, which makes use of the high-level semantics to the low-level fine details. We found that it improves the accuracy of 6D pose rather than using a single GenFlow module. + +# 3.3. Training + +Dataset Both the coarse and refiner models are trained using the large dataset provided by MegaPose [37]. It is a large synthetic dataset comprising 2 million RGB-D images generated by BlenderProc [11], which contains diverse 3D models with ground-truth 6D pose annotations. The 3D shapes are collected from ShapeNet [3] and Google Scanned Object (GSO) [14] datasets, and we trained our model on both of the datasets. + +Refiner Model Our refiner model is trained similarly to the previous iterative refinement methods [37, 38]. Given an image with the ground-truth 6D pose $\mathbf{P}_{gt}$ for the 3D model $\mathcal{M}$ , we perturb the pose to generate an initial pose $\mathbf{P}_{init}$ by applying random noises to the translation and rotation of the pose respectively. Translation noise is sampled from a normal distribution with a standard deviation of (0.01, 0.01, 0.05) centimeters, and rotation noise is sampled from another normal distribution of Euler angles with a standard deviation of 15 degrees in each axis. The refiner model + +learns to predict the pose $\mathbf{P}_{gt}$ from the perturbed initial pose $\mathbf{P}_{init}$ . + +Coarse Model Given an input RGB image $\mathcal{I}$ and a pose $\mathbf{P}$ , a coarse model is trained to classify whether the refiner could estimate the 6D pose in proximity to the ground truth pose from the initial pose $\mathbf{P}$ . We randomly generate the positive and negative pose samples and train the network to classify them using the binary cross entropy loss. For the positive samples, we use the same method used to generate the perturbed initial pose in the refiner model. Negative samples are generated using the method proposed in [37]. + +# 3.4. Implementation Details + +Coarse Estimation We use $224 \times 224$ resolution for the resized input crop and the synthetic image $\mathcal{I}_r$ . The coarse model is implemented using the ConvNeXt backbone [46] and fully connected head. We used the ConvNext model pre-trained on ImageNet [10] classification to initialize the networks' weights. For inference, we select top- $n$ hypotheses using our GMM-based sampling strategy with $M = 104$ . + +GenFlow Refinement First, we obtain a synthetic image $\mathcal{I}_r$ of which resolution is $256\times 256$ and an image crop resized to the same resolution as the synthetic image with adjusted intrinsic parameters. Like the coarse model, we use the pre-trained ConvNeXt for the feature and context encoders. The feature maps from the first two layers of the feature encoder are used to construct correlation volumes $\mathbf{C}_1\in R^{\frac{H}{4}\times \frac{W}{4}\times \frac{H}{4}\times \frac{W}{4}}$ and $\mathbf{C}_2\in R^{\frac{H}{8}\times \frac{W}{8}\times \frac{H}{8}\times \frac{W}{8}}$ . Our GRU is implemented with a ConvGRU unit [61] and following four CNN-based heads for the 2D optical flow $\mathbf{f}$ , flow upsampling mask $\mathbf{u}$ , certainty $\mathbf{c}$ and pose sensitivity $\mathbf{s}$ . Two GenFlow modules are attached to the correlation volumes for the cascade architecture. They are serially connected so that the estimations from higher-level features are utilized as initial values for the next module. + +The optical flow and its upsampling mask produce the upsampled flow, which is supervised by the L1 endpoint error loss with the ground-truth optical flow. We also utilize the disentangled point matching loss for the pose loss, supervising the rotation, 2D object center, and depth of the object individually. Specifically, given an estimated pose $\mathbf{P} = [\mathbf{R}][\mathbf{t}_x,\mathbf{t}_y,\mathbf{t}_z]^{\mathrm{T}}]$ and ground truth pose $\mathbf{P}_{gt} = [\mathbf{R}^{\star}][\mathbf{t}_x^{\star},\mathbf{t}_y^{\star},\mathbf{t}_z^{\star}]^{\mathrm{T}}]$ , the pose loss is computed as follows: + +$$ +\begin{array}{l} \mathcal {L} _ {p o s e} = \mathcal {D} ([ \mathbf {R} | [ \mathbf {t} _ {x} ^ {\star}, \mathbf {t} _ {y} ^ {\star}, \mathbf {t} _ {z} ^ {\star} ] ^ {\mathrm {T}} ], \mathbf {P} _ {g t}) \\ + \mathcal {D} \left( \right.\left[ \right. \mathbf {R} ^ {\star} \left. \right|\left[ \mathbf {t} _ {x}, \mathbf {t} _ {y}, \mathbf {t} _ {z} ^ {\star} \right] ^ {\mathrm {T}}\left. \right), \mathbf {P} _ {g t}\left. \right) \tag {4} \\ + \mathcal {D} ([ \mathbf {R} ^ {\star} | [ \mathbf {t} _ {x} ^ {\star}, \mathbf {t} _ {y} ^ {\star}, \mathbf {t} _ {z} ] ^ {\mathrm {T}} ], \mathbf {P} _ {g t}) \\ \end{array} +$$ + +where the distance $\mathcal{D}$ between two 6D poses $\mathbf{P}_1$ and $\mathbf{P}_2$ using the 3D points $\mathcal{X}_{\mathcal{M}}$ on the 3D model $\mathcal{M}$ is defined as + +$$ +\mathcal {D} \left(\mathbf {P} _ {1}, \mathbf {P} _ {2}\right) = \frac {1}{\left| \mathcal {X} _ {\mathcal {M}} \right|} \sum_ {x \in \mathcal {X} _ {\mathcal {M}}} \left| \mathbf {P} _ {1} x - \mathbf {P} _ {2} x \right| \tag {5} +$$ + +To supervise the certainty $\mathbf{c}$ , we introduce the certainty estimation by classifying the depth consistency similarly to [17]. Given 6D poses $\mathbf{P}_{init}$ and $\mathbf{P}_{gt}$ for the 3D model $\mathcal{M}$ and intrinsic camera parameters $\mathbf{K}$ , we compute the projected depth $\mathcal{D}_{r\rightarrow t}$ by warping the rendered image to the target input image using the depth $\mathcal{D}_r$ . We also obtain the depth map of warped pixels $\mathcal{D}_t$ given the synthetic depth of the target image. Then we train the networks by minimizing the binary cross entropy loss $\mathcal{L}_{cert}$ between $\mathbf{c}$ and $\mathbf{c}_{r\rightarrow t}$ which is a binary mask $\mathbf{c}_{r\rightarrow t} = |\mathcal{D}_{r\rightarrow t} - \mathcal{D}_t| < d_{th}$ where $d_{th}$ is the distance threshold. It is worth noting that we place a stop-gradient on the certainty mask $\mathbf{c}^j$ when computing the pose loss $\mathcal{L}_{pose}^j$ for disentangling the certainty $\mathbf{c}$ and pose sensitivity $\mathbf{s}$ . + +Given the RGB-D images with annotated ground truth 6D poses, the overall loss is defined as + +$$ +\mathcal {L} = \sum_ {j = 1} ^ {N} \gamma^ {j - N} \left(\mathcal {L} _ {\text {f l o w}} ^ {j} + \alpha \mathcal {L} _ {\text {c e r t}} ^ {j} + \beta \mathcal {L} _ {\text {p o s e}} ^ {j}\right) \tag {6} +$$ + +following the strategy of exponentially increasing weights [61] where we use the weight $\gamma = 0.8$ and the iteration of GenFlow updates $N = 8$ in this work. In specific, the GenFlow module assigned to $\mathbf{C}_2$ outputs $\{\mathbf{P}^j\}_{j=1}^{N/2}$ and the another module outputs $\{\mathbf{P}^j\}_{j=N/2}^N$ from $\mathbf{C}_1$ . + +Multi-Hypotheses Strategy We select $n$ best-scoring hypotheses in the coarse estimation. The selected hypotheses are refined, and we choose the best estimate by evaluating the refined hypotheses using the coarse model again. This strategy requires more inference time for evaluating $n$ iterations of the rendering and refinement process, but it can boost the accuracy by avoiding the risk of the "Winner-Takes-All" strategy. + +RGB-D Input If the observed image consists of depth information, we apply the depth refinement after the GenFlow loop for every refinement step. First, we obtain the optical flow $\mathbf{F}$ and upsampled certainty from the results of the last GenFlow update. 3D-3D correspondences are established by the optical flow and the input depth. We filter out the correspondences of certainty lower than a specific threshold to remove the outliers. Then we apply the RANSAC-Kabsch [18, 34] for the depth refinement. Using certainty for filtering out the outliers is more effective than using confidence weights since the high-confidence pixels are sparse and the input depth is prone to be noisy. + +
2D LocalizationPose InitializationPose RefinementAverage Recall (↑)
RGB-D InputMethodNovel ObjectsMethodNovel ObjectsMethodNovel ObjectsLM-O T-LESS TUD-L IC-BIN ITODD HB YCB-VMEAN
1XMask-RCNN [22, 38]XCosyPose [38]XCosyPose [38]X63.364.068.558.321.665.657.457.0
2XMask-RCNN [22, 38]XSurfEmb [21]XBFGSX66.373.571.558.841.379.164.765.0
3XYOLOX [19, 44]XZebraPose [56]X-X72.981.175.659.250.492.172.972.0
4XYOLOX [19, 44]XMegaPose [37]✓MegaPose+MH [37]✓64.878.174.156.942.286.370.267.5
5XYOLOX [19, 44]XOurs✓Ours+MH✓68.382.877.859.650.189.770.871.3
6✓YOLOX [19, 44]XWDR-Pose [30]XPFA [31]+KabschX79.284.996.370.652.686.789.980.0
7✓YOLOX [19, 44]XMegaPose [37]✓MegaPose+MH [37]+Teaserpp [67]✓70.471.891.659.255.387.285.574.4
8✓YOLOX [19, 44]XOurs✓Ours+MH+Kabsch✓74.278.392.864.965.292.088.379.4
9XOSOP [54]✓OSOP [54]✓OSOP+PnP+MH [54]✓31.2----49.233.2-
10XCNOS-det. [50]✓MegaPose [37]✓MegaPose+MH [37]✓56.050.868.741.934.670.662.054.9
11XCNOS-det. [50]✓Ours✓Ours+MH✓57.553.069.145.640.874.563.957.8
12✓OSOP [54]✓OSOP [54]✓OSOP+Kabsch+MH [54]+ICP✓48.2----60.557.2-
13✓CNOS-seg. [50]✓ZeroPose [5]✓MegaPose+MH [37]✓53.840.083.539.252.165.365.357.0
14✓CNOS-det. [50]✓MegaPose [37]✓MegaPose+MH [37]+Teaserpp [67]✓62.648.785.146.746.873.076.462.8
15✓CNOS-det. [50]✓Ours✓Ours+Kabsch+MH✓62.951.785.853.355.978.282.567.2
+ +# 4. Experiments + +# 4.1. Datasets and Evaluation Metrics + +Datasets We evaluate our method on seven datasets of the BOP challenge [28]: LM-O [25], T-LESS [26], TUD-L [27], IC-BIN [13], ITODD [16], HomebrewdDB [35] and YCB-V [63]. These datasets contain cluttered real-world scenes captured from different lighting and camera sensors. Each scene consists of multiple objects including occlusions between them. The objects vary in terms of texture richness, symmetry, and usage. + +Evaluation Metrics Following the evaluation methodology of BOP challenge [28], we report the average recall (AR) considering three pose-error functions: Visible Surface Discrepancy (VSD), Maximum Symmetry-Aware Surface Distance (MSSD), and Maximum Symmetry-Aware Projection Distance (MSPD). VSD measures the misalignment of the visible part by computing the ratio of pixels in which the difference between the estimated distance to the camera center and ground truth is under a specific tolerance. In consideration of the global symmetry information, MSSD computes the maximum distance between the estimated camera-space coordinates and corresponding ground truth. In contrast, MSPD computes the maximum distance between the estimated image-space coordinates and the corresponding ground truth. Concerning each pose-error function, the AR is obtained by averaging recall calculated for multiple settings of the correctness thresholds (and misalignment tolerances for VSD). + +# 4.2. BOP Benchmark Results + +Table 1 shows the results of our method on the BOP datasets. For a fair comparison, all reported pose initializa + +Table 1. 6D pose estimation results on the BOP challenge datasets. We report the Average Recall (AR) scores across the datasets for various methods. The higher score the better. The best results among the comparable methods are in bold. We denote the multi-hypotheses strategy as MH for simplicity. + +
Method # of renderings (↓)Average Recall (↑)
LM-OT-LESSTUD-LIC-BINYCB-VMEAN
Naïve57622.525.630.617.516.422.5
Ours14428.835.533.829.728.931.3
20832.940.440.031.134.735.8
+ +Table 2. Ablation study of the hypotheses generation strategy for the coarse estimation. The AR scores are reported for the best pose hypothesis from the candidates from different sampling methods. + +tion and pose refinement methods are trained with synthetic data containing images rendered with BlenderProc [11]. We reported the AR scores of MegaPose and ours by setting the number of hypotheses $n$ as 10 for the multi-hypotheses strategy and the number of outer updates as 5. For ours, the number of GenFlow updates per single outer update is set to 8. + +Unseen Objects For both RGB input (row 9-11) and RGB-D input (row 12-15), the results show that our approach outperforms all other methods for novel objects. + +Seen Objects From row 1 to 8, the pose estimation results on the 2D detection trained for target datasets are reported. Our method still outperforms MegaPose for the same detection results. It is noticeable that our method provides competitive results to the state-of-the-art methods, though it is not the best, even without fine-tuning on the target datasets. + +# 4.3. Ablation Study + +We conduct ablation studies to validate the efficacy of our method. The ablation results are reported for the RGB inputs we mainly focus on. All experiments were performed on 5 BOP datasets: LM-O, T-LESS, TUD-L, IC-BIN, and YCB-V. We used the 2D detection results of YOLOX [19, + +
MethodTrainingAverage Recall (↑)
LM-OT-LESSTUD-LIC-BINYCB-VMEAN
1No shape-constraint + RANSAC-PnP [39]Lflow61.777.672.454.265.766.3
2Shape-constraintLflow, Lpose64.981.875.157.070.169.8
3Shape-constraint + CascadeLflow, Lpose65.782.375.058.969.970.4
4Shape-constraint + Confidence factorizationLflow, Lcert, Lpose64.581.875.357.370.569.9
5Shape-constraint + Confidence factorization + CascadeLflow, Lcert, Lpose65.982.076.159.569.870.7
+ +Table 3. Ablation study of GenFlow design. The best result is in bold. Our method accomplishes the best performance (row 5). + +44] trained on target datasets to focus on the accuracy of 6D pose estimation by preventing the performance bottleneck caused by the detection methods for novel objects. + +# 4.3.1 Coarse Pose Estimation + +Table 2 shows the effectiveness of the GMM-based sampling method for generating pose hypotheses. Pre-defined rotations are generated following [68] for the Naive sampling and for creating GMM of our method. We used the $N = 576$ rotations for the Naive sampling, and $M = 72$ and $M = 104$ which conduct $2M$ renderings (144 and 208, respectively) for ours. The results show that our method can generate a better initial pose despite the fewer renderings and network inferences. + +# 4.3.2 GenFlow design + +Table 3 shows the results on different refiner designs. All experiments are conducted using the same result of coarse estimation, which uses our GMM-based sampling with $M = 104$ without the multi-hypotheses strategy. We run 5 iterations of GenFlow refinement, each containing the 8 iterations of the inner update. We evaluated the final 6D pose from the GenFlow refiner. + +Shape-Constraint Learning to estimate 2D correspondences with flow loss only is suboptimal for 6D pose estimation since it makes no use of the target shape (row 1). Imposing shape-constraint on the model benefits from the guide of 3D information without loss of generalization performance (row 2-5). + +Cascade Architecture The proposed cascade architecture with multiple GenFlow modules contributes to the performance improvement (row 3, 5). + +Confidence Factorization Factorizing the confidence weights to certainty and pose sensitivity helps to increase the accuracy of the output pose. Certainty estimation leverages the training depth data and makes our method robust to occlusion (row 4, 5). + +![](images/0da2290d5dcb4220551cdb25a3c6c9e59c5c1292f6be71943e10efbf3983b475.jpg) +Figure 6. GenFlow vs MegaPose refiner for the RGB inputs. + +# 4.3.3 MegaPose Refiner vs Ours + +We compare the performance of the MegaPose refiner and ours for the RGB input. The same pose initializations from our coarse estimation were used for a fair evaluation. We reported the changes in AR score according to the iteration of the outer updates for both methods in Fig 6. The reported score is computed as the arithmetic mean of the AR scores for the five datasets. A single outer update takes 66.5 milliseconds for the MegaPose refiner and 98.2 milliseconds for ours, respectively, on an RTX 3090 GPU and Core i9-10920X CPU. Although ours is slower than the MegaPose refiner for a single refinement step, it achieves better performance for the same and even fewer iterations. Note that ours performs effectively by comparing a single-view rendered image for an outer update, while MegaPose refiner uses multiple rendered views as input. + +# 5. Conclusion + +In this paper, we present GenFlow, an optical flow-based method for 6D pose refinement of novel objects. First, we generate pose hypotheses using a GMM-based sampling strategy for efficient and effective coarse estimation. Then, we refine the high-scoring hypotheses with our pose refiner, our main contribution. GenFlow iteratively refines the flow and confidence with the guidance of 3D shape. Our method achieves state-of-the-art performance of 6D pose estimation for novel objects concerning RGB and RGB-D inputs. + +# References + +[1] Christopher M. Bishop. Pattern recognition and machine learning, 5th Edition. Springer, 2007. 3 +[2] Michael J. Black and P. Anandan. A framework for the robust estimation of optical flow. In Fourth International Conference on Computer Vision, ICCV 1993, Berlin, Germany, 11-14 May, 1993, Proceedings, pages 231-236. IEEE Computer Society, 1993. 2 +[3] Angel X. Chang, Thomas A. Funkhouser, Leonidas J. Guibas, Pat Hanrahan, Qi-Xing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. Shapenet: An information-rich 3d model repository. CoRR, abs/1512.03012, 2015. 5 +[4] Hansheng Chen, Pichao Wang, Fan Wang, Wei Tian, Lu Xiong, and Hao Li. Epro-pnp: Generalized end-to-end probabilistic perspective-n-points for monocular object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2781-2790, 2022. 5 +[5] Jianqiu Chen, Mingshan Sun, Tianpeng Bao, Rui Zhao, Liwei Wu, and Zhenyu He. Zeropose: Cad-model-based zero-shot pose estimation, 2023. 1, 2, 7 +[6] Qifeng Chen and Vladlen Koltun. Full flow: Optical flow estimation by global optimization over regular grids. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 4706-4714. IEEE Computer Society, 2016. 2 +[7] Xu Chen, Zijian Dong, Jie Song, Andreas Geiger, and Otmar Hilliges. Category level object pose estimation via neural analysis-by-synthesis. 2020. 2 +[8] Yilun Chen, Zhicheng Wang, Yuxiang Peng, Zhiqiang Zhang, Gang Yu, and Jian Sun. Cascaded pyramid network for multi-person pose estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 5 +[9] Kyunghyun Cho, Bart van Merrienboer, Dzmitry Bahdanau, and Yoshua Bengio. On the properties of neural machine translation: Encoder-decoder approaches. In Proceedings of SSST@EMNLP 2014, Eighth Workshop on Syntax, Semantics and Structure in Statistical Translation, Doha, Qatar, 25 October 2014, pages 103-111. Association for Computational Linguistics, 2014. 4 +[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2009), 20-25 June 2009, Miami, Florida, USA, pages 248-255. IEEE Computer Society, 2009. 6 +[11] Maximilian Denninger, Dominik Winkelbauer, Martin Sundermeyer, Wout Boerdijk, Markus Knauer, Klaus H. Strobl, Matthias Humt, and Rudolph Triebel. Blenderproc2: A procedural pipeline for photorealistic rendering. Journal of Open Source Software, 8(82):4901, 2023. 5, 7 +[12] Alexey Dosovitskiy, Philipp Fischer, Eddy Ilg, Philip Hausser, Caner Hazirbas, Vladimir Golkov, Patrick van der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. In Proceedings + +of the IEEE International Conference on Computer Vision (ICCV), 2015. 2 +[13] Andreas Doumanoglou, Rigas Kouskouridas, Sotiris Malassiotis, and Tae-Kyun Kim. Recovering 6d object pose and predicting next-best-view in the crowd. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 3583-3592. IEEE Computer Society, 2016. 7 +[14] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas Barlow McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. In 2022 International Conference on Robotics and Automation, ICRA 2022, Philadelphia, PA, USA, May 23-27, 2022, pages 2553-2560. IEEE, 2022. 5 +[15] Bertram Drost, Markus Ulrich, Nassir Navab, and Slobodan Ilic. Model globally, match locally: Efficient and robust 3d object recognition. In 2010 IEEE computer society conference on computer vision and pattern recognition, pages 998-1005. IEEE, 2010. 2 +[16] Bertram Drost, Markus Ulrich, Paul Bergmann, Philipp Hartinger, and Carsten Steger. Introducing mvtec ITODD - A dataset for 3d object recognition in industry. In 2017 IEEE International Conference on Computer Vision Workshops, ICCV Workshops 2017, Venice, Italy, October 22-29, 2017, pages 2200-2208. IEEE Computer Society, 2017. 7 +[17] Johan Edstedt, Ioannis Athanasiadis, Mårten Wadenbäck, and Michael Felsberg. Dkm: Dense kernelized feature matching for geometry estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17765-17775, 2023. 6 +[18] Martin A. Fischler and Robert C. Bolles. Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM, 24(6):381-395, 1981. 2, 6 +[19] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. YOLOX: exceeding YOLO series in 2021. CoRR, abs/2107.08430, 2021. 3, 7 +[20] Yang Hai, Rui Song, Jiaojiao Li, and Yinlin Hu. Shape-constraint recurrent flow for 6d object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4831-4840, 2023. 1, 2, 5 +[21] Rasmus Laurvig Haugaard and Anders Glent Buch. Surfemb: Dense and continuous correspondence distributions for object pose estimation with learnt surface embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6749-6758, 2022. 7 +[22] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross B. Girshick. Mask R-CNN. In IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017, pages 2980-2988. IEEE Computer Society, 2017. 7 +[23] Xingyi He, Jiaming Sun, Yuang Wang, Di Huang, Hujun Bao, and Xiaowei Zhou. Onepose++: Keypoint-free one-shot object pose estimation without CAD models. In Advances in Neural Information Processing Systems, 2022. 2 + +[24] Yisheng He, Yao Wang, Haoqiang Fan, Jian Sun, and Qifeng Chen. Fs6d: Few-shot 6d pose estimation of novel objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6814-6824, 2022. 2 +[25] Stefan Hinterstoisser, Vincent Lepetit, Slobodan Ilic, Stefan Holzer, Gary R. Bradski, Kurt Konolige, and Nassir Navab. Model based training, detection and pose estimation of texture-less 3d objects in heavily cluttered scenes. In Computer Vision - ACCV 2012 - 11th Asian Conference on Computer Vision, Daejeon, Korea, November 5-9, 2012, Revised Selected Papers, Part I, pages 548-562. Springer, 2012. 7 +[26] Tomas Hodan, Pavel Haluza, Stepán Obdrzánek, Jiri Matas, Manolis I. A. Lourakis, and Xenophon Zabulis. T-LESS: an RGB-D dataset for 6d pose estimation of texture-less objects. In 2017 IEEE Winter Conference on Applications of Computer Vision, WACV 2017, Santa Rosa, CA, USA, March 24-31, 2017, pages 880-888. IEEE Computer Society, 2017. 7 +[27] Tomás Hodań, Frank Michel, Eric Brachmann, Wadim Kehl, Anders Glent Buch, Dirk Kraft, Bertram Drost, Joel Vidal, Stephan Ihrke, Xenophon Zabulis, Caner Sahin, Fabian Manhardt, Federico Tombari, Tae-Kyun Kim, Jiri Matas, and Carsten Rother. BOP: Benchmark for 6D object pose estimation. European Conference on Computer Vision (ECCV), 2018. 7 +[28] Tomás Hodan, Martin Sundermeyer, Bertram Drost, Yann Labbé, Eric Brachmann, Frank Michel, Carsten Rother, and Jiri Matas. BOP challenge 2020 on 6d object localization. In Computer Vision - ECCV 2020 Workshops - Glasgow, UK, August 23-28, 2020, Proceedings, Part II, pages 577-594. Springer, 2020. 2, 7 +[29] Berthold K. P. Horn and Brian G. Schunck. Determining optical flow. Artif. Intell., 17(1-3):185-203, 1981. 2 +[30] Yinlin Hu, Sébastien Speierer, Wenzel Jakob, Pascal Fua, and Mathieu Salzmann. Wide-depth-range 6d object pose estimation in space. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 15870-15879. Computer Vision Foundation / IEEE, 2021. 7 +[31] Yinlin Hu, Pascal Fua, and Mathieu Salzmann. Perspective flow aggregation for data-limited 6d object pose estimation. In ECCV, 2022. 1, 2, 7 +[32] Tak-Wai Hui, Xiaou Tang, and Chen Change Loy. Liteflownet: A lightweight convolutional neural network for optical flow estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2 +[33] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2 +[34] Wolfgang Kabsch. A solution for the best rotation to relate two sets of vectors. Acta Crystallographica Section A: Crystal Physics, Diffraction, Theoretical and General Crystallography, 32(5):922-923, 1976. 2, 6 + +[35] Roman Kaskman, Sergey Zakharov, Ivan Shugurov, and Slobodan Ilic. Homebrewdb: RGB-D dataset for 6d pose estimation of 3d objects. In 2019 IEEE/CVF International Conference on Computer Vision Workshops, ICCV Workshops 2019, Seoul, Korea (South), October 27-28, 2019, pages 2767-2776. IEEE, 2019. 7 +[36] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 3 +[37] Yann Labbe, Lucas Manuelli, Arsalan Mousavian, Stephen Tyree, Stan Birchfield, Jonathan Tremblay, Justin Carpentier, Mathieu Aubry, Dieter Fox, and Josef Sivic. MegaPose: 6D Pose Estimation of Novel Objects via Render & Compare. In CoRL. 1, 2, 3, 5, 6, 7 +[38] Y. Labbe, J. Carpentier, M. Aubry, and J. Sivic. Cosyposse: Consistent multi-view multi-object 6d pose estimation. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2, 5, 7 +[39] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Epnp: An accurate o(n) solution to the pnp problem. International Journal Of Computer Vision, 81:155-166, 2009. 8 +[40] Fu Li, Ivan Shugurov, Benjamin Busam, Minglong Li, Shaowu Yang, and Slobodan Ilic. Polarmesh: A star-convex 3d shape approximation for object pose estimation. IEEE Robotics Autom. Lett., 7(2):4416-4423, 2022. 2 +[41] Yi Li, Gu Wang, Xiangyang Ji, Yu Xiang, and Dieter Fox. Deepim: Deep iterative matching for 6d pose estimation. In European Conference on Computer Vision (ECCV), 2018. 2 +[42] Tsung-Yi Lin, Piotr Dollar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 5 +[43] Lahav Lipson, Zachary Teed, Ankit Goyal, and Jia Deng. Coupled iterative refinement for 6d multi-object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6728-6737, 2022. 2, 5 +[44] Xingyu Liu, Ruida Zhang, Chenyangguang Zhang, Bowen Fu, Jiwen Tang, Xiquan Liang, Jingyi Tang, Xiaotian Cheng, Yukang Zhang, Gu Wang, and Xiangyang Ji. Gdrnpp. https://github.com/shanice-1/gdrnpp_bop2022, 2022.3,7,8 +[45] Yuan Liu, Yilin Wen, Sida Peng, Cheng Lin, Xiaoxiao Long, Taku Komura, and Wenping Wang. Gen6d: Generalizable model-free 6-dof object pose estimation from rgb images. In ECCV, 2022. 2 +[46] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11976-11986, 2022. 6 +[47] David G. Lowe. Object recognition from local scale-invariant features. In Proceedings of the International Conference on Computer Vision, Kerkyra, Corfu, Greece, + +September 20-25, 1999, pages 1150-1157. IEEE Computer Society, 1999. 2 +[48] Fabian Manhardt, Gu Wang, Benjamin Busam, Manuel Nickel, Sven Meier, Luca Minciullo, Xiangyang Ji, and Nassir Navab. Cps++: Improving class-level 6d pose and shape estimation from monocular images with self-supervised learning, 2020. 2 +[49] Van Nguyen Nguyen, Yinlin Hu, Yang Xiao, Mathieu Salzmann, and Vincent Lepetit. Templates for 3d object pose estimation revisited: Generalization to new objects and robustness to occlusions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6771-6780, 2022. 2 +[50] Van Nguyen Nguyen, Thibault Groueix, Georgy Ponimatkin, Vincent Lepetit, and Tomas Hodan. Cnos: A strong baseline for cad-based novel object segmentation. arXiv preprint arXiv:2307.11067, 2023. 2, 3, 7 +[51] Brian Okorn, Qiao Gu, Martial Hebert, and David Held. Zephyr: Zero-shot pose hypothesis rating. In IEEE International Conference on Robotics and Automation, ICRA 2021, Xi'an, China, May 30 - June 5, 2021, pages 14141-14148. IEEE, 2021. 1, 2 +[52] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 3 +[53] Keunhong Park, Arsalan Mousavian, Yu Xiang, and Dieter Fox. Latentfusion: End-to-end differentiable reconstruction and rendering for unseen object pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2 +[54] Ivan Shugurov, Fu Li, Benjamin Busam, and Slobodan Ilic. Osop: A multi-stage one shot object pose estimation framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6835-6844, 2022. 1, 2, 7 +[55] Andrea Simonelli, Samuel Rota Bulo, Lorenzo Porzi, Manuel Lopez-Antequera, and Peter Kontschieder. Disentangling monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 2 +[56] Yongzhi Su, Mahdi Saleh, Torben Fetzer, Jason Rambach, Nassir Navab, Benjamin Busam, Didier Stricker, and Federico Tombari. Zebrapose: Coarse to fine surface encoding for 6 dof object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6738-6748, 2022. 7 +[57] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2 +[58] Jiaming Sun, Zihao Wang, Siyu Zhang, Xingyi He, Hongcheng Zhao, Guofeng Zhang, and Xiaowei Zhou. + +Onepose: One-shot object pose estimation without cad models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6825-6834, 2022. 2 +[59] Martin Sundermeyer, Maximilian Durner, En Yen Puang, Zoltan-Csaba Marton, Narunas Vaskevicius, Kai O. Arras, and Rudolph Triebel. Multi-path learning for object pose estimation across domains. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2 +[60] Mingxing Tan and Quoc V. Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In Proceedings of the 36th International Conference on Machine Learning, ICML 2019, 9-15 June 2019, Long Beach, California, USA, pages 6105-6114. PMLR, 2019. 2 +[61] Zachary Teed and Jia Deng. RAFT: recurrent all-pairs field transforms for optical flow. In Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II, pages 402-419. Springer, 2020. 2, 4, 6 +[62] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J. Guibas. Normalized object coordinate space for category-level 6d object pose and size estimation. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2 +[63] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes. 2018. 7 +[64] Yang Xiao, Xuchong Qiu, Pierre-Alain Langlois, Mathieu Aubry, and Renaud Marlet. Pose from shape: Deep pose estimation for arbitrary 3d objects. In 30th British Machine Vision Conference 2019, BMVC 2019, Cardiff, UK, September 9-12, 2019, page 61. BMVA Press, 2019. 2 +[65] Yang Xiao, Yuming Du, and Renaud Marlet. Posecontrast: Class-agnostic object viewpoint estimation in the wild with pose-aware contrastive learning. In International Conference on 3D Vision (3DV), 2021. 2 +[66] Gengshan Yang and Deva Ramanan. Volumetric correspondence networks for optical flow. In Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada, pages 793-803, 2019. 2 +[67] Heng Yang, Jingnan Shi, and Luca Carlone. TEASER: fast and certifiable point cloud registration. IEEE Trans. Robotics, 37(2):314-333, 2021. 7 +[68] Anna Yershova, Swati Jain, Steven M. LaValle, and Julie C. Mitchell. Generating uniform incremental grids on $SO(3)$ using the hopf fibration. Int. J. Robotics Res., 29(7):801-812, 2010. 8 +[69] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2 \ No newline at end of file diff --git a/2024/GenFlow_ Generalizable Recurrent Flow for 6D Pose Refinement of Novel Objects/images.zip b/2024/GenFlow_ Generalizable Recurrent Flow for 6D Pose Refinement of Novel Objects/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8fcd5b488ba124085e60d78448eae3a8211241a9 --- /dev/null +++ b/2024/GenFlow_ Generalizable Recurrent Flow for 6D Pose Refinement of Novel Objects/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:503c4c035ce9a3a35db4b9bf04b49f2296f8eec6d843b4c4b63ea69c99668ec9 +size 509191 diff --git a/2024/GenFlow_ Generalizable Recurrent Flow for 6D Pose Refinement of Novel Objects/layout.json b/2024/GenFlow_ Generalizable Recurrent Flow for 6D Pose Refinement of Novel Objects/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6a49095e9702d24a362d151e28697948a08d796e --- /dev/null +++ b/2024/GenFlow_ Generalizable Recurrent Flow for 6D Pose Refinement of Novel Objects/layout.json @@ -0,0 +1,10219 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 70, + 102, + 523, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 102, + 523, + 139 + ], + "spans": [ + { + "bbox": [ + 70, + 102, + 523, + 139 + ], + "type": "text", + "content": "GenFlow: Generalizable Recurrent Flow for 6D Pose Refinement of Novel Objects" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 101, + 161, + 189, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 161, + 189, + 175 + ], + "spans": [ + { + "bbox": [ + 101, + 161, + 189, + 175 + ], + "type": "text", + "content": "Sungphill Moon*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 212, + 161, + 288, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 161, + 288, + 175 + ], + "spans": [ + { + "bbox": [ + 212, + 161, + 288, + 175 + ], + "type": "text", + "content": "Hyeontae Son*" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 162, + 390, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 162, + 390, + 175 + ], + "spans": [ + { + "bbox": [ + 313, + 162, + 390, + 175 + ], + "type": "text", + "content": "Dongcheol Hur" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 414, + 162, + 491, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 414, + 162, + 491, + 175 + ], + "spans": [ + { + "bbox": [ + 414, + 162, + 491, + 175 + ], + "type": "text", + "content": "Sangwook Kim" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 258, + 176, + 334, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 176, + 334, + 188 + ], + "spans": [ + { + "bbox": [ + 258, + 176, + 334, + 188 + ], + "type": "text", + "content": "NAVER LABS" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 191, + 473, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 191, + 473, + 203 + ], + "spans": [ + { + "bbox": [ + 116, + 191, + 473, + 203 + ], + "type": "text", + "content": "{sungphill.moon, son.ht, dongcheol.hur, o.s.w.a.l.k}@naverlabs.com" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 231, + 191, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 191, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 191, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 256, + 290, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 290, + 508 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 290, + 508 + ], + "type": "text", + "content": "Despite the progress of learning-based methods for 6D object pose estimation, the trade-off between accuracy and scalability for novel objects still exists. Specifically, previous methods for novel objects do not make good use of the target object's 3D shape information since they focus on generalization by processing the shape indirectly, making them less effective. We present GenFlow, an approach that enables both accuracy and generalization to novel objects with the guidance of the target object's shape. Our method predicts optical flow between the rendered image and the observed image and refines the 6D pose iteratively. It boosts the performance by a constraint of the 3D shape and the generalizable geometric knowledge learned from an end-to-end differentiable system. We further improve our model by designing a cascade network architecture to exploit the multi-scale correlations and coarse-to-fine refinement. GenFlow ranked first on the unseen object pose estimation benchmarks in both the RGB and RGB-D cases. It also achieves performance competitive with existing state-of-the-art methods for the seen object pose estimation without any fine-tuning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 531, + 127, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 531, + 127, + 542 + ], + "spans": [ + { + "bbox": [ + 47, + 531, + 127, + 542 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 551, + 287, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 551, + 287, + 670 + ], + "spans": [ + { + "bbox": [ + 46, + 551, + 287, + 670 + ], + "type": "text", + "content": "Estimating accurate 6D object pose from an RGB or RGB-D image is crucial for robotic tasks and augmented reality applications. In the standard setup of 6D object pose estimation, 3D models of target objects are given and available in training and testing. Most state-of-the-art algorithms are learning-based and require training the model per object or dataset containing few objects. However, these methods are difficult to handle novel objects since they memorize the 3D shapes of the targets and suffer from the training cost to support them." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 670, + 287, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 287, + 694 + ], + "type": "text", + "content": "Recent works [5, 37, 51, 54] have proposed generalizable models that can estimate the 6D pose of novel objects, i.e.," + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 318, + 234, + 535, + 331 + ], + "blocks": [ + { + "bbox": [ + 318, + 234, + 535, + 331 + ], + "lines": [ + { + "bbox": [ + 318, + 234, + 535, + 331 + ], + "spans": [ + { + "bbox": [ + 318, + 234, + 535, + 331 + ], + "type": "image", + "image_path": "606dece7bed8cf402cafcfdde4e308e20119e9e6352721798c50b3c1c9d2d9f8.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 345, + 547, + 378 + ], + "lines": [ + { + "bbox": [ + 305, + 345, + 547, + 378 + ], + "spans": [ + { + "bbox": [ + 305, + 345, + 547, + 378 + ], + "type": "text", + "content": "Figure 1. Examples of 6D pose estimation of novel objects. Our method estimates correspondences between the input image and rendered image, and 6D object pose." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 400, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 400, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 400, + 545, + 544 + ], + "type": "text", + "content": "objects unseen during training. Specifically, OSOP [54] and MegaPose [37] show remarkable performance with the generalizable pose refinement. Their refinement methods have the scalability to novel objects since the networks learn to compare the rendered image and the observed image without memorizing 3D shapes. However, their refiners avail themselves of 3D shape information indirectly. OSOP refiner is trained by the surrogate 2D-2D matching loss, which makes it suboptimal for 6D pose estimation. MegaPose refiner regresses the relative 6D pose between the rendered image and the input, but the regression makes no use of the projective geometry." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 545, + 546, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 546, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 546, + 700 + ], + "type": "text", + "content": "Our research question is Is there a better design for generalizable pose refinement that takes advantage of shape information?. Recently, 2D optical flow-based methods [20, 31] show noticeable performance on 6D pose estimation. These methods estimate dense 2D-2D correspondences between the rendered image and the observed image. The estimated correspondences are lifted to 2D-3D correspondences on the 3D shape of the target object, then the 6D pose is recovered using a PnP solver [31] or pose regressor networks [20]. The optical flow-based pose refinement provides a clue for the generalization due to its robustness to the domain shift [31] and shows the potential to utilize 3D information in the model [20]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "type": "text", + "content": "We propose GenFlow, the optical flow-based iterative" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 703, + 178, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 703, + 178, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 703, + 178, + 713 + ], + "type": "text", + "content": "*These authors contributed equally." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10039" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "content": "refinement for the generalizable object pose estimation guided by the 3D shape information of target objects. Given an image and initial 6D pose of an object in the image, GenFlow estimates 2D dense correspondences between the rendered image and the input image and refined 6D pose through a PnP layer. This series of processes is performed iteratively, so it can improve accuracy and robustness to outliers [43]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 168, + 288, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 168, + 288, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 288, + 300 + ], + "type": "text", + "content": "We adopt the shape-constraint recurrent flow framework inspired by [20] for our model to use 3D information. This framework utilizes the shape information implicitly to enjoy both generalization and performance. Its end-to-end differentiable system learns to estimate the optical flow, confidence, and object pose. These estimations are updated iteratively complying with the 3D shape. We further introduce a cascade network design to make use of the multi-scale correlation volumes. It takes advantage of the coarse-to-fine refinement and helps the model estimate a more accurate pose." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 300, + 287, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 287, + 431 + ], + "type": "text", + "content": "We validate our method on various datasets provided in the BOP challenge [28]. Our model is designed to estimate 6D pose for the RGB input but is also applicable to the RGB-D case with the RANSAC-Kabsch algorithm [18, 34]. For both RGB and RGB-D cases, our method achieves the overall best performance for the unseen object pose estimation task using the default detection [50] provided in the BOP challenge 2023. It is also notable that our method achieves competitive performance with the state-of-the-art methods for seen objects when used with better detection results." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 442, + 134, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 442, + 134, + 454 + ], + "spans": [ + { + "bbox": [ + 47, + 442, + 134, + 454 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 462, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 287, + 676 + ], + "type": "text", + "content": "Iterative Refinement DeepIM [41] pioneered the iterative render-and-compare strategy for 6D pose estimation. It predicts the relative pose that better matches a rendered image against the observed image. By iterative refinement, the rendered and observed images become more similar. CosyPose [38] improved the idea by introducing better parameterization for rotation [69], network architecture [60], loss design [55, 62] and data augmentation. These iterative methods show accurate results and the possibility of being applied to the unseen objects [41]. In contrast to the previous works, CIR [43] introduces a novel differentiable PnP layer to utilize the projective geometry with deep learning. Unlike CIR, which assumes the RGB-D input, ours focuses on the RGB input. For the RGB input case, CIR obtains the rendered depth as the substitution of input depth for every flow update, so the inference suffers from time inefficiency. On the other hand, our refinement method requires fewer renderings." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Optical Flow Estimation The classical approaches [2, 6, 29] dealt with optical flow estimation as a hand-crafted optimization problem in pursuit of the balance of bright-" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "text", + "content": "ness consistency and motion plausibility. Deep learning methods [12, 32, 33, 57, 66] emerged as promising alternatives of the classical methods, achieving considerable performance with faster inference times. Recently, RAFT [61] introduced a recurrent deep network architecture which has shown significant improvement over the previous methods regarding accuracy, efficiency, and generalization. PFA [31] applied the optical flow estimation to 6D pose refinement for data-limited cases. Since learning to estimate dense optical flow helps the networks to focus on the lower-level information, it could be robust to domain shift. Following PFA, SCFlow [20] further improved the performance by introducing an end-to-end trainable system using a pose regressor and shape-constraint lookup. Our method capitalizes on a differentiable PnP layer to the framework in order to make use of projective geometry." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 264, + 545, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 264, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 304, + 264, + 545, + 610 + ], + "type": "text", + "content": "Unseen Object Pose Estimation Category-level 6D pose estimation [7, 40, 48, 62] tackled the case of unseen objects of which category is given in advance. Since these methods depend on the prior of the 3D shapes belonging to specific category, they struggle with handling the objects of unknown categories. Recently, several works proposed class-agnostic generalizable pose estimators that can handle unseen objects. A few works attempted the CAD model-free pose estimation [23, 24, 45, 53, 58] to cover everyday objects, but the results were only evaluated on datasets without occlusion and cluttered scenes. [49, 59, 64, 65] focused on estimating the orientation of novel objects by comparing the deep image features of the rendered image and the observed image. ZePHyR [51] proposed a generalizable scoring function combined with hand-crafted features [15, 47]. However, it uses the features selectively according to the richness of texture in the datasets. ZeroPose [5] estimates 6D pose by hierarchical feature matching which requires the depth input by its nature and pose refinement for better performance. Like ours, OSOP [54] and MegaPose [37] proposed CAD-based pose estimation methods that can process RGB and RGB-D inputs. Specifically, MegaPose achieved state-of-the-art performance by its iterative pose refiner. The pose refiner leverages the multi-view rendered images containing surface normals for utilizing rich information about the novel object's shape, anchor point, and coordinate system. Our method refines the pose with the guidance of the target object's shape without rendering multiview images per refinement step." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 621, + 361, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 621, + 361, + 633 + ], + "spans": [ + { + "bbox": [ + 306, + 621, + 361, + 633 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "This section describes our methodology for 6D pose estimation of novel objects. Given an RGB image and 3D models of target objects, which are unseen during training, our goal is to estimate the 6D poses of the objects with respect to the camera. Our pose estimation pipeline consists of 3 stages: object detection, coarse pose estimation, and" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10040" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 112, + 75, + 225, + 127 + ], + "blocks": [ + { + "bbox": [ + 112, + 75, + 225, + 127 + ], + "lines": [ + { + "bbox": [ + 112, + 75, + 225, + 127 + ], + "spans": [ + { + "bbox": [ + 112, + 75, + 225, + 127 + ], + "type": "image", + "image_path": "e0d38f96da9325429b40f9986f7b3827c355feef3851a1f5ea20f6bdaffe6686.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 126, + 132, + 208, + 140 + ], + "lines": [ + { + "bbox": [ + 126, + 132, + 208, + 140 + ], + "spans": [ + { + "bbox": [ + 126, + 132, + 208, + 140 + ], + "type": "text", + "content": "(a) Naïve Sampling " + }, + { + "bbox": [ + 126, + 132, + 208, + 140 + ], + "type": "inline_equation", + "content": "(N = 576)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 246, + 118, + 281, + 145 + ], + "blocks": [ + { + "bbox": [ + 246, + 118, + 281, + 145 + ], + "lines": [ + { + "bbox": [ + 246, + 118, + 281, + 145 + ], + "spans": [ + { + "bbox": [ + 246, + 118, + 281, + 145 + ], + "type": "image", + "image_path": "aa3ec160f48a69d74c34ece7005891027a3e8024f85f54a2ba9c7fab38ebd3ad.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 53, + 149, + 165, + 202 + ], + "blocks": [ + { + "bbox": [ + 53, + 149, + 165, + 202 + ], + "lines": [ + { + "bbox": [ + 53, + 149, + 165, + 202 + ], + "spans": [ + { + "bbox": [ + 53, + 149, + 165, + 202 + ], + "type": "image", + "image_path": "fe6f8b1333f676f5f84571ce1c3e959148008c987e410a03443a5ff38a2b045d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 116, + 206, + 217, + 216 + ], + "lines": [ + { + "bbox": [ + 116, + 206, + 217, + 216 + ], + "spans": [ + { + "bbox": [ + 116, + 206, + 217, + 216 + ], + "type": "text", + "content": "(b) Sampling using GMM " + }, + { + "bbox": [ + 116, + 206, + 217, + 216 + ], + "type": "inline_equation", + "content": "(M = 104)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 170, + 149, + 282, + 202 + ], + "blocks": [ + { + "bbox": [ + 170, + 149, + 282, + 202 + ], + "lines": [ + { + "bbox": [ + 170, + 149, + 282, + 202 + ], + "spans": [ + { + "bbox": [ + 170, + 149, + 282, + 202 + ], + "type": "image", + "image_path": "a8cca536b3a41d16a71c1c7c5b60f623554edb6dfe5c3b7167e90860d9de4301.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 227, + 288, + 293 + ], + "lines": [ + { + "bbox": [ + 46, + 227, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 227, + 288, + 293 + ], + "type": "text", + "content": "Figure 2. Visualization of the scores predicted by the coarse model. Each dot denotes the sampled rotation, and its size represents the relative score of the coarse model for the corresponding rotation. The green circles are boundaries of the area where the refiner is interested. Figures show that the high-scoring poses are clustered in the vicinity of the green circles." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 316, + 287, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 287, + 363 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 287, + 363 + ], + "type": "text", + "content": "pose refinement. Note that the detection of novel objects is out of the scope of this paper. We first provide an overview of the whole process and then focus on the pose refinement method, which is the main contribution of this work." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 371, + 133, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 371, + 133, + 383 + ], + "spans": [ + { + "bbox": [ + 47, + 371, + 133, + 383 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 389, + 287, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 389, + 287, + 473 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 287, + 473 + ], + "type": "text", + "content": "Given a calibrated RGB image " + }, + { + "bbox": [ + 46, + 389, + 287, + 473 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 46, + 389, + 287, + 473 + ], + "type": "text", + "content": " and a set of 3D models of the target objects " + }, + { + "bbox": [ + 46, + 389, + 287, + 473 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{M}_i\\}" + }, + { + "bbox": [ + 46, + 389, + 287, + 473 + ], + "type": "text", + "content": ", we estimate the 6D poses of the visible objects in the image. A 6D pose is defined by a matrix " + }, + { + "bbox": [ + 46, + 389, + 287, + 473 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 46, + 389, + 287, + 473 + ], + "type": "text", + "content": " representing the rigid transformation from the object space to the camera space. Note that we assume the input image is calibrated, i.e., the intrinsic camera matrix " + }, + { + "bbox": [ + 46, + 389, + 287, + 473 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 46, + 389, + 287, + 473 + ], + "type": "text", + "content": " is known." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 474, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 287, + 628 + ], + "type": "text", + "content": "2D Object Detection Given an RGB image, the object detection method produces a list of object detections consisting of 2D bounding boxes and associated labels. For the unseen object detection, we apply the CNOS [50] detection leveraging the foundation models [36, 52]. To show the extensive results of our work, we also apply the YOLOX detector [19] from GDRNPP [44] which is used for seen object detection. For each detection, the image " + }, + { + "bbox": [ + 46, + 474, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 46, + 474, + 287, + 628 + ], + "type": "text", + "content": " is cropped and further resized to the specific resolution in the following steps. The intrinsic parameters are adjusted according to the resized resolution. For convenience, the following explanation focuses on the 6D pose estimation of a single detected object of which the 3D model is " + }, + { + "bbox": [ + 46, + 474, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 474, + 287, + 628 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": "Coarse Pose Estimation Given object detection, the coarse pose estimator predicts rough initial poses. We used the classification-based method proposed in [37] for better generalization. Given an input image and an arbitrary pose " + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": " for the 3D model " + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": ", the coarse model classifies whether or not the pose " + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": " could be improved to the pose in proximity to the ground-truth by the refiner. It uses a pre-defined" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "text", + "content": "set of 3D rotations of which the cardinality is " + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "text", + "content": " to generate the various arbitrary pose hypotheses. For each rotation, a rough 3D translation is estimated for the points of the object's 3D model to be approximately inside the 2D bounding box. We extract scores of the pose hypotheses from the model and select the top-" + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "text", + "content": " scoring poses. Such a classification-based estimation can handle novel objects, but it suffers from the cost of rendering a large number of images for the pre-defined rotations. Inspired by the property of high-scoring rotations to be sparse and clustered as shown in Fig 2, we introduce sampling using a Gaussian mixture model (GMM) [1] for efficiency. First, we choose the top " + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "text", + "content": "-scoring samples " + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "inline_equation", + "content": "\\{R_i = (\\psi_i, \\theta_i, \\phi_i)\\}_{i=1}^k" + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "text", + "content": " pre-defined rotations (Fig 2b left). Then we create a density distribution " + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "inline_equation", + "content": "p(X)" + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "text", + "content": " for the Euler angle " + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "text", + "content": " using a GMM which satisfies that" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 349, + 271, + 545, + 303 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 271, + 545, + 303 + ], + "spans": [ + { + "bbox": [ + 349, + 271, + 545, + 303 + ], + "type": "interline_equation", + "content": "p (X) = \\sum_ {i = 1} ^ {k} \\frac {1}{k} \\mathcal {N} (X | \\mu_ {i}, \\Sigma), \\mu_ {i} = R _ {i} \\tag {1}", + "image_path": "ffd4b5c58e4965910f62285344694bf98dc728fdfa1bf9f547baa0941bfd789e.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 310, + 545, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 310, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 304, + 310, + 545, + 369 + ], + "type": "text", + "content": "Finally, we sample additional " + }, + { + "bbox": [ + 304, + 310, + 545, + 369 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 310, + 545, + 369 + ], + "type": "text", + "content": " rotations from the density distribution " + }, + { + "bbox": [ + 304, + 310, + 545, + 369 + ], + "type": "inline_equation", + "content": "p(X)" + }, + { + "bbox": [ + 304, + 310, + 545, + 369 + ], + "type": "text", + "content": " (Fig 2b right) and choose the best-scoring " + }, + { + "bbox": [ + 304, + 310, + 545, + 369 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 310, + 545, + 369 + ], + "type": "text", + "content": " hypotheses among the " + }, + { + "bbox": [ + 304, + 310, + 545, + 369 + ], + "type": "inline_equation", + "content": "2M" + }, + { + "bbox": [ + 304, + 310, + 545, + 369 + ], + "type": "text", + "content": " samples. This method renders " + }, + { + "bbox": [ + 304, + 310, + 545, + 369 + ], + "type": "inline_equation", + "content": "2M < N" + }, + { + "bbox": [ + 304, + 310, + 545, + 369 + ], + "type": "text", + "content": " images while achieving better performance. We used " + }, + { + "bbox": [ + 304, + 310, + 545, + 369 + ], + "type": "inline_equation", + "content": "M = 104" + }, + { + "bbox": [ + 304, + 310, + 545, + 369 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 310, + 545, + 369 + ], + "type": "inline_equation", + "content": "k = 16" + }, + { + "bbox": [ + 304, + 310, + 545, + 369 + ], + "type": "text", + "content": " for our experiments." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 370, + 545, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 370, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 370, + 545, + 502 + ], + "type": "text", + "content": "Pose Refinement Given a coarse pose estimate, we refine the pose for the rendered image to match against the input image crop. Starting from the coarse estimate " + }, + { + "bbox": [ + 304, + 370, + 545, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_0" + }, + { + "bbox": [ + 304, + 370, + 545, + 502 + ], + "type": "text", + "content": " as an initial pose, we iteratively refine the pose using our refiner. We denote the " + }, + { + "bbox": [ + 304, + 370, + 545, + 502 + ], + "type": "inline_equation", + "content": "k^{th}" + }, + { + "bbox": [ + 304, + 370, + 545, + 502 + ], + "type": "text", + "content": " 6D pose estimate of the refiner as " + }, + { + "bbox": [ + 304, + 370, + 545, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_k" + }, + { + "bbox": [ + 304, + 370, + 545, + 502 + ], + "type": "text", + "content": ". For the " + }, + { + "bbox": [ + 304, + 370, + 545, + 502 + ], + "type": "inline_equation", + "content": "k^{th}" + }, + { + "bbox": [ + 304, + 370, + 545, + 502 + ], + "type": "text", + "content": " iteration, we obtain an image " + }, + { + "bbox": [ + 304, + 370, + 545, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{r,k}" + }, + { + "bbox": [ + 304, + 370, + 545, + 502 + ], + "type": "text", + "content": " by rendering the 3D model " + }, + { + "bbox": [ + 304, + 370, + 545, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 370, + 545, + 502 + ], + "type": "text", + "content": " for the pose " + }, + { + "bbox": [ + 304, + 370, + 545, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{k-1}" + }, + { + "bbox": [ + 304, + 370, + 545, + 502 + ], + "type": "text", + "content": ". Our refiner then estimates the 2D optical flow " + }, + { + "bbox": [ + 304, + 370, + 545, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_k" + }, + { + "bbox": [ + 304, + 370, + 545, + 502 + ], + "type": "text", + "content": " between the image " + }, + { + "bbox": [ + 304, + 370, + 545, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{r,k}" + }, + { + "bbox": [ + 304, + 370, + 545, + 502 + ], + "type": "text", + "content": " and the input image crop and the confidence weights " + }, + { + "bbox": [ + 304, + 370, + 545, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_k" + }, + { + "bbox": [ + 304, + 370, + 545, + 502 + ], + "type": "text", + "content": " which are used to solve the pose " + }, + { + "bbox": [ + 304, + 370, + 545, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_k" + }, + { + "bbox": [ + 304, + 370, + 545, + 502 + ], + "type": "text", + "content": ". A detailed description is in the following section." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 509, + 427, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 509, + 427, + 521 + ], + "spans": [ + { + "bbox": [ + 306, + 509, + 427, + 521 + ], + "type": "text", + "content": "3.2. GenFlow Refinement" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 527, + 380, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 527, + 380, + 538 + ], + "spans": [ + { + "bbox": [ + 306, + 527, + 380, + 538 + ], + "type": "text", + "content": "3.2.1 Overview" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 545, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 676 + ], + "type": "text", + "content": "We describe the specific operation of GenFlow refiner to estimate the refined pose " + }, + { + "bbox": [ + 304, + 545, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_k" + }, + { + "bbox": [ + 304, + 545, + 545, + 676 + ], + "type": "text", + "content": " from the previous result " + }, + { + "bbox": [ + 304, + 545, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{k - 1}" + }, + { + "bbox": [ + 304, + 545, + 545, + 676 + ], + "type": "text", + "content": " but hereafter, we omit the " + }, + { + "bbox": [ + 304, + 545, + 545, + 676 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 545, + 545, + 676 + ], + "type": "text", + "content": " notation if possible for convenience. First, we obtain a synthetic image by rendering the 3D model " + }, + { + "bbox": [ + 304, + 545, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 545, + 545, + 676 + ], + "type": "text", + "content": " for the pose " + }, + { + "bbox": [ + 304, + 545, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 304, + 545, + 545, + 676 + ], + "type": "text", + "content": " and adjusted intrinsic parameters. Specifically, the results of the rendering consists of the RGB image, depth image and binary rendered mask, denoted by " + }, + { + "bbox": [ + 304, + 545, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_r\\in R^{H\\times W\\times 3}" + }, + { + "bbox": [ + 304, + 545, + 545, + 676 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 545, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r\\in R^{H\\times W}" + }, + { + "bbox": [ + 304, + 545, + 545, + 676 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 545, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_r\\in \\{0,1\\}^{H\\times W}" + }, + { + "bbox": [ + 304, + 545, + 545, + 676 + ], + "type": "text", + "content": " respectively. Then, we extract features from both images and refine the pose with the GenFlow module." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "type": "text", + "content": "Feature Extraction We extract the image features from the input image crop and the rendered image " + }, + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_r" + }, + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "type": "text", + "content": " with the same feature encoder and the context features from" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10041" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 68, + 545, + 350 + ], + "blocks": [ + { + "bbox": [ + 63, + 68, + 545, + 350 + ], + "lines": [ + { + "bbox": [ + 63, + 68, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 63, + 68, + 545, + 350 + ], + "type": "image", + "image_path": "53dcdc771a5c4b98eb9305338aa5bf17832de51b7f7e3bdb5588082b0b881c14.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 358, + 546, + 383 + ], + "lines": [ + { + "bbox": [ + 46, + 358, + 546, + 383 + ], + "spans": [ + { + "bbox": [ + 46, + 358, + 546, + 383 + ], + "type": "text", + "content": "Figure 3. Overview of GenFlow refinement. We visualize the process of feature extraction and GenFlow update in the " + }, + { + "bbox": [ + 46, + 358, + 546, + 383 + ], + "type": "inline_equation", + "content": "k^{th}" + }, + { + "bbox": [ + 46, + 358, + 546, + 383 + ], + "type": "text", + "content": " refinement. " + }, + { + "bbox": [ + 46, + 358, + 546, + 383 + ], + "type": "inline_equation", + "content": "T_{C2O}" + }, + { + "bbox": [ + 46, + 358, + 546, + 383 + ], + "type": "text", + "content": " refers to a transformation that maps the coordinates from camera space to object space." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 403, + 287, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 287, + 438 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 287, + 438 + ], + "type": "text", + "content": "the " + }, + { + "bbox": [ + 46, + 403, + 287, + 438 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_r" + }, + { + "bbox": [ + 46, + 403, + 287, + 438 + ], + "type": "text", + "content": ". The feature maps from the " + }, + { + "bbox": [ + 46, + 403, + 287, + 438 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 46, + 403, + 287, + 438 + ], + "type": "text", + "content": " layer of the feature encoder are used to construct the correlation volumes " + }, + { + "bbox": [ + 46, + 403, + 287, + 438 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_i \\in R^{H_i \\times W_i \\times H_i \\times W_i}" + }, + { + "bbox": [ + 46, + 403, + 287, + 438 + ], + "type": "text", + "content": " by the dot product." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 439, + 288, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 439, + 288, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 439, + 288, + 594 + ], + "type": "text", + "content": "GenFlow Module For each correlation volume and context feature map, we construct a module named \"GenFlow module\". A GenFlow module consists of a convolution-based Gated Recurrent Unit (GRU) [9], a differentiable PnP solver, and correlation lookup based on the pose-induced flow as shown in the Fig 3. GenFlow module iteratively updates the optical flow, confidence weights, and 6D pose. For the " + }, + { + "bbox": [ + 46, + 439, + 288, + 594 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 46, + 439, + 288, + 594 + ], + "type": "text", + "content": " update, the output of the module is the refined 6D pose " + }, + { + "bbox": [ + 46, + 439, + 288, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^j" + }, + { + "bbox": [ + 46, + 439, + 288, + 594 + ], + "type": "text", + "content": ". Note that we use subscripts for the outer updates and superscripts for the inner updates, i.e., GenFlow updates. The detailed illustration of updating the 6D pose with a GenFlow module for " + }, + { + "bbox": [ + 46, + 439, + 288, + 594 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 46, + 439, + 288, + 594 + ], + "type": "text", + "content": " correlation volume is as follows." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 609, + 153, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 153, + 621 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 153, + 621 + ], + "type": "text", + "content": "3.2.2 GenFlow Update" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 628, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 628, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 628, + 287, + 715 + ], + "type": "text", + "content": "GRU Update For the " + }, + { + "bbox": [ + 46, + 628, + 287, + 715 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 46, + 628, + 287, + 715 + ], + "type": "text", + "content": " update, the GRU of the GenFlow module consumes the previous hidden state " + }, + { + "bbox": [ + 46, + 628, + 287, + 715 + ], + "type": "inline_equation", + "content": "h^{j-1}" + }, + { + "bbox": [ + 46, + 628, + 287, + 715 + ], + "type": "text", + "content": ", correlation features " + }, + { + "bbox": [ + 46, + 628, + 287, + 715 + ], + "type": "inline_equation", + "content": "r^{j-1}" + }, + { + "bbox": [ + 46, + 628, + 287, + 715 + ], + "type": "text", + "content": " and context features. The output of the GRU " + }, + { + "bbox": [ + 46, + 628, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{h}^j, \\mathbf{f}^j, \\mathbf{u}^j, \\mathbf{c}^j, \\mathbf{s}^j\\}" + }, + { + "bbox": [ + 46, + 628, + 287, + 715 + ], + "type": "text", + "content": " includes the updated hidden state " + }, + { + "bbox": [ + 46, + 628, + 287, + 715 + ], + "type": "inline_equation", + "content": "h^j" + }, + { + "bbox": [ + 46, + 628, + 287, + 715 + ], + "type": "text", + "content": ", optical flow " + }, + { + "bbox": [ + 46, + 628, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{f}^j" + }, + { + "bbox": [ + 46, + 628, + 287, + 715 + ], + "type": "text", + "content": ", flow upsampling mask " + }, + { + "bbox": [ + 46, + 628, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{u}^j" + }, + { + "bbox": [ + 46, + 628, + 287, + 715 + ], + "type": "text", + "content": ", certainty " + }, + { + "bbox": [ + 46, + 628, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{c}^j" + }, + { + "bbox": [ + 46, + 628, + 287, + 715 + ], + "type": "text", + "content": " and pose sensitivity " + }, + { + "bbox": [ + 46, + 628, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{s}^j" + }, + { + "bbox": [ + 46, + 628, + 287, + 715 + ], + "type": "text", + "content": ". The flow " + }, + { + "bbox": [ + 46, + 628, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{f}^j \\in R^{H_i \\times W_i \\times 2}" + }, + { + "bbox": [ + 46, + 628, + 287, + 715 + ], + "type": "text", + "content": " is upsampled to " + }, + { + "bbox": [ + 46, + 628, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^j" + }, + { + "bbox": [ + 46, + 628, + 287, + 715 + ], + "type": "text", + "content": " to match the resolution of previous " + }, + { + "bbox": [ + 46, + 628, + 287, + 715 + ], + "type": "inline_equation", + "content": "(i-1)^{th}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 403, + 546, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 546, + 534 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 546, + 534 + ], + "type": "text", + "content": "feature map using the mask " + }, + { + "bbox": [ + 304, + 403, + 546, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{u}^j" + }, + { + "bbox": [ + 304, + 403, + 546, + 534 + ], + "type": "text", + "content": ". We apply the convex combination method of 3x3 neighbors for the optical flow proposed in [61]. The certainty " + }, + { + "bbox": [ + 304, + 403, + 546, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{c}^j" + }, + { + "bbox": [ + 304, + 403, + 546, + 534 + ], + "type": "text", + "content": " and pose sensitivity " + }, + { + "bbox": [ + 304, + 403, + 546, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{s}^j" + }, + { + "bbox": [ + 304, + 403, + 546, + 534 + ], + "type": "text", + "content": " are upsampled by bilinear interpolation, and then we multiply them to obtain the confidence weights " + }, + { + "bbox": [ + 304, + 403, + 546, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^j" + }, + { + "bbox": [ + 304, + 403, + 546, + 534 + ], + "type": "text", + "content": ". The head networks for certainty and pose sensitivity are supervised by different losses, respectively. This explicit factorization of the confidence weights helps to be robust to occlusion with the help of certainty estimation as shown in Fig 4. The upsampled flow " + }, + { + "bbox": [ + 304, + 403, + 546, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^j" + }, + { + "bbox": [ + 304, + 403, + 546, + 534 + ], + "type": "text", + "content": " and confidence weights " + }, + { + "bbox": [ + 304, + 403, + 546, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^j" + }, + { + "bbox": [ + 304, + 403, + 546, + 534 + ], + "type": "text", + "content": " are used to establish weighted 2D-3D correspondences." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 534, + 545, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 583 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 583 + ], + "type": "text", + "content": "Pose Update As we know the depth of synthetic image " + }, + { + "bbox": [ + 304, + 534, + 545, + 583 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 304, + 534, + 545, + 583 + ], + "type": "text", + "content": " for the pose " + }, + { + "bbox": [ + 304, + 534, + 545, + 583 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_k^0 = [\\mathbf{R}_k^0 |\\mathbf{t}_k^0 ]" + }, + { + "bbox": [ + 304, + 534, + 545, + 583 + ], + "type": "text", + "content": " and the intrinsic camera matrix " + }, + { + "bbox": [ + 304, + 534, + 545, + 583 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 304, + 534, + 545, + 583 + ], + "type": "text", + "content": ", we can compute the object-space 3D coordinate of each pixel by" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 346, + 586, + 545, + 625 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 586, + 545, + 625 + ], + "spans": [ + { + "bbox": [ + 346, + 586, + 545, + 625 + ], + "type": "interline_equation", + "content": "\\left( \\begin{array}{l} x \\\\ y \\\\ z \\end{array} \\right) = \\mathbf {R} _ {k} ^ {\\mathrm {T}} \\left(\\mathbf {K} ^ {- 1} \\mathcal {D} _ {r} (u, v) \\left( \\begin{array}{l} u \\\\ v \\\\ 1 \\end{array} \\right) - \\mathbf {t} _ {k}\\right) \\tag {2}", + "image_path": "699b626b2b0aa0d6e91820bcda126ad3e634a160a4112aa8c89a7a2e96fb6e54.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 629, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 629, + 545, + 689 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 304, + 629, + 545, + 689 + ], + "type": "text", + "content": " is the 2D coordinate of a pixel and the " + }, + { + "bbox": [ + 304, + 629, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r(u,v)" + }, + { + "bbox": [ + 304, + 629, + 545, + 689 + ], + "type": "text", + "content": " is depth of the pixel. Through this lifting process, we obtain the 2D-3D correspondences for the flow " + }, + { + "bbox": [ + 304, + 629, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^j" + }, + { + "bbox": [ + 304, + 629, + 545, + 689 + ], + "type": "text", + "content": ". We use " + }, + { + "bbox": [ + 304, + 629, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^j" + }, + { + "bbox": [ + 304, + 629, + 545, + 689 + ], + "type": "text", + "content": " for the confidence of correspondences and binary rendered mask " + }, + { + "bbox": [ + 304, + 629, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_r" + }, + { + "bbox": [ + 304, + 629, + 545, + 689 + ], + "type": "text", + "content": " to consider the object region only." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "To obtain the 6D pose " + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^j" + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": ", we utilize the iterative PnP solver based on the Levenberg-Marquardt (LM) algorithm." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10042" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 286, + 220 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 286, + 220 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 286, + 220 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 286, + 220 + ], + "type": "image", + "image_path": "0931a216aea849aa4cbfa6888d19dcc1afbd0d2381b50b917f04b5efd6ec7be1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 231, + 288, + 285 + ], + "lines": [ + { + "bbox": [ + 46, + 231, + 288, + 285 + ], + "spans": [ + { + "bbox": [ + 46, + 231, + 288, + 285 + ], + "type": "text", + "content": "Figure 4. Visualization of the outputs of GRU. Our method factorizes the confidence weights into two terms: certainty and pose sensitivity. Certainty estimation helps to the robustness to occlusion. Pose sensitivity highlights the rich texture regions and extremities of the object." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 307, + 287, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 307, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 47, + 307, + 287, + 331 + ], + "type": "text", + "content": "It optimizes the 6D pose by minimizing the sum of squared weighted reprojection error:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 339, + 287, + 406 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 339, + 287, + 406 + ], + "spans": [ + { + "bbox": [ + 62, + 339, + 287, + 406 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\underset {R, t} {\\operatorname {a r g m i n}} \\frac {1}{2} \\sum_ {u} \\sum_ {v} \\| \\mathbf {W} ^ {j} (u, v) \\times (\\pi (R \\left( \\begin{array}{l} x \\\\ y \\\\ z \\end{array} \\right) + t) \\tag {3} \\\\ - (\\left( \\begin{array}{c} u \\\\ v \\end{array} \\right) + \\mathbf {F} ^ {j} (u, v))) \\| ^ {2}, \\\\ \\end{array}", + "image_path": "efac913e5a42f4e6d78a4291c861f5c2a16c62eda0a129616a729756578f71bb.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 415, + 287, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 287, + 569 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 415, + 287, + 569 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 46, + 415, + 287, + 569 + ], + "type": "text", + "content": " is the projection function with the intrinsic camera parameters and " + }, + { + "bbox": [ + 46, + 415, + 287, + 569 + ], + "type": "inline_equation", + "content": "(x,y,z)^{\\mathrm{T}}" + }, + { + "bbox": [ + 46, + 415, + 287, + 569 + ], + "type": "text", + "content": " is the 3D coordinate of the pixel " + }, + { + "bbox": [ + 46, + 415, + 287, + 569 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 46, + 415, + 287, + 569 + ], + "type": "text", + "content": " computed by the equation 2. Due to its differentiable nature, the PnP layer enables end-to-end training. However, the backpropagation by unrolling the iterative optimization causes numerical instability, making the training process hard. For stable end-to-end training, we introduce the regularization method [4], which backpropagates the gradient through the last step of a differentiable optimization. Specifically, we first obtain a detached pose " + }, + { + "bbox": [ + 46, + 415, + 287, + 569 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^*" + }, + { + "bbox": [ + 46, + 415, + 287, + 569 + ], + "type": "text", + "content": " through 3 LM steps and update the pose via another iteration of the Gauss-Newton algorithm to obtain the pose " + }, + { + "bbox": [ + 46, + 415, + 287, + 569 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^j" + }, + { + "bbox": [ + 46, + 415, + 287, + 569 + ], + "type": "text", + "content": ". Note that the gradient is not backpropagated through " + }, + { + "bbox": [ + 46, + 415, + 287, + 569 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^*" + }, + { + "bbox": [ + 46, + 415, + 287, + 569 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "content": "Correlation Lookup We compute the pose-induced flow, which is the 2D optical flow between the " + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_r" + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "content": " and the image with respect to the pose " + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^j" + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "content": ". The pose-induced flow is calculated as the displacement of each pixel in the " + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_r" + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "content": " by lifting and reprojecting the pixel using the depth " + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "content": ". To embed the 3D shape of the target object in the lookup operation [20, 43], we use the pose-induced flow rather than the estimated flow " + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{f}^j" + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "content": " for indexing the correlation volume " + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "content": ". This way of imposing a shape constraint on the model induces the networks to predict dense matching to comply with the target's 3D shape information. From the lookup operation, we obtain the correlation feature " + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "inline_equation", + "content": "r^j" + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "content": ", which is the input of" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 545, + 188 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 545, + 188 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 545, + 188 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 545, + 188 + ], + "type": "image", + "image_path": "34a61526b8c2ffbc7928a0358b70caef434c7ad8c897d3a40e20098ab15be32f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 198, + 547, + 265 + ], + "lines": [ + { + "bbox": [ + 304, + 198, + 547, + 265 + ], + "spans": [ + { + "bbox": [ + 304, + 198, + 547, + 265 + ], + "type": "text", + "content": "Figure 5. Cascade design of GenFlow modules. Multiple GenFlow modules are attached to each level on the feature pyramid. The last updated flow " + }, + { + "bbox": [ + 304, + 198, + 547, + 265 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_k^M" + }, + { + "bbox": [ + 304, + 198, + 547, + 265 + ], + "type": "text", + "content": " and 6D pose " + }, + { + "bbox": [ + 304, + 198, + 547, + 265 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_k^M" + }, + { + "bbox": [ + 304, + 198, + 547, + 265 + ], + "type": "text", + "content": " from the higher-level GenFlow module are used to initialize the flow and 6D pose of the lower-level module. With the cascade architecture, the 6D pose is recovered in a coarse to fine manner." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 288, + 448, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 288, + 448, + 300 + ], + "spans": [ + { + "bbox": [ + 306, + 288, + 448, + 300 + ], + "type": "text", + "content": "GRU for the next GenFlow update." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 317, + 432, + 328 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 317, + 432, + 328 + ], + "spans": [ + { + "bbox": [ + 306, + 317, + 432, + 328 + ], + "type": "text", + "content": "3.2.3 Cascade Architecture" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 338, + 545, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 338, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 545, + 469 + ], + "type": "text", + "content": "Inspired by the previous methods that exploit the multiscale features [8, 42], we design a cascade architecture using multiple GenFlow modules. Given correlation volumes " + }, + { + "bbox": [ + 304, + 338, + 545, + 469 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{C}_i\\}" + }, + { + "bbox": [ + 304, + 338, + 545, + 469 + ], + "type": "text", + "content": ", we assign a GenFlow module to each volume. Then, we use the 2D optical flow and refined 6D pose of the last inner updates to initialize those of the next module as shown in Fig 5. This cascade architecture takes advantage of progressive refinement, which makes use of the high-level semantics to the low-level fine details. We found that it improves the accuracy of 6D pose rather than using a single GenFlow module." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 479, + 370, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 370, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 370, + 491 + ], + "type": "text", + "content": "3.3. Training" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 498, + 545, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 592 + ], + "type": "text", + "content": "Dataset Both the coarse and refiner models are trained using the large dataset provided by MegaPose [37]. It is a large synthetic dataset comprising 2 million RGB-D images generated by BlenderProc [11], which contains diverse 3D models with ground-truth 6D pose annotations. The 3D shapes are collected from ShapeNet [3] and Google Scanned Object (GSO) [14] datasets, and we trained our model on both of the datasets." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "content": "Refiner Model Our refiner model is trained similarly to the previous iterative refinement methods [37, 38]. Given an image with the ground-truth 6D pose " + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{gt}" + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "content": " for the 3D model " + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "content": ", we perturb the pose to generate an initial pose " + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{init}" + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "content": " by applying random noises to the translation and rotation of the pose respectively. Translation noise is sampled from a normal distribution with a standard deviation of (0.01, 0.01, 0.05) centimeters, and rotation noise is sampled from another normal distribution of Euler angles with a standard deviation of 15 degrees in each axis. The refiner model" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10043" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "content": "learns to predict the pose " + }, + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{gt}" + }, + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "content": " from the perturbed initial pose " + }, + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{init}" + }, + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 97, + 287, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 97, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 46, + 97, + 287, + 205 + ], + "type": "text", + "content": "Coarse Model Given an input RGB image " + }, + { + "bbox": [ + 46, + 97, + 287, + 205 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 46, + 97, + 287, + 205 + ], + "type": "text", + "content": " and a pose " + }, + { + "bbox": [ + 46, + 97, + 287, + 205 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 46, + 97, + 287, + 205 + ], + "type": "text", + "content": ", a coarse model is trained to classify whether the refiner could estimate the 6D pose in proximity to the ground truth pose from the initial pose " + }, + { + "bbox": [ + 46, + 97, + 287, + 205 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 46, + 97, + 287, + 205 + ], + "type": "text", + "content": ". We randomly generate the positive and negative pose samples and train the network to classify them using the binary cross entropy loss. For the positive samples, we use the same method used to generate the perturbed initial pose in the refiner model. Negative samples are generated using the method proposed in [37]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 217, + 180, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 217, + 180, + 228 + ], + "spans": [ + { + "bbox": [ + 47, + 217, + 180, + 228 + ], + "type": "text", + "content": "3.4. Implementation Details" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 236, + 287, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 236, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 287, + 331 + ], + "type": "text", + "content": "Coarse Estimation We use " + }, + { + "bbox": [ + 46, + 236, + 287, + 331 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 236, + 287, + 331 + ], + "type": "text", + "content": " resolution for the resized input crop and the synthetic image " + }, + { + "bbox": [ + 46, + 236, + 287, + 331 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_r" + }, + { + "bbox": [ + 46, + 236, + 287, + 331 + ], + "type": "text", + "content": ". The coarse model is implemented using the ConvNeXt backbone [46] and fully connected head. We used the ConvNext model pre-trained on ImageNet [10] classification to initialize the networks' weights. For inference, we select top-" + }, + { + "bbox": [ + 46, + 236, + 287, + 331 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 236, + 287, + 331 + ], + "type": "text", + "content": " hypotheses using our GMM-based sampling strategy with " + }, + { + "bbox": [ + 46, + 236, + 287, + 331 + ], + "type": "inline_equation", + "content": "M = 104" + }, + { + "bbox": [ + 46, + 236, + 287, + 331 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 333, + 287, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 333, + 287, + 511 + ], + "spans": [ + { + "bbox": [ + 46, + 333, + 287, + 511 + ], + "type": "text", + "content": "GenFlow Refinement First, we obtain a synthetic image " + }, + { + "bbox": [ + 46, + 333, + 287, + 511 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_r" + }, + { + "bbox": [ + 46, + 333, + 287, + 511 + ], + "type": "text", + "content": " of which resolution is " + }, + { + "bbox": [ + 46, + 333, + 287, + 511 + ], + "type": "inline_equation", + "content": "256\\times 256" + }, + { + "bbox": [ + 46, + 333, + 287, + 511 + ], + "type": "text", + "content": " and an image crop resized to the same resolution as the synthetic image with adjusted intrinsic parameters. Like the coarse model, we use the pre-trained ConvNeXt for the feature and context encoders. The feature maps from the first two layers of the feature encoder are used to construct correlation volumes " + }, + { + "bbox": [ + 46, + 333, + 287, + 511 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_1\\in R^{\\frac{H}{4}\\times \\frac{W}{4}\\times \\frac{H}{4}\\times \\frac{W}{4}}" + }, + { + "bbox": [ + 46, + 333, + 287, + 511 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 333, + 287, + 511 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_2\\in R^{\\frac{H}{8}\\times \\frac{W}{8}\\times \\frac{H}{8}\\times \\frac{W}{8}}" + }, + { + "bbox": [ + 46, + 333, + 287, + 511 + ], + "type": "text", + "content": ". Our GRU is implemented with a ConvGRU unit [61] and following four CNN-based heads for the 2D optical flow " + }, + { + "bbox": [ + 46, + 333, + 287, + 511 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 46, + 333, + 287, + 511 + ], + "type": "text", + "content": ", flow upsampling mask " + }, + { + "bbox": [ + 46, + 333, + 287, + 511 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 46, + 333, + 287, + 511 + ], + "type": "text", + "content": ", certainty " + }, + { + "bbox": [ + 46, + 333, + 287, + 511 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 46, + 333, + 287, + 511 + ], + "type": "text", + "content": " and pose sensitivity " + }, + { + "bbox": [ + 46, + 333, + 287, + 511 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 46, + 333, + 287, + 511 + ], + "type": "text", + "content": ". Two GenFlow modules are attached to the correlation volumes for the cascade architecture. They are serially connected so that the estimations from higher-level features are utilized as initial values for the next module." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 514, + 287, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 287, + 611 + ], + "type": "text", + "content": "The optical flow and its upsampling mask produce the upsampled flow, which is supervised by the L1 endpoint error loss with the ground-truth optical flow. We also utilize the disentangled point matching loss for the pose loss, supervising the rotation, 2D object center, and depth of the object individually. Specifically, given an estimated pose " + }, + { + "bbox": [ + 47, + 514, + 287, + 611 + ], + "type": "inline_equation", + "content": "\\mathbf{P} = [\\mathbf{R}][\\mathbf{t}_x,\\mathbf{t}_y,\\mathbf{t}_z]^{\\mathrm{T}}]" + }, + { + "bbox": [ + 47, + 514, + 287, + 611 + ], + "type": "text", + "content": " and ground truth pose " + }, + { + "bbox": [ + 47, + 514, + 287, + 611 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{gt} = [\\mathbf{R}^{\\star}][\\mathbf{t}_x^{\\star},\\mathbf{t}_y^{\\star},\\mathbf{t}_z^{\\star}]^{\\mathrm{T}}]" + }, + { + "bbox": [ + 47, + 514, + 287, + 611 + ], + "type": "text", + "content": ", the pose loss is computed as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 99, + 631, + 287, + 679 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 631, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 99, + 631, + 287, + 679 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {p o s e} = \\mathcal {D} ([ \\mathbf {R} | [ \\mathbf {t} _ {x} ^ {\\star}, \\mathbf {t} _ {y} ^ {\\star}, \\mathbf {t} _ {z} ^ {\\star} ] ^ {\\mathrm {T}} ], \\mathbf {P} _ {g t}) \\\\ + \\mathcal {D} \\left( \\right.\\left[ \\right. \\mathbf {R} ^ {\\star} \\left. \\right|\\left[ \\mathbf {t} _ {x}, \\mathbf {t} _ {y}, \\mathbf {t} _ {z} ^ {\\star} \\right] ^ {\\mathrm {T}}\\left. \\right), \\mathbf {P} _ {g t}\\left. \\right) \\tag {4} \\\\ + \\mathcal {D} ([ \\mathbf {R} ^ {\\star} | [ \\mathbf {t} _ {x} ^ {\\star}, \\mathbf {t} _ {y} ^ {\\star}, \\mathbf {t} _ {z} ] ^ {\\mathrm {T}} ], \\mathbf {P} _ {g t}) \\\\ \\end{array}", + "image_path": "360dd728bba0b3bc19baf8d7c0dcdef5cb45f035f3b76fb3773b5b67a0ec08d6.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "where the distance " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": " between two 6D poses " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_1" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_2" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": " using the 3D points " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{\\mathcal{M}}" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": " on the 3D model " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": " is defined as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 342, + 97, + 545, + 127 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 97, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 342, + 97, + 545, + 127 + ], + "type": "interline_equation", + "content": "\\mathcal {D} \\left(\\mathbf {P} _ {1}, \\mathbf {P} _ {2}\\right) = \\frac {1}{\\left| \\mathcal {X} _ {\\mathcal {M}} \\right|} \\sum_ {x \\in \\mathcal {X} _ {\\mathcal {M}}} \\left| \\mathbf {P} _ {1} x - \\mathbf {P} _ {2} x \\right| \\tag {5}", + "image_path": "b1a018f66bcdfa393359cdb5b41415fcef2ca3ad9eec1a9c51ca757e1bcf8ff8.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 145, + 545, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 145, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 145, + 545, + 312 + ], + "type": "text", + "content": "To supervise the certainty " + }, + { + "bbox": [ + 304, + 145, + 545, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 304, + 145, + 545, + 312 + ], + "type": "text", + "content": ", we introduce the certainty estimation by classifying the depth consistency similarly to [17]. Given 6D poses " + }, + { + "bbox": [ + 304, + 145, + 545, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{init}" + }, + { + "bbox": [ + 304, + 145, + 545, + 312 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 145, + 545, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{gt}" + }, + { + "bbox": [ + 304, + 145, + 545, + 312 + ], + "type": "text", + "content": " for the 3D model " + }, + { + "bbox": [ + 304, + 145, + 545, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 145, + 545, + 312 + ], + "type": "text", + "content": " and intrinsic camera parameters " + }, + { + "bbox": [ + 304, + 145, + 545, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 304, + 145, + 545, + 312 + ], + "type": "text", + "content": ", we compute the projected depth " + }, + { + "bbox": [ + 304, + 145, + 545, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{r\\rightarrow t}" + }, + { + "bbox": [ + 304, + 145, + 545, + 312 + ], + "type": "text", + "content": " by warping the rendered image to the target input image using the depth " + }, + { + "bbox": [ + 304, + 145, + 545, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 304, + 145, + 545, + 312 + ], + "type": "text", + "content": ". We also obtain the depth map of warped pixels " + }, + { + "bbox": [ + 304, + 145, + 545, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_t" + }, + { + "bbox": [ + 304, + 145, + 545, + 312 + ], + "type": "text", + "content": " given the synthetic depth of the target image. Then we train the networks by minimizing the binary cross entropy loss " + }, + { + "bbox": [ + 304, + 145, + 545, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cert}" + }, + { + "bbox": [ + 304, + 145, + 545, + 312 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 304, + 145, + 545, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 304, + 145, + 545, + 312 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 145, + 545, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{r\\rightarrow t}" + }, + { + "bbox": [ + 304, + 145, + 545, + 312 + ], + "type": "text", + "content": " which is a binary mask " + }, + { + "bbox": [ + 304, + 145, + 545, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{r\\rightarrow t} = |\\mathcal{D}_{r\\rightarrow t} - \\mathcal{D}_t| < d_{th}" + }, + { + "bbox": [ + 304, + 145, + 545, + 312 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 145, + 545, + 312 + ], + "type": "inline_equation", + "content": "d_{th}" + }, + { + "bbox": [ + 304, + 145, + 545, + 312 + ], + "type": "text", + "content": " is the distance threshold. It is worth noting that we place a stop-gradient on the certainty mask " + }, + { + "bbox": [ + 304, + 145, + 545, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{c}^j" + }, + { + "bbox": [ + 304, + 145, + 545, + 312 + ], + "type": "text", + "content": " when computing the pose loss " + }, + { + "bbox": [ + 304, + 145, + 545, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pose}^j" + }, + { + "bbox": [ + 304, + 145, + 545, + 312 + ], + "type": "text", + "content": " for disentangling the certainty " + }, + { + "bbox": [ + 304, + 145, + 545, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 304, + 145, + 545, + 312 + ], + "type": "text", + "content": " and pose sensitivity " + }, + { + "bbox": [ + 304, + 145, + 545, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 304, + 145, + 545, + 312 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 316, + 545, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 316, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 305, + 316, + 545, + 339 + ], + "type": "text", + "content": "Given the RGB-D images with annotated ground truth 6D poses, the overall loss is defined as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 337, + 355, + 545, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 355, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 337, + 355, + 545, + 389 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\sum_ {j = 1} ^ {N} \\gamma^ {j - N} \\left(\\mathcal {L} _ {\\text {f l o w}} ^ {j} + \\alpha \\mathcal {L} _ {\\text {c e r t}} ^ {j} + \\beta \\mathcal {L} _ {\\text {p o s e}} ^ {j}\\right) \\tag {6}", + "image_path": "e41e675d6dd821482e91f82375c201d328c9ec7ad828170e4395de9893e0939c.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 404, + 545, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 404, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 304, + 404, + 545, + 470 + ], + "type": "text", + "content": "following the strategy of exponentially increasing weights [61] where we use the weight " + }, + { + "bbox": [ + 304, + 404, + 545, + 470 + ], + "type": "inline_equation", + "content": "\\gamma = 0.8" + }, + { + "bbox": [ + 304, + 404, + 545, + 470 + ], + "type": "text", + "content": " and the iteration of GenFlow updates " + }, + { + "bbox": [ + 304, + 404, + 545, + 470 + ], + "type": "inline_equation", + "content": "N = 8" + }, + { + "bbox": [ + 304, + 404, + 545, + 470 + ], + "type": "text", + "content": " in this work. In specific, the GenFlow module assigned to " + }, + { + "bbox": [ + 304, + 404, + 545, + 470 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_2" + }, + { + "bbox": [ + 304, + 404, + 545, + 470 + ], + "type": "text", + "content": " outputs " + }, + { + "bbox": [ + 304, + 404, + 545, + 470 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{P}^j\\}_{j=1}^{N/2}" + }, + { + "bbox": [ + 304, + 404, + 545, + 470 + ], + "type": "text", + "content": " and the another module outputs " + }, + { + "bbox": [ + 304, + 404, + 545, + 470 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{P}^j\\}_{j=N/2}^N" + }, + { + "bbox": [ + 304, + 404, + 545, + 470 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 304, + 404, + 545, + 470 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_1" + }, + { + "bbox": [ + 304, + 404, + 545, + 470 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 471, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 471, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 471, + 545, + 567 + ], + "type": "text", + "content": "Multi-Hypotheses Strategy We select " + }, + { + "bbox": [ + 304, + 471, + 545, + 567 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 471, + 545, + 567 + ], + "type": "text", + "content": " best-scoring hypotheses in the coarse estimation. The selected hypotheses are refined, and we choose the best estimate by evaluating the refined hypotheses using the coarse model again. This strategy requires more inference time for evaluating " + }, + { + "bbox": [ + 304, + 471, + 545, + 567 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 471, + 545, + 567 + ], + "type": "text", + "content": " iterations of the rendering and refinement process, but it can boost the accuracy by avoiding the risk of the \"Winner-Takes-All\" strategy." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "RGB-D Input If the observed image consists of depth information, we apply the depth refinement after the GenFlow loop for every refinement step. First, we obtain the optical flow " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " and upsampled certainty from the results of the last GenFlow update. 3D-3D correspondences are established by the optical flow and the input depth. We filter out the correspondences of certainty lower than a specific threshold to remove the outliers. Then we apply the RANSAC-Kabsch [18, 34] for the depth refinement. Using certainty for filtering out the outliers is more effective than using confidence weights since the high-confidence pixels are sparse and the input depth is prone to be noisy." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10044" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 70, + 563, + 253 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 563, + 253 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 563, + 253 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 563, + 253 + ], + "type": "table", + "html": "
2D LocalizationPose InitializationPose RefinementAverage Recall (↑)
RGB-D InputMethodNovel ObjectsMethodNovel ObjectsMethodNovel ObjectsLM-O T-LESS TUD-L IC-BIN ITODD HB YCB-VMEAN
1XMask-RCNN [22, 38]XCosyPose [38]XCosyPose [38]X63.364.068.558.321.665.657.457.0
2XMask-RCNN [22, 38]XSurfEmb [21]XBFGSX66.373.571.558.841.379.164.765.0
3XYOLOX [19, 44]XZebraPose [56]X-X72.981.175.659.250.492.172.972.0
4XYOLOX [19, 44]XMegaPose [37]✓MegaPose+MH [37]✓64.878.174.156.942.286.370.267.5
5XYOLOX [19, 44]XOurs✓Ours+MH✓68.382.877.859.650.189.770.871.3
6✓YOLOX [19, 44]XWDR-Pose [30]XPFA [31]+KabschX79.284.996.370.652.686.789.980.0
7✓YOLOX [19, 44]XMegaPose [37]✓MegaPose+MH [37]+Teaserpp [67]✓70.471.891.659.255.387.285.574.4
8✓YOLOX [19, 44]XOurs✓Ours+MH+Kabsch✓74.278.392.864.965.292.088.379.4
9XOSOP [54]✓OSOP [54]✓OSOP+PnP+MH [54]✓31.2----49.233.2-
10XCNOS-det. [50]✓MegaPose [37]✓MegaPose+MH [37]✓56.050.868.741.934.670.662.054.9
11XCNOS-det. [50]✓Ours✓Ours+MH✓57.553.069.145.640.874.563.957.8
12✓OSOP [54]✓OSOP [54]✓OSOP+Kabsch+MH [54]+ICP✓48.2----60.557.2-
13✓CNOS-seg. [50]✓ZeroPose [5]✓MegaPose+MH [37]✓53.840.083.539.252.165.365.357.0
14✓CNOS-det. [50]✓MegaPose [37]✓MegaPose+MH [37]+Teaserpp [67]✓62.648.785.146.746.873.076.462.8
15✓CNOS-det. [50]✓Ours✓Ours+Kabsch+MH✓62.951.785.853.355.978.282.567.2
", + "image_path": "3ce43fc09fe0b6ed08000c39334057d5da74428f1443de099a7db8c1cd8bee73.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 315, + 128, + 328 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 315, + 128, + 328 + ], + "spans": [ + { + "bbox": [ + 46, + 315, + 128, + 328 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 334, + 222, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 334, + 222, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 334, + 222, + 346 + ], + "type": "text", + "content": "4.1. Datasets and Evaluation Metrics" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 352, + 287, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 352, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 352, + 287, + 448 + ], + "type": "text", + "content": "Datasets We evaluate our method on seven datasets of the BOP challenge [28]: LM-O [25], T-LESS [26], TUD-L [27], IC-BIN [13], ITODD [16], HomebrewdDB [35] and YCB-V [63]. These datasets contain cluttered real-world scenes captured from different lighting and camera sensors. Each scene consists of multiple objects including occlusions between them. The objects vary in terms of texture richness, symmetry, and usage." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 449, + 288, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 288, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 288, + 662 + ], + "type": "text", + "content": "Evaluation Metrics Following the evaluation methodology of BOP challenge [28], we report the average recall (AR) considering three pose-error functions: Visible Surface Discrepancy (VSD), Maximum Symmetry-Aware Surface Distance (MSSD), and Maximum Symmetry-Aware Projection Distance (MSPD). VSD measures the misalignment of the visible part by computing the ratio of pixels in which the difference between the estimated distance to the camera center and ground truth is under a specific tolerance. In consideration of the global symmetry information, MSSD computes the maximum distance between the estimated camera-space coordinates and corresponding ground truth. In contrast, MSPD computes the maximum distance between the estimated image-space coordinates and the corresponding ground truth. Concerning each pose-error function, the AR is obtained by averaging recall calculated for multiple settings of the correctness thresholds (and misalignment tolerances for VSD)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 671, + 187, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 671, + 187, + 683 + ], + "spans": [ + { + "bbox": [ + 46, + 671, + 187, + 683 + ], + "type": "text", + "content": "4.2. BOP Benchmark Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": "Table 1 shows the results of our method on the BOP datasets. For a fair comparison, all reported pose initializa" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 307, + 313, + 553, + 365 + ], + "blocks": [ + { + "bbox": [ + 45, + 261, + 547, + 297 + ], + "lines": [ + { + "bbox": [ + 45, + 261, + 547, + 297 + ], + "spans": [ + { + "bbox": [ + 45, + 261, + 547, + 297 + ], + "type": "text", + "content": "Table 1. 6D pose estimation results on the BOP challenge datasets. We report the Average Recall (AR) scores across the datasets for various methods. The higher score the better. The best results among the comparable methods are in bold. We denote the multi-hypotheses strategy as MH for simplicity." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 313, + 553, + 365 + ], + "lines": [ + { + "bbox": [ + 307, + 313, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 307, + 313, + 553, + 365 + ], + "type": "table", + "html": "
Method # of renderings (↓)Average Recall (↑)
LM-OT-LESSTUD-LIC-BINYCB-VMEAN
Naïve57622.525.630.617.516.422.5
Ours14428.835.533.829.728.931.3
20832.940.440.031.134.735.8
", + "image_path": "166b6174dd27fa33dfc3ca37065a687781515dec4e61ef5eb3c5c09847e7d3b0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 374, + 545, + 407 + ], + "lines": [ + { + "bbox": [ + 305, + 374, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 305, + 374, + 545, + 407 + ], + "type": "text", + "content": "Table 2. Ablation study of the hypotheses generation strategy for the coarse estimation. The AR scores are reported for the best pose hypothesis from the candidates from different sampling methods." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 430, + 545, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 430, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 304, + 430, + 545, + 514 + ], + "type": "text", + "content": "tion and pose refinement methods are trained with synthetic data containing images rendered with BlenderProc [11]. We reported the AR scores of MegaPose and ours by setting the number of hypotheses " + }, + { + "bbox": [ + 304, + 430, + 545, + 514 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 430, + 545, + 514 + ], + "type": "text", + "content": " as 10 for the multi-hypotheses strategy and the number of outer updates as 5. For ours, the number of GenFlow updates per single outer update is set to 8." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 515, + 545, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 515, + 545, + 551 + ], + "spans": [ + { + "bbox": [ + 304, + 515, + 545, + 551 + ], + "type": "text", + "content": "Unseen Objects For both RGB input (row 9-11) and RGB-D input (row 12-15), the results show that our approach outperforms all other methods for novel objects." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 552, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 552, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 552, + 545, + 624 + ], + "type": "text", + "content": "Seen Objects From row 1 to 8, the pose estimation results on the 2D detection trained for target datasets are reported. Our method still outperforms MegaPose for the same detection results. It is noticeable that our method provides competitive results to the state-of-the-art methods, though it is not the best, even without fine-tuning on the target datasets." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 634, + 400, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 634, + 400, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 634, + 400, + 647 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": "We conduct ablation studies to validate the efficacy of our method. The ablation results are reported for the RGB inputs we mainly focus on. All experiments were performed on 5 BOP datasets: LM-O, T-LESS, TUD-L, IC-BIN, and YCB-V. We used the 2D detection results of YOLOX [19," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10045" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 541, + 159 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 541, + 159 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 541, + 159 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 541, + 159 + ], + "type": "table", + "html": "
MethodTrainingAverage Recall (↑)
LM-OT-LESSTUD-LIC-BINYCB-VMEAN
1No shape-constraint + RANSAC-PnP [39]Lflow61.777.672.454.265.766.3
2Shape-constraintLflow, Lpose64.981.875.157.070.169.8
3Shape-constraint + CascadeLflow, Lpose65.782.375.058.969.970.4
4Shape-constraint + Confidence factorizationLflow, Lcert, Lpose64.581.875.357.370.569.9
5Shape-constraint + Confidence factorization + CascadeLflow, Lcert, Lpose65.982.076.159.569.870.7
", + "image_path": "a24a26797b84c15bcf8cc0e790a057de4a8b8059235839fac8e418bf973863eb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 66, + 166, + 525, + 178 + ], + "lines": [ + { + "bbox": [ + 66, + 166, + 525, + 178 + ], + "spans": [ + { + "bbox": [ + 66, + 166, + 525, + 178 + ], + "type": "text", + "content": "Table 3. Ablation study of GenFlow design. The best result is in bold. Our method accomplishes the best performance (row 5)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 198, + 287, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 198, + 287, + 236 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 287, + 236 + ], + "type": "text", + "content": "44] trained on target datasets to focus on the accuracy of 6D pose estimation by preventing the performance bottleneck caused by the detection methods for novel objects." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 266, + 181, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 266, + 181, + 277 + ], + "spans": [ + { + "bbox": [ + 47, + 266, + 181, + 277 + ], + "type": "text", + "content": "4.3.1 Coarse Pose Estimation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 290, + 288, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 290, + 288, + 398 + ], + "spans": [ + { + "bbox": [ + 46, + 290, + 288, + 398 + ], + "type": "text", + "content": "Table 2 shows the effectiveness of the GMM-based sampling method for generating pose hypotheses. Pre-defined rotations are generated following [68] for the Naive sampling and for creating GMM of our method. We used the " + }, + { + "bbox": [ + 46, + 290, + 288, + 398 + ], + "type": "inline_equation", + "content": "N = 576" + }, + { + "bbox": [ + 46, + 290, + 288, + 398 + ], + "type": "text", + "content": " rotations for the Naive sampling, and " + }, + { + "bbox": [ + 46, + 290, + 288, + 398 + ], + "type": "inline_equation", + "content": "M = 72" + }, + { + "bbox": [ + 46, + 290, + 288, + 398 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 290, + 288, + 398 + ], + "type": "inline_equation", + "content": "M = 104" + }, + { + "bbox": [ + 46, + 290, + 288, + 398 + ], + "type": "text", + "content": " which conduct " + }, + { + "bbox": [ + 46, + 290, + 288, + 398 + ], + "type": "inline_equation", + "content": "2M" + }, + { + "bbox": [ + 46, + 290, + 288, + 398 + ], + "type": "text", + "content": " renderings (144 and 208, respectively) for ours. The results show that our method can generate a better initial pose despite the fewer renderings and network inferences." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 429, + 149, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 429, + 149, + 441 + ], + "spans": [ + { + "bbox": [ + 47, + 429, + 149, + 441 + ], + "type": "text", + "content": "4.3.2 GenFlow design" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 453, + 287, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 453, + 287, + 537 + ], + "spans": [ + { + "bbox": [ + 46, + 453, + 287, + 537 + ], + "type": "text", + "content": "Table 3 shows the results on different refiner designs. All experiments are conducted using the same result of coarse estimation, which uses our GMM-based sampling with " + }, + { + "bbox": [ + 46, + 453, + 287, + 537 + ], + "type": "inline_equation", + "content": "M = 104" + }, + { + "bbox": [ + 46, + 453, + 287, + 537 + ], + "type": "text", + "content": " without the multi-hypotheses strategy. We run 5 iterations of GenFlow refinement, each containing the 8 iterations of the inner update. We evaluated the final 6D pose from the GenFlow refiner." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 540, + 287, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 287, + 612 + ], + "type": "text", + "content": "Shape-Constraint Learning to estimate 2D correspondences with flow loss only is suboptimal for 6D pose estimation since it makes no use of the target shape (row 1). Imposing shape-constraint on the model benefits from the guide of 3D information without loss of generalization performance (row 2-5)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 614, + 287, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 614, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 614, + 287, + 651 + ], + "type": "text", + "content": "Cascade Architecture The proposed cascade architecture with multiple GenFlow modules contributes to the performance improvement (row 3, 5)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "Confidence Factorization Factorizing the confidence weights to certainty and pose sensitivity helps to increase the accuracy of the output pose. Certainty estimation leverages the training depth data and makes our method robust to occlusion (row 4, 5)." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 310, + 200, + 542, + 326 + ], + "blocks": [ + { + "bbox": [ + 310, + 200, + 542, + 326 + ], + "lines": [ + { + "bbox": [ + 310, + 200, + 542, + 326 + ], + "spans": [ + { + "bbox": [ + 310, + 200, + 542, + 326 + ], + "type": "image", + "image_path": "0da2290d5dcb4220551cdb25a3c6c9e59c5c1292f6be71943e10efbf3983b475.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 338, + 535, + 350 + ], + "lines": [ + { + "bbox": [ + 315, + 338, + 535, + 350 + ], + "spans": [ + { + "bbox": [ + 315, + 338, + 535, + 350 + ], + "type": "text", + "content": "Figure 6. GenFlow vs MegaPose refiner for the RGB inputs." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 371, + 452, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 371, + 452, + 384 + ], + "spans": [ + { + "bbox": [ + 306, + 371, + 452, + 384 + ], + "type": "text", + "content": "4.3.3 MegaPose Refiner vs Ours" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 391, + 545, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 391, + 545, + 571 + ], + "spans": [ + { + "bbox": [ + 304, + 391, + 545, + 571 + ], + "type": "text", + "content": "We compare the performance of the MegaPose refiner and ours for the RGB input. The same pose initializations from our coarse estimation were used for a fair evaluation. We reported the changes in AR score according to the iteration of the outer updates for both methods in Fig 6. The reported score is computed as the arithmetic mean of the AR scores for the five datasets. A single outer update takes 66.5 milliseconds for the MegaPose refiner and 98.2 milliseconds for ours, respectively, on an RTX 3090 GPU and Core i9-10920X CPU. Although ours is slower than the MegaPose refiner for a single refinement step, it achieves better performance for the same and even fewer iterations. Note that ours performs effectively by comparing a single-view rendered image for an outer update, while MegaPose refiner uses multiple rendered views as input." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 583, + 379, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 583, + 379, + 596 + ], + "spans": [ + { + "bbox": [ + 306, + 583, + 379, + 596 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 602, + 545, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 602, + 545, + 703 + ], + "spans": [ + { + "bbox": [ + 304, + 602, + 545, + 703 + ], + "type": "text", + "content": "In this paper, we present GenFlow, an optical flow-based method for 6D pose refinement of novel objects. First, we generate pose hypotheses using a GMM-based sampling strategy for efficient and effective coarse estimation. Then, we refine the high-scoring hypotheses with our pose refiner, our main contribution. GenFlow iteratively refines the flow and confidence with the guidance of 3D shape. Our method achieves state-of-the-art performance of 6D pose estimation for novel objects concerning RGB and RGB-D inputs." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10046" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "type": "text", + "content": "[1] Christopher M. Bishop. Pattern recognition and machine learning, 5th Edition. Springer, 2007. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 114, + 286, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 286, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 286, + 168 + ], + "type": "text", + "content": "[2] Michael J. Black and P. Anandan. A framework for the robust estimation of optical flow. In Fourth International Conference on Computer Vision, ICCV 1993, Berlin, Germany, 11-14 May, 1993, Proceedings, pages 231-236. IEEE Computer Society, 1993. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 170, + 286, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 170, + 286, + 224 + ], + "spans": [ + { + "bbox": [ + 53, + 170, + 286, + 224 + ], + "type": "text", + "content": "[3] Angel X. Chang, Thomas A. Funkhouser, Leonidas J. Guibas, Pat Hanrahan, Qi-Xing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. Shapenet: An information-rich 3d model repository. CoRR, abs/1512.03012, 2015. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 225, + 286, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 225, + 286, + 289 + ], + "spans": [ + { + "bbox": [ + 53, + 225, + 286, + 289 + ], + "type": "text", + "content": "[4] Hansheng Chen, Pichao Wang, Fan Wang, Wei Tian, Lu Xiong, and Hao Li. Epro-pnp: Generalized end-to-end probabilistic perspective-n-points for monocular object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2781-2790, 2022. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 292, + 286, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 292, + 286, + 324 + ], + "spans": [ + { + "bbox": [ + 53, + 292, + 286, + 324 + ], + "type": "text", + "content": "[5] Jianqiu Chen, Mingshan Sun, Tianpeng Bao, Rui Zhao, Liwei Wu, and Zhenyu He. Zeropose: Cad-model-based zero-shot pose estimation, 2023. 1, 2, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 325, + 286, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 325, + 286, + 380 + ], + "spans": [ + { + "bbox": [ + 53, + 325, + 286, + 380 + ], + "type": "text", + "content": "[6] Qifeng Chen and Vladlen Koltun. Full flow: Optical flow estimation by global optimization over regular grids. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 4706-4714. IEEE Computer Society, 2016. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 381, + 286, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 381, + 286, + 414 + ], + "spans": [ + { + "bbox": [ + 53, + 381, + 286, + 414 + ], + "type": "text", + "content": "[7] Xu Chen, Zijian Dong, Jie Song, Andreas Geiger, and Otmar Hilliges. Category level object pose estimation via neural analysis-by-synthesis. 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 415, + 286, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 415, + 286, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 415, + 286, + 468 + ], + "type": "text", + "content": "[8] Yilun Chen, Zhicheng Wang, Yuxiang Peng, Zhiqiang Zhang, Gang Yu, and Jian Sun. Cascaded pyramid network for multi-person pose estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 470, + 286, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 470, + 286, + 546 + ], + "spans": [ + { + "bbox": [ + 53, + 470, + 286, + 546 + ], + "type": "text", + "content": "[9] Kyunghyun Cho, Bart van Merrienboer, Dzmitry Bahdanau, and Yoshua Bengio. On the properties of neural machine translation: Encoder-decoder approaches. In Proceedings of SSST@EMNLP 2014, Eighth Workshop on Syntax, Semantics and Structure in Statistical Translation, Doha, Qatar, 25 October 2014, pages 103-111. Association for Computational Linguistics, 2014. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 548, + 286, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 286, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 286, + 613 + ], + "type": "text", + "content": "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2009), 20-25 June 2009, Miami, Florida, USA, pages 248-255. IEEE Computer Society, 2009. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 614, + 286, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 286, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 286, + 668 + ], + "type": "text", + "content": "[11] Maximilian Denninger, Dominik Winkelbauer, Martin Sundermeyer, Wout Boerdijk, Markus Knauer, Klaus H. Strobl, Matthias Humt, and Rudolph Triebel. Blenderproc2: A procedural pipeline for photorealistic rendering. Journal of Open Source Software, 8(82):4901, 2023. 5, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 286, + 713 + ], + "type": "text", + "content": "[12] Alexey Dosovitskiy, Philipp Fischer, Eddy Ilg, Philip Hausser, Caner Hazirbas, Vladimir Golkov, Patrick van der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. In Proceedings" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "of the IEEE International Conference on Computer Vision (ICCV), 2015. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 96, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 161 + ], + "type": "text", + "content": "[13] Andreas Doumanoglou, Rigas Kouskouridas, Sotiris Malassiotis, and Tae-Kyun Kim. Recovering 6d object pose and predicting next-best-view in the crowd. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 3583-3592. IEEE Computer Society, 2016. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 162, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 237 + ], + "type": "text", + "content": "[14] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas Barlow McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. In 2022 International Conference on Robotics and Automation, ICRA 2022, Philadelphia, PA, USA, May 23-27, 2022, pages 2553-2560. IEEE, 2022. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 239, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 293 + ], + "type": "text", + "content": "[15] Bertram Drost, Markus Ulrich, Nassir Navab, and Slobodan Ilic. Model globally, match locally: Efficient and robust 3d object recognition. In 2010 IEEE computer society conference on computer vision and pattern recognition, pages 998-1005. IEEE, 2010. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 294, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 545, + 360 + ], + "type": "text", + "content": "[16] Bertram Drost, Markus Ulrich, Paul Bergmann, Philipp Hartinger, and Carsten Steger. Introducing mvtec ITODD - A dataset for 3d object recognition in industry. In 2017 IEEE International Conference on Computer Vision Workshops, ICCV Workshops 2017, Venice, Italy, October 22-29, 2017, pages 2200-2208. IEEE Computer Society, 2017. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 361, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 361, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 361, + 545, + 415 + ], + "type": "text", + "content": "[17] Johan Edstedt, Ioannis Athanasiadis, Mårten Wadenbäck, and Michael Felsberg. Dkm: Dense kernelized feature matching for geometry estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17765-17775, 2023. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 415, + 545, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 415, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 308, + 415, + 545, + 458 + ], + "type": "text", + "content": "[18] Martin A. Fischler and Robert C. Bolles. Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM, 24(6):381-395, 1981. 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 460, + 545, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 492 + ], + "type": "text", + "content": "[19] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. YOLOX: exceeding YOLO series in 2021. CoRR, abs/2107.08430, 2021. 3, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 493, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 493, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 308, + 493, + 545, + 547 + ], + "type": "text", + "content": "[20] Yang Hai, Rui Song, Jiaojiao Li, and Yinlin Hu. Shape-constraint recurrent flow for 6d object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4831-4840, 2023. 1, 2, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 548, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 613 + ], + "type": "text", + "content": "[21] Rasmus Laurvig Haugaard and Anders Glent Buch. Surfemb: Dense and continuous correspondence distributions for object pose estimation with learnt surface embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6749-6758, 2022. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 614, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 668 + ], + "type": "text", + "content": "[22] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross B. Girshick. Mask R-CNN. In IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017, pages 2980-2988. IEEE Computer Society, 2017. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "type": "text", + "content": "[23] Xingyi He, Jiaming Sun, Yuang Wang, Di Huang, Hujun Bao, and Xiaowei Zhou. Onepose++: Keypoint-free one-shot object pose estimation without CAD models. In Advances in Neural Information Processing Systems, 2022. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10047" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "text", + "content": "[24] Yisheng He, Yao Wang, Haoqiang Fan, Jian Sun, and Qifeng Chen. Fs6d: Few-shot 6d pose estimation of novel objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6814-6824, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 288, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 288, + 215 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 288, + 215 + ], + "type": "text", + "content": "[25] Stefan Hinterstoisser, Vincent Lepetit, Slobodan Ilic, Stefan Holzer, Gary R. Bradski, Kurt Konolige, and Nassir Navab. Model based training, detection and pose estimation of texture-less 3d objects in heavily cluttered scenes. In Computer Vision - ACCV 2012 - 11th Asian Conference on Computer Vision, Daejeon, Korea, November 5-9, 2012, Revised Selected Papers, Part I, pages 548-562. Springer, 2012. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 217, + 288, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 288, + 292 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 288, + 292 + ], + "type": "text", + "content": "[26] Tomas Hodan, Pavel Haluza, Stepán Obdrzánek, Jiri Matas, Manolis I. A. Lourakis, and Xenophon Zabulis. T-LESS: an RGB-D dataset for 6d pose estimation of texture-less objects. In 2017 IEEE Winter Conference on Applications of Computer Vision, WACV 2017, Santa Rosa, CA, USA, March 24-31, 2017, pages 880-888. IEEE Computer Society, 2017. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 293, + 287, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 293, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 48, + 293, + 287, + 369 + ], + "type": "text", + "content": "[27] Tomás Hodań, Frank Michel, Eric Brachmann, Wadim Kehl, Anders Glent Buch, Dirk Kraft, Bertram Drost, Joel Vidal, Stephan Ihrke, Xenophon Zabulis, Caner Sahin, Fabian Manhardt, Federico Tombari, Tae-Kyun Kim, Jiri Matas, and Carsten Rother. BOP: Benchmark for 6D object pose estimation. European Conference on Computer Vision (ECCV), 2018. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 370, + 287, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 370, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 48, + 370, + 287, + 435 + ], + "type": "text", + "content": "[28] Tomás Hodan, Martin Sundermeyer, Bertram Drost, Yann Labbé, Eric Brachmann, Frank Michel, Carsten Rother, and Jiri Matas. BOP challenge 2020 on 6d object localization. In Computer Vision - ECCV 2020 Workshops - Glasgow, UK, August 23-28, 2020, Proceedings, Part II, pages 577-594. Springer, 2020. 2, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 437, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 287, + 459 + ], + "type": "text", + "content": "[29] Berthold K. P. Horn and Brian G. Schunck. Determining optical flow. Artif. Intell., 17(1-3):185-203, 1981. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 460, + 287, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 287, + 524 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 287, + 524 + ], + "type": "text", + "content": "[30] Yinlin Hu, Sébastien Speierer, Wenzel Jakob, Pascal Fua, and Mathieu Salzmann. Wide-depth-range 6d object pose estimation in space. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 15870-15879. Computer Vision Foundation / IEEE, 2021. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 525, + 287, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 557 + ], + "type": "text", + "content": "[31] Yinlin Hu, Pascal Fua, and Mathieu Salzmann. Perspective flow aggregation for data-limited 6d object pose estimation. In ECCV, 2022. 1, 2, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 559, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 612 + ], + "type": "text", + "content": "[32] Tak-Wai Hui, Xiaou Tang, and Chen Change Loy. Liteflownet: A lightweight convolutional neural network for optical flow estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 614, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 668 + ], + "type": "text", + "content": "[33] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[34] Wolfgang Kabsch. A solution for the best rotation to relate two sets of vectors. Acta Crystallographica Section A: Crystal Physics, Diffraction, Theoretical and General Crystallography, 32(5):922-923, 1976. 2, 6" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "type": "text", + "content": "[35] Roman Kaskman, Sergey Zakharov, Ivan Shugurov, and Slobodan Ilic. Homebrewdb: RGB-D dataset for 6d pose estimation of 3d objects. In 2019 IEEE/CVF International Conference on Computer Vision Workshops, ICCV Workshops 2019, Seoul, Korea (South), October 27-28, 2019, pages 2767-2776. IEEE, 2019. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 140, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 194 + ], + "type": "text", + "content": "[36] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 197, + 545, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 250 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 250 + ], + "type": "text", + "content": "[37] Yann Labbe, Lucas Manuelli, Arsalan Mousavian, Stephen Tyree, Stan Birchfield, Jonathan Tremblay, Justin Carpentier, Mathieu Aubry, Dieter Fox, and Josef Sivic. MegaPose: 6D Pose Estimation of Novel Objects via Render & Compare. In CoRL. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 252, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 252, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 307, + 252, + 545, + 296 + ], + "type": "text", + "content": "[38] Y. Labbe, J. Carpentier, M. Aubry, and J. Sivic. Cosyposse: Consistent multi-view multi-object 6d pose estimation. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2, 5, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 297, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 545, + 330 + ], + "type": "text", + "content": "[39] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Epnp: An accurate o(n) solution to the pnp problem. International Journal Of Computer Vision, 81:155-166, 2009. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 332, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 332, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 307, + 332, + 545, + 374 + ], + "type": "text", + "content": "[40] Fu Li, Ivan Shugurov, Benjamin Busam, Minglong Li, Shaowu Yang, and Slobodan Ilic. Polarmesh: A star-convex 3d shape approximation for object pose estimation. IEEE Robotics Autom. Lett., 7(2):4416-4423, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 376, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 409 + ], + "type": "text", + "content": "[41] Yi Li, Gu Wang, Xiangyang Ji, Yu Xiang, and Dieter Fox. Deepim: Deep iterative matching for 6d pose estimation. In European Conference on Computer Vision (ECCV), 2018. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 411, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 464 + ], + "type": "text", + "content": "[42] Tsung-Yi Lin, Piotr Dollar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 468, + 545, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 468, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 307, + 468, + 545, + 520 + ], + "type": "text", + "content": "[43] Lahav Lipson, Zachary Teed, Ankit Goyal, and Jia Deng. Coupled iterative refinement for 6d multi-object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6728-6737, 2022. 2, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 523, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 545, + 578 + ], + "type": "text", + "content": "[44] Xingyu Liu, Ruida Zhang, Chenyangguang Zhang, Bowen Fu, Jiwen Tang, Xiquan Liang, Jingyi Tang, Xiaotian Cheng, Yukang Zhang, Gu Wang, and Xiangyang Ji. Gdrnpp. https://github.com/shanice-1/gdrnpp_bop2022, 2022.3,7,8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 579, + 545, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 545, + 621 + ], + "type": "text", + "content": "[45] Yuan Liu, Yilin Wen, Sida Peng, Cheng Lin, Xiaoxiao Long, Taku Komura, and Wenping Wang. Gen6d: Generalizable model-free 6-dof object pose estimation from rgb images. In ECCV, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 624, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 677 + ], + "type": "text", + "content": "[46] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11976-11986, 2022. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "text", + "content": "[47] David G. Lowe. Object recognition from local scale-invariant features. In Proceedings of the International Conference on Computer Vision, Kerkyra, Corfu, Greece," + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10048" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "text", + "content": "September 20-25, 1999, pages 1150-1157. IEEE Computer Society, 1999. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "type": "text", + "content": "[48] Fabian Manhardt, Gu Wang, Benjamin Busam, Manuel Nickel, Sven Meier, Luca Minciullo, Xiangyang Ji, and Nassir Navab. Cps++: Improving class-level 6d pose and shape estimation from monocular images with self-supervised learning, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 151, + 286, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 151, + 286, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 151, + 286, + 216 + ], + "type": "text", + "content": "[49] Van Nguyen Nguyen, Yinlin Hu, Yang Xiao, Mathieu Salzmann, and Vincent Lepetit. Templates for 3d object pose estimation revisited: Generalization to new objects and robustness to occlusions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6771-6780, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 217, + 286, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 286, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 286, + 259 + ], + "type": "text", + "content": "[50] Van Nguyen Nguyen, Thibault Groueix, Georgy Ponimatkin, Vincent Lepetit, and Tomas Hodan. Cnos: A strong baseline for cad-based novel object segmentation. arXiv preprint arXiv:2307.11067, 2023. 2, 3, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 261, + 286, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 286, + 315 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 286, + 315 + ], + "type": "text", + "content": "[51] Brian Okorn, Qiao Gu, Martial Hebert, and David Held. Zephyr: Zero-shot pose hypothesis rating. In IEEE International Conference on Robotics and Automation, ICRA 2021, Xi'an, China, May 30 - June 5, 2021, pages 14141-14148. IEEE, 2021. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 316, + 286, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 286, + 414 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 286, + 414 + ], + "type": "text", + "content": "[52] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 415, + 286, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 286, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 286, + 470 + ], + "type": "text", + "content": "[53] Keunhong Park, Arsalan Mousavian, Yu Xiang, and Dieter Fox. Latentfusion: End-to-end differentiable reconstruction and rendering for unseen object pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 471, + 286, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 286, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 286, + 525 + ], + "type": "text", + "content": "[54] Ivan Shugurov, Fu Li, Benjamin Busam, and Slobodan Ilic. Osop: A multi-stage one shot object pose estimation framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6835-6844, 2022. 1, 2, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 525, + 286, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 286, + 579 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 286, + 579 + ], + "type": "text", + "content": "[55] Andrea Simonelli, Samuel Rota Bulo, Lorenzo Porzi, Manuel Lopez-Antequera, and Peter Kontschieder. Disentangling monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 581, + 286, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 286, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 286, + 646 + ], + "type": "text", + "content": "[56] Yongzhi Su, Mahdi Saleh, Torben Fetzer, Jason Rambach, Nassir Navab, Benjamin Busam, Didier Stricker, and Federico Tombari. Zebrapose: Coarse to fine surface encoding for 6 dof object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6738-6748, 2022. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 647, + 286, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 286, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 286, + 690 + ], + "type": "text", + "content": "[57] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 692, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 286, + 713 + ], + "type": "text", + "content": "[58] Jiaming Sun, Zihao Wang, Siyu Zhang, Xingyi He, Hongcheng Zhao, Guofeng Zhang, and Xiaowei Zhou." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 665 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "type": "text", + "content": "Onepose: One-shot object pose estimation without cad models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6825-6834, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 118, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 545, + 173 + ], + "type": "text", + "content": "[59] Martin Sundermeyer, Maximilian Durner, En Yen Puang, Zoltan-Csaba Marton, Narunas Vaskevicius, Kai O. Arras, and Rudolph Triebel. Multi-path learning for object pose estimation across domains. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 175, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 175, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 308, + 175, + 545, + 228 + ], + "type": "text", + "content": "[60] Mingxing Tan and Quoc V. Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In Proceedings of the 36th International Conference on Machine Learning, ICML 2019, 9-15 June 2019, Long Beach, California, USA, pages 6105-6114. PMLR, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 230, + 545, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 230, + 545, + 283 + ], + "spans": [ + { + "bbox": [ + 308, + 230, + 545, + 283 + ], + "type": "text", + "content": "[61] Zachary Teed and Jia Deng. RAFT: recurrent all-pairs field transforms for optical flow. In Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II, pages 402-419. Springer, 2020. 2, 4, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 285, + 545, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 285, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 308, + 285, + 545, + 340 + ], + "type": "text", + "content": "[62] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J. Guibas. Normalized object coordinate space for category-level 6d object pose and size estimation. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 342, + 545, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 342, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 308, + 342, + 545, + 373 + ], + "type": "text", + "content": "[63] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes. 2018. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 375, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 375, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 308, + 375, + 545, + 430 + ], + "type": "text", + "content": "[64] Yang Xiao, Xuchong Qiu, Pierre-Alain Langlois, Mathieu Aubry, and Renaud Marlet. Pose from shape: Deep pose estimation for arbitrary 3d objects. In 30th British Machine Vision Conference 2019, BMVC 2019, Cardiff, UK, September 9-12, 2019, page 61. BMVA Press, 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 431, + 545, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 431, + 545, + 474 + ], + "spans": [ + { + "bbox": [ + 308, + 431, + 545, + 474 + ], + "type": "text", + "content": "[65] Yang Xiao, Yuming Du, and Renaud Marlet. Posecontrast: Class-agnostic object viewpoint estimation in the wild with pose-aware contrastive learning. In International Conference on 3D Vision (3DV), 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 475, + 545, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 475, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 308, + 475, + 545, + 540 + ], + "type": "text", + "content": "[66] Gengshan Yang and Deva Ramanan. Volumetric correspondence networks for optical flow. In Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada, pages 793-803, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 543, + 545, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 543, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 308, + 543, + 545, + 574 + ], + "type": "text", + "content": "[67] Heng Yang, Jingnan Shi, and Luca Carlone. TEASER: fast and certifiable point cloud registration. IEEE Trans. Robotics, 37(2):314-333, 2021. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 576, + 545, + 619 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 576, + 545, + 619 + ], + "spans": [ + { + "bbox": [ + 308, + 576, + 545, + 619 + ], + "type": "text", + "content": "[68] Anna Yershova, Swati Jain, Steven M. LaValle, and Julie C. Mitchell. Generating uniform incremental grids on " + }, + { + "bbox": [ + 308, + 576, + 545, + 619 + ], + "type": "inline_equation", + "content": "SO(3)" + }, + { + "bbox": [ + 308, + 576, + 545, + 619 + ], + "type": "text", + "content": " using the hopf fibration. Int. J. Robotics Res., 29(7):801-812, 2010. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 621, + 545, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 621, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 308, + 621, + 545, + 665 + ], + "type": "text", + "content": "[69] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "10049" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/GenH2R_ Learning Generalizable Human-to-Robot Handover via Scalable Simulation Demonstration and Imitation/40314ff6-3af8-45b8-b8fc-6f9ebdc7d993_content_list.json b/2024/GenH2R_ Learning Generalizable Human-to-Robot Handover via Scalable Simulation Demonstration and Imitation/40314ff6-3af8-45b8-b8fc-6f9ebdc7d993_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..93eca34e4b392f3ca3bda145ea566628dde23eb2 --- /dev/null +++ b/2024/GenH2R_ Learning Generalizable Human-to-Robot Handover via Scalable Simulation Demonstration and Imitation/40314ff6-3af8-45b8-b8fc-6f9ebdc7d993_content_list.json @@ -0,0 +1,1343 @@ +[ + { + "type": "text", + "text": "GenH2R: Learning Generalizable Human-to-Robot Handover via Scalable Simulation, Demonstration, and Imitation", + "text_level": 1, + "bbox": [ + 171, + 130, + 799, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zifan Wang $^{1,3}$ Junyu Chen $^{1,3}$ Ziqing Chen $^{1}$ Pengwei Xie $^{1}$ Rui Chen $^{1}$ Li Yi $^{\\dagger 1,2,3}$", + "bbox": [ + 127, + 202, + 833, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Tsinghua University $^{2}$ Shanghai Artificial Intelligence Laboratory $^{3}$ Shanghai Qi Zhi Institute https://GenH2R.github.io", + "bbox": [ + 109, + 223, + 859, + 260 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 295, + 313, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper presents GenH2R, a framework for learning generalizable vision-based human-to-robot (H2R) handover skills. The goal is to equip robots with the ability to reliably receive objects with unseen geometry handed over by humans in various complex trajectories. We acquire such generalizability by learning H2R handover at scale with a comprehensive solution including procedural simulation assets creation, automated demonstration generation, and effective imitation learning. We leverage large-scale 3D model repositories, dexterous grasp generation methods, and curve-based 3D animation to create an H2R handover simulation environment named GenH2R-Sim, surpassing the number of scenes in existing simulators by three orders of magnitude. We further introduce a distillation-friendly demonstration generation method that automatically generates a million high-quality demonstrations suitable for learning. Finally, we present a 4D imitation learning method augmented by a future forecasting objective to distill demonstrations into a visuo-motor handover policy. Experimental evaluations in both simulators and the real world demonstrate significant improvements (at least +10% success rate) over baselines in all cases.", + "bbox": [ + 75, + 327, + 473, + 659 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 688, + 209, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The embodied AI research community has long been driven by the goal of empowering robots to interact and collaborate with humans. A crucial aspect of this pursuit is equipping robots with the capability to reliably receive arbitrarily moving objects of varying geometry handed over by humans, based on dynamic visual observations. This human-to-robot (H2R) handover ability allows robots to seamlessly collaborate with humans across a wide range of tasks, including cooking, room tidying, and furniture assembly.", + "bbox": [ + 75, + 713, + 468, + 849 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, compared to learning human-free robot manip-", + "bbox": [ + 96, + 849, + 467, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c7a626db76ce77f60bec918fae38f7655a0f03c08f3c374596a4a8ea69b1bb4a.jpg", + "image_caption": [ + "Scalable Synthetic Handover Simulator" + ], + "image_footnote": [], + "bbox": [ + 514, + 294, + 687, + 398 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b65b3f3cc6277337bf282f44a073d65fbd1a0a2161d60153dc7902c157040165.jpg", + "image_caption": [ + "Distillation-friendly Expert Demonstrations" + ], + "image_footnote": [], + "bbox": [ + 697, + 294, + 872, + 398 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8365645783823546867e57ab810053920af0517be16c1c1c7e03db65bf22b0c2.jpg", + "image_caption": [ + "Sim-to-real Deployment" + ], + "image_footnote": [], + "bbox": [ + 514, + 411, + 689, + 505 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a65c3f3399fb73c24d783dffda82d53ffa58cbb47c7748061f6c5afbaefc24bf.jpg", + "image_caption": [ + "Forecast-Aided 4D Imitation Learning", + "Figure 1. The overview of GenH2R. We introduce a framework for learning generalizable vision-based human-to-robot handover via scalable synthetic simulation, distillation-friendly expert demonstration generation, and a forecast-aided 4D imitation learning method. Our models demonstrate strong generalization capabilities to real datasets and can be deployed to a real robot." + ], + "image_footnote": [], + "bbox": [ + 697, + 410, + 870, + 503 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ulation skills, the progress in scalably learning H2R handover that can generalize to various objects and versatile human behaviors has lagged due to its unique challenges. Training robots to interact with humans in real-world scenarios entails increased risks and expenses, rendering it inherently non-scalable. Therefore, it is demanded to simulate human behaviors and train robots in simulated environments prior to real-world deployment. However, creating a substantial number of assets for humans handing over objects poses a significant challenge. In a recent study [9] that employed motion capturing to drive virtual humans in a simulator, only 1000 unique human hand motion trajectories were provided for handing over 20 objects. Limited object geometry and human motion assets can hardly capture the complexities of the real world. Besides, the challenge extends to the demonstration side. The success of large language model [6, 32, 52] has suggested a recipe for scaling up learning through modeling large-scale training", + "bbox": [ + 496, + 628, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution with the order determined by rolling dice.", + "bbox": [ + 94, + 875, + 423, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 96, + 887, + 222, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16362", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "data. Nevertheless, collecting robot demonstrations receiving objects from real humans is very costly and unscalable. How to scale up the number of demonstrations while ensuring effective learning poses additional challenges.", + "bbox": [ + 75, + 90, + 467, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we aim to learn generalizable H2R handover at scale by tackling the above challenges. We present a comprehensive solution that scales up both the assets and demonstrations and effectively learns a closed-loop visuomotor policy through a novel imitation learning algorithm.", + "bbox": [ + 75, + 154, + 467, + 229 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, to scale up geometry and motion assets depicting humans handing over various objects, we leverage large-scale 3D model repositories [7, 16], dexterous grasp generation methods [46], and curve-based 3D animation. This enables us to procedurally generate millions of handover scenes, forming an environment named GenH2R-Sim to support generalizable H2R handover learning. GenH2R-Sim surpasses HandoverSim [9], an existing H2R simulator, in both scene quantity (by three orders of magnitude) and unique object involvement (by two orders of magnitude). In addition, scenes in GenH2R-Sim go beyond a straightforward giving and then receiving and cover cases when humans might keep transforming the object in a large range during the entire H2R handover process. This allows for studying complex behaviors such as humans hesitating before handing over.", + "bbox": [ + 75, + 231, + 467, + 473 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To scale up robot demonstrations, we draw inspiration from the Task and Motion Planning (TAMP) [22] literature and propose to automatically generate demonstrations with grasp and motion planning using privileged human motion and object state information. There are some straightforward ways to achieve this goal, such as using the privileged human handover destination information to plan a smooth demonstration. However, the problem is more challenging than it seems since the generated demonstrations need to be suitable for distilling into a visuomotor policy. We identify the vision-action correlation between visual observations and planned actions as the crucial factor influencing distillability and point out that due to the constraints of robot arm morphology one can easily generate observation-irrelevant actions and thus harm distillation. To tackle this challenge, we present a distillation-friendly demonstration generation method that sparsely samples handover animations for landmark states and periodically replans grasp and motion based on privileged future landmarks.", + "bbox": [ + 75, + 474, + 467, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, to distill the above demonstrations into a visuomotor policy, we utilize point cloud input for its richer geometric information and smaller sim-vs-real gap compared to images. We propose a 4D imitation learning method that factors the sequential point cloud observations into geometry and motion parts, facilitating policy learning by better revealing the current scene state. Furthermore, the imitation objective is augmented by a forecasting objective which predicts the future motion of the handover object. Since our", + "bbox": [ + 75, + 763, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "demonstrating actions are generated based on future landmarks, the forecasting objective can help further exploit the vision-action correlation.", + "bbox": [ + 498, + 90, + 890, + 135 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate our learned policy in simulators (Handover-Sim and our own GenH2R-Sim) and the real world. Remarkably, without any mocap assets or real-world demonstrations, our method achieves significantly better performance compared to baselines across all settings (at least $+10\\%$ success rate). Our experiments highlight that the scaling-up efforts bring substantial improvement in policy generalizability to novel geometry and complex motion. Furthermore, these efforts greatly facilitate skill transfer to real robotic systems.", + "bbox": [ + 496, + 136, + 890, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, the key contribution of this paper is a novel framework scaling up the learning of H2R handover with the following three components: i) a simulation environment named GenH2R-Sim consists of millions of human handover animations for generalizable H2R handover learning, ii) an empirically validated automatic robot demonstration generation pipeline for vision-based closed-loop control, iii) a forecast-aided 4D imitation learning method effective in distilling the large-scale demonstrations.", + "bbox": [ + 496, + 287, + 890, + 424 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 469, + 640, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Human-to-Robot Handovers", + "text_level": 1, + "bbox": [ + 500, + 494, + 756, + 508 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, significant progress in human-robot handovers [12, 33, 36] has been observed, driven by the increasing popularity of human-robot interaction [1, 38] and the emergence of extensive datasets [5, 8, 18, 25, 28, 50] capturing hand-object interactions. Some traditional methods [2, 4] require 3D object models and struggle to handle unseen objects. One possible way is to consider grasping and dynamic motion planning [19, 30, 49, 51]. However, these methods often exhibit constrained motions and perform poorly on large-scale datasets. HandoverSim [9], a physics-simulated environment, introduced a new simulation benchmark for human-to-robot object handovers. Leveraging DexYCB [8], a dataset of human grasping objects and performing handover attempts, this environment allows training learning-based handover policies such as [11]. However, it lacks large-scale and diverse handover scenes, which limits generalizable handovers. At the same time, SynH2R [10] proposes to use synthetic data but makes limited progress. Building on this, we propose GenH2R-Sim , aiming to benchmark generalizable handover.", + "bbox": [ + 496, + 518, + 890, + 820 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Scaling Up Robot Demonstrations", + "text_level": 1, + "bbox": [ + 500, + 830, + 797, + 847 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For robot learning, scaling up data collection for manipulation skills has spurred extensive research. Approaches include leveraging large language models [24] or hardware", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16363", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8d67f0d7bca143f2d5d11b877b11361c982ad3fea0621b6abb9be4e22168ebfd.jpg", + "image_caption": [ + "Figure 2. The overview of our framework. First, we propose a new simulation environment named GenH2R-Sim, featuring large-scale synthetic datasets with diversity in object geometry, grasp poses, and complex trajectories. Second, other than destination planning (move straight toward the final position) and dense planning (replan at each step), we propose a distillation-friendly demonstration generation method—landmark planning, predicting landmarks on the trajectory (as indicated by the dashed object above) and replanning based on those landmarks. Thirdly, our Forecast-aided 4D Imitation Learning leverages past flow information, and the forecasting objective enhances the exploitation of vision-action correlation." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 441, + 330 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/19062cc6078a2249cc73be71dfa6bb7489a75942007b3e060b4eeed8dfd8e240.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 442, + 88, + 651, + 330 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/05fec30c5092a9c1b790bc8952ca4c628d26fc02d4f4779e92ec95fdb4f60b01.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 88, + 870, + 330 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "capabilities [39], utilizing non-robotics datasets [23], and employing trial-and-error explorations [21]. As depicted in [24], one of the challenges is scaling up robot-complete data. A popular line of research scales up demonstration generation via Task and Motion Planning [13, 22, 31]. These works usually focus on fairly static scenes without active motion or object and task variety [44, 47] while our method extends to dynamic H2R handover by considering how to interpret human behavior and generate demonstrations easy to be distilled by closed-loop visuo-motor policy.", + "bbox": [ + 75, + 436, + 468, + 589 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Offline Learning from Demonstrations", + "text_level": 1, + "bbox": [ + 76, + 599, + 410, + 616 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Imitation Learning (IL) represents a methodology for training embodied agents in manipulation tasks by utilizing expert demonstrations. The commonly used Behavior Cloning (BC) [34] strategy directly trains the policy to imitate expert actions in a supervised learning manner. Despite its simplicity, this approach has demonstrated remarkable effectiveness in robotic manipulation [3, 20, 29, 53] especially when combined with a substantial number of high-quality demonstrations [15, 26]. Inspired by these works, we adopt an imitation learning paradigm, focusing on how to leverage spatial-temporal perception and future forecasting to better consume our distillation-friendly demonstrations.", + "bbox": [ + 75, + 623, + 468, + 805 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 820, + 166, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 76, + 845, + 186, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For the generalizable H2R handover task, we introduce GenH2R, a framework designed to learn control policies,", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "specifically 6D control actions for the robot gripper, using segmented point cloud data captured from an egocentric camera. We describe our method for synthesizing human handover animations in Section 3.2, generating expert demonstrations in Section 3.3, and distilling demonstrations to 4D vision-based neural networks by imitation learning in Section 3.4, as the pipeline depicted in Figure 2.", + "bbox": [ + 496, + 436, + 890, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. GenH2R-Sim", + "text_level": 1, + "bbox": [ + 500, + 556, + 638, + 570 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The size and quality of human-object datasets in simulators play a crucial role in generating high-quality handover demonstrations and training reliable policies for handover scenarios. The recent handover simulator, HandoverSim [9], utilizes the DexYCB [8] dataset, which captures real-world human grasping objects in a limited manner, comprising only 1000 scenes with 20 distinct objects. In the real world, scenarios can be more complex and may involve intricate trajectories and poses beyond those in DexYCB.", + "bbox": [ + 496, + 580, + 890, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address these limitations, we introduce a new environment, GenH2R-Sim, to overcome these deficiencies and facilitate generalizable handovers. To diversify geometry and motion assets depicting humans handing over various objects, we focus on two primary aspects: the hand grasping pose and the hand-object moving trajectory within a scene.", + "bbox": [ + 496, + 718, + 890, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In aspects of grasping poses, DexGraspNet [45] has made significant contributions by employing optimization techniques to generate a substantial dataset of human hand grasp poses. We utilize this method to generate approximately 1,000,000 grasp poses for 3,266 different objects sourced from Shapenet [7]. These objects span a wide range", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16364", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of categories, from larger items like computers to smaller ones like mobile phones, covering most sizes and shapes encountered in real-life handovers.", + "bbox": [ + 75, + 90, + 468, + 135 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In aspects of hand-object moving trajectories, we propose to use Bézier curves, which are one class of smooth curves determined by several control points, to generate complex yet smooth-transiting motion trajectories. We use multiple Bézier curves to model different stages of the motion, and link the ends of these curves to create a seamless track. We can generate scenes matching various scenarios of different complexity in the real world by adjusting the distribution of control points of the trajectory and the speed of the human hand. To enhance the trajectory's realism, we incorporate consistent object rotations, which also enhances the importance of choosing the appropriate grasp for the robotic arm. Since we can always attach a new segment of motion at the end of the current motion and the duration is much longer than DexYCB scenes, the destination of the hand-object is not a significant factor, so we just randomly select a point within the reach of the robotic arm.", + "bbox": [ + 75, + 137, + 470, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We do not guarantee that every item in the dataset we generate perfectly mimics the human-like characteristics of real-world data, but our approach ensures a significantly higher degree of domain randomization and provides greater diversity in terms of geometry and motion. Given the challenges in scaling up real-world motion capture datasets, we opt for a large-scale synthetic dataset for our handover simulations. Our key insight is that for both demonstrations and policy learning, having a substantial amount of synthetic data is more beneficial than relying on a small-scale real-world dataset.", + "bbox": [ + 75, + 395, + 468, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "GenH2R-Sim follows the setup of HandoverSim, which consists of a Panda 7DoF robotic arm with a gripper and a wrist-mounted RGB-D camera, and a simulated human hand. Just like HandoverSim, we switch from the prehandover kinematic phase to the handover dynamic phase when the object has been in contact with the gripper. HandoverSim is not adaptive to the robot's action and just loads and replays every frame of the data. To align with the real-world handover process more naturally in GenH2R-Sim, the simulated hand will stop from moving and wait for handover when the robot arm is close to the object.", + "bbox": [ + 75, + 561, + 468, + 728 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Generating Demonstrations for Distillation", + "text_level": 1, + "bbox": [ + 76, + 739, + 442, + 755 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we address a key question in learning visuomotor policy: how to efficiently generate robot demonstrations that incorporate paired vision-action data from successful task experiences. While distilling successful demonstrations into a single policy has proven effective for open-loop control tasks, the challenge lies in closed-loop visuo-motor control, where the quality of demonstrations becomes crucial for learning. Merely ensuring success is no longer sufficient. We present two examples of demon-", + "bbox": [ + 75, + 763, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c9d34d63fb5e2bdd134e280734397e7af3be6a5d1c059e612b63a981738d71ff.jpg", + "image_caption": [ + "Figure 3. Different demonstration generation methods for dynamic handover. The orange curve shows the hand-object trajectory. The blue, red, and green curves show the example trajectories generated by the foresighted planner, the shortsighted planner, and our planner, respectively." + ], + "image_footnote": [], + "bbox": [ + 501, + 89, + 890, + 301 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "stration generation with different grasp and motion planning strategies as shown in Figure 3. In the first example, a foresighted planner generates smooth, short demonstrations based on the privileged destination end state of a human handover animation. Though efficient, the planned path does not align actions with the dynamic visual observations during the handover. Distilling such demonstrations requires accurately forecasting the end state of the human trajectory, which can be extremely challenging in complex handover cases. The second example involves a shortsighted planner that independently replans grasp and motion at each time step using privileged hand and object states. Due to robot morphology constraints and the multisolution nature of common robot planners, smooth visual observations may correspond to unsmooth and multi-modal robot trajectories, increasing the difficulty of distillation. We emphasize the importance of distillability as a quality factor for handover demonstrations. An effective demonstration generation method must consider the vision-action correlation by jointly incorporating robot morphology and dynamic vision during grasp and motion planning.", + "bbox": [ + 496, + 398, + 892, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Along this line, we base our method on the foresighted and shortsighted planner mentioned above to combine the advantages of both sides while encouraging the demonstration distillability. We first improve the shortsighted planner so that sequentially smooth visual observations result in smooth grasp and motion plans. Then we improve the handover efficiency by looking toward the future while guaranteeing the vision-action correlation.", + "bbox": [ + 496, + 717, + 890, + 837 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To be specific, we build our method based on the OMG planner [41] for grasp and motion planning. This planner optimizes the grasp and motion path by considering the object's 6D pose and a set of candidate grasp poses. To support", + "bbox": [ + 496, + 839, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16365", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "this optimization, we provide privileged knowledge that includes the object's 6D pose, candidate grasps generated through physics simulation [16], and human hand poses for filtering out invalid grasps. However, independently calling the OMG planner for each time step may result in unsmooth trajectories, as it is designed for static scenarios. To address this, we sequentially plan the grasp and motion based on the privileged knowledge by: 1) sorting grasps based on their pose distance to the robot end effector and attempting inverse kinematics (IK) starting from the nearest grasp until success; 2) initializing IK based on the robot arm pose from the previous time step; 3) invoking the OMG planner only when IK can be successfully solved. By prioritizing closer grasps, we encourage the object to remain within the field of view of a wrist camera, reducing visually irrelevant actions when the object is not visible. Additionally, enforcing IK smoothness improves the overall trajectory smoothness. As a result, the enhanced vision-action correlation dramatically improves the demonstration quality.", + "bbox": [ + 75, + 90, + 472, + 380 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our approach modifies the OMG planner for dynamic grasp and motion planning. However, densely replanning at each time step leads to inefficient and non-smooth zigzag demonstrations, which does not align with how humans receive objects. Humans anticipate dynamic scene changes before taking action. On the other hand, a highly fore-sighted planner that directly plans grasp and motion based on the end state of a human handover animation can disrupt the vision-action correlation. To strike a balance between these extremes, we propose an algorithm that sparsely samples handover animations for landmark states and periodically replans grasp and motion based on future landmarks. The key idea is to select landmarks strategically so that the planner only considers visually foreseeable futures. Specifically, let $\\xi = (\\mathcal{T}_0,\\mathcal{T}_1,\\dots ,\\mathcal{T}_{T - 1})$ represent an object trajectory, where $\\mathcal{T}_t\\in \\mathbb{SE}(3)$ denotes the object pose in the $t$ -th frame within the world coordinate system. Based on all the object trajectories in the training set, we train an object pose forecasting network which consumes past and current object poses $(\\mathcal{T}_0,\\mathcal{T}_1,\\dots ,\\mathcal{T}_t)$ for each time step $t$ within each trajectory and forecasts the object poses $(\\mathcal{T}_{t + 1},\\mathcal{T}_{t + 2},\\dots ,\\mathcal{T}_{t + N})$ in future $N$ steps. By thresholding the forecasting error corresponding to each time step, we identify a set of endpoints where past observations cannot forecast the future very well and partition the complete trajectory $\\xi$ to several segments using endpoints $0 = l_{0} < l_{1} < \\dots < l_{k} = T$ . Within each segment, we assume the ability to predict the future object pose based on historical information. We then denote $P\\in \\mathbb{N}$ as the hyperparameter determining the replanning period. For each planning frame $t = 0,P,2P,\\dots$ , suppose the next endpoint is $l_{i + 1}$ , i.e., $l_{i}\\leq t < l_{i + 1}$ . Then we will plan based on the object pose at frame $\\hat{t} = \\min (t + P,l_{i + 1})$ , which serves as a landmark. Note here planning is based on the future states", + "bbox": [ + 75, + 387, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "but avoids bypassing the sharply transitioning points where human motion becomes unpredictable. Also worth mentioning, densely planning is a special case of our method, and landmark planning is a full version.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Forecast-Aided 4D Imitation Learning", + "text_level": 1, + "bbox": [ + 498, + 162, + 831, + 180 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Traditional methods for human-to-robot handover face challenges in gaining insights into dynamic scene perception. Approaches based on motion planning [42] often emphasize robot morphology and lack dynamic vision perception. They struggle to capture long-horizon information, mainly focusing on the current frame and failing to predict the future. Reinforcement Learning methods [11, 43], while powerful, require extensive training and may train unstably across different scenarios. To enhance the vision-action correlation and establish an efficient training paradigm, we introduce our forecast-aided 4D imitation learning approach.", + "bbox": [ + 496, + 186, + 890, + 352 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In robot perception, the 4D point cloud serves as the common representation. In the $t$ -th frame, we can define $M_t^i \\in \\mathbb{SE}(3)$ as the relative object pose between the current frame and the $i$ -th frame in the egocentric view. While frame stacking is a straightforward approach, it struggles to capture both motion and geometry effectively. Inspired by recent 4D learning methods [14, 40], we employ the Iterative Closest Point (ICP) registration algorithm [37] to efficiently compute transformation matrices $\\{\\hat{M}_t^{t-1}, \\hat{M}_t^{t-2}, \\dots, \\hat{M}_t^{t-L_1}\\}$ between the point cloud in the $t$ -th frame and the point clouds in previous $L_1$ frames. Applying these transformation matrices to a specific point in the current frame yields its rough coordinates in previous frames. Then we incorporate this flow feature into 3D PointNet++ [35] to encode a global spatial-temporal feature and use Multilayer Perceptron (MLP) to decode it into a 6D egocentric action. The loss function, denoted as $\\mathcal{L}_{action}$ , is computed as the L1 loss for aligning 3D points on the robot gripper as defined in [27]. We believe some sophisticated 4D backbones [17, 48] are suitable for 4D understanding, but they are often not suitable for robotic tasks that require a fast reference speed. Our method strikes a balance between effectiveness and simplicity.", + "bbox": [ + 496, + 353, + 892, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To enhance the responsiveness of our policy to human motion and extend the vision horizon into the future, we introduce an auxiliary task to predict the future motion $\\{M_t^{t + 1}, M_t^{t + 2}, \\dots, M_t^{t + L_2}\\}$ of objects in the next $L_{2}$ frames. Using the ground truth object poses from trajectories, we compute the motion prediction loss for the $t$ -th frame:", + "bbox": [ + 496, + 700, + 890, + 806 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p r e d} = \\sum_ {i = t + 1} ^ {t + L _ {2}} \\| \\hat {M} _ {t} ^ {i} - M _ {t} ^ {i} \\| \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 820, + 890, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In contrast to reinforcement learning, our imitation learning method requires only a few hours of training and", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16366", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "achieves great generalizability through large-scale, high-quality demonstrations. We acquire vision-action pairs and ground truth object states from demonstrations, and then supervise our policy using the loss function $\\mathcal{L} = \\mathcal{L}_{action} + \\lambda \\mathcal{L}_{pred}$ , where $\\lambda$ serves as a weighting hyper-parameter to balance the losses. This efficient distillation paradigm empowers our policy to naturally approach objects with a forecasting intention and to effectively generalize to a wide range of unseen objects and motions.", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 239, + 209, + 257 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset (1) HandoverSim [9] contains 1000 real-world H2R handover scenes and 20 objects from DexYCB [8]. We evaluate on the \"s0\" setup which contains 720 training and 144 testing scenes. Each handover motion has a duration of 3 seconds. Following the evaluation of Handover-Sim2real [11], we consider \"Sequential\" and \"Simultaneous\" settings. In \"s0 (Sequential)\", the robot is allowed to move when the hand reaches the handover location and remains static. In \"s0 (Simultaneous)\", the robot is allowed to move from the beginning of the episode. (2) GenH2R-Sim contains 1,000,000 complex synthetic H2R handover scenes and 3266 objects. We evaluate the \"t0\" setup which contains 1,000,000 training and 3260 testing scenes. Each handover motion has a duration of 8s and will stop when the robot gripper is close to the object. To introduce more real-world handover scenes into GenH2R-Sim for evaluation, we extract and clip the handover point cloud sequence from HOI4D [28], a real-world mocap dataset. This additional setup is referred to as \"t1\", which only contains 1000 testing scenes for evaluation.", + "bbox": [ + 75, + 263, + 472, + 565 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics We adhere to the HandoverSim evaluation protocol. A successful handover involves grasping the object from the human hand and moving it to a designated location. Failure cases involve hand contact, object drop, and timeout $(T_{max} = 13s)$ . We report the successful rate and the execution time. Given that some policies prioritize success over speed, potentially wasting considerable human time, and others prioritize speed without considering success, we aim to evaluate both success rate and completion efficiency. To achieve this, we introduce AS (Average Success), akin to AP (Average Precision):", + "bbox": [ + 75, + 566, + 472, + 733 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {A S} = \\int_ {0} ^ {1} \\operatorname {S u c c e s s} (t) \\mathrm {d} t \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 738, + 468, + 773 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\operatorname{Success}(t)$ is success rate considering only successful cases within $t \\cdot T_{\\max}$ . This method can better evaluate success-time relations which is more suitable in our handover scenarios.", + "bbox": [ + 75, + 779, + 468, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Evaluating on Different Benchmarks", + "text_level": 1, + "bbox": [ + 76, + 847, + 397, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Setup We have 2 training sets: small-scale real-world \"s0\" from HandoverSim and large-scale synthetic \"t0\" from our", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "GenH2R-Sim. Evaluation is conducted on four testing sets: \"s0 (Sequential)\"/“s0 (Simultaneous)” from HandoverSim and \"t0\"/“t1” from our GenH2R-Sim. We conduct experiments on our forecast-aid 4D imitation learning from different demonstration strategies including destination planning, dense planning, and landmark planning. As discussed in Section 3.3, destination planning denotes the foresighted planner, dense planning denotes the improved shortsighted planner and landmark planning is our proposed method.", + "bbox": [ + 496, + 90, + 890, + 227 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines We compare our methods with Handover-Sim2real*, the state-of-the-art method in HandoverSim. We additionally compare GA-DDPG which is designed for grasping objects, and OMG Planner.", + "bbox": [ + 496, + 227, + 890, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results on different datasets As depicted in Table 1, our method trained on \"t0\" outperform all methods trained on \"s0\" by a large margin. Compared with Handover-Sim2real trained on \"s0\", our landmark planning method trained on \"t0\" exhibits $11.34\\%$ , $16.90\\%$ , $12.26\\%$ , and $15.93\\%$ increase in the success rate across the four testing sets. Moreover, compared with our landmark planning method trained on \"s0\", the version trained on \"t0\" demonstrates notable improvements, achieving success rate increases of $8.79\\%$ , $6.48\\%$ , $11.80\\%$ , and $14.13\\%$ increase in the same testing sets. This underscores the importance of having a substantial amount of synthetic data for handover training in simulation, which is more beneficial than only relying on a small-scale real-world dataset. Our GenH2R-Sim, with its large-scale complex human hand behavior, generalizes effectively to real-world scenarios such as \"s0\" in DexYCB and \"t1\" in HOI4D.", + "bbox": [ + 496, + 289, + 892, + 545 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results for different methods We can compare our methods with the baseline HandoverSim2real within the same training set in different benchmarks. When trained on \"s0\", our landmark planning method demonstrates improvements of $2.55\\%$ , $10.42\\%$ , $0.46\\%$ , and $1.8\\%$ ( $13.43\\%$ , $53.48\\%$ , $1.07\\%$ , and $23.60\\%$ in our reproduced version) across the 4 test sets. Similarly, When trained on \"t0\", our landmark planning method gives substantial improvements of $20.78\\%$ , $23.15\\%$ , $7.72\\%$ , and $21.23\\%$ ( $23.02\\%$ , $46.76\\%$ , $8.12\\%$ , and $34.98\\%$ in our reproduced version). The last 3 benchmarks (\"s0\"(simultaneous), \"t0\", and \"t1\") closely resemble real-world scenarios. They greatly demonstrate the effectiveness of our pipeline from distillation-friendly demonstrations to forecast-aided 4D imitation learning, which is capable of handling dynamic robot perception in complex handover scenarios. We also show visualizations on different methods in Figure 4 (a)(b).", + "bbox": [ + 496, + 546, + 893, + 803 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "*Our approach strictly adheres to the simultaneous setting defined in the paper of HandoverSim and HandoverSim2real: the robot moves from the beginning of the handover episode. However, it's noteworthy that HandoverSim2real manually makes their policy hold still in the first 1.5 seconds in the code implementation, deviating from the simultaneous setting definition. To ensure a fair comparison, we reproduce their results in the true simultaneous setting.", + "bbox": [ + 496, + 815, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16367", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/bc3eff356f8890ab4be5d3580bfe55c3f53fb3ff08745488ba1669bce15349c2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
s0 (Sequential)s0 (Simultaneous)t0t1
STASSTASSTASSTAS
OMG Planner† [42]62.508.3122.5---------
train on s0GA-DDPG [43]50.007.1422.536.814.6623.623.597.3110.346.75.5026.9
Handover-Sim2real [11]75.237.7430.468.756.2335.829.176.2915.052.407.0923.8
Handover-Sim2real* [11]64.357.6126.725.695.4315.028.564.7317.930.605.9816.5
Destination Planning74.319.0122.876.166.9835.225.685.9614.148.48.9415.1
Dense Planning74.779.5419.875.457.3233.027.306.2614.152.39.2415.1
Landmark Planning77.789.2422.379.177.2634.929.636.2315.454.29.0216.6
train on t0GA-DDPG [43]54.767.2624.244.685.3026.524.054.7015.325.505.8614.1
Handover-Sim2real [11]65.977.1829.562.506.0433.533.715.9118.447.106.3524.1
Handover-Sim2real* [11]63.557.5826.538.895.2923.133.314.6421.433.355.8118.4
Destination Planning0.9312.800.016.4812.410.35.968.811.91.6012.030.1
Dense Planning81.489.5121.984.957.4536.338.047.1617.157.908.8518.4
Landmark Planning86.578.8128.085.656.5842.841.436.0122.368.337.7027.9
", + "bbox": [ + 76, + 75, + 916, + 292 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results for different Demonstrations Trained on \"s0\" which consists of relatively simple trajectories, demonstrations based on destination planning can offer a rudimentary cue for downstream visuo-motor policy. However, when trained on \"t0\" this strategy may lose focus on the object, leading to a failure in maintaining vision-action correlation and providing minimal gains for vision-friendly learning. There is a significant $73.38\\%$ / $69.68\\%$ decrease in success rate in the \"s0\" setting. Additionally, distillation from landmark planning slightly outperforms dense planning in success rate and completes the handover process more quickly in all benchmarks. While dense planning can sustain the success rate to some extent, it slows down the agent and may result in unnatural approaches to objects. To jointly consider the time efficiency and the success rate, we compare the Average Success in methods distilled from these two strategies and find that landmark planning is a more efficient and generalizable approach. For instance, when trained on \"t0\", landmark planning exhibits significant improvements of $6.1\\%$ , $6.5\\%$ , $5.2\\%$ , and $9.5\\%$ across the four testing sets.", + "bbox": [ + 75, + 398, + 472, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Evaluating on different Dataset Scales", + "text_level": 1, + "bbox": [ + 76, + 726, + 406, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We have proved the crucial role of large-scale datasets in handover generalization in Section 4.1. We can also reveal it by scaling down the usage of \"t0\" in GenH2R-Sim which contains 1,000,000 training scenes. With $10\\%$ data utilization, we observe a $5.93\\%$ drop in the success rate on the unseen \"t1\" test set. This result proves the significance of the dataset scale in our imitation learning method. Thanks to our large-scale data and efficient demonstration generation pipeline, concerns about limited datasets hindering generalization are alleviated.", + "bbox": [ + 75, + 750, + 472, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c43aca7a791fd1e06d4a5bf72e063a0468d876c2dcfa876690b068dade7bc06b.jpg", + "table_caption": [ + "Table 1. Evaluating on different benchmarks. We compare our method against baselines from the test set of HandoverSim [9] benchmark (\"s0 (sequential)\" and \"s0 (simultaneous)\") and our GenH2R-Sim benchmark (\"t0\" and \"t1\"). We use the best-pretrained models from the repositories of GA-DDPG [43] and Handover-Sim2real [11] for evaluation. The results for our method are averaged across 3 random seeds. Note that S means success rate(%) . T means time(s). AS means average success(%) as defined in Equation 2. †: This method [42] is evaluated with ground-truth states and cannot handle dynamic handover like \"s0 (Simultaneous)\", \"t0\" and \"t1\".*: We reproduce the results of HandoverSim2real in the true simultaneous setting as detailed in Section 4.1 to make a fair comparison." + ], + "table_footnote": [], + "table_body": "
MethodsSTAS
w/o Flow31.665.6717.9
w/o Prediction39.186.1120.7
w/o Flow & Prediction37.045.9320.1
Ours41.436.0122.3
", + "bbox": [ + 500, + 397, + 893, + 470 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Ablations on different modules. \"w/o Flow\" means do not use flow information in the input. \"w/o Prediction\" means do not add prediction loss in the output.", + "bbox": [ + 498, + 479, + 893, + 522 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 537, + 653, + 555 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Table 2, we prove the effectiveness of our well-designed 4D imitation learning method. The absence of flow information results in a $9.77\\%$ decrease (predicting without past information adversely affects the model performance). The absence of the prediction task leads to a $2.25\\%$ decrease, and the absence of both components results in a $4.39\\%$ decrease. The results demonstrate the model obtains improved performance in leveraging flow information, particularly when tasked with predicting the future object pose. More ablations about our demonstration generation and imitation learning are detailed in the supplementary material.", + "bbox": [ + 496, + 561, + 893, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Real World Experiments", + "text_level": 1, + "bbox": [ + 500, + 739, + 728, + 757 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Sim-to-Real Transfer In addition to simulation, we deploy the models trained in GenH2R-Sim on a real robotic platform. Using point cloud input from the wrist-mounted camera, we employ the output 6D egocentric action to update the end effector's target position. A user study compares our method against Handover-Sim2real [11]. The supplementary material provides further details.", + "bbox": [ + 496, + 763, + 893, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "User Study We recruited 6 users to compare our method (based on landmark planning) and Handover-Sim2real", + "bbox": [ + 500, + 869, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16368", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5bd4ab0a985eb662e2b4f955c7931630c2d156a702d31873371906194b88439e.jpg", + "image_caption": [ + "Figure 4. Qualitative results. We in detail compare different methods in simulators and deploy them in the real-world platform." + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 803, + 657 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1c374940059030824c8afde3c7d969889f8374acd9e7f062d3862535a71af059.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsSimple SettingComplex Setting
Handover-Sim2real56.7%33.3%
Ours90.0%70.0%
", + "bbox": [ + 78, + 686, + 470, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Sim-to-Real Experiments. We report the success rate of our method and HandoverSim2real in 2 different settings.", + "bbox": [ + 75, + 739, + 470, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "across 5 objects in 2 different settings. In the simple setting, users hand each object to the gripper without quick movements. In the complex setting, users execute a relatively long and quick trajectory. The results are reported in Table 3. We observe that our model gets better performance in completing the handover process across various objects and scenarios. Figure 4(c) shows examples of the real-world handover trials.", + "bbox": [ + 75, + 779, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 688, + 617, + 704 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we present a novel framework GenH2R for scaling up the learning of human-to-robot handover. We introduce a new simulator GenH2R-Sim and generate a million human handover animations to facilitate generalizable H2R handover learning. We then propose a distillation-friendly demonstration generation method that automatically produces a million high-quality demonstrations suitable for learning. We further introduce a forecast-aided 4D imitation learning method for effective demonstration distillation. Our experiments demonstrate that scaling-up efforts result in substantial improvement of generalizability to novel geometry and complex motion, both in the simulator and the real world.", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16369", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 121, + 173, + 136 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Christoph Bartneck, Tony Belpaeme, Friederike Eyssel, Takayuki Kanda, Merel Keijrsers, and Selma Šabanović. Human-robot interaction: An introduction. Cambridge University Press, 2020. 2", + "[2] Antonio Bicchi and Vijay Kumar. Robotic grasping and contact: A review. In Proceedings 2000 ICRA. Millennium conference. IEEE international conference on robotics and automation. Symposia proceedings (Cat. No. 00CH37065), pages 348-353. IEEE, 2000. 2", + "[3] Aude Billard, Sylvain Calinon, Ruediger Dillmann, and Stefan Schaal. Survey: Robot programming by demonstration. Technical report, Springer, 2008. 3", + "[4] Jeannette Bohg, Antonio Morales, Tamim Asfour, and Dania Kragic. Data-driven grasp synthesis—a survey. IEEE Transactions on robotics, 30(2):289-309, 2013. 2", + "[5] Samarth Brahmbhatt, Chengcheng Tang, Christopher D Twigg, Charles C Kemp, and James Hays. Contactpose: A dataset of grasps with object contact and hand pose. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16, pages 361-378. Springer, 2020. 2", + "[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1", + "[7] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 2, 3", + "[8] Yu-Wei Chao, Wei Yang, Yu Xiang, Pavlo Molchanov, Ankur Handa, Jonathan Tremblay, Yashraj S Narang, Karl Van Wyk, Umar Iqbal, Stan Birchfield, et al. Dexycb: A benchmark for capturing hand grasping of objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9044–9053, 2021. 2, 3, 6, 14, 17", + "[9] Yu-Wei Chao, Chris Paxton, Yu Xiang, Wei Yang, Balakumar Sundaralingam, Tao Chen, Adithyavairavan Murali, Maya Cakmak, and Dieter Fox. Handoversim: A simulation framework and benchmark for human-to-robot object handovers. In 2022 International Conference on Robotics and Automation (ICRA), pages 6941–6947. IEEE, 2022. 1, 2, 3, 6, 7, 12, 14, 16, 17, 18", + "[10] Sammy Christen, Lan Feng, Wei Yang, Yu-Wei Chao, Otmar Hilliges, and Jie Song. Synh2r: Synthesizing handobject motions for learning human-to-robot handovers. arXiv preprint arXiv:2311.05599, 2023. 2", + "[11] Sammy Christen, Wei Yang, Claudia Pérez-D'Arpino, Otmar Hilliges, Dieter Fox, and Yu-Wei Chao. Learning human-to-robot handovers from point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and" + ], + "bbox": [ + 78, + 146, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Pattern Recognition, pages 9654-9664, 2023. 2, 5, 6, 7, 13, 14, 16, 17", + "[12] Gianluca Corsini, Martin Jacquet, Hemjyoti Das, Amr Affi, Daniel Sidobre, and Antonio Franchi. Nonlinear model predictive control for human-robot handover with application to the aerial case. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7597–7604. IEEE, 2022. 2", + "[13] Murtaza Dalal, Ajay Mandlekar, Caelan Garrett, Ankur Handa, Ruslan Salakhutdinov, and Dieter Fox. Imitating task and motion planning with visuomotor transformers. arXiv preprint arXiv:2305.16309, 2023. 3", + "[14] Yuhao Dong, Zhuoyang Zhang, Yunze Liu, and Li Yi. Nsm4d: Neural scene model based online 4d point cloud sequence understanding. arXiv preprint arXiv:2310.08326, 2023.5", + "[15] Frederik Ebert, Yanlai Yang, Karl Schmeckpeper, Bernadette Bucher, Georgios Georgakis, Kostas Daniilidis, Chelsea Finn, and Sergey Levine. Bridge data: Boosting generalization of robotic skills with cross-domain datasets. arXiv preprint arXiv:2109.13396, 2021. 3", + "[16] Clemens Eppner, Arsalan Mousavian, and Dieter Fox. Acronym: A large-scale grasp dataset based on simulation. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 6222-6227. IEEE, 2021. 2, 5", + "[17] Hehe Fan, Yi Yang, and Mohan Kankanhalli. Point 4d transformer networks for spatio-temporal modeling in point cloud videos. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14204-14213, 2021. 5", + "[18] Zicong Fan, Omid Taheri, Dimitrios Tzionas, Muhammed Kocabas, Manuel Kaufmann, Michael J Black, and Otmar Hilliges. Arctic: A dataset for dexterous bimanual handobject manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12943-12954, 2023. 2", + "[19] Hao-Shu Fang, Chenxi Wang, Hongjie Fang, Minghao Gou, Jirong Liu, Hengxu Yan, Wenhai Liu, Yichen Xie, and Cewu Lu. Anygrasp: Robust and efficient grasp perception in spatial and temporal domains. IEEE Transactions on Robotics, 2023. 2", + "[20] Chelsea Finn, Tianhe Yu, Tianhao Zhang, Pieter Abbeel, and Sergey Levine. One-shot visual imitation learning via meta-learning. In Conference on robot learning, pages 357-368. PMLR, 2017. 3", + "[21] Justin Fu, Aviral Kumar, Ofir Nachum, George Tucker, and Sergey Levine. D4rl: Datasets for deep data-driven reinforcement learning. arXiv preprint arXiv:2004.07219, 2020.3", + "[22] Caelan Reed Garrett, Rohan Chitnis, Rachel Holladay, Beomjoon Kim, Tom Silver, Leslie Pack Kaelbling, and Tomás Lozano-Pérez. Integrated task and motion planning. Annual review of control, robotics, and autonomous systems, 4:265-293, 2021. 2, 3", + "[23] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d:" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16370", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022. 3", + "[24] Huy Ha, Pete Florence, and Shuran Song. Scaling up and distilling down: Language-guided robot skill acquisition. arXiv preprint arXiv:2307.14535, 2023. 2, 3", + "[25] Yana Hasson, Gul Varol, Dimitrios Tzionas, Igor Kalevatykh, Michael J Black, Ivan Laptev, and Cordelia Schmid. Learning joint reconstruction of hands and manipulated objects. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11807-11816, 2019. 2", + "[26] Eric Jang, Alex Irpan, Mohi Khansari, Daniel Kappler, Frederik Ebert, Corey Lynch, Sergey Levine, and Chelsea Finn. Bc-z: Zero-shot task generalization with robotic imitation learning. In Conference on Robot Learning, pages 991-1002. PMLR, 2022. 3", + "[27] Yi Li, Gu Wang, Xiangyang Ji, Yu Xiang, and Dieter Fox. Deepim: Deep iterative matching for 6d pose estimation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 683-698, 2018. 5, 14", + "[28] Yunze Liu, Yun Liu, Che Jiang, Kangbo Lyu, Weikang Wan, Hao Shen, Boqiang Liang, Zhoujie Fu, He Wang, and Li Yi. Hoi4d: A 4d egocentric dataset for category-level human-object interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21013-21022, 2022. 2, 6", + "[29] Ajay Mandlekar, Danfei Xu, Josiah Wong, Soroush Nasiriy, Chen Wang, Rohun Kulkarni, Li Fei-Fei, Silvio Savarese, Yuke Zhu, and Roberto Martin-Martin. What matters in learning from offline human demonstrations for robot manipulation. arXiv preprint arXiv:2108.03298, 2021. 3", + "[30] Naresh Marturi, Marek Kopicki, Alireza Rastegarpanah, Vijaykumar Rajasekaran, Maxime Adjigble, Rustam Stolkin, Ales Leonidis, and Yasemin Bekiroglu. Dynamic grasp and trajectory planning for moving objects. Autonomous Robots, 43:1241-1256, 2019. 2", + "[31] Michael James McDonald and Dylan Hadfield-Menell. Guided imitation of task and motion planning. In Conference on Robot Learning, pages 630-640. PMLR, 2022. 3", + "[32] OpenAI. Gpt-4 technical report, 2023. 1", + "[33] Valerio Ortenzi, Akansel Cosgun, Tommaso Pardi, Wesley P Chan, Elizabeth Croft, and Dana Kulić. Object handovers: a review for robotics. IEEE Transactions on Robotics, 37(6): 1855-1873, 2021. 2", + "[34] Dean A Pomerleau. Alvinn: An autonomous land vehicle in a neural network. Advances in neural information processing systems, 1, 1988. 3", + "[35] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 5, 14", + "[36] Patrick Rosenberger, Akansel Cosgun, Rhys Newbury, Jun Kwan, Valerio Ortenzi, Peter Corke, and Manfred Grafinger. Object-independent human-to-robot handovers using real time robotic vision. IEEE Robotics and Automation Letters, 6(1):17-23, 2020. 2" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Szymon Rusinkiewicz and Marc Levoy. Efficient variants of the icp algorithm. In Proceedings third international conference on 3-D digital imaging and modeling, pages 145-152. IEEE, 2001. 5, 14", + "[38] Thomas B Sheridan. Human–robot interaction: status and challenges. Human factors, 58(4):525-532, 2016. 2", + "[39] Shuran Song, Andy Zeng, Johnny Lee, and Thomas Funkhouser. Grasping in the wild: Learning 6dof closed-loop grasping from low-cost demonstrations. IEEE Robotics and Automation Letters, 5(3):4978-4985, 2020. 3", + "[40] Zachary Teed and Jia Deng. Raft-3d: Scene flow using rigid-motion embeddings. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8375–8384, 2021. 5", + "[41] Lirui Wang, Yu Xiang, and Dieter Fox. Manipulation trajectory optimization with online grasp synthesis and selection. In Robotics: Science and Systems (RSS), 2020. 4", + "[42] Lirui Wang, Yu Xiang, and Dieter Fox. Manipulation trajectory optimization with online grasp synthesis and selection. In Proceedings of Robotics: Science and Systems (RSS), 2020. 5, 7, 13, 14", + "[43] Lirui Wang, Yu Xiang, Wei Yang, Arsalan Mousavian, and Dieter Fox. Goal-auxiliary actor-critic for 6d robotic grasping with point clouds. In Conference on Robot Learning, pages 70-80. PMLR, 2022. 5, 7, 14, 16, 17", + "[44] Lirui Wang, Yiyang Ling, Zhecheng Yuan, Mohit Shridhar, Chen Bao, Yuzhe Qin, Bailin Wang, Huazhe Xu, and Xiaolong Wang. Gensim: Generating robotic simulation tasks via large language models. arXiv preprint arXiv:2310.01361, 2023. 3", + "[45] Ruicheng Wang, Jialiang Zhang, Jiayi Chen, Yinzhen Xu, Puhao Li, Tengyu Liu, and He Wang. Dexgraspnet: A large-scale robotic dexterous grasp dataset for general objects based on simulation. arXiv preprint arXiv:2210.02697, 2022.3", + "[46] Ruicheng Wang, Jialiang Zhang, Jiayi Chen, Yinzhen Xu, Puhao Li, Tengyu Liu, and He Wang. Dexgraspnet: A large-scale robotic dexterous grasp dataset for general objects based on simulation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 11359-11366. IEEE, 2023. 2", + "[47] Yufei Wang, Zhou Xian, Feng Chen, Tsun-Hsuan Wang, Yian Wang, Katerina Fragkiadaki, Zackory Erickson, David Held, and Chuang Gan. Robogen: Towards unleashing infinite data for automated robot learning via generative simulation. arXiv preprint arXiv:2311.01455, 2023. 3", + "[48] Hao Wen, Yunze Liu, Jingwei Huang, Bo Duan, and Li Yi. Point primitive transformer for long-term 4d point cloud video understanding. In European Conference on Computer Vision, pages 19-35. Springer, 2022. 5", + "[49] Wei Yang, Chris Paxton, Arsalan Mousavian, Yu-Wei Chao, Maya Cakmak, and Dieter Fox. Reactive human-to-robot handovers of arbitrary objects. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 3118–3124. IEEE, 2021. 2", + "[50] Ruolin Ye, Wenqiang Xu, Zhendong Xue, Tutian Tang, Yanfeng Wang, and Cewu Lu. H2o: A benchmark for visual" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16371", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "human-human object handover analysis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15762-15771, 2021. 2", + "[51] Gu Zhang, Hao-Shu Fang, Hongjie Fang, and Cewu Lu. Flexible handover with real-time robust dynamic grasp trajectory generation. arXiv preprint arXiv:2308.15622, 2023. 2", + "[52] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 1", + "[53] Tianhao Zhang, Zoe McCarthy, Owen Jow, Dennis Lee, Xi Chen, Ken Goldberg, and Pieter Abbeel. Deep imitation learning for complex manipulation tasks from virtual reality teleoperation. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pages 5628-5635. IEEE, 2018. 3" + ], + "bbox": [ + 78, + 90, + 468, + 343 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "16372", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/GenH2R_ Learning Generalizable Human-to-Robot Handover via Scalable Simulation Demonstration and Imitation/40314ff6-3af8-45b8-b8fc-6f9ebdc7d993_model.json b/2024/GenH2R_ Learning Generalizable Human-to-Robot Handover via Scalable Simulation Demonstration and Imitation/40314ff6-3af8-45b8-b8fc-6f9ebdc7d993_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9b7ea9cc3df2bf101ed88c2130aca66eb6b7913b --- /dev/null +++ b/2024/GenH2R_ Learning Generalizable Human-to-Robot Handover via Scalable Simulation Demonstration and Imitation/40314ff6-3af8-45b8-b8fc-6f9ebdc7d993_model.json @@ -0,0 +1,1960 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.172, + 0.131, + 0.8, + 0.175 + ], + "angle": 0, + "content": "GenH2R: Learning Generalizable Human-to-Robot Handover via Scalable Simulation, Demonstration, and Imitation" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.203, + 0.834, + 0.223 + ], + "angle": 0, + "content": "Zifan Wang\\(^{1,3}\\) Junyu Chen\\(^{1,3}\\) Ziqing Chen\\(^{1}\\) Pengwei Xie\\(^{1}\\) Rui Chen\\(^{1}\\) Li Yi\\(^{\\dagger 1,2,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.224, + 0.86, + 0.261 + ], + "angle": 0, + "content": "\\(^{1}\\)Tsinghua University \\(^{2}\\)Shanghai Artificial Intelligence Laboratory \\(^{3}\\)Shanghai Qi Zhi Institute https://GenH2R.github.io" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.296, + 0.314, + 0.312 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.328, + 0.474, + 0.66 + ], + "angle": 0, + "content": "This paper presents GenH2R, a framework for learning generalizable vision-based human-to-robot (H2R) handover skills. The goal is to equip robots with the ability to reliably receive objects with unseen geometry handed over by humans in various complex trajectories. We acquire such generalizability by learning H2R handover at scale with a comprehensive solution including procedural simulation assets creation, automated demonstration generation, and effective imitation learning. We leverage large-scale 3D model repositories, dexterous grasp generation methods, and curve-based 3D animation to create an H2R handover simulation environment named GenH2R-Sim, surpassing the number of scenes in existing simulators by three orders of magnitude. We further introduce a distillation-friendly demonstration generation method that automatically generates a million high-quality demonstrations suitable for learning. Finally, we present a 4D imitation learning method augmented by a future forecasting objective to distill demonstrations into a visuo-motor handover policy. Experimental evaluations in both simulators and the real world demonstrate significant improvements (at least +10% success rate) over baselines in all cases." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.689, + 0.21, + 0.705 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.714, + 0.47, + 0.85 + ], + "angle": 0, + "content": "The embodied AI research community has long been driven by the goal of empowering robots to interact and collaborate with humans. A crucial aspect of this pursuit is equipping robots with the capability to reliably receive arbitrarily moving objects of varying geometry handed over by humans, based on dynamic visual observations. This human-to-robot (H2R) handover ability allows robots to seamlessly collaborate with humans across a wide range of tasks, including cooking, room tidying, and furniture assembly." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.851, + 0.468, + 0.866 + ], + "angle": 0, + "content": "However, compared to learning human-free robot manip-" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.295, + 0.689, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.401, + 0.685, + 0.409 + ], + "angle": 0, + "content": "Scalable Synthetic Handover Simulator" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.295, + 0.874, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.697, + 0.401, + 0.882, + 0.409 + ], + "angle": 0, + "content": "Distillation-friendly Expert Demonstrations" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.412, + 0.69, + 0.506 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.549, + 0.508, + 0.655, + 0.518 + ], + "angle": 0, + "content": "Sim-to-real Deployment" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.411, + 0.872, + 0.505 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.702, + 0.507, + 0.867, + 0.518 + ], + "angle": 0, + "content": "Forecast-Aided 4D Imitation Learning" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.53, + 0.892, + 0.614 + ], + "angle": 0, + "content": "Figure 1. The overview of GenH2R. We introduce a framework for learning generalizable vision-based human-to-robot handover via scalable synthetic simulation, distillation-friendly expert demonstration generation, and a forecast-aided 4D imitation learning method. Our models demonstrate strong generalization capabilities to real datasets and can be deployed to a real robot." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.895, + 0.903 + ], + "angle": 0, + "content": "ulation skills, the progress in scalably learning H2R handover that can generalize to various objects and versatile human behaviors has lagged due to its unique challenges. Training robots to interact with humans in real-world scenarios entails increased risks and expenses, rendering it inherently non-scalable. Therefore, it is demanded to simulate human behaviors and train robots in simulated environments prior to real-world deployment. However, creating a substantial number of assets for humans handing over objects poses a significant challenge. In a recent study [9] that employed motion capturing to drive virtual humans in a simulator, only 1000 unique human hand motion trajectories were provided for handing over 20 objects. Limited object geometry and human motion assets can hardly capture the complexities of the real world. Besides, the challenge extends to the demonstration side. The success of large language model [6, 32, 52] has suggested a recipe for scaling up learning through modeling large-scale training" + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.875, + 0.424, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution with the order determined by rolling dice." + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.888, + 0.223, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.095, + 0.875, + 0.424, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16362" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.152 + ], + "angle": 0, + "content": "data. Nevertheless, collecting robot demonstrations receiving objects from real humans is very costly and unscalable. How to scale up the number of demonstrations while ensuring effective learning poses additional challenges." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.155, + 0.468, + 0.23 + ], + "angle": 0, + "content": "In this work, we aim to learn generalizable H2R handover at scale by tackling the above challenges. We present a comprehensive solution that scales up both the assets and demonstrations and effectively learns a closed-loop visuomotor policy through a novel imitation learning algorithm." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.232, + 0.468, + 0.474 + ], + "angle": 0, + "content": "Specifically, to scale up geometry and motion assets depicting humans handing over various objects, we leverage large-scale 3D model repositories [7, 16], dexterous grasp generation methods [46], and curve-based 3D animation. This enables us to procedurally generate millions of handover scenes, forming an environment named GenH2R-Sim to support generalizable H2R handover learning. GenH2R-Sim surpasses HandoverSim [9], an existing H2R simulator, in both scene quantity (by three orders of magnitude) and unique object involvement (by two orders of magnitude). In addition, scenes in GenH2R-Sim go beyond a straightforward giving and then receiving and cover cases when humans might keep transforming the object in a large range during the entire H2R handover process. This allows for studying complex behaviors such as humans hesitating before handing over." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.476, + 0.468, + 0.763 + ], + "angle": 0, + "content": "To scale up robot demonstrations, we draw inspiration from the Task and Motion Planning (TAMP) [22] literature and propose to automatically generate demonstrations with grasp and motion planning using privileged human motion and object state information. There are some straightforward ways to achieve this goal, such as using the privileged human handover destination information to plan a smooth demonstration. However, the problem is more challenging than it seems since the generated demonstrations need to be suitable for distilling into a visuomotor policy. We identify the vision-action correlation between visual observations and planned actions as the crucial factor influencing distillability and point out that due to the constraints of robot arm morphology one can easily generate observation-irrelevant actions and thus harm distillation. To tackle this challenge, we present a distillation-friendly demonstration generation method that sparsely samples handover animations for landmark states and periodically replans grasp and motion based on privileged future landmarks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Finally, to distill the above demonstrations into a visuomotor policy, we utilize point cloud input for its richer geometric information and smaller sim-vs-real gap compared to images. We propose a 4D imitation learning method that factors the sequential point cloud observations into geometry and motion parts, facilitating policy learning by better revealing the current scene state. Furthermore, the imitation objective is augmented by a forecasting objective which predicts the future motion of the handover object. Since our" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.891, + 0.136 + ], + "angle": 0, + "content": "demonstrating actions are generated based on future landmarks, the forecasting objective can help further exploit the vision-action correlation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.137, + 0.892, + 0.288 + ], + "angle": 0, + "content": "We evaluate our learned policy in simulators (Handover-Sim and our own GenH2R-Sim) and the real world. Remarkably, without any mocap assets or real-world demonstrations, our method achieves significantly better performance compared to baselines across all settings (at least \\(+10\\%\\) success rate). Our experiments highlight that the scaling-up efforts bring substantial improvement in policy generalizability to novel geometry and complex motion. Furthermore, these efforts greatly facilitate skill transfer to real robotic systems." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.289, + 0.892, + 0.425 + ], + "angle": 0, + "content": "In summary, the key contribution of this paper is a novel framework scaling up the learning of H2R handover with the following three components: i) a simulation environment named GenH2R-Sim consists of millions of human handover animations for generalizable H2R handover learning, ii) an empirically validated automatic robot demonstration generation pipeline for vision-based closed-loop control, iii) a forecast-aided 4D imitation learning method effective in distilling the large-scale demonstrations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.47, + 0.642, + 0.485 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.495, + 0.757, + 0.51 + ], + "angle": 0, + "content": "2.1. Human-to-Robot Handovers" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.519, + 0.892, + 0.821 + ], + "angle": 0, + "content": "Recently, significant progress in human-robot handovers [12, 33, 36] has been observed, driven by the increasing popularity of human-robot interaction [1, 38] and the emergence of extensive datasets [5, 8, 18, 25, 28, 50] capturing hand-object interactions. Some traditional methods [2, 4] require 3D object models and struggle to handle unseen objects. One possible way is to consider grasping and dynamic motion planning [19, 30, 49, 51]. However, these methods often exhibit constrained motions and perform poorly on large-scale datasets. HandoverSim [9], a physics-simulated environment, introduced a new simulation benchmark for human-to-robot object handovers. Leveraging DexYCB [8], a dataset of human grasping objects and performing handover attempts, this environment allows training learning-based handover policies such as [11]. However, it lacks large-scale and diverse handover scenes, which limits generalizable handovers. At the same time, SynH2R [10] proposes to use synthetic data but makes limited progress. Building on this, we propose GenH2R-Sim , aiming to benchmark generalizable handover." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.832, + 0.798, + 0.848 + ], + "angle": 0, + "content": "2.2. Scaling Up Robot Demonstrations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "For robot learning, scaling up data collection for manipulation skills has spurred extensive research. Approaches include leveraging large language models [24] or hardware" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16363" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.089, + 0.442, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.443, + 0.089, + 0.653, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.089, + 0.872, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.343, + 0.892, + 0.426 + ], + "angle": 0, + "content": "Figure 2. The overview of our framework. First, we propose a new simulation environment named GenH2R-Sim, featuring large-scale synthetic datasets with diversity in object geometry, grasp poses, and complex trajectories. Second, other than destination planning (move straight toward the final position) and dense planning (replan at each step), we propose a distillation-friendly demonstration generation method—landmark planning, predicting landmarks on the trajectory (as indicated by the dashed object above) and replanning based on those landmarks. Thirdly, our Forecast-aided 4D Imitation Learning leverages past flow information, and the forecasting objective enhances the exploitation of vision-action correlation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.438, + 0.47, + 0.59 + ], + "angle": 0, + "content": "capabilities [39], utilizing non-robotics datasets [23], and employing trial-and-error explorations [21]. As depicted in [24], one of the challenges is scaling up robot-complete data. A popular line of research scales up demonstration generation via Task and Motion Planning [13, 22, 31]. These works usually focus on fairly static scenes without active motion or object and task variety [44, 47] while our method extends to dynamic H2R handover by considering how to interpret human behavior and generate demonstrations easy to be distilled by closed-loop visuo-motor policy." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.601, + 0.411, + 0.617 + ], + "angle": 0, + "content": "2.3. Offline Learning from Demonstrations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.624, + 0.47, + 0.806 + ], + "angle": 0, + "content": "Imitation Learning (IL) represents a methodology for training embodied agents in manipulation tasks by utilizing expert demonstrations. The commonly used Behavior Cloning (BC) [34] strategy directly trains the policy to imitate expert actions in a supervised learning manner. Despite its simplicity, this approach has demonstrated remarkable effectiveness in robotic manipulation [3, 20, 29, 53] especially when combined with a substantial number of high-quality demonstrations [15, 26]. Inspired by these works, we adopt an imitation learning paradigm, focusing on how to leverage spatial-temporal perception and future forecasting to better consume our distillation-friendly demonstrations." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.821, + 0.168, + 0.836 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.847, + 0.187, + 0.861 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "For the generalizable H2R handover task, we introduce GenH2R, a framework designed to learn control policies," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.438, + 0.892, + 0.545 + ], + "angle": 0, + "content": "specifically 6D control actions for the robot gripper, using segmented point cloud data captured from an egocentric camera. We describe our method for synthesizing human handover animations in Section 3.2, generating expert demonstrations in Section 3.3, and distilling demonstrations to 4D vision-based neural networks by imitation learning in Section 3.4, as the pipeline depicted in Figure 2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.557, + 0.64, + 0.571 + ], + "angle": 0, + "content": "3.2. GenH2R-Sim" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.582, + 0.892, + 0.717 + ], + "angle": 0, + "content": "The size and quality of human-object datasets in simulators play a crucial role in generating high-quality handover demonstrations and training reliable policies for handover scenarios. The recent handover simulator, HandoverSim [9], utilizes the DexYCB [8] dataset, which captures real-world human grasping objects in a limited manner, comprising only 1000 scenes with 20 distinct objects. In the real world, scenarios can be more complex and may involve intricate trajectories and poses beyond those in DexYCB." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.719, + 0.892, + 0.809 + ], + "angle": 0, + "content": "To address these limitations, we introduce a new environment, GenH2R-Sim, to overcome these deficiencies and facilitate generalizable handovers. To diversify geometry and motion assets depicting humans handing over various objects, we focus on two primary aspects: the hand grasping pose and the hand-object moving trajectory within a scene." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In aspects of grasping poses, DexGraspNet [45] has made significant contributions by employing optimization techniques to generate a substantial dataset of human hand grasp poses. We utilize this method to generate approximately 1,000,000 grasp poses for 3,266 different objects sourced from Shapenet [7]. These objects span a wide range" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16364" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.136 + ], + "angle": 0, + "content": "of categories, from larger items like computers to smaller ones like mobile phones, covering most sizes and shapes encountered in real-life handovers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.138, + 0.471, + 0.394 + ], + "angle": 0, + "content": "In aspects of hand-object moving trajectories, we propose to use Bézier curves, which are one class of smooth curves determined by several control points, to generate complex yet smooth-transiting motion trajectories. We use multiple Bézier curves to model different stages of the motion, and link the ends of these curves to create a seamless track. We can generate scenes matching various scenarios of different complexity in the real world by adjusting the distribution of control points of the trajectory and the speed of the human hand. To enhance the trajectory's realism, we incorporate consistent object rotations, which also enhances the importance of choosing the appropriate grasp for the robotic arm. Since we can always attach a new segment of motion at the end of the current motion and the duration is much longer than DexYCB scenes, the destination of the hand-object is not a significant factor, so we just randomly select a point within the reach of the robotic arm." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.396, + 0.47, + 0.56 + ], + "angle": 0, + "content": "We do not guarantee that every item in the dataset we generate perfectly mimics the human-like characteristics of real-world data, but our approach ensures a significantly higher degree of domain randomization and provides greater diversity in terms of geometry and motion. Given the challenges in scaling up real-world motion capture datasets, we opt for a large-scale synthetic dataset for our handover simulations. Our key insight is that for both demonstrations and policy learning, having a substantial amount of synthetic data is more beneficial than relying on a small-scale real-world dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.563, + 0.47, + 0.729 + ], + "angle": 0, + "content": "GenH2R-Sim follows the setup of HandoverSim, which consists of a Panda 7DoF robotic arm with a gripper and a wrist-mounted RGB-D camera, and a simulated human hand. Just like HandoverSim, we switch from the prehandover kinematic phase to the handover dynamic phase when the object has been in contact with the gripper. HandoverSim is not adaptive to the robot's action and just loads and replays every frame of the data. To align with the real-world handover process more naturally in GenH2R-Sim, the simulated hand will stop from moving and wait for handover when the robot arm is close to the object." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.741, + 0.443, + 0.756 + ], + "angle": 0, + "content": "3.3. Generating Demonstrations for Distillation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.901 + ], + "angle": 0, + "content": "In this section, we address a key question in learning visuomotor policy: how to efficiently generate robot demonstrations that incorporate paired vision-action data from successful task experiences. While distilling successful demonstrations into a single policy has proven effective for open-loop control tasks, the challenge lies in closed-loop visuo-motor control, where the quality of demonstrations becomes crucial for learning. Merely ensuring success is no longer sufficient. We present two examples of demon-" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.09, + 0.892, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.312, + 0.892, + 0.382 + ], + "angle": 0, + "content": "Figure 3. Different demonstration generation methods for dynamic handover. The orange curve shows the hand-object trajectory. The blue, red, and green curves show the example trajectories generated by the foresighted planner, the shortsighted planner, and our planner, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.399, + 0.893, + 0.716 + ], + "angle": 0, + "content": "stration generation with different grasp and motion planning strategies as shown in Figure 3. In the first example, a foresighted planner generates smooth, short demonstrations based on the privileged destination end state of a human handover animation. Though efficient, the planned path does not align actions with the dynamic visual observations during the handover. Distilling such demonstrations requires accurately forecasting the end state of the human trajectory, which can be extremely challenging in complex handover cases. The second example involves a shortsighted planner that independently replans grasp and motion at each time step using privileged hand and object states. Due to robot morphology constraints and the multisolution nature of common robot planners, smooth visual observations may correspond to unsmooth and multi-modal robot trajectories, increasing the difficulty of distillation. We emphasize the importance of distillability as a quality factor for handover demonstrations. An effective demonstration generation method must consider the vision-action correlation by jointly incorporating robot morphology and dynamic vision during grasp and motion planning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.718, + 0.892, + 0.838 + ], + "angle": 0, + "content": "Along this line, we base our method on the foresighted and shortsighted planner mentioned above to combine the advantages of both sides while encouraging the demonstration distillability. We first improve the shortsighted planner so that sequentially smooth visual observations result in smooth grasp and motion plans. Then we improve the handover efficiency by looking toward the future while guaranteeing the vision-action correlation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.902 + ], + "angle": 0, + "content": "To be specific, we build our method based on the OMG planner [41] for grasp and motion planning. This planner optimizes the grasp and motion path by considering the object's 6D pose and a set of candidate grasp poses. To support" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16365" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.381 + ], + "angle": 0, + "content": "this optimization, we provide privileged knowledge that includes the object's 6D pose, candidate grasps generated through physics simulation [16], and human hand poses for filtering out invalid grasps. However, independently calling the OMG planner for each time step may result in unsmooth trajectories, as it is designed for static scenarios. To address this, we sequentially plan the grasp and motion based on the privileged knowledge by: 1) sorting grasps based on their pose distance to the robot end effector and attempting inverse kinematics (IK) starting from the nearest grasp until success; 2) initializing IK based on the robot arm pose from the previous time step; 3) invoking the OMG planner only when IK can be successfully solved. By prioritizing closer grasps, we encourage the object to remain within the field of view of a wrist camera, reducing visually irrelevant actions when the object is not visible. Additionally, enforcing IK smoothness improves the overall trajectory smoothness. As a result, the enhanced vision-action correlation dramatically improves the demonstration quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.388, + 0.473, + 0.904 + ], + "angle": 0, + "content": "Our approach modifies the OMG planner for dynamic grasp and motion planning. However, densely replanning at each time step leads to inefficient and non-smooth zigzag demonstrations, which does not align with how humans receive objects. Humans anticipate dynamic scene changes before taking action. On the other hand, a highly fore-sighted planner that directly plans grasp and motion based on the end state of a human handover animation can disrupt the vision-action correlation. To strike a balance between these extremes, we propose an algorithm that sparsely samples handover animations for landmark states and periodically replans grasp and motion based on future landmarks. The key idea is to select landmarks strategically so that the planner only considers visually foreseeable futures. Specifically, let \\(\\xi = (\\mathcal{T}_0,\\mathcal{T}_1,\\dots ,\\mathcal{T}_{T - 1})\\) represent an object trajectory, where \\(\\mathcal{T}_t\\in \\mathbb{SE}(3)\\) denotes the object pose in the \\(t\\)-th frame within the world coordinate system. Based on all the object trajectories in the training set, we train an object pose forecasting network which consumes past and current object poses \\((\\mathcal{T}_0,\\mathcal{T}_1,\\dots ,\\mathcal{T}_t)\\) for each time step \\(t\\) within each trajectory and forecasts the object poses \\((\\mathcal{T}_{t + 1},\\mathcal{T}_{t + 2},\\dots ,\\mathcal{T}_{t + N})\\) in future \\(N\\) steps. By thresholding the forecasting error corresponding to each time step, we identify a set of endpoints where past observations cannot forecast the future very well and partition the complete trajectory \\(\\xi\\) to several segments using endpoints \\(0 = l_{0} < l_{1} < \\dots < l_{k} = T\\). Within each segment, we assume the ability to predict the future object pose based on historical information. We then denote \\(P\\in \\mathbb{N}\\) as the hyperparameter determining the replanning period. For each planning frame \\(t = 0,P,2P,\\dots\\), suppose the next endpoint is \\(l_{i + 1}\\), i.e., \\(l_{i}\\leq t < l_{i + 1}\\). Then we will plan based on the object pose at frame \\(\\hat{t} = \\min (t + P,l_{i + 1})\\), which serves as a landmark. Note here planning is based on the future states" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "but avoids bypassing the sharply transitioning points where human motion becomes unpredictable. Also worth mentioning, densely planning is a special case of our method, and landmark planning is a full version." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.163, + 0.833, + 0.181 + ], + "angle": 0, + "content": "3.4. Forecast-Aided 4D Imitation Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.187, + 0.892, + 0.353 + ], + "angle": 0, + "content": "Traditional methods for human-to-robot handover face challenges in gaining insights into dynamic scene perception. Approaches based on motion planning [42] often emphasize robot morphology and lack dynamic vision perception. They struggle to capture long-horizon information, mainly focusing on the current frame and failing to predict the future. Reinforcement Learning methods [11, 43], while powerful, require extensive training and may train unstably across different scenarios. To enhance the vision-action correlation and establish an efficient training paradigm, we introduce our forecast-aided 4D imitation learning approach." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.354, + 0.893, + 0.701 + ], + "angle": 0, + "content": "In robot perception, the 4D point cloud serves as the common representation. In the \\(t\\)-th frame, we can define \\(M_t^i \\in \\mathbb{SE}(3)\\) as the relative object pose between the current frame and the \\(i\\)-th frame in the egocentric view. While frame stacking is a straightforward approach, it struggles to capture both motion and geometry effectively. Inspired by recent 4D learning methods [14, 40], we employ the Iterative Closest Point (ICP) registration algorithm [37] to efficiently compute transformation matrices \\(\\{\\hat{M}_t^{t-1}, \\hat{M}_t^{t-2}, \\dots, \\hat{M}_t^{t-L_1}\\}\\) between the point cloud in the \\(t\\)-th frame and the point clouds in previous \\(L_1\\) frames. Applying these transformation matrices to a specific point in the current frame yields its rough coordinates in previous frames. Then we incorporate this flow feature into 3D PointNet++ [35] to encode a global spatial-temporal feature and use Multilayer Perceptron (MLP) to decode it into a 6D egocentric action. The loss function, denoted as \\(\\mathcal{L}_{action}\\), is computed as the L1 loss for aligning 3D points on the robot gripper as defined in [27]. We believe some sophisticated 4D backbones [17, 48] are suitable for 4D understanding, but they are often not suitable for robotic tasks that require a fast reference speed. Our method strikes a balance between effectiveness and simplicity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.701, + 0.892, + 0.807 + ], + "angle": 0, + "content": "To enhance the responsiveness of our policy to human motion and extend the vision horizon into the future, we introduce an auxiliary task to predict the future motion \\(\\{M_t^{t + 1}, M_t^{t + 2}, \\dots, M_t^{t + L_2}\\}\\) of objects in the next \\(L_{2}\\) frames. Using the ground truth object poses from trajectories, we compute the motion prediction loss for the \\(t\\)-th frame:" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.821, + 0.892, + 0.862 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p r e d} = \\sum_ {i = t + 1} ^ {t + L _ {2}} \\| \\hat {M} _ {t} ^ {i} - M _ {t} ^ {i} \\| \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In contrast to reinforcement learning, our imitation learning method requires only a few hours of training and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16366" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "achieves great generalizability through large-scale, high-quality demonstrations. We acquire vision-action pairs and ground truth object states from demonstrations, and then supervise our policy using the loss function \\(\\mathcal{L} = \\mathcal{L}_{action} + \\lambda \\mathcal{L}_{pred}\\), where \\(\\lambda\\) serves as a weighting hyper-parameter to balance the losses. This efficient distillation paradigm empowers our policy to naturally approach objects with a forecasting intention and to effectively generalize to a wide range of unseen objects and motions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.24, + 0.21, + 0.258 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.264, + 0.473, + 0.566 + ], + "angle": 0, + "content": "Dataset (1) HandoverSim [9] contains 1000 real-world H2R handover scenes and 20 objects from DexYCB [8]. We evaluate on the \"s0\" setup which contains 720 training and 144 testing scenes. Each handover motion has a duration of 3 seconds. Following the evaluation of Handover-Sim2real [11], we consider \"Sequential\" and \"Simultaneous\" settings. In \"s0 (Sequential)\", the robot is allowed to move when the hand reaches the handover location and remains static. In \"s0 (Simultaneous)\", the robot is allowed to move from the beginning of the episode. (2) GenH2R-Sim contains 1,000,000 complex synthetic H2R handover scenes and 3266 objects. We evaluate the \"t0\" setup which contains 1,000,000 training and 3260 testing scenes. Each handover motion has a duration of 8s and will stop when the robot gripper is close to the object. To introduce more real-world handover scenes into GenH2R-Sim for evaluation, we extract and clip the handover point cloud sequence from HOI4D [28], a real-world mocap dataset. This additional setup is referred to as \"t1\", which only contains 1000 testing scenes for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.567, + 0.473, + 0.734 + ], + "angle": 0, + "content": "Metrics We adhere to the HandoverSim evaluation protocol. A successful handover involves grasping the object from the human hand and moving it to a designated location. Failure cases involve hand contact, object drop, and timeout \\((T_{max} = 13s)\\). We report the successful rate and the execution time. Given that some policies prioritize success over speed, potentially wasting considerable human time, and others prioritize speed without considering success, we aim to evaluate both success rate and completion efficiency. To achieve this, we introduce AS (Average Success), akin to AP (Average Precision):" + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.739, + 0.47, + 0.775 + ], + "angle": 0, + "content": "\\[\n\\mathrm {A S} = \\int_ {0} ^ {1} \\operatorname {S u c c e s s} (t) \\mathrm {d} t \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.47, + 0.84 + ], + "angle": 0, + "content": "where \\( \\operatorname{Success}(t) \\) is success rate considering only successful cases within \\( t \\cdot T_{\\max} \\). This method can better evaluate success-time relations which is more suitable in our handover scenarios." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.398, + 0.865 + ], + "angle": 0, + "content": "4.1. Evaluating on Different Benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.87, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Setup We have 2 training sets: small-scale real-world \"s0\" from HandoverSim and large-scale synthetic \"t0\" from our" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.228 + ], + "angle": 0, + "content": "GenH2R-Sim. Evaluation is conducted on four testing sets: \"s0 (Sequential)\"/“s0 (Simultaneous)” from HandoverSim and \"t0\"/“t1” from our GenH2R-Sim. We conduct experiments on our forecast-aid 4D imitation learning from different demonstration strategies including destination planning, dense planning, and landmark planning. As discussed in Section 3.3, destination planning denotes the foresighted planner, dense planning denotes the improved shortsighted planner and landmark planning is our proposed method." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.228, + 0.892, + 0.289 + ], + "angle": 0, + "content": "Baselines We compare our methods with Handover-Sim2real*, the state-of-the-art method in HandoverSim. We additionally compare GA-DDPG which is designed for grasping objects, and OMG Planner." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.29, + 0.893, + 0.546 + ], + "angle": 0, + "content": "Results on different datasets As depicted in Table 1, our method trained on \"t0\" outperform all methods trained on \"s0\" by a large margin. Compared with Handover-Sim2real trained on \"s0\", our landmark planning method trained on \"t0\" exhibits \\(11.34\\%\\), \\(16.90\\%\\), \\(12.26\\%\\), and \\(15.93\\%\\) increase in the success rate across the four testing sets. Moreover, compared with our landmark planning method trained on \"s0\", the version trained on \"t0\" demonstrates notable improvements, achieving success rate increases of \\(8.79\\%\\), \\(6.48\\%\\), \\(11.80\\%\\), and \\(14.13\\%\\) increase in the same testing sets. This underscores the importance of having a substantial amount of synthetic data for handover training in simulation, which is more beneficial than only relying on a small-scale real-world dataset. Our GenH2R-Sim, with its large-scale complex human hand behavior, generalizes effectively to real-world scenarios such as \"s0\" in DexYCB and \"t1\" in HOI4D." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.547, + 0.894, + 0.804 + ], + "angle": 0, + "content": "Results for different methods We can compare our methods with the baseline HandoverSim2real within the same training set in different benchmarks. When trained on \"s0\", our landmark planning method demonstrates improvements of \\(2.55\\%\\), \\(10.42\\%\\), \\(0.46\\%\\), and \\(1.8\\%\\) (\\(13.43\\%\\), \\(53.48\\%\\), \\(1.07\\%\\), and \\(23.60\\%\\) in our reproduced version) across the 4 test sets. Similarly, When trained on \"t0\", our landmark planning method gives substantial improvements of \\(20.78\\%\\), \\(23.15\\%\\), \\(7.72\\%\\), and \\(21.23\\%\\) (\\(23.02\\%\\), \\(46.76\\%\\), \\(8.12\\%\\), and \\(34.98\\%\\) in our reproduced version). The last 3 benchmarks (\"s0\"(simultaneous), \"t0\", and \"t1\") closely resemble real-world scenarios. They greatly demonstrate the effectiveness of our pipeline from distillation-friendly demonstrations to forecast-aided 4D imitation learning, which is capable of handling dynamic robot perception in complex handover scenarios. We also show visualizations on different methods in Figure 4 (a)(b)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.498, + 0.816, + 0.894, + 0.901 + ], + "angle": 0, + "content": "*Our approach strictly adheres to the simultaneous setting defined in the paper of HandoverSim and HandoverSim2real: the robot moves from the beginning of the handover episode. However, it's noteworthy that HandoverSim2real manually makes their policy hold still in the first 1.5 seconds in the code implementation, deviating from the simultaneous setting definition. To ensure a fair comparison, we reproduce their results in the true simultaneous setting." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16367" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.078, + 0.077, + 0.917, + 0.293 + ], + "angle": 0, + "content": "
s0 (Sequential)s0 (Simultaneous)t0t1
STASSTASSTASSTAS
OMG Planner† [42]62.508.3122.5---------
train on s0GA-DDPG [43]50.007.1422.536.814.6623.623.597.3110.346.75.5026.9
Handover-Sim2real [11]75.237.7430.468.756.2335.829.176.2915.052.407.0923.8
Handover-Sim2real* [11]64.357.6126.725.695.4315.028.564.7317.930.605.9816.5
Destination Planning74.319.0122.876.166.9835.225.685.9614.148.48.9415.1
Dense Planning74.779.5419.875.457.3233.027.306.2614.152.39.2415.1
Landmark Planning77.789.2422.379.177.2634.929.636.2315.454.29.0216.6
train on t0GA-DDPG [43]54.767.2624.244.685.3026.524.054.7015.325.505.8614.1
Handover-Sim2real [11]65.977.1829.562.506.0433.533.715.9118.447.106.3524.1
Handover-Sim2real* [11]63.557.5826.538.895.2923.133.314.6421.433.355.8118.4
Destination Planning0.9312.800.016.4812.410.35.968.811.91.6012.030.1
Dense Planning81.489.5121.984.957.4536.338.047.1617.157.908.8518.4
Landmark Planning86.578.8128.085.656.5842.841.436.0122.368.337.7027.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.302, + 0.895, + 0.388 + ], + "angle": 0, + "content": "Table 1. Evaluating on different benchmarks. We compare our method against baselines from the test set of HandoverSim [9] benchmark (\"s0 (sequential)\" and \"s0 (simultaneous)\") and our GenH2R-Sim benchmark (\"t0\" and \"t1\"). We use the best-pretrained models from the repositories of GA-DDPG [43] and Handover-Sim2real [11] for evaluation. The results for our method are averaged across 3 random seeds. Note that S means success rate(%) . T means time(s). AS means average success(%) as defined in Equation 2. †: This method [42] is evaluated with ground-truth states and cannot handle dynamic handover like \"s0 (Simultaneous)\", \"t0\" and \"t1\".*: We reproduce the results of HandoverSim2real in the true simultaneous setting as detailed in Section 4.1 to make a fair comparison." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.4, + 0.473, + 0.719 + ], + "angle": 0, + "content": "Results for different Demonstrations Trained on \"s0\" which consists of relatively simple trajectories, demonstrations based on destination planning can offer a rudimentary cue for downstream visuo-motor policy. However, when trained on \"t0\" this strategy may lose focus on the object, leading to a failure in maintaining vision-action correlation and providing minimal gains for vision-friendly learning. There is a significant \\(73.38\\%\\) / \\(69.68\\%\\) decrease in success rate in the \"s0\" setting. Additionally, distillation from landmark planning slightly outperforms dense planning in success rate and completes the handover process more quickly in all benchmarks. While dense planning can sustain the success rate to some extent, it slows down the agent and may result in unnatural approaches to objects. To jointly consider the time efficiency and the success rate, we compare the Average Success in methods distilled from these two strategies and find that landmark planning is a more efficient and generalizable approach. For instance, when trained on \"t0\", landmark planning exhibits significant improvements of \\(6.1\\%\\), \\(6.5\\%\\), \\(5.2\\%\\), and \\(9.5\\%\\) across the four testing sets." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.727, + 0.407, + 0.743 + ], + "angle": 0, + "content": "4.2. Evaluating on different Dataset Scales" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.901 + ], + "angle": 0, + "content": "We have proved the crucial role of large-scale datasets in handover generalization in Section 4.1. We can also reveal it by scaling down the usage of \"t0\" in GenH2R-Sim which contains 1,000,000 training scenes. With \\(10\\%\\) data utilization, we observe a \\(5.93\\%\\) drop in the success rate on the unseen \"t1\" test set. This result proves the significance of the dataset scale in our imitation learning method. Thanks to our large-scale data and efficient demonstration generation pipeline, concerns about limited datasets hindering generalization are alleviated." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.398, + 0.895, + 0.472 + ], + "angle": 0, + "content": "
MethodsSTAS
w/o Flow31.665.6717.9
w/o Prediction39.186.1120.7
w/o Flow & Prediction37.045.9320.1
Ours41.436.0122.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.48, + 0.895, + 0.523 + ], + "angle": 0, + "content": "Table 2. Ablations on different modules. \"w/o Flow\" means do not use flow information in the input. \"w/o Prediction\" means do not add prediction loss in the output." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.538, + 0.655, + 0.556 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.562, + 0.895, + 0.73 + ], + "angle": 0, + "content": "As shown in Table 2, we prove the effectiveness of our well-designed 4D imitation learning method. The absence of flow information results in a \\(9.77\\%\\) decrease (predicting without past information adversely affects the model performance). The absence of the prediction task leads to a \\(2.25\\%\\) decrease, and the absence of both components results in a \\(4.39\\%\\) decrease. The results demonstrate the model obtains improved performance in leveraging flow information, particularly when tasked with predicting the future object pose. More ablations about our demonstration generation and imitation learning are detailed in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.74, + 0.729, + 0.758 + ], + "angle": 0, + "content": "4.4. Real World Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.894, + 0.87 + ], + "angle": 0, + "content": "Sim-to-Real Transfer In addition to simulation, we deploy the models trained in GenH2R-Sim on a real robotic platform. Using point cloud input from the wrist-mounted camera, we employ the output 6D egocentric action to update the end effector's target position. A user study compares our method against Handover-Sim2real [11]. The supplementary material provides further details." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.894, + 0.903 + ], + "angle": 0, + "content": "User Study We recruited 6 users to compare our method (based on landmark planning) and Handover-Sim2real" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16368" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.805, + 0.658 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.105, + 0.662, + 0.863, + 0.677 + ], + "angle": 0, + "content": "Figure 4. Qualitative results. We in detail compare different methods in simulators and deploy them in the real-world platform." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.687, + 0.472, + 0.734 + ], + "angle": 0, + "content": "
MethodsSimple SettingComplex Setting
Handover-Sim2real56.7%33.3%
Ours90.0%70.0%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.741, + 0.471, + 0.77 + ], + "angle": 0, + "content": "Table 3. Sim-to-Real Experiments. We report the success rate of our method and HandoverSim2real in 2 different settings." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.901 + ], + "angle": 0, + "content": "across 5 objects in 2 different settings. In the simple setting, users hand each object to the gripper without quick movements. In the complex setting, users execute a relatively long and quick trajectory. The results are reported in Table 3. We observe that our model gets better performance in completing the handover process across various objects and scenarios. Figure 4(c) shows examples of the real-world handover trials." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.689, + 0.619, + 0.705 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this work, we present a novel framework GenH2R for scaling up the learning of human-to-robot handover. We introduce a new simulator GenH2R-Sim and generate a million human handover animations to facilitate generalizable H2R handover learning. We then propose a distillation-friendly demonstration generation method that automatically produces a million high-quality demonstrations suitable for learning. We further introduce a forecast-aided 4D imitation learning method for effective demonstration distillation. Our experiments demonstrate that scaling-up efforts result in substantial improvement of generalizability to novel geometry and complex motion, both in the simulator and the real world." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16369" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.122, + 0.174, + 0.137 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.147, + 0.47, + 0.203 + ], + "angle": 0, + "content": "[1] Christoph Bartneck, Tony Belpaeme, Friederike Eyssel, Takayuki Kanda, Merel Keijrsers, and Selma Šabanović. Human-robot interaction: An introduction. Cambridge University Press, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.205, + 0.472, + 0.274 + ], + "angle": 0, + "content": "[2] Antonio Bicchi and Vijay Kumar. Robotic grasping and contact: A review. In Proceedings 2000 ICRA. Millennium conference. IEEE international conference on robotics and automation. Symposia proceedings (Cat. No. 00CH37065), pages 348-353. IEEE, 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.276, + 0.47, + 0.317 + ], + "angle": 0, + "content": "[3] Aude Billard, Sylvain Calinon, Ruediger Dillmann, and Stefan Schaal. Survey: Robot programming by demonstration. Technical report, Springer, 2008. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.32, + 0.47, + 0.36 + ], + "angle": 0, + "content": "[4] Jeannette Bohg, Antonio Morales, Tamim Asfour, and Dania Kragic. Data-driven grasp synthesis—a survey. IEEE Transactions on robotics, 30(2):289-309, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.363, + 0.47, + 0.446 + ], + "angle": 0, + "content": "[5] Samarth Brahmbhatt, Chengcheng Tang, Christopher D Twigg, Charles C Kemp, and James Hays. Contactpose: A dataset of grasps with object contact and hand pose. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16, pages 361-378. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.448, + 0.47, + 0.517 + ], + "angle": 0, + "content": "[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.519, + 0.47, + 0.587 + ], + "angle": 0, + "content": "[7] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.591, + 0.47, + 0.686 + ], + "angle": 0, + "content": "[8] Yu-Wei Chao, Wei Yang, Yu Xiang, Pavlo Molchanov, Ankur Handa, Jonathan Tremblay, Yashraj S Narang, Karl Van Wyk, Umar Iqbal, Stan Birchfield, et al. Dexycb: A benchmark for capturing hand grasping of objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9044–9053, 2021. 2, 3, 6, 14, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.689, + 0.47, + 0.785 + ], + "angle": 0, + "content": "[9] Yu-Wei Chao, Chris Paxton, Yu Xiang, Wei Yang, Balakumar Sundaralingam, Tao Chen, Adithyavairavan Murali, Maya Cakmak, and Dieter Fox. Handoversim: A simulation framework and benchmark for human-to-robot object handovers. In 2022 International Conference on Robotics and Automation (ICRA), pages 6941–6947. IEEE, 2022. 1, 2, 3, 6, 7, 12, 14, 16, 17, 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[10] Sammy Christen, Lan Feng, Wei Yang, Yu-Wei Chao, Otmar Hilliges, and Jie Song. Synh2r: Synthesizing handobject motions for learning human-to-robot handovers. arXiv preprint arXiv:2311.05599, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[11] Sammy Christen, Wei Yang, Claudia Pérez-D'Arpino, Otmar Hilliges, Dieter Fox, and Yu-Wei Chao. Learning human-to-robot handovers from point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.147, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "Pattern Recognition, pages 9654-9664, 2023. 2, 5, 6, 7, 13, 14, 16, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[12] Gianluca Corsini, Martin Jacquet, Hemjyoti Das, Amr Affi, Daniel Sidobre, and Antonio Franchi. Nonlinear model predictive control for human-robot handover with application to the aerial case. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7597–7604. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[13] Murtaza Dalal, Ajay Mandlekar, Caelan Garrett, Ankur Handa, Ruslan Salakhutdinov, and Dieter Fox. Imitating task and motion planning with visuomotor transformers. arXiv preprint arXiv:2305.16309, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.264, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[14] Yuhao Dong, Zhuoyang Zhang, Yunze Liu, and Li Yi. Nsm4d: Neural scene model based online 4d point cloud sequence understanding. arXiv preprint arXiv:2310.08326, 2023.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.321, + 0.892, + 0.39 + ], + "angle": 0, + "content": "[15] Frederik Ebert, Yanlai Yang, Karl Schmeckpeper, Bernadette Bucher, Georgios Georgakis, Kostas Daniilidis, Chelsea Finn, and Sergey Levine. Bridge data: Boosting generalization of robotic skills with cross-domain datasets. arXiv preprint arXiv:2109.13396, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.392, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[16] Clemens Eppner, Arsalan Mousavian, and Dieter Fox. Acronym: A large-scale grasp dataset based on simulation. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 6222-6227. IEEE, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.449, + 0.892, + 0.516 + ], + "angle": 0, + "content": "[17] Hehe Fan, Yi Yang, and Mohan Kankanhalli. Point 4d transformer networks for spatio-temporal modeling in point cloud videos. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14204-14213, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.519, + 0.892, + 0.601 + ], + "angle": 0, + "content": "[18] Zicong Fan, Omid Taheri, Dimitrios Tzionas, Muhammed Kocabas, Manuel Kaufmann, Michael J Black, and Otmar Hilliges. Arctic: A dataset for dexterous bimanual handobject manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12943-12954, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.604, + 0.892, + 0.671 + ], + "angle": 0, + "content": "[19] Hao-Shu Fang, Chenxi Wang, Hongjie Fang, Minghao Gou, Jirong Liu, Hengxu Yan, Wenhai Liu, Yichen Xie, and Cewu Lu. Anygrasp: Robust and efficient grasp perception in spatial and temporal domains. IEEE Transactions on Robotics, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.675, + 0.892, + 0.729 + ], + "angle": 0, + "content": "[20] Chelsea Finn, Tianhe Yu, Tianhao Zhang, Pieter Abbeel, and Sergey Levine. One-shot visual imitation learning via meta-learning. In Conference on robot learning, pages 357-368. PMLR, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.732, + 0.892, + 0.785 + ], + "angle": 0, + "content": "[21] Justin Fu, Aviral Kumar, Ofir Nachum, George Tucker, and Sergey Levine. D4rl: Datasets for deep data-driven reinforcement learning. arXiv preprint arXiv:2004.07219, 2020.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.788, + 0.892, + 0.856 + ], + "angle": 0, + "content": "[22] Caelan Reed Garrett, Rohan Chitnis, Rachel Holladay, Beomjoon Kim, Tom Silver, Leslie Pack Kaelbling, and Tomás Lozano-Pérez. Integrated task and motion planning. Annual review of control, robotics, and autonomous systems, 4:265-293, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[23] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d:" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16370" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.177 + ], + "angle": 0, + "content": "[24] Huy Ha, Pete Florence, and Shuran Song. Scaling up and distilling down: Language-guided robot skill acquisition. arXiv preprint arXiv:2307.14535, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.26 + ], + "angle": 0, + "content": "[25] Yana Hasson, Gul Varol, Dimitrios Tzionas, Igor Kalevatykh, Michael J Black, Ivan Laptev, and Cordelia Schmid. Learning joint reconstruction of hands and manipulated objects. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11807-11816, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.469, + 0.331 + ], + "angle": 0, + "content": "[26] Eric Jang, Alex Irpan, Mohi Khansari, Daniel Kappler, Frederik Ebert, Corey Lynch, Sergey Levine, and Chelsea Finn. Bc-z: Zero-shot task generalization with robotic imitation learning. In Conference on Robot Learning, pages 991-1002. PMLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.469, + 0.389 + ], + "angle": 0, + "content": "[27] Yi Li, Gu Wang, Xiangyang Ji, Yu Xiang, and Dieter Fox. Deepim: Deep iterative matching for 6d pose estimation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 683-698, 2018. 5, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.391, + 0.469, + 0.473 + ], + "angle": 0, + "content": "[28] Yunze Liu, Yun Liu, Che Jiang, Kangbo Lyu, Weikang Wan, Hao Shen, Boqiang Liang, Zhoujie Fu, He Wang, and Li Yi. Hoi4d: A 4d egocentric dataset for category-level human-object interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21013-21022, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.476, + 0.469, + 0.544 + ], + "angle": 0, + "content": "[29] Ajay Mandlekar, Danfei Xu, Josiah Wong, Soroush Nasiriy, Chen Wang, Rohun Kulkarni, Li Fei-Fei, Silvio Savarese, Yuke Zhu, and Roberto Martin-Martin. What matters in learning from offline human demonstrations for robot manipulation. arXiv preprint arXiv:2108.03298, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.547, + 0.469, + 0.614 + ], + "angle": 0, + "content": "[30] Naresh Marturi, Marek Kopicki, Alireza Rastegarpanah, Vijaykumar Rajasekaran, Maxime Adjigble, Rustam Stolkin, Ales Leonidis, and Yasemin Bekiroglu. Dynamic grasp and trajectory planning for moving objects. Autonomous Robots, 43:1241-1256, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.617, + 0.469, + 0.657 + ], + "angle": 0, + "content": "[31] Michael James McDonald and Dylan Hadfield-Menell. Guided imitation of task and motion planning. In Conference on Robot Learning, pages 630-640. PMLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.352, + 0.674 + ], + "angle": 0, + "content": "[32] OpenAI. Gpt-4 technical report, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.676, + 0.469, + 0.729 + ], + "angle": 0, + "content": "[33] Valerio Ortenzi, Akansel Cosgun, Tommaso Pardi, Wesley P Chan, Elizabeth Croft, and Dana Kulić. Object handovers: a review for robotics. IEEE Transactions on Robotics, 37(6): 1855-1873, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.469, + 0.772 + ], + "angle": 0, + "content": "[34] Dean A Pomerleau. Alvinn: An autonomous land vehicle in a neural network. Advances in neural information processing systems, 1, 1988. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.469, + 0.83 + ], + "angle": 0, + "content": "[35] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 5, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[36] Patrick Rosenberger, Akansel Cosgun, Rhys Newbury, Jun Kwan, Valerio Ortenzi, Peter Corke, and Manfred Grafinger. Object-independent human-to-robot handovers using real time robotic vision. IEEE Robotics and Automation Letters, 6(1):17-23, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[37] Szymon Rusinkiewicz and Marc Levoy. Efficient variants of the icp algorithm. In Proceedings third international conference on 3-D digital imaging and modeling, pages 145-152. IEEE, 2001. 5, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.151, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[38] Thomas B Sheridan. Human–robot interaction: status and challenges. Human factors, 58(4):525-532, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[39] Shuran Song, Andy Zeng, Johnny Lee, and Thomas Funkhouser. Grasping in the wild: Learning 6dof closed-loop grasping from low-cost demonstrations. IEEE Robotics and Automation Letters, 5(3):4978-4985, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.236, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[40] Zachary Teed and Jia Deng. Raft-3d: Scene flow using rigid-motion embeddings. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8375–8384, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.293, + 0.892, + 0.333 + ], + "angle": 0, + "content": "[41] Lirui Wang, Yu Xiang, and Dieter Fox. Manipulation trajectory optimization with online grasp synthesis and selection. In Robotics: Science and Systems (RSS), 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.336, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[42] Lirui Wang, Yu Xiang, and Dieter Fox. Manipulation trajectory optimization with online grasp synthesis and selection. In Proceedings of Robotics: Science and Systems (RSS), 2020. 5, 7, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.393, + 0.892, + 0.447 + ], + "angle": 0, + "content": "[43] Lirui Wang, Yu Xiang, Wei Yang, Arsalan Mousavian, and Dieter Fox. Goal-auxiliary actor-critic for 6d robotic grasping with point clouds. In Conference on Robot Learning, pages 70-80. PMLR, 2022. 5, 7, 14, 16, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.45, + 0.892, + 0.516 + ], + "angle": 0, + "content": "[44] Lirui Wang, Yiyang Ling, Zhecheng Yuan, Mohit Shridhar, Chen Bao, Yuzhe Qin, Bailin Wang, Huazhe Xu, and Xiaolong Wang. Gensim: Generating robotic simulation tasks via large language models. arXiv preprint arXiv:2310.01361, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.52, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[45] Ruicheng Wang, Jialiang Zhang, Jiayi Chen, Yinzhen Xu, Puhao Li, Tengyu Liu, and He Wang. Dexgraspnet: A large-scale robotic dexterous grasp dataset for general objects based on simulation. arXiv preprint arXiv:2210.02697, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.591, + 0.892, + 0.672 + ], + "angle": 0, + "content": "[46] Ruicheng Wang, Jialiang Zhang, Jiayi Chen, Yinzhen Xu, Puhao Li, Tengyu Liu, and He Wang. Dexgraspnet: A large-scale robotic dexterous grasp dataset for general objects based on simulation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 11359-11366. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.676, + 0.892, + 0.743 + ], + "angle": 0, + "content": "[47] Yufei Wang, Zhou Xian, Feng Chen, Tsun-Hsuan Wang, Yian Wang, Katerina Fragkiadaki, Zackory Erickson, David Held, and Chuang Gan. Robogen: Towards unleashing infinite data for automated robot learning via generative simulation. arXiv preprint arXiv:2311.01455, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.746, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[48] Hao Wen, Yunze Liu, Jingwei Huang, Bo Duan, and Li Yi. Point primitive transformer for long-term 4d point cloud video understanding. In European Conference on Computer Vision, pages 19-35. Springer, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.803, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[49] Wei Yang, Chris Paxton, Arsalan Mousavian, Yu-Wei Chao, Maya Cakmak, and Dieter Fox. Reactive human-to-robot handovers of arbitrary objects. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 3118–3124. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[50] Ruolin Ye, Wenqiang Xu, Zhendong Xue, Tutian Tang, Yanfeng Wang, and Cewu Lu. H2o: A benchmark for visual" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "16371" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "human-human object handover analysis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15762-15771, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.189 + ], + "angle": 0, + "content": "[51] Gu Zhang, Hao-Shu Fang, Hongjie Fang, and Cewu Lu. Flexible handover with real-time robust dynamic grasp trajectory generation. arXiv preprint arXiv:2308.15622, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.26 + ], + "angle": 0, + "content": "[52] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.469, + 0.344 + ], + "angle": 0, + "content": "[53] Tianhao Zhang, Zoe McCarthy, Owen Jow, Dennis Lee, Xi Chen, Ken Goldberg, and Pieter Abbeel. Deep imitation learning for complex manipulation tasks from virtual reality teleoperation. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pages 5628-5635. IEEE, 2018. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "16372" + } + ] +] \ No newline at end of file diff --git a/2024/GenH2R_ Learning Generalizable Human-to-Robot Handover via Scalable Simulation Demonstration and Imitation/40314ff6-3af8-45b8-b8fc-6f9ebdc7d993_origin.pdf b/2024/GenH2R_ Learning Generalizable Human-to-Robot Handover via Scalable Simulation Demonstration and Imitation/40314ff6-3af8-45b8-b8fc-6f9ebdc7d993_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..31055e6b09be64e0868579f55a2fe76ff75f5eec --- /dev/null +++ b/2024/GenH2R_ Learning Generalizable Human-to-Robot Handover via Scalable Simulation Demonstration and Imitation/40314ff6-3af8-45b8-b8fc-6f9ebdc7d993_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b8068563ef3f44d02aec8b79cf920845db63cede131b5134ef51d31e2e51f8dc +size 2548588 diff --git a/2024/GenH2R_ Learning Generalizable Human-to-Robot Handover via Scalable Simulation Demonstration and Imitation/full.md b/2024/GenH2R_ Learning Generalizable Human-to-Robot Handover via Scalable Simulation Demonstration and Imitation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0bed2f2fc6500bdc84af2763a4837eb529181892 --- /dev/null +++ b/2024/GenH2R_ Learning Generalizable Human-to-Robot Handover via Scalable Simulation Demonstration and Imitation/full.md @@ -0,0 +1,252 @@ +# GenH2R: Learning Generalizable Human-to-Robot Handover via Scalable Simulation, Demonstration, and Imitation + +Zifan Wang $^{1,3}$ Junyu Chen $^{1,3}$ Ziqing Chen $^{1}$ Pengwei Xie $^{1}$ Rui Chen $^{1}$ Li Yi $^{\dagger 1,2,3}$ + +$^{1}$ Tsinghua University $^{2}$ Shanghai Artificial Intelligence Laboratory $^{3}$ Shanghai Qi Zhi Institute https://GenH2R.github.io + +# Abstract + +This paper presents GenH2R, a framework for learning generalizable vision-based human-to-robot (H2R) handover skills. The goal is to equip robots with the ability to reliably receive objects with unseen geometry handed over by humans in various complex trajectories. We acquire such generalizability by learning H2R handover at scale with a comprehensive solution including procedural simulation assets creation, automated demonstration generation, and effective imitation learning. We leverage large-scale 3D model repositories, dexterous grasp generation methods, and curve-based 3D animation to create an H2R handover simulation environment named GenH2R-Sim, surpassing the number of scenes in existing simulators by three orders of magnitude. We further introduce a distillation-friendly demonstration generation method that automatically generates a million high-quality demonstrations suitable for learning. Finally, we present a 4D imitation learning method augmented by a future forecasting objective to distill demonstrations into a visuo-motor handover policy. Experimental evaluations in both simulators and the real world demonstrate significant improvements (at least +10% success rate) over baselines in all cases. + +# 1. Introduction + +The embodied AI research community has long been driven by the goal of empowering robots to interact and collaborate with humans. A crucial aspect of this pursuit is equipping robots with the capability to reliably receive arbitrarily moving objects of varying geometry handed over by humans, based on dynamic visual observations. This human-to-robot (H2R) handover ability allows robots to seamlessly collaborate with humans across a wide range of tasks, including cooking, room tidying, and furniture assembly. + +However, compared to learning human-free robot manip- + +![](images/c7a626db76ce77f60bec918fae38f7655a0f03c08f3c374596a4a8ea69b1bb4a.jpg) +Scalable Synthetic Handover Simulator + +![](images/b65b3f3cc6277337bf282f44a073d65fbd1a0a2161d60153dc7902c157040165.jpg) +Distillation-friendly Expert Demonstrations + +![](images/8365645783823546867e57ab810053920af0517be16c1c1c7e03db65bf22b0c2.jpg) +Sim-to-real Deployment + +![](images/a65c3f3399fb73c24d783dffda82d53ffa58cbb47c7748061f6c5afbaefc24bf.jpg) +Forecast-Aided 4D Imitation Learning +Figure 1. The overview of GenH2R. We introduce a framework for learning generalizable vision-based human-to-robot handover via scalable synthetic simulation, distillation-friendly expert demonstration generation, and a forecast-aided 4D imitation learning method. Our models demonstrate strong generalization capabilities to real datasets and can be deployed to a real robot. + +ulation skills, the progress in scalably learning H2R handover that can generalize to various objects and versatile human behaviors has lagged due to its unique challenges. Training robots to interact with humans in real-world scenarios entails increased risks and expenses, rendering it inherently non-scalable. Therefore, it is demanded to simulate human behaviors and train robots in simulated environments prior to real-world deployment. However, creating a substantial number of assets for humans handing over objects poses a significant challenge. In a recent study [9] that employed motion capturing to drive virtual humans in a simulator, only 1000 unique human hand motion trajectories were provided for handing over 20 objects. Limited object geometry and human motion assets can hardly capture the complexities of the real world. Besides, the challenge extends to the demonstration side. The success of large language model [6, 32, 52] has suggested a recipe for scaling up learning through modeling large-scale training + +data. Nevertheless, collecting robot demonstrations receiving objects from real humans is very costly and unscalable. How to scale up the number of demonstrations while ensuring effective learning poses additional challenges. + +In this work, we aim to learn generalizable H2R handover at scale by tackling the above challenges. We present a comprehensive solution that scales up both the assets and demonstrations and effectively learns a closed-loop visuomotor policy through a novel imitation learning algorithm. + +Specifically, to scale up geometry and motion assets depicting humans handing over various objects, we leverage large-scale 3D model repositories [7, 16], dexterous grasp generation methods [46], and curve-based 3D animation. This enables us to procedurally generate millions of handover scenes, forming an environment named GenH2R-Sim to support generalizable H2R handover learning. GenH2R-Sim surpasses HandoverSim [9], an existing H2R simulator, in both scene quantity (by three orders of magnitude) and unique object involvement (by two orders of magnitude). In addition, scenes in GenH2R-Sim go beyond a straightforward giving and then receiving and cover cases when humans might keep transforming the object in a large range during the entire H2R handover process. This allows for studying complex behaviors such as humans hesitating before handing over. + +To scale up robot demonstrations, we draw inspiration from the Task and Motion Planning (TAMP) [22] literature and propose to automatically generate demonstrations with grasp and motion planning using privileged human motion and object state information. There are some straightforward ways to achieve this goal, such as using the privileged human handover destination information to plan a smooth demonstration. However, the problem is more challenging than it seems since the generated demonstrations need to be suitable for distilling into a visuomotor policy. We identify the vision-action correlation between visual observations and planned actions as the crucial factor influencing distillability and point out that due to the constraints of robot arm morphology one can easily generate observation-irrelevant actions and thus harm distillation. To tackle this challenge, we present a distillation-friendly demonstration generation method that sparsely samples handover animations for landmark states and periodically replans grasp and motion based on privileged future landmarks. + +Finally, to distill the above demonstrations into a visuomotor policy, we utilize point cloud input for its richer geometric information and smaller sim-vs-real gap compared to images. We propose a 4D imitation learning method that factors the sequential point cloud observations into geometry and motion parts, facilitating policy learning by better revealing the current scene state. Furthermore, the imitation objective is augmented by a forecasting objective which predicts the future motion of the handover object. Since our + +demonstrating actions are generated based on future landmarks, the forecasting objective can help further exploit the vision-action correlation. + +We evaluate our learned policy in simulators (Handover-Sim and our own GenH2R-Sim) and the real world. Remarkably, without any mocap assets or real-world demonstrations, our method achieves significantly better performance compared to baselines across all settings (at least $+10\%$ success rate). Our experiments highlight that the scaling-up efforts bring substantial improvement in policy generalizability to novel geometry and complex motion. Furthermore, these efforts greatly facilitate skill transfer to real robotic systems. + +In summary, the key contribution of this paper is a novel framework scaling up the learning of H2R handover with the following three components: i) a simulation environment named GenH2R-Sim consists of millions of human handover animations for generalizable H2R handover learning, ii) an empirically validated automatic robot demonstration generation pipeline for vision-based closed-loop control, iii) a forecast-aided 4D imitation learning method effective in distilling the large-scale demonstrations. + +# 2. Related Work + +# 2.1. Human-to-Robot Handovers + +Recently, significant progress in human-robot handovers [12, 33, 36] has been observed, driven by the increasing popularity of human-robot interaction [1, 38] and the emergence of extensive datasets [5, 8, 18, 25, 28, 50] capturing hand-object interactions. Some traditional methods [2, 4] require 3D object models and struggle to handle unseen objects. One possible way is to consider grasping and dynamic motion planning [19, 30, 49, 51]. However, these methods often exhibit constrained motions and perform poorly on large-scale datasets. HandoverSim [9], a physics-simulated environment, introduced a new simulation benchmark for human-to-robot object handovers. Leveraging DexYCB [8], a dataset of human grasping objects and performing handover attempts, this environment allows training learning-based handover policies such as [11]. However, it lacks large-scale and diverse handover scenes, which limits generalizable handovers. At the same time, SynH2R [10] proposes to use synthetic data but makes limited progress. Building on this, we propose GenH2R-Sim , aiming to benchmark generalizable handover. + +# 2.2. Scaling Up Robot Demonstrations + +For robot learning, scaling up data collection for manipulation skills has spurred extensive research. Approaches include leveraging large language models [24] or hardware + +![](images/8d67f0d7bca143f2d5d11b877b11361c982ad3fea0621b6abb9be4e22168ebfd.jpg) +Figure 2. The overview of our framework. First, we propose a new simulation environment named GenH2R-Sim, featuring large-scale synthetic datasets with diversity in object geometry, grasp poses, and complex trajectories. Second, other than destination planning (move straight toward the final position) and dense planning (replan at each step), we propose a distillation-friendly demonstration generation method—landmark planning, predicting landmarks on the trajectory (as indicated by the dashed object above) and replanning based on those landmarks. Thirdly, our Forecast-aided 4D Imitation Learning leverages past flow information, and the forecasting objective enhances the exploitation of vision-action correlation. + +![](images/19062cc6078a2249cc73be71dfa6bb7489a75942007b3e060b4eeed8dfd8e240.jpg) + +![](images/05fec30c5092a9c1b790bc8952ca4c628d26fc02d4f4779e92ec95fdb4f60b01.jpg) + +capabilities [39], utilizing non-robotics datasets [23], and employing trial-and-error explorations [21]. As depicted in [24], one of the challenges is scaling up robot-complete data. A popular line of research scales up demonstration generation via Task and Motion Planning [13, 22, 31]. These works usually focus on fairly static scenes without active motion or object and task variety [44, 47] while our method extends to dynamic H2R handover by considering how to interpret human behavior and generate demonstrations easy to be distilled by closed-loop visuo-motor policy. + +# 2.3. Offline Learning from Demonstrations + +Imitation Learning (IL) represents a methodology for training embodied agents in manipulation tasks by utilizing expert demonstrations. The commonly used Behavior Cloning (BC) [34] strategy directly trains the policy to imitate expert actions in a supervised learning manner. Despite its simplicity, this approach has demonstrated remarkable effectiveness in robotic manipulation [3, 20, 29, 53] especially when combined with a substantial number of high-quality demonstrations [15, 26]. Inspired by these works, we adopt an imitation learning paradigm, focusing on how to leverage spatial-temporal perception and future forecasting to better consume our distillation-friendly demonstrations. + +# 3. Method + +# 3.1. Overview + +For the generalizable H2R handover task, we introduce GenH2R, a framework designed to learn control policies, + +specifically 6D control actions for the robot gripper, using segmented point cloud data captured from an egocentric camera. We describe our method for synthesizing human handover animations in Section 3.2, generating expert demonstrations in Section 3.3, and distilling demonstrations to 4D vision-based neural networks by imitation learning in Section 3.4, as the pipeline depicted in Figure 2. + +# 3.2. GenH2R-Sim + +The size and quality of human-object datasets in simulators play a crucial role in generating high-quality handover demonstrations and training reliable policies for handover scenarios. The recent handover simulator, HandoverSim [9], utilizes the DexYCB [8] dataset, which captures real-world human grasping objects in a limited manner, comprising only 1000 scenes with 20 distinct objects. In the real world, scenarios can be more complex and may involve intricate trajectories and poses beyond those in DexYCB. + +To address these limitations, we introduce a new environment, GenH2R-Sim, to overcome these deficiencies and facilitate generalizable handovers. To diversify geometry and motion assets depicting humans handing over various objects, we focus on two primary aspects: the hand grasping pose and the hand-object moving trajectory within a scene. + +In aspects of grasping poses, DexGraspNet [45] has made significant contributions by employing optimization techniques to generate a substantial dataset of human hand grasp poses. We utilize this method to generate approximately 1,000,000 grasp poses for 3,266 different objects sourced from Shapenet [7]. These objects span a wide range + +of categories, from larger items like computers to smaller ones like mobile phones, covering most sizes and shapes encountered in real-life handovers. + +In aspects of hand-object moving trajectories, we propose to use Bézier curves, which are one class of smooth curves determined by several control points, to generate complex yet smooth-transiting motion trajectories. We use multiple Bézier curves to model different stages of the motion, and link the ends of these curves to create a seamless track. We can generate scenes matching various scenarios of different complexity in the real world by adjusting the distribution of control points of the trajectory and the speed of the human hand. To enhance the trajectory's realism, we incorporate consistent object rotations, which also enhances the importance of choosing the appropriate grasp for the robotic arm. Since we can always attach a new segment of motion at the end of the current motion and the duration is much longer than DexYCB scenes, the destination of the hand-object is not a significant factor, so we just randomly select a point within the reach of the robotic arm. + +We do not guarantee that every item in the dataset we generate perfectly mimics the human-like characteristics of real-world data, but our approach ensures a significantly higher degree of domain randomization and provides greater diversity in terms of geometry and motion. Given the challenges in scaling up real-world motion capture datasets, we opt for a large-scale synthetic dataset for our handover simulations. Our key insight is that for both demonstrations and policy learning, having a substantial amount of synthetic data is more beneficial than relying on a small-scale real-world dataset. + +GenH2R-Sim follows the setup of HandoverSim, which consists of a Panda 7DoF robotic arm with a gripper and a wrist-mounted RGB-D camera, and a simulated human hand. Just like HandoverSim, we switch from the prehandover kinematic phase to the handover dynamic phase when the object has been in contact with the gripper. HandoverSim is not adaptive to the robot's action and just loads and replays every frame of the data. To align with the real-world handover process more naturally in GenH2R-Sim, the simulated hand will stop from moving and wait for handover when the robot arm is close to the object. + +# 3.3. Generating Demonstrations for Distillation + +In this section, we address a key question in learning visuomotor policy: how to efficiently generate robot demonstrations that incorporate paired vision-action data from successful task experiences. While distilling successful demonstrations into a single policy has proven effective for open-loop control tasks, the challenge lies in closed-loop visuo-motor control, where the quality of demonstrations becomes crucial for learning. Merely ensuring success is no longer sufficient. We present two examples of demon- + +![](images/c9d34d63fb5e2bdd134e280734397e7af3be6a5d1c059e612b63a981738d71ff.jpg) +Figure 3. Different demonstration generation methods for dynamic handover. The orange curve shows the hand-object trajectory. The blue, red, and green curves show the example trajectories generated by the foresighted planner, the shortsighted planner, and our planner, respectively. + +stration generation with different grasp and motion planning strategies as shown in Figure 3. In the first example, a foresighted planner generates smooth, short demonstrations based on the privileged destination end state of a human handover animation. Though efficient, the planned path does not align actions with the dynamic visual observations during the handover. Distilling such demonstrations requires accurately forecasting the end state of the human trajectory, which can be extremely challenging in complex handover cases. The second example involves a shortsighted planner that independently replans grasp and motion at each time step using privileged hand and object states. Due to robot morphology constraints and the multisolution nature of common robot planners, smooth visual observations may correspond to unsmooth and multi-modal robot trajectories, increasing the difficulty of distillation. We emphasize the importance of distillability as a quality factor for handover demonstrations. An effective demonstration generation method must consider the vision-action correlation by jointly incorporating robot morphology and dynamic vision during grasp and motion planning. + +Along this line, we base our method on the foresighted and shortsighted planner mentioned above to combine the advantages of both sides while encouraging the demonstration distillability. We first improve the shortsighted planner so that sequentially smooth visual observations result in smooth grasp and motion plans. Then we improve the handover efficiency by looking toward the future while guaranteeing the vision-action correlation. + +To be specific, we build our method based on the OMG planner [41] for grasp and motion planning. This planner optimizes the grasp and motion path by considering the object's 6D pose and a set of candidate grasp poses. To support + +this optimization, we provide privileged knowledge that includes the object's 6D pose, candidate grasps generated through physics simulation [16], and human hand poses for filtering out invalid grasps. However, independently calling the OMG planner for each time step may result in unsmooth trajectories, as it is designed for static scenarios. To address this, we sequentially plan the grasp and motion based on the privileged knowledge by: 1) sorting grasps based on their pose distance to the robot end effector and attempting inverse kinematics (IK) starting from the nearest grasp until success; 2) initializing IK based on the robot arm pose from the previous time step; 3) invoking the OMG planner only when IK can be successfully solved. By prioritizing closer grasps, we encourage the object to remain within the field of view of a wrist camera, reducing visually irrelevant actions when the object is not visible. Additionally, enforcing IK smoothness improves the overall trajectory smoothness. As a result, the enhanced vision-action correlation dramatically improves the demonstration quality. + +Our approach modifies the OMG planner for dynamic grasp and motion planning. However, densely replanning at each time step leads to inefficient and non-smooth zigzag demonstrations, which does not align with how humans receive objects. Humans anticipate dynamic scene changes before taking action. On the other hand, a highly fore-sighted planner that directly plans grasp and motion based on the end state of a human handover animation can disrupt the vision-action correlation. To strike a balance between these extremes, we propose an algorithm that sparsely samples handover animations for landmark states and periodically replans grasp and motion based on future landmarks. The key idea is to select landmarks strategically so that the planner only considers visually foreseeable futures. Specifically, let $\xi = (\mathcal{T}_0,\mathcal{T}_1,\dots ,\mathcal{T}_{T - 1})$ represent an object trajectory, where $\mathcal{T}_t\in \mathbb{SE}(3)$ denotes the object pose in the $t$ -th frame within the world coordinate system. Based on all the object trajectories in the training set, we train an object pose forecasting network which consumes past and current object poses $(\mathcal{T}_0,\mathcal{T}_1,\dots ,\mathcal{T}_t)$ for each time step $t$ within each trajectory and forecasts the object poses $(\mathcal{T}_{t + 1},\mathcal{T}_{t + 2},\dots ,\mathcal{T}_{t + N})$ in future $N$ steps. By thresholding the forecasting error corresponding to each time step, we identify a set of endpoints where past observations cannot forecast the future very well and partition the complete trajectory $\xi$ to several segments using endpoints $0 = l_{0} < l_{1} < \dots < l_{k} = T$ . Within each segment, we assume the ability to predict the future object pose based on historical information. We then denote $P\in \mathbb{N}$ as the hyperparameter determining the replanning period. For each planning frame $t = 0,P,2P,\dots$ , suppose the next endpoint is $l_{i + 1}$ , i.e., $l_{i}\leq t < l_{i + 1}$ . Then we will plan based on the object pose at frame $\hat{t} = \min (t + P,l_{i + 1})$ , which serves as a landmark. Note here planning is based on the future states + +but avoids bypassing the sharply transitioning points where human motion becomes unpredictable. Also worth mentioning, densely planning is a special case of our method, and landmark planning is a full version. + +# 3.4. Forecast-Aided 4D Imitation Learning + +Traditional methods for human-to-robot handover face challenges in gaining insights into dynamic scene perception. Approaches based on motion planning [42] often emphasize robot morphology and lack dynamic vision perception. They struggle to capture long-horizon information, mainly focusing on the current frame and failing to predict the future. Reinforcement Learning methods [11, 43], while powerful, require extensive training and may train unstably across different scenarios. To enhance the vision-action correlation and establish an efficient training paradigm, we introduce our forecast-aided 4D imitation learning approach. + +In robot perception, the 4D point cloud serves as the common representation. In the $t$ -th frame, we can define $M_t^i \in \mathbb{SE}(3)$ as the relative object pose between the current frame and the $i$ -th frame in the egocentric view. While frame stacking is a straightforward approach, it struggles to capture both motion and geometry effectively. Inspired by recent 4D learning methods [14, 40], we employ the Iterative Closest Point (ICP) registration algorithm [37] to efficiently compute transformation matrices $\{\hat{M}_t^{t-1}, \hat{M}_t^{t-2}, \dots, \hat{M}_t^{t-L_1}\}$ between the point cloud in the $t$ -th frame and the point clouds in previous $L_1$ frames. Applying these transformation matrices to a specific point in the current frame yields its rough coordinates in previous frames. Then we incorporate this flow feature into 3D PointNet++ [35] to encode a global spatial-temporal feature and use Multilayer Perceptron (MLP) to decode it into a 6D egocentric action. The loss function, denoted as $\mathcal{L}_{action}$ , is computed as the L1 loss for aligning 3D points on the robot gripper as defined in [27]. We believe some sophisticated 4D backbones [17, 48] are suitable for 4D understanding, but they are often not suitable for robotic tasks that require a fast reference speed. Our method strikes a balance between effectiveness and simplicity. + +To enhance the responsiveness of our policy to human motion and extend the vision horizon into the future, we introduce an auxiliary task to predict the future motion $\{M_t^{t + 1}, M_t^{t + 2}, \dots, M_t^{t + L_2}\}$ of objects in the next $L_{2}$ frames. Using the ground truth object poses from trajectories, we compute the motion prediction loss for the $t$ -th frame: + +$$ +\mathcal {L} _ {p r e d} = \sum_ {i = t + 1} ^ {t + L _ {2}} \| \hat {M} _ {t} ^ {i} - M _ {t} ^ {i} \| \tag {1} +$$ + +In contrast to reinforcement learning, our imitation learning method requires only a few hours of training and + +achieves great generalizability through large-scale, high-quality demonstrations. We acquire vision-action pairs and ground truth object states from demonstrations, and then supervise our policy using the loss function $\mathcal{L} = \mathcal{L}_{action} + \lambda \mathcal{L}_{pred}$ , where $\lambda$ serves as a weighting hyper-parameter to balance the losses. This efficient distillation paradigm empowers our policy to naturally approach objects with a forecasting intention and to effectively generalize to a wide range of unseen objects and motions. + +# 4. Experiments + +Dataset (1) HandoverSim [9] contains 1000 real-world H2R handover scenes and 20 objects from DexYCB [8]. We evaluate on the "s0" setup which contains 720 training and 144 testing scenes. Each handover motion has a duration of 3 seconds. Following the evaluation of Handover-Sim2real [11], we consider "Sequential" and "Simultaneous" settings. In "s0 (Sequential)", the robot is allowed to move when the hand reaches the handover location and remains static. In "s0 (Simultaneous)", the robot is allowed to move from the beginning of the episode. (2) GenH2R-Sim contains 1,000,000 complex synthetic H2R handover scenes and 3266 objects. We evaluate the "t0" setup which contains 1,000,000 training and 3260 testing scenes. Each handover motion has a duration of 8s and will stop when the robot gripper is close to the object. To introduce more real-world handover scenes into GenH2R-Sim for evaluation, we extract and clip the handover point cloud sequence from HOI4D [28], a real-world mocap dataset. This additional setup is referred to as "t1", which only contains 1000 testing scenes for evaluation. + +Metrics We adhere to the HandoverSim evaluation protocol. A successful handover involves grasping the object from the human hand and moving it to a designated location. Failure cases involve hand contact, object drop, and timeout $(T_{max} = 13s)$ . We report the successful rate and the execution time. Given that some policies prioritize success over speed, potentially wasting considerable human time, and others prioritize speed without considering success, we aim to evaluate both success rate and completion efficiency. To achieve this, we introduce AS (Average Success), akin to AP (Average Precision): + +$$ +\mathrm {A S} = \int_ {0} ^ {1} \operatorname {S u c c e s s} (t) \mathrm {d} t \tag {2} +$$ + +where $\operatorname{Success}(t)$ is success rate considering only successful cases within $t \cdot T_{\max}$ . This method can better evaluate success-time relations which is more suitable in our handover scenarios. + +# 4.1. Evaluating on Different Benchmarks + +Setup We have 2 training sets: small-scale real-world "s0" from HandoverSim and large-scale synthetic "t0" from our + +GenH2R-Sim. Evaluation is conducted on four testing sets: "s0 (Sequential)"/“s0 (Simultaneous)” from HandoverSim and "t0"/“t1” from our GenH2R-Sim. We conduct experiments on our forecast-aid 4D imitation learning from different demonstration strategies including destination planning, dense planning, and landmark planning. As discussed in Section 3.3, destination planning denotes the foresighted planner, dense planning denotes the improved shortsighted planner and landmark planning is our proposed method. + +Baselines We compare our methods with Handover-Sim2real*, the state-of-the-art method in HandoverSim. We additionally compare GA-DDPG which is designed for grasping objects, and OMG Planner. + +Results on different datasets As depicted in Table 1, our method trained on "t0" outperform all methods trained on "s0" by a large margin. Compared with Handover-Sim2real trained on "s0", our landmark planning method trained on "t0" exhibits $11.34\%$ , $16.90\%$ , $12.26\%$ , and $15.93\%$ increase in the success rate across the four testing sets. Moreover, compared with our landmark planning method trained on "s0", the version trained on "t0" demonstrates notable improvements, achieving success rate increases of $8.79\%$ , $6.48\%$ , $11.80\%$ , and $14.13\%$ increase in the same testing sets. This underscores the importance of having a substantial amount of synthetic data for handover training in simulation, which is more beneficial than only relying on a small-scale real-world dataset. Our GenH2R-Sim, with its large-scale complex human hand behavior, generalizes effectively to real-world scenarios such as "s0" in DexYCB and "t1" in HOI4D. + +Results for different methods We can compare our methods with the baseline HandoverSim2real within the same training set in different benchmarks. When trained on "s0", our landmark planning method demonstrates improvements of $2.55\%$ , $10.42\%$ , $0.46\%$ , and $1.8\%$ ( $13.43\%$ , $53.48\%$ , $1.07\%$ , and $23.60\%$ in our reproduced version) across the 4 test sets. Similarly, When trained on "t0", our landmark planning method gives substantial improvements of $20.78\%$ , $23.15\%$ , $7.72\%$ , and $21.23\%$ ( $23.02\%$ , $46.76\%$ , $8.12\%$ , and $34.98\%$ in our reproduced version). The last 3 benchmarks ("s0"(simultaneous), "t0", and "t1") closely resemble real-world scenarios. They greatly demonstrate the effectiveness of our pipeline from distillation-friendly demonstrations to forecast-aided 4D imitation learning, which is capable of handling dynamic robot perception in complex handover scenarios. We also show visualizations on different methods in Figure 4 (a)(b). + +
s0 (Sequential)s0 (Simultaneous)t0t1
STASSTASSTASSTAS
OMG Planner† [42]62.508.3122.5---------
train on s0GA-DDPG [43]50.007.1422.536.814.6623.623.597.3110.346.75.5026.9
Handover-Sim2real [11]75.237.7430.468.756.2335.829.176.2915.052.407.0923.8
Handover-Sim2real* [11]64.357.6126.725.695.4315.028.564.7317.930.605.9816.5
Destination Planning74.319.0122.876.166.9835.225.685.9614.148.48.9415.1
Dense Planning74.779.5419.875.457.3233.027.306.2614.152.39.2415.1
Landmark Planning77.789.2422.379.177.2634.929.636.2315.454.29.0216.6
train on t0GA-DDPG [43]54.767.2624.244.685.3026.524.054.7015.325.505.8614.1
Handover-Sim2real [11]65.977.1829.562.506.0433.533.715.9118.447.106.3524.1
Handover-Sim2real* [11]63.557.5826.538.895.2923.133.314.6421.433.355.8118.4
Destination Planning0.9312.800.016.4812.410.35.968.811.91.6012.030.1
Dense Planning81.489.5121.984.957.4536.338.047.1617.157.908.8518.4
Landmark Planning86.578.8128.085.656.5842.841.436.0122.368.337.7027.9
+ +Results for different Demonstrations Trained on "s0" which consists of relatively simple trajectories, demonstrations based on destination planning can offer a rudimentary cue for downstream visuo-motor policy. However, when trained on "t0" this strategy may lose focus on the object, leading to a failure in maintaining vision-action correlation and providing minimal gains for vision-friendly learning. There is a significant $73.38\%$ / $69.68\%$ decrease in success rate in the "s0" setting. Additionally, distillation from landmark planning slightly outperforms dense planning in success rate and completes the handover process more quickly in all benchmarks. While dense planning can sustain the success rate to some extent, it slows down the agent and may result in unnatural approaches to objects. To jointly consider the time efficiency and the success rate, we compare the Average Success in methods distilled from these two strategies and find that landmark planning is a more efficient and generalizable approach. For instance, when trained on "t0", landmark planning exhibits significant improvements of $6.1\%$ , $6.5\%$ , $5.2\%$ , and $9.5\%$ across the four testing sets. + +# 4.2. Evaluating on different Dataset Scales + +We have proved the crucial role of large-scale datasets in handover generalization in Section 4.1. We can also reveal it by scaling down the usage of "t0" in GenH2R-Sim which contains 1,000,000 training scenes. With $10\%$ data utilization, we observe a $5.93\%$ drop in the success rate on the unseen "t1" test set. This result proves the significance of the dataset scale in our imitation learning method. Thanks to our large-scale data and efficient demonstration generation pipeline, concerns about limited datasets hindering generalization are alleviated. + +Table 1. Evaluating on different benchmarks. We compare our method against baselines from the test set of HandoverSim [9] benchmark ("s0 (sequential)" and "s0 (simultaneous)") and our GenH2R-Sim benchmark ("t0" and "t1"). We use the best-pretrained models from the repositories of GA-DDPG [43] and Handover-Sim2real [11] for evaluation. The results for our method are averaged across 3 random seeds. Note that S means success rate(%) . T means time(s). AS means average success(%) as defined in Equation 2. †: This method [42] is evaluated with ground-truth states and cannot handle dynamic handover like "s0 (Simultaneous)", "t0" and "t1".*: We reproduce the results of HandoverSim2real in the true simultaneous setting as detailed in Section 4.1 to make a fair comparison. + +
MethodsSTAS
w/o Flow31.665.6717.9
w/o Prediction39.186.1120.7
w/o Flow & Prediction37.045.9320.1
Ours41.436.0122.3
+ +Table 2. Ablations on different modules. "w/o Flow" means do not use flow information in the input. "w/o Prediction" means do not add prediction loss in the output. + +# 4.3. Ablation Study + +As shown in Table 2, we prove the effectiveness of our well-designed 4D imitation learning method. The absence of flow information results in a $9.77\%$ decrease (predicting without past information adversely affects the model performance). The absence of the prediction task leads to a $2.25\%$ decrease, and the absence of both components results in a $4.39\%$ decrease. The results demonstrate the model obtains improved performance in leveraging flow information, particularly when tasked with predicting the future object pose. More ablations about our demonstration generation and imitation learning are detailed in the supplementary material. + +# 4.4. Real World Experiments + +Sim-to-Real Transfer In addition to simulation, we deploy the models trained in GenH2R-Sim on a real robotic platform. Using point cloud input from the wrist-mounted camera, we employ the output 6D egocentric action to update the end effector's target position. A user study compares our method against Handover-Sim2real [11]. The supplementary material provides further details. + +User Study We recruited 6 users to compare our method (based on landmark planning) and Handover-Sim2real + +![](images/5bd4ab0a985eb662e2b4f955c7931630c2d156a702d31873371906194b88439e.jpg) +Figure 4. Qualitative results. We in detail compare different methods in simulators and deploy them in the real-world platform. + +
MethodsSimple SettingComplex Setting
Handover-Sim2real56.7%33.3%
Ours90.0%70.0%
+ +Table 3. Sim-to-Real Experiments. We report the success rate of our method and HandoverSim2real in 2 different settings. + +across 5 objects in 2 different settings. In the simple setting, users hand each object to the gripper without quick movements. In the complex setting, users execute a relatively long and quick trajectory. The results are reported in Table 3. We observe that our model gets better performance in completing the handover process across various objects and scenarios. Figure 4(c) shows examples of the real-world handover trials. + +# 5. Conclusion + +In this work, we present a novel framework GenH2R for scaling up the learning of human-to-robot handover. We introduce a new simulator GenH2R-Sim and generate a million human handover animations to facilitate generalizable H2R handover learning. We then propose a distillation-friendly demonstration generation method that automatically produces a million high-quality demonstrations suitable for learning. We further introduce a forecast-aided 4D imitation learning method for effective demonstration distillation. Our experiments demonstrate that scaling-up efforts result in substantial improvement of generalizability to novel geometry and complex motion, both in the simulator and the real world. + +# References + +[1] Christoph Bartneck, Tony Belpaeme, Friederike Eyssel, Takayuki Kanda, Merel Keijrsers, and Selma Šabanović. Human-robot interaction: An introduction. Cambridge University Press, 2020. 2 +[2] Antonio Bicchi and Vijay Kumar. Robotic grasping and contact: A review. In Proceedings 2000 ICRA. Millennium conference. IEEE international conference on robotics and automation. Symposia proceedings (Cat. No. 00CH37065), pages 348-353. IEEE, 2000. 2 +[3] Aude Billard, Sylvain Calinon, Ruediger Dillmann, and Stefan Schaal. Survey: Robot programming by demonstration. Technical report, Springer, 2008. 3 +[4] Jeannette Bohg, Antonio Morales, Tamim Asfour, and Dania Kragic. Data-driven grasp synthesis—a survey. IEEE Transactions on robotics, 30(2):289-309, 2013. 2 +[5] Samarth Brahmbhatt, Chengcheng Tang, Christopher D Twigg, Charles C Kemp, and James Hays. Contactpose: A dataset of grasps with object contact and hand pose. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16, pages 361-378. Springer, 2020. 2 +[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1 +[7] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 2, 3 +[8] Yu-Wei Chao, Wei Yang, Yu Xiang, Pavlo Molchanov, Ankur Handa, Jonathan Tremblay, Yashraj S Narang, Karl Van Wyk, Umar Iqbal, Stan Birchfield, et al. Dexycb: A benchmark for capturing hand grasping of objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9044–9053, 2021. 2, 3, 6, 14, 17 +[9] Yu-Wei Chao, Chris Paxton, Yu Xiang, Wei Yang, Balakumar Sundaralingam, Tao Chen, Adithyavairavan Murali, Maya Cakmak, and Dieter Fox. Handoversim: A simulation framework and benchmark for human-to-robot object handovers. In 2022 International Conference on Robotics and Automation (ICRA), pages 6941–6947. IEEE, 2022. 1, 2, 3, 6, 7, 12, 14, 16, 17, 18 +[10] Sammy Christen, Lan Feng, Wei Yang, Yu-Wei Chao, Otmar Hilliges, and Jie Song. Synh2r: Synthesizing handobject motions for learning human-to-robot handovers. arXiv preprint arXiv:2311.05599, 2023. 2 +[11] Sammy Christen, Wei Yang, Claudia Pérez-D'Arpino, Otmar Hilliges, Dieter Fox, and Yu-Wei Chao. Learning human-to-robot handovers from point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and + +Pattern Recognition, pages 9654-9664, 2023. 2, 5, 6, 7, 13, 14, 16, 17 +[12] Gianluca Corsini, Martin Jacquet, Hemjyoti Das, Amr Affi, Daniel Sidobre, and Antonio Franchi. Nonlinear model predictive control for human-robot handover with application to the aerial case. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7597–7604. IEEE, 2022. 2 +[13] Murtaza Dalal, Ajay Mandlekar, Caelan Garrett, Ankur Handa, Ruslan Salakhutdinov, and Dieter Fox. Imitating task and motion planning with visuomotor transformers. arXiv preprint arXiv:2305.16309, 2023. 3 +[14] Yuhao Dong, Zhuoyang Zhang, Yunze Liu, and Li Yi. Nsm4d: Neural scene model based online 4d point cloud sequence understanding. arXiv preprint arXiv:2310.08326, 2023.5 +[15] Frederik Ebert, Yanlai Yang, Karl Schmeckpeper, Bernadette Bucher, Georgios Georgakis, Kostas Daniilidis, Chelsea Finn, and Sergey Levine. Bridge data: Boosting generalization of robotic skills with cross-domain datasets. arXiv preprint arXiv:2109.13396, 2021. 3 +[16] Clemens Eppner, Arsalan Mousavian, and Dieter Fox. Acronym: A large-scale grasp dataset based on simulation. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 6222-6227. IEEE, 2021. 2, 5 +[17] Hehe Fan, Yi Yang, and Mohan Kankanhalli. Point 4d transformer networks for spatio-temporal modeling in point cloud videos. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14204-14213, 2021. 5 +[18] Zicong Fan, Omid Taheri, Dimitrios Tzionas, Muhammed Kocabas, Manuel Kaufmann, Michael J Black, and Otmar Hilliges. Arctic: A dataset for dexterous bimanual handobject manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12943-12954, 2023. 2 +[19] Hao-Shu Fang, Chenxi Wang, Hongjie Fang, Minghao Gou, Jirong Liu, Hengxu Yan, Wenhai Liu, Yichen Xie, and Cewu Lu. Anygrasp: Robust and efficient grasp perception in spatial and temporal domains. IEEE Transactions on Robotics, 2023. 2 +[20] Chelsea Finn, Tianhe Yu, Tianhao Zhang, Pieter Abbeel, and Sergey Levine. One-shot visual imitation learning via meta-learning. In Conference on robot learning, pages 357-368. PMLR, 2017. 3 +[21] Justin Fu, Aviral Kumar, Ofir Nachum, George Tucker, and Sergey Levine. D4rl: Datasets for deep data-driven reinforcement learning. arXiv preprint arXiv:2004.07219, 2020.3 +[22] Caelan Reed Garrett, Rohan Chitnis, Rachel Holladay, Beomjoon Kim, Tom Silver, Leslie Pack Kaelbling, and Tomás Lozano-Pérez. Integrated task and motion planning. Annual review of control, robotics, and autonomous systems, 4:265-293, 2021. 2, 3 +[23] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: + +Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022. 3 +[24] Huy Ha, Pete Florence, and Shuran Song. Scaling up and distilling down: Language-guided robot skill acquisition. arXiv preprint arXiv:2307.14535, 2023. 2, 3 +[25] Yana Hasson, Gul Varol, Dimitrios Tzionas, Igor Kalevatykh, Michael J Black, Ivan Laptev, and Cordelia Schmid. Learning joint reconstruction of hands and manipulated objects. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11807-11816, 2019. 2 +[26] Eric Jang, Alex Irpan, Mohi Khansari, Daniel Kappler, Frederik Ebert, Corey Lynch, Sergey Levine, and Chelsea Finn. Bc-z: Zero-shot task generalization with robotic imitation learning. In Conference on Robot Learning, pages 991-1002. PMLR, 2022. 3 +[27] Yi Li, Gu Wang, Xiangyang Ji, Yu Xiang, and Dieter Fox. Deepim: Deep iterative matching for 6d pose estimation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 683-698, 2018. 5, 14 +[28] Yunze Liu, Yun Liu, Che Jiang, Kangbo Lyu, Weikang Wan, Hao Shen, Boqiang Liang, Zhoujie Fu, He Wang, and Li Yi. Hoi4d: A 4d egocentric dataset for category-level human-object interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21013-21022, 2022. 2, 6 +[29] Ajay Mandlekar, Danfei Xu, Josiah Wong, Soroush Nasiriy, Chen Wang, Rohun Kulkarni, Li Fei-Fei, Silvio Savarese, Yuke Zhu, and Roberto Martin-Martin. What matters in learning from offline human demonstrations for robot manipulation. arXiv preprint arXiv:2108.03298, 2021. 3 +[30] Naresh Marturi, Marek Kopicki, Alireza Rastegarpanah, Vijaykumar Rajasekaran, Maxime Adjigble, Rustam Stolkin, Ales Leonidis, and Yasemin Bekiroglu. Dynamic grasp and trajectory planning for moving objects. Autonomous Robots, 43:1241-1256, 2019. 2 +[31] Michael James McDonald and Dylan Hadfield-Menell. Guided imitation of task and motion planning. In Conference on Robot Learning, pages 630-640. PMLR, 2022. 3 +[32] OpenAI. Gpt-4 technical report, 2023. 1 +[33] Valerio Ortenzi, Akansel Cosgun, Tommaso Pardi, Wesley P Chan, Elizabeth Croft, and Dana Kulić. Object handovers: a review for robotics. IEEE Transactions on Robotics, 37(6): 1855-1873, 2021. 2 +[34] Dean A Pomerleau. Alvinn: An autonomous land vehicle in a neural network. Advances in neural information processing systems, 1, 1988. 3 +[35] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 5, 14 +[36] Patrick Rosenberger, Akansel Cosgun, Rhys Newbury, Jun Kwan, Valerio Ortenzi, Peter Corke, and Manfred Grafinger. Object-independent human-to-robot handovers using real time robotic vision. IEEE Robotics and Automation Letters, 6(1):17-23, 2020. 2 + +[37] Szymon Rusinkiewicz and Marc Levoy. Efficient variants of the icp algorithm. In Proceedings third international conference on 3-D digital imaging and modeling, pages 145-152. IEEE, 2001. 5, 14 +[38] Thomas B Sheridan. Human–robot interaction: status and challenges. Human factors, 58(4):525-532, 2016. 2 +[39] Shuran Song, Andy Zeng, Johnny Lee, and Thomas Funkhouser. Grasping in the wild: Learning 6dof closed-loop grasping from low-cost demonstrations. IEEE Robotics and Automation Letters, 5(3):4978-4985, 2020. 3 +[40] Zachary Teed and Jia Deng. Raft-3d: Scene flow using rigid-motion embeddings. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8375–8384, 2021. 5 +[41] Lirui Wang, Yu Xiang, and Dieter Fox. Manipulation trajectory optimization with online grasp synthesis and selection. In Robotics: Science and Systems (RSS), 2020. 4 +[42] Lirui Wang, Yu Xiang, and Dieter Fox. Manipulation trajectory optimization with online grasp synthesis and selection. In Proceedings of Robotics: Science and Systems (RSS), 2020. 5, 7, 13, 14 +[43] Lirui Wang, Yu Xiang, Wei Yang, Arsalan Mousavian, and Dieter Fox. Goal-auxiliary actor-critic for 6d robotic grasping with point clouds. In Conference on Robot Learning, pages 70-80. PMLR, 2022. 5, 7, 14, 16, 17 +[44] Lirui Wang, Yiyang Ling, Zhecheng Yuan, Mohit Shridhar, Chen Bao, Yuzhe Qin, Bailin Wang, Huazhe Xu, and Xiaolong Wang. Gensim: Generating robotic simulation tasks via large language models. arXiv preprint arXiv:2310.01361, 2023. 3 +[45] Ruicheng Wang, Jialiang Zhang, Jiayi Chen, Yinzhen Xu, Puhao Li, Tengyu Liu, and He Wang. Dexgraspnet: A large-scale robotic dexterous grasp dataset for general objects based on simulation. arXiv preprint arXiv:2210.02697, 2022.3 +[46] Ruicheng Wang, Jialiang Zhang, Jiayi Chen, Yinzhen Xu, Puhao Li, Tengyu Liu, and He Wang. Dexgraspnet: A large-scale robotic dexterous grasp dataset for general objects based on simulation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 11359-11366. IEEE, 2023. 2 +[47] Yufei Wang, Zhou Xian, Feng Chen, Tsun-Hsuan Wang, Yian Wang, Katerina Fragkiadaki, Zackory Erickson, David Held, and Chuang Gan. Robogen: Towards unleashing infinite data for automated robot learning via generative simulation. arXiv preprint arXiv:2311.01455, 2023. 3 +[48] Hao Wen, Yunze Liu, Jingwei Huang, Bo Duan, and Li Yi. Point primitive transformer for long-term 4d point cloud video understanding. In European Conference on Computer Vision, pages 19-35. Springer, 2022. 5 +[49] Wei Yang, Chris Paxton, Arsalan Mousavian, Yu-Wei Chao, Maya Cakmak, and Dieter Fox. Reactive human-to-robot handovers of arbitrary objects. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 3118–3124. IEEE, 2021. 2 +[50] Ruolin Ye, Wenqiang Xu, Zhendong Xue, Tutian Tang, Yanfeng Wang, and Cewu Lu. H2o: A benchmark for visual + +human-human object handover analysis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15762-15771, 2021. 2 +[51] Gu Zhang, Hao-Shu Fang, Hongjie Fang, and Cewu Lu. Flexible handover with real-time robust dynamic grasp trajectory generation. arXiv preprint arXiv:2308.15622, 2023. 2 +[52] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 1 +[53] Tianhao Zhang, Zoe McCarthy, Owen Jow, Dennis Lee, Xi Chen, Ken Goldberg, and Pieter Abbeel. Deep imitation learning for complex manipulation tasks from virtual reality teleoperation. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pages 5628-5635. IEEE, 2018. 3 \ No newline at end of file diff --git a/2024/GenH2R_ Learning Generalizable Human-to-Robot Handover via Scalable Simulation Demonstration and Imitation/images.zip b/2024/GenH2R_ Learning Generalizable Human-to-Robot Handover via Scalable Simulation Demonstration and Imitation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..465f8f97e82fa4b446ad5d6cb73501448ea4b713 --- /dev/null +++ b/2024/GenH2R_ Learning Generalizable Human-to-Robot Handover via Scalable Simulation Demonstration and Imitation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:16f169bfb77a210427f9db5f8a409e82c2c401bc5d14e04b3856d38c70750a14 +size 526405 diff --git a/2024/GenH2R_ Learning Generalizable Human-to-Robot Handover via Scalable Simulation Demonstration and Imitation/layout.json b/2024/GenH2R_ Learning Generalizable Human-to-Robot Handover via Scalable Simulation Demonstration and Imitation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..baa568785380a029d8cb7c94e6fc978f5ac8d02e --- /dev/null +++ b/2024/GenH2R_ Learning Generalizable Human-to-Robot Handover via Scalable Simulation Demonstration and Imitation/layout.json @@ -0,0 +1,7416 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 105, + 103, + 489, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 103, + 489, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 103, + 489, + 138 + ], + "type": "text", + "content": "GenH2R: Learning Generalizable Human-to-Robot Handover via Scalable Simulation, Demonstration, and Imitation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 78, + 160, + 510, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 160, + 510, + 176 + ], + "spans": [ + { + "bbox": [ + 78, + 160, + 510, + 176 + ], + "type": "text", + "content": "Zifan Wang" + }, + { + "bbox": [ + 78, + 160, + 510, + 176 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 78, + 160, + 510, + 176 + ], + "type": "text", + "content": " Junyu Chen" + }, + { + "bbox": [ + 78, + 160, + 510, + 176 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 78, + 160, + 510, + 176 + ], + "type": "text", + "content": " Ziqing Chen" + }, + { + "bbox": [ + 78, + 160, + 510, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 78, + 160, + 510, + 176 + ], + "type": "text", + "content": " Pengwei Xie" + }, + { + "bbox": [ + 78, + 160, + 510, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 78, + 160, + 510, + 176 + ], + "type": "text", + "content": " Rui Chen" + }, + { + "bbox": [ + 78, + 160, + 510, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 78, + 160, + 510, + 176 + ], + "type": "text", + "content": " Li Yi" + }, + { + "bbox": [ + 78, + 160, + 510, + 176 + ], + "type": "inline_equation", + "content": "^{\\dagger 1,2,3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 177, + 526, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 177, + 526, + 206 + ], + "spans": [ + { + "bbox": [ + 67, + 177, + 526, + 206 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 177, + 526, + 206 + ], + "type": "text", + "content": "Tsinghua University " + }, + { + "bbox": [ + 67, + 177, + 526, + 206 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 177, + 526, + 206 + ], + "type": "text", + "content": "Shanghai Artificial Intelligence Laboratory " + }, + { + "bbox": [ + 67, + 177, + 526, + 206 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 67, + 177, + 526, + 206 + ], + "type": "text", + "content": "Shanghai Qi Zhi Institute https://GenH2R.github.io" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 234, + 192, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 234, + 192, + 247 + ], + "spans": [ + { + "bbox": [ + 143, + 234, + 192, + 247 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 259, + 290, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 259, + 290, + 522 + ], + "spans": [ + { + "bbox": [ + 46, + 259, + 290, + 522 + ], + "type": "text", + "content": "This paper presents GenH2R, a framework for learning generalizable vision-based human-to-robot (H2R) handover skills. The goal is to equip robots with the ability to reliably receive objects with unseen geometry handed over by humans in various complex trajectories. We acquire such generalizability by learning H2R handover at scale with a comprehensive solution including procedural simulation assets creation, automated demonstration generation, and effective imitation learning. We leverage large-scale 3D model repositories, dexterous grasp generation methods, and curve-based 3D animation to create an H2R handover simulation environment named GenH2R-Sim, surpassing the number of scenes in existing simulators by three orders of magnitude. We further introduce a distillation-friendly demonstration generation method that automatically generates a million high-quality demonstrations suitable for learning. Finally, we present a 4D imitation learning method augmented by a future forecasting objective to distill demonstrations into a visuo-motor handover policy. Experimental evaluations in both simulators and the real world demonstrate significant improvements (at least +10% success rate) over baselines in all cases." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 545, + 128, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 545, + 128, + 558 + ], + "spans": [ + { + "bbox": [ + 47, + 545, + 128, + 558 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 565, + 287, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 565, + 287, + 673 + ], + "spans": [ + { + "bbox": [ + 46, + 565, + 287, + 673 + ], + "type": "text", + "content": "The embodied AI research community has long been driven by the goal of empowering robots to interact and collaborate with humans. A crucial aspect of this pursuit is equipping robots with the capability to reliably receive arbitrarily moving objects of varying geometry handed over by humans, based on dynamic visual observations. This human-to-robot (H2R) handover ability allows robots to seamlessly collaborate with humans across a wide range of tasks, including cooking, room tidying, and furniture assembly." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 59, + 673, + 286, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 673, + 286, + 685 + ], + "spans": [ + { + "bbox": [ + 59, + 673, + 286, + 685 + ], + "type": "text", + "content": "However, compared to learning human-free robot manip-" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 315, + 233, + 421, + 316 + ], + "blocks": [ + { + "bbox": [ + 315, + 233, + 421, + 316 + ], + "lines": [ + { + "bbox": [ + 315, + 233, + 421, + 316 + ], + "spans": [ + { + "bbox": [ + 315, + 233, + 421, + 316 + ], + "type": "image", + "image_path": "c7a626db76ce77f60bec918fae38f7655a0f03c08f3c374596a4a8ea69b1bb4a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 317, + 419, + 323 + ], + "lines": [ + { + "bbox": [ + 317, + 317, + 419, + 323 + ], + "spans": [ + { + "bbox": [ + 317, + 317, + 419, + 323 + ], + "type": "text", + "content": "Scalable Synthetic Handover Simulator" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 427, + 233, + 534, + 316 + ], + "blocks": [ + { + "bbox": [ + 427, + 233, + 534, + 316 + ], + "lines": [ + { + "bbox": [ + 427, + 233, + 534, + 316 + ], + "spans": [ + { + "bbox": [ + 427, + 233, + 534, + 316 + ], + "type": "image", + "image_path": "b65b3f3cc6277337bf282f44a073d65fbd1a0a2161d60153dc7902c157040165.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 426, + 317, + 539, + 323 + ], + "lines": [ + { + "bbox": [ + 426, + 317, + 539, + 323 + ], + "spans": [ + { + "bbox": [ + 426, + 317, + 539, + 323 + ], + "type": "text", + "content": "Distillation-friendly Expert Demonstrations" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 315, + 326, + 422, + 400 + ], + "blocks": [ + { + "bbox": [ + 315, + 326, + 422, + 400 + ], + "lines": [ + { + "bbox": [ + 315, + 326, + 422, + 400 + ], + "spans": [ + { + "bbox": [ + 315, + 326, + 422, + 400 + ], + "type": "image", + "image_path": "8365645783823546867e57ab810053920af0517be16c1c1c7e03db65bf22b0c2.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 402, + 400, + 410 + ], + "lines": [ + { + "bbox": [ + 335, + 402, + 400, + 410 + ], + "spans": [ + { + "bbox": [ + 335, + 402, + 400, + 410 + ], + "type": "text", + "content": "Sim-to-real Deployment" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 427, + 325, + 533, + 399 + ], + "blocks": [ + { + "bbox": [ + 427, + 325, + 533, + 399 + ], + "lines": [ + { + "bbox": [ + 427, + 325, + 533, + 399 + ], + "spans": [ + { + "bbox": [ + 427, + 325, + 533, + 399 + ], + "type": "image", + "image_path": "a65c3f3399fb73c24d783dffda82d53ffa58cbb47c7748061f6c5afbaefc24bf.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 429, + 401, + 530, + 410 + ], + "lines": [ + { + "bbox": [ + 429, + 401, + 530, + 410 + ], + "spans": [ + { + "bbox": [ + 429, + 401, + 530, + 410 + ], + "type": "text", + "content": "Forecast-Aided 4D Imitation Learning" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 419, + 545, + 486 + ], + "lines": [ + { + "bbox": [ + 304, + 419, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 419, + 545, + 486 + ], + "type": "text", + "content": "Figure 1. The overview of GenH2R. We introduce a framework for learning generalizable vision-based human-to-robot handover via scalable synthetic simulation, distillation-friendly expert demonstration generation, and a forecast-aided 4D imitation learning method. Our models demonstrate strong generalization capabilities to real datasets and can be deployed to a real robot." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "text", + "content": "ulation skills, the progress in scalably learning H2R handover that can generalize to various objects and versatile human behaviors has lagged due to its unique challenges. Training robots to interact with humans in real-world scenarios entails increased risks and expenses, rendering it inherently non-scalable. Therefore, it is demanded to simulate human behaviors and train robots in simulated environments prior to real-world deployment. However, creating a substantial number of assets for humans handing over objects poses a significant challenge. In a recent study [9] that employed motion capturing to drive virtual humans in a simulator, only 1000 unique human hand motion trajectories were provided for handing over 20 objects. Limited object geometry and human motion assets can hardly capture the complexities of the real world. Besides, the challenge extends to the demonstration side. The success of large language model [6, 32, 52] has suggested a recipe for scaling up learning through modeling large-scale training" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 259, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 259, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 259, + 703 + ], + "type": "text", + "content": "*Equal contribution with the order determined by rolling dice." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 59, + 703, + 136, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 136, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 136, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16362" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 120 + ], + "type": "text", + "content": "data. Nevertheless, collecting robot demonstrations receiving objects from real humans is very costly and unscalable. How to scale up the number of demonstrations while ensuring effective learning poses additional challenges." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 122, + 286, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 122, + 286, + 182 + ], + "spans": [ + { + "bbox": [ + 46, + 122, + 286, + 182 + ], + "type": "text", + "content": "In this work, we aim to learn generalizable H2R handover at scale by tackling the above challenges. We present a comprehensive solution that scales up both the assets and demonstrations and effectively learns a closed-loop visuomotor policy through a novel imitation learning algorithm." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 183, + 286, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 183, + 286, + 375 + ], + "spans": [ + { + "bbox": [ + 46, + 183, + 286, + 375 + ], + "type": "text", + "content": "Specifically, to scale up geometry and motion assets depicting humans handing over various objects, we leverage large-scale 3D model repositories [7, 16], dexterous grasp generation methods [46], and curve-based 3D animation. This enables us to procedurally generate millions of handover scenes, forming an environment named GenH2R-Sim to support generalizable H2R handover learning. GenH2R-Sim surpasses HandoverSim [9], an existing H2R simulator, in both scene quantity (by three orders of magnitude) and unique object involvement (by two orders of magnitude). In addition, scenes in GenH2R-Sim go beyond a straightforward giving and then receiving and cover cases when humans might keep transforming the object in a large range during the entire H2R handover process. This allows for studying complex behaviors such as humans hesitating before handing over." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 376, + 286, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 286, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 286, + 604 + ], + "type": "text", + "content": "To scale up robot demonstrations, we draw inspiration from the Task and Motion Planning (TAMP) [22] literature and propose to automatically generate demonstrations with grasp and motion planning using privileged human motion and object state information. There are some straightforward ways to achieve this goal, such as using the privileged human handover destination information to plan a smooth demonstration. However, the problem is more challenging than it seems since the generated demonstrations need to be suitable for distilling into a visuomotor policy. We identify the vision-action correlation between visual observations and planned actions as the crucial factor influencing distillability and point out that due to the constraints of robot arm morphology one can easily generate observation-irrelevant actions and thus harm distillation. To tackle this challenge, we present a distillation-friendly demonstration generation method that sparsely samples handover animations for landmark states and periodically replans grasp and motion based on privileged future landmarks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "content": "Finally, to distill the above demonstrations into a visuomotor policy, we utilize point cloud input for its richer geometric information and smaller sim-vs-real gap compared to images. We propose a 4D imitation learning method that factors the sequential point cloud observations into geometry and motion parts, facilitating policy learning by better revealing the current scene state. Furthermore, the imitation objective is augmented by a forecasting objective which predicts the future motion of the handover object. Since our" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "text", + "content": "demonstrating actions are generated based on future landmarks, the forecasting objective can help further exploit the vision-action correlation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 108, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 108, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 108, + 545, + 228 + ], + "type": "text", + "content": "We evaluate our learned policy in simulators (Handover-Sim and our own GenH2R-Sim) and the real world. Remarkably, without any mocap assets or real-world demonstrations, our method achieves significantly better performance compared to baselines across all settings (at least " + }, + { + "bbox": [ + 304, + 108, + 545, + 228 + ], + "type": "inline_equation", + "content": "+10\\%" + }, + { + "bbox": [ + 304, + 108, + 545, + 228 + ], + "type": "text", + "content": " success rate). Our experiments highlight that the scaling-up efforts bring substantial improvement in policy generalizability to novel geometry and complex motion. Furthermore, these efforts greatly facilitate skill transfer to real robotic systems." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 228, + 545, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 545, + 336 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 545, + 336 + ], + "type": "text", + "content": "In summary, the key contribution of this paper is a novel framework scaling up the learning of H2R handover with the following three components: i) a simulation environment named GenH2R-Sim consists of millions of human handover animations for generalizable H2R handover learning, ii) an empirically validated automatic robot demonstration generation pipeline for vision-based closed-loop control, iii) a forecast-aided 4D imitation learning method effective in distilling the large-scale demonstrations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 372, + 392, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 372, + 392, + 384 + ], + "spans": [ + { + "bbox": [ + 306, + 372, + 392, + 384 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 392, + 463, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 392, + 463, + 403 + ], + "spans": [ + { + "bbox": [ + 306, + 392, + 463, + 403 + ], + "type": "text", + "content": "2.1. Human-to-Robot Handovers" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 411, + 545, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 545, + 650 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 545, + 650 + ], + "type": "text", + "content": "Recently, significant progress in human-robot handovers [12, 33, 36] has been observed, driven by the increasing popularity of human-robot interaction [1, 38] and the emergence of extensive datasets [5, 8, 18, 25, 28, 50] capturing hand-object interactions. Some traditional methods [2, 4] require 3D object models and struggle to handle unseen objects. One possible way is to consider grasping and dynamic motion planning [19, 30, 49, 51]. However, these methods often exhibit constrained motions and perform poorly on large-scale datasets. HandoverSim [9], a physics-simulated environment, introduced a new simulation benchmark for human-to-robot object handovers. Leveraging DexYCB [8], a dataset of human grasping objects and performing handover attempts, this environment allows training learning-based handover policies such as [11]. However, it lacks large-scale and diverse handover scenes, which limits generalizable handovers. At the same time, SynH2R [10] proposes to use synthetic data but makes limited progress. Building on this, we propose GenH2R-Sim , aiming to benchmark generalizable handover." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 658, + 488, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 488, + 671 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 488, + 671 + ], + "type": "text", + "content": "2.2. Scaling Up Robot Demonstrations" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "For robot learning, scaling up data collection for manipulation skills has spurred extensive research. Approaches include leveraging large language models [24] or hardware" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16363" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 270, + 262 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 270, + 262 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 270, + 262 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 270, + 262 + ], + "type": "image", + "image_path": "8d67f0d7bca143f2d5d11b877b11361c982ad3fea0621b6abb9be4e22168ebfd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 271, + 545, + 337 + ], + "lines": [ + { + "bbox": [ + 46, + 271, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 46, + 271, + 545, + 337 + ], + "type": "text", + "content": "Figure 2. The overview of our framework. First, we propose a new simulation environment named GenH2R-Sim, featuring large-scale synthetic datasets with diversity in object geometry, grasp poses, and complex trajectories. Second, other than destination planning (move straight toward the final position) and dense planning (replan at each step), we propose a distillation-friendly demonstration generation method—landmark planning, predicting landmarks on the trajectory (as indicated by the dashed object above) and replanning based on those landmarks. Thirdly, our Forecast-aided 4D Imitation Learning leverages past flow information, and the forecasting objective enhances the exploitation of vision-action correlation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 271, + 70, + 399, + 262 + ], + "blocks": [ + { + "bbox": [ + 271, + 70, + 399, + 262 + ], + "lines": [ + { + "bbox": [ + 271, + 70, + 399, + 262 + ], + "spans": [ + { + "bbox": [ + 271, + 70, + 399, + 262 + ], + "type": "image", + "image_path": "19062cc6078a2249cc73be71dfa6bb7489a75942007b3e060b4eeed8dfd8e240.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 399, + 70, + 533, + 262 + ], + "blocks": [ + { + "bbox": [ + 399, + 70, + 533, + 262 + ], + "lines": [ + { + "bbox": [ + 399, + 70, + 533, + 262 + ], + "spans": [ + { + "bbox": [ + 399, + 70, + 533, + 262 + ], + "type": "image", + "image_path": "05fec30c5092a9c1b790bc8952ca4c628d26fc02d4f4779e92ec95fdb4f60b01.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 346, + 287, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 346, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 287, + 467 + ], + "type": "text", + "content": "capabilities [39], utilizing non-robotics datasets [23], and employing trial-and-error explorations [21]. As depicted in [24], one of the challenges is scaling up robot-complete data. A popular line of research scales up demonstration generation via Task and Motion Planning [13, 22, 31]. These works usually focus on fairly static scenes without active motion or object and task variety [44, 47] while our method extends to dynamic H2R handover by considering how to interpret human behavior and generate demonstrations easy to be distilled by closed-loop visuo-motor policy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 475, + 251, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 475, + 251, + 488 + ], + "spans": [ + { + "bbox": [ + 47, + 475, + 251, + 488 + ], + "type": "text", + "content": "2.3. Offline Learning from Demonstrations" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 494, + 287, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 494, + 287, + 638 + ], + "spans": [ + { + "bbox": [ + 46, + 494, + 287, + 638 + ], + "type": "text", + "content": "Imitation Learning (IL) represents a methodology for training embodied agents in manipulation tasks by utilizing expert demonstrations. The commonly used Behavior Cloning (BC) [34] strategy directly trains the policy to imitate expert actions in a supervised learning manner. Despite its simplicity, this approach has demonstrated remarkable effectiveness in robotic manipulation [3, 20, 29, 53] especially when combined with a substantial number of high-quality demonstrations [15, 26]. Inspired by these works, we adopt an imitation learning paradigm, focusing on how to leverage spatial-temporal perception and future forecasting to better consume our distillation-friendly demonstrations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 650, + 102, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 650, + 102, + 662 + ], + "spans": [ + { + "bbox": [ + 47, + 650, + 102, + 662 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 670, + 114, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 114, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 114, + 681 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "For the generalizable H2R handover task, we introduce GenH2R, a framework designed to learn control policies," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 346, + 545, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 346, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 545, + 431 + ], + "type": "text", + "content": "specifically 6D control actions for the robot gripper, using segmented point cloud data captured from an egocentric camera. We describe our method for synthesizing human handover animations in Section 3.2, generating expert demonstrations in Section 3.3, and distilling demonstrations to 4D vision-based neural networks by imitation learning in Section 3.4, as the pipeline depicted in Figure 2." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 441, + 391, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 441, + 391, + 452 + ], + "spans": [ + { + "bbox": [ + 306, + 441, + 391, + 452 + ], + "type": "text", + "content": "3.2. GenH2R-Sim" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 460, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 567 + ], + "type": "text", + "content": "The size and quality of human-object datasets in simulators play a crucial role in generating high-quality handover demonstrations and training reliable policies for handover scenarios. The recent handover simulator, HandoverSim [9], utilizes the DexYCB [8] dataset, which captures real-world human grasping objects in a limited manner, comprising only 1000 scenes with 20 distinct objects. In the real world, scenarios can be more complex and may involve intricate trajectories and poses beyond those in DexYCB." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 569, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 569, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 545, + 640 + ], + "type": "text", + "content": "To address these limitations, we introduce a new environment, GenH2R-Sim, to overcome these deficiencies and facilitate generalizable handovers. To diversify geometry and motion assets depicting humans handing over various objects, we focus on two primary aspects: the hand grasping pose and the hand-object moving trajectory within a scene." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "In aspects of grasping poses, DexGraspNet [45] has made significant contributions by employing optimization techniques to generate a substantial dataset of human hand grasp poses. We utilize this method to generate approximately 1,000,000 grasp poses for 3,266 different objects sourced from Shapenet [7]. These objects span a wide range" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16364" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "content": "of categories, from larger items like computers to smaller ones like mobile phones, covering most sizes and shapes encountered in real-life handovers." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 109, + 288, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 109, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 109, + 288, + 312 + ], + "type": "text", + "content": "In aspects of hand-object moving trajectories, we propose to use Bézier curves, which are one class of smooth curves determined by several control points, to generate complex yet smooth-transiting motion trajectories. We use multiple Bézier curves to model different stages of the motion, and link the ends of these curves to create a seamless track. We can generate scenes matching various scenarios of different complexity in the real world by adjusting the distribution of control points of the trajectory and the speed of the human hand. To enhance the trajectory's realism, we incorporate consistent object rotations, which also enhances the importance of choosing the appropriate grasp for the robotic arm. Since we can always attach a new segment of motion at the end of the current motion and the duration is much longer than DexYCB scenes, the destination of the hand-object is not a significant factor, so we just randomly select a point within the reach of the robotic arm." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 313, + 287, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 313, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 46, + 313, + 287, + 443 + ], + "type": "text", + "content": "We do not guarantee that every item in the dataset we generate perfectly mimics the human-like characteristics of real-world data, but our approach ensures a significantly higher degree of domain randomization and provides greater diversity in terms of geometry and motion. Given the challenges in scaling up real-world motion capture datasets, we opt for a large-scale synthetic dataset for our handover simulations. Our key insight is that for both demonstrations and policy learning, having a substantial amount of synthetic data is more beneficial than relying on a small-scale real-world dataset." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 445, + 287, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 287, + 577 + ], + "type": "text", + "content": "GenH2R-Sim follows the setup of HandoverSim, which consists of a Panda 7DoF robotic arm with a gripper and a wrist-mounted RGB-D camera, and a simulated human hand. Just like HandoverSim, we switch from the prehandover kinematic phase to the handover dynamic phase when the object has been in contact with the gripper. HandoverSim is not adaptive to the robot's action and just loads and replays every frame of the data. To align with the real-world handover process more naturally in GenH2R-Sim, the simulated hand will stop from moving and wait for handover when the robot arm is close to the object." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 586, + 271, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 271, + 598 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 271, + 598 + ], + "type": "text", + "content": "3.3. Generating Demonstrations for Distillation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": "In this section, we address a key question in learning visuomotor policy: how to efficiently generate robot demonstrations that incorporate paired vision-action data from successful task experiences. While distilling successful demonstrations into a single policy has proven effective for open-loop control tasks, the challenge lies in closed-loop visuo-motor control, where the quality of demonstrations becomes crucial for learning. Merely ensuring success is no longer sufficient. We present two examples of demon-" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 307, + 71, + 545, + 239 + ], + "blocks": [ + { + "bbox": [ + 307, + 71, + 545, + 239 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 545, + 239 + ], + "type": "image", + "image_path": "c9d34d63fb5e2bdd134e280734397e7af3be6a5d1c059e612b63a981738d71ff.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 247, + 545, + 302 + ], + "lines": [ + { + "bbox": [ + 305, + 247, + 545, + 302 + ], + "spans": [ + { + "bbox": [ + 305, + 247, + 545, + 302 + ], + "type": "text", + "content": "Figure 3. Different demonstration generation methods for dynamic handover. The orange curve shows the hand-object trajectory. The blue, red, and green curves show the example trajectories generated by the foresighted planner, the shortsighted planner, and our planner, respectively." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 316, + 546, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 546, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 546, + 567 + ], + "type": "text", + "content": "stration generation with different grasp and motion planning strategies as shown in Figure 3. In the first example, a foresighted planner generates smooth, short demonstrations based on the privileged destination end state of a human handover animation. Though efficient, the planned path does not align actions with the dynamic visual observations during the handover. Distilling such demonstrations requires accurately forecasting the end state of the human trajectory, which can be extremely challenging in complex handover cases. The second example involves a shortsighted planner that independently replans grasp and motion at each time step using privileged hand and object states. Due to robot morphology constraints and the multisolution nature of common robot planners, smooth visual observations may correspond to unsmooth and multi-modal robot trajectories, increasing the difficulty of distillation. We emphasize the importance of distillability as a quality factor for handover demonstrations. An effective demonstration generation method must consider the vision-action correlation by jointly incorporating robot morphology and dynamic vision during grasp and motion planning." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 568, + 545, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 568, + 545, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 568, + 545, + 663 + ], + "type": "text", + "content": "Along this line, we base our method on the foresighted and shortsighted planner mentioned above to combine the advantages of both sides while encouraging the demonstration distillability. We first improve the shortsighted planner so that sequentially smooth visual observations result in smooth grasp and motion plans. Then we improve the handover efficiency by looking toward the future while guaranteeing the vision-action correlation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "content": "To be specific, we build our method based on the OMG planner [41] for grasp and motion planning. This planner optimizes the grasp and motion path by considering the object's 6D pose and a set of candidate grasp poses. To support" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16365" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 301 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 301 + ], + "type": "text", + "content": "this optimization, we provide privileged knowledge that includes the object's 6D pose, candidate grasps generated through physics simulation [16], and human hand poses for filtering out invalid grasps. However, independently calling the OMG planner for each time step may result in unsmooth trajectories, as it is designed for static scenarios. To address this, we sequentially plan the grasp and motion based on the privileged knowledge by: 1) sorting grasps based on their pose distance to the robot end effector and attempting inverse kinematics (IK) starting from the nearest grasp until success; 2) initializing IK based on the robot arm pose from the previous time step; 3) invoking the OMG planner only when IK can be successfully solved. By prioritizing closer grasps, we encourage the object to remain within the field of view of a wrist camera, reducing visually irrelevant actions when the object is not visible. Additionally, enforcing IK smoothness improves the overall trajectory smoothness. As a result, the enhanced vision-action correlation dramatically improves the demonstration quality." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 307, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 307, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 307, + 289, + 715 + ], + "type": "text", + "content": "Our approach modifies the OMG planner for dynamic grasp and motion planning. However, densely replanning at each time step leads to inefficient and non-smooth zigzag demonstrations, which does not align with how humans receive objects. Humans anticipate dynamic scene changes before taking action. On the other hand, a highly fore-sighted planner that directly plans grasp and motion based on the end state of a human handover animation can disrupt the vision-action correlation. To strike a balance between these extremes, we propose an algorithm that sparsely samples handover animations for landmark states and periodically replans grasp and motion based on future landmarks. The key idea is to select landmarks strategically so that the planner only considers visually foreseeable futures. Specifically, let " + }, + { + "bbox": [ + 46, + 307, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\xi = (\\mathcal{T}_0,\\mathcal{T}_1,\\dots ,\\mathcal{T}_{T - 1})" + }, + { + "bbox": [ + 46, + 307, + 289, + 715 + ], + "type": "text", + "content": " represent an object trajectory, where " + }, + { + "bbox": [ + 46, + 307, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_t\\in \\mathbb{SE}(3)" + }, + { + "bbox": [ + 46, + 307, + 289, + 715 + ], + "type": "text", + "content": " denotes the object pose in the " + }, + { + "bbox": [ + 46, + 307, + 289, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 307, + 289, + 715 + ], + "type": "text", + "content": "-th frame within the world coordinate system. Based on all the object trajectories in the training set, we train an object pose forecasting network which consumes past and current object poses " + }, + { + "bbox": [ + 46, + 307, + 289, + 715 + ], + "type": "inline_equation", + "content": "(\\mathcal{T}_0,\\mathcal{T}_1,\\dots ,\\mathcal{T}_t)" + }, + { + "bbox": [ + 46, + 307, + 289, + 715 + ], + "type": "text", + "content": " for each time step " + }, + { + "bbox": [ + 46, + 307, + 289, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 307, + 289, + 715 + ], + "type": "text", + "content": " within each trajectory and forecasts the object poses " + }, + { + "bbox": [ + 46, + 307, + 289, + 715 + ], + "type": "inline_equation", + "content": "(\\mathcal{T}_{t + 1},\\mathcal{T}_{t + 2},\\dots ,\\mathcal{T}_{t + N})" + }, + { + "bbox": [ + 46, + 307, + 289, + 715 + ], + "type": "text", + "content": " in future " + }, + { + "bbox": [ + 46, + 307, + 289, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 307, + 289, + 715 + ], + "type": "text", + "content": " steps. By thresholding the forecasting error corresponding to each time step, we identify a set of endpoints where past observations cannot forecast the future very well and partition the complete trajectory " + }, + { + "bbox": [ + 46, + 307, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 46, + 307, + 289, + 715 + ], + "type": "text", + "content": " to several segments using endpoints " + }, + { + "bbox": [ + 46, + 307, + 289, + 715 + ], + "type": "inline_equation", + "content": "0 = l_{0} < l_{1} < \\dots < l_{k} = T" + }, + { + "bbox": [ + 46, + 307, + 289, + 715 + ], + "type": "text", + "content": ". Within each segment, we assume the ability to predict the future object pose based on historical information. We then denote " + }, + { + "bbox": [ + 46, + 307, + 289, + 715 + ], + "type": "inline_equation", + "content": "P\\in \\mathbb{N}" + }, + { + "bbox": [ + 46, + 307, + 289, + 715 + ], + "type": "text", + "content": " as the hyperparameter determining the replanning period. For each planning frame " + }, + { + "bbox": [ + 46, + 307, + 289, + 715 + ], + "type": "inline_equation", + "content": "t = 0,P,2P,\\dots" + }, + { + "bbox": [ + 46, + 307, + 289, + 715 + ], + "type": "text", + "content": ", suppose the next endpoint is " + }, + { + "bbox": [ + 46, + 307, + 289, + 715 + ], + "type": "inline_equation", + "content": "l_{i + 1}" + }, + { + "bbox": [ + 46, + 307, + 289, + 715 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 46, + 307, + 289, + 715 + ], + "type": "inline_equation", + "content": "l_{i}\\leq t < l_{i + 1}" + }, + { + "bbox": [ + 46, + 307, + 289, + 715 + ], + "type": "text", + "content": ". Then we will plan based on the object pose at frame " + }, + { + "bbox": [ + 46, + 307, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\hat{t} = \\min (t + P,l_{i + 1})" + }, + { + "bbox": [ + 46, + 307, + 289, + 715 + ], + "type": "text", + "content": ", which serves as a landmark. Note here planning is based on the future states" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "but avoids bypassing the sharply transitioning points where human motion becomes unpredictable. Also worth mentioning, densely planning is a special case of our method, and landmark planning is a full version." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 305, + 129, + 509, + 143 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 129, + 509, + 143 + ], + "spans": [ + { + "bbox": [ + 305, + 129, + 509, + 143 + ], + "type": "text", + "content": "3.4. Forecast-Aided 4D Imitation Learning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 148, + 545, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 148, + 545, + 279 + ], + "spans": [ + { + "bbox": [ + 304, + 148, + 545, + 279 + ], + "type": "text", + "content": "Traditional methods for human-to-robot handover face challenges in gaining insights into dynamic scene perception. Approaches based on motion planning [42] often emphasize robot morphology and lack dynamic vision perception. They struggle to capture long-horizon information, mainly focusing on the current frame and failing to predict the future. Reinforcement Learning methods [11, 43], while powerful, require extensive training and may train unstably across different scenarios. To enhance the vision-action correlation and establish an efficient training paradigm, we introduce our forecast-aided 4D imitation learning approach." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 280, + 546, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 280, + 546, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 280, + 546, + 555 + ], + "type": "text", + "content": "In robot perception, the 4D point cloud serves as the common representation. In the " + }, + { + "bbox": [ + 304, + 280, + 546, + 555 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 280, + 546, + 555 + ], + "type": "text", + "content": "-th frame, we can define " + }, + { + "bbox": [ + 304, + 280, + 546, + 555 + ], + "type": "inline_equation", + "content": "M_t^i \\in \\mathbb{SE}(3)" + }, + { + "bbox": [ + 304, + 280, + 546, + 555 + ], + "type": "text", + "content": " as the relative object pose between the current frame and the " + }, + { + "bbox": [ + 304, + 280, + 546, + 555 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 280, + 546, + 555 + ], + "type": "text", + "content": "-th frame in the egocentric view. While frame stacking is a straightforward approach, it struggles to capture both motion and geometry effectively. Inspired by recent 4D learning methods [14, 40], we employ the Iterative Closest Point (ICP) registration algorithm [37] to efficiently compute transformation matrices " + }, + { + "bbox": [ + 304, + 280, + 546, + 555 + ], + "type": "inline_equation", + "content": "\\{\\hat{M}_t^{t-1}, \\hat{M}_t^{t-2}, \\dots, \\hat{M}_t^{t-L_1}\\}" + }, + { + "bbox": [ + 304, + 280, + 546, + 555 + ], + "type": "text", + "content": " between the point cloud in the " + }, + { + "bbox": [ + 304, + 280, + 546, + 555 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 280, + 546, + 555 + ], + "type": "text", + "content": "-th frame and the point clouds in previous " + }, + { + "bbox": [ + 304, + 280, + 546, + 555 + ], + "type": "inline_equation", + "content": "L_1" + }, + { + "bbox": [ + 304, + 280, + 546, + 555 + ], + "type": "text", + "content": " frames. Applying these transformation matrices to a specific point in the current frame yields its rough coordinates in previous frames. Then we incorporate this flow feature into 3D PointNet++ [35] to encode a global spatial-temporal feature and use Multilayer Perceptron (MLP) to decode it into a 6D egocentric action. The loss function, denoted as " + }, + { + "bbox": [ + 304, + 280, + 546, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{action}" + }, + { + "bbox": [ + 304, + 280, + 546, + 555 + ], + "type": "text", + "content": ", is computed as the L1 loss for aligning 3D points on the robot gripper as defined in [27]. We believe some sophisticated 4D backbones [17, 48] are suitable for 4D understanding, but they are often not suitable for robotic tasks that require a fast reference speed. Our method strikes a balance between effectiveness and simplicity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 555, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 545, + 639 + ], + "type": "text", + "content": "To enhance the responsiveness of our policy to human motion and extend the vision horizon into the future, we introduce an auxiliary task to predict the future motion " + }, + { + "bbox": [ + 304, + 555, + 545, + 639 + ], + "type": "inline_equation", + "content": "\\{M_t^{t + 1}, M_t^{t + 2}, \\dots, M_t^{t + L_2}\\}" + }, + { + "bbox": [ + 304, + 555, + 545, + 639 + ], + "type": "text", + "content": " of objects in the next " + }, + { + "bbox": [ + 304, + 555, + 545, + 639 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 304, + 555, + 545, + 639 + ], + "type": "text", + "content": " frames. Using the ground truth object poses from trajectories, we compute the motion prediction loss for the " + }, + { + "bbox": [ + 304, + 555, + 545, + 639 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 555, + 545, + 639 + ], + "type": "text", + "content": "-th frame:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 368, + 650, + 545, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 650, + 545, + 682 + ], + "spans": [ + { + "bbox": [ + 368, + 650, + 545, + 682 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p r e d} = \\sum_ {i = t + 1} ^ {t + L _ {2}} \\| \\hat {M} _ {t} ^ {i} - M _ {t} ^ {i} \\| \\tag {1}", + "image_path": "f2bbda85d7c0e84011cb3a6bfbb994b8c906f4f5d5d6fcf834626cdd24d20af1.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "In contrast to reinforcement learning, our imitation learning method requires only a few hours of training and" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16366" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "achieves great generalizability through large-scale, high-quality demonstrations. We acquire vision-action pairs and ground truth object states from demonstrations, and then supervise our policy using the loss function " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\mathcal{L}_{action} + \\lambda \\mathcal{L}_{pred}" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": " serves as a weighting hyper-parameter to balance the losses. This efficient distillation paradigm empowers our policy to naturally approach objects with a forecasting intention and to effectively generalize to a wide range of unseen objects and motions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 190, + 128, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 190, + 128, + 204 + ], + "spans": [ + { + "bbox": [ + 47, + 190, + 128, + 204 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 209, + 289, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 209, + 289, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 289, + 448 + ], + "type": "text", + "content": "Dataset (1) HandoverSim [9] contains 1000 real-world H2R handover scenes and 20 objects from DexYCB [8]. We evaluate on the \"s0\" setup which contains 720 training and 144 testing scenes. Each handover motion has a duration of 3 seconds. Following the evaluation of Handover-Sim2real [11], we consider \"Sequential\" and \"Simultaneous\" settings. In \"s0 (Sequential)\", the robot is allowed to move when the hand reaches the handover location and remains static. In \"s0 (Simultaneous)\", the robot is allowed to move from the beginning of the episode. (2) GenH2R-Sim contains 1,000,000 complex synthetic H2R handover scenes and 3266 objects. We evaluate the \"t0\" setup which contains 1,000,000 training and 3260 testing scenes. Each handover motion has a duration of 8s and will stop when the robot gripper is close to the object. To introduce more real-world handover scenes into GenH2R-Sim for evaluation, we extract and clip the handover point cloud sequence from HOI4D [28], a real-world mocap dataset. This additional setup is referred to as \"t1\", which only contains 1000 testing scenes for evaluation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 449, + 289, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 289, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 289, + 581 + ], + "type": "text", + "content": "Metrics We adhere to the HandoverSim evaluation protocol. A successful handover involves grasping the object from the human hand and moving it to a designated location. Failure cases involve hand contact, object drop, and timeout " + }, + { + "bbox": [ + 46, + 449, + 289, + 581 + ], + "type": "inline_equation", + "content": "(T_{max} = 13s)" + }, + { + "bbox": [ + 46, + 449, + 289, + 581 + ], + "type": "text", + "content": ". We report the successful rate and the execution time. Given that some policies prioritize success over speed, potentially wasting considerable human time, and others prioritize speed without considering success, we aim to evaluate both success rate and completion efficiency. To achieve this, we introduce AS (Average Success), akin to AP (Average Precision):" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 585, + 287, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 585, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 116, + 585, + 287, + 613 + ], + "type": "interline_equation", + "content": "\\mathrm {A S} = \\int_ {0} ^ {1} \\operatorname {S u c c e s s} (t) \\mathrm {d} t \\tag {2}", + "image_path": "43b6879bcc7255a7aa6b55f7d5750a5bbd0fbfe4fdde05372c03a2cd1028de23.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 617, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 617, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\operatorname{Success}(t)" + }, + { + "bbox": [ + 46, + 617, + 287, + 665 + ], + "type": "text", + "content": " is success rate considering only successful cases within " + }, + { + "bbox": [ + 46, + 617, + 287, + 665 + ], + "type": "inline_equation", + "content": "t \\cdot T_{\\max}" + }, + { + "bbox": [ + 46, + 617, + 287, + 665 + ], + "type": "text", + "content": ". This method can better evaluate success-time relations which is more suitable in our handover scenarios." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 671, + 243, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 243, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 243, + 685 + ], + "type": "text", + "content": "4.1. Evaluating on Different Benchmarks" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "Setup We have 2 training sets: small-scale real-world \"s0\" from HandoverSim and large-scale synthetic \"t0\" from our" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": "GenH2R-Sim. Evaluation is conducted on four testing sets: \"s0 (Sequential)\"/“s0 (Simultaneous)” from HandoverSim and \"t0\"/“t1” from our GenH2R-Sim. We conduct experiments on our forecast-aid 4D imitation learning from different demonstration strategies including destination planning, dense planning, and landmark planning. As discussed in Section 3.3, destination planning denotes the foresighted planner, dense planning denotes the improved shortsighted planner and landmark planning is our proposed method." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 180, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 180, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 545, + 228 + ], + "type": "text", + "content": "Baselines We compare our methods with Handover-Sim2real*, the state-of-the-art method in HandoverSim. We additionally compare GA-DDPG which is designed for grasping objects, and OMG Planner." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 229, + 546, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 229, + 546, + 432 + ], + "spans": [ + { + "bbox": [ + 304, + 229, + 546, + 432 + ], + "type": "text", + "content": "Results on different datasets As depicted in Table 1, our method trained on \"t0\" outperform all methods trained on \"s0\" by a large margin. Compared with Handover-Sim2real trained on \"s0\", our landmark planning method trained on \"t0\" exhibits " + }, + { + "bbox": [ + 304, + 229, + 546, + 432 + ], + "type": "inline_equation", + "content": "11.34\\%" + }, + { + "bbox": [ + 304, + 229, + 546, + 432 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 229, + 546, + 432 + ], + "type": "inline_equation", + "content": "16.90\\%" + }, + { + "bbox": [ + 304, + 229, + 546, + 432 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 229, + 546, + 432 + ], + "type": "inline_equation", + "content": "12.26\\%" + }, + { + "bbox": [ + 304, + 229, + 546, + 432 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 229, + 546, + 432 + ], + "type": "inline_equation", + "content": "15.93\\%" + }, + { + "bbox": [ + 304, + 229, + 546, + 432 + ], + "type": "text", + "content": " increase in the success rate across the four testing sets. Moreover, compared with our landmark planning method trained on \"s0\", the version trained on \"t0\" demonstrates notable improvements, achieving success rate increases of " + }, + { + "bbox": [ + 304, + 229, + 546, + 432 + ], + "type": "inline_equation", + "content": "8.79\\%" + }, + { + "bbox": [ + 304, + 229, + 546, + 432 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 229, + 546, + 432 + ], + "type": "inline_equation", + "content": "6.48\\%" + }, + { + "bbox": [ + 304, + 229, + 546, + 432 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 229, + 546, + 432 + ], + "type": "inline_equation", + "content": "11.80\\%" + }, + { + "bbox": [ + 304, + 229, + 546, + 432 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 229, + 546, + 432 + ], + "type": "inline_equation", + "content": "14.13\\%" + }, + { + "bbox": [ + 304, + 229, + 546, + 432 + ], + "type": "text", + "content": " increase in the same testing sets. This underscores the importance of having a substantial amount of synthetic data for handover training in simulation, which is more beneficial than only relying on a small-scale real-world dataset. Our GenH2R-Sim, with its large-scale complex human hand behavior, generalizes effectively to real-world scenarios such as \"s0\" in DexYCB and \"t1\" in HOI4D." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 433, + 547, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 433, + 547, + 636 + ], + "spans": [ + { + "bbox": [ + 304, + 433, + 547, + 636 + ], + "type": "text", + "content": "Results for different methods We can compare our methods with the baseline HandoverSim2real within the same training set in different benchmarks. When trained on \"s0\", our landmark planning method demonstrates improvements of " + }, + { + "bbox": [ + 304, + 433, + 547, + 636 + ], + "type": "inline_equation", + "content": "2.55\\%" + }, + { + "bbox": [ + 304, + 433, + 547, + 636 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 433, + 547, + 636 + ], + "type": "inline_equation", + "content": "10.42\\%" + }, + { + "bbox": [ + 304, + 433, + 547, + 636 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 433, + 547, + 636 + ], + "type": "inline_equation", + "content": "0.46\\%" + }, + { + "bbox": [ + 304, + 433, + 547, + 636 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 433, + 547, + 636 + ], + "type": "inline_equation", + "content": "1.8\\%" + }, + { + "bbox": [ + 304, + 433, + 547, + 636 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 433, + 547, + 636 + ], + "type": "inline_equation", + "content": "13.43\\%" + }, + { + "bbox": [ + 304, + 433, + 547, + 636 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 433, + 547, + 636 + ], + "type": "inline_equation", + "content": "53.48\\%" + }, + { + "bbox": [ + 304, + 433, + 547, + 636 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 433, + 547, + 636 + ], + "type": "inline_equation", + "content": "1.07\\%" + }, + { + "bbox": [ + 304, + 433, + 547, + 636 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 433, + 547, + 636 + ], + "type": "inline_equation", + "content": "23.60\\%" + }, + { + "bbox": [ + 304, + 433, + 547, + 636 + ], + "type": "text", + "content": " in our reproduced version) across the 4 test sets. Similarly, When trained on \"t0\", our landmark planning method gives substantial improvements of " + }, + { + "bbox": [ + 304, + 433, + 547, + 636 + ], + "type": "inline_equation", + "content": "20.78\\%" + }, + { + "bbox": [ + 304, + 433, + 547, + 636 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 433, + 547, + 636 + ], + "type": "inline_equation", + "content": "23.15\\%" + }, + { + "bbox": [ + 304, + 433, + 547, + 636 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 433, + 547, + 636 + ], + "type": "inline_equation", + "content": "7.72\\%" + }, + { + "bbox": [ + 304, + 433, + 547, + 636 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 433, + 547, + 636 + ], + "type": "inline_equation", + "content": "21.23\\%" + }, + { + "bbox": [ + 304, + 433, + 547, + 636 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 433, + 547, + 636 + ], + "type": "inline_equation", + "content": "23.02\\%" + }, + { + "bbox": [ + 304, + 433, + 547, + 636 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 433, + 547, + 636 + ], + "type": "inline_equation", + "content": "46.76\\%" + }, + { + "bbox": [ + 304, + 433, + 547, + 636 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 433, + 547, + 636 + ], + "type": "inline_equation", + "content": "8.12\\%" + }, + { + "bbox": [ + 304, + 433, + 547, + 636 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 433, + 547, + 636 + ], + "type": "inline_equation", + "content": "34.98\\%" + }, + { + "bbox": [ + 304, + 433, + 547, + 636 + ], + "type": "text", + "content": " in our reproduced version). The last 3 benchmarks (\"s0\"(simultaneous), \"t0\", and \"t1\") closely resemble real-world scenarios. They greatly demonstrate the effectiveness of our pipeline from distillation-friendly demonstrations to forecast-aided 4D imitation learning, which is capable of handling dynamic robot perception in complex handover scenarios. We also show visualizations on different methods in Figure 4 (a)(b)." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 304, + 646, + 547, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 646, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 646, + 547, + 713 + ], + "type": "text", + "content": "*Our approach strictly adheres to the simultaneous setting defined in the paper of HandoverSim and HandoverSim2real: the robot moves from the beginning of the handover episode. However, it's noteworthy that HandoverSim2real manually makes their policy hold still in the first 1.5 seconds in the code implementation, deviating from the simultaneous setting definition. To ensure a fair comparison, we reproduce their results in the true simultaneous setting." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16367" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 60, + 561, + 232 + ], + "blocks": [ + { + "bbox": [ + 47, + 60, + 561, + 232 + ], + "lines": [ + { + "bbox": [ + 47, + 60, + 561, + 232 + ], + "spans": [ + { + "bbox": [ + 47, + 60, + 561, + 232 + ], + "type": "table", + "html": "
s0 (Sequential)s0 (Simultaneous)t0t1
STASSTASSTASSTAS
OMG Planner† [42]62.508.3122.5---------
train on s0GA-DDPG [43]50.007.1422.536.814.6623.623.597.3110.346.75.5026.9
Handover-Sim2real [11]75.237.7430.468.756.2335.829.176.2915.052.407.0923.8
Handover-Sim2real* [11]64.357.6126.725.695.4315.028.564.7317.930.605.9816.5
Destination Planning74.319.0122.876.166.9835.225.685.9614.148.48.9415.1
Dense Planning74.779.5419.875.457.3233.027.306.2614.152.39.2415.1
Landmark Planning77.789.2422.379.177.2634.929.636.2315.454.29.0216.6
train on t0GA-DDPG [43]54.767.2624.244.685.3026.524.054.7015.325.505.8614.1
Handover-Sim2real [11]65.977.1829.562.506.0433.533.715.9118.447.106.3524.1
Handover-Sim2real* [11]63.557.5826.538.895.2923.133.314.6421.433.355.8118.4
Destination Planning0.9312.800.016.4812.410.35.968.811.91.6012.030.1
Dense Planning81.489.5121.984.957.4536.338.047.1617.157.908.8518.4
Landmark Planning86.578.8128.085.656.5842.841.436.0122.368.337.7027.9
", + "image_path": "bc3eff356f8890ab4be5d3580bfe55c3f53fb3ff08745488ba1669bce15349c2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 316, + 289, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 289, + 569 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 289, + 569 + ], + "type": "text", + "content": "Results for different Demonstrations Trained on \"s0\" which consists of relatively simple trajectories, demonstrations based on destination planning can offer a rudimentary cue for downstream visuo-motor policy. However, when trained on \"t0\" this strategy may lose focus on the object, leading to a failure in maintaining vision-action correlation and providing minimal gains for vision-friendly learning. There is a significant " + }, + { + "bbox": [ + 46, + 316, + 289, + 569 + ], + "type": "inline_equation", + "content": "73.38\\%" + }, + { + "bbox": [ + 46, + 316, + 289, + 569 + ], + "type": "text", + "content": " / " + }, + { + "bbox": [ + 46, + 316, + 289, + 569 + ], + "type": "inline_equation", + "content": "69.68\\%" + }, + { + "bbox": [ + 46, + 316, + 289, + 569 + ], + "type": "text", + "content": " decrease in success rate in the \"s0\" setting. Additionally, distillation from landmark planning slightly outperforms dense planning in success rate and completes the handover process more quickly in all benchmarks. While dense planning can sustain the success rate to some extent, it slows down the agent and may result in unnatural approaches to objects. To jointly consider the time efficiency and the success rate, we compare the Average Success in methods distilled from these two strategies and find that landmark planning is a more efficient and generalizable approach. For instance, when trained on \"t0\", landmark planning exhibits significant improvements of " + }, + { + "bbox": [ + 46, + 316, + 289, + 569 + ], + "type": "inline_equation", + "content": "6.1\\%" + }, + { + "bbox": [ + 46, + 316, + 289, + 569 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 316, + 289, + 569 + ], + "type": "inline_equation", + "content": "6.5\\%" + }, + { + "bbox": [ + 46, + 316, + 289, + 569 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 316, + 289, + 569 + ], + "type": "inline_equation", + "content": "5.2\\%" + }, + { + "bbox": [ + 46, + 316, + 289, + 569 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 316, + 289, + 569 + ], + "type": "inline_equation", + "content": "9.5\\%" + }, + { + "bbox": [ + 46, + 316, + 289, + 569 + ], + "type": "text", + "content": " across the four testing sets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 575, + 249, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 249, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 249, + 588 + ], + "type": "text", + "content": "4.2. Evaluating on different Dataset Scales" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 594, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 713 + ], + "type": "text", + "content": "We have proved the crucial role of large-scale datasets in handover generalization in Section 4.1. We can also reveal it by scaling down the usage of \"t0\" in GenH2R-Sim which contains 1,000,000 training scenes. With " + }, + { + "bbox": [ + 46, + 594, + 289, + 713 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 594, + 289, + 713 + ], + "type": "text", + "content": " data utilization, we observe a " + }, + { + "bbox": [ + 46, + 594, + 289, + 713 + ], + "type": "inline_equation", + "content": "5.93\\%" + }, + { + "bbox": [ + 46, + 594, + 289, + 713 + ], + "type": "text", + "content": " drop in the success rate on the unseen \"t1\" test set. This result proves the significance of the dataset scale in our imitation learning method. Thanks to our large-scale data and efficient demonstration generation pipeline, concerns about limited datasets hindering generalization are alleviated." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 306, + 315, + 547, + 373 + ], + "blocks": [ + { + "bbox": [ + 46, + 239, + 547, + 307 + ], + "lines": [ + { + "bbox": [ + 46, + 239, + 547, + 307 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 547, + 307 + ], + "type": "text", + "content": "Table 1. Evaluating on different benchmarks. We compare our method against baselines from the test set of HandoverSim [9] benchmark (\"s0 (sequential)\" and \"s0 (simultaneous)\") and our GenH2R-Sim benchmark (\"t0\" and \"t1\"). We use the best-pretrained models from the repositories of GA-DDPG [43] and Handover-Sim2real [11] for evaluation. The results for our method are averaged across 3 random seeds. Note that S means success rate(%) . T means time(s). AS means average success(%) as defined in Equation 2. †: This method [42] is evaluated with ground-truth states and cannot handle dynamic handover like \"s0 (Simultaneous)\", \"t0\" and \"t1\".*: We reproduce the results of HandoverSim2real in the true simultaneous setting as detailed in Section 4.1 to make a fair comparison." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 315, + 547, + 373 + ], + "lines": [ + { + "bbox": [ + 306, + 315, + 547, + 373 + ], + "spans": [ + { + "bbox": [ + 306, + 315, + 547, + 373 + ], + "type": "table", + "html": "
MethodsSTAS
w/o Flow31.665.6717.9
w/o Prediction39.186.1120.7
w/o Flow & Prediction37.045.9320.1
Ours41.436.0122.3
", + "image_path": "c43aca7a791fd1e06d4a5bf72e063a0468d876c2dcfa876690b068dade7bc06b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 380, + 547, + 414 + ], + "lines": [ + { + "bbox": [ + 305, + 380, + 547, + 414 + ], + "spans": [ + { + "bbox": [ + 305, + 380, + 547, + 414 + ], + "type": "text", + "content": "Table 2. Ablations on different modules. \"w/o Flow\" means do not use flow information in the input. \"w/o Prediction\" means do not add prediction loss in the output." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 426, + 400, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 426, + 400, + 440 + ], + "spans": [ + { + "bbox": [ + 306, + 426, + 400, + 440 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 445, + 547, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 547, + 578 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 547, + 578 + ], + "type": "text", + "content": "As shown in Table 2, we prove the effectiveness of our well-designed 4D imitation learning method. The absence of flow information results in a " + }, + { + "bbox": [ + 304, + 445, + 547, + 578 + ], + "type": "inline_equation", + "content": "9.77\\%" + }, + { + "bbox": [ + 304, + 445, + 547, + 578 + ], + "type": "text", + "content": " decrease (predicting without past information adversely affects the model performance). The absence of the prediction task leads to a " + }, + { + "bbox": [ + 304, + 445, + 547, + 578 + ], + "type": "inline_equation", + "content": "2.25\\%" + }, + { + "bbox": [ + 304, + 445, + 547, + 578 + ], + "type": "text", + "content": " decrease, and the absence of both components results in a " + }, + { + "bbox": [ + 304, + 445, + 547, + 578 + ], + "type": "inline_equation", + "content": "4.39\\%" + }, + { + "bbox": [ + 304, + 445, + 547, + 578 + ], + "type": "text", + "content": " decrease. The results demonstrate the model obtains improved performance in leveraging flow information, particularly when tasked with predicting the future object pose. More ablations about our demonstration generation and imitation learning are detailed in the supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 586, + 446, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 446, + 600 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 446, + 600 + ], + "type": "text", + "content": "4.4. Real World Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 605, + 547, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 689 + ], + "type": "text", + "content": "Sim-to-Real Transfer In addition to simulation, we deploy the models trained in GenH2R-Sim on a real robotic platform. Using point cloud input from the wrist-mounted camera, we employ the output 6D egocentric action to update the end effector's target position. A user study compares our method against Handover-Sim2real [11]. The supplementary material provides further details." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "type": "text", + "content": "User Study We recruited 6 users to compare our method (based on landmark planning) and Handover-Sim2real" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16368" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 492, + 521 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 492, + 521 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 492, + 521 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 492, + 521 + ], + "type": "image", + "image_path": "5bd4ab0a985eb662e2b4f955c7931630c2d156a702d31873371906194b88439e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 524, + 528, + 536 + ], + "lines": [ + { + "bbox": [ + 64, + 524, + 528, + 536 + ], + "spans": [ + { + "bbox": [ + 64, + 524, + 528, + 536 + ], + "type": "text", + "content": "Figure 4. Qualitative results. We in detail compare different methods in simulators and deploy them in the real-world platform." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 544, + 288, + 581 + ], + "blocks": [ + { + "bbox": [ + 48, + 544, + 288, + 581 + ], + "lines": [ + { + "bbox": [ + 48, + 544, + 288, + 581 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 288, + 581 + ], + "type": "table", + "html": "
MethodsSimple SettingComplex Setting
Handover-Sim2real56.7%33.3%
Ours90.0%70.0%
", + "image_path": "1c374940059030824c8afde3c7d969889f8374acd9e7f062d3862535a71af059.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 586, + 288, + 609 + ], + "lines": [ + { + "bbox": [ + 46, + 586, + 288, + 609 + ], + "spans": [ + { + "bbox": [ + 46, + 586, + 288, + 609 + ], + "type": "text", + "content": "Table 3. Sim-to-Real Experiments. We report the success rate of our method and HandoverSim2real in 2 different settings." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "content": "across 5 objects in 2 different settings. In the simple setting, users hand each object to the gripper without quick movements. In the complex setting, users execute a relatively long and quick trajectory. The results are reported in Table 3. We observe that our model gets better performance in completing the handover process across various objects and scenarios. Figure 4(c) shows examples of the real-world handover trials." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 545, + 378, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 545, + 378, + 558 + ], + "spans": [ + { + "bbox": [ + 306, + 545, + 378, + 558 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "In this work, we present a novel framework GenH2R for scaling up the learning of human-to-robot handover. We introduce a new simulator GenH2R-Sim and generate a million human handover animations to facilitate generalizable H2R handover learning. We then propose a distillation-friendly demonstration generation method that automatically produces a million high-quality demonstrations suitable for learning. We further introduce a forecast-aided 4D imitation learning method for effective demonstration distillation. Our experiments demonstrate that scaling-up efforts result in substantial improvement of generalizability to novel geometry and complex motion, both in the simulator and the real world." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16369" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 96, + 106, + 108 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 106, + 108 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 106, + 108 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 116, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 116, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 116, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 53, + 116, + 287, + 160 + ], + "type": "text", + "content": "[1] Christoph Bartneck, Tony Belpaeme, Friederike Eyssel, Takayuki Kanda, Merel Keijrsers, and Selma Šabanović. Human-robot interaction: An introduction. Cambridge University Press, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 162, + 288, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 162, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 53, + 162, + 288, + 217 + ], + "type": "text", + "content": "[2] Antonio Bicchi and Vijay Kumar. Robotic grasping and contact: A review. In Proceedings 2000 ICRA. Millennium conference. IEEE international conference on robotics and automation. Symposia proceedings (Cat. No. 00CH37065), pages 348-353. IEEE, 2000. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 218, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 218, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 53, + 218, + 287, + 251 + ], + "type": "text", + "content": "[3] Aude Billard, Sylvain Calinon, Ruediger Dillmann, and Stefan Schaal. Survey: Robot programming by demonstration. Technical report, Springer, 2008. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 253, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 253, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 53, + 253, + 287, + 285 + ], + "type": "text", + "content": "[4] Jeannette Bohg, Antonio Morales, Tamim Asfour, and Dania Kragic. Data-driven grasp synthesis—a survey. IEEE Transactions on robotics, 30(2):289-309, 2013. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 287, + 287, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 287, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 53, + 287, + 287, + 353 + ], + "type": "text", + "content": "[5] Samarth Brahmbhatt, Chengcheng Tang, Christopher D Twigg, Charles C Kemp, and James Hays. Contactpose: A dataset of grasps with object contact and hand pose. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16, pages 361-378. Springer, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 354, + 287, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 354, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 53, + 354, + 287, + 409 + ], + "type": "text", + "content": "[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 411, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 411, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 53, + 411, + 287, + 464 + ], + "type": "text", + "content": "[7] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 468, + 287, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 468, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 53, + 468, + 287, + 543 + ], + "type": "text", + "content": "[8] Yu-Wei Chao, Wei Yang, Yu Xiang, Pavlo Molchanov, Ankur Handa, Jonathan Tremblay, Yashraj S Narang, Karl Van Wyk, Umar Iqbal, Stan Birchfield, et al. Dexycb: A benchmark for capturing hand grasping of objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9044–9053, 2021. 2, 3, 6, 14, 17" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 545, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 545, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 53, + 545, + 287, + 621 + ], + "type": "text", + "content": "[9] Yu-Wei Chao, Chris Paxton, Yu Xiang, Wei Yang, Balakumar Sundaralingam, Tao Chen, Adithyavairavan Murali, Maya Cakmak, and Dieter Fox. Handoversim: A simulation framework and benchmark for human-to-robot object handovers. In 2022 International Conference on Robotics and Automation (ICRA), pages 6941–6947. IEEE, 2022. 1, 2, 3, 6, 7, 12, 14, 16, 17, 18" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "text", + "content": "[10] Sammy Christen, Lan Feng, Wei Yang, Yu-Wei Chao, Otmar Hilliges, and Jie Song. Synh2r: Synthesizing handobject motions for learning human-to-robot handovers. arXiv preprint arXiv:2311.05599, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[11] Sammy Christen, Wei Yang, Claudia Pérez-D'Arpino, Otmar Hilliges, Dieter Fox, and Yu-Wei Chao. Learning human-to-robot handovers from point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "text", + "content": "Pattern Recognition, pages 9654-9664, 2023. 2, 5, 6, 7, 13, 14, 16, 17" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 96, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 161 + ], + "type": "text", + "content": "[12] Gianluca Corsini, Martin Jacquet, Hemjyoti Das, Amr Affi, Daniel Sidobre, and Antonio Franchi. Nonlinear model predictive control for human-robot handover with application to the aerial case. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7597–7604. IEEE, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 163, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 207 + ], + "type": "text", + "content": "[13] Murtaza Dalal, Ajay Mandlekar, Caelan Garrett, Ankur Handa, Ruslan Salakhutdinov, and Dieter Fox. Imitating task and motion planning with visuomotor transformers. arXiv preprint arXiv:2305.16309, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 209, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 251 + ], + "type": "text", + "content": "[14] Yuhao Dong, Zhuoyang Zhang, Yunze Liu, and Li Yi. Nsm4d: Neural scene model based online 4d point cloud sequence understanding. arXiv preprint arXiv:2310.08326, 2023.5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 254, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 254, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 254, + 545, + 308 + ], + "type": "text", + "content": "[15] Frederik Ebert, Yanlai Yang, Karl Schmeckpeper, Bernadette Bucher, Georgios Georgakis, Kostas Daniilidis, Chelsea Finn, and Sergey Levine. Bridge data: Boosting generalization of robotic skills with cross-domain datasets. arXiv preprint arXiv:2109.13396, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 310, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 310, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 308, + 310, + 545, + 353 + ], + "type": "text", + "content": "[16] Clemens Eppner, Arsalan Mousavian, and Dieter Fox. Acronym: A large-scale grasp dataset based on simulation. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 6222-6227. IEEE, 2021. 2, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 355, + 545, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 355, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 308, + 355, + 545, + 408 + ], + "type": "text", + "content": "[17] Hehe Fan, Yi Yang, and Mohan Kankanhalli. Point 4d transformer networks for spatio-temporal modeling in point cloud videos. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14204-14213, 2021. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 411, + 545, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 475 + ], + "type": "text", + "content": "[18] Zicong Fan, Omid Taheri, Dimitrios Tzionas, Muhammed Kocabas, Manuel Kaufmann, Michael J Black, and Otmar Hilliges. Arctic: A dataset for dexterous bimanual handobject manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12943-12954, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 478, + 545, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 478, + 545, + 531 + ], + "spans": [ + { + "bbox": [ + 308, + 478, + 545, + 531 + ], + "type": "text", + "content": "[19] Hao-Shu Fang, Chenxi Wang, Hongjie Fang, Minghao Gou, Jirong Liu, Hengxu Yan, Wenhai Liu, Yichen Xie, and Cewu Lu. Anygrasp: Robust and efficient grasp perception in spatial and temporal domains. IEEE Transactions on Robotics, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 534, + 545, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 534, + 545, + 577 + ], + "spans": [ + { + "bbox": [ + 308, + 534, + 545, + 577 + ], + "type": "text", + "content": "[20] Chelsea Finn, Tianhe Yu, Tianhao Zhang, Pieter Abbeel, and Sergey Levine. One-shot visual imitation learning via meta-learning. In Conference on robot learning, pages 357-368. PMLR, 2017. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 579, + 545, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 579, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 308, + 579, + 545, + 621 + ], + "type": "text", + "content": "[21] Justin Fu, Aviral Kumar, Ofir Nachum, George Tucker, and Sergey Levine. D4rl: Datasets for deep data-driven reinforcement learning. arXiv preprint arXiv:2004.07219, 2020.3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 624, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 677 + ], + "type": "text", + "content": "[22] Caelan Reed Garrett, Rohan Chitnis, Rachel Holladay, Beomjoon Kim, Tom Silver, Leslie Pack Kaelbling, and Tomás Lozano-Pérez. Integrated task and motion planning. Annual review of control, robotics, and autonomous systems, 4:265-293, 2021. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "text", + "content": "[23] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d:" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16370" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "type": "text", + "content": "[24] Huy Ha, Pete Florence, and Shuran Song. Scaling up and distilling down: Language-guided robot skill acquisition. arXiv preprint arXiv:2307.14535, 2023. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 205 + ], + "type": "text", + "content": "[25] Yana Hasson, Gul Varol, Dimitrios Tzionas, Igor Kalevatykh, Michael J Black, Ivan Laptev, and Cordelia Schmid. Learning joint reconstruction of hands and manipulated objects. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11807-11816, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 262 + ], + "type": "text", + "content": "[26] Eric Jang, Alex Irpan, Mohi Khansari, Daniel Kappler, Frederik Ebert, Corey Lynch, Sergey Levine, and Chelsea Finn. Bc-z: Zero-shot task generalization with robotic imitation learning. In Conference on Robot Learning, pages 991-1002. PMLR, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 308 + ], + "type": "text", + "content": "[27] Yi Li, Gu Wang, Xiangyang Ji, Yu Xiang, and Dieter Fox. Deepim: Deep iterative matching for 6d pose estimation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 683-698, 2018. 5, 14" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 309, + 287, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 309, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 287, + 374 + ], + "type": "text", + "content": "[28] Yunze Liu, Yun Liu, Che Jiang, Kangbo Lyu, Weikang Wan, Hao Shen, Boqiang Liang, Zhoujie Fu, He Wang, and Li Yi. Hoi4d: A 4d egocentric dataset for category-level human-object interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21013-21022, 2022. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 376, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 287, + 430 + ], + "type": "text", + "content": "[29] Ajay Mandlekar, Danfei Xu, Josiah Wong, Soroush Nasiriy, Chen Wang, Rohun Kulkarni, Li Fei-Fei, Silvio Savarese, Yuke Zhu, and Roberto Martin-Martin. What matters in learning from offline human demonstrations for robot manipulation. arXiv preprint arXiv:2108.03298, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 433, + 287, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 433, + 287, + 486 + ], + "spans": [ + { + "bbox": [ + 48, + 433, + 287, + 486 + ], + "type": "text", + "content": "[30] Naresh Marturi, Marek Kopicki, Alireza Rastegarpanah, Vijaykumar Rajasekaran, Maxime Adjigble, Rustam Stolkin, Ales Leonidis, and Yasemin Bekiroglu. Dynamic grasp and trajectory planning for moving objects. Autonomous Robots, 43:1241-1256, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 488, + 287, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 488, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 488, + 287, + 520 + ], + "type": "text", + "content": "[31] Michael James McDonald and Dylan Hadfield-Menell. Guided imitation of task and motion planning. In Conference on Robot Learning, pages 630-640. PMLR, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 522, + 215, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 215, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 215, + 533 + ], + "type": "text", + "content": "[32] OpenAI. Gpt-4 technical report, 2023. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 535, + 287, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 535, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 48, + 535, + 287, + 577 + ], + "type": "text", + "content": "[33] Valerio Ortenzi, Akansel Cosgun, Tommaso Pardi, Wesley P Chan, Elizabeth Croft, and Dana Kulić. Object handovers: a review for robotics. IEEE Transactions on Robotics, 37(6): 1855-1873, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 580, + 287, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 287, + 611 + ], + "type": "text", + "content": "[34] Dean A Pomerleau. Alvinn: An autonomous land vehicle in a neural network. Advances in neural information processing systems, 1, 1988. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 613, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 657 + ], + "type": "text", + "content": "[35] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 5, 14" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[36] Patrick Rosenberger, Akansel Cosgun, Rhys Newbury, Jun Kwan, Valerio Ortenzi, Peter Corke, and Manfred Grafinger. Object-independent human-to-robot handovers using real time robotic vision. IEEE Robotics and Automation Letters, 6(1):17-23, 2020. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[37] Szymon Rusinkiewicz and Marc Levoy. Efficient variants of the icp algorithm. In Proceedings third international conference on 3-D digital imaging and modeling, pages 145-152. IEEE, 2001. 5, 14" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 119, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 119, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 119, + 545, + 139 + ], + "type": "text", + "content": "[38] Thomas B Sheridan. Human–robot interaction: status and challenges. Human factors, 58(4):525-532, 2016. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "type": "text", + "content": "[39] Shuran Song, Andy Zeng, Johnny Lee, and Thomas Funkhouser. Grasping in the wild: Learning 6dof closed-loop grasping from low-cost demonstrations. IEEE Robotics and Automation Letters, 5(3):4978-4985, 2020. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 186, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 228 + ], + "type": "text", + "content": "[40] Zachary Teed and Jia Deng. Raft-3d: Scene flow using rigid-motion embeddings. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8375–8384, 2021. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 232, + 545, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 232, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 307, + 232, + 545, + 263 + ], + "type": "text", + "content": "[41] Lirui Wang, Yu Xiang, and Dieter Fox. Manipulation trajectory optimization with online grasp synthesis and selection. In Robotics: Science and Systems (RSS), 2020. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 266, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 266, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 266, + 545, + 308 + ], + "type": "text", + "content": "[42] Lirui Wang, Yu Xiang, and Dieter Fox. Manipulation trajectory optimization with online grasp synthesis and selection. In Proceedings of Robotics: Science and Systems (RSS), 2020. 5, 7, 13, 14" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 311, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 311, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 307, + 311, + 545, + 354 + ], + "type": "text", + "content": "[43] Lirui Wang, Yu Xiang, Wei Yang, Arsalan Mousavian, and Dieter Fox. Goal-auxiliary actor-critic for 6d robotic grasping with point clouds. In Conference on Robot Learning, pages 70-80. PMLR, 2022. 5, 7, 14, 16, 17" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 356, + 545, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 356, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 307, + 356, + 545, + 408 + ], + "type": "text", + "content": "[44] Lirui Wang, Yiyang Ling, Zhecheng Yuan, Mohit Shridhar, Chen Bao, Yuzhe Qin, Bailin Wang, Huazhe Xu, and Xiaolong Wang. Gensim: Generating robotic simulation tasks via large language models. arXiv preprint arXiv:2310.01361, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 411, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 464 + ], + "type": "text", + "content": "[45] Ruicheng Wang, Jialiang Zhang, Jiayi Chen, Yinzhen Xu, Puhao Li, Tengyu Liu, and He Wang. Dexgraspnet: A large-scale robotic dexterous grasp dataset for general objects based on simulation. arXiv preprint arXiv:2210.02697, 2022.3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 468, + 545, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 468, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 307, + 468, + 545, + 532 + ], + "type": "text", + "content": "[46] Ruicheng Wang, Jialiang Zhang, Jiayi Chen, Yinzhen Xu, Puhao Li, Tengyu Liu, and He Wang. Dexgraspnet: A large-scale robotic dexterous grasp dataset for general objects based on simulation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 11359-11366. IEEE, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 535, + 545, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 535, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 307, + 535, + 545, + 588 + ], + "type": "text", + "content": "[47] Yufei Wang, Zhou Xian, Feng Chen, Tsun-Hsuan Wang, Yian Wang, Katerina Fragkiadaki, Zackory Erickson, David Held, and Chuang Gan. Robogen: Towards unleashing infinite data for automated robot learning via generative simulation. arXiv preprint arXiv:2311.01455, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 590, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 634 + ], + "type": "text", + "content": "[48] Hao Wen, Yunze Liu, Jingwei Huang, Bo Duan, and Li Yi. Point primitive transformer for long-term 4d point cloud video understanding. In European Conference on Computer Vision, pages 19-35. Springer, 2022. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "type": "text", + "content": "[49] Wei Yang, Chris Paxton, Arsalan Mousavian, Yu-Wei Chao, Maya Cakmak, and Dieter Fox. Reactive human-to-robot handovers of arbitrary objects. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 3118–3124. IEEE, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 691, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 712 + ], + "type": "text", + "content": "[50] Ruolin Ye, Wenqiang Xu, Zhendong Xue, Tutian Tang, Yanfeng Wang, and Cewu Lu. H2o: A benchmark for visual" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16371" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 272 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "human-human object handover analysis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15762-15771, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 149 + ], + "type": "text", + "content": "[51] Gu Zhang, Hao-Shu Fang, Hongjie Fang, and Cewu Lu. Flexible handover with real-time robust dynamic grasp trajectory generation. arXiv preprint arXiv:2308.15622, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 205 + ], + "type": "text", + "content": "[52] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 287, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 287, + 272 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 287, + 272 + ], + "type": "text", + "content": "[53] Tianhao Zhang, Zoe McCarthy, Owen Jow, Dennis Lee, Xi Chen, Ken Goldberg, and Pieter Abbeel. Deep imitation learning for complex manipulation tasks from virtual reality teleoperation. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pages 5628-5635. IEEE, 2018. 3" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16372" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/GenN2N_ Generative NeRF2NeRF Translation/720fe634-ca22-4125-bab8-c49c422e71c2_content_list.json b/2024/GenN2N_ Generative NeRF2NeRF Translation/720fe634-ca22-4125-bab8-c49c422e71c2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..30912f49a0ffb8d7d261dee38ef5d12fd3ea6e06 --- /dev/null +++ b/2024/GenN2N_ Generative NeRF2NeRF Translation/720fe634-ca22-4125-bab8-c49c422e71c2_content_list.json @@ -0,0 +1,2321 @@ +[ + { + "type": "text", + "text": "GenN2N: Generative NeRF2NeRF Translation", + "text_level": 1, + "bbox": [ + 248, + 130, + 722, + 150 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiangyue Liu $^{1}$ Han Xue $^{2}$ Kunming Luo $^{1}$ Ping Tan $^{1\\dagger}$ Li Yi $^{2,3,4\\dagger}$", + "bbox": [ + 212, + 178, + 759, + 200 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$^{1}$ Hong Kong University of Science and Technology $^{2}$ Tsinghua University", + "$^{3}$ Shanghai Artificial Intelligence Laboratory $^{4}$ Shanghai Qi Zhi Institute" + ], + "bbox": [ + 183, + 214, + 785, + 252 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8ca0ac5a4e3a682fee5599c6ab5ed0c5060ff70bc4ee10d2cfb5617aaa7aa150.jpg", + "image_caption": [ + "Input", + "Output 1", + "Output 2", + "Output 3" + ], + "image_footnote": [], + "bbox": [ + 138, + 286, + 478, + 440 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d2ed03faeb021d1cd882c010f8c11c49e4e6d3eb5f10640473784caef20e3b43.jpg", + "image_caption": [ + "Input", + "Input zoom in", + "Output", + "Output zoom in", + "Figure 1. We introduce GenN2N, a unified framework for NeRF-to-NeRF translation, enabling a range of 3D NeRF editing tasks, including text-driven editing, colorization, super-resolution, inpainting, etc. We show at least two rendering views of edited NeRF scenes at inference time. Given a 3D NeRF scene, GenN2N can produce high-quality editing results with suitable multi-view consistency." + ], + "image_footnote": [], + "bbox": [ + 138, + 449, + 478, + 556 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0c4653d18e861ac74f52cc629cb0c01ddec6feb4d59d833ca8a08d8348335202.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 287, + 581, + 364 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ced5e4e1b34e9e675d51c4effdee96482ae89d6c6181312791da3b8b44b820c1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 584, + 287, + 661, + 364 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bbe829216193b8afbe4b78fbd624a9f4100a2203e96ea105e3c7992549d36d9d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 665, + 287, + 743, + 364 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d73bb4246a5bbc4aaa5627ac0a4a65ec97e627397013eac5045b5c52ca9ff5c2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 746, + 287, + 825, + 364 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/334c92776c42b04e72183a28f470290c4c36b14817becffc441caecd7a1a763d.jpg", + "image_caption": [ + "aannnnnne aannnnnne", + "Input" + ], + "image_footnote": [], + "bbox": [ + 486, + 364, + 581, + 440 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/92fe8edd7c667b81fa33654f47f1efc84d0933ea7f8e8393eb274fcb58d702b2.jpg", + "image_caption": [ + "Output 1" + ], + "image_footnote": [], + "bbox": [ + 584, + 364, + 661, + 440 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3db7a8a2143ef22b17deadc2495d63a5b98d3aa22ffd082fd475b3b581374966.jpg", + "image_caption": [ + "Output 2" + ], + "image_footnote": [], + "bbox": [ + 665, + 364, + 743, + 440 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/de71d79b9a5a341e7ad1909c3bb5915cc710552ad11da465162dff7bb897b6b2.jpg", + "image_caption": [ + "Output 3" + ], + "image_footnote": [], + "bbox": [ + 746, + 364, + 825, + 440 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b1f6eff79e0b92e2dd375eb6d1e4695dd5221a8a1c1de8b36d184a78266158d4.jpg", + "image_caption": [ + "View 1", + "View 2" + ], + "image_footnote": [], + "bbox": [ + 486, + 449, + 581, + 556 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/76ce41d1ad5c40dfef353ad597af3b9249363f203cdd96812838d74c78398c01.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 584, + 449, + 661, + 502 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9b8679bb30949d69e7341a3fcf6c96328e98600ff1eae0f71a07aa65eed6ff82.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 665, + 449, + 743, + 502 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3c924143fbd79fdff13d90837596c4043014634b0243cf5d20145281e3eea6c2.jpg", + "image_caption": [ + "View 3" + ], + "image_footnote": [], + "bbox": [ + 665, + 503, + 743, + 556 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0bd289d18c6ebb9f6a2c63c2d2f309661a17c8aeeb751f76321ba5ed15242781.jpg", + "image_caption": [ + "View 4" + ], + "image_footnote": [], + "bbox": [ + 746, + 449, + 825, + 556 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 625, + 313, + 640 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present GenN2N, a unified NeRF-to-NeRF translation framework for various NeRF translation tasks such as text-driven NeRF editing, colorization, super-resolution, inpainting, etc. Unlike previous methods designed for individual translation tasks with task-specific schemes, GenN2N achieves all these NeRF editing tasks by employing a plug-and-play image-to-image translator to perform editing in the 2D domain and lifting 2D edits into the 3D NeRF space. Since the 3D consistency of 2D edits may not be assured, we propose to model the distribution of the underlying 3D edits through a generative model that can cover all possible edited NeRFs. To model the distribution of 3D edited NeRFs from 2D edited images, we carefully design a VAE-GAN that encodes images while decoding NeRFs. The la", + "bbox": [ + 75, + 657, + 472, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tent space is trained to align with a Gaussian distribution and the NeRFs are supervised through an adversarial loss on its renderings. To ensure the latent code does not depend on 2D viewpoints but truly reflects the 3D edits, we also regularize the latent code through a contrastive learning scheme. Extensive experiments on various editing tasks show GenN2N, as a universal framework, performs as well or better than task-specific specialists while possessing flexible generative power. More results on our project page: https://xiangyueliu.github.io/GenN2N/.", + "bbox": [ + 500, + 626, + 893, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 811, + 632, + 829 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Over the past few years, Neural radiance fields (NeRFs) [23] have brought a promising paradigm in the realm of 3D reconstruction, 3D generation, and novel view synthesis due to their unparalleled compactness,", + "bbox": [ + 498, + 839, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding authors.", + "bbox": [ + 94, + 886, + 227, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5105", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "high quality, and versatility. Extensive research efforts have been devoted to creating NeRF scenes from 2D images [4, 20, 22, 36, 38] or just text [13, 26] input. However, once the NeRF scenes have been created, these methods often lack further control over the generated geometry and appearance. NeRF editing has therefore become a notable research focus recently.", + "bbox": [ + 75, + 90, + 470, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing NeRF editing schemes are usually task-specific. For example, researchers have developed NeRF-SR [34], NeRF-In [19], PaletteNeRF [16] for NeRF super-resolution, inpainting, and color-editing respectively. These designs require a significant amount of domain knowledge for each specific task. On the other hand, in the field of 2D image editing, a growing trend is to develop universal image-to-image translation methods to support versatile image editing [25, 29, 43]. By leveraging foundational 2D generative models, e.g., stable diffusion [28], these methods achieve impressive editing results without task-specific customization or tuning. We then ask the question: can we conduct universal NeRF editing leveraging foundational 2D generative models as well?", + "bbox": [ + 75, + 199, + 470, + 411 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The first challenge is the representation gap between NeRFs and 2D images. It is not intuitive how to leverage image editing tools to edit NeRFs. A recent text-driven NeRF editing method [9] has shed some light on this. The method adopts a \"render-edit-aggregate\" pipeline. Specifically, it gradually updates a NeRF scene by iteratively rendering multi-view images, conducting text-driven visual editing on these images, and finally aggregating the edits in the NeRF scene. It seems that replacing the image editing tool with a universal image-to-image translator could lead to a universal NeRF editing method. However, the second challenge would then come. Image-to-image translators usually generate diverse and inconsistent edits for different views, e.g. turning a man into an elf might or might not put a hat on his head, making edits aggregation intricate. Regarding this challenge, Instruct-NeRF2NeRF [9] presents a complex optimization technique to pursue unblurred NeRF with inconsistent multi-view edits. Due to its complexity, the optimization cannot ensure the robustness of the outcomes. Additionally, the unique optimization outcome fails to reflect the stochastic nature of NeRF editing. Users typically anticipate a variety of edited NeRFs just like the diverse edited images.", + "bbox": [ + 75, + 414, + 468, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle the challenges above, we propose GenN2N, a unified NeRF-to-NeRF translation framework for various NeRF editing tasks such as text-driven editing, colorization, super-resolution, inpainting (see Fig. 1). In contrast to Instruct-NeRF2NeRF which adopts a \"render-edit-aggregate\" pipeline, we first render a NeRF scene into multi-view images, then exploit an image-to-image translator to edit different views, and finally learn a generative model to depict the distribution of NeRF edits. Instead of", + "bbox": [ + 75, + 763, + 470, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "aggregating all the image edits to form a single NeRF edit, our key idea is to embrace the stochastic nature of content editing by modeling the distribution of the edits in the 3D NeRF space.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically given a NeRF model or its multi-view images, along with the editing goal, we first generate edited multi-view images using a plug-and-play image-to-image translator. Each view corresponds to a unique 3D edit with some geometry or appearance variations. Conditioned on the input NeRF, GenN2N trains a conditional 3D generative model to reflect such content variations. At the core of GenN2N, we design a 3D VAE-GAN that incorporates a differentiable volume renderer to connect 3D content creation with 2D GAN losses, ensuring that the inconsistent multi-view renderings can still help each other regarding 3D generation. Moreover, we introduce a contrastive learning loss to ensure that the 3D content variation can be successfully understood just from edited 2D images without being influenced by the camera viewpoints. During inference, users can simply sample from the conditional generative model to obtain various 3D editing results aligned with the editing goal. We have conducted experiments on human, items, indoor and outdoor scenes for various editing tasks such as text-driven editing, colorization, superresolution and inpainting, demonstrating the effectiveness of GenN2N in supporting diverse NeRF editing tasks while keeping the multi-view consistency of the edited NeRF.", + "bbox": [ + 496, + 152, + 892, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We summarize the contribution of this paper as follows,", + "bbox": [ + 519, + 500, + 887, + 513 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A generative NeRF-to-NeRF translation formulation for the universal NeRF editing task together with a generic solution;", + "- a 3D VAE-GAN framework that can learn the distribution of all possible 3D NeRF edits corresponding to the a set of input edited 2D images;", + "- a contrastive learning framework that can disentangle the 3D edits and 2D camera views from edited images;", + "- extensive experiments demonstrating the superior efficiency, quality, and diversity of the NeRF-to-NeRF translation results." + ], + "bbox": [ + 500, + 513, + 890, + 678 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 694, + 640, + 709 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "NeRF Editing. Previous works such as EditNeRF [21] propose a conditional neural field that enables shape and appearance editing in the latent space. PaletteNeRF [16, 37] focuses on controlling color palette weights to manipulate appearance. Other approaches utilize bounding boxes [41], meshes [39], point clouds [6], key points [45], or feature volumes [18] to directly manipulate the spatial representation of NeRF. However, these methods either heavily rely on user interactions or have limitations in terms of spatial deformation and color transfer capabilities.", + "bbox": [ + 496, + 719, + 890, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "NeRF Stylization. Images-referenced stylization [7, 12, 42] often prioritize capturing texture style rather than de", + "bbox": [ + 496, + 869, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5106", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d47cf812fb1b9fdcafce9ba7a0f9bd66826f1f7cd6321f18674386b1d0abcafb.jpg", + "image_caption": [ + "Figure 2. Overview of GenN2N. We first edit the source image set $\\{\\mathbf{I}_i\\}_{i=0}^{N-1}$ using 2D image-to-image translation methods, e.g., text-driven editing, colorization, zoom out, etc. For each view $i \\in [0, N-1]$ , we generate $M$ edited images, resulting in a group of translated image set $\\{\\{\\mathbf{S}_i^j\\}_{j=0}^{M-1}\\}_{i=0}^{N-1}$ . Then we use the Latent Distill Module to learn $M \\times N$ edit code vectors from the translated image set, which serve as the input of the translated NeRF. To optimize our GenN2N, we design four loss functions: a KL loss to constrain the latent vectors to a Gaussian distribution; and $\\mathcal{L}_{\\mathrm{recon}}$ , $\\mathcal{L}_{\\mathrm{adv}}$ and $\\mathcal{L}_{\\mathrm{contr}}$ to optimize the appearance and geometry of the translated NeRF. At inference, we can sample a latent vector $\\mathbf{z}$ from Gaussian distribution and render a corresponding multi-view consistent 3D scene with high quality." + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 890, + 281 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tailed content, resulting in imprecise editing appearance of NeRF only. Text-guided works [33, 35], on the other hand, apply contrastive losses based on CLIP [27] to achieve the desired edits. While text references usually describe the global characteristics of the edited results, instructions offer a more convenient and precise expression.", + "bbox": [ + 75, + 402, + 470, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Instruct-driven NeRF editing. Among numerous image-to-image translation works, InstructPix2Pix [2] stands out by efficiently editing images following instructions. It leverages large pre-trained models in the language and image domains [3, 28] to generate paired data (before and after editing) for training. While editing NeRF solely based on edited images is problematic due to multi-view inconsistency. To address this, an intuitive yet heavy approach [9] is to iteratively edit the image and optimize NeRF. In addition, NeRF-Art [35] and DreamEditor [46] adopt a CLIP-based contrastive loss [27] and score distillation sampling [26] separately to supervise the optimization of editing NeRF. Inspired by Generative Radiance Fields [5, 30], We capture various possible NeRF editing in the generative space to solve it.", + "bbox": [ + 75, + 494, + 470, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 738, + 168, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a NeRF scene, we present a unified framework GenN2N to achieve various editing on the 3D scene leveraging geometry and texture priors from 2D image editing methods, such as text-driven editing, colorization, superresolution, inpainting, etc. While a universal image-to-image translator can theoretically accomplish these 2D editing tasks, we actually use a state-of-the-art translator for each task. Therefore, we formulate each 2D image editing method as a plug-and-play image-to-image translator", + "bbox": [ + 75, + 763, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and all NeRF editing tasks as our universal NeRF-to-NeRF translation, in which the given NeRF is translated into NeRF scenes with high rendering quality and 3D geometry consistency according to the user-selected editing target. The overview of GenN2N is illustrated in Fig. 2, we first perform image-to-image translation in the 2D domain and then lift 2D edits to 3D and achieve NeRF-to-NeRF translation.", + "bbox": [ + 496, + 402, + 892, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given $N$ multi-view images $\\{\\mathbf{I}_i\\}_{i = 0}^{N - 1}$ of a scene, we first use Nerfstudio [32] to train the original NeRF. Then we use a plug-and-play image-to-image translator to edit these source images. However, the content generated by the 2D translator may be inconsistent among multi-view images. For example, using different initial noise, the 2D translator [1] may generate different content for image editing, which makes it difficult to ensure the 3D consistency between different view directions in the 3D scene.", + "bbox": [ + 496, + 510, + 892, + 647 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To ensure the 3D consistency and rendering quality, we propose to model the distribution of the underlying 3D edits through a generative model that can cover all possible edited NeRFs, by learning an edit code for each edited image so that the generated content can be controlled by this edit code during the NeRF-to-NeRF translation process.", + "bbox": [ + 496, + 650, + 893, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For each view $\\in [0,N - 1]$ , we generate $M$ edited images, resulting in a group of the translated image set $\\{\\{\\mathbf{S}_i^j\\}_{j = 0}^{M - 1}\\}_{i = 0}^{N - 1}$ . Then we design a Latent Distill Module described in Sec. 3.1 to map each translated image $\\mathbf{S}_i^j$ into an edit code vector $\\mathbf{z}_i^j$ and design a KL loss $\\mathcal{L}_{\\mathrm{KL}}$ to constrain those edit code vectors to a Gaussian distribution. Conditioned on the edit code $\\mathbf{z}_i^j$ , we perform NeRF-to-NeRF translation in Sec. 3.2 by rendering multi-view images $\\{\\mathbf{C}_i\\}_{i = 0}^{N - 1}$ and optimize the translated NeRF by three loss functions: the reconstruction loss $\\mathcal{L}_{\\mathrm{recon}}$ , the adversarial", + "bbox": [ + 496, + 744, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5107", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/360c66cdad8c1dac2d8d0bec7f5be42a74c2b342e866fcd367df72d5cec8fdaf.jpg", + "image_caption": [ + "Figure 3. Illustration of our proposed contrastive loss functions. Regarding the multi-view rendered images $\\mathbf{C}_i^j$ and $\\mathbf{C}_l^j$ sharing the same edit code, we resend them to our Latent Distill Module to extract $\\mathbf{z}_i^j$ and $\\mathbf{z}_l^j$ , and aggregate them via $\\mathcal{L}_{\\mathrm{contr}}^{\\mathrm{att}}$ . In addition, for $\\mathbf{S}_i^k$ whose editing style varies from $\\mathbf{S}_i^j$ , $\\mathcal{L}_{\\mathrm{contr}}^{\\mathrm{rep}}$ increases the distance between edit codes of them." + ], + "image_footnote": [], + "bbox": [ + 101, + 89, + 449, + 299 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a9f17cc1a480904a0523620ddc0d9d809312deb467bf9480bd9e57c4f35e42ae.jpg", + "image_caption": [ + "Figure 4. Illustration of our proposed conditional adversarial loss functions. Our conditional discriminator distinguishes artifacts such as blur and distortion in novel-view rendered image $\\mathbf{C}_l^j$ compared with target image $\\mathbf{S}_l^j$ . $\\mathbf{S}_l^j$ and $\\mathbf{S}_l^k$ are edited with same view but various styles, the latter serves as the condition to concatenate with $\\mathbf{C}_l^j$ and $\\mathbf{S}_l^j$ and manufacture fake and real pairs." + ], + "image_footnote": [], + "bbox": [ + 89, + 412, + 457, + 525 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "loss $\\mathcal{L}_{\\mathrm{AD}}$ , and the contrastive loss $\\mathcal{L}_{\\mathrm{contr}}$ . After the optimization of the translated NeRF, as described in Sec. 3.3, we can sample an edit code $\\mathbf{z}$ from Gaussian distribution and render the corresponding edited 3D scene with high quality and multi-view consistency in the inference stage.", + "bbox": [ + 75, + 650, + 468, + 726 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Latent Distill Module", + "text_level": 1, + "bbox": [ + 76, + 739, + 277, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Image Translation. As illustrated in Fig. 2, GenN2N is a unified framework for NeRF-to-NeRF translation, in which the core is to perform a 2D image-to-image translation and lift 2D edits into universal 3D NeRF-to-NeRF translation. Given the source multi-view image set $\\{\\mathbf{I}_i\\}_{i=0}^{N-1}$ of a NeRF scene, we first perform image editing M times for each view using a plug-and-play 2D image-to-image translator, producing a group of translated image set $\\{\\{\\mathbf{S}_i^j\\}_{j=0}^{M-1}\\}_{i=0}^{N-1}$ . In this paper, we use several 2D translation tasks to show", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the adaptability of our GenN2N: text-driven editing, superresolution, colorization and inpainting. For more details about those 2D image editing methods, please refer to the supplementary materials.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Edit Code. Since 2D image-to-image translation may generate different content even with the same editing target, causing the inconsistency problem in the 3D scene. We propose to map each edited image $\\mathbf{S}_i^j$ into a latent feature vector named edit code to characterize these diverse editorings. We employ the off-the-shelf VAE encoder from stable diffusion [28] to extract the feature from $\\mathbf{S}_i^j$ and then apply a tiny MLP network to produce this edit code $\\mathbf{z}_i^j \\in \\mathbb{R}^{64}$ . During the training process, we keep the pre-trained encoder fixed and only optimize the parameters of the tiny MLP network. This mapping process can be formulated as follows:", + "bbox": [ + 496, + 152, + 892, + 320 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {i} ^ {j} = \\mathcal {D} \\left(\\mathbf {S} _ {i} ^ {j}\\right) = \\mathcal {M} \\left(\\mathcal {E} \\left(\\mathbf {S} _ {i} ^ {j}\\right)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 332, + 890, + 352 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{D}$ represent this mapping process, $\\mathcal{E}$ is the fixed encoder, and $\\mathcal{M}$ is the learnable tiny MLP.", + "bbox": [ + 496, + 364, + 890, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "KL loss. In order to facilitate effective sampling of the edit code so as to control the editing diversity of our NeRF-to-NeRF translation, we need to constrain the edit code to a well-defined distribution. Thus we design a KL loss to encourage $\\mathbf{z}_i^j$ to approximate a Gaussian distribution:", + "bbox": [ + 496, + 396, + 890, + 472 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {K L}} = \\mathbb {E} _ {\\mathbf {S} \\in \\{\\left\\{\\mathbf {S} _ {i} ^ {j} \\right\\} _ {j = 0} ^ {M - 1} \\} _ {i = 0} ^ {N - 1}} [ P (\\mathbf {z} _ {\\text {n o r m a l}}) \\log \\left(\\frac {P \\left(\\mathbf {z} _ {\\text {n o r m a l}}\\right)}{P \\left(\\mathcal {D} (\\mathbf {S})\\right)}\\right) ], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 503, + 484, + 890, + 531 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $P(\\mathbf{z}_{\\text{normal}})$ denotes probability distribution of the standard Gaussian distribution in $\\mathbb{R}^{64}$ and $P(\\mathcal{D}(\\mathbf{S}))$ means probability distribution of the extracted edit codes.", + "bbox": [ + 496, + 531, + 890, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Contrastive loss. It is not assured that edit codes $\\mathbf{z}$ obtained from the Latent Distill Module contain only the editing information while excluding viewpoint-related effects. However, since the translated NeRF utilizes $\\mathbf{z}$ to edit scenes, it yields instability if $\\mathbf{z}$ violently changes given images that are similar in appearance but different in viewpoints. To ensure the latent code does not depend on 2D viewpoints but truly reflects the 3D edits, we regularize the latent code through a contrastive learning scheme. Specifically, we reduce the distance between edit codes of different-view rendered images from a translated NeRF that share the same edit code, while increasing the distance between same-view images that are multi-time edited by the 2D image-to-image translator. As illustrated in Fig. 3, given an edit code $\\mathbf{z}_i^j$ extracted from the $i$ -th input view at the $j$ -th edited image $\\mathbf{S}_i^j$ , we render multi-view images $\\{\\mathbf{C}_l^j\\}_{l=0}^{N-1}$ using the translated NeRF conditioned on $\\mathbf{z}_i^j$ . Then we employ contrastive learning to encourage the edit code $\\mathbf{z}_i^j$ to be close to $\\{\\mathbf{z}_l^j\\}_{l=0}^{N-1}$ extracted from $\\{\\mathbf{C}_l^j\\}_{l=0}^{N-1}$ , while being distinct from the edit codes $\\{\\mathbf{z}_i^k\\}_{k=0}^{M-1}$ extracted from $\\{\\mathbf{S}_i^k\\}_{k=0}^{M-1}$ , where $k \\neq j$ .", + "bbox": [ + 496, + 577, + 892, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5108", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2a59429c383d15a77cd2d8111c9489f90c9975e3773858f91cdf8e17b9ca6414.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 106, + 168, + 204 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a85184f26f80b850d50b0ef12f278ac2b2ec723cd1f0ed303db3fe012b93d0d2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 181, + 106, + 267, + 203 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/98732b7a251c45d6ae2e6607ba03db46cfbe6812cfe8939993d4d23aee91b90e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 269, + 106, + 354, + 203 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f5697f4c13ae0b6187525bea0e5afce085e2f6b25e696ed76b1ff33e0b9cabb1.jpg", + "image_caption": [ + "\"Make it an Egyptian sculpture\"" + ], + "image_footnote": [], + "bbox": [ + 359, + 106, + 444, + 203 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5551ee4ed2848b775519e0d428a75b8c161bb2658b4b57c22de51584781157e3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 447, + 106, + 532, + 203 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0c076c18fb344ee4ed1f89a07b0027a2a2813dfff2fe39aec549fa29ecb7bddd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 535, + 106, + 620, + 203 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/bdf6fd6dceefa3cdcfbb5cb2e621dffc633583c58826e844f8c3836e76f977bb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 624, + 106, + 709, + 203 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/27d3cb7179f57174f11a26b833c9c166e3bb171a9b7440c44ba41a254b2148c0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 715, + 106, + 800, + 203 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/68c72715214e058cb0f2e3267afddc95bf1a646ba91ff7da689511af97d350a7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 803, + 106, + 887, + 203 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e25597512fa654b4a23cb6294a3ae57b7139848cb17c43e70f30c65fb8e2768c.jpg", + "image_caption": [ + "Original NeRF" + ], + "image_footnote": [], + "bbox": [ + 83, + 222, + 168, + 320 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/11939d6053dfc042d19524d56e7c064b6832670414dca9b13dbba17776f35179.jpg", + "image_caption": [ + "Inference 1" + ], + "image_footnote": [], + "bbox": [ + 183, + 222, + 267, + 320 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0308bd078b45e6b9fa17bb4b800cddd9452bd3877c20c917301b8d073103d902.jpg", + "image_caption": [ + "Figure 5. Text-Driven Editing. We sample 4 inference results for both text-driven editing tasks. The diversity of geometry and appearance showcases awesome generative ability of GenN2N, on the premise of maintaining the 3D consistency between different viewpoints." + ], + "image_footnote": [], + "bbox": [ + 269, + 222, + 354, + 320 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c6dafb7f106bd694903804ccc5d18a5615d7996d3a7b6f440088be450407316b.jpg", + "image_caption": [ + "Inference 2" + ], + "image_footnote": [], + "bbox": [ + 359, + 222, + 444, + 320 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fa33bc58d5b4297395423c609e762e0a71d75f3817b33c4935b04ddb246b1527.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 447, + 222, + 532, + 320 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8842118892dba43b2f3e21b506c5b8f6939f5d54979913852e941058b0b11156.jpg", + "image_caption": [ + "Inference 3" + ], + "image_footnote": [], + "bbox": [ + 535, + 222, + 620, + 320 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0d807275c45433d989082d0d688d823214358915658926e02c9694f2e47ab562.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 624, + 222, + 709, + 320 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8afa615510a22bd70d4b97463a0f0c94960329822ac70b51e206cc9cab6f85e4.jpg", + "image_caption": [ + "Inference 4" + ], + "image_footnote": [], + "bbox": [ + 715, + 222, + 800, + 320 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/693f9a1bbddab8ae2ec0745817b0d3445cc5bbae8e5299b157be7e5693b66b72.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 803, + 222, + 887, + 320 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/95d490a396cb3401e2eedb9da334e77f98f2c15cde96bd6215e5343719112fb7.jpg", + "image_caption": [ + "Original NeRF" + ], + "image_footnote": [], + "bbox": [ + 81, + 390, + 209, + 541 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0d31b9a9f2046581c21fcd7b55487da24655e421dd18554eaa002695c3724977.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 390, + 352, + 465 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/12de9452394cc1cdc540daee145d15972b4db943377d7993e27a777c2c3402b3.jpg", + "image_caption": [ + "Inference 1" + ], + "image_footnote": [], + "bbox": [ + 223, + 467, + 351, + 541 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7cea362638cc036d7a0c716a306bef32dd098ac75ece1fbcd79de2a5d2f5a33f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 357, + 390, + 485, + 465 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/925b7724b83d034c415d74f9012ec24b1ed4f5a702b33d18c945a6aa46c860f7.jpg", + "image_caption": [ + "Inference 2" + ], + "image_footnote": [], + "bbox": [ + 357, + 465, + 485, + 541 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b99df597dd4df24c43683a6ea3beeed898412fba61220d04390636665d35fda9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 390, + 619, + 465 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8bb7696f43673e9f7661b9a5c4f3659c75fa815ee89d7c604d24d93b7e46b038.jpg", + "image_caption": [ + "Inference 3", + "Figure 6. Colorization. Our method colorizes the gray-scale 3D scene consistently across views. By changing the edit code during inference, diverse colorized scenes can be rendered with satisfying photorealism and reasonably rich colors." + ], + "image_footnote": [], + "bbox": [ + 491, + 467, + 619, + 541 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/33f3a99405255176ebb939334391fd46cf6f69ce46e18f611595c6c978a2448a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 625, + 390, + 753, + 465 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6d1d848ca2fab9504e9191ad8fe3827df1a0e985deb1b96a43c7228875fe0f52.jpg", + "image_caption": [ + "Inference 4" + ], + "image_footnote": [], + "bbox": [ + 625, + 465, + 753, + 541 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/03192cd3579b71dc28ba8cd7d112187da2cdf2576eba6c5ef41aef4ecfd09b2e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 759, + 390, + 887, + 465 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/91f4e08eca197464bc3c0f30b58ad3525ad06b5bd92f2e83ec33615b442b4cd7.jpg", + "image_caption": [ + "Inference 5" + ], + "image_footnote": [], + "bbox": [ + 759, + 467, + 887, + 541 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, our contrastive loss is designed as follows:", + "bbox": [ + 94, + 621, + 464, + 636 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {c o n t r}} = \\mathcal {L} _ {\\text {c o n t r}} ^ {\\text {a t t}} + \\mathcal {L} _ {\\text {c o n t r}} ^ {\\text {r e p}} \\\\ = \\sum_ {l = 0} ^ {N - 1} \\left\\| \\mathbf {z} _ {i} ^ {j} - \\mathbf {z} _ {l} ^ {j} \\right\\| _ {2} ^ {2} + \\sum_ {k = 0} ^ {M - 1} \\max (0, \\alpha - \\left\\| \\mathbf {z} _ {i} ^ {j} - \\mathbf {z} _ {i} ^ {k} \\right\\| _ {2} ^ {2}), \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 643, + 468, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha$ represents the margin that encourages the difference in features, and $k\\neq j$", + "bbox": [ + 75, + 719, + 467, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. NeRF-to-NeRF translation", + "text_level": 1, + "bbox": [ + 76, + 756, + 318, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Translated NeRF. After 2D image-to-image translation, we need to lift these 2D edits to the 3D NeRF. For this purpose, we propose to modify the original NeRF as a translated NeRF that takes the edit code $\\mathbf{z}$ as input and generates the translated 3D scene according to the edit code. We refer readers to the supplementary for more details about the network architecture.", + "bbox": [ + 75, + 779, + 468, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Reconstruction loss. Given an edit code $\\mathbf{z}_i^j$ extracted from", + "bbox": [ + 76, + 885, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the edited image $\\mathbf{S}_i^j$ , we can generate a translated NeRF to render $\\mathbf{C}_i^j$ from the same viewpoint. Then we define the reconstruction loss as the L1 normalization and Learned Perceptual Image Patch Similarity (LPIPS) [44] between the rendered image $\\mathbf{C}_i^j$ and the edited image $\\mathbf{S}_i^j$ as follows:", + "bbox": [ + 496, + 619, + 890, + 699 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {r e c o n}} = \\mathcal {L} _ {\\mathrm {L} 1} + \\mathcal {L} _ {\\mathrm {L P I P S}} \\\\ = \\left\\| \\mathbf {C} _ {i} ^ {j} - \\mathbf {S} _ {i} ^ {j} \\right\\| _ {1} + L P I P S [ \\mathcal {P} (\\mathbf {C} _ {i} ^ {j}) - \\mathcal {P} (\\mathbf {S} _ {i} ^ {j}) ], \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 709, + 890, + 755 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{P}$ means a patch sampled from the image. Note that due to the lack of 3D consistency of the edited multiview image, the supervision of the edited image from other viewpoints $\\{\\mathbf{S}_l^j\\}_{l\\neq i}$ will lead to conflicts in pixel-space optimization. Therefore, we only employ reconstruction loss on the same view image $\\mathbf{S}_i^j$ to optimize the translated NeRF. Adversarial loss. Since the 3D consistency of edited multiview images is not assured, relying solely on the reconstruction loss on the same view often leads to blurry or distorted", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "\"What would he look like as a bearded man\"", + "bbox": [ + 354, + 88, + 642, + 102 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5109", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1903030e864e1d3f55178a367dcc206654f9470b734b5ec7c6524603a9755c09.jpg", + "image_caption": [ + "Original NeRF" + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 205, + 155 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b31dda698362d8b9ae293fdc2235c9086df7d7269612578a6c8b0de3388462fa.jpg", + "image_caption": [ + "Instruct-NeRF2NeRF 1" + ], + "image_footnote": [], + "bbox": [ + 210, + 89, + 334, + 154 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a5a0c06a3876536af5a979b53b435f76c3920d4cc176f4cd355c3d25c723cc22.jpg", + "image_caption": [ + "Instruct-NeRF2NeRF 2" + ], + "image_footnote": [], + "bbox": [ + 341, + 90, + 465, + 154 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/df92d6df55aa709bdda7be0e168885a51760fbd267ca4524d1425ae1c856f04d.jpg", + "image_caption": [ + "Ours inference 1", + "Figure 7. Comparisons with baselines of text-driven NeRF editing. We compare our method with Instruct-NeRF2NeRF [9] in the editing by using the text prompt \"Make it Spring\"." + ], + "image_footnote": [], + "bbox": [ + 81, + 166, + 205, + 229 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/86bfc0b53e1dff3aeab3555a77b800c0c79363fc58056e5633805bd7fdb56bf1.jpg", + "image_caption": [ + "Ours inference 2" + ], + "image_footnote": [], + "bbox": [ + 210, + 166, + 336, + 229 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d9b07c9875a004e5c9656be5ba738e6b5355b277fa061b07a93f895e9c67dcec.jpg", + "image_caption": [ + "Ours inference 3" + ], + "image_footnote": [], + "bbox": [ + 341, + 166, + 465, + 229 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/36074e7b16c6e6d15085d87da927d74deef179f06e72f4172939fd00dbe646a5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCLIP Text-Image Direction Similarity↑CLIP Direction Consistency ↑FID ↓
InstructPix2Pix [2]+NeRF0.16690.8475270.542
Instruct-NeRF2NeRF0.20210.9828148.021
Ours w/o Ladv0.19200.9657162.275
Ours w/o Lcontr0.20070.9749156.524
Ours0.20890.9864137.740
", + "bbox": [ + 81, + 306, + 483, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "artifacts on novel views. Previous research demonstrates the effectiveness of conditional adversarial training in preventing the production of blurry rendered images resulting from conflicts that arise from noise in the camera extrinsic when performing image supervision from different viewpoints [11]. The function of the condition is to guide discriminator with fine-grained information from the same viewpoint, thus preventing GAN mode collapse.", + "bbox": [ + 75, + 500, + 468, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "It inspires us to incorporate conditional adversarial loss on rendered images from the translated NeRF, which is conducive to distinguish artifacts in rendered images. As illustrated in Fig.4, the discriminator $\\mathbf{D}$ takes into real pairs and fake pairs. Each real pair $\\mathbf{R}$ consists of $\\mathbf{S}^j$ and $\\mathbf{S}^j -\\mathbf{S}^k$ where $\\mathbf{S}^j\\in \\{\\mathbf{S}_i^j\\}_{i = 0}^{N - 1}$ and $\\mathbf{S}^k\\in \\{\\mathbf{S}_i^k\\}_{i = 0}^{N - 1}$ are from two sets of edited images from the image translation. Similarly, each fake pair $\\mathbf{F}$ consists of $\\mathbf{C}^j$ and $\\mathbf{C}^j -\\mathbf{S}^k$ in which $\\mathbf{C}^j\\in \\{\\mathbf{C}_i^j\\}_{i = 0}^{N - 1}$ is generated by translated NeRF. Note that the images in the same pair come from the same viewpoint. The pairs are concatenated in RGB channels and fed into the discriminator. We optimize the discriminator $\\mathbf{D}$ and translated NeRF with the objective functions below:", + "bbox": [ + 75, + 619, + 468, + 816 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {A D - D}} = \\mathbb {E} _ {\\mathbf {R}} [ - l o g (\\mathbf {D} (\\mathbf {R})) ] + \\mathbb {E} _ {\\mathbf {F}} [ - l o g (1 - \\mathbf {D} (\\mathbf {F})) ], \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 825, + 468, + 849 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {A D - G}} = \\mathbb {E} _ {\\mathbf {F}} [ - l o g (\\mathbf {D} (\\mathbf {F})) ].\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 845, + 272, + 861 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Optimization. During the training process, we jointly optimize the loss functions mentioned above: $\\mathcal{L}_{\\mathrm{KL}}$ and $\\mathcal{L}_{\\mathrm{contr}}$", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8b3abbdfb1cc401f78083ee3e9d504f25ca9dfd926e18de2204502d389ab78ca.jpg", + "image_caption": [ + "Original NeRF" + ], + "image_footnote": [], + "bbox": [ + 504, + 90, + 627, + 159 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/84cc93b7444cb4f0729abcbc0ec4d39f361680c551f7427720cf6c1197de227b.jpg", + "image_caption": [ + "Instruct-NeRF2NeRF" + ], + "image_footnote": [], + "bbox": [ + 633, + 90, + 756, + 160 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dda52412dbaf0acaba32f1cfbfb5f51e7974b02aa213a2afd0feb29121574ff7.jpg", + "image_caption": [ + "PaletteNeRF" + ], + "image_footnote": [], + "bbox": [ + 764, + 90, + 887, + 160 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/159f0eda688041349c873d7f76f4316f35000945f3405e18f79ef75d96b19797.jpg", + "image_caption": [ + "Ours inference 1", + "Figure 8. Comparisons with baselines of NeRF colorization. We compare with PaletteNeRF[16] in colorization." + ], + "image_footnote": [], + "bbox": [ + 503, + 171, + 627, + 242 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ccc9308bb2cf789858b3dc4ffd8801748154ebf4b74d9d5d6879070c1c67ca90.jpg", + "image_caption": [ + "Ours inference 2" + ], + "image_footnote": [], + "bbox": [ + 633, + 171, + 756, + 242 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e2e1951d4fae3b38aff5be22e2642dcb7694a7fc7c3c92b557bd8a70d598b963.jpg", + "image_caption": [ + "Ours inference 3" + ], + "image_footnote": [], + "bbox": [ + 764, + 171, + 887, + 242 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d87e60f944ea68b8818327176abb4eafa830d090d0552b1905650a13a957d7f9.jpg", + "table_caption": [ + "Table 1. Quantitative results on text-driven editing. We compare our method with the naive method of directly combining InstructPix2Pix [2] with NeRF and the state-of-the-art method Instruct-NeRF2NeRF [9]." + ], + "table_footnote": [], + "table_body": "
MethodCF ↑FID ↓
DDColor [14]+NeRF40.435148.957
Instruct-NeRF2NeRF45.599201.456
PaletteNeRF [16]39.654-
Ours w/o Ladv35.031137.740
Ours w/o Lcontr34.829105.750
Ours65.09935.041
", + "bbox": [ + 504, + 303, + 906, + 422 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Quantitative results on colorization. We colorize images with the translator, DDcolor [14], and edit NeRF by directly optimizing NeRF, Instruct-NeRF2NeRF [9] and our NeRF translation method. The quantitative comparison is conducted between these methods as well as PaletteNeRF [16].", + "bbox": [ + 498, + 431, + 890, + 502 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "for the edit code, $\\mathcal{L}_{\\mathrm{recon}}$ and $\\mathcal{L}_{\\mathrm{AD - G}}$ for the translated NeRF, and $\\mathcal{L}_{\\mathrm{AD - D}}$ for the discriminator. The total loss formula is expressed as follows:", + "bbox": [ + 498, + 526, + 890, + 571 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {K L}} + \\mathcal {L} _ {\\text {r e c o n}} + \\mathcal {L} _ {\\mathrm {A D - G}} + \\mathcal {L} _ {\\mathrm {A D - D}} + \\mathcal {L} _ {\\text {c o n t r}}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 580, + 890, + 597 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where we assign each regularization term the weight of 1.0, 1.0, 0.1, 0.1, 0.1 in all of our experiments. The weights can be adjusted to prioritize different aspects of the training objective, such as reconstruction accuracy, adversarial training, and perceptual quality.", + "bbox": [ + 498, + 604, + 890, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3. Inference", + "text_level": 1, + "bbox": [ + 500, + 689, + 607, + 703 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After the optimization of our GenN2N, the translated NeRF is optimized to be able to render the target scene conditioned on the edit code. As shown in Fig. 1, users can simply sample an edit code from the Gaussian distribution and use the translated NeRF to render the 3D scene with high-quality and multi-view 3D consistency.", + "bbox": [ + 496, + 710, + 890, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 815, + 630, + 830 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our proposed GenN2N is a unified NeRF-to-NeRF translation framework which can support various NeRF editing tasks. In this paper, we demonstrate the effectiveness of GenN2N by a suite of challenging tasks:", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5110", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ea7a30b00b68fa9411a593ace664d3944b1ec609ad2b74c06b46571d549da071.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 89, + 176, + 145 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ea07cdbf5005153b3fefbaf6cbbc4ab178f07a04305b7d8b8fc23560a8262968.jpg", + "image_caption": [ + "Original NeRF" + ], + "image_footnote": [], + "bbox": [ + 176, + 90, + 271, + 146 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fc446adaa1a09002755507642964878b5a058cab2d5b9896f16222960353f499.jpg", + "image_caption": [ + "Instruct-NeRF2NeRF" + ], + "image_footnote": [], + "bbox": [ + 276, + 90, + 370, + 146 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8c734d8649c9850ae52c36c4655967ffdb6759f5ec469f76b577323ab321ed99.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 90, + 465, + 146 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f33b2e818b8f36d411da9f6cbd9a7dfd67cf7f89580d02ff866ba9ce7f9adbf5.jpg", + "image_caption": [ + "NeRF-SR", + "Figure 9. Comparisons with baselines of NeRF super-resolution. We compare with NeRF-SR [34] in super-resolution." + ], + "image_footnote": [], + "bbox": [ + 81, + 155, + 174, + 210 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/79f4eb3b5408230337924aa9c63495fd66cdd5254acd6196aaee3ad6c651e027.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 176, + 155, + 271, + 210 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5cbb8ff671d1809fabcc856eb267883b8cc5bceb9655e76da1c52bbc7511d1a7.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 276, + 155, + 370, + 210 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/14fa18a5859480ae72322936d6e46c5f13087490b5647c2789da03ad0b9ea568.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 155, + 465, + 210 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ffb5aefc78086ae3de55974ed7d2c874d0bb8dcbb7599b8160bc4d4171cf8a3e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPSNR ↑SSIM ↑LPIPS ↓
ResShift [40]+NeRF19.9780.5350.1156
Instruct-NeRF2NeRF20.2990.6420.2732
NeRF-SR [34]27.9570.8970.0937
Ours w/o Ladv12.5550.6630.2001
Ours w/o Lcontr15.3720.6620.1834
Ours28.5010.9130.0748
", + "bbox": [ + 81, + 273, + 483, + 366 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Quantitative results on super-resolution. We improve image resolution with ResShift [40] and edit NeRF by directly optimizing NeRF, Instruct-NeRF2NeRF [9] and our NeRF translation method. The quantitative comparison is conducted between these methods as well as NeRF-SR [34].", + "bbox": [ + 75, + 377, + 468, + 446 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) Text-driven Editing edits the given NeRF scene to a set of NeRF scenes according to the text instruction.", + "(2) Colorization transforms a gray-scale NeRF scene to a set of plausible color NeRF scenes.", + "(3) Super-resolution enhances the resolution of NeRF and enables multiple plausible outcomes.", + "(4) Inpainting fills in user-specified masked regions in the NeRF scene with realistic content." + ], + "bbox": [ + 81, + 473, + 468, + 593 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We achieve those tasks by simply changing the plug-and-play 2D image translator in our framework, without any additional task-specific design. Previous studies have extensively explored some of these issues like text-driven editing, colorization, super-resolution, and inpainting. However, there is rarely a unified framework that can achieve all these tasks with strong performance, high quality, and plausible multi-view consistent 3D structure. Furthermore, GenN2N can also perform zooming out and text-driven inpainting in NeRF-to-NeRF translation, which were not explored in prior research. We refer readers to the supplementary materials for detailed experiment settings, dataset settings and implementation details.", + "bbox": [ + 75, + 595, + 468, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Comparisons", + "text_level": 1, + "bbox": [ + 76, + 801, + 215, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Text-driven Editing. We achieve text-driven editing of the given NeRF by using InstructPix2Pix [2] as the 2D image-to-image translator in our framework. We compared our approach to a naive solution, which involves optimizing a NeRF with edited images via InstructPix2Pix. However,", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "this naive approach leads to a 3D inconsistency problem among different edits. While Instruct-NeRF2NeRF [9] proposed an iterative updating mechanism to address this issue, it falls short in capturing the diversity of different edits, making it challenging to ensure the quality of the outcomes. To evaluate our method, we conducted experiments on the Face [9] and Fangzhou [35] self-portrait datasets, and Farm [9] dataset, comparing GenN2N with the state-of-the-art NeRF editing method Instruct-NeRF2NeRF [9].", + "bbox": [ + 496, + 90, + 890, + 227 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative results are presented in Table 1, where we employed CLIP Text-Image Direction Similarity [9], CLIP Direction Consistency [9], and Fréchet Inception Distance (FID) [10] as evaluation metrics. The results highlight the superior performance of GenN2N over other methods, demonstrating its effectiveness in producing high-quality 3D text-driven editing results. Furthermore, we provide a qualitative comparison between GenN2N and Instruct-NeRF2NeRF [9] in Figure 7. Notably, Instruct-NeRF2NeRF was trained twice, but the results are nearly identical. In contrast, GenN2N can render edited scenes with various effects, consistent with the input text instruction, by inferring with different edit codes sampled from a Gaussian distribution. To further illustrate the generative capability of GenN2N, we showcase additional results of our inference renderings in Figure 5, demonstrating its diverse generative ability in terms of appearance and geometry.", + "bbox": [ + 496, + 229, + 892, + 502 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Colorization. For NeRF colorization, GenN2N uses DD-Color [14] as the 2D image-to-image translator. CoRF [8] and Palette-NeRF [16] do a similar task and we compare with Palette-NeRF in Table 2. The Fréchet Inception Distance (FID) [10] and colorfulness score (CF) [17] are used to measure the distribution similarity and vividness of generated images. We show visual comparison results in Fig. 6 and Fig. 8. We can find that with different edit codes, the scene can be rendered in different color styles. It is noticeable that with the same edit code, the color rendered from different views is consistent. This strongly demonstrates the effectiveness of our method in translating NeRF while keeping the 3D consistency of the scene.", + "bbox": [ + 496, + 503, + 892, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Super-resolution. When only low-resolution images are available, our methods can boost NeRF in reconstructing scenes at higher resolution, while keeping view consistency and avoiding blurry outputs. We achieve this by employing ResShift [40] as the 2D image-to-image translator in GenN2N. Following state-of-the-art method NeRF-SR [34], we conduct experiments on $LLFF$ dataset [23], using PSNR, SSIM, and LPIPS as evaluation metrics. As shown in Table 3, GenN2N obtains NeRF-to-NeRF translation with higher performance than NeRF-SR [34]. Moreover, we also provide qualitative comparison results in Fig. 9, where GenN2N produces clearer and more realistic rendering results than previous methods.", + "bbox": [ + 496, + 704, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5111", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2d395f82bc64109b08b3631b093c0feaadb2158aa8a700af2aa2a6ccfdbf0e1c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 90, + 174, + 146 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0e9e79b1be9ac459c1c1ae539d5a4d29ca56248b322a6e3a99dc13ae8cae17be.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 174, + 90, + 271, + 146 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/03915de0884441efb6c533790dc10dd5779d3f3d7b1f48e254f6fc42dad1d4d1.jpg", + "image_caption": [ + "Instruct-NeRF2NeRF" + ], + "image_footnote": [], + "bbox": [ + 274, + 90, + 370, + 146 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/78479ea566ffe1847a74ba82b42f0b4ba8494975fec2cf0053a6a588f567fb1b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 370, + 90, + 465, + 146 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9dff90df2b08ff280aea834c0947a49fa62b04f3625e1716db0dbda811e74005.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 159, + 174, + 212 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/484cba32e0cd42a5a3d692f9500a6699281318e62d26b4afa219ff24723d5e29.jpg", + "image_caption": [ + "Original NeRF", + "SPin-NeRF", + "Figure 10. Comparisons with baselines of NeRF inpainting. We compare with SPIn-NeRF [24] in inpainting." + ], + "image_footnote": [], + "bbox": [ + 174, + 157, + 271, + 214 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/04848002808598c140cd62eb2f3e2486767f21657c5c122071169ba1305fbd8b.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 274, + 157, + 370, + 214 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7ed027a0cbd9386e6dfe4cb4901dd88152219e7d5dc64eea4761285f4f21f6c8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 370, + 157, + 465, + 214 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3b5fb1001a93f589467568ab549c13d6f0cc6e4d5ed767d7b1cfc1e9f8a90d24.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPSNR ↑SSIM ↑LPIPS ↓
LaMa [31]+NeRF18.9830.37060.1730
Instruct-NeRF2NeRF16.7340.30880.2750
SPin-NeRF [24]24.3690.72170.1754
Ours26.8680.81370.1284
", + "bbox": [ + 81, + 276, + 483, + 349 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Quantitative results on NeRF inpainting.", + "bbox": [ + 117, + 359, + 428, + 375 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Inpainting. The goal of NeRF Inpainting is to fill the 3D content of regions specified by users. SPin-NeRF [24] achieves this through a multi-step process: it employs SAM [15] for object segmentation, utilizes LaMa [31] to paint the background content in multi-view images, and subsequently trains the NeRF model with color, depth, and perceptual cues. In our experiments, we use SAM and LaMa as the 2D image-to-image translator in our GenN2N, which is the same setting as SPin-NeRF [24]. Quantitative comparisons on statue dataset [24] are shown in Table 4, where GenN2N achieves superior PSNR and SSIM scores than SPin-NeRF [24], highlighting the effectiveness of our GenN2N framework. In addition, qualitative results are showcased in Fig. 10 revealing that while SPin-NeRF [24] fails to generate reasonable content behind the masked object, our GenN2N produces realistic content in the same region with fine multi-view consistency.", + "bbox": [ + 75, + 398, + 472, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 666, + 240, + 680 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct comprehensive ablation experiments to validate the designs of each component in GenN2N. Due to space limitations, we only highlight the essential aspects below. Please refer to supplementary for more details.", + "bbox": [ + 75, + 689, + 468, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The Contrastive Loss. We demonstrate the advantages of incorporating our proposed contrastive loss in Table 1, 2, 3. The motivation is to disentangle the camera view and edit information present in the latent space. We achieve this by reducing the distance between edit codes of different-view rendered images from a translated NeRF that shares the same edit code, while increasing the distance between same-view images that are edited by the 2D translator with diverse edit styles. As demonstrated in Tables, the absence of contrastive loss leads to the generation of blurry areas", + "bbox": [ + 75, + 750, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "in the rendered images, resulting in a decrease in the metric scores. This blurriness can be attributed to the inclusion of pose information within the edit code z. By incorporating the contrastive loss, our method successfully achieves a uniform appearance with different observing views under the same style latent z.", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Discriminator for Novel Views. We demonstrate the effectiveness of employing a conditional discriminator to address artifacts caused by inconsistent cross-view edited images and to enhance the quality of novel view rendering images, as depicted in Table 1, 2, 3. The removal of this conditional discriminator results in blurry novel view images with artifacts in the background region. We attribute these undesirable effects to the inability of current image-to-image translation methods, such as InstructPix2Pix [2], to produce image editing consistently across multi-view images. To mitigate these issues, we introduce a conditional discriminator between rendered images from the translated NeRF and edited images from the 2D image-to-image translator. This inclusion successfully eliminates artifacts and enhances the image quality of rendered images from the translated NeRF.", + "bbox": [ + 496, + 186, + 892, + 412 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Applications", + "text_level": 1, + "bbox": [ + 500, + 436, + 633, + 453 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We demonstrate the versatility and robustness of GenN2N by exploring two translation applications: Zoom Out and Text-Driven Inpainting. While existing 2D translators [1, 29] can complete these tasks, 3D editing has not been explored. We achieve these tasks by incorporating Blended Latent Diffusion [1] as the 2D image-to-image translator, enabling us to generate diverse and high-quality content with multi-view consistency. Please refer to our supplementary materials for the results of these applications due to space constraints.", + "bbox": [ + 496, + 464, + 890, + 616 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 643, + 625, + 659 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce GenN2N, a unified NeRF-to-NeRF translation framework that can handle various NeRF editing tasks. Unlike previous task-specific approaches, our framework uses an image-to-image translator for 2D editing and integrates the results into 3D NeRF space. To address the challenge of ensuring 3D consistency, we propose modeling the distribution of 3D edited NeRFs from 2D edited images using our novel techniques. After optimization, users can sample from the conditional generative model to obtain diverse 3D editing results with high rendering quality and multi-view consistency. Our experiments demonstrate that GenN2N outperforms existing task-specific methods on various editing tasks, including text-driven editing, colorization, super-resolution, and inpainting, in terms of efficiency, quality, and diversity.", + "bbox": [ + 496, + 674, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5112", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM Transactions on Graphics (TOG), 42 (4):1-11, 2023. 3, 8", + "[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. arXiv preprint arXiv:2211.09800, 2022. 3, 6, 7, 8", + "[3] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 3", + "[4] Shengqu Cai, Anton Obukhov, Dengxin Dai, and Luc Van Gool. Pix2nerf: Unsupervised conditional p-gan for single image to neural radiance fields translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3981-3990, 2022. 2", + "[5] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 3", + "[6] Jun-Kun Chen, Jipeng Lyu, and Yu-Xiong Wang. Neuraleditor: Editing neural radiance fields via manipulating point clouds. arXiv preprint arXiv:2305.03049, 2023. 2", + "[7] Pei-Ze Chiang, Meng-Shiun Tsai, Hung-Yu Tseng, Wei-Sheng Lai, and Wei-Chen Chiu. Stylizing 3d scene via implicit representation and hypernetwork. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1475-1484, 2022. 2", + "[8] Ankit Dhiman, R Srinath, Srinjay Sarkar, Lokesh R. Boregowda, and R Venkatesh Babu. Corf: Colorizing radiance fields using knowledge distillation. ArXiv, abs/2309.07668, 2023. 7", + "[9] Ayaan Haque, Matthew Tancik, Alexei A Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. arXiv preprint arXiv:2303.12789, 2023.2,3,6,7", + "[10] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7", + "[11] Jingwei Huang, Justus Thies, Angela Dai, Abhijit Kundu, Chiyu Jiang, Leonidas J Guibas, Matthias Nießner, Thomas Funkhouser, et al. Adversarial texture optimization from rgb-d scans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1559-1568, 2020. 6", + "[12] Yi-Hua Huang, Yue He, Yu-Jie Yuan, Yu-Kun Lai, and Lin Gao. Stylizednerf: consistent 3d scene stylization as stylized nerf via 2d-3d mutual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18342–18352, 2022. 2", + "[13] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object genera" + ], + "bbox": [ + 78, + 114, + 470, + 902 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tion with dream fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 867-876, 2022. 2", + "[14] Xiaoyang Kang, Tao Yang, Wenqi Ouyang, Peiran Ren, Lingzhi Li, and Xuansong Xie. Ddcolor: Towards photorealistic and semantic-aware image colorization via dual decoders. arXiv preprint arXiv:2212.11613, 2022. 6, 7", + "[15] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 8", + "[16] Zhengfei Kuang, Fujun Luan, Sai Bi, Zhixin Shu, Gordon Wetzstein, and Kalyan Sunkavalli. Palettenerf: Palette-based appearance editing of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20691-20700, 2023. 2, 6, 7", + "[17] Gustav Larsson, Michael Maire, and Gregory Shakhnarovich. Learning representations for automatic colorization. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 577-593. Springer, 2016. 7", + "[18] Verica Lazova, Vladimir Guzov, Kyle Olszewski, Sergey Tulyakov, and Gerard Pons-Moll. Control-nerf: Editable feature volumes for scene rendering and manipulation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4340–4350, 2023. 2", + "[19] Hao-Kang Liu, I Shen, Bing-Yu Chen, et al. Nerf-in: Free-form nerf inpainting with rgb-d priors. arXiv preprint arXiv:2206.04901, 2022. 2", + "[20] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. arXiv preprint arXiv:2303.11328, 2023. 2", + "[21] Steven Liu, Xiuming Zhang, Zhoutong Zhang, Richard Zhang, Jun-Yan Zhu, and Bryan Russell. Editing conditional radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5773-5783, 2021. 2", + "[22] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Realfusion: $360^{\\circ}$ reconstruction of any object from a single image. arXiv e-prints, pages arXiv-2302, 2023. 2", + "[23] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 7", + "[24] Ashkan Mirzaei, Tristan Aumentado-Armstrong, Konstantinos G Derpanis, Jonathan Kelly, Marcus A Brubaker, Igor Gilitschenski, and Alex Levinshtein. Spin-nerf: Multiview segmentation and perceptual inpainting with neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20669–20679, 2023. 8", + "[25] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image" + ], + "bbox": [ + 503, + 92, + 890, + 902 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5113", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "translation. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2", + "[26] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2, 3", + "[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3", + "[28] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 3, 4", + "[29] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 2, 8", + "[30] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. Advances in Neural Information Processing Systems, 33:20154-20166, 2020. 3", + "[31] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. arXiv preprint arXiv:2109.07161, 2021.8", + "[32] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, et al. Nerfstudio: A modular framework for neural radiance field development. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-12, 2023. 3", + "[33] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3835–3844, 2022. 3", + "[34] Chen Wang, Xian Wu, Yuan-Chen Guo, Song-Hai Zhang, Yu-Wing Tai, and Shi-Min Hu. Nerf-sr: High quality neural radiance fields using supersampling. In Proceedings of the 30th ACM International Conference on Multimedia, pages 6445-6454, 2022. 2, 7", + "[35] Can Wang, Ruixiang Jiang, Mengei Chai, Mingming He, Dongdong Chen, and Jing Liao. Nerf-art: Text-driven neural radiance fields stylization. IEEE Transactions on Visualization and Computer Graphics, 2023. 3, 7", + "[36] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3d digital avatars using diffusion. arXiv preprint arXiv:2212.06135, 2022. 2", + "[37] Qiling Wu, Jianchao Tan, and Kun Xu. Palettenerf:" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Palette-based color editing for nerfs. arXiv preprint arXiv:2212.12871, 2022. 2", + "[38] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021. 2", + "[39] Yu-Jie Yuan, Yang-Tian Sun, Yu-Kun Lai, Yuewen Ma, Rongfei Jia, and Lin Gao. Nerf-editing: geometry editing of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18353-18364, 2022. 2", + "[40] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Resshift: Efficient diffusion model for image super-resolution by residual shifting. arXiv preprint arXiv:2307.12348, 2023. 7", + "[41] Jiakai Zhang, Xinhang Liu, Xinyi Ye, Fuqiang Zhao, Yanshun Zhang, Minye Wu, Yingliang Zhang, Lan Xu, and Jingyi Yu. Editable free-viewpoint video using a layered neural representation. ACM Transactions on Graphics (TOG), 40(4):1-18, 2021. 2", + "[42] Kai Zhang, Nick Kolkin, Sai Bi, Fujun Luan, Zexiang Xu, Eli Shechtman, and Noah Snavely. Arf: Artistic radiance fields. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXI, pages 717-733. Springer, 2022. 2", + "[43] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543, 2023. 2", + "[44] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5", + "[45] Chengwei Zheng, Wenbin Lin, and Feng Xu. Editablenerf: Editing topologically varying neural radiance fields by key points. arXiv preprint arXiv:2212.04247, 2022. 2", + "[46] Jingyu Zhuang, Chen Wang, Lingjie Liu, Liang Lin, and Guanbin Li. Dreameditor: Text-driven 3d scene editing with neural fields. arXiv preprint arXiv:2306.13455, 2023. 3" + ], + "bbox": [ + 501, + 92, + 893, + 643 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5114", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/GenN2N_ Generative NeRF2NeRF Translation/720fe634-ca22-4125-bab8-c49c422e71c2_model.json b/2024/GenN2N_ Generative NeRF2NeRF Translation/720fe634-ca22-4125-bab8-c49c422e71c2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8f78fd4f288c80762ad41550ffecd8c5313916c2 --- /dev/null +++ b/2024/GenN2N_ Generative NeRF2NeRF Translation/720fe634-ca22-4125-bab8-c49c422e71c2_model.json @@ -0,0 +1,3289 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.131, + 0.723, + 0.151 + ], + "angle": 0, + "content": "GenN2N: Generative NeRF2NeRF Translation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.761, + 0.201 + ], + "angle": 0, + "content": "Xiangyue Liu\\(^{1}\\) Han Xue\\(^{2}\\) Kunming Luo\\(^{1}\\) Ping Tan\\(^{1\\dagger}\\) Li Yi\\(^{2,3,4\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.184, + 0.215, + 0.787, + 0.234 + ], + "angle": 0, + "content": "\\(^{1}\\) Hong Kong University of Science and Technology \\(^{2}\\) Tsinghua University" + }, + { + "type": "text", + "bbox": [ + 0.189, + 0.234, + 0.776, + 0.253 + ], + "angle": 0, + "content": "\\(^{3}\\) Shanghai Artificial Intelligence Laboratory \\(^{4}\\) Shanghai Qi Zhi Institute" + }, + { + "type": "list", + "bbox": [ + 0.184, + 0.215, + 0.787, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.139, + 0.287, + 0.48, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.182, + 0.441, + 0.209, + 0.449 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image_caption", + "bbox": [ + 0.234, + 0.441, + 0.297, + 0.449 + ], + "angle": 0, + "content": "Output 1" + }, + { + "type": "image_caption", + "bbox": [ + 0.318, + 0.441, + 0.379, + 0.449 + ], + "angle": 0, + "content": "Output 2" + }, + { + "type": "image_caption", + "bbox": [ + 0.398, + 0.441, + 0.479, + 0.449 + ], + "angle": 0, + "content": "Output 3" + }, + { + "type": "image", + "bbox": [ + 0.139, + 0.45, + 0.48, + 0.558 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.183, + 0.559, + 0.209, + 0.567 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image_caption", + "bbox": [ + 0.237, + 0.558, + 0.309, + 0.567 + ], + "angle": 0, + "content": "Input zoom in" + }, + { + "type": "image_caption", + "bbox": [ + 0.343, + 0.559, + 0.374, + 0.567 + ], + "angle": 0, + "content": "Output" + }, + { + "type": "image_caption", + "bbox": [ + 0.398, + 0.558, + 0.475, + 0.567 + ], + "angle": 0, + "content": "Output zoom in" + }, + { + "type": "image_caption", + "bbox": [ + 0.486, + 0.342, + 0.501, + 0.388 + ], + "angle": 0, + "content": "aannnnnne aannnnnne" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.288, + 0.582, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.585, + 0.289, + 0.663, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.666, + 0.289, + 0.744, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.748, + 0.289, + 0.826, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.488, + 0.365, + 0.582, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.441, + 0.556, + 0.449 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.585, + 0.365, + 0.663, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.604, + 0.441, + 0.644, + 0.449 + ], + "angle": 0, + "content": "Output 1" + }, + { + "type": "image", + "bbox": [ + 0.666, + 0.365, + 0.744, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.685, + 0.441, + 0.726, + 0.449 + ], + "angle": 0, + "content": "Output 2" + }, + { + "type": "image", + "bbox": [ + 0.748, + 0.365, + 0.826, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.766, + 0.441, + 0.807, + 0.449 + ], + "angle": 0, + "content": "Output 3" + }, + { + "type": "image", + "bbox": [ + 0.488, + 0.45, + 0.582, + 0.557 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.558, + 0.56, + 0.566 + ], + "angle": 0, + "content": "View 1" + }, + { + "type": "image", + "bbox": [ + 0.585, + 0.45, + 0.663, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.604, + 0.558, + 0.643, + 0.566 + ], + "angle": 0, + "content": "View 2" + }, + { + "type": "image", + "bbox": [ + 0.666, + 0.45, + 0.744, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.666, + 0.504, + 0.744, + 0.557 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.688, + 0.558, + 0.722, + 0.566 + ], + "angle": 0, + "content": "View 3" + }, + { + "type": "image", + "bbox": [ + 0.748, + 0.45, + 0.826, + 0.557 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.769, + 0.558, + 0.805, + 0.566 + ], + "angle": 0, + "content": "View 4" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.571, + 0.894, + 0.613 + ], + "angle": 0, + "content": "Figure 1. We introduce GenN2N, a unified framework for NeRF-to-NeRF translation, enabling a range of 3D NeRF editing tasks, including text-driven editing, colorization, super-resolution, inpainting, etc. We show at least two rendering views of edited NeRF scenes at inference time. Given a 3D NeRF scene, GenN2N can produce high-quality editing results with suitable multi-view consistency." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.625, + 0.314, + 0.641 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.473, + 0.871 + ], + "angle": 0, + "content": "We present GenN2N, a unified NeRF-to-NeRF translation framework for various NeRF translation tasks such as text-driven NeRF editing, colorization, super-resolution, inpainting, etc. Unlike previous methods designed for individual translation tasks with task-specific schemes, GenN2N achieves all these NeRF editing tasks by employing a plug-and-play image-to-image translator to perform editing in the 2D domain and lifting 2D edits into the 3D NeRF space. Since the 3D consistency of 2D edits may not be assured, we propose to model the distribution of the underlying 3D edits through a generative model that can cover all possible edited NeRFs. To model the distribution of 3D edited NeRFs from 2D edited images, we carefully design a VAE-GAN that encodes images while decoding NeRFs. The la" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.627, + 0.895, + 0.779 + ], + "angle": 0, + "content": "tent space is trained to align with a Gaussian distribution and the NeRFs are supervised through an adversarial loss on its renderings. To ensure the latent code does not depend on 2D viewpoints but truly reflects the 3D edits, we also regularize the latent code through a contrastive learning scheme. Extensive experiments on various editing tasks show GenN2N, as a universal framework, performs as well or better than task-specific specialists while possessing flexible generative power. More results on our project page: https://xiangyueliu.github.io/GenN2N/." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.813, + 0.633, + 0.83 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Over the past few years, Neural radiance fields (NeRFs) [23] have brought a promising paradigm in the realm of 3D reconstruction, 3D generation, and novel view synthesis due to their unparalleled compactness," + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.228, + 0.9 + ], + "angle": 0, + "content": "†Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5105" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.198 + ], + "angle": 0, + "content": "high quality, and versatility. Extensive research efforts have been devoted to creating NeRF scenes from 2D images [4, 20, 22, 36, 38] or just text [13, 26] input. However, once the NeRF scenes have been created, these methods often lack further control over the generated geometry and appearance. NeRF editing has therefore become a notable research focus recently." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.2, + 0.471, + 0.412 + ], + "angle": 0, + "content": "Existing NeRF editing schemes are usually task-specific. For example, researchers have developed NeRF-SR [34], NeRF-In [19], PaletteNeRF [16] for NeRF super-resolution, inpainting, and color-editing respectively. These designs require a significant amount of domain knowledge for each specific task. On the other hand, in the field of 2D image editing, a growing trend is to develop universal image-to-image translation methods to support versatile image editing [25, 29, 43]. By leveraging foundational 2D generative models, e.g., stable diffusion [28], these methods achieve impressive editing results without task-specific customization or tuning. We then ask the question: can we conduct universal NeRF editing leveraging foundational 2D generative models as well?" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.415, + 0.47, + 0.763 + ], + "angle": 0, + "content": "The first challenge is the representation gap between NeRFs and 2D images. It is not intuitive how to leverage image editing tools to edit NeRFs. A recent text-driven NeRF editing method [9] has shed some light on this. The method adopts a \"render-edit-aggregate\" pipeline. Specifically, it gradually updates a NeRF scene by iteratively rendering multi-view images, conducting text-driven visual editing on these images, and finally aggregating the edits in the NeRF scene. It seems that replacing the image editing tool with a universal image-to-image translator could lead to a universal NeRF editing method. However, the second challenge would then come. Image-to-image translators usually generate diverse and inconsistent edits for different views, e.g. turning a man into an elf might or might not put a hat on his head, making edits aggregation intricate. Regarding this challenge, Instruct-NeRF2NeRF [9] presents a complex optimization technique to pursue unblurred NeRF with inconsistent multi-view edits. Due to its complexity, the optimization cannot ensure the robustness of the outcomes. Additionally, the unique optimization outcome fails to reflect the stochastic nature of NeRF editing. Users typically anticipate a variety of edited NeRFs just like the diverse edited images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.901 + ], + "angle": 0, + "content": "To tackle the challenges above, we propose GenN2N, a unified NeRF-to-NeRF translation framework for various NeRF editing tasks such as text-driven editing, colorization, super-resolution, inpainting (see Fig. 1). In contrast to Instruct-NeRF2NeRF which adopts a \"render-edit-aggregate\" pipeline, we first render a NeRF scene into multi-view images, then exploit an image-to-image translator to edit different views, and finally learn a generative model to depict the distribution of NeRF edits. Instead of" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "aggregating all the image edits to form a single NeRF edit, our key idea is to embrace the stochastic nature of content editing by modeling the distribution of the edits in the 3D NeRF space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.153, + 0.893, + 0.499 + ], + "angle": 0, + "content": "Specifically given a NeRF model or its multi-view images, along with the editing goal, we first generate edited multi-view images using a plug-and-play image-to-image translator. Each view corresponds to a unique 3D edit with some geometry or appearance variations. Conditioned on the input NeRF, GenN2N trains a conditional 3D generative model to reflect such content variations. At the core of GenN2N, we design a 3D VAE-GAN that incorporates a differentiable volume renderer to connect 3D content creation with 2D GAN losses, ensuring that the inconsistent multi-view renderings can still help each other regarding 3D generation. Moreover, we introduce a contrastive learning loss to ensure that the 3D content variation can be successfully understood just from edited 2D images without being influenced by the camera viewpoints. During inference, users can simply sample from the conditional generative model to obtain various 3D editing results aligned with the editing goal. We have conducted experiments on human, items, indoor and outdoor scenes for various editing tasks such as text-driven editing, colorization, superresolution and inpainting, demonstrating the effectiveness of GenN2N in supporting diverse NeRF editing tasks while keeping the multi-view consistency of the edited NeRF." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.5, + 0.888, + 0.515 + ], + "angle": 0, + "content": "We summarize the contribution of this paper as follows," + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.515, + 0.891, + 0.558 + ], + "angle": 0, + "content": "- A generative NeRF-to-NeRF translation formulation for the universal NeRF editing task together with a generic solution;" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.56, + 0.892, + 0.605 + ], + "angle": 0, + "content": "- a 3D VAE-GAN framework that can learn the distribution of all possible 3D NeRF edits corresponding to the a set of input edited 2D images;" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.606, + 0.891, + 0.636 + ], + "angle": 0, + "content": "- a contrastive learning framework that can disentangle the 3D edits and 2D camera views from edited images;" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.636, + 0.891, + 0.679 + ], + "angle": 0, + "content": "- extensive experiments demonstrating the superior efficiency, quality, and diversity of the NeRF-to-NeRF translation results." + }, + { + "type": "list", + "bbox": [ + 0.501, + 0.515, + 0.892, + 0.679 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.695, + 0.642, + 0.71 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.87 + ], + "angle": 0, + "content": "NeRF Editing. Previous works such as EditNeRF [21] propose a conditional neural field that enables shape and appearance editing in the latent space. PaletteNeRF [16, 37] focuses on controlling color palette weights to manipulate appearance. Other approaches utilize bounding boxes [41], meshes [39], point clouds [6], key points [45], or feature volumes [18] to directly manipulate the spatial representation of NeRF. However, these methods either heavily rely on user interactions or have limitations in terms of spatial deformation and color transfer capabilities." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "NeRF Stylization. Images-referenced stylization [7, 12, 42] often prioritize capturing texture style rather than de" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5106" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.891, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.293, + 0.896, + 0.381 + ], + "angle": 0, + "content": "Figure 2. Overview of GenN2N. We first edit the source image set \\(\\{\\mathbf{I}_i\\}_{i=0}^{N-1}\\) using 2D image-to-image translation methods, e.g., text-driven editing, colorization, zoom out, etc. For each view \\(i \\in [0, N-1]\\), we generate \\(M\\) edited images, resulting in a group of translated image set \\(\\{\\{\\mathbf{S}_i^j\\}_{j=0}^{M-1}\\}_{i=0}^{N-1}\\). Then we use the Latent Distill Module to learn \\(M \\times N\\) edit code vectors from the translated image set, which serve as the input of the translated NeRF. To optimize our GenN2N, we design four loss functions: a KL loss to constrain the latent vectors to a Gaussian distribution; and \\(\\mathcal{L}_{\\mathrm{recon}}\\), \\(\\mathcal{L}_{\\mathrm{adv}}\\) and \\(\\mathcal{L}_{\\mathrm{contr}}\\) to optimize the appearance and geometry of the translated NeRF. At inference, we can sample a latent vector \\(\\mathbf{z}\\) from Gaussian distribution and render a corresponding multi-view consistent 3D scene with high quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.404, + 0.471, + 0.495 + ], + "angle": 0, + "content": "tailed content, resulting in imprecise editing appearance of NeRF only. Text-guided works [33, 35], on the other hand, apply contrastive losses based on CLIP [27] to achieve the desired edits. While text references usually describe the global characteristics of the edited results, instructions offer a more convenient and precise expression." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.496, + 0.471, + 0.722 + ], + "angle": 0, + "content": "Instruct-driven NeRF editing. Among numerous image-to-image translation works, InstructPix2Pix [2] stands out by efficiently editing images following instructions. It leverages large pre-trained models in the language and image domains [3, 28] to generate paired data (before and after editing) for training. While editing NeRF solely based on edited images is problematic due to multi-view inconsistency. To address this, an intuitive yet heavy approach [9] is to iteratively edit the image and optimize NeRF. In addition, NeRF-Art [35] and DreamEditor [46] adopt a CLIP-based contrastive loss [27] and score distillation sampling [26] separately to supervise the optimization of editing NeRF. Inspired by Generative Radiance Fields [5, 30], We capture various possible NeRF editing in the generative space to solve it." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.739, + 0.169, + 0.754 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Given a NeRF scene, we present a unified framework GenN2N to achieve various editing on the 3D scene leveraging geometry and texture priors from 2D image editing methods, such as text-driven editing, colorization, superresolution, inpainting, etc. While a universal image-to-image translator can theoretically accomplish these 2D editing tasks, we actually use a state-of-the-art translator for each task. Therefore, we formulate each 2D image editing method as a plug-and-play image-to-image translator" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.404, + 0.893, + 0.509 + ], + "angle": 0, + "content": "and all NeRF editing tasks as our universal NeRF-to-NeRF translation, in which the given NeRF is translated into NeRF scenes with high rendering quality and 3D geometry consistency according to the user-selected editing target. The overview of GenN2N is illustrated in Fig. 2, we first perform image-to-image translation in the 2D domain and then lift 2D edits to 3D and achieve NeRF-to-NeRF translation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.511, + 0.893, + 0.648 + ], + "angle": 0, + "content": "Given \\(N\\) multi-view images \\(\\{\\mathbf{I}_i\\}_{i = 0}^{N - 1}\\) of a scene, we first use Nerfstudio [32] to train the original NeRF. Then we use a plug-and-play image-to-image translator to edit these source images. However, the content generated by the 2D translator may be inconsistent among multi-view images. For example, using different initial noise, the 2D translator [1] may generate different content for image editing, which makes it difficult to ensure the 3D consistency between different view directions in the 3D scene." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.651, + 0.894, + 0.743 + ], + "angle": 0, + "content": "To ensure the 3D consistency and rendering quality, we propose to model the distribution of the underlying 3D edits through a generative model that can cover all possible edited NeRFs, by learning an edit code for each edited image so that the generated content can be controlled by this edit code during the NeRF-to-NeRF translation process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.745, + 0.895, + 0.901 + ], + "angle": 0, + "content": "For each view \\(\\in [0,N - 1]\\), we generate \\(M\\) edited images, resulting in a group of the translated image set \\(\\{\\{\\mathbf{S}_i^j\\}_{j = 0}^{M - 1}\\}_{i = 0}^{N - 1}\\). Then we design a Latent Distill Module described in Sec. 3.1 to map each translated image \\(\\mathbf{S}_i^j\\) into an edit code vector \\(\\mathbf{z}_i^j\\) and design a KL loss \\(\\mathcal{L}_{\\mathrm{KL}}\\) to constrain those edit code vectors to a Gaussian distribution. Conditioned on the edit code \\(\\mathbf{z}_i^j\\), we perform NeRF-to-NeRF translation in Sec. 3.2 by rendering multi-view images \\(\\{\\mathbf{C}_i\\}_{i = 0}^{N - 1}\\) and optimize the translated NeRF by three loss functions: the reconstruction loss \\(\\mathcal{L}_{\\mathrm{recon}}\\), the adversarial" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5107" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.09, + 0.45, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.311, + 0.471, + 0.395 + ], + "angle": 0, + "content": "Figure 3. Illustration of our proposed contrastive loss functions. Regarding the multi-view rendered images \\(\\mathbf{C}_i^j\\) and \\(\\mathbf{C}_l^j\\) sharing the same edit code, we resend them to our Latent Distill Module to extract \\(\\mathbf{z}_i^j\\) and \\(\\mathbf{z}_l^j\\), and aggregate them via \\(\\mathcal{L}_{\\mathrm{contr}}^{\\mathrm{att}}\\). In addition, for \\(\\mathbf{S}_i^k\\) whose editing style varies from \\(\\mathbf{S}_i^j\\), \\(\\mathcal{L}_{\\mathrm{contr}}^{\\mathrm{rep}}\\) increases the distance between edit codes of them." + }, + { + "type": "image", + "bbox": [ + 0.09, + 0.414, + 0.459, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.538, + 0.47, + 0.622 + ], + "angle": 0, + "content": "Figure 4. Illustration of our proposed conditional adversarial loss functions. Our conditional discriminator distinguishes artifacts such as blur and distortion in novel-view rendered image \\(\\mathbf{C}_l^j\\) compared with target image \\(\\mathbf{S}_l^j\\). \\(\\mathbf{S}_l^j\\) and \\(\\mathbf{S}_l^k\\) are edited with same view but various styles, the latter serves as the condition to concatenate with \\(\\mathbf{C}_l^j\\) and \\(\\mathbf{S}_l^j\\) and manufacture fake and real pairs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.651, + 0.47, + 0.727 + ], + "angle": 0, + "content": "loss \\(\\mathcal{L}_{\\mathrm{AD}}\\), and the contrastive loss \\(\\mathcal{L}_{\\mathrm{contr}}\\). After the optimization of the translated NeRF, as described in Sec. 3.3, we can sample an edit code \\(\\mathbf{z}\\) from Gaussian distribution and render the corresponding edited 3D scene with high quality and multi-view consistency in the inference stage." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.74, + 0.279, + 0.755 + ], + "angle": 0, + "content": "3.1. Latent Distill Module" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Image Translation. As illustrated in Fig. 2, GenN2N is a unified framework for NeRF-to-NeRF translation, in which the core is to perform a 2D image-to-image translation and lift 2D edits into universal 3D NeRF-to-NeRF translation. Given the source multi-view image set \\(\\{\\mathbf{I}_i\\}_{i=0}^{N-1}\\) of a NeRF scene, we first perform image editing M times for each view using a plug-and-play 2D image-to-image translator, producing a group of translated image set \\(\\{\\{\\mathbf{S}_i^j\\}_{j=0}^{M-1}\\}_{i=0}^{N-1}\\). In this paper, we use several 2D translation tasks to show" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "the adaptability of our GenN2N: text-driven editing, superresolution, colorization and inpainting. For more details about those 2D image editing methods, please refer to the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.153, + 0.893, + 0.321 + ], + "angle": 0, + "content": "Edit Code. Since 2D image-to-image translation may generate different content even with the same editing target, causing the inconsistency problem in the 3D scene. We propose to map each edited image \\(\\mathbf{S}_i^j\\) into a latent feature vector named edit code to characterize these diverse editorings. We employ the off-the-shelf VAE encoder from stable diffusion [28] to extract the feature from \\(\\mathbf{S}_i^j\\) and then apply a tiny MLP network to produce this edit code \\(\\mathbf{z}_i^j \\in \\mathbb{R}^{64}\\). During the training process, we keep the pre-trained encoder fixed and only optimize the parameters of the tiny MLP network. This mapping process can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.603, + 0.333, + 0.892, + 0.353 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {i} ^ {j} = \\mathcal {D} \\left(\\mathbf {S} _ {i} ^ {j}\\right) = \\mathcal {M} \\left(\\mathcal {E} \\left(\\mathbf {S} _ {i} ^ {j}\\right)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.365, + 0.892, + 0.395 + ], + "angle": 0, + "content": "where \\(\\mathcal{D}\\) represent this mapping process, \\(\\mathcal{E}\\) is the fixed encoder, and \\(\\mathcal{M}\\) is the learnable tiny MLP." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.397, + 0.892, + 0.473 + ], + "angle": 0, + "content": "KL loss. In order to facilitate effective sampling of the edit code so as to control the editing diversity of our NeRF-to-NeRF translation, we need to constrain the edit code to a well-defined distribution. Thus we design a KL loss to encourage \\(\\mathbf{z}_i^j\\) to approximate a Gaussian distribution:" + }, + { + "type": "equation", + "bbox": [ + 0.504, + 0.486, + 0.891, + 0.532 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {K L}} = \\mathbb {E} _ {\\mathbf {S} \\in \\{\\left\\{\\mathbf {S} _ {i} ^ {j} \\right\\} _ {j = 0} ^ {M - 1} \\} _ {i = 0} ^ {N - 1}} [ P (\\mathbf {z} _ {\\text {n o r m a l}}) \\log \\left(\\frac {P \\left(\\mathbf {z} _ {\\text {n o r m a l}}\\right)}{P \\left(\\mathcal {D} (\\mathbf {S})\\right)}\\right) ], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.532, + 0.892, + 0.576 + ], + "angle": 0, + "content": "where \\( P(\\mathbf{z}_{\\text{normal}}) \\) denotes probability distribution of the standard Gaussian distribution in \\( \\mathbb{R}^{64} \\) and \\( P(\\mathcal{D}(\\mathbf{S})) \\) means probability distribution of the extracted edit codes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.578, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Contrastive loss. It is not assured that edit codes \\(\\mathbf{z}\\) obtained from the Latent Distill Module contain only the editing information while excluding viewpoint-related effects. However, since the translated NeRF utilizes \\(\\mathbf{z}\\) to edit scenes, it yields instability if \\(\\mathbf{z}\\) violently changes given images that are similar in appearance but different in viewpoints. To ensure the latent code does not depend on 2D viewpoints but truly reflects the 3D edits, we regularize the latent code through a contrastive learning scheme. Specifically, we reduce the distance between edit codes of different-view rendered images from a translated NeRF that share the same edit code, while increasing the distance between same-view images that are multi-time edited by the 2D image-to-image translator. As illustrated in Fig. 3, given an edit code \\(\\mathbf{z}_i^j\\) extracted from the \\(i\\)-th input view at the \\(j\\)-th edited image \\(\\mathbf{S}_i^j\\), we render multi-view images \\(\\{\\mathbf{C}_l^j\\}_{l=0}^{N-1}\\) using the translated NeRF conditioned on \\(\\mathbf{z}_i^j\\). Then we employ contrastive learning to encourage the edit code \\(\\mathbf{z}_i^j\\) to be close to \\(\\{\\mathbf{z}_l^j\\}_{l=0}^{N-1}\\) extracted from \\(\\{\\mathbf{C}_l^j\\}_{l=0}^{N-1}\\), while being distinct from the edit codes \\(\\{\\mathbf{z}_i^k\\}_{k=0}^{M-1}\\) extracted from \\(\\{\\mathbf{S}_i^k\\}_{k=0}^{M-1}\\), where \\(k \\neq j\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5108" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.356, + 0.089, + 0.643, + 0.103 + ], + "angle": 0, + "content": "\"What would he look like as a bearded man\"" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.107, + 0.169, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.107, + 0.268, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.271, + 0.107, + 0.356, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.107, + 0.446, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.448, + 0.107, + 0.533, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.107, + 0.622, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.107, + 0.71, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.107, + 0.802, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.804, + 0.107, + 0.888, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.391, + 0.206, + 0.597, + 0.22 + ], + "angle": 0, + "content": "\"Make it an Egyptian sculpture\"" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.223, + 0.169, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.083, + 0.322, + 0.168, + 0.335 + ], + "angle": 0, + "content": "Original NeRF" + }, + { + "type": "image", + "bbox": [ + 0.184, + 0.223, + 0.268, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.238, + 0.323, + 0.3, + 0.334 + ], + "angle": 0, + "content": "Inference 1" + }, + { + "type": "image", + "bbox": [ + 0.271, + 0.223, + 0.356, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.223, + 0.446, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.448, + 0.223, + 0.533, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.223, + 0.622, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.223, + 0.71, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.223, + 0.802, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.804, + 0.223, + 0.888, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.415, + 0.323, + 0.479, + 0.334 + ], + "angle": 0, + "content": "Inference 2" + }, + { + "type": "image_caption", + "bbox": [ + 0.593, + 0.323, + 0.656, + 0.333 + ], + "angle": 0, + "content": "Inference 3" + }, + { + "type": "image_caption", + "bbox": [ + 0.77, + 0.323, + 0.835, + 0.333 + ], + "angle": 0, + "content": "Inference 4" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.348, + 0.892, + 0.377 + ], + "angle": 0, + "content": "Figure 5. Text-Driven Editing. We sample 4 inference results for both text-driven editing tasks. The diversity of geometry and appearance showcases awesome generative ability of GenN2N, on the premise of maintaining the 3D consistency between different viewpoints." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.391, + 0.21, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.543, + 0.186, + 0.554 + ], + "angle": 0, + "content": "Original NeRF" + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.391, + 0.353, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.468, + 0.352, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.258, + 0.543, + 0.318, + 0.553 + ], + "angle": 0, + "content": "Inference 1" + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.391, + 0.486, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.467, + 0.486, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.392, + 0.543, + 0.453, + 0.553 + ], + "angle": 0, + "content": "Inference 2" + }, + { + "type": "image", + "bbox": [ + 0.492, + 0.391, + 0.62, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.492, + 0.468, + 0.62, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.525, + 0.543, + 0.587, + 0.553 + ], + "angle": 0, + "content": "Inference 3" + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.391, + 0.754, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.467, + 0.754, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.66, + 0.543, + 0.721, + 0.553 + ], + "angle": 0, + "content": "Inference 4" + }, + { + "type": "image", + "bbox": [ + 0.761, + 0.391, + 0.888, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.761, + 0.468, + 0.888, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.794, + 0.543, + 0.855, + 0.553 + ], + "angle": 0, + "content": "Inference 5" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.567, + 0.892, + 0.595 + ], + "angle": 0, + "content": "Figure 6. Colorization. Our method colorizes the gray-scale 3D scene consistently across views. By changing the edit code during inference, diverse colorized scenes can be rendered with satisfying photorealism and reasonably rich colors." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.622, + 0.465, + 0.637 + ], + "angle": 0, + "content": "Specifically, our contrastive loss is designed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.644, + 0.47, + 0.718 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {c o n t r}} = \\mathcal {L} _ {\\text {c o n t r}} ^ {\\text {a t t}} + \\mathcal {L} _ {\\text {c o n t r}} ^ {\\text {r e p}} \\\\ = \\sum_ {l = 0} ^ {N - 1} \\left\\| \\mathbf {z} _ {i} ^ {j} - \\mathbf {z} _ {l} ^ {j} \\right\\| _ {2} ^ {2} + \\sum_ {k = 0} ^ {M - 1} \\max (0, \\alpha - \\left\\| \\mathbf {z} _ {i} ^ {j} - \\mathbf {z} _ {i} ^ {k} \\right\\| _ {2} ^ {2}), \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.468, + 0.75 + ], + "angle": 0, + "content": "where \\(\\alpha\\) represents the margin that encourages the difference in features, and \\(k\\neq j\\)" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.757, + 0.319, + 0.772 + ], + "angle": 0, + "content": "3.2. NeRF-to-NeRF translation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.469, + 0.884 + ], + "angle": 0, + "content": "Translated NeRF. After 2D image-to-image translation, we need to lift these 2D edits to the 3D NeRF. For this purpose, we propose to modify the original NeRF as a translated NeRF that takes the edit code \\(\\mathbf{z}\\) as input and generates the translated 3D scene according to the edit code. We refer readers to the supplementary for more details about the network architecture." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Reconstruction loss. Given an edit code \\(\\mathbf{z}_i^j\\) extracted from" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.621, + 0.892, + 0.7 + ], + "angle": 0, + "content": "the edited image \\(\\mathbf{S}_i^j\\), we can generate a translated NeRF to render \\(\\mathbf{C}_i^j\\) from the same viewpoint. Then we define the reconstruction loss as the L1 normalization and Learned Perceptual Image Patch Similarity (LPIPS) [44] between the rendered image \\(\\mathbf{C}_i^j\\) and the edited image \\(\\mathbf{S}_i^j\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.71, + 0.892, + 0.756 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {r e c o n}} = \\mathcal {L} _ {\\mathrm {L} 1} + \\mathcal {L} _ {\\mathrm {L P I P S}} \\\\ = \\left\\| \\mathbf {C} _ {i} ^ {j} - \\mathbf {S} _ {i} ^ {j} \\right\\| _ {1} + L P I P S [ \\mathcal {P} (\\mathbf {C} _ {i} ^ {j}) - \\mathcal {P} (\\mathbf {S} _ {i} ^ {j}) ], \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathcal{P}\\) means a patch sampled from the image. Note that due to the lack of 3D consistency of the edited multiview image, the supervision of the edited image from other viewpoints \\(\\{\\mathbf{S}_l^j\\}_{l\\neq i}\\) will lead to conflicts in pixel-space optimization. Therefore, we only employ reconstruction loss on the same view image \\(\\mathbf{S}_i^j\\) to optimize the translated NeRF. Adversarial loss. Since the 3D consistency of edited multiview images is not assured, relying solely on the reconstruction loss on the same view often leads to blurry or distorted" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5109" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.206, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.106, + 0.156, + 0.18, + 0.166 + ], + "angle": 0, + "content": "Original NeRF" + }, + { + "type": "image", + "bbox": [ + 0.212, + 0.09, + 0.336, + 0.155 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.156, + 0.332, + 0.165 + ], + "angle": 0, + "content": "Instruct-NeRF2NeRF 1" + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.091, + 0.467, + 0.155 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.346, + 0.156, + 0.464, + 0.165 + ], + "angle": 0, + "content": "Instruct-NeRF2NeRF 2" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.167, + 0.206, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.232, + 0.185, + 0.242 + ], + "angle": 0, + "content": "Ours inference 1" + }, + { + "type": "image", + "bbox": [ + 0.212, + 0.167, + 0.337, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.231, + 0.232, + 0.316, + 0.242 + ], + "angle": 0, + "content": "Ours inference 2" + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.167, + 0.467, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.361, + 0.232, + 0.446, + 0.242 + ], + "angle": 0, + "content": "Ours inference 3" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.253, + 0.47, + 0.296 + ], + "angle": 0, + "content": "Figure 7. Comparisons with baselines of text-driven NeRF editing. We compare our method with Instruct-NeRF2NeRF [9] in the editing by using the text prompt \"Make it Spring\"." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.308, + 0.485, + 0.408 + ], + "angle": 0, + "content": "
MethodCLIP Text-Image Direction Similarity↑CLIP Direction Consistency ↑FID ↓
InstructPix2Pix [2]+NeRF0.16690.8475270.542
Instruct-NeRF2NeRF0.20210.9828148.021
Ours w/o Ladv0.19200.9657162.275
Ours w/o Lcontr0.20070.9749156.524
Ours0.20890.9864137.740
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.419, + 0.47, + 0.475 + ], + "angle": 0, + "content": "Table 1. Quantitative results on text-driven editing. We compare our method with the naive method of directly combining InstructPix2Pix [2] with NeRF and the state-of-the-art method Instruct-NeRF2NeRF [9]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.5, + 0.469, + 0.621 + ], + "angle": 0, + "content": "artifacts on novel views. Previous research demonstrates the effectiveness of conditional adversarial training in preventing the production of blurry rendered images resulting from conflicts that arise from noise in the camera extrinsic when performing image supervision from different viewpoints [11]. The function of the condition is to guide discriminator with fine-grained information from the same viewpoint, thus preventing GAN mode collapse." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.621, + 0.469, + 0.817 + ], + "angle": 0, + "content": "It inspires us to incorporate conditional adversarial loss on rendered images from the translated NeRF, which is conducive to distinguish artifacts in rendered images. As illustrated in Fig.4, the discriminator \\(\\mathbf{D}\\) takes into real pairs and fake pairs. Each real pair \\(\\mathbf{R}\\) consists of \\(\\mathbf{S}^j\\) and \\(\\mathbf{S}^j -\\mathbf{S}^k\\) where \\(\\mathbf{S}^j\\in \\{\\mathbf{S}_i^j\\}_{i = 0}^{N - 1}\\) and \\(\\mathbf{S}^k\\in \\{\\mathbf{S}_i^k\\}_{i = 0}^{N - 1}\\) are from two sets of edited images from the image translation. Similarly, each fake pair \\(\\mathbf{F}\\) consists of \\(\\mathbf{C}^j\\) and \\(\\mathbf{C}^j -\\mathbf{S}^k\\) in which \\(\\mathbf{C}^j\\in \\{\\mathbf{C}_i^j\\}_{i = 0}^{N - 1}\\) is generated by translated NeRF. Note that the images in the same pair come from the same viewpoint. The pairs are concatenated in RGB channels and fed into the discriminator. We optimize the discriminator \\(\\mathbf{D}\\) and translated NeRF with the objective functions below:" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.826, + 0.469, + 0.85 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {A D - D}} = \\mathbb {E} _ {\\mathbf {R}} [ - l o g (\\mathbf {D} (\\mathbf {R})) ] + \\mathbb {E} _ {\\mathbf {F}} [ - l o g (1 - \\mathbf {D} (\\mathbf {F})) ], \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.846, + 0.273, + 0.862 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {A D - G}} = \\mathbb {E} _ {\\mathbf {F}} [ - l o g (\\mathbf {D} (\\mathbf {F})) ].\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Optimization. During the training process, we jointly optimize the loss functions mentioned above: \\(\\mathcal{L}_{\\mathrm{KL}}\\) and \\(\\mathcal{L}_{\\mathrm{contr}}\\)" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.091, + 0.628, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.524, + 0.161, + 0.598, + 0.171 + ], + "angle": 0, + "content": "Original NeRF" + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.091, + 0.758, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.643, + 0.161, + 0.75, + 0.17 + ], + "angle": 0, + "content": "Instruct-NeRF2NeRF" + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.091, + 0.888, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.795, + 0.161, + 0.858, + 0.17 + ], + "angle": 0, + "content": "PaletteNeRF" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.172, + 0.628, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.524, + 0.243, + 0.607, + 0.252 + ], + "angle": 0, + "content": "Ours inference 1" + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.172, + 0.758, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.654, + 0.243, + 0.737, + 0.252 + ], + "angle": 0, + "content": "Ours inference 2" + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.172, + 0.888, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.785, + 0.243, + 0.867, + 0.252 + ], + "angle": 0, + "content": "Ours inference 3" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.264, + 0.892, + 0.292 + ], + "angle": 0, + "content": "Figure 8. Comparisons with baselines of NeRF colorization. We compare with PaletteNeRF[16] in colorization." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.304, + 0.907, + 0.423 + ], + "angle": 0, + "content": "
MethodCF ↑FID ↓
DDColor [14]+NeRF40.435148.957
Instruct-NeRF2NeRF45.599201.456
PaletteNeRF [16]39.654-
Ours w/o Ladv35.031137.740
Ours w/o Lcontr34.829105.750
Ours65.09935.041
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.433, + 0.892, + 0.503 + ], + "angle": 0, + "content": "Table 2. Quantitative results on colorization. We colorize images with the translator, DDcolor [14], and edit NeRF by directly optimizing NeRF, Instruct-NeRF2NeRF [9] and our NeRF translation method. The quantitative comparison is conducted between these methods as well as PaletteNeRF [16]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.527, + 0.892, + 0.572 + ], + "angle": 0, + "content": "for the edit code, \\(\\mathcal{L}_{\\mathrm{recon}}\\) and \\(\\mathcal{L}_{\\mathrm{AD - G}}\\) for the translated NeRF, and \\(\\mathcal{L}_{\\mathrm{AD - D}}\\) for the discriminator. The total loss formula is expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.544, + 0.582, + 0.891, + 0.598 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {K L}} + \\mathcal {L} _ {\\text {r e c o n}} + \\mathcal {L} _ {\\mathrm {A D - G}} + \\mathcal {L} _ {\\mathrm {A D - D}} + \\mathcal {L} _ {\\text {c o n t r}}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.606, + 0.892, + 0.682 + ], + "angle": 0, + "content": "where we assign each regularization term the weight of 1.0, 1.0, 0.1, 0.1, 0.1 in all of our experiments. The weights can be adjusted to prioritize different aspects of the training objective, such as reconstruction accuracy, adversarial training, and perceptual quality." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.69, + 0.609, + 0.704 + ], + "angle": 0, + "content": "3.3. Inference" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.712, + 0.892, + 0.803 + ], + "angle": 0, + "content": "After the optimization of our GenN2N, the translated NeRF is optimized to be able to render the target scene conditioned on the edit code. As shown in Fig. 1, users can simply sample an edit code from the Gaussian distribution and use the translated NeRF to render the 3D scene with high-quality and multi-view 3D consistency." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.816, + 0.632, + 0.832 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Our proposed GenN2N is a unified NeRF-to-NeRF translation framework which can support various NeRF editing tasks. In this paper, we demonstrate the effectiveness of GenN2N by a suite of challenging tasks:" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5110" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.177, + 0.146 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.143, + 0.147, + 0.21, + 0.155 + ], + "angle": 0, + "content": "Original NeRF" + }, + { + "type": "image", + "bbox": [ + 0.178, + 0.091, + 0.272, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.091, + 0.371, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.091, + 0.467, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.322, + 0.147, + 0.419, + 0.155 + ], + "angle": 0, + "content": "Instruct-NeRF2NeRF" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.156, + 0.176, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.153, + 0.212, + 0.199, + 0.221 + ], + "angle": 0, + "content": "NeRF-SR" + }, + { + "type": "image", + "bbox": [ + 0.178, + 0.156, + 0.272, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.156, + 0.371, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.156, + 0.467, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.359, + 0.213, + 0.383, + 0.221 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.232, + 0.47, + 0.261 + ], + "angle": 0, + "content": "Figure 9. Comparisons with baselines of NeRF super-resolution. We compare with NeRF-SR [34] in super-resolution." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.274, + 0.485, + 0.367 + ], + "angle": 0, + "content": "
MethodPSNR ↑SSIM ↑LPIPS ↓
ResShift [40]+NeRF19.9780.5350.1156
Instruct-NeRF2NeRF20.2990.6420.2732
NeRF-SR [34]27.9570.8970.0937
Ours w/o Ladv12.5550.6630.2001
Ours w/o Lcontr15.3720.6620.1834
Ours28.5010.9130.0748
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.378, + 0.47, + 0.448 + ], + "angle": 0, + "content": "Table 3. Quantitative results on super-resolution. We improve image resolution with ResShift [40] and edit NeRF by directly optimizing NeRF, Instruct-NeRF2NeRF [9] and our NeRF translation method. The quantitative comparison is conducted between these methods as well as NeRF-SR [34]." + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.474, + 0.469, + 0.504 + ], + "angle": 0, + "content": "(1) Text-driven Editing edits the given NeRF scene to a set of NeRF scenes according to the text instruction." + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.506, + 0.469, + 0.534 + ], + "angle": 0, + "content": "(2) Colorization transforms a gray-scale NeRF scene to a set of plausible color NeRF scenes." + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.535, + 0.469, + 0.565 + ], + "angle": 0, + "content": "(3) Super-resolution enhances the resolution of NeRF and enables multiple plausible outcomes." + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.566, + 0.469, + 0.594 + ], + "angle": 0, + "content": "(4) Inpainting fills in user-specified masked regions in the NeRF scene with realistic content." + }, + { + "type": "list", + "bbox": [ + 0.082, + 0.474, + 0.469, + 0.594 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.596, + 0.47, + 0.793 + ], + "angle": 0, + "content": "We achieve those tasks by simply changing the plug-and-play 2D image translator in our framework, without any additional task-specific design. Previous studies have extensively explored some of these issues like text-driven editing, colorization, super-resolution, and inpainting. However, there is rarely a unified framework that can achieve all these tasks with strong performance, high quality, and plausible multi-view consistent 3D structure. Furthermore, GenN2N can also perform zooming out and text-driven inpainting in NeRF-to-NeRF translation, which were not explored in prior research. We refer readers to the supplementary materials for detailed experiment settings, dataset settings and implementation details." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.217, + 0.818 + ], + "angle": 0, + "content": "4.1. Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Text-driven Editing. We achieve text-driven editing of the given NeRF by using InstructPix2Pix [2] as the 2D image-to-image translator in our framework. We compared our approach to a naive solution, which involves optimizing a NeRF with edited images via InstructPix2Pix. However," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.228 + ], + "angle": 0, + "content": "this naive approach leads to a 3D inconsistency problem among different edits. While Instruct-NeRF2NeRF [9] proposed an iterative updating mechanism to address this issue, it falls short in capturing the diversity of different edits, making it challenging to ensure the quality of the outcomes. To evaluate our method, we conducted experiments on the Face [9] and Fangzhou [35] self-portrait datasets, and Farm [9] dataset, comparing GenN2N with the state-of-the-art NeRF editing method Instruct-NeRF2NeRF [9]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.23, + 0.893, + 0.503 + ], + "angle": 0, + "content": "Quantitative results are presented in Table 1, where we employed CLIP Text-Image Direction Similarity [9], CLIP Direction Consistency [9], and Fréchet Inception Distance (FID) [10] as evaluation metrics. The results highlight the superior performance of GenN2N over other methods, demonstrating its effectiveness in producing high-quality 3D text-driven editing results. Furthermore, we provide a qualitative comparison between GenN2N and Instruct-NeRF2NeRF [9] in Figure 7. Notably, Instruct-NeRF2NeRF was trained twice, but the results are nearly identical. In contrast, GenN2N can render edited scenes with various effects, consistent with the input text instruction, by inferring with different edit codes sampled from a Gaussian distribution. To further illustrate the generative capability of GenN2N, we showcase additional results of our inference renderings in Figure 5, demonstrating its diverse generative ability in terms of appearance and geometry." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.505, + 0.893, + 0.702 + ], + "angle": 0, + "content": "Colorization. For NeRF colorization, GenN2N uses DD-Color [14] as the 2D image-to-image translator. CoRF [8] and Palette-NeRF [16] do a similar task and we compare with Palette-NeRF in Table 2. The Fréchet Inception Distance (FID) [10] and colorfulness score (CF) [17] are used to measure the distribution similarity and vividness of generated images. We show visual comparison results in Fig. 6 and Fig. 8. We can find that with different edit codes, the scene can be rendered in different color styles. It is noticeable that with the same edit code, the color rendered from different views is consistent. This strongly demonstrates the effectiveness of our method in translating NeRF while keeping the 3D consistency of the scene." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Super-resolution. When only low-resolution images are available, our methods can boost NeRF in reconstructing scenes at higher resolution, while keeping view consistency and avoiding blurry outputs. We achieve this by employing ResShift [40] as the 2D image-to-image translator in GenN2N. Following state-of-the-art method NeRF-SR [34], we conduct experiments on \\(LLFF\\) dataset [23], using PSNR, SSIM, and LPIPS as evaluation metrics. As shown in Table 3, GenN2N obtains NeRF-to-NeRF translation with higher performance than NeRF-SR [34]. Moreover, we also provide qualitative comparison results in Fig. 9, where GenN2N produces clearer and more realistic rendering results than previous methods." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "5111" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.091, + 0.175, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.176, + 0.091, + 0.272, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.091, + 0.371, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.091, + 0.467, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.137, + 0.148, + 0.216, + 0.158 + ], + "angle": 0, + "content": "Original NeRF" + }, + { + "type": "image_caption", + "bbox": [ + 0.314, + 0.147, + 0.428, + 0.157 + ], + "angle": 0, + "content": "Instruct-NeRF2NeRF" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.16, + 0.175, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.176, + 0.159, + 0.272, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.159, + 0.371, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.159, + 0.467, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.146, + 0.215, + 0.208, + 0.225 + ], + "angle": 0, + "content": "SPin-NeRF" + }, + { + "type": "image_caption", + "bbox": [ + 0.356, + 0.215, + 0.384, + 0.225 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.236, + 0.47, + 0.265 + ], + "angle": 0, + "content": "Figure 10. Comparisons with baselines of NeRF inpainting. We compare with SPIn-NeRF [24] in inpainting." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.277, + 0.485, + 0.35 + ], + "angle": 0, + "content": "
MethodPSNR ↑SSIM ↑LPIPS ↓
LaMa [31]+NeRF18.9830.37060.1730
Instruct-NeRF2NeRF16.7340.30880.2750
SPin-NeRF [24]24.3690.72170.1754
Ours26.8680.81370.1284
" + }, + { + "type": "table_caption", + "bbox": [ + 0.118, + 0.361, + 0.429, + 0.375 + ], + "angle": 0, + "content": "Table 4. Quantitative results on NeRF inpainting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.4, + 0.473, + 0.658 + ], + "angle": 0, + "content": "Inpainting. The goal of NeRF Inpainting is to fill the 3D content of regions specified by users. SPin-NeRF [24] achieves this through a multi-step process: it employs SAM [15] for object segmentation, utilizes LaMa [31] to paint the background content in multi-view images, and subsequently trains the NeRF model with color, depth, and perceptual cues. In our experiments, we use SAM and LaMa as the 2D image-to-image translator in our GenN2N, which is the same setting as SPin-NeRF [24]. Quantitative comparisons on statue dataset [24] are shown in Table 4, where GenN2N achieves superior PSNR and SSIM scores than SPin-NeRF [24], highlighting the effectiveness of our GenN2N framework. In addition, qualitative results are showcased in Fig. 10 revealing that while SPin-NeRF [24] fails to generate reasonable content behind the masked object, our GenN2N produces realistic content in the same region with fine multi-view consistency." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.667, + 0.241, + 0.681 + ], + "angle": 0, + "content": "4.2. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.69, + 0.469, + 0.749 + ], + "angle": 0, + "content": "We conduct comprehensive ablation experiments to validate the designs of each component in GenN2N. Due to space limitations, we only highlight the essential aspects below. Please refer to supplementary for more details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.902 + ], + "angle": 0, + "content": "The Contrastive Loss. We demonstrate the advantages of incorporating our proposed contrastive loss in Table 1, 2, 3. The motivation is to disentangle the camera view and edit information present in the latent space. We achieve this by reducing the distance between edit codes of different-view rendered images from a translated NeRF that shares the same edit code, while increasing the distance between same-view images that are edited by the 2D translator with diverse edit styles. As demonstrated in Tables, the absence of contrastive loss leads to the generation of blurry areas" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.182 + ], + "angle": 0, + "content": "in the rendered images, resulting in a decrease in the metric scores. This blurriness can be attributed to the inclusion of pose information within the edit code z. By incorporating the contrastive loss, our method successfully achieves a uniform appearance with different observing views under the same style latent z." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.187, + 0.893, + 0.414 + ], + "angle": 0, + "content": "Discriminator for Novel Views. We demonstrate the effectiveness of employing a conditional discriminator to address artifacts caused by inconsistent cross-view edited images and to enhance the quality of novel view rendering images, as depicted in Table 1, 2, 3. The removal of this conditional discriminator results in blurry novel view images with artifacts in the background region. We attribute these undesirable effects to the inability of current image-to-image translation methods, such as InstructPix2Pix [2], to produce image editing consistently across multi-view images. To mitigate these issues, we introduce a conditional discriminator between rendered images from the translated NeRF and edited images from the 2D image-to-image translator. This inclusion successfully eliminates artifacts and enhances the image quality of rendered images from the translated NeRF." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.438, + 0.634, + 0.454 + ], + "angle": 0, + "content": "4.3. Applications" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.465, + 0.892, + 0.617 + ], + "angle": 0, + "content": "We demonstrate the versatility and robustness of GenN2N by exploring two translation applications: Zoom Out and Text-Driven Inpainting. While existing 2D translators [1, 29] can complete these tasks, 3D editing has not been explored. We achieve these tasks by incorporating Blended Latent Diffusion [1] as the 2D image-to-image translator, enabling us to generate diverse and high-quality content with multi-view consistency. Please refer to our supplementary materials for the results of these applications due to space constraints." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.645, + 0.627, + 0.66 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We introduce GenN2N, a unified NeRF-to-NeRF translation framework that can handle various NeRF editing tasks. Unlike previous task-specific approaches, our framework uses an image-to-image translator for 2D editing and integrates the results into 3D NeRF space. To address the challenge of ensuring 3D consistency, we propose modeling the distribution of 3D edited NeRFs from 2D edited images using our novel techniques. After optimization, users can sample from the conditional generative model to obtain diverse 3D editing results with high rendering quality and multi-view consistency. Our experiments demonstrate that GenN2N outperforms existing task-specific methods on various editing tasks, including text-driven editing, colorization, super-resolution, and inpainting, in terms of efficiency, quality, and diversity." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5112" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM Transactions on Graphics (TOG), 42 (4):1-11, 2023. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.158, + 0.47, + 0.2 + ], + "angle": 0, + "content": "[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. arXiv preprint arXiv:2211.09800, 2022. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.471, + 0.27 + ], + "angle": 0, + "content": "[3] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.271, + 0.471, + 0.34 + ], + "angle": 0, + "content": "[4] Shengqu Cai, Anton Obukhov, Dengxin Dai, and Luc Van Gool. Pix2nerf: Unsupervised conditional p-gan for single image to neural radiance fields translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3981-3990, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.34, + 0.471, + 0.424 + ], + "angle": 0, + "content": "[5] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.471, + 0.465 + ], + "angle": 0, + "content": "[6] Jun-Kun Chen, Jipeng Lyu, and Yu-Xiong Wang. Neuraleditor: Editing neural radiance fields via manipulating point clouds. arXiv preprint arXiv:2305.03049, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.466, + 0.471, + 0.536 + ], + "angle": 0, + "content": "[7] Pei-Ze Chiang, Meng-Shiun Tsai, Hung-Yu Tseng, Wei-Sheng Lai, and Wei-Chen Chiu. Stylizing 3d scene via implicit representation and hypernetwork. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1475-1484, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.536, + 0.471, + 0.592 + ], + "angle": 0, + "content": "[8] Ankit Dhiman, R Srinath, Srinjay Sarkar, Lokesh R. Boregowda, and R Venkatesh Babu. Corf: Colorizing radiance fields using knowledge distillation. ArXiv, abs/2309.07668, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.593, + 0.471, + 0.648 + ], + "angle": 0, + "content": "[9] Ayaan Haque, Matthew Tancik, Alexei A Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. arXiv preprint arXiv:2303.12789, 2023.2,3,6,7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.649, + 0.47, + 0.718 + ], + "angle": 0, + "content": "[10] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.719, + 0.471, + 0.802 + ], + "angle": 0, + "content": "[11] Jingwei Huang, Justus Thies, Angela Dai, Abhijit Kundu, Chiyu Jiang, Leonidas J Guibas, Matthias Nießner, Thomas Funkhouser, et al. Adversarial texture optimization from rgb-d scans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1559-1568, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.471, + 0.873 + ], + "angle": 0, + "content": "[12] Yi-Hua Huang, Yue He, Yu-Jie Yuan, Yu-Kun Lai, and Lin Gao. Stylizednerf: consistent 3d scene stylization as stylized nerf via 2d-3d mutual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18342–18352, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.903 + ], + "angle": 0, + "content": "[13] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object genera" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "tion with dream fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 867-876, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.192 + ], + "angle": 0, + "content": "[14] Xiaoyang Kang, Tao Yang, Wenqi Ouyang, Peiran Ren, Lingzhi Li, and Xuansong Xie. Ddcolor: Towards photorealistic and semantic-aware image colorization via dual decoders. arXiv preprint arXiv:2212.11613, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.192, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[15] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.249, + 0.892, + 0.319 + ], + "angle": 0, + "content": "[16] Zhengfei Kuang, Fujun Luan, Sai Bi, Zhixin Shu, Gordon Wetzstein, and Kalyan Sunkavalli. Palettenerf: Palette-based appearance editing of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20691-20700, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.32, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[17] Gustav Larsson, Michael Maire, and Gregory Shakhnarovich. Learning representations for automatic colorization. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 577-593. Springer, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.405, + 0.892, + 0.475 + ], + "angle": 0, + "content": "[18] Verica Lazova, Vladimir Guzov, Kyle Olszewski, Sergey Tulyakov, and Gerard Pons-Moll. Control-nerf: Editable feature volumes for scene rendering and manipulation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4340–4350, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.476, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[19] Hao-Kang Liu, I Shen, Bing-Yu Chen, et al. Nerf-in: Free-form nerf inpainting with rgb-d priors. arXiv preprint arXiv:2206.04901, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.518, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[20] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. arXiv preprint arXiv:2303.11328, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.576, + 0.892, + 0.645 + ], + "angle": 0, + "content": "[21] Steven Liu, Xiuming Zhang, Zhoutong Zhang, Richard Zhang, Jun-Yan Zhu, and Bryan Russell. Editing conditional radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5773-5783, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.646, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[22] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Realfusion: \\(360^{\\circ}\\) reconstruction of any object from a single image. arXiv e-prints, pages arXiv-2302, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.703, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[23] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.774, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[24] Ashkan Mirzaei, Tristan Aumentado-Armstrong, Konstantinos G Derpanis, Jonathan Kelly, Marcus A Brubaker, Igor Gilitschenski, and Alex Levinshtein. Spin-nerf: Multiview segmentation and perceptual inpainting with neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20669–20679, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.872, + 0.892, + 0.903 + ], + "angle": 0, + "content": "[25] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5113" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "translation. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.164 + ], + "angle": 0, + "content": "[26] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.165, + 0.471, + 0.248 + ], + "angle": 0, + "content": "[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.25, + 0.472, + 0.319 + ], + "angle": 0, + "content": "[28] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.471, + 0.39 + ], + "angle": 0, + "content": "[29] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.392, + 0.471, + 0.447 + ], + "angle": 0, + "content": "[30] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. Advances in Neural Information Processing Systems, 33:20154-20166, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.449, + 0.47, + 0.531 + ], + "angle": 0, + "content": "[31] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. arXiv preprint arXiv:2109.07161, 2021.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.533, + 0.471, + 0.615 + ], + "angle": 0, + "content": "[32] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, et al. Nerfstudio: A modular framework for neural radiance field development. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-12, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.618, + 0.471, + 0.687 + ], + "angle": 0, + "content": "[33] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3835–3844, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.471, + 0.757 + ], + "angle": 0, + "content": "[34] Chen Wang, Xian Wu, Yuan-Chen Guo, Song-Hai Zhang, Yu-Wing Tai, and Shi-Min Hu. Nerf-sr: High quality neural radiance fields using supersampling. In Proceedings of the 30th ACM International Conference on Multimedia, pages 6445-6454, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.759, + 0.47, + 0.815 + ], + "angle": 0, + "content": "[35] Can Wang, Ruixiang Jiang, Mengei Chai, Mingming He, Dongdong Chen, and Jing Liao. Nerf-art: Text-driven neural radiance fields stylization. IEEE Transactions on Visualization and Computer Graphics, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.471, + 0.884 + ], + "angle": 0, + "content": "[36] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3d digital avatars using diffusion. arXiv preprint arXiv:2212.06135, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.887, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[37] Qiling Wu, Jianchao Tan, and Kun Xu. Palettenerf:" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Palette-based color editing for nerfs. arXiv preprint arXiv:2212.12871, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.178 + ], + "angle": 0, + "content": "[38] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.894, + 0.247 + ], + "angle": 0, + "content": "[39] Yu-Jie Yuan, Yang-Tian Sun, Yu-Kun Lai, Yuewen Ma, Rongfei Jia, and Lin Gao. Nerf-editing: geometry editing of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18353-18364, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.894, + 0.303 + ], + "angle": 0, + "content": "[40] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Resshift: Efficient diffusion model for image super-resolution by residual shifting. arXiv preprint arXiv:2307.12348, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.894, + 0.374 + ], + "angle": 0, + "content": "[41] Jiakai Zhang, Xinhang Liu, Xinyi Ye, Fuqiang Zhao, Yanshun Zhang, Minye Wu, Yingliang Zhang, Lan Xu, and Jingyi Yu. Editable free-viewpoint video using a layered neural representation. ACM Transactions on Graphics (TOG), 40(4):1-18, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.376, + 0.894, + 0.445 + ], + "angle": 0, + "content": "[42] Kai Zhang, Nick Kolkin, Sai Bi, Fujun Luan, Zexiang Xu, Eli Shechtman, and Noah Snavely. Arf: Artistic radiance fields. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXI, pages 717-733. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.447, + 0.894, + 0.487 + ], + "angle": 0, + "content": "[43] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.489, + 0.894, + 0.559 + ], + "angle": 0, + "content": "[44] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.56, + 0.894, + 0.602 + ], + "angle": 0, + "content": "[45] Chengwei Zheng, Wenbin Lin, and Feng Xu. Editablenerf: Editing topologically varying neural radiance fields by key points. arXiv preprint arXiv:2212.04247, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.603, + 0.894, + 0.644 + ], + "angle": 0, + "content": "[46] Jingyu Zhuang, Chen Wang, Lingjie Liu, Liang Lin, and Guanbin Li. Dreameditor: Text-driven 3d scene editing with neural fields. arXiv preprint arXiv:2306.13455, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.644 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "5114" + } + ] +] \ No newline at end of file diff --git a/2024/GenN2N_ Generative NeRF2NeRF Translation/720fe634-ca22-4125-bab8-c49c422e71c2_origin.pdf b/2024/GenN2N_ Generative NeRF2NeRF Translation/720fe634-ca22-4125-bab8-c49c422e71c2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2f374d2609c18d7ef70b4617ba3cc1a69b66509f --- /dev/null +++ b/2024/GenN2N_ Generative NeRF2NeRF Translation/720fe634-ca22-4125-bab8-c49c422e71c2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:45cc70c7c162bd2bf3613ef45f121bf4a49fcceaddff526fcda78aa840c7dea5 +size 10448684 diff --git a/2024/GenN2N_ Generative NeRF2NeRF Translation/full.md b/2024/GenN2N_ Generative NeRF2NeRF Translation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..cc21a603939f61f17124258f733ab417c0e46a0b --- /dev/null +++ b/2024/GenN2N_ Generative NeRF2NeRF Translation/full.md @@ -0,0 +1,454 @@ +# GenN2N: Generative NeRF2NeRF Translation + +Xiangyue Liu $^{1}$ Han Xue $^{2}$ Kunming Luo $^{1}$ Ping Tan $^{1\dagger}$ Li Yi $^{2,3,4\dagger}$ + +$^{1}$ Hong Kong University of Science and Technology $^{2}$ Tsinghua University +$^{3}$ Shanghai Artificial Intelligence Laboratory $^{4}$ Shanghai Qi Zhi Institute + +![](images/8ca0ac5a4e3a682fee5599c6ab5ed0c5060ff70bc4ee10d2cfb5617aaa7aa150.jpg) +Input +Output 1 +Output 2 +Output 3 + +![](images/d2ed03faeb021d1cd882c010f8c11c49e4e6d3eb5f10640473784caef20e3b43.jpg) +Input +Input zoom in +Output +Output zoom in +Figure 1. We introduce GenN2N, a unified framework for NeRF-to-NeRF translation, enabling a range of 3D NeRF editing tasks, including text-driven editing, colorization, super-resolution, inpainting, etc. We show at least two rendering views of edited NeRF scenes at inference time. Given a 3D NeRF scene, GenN2N can produce high-quality editing results with suitable multi-view consistency. + +![](images/0c4653d18e861ac74f52cc629cb0c01ddec6feb4d59d833ca8a08d8348335202.jpg) + +![](images/ced5e4e1b34e9e675d51c4effdee96482ae89d6c6181312791da3b8b44b820c1.jpg) + +![](images/bbe829216193b8afbe4b78fbd624a9f4100a2203e96ea105e3c7992549d36d9d.jpg) + +![](images/d73bb4246a5bbc4aaa5627ac0a4a65ec97e627397013eac5045b5c52ca9ff5c2.jpg) + +![](images/334c92776c42b04e72183a28f470290c4c36b14817becffc441caecd7a1a763d.jpg) +aannnnnne aannnnnne +Input + +![](images/92fe8edd7c667b81fa33654f47f1efc84d0933ea7f8e8393eb274fcb58d702b2.jpg) +Output 1 + +![](images/3db7a8a2143ef22b17deadc2495d63a5b98d3aa22ffd082fd475b3b581374966.jpg) +Output 2 + +![](images/de71d79b9a5a341e7ad1909c3bb5915cc710552ad11da465162dff7bb897b6b2.jpg) +Output 3 + +![](images/b1f6eff79e0b92e2dd375eb6d1e4695dd5221a8a1c1de8b36d184a78266158d4.jpg) +View 1 +View 2 + +![](images/76ce41d1ad5c40dfef353ad597af3b9249363f203cdd96812838d74c78398c01.jpg) + +![](images/9b8679bb30949d69e7341a3fcf6c96328e98600ff1eae0f71a07aa65eed6ff82.jpg) + +![](images/3c924143fbd79fdff13d90837596c4043014634b0243cf5d20145281e3eea6c2.jpg) +View 3 + +![](images/0bd289d18c6ebb9f6a2c63c2d2f309661a17c8aeeb751f76321ba5ed15242781.jpg) +View 4 + +# Abstract + +We present GenN2N, a unified NeRF-to-NeRF translation framework for various NeRF translation tasks such as text-driven NeRF editing, colorization, super-resolution, inpainting, etc. Unlike previous methods designed for individual translation tasks with task-specific schemes, GenN2N achieves all these NeRF editing tasks by employing a plug-and-play image-to-image translator to perform editing in the 2D domain and lifting 2D edits into the 3D NeRF space. Since the 3D consistency of 2D edits may not be assured, we propose to model the distribution of the underlying 3D edits through a generative model that can cover all possible edited NeRFs. To model the distribution of 3D edited NeRFs from 2D edited images, we carefully design a VAE-GAN that encodes images while decoding NeRFs. The la + +tent space is trained to align with a Gaussian distribution and the NeRFs are supervised through an adversarial loss on its renderings. To ensure the latent code does not depend on 2D viewpoints but truly reflects the 3D edits, we also regularize the latent code through a contrastive learning scheme. Extensive experiments on various editing tasks show GenN2N, as a universal framework, performs as well or better than task-specific specialists while possessing flexible generative power. More results on our project page: https://xiangyueliu.github.io/GenN2N/. + +# 1. Introduction + +Over the past few years, Neural radiance fields (NeRFs) [23] have brought a promising paradigm in the realm of 3D reconstruction, 3D generation, and novel view synthesis due to their unparalleled compactness, + +high quality, and versatility. Extensive research efforts have been devoted to creating NeRF scenes from 2D images [4, 20, 22, 36, 38] or just text [13, 26] input. However, once the NeRF scenes have been created, these methods often lack further control over the generated geometry and appearance. NeRF editing has therefore become a notable research focus recently. + +Existing NeRF editing schemes are usually task-specific. For example, researchers have developed NeRF-SR [34], NeRF-In [19], PaletteNeRF [16] for NeRF super-resolution, inpainting, and color-editing respectively. These designs require a significant amount of domain knowledge for each specific task. On the other hand, in the field of 2D image editing, a growing trend is to develop universal image-to-image translation methods to support versatile image editing [25, 29, 43]. By leveraging foundational 2D generative models, e.g., stable diffusion [28], these methods achieve impressive editing results without task-specific customization or tuning. We then ask the question: can we conduct universal NeRF editing leveraging foundational 2D generative models as well? + +The first challenge is the representation gap between NeRFs and 2D images. It is not intuitive how to leverage image editing tools to edit NeRFs. A recent text-driven NeRF editing method [9] has shed some light on this. The method adopts a "render-edit-aggregate" pipeline. Specifically, it gradually updates a NeRF scene by iteratively rendering multi-view images, conducting text-driven visual editing on these images, and finally aggregating the edits in the NeRF scene. It seems that replacing the image editing tool with a universal image-to-image translator could lead to a universal NeRF editing method. However, the second challenge would then come. Image-to-image translators usually generate diverse and inconsistent edits for different views, e.g. turning a man into an elf might or might not put a hat on his head, making edits aggregation intricate. Regarding this challenge, Instruct-NeRF2NeRF [9] presents a complex optimization technique to pursue unblurred NeRF with inconsistent multi-view edits. Due to its complexity, the optimization cannot ensure the robustness of the outcomes. Additionally, the unique optimization outcome fails to reflect the stochastic nature of NeRF editing. Users typically anticipate a variety of edited NeRFs just like the diverse edited images. + +To tackle the challenges above, we propose GenN2N, a unified NeRF-to-NeRF translation framework for various NeRF editing tasks such as text-driven editing, colorization, super-resolution, inpainting (see Fig. 1). In contrast to Instruct-NeRF2NeRF which adopts a "render-edit-aggregate" pipeline, we first render a NeRF scene into multi-view images, then exploit an image-to-image translator to edit different views, and finally learn a generative model to depict the distribution of NeRF edits. Instead of + +aggregating all the image edits to form a single NeRF edit, our key idea is to embrace the stochastic nature of content editing by modeling the distribution of the edits in the 3D NeRF space. + +Specifically given a NeRF model or its multi-view images, along with the editing goal, we first generate edited multi-view images using a plug-and-play image-to-image translator. Each view corresponds to a unique 3D edit with some geometry or appearance variations. Conditioned on the input NeRF, GenN2N trains a conditional 3D generative model to reflect such content variations. At the core of GenN2N, we design a 3D VAE-GAN that incorporates a differentiable volume renderer to connect 3D content creation with 2D GAN losses, ensuring that the inconsistent multi-view renderings can still help each other regarding 3D generation. Moreover, we introduce a contrastive learning loss to ensure that the 3D content variation can be successfully understood just from edited 2D images without being influenced by the camera viewpoints. During inference, users can simply sample from the conditional generative model to obtain various 3D editing results aligned with the editing goal. We have conducted experiments on human, items, indoor and outdoor scenes for various editing tasks such as text-driven editing, colorization, superresolution and inpainting, demonstrating the effectiveness of GenN2N in supporting diverse NeRF editing tasks while keeping the multi-view consistency of the edited NeRF. + +We summarize the contribution of this paper as follows, + +- A generative NeRF-to-NeRF translation formulation for the universal NeRF editing task together with a generic solution; +- a 3D VAE-GAN framework that can learn the distribution of all possible 3D NeRF edits corresponding to the a set of input edited 2D images; +- a contrastive learning framework that can disentangle the 3D edits and 2D camera views from edited images; +- extensive experiments demonstrating the superior efficiency, quality, and diversity of the NeRF-to-NeRF translation results. + +# 2. Related Work + +NeRF Editing. Previous works such as EditNeRF [21] propose a conditional neural field that enables shape and appearance editing in the latent space. PaletteNeRF [16, 37] focuses on controlling color palette weights to manipulate appearance. Other approaches utilize bounding boxes [41], meshes [39], point clouds [6], key points [45], or feature volumes [18] to directly manipulate the spatial representation of NeRF. However, these methods either heavily rely on user interactions or have limitations in terms of spatial deformation and color transfer capabilities. + +NeRF Stylization. Images-referenced stylization [7, 12, 42] often prioritize capturing texture style rather than de + +![](images/d47cf812fb1b9fdcafce9ba7a0f9bd66826f1f7cd6321f18674386b1d0abcafb.jpg) +Figure 2. Overview of GenN2N. We first edit the source image set $\{\mathbf{I}_i\}_{i=0}^{N-1}$ using 2D image-to-image translation methods, e.g., text-driven editing, colorization, zoom out, etc. For each view $i \in [0, N-1]$ , we generate $M$ edited images, resulting in a group of translated image set $\{\{\mathbf{S}_i^j\}_{j=0}^{M-1}\}_{i=0}^{N-1}$ . Then we use the Latent Distill Module to learn $M \times N$ edit code vectors from the translated image set, which serve as the input of the translated NeRF. To optimize our GenN2N, we design four loss functions: a KL loss to constrain the latent vectors to a Gaussian distribution; and $\mathcal{L}_{\mathrm{recon}}$ , $\mathcal{L}_{\mathrm{adv}}$ and $\mathcal{L}_{\mathrm{contr}}$ to optimize the appearance and geometry of the translated NeRF. At inference, we can sample a latent vector $\mathbf{z}$ from Gaussian distribution and render a corresponding multi-view consistent 3D scene with high quality. + +tailed content, resulting in imprecise editing appearance of NeRF only. Text-guided works [33, 35], on the other hand, apply contrastive losses based on CLIP [27] to achieve the desired edits. While text references usually describe the global characteristics of the edited results, instructions offer a more convenient and precise expression. + +Instruct-driven NeRF editing. Among numerous image-to-image translation works, InstructPix2Pix [2] stands out by efficiently editing images following instructions. It leverages large pre-trained models in the language and image domains [3, 28] to generate paired data (before and after editing) for training. While editing NeRF solely based on edited images is problematic due to multi-view inconsistency. To address this, an intuitive yet heavy approach [9] is to iteratively edit the image and optimize NeRF. In addition, NeRF-Art [35] and DreamEditor [46] adopt a CLIP-based contrastive loss [27] and score distillation sampling [26] separately to supervise the optimization of editing NeRF. Inspired by Generative Radiance Fields [5, 30], We capture various possible NeRF editing in the generative space to solve it. + +# 3. Method + +Given a NeRF scene, we present a unified framework GenN2N to achieve various editing on the 3D scene leveraging geometry and texture priors from 2D image editing methods, such as text-driven editing, colorization, superresolution, inpainting, etc. While a universal image-to-image translator can theoretically accomplish these 2D editing tasks, we actually use a state-of-the-art translator for each task. Therefore, we formulate each 2D image editing method as a plug-and-play image-to-image translator + +and all NeRF editing tasks as our universal NeRF-to-NeRF translation, in which the given NeRF is translated into NeRF scenes with high rendering quality and 3D geometry consistency according to the user-selected editing target. The overview of GenN2N is illustrated in Fig. 2, we first perform image-to-image translation in the 2D domain and then lift 2D edits to 3D and achieve NeRF-to-NeRF translation. + +Given $N$ multi-view images $\{\mathbf{I}_i\}_{i = 0}^{N - 1}$ of a scene, we first use Nerfstudio [32] to train the original NeRF. Then we use a plug-and-play image-to-image translator to edit these source images. However, the content generated by the 2D translator may be inconsistent among multi-view images. For example, using different initial noise, the 2D translator [1] may generate different content for image editing, which makes it difficult to ensure the 3D consistency between different view directions in the 3D scene. + +To ensure the 3D consistency and rendering quality, we propose to model the distribution of the underlying 3D edits through a generative model that can cover all possible edited NeRFs, by learning an edit code for each edited image so that the generated content can be controlled by this edit code during the NeRF-to-NeRF translation process. + +For each view $\in [0,N - 1]$ , we generate $M$ edited images, resulting in a group of the translated image set $\{\{\mathbf{S}_i^j\}_{j = 0}^{M - 1}\}_{i = 0}^{N - 1}$ . Then we design a Latent Distill Module described in Sec. 3.1 to map each translated image $\mathbf{S}_i^j$ into an edit code vector $\mathbf{z}_i^j$ and design a KL loss $\mathcal{L}_{\mathrm{KL}}$ to constrain those edit code vectors to a Gaussian distribution. Conditioned on the edit code $\mathbf{z}_i^j$ , we perform NeRF-to-NeRF translation in Sec. 3.2 by rendering multi-view images $\{\mathbf{C}_i\}_{i = 0}^{N - 1}$ and optimize the translated NeRF by three loss functions: the reconstruction loss $\mathcal{L}_{\mathrm{recon}}$ , the adversarial + +![](images/360c66cdad8c1dac2d8d0bec7f5be42a74c2b342e866fcd367df72d5cec8fdaf.jpg) +Figure 3. Illustration of our proposed contrastive loss functions. Regarding the multi-view rendered images $\mathbf{C}_i^j$ and $\mathbf{C}_l^j$ sharing the same edit code, we resend them to our Latent Distill Module to extract $\mathbf{z}_i^j$ and $\mathbf{z}_l^j$ , and aggregate them via $\mathcal{L}_{\mathrm{contr}}^{\mathrm{att}}$ . In addition, for $\mathbf{S}_i^k$ whose editing style varies from $\mathbf{S}_i^j$ , $\mathcal{L}_{\mathrm{contr}}^{\mathrm{rep}}$ increases the distance between edit codes of them. + +![](images/a9f17cc1a480904a0523620ddc0d9d809312deb467bf9480bd9e57c4f35e42ae.jpg) +Figure 4. Illustration of our proposed conditional adversarial loss functions. Our conditional discriminator distinguishes artifacts such as blur and distortion in novel-view rendered image $\mathbf{C}_l^j$ compared with target image $\mathbf{S}_l^j$ . $\mathbf{S}_l^j$ and $\mathbf{S}_l^k$ are edited with same view but various styles, the latter serves as the condition to concatenate with $\mathbf{C}_l^j$ and $\mathbf{S}_l^j$ and manufacture fake and real pairs. + +loss $\mathcal{L}_{\mathrm{AD}}$ , and the contrastive loss $\mathcal{L}_{\mathrm{contr}}$ . After the optimization of the translated NeRF, as described in Sec. 3.3, we can sample an edit code $\mathbf{z}$ from Gaussian distribution and render the corresponding edited 3D scene with high quality and multi-view consistency in the inference stage. + +# 3.1. Latent Distill Module + +Image Translation. As illustrated in Fig. 2, GenN2N is a unified framework for NeRF-to-NeRF translation, in which the core is to perform a 2D image-to-image translation and lift 2D edits into universal 3D NeRF-to-NeRF translation. Given the source multi-view image set $\{\mathbf{I}_i\}_{i=0}^{N-1}$ of a NeRF scene, we first perform image editing M times for each view using a plug-and-play 2D image-to-image translator, producing a group of translated image set $\{\{\mathbf{S}_i^j\}_{j=0}^{M-1}\}_{i=0}^{N-1}$ . In this paper, we use several 2D translation tasks to show + +the adaptability of our GenN2N: text-driven editing, superresolution, colorization and inpainting. For more details about those 2D image editing methods, please refer to the supplementary materials. + +Edit Code. Since 2D image-to-image translation may generate different content even with the same editing target, causing the inconsistency problem in the 3D scene. We propose to map each edited image $\mathbf{S}_i^j$ into a latent feature vector named edit code to characterize these diverse editorings. We employ the off-the-shelf VAE encoder from stable diffusion [28] to extract the feature from $\mathbf{S}_i^j$ and then apply a tiny MLP network to produce this edit code $\mathbf{z}_i^j \in \mathbb{R}^{64}$ . During the training process, we keep the pre-trained encoder fixed and only optimize the parameters of the tiny MLP network. This mapping process can be formulated as follows: + +$$ +\mathbf {z} _ {i} ^ {j} = \mathcal {D} \left(\mathbf {S} _ {i} ^ {j}\right) = \mathcal {M} \left(\mathcal {E} \left(\mathbf {S} _ {i} ^ {j}\right)\right), \tag {1} +$$ + +where $\mathcal{D}$ represent this mapping process, $\mathcal{E}$ is the fixed encoder, and $\mathcal{M}$ is the learnable tiny MLP. + +KL loss. In order to facilitate effective sampling of the edit code so as to control the editing diversity of our NeRF-to-NeRF translation, we need to constrain the edit code to a well-defined distribution. Thus we design a KL loss to encourage $\mathbf{z}_i^j$ to approximate a Gaussian distribution: + +$$ +\mathcal {L} _ {\mathrm {K L}} = \mathbb {E} _ {\mathbf {S} \in \{\left\{\mathbf {S} _ {i} ^ {j} \right\} _ {j = 0} ^ {M - 1} \} _ {i = 0} ^ {N - 1}} [ P (\mathbf {z} _ {\text {n o r m a l}}) \log \left(\frac {P \left(\mathbf {z} _ {\text {n o r m a l}}\right)}{P \left(\mathcal {D} (\mathbf {S})\right)}\right) ], \tag {2} +$$ + +where $P(\mathbf{z}_{\text{normal}})$ denotes probability distribution of the standard Gaussian distribution in $\mathbb{R}^{64}$ and $P(\mathcal{D}(\mathbf{S}))$ means probability distribution of the extracted edit codes. + +Contrastive loss. It is not assured that edit codes $\mathbf{z}$ obtained from the Latent Distill Module contain only the editing information while excluding viewpoint-related effects. However, since the translated NeRF utilizes $\mathbf{z}$ to edit scenes, it yields instability if $\mathbf{z}$ violently changes given images that are similar in appearance but different in viewpoints. To ensure the latent code does not depend on 2D viewpoints but truly reflects the 3D edits, we regularize the latent code through a contrastive learning scheme. Specifically, we reduce the distance between edit codes of different-view rendered images from a translated NeRF that share the same edit code, while increasing the distance between same-view images that are multi-time edited by the 2D image-to-image translator. As illustrated in Fig. 3, given an edit code $\mathbf{z}_i^j$ extracted from the $i$ -th input view at the $j$ -th edited image $\mathbf{S}_i^j$ , we render multi-view images $\{\mathbf{C}_l^j\}_{l=0}^{N-1}$ using the translated NeRF conditioned on $\mathbf{z}_i^j$ . Then we employ contrastive learning to encourage the edit code $\mathbf{z}_i^j$ to be close to $\{\mathbf{z}_l^j\}_{l=0}^{N-1}$ extracted from $\{\mathbf{C}_l^j\}_{l=0}^{N-1}$ , while being distinct from the edit codes $\{\mathbf{z}_i^k\}_{k=0}^{M-1}$ extracted from $\{\mathbf{S}_i^k\}_{k=0}^{M-1}$ , where $k \neq j$ . + +![](images/2a59429c383d15a77cd2d8111c9489f90c9975e3773858f91cdf8e17b9ca6414.jpg) + +![](images/a85184f26f80b850d50b0ef12f278ac2b2ec723cd1f0ed303db3fe012b93d0d2.jpg) + +![](images/98732b7a251c45d6ae2e6607ba03db46cfbe6812cfe8939993d4d23aee91b90e.jpg) + +![](images/f5697f4c13ae0b6187525bea0e5afce085e2f6b25e696ed76b1ff33e0b9cabb1.jpg) +"Make it an Egyptian sculpture" + +![](images/5551ee4ed2848b775519e0d428a75b8c161bb2658b4b57c22de51584781157e3.jpg) + +![](images/0c076c18fb344ee4ed1f89a07b0027a2a2813dfff2fe39aec549fa29ecb7bddd.jpg) + +![](images/bdf6fd6dceefa3cdcfbb5cb2e621dffc633583c58826e844f8c3836e76f977bb.jpg) + +![](images/27d3cb7179f57174f11a26b833c9c166e3bb171a9b7440c44ba41a254b2148c0.jpg) + +![](images/68c72715214e058cb0f2e3267afddc95bf1a646ba91ff7da689511af97d350a7.jpg) + +![](images/e25597512fa654b4a23cb6294a3ae57b7139848cb17c43e70f30c65fb8e2768c.jpg) +Original NeRF + +![](images/11939d6053dfc042d19524d56e7c064b6832670414dca9b13dbba17776f35179.jpg) +Inference 1 + +![](images/0308bd078b45e6b9fa17bb4b800cddd9452bd3877c20c917301b8d073103d902.jpg) +Figure 5. Text-Driven Editing. We sample 4 inference results for both text-driven editing tasks. The diversity of geometry and appearance showcases awesome generative ability of GenN2N, on the premise of maintaining the 3D consistency between different viewpoints. + +![](images/c6dafb7f106bd694903804ccc5d18a5615d7996d3a7b6f440088be450407316b.jpg) +Inference 2 + +![](images/fa33bc58d5b4297395423c609e762e0a71d75f3817b33c4935b04ddb246b1527.jpg) + +![](images/8842118892dba43b2f3e21b506c5b8f6939f5d54979913852e941058b0b11156.jpg) +Inference 3 + +![](images/0d807275c45433d989082d0d688d823214358915658926e02c9694f2e47ab562.jpg) + +![](images/8afa615510a22bd70d4b97463a0f0c94960329822ac70b51e206cc9cab6f85e4.jpg) +Inference 4 + +![](images/693f9a1bbddab8ae2ec0745817b0d3445cc5bbae8e5299b157be7e5693b66b72.jpg) + +![](images/95d490a396cb3401e2eedb9da334e77f98f2c15cde96bd6215e5343719112fb7.jpg) +Original NeRF + +![](images/0d31b9a9f2046581c21fcd7b55487da24655e421dd18554eaa002695c3724977.jpg) + +![](images/12de9452394cc1cdc540daee145d15972b4db943377d7993e27a777c2c3402b3.jpg) +Inference 1 + +![](images/7cea362638cc036d7a0c716a306bef32dd098ac75ece1fbcd79de2a5d2f5a33f.jpg) + +![](images/925b7724b83d034c415d74f9012ec24b1ed4f5a702b33d18c945a6aa46c860f7.jpg) +Inference 2 + +![](images/b99df597dd4df24c43683a6ea3beeed898412fba61220d04390636665d35fda9.jpg) + +![](images/8bb7696f43673e9f7661b9a5c4f3659c75fa815ee89d7c604d24d93b7e46b038.jpg) +Inference 3 +Figure 6. Colorization. Our method colorizes the gray-scale 3D scene consistently across views. By changing the edit code during inference, diverse colorized scenes can be rendered with satisfying photorealism and reasonably rich colors. + +![](images/33f3a99405255176ebb939334391fd46cf6f69ce46e18f611595c6c978a2448a.jpg) + +![](images/6d1d848ca2fab9504e9191ad8fe3827df1a0e985deb1b96a43c7228875fe0f52.jpg) +Inference 4 + +![](images/03192cd3579b71dc28ba8cd7d112187da2cdf2576eba6c5ef41aef4ecfd09b2e.jpg) + +![](images/91f4e08eca197464bc3c0f30b58ad3525ad06b5bd92f2e83ec33615b442b4cd7.jpg) +Inference 5 + +Specifically, our contrastive loss is designed as follows: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {c o n t r}} = \mathcal {L} _ {\text {c o n t r}} ^ {\text {a t t}} + \mathcal {L} _ {\text {c o n t r}} ^ {\text {r e p}} \\ = \sum_ {l = 0} ^ {N - 1} \left\| \mathbf {z} _ {i} ^ {j} - \mathbf {z} _ {l} ^ {j} \right\| _ {2} ^ {2} + \sum_ {k = 0} ^ {M - 1} \max (0, \alpha - \left\| \mathbf {z} _ {i} ^ {j} - \mathbf {z} _ {i} ^ {k} \right\| _ {2} ^ {2}), \tag {3} \\ \end{array} +$$ + +where $\alpha$ represents the margin that encourages the difference in features, and $k\neq j$ + +# 3.2. NeRF-to-NeRF translation + +Translated NeRF. After 2D image-to-image translation, we need to lift these 2D edits to the 3D NeRF. For this purpose, we propose to modify the original NeRF as a translated NeRF that takes the edit code $\mathbf{z}$ as input and generates the translated 3D scene according to the edit code. We refer readers to the supplementary for more details about the network architecture. + +Reconstruction loss. Given an edit code $\mathbf{z}_i^j$ extracted from + +the edited image $\mathbf{S}_i^j$ , we can generate a translated NeRF to render $\mathbf{C}_i^j$ from the same viewpoint. Then we define the reconstruction loss as the L1 normalization and Learned Perceptual Image Patch Similarity (LPIPS) [44] between the rendered image $\mathbf{C}_i^j$ and the edited image $\mathbf{S}_i^j$ as follows: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {r e c o n}} = \mathcal {L} _ {\mathrm {L} 1} + \mathcal {L} _ {\mathrm {L P I P S}} \\ = \left\| \mathbf {C} _ {i} ^ {j} - \mathbf {S} _ {i} ^ {j} \right\| _ {1} + L P I P S [ \mathcal {P} (\mathbf {C} _ {i} ^ {j}) - \mathcal {P} (\mathbf {S} _ {i} ^ {j}) ], \tag {4} \\ \end{array} +$$ + +where $\mathcal{P}$ means a patch sampled from the image. Note that due to the lack of 3D consistency of the edited multiview image, the supervision of the edited image from other viewpoints $\{\mathbf{S}_l^j\}_{l\neq i}$ will lead to conflicts in pixel-space optimization. Therefore, we only employ reconstruction loss on the same view image $\mathbf{S}_i^j$ to optimize the translated NeRF. Adversarial loss. Since the 3D consistency of edited multiview images is not assured, relying solely on the reconstruction loss on the same view often leads to blurry or distorted + +![](images/1903030e864e1d3f55178a367dcc206654f9470b734b5ec7c6524603a9755c09.jpg) +Original NeRF + +![](images/b31dda698362d8b9ae293fdc2235c9086df7d7269612578a6c8b0de3388462fa.jpg) +Instruct-NeRF2NeRF 1 + +![](images/a5a0c06a3876536af5a979b53b435f76c3920d4cc176f4cd355c3d25c723cc22.jpg) +Instruct-NeRF2NeRF 2 + +![](images/df92d6df55aa709bdda7be0e168885a51760fbd267ca4524d1425ae1c856f04d.jpg) +Ours inference 1 +Figure 7. Comparisons with baselines of text-driven NeRF editing. We compare our method with Instruct-NeRF2NeRF [9] in the editing by using the text prompt "Make it Spring". + +![](images/86bfc0b53e1dff3aeab3555a77b800c0c79363fc58056e5633805bd7fdb56bf1.jpg) +Ours inference 2 + +![](images/d9b07c9875a004e5c9656be5ba738e6b5355b277fa061b07a93f895e9c67dcec.jpg) +Ours inference 3 + +
MethodCLIP Text-Image Direction Similarity↑CLIP Direction Consistency ↑FID ↓
InstructPix2Pix [2]+NeRF0.16690.8475270.542
Instruct-NeRF2NeRF0.20210.9828148.021
Ours w/o Ladv0.19200.9657162.275
Ours w/o Lcontr0.20070.9749156.524
Ours0.20890.9864137.740
+ +artifacts on novel views. Previous research demonstrates the effectiveness of conditional adversarial training in preventing the production of blurry rendered images resulting from conflicts that arise from noise in the camera extrinsic when performing image supervision from different viewpoints [11]. The function of the condition is to guide discriminator with fine-grained information from the same viewpoint, thus preventing GAN mode collapse. + +It inspires us to incorporate conditional adversarial loss on rendered images from the translated NeRF, which is conducive to distinguish artifacts in rendered images. As illustrated in Fig.4, the discriminator $\mathbf{D}$ takes into real pairs and fake pairs. Each real pair $\mathbf{R}$ consists of $\mathbf{S}^j$ and $\mathbf{S}^j -\mathbf{S}^k$ where $\mathbf{S}^j\in \{\mathbf{S}_i^j\}_{i = 0}^{N - 1}$ and $\mathbf{S}^k\in \{\mathbf{S}_i^k\}_{i = 0}^{N - 1}$ are from two sets of edited images from the image translation. Similarly, each fake pair $\mathbf{F}$ consists of $\mathbf{C}^j$ and $\mathbf{C}^j -\mathbf{S}^k$ in which $\mathbf{C}^j\in \{\mathbf{C}_i^j\}_{i = 0}^{N - 1}$ is generated by translated NeRF. Note that the images in the same pair come from the same viewpoint. The pairs are concatenated in RGB channels and fed into the discriminator. We optimize the discriminator $\mathbf{D}$ and translated NeRF with the objective functions below: + +$$ +\mathcal {L} _ {\mathrm {A D - D}} = \mathbb {E} _ {\mathbf {R}} [ - l o g (\mathbf {D} (\mathbf {R})) ] + \mathbb {E} _ {\mathbf {F}} [ - l o g (1 - \mathbf {D} (\mathbf {F})) ], \tag {5} +$$ + +$$ +\mathcal {L} _ {\mathrm {A D - G}} = \mathbb {E} _ {\mathbf {F}} [ - l o g (\mathbf {D} (\mathbf {F})) ]. +$$ + +Optimization. During the training process, we jointly optimize the loss functions mentioned above: $\mathcal{L}_{\mathrm{KL}}$ and $\mathcal{L}_{\mathrm{contr}}$ + +![](images/8b3abbdfb1cc401f78083ee3e9d504f25ca9dfd926e18de2204502d389ab78ca.jpg) +Original NeRF + +![](images/84cc93b7444cb4f0729abcbc0ec4d39f361680c551f7427720cf6c1197de227b.jpg) +Instruct-NeRF2NeRF + +![](images/dda52412dbaf0acaba32f1cfbfb5f51e7974b02aa213a2afd0feb29121574ff7.jpg) +PaletteNeRF + +![](images/159f0eda688041349c873d7f76f4316f35000945f3405e18f79ef75d96b19797.jpg) +Ours inference 1 +Figure 8. Comparisons with baselines of NeRF colorization. We compare with PaletteNeRF[16] in colorization. + +![](images/ccc9308bb2cf789858b3dc4ffd8801748154ebf4b74d9d5d6879070c1c67ca90.jpg) +Ours inference 2 + +![](images/e2e1951d4fae3b38aff5be22e2642dcb7694a7fc7c3c92b557bd8a70d598b963.jpg) +Ours inference 3 + +Table 1. Quantitative results on text-driven editing. We compare our method with the naive method of directly combining InstructPix2Pix [2] with NeRF and the state-of-the-art method Instruct-NeRF2NeRF [9]. + +
MethodCF ↑FID ↓
DDColor [14]+NeRF40.435148.957
Instruct-NeRF2NeRF45.599201.456
PaletteNeRF [16]39.654-
Ours w/o Ladv35.031137.740
Ours w/o Lcontr34.829105.750
Ours65.09935.041
+ +Table 2. Quantitative results on colorization. We colorize images with the translator, DDcolor [14], and edit NeRF by directly optimizing NeRF, Instruct-NeRF2NeRF [9] and our NeRF translation method. The quantitative comparison is conducted between these methods as well as PaletteNeRF [16]. + +for the edit code, $\mathcal{L}_{\mathrm{recon}}$ and $\mathcal{L}_{\mathrm{AD - G}}$ for the translated NeRF, and $\mathcal{L}_{\mathrm{AD - D}}$ for the discriminator. The total loss formula is expressed as follows: + +$$ +\mathcal {L} = \mathcal {L} _ {\mathrm {K L}} + \mathcal {L} _ {\text {r e c o n}} + \mathcal {L} _ {\mathrm {A D - G}} + \mathcal {L} _ {\mathrm {A D - D}} + \mathcal {L} _ {\text {c o n t r}}. \tag {6} +$$ + +where we assign each regularization term the weight of 1.0, 1.0, 0.1, 0.1, 0.1 in all of our experiments. The weights can be adjusted to prioritize different aspects of the training objective, such as reconstruction accuracy, adversarial training, and perceptual quality. + +# 3.3. Inference + +After the optimization of our GenN2N, the translated NeRF is optimized to be able to render the target scene conditioned on the edit code. As shown in Fig. 1, users can simply sample an edit code from the Gaussian distribution and use the translated NeRF to render the 3D scene with high-quality and multi-view 3D consistency. + +# 4. Experiments + +Our proposed GenN2N is a unified NeRF-to-NeRF translation framework which can support various NeRF editing tasks. In this paper, we demonstrate the effectiveness of GenN2N by a suite of challenging tasks: + +![](images/ea7a30b00b68fa9411a593ace664d3944b1ec609ad2b74c06b46571d549da071.jpg) + +![](images/ea07cdbf5005153b3fefbaf6cbbc4ab178f07a04305b7d8b8fc23560a8262968.jpg) +Original NeRF + +![](images/fc446adaa1a09002755507642964878b5a058cab2d5b9896f16222960353f499.jpg) +Instruct-NeRF2NeRF + +![](images/8c734d8649c9850ae52c36c4655967ffdb6759f5ec469f76b577323ab321ed99.jpg) + +![](images/f33b2e818b8f36d411da9f6cbd9a7dfd67cf7f89580d02ff866ba9ce7f9adbf5.jpg) +NeRF-SR +Figure 9. Comparisons with baselines of NeRF super-resolution. We compare with NeRF-SR [34] in super-resolution. + +![](images/79f4eb3b5408230337924aa9c63495fd66cdd5254acd6196aaee3ad6c651e027.jpg) + +![](images/5cbb8ff671d1809fabcc856eb267883b8cc5bceb9655e76da1c52bbc7511d1a7.jpg) +Ours + +![](images/14fa18a5859480ae72322936d6e46c5f13087490b5647c2789da03ad0b9ea568.jpg) + +
MethodPSNR ↑SSIM ↑LPIPS ↓
ResShift [40]+NeRF19.9780.5350.1156
Instruct-NeRF2NeRF20.2990.6420.2732
NeRF-SR [34]27.9570.8970.0937
Ours w/o Ladv12.5550.6630.2001
Ours w/o Lcontr15.3720.6620.1834
Ours28.5010.9130.0748
+ +Table 3. Quantitative results on super-resolution. We improve image resolution with ResShift [40] and edit NeRF by directly optimizing NeRF, Instruct-NeRF2NeRF [9] and our NeRF translation method. The quantitative comparison is conducted between these methods as well as NeRF-SR [34]. + +(1) Text-driven Editing edits the given NeRF scene to a set of NeRF scenes according to the text instruction. +(2) Colorization transforms a gray-scale NeRF scene to a set of plausible color NeRF scenes. +(3) Super-resolution enhances the resolution of NeRF and enables multiple plausible outcomes. +(4) Inpainting fills in user-specified masked regions in the NeRF scene with realistic content. + +We achieve those tasks by simply changing the plug-and-play 2D image translator in our framework, without any additional task-specific design. Previous studies have extensively explored some of these issues like text-driven editing, colorization, super-resolution, and inpainting. However, there is rarely a unified framework that can achieve all these tasks with strong performance, high quality, and plausible multi-view consistent 3D structure. Furthermore, GenN2N can also perform zooming out and text-driven inpainting in NeRF-to-NeRF translation, which were not explored in prior research. We refer readers to the supplementary materials for detailed experiment settings, dataset settings and implementation details. + +# 4.1. Comparisons + +Text-driven Editing. We achieve text-driven editing of the given NeRF by using InstructPix2Pix [2] as the 2D image-to-image translator in our framework. We compared our approach to a naive solution, which involves optimizing a NeRF with edited images via InstructPix2Pix. However, + +this naive approach leads to a 3D inconsistency problem among different edits. While Instruct-NeRF2NeRF [9] proposed an iterative updating mechanism to address this issue, it falls short in capturing the diversity of different edits, making it challenging to ensure the quality of the outcomes. To evaluate our method, we conducted experiments on the Face [9] and Fangzhou [35] self-portrait datasets, and Farm [9] dataset, comparing GenN2N with the state-of-the-art NeRF editing method Instruct-NeRF2NeRF [9]. + +Quantitative results are presented in Table 1, where we employed CLIP Text-Image Direction Similarity [9], CLIP Direction Consistency [9], and Fréchet Inception Distance (FID) [10] as evaluation metrics. The results highlight the superior performance of GenN2N over other methods, demonstrating its effectiveness in producing high-quality 3D text-driven editing results. Furthermore, we provide a qualitative comparison between GenN2N and Instruct-NeRF2NeRF [9] in Figure 7. Notably, Instruct-NeRF2NeRF was trained twice, but the results are nearly identical. In contrast, GenN2N can render edited scenes with various effects, consistent with the input text instruction, by inferring with different edit codes sampled from a Gaussian distribution. To further illustrate the generative capability of GenN2N, we showcase additional results of our inference renderings in Figure 5, demonstrating its diverse generative ability in terms of appearance and geometry. + +Colorization. For NeRF colorization, GenN2N uses DD-Color [14] as the 2D image-to-image translator. CoRF [8] and Palette-NeRF [16] do a similar task and we compare with Palette-NeRF in Table 2. The Fréchet Inception Distance (FID) [10] and colorfulness score (CF) [17] are used to measure the distribution similarity and vividness of generated images. We show visual comparison results in Fig. 6 and Fig. 8. We can find that with different edit codes, the scene can be rendered in different color styles. It is noticeable that with the same edit code, the color rendered from different views is consistent. This strongly demonstrates the effectiveness of our method in translating NeRF while keeping the 3D consistency of the scene. + +Super-resolution. When only low-resolution images are available, our methods can boost NeRF in reconstructing scenes at higher resolution, while keeping view consistency and avoiding blurry outputs. We achieve this by employing ResShift [40] as the 2D image-to-image translator in GenN2N. Following state-of-the-art method NeRF-SR [34], we conduct experiments on $LLFF$ dataset [23], using PSNR, SSIM, and LPIPS as evaluation metrics. As shown in Table 3, GenN2N obtains NeRF-to-NeRF translation with higher performance than NeRF-SR [34]. Moreover, we also provide qualitative comparison results in Fig. 9, where GenN2N produces clearer and more realistic rendering results than previous methods. + +![](images/2d395f82bc64109b08b3631b093c0feaadb2158aa8a700af2aa2a6ccfdbf0e1c.jpg) + +![](images/0e9e79b1be9ac459c1c1ae539d5a4d29ca56248b322a6e3a99dc13ae8cae17be.jpg) + +![](images/03915de0884441efb6c533790dc10dd5779d3f3d7b1f48e254f6fc42dad1d4d1.jpg) +Instruct-NeRF2NeRF + +![](images/78479ea566ffe1847a74ba82b42f0b4ba8494975fec2cf0053a6a588f567fb1b.jpg) + +![](images/9dff90df2b08ff280aea834c0947a49fa62b04f3625e1716db0dbda811e74005.jpg) + +![](images/484cba32e0cd42a5a3d692f9500a6699281318e62d26b4afa219ff24723d5e29.jpg) +Original NeRF +SPin-NeRF +Figure 10. Comparisons with baselines of NeRF inpainting. We compare with SPIn-NeRF [24] in inpainting. + +![](images/04848002808598c140cd62eb2f3e2486767f21657c5c122071169ba1305fbd8b.jpg) +Ours + +![](images/7ed027a0cbd9386e6dfe4cb4901dd88152219e7d5dc64eea4761285f4f21f6c8.jpg) + +
MethodPSNR ↑SSIM ↑LPIPS ↓
LaMa [31]+NeRF18.9830.37060.1730
Instruct-NeRF2NeRF16.7340.30880.2750
SPin-NeRF [24]24.3690.72170.1754
Ours26.8680.81370.1284
+ +Table 4. Quantitative results on NeRF inpainting. + +Inpainting. The goal of NeRF Inpainting is to fill the 3D content of regions specified by users. SPin-NeRF [24] achieves this through a multi-step process: it employs SAM [15] for object segmentation, utilizes LaMa [31] to paint the background content in multi-view images, and subsequently trains the NeRF model with color, depth, and perceptual cues. In our experiments, we use SAM and LaMa as the 2D image-to-image translator in our GenN2N, which is the same setting as SPin-NeRF [24]. Quantitative comparisons on statue dataset [24] are shown in Table 4, where GenN2N achieves superior PSNR and SSIM scores than SPin-NeRF [24], highlighting the effectiveness of our GenN2N framework. In addition, qualitative results are showcased in Fig. 10 revealing that while SPin-NeRF [24] fails to generate reasonable content behind the masked object, our GenN2N produces realistic content in the same region with fine multi-view consistency. + +# 4.2. Ablation Studies + +We conduct comprehensive ablation experiments to validate the designs of each component in GenN2N. Due to space limitations, we only highlight the essential aspects below. Please refer to supplementary for more details. + +The Contrastive Loss. We demonstrate the advantages of incorporating our proposed contrastive loss in Table 1, 2, 3. The motivation is to disentangle the camera view and edit information present in the latent space. We achieve this by reducing the distance between edit codes of different-view rendered images from a translated NeRF that shares the same edit code, while increasing the distance between same-view images that are edited by the 2D translator with diverse edit styles. As demonstrated in Tables, the absence of contrastive loss leads to the generation of blurry areas + +in the rendered images, resulting in a decrease in the metric scores. This blurriness can be attributed to the inclusion of pose information within the edit code z. By incorporating the contrastive loss, our method successfully achieves a uniform appearance with different observing views under the same style latent z. + +Discriminator for Novel Views. We demonstrate the effectiveness of employing a conditional discriminator to address artifacts caused by inconsistent cross-view edited images and to enhance the quality of novel view rendering images, as depicted in Table 1, 2, 3. The removal of this conditional discriminator results in blurry novel view images with artifacts in the background region. We attribute these undesirable effects to the inability of current image-to-image translation methods, such as InstructPix2Pix [2], to produce image editing consistently across multi-view images. To mitigate these issues, we introduce a conditional discriminator between rendered images from the translated NeRF and edited images from the 2D image-to-image translator. This inclusion successfully eliminates artifacts and enhances the image quality of rendered images from the translated NeRF. + +# 4.3. Applications + +We demonstrate the versatility and robustness of GenN2N by exploring two translation applications: Zoom Out and Text-Driven Inpainting. While existing 2D translators [1, 29] can complete these tasks, 3D editing has not been explored. We achieve these tasks by incorporating Blended Latent Diffusion [1] as the 2D image-to-image translator, enabling us to generate diverse and high-quality content with multi-view consistency. Please refer to our supplementary materials for the results of these applications due to space constraints. + +# 5. Conclusions + +We introduce GenN2N, a unified NeRF-to-NeRF translation framework that can handle various NeRF editing tasks. Unlike previous task-specific approaches, our framework uses an image-to-image translator for 2D editing and integrates the results into 3D NeRF space. To address the challenge of ensuring 3D consistency, we propose modeling the distribution of 3D edited NeRFs from 2D edited images using our novel techniques. After optimization, users can sample from the conditional generative model to obtain diverse 3D editing results with high rendering quality and multi-view consistency. Our experiments demonstrate that GenN2N outperforms existing task-specific methods on various editing tasks, including text-driven editing, colorization, super-resolution, and inpainting, in terms of efficiency, quality, and diversity. + +# References + +[1] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM Transactions on Graphics (TOG), 42 (4):1-11, 2023. 3, 8 +[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. arXiv preprint arXiv:2211.09800, 2022. 3, 6, 7, 8 +[3] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 3 +[4] Shengqu Cai, Anton Obukhov, Dengxin Dai, and Luc Van Gool. Pix2nerf: Unsupervised conditional p-gan for single image to neural radiance fields translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3981-3990, 2022. 2 +[5] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 3 +[6] Jun-Kun Chen, Jipeng Lyu, and Yu-Xiong Wang. Neuraleditor: Editing neural radiance fields via manipulating point clouds. arXiv preprint arXiv:2305.03049, 2023. 2 +[7] Pei-Ze Chiang, Meng-Shiun Tsai, Hung-Yu Tseng, Wei-Sheng Lai, and Wei-Chen Chiu. Stylizing 3d scene via implicit representation and hypernetwork. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1475-1484, 2022. 2 +[8] Ankit Dhiman, R Srinath, Srinjay Sarkar, Lokesh R. Boregowda, and R Venkatesh Babu. Corf: Colorizing radiance fields using knowledge distillation. ArXiv, abs/2309.07668, 2023. 7 +[9] Ayaan Haque, Matthew Tancik, Alexei A Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. arXiv preprint arXiv:2303.12789, 2023.2,3,6,7 +[10] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7 +[11] Jingwei Huang, Justus Thies, Angela Dai, Abhijit Kundu, Chiyu Jiang, Leonidas J Guibas, Matthias Nießner, Thomas Funkhouser, et al. Adversarial texture optimization from rgb-d scans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1559-1568, 2020. 6 +[12] Yi-Hua Huang, Yue He, Yu-Jie Yuan, Yu-Kun Lai, and Lin Gao. Stylizednerf: consistent 3d scene stylization as stylized nerf via 2d-3d mutual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18342–18352, 2022. 2 +[13] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object genera + +tion with dream fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 867-876, 2022. 2 +[14] Xiaoyang Kang, Tao Yang, Wenqi Ouyang, Peiran Ren, Lingzhi Li, and Xuansong Xie. Ddcolor: Towards photorealistic and semantic-aware image colorization via dual decoders. arXiv preprint arXiv:2212.11613, 2022. 6, 7 +[15] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 8 +[16] Zhengfei Kuang, Fujun Luan, Sai Bi, Zhixin Shu, Gordon Wetzstein, and Kalyan Sunkavalli. Palettenerf: Palette-based appearance editing of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20691-20700, 2023. 2, 6, 7 +[17] Gustav Larsson, Michael Maire, and Gregory Shakhnarovich. Learning representations for automatic colorization. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 577-593. Springer, 2016. 7 +[18] Verica Lazova, Vladimir Guzov, Kyle Olszewski, Sergey Tulyakov, and Gerard Pons-Moll. Control-nerf: Editable feature volumes for scene rendering and manipulation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4340–4350, 2023. 2 +[19] Hao-Kang Liu, I Shen, Bing-Yu Chen, et al. Nerf-in: Free-form nerf inpainting with rgb-d priors. arXiv preprint arXiv:2206.04901, 2022. 2 +[20] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. arXiv preprint arXiv:2303.11328, 2023. 2 +[21] Steven Liu, Xiuming Zhang, Zhoutong Zhang, Richard Zhang, Jun-Yan Zhu, and Bryan Russell. Editing conditional radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5773-5783, 2021. 2 +[22] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Realfusion: $360^{\circ}$ reconstruction of any object from a single image. arXiv e-prints, pages arXiv-2302, 2023. 2 +[23] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 7 +[24] Ashkan Mirzaei, Tristan Aumentado-Armstrong, Konstantinos G Derpanis, Jonathan Kelly, Marcus A Brubaker, Igor Gilitschenski, and Alex Levinshtein. Spin-nerf: Multiview segmentation and perceptual inpainting with neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20669–20679, 2023. 8 +[25] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image + +translation. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2 +[26] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2, 3 +[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3 +[28] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 3, 4 +[29] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 2, 8 +[30] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. Advances in Neural Information Processing Systems, 33:20154-20166, 2020. 3 +[31] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. arXiv preprint arXiv:2109.07161, 2021.8 +[32] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, et al. Nerfstudio: A modular framework for neural radiance field development. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-12, 2023. 3 +[33] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3835–3844, 2022. 3 +[34] Chen Wang, Xian Wu, Yuan-Chen Guo, Song-Hai Zhang, Yu-Wing Tai, and Shi-Min Hu. Nerf-sr: High quality neural radiance fields using supersampling. In Proceedings of the 30th ACM International Conference on Multimedia, pages 6445-6454, 2022. 2, 7 +[35] Can Wang, Ruixiang Jiang, Mengei Chai, Mingming He, Dongdong Chen, and Jing Liao. Nerf-art: Text-driven neural radiance fields stylization. IEEE Transactions on Visualization and Computer Graphics, 2023. 3, 7 +[36] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3d digital avatars using diffusion. arXiv preprint arXiv:2212.06135, 2022. 2 +[37] Qiling Wu, Jianchao Tan, and Kun Xu. Palettenerf: + +Palette-based color editing for nerfs. arXiv preprint arXiv:2212.12871, 2022. 2 +[38] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021. 2 +[39] Yu-Jie Yuan, Yang-Tian Sun, Yu-Kun Lai, Yuewen Ma, Rongfei Jia, and Lin Gao. Nerf-editing: geometry editing of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18353-18364, 2022. 2 +[40] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Resshift: Efficient diffusion model for image super-resolution by residual shifting. arXiv preprint arXiv:2307.12348, 2023. 7 +[41] Jiakai Zhang, Xinhang Liu, Xinyi Ye, Fuqiang Zhao, Yanshun Zhang, Minye Wu, Yingliang Zhang, Lan Xu, and Jingyi Yu. Editable free-viewpoint video using a layered neural representation. ACM Transactions on Graphics (TOG), 40(4):1-18, 2021. 2 +[42] Kai Zhang, Nick Kolkin, Sai Bi, Fujun Luan, Zexiang Xu, Eli Shechtman, and Noah Snavely. Arf: Artistic radiance fields. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXI, pages 717-733. Springer, 2022. 2 +[43] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543, 2023. 2 +[44] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5 +[45] Chengwei Zheng, Wenbin Lin, and Feng Xu. Editablenerf: Editing topologically varying neural radiance fields by key points. arXiv preprint arXiv:2212.04247, 2022. 2 +[46] Jingyu Zhuang, Chen Wang, Lingjie Liu, Liang Lin, and Guanbin Li. Dreameditor: Text-driven 3d scene editing with neural fields. arXiv preprint arXiv:2306.13455, 2023. 3 \ No newline at end of file diff --git a/2024/GenN2N_ Generative NeRF2NeRF Translation/images.zip b/2024/GenN2N_ Generative NeRF2NeRF Translation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a2fb1d1d0bd54c8ebb9b19243b2ed33917ee13db --- /dev/null +++ b/2024/GenN2N_ Generative NeRF2NeRF Translation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9972bb3f82016b4e3efb1663845d04674fd4749a8e9c7a0515c742f2f1b9da2a +size 868860 diff --git a/2024/GenN2N_ Generative NeRF2NeRF Translation/layout.json b/2024/GenN2N_ Generative NeRF2NeRF Translation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7746729b41f2f05e8596d77e51c12c3352cfdb39 --- /dev/null +++ b/2024/GenN2N_ Generative NeRF2NeRF Translation/layout.json @@ -0,0 +1,12733 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 152, + 103, + 442, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 103, + 442, + 119 + ], + "spans": [ + { + "bbox": [ + 152, + 103, + 442, + 119 + ], + "type": "text", + "content": "GenN2N: Generative NeRF2NeRF Translation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 465, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 465, + 159 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 465, + 159 + ], + "type": "text", + "content": "Xiangyue Liu" + }, + { + "bbox": [ + 130, + 141, + 465, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 130, + 141, + 465, + 159 + ], + "type": "text", + "content": " Han Xue" + }, + { + "bbox": [ + 130, + 141, + 465, + 159 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 130, + 141, + 465, + 159 + ], + "type": "text", + "content": " Kunming Luo" + }, + { + "bbox": [ + 130, + 141, + 465, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 130, + 141, + 465, + 159 + ], + "type": "text", + "content": " Ping Tan" + }, + { + "bbox": [ + 130, + 141, + 465, + 159 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 130, + 141, + 465, + 159 + ], + "type": "text", + "content": " Li Yi" + }, + { + "bbox": [ + 130, + 141, + 465, + 159 + ], + "type": "inline_equation", + "content": "^{2,3,4\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 112, + 170, + 481, + 200 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 112, + 170, + 481, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 170, + 481, + 185 + ], + "spans": [ + { + "bbox": [ + 112, + 170, + 481, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 112, + 170, + 481, + 185 + ], + "type": "text", + "content": " Hong Kong University of Science and Technology " + }, + { + "bbox": [ + 112, + 170, + 481, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 112, + 170, + 481, + 185 + ], + "type": "text", + "content": " Tsinghua University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 185, + 474, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 185, + 474, + 200 + ], + "spans": [ + { + "bbox": [ + 115, + 185, + 474, + 200 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 115, + 185, + 474, + 200 + ], + "type": "text", + "content": " Shanghai Artificial Intelligence Laboratory " + }, + { + "bbox": [ + 115, + 185, + 474, + 200 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 115, + 185, + 474, + 200 + ], + "type": "text", + "content": " Shanghai Qi Zhi Institute" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 85, + 227, + 293, + 349 + ], + "blocks": [ + { + "bbox": [ + 85, + 227, + 293, + 349 + ], + "lines": [ + { + "bbox": [ + 85, + 227, + 293, + 349 + ], + "spans": [ + { + "bbox": [ + 85, + 227, + 293, + 349 + ], + "type": "image", + "image_path": "8ca0ac5a4e3a682fee5599c6ab5ed0c5060ff70bc4ee10d2cfb5617aaa7aa150.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 111, + 349, + 127, + 355 + ], + "lines": [ + { + "bbox": [ + 111, + 349, + 127, + 355 + ], + "spans": [ + { + "bbox": [ + 111, + 349, + 127, + 355 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 143, + 349, + 181, + 355 + ], + "lines": [ + { + "bbox": [ + 143, + 349, + 181, + 355 + ], + "spans": [ + { + "bbox": [ + 143, + 349, + 181, + 355 + ], + "type": "text", + "content": "Output 1" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 194, + 349, + 231, + 355 + ], + "lines": [ + { + "bbox": [ + 194, + 349, + 231, + 355 + ], + "spans": [ + { + "bbox": [ + 194, + 349, + 231, + 355 + ], + "type": "text", + "content": "Output 2" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 243, + 349, + 293, + 355 + ], + "lines": [ + { + "bbox": [ + 243, + 349, + 293, + 355 + ], + "spans": [ + { + "bbox": [ + 243, + 349, + 293, + 355 + ], + "type": "text", + "content": "Output 3" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 85, + 356, + 293, + 441 + ], + "blocks": [ + { + "bbox": [ + 85, + 356, + 293, + 441 + ], + "lines": [ + { + "bbox": [ + 85, + 356, + 293, + 441 + ], + "spans": [ + { + "bbox": [ + 85, + 356, + 293, + 441 + ], + "type": "image", + "image_path": "d2ed03faeb021d1cd882c010f8c11c49e4e6d3eb5f10640473784caef20e3b43.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 111, + 442, + 127, + 449 + ], + "lines": [ + { + "bbox": [ + 111, + 442, + 127, + 449 + ], + "spans": [ + { + "bbox": [ + 111, + 442, + 127, + 449 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 145, + 441, + 189, + 449 + ], + "lines": [ + { + "bbox": [ + 145, + 441, + 189, + 449 + ], + "spans": [ + { + "bbox": [ + 145, + 441, + 189, + 449 + ], + "type": "text", + "content": "Input zoom in" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 209, + 442, + 228, + 449 + ], + "lines": [ + { + "bbox": [ + 209, + 442, + 228, + 449 + ], + "spans": [ + { + "bbox": [ + 209, + 442, + 228, + 449 + ], + "type": "text", + "content": "Output" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 243, + 441, + 290, + 449 + ], + "lines": [ + { + "bbox": [ + 243, + 441, + 290, + 449 + ], + "spans": [ + { + "bbox": [ + 243, + 441, + 290, + 449 + ], + "type": "text", + "content": "Output zoom in" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 452, + 547, + 485 + ], + "lines": [ + { + "bbox": [ + 46, + 452, + 547, + 485 + ], + "spans": [ + { + "bbox": [ + 46, + 452, + 547, + 485 + ], + "type": "text", + "content": "Figure 1. We introduce GenN2N, a unified framework for NeRF-to-NeRF translation, enabling a range of 3D NeRF editing tasks, including text-driven editing, colorization, super-resolution, inpainting, etc. We show at least two rendering views of edited NeRF scenes at inference time. Given a 3D NeRF scene, GenN2N can produce high-quality editing results with suitable multi-view consistency." + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 309, + 228, + 356, + 289 + ], + "blocks": [ + { + "bbox": [ + 309, + 228, + 356, + 289 + ], + "lines": [ + { + "bbox": [ + 309, + 228, + 356, + 289 + ], + "spans": [ + { + "bbox": [ + 309, + 228, + 356, + 289 + ], + "type": "image", + "image_path": "0c4653d18e861ac74f52cc629cb0c01ddec6feb4d59d833ca8a08d8348335202.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 358, + 228, + 405, + 289 + ], + "blocks": [ + { + "bbox": [ + 358, + 228, + 405, + 289 + ], + "lines": [ + { + "bbox": [ + 358, + 228, + 405, + 289 + ], + "spans": [ + { + "bbox": [ + 358, + 228, + 405, + 289 + ], + "type": "image", + "image_path": "ced5e4e1b34e9e675d51c4effdee96482ae89d6c6181312791da3b8b44b820c1.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 407, + 228, + 455, + 289 + ], + "blocks": [ + { + "bbox": [ + 407, + 228, + 455, + 289 + ], + "lines": [ + { + "bbox": [ + 407, + 228, + 455, + 289 + ], + "spans": [ + { + "bbox": [ + 407, + 228, + 455, + 289 + ], + "type": "image", + "image_path": "bbe829216193b8afbe4b78fbd624a9f4100a2203e96ea105e3c7992549d36d9d.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 457, + 228, + 505, + 289 + ], + "blocks": [ + { + "bbox": [ + 457, + 228, + 505, + 289 + ], + "lines": [ + { + "bbox": [ + 457, + 228, + 505, + 289 + ], + "spans": [ + { + "bbox": [ + 457, + 228, + 505, + 289 + ], + "type": "image", + "image_path": "d73bb4246a5bbc4aaa5627ac0a4a65ec97e627397013eac5045b5c52ca9ff5c2.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 298, + 289, + 356, + 349 + ], + "blocks": [ + { + "bbox": [ + 297, + 270, + 306, + 307 + ], + "lines": [ + { + "bbox": [ + 297, + 270, + 306, + 307 + ], + "spans": [ + { + "bbox": [ + 297, + 270, + 306, + 307 + ], + "type": "text", + "content": "aannnnnne aannnnnne" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 298, + 289, + 356, + 349 + ], + "lines": [ + { + "bbox": [ + 298, + 289, + 356, + 349 + ], + "spans": [ + { + "bbox": [ + 298, + 289, + 356, + 349 + ], + "type": "image", + "image_path": "334c92776c42b04e72183a28f470290c4c36b14817becffc441caecd7a1a763d.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 325, + 349, + 340, + 355 + ], + "lines": [ + { + "bbox": [ + 325, + 349, + 340, + 355 + ], + "spans": [ + { + "bbox": [ + 325, + 349, + 340, + 355 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 358, + 289, + 405, + 349 + ], + "blocks": [ + { + "bbox": [ + 358, + 289, + 405, + 349 + ], + "lines": [ + { + "bbox": [ + 358, + 289, + 405, + 349 + ], + "spans": [ + { + "bbox": [ + 358, + 289, + 405, + 349 + ], + "type": "image", + "image_path": "92fe8edd7c667b81fa33654f47f1efc84d0933ea7f8e8393eb274fcb58d702b2.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 369, + 349, + 394, + 355 + ], + "lines": [ + { + "bbox": [ + 369, + 349, + 394, + 355 + ], + "spans": [ + { + "bbox": [ + 369, + 349, + 394, + 355 + ], + "type": "text", + "content": "Output 1" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 407, + 289, + 455, + 349 + ], + "blocks": [ + { + "bbox": [ + 407, + 289, + 455, + 349 + ], + "lines": [ + { + "bbox": [ + 407, + 289, + 455, + 349 + ], + "spans": [ + { + "bbox": [ + 407, + 289, + 455, + 349 + ], + "type": "image", + "image_path": "3db7a8a2143ef22b17deadc2495d63a5b98d3aa22ffd082fd475b3b581374966.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 419, + 349, + 444, + 355 + ], + "lines": [ + { + "bbox": [ + 419, + 349, + 444, + 355 + ], + "spans": [ + { + "bbox": [ + 419, + 349, + 444, + 355 + ], + "type": "text", + "content": "Output 2" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 457, + 289, + 505, + 349 + ], + "blocks": [ + { + "bbox": [ + 457, + 289, + 505, + 349 + ], + "lines": [ + { + "bbox": [ + 457, + 289, + 505, + 349 + ], + "spans": [ + { + "bbox": [ + 457, + 289, + 505, + 349 + ], + "type": "image", + "image_path": "de71d79b9a5a341e7ad1909c3bb5915cc710552ad11da465162dff7bb897b6b2.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 468, + 349, + 493, + 355 + ], + "lines": [ + { + "bbox": [ + 468, + 349, + 493, + 355 + ], + "spans": [ + { + "bbox": [ + 468, + 349, + 493, + 355 + ], + "type": "text", + "content": "Output 3" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 298, + 356, + 356, + 441 + ], + "blocks": [ + { + "bbox": [ + 298, + 356, + 356, + 441 + ], + "lines": [ + { + "bbox": [ + 298, + 356, + 356, + 441 + ], + "spans": [ + { + "bbox": [ + 298, + 356, + 356, + 441 + ], + "type": "image", + "image_path": "b1f6eff79e0b92e2dd375eb6d1e4695dd5221a8a1c1de8b36d184a78266158d4.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 322, + 441, + 342, + 448 + ], + "lines": [ + { + "bbox": [ + 322, + 441, + 342, + 448 + ], + "spans": [ + { + "bbox": [ + 322, + 441, + 342, + 448 + ], + "type": "text", + "content": "View 1" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 369, + 441, + 393, + 448 + ], + "lines": [ + { + "bbox": [ + 369, + 441, + 393, + 448 + ], + "spans": [ + { + "bbox": [ + 369, + 441, + 393, + 448 + ], + "type": "text", + "content": "View 2" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 358, + 356, + 405, + 398 + ], + "blocks": [ + { + "bbox": [ + 358, + 356, + 405, + 398 + ], + "lines": [ + { + "bbox": [ + 358, + 356, + 405, + 398 + ], + "spans": [ + { + "bbox": [ + 358, + 356, + 405, + 398 + ], + "type": "image", + "image_path": "76ce41d1ad5c40dfef353ad597af3b9249363f203cdd96812838d74c78398c01.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 407, + 356, + 455, + 398 + ], + "blocks": [ + { + "bbox": [ + 407, + 356, + 455, + 398 + ], + "lines": [ + { + "bbox": [ + 407, + 356, + 455, + 398 + ], + "spans": [ + { + "bbox": [ + 407, + 356, + 455, + 398 + ], + "type": "image", + "image_path": "9b8679bb30949d69e7341a3fcf6c96328e98600ff1eae0f71a07aa65eed6ff82.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 407, + 399, + 455, + 441 + ], + "blocks": [ + { + "bbox": [ + 407, + 399, + 455, + 441 + ], + "lines": [ + { + "bbox": [ + 407, + 399, + 455, + 441 + ], + "spans": [ + { + "bbox": [ + 407, + 399, + 455, + 441 + ], + "type": "image", + "image_path": "3c924143fbd79fdff13d90837596c4043014634b0243cf5d20145281e3eea6c2.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 421, + 441, + 441, + 448 + ], + "lines": [ + { + "bbox": [ + 421, + 441, + 441, + 448 + ], + "spans": [ + { + "bbox": [ + 421, + 441, + 441, + 448 + ], + "type": "text", + "content": "View 3" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 457, + 356, + 505, + 441 + ], + "blocks": [ + { + "bbox": [ + 457, + 356, + 505, + 441 + ], + "lines": [ + { + "bbox": [ + 457, + 356, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 457, + 356, + 505, + 441 + ], + "type": "image", + "image_path": "0bd289d18c6ebb9f6a2c63c2d2f309661a17c8aeeb751f76321ba5ed15242781.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 470, + 441, + 492, + 448 + ], + "lines": [ + { + "bbox": [ + 470, + 441, + 492, + 448 + ], + "spans": [ + { + "bbox": [ + 470, + 441, + 492, + 448 + ], + "type": "text", + "content": "View 4" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "bbox": [ + 143, + 495, + 192, + 507 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 495, + 192, + 507 + ], + "spans": [ + { + "bbox": [ + 143, + 495, + 192, + 507 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 46, + 521, + 289, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 289, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 289, + 689 + ], + "type": "text", + "content": "We present GenN2N, a unified NeRF-to-NeRF translation framework for various NeRF translation tasks such as text-driven NeRF editing, colorization, super-resolution, inpainting, etc. Unlike previous methods designed for individual translation tasks with task-specific schemes, GenN2N achieves all these NeRF editing tasks by employing a plug-and-play image-to-image translator to perform editing in the 2D domain and lifting 2D edits into the 3D NeRF space. Since the 3D consistency of 2D edits may not be assured, we propose to model the distribution of the underlying 3D edits through a generative model that can cover all possible edited NeRFs. To model the distribution of 3D edited NeRFs from 2D edited images, we carefully design a VAE-GAN that encodes images while decoding NeRFs. The la" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 306, + 496, + 547, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 496, + 547, + 616 + ], + "spans": [ + { + "bbox": [ + 306, + 496, + 547, + 616 + ], + "type": "text", + "content": "tent space is trained to align with a Gaussian distribution and the NeRFs are supervised through an adversarial loss on its renderings. To ensure the latent code does not depend on 2D viewpoints but truly reflects the 3D edits, we also regularize the latent code through a contrastive learning scheme. Extensive experiments on various editing tasks show GenN2N, as a universal framework, performs as well or better than task-specific specialists while possessing flexible generative power. More results on our project page: https://xiangyueliu.github.io/GenN2N/." + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 306, + 643, + 387, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 643, + 387, + 657 + ], + "spans": [ + { + "bbox": [ + 306, + 643, + 387, + 657 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "text", + "content": "Over the past few years, Neural radiance fields (NeRFs) [23] have brought a promising paradigm in the realm of 3D reconstruction, 3D generation, and novel view synthesis due to their unparalleled compactness," + } + ] + } + ], + "index": 44 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 139, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 139, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 139, + 712 + ], + "type": "text", + "content": "†Corresponding authors." + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5105" + } + ] + } + ], + "index": 46 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "type": "text", + "content": "high quality, and versatility. Extensive research efforts have been devoted to creating NeRF scenes from 2D images [4, 20, 22, 36, 38] or just text [13, 26] input. However, once the NeRF scenes have been created, these methods often lack further control over the generated geometry and appearance. NeRF editing has therefore become a notable research focus recently." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 158, + 288, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 158, + 288, + 326 + ], + "spans": [ + { + "bbox": [ + 46, + 158, + 288, + 326 + ], + "type": "text", + "content": "Existing NeRF editing schemes are usually task-specific. For example, researchers have developed NeRF-SR [34], NeRF-In [19], PaletteNeRF [16] for NeRF super-resolution, inpainting, and color-editing respectively. These designs require a significant amount of domain knowledge for each specific task. On the other hand, in the field of 2D image editing, a growing trend is to develop universal image-to-image translation methods to support versatile image editing [25, 29, 43]. By leveraging foundational 2D generative models, e.g., stable diffusion [28], these methods achieve impressive editing results without task-specific customization or tuning. We then ask the question: can we conduct universal NeRF editing leveraging foundational 2D generative models as well?" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 328, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 328, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 328, + 287, + 604 + ], + "type": "text", + "content": "The first challenge is the representation gap between NeRFs and 2D images. It is not intuitive how to leverage image editing tools to edit NeRFs. A recent text-driven NeRF editing method [9] has shed some light on this. The method adopts a \"render-edit-aggregate\" pipeline. Specifically, it gradually updates a NeRF scene by iteratively rendering multi-view images, conducting text-driven visual editing on these images, and finally aggregating the edits in the NeRF scene. It seems that replacing the image editing tool with a universal image-to-image translator could lead to a universal NeRF editing method. However, the second challenge would then come. Image-to-image translators usually generate diverse and inconsistent edits for different views, e.g. turning a man into an elf might or might not put a hat on his head, making edits aggregation intricate. Regarding this challenge, Instruct-NeRF2NeRF [9] presents a complex optimization technique to pursue unblurred NeRF with inconsistent multi-view edits. Due to its complexity, the optimization cannot ensure the robustness of the outcomes. Additionally, the unique optimization outcome fails to reflect the stochastic nature of NeRF editing. Users typically anticipate a variety of edited NeRFs just like the diverse edited images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "text", + "content": "To tackle the challenges above, we propose GenN2N, a unified NeRF-to-NeRF translation framework for various NeRF editing tasks such as text-driven editing, colorization, super-resolution, inpainting (see Fig. 1). In contrast to Instruct-NeRF2NeRF which adopts a \"render-edit-aggregate\" pipeline, we first render a NeRF scene into multi-view images, then exploit an image-to-image translator to edit different views, and finally learn a generative model to depict the distribution of NeRF edits. Instead of" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "aggregating all the image edits to form a single NeRF edit, our key idea is to embrace the stochastic nature of content editing by modeling the distribution of the edits in the 3D NeRF space." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 121, + 546, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 121, + 546, + 395 + ], + "spans": [ + { + "bbox": [ + 304, + 121, + 546, + 395 + ], + "type": "text", + "content": "Specifically given a NeRF model or its multi-view images, along with the editing goal, we first generate edited multi-view images using a plug-and-play image-to-image translator. Each view corresponds to a unique 3D edit with some geometry or appearance variations. Conditioned on the input NeRF, GenN2N trains a conditional 3D generative model to reflect such content variations. At the core of GenN2N, we design a 3D VAE-GAN that incorporates a differentiable volume renderer to connect 3D content creation with 2D GAN losses, ensuring that the inconsistent multi-view renderings can still help each other regarding 3D generation. Moreover, we introduce a contrastive learning loss to ensure that the 3D content variation can be successfully understood just from edited 2D images without being influenced by the camera viewpoints. During inference, users can simply sample from the conditional generative model to obtain various 3D editing results aligned with the editing goal. We have conducted experiments on human, items, indoor and outdoor scenes for various editing tasks such as text-driven editing, colorization, superresolution and inpainting, demonstrating the effectiveness of GenN2N in supporting diverse NeRF editing tasks while keeping the multi-view consistency of the edited NeRF." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 318, + 396, + 543, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 396, + 543, + 407 + ], + "spans": [ + { + "bbox": [ + 318, + 396, + 543, + 407 + ], + "type": "text", + "content": "We summarize the contribution of this paper as follows," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 407, + 545, + 537 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 306, + 407, + 545, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 407, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 306, + 407, + 545, + 441 + ], + "type": "text", + "content": "- A generative NeRF-to-NeRF translation formulation for the universal NeRF editing task together with a generic solution;" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 443, + 545, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 443, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 306, + 443, + 545, + 479 + ], + "type": "text", + "content": "- a 3D VAE-GAN framework that can learn the distribution of all possible 3D NeRF edits corresponding to the a set of input edited 2D images;" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 479, + 545, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 545, + 503 + ], + "type": "text", + "content": "- a contrastive learning framework that can disentangle the 3D edits and 2D camera views from edited images;" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 503, + 545, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 503, + 545, + 537 + ], + "spans": [ + { + "bbox": [ + 306, + 503, + 545, + 537 + ], + "type": "text", + "content": "- extensive experiments demonstrating the superior efficiency, quality, and diversity of the NeRF-to-NeRF translation results." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 550, + 392, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 550, + 392, + 562 + ], + "spans": [ + { + "bbox": [ + 306, + 550, + 392, + 562 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "text", + "content": "NeRF Editing. Previous works such as EditNeRF [21] propose a conditional neural field that enables shape and appearance editing in the latent space. PaletteNeRF [16, 37] focuses on controlling color palette weights to manipulate appearance. Other approaches utilize bounding boxes [41], meshes [39], point clouds [6], key points [45], or feature volumes [18] to directly manipulate the spatial representation of NeRF. However, these methods either heavily rely on user interactions or have limitations in terms of spatial deformation and color transfer capabilities." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": "NeRF Stylization. Images-referenced stylization [7, 12, 42] often prioritize capturing texture style rather than de" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5106" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 545, + 223 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 545, + 223 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 545, + 223 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 545, + 223 + ], + "type": "image", + "image_path": "d47cf812fb1b9fdcafce9ba7a0f9bd66826f1f7cd6321f18674386b1d0abcafb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 232, + 548, + 301 + ], + "lines": [ + { + "bbox": [ + 46, + 232, + 548, + 301 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 548, + 301 + ], + "type": "text", + "content": "Figure 2. Overview of GenN2N. We first edit the source image set " + }, + { + "bbox": [ + 46, + 232, + 548, + 301 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{I}_i\\}_{i=0}^{N-1}" + }, + { + "bbox": [ + 46, + 232, + 548, + 301 + ], + "type": "text", + "content": " using 2D image-to-image translation methods, e.g., text-driven editing, colorization, zoom out, etc. For each view " + }, + { + "bbox": [ + 46, + 232, + 548, + 301 + ], + "type": "inline_equation", + "content": "i \\in [0, N-1]" + }, + { + "bbox": [ + 46, + 232, + 548, + 301 + ], + "type": "text", + "content": ", we generate " + }, + { + "bbox": [ + 46, + 232, + 548, + 301 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 232, + 548, + 301 + ], + "type": "text", + "content": " edited images, resulting in a group of translated image set " + }, + { + "bbox": [ + 46, + 232, + 548, + 301 + ], + "type": "inline_equation", + "content": "\\{\\{\\mathbf{S}_i^j\\}_{j=0}^{M-1}\\}_{i=0}^{N-1}" + }, + { + "bbox": [ + 46, + 232, + 548, + 301 + ], + "type": "text", + "content": ". Then we use the Latent Distill Module to learn " + }, + { + "bbox": [ + 46, + 232, + 548, + 301 + ], + "type": "inline_equation", + "content": "M \\times N" + }, + { + "bbox": [ + 46, + 232, + 548, + 301 + ], + "type": "text", + "content": " edit code vectors from the translated image set, which serve as the input of the translated NeRF. To optimize our GenN2N, we design four loss functions: a KL loss to constrain the latent vectors to a Gaussian distribution; and " + }, + { + "bbox": [ + 46, + 232, + 548, + 301 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{recon}}" + }, + { + "bbox": [ + 46, + 232, + 548, + 301 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 232, + 548, + 301 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{adv}}" + }, + { + "bbox": [ + 46, + 232, + 548, + 301 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 232, + 548, + 301 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{contr}}" + }, + { + "bbox": [ + 46, + 232, + 548, + 301 + ], + "type": "text", + "content": " to optimize the appearance and geometry of the translated NeRF. At inference, we can sample a latent vector " + }, + { + "bbox": [ + 46, + 232, + 548, + 301 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 46, + 232, + 548, + 301 + ], + "type": "text", + "content": " from Gaussian distribution and render a corresponding multi-view consistent 3D scene with high quality." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 319, + 288, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 319, + 288, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 288, + 392 + ], + "type": "text", + "content": "tailed content, resulting in imprecise editing appearance of NeRF only. Text-guided works [33, 35], on the other hand, apply contrastive losses based on CLIP [27] to achieve the desired edits. While text references usually describe the global characteristics of the edited results, instructions offer a more convenient and precise expression." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 392, + 288, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 392, + 288, + 571 + ], + "spans": [ + { + "bbox": [ + 46, + 392, + 288, + 571 + ], + "type": "text", + "content": "Instruct-driven NeRF editing. Among numerous image-to-image translation works, InstructPix2Pix [2] stands out by efficiently editing images following instructions. It leverages large pre-trained models in the language and image domains [3, 28] to generate paired data (before and after editing) for training. While editing NeRF solely based on edited images is problematic due to multi-view inconsistency. To address this, an intuitive yet heavy approach [9] is to iteratively edit the image and optimize NeRF. In addition, NeRF-Art [35] and DreamEditor [46] adopt a CLIP-based contrastive loss [27] and score distillation sampling [26] separately to supervise the optimization of editing NeRF. Inspired by Generative Radiance Fields [5, 30], We capture various possible NeRF editing in the generative space to solve it." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 585, + 103, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 585, + 103, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 585, + 103, + 597 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": "Given a NeRF scene, we present a unified framework GenN2N to achieve various editing on the 3D scene leveraging geometry and texture priors from 2D image editing methods, such as text-driven editing, colorization, superresolution, inpainting, etc. While a universal image-to-image translator can theoretically accomplish these 2D editing tasks, we actually use a state-of-the-art translator for each task. Therefore, we formulate each 2D image editing method as a plug-and-play image-to-image translator" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 319, + 546, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 546, + 403 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 546, + 403 + ], + "type": "text", + "content": "and all NeRF editing tasks as our universal NeRF-to-NeRF translation, in which the given NeRF is translated into NeRF scenes with high rendering quality and 3D geometry consistency according to the user-selected editing target. The overview of GenN2N is illustrated in Fig. 2, we first perform image-to-image translation in the 2D domain and then lift 2D edits to 3D and achieve NeRF-to-NeRF translation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 404, + 546, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 404, + 546, + 513 + ], + "spans": [ + { + "bbox": [ + 304, + 404, + 546, + 513 + ], + "type": "text", + "content": "Given " + }, + { + "bbox": [ + 304, + 404, + 546, + 513 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 404, + 546, + 513 + ], + "type": "text", + "content": " multi-view images " + }, + { + "bbox": [ + 304, + 404, + 546, + 513 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{I}_i\\}_{i = 0}^{N - 1}" + }, + { + "bbox": [ + 304, + 404, + 546, + 513 + ], + "type": "text", + "content": " of a scene, we first use Nerfstudio [32] to train the original NeRF. Then we use a plug-and-play image-to-image translator to edit these source images. However, the content generated by the 2D translator may be inconsistent among multi-view images. For example, using different initial noise, the 2D translator [1] may generate different content for image editing, which makes it difficult to ensure the 3D consistency between different view directions in the 3D scene." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 515, + 547, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 515, + 547, + 588 + ], + "spans": [ + { + "bbox": [ + 304, + 515, + 547, + 588 + ], + "type": "text", + "content": "To ensure the 3D consistency and rendering quality, we propose to model the distribution of the underlying 3D edits through a generative model that can cover all possible edited NeRFs, by learning an edit code for each edited image so that the generated content can be controlled by this edit code during the NeRF-to-NeRF translation process." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 590, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 590, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 590, + 547, + 713 + ], + "type": "text", + "content": "For each view " + }, + { + "bbox": [ + 304, + 590, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\in [0,N - 1]" + }, + { + "bbox": [ + 304, + 590, + 547, + 713 + ], + "type": "text", + "content": ", we generate " + }, + { + "bbox": [ + 304, + 590, + 547, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 590, + 547, + 713 + ], + "type": "text", + "content": " edited images, resulting in a group of the translated image set " + }, + { + "bbox": [ + 304, + 590, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\{\\{\\mathbf{S}_i^j\\}_{j = 0}^{M - 1}\\}_{i = 0}^{N - 1}" + }, + { + "bbox": [ + 304, + 590, + 547, + 713 + ], + "type": "text", + "content": ". Then we design a Latent Distill Module described in Sec. 3.1 to map each translated image " + }, + { + "bbox": [ + 304, + 590, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_i^j" + }, + { + "bbox": [ + 304, + 590, + 547, + 713 + ], + "type": "text", + "content": " into an edit code vector " + }, + { + "bbox": [ + 304, + 590, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i^j" + }, + { + "bbox": [ + 304, + 590, + 547, + 713 + ], + "type": "text", + "content": " and design a KL loss " + }, + { + "bbox": [ + 304, + 590, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{KL}}" + }, + { + "bbox": [ + 304, + 590, + 547, + 713 + ], + "type": "text", + "content": " to constrain those edit code vectors to a Gaussian distribution. Conditioned on the edit code " + }, + { + "bbox": [ + 304, + 590, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i^j" + }, + { + "bbox": [ + 304, + 590, + 547, + 713 + ], + "type": "text", + "content": ", we perform NeRF-to-NeRF translation in Sec. 3.2 by rendering multi-view images " + }, + { + "bbox": [ + 304, + 590, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{C}_i\\}_{i = 0}^{N - 1}" + }, + { + "bbox": [ + 304, + 590, + 547, + 713 + ], + "type": "text", + "content": " and optimize the translated NeRF by three loss functions: the reconstruction loss " + }, + { + "bbox": [ + 304, + 590, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{recon}}" + }, + { + "bbox": [ + 304, + 590, + 547, + 713 + ], + "type": "text", + "content": ", the adversarial" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5107" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 71, + 275, + 237 + ], + "blocks": [ + { + "bbox": [ + 62, + 71, + 275, + 237 + ], + "lines": [ + { + "bbox": [ + 62, + 71, + 275, + 237 + ], + "spans": [ + { + "bbox": [ + 62, + 71, + 275, + 237 + ], + "type": "image", + "image_path": "360c66cdad8c1dac2d8d0bec7f5be42a74c2b342e866fcd367df72d5cec8fdaf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 246, + 288, + 312 + ], + "lines": [ + { + "bbox": [ + 46, + 246, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 246, + 288, + 312 + ], + "type": "text", + "content": "Figure 3. Illustration of our proposed contrastive loss functions. Regarding the multi-view rendered images " + }, + { + "bbox": [ + 46, + 246, + 288, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_i^j" + }, + { + "bbox": [ + 46, + 246, + 288, + 312 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 246, + 288, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_l^j" + }, + { + "bbox": [ + 46, + 246, + 288, + 312 + ], + "type": "text", + "content": " sharing the same edit code, we resend them to our Latent Distill Module to extract " + }, + { + "bbox": [ + 46, + 246, + 288, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i^j" + }, + { + "bbox": [ + 46, + 246, + 288, + 312 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 246, + 288, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_l^j" + }, + { + "bbox": [ + 46, + 246, + 288, + 312 + ], + "type": "text", + "content": ", and aggregate them via " + }, + { + "bbox": [ + 46, + 246, + 288, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{contr}}^{\\mathrm{att}}" + }, + { + "bbox": [ + 46, + 246, + 288, + 312 + ], + "type": "text", + "content": ". In addition, for " + }, + { + "bbox": [ + 46, + 246, + 288, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_i^k" + }, + { + "bbox": [ + 46, + 246, + 288, + 312 + ], + "type": "text", + "content": " whose editing style varies from " + }, + { + "bbox": [ + 46, + 246, + 288, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_i^j" + }, + { + "bbox": [ + 46, + 246, + 288, + 312 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 246, + 288, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{contr}}^{\\mathrm{rep}}" + }, + { + "bbox": [ + 46, + 246, + 288, + 312 + ], + "type": "text", + "content": " increases the distance between edit codes of them." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 55, + 327, + 280, + 416 + ], + "blocks": [ + { + "bbox": [ + 55, + 327, + 280, + 416 + ], + "lines": [ + { + "bbox": [ + 55, + 327, + 280, + 416 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 280, + 416 + ], + "type": "image", + "image_path": "a9f17cc1a480904a0523620ddc0d9d809312deb467bf9480bd9e57c4f35e42ae.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 426, + 287, + 492 + ], + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 492 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 492 + ], + "type": "text", + "content": "Figure 4. Illustration of our proposed conditional adversarial loss functions. Our conditional discriminator distinguishes artifacts such as blur and distortion in novel-view rendered image " + }, + { + "bbox": [ + 46, + 426, + 287, + 492 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_l^j" + }, + { + "bbox": [ + 46, + 426, + 287, + 492 + ], + "type": "text", + "content": " compared with target image " + }, + { + "bbox": [ + 46, + 426, + 287, + 492 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_l^j" + }, + { + "bbox": [ + 46, + 426, + 287, + 492 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 426, + 287, + 492 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_l^j" + }, + { + "bbox": [ + 46, + 426, + 287, + 492 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 426, + 287, + 492 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_l^k" + }, + { + "bbox": [ + 46, + 426, + 287, + 492 + ], + "type": "text", + "content": " are edited with same view but various styles, the latter serves as the condition to concatenate with " + }, + { + "bbox": [ + 46, + 426, + 287, + 492 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_l^j" + }, + { + "bbox": [ + 46, + 426, + 287, + 492 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 426, + 287, + 492 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_l^j" + }, + { + "bbox": [ + 46, + 426, + 287, + 492 + ], + "type": "text", + "content": " and manufacture fake and real pairs." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 515, + 287, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 515, + 287, + 575 + ], + "spans": [ + { + "bbox": [ + 46, + 515, + 287, + 575 + ], + "type": "text", + "content": "loss " + }, + { + "bbox": [ + 46, + 515, + 287, + 575 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{AD}}" + }, + { + "bbox": [ + 46, + 515, + 287, + 575 + ], + "type": "text", + "content": ", and the contrastive loss " + }, + { + "bbox": [ + 46, + 515, + 287, + 575 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{contr}}" + }, + { + "bbox": [ + 46, + 515, + 287, + 575 + ], + "type": "text", + "content": ". After the optimization of the translated NeRF, as described in Sec. 3.3, we can sample an edit code " + }, + { + "bbox": [ + 46, + 515, + 287, + 575 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 46, + 515, + 287, + 575 + ], + "type": "text", + "content": " from Gaussian distribution and render the corresponding edited 3D scene with high quality and multi-view consistency in the inference stage." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 586, + 170, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 170, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 170, + 597 + ], + "type": "text", + "content": "3.1. Latent Distill Module" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "Image Translation. As illustrated in Fig. 2, GenN2N is a unified framework for NeRF-to-NeRF translation, in which the core is to perform a 2D image-to-image translation and lift 2D edits into universal 3D NeRF-to-NeRF translation. Given the source multi-view image set " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{I}_i\\}_{i=0}^{N-1}" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " of a NeRF scene, we first perform image editing M times for each view using a plug-and-play 2D image-to-image translator, producing a group of translated image set " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\{\\{\\mathbf{S}_i^j\\}_{j=0}^{M-1}\\}_{i=0}^{N-1}" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": ". In this paper, we use several 2D translation tasks to show" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "the adaptability of our GenN2N: text-driven editing, superresolution, colorization and inpainting. For more details about those 2D image editing methods, please refer to the supplementary materials." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 121, + 546, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 121, + 546, + 254 + ], + "spans": [ + { + "bbox": [ + 304, + 121, + 546, + 254 + ], + "type": "text", + "content": "Edit Code. Since 2D image-to-image translation may generate different content even with the same editing target, causing the inconsistency problem in the 3D scene. We propose to map each edited image " + }, + { + "bbox": [ + 304, + 121, + 546, + 254 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_i^j" + }, + { + "bbox": [ + 304, + 121, + 546, + 254 + ], + "type": "text", + "content": " into a latent feature vector named edit code to characterize these diverse editorings. We employ the off-the-shelf VAE encoder from stable diffusion [28] to extract the feature from " + }, + { + "bbox": [ + 304, + 121, + 546, + 254 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_i^j" + }, + { + "bbox": [ + 304, + 121, + 546, + 254 + ], + "type": "text", + "content": " and then apply a tiny MLP network to produce this edit code " + }, + { + "bbox": [ + 304, + 121, + 546, + 254 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i^j \\in \\mathbb{R}^{64}" + }, + { + "bbox": [ + 304, + 121, + 546, + 254 + ], + "type": "text", + "content": ". During the training process, we keep the pre-trained encoder fixed and only optimize the parameters of the tiny MLP network. This mapping process can be formulated as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 369, + 263, + 545, + 279 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 263, + 545, + 279 + ], + "spans": [ + { + "bbox": [ + 369, + 263, + 545, + 279 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {i} ^ {j} = \\mathcal {D} \\left(\\mathbf {S} _ {i} ^ {j}\\right) = \\mathcal {M} \\left(\\mathcal {E} \\left(\\mathbf {S} _ {i} ^ {j}\\right)\\right), \\tag {1}", + "image_path": "a64149b10ec0d9d98727afee1faf202142d1ce5a90f7d715f1eb6ae720c42bf2.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 289, + 545, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 545, + 312 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 289, + 545, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 289, + 545, + 312 + ], + "type": "text", + "content": " represent this mapping process, " + }, + { + "bbox": [ + 304, + 289, + 545, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 304, + 289, + 545, + 312 + ], + "type": "text", + "content": " is the fixed encoder, and " + }, + { + "bbox": [ + 304, + 289, + 545, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 289, + 545, + 312 + ], + "type": "text", + "content": " is the learnable tiny MLP." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 314, + 545, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 314, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 304, + 314, + 545, + 374 + ], + "type": "text", + "content": "KL loss. In order to facilitate effective sampling of the edit code so as to control the editing diversity of our NeRF-to-NeRF translation, we need to constrain the edit code to a well-defined distribution. Thus we design a KL loss to encourage " + }, + { + "bbox": [ + 304, + 314, + 545, + 374 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i^j" + }, + { + "bbox": [ + 304, + 314, + 545, + 374 + ], + "type": "text", + "content": " to approximate a Gaussian distribution:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 308, + 384, + 545, + 421 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 384, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 308, + 384, + 545, + 421 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {K L}} = \\mathbb {E} _ {\\mathbf {S} \\in \\{\\left\\{\\mathbf {S} _ {i} ^ {j} \\right\\} _ {j = 0} ^ {M - 1} \\} _ {i = 0} ^ {N - 1}} [ P (\\mathbf {z} _ {\\text {n o r m a l}}) \\log \\left(\\frac {P \\left(\\mathbf {z} _ {\\text {n o r m a l}}\\right)}{P \\left(\\mathcal {D} (\\mathbf {S})\\right)}\\right) ], \\tag {2}", + "image_path": "68494b0bb4d03c7b38343192b7bb37f2b72f85649d88d17f18f58fa964ba1a62.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 421, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 421, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 421, + 545, + 456 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 421, + 545, + 456 + ], + "type": "inline_equation", + "content": "P(\\mathbf{z}_{\\text{normal}})" + }, + { + "bbox": [ + 304, + 421, + 545, + 456 + ], + "type": "text", + "content": " denotes probability distribution of the standard Gaussian distribution in " + }, + { + "bbox": [ + 304, + 421, + 545, + 456 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{64}" + }, + { + "bbox": [ + 304, + 421, + 545, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 421, + 545, + 456 + ], + "type": "inline_equation", + "content": "P(\\mathcal{D}(\\mathbf{S}))" + }, + { + "bbox": [ + 304, + 421, + 545, + 456 + ], + "type": "text", + "content": " means probability distribution of the extracted edit codes." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 457, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 457, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 457, + 546, + 714 + ], + "type": "text", + "content": "Contrastive loss. It is not assured that edit codes " + }, + { + "bbox": [ + 304, + 457, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 304, + 457, + 546, + 714 + ], + "type": "text", + "content": " obtained from the Latent Distill Module contain only the editing information while excluding viewpoint-related effects. However, since the translated NeRF utilizes " + }, + { + "bbox": [ + 304, + 457, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 304, + 457, + 546, + 714 + ], + "type": "text", + "content": " to edit scenes, it yields instability if " + }, + { + "bbox": [ + 304, + 457, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 304, + 457, + 546, + 714 + ], + "type": "text", + "content": " violently changes given images that are similar in appearance but different in viewpoints. To ensure the latent code does not depend on 2D viewpoints but truly reflects the 3D edits, we regularize the latent code through a contrastive learning scheme. Specifically, we reduce the distance between edit codes of different-view rendered images from a translated NeRF that share the same edit code, while increasing the distance between same-view images that are multi-time edited by the 2D image-to-image translator. As illustrated in Fig. 3, given an edit code " + }, + { + "bbox": [ + 304, + 457, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i^j" + }, + { + "bbox": [ + 304, + 457, + 546, + 714 + ], + "type": "text", + "content": " extracted from the " + }, + { + "bbox": [ + 304, + 457, + 546, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 457, + 546, + 714 + ], + "type": "text", + "content": "-th input view at the " + }, + { + "bbox": [ + 304, + 457, + 546, + 714 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 457, + 546, + 714 + ], + "type": "text", + "content": "-th edited image " + }, + { + "bbox": [ + 304, + 457, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_i^j" + }, + { + "bbox": [ + 304, + 457, + 546, + 714 + ], + "type": "text", + "content": ", we render multi-view images " + }, + { + "bbox": [ + 304, + 457, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{C}_l^j\\}_{l=0}^{N-1}" + }, + { + "bbox": [ + 304, + 457, + 546, + 714 + ], + "type": "text", + "content": " using the translated NeRF conditioned on " + }, + { + "bbox": [ + 304, + 457, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i^j" + }, + { + "bbox": [ + 304, + 457, + 546, + 714 + ], + "type": "text", + "content": ". Then we employ contrastive learning to encourage the edit code " + }, + { + "bbox": [ + 304, + 457, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i^j" + }, + { + "bbox": [ + 304, + 457, + 546, + 714 + ], + "type": "text", + "content": " to be close to " + }, + { + "bbox": [ + 304, + 457, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{z}_l^j\\}_{l=0}^{N-1}" + }, + { + "bbox": [ + 304, + 457, + 546, + 714 + ], + "type": "text", + "content": " extracted from " + }, + { + "bbox": [ + 304, + 457, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{C}_l^j\\}_{l=0}^{N-1}" + }, + { + "bbox": [ + 304, + 457, + 546, + 714 + ], + "type": "text", + "content": ", while being distinct from the edit codes " + }, + { + "bbox": [ + 304, + 457, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{z}_i^k\\}_{k=0}^{M-1}" + }, + { + "bbox": [ + 304, + 457, + 546, + 714 + ], + "type": "text", + "content": " extracted from " + }, + { + "bbox": [ + 304, + 457, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{S}_i^k\\}_{k=0}^{M-1}" + }, + { + "bbox": [ + 304, + 457, + 546, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 457, + 546, + 714 + ], + "type": "inline_equation", + "content": "k \\neq j" + }, + { + "bbox": [ + 304, + 457, + 546, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5108" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 84, + 103, + 162 + ], + "blocks": [ + { + "bbox": [ + 51, + 84, + 103, + 162 + ], + "lines": [ + { + "bbox": [ + 51, + 84, + 103, + 162 + ], + "spans": [ + { + "bbox": [ + 51, + 84, + 103, + 162 + ], + "type": "image", + "image_path": "2a59429c383d15a77cd2d8111c9489f90c9975e3773858f91cdf8e17b9ca6414.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 111, + 84, + 164, + 161 + ], + "blocks": [ + { + "bbox": [ + 111, + 84, + 164, + 161 + ], + "lines": [ + { + "bbox": [ + 111, + 84, + 164, + 161 + ], + "spans": [ + { + "bbox": [ + 111, + 84, + 164, + 161 + ], + "type": "image", + "image_path": "a85184f26f80b850d50b0ef12f278ac2b2ec723cd1f0ed303db3fe012b93d0d2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 165, + 84, + 217, + 161 + ], + "blocks": [ + { + "bbox": [ + 165, + 84, + 217, + 161 + ], + "lines": [ + { + "bbox": [ + 165, + 84, + 217, + 161 + ], + "spans": [ + { + "bbox": [ + 165, + 84, + 217, + 161 + ], + "type": "image", + "image_path": "98732b7a251c45d6ae2e6607ba03db46cfbe6812cfe8939993d4d23aee91b90e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 220, + 84, + 272, + 161 + ], + "blocks": [ + { + "bbox": [ + 220, + 84, + 272, + 161 + ], + "lines": [ + { + "bbox": [ + 220, + 84, + 272, + 161 + ], + "spans": [ + { + "bbox": [ + 220, + 84, + 272, + 161 + ], + "type": "image", + "image_path": "f5697f4c13ae0b6187525bea0e5afce085e2f6b25e696ed76b1ff33e0b9cabb1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 239, + 163, + 365, + 174 + ], + "lines": [ + { + "bbox": [ + 239, + 163, + 365, + 174 + ], + "spans": [ + { + "bbox": [ + 239, + 163, + 365, + 174 + ], + "type": "text", + "content": "\"Make it an Egyptian sculpture\"" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 274, + 84, + 326, + 161 + ], + "blocks": [ + { + "bbox": [ + 274, + 84, + 326, + 161 + ], + "lines": [ + { + "bbox": [ + 274, + 84, + 326, + 161 + ], + "spans": [ + { + "bbox": [ + 274, + 84, + 326, + 161 + ], + "type": "image", + "image_path": "5551ee4ed2848b775519e0d428a75b8c161bb2658b4b57c22de51584781157e3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 328, + 84, + 380, + 161 + ], + "blocks": [ + { + "bbox": [ + 328, + 84, + 380, + 161 + ], + "lines": [ + { + "bbox": [ + 328, + 84, + 380, + 161 + ], + "spans": [ + { + "bbox": [ + 328, + 84, + 380, + 161 + ], + "type": "image", + "image_path": "0c076c18fb344ee4ed1f89a07b0027a2a2813dfff2fe39aec549fa29ecb7bddd.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 382, + 84, + 434, + 161 + ], + "blocks": [ + { + "bbox": [ + 382, + 84, + 434, + 161 + ], + "lines": [ + { + "bbox": [ + 382, + 84, + 434, + 161 + ], + "spans": [ + { + "bbox": [ + 382, + 84, + 434, + 161 + ], + "type": "image", + "image_path": "bdf6fd6dceefa3cdcfbb5cb2e621dffc633583c58826e844f8c3836e76f977bb.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 438, + 84, + 490, + 161 + ], + "blocks": [ + { + "bbox": [ + 438, + 84, + 490, + 161 + ], + "lines": [ + { + "bbox": [ + 438, + 84, + 490, + 161 + ], + "spans": [ + { + "bbox": [ + 438, + 84, + 490, + 161 + ], + "type": "image", + "image_path": "27d3cb7179f57174f11a26b833c9c166e3bb171a9b7440c44ba41a254b2148c0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 492, + 84, + 543, + 161 + ], + "blocks": [ + { + "bbox": [ + 492, + 84, + 543, + 161 + ], + "lines": [ + { + "bbox": [ + 492, + 84, + 543, + 161 + ], + "spans": [ + { + "bbox": [ + 492, + 84, + 543, + 161 + ], + "type": "image", + "image_path": "68c72715214e058cb0f2e3267afddc95bf1a646ba91ff7da689511af97d350a7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 51, + 176, + 103, + 254 + ], + "blocks": [ + { + "bbox": [ + 51, + 176, + 103, + 254 + ], + "lines": [ + { + "bbox": [ + 51, + 176, + 103, + 254 + ], + "spans": [ + { + "bbox": [ + 51, + 176, + 103, + 254 + ], + "type": "image", + "image_path": "e25597512fa654b4a23cb6294a3ae57b7139848cb17c43e70f30c65fb8e2768c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 50, + 255, + 102, + 265 + ], + "lines": [ + { + "bbox": [ + 50, + 255, + 102, + 265 + ], + "spans": [ + { + "bbox": [ + 50, + 255, + 102, + 265 + ], + "type": "text", + "content": "Original NeRF" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 112, + 176, + 164, + 254 + ], + "blocks": [ + { + "bbox": [ + 112, + 176, + 164, + 254 + ], + "lines": [ + { + "bbox": [ + 112, + 176, + 164, + 254 + ], + "spans": [ + { + "bbox": [ + 112, + 176, + 164, + 254 + ], + "type": "image", + "image_path": "11939d6053dfc042d19524d56e7c064b6832670414dca9b13dbba17776f35179.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 145, + 255, + 183, + 264 + ], + "lines": [ + { + "bbox": [ + 145, + 255, + 183, + 264 + ], + "spans": [ + { + "bbox": [ + 145, + 255, + 183, + 264 + ], + "type": "text", + "content": "Inference 1" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 165, + 176, + 217, + 254 + ], + "blocks": [ + { + "bbox": [ + 165, + 176, + 217, + 254 + ], + "lines": [ + { + "bbox": [ + 165, + 176, + 217, + 254 + ], + "spans": [ + { + "bbox": [ + 165, + 176, + 217, + 254 + ], + "type": "image", + "image_path": "0308bd078b45e6b9fa17bb4b800cddd9452bd3877c20c917301b8d073103d902.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 275, + 545, + 298 + ], + "lines": [ + { + "bbox": [ + 46, + 275, + 545, + 298 + ], + "spans": [ + { + "bbox": [ + 46, + 275, + 545, + 298 + ], + "type": "text", + "content": "Figure 5. Text-Driven Editing. We sample 4 inference results for both text-driven editing tasks. The diversity of geometry and appearance showcases awesome generative ability of GenN2N, on the premise of maintaining the 3D consistency between different viewpoints." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 220, + 176, + 272, + 254 + ], + "blocks": [ + { + "bbox": [ + 220, + 176, + 272, + 254 + ], + "lines": [ + { + "bbox": [ + 220, + 176, + 272, + 254 + ], + "spans": [ + { + "bbox": [ + 220, + 176, + 272, + 254 + ], + "type": "image", + "image_path": "c6dafb7f106bd694903804ccc5d18a5615d7996d3a7b6f440088be450407316b.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 253, + 255, + 293, + 264 + ], + "lines": [ + { + "bbox": [ + 253, + 255, + 293, + 264 + ], + "spans": [ + { + "bbox": [ + 253, + 255, + 293, + 264 + ], + "type": "text", + "content": "Inference 2" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 274, + 176, + 326, + 254 + ], + "blocks": [ + { + "bbox": [ + 274, + 176, + 326, + 254 + ], + "lines": [ + { + "bbox": [ + 274, + 176, + 326, + 254 + ], + "spans": [ + { + "bbox": [ + 274, + 176, + 326, + 254 + ], + "type": "image", + "image_path": "fa33bc58d5b4297395423c609e762e0a71d75f3817b33c4935b04ddb246b1527.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 328, + 176, + 380, + 254 + ], + "blocks": [ + { + "bbox": [ + 328, + 176, + 380, + 254 + ], + "lines": [ + { + "bbox": [ + 328, + 176, + 380, + 254 + ], + "spans": [ + { + "bbox": [ + 328, + 176, + 380, + 254 + ], + "type": "image", + "image_path": "8842118892dba43b2f3e21b506c5b8f6939f5d54979913852e941058b0b11156.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 362, + 255, + 401, + 263 + ], + "lines": [ + { + "bbox": [ + 362, + 255, + 401, + 263 + ], + "spans": [ + { + "bbox": [ + 362, + 255, + 401, + 263 + ], + "type": "text", + "content": "Inference 3" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 382, + 176, + 434, + 254 + ], + "blocks": [ + { + "bbox": [ + 382, + 176, + 434, + 254 + ], + "lines": [ + { + "bbox": [ + 382, + 176, + 434, + 254 + ], + "spans": [ + { + "bbox": [ + 382, + 176, + 434, + 254 + ], + "type": "image", + "image_path": "0d807275c45433d989082d0d688d823214358915658926e02c9694f2e47ab562.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 438, + 176, + 490, + 254 + ], + "blocks": [ + { + "bbox": [ + 438, + 176, + 490, + 254 + ], + "lines": [ + { + "bbox": [ + 438, + 176, + 490, + 254 + ], + "spans": [ + { + "bbox": [ + 438, + 176, + 490, + 254 + ], + "type": "image", + "image_path": "8afa615510a22bd70d4b97463a0f0c94960329822ac70b51e206cc9cab6f85e4.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 471, + 255, + 511, + 263 + ], + "lines": [ + { + "bbox": [ + 471, + 255, + 511, + 263 + ], + "spans": [ + { + "bbox": [ + 471, + 255, + 511, + 263 + ], + "type": "text", + "content": "Inference 4" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 492, + 176, + 543, + 254 + ], + "blocks": [ + { + "bbox": [ + 492, + 176, + 543, + 254 + ], + "lines": [ + { + "bbox": [ + 492, + 176, + 543, + 254 + ], + "spans": [ + { + "bbox": [ + 492, + 176, + 543, + 254 + ], + "type": "image", + "image_path": "693f9a1bbddab8ae2ec0745817b0d3445cc5bbae8e5299b157be7e5693b66b72.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 50, + 309, + 128, + 429 + ], + "blocks": [ + { + "bbox": [ + 50, + 309, + 128, + 429 + ], + "lines": [ + { + "bbox": [ + 50, + 309, + 128, + 429 + ], + "spans": [ + { + "bbox": [ + 50, + 309, + 128, + 429 + ], + "type": "image", + "image_path": "95d490a396cb3401e2eedb9da334e77f98f2c15cde96bd6215e5343719112fb7.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 430, + 113, + 438 + ], + "lines": [ + { + "bbox": [ + 63, + 430, + 113, + 438 + ], + "spans": [ + { + "bbox": [ + 63, + 430, + 113, + 438 + ], + "type": "text", + "content": "Original NeRF" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 137, + 309, + 216, + 369 + ], + "blocks": [ + { + "bbox": [ + 137, + 309, + 216, + 369 + ], + "lines": [ + { + "bbox": [ + 137, + 309, + 216, + 369 + ], + "spans": [ + { + "bbox": [ + 137, + 309, + 216, + 369 + ], + "type": "image", + "image_path": "0d31b9a9f2046581c21fcd7b55487da24655e421dd18554eaa002695c3724977.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 137, + 370, + 215, + 429 + ], + "blocks": [ + { + "bbox": [ + 137, + 370, + 215, + 429 + ], + "lines": [ + { + "bbox": [ + 137, + 370, + 215, + 429 + ], + "spans": [ + { + "bbox": [ + 137, + 370, + 215, + 429 + ], + "type": "image", + "image_path": "12de9452394cc1cdc540daee145d15972b4db943377d7993e27a777c2c3402b3.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 157, + 430, + 194, + 437 + ], + "lines": [ + { + "bbox": [ + 157, + 430, + 194, + 437 + ], + "spans": [ + { + "bbox": [ + 157, + 430, + 194, + 437 + ], + "type": "text", + "content": "Inference 1" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 219, + 309, + 297, + 369 + ], + "blocks": [ + { + "bbox": [ + 219, + 309, + 297, + 369 + ], + "lines": [ + { + "bbox": [ + 219, + 309, + 297, + 369 + ], + "spans": [ + { + "bbox": [ + 219, + 309, + 297, + 369 + ], + "type": "image", + "image_path": "7cea362638cc036d7a0c716a306bef32dd098ac75ece1fbcd79de2a5d2f5a33f.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 219, + 369, + 297, + 429 + ], + "blocks": [ + { + "bbox": [ + 219, + 369, + 297, + 429 + ], + "lines": [ + { + "bbox": [ + 219, + 369, + 297, + 429 + ], + "spans": [ + { + "bbox": [ + 219, + 369, + 297, + 429 + ], + "type": "image", + "image_path": "925b7724b83d034c415d74f9012ec24b1ed4f5a702b33d18c945a6aa46c860f7.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 239, + 430, + 277, + 437 + ], + "lines": [ + { + "bbox": [ + 239, + 430, + 277, + 437 + ], + "spans": [ + { + "bbox": [ + 239, + 430, + 277, + 437 + ], + "type": "text", + "content": "Inference 2" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 301, + 309, + 379, + 369 + ], + "blocks": [ + { + "bbox": [ + 301, + 309, + 379, + 369 + ], + "lines": [ + { + "bbox": [ + 301, + 309, + 379, + 369 + ], + "spans": [ + { + "bbox": [ + 301, + 309, + 379, + 369 + ], + "type": "image", + "image_path": "b99df597dd4df24c43683a6ea3beeed898412fba61220d04390636665d35fda9.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 301, + 370, + 379, + 429 + ], + "blocks": [ + { + "bbox": [ + 301, + 370, + 379, + 429 + ], + "lines": [ + { + "bbox": [ + 301, + 370, + 379, + 429 + ], + "spans": [ + { + "bbox": [ + 301, + 370, + 379, + 429 + ], + "type": "image", + "image_path": "8bb7696f43673e9f7661b9a5c4f3659c75fa815ee89d7c604d24d93b7e46b038.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 430, + 359, + 437 + ], + "lines": [ + { + "bbox": [ + 321, + 430, + 359, + 437 + ], + "spans": [ + { + "bbox": [ + 321, + 430, + 359, + 437 + ], + "type": "text", + "content": "Inference 3" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 449, + 545, + 471 + ], + "lines": [ + { + "bbox": [ + 46, + 449, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 545, + 471 + ], + "type": "text", + "content": "Figure 6. Colorization. Our method colorizes the gray-scale 3D scene consistently across views. By changing the edit code during inference, diverse colorized scenes can be rendered with satisfying photorealism and reasonably rich colors." + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 383, + 309, + 461, + 369 + ], + "blocks": [ + { + "bbox": [ + 383, + 309, + 461, + 369 + ], + "lines": [ + { + "bbox": [ + 383, + 309, + 461, + 369 + ], + "spans": [ + { + "bbox": [ + 383, + 309, + 461, + 369 + ], + "type": "image", + "image_path": "33f3a99405255176ebb939334391fd46cf6f69ce46e18f611595c6c978a2448a.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 383, + 369, + 461, + 429 + ], + "blocks": [ + { + "bbox": [ + 383, + 369, + 461, + 429 + ], + "lines": [ + { + "bbox": [ + 383, + 369, + 461, + 429 + ], + "spans": [ + { + "bbox": [ + 383, + 369, + 461, + 429 + ], + "type": "image", + "image_path": "6d1d848ca2fab9504e9191ad8fe3827df1a0e985deb1b96a43c7228875fe0f52.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 430, + 441, + 437 + ], + "lines": [ + { + "bbox": [ + 403, + 430, + 441, + 437 + ], + "spans": [ + { + "bbox": [ + 403, + 430, + 441, + 437 + ], + "type": "text", + "content": "Inference 4" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 465, + 309, + 543, + 369 + ], + "blocks": [ + { + "bbox": [ + 465, + 309, + 543, + 369 + ], + "lines": [ + { + "bbox": [ + 465, + 309, + 543, + 369 + ], + "spans": [ + { + "bbox": [ + 465, + 309, + 543, + 369 + ], + "type": "image", + "image_path": "03192cd3579b71dc28ba8cd7d112187da2cdf2576eba6c5ef41aef4ecfd09b2e.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 465, + 370, + 543, + 429 + ], + "blocks": [ + { + "bbox": [ + 465, + 370, + 543, + 429 + ], + "lines": [ + { + "bbox": [ + 465, + 370, + 543, + 429 + ], + "spans": [ + { + "bbox": [ + 465, + 370, + 543, + 429 + ], + "type": "image", + "image_path": "91f4e08eca197464bc3c0f30b58ad3525ad06b5bd92f2e83ec33615b442b4cd7.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 485, + 430, + 523, + 437 + ], + "lines": [ + { + "bbox": [ + 485, + 430, + 523, + 437 + ], + "spans": [ + { + "bbox": [ + 485, + 430, + 523, + 437 + ], + "type": "text", + "content": "Inference 5" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 41 + }, + { + "bbox": [ + 58, + 492, + 284, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 492, + 284, + 504 + ], + "spans": [ + { + "bbox": [ + 58, + 492, + 284, + 504 + ], + "type": "text", + "content": "Specifically, our contrastive loss is designed as follows:" + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 47, + 510, + 287, + 568 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 510, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 47, + 510, + 287, + 568 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {c o n t r}} = \\mathcal {L} _ {\\text {c o n t r}} ^ {\\text {a t t}} + \\mathcal {L} _ {\\text {c o n t r}} ^ {\\text {r e p}} \\\\ = \\sum_ {l = 0} ^ {N - 1} \\left\\| \\mathbf {z} _ {i} ^ {j} - \\mathbf {z} _ {l} ^ {j} \\right\\| _ {2} ^ {2} + \\sum_ {k = 0} ^ {M - 1} \\max (0, \\alpha - \\left\\| \\mathbf {z} _ {i} ^ {j} - \\mathbf {z} _ {i} ^ {k} \\right\\| _ {2} ^ {2}), \\tag {3} \\\\ \\end{array}", + "image_path": "ab6fc331862681a31f5a022246ccd2c9dd1c3ab3f3f2254dd1246af89170033d.jpg" + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 46, + 570, + 286, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 286, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 286, + 594 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 570, + 286, + 594 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 570, + 286, + 594 + ], + "type": "text", + "content": " represents the margin that encourages the difference in features, and " + }, + { + "bbox": [ + 46, + 570, + 286, + 594 + ], + "type": "inline_equation", + "content": "k\\neq j" + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 47, + 599, + 195, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 195, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 195, + 611 + ], + "type": "text", + "content": "3.2. NeRF-to-NeRF translation" + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 46, + 617, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 700 + ], + "type": "text", + "content": "Translated NeRF. After 2D image-to-image translation, we need to lift these 2D edits to the 3D NeRF. For this purpose, we propose to modify the original NeRF as a translated NeRF that takes the edit code " + }, + { + "bbox": [ + 46, + 617, + 287, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 46, + 617, + 287, + 700 + ], + "type": "text", + "content": " as input and generates the translated 3D scene according to the edit code. We refer readers to the supplementary for more details about the network architecture." + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "content": "Reconstruction loss. Given an edit code " + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i^j" + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "content": " extracted from" + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 304, + 491, + 545, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 491, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 304, + 491, + 545, + 554 + ], + "type": "text", + "content": "the edited image " + }, + { + "bbox": [ + 304, + 491, + 545, + 554 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_i^j" + }, + { + "bbox": [ + 304, + 491, + 545, + 554 + ], + "type": "text", + "content": ", we can generate a translated NeRF to render " + }, + { + "bbox": [ + 304, + 491, + 545, + 554 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_i^j" + }, + { + "bbox": [ + 304, + 491, + 545, + 554 + ], + "type": "text", + "content": " from the same viewpoint. Then we define the reconstruction loss as the L1 normalization and Learned Perceptual Image Patch Similarity (LPIPS) [44] between the rendered image " + }, + { + "bbox": [ + 304, + 491, + 545, + 554 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_i^j" + }, + { + "bbox": [ + 304, + 491, + 545, + 554 + ], + "type": "text", + "content": " and the edited image " + }, + { + "bbox": [ + 304, + 491, + 545, + 554 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_i^j" + }, + { + "bbox": [ + 304, + 491, + 545, + 554 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 315, + 562, + 545, + 598 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 562, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 315, + 562, + 545, + 598 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {r e c o n}} = \\mathcal {L} _ {\\mathrm {L} 1} + \\mathcal {L} _ {\\mathrm {L P I P S}} \\\\ = \\left\\| \\mathbf {C} _ {i} ^ {j} - \\mathbf {S} _ {i} ^ {j} \\right\\| _ {1} + L P I P S [ \\mathcal {P} (\\mathbf {C} _ {i} ^ {j}) - \\mathcal {P} (\\mathbf {S} _ {i} ^ {j}) ], \\tag {4} \\\\ \\end{array}", + "image_path": "9ca53dc03a16fef08a7830d3545b1ec4b1c66c98b515c41b81a11dac5ba90b6f.jpg" + } + ] + } + ], + "index": 51 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " means a patch sampled from the image. Note that due to the lack of 3D consistency of the edited multiview image, the supervision of the edited image from other viewpoints " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{S}_l^j\\}_{l\\neq i}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " will lead to conflicts in pixel-space optimization. Therefore, we only employ reconstruction loss on the same view image " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_i^j" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " to optimize the translated NeRF. Adversarial loss. Since the 3D consistency of edited multiview images is not assured, relying solely on the reconstruction loss on the same view often leads to blurry or distorted" + } + ] + } + ], + "index": 52 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 217, + 70, + 393, + 81 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 70, + 393, + 81 + ], + "spans": [ + { + "bbox": [ + 217, + 70, + 393, + 81 + ], + "type": "text", + "content": "\"What would he look like as a bearded man\"" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5109" + } + ] + } + ], + "index": 53 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 126, + 123 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 126, + 123 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 126, + 123 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 126, + 123 + ], + "type": "image", + "image_path": "1903030e864e1d3f55178a367dcc206654f9470b734b5ec7c6524603a9755c09.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 123, + 110, + 131 + ], + "lines": [ + { + "bbox": [ + 64, + 123, + 110, + 131 + ], + "spans": [ + { + "bbox": [ + 64, + 123, + 110, + 131 + ], + "type": "text", + "content": "Original NeRF" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 129, + 71, + 205, + 122 + ], + "blocks": [ + { + "bbox": [ + 129, + 71, + 205, + 122 + ], + "lines": [ + { + "bbox": [ + 129, + 71, + 205, + 122 + ], + "spans": [ + { + "bbox": [ + 129, + 71, + 205, + 122 + ], + "type": "image", + "image_path": "b31dda698362d8b9ae293fdc2235c9086df7d7269612578a6c8b0de3388462fa.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 123, + 203, + 130 + ], + "lines": [ + { + "bbox": [ + 131, + 123, + 203, + 130 + ], + "spans": [ + { + "bbox": [ + 131, + 123, + 203, + 130 + ], + "type": "text", + "content": "Instruct-NeRF2NeRF 1" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 209, + 72, + 285, + 122 + ], + "blocks": [ + { + "bbox": [ + 209, + 72, + 285, + 122 + ], + "lines": [ + { + "bbox": [ + 209, + 72, + 285, + 122 + ], + "spans": [ + { + "bbox": [ + 209, + 72, + 285, + 122 + ], + "type": "image", + "image_path": "a5a0c06a3876536af5a979b53b435f76c3920d4cc176f4cd355c3d25c723cc22.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 211, + 123, + 283, + 130 + ], + "lines": [ + { + "bbox": [ + 211, + 123, + 283, + 130 + ], + "spans": [ + { + "bbox": [ + 211, + 123, + 283, + 130 + ], + "type": "text", + "content": "Instruct-NeRF2NeRF 2" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 50, + 132, + 126, + 182 + ], + "blocks": [ + { + "bbox": [ + 50, + 132, + 126, + 182 + ], + "lines": [ + { + "bbox": [ + 50, + 132, + 126, + 182 + ], + "spans": [ + { + "bbox": [ + 50, + 132, + 126, + 182 + ], + "type": "image", + "image_path": "df92d6df55aa709bdda7be0e168885a51760fbd267ca4524d1425ae1c856f04d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 183, + 113, + 191 + ], + "lines": [ + { + "bbox": [ + 61, + 183, + 113, + 191 + ], + "spans": [ + { + "bbox": [ + 61, + 183, + 113, + 191 + ], + "type": "text", + "content": "Ours inference 1" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 200, + 287, + 234 + ], + "lines": [ + { + "bbox": [ + 46, + 200, + 287, + 234 + ], + "spans": [ + { + "bbox": [ + 46, + 200, + 287, + 234 + ], + "type": "text", + "content": "Figure 7. Comparisons with baselines of text-driven NeRF editing. We compare our method with Instruct-NeRF2NeRF [9] in the editing by using the text prompt \"Make it Spring\"." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 129, + 132, + 206, + 182 + ], + "blocks": [ + { + "bbox": [ + 129, + 132, + 206, + 182 + ], + "lines": [ + { + "bbox": [ + 129, + 132, + 206, + 182 + ], + "spans": [ + { + "bbox": [ + 129, + 132, + 206, + 182 + ], + "type": "image", + "image_path": "86bfc0b53e1dff3aeab3555a77b800c0c79363fc58056e5633805bd7fdb56bf1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 141, + 183, + 193, + 191 + ], + "lines": [ + { + "bbox": [ + 141, + 183, + 193, + 191 + ], + "spans": [ + { + "bbox": [ + 141, + 183, + 193, + 191 + ], + "type": "text", + "content": "Ours inference 2" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 209, + 132, + 285, + 182 + ], + "blocks": [ + { + "bbox": [ + 209, + 132, + 285, + 182 + ], + "lines": [ + { + "bbox": [ + 209, + 132, + 285, + 182 + ], + "spans": [ + { + "bbox": [ + 209, + 132, + 285, + 182 + ], + "type": "image", + "image_path": "d9b07c9875a004e5c9656be5ba738e6b5355b277fa061b07a93f895e9c67dcec.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 183, + 272, + 191 + ], + "lines": [ + { + "bbox": [ + 220, + 183, + 272, + 191 + ], + "spans": [ + { + "bbox": [ + 220, + 183, + 272, + 191 + ], + "type": "text", + "content": "Ours inference 3" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 50, + 243, + 296, + 323 + ], + "blocks": [ + { + "bbox": [ + 50, + 243, + 296, + 323 + ], + "lines": [ + { + "bbox": [ + 50, + 243, + 296, + 323 + ], + "spans": [ + { + "bbox": [ + 50, + 243, + 296, + 323 + ], + "type": "table", + "html": "
MethodCLIP Text-Image Direction Similarity↑CLIP Direction Consistency ↑FID ↓
InstructPix2Pix [2]+NeRF0.16690.8475270.542
Instruct-NeRF2NeRF0.20210.9828148.021
Ours w/o Ladv0.19200.9657162.275
Ours w/o Lcontr0.20070.9749156.524
Ours0.20890.9864137.740
", + "image_path": "36074e7b16c6e6d15085d87da927d74deef179f06e72f4172939fd00dbe646a5.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 396, + 287, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 287, + 491 + ], + "type": "text", + "content": "artifacts on novel views. Previous research demonstrates the effectiveness of conditional adversarial training in preventing the production of blurry rendered images resulting from conflicts that arise from noise in the camera extrinsic when performing image supervision from different viewpoints [11]. The function of the condition is to guide discriminator with fine-grained information from the same viewpoint, thus preventing GAN mode collapse." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 491, + 287, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 491, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 491, + 287, + 647 + ], + "type": "text", + "content": "It inspires us to incorporate conditional adversarial loss on rendered images from the translated NeRF, which is conducive to distinguish artifacts in rendered images. As illustrated in Fig.4, the discriminator " + }, + { + "bbox": [ + 46, + 491, + 287, + 647 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 46, + 491, + 287, + 647 + ], + "type": "text", + "content": " takes into real pairs and fake pairs. Each real pair " + }, + { + "bbox": [ + 46, + 491, + 287, + 647 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 46, + 491, + 287, + 647 + ], + "type": "text", + "content": " consists of " + }, + { + "bbox": [ + 46, + 491, + 287, + 647 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^j" + }, + { + "bbox": [ + 46, + 491, + 287, + 647 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 491, + 287, + 647 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^j -\\mathbf{S}^k" + }, + { + "bbox": [ + 46, + 491, + 287, + 647 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 491, + 287, + 647 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^j\\in \\{\\mathbf{S}_i^j\\}_{i = 0}^{N - 1}" + }, + { + "bbox": [ + 46, + 491, + 287, + 647 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 491, + 287, + 647 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^k\\in \\{\\mathbf{S}_i^k\\}_{i = 0}^{N - 1}" + }, + { + "bbox": [ + 46, + 491, + 287, + 647 + ], + "type": "text", + "content": " are from two sets of edited images from the image translation. Similarly, each fake pair " + }, + { + "bbox": [ + 46, + 491, + 287, + 647 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 46, + 491, + 287, + 647 + ], + "type": "text", + "content": " consists of " + }, + { + "bbox": [ + 46, + 491, + 287, + 647 + ], + "type": "inline_equation", + "content": "\\mathbf{C}^j" + }, + { + "bbox": [ + 46, + 491, + 287, + 647 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 491, + 287, + 647 + ], + "type": "inline_equation", + "content": "\\mathbf{C}^j -\\mathbf{S}^k" + }, + { + "bbox": [ + 46, + 491, + 287, + 647 + ], + "type": "text", + "content": " in which " + }, + { + "bbox": [ + 46, + 491, + 287, + 647 + ], + "type": "inline_equation", + "content": "\\mathbf{C}^j\\in \\{\\mathbf{C}_i^j\\}_{i = 0}^{N - 1}" + }, + { + "bbox": [ + 46, + 491, + 287, + 647 + ], + "type": "text", + "content": " is generated by translated NeRF. Note that the images in the same pair come from the same viewpoint. The pairs are concatenated in RGB channels and fed into the discriminator. We optimize the discriminator " + }, + { + "bbox": [ + 46, + 491, + 287, + 647 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 46, + 491, + 287, + 647 + ], + "type": "text", + "content": " and translated NeRF with the objective functions below:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 52, + 654, + 287, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 654, + 287, + 673 + ], + "spans": [ + { + "bbox": [ + 52, + 654, + 287, + 673 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {A D - D}} = \\mathbb {E} _ {\\mathbf {R}} [ - l o g (\\mathbf {D} (\\mathbf {R})) ] + \\mathbb {E} _ {\\mathbf {F}} [ - l o g (1 - \\mathbf {D} (\\mathbf {F})) ], \\tag {5}", + "image_path": "c906d9b87c456d02fb671ca35418d3c127e4281d21c7f0fee659fc1a019be580.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 53, + 670, + 167, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 670, + 167, + 682 + ], + "spans": [ + { + "bbox": [ + 53, + 670, + 167, + 682 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {A D - G}} = \\mathbb {E} _ {\\mathbf {F}} [ - l o g (\\mathbf {D} (\\mathbf {F})) ].", + "image_path": "3b0aaf68d32cf9425a1d081cc105ac251abd41284941421e2aa83e818048a26b.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "Optimization. During the training process, we jointly optimize the loss functions mentioned above: " + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{KL}}" + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{contr}}" + } + ] + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 384, + 126 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 384, + 126 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 384, + 126 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 384, + 126 + ], + "type": "image", + "image_path": "8b3abbdfb1cc401f78083ee3e9d504f25ca9dfd926e18de2204502d389ab78ca.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 127, + 365, + 135 + ], + "lines": [ + { + "bbox": [ + 320, + 127, + 365, + 135 + ], + "spans": [ + { + "bbox": [ + 320, + 127, + 365, + 135 + ], + "type": "text", + "content": "Original NeRF" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 388, + 72, + 463, + 127 + ], + "blocks": [ + { + "bbox": [ + 388, + 72, + 463, + 127 + ], + "lines": [ + { + "bbox": [ + 388, + 72, + 463, + 127 + ], + "spans": [ + { + "bbox": [ + 388, + 72, + 463, + 127 + ], + "type": "image", + "image_path": "84cc93b7444cb4f0729abcbc0ec4d39f361680c551f7427720cf6c1197de227b.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 393, + 127, + 459, + 134 + ], + "lines": [ + { + "bbox": [ + 393, + 127, + 459, + 134 + ], + "spans": [ + { + "bbox": [ + 393, + 127, + 459, + 134 + ], + "type": "text", + "content": "Instruct-NeRF2NeRF" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 468, + 72, + 543, + 127 + ], + "blocks": [ + { + "bbox": [ + 468, + 72, + 543, + 127 + ], + "lines": [ + { + "bbox": [ + 468, + 72, + 543, + 127 + ], + "spans": [ + { + "bbox": [ + 468, + 72, + 543, + 127 + ], + "type": "image", + "image_path": "dda52412dbaf0acaba32f1cfbfb5f51e7974b02aa213a2afd0feb29121574ff7.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 486, + 127, + 525, + 134 + ], + "lines": [ + { + "bbox": [ + 486, + 127, + 525, + 134 + ], + "spans": [ + { + "bbox": [ + 486, + 127, + 525, + 134 + ], + "type": "text", + "content": "PaletteNeRF" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 308, + 136, + 384, + 192 + ], + "blocks": [ + { + "bbox": [ + 308, + 136, + 384, + 192 + ], + "lines": [ + { + "bbox": [ + 308, + 136, + 384, + 192 + ], + "spans": [ + { + "bbox": [ + 308, + 136, + 384, + 192 + ], + "type": "image", + "image_path": "159f0eda688041349c873d7f76f4316f35000945f3405e18f79ef75d96b19797.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 192, + 371, + 199 + ], + "lines": [ + { + "bbox": [ + 320, + 192, + 371, + 199 + ], + "spans": [ + { + "bbox": [ + 320, + 192, + 371, + 199 + ], + "type": "text", + "content": "Ours inference 1" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 209, + 545, + 231 + ], + "lines": [ + { + "bbox": [ + 306, + 209, + 545, + 231 + ], + "spans": [ + { + "bbox": [ + 306, + 209, + 545, + 231 + ], + "type": "text", + "content": "Figure 8. Comparisons with baselines of NeRF colorization. We compare with PaletteNeRF[16] in colorization." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 388, + 136, + 463, + 192 + ], + "blocks": [ + { + "bbox": [ + 388, + 136, + 463, + 192 + ], + "lines": [ + { + "bbox": [ + 388, + 136, + 463, + 192 + ], + "spans": [ + { + "bbox": [ + 388, + 136, + 463, + 192 + ], + "type": "image", + "image_path": "ccc9308bb2cf789858b3dc4ffd8801748154ebf4b74d9d5d6879070c1c67ca90.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 400, + 192, + 451, + 199 + ], + "lines": [ + { + "bbox": [ + 400, + 192, + 451, + 199 + ], + "spans": [ + { + "bbox": [ + 400, + 192, + 451, + 199 + ], + "type": "text", + "content": "Ours inference 2" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 468, + 136, + 543, + 192 + ], + "blocks": [ + { + "bbox": [ + 468, + 136, + 543, + 192 + ], + "lines": [ + { + "bbox": [ + 468, + 136, + 543, + 192 + ], + "spans": [ + { + "bbox": [ + 468, + 136, + 543, + 192 + ], + "type": "image", + "image_path": "e2e1951d4fae3b38aff5be22e2642dcb7694a7fc7c3c92b557bd8a70d598b963.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 480, + 192, + 530, + 199 + ], + "lines": [ + { + "bbox": [ + 480, + 192, + 530, + 199 + ], + "spans": [ + { + "bbox": [ + 480, + 192, + 530, + 199 + ], + "type": "text", + "content": "Ours inference 3" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "table", + "bbox": [ + 309, + 240, + 555, + 335 + ], + "blocks": [ + { + "bbox": [ + 46, + 331, + 287, + 376 + ], + "lines": [ + { + "bbox": [ + 46, + 331, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 287, + 376 + ], + "type": "text", + "content": "Table 1. Quantitative results on text-driven editing. We compare our method with the naive method of directly combining InstructPix2Pix [2] with NeRF and the state-of-the-art method Instruct-NeRF2NeRF [9]." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 240, + 555, + 335 + ], + "lines": [ + { + "bbox": [ + 309, + 240, + 555, + 335 + ], + "spans": [ + { + "bbox": [ + 309, + 240, + 555, + 335 + ], + "type": "table", + "html": "
MethodCF ↑FID ↓
DDColor [14]+NeRF40.435148.957
Instruct-NeRF2NeRF45.599201.456
PaletteNeRF [16]39.654-
Ours w/o Ladv35.031137.740
Ours w/o Lcontr34.829105.750
Ours65.09935.041
", + "image_path": "d87e60f944ea68b8818327176abb4eafa830d090d0552b1905650a13a957d7f9.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "table_body" + } + ], + "index": 33 + }, + { + "bbox": [ + 305, + 342, + 545, + 398 + ], + "lines": [ + { + "bbox": [ + 305, + 342, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 305, + 342, + 545, + 398 + ], + "type": "text", + "content": "Table 2. Quantitative results on colorization. We colorize images with the translator, DDcolor [14], and edit NeRF by directly optimizing NeRF, Instruct-NeRF2NeRF [9] and our NeRF translation method. The quantitative comparison is conducted between these methods as well as PaletteNeRF [16]." + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 417, + 545, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 417, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 305, + 417, + 545, + 453 + ], + "type": "text", + "content": "for the edit code, " + }, + { + "bbox": [ + 305, + 417, + 545, + 453 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{recon}}" + }, + { + "bbox": [ + 305, + 417, + 545, + 453 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 417, + 545, + 453 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{AD - G}}" + }, + { + "bbox": [ + 305, + 417, + 545, + 453 + ], + "type": "text", + "content": " for the translated NeRF, and " + }, + { + "bbox": [ + 305, + 417, + 545, + 453 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{AD - D}}" + }, + { + "bbox": [ + 305, + 417, + 545, + 453 + ], + "type": "text", + "content": " for the discriminator. The total loss formula is expressed as follows:" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 332, + 460, + 545, + 473 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 460, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 332, + 460, + 545, + 473 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\mathrm {K L}} + \\mathcal {L} _ {\\text {r e c o n}} + \\mathcal {L} _ {\\mathrm {A D - G}} + \\mathcal {L} _ {\\mathrm {A D - D}} + \\mathcal {L} _ {\\text {c o n t r}}. \\tag {6}", + "image_path": "b678af2456ae07301bbbdaa62f8e2e0505a9bb93052ed468ace4fff4f20e2b7e.jpg" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 305, + 479, + 545, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 479, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 305, + 479, + 545, + 540 + ], + "type": "text", + "content": "where we assign each regularization term the weight of 1.0, 1.0, 0.1, 0.1, 0.1 in all of our experiments. The weights can be adjusted to prioritize different aspects of the training objective, such as reconstruction accuracy, adversarial training, and perceptual quality." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 306, + 546, + 372, + 557 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 546, + 372, + 557 + ], + "spans": [ + { + "bbox": [ + 306, + 546, + 372, + 557 + ], + "type": "text", + "content": "3.3. Inference" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 304, + 563, + 545, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 563, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 563, + 545, + 635 + ], + "type": "text", + "content": "After the optimization of our GenN2N, the translated NeRF is optimized to be able to render the target scene conditioned on the edit code. As shown in Fig. 1, users can simply sample an edit code from the Gaussian distribution and use the translated NeRF to render the 3D scene with high-quality and multi-view 3D consistency." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 306, + 646, + 386, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 646, + 386, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 646, + 386, + 658 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Our proposed GenN2N is a unified NeRF-to-NeRF translation framework which can support various NeRF editing tasks. In this paper, we demonstrate the effectiveness of GenN2N by a suite of challenging tasks:" + } + ] + } + ], + "index": 41 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5110" + } + ] + } + ], + "index": 42 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 108, + 115 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 108, + 115 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 108, + 115 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 108, + 115 + ], + "type": "image", + "image_path": "ea7a30b00b68fa9411a593ace664d3944b1ec609ad2b74c06b46571d549da071.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 108, + 72, + 166, + 116 + ], + "blocks": [ + { + "bbox": [ + 87, + 116, + 128, + 122 + ], + "lines": [ + { + "bbox": [ + 87, + 116, + 128, + 122 + ], + "spans": [ + { + "bbox": [ + 87, + 116, + 128, + 122 + ], + "type": "text", + "content": "Original NeRF" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 108, + 72, + 166, + 116 + ], + "lines": [ + { + "bbox": [ + 108, + 72, + 166, + 116 + ], + "spans": [ + { + "bbox": [ + 108, + 72, + 166, + 116 + ], + "type": "image", + "image_path": "ea07cdbf5005153b3fefbaf6cbbc4ab178f07a04305b7d8b8fc23560a8262968.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 169, + 72, + 227, + 116 + ], + "blocks": [ + { + "bbox": [ + 169, + 72, + 227, + 116 + ], + "lines": [ + { + "bbox": [ + 169, + 72, + 227, + 116 + ], + "spans": [ + { + "bbox": [ + 169, + 72, + 227, + 116 + ], + "type": "image", + "image_path": "fc446adaa1a09002755507642964878b5a058cab2d5b9896f16222960353f499.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 197, + 116, + 256, + 122 + ], + "lines": [ + { + "bbox": [ + 197, + 116, + 256, + 122 + ], + "spans": [ + { + "bbox": [ + 197, + 116, + 256, + 122 + ], + "type": "text", + "content": "Instruct-NeRF2NeRF" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 228, + 72, + 285, + 116 + ], + "blocks": [ + { + "bbox": [ + 228, + 72, + 285, + 116 + ], + "lines": [ + { + "bbox": [ + 228, + 72, + 285, + 116 + ], + "spans": [ + { + "bbox": [ + 228, + 72, + 285, + 116 + ], + "type": "image", + "image_path": "8c734d8649c9850ae52c36c4655967ffdb6759f5ec469f76b577323ab321ed99.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 50, + 123, + 107, + 167 + ], + "blocks": [ + { + "bbox": [ + 50, + 123, + 107, + 167 + ], + "lines": [ + { + "bbox": [ + 50, + 123, + 107, + 167 + ], + "spans": [ + { + "bbox": [ + 50, + 123, + 107, + 167 + ], + "type": "image", + "image_path": "f33b2e818b8f36d411da9f6cbd9a7dfd67cf7f89580d02ff866ba9ce7f9adbf5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 93, + 167, + 121, + 175 + ], + "lines": [ + { + "bbox": [ + 93, + 167, + 121, + 175 + ], + "spans": [ + { + "bbox": [ + 93, + 167, + 121, + 175 + ], + "type": "text", + "content": "NeRF-SR" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 183, + 287, + 206 + ], + "lines": [ + { + "bbox": [ + 47, + 183, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 47, + 183, + 287, + 206 + ], + "type": "text", + "content": "Figure 9. Comparisons with baselines of NeRF super-resolution. We compare with NeRF-SR [34] in super-resolution." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 108, + 123, + 166, + 167 + ], + "blocks": [ + { + "bbox": [ + 108, + 123, + 166, + 167 + ], + "lines": [ + { + "bbox": [ + 108, + 123, + 166, + 167 + ], + "spans": [ + { + "bbox": [ + 108, + 123, + 166, + 167 + ], + "type": "image", + "image_path": "79f4eb3b5408230337924aa9c63495fd66cdd5254acd6196aaee3ad6c651e027.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 169, + 123, + 227, + 167 + ], + "blocks": [ + { + "bbox": [ + 169, + 123, + 227, + 167 + ], + "lines": [ + { + "bbox": [ + 169, + 123, + 227, + 167 + ], + "spans": [ + { + "bbox": [ + 169, + 123, + 227, + 167 + ], + "type": "image", + "image_path": "5cbb8ff671d1809fabcc856eb267883b8cc5bceb9655e76da1c52bbc7511d1a7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 219, + 168, + 234, + 175 + ], + "lines": [ + { + "bbox": [ + 219, + 168, + 234, + 175 + ], + "spans": [ + { + "bbox": [ + 219, + 168, + 234, + 175 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 228, + 123, + 285, + 167 + ], + "blocks": [ + { + "bbox": [ + 228, + 123, + 285, + 167 + ], + "lines": [ + { + "bbox": [ + 228, + 123, + 285, + 167 + ], + "spans": [ + { + "bbox": [ + 228, + 123, + 285, + 167 + ], + "type": "image", + "image_path": "14fa18a5859480ae72322936d6e46c5f13087490b5647c2789da03ad0b9ea568.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 50, + 217, + 296, + 290 + ], + "blocks": [ + { + "bbox": [ + 50, + 217, + 296, + 290 + ], + "lines": [ + { + "bbox": [ + 50, + 217, + 296, + 290 + ], + "spans": [ + { + "bbox": [ + 50, + 217, + 296, + 290 + ], + "type": "table", + "html": "
MethodPSNR ↑SSIM ↑LPIPS ↓
ResShift [40]+NeRF19.9780.5350.1156
Instruct-NeRF2NeRF20.2990.6420.2732
NeRF-SR [34]27.9570.8970.0937
Ours w/o Ladv12.5550.6630.2001
Ours w/o Lcontr15.3720.6620.1834
Ours28.5010.9130.0748
", + "image_path": "ffb5aefc78086ae3de55974ed7d2c874d0bb8dcbb7599b8160bc4d4171cf8a3e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 299, + 287, + 354 + ], + "lines": [ + { + "bbox": [ + 46, + 299, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 299, + 287, + 354 + ], + "type": "text", + "content": "Table 3. Quantitative results on super-resolution. We improve image resolution with ResShift [40] and edit NeRF by directly optimizing NeRF, Instruct-NeRF2NeRF [9] and our NeRF translation method. The quantitative comparison is conducted between these methods as well as NeRF-SR [34]." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 50, + 375, + 287, + 470 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 50, + 375, + 287, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 375, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 50, + 375, + 287, + 399 + ], + "type": "text", + "content": "(1) Text-driven Editing edits the given NeRF scene to a set of NeRF scenes according to the text instruction." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 50, + 400, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 400, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 50, + 400, + 287, + 422 + ], + "type": "text", + "content": "(2) Colorization transforms a gray-scale NeRF scene to a set of plausible color NeRF scenes." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 50, + 423, + 287, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 423, + 287, + 447 + ], + "spans": [ + { + "bbox": [ + 50, + 423, + 287, + 447 + ], + "type": "text", + "content": "(3) Super-resolution enhances the resolution of NeRF and enables multiple plausible outcomes." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 50, + 448, + 287, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 448, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 50, + 448, + 287, + 470 + ], + "type": "text", + "content": "(4) Inpainting fills in user-specified masked regions in the NeRF scene with realistic content." + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 472, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 628 + ], + "type": "text", + "content": "We achieve those tasks by simply changing the plug-and-play 2D image translator in our framework, without any additional task-specific design. Previous studies have extensively explored some of these issues like text-driven editing, colorization, super-resolution, and inpainting. However, there is rarely a unified framework that can achieve all these tasks with strong performance, high quality, and plausible multi-view consistent 3D structure. Furthermore, GenN2N can also perform zooming out and text-driven inpainting in NeRF-to-NeRF translation, which were not explored in prior research. We refer readers to the supplementary materials for detailed experiment settings, dataset settings and implementation details." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 47, + 635, + 132, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 132, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 132, + 647 + ], + "type": "text", + "content": "4.1. Comparisons" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "Text-driven Editing. We achieve text-driven editing of the given NeRF by using InstructPix2Pix [2] as the 2D image-to-image translator in our framework. We compared our approach to a naive solution, which involves optimizing a NeRF with edited images via InstructPix2Pix. However," + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": "this naive approach leads to a 3D inconsistency problem among different edits. While Instruct-NeRF2NeRF [9] proposed an iterative updating mechanism to address this issue, it falls short in capturing the diversity of different edits, making it challenging to ensure the quality of the outcomes. To evaluate our method, we conducted experiments on the Face [9] and Fangzhou [35] self-portrait datasets, and Farm [9] dataset, comparing GenN2N with the state-of-the-art NeRF editing method Instruct-NeRF2NeRF [9]." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 182, + 546, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 182, + 546, + 398 + ], + "spans": [ + { + "bbox": [ + 304, + 182, + 546, + 398 + ], + "type": "text", + "content": "Quantitative results are presented in Table 1, where we employed CLIP Text-Image Direction Similarity [9], CLIP Direction Consistency [9], and Fréchet Inception Distance (FID) [10] as evaluation metrics. The results highlight the superior performance of GenN2N over other methods, demonstrating its effectiveness in producing high-quality 3D text-driven editing results. Furthermore, we provide a qualitative comparison between GenN2N and Instruct-NeRF2NeRF [9] in Figure 7. Notably, Instruct-NeRF2NeRF was trained twice, but the results are nearly identical. In contrast, GenN2N can render edited scenes with various effects, consistent with the input text instruction, by inferring with different edit codes sampled from a Gaussian distribution. To further illustrate the generative capability of GenN2N, we showcase additional results of our inference renderings in Figure 5, demonstrating its diverse generative ability in terms of appearance and geometry." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 399, + 546, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 546, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 546, + 555 + ], + "type": "text", + "content": "Colorization. For NeRF colorization, GenN2N uses DD-Color [14] as the 2D image-to-image translator. CoRF [8] and Palette-NeRF [16] do a similar task and we compare with Palette-NeRF in Table 2. The Fréchet Inception Distance (FID) [10] and colorfulness score (CF) [17] are used to measure the distribution similarity and vividness of generated images. We show visual comparison results in Fig. 6 and Fig. 8. We can find that with different edit codes, the scene can be rendered in different color styles. It is noticeable that with the same edit code, the color rendered from different views is consistent. This strongly demonstrates the effectiveness of our method in translating NeRF while keeping the 3D consistency of the scene." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": "Super-resolution. When only low-resolution images are available, our methods can boost NeRF in reconstructing scenes at higher resolution, while keeping view consistency and avoiding blurry outputs. We achieve this by employing ResShift [40] as the 2D image-to-image translator in GenN2N. Following state-of-the-art method NeRF-SR [34], we conduct experiments on " + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "inline_equation", + "content": "LLFF" + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": " dataset [23], using PSNR, SSIM, and LPIPS as evaluation metrics. As shown in Table 3, GenN2N obtains NeRF-to-NeRF translation with higher performance than NeRF-SR [34]. Moreover, we also provide qualitative comparison results in Fig. 9, where GenN2N produces clearer and more realistic rendering results than previous methods." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "5111" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 107, + 116 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 107, + 116 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 107, + 116 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 107, + 116 + ], + "type": "image", + "image_path": "2d395f82bc64109b08b3631b093c0feaadb2158aa8a700af2aa2a6ccfdbf0e1c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 107, + 72, + 166, + 116 + ], + "blocks": [ + { + "bbox": [ + 107, + 72, + 166, + 116 + ], + "lines": [ + { + "bbox": [ + 107, + 72, + 166, + 116 + ], + "spans": [ + { + "bbox": [ + 107, + 72, + 166, + 116 + ], + "type": "image", + "image_path": "0e9e79b1be9ac459c1c1ae539d5a4d29ca56248b322a6e3a99dc13ae8cae17be.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 168, + 72, + 227, + 116 + ], + "blocks": [ + { + "bbox": [ + 168, + 72, + 227, + 116 + ], + "lines": [ + { + "bbox": [ + 168, + 72, + 227, + 116 + ], + "spans": [ + { + "bbox": [ + 168, + 72, + 227, + 116 + ], + "type": "image", + "image_path": "03915de0884441efb6c533790dc10dd5779d3f3d7b1f48e254f6fc42dad1d4d1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 116, + 261, + 124 + ], + "lines": [ + { + "bbox": [ + 192, + 116, + 261, + 124 + ], + "spans": [ + { + "bbox": [ + 192, + 116, + 261, + 124 + ], + "type": "text", + "content": "Instruct-NeRF2NeRF" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 227, + 72, + 285, + 116 + ], + "blocks": [ + { + "bbox": [ + 227, + 72, + 285, + 116 + ], + "lines": [ + { + "bbox": [ + 227, + 72, + 285, + 116 + ], + "spans": [ + { + "bbox": [ + 227, + 72, + 285, + 116 + ], + "type": "image", + "image_path": "78479ea566ffe1847a74ba82b42f0b4ba8494975fec2cf0053a6a588f567fb1b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 50, + 126, + 107, + 168 + ], + "blocks": [ + { + "bbox": [ + 50, + 126, + 107, + 168 + ], + "lines": [ + { + "bbox": [ + 50, + 126, + 107, + 168 + ], + "spans": [ + { + "bbox": [ + 50, + 126, + 107, + 168 + ], + "type": "image", + "image_path": "9dff90df2b08ff280aea834c0947a49fa62b04f3625e1716db0dbda811e74005.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 107, + 125, + 166, + 170 + ], + "blocks": [ + { + "bbox": [ + 83, + 117, + 132, + 125 + ], + "lines": [ + { + "bbox": [ + 83, + 117, + 132, + 125 + ], + "spans": [ + { + "bbox": [ + 83, + 117, + 132, + 125 + ], + "type": "text", + "content": "Original NeRF" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 107, + 125, + 166, + 170 + ], + "lines": [ + { + "bbox": [ + 107, + 125, + 166, + 170 + ], + "spans": [ + { + "bbox": [ + 107, + 125, + 166, + 170 + ], + "type": "image", + "image_path": "484cba32e0cd42a5a3d692f9500a6699281318e62d26b4afa219ff24723d5e29.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 89, + 170, + 127, + 178 + ], + "lines": [ + { + "bbox": [ + 89, + 170, + 127, + 178 + ], + "spans": [ + { + "bbox": [ + 89, + 170, + 127, + 178 + ], + "type": "text", + "content": "SPin-NeRF" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 186, + 287, + 209 + ], + "lines": [ + { + "bbox": [ + 47, + 186, + 287, + 209 + ], + "spans": [ + { + "bbox": [ + 47, + 186, + 287, + 209 + ], + "type": "text", + "content": "Figure 10. Comparisons with baselines of NeRF inpainting. We compare with SPIn-NeRF [24] in inpainting." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 168, + 125, + 227, + 170 + ], + "blocks": [ + { + "bbox": [ + 168, + 125, + 227, + 170 + ], + "lines": [ + { + "bbox": [ + 168, + 125, + 227, + 170 + ], + "spans": [ + { + "bbox": [ + 168, + 125, + 227, + 170 + ], + "type": "image", + "image_path": "04848002808598c140cd62eb2f3e2486767f21657c5c122071169ba1305fbd8b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 217, + 170, + 235, + 178 + ], + "lines": [ + { + "bbox": [ + 217, + 170, + 235, + 178 + ], + "spans": [ + { + "bbox": [ + 217, + 170, + 235, + 178 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 227, + 125, + 285, + 170 + ], + "blocks": [ + { + "bbox": [ + 227, + 125, + 285, + 170 + ], + "lines": [ + { + "bbox": [ + 227, + 125, + 285, + 170 + ], + "spans": [ + { + "bbox": [ + 227, + 125, + 285, + 170 + ], + "type": "image", + "image_path": "7ed027a0cbd9386e6dfe4cb4901dd88152219e7d5dc64eea4761285f4f21f6c8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 50, + 219, + 296, + 277 + ], + "blocks": [ + { + "bbox": [ + 50, + 219, + 296, + 277 + ], + "lines": [ + { + "bbox": [ + 50, + 219, + 296, + 277 + ], + "spans": [ + { + "bbox": [ + 50, + 219, + 296, + 277 + ], + "type": "table", + "html": "
MethodPSNR ↑SSIM ↑LPIPS ↓
LaMa [31]+NeRF18.9830.37060.1730
Instruct-NeRF2NeRF16.7340.30880.2750
SPin-NeRF [24]24.3690.72170.1754
Ours26.8680.81370.1284
", + "image_path": "3b5fb1001a93f589467568ab549c13d6f0cc6e4d5ed767d7b1cfc1e9f8a90d24.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 72, + 285, + 262, + 297 + ], + "lines": [ + { + "bbox": [ + 72, + 285, + 262, + 297 + ], + "spans": [ + { + "bbox": [ + 72, + 285, + 262, + 297 + ], + "type": "text", + "content": "Table 4. Quantitative results on NeRF inpainting." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 316, + 289, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 289, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 289, + 521 + ], + "type": "text", + "content": "Inpainting. The goal of NeRF Inpainting is to fill the 3D content of regions specified by users. SPin-NeRF [24] achieves this through a multi-step process: it employs SAM [15] for object segmentation, utilizes LaMa [31] to paint the background content in multi-view images, and subsequently trains the NeRF model with color, depth, and perceptual cues. In our experiments, we use SAM and LaMa as the 2D image-to-image translator in our GenN2N, which is the same setting as SPin-NeRF [24]. Quantitative comparisons on statue dataset [24] are shown in Table 4, where GenN2N achieves superior PSNR and SSIM scores than SPin-NeRF [24], highlighting the effectiveness of our GenN2N framework. In addition, qualitative results are showcased in Fig. 10 revealing that while SPin-NeRF [24] fails to generate reasonable content behind the masked object, our GenN2N produces realistic content in the same region with fine multi-view consistency." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 528, + 147, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 528, + 147, + 539 + ], + "spans": [ + { + "bbox": [ + 47, + 528, + 147, + 539 + ], + "type": "text", + "content": "4.2. Ablation Studies" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 546, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 546, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 546, + 287, + 593 + ], + "type": "text", + "content": "We conduct comprehensive ablation experiments to validate the designs of each component in GenN2N. Due to space limitations, we only highlight the essential aspects below. Please refer to supplementary for more details." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": "The Contrastive Loss. We demonstrate the advantages of incorporating our proposed contrastive loss in Table 1, 2, 3. The motivation is to disentangle the camera view and edit information present in the latent space. We achieve this by reducing the distance between edit codes of different-view rendered images from a translated NeRF that shares the same edit code, while increasing the distance between same-view images that are edited by the 2D translator with diverse edit styles. As demonstrated in Tables, the absence of contrastive loss leads to the generation of blurry areas" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "in the rendered images, resulting in a decrease in the metric scores. This blurriness can be attributed to the inclusion of pose information within the edit code z. By incorporating the contrastive loss, our method successfully achieves a uniform appearance with different observing views under the same style latent z." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 148, + 546, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 148, + 546, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 148, + 546, + 327 + ], + "type": "text", + "content": "Discriminator for Novel Views. We demonstrate the effectiveness of employing a conditional discriminator to address artifacts caused by inconsistent cross-view edited images and to enhance the quality of novel view rendering images, as depicted in Table 1, 2, 3. The removal of this conditional discriminator results in blurry novel view images with artifacts in the background region. We attribute these undesirable effects to the inability of current image-to-image translation methods, such as InstructPix2Pix [2], to produce image editing consistently across multi-view images. To mitigate these issues, we introduce a conditional discriminator between rendered images from the translated NeRF and edited images from the 2D image-to-image translator. This inclusion successfully eliminates artifacts and enhances the image quality of rendered images from the translated NeRF." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 346, + 388, + 359 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 346, + 388, + 359 + ], + "spans": [ + { + "bbox": [ + 306, + 346, + 388, + 359 + ], + "type": "text", + "content": "4.3. Applications" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 368, + 545, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 368, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 304, + 368, + 545, + 488 + ], + "type": "text", + "content": "We demonstrate the versatility and robustness of GenN2N by exploring two translation applications: Zoom Out and Text-Driven Inpainting. While existing 2D translators [1, 29] can complete these tasks, 3D editing has not been explored. We achieve these tasks by incorporating Blended Latent Diffusion [1] as the 2D image-to-image translator, enabling us to generate diverse and high-quality content with multi-view consistency. Please refer to our supplementary materials for the results of these applications due to space constraints." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 510, + 383, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 510, + 383, + 522 + ], + "spans": [ + { + "bbox": [ + 306, + 510, + 383, + 522 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": "We introduce GenN2N, a unified NeRF-to-NeRF translation framework that can handle various NeRF editing tasks. Unlike previous task-specific approaches, our framework uses an image-to-image translator for 2D editing and integrates the results into 3D NeRF space. To address the challenge of ensuring 3D consistency, we propose modeling the distribution of 3D edited NeRFs from 2D edited images using our novel techniques. After optimization, users can sample from the conditional generative model to obtain diverse 3D editing results with high rendering quality and multi-view consistency. Our experiments demonstrate that GenN2N outperforms existing task-specific methods on various editing tasks, including text-driven editing, colorization, super-resolution, and inpainting, in terms of efficiency, quality, and diversity." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5112" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM Transactions on Graphics (TOG), 42 (4):1-11, 2023. 3, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 287, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 287, + 158 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 287, + 158 + ], + "type": "text", + "content": "[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. arXiv preprint arXiv:2211.09800, 2022. 3, 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 158, + 288, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 288, + 213 + ], + "type": "text", + "content": "[3] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 214, + 288, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 214, + 288, + 269 + ], + "spans": [ + { + "bbox": [ + 53, + 214, + 288, + 269 + ], + "type": "text", + "content": "[4] Shengqu Cai, Anton Obukhov, Dengxin Dai, and Luc Van Gool. Pix2nerf: Unsupervised conditional p-gan for single image to neural radiance fields translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3981-3990, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 269, + 288, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 269, + 288, + 335 + ], + "spans": [ + { + "bbox": [ + 53, + 269, + 288, + 335 + ], + "type": "text", + "content": "[5] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 335, + 288, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 288, + 368 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 288, + 368 + ], + "type": "text", + "content": "[6] Jun-Kun Chen, Jipeng Lyu, and Yu-Xiong Wang. Neuraleditor: Editing neural radiance fields via manipulating point clouds. arXiv preprint arXiv:2305.03049, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 369, + 288, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 288, + 424 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 288, + 424 + ], + "type": "text", + "content": "[7] Pei-Ze Chiang, Meng-Shiun Tsai, Hung-Yu Tseng, Wei-Sheng Lai, and Wei-Chen Chiu. Stylizing 3d scene via implicit representation and hypernetwork. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1475-1484, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 424, + 288, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 424, + 288, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 424, + 288, + 468 + ], + "type": "text", + "content": "[8] Ankit Dhiman, R Srinath, Srinjay Sarkar, Lokesh R. Boregowda, and R Venkatesh Babu. Corf: Colorizing radiance fields using knowledge distillation. ArXiv, abs/2309.07668, 2023. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 469, + 288, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 469, + 288, + 513 + ], + "spans": [ + { + "bbox": [ + 53, + 469, + 288, + 513 + ], + "type": "text", + "content": "[9] Ayaan Haque, Matthew Tancik, Alexei A Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. arXiv preprint arXiv:2303.12789, 2023.2,3,6,7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 514, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 568 + ], + "type": "text", + "content": "[10] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 569, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 288, + 635 + ], + "type": "text", + "content": "[11] Jingwei Huang, Justus Thies, Angela Dai, Abhijit Kundu, Chiyu Jiang, Leonidas J Guibas, Matthias Nießner, Thomas Funkhouser, et al. Adversarial texture optimization from rgb-d scans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1559-1568, 2020. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 288, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 691 + ], + "type": "text", + "content": "[12] Yi-Hua Huang, Yue He, Yu-Jie Yuan, Yu-Kun Lai, and Lin Gao. Stylizednerf: consistent 3d scene stylization as stylized nerf via 2d-3d mutual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18342–18352, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "text", + "content": "[13] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object genera" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 715 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "tion with dream fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 867-876, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 107, + 545, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 152 + ], + "type": "text", + "content": "[14] Xiaoyang Kang, Tao Yang, Wenqi Ouyang, Peiran Ren, Lingzhi Li, and Xuansong Xie. Ddcolor: Towards photorealistic and semantic-aware image colorization via dual decoders. arXiv preprint arXiv:2212.11613, 2022. 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 152, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 545, + 196 + ], + "type": "text", + "content": "[15] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 197, + 545, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 197, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 308, + 197, + 545, + 252 + ], + "type": "text", + "content": "[16] Zhengfei Kuang, Fujun Luan, Sai Bi, Zhixin Shu, Gordon Wetzstein, and Kalyan Sunkavalli. Palettenerf: Palette-based appearance editing of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20691-20700, 2023. 2, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 253, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 253, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 253, + 545, + 319 + ], + "type": "text", + "content": "[17] Gustav Larsson, Michael Maire, and Gregory Shakhnarovich. Learning representations for automatic colorization. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 577-593. Springer, 2016. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 320, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 320, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 308, + 320, + 545, + 376 + ], + "type": "text", + "content": "[18] Verica Lazova, Vladimir Guzov, Kyle Olszewski, Sergey Tulyakov, and Gerard Pons-Moll. Control-nerf: Editable feature volumes for scene rendering and manipulation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4340–4350, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 376, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 545, + 409 + ], + "type": "text", + "content": "[19] Hao-Kang Liu, I Shen, Bing-Yu Chen, et al. Nerf-in: Free-form nerf inpainting with rgb-d priors. arXiv preprint arXiv:2206.04901, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 410, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 410, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 308, + 410, + 545, + 454 + ], + "type": "text", + "content": "[20] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. arXiv preprint arXiv:2303.11328, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 456, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 456, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 456, + 545, + 510 + ], + "type": "text", + "content": "[21] Steven Liu, Xiuming Zhang, Zhoutong Zhang, Richard Zhang, Jun-Yan Zhu, and Bryan Russell. Editing conditional radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5773-5783, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 511, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 511, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 511, + 545, + 555 + ], + "type": "text", + "content": "[22] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Realfusion: " + }, + { + "bbox": [ + 308, + 511, + 545, + 555 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 308, + 511, + 545, + 555 + ], + "type": "text", + "content": " reconstruction of any object from a single image. arXiv e-prints, pages arXiv-2302, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 556, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 556, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 556, + 545, + 612 + ], + "type": "text", + "content": "[23] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 613, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 689 + ], + "type": "text", + "content": "[24] Ashkan Mirzaei, Tristan Aumentado-Armstrong, Konstantinos G Derpanis, Jonathan Kelly, Marcus A Brubaker, Igor Gilitschenski, and Alex Levinshtein. Spin-nerf: Multiview segmentation and perceptual inpainting with neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20669–20679, 2023. 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 690, + 545, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 690, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 308, + 690, + 545, + 715 + ], + "type": "text", + "content": "[25] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5113" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "text", + "content": "translation. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "type": "text", + "content": "[26] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 130, + 288, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 130, + 288, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 130, + 288, + 196 + ], + "type": "text", + "content": "[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 198, + 288, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 288, + 252 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 288, + 252 + ], + "type": "text", + "content": "[28] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 3, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 254, + 288, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 288, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 288, + 308 + ], + "type": "text", + "content": "[29] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 2, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 310, + 288, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 310, + 288, + 354 + ], + "spans": [ + { + "bbox": [ + 48, + 310, + 288, + 354 + ], + "type": "text", + "content": "[30] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. Advances in Neural Information Processing Systems, 33:20154-20166, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 355, + 287, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 355, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 48, + 355, + 287, + 420 + ], + "type": "text", + "content": "[31] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. arXiv preprint arXiv:2109.07161, 2021.8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 422, + 288, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 288, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 288, + 487 + ], + "type": "text", + "content": "[32] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, et al. Nerfstudio: A modular framework for neural radiance field development. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-12, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 489, + 288, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 489, + 288, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 489, + 288, + 544 + ], + "type": "text", + "content": "[33] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3835–3844, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 545, + 288, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 288, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 288, + 599 + ], + "type": "text", + "content": "[34] Chen Wang, Xian Wu, Yuan-Chen Guo, Song-Hai Zhang, Yu-Wing Tai, and Shi-Min Hu. Nerf-sr: High quality neural radiance fields using supersampling. In Proceedings of the 30th ACM International Conference on Multimedia, pages 6445-6454, 2022. 2, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 601, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 645 + ], + "type": "text", + "content": "[35] Can Wang, Ruixiang Jiang, Mengei Chai, Mingming He, Dongdong Chen, and Jing Liao. Nerf-art: Text-driven neural radiance fields stylization. IEEE Transactions on Visualization and Computer Graphics, 2023. 3, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 647, + 288, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 700 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 700 + ], + "type": "text", + "content": "[36] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3d digital avatars using diffusion. arXiv preprint arXiv:2212.06135, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 702, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 702, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 702, + 288, + 714 + ], + "type": "text", + "content": "[37] Qiling Wu, Jianchao Tan, and Kun Xu. Palettenerf:" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 510 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "Palette-based color editing for nerfs. arXiv preprint arXiv:2212.12871, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 547, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 140 + ], + "type": "text", + "content": "[38] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 141, + 547, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 547, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 547, + 195 + ], + "type": "text", + "content": "[39] Yu-Jie Yuan, Yang-Tian Sun, Yu-Kun Lai, Yuewen Ma, Rongfei Jia, and Lin Gao. Nerf-editing: geometry editing of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18353-18364, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 197, + 547, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 547, + 239 + ], + "type": "text", + "content": "[40] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Resshift: Efficient diffusion model for image super-resolution by residual shifting. arXiv preprint arXiv:2307.12348, 2023. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 241, + 547, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 547, + 296 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 547, + 296 + ], + "type": "text", + "content": "[41] Jiakai Zhang, Xinhang Liu, Xinyi Ye, Fuqiang Zhao, Yanshun Zhang, Minye Wu, Yingliang Zhang, Lan Xu, and Jingyi Yu. Editable free-viewpoint video using a layered neural representation. ACM Transactions on Graphics (TOG), 40(4):1-18, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 297, + 547, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 547, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 547, + 352 + ], + "type": "text", + "content": "[42] Kai Zhang, Nick Kolkin, Sai Bi, Fujun Luan, Zexiang Xu, Eli Shechtman, and Noah Snavely. Arf: Artistic radiance fields. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXI, pages 717-733. Springer, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 354, + 547, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 547, + 385 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 547, + 385 + ], + "type": "text", + "content": "[43] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 387, + 547, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 387, + 547, + 442 + ], + "spans": [ + { + "bbox": [ + 307, + 387, + 547, + 442 + ], + "type": "text", + "content": "[44] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 443, + 547, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 443, + 547, + 476 + ], + "spans": [ + { + "bbox": [ + 307, + 443, + 547, + 476 + ], + "type": "text", + "content": "[45] Chengwei Zheng, Wenbin Lin, and Feng Xu. Editablenerf: Editing topologically varying neural radiance fields by key points. arXiv preprint arXiv:2212.04247, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 477, + 547, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 477, + 547, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 477, + 547, + 510 + ], + "type": "text", + "content": "[46] Jingyu Zhuang, Chen Wang, Lingjie Liu, Liang Lin, and Guanbin Li. Dreameditor: Text-driven 3d scene editing with neural fields. arXiv preprint arXiv:2306.13455, 2023. 3" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "5114" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/GenNBV_ Generalizable Next-Best-View Policy for Active 3D Reconstruction/171b2e0e-83b0-4811-a381-9c85825e51ab_content_list.json b/2024/GenNBV_ Generalizable Next-Best-View Policy for Active 3D Reconstruction/171b2e0e-83b0-4811-a381-9c85825e51ab_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2946ab986bb7e5700be70e6b4fa4781c99378b28 --- /dev/null +++ b/2024/GenNBV_ Generalizable Next-Best-View Policy for Active 3D Reconstruction/171b2e0e-83b0-4811-a381-9c85825e51ab_content_list.json @@ -0,0 +1,1456 @@ +[ + { + "type": "text", + "text": "GenNBV: Generalizable Next-Best-View Policy for Active 3D Reconstruction", + "text_level": 1, + "bbox": [ + 98, + 130, + 870, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiao Chen $^{1,2}$ Quanyi Li $^{1}$ Tai Wang $^{1}$ Tianfan Xue $^{2,\\dagger}$ Jiangmiao Pang $^{1,\\dagger}$", + "bbox": [ + 176, + 178, + 799, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Shanghai AI Laboratory $^{2}$ The Chinese University of Hong Kong", + "bbox": [ + 223, + 200, + 761, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{cx123,tfxue}@ie.cuhk.edu.hk quanyili0057@gmail.com {wangtai,pangjiangmiao}@pjlab.org.cn", + "bbox": [ + 86, + 222, + 890, + 238 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/97e6c475f76b6f9fa9f7dbf28d5c5131ec961abd99ab0e3a9e7143b9500460d2.jpg", + "image_caption": [ + "Figure 1. To determine the best view for 3D reconstruction, previous methods only chose from hand-crafted action space or based on object-centric capturing, lacking the ability to generalize to unforeseen scenes (Left). With our end-to-end trained, generalized free-space policy, it can generalize to unseen objects, enabling the captured drone to image from any viewpoint (Right)." + ], + "image_footnote": [], + "bbox": [ + 81, + 253, + 436, + 402 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ded81a23d7169a288f767c4332cb809663f91477947822176507cd55a0374d8c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 450, + 255, + 875, + 404 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 472, + 313, + 488 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While recent advances in neural radiance field enable realistic digitization for large-scale scenes, the image-capturing process is still time-consuming and labor-intensive. Previous works attempt to automate this process using the Next-Best-View (NBV) policy for active 3D reconstruction. However, the existing NBV policies heavily rely on hand-crafted criteria, limited action space, or perscense optimized representations. These constraints limit their cross-dataset generalizability. To overcome them, we propose GenNBV, an end-to-end generalizable NBV policy. Our policy adopts a reinforcement learning (RL)-based framework and extends typical limited action space to 5D free space. It empowers our agent drone to scan from any viewpoint, and even interact with unseen geometries during training. To boost the cross-dataset generalizability, we also propose a novel multi-source state embedding, including geometric, semantic, and action representations. We establish a benchmark using the Isaac Gym simulator with the Houses3K and OmniObject3D datasets to evaluate this NBV policy. Experiments demonstrate that our policy achieves a $98.26\\%$ and $97.12\\%$ coverage ratio on unseen building-scale objects from these datasets, respectively, outperforming prior solutions.", + "bbox": [ + 73, + 498, + 473, + 845 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 472, + 630, + 488 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in 3D reconstruction [23, 38, 39] and neural rendering [20, 37, 48] have significantly enhanced the quality of 3D digitization of large scenes, such as buildings and city landmarks [11, 17, 18, 43, 47]. However, image-capturing process still remains time-consuming and labor-intensive. To scan a building-scale scene using a drone, it may take several days' effort of a professional team to ensure a full coverage. Moreover, even professional pilots may miss some areas at the first scan, leading to multiple rounds for rescanning.", + "bbox": [ + 496, + 498, + 890, + 651 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To alleviate the effort in manual scanning, active 3D reconstruction algorithms, especially the Next-Best-View (NBV) policy, have emerged as a promising approach to automate view planning. However, typical rule-based NBV policies [9, 14, 22, 26, 28, 46] heavily rely on empirically designed view-selection criteria and hand-crafted action space, such as a hemisphere as shown in Fig. 1, which limit their generalizability to unseen scenes. To enhance generalization capacity, pioneering works on learning-based NBV policy have been proposed [3, 4, 26, 46]. Nevertheless, they are constrained into impractical setups in real-world scenarios, such as object-centric capturing and limited action space for ground robots, resulting in incomplete 3D reconstruction because of significant self-occlusion.", + "bbox": [ + 496, + 654, + 892, + 866 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Therefore, in this work, we study this problem: \"How to design an NBV policy that supports exploration in any", + "bbox": [ + 500, + 869, + 890, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Project website: gennbv.tech", + "bbox": [ + 94, + 875, + 261, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding authors.", + "bbox": [ + 96, + 887, + 225, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16436", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "space and also can generalize to unseen building-scale geometries?\" Building upon the design of previous reinforcement learning-based (RL-based) NBV policies [3, 4, 26], our setup faces the following new challenges that were not fully studied before: 1) Designing an easy-to-generalize action space, rather than object-specific spaces used in previous work, such as hand-crafted candidates or a constrained hemisphere; 2) With a large action space, an informative representation is necessary to guide the policy to efficiently find an optimal capturing trajectory; 3) Proposing generic reward functions and developing the distributed training procedure for better generalizability.", + "bbox": [ + 75, + 90, + 472, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome these two challenges, we propose GenNBV, an end-to-end generalizable NBV policy. It has a novel design of action spaces and state embeddings, which allows free space exploration and ensures a large coverage when applying to unseen objects. We firstly extend former limited action space, like a hemisphere, to 5D free space. As shown in Fig. 1, this new space is composed of a position subspace of approximately $20m \\times 20m \\times 10m$ and an omnidirectional heading subspace. This free space enables our agent drone to scan from any viewpoint. Moreover, with much larger action space, our RL-based policy can be easily generalized to building-scale geometries without concerning the potential scale shift from training to evaluation. On the contrary, many hand-crafted designs [14, 22] limit available viewpoints to a small predefined set, and thus cannot generalize if there is drastic change in scales.", + "bbox": [ + 75, + 275, + 472, + 517 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Second, in order to generalize a well-trained NBV policy to unseen scenes during evaluation, we propose a novel state representation directly from raw sensory inputs to guide the policy. In contrast to the widely-used neural radiance field (NeRF) [14, 22, 28, 34, 46], our representation does not need time-consuming per-scene optimization. Specifically, our NBV policy incorporates a multi-source state embedding, which comprises three key components: a novel geometric embedding, a semantic embedding, and an action embedding. The geometric embedding is derived from multi-view depth maps and represents an encoded probabilistic 3D grid, while the semantic embedding is extracted from RGB images, and the action embedding is an encoded viewpoint sequence. In contrast to 2D representations used in previous RL-based NBV policies [3, 4], our multi-source state embedding, extracted from a probabilistic 3D grid, is a more comprehensive representation that supports a robust prediction of next viewpoints.", + "bbox": [ + 75, + 518, + 472, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To validate the effectiveness of GenNBV, we construct a benchmark for training and evaluation with Houses3K [25] dataset from the NVIDIA Isaac Gym [19] simulator. We further test its generalization ability on novel buildings and object categories from OmniObject3D [42] and Objaverse [6] datasets. Our evaluation metrics quantify the completeness, accuracy, and efficiency for active 3D re", + "bbox": [ + 75, + 794, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "construction. To consolidate completeness and efficiency into a single number, we also propose a metric, Area Under the Curve (AUC) of the coverage. Compared with heuristic [12], information gain-based [12, 14, 46], and RL-based baselines [26], our method achieves the best result under different metrics. We also provide visualizations to demonstrate the generalizability of GenNBV. A demo video is attached in the appendix to further illustrate our method.", + "bbox": [ + 496, + 90, + 893, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 224, + 640, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Traditional 3D Reconstruction. Photometry and geometry are two crucial aspects of reconstruction evaluation. Neural implicit representations [20, 23] have shown progress in photometric rendering but faces challenges like time-consuming optimization and poor generalization, hindering its application in real-time reconstruction scenarios like 3D SLAM [21, 30, 36, 50]. Geometry, in contrast, typically corresponds to 3D representations such as point clouds and 3D mesh and is directly related to issues like collision avoidance. Considering these properties, researchers tend to focus on geometric reconstruction in practical applications and we also follow this stream.", + "bbox": [ + 496, + 251, + 893, + 431 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Active 3D Reconstruction. Active 3D reconstruction is a promising field that has not been thoroughly benchmarked yet. The pipeline of active 3D reconstruction frameworks alternates between inferring optimal viewpoints, capturing new data, and updating the rebuilt 3D model. With the optimal sequence of key viewpoints, a continuous planning path can be generated using the classic planners, such as Fast Marching Method (FMM) [33].", + "bbox": [ + 496, + 431, + 893, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing works can be pivotally differentiated based on the paradigm of NBV policies: rule-based or learning-based. Rule-based approaches like uncertainty-driven works [12, 14, 28, 46] typically yield the next best view from hand-crafted rules from scene representations, which tends to overfit specific scenes. Most learning-based policies [3, 4, 26] use a deep reinforcement learning algorithm like PPO [32] to sequentially predict the optimal viewpoints based on observation. They must obtain feedback from task-related rewards such as coverage ratio [9, 26] and optimize during massive iteration with task environment.", + "bbox": [ + 496, + 553, + 895, + 718 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The action space is designed based on the paradigm of NBV policies. Most rule-based NBV policies select views from a limited action space, such as a set composed of only one hundred candidate viewpoints [22], a tiny pre-defined space like a hemisphere [9, 14, 28, 46], making it possible to overlook some important viewpoints due to unavailability. Even though learning-based categories further explore the larger action space like a 2D plain [3, 4] or a constrained 3D space [26], their limited action space still prevents them from capturing sufficient details for 3D reconstruction.", + "bbox": [ + 496, + 719, + 895, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Scene representation built from history observations, which directly provide reconstruction progress to NBV", + "bbox": [ + 500, + 869, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16437", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/78664c646a8f660fbe0d055e673107469844db204a1d0eab6e3db8e59a6e10a6.jpg", + "image_caption": [ + "Figure 2. Overview of our proposed framework GenNBV. Our end-to-end policy takes the historical multi-source observations as input, transforms them into a more informative scene representation, and produces the next viewpoint position. A reward signal will be returned at training time to optimize the end-to-end policy for maximizing the expected cumulative reward in one episode. Specifically, the signal is the increased coverage ratio after collecting a new viewpoint." + ], + "image_footnote": [], + "bbox": [ + 81, + 92, + 883, + 277 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "policies, is also a crucial aspect of the active 3D reconstruction framework. Previous works have explored visual representations for NBV policies, such as TSDF [11], neural radiance field [1], and 2D BEV maps [3, 10, 44]. However, implicit representations are hard to jointly with learning-based NBV policies, and 2D BEV map lacks sufficient information for large-scale outdoor scenes that contain numerous geometric details in 3D space.", + "bbox": [ + 75, + 358, + 468, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our benchmark, we select Scan-RL [26] as our main RL-based baseline. Compared to it, we introduce a much larger action space and informative visual representations for better generalizability. Additionally, we don't benchmark the mentioned baselines [3, 4], mainly due to their reliance on pretraining from human demonstration using imitation learning, while our focus lies on learning through iterations and end-to-end training. Different from our 3D free-space movement, they only explore the traversable 2D areas of indoor scenes using ground robots, thus limiting their motion space in 2D.", + "bbox": [ + 75, + 484, + 470, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generalizable Reinforcement Learning. Increasing the diversity of training data leads to better generalizability [5]. In the robotics field, domain randomization [41] allows legged robots to walk on various terrains absent from the training environment. For end-to-end driving policy, training in large-scale synthetic and realistic scenarios improves the safety of autonomous cars [15, 16] in unseen test scenarios. The advanced work [7] presents an RL-based formulation for the view planning problem. Based on similar frameworks of active 3D reconstruction, Scan-RL [26] and SCONE [9] are two pioneering works showing the generalizability of their frameworks, while both of them suffer from the constrain of viewpoint sampling in limited space and lack of in-depth analysis. In this work, we tackle the aforementioned issues by predicting NBV in free 3D space with a learning-based planner, which is trained with a", + "bbox": [ + 75, + 659, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "dataset containing buildings in various shapes and poses for acquiring generalizability for unseen buildings.", + "bbox": [ + 498, + 358, + 890, + 390 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 500, + 401, + 635, + 419 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we deliver our active 3D reconstruction framework GenNBV, especially the pivotal Next-Best-View policy. An overview of our framework is illustrated in Fig. 2. Firstly, we formulate the NBV problem as a Markov Decision Process (MDP), with a novel design of observations (blue boxes in Fig. 2) and action space in Sec. 3.1. Next, we elaborate our end-to-end NBV policy $\\pi$ (orange box in Fig. 2) in Sec. 3.2. Inspired by Curl [13], we point out that generalizability greatly depends on how to extract embeddings (green box in Fig. 2), which reflect the reconstruction progress, from raw sensory observations like RGB images and camera poses. In Sec. 3.3, we introduce the reward function (right-hand side of Fig. 2) reflecting the optimization objective and the details of policy optimization.", + "bbox": [ + 496, + 426, + 892, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Formulation of the Next-Best-View Problem", + "text_level": 1, + "bbox": [ + 500, + 650, + 872, + 666 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We formulate the NBV problem as learning an optimal policy $\\pi$ that controls the capturing process, such that enough information is captured for large-scale scene reconstruction, with limited decision-making budgets. As capturing, transferring, and processing a large set of captured images introduce significant computational costs, we also want to design a policy that also minimizes the number of captured images. Therefore, our policy only captures sparse keyframes that sufficiently record all details of objects.", + "bbox": [ + 496, + 672, + 890, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 2, at each time step $t$ , the agent receives a visual observation $o_t$ , takes an action $a_t$ , infers the action at the next time step $t + 1$ to move to a new location, and then receives a new visual observation, repeating this interaction with the environment until the episode ends. Our simulated agent is embodied as CrazyFlie [8], a type of unmanned", + "bbox": [ + 496, + 810, + 892, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16438", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "aerial vehicle equipped with various sensors, including an RGB-D camera and an IMU, to execute data collection for reconstruction. We discussed the details of the observation space and action space below.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Observation Space. As shown in the left column of Fig. 2, at each time step, the agent receives an RGB image $I_{t}$ , a depth map $D_{t}$ , and a state vector including the heading (yaw and pitch) and position $(\\mathrm{x},\\mathrm{y},\\mathrm{z})$ of the onboard camera. The observation $o_{t}$ (yellow box in Fig. 2) consists of all previously captured images in one episode, historical actions, and the current captures and actions. With this information as input, the policy network can estimate the progress of data collection and determine where to scan next.", + "bbox": [ + 75, + 152, + 470, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Action Space. Unlike most previous NBV algorithms, we use larger action space in order to cover all details of objects. Specifically, we use the camera location and camera angle as our action space, which is a 5-dimension vector consisting of 3D position coordinates and 2D rotation angles (yaw-axis and pitch-axis). We restrict the roll-axis rotation as it is not supported in all drone platforms.", + "bbox": [ + 75, + 290, + 468, + 397 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Generalizable State Embedding", + "text_level": 1, + "bbox": [ + 76, + 406, + 357, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To ensure that the policy learned on one set of 3D objects can generalize to objects with different appearances and structures, a smart state embedding of raw sensor observations is needed to capture invariant features across different objects. Previous methods [3, 26] only extract representations in 2D space, which are very sensible to appearance changes, and our experiments have shown that policies only trained on 2D features may not generalize to different objects. Instead, we propose a multi-source representation that has better generalizability.", + "bbox": [ + 75, + 430, + 468, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, we first build two mid-level representations that can better model the relationship between the objects and our agent: a 3D geometric representation $F^G$ from depth maps and a semantic representation $F^S$ from RGB images. Then we encode these representations and concatenate them with a pose embedding into state embedding $s_t$ , guiding the NBV policy for subsequent decision-making. The overall encoder for state embedding is shown in Fig. 2, and details of each representation are discussed below.", + "bbox": [ + 75, + 580, + 468, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Geometric Representation One simple way to record the geometry of a 3D object is using a binary 3D occupancy map [3], where the value of each cell in the 3D cube indicates whether a cell contains a 3D object or not. However, since the 3D representation is gradually updated with newly captured data, this simple binary occupancy map cannot differentiate a real unoccupied cell from an unscanned cell. An unscanned cell is a strong indicator that the agent should capture more data in this region, while no further scan is needed for a real unoccupied cell.", + "bbox": [ + 75, + 719, + 468, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Previous works [3, 10, 44] simplify the 3D scene representation into a 2D Bird's Eye View map for actively recon", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "structuring indoor scenes. However, the 2D BEV map lacks sufficient information for our large-scale outdoor scenes that contain many geometric details in 3D space. In addition, their wheeled robotic platforms are constrained to freely scan and represent these outdoor scenes. Therefore, to model the scanning process in 3D free space, we employ the probabilistic 3D grid [40] as our geometric 3D representation, which records the probability of each 3D voxel being captured or not. Specifically, we first obtain a 3D point cloud in the world coordinate by back-projecting all 2D pixels to 3D points, using the depth map $D_{t}$ , camera intrinsic parameters, and camera pose $a_{t}$ . By voxelizing the obtained point cloud, we then build a 3D occupancy grid that explicitly indicates the binary state (occupied or free) in this 3D space. Subsequently, we represent this 3D grid as a probabilistic occupancy grid $F_{t}^{G}$ and extend the state space of voxels with three states (occupied, free, unknown).", + "bbox": [ + 496, + 90, + 890, + 347 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During each scanning (one episode in RL), at each step $t + 1$ , we update the probabilistic occupancy grid $F_{t + 1}^{G}$ based on the preceding grid $F_{t}^{G}$ and current observation. Specifically, the grid is updated through Bresenham's line algorithm [2], which casts the ray path in 3D space between the camera viewpoint and the endpoints among the point cloud back-projected from depth $D_{t + 1}$ . Following the classical occupancy grid mapping algorithm [40], we have the log-odds formulation of occupancy probability:", + "bbox": [ + 496, + 347, + 890, + 484 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\log \\operatorname {O d d} \\left(v _ {i} \\mid z _ {j}\\right) = \\log \\operatorname {O d d} \\left(v _ {i}\\right) + C, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 494, + 890, + 511 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $v_{i}$ denotes the occupancy probability of $i^{th}$ voxel in the grid $F_{t}^{G}$ , $z_{j}$ is the measurement event that $j^{th}$ camera ray passes through this voxel and $C$ is an empirical constant. The derivation can be found in Appendix. Thus, we update the log-odds occupancy probability of each voxel in the grid $F_{t}^{G}$ by adding a constant one time when a single camera ray passes through this voxel. Note that the probabilistic occupancy grid $F^{G}$ is continually updated within one episode. Finally, the occupancy status of voxels is classified into three categories: unknown, occupied, and free, using preset probability thresholds.", + "bbox": [ + 496, + 521, + 890, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Semantic Representation. Geometric representation enables agents to comprehend spatial occupancy, yet it's insufficient for perceiving the environment. For example, when observing a hole in an object, the agent may struggle to differentiate between incomplete scanning and the actual presence of a hole in the object. In such cases, the semantic information contained in the captured RGB images can help the agent distinguish between these two scenarios.", + "bbox": [ + 496, + 688, + 890, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To provide semantic information, we employ a preprocessing module that takes as input the current frame of RGB image $I_{t}$ and preceding $k$ frames and converts these frames $[I_t,I_{t - 1},\\dots,I_{t - k}]$ to grayscale, and concatenate them as output, following [26]. Then the preprocessed frames are fed into a two-layer convolutional network for extracting", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16439", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the semantic representation $F_{t}^{S}$", + "bbox": [ + 76, + 90, + 292, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "State Embedding. To further combine the semantic and geometric embeddings, we first encode them to $s_t^S = f^S (F_t^S)$ and $s_t^G = f^G (F_t^G)$ where $f^{*}$ are learnable networks Linear(Flatten(x)), as shown in Fig. 2. Subsequently, we combine them with the historical action embedding $s_t^A = \\mathrm{Linear}(a_{1:t})$ to generate the final state embedding $s_t$ , as the input to the policy network. This process can be formulated as: $s_t = \\mathrm{Linear}(s_t^G;s_t^S;s_t^A)$ .", + "bbox": [ + 75, + 107, + 468, + 228 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Policy Network. Taking the state embedding $s_t$ as input, the policy network is a 3-layer multi-layer perceptron network (MLP) whose output is used to parameterize a normal distribution over action space. In this way, the action can be drawn from the stochastic policy $a \\sim \\pi(\\cdot | o_t)$ .", + "bbox": [ + 75, + 229, + 468, + 305 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Reward Function and Optimization", + "text_level": 1, + "bbox": [ + 76, + 313, + 387, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We train the end-to-end policy with reinforcement learning (RL) and hence design reward functions to precisely reflect the task objective for 3D reconstruction. The policy is optimized with proximal policy optimization [32] (PPO) for parallelizing sampling.", + "bbox": [ + 75, + 335, + 468, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Reward Functions. With the occupancy probability $F_{t}^{G}$ at time step $t$ , we can threshold each voxel with an empirical bound to determine if it is occupied. This discrimination process outputs a binary occupancy grid with $\\tilde{N}_{t}$ voxels being occupied, which is used to calculate the coverage ratio:", + "bbox": [ + 75, + 412, + 468, + 489 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm{CR}_{t} = \\frac{\\tilde{N}_{t}}{N^{*}}\\cdot 100\\% , \\tag{2}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 501, + 468, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $N^{*}$ is the number of ground-truth occupied voxels representing the surface of objects. To encourage our NBV policy to cover as many unseen areas of objects as possible, we use the difference of coverage ratio (CR) between two consecutive steps as the main reward function $r^{CR}$ :", + "bbox": [ + 75, + 536, + 468, + 609 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr _ {t + 1} ^ {\\mathrm {C R}} = \\mathrm {C R} _ {t + 1} - \\mathrm {C R} _ {t}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 619, + 468, + 640 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In free-space exploration, we also need to avoid collision. Previous limited-space agents [9, 14, 46] do not consider collision avoidance since their search space, like hemisphere, is by-design safe. Thus, we add a negative reward for collision and terminate the episode if a collision happens. We also implement a negative reward when the number of captured keyframes is over an empirical threshold to improve the path efficiency.", + "bbox": [ + 75, + 648, + 468, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Policy Optimization. Once the reward functions has been specified, the policy can be learned through any off-the-shelf RL algorithm. In this work, we use PPO as thousands of workers can be parallelized to improve the sample efficiency. Specifically, given our parameterized policy $\\pi_{\\theta}$ , the objective of PPO is to maximize the following function:", + "bbox": [ + 75, + 771, + 468, + 861 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL (\\theta) = \\mathbb {E} _ {t} \\left[ \\frac {\\pi_ {\\theta} (a _ {t} | s _ {t})}{\\pi_ {\\theta_ {\\mathrm {o l d}}} (a _ {t} | s _ {t})} A ^ {\\pi_ {\\theta_ {\\mathrm {o l d}}}} (s _ {t}, a _ {t}) \\right], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 869, + 468, + 905 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $A^{\\pi_{\\theta_{\\mathrm{old}}}}(s_t, a_t)$ is the advantage function that measures the value of taking action $a_t$ at state $s_t$ under the current policy $\\pi_{\\theta_{\\mathrm{old}}}$ . To prevent significant deviation of the new policy from the old policy, PPO incorporates a clipped surrogate objective function:", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L ^ {\\mathrm {C L I P}} (\\theta) = \\mathbb {E} _ {t} [ \\min (\\eta_ {t} (\\theta) A ^ {\\pi_ {\\theta_ {\\text {o l d}}}} (s _ {t}, a _ {t}), \\tag {5} \\\\ \\left. \\operatorname {c l i p} \\left(r _ {t} (\\theta), 1 - \\epsilon , 1 + \\epsilon\\right) A ^ {\\pi_ {\\theta_ {\\text {o l d}}}} \\left(s _ {t}, a _ {t}\\right) \\right], \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 172, + 890, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\eta_t(\\theta) = \\frac{\\pi_\\theta(a_t|s_t)}{\\pi_{\\theta_{\\mathrm{old}}}(a_t|s_t)}$ and $\\epsilon$ is a hyper-parameter that controls the size of the trust region.", + "bbox": [ + 496, + 218, + 890, + 253 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 498, + 268, + 630, + 286 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we conduct experiments on Houses3K [25], OmniObject3D [42], and Objaverse [6]. For our policy and other learning-based policies, we train them on Houses3K training set. Then we evaluate all policies on the Houses3K test dataset and the OmniObject3D dataset for quantifying the in-distribution and out-of-distribution generalizability. Finally, we show the visualization results on three datasets, demonstrating the effectiveness of the proposed method.", + "bbox": [ + 496, + 294, + 890, + 415 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 498, + 422, + 689, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Simulation Environment. We conduct all experiments in NVIDIA Isaac Gym [19], a physics simulation platform designed for reinforcement learning and robotics research. With GPU-accelerated tensor API and sensor interaction API, we can easily implement customized CUDA kernels to calculate our geometric representation efficiently. Moreover, the sensor simulation is efficient. It can run up to 1000 FPS, significantly reducing the training time. We create the agent drone CrazyFlie [8] and equip it with sensors including RGB-D cameras and IMU. Considering memory limitations, we downsample the image resolution to $400 \\times 400$ . The vertical field of view of this onboard camera is $90^{\\circ}$ . We also set up a point light source with a fixed position and constant intensity in Isaac Gym.", + "bbox": [ + 496, + 446, + 890, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. We conduct our experiments on Houses3K [25], OmniObject3D [42], Objverse 1.0 [6] and Replica [35]. Our model is trained on large-scale 3D objects from Houses3K. To validate the effectiveness and generalizability of our model, we evaluate it on unseen 3D objects from Houses3K and further on batches of unseen cross-domain 3D objects from OmniObject3D, Objverse, and Replica, which include both house and non-house categories.", + "bbox": [ + 496, + 657, + 890, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Houses3K contains 3,000 textured 3D building models. They are divided into 12 batches, with each batch featuring 50 distinct geometries and 5 varying textures. We further eliminate poorly designed geometries based on the following two criteria: 1) objects with complex internal structures (as we are mostly focusing on the surface reconstruction), and 2) objects with redundant bases at the bottom. Following these criteria, the training set consists of 256 objects", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16440", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/9fd38b9c2299fe3e2cc7cb2129c41965695f82d98f8171dfc07b084694c73185.jpg", + "table_caption": [ + "Table 1. Evaluation results of Next-Best-View policies for active 3D reconstruction on Houses3K and the house category from OmniObject3D (cross-dataset generalization). The number of views is set to 30 and 20 for Houses3K and OmniObject3D, respectively.", + "\\*\\*: the policy is trained with the Houses3K training set and evaluated with holdout Houses3K test set and OmniObject3D house category. $\\ddagger$ : the policy heavily relies on optimized per-scene representation (NeRF), and thus is directly trained and evaluated on testing objects." + ], + "table_footnote": [], + "table_body": "
Houses3K Test SetOmniObject3D
NBV PolicyMean AUC ↑Final Coverage Ratio ↑Accuracy (cm) ↓Mean AUC ↑Final Coverage Ratio ↑Accuracy (cm) ↓
HeuristicRandom48.53%58.24%1.3860.13%73.73%0.62
Random Hemisphere71.19%79.72%0.6374.03%84.74%0.48
Uniform Hemisphere82.91%89.71%0.4483.09%92.90%0.41
InfoGain†Uncertainty-Guided [14]83.13%89.31%0.4469.08%92.85%0.42
†ActiveRMap [46]84.86%90.77%0.4469.47%93.15%0.42
RL-based*Scan-RL [26]84.49%91.61%0.4082.17%92.53%0.34
*Scan-RL w/ Our Repre.86.48%92.20%0.3786.14%93.73%0.35
*Ours w/ Scan-RL Repre.87.39%95.63%0.3886.81%93.64%0.34
*GenNBV (Ours)91.19%98.26%0.3788.63%97.12%0.33
", + "bbox": [ + 76, + 157, + 893, + 415 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8f31a7290e347040cf3971def819014c995cf0e18c2b5356767c8d133dacfa0c.jpg", + "table_caption": [ + "Table 2. The cross-dataset generalization for non-house categories. We train the baseline Scan-RL and our GenNBV on Houses3K and generalize them to non-house categories from OmniObject3D and an indoor scene from Replica." + ], + "table_footnote": [], + "table_body": "
CategoryDatasetNBV PolicyAUC ↑FCR ↑Acc. ↓
AnimalsOmniObject3DScan-RL76.43%84.98%0.17
GenNBV84.28%94.07%0.16
TrucksOmniObject3DScan-RL69.58%76.37%0.06
GenNBV75.54%83.47%0.03
DinosaursOmniObject3DScan-RL78.47%86.89%0.88
GenNBV84.24%94.03%0.81
Room 0ReplicaScan-RL69.21%80.75%3.12
GenNBV88.53%99.16%3.59
", + "bbox": [ + 81, + 498, + 468, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "from 6 selected batches, and the test set consists of 50 objects from another batch. All these selected training and test objects have distinct geometries.", + "bbox": [ + 75, + 666, + 468, + 712 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "OmniObject3D offers a collection of 6K high-fidelity objects scanned from real-world sources across 190 typical categories. The house category from OmniObject3D for evaluation has 43 diverse objects. In contrast, Objaverse 1.0 comprises a vast library of 818K 3D synthetic objects spanning 21K categories. The diverse and high-quality nature of these datasets makes them ideal for evaluating our models and visualizing the results.", + "bbox": [ + 75, + 714, + 468, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. The objective of NBV policies is to capture the most useful information for reconstruction, with the least number of views. We report the (1) Final Coverage Ratio $(\\%)$ , (2) Mean AUC $(\\%)$ of coverage ratio and", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(3) Reconstruction Accuracy (cm) along with the consistent number of views in all tables. Specifically, most prior works [9, 26] separately use the coverage ratio (\\%) and the number of views to evaluate the reconstruction completeness and efficiency for NBV policies. However, the coverage ratio is highly correlated with the number of views. Thus we propose to unify the number of views to a fixed value for all NBV policies during evaluation and use the area under the curve (AUC) of coverage ratio as the main metric for comparison. Also, we calculate the Chamfer Distance between scanned and ground-truth point clouds to represent the reconstruction accuracy (cm).", + "bbox": [ + 496, + 433, + 890, + 616 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We conduct all experiments in Isaac Gym simulation engine with one NVIDIA Tesla V100 GPU. Our NBV policy is optimized through over 32 million iterations and uses approximately 24 hours of training time on an NVIDIA V100 GPU. All networks are randomly initialized and trained end-to-end. Our implementation refers to the codebase of Legged Gym [31] and the PPO implementation in Stable Baseline3 [27], which is implemented by PyTorch [24]. The ground-truth point clouds on objects' surfaces are sampled by the Poisson Disk sampling method [45] using Open3D API [49]. The Chamfer Distance between the scanned point cloud and the ground-truth point cloud is calculated using PyTorch3D API [29]. Please refer to the Appendix for further details and experiments.", + "bbox": [ + 496, + 616, + 892, + 828 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Performance Comparison", + "text_level": 1, + "bbox": [ + 500, + 834, + 735, + 849 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To comprehensively demonstrate the effectiveness and generalizability of GenNBV, we design three levels of evaluation experiments: 1) As shown in Table 1, we show the per", + "bbox": [ + 496, + 854, + 890, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16441", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b0d176c1467ed0538f742e96a7dacf2236e75e6ab7466bd292693e05d4f08725.jpg", + "image_caption": [ + "(a) Houses3K" + ], + "image_footnote": [], + "bbox": [ + 86, + 87, + 495, + 277 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a7283e936288c6a87f1ba831d3104b034d1b5a6a91f72dd06461e157e0ac2cf4.jpg", + "image_caption": [ + "(b) OmniObject3D", + "Figure 3. The visualization results of unseen 3D objects reconstructed by Scan-RL [26] and our model to compare the generalizability. (a) Unseen buildings from the test set of Houses3K. (b) Unseen buildings from OmniObject3D. It's quite obvious that some parts of the model reconstructed by Scan-RL are wrong or missing. For example, the second object in the first row has a pillar in the wrong shape. Scan-RL fails to reconstruct the roof edge for the fourth object from OmniObject3D, as shown in the third row." + ], + "image_footnote": [], + "bbox": [ + 509, + 95, + 880, + 277 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "formance of our NBV policy on the Houses3K test set; 2) We generalize the policy trained on Houses3K to the house category from OmniObject3D that has completely different geometric structures and textures compared to the training set. The quantitative result is at Table 1 and the visualization result is as shown in Fig. 3; 3) We also generalize GenNBV to non-house categories from OmniObject3D and scenes with enormous details from Objaverse and Replica [35] to demonstrate its generalizability potential.", + "bbox": [ + 75, + 378, + 472, + 516 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We implement the following policies as our baselines. 1) Random: This policy randomly generates 5-dim vector $(x,y,z,\\text{pitch},\\text{yaw})$ among the action space as the next viewpoint. 2) Random Hemisphere: This policy randomly generates the next positions on a pre-defined hemisphere that sufficiently covers all objects of the test set. The headings are constrained to point to the center of the hemisphere. 3) Uniform Hemisphere: All positions are evenly distributed on the previously mentioned hemispheres. 4) Uncertainty-Guided Policy [14]: This NBV policy iteratively selects the next view from a pre-defined viewpoint set according to the uncertainty based on a continually optimized neural radiance field. 5) ActiveRMap [46]: This policy also adopts an iterative selection framework, with multiple objectives including information gain, to select the next viewpoints from the candidate set. 6) Scan-RL [26]: This RL-based policy predicts the next viewpoint from a hemisphere space, relying on the historical RGB images. 7) Ours with Scan-RL's Representation: This free-space NBV policy replaces our representation with Scan-RL's representation only extracted from RGB images.", + "bbox": [ + 75, + 518, + 472, + 837 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Table 1, our GenNBV shows the best in-distribution and out-of-distribution generalizability in both coverage sufficiency and view efficiency, when evaluated on test sets consisting of unseen objects from Houses3K", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/244c6c11162acbb41304432ebcc17153ccfe7736e2b67f290985c5cf2dc9f7f2.jpg", + "table_caption": [ + "Table 3. Ablation studies of representation categories in our framework on unseen Houses3K test set." + ], + "table_footnote": [], + "table_body": "
Representation CategoryEvaluation Metrics
Probabilistic 3D Grid5-DoF PoseSemantic 2D MapMean AUCFinal Coverage Ratio
Uni-source✓81.06%84.56%
✓69.53%76.61%
✓81.24%87.90%
Multi-source✓✓88.66%96.67%
✓✓89.77%95.31%
✓✓88.30%96.29%
✓✓✓91.19%98.26%
", + "bbox": [ + 503, + 416, + 890, + 616 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and OmniObject3D. In particular, the comparison in Table 1 demonstrates the strength of the proposed representations and action space. Note that the Scan-RL's representation (\\* in the table) uses 6 frames, while our semantic representation only needs 2 frames. Moreover, learning-based NBV policies such as GenNBV and Scan-RL, with much larger action space, outperform all rule-based baselines, even if ActiveRMap and Uncertainty-Guided baselines are trained and evaluated on the same 3D objects.", + "bbox": [ + 496, + 623, + 893, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Non-house Generalization In Table 2, we introduce an experiment to evaluate the generalizability of our GenNBV trained on Houses3K to non-house categories. The targets include 74 animals, 43 trucks and 33 dinosaurs from OmniObject3D, and a challenging indoor scene (Room 0) from Replica. The number of views is set to 20.", + "bbox": [ + 496, + 761, + 893, + 852 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our GenNBV almost outperforms Scan-RL in all metrics. Particularly, on Replica Room 0, which consists of numerous multi-sized objects with complex occlusion, our", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16442", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b0c3e919f63a980a9b583f9aa7e2dde3596478edfd85ffd2421651dcb5de0174.jpg", + "image_caption": [ + "Uncertainty-guided" + ], + "image_footnote": [], + "bbox": [ + 81, + 94, + 334, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/de115d3f6207bdc922bc1c62d4fd16d421e2b26f8d9a89687ba5ce3dbb6c4e90.jpg", + "image_caption": [ + "Scan-RL", + "Figure 4. The visualization results of an unseen 3D outdoor scene with enormous details from Objaverse, reconstructed by Uncertainty-guided, Scan-RL and our model. Compared to the uncertainty-guided method and Scan-RL, the scene reconstructed by our method is more watertight and has fewer holes on the ground and building surface, especially in the region highlighted by the red box." + ], + "image_footnote": [], + "bbox": [ + 348, + 90, + 609, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/80d0874bac5a1455c60639ed88a12461439c227ea43b987e895652f906d81521.jpg", + "image_caption": [ + "GenNBV (Ours)" + ], + "image_footnote": [], + "bbox": [ + 627, + 90, + 890, + 244 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "GenNBV is significantly better than Scan-RL. This shows the effectiveness of our free-space NBV policy.", + "bbox": [ + 75, + 356, + 468, + 386 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 397, + 230, + 412 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We implement the ablation study about multi-source state embedding and depth-based representation. Please refer to the Appendix for further experiments.", + "bbox": [ + 75, + 421, + 468, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multi-source State Embedding. In Sec. 3.2, we introduce how to build our multi-source state embedding. Thanks to the same shape of representations and state embedding, we only need to adjust the input dimension of the linear layer of state embedding according to the modal type when ablating the model. Here we reveal the importance of specific representations with the ablation results shown in Table 3. In uni-source experiments, we design the 5-DoF Pose baseline that learns an empirically optimal trajectory, which helps understand the effectiveness of domain knowledge of the dataset. We also evaluate different combinations of representation categories in Table 3, demonstrating the effectiveness of our multi-source state embedding for policy learning in multi-source experiments.", + "bbox": [ + 75, + 465, + 470, + 678 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b4cc58cbf9dd695a50ab5cf5ca4402ec145e236b3b13ce75833be940143aaa01.jpg", + "table_caption": [ + "Table 4. Ablation studies of depth-based representations on unseen Houses3K test set, where depth map is the only sensory source." + ], + "table_footnote": [], + "table_body": "
Depth-based RepresentationMean AUCFinal Coverage Ratio
Geometric 2D Map71.86%74.88%
Binary 3D Grid77.28%80.85%
Probabilistic 3D Grid81.06%84.56%
", + "bbox": [ + 81, + 729, + 468, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Depth-based Representation. Furthermore, we ablate the depth-based representations in Table 4. The experimental results demonstrate that our probabilistic 3D grid is more comprehensive than a binary 3D grid and a geometric 2D map to support the NBV prediction.", + "bbox": [ + 75, + 825, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Qualitative Results", + "text_level": 1, + "bbox": [ + 500, + 354, + 681, + 371 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We visualize the reconstruction results generated from the scanning trajectory of a single episode in Fig. 3. It demonstrates that our next-best-view policy can reconstruct objects better in terms of completeness and appearance quality compared to Scan-RL. We further visualize the reconstruction results of an outdoor scene from Objaverse using 30 collected views in Fig. 4.", + "bbox": [ + 496, + 378, + 892, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 498, + 617, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This study presents an end-to-end approach for active 3D scene reconstruction, reducing the need for manual intervention. Specifically, the learning-based policy explores how to reconstruct diverse objects in the training stage and thus can generalize to reconstruct unseen objects in a fully autonomous manner. Our controller maneuvers in free space and selects the next best view based on a hybrid scene representation which conveys scene coverage status and thus reconstruction progress. We show the effectiveness of our approach by generalizing it on multiple datasets. The quantitative and qualitative generalization results on holdout Houses3K test set and cross-domain OmniObject3D, including house category and non-house categories, show that our method outperforms other baselines in terms of reconstruction completeness, efficiency and accuracy. Furthermore, the experiment conducted on Objaverse and Replica shows that the policy trained in house-only settings can even generalize to complicated outdoor and indoor scenes.", + "bbox": [ + 496, + 523, + 892, + 795 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 809, + 666, + 825 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is supported by Shanghai Artificial Intelligence Laboratory and CUHK Direct Grants (RCFUS) 4055189. We're very grateful to Tao Huang, Zeqi Xiao, Junfeng Long, Mulin Yu, and all reviewers for their insightful advice.", + "bbox": [ + 496, + 834, + 890, + 893 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16443", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Michal Adamkiewicz, Timothy Chen, Adam Caccavale, Rachel Gardner, Preston Culbertson, Jeannette Bohg, and Mac Schwager. Vision-only robot navigation in a neural radiance world. IEEE Robotics and Automation Letters, 7(2): 4606-4613, 2022. 3", + "[2] Jack E Bresenham. Algorithm for computer control of a digital plotter. IBM Systems journal, 4(1):25-30, 1965. 4", + "[3] Devendra Singh Chaplot, Dhiraj Gandhi, Saurabh Gupta, Abhinav Gupta, and Ruslan Salakhutdinov. Learning to explore using active neural slam. In International Conference on Learning Representations, 2020. 1, 2, 3, 4", + "[4] Tao Chen, Saurabh Gupta, and Abhinav Gupta. Learning exploration policies for navigation. In International Conference on Learning Representations, 2019. 1, 2, 3", + "[5] Karl Cobbe, Chris Hesse, Jacob Hilton, and John Schulman. Leveraging procedural generation to benchmark reinforcement learning. In International conference on machine learning, pages 2048-2056. PMLR, 2020. 3", + "[6] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 2, 5", + "[7] Mustafa Devrim Kaba, Mustafa Gokhan Uzunbas, and Ser Nam Lim. A reinforcement learning approach to the view planning problem. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6933-6941, 2017. 3", + "[8] Wojciech Giernacki, Mateusz Skwierczyński, Wojciech Witwicki, Paweł Wronski, and Piotr Kozierski. Crazyflie 2.0 quadrotor as a platform for research and education in robotics and control engineering. In 2017 22nd International Conference on Methods and Models in Automation and Robotics (MMAR), pages 37-42. IEEE, 2017. 3, 5", + "[9] Antoine Guedon, Pascal Monasse, and Vincent Lepetit. SCONE: Surface coverage optimization in unknown environments by volumetric integration. In Advances in Neural Information Processing Systems, 2022. 1, 2, 3, 5, 6", + "[10] Junfu Guo, Changhao Li, Xi Xia, Ruizhen Hu, and Ligang Liu. Asynchronous collaborative autoscanning with mode switching for multi-robot scene reconstruction. ACM Transactions on Graphics (TOG), 41(6):1-13, 2022. 3, 4", + "[11] Guillaume Hardouin, Julien Moras, Fabio Morbidi, Julien Marzat, and El Mustapha Mouaddib. Next-Best-View planning for surface reconstruction of large-scale 3d environments with multiple uavs. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1567–1574. IEEE, 2020. 1, 3", + "[12] Stefan Isler, Reza Sabzevari, Jeffrey Delmerico, and Davide Scaramuzza. An information gain formulation for active volumetric 3d reconstruction. In 2016 IEEE International Conference on Robotics and Automation (ICRA), pages 3477-3484. IEEE, 2016. 2", + "[13] Michael Laskin, Aravind Srinivas, and Pieter Abbeel. Curl: Contrastive unsupervised representations for reinforcement" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "learning. In International Conference on Machine Learning, pages 5639-5650. PMLR, 2020. 3", + "[14] Soomin Lee, Le Chen, Jiahao Wang, Alexander Liniger, Suryansh Kumar, and Fisher Yu. Uncertainty guided policy for active robotic 3d reconstruction using neural radiance fields. IEEE Robotics and Automation Letters, 7(4):12070-12077, 2022. 1, 2, 5, 6, 7", + "[15] Quanyi Li, Zhenghao Peng, Lan Feng, Chenda Duan, Wenjie Mo, Bolei Zhou, et al. ScenarioNet: Open-source platform for large-scale traffic scenario simulation and modeling. In Advances in Neural Information Processing Systems, 2022. 3", + "[16] Quanyi Li, Zhenghao Peng, Lan Feng, Qihang Zhang, Zhenghai Xue, and Bolei Zhou. MetaDrive: Composing diverse driving scenarios for generalizable reinforcement learning. IEEE transactions on pattern analysis and machine intelligence, 45(3):3461-3475, 2022. 3", + "[17] Yixuan Li, Lihan Jiang, Linning Xu, Yuanbo Xiangli, Zhenzhi Wang, Dahua Lin, and Bo Dai. Matrixcity: A large-scale city dataset for city-scale neural rendering and beyond. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3205-3215, 2023. 1", + "[18] Yilin Liu, Liqiang Lin, Yue Hu, Ke Xie, Chi-Wing Fu, Hao Zhang, and Hui Huang. Learning reconstructability for drone aerial path planning. ACM Transactions on Graphics (TOG), 41(6):1-17, 2022. 1", + "[19] Viktor Makoviychuk, Lukasz Wawrzyniak, Yunrong Guo, Michelle Lu, Kier Storey, Miles Macklin, David Hoeller, Nikita Rudin, Arthur Allshire, Ankur Handa, and Gavriel State. Isaac Gym: High performancegpu-based physics simulation for robot learning, 2021. 2, 5", + "[20] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 405-421. Springer, 2020. 1, 2", + "[21] Joseph Ortiz, Alexander Clegg, Jing Dong, Edgar Sucar, David Novotny, Michael Zollhoefer, and Mustafa Mukadam. iSDF: Real-time neural signed distance fields for robot perception. Robotics: Science and Systems, 2022. 2", + "[22] Xuran Pan, Zihang Lai, Shiji Song, and Gao Huang. ActiveNeRF: Learning where to see with uncertainty estimation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIII, pages 230-246. Springer, 2022. 1, 2", + "[23] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019. 1, 2", + "[24] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. PyTorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16444", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Daryl Peralta, Joel Casimiro, Aldrin Michael Nilles, Justine Aletta Aguilar, Rowel Atienza, and Rhandley Cajote. Houses3K dataset. https://github.com/darylperalta/Houses3K, 2020.2.5", + "[26] Daryl Peralta, Joel Casimiro, Aldrin Michael Nilles, Justine Aletta Aguilar, Rowel Atienza, and Rhandley Cajote. Next-best view policy for 3d reconstruction. In Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16, pages 558-573. Springer, 2020. 1, 2, 3, 4, 6, 7", + "[27] Antonin Raffin, Ashley Hill, Adam Gleave, Anssi Kanervisto, Maximilian Ernestus, and Noah Dormann. Stable-baselines3: Reliable reinforcement learning implementations. The Journal of Machine Learning Research, 22(1): 12348-12355, 2021. 6", + "[28] Yunlong Ran, Jing Zeng, Shibo He, Jiming Chen, Lincheng Li, Yingfeng Chen, Gimhee Lee, and Qi Ye. NeurAR: Neural uncertainty for autonomous 3d reconstruction with implicit neural representations. IEEE Robotics and Automation Letters, 2023. 1, 2", + "[29] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv:2007.08501, 2020.6", + "[30] Antoni Rosinol, John J Leonard, and Luca Carlone. NeRF-SLAM: Real-time dense monocular slam with neural radiance fields. arXiv preprint arXiv:2210.13641, 2022. 2", + "[31] Nikita Rudin, David Hoeller, Philipp Reist, and Marco Hutter. Learning to walk in minutes using massively parallel deep reinforcement learning. In 5th Annual Conference on Robot Learning, 2021. 6", + "[32] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 2, 5", + "[33] James A Sethian. A fast marching level set method for monotonically advancing fronts. proceedings of the National Academy of Sciences, 93(4):1591-1595, 1996. 2", + "[34] Edward J Smith, Michal Drozdal, Derek Nowrouzezahrai, David Meger, and Adriana Romero-Soriano. Uncertainty-driven active vision for implicit scene reconstruction. arXiv preprint arXiv:2210.00978, 2022. 2", + "[35] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 5, 7", + "[36] Edgar Sucar, Shikun Liu, Joseph Ortiz, and Andrew J Davison. iMAP: Implicit mapping and positioning in real-time. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6229-6238, 2021. 2", + "[37] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 1", + "[38] Jiaming Sun, Yiming Xie, Linghao Chen, Xiaowei Zhou, and Hujun Bao. NeuralRecon: Real-time coherent 3d reconstruction from monocular video. In Proceedings of the IEEE/CVF" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Computer Vision and Pattern Recognition, pages 15598-15607, 2021. 1", + "[39] Jiaming Sun, Xi Chen, Qianqian Wang, Zhengqi Li, Hadar Averbuch-Elor, Xiaowei Zhou, and Noah Snavely. Neural 3D reconstruction in the wild. In SIGGRAPH Conference Proceedings, 2022. 1", + "[40] Sebastian Thrun. Probabilistic robotics. Communications of the ACM, 45(3):52-57, 2002. 4", + "[41] Josh Tobin, Rachel Fong, Alex Ray, Jonas Schneider, Wojciech Zaremba, and Pieter Abbeel. Domain randomization for transferring deep neural networks from simulation to the real world. In 2017 IEEE/RSJ international conference on intelligent robots and systems (IROS), pages 23-30. IEEE, 2017. 3", + "[42] Tong Wu, Jiarui Zhang, Xiao Fu, Yuxin Wang, Jiawei Ren, Liang Pan, Wayne Wu, Lei Yang, Jiaqi Wang, Chen Qian, et al. Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 803-814, 2023. 2, 5", + "[43] Yuanbo Xiangli, Linning Xu, Xingang Pan, Nanxuan Zhao, Anyi Rao, Christian Theobalt, Bo Dai, and Dahua Lin. BungeeNeRF: Progressive neural radiance field for extreme multi-scale scene rendering. In The European Conference on Computer Vision (ECCV), 2022. 1", + "[44] Kai Ye, Siyan Dong, Qingnan Fan, He Wang, Li Yi, Fei Xia, Jue Wang, and Baoquan Chen. Multi-robot active mapping via neural bipartite graph matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14839–14848, 2022. 3, 4", + "[45] Cem Yuksel. Sample elimination for generating poisson disk sample sets. Computer Graphics Forum, 34(2):25-32, 2015. 6", + "[46] Huangying Zhan, Jiyang Zheng, Yi Xu, Ian Reid, and Hamid Rezatofighi. ActiveRMAP: Radiance field for active mapping and planning. arXiv preprint arXiv:2211.12656, 2022. 1, 2, 5, 6, 7", + "[47] Han Zhang, Yuong Yao, Ke Xie, Chi-Wing Fu, Hao Zhang, and Hui Huang. Continuous aerial path planning for 3d urban scene reconstruction. ACM Trans. Graph., 40(6):225-1, 2021. 1", + "[48] Xiaoshuai Zhang, Sai Bi, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. NeRFusion: Fusing radiance fields for large-scale scene reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5449-5458, 2022. 1", + "[49] Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3D: A modern library for 3D data processing. arXiv:1801.09847, 2018.6", + "[50] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. Nice-SLAM: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12786-12796, 2022. 2" + ], + "bbox": [ + 501, + 92, + 893, + 867 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16445", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/GenNBV_ Generalizable Next-Best-View Policy for Active 3D Reconstruction/171b2e0e-83b0-4811-a381-9c85825e51ab_model.json b/2024/GenNBV_ Generalizable Next-Best-View Policy for Active 3D Reconstruction/171b2e0e-83b0-4811-a381-9c85825e51ab_model.json new file mode 100644 index 0000000000000000000000000000000000000000..431fd73de8987129f90ecc874a5af3bc3fa1a96f --- /dev/null +++ b/2024/GenNBV_ Generalizable Next-Best-View Policy for Active 3D Reconstruction/171b2e0e-83b0-4811-a381-9c85825e51ab_model.json @@ -0,0 +1,2079 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.099, + 0.131, + 0.872, + 0.152 + ], + "angle": 0, + "content": "GenNBV: Generalizable Next-Best-View Policy for Active 3D Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.178, + 0.179, + 0.8, + 0.201 + ], + "angle": 0, + "content": "Xiao Chen\\(^{1,2}\\) Quanyi Li\\(^{1}\\) Tai Wang\\(^{1}\\) Tianfan Xue\\(^{2,\\dagger}\\) Jiangmiao Pang\\(^{1,\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.224, + 0.202, + 0.762, + 0.221 + ], + "angle": 0, + "content": "\\(^{1}\\)Shanghai AI Laboratory \\(^{2}\\)The Chinese University of Hong Kong" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.223, + 0.892, + 0.239 + ], + "angle": 0, + "content": "{cx123,tfxue}@ie.cuhk.edu.hk quanyili0057@gmail.com {wangtai,pangjiangmiao}@pjlab.org.cn" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.255, + 0.437, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.451, + 0.256, + 0.877, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.419, + 0.893, + 0.462 + ], + "angle": 0, + "content": "Figure 1. To determine the best view for 3D reconstruction, previous methods only chose from hand-crafted action space or based on object-centric capturing, lacking the ability to generalize to unforeseen scenes (Left). With our end-to-end trained, generalized free-space policy, it can generalize to unseen objects, enabling the captured drone to image from any viewpoint (Right)." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.473, + 0.314, + 0.489 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.499, + 0.474, + 0.846 + ], + "angle": 0, + "content": "While recent advances in neural radiance field enable realistic digitization for large-scale scenes, the image-capturing process is still time-consuming and labor-intensive. Previous works attempt to automate this process using the Next-Best-View (NBV) policy for active 3D reconstruction. However, the existing NBV policies heavily rely on hand-crafted criteria, limited action space, or perscense optimized representations. These constraints limit their cross-dataset generalizability. To overcome them, we propose GenNBV, an end-to-end generalizable NBV policy. Our policy adopts a reinforcement learning (RL)-based framework and extends typical limited action space to 5D free space. It empowers our agent drone to scan from any viewpoint, and even interact with unseen geometries during training. To boost the cross-dataset generalizability, we also propose a novel multi-source state embedding, including geometric, semantic, and action representations. We establish a benchmark using the Isaac Gym simulator with the Houses3K and OmniObject3D datasets to evaluate this NBV policy. Experiments demonstrate that our policy achieves a \\(98.26\\%\\) and \\(97.12\\%\\) coverage ratio on unseen building-scale objects from these datasets, respectively, outperforming prior solutions." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.473, + 0.631, + 0.489 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.499, + 0.892, + 0.652 + ], + "angle": 0, + "content": "Recent advances in 3D reconstruction [23, 38, 39] and neural rendering [20, 37, 48] have significantly enhanced the quality of 3D digitization of large scenes, such as buildings and city landmarks [11, 17, 18, 43, 47]. However, image-capturing process still remains time-consuming and labor-intensive. To scan a building-scale scene using a drone, it may take several days' effort of a professional team to ensure a full coverage. Moreover, even professional pilots may miss some areas at the first scan, leading to multiple rounds for rescanning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.655, + 0.893, + 0.867 + ], + "angle": 0, + "content": "To alleviate the effort in manual scanning, active 3D reconstruction algorithms, especially the Next-Best-View (NBV) policy, have emerged as a promising approach to automate view planning. However, typical rule-based NBV policies [9, 14, 22, 26, 28, 46] heavily rely on empirically designed view-selection criteria and hand-crafted action space, such as a hemisphere as shown in Fig. 1, which limit their generalizability to unseen scenes. To enhance generalization capacity, pioneering works on learning-based NBV policy have been proposed [3, 4, 26, 46]. Nevertheless, they are constrained into impractical setups in real-world scenarios, such as object-centric capturing and limited action space for ground robots, resulting in incomplete 3D reconstruction because of significant self-occlusion." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Therefore, in this work, we study this problem: \"How to design an NBV policy that supports exploration in any" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.262, + 0.888 + ], + "angle": 0, + "content": "*Project website: gennbv.tech" + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.888, + 0.227, + 0.9 + ], + "angle": 0, + "content": "† Corresponding authors." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.871, + 0.262, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16436" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.273 + ], + "angle": 0, + "content": "space and also can generalize to unseen building-scale geometries?\" Building upon the design of previous reinforcement learning-based (RL-based) NBV policies [3, 4, 26], our setup faces the following new challenges that were not fully studied before: 1) Designing an easy-to-generalize action space, rather than object-specific spaces used in previous work, such as hand-crafted candidates or a constrained hemisphere; 2) With a large action space, an informative representation is necessary to guide the policy to efficiently find an optimal capturing trajectory; 3) Proposing generic reward functions and developing the distributed training procedure for better generalizability." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.276, + 0.473, + 0.518 + ], + "angle": 0, + "content": "To overcome these two challenges, we propose GenNBV, an end-to-end generalizable NBV policy. It has a novel design of action spaces and state embeddings, which allows free space exploration and ensures a large coverage when applying to unseen objects. We firstly extend former limited action space, like a hemisphere, to 5D free space. As shown in Fig. 1, this new space is composed of a position subspace of approximately \\(20m \\times 20m \\times 10m\\) and an omnidirectional heading subspace. This free space enables our agent drone to scan from any viewpoint. Moreover, with much larger action space, our RL-based policy can be easily generalized to building-scale geometries without concerning the potential scale shift from training to evaluation. On the contrary, many hand-crafted designs [14, 22] limit available viewpoints to a small predefined set, and thus cannot generalize if there is drastic change in scales." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.52, + 0.473, + 0.793 + ], + "angle": 0, + "content": "Second, in order to generalize a well-trained NBV policy to unseen scenes during evaluation, we propose a novel state representation directly from raw sensory inputs to guide the policy. In contrast to the widely-used neural radiance field (NeRF) [14, 22, 28, 34, 46], our representation does not need time-consuming per-scene optimization. Specifically, our NBV policy incorporates a multi-source state embedding, which comprises three key components: a novel geometric embedding, a semantic embedding, and an action embedding. The geometric embedding is derived from multi-view depth maps and represents an encoded probabilistic 3D grid, while the semantic embedding is extracted from RGB images, and the action embedding is an encoded viewpoint sequence. In contrast to 2D representations used in previous RL-based NBV policies [3, 4], our multi-source state embedding, extracted from a probabilistic 3D grid, is a more comprehensive representation that supports a robust prediction of next viewpoints." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.473, + 0.903 + ], + "angle": 0, + "content": "To validate the effectiveness of GenNBV, we construct a benchmark for training and evaluation with Houses3K [25] dataset from the NVIDIA Isaac Gym [19] simulator. We further test its generalization ability on novel buildings and object categories from OmniObject3D [42] and Objaverse [6] datasets. Our evaluation metrics quantify the completeness, accuracy, and efficiency for active 3D re" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.213 + ], + "angle": 0, + "content": "construction. To consolidate completeness and efficiency into a single number, we also propose a metric, Area Under the Curve (AUC) of the coverage. Compared with heuristic [12], information gain-based [12, 14, 46], and RL-based baselines [26], our method achieves the best result under different metrics. We also provide visualizations to demonstrate the generalizability of GenNBV. A demo video is attached in the appendix to further illustrate our method." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.226, + 0.642, + 0.243 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.252, + 0.895, + 0.432 + ], + "angle": 0, + "content": "Traditional 3D Reconstruction. Photometry and geometry are two crucial aspects of reconstruction evaluation. Neural implicit representations [20, 23] have shown progress in photometric rendering but faces challenges like time-consuming optimization and poor generalization, hindering its application in real-time reconstruction scenarios like 3D SLAM [21, 30, 36, 50]. Geometry, in contrast, typically corresponds to 3D representations such as point clouds and 3D mesh and is directly related to issues like collision avoidance. Considering these properties, researchers tend to focus on geometric reconstruction in practical applications and we also follow this stream." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.433, + 0.895, + 0.553 + ], + "angle": 0, + "content": "Active 3D Reconstruction. Active 3D reconstruction is a promising field that has not been thoroughly benchmarked yet. The pipeline of active 3D reconstruction frameworks alternates between inferring optimal viewpoints, capturing new data, and updating the rebuilt 3D model. With the optimal sequence of key viewpoints, a continuous planning path can be generated using the classic planners, such as Fast Marching Method (FMM) [33]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.896, + 0.719 + ], + "angle": 0, + "content": "Existing works can be pivotally differentiated based on the paradigm of NBV policies: rule-based or learning-based. Rule-based approaches like uncertainty-driven works [12, 14, 28, 46] typically yield the next best view from hand-crafted rules from scene representations, which tends to overfit specific scenes. Most learning-based policies [3, 4, 26] use a deep reinforcement learning algorithm like PPO [32] to sequentially predict the optimal viewpoints based on observation. They must obtain feedback from task-related rewards such as coverage ratio [9, 26] and optimize during massive iteration with task environment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.896, + 0.87 + ], + "angle": 0, + "content": "The action space is designed based on the paradigm of NBV policies. Most rule-based NBV policies select views from a limited action space, such as a set composed of only one hundred candidate viewpoints [22], a tiny pre-defined space like a hemisphere [9, 14, 28, 46], making it possible to overlook some important viewpoints due to unavailability. Even though learning-based categories further explore the larger action space like a 2D plain [3, 4] or a constrained 3D space [26], their limited action space still prevents them from capturing sufficient details for 3D reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Scene representation built from history observations, which directly provide reconstruction progress to NBV" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16437" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.093, + 0.885, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.289, + 0.893, + 0.348 + ], + "angle": 0, + "content": "Figure 2. Overview of our proposed framework GenNBV. Our end-to-end policy takes the historical multi-source observations as input, transforms them into a more informative scene representation, and produces the next viewpoint position. A reward signal will be returned at training time to optimize the end-to-end policy for maximizing the expected cumulative reward in one episode. Specifically, the signal is the increased coverage ratio after collecting a new viewpoint." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.359, + 0.47, + 0.48 + ], + "angle": 0, + "content": "policies, is also a crucial aspect of the active 3D reconstruction framework. Previous works have explored visual representations for NBV policies, such as TSDF [11], neural radiance field [1], and 2D BEV maps [3, 10, 44]. However, implicit representations are hard to jointly with learning-based NBV policies, and 2D BEV map lacks sufficient information for large-scale outdoor scenes that contain numerous geometric details in 3D space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.486, + 0.471, + 0.653 + ], + "angle": 0, + "content": "In our benchmark, we select Scan-RL [26] as our main RL-based baseline. Compared to it, we introduce a much larger action space and informative visual representations for better generalizability. Additionally, we don't benchmark the mentioned baselines [3, 4], mainly due to their reliance on pretraining from human demonstration using imitation learning, while our focus lies on learning through iterations and end-to-end training. Different from our 3D free-space movement, they only explore the traversable 2D areas of indoor scenes using ground robots, thus limiting their motion space in 2D." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Generalizable Reinforcement Learning. Increasing the diversity of training data leads to better generalizability [5]. In the robotics field, domain randomization [41] allows legged robots to walk on various terrains absent from the training environment. For end-to-end driving policy, training in large-scale synthetic and realistic scenarios improves the safety of autonomous cars [15, 16] in unseen test scenarios. The advanced work [7] presents an RL-based formulation for the view planning problem. Based on similar frameworks of active 3D reconstruction, Scan-RL [26] and SCONE [9] are two pioneering works showing the generalizability of their frameworks, while both of them suffer from the constrain of viewpoint sampling in limited space and lack of in-depth analysis. In this work, we tackle the aforementioned issues by predicting NBV in free 3D space with a learning-based planner, which is trained with a" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.359, + 0.892, + 0.391 + ], + "angle": 0, + "content": "dataset containing buildings in various shapes and poses for acquiring generalizability for unseen buildings." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.402, + 0.636, + 0.42 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.427, + 0.893, + 0.64 + ], + "angle": 0, + "content": "In this section, we deliver our active 3D reconstruction framework GenNBV, especially the pivotal Next-Best-View policy. An overview of our framework is illustrated in Fig. 2. Firstly, we formulate the NBV problem as a Markov Decision Process (MDP), with a novel design of observations (blue boxes in Fig. 2) and action space in Sec. 3.1. Next, we elaborate our end-to-end NBV policy \\(\\pi\\) (orange box in Fig. 2) in Sec. 3.2. Inspired by Curl [13], we point out that generalizability greatly depends on how to extract embeddings (green box in Fig. 2), which reflect the reconstruction progress, from raw sensory observations like RGB images and camera poses. In Sec. 3.3, we introduce the reward function (right-hand side of Fig. 2) reflecting the optimization objective and the details of policy optimization." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.651, + 0.874, + 0.667 + ], + "angle": 0, + "content": "3.1. Formulation of the Next-Best-View Problem" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.892, + 0.81 + ], + "angle": 0, + "content": "We formulate the NBV problem as learning an optimal policy \\(\\pi\\) that controls the capturing process, such that enough information is captured for large-scale scene reconstruction, with limited decision-making budgets. As capturing, transferring, and processing a large set of captured images introduce significant computational costs, we also want to design a policy that also minimizes the number of captured images. Therefore, our policy only captures sparse keyframes that sufficiently record all details of objects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.893, + 0.901 + ], + "angle": 0, + "content": "As shown in Fig. 2, at each time step \\( t \\), the agent receives a visual observation \\( o_t \\), takes an action \\( a_t \\), infers the action at the next time step \\( t + 1 \\) to move to a new location, and then receives a new visual observation, repeating this interaction with the environment until the episode ends. Our simulated agent is embodied as CrazyFlie [8], a type of unmanned" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16438" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "aerial vehicle equipped with various sensors, including an RGB-D camera and an IMU, to execute data collection for reconstruction. We discussed the details of the observation space and action space below." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.153, + 0.471, + 0.289 + ], + "angle": 0, + "content": "Observation Space. As shown in the left column of Fig. 2, at each time step, the agent receives an RGB image \\(I_{t}\\), a depth map \\(D_{t}\\), and a state vector including the heading (yaw and pitch) and position \\((\\mathrm{x},\\mathrm{y},\\mathrm{z})\\) of the onboard camera. The observation \\(o_{t}\\) (yellow box in Fig. 2) consists of all previously captured images in one episode, historical actions, and the current captures and actions. With this information as input, the policy network can estimate the progress of data collection and determine where to scan next." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.291, + 0.47, + 0.398 + ], + "angle": 0, + "content": "Action Space. Unlike most previous NBV algorithms, we use larger action space in order to cover all details of objects. Specifically, we use the camera location and camera angle as our action space, which is a 5-dimension vector consisting of 3D position coordinates and 2D rotation angles (yaw-axis and pitch-axis). We restrict the roll-axis rotation as it is not supported in all drone platforms." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.407, + 0.358, + 0.423 + ], + "angle": 0, + "content": "3.2. Generalizable State Embedding" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.431, + 0.469, + 0.581 + ], + "angle": 0, + "content": "To ensure that the policy learned on one set of 3D objects can generalize to objects with different appearances and structures, a smart state embedding of raw sensor observations is needed to capture invariant features across different objects. Previous methods [3, 26] only extract representations in 2D space, which are very sensible to appearance changes, and our experiments have shown that policies only trained on 2D features may not generalize to different objects. Instead, we propose a multi-source representation that has better generalizability." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.582, + 0.469, + 0.717 + ], + "angle": 0, + "content": "Specifically, we first build two mid-level representations that can better model the relationship between the objects and our agent: a 3D geometric representation \\( F^G \\) from depth maps and a semantic representation \\( F^S \\) from RGB images. Then we encode these representations and concatenate them with a pose embedding into state embedding \\( s_t \\), guiding the NBV policy for subsequent decision-making. The overall encoder for state embedding is shown in Fig. 2, and details of each representation are discussed below." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.469, + 0.87 + ], + "angle": 0, + "content": "Geometric Representation One simple way to record the geometry of a 3D object is using a binary 3D occupancy map [3], where the value of each cell in the 3D cube indicates whether a cell contains a 3D object or not. However, since the 3D representation is gradually updated with newly captured data, this simple binary occupancy map cannot differentiate a real unoccupied cell from an unscanned cell. An unscanned cell is a strong indicator that the agent should capture more data in this region, while no further scan is needed for a real unoccupied cell." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Previous works [3, 10, 44] simplify the 3D scene representation into a 2D Bird's Eye View map for actively recon" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.348 + ], + "angle": 0, + "content": "structuring indoor scenes. However, the 2D BEV map lacks sufficient information for our large-scale outdoor scenes that contain many geometric details in 3D space. In addition, their wheeled robotic platforms are constrained to freely scan and represent these outdoor scenes. Therefore, to model the scanning process in 3D free space, we employ the probabilistic 3D grid [40] as our geometric 3D representation, which records the probability of each 3D voxel being captured or not. Specifically, we first obtain a 3D point cloud in the world coordinate by back-projecting all 2D pixels to 3D points, using the depth map \\( D_{t} \\), camera intrinsic parameters, and camera pose \\( a_{t} \\). By voxelizing the obtained point cloud, we then build a 3D occupancy grid that explicitly indicates the binary state (occupied or free) in this 3D space. Subsequently, we represent this 3D grid as a probabilistic occupancy grid \\( F_{t}^{G} \\) and extend the state space of voxels with three states (occupied, free, unknown)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.348, + 0.892, + 0.485 + ], + "angle": 0, + "content": "During each scanning (one episode in RL), at each step \\( t + 1 \\), we update the probabilistic occupancy grid \\( F_{t + 1}^{G} \\) based on the preceding grid \\( F_{t}^{G} \\) and current observation. Specifically, the grid is updated through Bresenham's line algorithm [2], which casts the ray path in 3D space between the camera viewpoint and the endpoints among the point cloud back-projected from depth \\( D_{t + 1} \\). Following the classical occupancy grid mapping algorithm [40], we have the log-odds formulation of occupancy probability:" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.495, + 0.891, + 0.512 + ], + "angle": 0, + "content": "\\[\n\\log \\operatorname {O d d} \\left(v _ {i} \\mid z _ {j}\\right) = \\log \\operatorname {O d d} \\left(v _ {i}\\right) + C, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.892, + 0.688 + ], + "angle": 0, + "content": "where \\( v_{i} \\) denotes the occupancy probability of \\( i^{th} \\) voxel in the grid \\( F_{t}^{G} \\), \\( z_{j} \\) is the measurement event that \\( j^{th} \\) camera ray passes through this voxel and \\( C \\) is an empirical constant. The derivation can be found in Appendix. Thus, we update the log-odds occupancy probability of each voxel in the grid \\( F_{t}^{G} \\) by adding a constant one time when a single camera ray passes through this voxel. Note that the probabilistic occupancy grid \\( F^{G} \\) is continually updated within one episode. Finally, the occupancy status of voxels is classified into three categories: unknown, occupied, and free, using preset probability thresholds." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.81 + ], + "angle": 0, + "content": "Semantic Representation. Geometric representation enables agents to comprehend spatial occupancy, yet it's insufficient for perceiving the environment. For example, when observing a hole in an object, the agent may struggle to differentiate between incomplete scanning and the actual presence of a hole in the object. In such cases, the semantic information contained in the captured RGB images can help the agent distinguish between these two scenarios." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To provide semantic information, we employ a preprocessing module that takes as input the current frame of RGB image \\(I_{t}\\) and preceding \\(k\\) frames and converts these frames \\([I_t,I_{t - 1},\\dots,I_{t - k}]\\) to grayscale, and concatenate them as output, following [26]. Then the preprocessed frames are fed into a two-layer convolutional network for extracting" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16439" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.091, + 0.293, + 0.107 + ], + "angle": 0, + "content": "the semantic representation \\(F_{t}^{S}\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.108, + 0.469, + 0.229 + ], + "angle": 0, + "content": "State Embedding. To further combine the semantic and geometric embeddings, we first encode them to \\( s_t^S = f^S (F_t^S) \\) and \\( s_t^G = f^G (F_t^G) \\) where \\( f^{*} \\) are learnable networks Linear(Flatten(x)), as shown in Fig. 2. Subsequently, we combine them with the historical action embedding \\( s_t^A = \\mathrm{Linear}(a_{1:t}) \\) to generate the final state embedding \\( s_t \\), as the input to the policy network. This process can be formulated as: \\( s_t = \\mathrm{Linear}(s_t^G;s_t^S;s_t^A) \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.23, + 0.469, + 0.306 + ], + "angle": 0, + "content": "Policy Network. Taking the state embedding \\( s_t \\) as input, the policy network is a 3-layer multi-layer perceptron network (MLP) whose output is used to parameterize a normal distribution over action space. In this way, the action can be drawn from the stochastic policy \\( a \\sim \\pi(\\cdot | o_t) \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.314, + 0.388, + 0.331 + ], + "angle": 0, + "content": "3.3. Reward Function and Optimization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.337, + 0.469, + 0.413 + ], + "angle": 0, + "content": "We train the end-to-end policy with reinforcement learning (RL) and hence design reward functions to precisely reflect the task objective for 3D reconstruction. The policy is optimized with proximal policy optimization [32] (PPO) for parallelizing sampling." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.414, + 0.469, + 0.49 + ], + "angle": 0, + "content": "Reward Functions. With the occupancy probability \\( F_{t}^{G} \\) at time step \\( t \\), we can threshold each voxel with an empirical bound to determine if it is occupied. This discrimination process outputs a binary occupancy grid with \\( \\tilde{N}_{t} \\) voxels being occupied, which is used to calculate the coverage ratio:" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.502, + 0.469, + 0.533 + ], + "angle": 0, + "content": "\\[\n\\mathrm{CR}_{t} = \\frac{\\tilde{N}_{t}}{N^{*}}\\cdot 100\\% , \\tag{2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.537, + 0.469, + 0.611 + ], + "angle": 0, + "content": "where \\( N^{*} \\) is the number of ground-truth occupied voxels representing the surface of objects. To encourage our NBV policy to cover as many unseen areas of objects as possible, we use the difference of coverage ratio (CR) between two consecutive steps as the main reward function \\( r^{CR} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.196, + 0.621, + 0.469, + 0.641 + ], + "angle": 0, + "content": "\\[\nr _ {t + 1} ^ {\\mathrm {C R}} = \\mathrm {C R} _ {t + 1} - \\mathrm {C R} _ {t}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.65, + 0.469, + 0.77 + ], + "angle": 0, + "content": "In free-space exploration, we also need to avoid collision. Previous limited-space agents [9, 14, 46] do not consider collision avoidance since their search space, like hemisphere, is by-design safe. Thus, we add a negative reward for collision and terminate the episode if a collision happens. We also implement a negative reward when the number of captured keyframes is over an empirical threshold to improve the path efficiency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.772, + 0.469, + 0.862 + ], + "angle": 0, + "content": "Policy Optimization. Once the reward functions has been specified, the policy can be learned through any off-the-shelf RL algorithm. In this work, we use PPO as thousands of workers can be parallelized to improve the sample efficiency. Specifically, given our parameterized policy \\(\\pi_{\\theta}\\), the objective of PPO is to maximize the following function:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.871, + 0.469, + 0.906 + ], + "angle": 0, + "content": "\\[\nL (\\theta) = \\mathbb {E} _ {t} \\left[ \\frac {\\pi_ {\\theta} (a _ {t} | s _ {t})}{\\pi_ {\\theta_ {\\mathrm {o l d}}} (a _ {t} | s _ {t})} A ^ {\\pi_ {\\theta_ {\\mathrm {o l d}}}} (s _ {t}, a _ {t}) \\right], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "where \\( A^{\\pi_{\\theta_{\\mathrm{old}}}}(s_t, a_t) \\) is the advantage function that measures the value of taking action \\( a_t \\) at state \\( s_t \\) under the current policy \\( \\pi_{\\theta_{\\mathrm{old}}} \\). To prevent significant deviation of the new policy from the old policy, PPO incorporates a clipped surrogate objective function:" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.173, + 0.892, + 0.211 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L ^ {\\mathrm {C L I P}} (\\theta) = \\mathbb {E} _ {t} [ \\min (\\eta_ {t} (\\theta) A ^ {\\pi_ {\\theta_ {\\text {o l d}}}} (s _ {t}, a _ {t}), \\tag {5} \\\\ \\left. \\operatorname {c l i p} \\left(r _ {t} (\\theta), 1 - \\epsilon , 1 + \\epsilon\\right) A ^ {\\pi_ {\\theta_ {\\text {o l d}}}} \\left(s _ {t}, a _ {t}\\right) \\right], \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.219, + 0.892, + 0.255 + ], + "angle": 0, + "content": "where \\(\\eta_t(\\theta) = \\frac{\\pi_\\theta(a_t|s_t)}{\\pi_{\\theta_{\\mathrm{old}}}(a_t|s_t)}\\) and \\(\\epsilon\\) is a hyper-parameter that controls the size of the trust region." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.27, + 0.632, + 0.287 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.295, + 0.892, + 0.416 + ], + "angle": 0, + "content": "In this section, we conduct experiments on Houses3K [25], OmniObject3D [42], and Objaverse [6]. For our policy and other learning-based policies, we train them on Houses3K training set. Then we evaluate all policies on the Houses3K test dataset and the OmniObject3D dataset for quantifying the in-distribution and out-of-distribution generalizability. Finally, we show the visualization results on three datasets, demonstrating the effectiveness of the proposed method." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.424, + 0.691, + 0.44 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.447, + 0.892, + 0.658 + ], + "angle": 0, + "content": "Simulation Environment. We conduct all experiments in NVIDIA Isaac Gym [19], a physics simulation platform designed for reinforcement learning and robotics research. With GPU-accelerated tensor API and sensor interaction API, we can easily implement customized CUDA kernels to calculate our geometric representation efficiently. Moreover, the sensor simulation is efficient. It can run up to 1000 FPS, significantly reducing the training time. We create the agent drone CrazyFlie [8] and equip it with sensors including RGB-D cameras and IMU. Considering memory limitations, we downsample the image resolution to \\(400 \\times 400\\). The vertical field of view of this onboard camera is \\(90^{\\circ}\\). We also set up a point light source with a fixed position and constant intensity in Isaac Gym." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.78 + ], + "angle": 0, + "content": "Dataset. We conduct our experiments on Houses3K [25], OmniObject3D [42], Objverse 1.0 [6] and Replica [35]. Our model is trained on large-scale 3D objects from Houses3K. To validate the effectiveness and generalizability of our model, we evaluate it on unseen 3D objects from Houses3K and further on batches of unseen cross-domain 3D objects from OmniObject3D, Objverse, and Replica, which include both house and non-house categories." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Houses3K contains 3,000 textured 3D building models. They are divided into 12 batches, with each batch featuring 50 distinct geometries and 5 varying textures. We further eliminate poorly designed geometries based on the following two criteria: 1) objects with complex internal structures (as we are mostly focusing on the surface reconstruction), and 2) objects with redundant bases at the bottom. Following these criteria, the training set consists of 256 objects" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16440" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.892, + 0.118 + ], + "angle": 0, + "content": "Table 1. Evaluation results of Next-Best-View policies for active 3D reconstruction on Houses3K and the house category from OmniObject3D (cross-dataset generalization). The number of views is set to 30 and 20 for Houses3K and OmniObject3D, respectively." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.119, + 0.892, + 0.146 + ], + "angle": 0, + "content": "\\*\\*: the policy is trained with the Houses3K training set and evaluated with holdout Houses3K test set and OmniObject3D house category. \\(\\ddagger\\) : the policy heavily relies on optimized per-scene representation (NeRF), and thus is directly trained and evaluated on testing objects." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.158, + 0.895, + 0.416 + ], + "angle": 0, + "content": "
Houses3K Test SetOmniObject3D
NBV PolicyMean AUC ↑Final Coverage Ratio ↑Accuracy (cm) ↓Mean AUC ↑Final Coverage Ratio ↑Accuracy (cm) ↓
HeuristicRandom48.53%58.24%1.3860.13%73.73%0.62
Random Hemisphere71.19%79.72%0.6374.03%84.74%0.48
Uniform Hemisphere82.91%89.71%0.4483.09%92.90%0.41
InfoGain†Uncertainty-Guided [14]83.13%89.31%0.4469.08%92.85%0.42
†ActiveRMap [46]84.86%90.77%0.4469.47%93.15%0.42
RL-based*Scan-RL [26]84.49%91.61%0.4082.17%92.53%0.34
*Scan-RL w/ Our Repre.86.48%92.20%0.3786.14%93.73%0.35
*Ours w/ Scan-RL Repre.87.39%95.63%0.3886.81%93.64%0.34
*GenNBV (Ours)91.19%98.26%0.3788.63%97.12%0.33
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.432, + 0.471, + 0.488 + ], + "angle": 0, + "content": "Table 2. The cross-dataset generalization for non-house categories. We train the baseline Scan-RL and our GenNBV on Houses3K and generalize them to non-house categories from OmniObject3D and an indoor scene from Replica." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.499, + 0.469, + 0.653 + ], + "angle": 0, + "content": "
CategoryDatasetNBV PolicyAUC ↑FCR ↑Acc. ↓
AnimalsOmniObject3DScan-RL76.43%84.98%0.17
GenNBV84.28%94.07%0.16
TrucksOmniObject3DScan-RL69.58%76.37%0.06
GenNBV75.54%83.47%0.03
DinosaursOmniObject3DScan-RL78.47%86.89%0.88
GenNBV84.24%94.03%0.81
Room 0ReplicaScan-RL69.21%80.75%3.12
GenNBV88.53%99.16%3.59
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.667, + 0.47, + 0.713 + ], + "angle": 0, + "content": "from 6 selected batches, and the test set consists of 50 objects from another batch. All these selected training and test objects have distinct geometries." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.715, + 0.47, + 0.837 + ], + "angle": 0, + "content": "OmniObject3D offers a collection of 6K high-fidelity objects scanned from real-world sources across 190 typical categories. The house category from OmniObject3D for evaluation has 43 diverse objects. In contrast, Objaverse 1.0 comprises a vast library of 818K 3D synthetic objects spanning 21K categories. The diverse and high-quality nature of these datasets makes them ideal for evaluating our models and visualizing the results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Evaluation Metrics. The objective of NBV policies is to capture the most useful information for reconstruction, with the least number of views. We report the (1) Final Coverage Ratio \\((\\%)\\), (2) Mean AUC \\((\\%)\\) of coverage ratio and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.434, + 0.892, + 0.617 + ], + "angle": 0, + "content": "(3) Reconstruction Accuracy (cm) along with the consistent number of views in all tables. Specifically, most prior works [9, 26] separately use the coverage ratio (\\%) and the number of views to evaluate the reconstruction completeness and efficiency for NBV policies. However, the coverage ratio is highly correlated with the number of views. Thus we propose to unify the number of views to a fixed value for all NBV policies during evaluation and use the area under the curve (AUC) of coverage ratio as the main metric for comparison. Also, we calculate the Chamfer Distance between scanned and ground-truth point clouds to represent the reconstruction accuracy (cm)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.617, + 0.893, + 0.829 + ], + "angle": 0, + "content": "Implementation Details. We conduct all experiments in Isaac Gym simulation engine with one NVIDIA Tesla V100 GPU. Our NBV policy is optimized through over 32 million iterations and uses approximately 24 hours of training time on an NVIDIA V100 GPU. All networks are randomly initialized and trained end-to-end. Our implementation refers to the codebase of Legged Gym [31] and the PPO implementation in Stable Baseline3 [27], which is implemented by PyTorch [24]. The ground-truth point clouds on objects' surfaces are sampled by the Poisson Disk sampling method [45] using Open3D API [49]. The Chamfer Distance between the scanned point cloud and the ground-truth point cloud is calculated using PyTorch3D API [29]. Please refer to the Appendix for further details and experiments." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.835, + 0.736, + 0.851 + ], + "angle": 0, + "content": "4.2. Performance Comparison" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.903 + ], + "angle": 0, + "content": "To comprehensively demonstrate the effectiveness and generalizability of GenNBV, we design three levels of evaluation experiments: 1) As shown in Table 1, we show the per" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16441" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.088, + 0.496, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.289, + 0.281, + 0.39, + 0.296 + ], + "angle": 0, + "content": "(a) Houses3K" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.097, + 0.882, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.636, + 0.281, + 0.766, + 0.296 + ], + "angle": 0, + "content": "(b) OmniObject3D" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.308, + 0.895, + 0.366 + ], + "angle": 0, + "content": "Figure 3. The visualization results of unseen 3D objects reconstructed by Scan-RL [26] and our model to compare the generalizability. (a) Unseen buildings from the test set of Houses3K. (b) Unseen buildings from OmniObject3D. It's quite obvious that some parts of the model reconstructed by Scan-RL are wrong or missing. For example, the second object in the first row has a pillar in the wrong shape. Scan-RL fails to reconstruct the roof edge for the fourth object from OmniObject3D, as shown in the third row." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.38, + 0.473, + 0.517 + ], + "angle": 0, + "content": "formance of our NBV policy on the Houses3K test set; 2) We generalize the policy trained on Houses3K to the house category from OmniObject3D that has completely different geometric structures and textures compared to the training set. The quantitative result is at Table 1 and the visualization result is as shown in Fig. 3; 3) We also generalize GenNBV to non-house categories from OmniObject3D and scenes with enormous details from Objaverse and Replica [35] to demonstrate its generalizability potential." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.52, + 0.473, + 0.838 + ], + "angle": 0, + "content": "We implement the following policies as our baselines. 1) Random: This policy randomly generates 5-dim vector \\((x,y,z,\\text{pitch},\\text{yaw})\\) among the action space as the next viewpoint. 2) Random Hemisphere: This policy randomly generates the next positions on a pre-defined hemisphere that sufficiently covers all objects of the test set. The headings are constrained to point to the center of the hemisphere. 3) Uniform Hemisphere: All positions are evenly distributed on the previously mentioned hemispheres. 4) Uncertainty-Guided Policy [14]: This NBV policy iteratively selects the next view from a pre-defined viewpoint set according to the uncertainty based on a continually optimized neural radiance field. 5) ActiveRMap [46]: This policy also adopts an iterative selection framework, with multiple objectives including information gain, to select the next viewpoints from the candidate set. 6) Scan-RL [26]: This RL-based policy predicts the next viewpoint from a hemisphere space, relying on the historical RGB images. 7) Ours with Scan-RL's Representation: This free-space NBV policy replaces our representation with Scan-RL's representation only extracted from RGB images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.472, + 0.902 + ], + "angle": 0, + "content": "As shown in Table 1, our GenNBV shows the best in-distribution and out-of-distribution generalizability in both coverage sufficiency and view efficiency, when evaluated on test sets consisting of unseen objects from Houses3K" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.378, + 0.892, + 0.406 + ], + "angle": 0, + "content": "Table 3. Ablation studies of representation categories in our framework on unseen Houses3K test set." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.417, + 0.892, + 0.617 + ], + "angle": 0, + "content": "
Representation CategoryEvaluation Metrics
Probabilistic 3D Grid5-DoF PoseSemantic 2D MapMean AUCFinal Coverage Ratio
Uni-source✓81.06%84.56%
✓69.53%76.61%
✓81.24%87.90%
Multi-source✓✓88.66%96.67%
✓✓89.77%95.31%
✓✓88.30%96.29%
✓✓✓91.19%98.26%
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.624, + 0.895, + 0.76 + ], + "angle": 0, + "content": "and OmniObject3D. In particular, the comparison in Table 1 demonstrates the strength of the proposed representations and action space. Note that the Scan-RL's representation (\\* in the table) uses 6 frames, while our semantic representation only needs 2 frames. Moreover, learning-based NBV policies such as GenNBV and Scan-RL, with much larger action space, outperform all rule-based baselines, even if ActiveRMap and Uncertainty-Guided baselines are trained and evaluated on the same 3D objects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.762, + 0.895, + 0.853 + ], + "angle": 0, + "content": "Non-house Generalization In Table 2, we introduce an experiment to evaluate the generalizability of our GenNBV trained on Houses3K to non-house categories. The targets include 74 animals, 43 trucks and 33 dinosaurs from OmniObject3D, and a challenging indoor scene (Room 0) from Replica. The number of views is set to 20." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Our GenNBV almost outperforms Scan-RL in all metrics. Particularly, on Replica Room 0, which consists of numerous multi-sized objects with complex occlusion, our" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16442" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.095, + 0.335, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.131, + 0.258, + 0.265, + 0.272 + ], + "angle": 0, + "content": "Uncertainty-guided" + }, + { + "type": "image", + "bbox": [ + 0.349, + 0.092, + 0.611, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.444, + 0.258, + 0.504, + 0.271 + ], + "angle": 0, + "content": "Scan-RL" + }, + { + "type": "image", + "bbox": [ + 0.629, + 0.092, + 0.891, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.711, + 0.258, + 0.819, + 0.272 + ], + "angle": 0, + "content": "GenNBV (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.288, + 0.893, + 0.331 + ], + "angle": 0, + "content": "Figure 4. The visualization results of an unseen 3D outdoor scene with enormous details from Objaverse, reconstructed by Uncertainty-guided, Scan-RL and our model. Compared to the uncertainty-guided method and Scan-RL, the scene reconstructed by our method is more watertight and has fewer holes on the ground and building surface, especially in the region highlighted by the red box." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.357, + 0.47, + 0.387 + ], + "angle": 0, + "content": "GenNBV is significantly better than Scan-RL. This shows the effectiveness of our free-space NBV policy." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.398, + 0.231, + 0.414 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.422, + 0.47, + 0.466 + ], + "angle": 0, + "content": "We implement the ablation study about multi-source state embedding and depth-based representation. Please refer to the Appendix for further experiments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.467, + 0.471, + 0.679 + ], + "angle": 0, + "content": "Multi-source State Embedding. In Sec. 3.2, we introduce how to build our multi-source state embedding. Thanks to the same shape of representations and state embedding, we only need to adjust the input dimension of the linear layer of state embedding according to the modal type when ablating the model. Here we reveal the importance of specific representations with the ablation results shown in Table 3. In uni-source experiments, we design the 5-DoF Pose baseline that learns an empirically optimal trajectory, which helps understand the effectiveness of domain knowledge of the dataset. We also evaluate different combinations of representation categories in Table 3, demonstrating the effectiveness of our multi-source state embedding for policy learning in multi-source experiments." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.69, + 0.469, + 0.719 + ], + "angle": 0, + "content": "Table 4. Ablation studies of depth-based representations on unseen Houses3K test set, where depth map is the only sensory source." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.731, + 0.47, + 0.809 + ], + "angle": 0, + "content": "
Depth-based RepresentationMean AUCFinal Coverage Ratio
Geometric 2D Map71.86%74.88%
Binary 3D Grid77.28%80.85%
Probabilistic 3D Grid81.06%84.56%
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Depth-based Representation. Furthermore, we ablate the depth-based representations in Table 4. The experimental results demonstrate that our probabilistic 3D grid is more comprehensive than a binary 3D grid and a geometric 2D map to support the NBV prediction." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.356, + 0.683, + 0.372 + ], + "angle": 0, + "content": "4.4. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.379, + 0.893, + 0.485 + ], + "angle": 0, + "content": "We visualize the reconstruction results generated from the scanning trajectory of a single episode in Fig. 3. It demonstrates that our next-best-view policy can reconstruct objects better in terms of completeness and appearance quality compared to Scan-RL. We further visualize the reconstruction results of an outdoor scene from Objaverse using 30 collected views in Fig. 4." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.499, + 0.619, + 0.514 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.893, + 0.796 + ], + "angle": 0, + "content": "This study presents an end-to-end approach for active 3D scene reconstruction, reducing the need for manual intervention. Specifically, the learning-based policy explores how to reconstruct diverse objects in the training stage and thus can generalize to reconstruct unseen objects in a fully autonomous manner. Our controller maneuvers in free space and selects the next best view based on a hybrid scene representation which conveys scene coverage status and thus reconstruction progress. We show the effectiveness of our approach by generalizing it on multiple datasets. The quantitative and qualitative generalization results on holdout Houses3K test set and cross-domain OmniObject3D, including house category and non-house categories, show that our method outperforms other baselines in terms of reconstruction completeness, efficiency and accuracy. Furthermore, the experiment conducted on Objaverse and Replica shows that the policy trained in house-only settings can even generalize to complicated outdoor and indoor scenes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.81, + 0.667, + 0.827 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.835, + 0.892, + 0.895 + ], + "angle": 0, + "content": "This work is supported by Shanghai Artificial Intelligence Laboratory and CUHK Direct Grants (RCFUS) 4055189. We're very grateful to Tao Huang, Zeqi Xiao, Junfeng Long, Mulin Yu, and all reviewers for their insightful advice." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16443" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Michal Adamkiewicz, Timothy Chen, Adam Caccavale, Rachel Gardner, Preston Culbertson, Jeannette Bohg, and Mac Schwager. Vision-only robot navigation in a neural radiance world. IEEE Robotics and Automation Letters, 7(2): 4606-4613, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.185, + 0.47, + 0.213 + ], + "angle": 0, + "content": "[2] Jack E Bresenham. Algorithm for computer control of a digital plotter. IBM Systems journal, 4(1):25-30, 1965. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.213, + 0.471, + 0.269 + ], + "angle": 0, + "content": "[3] Devendra Singh Chaplot, Dhiraj Gandhi, Saurabh Gupta, Abhinav Gupta, and Ruslan Salakhutdinov. Learning to explore using active neural slam. In International Conference on Learning Representations, 2020. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.27, + 0.47, + 0.312 + ], + "angle": 0, + "content": "[4] Tao Chen, Saurabh Gupta, and Abhinav Gupta. Learning exploration policies for navigation. In International Conference on Learning Representations, 2019. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.313, + 0.471, + 0.368 + ], + "angle": 0, + "content": "[5] Karl Cobbe, Chris Hesse, Jacob Hilton, and John Schulman. Leveraging procedural generation to benchmark reinforcement learning. In International conference on machine learning, pages 2048-2056. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.369, + 0.471, + 0.451 + ], + "angle": 0, + "content": "[6] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.453, + 0.471, + 0.521 + ], + "angle": 0, + "content": "[7] Mustafa Devrim Kaba, Mustafa Gokhan Uzunbas, and Ser Nam Lim. A reinforcement learning approach to the view planning problem. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6933-6941, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.523, + 0.471, + 0.606 + ], + "angle": 0, + "content": "[8] Wojciech Giernacki, Mateusz Skwierczyński, Wojciech Witwicki, Paweł Wronski, and Piotr Kozierski. Crazyflie 2.0 quadrotor as a platform for research and education in robotics and control engineering. In 2017 22nd International Conference on Methods and Models in Automation and Robotics (MMAR), pages 37-42. IEEE, 2017. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.607, + 0.471, + 0.662 + ], + "angle": 0, + "content": "[9] Antoine Guedon, Pascal Monasse, and Vincent Lepetit. SCONE: Surface coverage optimization in unknown environments by volumetric integration. In Advances in Neural Information Processing Systems, 2022. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.663, + 0.47, + 0.718 + ], + "angle": 0, + "content": "[10] Junfu Guo, Changhao Li, Xi Xia, Ruizhen Hu, and Ligang Liu. Asynchronous collaborative autoscanning with mode switching for multi-robot scene reconstruction. ACM Transactions on Graphics (TOG), 41(6):1-13, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.719, + 0.471, + 0.802 + ], + "angle": 0, + "content": "[11] Guillaume Hardouin, Julien Moras, Fabio Morbidi, Julien Marzat, and El Mustapha Mouaddib. Next-Best-View planning for surface reconstruction of large-scale 3d environments with multiple uavs. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1567–1574. IEEE, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.803, + 0.471, + 0.871 + ], + "angle": 0, + "content": "[12] Stefan Isler, Reza Sabzevari, Jeffrey Delmerico, and Davide Scaramuzza. An information gain formulation for active volumetric 3d reconstruction. In 2016 IEEE International Conference on Robotics and Automation (ICRA), pages 3477-3484. IEEE, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.872, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[13] Michael Laskin, Aravind Srinivas, and Pieter Abbeel. Curl: Contrastive unsupervised representations for reinforcement" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "learning. In International Conference on Machine Learning, pages 5639-5650. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[14] Soomin Lee, Le Chen, Jiahao Wang, Alexander Liniger, Suryansh Kumar, and Fisher Yu. Uncertainty guided policy for active robotic 3d reconstruction using neural radiance fields. IEEE Robotics and Automation Letters, 7(4):12070-12077, 2022. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.19, + 0.892, + 0.258 + ], + "angle": 0, + "content": "[15] Quanyi Li, Zhenghao Peng, Lan Feng, Chenda Duan, Wenjie Mo, Bolei Zhou, et al. ScenarioNet: Open-source platform for large-scale traffic scenario simulation and modeling. In Advances in Neural Information Processing Systems, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.26, + 0.892, + 0.329 + ], + "angle": 0, + "content": "[16] Quanyi Li, Zhenghao Peng, Lan Feng, Qihang Zhang, Zhenghai Xue, and Bolei Zhou. MetaDrive: Composing diverse driving scenarios for generalizable reinforcement learning. IEEE transactions on pattern analysis and machine intelligence, 45(3):3461-3475, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[17] Yixuan Li, Lihan Jiang, Linning Xu, Yuanbo Xiangli, Zhenzhi Wang, Dahua Lin, and Bo Dai. Matrixcity: A large-scale city dataset for city-scale neural rendering and beyond. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3205-3215, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[18] Yilin Liu, Liqiang Lin, Yue Hu, Ke Xie, Chi-Wing Fu, Hao Zhang, and Hui Huang. Learning reconstructability for drone aerial path planning. ACM Transactions on Graphics (TOG), 41(6):1-17, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.456, + 0.892, + 0.525 + ], + "angle": 0, + "content": "[19] Viktor Makoviychuk, Lukasz Wawrzyniak, Yunrong Guo, Michelle Lu, Kier Storey, Miles Macklin, David Hoeller, Nikita Rudin, Arthur Allshire, Ankur Handa, and Gavriel State. Isaac Gym: High performancegpu-based physics simulation for robot learning, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.526, + 0.892, + 0.608 + ], + "angle": 0, + "content": "[20] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 405-421. Springer, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.892, + 0.664 + ], + "angle": 0, + "content": "[21] Joseph Ortiz, Alexander Clegg, Jing Dong, Edgar Sucar, David Novotny, Michael Zollhoefer, and Mustafa Mukadam. iSDF: Real-time neural signed distance fields for robot perception. Robotics: Science and Systems, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.665, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[22] Xuran Pan, Zihang Lai, Shiji Song, and Gao Huang. ActiveNeRF: Learning where to see with uncertainty estimation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIII, pages 230-246. Springer, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[23] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[24] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. PyTorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16444" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[25] Daryl Peralta, Joel Casimiro, Aldrin Michael Nilles, Justine Aletta Aguilar, Rowel Atienza, and Rhandley Cajote. Houses3K dataset. https://github.com/darylperalta/Houses3K, 2020.2.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.148, + 0.472, + 0.232 + ], + "angle": 0, + "content": "[26] Daryl Peralta, Joel Casimiro, Aldrin Michael Nilles, Justine Aletta Aguilar, Rowel Atienza, and Rhandley Cajote. Next-best view policy for 3d reconstruction. In Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16, pages 558-573. Springer, 2020. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.471, + 0.3 + ], + "angle": 0, + "content": "[27] Antonin Raffin, Ashley Hill, Adam Gleave, Anssi Kanervisto, Maximilian Ernestus, and Noah Dormann. Stable-baselines3: Reliable reinforcement learning implementations. The Journal of Machine Learning Research, 22(1): 12348-12355, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.301, + 0.471, + 0.37 + ], + "angle": 0, + "content": "[28] Yunlong Ran, Jing Zeng, Shibo He, Jiming Chen, Lincheng Li, Yingfeng Chen, Gimhee Lee, and Qi Ye. NeurAR: Neural uncertainty for autonomous 3d reconstruction with implicit neural representations. IEEE Robotics and Automation Letters, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.371, + 0.471, + 0.426 + ], + "angle": 0, + "content": "[29] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv:2007.08501, 2020.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.427, + 0.471, + 0.468 + ], + "angle": 0, + "content": "[30] Antoni Rosinol, John J Leonard, and Luca Carlone. NeRF-SLAM: Real-time dense monocular slam with neural radiance fields. arXiv preprint arXiv:2210.13641, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.471, + 0.524 + ], + "angle": 0, + "content": "[31] Nikita Rudin, David Hoeller, Philipp Reist, and Marco Hutter. Learning to walk in minutes using massively parallel deep reinforcement learning. In 5th Annual Conference on Robot Learning, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.471, + 0.566 + ], + "angle": 0, + "content": "[32] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.471, + 0.607 + ], + "angle": 0, + "content": "[33] James A Sethian. A fast marching level set method for monotonically advancing fronts. proceedings of the National Academy of Sciences, 93(4):1591-1595, 1996. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.471, + 0.663 + ], + "angle": 0, + "content": "[34] Edward J Smith, Michal Drozdal, Derek Nowrouzezahrai, David Meger, and Adriana Romero-Soriano. Uncertainty-driven active vision for implicit scene reconstruction. arXiv preprint arXiv:2210.00978, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.471, + 0.733 + ], + "angle": 0, + "content": "[35] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.471, + 0.789 + ], + "angle": 0, + "content": "[36] Edgar Sucar, Shikun Liu, Joseph Ortiz, and Andrew J Davison. iMAP: Implicit mapping and positioning in real-time. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6229-6238, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.471, + 0.858 + ], + "angle": 0, + "content": "[37] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[38] Jiaming Sun, Yiming Xie, Linghao Chen, Xiaowei Zhou, and Hujun Bao. NeuralRecon: Real-time coherent 3d reconstruction from monocular video. In Proceedings of the IEEE/CVF" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Conference on Computer Vision and Pattern Recognition, pages 15598-15607, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[39] Jiaming Sun, Xi Chen, Qianqian Wang, Zhengqi Li, Hadar Averbuch-Elor, Xiaowei Zhou, and Noah Snavely. Neural 3D reconstruction in the wild. In SIGGRAPH Conference Proceedings, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.894, + 0.205 + ], + "angle": 0, + "content": "[40] Sebastian Thrun. Probabilistic robotics. Communications of the ACM, 45(3):52-57, 2002. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[41] Josh Tobin, Rachel Fong, Alex Ray, Jonas Schneider, Wojciech Zaremba, and Pieter Abbeel. Domain randomization for transferring deep neural networks from simulation to the real world. In 2017 IEEE/RSJ international conference on intelligent robots and systems (IROS), pages 23-30. IEEE, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[42] Tong Wu, Jiarui Zhang, Xiao Fu, Yuxin Wang, Jiawei Ren, Liang Pan, Wayne Wu, Lei Yang, Jiaqi Wang, Chen Qian, et al. Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 803-814, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.375, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[43] Yuanbo Xiangli, Linning Xu, Xingang Pan, Nanxuan Zhao, Anyi Rao, Christian Theobalt, Bo Dai, and Dahua Lin. BungeeNeRF: Progressive neural radiance field for extreme multi-scale scene rendering. In The European Conference on Computer Vision (ECCV), 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.446, + 0.894, + 0.515 + ], + "angle": 0, + "content": "[44] Kai Ye, Siyan Dong, Qingnan Fan, He Wang, Li Yi, Fei Xia, Jue Wang, and Baoquan Chen. Multi-robot active mapping via neural bipartite graph matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14839–14848, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.517, + 0.892, + 0.557 + ], + "angle": 0, + "content": "[45] Cem Yuksel. Sample elimination for generating poisson disk sample sets. Computer Graphics Forum, 34(2):25-32, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.559, + 0.892, + 0.613 + ], + "angle": 0, + "content": "[46] Huangying Zhan, Jiyang Zheng, Yi Xu, Ian Reid, and Hamid Rezatofighi. ActiveRMAP: Radiance field for active mapping and planning. arXiv preprint arXiv:2211.12656, 2022. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.615, + 0.892, + 0.671 + ], + "angle": 0, + "content": "[47] Han Zhang, Yuong Yao, Ke Xie, Chi-Wing Fu, Hao Zhang, and Hui Huang. Continuous aerial path planning for 3d urban scene reconstruction. ACM Trans. Graph., 40(6):225-1, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.672, + 0.892, + 0.741 + ], + "angle": 0, + "content": "[48] Xiaoshuai Zhang, Sai Bi, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. NeRFusion: Fusing radiance fields for large-scale scene reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5449-5458, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.742, + 0.892, + 0.783 + ], + "angle": 0, + "content": "[49] Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3D: A modern library for 3D data processing. arXiv:1801.09847, 2018.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.786, + 0.892, + 0.868 + ], + "angle": 0, + "content": "[50] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. Nice-SLAM: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12786-12796, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.868 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16445" + } + ] +] \ No newline at end of file diff --git a/2024/GenNBV_ Generalizable Next-Best-View Policy for Active 3D Reconstruction/171b2e0e-83b0-4811-a381-9c85825e51ab_origin.pdf b/2024/GenNBV_ Generalizable Next-Best-View Policy for Active 3D Reconstruction/171b2e0e-83b0-4811-a381-9c85825e51ab_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b15ce06329a125ae11fcd11a89014681737c467c --- /dev/null +++ b/2024/GenNBV_ Generalizable Next-Best-View Policy for Active 3D Reconstruction/171b2e0e-83b0-4811-a381-9c85825e51ab_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4651f4752beecce770bc6fcc2a642c4cc5cbc3fe8bf798103e3e856df89d92cd +size 4641292 diff --git a/2024/GenNBV_ Generalizable Next-Best-View Policy for Active 3D Reconstruction/full.md b/2024/GenNBV_ Generalizable Next-Best-View Policy for Active 3D Reconstruction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7a0b307014bc8301479631eefac39114c1aef896 --- /dev/null +++ b/2024/GenNBV_ Generalizable Next-Best-View Policy for Active 3D Reconstruction/full.md @@ -0,0 +1,287 @@ +# GenNBV: Generalizable Next-Best-View Policy for Active 3D Reconstruction + +Xiao Chen $^{1,2}$ Quanyi Li $^{1}$ Tai Wang $^{1}$ Tianfan Xue $^{2,\dagger}$ Jiangmiao Pang $^{1,\dagger}$ + +$^{1}$ Shanghai AI Laboratory $^{2}$ The Chinese University of Hong Kong + +{cx123,tfxue}@ie.cuhk.edu.hk quanyili0057@gmail.com {wangtai,pangjiangmiao}@pjlab.org.cn + +![](images/97e6c475f76b6f9fa9f7dbf28d5c5131ec961abd99ab0e3a9e7143b9500460d2.jpg) +Figure 1. To determine the best view for 3D reconstruction, previous methods only chose from hand-crafted action space or based on object-centric capturing, lacking the ability to generalize to unforeseen scenes (Left). With our end-to-end trained, generalized free-space policy, it can generalize to unseen objects, enabling the captured drone to image from any viewpoint (Right). + +![](images/ded81a23d7169a288f767c4332cb809663f91477947822176507cd55a0374d8c.jpg) + +# Abstract + +While recent advances in neural radiance field enable realistic digitization for large-scale scenes, the image-capturing process is still time-consuming and labor-intensive. Previous works attempt to automate this process using the Next-Best-View (NBV) policy for active 3D reconstruction. However, the existing NBV policies heavily rely on hand-crafted criteria, limited action space, or perscense optimized representations. These constraints limit their cross-dataset generalizability. To overcome them, we propose GenNBV, an end-to-end generalizable NBV policy. Our policy adopts a reinforcement learning (RL)-based framework and extends typical limited action space to 5D free space. It empowers our agent drone to scan from any viewpoint, and even interact with unseen geometries during training. To boost the cross-dataset generalizability, we also propose a novel multi-source state embedding, including geometric, semantic, and action representations. We establish a benchmark using the Isaac Gym simulator with the Houses3K and OmniObject3D datasets to evaluate this NBV policy. Experiments demonstrate that our policy achieves a $98.26\%$ and $97.12\%$ coverage ratio on unseen building-scale objects from these datasets, respectively, outperforming prior solutions. + +# 1. Introduction + +Recent advances in 3D reconstruction [23, 38, 39] and neural rendering [20, 37, 48] have significantly enhanced the quality of 3D digitization of large scenes, such as buildings and city landmarks [11, 17, 18, 43, 47]. However, image-capturing process still remains time-consuming and labor-intensive. To scan a building-scale scene using a drone, it may take several days' effort of a professional team to ensure a full coverage. Moreover, even professional pilots may miss some areas at the first scan, leading to multiple rounds for rescanning. + +To alleviate the effort in manual scanning, active 3D reconstruction algorithms, especially the Next-Best-View (NBV) policy, have emerged as a promising approach to automate view planning. However, typical rule-based NBV policies [9, 14, 22, 26, 28, 46] heavily rely on empirically designed view-selection criteria and hand-crafted action space, such as a hemisphere as shown in Fig. 1, which limit their generalizability to unseen scenes. To enhance generalization capacity, pioneering works on learning-based NBV policy have been proposed [3, 4, 26, 46]. Nevertheless, they are constrained into impractical setups in real-world scenarios, such as object-centric capturing and limited action space for ground robots, resulting in incomplete 3D reconstruction because of significant self-occlusion. + +Therefore, in this work, we study this problem: "How to design an NBV policy that supports exploration in any + +space and also can generalize to unseen building-scale geometries?" Building upon the design of previous reinforcement learning-based (RL-based) NBV policies [3, 4, 26], our setup faces the following new challenges that were not fully studied before: 1) Designing an easy-to-generalize action space, rather than object-specific spaces used in previous work, such as hand-crafted candidates or a constrained hemisphere; 2) With a large action space, an informative representation is necessary to guide the policy to efficiently find an optimal capturing trajectory; 3) Proposing generic reward functions and developing the distributed training procedure for better generalizability. + +To overcome these two challenges, we propose GenNBV, an end-to-end generalizable NBV policy. It has a novel design of action spaces and state embeddings, which allows free space exploration and ensures a large coverage when applying to unseen objects. We firstly extend former limited action space, like a hemisphere, to 5D free space. As shown in Fig. 1, this new space is composed of a position subspace of approximately $20m \times 20m \times 10m$ and an omnidirectional heading subspace. This free space enables our agent drone to scan from any viewpoint. Moreover, with much larger action space, our RL-based policy can be easily generalized to building-scale geometries without concerning the potential scale shift from training to evaluation. On the contrary, many hand-crafted designs [14, 22] limit available viewpoints to a small predefined set, and thus cannot generalize if there is drastic change in scales. + +Second, in order to generalize a well-trained NBV policy to unseen scenes during evaluation, we propose a novel state representation directly from raw sensory inputs to guide the policy. In contrast to the widely-used neural radiance field (NeRF) [14, 22, 28, 34, 46], our representation does not need time-consuming per-scene optimization. Specifically, our NBV policy incorporates a multi-source state embedding, which comprises three key components: a novel geometric embedding, a semantic embedding, and an action embedding. The geometric embedding is derived from multi-view depth maps and represents an encoded probabilistic 3D grid, while the semantic embedding is extracted from RGB images, and the action embedding is an encoded viewpoint sequence. In contrast to 2D representations used in previous RL-based NBV policies [3, 4], our multi-source state embedding, extracted from a probabilistic 3D grid, is a more comprehensive representation that supports a robust prediction of next viewpoints. + +To validate the effectiveness of GenNBV, we construct a benchmark for training and evaluation with Houses3K [25] dataset from the NVIDIA Isaac Gym [19] simulator. We further test its generalization ability on novel buildings and object categories from OmniObject3D [42] and Objaverse [6] datasets. Our evaluation metrics quantify the completeness, accuracy, and efficiency for active 3D re + +construction. To consolidate completeness and efficiency into a single number, we also propose a metric, Area Under the Curve (AUC) of the coverage. Compared with heuristic [12], information gain-based [12, 14, 46], and RL-based baselines [26], our method achieves the best result under different metrics. We also provide visualizations to demonstrate the generalizability of GenNBV. A demo video is attached in the appendix to further illustrate our method. + +# 2. Related Work + +Traditional 3D Reconstruction. Photometry and geometry are two crucial aspects of reconstruction evaluation. Neural implicit representations [20, 23] have shown progress in photometric rendering but faces challenges like time-consuming optimization and poor generalization, hindering its application in real-time reconstruction scenarios like 3D SLAM [21, 30, 36, 50]. Geometry, in contrast, typically corresponds to 3D representations such as point clouds and 3D mesh and is directly related to issues like collision avoidance. Considering these properties, researchers tend to focus on geometric reconstruction in practical applications and we also follow this stream. + +Active 3D Reconstruction. Active 3D reconstruction is a promising field that has not been thoroughly benchmarked yet. The pipeline of active 3D reconstruction frameworks alternates between inferring optimal viewpoints, capturing new data, and updating the rebuilt 3D model. With the optimal sequence of key viewpoints, a continuous planning path can be generated using the classic planners, such as Fast Marching Method (FMM) [33]. + +Existing works can be pivotally differentiated based on the paradigm of NBV policies: rule-based or learning-based. Rule-based approaches like uncertainty-driven works [12, 14, 28, 46] typically yield the next best view from hand-crafted rules from scene representations, which tends to overfit specific scenes. Most learning-based policies [3, 4, 26] use a deep reinforcement learning algorithm like PPO [32] to sequentially predict the optimal viewpoints based on observation. They must obtain feedback from task-related rewards such as coverage ratio [9, 26] and optimize during massive iteration with task environment. + +The action space is designed based on the paradigm of NBV policies. Most rule-based NBV policies select views from a limited action space, such as a set composed of only one hundred candidate viewpoints [22], a tiny pre-defined space like a hemisphere [9, 14, 28, 46], making it possible to overlook some important viewpoints due to unavailability. Even though learning-based categories further explore the larger action space like a 2D plain [3, 4] or a constrained 3D space [26], their limited action space still prevents them from capturing sufficient details for 3D reconstruction. + +Scene representation built from history observations, which directly provide reconstruction progress to NBV + +![](images/78664c646a8f660fbe0d055e673107469844db204a1d0eab6e3db8e59a6e10a6.jpg) +Figure 2. Overview of our proposed framework GenNBV. Our end-to-end policy takes the historical multi-source observations as input, transforms them into a more informative scene representation, and produces the next viewpoint position. A reward signal will be returned at training time to optimize the end-to-end policy for maximizing the expected cumulative reward in one episode. Specifically, the signal is the increased coverage ratio after collecting a new viewpoint. + +policies, is also a crucial aspect of the active 3D reconstruction framework. Previous works have explored visual representations for NBV policies, such as TSDF [11], neural radiance field [1], and 2D BEV maps [3, 10, 44]. However, implicit representations are hard to jointly with learning-based NBV policies, and 2D BEV map lacks sufficient information for large-scale outdoor scenes that contain numerous geometric details in 3D space. + +In our benchmark, we select Scan-RL [26] as our main RL-based baseline. Compared to it, we introduce a much larger action space and informative visual representations for better generalizability. Additionally, we don't benchmark the mentioned baselines [3, 4], mainly due to their reliance on pretraining from human demonstration using imitation learning, while our focus lies on learning through iterations and end-to-end training. Different from our 3D free-space movement, they only explore the traversable 2D areas of indoor scenes using ground robots, thus limiting their motion space in 2D. + +Generalizable Reinforcement Learning. Increasing the diversity of training data leads to better generalizability [5]. In the robotics field, domain randomization [41] allows legged robots to walk on various terrains absent from the training environment. For end-to-end driving policy, training in large-scale synthetic and realistic scenarios improves the safety of autonomous cars [15, 16] in unseen test scenarios. The advanced work [7] presents an RL-based formulation for the view planning problem. Based on similar frameworks of active 3D reconstruction, Scan-RL [26] and SCONE [9] are two pioneering works showing the generalizability of their frameworks, while both of them suffer from the constrain of viewpoint sampling in limited space and lack of in-depth analysis. In this work, we tackle the aforementioned issues by predicting NBV in free 3D space with a learning-based planner, which is trained with a + +dataset containing buildings in various shapes and poses for acquiring generalizability for unseen buildings. + +# 3. Methodology + +In this section, we deliver our active 3D reconstruction framework GenNBV, especially the pivotal Next-Best-View policy. An overview of our framework is illustrated in Fig. 2. Firstly, we formulate the NBV problem as a Markov Decision Process (MDP), with a novel design of observations (blue boxes in Fig. 2) and action space in Sec. 3.1. Next, we elaborate our end-to-end NBV policy $\pi$ (orange box in Fig. 2) in Sec. 3.2. Inspired by Curl [13], we point out that generalizability greatly depends on how to extract embeddings (green box in Fig. 2), which reflect the reconstruction progress, from raw sensory observations like RGB images and camera poses. In Sec. 3.3, we introduce the reward function (right-hand side of Fig. 2) reflecting the optimization objective and the details of policy optimization. + +# 3.1. Formulation of the Next-Best-View Problem + +We formulate the NBV problem as learning an optimal policy $\pi$ that controls the capturing process, such that enough information is captured for large-scale scene reconstruction, with limited decision-making budgets. As capturing, transferring, and processing a large set of captured images introduce significant computational costs, we also want to design a policy that also minimizes the number of captured images. Therefore, our policy only captures sparse keyframes that sufficiently record all details of objects. + +As shown in Fig. 2, at each time step $t$ , the agent receives a visual observation $o_t$ , takes an action $a_t$ , infers the action at the next time step $t + 1$ to move to a new location, and then receives a new visual observation, repeating this interaction with the environment until the episode ends. Our simulated agent is embodied as CrazyFlie [8], a type of unmanned + +aerial vehicle equipped with various sensors, including an RGB-D camera and an IMU, to execute data collection for reconstruction. We discussed the details of the observation space and action space below. + +Observation Space. As shown in the left column of Fig. 2, at each time step, the agent receives an RGB image $I_{t}$ , a depth map $D_{t}$ , and a state vector including the heading (yaw and pitch) and position $(\mathrm{x},\mathrm{y},\mathrm{z})$ of the onboard camera. The observation $o_{t}$ (yellow box in Fig. 2) consists of all previously captured images in one episode, historical actions, and the current captures and actions. With this information as input, the policy network can estimate the progress of data collection and determine where to scan next. + +Action Space. Unlike most previous NBV algorithms, we use larger action space in order to cover all details of objects. Specifically, we use the camera location and camera angle as our action space, which is a 5-dimension vector consisting of 3D position coordinates and 2D rotation angles (yaw-axis and pitch-axis). We restrict the roll-axis rotation as it is not supported in all drone platforms. + +# 3.2. Generalizable State Embedding + +To ensure that the policy learned on one set of 3D objects can generalize to objects with different appearances and structures, a smart state embedding of raw sensor observations is needed to capture invariant features across different objects. Previous methods [3, 26] only extract representations in 2D space, which are very sensible to appearance changes, and our experiments have shown that policies only trained on 2D features may not generalize to different objects. Instead, we propose a multi-source representation that has better generalizability. + +Specifically, we first build two mid-level representations that can better model the relationship between the objects and our agent: a 3D geometric representation $F^G$ from depth maps and a semantic representation $F^S$ from RGB images. Then we encode these representations and concatenate them with a pose embedding into state embedding $s_t$ , guiding the NBV policy for subsequent decision-making. The overall encoder for state embedding is shown in Fig. 2, and details of each representation are discussed below. + +Geometric Representation One simple way to record the geometry of a 3D object is using a binary 3D occupancy map [3], where the value of each cell in the 3D cube indicates whether a cell contains a 3D object or not. However, since the 3D representation is gradually updated with newly captured data, this simple binary occupancy map cannot differentiate a real unoccupied cell from an unscanned cell. An unscanned cell is a strong indicator that the agent should capture more data in this region, while no further scan is needed for a real unoccupied cell. + +Previous works [3, 10, 44] simplify the 3D scene representation into a 2D Bird's Eye View map for actively recon + +structuring indoor scenes. However, the 2D BEV map lacks sufficient information for our large-scale outdoor scenes that contain many geometric details in 3D space. In addition, their wheeled robotic platforms are constrained to freely scan and represent these outdoor scenes. Therefore, to model the scanning process in 3D free space, we employ the probabilistic 3D grid [40] as our geometric 3D representation, which records the probability of each 3D voxel being captured or not. Specifically, we first obtain a 3D point cloud in the world coordinate by back-projecting all 2D pixels to 3D points, using the depth map $D_{t}$ , camera intrinsic parameters, and camera pose $a_{t}$ . By voxelizing the obtained point cloud, we then build a 3D occupancy grid that explicitly indicates the binary state (occupied or free) in this 3D space. Subsequently, we represent this 3D grid as a probabilistic occupancy grid $F_{t}^{G}$ and extend the state space of voxels with three states (occupied, free, unknown). + +During each scanning (one episode in RL), at each step $t + 1$ , we update the probabilistic occupancy grid $F_{t + 1}^{G}$ based on the preceding grid $F_{t}^{G}$ and current observation. Specifically, the grid is updated through Bresenham's line algorithm [2], which casts the ray path in 3D space between the camera viewpoint and the endpoints among the point cloud back-projected from depth $D_{t + 1}$ . Following the classical occupancy grid mapping algorithm [40], we have the log-odds formulation of occupancy probability: + +$$ +\log \operatorname {O d d} \left(v _ {i} \mid z _ {j}\right) = \log \operatorname {O d d} \left(v _ {i}\right) + C, \tag {1} +$$ + +where $v_{i}$ denotes the occupancy probability of $i^{th}$ voxel in the grid $F_{t}^{G}$ , $z_{j}$ is the measurement event that $j^{th}$ camera ray passes through this voxel and $C$ is an empirical constant. The derivation can be found in Appendix. Thus, we update the log-odds occupancy probability of each voxel in the grid $F_{t}^{G}$ by adding a constant one time when a single camera ray passes through this voxel. Note that the probabilistic occupancy grid $F^{G}$ is continually updated within one episode. Finally, the occupancy status of voxels is classified into three categories: unknown, occupied, and free, using preset probability thresholds. + +Semantic Representation. Geometric representation enables agents to comprehend spatial occupancy, yet it's insufficient for perceiving the environment. For example, when observing a hole in an object, the agent may struggle to differentiate between incomplete scanning and the actual presence of a hole in the object. In such cases, the semantic information contained in the captured RGB images can help the agent distinguish between these two scenarios. + +To provide semantic information, we employ a preprocessing module that takes as input the current frame of RGB image $I_{t}$ and preceding $k$ frames and converts these frames $[I_t,I_{t - 1},\dots,I_{t - k}]$ to grayscale, and concatenate them as output, following [26]. Then the preprocessed frames are fed into a two-layer convolutional network for extracting + +the semantic representation $F_{t}^{S}$ + +State Embedding. To further combine the semantic and geometric embeddings, we first encode them to $s_t^S = f^S (F_t^S)$ and $s_t^G = f^G (F_t^G)$ where $f^{*}$ are learnable networks Linear(Flatten(x)), as shown in Fig. 2. Subsequently, we combine them with the historical action embedding $s_t^A = \mathrm{Linear}(a_{1:t})$ to generate the final state embedding $s_t$ , as the input to the policy network. This process can be formulated as: $s_t = \mathrm{Linear}(s_t^G;s_t^S;s_t^A)$ . + +Policy Network. Taking the state embedding $s_t$ as input, the policy network is a 3-layer multi-layer perceptron network (MLP) whose output is used to parameterize a normal distribution over action space. In this way, the action can be drawn from the stochastic policy $a \sim \pi(\cdot | o_t)$ . + +# 3.3. Reward Function and Optimization + +We train the end-to-end policy with reinforcement learning (RL) and hence design reward functions to precisely reflect the task objective for 3D reconstruction. The policy is optimized with proximal policy optimization [32] (PPO) for parallelizing sampling. + +Reward Functions. With the occupancy probability $F_{t}^{G}$ at time step $t$ , we can threshold each voxel with an empirical bound to determine if it is occupied. This discrimination process outputs a binary occupancy grid with $\tilde{N}_{t}$ voxels being occupied, which is used to calculate the coverage ratio: + +$$ +\mathrm{CR}_{t} = \frac{\tilde{N}_{t}}{N^{*}}\cdot 100\% , \tag{2} +$$ + +where $N^{*}$ is the number of ground-truth occupied voxels representing the surface of objects. To encourage our NBV policy to cover as many unseen areas of objects as possible, we use the difference of coverage ratio (CR) between two consecutive steps as the main reward function $r^{CR}$ : + +$$ +r _ {t + 1} ^ {\mathrm {C R}} = \mathrm {C R} _ {t + 1} - \mathrm {C R} _ {t}. \tag {3} +$$ + +In free-space exploration, we also need to avoid collision. Previous limited-space agents [9, 14, 46] do not consider collision avoidance since their search space, like hemisphere, is by-design safe. Thus, we add a negative reward for collision and terminate the episode if a collision happens. We also implement a negative reward when the number of captured keyframes is over an empirical threshold to improve the path efficiency. + +Policy Optimization. Once the reward functions has been specified, the policy can be learned through any off-the-shelf RL algorithm. In this work, we use PPO as thousands of workers can be parallelized to improve the sample efficiency. Specifically, given our parameterized policy $\pi_{\theta}$ , the objective of PPO is to maximize the following function: + +$$ +L (\theta) = \mathbb {E} _ {t} \left[ \frac {\pi_ {\theta} (a _ {t} | s _ {t})}{\pi_ {\theta_ {\mathrm {o l d}}} (a _ {t} | s _ {t})} A ^ {\pi_ {\theta_ {\mathrm {o l d}}}} (s _ {t}, a _ {t}) \right], \tag {4} +$$ + +where $A^{\pi_{\theta_{\mathrm{old}}}}(s_t, a_t)$ is the advantage function that measures the value of taking action $a_t$ at state $s_t$ under the current policy $\pi_{\theta_{\mathrm{old}}}$ . To prevent significant deviation of the new policy from the old policy, PPO incorporates a clipped surrogate objective function: + +$$ +\begin{array}{l} L ^ {\mathrm {C L I P}} (\theta) = \mathbb {E} _ {t} [ \min (\eta_ {t} (\theta) A ^ {\pi_ {\theta_ {\text {o l d}}}} (s _ {t}, a _ {t}), \tag {5} \\ \left. \operatorname {c l i p} \left(r _ {t} (\theta), 1 - \epsilon , 1 + \epsilon\right) A ^ {\pi_ {\theta_ {\text {o l d}}}} \left(s _ {t}, a _ {t}\right) \right], \\ \end{array} +$$ + +where $\eta_t(\theta) = \frac{\pi_\theta(a_t|s_t)}{\pi_{\theta_{\mathrm{old}}}(a_t|s_t)}$ and $\epsilon$ is a hyper-parameter that controls the size of the trust region. + +# 4. Experiments + +In this section, we conduct experiments on Houses3K [25], OmniObject3D [42], and Objaverse [6]. For our policy and other learning-based policies, we train them on Houses3K training set. Then we evaluate all policies on the Houses3K test dataset and the OmniObject3D dataset for quantifying the in-distribution and out-of-distribution generalizability. Finally, we show the visualization results on three datasets, demonstrating the effectiveness of the proposed method. + +# 4.1. Experimental Setup + +Simulation Environment. We conduct all experiments in NVIDIA Isaac Gym [19], a physics simulation platform designed for reinforcement learning and robotics research. With GPU-accelerated tensor API and sensor interaction API, we can easily implement customized CUDA kernels to calculate our geometric representation efficiently. Moreover, the sensor simulation is efficient. It can run up to 1000 FPS, significantly reducing the training time. We create the agent drone CrazyFlie [8] and equip it with sensors including RGB-D cameras and IMU. Considering memory limitations, we downsample the image resolution to $400 \times 400$ . The vertical field of view of this onboard camera is $90^{\circ}$ . We also set up a point light source with a fixed position and constant intensity in Isaac Gym. + +Dataset. We conduct our experiments on Houses3K [25], OmniObject3D [42], Objverse 1.0 [6] and Replica [35]. Our model is trained on large-scale 3D objects from Houses3K. To validate the effectiveness and generalizability of our model, we evaluate it on unseen 3D objects from Houses3K and further on batches of unseen cross-domain 3D objects from OmniObject3D, Objverse, and Replica, which include both house and non-house categories. + +Houses3K contains 3,000 textured 3D building models. They are divided into 12 batches, with each batch featuring 50 distinct geometries and 5 varying textures. We further eliminate poorly designed geometries based on the following two criteria: 1) objects with complex internal structures (as we are mostly focusing on the surface reconstruction), and 2) objects with redundant bases at the bottom. Following these criteria, the training set consists of 256 objects + +Table 1. Evaluation results of Next-Best-View policies for active 3D reconstruction on Houses3K and the house category from OmniObject3D (cross-dataset generalization). The number of views is set to 30 and 20 for Houses3K and OmniObject3D, respectively. +\*\*: the policy is trained with the Houses3K training set and evaluated with holdout Houses3K test set and OmniObject3D house category. $\ddagger$ : the policy heavily relies on optimized per-scene representation (NeRF), and thus is directly trained and evaluated on testing objects. + +
Houses3K Test SetOmniObject3D
NBV PolicyMean AUC ↑Final Coverage Ratio ↑Accuracy (cm) ↓Mean AUC ↑Final Coverage Ratio ↑Accuracy (cm) ↓
HeuristicRandom48.53%58.24%1.3860.13%73.73%0.62
Random Hemisphere71.19%79.72%0.6374.03%84.74%0.48
Uniform Hemisphere82.91%89.71%0.4483.09%92.90%0.41
InfoGain†Uncertainty-Guided [14]83.13%89.31%0.4469.08%92.85%0.42
†ActiveRMap [46]84.86%90.77%0.4469.47%93.15%0.42
RL-based*Scan-RL [26]84.49%91.61%0.4082.17%92.53%0.34
*Scan-RL w/ Our Repre.86.48%92.20%0.3786.14%93.73%0.35
*Ours w/ Scan-RL Repre.87.39%95.63%0.3886.81%93.64%0.34
*GenNBV (Ours)91.19%98.26%0.3788.63%97.12%0.33
+ +Table 2. The cross-dataset generalization for non-house categories. We train the baseline Scan-RL and our GenNBV on Houses3K and generalize them to non-house categories from OmniObject3D and an indoor scene from Replica. + +
CategoryDatasetNBV PolicyAUC ↑FCR ↑Acc. ↓
AnimalsOmniObject3DScan-RL76.43%84.98%0.17
GenNBV84.28%94.07%0.16
TrucksOmniObject3DScan-RL69.58%76.37%0.06
GenNBV75.54%83.47%0.03
DinosaursOmniObject3DScan-RL78.47%86.89%0.88
GenNBV84.24%94.03%0.81
Room 0ReplicaScan-RL69.21%80.75%3.12
GenNBV88.53%99.16%3.59
+ +from 6 selected batches, and the test set consists of 50 objects from another batch. All these selected training and test objects have distinct geometries. + +OmniObject3D offers a collection of 6K high-fidelity objects scanned from real-world sources across 190 typical categories. The house category from OmniObject3D for evaluation has 43 diverse objects. In contrast, Objaverse 1.0 comprises a vast library of 818K 3D synthetic objects spanning 21K categories. The diverse and high-quality nature of these datasets makes them ideal for evaluating our models and visualizing the results. + +Evaluation Metrics. The objective of NBV policies is to capture the most useful information for reconstruction, with the least number of views. We report the (1) Final Coverage Ratio $(\%)$ , (2) Mean AUC $(\%)$ of coverage ratio and + +(3) Reconstruction Accuracy (cm) along with the consistent number of views in all tables. Specifically, most prior works [9, 26] separately use the coverage ratio (\%) and the number of views to evaluate the reconstruction completeness and efficiency for NBV policies. However, the coverage ratio is highly correlated with the number of views. Thus we propose to unify the number of views to a fixed value for all NBV policies during evaluation and use the area under the curve (AUC) of coverage ratio as the main metric for comparison. Also, we calculate the Chamfer Distance between scanned and ground-truth point clouds to represent the reconstruction accuracy (cm). + +Implementation Details. We conduct all experiments in Isaac Gym simulation engine with one NVIDIA Tesla V100 GPU. Our NBV policy is optimized through over 32 million iterations and uses approximately 24 hours of training time on an NVIDIA V100 GPU. All networks are randomly initialized and trained end-to-end. Our implementation refers to the codebase of Legged Gym [31] and the PPO implementation in Stable Baseline3 [27], which is implemented by PyTorch [24]. The ground-truth point clouds on objects' surfaces are sampled by the Poisson Disk sampling method [45] using Open3D API [49]. The Chamfer Distance between the scanned point cloud and the ground-truth point cloud is calculated using PyTorch3D API [29]. Please refer to the Appendix for further details and experiments. + +# 4.2. Performance Comparison + +To comprehensively demonstrate the effectiveness and generalizability of GenNBV, we design three levels of evaluation experiments: 1) As shown in Table 1, we show the per + +![](images/b0d176c1467ed0538f742e96a7dacf2236e75e6ab7466bd292693e05d4f08725.jpg) +(a) Houses3K + +![](images/a7283e936288c6a87f1ba831d3104b034d1b5a6a91f72dd06461e157e0ac2cf4.jpg) +(b) OmniObject3D +Figure 3. The visualization results of unseen 3D objects reconstructed by Scan-RL [26] and our model to compare the generalizability. (a) Unseen buildings from the test set of Houses3K. (b) Unseen buildings from OmniObject3D. It's quite obvious that some parts of the model reconstructed by Scan-RL are wrong or missing. For example, the second object in the first row has a pillar in the wrong shape. Scan-RL fails to reconstruct the roof edge for the fourth object from OmniObject3D, as shown in the third row. + +formance of our NBV policy on the Houses3K test set; 2) We generalize the policy trained on Houses3K to the house category from OmniObject3D that has completely different geometric structures and textures compared to the training set. The quantitative result is at Table 1 and the visualization result is as shown in Fig. 3; 3) We also generalize GenNBV to non-house categories from OmniObject3D and scenes with enormous details from Objaverse and Replica [35] to demonstrate its generalizability potential. + +We implement the following policies as our baselines. 1) Random: This policy randomly generates 5-dim vector $(x,y,z,\text{pitch},\text{yaw})$ among the action space as the next viewpoint. 2) Random Hemisphere: This policy randomly generates the next positions on a pre-defined hemisphere that sufficiently covers all objects of the test set. The headings are constrained to point to the center of the hemisphere. 3) Uniform Hemisphere: All positions are evenly distributed on the previously mentioned hemispheres. 4) Uncertainty-Guided Policy [14]: This NBV policy iteratively selects the next view from a pre-defined viewpoint set according to the uncertainty based on a continually optimized neural radiance field. 5) ActiveRMap [46]: This policy also adopts an iterative selection framework, with multiple objectives including information gain, to select the next viewpoints from the candidate set. 6) Scan-RL [26]: This RL-based policy predicts the next viewpoint from a hemisphere space, relying on the historical RGB images. 7) Ours with Scan-RL's Representation: This free-space NBV policy replaces our representation with Scan-RL's representation only extracted from RGB images. + +As shown in Table 1, our GenNBV shows the best in-distribution and out-of-distribution generalizability in both coverage sufficiency and view efficiency, when evaluated on test sets consisting of unseen objects from Houses3K + +Table 3. Ablation studies of representation categories in our framework on unseen Houses3K test set. + +
Representation CategoryEvaluation Metrics
Probabilistic 3D Grid5-DoF PoseSemantic 2D MapMean AUCFinal Coverage Ratio
Uni-source✓81.06%84.56%
✓69.53%76.61%
✓81.24%87.90%
Multi-source✓✓88.66%96.67%
✓✓89.77%95.31%
✓✓88.30%96.29%
✓✓✓91.19%98.26%
+ +and OmniObject3D. In particular, the comparison in Table 1 demonstrates the strength of the proposed representations and action space. Note that the Scan-RL's representation (\* in the table) uses 6 frames, while our semantic representation only needs 2 frames. Moreover, learning-based NBV policies such as GenNBV and Scan-RL, with much larger action space, outperform all rule-based baselines, even if ActiveRMap and Uncertainty-Guided baselines are trained and evaluated on the same 3D objects. + +Non-house Generalization In Table 2, we introduce an experiment to evaluate the generalizability of our GenNBV trained on Houses3K to non-house categories. The targets include 74 animals, 43 trucks and 33 dinosaurs from OmniObject3D, and a challenging indoor scene (Room 0) from Replica. The number of views is set to 20. + +Our GenNBV almost outperforms Scan-RL in all metrics. Particularly, on Replica Room 0, which consists of numerous multi-sized objects with complex occlusion, our + +![](images/b0c3e919f63a980a9b583f9aa7e2dde3596478edfd85ffd2421651dcb5de0174.jpg) +Uncertainty-guided + +![](images/de115d3f6207bdc922bc1c62d4fd16d421e2b26f8d9a89687ba5ce3dbb6c4e90.jpg) +Scan-RL +Figure 4. The visualization results of an unseen 3D outdoor scene with enormous details from Objaverse, reconstructed by Uncertainty-guided, Scan-RL and our model. Compared to the uncertainty-guided method and Scan-RL, the scene reconstructed by our method is more watertight and has fewer holes on the ground and building surface, especially in the region highlighted by the red box. + +![](images/80d0874bac5a1455c60639ed88a12461439c227ea43b987e895652f906d81521.jpg) +GenNBV (Ours) + +GenNBV is significantly better than Scan-RL. This shows the effectiveness of our free-space NBV policy. + +# 4.3. Ablation Study + +We implement the ablation study about multi-source state embedding and depth-based representation. Please refer to the Appendix for further experiments. + +Multi-source State Embedding. In Sec. 3.2, we introduce how to build our multi-source state embedding. Thanks to the same shape of representations and state embedding, we only need to adjust the input dimension of the linear layer of state embedding according to the modal type when ablating the model. Here we reveal the importance of specific representations with the ablation results shown in Table 3. In uni-source experiments, we design the 5-DoF Pose baseline that learns an empirically optimal trajectory, which helps understand the effectiveness of domain knowledge of the dataset. We also evaluate different combinations of representation categories in Table 3, demonstrating the effectiveness of our multi-source state embedding for policy learning in multi-source experiments. + +Table 4. Ablation studies of depth-based representations on unseen Houses3K test set, where depth map is the only sensory source. + +
Depth-based RepresentationMean AUCFinal Coverage Ratio
Geometric 2D Map71.86%74.88%
Binary 3D Grid77.28%80.85%
Probabilistic 3D Grid81.06%84.56%
+ +Depth-based Representation. Furthermore, we ablate the depth-based representations in Table 4. The experimental results demonstrate that our probabilistic 3D grid is more comprehensive than a binary 3D grid and a geometric 2D map to support the NBV prediction. + +# 4.4. Qualitative Results + +We visualize the reconstruction results generated from the scanning trajectory of a single episode in Fig. 3. It demonstrates that our next-best-view policy can reconstruct objects better in terms of completeness and appearance quality compared to Scan-RL. We further visualize the reconstruction results of an outdoor scene from Objaverse using 30 collected views in Fig. 4. + +# 5. Conclusion + +This study presents an end-to-end approach for active 3D scene reconstruction, reducing the need for manual intervention. Specifically, the learning-based policy explores how to reconstruct diverse objects in the training stage and thus can generalize to reconstruct unseen objects in a fully autonomous manner. Our controller maneuvers in free space and selects the next best view based on a hybrid scene representation which conveys scene coverage status and thus reconstruction progress. We show the effectiveness of our approach by generalizing it on multiple datasets. The quantitative and qualitative generalization results on holdout Houses3K test set and cross-domain OmniObject3D, including house category and non-house categories, show that our method outperforms other baselines in terms of reconstruction completeness, efficiency and accuracy. Furthermore, the experiment conducted on Objaverse and Replica shows that the policy trained in house-only settings can even generalize to complicated outdoor and indoor scenes. + +# Acknowledgements + +This work is supported by Shanghai Artificial Intelligence Laboratory and CUHK Direct Grants (RCFUS) 4055189. We're very grateful to Tao Huang, Zeqi Xiao, Junfeng Long, Mulin Yu, and all reviewers for their insightful advice. + +# References + +[1] Michal Adamkiewicz, Timothy Chen, Adam Caccavale, Rachel Gardner, Preston Culbertson, Jeannette Bohg, and Mac Schwager. Vision-only robot navigation in a neural radiance world. IEEE Robotics and Automation Letters, 7(2): 4606-4613, 2022. 3 +[2] Jack E Bresenham. Algorithm for computer control of a digital plotter. IBM Systems journal, 4(1):25-30, 1965. 4 +[3] Devendra Singh Chaplot, Dhiraj Gandhi, Saurabh Gupta, Abhinav Gupta, and Ruslan Salakhutdinov. Learning to explore using active neural slam. In International Conference on Learning Representations, 2020. 1, 2, 3, 4 +[4] Tao Chen, Saurabh Gupta, and Abhinav Gupta. Learning exploration policies for navigation. In International Conference on Learning Representations, 2019. 1, 2, 3 +[5] Karl Cobbe, Chris Hesse, Jacob Hilton, and John Schulman. Leveraging procedural generation to benchmark reinforcement learning. In International conference on machine learning, pages 2048-2056. PMLR, 2020. 3 +[6] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 2, 5 +[7] Mustafa Devrim Kaba, Mustafa Gokhan Uzunbas, and Ser Nam Lim. A reinforcement learning approach to the view planning problem. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6933-6941, 2017. 3 +[8] Wojciech Giernacki, Mateusz Skwierczyński, Wojciech Witwicki, Paweł Wronski, and Piotr Kozierski. Crazyflie 2.0 quadrotor as a platform for research and education in robotics and control engineering. In 2017 22nd International Conference on Methods and Models in Automation and Robotics (MMAR), pages 37-42. IEEE, 2017. 3, 5 +[9] Antoine Guedon, Pascal Monasse, and Vincent Lepetit. SCONE: Surface coverage optimization in unknown environments by volumetric integration. In Advances in Neural Information Processing Systems, 2022. 1, 2, 3, 5, 6 +[10] Junfu Guo, Changhao Li, Xi Xia, Ruizhen Hu, and Ligang Liu. Asynchronous collaborative autoscanning with mode switching for multi-robot scene reconstruction. ACM Transactions on Graphics (TOG), 41(6):1-13, 2022. 3, 4 +[11] Guillaume Hardouin, Julien Moras, Fabio Morbidi, Julien Marzat, and El Mustapha Mouaddib. Next-Best-View planning for surface reconstruction of large-scale 3d environments with multiple uavs. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1567–1574. IEEE, 2020. 1, 3 +[12] Stefan Isler, Reza Sabzevari, Jeffrey Delmerico, and Davide Scaramuzza. An information gain formulation for active volumetric 3d reconstruction. In 2016 IEEE International Conference on Robotics and Automation (ICRA), pages 3477-3484. IEEE, 2016. 2 +[13] Michael Laskin, Aravind Srinivas, and Pieter Abbeel. Curl: Contrastive unsupervised representations for reinforcement + +learning. In International Conference on Machine Learning, pages 5639-5650. PMLR, 2020. 3 +[14] Soomin Lee, Le Chen, Jiahao Wang, Alexander Liniger, Suryansh Kumar, and Fisher Yu. Uncertainty guided policy for active robotic 3d reconstruction using neural radiance fields. IEEE Robotics and Automation Letters, 7(4):12070-12077, 2022. 1, 2, 5, 6, 7 +[15] Quanyi Li, Zhenghao Peng, Lan Feng, Chenda Duan, Wenjie Mo, Bolei Zhou, et al. ScenarioNet: Open-source platform for large-scale traffic scenario simulation and modeling. In Advances in Neural Information Processing Systems, 2022. 3 +[16] Quanyi Li, Zhenghao Peng, Lan Feng, Qihang Zhang, Zhenghai Xue, and Bolei Zhou. MetaDrive: Composing diverse driving scenarios for generalizable reinforcement learning. IEEE transactions on pattern analysis and machine intelligence, 45(3):3461-3475, 2022. 3 +[17] Yixuan Li, Lihan Jiang, Linning Xu, Yuanbo Xiangli, Zhenzhi Wang, Dahua Lin, and Bo Dai. Matrixcity: A large-scale city dataset for city-scale neural rendering and beyond. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3205-3215, 2023. 1 +[18] Yilin Liu, Liqiang Lin, Yue Hu, Ke Xie, Chi-Wing Fu, Hao Zhang, and Hui Huang. Learning reconstructability for drone aerial path planning. ACM Transactions on Graphics (TOG), 41(6):1-17, 2022. 1 +[19] Viktor Makoviychuk, Lukasz Wawrzyniak, Yunrong Guo, Michelle Lu, Kier Storey, Miles Macklin, David Hoeller, Nikita Rudin, Arthur Allshire, Ankur Handa, and Gavriel State. Isaac Gym: High performancegpu-based physics simulation for robot learning, 2021. 2, 5 +[20] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 405-421. Springer, 2020. 1, 2 +[21] Joseph Ortiz, Alexander Clegg, Jing Dong, Edgar Sucar, David Novotny, Michael Zollhoefer, and Mustafa Mukadam. iSDF: Real-time neural signed distance fields for robot perception. Robotics: Science and Systems, 2022. 2 +[22] Xuran Pan, Zihang Lai, Shiji Song, and Gao Huang. ActiveNeRF: Learning where to see with uncertainty estimation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIII, pages 230-246. Springer, 2022. 1, 2 +[23] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019. 1, 2 +[24] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. PyTorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6 + +[25] Daryl Peralta, Joel Casimiro, Aldrin Michael Nilles, Justine Aletta Aguilar, Rowel Atienza, and Rhandley Cajote. Houses3K dataset. https://github.com/darylperalta/Houses3K, 2020.2.5 +[26] Daryl Peralta, Joel Casimiro, Aldrin Michael Nilles, Justine Aletta Aguilar, Rowel Atienza, and Rhandley Cajote. Next-best view policy for 3d reconstruction. In Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16, pages 558-573. Springer, 2020. 1, 2, 3, 4, 6, 7 +[27] Antonin Raffin, Ashley Hill, Adam Gleave, Anssi Kanervisto, Maximilian Ernestus, and Noah Dormann. Stable-baselines3: Reliable reinforcement learning implementations. The Journal of Machine Learning Research, 22(1): 12348-12355, 2021. 6 +[28] Yunlong Ran, Jing Zeng, Shibo He, Jiming Chen, Lincheng Li, Yingfeng Chen, Gimhee Lee, and Qi Ye. NeurAR: Neural uncertainty for autonomous 3d reconstruction with implicit neural representations. IEEE Robotics and Automation Letters, 2023. 1, 2 +[29] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv:2007.08501, 2020.6 +[30] Antoni Rosinol, John J Leonard, and Luca Carlone. NeRF-SLAM: Real-time dense monocular slam with neural radiance fields. arXiv preprint arXiv:2210.13641, 2022. 2 +[31] Nikita Rudin, David Hoeller, Philipp Reist, and Marco Hutter. Learning to walk in minutes using massively parallel deep reinforcement learning. In 5th Annual Conference on Robot Learning, 2021. 6 +[32] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 2, 5 +[33] James A Sethian. A fast marching level set method for monotonically advancing fronts. proceedings of the National Academy of Sciences, 93(4):1591-1595, 1996. 2 +[34] Edward J Smith, Michal Drozdal, Derek Nowrouzezahrai, David Meger, and Adriana Romero-Soriano. Uncertainty-driven active vision for implicit scene reconstruction. arXiv preprint arXiv:2210.00978, 2022. 2 +[35] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 5, 7 +[36] Edgar Sucar, Shikun Liu, Joseph Ortiz, and Andrew J Davison. iMAP: Implicit mapping and positioning in real-time. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6229-6238, 2021. 2 +[37] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 1 +[38] Jiaming Sun, Yiming Xie, Linghao Chen, Xiaowei Zhou, and Hujun Bao. NeuralRecon: Real-time coherent 3d reconstruction from monocular video. In Proceedings of the IEEE/CVF + +Conference on Computer Vision and Pattern Recognition, pages 15598-15607, 2021. 1 +[39] Jiaming Sun, Xi Chen, Qianqian Wang, Zhengqi Li, Hadar Averbuch-Elor, Xiaowei Zhou, and Noah Snavely. Neural 3D reconstruction in the wild. In SIGGRAPH Conference Proceedings, 2022. 1 +[40] Sebastian Thrun. Probabilistic robotics. Communications of the ACM, 45(3):52-57, 2002. 4 +[41] Josh Tobin, Rachel Fong, Alex Ray, Jonas Schneider, Wojciech Zaremba, and Pieter Abbeel. Domain randomization for transferring deep neural networks from simulation to the real world. In 2017 IEEE/RSJ international conference on intelligent robots and systems (IROS), pages 23-30. IEEE, 2017. 3 +[42] Tong Wu, Jiarui Zhang, Xiao Fu, Yuxin Wang, Jiawei Ren, Liang Pan, Wayne Wu, Lei Yang, Jiaqi Wang, Chen Qian, et al. Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 803-814, 2023. 2, 5 +[43] Yuanbo Xiangli, Linning Xu, Xingang Pan, Nanxuan Zhao, Anyi Rao, Christian Theobalt, Bo Dai, and Dahua Lin. BungeeNeRF: Progressive neural radiance field for extreme multi-scale scene rendering. In The European Conference on Computer Vision (ECCV), 2022. 1 +[44] Kai Ye, Siyan Dong, Qingnan Fan, He Wang, Li Yi, Fei Xia, Jue Wang, and Baoquan Chen. Multi-robot active mapping via neural bipartite graph matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14839–14848, 2022. 3, 4 +[45] Cem Yuksel. Sample elimination for generating poisson disk sample sets. Computer Graphics Forum, 34(2):25-32, 2015. 6 +[46] Huangying Zhan, Jiyang Zheng, Yi Xu, Ian Reid, and Hamid Rezatofighi. ActiveRMAP: Radiance field for active mapping and planning. arXiv preprint arXiv:2211.12656, 2022. 1, 2, 5, 6, 7 +[47] Han Zhang, Yuong Yao, Ke Xie, Chi-Wing Fu, Hao Zhang, and Hui Huang. Continuous aerial path planning for 3d urban scene reconstruction. ACM Trans. Graph., 40(6):225-1, 2021. 1 +[48] Xiaoshuai Zhang, Sai Bi, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. NeRFusion: Fusing radiance fields for large-scale scene reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5449-5458, 2022. 1 +[49] Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3D: A modern library for 3D data processing. arXiv:1801.09847, 2018.6 +[50] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. Nice-SLAM: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12786-12796, 2022. 2 \ No newline at end of file diff --git a/2024/GenNBV_ Generalizable Next-Best-View Policy for Active 3D Reconstruction/images.zip b/2024/GenNBV_ Generalizable Next-Best-View Policy for Active 3D Reconstruction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b3b4861f7512023c7cb522ea690d86ee11961c04 --- /dev/null +++ b/2024/GenNBV_ Generalizable Next-Best-View Policy for Active 3D Reconstruction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f7a7f77cad773bdda530bfcd9a0ced5a588e96741826dc2de541e3dc9064f4b8 +size 522204 diff --git a/2024/GenNBV_ Generalizable Next-Best-View Policy for Active 3D Reconstruction/layout.json b/2024/GenNBV_ Generalizable Next-Best-View Policy for Active 3D Reconstruction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3110a9cfaf4f06fb0e546506e9411e650ab8e2c2 --- /dev/null +++ b/2024/GenNBV_ Generalizable Next-Best-View Policy for Active 3D Reconstruction/layout.json @@ -0,0 +1,7661 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 60, + 103, + 533, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 103, + 533, + 120 + ], + "spans": [ + { + "bbox": [ + 60, + 103, + 533, + 120 + ], + "type": "text", + "content": "GenNBV: Generalizable Next-Best-View Policy for Active 3D Reconstruction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 108, + 141, + 489, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 141, + 489, + 159 + ], + "spans": [ + { + "bbox": [ + 108, + 141, + 489, + 159 + ], + "type": "text", + "content": "Xiao Chen" + }, + { + "bbox": [ + 108, + 141, + 489, + 159 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 108, + 141, + 489, + 159 + ], + "type": "text", + "content": " Quanyi Li" + }, + { + "bbox": [ + 108, + 141, + 489, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 108, + 141, + 489, + 159 + ], + "type": "text", + "content": " Tai Wang" + }, + { + "bbox": [ + 108, + 141, + 489, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 108, + 141, + 489, + 159 + ], + "type": "text", + "content": " Tianfan Xue" + }, + { + "bbox": [ + 108, + 141, + 489, + 159 + ], + "type": "inline_equation", + "content": "^{2,\\dagger}" + }, + { + "bbox": [ + 108, + 141, + 489, + 159 + ], + "type": "text", + "content": " Jiangmiao Pang" + }, + { + "bbox": [ + 108, + 141, + 489, + 159 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 159, + 466, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 159, + 466, + 175 + ], + "spans": [ + { + "bbox": [ + 137, + 159, + 466, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 137, + 159, + 466, + 175 + ], + "type": "text", + "content": "Shanghai AI Laboratory " + }, + { + "bbox": [ + 137, + 159, + 466, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 137, + 159, + 466, + 175 + ], + "type": "text", + "content": "The Chinese University of Hong Kong" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 176, + 545, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 176, + 545, + 189 + ], + "spans": [ + { + "bbox": [ + 53, + 176, + 545, + 189 + ], + "type": "text", + "content": "{cx123,tfxue}@ie.cuhk.edu.hk quanyili0057@gmail.com {wangtai,pangjiangmiao}@pjlab.org.cn" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 50, + 201, + 267, + 319 + ], + "blocks": [ + { + "bbox": [ + 50, + 201, + 267, + 319 + ], + "lines": [ + { + "bbox": [ + 50, + 201, + 267, + 319 + ], + "spans": [ + { + "bbox": [ + 50, + 201, + 267, + 319 + ], + "type": "image", + "image_path": "97e6c475f76b6f9fa9f7dbf28d5c5131ec961abd99ab0e3a9e7143b9500460d2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 331, + 546, + 365 + ], + "lines": [ + { + "bbox": [ + 46, + 331, + 546, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 546, + 365 + ], + "type": "text", + "content": "Figure 1. To determine the best view for 3D reconstruction, previous methods only chose from hand-crafted action space or based on object-centric capturing, lacking the ability to generalize to unforeseen scenes (Left). With our end-to-end trained, generalized free-space policy, it can generalize to unseen objects, enabling the captured drone to image from any viewpoint (Right)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 276, + 202, + 536, + 320 + ], + "blocks": [ + { + "bbox": [ + 276, + 202, + 536, + 320 + ], + "lines": [ + { + "bbox": [ + 276, + 202, + 536, + 320 + ], + "spans": [ + { + "bbox": [ + 276, + 202, + 536, + 320 + ], + "type": "image", + "image_path": "ded81a23d7169a288f767c4332cb809663f91477947822176507cd55a0374d8c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 374, + 192, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 374, + 192, + 387 + ], + "spans": [ + { + "bbox": [ + 143, + 374, + 192, + 387 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 45, + 395, + 290, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 395, + 290, + 670 + ], + "spans": [ + { + "bbox": [ + 45, + 395, + 290, + 670 + ], + "type": "text", + "content": "While recent advances in neural radiance field enable realistic digitization for large-scale scenes, the image-capturing process is still time-consuming and labor-intensive. Previous works attempt to automate this process using the Next-Best-View (NBV) policy for active 3D reconstruction. However, the existing NBV policies heavily rely on hand-crafted criteria, limited action space, or perscense optimized representations. These constraints limit their cross-dataset generalizability. To overcome them, we propose GenNBV, an end-to-end generalizable NBV policy. Our policy adopts a reinforcement learning (RL)-based framework and extends typical limited action space to 5D free space. It empowers our agent drone to scan from any viewpoint, and even interact with unseen geometries during training. To boost the cross-dataset generalizability, we also propose a novel multi-source state embedding, including geometric, semantic, and action representations. We establish a benchmark using the Isaac Gym simulator with the Houses3K and OmniObject3D datasets to evaluate this NBV policy. Experiments demonstrate that our policy achieves a " + }, + { + "bbox": [ + 45, + 395, + 290, + 670 + ], + "type": "inline_equation", + "content": "98.26\\%" + }, + { + "bbox": [ + 45, + 395, + 290, + 670 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 395, + 290, + 670 + ], + "type": "inline_equation", + "content": "97.12\\%" + }, + { + "bbox": [ + 45, + 395, + 290, + 670 + ], + "type": "text", + "content": " coverage ratio on unseen building-scale objects from these datasets, respectively, outperforming prior solutions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 374, + 386, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 374, + 386, + 387 + ], + "spans": [ + { + "bbox": [ + 306, + 374, + 386, + 387 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 395, + 545, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 395, + 545, + 516 + ], + "spans": [ + { + "bbox": [ + 304, + 395, + 545, + 516 + ], + "type": "text", + "content": "Recent advances in 3D reconstruction [23, 38, 39] and neural rendering [20, 37, 48] have significantly enhanced the quality of 3D digitization of large scenes, such as buildings and city landmarks [11, 17, 18, 43, 47]. However, image-capturing process still remains time-consuming and labor-intensive. To scan a building-scale scene using a drone, it may take several days' effort of a professional team to ensure a full coverage. Moreover, even professional pilots may miss some areas at the first scan, leading to multiple rounds for rescanning." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 518, + 546, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 518, + 546, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 518, + 546, + 686 + ], + "type": "text", + "content": "To alleviate the effort in manual scanning, active 3D reconstruction algorithms, especially the Next-Best-View (NBV) policy, have emerged as a promising approach to automate view planning. However, typical rule-based NBV policies [9, 14, 22, 26, 28, 46] heavily rely on empirically designed view-selection criteria and hand-crafted action space, such as a hemisphere as shown in Fig. 1, which limit their generalizability to unseen scenes. To enhance generalization capacity, pioneering works on learning-based NBV policy have been proposed [3, 4, 26, 46]. Nevertheless, they are constrained into impractical setups in real-world scenarios, such as object-centric capturing and limited action space for ground robots, resulting in incomplete 3D reconstruction because of significant self-occlusion." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": "Therefore, in this work, we study this problem: \"How to design an NBV policy that supports exploration in any" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 160, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 160, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 160, + 703 + ], + "type": "text", + "content": "*Project website: gennbv.tech" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 59, + 703, + 138, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 138, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 138, + 712 + ], + "type": "text", + "content": "† Corresponding authors." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16436" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": "space and also can generalize to unseen building-scale geometries?\" Building upon the design of previous reinforcement learning-based (RL-based) NBV policies [3, 4, 26], our setup faces the following new challenges that were not fully studied before: 1) Designing an easy-to-generalize action space, rather than object-specific spaces used in previous work, such as hand-crafted candidates or a constrained hemisphere; 2) With a large action space, an informative representation is necessary to guide the policy to efficiently find an optimal capturing trajectory; 3) Proposing generic reward functions and developing the distributed training procedure for better generalizability." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 218, + 289, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 218, + 289, + 410 + ], + "spans": [ + { + "bbox": [ + 46, + 218, + 289, + 410 + ], + "type": "text", + "content": "To overcome these two challenges, we propose GenNBV, an end-to-end generalizable NBV policy. It has a novel design of action spaces and state embeddings, which allows free space exploration and ensures a large coverage when applying to unseen objects. We firstly extend former limited action space, like a hemisphere, to 5D free space. As shown in Fig. 1, this new space is composed of a position subspace of approximately " + }, + { + "bbox": [ + 46, + 218, + 289, + 410 + ], + "type": "inline_equation", + "content": "20m \\times 20m \\times 10m" + }, + { + "bbox": [ + 46, + 218, + 289, + 410 + ], + "type": "text", + "content": " and an omnidirectional heading subspace. This free space enables our agent drone to scan from any viewpoint. Moreover, with much larger action space, our RL-based policy can be easily generalized to building-scale geometries without concerning the potential scale shift from training to evaluation. On the contrary, many hand-crafted designs [14, 22] limit available viewpoints to a small predefined set, and thus cannot generalize if there is drastic change in scales." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 411, + 289, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 289, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 289, + 628 + ], + "type": "text", + "content": "Second, in order to generalize a well-trained NBV policy to unseen scenes during evaluation, we propose a novel state representation directly from raw sensory inputs to guide the policy. In contrast to the widely-used neural radiance field (NeRF) [14, 22, 28, 34, 46], our representation does not need time-consuming per-scene optimization. Specifically, our NBV policy incorporates a multi-source state embedding, which comprises three key components: a novel geometric embedding, a semantic embedding, and an action embedding. The geometric embedding is derived from multi-view depth maps and represents an encoded probabilistic 3D grid, while the semantic embedding is extracted from RGB images, and the action embedding is an encoded viewpoint sequence. In contrast to 2D representations used in previous RL-based NBV policies [3, 4], our multi-source state embedding, extracted from a probabilistic 3D grid, is a more comprehensive representation that supports a robust prediction of next viewpoints." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "type": "text", + "content": "To validate the effectiveness of GenNBV, we construct a benchmark for training and evaluation with Houses3K [25] dataset from the NVIDIA Isaac Gym [19] simulator. We further test its generalization ability on novel buildings and object categories from OmniObject3D [42] and Objaverse [6] datasets. Our evaluation metrics quantify the completeness, accuracy, and efficiency for active 3D re" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 168 + ], + "type": "text", + "content": "construction. To consolidate completeness and efficiency into a single number, we also propose a metric, Area Under the Curve (AUC) of the coverage. Compared with heuristic [12], information gain-based [12, 14, 46], and RL-based baselines [26], our method achieves the best result under different metrics. We also provide visualizations to demonstrate the generalizability of GenNBV. A demo video is attached in the appendix to further illustrate our method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 178, + 392, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 178, + 392, + 192 + ], + "spans": [ + { + "bbox": [ + 306, + 178, + 392, + 192 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 199, + 547, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 199, + 547, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 199, + 547, + 342 + ], + "type": "text", + "content": "Traditional 3D Reconstruction. Photometry and geometry are two crucial aspects of reconstruction evaluation. Neural implicit representations [20, 23] have shown progress in photometric rendering but faces challenges like time-consuming optimization and poor generalization, hindering its application in real-time reconstruction scenarios like 3D SLAM [21, 30, 36, 50]. Geometry, in contrast, typically corresponds to 3D representations such as point clouds and 3D mesh and is directly related to issues like collision avoidance. Considering these properties, researchers tend to focus on geometric reconstruction in practical applications and we also follow this stream." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 342, + 547, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 547, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 547, + 437 + ], + "type": "text", + "content": "Active 3D Reconstruction. Active 3D reconstruction is a promising field that has not been thoroughly benchmarked yet. The pipeline of active 3D reconstruction frameworks alternates between inferring optimal viewpoints, capturing new data, and updating the rebuilt 3D model. With the optimal sequence of key viewpoints, a continuous planning path can be generated using the classic planners, such as Fast Marching Method (FMM) [33]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 438, + 548, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 548, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 548, + 569 + ], + "type": "text", + "content": "Existing works can be pivotally differentiated based on the paradigm of NBV policies: rule-based or learning-based. Rule-based approaches like uncertainty-driven works [12, 14, 28, 46] typically yield the next best view from hand-crafted rules from scene representations, which tends to overfit specific scenes. Most learning-based policies [3, 4, 26] use a deep reinforcement learning algorithm like PPO [32] to sequentially predict the optimal viewpoints based on observation. They must obtain feedback from task-related rewards such as coverage ratio [9, 26] and optimize during massive iteration with task environment." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 570, + 548, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 548, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 548, + 689 + ], + "type": "text", + "content": "The action space is designed based on the paradigm of NBV policies. Most rule-based NBV policies select views from a limited action space, such as a set composed of only one hundred candidate viewpoints [22], a tiny pre-defined space like a hemisphere [9, 14, 28, 46], making it possible to overlook some important viewpoints due to unavailability. Even though learning-based categories further explore the larger action space like a 2D plain [3, 4] or a constrained 3D space [26], their limited action space still prevents them from capturing sufficient details for 3D reconstruction." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "type": "text", + "content": "Scene representation built from history observations, which directly provide reconstruction progress to NBV" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16437" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 73, + 541, + 220 + ], + "blocks": [ + { + "bbox": [ + 50, + 73, + 541, + 220 + ], + "lines": [ + { + "bbox": [ + 50, + 73, + 541, + 220 + ], + "spans": [ + { + "bbox": [ + 50, + 73, + 541, + 220 + ], + "type": "image", + "image_path": "78664c646a8f660fbe0d055e673107469844db204a1d0eab6e3db8e59a6e10a6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 228, + 546, + 275 + ], + "lines": [ + { + "bbox": [ + 46, + 228, + 546, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 546, + 275 + ], + "type": "text", + "content": "Figure 2. Overview of our proposed framework GenNBV. Our end-to-end policy takes the historical multi-source observations as input, transforms them into a more informative scene representation, and produces the next viewpoint position. A reward signal will be returned at training time to optimize the end-to-end policy for maximizing the expected cumulative reward in one episode. Specifically, the signal is the increased coverage ratio after collecting a new viewpoint." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 284, + 287, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 284, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 46, + 284, + 287, + 380 + ], + "type": "text", + "content": "policies, is also a crucial aspect of the active 3D reconstruction framework. Previous works have explored visual representations for NBV policies, such as TSDF [11], neural radiance field [1], and 2D BEV maps [3, 10, 44]. However, implicit representations are hard to jointly with learning-based NBV policies, and 2D BEV map lacks sufficient information for large-scale outdoor scenes that contain numerous geometric details in 3D space." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 384, + 288, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 288, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 288, + 517 + ], + "type": "text", + "content": "In our benchmark, we select Scan-RL [26] as our main RL-based baseline. Compared to it, we introduce a much larger action space and informative visual representations for better generalizability. Additionally, we don't benchmark the mentioned baselines [3, 4], mainly due to their reliance on pretraining from human demonstration using imitation learning, while our focus lies on learning through iterations and end-to-end training. Different from our 3D free-space movement, they only explore the traversable 2D areas of indoor scenes using ground robots, thus limiting their motion space in 2D." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "text", + "content": "Generalizable Reinforcement Learning. Increasing the diversity of training data leads to better generalizability [5]. In the robotics field, domain randomization [41] allows legged robots to walk on various terrains absent from the training environment. For end-to-end driving policy, training in large-scale synthetic and realistic scenarios improves the safety of autonomous cars [15, 16] in unseen test scenarios. The advanced work [7] presents an RL-based formulation for the view planning problem. Based on similar frameworks of active 3D reconstruction, Scan-RL [26] and SCONE [9] are two pioneering works showing the generalizability of their frameworks, while both of them suffer from the constrain of viewpoint sampling in limited space and lack of in-depth analysis. In this work, we tackle the aforementioned issues by predicting NBV in free 3D space with a learning-based planner, which is trained with a" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 284, + 545, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 284, + 545, + 309 + ], + "spans": [ + { + "bbox": [ + 305, + 284, + 545, + 309 + ], + "type": "text", + "content": "dataset containing buildings in various shapes and poses for acquiring generalizability for unseen buildings." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 318, + 389, + 332 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 318, + 389, + 332 + ], + "spans": [ + { + "bbox": [ + 306, + 318, + 389, + 332 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 338, + 546, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 338, + 546, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 546, + 506 + ], + "type": "text", + "content": "In this section, we deliver our active 3D reconstruction framework GenNBV, especially the pivotal Next-Best-View policy. An overview of our framework is illustrated in Fig. 2. Firstly, we formulate the NBV problem as a Markov Decision Process (MDP), with a novel design of observations (blue boxes in Fig. 2) and action space in Sec. 3.1. Next, we elaborate our end-to-end NBV policy " + }, + { + "bbox": [ + 304, + 338, + 546, + 506 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 304, + 338, + 546, + 506 + ], + "type": "text", + "content": " (orange box in Fig. 2) in Sec. 3.2. Inspired by Curl [13], we point out that generalizability greatly depends on how to extract embeddings (green box in Fig. 2), which reflect the reconstruction progress, from raw sensory observations like RGB images and camera poses. In Sec. 3.3, we introduce the reward function (right-hand side of Fig. 2) reflecting the optimization objective and the details of policy optimization." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 515, + 534, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 515, + 534, + 528 + ], + "spans": [ + { + "bbox": [ + 306, + 515, + 534, + 528 + ], + "type": "text", + "content": "3.1. Formulation of the Next-Best-View Problem" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 533, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 641 + ], + "type": "text", + "content": "We formulate the NBV problem as learning an optimal policy " + }, + { + "bbox": [ + 304, + 533, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 304, + 533, + 545, + 641 + ], + "type": "text", + "content": " that controls the capturing process, such that enough information is captured for large-scale scene reconstruction, with limited decision-making budgets. As capturing, transferring, and processing a large set of captured images introduce significant computational costs, we also want to design a policy that also minimizes the number of captured images. Therefore, our policy only captures sparse keyframes that sufficiently record all details of objects." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 642, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 546, + 713 + ], + "type": "text", + "content": "As shown in Fig. 2, at each time step " + }, + { + "bbox": [ + 304, + 642, + 546, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 642, + 546, + 713 + ], + "type": "text", + "content": ", the agent receives a visual observation " + }, + { + "bbox": [ + 304, + 642, + 546, + 713 + ], + "type": "inline_equation", + "content": "o_t" + }, + { + "bbox": [ + 304, + 642, + 546, + 713 + ], + "type": "text", + "content": ", takes an action " + }, + { + "bbox": [ + 304, + 642, + 546, + 713 + ], + "type": "inline_equation", + "content": "a_t" + }, + { + "bbox": [ + 304, + 642, + 546, + 713 + ], + "type": "text", + "content": ", infers the action at the next time step " + }, + { + "bbox": [ + 304, + 642, + 546, + 713 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 304, + 642, + 546, + 713 + ], + "type": "text", + "content": " to move to a new location, and then receives a new visual observation, repeating this interaction with the environment until the episode ends. Our simulated agent is embodied as CrazyFlie [8], a type of unmanned" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16438" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "aerial vehicle equipped with various sensors, including an RGB-D camera and an IMU, to execute data collection for reconstruction. We discussed the details of the observation space and action space below." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 121, + 288, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 288, + 228 + ], + "type": "text", + "content": "Observation Space. As shown in the left column of Fig. 2, at each time step, the agent receives an RGB image " + }, + { + "bbox": [ + 46, + 121, + 288, + 228 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 46, + 121, + 288, + 228 + ], + "type": "text", + "content": ", a depth map " + }, + { + "bbox": [ + 46, + 121, + 288, + 228 + ], + "type": "inline_equation", + "content": "D_{t}" + }, + { + "bbox": [ + 46, + 121, + 288, + 228 + ], + "type": "text", + "content": ", and a state vector including the heading (yaw and pitch) and position " + }, + { + "bbox": [ + 46, + 121, + 288, + 228 + ], + "type": "inline_equation", + "content": "(\\mathrm{x},\\mathrm{y},\\mathrm{z})" + }, + { + "bbox": [ + 46, + 121, + 288, + 228 + ], + "type": "text", + "content": " of the onboard camera. The observation " + }, + { + "bbox": [ + 46, + 121, + 288, + 228 + ], + "type": "inline_equation", + "content": "o_{t}" + }, + { + "bbox": [ + 46, + 121, + 288, + 228 + ], + "type": "text", + "content": " (yellow box in Fig. 2) consists of all previously captured images in one episode, historical actions, and the current captures and actions. With this information as input, the policy network can estimate the progress of data collection and determine where to scan next." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 230, + 287, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 230, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 46, + 230, + 287, + 315 + ], + "type": "text", + "content": "Action Space. Unlike most previous NBV algorithms, we use larger action space in order to cover all details of objects. Specifically, we use the camera location and camera angle as our action space, which is a 5-dimension vector consisting of 3D position coordinates and 2D rotation angles (yaw-axis and pitch-axis). We restrict the roll-axis rotation as it is not supported in all drone platforms." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 322, + 219, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 322, + 219, + 335 + ], + "spans": [ + { + "bbox": [ + 47, + 322, + 219, + 335 + ], + "type": "text", + "content": "3.2. Generalizable State Embedding" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 341, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 341, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 341, + 287, + 460 + ], + "type": "text", + "content": "To ensure that the policy learned on one set of 3D objects can generalize to objects with different appearances and structures, a smart state embedding of raw sensor observations is needed to capture invariant features across different objects. Previous methods [3, 26] only extract representations in 2D space, which are very sensible to appearance changes, and our experiments have shown that policies only trained on 2D features may not generalize to different objects. Instead, we propose a multi-source representation that has better generalizability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 460, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 567 + ], + "type": "text", + "content": "Specifically, we first build two mid-level representations that can better model the relationship between the objects and our agent: a 3D geometric representation " + }, + { + "bbox": [ + 46, + 460, + 287, + 567 + ], + "type": "inline_equation", + "content": "F^G" + }, + { + "bbox": [ + 46, + 460, + 287, + 567 + ], + "type": "text", + "content": " from depth maps and a semantic representation " + }, + { + "bbox": [ + 46, + 460, + 287, + 567 + ], + "type": "inline_equation", + "content": "F^S" + }, + { + "bbox": [ + 46, + 460, + 287, + 567 + ], + "type": "text", + "content": " from RGB images. Then we encode these representations and concatenate them with a pose embedding into state embedding " + }, + { + "bbox": [ + 46, + 460, + 287, + 567 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 46, + 460, + 287, + 567 + ], + "type": "text", + "content": ", guiding the NBV policy for subsequent decision-making. The overall encoder for state embedding is shown in Fig. 2, and details of each representation are discussed below." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "type": "text", + "content": "Geometric Representation One simple way to record the geometry of a 3D object is using a binary 3D occupancy map [3], where the value of each cell in the 3D cube indicates whether a cell contains a 3D object or not. However, since the 3D representation is gradually updated with newly captured data, this simple binary occupancy map cannot differentiate a real unoccupied cell from an unscanned cell. An unscanned cell is a strong indicator that the agent should capture more data in this region, while no further scan is needed for a real unoccupied cell." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Previous works [3, 10, 44] simplify the 3D scene representation into a 2D Bird's Eye View map for actively recon" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 275 + ], + "type": "text", + "content": "structuring indoor scenes. However, the 2D BEV map lacks sufficient information for our large-scale outdoor scenes that contain many geometric details in 3D space. In addition, their wheeled robotic platforms are constrained to freely scan and represent these outdoor scenes. Therefore, to model the scanning process in 3D free space, we employ the probabilistic 3D grid [40] as our geometric 3D representation, which records the probability of each 3D voxel being captured or not. Specifically, we first obtain a 3D point cloud in the world coordinate by back-projecting all 2D pixels to 3D points, using the depth map " + }, + { + "bbox": [ + 304, + 72, + 545, + 275 + ], + "type": "inline_equation", + "content": "D_{t}" + }, + { + "bbox": [ + 304, + 72, + 545, + 275 + ], + "type": "text", + "content": ", camera intrinsic parameters, and camera pose " + }, + { + "bbox": [ + 304, + 72, + 545, + 275 + ], + "type": "inline_equation", + "content": "a_{t}" + }, + { + "bbox": [ + 304, + 72, + 545, + 275 + ], + "type": "text", + "content": ". By voxelizing the obtained point cloud, we then build a 3D occupancy grid that explicitly indicates the binary state (occupied or free) in this 3D space. Subsequently, we represent this 3D grid as a probabilistic occupancy grid " + }, + { + "bbox": [ + 304, + 72, + 545, + 275 + ], + "type": "inline_equation", + "content": "F_{t}^{G}" + }, + { + "bbox": [ + 304, + 72, + 545, + 275 + ], + "type": "text", + "content": " and extend the state space of voxels with three states (occupied, free, unknown)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 275, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 275, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 275, + 545, + 384 + ], + "type": "text", + "content": "During each scanning (one episode in RL), at each step " + }, + { + "bbox": [ + 304, + 275, + 545, + 384 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 304, + 275, + 545, + 384 + ], + "type": "text", + "content": ", we update the probabilistic occupancy grid " + }, + { + "bbox": [ + 304, + 275, + 545, + 384 + ], + "type": "inline_equation", + "content": "F_{t + 1}^{G}" + }, + { + "bbox": [ + 304, + 275, + 545, + 384 + ], + "type": "text", + "content": " based on the preceding grid " + }, + { + "bbox": [ + 304, + 275, + 545, + 384 + ], + "type": "inline_equation", + "content": "F_{t}^{G}" + }, + { + "bbox": [ + 304, + 275, + 545, + 384 + ], + "type": "text", + "content": " and current observation. Specifically, the grid is updated through Bresenham's line algorithm [2], which casts the ray path in 3D space between the camera viewpoint and the endpoints among the point cloud back-projected from depth " + }, + { + "bbox": [ + 304, + 275, + 545, + 384 + ], + "type": "inline_equation", + "content": "D_{t + 1}" + }, + { + "bbox": [ + 304, + 275, + 545, + 384 + ], + "type": "text", + "content": ". Following the classical occupancy grid mapping algorithm [40], we have the log-odds formulation of occupancy probability:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 350, + 392, + 545, + 405 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 392, + 545, + 405 + ], + "spans": [ + { + "bbox": [ + 350, + 392, + 545, + 405 + ], + "type": "interline_equation", + "content": "\\log \\operatorname {O d d} \\left(v _ {i} \\mid z _ {j}\\right) = \\log \\operatorname {O d d} \\left(v _ {i}\\right) + C, \\tag {1}", + "image_path": "af77ed000b0942ede67ac9720a7adc81acf2204870024bea5f1f679f96e2d3ed.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 413, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 545, + 544 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 413, + 545, + 544 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 304, + 413, + 545, + 544 + ], + "type": "text", + "content": " denotes the occupancy probability of " + }, + { + "bbox": [ + 304, + 413, + 545, + 544 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 304, + 413, + 545, + 544 + ], + "type": "text", + "content": " voxel in the grid " + }, + { + "bbox": [ + 304, + 413, + 545, + 544 + ], + "type": "inline_equation", + "content": "F_{t}^{G}" + }, + { + "bbox": [ + 304, + 413, + 545, + 544 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 413, + 545, + 544 + ], + "type": "inline_equation", + "content": "z_{j}" + }, + { + "bbox": [ + 304, + 413, + 545, + 544 + ], + "type": "text", + "content": " is the measurement event that " + }, + { + "bbox": [ + 304, + 413, + 545, + 544 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 304, + 413, + 545, + 544 + ], + "type": "text", + "content": " camera ray passes through this voxel and " + }, + { + "bbox": [ + 304, + 413, + 545, + 544 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 413, + 545, + 544 + ], + "type": "text", + "content": " is an empirical constant. The derivation can be found in Appendix. Thus, we update the log-odds occupancy probability of each voxel in the grid " + }, + { + "bbox": [ + 304, + 413, + 545, + 544 + ], + "type": "inline_equation", + "content": "F_{t}^{G}" + }, + { + "bbox": [ + 304, + 413, + 545, + 544 + ], + "type": "text", + "content": " by adding a constant one time when a single camera ray passes through this voxel. Note that the probabilistic occupancy grid " + }, + { + "bbox": [ + 304, + 413, + 545, + 544 + ], + "type": "inline_equation", + "content": "F^{G}" + }, + { + "bbox": [ + 304, + 413, + 545, + 544 + ], + "type": "text", + "content": " is continually updated within one episode. Finally, the occupancy status of voxels is classified into three categories: unknown, occupied, and free, using preset probability thresholds." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 545, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 641 + ], + "type": "text", + "content": "Semantic Representation. Geometric representation enables agents to comprehend spatial occupancy, yet it's insufficient for perceiving the environment. For example, when observing a hole in an object, the agent may struggle to differentiate between incomplete scanning and the actual presence of a hole in the object. In such cases, the semantic information contained in the captured RGB images can help the agent distinguish between these two scenarios." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "To provide semantic information, we employ a preprocessing module that takes as input the current frame of RGB image " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": " and preceding " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": " frames and converts these frames " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "[I_t,I_{t - 1},\\dots,I_{t - k}]" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": " to grayscale, and concatenate them as output, following [26]. Then the preprocessed frames are fed into a two-layer convolutional network for extracting" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16439" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 179, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 179, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 179, + 84 + ], + "type": "text", + "content": "the semantic representation " + }, + { + "bbox": [ + 47, + 72, + 179, + 84 + ], + "type": "inline_equation", + "content": "F_{t}^{S}" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 85, + 287, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 85, + 287, + 181 + ], + "spans": [ + { + "bbox": [ + 46, + 85, + 287, + 181 + ], + "type": "text", + "content": "State Embedding. To further combine the semantic and geometric embeddings, we first encode them to " + }, + { + "bbox": [ + 46, + 85, + 287, + 181 + ], + "type": "inline_equation", + "content": "s_t^S = f^S (F_t^S)" + }, + { + "bbox": [ + 46, + 85, + 287, + 181 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 85, + 287, + 181 + ], + "type": "inline_equation", + "content": "s_t^G = f^G (F_t^G)" + }, + { + "bbox": [ + 46, + 85, + 287, + 181 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 85, + 287, + 181 + ], + "type": "inline_equation", + "content": "f^{*}" + }, + { + "bbox": [ + 46, + 85, + 287, + 181 + ], + "type": "text", + "content": " are learnable networks Linear(Flatten(x)), as shown in Fig. 2. Subsequently, we combine them with the historical action embedding " + }, + { + "bbox": [ + 46, + 85, + 287, + 181 + ], + "type": "inline_equation", + "content": "s_t^A = \\mathrm{Linear}(a_{1:t})" + }, + { + "bbox": [ + 46, + 85, + 287, + 181 + ], + "type": "text", + "content": " to generate the final state embedding " + }, + { + "bbox": [ + 46, + 85, + 287, + 181 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 46, + 85, + 287, + 181 + ], + "type": "text", + "content": ", as the input to the policy network. This process can be formulated as: " + }, + { + "bbox": [ + 46, + 85, + 287, + 181 + ], + "type": "inline_equation", + "content": "s_t = \\mathrm{Linear}(s_t^G;s_t^S;s_t^A)" + }, + { + "bbox": [ + 46, + 85, + 287, + 181 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 182, + 287, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 182, + 287, + 242 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 287, + 242 + ], + "type": "text", + "content": "Policy Network. Taking the state embedding " + }, + { + "bbox": [ + 46, + 182, + 287, + 242 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 46, + 182, + 287, + 242 + ], + "type": "text", + "content": " as input, the policy network is a 3-layer multi-layer perceptron network (MLP) whose output is used to parameterize a normal distribution over action space. In this way, the action can be drawn from the stochastic policy " + }, + { + "bbox": [ + 46, + 182, + 287, + 242 + ], + "type": "inline_equation", + "content": "a \\sim \\pi(\\cdot | o_t)" + }, + { + "bbox": [ + 46, + 182, + 287, + 242 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 248, + 237, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 248, + 237, + 262 + ], + "spans": [ + { + "bbox": [ + 47, + 248, + 237, + 262 + ], + "type": "text", + "content": "3.3. Reward Function and Optimization" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 266, + 287, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 287, + 327 + ], + "type": "text", + "content": "We train the end-to-end policy with reinforcement learning (RL) and hence design reward functions to precisely reflect the task objective for 3D reconstruction. The policy is optimized with proximal policy optimization [32] (PPO) for parallelizing sampling." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 327, + 287, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 287, + 388 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 287, + 388 + ], + "type": "text", + "content": "Reward Functions. With the occupancy probability " + }, + { + "bbox": [ + 46, + 327, + 287, + 388 + ], + "type": "inline_equation", + "content": "F_{t}^{G}" + }, + { + "bbox": [ + 46, + 327, + 287, + 388 + ], + "type": "text", + "content": " at time step " + }, + { + "bbox": [ + 46, + 327, + 287, + 388 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 327, + 287, + 388 + ], + "type": "text", + "content": ", we can threshold each voxel with an empirical bound to determine if it is occupied. This discrimination process outputs a binary occupancy grid with " + }, + { + "bbox": [ + 46, + 327, + 287, + 388 + ], + "type": "inline_equation", + "content": "\\tilde{N}_{t}" + }, + { + "bbox": [ + 46, + 327, + 287, + 388 + ], + "type": "text", + "content": " voxels being occupied, which is used to calculate the coverage ratio:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 125, + 397, + 287, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 397, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 125, + 397, + 287, + 422 + ], + "type": "interline_equation", + "content": "\\mathrm{CR}_{t} = \\frac{\\tilde{N}_{t}}{N^{*}}\\cdot 100\\% , \\tag{2}", + "image_path": "c7dde36de5085f61851a8568a8a7d7b990874430a8ce7cb2fe64e898faf30ff9.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 425, + 287, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 425, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 287, + 483 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 425, + 287, + 483 + ], + "type": "inline_equation", + "content": "N^{*}" + }, + { + "bbox": [ + 46, + 425, + 287, + 483 + ], + "type": "text", + "content": " is the number of ground-truth occupied voxels representing the surface of objects. To encourage our NBV policy to cover as many unseen areas of objects as possible, we use the difference of coverage ratio (CR) between two consecutive steps as the main reward function " + }, + { + "bbox": [ + 46, + 425, + 287, + 483 + ], + "type": "inline_equation", + "content": "r^{CR}" + }, + { + "bbox": [ + 46, + 425, + 287, + 483 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 119, + 491, + 287, + 507 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 491, + 287, + 507 + ], + "spans": [ + { + "bbox": [ + 119, + 491, + 287, + 507 + ], + "type": "interline_equation", + "content": "r _ {t + 1} ^ {\\mathrm {C R}} = \\mathrm {C R} _ {t + 1} - \\mathrm {C R} _ {t}. \\tag {3}", + "image_path": "6f2b6c44e83bb1eab8608527ff378e42470500b55d651cebb7fee18db734cf55.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 514, + 287, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 514, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 46, + 514, + 287, + 609 + ], + "type": "text", + "content": "In free-space exploration, we also need to avoid collision. Previous limited-space agents [9, 14, 46] do not consider collision avoidance since their search space, like hemisphere, is by-design safe. Thus, we add a negative reward for collision and terminate the episode if a collision happens. We also implement a negative reward when the number of captured keyframes is over an empirical threshold to improve the path efficiency." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 611, + 287, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 611, + 287, + 682 + ], + "spans": [ + { + "bbox": [ + 46, + 611, + 287, + 682 + ], + "type": "text", + "content": "Policy Optimization. Once the reward functions has been specified, the policy can be learned through any off-the-shelf RL algorithm. In this work, we use PPO as thousands of workers can be parallelized to improve the sample efficiency. Specifically, given our parameterized policy " + }, + { + "bbox": [ + 46, + 611, + 287, + 682 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}" + }, + { + "bbox": [ + 46, + 611, + 287, + 682 + ], + "type": "text", + "content": ", the objective of PPO is to maximize the following function:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 85, + 689, + 287, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 689, + 287, + 717 + ], + "spans": [ + { + "bbox": [ + 85, + 689, + 287, + 717 + ], + "type": "interline_equation", + "content": "L (\\theta) = \\mathbb {E} _ {t} \\left[ \\frac {\\pi_ {\\theta} (a _ {t} | s _ {t})}{\\pi_ {\\theta_ {\\mathrm {o l d}}} (a _ {t} | s _ {t})} A ^ {\\pi_ {\\theta_ {\\mathrm {o l d}}}} (s _ {t}, a _ {t}) \\right], \\tag {4}", + "image_path": "8de1a518882bac35da10a9eb89ea94e2bd1979847ba17076ba2e50d80124b139.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "A^{\\pi_{\\theta_{\\mathrm{old}}}}(s_t, a_t)" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " is the advantage function that measures the value of taking action " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "a_t" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " at state " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " under the current policy " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta_{\\mathrm{old}}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": ". To prevent significant deviation of the new policy from the old policy, PPO incorporates a clipped surrogate objective function:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 137, + 545, + 167 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 137, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 317, + 137, + 545, + 167 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L ^ {\\mathrm {C L I P}} (\\theta) = \\mathbb {E} _ {t} [ \\min (\\eta_ {t} (\\theta) A ^ {\\pi_ {\\theta_ {\\text {o l d}}}} (s _ {t}, a _ {t}), \\tag {5} \\\\ \\left. \\operatorname {c l i p} \\left(r _ {t} (\\theta), 1 - \\epsilon , 1 + \\epsilon\\right) A ^ {\\pi_ {\\theta_ {\\text {o l d}}}} \\left(s _ {t}, a _ {t}\\right) \\right], \\\\ \\end{array}", + "image_path": "98d5b064b3f65eb47249c25db23357ef965bec3e26d8bbb8f506a708b644be0d.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 173, + 545, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 173, + 545, + 201 + ], + "spans": [ + { + "bbox": [ + 304, + 173, + 545, + 201 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 173, + 545, + 201 + ], + "type": "inline_equation", + "content": "\\eta_t(\\theta) = \\frac{\\pi_\\theta(a_t|s_t)}{\\pi_{\\theta_{\\mathrm{old}}}(a_t|s_t)}" + }, + { + "bbox": [ + 304, + 173, + 545, + 201 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 173, + 545, + 201 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 173, + 545, + 201 + ], + "type": "text", + "content": " is a hyper-parameter that controls the size of the trust region." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 213, + 386, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 213, + 386, + 227 + ], + "spans": [ + { + "bbox": [ + 305, + 213, + 386, + 227 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 233, + 545, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 233, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 304, + 233, + 545, + 329 + ], + "type": "text", + "content": "In this section, we conduct experiments on Houses3K [25], OmniObject3D [42], and Objaverse [6]. For our policy and other learning-based policies, we train them on Houses3K training set. Then we evaluate all policies on the Houses3K test dataset and the OmniObject3D dataset for quantifying the in-distribution and out-of-distribution generalizability. Finally, we show the visualization results on three datasets, demonstrating the effectiveness of the proposed method." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 335, + 422, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 335, + 422, + 348 + ], + "spans": [ + { + "bbox": [ + 305, + 335, + 422, + 348 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 354, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 545, + 521 + ], + "type": "text", + "content": "Simulation Environment. We conduct all experiments in NVIDIA Isaac Gym [19], a physics simulation platform designed for reinforcement learning and robotics research. With GPU-accelerated tensor API and sensor interaction API, we can easily implement customized CUDA kernels to calculate our geometric representation efficiently. Moreover, the sensor simulation is efficient. It can run up to 1000 FPS, significantly reducing the training time. We create the agent drone CrazyFlie [8] and equip it with sensors including RGB-D cameras and IMU. Considering memory limitations, we downsample the image resolution to " + }, + { + "bbox": [ + 304, + 354, + 545, + 521 + ], + "type": "inline_equation", + "content": "400 \\times 400" + }, + { + "bbox": [ + 304, + 354, + 545, + 521 + ], + "type": "text", + "content": ". The vertical field of view of this onboard camera is " + }, + { + "bbox": [ + 304, + 354, + 545, + 521 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 304, + 354, + 545, + 521 + ], + "type": "text", + "content": ". We also set up a point light source with a fixed position and constant intensity in Isaac Gym." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 521, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 617 + ], + "type": "text", + "content": "Dataset. We conduct our experiments on Houses3K [25], OmniObject3D [42], Objverse 1.0 [6] and Replica [35]. Our model is trained on large-scale 3D objects from Houses3K. To validate the effectiveness and generalizability of our model, we evaluate it on unseen 3D objects from Houses3K and further on batches of unseen cross-domain 3D objects from OmniObject3D, Objverse, and Replica, which include both house and non-house categories." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Houses3K contains 3,000 textured 3D building models. They are divided into 12 batches, with each batch featuring 50 distinct geometries and 5 varying textures. We further eliminate poorly designed geometries based on the following two criteria: 1) objects with complex internal structures (as we are mostly focusing on the surface reconstruction), and 2) objects with redundant bases at the bottom. Following these criteria, the training set consists of 256 objects" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16440" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 125, + 547, + 329 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 545, + 93 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 545, + 93 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 545, + 93 + ], + "type": "text", + "content": "Table 1. Evaluation results of Next-Best-View policies for active 3D reconstruction on Houses3K and the house category from OmniObject3D (cross-dataset generalization). The number of views is set to 30 and 20 for Houses3K and OmniObject3D, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 94, + 545, + 115 + ], + "lines": [ + { + "bbox": [ + 47, + 94, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 47, + 94, + 545, + 115 + ], + "type": "text", + "content": "\\*\\*: the policy is trained with the Houses3K training set and evaluated with holdout Houses3K test set and OmniObject3D house category. " + }, + { + "bbox": [ + 47, + 94, + 545, + 115 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 47, + 94, + 545, + 115 + ], + "type": "text", + "content": " : the policy heavily relies on optimized per-scene representation (NeRF), and thus is directly trained and evaluated on testing objects." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 125, + 547, + 329 + ], + "lines": [ + { + "bbox": [ + 47, + 125, + 547, + 329 + ], + "spans": [ + { + "bbox": [ + 47, + 125, + 547, + 329 + ], + "type": "table", + "html": "
Houses3K Test SetOmniObject3D
NBV PolicyMean AUC ↑Final Coverage Ratio ↑Accuracy (cm) ↓Mean AUC ↑Final Coverage Ratio ↑Accuracy (cm) ↓
HeuristicRandom48.53%58.24%1.3860.13%73.73%0.62
Random Hemisphere71.19%79.72%0.6374.03%84.74%0.48
Uniform Hemisphere82.91%89.71%0.4483.09%92.90%0.41
InfoGain†Uncertainty-Guided [14]83.13%89.31%0.4469.08%92.85%0.42
†ActiveRMap [46]84.86%90.77%0.4469.47%93.15%0.42
RL-based*Scan-RL [26]84.49%91.61%0.4082.17%92.53%0.34
*Scan-RL w/ Our Repre.86.48%92.20%0.3786.14%93.73%0.35
*Ours w/ Scan-RL Repre.87.39%95.63%0.3886.81%93.64%0.34
*GenNBV (Ours)91.19%98.26%0.3788.63%97.12%0.33
", + "image_path": "9fd38b9c2299fe3e2cc7cb2129c41965695f82d98f8171dfc07b084694c73185.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 50, + 395, + 287, + 517 + ], + "blocks": [ + { + "bbox": [ + 46, + 342, + 288, + 386 + ], + "lines": [ + { + "bbox": [ + 46, + 342, + 288, + 386 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 288, + 386 + ], + "type": "text", + "content": "Table 2. The cross-dataset generalization for non-house categories. We train the baseline Scan-RL and our GenNBV on Houses3K and generalize them to non-house categories from OmniObject3D and an indoor scene from Replica." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 395, + 287, + 517 + ], + "lines": [ + { + "bbox": [ + 50, + 395, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 50, + 395, + 287, + 517 + ], + "type": "table", + "html": "
CategoryDatasetNBV PolicyAUC ↑FCR ↑Acc. ↓
AnimalsOmniObject3DScan-RL76.43%84.98%0.17
GenNBV84.28%94.07%0.16
TrucksOmniObject3DScan-RL69.58%76.37%0.06
GenNBV75.54%83.47%0.03
DinosaursOmniObject3DScan-RL78.47%86.89%0.88
GenNBV84.24%94.03%0.81
Room 0ReplicaScan-RL69.21%80.75%3.12
GenNBV88.53%99.16%3.59
", + "image_path": "8f31a7290e347040cf3971def819014c995cf0e18c2b5356767c8d133dacfa0c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 528, + 287, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 528, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 46, + 528, + 287, + 564 + ], + "type": "text", + "content": "from 6 selected batches, and the test set consists of 50 objects from another batch. All these selected training and test objects have distinct geometries." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 566, + 287, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 566, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 566, + 287, + 662 + ], + "type": "text", + "content": "OmniObject3D offers a collection of 6K high-fidelity objects scanned from real-world sources across 190 typical categories. The house category from OmniObject3D for evaluation has 43 diverse objects. In contrast, Objaverse 1.0 comprises a vast library of 818K 3D synthetic objects spanning 21K categories. The diverse and high-quality nature of these datasets makes them ideal for evaluating our models and visualizing the results." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "Evaluation Metrics. The objective of NBV policies is to capture the most useful information for reconstruction, with the least number of views. We report the (1) Final Coverage Ratio " + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": ", (2) Mean AUC " + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": " of coverage ratio and" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 343, + 545, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 343, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 304, + 343, + 545, + 488 + ], + "type": "text", + "content": "(3) Reconstruction Accuracy (cm) along with the consistent number of views in all tables. Specifically, most prior works [9, 26] separately use the coverage ratio (\\%) and the number of views to evaluate the reconstruction completeness and efficiency for NBV policies. However, the coverage ratio is highly correlated with the number of views. Thus we propose to unify the number of views to a fixed value for all NBV policies during evaluation and use the area under the curve (AUC) of coverage ratio as the main metric for comparison. Also, we calculate the Chamfer Distance between scanned and ground-truth point clouds to represent the reconstruction accuracy (cm)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 488, + 546, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 488, + 546, + 656 + ], + "spans": [ + { + "bbox": [ + 304, + 488, + 546, + 656 + ], + "type": "text", + "content": "Implementation Details. We conduct all experiments in Isaac Gym simulation engine with one NVIDIA Tesla V100 GPU. Our NBV policy is optimized through over 32 million iterations and uses approximately 24 hours of training time on an NVIDIA V100 GPU. All networks are randomly initialized and trained end-to-end. Our implementation refers to the codebase of Legged Gym [31] and the PPO implementation in Stable Baseline3 [27], which is implemented by PyTorch [24]. The ground-truth point clouds on objects' surfaces are sampled by the Poisson Disk sampling method [45] using Open3D API [49]. The Chamfer Distance between the scanned point cloud and the ground-truth point cloud is calculated using PyTorch3D API [29]. Please refer to the Appendix for further details and experiments." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 661, + 450, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 661, + 450, + 673 + ], + "spans": [ + { + "bbox": [ + 306, + 661, + 450, + 673 + ], + "type": "text", + "content": "4.2. Performance Comparison" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 677, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 715 + ], + "type": "text", + "content": "To comprehensively demonstrate the effectiveness and generalizability of GenNBV, we design three levels of evaluation experiments: 1) As shown in Table 1, we show the per" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16441" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 69, + 303, + 220 + ], + "blocks": [ + { + "bbox": [ + 53, + 69, + 303, + 220 + ], + "lines": [ + { + "bbox": [ + 53, + 69, + 303, + 220 + ], + "spans": [ + { + "bbox": [ + 53, + 69, + 303, + 220 + ], + "type": "image", + "image_path": "b0d176c1467ed0538f742e96a7dacf2236e75e6ab7466bd292693e05d4f08725.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 176, + 222, + 238, + 234 + ], + "lines": [ + { + "bbox": [ + 176, + 222, + 238, + 234 + ], + "spans": [ + { + "bbox": [ + 176, + 222, + 238, + 234 + ], + "type": "text", + "content": "(a) Houses3K" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 312, + 76, + 539, + 220 + ], + "blocks": [ + { + "bbox": [ + 312, + 76, + 539, + 220 + ], + "lines": [ + { + "bbox": [ + 312, + 76, + 539, + 220 + ], + "spans": [ + { + "bbox": [ + 312, + 76, + 539, + 220 + ], + "type": "image", + "image_path": "a7283e936288c6a87f1ba831d3104b034d1b5a6a91f72dd06461e157e0ac2cf4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 389, + 222, + 468, + 234 + ], + "lines": [ + { + "bbox": [ + 389, + 222, + 468, + 234 + ], + "spans": [ + { + "bbox": [ + 389, + 222, + 468, + 234 + ], + "type": "text", + "content": "(b) OmniObject3D" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 243, + 547, + 289 + ], + "lines": [ + { + "bbox": [ + 46, + 243, + 547, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 547, + 289 + ], + "type": "text", + "content": "Figure 3. The visualization results of unseen 3D objects reconstructed by Scan-RL [26] and our model to compare the generalizability. (a) Unseen buildings from the test set of Houses3K. (b) Unseen buildings from OmniObject3D. It's quite obvious that some parts of the model reconstructed by Scan-RL are wrong or missing. For example, the second object in the first row has a pillar in the wrong shape. Scan-RL fails to reconstruct the roof edge for the fourth object from OmniObject3D, as shown in the third row." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 300, + 289, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 289, + 409 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 289, + 409 + ], + "type": "text", + "content": "formance of our NBV policy on the Houses3K test set; 2) We generalize the policy trained on Houses3K to the house category from OmniObject3D that has completely different geometric structures and textures compared to the training set. The quantitative result is at Table 1 and the visualization result is as shown in Fig. 3; 3) We also generalize GenNBV to non-house categories from OmniObject3D and scenes with enormous details from Objaverse and Replica [35] to demonstrate its generalizability potential." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 411, + 289, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 289, + 663 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 289, + 663 + ], + "type": "text", + "content": "We implement the following policies as our baselines. 1) Random: This policy randomly generates 5-dim vector " + }, + { + "bbox": [ + 46, + 411, + 289, + 663 + ], + "type": "inline_equation", + "content": "(x,y,z,\\text{pitch},\\text{yaw})" + }, + { + "bbox": [ + 46, + 411, + 289, + 663 + ], + "type": "text", + "content": " among the action space as the next viewpoint. 2) Random Hemisphere: This policy randomly generates the next positions on a pre-defined hemisphere that sufficiently covers all objects of the test set. The headings are constrained to point to the center of the hemisphere. 3) Uniform Hemisphere: All positions are evenly distributed on the previously mentioned hemispheres. 4) Uncertainty-Guided Policy [14]: This NBV policy iteratively selects the next view from a pre-defined viewpoint set according to the uncertainty based on a continually optimized neural radiance field. 5) ActiveRMap [46]: This policy also adopts an iterative selection framework, with multiple objectives including information gain, to select the next viewpoints from the candidate set. 6) Scan-RL [26]: This RL-based policy predicts the next viewpoint from a hemisphere space, relying on the historical RGB images. 7) Ours with Scan-RL's Representation: This free-space NBV policy replaces our representation with Scan-RL's representation only extracted from RGB images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "As shown in Table 1, our GenNBV shows the best in-distribution and out-of-distribution generalizability in both coverage sufficiency and view efficiency, when evaluated on test sets consisting of unseen objects from Houses3K" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 308, + 330, + 545, + 488 + ], + "blocks": [ + { + "bbox": [ + 305, + 299, + 545, + 321 + ], + "lines": [ + { + "bbox": [ + 305, + 299, + 545, + 321 + ], + "spans": [ + { + "bbox": [ + 305, + 299, + 545, + 321 + ], + "type": "text", + "content": "Table 3. Ablation studies of representation categories in our framework on unseen Houses3K test set." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 330, + 545, + 488 + ], + "lines": [ + { + "bbox": [ + 308, + 330, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 308, + 330, + 545, + 488 + ], + "type": "table", + "html": "
Representation CategoryEvaluation Metrics
Probabilistic 3D Grid5-DoF PoseSemantic 2D MapMean AUCFinal Coverage Ratio
Uni-source✓81.06%84.56%
✓69.53%76.61%
✓81.24%87.90%
Multi-source✓✓88.66%96.67%
✓✓89.77%95.31%
✓✓88.30%96.29%
✓✓✓91.19%98.26%
", + "image_path": "244c6c11162acbb41304432ebcc17153ccfe7736e2b67f290985c5cf2dc9f7f2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 494, + 547, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 494, + 547, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 494, + 547, + 601 + ], + "type": "text", + "content": "and OmniObject3D. In particular, the comparison in Table 1 demonstrates the strength of the proposed representations and action space. Note that the Scan-RL's representation (\\* in the table) uses 6 frames, while our semantic representation only needs 2 frames. Moreover, learning-based NBV policies such as GenNBV and Scan-RL, with much larger action space, outperform all rule-based baselines, even if ActiveRMap and Uncertainty-Guided baselines are trained and evaluated on the same 3D objects." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 603, + 547, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 603, + 547, + 675 + ], + "spans": [ + { + "bbox": [ + 304, + 603, + 547, + 675 + ], + "type": "text", + "content": "Non-house Generalization In Table 2, we introduce an experiment to evaluate the generalizability of our GenNBV trained on Houses3K to non-house categories. The targets include 74 animals, 43 trucks and 33 dinosaurs from OmniObject3D, and a challenging indoor scene (Room 0) from Replica. The number of views is set to 20." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "Our GenNBV almost outperforms Scan-RL in all metrics. Particularly, on Replica Room 0, which consists of numerous multi-sized objects with complex occlusion, our" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16442" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 75, + 205, + 194 + ], + "blocks": [ + { + "bbox": [ + 50, + 75, + 205, + 194 + ], + "lines": [ + { + "bbox": [ + 50, + 75, + 205, + 194 + ], + "spans": [ + { + "bbox": [ + 50, + 75, + 205, + 194 + ], + "type": "image", + "image_path": "b0c3e919f63a980a9b583f9aa7e2dde3596478edfd85ffd2421651dcb5de0174.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 204, + 162, + 215 + ], + "lines": [ + { + "bbox": [ + 80, + 204, + 162, + 215 + ], + "spans": [ + { + "bbox": [ + 80, + 204, + 162, + 215 + ], + "type": "text", + "content": "Uncertainty-guided" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 213, + 72, + 373, + 194 + ], + "blocks": [ + { + "bbox": [ + 213, + 72, + 373, + 194 + ], + "lines": [ + { + "bbox": [ + 213, + 72, + 373, + 194 + ], + "spans": [ + { + "bbox": [ + 213, + 72, + 373, + 194 + ], + "type": "image", + "image_path": "de115d3f6207bdc922bc1c62d4fd16d421e2b26f8d9a89687ba5ce3dbb6c4e90.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 271, + 204, + 308, + 214 + ], + "lines": [ + { + "bbox": [ + 271, + 204, + 308, + 214 + ], + "spans": [ + { + "bbox": [ + 271, + 204, + 308, + 214 + ], + "type": "text", + "content": "Scan-RL" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 228, + 546, + 262 + ], + "lines": [ + { + "bbox": [ + 46, + 228, + 546, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 546, + 262 + ], + "type": "text", + "content": "Figure 4. The visualization results of an unseen 3D outdoor scene with enormous details from Objaverse, reconstructed by Uncertainty-guided, Scan-RL and our model. Compared to the uncertainty-guided method and Scan-RL, the scene reconstructed by our method is more watertight and has fewer holes on the ground and building surface, especially in the region highlighted by the red box." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 384, + 72, + 545, + 194 + ], + "blocks": [ + { + "bbox": [ + 384, + 72, + 545, + 194 + ], + "lines": [ + { + "bbox": [ + 384, + 72, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 384, + 72, + 545, + 194 + ], + "type": "image", + "image_path": "80d0874bac5a1455c60639ed88a12461439c227ea43b987e895652f906d81521.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 435, + 204, + 501, + 215 + ], + "lines": [ + { + "bbox": [ + 435, + 204, + 501, + 215 + ], + "spans": [ + { + "bbox": [ + 435, + 204, + 501, + 215 + ], + "type": "text", + "content": "GenNBV (Ours)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 282, + 287, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 282, + 287, + 306 + ], + "spans": [ + { + "bbox": [ + 46, + 282, + 287, + 306 + ], + "type": "text", + "content": "GenNBV is significantly better than Scan-RL. This shows the effectiveness of our free-space NBV policy." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 315, + 141, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 315, + 141, + 327 + ], + "spans": [ + { + "bbox": [ + 47, + 315, + 141, + 327 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 334, + 287, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 334, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 334, + 287, + 369 + ], + "type": "text", + "content": "We implement the ablation study about multi-source state embedding and depth-based representation. Please refer to the Appendix for further experiments." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 369, + 288, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 288, + 537 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 288, + 537 + ], + "type": "text", + "content": "Multi-source State Embedding. In Sec. 3.2, we introduce how to build our multi-source state embedding. Thanks to the same shape of representations and state embedding, we only need to adjust the input dimension of the linear layer of state embedding according to the modal type when ablating the model. Here we reveal the importance of specific representations with the ablation results shown in Table 3. In uni-source experiments, we design the 5-DoF Pose baseline that learns an empirically optimal trajectory, which helps understand the effectiveness of domain knowledge of the dataset. We also evaluate different combinations of representation categories in Table 3, demonstrating the effectiveness of our multi-source state embedding for policy learning in multi-source experiments." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 50, + 578, + 287, + 640 + ], + "blocks": [ + { + "bbox": [ + 47, + 546, + 287, + 569 + ], + "lines": [ + { + "bbox": [ + 47, + 546, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 47, + 546, + 287, + 569 + ], + "type": "text", + "content": "Table 4. Ablation studies of depth-based representations on unseen Houses3K test set, where depth map is the only sensory source." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 578, + 287, + 640 + ], + "lines": [ + { + "bbox": [ + 50, + 578, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 50, + 578, + 287, + 640 + ], + "type": "table", + "html": "
Depth-based RepresentationMean AUCFinal Coverage Ratio
Geometric 2D Map71.86%74.88%
Binary 3D Grid77.28%80.85%
Probabilistic 3D Grid81.06%84.56%
", + "image_path": "b4cc58cbf9dd695a50ab5cf5ca4402ec145e236b3b13ce75833be940143aaa01.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "type": "text", + "content": "Depth-based Representation. Furthermore, we ablate the depth-based representations in Table 4. The experimental results demonstrate that our probabilistic 3D grid is more comprehensive than a binary 3D grid and a geometric 2D map to support the NBV prediction." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 281, + 417, + 294 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 281, + 417, + 294 + ], + "spans": [ + { + "bbox": [ + 306, + 281, + 417, + 294 + ], + "type": "text", + "content": "4.4. Qualitative Results" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 300, + 546, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 546, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 546, + 384 + ], + "type": "text", + "content": "We visualize the reconstruction results generated from the scanning trajectory of a single episode in Fig. 3. It demonstrates that our next-best-view policy can reconstruct objects better in terms of completeness and appearance quality compared to Scan-RL. We further visualize the reconstruction results of an outdoor scene from Objaverse using 30 collected views in Fig. 4." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 395, + 378, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 395, + 378, + 407 + ], + "spans": [ + { + "bbox": [ + 306, + 395, + 378, + 407 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 415, + 546, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 546, + 630 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 546, + 630 + ], + "type": "text", + "content": "This study presents an end-to-end approach for active 3D scene reconstruction, reducing the need for manual intervention. Specifically, the learning-based policy explores how to reconstruct diverse objects in the training stage and thus can generalize to reconstruct unseen objects in a fully autonomous manner. Our controller maneuvers in free space and selects the next best view based on a hybrid scene representation which conveys scene coverage status and thus reconstruction progress. We show the effectiveness of our approach by generalizing it on multiple datasets. The quantitative and qualitative generalization results on holdout Houses3K test set and cross-domain OmniObject3D, including house category and non-house categories, show that our method outperforms other baselines in terms of reconstruction completeness, efficiency and accuracy. Furthermore, the experiment conducted on Objaverse and Replica shows that the policy trained in house-only settings can even generalize to complicated outdoor and indoor scenes." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 641, + 408, + 654 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 641, + 408, + 654 + ], + "spans": [ + { + "bbox": [ + 306, + 641, + 408, + 654 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 661, + 545, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 661, + 545, + 708 + ], + "spans": [ + { + "bbox": [ + 304, + 661, + 545, + 708 + ], + "type": "text", + "content": "This work is supported by Shanghai Artificial Intelligence Laboratory and CUHK Direct Grants (RCFUS) 4055189. We're very grateful to Tao Huang, Zeqi Xiao, Junfeng Long, Mulin Yu, and all reviewers for their insightful advice." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16443" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Michal Adamkiewicz, Timothy Chen, Adam Caccavale, Rachel Gardner, Preston Culbertson, Jeannette Bohg, and Mac Schwager. Vision-only robot navigation in a neural radiance world. IEEE Robotics and Automation Letters, 7(2): 4606-4613, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 146, + 287, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 146, + 287, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 146, + 287, + 168 + ], + "type": "text", + "content": "[2] Jack E Bresenham. Algorithm for computer control of a digital plotter. IBM Systems journal, 4(1):25-30, 1965. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 168, + 288, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 168, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 168, + 288, + 213 + ], + "type": "text", + "content": "[3] Devendra Singh Chaplot, Dhiraj Gandhi, Saurabh Gupta, Abhinav Gupta, and Ruslan Salakhutdinov. Learning to explore using active neural slam. In International Conference on Learning Representations, 2020. 1, 2, 3, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 213, + 287, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 213, + 287, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 213, + 287, + 247 + ], + "type": "text", + "content": "[4] Tao Chen, Saurabh Gupta, and Abhinav Gupta. Learning exploration policies for navigation. In International Conference on Learning Representations, 2019. 1, 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 247, + 288, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 288, + 291 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 288, + 291 + ], + "type": "text", + "content": "[5] Karl Cobbe, Chris Hesse, Jacob Hilton, and John Schulman. Leveraging procedural generation to benchmark reinforcement learning. In International conference on machine learning, pages 2048-2056. PMLR, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 292, + 288, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 292, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 292, + 288, + 357 + ], + "type": "text", + "content": "[6] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 2, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 358, + 288, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 358, + 288, + 412 + ], + "spans": [ + { + "bbox": [ + 53, + 358, + 288, + 412 + ], + "type": "text", + "content": "[7] Mustafa Devrim Kaba, Mustafa Gokhan Uzunbas, and Ser Nam Lim. A reinforcement learning approach to the view planning problem. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6933-6941, 2017. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 414, + 288, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 414, + 288, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 414, + 288, + 479 + ], + "type": "text", + "content": "[8] Wojciech Giernacki, Mateusz Skwierczyński, Wojciech Witwicki, Paweł Wronski, and Piotr Kozierski. Crazyflie 2.0 quadrotor as a platform for research and education in robotics and control engineering. In 2017 22nd International Conference on Methods and Models in Automation and Robotics (MMAR), pages 37-42. IEEE, 2017. 3, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 480, + 288, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 480, + 288, + 524 + ], + "spans": [ + { + "bbox": [ + 53, + 480, + 288, + 524 + ], + "type": "text", + "content": "[9] Antoine Guedon, Pascal Monasse, and Vincent Lepetit. SCONE: Surface coverage optimization in unknown environments by volumetric integration. In Advances in Neural Information Processing Systems, 2022. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 525, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 568 + ], + "type": "text", + "content": "[10] Junfu Guo, Changhao Li, Xi Xia, Ruizhen Hu, and Ligang Liu. Asynchronous collaborative autoscanning with mode switching for multi-robot scene reconstruction. ACM Transactions on Graphics (TOG), 41(6):1-13, 2022. 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 569, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 288, + 635 + ], + "type": "text", + "content": "[11] Guillaume Hardouin, Julien Moras, Fabio Morbidi, Julien Marzat, and El Mustapha Mouaddib. Next-Best-View planning for surface reconstruction of large-scale 3d environments with multiple uavs. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1567–1574. IEEE, 2020. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "type": "text", + "content": "[12] Stefan Isler, Reza Sabzevari, Jeffrey Delmerico, and Davide Scaramuzza. An information gain formulation for active volumetric 3d reconstruction. In 2016 IEEE International Conference on Robotics and Automation (ICRA), pages 3477-3484. IEEE, 2016. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 690, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 690, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 690, + 288, + 713 + ], + "type": "text", + "content": "[13] Michael Laskin, Aravind Srinivas, and Pieter Abbeel. Curl: Contrastive unsupervised representations for reinforcement" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "learning. In International Conference on Machine Learning, pages 5639-5650. PMLR, 2020. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 95, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 149 + ], + "type": "text", + "content": "[14] Soomin Lee, Le Chen, Jiahao Wang, Alexander Liniger, Suryansh Kumar, and Fisher Yu. Uncertainty guided policy for active robotic 3d reconstruction using neural radiance fields. IEEE Robotics and Automation Letters, 7(4):12070-12077, 2022. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 150, + 545, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 150, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 308, + 150, + 545, + 204 + ], + "type": "text", + "content": "[15] Quanyi Li, Zhenghao Peng, Lan Feng, Chenda Duan, Wenjie Mo, Bolei Zhou, et al. ScenarioNet: Open-source platform for large-scale traffic scenario simulation and modeling. In Advances in Neural Information Processing Systems, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 205, + 545, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 205, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 308, + 205, + 545, + 260 + ], + "type": "text", + "content": "[16] Quanyi Li, Zhenghao Peng, Lan Feng, Qihang Zhang, Zhenghai Xue, and Bolei Zhou. MetaDrive: Composing diverse driving scenarios for generalizable reinforcement learning. IEEE transactions on pattern analysis and machine intelligence, 45(3):3461-3475, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 261, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 316 + ], + "type": "text", + "content": "[17] Yixuan Li, Lihan Jiang, Linning Xu, Yuanbo Xiangli, Zhenzhi Wang, Dahua Lin, and Bo Dai. Matrixcity: A large-scale city dataset for city-scale neural rendering and beyond. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3205-3215, 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 316, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 360 + ], + "type": "text", + "content": "[18] Yilin Liu, Liqiang Lin, Yue Hu, Ke Xie, Chi-Wing Fu, Hao Zhang, and Hui Huang. Learning reconstructability for drone aerial path planning. ACM Transactions on Graphics (TOG), 41(6):1-17, 2022. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 361, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 361, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 361, + 545, + 415 + ], + "type": "text", + "content": "[19] Viktor Makoviychuk, Lukasz Wawrzyniak, Yunrong Guo, Michelle Lu, Kier Storey, Miles Macklin, David Hoeller, Nikita Rudin, Arthur Allshire, Ankur Handa, and Gavriel State. Isaac Gym: High performancegpu-based physics simulation for robot learning, 2021. 2, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 416, + 545, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 416, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 308, + 416, + 545, + 481 + ], + "type": "text", + "content": "[20] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 405-421. Springer, 2020. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 482, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 545, + 525 + ], + "type": "text", + "content": "[21] Joseph Ortiz, Alexander Clegg, Jing Dong, Edgar Sucar, David Novotny, Michael Zollhoefer, and Mustafa Mukadam. iSDF: Real-time neural signed distance fields for robot perception. Robotics: Science and Systems, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 526, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 526, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 308, + 526, + 545, + 581 + ], + "type": "text", + "content": "[22] Xuran Pan, Zihang Lai, Shiji Song, and Gao Huang. ActiveNeRF: Learning where to see with uncertainty estimation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIII, pages 230-246. Springer, 2022. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 582, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 545, + 645 + ], + "type": "text", + "content": "[23] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "type": "text", + "content": "[24] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. PyTorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16444" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[25] Daryl Peralta, Joel Casimiro, Aldrin Michael Nilles, Justine Aletta Aguilar, Rowel Atienza, and Rhandley Cajote. Houses3K dataset. https://github.com/darylperalta/Houses3K, 2020.2.5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 117, + 288, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 117, + 288, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 117, + 288, + 183 + ], + "type": "text", + "content": "[26] Daryl Peralta, Joel Casimiro, Aldrin Michael Nilles, Justine Aletta Aguilar, Rowel Atienza, and Rhandley Cajote. Next-best view policy for 3d reconstruction. In Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16, pages 558-573. Springer, 2020. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 184, + 288, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 288, + 237 + ], + "type": "text", + "content": "[27] Antonin Raffin, Ashley Hill, Adam Gleave, Anssi Kanervisto, Maximilian Ernestus, and Noah Dormann. Stable-baselines3: Reliable reinforcement learning implementations. The Journal of Machine Learning Research, 22(1): 12348-12355, 2021. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 238, + 288, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 238, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 238, + 288, + 293 + ], + "type": "text", + "content": "[28] Yunlong Ran, Jing Zeng, Shibo He, Jiming Chen, Lincheng Li, Yingfeng Chen, Gimhee Lee, and Qi Ye. NeurAR: Neural uncertainty for autonomous 3d reconstruction with implicit neural representations. IEEE Robotics and Automation Letters, 2023. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 293, + 288, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 293, + 288, + 337 + ], + "spans": [ + { + "bbox": [ + 48, + 293, + 288, + 337 + ], + "type": "text", + "content": "[29] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv:2007.08501, 2020.6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 338, + 288, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 288, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 288, + 370 + ], + "type": "text", + "content": "[30] Antoni Rosinol, John J Leonard, and Luca Carlone. NeRF-SLAM: Real-time dense monocular slam with neural radiance fields. arXiv preprint arXiv:2210.13641, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 371, + 288, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 288, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 288, + 415 + ], + "type": "text", + "content": "[31] Nikita Rudin, David Hoeller, Philipp Reist, and Marco Hutter. Learning to walk in minutes using massively parallel deep reinforcement learning. In 5th Annual Conference on Robot Learning, 2021. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 415, + 288, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 288, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 288, + 448 + ], + "type": "text", + "content": "[32] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 449, + 288, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 288, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 288, + 480 + ], + "type": "text", + "content": "[33] James A Sethian. A fast marching level set method for monotonically advancing fronts. proceedings of the National Academy of Sciences, 93(4):1591-1595, 1996. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 481, + 288, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 288, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 288, + 525 + ], + "type": "text", + "content": "[34] Edward J Smith, Michal Drozdal, Derek Nowrouzezahrai, David Meger, and Adriana Romero-Soriano. Uncertainty-driven active vision for implicit scene reconstruction. arXiv preprint arXiv:2210.00978, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 525, + 288, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 288, + 580 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 288, + 580 + ], + "type": "text", + "content": "[35] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 5, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 581, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 288, + 624 + ], + "type": "text", + "content": "[36] Edgar Sucar, Shikun Liu, Joseph Ortiz, and Andrew J Davison. iMAP: Implicit mapping and positioning in real-time. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6229-6238, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 625, + 288, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 288, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 288, + 679 + ], + "type": "text", + "content": "[37] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "type": "text", + "content": "[38] Jiaming Sun, Yiming Xie, Linghao Chen, Xiaowei Zhou, and Hujun Bao. NeuralRecon: Real-time coherent 3d reconstruction from monocular video. In Proceedings of the IEEE/CVF" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 687 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "Conference on Computer Vision and Pattern Recognition, pages 15598-15607, 2021. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 95, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 545, + 140 + ], + "type": "text", + "content": "[39] Jiaming Sun, Xi Chen, Qianqian Wang, Zhengqi Li, Hadar Averbuch-Elor, Xiaowei Zhou, and Noah Snavely. Neural 3D reconstruction in the wild. In SIGGRAPH Conference Proceedings, 2022. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 141, + 547, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 547, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 547, + 162 + ], + "type": "text", + "content": "[40] Sebastian Thrun. Probabilistic robotics. Communications of the ACM, 45(3):52-57, 2002. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 163, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 228 + ], + "type": "text", + "content": "[41] Josh Tobin, Rachel Fong, Alex Ray, Jonas Schneider, Wojciech Zaremba, and Pieter Abbeel. Domain randomization for transferring deep neural networks from simulation to the real world. In 2017 IEEE/RSJ international conference on intelligent robots and systems (IROS), pages 23-30. IEEE, 2017. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 230, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 545, + 296 + ], + "type": "text", + "content": "[42] Tong Wu, Jiarui Zhang, Xiao Fu, Yuxin Wang, Jiawei Ren, Liang Pan, Wayne Wu, Lei Yang, Jiaqi Wang, Chen Qian, et al. Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 803-814, 2023. 2, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 297, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 545, + 352 + ], + "type": "text", + "content": "[43] Yuanbo Xiangli, Linning Xu, Xingang Pan, Nanxuan Zhao, Anyi Rao, Christian Theobalt, Bo Dai, and Dahua Lin. BungeeNeRF: Progressive neural radiance field for extreme multi-scale scene rendering. In The European Conference on Computer Vision (ECCV), 2022. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 353, + 547, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 353, + 547, + 407 + ], + "spans": [ + { + "bbox": [ + 307, + 353, + 547, + 407 + ], + "type": "text", + "content": "[44] Kai Ye, Siyan Dong, Qingnan Fan, He Wang, Li Yi, Fei Xia, Jue Wang, and Baoquan Chen. Multi-robot active mapping via neural bipartite graph matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14839–14848, 2022. 3, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 409, + 545, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 409, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 307, + 409, + 545, + 441 + ], + "type": "text", + "content": "[45] Cem Yuksel. Sample elimination for generating poisson disk sample sets. Computer Graphics Forum, 34(2):25-32, 2015. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 442, + 545, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 442, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 307, + 442, + 545, + 485 + ], + "type": "text", + "content": "[46] Huangying Zhan, Jiyang Zheng, Yi Xu, Ian Reid, and Hamid Rezatofighi. ActiveRMAP: Radiance field for active mapping and planning. arXiv preprint arXiv:2211.12656, 2022. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 487, + 545, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 487, + 545, + 531 + ], + "spans": [ + { + "bbox": [ + 307, + 487, + 545, + 531 + ], + "type": "text", + "content": "[47] Han Zhang, Yuong Yao, Ke Xie, Chi-Wing Fu, Hao Zhang, and Hui Huang. Continuous aerial path planning for 3d urban scene reconstruction. ACM Trans. Graph., 40(6):225-1, 2021. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 532, + 545, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 532, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 307, + 532, + 545, + 586 + ], + "type": "text", + "content": "[48] Xiaoshuai Zhang, Sai Bi, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. NeRFusion: Fusing radiance fields for large-scale scene reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5449-5458, 2022. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 587, + 545, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 587, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 307, + 587, + 545, + 620 + ], + "type": "text", + "content": "[49] Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3D: A modern library for 3D data processing. arXiv:1801.09847, 2018.6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 622, + 545, + 687 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 622, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 307, + 622, + 545, + 687 + ], + "type": "text", + "content": "[50] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. Nice-SLAM: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12786-12796, 2022. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16445" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/GenTron_ Diffusion Transformers for Image and Video Generation/2bb6a697-7387-4091-9b21-cb7c548eb341_content_list.json b/2024/GenTron_ Diffusion Transformers for Image and Video Generation/2bb6a697-7387-4091-9b21-cb7c548eb341_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..320faab98b5a3589a556ebce724a44651711bc39 --- /dev/null +++ b/2024/GenTron_ Diffusion Transformers for Image and Video Generation/2bb6a697-7387-4091-9b21-cb7c548eb341_content_list.json @@ -0,0 +1,1557 @@ +[ + { + "type": "text", + "text": "GenTron: Diffusion Transformers for Image and Video Generation", + "text_level": 1, + "bbox": [ + 145, + 130, + 821, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shoufa Chen $^{1,2*}$ Mengmeng Xu $^{2*}$ Jiawei Ren $^{2}$ Yuren Cong $^{2}$ Sen He $^{2}$ Yanping Xie $^{2}$ Animesh Sinha $^{2}$ Ping Luo $^{1}$ Tao Xiang $^{2}$ Juan-Manuel Perez-Rua $^{2}$ The University of Hong Kong ${}^{2}$ Meta", + "bbox": [ + 133, + 179, + 833, + 233 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e449d602918cf754384afa929d4d33a7802872219b2287935e6d1f3ca0b15005.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 104, + 250, + 359, + 445 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/39dacd659364b26924a31c15f7b7fc3868d0030c53794ccbee361706bf98cc51.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 250, + 612, + 446 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/73f12c5e67ca1fa52f149fb37427e7d7bea9fc8b528384069d13341f24df0b21.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 612, + 250, + 864, + 445 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b5ef7783ec1b8f511faba7baa8c3cf25d763f8b7b17ca461816b78352feeb43f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 102, + 446, + 485, + 532 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a98345756a35f127066decf21da3bccba3d4704694f70d03c926322979c23747.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 485, + 446, + 864, + 531 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a5baec7215de49c1a96add6f7dd16cf2926df3cdb676c96a97e82ca87ad754b4.jpg", + "image_caption": [ + "Figure 1. GenTron: Transformer based diffusion model for high-quality text-to-image/video generation." + ], + "image_footnote": [], + "bbox": [ + 102, + 532, + 485, + 619 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a74bbb5fe69b6e7559bc86c5951149c6d0403ac223ce685f030759ab27622002.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 485, + 532, + 864, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 669, + 310, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this study, we explore Transformer-based diffusion models for image and video generation. Despite the dominance of Transformer architectures in various fields due to their flexibility and scalability, the visual generative domain primarily utilizes CNN-based U-Net architectures, particularly in diffusion-based models. We introduce GenTron, a family of Generative models employing Transformer-based diffusion, to address this gap. Our initial step was to adapt Diffusion Transformers (DiTs) from class to text conditioning, a process involving thorough empirical exploration of the conditioning mechanism. We then scale GenTron from approximately 900M to over 3B parameters, observing improvements in visual quality. Furthermore, we extend", + "bbox": [ + 73, + 704, + 470, + 900 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "GenTron to text-to-video generation, incorporating novel motion-free guidance to enhance video quality. In human evaluations against SDXL, GenTron achieves a $51.1\\%$ win rate in visual quality (with a $19.8\\%$ draw rate), and a $42.3\\%$ win rate in text alignment (with a $42.9\\%$ draw rate). GenTron notably performs well in T2I-CompBench, highlighting its compositional generation ability. We hope GenTron could provide meaningful insights and serve as a valuable reference for future research. Please refer to the website1 and the arXiv version for the most up-to-date results: https://arxiv.org/abs/2312.04557.", + "bbox": [ + 496, + 670, + 890, + 835 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{1}$ https://www.shoufachen.com/gentron網站/ \n*Equal contribution.", + "bbox": [ + 522, + 875, + 869, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6441", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 78, + 89, + 207, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion models have recently shown remarkable progress in content creation, impacting areas such as image generation [27, 53, 55], video generation [5, 7, 28, 57], and editing [14, 33]. Among these notable developments, the CNN-based U-Net architecture has emerged as the predominant backbone design, a choice that stands in contrast to the prevailing trend in natural language processing [8, 19, 59] and computer visual perception [20, 37, 66] domains, where attention-based transformer architectures [60] have empowered a renaissance and become increasingly dominant. To provide a comprehensive understanding of Transformers in diffusion generation and to bridge the gap in architectural choices between visual generation and the other two domains — visual perception and NLP — a thorough investigation of visual generation using Transformers is of substantial scientific value.", + "bbox": [ + 76, + 114, + 468, + 354 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We focus on diffusion models with Transformers in this work. Specifically, our starting point is the foundational work known as DiT [44], which introduced a class-conditioned latent diffusion model that employs a Transformer to replace the traditionally used U-Net architecture. We overcome the limitation of the original DiT model, which is constrained to handling only a restricted number (e.g., 1000) of predefined classes, by utilizing language embeddings derived from open-world, free-form text captions instead of predefined one-hot class embeddings. Along the way, we comprehensively investigate conditioning strategies, including (1) conditioning architectures: adaptive layer norm (adaLN) [45] vs. cross-attention [60]; and (2) text encoding methods: a generic large language model [13] vs. the language tower of multimodal models [50], or the combination of both of them. We additionally carry out comparative experiments and offer detailed empirical analyses to evaluate the effectiveness of these conditioning strategies.", + "bbox": [ + 76, + 356, + 470, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Next, we explore the scaling-up properties of GenTron. The Transformer architectures have been demonstrated to possess significant scalability in both visual perception [11, 17, 51, 66] and language [8, 19, 48, 49, 59] tasks. For example, the largest dense language model has 540B parameters [12], and the largest vision model has 22B [17] parameters. In contrast, the largest diffusion transformer, DiT-XL [44], only has about 675M parameters, trailed far behind both the Transformers utilized in other domains (e.g., NLP) and recent diffusion arts with convolutional U-Net architectures [15, 46]. To compensate for this considerable lagging, we scale up GenTron in two dimensions, the number of transformer blocks and hidden dimension size, following the scaling strategy in [66]. As a result, our largest model, GenTron-G/2, has more than 3B parameters and achieves significant visual quality improvement compared with the smaller one.", + "bbox": [ + 76, + 643, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, we have advanced GenTron from a T2I to a T2V model by inserting a temporal self-attention layer into each transformer block, making the first attempt to use transformers as the exclusive building block for video diffusion models. We also discuss existing challenges in video generation and introduce our solution, the motion-free guidance (MFG). Specifically, This approach involves intermittently disabling motion modeling during training by setting the temporal self-attention mask to an identity matrix. Besides, MFG seamlessly integrates with the joint image-video strategy [9, 16, 30, 62], where images are used as training samples whenever motion is deactivated. Our experiments indicate that this approach clearly improves the visual quality of generated videos.", + "bbox": [ + 496, + 90, + 890, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In human evaluations, GenTron outperforms SDXL, achieving a $51.1\\%$ win rate in visual quality (with a $19.8\\%$ draw rate), and a $42.3\\%$ win rate in text alignment (with a $42.9\\%$ draw rate). Furthermore, when compared to previous studies, particularly as benchmarked against T2I-CompBench [31]-a comprehensive framework for evaluating open-world compositional T2I generation-GenTron demonstrates superior performance across various criteria. These include attribute binding, object relationships, and handling of complex compositions.", + "bbox": [ + 496, + 301, + 890, + 453 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows: (1) We have conducted a thorough and systematic investigation of transformer-based T2I generation with diffusion models. This study encompasses various conditioning choices and aspects of model scaling. (2) In a pioneering effort, we explore a purely transformer-based diffusion model for T2V generation. We introduce motion-free guidance, an innovative technique that efficiently fine-tunes T2I generation models for producing high-quality videos. (3) Experimental results indicate a clear preference for GenTron over SDXL in human evaluations. Furthermore, GenTron demonstrates superior performance compared to existing methods in the T2I-CompBench evaluations.", + "bbox": [ + 496, + 454, + 890, + 650 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 664, + 640, + 679 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion models for T2I and T2V generation. Diffusion models [27, 42] are a type of generative model that creates data samples from random noise. Later, latent diffusion models [44, 46, 53] are proposed for efficient T2I generation. These designs usually have $I$ a pre-trained Variational Autoencoder [35] that maps images to a compact latent space, 2) a conditioner modeled by cross-attention [28, 53] to process text as conditions with a strength control [26], and 3) a backbone network, U-Net [54] in particular, to process image features. The success of diffusion on T2I generation tasks underscores the promising potential for text-to-video (T2V) generation [34, 41, 47, 57]. VDM [30] and Imagen Video [28] extend the image diffusion architecture on the temporal dimension with promising initial results. To", + "bbox": [ + 496, + 688, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6442", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "avoid excessive computing demands, video latent diffusion models [5, 9, 25, 40, 61, 69] implement the video diffusion process in a low-dimensional latent space.", + "bbox": [ + 75, + 90, + 468, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Transformer-based Diffusion. Recently, Transformer-based Diffusion models have attracted increasing research interest. Among these, U-ViT [3] treats all inputs as tokens by integrating transformer blocks with a U-net architecture. In contrast, DiT [44] employs a simpler, non-hierarchical transformer structure. MDT [22] and MaskDiT [68] enhance DiT's training efficiency by incorporating the mask strategy [24]. Dolfin [63] is a transformer-based model for layout generation. Concurrently to this work, PixArt- $\\alpha$ [10] demonstrates promising outcomes in Transformer-based T2I diffusion. It's trained using a three-stage decomposition process with high-quality data. Our work diverges from PixArt- $\\alpha$ in key aspects. Firstly, while PixArt- $\\alpha$ emphasizes training efficiency, our focus is on the design choice of conditioning strategy and scalability in T2I Transformer diffusion models. Secondly, we extend our exploration beyond image generation to video diffusion. We propose an innovative approach in video domain, which is not covered by PixArt- $\\alpha$ .", + "bbox": [ + 75, + 157, + 472, + 448 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 460, + 168, + 477 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first introduce the preliminaries in Section 3.1, and then present the details of GenTron for text-to-image generation in Section 3.2, which includes text encoder models, embedding integration methods, and scaling up strategy of GenTron. Lastly, in Section 3.3, we extend GenTron's application to video generation, building on top of the T2I foundations laid in previous sections.", + "bbox": [ + 75, + 487, + 468, + 594 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 604, + 218, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models. Diffusion models [27] have emerged as a family of generative models that generate data by performing a series of transformations on random noise. They are characterized by a forward and a backward process. Given an instance from the data distribution $x_0 \\sim p(x_0)$ , random Gaussian noise is iteratively added to the instance in the forward noisig process to create a Markov Chain of random latent variable $x_1, x_2, \\ldots, x_T$ following:", + "bbox": [ + 75, + 628, + 468, + 750 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(x _ {t} \\mid x _ {t - 1}\\right) = \\mathcal {N} \\left(x _ {t}; \\sqrt {1 - \\beta_ {i}} x _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 763, + 468, + 782 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\beta_{1},\\ldots ,\\beta_{T}$ are hyperparameters corresponding to the noise schedule. After a large enough number of diffusion steps, $x_{T}$ can be viewed as a standard Gaussian noise. A denoising network $\\epsilon_{\\theta}$ is further trained to learn the backward process, i.e., how to remove the noise from a noisy input [27]. For inference, an instance can be sampled starting from a random Gaussian noise $x_{T}\\sim \\mathcal{N}(0;\\mathbf{I})$ and denoised", + "bbox": [ + 75, + 794, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6f9293ac2ca6ee4a494efe39f6e5459242bbd1cc1d489750c3d4f6e0a26422f0.jpg", + "image_caption": [ + "(a)adaLN-Zero", + "Figure 2. Text embedding integration architecture. We directly adapt adaLN from DiT [44], substituting the one-hot class embedding with text embedding. For cross attention, different from the approach in [44], we maintain the use of adaLN to model the combination of time embedding and the aggregated text embedding." + ], + "image_footnote": [], + "bbox": [ + 522, + 88, + 684, + 358 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/bfda24d2d5047514ccd5cfd22f5ccb8b82a8b15d1c13e105f1e2b97f47dc2ace.jpg", + "image_caption": [ + "(b) Cross attention" + ], + "image_footnote": [], + "bbox": [ + 705, + 88, + 867, + 357 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "step-by-step following the Markov Chain, i.e., by sequentially sampling $x_{t-1}$ to $x_0$ with $p_\\theta(x_{t-1}|x_t)$ :", + "bbox": [ + 498, + 469, + 888, + 501 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(x _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta} (x _ {t}, t)\\right) + \\sigma_ {t} \\mathbf {z}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 511, + 890, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\bar{\\alpha}_t = \\prod_{s=1}^t \\alpha_s$ , $\\alpha_t = 1 - \\beta_t$ and $\\sigma_t$ is the noise scale. In practical application, the diffusion sampling process can be further accelerated using different sampling techniques [39, 58].", + "bbox": [ + 498, + 556, + 890, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Latent diffusion model architectures. Latent diffusion models (LDMs) [53] reduce the high computational cost by conducting the diffusion process in the latent space. First, a pre-trained autoencoder [21, 35] is utilized to compress the raw image from pixel to latent space, then the diffusion models, which are commonly implemented with a U-Net [54] backbone, work on the latent space. Peebles et al. proposed DiT [44] to leverage the transformer architecture as an alternative to the traditional U-Net backbone for class-conditioned image generation, adopting the adaptive layernorm (adaLN [45]) for class conditioning mechanism, as shown in Figure 2a.", + "bbox": [ + 496, + 638, + 890, + 820 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Text-to-Image GenTron", + "text_level": 1, + "bbox": [ + 500, + 830, + 718, + 847 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our GenTron is built upon the DiT-XL/2 [44], which converts the latent of shape $32 \\times 32 \\times 4$ to a sequence of non-overlapping tokens with a $2 \\times 2$ patchify layer [20]. Then,", + "bbox": [ + 498, + 854, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6443", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "these tokens are sent into a series of transformer blocks. Finally, a standard linear decoder is applied to convert these image tokens into latent space.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While DiT has shown that transformer-based models yield promising results in class-conditioned scenarios, it did not explore the realm of T2I generation. This field poses a considerable challenge, given its less constrained conditioning format. Moreover, even the largest DiT model, DiT-XL/2, with its 675 million parameters, is significantly overshadowed by current U-Nets [15, 46], which boast over 3 billion parameters. To address these limitations, our research conducts a thorough investigation of transformer-based T2I diffusion models, focusing specifically on text conditioning approaches and assessing the scalability of the transformer architecture by expanding GenTron to more than 3 billion parameters.", + "bbox": [ + 75, + 137, + 470, + 333 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1 From Class to Text Condition", + "text_level": 1, + "bbox": [ + 76, + 353, + 334, + 367 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "T2I diffusion models rely on textual inputs to steer the process of image generation. The mechanism of text conditioning involves two critical components: firstly, the selection of a text encoder, which is responsible for converting raw text into text embeddings, and secondly, the method of integrating these embeddings into the diffusion process. For a complete understanding, we have included in the appendix a detailed presentation of the decisions made in existing works concerning these two components.", + "bbox": [ + 75, + 377, + 468, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Text encoder model. Current advancements in T2I diffusion techniques employ a variety of language models, each with its unique strengths and limitations. To thoroughly assess which model best complements transformer-based diffusion methods, we have integrated several models into GenTron. This includes the text towers from multimodal models, CLIP [50], as well as a pure large language model, Flan-T5 [13]. Our approach explores the effectiveness of these language models by integrating each model independently with GenTron to evaluate their individual performance and combinations of them to assess the potential properties they may offer when used together.", + "bbox": [ + 75, + 532, + 468, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "**Embedding integration.** In our study, we focused on two methods of embedding integration: adaptive layernorm and cross-attention. (1) Adaptive layernorm (adaLN). As shown in Figure 2a, this method integrates conditioning embeddings as normalization parameters on the feature channel. Widely used in conditional generative modeling, such as in StyleGAN [32], adaLN serves as the standard approach in DiT [44] for managing class conditions. (2) Cross-attention. As illustrated in Figure 2b, the image feature acts as the query, with textual embedding serving as key and value. This setup allows for direct interaction", + "bbox": [ + 75, + 734, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/1f26b43e6452aa6ecb59324495645ea82e4cec0876a45aca968ef41abe52f1ab.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelDepthWidthMLP Width#Param.
GenTron-XL/22811524608930.0M
GenTron-G/248166466563083.8M
", + "bbox": [ + 501, + 88, + 885, + 132 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. Configuration details of GenTron models.", + "bbox": [ + 537, + 141, + 854, + 155 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "between the image feature and textual embedding through an attention mechanism [60]. Besides, different from the cross-attention discussed in [44], which processes the class embedding and time embedding together by firstly concatenating them, we maintain the use of adaLN in conjunction with the cross-attention to separately model the time embedding. The underlying rationale for this design is our belief that the time embedding, which is consistent across all spatial positions, benefits from the global modulation capabilities of adaLN. Moreover, we also add the pooled text embeddings to the time embedding followings [2, 29, 43, 46].", + "bbox": [ + 496, + 175, + 890, + 340 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2 Scaling Up GenTron", + "text_level": 1, + "bbox": [ + 500, + 359, + 696, + 376 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To explore the impact of substantially scaling up the model size, we have developed an advanced version of GenTron, which we refer to as GenTron-G/2. This model was constructed in accordance with the scaling principles outlined in [66]. We focused on expanding three critical aspects: the number of transformer blocks (depth), the dimensionality of patch embeddings (width), and the hidden dimension of the MLP (MLP-width). The specifications and configurations of the GenTron models are detailed in Table 1. Significantly, the GenTron-G/2 model boasts over 3 billion parameters. To our knowledge, this represents the largest transformer-based diffusion architecture developed to date.", + "bbox": [ + 496, + 383, + 890, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Text-to-Video GenTron", + "text_level": 1, + "bbox": [ + 500, + 575, + 714, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this subsection, we elaborate on the process of adapting GenTron from a T2I framework to a T2V framework. Sec. 3.3.1 will detail the modifications made to the model's architecture, enabling GenTron to process video data. Furthermore, Sec. 3.3.2 will discuss the challenges encountered in the domain of video generation and the innovative solutions we have proposed to address them.", + "bbox": [ + 496, + 599, + 890, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1 GenTron-T2V Architecture", + "text_level": 1, + "bbox": [ + 500, + 724, + 745, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Transformer block with temporal self-attention. It is typically a common practice to train video diffusion models from image diffusion models by adding new temporal modeling modules [5, 23, 30, 69]. These usually consist of 3D convolutional layers and temporal transformer blocks that focus on calculating attention along the temporal dimension. In contrast to the traditional approach [5], which involves adding both temporal convolution layers and temporal transformer blocks to the T2I U-Net, our method integrates only lightweight temporal", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6444", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d533bf56cb27eae85c1f8aa3ff697ebd95e0b758f57f208edc53ad3567e2c78c.jpg", + "image_caption": [ + "Figure 3. GenTron Transformer block with TempSelfAttn and Motion-Free Mask. The temporal self-attention layer is inserted between the cross-attention and the MLPs. The motion-free mask, which is an identity matrix, will be utilized in the TempSelfAttn with a probability of $p_{\\text{motion\\_free}}$ . We omit details like text conditioning, LN here for simplicity, which could be found in Figure 2." + ], + "image_footnote": [], + "bbox": [ + 135, + 90, + 405, + 294 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "self-attention (TempSelfAttn) layers into each transformer block. As depicted in Figure 3, the TempSelfAttn layer is placed right after the cross-attention layer and before the MLP layer. Additionally, we modify the output of the cross-attention layer by reshaping it before it enters the TempSelfAttn layer and then reshape it back to its original format once it has passed through. This process can be formally represented as:", + "bbox": [ + 75, + 411, + 468, + 532 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} = \\text {r e a r r a n g e} (\\mathbf {x}, (\\mathrm {b} \\quad \\mathrm {t}) \\quad \\mathrm {n} \\quad \\mathrm {d} \\rightarrow (\\mathrm {b} \\quad \\mathrm {n}) \\quad \\mathrm {t} \\quad \\mathrm {d}) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 83, + 550, + 467, + 566 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} = \\mathbf {x} + \\operatorname {T e m p S e l f A t t n} (\\mathrm {L N} (\\mathbf {x})) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 568, + 468, + 585 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} = \\text {r e a r r a n g e} (\\mathbf {x}, (\\mathrm {b} \\quad \\mathrm {n}) \\quad \\mathrm {t} \\quad \\mathrm {d} \\rightarrow (\\mathrm {b} \\quad \\mathrm {t}) \\quad \\mathrm {n} \\quad \\mathrm {d}) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 588, + 468, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $b, t, n, d$ represent the batch size, number of frames, number of patches per frame, and channel dimension, respectively. rearrange is a notation from [52]. We discovered that a simple TempSelfAttn layer suffices to capture motion, a finding that aligns with observations in a recent study [62]. In addition, only using TempSelfAttn makes it convenient to turn on and turn off the temporal modeling, which would be discussed in Sec. 3.3.2.", + "bbox": [ + 75, + 622, + 468, + 743 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Initialization. We use the pre-trained T2I model as a basis for initializing the shared layers between T2I and T2V models. In addition, for the newly added TempSelfAttn layers, we initialize the weights and biases of the output project layers to zero. This ensures that at the beginning of the T2V fine-tuning stage, these layers produce a zero output, effectively functioning as an identity mapping in conjunction with the shortcut connection.", + "bbox": [ + 75, + 779, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.2 Motion-Free Guidance", + "text_level": 1, + "bbox": [ + 500, + 90, + 714, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Challenges encountered. We observed a notable phenomenon in the current T2V diffusion models [5, 30] where the per-frame visual quality significantly lags behind that of T2I models [15, 46, 53, 64]. Furthermore, our analysis revealed a remarkable degradation in visual quality in the T2V models post-fine-tuning, especially when compared to their original T2I counterparts. We note that these problems generally exist in current T2V diffusion models, not limited to our transformer-based T2V.", + "bbox": [ + 496, + 116, + 893, + 252 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Problem analysis and insights. We presume that the observed lag in the visual quality of T2V primarily stems from two factors: the nature of video data and the fine-tuning approach. Firstly, publicly available video datasets often fall short in both quality and quantity compared to image datasets. For instance, [56] has more than 2B English image-text pairs, whereas the current widely used video dataset, WebVid-10M [1] contains only 10.7M video-text pairs. Additionally, many video frames are compromised by motion blur and watermarks, further reducing their visual quality. This limited availability hampers the development of robust and versatile video diffusion models. Secondly, the focus on optimizing temporal aspects during video fine-tuning can inadvertently compromise the spatial visual quality, resulting in a decline in the overall quality of the generated videos.", + "bbox": [ + 496, + 277, + 892, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Solution I: joint image-video training. From the data aspect, we adopt the joint image-video training strategy [9, 16, 30, 62] to mitigate the video data shortages. Furthermore, joint training helps to alleviate the problem of domain discrepancy between video and image datasets by integrating both data types for training.", + "bbox": [ + 496, + 542, + 890, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Solution II: motion-free guidance. We treat the temporal motion within a video clip as a special conditioning signal, which can be analogized to the textual conditioning in T2I/T2V diffusion models. Based on this analogy, we propose a novel approach, motion-free guidance (MFG), inspired by classifier-free guidance [6, 26], to modulate the weight of motion information in the generated video.", + "bbox": [ + 496, + 657, + 890, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In a particular training iteration, our approach mirrors the concept used in classifier-free guidance, where conditioned text is replaced with an empty string. The difference is that we employ an identity matrix to nullify the temporal attention with a probability of $p_{\\text{motion\\_free}}$ . This identity matrix, which is depicted in Figure 3 (Motion-Free Mask), is structured such that its diagonal is populated with ones, while all other positions are zeroes. This configuration confines the temporal self-attention to work within a single", + "bbox": [ + 496, + 763, + 892, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6445", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/57ba0ea001525f1cf1173ffe98c6fb672bb8935525ec7f17680f50e802b09b41.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ConditioningScaleAttribute BindingObject RelationshipComplexMean
Text EncoderTypeIntegrationColorShapeTextureSpatialNon-spatial
CLIP-L [50]MMadaLN-zeroXL/236.9442.0650.739.4130.3836.4134.32
CLIP-L [50]MMcross-attnXL/273.9151.8168.7619.2631.8041.5247.84
T5-XXL [13]LLMcross-attnXL/274.9055.4070.0520.5231.6841.0148.93
CLIP-T5XXLMM + LLMcross-attnXL/275.6555.7469.4820.6731.7941.4449.13
CLIP-T5XXLMM + LLMcross-attnG/276.7457.0071.5020.9832.0241.6749.99
", + "bbox": [ + 125, + 88, + 841, + 207 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "frame. Furthermore, as introduced in Sec. 3.3.1, temporal self-attention is the sole operator for temporal modeling. Thus, using a motion-free attention mask suffices to disable temporal modeling in the video diffusion process.", + "bbox": [ + 75, + 273, + 468, + 334 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During inference, we have text and motion conditionings. Inspired by [6], we can modify the score estimate as:", + "bbox": [ + 75, + 335, + 468, + 364 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\tilde {\\epsilon} _ {\\theta} = \\epsilon_ {\\theta} (x _ {t}, \\varnothing , \\varnothing) \\\\ + \\lambda_ {T} \\cdot \\left(\\epsilon_ {\\theta} \\left(x _ {t}, c _ {T}, c _ {M}\\right) - \\epsilon_ {\\theta} \\left(x _ {t}, \\varnothing , c _ {M}\\right)\\right) \\\\ + \\lambda_ {M} \\cdot \\left(\\epsilon_ {\\theta} \\left(x _ {t}, \\varnothing , c _ {M}\\right) - \\epsilon_ {\\theta} \\left(x _ {t}, \\varnothing , \\varnothing\\right)\\right) \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 393, + 467, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $c_{T}$ and $c_{M}$ represent the text conditioning and motion conditioning. $\\lambda_{T}$ and $\\lambda_{M}$ are the guidance scale of standard text and that of motion, controlling how strongly the generated samples correspond with the text condition and the motion strength, respectively. We empirically found that fixing $\\lambda_{T} = 7.5$ and adjusting $\\lambda_{M} \\in [1.0, 1.3]$ for each example tend to achieve the best result. This finding is similar to [6], although our study utilizes a narrower range for $\\lambda_{M}$ .", + "bbox": [ + 75, + 458, + 468, + 579 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Putting solutions together. We can integrate solution I and II together in the following way: when the motion is omitted at a training step, we load an image-text pair and repeat the image $T - 1$ times to create a pseudo video. Conversely, if motion is included, we instead load a video clip and extract it into $T$ frames.", + "bbox": [ + 75, + 597, + 468, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 700, + 207, + 718 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 726, + 292, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training scheme For all GenTron model variations, we employ the AdamW [38] optimizer, maintaining a constant learning rate of $1 \\times 10^{-4}$ . We train our T2I GenTron models in a multi-stage procedure [46, 53] with an internal dataset, including a low-resolution $(256 \\times 256)$ training with a batch size of 2048 and 500K optimization steps, as well as high-resolution $(512 \\times 512)$ with a batch size of 784 and 300K steps. For the GenTron-G/2 model, we further integrate Fully Sharded Data Parallel (FSDP) [67] and activation checkpointing (AC), strategies specifically adopted to", + "bbox": [ + 75, + 750, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/150d925731fbe700347e9b39043c2a0ae10a9dcdc97b1f1449330355d75cb7ca.jpg", + "table_caption": [ + "Table 2. Conditioning and model scale in GenTron. We compare GenTron model variants with different design choices on T2I-CompBench [31]. The text encoders are from the language tower of the multi-modal (MM) model, the large language model (LLM), or a combination of them. The GenTron-G/2 with CLIP-T5XXL performs best. Detailed discussions can be found in Sec. 4.2." + ], + "table_footnote": [], + "table_body": "
ModelAttribute BindingObj. RelationComp.Mean
ColorShapeTextureSpat.Non-spat.
LDM v1.437.6535.7641.5612.4630.7930.8031.50
LDM v250.6542.2149.2213.4230.9633.8636.72
Composable v240.6332.9936.458.0029.8028.9829.47
Structured v249.9042.1849.0013.8631.1133.5536.60
Attn-Exact v264.0045.1759.6314.5531.0934.0141.41
GORS66.0347.8562.8718.1531.9333.2843.35
DALL-E 257.5054.6463.7412.8330.4336.9642.68
LDM XL63.6954.0856.3720.3231.1040.9144.41
PixArt-α68.8655.8270.4420.8231.7941.1748.15
GenTron76.7457.0071.5020.9832.0241.6749.99
", + "bbox": [ + 501, + 271, + 888, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Comparison of alignment evaluation on T2I-CompBench [31]. Results show our advanced model, GenTron-CLIPT5XXL-G/2, achieves superior performance across multiple compositional metrics compared to previous methods.", + "bbox": [ + 498, + 435, + 890, + 491 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "optimize GPU memory usage. In our video experiments, we train videos on a video dataset that comprises approximately 34M videos. To optimize storage usage and enhance data loading efficiency, the videos are pre-processed to a resolution with a short side of 512 pixels and a frame rate of 24 FPS. We process batches of 128 video clips. Each clip comprises 8 frames, captured at a sampling rate of 4 FPS.", + "bbox": [ + 496, + 513, + 890, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics. We mainly adopt the recent T2I-CompBench [31] to compare GenTron model variants, following [4, 10]. Specifically, we compare the attribute binding aspects, which include color, shape, and texture. We also compare the spatial and non-spatial object relationships. Moreover, user studies are conducted to compare visual quality and text alignment.", + "bbox": [ + 496, + 648, + 890, + 755 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Main Results of GenTron-T2I", + "text_level": 1, + "bbox": [ + 498, + 768, + 764, + 784 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this subsection, we discuss our experimental results, focusing on how various conditioning factors and model sizes impact GenTron's performance. Table 2 presents the quantitative findings of our study. Additionally, we offer a comparative visualization, illustrating the effects of each conditioning factor we explored. A comparison to prior art is also provided.", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6446", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a3324278740a7af051c211a79feffb5c48d2b940628ce4edc5f70c590d7e53c8.jpg", + "image_caption": [ + "(a)adaLN-Zero" + ], + "image_footnote": [], + "bbox": [ + 86, + 88, + 271, + 229 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4559b4fdbd5e73a5d760c13e1fb6f6c0241445b4dd33bcd5f42faa738400c22d.jpg", + "image_caption": [ + "(b) Cross attention" + ], + "image_footnote": [], + "bbox": [ + 272, + 88, + 455, + 229 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Cross attention vs. adaLN-Zero. Recent findings [44] conclude that the adaLN design yields superior results in terms of the FID, outperforming both cross-attention and in-context conditioning in efficiency for class-based scenarios. However, our observations reveal a limitation of adaLN in handling free-form text conditioning. This shortcoming is evident in Figure 4, where adaLN's attempt to generate a panda image falls short, with cross-attention demonstrating a clear advantage. This is further verified quantitatively in the first two rows of Table 2, where cross-attention uniformly excels over adaLN in all evaluated metrics.", + "bbox": [ + 75, + 308, + 467, + 473 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This outcome is reasonable considering the nature of class conditioning, which typically involves a limited set of fixed signals (e.g., the 1000 one-hot class embeddings for ImageNet [18]). In such contexts, the adaLN approach, operating at a spatially global level, adjusts the image features uniformly across all positions through the normalization layer, making it adequate for the static signals. In contrast, cross-attention treats spatial positions with more granularity. It differentiates between various spatial locations by dynamically modulating image features based on the cross-attention map between the text embedding and the image features. This spatial-sensitive processing is essential for free-from text conditioning, where the conditioning signals are infinitely diverse and demand detailed representation in line with the specific content of textual descriptions.", + "bbox": [ + 75, + 474, + 468, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparative analysis of text encoders. In Table 2 (rows two to four), we conduct a quantitative evaluation of various text encoders on T2I-CompBench, ensuring a fair comparison by maintaining a consistent XL/2 size across models. Results reveal that GenTron-T5XXL outperforms GenTron-CLIP-L across all three attribute binding and spatial relationship metrics, while it demonstrates comparable performance in the remaining two metrics. This suggests that T5 embeddings are superior in terms of compositional ability. These observations are in line with [2], which utilizes both CLIP and T5 embeddings for training but tests them individually or in combination during inference. Unlike eDiff,", + "bbox": [ + 75, + 719, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b567381917800d47279ed87745988bb05d38e90981403c764900d101ab7009be.jpg", + "image_caption": [ + "Figure 4. adaLN-Zero vs. cross attention. The prompt is \"A panda standing on a surfboard in the ocean in sunset.\" Cross attention exhibits a distinct advantage in the text-conditioned scenario.", + "(a) GenTron-XL/2", + "Figure 5. Effect of model scale. The prompt is \"a cat reading a newspaper\". The larger model GenTron-G/2 excels in rendering finer details and rationalization in the layout of the cat and newspaper. More comparisons can be found in the appendix." + ], + "image_footnote": [], + "bbox": [ + 509, + 88, + 694, + 229 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ae1c4fcef810c0358b38094c6070e77db9ef7cb1543a602ccdb72fa6e343df07.jpg", + "image_caption": [ + "(b) GenTron-G/2" + ], + "image_footnote": [], + "bbox": [ + 696, + 89, + 879, + 229 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1a556455ab05ef8cc627cd0cd1449d53214184572d68f1962cdabcf2484da7f5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method#Param.FID-30K↓CLIP-Score↑T2I-CompBen↑
Imagen3.0B7.270.27-
Parti-750M0.8B10.71--
Parti-3B3.0B8.10--
GigaGAN1.0B9.090.322-
MUSE-3B3.0B7.880.320-
SD v1.40.9B12.940.32531.50
SDXL2.6B17.820.32944.41
GenTron-XL/20.9B14.210.32649.13
GenTron-G/23.1B14.530.33549.99
", + "bbox": [ + 501, + 310, + 890, + 452 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Comparison to T2I models.", + "bbox": [ + 581, + 464, + 808, + 477 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "our approach maintains the same settings for both training and inference. Notably, GenTron demonstrates enhanced performance when combining CLIP-L and T5XXL embeddings, indicating the model's ability to leverage the distinct advantages of each text embedding type.", + "bbox": [ + 498, + 500, + 890, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Scaling GenTron up. In Figure 5, we showcase examples from the PartiPrompts benchmark [65] to illustrate the qualitative enhancements achieved by scaling our model up from approximately 900 million to 3 billion parameters. Both models operate under the same CLIP-T5XXL condition. The larger GenTron-G/2 model excels in rendering finer details and more accurate representations, particularly in rationalizing the layout of objects like cats and newspapers. This results in image compositions that are both more coherent and more realistic. In comparison, the smaller GenTron-XL/2 model, while producing recognizable images with similar color schemes, falls short in terms of precision and visual appeal.", + "bbox": [ + 496, + 597, + 890, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Furthermore, the superiority of GenTron-G/2 is quantitatively affirmed through its performance on T2I-CompBench, as detailed in Table 2. The increase in model size correlates with significant improvements across all evaluative criteria for object composition, including attributes and relationships. Additional comparative examples are provided in the appendix for further illustration.", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6447", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9e76d4ee6a95af33bf39178cf9f107e18420b7cecdf7bb753ce13525ff201ac7.jpg", + "image_caption": [ + "Figure 6. GenTron-T2V examples. Prompts are \"A giant tortoise is making its way across the beach\" and \"A dog swimming\"." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 874, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b8916973999bebc8e3ed364d550f718a65574faeb7be3595a41fccc7f36f2ae4.jpg", + "image_caption": [ + "Figure 7. Visualization of the human preference of our method vs. Latent Diffusion XL. Our method received a significantly higher number of votes as the winner in comparisons of visual quality and text faithfulness, with a total of 3000 answers." + ], + "image_footnote": [], + "bbox": [ + 84, + 208, + 459, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison to prior work. In Table 3, we showcase the alignment evaluation results from T2I-CompBench. Our method demonstrates outstanding performance in all areas, including attribute binding, object relationships, and complex compositions. This indicates a heightened proficiency in compositional generation, with a notable strength in color binding. In this aspect, our approach surpasses the previous state-of-the-art (SoTA) work [10] by over $7\\%$ .", + "bbox": [ + 75, + 373, + 468, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Table 4, we further thoroughly compare the zero-shot FID-30K, CLIP-Score, and T2I-CompBench metrics with previous T2I models. In contrast to U-Net based SDv1.4, GenTron uses about four times less data (550M vs. 2B), while achieving better CLIP-scores and T2I-CompBench results. Although GenTron does not achieve a superior FID score compared to SDv1.4, it's important to note that recent studies, including Pick-a-Pic [36] and SDXL [46], have highlighted that FID scores often misrepresent human preferences in generative models, sometimes even negatively correlating with visual aesthetics.", + "bbox": [ + 75, + 496, + 468, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "User study Figure 7 shows the human preference of our method versus SDXL. We used standard prompts in PartiPrompt2 [65] to generate 100 images using both methods and ask people for their preference blindly after shuffling. We received a total of three thousand responses on the comparisons of visual quality and text faithfulness, with our method emerging as the clear winner by a clear margin.", + "bbox": [ + 75, + 684, + 468, + 791 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. GenTron-T2V Results", + "text_level": 1, + "bbox": [ + 76, + 800, + 284, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Figure 6, We showcase several samples generated by GenTron-T2V, which are not only visually striking but also temporally coherent. This highlights GenTron-T2V's effectiveness in creating videos that are both aesthetically pleasing and consistent over time.", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ea2c668e91e1c69d5f86bda74e77889ad223d7c571b7e9a9694b7bfefd704206.jpg", + "image_caption": [ + "Figure 8. Effect of motion-free guidance. GenTron-T2V with motion-free guidance has a clear visual appearance improvement. The prompt is \"A lion standing on a surfboard in the ocean in sunset\"." + ], + "image_footnote": [], + "bbox": [ + 506, + 208, + 883, + 348 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of motion-free guidance. In Figure 8, we present a comparison between our GenTron variants, with motion-free guidance (MFG) and without. For this comparison, critical factors such as the pre-trained T2I model, training data, and the number of training iterations were kept constant to ensure a fair evaluation. The results clearly indicate that GenTron-T2V, when integrated with MFG, shows a marked tendency to focus on the central object mentioned in the prompt, often rendering it in greater detail. Specifically, the object typically occupies a more prominent, central position in the generated video, thereby dominating the visual focus across video frames.", + "bbox": [ + 498, + 414, + 892, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 616, + 619, + 632 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we provide a thorough exploration of transformer-based diffusion models for text-conditioned image and video generation. Our findings shed light on the properties of various conditioning approaches and offer compelling evidence of quality improvement when scaling up the model. A notable contribution of our work is the development of GenTron for video generation, where we introduce motion-free guidance. This innovative approach has demonstrably enhanced the visual quality of generated videos. We hope that our research will contribute to bridging the existing gap in applying transformers to diffusion models and their broader use in other domains.", + "bbox": [ + 498, + 643, + 893, + 825 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This paper is partially supported by the National Key R&D Program of China No.2022ZD0161000 and the General Research Fund of Hong Kong No.17200622.", + "bbox": [ + 498, + 839, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6448", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Max Bain, Arsha Nagrani, Güll Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1728-1738, 2021. 5", + "[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 4, 7", + "[3] Fan Bao, Shen Nie, Kaiwen Xue, Yue Cao, Chongxuan Li, Hang Su, and Jun Zhu. All are worth words: A vit backbone for diffusion models. In CVPR, 2023. 3", + "[4] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions, 2023. 6", + "[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 2, 3, 4, 5", + "[6] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 5, 6", + "[7] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. OpenAI, 2024. 2", + "[8] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 2", + "[9] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. Videocrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 2, 3, 5", + "[10] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 3, 6, 8", + "[11] Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish V Thapliyal, James Bradbury, Weicheng Kuo, Mojtaba Seyedhosseini, Chao Jia, Burcu Karagol Ayan, Carlos Riquelme Ruiz, Andreas Peter Steiner, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Radu Soricut. PaLI: A jointly-scaled multilingual language-image model. In The Eleventh International Conference on Learning Representations, 2023. 2", + "[12] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 2", + "[13] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416, 2022. 2, 4, 6", + "[14] Yuren Cong, Mengmeng Xu, christian simon, Shoufa Chen, Jiawei Ren, Yanping Xie, Juan-Manuel Perez-Rua, Bodo Rosenhahn, Tao Xiang, and Sen He. FLATTEN: optical FLOW-guided ATTENTION for consistent text-to-video editing. In The Twelfth International Conference on Learning Representations, 2024. 2", + "[15] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xiaofang Wang, Abhimanyu Dubey, et al. Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807, 2023. 2, 4, 5", + "[16] Yatin Dandi, Aniket Das, Soumye Singhal, Vinay Nambodiri, and Piyush Rai. Jointly trained image and video generation using residual vectors. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3028-3042, 2020. 2, 5", + "[17] Mostafa Dehghani, Josip Djolonga, Basil Mustafa, Piotr Padlewski, Jonathan Heek, Justin Gilmer, Andreas Peter Steiner, Mathilde Caron, Robert Geirhos, Ibrahim Alabdulmohsin, et al. Scaling vision transformers to 22 billion parameters. In International Conference on Machine Learning, pages 7480-7512. PMLR, 2023. 2", + "[18] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 7", + "[19] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers), pages 4171-4186. Association for Computational Linguistics, 2019. 2", + "[20] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 2, 3", + "[21] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Pro" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6449", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 3", + "[22] Shanghua Gao, Pan Zhou, Ming-Ming Cheng, and Shuicheng Yan. Masked diffusion transformer is a strong image synthesizer. arXiv preprint arXiv:2303.14389, 2023. 3", + "[23] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 4", + "[24] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 3", + "[25] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2022. 3", + "[26] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 2, 5", + "[27] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 3", + "[28] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2", + "[29] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. The Journal of Machine Learning Research, 23(1):2249-2281, 2022. 4", + "[30] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video diffusion models, 2022. 2, 4, 5", + "[31] Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 2, 6", + "[32] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 4", + "[33] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6007-6017, 2023. 2", + "[34] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439, 2023. 2", + "[35] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 2, 3" + ], + "bbox": [ + 78, + 92, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Matiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36, 2024. 8", + "[37] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2", + "[38] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6", + "[39] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 3", + "[40] Haoyu Lu, Guoxing Yang, Nanyi Fei, Yuqi Huo, Zhiwu Lu, Ping Luo, and Mingyu Ding. Vdt: An empirical study on video diffusion with transformers. arXiv preprint arXiv:2305.13311, 2023. 3", + "[41] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Tieniu Tan. Videofusion: Decomposed diffusion models for high-quality video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10209-10218, 2023. 2", + "[42] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2", + "[43] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob Mcgrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In Proceedings of the 39th International Conference on Machine Learning, pages 16784-16804. PMLR, 2022. 4", + "[44] William Peebles and Saining Xie. Scalable diffusion models with transformers. arXiv preprint arXiv:2212.09748, 2022. 2,3,4,7", + "[45] Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. In Proceedings of the AAAI conference on artificial intelligence, 2018. 2, 3", + "[46] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 4, 5, 6, 8", + "[47] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. arXiv preprint arXiv:2303.09535, 2023. 2", + "[48] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. OpenAI blog, 2018. 2" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6450", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2", + "[50] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 4, 6", + "[51] Carlos Riquelme, Joan Puigcerver, Basil Mustafa, Maxim Neumann, Rodolphe Jenatton, André Susano Pinto, Daniel Keysers, and Neil Houlsby. Scaling vision with sparse mixture of experts. Advances in Neural Information Processing Systems, 34:8583-8595, 2021. 2", + "[52] Alex Rogozhnikov. Einops: Clear and reliable tensor manipulations with einstein-like notation. In International Conference on Learning Representations, 2021. 5", + "[53] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3, 5, 6", + "[54] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 2, 3", + "[55] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 2", + "[56] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 5", + "[57] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-video: Text-to-video generation without text-video data. In The Eleventh International Conference on Learning Representations, 2023. 2", + "[58] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3", + "[59] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2", + "[60] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2, 4" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[61] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 3", + "[62] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 2, 5", + "[63] Yilin Wang, Zeyuan Chen, Liangjun Zhong, Zheng Ding, Zhizhou Sha, and Zhuowen Tu. Dolfin: Diffusion layout transformers without autoencoder. arXiv preprint arXiv:2310.16305, 2023. 3", + "[64] Zeyue Xue, Guanglu Song, Qiushan Guo, Boxiao Liu, Zhuofan Zong, Yu Liu, and Ping Luo. RAPHAEL: Text-to-image generation via large mixture of diffusion paths. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 5", + "[65] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. Transactions on Machine Learning Research, 2022. Featured Certification. 7, 8", + "[66] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12104-12113, 2022. 2, 4", + "[67] Yanli Zhao, Andrew Gu, Rohan Varma, Liang Luo, Chien-Chin Huang, Min Xu, Less Wright, Hamid Shojanazeri, Myle Ott, Sam Shleifer, et al. Pytorch fsdp: experiences on scaling fully sharded data parallel. arXiv preprint arXiv:2304.11277, 2023. 6", + "[68] Hongkai Zheng, Weili Nie, Arash Vahdat, and Anima Anandkumar. Fast training of diffusion models with masked transformers. arXiv preprint arXiv:2306.09305, 2023. 3", + "[69] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 3, 4" + ], + "bbox": [ + 501, + 92, + 890, + 654 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "6451", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/GenTron_ Diffusion Transformers for Image and Video Generation/2bb6a697-7387-4091-9b21-cb7c548eb341_model.json b/2024/GenTron_ Diffusion Transformers for Image and Video Generation/2bb6a697-7387-4091-9b21-cb7c548eb341_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b759f5791e1b136f30544d50b85b2398d2fe34de --- /dev/null +++ b/2024/GenTron_ Diffusion Transformers for Image and Video Generation/2bb6a697-7387-4091-9b21-cb7c548eb341_model.json @@ -0,0 +1,2345 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.147, + 0.131, + 0.823, + 0.152 + ], + "angle": 0, + "content": "GenTron: Diffusion Transformers for Image and Video Generation" + }, + { + "type": "text", + "bbox": [ + 0.134, + 0.18, + 0.834, + 0.234 + ], + "angle": 0, + "content": "Shoufa Chen\\(^{1,2*}\\) Mengmeng Xu\\(^{2*}\\) Jiawei Ren\\(^{2}\\) Yuren Cong\\(^{2}\\) Sen He\\(^{2}\\) Yanping Xie\\(^{2}\\) Animesh Sinha\\(^{2}\\) Ping Luo\\(^{1}\\) Tao Xiang\\(^{2}\\) Juan-Manuel Perez-Rua\\(^{2}\\) The University of Hong Kong \\({}^{2}\\)Meta" + }, + { + "type": "image", + "bbox": [ + 0.105, + 0.25, + 0.36, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.251, + 0.614, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.614, + 0.251, + 0.865, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.447, + 0.486, + 0.533 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.486, + 0.447, + 0.866, + 0.532 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.533, + 0.486, + 0.621 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.486, + 0.533, + 0.866, + 0.621 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.158, + 0.642, + 0.808, + 0.657 + ], + "angle": 0, + "content": "Figure 1. GenTron: Transformer based diffusion model for high-quality text-to-image/video generation." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.67, + 0.312, + 0.685 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.705, + 0.471, + 0.901 + ], + "angle": 0, + "content": "In this study, we explore Transformer-based diffusion models for image and video generation. Despite the dominance of Transformer architectures in various fields due to their flexibility and scalability, the visual generative domain primarily utilizes CNN-based U-Net architectures, particularly in diffusion-based models. We introduce GenTron, a family of Generative models employing Transformer-based diffusion, to address this gap. Our initial step was to adapt Diffusion Transformers (DiTs) from class to text conditioning, a process involving thorough empirical exploration of the conditioning mechanism. We then scale GenTron from approximately 900M to over 3B parameters, observing improvements in visual quality. Furthermore, we extend" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.671, + 0.892, + 0.837 + ], + "angle": 0, + "content": "GenTron to text-to-video generation, incorporating novel motion-free guidance to enhance video quality. In human evaluations against SDXL, GenTron achieves a \\(51.1\\%\\) win rate in visual quality (with a \\(19.8\\%\\) draw rate), and a \\(42.3\\%\\) win rate in text alignment (with a \\(42.9\\%\\) draw rate). GenTron notably performs well in T2I-CompBench, highlighting its compositional generation ability. We hope GenTron could provide meaningful insights and serve as a valuable reference for future research. Please refer to the website1 and the arXiv version for the most up-to-date results: https://arxiv.org/abs/2312.04557." + }, + { + "type": "page_footnote", + "bbox": [ + 0.524, + 0.875, + 0.87, + 0.901 + ], + "angle": 0, + "content": "\\(^{1}\\)https://www.shoufachen.com/gentron網站/ \n*Equal contribution." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "6441" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.208, + 0.106 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.115, + 0.47, + 0.356 + ], + "angle": 0, + "content": "Diffusion models have recently shown remarkable progress in content creation, impacting areas such as image generation [27, 53, 55], video generation [5, 7, 28, 57], and editing [14, 33]. Among these notable developments, the CNN-based U-Net architecture has emerged as the predominant backbone design, a choice that stands in contrast to the prevailing trend in natural language processing [8, 19, 59] and computer visual perception [20, 37, 66] domains, where attention-based transformer architectures [60] have empowered a renaissance and become increasingly dominant. To provide a comprehensive understanding of Transformers in diffusion generation and to bridge the gap in architectural choices between visual generation and the other two domains — visual perception and NLP — a thorough investigation of visual generation using Transformers is of substantial scientific value." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.357, + 0.471, + 0.644 + ], + "angle": 0, + "content": "We focus on diffusion models with Transformers in this work. Specifically, our starting point is the foundational work known as DiT [44], which introduced a class-conditioned latent diffusion model that employs a Transformer to replace the traditionally used U-Net architecture. We overcome the limitation of the original DiT model, which is constrained to handling only a restricted number (e.g., 1000) of predefined classes, by utilizing language embeddings derived from open-world, free-form text captions instead of predefined one-hot class embeddings. Along the way, we comprehensively investigate conditioning strategies, including (1) conditioning architectures: adaptive layer norm (adaLN) [45] vs. cross-attention [60]; and (2) text encoding methods: a generic large language model [13] vs. the language tower of multimodal models [50], or the combination of both of them. We additionally carry out comparative experiments and offer detailed empirical analyses to evaluate the effectiveness of these conditioning strategies." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.645, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Next, we explore the scaling-up properties of GenTron. The Transformer architectures have been demonstrated to possess significant scalability in both visual perception [11, 17, 51, 66] and language [8, 19, 48, 49, 59] tasks. For example, the largest dense language model has 540B parameters [12], and the largest vision model has 22B [17] parameters. In contrast, the largest diffusion transformer, DiT-XL [44], only has about 675M parameters, trailed far behind both the Transformers utilized in other domains (e.g., NLP) and recent diffusion arts with convolutional U-Net architectures [15, 46]. To compensate for this considerable lagging, we scale up GenTron in two dimensions, the number of transformer blocks and hidden dimension size, following the scaling strategy in [66]. As a result, our largest model, GenTron-G/2, has more than 3B parameters and achieves significant visual quality improvement compared with the smaller one." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.302 + ], + "angle": 0, + "content": "Furthermore, we have advanced GenTron from a T2I to a T2V model by inserting a temporal self-attention layer into each transformer block, making the first attempt to use transformers as the exclusive building block for video diffusion models. We also discuss existing challenges in video generation and introduce our solution, the motion-free guidance (MFG). Specifically, This approach involves intermittently disabling motion modeling during training by setting the temporal self-attention mask to an identity matrix. Besides, MFG seamlessly integrates with the joint image-video strategy [9, 16, 30, 62], where images are used as training samples whenever motion is deactivated. Our experiments indicate that this approach clearly improves the visual quality of generated videos." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.303, + 0.892, + 0.454 + ], + "angle": 0, + "content": "In human evaluations, GenTron outperforms SDXL, achieving a \\(51.1\\%\\) win rate in visual quality (with a \\(19.8\\%\\) draw rate), and a \\(42.3\\%\\) win rate in text alignment (with a \\(42.9\\%\\) draw rate). Furthermore, when compared to previous studies, particularly as benchmarked against T2I-CompBench [31]-a comprehensive framework for evaluating open-world compositional T2I generation-GenTron demonstrates superior performance across various criteria. These include attribute binding, object relationships, and handling of complex compositions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.455, + 0.892, + 0.651 + ], + "angle": 0, + "content": "Our contributions are summarized as follows: (1) We have conducted a thorough and systematic investigation of transformer-based T2I generation with diffusion models. This study encompasses various conditioning choices and aspects of model scaling. (2) In a pioneering effort, we explore a purely transformer-based diffusion model for T2V generation. We introduce motion-free guidance, an innovative technique that efficiently fine-tunes T2I generation models for producing high-quality videos. (3) Experimental results indicate a clear preference for GenTron over SDXL in human evaluations. Furthermore, GenTron demonstrates superior performance compared to existing methods in the T2I-CompBench evaluations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.665, + 0.642, + 0.68 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Diffusion models for T2I and T2V generation. Diffusion models [27, 42] are a type of generative model that creates data samples from random noise. Later, latent diffusion models [44, 46, 53] are proposed for efficient T2I generation. These designs usually have \\( I \\) a pre-trained Variational Autoencoder [35] that maps images to a compact latent space, 2) a conditioner modeled by cross-attention [28, 53] to process text as conditions with a strength control [26], and 3) a backbone network, U-Net [54] in particular, to process image features. The success of diffusion on T2I generation tasks underscores the promising potential for text-to-video (T2V) generation [34, 41, 47, 57]. VDM [30] and Imagen Video [28] extend the image diffusion architecture on the temporal dimension with promising initial results. To" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6442" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.138 + ], + "angle": 0, + "content": "avoid excessive computing demands, video latent diffusion models [5, 9, 25, 40, 61, 69] implement the video diffusion process in a low-dimensional latent space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.159, + 0.473, + 0.449 + ], + "angle": 0, + "content": "Transformer-based Diffusion. Recently, Transformer-based Diffusion models have attracted increasing research interest. Among these, U-ViT [3] treats all inputs as tokens by integrating transformer blocks with a U-net architecture. In contrast, DiT [44] employs a simpler, non-hierarchical transformer structure. MDT [22] and MaskDiT [68] enhance DiT's training efficiency by incorporating the mask strategy [24]. Dolfin [63] is a transformer-based model for layout generation. Concurrently to this work, PixArt-\\(\\alpha\\) [10] demonstrates promising outcomes in Transformer-based T2I diffusion. It's trained using a three-stage decomposition process with high-quality data. Our work diverges from PixArt-\\(\\alpha\\) in key aspects. Firstly, while PixArt-\\(\\alpha\\) emphasizes training efficiency, our focus is on the design choice of conditioning strategy and scalability in T2I Transformer diffusion models. Secondly, we extend our exploration beyond image generation to video diffusion. We propose an innovative approach in video domain, which is not covered by PixArt-\\(\\alpha\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.462, + 0.169, + 0.478 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.488, + 0.47, + 0.595 + ], + "angle": 0, + "content": "We first introduce the preliminaries in Section 3.1, and then present the details of GenTron for text-to-image generation in Section 3.2, which includes text encoder models, embedding integration methods, and scaling up strategy of GenTron. Lastly, in Section 3.3, we extend GenTron's application to video generation, building on top of the T2I foundations laid in previous sections." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.606, + 0.219, + 0.62 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.63, + 0.47, + 0.751 + ], + "angle": 0, + "content": "Diffusion models. Diffusion models [27] have emerged as a family of generative models that generate data by performing a series of transformations on random noise. They are characterized by a forward and a backward process. Given an instance from the data distribution \\( x_0 \\sim p(x_0) \\), random Gaussian noise is iteratively added to the instance in the forward noisig process to create a Markov Chain of random latent variable \\( x_1, x_2, \\ldots, x_T \\) following:" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.764, + 0.47, + 0.783 + ], + "angle": 0, + "content": "\\[\nq \\left(x _ {t} \\mid x _ {t - 1}\\right) = \\mathcal {N} \\left(x _ {t}; \\sqrt {1 - \\beta_ {i}} x _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.901 + ], + "angle": 0, + "content": "where \\(\\beta_{1},\\ldots ,\\beta_{T}\\) are hyperparameters corresponding to the noise schedule. After a large enough number of diffusion steps, \\(x_{T}\\) can be viewed as a standard Gaussian noise. A denoising network \\(\\epsilon_{\\theta}\\) is further trained to learn the backward process, i.e., how to remove the noise from a noisy input [27]. For inference, an instance can be sampled starting from a random Gaussian noise \\(x_{T}\\sim \\mathcal{N}(0;\\mathbf{I})\\) and denoised" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.089, + 0.685, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.559, + 0.361, + 0.645, + 0.373 + ], + "angle": 0, + "content": "(a)adaLN-Zero" + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.089, + 0.868, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.738, + 0.361, + 0.838, + 0.372 + ], + "angle": 0, + "content": "(b) Cross attention" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.38, + 0.892, + 0.451 + ], + "angle": 0, + "content": "Figure 2. Text embedding integration architecture. We directly adapt adaLN from DiT [44], substituting the one-hot class embedding with text embedding. For cross attention, different from the approach in [44], we maintain the use of adaLN to model the combination of time embedding and the aggregated text embedding." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.47, + 0.89, + 0.502 + ], + "angle": 0, + "content": "step-by-step following the Markov Chain, i.e., by sequentially sampling \\( x_{t-1} \\) to \\( x_0 \\) with \\( p_\\theta(x_{t-1}|x_t) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.543, + 0.512, + 0.892, + 0.545 + ], + "angle": 0, + "content": "\\[\nx _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(x _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta} (x _ {t}, t)\\right) + \\sigma_ {t} \\mathbf {z}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.558, + 0.892, + 0.62 + ], + "angle": 0, + "content": "where \\(\\bar{\\alpha}_t = \\prod_{s=1}^t \\alpha_s\\), \\(\\alpha_t = 1 - \\beta_t\\) and \\(\\sigma_t\\) is the noise scale. In practical application, the diffusion sampling process can be further accelerated using different sampling techniques [39, 58]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.64, + 0.892, + 0.821 + ], + "angle": 0, + "content": "Latent diffusion model architectures. Latent diffusion models (LDMs) [53] reduce the high computational cost by conducting the diffusion process in the latent space. First, a pre-trained autoencoder [21, 35] is utilized to compress the raw image from pixel to latent space, then the diffusion models, which are commonly implemented with a U-Net [54] backbone, work on the latent space. Peebles et al. proposed DiT [44] to leverage the transformer architecture as an alternative to the traditional U-Net backbone for class-conditioned image generation, adopting the adaptive layernorm (adaLN [45]) for class conditioning mechanism, as shown in Figure 2a." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.832, + 0.719, + 0.848 + ], + "angle": 0, + "content": "3.2. Text-to-Image GenTron" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Our GenTron is built upon the DiT-XL/2 [44], which converts the latent of shape \\(32 \\times 32 \\times 4\\) to a sequence of non-overlapping tokens with a \\(2 \\times 2\\) patchify layer [20]. Then," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6443" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "these tokens are sent into a series of transformer blocks. Finally, a standard linear decoder is applied to convert these image tokens into latent space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.138, + 0.471, + 0.334 + ], + "angle": 0, + "content": "While DiT has shown that transformer-based models yield promising results in class-conditioned scenarios, it did not explore the realm of T2I generation. This field poses a considerable challenge, given its less constrained conditioning format. Moreover, even the largest DiT model, DiT-XL/2, with its 675 million parameters, is significantly overshadowed by current U-Nets [15, 46], which boast over 3 billion parameters. To address these limitations, our research conducts a thorough investigation of transformer-based T2I diffusion models, focusing specifically on text conditioning approaches and assessing the scalability of the transformer architecture by expanding GenTron to more than 3 billion parameters." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.354, + 0.335, + 0.368 + ], + "angle": 0, + "content": "3.2.1 From Class to Text Condition" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.378, + 0.47, + 0.514 + ], + "angle": 0, + "content": "T2I diffusion models rely on textual inputs to steer the process of image generation. The mechanism of text conditioning involves two critical components: firstly, the selection of a text encoder, which is responsible for converting raw text into text embeddings, and secondly, the method of integrating these embeddings into the diffusion process. For a complete understanding, we have included in the appendix a detailed presentation of the decisions made in existing works concerning these two components." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.534, + 0.47, + 0.716 + ], + "angle": 0, + "content": "Text encoder model. Current advancements in T2I diffusion techniques employ a variety of language models, each with its unique strengths and limitations. To thoroughly assess which model best complements transformer-based diffusion methods, we have integrated several models into GenTron. This includes the text towers from multimodal models, CLIP [50], as well as a pure large language model, Flan-T5 [13]. Our approach explores the effectiveness of these language models by integrating each model independently with GenTron to evaluate their individual performance and combinations of them to assess the potential properties they may offer when used together." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.902 + ], + "angle": 0, + "content": "**Embedding integration.** In our study, we focused on two methods of embedding integration: adaptive layernorm and cross-attention. (1) Adaptive layernorm (adaLN). As shown in Figure 2a, this method integrates conditioning embeddings as normalization parameters on the feature channel. Widely used in conditional generative modeling, such as in StyleGAN [32], adaLN serves as the standard approach in DiT [44] for managing class conditions. (2) Cross-attention. As illustrated in Figure 2b, the image feature acts as the query, with textual embedding serving as key and value. This setup allows for direct interaction" + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.089, + 0.887, + 0.133 + ], + "angle": 0, + "content": "
ModelDepthWidthMLP Width#Param.
GenTron-XL/22811524608930.0M
GenTron-G/248166466563083.8M
" + }, + { + "type": "table_caption", + "bbox": [ + 0.538, + 0.142, + 0.855, + 0.156 + ], + "angle": 0, + "content": "Table 1. Configuration details of GenTron models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.176, + 0.892, + 0.342 + ], + "angle": 0, + "content": "between the image feature and textual embedding through an attention mechanism [60]. Besides, different from the cross-attention discussed in [44], which processes the class embedding and time embedding together by firstly concatenating them, we maintain the use of adaLN in conjunction with the cross-attention to separately model the time embedding. The underlying rationale for this design is our belief that the time embedding, which is consistent across all spatial positions, benefits from the global modulation capabilities of adaLN. Moreover, we also add the pooled text embeddings to the time embedding followings [2, 29, 43, 46]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.361, + 0.697, + 0.377 + ], + "angle": 0, + "content": "3.2.2 Scaling Up GenTron" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.385, + 0.892, + 0.568 + ], + "angle": 0, + "content": "To explore the impact of substantially scaling up the model size, we have developed an advanced version of GenTron, which we refer to as GenTron-G/2. This model was constructed in accordance with the scaling principles outlined in [66]. We focused on expanding three critical aspects: the number of transformer blocks (depth), the dimensionality of patch embeddings (width), and the hidden dimension of the MLP (MLP-width). The specifications and configurations of the GenTron models are detailed in Table 1. Significantly, the GenTron-G/2 model boasts over 3 billion parameters. To our knowledge, this represents the largest transformer-based diffusion architecture developed to date." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.577, + 0.715, + 0.592 + ], + "angle": 0, + "content": "3.3. Text-to-Video GenTron" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.6, + 0.892, + 0.706 + ], + "angle": 0, + "content": "In this subsection, we elaborate on the process of adapting GenTron from a T2I framework to a T2V framework. Sec. 3.3.1 will detail the modifications made to the model's architecture, enabling GenTron to process video data. Furthermore, Sec. 3.3.2 will discuss the challenges encountered in the domain of video generation and the innovative solutions we have proposed to address them." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.726, + 0.746, + 0.74 + ], + "angle": 0, + "content": "3.3.1 GenTron-T2V Architecture" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Transformer block with temporal self-attention. It is typically a common practice to train video diffusion models from image diffusion models by adding new temporal modeling modules [5, 23, 30, 69]. These usually consist of 3D convolutional layers and temporal transformer blocks that focus on calculating attention along the temporal dimension. In contrast to the traditional approach [5], which involves adding both temporal convolution layers and temporal transformer blocks to the T2I U-Net, our method integrates only lightweight temporal" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6444" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.137, + 0.092, + 0.406, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.305, + 0.473, + 0.39 + ], + "angle": 0, + "content": "Figure 3. GenTron Transformer block with TempSelfAttn and Motion-Free Mask. The temporal self-attention layer is inserted between the cross-attention and the MLPs. The motion-free mask, which is an identity matrix, will be utilized in the TempSelfAttn with a probability of \\( p_{\\text{motion\\_free}} \\). We omit details like text conditioning, LN here for simplicity, which could be found in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.412, + 0.47, + 0.533 + ], + "angle": 0, + "content": "self-attention (TempSelfAttn) layers into each transformer block. As depicted in Figure 3, the TempSelfAttn layer is placed right after the cross-attention layer and before the MLP layer. Additionally, we modify the output of the cross-attention layer by reshaping it before it enters the TempSelfAttn layer and then reshape it back to its original format once it has passed through. This process can be formally represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.084, + 0.551, + 0.468, + 0.567 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} = \\text {r e a r r a n g e} (\\mathbf {x}, (\\mathrm {b} \\quad \\mathrm {t}) \\quad \\mathrm {n} \\quad \\mathrm {d} \\rightarrow (\\mathrm {b} \\quad \\mathrm {n}) \\quad \\mathrm {t} \\quad \\mathrm {d}) \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.569, + 0.469, + 0.586 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} = \\mathbf {x} + \\operatorname {T e m p S e l f A t t n} (\\mathrm {L N} (\\mathbf {x})) \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.589, + 0.469, + 0.605 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} = \\text {r e a r r a n g e} (\\mathbf {x}, (\\mathrm {b} \\quad \\mathrm {n}) \\quad \\mathrm {t} \\quad \\mathrm {d} \\rightarrow (\\mathrm {b} \\quad \\mathrm {t}) \\quad \\mathrm {n} \\quad \\mathrm {d}) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.623, + 0.47, + 0.744 + ], + "angle": 0, + "content": "where \\( b, t, n, d \\) represent the batch size, number of frames, number of patches per frame, and channel dimension, respectively. rearrange is a notation from [52]. We discovered that a simple TempSelfAttn layer suffices to capture motion, a finding that aligns with observations in a recent study [62]. In addition, only using TempSelfAttn makes it convenient to turn on and turn off the temporal modeling, which would be discussed in Sec. 3.3.2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Initialization. We use the pre-trained T2I model as a basis for initializing the shared layers between T2I and T2V models. In addition, for the newly added TempSelfAttn layers, we initialize the weights and biases of the output project layers to zero. This ensures that at the beginning of the T2V fine-tuning stage, these layers produce a zero output, effectively functioning as an identity mapping in conjunction with the shortcut connection." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.092, + 0.715, + 0.106 + ], + "angle": 0, + "content": "3.3.2 Motion-Free Guidance" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.117, + 0.894, + 0.253 + ], + "angle": 0, + "content": "Challenges encountered. We observed a notable phenomenon in the current T2V diffusion models [5, 30] where the per-frame visual quality significantly lags behind that of T2I models [15, 46, 53, 64]. Furthermore, our analysis revealed a remarkable degradation in visual quality in the T2V models post-fine-tuning, especially when compared to their original T2I counterparts. We note that these problems generally exist in current T2V diffusion models, not limited to our transformer-based T2V." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.278, + 0.893, + 0.52 + ], + "angle": 0, + "content": "Problem analysis and insights. We presume that the observed lag in the visual quality of T2V primarily stems from two factors: the nature of video data and the fine-tuning approach. Firstly, publicly available video datasets often fall short in both quality and quantity compared to image datasets. For instance, [56] has more than 2B English image-text pairs, whereas the current widely used video dataset, WebVid-10M [1] contains only 10.7M video-text pairs. Additionally, many video frames are compromised by motion blur and watermarks, further reducing their visual quality. This limited availability hampers the development of robust and versatile video diffusion models. Secondly, the focus on optimizing temporal aspects during video fine-tuning can inadvertently compromise the spatial visual quality, resulting in a decline in the overall quality of the generated videos." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.544, + 0.892, + 0.635 + ], + "angle": 0, + "content": "Solution I: joint image-video training. From the data aspect, we adopt the joint image-video training strategy [9, 16, 30, 62] to mitigate the video data shortages. Furthermore, joint training helps to alleviate the problem of domain discrepancy between video and image datasets by integrating both data types for training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.892, + 0.764 + ], + "angle": 0, + "content": "Solution II: motion-free guidance. We treat the temporal motion within a video clip as a special conditioning signal, which can be analogized to the textual conditioning in T2I/T2V diffusion models. Based on this analogy, we propose a novel approach, motion-free guidance (MFG), inspired by classifier-free guidance [6, 26], to modulate the weight of motion information in the generated video." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.893, + 0.902 + ], + "angle": 0, + "content": "In a particular training iteration, our approach mirrors the concept used in classifier-free guidance, where conditioned text is replaced with an empty string. The difference is that we employ an identity matrix to nullify the temporal attention with a probability of \\( p_{\\text{motion\\_free}} \\). This identity matrix, which is depicted in Figure 3 (Motion-Free Mask), is structured such that its diagonal is populated with ones, while all other positions are zeroes. This configuration confines the temporal self-attention to work within a single" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6445" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.126, + 0.089, + 0.842, + 0.208 + ], + "angle": 0, + "content": "
ConditioningScaleAttribute BindingObject RelationshipComplexMean
Text EncoderTypeIntegrationColorShapeTextureSpatialNon-spatial
CLIP-L [50]MMadaLN-zeroXL/236.9442.0650.739.4130.3836.4134.32
CLIP-L [50]MMcross-attnXL/273.9151.8168.7619.2631.8041.5247.84
T5-XXL [13]LLMcross-attnXL/274.9055.4070.0520.5231.6841.0148.93
CLIP-T5XXLMM + LLMcross-attnXL/275.6555.7469.4820.6731.7941.4449.13
CLIP-T5XXLMM + LLMcross-attnG/276.7457.0071.5020.9832.0241.6749.99
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.214, + 0.894, + 0.257 + ], + "angle": 0, + "content": "Table 2. Conditioning and model scale in GenTron. We compare GenTron model variants with different design choices on T2I-CompBench [31]. The text encoders are from the language tower of the multi-modal (MM) model, the large language model (LLM), or a combination of them. The GenTron-G/2 with CLIP-T5XXL performs best. Detailed discussions can be found in Sec. 4.2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.275, + 0.469, + 0.335 + ], + "angle": 0, + "content": "frame. Furthermore, as introduced in Sec. 3.3.1, temporal self-attention is the sole operator for temporal modeling. Thus, using a motion-free attention mask suffices to disable temporal modeling in the video diffusion process." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.336, + 0.469, + 0.366 + ], + "angle": 0, + "content": "During inference, we have text and motion conditionings. Inspired by [6], we can modify the score estimate as:" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.395, + 0.468, + 0.447 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\tilde {\\epsilon} _ {\\theta} = \\epsilon_ {\\theta} (x _ {t}, \\varnothing , \\varnothing) \\\\ + \\lambda_ {T} \\cdot \\left(\\epsilon_ {\\theta} \\left(x _ {t}, c _ {T}, c _ {M}\\right) - \\epsilon_ {\\theta} \\left(x _ {t}, \\varnothing , c _ {M}\\right)\\right) \\\\ + \\lambda_ {M} \\cdot \\left(\\epsilon_ {\\theta} \\left(x _ {t}, \\varnothing , c _ {M}\\right) - \\epsilon_ {\\theta} \\left(x _ {t}, \\varnothing , \\varnothing\\right)\\right) \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.459, + 0.469, + 0.58 + ], + "angle": 0, + "content": "where \\( c_{T} \\) and \\( c_{M} \\) represent the text conditioning and motion conditioning. \\( \\lambda_{T} \\) and \\( \\lambda_{M} \\) are the guidance scale of standard text and that of motion, controlling how strongly the generated samples correspond with the text condition and the motion strength, respectively. We empirically found that fixing \\( \\lambda_{T} = 7.5 \\) and adjusting \\( \\lambda_{M} \\in [1.0, 1.3] \\) for each example tend to achieve the best result. This finding is similar to [6], although our study utilizes a narrower range for \\( \\lambda_{M} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.469, + 0.687 + ], + "angle": 0, + "content": "Putting solutions together. We can integrate solution I and II together in the following way: when the motion is omitted at a training step, we load an image-text pair and repeat the image \\( T - 1 \\) times to create a pseudo video. Conversely, if motion is included, we instead load a video clip and extract it into \\( T \\) frames." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.702, + 0.209, + 0.719 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.727, + 0.294, + 0.743 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Training scheme For all GenTron model variations, we employ the AdamW [38] optimizer, maintaining a constant learning rate of \\(1 \\times 10^{-4}\\). We train our T2I GenTron models in a multi-stage procedure [46, 53] with an internal dataset, including a low-resolution \\((256 \\times 256)\\) training with a batch size of 2048 and 500K optimization steps, as well as high-resolution \\((512 \\times 512)\\) with a batch size of 784 and 300K steps. For the GenTron-G/2 model, we further integrate Fully Sharded Data Parallel (FSDP) [67] and activation checkpointing (AC), strategies specifically adopted to" + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.272, + 0.89, + 0.432 + ], + "angle": 0, + "content": "
ModelAttribute BindingObj. RelationComp.Mean
ColorShapeTextureSpat.Non-spat.
LDM v1.437.6535.7641.5612.4630.7930.8031.50
LDM v250.6542.2149.2213.4230.9633.8636.72
Composable v240.6332.9936.458.0029.8028.9829.47
Structured v249.9042.1849.0013.8631.1133.5536.60
Attn-Exact v264.0045.1759.6314.5531.0934.0141.41
GORS66.0347.8562.8718.1531.9333.2843.35
DALL-E 257.5054.6463.7412.8330.4336.9642.68
LDM XL63.6954.0856.3720.3231.1040.9144.41
PixArt-α68.8655.8270.4420.8231.7941.1748.15
GenTron76.7457.0071.5020.9832.0241.6749.99
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.436, + 0.892, + 0.492 + ], + "angle": 0, + "content": "Table 3. Comparison of alignment evaluation on T2I-CompBench [31]. Results show our advanced model, GenTron-CLIPT5XXL-G/2, achieves superior performance across multiple compositional metrics compared to previous methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.515, + 0.892, + 0.621 + ], + "angle": 0, + "content": "optimize GPU memory usage. In our video experiments, we train videos on a video dataset that comprises approximately 34M videos. To optimize storage usage and enhance data loading efficiency, the videos are pre-processed to a resolution with a short side of 512 pixels and a frame rate of 24 FPS. We process batches of 128 video clips. Each clip comprises 8 frames, captured at a sampling rate of 4 FPS." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.65, + 0.892, + 0.756 + ], + "angle": 0, + "content": "Evaluation metrics. We mainly adopt the recent T2I-CompBench [31] to compare GenTron model variants, following [4, 10]. Specifically, we compare the attribute binding aspects, which include color, shape, and texture. We also compare the spatial and non-spatial object relationships. Moreover, user studies are conducted to compare visual quality and text alignment." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.77, + 0.765, + 0.785 + ], + "angle": 0, + "content": "4.2. Main Results of GenTron-T2I" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this subsection, we discuss our experimental results, focusing on how various conditioning factors and model sizes impact GenTron's performance. Table 2 presents the quantitative findings of our study. Additionally, we offer a comparative visualization, illustrating the effects of each conditioning factor we explored. A comparison to prior art is also provided." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6446" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.089, + 0.272, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.137, + 0.234, + 0.221, + 0.245 + ], + "angle": 0, + "content": "(a)adaLN-Zero" + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.089, + 0.457, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.315, + 0.234, + 0.414, + 0.245 + ], + "angle": 0, + "content": "(b) Cross attention" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.25, + 0.47, + 0.293 + ], + "angle": 0, + "content": "Figure 4. adaLN-Zero vs. cross attention. The prompt is \"A panda standing on a surfboard in the ocean in sunset.\" Cross attention exhibits a distinct advantage in the text-conditioned scenario." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.309, + 0.468, + 0.474 + ], + "angle": 0, + "content": "Cross attention vs. adaLN-Zero. Recent findings [44] conclude that the adaLN design yields superior results in terms of the FID, outperforming both cross-attention and in-context conditioning in efficiency for class-based scenarios. However, our observations reveal a limitation of adaLN in handling free-form text conditioning. This shortcoming is evident in Figure 4, where adaLN's attempt to generate a panda image falls short, with cross-attention demonstrating a clear advantage. This is further verified quantitatively in the first two rows of Table 2, where cross-attention uniformly excels over adaLN in all evaluated metrics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.476, + 0.47, + 0.703 + ], + "angle": 0, + "content": "This outcome is reasonable considering the nature of class conditioning, which typically involves a limited set of fixed signals (e.g., the 1000 one-hot class embeddings for ImageNet [18]). In such contexts, the adaLN approach, operating at a spatially global level, adjusts the image features uniformly across all positions through the normalization layer, making it adequate for the static signals. In contrast, cross-attention treats spatial positions with more granularity. It differentiates between various spatial locations by dynamically modulating image features based on the cross-attention map between the text embedding and the image features. This spatial-sensitive processing is essential for free-from text conditioning, where the conditioning signals are infinitely diverse and demand detailed representation in line with the specific content of textual descriptions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Comparative analysis of text encoders. In Table 2 (rows two to four), we conduct a quantitative evaluation of various text encoders on T2I-CompBench, ensuring a fair comparison by maintaining a consistent XL/2 size across models. Results reveal that GenTron-T5XXL outperforms GenTron-CLIP-L across all three attribute binding and spatial relationship metrics, while it demonstrates comparable performance in the remaining two metrics. This suggests that T5 embeddings are superior in terms of compositional ability. These observations are in line with [2], which utilizes both CLIP and T5 embeddings for training but tests them individually or in combination during inference. Unlike eDiff," + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.089, + 0.695, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.553, + 0.234, + 0.651, + 0.245 + ], + "angle": 0, + "content": "(a) GenTron-XL/2" + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.09, + 0.88, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.743, + 0.234, + 0.833, + 0.245 + ], + "angle": 0, + "content": "(b) GenTron-G/2" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.25, + 0.892, + 0.307 + ], + "angle": 0, + "content": "Figure 5. Effect of model scale. The prompt is \"a cat reading a newspaper\". The larger model GenTron-G/2 excels in rendering finer details and rationalization in the layout of the cat and newspaper. More comparisons can be found in the appendix." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.311, + 0.892, + 0.453 + ], + "angle": 0, + "content": "
Method#Param.FID-30K↓CLIP-Score↑T2I-CompBen↑
Imagen3.0B7.270.27-
Parti-750M0.8B10.71--
Parti-3B3.0B8.10--
GigaGAN1.0B9.090.322-
MUSE-3B3.0B7.880.320-
SD v1.40.9B12.940.32531.50
SDXL2.6B17.820.32944.41
GenTron-XL/20.9B14.210.32649.13
GenTron-G/23.1B14.530.33549.99
" + }, + { + "type": "table_caption", + "bbox": [ + 0.582, + 0.465, + 0.809, + 0.478 + ], + "angle": 0, + "content": "Table 4. Comparison to T2I models." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.501, + 0.892, + 0.577 + ], + "angle": 0, + "content": "our approach maintains the same settings for both training and inference. Notably, GenTron demonstrates enhanced performance when combining CLIP-L and T5XXL embeddings, indicating the model's ability to leverage the distinct advantages of each text embedding type." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.892, + 0.794 + ], + "angle": 0, + "content": "Scaling GenTron up. In Figure 5, we showcase examples from the PartiPrompts benchmark [65] to illustrate the qualitative enhancements achieved by scaling our model up from approximately 900 million to 3 billion parameters. Both models operate under the same CLIP-T5XXL condition. The larger GenTron-G/2 model excels in rendering finer details and more accurate representations, particularly in rationalizing the layout of objects like cats and newspapers. This results in image compositions that are both more coherent and more realistic. In comparison, the smaller GenTron-XL/2 model, while producing recognizable images with similar color schemes, falls short in terms of precision and visual appeal." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Furthermore, the superiority of GenTron-G/2 is quantitatively affirmed through its performance on T2I-CompBench, as detailed in Table 2. The increase in model size correlates with significant improvements across all evaluative criteria for object composition, including attributes and relationships. Additional comparative examples are provided in the appendix for further illustration." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6447" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.875, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.108, + 0.169, + 0.86, + 0.184 + ], + "angle": 0, + "content": "Figure 6. GenTron-T2V examples. Prompts are \"A giant tortoise is making its way across the beach\" and \"A dog swimming\"." + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.209, + 0.46, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.293, + 0.471, + 0.351 + ], + "angle": 0, + "content": "Figure 7. Visualization of the human preference of our method vs. Latent Diffusion XL. Our method received a significantly higher number of votes as the winner in comparisons of visual quality and text faithfulness, with a total of 3000 answers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.374, + 0.469, + 0.496 + ], + "angle": 0, + "content": "Comparison to prior work. In Table 3, we showcase the alignment evaluation results from T2I-CompBench. Our method demonstrates outstanding performance in all areas, including attribute binding, object relationships, and complex compositions. This indicates a heightened proficiency in compositional generation, with a notable strength in color binding. In this aspect, our approach surpasses the previous state-of-the-art (SoTA) work [10] by over \\(7\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.497, + 0.47, + 0.664 + ], + "angle": 0, + "content": "In Table 4, we further thoroughly compare the zero-shot FID-30K, CLIP-Score, and T2I-CompBench metrics with previous T2I models. In contrast to U-Net based SDv1.4, GenTron uses about four times less data (550M vs. 2B), while achieving better CLIP-scores and T2I-CompBench results. Although GenTron does not achieve a superior FID score compared to SDv1.4, it's important to note that recent studies, including Pick-a-Pic [36] and SDXL [46], have highlighted that FID scores often misrepresent human preferences in generative models, sometimes even negatively correlating with visual aesthetics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.685, + 0.47, + 0.792 + ], + "angle": 0, + "content": "User study Figure 7 shows the human preference of our method versus SDXL. We used standard prompts in PartiPrompt2 [65] to generate 100 images using both methods and ask people for their preference blindly after shuffling. We received a total of three thousand responses on the comparisons of visual quality and text faithfulness, with our method emerging as the clear winner by a clear margin." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.801, + 0.285, + 0.817 + ], + "angle": 0, + "content": "4.3. GenTron-T2V Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In Figure 6, We showcase several samples generated by GenTron-T2V, which are not only visually striking but also temporally coherent. This highlights GenTron-T2V's effectiveness in creating videos that are both aesthetically pleasing and consistent over time." + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.209, + 0.885, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.35, + 0.892, + 0.392 + ], + "angle": 0, + "content": "Figure 8. Effect of motion-free guidance. GenTron-T2V with motion-free guidance has a clear visual appearance improvement. The prompt is \"A lion standing on a surfboard in the ocean in sunset\"." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.415, + 0.893, + 0.596 + ], + "angle": 0, + "content": "Effect of motion-free guidance. In Figure 8, we present a comparison between our GenTron variants, with motion-free guidance (MFG) and without. For this comparison, critical factors such as the pre-trained T2I model, training data, and the number of training iterations were kept constant to ensure a fair evaluation. The results clearly indicate that GenTron-T2V, when integrated with MFG, shows a marked tendency to focus on the central object mentioned in the prompt, often rendering it in greater detail. Specifically, the object typically occupies a more prominent, central position in the generated video, thereby dominating the visual focus across video frames." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.617, + 0.62, + 0.633 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.645, + 0.894, + 0.826 + ], + "angle": 0, + "content": "In this work, we provide a thorough exploration of transformer-based diffusion models for text-conditioned image and video generation. Our findings shed light on the properties of various conditioning approaches and offer compelling evidence of quality improvement when scaling up the model. A notable contribution of our work is the development of GenTron for video generation, where we introduce motion-free guidance. This innovative approach has demonstrably enhanced the visual quality of generated videos. We hope that our research will contribute to bridging the existing gap in applying transformers to diffusion models and their broader use in other domains." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement. This paper is partially supported by the National Key R&D Program of China No.2022ZD0161000 and the General Research Fund of Hong Kong No.17200622." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6448" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Max Bain, Arsha Nagrani, Güll Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1728-1738, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.185, + 0.472, + 0.257 + ], + "angle": 0, + "content": "[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.257, + 0.471, + 0.298 + ], + "angle": 0, + "content": "[3] Fan Bao, Shen Nie, Kaiwen Xue, Yue Cao, Chongxuan Li, Hang Su, and Jun Zhu. All are worth words: A vit backbone for diffusion models. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.471, + 0.354 + ], + "angle": 0, + "content": "[4] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.354, + 0.471, + 0.438 + ], + "angle": 0, + "content": "[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.439, + 0.471, + 0.507 + ], + "angle": 0, + "content": "[6] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.509, + 0.471, + 0.579 + ], + "angle": 0, + "content": "[7] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. OpenAI, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.58, + 0.471, + 0.649 + ], + "angle": 0, + "content": "[8] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.65, + 0.471, + 0.719 + ], + "angle": 0, + "content": "[9] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. Videocrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.72, + 0.471, + 0.789 + ], + "angle": 0, + "content": "[10] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.79, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[11] Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish V Thapliyal, James Bradbury, Weicheng Kuo, Mojtaba Seyedhosseini, Chao Jia, Burcu Karagol Ayan, Carlos Riquelme Ruiz, Andreas Peter Steiner, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "Radu Soricut. PaLI: A jointly-scaled multilingual language-image model. In The Eleventh International Conference on Learning Representations, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.894, + 0.207 + ], + "angle": 0, + "content": "[12] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.208, + 0.894, + 0.276 + ], + "angle": 0, + "content": "[13] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416, 2022. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.361 + ], + "angle": 0, + "content": "[14] Yuren Cong, Mengmeng Xu, christian simon, Shoufa Chen, Jiawei Ren, Yanping Xie, Juan-Manuel Perez-Rua, Bodo Rosenhahn, Tao Xiang, and Sen He. FLATTEN: optical FLOW-guided ATTENTION for consistent text-to-video editing. In The Twelfth International Conference on Learning Representations, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.892, + 0.433 + ], + "angle": 0, + "content": "[15] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xiaofang Wang, Abhimanyu Dubey, et al. Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807, 2023. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.892, + 0.504 + ], + "angle": 0, + "content": "[16] Yatin Dandi, Aniket Das, Soumye Singhal, Vinay Nambodiri, and Piyush Rai. Jointly trained image and video generation using residual vectors. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3028-3042, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.505, + 0.892, + 0.589 + ], + "angle": 0, + "content": "[17] Mostafa Dehghani, Josip Djolonga, Basil Mustafa, Piotr Padlewski, Jonathan Heek, Justin Gilmer, Andreas Peter Steiner, Mathilde Caron, Robert Geirhos, Ibrahim Alabdulmohsin, et al. Scaling vision transformers to 22 billion parameters. In International Conference on Machine Learning, pages 7480-7512. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.59, + 0.892, + 0.647 + ], + "angle": 0, + "content": "[18] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.648, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[19] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers), pages 4171-4186. Association for Computational Linguistics, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[20] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.872, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[21] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Pro" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "6449" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.093, + 0.469, + 0.12 + ], + "angle": 0, + "content": "ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.47, + 0.174 + ], + "angle": 0, + "content": "[22] Shanghua Gao, Pan Zhou, Ming-Ming Cheng, and Shuicheng Yan. Masked diffusion transformer is a strong image synthesizer. arXiv preprint arXiv:2303.14389, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.176, + 0.469, + 0.232 + ], + "angle": 0, + "content": "[23] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.469, + 0.3 + ], + "angle": 0, + "content": "[24] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.469, + 0.357 + ], + "angle": 0, + "content": "[25] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.358, + 0.469, + 0.399 + ], + "angle": 0, + "content": "[26] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.401, + 0.469, + 0.44 + ], + "angle": 0, + "content": "[27] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.469, + 0.51 + ], + "angle": 0, + "content": "[28] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.47, + 0.566 + ], + "angle": 0, + "content": "[29] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. The Journal of Machine Learning Research, 23(1):2249-2281, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.469, + 0.607 + ], + "angle": 0, + "content": "[30] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video diffusion models, 2022. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.609, + 0.469, + 0.678 + ], + "angle": 0, + "content": "[31] Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.469, + 0.734 + ], + "angle": 0, + "content": "[32] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.735, + 0.469, + 0.803 + ], + "angle": 0, + "content": "[33] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6007-6017, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.805, + 0.469, + 0.873 + ], + "angle": 0, + "content": "[34] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[35] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.16 + ], + "angle": 0, + "content": "[36] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Matiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[37] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[38] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.264, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[39] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[40] Haoyu Lu, Guoxing Yang, Nanyi Fei, Yuqi Huo, Zhiwu Lu, Ping Luo, and Mingyu Ding. Vdt: An empirical study on video diffusion with transformers. arXiv preprint arXiv:2305.13311, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.391, + 0.892, + 0.474 + ], + "angle": 0, + "content": "[41] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Tieniu Tan. Videofusion: Decomposed diffusion models for high-quality video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10209-10218, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.476, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[42] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.547, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[43] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob Mcgrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In Proceedings of the 39th International Conference on Machine Learning, pages 16784-16804. PMLR, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.632, + 0.892, + 0.672 + ], + "angle": 0, + "content": "[44] William Peebles and Saining Xie. Scalable diffusion models with transformers. arXiv preprint arXiv:2212.09748, 2022. 2,3,4,7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.675, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[45] Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. In Proceedings of the AAAI conference on artificial intelligence, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.732, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[46] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.802, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[47] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. arXiv preprint arXiv:2303.09535, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[48] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. OpenAI blog, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6450" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "[49] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.469, + 0.217 + ], + "angle": 0, + "content": "[50] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.469, + 0.287 + ], + "angle": 0, + "content": "[51] Carlos Riquelme, Joan Puigcerver, Basil Mustafa, Maxim Neumann, Rodolphe Jenatton, André Susano Pinto, Daniel Keysers, and Neil Houlsby. Scaling vision with sparse mixture of experts. Advances in Neural Information Processing Systems, 34:8583-8595, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.469, + 0.329 + ], + "angle": 0, + "content": "[52] Alex Rogozhnikov. Einops: Clear and reliable tensor manipulations with einstein-like notation. In International Conference on Learning Representations, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.469, + 0.399 + ], + "angle": 0, + "content": "[53] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.469, + 0.482 + ], + "angle": 0, + "content": "[54] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.469, + 0.566 + ], + "angle": 0, + "content": "[55] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.469, + 0.649 + ], + "angle": 0, + "content": "[56] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.469, + 0.732 + ], + "angle": 0, + "content": "[57] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-video: Text-to-video generation without text-video data. In The Eleventh International Conference on Learning Representations, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.469, + 0.774 + ], + "angle": 0, + "content": "[58] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[59] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[60] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[61] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[62] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[63] Yilin Wang, Zeyuan Chen, Liangjun Zhong, Zheng Ding, Zhizhou Sha, and Zhuowen Tu. Dolfin: Diffusion layout transformers without autoencoder. arXiv preprint arXiv:2310.16305, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.262, + 0.892, + 0.33 + ], + "angle": 0, + "content": "[64] Zeyue Xue, Guanglu Song, Qiushan Guo, Boxiao Liu, Zhuofan Zong, Yu Liu, and Ping Luo. RAPHAEL: Text-to-image generation via large mixture of diffusion paths. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.332, + 0.892, + 0.43 + ], + "angle": 0, + "content": "[65] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. Transactions on Machine Learning Research, 2022. Featured Certification. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.431, + 0.892, + 0.487 + ], + "angle": 0, + "content": "[66] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12104-12113, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.488, + 0.892, + 0.556 + ], + "angle": 0, + "content": "[67] Yanli Zhao, Andrew Gu, Rohan Varma, Liang Luo, Chien-Chin Huang, Min Xu, Less Wright, Hamid Shojanazeri, Myle Ott, Sam Shleifer, et al. Pytorch fsdp: experiences on scaling fully sharded data parallel. arXiv preprint arXiv:2304.11277, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.558, + 0.892, + 0.6 + ], + "angle": 0, + "content": "[68] Hongkai Zheng, Weili Nie, Arash Vahdat, and Anima Anandkumar. Fast training of diffusion models with masked transformers. arXiv preprint arXiv:2306.09305, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.601, + 0.892, + 0.655 + ], + "angle": 0, + "content": "[69] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.655 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.956 + ], + "angle": 0, + "content": "6451" + } + ] +] \ No newline at end of file diff --git a/2024/GenTron_ Diffusion Transformers for Image and Video Generation/2bb6a697-7387-4091-9b21-cb7c548eb341_origin.pdf b/2024/GenTron_ Diffusion Transformers for Image and Video Generation/2bb6a697-7387-4091-9b21-cb7c548eb341_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f390339fae4300b4513bb3b1e99e9dfd11ed7030 --- /dev/null +++ b/2024/GenTron_ Diffusion Transformers for Image and Video Generation/2bb6a697-7387-4091-9b21-cb7c548eb341_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9f57071d84c41152dfc38c378eba4f02379441026ab65ce42d5a7cb0b5c8b161 +size 3695374 diff --git a/2024/GenTron_ Diffusion Transformers for Image and Video Generation/full.md b/2024/GenTron_ Diffusion Transformers for Image and Video Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..80a1b606b0b49444d218d01f2a190829d85f7e38 --- /dev/null +++ b/2024/GenTron_ Diffusion Transformers for Image and Video Generation/full.md @@ -0,0 +1,314 @@ +# GenTron: Diffusion Transformers for Image and Video Generation + +Shoufa Chen $^{1,2*}$ Mengmeng Xu $^{2*}$ Jiawei Ren $^{2}$ Yuren Cong $^{2}$ Sen He $^{2}$ Yanping Xie $^{2}$ Animesh Sinha $^{2}$ Ping Luo $^{1}$ Tao Xiang $^{2}$ Juan-Manuel Perez-Rua $^{2}$ The University of Hong Kong ${}^{2}$ Meta + +![](images/e449d602918cf754384afa929d4d33a7802872219b2287935e6d1f3ca0b15005.jpg) + +![](images/39dacd659364b26924a31c15f7b7fc3868d0030c53794ccbee361706bf98cc51.jpg) + +![](images/73f12c5e67ca1fa52f149fb37427e7d7bea9fc8b528384069d13341f24df0b21.jpg) + +![](images/b5ef7783ec1b8f511faba7baa8c3cf25d763f8b7b17ca461816b78352feeb43f.jpg) + +![](images/a98345756a35f127066decf21da3bccba3d4704694f70d03c926322979c23747.jpg) + +![](images/a5baec7215de49c1a96add6f7dd16cf2926df3cdb676c96a97e82ca87ad754b4.jpg) +Figure 1. GenTron: Transformer based diffusion model for high-quality text-to-image/video generation. + +![](images/a74bbb5fe69b6e7559bc86c5951149c6d0403ac223ce685f030759ab27622002.jpg) + +# Abstract + +In this study, we explore Transformer-based diffusion models for image and video generation. Despite the dominance of Transformer architectures in various fields due to their flexibility and scalability, the visual generative domain primarily utilizes CNN-based U-Net architectures, particularly in diffusion-based models. We introduce GenTron, a family of Generative models employing Transformer-based diffusion, to address this gap. Our initial step was to adapt Diffusion Transformers (DiTs) from class to text conditioning, a process involving thorough empirical exploration of the conditioning mechanism. We then scale GenTron from approximately 900M to over 3B parameters, observing improvements in visual quality. Furthermore, we extend + +GenTron to text-to-video generation, incorporating novel motion-free guidance to enhance video quality. In human evaluations against SDXL, GenTron achieves a $51.1\%$ win rate in visual quality (with a $19.8\%$ draw rate), and a $42.3\%$ win rate in text alignment (with a $42.9\%$ draw rate). GenTron notably performs well in T2I-CompBench, highlighting its compositional generation ability. We hope GenTron could provide meaningful insights and serve as a valuable reference for future research. Please refer to the website1 and the arXiv version for the most up-to-date results: https://arxiv.org/abs/2312.04557. + +# 1. Introduction + +Diffusion models have recently shown remarkable progress in content creation, impacting areas such as image generation [27, 53, 55], video generation [5, 7, 28, 57], and editing [14, 33]. Among these notable developments, the CNN-based U-Net architecture has emerged as the predominant backbone design, a choice that stands in contrast to the prevailing trend in natural language processing [8, 19, 59] and computer visual perception [20, 37, 66] domains, where attention-based transformer architectures [60] have empowered a renaissance and become increasingly dominant. To provide a comprehensive understanding of Transformers in diffusion generation and to bridge the gap in architectural choices between visual generation and the other two domains — visual perception and NLP — a thorough investigation of visual generation using Transformers is of substantial scientific value. + +We focus on diffusion models with Transformers in this work. Specifically, our starting point is the foundational work known as DiT [44], which introduced a class-conditioned latent diffusion model that employs a Transformer to replace the traditionally used U-Net architecture. We overcome the limitation of the original DiT model, which is constrained to handling only a restricted number (e.g., 1000) of predefined classes, by utilizing language embeddings derived from open-world, free-form text captions instead of predefined one-hot class embeddings. Along the way, we comprehensively investigate conditioning strategies, including (1) conditioning architectures: adaptive layer norm (adaLN) [45] vs. cross-attention [60]; and (2) text encoding methods: a generic large language model [13] vs. the language tower of multimodal models [50], or the combination of both of them. We additionally carry out comparative experiments and offer detailed empirical analyses to evaluate the effectiveness of these conditioning strategies. + +Next, we explore the scaling-up properties of GenTron. The Transformer architectures have been demonstrated to possess significant scalability in both visual perception [11, 17, 51, 66] and language [8, 19, 48, 49, 59] tasks. For example, the largest dense language model has 540B parameters [12], and the largest vision model has 22B [17] parameters. In contrast, the largest diffusion transformer, DiT-XL [44], only has about 675M parameters, trailed far behind both the Transformers utilized in other domains (e.g., NLP) and recent diffusion arts with convolutional U-Net architectures [15, 46]. To compensate for this considerable lagging, we scale up GenTron in two dimensions, the number of transformer blocks and hidden dimension size, following the scaling strategy in [66]. As a result, our largest model, GenTron-G/2, has more than 3B parameters and achieves significant visual quality improvement compared with the smaller one. + +Furthermore, we have advanced GenTron from a T2I to a T2V model by inserting a temporal self-attention layer into each transformer block, making the first attempt to use transformers as the exclusive building block for video diffusion models. We also discuss existing challenges in video generation and introduce our solution, the motion-free guidance (MFG). Specifically, This approach involves intermittently disabling motion modeling during training by setting the temporal self-attention mask to an identity matrix. Besides, MFG seamlessly integrates with the joint image-video strategy [9, 16, 30, 62], where images are used as training samples whenever motion is deactivated. Our experiments indicate that this approach clearly improves the visual quality of generated videos. + +In human evaluations, GenTron outperforms SDXL, achieving a $51.1\%$ win rate in visual quality (with a $19.8\%$ draw rate), and a $42.3\%$ win rate in text alignment (with a $42.9\%$ draw rate). Furthermore, when compared to previous studies, particularly as benchmarked against T2I-CompBench [31]-a comprehensive framework for evaluating open-world compositional T2I generation-GenTron demonstrates superior performance across various criteria. These include attribute binding, object relationships, and handling of complex compositions. + +Our contributions are summarized as follows: (1) We have conducted a thorough and systematic investigation of transformer-based T2I generation with diffusion models. This study encompasses various conditioning choices and aspects of model scaling. (2) In a pioneering effort, we explore a purely transformer-based diffusion model for T2V generation. We introduce motion-free guidance, an innovative technique that efficiently fine-tunes T2I generation models for producing high-quality videos. (3) Experimental results indicate a clear preference for GenTron over SDXL in human evaluations. Furthermore, GenTron demonstrates superior performance compared to existing methods in the T2I-CompBench evaluations. + +# 2. Related Work + +Diffusion models for T2I and T2V generation. Diffusion models [27, 42] are a type of generative model that creates data samples from random noise. Later, latent diffusion models [44, 46, 53] are proposed for efficient T2I generation. These designs usually have $I$ a pre-trained Variational Autoencoder [35] that maps images to a compact latent space, 2) a conditioner modeled by cross-attention [28, 53] to process text as conditions with a strength control [26], and 3) a backbone network, U-Net [54] in particular, to process image features. The success of diffusion on T2I generation tasks underscores the promising potential for text-to-video (T2V) generation [34, 41, 47, 57]. VDM [30] and Imagen Video [28] extend the image diffusion architecture on the temporal dimension with promising initial results. To + +avoid excessive computing demands, video latent diffusion models [5, 9, 25, 40, 61, 69] implement the video diffusion process in a low-dimensional latent space. + +Transformer-based Diffusion. Recently, Transformer-based Diffusion models have attracted increasing research interest. Among these, U-ViT [3] treats all inputs as tokens by integrating transformer blocks with a U-net architecture. In contrast, DiT [44] employs a simpler, non-hierarchical transformer structure. MDT [22] and MaskDiT [68] enhance DiT's training efficiency by incorporating the mask strategy [24]. Dolfin [63] is a transformer-based model for layout generation. Concurrently to this work, PixArt- $\alpha$ [10] demonstrates promising outcomes in Transformer-based T2I diffusion. It's trained using a three-stage decomposition process with high-quality data. Our work diverges from PixArt- $\alpha$ in key aspects. Firstly, while PixArt- $\alpha$ emphasizes training efficiency, our focus is on the design choice of conditioning strategy and scalability in T2I Transformer diffusion models. Secondly, we extend our exploration beyond image generation to video diffusion. We propose an innovative approach in video domain, which is not covered by PixArt- $\alpha$ . + +# 3. Method + +We first introduce the preliminaries in Section 3.1, and then present the details of GenTron for text-to-image generation in Section 3.2, which includes text encoder models, embedding integration methods, and scaling up strategy of GenTron. Lastly, in Section 3.3, we extend GenTron's application to video generation, building on top of the T2I foundations laid in previous sections. + +# 3.1. Preliminaries + +Diffusion models. Diffusion models [27] have emerged as a family of generative models that generate data by performing a series of transformations on random noise. They are characterized by a forward and a backward process. Given an instance from the data distribution $x_0 \sim p(x_0)$ , random Gaussian noise is iteratively added to the instance in the forward noisig process to create a Markov Chain of random latent variable $x_1, x_2, \ldots, x_T$ following: + +$$ +q \left(x _ {t} \mid x _ {t - 1}\right) = \mathcal {N} \left(x _ {t}; \sqrt {1 - \beta_ {i}} x _ {t - 1}, \beta_ {t} \mathbf {I}\right), \tag {1} +$$ + +where $\beta_{1},\ldots ,\beta_{T}$ are hyperparameters corresponding to the noise schedule. After a large enough number of diffusion steps, $x_{T}$ can be viewed as a standard Gaussian noise. A denoising network $\epsilon_{\theta}$ is further trained to learn the backward process, i.e., how to remove the noise from a noisy input [27]. For inference, an instance can be sampled starting from a random Gaussian noise $x_{T}\sim \mathcal{N}(0;\mathbf{I})$ and denoised + +![](images/6f9293ac2ca6ee4a494efe39f6e5459242bbd1cc1d489750c3d4f6e0a26422f0.jpg) +(a)adaLN-Zero +Figure 2. Text embedding integration architecture. We directly adapt adaLN from DiT [44], substituting the one-hot class embedding with text embedding. For cross attention, different from the approach in [44], we maintain the use of adaLN to model the combination of time embedding and the aggregated text embedding. + +![](images/bfda24d2d5047514ccd5cfd22f5ccb8b82a8b15d1c13e105f1e2b97f47dc2ace.jpg) +(b) Cross attention + +step-by-step following the Markov Chain, i.e., by sequentially sampling $x_{t-1}$ to $x_0$ with $p_\theta(x_{t-1}|x_t)$ : + +$$ +x _ {t - 1} = \frac {1}{\sqrt {\alpha_ {t}}} \left(x _ {t} - \frac {1 - \alpha_ {t}}{\sqrt {1 - \bar {\alpha} _ {t}}} \epsilon_ {\theta} (x _ {t}, t)\right) + \sigma_ {t} \mathbf {z}, \tag {2} +$$ + +where $\bar{\alpha}_t = \prod_{s=1}^t \alpha_s$ , $\alpha_t = 1 - \beta_t$ and $\sigma_t$ is the noise scale. In practical application, the diffusion sampling process can be further accelerated using different sampling techniques [39, 58]. + +Latent diffusion model architectures. Latent diffusion models (LDMs) [53] reduce the high computational cost by conducting the diffusion process in the latent space. First, a pre-trained autoencoder [21, 35] is utilized to compress the raw image from pixel to latent space, then the diffusion models, which are commonly implemented with a U-Net [54] backbone, work on the latent space. Peebles et al. proposed DiT [44] to leverage the transformer architecture as an alternative to the traditional U-Net backbone for class-conditioned image generation, adopting the adaptive layernorm (adaLN [45]) for class conditioning mechanism, as shown in Figure 2a. + +# 3.2. Text-to-Image GenTron + +Our GenTron is built upon the DiT-XL/2 [44], which converts the latent of shape $32 \times 32 \times 4$ to a sequence of non-overlapping tokens with a $2 \times 2$ patchify layer [20]. Then, + +these tokens are sent into a series of transformer blocks. Finally, a standard linear decoder is applied to convert these image tokens into latent space. + +While DiT has shown that transformer-based models yield promising results in class-conditioned scenarios, it did not explore the realm of T2I generation. This field poses a considerable challenge, given its less constrained conditioning format. Moreover, even the largest DiT model, DiT-XL/2, with its 675 million parameters, is significantly overshadowed by current U-Nets [15, 46], which boast over 3 billion parameters. To address these limitations, our research conducts a thorough investigation of transformer-based T2I diffusion models, focusing specifically on text conditioning approaches and assessing the scalability of the transformer architecture by expanding GenTron to more than 3 billion parameters. + +# 3.2.1 From Class to Text Condition + +T2I diffusion models rely on textual inputs to steer the process of image generation. The mechanism of text conditioning involves two critical components: firstly, the selection of a text encoder, which is responsible for converting raw text into text embeddings, and secondly, the method of integrating these embeddings into the diffusion process. For a complete understanding, we have included in the appendix a detailed presentation of the decisions made in existing works concerning these two components. + +Text encoder model. Current advancements in T2I diffusion techniques employ a variety of language models, each with its unique strengths and limitations. To thoroughly assess which model best complements transformer-based diffusion methods, we have integrated several models into GenTron. This includes the text towers from multimodal models, CLIP [50], as well as a pure large language model, Flan-T5 [13]. Our approach explores the effectiveness of these language models by integrating each model independently with GenTron to evaluate their individual performance and combinations of them to assess the potential properties they may offer when used together. + +**Embedding integration.** In our study, we focused on two methods of embedding integration: adaptive layernorm and cross-attention. (1) Adaptive layernorm (adaLN). As shown in Figure 2a, this method integrates conditioning embeddings as normalization parameters on the feature channel. Widely used in conditional generative modeling, such as in StyleGAN [32], adaLN serves as the standard approach in DiT [44] for managing class conditions. (2) Cross-attention. As illustrated in Figure 2b, the image feature acts as the query, with textual embedding serving as key and value. This setup allows for direct interaction + +
ModelDepthWidthMLP Width#Param.
GenTron-XL/22811524608930.0M
GenTron-G/248166466563083.8M
+ +Table 1. Configuration details of GenTron models. + +between the image feature and textual embedding through an attention mechanism [60]. Besides, different from the cross-attention discussed in [44], which processes the class embedding and time embedding together by firstly concatenating them, we maintain the use of adaLN in conjunction with the cross-attention to separately model the time embedding. The underlying rationale for this design is our belief that the time embedding, which is consistent across all spatial positions, benefits from the global modulation capabilities of adaLN. Moreover, we also add the pooled text embeddings to the time embedding followings [2, 29, 43, 46]. + +# 3.2.2 Scaling Up GenTron + +To explore the impact of substantially scaling up the model size, we have developed an advanced version of GenTron, which we refer to as GenTron-G/2. This model was constructed in accordance with the scaling principles outlined in [66]. We focused on expanding three critical aspects: the number of transformer blocks (depth), the dimensionality of patch embeddings (width), and the hidden dimension of the MLP (MLP-width). The specifications and configurations of the GenTron models are detailed in Table 1. Significantly, the GenTron-G/2 model boasts over 3 billion parameters. To our knowledge, this represents the largest transformer-based diffusion architecture developed to date. + +# 3.3. Text-to-Video GenTron + +In this subsection, we elaborate on the process of adapting GenTron from a T2I framework to a T2V framework. Sec. 3.3.1 will detail the modifications made to the model's architecture, enabling GenTron to process video data. Furthermore, Sec. 3.3.2 will discuss the challenges encountered in the domain of video generation and the innovative solutions we have proposed to address them. + +# 3.3.1 GenTron-T2V Architecture + +Transformer block with temporal self-attention. It is typically a common practice to train video diffusion models from image diffusion models by adding new temporal modeling modules [5, 23, 30, 69]. These usually consist of 3D convolutional layers and temporal transformer blocks that focus on calculating attention along the temporal dimension. In contrast to the traditional approach [5], which involves adding both temporal convolution layers and temporal transformer blocks to the T2I U-Net, our method integrates only lightweight temporal + +![](images/d533bf56cb27eae85c1f8aa3ff697ebd95e0b758f57f208edc53ad3567e2c78c.jpg) +Figure 3. GenTron Transformer block with TempSelfAttn and Motion-Free Mask. The temporal self-attention layer is inserted between the cross-attention and the MLPs. The motion-free mask, which is an identity matrix, will be utilized in the TempSelfAttn with a probability of $p_{\text{motion\_free}}$ . We omit details like text conditioning, LN here for simplicity, which could be found in Figure 2. + +self-attention (TempSelfAttn) layers into each transformer block. As depicted in Figure 3, the TempSelfAttn layer is placed right after the cross-attention layer and before the MLP layer. Additionally, we modify the output of the cross-attention layer by reshaping it before it enters the TempSelfAttn layer and then reshape it back to its original format once it has passed through. This process can be formally represented as: + +$$ +\mathbf {x} = \text {r e a r r a n g e} (\mathbf {x}, (\mathrm {b} \quad \mathrm {t}) \quad \mathrm {n} \quad \mathrm {d} \rightarrow (\mathrm {b} \quad \mathrm {n}) \quad \mathrm {t} \quad \mathrm {d}) \tag {3} +$$ + +$$ +\mathbf {x} = \mathbf {x} + \operatorname {T e m p S e l f A t t n} (\mathrm {L N} (\mathbf {x})) \tag {4} +$$ + +$$ +\mathbf {x} = \text {r e a r r a n g e} (\mathbf {x}, (\mathrm {b} \quad \mathrm {n}) \quad \mathrm {t} \quad \mathrm {d} \rightarrow (\mathrm {b} \quad \mathrm {t}) \quad \mathrm {n} \quad \mathrm {d}) \tag {5} +$$ + +where $b, t, n, d$ represent the batch size, number of frames, number of patches per frame, and channel dimension, respectively. rearrange is a notation from [52]. We discovered that a simple TempSelfAttn layer suffices to capture motion, a finding that aligns with observations in a recent study [62]. In addition, only using TempSelfAttn makes it convenient to turn on and turn off the temporal modeling, which would be discussed in Sec. 3.3.2. + +Initialization. We use the pre-trained T2I model as a basis for initializing the shared layers between T2I and T2V models. In addition, for the newly added TempSelfAttn layers, we initialize the weights and biases of the output project layers to zero. This ensures that at the beginning of the T2V fine-tuning stage, these layers produce a zero output, effectively functioning as an identity mapping in conjunction with the shortcut connection. + +# 3.3.2 Motion-Free Guidance + +Challenges encountered. We observed a notable phenomenon in the current T2V diffusion models [5, 30] where the per-frame visual quality significantly lags behind that of T2I models [15, 46, 53, 64]. Furthermore, our analysis revealed a remarkable degradation in visual quality in the T2V models post-fine-tuning, especially when compared to their original T2I counterparts. We note that these problems generally exist in current T2V diffusion models, not limited to our transformer-based T2V. + +Problem analysis and insights. We presume that the observed lag in the visual quality of T2V primarily stems from two factors: the nature of video data and the fine-tuning approach. Firstly, publicly available video datasets often fall short in both quality and quantity compared to image datasets. For instance, [56] has more than 2B English image-text pairs, whereas the current widely used video dataset, WebVid-10M [1] contains only 10.7M video-text pairs. Additionally, many video frames are compromised by motion blur and watermarks, further reducing their visual quality. This limited availability hampers the development of robust and versatile video diffusion models. Secondly, the focus on optimizing temporal aspects during video fine-tuning can inadvertently compromise the spatial visual quality, resulting in a decline in the overall quality of the generated videos. + +Solution I: joint image-video training. From the data aspect, we adopt the joint image-video training strategy [9, 16, 30, 62] to mitigate the video data shortages. Furthermore, joint training helps to alleviate the problem of domain discrepancy between video and image datasets by integrating both data types for training. + +Solution II: motion-free guidance. We treat the temporal motion within a video clip as a special conditioning signal, which can be analogized to the textual conditioning in T2I/T2V diffusion models. Based on this analogy, we propose a novel approach, motion-free guidance (MFG), inspired by classifier-free guidance [6, 26], to modulate the weight of motion information in the generated video. + +In a particular training iteration, our approach mirrors the concept used in classifier-free guidance, where conditioned text is replaced with an empty string. The difference is that we employ an identity matrix to nullify the temporal attention with a probability of $p_{\text{motion\_free}}$ . This identity matrix, which is depicted in Figure 3 (Motion-Free Mask), is structured such that its diagonal is populated with ones, while all other positions are zeroes. This configuration confines the temporal self-attention to work within a single + +
ConditioningScaleAttribute BindingObject RelationshipComplexMean
Text EncoderTypeIntegrationColorShapeTextureSpatialNon-spatial
CLIP-L [50]MMadaLN-zeroXL/236.9442.0650.739.4130.3836.4134.32
CLIP-L [50]MMcross-attnXL/273.9151.8168.7619.2631.8041.5247.84
T5-XXL [13]LLMcross-attnXL/274.9055.4070.0520.5231.6841.0148.93
CLIP-T5XXLMM + LLMcross-attnXL/275.6555.7469.4820.6731.7941.4449.13
CLIP-T5XXLMM + LLMcross-attnG/276.7457.0071.5020.9832.0241.6749.99
+ +frame. Furthermore, as introduced in Sec. 3.3.1, temporal self-attention is the sole operator for temporal modeling. Thus, using a motion-free attention mask suffices to disable temporal modeling in the video diffusion process. + +During inference, we have text and motion conditionings. Inspired by [6], we can modify the score estimate as: + +$$ +\begin{array}{l} \tilde {\epsilon} _ {\theta} = \epsilon_ {\theta} (x _ {t}, \varnothing , \varnothing) \\ + \lambda_ {T} \cdot \left(\epsilon_ {\theta} \left(x _ {t}, c _ {T}, c _ {M}\right) - \epsilon_ {\theta} \left(x _ {t}, \varnothing , c _ {M}\right)\right) \\ + \lambda_ {M} \cdot \left(\epsilon_ {\theta} \left(x _ {t}, \varnothing , c _ {M}\right) - \epsilon_ {\theta} \left(x _ {t}, \varnothing , \varnothing\right)\right) \tag {6} \\ \end{array} +$$ + +where $c_{T}$ and $c_{M}$ represent the text conditioning and motion conditioning. $\lambda_{T}$ and $\lambda_{M}$ are the guidance scale of standard text and that of motion, controlling how strongly the generated samples correspond with the text condition and the motion strength, respectively. We empirically found that fixing $\lambda_{T} = 7.5$ and adjusting $\lambda_{M} \in [1.0, 1.3]$ for each example tend to achieve the best result. This finding is similar to [6], although our study utilizes a narrower range for $\lambda_{M}$ . + +Putting solutions together. We can integrate solution I and II together in the following way: when the motion is omitted at a training step, we load an image-text pair and repeat the image $T - 1$ times to create a pseudo video. Conversely, if motion is included, we instead load a video clip and extract it into $T$ frames. + +# 4. Experiments + +# 4.1. Implementation Details + +Training scheme For all GenTron model variations, we employ the AdamW [38] optimizer, maintaining a constant learning rate of $1 \times 10^{-4}$ . We train our T2I GenTron models in a multi-stage procedure [46, 53] with an internal dataset, including a low-resolution $(256 \times 256)$ training with a batch size of 2048 and 500K optimization steps, as well as high-resolution $(512 \times 512)$ with a batch size of 784 and 300K steps. For the GenTron-G/2 model, we further integrate Fully Sharded Data Parallel (FSDP) [67] and activation checkpointing (AC), strategies specifically adopted to + +Table 2. Conditioning and model scale in GenTron. We compare GenTron model variants with different design choices on T2I-CompBench [31]. The text encoders are from the language tower of the multi-modal (MM) model, the large language model (LLM), or a combination of them. The GenTron-G/2 with CLIP-T5XXL performs best. Detailed discussions can be found in Sec. 4.2. + +
ModelAttribute BindingObj. RelationComp.Mean
ColorShapeTextureSpat.Non-spat.
LDM v1.437.6535.7641.5612.4630.7930.8031.50
LDM v250.6542.2149.2213.4230.9633.8636.72
Composable v240.6332.9936.458.0029.8028.9829.47
Structured v249.9042.1849.0013.8631.1133.5536.60
Attn-Exact v264.0045.1759.6314.5531.0934.0141.41
GORS66.0347.8562.8718.1531.9333.2843.35
DALL-E 257.5054.6463.7412.8330.4336.9642.68
LDM XL63.6954.0856.3720.3231.1040.9144.41
PixArt-α68.8655.8270.4420.8231.7941.1748.15
GenTron76.7457.0071.5020.9832.0241.6749.99
+ +Table 3. Comparison of alignment evaluation on T2I-CompBench [31]. Results show our advanced model, GenTron-CLIPT5XXL-G/2, achieves superior performance across multiple compositional metrics compared to previous methods. + +optimize GPU memory usage. In our video experiments, we train videos on a video dataset that comprises approximately 34M videos. To optimize storage usage and enhance data loading efficiency, the videos are pre-processed to a resolution with a short side of 512 pixels and a frame rate of 24 FPS. We process batches of 128 video clips. Each clip comprises 8 frames, captured at a sampling rate of 4 FPS. + +Evaluation metrics. We mainly adopt the recent T2I-CompBench [31] to compare GenTron model variants, following [4, 10]. Specifically, we compare the attribute binding aspects, which include color, shape, and texture. We also compare the spatial and non-spatial object relationships. Moreover, user studies are conducted to compare visual quality and text alignment. + +# 4.2. Main Results of GenTron-T2I + +In this subsection, we discuss our experimental results, focusing on how various conditioning factors and model sizes impact GenTron's performance. Table 2 presents the quantitative findings of our study. Additionally, we offer a comparative visualization, illustrating the effects of each conditioning factor we explored. A comparison to prior art is also provided. + +![](images/a3324278740a7af051c211a79feffb5c48d2b940628ce4edc5f70c590d7e53c8.jpg) +(a)adaLN-Zero + +![](images/4559b4fdbd5e73a5d760c13e1fb6f6c0241445b4dd33bcd5f42faa738400c22d.jpg) +(b) Cross attention + +Cross attention vs. adaLN-Zero. Recent findings [44] conclude that the adaLN design yields superior results in terms of the FID, outperforming both cross-attention and in-context conditioning in efficiency for class-based scenarios. However, our observations reveal a limitation of adaLN in handling free-form text conditioning. This shortcoming is evident in Figure 4, where adaLN's attempt to generate a panda image falls short, with cross-attention demonstrating a clear advantage. This is further verified quantitatively in the first two rows of Table 2, where cross-attention uniformly excels over adaLN in all evaluated metrics. + +This outcome is reasonable considering the nature of class conditioning, which typically involves a limited set of fixed signals (e.g., the 1000 one-hot class embeddings for ImageNet [18]). In such contexts, the adaLN approach, operating at a spatially global level, adjusts the image features uniformly across all positions through the normalization layer, making it adequate for the static signals. In contrast, cross-attention treats spatial positions with more granularity. It differentiates between various spatial locations by dynamically modulating image features based on the cross-attention map between the text embedding and the image features. This spatial-sensitive processing is essential for free-from text conditioning, where the conditioning signals are infinitely diverse and demand detailed representation in line with the specific content of textual descriptions. + +Comparative analysis of text encoders. In Table 2 (rows two to four), we conduct a quantitative evaluation of various text encoders on T2I-CompBench, ensuring a fair comparison by maintaining a consistent XL/2 size across models. Results reveal that GenTron-T5XXL outperforms GenTron-CLIP-L across all three attribute binding and spatial relationship metrics, while it demonstrates comparable performance in the remaining two metrics. This suggests that T5 embeddings are superior in terms of compositional ability. These observations are in line with [2], which utilizes both CLIP and T5 embeddings for training but tests them individually or in combination during inference. Unlike eDiff, + +![](images/b567381917800d47279ed87745988bb05d38e90981403c764900d101ab7009be.jpg) +Figure 4. adaLN-Zero vs. cross attention. The prompt is "A panda standing on a surfboard in the ocean in sunset." Cross attention exhibits a distinct advantage in the text-conditioned scenario. +(a) GenTron-XL/2 +Figure 5. Effect of model scale. The prompt is "a cat reading a newspaper". The larger model GenTron-G/2 excels in rendering finer details and rationalization in the layout of the cat and newspaper. More comparisons can be found in the appendix. + +![](images/ae1c4fcef810c0358b38094c6070e77db9ef7cb1543a602ccdb72fa6e343df07.jpg) +(b) GenTron-G/2 + +
Method#Param.FID-30K↓CLIP-Score↑T2I-CompBen↑
Imagen3.0B7.270.27-
Parti-750M0.8B10.71--
Parti-3B3.0B8.10--
GigaGAN1.0B9.090.322-
MUSE-3B3.0B7.880.320-
SD v1.40.9B12.940.32531.50
SDXL2.6B17.820.32944.41
GenTron-XL/20.9B14.210.32649.13
GenTron-G/23.1B14.530.33549.99
+ +Table 4. Comparison to T2I models. + +our approach maintains the same settings for both training and inference. Notably, GenTron demonstrates enhanced performance when combining CLIP-L and T5XXL embeddings, indicating the model's ability to leverage the distinct advantages of each text embedding type. + +Scaling GenTron up. In Figure 5, we showcase examples from the PartiPrompts benchmark [65] to illustrate the qualitative enhancements achieved by scaling our model up from approximately 900 million to 3 billion parameters. Both models operate under the same CLIP-T5XXL condition. The larger GenTron-G/2 model excels in rendering finer details and more accurate representations, particularly in rationalizing the layout of objects like cats and newspapers. This results in image compositions that are both more coherent and more realistic. In comparison, the smaller GenTron-XL/2 model, while producing recognizable images with similar color schemes, falls short in terms of precision and visual appeal. + +Furthermore, the superiority of GenTron-G/2 is quantitatively affirmed through its performance on T2I-CompBench, as detailed in Table 2. The increase in model size correlates with significant improvements across all evaluative criteria for object composition, including attributes and relationships. Additional comparative examples are provided in the appendix for further illustration. + +![](images/9e76d4ee6a95af33bf39178cf9f107e18420b7cecdf7bb753ce13525ff201ac7.jpg) +Figure 6. GenTron-T2V examples. Prompts are "A giant tortoise is making its way across the beach" and "A dog swimming". + +![](images/b8916973999bebc8e3ed364d550f718a65574faeb7be3595a41fccc7f36f2ae4.jpg) +Figure 7. Visualization of the human preference of our method vs. Latent Diffusion XL. Our method received a significantly higher number of votes as the winner in comparisons of visual quality and text faithfulness, with a total of 3000 answers. + +Comparison to prior work. In Table 3, we showcase the alignment evaluation results from T2I-CompBench. Our method demonstrates outstanding performance in all areas, including attribute binding, object relationships, and complex compositions. This indicates a heightened proficiency in compositional generation, with a notable strength in color binding. In this aspect, our approach surpasses the previous state-of-the-art (SoTA) work [10] by over $7\%$ . + +In Table 4, we further thoroughly compare the zero-shot FID-30K, CLIP-Score, and T2I-CompBench metrics with previous T2I models. In contrast to U-Net based SDv1.4, GenTron uses about four times less data (550M vs. 2B), while achieving better CLIP-scores and T2I-CompBench results. Although GenTron does not achieve a superior FID score compared to SDv1.4, it's important to note that recent studies, including Pick-a-Pic [36] and SDXL [46], have highlighted that FID scores often misrepresent human preferences in generative models, sometimes even negatively correlating with visual aesthetics. + +User study Figure 7 shows the human preference of our method versus SDXL. We used standard prompts in PartiPrompt2 [65] to generate 100 images using both methods and ask people for their preference blindly after shuffling. We received a total of three thousand responses on the comparisons of visual quality and text faithfulness, with our method emerging as the clear winner by a clear margin. + +# 4.3. GenTron-T2V Results + +In Figure 6, We showcase several samples generated by GenTron-T2V, which are not only visually striking but also temporally coherent. This highlights GenTron-T2V's effectiveness in creating videos that are both aesthetically pleasing and consistent over time. + +![](images/ea2c668e91e1c69d5f86bda74e77889ad223d7c571b7e9a9694b7bfefd704206.jpg) +Figure 8. Effect of motion-free guidance. GenTron-T2V with motion-free guidance has a clear visual appearance improvement. The prompt is "A lion standing on a surfboard in the ocean in sunset". + +Effect of motion-free guidance. In Figure 8, we present a comparison between our GenTron variants, with motion-free guidance (MFG) and without. For this comparison, critical factors such as the pre-trained T2I model, training data, and the number of training iterations were kept constant to ensure a fair evaluation. The results clearly indicate that GenTron-T2V, when integrated with MFG, shows a marked tendency to focus on the central object mentioned in the prompt, often rendering it in greater detail. Specifically, the object typically occupies a more prominent, central position in the generated video, thereby dominating the visual focus across video frames. + +# 5. Conclusion + +In this work, we provide a thorough exploration of transformer-based diffusion models for text-conditioned image and video generation. Our findings shed light on the properties of various conditioning approaches and offer compelling evidence of quality improvement when scaling up the model. A notable contribution of our work is the development of GenTron for video generation, where we introduce motion-free guidance. This innovative approach has demonstrably enhanced the visual quality of generated videos. We hope that our research will contribute to bridging the existing gap in applying transformers to diffusion models and their broader use in other domains. + +Acknowledgement. This paper is partially supported by the National Key R&D Program of China No.2022ZD0161000 and the General Research Fund of Hong Kong No.17200622. + +# References + +[1] Max Bain, Arsha Nagrani, Güll Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1728-1738, 2021. 5 +[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 4, 7 +[3] Fan Bao, Shen Nie, Kaiwen Xue, Yue Cao, Chongxuan Li, Hang Su, and Jun Zhu. All are worth words: A vit backbone for diffusion models. In CVPR, 2023. 3 +[4] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions, 2023. 6 +[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 2, 3, 4, 5 +[6] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 5, 6 +[7] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. OpenAI, 2024. 2 +[8] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 2 +[9] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. Videocrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 2, 3, 5 +[10] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 3, 6, 8 +[11] Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish V Thapliyal, James Bradbury, Weicheng Kuo, Mojtaba Seyedhosseini, Chao Jia, Burcu Karagol Ayan, Carlos Riquelme Ruiz, Andreas Peter Steiner, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and + +Radu Soricut. PaLI: A jointly-scaled multilingual language-image model. In The Eleventh International Conference on Learning Representations, 2023. 2 +[12] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 2 +[13] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416, 2022. 2, 4, 6 +[14] Yuren Cong, Mengmeng Xu, christian simon, Shoufa Chen, Jiawei Ren, Yanping Xie, Juan-Manuel Perez-Rua, Bodo Rosenhahn, Tao Xiang, and Sen He. FLATTEN: optical FLOW-guided ATTENTION for consistent text-to-video editing. In The Twelfth International Conference on Learning Representations, 2024. 2 +[15] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xiaofang Wang, Abhimanyu Dubey, et al. Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807, 2023. 2, 4, 5 +[16] Yatin Dandi, Aniket Das, Soumye Singhal, Vinay Nambodiri, and Piyush Rai. Jointly trained image and video generation using residual vectors. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3028-3042, 2020. 2, 5 +[17] Mostafa Dehghani, Josip Djolonga, Basil Mustafa, Piotr Padlewski, Jonathan Heek, Justin Gilmer, Andreas Peter Steiner, Mathilde Caron, Robert Geirhos, Ibrahim Alabdulmohsin, et al. Scaling vision transformers to 22 billion parameters. In International Conference on Machine Learning, pages 7480-7512. PMLR, 2023. 2 +[18] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 7 +[19] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers), pages 4171-4186. Association for Computational Linguistics, 2019. 2 +[20] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 2, 3 +[21] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Pro + +ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 3 +[22] Shanghua Gao, Pan Zhou, Ming-Ming Cheng, and Shuicheng Yan. Masked diffusion transformer is a strong image synthesizer. arXiv preprint arXiv:2303.14389, 2023. 3 +[23] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 4 +[24] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 3 +[25] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2022. 3 +[26] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 2, 5 +[27] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 3 +[28] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2 +[29] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. The Journal of Machine Learning Research, 23(1):2249-2281, 2022. 4 +[30] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video diffusion models, 2022. 2, 4, 5 +[31] Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 2, 6 +[32] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 4 +[33] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6007-6017, 2023. 2 +[34] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439, 2023. 2 +[35] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 2, 3 + +[36] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Matiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36, 2024. 8 +[37] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2 +[38] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6 +[39] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 3 +[40] Haoyu Lu, Guoxing Yang, Nanyi Fei, Yuqi Huo, Zhiwu Lu, Ping Luo, and Mingyu Ding. Vdt: An empirical study on video diffusion with transformers. arXiv preprint arXiv:2305.13311, 2023. 3 +[41] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Tieniu Tan. Videofusion: Decomposed diffusion models for high-quality video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10209-10218, 2023. 2 +[42] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2 +[43] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob Mcgrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In Proceedings of the 39th International Conference on Machine Learning, pages 16784-16804. PMLR, 2022. 4 +[44] William Peebles and Saining Xie. Scalable diffusion models with transformers. arXiv preprint arXiv:2212.09748, 2022. 2,3,4,7 +[45] Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. In Proceedings of the AAAI conference on artificial intelligence, 2018. 2, 3 +[46] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 4, 5, 6, 8 +[47] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. arXiv preprint arXiv:2303.09535, 2023. 2 +[48] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. OpenAI blog, 2018. 2 + +[49] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2 +[50] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 4, 6 +[51] Carlos Riquelme, Joan Puigcerver, Basil Mustafa, Maxim Neumann, Rodolphe Jenatton, André Susano Pinto, Daniel Keysers, and Neil Houlsby. Scaling vision with sparse mixture of experts. Advances in Neural Information Processing Systems, 34:8583-8595, 2021. 2 +[52] Alex Rogozhnikov. Einops: Clear and reliable tensor manipulations with einstein-like notation. In International Conference on Learning Representations, 2021. 5 +[53] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3, 5, 6 +[54] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 2, 3 +[55] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 2 +[56] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 5 +[57] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-video: Text-to-video generation without text-video data. In The Eleventh International Conference on Learning Representations, 2023. 2 +[58] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3 +[59] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2 +[60] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2, 4 + +[61] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 3 +[62] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 2, 5 +[63] Yilin Wang, Zeyuan Chen, Liangjun Zhong, Zheng Ding, Zhizhou Sha, and Zhuowen Tu. Dolfin: Diffusion layout transformers without autoencoder. arXiv preprint arXiv:2310.16305, 2023. 3 +[64] Zeyue Xue, Guanglu Song, Qiushan Guo, Boxiao Liu, Zhuofan Zong, Yu Liu, and Ping Luo. RAPHAEL: Text-to-image generation via large mixture of diffusion paths. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 5 +[65] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. Transactions on Machine Learning Research, 2022. Featured Certification. 7, 8 +[66] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12104-12113, 2022. 2, 4 +[67] Yanli Zhao, Andrew Gu, Rohan Varma, Liang Luo, Chien-Chin Huang, Min Xu, Less Wright, Hamid Shojanazeri, Myle Ott, Sam Shleifer, et al. Pytorch fsdp: experiences on scaling fully sharded data parallel. arXiv preprint arXiv:2304.11277, 2023. 6 +[68] Hongkai Zheng, Weili Nie, Arash Vahdat, and Anima Anandkumar. Fast training of diffusion models with masked transformers. arXiv preprint arXiv:2306.09305, 2023. 3 +[69] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 3, 4 \ No newline at end of file diff --git a/2024/GenTron_ Diffusion Transformers for Image and Video Generation/images.zip b/2024/GenTron_ Diffusion Transformers for Image and Video Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7d8e0bd293ea8c280ebd541b5358b5d42d6415d4 --- /dev/null +++ b/2024/GenTron_ Diffusion Transformers for Image and Video Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fc87652770e079671f13672eeba99c77cd42b99a461947fc480e21a9f853c112 +size 627929 diff --git a/2024/GenTron_ Diffusion Transformers for Image and Video Generation/layout.json b/2024/GenTron_ Diffusion Transformers for Image and Video Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7cd460640ab36cadc51cadd08b03db753d53ee9b --- /dev/null +++ b/2024/GenTron_ Diffusion Transformers for Image and Video Generation/layout.json @@ -0,0 +1,8328 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 89, + 103, + 503, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 103, + 503, + 120 + ], + "spans": [ + { + "bbox": [ + 89, + 103, + 503, + 120 + ], + "type": "text", + "content": "GenTron: Diffusion Transformers for Image and Video Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 82, + 142, + 510, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 142, + 510, + 185 + ], + "spans": [ + { + "bbox": [ + 82, + 142, + 510, + 185 + ], + "type": "text", + "content": "Shoufa Chen" + }, + { + "bbox": [ + 82, + 142, + 510, + 185 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 82, + 142, + 510, + 185 + ], + "type": "text", + "content": " Mengmeng Xu" + }, + { + "bbox": [ + 82, + 142, + 510, + 185 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 82, + 142, + 510, + 185 + ], + "type": "text", + "content": " Jiawei Ren" + }, + { + "bbox": [ + 82, + 142, + 510, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 82, + 142, + 510, + 185 + ], + "type": "text", + "content": " Yuren Cong" + }, + { + "bbox": [ + 82, + 142, + 510, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 82, + 142, + 510, + 185 + ], + "type": "text", + "content": " Sen He" + }, + { + "bbox": [ + 82, + 142, + 510, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 82, + 142, + 510, + 185 + ], + "type": "text", + "content": " Yanping Xie" + }, + { + "bbox": [ + 82, + 142, + 510, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 82, + 142, + 510, + 185 + ], + "type": "text", + "content": " Animesh Sinha" + }, + { + "bbox": [ + 82, + 142, + 510, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 82, + 142, + 510, + 185 + ], + "type": "text", + "content": " Ping Luo" + }, + { + "bbox": [ + 82, + 142, + 510, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 82, + 142, + 510, + 185 + ], + "type": "text", + "content": " Tao Xiang" + }, + { + "bbox": [ + 82, + 142, + 510, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 82, + 142, + 510, + 185 + ], + "type": "text", + "content": " Juan-Manuel Perez-Rua" + }, + { + "bbox": [ + 82, + 142, + 510, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 82, + 142, + 510, + 185 + ], + "type": "text", + "content": " The University of Hong Kong " + }, + { + "bbox": [ + 82, + 142, + 510, + 185 + ], + "type": "inline_equation", + "content": "{}^{2}" + }, + { + "bbox": [ + 82, + 142, + 510, + 185 + ], + "type": "text", + "content": "Meta" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 64, + 198, + 220, + 353 + ], + "blocks": [ + { + "bbox": [ + 64, + 198, + 220, + 353 + ], + "lines": [ + { + "bbox": [ + 64, + 198, + 220, + 353 + ], + "spans": [ + { + "bbox": [ + 64, + 198, + 220, + 353 + ], + "type": "image", + "image_path": "e449d602918cf754384afa929d4d33a7802872219b2287935e6d1f3ca0b15005.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 220, + 198, + 375, + 354 + ], + "blocks": [ + { + "bbox": [ + 220, + 198, + 375, + 354 + ], + "lines": [ + { + "bbox": [ + 220, + 198, + 375, + 354 + ], + "spans": [ + { + "bbox": [ + 220, + 198, + 375, + 354 + ], + "type": "image", + "image_path": "39dacd659364b26924a31c15f7b7fc3868d0030c53794ccbee361706bf98cc51.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 375, + 198, + 529, + 353 + ], + "blocks": [ + { + "bbox": [ + 375, + 198, + 529, + 353 + ], + "lines": [ + { + "bbox": [ + 375, + 198, + 529, + 353 + ], + "spans": [ + { + "bbox": [ + 375, + 198, + 529, + 353 + ], + "type": "image", + "image_path": "73f12c5e67ca1fa52f149fb37427e7d7bea9fc8b528384069d13341f24df0b21.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 63, + 354, + 297, + 422 + ], + "blocks": [ + { + "bbox": [ + 63, + 354, + 297, + 422 + ], + "lines": [ + { + "bbox": [ + 63, + 354, + 297, + 422 + ], + "spans": [ + { + "bbox": [ + 63, + 354, + 297, + 422 + ], + "type": "image", + "image_path": "b5ef7783ec1b8f511faba7baa8c3cf25d763f8b7b17ca461816b78352feeb43f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 297, + 354, + 529, + 421 + ], + "blocks": [ + { + "bbox": [ + 297, + 354, + 529, + 421 + ], + "lines": [ + { + "bbox": [ + 297, + 354, + 529, + 421 + ], + "spans": [ + { + "bbox": [ + 297, + 354, + 529, + 421 + ], + "type": "image", + "image_path": "a98345756a35f127066decf21da3bccba3d4704694f70d03c926322979c23747.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 63, + 422, + 297, + 491 + ], + "blocks": [ + { + "bbox": [ + 63, + 422, + 297, + 491 + ], + "lines": [ + { + "bbox": [ + 63, + 422, + 297, + 491 + ], + "spans": [ + { + "bbox": [ + 63, + 422, + 297, + 491 + ], + "type": "image", + "image_path": "a5baec7215de49c1a96add6f7dd16cf2926df3cdb676c96a97e82ca87ad754b4.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 96, + 508, + 494, + 520 + ], + "lines": [ + { + "bbox": [ + 96, + 508, + 494, + 520 + ], + "spans": [ + { + "bbox": [ + 96, + 508, + 494, + 520 + ], + "type": "text", + "content": "Figure 1. GenTron: Transformer based diffusion model for high-quality text-to-image/video generation." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 297, + 422, + 529, + 491 + ], + "blocks": [ + { + "bbox": [ + 297, + 422, + 529, + 491 + ], + "lines": [ + { + "bbox": [ + 297, + 422, + 529, + 491 + ], + "spans": [ + { + "bbox": [ + 297, + 422, + 529, + 491 + ], + "type": "image", + "image_path": "a74bbb5fe69b6e7559bc86c5951149c6d0403ac223ce685f030759ab27622002.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 143, + 530, + 190, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 530, + 190, + 542 + ], + "spans": [ + { + "bbox": [ + 143, + 530, + 190, + 542 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 45, + 558, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 558, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 45, + 558, + 288, + 713 + ], + "type": "text", + "content": "In this study, we explore Transformer-based diffusion models for image and video generation. Despite the dominance of Transformer architectures in various fields due to their flexibility and scalability, the visual generative domain primarily utilizes CNN-based U-Net architectures, particularly in diffusion-based models. We introduce GenTron, a family of Generative models employing Transformer-based diffusion, to address this gap. Our initial step was to adapt Diffusion Transformers (DiTs) from class to text conditioning, a process involving thorough empirical exploration of the conditioning mechanism. We then scale GenTron from approximately 900M to over 3B parameters, observing improvements in visual quality. Furthermore, we extend" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 531, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 531, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 545, + 662 + ], + "type": "text", + "content": "GenTron to text-to-video generation, incorporating novel motion-free guidance to enhance video quality. In human evaluations against SDXL, GenTron achieves a " + }, + { + "bbox": [ + 304, + 531, + 545, + 662 + ], + "type": "inline_equation", + "content": "51.1\\%" + }, + { + "bbox": [ + 304, + 531, + 545, + 662 + ], + "type": "text", + "content": " win rate in visual quality (with a " + }, + { + "bbox": [ + 304, + 531, + 545, + 662 + ], + "type": "inline_equation", + "content": "19.8\\%" + }, + { + "bbox": [ + 304, + 531, + 545, + 662 + ], + "type": "text", + "content": " draw rate), and a " + }, + { + "bbox": [ + 304, + 531, + 545, + 662 + ], + "type": "inline_equation", + "content": "42.3\\%" + }, + { + "bbox": [ + 304, + 531, + 545, + 662 + ], + "type": "text", + "content": " win rate in text alignment (with a " + }, + { + "bbox": [ + 304, + 531, + 545, + 662 + ], + "type": "inline_equation", + "content": "42.9\\%" + }, + { + "bbox": [ + 304, + 531, + 545, + 662 + ], + "type": "text", + "content": " draw rate). GenTron notably performs well in T2I-CompBench, highlighting its compositional generation ability. We hope GenTron could provide meaningful insights and serve as a valuable reference for future research. Please refer to the website1 and the arXiv version for the most up-to-date results: https://arxiv.org/abs/2312.04557." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 320, + 693, + 532, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 693, + 532, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 693, + 532, + 713 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 320, + 693, + 532, + 713 + ], + "type": "text", + "content": "https://www.shoufachen.com/gentron網站/ \n*Equal contribution." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "6441" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 127, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 127, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 127, + 83 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 287, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 287, + 281 + ], + "type": "text", + "content": "Diffusion models have recently shown remarkable progress in content creation, impacting areas such as image generation [27, 53, 55], video generation [5, 7, 28, 57], and editing [14, 33]. Among these notable developments, the CNN-based U-Net architecture has emerged as the predominant backbone design, a choice that stands in contrast to the prevailing trend in natural language processing [8, 19, 59] and computer visual perception [20, 37, 66] domains, where attention-based transformer architectures [60] have empowered a renaissance and become increasingly dominant. To provide a comprehensive understanding of Transformers in diffusion generation and to bridge the gap in architectural choices between visual generation and the other two domains — visual perception and NLP — a thorough investigation of visual generation using Transformers is of substantial scientific value." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 282, + 288, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 282, + 288, + 510 + ], + "spans": [ + { + "bbox": [ + 47, + 282, + 288, + 510 + ], + "type": "text", + "content": "We focus on diffusion models with Transformers in this work. Specifically, our starting point is the foundational work known as DiT [44], which introduced a class-conditioned latent diffusion model that employs a Transformer to replace the traditionally used U-Net architecture. We overcome the limitation of the original DiT model, which is constrained to handling only a restricted number (e.g., 1000) of predefined classes, by utilizing language embeddings derived from open-world, free-form text captions instead of predefined one-hot class embeddings. Along the way, we comprehensively investigate conditioning strategies, including (1) conditioning architectures: adaptive layer norm (adaLN) [45] vs. cross-attention [60]; and (2) text encoding methods: a generic large language model [13] vs. the language tower of multimodal models [50], or the combination of both of them. We additionally carry out comparative experiments and offer detailed empirical analyses to evaluate the effectiveness of these conditioning strategies." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 510, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 510, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 510, + 288, + 714 + ], + "type": "text", + "content": "Next, we explore the scaling-up properties of GenTron. The Transformer architectures have been demonstrated to possess significant scalability in both visual perception [11, 17, 51, 66] and language [8, 19, 48, 49, 59] tasks. For example, the largest dense language model has 540B parameters [12], and the largest vision model has 22B [17] parameters. In contrast, the largest diffusion transformer, DiT-XL [44], only has about 675M parameters, trailed far behind both the Transformers utilized in other domains (e.g., NLP) and recent diffusion arts with convolutional U-Net architectures [15, 46]. To compensate for this considerable lagging, we scale up GenTron in two dimensions, the number of transformer blocks and hidden dimension size, following the scaling strategy in [66]. As a result, our largest model, GenTron-G/2, has more than 3B parameters and achieves significant visual quality improvement compared with the smaller one." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 239 + ], + "type": "text", + "content": "Furthermore, we have advanced GenTron from a T2I to a T2V model by inserting a temporal self-attention layer into each transformer block, making the first attempt to use transformers as the exclusive building block for video diffusion models. We also discuss existing challenges in video generation and introduce our solution, the motion-free guidance (MFG). Specifically, This approach involves intermittently disabling motion modeling during training by setting the temporal self-attention mask to an identity matrix. Besides, MFG seamlessly integrates with the joint image-video strategy [9, 16, 30, 62], where images are used as training samples whenever motion is deactivated. Our experiments indicate that this approach clearly improves the visual quality of generated videos." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 239, + 545, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 239, + 545, + 359 + ], + "spans": [ + { + "bbox": [ + 304, + 239, + 545, + 359 + ], + "type": "text", + "content": "In human evaluations, GenTron outperforms SDXL, achieving a " + }, + { + "bbox": [ + 304, + 239, + 545, + 359 + ], + "type": "inline_equation", + "content": "51.1\\%" + }, + { + "bbox": [ + 304, + 239, + 545, + 359 + ], + "type": "text", + "content": " win rate in visual quality (with a " + }, + { + "bbox": [ + 304, + 239, + 545, + 359 + ], + "type": "inline_equation", + "content": "19.8\\%" + }, + { + "bbox": [ + 304, + 239, + 545, + 359 + ], + "type": "text", + "content": " draw rate), and a " + }, + { + "bbox": [ + 304, + 239, + 545, + 359 + ], + "type": "inline_equation", + "content": "42.3\\%" + }, + { + "bbox": [ + 304, + 239, + 545, + 359 + ], + "type": "text", + "content": " win rate in text alignment (with a " + }, + { + "bbox": [ + 304, + 239, + 545, + 359 + ], + "type": "inline_equation", + "content": "42.9\\%" + }, + { + "bbox": [ + 304, + 239, + 545, + 359 + ], + "type": "text", + "content": " draw rate). Furthermore, when compared to previous studies, particularly as benchmarked against T2I-CompBench [31]-a comprehensive framework for evaluating open-world compositional T2I generation-GenTron demonstrates superior performance across various criteria. These include attribute binding, object relationships, and handling of complex compositions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 360, + 545, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 360, + 545, + 515 + ], + "spans": [ + { + "bbox": [ + 304, + 360, + 545, + 515 + ], + "type": "text", + "content": "Our contributions are summarized as follows: (1) We have conducted a thorough and systematic investigation of transformer-based T2I generation with diffusion models. This study encompasses various conditioning choices and aspects of model scaling. (2) In a pioneering effort, we explore a purely transformer-based diffusion model for T2V generation. We introduce motion-free guidance, an innovative technique that efficiently fine-tunes T2I generation models for producing high-quality videos. (3) Experimental results indicate a clear preference for GenTron over SDXL in human evaluations. Furthermore, GenTron demonstrates superior performance compared to existing methods in the T2I-CompBench evaluations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 526, + 392, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 526, + 392, + 538 + ], + "spans": [ + { + "bbox": [ + 306, + 526, + 392, + 538 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": "Diffusion models for T2I and T2V generation. Diffusion models [27, 42] are a type of generative model that creates data samples from random noise. Later, latent diffusion models [44, 46, 53] are proposed for efficient T2I generation. These designs usually have " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": " a pre-trained Variational Autoencoder [35] that maps images to a compact latent space, 2) a conditioner modeled by cross-attention [28, 53] to process text as conditions with a strength control [26], and 3) a backbone network, U-Net [54] in particular, to process image features. The success of diffusion on T2I generation tasks underscores the promising potential for text-to-video (T2V) generation [34, 41, 47, 57]. VDM [30] and Imagen Video [28] extend the image diffusion architecture on the temporal dimension with promising initial results. To" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6442" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "content": "avoid excessive computing demands, video latent diffusion models [5, 9, 25, 40, 61, 69] implement the video diffusion process in a low-dimensional latent space." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 125, + 289, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 125, + 289, + 355 + ], + "spans": [ + { + "bbox": [ + 46, + 125, + 289, + 355 + ], + "type": "text", + "content": "Transformer-based Diffusion. Recently, Transformer-based Diffusion models have attracted increasing research interest. Among these, U-ViT [3] treats all inputs as tokens by integrating transformer blocks with a U-net architecture. In contrast, DiT [44] employs a simpler, non-hierarchical transformer structure. MDT [22] and MaskDiT [68] enhance DiT's training efficiency by incorporating the mask strategy [24]. Dolfin [63] is a transformer-based model for layout generation. Concurrently to this work, PixArt-" + }, + { + "bbox": [ + 46, + 125, + 289, + 355 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 125, + 289, + 355 + ], + "type": "text", + "content": " [10] demonstrates promising outcomes in Transformer-based T2I diffusion. It's trained using a three-stage decomposition process with high-quality data. Our work diverges from PixArt-" + }, + { + "bbox": [ + 46, + 125, + 289, + 355 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 125, + 289, + 355 + ], + "type": "text", + "content": " in key aspects. Firstly, while PixArt-" + }, + { + "bbox": [ + 46, + 125, + 289, + 355 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 125, + 289, + 355 + ], + "type": "text", + "content": " emphasizes training efficiency, our focus is on the design choice of conditioning strategy and scalability in T2I Transformer diffusion models. Secondly, we extend our exploration beyond image generation to video diffusion. We propose an innovative approach in video domain, which is not covered by PixArt-" + }, + { + "bbox": [ + 46, + 125, + 289, + 355 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 125, + 289, + 355 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 365, + 103, + 378 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 365, + 103, + 378 + ], + "spans": [ + { + "bbox": [ + 47, + 365, + 103, + 378 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 386, + 287, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 386, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 386, + 287, + 471 + ], + "type": "text", + "content": "We first introduce the preliminaries in Section 3.1, and then present the details of GenTron for text-to-image generation in Section 3.2, which includes text encoder models, embedding integration methods, and scaling up strategy of GenTron. Lastly, in Section 3.3, we extend GenTron's application to video generation, building on top of the T2I foundations laid in previous sections." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 479, + 134, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 134, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 134, + 491 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 498, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 594 + ], + "type": "text", + "content": "Diffusion models. Diffusion models [27] have emerged as a family of generative models that generate data by performing a series of transformations on random noise. They are characterized by a forward and a backward process. Given an instance from the data distribution " + }, + { + "bbox": [ + 46, + 498, + 287, + 594 + ], + "type": "inline_equation", + "content": "x_0 \\sim p(x_0)" + }, + { + "bbox": [ + 46, + 498, + 287, + 594 + ], + "type": "text", + "content": ", random Gaussian noise is iteratively added to the instance in the forward noisig process to create a Markov Chain of random latent variable " + }, + { + "bbox": [ + 46, + 498, + 287, + 594 + ], + "type": "inline_equation", + "content": "x_1, x_2, \\ldots, x_T" + }, + { + "bbox": [ + 46, + 498, + 287, + 594 + ], + "type": "text", + "content": " following:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 82, + 605, + 287, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 605, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 82, + 605, + 287, + 620 + ], + "type": "interline_equation", + "content": "q \\left(x _ {t} \\mid x _ {t - 1}\\right) = \\mathcal {N} \\left(x _ {t}; \\sqrt {1 - \\beta_ {i}} x _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right), \\tag {1}", + "image_path": "b6797dca507344619062418a0d6828c87043654bdf95676e08cdee32e8cdcf80.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\beta_{1},\\ldots ,\\beta_{T}" + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "content": " are hyperparameters corresponding to the noise schedule. After a large enough number of diffusion steps, " + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "content": " can be viewed as a standard Gaussian noise. A denoising network " + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "content": " is further trained to learn the backward process, i.e., how to remove the noise from a noisy input [27]. For inference, an instance can be sampled starting from a random Gaussian noise " + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "inline_equation", + "content": "x_{T}\\sim \\mathcal{N}(0;\\mathbf{I})" + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "content": " and denoised" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 320, + 70, + 419, + 284 + ], + "blocks": [ + { + "bbox": [ + 320, + 70, + 419, + 284 + ], + "lines": [ + { + "bbox": [ + 320, + 70, + 419, + 284 + ], + "spans": [ + { + "bbox": [ + 320, + 70, + 419, + 284 + ], + "type": "image", + "image_path": "6f9293ac2ca6ee4a494efe39f6e5459242bbd1cc1d489750c3d4f6e0a26422f0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 342, + 285, + 394, + 295 + ], + "lines": [ + { + "bbox": [ + 342, + 285, + 394, + 295 + ], + "spans": [ + { + "bbox": [ + 342, + 285, + 394, + 295 + ], + "type": "text", + "content": "(a)adaLN-Zero" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 300, + 545, + 357 + ], + "lines": [ + { + "bbox": [ + 305, + 300, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 305, + 300, + 545, + 357 + ], + "type": "text", + "content": "Figure 2. Text embedding integration architecture. We directly adapt adaLN from DiT [44], substituting the one-hot class embedding with text embedding. For cross attention, different from the approach in [44], we maintain the use of adaLN to model the combination of time embedding and the aggregated text embedding." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 432, + 70, + 531, + 283 + ], + "blocks": [ + { + "bbox": [ + 432, + 70, + 531, + 283 + ], + "lines": [ + { + "bbox": [ + 432, + 70, + 531, + 283 + ], + "spans": [ + { + "bbox": [ + 432, + 70, + 531, + 283 + ], + "type": "image", + "image_path": "bfda24d2d5047514ccd5cfd22f5ccb8b82a8b15d1c13e105f1e2b97f47dc2ace.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 451, + 285, + 512, + 294 + ], + "lines": [ + { + "bbox": [ + 451, + 285, + 512, + 294 + ], + "spans": [ + { + "bbox": [ + 451, + 285, + 512, + 294 + ], + "type": "text", + "content": "(b) Cross attention" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 372, + 544, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 372, + 544, + 397 + ], + "spans": [ + { + "bbox": [ + 305, + 372, + 544, + 397 + ], + "type": "text", + "content": "step-by-step following the Markov Chain, i.e., by sequentially sampling " + }, + { + "bbox": [ + 305, + 372, + 544, + 397 + ], + "type": "inline_equation", + "content": "x_{t-1}" + }, + { + "bbox": [ + 305, + 372, + 544, + 397 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 305, + 372, + 544, + 397 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 305, + 372, + 544, + 397 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 305, + 372, + 544, + 397 + ], + "type": "inline_equation", + "content": "p_\\theta(x_{t-1}|x_t)" + }, + { + "bbox": [ + 305, + 372, + 544, + 397 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 332, + 405, + 545, + 431 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 405, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 332, + 405, + 545, + 431 + ], + "type": "interline_equation", + "content": "x _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(x _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta} (x _ {t}, t)\\right) + \\sigma_ {t} \\mathbf {z}, \\tag {2}", + "image_path": "ed85c177a5803e2fd2d6d8d78b81df3943b0979aebf0b06dc5526d7ffdff7dac.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 441, + 545, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 441, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 305, + 441, + 545, + 491 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 441, + 545, + 491 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t = \\prod_{s=1}^t \\alpha_s" + }, + { + "bbox": [ + 305, + 441, + 545, + 491 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 441, + 545, + 491 + ], + "type": "inline_equation", + "content": "\\alpha_t = 1 - \\beta_t" + }, + { + "bbox": [ + 305, + 441, + 545, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 441, + 545, + 491 + ], + "type": "inline_equation", + "content": "\\sigma_t" + }, + { + "bbox": [ + 305, + 441, + 545, + 491 + ], + "type": "text", + "content": " is the noise scale. In practical application, the diffusion sampling process can be further accelerated using different sampling techniques [39, 58]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 506, + 545, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 545, + 650 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 545, + 650 + ], + "type": "text", + "content": "Latent diffusion model architectures. Latent diffusion models (LDMs) [53] reduce the high computational cost by conducting the diffusion process in the latent space. First, a pre-trained autoencoder [21, 35] is utilized to compress the raw image from pixel to latent space, then the diffusion models, which are commonly implemented with a U-Net [54] backbone, work on the latent space. Peebles et al. proposed DiT [44] to leverage the transformer architecture as an alternative to the traditional U-Net backbone for class-conditioned image generation, adopting the adaptive layernorm (adaLN [45]) for class conditioning mechanism, as shown in Figure 2a." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 658, + 440, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 440, + 671 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 440, + 671 + ], + "type": "text", + "content": "3.2. Text-to-Image GenTron" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "text", + "content": "Our GenTron is built upon the DiT-XL/2 [44], which converts the latent of shape " + }, + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "32 \\times 32 \\times 4" + }, + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "text", + "content": " to a sequence of non-overlapping tokens with a " + }, + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "text", + "content": " patchify layer [20]. Then," + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6443" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "these tokens are sent into a series of transformer blocks. Finally, a standard linear decoder is applied to convert these image tokens into latent space." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 109, + 288, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 109, + 288, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 109, + 288, + 264 + ], + "type": "text", + "content": "While DiT has shown that transformer-based models yield promising results in class-conditioned scenarios, it did not explore the realm of T2I generation. This field poses a considerable challenge, given its less constrained conditioning format. Moreover, even the largest DiT model, DiT-XL/2, with its 675 million parameters, is significantly overshadowed by current U-Nets [15, 46], which boast over 3 billion parameters. To address these limitations, our research conducts a thorough investigation of transformer-based T2I diffusion models, focusing specifically on text conditioning approaches and assessing the scalability of the transformer architecture by expanding GenTron to more than 3 billion parameters." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 280, + 205, + 291 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 280, + 205, + 291 + ], + "spans": [ + { + "bbox": [ + 47, + 280, + 205, + 291 + ], + "type": "text", + "content": "3.2.1 From Class to Text Condition" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 299, + 287, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 299, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 299, + 287, + 407 + ], + "type": "text", + "content": "T2I diffusion models rely on textual inputs to steer the process of image generation. The mechanism of text conditioning involves two critical components: firstly, the selection of a text encoder, which is responsible for converting raw text into text embeddings, and secondly, the method of integrating these embeddings into the diffusion process. For a complete understanding, we have included in the appendix a detailed presentation of the decisions made in existing works concerning these two components." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 422, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 287, + 567 + ], + "type": "text", + "content": "Text encoder model. Current advancements in T2I diffusion techniques employ a variety of language models, each with its unique strengths and limitations. To thoroughly assess which model best complements transformer-based diffusion methods, we have integrated several models into GenTron. This includes the text towers from multimodal models, CLIP [50], as well as a pure large language model, Flan-T5 [13]. Our approach explores the effectiveness of these language models by integrating each model independently with GenTron to evaluate their individual performance and combinations of them to assess the potential properties they may offer when used together." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": "**Embedding integration.** In our study, we focused on two methods of embedding integration: adaptive layernorm and cross-attention. (1) Adaptive layernorm (adaLN). As shown in Figure 2a, this method integrates conditioning embeddings as normalization parameters on the feature channel. Widely used in conditional generative modeling, such as in StyleGAN [32], adaLN serves as the standard approach in DiT [44] for managing class conditions. (2) Cross-attention. As illustrated in Figure 2b, the image feature acts as the query, with textual embedding serving as key and value. This setup allows for direct interaction" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 307, + 70, + 542, + 105 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 542, + 105 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 542, + 105 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 542, + 105 + ], + "type": "table", + "html": "
ModelDepthWidthMLP Width#Param.
GenTron-XL/22811524608930.0M
GenTron-G/248166466563083.8M
", + "image_path": "1f26b43e6452aa6ecb59324495645ea82e4cec0876a45aca968ef41abe52f1ab.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 329, + 112, + 523, + 123 + ], + "lines": [ + { + "bbox": [ + 329, + 112, + 523, + 123 + ], + "spans": [ + { + "bbox": [ + 329, + 112, + 523, + 123 + ], + "type": "text", + "content": "Table 1. Configuration details of GenTron models." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 139, + 545, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 139, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 304, + 139, + 545, + 270 + ], + "type": "text", + "content": "between the image feature and textual embedding through an attention mechanism [60]. Besides, different from the cross-attention discussed in [44], which processes the class embedding and time embedding together by firstly concatenating them, we maintain the use of adaLN in conjunction with the cross-attention to separately model the time embedding. The underlying rationale for this design is our belief that the time embedding, which is consistent across all spatial positions, benefits from the global modulation capabilities of adaLN. Moreover, we also add the pooled text embeddings to the time embedding followings [2, 29, 43, 46]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 285, + 426, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 285, + 426, + 298 + ], + "spans": [ + { + "bbox": [ + 306, + 285, + 426, + 298 + ], + "type": "text", + "content": "3.2.2 Scaling Up GenTron" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 304, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 545, + 449 + ], + "type": "text", + "content": "To explore the impact of substantially scaling up the model size, we have developed an advanced version of GenTron, which we refer to as GenTron-G/2. This model was constructed in accordance with the scaling principles outlined in [66]. We focused on expanding three critical aspects: the number of transformer blocks (depth), the dimensionality of patch embeddings (width), and the hidden dimension of the MLP (MLP-width). The specifications and configurations of the GenTron models are detailed in Table 1. Significantly, the GenTron-G/2 model boasts over 3 billion parameters. To our knowledge, this represents the largest transformer-based diffusion architecture developed to date." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 456, + 437, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 456, + 437, + 468 + ], + "spans": [ + { + "bbox": [ + 306, + 456, + 437, + 468 + ], + "type": "text", + "content": "3.3. Text-to-Video GenTron" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 475, + 545, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 475, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 545, + 559 + ], + "type": "text", + "content": "In this subsection, we elaborate on the process of adapting GenTron from a T2I framework to a T2V framework. Sec. 3.3.1 will detail the modifications made to the model's architecture, enabling GenTron to process video data. Furthermore, Sec. 3.3.2 will discuss the challenges encountered in the domain of video generation and the innovative solutions we have proposed to address them." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 574, + 456, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 456, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 456, + 586 + ], + "type": "text", + "content": "3.3.1 GenTron-T2V Architecture" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "Transformer block with temporal self-attention. It is typically a common practice to train video diffusion models from image diffusion models by adding new temporal modeling modules [5, 23, 30, 69]. These usually consist of 3D convolutional layers and temporal transformer blocks that focus on calculating attention along the temporal dimension. In contrast to the traditional approach [5], which involves adding both temporal convolution layers and temporal transformer blocks to the T2I U-Net, our method integrates only lightweight temporal" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6444" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 83, + 72, + 248, + 233 + ], + "blocks": [ + { + "bbox": [ + 83, + 72, + 248, + 233 + ], + "lines": [ + { + "bbox": [ + 83, + 72, + 248, + 233 + ], + "spans": [ + { + "bbox": [ + 83, + 72, + 248, + 233 + ], + "type": "image", + "image_path": "d533bf56cb27eae85c1f8aa3ff697ebd95e0b758f57f208edc53ad3567e2c78c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 241, + 289, + 308 + ], + "lines": [ + { + "bbox": [ + 46, + 241, + 289, + 308 + ], + "spans": [ + { + "bbox": [ + 46, + 241, + 289, + 308 + ], + "type": "text", + "content": "Figure 3. GenTron Transformer block with TempSelfAttn and Motion-Free Mask. The temporal self-attention layer is inserted between the cross-attention and the MLPs. The motion-free mask, which is an identity matrix, will be utilized in the TempSelfAttn with a probability of " + }, + { + "bbox": [ + 46, + 241, + 289, + 308 + ], + "type": "inline_equation", + "content": "p_{\\text{motion\\_free}}" + }, + { + "bbox": [ + 46, + 241, + 289, + 308 + ], + "type": "text", + "content": ". We omit details like text conditioning, LN here for simplicity, which could be found in Figure 2." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 326, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 326, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 326, + 287, + 422 + ], + "type": "text", + "content": "self-attention (TempSelfAttn) layers into each transformer block. As depicted in Figure 3, the TempSelfAttn layer is placed right after the cross-attention layer and before the MLP layer. Additionally, we modify the output of the cross-attention layer by reshaping it before it enters the TempSelfAttn layer and then reshape it back to its original format once it has passed through. This process can be formally represented as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 51, + 436, + 286, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 436, + 286, + 449 + ], + "spans": [ + { + "bbox": [ + 51, + 436, + 286, + 449 + ], + "type": "interline_equation", + "content": "\\mathbf {x} = \\text {r e a r r a n g e} (\\mathbf {x}, (\\mathrm {b} \\quad \\mathrm {t}) \\quad \\mathrm {n} \\quad \\mathrm {d} \\rightarrow (\\mathrm {b} \\quad \\mathrm {n}) \\quad \\mathrm {t} \\quad \\mathrm {d}) \\tag {3}", + "image_path": "ffb2634d5ebf36757a958101071bfbc3956dbd83e0a5b4987f9618fea86ac1b7.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 52, + 450, + 287, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 450, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 52, + 450, + 287, + 464 + ], + "type": "interline_equation", + "content": "\\mathbf {x} = \\mathbf {x} + \\operatorname {T e m p S e l f A t t n} (\\mathrm {L N} (\\mathbf {x})) \\tag {4}", + "image_path": "761a32a6f0cea73a99d2096f9d3c8bc7892a39f65ef097abe10720b51b1bcbbe.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 52, + 466, + 287, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 466, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 52, + 466, + 287, + 479 + ], + "type": "interline_equation", + "content": "\\mathbf {x} = \\text {r e a r r a n g e} (\\mathbf {x}, (\\mathrm {b} \\quad \\mathrm {n}) \\quad \\mathrm {t} \\quad \\mathrm {d} \\rightarrow (\\mathrm {b} \\quad \\mathrm {t}) \\quad \\mathrm {n} \\quad \\mathrm {d}) \\tag {5}", + "image_path": "1a4e766aebde08faa2add9fb31fea2ff4e5f973df4b542b30b1cabb3b29d98a4.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 493, + 287, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 493, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 493, + 287, + 589 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 493, + 287, + 589 + ], + "type": "inline_equation", + "content": "b, t, n, d" + }, + { + "bbox": [ + 46, + 493, + 287, + 589 + ], + "type": "text", + "content": " represent the batch size, number of frames, number of patches per frame, and channel dimension, respectively. rearrange is a notation from [52]. We discovered that a simple TempSelfAttn layer suffices to capture motion, a finding that aligns with observations in a recent study [62]. In addition, only using TempSelfAttn makes it convenient to turn on and turn off the temporal modeling, which would be discussed in Sec. 3.3.2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": "Initialization. We use the pre-trained T2I model as a basis for initializing the shared layers between T2I and T2V models. In addition, for the newly added TempSelfAttn layers, we initialize the weights and biases of the output project layers to zero. This ensures that at the beginning of the T2V fine-tuning stage, these layers produce a zero output, effectively functioning as an identity mapping in conjunction with the shortcut connection." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 72, + 437, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 437, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 437, + 83 + ], + "type": "text", + "content": "3.3.2 Motion-Free Guidance" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 92, + 547, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 92, + 547, + 200 + ], + "spans": [ + { + "bbox": [ + 304, + 92, + 547, + 200 + ], + "type": "text", + "content": "Challenges encountered. We observed a notable phenomenon in the current T2V diffusion models [5, 30] where the per-frame visual quality significantly lags behind that of T2I models [15, 46, 53, 64]. Furthermore, our analysis revealed a remarkable degradation in visual quality in the T2V models post-fine-tuning, especially when compared to their original T2I counterparts. We note that these problems generally exist in current T2V diffusion models, not limited to our transformer-based T2V." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 220, + 546, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 220, + 546, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 220, + 546, + 411 + ], + "type": "text", + "content": "Problem analysis and insights. We presume that the observed lag in the visual quality of T2V primarily stems from two factors: the nature of video data and the fine-tuning approach. Firstly, publicly available video datasets often fall short in both quality and quantity compared to image datasets. For instance, [56] has more than 2B English image-text pairs, whereas the current widely used video dataset, WebVid-10M [1] contains only 10.7M video-text pairs. Additionally, many video frames are compromised by motion blur and watermarks, further reducing their visual quality. This limited availability hampers the development of robust and versatile video diffusion models. Secondly, the focus on optimizing temporal aspects during video fine-tuning can inadvertently compromise the spatial visual quality, resulting in a decline in the overall quality of the generated videos." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 430, + 545, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 430, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 430, + 545, + 502 + ], + "type": "text", + "content": "Solution I: joint image-video training. From the data aspect, we adopt the joint image-video training strategy [9, 16, 30, 62] to mitigate the video data shortages. Furthermore, joint training helps to alleviate the problem of domain discrepancy between video and image datasets by integrating both data types for training." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 521, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 605 + ], + "type": "text", + "content": "Solution II: motion-free guidance. We treat the temporal motion within a video clip as a special conditioning signal, which can be analogized to the textual conditioning in T2I/T2V diffusion models. Based on this analogy, we propose a novel approach, motion-free guidance (MFG), inspired by classifier-free guidance [6, 26], to modulate the weight of motion information in the generated video." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "type": "text", + "content": "In a particular training iteration, our approach mirrors the concept used in classifier-free guidance, where conditioned text is replaced with an empty string. The difference is that we employ an identity matrix to nullify the temporal attention with a probability of " + }, + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "type": "inline_equation", + "content": "p_{\\text{motion\\_free}}" + }, + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "type": "text", + "content": ". This identity matrix, which is depicted in Figure 3 (Motion-Free Mask), is structured such that its diagonal is populated with ones, while all other positions are zeroes. This configuration confines the temporal self-attention to work within a single" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6445" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 77, + 70, + 515, + 164 + ], + "blocks": [ + { + "bbox": [ + 77, + 70, + 515, + 164 + ], + "lines": [ + { + "bbox": [ + 77, + 70, + 515, + 164 + ], + "spans": [ + { + "bbox": [ + 77, + 70, + 515, + 164 + ], + "type": "table", + "html": "
ConditioningScaleAttribute BindingObject RelationshipComplexMean
Text EncoderTypeIntegrationColorShapeTextureSpatialNon-spatial
CLIP-L [50]MMadaLN-zeroXL/236.9442.0650.739.4130.3836.4134.32
CLIP-L [50]MMcross-attnXL/273.9151.8168.7619.2631.8041.5247.84
T5-XXL [13]LLMcross-attnXL/274.9055.4070.0520.5231.6841.0148.93
CLIP-T5XXLMM + LLMcross-attnXL/275.6555.7469.4820.6731.7941.4449.13
CLIP-T5XXLMM + LLMcross-attnG/276.7457.0071.5020.9832.0241.6749.99
", + "image_path": "57ba0ea001525f1cf1173ffe98c6fb672bb8935525ec7f17680f50e802b09b41.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 217, + 287, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 217, + 287, + 265 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 287, + 265 + ], + "type": "text", + "content": "frame. Furthermore, as introduced in Sec. 3.3.1, temporal self-attention is the sole operator for temporal modeling. Thus, using a motion-free attention mask suffices to disable temporal modeling in the video diffusion process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 266, + 287, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 287, + 289 + ], + "type": "text", + "content": "During inference, we have text and motion conditionings. Inspired by [6], we can modify the score estimate as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 80, + 312, + 286, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 312, + 286, + 354 + ], + "spans": [ + { + "bbox": [ + 80, + 312, + 286, + 354 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\tilde {\\epsilon} _ {\\theta} = \\epsilon_ {\\theta} (x _ {t}, \\varnothing , \\varnothing) \\\\ + \\lambda_ {T} \\cdot \\left(\\epsilon_ {\\theta} \\left(x _ {t}, c _ {T}, c _ {M}\\right) - \\epsilon_ {\\theta} \\left(x _ {t}, \\varnothing , c _ {M}\\right)\\right) \\\\ + \\lambda_ {M} \\cdot \\left(\\epsilon_ {\\theta} \\left(x _ {t}, \\varnothing , c _ {M}\\right) - \\epsilon_ {\\theta} \\left(x _ {t}, \\varnothing , \\varnothing\\right)\\right) \\tag {6} \\\\ \\end{array}", + "image_path": "d20afea9a5f19c340ad30f96aff7ef9cecb1e06a8c9b0109288d11ed414cc3d4.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 363, + 287, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 363, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 46, + 363, + 287, + 459 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 363, + 287, + 459 + ], + "type": "inline_equation", + "content": "c_{T}" + }, + { + "bbox": [ + 46, + 363, + 287, + 459 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 363, + 287, + 459 + ], + "type": "inline_equation", + "content": "c_{M}" + }, + { + "bbox": [ + 46, + 363, + 287, + 459 + ], + "type": "text", + "content": " represent the text conditioning and motion conditioning. " + }, + { + "bbox": [ + 46, + 363, + 287, + 459 + ], + "type": "inline_equation", + "content": "\\lambda_{T}" + }, + { + "bbox": [ + 46, + 363, + 287, + 459 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 363, + 287, + 459 + ], + "type": "inline_equation", + "content": "\\lambda_{M}" + }, + { + "bbox": [ + 46, + 363, + 287, + 459 + ], + "type": "text", + "content": " are the guidance scale of standard text and that of motion, controlling how strongly the generated samples correspond with the text condition and the motion strength, respectively. We empirically found that fixing " + }, + { + "bbox": [ + 46, + 363, + 287, + 459 + ], + "type": "inline_equation", + "content": "\\lambda_{T} = 7.5" + }, + { + "bbox": [ + 46, + 363, + 287, + 459 + ], + "type": "text", + "content": " and adjusting " + }, + { + "bbox": [ + 46, + 363, + 287, + 459 + ], + "type": "inline_equation", + "content": "\\lambda_{M} \\in [1.0, 1.3]" + }, + { + "bbox": [ + 46, + 363, + 287, + 459 + ], + "type": "text", + "content": " for each example tend to achieve the best result. This finding is similar to [6], although our study utilizes a narrower range for " + }, + { + "bbox": [ + 46, + 363, + 287, + 459 + ], + "type": "inline_equation", + "content": "\\lambda_{M}" + }, + { + "bbox": [ + 46, + 363, + 287, + 459 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 473, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 287, + 544 + ], + "type": "text", + "content": "Putting solutions together. We can integrate solution I and II together in the following way: when the motion is omitted at a training step, we load an image-text pair and repeat the image " + }, + { + "bbox": [ + 46, + 473, + 287, + 544 + ], + "type": "inline_equation", + "content": "T - 1" + }, + { + "bbox": [ + 46, + 473, + 287, + 544 + ], + "type": "text", + "content": " times to create a pseudo video. Conversely, if motion is included, we instead load a video clip and extract it into " + }, + { + "bbox": [ + 46, + 473, + 287, + 544 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 473, + 287, + 544 + ], + "type": "text", + "content": " frames." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 555, + 127, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 555, + 127, + 569 + ], + "spans": [ + { + "bbox": [ + 47, + 555, + 127, + 569 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 575, + 179, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 179, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 179, + 588 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": "Training scheme For all GenTron model variations, we employ the AdamW [38] optimizer, maintaining a constant learning rate of " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": ". We train our T2I GenTron models in a multi-stage procedure [46, 53] with an internal dataset, including a low-resolution " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "(256 \\times 256)" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": " training with a batch size of 2048 and 500K optimization steps, as well as high-resolution " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "(512 \\times 512)" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": " with a batch size of 784 and 300K steps. For the GenTron-G/2 model, we further integrate Fully Sharded Data Parallel (FSDP) [67] and activation checkpointing (AC), strategies specifically adopted to" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 307, + 215, + 544, + 342 + ], + "blocks": [ + { + "bbox": [ + 46, + 169, + 547, + 203 + ], + "lines": [ + { + "bbox": [ + 46, + 169, + 547, + 203 + ], + "spans": [ + { + "bbox": [ + 46, + 169, + 547, + 203 + ], + "type": "text", + "content": "Table 2. Conditioning and model scale in GenTron. We compare GenTron model variants with different design choices on T2I-CompBench [31]. The text encoders are from the language tower of the multi-modal (MM) model, the large language model (LLM), or a combination of them. The GenTron-G/2 with CLIP-T5XXL performs best. Detailed discussions can be found in Sec. 4.2." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 215, + 544, + 342 + ], + "lines": [ + { + "bbox": [ + 307, + 215, + 544, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 215, + 544, + 342 + ], + "type": "table", + "html": "
ModelAttribute BindingObj. RelationComp.Mean
ColorShapeTextureSpat.Non-spat.
LDM v1.437.6535.7641.5612.4630.7930.8031.50
LDM v250.6542.2149.2213.4230.9633.8636.72
Composable v240.6332.9936.458.0029.8028.9829.47
Structured v249.9042.1849.0013.8631.1133.5536.60
Attn-Exact v264.0045.1759.6314.5531.0934.0141.41
GORS66.0347.8562.8718.1531.9333.2843.35
DALL-E 257.5054.6463.7412.8330.4336.9642.68
LDM XL63.6954.0856.3720.3231.1040.9144.41
PixArt-α68.8655.8270.4420.8231.7941.1748.15
GenTron76.7457.0071.5020.9832.0241.6749.99
", + "image_path": "150d925731fbe700347e9b39043c2a0ae10a9dcdc97b1f1449330355d75cb7ca.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 345, + 545, + 389 + ], + "lines": [ + { + "bbox": [ + 305, + 345, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 305, + 345, + 545, + 389 + ], + "type": "text", + "content": "Table 3. Comparison of alignment evaluation on T2I-CompBench [31]. Results show our advanced model, GenTron-CLIPT5XXL-G/2, achieves superior performance across multiple compositional metrics compared to previous methods." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 407, + 545, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 407, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 407, + 545, + 491 + ], + "type": "text", + "content": "optimize GPU memory usage. In our video experiments, we train videos on a video dataset that comprises approximately 34M videos. To optimize storage usage and enhance data loading efficiency, the videos are pre-processed to a resolution with a short side of 512 pixels and a frame rate of 24 FPS. We process batches of 128 video clips. Each clip comprises 8 frames, captured at a sampling rate of 4 FPS." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 514, + 545, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 514, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 304, + 514, + 545, + 598 + ], + "type": "text", + "content": "Evaluation metrics. We mainly adopt the recent T2I-CompBench [31] to compare GenTron model variants, following [4, 10]. Specifically, we compare the attribute binding aspects, which include color, shape, and texture. We also compare the spatial and non-spatial object relationships. Moreover, user studies are conducted to compare visual quality and text alignment." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 609, + 468, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 609, + 468, + 621 + ], + "spans": [ + { + "bbox": [ + 305, + 609, + 468, + 621 + ], + "type": "text", + "content": "4.2. Main Results of GenTron-T2I" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "In this subsection, we discuss our experimental results, focusing on how various conditioning factors and model sizes impact GenTron's performance. Table 2 presents the quantitative findings of our study. Additionally, we offer a comparative visualization, illustrating the effects of each conditioning factor we explored. A comparison to prior art is also provided." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6446" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 70, + 166, + 182 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 166, + 182 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 166, + 182 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 166, + 182 + ], + "type": "image", + "image_path": "a3324278740a7af051c211a79feffb5c48d2b940628ce4edc5f70c590d7e53c8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 185, + 135, + 194 + ], + "lines": [ + { + "bbox": [ + 83, + 185, + 135, + 194 + ], + "spans": [ + { + "bbox": [ + 83, + 185, + 135, + 194 + ], + "type": "text", + "content": "(a)adaLN-Zero" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 167, + 70, + 279, + 182 + ], + "blocks": [ + { + "bbox": [ + 167, + 70, + 279, + 182 + ], + "lines": [ + { + "bbox": [ + 167, + 70, + 279, + 182 + ], + "spans": [ + { + "bbox": [ + 167, + 70, + 279, + 182 + ], + "type": "image", + "image_path": "4559b4fdbd5e73a5d760c13e1fb6f6c0241445b4dd33bcd5f42faa738400c22d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 185, + 253, + 194 + ], + "lines": [ + { + "bbox": [ + 192, + 185, + 253, + 194 + ], + "spans": [ + { + "bbox": [ + 192, + 185, + 253, + 194 + ], + "type": "text", + "content": "(b) Cross attention" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 244, + 286, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 244, + 286, + 375 + ], + "spans": [ + { + "bbox": [ + 46, + 244, + 286, + 375 + ], + "type": "text", + "content": "Cross attention vs. adaLN-Zero. Recent findings [44] conclude that the adaLN design yields superior results in terms of the FID, outperforming both cross-attention and in-context conditioning in efficiency for class-based scenarios. However, our observations reveal a limitation of adaLN in handling free-form text conditioning. This shortcoming is evident in Figure 4, where adaLN's attempt to generate a panda image falls short, with cross-attention demonstrating a clear advantage. This is further verified quantitatively in the first two rows of Table 2, where cross-attention uniformly excels over adaLN in all evaluated metrics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 376, + 287, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 287, + 556 + ], + "type": "text", + "content": "This outcome is reasonable considering the nature of class conditioning, which typically involves a limited set of fixed signals (e.g., the 1000 one-hot class embeddings for ImageNet [18]). In such contexts, the adaLN approach, operating at a spatially global level, adjusts the image features uniformly across all positions through the normalization layer, making it adequate for the static signals. In contrast, cross-attention treats spatial positions with more granularity. It differentiates between various spatial locations by dynamically modulating image features based on the cross-attention map between the text embedding and the image features. This spatial-sensitive processing is essential for free-from text conditioning, where the conditioning signals are infinitely diverse and demand detailed representation in line with the specific content of textual descriptions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "content": "Comparative analysis of text encoders. In Table 2 (rows two to four), we conduct a quantitative evaluation of various text encoders on T2I-CompBench, ensuring a fair comparison by maintaining a consistent XL/2 size across models. Results reveal that GenTron-T5XXL outperforms GenTron-CLIP-L across all three attribute binding and spatial relationship metrics, while it demonstrates comparable performance in the remaining two metrics. This suggests that T5 embeddings are superior in terms of compositional ability. These observations are in line with [2], which utilizes both CLIP and T5 embeddings for training but tests them individually or in combination during inference. Unlike eDiff," + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 312, + 70, + 425, + 182 + ], + "blocks": [ + { + "bbox": [ + 46, + 198, + 287, + 232 + ], + "lines": [ + { + "bbox": [ + 46, + 198, + 287, + 232 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 287, + 232 + ], + "type": "text", + "content": "Figure 4. adaLN-Zero vs. cross attention. The prompt is \"A panda standing on a surfboard in the ocean in sunset.\" Cross attention exhibits a distinct advantage in the text-conditioned scenario." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 312, + 70, + 425, + 182 + ], + "lines": [ + { + "bbox": [ + 312, + 70, + 425, + 182 + ], + "spans": [ + { + "bbox": [ + 312, + 70, + 425, + 182 + ], + "type": "image", + "image_path": "b567381917800d47279ed87745988bb05d38e90981403c764900d101ab7009be.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 338, + 185, + 398, + 194 + ], + "lines": [ + { + "bbox": [ + 338, + 185, + 398, + 194 + ], + "spans": [ + { + "bbox": [ + 338, + 185, + 398, + 194 + ], + "type": "text", + "content": "(a) GenTron-XL/2" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 198, + 545, + 243 + ], + "lines": [ + { + "bbox": [ + 305, + 198, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 305, + 198, + 545, + 243 + ], + "type": "text", + "content": "Figure 5. Effect of model scale. The prompt is \"a cat reading a newspaper\". The larger model GenTron-G/2 excels in rendering finer details and rationalization in the layout of the cat and newspaper. More comparisons can be found in the appendix." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 426, + 71, + 538, + 182 + ], + "blocks": [ + { + "bbox": [ + 426, + 71, + 538, + 182 + ], + "lines": [ + { + "bbox": [ + 426, + 71, + 538, + 182 + ], + "spans": [ + { + "bbox": [ + 426, + 71, + 538, + 182 + ], + "type": "image", + "image_path": "ae1c4fcef810c0358b38094c6070e77db9ef7cb1543a602ccdb72fa6e343df07.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 454, + 185, + 509, + 194 + ], + "lines": [ + { + "bbox": [ + 454, + 185, + 509, + 194 + ], + "spans": [ + { + "bbox": [ + 454, + 185, + 509, + 194 + ], + "type": "text", + "content": "(b) GenTron-G/2" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 307, + 246, + 545, + 358 + ], + "blocks": [ + { + "bbox": [ + 307, + 246, + 545, + 358 + ], + "lines": [ + { + "bbox": [ + 307, + 246, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 307, + 246, + 545, + 358 + ], + "type": "table", + "html": "
Method#Param.FID-30K↓CLIP-Score↑T2I-CompBen↑
Imagen3.0B7.270.27-
Parti-750M0.8B10.71--
Parti-3B3.0B8.10--
GigaGAN1.0B9.090.322-
MUSE-3B3.0B7.880.320-
SD v1.40.9B12.940.32531.50
SDXL2.6B17.820.32944.41
GenTron-XL/20.9B14.210.32649.13
GenTron-G/23.1B14.530.33549.99
", + "image_path": "1a556455ab05ef8cc627cd0cd1449d53214184572d68f1962cdabcf2484da7f5.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 356, + 368, + 495, + 378 + ], + "lines": [ + { + "bbox": [ + 356, + 368, + 495, + 378 + ], + "spans": [ + { + "bbox": [ + 356, + 368, + 495, + 378 + ], + "type": "text", + "content": "Table 4. Comparison to T2I models." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 396, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 396, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 305, + 396, + 545, + 456 + ], + "type": "text", + "content": "our approach maintains the same settings for both training and inference. Notably, GenTron demonstrates enhanced performance when combining CLIP-L and T5XXL embeddings, indicating the model's ability to leverage the distinct advantages of each text embedding type." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 473, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 545, + 628 + ], + "type": "text", + "content": "Scaling GenTron up. In Figure 5, we showcase examples from the PartiPrompts benchmark [65] to illustrate the qualitative enhancements achieved by scaling our model up from approximately 900 million to 3 billion parameters. Both models operate under the same CLIP-T5XXL condition. The larger GenTron-G/2 model excels in rendering finer details and more accurate representations, particularly in rationalizing the layout of objects like cats and newspapers. This results in image compositions that are both more coherent and more realistic. In comparison, the smaller GenTron-XL/2 model, while producing recognizable images with similar color schemes, falls short in terms of precision and visual appeal." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Furthermore, the superiority of GenTron-G/2 is quantitatively affirmed through its performance on T2I-CompBench, as detailed in Table 2. The increase in model size correlates with significant improvements across all evaluative criteria for object composition, including attributes and relationships. Additional comparative examples are provided in the appendix for further illustration." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6447" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 535, + 129 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 535, + 129 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 535, + 129 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 535, + 129 + ], + "type": "image", + "image_path": "9e76d4ee6a95af33bf39178cf9f107e18420b7cecdf7bb753ce13525ff201ac7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 133, + 526, + 145 + ], + "lines": [ + { + "bbox": [ + 66, + 133, + 526, + 145 + ], + "spans": [ + { + "bbox": [ + 66, + 133, + 526, + 145 + ], + "type": "text", + "content": "Figure 6. GenTron-T2V examples. Prompts are \"A giant tortoise is making its way across the beach\" and \"A dog swimming\"." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 52, + 165, + 281, + 224 + ], + "blocks": [ + { + "bbox": [ + 52, + 165, + 281, + 224 + ], + "lines": [ + { + "bbox": [ + 52, + 165, + 281, + 224 + ], + "spans": [ + { + "bbox": [ + 52, + 165, + 281, + 224 + ], + "type": "image", + "image_path": "b8916973999bebc8e3ed364d550f718a65574faeb7be3595a41fccc7f36f2ae4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 232, + 288, + 277 + ], + "lines": [ + { + "bbox": [ + 46, + 232, + 288, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 288, + 277 + ], + "type": "text", + "content": "Figure 7. Visualization of the human preference of our method vs. Latent Diffusion XL. Our method received a significantly higher number of votes as the winner in comparisons of visual quality and text faithfulness, with a total of 3000 answers." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 296, + 287, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 296, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 296, + 287, + 392 + ], + "type": "text", + "content": "Comparison to prior work. In Table 3, we showcase the alignment evaluation results from T2I-CompBench. Our method demonstrates outstanding performance in all areas, including attribute binding, object relationships, and complex compositions. This indicates a heightened proficiency in compositional generation, with a notable strength in color binding. In this aspect, our approach surpasses the previous state-of-the-art (SoTA) work [10] by over " + }, + { + "bbox": [ + 46, + 296, + 287, + 392 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 46, + 296, + 287, + 392 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 393, + 287, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 393, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 46, + 393, + 287, + 525 + ], + "type": "text", + "content": "In Table 4, we further thoroughly compare the zero-shot FID-30K, CLIP-Score, and T2I-CompBench metrics with previous T2I models. In contrast to U-Net based SDv1.4, GenTron uses about four times less data (550M vs. 2B), while achieving better CLIP-scores and T2I-CompBench results. Although GenTron does not achieve a superior FID score compared to SDv1.4, it's important to note that recent studies, including Pick-a-Pic [36] and SDXL [46], have highlighted that FID scores often misrepresent human preferences in generative models, sometimes even negatively correlating with visual aesthetics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 542, + 287, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 542, + 287, + 627 + ], + "spans": [ + { + "bbox": [ + 46, + 542, + 287, + 627 + ], + "type": "text", + "content": "User study Figure 7 shows the human preference of our method versus SDXL. We used standard prompts in PartiPrompt2 [65] to generate 100 images using both methods and ask people for their preference blindly after shuffling. We received a total of three thousand responses on the comparisons of visual quality and text faithfulness, with our method emerging as the clear winner by a clear margin." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 634, + 174, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 174, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 174, + 647 + ], + "type": "text", + "content": "4.3. GenTron-T2V Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "In Figure 6, We showcase several samples generated by GenTron-T2V, which are not only visually striking but also temporally coherent. This highlights GenTron-T2V's effectiveness in creating videos that are both aesthetically pleasing and consistent over time." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 165, + 541, + 276 + ], + "blocks": [ + { + "bbox": [ + 310, + 165, + 541, + 276 + ], + "lines": [ + { + "bbox": [ + 310, + 165, + 541, + 276 + ], + "spans": [ + { + "bbox": [ + 310, + 165, + 541, + 276 + ], + "type": "image", + "image_path": "ea2c668e91e1c69d5f86bda74e77889ad223d7c571b7e9a9694b7bfefd704206.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 277, + 545, + 310 + ], + "lines": [ + { + "bbox": [ + 305, + 277, + 545, + 310 + ], + "spans": [ + { + "bbox": [ + 305, + 277, + 545, + 310 + ], + "type": "text", + "content": "Figure 8. Effect of motion-free guidance. GenTron-T2V with motion-free guidance has a clear visual appearance improvement. The prompt is \"A lion standing on a surfboard in the ocean in sunset\"." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 328, + 546, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 328, + 546, + 472 + ], + "spans": [ + { + "bbox": [ + 305, + 328, + 546, + 472 + ], + "type": "text", + "content": "Effect of motion-free guidance. In Figure 8, we present a comparison between our GenTron variants, with motion-free guidance (MFG) and without. For this comparison, critical factors such as the pre-trained T2I model, training data, and the number of training iterations were kept constant to ensure a fair evaluation. The results clearly indicate that GenTron-T2V, when integrated with MFG, shows a marked tendency to focus on the central object mentioned in the prompt, often rendering it in greater detail. Specifically, the object typically occupies a more prominent, central position in the generated video, thereby dominating the visual focus across video frames." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 488, + 379, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 488, + 379, + 501 + ], + "spans": [ + { + "bbox": [ + 306, + 488, + 379, + 501 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 510, + 547, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 510, + 547, + 654 + ], + "spans": [ + { + "bbox": [ + 305, + 510, + 547, + 654 + ], + "type": "text", + "content": "In this work, we provide a thorough exploration of transformer-based diffusion models for text-conditioned image and video generation. Our findings shed light on the properties of various conditioning approaches and offer compelling evidence of quality improvement when scaling up the model. A notable contribution of our work is the development of GenTron for video generation, where we introduce motion-free guidance. This innovative approach has demonstrably enhanced the visual quality of generated videos. We hope that our research will contribute to bridging the existing gap in applying transformers to diffusion models and their broader use in other domains." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 547, + 713 + ], + "type": "text", + "content": "Acknowledgement. This paper is partially supported by the National Key R&D Program of China No.2022ZD0161000 and the General Research Fund of Hong Kong No.17200622." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6448" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Max Bain, Arsha Nagrani, Güll Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1728-1738, 2021. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 146, + 288, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 146, + 288, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 146, + 288, + 203 + ], + "type": "text", + "content": "[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 4, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 203, + 288, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 203, + 288, + 236 + ], + "spans": [ + { + "bbox": [ + 53, + 203, + 288, + 236 + ], + "type": "text", + "content": "[3] Fan Bao, Shen Nie, Kaiwen Xue, Yue Cao, Chongxuan Li, Hang Su, and Jun Zhu. All are worth words: A vit backbone for diffusion models. In CVPR, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 288, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 288, + 280 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 288, + 280 + ], + "type": "text", + "content": "[4] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions, 2023. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 280, + 288, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 280, + 288, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 280, + 288, + 346 + ], + "type": "text", + "content": "[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 2, 3, 4, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 347, + 288, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 347, + 288, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 347, + 288, + 401 + ], + "type": "text", + "content": "[6] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 403, + 288, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 288, + 458 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 288, + 458 + ], + "type": "text", + "content": "[7] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. OpenAI, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 459, + 288, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 459, + 288, + 514 + ], + "spans": [ + { + "bbox": [ + 53, + 459, + 288, + 514 + ], + "type": "text", + "content": "[8] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 514, + 288, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 514, + 288, + 569 + ], + "spans": [ + { + "bbox": [ + 53, + 514, + 288, + 569 + ], + "type": "text", + "content": "[9] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. Videocrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 2, 3, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 570, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 288, + 624 + ], + "type": "text", + "content": "[10] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 3, 6, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 625, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 288, + 714 + ], + "type": "text", + "content": "[11] Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish V Thapliyal, James Bradbury, Weicheng Kuo, Mojtaba Seyedhosseini, Chao Jia, Burcu Karagol Ayan, Carlos Riquelme Ruiz, Andreas Peter Steiner, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "Radu Soricut. PaLI: A jointly-scaled multilingual language-image model. In The Eleventh International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 107, + 547, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 547, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 547, + 163 + ], + "type": "text", + "content": "[12] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 164, + 547, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 164, + 547, + 218 + ], + "spans": [ + { + "bbox": [ + 307, + 164, + 547, + 218 + ], + "type": "text", + "content": "[13] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416, 2022. 2, 4, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 220, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 285 + ], + "type": "text", + "content": "[14] Yuren Cong, Mengmeng Xu, christian simon, Shoufa Chen, Jiawei Ren, Yanping Xie, Juan-Manuel Perez-Rua, Bodo Rosenhahn, Tao Xiang, and Sen He. FLATTEN: optical FLOW-guided ATTENTION for consistent text-to-video editing. In The Twelfth International Conference on Learning Representations, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 286, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 545, + 342 + ], + "type": "text", + "content": "[15] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xiaofang Wang, Abhimanyu Dubey, et al. Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807, 2023. 2, 4, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 343, + 545, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 545, + 399 + ], + "type": "text", + "content": "[16] Yatin Dandi, Aniket Das, Soumye Singhal, Vinay Nambodiri, and Piyush Rai. Jointly trained image and video generation using residual vectors. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3028-3042, 2020. 2, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 399, + 545, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 466 + ], + "type": "text", + "content": "[17] Mostafa Dehghani, Josip Djolonga, Basil Mustafa, Piotr Padlewski, Jonathan Heek, Justin Gilmer, Andreas Peter Steiner, Mathilde Caron, Robert Geirhos, Ibrahim Alabdulmohsin, et al. Scaling vision transformers to 22 billion parameters. In International Conference on Machine Learning, pages 7480-7512. PMLR, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 467, + 545, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 467, + 545, + 512 + ], + "spans": [ + { + "bbox": [ + 307, + 467, + 545, + 512 + ], + "type": "text", + "content": "[18] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 513, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 513, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 307, + 513, + 545, + 611 + ], + "type": "text", + "content": "[19] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers), pages 4171-4186. Association for Computational Linguistics, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 613, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 689 + ], + "type": "text", + "content": "[20] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 690, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 690, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 690, + 545, + 714 + ], + "type": "text", + "content": "[21] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Pro" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "6449" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 66, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 73, + 287, + 95 + ], + "type": "text", + "content": "ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 137 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 137 + ], + "type": "text", + "content": "[22] Shanghua Gao, Pan Zhou, Ming-Ming Cheng, and Shuicheng Yan. Masked diffusion transformer is a strong image synthesizer. arXiv preprint arXiv:2303.14389, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 139, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 139, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 139, + 287, + 183 + ], + "type": "text", + "content": "[23] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 184, + 287, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 287, + 237 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 287, + 237 + ], + "type": "text", + "content": "[24] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 239, + 287, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 287, + 282 + ], + "type": "text", + "content": "[25] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 283, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 283, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 48, + 283, + 287, + 316 + ], + "type": "text", + "content": "[26] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 317, + 287, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 317, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 317, + 287, + 348 + ], + "type": "text", + "content": "[27] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 350, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 287, + 403 + ], + "type": "text", + "content": "[28] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 404, + 287, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 287, + 448 + ], + "type": "text", + "content": "[29] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. The Journal of Machine Learning Research, 23(1):2249-2281, 2022. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 449, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 480 + ], + "type": "text", + "content": "[30] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video diffusion models, 2022. 2, 4, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 482, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 482, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 482, + 287, + 536 + ], + "type": "text", + "content": "[31] Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 537, + 287, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 287, + 581 + ], + "type": "text", + "content": "[32] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 582, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 582, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 582, + 287, + 635 + ], + "type": "text", + "content": "[33] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6007-6017, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 637, + 287, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 637, + 287, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 637, + 287, + 691 + ], + "type": "text", + "content": "[34] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "text", + "content": "[35] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 2, 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "type": "text", + "content": "[36] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Matiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36, 2024. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 129, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 184 + ], + "type": "text", + "content": "[37] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 186, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 207 + ], + "type": "text", + "content": "[38] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 209, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 262 + ], + "type": "text", + "content": "[39] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 264, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 308 + ], + "type": "text", + "content": "[40] Haoyu Lu, Guoxing Yang, Nanyi Fei, Yuqi Huo, Zhiwu Lu, Ping Luo, and Mingyu Ding. Vdt: An empirical study on video diffusion with transformers. arXiv preprint arXiv:2305.13311, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 309, + 545, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 309, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 308, + 309, + 545, + 375 + ], + "type": "text", + "content": "[41] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Tieniu Tan. Videofusion: Decomposed diffusion models for high-quality video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10209-10218, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 376, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 545, + 431 + ], + "type": "text", + "content": "[42] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 433, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 433, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 308, + 433, + 545, + 498 + ], + "type": "text", + "content": "[43] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob Mcgrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In Proceedings of the 39th International Conference on Machine Learning, pages 16784-16804. PMLR, 2022. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 500, + 545, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 500, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 308, + 500, + 545, + 532 + ], + "type": "text", + "content": "[44] William Peebles and Saining Xie. Scalable diffusion models with transformers. arXiv preprint arXiv:2212.09748, 2022. 2,3,4,7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 534, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 534, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 534, + 545, + 578 + ], + "type": "text", + "content": "[45] Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. In Proceedings of the AAAI conference on artificial intelligence, 2018. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 579, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 579, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 308, + 579, + 545, + 633 + ], + "type": "text", + "content": "[46] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 4, 5, 6, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 635, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 679 + ], + "type": "text", + "content": "[47] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. arXiv preprint arXiv:2303.09535, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "text", + "content": "[48] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. OpenAI blog, 2018. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6450" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[49] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 171 + ], + "type": "text", + "content": "[50] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 4, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 287, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 287, + 227 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 287, + 227 + ], + "type": "text", + "content": "[51] Carlos Riquelme, Joan Puigcerver, Basil Mustafa, Maxim Neumann, Rodolphe Jenatton, André Susano Pinto, Daniel Keysers, and Neil Houlsby. Scaling vision with sparse mixture of experts. Advances in Neural Information Processing Systems, 34:8583-8595, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 228, + 287, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 287, + 260 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 287, + 260 + ], + "type": "text", + "content": "[52] Alex Rogozhnikov. Einops: Clear and reliable tensor manipulations with einstein-like notation. In International Conference on Learning Representations, 2021. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 261, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 287, + 316 + ], + "type": "text", + "content": "[53] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 316, + 287, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 287, + 381 + ], + "type": "text", + "content": "[54] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 383, + 287, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 287, + 448 + ], + "type": "text", + "content": "[55] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 449, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 514 + ], + "type": "text", + "content": "[56] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 514, + 287, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 579 + ], + "type": "text", + "content": "[57] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-video: Text-to-video generation without text-video data. In The Eleventh International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 581, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 287, + 613 + ], + "type": "text", + "content": "[58] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 614, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 668 + ], + "type": "text", + "content": "[59] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[60] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2, 4" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 518 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[61] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "type": "text", + "content": "[62] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 2, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "type": "text", + "content": "[63] Yilin Wang, Zeyuan Chen, Liangjun Zhong, Zheng Ding, Zhizhou Sha, and Zhuowen Tu. Dolfin: Diffusion layout transformers without autoencoder. arXiv preprint arXiv:2310.16305, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 207, + 545, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 207, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 307, + 207, + 545, + 261 + ], + "type": "text", + "content": "[64] Zeyue Xue, Guanglu Song, Qiushan Guo, Boxiao Liu, Zhuofan Zong, Yu Liu, and Ping Luo. RAPHAEL: Text-to-image generation via large mixture of diffusion paths. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 262, + 545, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 262, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 307, + 262, + 545, + 340 + ], + "type": "text", + "content": "[65] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. Transactions on Machine Learning Research, 2022. Featured Certification. 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 341, + 545, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 341, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 307, + 341, + 545, + 385 + ], + "type": "text", + "content": "[66] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12104-12113, 2022. 2, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 386, + 545, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 386, + 545, + 440 + ], + "spans": [ + { + "bbox": [ + 307, + 386, + 545, + 440 + ], + "type": "text", + "content": "[67] Yanli Zhao, Andrew Gu, Rohan Varma, Liang Luo, Chien-Chin Huang, Min Xu, Less Wright, Hamid Shojanazeri, Myle Ott, Sam Shleifer, et al. Pytorch fsdp: experiences on scaling fully sharded data parallel. arXiv preprint arXiv:2304.11277, 2023. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 441, + 545, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 441, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 307, + 441, + 545, + 475 + ], + "type": "text", + "content": "[68] Hongkai Zheng, Weili Nie, Arash Vahdat, and Anima Anandkumar. Fast training of diffusion models with masked transformers. arXiv preprint arXiv:2306.09305, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 475, + 545, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 475, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 307, + 475, + 545, + 518 + ], + "type": "text", + "content": "[69] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 3, 4" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "6451" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/GenZI_ Zero-Shot 3D Human-Scene Interaction Generation/3823a91d-f3af-49d1-a8df-4cb9a1c22737_content_list.json b/2024/GenZI_ Zero-Shot 3D Human-Scene Interaction Generation/3823a91d-f3af-49d1-a8df-4cb9a1c22737_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d70ae26bc4ab85dd8bde7e5cdb340381d00b5181 --- /dev/null +++ b/2024/GenZI_ Zero-Shot 3D Human-Scene Interaction Generation/3823a91d-f3af-49d1-a8df-4cb9a1c22737_content_list.json @@ -0,0 +1,1880 @@ +[ + { + "type": "text", + "text": "GenZI: Zero-Shot 3D Human-Scene Interaction Generation", + "text_level": 1, + "bbox": [ + 181, + 130, + 785, + 150 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lei Li Angela Dai Technical University of Munich", + "bbox": [ + 354, + 181, + 609, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "\"riding a motorcycle, sitting\"", + "bbox": [ + 102, + 256, + 264, + 271 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "\"hands knocking on the closed door, facing the door, standing\"", + "bbox": [ + 295, + 251, + 475, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "\"stretching legs and sitting on the saddle on a standing cow\"", + "bbox": [ + 498, + 251, + 673, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "\"picking up dumbbells on a shelf, standing, bending over\"", + "bbox": [ + 700, + 251, + 870, + 276 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7fb1eb99e239083e4c92aa15a562c585e889e82f9690fca9541c91db3f6ce443.jpg", + "image_caption": [ + "Figure 1. Given an arbitrary 3D scene, GenZI can synthesize virtual humans interacting with the 3D environment at specified locations from a brief text description. Our approach does not require any 3D human-scene interaction training data or 3D learning. By distilling interaction priors from powerful 2D vision-language models, we optimize for 3D human-scene interaction synthesis in a flexible fashion, with simple language-based control and high generality to various types of scene environments." + ], + "image_footnote": [], + "bbox": [ + 81, + 280, + 281, + 433 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cd67976967625ecb4d433022798ca86154c59cf857218566c4b55c3e95759441.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 284, + 281, + 483, + 433 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/19bdeca1d364be38016848ec3d9a7e77e40a1a765103878f229318030c6476fe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 281, + 684, + 433 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e649850e33763e06dc50ad25785d560cd3fa23e67c0c638c07bba83aa33b4df6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 686, + 281, + 885, + 433 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 518, + 312, + 534 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Can we synthesize 3D humans interacting with scenes without learning from any 3D human-scene interaction data? We propose GenZI1, the first zero-shot approach to generating 3D human-scene interactions. Key to GenZI is our distillation of interaction priors from large vision-language models (VLMs), which have learned a rich semantic space of 2D human-scene compositions. Given a natural language description and a coarse point location of the desired interaction in a 3D scene, we first leverage VLMs to imagine plausible 2D human interactions inpainted into multiple rendered views of the scene. We then formulate a robust iterative optimization to synthesize the pose and shape of a 3D human model in the scene, guided by consistency with the 2D interaction hypotheses. In contrast to existing learning-based approaches, GenZI circumvents the conventional need for captured 3D interaction data, and allows for flexible control of the 3D interaction synthesis with easy-to-use text prompts. Extensive experiments show that our zero-shot approach has high flexibility and generality, making it applicable to diverse scene types, including both indoor and outdoor environments.", + "bbox": [ + 75, + 546, + 472, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 518, + 630, + 534 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Scenes are typically constructed to enable human interactions with the environment, like sitting on a couch, playing a piano, or opening a mailbox. Understanding these interactions between humans and scenes, also known as affordances [9, 17], has gained increasing attention in the fields of computer vision and graphics. In particular, achieving controllable and generalizable synthesis of human-scene interactions (or HSIs) holds immense potential for various applications, such as robotics, architectural design, video games, and virtual reality experiences, among many others.", + "bbox": [ + 496, + 547, + 890, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generating realistic humans interacting with a 3D scene is a challenging task, requiring holistic semantic understanding of both the environment and possible human actions therein. Existing approaches to HSI synthesis [11, 46, 50, 51] rely heavily on supervised training using meticulously captured data of real people interacting in 3D environments. Unfortunately, collecting large-scale datasets of 3D scenes and human interactions is exorbitantly difficult. It not only demands accurate tracking and reconstruction of both people and their environments, but also needs to ensure sufficient diversity in subjects and scenes. Existing HSI datasets currently contain very limited quantities of scenes and actions, for example, PROX [10, 51] only consisting", + "bbox": [ + 496, + 703, + 892, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1Project page: craigileli.github.io/projects/genzi", + "bbox": [ + 94, + 886, + 351, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20465", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of 12 indoor scenes and humans interacting with 11 object categories. The scarcity of ground truth data for supervision has thus strongly limited the applicability and generalization of the learning-based approaches for synthesizing diverse sets of actions in arbitrary 3D scenes.", + "bbox": [ + 75, + 90, + 470, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We thus consider an alternative perspective to HSI synthesis and pose the following question: Can we achieve plausible HSI synthesis without using any captured 3D interaction data? To this end, we present a novel zero-shot approach to 3D HSI generation. We propose the first method to leverage the powerful capabilities exhibited by recent vision-language models (VLMs) [19, 20, 30, 33, 34, 37] to synthesize plausible 2D images of human interactions, and introduce a robust optimization to distill inferred 2D pose information into 3D human synthesis in a 3D scene.", + "bbox": [ + 75, + 167, + 468, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "More concretely, given a 3D scene, a text prompt and a coarse point location of the desired interaction, GenZI optimizes for the pose and shape of a 3D human performing the action in the scene, guided by a large VLM [34]. We first leverage the VLM to imagine possible 2D humans by inpainting images from multiple rendered views of the scene. We automate this 2D human insertion process with a dynamic masking scheme that automatically updates proposed masks through the inpainting process, eliminating the need for manual specifications of human inpainting regions. We then lift these 2D interaction hypotheses to 3D and optimize for a parametric 3D human body model [23, 29] that is most consistent with the 2D pose guidance. We further refine the generated 3D human in the scene by iterating through the VLM-based 2D inpainting and robust 3D lifting stages. We demonstrate the flexibility and generality of GenZI in various types of 3D environments (Fig. 1), encompassing both indoor and outdoor scenes.", + "bbox": [ + 75, + 335, + 468, + 606 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are as follows:", + "bbox": [ + 94, + 609, + 395, + 623 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce GenZI, the first zero-shot approach to generating realistic 3D humans interacting with a 3D scene from natural language prompts. GenZI does not require any supervision from 3D interaction data, thus enabling flexible synthesis across diverse scenes and actions.", + "- We propose a dynamically-masked inpainting scheme that allows for the synthesis of plausible 2D human-scene compositions via VLMs without requiring manually-specified human inpainting masks.", + "- We develop a robust 3D pose optimization to lift various inferred images of human interactions to a view-consistent, realistic 3D HSI synthesis." + ], + "bbox": [ + 76, + 625, + 467, + 805 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 824, + 217, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Human-Scene Interaction Synthesis. Synthesizing humans in scenes is an important, challenging task in computer vision and graphics, as it models complex high-level", + "bbox": [ + 75, + 854, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "semantic understanding such as affordances and interactions. Existing approaches for human-scene interaction synthesis [11, 12, 16, 21, 27, 39, 43, 44, 50, 51] focus on learning priors from available collected data of people interacting with 3D indoor scenes [10] in a supervised manner.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The early work PiGraphs [39] introduced a probabilistic graphical model of human-object interactions, generating static human-object placements by pose sampling and model retrieval. Zhang et al. [50] introduced a generative model of human-scene interaction, using a conditional variational autoencoder (CVAE) to model the distribution of 3D human poses conditioned on scene depth and semantics. PLACE [49] explicitly represented human-scene interactions with Basis Point Sets (BPS) encoding [32], training a CVAE to synthesize such representations for generating natural human poses and contact relations within scenes. POSA [11] proposed an ego-centric representation by augmenting the SMPL-X model [29] with contact labels and scene semantics, using a CVAE conditioned on SMPL-X vertex positions to model potential interactions. Recently, COINS [51] presented a method for compositional human-scene interaction synthesis with high-level semantic control, using transformer-based CVAEs conditioned on provided 3D objects and interaction semantics to regress human body poses and contact features.", + "bbox": [ + 496, + 167, + 890, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, these works all require supervision from 3D human-scene interactions – such datasets are difficult and expensive to capture and annotate, for instance, the widely used PROX dataset [10, 51] only has 8 reconstructed indoor scenes for training and 4 for testing, capturing human interactions with 11 object categories. As such, these methods are limited to in-domain synthesis, prohibiting applicability to general 3D scene settings with arbitrary objects and arrangements. In this work, we deviate from the learning-based approach and propose to take advantage of established large vision-language models [20, 33, 34, 37] for human-scene interaction synthesis, thereby bypassing the requirement for data capture and 3D learning.", + "bbox": [ + 496, + 472, + 890, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Human Estimation from RGB Images. Over the past decades, significant progress has been made in 3D human estimation from RGB images [7, 26, 31, 38]. Among those, the prominent work SMPLify-X [29] fits the SMPL-X model to 2D joints estimated from a single image through optimization. Hassan et al. [10] build upon SMPLify-X and estimate 3D human poses by incorporating additional physical contact and penetration constraints between human and 3D scene. Learning-based approaches [3, 5, 18, 22, 35, 48, 53] have received much research attention in recent years towards tackling 3D human estimation from monocular images with improved hand and face estimation. Our work leverages 2D pose reasoning from multiple different 2D view hypotheses, and proposes a robust 3D formulation for aggregating the various 2D hypotheses to a consistent", + "bbox": [ + 496, + 674, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20466", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/486a6430454ccd00503ed6038de0c7e738b69f3716bb2625c649fb57b26d31fe.jpg", + "image_caption": [ + "Figure 2. GenZI distills information from vision-language model for 3D human-scene interaction. We first leverage large vision-language models to synthesize possible 2D humans interactions with the 3D scene $S$ by employing latent diffusion inpainting [34] on multiple rendered views of the environment at location $\\mathbf{p}$ using our dynamic masking scheme to automatically estimate inpainting masks. We then lift these 2D hypotheses to 3D in a robust optimization for a 3D parametric body model $\\mathcal{B}$ (SMPL-X [29]) that is most consistent with detected 2D poses in the inpainted 2D hypotheses. This produces a semantically consistent interaction that respects the scene context, without requiring any 3D human-scene interaction data." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 890, + 369 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D human body interacting with a 3D scene.", + "bbox": [ + 76, + 481, + 374, + 496 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Distilling Prior from Vision-Language Models. Recent advances in powerful vision-language models [20, 33, 34, 37] have also inspired various works aiming to distill information learned through the models to various tasks, including 2D panoptic segmentation [45], 3D semantic segmentation [36], 3D scene generation [15], and synthesis of images with hand-object interactions [47]. Our approach leverages the 2D generative capacity of these models to convey information about possible human-scene interactions, which we then lift to a 3D, consistent interaction.", + "bbox": [ + 75, + 500, + 470, + 651 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 667, + 168, + 684 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 76, + 694, + 187, + 708 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our objective is to synthesize plausible 3D humans interacting with a 3D scene, guided by input text descriptions, in the absence of 3D interaction data capture for learning. We present GenZI, a novel optimization-based multi-view approach that leverages large VLMs to infer spatial relations of interactions between a human and the scene. Fig. 2 shows an illustration of our approach.", + "bbox": [ + 75, + 718, + 468, + 823 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "GenZI takes as input a 3D scene $S$ , a text prompt $\\Gamma$ describing the desired interaction, and an approximate point location $\\mathbf{p} \\in \\mathbb{R}^3$ in the scene around which the interaction should occur. Our approach generates a posed 3D human $\\mathcal{B}$ as output, performing the specified action in the scene. We", + "bbox": [ + 75, + 824, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "adopt SMPL-X [23, 29] to parameterize the 3D human $\\mathcal{B}$ as it provides a fully differentiable function mapping a set of pose and shape parameters $(\\mathbf{R},\\mathbf{t},\\Theta ,\\Phi)$ to a 3D human mesh with vertices $\\mathbf{V}$ and faces $\\mathbf{F}$ . We thus optimize for $(\\mathbf{R},\\mathbf{t},\\Theta ,\\Phi)$ characterizing $\\mathcal{B}$ in the scene $S$ , with $\\mathbf{R} \\in \\mathbb{R}^6$ denoting the global orientation as a continuous rotation representation [52], $\\mathbf{t} \\in \\mathbb{R}^3$ the global translation, $\\Theta \\in \\mathbb{R}^{32}$ the body pose in the latent space of VPoser [29] which decodes to body joint rotations $\\hat{\\Theta} \\in \\mathbb{R}^{21\\times 3}$ , and $\\Phi \\in \\mathbb{R}^{10}$ representing the body shape as blend shape coefficients.", + "bbox": [ + 496, + 481, + 892, + 632 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "GenZI synthesizes the desired interaction between the 3D human $\\mathcal{B}$ and the scene $S$ by distilling information from VLMs through 2D human inpainting followed by robust 3D lifting. We first generate a collection of plausible 2D human-scene compositions by employing a large VLM to inpaint humans into multiple rendered views of the scene $S$ (Sec. 3.2). We then introduce a robust 3D lifting procedure that optimizes for the pose and shape of the human $\\mathcal{B}$ , guided by consistency with the 2D interaction hypotheses (Sec. 3.3). We further refine the posed 3D human $\\mathcal{B}$ by iteratively updating the 2D inpaintings and 3D optimization.", + "bbox": [ + 496, + 633, + 893, + 800 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Inpainting Multi-view Interaction Hypotheses", + "text_level": 1, + "bbox": [ + 498, + 811, + 887, + 828 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first leverage a VLM to generate 2D hypotheses of potential human interactions in the scene by automatically inpainting humans into multiple rendered views of $S$ .", + "bbox": [ + 496, + 835, + 890, + 881 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-view Rendering. To capture scene context for 2D", + "bbox": [ + 500, + 885, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20467", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "human inpainting, we render multiple views of the 3D scene $S$ from $k$ virtual cameras looking at $\\mathbf{p}$ . Cameras are randomly sampled on a hemisphere and filtered according to the visibility of $\\mathbf{p}$ ; we refer to the supplementary material for additional camera setup details. We denote the rendered scene images as $\\{\\mathbf{I}^i\\}_{i\\in [1,k]}$ . To simplify notation, we omit the superscript $i$ indexing each view in this section.", + "bbox": [ + 75, + 90, + 472, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inpainting with Dynamic Masking. Given a rendered scene image $\\mathbf{I}$ and the text prompt $\\Gamma$ , we leverage a state-of-the-art latent 2D diffusion model [34] to generate a new image $\\overline{\\mathbf{I}}$ , where a human is inpainted into the scene image $\\mathbf{I}$ while adhering to the specified interaction and 2D scene context. In practice, we opt for the popular Stable Diffusion Inpainting model [1] as the latent diffusion implementation.", + "bbox": [ + 75, + 205, + 468, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The latent diffusion model, denoted as $\\varOmega(\\mathbf{z}_t,\\mathbf{M},\\mathbf{I},\\Gamma,t)$ , performs image inpainting by progressively denoising a noisy latent $\\mathbf{z}_t$ at each time step $t$ . During the denoising diffusion process, the binary mask $\\mathbf{M}$ defines the inpainting region in the image $\\mathbf{I}$ ; however, specifying this mask typically requires manual effort [1]. Thus, we develop a fully-automated inpainting process by automatically generating the mask. Note that using a random or fixed human mask naively can lead to incorrect inpaintings, as this often results in scene context incorrectly masked, e.g., the objects affording the desired action may be entirely masked out, leading to the generated 2D human poses to be incoherent with the scene, producing undesirable 3D HSI synthesis.", + "bbox": [ + 75, + 316, + 470, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose a masking scheme that dynamically adapts the mask through the denoising process by leveraging the internal cross-attention maps [4, 13, 42] from $\\Omega$ to propose masks. The cross-attention maps capture rich semantic correlations between image pixels and input text tokens, playing a crucial role in guiding image generation. Let $\\mathbf{A} \\in \\mathbb{R}^{hw \\times n}$ denote the cross-attention map between an image feature map of $hw$ pixels and the text prompt $\\Gamma$ with $n$ tokens, normalized row-wise using softmax. Here, $\\mathbf{A}[i,j]$ signifies the influence of the $j$ -th token on the $i$ -th pixel, and $\\mathbf{A}[:,j]$ forms a heat map of image regions to be filled with content related to the $j$ -th token.", + "bbox": [ + 75, + 518, + 468, + 699 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Using the cross-attention map $\\mathbf{A}_t$ at each time step $t$ of the diffusion process, we can dynamically derive a mask related to human tokens. Specifically, at time step $t$ , we obtain $\\mathbf{z}_{t - 1},\\mathbf{A}_t\\gets \\varOmega (\\mathbf{z}_t,\\mathbf{M}_t,\\mathbf{I},\\varGamma ,t)$ after denoising. To create a human inpainting mask $\\mathbf{M}_{t - 1}$ for time step $t - 1$ , we extract heat maps from $\\mathbf{A}_t$ corresponding to the tokens referring to the human (e.g., \"woman\" or \"man\"), followed by summation of these extracted heat maps across the tokens and binarization. At the initial time step $T$ , we initialize $\\mathbf{M}_T$ as an empty mask. Our dynamic masking approach enables the synthesis of a 2D human-scene composition image $\\bar{\\mathbf{I}}$ without the need for manually-specified human inpainting masks. We illustrate our dynamic masking scheme in Fig. 3.", + "bbox": [ + 75, + 704, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6dfdd9debcfca513b105d50197d499e7cbceac967235f155bbf62904d0ca044b.jpg", + "image_caption": [ + "Figure 3. Human inpainting with dynamic masking. Top: Given a scene image and a text prompt, a human is inpainted into the image without a mask specifying the inpainting region for latent diffusion. Bottom: The masks generated by our dynamic masking scheme based on cross-attention maps at different diffusion time steps adaptively shift to find the region of interest." + ], + "image_footnote": [], + "bbox": [ + 506, + 89, + 890, + 244 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Robust Lifting to a 3D Interaction", + "text_level": 1, + "bbox": [ + 500, + 354, + 797, + 369 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the multi-view scene images inpainted with human subjects $\\{\\bar{\\mathbf{I}}^i\\}_{i\\in [1,k]}$ , our aim is to optimize for the pose and shape parameters $(\\mathbf{R},\\mathbf{t},\\boldsymbol {\\Theta},\\boldsymbol {\\Phi})$ of a 3D human $\\mathcal{B}$ , guided by the multi-view interaction cues. We optimize for $\\mathcal{B}$ by matching it with the 2D poses extracted from $\\{\\bar{\\mathbf{I}}^i\\}$ . Since the 2D hypotheses may not be consistent across views, we formulate this as a robust optimization, simultaneously optimizing for the set of views most consistent with $\\mathcal{B}$ .", + "bbox": [ + 496, + 376, + 890, + 497 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2D Pose Estimation To distill interaction guidance for the 3D HSI synthesis, we compute a 2D pose representation for the inpainted human in each image $\\bar{\\mathbf{I}}^i$ . We use AlphaPose [6], an off-the-shelf pose estimation approach that infers a set of 2D joint positions $\\mathbf{J}^i$ and the corresponding joint confidence scores $\\mathbf{c}^i$ for the human subject in image $\\bar{\\mathbf{I}}^i$ .", + "bbox": [ + 496, + 501, + 890, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "These 2D pose hypotheses $\\{\\mathbf{J}^i,\\mathbf{c}^i\\}_{i\\in [1,k]}$ are then used to steer the interaction synthesis between the 3D human $\\mathcal{B}$ and the scene $S$ . We aim to minimize the following objective function $\\mathcal{E}_{\\mathrm{total}}$ :", + "bbox": [ + 496, + 590, + 890, + 652 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {E} _ {\\text {t o t a l}} = \\lambda_ {\\mathrm {P F}} \\mathcal {E} _ {\\mathrm {P F}} + \\lambda_ {\\mathrm {V S}} \\mathcal {E} _ {\\mathrm {V S}} + \\lambda_ {\\mathrm {B P}} \\mathcal {E} _ {\\mathrm {B P}} + \\tag {1} \\\\ \\lambda_ {\\mathrm {B S}} \\mathcal {E} _ {\\mathrm {B S}} + \\lambda_ {\\mathrm {S C}} \\mathcal {E} _ {\\mathrm {S C}} + \\lambda_ {\\mathrm {S P}} \\mathcal {E} _ {\\mathrm {S P}}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 657, + 890, + 693 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the $\\lambda$ denote scalar weights for the energy terms: pose fitting $\\mathcal{E}_{\\mathrm{PF}}$ , view selection $\\mathcal{E}_{\\mathrm{VS}}$ , body pose $\\mathcal{E}_{\\mathrm{BP}}$ , body shape $\\mathcal{E}_{\\mathrm{BS}}$ , scene penetration $\\mathcal{E}_{\\mathrm{SC}}$ , and self-penetration $\\mathcal{E}_{\\mathrm{SP}}$ .", + "bbox": [ + 496, + 700, + 890, + 746 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Robust View-consistent 3D Pose Fitting. Our primary energy term $\\mathcal{E}_{\\mathrm{PF}}$ minimizes the discrepancy between the projections of the 3D pose of $\\mathcal{B}$ and the inpainted poses from multiple views. However, 2D pose hypotheses can often be inconsistent across different views, due to the stochastic nature of diffusion models, leading to conflicting optimization signals for $\\mathcal{B}$ . To address this, we employ a robust optimization strategy that optimizes additionally for view selection weights in $\\mathcal{E}_{\\mathrm{PF}}$ that promote consistency between the 3D pose and the most consistent 2D pose hypotheses.", + "bbox": [ + 496, + 750, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20468", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We thus introduce a new set of estimizable variables $\\mathbf{w} = \\{w^{i}|w^{i}\\in [0,1]\\}_{i\\in [1,k]}$ , representing view consistency scores, and apply a robust kernel $\\rho$ to the per-view joint fitting constraints. The weights $\\mathbf{w}$ allow the solver to adaptively focus on views with consistent inpainted poses and downweight the inconsistent ones. Our pose fitting energy $\\mathcal{E}_{\\mathrm{PF}}$ with view consistency is formulated as follows:", + "bbox": [ + 76, + 90, + 472, + 196 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} _ {\\mathrm {P F}} = \\frac {\\sum_ {i} w ^ {i} \\sum_ {j} \\mathbf {c} _ {j} ^ {i} \\rho (\\Pi (\\hat {\\mathbf {J}}) _ {j} ^ {i} - \\mathbf {J} _ {j} ^ {i})}{\\sum_ {i} w ^ {i}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 208, + 468, + 247 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{J}}$ denotes the 3D joint positions of the SMPL-X model [29] differentiably computed from the pose and shape parameters $(\\mathbf{R},\\mathbf{t},\\Theta ,\\Phi)$ . The function $\\varPi(\\cdot)_j^i$ represents projection of the $j$ -th joint in the $i$ -th camera view, and $\\rho$ is the robust Geman-McClure error function [8].", + "bbox": [ + 76, + 258, + 468, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Regularizations. Below, we describe the regularization terms adopted in our optimization.", + "bbox": [ + 76, + 339, + 468, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "1) We encourage the per-view weights $\\mathbf{w}$ to focus on at least $\\tau$ views:", + "bbox": [ + 76, + 373, + 470, + 402 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} _ {\\mathrm {V S}} = \\max (\\tau - \\sum_ {i} w ^ {i}, 0). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 178, + 412, + 468, + 444 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2) To ensure a natural 3D pose, we impose the following energy term on the body pose parameter:", + "bbox": [ + 76, + 458, + 468, + 489 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} _ {\\mathrm {B P}} = \\left\\| \\boldsymbol {\\Theta} \\right\\| ^ {2} + \\mathcal {E} _ {\\mathrm {J A}} (\\hat {\\boldsymbol {\\Theta}}). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 500, + 468, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The first term is the VPoser body prior [29] regularizing the latent pose $\\Theta$ , and the second term $\\mathcal{E}_{\\mathrm{JA}}$ is a simple angle prior on the body joint rotations $\\hat{\\Theta}$ , decoded from the latent pose $\\Theta$ by VPoser, to penalize extreme bending. The formulation of $\\mathcal{E}_{\\mathrm{JA}}$ is provided in the supplementary material.", + "bbox": [ + 76, + 530, + 468, + 606 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3) We regularize the shape parameter $\\Phi$ to obtain a plausible body shape via", + "bbox": [ + 76, + 609, + 468, + 640 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} _ {\\mathrm {B S}} = \\left\\| \\boldsymbol {\\Phi} \\right\\| ^ {2}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 225, + 651, + 468, + 670 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which measures the Mahalanobis distance between $\\Phi$ and the body shape distribution used in SMPL-X [29].", + "bbox": [ + 76, + 681, + 468, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4) To ensure physical contact but also avoid penetration between the 3D human $\\mathcal{B}$ and the scene $S$ , we formulate the spatial constraints as:", + "bbox": [ + 76, + 714, + 468, + 760 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} _ {\\mathrm {S C}} = \\left\\{ \\begin{array}{l l} \\min _ {\\mathbf {v} \\in \\mathbf {V}} \\Psi (\\mathbf {v}), & \\Psi (\\mathbf {v}) > 0 \\forall \\mathbf {v} \\in \\mathbf {V} \\\\ \\sum_ {\\mathbf {v} \\in \\mathbf {V}} | \\min (\\Psi (\\mathbf {v}), 0) |, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 772, + 468, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\varPsi$ is a pre-computed signed distance field (SDF) [10] of the scene $S$ . When $\\varPsi(\\mathbf{v})$ has a negative sign, it indicates that the body vertex $\\mathbf{v}$ is located inside the nearest scene object (i.e., penetration). Conversely, a positive sign means that $\\mathbf{v}$ is positioned on the outside.", + "bbox": [ + 76, + 825, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5) Finally, to resolve penetration within the human body $\\mathcal{B}$ itself, we include a self-penetration energy $\\mathcal{E}_{\\mathrm{SP}}$ based on detecting colliding body triangles using Bounding Volume Hierarchies (BVH) [40]. We refer the reader to [2, 10, 41] for more formulation details on $\\mathcal{E}_{\\mathrm{SP}}$ .", + "bbox": [ + 496, + 90, + 893, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Iterative Refinement", + "text_level": 1, + "bbox": [ + 500, + 181, + 694, + 195 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A posed 3D human $\\mathcal{B}$ interacting with the 3D scene $S$ is generated after applying the VLM-based 2D inpainting (Sec. 3.2) and the robust 3D lifting (Sec. 3.3). To improve the synthesis and consistency of the interaction results, we employ an iterative refinement scheme over the aforementioned inpainting and 3D lifting. During refinement, we render the silhouette of the posed 3D human $\\mathcal{B}$ in each camera view, and use it as a more precise and consistent mask $\\mathbf{M}$ in the latent diffusion inpainting $\\Omega$ [34], replacing dynamic masking. By doing so, the consistency among the 2D pose hypotheses inpainted into the multi-view scene images gradually improves, thus leading to an improved 3D HSI synthesis outcome.", + "bbox": [ + 496, + 205, + 893, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 416, + 718, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We implemented GenZI with PyTorch [28]. For VLM-based 2D inpainting, we use $k = 16$ cameras for multiview rendering. We use 50 denoising steps in Stable Diffusion Inpainting with a state-of-the-art diffusion sampler [25]. For dynamic masking during inpainting, we aggregate cross-attention maps with a resolution of $16 \\times 16$ . Input text prompts $\\Gamma$ (e.g., \"sitting on a bench\") are all appended with fixed prefix \"a woman\" and fixed suffix \"wearing a white shirt and blue pants, full body\" to better constrain generation. For robust 3D lifting, we set $\\tau = 3$ , and optimize the energy $\\mathcal{E}_{\\mathrm{total}}$ using gradient descent [24] for 1.6K steps, which takes $\\sim 3$ minutes on an NVIDIA A100 GPU. In the iterative refinement, we dilate the rendered human silhouette with a kernel size of $11 \\times 11$ for mask generation, and the refinement is performed once.", + "bbox": [ + 496, + 441, + 893, + 669 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 686, + 633, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We demonstrate the effectiveness and generality of our approach GenZI on a diverse collection of 3D scene models from Sketchfab.com. We conduct both quantitative and qualitative evaluations to compare GenZI with alternative baselines approaches [10, 51] to our new task.", + "bbox": [ + 496, + 714, + 893, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. In our Sketchfab dataset, we gathered 8 large-scale 3D scenes encompassing a variety of indoor and outdoor environments with diverse geometric structures, including a realistic Venice city, a gym, and a cartoon-style food truck. We collected 4-5 text prompts per scene describing human interactions with the scene for specified approximate point locations, resulting in 38 actions for evaluation.", + "bbox": [ + 496, + 794, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "footer", + "text": "20469", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. To the best of our knowledge, there are no baselines that estimate 3D human-scene interactions based on natural language text input in a zero-shot fashion. We thus perform comparisons with related baseline approaches:", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- COINS [51] is a state-of-the-art approach estimating 3D humans in indoor 3D scans with a fixed vocabulary of actions and objects, given object segmentations. It takes as input (action, object) pairs for semantic control, and requires full supervision in its CVAE training, using captured 3D interaction data with both instance segmentations of 3D objects and action labelling. Due to being trained on a small, closed set of indoor interactions, we adapt COINS to a subset of the most similar Sketchfab actions by manually segmenting corresponding 3D objects from the scenes similar to its indoor training data and mapping the text prompts to its (action, object) input. Note that our approach GenZI does not require any 3D scene segmentations.", + "- Hassan et al. [10] perform 3D human estimation from a single RGB image. To adapt this to Sketchfab, we reuse the multi-view scene images $\\{\\bar{\\mathbf{T}}^i\\}_{i\\in [1,k]}$ inpainted with humans from our dynamic masking scheme (Sec. 3.2), where the view with the best image-text cosine similarity (i.e., CLIP score [33]) is used as the input for 3D human estimation under the known virtual camera parameters.", + "- Ours-Single View: Finally, we consider a baseline leveraging 3D body estimation from our method, but limited to only one inpainted view. The same best view image, as described above, is also as input for this baseline." + ], + "bbox": [ + 76, + 152, + 470, + 530 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. To measure 3D HSI quality, we conduct perceptual studies and compute metrics including semantic consistency, diversity, and physical plausibility.", + "bbox": [ + 75, + 535, + 468, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first carry out two perceptual studies to assess the realism and semantic accuracy of the synthesized interactions. The first is a binary-choice study, where interaction samples generated by two different methods based on the same text prompt are shown, and the participants are asked to choose the sample that is more realistic and better matched the text. The second study is a unary test, where for each interaction sample, the participants are asked to rate the realism and consistency between the shown sample and text prompt from 1 (strongly disagree) to 5 (strongly agree).", + "bbox": [ + 75, + 580, + 468, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate the semantic consistency between a synthesized 3D interaction and the input text prompt, we calculate the CLIP score [33], where the 3D interaction is re-rendered into $k$ view images, and the image-text cosine similarities from CLIP ViT-B/32 are averaged across all views.", + "bbox": [ + 75, + 733, + 468, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Additionally, we include quantitative metrics from existing works to evaluate diversity and physical plausibility. We observe that these metrics often do not reflect perceptual quality, as generated bodies can be diverse but have interpenetrations without demonstrating any physical or semantic coherence. Nevertheless, to measure the synthesis", + "bbox": [ + 75, + 810, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4a2e2d5de573cabf7141f16eee60de174c3e181152b5d9faf737bdeec3b57731.jpg", + "image_caption": [ + "(a) Binary Perceptual Study" + ], + "image_footnote": [], + "bbox": [ + 519, + 107, + 870, + 200 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/253e1bf636820898e186febf3a12fdddc4e6a75d2b7a668080b2eeffcd357e9a.jpg", + "image_caption": [ + "Figure 4. User study of 3D human-scene interaction synthesis on the Sketchfab dataset, where participants show a strong preference for the generations by our approach, in comparison with all baselines, COINS [51], Hassan et al. [10], and Ours-Single View." + ], + "image_footnote": [], + "bbox": [ + 522, + 213, + 859, + 343 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "diversity, we follow [50, 51] and cluster the SMPL-X parameters of the generated humans into 20 clusters with K-means and compute the entropy of the cluster ID histogram of all the samples. We also calculate the cluster size as the mean distance to cluster centers. For the physical plausibility, we evaluate the collision and contact between body meshes and scene meshes, following [50, 51]. The noncollision score is computed as the ratio between the number of body mesh vertices with positive SDF values (Sec. 3.3) and the number of all body mesh vertices. The contact score is defined as the ratio between the number of body meshes with scene contact and the number of all generated body meshes. A body mesh is in contact with the scene if any body mesh vertex has a non-positive SDF value.", + "bbox": [ + 496, + 421, + 890, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Comparison to Alternative Approaches", + "text_level": 1, + "bbox": [ + 500, + 645, + 834, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Evaluation. Fig. 4 shows the results of the perceptual studies collected from 30 participants across a binary study and a unary study. In the binary study, we observe that participants overwhelmingly favor the generations by our GenZI compared to all baselines – more than $87\\%$ of the time. In the unary test, the average realism rating for our interaction generations is 3.6, the highest compared to the baselines, which are all below 2.0. These perceptual results strongly indicate that our GenZI can synthesize realistic 3D humans interacting with various 3D scenes without requiring any captured 3D interaction data.", + "bbox": [ + 496, + 672, + 890, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tab. 1 presents the quantitative evaluation of semantic consistency, diversity, and physical plausibility on the full Sketchfab dataset. Our approach achieves the best semantic consistency score, echoing the strong user preference for", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20470", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3bcc7b441772f78ba6d3906bc8f2dcbd3064268668baa1e6b2852e13680f45b2.jpg", + "image_caption": [ + "\"sitting and playing piano, hands on piano keys, feet flat on the ground\"" + ], + "image_footnote": [], + "bbox": [ + 140, + 143, + 236, + 209 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0b553c4a2ca70c64b73a76a174629046e40d0d9a3da958dc2f692f91201696e8.jpg", + "image_caption": [ + "\"climbing a ladder, hands gripping side rails\"" + ], + "image_footnote": [], + "bbox": [ + 138, + 239, + 238, + 314 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/95d548d2c559076cf004401f95022d748ffcc1c3237f524156b6ae33f3f01379.jpg", + "image_caption": [ + "\"sitting and paddling a wooden boat, spreading arms\"" + ], + "image_footnote": [], + "bbox": [ + 140, + 344, + 238, + 417 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/62f8a529f4ba054258b7ee10a315fb638e2269d74295ebd845d966f3b64bd4c0.jpg", + "image_caption": [ + "\"walking on a bridge\"" + ], + "image_footnote": [], + "bbox": [ + 148, + 443, + 232, + 520 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/783e81a1c61832f4b15955a1997135e0fe0a3938cec6b9e83cbf06fc1757d1fb.jpg", + "image_caption": [ + "\"picking up a trash bag, standing, bending over\"" + ], + "image_footnote": [], + "bbox": [ + 138, + 558, + 238, + 614 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9b3d53a5b246156290a7403c0b46b64ce7c438769e9a329c849828c7d86b618c.jpg", + "image_caption": [ + "\"walking out of a door\"" + ], + "image_footnote": [], + "bbox": [ + 138, + 652, + 238, + 729 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/987e6194f7d98f4c41a880b096dca735b21210925e942c07a15fcbdd926233f0.jpg", + "image_caption": [ + "\"riding a recumbent bike, sitting\"" + ], + "image_footnote": [], + "bbox": [ + 140, + 782, + 238, + 824 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6bc881151455391110e66bfd6aea11299fc44fbd733ccf7c92aeee27f137d351.jpg", + "image_caption": [ + "Text Prompt", + "COINS" + ], + "image_footnote": [], + "bbox": [ + 277, + 104, + 411, + 207 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f01f189760f63e9e1e32e1807e107e25f04586c9ded521ce3e99d4ee76c7c9d1.jpg", + "image_caption": [ + "Hassan et al." + ], + "image_footnote": [], + "bbox": [ + 424, + 104, + 558, + 205 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e162de5ffe520c1abdb1642c60d428332d0640f4223dac9cd529b76290399663.jpg", + "image_caption": [ + "Ours-SV" + ], + "image_footnote": [], + "bbox": [ + 571, + 104, + 704, + 205 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/04b88ee2ce0ea1478cea3f8a4861966dece09382e8b8480a3865f5025ca5e9b8.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 718, + 104, + 851, + 205 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/91883f8c53c2804f6d0f3c9b8bb7995148ee28774ac3cb62a70ca1f92c65f618.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 208, + 558, + 310 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ad9c525edd94a4eaa8e95a422aa4c0ba4f25f1819dd8c4c00a80c46c06c63371.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 571, + 208, + 704, + 310 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/026ee0159bacf51b95ee2369cf4d7fe81f0fcf6790f735a950cecc523592230c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 208, + 851, + 310 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/585ae1923e61e26e622c1d05e5dd8861825ea6793e0019bef976b875f15387df.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 311, + 558, + 415 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/347c82f5728a9ca6ac14e404b6d8d3493f37eef3c244e6af5e9bc08313beade7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 571, + 311, + 704, + 415 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0c9307135aada9be1e64f6dfe7f488493a14c300ae4bac7aa8cfca63c63467fb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 311, + 851, + 415 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/799fe841691ee3c412d33497b7e620a8087bcdb53666489310b98497dc74e68a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 277, + 417, + 411, + 521 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/16c317b5dbba0f25b8d1c3938f82bf5305b39b7af39f90d4fb648a4fe027ffac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 417, + 558, + 520 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1eb9f36e63879ba59c68a502439c6ddaf41925c600bc7d8d93fcc7879f7d98bf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 571, + 419, + 704, + 520 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8d23c4ee33ef316b5a6e5cd0e0658ba908b42ad2066dfbbcca2f4cd9cee0cd16.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 419, + 851, + 520 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1f7eb8ea742aae73c555ff5f2fba8cb5a1f4df94529a6d4674cea1c82cea15e9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 522, + 558, + 625 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/005844162b292bdf668d22ac84f462409c58a86c7f1e37b8dd6b26d36a17d9cc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 571, + 523, + 704, + 625 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/49e321fb2240ec8dd772d47048522b37fa8e2c947fb811d5b255e81f8d6e42ac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 523, + 851, + 625 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9145aaf37bf9c5c3ac2505f2c4b5bfd4304c0238bf003297bd9facc2f7212263.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 627, + 558, + 729 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cd2327af6bfd58ee031f311cddba687dbf504298e267d65faadc5b9d54370eb0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 571, + 627, + 704, + 729 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c68d305062d3c01a8d9e6ddc2b32f134bf26dbcb7b1fd2bf024d8b048b3a7967.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 627, + 851, + 729 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/195c72893bf9b901b89ddcf561ef1990dc1f4847071cb3651c225dfa30ab8004.jpg", + "image_caption": [ + "Figure 5. Qualitative results on the Sketchfab dataset. Our GenZI synthesizes more realistic 3D human-scene interactions and generalizes better across diverse scene types, compared to the baselines COINS [51], Hassan et al. [10], and Ours-Single View. For COINS, we show the used (action, object) labels from its closed set of indoor interactions; its closed setting can lead to degraded results from out-of-distribution object classes (e.g., curved bridge deck as the floor, chair at a different height or shape than those in the training set)." + ], + "image_footnote": [], + "bbox": [ + 277, + 732, + 411, + 835 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a398ac2ebd84c478cc3bfb296599c800f0137c4c802293d9fdbf47f12a443de8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 732, + 558, + 835 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e85d45af699882bcd1def9e555d2cd2bf10139ef73e17d47ffa1c82237fb6f36.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 571, + 732, + 704, + 835 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/19c9b6eeaceb1c10ba439b354a5185d46cf6c8f4c1f4a9e62a2333cf0486d34c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 732, + 851, + 835 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20471", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/05c45a05af4b7b36df8bf598d94c8ce986e97a4de409f166d6c0e8741255cc5a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSemanticsDiversityPhysical Plausibility
CLIP ↑Entropy ↑Cluster Size ↑Non-collision ↑Contact ↑
Hassan et al. [10]0.25982.70141.19070.88240.9669
Ours-SV0.26132.64521.58130.97651.0000
Ours0.27102.73041.05000.97670.9868
", + "bbox": [ + 78, + 88, + 467, + 138 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/09aae217713a2af14007ec9f06693fb8b26a61d24638003c63356b2308927337.jpg", + "table_caption": [ + "Table 1. Quantitative comparisons on the Sketchfab dataset. Our approach achieves the best semantic consistency, diversity entropy, and non-collision scores, with the contact score on par. Note that single view methods Hassan et al. [10] and Ours-SV tend to produce increased diversity at the cost of semantic plausibility." + ], + "table_footnote": [], + "table_body": "
w/o DMw/o VCw/o IROurs
CLIP ↑0.26390.26940.26640.2710
", + "bbox": [ + 133, + 227, + 405, + 253 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Ablation study on the Sketchfab dataset. The semantic consistency of 3D interaction generations degrades without dynamic masking (DM), view consistency (VC), or iterative refinement (IR), compared to our full approach.", + "bbox": [ + 75, + 258, + 468, + 315 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e1f5615d87b51a9a6de42d9775a816c93365f793d43cda462f9937ee3c946080.jpg", + "image_caption": [ + "Figure 6. Visualization of our method ablations on Sketchfab dataset for the input text: \"sitting on a bar stool\". Without dynamic masking (DM) or view consistency (VC), the person floats above the middle stool. Without iterative refinement (IR), the person penetrates the stool. Our full approach results in a more realistic synthesis." + ], + "image_footnote": [], + "bbox": [ + 78, + 329, + 467, + 415 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "GenZI in the perceptual studies. We note that since both Hassan et al. [10] and Ours-Single View operate on single inpainted view samples, results can be very diverse but lack 3D plausibility. This suggests that the CLIP score is a more reliable metric for assessing the HSI synthesis quality, compared to the diversity and physical plausibility metrics, which reflect less about the generation realism. Nevertheless, our approach has the best diversity entropy and noncollision scores, with the contact score on par.", + "bbox": [ + 75, + 513, + 467, + 650 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative Evaluation. We show qualitative comparisons in Fig. 5. COINS is severely limited by its training on the closed set of indoor interactions, and thus fails to generalize to outdoor scenes and unseen objects (e.g., no curved floors exist during training, and limited sets of heights and shapes of chairs). As Hassan et al. [10] and Ours-Single View operate from single views, they both suffer from insufficient pose constraints from other views for plausible interaction generation. In contrast, our approach demonstrates high flexibility and generality to a diverse set of 3D indoor and outdoor scenes by leveraging large VLMs to imagine multiview interaction hypotheses and then robust 3D lifting.", + "bbox": [ + 75, + 654, + 467, + 835 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 845, + 240, + 861 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct ablation studies on the Sketchfab dataset to validate the effectiveness of our proposed dynamic masking", + "bbox": [ + 75, + 869, + 467, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "scheme (Sec. 3.2), robust 3D lifting with view consistency (Sec. 3.3), and iterative refinement (Sec. 3.4). Results are presented in Fig. 6 and Tab. 2.", + "bbox": [ + 498, + 90, + 890, + 136 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Dynamic Masking (DM). We replace our dynamic masking, used during latent diffusion inpainting, with random masking. We sample a random mask around the image center covering at least $30\\%$ of the image area, and use it as a fixed mask input to the latent diffusion model $\\Omega$ . We observe that random masking results in noticeably worse quality of interaction synthesis (Tab. 2), and incoherence with the scene (Fig. 6, floating on the stool). This indicates that our dynamic masking is effective in incorporating sufficient scene context for human inpainting.", + "bbox": [ + 496, + 138, + 890, + 291 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "View Consistency (VC). We evaluate the role of view consistency in robust 3D lifting by fixing the estimizable scores to $\\mathbf{w} = \\mathbf{1}$ . Using all inpainted views leads to averaged, less expressive 3D human poses (Tab. 2, Fig. 6) due to potential inconsistent 2D pose hypotheses across views. By allowing the solver to adaptively focus on views with consistent inpaintings, our approach generates more realistic 3D HSIs.", + "bbox": [ + 496, + 294, + 890, + 400 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Iterative Refinement (IR). Finally, we show the effectiveness of iterative refinement by applying our VLM-based 2D inpainting and robust 3D lifting only once. Tab. 2 and Fig. 6 show that iterative refinement improves synthesis quality.", + "bbox": [ + 496, + 402, + 890, + 464 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Our approach is limited by the inpainting capability of latent diffusion models to imagine possible 2D human-scene compositions, and the diffusion models are also known to be slow at inference time due to their iterative nature [14]. Nevertheless, we believe that our approach can directly benefit from the rapid advancement of VLMs for improved HSI synthesis.", + "bbox": [ + 496, + 467, + 890, + 573 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 585, + 617, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have presented the first approach to synthesize general 3D human-scene interactions guided by text inputs. Key to our approach is effective distillation of knowledge from large vision-language models, enabling generating 3D humans in scenes without requiring any 3D interaction data for training. We leverage these powerful vision-language models to generate hypotheses for inpainted 2D human-scene interactions. We then formulate a robust optimization to lift the hypotheses to 3D in a view-consistent fashion by simultaneously optimizing for the most informative 2D hypotheses. Our approach is flexible and can be applied to general scene settings for a variety of actions. We believe this opens up new opportunities for 3D understanding without requiring expensive capture of 3D/4D data.", + "bbox": [ + 496, + 609, + 890, + 820 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This project is funded by the ERC Starting Grant SpatialSem (101076253), and the German Research Foundation (DFG) Grant \"Learning How to Interact with Scenes through Part-Based Understanding\", and supported in part by a Google research gift.", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20472", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Stable Diffusion v2. https://huggingface.co/stabilityai/stable-diffusion-2-inpainting.Accessed:2023-09-27.4", + "[2] Luca Ballan, Aparna Taneja, Jürgen Gall, Luc Van Gool, and Marc Pollefeys. Motion capture of hands in action using discriminative salient points. In ECCV, 2012. 5", + "[3] Zhongang Cai, Wanqi Yin, Ailing Zeng, Chen Wei, Qingping Sun, Yanjun Wang, Hui En Pang, Haiyi Mei, Mingyuan Zhang, Lei Zhang, et al. SMPLer-X: Scaling up expressive human pose and shape estimation. arXiv, 2023. 2", + "[4] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-Excite: Attention-based semantic guidance for text-to-image diffusion models. ACM TOG, 2023. 4", + "[5] Vasileios Choutas, Georgios Pavlakos, Timo Bolkart, Dimitrios Tzionas, and Michael J Black. Monocular expressive body regression through body-driven attention. In ECCV, 2020. 2", + "[6] Hao-Shu Fang, Jiefeng Li, Hongyang Tang, Chao Xu, Haoyi Zhu, Yuliang Xiu, Yong-Lu Li, and Cewu Lu. Alpha-Pose: Whole-body regional multi-person pose estimation and tracking in real-time. IEEE TPAMI, 2022. 4", + "[7] Dariu M Gavrila. The visual analysis of human movement: A survey. CVIU, 1999. 2", + "[8] Stuart Geman and Donald E. McClure. Statistical methods for tomographic image reconstruction. 1987. 5", + "[9] James J Gibson. The ecological approach to the visual perception of pictures. *Leonardo*, 11(3):227-235, 1978. 1", + "[10] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J. Black. Resolving 3D human pose ambiguities with 3D scene constraints. In ICCV, 2019. 1, 2, 5, 6, 7, 8", + "[11] Mohamed Hassan, Partha Ghosh, Joachim Tesch, Dimitrios Tzionas, and Michael J. Black. Populating 3D scenes by learning human-scene interaction. In CVPR, 2021. 1, 2", + "[12] Mohamed Hassan, Yunrong Guo, Tingwu Wang, Michael Black, Sanja Fidler, and Xue Bin Peng. Synthesizing physical character-scene interactions. In SIGGRAPH, 2023. 2", + "[13] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-prompt image editing with cross-attention control. In ICLR, 2023. 4", + "[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 8", + "[15] Lukas Hollein, Ang Cao, Andrew Owens, Justin Johnson, and Matthias Nießner. Text2Room: Extracting textured 3d meshes from 2d text-to-image models. In ICCV, 2023. 3", + "[16] Siyuan Huang, Zan Wang, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, and Song-Chun Zhu. Diffusion-based generation, optimization, and planning in 3D scenes. In CVPR, 2023. 2", + "[17] Kurt Koffka. Principles of gestalt psychology. 1935. 1", + "[18] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In ICCV, 2019. 2" + ], + "bbox": [ + 78, + 114, + 467, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[19] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. BLIP: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 2", + "[20] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 2, 3", + "[21] Xueting Li, Sifei Liu, Kihwan Kim, Xiaolong Wang, Ming-Hsuan Yang, and Jan Kautz. Putting humans in a scene: Learning affordance in 3D indoor environments. In CVPR, 2019. 2", + "[22] Jing Lin, Ailing Zeng, Haoqian Wang, Lei Zhang, and Yu Li. One-stage 3d whole-body mesh recovery with component aware transformer. In CVPR, pages 21159-21168, 2023. 2", + "[23] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. SMPL: A skinned multiperson linear model. ACM TOG, 34(6):1-16, 2015. 2, 3", + "[24] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 5", + "[25] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. DPM-Solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv, 2022. 5", + "[26] Thomas B Moeslund, Adrian Hilton, and Volker Krüger. A survey of advances in vision-based human motion capture and analysis. CVIU, 2006. 2", + "[27] Aron Monszpart, Paul Guerrero, Duygu Ceylan, Ersin Yumer, and Niloy J Mitra. iMapper: interaction-guided scene mapping from monocular videos. ACM Transactions On Graphics (TOG), 38(4):1-15, 2019. 2", + "[28] A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. Kopf, E. Yang, Z. DeVito, M. Raison, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, and S. Chintala. PyTorch: An imperative style, high-performance deep learning library. NeurIPS, 2019. 5", + "[29] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3D hands, face, and body from a single image. In CVPR, 2019. 2, 3, 5", + "[30] R. Po, W. Yifan, V. Golyanik, K. Aberman, J. T. Barron, A. H. Bermano, E. R. Chan, T. Dekel, A. Holynski, A. Kanazawa, C. K. Liu, L. Liu, B. Mildenhall, M. Nießner, B. Ommer, C. Theobalt, P. Wonka, and G. Wetzstein. State of the art on diffusion models for visual computing. arXiv, 2023. 2", + "[31] Ronald Poppe. Vision-based human motion analysis: An overview. CVIU, 2007. 2", + "[32] Sergey Prokudin, Christoph Lassner, and Javier Romero. Efficient learning on point clouds with basis point sets. In ICCV, 2019. 2", + "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3, 6" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20473", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3, 4, 5", + "[35] Yu Rong, Takaaki Shiratori, and Hanbyul Joo. FrankMocap: A monocular 3d whole-body pose estimation system via regression and integration. In ICCV, 2021. 2", + "[36] David Rozenberszki, Or Litany, and Angela Dai. Language-grounded indoor 3d semantic segmentation in the wild. In ECCV, 2022. 3", + "[37] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 2022. 2, 3", + "[38] Nikolaos Sarafianos, Bogdan Boteanu, Bogdan Ionescu, and Ioannis A Kakadiaris. 3D human pose estimation: A review of the literature and analysis of covariates. CVIU, 2016. 2", + "[39] Manolis Savva, Angel X Chang, Pat Hanrahan, Matthew Fisher, and Matthias Nießner. PiGraphs: learning interaction snapshots from observations. ACM TOG, 35(4):1-12, 2016. 2", + "[40] Matthias Teschner, Stefan Kummerle, Bruno Heidelberger, Gabriel Zachmann, Laks Raghupathi, Arnulph Fuhrmann, M-P Cani, François Faure, Nadia Magnenat-Thalmann, Wolfgang Strasser, et al. Collision detection for deformable objects. In CGF, 2005. 5", + "[41] Dimitrios Tzionas, Luca Ballan, Abhilash Srikantha, Pablo Aponte, Marc Pollefeys, and Juergen Gall. Capturing hands in action using discriminative salient points and physics simulation. IJCV, 2016. 5", + "[42] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 4", + "[43] Jingbo Wang, Yu Rong, Jingyuan Liu, Sijie Yan, Dahua Lin, and Bo Dai. Towards diverse and natural scene-aware 3D human motion synthesis. In CVPR, 2022. 2", + "[44] Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, Wei Liang, and Siyuan Huang. HUMANISE: Language-conditioned human motion generation in 3D scenes. NeurIPS, 2022. 2", + "[45] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In CVPR, 2023. 3", + "[46] Haibiao Xuan, Xiongzheng Li, Jinsong Zhang, Hongwen Zhang, Yebin Liu, and Kun Li. Narrator: Towards natural control of human-scene interaction generation via relationship reasoning. arXiv, 2023. 1", + "[47] Yufei Ye, Xueting Li, Abhinav Gupta, Shalini De Mello, Stan Birchfield, Jiaming Song, Shubham Tulsiani, and Sifei Liu. Affordance diffusion: Synthesizing hand-object interactions. In CVPR, 2023. 3", + "[48] Hongwen Zhang, Yating Tian, Yuxiang Zhang, Mengcheng Li, Liang An, Zhenan Sun, and Yebin Liu. PyMAF-X: Towards well-aligned full-body model regression from monocular images. IEEE TPAMI, 2023. 2" + ], + "bbox": [ + 78, + 92, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Siwei Zhang, Yan Zhang, Qianli Ma, Michael J Black, and Siyu Tang. Place: Proximity learning of articulation and contact in 3D environments. In 3DV, 2020. 2", + "[50] Yan Zhang, Mohamed Hassan, Heiko Neumann, Michael J. Black, and Siyu Tang. Generating 3D people in scenes without people. In CVPR, 2020. 1, 2, 6", + "[51] Kaifeng Zhao, Shaofei Wang, Yan Zhang, Thabo Beeler, and Siyu Tang. Compositional human-scene interaction synthesis with semantic control. In ECCV, 2022. 1, 2, 5, 6, 7", + "[52] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In CVPR, 2019. 3", + "[53] Yuxiao Zhou, Marc Habermann, Ikhsanul Habibie, Ayush Tewari, Christian Theobalt, and Feng Xu. Monocular real-time full body capture with inter-part correlations. In CVPR, 2021. 2" + ], + "bbox": [ + 501, + 92, + 890, + 316 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20474", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/GenZI_ Zero-Shot 3D Human-Scene Interaction Generation/3823a91d-f3af-49d1-a8df-4cb9a1c22737_model.json b/2024/GenZI_ Zero-Shot 3D Human-Scene Interaction Generation/3823a91d-f3af-49d1-a8df-4cb9a1c22737_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d7bdf28b6e740d156a3dff5bfb2b13319f71a8b6 --- /dev/null +++ b/2024/GenZI_ Zero-Shot 3D Human-Scene Interaction Generation/3823a91d-f3af-49d1-a8df-4cb9a1c22737_model.json @@ -0,0 +1,2552 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.183, + 0.131, + 0.787, + 0.151 + ], + "angle": 0, + "content": "GenZI: Zero-Shot 3D Human-Scene Interaction Generation" + }, + { + "type": "text", + "bbox": [ + 0.356, + 0.182, + 0.61, + 0.218 + ], + "angle": 0, + "content": "Lei Li Angela Dai Technical University of Munich" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.257, + 0.266, + 0.272 + ], + "angle": 0, + "content": "\"riding a motorcycle, sitting\"" + }, + { + "type": "text", + "bbox": [ + 0.296, + 0.252, + 0.476, + 0.277 + ], + "angle": 0, + "content": "\"hands knocking on the closed door, facing the door, standing\"" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.252, + 0.674, + 0.277 + ], + "angle": 0, + "content": "\"stretching legs and sitting on the saddle on a standing cow\"" + }, + { + "type": "text", + "bbox": [ + 0.701, + 0.252, + 0.872, + 0.277 + ], + "angle": 0, + "content": "\"picking up dumbbells on a shelf, standing, bending over\"" + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.281, + 0.282, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.285, + 0.282, + 0.484, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.282, + 0.685, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.282, + 0.886, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.439, + 0.893, + 0.496 + ], + "angle": 0, + "content": "Figure 1. Given an arbitrary 3D scene, GenZI can synthesize virtual humans interacting with the 3D environment at specified locations from a brief text description. Our approach does not require any 3D human-scene interaction training data or 3D learning. By distilling interaction priors from powerful 2D vision-language models, we optimize for 3D human-scene interaction synthesis in a flexible fashion, with simple language-based control and high generality to various types of scene environments." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.519, + 0.313, + 0.535 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.547, + 0.473, + 0.865 + ], + "angle": 0, + "content": "Can we synthesize 3D humans interacting with scenes without learning from any 3D human-scene interaction data? We propose GenZI1, the first zero-shot approach to generating 3D human-scene interactions. Key to GenZI is our distillation of interaction priors from large vision-language models (VLMs), which have learned a rich semantic space of 2D human-scene compositions. Given a natural language description and a coarse point location of the desired interaction in a 3D scene, we first leverage VLMs to imagine plausible 2D human interactions inpainted into multiple rendered views of the scene. We then formulate a robust iterative optimization to synthesize the pose and shape of a 3D human model in the scene, guided by consistency with the 2D interaction hypotheses. In contrast to existing learning-based approaches, GenZI circumvents the conventional need for captured 3D interaction data, and allows for flexible control of the 3D interaction synthesis with easy-to-use text prompts. Extensive experiments show that our zero-shot approach has high flexibility and generality, making it applicable to diverse scene types, including both indoor and outdoor environments." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.519, + 0.631, + 0.535 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.549, + 0.892, + 0.701 + ], + "angle": 0, + "content": "Scenes are typically constructed to enable human interactions with the environment, like sitting on a couch, playing a piano, or opening a mailbox. Understanding these interactions between humans and scenes, also known as affordances [9, 17], has gained increasing attention in the fields of computer vision and graphics. In particular, achieving controllable and generalizable synthesis of human-scene interactions (or HSIs) holds immense potential for various applications, such as robotics, architectural design, video games, and virtual reality experiences, among many others." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Generating realistic humans interacting with a 3D scene is a challenging task, requiring holistic semantic understanding of both the environment and possible human actions therein. Existing approaches to HSI synthesis [11, 46, 50, 51] rely heavily on supervised training using meticulously captured data of real people interacting in 3D environments. Unfortunately, collecting large-scale datasets of 3D scenes and human interactions is exorbitantly difficult. It not only demands accurate tracking and reconstruction of both people and their environments, but also needs to ensure sufficient diversity in subjects and scenes. Existing HSI datasets currently contain very limited quantities of scenes and actions, for example, PROX [10, 51] only consisting" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.352, + 0.902 + ], + "angle": 0, + "content": "1Project page: craigileli.github.io/projects/genzi" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "20465" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.168 + ], + "angle": 0, + "content": "of 12 indoor scenes and humans interacting with 11 object categories. The scarcity of ground truth data for supervision has thus strongly limited the applicability and generalization of the learning-based approaches for synthesizing diverse sets of actions in arbitrary 3D scenes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.169, + 0.47, + 0.333 + ], + "angle": 0, + "content": "We thus consider an alternative perspective to HSI synthesis and pose the following question: Can we achieve plausible HSI synthesis without using any captured 3D interaction data? To this end, we present a novel zero-shot approach to 3D HSI generation. We propose the first method to leverage the powerful capabilities exhibited by recent vision-language models (VLMs) [19, 20, 30, 33, 34, 37] to synthesize plausible 2D images of human interactions, and introduce a robust optimization to distill inferred 2D pose information into 3D human synthesis in a 3D scene." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.337, + 0.47, + 0.607 + ], + "angle": 0, + "content": "More concretely, given a 3D scene, a text prompt and a coarse point location of the desired interaction, GenZI optimizes for the pose and shape of a 3D human performing the action in the scene, guided by a large VLM [34]. We first leverage the VLM to imagine possible 2D humans by inpainting images from multiple rendered views of the scene. We automate this 2D human insertion process with a dynamic masking scheme that automatically updates proposed masks through the inpainting process, eliminating the need for manual specifications of human inpainting regions. We then lift these 2D interaction hypotheses to 3D and optimize for a parametric 3D human body model [23, 29] that is most consistent with the 2D pose guidance. We further refine the generated 3D human in the scene by iterating through the VLM-based 2D inpainting and robust 3D lifting stages. We demonstrate the flexibility and generality of GenZI in various types of 3D environments (Fig. 1), encompassing both indoor and outdoor scenes." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.61, + 0.397, + 0.624 + ], + "angle": 0, + "content": "In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.626, + 0.468, + 0.7 + ], + "angle": 0, + "content": "- We introduce GenZI, the first zero-shot approach to generating realistic 3D humans interacting with a 3D scene from natural language prompts. GenZI does not require any supervision from 3D interaction data, thus enabling flexible synthesis across diverse scenes and actions." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.702, + 0.468, + 0.761 + ], + "angle": 0, + "content": "- We propose a dynamically-masked inpainting scheme that allows for the synthesis of plausible 2D human-scene compositions via VLMs without requiring manually-specified human inpainting masks." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.762, + 0.468, + 0.806 + ], + "angle": 0, + "content": "- We develop a robust 3D pose optimization to lift various inferred images of human interactions to a view-consistent, realistic 3D HSI synthesis." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.626, + 0.468, + 0.806 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.825, + 0.218, + 0.84 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "3D Human-Scene Interaction Synthesis. Synthesizing humans in scenes is an important, challenging task in computer vision and graphics, as it models complex high-level" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "semantic understanding such as affordances and interactions. Existing approaches for human-scene interaction synthesis [11, 12, 16, 21, 27, 39, 43, 44, 50, 51] focus on learning priors from available collected data of people interacting with 3D indoor scenes [10] in a supervised manner." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.169, + 0.892, + 0.472 + ], + "angle": 0, + "content": "The early work PiGraphs [39] introduced a probabilistic graphical model of human-object interactions, generating static human-object placements by pose sampling and model retrieval. Zhang et al. [50] introduced a generative model of human-scene interaction, using a conditional variational autoencoder (CVAE) to model the distribution of 3D human poses conditioned on scene depth and semantics. PLACE [49] explicitly represented human-scene interactions with Basis Point Sets (BPS) encoding [32], training a CVAE to synthesize such representations for generating natural human poses and contact relations within scenes. POSA [11] proposed an ego-centric representation by augmenting the SMPL-X model [29] with contact labels and scene semantics, using a CVAE conditioned on SMPL-X vertex positions to model potential interactions. Recently, COINS [51] presented a method for compositional human-scene interaction synthesis with high-level semantic control, using transformer-based CVAEs conditioned on provided 3D objects and interaction semantics to regress human body poses and contact features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.473, + 0.892, + 0.669 + ], + "angle": 0, + "content": "However, these works all require supervision from 3D human-scene interactions – such datasets are difficult and expensive to capture and annotate, for instance, the widely used PROX dataset [10, 51] only has 8 reconstructed indoor scenes for training and 4 for testing, capturing human interactions with 11 object categories. As such, these methods are limited to in-domain synthesis, prohibiting applicability to general 3D scene settings with arbitrary objects and arrangements. In this work, we deviate from the learning-based approach and propose to take advantage of established large vision-language models [20, 33, 34, 37] for human-scene interaction synthesis, thereby bypassing the requirement for data capture and 3D learning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.901 + ], + "angle": 0, + "content": "3D Human Estimation from RGB Images. Over the past decades, significant progress has been made in 3D human estimation from RGB images [7, 26, 31, 38]. Among those, the prominent work SMPLify-X [29] fits the SMPL-X model to 2D joints estimated from a single image through optimization. Hassan et al. [10] build upon SMPLify-X and estimate 3D human poses by incorporating additional physical contact and penetration constraints between human and 3D scene. Learning-based approaches [3, 5, 18, 22, 35, 48, 53] have received much research attention in recent years towards tackling 3D human estimation from monocular images with improved hand and face estimation. Our work leverages 2D pose reasoning from multiple different 2D view hypotheses, and proposes a robust 3D formulation for aggregating the various 2D hypotheses to a consistent" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20466" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.891, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.383, + 0.893, + 0.466 + ], + "angle": 0, + "content": "Figure 2. GenZI distills information from vision-language model for 3D human-scene interaction. We first leverage large vision-language models to synthesize possible 2D humans interactions with the 3D scene \\( S \\) by employing latent diffusion inpainting [34] on multiple rendered views of the environment at location \\( \\mathbf{p} \\) using our dynamic masking scheme to automatically estimate inpainting masks. We then lift these 2D hypotheses to 3D in a robust optimization for a 3D parametric body model \\( \\mathcal{B} \\) (SMPL-X [29]) that is most consistent with detected 2D poses in the inpainted 2D hypotheses. This produces a semantically consistent interaction that respects the scene context, without requiring any 3D human-scene interaction data." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.482, + 0.375, + 0.497 + ], + "angle": 0, + "content": "3D human body interacting with a 3D scene." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.501, + 0.471, + 0.652 + ], + "angle": 0, + "content": "Distilling Prior from Vision-Language Models. Recent advances in powerful vision-language models [20, 33, 34, 37] have also inspired various works aiming to distill information learned through the models to various tasks, including 2D panoptic segmentation [45], 3D semantic segmentation [36], 3D scene generation [15], and synthesis of images with hand-object interactions [47]. Our approach leverages the 2D generative capacity of these models to convey information about possible human-scene interactions, which we then lift to a 3D, consistent interaction." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.669, + 0.169, + 0.685 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.695, + 0.188, + 0.709 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.719, + 0.47, + 0.824 + ], + "angle": 0, + "content": "Our objective is to synthesize plausible 3D humans interacting with a 3D scene, guided by input text descriptions, in the absence of 3D interaction data capture for learning. We present GenZI, a novel optimization-based multi-view approach that leverages large VLMs to infer spatial relations of interactions between a human and the scene. Fig. 2 shows an illustration of our approach." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.901 + ], + "angle": 0, + "content": "GenZI takes as input a 3D scene \\(S\\), a text prompt \\(\\Gamma\\) describing the desired interaction, and an approximate point location \\(\\mathbf{p} \\in \\mathbb{R}^3\\) in the scene around which the interaction should occur. Our approach generates a posed 3D human \\(\\mathcal{B}\\) as output, performing the specified action in the scene. We" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.482, + 0.893, + 0.633 + ], + "angle": 0, + "content": "adopt SMPL-X [23, 29] to parameterize the 3D human \\(\\mathcal{B}\\) as it provides a fully differentiable function mapping a set of pose and shape parameters \\((\\mathbf{R},\\mathbf{t},\\Theta ,\\Phi)\\) to a 3D human mesh with vertices \\(\\mathbf{V}\\) and faces \\(\\mathbf{F}\\). We thus optimize for \\((\\mathbf{R},\\mathbf{t},\\Theta ,\\Phi)\\) characterizing \\(\\mathcal{B}\\) in the scene \\(S\\), with \\(\\mathbf{R} \\in \\mathbb{R}^6\\) denoting the global orientation as a continuous rotation representation [52], \\(\\mathbf{t} \\in \\mathbb{R}^3\\) the global translation, \\(\\Theta \\in \\mathbb{R}^{32}\\) the body pose in the latent space of VPoser [29] which decodes to body joint rotations \\(\\hat{\\Theta} \\in \\mathbb{R}^{21\\times 3}\\), and \\(\\Phi \\in \\mathbb{R}^{10}\\) representing the body shape as blend shape coefficients." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.634, + 0.895, + 0.801 + ], + "angle": 0, + "content": "GenZI synthesizes the desired interaction between the 3D human \\(\\mathcal{B}\\) and the scene \\(S\\) by distilling information from VLMs through 2D human inpainting followed by robust 3D lifting. We first generate a collection of plausible 2D human-scene compositions by employing a large VLM to inpaint humans into multiple rendered views of the scene \\(S\\) (Sec. 3.2). We then introduce a robust 3D lifting procedure that optimizes for the pose and shape of the human \\(\\mathcal{B}\\), guided by consistency with the 2D interaction hypotheses (Sec. 3.3). We further refine the posed 3D human \\(\\mathcal{B}\\) by iteratively updating the 2D inpaintings and 3D optimization." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.812, + 0.888, + 0.829 + ], + "angle": 0, + "content": "3.2. Inpainting Multi-view Interaction Hypotheses" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.836, + 0.892, + 0.882 + ], + "angle": 0, + "content": "We first leverage a VLM to generate 2D hypotheses of potential human interactions in the scene by automatically inpainting humans into multiple rendered views of \\( S \\)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Multi-view Rendering. To capture scene context for 2D" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20467" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.198 + ], + "angle": 0, + "content": "human inpainting, we render multiple views of the 3D scene \\( S \\) from \\( k \\) virtual cameras looking at \\( \\mathbf{p} \\). Cameras are randomly sampled on a hemisphere and filtered according to the visibility of \\( \\mathbf{p} \\); we refer to the supplementary material for additional camera setup details. We denote the rendered scene images as \\( \\{\\mathbf{I}^i\\}_{i\\in [1,k]} \\). To simplify notation, we omit the superscript \\( i \\) indexing each view in this section." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.206, + 0.47, + 0.313 + ], + "angle": 0, + "content": "Inpainting with Dynamic Masking. Given a rendered scene image \\(\\mathbf{I}\\) and the text prompt \\(\\Gamma\\), we leverage a state-of-the-art latent 2D diffusion model [34] to generate a new image \\(\\overline{\\mathbf{I}}\\), where a human is inpainted into the scene image \\(\\mathbf{I}\\) while adhering to the specified interaction and 2D scene context. In practice, we opt for the popular Stable Diffusion Inpainting model [1] as the latent diffusion implementation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.317, + 0.471, + 0.514 + ], + "angle": 0, + "content": "The latent diffusion model, denoted as \\(\\varOmega(\\mathbf{z}_t,\\mathbf{M},\\mathbf{I},\\Gamma,t)\\), performs image inpainting by progressively denoising a noisy latent \\(\\mathbf{z}_t\\) at each time step \\(t\\). During the denoising diffusion process, the binary mask \\(\\mathbf{M}\\) defines the inpainting region in the image \\(\\mathbf{I}\\); however, specifying this mask typically requires manual effort [1]. Thus, we develop a fully-automated inpainting process by automatically generating the mask. Note that using a random or fixed human mask naively can lead to incorrect inpaintings, as this often results in scene context incorrectly masked, e.g., the objects affording the desired action may be entirely masked out, leading to the generated 2D human poses to be incoherent with the scene, producing undesirable 3D HSI synthesis." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.519, + 0.47, + 0.7 + ], + "angle": 0, + "content": "We propose a masking scheme that dynamically adapts the mask through the denoising process by leveraging the internal cross-attention maps [4, 13, 42] from \\(\\Omega\\) to propose masks. The cross-attention maps capture rich semantic correlations between image pixels and input text tokens, playing a crucial role in guiding image generation. Let \\(\\mathbf{A} \\in \\mathbb{R}^{hw \\times n}\\) denote the cross-attention map between an image feature map of \\(hw\\) pixels and the text prompt \\(\\Gamma\\) with \\(n\\) tokens, normalized row-wise using softmax. Here, \\(\\mathbf{A}[i,j]\\) signifies the influence of the \\(j\\)-th token on the \\(i\\)-th pixel, and \\(\\mathbf{A}[:,j]\\) forms a heat map of image regions to be filled with content related to the \\(j\\)-th token." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Using the cross-attention map \\(\\mathbf{A}_t\\) at each time step \\(t\\) of the diffusion process, we can dynamically derive a mask related to human tokens. Specifically, at time step \\(t\\), we obtain \\(\\mathbf{z}_{t - 1},\\mathbf{A}_t\\gets \\varOmega (\\mathbf{z}_t,\\mathbf{M}_t,\\mathbf{I},\\varGamma ,t)\\) after denoising. To create a human inpainting mask \\(\\mathbf{M}_{t - 1}\\) for time step \\(t - 1\\), we extract heat maps from \\(\\mathbf{A}_t\\) corresponding to the tokens referring to the human (e.g., \"woman\" or \"man\"), followed by summation of these extracted heat maps across the tokens and binarization. At the initial time step \\(T\\), we initialize \\(\\mathbf{M}_T\\) as an empty mask. Our dynamic masking approach enables the synthesis of a 2D human-scene composition image \\(\\bar{\\mathbf{I}}\\) without the need for manually-specified human inpainting masks. We illustrate our dynamic masking scheme in Fig. 3." + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.09, + 0.891, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.255, + 0.894, + 0.34 + ], + "angle": 0, + "content": "Figure 3. Human inpainting with dynamic masking. Top: Given a scene image and a text prompt, a human is inpainted into the image without a mask specifying the inpainting region for latent diffusion. Bottom: The masks generated by our dynamic masking scheme based on cross-attention maps at different diffusion time steps adaptively shift to find the region of interest." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.355, + 0.799, + 0.371 + ], + "angle": 0, + "content": "3.3. Robust Lifting to a 3D Interaction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.377, + 0.892, + 0.498 + ], + "angle": 0, + "content": "Given the multi-view scene images inpainted with human subjects \\(\\{\\bar{\\mathbf{I}}^i\\}_{i\\in [1,k]}\\), our aim is to optimize for the pose and shape parameters \\((\\mathbf{R},\\mathbf{t},\\boldsymbol {\\Theta},\\boldsymbol {\\Phi})\\) of a 3D human \\(\\mathcal{B}\\), guided by the multi-view interaction cues. We optimize for \\(\\mathcal{B}\\) by matching it with the 2D poses extracted from \\(\\{\\bar{\\mathbf{I}}^i\\}\\). Since the 2D hypotheses may not be consistent across views, we formulate this as a robust optimization, simultaneously optimizing for the set of views most consistent with \\(\\mathcal{B}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.502, + 0.892, + 0.592 + ], + "angle": 0, + "content": "2D Pose Estimation To distill interaction guidance for the 3D HSI synthesis, we compute a 2D pose representation for the inpainted human in each image \\(\\bar{\\mathbf{I}}^i\\). We use AlphaPose [6], an off-the-shelf pose estimation approach that infers a set of 2D joint positions \\(\\mathbf{J}^i\\) and the corresponding joint confidence scores \\(\\mathbf{c}^i\\) for the human subject in image \\(\\bar{\\mathbf{I}}^i\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.592, + 0.892, + 0.654 + ], + "angle": 0, + "content": "These 2D pose hypotheses \\(\\{\\mathbf{J}^i,\\mathbf{c}^i\\}_{i\\in [1,k]}\\) are then used to steer the interaction synthesis between the 3D human \\(\\mathcal{B}\\) and the scene \\(S\\). We aim to minimize the following objective function \\(\\mathcal{E}_{\\mathrm{total}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.659, + 0.892, + 0.694 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {E} _ {\\text {t o t a l}} = \\lambda_ {\\mathrm {P F}} \\mathcal {E} _ {\\mathrm {P F}} + \\lambda_ {\\mathrm {V S}} \\mathcal {E} _ {\\mathrm {V S}} + \\lambda_ {\\mathrm {B P}} \\mathcal {E} _ {\\mathrm {B P}} + \\tag {1} \\\\ \\lambda_ {\\mathrm {B S}} \\mathcal {E} _ {\\mathrm {B S}} + \\lambda_ {\\mathrm {S C}} \\mathcal {E} _ {\\mathrm {S C}} + \\lambda_ {\\mathrm {S P}} \\mathcal {E} _ {\\mathrm {S P}}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.702, + 0.891, + 0.747 + ], + "angle": 0, + "content": "where the \\(\\lambda\\) denote scalar weights for the energy terms: pose fitting \\(\\mathcal{E}_{\\mathrm{PF}}\\), view selection \\(\\mathcal{E}_{\\mathrm{VS}}\\), body pose \\(\\mathcal{E}_{\\mathrm{BP}}\\), body shape \\(\\mathcal{E}_{\\mathrm{BS}}\\), scene penetration \\(\\mathcal{E}_{\\mathrm{SC}}\\), and self-penetration \\(\\mathcal{E}_{\\mathrm{SP}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Robust View-consistent 3D Pose Fitting. Our primary energy term \\(\\mathcal{E}_{\\mathrm{PF}}\\) minimizes the discrepancy between the projections of the 3D pose of \\(\\mathcal{B}\\) and the inpainted poses from multiple views. However, 2D pose hypotheses can often be inconsistent across different views, due to the stochastic nature of diffusion models, leading to conflicting optimization signals for \\(\\mathcal{B}\\). To address this, we employ a robust optimization strategy that optimizes additionally for view selection weights in \\(\\mathcal{E}_{\\mathrm{PF}}\\) that promote consistency between the 3D pose and the most consistent 2D pose hypotheses." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20468" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.198 + ], + "angle": 0, + "content": "We thus introduce a new set of estimizable variables \\(\\mathbf{w} = \\{w^{i}|w^{i}\\in [0,1]\\}_{i\\in [1,k]}\\), representing view consistency scores, and apply a robust kernel \\(\\rho\\) to the per-view joint fitting constraints. The weights \\(\\mathbf{w}\\) allow the solver to adaptively focus on views with consistent inpainted poses and downweight the inconsistent ones. Our pose fitting energy \\(\\mathcal{E}_{\\mathrm{PF}}\\) with view consistency is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.209, + 0.47, + 0.248 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} _ {\\mathrm {P F}} = \\frac {\\sum_ {i} w ^ {i} \\sum_ {j} \\mathbf {c} _ {j} ^ {i} \\rho (\\Pi (\\hat {\\mathbf {J}}) _ {j} ^ {i} - \\mathbf {J} _ {j} ^ {i})}{\\sum_ {i} w ^ {i}}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.259, + 0.47, + 0.336 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{J}}\\) denotes the 3D joint positions of the SMPL-X model [29] differentiably computed from the pose and shape parameters \\((\\mathbf{R},\\mathbf{t},\\Theta ,\\Phi)\\). The function \\(\\varPi(\\cdot)_j^i\\) represents projection of the \\(j\\)-th joint in the \\(i\\)-th camera view, and \\(\\rho\\) is the robust Geman-McClure error function [8]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.34, + 0.47, + 0.37 + ], + "angle": 0, + "content": "Regularizations. Below, we describe the regularization terms adopted in our optimization." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.374, + 0.471, + 0.403 + ], + "angle": 0, + "content": "1) We encourage the per-view weights \\(\\mathbf{w}\\) to focus on at least \\(\\tau\\) views:" + }, + { + "type": "equation", + "bbox": [ + 0.179, + 0.414, + 0.47, + 0.445 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} _ {\\mathrm {V S}} = \\max (\\tau - \\sum_ {i} w ^ {i}, 0). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.459, + 0.47, + 0.49 + ], + "angle": 0, + "content": "2) To ensure a natural 3D pose, we impose the following energy term on the body pose parameter:" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.501, + 0.47, + 0.52 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} _ {\\mathrm {B P}} = \\left\\| \\boldsymbol {\\Theta} \\right\\| ^ {2} + \\mathcal {E} _ {\\mathrm {J A}} (\\hat {\\boldsymbol {\\Theta}}). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.531, + 0.47, + 0.607 + ], + "angle": 0, + "content": "The first term is the VPoser body prior [29] regularizing the latent pose \\(\\Theta\\), and the second term \\(\\mathcal{E}_{\\mathrm{JA}}\\) is a simple angle prior on the body joint rotations \\(\\hat{\\Theta}\\), decoded from the latent pose \\(\\Theta\\) by VPoser, to penalize extreme bending. The formulation of \\(\\mathcal{E}_{\\mathrm{JA}}\\) is provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.61, + 0.469, + 0.641 + ], + "angle": 0, + "content": "3) We regularize the shape parameter \\(\\Phi\\) to obtain a plausible body shape via" + }, + { + "type": "equation", + "bbox": [ + 0.227, + 0.652, + 0.47, + 0.671 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} _ {\\mathrm {B S}} = \\left\\| \\boldsymbol {\\Phi} \\right\\| ^ {2}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.682, + 0.47, + 0.712 + ], + "angle": 0, + "content": "which measures the Mahalanobis distance between \\(\\Phi\\) and the body shape distribution used in SMPL-X [29]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.715, + 0.47, + 0.761 + ], + "angle": 0, + "content": "4) To ensure physical contact but also avoid penetration between the 3D human \\(\\mathcal{B}\\) and the scene \\(S\\), we formulate the spatial constraints as:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.773, + 0.47, + 0.825 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} _ {\\mathrm {S C}} = \\left\\{ \\begin{array}{l l} \\min _ {\\mathbf {v} \\in \\mathbf {V}} \\Psi (\\mathbf {v}), & \\Psi (\\mathbf {v}) > 0 \\forall \\mathbf {v} \\in \\mathbf {V} \\\\ \\sum_ {\\mathbf {v} \\in \\mathbf {V}} | \\min (\\Psi (\\mathbf {v}), 0) |, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.826, + 0.47, + 0.901 + ], + "angle": 0, + "content": "where \\(\\varPsi\\) is a pre-computed signed distance field (SDF) [10] of the scene \\(S\\). When \\(\\varPsi(\\mathbf{v})\\) has a negative sign, it indicates that the body vertex \\(\\mathbf{v}\\) is located inside the nearest scene object (i.e., penetration). Conversely, a positive sign means that \\(\\mathbf{v}\\) is positioned on the outside." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.894, + 0.168 + ], + "angle": 0, + "content": "5) Finally, to resolve penetration within the human body \\(\\mathcal{B}\\) itself, we include a self-penetration energy \\(\\mathcal{E}_{\\mathrm{SP}}\\) based on detecting colliding body triangles using Bounding Volume Hierarchies (BVH) [40]. We refer the reader to [2, 10, 41] for more formulation details on \\(\\mathcal{E}_{\\mathrm{SP}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.182, + 0.695, + 0.196 + ], + "angle": 0, + "content": "3.4. Iterative Refinement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.206, + 0.894, + 0.403 + ], + "angle": 0, + "content": "A posed 3D human \\(\\mathcal{B}\\) interacting with the 3D scene \\(S\\) is generated after applying the VLM-based 2D inpainting (Sec. 3.2) and the robust 3D lifting (Sec. 3.3). To improve the synthesis and consistency of the interaction results, we employ an iterative refinement scheme over the aforementioned inpainting and 3D lifting. During refinement, we render the silhouette of the posed 3D human \\(\\mathcal{B}\\) in each camera view, and use it as a more precise and consistent mask \\(\\mathbf{M}\\) in the latent diffusion inpainting \\(\\Omega\\) [34], replacing dynamic masking. By doing so, the consistency among the 2D pose hypotheses inpainted into the multi-view scene images gradually improves, thus leading to an improved 3D HSI synthesis outcome." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.417, + 0.719, + 0.433 + ], + "angle": 0, + "content": "3.5. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.442, + 0.895, + 0.67 + ], + "angle": 0, + "content": "We implemented GenZI with PyTorch [28]. For VLM-based 2D inpainting, we use \\( k = 16 \\) cameras for multiview rendering. We use 50 denoising steps in Stable Diffusion Inpainting with a state-of-the-art diffusion sampler [25]. For dynamic masking during inpainting, we aggregate cross-attention maps with a resolution of \\( 16 \\times 16 \\). Input text prompts \\( \\Gamma \\) (e.g., \"sitting on a bench\") are all appended with fixed prefix \"a woman\" and fixed suffix \"wearing a white shirt and blue pants, full body\" to better constrain generation. For robust 3D lifting, we set \\( \\tau = 3 \\), and optimize the energy \\( \\mathcal{E}_{\\mathrm{total}} \\) using gradient descent [24] for 1.6K steps, which takes \\( \\sim 3 \\) minutes on an NVIDIA A100 GPU. In the iterative refinement, we dilate the rendered human silhouette with a kernel size of \\( 11 \\times 11 \\) for mask generation, and the refinement is performed once." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.688, + 0.634, + 0.705 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.715, + 0.894, + 0.79 + ], + "angle": 0, + "content": "We demonstrate the effectiveness and generality of our approach GenZI on a diverse collection of 3D scene models from Sketchfab.com. We conduct both quantitative and qualitative evaluations to compare GenZI with alternative baselines approaches [10, 51] to our new task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Dataset. In our Sketchfab dataset, we gathered 8 large-scale 3D scenes encompassing a variety of indoor and outdoor environments with diverse geometric structures, including a realistic Venice city, a gym, and a cartoon-style food truck. We collected 4-5 text prompts per scene describing human interactions with the scene for specified approximate point locations, resulting in 38 actions for evaluation." + }, + { + "type": "footer", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20469" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "Baselines. To the best of our knowledge, there are no baselines that estimate 3D human-scene interactions based on natural language text input in a zero-shot fashion. We thus perform comparisons with related baseline approaches:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.153, + 0.471, + 0.365 + ], + "angle": 0, + "content": "- COINS [51] is a state-of-the-art approach estimating 3D humans in indoor 3D scans with a fixed vocabulary of actions and objects, given object segmentations. It takes as input (action, object) pairs for semantic control, and requires full supervision in its CVAE training, using captured 3D interaction data with both instance segmentations of 3D objects and action labelling. Due to being trained on a small, closed set of indoor interactions, we adapt COINS to a subset of the most similar Sketchfab actions by manually segmenting corresponding 3D objects from the scenes similar to its indoor training data and mapping the text prompts to its (action, object) input. Note that our approach GenZI does not require any 3D scene segmentations." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.365, + 0.47, + 0.47 + ], + "angle": 0, + "content": "- Hassan et al. [10] perform 3D human estimation from a single RGB image. To adapt this to Sketchfab, we reuse the multi-view scene images \\(\\{\\bar{\\mathbf{T}}^i\\}_{i\\in [1,k]}\\) inpainted with humans from our dynamic masking scheme (Sec. 3.2), where the view with the best image-text cosine similarity (i.e., CLIP score [33]) is used as the input for 3D human estimation under the known virtual camera parameters." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.47, + 0.469, + 0.531 + ], + "angle": 0, + "content": "- Ours-Single View: Finally, we consider a baseline leveraging 3D body estimation from our method, but limited to only one inpainted view. The same best view image, as described above, is also as input for this baseline." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.153, + 0.471, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.536, + 0.47, + 0.581 + ], + "angle": 0, + "content": "Evaluation Metrics. To measure 3D HSI quality, we conduct perceptual studies and compute metrics including semantic consistency, diversity, and physical plausibility." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.582, + 0.469, + 0.734 + ], + "angle": 0, + "content": "We first carry out two perceptual studies to assess the realism and semantic accuracy of the synthesized interactions. The first is a binary-choice study, where interaction samples generated by two different methods based on the same text prompt are shown, and the participants are asked to choose the sample that is more realistic and better matched the text. The second study is a unary test, where for each interaction sample, the participants are asked to rate the realism and consistency between the shown sample and text prompt from 1 (strongly disagree) to 5 (strongly agree)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.734, + 0.469, + 0.809 + ], + "angle": 0, + "content": "To evaluate the semantic consistency between a synthesized 3D interaction and the input text prompt, we calculate the CLIP score [33], where the 3D interaction is re-rendered into \\( k \\) view images, and the image-text cosine similarities from CLIP ViT-B/32 are averaged across all views." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Additionally, we include quantitative metrics from existing works to evaluate diversity and physical plausibility. We observe that these metrics often do not reflect perceptual quality, as generated bodies can be diverse but have interpenetrations without demonstrating any physical or semantic coherence. Nevertheless, to measure the synthesis" + }, + { + "type": "image_caption", + "bbox": [ + 0.609, + 0.092, + 0.783, + 0.107 + ], + "angle": 0, + "content": "(a) Binary Perceptual Study" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.108, + 0.871, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.524, + 0.214, + 0.861, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.346, + 0.892, + 0.402 + ], + "angle": 0, + "content": "Figure 4. User study of 3D human-scene interaction synthesis on the Sketchfab dataset, where participants show a strong preference for the generations by our approach, in comparison with all baselines, COINS [51], Hassan et al. [10], and Ours-Single View." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.422, + 0.892, + 0.634 + ], + "angle": 0, + "content": "diversity, we follow [50, 51] and cluster the SMPL-X parameters of the generated humans into 20 clusters with K-means and compute the entropy of the cluster ID histogram of all the samples. We also calculate the cluster size as the mean distance to cluster centers. For the physical plausibility, we evaluate the collision and contact between body meshes and scene meshes, following [50, 51]. The noncollision score is computed as the ratio between the number of body mesh vertices with positive SDF values (Sec. 3.3) and the number of all body mesh vertices. The contact score is defined as the ratio between the number of body meshes with scene contact and the number of all generated body meshes. A body mesh is in contact with the scene if any body mesh vertex has a non-positive SDF value." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.646, + 0.836, + 0.663 + ], + "angle": 0, + "content": "4.1. Comparison to Alternative Approaches" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.673, + 0.892, + 0.839 + ], + "angle": 0, + "content": "Quantitative Evaluation. Fig. 4 shows the results of the perceptual studies collected from 30 participants across a binary study and a unary study. In the binary study, we observe that participants overwhelmingly favor the generations by our GenZI compared to all baselines – more than \\(87\\%\\) of the time. In the unary test, the average realism rating for our interaction generations is 3.6, the highest compared to the baselines, which are all below 2.0. These perceptual results strongly indicate that our GenZI can synthesize realistic 3D humans interacting with various 3D scenes without requiring any captured 3D interaction data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Tab. 1 presents the quantitative evaluation of semantic consistency, diversity, and physical plausibility on the full Sketchfab dataset. Our approach achieves the best semantic consistency score, echoing the strong user preference for" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20470" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.145, + 0.092, + 0.24, + 0.107 + ], + "angle": 0, + "content": "Text Prompt" + }, + { + "type": "image_caption", + "bbox": [ + 0.109, + 0.11, + 0.27, + 0.144 + ], + "angle": 0, + "content": "\"sitting and playing piano, hands on piano keys, feet flat on the ground\"" + }, + { + "type": "image", + "bbox": [ + 0.141, + 0.145, + 0.237, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.215, + 0.258, + 0.24 + ], + "angle": 0, + "content": "\"climbing a ladder, hands gripping side rails\"" + }, + { + "type": "image", + "bbox": [ + 0.14, + 0.24, + 0.24, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.106, + 0.321, + 0.273, + 0.344 + ], + "angle": 0, + "content": "\"sitting and paddling a wooden boat, spreading arms\"" + }, + { + "type": "image", + "bbox": [ + 0.142, + 0.345, + 0.239, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.132, + 0.43, + 0.248, + 0.443 + ], + "angle": 0, + "content": "\"walking on a bridge\"" + }, + { + "type": "image", + "bbox": [ + 0.15, + 0.444, + 0.233, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.126, + 0.534, + 0.255, + 0.557 + ], + "angle": 0, + "content": "\"picking up a trash bag, standing, bending over\"" + }, + { + "type": "image", + "bbox": [ + 0.14, + 0.559, + 0.24, + 0.616 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.641, + 0.253, + 0.653 + ], + "angle": 0, + "content": "\"walking out of a door\"" + }, + { + "type": "image", + "bbox": [ + 0.14, + 0.654, + 0.24, + 0.73 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.122, + 0.746, + 0.257, + 0.769 + ], + "angle": 0, + "content": "\"riding a recumbent bike, sitting\"" + }, + { + "type": "image", + "bbox": [ + 0.141, + 0.784, + 0.24, + 0.825 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.321, + 0.092, + 0.371, + 0.104 + ], + "angle": 0, + "content": "COINS" + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.105, + 0.413, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.446, + 0.092, + 0.543, + 0.104 + ], + "angle": 0, + "content": "Hassan et al." + }, + { + "type": "image", + "bbox": [ + 0.426, + 0.105, + 0.559, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.607, + 0.092, + 0.67, + 0.104 + ], + "angle": 0, + "content": "Ours-SV" + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.105, + 0.705, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.766, + 0.092, + 0.804, + 0.104 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.105, + 0.852, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.426, + 0.209, + 0.559, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.209, + 0.705, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.209, + 0.852, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.426, + 0.313, + 0.559, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.313, + 0.705, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.313, + 0.852, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.419, + 0.413, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.426, + 0.419, + 0.559, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.42, + 0.705, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.42, + 0.852, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.426, + 0.523, + 0.559, + 0.625 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.524, + 0.705, + 0.625 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.524, + 0.852, + 0.626 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.426, + 0.628, + 0.559, + 0.73 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.628, + 0.705, + 0.73 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.628, + 0.852, + 0.73 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.733, + 0.413, + 0.836 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.426, + 0.733, + 0.559, + 0.836 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.733, + 0.705, + 0.836 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.733, + 0.852, + 0.836 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.841, + 0.892, + 0.896 + ], + "angle": 0, + "content": "Figure 5. Qualitative results on the Sketchfab dataset. Our GenZI synthesizes more realistic 3D human-scene interactions and generalizes better across diverse scene types, compared to the baselines COINS [51], Hassan et al. [10], and Ours-Single View. For COINS, we show the used (action, object) labels from its closed set of indoor interactions; its closed setting can lead to degraded results from out-of-distribution object classes (e.g., curved bridge deck as the floor, chair at a different height or shape than those in the training set)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20471" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.079, + 0.089, + 0.468, + 0.14 + ], + "angle": 0, + "content": "
MethodSemanticsDiversityPhysical Plausibility
CLIP ↑Entropy ↑Cluster Size ↑Non-collision ↑Contact ↑
Hassan et al. [10]0.25982.70141.19070.88240.9669
Ours-SV0.26132.64521.58130.97651.0000
Ours0.27102.73041.05000.97670.9868
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.144, + 0.47, + 0.214 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons on the Sketchfab dataset. Our approach achieves the best semantic consistency, diversity entropy, and non-collision scores, with the contact score on par. Note that single view methods Hassan et al. [10] and Ours-SV tend to produce increased diversity at the cost of semantic plausibility." + }, + { + "type": "table", + "bbox": [ + 0.134, + 0.228, + 0.406, + 0.255 + ], + "angle": 0, + "content": "
w/o DMw/o VCw/o IROurs
CLIP ↑0.26390.26940.26640.2710
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.259, + 0.47, + 0.316 + ], + "angle": 0, + "content": "Table 2. Ablation study on the Sketchfab dataset. The semantic consistency of 3D interaction generations degrades without dynamic masking (DM), view consistency (VC), or iterative refinement (IR), compared to our full approach." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.33, + 0.468, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.418, + 0.468, + 0.501 + ], + "angle": 0, + "content": "Figure 6. Visualization of our method ablations on Sketchfab dataset for the input text: \"sitting on a bar stool\". Without dynamic masking (DM) or view consistency (VC), the person floats above the middle stool. Without iterative refinement (IR), the person penetrates the stool. Our full approach results in a more realistic synthesis." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.515, + 0.468, + 0.651 + ], + "angle": 0, + "content": "GenZI in the perceptual studies. We note that since both Hassan et al. [10] and Ours-Single View operate on single inpainted view samples, results can be very diverse but lack 3D plausibility. This suggests that the CLIP score is a more reliable metric for assessing the HSI synthesis quality, compared to the diversity and physical plausibility metrics, which reflect less about the generation realism. Nevertheless, our approach has the best diversity entropy and noncollision scores, with the contact score on par." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.655, + 0.468, + 0.837 + ], + "angle": 0, + "content": "Qualitative Evaluation. We show qualitative comparisons in Fig. 5. COINS is severely limited by its training on the closed set of indoor interactions, and thus fails to generalize to outdoor scenes and unseen objects (e.g., no curved floors exist during training, and limited sets of heights and shapes of chairs). As Hassan et al. [10] and Ours-Single View operate from single views, they both suffer from insufficient pose constraints from other views for plausible interaction generation. In contrast, our approach demonstrates high flexibility and generality to a diverse set of 3D indoor and outdoor scenes by leveraging large VLMs to imagine multiview interaction hypotheses and then robust 3D lifting." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.847, + 0.241, + 0.862 + ], + "angle": 0, + "content": "4.2. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.468, + 0.902 + ], + "angle": 0, + "content": "We conduct ablation studies on the Sketchfab dataset to validate the effectiveness of our proposed dynamic masking" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.891, + 0.137 + ], + "angle": 0, + "content": "scheme (Sec. 3.2), robust 3D lifting with view consistency (Sec. 3.3), and iterative refinement (Sec. 3.4). Results are presented in Fig. 6 and Tab. 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.14, + 0.892, + 0.292 + ], + "angle": 0, + "content": "Dynamic Masking (DM). We replace our dynamic masking, used during latent diffusion inpainting, with random masking. We sample a random mask around the image center covering at least \\(30\\%\\) of the image area, and use it as a fixed mask input to the latent diffusion model \\(\\Omega\\). We observe that random masking results in noticeably worse quality of interaction synthesis (Tab. 2), and incoherence with the scene (Fig. 6, floating on the stool). This indicates that our dynamic masking is effective in incorporating sufficient scene context for human inpainting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.295, + 0.892, + 0.401 + ], + "angle": 0, + "content": "View Consistency (VC). We evaluate the role of view consistency in robust 3D lifting by fixing the estimizable scores to \\(\\mathbf{w} = \\mathbf{1}\\). Using all inpainted views leads to averaged, less expressive 3D human poses (Tab. 2, Fig. 6) due to potential inconsistent 2D pose hypotheses across views. By allowing the solver to adaptively focus on views with consistent inpaintings, our approach generates more realistic 3D HSIs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.404, + 0.892, + 0.465 + ], + "angle": 0, + "content": "Iterative Refinement (IR). Finally, we show the effectiveness of iterative refinement by applying our VLM-based 2D inpainting and robust 3D lifting only once. Tab. 2 and Fig. 6 show that iterative refinement improves synthesis quality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.468, + 0.892, + 0.574 + ], + "angle": 0, + "content": "Limitations. Our approach is limited by the inpainting capability of latent diffusion models to imagine possible 2D human-scene compositions, and the diffusion models are also known to be slow at inference time due to their iterative nature [14]. Nevertheless, we believe that our approach can directly benefit from the rapid advancement of VLMs for improved HSI synthesis." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.586, + 0.619, + 0.601 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.61, + 0.892, + 0.821 + ], + "angle": 0, + "content": "We have presented the first approach to synthesize general 3D human-scene interactions guided by text inputs. Key to our approach is effective distillation of knowledge from large vision-language models, enabling generating 3D humans in scenes without requiring any 3D interaction data for training. We leverage these powerful vision-language models to generate hypotheses for inpainted 2D human-scene interactions. We then formulate a robust optimization to lift the hypotheses to 3D in a view-consistent fashion by simultaneously optimizing for the most informative 2D hypotheses. Our approach is flexible and can be applied to general scene settings for a variety of actions. We believe this opens up new opportunities for 3D understanding without requiring expensive capture of 3D/4D data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. This project is funded by the ERC Starting Grant SpatialSem (101076253), and the German Research Foundation (DFG) Grant \"Learning How to Interact with Scenes through Part-Based Understanding\", and supported in part by a Google research gift." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20472" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.156 + ], + "angle": 0, + "content": "[1] Stable Diffusion v2. https://huggingface.co/stabilityai/stable-diffusion-2-inpainting.Accessed:2023-09-27.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.468, + 0.2 + ], + "angle": 0, + "content": "[2] Luca Ballan, Aparna Taneja, Jürgen Gall, Luc Van Gool, and Marc Pollefeys. Motion capture of hands in action using discriminative salient points. In ECCV, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.203, + 0.468, + 0.258 + ], + "angle": 0, + "content": "[3] Zhongang Cai, Wanqi Yin, Ailing Zeng, Chen Wei, Qingping Sun, Yanjun Wang, Hui En Pang, Haiyi Mei, Mingyuan Zhang, Lei Zhang, et al. SMPLer-X: Scaling up expressive human pose and shape estimation. arXiv, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.261, + 0.468, + 0.315 + ], + "angle": 0, + "content": "[4] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-Excite: Attention-based semantic guidance for text-to-image diffusion models. ACM TOG, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.318, + 0.468, + 0.372 + ], + "angle": 0, + "content": "[5] Vasileios Choutas, Georgios Pavlakos, Timo Bolkart, Dimitrios Tzionas, and Michael J Black. Monocular expressive body regression through body-driven attention. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.376, + 0.468, + 0.431 + ], + "angle": 0, + "content": "[6] Hao-Shu Fang, Jiefeng Li, Hongyang Tang, Chao Xu, Haoyi Zhu, Yuliang Xiu, Yong-Lu Li, and Cewu Lu. Alpha-Pose: Whole-body regional multi-person pose estimation and tracking in real-time. IEEE TPAMI, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.434, + 0.468, + 0.46 + ], + "angle": 0, + "content": "[7] Dariu M Gavrila. The visual analysis of human movement: A survey. CVIU, 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.464, + 0.468, + 0.49 + ], + "angle": 0, + "content": "[8] Stuart Geman and Donald E. McClure. Statistical methods for tomographic image reconstruction. 1987. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.494, + 0.468, + 0.521 + ], + "angle": 0, + "content": "[9] James J Gibson. The ecological approach to the visual perception of pictures. *Leonardo*, 11(3):227-235, 1978. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.524, + 0.468, + 0.577 + ], + "angle": 0, + "content": "[10] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J. Black. Resolving 3D human pose ambiguities with 3D scene constraints. In ICCV, 2019. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.468, + 0.622 + ], + "angle": 0, + "content": "[11] Mohamed Hassan, Partha Ghosh, Joachim Tesch, Dimitrios Tzionas, and Michael J. Black. Populating 3D scenes by learning human-scene interaction. In CVPR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.625, + 0.468, + 0.665 + ], + "angle": 0, + "content": "[12] Mohamed Hassan, Yunrong Guo, Tingwu Wang, Michael Black, Sanja Fidler, and Xue Bin Peng. Synthesizing physical character-scene interactions. In SIGGRAPH, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.669, + 0.468, + 0.709 + ], + "angle": 0, + "content": "[13] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-prompt image editing with cross-attention control. In ICLR, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.713, + 0.468, + 0.739 + ], + "angle": 0, + "content": "[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.743, + 0.468, + 0.783 + ], + "angle": 0, + "content": "[15] Lukas Hollein, Ang Cao, Andrew Owens, Justin Johnson, and Matthias Nießner. Text2Room: Extracting textured 3d meshes from 2d text-to-image models. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.786, + 0.468, + 0.84 + ], + "angle": 0, + "content": "[16] Siyuan Huang, Zan Wang, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, and Song-Chun Zhu. Diffusion-based generation, optimization, and planning in 3D scenes. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.844, + 0.434, + 0.857 + ], + "angle": 0, + "content": "[17] Kurt Koffka. Principles of gestalt psychology. 1935. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[18] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In ICCV, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[19] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. BLIP: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.151, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[20] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.209, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[21] Xueting Li, Sifei Liu, Kihwan Kim, Xiaolong Wang, Ming-Hsuan Yang, and Jan Kautz. Putting humans in a scene: Learning affordance in 3D indoor environments. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.266, + 0.892, + 0.307 + ], + "angle": 0, + "content": "[22] Jing Lin, Ailing Zeng, Haoqian Wang, Lei Zhang, and Yu Li. One-stage 3d whole-body mesh recovery with component aware transformer. In CVPR, pages 21159-21168, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.31, + 0.892, + 0.351 + ], + "angle": 0, + "content": "[23] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. SMPL: A skinned multiperson linear model. ACM TOG, 34(6):1-16, 2015. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.354, + 0.892, + 0.381 + ], + "angle": 0, + "content": "[24] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.384, + 0.892, + 0.425 + ], + "angle": 0, + "content": "[25] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. DPM-Solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.428, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[26] Thomas B Moeslund, Adrian Hilton, and Volker Krüger. A survey of advances in vision-based human motion capture and analysis. CVIU, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.472, + 0.892, + 0.526 + ], + "angle": 0, + "content": "[27] Aron Monszpart, Paul Guerrero, Duygu Ceylan, Ersin Yumer, and Niloy J Mitra. iMapper: interaction-guided scene mapping from monocular videos. ACM Transactions On Graphics (TOG), 38(4):1-15, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.529, + 0.892, + 0.612 + ], + "angle": 0, + "content": "[28] A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. Kopf, E. Yang, Z. DeVito, M. Raison, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, and S. Chintala. PyTorch: An imperative style, high-performance deep learning library. NeurIPS, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.615, + 0.892, + 0.669 + ], + "angle": 0, + "content": "[29] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3D hands, face, and body from a single image. In CVPR, 2019. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.673, + 0.892, + 0.755 + ], + "angle": 0, + "content": "[30] R. Po, W. Yifan, V. Golyanik, K. Aberman, J. T. Barron, A. H. Bermano, E. R. Chan, T. Dekel, A. Holynski, A. Kanazawa, C. K. Liu, L. Liu, B. Mildenhall, M. Nießner, B. Ommer, C. Theobalt, P. Wonka, and G. Wetzstein. State of the art on diffusion models for visual computing. arXiv, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.758, + 0.892, + 0.784 + ], + "angle": 0, + "content": "[31] Ronald Poppe. Vision-based human motion analysis: An overview. CVIU, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.788, + 0.892, + 0.828 + ], + "angle": 0, + "content": "[32] Sergey Prokudin, Christoph Lassner, and Javier Romero. Efficient learning on point clouds with basis point sets. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20473" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.093, + 0.469, + 0.146 + ], + "angle": 0, + "content": "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.151, + 0.468, + 0.192 + ], + "angle": 0, + "content": "[35] Yu Rong, Takaaki Shiratori, and Hanbyul Joo. FrankMocap: A monocular 3d whole-body pose estimation system via regression and integration. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.195, + 0.469, + 0.234 + ], + "angle": 0, + "content": "[36] David Rozenberszki, Or Litany, and Angela Dai. Language-grounded indoor 3d semantic segmentation in the wild. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.239, + 0.469, + 0.307 + ], + "angle": 0, + "content": "[37] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.31, + 0.468, + 0.351 + ], + "angle": 0, + "content": "[38] Nikolaos Sarafianos, Bogdan Boteanu, Bogdan Ionescu, and Ioannis A Kakadiaris. 3D human pose estimation: A review of the literature and analysis of covariates. CVIU, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.354, + 0.468, + 0.407 + ], + "angle": 0, + "content": "[39] Manolis Savva, Angel X Chang, Pat Hanrahan, Matthew Fisher, and Matthias Nießner. PiGraphs: learning interaction snapshots from observations. ACM TOG, 35(4):1-12, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.412, + 0.468, + 0.48 + ], + "angle": 0, + "content": "[40] Matthias Teschner, Stefan Kummerle, Bruno Heidelberger, Gabriel Zachmann, Laks Raghupathi, Arnulph Fuhrmann, M-P Cani, François Faure, Nadia Magnenat-Thalmann, Wolfgang Strasser, et al. Collision detection for deformable objects. In CGF, 2005. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.483, + 0.468, + 0.537 + ], + "angle": 0, + "content": "[41] Dimitrios Tzionas, Luca Ballan, Abhilash Srikantha, Pablo Aponte, Marc Pollefeys, and Juergen Gall. Capturing hands in action using discriminative salient points and physics simulation. IJCV, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.541, + 0.468, + 0.581 + ], + "angle": 0, + "content": "[42] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.585, + 0.468, + 0.625 + ], + "angle": 0, + "content": "[43] Jingbo Wang, Yu Rong, Jingyuan Liu, Sijie Yan, Dahua Lin, and Bo Dai. Towards diverse and natural scene-aware 3D human motion synthesis. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.629, + 0.468, + 0.669 + ], + "angle": 0, + "content": "[44] Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, Wei Liang, and Siyuan Huang. HUMANISE: Language-conditioned human motion generation in 3D scenes. NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.673, + 0.468, + 0.726 + ], + "angle": 0, + "content": "[45] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.731, + 0.468, + 0.785 + ], + "angle": 0, + "content": "[46] Haibiao Xuan, Xiongzheng Li, Jinsong Zhang, Hongwen Zhang, Yebin Liu, and Kun Li. Narrator: Towards natural control of human-scene interaction generation via relationship reasoning. arXiv, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.788, + 0.468, + 0.842 + ], + "angle": 0, + "content": "[47] Yufei Ye, Xueting Li, Abhinav Gupta, Shalini De Mello, Stan Birchfield, Jiaming Song, Shubham Tulsiani, and Sifei Liu. Affordance diffusion: Synthesizing hand-object interactions. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.846, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[48] Hongwen Zhang, Yating Tian, Yuxiang Zhang, Mengcheng Li, Liang An, Zhenan Sun, and Yebin Liu. PyMAF-X: Towards well-aligned full-body model regression from monocular images. IEEE TPAMI, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[49] Siwei Zhang, Yan Zhang, Qianli Ma, Michael J Black, and Siyu Tang. Place: Proximity learning of articulation and contact in 3D environments. In 3DV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[50] Yan Zhang, Mohamed Hassan, Heiko Neumann, Michael J. Black, and Siyu Tang. Generating 3D people in scenes without people. In CVPR, 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[51] Kaifeng Zhao, Shaofei Wang, Yan Zhang, Thabo Beeler, and Siyu Tang. Compositional human-scene interaction synthesis with semantic control. In ECCV, 2022. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.222, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[52] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.264, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[53] Yuxiao Zhou, Marc Habermann, Ikhsanul Habibie, Ayush Tewari, Christian Theobalt, and Feng Xu. Monocular real-time full body capture with inter-part correlations. In CVPR, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20474" + } + ] +] \ No newline at end of file diff --git a/2024/GenZI_ Zero-Shot 3D Human-Scene Interaction Generation/3823a91d-f3af-49d1-a8df-4cb9a1c22737_origin.pdf b/2024/GenZI_ Zero-Shot 3D Human-Scene Interaction Generation/3823a91d-f3af-49d1-a8df-4cb9a1c22737_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..738a4cb1c65f25a8423864cc1aa67596a6e8485d --- /dev/null +++ b/2024/GenZI_ Zero-Shot 3D Human-Scene Interaction Generation/3823a91d-f3af-49d1-a8df-4cb9a1c22737_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8f740b0673e7010678695fcdaea529b63d06795ba0dbbc640f73cc7d1357e4f9 +size 7875741 diff --git a/2024/GenZI_ Zero-Shot 3D Human-Scene Interaction Generation/full.md b/2024/GenZI_ Zero-Shot 3D Human-Scene Interaction Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..85babdba9f090d6a167b630d82b7e40f8ec84075 --- /dev/null +++ b/2024/GenZI_ Zero-Shot 3D Human-Scene Interaction Generation/full.md @@ -0,0 +1,361 @@ +# GenZI: Zero-Shot 3D Human-Scene Interaction Generation + +Lei Li Angela Dai Technical University of Munich + +"riding a motorcycle, sitting" + +"hands knocking on the closed door, facing the door, standing" + +"stretching legs and sitting on the saddle on a standing cow" + +"picking up dumbbells on a shelf, standing, bending over" + +![](images/7fb1eb99e239083e4c92aa15a562c585e889e82f9690fca9541c91db3f6ce443.jpg) +Figure 1. Given an arbitrary 3D scene, GenZI can synthesize virtual humans interacting with the 3D environment at specified locations from a brief text description. Our approach does not require any 3D human-scene interaction training data or 3D learning. By distilling interaction priors from powerful 2D vision-language models, we optimize for 3D human-scene interaction synthesis in a flexible fashion, with simple language-based control and high generality to various types of scene environments. + +![](images/cd67976967625ecb4d433022798ca86154c59cf857218566c4b55c3e95759441.jpg) + +![](images/19bdeca1d364be38016848ec3d9a7e77e40a1a765103878f229318030c6476fe.jpg) + +![](images/e649850e33763e06dc50ad25785d560cd3fa23e67c0c638c07bba83aa33b4df6.jpg) + +# Abstract + +Can we synthesize 3D humans interacting with scenes without learning from any 3D human-scene interaction data? We propose GenZI1, the first zero-shot approach to generating 3D human-scene interactions. Key to GenZI is our distillation of interaction priors from large vision-language models (VLMs), which have learned a rich semantic space of 2D human-scene compositions. Given a natural language description and a coarse point location of the desired interaction in a 3D scene, we first leverage VLMs to imagine plausible 2D human interactions inpainted into multiple rendered views of the scene. We then formulate a robust iterative optimization to synthesize the pose and shape of a 3D human model in the scene, guided by consistency with the 2D interaction hypotheses. In contrast to existing learning-based approaches, GenZI circumvents the conventional need for captured 3D interaction data, and allows for flexible control of the 3D interaction synthesis with easy-to-use text prompts. Extensive experiments show that our zero-shot approach has high flexibility and generality, making it applicable to diverse scene types, including both indoor and outdoor environments. + +# 1. Introduction + +Scenes are typically constructed to enable human interactions with the environment, like sitting on a couch, playing a piano, or opening a mailbox. Understanding these interactions between humans and scenes, also known as affordances [9, 17], has gained increasing attention in the fields of computer vision and graphics. In particular, achieving controllable and generalizable synthesis of human-scene interactions (or HSIs) holds immense potential for various applications, such as robotics, architectural design, video games, and virtual reality experiences, among many others. + +Generating realistic humans interacting with a 3D scene is a challenging task, requiring holistic semantic understanding of both the environment and possible human actions therein. Existing approaches to HSI synthesis [11, 46, 50, 51] rely heavily on supervised training using meticulously captured data of real people interacting in 3D environments. Unfortunately, collecting large-scale datasets of 3D scenes and human interactions is exorbitantly difficult. It not only demands accurate tracking and reconstruction of both people and their environments, but also needs to ensure sufficient diversity in subjects and scenes. Existing HSI datasets currently contain very limited quantities of scenes and actions, for example, PROX [10, 51] only consisting + +of 12 indoor scenes and humans interacting with 11 object categories. The scarcity of ground truth data for supervision has thus strongly limited the applicability and generalization of the learning-based approaches for synthesizing diverse sets of actions in arbitrary 3D scenes. + +We thus consider an alternative perspective to HSI synthesis and pose the following question: Can we achieve plausible HSI synthesis without using any captured 3D interaction data? To this end, we present a novel zero-shot approach to 3D HSI generation. We propose the first method to leverage the powerful capabilities exhibited by recent vision-language models (VLMs) [19, 20, 30, 33, 34, 37] to synthesize plausible 2D images of human interactions, and introduce a robust optimization to distill inferred 2D pose information into 3D human synthesis in a 3D scene. + +More concretely, given a 3D scene, a text prompt and a coarse point location of the desired interaction, GenZI optimizes for the pose and shape of a 3D human performing the action in the scene, guided by a large VLM [34]. We first leverage the VLM to imagine possible 2D humans by inpainting images from multiple rendered views of the scene. We automate this 2D human insertion process with a dynamic masking scheme that automatically updates proposed masks through the inpainting process, eliminating the need for manual specifications of human inpainting regions. We then lift these 2D interaction hypotheses to 3D and optimize for a parametric 3D human body model [23, 29] that is most consistent with the 2D pose guidance. We further refine the generated 3D human in the scene by iterating through the VLM-based 2D inpainting and robust 3D lifting stages. We demonstrate the flexibility and generality of GenZI in various types of 3D environments (Fig. 1), encompassing both indoor and outdoor scenes. + +In summary, our contributions are as follows: + +- We introduce GenZI, the first zero-shot approach to generating realistic 3D humans interacting with a 3D scene from natural language prompts. GenZI does not require any supervision from 3D interaction data, thus enabling flexible synthesis across diverse scenes and actions. +- We propose a dynamically-masked inpainting scheme that allows for the synthesis of plausible 2D human-scene compositions via VLMs without requiring manually-specified human inpainting masks. +- We develop a robust 3D pose optimization to lift various inferred images of human interactions to a view-consistent, realistic 3D HSI synthesis. + +# 2. Related Work + +3D Human-Scene Interaction Synthesis. Synthesizing humans in scenes is an important, challenging task in computer vision and graphics, as it models complex high-level + +semantic understanding such as affordances and interactions. Existing approaches for human-scene interaction synthesis [11, 12, 16, 21, 27, 39, 43, 44, 50, 51] focus on learning priors from available collected data of people interacting with 3D indoor scenes [10] in a supervised manner. + +The early work PiGraphs [39] introduced a probabilistic graphical model of human-object interactions, generating static human-object placements by pose sampling and model retrieval. Zhang et al. [50] introduced a generative model of human-scene interaction, using a conditional variational autoencoder (CVAE) to model the distribution of 3D human poses conditioned on scene depth and semantics. PLACE [49] explicitly represented human-scene interactions with Basis Point Sets (BPS) encoding [32], training a CVAE to synthesize such representations for generating natural human poses and contact relations within scenes. POSA [11] proposed an ego-centric representation by augmenting the SMPL-X model [29] with contact labels and scene semantics, using a CVAE conditioned on SMPL-X vertex positions to model potential interactions. Recently, COINS [51] presented a method for compositional human-scene interaction synthesis with high-level semantic control, using transformer-based CVAEs conditioned on provided 3D objects and interaction semantics to regress human body poses and contact features. + +However, these works all require supervision from 3D human-scene interactions – such datasets are difficult and expensive to capture and annotate, for instance, the widely used PROX dataset [10, 51] only has 8 reconstructed indoor scenes for training and 4 for testing, capturing human interactions with 11 object categories. As such, these methods are limited to in-domain synthesis, prohibiting applicability to general 3D scene settings with arbitrary objects and arrangements. In this work, we deviate from the learning-based approach and propose to take advantage of established large vision-language models [20, 33, 34, 37] for human-scene interaction synthesis, thereby bypassing the requirement for data capture and 3D learning. + +3D Human Estimation from RGB Images. Over the past decades, significant progress has been made in 3D human estimation from RGB images [7, 26, 31, 38]. Among those, the prominent work SMPLify-X [29] fits the SMPL-X model to 2D joints estimated from a single image through optimization. Hassan et al. [10] build upon SMPLify-X and estimate 3D human poses by incorporating additional physical contact and penetration constraints between human and 3D scene. Learning-based approaches [3, 5, 18, 22, 35, 48, 53] have received much research attention in recent years towards tackling 3D human estimation from monocular images with improved hand and face estimation. Our work leverages 2D pose reasoning from multiple different 2D view hypotheses, and proposes a robust 3D formulation for aggregating the various 2D hypotheses to a consistent + +![](images/486a6430454ccd00503ed6038de0c7e738b69f3716bb2625c649fb57b26d31fe.jpg) +Figure 2. GenZI distills information from vision-language model for 3D human-scene interaction. We first leverage large vision-language models to synthesize possible 2D humans interactions with the 3D scene $S$ by employing latent diffusion inpainting [34] on multiple rendered views of the environment at location $\mathbf{p}$ using our dynamic masking scheme to automatically estimate inpainting masks. We then lift these 2D hypotheses to 3D in a robust optimization for a 3D parametric body model $\mathcal{B}$ (SMPL-X [29]) that is most consistent with detected 2D poses in the inpainted 2D hypotheses. This produces a semantically consistent interaction that respects the scene context, without requiring any 3D human-scene interaction data. + +3D human body interacting with a 3D scene. + +Distilling Prior from Vision-Language Models. Recent advances in powerful vision-language models [20, 33, 34, 37] have also inspired various works aiming to distill information learned through the models to various tasks, including 2D panoptic segmentation [45], 3D semantic segmentation [36], 3D scene generation [15], and synthesis of images with hand-object interactions [47]. Our approach leverages the 2D generative capacity of these models to convey information about possible human-scene interactions, which we then lift to a 3D, consistent interaction. + +# 3. Method + +# 3.1. Overview + +Our objective is to synthesize plausible 3D humans interacting with a 3D scene, guided by input text descriptions, in the absence of 3D interaction data capture for learning. We present GenZI, a novel optimization-based multi-view approach that leverages large VLMs to infer spatial relations of interactions between a human and the scene. Fig. 2 shows an illustration of our approach. + +GenZI takes as input a 3D scene $S$ , a text prompt $\Gamma$ describing the desired interaction, and an approximate point location $\mathbf{p} \in \mathbb{R}^3$ in the scene around which the interaction should occur. Our approach generates a posed 3D human $\mathcal{B}$ as output, performing the specified action in the scene. We + +adopt SMPL-X [23, 29] to parameterize the 3D human $\mathcal{B}$ as it provides a fully differentiable function mapping a set of pose and shape parameters $(\mathbf{R},\mathbf{t},\Theta ,\Phi)$ to a 3D human mesh with vertices $\mathbf{V}$ and faces $\mathbf{F}$ . We thus optimize for $(\mathbf{R},\mathbf{t},\Theta ,\Phi)$ characterizing $\mathcal{B}$ in the scene $S$ , with $\mathbf{R} \in \mathbb{R}^6$ denoting the global orientation as a continuous rotation representation [52], $\mathbf{t} \in \mathbb{R}^3$ the global translation, $\Theta \in \mathbb{R}^{32}$ the body pose in the latent space of VPoser [29] which decodes to body joint rotations $\hat{\Theta} \in \mathbb{R}^{21\times 3}$ , and $\Phi \in \mathbb{R}^{10}$ representing the body shape as blend shape coefficients. + +GenZI synthesizes the desired interaction between the 3D human $\mathcal{B}$ and the scene $S$ by distilling information from VLMs through 2D human inpainting followed by robust 3D lifting. We first generate a collection of plausible 2D human-scene compositions by employing a large VLM to inpaint humans into multiple rendered views of the scene $S$ (Sec. 3.2). We then introduce a robust 3D lifting procedure that optimizes for the pose and shape of the human $\mathcal{B}$ , guided by consistency with the 2D interaction hypotheses (Sec. 3.3). We further refine the posed 3D human $\mathcal{B}$ by iteratively updating the 2D inpaintings and 3D optimization. + +# 3.2. Inpainting Multi-view Interaction Hypotheses + +We first leverage a VLM to generate 2D hypotheses of potential human interactions in the scene by automatically inpainting humans into multiple rendered views of $S$ . + +Multi-view Rendering. To capture scene context for 2D + +human inpainting, we render multiple views of the 3D scene $S$ from $k$ virtual cameras looking at $\mathbf{p}$ . Cameras are randomly sampled on a hemisphere and filtered according to the visibility of $\mathbf{p}$ ; we refer to the supplementary material for additional camera setup details. We denote the rendered scene images as $\{\mathbf{I}^i\}_{i\in [1,k]}$ . To simplify notation, we omit the superscript $i$ indexing each view in this section. + +Inpainting with Dynamic Masking. Given a rendered scene image $\mathbf{I}$ and the text prompt $\Gamma$ , we leverage a state-of-the-art latent 2D diffusion model [34] to generate a new image $\overline{\mathbf{I}}$ , where a human is inpainted into the scene image $\mathbf{I}$ while adhering to the specified interaction and 2D scene context. In practice, we opt for the popular Stable Diffusion Inpainting model [1] as the latent diffusion implementation. + +The latent diffusion model, denoted as $\varOmega(\mathbf{z}_t,\mathbf{M},\mathbf{I},\Gamma,t)$ , performs image inpainting by progressively denoising a noisy latent $\mathbf{z}_t$ at each time step $t$ . During the denoising diffusion process, the binary mask $\mathbf{M}$ defines the inpainting region in the image $\mathbf{I}$ ; however, specifying this mask typically requires manual effort [1]. Thus, we develop a fully-automated inpainting process by automatically generating the mask. Note that using a random or fixed human mask naively can lead to incorrect inpaintings, as this often results in scene context incorrectly masked, e.g., the objects affording the desired action may be entirely masked out, leading to the generated 2D human poses to be incoherent with the scene, producing undesirable 3D HSI synthesis. + +We propose a masking scheme that dynamically adapts the mask through the denoising process by leveraging the internal cross-attention maps [4, 13, 42] from $\Omega$ to propose masks. The cross-attention maps capture rich semantic correlations between image pixels and input text tokens, playing a crucial role in guiding image generation. Let $\mathbf{A} \in \mathbb{R}^{hw \times n}$ denote the cross-attention map between an image feature map of $hw$ pixels and the text prompt $\Gamma$ with $n$ tokens, normalized row-wise using softmax. Here, $\mathbf{A}[i,j]$ signifies the influence of the $j$ -th token on the $i$ -th pixel, and $\mathbf{A}[:,j]$ forms a heat map of image regions to be filled with content related to the $j$ -th token. + +Using the cross-attention map $\mathbf{A}_t$ at each time step $t$ of the diffusion process, we can dynamically derive a mask related to human tokens. Specifically, at time step $t$ , we obtain $\mathbf{z}_{t - 1},\mathbf{A}_t\gets \varOmega (\mathbf{z}_t,\mathbf{M}_t,\mathbf{I},\varGamma ,t)$ after denoising. To create a human inpainting mask $\mathbf{M}_{t - 1}$ for time step $t - 1$ , we extract heat maps from $\mathbf{A}_t$ corresponding to the tokens referring to the human (e.g., "woman" or "man"), followed by summation of these extracted heat maps across the tokens and binarization. At the initial time step $T$ , we initialize $\mathbf{M}_T$ as an empty mask. Our dynamic masking approach enables the synthesis of a 2D human-scene composition image $\bar{\mathbf{I}}$ without the need for manually-specified human inpainting masks. We illustrate our dynamic masking scheme in Fig. 3. + +![](images/6dfdd9debcfca513b105d50197d499e7cbceac967235f155bbf62904d0ca044b.jpg) +Figure 3. Human inpainting with dynamic masking. Top: Given a scene image and a text prompt, a human is inpainted into the image without a mask specifying the inpainting region for latent diffusion. Bottom: The masks generated by our dynamic masking scheme based on cross-attention maps at different diffusion time steps adaptively shift to find the region of interest. + +# 3.3. Robust Lifting to a 3D Interaction + +Given the multi-view scene images inpainted with human subjects $\{\bar{\mathbf{I}}^i\}_{i\in [1,k]}$ , our aim is to optimize for the pose and shape parameters $(\mathbf{R},\mathbf{t},\boldsymbol {\Theta},\boldsymbol {\Phi})$ of a 3D human $\mathcal{B}$ , guided by the multi-view interaction cues. We optimize for $\mathcal{B}$ by matching it with the 2D poses extracted from $\{\bar{\mathbf{I}}^i\}$ . Since the 2D hypotheses may not be consistent across views, we formulate this as a robust optimization, simultaneously optimizing for the set of views most consistent with $\mathcal{B}$ . + +2D Pose Estimation To distill interaction guidance for the 3D HSI synthesis, we compute a 2D pose representation for the inpainted human in each image $\bar{\mathbf{I}}^i$ . We use AlphaPose [6], an off-the-shelf pose estimation approach that infers a set of 2D joint positions $\mathbf{J}^i$ and the corresponding joint confidence scores $\mathbf{c}^i$ for the human subject in image $\bar{\mathbf{I}}^i$ . + +These 2D pose hypotheses $\{\mathbf{J}^i,\mathbf{c}^i\}_{i\in [1,k]}$ are then used to steer the interaction synthesis between the 3D human $\mathcal{B}$ and the scene $S$ . We aim to minimize the following objective function $\mathcal{E}_{\mathrm{total}}$ : + +$$ +\begin{array}{l} \mathcal {E} _ {\text {t o t a l}} = \lambda_ {\mathrm {P F}} \mathcal {E} _ {\mathrm {P F}} + \lambda_ {\mathrm {V S}} \mathcal {E} _ {\mathrm {V S}} + \lambda_ {\mathrm {B P}} \mathcal {E} _ {\mathrm {B P}} + \tag {1} \\ \lambda_ {\mathrm {B S}} \mathcal {E} _ {\mathrm {B S}} + \lambda_ {\mathrm {S C}} \mathcal {E} _ {\mathrm {S C}} + \lambda_ {\mathrm {S P}} \mathcal {E} _ {\mathrm {S P}}, \\ \end{array} +$$ + +where the $\lambda$ denote scalar weights for the energy terms: pose fitting $\mathcal{E}_{\mathrm{PF}}$ , view selection $\mathcal{E}_{\mathrm{VS}}$ , body pose $\mathcal{E}_{\mathrm{BP}}$ , body shape $\mathcal{E}_{\mathrm{BS}}$ , scene penetration $\mathcal{E}_{\mathrm{SC}}$ , and self-penetration $\mathcal{E}_{\mathrm{SP}}$ . + +Robust View-consistent 3D Pose Fitting. Our primary energy term $\mathcal{E}_{\mathrm{PF}}$ minimizes the discrepancy between the projections of the 3D pose of $\mathcal{B}$ and the inpainted poses from multiple views. However, 2D pose hypotheses can often be inconsistent across different views, due to the stochastic nature of diffusion models, leading to conflicting optimization signals for $\mathcal{B}$ . To address this, we employ a robust optimization strategy that optimizes additionally for view selection weights in $\mathcal{E}_{\mathrm{PF}}$ that promote consistency between the 3D pose and the most consistent 2D pose hypotheses. + +We thus introduce a new set of estimizable variables $\mathbf{w} = \{w^{i}|w^{i}\in [0,1]\}_{i\in [1,k]}$ , representing view consistency scores, and apply a robust kernel $\rho$ to the per-view joint fitting constraints. The weights $\mathbf{w}$ allow the solver to adaptively focus on views with consistent inpainted poses and downweight the inconsistent ones. Our pose fitting energy $\mathcal{E}_{\mathrm{PF}}$ with view consistency is formulated as follows: + +$$ +\mathcal {E} _ {\mathrm {P F}} = \frac {\sum_ {i} w ^ {i} \sum_ {j} \mathbf {c} _ {j} ^ {i} \rho (\Pi (\hat {\mathbf {J}}) _ {j} ^ {i} - \mathbf {J} _ {j} ^ {i})}{\sum_ {i} w ^ {i}}, \tag {2} +$$ + +where $\hat{\mathbf{J}}$ denotes the 3D joint positions of the SMPL-X model [29] differentiably computed from the pose and shape parameters $(\mathbf{R},\mathbf{t},\Theta ,\Phi)$ . The function $\varPi(\cdot)_j^i$ represents projection of the $j$ -th joint in the $i$ -th camera view, and $\rho$ is the robust Geman-McClure error function [8]. + +Regularizations. Below, we describe the regularization terms adopted in our optimization. + +1) We encourage the per-view weights $\mathbf{w}$ to focus on at least $\tau$ views: + +$$ +\mathcal {E} _ {\mathrm {V S}} = \max (\tau - \sum_ {i} w ^ {i}, 0). \tag {3} +$$ + +2) To ensure a natural 3D pose, we impose the following energy term on the body pose parameter: + +$$ +\mathcal {E} _ {\mathrm {B P}} = \left\| \boldsymbol {\Theta} \right\| ^ {2} + \mathcal {E} _ {\mathrm {J A}} (\hat {\boldsymbol {\Theta}}). \tag {4} +$$ + +The first term is the VPoser body prior [29] regularizing the latent pose $\Theta$ , and the second term $\mathcal{E}_{\mathrm{JA}}$ is a simple angle prior on the body joint rotations $\hat{\Theta}$ , decoded from the latent pose $\Theta$ by VPoser, to penalize extreme bending. The formulation of $\mathcal{E}_{\mathrm{JA}}$ is provided in the supplementary material. + +3) We regularize the shape parameter $\Phi$ to obtain a plausible body shape via + +$$ +\mathcal {E} _ {\mathrm {B S}} = \left\| \boldsymbol {\Phi} \right\| ^ {2}, \tag {5} +$$ + +which measures the Mahalanobis distance between $\Phi$ and the body shape distribution used in SMPL-X [29]. + +4) To ensure physical contact but also avoid penetration between the 3D human $\mathcal{B}$ and the scene $S$ , we formulate the spatial constraints as: + +$$ +\mathcal {E} _ {\mathrm {S C}} = \left\{ \begin{array}{l l} \min _ {\mathbf {v} \in \mathbf {V}} \Psi (\mathbf {v}), & \Psi (\mathbf {v}) > 0 \forall \mathbf {v} \in \mathbf {V} \\ \sum_ {\mathbf {v} \in \mathbf {V}} | \min (\Psi (\mathbf {v}), 0) |, & \text {o t h e r w i s e} \end{array} \right. \tag {6} +$$ + +where $\varPsi$ is a pre-computed signed distance field (SDF) [10] of the scene $S$ . When $\varPsi(\mathbf{v})$ has a negative sign, it indicates that the body vertex $\mathbf{v}$ is located inside the nearest scene object (i.e., penetration). Conversely, a positive sign means that $\mathbf{v}$ is positioned on the outside. + +5) Finally, to resolve penetration within the human body $\mathcal{B}$ itself, we include a self-penetration energy $\mathcal{E}_{\mathrm{SP}}$ based on detecting colliding body triangles using Bounding Volume Hierarchies (BVH) [40]. We refer the reader to [2, 10, 41] for more formulation details on $\mathcal{E}_{\mathrm{SP}}$ . + +# 3.4. Iterative Refinement + +A posed 3D human $\mathcal{B}$ interacting with the 3D scene $S$ is generated after applying the VLM-based 2D inpainting (Sec. 3.2) and the robust 3D lifting (Sec. 3.3). To improve the synthesis and consistency of the interaction results, we employ an iterative refinement scheme over the aforementioned inpainting and 3D lifting. During refinement, we render the silhouette of the posed 3D human $\mathcal{B}$ in each camera view, and use it as a more precise and consistent mask $\mathbf{M}$ in the latent diffusion inpainting $\Omega$ [34], replacing dynamic masking. By doing so, the consistency among the 2D pose hypotheses inpainted into the multi-view scene images gradually improves, thus leading to an improved 3D HSI synthesis outcome. + +# 3.5. Implementation Details + +We implemented GenZI with PyTorch [28]. For VLM-based 2D inpainting, we use $k = 16$ cameras for multiview rendering. We use 50 denoising steps in Stable Diffusion Inpainting with a state-of-the-art diffusion sampler [25]. For dynamic masking during inpainting, we aggregate cross-attention maps with a resolution of $16 \times 16$ . Input text prompts $\Gamma$ (e.g., "sitting on a bench") are all appended with fixed prefix "a woman" and fixed suffix "wearing a white shirt and blue pants, full body" to better constrain generation. For robust 3D lifting, we set $\tau = 3$ , and optimize the energy $\mathcal{E}_{\mathrm{total}}$ using gradient descent [24] for 1.6K steps, which takes $\sim 3$ minutes on an NVIDIA A100 GPU. In the iterative refinement, we dilate the rendered human silhouette with a kernel size of $11 \times 11$ for mask generation, and the refinement is performed once. + +# 4. Experiments + +We demonstrate the effectiveness and generality of our approach GenZI on a diverse collection of 3D scene models from Sketchfab.com. We conduct both quantitative and qualitative evaluations to compare GenZI with alternative baselines approaches [10, 51] to our new task. + +Dataset. In our Sketchfab dataset, we gathered 8 large-scale 3D scenes encompassing a variety of indoor and outdoor environments with diverse geometric structures, including a realistic Venice city, a gym, and a cartoon-style food truck. We collected 4-5 text prompts per scene describing human interactions with the scene for specified approximate point locations, resulting in 38 actions for evaluation. + +Baselines. To the best of our knowledge, there are no baselines that estimate 3D human-scene interactions based on natural language text input in a zero-shot fashion. We thus perform comparisons with related baseline approaches: + +- COINS [51] is a state-of-the-art approach estimating 3D humans in indoor 3D scans with a fixed vocabulary of actions and objects, given object segmentations. It takes as input (action, object) pairs for semantic control, and requires full supervision in its CVAE training, using captured 3D interaction data with both instance segmentations of 3D objects and action labelling. Due to being trained on a small, closed set of indoor interactions, we adapt COINS to a subset of the most similar Sketchfab actions by manually segmenting corresponding 3D objects from the scenes similar to its indoor training data and mapping the text prompts to its (action, object) input. Note that our approach GenZI does not require any 3D scene segmentations. +- Hassan et al. [10] perform 3D human estimation from a single RGB image. To adapt this to Sketchfab, we reuse the multi-view scene images $\{\bar{\mathbf{T}}^i\}_{i\in [1,k]}$ inpainted with humans from our dynamic masking scheme (Sec. 3.2), where the view with the best image-text cosine similarity (i.e., CLIP score [33]) is used as the input for 3D human estimation under the known virtual camera parameters. +- Ours-Single View: Finally, we consider a baseline leveraging 3D body estimation from our method, but limited to only one inpainted view. The same best view image, as described above, is also as input for this baseline. + +Evaluation Metrics. To measure 3D HSI quality, we conduct perceptual studies and compute metrics including semantic consistency, diversity, and physical plausibility. + +We first carry out two perceptual studies to assess the realism and semantic accuracy of the synthesized interactions. The first is a binary-choice study, where interaction samples generated by two different methods based on the same text prompt are shown, and the participants are asked to choose the sample that is more realistic and better matched the text. The second study is a unary test, where for each interaction sample, the participants are asked to rate the realism and consistency between the shown sample and text prompt from 1 (strongly disagree) to 5 (strongly agree). + +To evaluate the semantic consistency between a synthesized 3D interaction and the input text prompt, we calculate the CLIP score [33], where the 3D interaction is re-rendered into $k$ view images, and the image-text cosine similarities from CLIP ViT-B/32 are averaged across all views. + +Additionally, we include quantitative metrics from existing works to evaluate diversity and physical plausibility. We observe that these metrics often do not reflect perceptual quality, as generated bodies can be diverse but have interpenetrations without demonstrating any physical or semantic coherence. Nevertheless, to measure the synthesis + +![](images/4a2e2d5de573cabf7141f16eee60de174c3e181152b5d9faf737bdeec3b57731.jpg) +(a) Binary Perceptual Study + +![](images/253e1bf636820898e186febf3a12fdddc4e6a75d2b7a668080b2eeffcd357e9a.jpg) +Figure 4. User study of 3D human-scene interaction synthesis on the Sketchfab dataset, where participants show a strong preference for the generations by our approach, in comparison with all baselines, COINS [51], Hassan et al. [10], and Ours-Single View. + +diversity, we follow [50, 51] and cluster the SMPL-X parameters of the generated humans into 20 clusters with K-means and compute the entropy of the cluster ID histogram of all the samples. We also calculate the cluster size as the mean distance to cluster centers. For the physical plausibility, we evaluate the collision and contact between body meshes and scene meshes, following [50, 51]. The noncollision score is computed as the ratio between the number of body mesh vertices with positive SDF values (Sec. 3.3) and the number of all body mesh vertices. The contact score is defined as the ratio between the number of body meshes with scene contact and the number of all generated body meshes. A body mesh is in contact with the scene if any body mesh vertex has a non-positive SDF value. + +# 4.1. Comparison to Alternative Approaches + +Quantitative Evaluation. Fig. 4 shows the results of the perceptual studies collected from 30 participants across a binary study and a unary study. In the binary study, we observe that participants overwhelmingly favor the generations by our GenZI compared to all baselines – more than $87\%$ of the time. In the unary test, the average realism rating for our interaction generations is 3.6, the highest compared to the baselines, which are all below 2.0. These perceptual results strongly indicate that our GenZI can synthesize realistic 3D humans interacting with various 3D scenes without requiring any captured 3D interaction data. + +Tab. 1 presents the quantitative evaluation of semantic consistency, diversity, and physical plausibility on the full Sketchfab dataset. Our approach achieves the best semantic consistency score, echoing the strong user preference for + +![](images/3bcc7b441772f78ba6d3906bc8f2dcbd3064268668baa1e6b2852e13680f45b2.jpg) +"sitting and playing piano, hands on piano keys, feet flat on the ground" + +![](images/0b553c4a2ca70c64b73a76a174629046e40d0d9a3da958dc2f692f91201696e8.jpg) +"climbing a ladder, hands gripping side rails" + +![](images/95d548d2c559076cf004401f95022d748ffcc1c3237f524156b6ae33f3f01379.jpg) +"sitting and paddling a wooden boat, spreading arms" + +![](images/62f8a529f4ba054258b7ee10a315fb638e2269d74295ebd845d966f3b64bd4c0.jpg) +"walking on a bridge" + +![](images/783e81a1c61832f4b15955a1997135e0fe0a3938cec6b9e83cbf06fc1757d1fb.jpg) +"picking up a trash bag, standing, bending over" + +![](images/9b3d53a5b246156290a7403c0b46b64ce7c438769e9a329c849828c7d86b618c.jpg) +"walking out of a door" + +![](images/987e6194f7d98f4c41a880b096dca735b21210925e942c07a15fcbdd926233f0.jpg) +"riding a recumbent bike, sitting" + +![](images/6bc881151455391110e66bfd6aea11299fc44fbd733ccf7c92aeee27f137d351.jpg) +Text Prompt +COINS + +![](images/f01f189760f63e9e1e32e1807e107e25f04586c9ded521ce3e99d4ee76c7c9d1.jpg) +Hassan et al. + +![](images/e162de5ffe520c1abdb1642c60d428332d0640f4223dac9cd529b76290399663.jpg) +Ours-SV + +![](images/04b88ee2ce0ea1478cea3f8a4861966dece09382e8b8480a3865f5025ca5e9b8.jpg) +Ours + +![](images/91883f8c53c2804f6d0f3c9b8bb7995148ee28774ac3cb62a70ca1f92c65f618.jpg) + +![](images/ad9c525edd94a4eaa8e95a422aa4c0ba4f25f1819dd8c4c00a80c46c06c63371.jpg) + +![](images/026ee0159bacf51b95ee2369cf4d7fe81f0fcf6790f735a950cecc523592230c.jpg) + +![](images/585ae1923e61e26e622c1d05e5dd8861825ea6793e0019bef976b875f15387df.jpg) + +![](images/347c82f5728a9ca6ac14e404b6d8d3493f37eef3c244e6af5e9bc08313beade7.jpg) + +![](images/0c9307135aada9be1e64f6dfe7f488493a14c300ae4bac7aa8cfca63c63467fb.jpg) + +![](images/799fe841691ee3c412d33497b7e620a8087bcdb53666489310b98497dc74e68a.jpg) + +![](images/16c317b5dbba0f25b8d1c3938f82bf5305b39b7af39f90d4fb648a4fe027ffac.jpg) + +![](images/1eb9f36e63879ba59c68a502439c6ddaf41925c600bc7d8d93fcc7879f7d98bf.jpg) + +![](images/8d23c4ee33ef316b5a6e5cd0e0658ba908b42ad2066dfbbcca2f4cd9cee0cd16.jpg) + +![](images/1f7eb8ea742aae73c555ff5f2fba8cb5a1f4df94529a6d4674cea1c82cea15e9.jpg) + +![](images/005844162b292bdf668d22ac84f462409c58a86c7f1e37b8dd6b26d36a17d9cc.jpg) + +![](images/49e321fb2240ec8dd772d47048522b37fa8e2c947fb811d5b255e81f8d6e42ac.jpg) + +![](images/9145aaf37bf9c5c3ac2505f2c4b5bfd4304c0238bf003297bd9facc2f7212263.jpg) + +![](images/cd2327af6bfd58ee031f311cddba687dbf504298e267d65faadc5b9d54370eb0.jpg) + +![](images/c68d305062d3c01a8d9e6ddc2b32f134bf26dbcb7b1fd2bf024d8b048b3a7967.jpg) + +![](images/195c72893bf9b901b89ddcf561ef1990dc1f4847071cb3651c225dfa30ab8004.jpg) +Figure 5. Qualitative results on the Sketchfab dataset. Our GenZI synthesizes more realistic 3D human-scene interactions and generalizes better across diverse scene types, compared to the baselines COINS [51], Hassan et al. [10], and Ours-Single View. For COINS, we show the used (action, object) labels from its closed set of indoor interactions; its closed setting can lead to degraded results from out-of-distribution object classes (e.g., curved bridge deck as the floor, chair at a different height or shape than those in the training set). + +![](images/a398ac2ebd84c478cc3bfb296599c800f0137c4c802293d9fdbf47f12a443de8.jpg) + +![](images/e85d45af699882bcd1def9e555d2cd2bf10139ef73e17d47ffa1c82237fb6f36.jpg) + +![](images/19c9b6eeaceb1c10ba439b354a5185d46cf6c8f4c1f4a9e62a2333cf0486d34c.jpg) + +
MethodSemanticsDiversityPhysical Plausibility
CLIP ↑Entropy ↑Cluster Size ↑Non-collision ↑Contact ↑
Hassan et al. [10]0.25982.70141.19070.88240.9669
Ours-SV0.26132.64521.58130.97651.0000
Ours0.27102.73041.05000.97670.9868
+ +Table 1. Quantitative comparisons on the Sketchfab dataset. Our approach achieves the best semantic consistency, diversity entropy, and non-collision scores, with the contact score on par. Note that single view methods Hassan et al. [10] and Ours-SV tend to produce increased diversity at the cost of semantic plausibility. + +
w/o DMw/o VCw/o IROurs
CLIP ↑0.26390.26940.26640.2710
+ +Table 2. Ablation study on the Sketchfab dataset. The semantic consistency of 3D interaction generations degrades without dynamic masking (DM), view consistency (VC), or iterative refinement (IR), compared to our full approach. + +![](images/e1f5615d87b51a9a6de42d9775a816c93365f793d43cda462f9937ee3c946080.jpg) +Figure 6. Visualization of our method ablations on Sketchfab dataset for the input text: "sitting on a bar stool". Without dynamic masking (DM) or view consistency (VC), the person floats above the middle stool. Without iterative refinement (IR), the person penetrates the stool. Our full approach results in a more realistic synthesis. + +GenZI in the perceptual studies. We note that since both Hassan et al. [10] and Ours-Single View operate on single inpainted view samples, results can be very diverse but lack 3D plausibility. This suggests that the CLIP score is a more reliable metric for assessing the HSI synthesis quality, compared to the diversity and physical plausibility metrics, which reflect less about the generation realism. Nevertheless, our approach has the best diversity entropy and noncollision scores, with the contact score on par. + +Qualitative Evaluation. We show qualitative comparisons in Fig. 5. COINS is severely limited by its training on the closed set of indoor interactions, and thus fails to generalize to outdoor scenes and unseen objects (e.g., no curved floors exist during training, and limited sets of heights and shapes of chairs). As Hassan et al. [10] and Ours-Single View operate from single views, they both suffer from insufficient pose constraints from other views for plausible interaction generation. In contrast, our approach demonstrates high flexibility and generality to a diverse set of 3D indoor and outdoor scenes by leveraging large VLMs to imagine multiview interaction hypotheses and then robust 3D lifting. + +# 4.2. Ablation Studies + +We conduct ablation studies on the Sketchfab dataset to validate the effectiveness of our proposed dynamic masking + +scheme (Sec. 3.2), robust 3D lifting with view consistency (Sec. 3.3), and iterative refinement (Sec. 3.4). Results are presented in Fig. 6 and Tab. 2. + +Dynamic Masking (DM). We replace our dynamic masking, used during latent diffusion inpainting, with random masking. We sample a random mask around the image center covering at least $30\%$ of the image area, and use it as a fixed mask input to the latent diffusion model $\Omega$ . We observe that random masking results in noticeably worse quality of interaction synthesis (Tab. 2), and incoherence with the scene (Fig. 6, floating on the stool). This indicates that our dynamic masking is effective in incorporating sufficient scene context for human inpainting. + +View Consistency (VC). We evaluate the role of view consistency in robust 3D lifting by fixing the estimizable scores to $\mathbf{w} = \mathbf{1}$ . Using all inpainted views leads to averaged, less expressive 3D human poses (Tab. 2, Fig. 6) due to potential inconsistent 2D pose hypotheses across views. By allowing the solver to adaptively focus on views with consistent inpaintings, our approach generates more realistic 3D HSIs. + +Iterative Refinement (IR). Finally, we show the effectiveness of iterative refinement by applying our VLM-based 2D inpainting and robust 3D lifting only once. Tab. 2 and Fig. 6 show that iterative refinement improves synthesis quality. + +Limitations. Our approach is limited by the inpainting capability of latent diffusion models to imagine possible 2D human-scene compositions, and the diffusion models are also known to be slow at inference time due to their iterative nature [14]. Nevertheless, we believe that our approach can directly benefit from the rapid advancement of VLMs for improved HSI synthesis. + +# 5. Conclusion + +We have presented the first approach to synthesize general 3D human-scene interactions guided by text inputs. Key to our approach is effective distillation of knowledge from large vision-language models, enabling generating 3D humans in scenes without requiring any 3D interaction data for training. We leverage these powerful vision-language models to generate hypotheses for inpainted 2D human-scene interactions. We then formulate a robust optimization to lift the hypotheses to 3D in a view-consistent fashion by simultaneously optimizing for the most informative 2D hypotheses. Our approach is flexible and can be applied to general scene settings for a variety of actions. We believe this opens up new opportunities for 3D understanding without requiring expensive capture of 3D/4D data. + +Acknowledgements. This project is funded by the ERC Starting Grant SpatialSem (101076253), and the German Research Foundation (DFG) Grant "Learning How to Interact with Scenes through Part-Based Understanding", and supported in part by a Google research gift. + +# References + +[1] Stable Diffusion v2. https://huggingface.co/stabilityai/stable-diffusion-2-inpainting.Accessed:2023-09-27.4 +[2] Luca Ballan, Aparna Taneja, Jürgen Gall, Luc Van Gool, and Marc Pollefeys. Motion capture of hands in action using discriminative salient points. In ECCV, 2012. 5 +[3] Zhongang Cai, Wanqi Yin, Ailing Zeng, Chen Wei, Qingping Sun, Yanjun Wang, Hui En Pang, Haiyi Mei, Mingyuan Zhang, Lei Zhang, et al. SMPLer-X: Scaling up expressive human pose and shape estimation. arXiv, 2023. 2 +[4] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-Excite: Attention-based semantic guidance for text-to-image diffusion models. ACM TOG, 2023. 4 +[5] Vasileios Choutas, Georgios Pavlakos, Timo Bolkart, Dimitrios Tzionas, and Michael J Black. Monocular expressive body regression through body-driven attention. In ECCV, 2020. 2 +[6] Hao-Shu Fang, Jiefeng Li, Hongyang Tang, Chao Xu, Haoyi Zhu, Yuliang Xiu, Yong-Lu Li, and Cewu Lu. Alpha-Pose: Whole-body regional multi-person pose estimation and tracking in real-time. IEEE TPAMI, 2022. 4 +[7] Dariu M Gavrila. The visual analysis of human movement: A survey. CVIU, 1999. 2 +[8] Stuart Geman and Donald E. McClure. Statistical methods for tomographic image reconstruction. 1987. 5 +[9] James J Gibson. The ecological approach to the visual perception of pictures. *Leonardo*, 11(3):227-235, 1978. 1 +[10] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J. Black. Resolving 3D human pose ambiguities with 3D scene constraints. In ICCV, 2019. 1, 2, 5, 6, 7, 8 +[11] Mohamed Hassan, Partha Ghosh, Joachim Tesch, Dimitrios Tzionas, and Michael J. Black. Populating 3D scenes by learning human-scene interaction. In CVPR, 2021. 1, 2 +[12] Mohamed Hassan, Yunrong Guo, Tingwu Wang, Michael Black, Sanja Fidler, and Xue Bin Peng. Synthesizing physical character-scene interactions. In SIGGRAPH, 2023. 2 +[13] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-prompt image editing with cross-attention control. In ICLR, 2023. 4 +[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 8 +[15] Lukas Hollein, Ang Cao, Andrew Owens, Justin Johnson, and Matthias Nießner. Text2Room: Extracting textured 3d meshes from 2d text-to-image models. In ICCV, 2023. 3 +[16] Siyuan Huang, Zan Wang, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, and Song-Chun Zhu. Diffusion-based generation, optimization, and planning in 3D scenes. In CVPR, 2023. 2 +[17] Kurt Koffka. Principles of gestalt psychology. 1935. 1 +[18] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In ICCV, 2019. 2 + +[19] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. BLIP: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 2 +[20] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 2, 3 +[21] Xueting Li, Sifei Liu, Kihwan Kim, Xiaolong Wang, Ming-Hsuan Yang, and Jan Kautz. Putting humans in a scene: Learning affordance in 3D indoor environments. In CVPR, 2019. 2 +[22] Jing Lin, Ailing Zeng, Haoqian Wang, Lei Zhang, and Yu Li. One-stage 3d whole-body mesh recovery with component aware transformer. In CVPR, pages 21159-21168, 2023. 2 +[23] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. SMPL: A skinned multiperson linear model. ACM TOG, 34(6):1-16, 2015. 2, 3 +[24] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 5 +[25] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. DPM-Solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv, 2022. 5 +[26] Thomas B Moeslund, Adrian Hilton, and Volker Krüger. A survey of advances in vision-based human motion capture and analysis. CVIU, 2006. 2 +[27] Aron Monszpart, Paul Guerrero, Duygu Ceylan, Ersin Yumer, and Niloy J Mitra. iMapper: interaction-guided scene mapping from monocular videos. ACM Transactions On Graphics (TOG), 38(4):1-15, 2019. 2 +[28] A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. Kopf, E. Yang, Z. DeVito, M. Raison, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, and S. Chintala. PyTorch: An imperative style, high-performance deep learning library. NeurIPS, 2019. 5 +[29] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3D hands, face, and body from a single image. In CVPR, 2019. 2, 3, 5 +[30] R. Po, W. Yifan, V. Golyanik, K. Aberman, J. T. Barron, A. H. Bermano, E. R. Chan, T. Dekel, A. Holynski, A. Kanazawa, C. K. Liu, L. Liu, B. Mildenhall, M. Nießner, B. Ommer, C. Theobalt, P. Wonka, and G. Wetzstein. State of the art on diffusion models for visual computing. arXiv, 2023. 2 +[31] Ronald Poppe. Vision-based human motion analysis: An overview. CVIU, 2007. 2 +[32] Sergey Prokudin, Christoph Lassner, and Javier Romero. Efficient learning on point clouds with basis point sets. In ICCV, 2019. 2 +[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3, 6 + +[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3, 4, 5 +[35] Yu Rong, Takaaki Shiratori, and Hanbyul Joo. FrankMocap: A monocular 3d whole-body pose estimation system via regression and integration. In ICCV, 2021. 2 +[36] David Rozenberszki, Or Litany, and Angela Dai. Language-grounded indoor 3d semantic segmentation in the wild. In ECCV, 2022. 3 +[37] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 2022. 2, 3 +[38] Nikolaos Sarafianos, Bogdan Boteanu, Bogdan Ionescu, and Ioannis A Kakadiaris. 3D human pose estimation: A review of the literature and analysis of covariates. CVIU, 2016. 2 +[39] Manolis Savva, Angel X Chang, Pat Hanrahan, Matthew Fisher, and Matthias Nießner. PiGraphs: learning interaction snapshots from observations. ACM TOG, 35(4):1-12, 2016. 2 +[40] Matthias Teschner, Stefan Kummerle, Bruno Heidelberger, Gabriel Zachmann, Laks Raghupathi, Arnulph Fuhrmann, M-P Cani, François Faure, Nadia Magnenat-Thalmann, Wolfgang Strasser, et al. Collision detection for deformable objects. In CGF, 2005. 5 +[41] Dimitrios Tzionas, Luca Ballan, Abhilash Srikantha, Pablo Aponte, Marc Pollefeys, and Juergen Gall. Capturing hands in action using discriminative salient points and physics simulation. IJCV, 2016. 5 +[42] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 4 +[43] Jingbo Wang, Yu Rong, Jingyuan Liu, Sijie Yan, Dahua Lin, and Bo Dai. Towards diverse and natural scene-aware 3D human motion synthesis. In CVPR, 2022. 2 +[44] Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, Wei Liang, and Siyuan Huang. HUMANISE: Language-conditioned human motion generation in 3D scenes. NeurIPS, 2022. 2 +[45] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In CVPR, 2023. 3 +[46] Haibiao Xuan, Xiongzheng Li, Jinsong Zhang, Hongwen Zhang, Yebin Liu, and Kun Li. Narrator: Towards natural control of human-scene interaction generation via relationship reasoning. arXiv, 2023. 1 +[47] Yufei Ye, Xueting Li, Abhinav Gupta, Shalini De Mello, Stan Birchfield, Jiaming Song, Shubham Tulsiani, and Sifei Liu. Affordance diffusion: Synthesizing hand-object interactions. In CVPR, 2023. 3 +[48] Hongwen Zhang, Yating Tian, Yuxiang Zhang, Mengcheng Li, Liang An, Zhenan Sun, and Yebin Liu. PyMAF-X: Towards well-aligned full-body model regression from monocular images. IEEE TPAMI, 2023. 2 + +[49] Siwei Zhang, Yan Zhang, Qianli Ma, Michael J Black, and Siyu Tang. Place: Proximity learning of articulation and contact in 3D environments. In 3DV, 2020. 2 +[50] Yan Zhang, Mohamed Hassan, Heiko Neumann, Michael J. Black, and Siyu Tang. Generating 3D people in scenes without people. In CVPR, 2020. 1, 2, 6 +[51] Kaifeng Zhao, Shaofei Wang, Yan Zhang, Thabo Beeler, and Siyu Tang. Compositional human-scene interaction synthesis with semantic control. In ECCV, 2022. 1, 2, 5, 6, 7 +[52] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In CVPR, 2019. 3 +[53] Yuxiao Zhou, Marc Habermann, Ikhsanul Habibie, Ayush Tewari, Christian Theobalt, and Feng Xu. Monocular real-time full body capture with inter-part correlations. In CVPR, 2021. 2 \ No newline at end of file diff --git a/2024/GenZI_ Zero-Shot 3D Human-Scene Interaction Generation/images.zip b/2024/GenZI_ Zero-Shot 3D Human-Scene Interaction Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..16c89457cb0f2779226db0b86fcf7210d273626f --- /dev/null +++ b/2024/GenZI_ Zero-Shot 3D Human-Scene Interaction Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab9e474891e0a684ce681541817a4006c45109aee7eee5b18d49b9e1bdf04186 +size 650469 diff --git a/2024/GenZI_ Zero-Shot 3D Human-Scene Interaction Generation/layout.json b/2024/GenZI_ Zero-Shot 3D Human-Scene Interaction Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c889675568fa95aabf9b3601d5be9fcadd6eb790 --- /dev/null +++ b/2024/GenZI_ Zero-Shot 3D Human-Scene Interaction Generation/layout.json @@ -0,0 +1,10744 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 111, + 103, + 481, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 103, + 481, + 119 + ], + "spans": [ + { + "bbox": [ + 111, + 103, + 481, + 119 + ], + "type": "text", + "content": "GenZI: Zero-Shot 3D Human-Scene Interaction Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 217, + 144, + 373, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 144, + 373, + 172 + ], + "spans": [ + { + "bbox": [ + 217, + 144, + 373, + 172 + ], + "type": "text", + "content": "Lei Li Angela Dai Technical University of Munich" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 63, + 203, + 162, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 203, + 162, + 215 + ], + "spans": [ + { + "bbox": [ + 63, + 203, + 162, + 215 + ], + "type": "text", + "content": "\"riding a motorcycle, sitting\"" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 181, + 199, + 291, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 199, + 291, + 219 + ], + "spans": [ + { + "bbox": [ + 181, + 199, + 291, + 219 + ], + "type": "text", + "content": "\"hands knocking on the closed door, facing the door, standing\"" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 199, + 412, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 199, + 412, + 219 + ], + "spans": [ + { + "bbox": [ + 305, + 199, + 412, + 219 + ], + "type": "text", + "content": "\"stretching legs and sitting on the saddle on a standing cow\"" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 429, + 199, + 533, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 429, + 199, + 533, + 219 + ], + "spans": [ + { + "bbox": [ + 429, + 199, + 533, + 219 + ], + "type": "text", + "content": "\"picking up dumbbells on a shelf, standing, bending over\"" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 50, + 222, + 172, + 343 + ], + "blocks": [ + { + "bbox": [ + 50, + 222, + 172, + 343 + ], + "lines": [ + { + "bbox": [ + 50, + 222, + 172, + 343 + ], + "spans": [ + { + "bbox": [ + 50, + 222, + 172, + 343 + ], + "type": "image", + "image_path": "7fb1eb99e239083e4c92aa15a562c585e889e82f9690fca9541c91db3f6ce443.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 347, + 546, + 392 + ], + "lines": [ + { + "bbox": [ + 46, + 347, + 546, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 347, + 546, + 392 + ], + "type": "text", + "content": "Figure 1. Given an arbitrary 3D scene, GenZI can synthesize virtual humans interacting with the 3D environment at specified locations from a brief text description. Our approach does not require any 3D human-scene interaction training data or 3D learning. By distilling interaction priors from powerful 2D vision-language models, we optimize for 3D human-scene interaction synthesis in a flexible fashion, with simple language-based control and high generality to various types of scene environments." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 174, + 223, + 296, + 343 + ], + "blocks": [ + { + "bbox": [ + 174, + 223, + 296, + 343 + ], + "lines": [ + { + "bbox": [ + 174, + 223, + 296, + 343 + ], + "spans": [ + { + "bbox": [ + 174, + 223, + 296, + 343 + ], + "type": "image", + "image_path": "cd67976967625ecb4d433022798ca86154c59cf857218566c4b55c3e95759441.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 298, + 223, + 419, + 343 + ], + "blocks": [ + { + "bbox": [ + 298, + 223, + 419, + 343 + ], + "lines": [ + { + "bbox": [ + 298, + 223, + 419, + 343 + ], + "spans": [ + { + "bbox": [ + 298, + 223, + 419, + 343 + ], + "type": "image", + "image_path": "19bdeca1d364be38016848ec3d9a7e77e40a1a765103878f229318030c6476fe.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 420, + 223, + 542, + 343 + ], + "blocks": [ + { + "bbox": [ + 420, + 223, + 542, + 343 + ], + "lines": [ + { + "bbox": [ + 420, + 223, + 542, + 343 + ], + "spans": [ + { + "bbox": [ + 420, + 223, + 542, + 343 + ], + "type": "image", + "image_path": "e649850e33763e06dc50ad25785d560cd3fa23e67c0c638c07bba83aa33b4df6.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 143, + 411, + 191, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 411, + 191, + 423 + ], + "spans": [ + { + "bbox": [ + 143, + 411, + 191, + 423 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 433, + 289, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 433, + 289, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 433, + 289, + 685 + ], + "type": "text", + "content": "Can we synthesize 3D humans interacting with scenes without learning from any 3D human-scene interaction data? We propose GenZI1, the first zero-shot approach to generating 3D human-scene interactions. Key to GenZI is our distillation of interaction priors from large vision-language models (VLMs), which have learned a rich semantic space of 2D human-scene compositions. Given a natural language description and a coarse point location of the desired interaction in a 3D scene, we first leverage VLMs to imagine plausible 2D human interactions inpainted into multiple rendered views of the scene. We then formulate a robust iterative optimization to synthesize the pose and shape of a 3D human model in the scene, guided by consistency with the 2D interaction hypotheses. In contrast to existing learning-based approaches, GenZI circumvents the conventional need for captured 3D interaction data, and allows for flexible control of the 3D interaction synthesis with easy-to-use text prompts. Extensive experiments show that our zero-shot approach has high flexibility and generality, making it applicable to diverse scene types, including both indoor and outdoor environments." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 411, + 386, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 386, + 423 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 386, + 423 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 434, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 545, + 555 + ], + "type": "text", + "content": "Scenes are typically constructed to enable human interactions with the environment, like sitting on a couch, playing a piano, or opening a mailbox. Understanding these interactions between humans and scenes, also known as affordances [9, 17], has gained increasing attention in the fields of computer vision and graphics. In particular, achieving controllable and generalizable synthesis of human-scene interactions (or HSIs) holds immense potential for various applications, such as robotics, architectural design, video games, and virtual reality experiences, among many others." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 557, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 546, + 715 + ], + "type": "text", + "content": "Generating realistic humans interacting with a 3D scene is a challenging task, requiring holistic semantic understanding of both the environment and possible human actions therein. Existing approaches to HSI synthesis [11, 46, 50, 51] rely heavily on supervised training using meticulously captured data of real people interacting in 3D environments. Unfortunately, collecting large-scale datasets of 3D scenes and human interactions is exorbitantly difficult. It not only demands accurate tracking and reconstruction of both people and their environments, but also needs to ensure sufficient diversity in subjects and scenes. Existing HSI datasets currently contain very limited quantities of scenes and actions, for example, PROX [10, 51] only consisting" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 215, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 215, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 215, + 714 + ], + "type": "text", + "content": "1Project page: craigileli.github.io/projects/genzi" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "20465" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": "of 12 indoor scenes and humans interacting with 11 object categories. The scarcity of ground truth data for supervision has thus strongly limited the applicability and generalization of the learning-based approaches for synthesizing diverse sets of actions in arbitrary 3D scenes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 287, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 287, + 263 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 287, + 263 + ], + "type": "text", + "content": "We thus consider an alternative perspective to HSI synthesis and pose the following question: Can we achieve plausible HSI synthesis without using any captured 3D interaction data? To this end, we present a novel zero-shot approach to 3D HSI generation. We propose the first method to leverage the powerful capabilities exhibited by recent vision-language models (VLMs) [19, 20, 30, 33, 34, 37] to synthesize plausible 2D images of human interactions, and introduce a robust optimization to distill inferred 2D pose information into 3D human synthesis in a 3D scene." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 266, + 287, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 287, + 480 + ], + "type": "text", + "content": "More concretely, given a 3D scene, a text prompt and a coarse point location of the desired interaction, GenZI optimizes for the pose and shape of a 3D human performing the action in the scene, guided by a large VLM [34]. We first leverage the VLM to imagine possible 2D humans by inpainting images from multiple rendered views of the scene. We automate this 2D human insertion process with a dynamic masking scheme that automatically updates proposed masks through the inpainting process, eliminating the need for manual specifications of human inpainting regions. We then lift these 2D interaction hypotheses to 3D and optimize for a parametric 3D human body model [23, 29] that is most consistent with the 2D pose guidance. We further refine the generated 3D human in the scene by iterating through the VLM-based 2D inpainting and robust 3D lifting stages. We demonstrate the flexibility and generality of GenZI in various types of 3D environments (Fig. 1), encompassing both indoor and outdoor scenes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 483, + 242, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 483, + 242, + 494 + ], + "spans": [ + { + "bbox": [ + 58, + 483, + 242, + 494 + ], + "type": "text", + "content": "In summary, our contributions are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 495, + 286, + 638 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 47, + 495, + 286, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 495, + 286, + 554 + ], + "spans": [ + { + "bbox": [ + 47, + 495, + 286, + 554 + ], + "type": "text", + "content": "- We introduce GenZI, the first zero-shot approach to generating realistic 3D humans interacting with a 3D scene from natural language prompts. GenZI does not require any supervision from 3D interaction data, thus enabling flexible synthesis across diverse scenes and actions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 555, + 286, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 555, + 286, + 602 + ], + "spans": [ + { + "bbox": [ + 47, + 555, + 286, + 602 + ], + "type": "text", + "content": "- We propose a dynamically-masked inpainting scheme that allows for the synthesis of plausible 2D human-scene compositions via VLMs without requiring manually-specified human inpainting masks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 603, + 286, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 603, + 286, + 638 + ], + "spans": [ + { + "bbox": [ + 47, + 603, + 286, + 638 + ], + "type": "text", + "content": "- We develop a robust 3D pose optimization to lift various inferred images of human interactions to a view-consistent, realistic 3D HSI synthesis." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 653, + 133, + 665 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 653, + 133, + 665 + ], + "spans": [ + { + "bbox": [ + 47, + 653, + 133, + 665 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": "3D Human-Scene Interaction Synthesis. Synthesizing humans in scenes is an important, challenging task in computer vision and graphics, as it models complex high-level" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "semantic understanding such as affordances and interactions. Existing approaches for human-scene interaction synthesis [11, 12, 16, 21, 27, 39, 43, 44, 50, 51] focus on learning priors from available collected data of people interacting with 3D indoor scenes [10] in a supervised manner." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 133, + 545, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 545, + 373 + ], + "type": "text", + "content": "The early work PiGraphs [39] introduced a probabilistic graphical model of human-object interactions, generating static human-object placements by pose sampling and model retrieval. Zhang et al. [50] introduced a generative model of human-scene interaction, using a conditional variational autoencoder (CVAE) to model the distribution of 3D human poses conditioned on scene depth and semantics. PLACE [49] explicitly represented human-scene interactions with Basis Point Sets (BPS) encoding [32], training a CVAE to synthesize such representations for generating natural human poses and contact relations within scenes. POSA [11] proposed an ego-centric representation by augmenting the SMPL-X model [29] with contact labels and scene semantics, using a CVAE conditioned on SMPL-X vertex positions to model potential interactions. Recently, COINS [51] presented a method for compositional human-scene interaction synthesis with high-level semantic control, using transformer-based CVAEs conditioned on provided 3D objects and interaction semantics to regress human body poses and contact features." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 374, + 545, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 374, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 374, + 545, + 529 + ], + "type": "text", + "content": "However, these works all require supervision from 3D human-scene interactions – such datasets are difficult and expensive to capture and annotate, for instance, the widely used PROX dataset [10, 51] only has 8 reconstructed indoor scenes for training and 4 for testing, capturing human interactions with 11 object categories. As such, these methods are limited to in-domain synthesis, prohibiting applicability to general 3D scene settings with arbitrary objects and arrangements. In this work, we deviate from the learning-based approach and propose to take advantage of established large vision-language models [20, 33, 34, 37] for human-scene interaction synthesis, thereby bypassing the requirement for data capture and 3D learning." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": "3D Human Estimation from RGB Images. Over the past decades, significant progress has been made in 3D human estimation from RGB images [7, 26, 31, 38]. Among those, the prominent work SMPLify-X [29] fits the SMPL-X model to 2D joints estimated from a single image through optimization. Hassan et al. [10] build upon SMPLify-X and estimate 3D human poses by incorporating additional physical contact and penetration constraints between human and 3D scene. Learning-based approaches [3, 5, 18, 22, 35, 48, 53] have received much research attention in recent years towards tackling 3D human estimation from monocular images with improved hand and face estimation. Our work leverages 2D pose reasoning from multiple different 2D view hypotheses, and proposes a robust 3D formulation for aggregating the various 2D hypotheses to a consistent" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20466" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 545, + 293 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 545, + 293 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 545, + 293 + ], + "type": "image", + "image_path": "486a6430454ccd00503ed6038de0c7e738b69f3716bb2625c649fb57b26d31fe.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 303, + 546, + 369 + ], + "lines": [ + { + "bbox": [ + 46, + 303, + 546, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 303, + 546, + 369 + ], + "type": "text", + "content": "Figure 2. GenZI distills information from vision-language model for 3D human-scene interaction. We first leverage large vision-language models to synthesize possible 2D humans interactions with the 3D scene " + }, + { + "bbox": [ + 46, + 303, + 546, + 369 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 303, + 546, + 369 + ], + "type": "text", + "content": " by employing latent diffusion inpainting [34] on multiple rendered views of the environment at location " + }, + { + "bbox": [ + 46, + 303, + 546, + 369 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 46, + 303, + 546, + 369 + ], + "type": "text", + "content": " using our dynamic masking scheme to automatically estimate inpainting masks. We then lift these 2D hypotheses to 3D in a robust optimization for a 3D parametric body model " + }, + { + "bbox": [ + 46, + 303, + 546, + 369 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 46, + 303, + 546, + 369 + ], + "type": "text", + "content": " (SMPL-X [29]) that is most consistent with detected 2D poses in the inpainted 2D hypotheses. This produces a semantically consistent interaction that respects the scene context, without requiring any 3D human-scene interaction data." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 381, + 229, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 381, + 229, + 393 + ], + "spans": [ + { + "bbox": [ + 47, + 381, + 229, + 393 + ], + "type": "text", + "content": "3D human body interacting with a 3D scene." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 396, + 288, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 288, + 516 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 288, + 516 + ], + "type": "text", + "content": "Distilling Prior from Vision-Language Models. Recent advances in powerful vision-language models [20, 33, 34, 37] have also inspired various works aiming to distill information learned through the models to various tasks, including 2D panoptic segmentation [45], 3D semantic segmentation [36], 3D scene generation [15], and synthesis of images with hand-object interactions [47]. Our approach leverages the 2D generative capacity of these models to convey information about possible human-scene interactions, which we then lift to a 3D, consistent interaction." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 529, + 103, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 529, + 103, + 542 + ], + "spans": [ + { + "bbox": [ + 47, + 529, + 103, + 542 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 550, + 115, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 550, + 115, + 561 + ], + "spans": [ + { + "bbox": [ + 47, + 550, + 115, + 561 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 569, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 569, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 569, + 287, + 652 + ], + "type": "text", + "content": "Our objective is to synthesize plausible 3D humans interacting with a 3D scene, guided by input text descriptions, in the absence of 3D interaction data capture for learning. We present GenZI, a novel optimization-based multi-view approach that leverages large VLMs to infer spatial relations of interactions between a human and the scene. Fig. 2 shows an illustration of our approach." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 653, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 713 + ], + "type": "text", + "content": "GenZI takes as input a 3D scene " + }, + { + "bbox": [ + 46, + 653, + 288, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 653, + 288, + 713 + ], + "type": "text", + "content": ", a text prompt " + }, + { + "bbox": [ + 46, + 653, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 46, + 653, + 288, + 713 + ], + "type": "text", + "content": " describing the desired interaction, and an approximate point location " + }, + { + "bbox": [ + 46, + 653, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{p} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 653, + 288, + 713 + ], + "type": "text", + "content": " in the scene around which the interaction should occur. Our approach generates a posed 3D human " + }, + { + "bbox": [ + 46, + 653, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 46, + 653, + 288, + 713 + ], + "type": "text", + "content": " as output, performing the specified action in the scene. We" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 381, + 546, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 381, + 546, + 501 + ], + "spans": [ + { + "bbox": [ + 304, + 381, + 546, + 501 + ], + "type": "text", + "content": "adopt SMPL-X [23, 29] to parameterize the 3D human " + }, + { + "bbox": [ + 304, + 381, + 546, + 501 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 304, + 381, + 546, + 501 + ], + "type": "text", + "content": " as it provides a fully differentiable function mapping a set of pose and shape parameters " + }, + { + "bbox": [ + 304, + 381, + 546, + 501 + ], + "type": "inline_equation", + "content": "(\\mathbf{R},\\mathbf{t},\\Theta ,\\Phi)" + }, + { + "bbox": [ + 304, + 381, + 546, + 501 + ], + "type": "text", + "content": " to a 3D human mesh with vertices " + }, + { + "bbox": [ + 304, + 381, + 546, + 501 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 304, + 381, + 546, + 501 + ], + "type": "text", + "content": " and faces " + }, + { + "bbox": [ + 304, + 381, + 546, + 501 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 304, + 381, + 546, + 501 + ], + "type": "text", + "content": ". We thus optimize for " + }, + { + "bbox": [ + 304, + 381, + 546, + 501 + ], + "type": "inline_equation", + "content": "(\\mathbf{R},\\mathbf{t},\\Theta ,\\Phi)" + }, + { + "bbox": [ + 304, + 381, + 546, + 501 + ], + "type": "text", + "content": " characterizing " + }, + { + "bbox": [ + 304, + 381, + 546, + 501 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 304, + 381, + 546, + 501 + ], + "type": "text", + "content": " in the scene " + }, + { + "bbox": [ + 304, + 381, + 546, + 501 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 381, + 546, + 501 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 304, + 381, + 546, + 501 + ], + "type": "inline_equation", + "content": "\\mathbf{R} \\in \\mathbb{R}^6" + }, + { + "bbox": [ + 304, + 381, + 546, + 501 + ], + "type": "text", + "content": " denoting the global orientation as a continuous rotation representation [52], " + }, + { + "bbox": [ + 304, + 381, + 546, + 501 + ], + "type": "inline_equation", + "content": "\\mathbf{t} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 381, + 546, + 501 + ], + "type": "text", + "content": " the global translation, " + }, + { + "bbox": [ + 304, + 381, + 546, + 501 + ], + "type": "inline_equation", + "content": "\\Theta \\in \\mathbb{R}^{32}" + }, + { + "bbox": [ + 304, + 381, + 546, + 501 + ], + "type": "text", + "content": " the body pose in the latent space of VPoser [29] which decodes to body joint rotations " + }, + { + "bbox": [ + 304, + 381, + 546, + 501 + ], + "type": "inline_equation", + "content": "\\hat{\\Theta} \\in \\mathbb{R}^{21\\times 3}" + }, + { + "bbox": [ + 304, + 381, + 546, + 501 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 381, + 546, + 501 + ], + "type": "inline_equation", + "content": "\\Phi \\in \\mathbb{R}^{10}" + }, + { + "bbox": [ + 304, + 381, + 546, + 501 + ], + "type": "text", + "content": " representing the body shape as blend shape coefficients." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 502, + 547, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 502, + 547, + 634 + ], + "spans": [ + { + "bbox": [ + 304, + 502, + 547, + 634 + ], + "type": "text", + "content": "GenZI synthesizes the desired interaction between the 3D human " + }, + { + "bbox": [ + 304, + 502, + 547, + 634 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 304, + 502, + 547, + 634 + ], + "type": "text", + "content": " and the scene " + }, + { + "bbox": [ + 304, + 502, + 547, + 634 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 502, + 547, + 634 + ], + "type": "text", + "content": " by distilling information from VLMs through 2D human inpainting followed by robust 3D lifting. We first generate a collection of plausible 2D human-scene compositions by employing a large VLM to inpaint humans into multiple rendered views of the scene " + }, + { + "bbox": [ + 304, + 502, + 547, + 634 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 502, + 547, + 634 + ], + "type": "text", + "content": " (Sec. 3.2). We then introduce a robust 3D lifting procedure that optimizes for the pose and shape of the human " + }, + { + "bbox": [ + 304, + 502, + 547, + 634 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 304, + 502, + 547, + 634 + ], + "type": "text", + "content": ", guided by consistency with the 2D interaction hypotheses (Sec. 3.3). We further refine the posed 3D human " + }, + { + "bbox": [ + 304, + 502, + 547, + 634 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 304, + 502, + 547, + 634 + ], + "type": "text", + "content": " by iteratively updating the 2D inpaintings and 3D optimization." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 643, + 543, + 656 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 643, + 543, + 656 + ], + "spans": [ + { + "bbox": [ + 305, + 643, + 543, + 656 + ], + "type": "text", + "content": "3.2. Inpainting Multi-view Interaction Hypotheses" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 662, + 545, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 662, + 545, + 698 + ], + "spans": [ + { + "bbox": [ + 304, + 662, + 545, + 698 + ], + "type": "text", + "content": "We first leverage a VLM to generate 2D hypotheses of potential human interactions in the scene by automatically inpainting humans into multiple rendered views of " + }, + { + "bbox": [ + 304, + 662, + 545, + 698 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 662, + 545, + 698 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "content": "Multi-view Rendering. To capture scene context for 2D" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20467" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": "human inpainting, we render multiple views of the 3D scene " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " virtual cameras looking at " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": ". Cameras are randomly sampled on a hemisphere and filtered according to the visibility of " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": "; we refer to the supplementary material for additional camera setup details. We denote the rendered scene images as " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{I}^i\\}_{i\\in [1,k]}" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": ". To simplify notation, we omit the superscript " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " indexing each view in this section." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 163, + 287, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 163, + 287, + 247 + ], + "spans": [ + { + "bbox": [ + 46, + 163, + 287, + 247 + ], + "type": "text", + "content": "Inpainting with Dynamic Masking. Given a rendered scene image " + }, + { + "bbox": [ + 46, + 163, + 287, + 247 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 46, + 163, + 287, + 247 + ], + "type": "text", + "content": " and the text prompt " + }, + { + "bbox": [ + 46, + 163, + 287, + 247 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 46, + 163, + 287, + 247 + ], + "type": "text", + "content": ", we leverage a state-of-the-art latent 2D diffusion model [34] to generate a new image " + }, + { + "bbox": [ + 46, + 163, + 287, + 247 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{I}}" + }, + { + "bbox": [ + 46, + 163, + 287, + 247 + ], + "type": "text", + "content": ", where a human is inpainted into the scene image " + }, + { + "bbox": [ + 46, + 163, + 287, + 247 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 46, + 163, + 287, + 247 + ], + "type": "text", + "content": " while adhering to the specified interaction and 2D scene context. In practice, we opt for the popular Stable Diffusion Inpainting model [1] as the latent diffusion implementation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 251, + 288, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 288, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 288, + 407 + ], + "type": "text", + "content": "The latent diffusion model, denoted as " + }, + { + "bbox": [ + 46, + 251, + 288, + 407 + ], + "type": "inline_equation", + "content": "\\varOmega(\\mathbf{z}_t,\\mathbf{M},\\mathbf{I},\\Gamma,t)" + }, + { + "bbox": [ + 46, + 251, + 288, + 407 + ], + "type": "text", + "content": ", performs image inpainting by progressively denoising a noisy latent " + }, + { + "bbox": [ + 46, + 251, + 288, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 46, + 251, + 288, + 407 + ], + "type": "text", + "content": " at each time step " + }, + { + "bbox": [ + 46, + 251, + 288, + 407 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 251, + 288, + 407 + ], + "type": "text", + "content": ". During the denoising diffusion process, the binary mask " + }, + { + "bbox": [ + 46, + 251, + 288, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 46, + 251, + 288, + 407 + ], + "type": "text", + "content": " defines the inpainting region in the image " + }, + { + "bbox": [ + 46, + 251, + 288, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 46, + 251, + 288, + 407 + ], + "type": "text", + "content": "; however, specifying this mask typically requires manual effort [1]. Thus, we develop a fully-automated inpainting process by automatically generating the mask. Note that using a random or fixed human mask naively can lead to incorrect inpaintings, as this often results in scene context incorrectly masked, e.g., the objects affording the desired action may be entirely masked out, leading to the generated 2D human poses to be incoherent with the scene, producing undesirable 3D HSI synthesis." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 411, + 287, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 287, + 554 + ], + "type": "text", + "content": "We propose a masking scheme that dynamically adapts the mask through the denoising process by leveraging the internal cross-attention maps [4, 13, 42] from " + }, + { + "bbox": [ + 46, + 411, + 287, + 554 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 46, + 411, + 287, + 554 + ], + "type": "text", + "content": " to propose masks. The cross-attention maps capture rich semantic correlations between image pixels and input text tokens, playing a crucial role in guiding image generation. Let " + }, + { + "bbox": [ + 46, + 411, + 287, + 554 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\in \\mathbb{R}^{hw \\times n}" + }, + { + "bbox": [ + 46, + 411, + 287, + 554 + ], + "type": "text", + "content": " denote the cross-attention map between an image feature map of " + }, + { + "bbox": [ + 46, + 411, + 287, + 554 + ], + "type": "inline_equation", + "content": "hw" + }, + { + "bbox": [ + 46, + 411, + 287, + 554 + ], + "type": "text", + "content": " pixels and the text prompt " + }, + { + "bbox": [ + 46, + 411, + 287, + 554 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 46, + 411, + 287, + 554 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 411, + 287, + 554 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 411, + 287, + 554 + ], + "type": "text", + "content": " tokens, normalized row-wise using softmax. Here, " + }, + { + "bbox": [ + 46, + 411, + 287, + 554 + ], + "type": "inline_equation", + "content": "\\mathbf{A}[i,j]" + }, + { + "bbox": [ + 46, + 411, + 287, + 554 + ], + "type": "text", + "content": " signifies the influence of the " + }, + { + "bbox": [ + 46, + 411, + 287, + 554 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 411, + 287, + 554 + ], + "type": "text", + "content": "-th token on the " + }, + { + "bbox": [ + 46, + 411, + 287, + 554 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 411, + 287, + 554 + ], + "type": "text", + "content": "-th pixel, and " + }, + { + "bbox": [ + 46, + 411, + 287, + 554 + ], + "type": "inline_equation", + "content": "\\mathbf{A}[:,j]" + }, + { + "bbox": [ + 46, + 411, + 287, + 554 + ], + "type": "text", + "content": " forms a heat map of image regions to be filled with content related to the " + }, + { + "bbox": [ + 46, + 411, + 287, + 554 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 411, + 287, + 554 + ], + "type": "text", + "content": "-th token." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": "Using the cross-attention map " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_t" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": " at each time step " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": " of the diffusion process, we can dynamically derive a mask related to human tokens. Specifically, at time step " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": ", we obtain " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{t - 1},\\mathbf{A}_t\\gets \\varOmega (\\mathbf{z}_t,\\mathbf{M}_t,\\mathbf{I},\\varGamma ,t)" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": " after denoising. To create a human inpainting mask " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{t - 1}" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": " for time step " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": ", we extract heat maps from " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_t" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": " corresponding to the tokens referring to the human (e.g., \"woman\" or \"man\"), followed by summation of these extracted heat maps across the tokens and binarization. At the initial time step " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": ", we initialize " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_T" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": " as an empty mask. Our dynamic masking approach enables the synthesis of a 2D human-scene composition image " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{I}}" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": " without the need for manually-specified human inpainting masks. We illustrate our dynamic masking scheme in Fig. 3." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 310, + 71, + 545, + 194 + ], + "blocks": [ + { + "bbox": [ + 310, + 71, + 545, + 194 + ], + "lines": [ + { + "bbox": [ + 310, + 71, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 310, + 71, + 545, + 194 + ], + "type": "image", + "image_path": "6dfdd9debcfca513b105d50197d499e7cbceac967235f155bbf62904d0ca044b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 201, + 547, + 269 + ], + "lines": [ + { + "bbox": [ + 304, + 201, + 547, + 269 + ], + "spans": [ + { + "bbox": [ + 304, + 201, + 547, + 269 + ], + "type": "text", + "content": "Figure 3. Human inpainting with dynamic masking. Top: Given a scene image and a text prompt, a human is inpainted into the image without a mask specifying the inpainting region for latent diffusion. Bottom: The masks generated by our dynamic masking scheme based on cross-attention maps at different diffusion time steps adaptively shift to find the region of interest." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 281, + 488, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 281, + 488, + 293 + ], + "spans": [ + { + "bbox": [ + 306, + 281, + 488, + 293 + ], + "type": "text", + "content": "3.3. Robust Lifting to a 3D Interaction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 298, + 545, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 298, + 545, + 394 + ], + "spans": [ + { + "bbox": [ + 304, + 298, + 545, + 394 + ], + "type": "text", + "content": "Given the multi-view scene images inpainted with human subjects " + }, + { + "bbox": [ + 304, + 298, + 545, + 394 + ], + "type": "inline_equation", + "content": "\\{\\bar{\\mathbf{I}}^i\\}_{i\\in [1,k]}" + }, + { + "bbox": [ + 304, + 298, + 545, + 394 + ], + "type": "text", + "content": ", our aim is to optimize for the pose and shape parameters " + }, + { + "bbox": [ + 304, + 298, + 545, + 394 + ], + "type": "inline_equation", + "content": "(\\mathbf{R},\\mathbf{t},\\boldsymbol {\\Theta},\\boldsymbol {\\Phi})" + }, + { + "bbox": [ + 304, + 298, + 545, + 394 + ], + "type": "text", + "content": " of a 3D human " + }, + { + "bbox": [ + 304, + 298, + 545, + 394 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 304, + 298, + 545, + 394 + ], + "type": "text", + "content": ", guided by the multi-view interaction cues. We optimize for " + }, + { + "bbox": [ + 304, + 298, + 545, + 394 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 304, + 298, + 545, + 394 + ], + "type": "text", + "content": " by matching it with the 2D poses extracted from " + }, + { + "bbox": [ + 304, + 298, + 545, + 394 + ], + "type": "inline_equation", + "content": "\\{\\bar{\\mathbf{I}}^i\\}" + }, + { + "bbox": [ + 304, + 298, + 545, + 394 + ], + "type": "text", + "content": ". Since the 2D hypotheses may not be consistent across views, we formulate this as a robust optimization, simultaneously optimizing for the set of views most consistent with " + }, + { + "bbox": [ + 304, + 298, + 545, + 394 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 304, + 298, + 545, + 394 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 397, + 545, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 397, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 304, + 397, + 545, + 468 + ], + "type": "text", + "content": "2D Pose Estimation To distill interaction guidance for the 3D HSI synthesis, we compute a 2D pose representation for the inpainted human in each image " + }, + { + "bbox": [ + 304, + 397, + 545, + 468 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{I}}^i" + }, + { + "bbox": [ + 304, + 397, + 545, + 468 + ], + "type": "text", + "content": ". We use AlphaPose [6], an off-the-shelf pose estimation approach that infers a set of 2D joint positions " + }, + { + "bbox": [ + 304, + 397, + 545, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{J}^i" + }, + { + "bbox": [ + 304, + 397, + 545, + 468 + ], + "type": "text", + "content": " and the corresponding joint confidence scores " + }, + { + "bbox": [ + 304, + 397, + 545, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{c}^i" + }, + { + "bbox": [ + 304, + 397, + 545, + 468 + ], + "type": "text", + "content": " for the human subject in image " + }, + { + "bbox": [ + 304, + 397, + 545, + 468 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{I}}^i" + }, + { + "bbox": [ + 304, + 397, + 545, + 468 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 468, + 545, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 468, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 468, + 545, + 517 + ], + "type": "text", + "content": "These 2D pose hypotheses " + }, + { + "bbox": [ + 304, + 468, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{J}^i,\\mathbf{c}^i\\}_{i\\in [1,k]}" + }, + { + "bbox": [ + 304, + 468, + 545, + 517 + ], + "type": "text", + "content": " are then used to steer the interaction synthesis between the 3D human " + }, + { + "bbox": [ + 304, + 468, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 304, + 468, + 545, + 517 + ], + "type": "text", + "content": " and the scene " + }, + { + "bbox": [ + 304, + 468, + 545, + 517 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 468, + 545, + 517 + ], + "type": "text", + "content": ". We aim to minimize the following objective function " + }, + { + "bbox": [ + 304, + 468, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{total}}" + }, + { + "bbox": [ + 304, + 468, + 545, + 517 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 349, + 521, + 545, + 549 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 521, + 545, + 549 + ], + "spans": [ + { + "bbox": [ + 349, + 521, + 545, + 549 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {E} _ {\\text {t o t a l}} = \\lambda_ {\\mathrm {P F}} \\mathcal {E} _ {\\mathrm {P F}} + \\lambda_ {\\mathrm {V S}} \\mathcal {E} _ {\\mathrm {V S}} + \\lambda_ {\\mathrm {B P}} \\mathcal {E} _ {\\mathrm {B P}} + \\tag {1} \\\\ \\lambda_ {\\mathrm {B S}} \\mathcal {E} _ {\\mathrm {B S}} + \\lambda_ {\\mathrm {S C}} \\mathcal {E} _ {\\mathrm {S C}} + \\lambda_ {\\mathrm {S P}} \\mathcal {E} _ {\\mathrm {S P}}, \\\\ \\end{array}", + "image_path": "510b622dbc7b901b32abf9236b4a56372d219de0437a883bf49b16d276888c1d.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 555, + 545, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 545, + 591 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 304, + 555, + 545, + 591 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 555, + 545, + 591 + ], + "type": "text", + "content": " denote scalar weights for the energy terms: pose fitting " + }, + { + "bbox": [ + 304, + 555, + 545, + 591 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{PF}}" + }, + { + "bbox": [ + 304, + 555, + 545, + 591 + ], + "type": "text", + "content": ", view selection " + }, + { + "bbox": [ + 304, + 555, + 545, + 591 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{VS}}" + }, + { + "bbox": [ + 304, + 555, + 545, + 591 + ], + "type": "text", + "content": ", body pose " + }, + { + "bbox": [ + 304, + 555, + 545, + 591 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{BP}}" + }, + { + "bbox": [ + 304, + 555, + 545, + 591 + ], + "type": "text", + "content": ", body shape " + }, + { + "bbox": [ + 304, + 555, + 545, + 591 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{BS}}" + }, + { + "bbox": [ + 304, + 555, + 545, + 591 + ], + "type": "text", + "content": ", scene penetration " + }, + { + "bbox": [ + 304, + 555, + 545, + 591 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{SC}}" + }, + { + "bbox": [ + 304, + 555, + 545, + 591 + ], + "type": "text", + "content": ", and self-penetration " + }, + { + "bbox": [ + 304, + 555, + 545, + 591 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{SP}}" + }, + { + "bbox": [ + 304, + 555, + 545, + 591 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "text", + "content": "Robust View-consistent 3D Pose Fitting. Our primary energy term " + }, + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{PF}}" + }, + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "text", + "content": " minimizes the discrepancy between the projections of the 3D pose of " + }, + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "text", + "content": " and the inpainted poses from multiple views. However, 2D pose hypotheses can often be inconsistent across different views, due to the stochastic nature of diffusion models, leading to conflicting optimization signals for " + }, + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "text", + "content": ". To address this, we employ a robust optimization strategy that optimizes additionally for view selection weights in " + }, + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{PF}}" + }, + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "text", + "content": " that promote consistency between the 3D pose and the most consistent 2D pose hypotheses." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20468" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 156 + ], + "type": "text", + "content": "We thus introduce a new set of estimizable variables " + }, + { + "bbox": [ + 47, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{w} = \\{w^{i}|w^{i}\\in [0,1]\\}_{i\\in [1,k]}" + }, + { + "bbox": [ + 47, + 72, + 289, + 156 + ], + "type": "text", + "content": ", representing view consistency scores, and apply a robust kernel " + }, + { + "bbox": [ + 47, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 47, + 72, + 289, + 156 + ], + "type": "text", + "content": " to the per-view joint fitting constraints. The weights " + }, + { + "bbox": [ + 47, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 47, + 72, + 289, + 156 + ], + "type": "text", + "content": " allow the solver to adaptively focus on views with consistent inpainted poses and downweight the inconsistent ones. Our pose fitting energy " + }, + { + "bbox": [ + 47, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{PF}}" + }, + { + "bbox": [ + 47, + 72, + 289, + 156 + ], + "type": "text", + "content": " with view consistency is formulated as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 90, + 165, + 287, + 196 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 165, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 90, + 165, + 287, + 196 + ], + "type": "interline_equation", + "content": "\\mathcal {E} _ {\\mathrm {P F}} = \\frac {\\sum_ {i} w ^ {i} \\sum_ {j} \\mathbf {c} _ {j} ^ {i} \\rho (\\Pi (\\hat {\\mathbf {J}}) _ {j} ^ {i} - \\mathbf {J} _ {j} ^ {i})}{\\sum_ {i} w ^ {i}}, \\tag {2}", + "image_path": "3ba686b56f3564e89f2a2bd9c335cee0e3482b235e479d7f9c4675109348cc5a.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 205, + 287, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 205, + 287, + 266 + ], + "spans": [ + { + "bbox": [ + 47, + 205, + 287, + 266 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 205, + 287, + 266 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{J}}" + }, + { + "bbox": [ + 47, + 205, + 287, + 266 + ], + "type": "text", + "content": " denotes the 3D joint positions of the SMPL-X model [29] differentiably computed from the pose and shape parameters " + }, + { + "bbox": [ + 47, + 205, + 287, + 266 + ], + "type": "inline_equation", + "content": "(\\mathbf{R},\\mathbf{t},\\Theta ,\\Phi)" + }, + { + "bbox": [ + 47, + 205, + 287, + 266 + ], + "type": "text", + "content": ". The function " + }, + { + "bbox": [ + 47, + 205, + 287, + 266 + ], + "type": "inline_equation", + "content": "\\varPi(\\cdot)_j^i" + }, + { + "bbox": [ + 47, + 205, + 287, + 266 + ], + "type": "text", + "content": " represents projection of the " + }, + { + "bbox": [ + 47, + 205, + 287, + 266 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 205, + 287, + 266 + ], + "type": "text", + "content": "-th joint in the " + }, + { + "bbox": [ + 47, + 205, + 287, + 266 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 205, + 287, + 266 + ], + "type": "text", + "content": "-th camera view, and " + }, + { + "bbox": [ + 47, + 205, + 287, + 266 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 47, + 205, + 287, + 266 + ], + "type": "text", + "content": " is the robust Geman-McClure error function [8]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 269, + 287, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 269, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 47, + 269, + 287, + 293 + ], + "type": "text", + "content": "Regularizations. Below, we describe the regularization terms adopted in our optimization." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 296, + 288, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 296, + 288, + 319 + ], + "spans": [ + { + "bbox": [ + 47, + 296, + 288, + 319 + ], + "type": "text", + "content": "1) We encourage the per-view weights " + }, + { + "bbox": [ + 47, + 296, + 288, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 47, + 296, + 288, + 319 + ], + "type": "text", + "content": " to focus on at least " + }, + { + "bbox": [ + 47, + 296, + 288, + 319 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 47, + 296, + 288, + 319 + ], + "type": "text", + "content": " views:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 109, + 327, + 287, + 352 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 327, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 109, + 327, + 287, + 352 + ], + "type": "interline_equation", + "content": "\\mathcal {E} _ {\\mathrm {V S}} = \\max (\\tau - \\sum_ {i} w ^ {i}, 0). \\tag {3}", + "image_path": "581b52d6f273ae18c15d3e753f2459818aa3b9cc93823809348dbc2c5995cf22.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 363, + 287, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 363, + 287, + 388 + ], + "spans": [ + { + "bbox": [ + 47, + 363, + 287, + 388 + ], + "type": "text", + "content": "2) To ensure a natural 3D pose, we impose the following energy term on the body pose parameter:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 117, + 396, + 287, + 411 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 396, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 117, + 396, + 287, + 411 + ], + "type": "interline_equation", + "content": "\\mathcal {E} _ {\\mathrm {B P}} = \\left\\| \\boldsymbol {\\Theta} \\right\\| ^ {2} + \\mathcal {E} _ {\\mathrm {J A}} (\\hat {\\boldsymbol {\\Theta}}). \\tag {4}", + "image_path": "63bfc503501dcd5a4f00e7e871df3aeb81c3a6849fbca8fa689509858667bf75.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 420, + 287, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 420, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 47, + 420, + 287, + 480 + ], + "type": "text", + "content": "The first term is the VPoser body prior [29] regularizing the latent pose " + }, + { + "bbox": [ + 47, + 420, + 287, + 480 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 47, + 420, + 287, + 480 + ], + "type": "text", + "content": ", and the second term " + }, + { + "bbox": [ + 47, + 420, + 287, + 480 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{JA}}" + }, + { + "bbox": [ + 47, + 420, + 287, + 480 + ], + "type": "text", + "content": " is a simple angle prior on the body joint rotations " + }, + { + "bbox": [ + 47, + 420, + 287, + 480 + ], + "type": "inline_equation", + "content": "\\hat{\\Theta}" + }, + { + "bbox": [ + 47, + 420, + 287, + 480 + ], + "type": "text", + "content": ", decoded from the latent pose " + }, + { + "bbox": [ + 47, + 420, + 287, + 480 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 47, + 420, + 287, + 480 + ], + "type": "text", + "content": " by VPoser, to penalize extreme bending. The formulation of " + }, + { + "bbox": [ + 47, + 420, + 287, + 480 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{JA}}" + }, + { + "bbox": [ + 47, + 420, + 287, + 480 + ], + "type": "text", + "content": " is provided in the supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 483, + 287, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 483, + 287, + 507 + ], + "spans": [ + { + "bbox": [ + 47, + 483, + 287, + 507 + ], + "type": "text", + "content": "3) We regularize the shape parameter " + }, + { + "bbox": [ + 47, + 483, + 287, + 507 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 47, + 483, + 287, + 507 + ], + "type": "text", + "content": " to obtain a plausible body shape via" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 516, + 287, + 531 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 516, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 138, + 516, + 287, + 531 + ], + "type": "interline_equation", + "content": "\\mathcal {E} _ {\\mathrm {B S}} = \\left\\| \\boldsymbol {\\Phi} \\right\\| ^ {2}, \\tag {5}", + "image_path": "6cb6750e470efe9f46862e436d0c2750593d6ed61b3d021e85a514e0e103e625.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 540, + 287, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 540, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 540, + 287, + 563 + ], + "type": "text", + "content": "which measures the Mahalanobis distance between " + }, + { + "bbox": [ + 47, + 540, + 287, + 563 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 47, + 540, + 287, + 563 + ], + "type": "text", + "content": " and the body shape distribution used in SMPL-X [29]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 566, + 287, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 566, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 47, + 566, + 287, + 602 + ], + "type": "text", + "content": "4) To ensure physical contact but also avoid penetration between the 3D human " + }, + { + "bbox": [ + 47, + 566, + 287, + 602 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 47, + 566, + 287, + 602 + ], + "type": "text", + "content": " and the scene " + }, + { + "bbox": [ + 47, + 566, + 287, + 602 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 47, + 566, + 287, + 602 + ], + "type": "text", + "content": ", we formulate the spatial constraints as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 612, + 287, + 653 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 287, + 653 + ], + "type": "interline_equation", + "content": "\\mathcal {E} _ {\\mathrm {S C}} = \\left\\{ \\begin{array}{l l} \\min _ {\\mathbf {v} \\in \\mathbf {V}} \\Psi (\\mathbf {v}), & \\Psi (\\mathbf {v}) > 0 \\forall \\mathbf {v} \\in \\mathbf {V} \\\\ \\sum_ {\\mathbf {v} \\in \\mathbf {V}} | \\min (\\Psi (\\mathbf {v}), 0) |, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {6}", + "image_path": "49a6caa283afff4ad0be78a9027af5de005dfdc564f55d66f8432b2ae7eda72f.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\varPsi" + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "text", + "content": " is a pre-computed signed distance field (SDF) [10] of the scene " + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "text", + "content": ". When " + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\varPsi(\\mathbf{v})" + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "text", + "content": " has a negative sign, it indicates that the body vertex " + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "text", + "content": " is located inside the nearest scene object (i.e., penetration). Conversely, a positive sign means that " + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "text", + "content": " is positioned on the outside." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "content": "5) Finally, to resolve penetration within the human body " + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "content": " itself, we include a self-penetration energy " + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{SP}}" + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "content": " based on detecting colliding body triangles using Bounding Volume Hierarchies (BVH) [40]. We refer the reader to [2, 10, 41] for more formulation details on " + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{SP}}" + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 144, + 425, + 155 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 144, + 425, + 155 + ], + "spans": [ + { + "bbox": [ + 306, + 144, + 425, + 155 + ], + "type": "text", + "content": "3.4. Iterative Refinement" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 163, + 547, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 163, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 304, + 163, + 547, + 319 + ], + "type": "text", + "content": "A posed 3D human " + }, + { + "bbox": [ + 304, + 163, + 547, + 319 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 304, + 163, + 547, + 319 + ], + "type": "text", + "content": " interacting with the 3D scene " + }, + { + "bbox": [ + 304, + 163, + 547, + 319 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 163, + 547, + 319 + ], + "type": "text", + "content": " is generated after applying the VLM-based 2D inpainting (Sec. 3.2) and the robust 3D lifting (Sec. 3.3). To improve the synthesis and consistency of the interaction results, we employ an iterative refinement scheme over the aforementioned inpainting and 3D lifting. During refinement, we render the silhouette of the posed 3D human " + }, + { + "bbox": [ + 304, + 163, + 547, + 319 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 304, + 163, + 547, + 319 + ], + "type": "text", + "content": " in each camera view, and use it as a more precise and consistent mask " + }, + { + "bbox": [ + 304, + 163, + 547, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 304, + 163, + 547, + 319 + ], + "type": "text", + "content": " in the latent diffusion inpainting " + }, + { + "bbox": [ + 304, + 163, + 547, + 319 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 304, + 163, + 547, + 319 + ], + "type": "text", + "content": " [34], replacing dynamic masking. By doing so, the consistency among the 2D pose hypotheses inpainted into the multi-view scene images gradually improves, thus leading to an improved 3D HSI synthesis outcome." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 330, + 440, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 330, + 440, + 342 + ], + "spans": [ + { + "bbox": [ + 306, + 330, + 440, + 342 + ], + "type": "text", + "content": "3.5. Implementation Details" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 350, + 547, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 350, + 547, + 530 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 547, + 530 + ], + "type": "text", + "content": "We implemented GenZI with PyTorch [28]. For VLM-based 2D inpainting, we use " + }, + { + "bbox": [ + 304, + 350, + 547, + 530 + ], + "type": "inline_equation", + "content": "k = 16" + }, + { + "bbox": [ + 304, + 350, + 547, + 530 + ], + "type": "text", + "content": " cameras for multiview rendering. We use 50 denoising steps in Stable Diffusion Inpainting with a state-of-the-art diffusion sampler [25]. For dynamic masking during inpainting, we aggregate cross-attention maps with a resolution of " + }, + { + "bbox": [ + 304, + 350, + 547, + 530 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 304, + 350, + 547, + 530 + ], + "type": "text", + "content": ". Input text prompts " + }, + { + "bbox": [ + 304, + 350, + 547, + 530 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 304, + 350, + 547, + 530 + ], + "type": "text", + "content": " (e.g., \"sitting on a bench\") are all appended with fixed prefix \"a woman\" and fixed suffix \"wearing a white shirt and blue pants, full body\" to better constrain generation. For robust 3D lifting, we set " + }, + { + "bbox": [ + 304, + 350, + 547, + 530 + ], + "type": "inline_equation", + "content": "\\tau = 3" + }, + { + "bbox": [ + 304, + 350, + 547, + 530 + ], + "type": "text", + "content": ", and optimize the energy " + }, + { + "bbox": [ + 304, + 350, + 547, + 530 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{total}}" + }, + { + "bbox": [ + 304, + 350, + 547, + 530 + ], + "type": "text", + "content": " using gradient descent [24] for 1.6K steps, which takes " + }, + { + "bbox": [ + 304, + 350, + 547, + 530 + ], + "type": "inline_equation", + "content": "\\sim 3" + }, + { + "bbox": [ + 304, + 350, + 547, + 530 + ], + "type": "text", + "content": " minutes on an NVIDIA A100 GPU. In the iterative refinement, we dilate the rendered human silhouette with a kernel size of " + }, + { + "bbox": [ + 304, + 350, + 547, + 530 + ], + "type": "inline_equation", + "content": "11 \\times 11" + }, + { + "bbox": [ + 304, + 350, + 547, + 530 + ], + "type": "text", + "content": " for mask generation, and the refinement is performed once." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 544, + 388, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 544, + 388, + 558 + ], + "spans": [ + { + "bbox": [ + 306, + 544, + 388, + 558 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 566, + 547, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 566, + 547, + 625 + ], + "spans": [ + { + "bbox": [ + 304, + 566, + 547, + 625 + ], + "type": "text", + "content": "We demonstrate the effectiveness and generality of our approach GenZI on a diverse collection of 3D scene models from Sketchfab.com. We conduct both quantitative and qualitative evaluations to compare GenZI with alternative baselines approaches [10, 51] to our new task." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "content": "Dataset. In our Sketchfab dataset, we gathered 8 large-scale 3D scenes encompassing a variety of indoor and outdoor environments with diverse geometric structures, including a realistic Venice city, a gym, and a cartoon-style food truck. We collected 4-5 text prompts per scene describing human interactions with the scene for specified approximate point locations, resulting in 38 actions for evaluation." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20469" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "Baselines. To the best of our knowledge, there are no baselines that estimate 3D human-scene interactions based on natural language text input in a zero-shot fashion. We thus perform comparisons with related baseline approaches:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 121, + 288, + 420 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 47, + 121, + 288, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 121, + 288, + 289 + ], + "spans": [ + { + "bbox": [ + 47, + 121, + 288, + 289 + ], + "type": "text", + "content": "- COINS [51] is a state-of-the-art approach estimating 3D humans in indoor 3D scans with a fixed vocabulary of actions and objects, given object segmentations. It takes as input (action, object) pairs for semantic control, and requires full supervision in its CVAE training, using captured 3D interaction data with both instance segmentations of 3D objects and action labelling. Due to being trained on a small, closed set of indoor interactions, we adapt COINS to a subset of the most similar Sketchfab actions by manually segmenting corresponding 3D objects from the scenes similar to its indoor training data and mapping the text prompts to its (action, object) input. Note that our approach GenZI does not require any 3D scene segmentations." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 289, + 287, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 289, + 287, + 372 + ], + "spans": [ + { + "bbox": [ + 47, + 289, + 287, + 372 + ], + "type": "text", + "content": "- Hassan et al. [10] perform 3D human estimation from a single RGB image. To adapt this to Sketchfab, we reuse the multi-view scene images " + }, + { + "bbox": [ + 47, + 289, + 287, + 372 + ], + "type": "inline_equation", + "content": "\\{\\bar{\\mathbf{T}}^i\\}_{i\\in [1,k]}" + }, + { + "bbox": [ + 47, + 289, + 287, + 372 + ], + "type": "text", + "content": " inpainted with humans from our dynamic masking scheme (Sec. 3.2), where the view with the best image-text cosine similarity (i.e., CLIP score [33]) is used as the input for 3D human estimation under the known virtual camera parameters." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 372, + 287, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 372, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 47, + 372, + 287, + 420 + ], + "type": "text", + "content": "- Ours-Single View: Finally, we consider a baseline leveraging 3D body estimation from our method, but limited to only one inpainted view. The same best view image, as described above, is also as input for this baseline." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 424, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 424, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 424, + 287, + 460 + ], + "type": "text", + "content": "Evaluation Metrics. To measure 3D HSI quality, we conduct perceptual studies and compute metrics including semantic consistency, diversity, and physical plausibility." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 460, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 581 + ], + "type": "text", + "content": "We first carry out two perceptual studies to assess the realism and semantic accuracy of the synthesized interactions. The first is a binary-choice study, where interaction samples generated by two different methods based on the same text prompt are shown, and the participants are asked to choose the sample that is more realistic and better matched the text. The second study is a unary test, where for each interaction sample, the participants are asked to rate the realism and consistency between the shown sample and text prompt from 1 (strongly disagree) to 5 (strongly agree)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 581, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 581, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 581, + 287, + 640 + ], + "type": "text", + "content": "To evaluate the semantic consistency between a synthesized 3D interaction and the input text prompt, we calculate the CLIP score [33], where the 3D interaction is re-rendered into " + }, + { + "bbox": [ + 46, + 581, + 287, + 640 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 581, + 287, + 640 + ], + "type": "text", + "content": " view images, and the image-text cosine similarities from CLIP ViT-B/32 are averaged across all views." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": "Additionally, we include quantitative metrics from existing works to evaluate diversity and physical plausibility. We observe that these metrics often do not reflect perceptual quality, as generated bodies can be diverse but have interpenetrations without demonstrating any physical or semantic coherence. Nevertheless, to measure the synthesis" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 318, + 85, + 533, + 159 + ], + "blocks": [ + { + "bbox": [ + 372, + 72, + 479, + 84 + ], + "lines": [ + { + "bbox": [ + 372, + 72, + 479, + 84 + ], + "spans": [ + { + "bbox": [ + 372, + 72, + 479, + 84 + ], + "type": "text", + "content": "(a) Binary Perceptual Study" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 318, + 85, + 533, + 159 + ], + "lines": [ + { + "bbox": [ + 318, + 85, + 533, + 159 + ], + "spans": [ + { + "bbox": [ + 318, + 85, + 533, + 159 + ], + "type": "image", + "image_path": "4a2e2d5de573cabf7141f16eee60de174c3e181152b5d9faf737bdeec3b57731.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 320, + 169, + 526, + 272 + ], + "blocks": [ + { + "bbox": [ + 320, + 169, + 526, + 272 + ], + "lines": [ + { + "bbox": [ + 320, + 169, + 526, + 272 + ], + "spans": [ + { + "bbox": [ + 320, + 169, + 526, + 272 + ], + "type": "image", + "image_path": "253e1bf636820898e186febf3a12fdddc4e6a75d2b7a668080b2eeffcd357e9a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 274, + 545, + 318 + ], + "lines": [ + { + "bbox": [ + 305, + 274, + 545, + 318 + ], + "spans": [ + { + "bbox": [ + 305, + 274, + 545, + 318 + ], + "type": "text", + "content": "Figure 4. User study of 3D human-scene interaction synthesis on the Sketchfab dataset, where participants show a strong preference for the generations by our approach, in comparison with all baselines, COINS [51], Hassan et al. [10], and Ours-Single View." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 334, + 545, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 334, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 334, + 545, + 502 + ], + "type": "text", + "content": "diversity, we follow [50, 51] and cluster the SMPL-X parameters of the generated humans into 20 clusters with K-means and compute the entropy of the cluster ID histogram of all the samples. We also calculate the cluster size as the mean distance to cluster centers. For the physical plausibility, we evaluate the collision and contact between body meshes and scene meshes, following [50, 51]. The noncollision score is computed as the ratio between the number of body mesh vertices with positive SDF values (Sec. 3.3) and the number of all body mesh vertices. The contact score is defined as the ratio between the number of body meshes with scene contact and the number of all generated body meshes. A body mesh is in contact with the scene if any body mesh vertex has a non-positive SDF value." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 511, + 511, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 511, + 511, + 525 + ], + "spans": [ + { + "bbox": [ + 306, + 511, + 511, + 525 + ], + "type": "text", + "content": "4.1. Comparison to Alternative Approaches" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 533, + 545, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 664 + ], + "type": "text", + "content": "Quantitative Evaluation. Fig. 4 shows the results of the perceptual studies collected from 30 participants across a binary study and a unary study. In the binary study, we observe that participants overwhelmingly favor the generations by our GenZI compared to all baselines – more than " + }, + { + "bbox": [ + 304, + 533, + 545, + 664 + ], + "type": "inline_equation", + "content": "87\\%" + }, + { + "bbox": [ + 304, + 533, + 545, + 664 + ], + "type": "text", + "content": " of the time. In the unary test, the average realism rating for our interaction generations is 3.6, the highest compared to the baselines, which are all below 2.0. These perceptual results strongly indicate that our GenZI can synthesize realistic 3D humans interacting with various 3D scenes without requiring any captured 3D interaction data." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Tab. 1 presents the quantitative evaluation of semantic consistency, diversity, and physical plausibility on the full Sketchfab dataset. Our approach achieves the best semantic consistency score, echoing the strong user preference for" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20470" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 86, + 114, + 145, + 166 + ], + "blocks": [ + { + "bbox": [ + 66, + 87, + 165, + 114 + ], + "lines": [ + { + "bbox": [ + 66, + 87, + 165, + 114 + ], + "spans": [ + { + "bbox": [ + 66, + 87, + 165, + 114 + ], + "type": "text", + "content": "\"sitting and playing piano, hands on piano keys, feet flat on the ground\"" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 86, + 114, + 145, + 166 + ], + "lines": [ + { + "bbox": [ + 86, + 114, + 145, + 166 + ], + "spans": [ + { + "bbox": [ + 86, + 114, + 145, + 166 + ], + "type": "image", + "image_path": "3bcc7b441772f78ba6d3906bc8f2dcbd3064268668baa1e6b2852e13680f45b2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 85, + 190, + 146, + 249 + ], + "blocks": [ + { + "bbox": [ + 74, + 170, + 157, + 190 + ], + "lines": [ + { + "bbox": [ + 74, + 170, + 157, + 190 + ], + "spans": [ + { + "bbox": [ + 74, + 170, + 157, + 190 + ], + "type": "text", + "content": "\"climbing a ladder, hands gripping side rails\"" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 85, + 190, + 146, + 249 + ], + "lines": [ + { + "bbox": [ + 85, + 190, + 146, + 249 + ], + "spans": [ + { + "bbox": [ + 85, + 190, + 146, + 249 + ], + "type": "image", + "image_path": "0b553c4a2ca70c64b73a76a174629046e40d0d9a3da958dc2f692f91201696e8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 86, + 273, + 146, + 331 + ], + "blocks": [ + { + "bbox": [ + 64, + 254, + 167, + 272 + ], + "lines": [ + { + "bbox": [ + 64, + 254, + 167, + 272 + ], + "spans": [ + { + "bbox": [ + 64, + 254, + 167, + 272 + ], + "type": "text", + "content": "\"sitting and paddling a wooden boat, spreading arms\"" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 86, + 273, + 146, + 331 + ], + "lines": [ + { + "bbox": [ + 86, + 273, + 146, + 331 + ], + "spans": [ + { + "bbox": [ + 86, + 273, + 146, + 331 + ], + "type": "image", + "image_path": "95d548d2c559076cf004401f95022d748ffcc1c3237f524156b6ae33f3f01379.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 91, + 351, + 142, + 412 + ], + "blocks": [ + { + "bbox": [ + 80, + 340, + 151, + 350 + ], + "lines": [ + { + "bbox": [ + 80, + 340, + 151, + 350 + ], + "spans": [ + { + "bbox": [ + 80, + 340, + 151, + 350 + ], + "type": "text", + "content": "\"walking on a bridge\"" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 91, + 351, + 142, + 412 + ], + "lines": [ + { + "bbox": [ + 91, + 351, + 142, + 412 + ], + "spans": [ + { + "bbox": [ + 91, + 351, + 142, + 412 + ], + "type": "image", + "image_path": "62f8a529f4ba054258b7ee10a315fb638e2269d74295ebd845d966f3b64bd4c0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 85, + 442, + 146, + 487 + ], + "blocks": [ + { + "bbox": [ + 77, + 422, + 156, + 441 + ], + "lines": [ + { + "bbox": [ + 77, + 422, + 156, + 441 + ], + "spans": [ + { + "bbox": [ + 77, + 422, + 156, + 441 + ], + "type": "text", + "content": "\"picking up a trash bag, standing, bending over\"" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 85, + 442, + 146, + 487 + ], + "lines": [ + { + "bbox": [ + 85, + 442, + 146, + 487 + ], + "spans": [ + { + "bbox": [ + 85, + 442, + 146, + 487 + ], + "type": "image", + "image_path": "783e81a1c61832f4b15955a1997135e0fe0a3938cec6b9e83cbf06fc1757d1fb.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 85, + 517, + 146, + 578 + ], + "blocks": [ + { + "bbox": [ + 78, + 507, + 154, + 517 + ], + "lines": [ + { + "bbox": [ + 78, + 507, + 154, + 517 + ], + "spans": [ + { + "bbox": [ + 78, + 507, + 154, + 517 + ], + "type": "text", + "content": "\"walking out of a door\"" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 85, + 517, + 146, + 578 + ], + "lines": [ + { + "bbox": [ + 85, + 517, + 146, + 578 + ], + "spans": [ + { + "bbox": [ + 85, + 517, + 146, + 578 + ], + "type": "image", + "image_path": "9b3d53a5b246156290a7403c0b46b64ce7c438769e9a329c849828c7d86b618c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 86, + 620, + 146, + 653 + ], + "blocks": [ + { + "bbox": [ + 74, + 590, + 157, + 609 + ], + "lines": [ + { + "bbox": [ + 74, + 590, + 157, + 609 + ], + "spans": [ + { + "bbox": [ + 74, + 590, + 157, + 609 + ], + "type": "text", + "content": "\"riding a recumbent bike, sitting\"" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 86, + 620, + 146, + 653 + ], + "lines": [ + { + "bbox": [ + 86, + 620, + 146, + 653 + ], + "spans": [ + { + "bbox": [ + 86, + 620, + 146, + 653 + ], + "type": "image", + "image_path": "987e6194f7d98f4c41a880b096dca735b21210925e942c07a15fcbdd926233f0.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 170, + 83, + 252, + 164 + ], + "blocks": [ + { + "bbox": [ + 88, + 72, + 146, + 84 + ], + "lines": [ + { + "bbox": [ + 88, + 72, + 146, + 84 + ], + "spans": [ + { + "bbox": [ + 88, + 72, + 146, + 84 + ], + "type": "text", + "content": "Text Prompt" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 196, + 72, + 227, + 82 + ], + "lines": [ + { + "bbox": [ + 196, + 72, + 227, + 82 + ], + "spans": [ + { + "bbox": [ + 196, + 72, + 227, + 82 + ], + "type": "text", + "content": "COINS" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 170, + 83, + 252, + 164 + ], + "lines": [ + { + "bbox": [ + 170, + 83, + 252, + 164 + ], + "spans": [ + { + "bbox": [ + 170, + 83, + 252, + 164 + ], + "type": "image", + "image_path": "6bc881151455391110e66bfd6aea11299fc44fbd733ccf7c92aeee27f137d351.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 260, + 83, + 342, + 163 + ], + "blocks": [ + { + "bbox": [ + 272, + 72, + 332, + 82 + ], + "lines": [ + { + "bbox": [ + 272, + 72, + 332, + 82 + ], + "spans": [ + { + "bbox": [ + 272, + 72, + 332, + 82 + ], + "type": "text", + "content": "Hassan et al." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 260, + 83, + 342, + 163 + ], + "lines": [ + { + "bbox": [ + 260, + 83, + 342, + 163 + ], + "spans": [ + { + "bbox": [ + 260, + 83, + 342, + 163 + ], + "type": "image", + "image_path": "f01f189760f63e9e1e32e1807e107e25f04586c9ded521ce3e99d4ee76c7c9d1.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 350, + 83, + 431, + 163 + ], + "blocks": [ + { + "bbox": [ + 371, + 72, + 410, + 82 + ], + "lines": [ + { + "bbox": [ + 371, + 72, + 410, + 82 + ], + "spans": [ + { + "bbox": [ + 371, + 72, + 410, + 82 + ], + "type": "text", + "content": "Ours-SV" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 350, + 83, + 431, + 163 + ], + "lines": [ + { + "bbox": [ + 350, + 83, + 431, + 163 + ], + "spans": [ + { + "bbox": [ + 350, + 83, + 431, + 163 + ], + "type": "image", + "image_path": "e162de5ffe520c1abdb1642c60d428332d0640f4223dac9cd529b76290399663.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 440, + 83, + 521, + 163 + ], + "blocks": [ + { + "bbox": [ + 468, + 72, + 492, + 82 + ], + "lines": [ + { + "bbox": [ + 468, + 72, + 492, + 82 + ], + "spans": [ + { + "bbox": [ + 468, + 72, + 492, + 82 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 440, + 83, + 521, + 163 + ], + "lines": [ + { + "bbox": [ + 440, + 83, + 521, + 163 + ], + "spans": [ + { + "bbox": [ + 440, + 83, + 521, + 163 + ], + "type": "image", + "image_path": "04b88ee2ce0ea1478cea3f8a4861966dece09382e8b8480a3865f5025ca5e9b8.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 260, + 165, + 342, + 246 + ], + "blocks": [ + { + "bbox": [ + 260, + 165, + 342, + 246 + ], + "lines": [ + { + "bbox": [ + 260, + 165, + 342, + 246 + ], + "spans": [ + { + "bbox": [ + 260, + 165, + 342, + 246 + ], + "type": "image", + "image_path": "91883f8c53c2804f6d0f3c9b8bb7995148ee28774ac3cb62a70ca1f92c65f618.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 350, + 165, + 431, + 246 + ], + "blocks": [ + { + "bbox": [ + 350, + 165, + 431, + 246 + ], + "lines": [ + { + "bbox": [ + 350, + 165, + 431, + 246 + ], + "spans": [ + { + "bbox": [ + 350, + 165, + 431, + 246 + ], + "type": "image", + "image_path": "ad9c525edd94a4eaa8e95a422aa4c0ba4f25f1819dd8c4c00a80c46c06c63371.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 440, + 165, + 521, + 246 + ], + "blocks": [ + { + "bbox": [ + 440, + 165, + 521, + 246 + ], + "lines": [ + { + "bbox": [ + 440, + 165, + 521, + 246 + ], + "spans": [ + { + "bbox": [ + 440, + 165, + 521, + 246 + ], + "type": "image", + "image_path": "026ee0159bacf51b95ee2369cf4d7fe81f0fcf6790f735a950cecc523592230c.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 260, + 247, + 342, + 329 + ], + "blocks": [ + { + "bbox": [ + 260, + 247, + 342, + 329 + ], + "lines": [ + { + "bbox": [ + 260, + 247, + 342, + 329 + ], + "spans": [ + { + "bbox": [ + 260, + 247, + 342, + 329 + ], + "type": "image", + "image_path": "585ae1923e61e26e622c1d05e5dd8861825ea6793e0019bef976b875f15387df.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 350, + 247, + 431, + 329 + ], + "blocks": [ + { + "bbox": [ + 350, + 247, + 431, + 329 + ], + "lines": [ + { + "bbox": [ + 350, + 247, + 431, + 329 + ], + "spans": [ + { + "bbox": [ + 350, + 247, + 431, + 329 + ], + "type": "image", + "image_path": "347c82f5728a9ca6ac14e404b6d8d3493f37eef3c244e6af5e9bc08313beade7.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 440, + 247, + 521, + 329 + ], + "blocks": [ + { + "bbox": [ + 440, + 247, + 521, + 329 + ], + "lines": [ + { + "bbox": [ + 440, + 247, + 521, + 329 + ], + "spans": [ + { + "bbox": [ + 440, + 247, + 521, + 329 + ], + "type": "image", + "image_path": "0c9307135aada9be1e64f6dfe7f488493a14c300ae4bac7aa8cfca63c63467fb.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 170, + 331, + 252, + 413 + ], + "blocks": [ + { + "bbox": [ + 170, + 331, + 252, + 413 + ], + "lines": [ + { + "bbox": [ + 170, + 331, + 252, + 413 + ], + "spans": [ + { + "bbox": [ + 170, + 331, + 252, + 413 + ], + "type": "image", + "image_path": "799fe841691ee3c412d33497b7e620a8087bcdb53666489310b98497dc74e68a.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 260, + 331, + 342, + 412 + ], + "blocks": [ + { + "bbox": [ + 260, + 331, + 342, + 412 + ], + "lines": [ + { + "bbox": [ + 260, + 331, + 342, + 412 + ], + "spans": [ + { + "bbox": [ + 260, + 331, + 342, + 412 + ], + "type": "image", + "image_path": "16c317b5dbba0f25b8d1c3938f82bf5305b39b7af39f90d4fb648a4fe027ffac.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 350, + 332, + 431, + 412 + ], + "blocks": [ + { + "bbox": [ + 350, + 332, + 431, + 412 + ], + "lines": [ + { + "bbox": [ + 350, + 332, + 431, + 412 + ], + "spans": [ + { + "bbox": [ + 350, + 332, + 431, + 412 + ], + "type": "image", + "image_path": "1eb9f36e63879ba59c68a502439c6ddaf41925c600bc7d8d93fcc7879f7d98bf.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 440, + 332, + 521, + 412 + ], + "blocks": [ + { + "bbox": [ + 440, + 332, + 521, + 412 + ], + "lines": [ + { + "bbox": [ + 440, + 332, + 521, + 412 + ], + "spans": [ + { + "bbox": [ + 440, + 332, + 521, + 412 + ], + "type": "image", + "image_path": "8d23c4ee33ef316b5a6e5cd0e0658ba908b42ad2066dfbbcca2f4cd9cee0cd16.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 260, + 414, + 342, + 495 + ], + "blocks": [ + { + "bbox": [ + 260, + 414, + 342, + 495 + ], + "lines": [ + { + "bbox": [ + 260, + 414, + 342, + 495 + ], + "spans": [ + { + "bbox": [ + 260, + 414, + 342, + 495 + ], + "type": "image", + "image_path": "1f7eb8ea742aae73c555ff5f2fba8cb5a1f4df94529a6d4674cea1c82cea15e9.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 350, + 415, + 431, + 495 + ], + "blocks": [ + { + "bbox": [ + 350, + 415, + 431, + 495 + ], + "lines": [ + { + "bbox": [ + 350, + 415, + 431, + 495 + ], + "spans": [ + { + "bbox": [ + 350, + 415, + 431, + 495 + ], + "type": "image", + "image_path": "005844162b292bdf668d22ac84f462409c58a86c7f1e37b8dd6b26d36a17d9cc.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 440, + 415, + 521, + 495 + ], + "blocks": [ + { + "bbox": [ + 440, + 415, + 521, + 495 + ], + "lines": [ + { + "bbox": [ + 440, + 415, + 521, + 495 + ], + "spans": [ + { + "bbox": [ + 440, + 415, + 521, + 495 + ], + "type": "image", + "image_path": "49e321fb2240ec8dd772d47048522b37fa8e2c947fb811d5b255e81f8d6e42ac.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 260, + 497, + 342, + 578 + ], + "blocks": [ + { + "bbox": [ + 260, + 497, + 342, + 578 + ], + "lines": [ + { + "bbox": [ + 260, + 497, + 342, + 578 + ], + "spans": [ + { + "bbox": [ + 260, + 497, + 342, + 578 + ], + "type": "image", + "image_path": "9145aaf37bf9c5c3ac2505f2c4b5bfd4304c0238bf003297bd9facc2f7212263.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 350, + 497, + 431, + 578 + ], + "blocks": [ + { + "bbox": [ + 350, + 497, + 431, + 578 + ], + "lines": [ + { + "bbox": [ + 350, + 497, + 431, + 578 + ], + "spans": [ + { + "bbox": [ + 350, + 497, + 431, + 578 + ], + "type": "image", + "image_path": "cd2327af6bfd58ee031f311cddba687dbf504298e267d65faadc5b9d54370eb0.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 440, + 497, + 521, + 578 + ], + "blocks": [ + { + "bbox": [ + 440, + 497, + 521, + 578 + ], + "lines": [ + { + "bbox": [ + 440, + 497, + 521, + 578 + ], + "spans": [ + { + "bbox": [ + 440, + 497, + 521, + 578 + ], + "type": "image", + "image_path": "c68d305062d3c01a8d9e6ddc2b32f134bf26dbcb7b1fd2bf024d8b048b3a7967.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 170, + 580, + 252, + 662 + ], + "blocks": [ + { + "bbox": [ + 170, + 580, + 252, + 662 + ], + "lines": [ + { + "bbox": [ + 170, + 580, + 252, + 662 + ], + "spans": [ + { + "bbox": [ + 170, + 580, + 252, + 662 + ], + "type": "image", + "image_path": "195c72893bf9b901b89ddcf561ef1990dc1f4847071cb3651c225dfa30ab8004.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 666, + 545, + 709 + ], + "lines": [ + { + "bbox": [ + 46, + 666, + 545, + 709 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 545, + 709 + ], + "type": "text", + "content": "Figure 5. Qualitative results on the Sketchfab dataset. Our GenZI synthesizes more realistic 3D human-scene interactions and generalizes better across diverse scene types, compared to the baselines COINS [51], Hassan et al. [10], and Ours-Single View. For COINS, we show the used (action, object) labels from its closed set of indoor interactions; its closed setting can lead to degraded results from out-of-distribution object classes (e.g., curved bridge deck as the floor, chair at a different height or shape than those in the training set)." + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 260, + 580, + 342, + 662 + ], + "blocks": [ + { + "bbox": [ + 260, + 580, + 342, + 662 + ], + "lines": [ + { + "bbox": [ + 260, + 580, + 342, + 662 + ], + "spans": [ + { + "bbox": [ + 260, + 580, + 342, + 662 + ], + "type": "image", + "image_path": "a398ac2ebd84c478cc3bfb296599c800f0137c4c802293d9fdbf47f12a443de8.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 350, + 580, + 431, + 662 + ], + "blocks": [ + { + "bbox": [ + 350, + 580, + 431, + 662 + ], + "lines": [ + { + "bbox": [ + 350, + 580, + 431, + 662 + ], + "spans": [ + { + "bbox": [ + 350, + 580, + 431, + 662 + ], + "type": "image", + "image_path": "e85d45af699882bcd1def9e555d2cd2bf10139ef73e17d47ffa1c82237fb6f36.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 440, + 580, + 521, + 662 + ], + "blocks": [ + { + "bbox": [ + 440, + 580, + 521, + 662 + ], + "lines": [ + { + "bbox": [ + 440, + 580, + 521, + 662 + ], + "spans": [ + { + "bbox": [ + 440, + 580, + 521, + 662 + ], + "type": "image", + "image_path": "19c9b6eeaceb1c10ba439b354a5185d46cf6c8f4c1f4a9e62a2333cf0486d34c.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + } + ], + "index": 42 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20471" + } + ] + } + ], + "index": 44 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 286, + 110 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 286, + 110 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 286, + 110 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 286, + 110 + ], + "type": "table", + "html": "
MethodSemanticsDiversityPhysical Plausibility
CLIP ↑Entropy ↑Cluster Size ↑Non-collision ↑Contact ↑
Hassan et al. [10]0.25982.70141.19070.88240.9669
Ours-SV0.26132.64521.58130.97651.0000
Ours0.27102.73041.05000.97670.9868
", + "image_path": "05c45a05af4b7b36df8bf598d94c8ce986e97a4de409f166d6c0e8741255cc5a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 82, + 180, + 248, + 201 + ], + "blocks": [ + { + "bbox": [ + 46, + 114, + 287, + 169 + ], + "lines": [ + { + "bbox": [ + 46, + 114, + 287, + 169 + ], + "spans": [ + { + "bbox": [ + 46, + 114, + 287, + 169 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons on the Sketchfab dataset. Our approach achieves the best semantic consistency, diversity entropy, and non-collision scores, with the contact score on par. Note that single view methods Hassan et al. [10] and Ours-SV tend to produce increased diversity at the cost of semantic plausibility." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 82, + 180, + 248, + 201 + ], + "lines": [ + { + "bbox": [ + 82, + 180, + 248, + 201 + ], + "spans": [ + { + "bbox": [ + 82, + 180, + 248, + 201 + ], + "type": "table", + "html": "
w/o DMw/o VCw/o IROurs
CLIP ↑0.26390.26940.26640.2710
", + "image_path": "09aae217713a2af14007ec9f06693fb8b26a61d24638003c63356b2308927337.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 205, + 287, + 250 + ], + "lines": [ + { + "bbox": [ + 46, + 205, + 287, + 250 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 287, + 250 + ], + "type": "text", + "content": "Table 2. Ablation study on the Sketchfab dataset. The semantic consistency of 3D interaction generations degrades without dynamic masking (DM), view consistency (VC), or iterative refinement (IR), compared to our full approach." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 48, + 261, + 286, + 329 + ], + "blocks": [ + { + "bbox": [ + 48, + 261, + 286, + 329 + ], + "lines": [ + { + "bbox": [ + 48, + 261, + 286, + 329 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 286, + 329 + ], + "type": "image", + "image_path": "e1f5615d87b51a9a6de42d9775a816c93365f793d43cda462f9937ee3c946080.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 331, + 286, + 396 + ], + "lines": [ + { + "bbox": [ + 46, + 331, + 286, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 286, + 396 + ], + "type": "text", + "content": "Figure 6. Visualization of our method ablations on Sketchfab dataset for the input text: \"sitting on a bar stool\". Without dynamic masking (DM) or view consistency (VC), the person floats above the middle stool. Without iterative refinement (IR), the person penetrates the stool. Our full approach results in a more realistic synthesis." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 407, + 286, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 407, + 286, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 286, + 515 + ], + "type": "text", + "content": "GenZI in the perceptual studies. We note that since both Hassan et al. [10] and Ours-Single View operate on single inpainted view samples, results can be very diverse but lack 3D plausibility. This suggests that the CLIP score is a more reliable metric for assessing the HSI synthesis quality, compared to the diversity and physical plausibility metrics, which reflect less about the generation realism. Nevertheless, our approach has the best diversity entropy and noncollision scores, with the contact score on par." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 518, + 286, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 518, + 286, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 518, + 286, + 662 + ], + "type": "text", + "content": "Qualitative Evaluation. We show qualitative comparisons in Fig. 5. COINS is severely limited by its training on the closed set of indoor interactions, and thus fails to generalize to outdoor scenes and unseen objects (e.g., no curved floors exist during training, and limited sets of heights and shapes of chairs). As Hassan et al. [10] and Ours-Single View operate from single views, they both suffer from insufficient pose constraints from other views for plausible interaction generation. In contrast, our approach demonstrates high flexibility and generality to a diverse set of 3D indoor and outdoor scenes by leveraging large VLMs to imagine multiview interaction hypotheses and then robust 3D lifting." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 670, + 147, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 147, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 147, + 682 + ], + "type": "text", + "content": "4.2. Ablation Studies" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 689, + 286, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 286, + 714 + ], + "type": "text", + "content": "We conduct ablation studies on the Sketchfab dataset to validate the effectiveness of our proposed dynamic masking" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "content": "scheme (Sec. 3.2), robust 3D lifting with view consistency (Sec. 3.3), and iterative refinement (Sec. 3.4). Results are presented in Fig. 6 and Tab. 2." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 110, + 545, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 110, + 545, + 231 + ], + "spans": [ + { + "bbox": [ + 304, + 110, + 545, + 231 + ], + "type": "text", + "content": "Dynamic Masking (DM). We replace our dynamic masking, used during latent diffusion inpainting, with random masking. We sample a random mask around the image center covering at least " + }, + { + "bbox": [ + 304, + 110, + 545, + 231 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 304, + 110, + 545, + 231 + ], + "type": "text", + "content": " of the image area, and use it as a fixed mask input to the latent diffusion model " + }, + { + "bbox": [ + 304, + 110, + 545, + 231 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 304, + 110, + 545, + 231 + ], + "type": "text", + "content": ". We observe that random masking results in noticeably worse quality of interaction synthesis (Tab. 2), and incoherence with the scene (Fig. 6, floating on the stool). This indicates that our dynamic masking is effective in incorporating sufficient scene context for human inpainting." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 233, + 545, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 233, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 304, + 233, + 545, + 317 + ], + "type": "text", + "content": "View Consistency (VC). We evaluate the role of view consistency in robust 3D lifting by fixing the estimizable scores to " + }, + { + "bbox": [ + 304, + 233, + 545, + 317 + ], + "type": "inline_equation", + "content": "\\mathbf{w} = \\mathbf{1}" + }, + { + "bbox": [ + 304, + 233, + 545, + 317 + ], + "type": "text", + "content": ". Using all inpainted views leads to averaged, less expressive 3D human poses (Tab. 2, Fig. 6) due to potential inconsistent 2D pose hypotheses across views. By allowing the solver to adaptively focus on views with consistent inpaintings, our approach generates more realistic 3D HSIs." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 319, + 545, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 545, + 368 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 545, + 368 + ], + "type": "text", + "content": "Iterative Refinement (IR). Finally, we show the effectiveness of iterative refinement by applying our VLM-based 2D inpainting and robust 3D lifting only once. Tab. 2 and Fig. 6 show that iterative refinement improves synthesis quality." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 370, + 545, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 370, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 304, + 370, + 545, + 454 + ], + "type": "text", + "content": "Limitations. Our approach is limited by the inpainting capability of latent diffusion models to imagine possible 2D human-scene compositions, and the diffusion models are also known to be slow at inference time due to their iterative nature [14]. Nevertheless, we believe that our approach can directly benefit from the rapid advancement of VLMs for improved HSI synthesis." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 464, + 378, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 464, + 378, + 475 + ], + "spans": [ + { + "bbox": [ + 306, + 464, + 378, + 475 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 483, + 545, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 545, + 650 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 545, + 650 + ], + "type": "text", + "content": "We have presented the first approach to synthesize general 3D human-scene interactions guided by text inputs. Key to our approach is effective distillation of knowledge from large vision-language models, enabling generating 3D humans in scenes without requiring any 3D interaction data for training. We leverage these powerful vision-language models to generate hypotheses for inpainted 2D human-scene interactions. We then formulate a robust optimization to lift the hypotheses to 3D in a view-consistent fashion by simultaneously optimizing for the most informative 2D hypotheses. Our approach is flexible and can be applied to general scene settings for a variety of actions. We believe this opens up new opportunities for 3D understanding without requiring expensive capture of 3D/4D data." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgements. This project is funded by the ERC Starting Grant SpatialSem (101076253), and the German Research Foundation (DFG) Grant \"Learning How to Interact with Scenes through Part-Based Understanding\", and supported in part by a Google research gift." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20472" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "text", + "content": "[1] Stable Diffusion v2. https://huggingface.co/stabilityai/stable-diffusion-2-inpainting.Accessed:2023-09-27.4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 286, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 286, + 158 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 286, + 158 + ], + "type": "text", + "content": "[2] Luca Ballan, Aparna Taneja, Jürgen Gall, Luc Van Gool, and Marc Pollefeys. Motion capture of hands in action using discriminative salient points. In ECCV, 2012. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 160, + 286, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 160, + 286, + 204 + ], + "spans": [ + { + "bbox": [ + 53, + 160, + 286, + 204 + ], + "type": "text", + "content": "[3] Zhongang Cai, Wanqi Yin, Ailing Zeng, Chen Wei, Qingping Sun, Yanjun Wang, Hui En Pang, Haiyi Mei, Mingyuan Zhang, Lei Zhang, et al. SMPLer-X: Scaling up expressive human pose and shape estimation. arXiv, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 206, + 286, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 206, + 286, + 249 + ], + "spans": [ + { + "bbox": [ + 53, + 206, + 286, + 249 + ], + "type": "text", + "content": "[4] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-Excite: Attention-based semantic guidance for text-to-image diffusion models. ACM TOG, 2023. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 251, + 286, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 251, + 286, + 294 + ], + "spans": [ + { + "bbox": [ + 53, + 251, + 286, + 294 + ], + "type": "text", + "content": "[5] Vasileios Choutas, Georgios Pavlakos, Timo Bolkart, Dimitrios Tzionas, and Michael J Black. Monocular expressive body regression through body-driven attention. In ECCV, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 297, + 286, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 297, + 286, + 341 + ], + "spans": [ + { + "bbox": [ + 53, + 297, + 286, + 341 + ], + "type": "text", + "content": "[6] Hao-Shu Fang, Jiefeng Li, Hongyang Tang, Chao Xu, Haoyi Zhu, Yuliang Xiu, Yong-Lu Li, and Cewu Lu. Alpha-Pose: Whole-body regional multi-person pose estimation and tracking in real-time. IEEE TPAMI, 2022. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 343, + 286, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 343, + 286, + 364 + ], + "spans": [ + { + "bbox": [ + 53, + 343, + 286, + 364 + ], + "type": "text", + "content": "[7] Dariu M Gavrila. The visual analysis of human movement: A survey. CVIU, 1999. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 367, + 286, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 367, + 286, + 388 + ], + "spans": [ + { + "bbox": [ + 53, + 367, + 286, + 388 + ], + "type": "text", + "content": "[8] Stuart Geman and Donald E. McClure. Statistical methods for tomographic image reconstruction. 1987. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 391, + 286, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 391, + 286, + 412 + ], + "spans": [ + { + "bbox": [ + 53, + 391, + 286, + 412 + ], + "type": "text", + "content": "[9] James J Gibson. The ecological approach to the visual perception of pictures. *Leonardo*, 11(3):227-235, 1978. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 415, + 286, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 286, + 456 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 286, + 456 + ], + "type": "text", + "content": "[10] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J. Black. Resolving 3D human pose ambiguities with 3D scene constraints. In ICCV, 2019. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 460, + 286, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 286, + 492 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 286, + 492 + ], + "type": "text", + "content": "[11] Mohamed Hassan, Partha Ghosh, Joachim Tesch, Dimitrios Tzionas, and Michael J. Black. Populating 3D scenes by learning human-scene interaction. In CVPR, 2021. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 495, + 286, + 526 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 495, + 286, + 526 + ], + "spans": [ + { + "bbox": [ + 48, + 495, + 286, + 526 + ], + "type": "text", + "content": "[12] Mohamed Hassan, Yunrong Guo, Tingwu Wang, Michael Black, Sanja Fidler, and Xue Bin Peng. Synthesizing physical character-scene interactions. In SIGGRAPH, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 529, + 286, + 561 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 529, + 286, + 561 + ], + "spans": [ + { + "bbox": [ + 48, + 529, + 286, + 561 + ], + "type": "text", + "content": "[13] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-prompt image editing with cross-attention control. In ICLR, 2023. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 564, + 286, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 564, + 286, + 585 + ], + "spans": [ + { + "bbox": [ + 48, + 564, + 286, + 585 + ], + "type": "text", + "content": "[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 588, + 286, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 588, + 286, + 620 + ], + "spans": [ + { + "bbox": [ + 48, + 588, + 286, + 620 + ], + "type": "text", + "content": "[15] Lukas Hollein, Ang Cao, Andrew Owens, Justin Johnson, and Matthias Nießner. Text2Room: Extracting textured 3d meshes from 2d text-to-image models. In ICCV, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 622, + 286, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 622, + 286, + 665 + ], + "spans": [ + { + "bbox": [ + 48, + 622, + 286, + 665 + ], + "type": "text", + "content": "[16] Siyuan Huang, Zan Wang, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, and Song-Chun Zhu. Diffusion-based generation, optimization, and planning in 3D scenes. In CVPR, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 668, + 265, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 668, + 265, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 668, + 265, + 678 + ], + "type": "text", + "content": "[17] Kurt Koffka. Principles of gestalt psychology. 1935. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 48, + 681, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 286, + 712 + ], + "type": "text", + "content": "[18] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In ICCV, 2019. 2" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[19] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. BLIP: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 119, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 119, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 119, + 545, + 162 + ], + "type": "text", + "content": "[20] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 165, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 165, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 165, + 545, + 207 + ], + "type": "text", + "content": "[21] Xueting Li, Sifei Liu, Kihwan Kim, Xiaolong Wang, Ming-Hsuan Yang, and Jan Kautz. Putting humans in a scene: Learning affordance in 3D indoor environments. In CVPR, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 210, + 545, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 210, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 307, + 210, + 545, + 243 + ], + "type": "text", + "content": "[22] Jing Lin, Ailing Zeng, Haoqian Wang, Lei Zhang, and Yu Li. One-stage 3d whole-body mesh recovery with component aware transformer. In CVPR, pages 21159-21168, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 245, + 545, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 245, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 307, + 245, + 545, + 277 + ], + "type": "text", + "content": "[23] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. SMPL: A skinned multiperson linear model. ACM TOG, 34(6):1-16, 2015. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 280, + 545, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 280, + 545, + 301 + ], + "spans": [ + { + "bbox": [ + 307, + 280, + 545, + 301 + ], + "type": "text", + "content": "[24] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 304, + 545, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 304, + 545, + 336 + ], + "spans": [ + { + "bbox": [ + 307, + 304, + 545, + 336 + ], + "type": "text", + "content": "[25] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. DPM-Solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv, 2022. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 338, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 338, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 307, + 338, + 545, + 370 + ], + "type": "text", + "content": "[26] Thomas B Moeslund, Adrian Hilton, and Volker Krüger. A survey of advances in vision-based human motion capture and analysis. CVIU, 2006. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 373, + 545, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 373, + 545, + 416 + ], + "spans": [ + { + "bbox": [ + 307, + 373, + 545, + 416 + ], + "type": "text", + "content": "[27] Aron Monszpart, Paul Guerrero, Duygu Ceylan, Ersin Yumer, and Niloy J Mitra. iMapper: interaction-guided scene mapping from monocular videos. ACM Transactions On Graphics (TOG), 38(4):1-15, 2019. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 418, + 545, + 484 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 418, + 545, + 484 + ], + "spans": [ + { + "bbox": [ + 307, + 418, + 545, + 484 + ], + "type": "text", + "content": "[28] A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. Kopf, E. Yang, Z. DeVito, M. Raison, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, and S. Chintala. PyTorch: An imperative style, high-performance deep learning library. NeurIPS, 2019. 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 487, + 545, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 487, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 307, + 487, + 545, + 529 + ], + "type": "text", + "content": "[29] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3D hands, face, and body from a single image. In CVPR, 2019. 2, 3, 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 533, + 545, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 533, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 307, + 533, + 545, + 597 + ], + "type": "text", + "content": "[30] R. Po, W. Yifan, V. Golyanik, K. Aberman, J. T. Barron, A. H. Bermano, E. R. Chan, T. Dekel, A. Holynski, A. Kanazawa, C. K. Liu, L. Liu, B. Mildenhall, M. Nießner, B. Ommer, C. Theobalt, P. Wonka, and G. Wetzstein. State of the art on diffusion models for visual computing. arXiv, 2023. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 600, + 545, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 600, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 307, + 600, + 545, + 620 + ], + "type": "text", + "content": "[31] Ronald Poppe. Vision-based human motion analysis: An overview. CVIU, 2007. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 624, + 545, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 655 + ], + "type": "text", + "content": "[32] Sergey Prokudin, Christoph Lassner, and Javier Romero. Efficient learning on point clouds with basis point sets. In ICCV, 2019. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "text", + "content": "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3, 6" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20473" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 73, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 73, + 287, + 115 + ], + "type": "text", + "content": "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3, 4, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 119, + 286, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 119, + 286, + 152 + ], + "spans": [ + { + "bbox": [ + 49, + 119, + 286, + 152 + ], + "type": "text", + "content": "[35] Yu Rong, Takaaki Shiratori, and Hanbyul Joo. FrankMocap: A monocular 3d whole-body pose estimation system via regression and integration. In ICCV, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 154, + 287, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 154, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 49, + 154, + 287, + 185 + ], + "type": "text", + "content": "[36] David Rozenberszki, Or Litany, and Angela Dai. Language-grounded indoor 3d semantic segmentation in the wild. In ECCV, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 189, + 287, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 189, + 287, + 243 + ], + "spans": [ + { + "bbox": [ + 49, + 189, + 287, + 243 + ], + "type": "text", + "content": "[37] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 2022. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 245, + 286, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 245, + 286, + 277 + ], + "spans": [ + { + "bbox": [ + 49, + 245, + 286, + 277 + ], + "type": "text", + "content": "[38] Nikolaos Sarafianos, Bogdan Boteanu, Bogdan Ionescu, and Ioannis A Kakadiaris. 3D human pose estimation: A review of the literature and analysis of covariates. CVIU, 2016. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 280, + 286, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 280, + 286, + 322 + ], + "spans": [ + { + "bbox": [ + 49, + 280, + 286, + 322 + ], + "type": "text", + "content": "[39] Manolis Savva, Angel X Chang, Pat Hanrahan, Matthew Fisher, and Matthias Nießner. PiGraphs: learning interaction snapshots from observations. ACM TOG, 35(4):1-12, 2016. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 326, + 286, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 326, + 286, + 380 + ], + "spans": [ + { + "bbox": [ + 49, + 326, + 286, + 380 + ], + "type": "text", + "content": "[40] Matthias Teschner, Stefan Kummerle, Bruno Heidelberger, Gabriel Zachmann, Laks Raghupathi, Arnulph Fuhrmann, M-P Cani, François Faure, Nadia Magnenat-Thalmann, Wolfgang Strasser, et al. Collision detection for deformable objects. In CGF, 2005. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 382, + 286, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 382, + 286, + 425 + ], + "spans": [ + { + "bbox": [ + 49, + 382, + 286, + 425 + ], + "type": "text", + "content": "[41] Dimitrios Tzionas, Luca Ballan, Abhilash Srikantha, Pablo Aponte, Marc Pollefeys, and Juergen Gall. Capturing hands in action using discriminative salient points and physics simulation. IJCV, 2016. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 428, + 286, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 428, + 286, + 460 + ], + "spans": [ + { + "bbox": [ + 49, + 428, + 286, + 460 + ], + "type": "text", + "content": "[42] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 463, + 286, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 463, + 286, + 495 + ], + "spans": [ + { + "bbox": [ + 49, + 463, + 286, + 495 + ], + "type": "text", + "content": "[43] Jingbo Wang, Yu Rong, Jingyuan Liu, Sijie Yan, Dahua Lin, and Bo Dai. Towards diverse and natural scene-aware 3D human motion synthesis. In CVPR, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 498, + 286, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 498, + 286, + 529 + ], + "spans": [ + { + "bbox": [ + 49, + 498, + 286, + 529 + ], + "type": "text", + "content": "[44] Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, Wei Liang, and Siyuan Huang. HUMANISE: Language-conditioned human motion generation in 3D scenes. NeurIPS, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 533, + 286, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 533, + 286, + 574 + ], + "spans": [ + { + "bbox": [ + 49, + 533, + 286, + 574 + ], + "type": "text", + "content": "[45] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In CVPR, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 578, + 286, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 578, + 286, + 621 + ], + "spans": [ + { + "bbox": [ + 49, + 578, + 286, + 621 + ], + "type": "text", + "content": "[46] Haibiao Xuan, Xiongzheng Li, Jinsong Zhang, Hongwen Zhang, Yebin Liu, and Kun Li. Narrator: Towards natural control of human-scene interaction generation via relationship reasoning. arXiv, 2023. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 624, + 286, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 624, + 286, + 666 + ], + "spans": [ + { + "bbox": [ + 49, + 624, + 286, + 666 + ], + "type": "text", + "content": "[47] Yufei Ye, Xueting Li, Abhinav Gupta, Shalini De Mello, Stan Birchfield, Jiaming Song, Shubham Tulsiani, and Sifei Liu. Affordance diffusion: Synthesizing hand-object interactions. In CVPR, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 670, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 670, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 670, + 286, + 712 + ], + "type": "text", + "content": "[48] Hongwen Zhang, Yating Tian, Yuxiang Zhang, Mengcheng Li, Liang An, Zhenan Sun, and Yebin Liu. PyMAF-X: Towards well-aligned full-body model regression from monocular images. IEEE TPAMI, 2023. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 251 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[49] Siwei Zhang, Yan Zhang, Qianli Ma, Michael J Black, and Siyu Tang. Place: Proximity learning of articulation and contact in 3D environments. In 3DV, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 107, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 140 + ], + "type": "text", + "content": "[50] Yan Zhang, Mohamed Hassan, Heiko Neumann, Michael J. Black, and Siyu Tang. Generating 3D people in scenes without people. In CVPR, 2020. 1, 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 141, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 545, + 173 + ], + "type": "text", + "content": "[51] Kaifeng Zhao, Shaofei Wang, Yan Zhang, Thabo Beeler, and Siyu Tang. Compositional human-scene interaction synthesis with semantic control. In ECCV, 2022. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 175, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 175, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 308, + 175, + 545, + 206 + ], + "type": "text", + "content": "[52] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In CVPR, 2019. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 209, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 251 + ], + "type": "text", + "content": "[53] Yuxiao Zhou, Marc Habermann, Ikhsanul Habibie, Ayush Tewari, Christian Theobalt, and Feng Xu. Monocular real-time full body capture with inter-part correlations. In CVPR, 2021. 2" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20474" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/GeneAvatar_ Generic Expression-Aware Volumetric Head Avatar Editing from a Single Image/2d50005a-8c8d-4327-bf7a-e48503a1d824_content_list.json b/2024/GeneAvatar_ Generic Expression-Aware Volumetric Head Avatar Editing from a Single Image/2d50005a-8c8d-4327-bf7a-e48503a1d824_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7d8165c8c88938a4a4f99baa685703a49fbb8402 --- /dev/null +++ b/2024/GeneAvatar_ Generic Expression-Aware Volumetric Head Avatar Editing from a Single Image/2d50005a-8c8d-4327-bf7a-e48503a1d824_content_list.json @@ -0,0 +1,3764 @@ +[ + { + "type": "text", + "text": "GeneAvatar: Generic Expression-Aware Volumetric Head Avatar Editing from a Single Image", + "text_level": 1, + "bbox": [ + 116, + 84, + 854, + 130 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chong Bao $^{1*§}$ Yinda Zhang $^{2*}$ Yuan Li $^{1*}$ Xiyu Zhang $^{1}$ Bangbang Yang $^{4}$", + "bbox": [ + 171, + 157, + 790, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hujun Bao $^{1}$ Marc Pollefeys $^{3}$ Guofeng Zhang $^{1}$ Zhaopeng Cui $^{1\\dagger}$", + "bbox": [ + 210, + 176, + 763, + 194 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ State Key Lab of CAD&CG, Zhejiang University $^{2}$ Google $^{3}$ ETH Zürich $^{4}$ ByteDance", + "bbox": [ + 119, + 193, + 852, + 212 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2e0986f0a2bda4ee93e297c0c43264b7365ebb9b2b4eb81632b5107250b134b6.jpg", + "image_caption": [ + "GeneAvatar + NeBShape" + ], + "image_footnote": [], + "bbox": [ + 86, + 234, + 344, + 358 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7d863c72e7ab48cd31396cafe0145ba7a2de47719d4017b9f45a880d8a6f4308.jpg", + "image_caption": [ + "GeneAvatar + INSTA", + "Figure 1. We propose a generic approach to edit 3D avatars in various volumetric representations (NeRFBblendShape [16], INSTA [76], Next3D [50]) from a single perspective using 2D editing methods with drag-style, text-prompt and pattern painting. Our editing results are consistent across multiple facial expression and camera viewpoints." + ], + "image_footnote": [], + "bbox": [ + 367, + 231, + 612, + 358 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c0c1a36e134b9c478fe95ae59a9424db14c2501c633a91a929b807ea926e0322.jpg", + "image_caption": [ + "GeneAvatar + Next3D" + ], + "image_footnote": [], + "bbox": [ + 622, + 232, + 887, + 357 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 409, + 313, + 425 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, we have witnessed the explosive growth of various volumetric representations in modeling animatable head avatars. However, due to the diversity of frameworks, there is no practical method to support high-level applications like 3D head avatar editing across different representations. In this paper, we propose a generic avatar editing approach that can be universally applied to various 3DMM-driving volumetric head avatars. To achieve this goal, we design a novel expression-aware modification generative model, which enables lift 2D editing from a single image to a consistent 3D modification field. To ensure the effectiveness of the generative modification process, we develop several techniques, including an expression-dependent modification distillation scheme to draw knowledge from the large-scale head avatar model and 2D facial texture editing tools, implicit latent space guidance to enhance model convergence, and a segmentation-based loss reweight strategy for fine-grained texture inversion. Extensive experiments demonstrate that our method delivers high-quality and consistent results across multiple expression and viewpoints. Project page: https://zju3dv.github.io/geneavatar/.", + "bbox": [ + 75, + 428, + 473, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 753, + 209, + 768 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently various volumetric representations [3, 4, 15, 16, 57, 68, 69, 76] have achieved remarkable success in reconstructing personalized, animatable, and photorealistic head avatars using implicit [15, 16, 57, 68, 69] or explicit [3, 4, 76] conditioning of 3D Morphable Models", + "bbox": [ + 75, + 779, + 468, + 856 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(3DMM) [6]. A popular demand, once with a created avatar model, is to edit the avatar, e.g., for face shape, facial makeup, or apply artistic effects, for the downstream applications, e.g., in virtual/augmented reality.", + "bbox": [ + 500, + 410, + 890, + 470 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ideally, the desired editing functionality on the animatable avatar should have the following properties. (1) Adaptable: The editing method should be applicable across various volumetric avatar representations. This is particularly valuable in light of the growing diversity of avatar frameworks [16, 50, 76]. (2) User-friendly: The editing should be user-friendly and intuitive. Preferably, the editing of geometry and texture of the 3D avatar could be accomplished on a single-perspective rendered image. (3) Faithful: The editing results should be consistent across various facial expression and camera viewpoints. (4) Flexible: Both intensive editing (e.g., global appearance transfer following style prompts) and delicate local editing (e.g., dragging to enlarge eyes or ears) should be supported as illustrated in Fig. 1.", + "bbox": [ + 496, + 473, + 892, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, 3D-aware avatar editing is still underexplored in both geometry and texture. One plausible way is to perform 3D editing via animatable 3D GAN [50, 52, 54], but the editing results may not be consistently reflected when expression and camera viewpoint change. Alternatively, the editing can be done on the generated 2D video using 2D personalized StyleGAN [28]; however, the identity shift is often observed. Some face-swapping methods [11, 12, 42] are capable of substituting the face in a video with another face derived from a reference image or video; however, they do not support texture editing and local geometry editing.", + "bbox": [ + 496, + 686, + 893, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To this end, we propose GeneAvatar - a generic approach to support fine-grained 3D editing in various volumetric avatar representations from a single perspective by", + "bbox": [ + 498, + 854, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 236, + 0, + 808, + 18 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 323, + 16, + 722, + 30 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 292, + 31, + 753, + 45 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Authors contributed equally.", + "bbox": [ + 93, + 862, + 253, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding authors.", + "bbox": [ + 96, + 875, + 227, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\S$ The work was partially done when visiting ETHZ.", + "bbox": [ + 94, + 887, + 367, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8952", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "leveraging 2D editing methods, such as drag-based methods [30, 34, 39, 47], text-driven methods [7, 17, 20, 40, 41], or image editing tools like Photoshop (see Fig. 1). We adopt a novel editing framework that formulates the editing as predicting expression-aware 3D modification fields applied in the geometry and texture space of the volumetric avatars, which makes editing independent with the original representation as long as they are in parametric-driven radiance field, e.g., 3DMM-based neural avatar. Second, to ensure that the 2D image editing can be faithfully transferred into the 3D space, we propose to learn a generative model for modification fields, which produces 3DMM conditional modification fields from a compact latent space. Given the rendered avatar image and its edited counterpart, we conduct auto-decoding optimization on this generative model to search for the latent code that best explains the editing, obtaining consistent 3DMM conditional modification fields across various viewpoints and expression. Third, inspired by the spirit of learning from the pre-trained large-scale generative model [7, 18, 32, 74], we design a novel distillation scheme to learn the expression-dependent modification from a 3DMM-based GAN [50] and 2D face editing tools [22, 27, 36]. The scheme addresses the issue of insufficient real training data (i.e., avatars with a wide range of geometry and texture changes). Besides, we develop several techniques to enhance the editing effects, including the implicit latent space guidance to stabilize the initialization and convergence of learning, and a segmentation-based loss reweight strategy for fine-grained texture inversion.", + "bbox": [ + 76, + 90, + 472, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contributions of our paper are summarized as follows. 1) We propose a generic avatar editing approach that can be applied to various 3DMM driving head avatars in the neural radiance field. To achieve this, we design a novel expression-aware modification generative model, which lifts the geometry and texture editing from a single image to a consistent 3D modification field. 2) To bootstrap the training of the modification generator with limited real paired training data, we design a distillation scheme to learn the expression-dependent geometry and texture modification from the large-scale head avatar generative model [50] and 2D face texture editing tools [22, 27, 36], and develop several techniques, including implicit guidance in latent space to improve training convergence, and a loss reweight strategy based on segmentation for fine-grained texture inversion. 3) Extensive experiments on various head avatar representations demonstrate that our method delivers high-quality editing results and the editing effects are consistent under different viewpoints and expression.", + "bbox": [ + 76, + 529, + 472, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 821, + 218, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2D Head Avatar Editing. The manipulation of 2D head avatars has made significant strides in recent years. Various GAN-based methods [23-25] can result in precise and high-resolution human face editing by leveraging image space se", + "bbox": [ + 76, + 839, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "mantic information [29, 72] or controlling latent space explorations [10, 19, 46, 70]. Some approaches [22, 27, 36, 61] focus on the task of makeup transferring by exploiting the GAN to learn the transferring ability from a large unaligned makeup and non-makeup face datasets. The drag-based GAN editing approach [39, 71] gained vast popularity due to providing a user-friendly editing way. PVP [28] uses a monocular video to fine-tune StyleGAN [23, 24] to obtain the personalized image generator and provide various editing functions. The diffusion models [7, 21] also show the capability of achieving fine-grained face editing with text prompt and other conditional input. For editing an avatar in a video, lots of face-swapping methods [11, 12, 42] have emerged to provide high-quality and properly aligned face-swapping results. However, these approaches typically suffer from multi-view consistency and identity preservation.", + "bbox": [ + 496, + 90, + 893, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Head Avatar Editing. Neural Radiance Field [33] has exhibited great reconstruction and rendering qualities in SLAM [62, 73], scene editing [5, 58-60, 64] and relighting [63, 66, 67], especially promoting the emergence of many 3D avatar reconstruction [4, 16, 53, 68, 69, 76] and generation [50, 52, 54]. Some methods [2, 49, 56, 65] exploit the powerful editing ability of GAN to edit a 3D static head portrait. However, they cannot be trivially extended to the dynamic avatars. The methods [37, 38, 45] focus on style transfer of the avatar using text prompt or style image but reach a poor identity-preserving. We propose a novel 3D avatar editing approach with an expression-aware modification generative model, which can be applied to various 3DMM-based volumetric avatars and render consistent novel views with fine-grained editing across multiple viewpoints and expression while preserving identity of person.", + "bbox": [ + 496, + 332, + 893, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 579, + 591, + 594 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As shown in Fig. 2, given a volumetric head avatar, we edit the avatar using a single-view image and synthesize consistent novel views across multiple expression and viewpoints. To achieve this goal, we propose a novel expression-aware modification generator to generate 3D modification fields, which can be seamlessly integrated into various representations and animated with facial expression (see Sec. 3.2). Furthermore, to bootstrap the training with limited pairwise data, we propose a novel expression-aware distillation scheme to learn the expression-dependent modifications from large-scale generative models [7, 50] (see Sec. 3.3). During the editing process, given a single edited image of a 3D avatar, we perform an auto-decoding optimization to lift 2D editing effect to the 3D space (see Sec. 3.4).", + "bbox": [ + 496, + 604, + 893, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 500, + 823, + 640, + 837 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The current implicit volumetric representations of head avatar [16, 50, 76] are mostly built upon the NeRF [33] or its variants [8, 9, 31, 35, 44, 48]. In general, the neural architecture can be simplified as an implicit field $\\mathbf{F}$ that", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8953", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f924cb4273c9bd6a9b02ba279b7b66597d0d3146b5b1a78fbc4dc60523bdb6cb.jpg", + "image_caption": [ + "Figure 2. We use an expression-aware generative model that accepts a modification latent code $\\mathbf{z}_{g / t}$ and 3DMM coefficients and outputs a modification field of a tri-plane structure. The modification field modifies the geometry and texture of the template avatar by deforming the sample points $\\mathbf{x}$ and blending the color $\\mathbf{c}_o$ with the modification color $\\mathbf{c}_{\\Delta}$ respectively. We lift the 2D editing effect to 3D using an auto-decoding optimization and synthesize novel views across different expression." + ], + "image_footnote": [], + "bbox": [ + 89, + 66, + 253, + 273 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/eba85a7295be3c26abebcef3021db96d255b1c5190a1691a3b83a9a22cd42097.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 258, + 66, + 709, + 273 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/527ac188974e9bd8d25909b1767c6a81f690304dd61d83fcb1618a945ca45345.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 66, + 875, + 273 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "takes position $\\mathbf{x}$ and view direction $\\mathbf{d}$ as inputs and predicts the geometry $\\sigma$ and the texture $\\mathbf{c}$ of the avatar, i.e., $(\\sigma, \\mathbf{c}) = \\mathbf{F}(\\mathbf{x}, \\mathbf{d})$ . Then, the volume rendering technique is used to render images as follows:", + "bbox": [ + 76, + 342, + 470, + 402 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {C} (\\boldsymbol {r}) = \\sum_ {i = 1} ^ {N} T _ {i} \\alpha_ {i} \\mathbf {c} _ {i}, \\quad T _ {i} = \\exp \\left(- \\sum_ {j = 1} ^ {i - 1} \\sigma^ {\\prime} _ {j} \\delta_ {j}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 412, + 468, + 463 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha_{i} = 1 - \\exp \\left(-\\sigma^{\\prime}_{i}\\delta_{i}\\right)$ , and $\\delta_{i}$ is the distance between adjacent samples along the ray. In order to animate the head avatar, 3DMM [6] is incorporated to describe the deformation implicitly [15, 16, 57, 68, 69] or explicitly [4, 76].", + "bbox": [ + 75, + 472, + 470, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Expression-aware Modification Generator", + "text_level": 1, + "bbox": [ + 76, + 542, + 437, + 556 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To enable the modification animated with the facial expressions, we follow the architecture of 3DMM-based 3D GAN [50] to build our expression-aware modification generator. As shown in Fig. 2, our generator consists of a geometry generator $\\mathbf{G}_{\\Delta g}$ and a texture generator $\\mathbf{G}_{\\Delta t}$ . $\\mathbf{G}_{\\Delta g}$ encodes the expression-dependent geometry modification by deforming the query points in the edited space to the original template space under each expression. $\\mathbf{G}_{\\Delta t}$ encodes the expression-dependent modification color of query points under each expression:", + "bbox": [ + 75, + 559, + 468, + 710 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} ^ {\\prime} = \\mathbf {G} _ {\\Delta g} (\\mathbf {x}, \\mathbf {z} _ {g}, \\mathbf {v}), \\left(\\mathbf {c} _ {\\Delta}, \\beta_ {\\Delta}\\right) = \\mathbf {G} _ {\\Delta t} (\\mathbf {x}, \\mathbf {z} _ {t}, \\mathbf {v}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 720, + 468, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{x}$ is the query points in the edited space, and $\\mathbf{x}'$ is the deformed point in the space of original avatar under current expression $\\mathbf{e}$ . $\\mathbf{c}_{\\Delta}$ is the modification color and $\\beta_{\\Delta}$ determines the blending weights with the original color. $\\mathbf{z}_g,\\mathbf{z}_t$ are the geometry and texture modification latent code respectively, where $\\mathbf{z}_g,\\mathbf{z}_t\\in R^{1024}$ . They control the generation of modification feature maps in the UV space. $\\mathbf{v}$ is the 3DMM mesh vertices [26] that condition the current expression $\\mathbf{e}$ . Each mesh vertex has a neural feature that is retrieved in the modification feature map using pre-defined", + "bbox": [ + 75, + 750, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "UV mapping. We rasterize the vertex features to the three axis-aligned planes to generate the tri-plane feature. The modification information of the query point $\\mathbf{x}$ is first collected by bilinear interpolation on the tri-plane feature and then decoded by the neural feature decoder [8]. Since the modification is defined as decoupled fields without relying on the original field, our generated modification field can be integrated into various volumetric avatar representations and be animated following the facial expression.", + "bbox": [ + 496, + 342, + 890, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Expression-dependent Modification Learning", + "text_level": 1, + "bbox": [ + 498, + 487, + 883, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To learn the proposed expression-aware modification, we need extensive training data on avatars with a wide range of geometry and texture changes, which is hard to obtain in practice. Following the spirit of learning high-fidelity editing ability from the large-scale generative model [7, 18, 32, 74], we propose a novel expression-aware distillation scheme to deal with insufficient real training data. We leverage the ability of 3DMM-based 3D GAN [50] and 2D face texture editing tools to generate facial editing data, which encompasses a wide range of geometry and texture editing across various expression and viewpoints.", + "bbox": [ + 496, + 506, + 890, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Geometry Distillation. We use the teacher 3DMM-based 3D GAN [50] $\\mathbf{G}_n$ to synthesize two volumetric avatars with different geometry (an original avatar $\\mathbf{F}$ and an edited avatar $\\mathbf{F}^{\\prime}$ ) by modifying the 3DMM shape parameter of the original avatar. This provides the paired editing data for our generator to learn how to modify the geometry of the avatar while maintaining consistency across various expression and viewpoints. Specifically, we randomly sample the latent code $\\mathbf{z}_n$ in the latent space of $\\mathbf{G}_n$ as well as 3DMM shape parameter $\\beta$ , expression parameter $\\psi$ , and pose parameter $\\theta$ . $\\beta$ , $\\psi$ are sampled from a normal distribution whose absolute mean and standard deviation are within $[0,1]$ . $\\theta$ are a group of rotation vectors that have random directions within a unit sphere and magnitude within $[-6,6]$ degrees. Then, we sample an edit vector $\\beta_{\\Delta}$ from a uniform distribution", + "bbox": [ + 496, + 674, + 892, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8954", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\mathcal{U}(-3,3)$ and apply it to the original shape parameter by $\\beta' = \\beta_{\\Delta} + \\beta$ . These hyperparameters w.r.t. 3DMM coefficients sampling are selected empirically to maintain the shape definition of the human head. Please refer to our supplementary Sec. B.2 for more details on 3DMM sampling. The original avatar $\\mathbf{F}$ and paired edited avatar $\\mathbf{F}'$ are generated by $\\mathbf{F} = \\mathbf{G}_n(\\mathbf{z}_n, \\boldsymbol{\\beta}, \\boldsymbol{\\psi}, \\boldsymbol{\\theta})$ , $\\mathbf{F}' = \\mathbf{G}(\\mathbf{z}_n, \\boldsymbol{\\beta}', \\boldsymbol{\\psi}, \\boldsymbol{\\theta})$ . During training, we will apply our modification generator to modify the geometry of $\\mathbf{F}$ such that $F$ and $F'$ render the face with the same geometry.", + "bbox": [ + 75, + 90, + 472, + 243 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Texture Distillation. We distill the capabilities of fine-grained texture editing from 2D face editing algorithms by generating texture-modified avatar $\\mathbf{F}'$ with the teacher 3DMM GAN [50] $\\mathbf{G}_n$ . Specifically, we sample an original avatar $\\mathbf{F} = \\mathbf{G}_n(\\mathbf{z}_n, \\beta, \\psi, \\theta)$ from the teacher generator and render the image of its positive face. A segmentation-based 2D face texture editing algorithm (SBA) [77] and two makeup transfer algorithms (MTA) [22, 27, 36] are referred to in the distillation. We randomly choose one of them to edit the texture of the rendered face image. For SBA, we define several editable semantic regions of the face. A subset of these regions is selected randomly for texture painting using hues randomly sampled from the HSV color spectrum. For MTA, we randomly choose a makeup image as a reference from the open-sourced makeup dataset [22, 27, 36] and transfer the reference makeup to the rendered face image. The makeup dataset [36] contains complex makeup, such as blushes and makeup jewelry, which allow our generator to learn complicated texture editing patterns. Then, we perform the PTI inversion [43] on the texture-modified face image to lift the 2D texture editing to 3D space and obtain a texture-modified avatar $\\mathbf{F}'$ .", + "bbox": [ + 75, + 246, + 472, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Modification Learning. Following the training style of StyleGAN [24], we sample a modification latent code $\\mathbf{z}_{g / t} \\in R^{1024}$ in $\\mathcal{Z}$ latent space for each paired editing data. We do not fully sample a 1024-dimensional modification code but sample a reduced code $\\bar{\\mathbf{z}}$ from a standard normal distribution and concatenate it with the latent code $\\mathbf{z}_n$ of the original avatar $\\mathbf{F}$ that is sampled from the teacher model, i.e., $\\mathbf{z}_{g / t} = (\\mathbf{z}_n, \\bar{\\mathbf{z}}), \\mathbf{z}_n, \\bar{\\mathbf{z}} \\in R^{512}$ . This design is regarded as implicit code guidance that decently integrates knowledge from the teacher model to facilitate the model convergence. $\\mathbf{z}_n$ encodes the facial appearances of avatar $\\mathbf{F}$ , serving as a reference to the superimposition of the modification onto the avatar $\\mathbf{F}$ . Note that during inference, we do not require the concatenation of latent code from the teacher model and directly optimize the full modification code from the edited image using an auto-decoding manner. Our generator generates the modification field following Eq. (2) where $\\mathbf{v}$ is decoded from the 3DMM parameters of the avatar $\\mathbf{F}$ using FLAME model [26] $\\mathbf{E}$ , i.e., $\\mathbf{v} = \\mathbf{E}(\\boldsymbol{\\beta}, \\boldsymbol{\\psi}, \\boldsymbol{\\theta})$ . To apply the modification field, we feed the deformed query points $\\mathbf{x}'$ to the original avatar $\\mathbf{F}$ to", + "bbox": [ + 75, + 583, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "obtain the density and color, and composite the color with modification color by:", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} = \\left(1 - \\beta_ {\\Delta}\\right) * \\mathbf {c} _ {o} + \\beta_ {\\Delta} * \\mathbf {c} _ {\\Delta}, \\quad (\\sigma , \\mathbf {c} _ {o}) = \\mathbf {F} \\left(\\mathbf {x} ^ {\\prime}, \\mathbf {d}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 132, + 890, + 150 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, we perform volume rendering on the density $\\sigma$ and color $\\mathbf{c}$ using Eq. (1) to render the modified image $\\hat{I}_e$ of avatar $\\mathbf{F}$ . We use the photometric loss to supervise the modified image with the rendered image $\\hat{I}'$ from the edited avatar $\\mathbf{F}'$ under the same camera parameters.", + "bbox": [ + 498, + 160, + 893, + 236 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\left\\| \\hat {I} _ {e} - \\hat {I} ^ {\\prime} \\right\\| _ {2} ^ {2}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 638, + 244, + 890, + 263 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During training, we sample multiple viewpoints and 3DMM expression parameters $\\psi$ for each editing pair $(\\mathbf{F},\\mathbf{F}^{\\prime})$ to enhance the spatial consistency under different expressions.", + "bbox": [ + 498, + 273, + 890, + 319 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Avatar Editing with Single Image", + "text_level": 1, + "bbox": [ + 498, + 321, + 792, + 338 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our task, users are allowed to edit a single image with various out-of-box face editing tools, such as Photoshop, drag-based editing [30, 39], text-driven editing [7]. For each editing input, we use the auto-decoding optimization on modification code to lift 2D edits into a 3D expression-aware modification field generated by our model. This field adapts to expression and viewpoint changes and is not tied to the specific avatar representation. The StyleGAN-based generator [24, 43, 55] features a latent space mapping from $\\mathbf{z} \\in R^{Z}$ in $\\mathcal{Z}$ to $\\mathbf{w} \\in R^{W_{n} \\times W_{d}}$ in $\\mathcal{W}$ , where $\\mathbf{w}$ is more influential as it conditions the generator. Therefore, we perform code inversion in $\\mathcal{W}$ space by randomly sampling a modification code $\\mathbf{w}_{g/t}$ during editing. This code conditions a modification field $\\mathbf{G}_{\\Delta g/t}(\\mathbf{x}, \\mathbf{w}_{g/t}, \\mathbf{v})$ following Eq. (2). We apply the modification field to the original avatar using Eq. (4). The modified image $\\hat{I}_{e}$ is rendered following the original avatar's rendering pipeline and is encouraged to match the user-edited image $I_{e}$ by optimizing $\\mathbf{w}_{g/t}$ with the following loss terms:", + "bbox": [ + 496, + 339, + 890, + 628 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {i} = \\lambda_ {1} \\mathcal {L} _ {2} (\\hat {I} _ {e}, I _ {e}) + \\lambda_ {2} \\mathcal {L} _ {\\mathrm {l p i p s}} (\\hat {I} _ {e}, I _ {e}) + \\lambda_ {3} \\mathcal {L} _ {\\mathrm {r e g}} (\\mathbf {w}, \\mathbf {w} _ {\\mathrm {a v g}}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 504, + 638, + 890, + 670 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The L2 loss term $\\mathcal{L}_2$ and LPIPS perceptual loss term $\\mathcal{L}_{\\mathrm{lpips}}$ encourage the rendered face close to the appearance and structure of the edited face. We set $\\lambda_1 = 1000, \\lambda_2 = 1, \\lambda_3 = 1$ . The regularization term $\\mathcal{L}_{\\mathrm{reg}}$ applied to the latent code $\\mathbf{w}$ enforces alignment with the distribution of the $\\mathcal{W}$ space, with $\\mathbf{w}_{\\mathrm{avg}}$ representing the mean latent code computed from 1000 random samples within the $\\mathcal{W}$ space. Besides, we observe that the L2 loss on the whole face will give an underfitting result for the fine-grained makeup on the eyes, eyebrows and lips. Therefore, we reweight the L2 loss on the facial features by face segmentation mask $M = \\{N_i | i = 1, \\dots, m\\}$ for texture editing:", + "bbox": [ + 498, + 670, + 890, + 853 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {2} = \\sum_ {i = 1} ^ {m} \\sum_ {r \\in N _ {i}} \\frac {1}{| N _ {i} |} \\left\\| \\hat {C} _ {e} (r) - C _ {e} (r) \\right\\| _ {2} ^ {2}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 863, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8955", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/3c0cd1832132cae32f371e627d8cd9ead3046dd6070fc7d7093364b0b412071d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsGeometryTexture
RoopPVPNext3DOursRoopPVPNext3DOurs
Editing preservation ↑29.44%23.33%5.56%41.67%8.33%10.56%5.00%76.11%
Identity preservation ↑31.11%21.11%5.56%42.22%7.78%12.78%3.33%76.11%
Temporal consistency ↑30.00%23.89%4.44%41.67%5.56%12.78%1.67%80.00%
Overall ↑29.44%23.33%3.89%43.33%5.56%12.78%2.22%79.44%
image identity similarity ↑0.83730.87040.85470.88450.73200.84760.85000.9147
", + "bbox": [ + 80, + 69, + 467, + 141 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. We quantitatively compare with the PVP [28], Roop [12], Next3D [50] by user study and image identity similarity [13].", + "bbox": [ + 76, + 147, + 468, + 175 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{C}_e(r), C_e(r)$ are the rendered and target color respectively, $N_i$ are the rays within the $i$ -th semantic part of the face. Generally, we freeze the weight of the modification generator and only optimize the modification latent code w to reach a satisfied 3D modification result. When an intense makeup or complicated pattern is painted onto the human face, we will continuously fine-tune the weight of our generator and freeze the latent code w to achieve more accurate editing results. To animate the edited avatar, users can input the new 3DMM expression parameter to the original avatar and our modification generator simultaneously. In this way, the generated modification field tightly sticks to the original avatar and presents reasonable editing results under different expression and viewpoints.", + "bbox": [ + 75, + 189, + 472, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 412, + 209, + 429 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we evaluate our avatar editing capability from a single perspective. One major difference with static NeRF editing is that we focus on showing how the edits are correctly lifted to 3D avatars under various expression and camera viewpoints.", + "bbox": [ + 75, + 438, + 468, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Dataset and Baselines", + "text_level": 1, + "bbox": [ + 76, + 518, + 279, + 534 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We use a total of 19 neural implicit head avatars from three methods, i.e., 7 from INSTA [76], 8 from NeRF-BlendShape [16], 4 from Next3D [50], and show editing results on them. For INSTA [76] and NeRFBlendShape [16], we use the human head data (i.e., a monocular video of a head) provided by their methods to reconstruct the volumetric avatar using their respective representations. For Next3D [50], we random sample its latent space to generate volumetric avatars and perform editing on them. The evaluation datasets exhibit a substantial variation in identities, encompassing a diverse range of races, ages, and genders.", + "bbox": [ + 75, + 537, + 468, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We pick several baseline methods that can support single-view-based avatar editing. Roop [12] is a face-swapping method that can swap the human face in a video from a single reference view. To compare with Roop, we generate videos of the original avatar rendered in driving signals and single edited frames, and perform face swap. PVP [28] learns a personalized avatar image generator from a monocular video by fine-tuning the latent space of StyleGAN [24], and performs GAN-inversion style optimization [41] to edit the shape and appearance of the avatar. Next3D [50] is a 3DMM-based 3D GAN. To make a fair comparison on the same input data (i.e., a monocular video of avatar and an edited image), we perform GAN inver", + "bbox": [ + 75, + 704, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/43bbe264c875fe4dc42c506fb6c8a978b42bcd69db9f9f7e040985c836fe9a9f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 64, + 573, + 123 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3b5f7267aed76fbc90f89380c0630a0aa93e792934409ac98b04c6464f4009fa.jpg", + "image_caption": [ + "Hair++, Eyes++, Mouse" + ], + "image_footnote": [], + "bbox": [ + 506, + 127, + 558, + 179 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/786fd0f0a878b9e63eab588746a0ecf80e36ab8dd8c910fb3c704666189f0bb4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 576, + 63, + 660, + 119 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cf11cf46baf36d4d082e737f1a829d7539e40a5435a39547ee4070fef0e1eeab.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 576, + 119, + 692, + 140 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f1872ea41b792926f5f3629b6e6567d860f1aa245ae580afa212ad58d2f0dfc9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 576, + 141, + 661, + 198 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3428f2123db18035433243704f6d6d4d11708e2e0a46f9bf121a567250b8ce34.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 666, + 63, + 754, + 119 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3582093e97877d4d5a7bd0dd3b667ba7d0db2b5279fd6ac022bcef59fcaacf12.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 671, + 119, + 754, + 140 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b84f3ed2e81320d68371589cdebb5b812c4124e2d47711908666447b83e7b9e9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 763, + 64, + 813, + 119 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d6ee9dd647e17db201d638acd8edc83ebd77e3c24fe38d288159308f95b809ad.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 761, + 119, + 816, + 140 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ddad9fc09aa57dee0b664813232dff2635803507a570c156c59cdcd5f45f23de.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 821, + 64, + 870, + 119 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/320f79daa0c6adc0b20bbd09f87a86ff090e84ee04597f21ff749b1f70595a75.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 821, + 119, + 875, + 140 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9c4d61b461c9bc294ee5c35bae04ffb8722fccba24a2b9fd8ad8221c9c884b56.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 508, + 199, + 560, + 262 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b172bd92bdd7480f5fb4f432d5372048e3e458268ff6519d6a3177af230f5ebc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 262, + 560, + 324 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cc37aae23a265628ef1e8214f5400399ac17452f086fc21d9978bf7e80bb2417.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 580, + 199, + 632, + 257 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8d9b606642f75a4de68acafcbafaaf7bf581d2bfb4e9e6e0e6ea588cc816c69a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 576, + 257, + 635, + 277 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/73bfa2b7e1dfb4edd5bcf8e0941fbe4fb555e89db671424ebb5b55170dd14b4d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 199, + 694, + 257 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5fa147eda84fe98bbdc0733ca29629021f620c9a2765335de70f117ab1eec5b2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 257, + 692, + 277 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/15c450b3883804d63d2659e01ce3b8ca0d27e38fe587d9954da3a2efb8c61734.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 199, + 750, + 257 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/097146ef864b05f192edd0440c381450d72919bc7524c779c793d63f593cd816.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 257, + 756, + 277 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/95b827a23a7c086a6528271f59bfe83d88718c13b13dc0f9d65ed854fd8473f0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 792, + 199, + 874, + 257 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/545ff9b4f7f404d803c96ff05aae0dcf6137497646caabd7d46427fe8295eb03.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 761, + 257, + 818, + 277 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/62424f6e28e96a5024f72d5b028da3da1241069b9eb2b2bcfb98c52f40a66356.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 823, + 199, + 875, + 257 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a4276534daa3eb40310b271f3f47743fd7d664b00f58e13c3b74591fb8cd52f3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 823, + 257, + 880, + 277 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/77764d4e5ce0b53bbf8b92f29ce3a0c74862dd6247e63aa924ca21229a625b63.jpg", + "image_caption": [ + "Nose-, Lip++" + ], + "image_footnote": [], + "bbox": [ + 508, + 340, + 560, + 402 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e343c5cfc35b8bc98da8cc93c4f712ed9a4f5ffb1e93fda5798231f45557638b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 402, + 562, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ForeHead--", + "bbox": [ + 516, + 469, + 563, + 478 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e599c41ff5a5c6c4ffd7f8c148c18ba330cd054fe3a6ea520397f8cea250430b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 581, + 279, + 633, + 339 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/433eee09d0af752d581b901ca537b87053ba59982834cb1afd541e8af7e1cc00.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 581, + 340, + 635, + 398 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c7e387714b649051414e74ebb46e0f381cf237632e68a1711cd4ae104837647c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 280, + 696, + 338 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4c8496bd89347c4ac6a8187b189141f0ad25e9d0b0fc8af82257161c1df186a7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 340, + 697, + 398 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4bf9e90bb7e4ee5aefb9ad000f83880f90548dbe158956b074d4841bc95df597.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 280, + 753, + 338 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0ef874137b0767838154229871914045dc291d7a89d872be87657e2574e41ccc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 340, + 756, + 398 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0a2152595ce87420e6934bee87e2b3eeb54938e03e4d92ba2230961217520011.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 280, + 813, + 338 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/303976365356cdcd402ee5509a97778b231b9f94a20ce7a7e573d45f1904ee43.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 763, + 340, + 816, + 397 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5d3e0e2362c64a7f915cf5b6a9362e5f26153b12487181d30ce8818b25cf7728.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 823, + 280, + 875, + 338 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d041762c5b6d3f7aede17747b09e0b42fae10dfc74891eb89698fc71a5641f4f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 823, + 340, + 874, + 397 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f1044c94474c383d609feead99c75b0dbc0c4f8e56c933326cbaa7c123f0d0cc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 578, + 400, + 633, + 420 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/137be98d9d39f9494c4ff6735b8f9e96ccecc2f250901dc8e4c4baf5ab1faae9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 400, + 694, + 420 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/62b9861c4201e18303d031ef72fd9ca9664776cb3f80297a8879638ce7beac79.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 400, + 756, + 420 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/186b222fd37d76cabe5ba45a4f44f43d7ec07d595432c00d32b7db8325f97454.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 763, + 400, + 818, + 420 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/864e9deda6c8595f366462ba94874f501538d386551e6e8189945a824fb97f13.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 823, + 400, + 880, + 420 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "", + "bbox": [ + 506, + 421, + 560, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ForeHead--", + "bbox": [ + 516, + 469, + 563, + 478 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e8466da5d17f01e6573589d016afa53107bbb8b7ddf94167d77cbdb6a141b79a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 581, + 422, + 635, + 479 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6119cb31199c31852d9ecddbb8275798e5b61c478bdb6d7e59bd2e6aad691475.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 581, + 479, + 635, + 541 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/78fb73bac479c0a577c1a4689bdd01c7b21ac0d1f1b61cdf075f338a26b8c244.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 422, + 697, + 479 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e32b9abf2967bef31e638841f549e659f85dbcbe383dc863508f74edbc581560.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 479, + 696, + 541 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b5190ed077d108b883734f4cb89a49043dfa789174838bd18057d1074e5fe878.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 704, + 422, + 758, + 479 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e31bdf143f5a053b03cf5f5194beb4b2adceee34721a0b38251a7cb80d38bae5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 482, + 756, + 541 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/bc3ecca9b0e87ddfb2ff4e88ba23117a2302de9ad39402887453bae38bd15010.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 424, + 816, + 479 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0e020db4fc3fcec971db7d04b39a60adf2935d1d546d66f499214514822c3f21.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 821, + 424, + 875, + 479 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a4115f15bb39d540bed289fd1895e05c190a30d46de764b516f38a0731090895.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 576, + 542, + 633, + 563 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1d4e24d3f38814b6b71574a413b355b5264d28b43d137e35a4964c45b814c2b6.jpg", + "image_caption": [ + "2D Editing", + "Reference Animation", + "Figure 3. We compare geometry editing with PVP [28], Roop [12], Next3D [50] on INSTA [76] and NeRFBlendshape [16] avatars. The \"Reference Animation\" denotes the image of the original avatar under the same expression with the rendered edited view." + ], + "image_footnote": [], + "bbox": [ + 581, + 563, + 635, + 617 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e790a4536369185f73ebdebc5f0426bf2622c17a280a93ebb85fb9a7437dac1b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 542, + 694, + 563 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f8b85e621ee8c72b2701d193d5b3253e043a529ed520955369ddeb50d53f1031.jpg", + "image_caption": [ + "PVP" + ], + "image_footnote": [], + "bbox": [ + 645, + 564, + 692, + 617 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/34ddfcb176d2747a0a68641541ce75f078962dc5082bcfb73be22f4d8dc03b61.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 704, + 542, + 756, + 563 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a653d4bbd77d58931dcadf6deba500da701bfd11db45fc7eb314d141c5a641c6.jpg", + "image_caption": [ + "Roop", + "Next3D" + ], + "image_footnote": [], + "bbox": [ + 702, + 564, + 753, + 617 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d6697f8dbf19f3dfbb45a456787d167f8012a432afa93307f33601d781409d5e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 767, + 482, + 816, + 541 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f2455b140aa55f770aa54608605506466a3354d0e990ca92987987f53727990e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 763, + 542, + 816, + 563 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c7fa26a7185bbf1e5f23cdc73c2a0c3c826e12c5c067789d011820177a272215.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 821, + 482, + 874, + 541 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5dbd9bc2784406fd5f6ce41d0d65018dadc64176a90a060983706032eabeb524.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 823, + 542, + 880, + 563 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "sion [43] with Next3D twice. First, we fine-tune Next3D with the input video to make it learn the original geometry and texture of the avatar. Second, we fine-tune Next3D on the edited image based on the weights of the code and generator from the first fine-tuning.", + "bbox": [ + 496, + 705, + 890, + 782 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Qualitative Comparison", + "text_level": 1, + "bbox": [ + 500, + 791, + 720, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Geometry Editing. We first compare our method with baselines on geometry editing, e.g., changing the size of the eyes or the contour of the cheek. We use 2D editing tools, e.g., Photoshop and DragGAN [39], to modify the shape of various facial features. Figure. 3 shows qualitative results on avatars from INSTA [76] (the top two) and", + "bbox": [ + 496, + 810, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8956", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b90f233289b3424296c9a976ab097c0d45a1186695d1585ded262f1d4bb18c9e.jpg", + "image_caption": [ + "Source" + ], + "image_footnote": [], + "bbox": [ + 98, + 63, + 171, + 143 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bbca215fca18c5cbbe146b5d53edc3779f4e7a5ea246ee16d1bacd9035d3b7a3.jpg", + "image_caption": [ + "Eyes++" + ], + "image_footnote": [], + "bbox": [ + 181, + 63, + 256, + 143 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2a722f651cd02ed7b58d5fd25dd1fb0977a1a9439bb44fdb04ffbaec3540e725.jpg", + "image_caption": [ + "Nose++" + ], + "image_footnote": [], + "bbox": [ + 259, + 63, + 331, + 143 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/11f2fdbb504c0b0d37bfcd4b4ace6260d93a1af5e31f55992afe19d72d1ee635.jpg", + "image_caption": [ + "Face++" + ], + "image_footnote": [], + "bbox": [ + 334, + 65, + 406, + 145 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f0f487863b3cfeabcb9e4247c4c445967e9b20aa062a96e850c72935064d5c18.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 98, + 157, + 176, + 215 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1304beb1885bde3f4d9a7ca7dfec5b8fa96ff1c142769379cbf9ac54025d0380.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 183, + 157, + 256, + 215 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/efd6a7c719849eb243cf4f786f340f814a920aa7f733f736fbae3cdf577ad52d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 259, + 157, + 334, + 215 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5eded29fc1b7998dc77200c879b3f7471ef3c48a939c288e0422dbe487731404.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 336, + 157, + 410, + 215 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a3fc954fc4aed7d2eee24ebba4d068c2a33c7c309221cb7e0bf3cbdacafbb7ec.jpg", + "image_caption": [ + "Source", + "Source" + ], + "image_footnote": [], + "bbox": [ + 104, + 228, + 169, + 294 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b9e3dc8e03f02ff5d2f03e4fe9ef0942679dbcffe76ff0ccd9053d06e3c0214e.jpg", + "image_caption": [ + "Eyes++" + ], + "image_footnote": [], + "bbox": [ + 184, + 228, + 251, + 294 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7197b1c39086bb37705a0fbded774806a2513e754af84ad1b676c586137f2d1a.jpg", + "image_caption": [ + "Eyes--", + "Mouth++" + ], + "image_footnote": [], + "bbox": [ + 183, + 305, + 250, + 371 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/478bff7e3da760bd9cef87f5f41d37076622f78a5b19d90044bc6fc0ed0d8465.jpg", + "image_caption": [ + "Face—" + ], + "image_footnote": [], + "bbox": [ + 261, + 228, + 326, + 294 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ca7a3b80bc3f0171f58e6a3ba841465a1b5301fa1c19326bb7ac360b64eedf67.jpg", + "image_caption": [ + "Nose++", + "Nose++", + "2D Editing" + ], + "image_footnote": [], + "bbox": [ + 259, + 305, + 326, + 371 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e9e712a6349a7975b87c7cebe0d43030027302496e5ee87ba44d8446a222de07.jpg", + "image_caption": [ + "Face++" + ], + "image_footnote": [], + "bbox": [ + 339, + 228, + 401, + 294 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ef9338c97ff64aca72b723801fa053ab1152ba1065bf5d7e4e0bbb8b98739901.jpg", + "image_caption": [ + "Jaw--", + "Jaw--" + ], + "image_footnote": [], + "bbox": [ + 336, + 305, + 398, + 371 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bb9011ed37410fd69b9793bc984434283713980fd18fdca3222b95cfc78906dd.jpg", + "image_caption": [ + "Eyes++" + ], + "image_footnote": [], + "bbox": [ + 418, + 65, + 488, + 145 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bca0d99aa388cdd58bce78abebb3ea0cf1ecc971a08ef9c6adaba7d2d4f10c19.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 413, + 157, + 490, + 215 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ca681e99a62d0ac892a7186c8de3ee5f00e978a820fa68d74385290840c2e5f0.jpg", + "image_caption": [ + "Eyes++" + ], + "image_footnote": [], + "bbox": [ + 423, + 229, + 486, + 292 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b8997f3637d63b68b8a7826e899b88bee251f410a2a7b40ae5734a480190a727.jpg", + "image_caption": [ + "Eyes", + "Mouth++" + ], + "image_footnote": [], + "bbox": [ + 421, + 305, + 480, + 371 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/230766ffce13445e1c592c48d7dd848e97c543427124fb798e4c5f1b60bc7850.jpg", + "image_caption": [ + "Nose++" + ], + "image_footnote": [], + "bbox": [ + 493, + 65, + 563, + 145 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7232a9d916ef95bdabd8f8a74053b12449a71836f43b80907fa46538b2f2d258.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 157, + 566, + 215 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e79f0fd28e428d35226fb4860bcccc720a4a9a5dc583dd015027b6a37277b20d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 566, + 66, + 638, + 215 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Face++", + "bbox": [ + 584, + 218, + 624, + 228 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f26317241bbdd62fa66d7a8deba6a4c7acf22955db5a06069d7af6d1f28aebbb.jpg", + "image_caption": [ + "Face--" + ], + "image_footnote": [], + "bbox": [ + 500, + 229, + 560, + 294 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/23bae771951b2c1798856ab0617da82f1d23c4fec4095f38e7955573d719acb2.jpg", + "image_caption": [ + "Nose++", + "Nose++" + ], + "image_footnote": [], + "bbox": [ + 496, + 305, + 555, + 371 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6e7ddb7187d0d19303cdf76649ce0ad12ce1878c3027f02d6bf5315636df2588.jpg", + "image_caption": [ + "Face++" + ], + "image_footnote": [], + "bbox": [ + 573, + 229, + 635, + 294 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/afb64723be416433deff85c3a4665c6bd1f542fb0ec4f2a9cc66fda6d329fd28.jpg", + "image_caption": [ + "Jaw", + "Jaw" + ], + "image_footnote": [], + "bbox": [ + 571, + 305, + 630, + 371 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6e24651350dee56a36acf41eed355e22a704a1be2906b303ae65b268ac9faebc.jpg", + "image_caption": [ + "Eyes++" + ], + "image_footnote": [], + "bbox": [ + 656, + 66, + 723, + 145 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d0afbee0d8c01e1e0d6ba342c2c77f189db98cafe07124cd87a9fbf4c9c31bc8.jpg", + "image_caption": [ + "Nose++" + ], + "image_footnote": [], + "bbox": [ + 728, + 66, + 795, + 145 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ce2b50b1fed1ee519436a36d8dc0e293053d0d6bd9fc898d83a09a9389e69746.jpg", + "image_caption": [ + "Face++" + ], + "image_footnote": [], + "bbox": [ + 802, + 66, + 870, + 143 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/741ece174c8853b3d206655524ff8cb1e47fec356b7f6b0d038a5f51f33faeec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 157, + 725, + 215 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/80870e6918668f9108fe218213f2bad7a3dc1773af0c1dad14816f2e4bfd4c1d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 727, + 157, + 797, + 214 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cd14ad760734a44e717880f6c6cc2f83cf676c3ab6ea6e374b5801c9eac75a20.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 800, + 157, + 874, + 214 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d5da76eca8a595452a5c7e83a8c96eaf23afafb04714f8f932270492c44c3bb9.jpg", + "image_caption": [ + "Eyes++" + ], + "image_footnote": [], + "bbox": [ + 656, + 229, + 717, + 294 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d3ced2f2959cbd833307ed98ca40876f76275fbb348555d27d6ba9ffb6ceed57.jpg", + "image_caption": [ + "Face--" + ], + "image_footnote": [], + "bbox": [ + 730, + 229, + 790, + 294 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/667ffc8d7c57b39661e47dd63b4c79a6078676d3934bd0c3663c52aafc19ec17.jpg", + "image_caption": [ + "Face++" + ], + "image_footnote": [], + "bbox": [ + 805, + 229, + 866, + 294 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/38b33b46641b59352520a050b71f1b011a91f147b55303bcbc1bfe09afff9de9.jpg", + "image_caption": [ + "Eyes", + "Mouth++", + "Rendered View 2" + ], + "image_footnote": [], + "bbox": [ + 655, + 305, + 717, + 371 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f4e33dede41e3e56575e3524a713d8893388a0f60e988117d82079771b19903c.jpg", + "image_caption": [ + "Nose++", + "Nose++" + ], + "image_footnote": [], + "bbox": [ + 730, + 305, + 792, + 371 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/da76fc00a6ee01e36bef667b915c92dcaa254ed633381314c3a159a39cc5a236.jpg", + "image_caption": [ + "Jaw--", + "Jaw" + ], + "image_footnote": [], + "bbox": [ + 803, + 305, + 867, + 371 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/02c7af284df2b8168629893e30005183707e1eafc19abef6404e32aa97f73b6c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 417, + 186, + 478 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1a31f7924d28469fe607b650ed390b34495eec10a8db058716820f4926348675.jpg", + "image_caption": [ + "Nose++", + "Hair++" + ], + "image_footnote": [], + "bbox": [ + 91, + 489, + 179, + 566 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1cf75bf3815a98bd2b46912a9e6cb75c9bc87a62adc7af0cd7df841903dbfec4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 196, + 417, + 264, + 478 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a22ceb8d96f639e53a0ed1a2e1ccbb494d5bd0e8e143744c47000fdf8de9fd0b.jpg", + "image_caption": [ + "Rendering" + ], + "image_footnote": [], + "bbox": [ + 194, + 492, + 261, + 568 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Rendering", + "bbox": [ + 230, + 570, + 295, + 582 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/58e8c916ead852d59a7ca5c2648b45bce0c0e46562e3b7e1be13ef4a1a6d5fd9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 272, + 417, + 336, + 477 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ering", + "bbox": [ + 272, + 479, + 299, + 491 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e5a34c470e88eeff2443a2a861e733e8b673ae20804126433159de6e552c46dc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 272, + 492, + 334, + 568 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Rendering", + "bbox": [ + 230, + 570, + 295, + 582 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3a5417ce0ca53138a5b8bf3815abeb18e0aa8b55d86e66a4daed3ffc8b0718c5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 343, + 417, + 388, + 468 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ef7ad31c9a02a5c9dab557d9fafb8503df357b9163df2da4225f2f63f2a34dc9.jpg", + "image_caption": [ + "Eyes++" + ], + "image_footnote": [], + "bbox": [ + 383, + 431, + 429, + 482 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bd5fed5b3876ebcd03d8e399050945554b3a5be9adfa561e16a59f9265ebeecf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 343, + 492, + 388, + 542 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cbfd92e8c331c4dd5056e115b36e6fe125d690618614ee1c97f8526b1a8d962f.jpg", + "image_caption": [ + "Mouth++" + ], + "image_footnote": [], + "bbox": [ + 390, + 492, + 434, + 568 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ea2a0146457762a08953182b199d9af2ddd7ef22a452e89e00f7871c1cd412de.jpg", + "image_caption": [ + "Rendering" + ], + "image_footnote": [], + "bbox": [ + 434, + 417, + 500, + 479 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2f329292d92306774e6b1cf0c0645a91ebf8f09b27728936ca1a1595ef11f7f5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 439, + 492, + 504, + 568 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/35326647f75ba52cb62ca8d4d0e92ade8e1c8673062fb979a50454d465082e65.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 417, + 573, + 478 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ering", + "bbox": [ + 509, + 479, + 537, + 491 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/173c0476fffde8cb7cca357ce4246bd91fb28e70e1aea04980b3bf0d8c7496f1.jpg", + "image_caption": [ + "Rendering", + "(c) INSTA & Next3D Geometry Editing Results", + "Figure 4. Our geometry editing results with the drag-style 2D editing on INSTA [76], NeRFBlendshape [16], and Next3D [50] avatars." + ], + "image_footnote": [], + "bbox": [ + 500, + 492, + 576, + 570 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/92381cd1262c3da770ddd8919d31a415d9ddad5a19ffbdf35a98a126574ca716.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 578, + 422, + 651, + 481 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9c9f40b2e66d389cf05f9cb2f47cb9f0f18b20ce821eadfbdc0914bd0b706701.jpg", + "image_caption": [ + "Source", + "Source", + "Mouse-,Face-", + "Eyes++, Nose-, Hair++" + ], + "image_footnote": [], + "bbox": [ + 591, + 496, + 653, + 561 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e07791129d7142fcad928fc6c93713be12517e7fff1c476287590021ec7941fc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 655, + 422, + 727, + 481 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3f0b04bfad01a4bd5db9ef5d2be505e0c19cb96af43bac7fb2f46f855b9000a7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 653, + 496, + 720, + 561 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9f90c5f4153f2249ae183c84f8946c59068cfe25bb24c161b85ecef8953ed105.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 738, + 422, + 800, + 481 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3df63afd676cd555cd205036f5c1aab31fe77560826ac343268183e7485dfe5c.jpg", + "image_caption": [ + "Rendering", + "Rendering" + ], + "image_footnote": [], + "bbox": [ + 730, + 497, + 792, + 561 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2c119aca7181f935b6918b36f113cef2e15bd6ac3695574a3494acf4c82d4c6b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 805, + 422, + 872, + 481 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0684e4d46d041e57c89c0306991a29584c03dbcef7d86208bce1fe1ea9b1d28a.jpg", + "image_caption": [ + "ering", + "dering" + ], + "image_footnote": [], + "bbox": [ + 807, + 497, + 869, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "NeRFBlendshape [16] (the bottom two). Roop [12] fails to handle the find-grained geometry change, like hairlines, and lips. Nex3D [50] is able to successfully update the avatar based on the editing, however, changes the untouched part and causes an obvious identity shift. PVP [28] can make edits while preserving the identity, however, the magnitude of change tends to be smaller than the given image. In contrast, our method produces the desired editing effect from the edited image and preserves the multiview consistency and identity of the original face.", + "bbox": [ + 75, + 618, + 468, + 770 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We further show more geometry editing results of our method on avatars in various representations in Fig. 4. Our method supports a convenient way to adjust the size of diverse facial features, such as eyes, mouths, jaw, etc., by editing a single rendered image from the avatars. The edits on 2D images are successfully lifted onto the avatar and rendered across different viewpoints and expression. Please refer to the supplementary Sec. C.3 for detailed visualization.", + "bbox": [ + 75, + 779, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tions of modified head geometry.", + "bbox": [ + 498, + 618, + 718, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Texutre Editing. We then show our capability in texture editing. We utilize Photoshop, an online makeup app WebBeauty [1], and text-driven editing method Instructpix2pix [7] to modify the texture on 2D renderings. We show comparisons with PVP [28], Roop [12] and Next3D [50] on four distinct heads avatars (INSTA [76] for the upper three and NeRFBblendshape [16] for bottom one) in Fig. 5. Roop [12] is ineffective in transferring non-human-face-like texture, thus failing in all examples. PVP [28] only transfers partial or blurry textures, and also causes shifts across the expression and head poses. Next3D [50] successfully uplifts the texture editing in 2D images sharply. However, it still suffers from the identity shift issue in the lower two heads and a blurred pattern in the upper two heads in Fig. 5. In contrast, our method faithfully paints the complicated texture following the edited image and preserves the identity of the original avatar and consis", + "bbox": [ + 496, + 643, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8957", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/93c76c917b283731e84b3bf7e8db2622a160ae3e98a088cb642f83e46ad9ea4b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 98, + 92, + 158, + 152 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3cf60b8772dd99d5e00a0e2b714a2e6c95bdc15d6ce8270373f5775c1eab50d6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 161, + 93, + 220, + 152 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6a2b6efd56c79775607a700590ac5a5e1914bfbae6ffad9cde60200f89298f1c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 92, + 292, + 151 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/16507e58984ee4f84362d69add2b356d5cdbb7487ce1d23ba60acb5d13121634.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 92, + 352, + 151 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5dbf690b0d43d237d0804d41963e940875a70cc1e71fec4a073fd96c4a630f02.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 357, + 92, + 416, + 151 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5f407eed9fc984e8a369b44e7c274172a6118b7d79f465e0ba278f9ff6a26f94.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 419, + 92, + 475, + 151 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5569949685e0a01f72da09cb42840b7d03ab727a447ea5faa8a1a8b8cc2313b5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 478, + 92, + 539, + 151 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/870a96d8475708e0b1fef9b511f58ac496c7d68d8a93e8b2657eb0b8327cfcf8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 104, + 154, + 158, + 212 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b8a0ec6a0aae99d4ba145425c96ab06f38956b5e93d7ad767cd2b9b0ee68996c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 166, + 154, + 220, + 212 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/82b070bebc38f5049677adf569d8484ed060ce1daa225b9878342df6a445a877.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 154, + 290, + 214 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/274d3feccbfff7783ef952bb15d03656d93339cc11f338bd2f481e956786b59d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 297, + 154, + 352, + 214 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3736910badfff332d0ef1a64ee9f5b5f07fe25971bc8ec3f5f269bba8a75c553.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 357, + 154, + 413, + 214 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8f6b0fe5950c661de7baaf0854f5a3a591797b99c17d77fda0844f5b90a37997.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 416, + 154, + 470, + 214 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/00fa592bb222b620a7d2ef30e5e4a14dfb165ade6207ad3ab852bb5ecfbe1a17.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 475, + 154, + 535, + 214 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/da3856b52035ad07ab8f2f23f97db60d8a6be9c4db9061c041193480bda3bb4c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 102, + 215, + 158, + 273 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4c6537a3fa5d70e15d107dbb7c5808be7df9095fbf9cbb0153ece89c28f9a2dd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 166, + 215, + 220, + 273 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/854c79f2f36cebfe8c56e6ac71d89066d0f131709d9de4633dda669c36b5551a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 230, + 215, + 289, + 275 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/03a44820aecb534d32305107396b568443e567ebeccca15ef81a57eb4d5b2e6f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 215, + 352, + 275 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/40a09dfe949365a35d78d66ec57f0a2eebe8505cdb1b3cedc543739c52e03b2a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 356, + 215, + 415, + 275 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3dc071161141e8dd00b1e90a53b2d64f9c0a7824bbf3126ced18bf205380c7f6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 419, + 215, + 473, + 275 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e2422af44a97c889265828ae743c16a0d75fd2383ef1191c3916bc33273ec7b1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 477, + 215, + 535, + 275 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e41fef444d7908c0cbc5cee7e576dfda7f4f6e6662fb00a3198502cc6b68eb0f.jpg", + "image_caption": [ + "Source" + ], + "image_footnote": [], + "bbox": [ + 104, + 282, + 153, + 334 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/68b7b3ad5625c1d3e491ff00895803e4f226c18aed515158736b2f6d5eb7b40d.jpg", + "image_caption": [ + "2D Editing" + ], + "image_footnote": [], + "bbox": [ + 158, + 282, + 218, + 334 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/443fe59e0873cc5e8c6833a9c528e78386bfaab33c93d6ed114aeff22796b102.jpg", + "image_caption": [ + "Reference Animation" + ], + "image_footnote": [], + "bbox": [ + 232, + 277, + 289, + 337 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f10a02b46c3033b3f461b036bc7bfe28904d20f589cd981a55db39059ffdc4a9.jpg", + "image_caption": [ + "PVP" + ], + "image_footnote": [], + "bbox": [ + 294, + 277, + 352, + 337 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6c01f410daf0aa7c524a0126b7aeb87c6582df5350e211cf675a62d16a7713fc.jpg", + "image_caption": [ + "Roop" + ], + "image_footnote": [], + "bbox": [ + 356, + 277, + 413, + 337 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ce6087c99512b74f1caf5b96fdcb54815366c619a23c28fb51135d334a08fbf6.jpg", + "image_caption": [ + "Next3D" + ], + "image_footnote": [], + "bbox": [ + 419, + 282, + 473, + 337 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/94129cc35166d5abf1e6ca41942d34e597ce4f1a01432dddc92af159950105ec.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 477, + 279, + 535, + 337 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/69037e33481230528a1f2921e635a940c72097da1718e3eccdc3081d9fadacd4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 547, + 90, + 607, + 150 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1c4472994369a25d4b49ac36951319e3ec5c9d31a5fd1108fcaebc98a7ce2f21.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 612, + 90, + 671, + 150 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/419fdc01d70ab579a01c704d9c297dbfe7ecd98668dba16c3fa0a2e05f664304.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 674, + 90, + 733, + 150 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d67de9d18779d38cfd6e51fdebe8467191d75ddef42b6c71a61eedc544cc590b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 736, + 90, + 797, + 150 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/693849cfa7572135edd57630e46cb8c2527460ecc4b70ed938ba416dc67d9b77.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 800, + 90, + 861, + 150 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/270d379d3cf99c131a41185076bafd2e2e5b1d196cb43dea981f6778bdc7c310.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 550, + 154, + 607, + 215 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eaf86ffe2902db415b71848c54e3d37fb0c321425fc3f59801d9da481ac9fa6a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 611, + 154, + 669, + 215 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bedeca8294a3af5e040ffe291d01a1871e6b37e268aa137bb455c56097712ff2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 676, + 154, + 733, + 215 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a998e56ea963608f1354780e84f24a4cb7fded5693d42fb2399b7ac4a6b23c78.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 738, + 154, + 794, + 215 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3a3604e48aa3631cb242f6388bc15ba7f3f6ad7e3599665bb2c774f387807f59.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 800, + 154, + 857, + 215 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1365f99133e2e83108312a66de1441267bdef4642a87768a8b91d09594682ff1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 552, + 215, + 612, + 276 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/791fe9d51ead790c5ea66b7808dafa3129ac078a02feae920e5d23c49d749670.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 215, + 673, + 276 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d0217c83307093b2e54ce669def02d63fe37926f31c14f5bcb307a1e28cc874d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 679, + 215, + 736, + 276 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4f5e235dc0659090ee67172c561047d24efc656044a9d75829c5ca27dbea4765.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 741, + 215, + 800, + 276 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2bb2d62cd2e78353554e4f9de7efa3d166f11675bc07c2a566f07eb1756ec4ca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 803, + 215, + 866, + 276 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5ce42b0aefb7fd52b678606a1827bd9e434285c0cc3b57b8cab40c7e46939a3e.jpg", + "image_caption": [ + "Reference Animation" + ], + "image_footnote": [], + "bbox": [ + 555, + 279, + 609, + 338 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/449238b1e5b10ba421609285fd247a627168f4d10c21cb99ccc664e9dc47cf0d.jpg", + "image_caption": [ + "PVP" + ], + "image_footnote": [], + "bbox": [ + 614, + 279, + 671, + 338 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/828046c3acfee0c3ec5a73b6a232e84144d84d0b42149de9fc1bf4d86c720457.jpg", + "image_caption": [ + "Roop" + ], + "image_footnote": [], + "bbox": [ + 684, + 279, + 735, + 338 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8ecc9323bb53efe8184ae53663e39c7df5791349fd916a49bba37924ca796313.jpg", + "image_caption": [ + "Next3D" + ], + "image_footnote": [], + "bbox": [ + 740, + 279, + 800, + 338 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ba6727572b4d052774b7b5d34fca0e70fdc5b1184c972def0aa0b5f012398fc8.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 805, + 279, + 862, + 338 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1c072a5148be7a0fdee133d39d607944ffcc76662bc86df7feb7d07d50441b39.jpg", + "image_caption": [ + "Figure 5. We compare texture editing with PVP [28], Roop [12], Next3D [50] on INSTA [76] and NeRFBlendshape [16] avatars. The \"Reference Animation\" denotes the image of the original avatar under the same expression with the rendered edited view." + ], + "image_footnote": [], + "bbox": [ + 84, + 402, + 459, + 545 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2d144f5947784f4cdaaa08a753eb470a5db0c221e0ab5b7d24c1612c7250583e.jpg", + "image_caption": [ + "Figure 6. Analysis of our effectiveness with the naive baselines that can accomplish the single-view avatar editing." + ], + "image_footnote": [], + "bbox": [ + 84, + 546, + 460, + 621 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tency across multiple viewpoints and expressions.", + "bbox": [ + 75, + 667, + 406, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We show extensive texture editing results in three avatar representations in Fig. 7. Our method supports a wide range of texture editing, including global style transfer (\"Add a clown makeup\" via text-driven editing), semantic-driven editing (changing the hair color), and free-form sketch (painting on the face). Please refer to the supplementary Sec. C.1/2 for hybrid editing and face reenactment results.", + "bbox": [ + 75, + 681, + 468, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.Quantitative Comparison", + "text_level": 1, + "bbox": [ + 76, + 792, + 308, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "User Study. We conducted a user study to validate our method further quantitatively. Following the evaluation protocol of Avatarstudio [38], users are required to watch the rendered videos of different methods side by side and answer each question by picking up one of the methods. For each group of editing results, we will ask four questions the", + "bbox": [ + 75, + 810, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "same as Avatarstudio [38] on editing preservation, identity preservation, temporal consistency and overall performance in Tab. 1. We collected statistics from 30 participants in 12 groups of edit results. Results are reported in Tab. 1. Our method exhibits the best editing ability while keeping the best consistency across different facial expressions for geometry and texture editing. Moreover, our method performs the best in keeping non-edited parts untouched and making the human heads still recognizable after being edited e.g., identity preservation metric in Tab. 1. Please refer to the supplementary Sec. B.4 for more details.", + "bbox": [ + 496, + 402, + 890, + 568 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Image Identity Similarity Evaluation. Following the VoLux-GAN [51], we further evaluate the cross-view identity consistency in our edit results. We take 7 groups of geometry editing results and 7 groups of texture editing results, and render the 350 images of edited avatars with different viewpoints and expressions. We calculate the cosine similarities between each rendered image and the single-view 2D edited image and average the similarities on all rendered images as metrics. As reported in Tab. 1, our method outperforms all baselines in recovering desired editing effect and retaining identity consistency.", + "bbox": [ + 496, + 570, + 890, + 736 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Native Editing Capability of Avatar", + "text_level": 1, + "bbox": [ + 498, + 744, + 807, + 760 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we analyze the editing capability from the original avatar model and show the necessity of our design.", + "bbox": [ + 496, + 763, + 890, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison to 3DMM-based Geometry Editing. One intuitive way to support geometry editing is updating the underlying 3DMM geometry [14]. Here, we investigate this method and verify its capability. Specifically, we run state-of-the-art single image-based 3DMM reconstruction method [75], and update the 3DMM shape parameter of the avatar model with the estimated one. Note that such an ap-", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8958", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b4d2729bd274985130b2a1b9d63c175f1cd732ccc0cc4c0272d23a83936acbf6.jpg", + "image_caption": [ + "Figure 7. We show our texture editing results using the 2D editing method with text-prompt, pattern painting and makeup drawing on INSTA [76] and NeRFBlendshape [16], Next3D [50] avatars." + ], + "image_footnote": [], + "bbox": [ + 101, + 87, + 869, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "proach is only available for those avatars that use 3DMM explicitly, and we take INSTA [76] as an example and show the results in Fig. 6 (a). The 3DMM fitting tends to fail when the edited face is out-of-distribution, so the magnitude of editing may not be correct (e.g. the face shape). Changing the 3DMM parameter could also result in blurry rendering from the pre-trained avatar model, which cannot be trivially fixed even after fine-tuning the rendering decoder.", + "bbox": [ + 75, + 680, + 472, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison to Fine-tuning for Texture Editing. Texture editing could be done by fine-tuning the avatar with the one-shot edited image. We test this method on avatars from NeRFBlendShape [16] and show the results in Fig. 6 (b). While large structured changes, e.g., hair color, can be edited, detailed editing is largely ignored. Please refer to", + "bbox": [ + 75, + 809, + 472, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the supplementary Sec. C.4 for more ablation studies.", + "bbox": [ + 498, + 680, + 854, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 700, + 619, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have proposed a novel generic editing approach that allows users to edit various volumetric head avatar representations from a single image, where an expression-aware modification generator lifts the editing to the 3D avatar while maintaining consistency across multiple expression and viewpoints. As a limitation, we cannot add additional objects (e.g., hat) or modify the hairstyle as shown in our supplementary Sec. C.5, which may be improved by learning extra specialized geometry addition and hair modification generators.", + "bbox": [ + 496, + 719, + 890, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment: This work was partially supported by the NSFC (No. 62102356) and Ant Group.", + "bbox": [ + 498, + 869, + 890, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8959", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Webbar/beauty demo app. 6", + "[2] Rameen Abdal, Hsin-Ying Lee, Peihao Zhu, Mengei Chai, Aliaksandr Siarohin, Peter Wonka, and Sergey Tulyakov. 3davatargan: Bridging domains for personalized editable avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4552-4562, 2023. 2", + "[3] ShahRukh Athar, Zexiang Xu, Kalyan Sunkavalli, Eli Shechtman, and Zhixin Shu. Rignerf: Fully controllable neural 3d portraits. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 20364-20373, 2022. 1", + "[4] Ziqian Bai, Feitong Tan, Zeng Huang, Kripasindhu Sarkar, Danhang Tang, Di Qiu, Abhimitra Meka, Ruofei Du, Mingsong Dou, Sergio Orts-Escalano, et al. Learning personalized high quality volumetric head avatars from monocular rgb videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16890-16900, 2023. 1, 2, 3", + "[5] Chong Bao, Yinda Zhang, Bangbang Yang, Tianxing Fan, Zesong Yang, Hujun Bao, Guofeng Zhang, and Zhaopeng Cui. Sine: Semantic-driven image-based nef editing with prior-guided editing field. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20919-20929, 2023. 2", + "[6] V Blanz and T Vetter. A morphable model for the synthesis of 3d faces. In 26th Annual Conference on Computer Graphics and Interactive Techniques (SIGGRAPH 1999), pages 187-194. ACM Press, 1999. 1, 3", + "[7] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 2, 3, 4, 6", + "[8] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3", + "[9] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In Proceedings of the European Conference on Computer Vision, pages 333-350. Springer, 2022. 2", + "[10] Anton Cherepkov, Andrey Voynov, and Artem Babenko. Navigating the gan parameter space for semantic image editing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3671-3680, 2021. 2", + "[11] deepfakes. faceswap. https://github.com/deepfakes/faceswap, 2023. Accessed: 2023-10-10. 1, 2", + "[12] deepfakes. roop. SomdevSangwan, 2023. Accessed: 2023-10-10. 1, 2, 5, 6, 7" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690-4699, 2019. 5", + "[14] Yao Feng, Haiwen Feng, Michael J Black, and Timo Bolkart. Learning an animatable detailed 3d face model from in-the-wild images. ACM Transactions on Graphics (ToG), 40(4): 1-13, 2021. 7", + "[15] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 1, 3", + "[16] Xuan Gao, Chenglai Zhong, Jun Xiang, Yang Hong, Yudong Guo, and Juyong Zhang. Reconstructing personalized semantic facial nerf models from monocular video. ACM Transactions on Graphics (TOG), 41(6):1-12, 2022. 1, 2, 3, 5, 6, 7, 8", + "[17] Songwei Ge, Taesung Park, Jun-Yan Zhu, and Jia-Bin Huang. Expressive text-to-image generation with rich text. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7545-7556, 2023. 2", + "[18] Ayaan Haque, Matthew Tancik, Alexei A Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. arXiv preprint arXiv:2303.12789, 2023. 2, 3", + "[19] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. Advances in neural information processing systems, 33:9841-9850, 2020. 2", + "[20] Amir Hertz, Kfir Aberman, and Daniel Cohen-Or. Delta denoising score. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2328–2337, 2023. 2", + "[21] Lianghua Huang, Di Chen, Yu Liu, Yujun Shen, Deli Zhao, and Jingren Zhou.Composer: Creative and controllable image synthesis with composable conditions. arXiv preprint arXiv:2302.09778, 2023. 2", + "[22] Wentao Jiang, Si Liu, Chen Gao, Jie Cao, Ran He, Jiashi Feng, and Shuicheng Yan. Psgan: Pose and expression robust spatial-aware gan for customizable makeup transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5194-5202, 2020. 2, 4", + "[23] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2", + "[24] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 2, 4, 5", + "[25] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021. 2" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8960", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194-1, 2017. 3, 4", + "[27] Tingting Li, Ruihe Qian, Chao Dong, Si Liu, Qiong Yan, Wenwu Zhu, and Liang Lin. Beautygan: Instance-level facial makeup transfer with deep generative adversarial network. In Proceedings of the 26th ACM international conference on Multimedia, pages 645-653, 2018. 2, 4", + "[28] K-E Lin, Alex Trevithick, Keli Cheng, Michel Sarkis, Mohsen Ghafoorian, Ning Bi, Gerhard Reitmayr, and Ravi Ramamoorthi. Pvp: Personalized video prior for editable dynamic portraits using stylegan. In Computer Graphics Forum, page e14890. Wiley Online Library, 2023. 1, 2, 5, 6, 7", + "[29] Huan Ling, Karsten Kreis, Daiqing Li, Seung Wook Kim, Antonio Torralba, and Sanja Fidler. Editgan: High-precision semantic image editing. Advances in Neural Information Processing Systems, 34:16331-16345, 2021. 2", + "[30] Pengyang Ling, Lin Chen, Pan Zhang, Huaian Chen, and Yi Jin. Freedrag: Point tracking is not you need for interactive point-based image editing. arXiv preprint arXiv:2307.04684, 2023. 2, 4", + "[31] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 2", + "[32] Aryan Mikaeili, Or Perel, Mehdi Safaee, Daniel Cohen-Or, and Ali Mahdavi-Amiri. Sked: Sketch-guided text-based 3d editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14607–14619, 2023. 2, 3", + "[33] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2", + "[34] Chong Mou, Xintao Wang, Jiechong Song, Ying Shan, and Jian Zhang. Dragondiffusion: Enabling drag-style manipulation on diffusion models. arXiv preprint arXiv:2307.02421, 2023. 2", + "[35] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. arXiv preprint arXiv:2201.05989, 2022. 2", + "[36] Thao Nguyen, Anh Tuan Tran, and Minh Hoai. Lipstick ain't enough: beyond color matching for in-the-wild makeup transfer. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 13305-13314, 2021. 2, 4", + "[37] Thu Nguyen-Phuoc, Gabriel Schwartz, Yuting Ye, Stephen Lombardi, and Lei Xiao. Alteredavatar: Stylizing dynamic 3d avatars with fast style adaptation. arXiv preprint arXiv:2305.19245, 2023. 2", + "[38] Mohit Mendiratta Pan, Mohamed Elgharib, Kartik Teotia, Ayush Tewari, Vladislav Golyanik, Adam Kortylewski, Christian Theobalt, et al. Avatarstudio: Text-driven editing of 3d dynamic human head avatars. arXiv preprint arXiv:2306.00547, 2023. 2, 7" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Xingang Pan, Ayush Tewari, Thomas Leimkuhler, Lingjie Liu, Abhimitra Meka, and Christian Theobalt. Drag your gan: Interactive point-based manipulation on the generative image manifold. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2, 4, 5", + "[40] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2", + "[41] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021. 2, 5", + "[42] Ivan Perov, Daiheng Gao, Nikolay Chervoniy, Kunlin Liu, Sugasa Marangonda, Chris Umé, Mr Dpfks, Carl Shift Facenheim, Luis RP, Jian Jiang, et al. Deepfacelab: Integrated, flexible and extensible face-swapping framework. arXiv preprint arXiv:2005.05535, 2020. 1, 2", + "[43] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on graphics (TOG), 42(1):1-13, 2022. 4, 5", + "[44] Sara Fridovich-Keil and Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance Fields without Neural Networks. In CVPR, 2022. 2", + "[45] Ruizhi Shao, Jingxiang Sun, Cheng Peng, Zerong Zheng, Boyao Zhou, Hongwen Zhang, and Yebin Liu. Control4d: Dynamic portrait editing by learning 4d gan from 2d diffusion-based editor. arXiv preprint arXiv:2305.20082, 2023. 2", + "[46] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9243-9252, 2020. 2", + "[47] Yujun Shi, Chuhui Xue, Jiachun Pan, Wenqing Zhang, Vincent YF Tan, and Song Bai. Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. arXiv preprint arXiv:2306.14435, 2023. 2", + "[48] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 2", + "[49] Jingxiang Sun, Xuan Wang, Yichun Shi, Lizhen Wang, Jue Wang, and Yebin Liu. Ide-3d: Interactive disentangled editing for high-resolution 3d-aware portrait synthesis. arXiv preprint arXiv:2205.15517, 2022. 2", + "[50] Jingxiang Sun, Xuan Wang, Lizhen Wang, Xiaoyu Li, Yong Zhang, Hongwen Zhang, and Yebin Liu. Next3d: Generative neural texture rasterization for 3d-aware head avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20991-21002, 2023. 1, 2, 3, 4, 5, 6, 7, 8", + "[51] Feitong Tan, Sean Fanello, Abhimitra Meka, Sergio Orts-Escolano, Danhang Tang, Rohit Pandey, Jonathan Taylor," + ], + "bbox": [ + 501, + 92, + 890, + 902 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8961", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Ping Tan, and Yinda Zhang. Volux-gan: A generative model for 3d face synthesis with hdri relighting. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 7", + "[52] Junshu Tang, Bo Zhang, Binxin Yang, Ting Zhang, Dong Chen, Lizhuang Ma, and Fang Wen. 3dfaceshop: Explicitly controllable 3d-aware portrait generation. IEEE Transactions on Visualization and Computer Graphics, 2023. 1, 2", + "[53] Ziyan Wang, Timur Bagautdinov, Stephen Lombardi, Tomas Simon, Jason Saragih, Jessica Hodgins, and Michael Zollhofer. Learning compositional radiance fields of dynamic human heads. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5704-5713, 2021. 2", + "[54] Yue Wu, Yu Deng, Jiaolong Yang, Fangyun Wei, Qifeng Chen, and Xin Tong. Anifacegan: Animatable 3d-aware face image generation for video avatars. Advances in Neural Information Processing Systems, 35:36188-36201, 2022. 1, 2", + "[55] Weihao Xia, Yulun Zhang, Yujiu Yang, Jing-Hao Xue, Bolei Zhou, and Ming-Hsuan Yang. Gan inversion: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3121-3138, 2022. 4", + "[56] Yinghao Xu, Sida Peng, Ceyuan Yang, Yujun Shen, and Bolei Zhou. 3d-aware image synthesis via learning structural and textural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18430-18439, 2022. 2", + "[57] Yuelang Xu, Lizhen Wang, Xiaochen Zhao, Hongwen Zhang, and Yebin Liu. Avatarmav: Fast 3d head avatar reconstruction using motion-aware neural voxels. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-10, 2023. 1, 3", + "[58] Bangbang Yang, Yinda Zhang, Yinghao Xu, Yijin Li, Han Zhou, Hujun Bao, Guofeng Zhang, and Zhaopeng Cui. Learning object-compositional neural radiance field for ed-itable scene rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13779-13788, 2021. 2", + "[59] Bangbang Yang, Chong Bao, Junyi Zeng, Hujun Bao, Yinda Zhang, Zhaopeng Cui, and Guofeng Zhang. Neumesh: Learning disentangled neural mesh-based implicit field for geometry and texture editing. In European Conference on Computer Vision, pages 597-614. Springer, 2022.", + "[60] Bangbang Yang, Yinda Zhang, Yijin Li, Zhaopeng Cui, Sean Fanello, Hujun Bao, and Guofeng Zhang. Neural rendering in a room: amodal 3d understanding and free-viewpoint rendering for the closed scene composed of pre-captured objects. ACM Transactions on Graphics (TOG), 41(4):1-10, 2022. 2", + "[61] Chenyu Yang, Wanrong He, Yingqing Xu, and Yang Gao. Elegant: Exquisite and locally editable gan for makeup transfer. In European Conference on Computer Vision, pages 737-754. Springer, 2022. 2", + "[62] Xingrui Yang, Hai Li, Hongjia Zhai, Yuhang Ming, Yuqian Liu, and Guofeng Zhang. Vox-fusion: Dense tracking and mapping with voxel-based neural implicit representation. In" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "2022 IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 499-507. IEEE, 2022. 2", + "[63] Weicai Ye, Shuo Chen, Chong Bao, Hujun Bao, Marc Pollefeys, Zhaopeng Cui, and Guofeng Zhang. Intrinsicerf: Learning intrinsic neural radiance fields for editable novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 339-351, 2023. 2", + "[64] Deheng Zhang, Clara Fernandez-Labrador, and Christopher Schroers. Coarf: Controllable 3d artistic style transfer for radiance fields. In 2024 International Conference on 3D Vision (3DV). IEEE, 2024. 2", + "[65] Hao Zhang, Yanbo Xu, Tianyuan Dai, Tai Chi-Keung Tang, et al. Fdnerf: Semantics-driven face reconstruction, prompt editing and relighting with diffusion models. arXiv preprint arXiv:2306.00783, 2023. 2", + "[66] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Nefactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (TOG), 40(6):1-18, 2021. 2", + "[67] Boming Zhao, Bangbang Yang, Zhenyang Li, Zuoyue Li, Guofeng Zhang, Jiashu Zhao, Dawei Yin, Zhaopeng Cui, and Hujun Bao. Factorized and controllable neural re-rendering of outdoor scene for photo extrapolation. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1455-1464, 2022. 2", + "[68] Yufeng Zheng, Victoria Fernández Abrevaya, Marcel C Bühler, Xu Chen, Michael J Black, and Otmar Hilliges. Im avatar: Implicit morphable head avatars from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13545-13555, 2022. 1, 2, 3", + "[69] Yufeng Zheng, Wang Yifan, Gordon Wetzstein, Michael J Black, and Otmar Hilliges. Pointavatar: Deformable point-based head avatars from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21057-21067, 2023. 1, 2, 3", + "[70] Jiapeng Zhu, Yujun Shen, Deli Zhao, and Bolei Zhou. Indomain gan inversion for real image editing. In European conference on computer vision, pages 592-608. Springer, 2020. 2", + "[71] Jiapeng Zhu, Ceyuan Yang, Yujun Shen, Zifan Shi, Bo Dai, Deli Zhao, and Qifeng Chen. Linkgan: Linking gan latents to pixels for controllable image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7656-7666, 2023. 2", + "[72] Peihao Zhu, Rameen Abdal, Yipeng Qin, and Peter Wonka. Sean: Image synthesis with semantic region-adaptive normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5104-5113, 2020. 2", + "[73] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12786-12796, 2022. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8962", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[74] Jingyu Zhuang, Chen Wang, Lingjie Liu, Liang Lin, and Guanbin Li. Dreameditor: Text-driven 3d scene editing with neural fields. arXiv preprint arXiv:2306.13455, 2023. 2, 3", + "[75] Wojciech Zielonka, Timo Bolkart, and Justus Thies. Towards metrical reconstruction of human faces. In European Conference on Computer Vision, pages 250–269. Springer, 2022. 7", + "[76] Wojciech Zielonka, Timo Bolkart, and Justus Thies. Instant volumetric head avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4574-4584, 2023. 1, 2, 3, 5, 6, 7, 8", + "[77] zllrunning. face-makeup.pytorch. https://github.com/zllrunning/face-makeup.PyTorch, 2023. Accessed: 2023-10-10. 4" + ], + "bbox": [ + 78, + 90, + 468, + 289 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "8963", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2024/GeneAvatar_ Generic Expression-Aware Volumetric Head Avatar Editing from a Single Image/2d50005a-8c8d-4327-bf7a-e48503a1d824_model.json b/2024/GeneAvatar_ Generic Expression-Aware Volumetric Head Avatar Editing from a Single Image/2d50005a-8c8d-4327-bf7a-e48503a1d824_model.json new file mode 100644 index 0000000000000000000000000000000000000000..87417e5dbb320798646b9f979d7249772067f02f --- /dev/null +++ b/2024/GeneAvatar_ Generic Expression-Aware Volumetric Head Avatar Editing from a Single Image/2d50005a-8c8d-4327-bf7a-e48503a1d824_model.json @@ -0,0 +1,4998 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.809, + 0.02 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.325, + 0.017, + 0.723, + 0.031 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.294, + 0.032, + 0.754, + 0.046 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.117, + 0.085, + 0.856, + 0.131 + ], + "angle": 0, + "content": "GeneAvatar: Generic Expression-Aware Volumetric Head Avatar Editing from a Single Image" + }, + { + "type": "text", + "bbox": [ + 0.173, + 0.158, + 0.792, + 0.177 + ], + "angle": 0, + "content": "Chong Bao\\(^{1*§}\\) Yinda Zhang\\(^{2*}\\) Yuan Li\\(^{1*}\\) Xiyu Zhang\\(^{1}\\) Bangbang Yang\\(^{4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.212, + 0.177, + 0.764, + 0.195 + ], + "angle": 0, + "content": "Hujun Bao\\(^{1}\\) Marc Pollefeys\\(^{3}\\) Guofeng Zhang\\(^{1}\\) Zhaopeng Cui\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.12, + 0.194, + 0.853, + 0.213 + ], + "angle": 0, + "content": "\\(^{1}\\)State Key Lab of CAD&CG, Zhejiang University \\(^{2}\\)Google \\(^{3}\\)ETH Zürich \\(^{4}\\)ByteDance" + }, + { + "type": "image_caption", + "bbox": [ + 0.126, + 0.219, + 0.299, + 0.233 + ], + "angle": 0, + "content": "GeneAvatar + NeBShape" + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.235, + 0.346, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.408, + 0.219, + 0.556, + 0.232 + ], + "angle": 0, + "content": "GeneAvatar + INSTA" + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.232, + 0.614, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.676, + 0.219, + 0.829, + 0.232 + ], + "angle": 0, + "content": "GeneAvatar + Next3D" + }, + { + "type": "image", + "bbox": [ + 0.623, + 0.233, + 0.888, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.362, + 0.894, + 0.404 + ], + "angle": 0, + "content": "Figure 1. We propose a generic approach to edit 3D avatars in various volumetric representations (NeRFBblendShape [16], INSTA [76], Next3D [50]) from a single perspective using 2D editing methods with drag-style, text-prompt and pattern painting. Our editing results are consistent across multiple facial expression and camera viewpoints." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.41, + 0.314, + 0.426 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.429, + 0.474, + 0.749 + ], + "angle": 0, + "content": "Recently, we have witnessed the explosive growth of various volumetric representations in modeling animatable head avatars. However, due to the diversity of frameworks, there is no practical method to support high-level applications like 3D head avatar editing across different representations. In this paper, we propose a generic avatar editing approach that can be universally applied to various 3DMM-driving volumetric head avatars. To achieve this goal, we design a novel expression-aware modification generative model, which enables lift 2D editing from a single image to a consistent 3D modification field. To ensure the effectiveness of the generative modification process, we develop several techniques, including an expression-dependent modification distillation scheme to draw knowledge from the large-scale head avatar model and 2D facial texture editing tools, implicit latent space guidance to enhance model convergence, and a segmentation-based loss reweight strategy for fine-grained texture inversion. Extensive experiments demonstrate that our method delivers high-quality and consistent results across multiple expression and viewpoints. Project page: https://zju3dv.github.io/geneavatar/." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.755, + 0.21, + 0.77 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.47, + 0.857 + ], + "angle": 0, + "content": "Recently various volumetric representations [3, 4, 15, 16, 57, 68, 69, 76] have achieved remarkable success in reconstructing personalized, animatable, and photorealistic head avatars using implicit [15, 16, 57, 68, 69] or explicit [3, 4, 76] conditioning of 3D Morphable Models" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.411, + 0.892, + 0.472 + ], + "angle": 0, + "content": "(3DMM) [6]. A popular demand, once with a created avatar model, is to edit the avatar, e.g., for face shape, facial makeup, or apply artistic effects, for the downstream applications, e.g., in virtual/augmented reality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.474, + 0.893, + 0.685 + ], + "angle": 0, + "content": "Ideally, the desired editing functionality on the animatable avatar should have the following properties. (1) Adaptable: The editing method should be applicable across various volumetric avatar representations. This is particularly valuable in light of the growing diversity of avatar frameworks [16, 50, 76]. (2) User-friendly: The editing should be user-friendly and intuitive. Preferably, the editing of geometry and texture of the 3D avatar could be accomplished on a single-perspective rendered image. (3) Faithful: The editing results should be consistent across various facial expression and camera viewpoints. (4) Flexible: Both intensive editing (e.g., global appearance transfer following style prompts) and delicate local editing (e.g., dragging to enlarge eyes or ears) should be supported as illustrated in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.687, + 0.895, + 0.854 + ], + "angle": 0, + "content": "However, 3D-aware avatar editing is still underexplored in both geometry and texture. One plausible way is to perform 3D editing via animatable 3D GAN [50, 52, 54], but the editing results may not be consistently reflected when expression and camera viewpoint change. Alternatively, the editing can be done on the generated 2D video using 2D personalized StyleGAN [28]; however, the identity shift is often observed. Some face-swapping methods [11, 12, 42] are capable of substituting the face in a video with another face derived from a reference image or video; however, they do not support texture editing and local geometry editing." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.894, + 0.903 + ], + "angle": 0, + "content": "To this end, we propose GeneAvatar - a generic approach to support fine-grained 3D editing in various volumetric avatar representations from a single perspective by" + }, + { + "type": "page_footnote", + "bbox": [ + 0.094, + 0.863, + 0.254, + 0.876 + ], + "angle": 0, + "content": "*Authors contributed equally." + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.876, + 0.228, + 0.888 + ], + "angle": 0, + "content": "† Corresponding authors." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.368, + 0.9 + ], + "angle": 0, + "content": "\\(\\S\\) The work was partially done when visiting ETHZ." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.863, + 0.368, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8952" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.529 + ], + "angle": 0, + "content": "leveraging 2D editing methods, such as drag-based methods [30, 34, 39, 47], text-driven methods [7, 17, 20, 40, 41], or image editing tools like Photoshop (see Fig. 1). We adopt a novel editing framework that formulates the editing as predicting expression-aware 3D modification fields applied in the geometry and texture space of the volumetric avatars, which makes editing independent with the original representation as long as they are in parametric-driven radiance field, e.g., 3DMM-based neural avatar. Second, to ensure that the 2D image editing can be faithfully transferred into the 3D space, we propose to learn a generative model for modification fields, which produces 3DMM conditional modification fields from a compact latent space. Given the rendered avatar image and its edited counterpart, we conduct auto-decoding optimization on this generative model to search for the latent code that best explains the editing, obtaining consistent 3DMM conditional modification fields across various viewpoints and expression. Third, inspired by the spirit of learning from the pre-trained large-scale generative model [7, 18, 32, 74], we design a novel distillation scheme to learn the expression-dependent modification from a 3DMM-based GAN [50] and 2D face editing tools [22, 27, 36]. The scheme addresses the issue of insufficient real training data (i.e., avatars with a wide range of geometry and texture changes). Besides, we develop several techniques to enhance the editing effects, including the implicit latent space guidance to stabilize the initialization and convergence of learning, and a segmentation-based loss reweight strategy for fine-grained texture inversion." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.53, + 0.473, + 0.818 + ], + "angle": 0, + "content": "The contributions of our paper are summarized as follows. 1) We propose a generic avatar editing approach that can be applied to various 3DMM driving head avatars in the neural radiance field. To achieve this, we design a novel expression-aware modification generative model, which lifts the geometry and texture editing from a single image to a consistent 3D modification field. 2) To bootstrap the training of the modification generator with limited real paired training data, we design a distillation scheme to learn the expression-dependent geometry and texture modification from the large-scale head avatar generative model [50] and 2D face texture editing tools [22, 27, 36], and develop several techniques, including implicit guidance in latent space to improve training convergence, and a loss reweight strategy based on segmentation for fine-grained texture inversion. 3) Extensive experiments on various head avatar representations demonstrate that our method delivers high-quality editing results and the editing effects are consistent under different viewpoints and expression." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.822, + 0.22, + 0.837 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.472, + 0.903 + ], + "angle": 0, + "content": "2D Head Avatar Editing. The manipulation of 2D head avatars has made significant strides in recent years. Various GAN-based methods [23-25] can result in precise and high-resolution human face editing by leveraging image space se" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.332 + ], + "angle": 0, + "content": "mantic information [29, 72] or controlling latent space explorations [10, 19, 46, 70]. Some approaches [22, 27, 36, 61] focus on the task of makeup transferring by exploiting the GAN to learn the transferring ability from a large unaligned makeup and non-makeup face datasets. The drag-based GAN editing approach [39, 71] gained vast popularity due to providing a user-friendly editing way. PVP [28] uses a monocular video to fine-tune StyleGAN [23, 24] to obtain the personalized image generator and provide various editing functions. The diffusion models [7, 21] also show the capability of achieving fine-grained face editing with text prompt and other conditional input. For editing an avatar in a video, lots of face-swapping methods [11, 12, 42] have emerged to provide high-quality and properly aligned face-swapping results. However, these approaches typically suffer from multi-view consistency and identity preservation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.333, + 0.895, + 0.576 + ], + "angle": 0, + "content": "3D Head Avatar Editing. Neural Radiance Field [33] has exhibited great reconstruction and rendering qualities in SLAM [62, 73], scene editing [5, 58-60, 64] and relighting [63, 66, 67], especially promoting the emergence of many 3D avatar reconstruction [4, 16, 53, 68, 69, 76] and generation [50, 52, 54]. Some methods [2, 49, 56, 65] exploit the powerful editing ability of GAN to edit a 3D static head portrait. However, they cannot be trivially extended to the dynamic avatars. The methods [37, 38, 45] focus on style transfer of the avatar using text prompt or style image but reach a poor identity-preserving. We propose a novel 3D avatar editing approach with an expression-aware modification generative model, which can be applied to various 3DMM-based volumetric avatars and render consistent novel views with fine-grained editing across multiple viewpoints and expression while preserving identity of person." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.58, + 0.593, + 0.595 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.605, + 0.895, + 0.818 + ], + "angle": 0, + "content": "As shown in Fig. 2, given a volumetric head avatar, we edit the avatar using a single-view image and synthesize consistent novel views across multiple expression and viewpoints. To achieve this goal, we propose a novel expression-aware modification generator to generate 3D modification fields, which can be seamlessly integrated into various representations and animated with facial expression (see Sec. 3.2). Furthermore, to bootstrap the training with limited pairwise data, we propose a novel expression-aware distillation scheme to learn the expression-dependent modifications from large-scale generative models [7, 50] (see Sec. 3.3). During the editing process, given a single edited image of a 3D avatar, we perform an auto-decoding optimization to lift 2D editing effect to the 3D space (see Sec. 3.4)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.824, + 0.642, + 0.838 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.903 + ], + "angle": 0, + "content": "The current implicit volumetric representations of head avatar [16, 50, 76] are mostly built upon the NeRF [33] or its variants [8, 9, 31, 35, 44, 48]. In general, the neural architecture can be simplified as an implicit field \\(\\mathbf{F}\\) that" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8953" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.068, + 0.254, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.259, + 0.068, + 0.71, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.067, + 0.877, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.278, + 0.895, + 0.335 + ], + "angle": 0, + "content": "Figure 2. We use an expression-aware generative model that accepts a modification latent code \\(\\mathbf{z}_{g / t}\\) and 3DMM coefficients and outputs a modification field of a tri-plane structure. The modification field modifies the geometry and texture of the template avatar by deforming the sample points \\(\\mathbf{x}\\) and blending the color \\(\\mathbf{c}_o\\) with the modification color \\(\\mathbf{c}_{\\Delta}\\) respectively. We lift the 2D editing effect to 3D using an auto-decoding optimization and synthesize novel views across different expression." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.343, + 0.471, + 0.404 + ], + "angle": 0, + "content": "takes position \\(\\mathbf{x}\\) and view direction \\(\\mathbf{d}\\) as inputs and predicts the geometry \\(\\sigma\\) and the texture \\(\\mathbf{c}\\) of the avatar, i.e., \\((\\sigma, \\mathbf{c}) = \\mathbf{F}(\\mathbf{x}, \\mathbf{d})\\). Then, the volume rendering technique is used to render images as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.414, + 0.47, + 0.464 + ], + "angle": 0, + "content": "\\[\n\\hat {C} (\\boldsymbol {r}) = \\sum_ {i = 1} ^ {N} T _ {i} \\alpha_ {i} \\mathbf {c} _ {i}, \\quad T _ {i} = \\exp \\left(- \\sum_ {j = 1} ^ {i - 1} \\sigma^ {\\prime} _ {j} \\delta_ {j}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.473, + 0.471, + 0.534 + ], + "angle": 0, + "content": "where \\(\\alpha_{i} = 1 - \\exp \\left(-\\sigma^{\\prime}_{i}\\delta_{i}\\right)\\), and \\(\\delta_{i}\\) is the distance between adjacent samples along the ray. In order to animate the head avatar, 3DMM [6] is incorporated to describe the deformation implicitly [15, 16, 57, 68, 69] or explicitly [4, 76]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.543, + 0.439, + 0.558 + ], + "angle": 0, + "content": "3.2. Expression-aware Modification Generator" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.56, + 0.47, + 0.711 + ], + "angle": 0, + "content": "To enable the modification animated with the facial expressions, we follow the architecture of 3DMM-based 3D GAN [50] to build our expression-aware modification generator. As shown in Fig. 2, our generator consists of a geometry generator \\(\\mathbf{G}_{\\Delta g}\\) and a texture generator \\(\\mathbf{G}_{\\Delta t}\\). \\(\\mathbf{G}_{\\Delta g}\\) encodes the expression-dependent geometry modification by deforming the query points in the edited space to the original template space under each expression. \\(\\mathbf{G}_{\\Delta t}\\) encodes the expression-dependent modification color of query points under each expression:" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.722, + 0.47, + 0.74 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} ^ {\\prime} = \\mathbf {G} _ {\\Delta g} (\\mathbf {x}, \\mathbf {z} _ {g}, \\mathbf {v}), \\left(\\mathbf {c} _ {\\Delta}, \\beta_ {\\Delta}\\right) = \\mathbf {G} _ {\\Delta t} (\\mathbf {x}, \\mathbf {z} _ {t}, \\mathbf {v}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}\\) is the query points in the edited space, and \\(\\mathbf{x}'\\) is the deformed point in the space of original avatar under current expression \\(\\mathbf{e}\\). \\(\\mathbf{c}_{\\Delta}\\) is the modification color and \\(\\beta_{\\Delta}\\) determines the blending weights with the original color. \\(\\mathbf{z}_g,\\mathbf{z}_t\\) are the geometry and texture modification latent code respectively, where \\(\\mathbf{z}_g,\\mathbf{z}_t\\in R^{1024}\\). They control the generation of modification feature maps in the UV space. \\(\\mathbf{v}\\) is the 3DMM mesh vertices [26] that condition the current expression \\(\\mathbf{e}\\). Each mesh vertex has a neural feature that is retrieved in the modification feature map using pre-defined" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.343, + 0.892, + 0.48 + ], + "angle": 0, + "content": "UV mapping. We rasterize the vertex features to the three axis-aligned planes to generate the tri-plane feature. The modification information of the query point \\(\\mathbf{x}\\) is first collected by bilinear interpolation on the tri-plane feature and then decoded by the neural feature decoder [8]. Since the modification is defined as decoupled fields without relying on the original field, our generated modification field can be integrated into various volumetric avatar representations and be animated following the facial expression." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.488, + 0.885, + 0.504 + ], + "angle": 0, + "content": "3.3. Expression-dependent Modification Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.507, + 0.892, + 0.674 + ], + "angle": 0, + "content": "To learn the proposed expression-aware modification, we need extensive training data on avatars with a wide range of geometry and texture changes, which is hard to obtain in practice. Following the spirit of learning high-fidelity editing ability from the large-scale generative model [7, 18, 32, 74], we propose a novel expression-aware distillation scheme to deal with insufficient real training data. We leverage the ability of 3DMM-based 3D GAN [50] and 2D face texture editing tools to generate facial editing data, which encompasses a wide range of geometry and texture editing across various expression and viewpoints." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Geometry Distillation. We use the teacher 3DMM-based 3D GAN [50] \\(\\mathbf{G}_n\\) to synthesize two volumetric avatars with different geometry (an original avatar \\(\\mathbf{F}\\) and an edited avatar \\(\\mathbf{F}^{\\prime}\\)) by modifying the 3DMM shape parameter of the original avatar. This provides the paired editing data for our generator to learn how to modify the geometry of the avatar while maintaining consistency across various expression and viewpoints. Specifically, we randomly sample the latent code \\(\\mathbf{z}_n\\) in the latent space of \\(\\mathbf{G}_n\\) as well as 3DMM shape parameter \\(\\beta\\), expression parameter \\(\\psi\\), and pose parameter \\(\\theta\\). \\(\\beta\\), \\(\\psi\\) are sampled from a normal distribution whose absolute mean and standard deviation are within \\([0,1]\\). \\(\\theta\\) are a group of rotation vectors that have random directions within a unit sphere and magnitude within \\([-6,6]\\) degrees. Then, we sample an edit vector \\(\\beta_{\\Delta}\\) from a uniform distribution" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8954" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.244 + ], + "angle": 0, + "content": "\\(\\mathcal{U}(-3,3)\\) and apply it to the original shape parameter by \\(\\beta' = \\beta_{\\Delta} + \\beta\\). These hyperparameters w.r.t. 3DMM coefficients sampling are selected empirically to maintain the shape definition of the human head. Please refer to our supplementary Sec. B.2 for more details on 3DMM sampling. The original avatar \\(\\mathbf{F}\\) and paired edited avatar \\(\\mathbf{F}'\\) are generated by \\(\\mathbf{F} = \\mathbf{G}_n(\\mathbf{z}_n, \\boldsymbol{\\beta}, \\boldsymbol{\\psi}, \\boldsymbol{\\theta})\\), \\(\\mathbf{F}' = \\mathbf{G}(\\mathbf{z}_n, \\boldsymbol{\\beta}', \\boldsymbol{\\psi}, \\boldsymbol{\\theta})\\). During training, we will apply our modification generator to modify the geometry of \\(\\mathbf{F}\\) such that \\(F\\) and \\(F'\\) render the face with the same geometry." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.247, + 0.473, + 0.581 + ], + "angle": 0, + "content": "Texture Distillation. We distill the capabilities of fine-grained texture editing from 2D face editing algorithms by generating texture-modified avatar \\(\\mathbf{F}'\\) with the teacher 3DMM GAN [50] \\(\\mathbf{G}_n\\). Specifically, we sample an original avatar \\(\\mathbf{F} = \\mathbf{G}_n(\\mathbf{z}_n, \\beta, \\psi, \\theta)\\) from the teacher generator and render the image of its positive face. A segmentation-based 2D face texture editing algorithm (SBA) [77] and two makeup transfer algorithms (MTA) [22, 27, 36] are referred to in the distillation. We randomly choose one of them to edit the texture of the rendered face image. For SBA, we define several editable semantic regions of the face. A subset of these regions is selected randomly for texture painting using hues randomly sampled from the HSV color spectrum. For MTA, we randomly choose a makeup image as a reference from the open-sourced makeup dataset [22, 27, 36] and transfer the reference makeup to the rendered face image. The makeup dataset [36] contains complex makeup, such as blushes and makeup jewelry, which allow our generator to learn complicated texture editing patterns. Then, we perform the PTI inversion [43] on the texture-modified face image to lift the 2D texture editing to 3D space and obtain a texture-modified avatar \\(\\mathbf{F}'\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Modification Learning. Following the training style of StyleGAN [24], we sample a modification latent code \\(\\mathbf{z}_{g / t} \\in R^{1024}\\) in \\(\\mathcal{Z}\\) latent space for each paired editing data. We do not fully sample a 1024-dimensional modification code but sample a reduced code \\(\\bar{\\mathbf{z}}\\) from a standard normal distribution and concatenate it with the latent code \\(\\mathbf{z}_n\\) of the original avatar \\(\\mathbf{F}\\) that is sampled from the teacher model, i.e., \\(\\mathbf{z}_{g / t} = (\\mathbf{z}_n, \\bar{\\mathbf{z}}), \\mathbf{z}_n, \\bar{\\mathbf{z}} \\in R^{512}\\). This design is regarded as implicit code guidance that decently integrates knowledge from the teacher model to facilitate the model convergence. \\(\\mathbf{z}_n\\) encodes the facial appearances of avatar \\(\\mathbf{F}\\), serving as a reference to the superimposition of the modification onto the avatar \\(\\mathbf{F}\\). Note that during inference, we do not require the concatenation of latent code from the teacher model and directly optimize the full modification code from the edited image using an auto-decoding manner. Our generator generates the modification field following Eq. (2) where \\(\\mathbf{v}\\) is decoded from the 3DMM parameters of the avatar \\(\\mathbf{F}\\) using FLAME model [26] \\(\\mathbf{E}\\), i.e., \\(\\mathbf{v} = \\mathbf{E}(\\boldsymbol{\\beta}, \\boldsymbol{\\psi}, \\boldsymbol{\\theta})\\). To apply the modification field, we feed the deformed query points \\(\\mathbf{x}'\\) to the original avatar \\(\\mathbf{F}\\) to" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "obtain the density and color, and composite the color with modification color by:" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.133, + 0.892, + 0.151 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} = \\left(1 - \\beta_ {\\Delta}\\right) * \\mathbf {c} _ {o} + \\beta_ {\\Delta} * \\mathbf {c} _ {\\Delta}, \\quad (\\sigma , \\mathbf {c} _ {o}) = \\mathbf {F} \\left(\\mathbf {x} ^ {\\prime}, \\mathbf {d}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.161, + 0.894, + 0.237 + ], + "angle": 0, + "content": "Then, we perform volume rendering on the density \\(\\sigma\\) and color \\(\\mathbf{c}\\) using Eq. (1) to render the modified image \\(\\hat{I}_e\\) of avatar \\(\\mathbf{F}\\). We use the photometric loss to supervise the modified image with the rendered image \\(\\hat{I}'\\) from the edited avatar \\(\\mathbf{F}'\\) under the same camera parameters." + }, + { + "type": "equation", + "bbox": [ + 0.639, + 0.246, + 0.892, + 0.265 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\left\\| \\hat {I} _ {e} - \\hat {I} ^ {\\prime} \\right\\| _ {2} ^ {2}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.275, + 0.892, + 0.32 + ], + "angle": 0, + "content": "During training, we sample multiple viewpoints and 3DMM expression parameters \\(\\psi\\) for each editing pair \\((\\mathbf{F},\\mathbf{F}^{\\prime})\\) to enhance the spatial consistency under different expressions." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.323, + 0.794, + 0.339 + ], + "angle": 0, + "content": "3.4. Avatar Editing with Single Image" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.34, + 0.892, + 0.629 + ], + "angle": 0, + "content": "In our task, users are allowed to edit a single image with various out-of-box face editing tools, such as Photoshop, drag-based editing [30, 39], text-driven editing [7]. For each editing input, we use the auto-decoding optimization on modification code to lift 2D edits into a 3D expression-aware modification field generated by our model. This field adapts to expression and viewpoint changes and is not tied to the specific avatar representation. The StyleGAN-based generator [24, 43, 55] features a latent space mapping from \\(\\mathbf{z} \\in R^{Z}\\) in \\(\\mathcal{Z}\\) to \\(\\mathbf{w} \\in R^{W_{n} \\times W_{d}}\\) in \\(\\mathcal{W}\\), where \\(\\mathbf{w}\\) is more influential as it conditions the generator. Therefore, we perform code inversion in \\(\\mathcal{W}\\) space by randomly sampling a modification code \\(\\mathbf{w}_{g/t}\\) during editing. This code conditions a modification field \\(\\mathbf{G}_{\\Delta g/t}(\\mathbf{x}, \\mathbf{w}_{g/t}, \\mathbf{v})\\) following Eq. (2). We apply the modification field to the original avatar using Eq. (4). The modified image \\(\\hat{I}_{e}\\) is rendered following the original avatar's rendering pipeline and is encouraged to match the user-edited image \\(I_{e}\\) by optimizing \\(\\mathbf{w}_{g/t}\\) with the following loss terms:" + }, + { + "type": "equation", + "bbox": [ + 0.506, + 0.639, + 0.892, + 0.671 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {i} = \\lambda_ {1} \\mathcal {L} _ {2} (\\hat {I} _ {e}, I _ {e}) + \\lambda_ {2} \\mathcal {L} _ {\\mathrm {l p i p s}} (\\hat {I} _ {e}, I _ {e}) + \\lambda_ {3} \\mathcal {L} _ {\\mathrm {r e g}} (\\mathbf {w}, \\mathbf {w} _ {\\mathrm {a v g}}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.671, + 0.892, + 0.854 + ], + "angle": 0, + "content": "The L2 loss term \\(\\mathcal{L}_2\\) and LPIPS perceptual loss term \\(\\mathcal{L}_{\\mathrm{lpips}}\\) encourage the rendered face close to the appearance and structure of the edited face. We set \\(\\lambda_1 = 1000, \\lambda_2 = 1, \\lambda_3 = 1\\). The regularization term \\(\\mathcal{L}_{\\mathrm{reg}}\\) applied to the latent code \\(\\mathbf{w}\\) enforces alignment with the distribution of the \\(\\mathcal{W}\\) space, with \\(\\mathbf{w}_{\\mathrm{avg}}\\) representing the mean latent code computed from 1000 random samples within the \\(\\mathcal{W}\\) space. Besides, we observe that the L2 loss on the whole face will give an underfitting result for the fine-grained makeup on the eyes, eyebrows and lips. Therefore, we reweight the L2 loss on the facial features by face segmentation mask \\(M = \\{N_i | i = 1, \\dots, m\\}\\) for texture editing:" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.864, + 0.892, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {2} = \\sum_ {i = 1} ^ {m} \\sum_ {r \\in N _ {i}} \\frac {1}{| N _ {i} |} \\left\\| \\hat {C} _ {e} (r) - C _ {e} (r) \\right\\| _ {2} ^ {2}, \\tag {6}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8955" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.07, + 0.468, + 0.142 + ], + "angle": 0, + "content": "
MethodsGeometryTexture
RoopPVPNext3DOursRoopPVPNext3DOurs
Editing preservation ↑29.44%23.33%5.56%41.67%8.33%10.56%5.00%76.11%
Identity preservation ↑31.11%21.11%5.56%42.22%7.78%12.78%3.33%76.11%
Temporal consistency ↑30.00%23.89%4.44%41.67%5.56%12.78%1.67%80.00%
Overall ↑29.44%23.33%3.89%43.33%5.56%12.78%2.22%79.44%
image identity similarity ↑0.83730.87040.85470.88450.73200.84760.85000.9147
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.148, + 0.47, + 0.176 + ], + "angle": 0, + "content": "Table 1. We quantitatively compare with the PVP [28], Roop [12], Next3D [50] by user study and image identity similarity [13]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.19, + 0.473, + 0.404 + ], + "angle": 0, + "content": "where \\(\\hat{C}_e(r), C_e(r)\\) are the rendered and target color respectively, \\(N_i\\) are the rays within the \\(i\\)-th semantic part of the face. Generally, we freeze the weight of the modification generator and only optimize the modification latent code w to reach a satisfied 3D modification result. When an intense makeup or complicated pattern is painted onto the human face, we will continuously fine-tune the weight of our generator and freeze the latent code w to achieve more accurate editing results. To animate the edited avatar, users can input the new 3DMM expression parameter to the original avatar and our modification generator simultaneously. In this way, the generated modification field tightly sticks to the original avatar and presents reasonable editing results under different expression and viewpoints." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.413, + 0.21, + 0.43 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.439, + 0.47, + 0.514 + ], + "angle": 0, + "content": "In this section, we evaluate our avatar editing capability from a single perspective. One major difference with static NeRF editing is that we focus on showing how the edits are correctly lifted to 3D avatars under various expression and camera viewpoints." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.52, + 0.281, + 0.535 + ], + "angle": 0, + "content": "4.1. Dataset and Baselines" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.47, + 0.704 + ], + "angle": 0, + "content": "Datasets. We use a total of 19 neural implicit head avatars from three methods, i.e., 7 from INSTA [76], 8 from NeRF-BlendShape [16], 4 from Next3D [50], and show editing results on them. For INSTA [76] and NeRFBlendShape [16], we use the human head data (i.e., a monocular video of a head) provided by their methods to reconstruct the volumetric avatar using their respective representations. For Next3D [50], we random sample its latent space to generate volumetric avatars and perform editing on them. The evaluation datasets exhibit a substantial variation in identities, encompassing a diverse range of races, ages, and genders." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Baselines. We pick several baseline methods that can support single-view-based avatar editing. Roop [12] is a face-swapping method that can swap the human face in a video from a single reference view. To compare with Roop, we generate videos of the original avatar rendered in driving signals and single edited frames, and perform face swap. PVP [28] learns a personalized avatar image generator from a monocular video by fine-tuning the latent space of StyleGAN [24], and performs GAN-inversion style optimization [41] to edit the shape and appearance of the avatar. Next3D [50] is a 3DMM-based 3D GAN. To make a fair comparison on the same input data (i.e., a monocular video of avatar and an edited image), we perform GAN inver" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.065, + 0.574, + 0.124 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.128, + 0.56, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.181, + 0.574, + 0.197 + ], + "angle": 0, + "content": "Hair++, Eyes++, Mouse" + }, + { + "type": "image", + "bbox": [ + 0.577, + 0.064, + 0.661, + 0.121 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.577, + 0.121, + 0.693, + 0.141 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.577, + 0.142, + 0.663, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.064, + 0.755, + 0.121 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.672, + 0.121, + 0.755, + 0.141 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.065, + 0.814, + 0.121 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.762, + 0.121, + 0.817, + 0.141 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.822, + 0.065, + 0.871, + 0.121 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.822, + 0.121, + 0.877, + 0.141 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.2, + 0.562, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.263, + 0.562, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.522, + 0.326, + 0.553, + 0.342 + ], + "angle": 0, + "content": "Nose-, Lip++" + }, + { + "type": "image", + "bbox": [ + 0.581, + 0.2, + 0.633, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.577, + 0.258, + 0.636, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.2, + 0.695, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.258, + 0.694, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.2, + 0.751, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.258, + 0.757, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.794, + 0.2, + 0.875, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.762, + 0.258, + 0.819, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.825, + 0.2, + 0.876, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.825, + 0.258, + 0.881, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.342, + 0.562, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.403, + 0.563, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.47, + 0.565, + 0.479 + ], + "angle": 0, + "content": "ForeHead--" + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.28, + 0.635, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.342, + 0.637, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.644, + 0.281, + 0.697, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.644, + 0.342, + 0.698, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.281, + 0.754, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.342, + 0.758, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.281, + 0.815, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.342, + 0.817, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.825, + 0.281, + 0.876, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.824, + 0.342, + 0.875, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.401, + 0.634, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.401, + 0.696, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.401, + 0.757, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.401, + 0.819, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.825, + 0.401, + 0.882, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.422, + 0.562, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.47, + 0.565, + 0.479 + ], + "angle": 0, + "content": "ForeHead--" + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.423, + 0.636, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.481, + 0.637, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.644, + 0.424, + 0.698, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.644, + 0.481, + 0.697, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.424, + 0.759, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.483, + 0.757, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.425, + 0.817, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.823, + 0.425, + 0.876, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.577, + 0.543, + 0.634, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.583, + 0.564, + 0.636, + 0.618 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.543, + 0.696, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.646, + 0.565, + 0.694, + 0.618 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.543, + 0.757, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.565, + 0.754, + 0.618 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.769, + 0.483, + 0.817, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.543, + 0.818, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.823, + 0.483, + 0.875, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.825, + 0.543, + 0.882, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.622, + 0.57, + 0.636 + ], + "angle": 0, + "content": "2D Editing" + }, + { + "type": "image_caption", + "bbox": [ + 0.584, + 0.619, + 0.64, + 0.641 + ], + "angle": 0, + "content": "Reference Animation" + }, + { + "type": "image_caption", + "bbox": [ + 0.659, + 0.625, + 0.686, + 0.635 + ], + "angle": 0, + "content": "PVP" + }, + { + "type": "image_caption", + "bbox": [ + 0.713, + 0.625, + 0.742, + 0.635 + ], + "angle": 0, + "content": "Roop" + }, + { + "type": "image_caption", + "bbox": [ + 0.77, + 0.625, + 0.812, + 0.635 + ], + "angle": 0, + "content": "Next3D" + }, + { + "type": "image_caption", + "bbox": [ + 0.844, + 0.626, + 0.869, + 0.634 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.64, + 0.892, + 0.696 + ], + "angle": 0, + "content": "Figure 3. We compare geometry editing with PVP [28], Roop [12], Next3D [50] on INSTA [76] and NeRFBlendshape [16] avatars. The \"Reference Animation\" denotes the image of the original avatar under the same expression with the rendered edited view." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.707, + 0.892, + 0.784 + ], + "angle": 0, + "content": "sion [43] with Next3D twice. First, we fine-tune Next3D with the input video to make it learn the original geometry and texture of the avatar. Second, we fine-tune Next3D on the edited image based on the weights of the code and generator from the first fine-tuning." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.792, + 0.722, + 0.809 + ], + "angle": 0, + "content": "4.2. Qualitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Geometry Editing. We first compare our method with baselines on geometry editing, e.g., changing the size of the eyes or the contour of the cheek. We use 2D editing tools, e.g., Photoshop and DragGAN [39], to modify the shape of various facial features. Figure. 3 shows qualitative results on avatars from INSTA [76] (the top two) and" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8956" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.064, + 0.173, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.115, + 0.146, + 0.16, + 0.156 + ], + "angle": 0, + "content": "Source" + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.064, + 0.257, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.205, + 0.146, + 0.245, + 0.156 + ], + "angle": 0, + "content": "Eyes++" + }, + { + "type": "image", + "bbox": [ + 0.261, + 0.064, + 0.332, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.276, + 0.146, + 0.318, + 0.156 + ], + "angle": 0, + "content": "Nose++" + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.066, + 0.408, + 0.146 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.35, + 0.147, + 0.39, + 0.156 + ], + "angle": 0, + "content": "Face++" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.159, + 0.177, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.115, + 0.219, + 0.159, + 0.229 + ], + "angle": 0, + "content": "Source" + }, + { + "type": "image", + "bbox": [ + 0.184, + 0.159, + 0.258, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.189, + 0.219, + 0.245, + 0.229 + ], + "angle": 0, + "content": "Eyes++" + }, + { + "type": "image", + "bbox": [ + 0.261, + 0.159, + 0.335, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.277, + 0.219, + 0.311, + 0.229 + ], + "angle": 0, + "content": "Face—" + }, + { + "type": "image", + "bbox": [ + 0.337, + 0.159, + 0.411, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.35, + 0.219, + 0.39, + 0.229 + ], + "angle": 0, + "content": "Face++" + }, + { + "type": "image", + "bbox": [ + 0.106, + 0.229, + 0.17, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.115, + 0.296, + 0.159, + 0.306 + ], + "angle": 0, + "content": "Source" + }, + { + "type": "image", + "bbox": [ + 0.185, + 0.229, + 0.252, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.205, + 0.296, + 0.24, + 0.306 + ], + "angle": 0, + "content": "Eyes--" + }, + { + "type": "image", + "bbox": [ + 0.184, + 0.306, + 0.25, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.197, + 0.374, + 0.247, + 0.384 + ], + "angle": 0, + "content": "Mouth++" + }, + { + "type": "image", + "bbox": [ + 0.262, + 0.229, + 0.328, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.277, + 0.296, + 0.318, + 0.306 + ], + "angle": 0, + "content": "Nose++" + }, + { + "type": "image", + "bbox": [ + 0.261, + 0.306, + 0.327, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.276, + 0.374, + 0.318, + 0.384 + ], + "angle": 0, + "content": "Nose++" + }, + { + "type": "image_caption", + "bbox": [ + 0.262, + 0.387, + 0.33, + 0.402 + ], + "angle": 0, + "content": "2D Editing" + }, + { + "type": "image", + "bbox": [ + 0.34, + 0.229, + 0.402, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.357, + 0.296, + 0.388, + 0.306 + ], + "angle": 0, + "content": "Jaw--" + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.306, + 0.399, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.357, + 0.374, + 0.387, + 0.384 + ], + "angle": 0, + "content": "Jaw--" + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.066, + 0.49, + 0.146 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.438, + 0.147, + 0.478, + 0.156 + ], + "angle": 0, + "content": "Eyes++" + }, + { + "type": "image", + "bbox": [ + 0.415, + 0.159, + 0.491, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.438, + 0.219, + 0.478, + 0.229 + ], + "angle": 0, + "content": "Eyes++" + }, + { + "type": "image", + "bbox": [ + 0.424, + 0.23, + 0.487, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.438, + 0.296, + 0.471, + 0.306 + ], + "angle": 0, + "content": "Eyes" + }, + { + "type": "image", + "bbox": [ + 0.422, + 0.306, + 0.482, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.431, + 0.374, + 0.482, + 0.384 + ], + "angle": 0, + "content": "Mouth++" + }, + { + "type": "image", + "bbox": [ + 0.495, + 0.066, + 0.564, + 0.146 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.511, + 0.147, + 0.552, + 0.156 + ], + "angle": 0, + "content": "Nose++" + }, + { + "type": "image", + "bbox": [ + 0.492, + 0.159, + 0.567, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.219, + 0.547, + 0.229 + ], + "angle": 0, + "content": "Face--" + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.067, + 0.639, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.586, + 0.219, + 0.625, + 0.229 + ], + "angle": 0, + "content": "Face++" + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.23, + 0.562, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.511, + 0.296, + 0.552, + 0.306 + ], + "angle": 0, + "content": "Nose++" + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.306, + 0.556, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.511, + 0.374, + 0.552, + 0.383 + ], + "angle": 0, + "content": "Nose++" + }, + { + "type": "image_caption", + "bbox": [ + 0.57, + 0.219, + 0.628, + 0.229 + ], + "angle": 0, + "content": "Face++" + }, + { + "type": "image", + "bbox": [ + 0.575, + 0.23, + 0.636, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.589, + 0.296, + 0.619, + 0.306 + ], + "angle": 0, + "content": "Jaw" + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.306, + 0.631, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.59, + 0.374, + 0.619, + 0.383 + ], + "angle": 0, + "content": "Jaw" + }, + { + "type": "image", + "bbox": [ + 0.657, + 0.067, + 0.724, + 0.146 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.671, + 0.147, + 0.712, + 0.156 + ], + "angle": 0, + "content": "Eyes++" + }, + { + "type": "image", + "bbox": [ + 0.73, + 0.068, + 0.796, + 0.146 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.741, + 0.147, + 0.783, + 0.156 + ], + "angle": 0, + "content": "Nose++" + }, + { + "type": "image", + "bbox": [ + 0.803, + 0.068, + 0.872, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.819, + 0.147, + 0.858, + 0.156 + ], + "angle": 0, + "content": "Face++" + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.159, + 0.727, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.671, + 0.219, + 0.712, + 0.229 + ], + "angle": 0, + "content": "Eyes++" + }, + { + "type": "image", + "bbox": [ + 0.728, + 0.159, + 0.799, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.746, + 0.219, + 0.784, + 0.229 + ], + "angle": 0, + "content": "Face--" + }, + { + "type": "image", + "bbox": [ + 0.802, + 0.159, + 0.875, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.819, + 0.219, + 0.858, + 0.229 + ], + "angle": 0, + "content": "Face++" + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.23, + 0.718, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.671, + 0.296, + 0.706, + 0.306 + ], + "angle": 0, + "content": "Eyes" + }, + { + "type": "image", + "bbox": [ + 0.732, + 0.23, + 0.792, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.741, + 0.296, + 0.783, + 0.306 + ], + "angle": 0, + "content": "Nose++" + }, + { + "type": "image", + "bbox": [ + 0.806, + 0.23, + 0.867, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.823, + 0.296, + 0.854, + 0.306 + ], + "angle": 0, + "content": "Jaw--" + }, + { + "type": "image", + "bbox": [ + 0.656, + 0.306, + 0.718, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.374, + 0.714, + 0.383 + ], + "angle": 0, + "content": "Mouth++" + }, + { + "type": "image", + "bbox": [ + 0.732, + 0.306, + 0.793, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.741, + 0.374, + 0.783, + 0.383 + ], + "angle": 0, + "content": "Nose++" + }, + { + "type": "image", + "bbox": [ + 0.805, + 0.306, + 0.868, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.824, + 0.374, + 0.854, + 0.383 + ], + "angle": 0, + "content": "Jaw" + }, + { + "type": "image_caption", + "bbox": [ + 0.711, + 0.387, + 0.818, + 0.4 + ], + "angle": 0, + "content": "Rendered View 2" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.418, + 0.187, + 0.479 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.12, + 0.482, + 0.163, + 0.492 + ], + "angle": 0, + "content": "Nose++" + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.491, + 0.181, + 0.568 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.12, + 0.571, + 0.158, + 0.58 + ], + "angle": 0, + "content": "Hair++" + }, + { + "type": "image", + "bbox": [ + 0.197, + 0.418, + 0.266, + 0.479 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.235, + 0.481, + 0.3, + 0.493 + ], + "angle": 0, + "content": "Rendering" + }, + { + "type": "image", + "bbox": [ + 0.196, + 0.493, + 0.262, + 0.569 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.232, + 0.571, + 0.297, + 0.583 + ], + "angle": 0, + "content": "Rendering" + }, + { + "type": "image", + "bbox": [ + 0.273, + 0.418, + 0.338, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.273, + 0.481, + 0.3, + 0.492 + ], + "angle": 0, + "content": "ering" + }, + { + "type": "image", + "bbox": [ + 0.273, + 0.493, + 0.336, + 0.569 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.232, + 0.571, + 0.297, + 0.583 + ], + "angle": 0, + "content": "Rendering" + }, + { + "type": "image", + "bbox": [ + 0.344, + 0.418, + 0.389, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.385, + 0.432, + 0.431, + 0.483 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.375, + 0.482, + 0.416, + 0.492 + ], + "angle": 0, + "content": "Eyes++" + }, + { + "type": "image", + "bbox": [ + 0.344, + 0.493, + 0.389, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.391, + 0.493, + 0.435, + 0.569 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.37, + 0.571, + 0.419, + 0.581 + ], + "angle": 0, + "content": "Mouth++" + }, + { + "type": "image", + "bbox": [ + 0.435, + 0.418, + 0.5, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.473, + 0.481, + 0.538, + 0.493 + ], + "angle": 0, + "content": "Rendering" + }, + { + "type": "image", + "bbox": [ + 0.441, + 0.493, + 0.506, + 0.569 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.47, + 0.571, + 0.535, + 0.583 + ], + "angle": 0, + "content": "Rendering" + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.418, + 0.574, + 0.479 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.511, + 0.481, + 0.538, + 0.492 + ], + "angle": 0, + "content": "ering" + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.493, + 0.578, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.423, + 0.653, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.597, + 0.484, + 0.642, + 0.495 + ], + "angle": 0, + "content": "Source" + }, + { + "type": "image", + "bbox": [ + 0.592, + 0.497, + 0.654, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.597, + 0.567, + 0.642, + 0.578 + ], + "angle": 0, + "content": "Source" + }, + { + "type": "image", + "bbox": [ + 0.656, + 0.424, + 0.728, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.653, + 0.484, + 0.738, + 0.495 + ], + "angle": 0, + "content": "Mouse-,Face-" + }, + { + "type": "image", + "bbox": [ + 0.654, + 0.497, + 0.722, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.655, + 0.565, + 0.74, + 0.585 + ], + "angle": 0, + "content": "Eyes++, Nose-, Hair++" + }, + { + "type": "image", + "bbox": [ + 0.739, + 0.424, + 0.801, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.769, + 0.484, + 0.834, + 0.497 + ], + "angle": 0, + "content": "Rendering" + }, + { + "type": "image", + "bbox": [ + 0.732, + 0.498, + 0.794, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.771, + 0.567, + 0.836, + 0.58 + ], + "angle": 0, + "content": "Rendering" + }, + { + "type": "image", + "bbox": [ + 0.806, + 0.424, + 0.874, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.8, + 0.484, + 0.835, + 0.497 + ], + "angle": 0, + "content": "ering" + }, + { + "type": "image", + "bbox": [ + 0.808, + 0.498, + 0.87, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.8, + 0.567, + 0.864, + 0.58 + ], + "angle": 0, + "content": "dering" + }, + { + "type": "image_caption", + "bbox": [ + 0.58, + 0.584, + 0.873, + 0.597 + ], + "angle": 0, + "content": "(c) INSTA & Next3D Geometry Editing Results" + }, + { + "type": "image_caption", + "bbox": [ + 0.086, + 0.598, + 0.882, + 0.612 + ], + "angle": 0, + "content": "Figure 4. Our geometry editing results with the drag-style 2D editing on INSTA [76], NeRFBlendshape [16], and Next3D [50] avatars." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.619, + 0.47, + 0.771 + ], + "angle": 0, + "content": "NeRFBlendshape [16] (the bottom two). Roop [12] fails to handle the find-grained geometry change, like hairlines, and lips. Nex3D [50] is able to successfully update the avatar based on the editing, however, changes the untouched part and causes an obvious identity shift. PVP [28] can make edits while preserving the identity, however, the magnitude of change tends to be smaller than the given image. In contrast, our method produces the desired editing effect from the edited image and preserves the multiview consistency and identity of the original face." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.901 + ], + "angle": 0, + "content": "We further show more geometry editing results of our method on avatars in various representations in Fig. 4. Our method supports a convenient way to adjust the size of diverse facial features, such as eyes, mouths, jaw, etc., by editing a single rendered image from the avatars. The edits on 2D images are successfully lifted onto the avatar and rendered across different viewpoints and expression. Please refer to the supplementary Sec. C.3 for detailed visualization." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.619, + 0.72, + 0.635 + ], + "angle": 0, + "content": "tions of modified head geometry." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Texutre Editing. We then show our capability in texture editing. We utilize Photoshop, an online makeup app WebBeauty [1], and text-driven editing method Instructpix2pix [7] to modify the texture on 2D renderings. We show comparisons with PVP [28], Roop [12] and Next3D [50] on four distinct heads avatars (INSTA [76] for the upper three and NeRFBblendshape [16] for bottom one) in Fig. 5. Roop [12] is ineffective in transferring non-human-face-like texture, thus failing in all examples. PVP [28] only transfers partial or blurry textures, and also causes shifts across the expression and head poses. Next3D [50] successfully uplifts the texture editing in 2D images sharply. However, it still suffers from the identity shift issue in the lower two heads and a blurred pattern in the upper two heads in Fig. 5. In contrast, our method faithfully paints the complicated texture following the edited image and preserves the identity of the original avatar and consis" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8957" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.093, + 0.159, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.163, + 0.094, + 0.221, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.093, + 0.293, + 0.152 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.093, + 0.354, + 0.152 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.093, + 0.417, + 0.152 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.42, + 0.093, + 0.476, + 0.152 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.48, + 0.093, + 0.54, + 0.152 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.106, + 0.155, + 0.159, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.168, + 0.155, + 0.221, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.155, + 0.291, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.298, + 0.155, + 0.354, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.155, + 0.415, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.418, + 0.155, + 0.472, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.477, + 0.155, + 0.537, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.217, + 0.159, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.168, + 0.217, + 0.221, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.217, + 0.29, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.217, + 0.353, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.357, + 0.217, + 0.416, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.42, + 0.217, + 0.474, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.478, + 0.217, + 0.537, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.106, + 0.284, + 0.155, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.16, + 0.284, + 0.22, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.279, + 0.29, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.295, + 0.279, + 0.353, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.357, + 0.279, + 0.415, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.42, + 0.283, + 0.474, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.478, + 0.28, + 0.537, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.345, + 0.146, + 0.354 + ], + "angle": 0, + "content": "Source" + }, + { + "type": "image_caption", + "bbox": [ + 0.159, + 0.345, + 0.22, + 0.356 + ], + "angle": 0, + "content": "2D Editing" + }, + { + "type": "image_caption", + "bbox": [ + 0.239, + 0.34, + 0.296, + 0.362 + ], + "angle": 0, + "content": "Reference Animation" + }, + { + "type": "image_caption", + "bbox": [ + 0.316, + 0.345, + 0.342, + 0.356 + ], + "angle": 0, + "content": "PVP" + }, + { + "type": "image_caption", + "bbox": [ + 0.372, + 0.345, + 0.402, + 0.356 + ], + "angle": 0, + "content": "Roop" + }, + { + "type": "image_caption", + "bbox": [ + 0.425, + 0.345, + 0.469, + 0.356 + ], + "angle": 0, + "content": "Next3D" + }, + { + "type": "image_caption", + "bbox": [ + 0.494, + 0.346, + 0.522, + 0.359 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.549, + 0.091, + 0.609, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.613, + 0.091, + 0.672, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.675, + 0.091, + 0.734, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.738, + 0.092, + 0.798, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.802, + 0.091, + 0.862, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.551, + 0.155, + 0.608, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.155, + 0.671, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.677, + 0.155, + 0.735, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.739, + 0.155, + 0.795, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.802, + 0.155, + 0.859, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.553, + 0.217, + 0.614, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.616, + 0.217, + 0.674, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.68, + 0.217, + 0.738, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.742, + 0.217, + 0.801, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.805, + 0.217, + 0.867, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.28, + 0.611, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.615, + 0.28, + 0.673, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.685, + 0.28, + 0.736, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.741, + 0.28, + 0.801, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.806, + 0.28, + 0.864, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.341, + 0.613, + 0.362 + ], + "angle": 0, + "content": "Reference Animation" + }, + { + "type": "image_caption", + "bbox": [ + 0.635, + 0.345, + 0.661, + 0.356 + ], + "angle": 0, + "content": "PVP" + }, + { + "type": "image_caption", + "bbox": [ + 0.692, + 0.345, + 0.722, + 0.357 + ], + "angle": 0, + "content": "Roop" + }, + { + "type": "image_caption", + "bbox": [ + 0.749, + 0.345, + 0.793, + 0.356 + ], + "angle": 0, + "content": "Next3D" + }, + { + "type": "image_caption", + "bbox": [ + 0.819, + 0.345, + 0.848, + 0.356 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.361, + 0.892, + 0.39 + ], + "angle": 0, + "content": "Figure 5. We compare texture editing with PVP [28], Roop [12], Next3D [50] on INSTA [76] and NeRFBlendshape [16] avatars. The \"Reference Animation\" denotes the image of the original avatar under the same expression with the rendered edited view." + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.404, + 0.46, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.547, + 0.461, + 0.622 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.624, + 0.468, + 0.652 + ], + "angle": 0, + "content": "Figure 6. Analysis of our effectiveness with the naive baselines that can accomplish the single-view avatar editing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.668, + 0.407, + 0.683 + ], + "angle": 0, + "content": "tency across multiple viewpoints and expressions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.683, + 0.469, + 0.788 + ], + "angle": 0, + "content": "We show extensive texture editing results in three avatar representations in Fig. 7. Our method supports a wide range of texture editing, including global style transfer (\"Add a clown makeup\" via text-driven editing), semantic-driven editing (changing the hair color), and free-form sketch (painting on the face). Please refer to the supplementary Sec. C.1/2 for hybrid editing and face reenactment results." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.793, + 0.31, + 0.809 + ], + "angle": 0, + "content": "4.3.Quantitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.469, + 0.901 + ], + "angle": 0, + "content": "User Study. We conducted a user study to validate our method further quantitatively. Following the evaluation protocol of Avatarstudio [38], users are required to watch the rendered videos of different methods side by side and answer each question by picking up one of the methods. For each group of editing results, we will ask four questions the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.403, + 0.892, + 0.569 + ], + "angle": 0, + "content": "same as Avatarstudio [38] on editing preservation, identity preservation, temporal consistency and overall performance in Tab. 1. We collected statistics from 30 participants in 12 groups of edit results. Results are reported in Tab. 1. Our method exhibits the best editing ability while keeping the best consistency across different facial expressions for geometry and texture editing. Moreover, our method performs the best in keeping non-edited parts untouched and making the human heads still recognizable after being edited e.g., identity preservation metric in Tab. 1. Please refer to the supplementary Sec. B.4 for more details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.571, + 0.892, + 0.737 + ], + "angle": 0, + "content": "Image Identity Similarity Evaluation. Following the VoLux-GAN [51], we further evaluate the cross-view identity consistency in our edit results. We take 7 groups of geometry editing results and 7 groups of texture editing results, and render the 350 images of edited avatars with different viewpoints and expressions. We calculate the cosine similarities between each rendered image and the single-view 2D edited image and average the similarities on all rendered images as metrics. As reported in Tab. 1, our method outperforms all baselines in recovering desired editing effect and retaining identity consistency." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.745, + 0.808, + 0.761 + ], + "angle": 0, + "content": "4.4. Native Editing Capability of Avatar" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.764, + 0.892, + 0.794 + ], + "angle": 0, + "content": "In this section, we analyze the editing capability from the original avatar model and show the necessity of our design." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Comparison to 3DMM-based Geometry Editing. One intuitive way to support geometry editing is updating the underlying 3DMM geometry [14]. Here, we investigate this method and verify its capability. Specifically, we run state-of-the-art single image-based 3DMM reconstruction method [75], and update the 3DMM shape parameter of the avatar model with the estimated one. Note that such an ap-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8958" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.088, + 0.87, + 0.635 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.639, + 0.895, + 0.668 + ], + "angle": 0, + "content": "Figure 7. We show our texture editing results using the 2D editing method with text-prompt, pattern painting and makeup drawing on INSTA [76] and NeRFBlendshape [16], Next3D [50] avatars." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.681, + 0.473, + 0.803 + ], + "angle": 0, + "content": "proach is only available for those avatars that use 3DMM explicitly, and we take INSTA [76] as an example and show the results in Fig. 6 (a). The 3DMM fitting tends to fail when the edited face is out-of-distribution, so the magnitude of editing may not be correct (e.g. the face shape). Changing the 3DMM parameter could also result in blurry rendering from the pre-trained avatar model, which cannot be trivially fixed even after fine-tuning the rendering decoder." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Comparison to Fine-tuning for Texture Editing. Texture editing could be done by fine-tuning the avatar with the one-shot edited image. We test this method on avatars from NeRFBlendShape [16] and show the results in Fig. 6 (b). While large structured changes, e.g., hair color, can be edited, detailed editing is largely ignored. Please refer to" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.681, + 0.855, + 0.697 + ], + "angle": 0, + "content": "the supplementary Sec. C.4 for more ablation studies." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.701, + 0.62, + 0.716 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.87 + ], + "angle": 0, + "content": "We have proposed a novel generic editing approach that allows users to edit various volumetric head avatar representations from a single image, where an expression-aware modification generator lifts the editing to the 3D avatar while maintaining consistency across multiple expression and viewpoints. As a limitation, we cannot add additional objects (e.g., hat) or modify the hairstyle as shown in our supplementary Sec. C.5, which may be improved by learning extra specialized geometry addition and hair modification generators." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Acknowledgment: This work was partially supported by the NSFC (No. 62102356) and Ant Group." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8959" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.275, + 0.129 + ], + "angle": 0, + "content": "[1] Webbar/beauty demo app. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.131, + 0.47, + 0.214 + ], + "angle": 0, + "content": "[2] Rameen Abdal, Hsin-Ying Lee, Peihao Zhu, Mengei Chai, Aliaksandr Siarohin, Peter Wonka, and Sergey Tulyakov. 3davatargan: Bridging domains for personalized editable avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4552-4562, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.216, + 0.471, + 0.285 + ], + "angle": 0, + "content": "[3] ShahRukh Athar, Zexiang Xu, Kalyan Sunkavalli, Eli Shechtman, and Zhixin Shu. Rignerf: Fully controllable neural 3d portraits. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 20364-20373, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.288, + 0.471, + 0.384 + ], + "angle": 0, + "content": "[4] Ziqian Bai, Feitong Tan, Zeng Huang, Kripasindhu Sarkar, Danhang Tang, Di Qiu, Abhimitra Meka, Ruofei Du, Mingsong Dou, Sergio Orts-Escalano, et al. Learning personalized high quality volumetric head avatars from monocular rgb videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16890-16900, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.387, + 0.471, + 0.47 + ], + "angle": 0, + "content": "[5] Chong Bao, Yinda Zhang, Bangbang Yang, Tianxing Fan, Zesong Yang, Hujun Bao, Guofeng Zhang, and Zhaopeng Cui. Sine: Semantic-driven image-based nef editing with prior-guided editing field. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20919-20929, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.472, + 0.47, + 0.527 + ], + "angle": 0, + "content": "[6] V Blanz and T Vetter. A morphable model for the synthesis of 3d faces. In 26th Annual Conference on Computer Graphics and Interactive Techniques (SIGGRAPH 1999), pages 187-194. ACM Press, 1999. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.53, + 0.471, + 0.599 + ], + "angle": 0, + "content": "[7] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.601, + 0.471, + 0.698 + ], + "angle": 0, + "content": "[8] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.7, + 0.471, + 0.756 + ], + "angle": 0, + "content": "[9] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In Proceedings of the European Conference on Computer Vision, pages 333-350. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.758, + 0.47, + 0.826 + ], + "angle": 0, + "content": "[10] Anton Cherepkov, Andrey Voynov, and Artem Babenko. Navigating the gan parameter space for semantic image editing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3671-3680, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.829, + 0.47, + 0.87 + ], + "angle": 0, + "content": "[11] deepfakes. faceswap. https://github.com/deepfakes/faceswap, 2023. Accessed: 2023-10-10. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.872, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[12] deepfakes. roop. SomdevSangwan, 2023. Accessed: 2023-10-10. 1, 2, 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[13] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690-4699, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[14] Yao Feng, Haiwen Feng, Michael J Black, and Timo Bolkart. Learning an animatable detailed 3d face model from in-the-wild images. ACM Transactions on Graphics (ToG), 40(4): 1-13, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[15] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.292, + 0.892, + 0.359 + ], + "angle": 0, + "content": "[16] Xuan Gao, Chenglai Zhong, Jun Xiang, Yang Hong, Yudong Guo, and Juyong Zhang. Reconstructing personalized semantic facial nerf models from monocular video. ACM Transactions on Graphics (TOG), 41(6):1-12, 2022. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[17] Songwei Ge, Taesung Park, Jun-Yan Zhu, and Jia-Bin Huang. Expressive text-to-image generation with rich text. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7545-7556, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.42, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[18] Ayaan Haque, Matthew Tancik, Alexei A Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. arXiv preprint arXiv:2303.12789, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.892, + 0.53 + ], + "angle": 0, + "content": "[19] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. Advances in neural information processing systems, 33:9841-9850, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.533, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[20] Amir Hertz, Kfir Aberman, and Daniel Cohen-Or. Delta denoising score. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2328–2337, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.577, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[21] Lianghua Huang, Di Chen, Yu Liu, Yujun Shen, Deli Zhao, and Jingren Zhou.Composer: Creative and controllable image synthesis with composable conditions. arXiv preprint arXiv:2302.09778, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.633, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[22] Wentao Jiang, Si Liu, Chen Gao, Jie Cao, Ran He, Jiashi Feng, and Shuicheng Yan. Psgan: Pose and expression robust spatial-aware gan for customizable makeup transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5194-5202, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.718, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[23] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.775, + 0.894, + 0.844 + ], + "angle": 0, + "content": "[24] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "8960" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[26] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194-1, 2017. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.47, + 0.204 + ], + "angle": 0, + "content": "[27] Tingting Li, Ruihe Qian, Chao Dong, Si Liu, Qiong Yan, Wenwu Zhu, and Liang Lin. Beautygan: Instance-level facial makeup transfer with deep generative adversarial network. In Proceedings of the 26th ACM international conference on Multimedia, pages 645-653, 2018. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.204, + 0.471, + 0.285 + ], + "angle": 0, + "content": "[28] K-E Lin, Alex Trevithick, Keli Cheng, Michel Sarkis, Mohsen Ghafoorian, Ning Bi, Gerhard Reitmayr, and Ravi Ramamoorthi. Pvp: Personalized video prior for editable dynamic portraits using stylegan. In Computer Graphics Forum, page e14890. Wiley Online Library, 2023. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.471, + 0.343 + ], + "angle": 0, + "content": "[29] Huan Ling, Karsten Kreis, Daiqing Li, Seung Wook Kim, Antonio Torralba, and Sanja Fidler. Editgan: High-precision semantic image editing. Advances in Neural Information Processing Systems, 34:16331-16345, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.471, + 0.399 + ], + "angle": 0, + "content": "[30] Pengyang Ling, Lin Chen, Pan Zhang, Huaian Chen, and Yi Jin. Freedrag: Point tracking is not you need for interactive point-based image editing. arXiv preprint arXiv:2307.04684, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.471, + 0.453 + ], + "angle": 0, + "content": "[31] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.471, + 0.523 + ], + "angle": 0, + "content": "[32] Aryan Mikaeili, Or Perel, Mehdi Safaee, Daniel Cohen-Or, and Ali Mahdavi-Amiri. Sked: Sketch-guided text-based 3d editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14607–14619, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.471, + 0.593 + ], + "angle": 0, + "content": "[33] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.471, + 0.649 + ], + "angle": 0, + "content": "[34] Chong Mou, Xintao Wang, Jiechong Song, Ying Shan, and Jian Zhang. Dragondiffusion: Enabling drag-style manipulation on diffusion models. arXiv preprint arXiv:2307.02421, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.471, + 0.704 + ], + "angle": 0, + "content": "[35] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. arXiv preprint arXiv:2201.05989, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.471, + 0.775 + ], + "angle": 0, + "content": "[36] Thao Nguyen, Anh Tuan Tran, and Minh Hoai. Lipstick ain't enough: beyond color matching for in-the-wild makeup transfer. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 13305-13314, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.471, + 0.831 + ], + "angle": 0, + "content": "[37] Thu Nguyen-Phuoc, Gabriel Schwartz, Yuting Ye, Stephen Lombardi, and Lei Xiao. Alteredavatar: Stylizing dynamic 3d avatars with fast style adaptation. arXiv preprint arXiv:2305.19245, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[38] Mohit Mendiratta Pan, Mohamed Elgharib, Kartik Teotia, Ayush Tewari, Vladislav Golyanik, Adam Kortylewski, Christian Theobalt, et al. Avatarstudio: Text-driven editing of 3d dynamic human head avatars. arXiv preprint arXiv:2306.00547, 2023. 2, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[39] Xingang Pan, Ayush Tewari, Thomas Leimkuhler, Lingjie Liu, Abhimitra Meka, and Christian Theobalt. Drag your gan: Interactive point-based manipulation on the generative image manifold. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.22 + ], + "angle": 0, + "content": "[40] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[41] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[42] Ivan Perov, Daiheng Gao, Nikolay Chervoniy, Kunlin Liu, Sugasa Marangonda, Chris Umé, Mr Dpfks, Carl Shift Facenheim, Luis RP, Jian Jiang, et al. Deepfacelab: Integrated, flexible and extensible face-swapping framework. arXiv preprint arXiv:2005.05535, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[43] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on graphics (TOG), 42(1):1-13, 2022. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.419, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[44] Sara Fridovich-Keil and Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance Fields without Neural Networks. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[45] Ruizhi Shao, Jingxiang Sun, Cheng Peng, Zerong Zheng, Boyao Zhou, Hongwen Zhang, and Yebin Liu. Control4d: Dynamic portrait editing by learning 4d gan from 2d diffusion-based editor. arXiv preprint arXiv:2305.20082, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.547, + 0.892, + 0.603 + ], + "angle": 0, + "content": "[46] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9243-9252, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.604, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[47] Yujun Shi, Chuhui Xue, Jiachun Pan, Wenqing Zhang, Vincent YF Tan, and Song Bai. Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. arXiv preprint arXiv:2306.14435, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.66, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[48] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[49] Jingxiang Sun, Xuan Wang, Yichun Shi, Lizhen Wang, Jue Wang, and Yebin Liu. Ide-3d: Interactive disentangled editing for high-resolution 3d-aware portrait synthesis. arXiv preprint arXiv:2205.15517, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[50] Jingxiang Sun, Xuan Wang, Lizhen Wang, Xiaoyu Li, Yong Zhang, Hongwen Zhang, and Yebin Liu. Next3d: Generative neural texture rasterization for 3d-aware head avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20991-21002, 2023. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.903 + ], + "angle": 0, + "content": "[51] Feitong Tan, Sean Fanello, Abhimitra Meka, Sergio Orts-Escolano, Danhang Tang, Rohit Pandey, Jonathan Taylor," + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8961" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.469, + 0.135 + ], + "angle": 0, + "content": "Ping Tan, and Yinda Zhang. Volux-gan: A generative model for 3d face synthesis with hdri relighting. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.469, + 0.205 + ], + "angle": 0, + "content": "[52] Junshu Tang, Bo Zhang, Binxin Yang, Ting Zhang, Dong Chen, Lizhuang Ma, and Fang Wen. 3dfaceshop: Explicitly controllable 3d-aware portrait generation. IEEE Transactions on Visualization and Computer Graphics, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.469, + 0.289 + ], + "angle": 0, + "content": "[53] Ziyan Wang, Timur Bagautdinov, Stephen Lombardi, Tomas Simon, Jason Saragih, Jessica Hodgins, and Michael Zollhofer. Learning compositional radiance fields of dynamic human heads. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5704-5713, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.469, + 0.359 + ], + "angle": 0, + "content": "[54] Yue Wu, Yu Deng, Jiaolong Yang, Fangyun Wei, Qifeng Chen, and Xin Tong. Anifacegan: Animatable 3d-aware face image generation for video avatars. Advances in Neural Information Processing Systems, 35:36188-36201, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.469, + 0.417 + ], + "angle": 0, + "content": "[55] Weihao Xia, Yulun Zhang, Yujiu Yang, Jing-Hao Xue, Bolei Zhou, and Ming-Hsuan Yang. Gan inversion: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3121-3138, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.42, + 0.469, + 0.489 + ], + "angle": 0, + "content": "[56] Yinghao Xu, Sida Peng, Ceyuan Yang, Yujun Shen, and Bolei Zhou. 3d-aware image synthesis via learning structural and textural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18430-18439, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.491, + 0.469, + 0.559 + ], + "angle": 0, + "content": "[57] Yuelang Xu, Lizhen Wang, Xiaochen Zhao, Hongwen Zhang, and Yebin Liu. Avatarmav: Fast 3d head avatar reconstruction using motion-aware neural voxels. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-10, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.562, + 0.469, + 0.644 + ], + "angle": 0, + "content": "[58] Bangbang Yang, Yinda Zhang, Yinghao Xu, Yijin Li, Han Zhou, Hujun Bao, Guofeng Zhang, and Zhaopeng Cui. Learning object-compositional neural radiance field for ed-itable scene rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13779-13788, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.469, + 0.716 + ], + "angle": 0, + "content": "[59] Bangbang Yang, Chong Bao, Junyi Zeng, Hujun Bao, Yinda Zhang, Zhaopeng Cui, and Guofeng Zhang. Neumesh: Learning disentangled neural mesh-based implicit field for geometry and texture editing. In European Conference on Computer Vision, pages 597-614. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.469, + 0.799 + ], + "angle": 0, + "content": "[60] Bangbang Yang, Yinda Zhang, Yijin Li, Zhaopeng Cui, Sean Fanello, Hujun Bao, and Guofeng Zhang. Neural rendering in a room: amodal 3d understanding and free-viewpoint rendering for the closed scene composed of pre-captured objects. ACM Transactions on Graphics (TOG), 41(4):1-10, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[61] Chenyu Yang, Wanrong He, Yingqing Xu, and Yang Gao. Elegant: Exquisite and locally editable gan for makeup transfer. In European Conference on Computer Vision, pages 737-754. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[62] Xingrui Yang, Hai Li, Hongjia Zhai, Yuhang Ming, Yuqian Liu, and Guofeng Zhang. Vox-fusion: Dense tracking and mapping with voxel-based neural implicit representation. In" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "2022 IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 499-507. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.123, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[63] Weicai Ye, Shuo Chen, Chong Bao, Hujun Bao, Marc Pollefeys, Zhaopeng Cui, and Guofeng Zhang. Intrinsicerf: Learning intrinsic neural radiance fields for editable novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 339-351, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[64] Deheng Zhang, Clara Fernandez-Labrador, and Christopher Schroers. Coarf: Controllable 3d artistic style transfer for radiance fields. In 2024 International Conference on 3D Vision (3DV). IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.264, + 0.892, + 0.319 + ], + "angle": 0, + "content": "[65] Hao Zhang, Yanbo Xu, Tianyuan Dai, Tai Chi-Keung Tang, et al. Fdnerf: Semantics-driven face reconstruction, prompt editing and relighting with diffusion models. arXiv preprint arXiv:2306.00783, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.321, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[66] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Nefactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (TOG), 40(6):1-18, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.392, + 0.892, + 0.475 + ], + "angle": 0, + "content": "[67] Boming Zhao, Bangbang Yang, Zhenyang Li, Zuoyue Li, Guofeng Zhang, Jiashu Zhao, Dawei Yin, Zhaopeng Cui, and Hujun Bao. Factorized and controllable neural re-rendering of outdoor scene for photo extrapolation. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1455-1464, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.477, + 0.892, + 0.546 + ], + "angle": 0, + "content": "[68] Yufeng Zheng, Victoria Fernández Abrevaya, Marcel C Bühler, Xu Chen, Michael J Black, and Otmar Hilliges. Im avatar: Implicit morphable head avatars from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13545-13555, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.548, + 0.892, + 0.617 + ], + "angle": 0, + "content": "[69] Yufeng Zheng, Wang Yifan, Gordon Wetzstein, Michael J Black, and Otmar Hilliges. Pointavatar: Deformable point-based head avatars from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21057-21067, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.619, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[70] Jiapeng Zhu, Yujun Shen, Deli Zhao, and Bolei Zhou. Indomain gan inversion for real image editing. In European conference on computer vision, pages 592-608. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.676, + 0.892, + 0.745 + ], + "angle": 0, + "content": "[71] Jiapeng Zhu, Ceyuan Yang, Yujun Shen, Zifan Shi, Bo Dai, Deli Zhao, and Qifeng Chen. Linkgan: Linking gan latents to pixels for controllable image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7656-7666, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.747, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[72] Peihao Zhu, Rameen Abdal, Yipeng Qin, and Peter Wonka. Sean: Image synthesis with semantic region-adaptive normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5104-5113, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[73] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12786-12796, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8962" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[74] Jingyu Zhuang, Chen Wang, Lingjie Liu, Liang Lin, and Guanbin Li. Dreameditor: Text-driven 3d scene editing with neural fields. arXiv preprint arXiv:2306.13455, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.188 + ], + "angle": 0, + "content": "[75] Wojciech Zielonka, Timo Bolkart, and Justus Thies. Towards metrical reconstruction of human faces. In European Conference on Computer Vision, pages 250–269. Springer, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.47, + 0.248 + ], + "angle": 0, + "content": "[76] Wojciech Zielonka, Timo Bolkart, and Justus Thies. Instant volumetric head avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4574-4584, 2023. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.25, + 0.469, + 0.29 + ], + "angle": 0, + "content": "[77] zllrunning. face-makeup.pytorch. https://github.com/zllrunning/face-makeup.PyTorch, 2023. Accessed: 2023-10-10. 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "8963" + } + ] +] \ No newline at end of file diff --git a/2024/GeneAvatar_ Generic Expression-Aware Volumetric Head Avatar Editing from a Single Image/2d50005a-8c8d-4327-bf7a-e48503a1d824_origin.pdf b/2024/GeneAvatar_ Generic Expression-Aware Volumetric Head Avatar Editing from a Single Image/2d50005a-8c8d-4327-bf7a-e48503a1d824_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ea051167ae9a5be9618ac2ac0e60576d7b23849a --- /dev/null +++ b/2024/GeneAvatar_ Generic Expression-Aware Volumetric Head Avatar Editing from a Single Image/2d50005a-8c8d-4327-bf7a-e48503a1d824_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:82c8e47be6328f438786c55374300243583de704bddf26ff78a9d1f30911595f +size 6020188 diff --git a/2024/GeneAvatar_ Generic Expression-Aware Volumetric Head Avatar Editing from a Single Image/full.md b/2024/GeneAvatar_ Generic Expression-Aware Volumetric Head Avatar Editing from a Single Image/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8d5566a4b16b345fd3169c828849be1e214e9c46 --- /dev/null +++ b/2024/GeneAvatar_ Generic Expression-Aware Volumetric Head Avatar Editing from a Single Image/full.md @@ -0,0 +1,701 @@ +# GeneAvatar: Generic Expression-Aware Volumetric Head Avatar Editing from a Single Image + +Chong Bao $^{1*§}$ Yinda Zhang $^{2*}$ Yuan Li $^{1*}$ Xiyu Zhang $^{1}$ Bangbang Yang $^{4}$ + +Hujun Bao $^{1}$ Marc Pollefeys $^{3}$ Guofeng Zhang $^{1}$ Zhaopeng Cui $^{1\dagger}$ + +$^{1}$ State Key Lab of CAD&CG, Zhejiang University $^{2}$ Google $^{3}$ ETH Zürich $^{4}$ ByteDance + +![](images/2e0986f0a2bda4ee93e297c0c43264b7365ebb9b2b4eb81632b5107250b134b6.jpg) +GeneAvatar + NeBShape + +![](images/7d863c72e7ab48cd31396cafe0145ba7a2de47719d4017b9f45a880d8a6f4308.jpg) +GeneAvatar + INSTA +Figure 1. We propose a generic approach to edit 3D avatars in various volumetric representations (NeRFBblendShape [16], INSTA [76], Next3D [50]) from a single perspective using 2D editing methods with drag-style, text-prompt and pattern painting. Our editing results are consistent across multiple facial expression and camera viewpoints. + +![](images/c0c1a36e134b9c478fe95ae59a9424db14c2501c633a91a929b807ea926e0322.jpg) +GeneAvatar + Next3D + +# Abstract + +Recently, we have witnessed the explosive growth of various volumetric representations in modeling animatable head avatars. However, due to the diversity of frameworks, there is no practical method to support high-level applications like 3D head avatar editing across different representations. In this paper, we propose a generic avatar editing approach that can be universally applied to various 3DMM-driving volumetric head avatars. To achieve this goal, we design a novel expression-aware modification generative model, which enables lift 2D editing from a single image to a consistent 3D modification field. To ensure the effectiveness of the generative modification process, we develop several techniques, including an expression-dependent modification distillation scheme to draw knowledge from the large-scale head avatar model and 2D facial texture editing tools, implicit latent space guidance to enhance model convergence, and a segmentation-based loss reweight strategy for fine-grained texture inversion. Extensive experiments demonstrate that our method delivers high-quality and consistent results across multiple expression and viewpoints. Project page: https://zju3dv.github.io/geneavatar/. + +# 1. Introduction + +Recently various volumetric representations [3, 4, 15, 16, 57, 68, 69, 76] have achieved remarkable success in reconstructing personalized, animatable, and photorealistic head avatars using implicit [15, 16, 57, 68, 69] or explicit [3, 4, 76] conditioning of 3D Morphable Models + +(3DMM) [6]. A popular demand, once with a created avatar model, is to edit the avatar, e.g., for face shape, facial makeup, or apply artistic effects, for the downstream applications, e.g., in virtual/augmented reality. + +Ideally, the desired editing functionality on the animatable avatar should have the following properties. (1) Adaptable: The editing method should be applicable across various volumetric avatar representations. This is particularly valuable in light of the growing diversity of avatar frameworks [16, 50, 76]. (2) User-friendly: The editing should be user-friendly and intuitive. Preferably, the editing of geometry and texture of the 3D avatar could be accomplished on a single-perspective rendered image. (3) Faithful: The editing results should be consistent across various facial expression and camera viewpoints. (4) Flexible: Both intensive editing (e.g., global appearance transfer following style prompts) and delicate local editing (e.g., dragging to enlarge eyes or ears) should be supported as illustrated in Fig. 1. + +However, 3D-aware avatar editing is still underexplored in both geometry and texture. One plausible way is to perform 3D editing via animatable 3D GAN [50, 52, 54], but the editing results may not be consistently reflected when expression and camera viewpoint change. Alternatively, the editing can be done on the generated 2D video using 2D personalized StyleGAN [28]; however, the identity shift is often observed. Some face-swapping methods [11, 12, 42] are capable of substituting the face in a video with another face derived from a reference image or video; however, they do not support texture editing and local geometry editing. + +To this end, we propose GeneAvatar - a generic approach to support fine-grained 3D editing in various volumetric avatar representations from a single perspective by + +leveraging 2D editing methods, such as drag-based methods [30, 34, 39, 47], text-driven methods [7, 17, 20, 40, 41], or image editing tools like Photoshop (see Fig. 1). We adopt a novel editing framework that formulates the editing as predicting expression-aware 3D modification fields applied in the geometry and texture space of the volumetric avatars, which makes editing independent with the original representation as long as they are in parametric-driven radiance field, e.g., 3DMM-based neural avatar. Second, to ensure that the 2D image editing can be faithfully transferred into the 3D space, we propose to learn a generative model for modification fields, which produces 3DMM conditional modification fields from a compact latent space. Given the rendered avatar image and its edited counterpart, we conduct auto-decoding optimization on this generative model to search for the latent code that best explains the editing, obtaining consistent 3DMM conditional modification fields across various viewpoints and expression. Third, inspired by the spirit of learning from the pre-trained large-scale generative model [7, 18, 32, 74], we design a novel distillation scheme to learn the expression-dependent modification from a 3DMM-based GAN [50] and 2D face editing tools [22, 27, 36]. The scheme addresses the issue of insufficient real training data (i.e., avatars with a wide range of geometry and texture changes). Besides, we develop several techniques to enhance the editing effects, including the implicit latent space guidance to stabilize the initialization and convergence of learning, and a segmentation-based loss reweight strategy for fine-grained texture inversion. + +The contributions of our paper are summarized as follows. 1) We propose a generic avatar editing approach that can be applied to various 3DMM driving head avatars in the neural radiance field. To achieve this, we design a novel expression-aware modification generative model, which lifts the geometry and texture editing from a single image to a consistent 3D modification field. 2) To bootstrap the training of the modification generator with limited real paired training data, we design a distillation scheme to learn the expression-dependent geometry and texture modification from the large-scale head avatar generative model [50] and 2D face texture editing tools [22, 27, 36], and develop several techniques, including implicit guidance in latent space to improve training convergence, and a loss reweight strategy based on segmentation for fine-grained texture inversion. 3) Extensive experiments on various head avatar representations demonstrate that our method delivers high-quality editing results and the editing effects are consistent under different viewpoints and expression. + +# 2. Related Work + +2D Head Avatar Editing. The manipulation of 2D head avatars has made significant strides in recent years. Various GAN-based methods [23-25] can result in precise and high-resolution human face editing by leveraging image space se + +mantic information [29, 72] or controlling latent space explorations [10, 19, 46, 70]. Some approaches [22, 27, 36, 61] focus on the task of makeup transferring by exploiting the GAN to learn the transferring ability from a large unaligned makeup and non-makeup face datasets. The drag-based GAN editing approach [39, 71] gained vast popularity due to providing a user-friendly editing way. PVP [28] uses a monocular video to fine-tune StyleGAN [23, 24] to obtain the personalized image generator and provide various editing functions. The diffusion models [7, 21] also show the capability of achieving fine-grained face editing with text prompt and other conditional input. For editing an avatar in a video, lots of face-swapping methods [11, 12, 42] have emerged to provide high-quality and properly aligned face-swapping results. However, these approaches typically suffer from multi-view consistency and identity preservation. + +3D Head Avatar Editing. Neural Radiance Field [33] has exhibited great reconstruction and rendering qualities in SLAM [62, 73], scene editing [5, 58-60, 64] and relighting [63, 66, 67], especially promoting the emergence of many 3D avatar reconstruction [4, 16, 53, 68, 69, 76] and generation [50, 52, 54]. Some methods [2, 49, 56, 65] exploit the powerful editing ability of GAN to edit a 3D static head portrait. However, they cannot be trivially extended to the dynamic avatars. The methods [37, 38, 45] focus on style transfer of the avatar using text prompt or style image but reach a poor identity-preserving. We propose a novel 3D avatar editing approach with an expression-aware modification generative model, which can be applied to various 3DMM-based volumetric avatars and render consistent novel views with fine-grained editing across multiple viewpoints and expression while preserving identity of person. + +# 3. Method + +As shown in Fig. 2, given a volumetric head avatar, we edit the avatar using a single-view image and synthesize consistent novel views across multiple expression and viewpoints. To achieve this goal, we propose a novel expression-aware modification generator to generate 3D modification fields, which can be seamlessly integrated into various representations and animated with facial expression (see Sec. 3.2). Furthermore, to bootstrap the training with limited pairwise data, we propose a novel expression-aware distillation scheme to learn the expression-dependent modifications from large-scale generative models [7, 50] (see Sec. 3.3). During the editing process, given a single edited image of a 3D avatar, we perform an auto-decoding optimization to lift 2D editing effect to the 3D space (see Sec. 3.4). + +# 3.1. Preliminaries + +The current implicit volumetric representations of head avatar [16, 50, 76] are mostly built upon the NeRF [33] or its variants [8, 9, 31, 35, 44, 48]. In general, the neural architecture can be simplified as an implicit field $\mathbf{F}$ that + +![](images/f924cb4273c9bd6a9b02ba279b7b66597d0d3146b5b1a78fbc4dc60523bdb6cb.jpg) +Figure 2. We use an expression-aware generative model that accepts a modification latent code $\mathbf{z}_{g / t}$ and 3DMM coefficients and outputs a modification field of a tri-plane structure. The modification field modifies the geometry and texture of the template avatar by deforming the sample points $\mathbf{x}$ and blending the color $\mathbf{c}_o$ with the modification color $\mathbf{c}_{\Delta}$ respectively. We lift the 2D editing effect to 3D using an auto-decoding optimization and synthesize novel views across different expression. + +![](images/eba85a7295be3c26abebcef3021db96d255b1c5190a1691a3b83a9a22cd42097.jpg) + +![](images/527ac188974e9bd8d25909b1767c6a81f690304dd61d83fcb1618a945ca45345.jpg) + +takes position $\mathbf{x}$ and view direction $\mathbf{d}$ as inputs and predicts the geometry $\sigma$ and the texture $\mathbf{c}$ of the avatar, i.e., $(\sigma, \mathbf{c}) = \mathbf{F}(\mathbf{x}, \mathbf{d})$ . Then, the volume rendering technique is used to render images as follows: + +$$ +\hat {C} (\boldsymbol {r}) = \sum_ {i = 1} ^ {N} T _ {i} \alpha_ {i} \mathbf {c} _ {i}, \quad T _ {i} = \exp \left(- \sum_ {j = 1} ^ {i - 1} \sigma^ {\prime} _ {j} \delta_ {j}\right), \tag {1} +$$ + +where $\alpha_{i} = 1 - \exp \left(-\sigma^{\prime}_{i}\delta_{i}\right)$ , and $\delta_{i}$ is the distance between adjacent samples along the ray. In order to animate the head avatar, 3DMM [6] is incorporated to describe the deformation implicitly [15, 16, 57, 68, 69] or explicitly [4, 76]. + +# 3.2. Expression-aware Modification Generator + +To enable the modification animated with the facial expressions, we follow the architecture of 3DMM-based 3D GAN [50] to build our expression-aware modification generator. As shown in Fig. 2, our generator consists of a geometry generator $\mathbf{G}_{\Delta g}$ and a texture generator $\mathbf{G}_{\Delta t}$ . $\mathbf{G}_{\Delta g}$ encodes the expression-dependent geometry modification by deforming the query points in the edited space to the original template space under each expression. $\mathbf{G}_{\Delta t}$ encodes the expression-dependent modification color of query points under each expression: + +$$ +\mathbf {x} ^ {\prime} = \mathbf {G} _ {\Delta g} (\mathbf {x}, \mathbf {z} _ {g}, \mathbf {v}), \left(\mathbf {c} _ {\Delta}, \beta_ {\Delta}\right) = \mathbf {G} _ {\Delta t} (\mathbf {x}, \mathbf {z} _ {t}, \mathbf {v}), \tag {2} +$$ + +where $\mathbf{x}$ is the query points in the edited space, and $\mathbf{x}'$ is the deformed point in the space of original avatar under current expression $\mathbf{e}$ . $\mathbf{c}_{\Delta}$ is the modification color and $\beta_{\Delta}$ determines the blending weights with the original color. $\mathbf{z}_g,\mathbf{z}_t$ are the geometry and texture modification latent code respectively, where $\mathbf{z}_g,\mathbf{z}_t\in R^{1024}$ . They control the generation of modification feature maps in the UV space. $\mathbf{v}$ is the 3DMM mesh vertices [26] that condition the current expression $\mathbf{e}$ . Each mesh vertex has a neural feature that is retrieved in the modification feature map using pre-defined + +UV mapping. We rasterize the vertex features to the three axis-aligned planes to generate the tri-plane feature. The modification information of the query point $\mathbf{x}$ is first collected by bilinear interpolation on the tri-plane feature and then decoded by the neural feature decoder [8]. Since the modification is defined as decoupled fields without relying on the original field, our generated modification field can be integrated into various volumetric avatar representations and be animated following the facial expression. + +# 3.3. Expression-dependent Modification Learning + +To learn the proposed expression-aware modification, we need extensive training data on avatars with a wide range of geometry and texture changes, which is hard to obtain in practice. Following the spirit of learning high-fidelity editing ability from the large-scale generative model [7, 18, 32, 74], we propose a novel expression-aware distillation scheme to deal with insufficient real training data. We leverage the ability of 3DMM-based 3D GAN [50] and 2D face texture editing tools to generate facial editing data, which encompasses a wide range of geometry and texture editing across various expression and viewpoints. + +Geometry Distillation. We use the teacher 3DMM-based 3D GAN [50] $\mathbf{G}_n$ to synthesize two volumetric avatars with different geometry (an original avatar $\mathbf{F}$ and an edited avatar $\mathbf{F}^{\prime}$ ) by modifying the 3DMM shape parameter of the original avatar. This provides the paired editing data for our generator to learn how to modify the geometry of the avatar while maintaining consistency across various expression and viewpoints. Specifically, we randomly sample the latent code $\mathbf{z}_n$ in the latent space of $\mathbf{G}_n$ as well as 3DMM shape parameter $\beta$ , expression parameter $\psi$ , and pose parameter $\theta$ . $\beta$ , $\psi$ are sampled from a normal distribution whose absolute mean and standard deviation are within $[0,1]$ . $\theta$ are a group of rotation vectors that have random directions within a unit sphere and magnitude within $[-6,6]$ degrees. Then, we sample an edit vector $\beta_{\Delta}$ from a uniform distribution + +$\mathcal{U}(-3,3)$ and apply it to the original shape parameter by $\beta' = \beta_{\Delta} + \beta$ . These hyperparameters w.r.t. 3DMM coefficients sampling are selected empirically to maintain the shape definition of the human head. Please refer to our supplementary Sec. B.2 for more details on 3DMM sampling. The original avatar $\mathbf{F}$ and paired edited avatar $\mathbf{F}'$ are generated by $\mathbf{F} = \mathbf{G}_n(\mathbf{z}_n, \boldsymbol{\beta}, \boldsymbol{\psi}, \boldsymbol{\theta})$ , $\mathbf{F}' = \mathbf{G}(\mathbf{z}_n, \boldsymbol{\beta}', \boldsymbol{\psi}, \boldsymbol{\theta})$ . During training, we will apply our modification generator to modify the geometry of $\mathbf{F}$ such that $F$ and $F'$ render the face with the same geometry. + +Texture Distillation. We distill the capabilities of fine-grained texture editing from 2D face editing algorithms by generating texture-modified avatar $\mathbf{F}'$ with the teacher 3DMM GAN [50] $\mathbf{G}_n$ . Specifically, we sample an original avatar $\mathbf{F} = \mathbf{G}_n(\mathbf{z}_n, \beta, \psi, \theta)$ from the teacher generator and render the image of its positive face. A segmentation-based 2D face texture editing algorithm (SBA) [77] and two makeup transfer algorithms (MTA) [22, 27, 36] are referred to in the distillation. We randomly choose one of them to edit the texture of the rendered face image. For SBA, we define several editable semantic regions of the face. A subset of these regions is selected randomly for texture painting using hues randomly sampled from the HSV color spectrum. For MTA, we randomly choose a makeup image as a reference from the open-sourced makeup dataset [22, 27, 36] and transfer the reference makeup to the rendered face image. The makeup dataset [36] contains complex makeup, such as blushes and makeup jewelry, which allow our generator to learn complicated texture editing patterns. Then, we perform the PTI inversion [43] on the texture-modified face image to lift the 2D texture editing to 3D space and obtain a texture-modified avatar $\mathbf{F}'$ . + +Modification Learning. Following the training style of StyleGAN [24], we sample a modification latent code $\mathbf{z}_{g / t} \in R^{1024}$ in $\mathcal{Z}$ latent space for each paired editing data. We do not fully sample a 1024-dimensional modification code but sample a reduced code $\bar{\mathbf{z}}$ from a standard normal distribution and concatenate it with the latent code $\mathbf{z}_n$ of the original avatar $\mathbf{F}$ that is sampled from the teacher model, i.e., $\mathbf{z}_{g / t} = (\mathbf{z}_n, \bar{\mathbf{z}}), \mathbf{z}_n, \bar{\mathbf{z}} \in R^{512}$ . This design is regarded as implicit code guidance that decently integrates knowledge from the teacher model to facilitate the model convergence. $\mathbf{z}_n$ encodes the facial appearances of avatar $\mathbf{F}$ , serving as a reference to the superimposition of the modification onto the avatar $\mathbf{F}$ . Note that during inference, we do not require the concatenation of latent code from the teacher model and directly optimize the full modification code from the edited image using an auto-decoding manner. Our generator generates the modification field following Eq. (2) where $\mathbf{v}$ is decoded from the 3DMM parameters of the avatar $\mathbf{F}$ using FLAME model [26] $\mathbf{E}$ , i.e., $\mathbf{v} = \mathbf{E}(\boldsymbol{\beta}, \boldsymbol{\psi}, \boldsymbol{\theta})$ . To apply the modification field, we feed the deformed query points $\mathbf{x}'$ to the original avatar $\mathbf{F}$ to + +obtain the density and color, and composite the color with modification color by: + +$$ +\mathbf {c} = \left(1 - \beta_ {\Delta}\right) * \mathbf {c} _ {o} + \beta_ {\Delta} * \mathbf {c} _ {\Delta}, \quad (\sigma , \mathbf {c} _ {o}) = \mathbf {F} \left(\mathbf {x} ^ {\prime}, \mathbf {d}\right). \tag {3} +$$ + +Then, we perform volume rendering on the density $\sigma$ and color $\mathbf{c}$ using Eq. (1) to render the modified image $\hat{I}_e$ of avatar $\mathbf{F}$ . We use the photometric loss to supervise the modified image with the rendered image $\hat{I}'$ from the edited avatar $\mathbf{F}'$ under the same camera parameters. + +$$ +\mathcal {L} = \left\| \hat {I} _ {e} - \hat {I} ^ {\prime} \right\| _ {2} ^ {2}. \tag {4} +$$ + +During training, we sample multiple viewpoints and 3DMM expression parameters $\psi$ for each editing pair $(\mathbf{F},\mathbf{F}^{\prime})$ to enhance the spatial consistency under different expressions. + +# 3.4. Avatar Editing with Single Image + +In our task, users are allowed to edit a single image with various out-of-box face editing tools, such as Photoshop, drag-based editing [30, 39], text-driven editing [7]. For each editing input, we use the auto-decoding optimization on modification code to lift 2D edits into a 3D expression-aware modification field generated by our model. This field adapts to expression and viewpoint changes and is not tied to the specific avatar representation. The StyleGAN-based generator [24, 43, 55] features a latent space mapping from $\mathbf{z} \in R^{Z}$ in $\mathcal{Z}$ to $\mathbf{w} \in R^{W_{n} \times W_{d}}$ in $\mathcal{W}$ , where $\mathbf{w}$ is more influential as it conditions the generator. Therefore, we perform code inversion in $\mathcal{W}$ space by randomly sampling a modification code $\mathbf{w}_{g/t}$ during editing. This code conditions a modification field $\mathbf{G}_{\Delta g/t}(\mathbf{x}, \mathbf{w}_{g/t}, \mathbf{v})$ following Eq. (2). We apply the modification field to the original avatar using Eq. (4). The modified image $\hat{I}_{e}$ is rendered following the original avatar's rendering pipeline and is encouraged to match the user-edited image $I_{e}$ by optimizing $\mathbf{w}_{g/t}$ with the following loss terms: + +$$ +\mathcal {L} _ {i} = \lambda_ {1} \mathcal {L} _ {2} (\hat {I} _ {e}, I _ {e}) + \lambda_ {2} \mathcal {L} _ {\mathrm {l p i p s}} (\hat {I} _ {e}, I _ {e}) + \lambda_ {3} \mathcal {L} _ {\mathrm {r e g}} (\mathbf {w}, \mathbf {w} _ {\mathrm {a v g}}), \tag {5} +$$ + +The L2 loss term $\mathcal{L}_2$ and LPIPS perceptual loss term $\mathcal{L}_{\mathrm{lpips}}$ encourage the rendered face close to the appearance and structure of the edited face. We set $\lambda_1 = 1000, \lambda_2 = 1, \lambda_3 = 1$ . The regularization term $\mathcal{L}_{\mathrm{reg}}$ applied to the latent code $\mathbf{w}$ enforces alignment with the distribution of the $\mathcal{W}$ space, with $\mathbf{w}_{\mathrm{avg}}$ representing the mean latent code computed from 1000 random samples within the $\mathcal{W}$ space. Besides, we observe that the L2 loss on the whole face will give an underfitting result for the fine-grained makeup on the eyes, eyebrows and lips. Therefore, we reweight the L2 loss on the facial features by face segmentation mask $M = \{N_i | i = 1, \dots, m\}$ for texture editing: + +$$ +\mathcal {L} _ {2} = \sum_ {i = 1} ^ {m} \sum_ {r \in N _ {i}} \frac {1}{| N _ {i} |} \left\| \hat {C} _ {e} (r) - C _ {e} (r) \right\| _ {2} ^ {2}, \tag {6} +$$ + +
MethodsGeometryTexture
RoopPVPNext3DOursRoopPVPNext3DOurs
Editing preservation ↑29.44%23.33%5.56%41.67%8.33%10.56%5.00%76.11%
Identity preservation ↑31.11%21.11%5.56%42.22%7.78%12.78%3.33%76.11%
Temporal consistency ↑30.00%23.89%4.44%41.67%5.56%12.78%1.67%80.00%
Overall ↑29.44%23.33%3.89%43.33%5.56%12.78%2.22%79.44%
image identity similarity ↑0.83730.87040.85470.88450.73200.84760.85000.9147
+ +Table 1. We quantitatively compare with the PVP [28], Roop [12], Next3D [50] by user study and image identity similarity [13]. + +where $\hat{C}_e(r), C_e(r)$ are the rendered and target color respectively, $N_i$ are the rays within the $i$ -th semantic part of the face. Generally, we freeze the weight of the modification generator and only optimize the modification latent code w to reach a satisfied 3D modification result. When an intense makeup or complicated pattern is painted onto the human face, we will continuously fine-tune the weight of our generator and freeze the latent code w to achieve more accurate editing results. To animate the edited avatar, users can input the new 3DMM expression parameter to the original avatar and our modification generator simultaneously. In this way, the generated modification field tightly sticks to the original avatar and presents reasonable editing results under different expression and viewpoints. + +# 4. Experiments + +In this section, we evaluate our avatar editing capability from a single perspective. One major difference with static NeRF editing is that we focus on showing how the edits are correctly lifted to 3D avatars under various expression and camera viewpoints. + +# 4.1. Dataset and Baselines + +Datasets. We use a total of 19 neural implicit head avatars from three methods, i.e., 7 from INSTA [76], 8 from NeRF-BlendShape [16], 4 from Next3D [50], and show editing results on them. For INSTA [76] and NeRFBlendShape [16], we use the human head data (i.e., a monocular video of a head) provided by their methods to reconstruct the volumetric avatar using their respective representations. For Next3D [50], we random sample its latent space to generate volumetric avatars and perform editing on them. The evaluation datasets exhibit a substantial variation in identities, encompassing a diverse range of races, ages, and genders. + +Baselines. We pick several baseline methods that can support single-view-based avatar editing. Roop [12] is a face-swapping method that can swap the human face in a video from a single reference view. To compare with Roop, we generate videos of the original avatar rendered in driving signals and single edited frames, and perform face swap. PVP [28] learns a personalized avatar image generator from a monocular video by fine-tuning the latent space of StyleGAN [24], and performs GAN-inversion style optimization [41] to edit the shape and appearance of the avatar. Next3D [50] is a 3DMM-based 3D GAN. To make a fair comparison on the same input data (i.e., a monocular video of avatar and an edited image), we perform GAN inver + +![](images/43bbe264c875fe4dc42c506fb6c8a978b42bcd69db9f9f7e040985c836fe9a9f.jpg) + +![](images/3b5f7267aed76fbc90f89380c0630a0aa93e792934409ac98b04c6464f4009fa.jpg) +Hair++, Eyes++, Mouse + +![](images/786fd0f0a878b9e63eab588746a0ecf80e36ab8dd8c910fb3c704666189f0bb4.jpg) + +![](images/cf11cf46baf36d4d082e737f1a829d7539e40a5435a39547ee4070fef0e1eeab.jpg) + +![](images/f1872ea41b792926f5f3629b6e6567d860f1aa245ae580afa212ad58d2f0dfc9.jpg) + +![](images/3428f2123db18035433243704f6d6d4d11708e2e0a46f9bf121a567250b8ce34.jpg) + +![](images/3582093e97877d4d5a7bd0dd3b667ba7d0db2b5279fd6ac022bcef59fcaacf12.jpg) + +![](images/b84f3ed2e81320d68371589cdebb5b812c4124e2d47711908666447b83e7b9e9.jpg) + +![](images/d6ee9dd647e17db201d638acd8edc83ebd77e3c24fe38d288159308f95b809ad.jpg) + +![](images/ddad9fc09aa57dee0b664813232dff2635803507a570c156c59cdcd5f45f23de.jpg) + +![](images/320f79daa0c6adc0b20bbd09f87a86ff090e84ee04597f21ff749b1f70595a75.jpg) + +![](images/9c4d61b461c9bc294ee5c35bae04ffb8722fccba24a2b9fd8ad8221c9c884b56.jpg) + +![](images/b172bd92bdd7480f5fb4f432d5372048e3e458268ff6519d6a3177af230f5ebc.jpg) + +![](images/cc37aae23a265628ef1e8214f5400399ac17452f086fc21d9978bf7e80bb2417.jpg) + +![](images/8d9b606642f75a4de68acafcbafaaf7bf581d2bfb4e9e6e0e6ea588cc816c69a.jpg) + +![](images/73bfa2b7e1dfb4edd5bcf8e0941fbe4fb555e89db671424ebb5b55170dd14b4d.jpg) + +![](images/5fa147eda84fe98bbdc0733ca29629021f620c9a2765335de70f117ab1eec5b2.jpg) + +![](images/15c450b3883804d63d2659e01ce3b8ca0d27e38fe587d9954da3a2efb8c61734.jpg) + +![](images/097146ef864b05f192edd0440c381450d72919bc7524c779c793d63f593cd816.jpg) + +![](images/95b827a23a7c086a6528271f59bfe83d88718c13b13dc0f9d65ed854fd8473f0.jpg) + +![](images/545ff9b4f7f404d803c96ff05aae0dcf6137497646caabd7d46427fe8295eb03.jpg) + +![](images/62424f6e28e96a5024f72d5b028da3da1241069b9eb2b2bcfb98c52f40a66356.jpg) + +![](images/a4276534daa3eb40310b271f3f47743fd7d664b00f58e13c3b74591fb8cd52f3.jpg) + +![](images/77764d4e5ce0b53bbf8b92f29ce3a0c74862dd6247e63aa924ca21229a625b63.jpg) +Nose-, Lip++ + +![](images/e343c5cfc35b8bc98da8cc93c4f712ed9a4f5ffb1e93fda5798231f45557638b.jpg) + +ForeHead-- + +![](images/e599c41ff5a5c6c4ffd7f8c148c18ba330cd054fe3a6ea520397f8cea250430b.jpg) + +![](images/433eee09d0af752d581b901ca537b87053ba59982834cb1afd541e8af7e1cc00.jpg) + +![](images/c7e387714b649051414e74ebb46e0f381cf237632e68a1711cd4ae104837647c.jpg) + +![](images/4c8496bd89347c4ac6a8187b189141f0ad25e9d0b0fc8af82257161c1df186a7.jpg) + +![](images/4bf9e90bb7e4ee5aefb9ad000f83880f90548dbe158956b074d4841bc95df597.jpg) + +![](images/0ef874137b0767838154229871914045dc291d7a89d872be87657e2574e41ccc.jpg) + +![](images/0a2152595ce87420e6934bee87e2b3eeb54938e03e4d92ba2230961217520011.jpg) + +![](images/303976365356cdcd402ee5509a97778b231b9f94a20ce7a7e573d45f1904ee43.jpg) + +![](images/5d3e0e2362c64a7f915cf5b6a9362e5f26153b12487181d30ce8818b25cf7728.jpg) + +![](images/d041762c5b6d3f7aede17747b09e0b42fae10dfc74891eb89698fc71a5641f4f.jpg) + +![](images/f1044c94474c383d609feead99c75b0dbc0c4f8e56c933326cbaa7c123f0d0cc.jpg) + +![](images/137be98d9d39f9494c4ff6735b8f9e96ccecc2f250901dc8e4c4baf5ab1faae9.jpg) + +![](images/62b9861c4201e18303d031ef72fd9ca9664776cb3f80297a8879638ce7beac79.jpg) + +![](images/186b222fd37d76cabe5ba45a4f44f43d7ec07d595432c00d32b7db8325f97454.jpg) + +![](images/864e9deda6c8595f366462ba94874f501538d386551e6e8189945a824fb97f13.jpg) + +ForeHead-- + +![](images/e8466da5d17f01e6573589d016afa53107bbb8b7ddf94167d77cbdb6a141b79a.jpg) + +![](images/6119cb31199c31852d9ecddbb8275798e5b61c478bdb6d7e59bd2e6aad691475.jpg) + +![](images/78fb73bac479c0a577c1a4689bdd01c7b21ac0d1f1b61cdf075f338a26b8c244.jpg) + +![](images/e32b9abf2967bef31e638841f549e659f85dbcbe383dc863508f74edbc581560.jpg) + +![](images/b5190ed077d108b883734f4cb89a49043dfa789174838bd18057d1074e5fe878.jpg) + +![](images/e31bdf143f5a053b03cf5f5194beb4b2adceee34721a0b38251a7cb80d38bae5.jpg) + +![](images/bc3ecca9b0e87ddfb2ff4e88ba23117a2302de9ad39402887453bae38bd15010.jpg) + +![](images/0e020db4fc3fcec971db7d04b39a60adf2935d1d546d66f499214514822c3f21.jpg) + +![](images/a4115f15bb39d540bed289fd1895e05c190a30d46de764b516f38a0731090895.jpg) + +![](images/1d4e24d3f38814b6b71574a413b355b5264d28b43d137e35a4964c45b814c2b6.jpg) +2D Editing +Reference Animation +Figure 3. We compare geometry editing with PVP [28], Roop [12], Next3D [50] on INSTA [76] and NeRFBlendshape [16] avatars. The "Reference Animation" denotes the image of the original avatar under the same expression with the rendered edited view. + +![](images/e790a4536369185f73ebdebc5f0426bf2622c17a280a93ebb85fb9a7437dac1b.jpg) + +![](images/f8b85e621ee8c72b2701d193d5b3253e043a529ed520955369ddeb50d53f1031.jpg) +PVP + +![](images/34ddfcb176d2747a0a68641541ce75f078962dc5082bcfb73be22f4d8dc03b61.jpg) + +![](images/a653d4bbd77d58931dcadf6deba500da701bfd11db45fc7eb314d141c5a641c6.jpg) +Roop +Next3D + +![](images/d6697f8dbf19f3dfbb45a456787d167f8012a432afa93307f33601d781409d5e.jpg) + +![](images/f2455b140aa55f770aa54608605506466a3354d0e990ca92987987f53727990e.jpg) + +![](images/c7fa26a7185bbf1e5f23cdc73c2a0c3c826e12c5c067789d011820177a272215.jpg) + +![](images/5dbd9bc2784406fd5f6ce41d0d65018dadc64176a90a060983706032eabeb524.jpg) +Ours + +sion [43] with Next3D twice. First, we fine-tune Next3D with the input video to make it learn the original geometry and texture of the avatar. Second, we fine-tune Next3D on the edited image based on the weights of the code and generator from the first fine-tuning. + +# 4.2. Qualitative Comparison + +Geometry Editing. We first compare our method with baselines on geometry editing, e.g., changing the size of the eyes or the contour of the cheek. We use 2D editing tools, e.g., Photoshop and DragGAN [39], to modify the shape of various facial features. Figure. 3 shows qualitative results on avatars from INSTA [76] (the top two) and + +![](images/b90f233289b3424296c9a976ab097c0d45a1186695d1585ded262f1d4bb18c9e.jpg) +Source + +![](images/bbca215fca18c5cbbe146b5d53edc3779f4e7a5ea246ee16d1bacd9035d3b7a3.jpg) +Eyes++ + +![](images/2a722f651cd02ed7b58d5fd25dd1fb0977a1a9439bb44fdb04ffbaec3540e725.jpg) +Nose++ + +![](images/11f2fdbb504c0b0d37bfcd4b4ace6260d93a1af5e31f55992afe19d72d1ee635.jpg) +Face++ + +![](images/f0f487863b3cfeabcb9e4247c4c445967e9b20aa062a96e850c72935064d5c18.jpg) + +![](images/1304beb1885bde3f4d9a7ca7dfec5b8fa96ff1c142769379cbf9ac54025d0380.jpg) + +![](images/efd6a7c719849eb243cf4f786f340f814a920aa7f733f736fbae3cdf577ad52d.jpg) + +![](images/5eded29fc1b7998dc77200c879b3f7471ef3c48a939c288e0422dbe487731404.jpg) + +![](images/a3fc954fc4aed7d2eee24ebba4d068c2a33c7c309221cb7e0bf3cbdacafbb7ec.jpg) +Source +Source + +![](images/b9e3dc8e03f02ff5d2f03e4fe9ef0942679dbcffe76ff0ccd9053d06e3c0214e.jpg) +Eyes++ + +![](images/7197b1c39086bb37705a0fbded774806a2513e754af84ad1b676c586137f2d1a.jpg) +Eyes-- +Mouth++ + +![](images/478bff7e3da760bd9cef87f5f41d37076622f78a5b19d90044bc6fc0ed0d8465.jpg) +Face— + +![](images/ca7a3b80bc3f0171f58e6a3ba841465a1b5301fa1c19326bb7ac360b64eedf67.jpg) +Nose++ +Nose++ +2D Editing + +![](images/e9e712a6349a7975b87c7cebe0d43030027302496e5ee87ba44d8446a222de07.jpg) +Face++ + +![](images/ef9338c97ff64aca72b723801fa053ab1152ba1065bf5d7e4e0bbb8b98739901.jpg) +Jaw-- +Jaw-- + +![](images/bb9011ed37410fd69b9793bc984434283713980fd18fdca3222b95cfc78906dd.jpg) +Eyes++ + +![](images/bca0d99aa388cdd58bce78abebb3ea0cf1ecc971a08ef9c6adaba7d2d4f10c19.jpg) + +![](images/ca681e99a62d0ac892a7186c8de3ee5f00e978a820fa68d74385290840c2e5f0.jpg) +Eyes++ + +![](images/b8997f3637d63b68b8a7826e899b88bee251f410a2a7b40ae5734a480190a727.jpg) +Eyes +Mouth++ + +![](images/230766ffce13445e1c592c48d7dd848e97c543427124fb798e4c5f1b60bc7850.jpg) +Nose++ + +![](images/7232a9d916ef95bdabd8f8a74053b12449a71836f43b80907fa46538b2f2d258.jpg) + +![](images/e79f0fd28e428d35226fb4860bcccc720a4a9a5dc583dd015027b6a37277b20d.jpg) + +Face++ + +![](images/f26317241bbdd62fa66d7a8deba6a4c7acf22955db5a06069d7af6d1f28aebbb.jpg) +Face-- + +![](images/23bae771951b2c1798856ab0617da82f1d23c4fec4095f38e7955573d719acb2.jpg) +Nose++ +Nose++ + +![](images/6e7ddb7187d0d19303cdf76649ce0ad12ce1878c3027f02d6bf5315636df2588.jpg) +Face++ + +![](images/afb64723be416433deff85c3a4665c6bd1f542fb0ec4f2a9cc66fda6d329fd28.jpg) +Jaw +Jaw + +![](images/6e24651350dee56a36acf41eed355e22a704a1be2906b303ae65b268ac9faebc.jpg) +Eyes++ + +![](images/d0afbee0d8c01e1e0d6ba342c2c77f189db98cafe07124cd87a9fbf4c9c31bc8.jpg) +Nose++ + +![](images/ce2b50b1fed1ee519436a36d8dc0e293053d0d6bd9fc898d83a09a9389e69746.jpg) +Face++ + +![](images/741ece174c8853b3d206655524ff8cb1e47fec356b7f6b0d038a5f51f33faeec.jpg) + +![](images/80870e6918668f9108fe218213f2bad7a3dc1773af0c1dad14816f2e4bfd4c1d.jpg) + +![](images/cd14ad760734a44e717880f6c6cc2f83cf676c3ab6ea6e374b5801c9eac75a20.jpg) + +![](images/d5da76eca8a595452a5c7e83a8c96eaf23afafb04714f8f932270492c44c3bb9.jpg) +Eyes++ + +![](images/d3ced2f2959cbd833307ed98ca40876f76275fbb348555d27d6ba9ffb6ceed57.jpg) +Face-- + +![](images/667ffc8d7c57b39661e47dd63b4c79a6078676d3934bd0c3663c52aafc19ec17.jpg) +Face++ + +![](images/38b33b46641b59352520a050b71f1b011a91f147b55303bcbc1bfe09afff9de9.jpg) +Eyes +Mouth++ +Rendered View 2 + +![](images/f4e33dede41e3e56575e3524a713d8893388a0f60e988117d82079771b19903c.jpg) +Nose++ +Nose++ + +![](images/da76fc00a6ee01e36bef667b915c92dcaa254ed633381314c3a159a39cc5a236.jpg) +Jaw-- +Jaw + +![](images/02c7af284df2b8168629893e30005183707e1eafc19abef6404e32aa97f73b6c.jpg) + +![](images/1a31f7924d28469fe607b650ed390b34495eec10a8db058716820f4926348675.jpg) +Nose++ +Hair++ + +![](images/1cf75bf3815a98bd2b46912a9e6cb75c9bc87a62adc7af0cd7df841903dbfec4.jpg) + +![](images/a22ceb8d96f639e53a0ed1a2e1ccbb494d5bd0e8e143744c47000fdf8de9fd0b.jpg) +Rendering + +Rendering + +![](images/58e8c916ead852d59a7ca5c2648b45bce0c0e46562e3b7e1be13ef4a1a6d5fd9.jpg) + +ering + +![](images/e5a34c470e88eeff2443a2a861e733e8b673ae20804126433159de6e552c46dc.jpg) + +Rendering + +![](images/3a5417ce0ca53138a5b8bf3815abeb18e0aa8b55d86e66a4daed3ffc8b0718c5.jpg) + +![](images/ef7ad31c9a02a5c9dab557d9fafb8503df357b9163df2da4225f2f63f2a34dc9.jpg) +Eyes++ + +![](images/bd5fed5b3876ebcd03d8e399050945554b3a5be9adfa561e16a59f9265ebeecf.jpg) + +![](images/cbfd92e8c331c4dd5056e115b36e6fe125d690618614ee1c97f8526b1a8d962f.jpg) +Mouth++ + +![](images/ea2a0146457762a08953182b199d9af2ddd7ef22a452e89e00f7871c1cd412de.jpg) +Rendering + +![](images/2f329292d92306774e6b1cf0c0645a91ebf8f09b27728936ca1a1595ef11f7f5.jpg) + +![](images/35326647f75ba52cb62ca8d4d0e92ade8e1c8673062fb979a50454d465082e65.jpg) + +ering + +![](images/173c0476fffde8cb7cca357ce4246bd91fb28e70e1aea04980b3bf0d8c7496f1.jpg) +Rendering +(c) INSTA & Next3D Geometry Editing Results +Figure 4. Our geometry editing results with the drag-style 2D editing on INSTA [76], NeRFBlendshape [16], and Next3D [50] avatars. + +![](images/92381cd1262c3da770ddd8919d31a415d9ddad5a19ffbdf35a98a126574ca716.jpg) + +![](images/9c9f40b2e66d389cf05f9cb2f47cb9f0f18b20ce821eadfbdc0914bd0b706701.jpg) +Source +Source +Mouse-,Face- +Eyes++, Nose-, Hair++ + +![](images/e07791129d7142fcad928fc6c93713be12517e7fff1c476287590021ec7941fc.jpg) + +![](images/3f0b04bfad01a4bd5db9ef5d2be505e0c19cb96af43bac7fb2f46f855b9000a7.jpg) + +![](images/9f90c5f4153f2249ae183c84f8946c59068cfe25bb24c161b85ecef8953ed105.jpg) + +![](images/3df63afd676cd555cd205036f5c1aab31fe77560826ac343268183e7485dfe5c.jpg) +Rendering +Rendering + +![](images/2c119aca7181f935b6918b36f113cef2e15bd6ac3695574a3494acf4c82d4c6b.jpg) + +![](images/0684e4d46d041e57c89c0306991a29584c03dbcef7d86208bce1fe1ea9b1d28a.jpg) +ering +dering + +NeRFBlendshape [16] (the bottom two). Roop [12] fails to handle the find-grained geometry change, like hairlines, and lips. Nex3D [50] is able to successfully update the avatar based on the editing, however, changes the untouched part and causes an obvious identity shift. PVP [28] can make edits while preserving the identity, however, the magnitude of change tends to be smaller than the given image. In contrast, our method produces the desired editing effect from the edited image and preserves the multiview consistency and identity of the original face. + +We further show more geometry editing results of our method on avatars in various representations in Fig. 4. Our method supports a convenient way to adjust the size of diverse facial features, such as eyes, mouths, jaw, etc., by editing a single rendered image from the avatars. The edits on 2D images are successfully lifted onto the avatar and rendered across different viewpoints and expression. Please refer to the supplementary Sec. C.3 for detailed visualization. + +tions of modified head geometry. + +Texutre Editing. We then show our capability in texture editing. We utilize Photoshop, an online makeup app WebBeauty [1], and text-driven editing method Instructpix2pix [7] to modify the texture on 2D renderings. We show comparisons with PVP [28], Roop [12] and Next3D [50] on four distinct heads avatars (INSTA [76] for the upper three and NeRFBblendshape [16] for bottom one) in Fig. 5. Roop [12] is ineffective in transferring non-human-face-like texture, thus failing in all examples. PVP [28] only transfers partial or blurry textures, and also causes shifts across the expression and head poses. Next3D [50] successfully uplifts the texture editing in 2D images sharply. However, it still suffers from the identity shift issue in the lower two heads and a blurred pattern in the upper two heads in Fig. 5. In contrast, our method faithfully paints the complicated texture following the edited image and preserves the identity of the original avatar and consis + +![](images/93c76c917b283731e84b3bf7e8db2622a160ae3e98a088cb642f83e46ad9ea4b.jpg) + +![](images/3cf60b8772dd99d5e00a0e2b714a2e6c95bdc15d6ce8270373f5775c1eab50d6.jpg) + +![](images/6a2b6efd56c79775607a700590ac5a5e1914bfbae6ffad9cde60200f89298f1c.jpg) + +![](images/16507e58984ee4f84362d69add2b356d5cdbb7487ce1d23ba60acb5d13121634.jpg) + +![](images/5dbf690b0d43d237d0804d41963e940875a70cc1e71fec4a073fd96c4a630f02.jpg) + +![](images/5f407eed9fc984e8a369b44e7c274172a6118b7d79f465e0ba278f9ff6a26f94.jpg) + +![](images/5569949685e0a01f72da09cb42840b7d03ab727a447ea5faa8a1a8b8cc2313b5.jpg) + +![](images/870a96d8475708e0b1fef9b511f58ac496c7d68d8a93e8b2657eb0b8327cfcf8.jpg) + +![](images/b8a0ec6a0aae99d4ba145425c96ab06f38956b5e93d7ad767cd2b9b0ee68996c.jpg) + +![](images/82b070bebc38f5049677adf569d8484ed060ce1daa225b9878342df6a445a877.jpg) + +![](images/274d3feccbfff7783ef952bb15d03656d93339cc11f338bd2f481e956786b59d.jpg) + +![](images/3736910badfff332d0ef1a64ee9f5b5f07fe25971bc8ec3f5f269bba8a75c553.jpg) + +![](images/8f6b0fe5950c661de7baaf0854f5a3a591797b99c17d77fda0844f5b90a37997.jpg) + +![](images/00fa592bb222b620a7d2ef30e5e4a14dfb165ade6207ad3ab852bb5ecfbe1a17.jpg) + +![](images/da3856b52035ad07ab8f2f23f97db60d8a6be9c4db9061c041193480bda3bb4c.jpg) + +![](images/4c6537a3fa5d70e15d107dbb7c5808be7df9095fbf9cbb0153ece89c28f9a2dd.jpg) + +![](images/854c79f2f36cebfe8c56e6ac71d89066d0f131709d9de4633dda669c36b5551a.jpg) + +![](images/03a44820aecb534d32305107396b568443e567ebeccca15ef81a57eb4d5b2e6f.jpg) + +![](images/40a09dfe949365a35d78d66ec57f0a2eebe8505cdb1b3cedc543739c52e03b2a.jpg) + +![](images/3dc071161141e8dd00b1e90a53b2d64f9c0a7824bbf3126ced18bf205380c7f6.jpg) + +![](images/e2422af44a97c889265828ae743c16a0d75fd2383ef1191c3916bc33273ec7b1.jpg) + +![](images/e41fef444d7908c0cbc5cee7e576dfda7f4f6e6662fb00a3198502cc6b68eb0f.jpg) +Source + +![](images/68b7b3ad5625c1d3e491ff00895803e4f226c18aed515158736b2f6d5eb7b40d.jpg) +2D Editing + +![](images/443fe59e0873cc5e8c6833a9c528e78386bfaab33c93d6ed114aeff22796b102.jpg) +Reference Animation + +![](images/f10a02b46c3033b3f461b036bc7bfe28904d20f589cd981a55db39059ffdc4a9.jpg) +PVP + +![](images/6c01f410daf0aa7c524a0126b7aeb87c6582df5350e211cf675a62d16a7713fc.jpg) +Roop + +![](images/ce6087c99512b74f1caf5b96fdcb54815366c619a23c28fb51135d334a08fbf6.jpg) +Next3D + +![](images/94129cc35166d5abf1e6ca41942d34e597ce4f1a01432dddc92af159950105ec.jpg) +Ours + +![](images/69037e33481230528a1f2921e635a940c72097da1718e3eccdc3081d9fadacd4.jpg) + +![](images/1c4472994369a25d4b49ac36951319e3ec5c9d31a5fd1108fcaebc98a7ce2f21.jpg) + +![](images/419fdc01d70ab579a01c704d9c297dbfe7ecd98668dba16c3fa0a2e05f664304.jpg) + +![](images/d67de9d18779d38cfd6e51fdebe8467191d75ddef42b6c71a61eedc544cc590b.jpg) + +![](images/693849cfa7572135edd57630e46cb8c2527460ecc4b70ed938ba416dc67d9b77.jpg) + +![](images/270d379d3cf99c131a41185076bafd2e2e5b1d196cb43dea981f6778bdc7c310.jpg) + +![](images/eaf86ffe2902db415b71848c54e3d37fb0c321425fc3f59801d9da481ac9fa6a.jpg) + +![](images/bedeca8294a3af5e040ffe291d01a1871e6b37e268aa137bb455c56097712ff2.jpg) + +![](images/a998e56ea963608f1354780e84f24a4cb7fded5693d42fb2399b7ac4a6b23c78.jpg) + +![](images/3a3604e48aa3631cb242f6388bc15ba7f3f6ad7e3599665bb2c774f387807f59.jpg) + +![](images/1365f99133e2e83108312a66de1441267bdef4642a87768a8b91d09594682ff1.jpg) + +![](images/791fe9d51ead790c5ea66b7808dafa3129ac078a02feae920e5d23c49d749670.jpg) + +![](images/d0217c83307093b2e54ce669def02d63fe37926f31c14f5bcb307a1e28cc874d.jpg) + +![](images/4f5e235dc0659090ee67172c561047d24efc656044a9d75829c5ca27dbea4765.jpg) + +![](images/2bb2d62cd2e78353554e4f9de7efa3d166f11675bc07c2a566f07eb1756ec4ca.jpg) + +![](images/5ce42b0aefb7fd52b678606a1827bd9e434285c0cc3b57b8cab40c7e46939a3e.jpg) +Reference Animation + +![](images/449238b1e5b10ba421609285fd247a627168f4d10c21cb99ccc664e9dc47cf0d.jpg) +PVP + +![](images/828046c3acfee0c3ec5a73b6a232e84144d84d0b42149de9fc1bf4d86c720457.jpg) +Roop + +![](images/8ecc9323bb53efe8184ae53663e39c7df5791349fd916a49bba37924ca796313.jpg) +Next3D + +![](images/ba6727572b4d052774b7b5d34fca0e70fdc5b1184c972def0aa0b5f012398fc8.jpg) +Ours + +![](images/1c072a5148be7a0fdee133d39d607944ffcc76662bc86df7feb7d07d50441b39.jpg) +Figure 5. We compare texture editing with PVP [28], Roop [12], Next3D [50] on INSTA [76] and NeRFBlendshape [16] avatars. The "Reference Animation" denotes the image of the original avatar under the same expression with the rendered edited view. + +![](images/2d144f5947784f4cdaaa08a753eb470a5db0c221e0ab5b7d24c1612c7250583e.jpg) +Figure 6. Analysis of our effectiveness with the naive baselines that can accomplish the single-view avatar editing. + +tency across multiple viewpoints and expressions. + +We show extensive texture editing results in three avatar representations in Fig. 7. Our method supports a wide range of texture editing, including global style transfer ("Add a clown makeup" via text-driven editing), semantic-driven editing (changing the hair color), and free-form sketch (painting on the face). Please refer to the supplementary Sec. C.1/2 for hybrid editing and face reenactment results. + +# 4.3.Quantitative Comparison + +User Study. We conducted a user study to validate our method further quantitatively. Following the evaluation protocol of Avatarstudio [38], users are required to watch the rendered videos of different methods side by side and answer each question by picking up one of the methods. For each group of editing results, we will ask four questions the + +same as Avatarstudio [38] on editing preservation, identity preservation, temporal consistency and overall performance in Tab. 1. We collected statistics from 30 participants in 12 groups of edit results. Results are reported in Tab. 1. Our method exhibits the best editing ability while keeping the best consistency across different facial expressions for geometry and texture editing. Moreover, our method performs the best in keeping non-edited parts untouched and making the human heads still recognizable after being edited e.g., identity preservation metric in Tab. 1. Please refer to the supplementary Sec. B.4 for more details. + +Image Identity Similarity Evaluation. Following the VoLux-GAN [51], we further evaluate the cross-view identity consistency in our edit results. We take 7 groups of geometry editing results and 7 groups of texture editing results, and render the 350 images of edited avatars with different viewpoints and expressions. We calculate the cosine similarities between each rendered image and the single-view 2D edited image and average the similarities on all rendered images as metrics. As reported in Tab. 1, our method outperforms all baselines in recovering desired editing effect and retaining identity consistency. + +# 4.4. Native Editing Capability of Avatar + +In this section, we analyze the editing capability from the original avatar model and show the necessity of our design. + +Comparison to 3DMM-based Geometry Editing. One intuitive way to support geometry editing is updating the underlying 3DMM geometry [14]. Here, we investigate this method and verify its capability. Specifically, we run state-of-the-art single image-based 3DMM reconstruction method [75], and update the 3DMM shape parameter of the avatar model with the estimated one. Note that such an ap- + +![](images/b4d2729bd274985130b2a1b9d63c175f1cd732ccc0cc4c0272d23a83936acbf6.jpg) +Figure 7. We show our texture editing results using the 2D editing method with text-prompt, pattern painting and makeup drawing on INSTA [76] and NeRFBlendshape [16], Next3D [50] avatars. + +proach is only available for those avatars that use 3DMM explicitly, and we take INSTA [76] as an example and show the results in Fig. 6 (a). The 3DMM fitting tends to fail when the edited face is out-of-distribution, so the magnitude of editing may not be correct (e.g. the face shape). Changing the 3DMM parameter could also result in blurry rendering from the pre-trained avatar model, which cannot be trivially fixed even after fine-tuning the rendering decoder. + +Comparison to Fine-tuning for Texture Editing. Texture editing could be done by fine-tuning the avatar with the one-shot edited image. We test this method on avatars from NeRFBlendShape [16] and show the results in Fig. 6 (b). While large structured changes, e.g., hair color, can be edited, detailed editing is largely ignored. Please refer to + +the supplementary Sec. C.4 for more ablation studies. + +# 5. Conclusion + +We have proposed a novel generic editing approach that allows users to edit various volumetric head avatar representations from a single image, where an expression-aware modification generator lifts the editing to the 3D avatar while maintaining consistency across multiple expression and viewpoints. As a limitation, we cannot add additional objects (e.g., hat) or modify the hairstyle as shown in our supplementary Sec. C.5, which may be improved by learning extra specialized geometry addition and hair modification generators. + +Acknowledgment: This work was partially supported by the NSFC (No. 62102356) and Ant Group. + +# References + +[1] Webbar/beauty demo app. 6 +[2] Rameen Abdal, Hsin-Ying Lee, Peihao Zhu, Mengei Chai, Aliaksandr Siarohin, Peter Wonka, and Sergey Tulyakov. 3davatargan: Bridging domains for personalized editable avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4552-4562, 2023. 2 +[3] ShahRukh Athar, Zexiang Xu, Kalyan Sunkavalli, Eli Shechtman, and Zhixin Shu. Rignerf: Fully controllable neural 3d portraits. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 20364-20373, 2022. 1 +[4] Ziqian Bai, Feitong Tan, Zeng Huang, Kripasindhu Sarkar, Danhang Tang, Di Qiu, Abhimitra Meka, Ruofei Du, Mingsong Dou, Sergio Orts-Escalano, et al. Learning personalized high quality volumetric head avatars from monocular rgb videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16890-16900, 2023. 1, 2, 3 +[5] Chong Bao, Yinda Zhang, Bangbang Yang, Tianxing Fan, Zesong Yang, Hujun Bao, Guofeng Zhang, and Zhaopeng Cui. Sine: Semantic-driven image-based nef editing with prior-guided editing field. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20919-20929, 2023. 2 +[6] V Blanz and T Vetter. A morphable model for the synthesis of 3d faces. In 26th Annual Conference on Computer Graphics and Interactive Techniques (SIGGRAPH 1999), pages 187-194. ACM Press, 1999. 1, 3 +[7] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 2, 3, 4, 6 +[8] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3 +[9] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In Proceedings of the European Conference on Computer Vision, pages 333-350. Springer, 2022. 2 +[10] Anton Cherepkov, Andrey Voynov, and Artem Babenko. Navigating the gan parameter space for semantic image editing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3671-3680, 2021. 2 +[11] deepfakes. faceswap. https://github.com/deepfakes/faceswap, 2023. Accessed: 2023-10-10. 1, 2 +[12] deepfakes. roop. SomdevSangwan, 2023. Accessed: 2023-10-10. 1, 2, 5, 6, 7 + +[13] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690-4699, 2019. 5 +[14] Yao Feng, Haiwen Feng, Michael J Black, and Timo Bolkart. Learning an animatable detailed 3d face model from in-the-wild images. ACM Transactions on Graphics (ToG), 40(4): 1-13, 2021. 7 +[15] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 1, 3 +[16] Xuan Gao, Chenglai Zhong, Jun Xiang, Yang Hong, Yudong Guo, and Juyong Zhang. Reconstructing personalized semantic facial nerf models from monocular video. ACM Transactions on Graphics (TOG), 41(6):1-12, 2022. 1, 2, 3, 5, 6, 7, 8 +[17] Songwei Ge, Taesung Park, Jun-Yan Zhu, and Jia-Bin Huang. Expressive text-to-image generation with rich text. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7545-7556, 2023. 2 +[18] Ayaan Haque, Matthew Tancik, Alexei A Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. arXiv preprint arXiv:2303.12789, 2023. 2, 3 +[19] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. Advances in neural information processing systems, 33:9841-9850, 2020. 2 +[20] Amir Hertz, Kfir Aberman, and Daniel Cohen-Or. Delta denoising score. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2328–2337, 2023. 2 +[21] Lianghua Huang, Di Chen, Yu Liu, Yujun Shen, Deli Zhao, and Jingren Zhou.Composer: Creative and controllable image synthesis with composable conditions. arXiv preprint arXiv:2302.09778, 2023. 2 +[22] Wentao Jiang, Si Liu, Chen Gao, Jie Cao, Ran He, Jiashi Feng, and Shuicheng Yan. Psgan: Pose and expression robust spatial-aware gan for customizable makeup transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5194-5202, 2020. 2, 4 +[23] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2 +[24] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 2, 4, 5 +[25] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021. 2 + +[26] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194-1, 2017. 3, 4 +[27] Tingting Li, Ruihe Qian, Chao Dong, Si Liu, Qiong Yan, Wenwu Zhu, and Liang Lin. Beautygan: Instance-level facial makeup transfer with deep generative adversarial network. In Proceedings of the 26th ACM international conference on Multimedia, pages 645-653, 2018. 2, 4 +[28] K-E Lin, Alex Trevithick, Keli Cheng, Michel Sarkis, Mohsen Ghafoorian, Ning Bi, Gerhard Reitmayr, and Ravi Ramamoorthi. Pvp: Personalized video prior for editable dynamic portraits using stylegan. In Computer Graphics Forum, page e14890. Wiley Online Library, 2023. 1, 2, 5, 6, 7 +[29] Huan Ling, Karsten Kreis, Daiqing Li, Seung Wook Kim, Antonio Torralba, and Sanja Fidler. Editgan: High-precision semantic image editing. Advances in Neural Information Processing Systems, 34:16331-16345, 2021. 2 +[30] Pengyang Ling, Lin Chen, Pan Zhang, Huaian Chen, and Yi Jin. Freedrag: Point tracking is not you need for interactive point-based image editing. arXiv preprint arXiv:2307.04684, 2023. 2, 4 +[31] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 2 +[32] Aryan Mikaeili, Or Perel, Mehdi Safaee, Daniel Cohen-Or, and Ali Mahdavi-Amiri. Sked: Sketch-guided text-based 3d editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14607–14619, 2023. 2, 3 +[33] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2 +[34] Chong Mou, Xintao Wang, Jiechong Song, Ying Shan, and Jian Zhang. Dragondiffusion: Enabling drag-style manipulation on diffusion models. arXiv preprint arXiv:2307.02421, 2023. 2 +[35] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. arXiv preprint arXiv:2201.05989, 2022. 2 +[36] Thao Nguyen, Anh Tuan Tran, and Minh Hoai. Lipstick ain't enough: beyond color matching for in-the-wild makeup transfer. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 13305-13314, 2021. 2, 4 +[37] Thu Nguyen-Phuoc, Gabriel Schwartz, Yuting Ye, Stephen Lombardi, and Lei Xiao. Alteredavatar: Stylizing dynamic 3d avatars with fast style adaptation. arXiv preprint arXiv:2305.19245, 2023. 2 +[38] Mohit Mendiratta Pan, Mohamed Elgharib, Kartik Teotia, Ayush Tewari, Vladislav Golyanik, Adam Kortylewski, Christian Theobalt, et al. Avatarstudio: Text-driven editing of 3d dynamic human head avatars. arXiv preprint arXiv:2306.00547, 2023. 2, 7 + +[39] Xingang Pan, Ayush Tewari, Thomas Leimkuhler, Lingjie Liu, Abhimitra Meka, and Christian Theobalt. Drag your gan: Interactive point-based manipulation on the generative image manifold. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2, 4, 5 +[40] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2 +[41] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021. 2, 5 +[42] Ivan Perov, Daiheng Gao, Nikolay Chervoniy, Kunlin Liu, Sugasa Marangonda, Chris Umé, Mr Dpfks, Carl Shift Facenheim, Luis RP, Jian Jiang, et al. Deepfacelab: Integrated, flexible and extensible face-swapping framework. arXiv preprint arXiv:2005.05535, 2020. 1, 2 +[43] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on graphics (TOG), 42(1):1-13, 2022. 4, 5 +[44] Sara Fridovich-Keil and Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance Fields without Neural Networks. In CVPR, 2022. 2 +[45] Ruizhi Shao, Jingxiang Sun, Cheng Peng, Zerong Zheng, Boyao Zhou, Hongwen Zhang, and Yebin Liu. Control4d: Dynamic portrait editing by learning 4d gan from 2d diffusion-based editor. arXiv preprint arXiv:2305.20082, 2023. 2 +[46] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9243-9252, 2020. 2 +[47] Yujun Shi, Chuhui Xue, Jiachun Pan, Wenqing Zhang, Vincent YF Tan, and Song Bai. Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. arXiv preprint arXiv:2306.14435, 2023. 2 +[48] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 2 +[49] Jingxiang Sun, Xuan Wang, Yichun Shi, Lizhen Wang, Jue Wang, and Yebin Liu. Ide-3d: Interactive disentangled editing for high-resolution 3d-aware portrait synthesis. arXiv preprint arXiv:2205.15517, 2022. 2 +[50] Jingxiang Sun, Xuan Wang, Lizhen Wang, Xiaoyu Li, Yong Zhang, Hongwen Zhang, and Yebin Liu. Next3d: Generative neural texture rasterization for 3d-aware head avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20991-21002, 2023. 1, 2, 3, 4, 5, 6, 7, 8 +[51] Feitong Tan, Sean Fanello, Abhimitra Meka, Sergio Orts-Escolano, Danhang Tang, Rohit Pandey, Jonathan Taylor, + +Ping Tan, and Yinda Zhang. Volux-gan: A generative model for 3d face synthesis with hdri relighting. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 7 +[52] Junshu Tang, Bo Zhang, Binxin Yang, Ting Zhang, Dong Chen, Lizhuang Ma, and Fang Wen. 3dfaceshop: Explicitly controllable 3d-aware portrait generation. IEEE Transactions on Visualization and Computer Graphics, 2023. 1, 2 +[53] Ziyan Wang, Timur Bagautdinov, Stephen Lombardi, Tomas Simon, Jason Saragih, Jessica Hodgins, and Michael Zollhofer. Learning compositional radiance fields of dynamic human heads. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5704-5713, 2021. 2 +[54] Yue Wu, Yu Deng, Jiaolong Yang, Fangyun Wei, Qifeng Chen, and Xin Tong. Anifacegan: Animatable 3d-aware face image generation for video avatars. Advances in Neural Information Processing Systems, 35:36188-36201, 2022. 1, 2 +[55] Weihao Xia, Yulun Zhang, Yujiu Yang, Jing-Hao Xue, Bolei Zhou, and Ming-Hsuan Yang. Gan inversion: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3121-3138, 2022. 4 +[56] Yinghao Xu, Sida Peng, Ceyuan Yang, Yujun Shen, and Bolei Zhou. 3d-aware image synthesis via learning structural and textural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18430-18439, 2022. 2 +[57] Yuelang Xu, Lizhen Wang, Xiaochen Zhao, Hongwen Zhang, and Yebin Liu. Avatarmav: Fast 3d head avatar reconstruction using motion-aware neural voxels. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-10, 2023. 1, 3 +[58] Bangbang Yang, Yinda Zhang, Yinghao Xu, Yijin Li, Han Zhou, Hujun Bao, Guofeng Zhang, and Zhaopeng Cui. Learning object-compositional neural radiance field for ed-itable scene rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13779-13788, 2021. 2 +[59] Bangbang Yang, Chong Bao, Junyi Zeng, Hujun Bao, Yinda Zhang, Zhaopeng Cui, and Guofeng Zhang. Neumesh: Learning disentangled neural mesh-based implicit field for geometry and texture editing. In European Conference on Computer Vision, pages 597-614. Springer, 2022. +[60] Bangbang Yang, Yinda Zhang, Yijin Li, Zhaopeng Cui, Sean Fanello, Hujun Bao, and Guofeng Zhang. Neural rendering in a room: amodal 3d understanding and free-viewpoint rendering for the closed scene composed of pre-captured objects. ACM Transactions on Graphics (TOG), 41(4):1-10, 2022. 2 +[61] Chenyu Yang, Wanrong He, Yingqing Xu, and Yang Gao. Elegant: Exquisite and locally editable gan for makeup transfer. In European Conference on Computer Vision, pages 737-754. Springer, 2022. 2 +[62] Xingrui Yang, Hai Li, Hongjia Zhai, Yuhang Ming, Yuqian Liu, and Guofeng Zhang. Vox-fusion: Dense tracking and mapping with voxel-based neural implicit representation. In + +2022 IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 499-507. IEEE, 2022. 2 +[63] Weicai Ye, Shuo Chen, Chong Bao, Hujun Bao, Marc Pollefeys, Zhaopeng Cui, and Guofeng Zhang. Intrinsicerf: Learning intrinsic neural radiance fields for editable novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 339-351, 2023. 2 +[64] Deheng Zhang, Clara Fernandez-Labrador, and Christopher Schroers. Coarf: Controllable 3d artistic style transfer for radiance fields. In 2024 International Conference on 3D Vision (3DV). IEEE, 2024. 2 +[65] Hao Zhang, Yanbo Xu, Tianyuan Dai, Tai Chi-Keung Tang, et al. Fdnerf: Semantics-driven face reconstruction, prompt editing and relighting with diffusion models. arXiv preprint arXiv:2306.00783, 2023. 2 +[66] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Nefactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (TOG), 40(6):1-18, 2021. 2 +[67] Boming Zhao, Bangbang Yang, Zhenyang Li, Zuoyue Li, Guofeng Zhang, Jiashu Zhao, Dawei Yin, Zhaopeng Cui, and Hujun Bao. Factorized and controllable neural re-rendering of outdoor scene for photo extrapolation. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1455-1464, 2022. 2 +[68] Yufeng Zheng, Victoria Fernández Abrevaya, Marcel C Bühler, Xu Chen, Michael J Black, and Otmar Hilliges. Im avatar: Implicit morphable head avatars from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13545-13555, 2022. 1, 2, 3 +[69] Yufeng Zheng, Wang Yifan, Gordon Wetzstein, Michael J Black, and Otmar Hilliges. Pointavatar: Deformable point-based head avatars from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21057-21067, 2023. 1, 2, 3 +[70] Jiapeng Zhu, Yujun Shen, Deli Zhao, and Bolei Zhou. Indomain gan inversion for real image editing. In European conference on computer vision, pages 592-608. Springer, 2020. 2 +[71] Jiapeng Zhu, Ceyuan Yang, Yujun Shen, Zifan Shi, Bo Dai, Deli Zhao, and Qifeng Chen. Linkgan: Linking gan latents to pixels for controllable image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7656-7666, 2023. 2 +[72] Peihao Zhu, Rameen Abdal, Yipeng Qin, and Peter Wonka. Sean: Image synthesis with semantic region-adaptive normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5104-5113, 2020. 2 +[73] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12786-12796, 2022. 2 + +[74] Jingyu Zhuang, Chen Wang, Lingjie Liu, Liang Lin, and Guanbin Li. Dreameditor: Text-driven 3d scene editing with neural fields. arXiv preprint arXiv:2306.13455, 2023. 2, 3 +[75] Wojciech Zielonka, Timo Bolkart, and Justus Thies. Towards metrical reconstruction of human faces. In European Conference on Computer Vision, pages 250–269. Springer, 2022. 7 +[76] Wojciech Zielonka, Timo Bolkart, and Justus Thies. Instant volumetric head avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4574-4584, 2023. 1, 2, 3, 5, 6, 7, 8 +[77] zllrunning. face-makeup.pytorch. https://github.com/zllrunning/face-makeup.PyTorch, 2023. Accessed: 2023-10-10. 4 \ No newline at end of file diff --git a/2024/GeneAvatar_ Generic Expression-Aware Volumetric Head Avatar Editing from a Single Image/images.zip b/2024/GeneAvatar_ Generic Expression-Aware Volumetric Head Avatar Editing from a Single Image/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b34b7d6f673de784e438f8d1b2946aa22ed66ee7 --- /dev/null +++ b/2024/GeneAvatar_ Generic Expression-Aware Volumetric Head Avatar Editing from a Single Image/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0b83409cd15b18c538da429238de0cd9b883831f7ccd29073cf04a585db8ce35 +size 1037847 diff --git a/2024/GeneAvatar_ Generic Expression-Aware Volumetric Head Avatar Editing from a Single Image/layout.json b/2024/GeneAvatar_ Generic Expression-Aware Volumetric Head Avatar Editing from a Single Image/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..17e0f93ef1590c62daba3cf9ee0ad868f2a86a16 --- /dev/null +++ b/2024/GeneAvatar_ Generic Expression-Aware Volumetric Head Avatar Editing from a Single Image/layout.json @@ -0,0 +1,19217 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 71, + 67, + 523, + 103 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 67, + 523, + 103 + ], + "spans": [ + { + "bbox": [ + 71, + 67, + 523, + 103 + ], + "type": "text", + "content": "GeneAvatar: Generic Expression-Aware Volumetric Head Avatar Editing from a Single Image" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 125, + 484, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 125, + 484, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 125, + 484, + 140 + ], + "type": "text", + "content": "Chong Bao" + }, + { + "bbox": [ + 105, + 125, + 484, + 140 + ], + "type": "inline_equation", + "content": "^{1*§}" + }, + { + "bbox": [ + 105, + 125, + 484, + 140 + ], + "type": "text", + "content": " Yinda Zhang" + }, + { + "bbox": [ + 105, + 125, + 484, + 140 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 105, + 125, + 484, + 140 + ], + "type": "text", + "content": " Yuan Li" + }, + { + "bbox": [ + 105, + 125, + 484, + 140 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 105, + 125, + 484, + 140 + ], + "type": "text", + "content": " Xiyu Zhang" + }, + { + "bbox": [ + 105, + 125, + 484, + 140 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 105, + 125, + 484, + 140 + ], + "type": "text", + "content": " Bangbang Yang" + }, + { + "bbox": [ + 105, + 125, + 484, + 140 + ], + "type": "inline_equation", + "content": "^{4}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 129, + 140, + 467, + 154 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 140, + 467, + 154 + ], + "spans": [ + { + "bbox": [ + 129, + 140, + 467, + 154 + ], + "type": "text", + "content": "Hujun Bao" + }, + { + "bbox": [ + 129, + 140, + 467, + 154 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 129, + 140, + 467, + 154 + ], + "type": "text", + "content": " Marc Pollefeys" + }, + { + "bbox": [ + 129, + 140, + 467, + 154 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 129, + 140, + 467, + 154 + ], + "type": "text", + "content": " Guofeng Zhang" + }, + { + "bbox": [ + 129, + 140, + 467, + 154 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 129, + 140, + 467, + 154 + ], + "type": "text", + "content": " Zhaopeng Cui" + }, + { + "bbox": [ + 129, + 140, + 467, + 154 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 73, + 153, + 522, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 153, + 522, + 168 + ], + "spans": [ + { + "bbox": [ + 73, + 153, + 522, + 168 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 73, + 153, + 522, + 168 + ], + "type": "text", + "content": "State Key Lab of CAD&CG, Zhejiang University " + }, + { + "bbox": [ + 73, + 153, + 522, + 168 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 73, + 153, + 522, + 168 + ], + "type": "text", + "content": "Google " + }, + { + "bbox": [ + 73, + 153, + 522, + 168 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 73, + 153, + 522, + 168 + ], + "type": "text", + "content": "ETH Zürich " + }, + { + "bbox": [ + 73, + 153, + 522, + 168 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 73, + 153, + 522, + 168 + ], + "type": "text", + "content": "ByteDance" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 53, + 186, + 211, + 284 + ], + "blocks": [ + { + "bbox": [ + 77, + 173, + 182, + 184 + ], + "lines": [ + { + "bbox": [ + 77, + 173, + 182, + 184 + ], + "spans": [ + { + "bbox": [ + 77, + 173, + 182, + 184 + ], + "type": "text", + "content": "GeneAvatar + NeBShape" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 53, + 186, + 211, + 284 + ], + "lines": [ + { + "bbox": [ + 53, + 186, + 211, + 284 + ], + "spans": [ + { + "bbox": [ + 53, + 186, + 211, + 284 + ], + "type": "image", + "image_path": "2e0986f0a2bda4ee93e297c0c43264b7365ebb9b2b4eb81632b5107250b134b6.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 225, + 183, + 375, + 284 + ], + "blocks": [ + { + "bbox": [ + 249, + 173, + 340, + 183 + ], + "lines": [ + { + "bbox": [ + 249, + 173, + 340, + 183 + ], + "spans": [ + { + "bbox": [ + 249, + 173, + 340, + 183 + ], + "type": "text", + "content": "GeneAvatar + INSTA" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 225, + 183, + 375, + 284 + ], + "lines": [ + { + "bbox": [ + 225, + 183, + 375, + 284 + ], + "spans": [ + { + "bbox": [ + 225, + 183, + 375, + 284 + ], + "type": "image", + "image_path": "7d863c72e7ab48cd31396cafe0145ba7a2de47719d4017b9f45a880d8a6f4308.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 286, + 547, + 319 + ], + "lines": [ + { + "bbox": [ + 46, + 286, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 286, + 547, + 319 + ], + "type": "text", + "content": "Figure 1. We propose a generic approach to edit 3D avatars in various volumetric representations (NeRFBblendShape [16], INSTA [76], Next3D [50]) from a single perspective using 2D editing methods with drag-style, text-prompt and pattern painting. Our editing results are consistent across multiple facial expression and camera viewpoints." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 381, + 184, + 543, + 283 + ], + "blocks": [ + { + "bbox": [ + 413, + 173, + 507, + 183 + ], + "lines": [ + { + "bbox": [ + 413, + 173, + 507, + 183 + ], + "spans": [ + { + "bbox": [ + 413, + 173, + 507, + 183 + ], + "type": "text", + "content": "GeneAvatar + Next3D" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 381, + 184, + 543, + 283 + ], + "lines": [ + { + "bbox": [ + 381, + 184, + 543, + 283 + ], + "spans": [ + { + "bbox": [ + 381, + 184, + 543, + 283 + ], + "type": "image", + "image_path": "c0c1a36e134b9c478fe95ae59a9424db14c2501c633a91a929b807ea926e0322.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 143, + 324, + 192, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 324, + 192, + 337 + ], + "spans": [ + { + "bbox": [ + 143, + 324, + 192, + 337 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 339, + 290, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 339, + 290, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 339, + 290, + 593 + ], + "type": "text", + "content": "Recently, we have witnessed the explosive growth of various volumetric representations in modeling animatable head avatars. However, due to the diversity of frameworks, there is no practical method to support high-level applications like 3D head avatar editing across different representations. In this paper, we propose a generic avatar editing approach that can be universally applied to various 3DMM-driving volumetric head avatars. To achieve this goal, we design a novel expression-aware modification generative model, which enables lift 2D editing from a single image to a consistent 3D modification field. To ensure the effectiveness of the generative modification process, we develop several techniques, including an expression-dependent modification distillation scheme to draw knowledge from the large-scale head avatar model and 2D facial texture editing tools, implicit latent space guidance to enhance model convergence, and a segmentation-based loss reweight strategy for fine-grained texture inversion. Extensive experiments demonstrate that our method delivers high-quality and consistent results across multiple expression and viewpoints. Project page: https://zju3dv.github.io/geneavatar/." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 47, + 597, + 128, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 128, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 128, + 609 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 617, + 287, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 678 + ], + "type": "text", + "content": "Recently various volumetric representations [3, 4, 15, 16, 57, 68, 69, 76] have achieved remarkable success in reconstructing personalized, animatable, and photorealistic head avatars using implicit [15, 16, 57, 68, 69] or explicit [3, 4, 76] conditioning of 3D Morphable Models" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 325, + 545, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 325, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 306, + 325, + 545, + 373 + ], + "type": "text", + "content": "(3DMM) [6]. A popular demand, once with a created avatar model, is to edit the avatar, e.g., for face shape, facial makeup, or apply artistic effects, for the downstream applications, e.g., in virtual/augmented reality." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 375, + 546, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 375, + 546, + 542 + ], + "spans": [ + { + "bbox": [ + 304, + 375, + 546, + 542 + ], + "type": "text", + "content": "Ideally, the desired editing functionality on the animatable avatar should have the following properties. (1) Adaptable: The editing method should be applicable across various volumetric avatar representations. This is particularly valuable in light of the growing diversity of avatar frameworks [16, 50, 76]. (2) User-friendly: The editing should be user-friendly and intuitive. Preferably, the editing of geometry and texture of the 3D avatar could be accomplished on a single-perspective rendered image. (3) Faithful: The editing results should be consistent across various facial expression and camera viewpoints. (4) Flexible: Both intensive editing (e.g., global appearance transfer following style prompts) and delicate local editing (e.g., dragging to enlarge eyes or ears) should be supported as illustrated in Fig. 1." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 544, + 547, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 547, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 547, + 676 + ], + "type": "text", + "content": "However, 3D-aware avatar editing is still underexplored in both geometry and texture. One plausible way is to perform 3D editing via animatable 3D GAN [50, 52, 54], but the editing results may not be consistently reflected when expression and camera viewpoint change. Alternatively, the editing can be done on the generated 2D video using 2D personalized StyleGAN [28]; however, the identity shift is often observed. Some face-swapping methods [11, 12, 42] are capable of substituting the face in a video with another face derived from a reference image or video; however, they do not support texture editing and local geometry editing." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "content": "To this end, we propose GeneAvatar - a generic approach to support fine-grained 3D editing in various volumetric avatar representations from a single perspective by" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 495, + 15 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 495, + 15 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 495, + 15 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "spans": [ + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 25, + 461, + 36 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 25, + 461, + 36 + ], + "spans": [ + { + "bbox": [ + 179, + 25, + 461, + 36 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 683, + 155, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 683, + 155, + 693 + ], + "spans": [ + { + "bbox": [ + 57, + 683, + 155, + 693 + ], + "type": "text", + "content": "*Authors contributed equally." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 59, + 693, + 139, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 693, + 139, + 703 + ], + "spans": [ + { + "bbox": [ + 59, + 693, + 139, + 703 + ], + "type": "text", + "content": "† Corresponding authors." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 58, + 703, + 225, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 225, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 225, + 712 + ], + "type": "inline_equation", + "content": "\\S" + }, + { + "bbox": [ + 58, + 703, + 225, + 712 + ], + "type": "text", + "content": " The work was partially done when visiting ETHZ." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8952" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 418 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 418 + ], + "type": "text", + "content": "leveraging 2D editing methods, such as drag-based methods [30, 34, 39, 47], text-driven methods [7, 17, 20, 40, 41], or image editing tools like Photoshop (see Fig. 1). We adopt a novel editing framework that formulates the editing as predicting expression-aware 3D modification fields applied in the geometry and texture space of the volumetric avatars, which makes editing independent with the original representation as long as they are in parametric-driven radiance field, e.g., 3DMM-based neural avatar. Second, to ensure that the 2D image editing can be faithfully transferred into the 3D space, we propose to learn a generative model for modification fields, which produces 3DMM conditional modification fields from a compact latent space. Given the rendered avatar image and its edited counterpart, we conduct auto-decoding optimization on this generative model to search for the latent code that best explains the editing, obtaining consistent 3DMM conditional modification fields across various viewpoints and expression. Third, inspired by the spirit of learning from the pre-trained large-scale generative model [7, 18, 32, 74], we design a novel distillation scheme to learn the expression-dependent modification from a 3DMM-based GAN [50] and 2D face editing tools [22, 27, 36]. The scheme addresses the issue of insufficient real training data (i.e., avatars with a wide range of geometry and texture changes). Besides, we develop several techniques to enhance the editing effects, including the implicit latent space guidance to stabilize the initialization and convergence of learning, and a segmentation-based loss reweight strategy for fine-grained texture inversion." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 419, + 289, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 419, + 289, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 419, + 289, + 647 + ], + "type": "text", + "content": "The contributions of our paper are summarized as follows. 1) We propose a generic avatar editing approach that can be applied to various 3DMM driving head avatars in the neural radiance field. To achieve this, we design a novel expression-aware modification generative model, which lifts the geometry and texture editing from a single image to a consistent 3D modification field. 2) To bootstrap the training of the modification generator with limited real paired training data, we design a distillation scheme to learn the expression-dependent geometry and texture modification from the large-scale head avatar generative model [50] and 2D face texture editing tools [22, 27, 36], and develop several techniques, including implicit guidance in latent space to improve training convergence, and a loss reweight strategy based on segmentation for fine-grained texture inversion. 3) Extensive experiments on various head avatar representations demonstrate that our method delivers high-quality editing results and the editing effects are consistent under different viewpoints and expression." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 651, + 134, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 651, + 134, + 662 + ], + "spans": [ + { + "bbox": [ + 47, + 651, + 134, + 662 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 288, + 715 + ], + "type": "text", + "content": "2D Head Avatar Editing. The manipulation of 2D head avatars has made significant strides in recent years. Various GAN-based methods [23-25] can result in precise and high-resolution human face editing by leveraging image space se" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 262 + ], + "type": "text", + "content": "mantic information [29, 72] or controlling latent space explorations [10, 19, 46, 70]. Some approaches [22, 27, 36, 61] focus on the task of makeup transferring by exploiting the GAN to learn the transferring ability from a large unaligned makeup and non-makeup face datasets. The drag-based GAN editing approach [39, 71] gained vast popularity due to providing a user-friendly editing way. PVP [28] uses a monocular video to fine-tune StyleGAN [23, 24] to obtain the personalized image generator and provide various editing functions. The diffusion models [7, 21] also show the capability of achieving fine-grained face editing with text prompt and other conditional input. For editing an avatar in a video, lots of face-swapping methods [11, 12, 42] have emerged to provide high-quality and properly aligned face-swapping results. However, these approaches typically suffer from multi-view consistency and identity preservation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 263, + 547, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 263, + 547, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 263, + 547, + 456 + ], + "type": "text", + "content": "3D Head Avatar Editing. Neural Radiance Field [33] has exhibited great reconstruction and rendering qualities in SLAM [62, 73], scene editing [5, 58-60, 64] and relighting [63, 66, 67], especially promoting the emergence of many 3D avatar reconstruction [4, 16, 53, 68, 69, 76] and generation [50, 52, 54]. Some methods [2, 49, 56, 65] exploit the powerful editing ability of GAN to edit a 3D static head portrait. However, they cannot be trivially extended to the dynamic avatars. The methods [37, 38, 45] focus on style transfer of the avatar using text prompt or style image but reach a poor identity-preserving. We propose a novel 3D avatar editing approach with an expression-aware modification generative model, which can be applied to various 3DMM-based volumetric avatars and render consistent novel views with fine-grained editing across multiple viewpoints and expression while preserving identity of person." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 459, + 362, + 471 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 459, + 362, + 471 + ], + "spans": [ + { + "bbox": [ + 306, + 459, + 362, + 471 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 479, + 547, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 547, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 547, + 647 + ], + "type": "text", + "content": "As shown in Fig. 2, given a volumetric head avatar, we edit the avatar using a single-view image and synthesize consistent novel views across multiple expression and viewpoints. To achieve this goal, we propose a novel expression-aware modification generator to generate 3D modification fields, which can be seamlessly integrated into various representations and animated with facial expression (see Sec. 3.2). Furthermore, to bootstrap the training with limited pairwise data, we propose a novel expression-aware distillation scheme to learn the expression-dependent modifications from large-scale generative models [7, 50] (see Sec. 3.3). During the editing process, given a single edited image of a 3D avatar, we perform an auto-decoding optimization to lift 2D editing effect to the 3D space (see Sec. 3.4)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 652, + 392, + 663 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 652, + 392, + 663 + ], + "spans": [ + { + "bbox": [ + 306, + 652, + 392, + 663 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "The current implicit volumetric representations of head avatar [16, 50, 76] are mostly built upon the NeRF [33] or its variants [8, 9, 31, 35, 44, 48]. In general, the neural architecture can be simplified as an implicit field " + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": " that" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8953" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 53, + 155, + 217 + ], + "blocks": [ + { + "bbox": [ + 55, + 53, + 155, + 217 + ], + "lines": [ + { + "bbox": [ + 55, + 53, + 155, + 217 + ], + "spans": [ + { + "bbox": [ + 55, + 53, + 155, + 217 + ], + "type": "image", + "image_path": "f924cb4273c9bd6a9b02ba279b7b66597d0d3146b5b1a78fbc4dc60523bdb6cb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 220, + 547, + 265 + ], + "lines": [ + { + "bbox": [ + 46, + 220, + 547, + 265 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 547, + 265 + ], + "type": "text", + "content": "Figure 2. We use an expression-aware generative model that accepts a modification latent code " + }, + { + "bbox": [ + 46, + 220, + 547, + 265 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{g / t}" + }, + { + "bbox": [ + 46, + 220, + 547, + 265 + ], + "type": "text", + "content": " and 3DMM coefficients and outputs a modification field of a tri-plane structure. The modification field modifies the geometry and texture of the template avatar by deforming the sample points " + }, + { + "bbox": [ + 46, + 220, + 547, + 265 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 220, + 547, + 265 + ], + "type": "text", + "content": " and blending the color " + }, + { + "bbox": [ + 46, + 220, + 547, + 265 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_o" + }, + { + "bbox": [ + 46, + 220, + 547, + 265 + ], + "type": "text", + "content": " with the modification color " + }, + { + "bbox": [ + 46, + 220, + 547, + 265 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{\\Delta}" + }, + { + "bbox": [ + 46, + 220, + 547, + 265 + ], + "type": "text", + "content": " respectively. We lift the 2D editing effect to 3D using an auto-decoding optimization and synthesize novel views across different expression." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 158, + 53, + 434, + 217 + ], + "blocks": [ + { + "bbox": [ + 158, + 53, + 434, + 217 + ], + "lines": [ + { + "bbox": [ + 158, + 53, + 434, + 217 + ], + "spans": [ + { + "bbox": [ + 158, + 53, + 434, + 217 + ], + "type": "image", + "image_path": "eba85a7295be3c26abebcef3021db96d255b1c5190a1691a3b83a9a22cd42097.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 434, + 53, + 536, + 217 + ], + "blocks": [ + { + "bbox": [ + 434, + 53, + 536, + 217 + ], + "lines": [ + { + "bbox": [ + 434, + 53, + 536, + 217 + ], + "spans": [ + { + "bbox": [ + 434, + 53, + 536, + 217 + ], + "type": "image", + "image_path": "527ac188974e9bd8d25909b1767c6a81f690304dd61d83fcb1618a945ca45345.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 271, + 288, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 271, + 288, + 319 + ], + "spans": [ + { + "bbox": [ + 47, + 271, + 288, + 319 + ], + "type": "text", + "content": "takes position " + }, + { + "bbox": [ + 47, + 271, + 288, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 47, + 271, + 288, + 319 + ], + "type": "text", + "content": " and view direction " + }, + { + "bbox": [ + 47, + 271, + 288, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 47, + 271, + 288, + 319 + ], + "type": "text", + "content": " as inputs and predicts the geometry " + }, + { + "bbox": [ + 47, + 271, + 288, + 319 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 47, + 271, + 288, + 319 + ], + "type": "text", + "content": " and the texture " + }, + { + "bbox": [ + 47, + 271, + 288, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 47, + 271, + 288, + 319 + ], + "type": "text", + "content": " of the avatar, i.e., " + }, + { + "bbox": [ + 47, + 271, + 288, + 319 + ], + "type": "inline_equation", + "content": "(\\sigma, \\mathbf{c}) = \\mathbf{F}(\\mathbf{x}, \\mathbf{d})" + }, + { + "bbox": [ + 47, + 271, + 288, + 319 + ], + "type": "text", + "content": ". Then, the volume rendering technique is used to render images as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 60, + 327, + 287, + 367 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 327, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 60, + 327, + 287, + 367 + ], + "type": "interline_equation", + "content": "\\hat {C} (\\boldsymbol {r}) = \\sum_ {i = 1} ^ {N} T _ {i} \\alpha_ {i} \\mathbf {c} _ {i}, \\quad T _ {i} = \\exp \\left(- \\sum_ {j = 1} ^ {i - 1} \\sigma^ {\\prime} _ {j} \\delta_ {j}\\right), \\tag {1}", + "image_path": "ede91bcdcba07ec1fbe16c633e8d48a407596493bd27b2dab2d2bd307fc5a403.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 374, + 288, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 374, + 288, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 374, + 288, + 422 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 374, + 288, + 422 + ], + "type": "inline_equation", + "content": "\\alpha_{i} = 1 - \\exp \\left(-\\sigma^{\\prime}_{i}\\delta_{i}\\right)" + }, + { + "bbox": [ + 46, + 374, + 288, + 422 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 374, + 288, + 422 + ], + "type": "inline_equation", + "content": "\\delta_{i}" + }, + { + "bbox": [ + 46, + 374, + 288, + 422 + ], + "type": "text", + "content": " is the distance between adjacent samples along the ray. In order to animate the head avatar, 3DMM [6] is incorporated to describe the deformation implicitly [15, 16, 57, 68, 69] or explicitly [4, 76]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 430, + 268, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 430, + 268, + 441 + ], + "spans": [ + { + "bbox": [ + 47, + 430, + 268, + 441 + ], + "type": "text", + "content": "3.2. Expression-aware Modification Generator" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 443, + 287, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 443, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 46, + 443, + 287, + 563 + ], + "type": "text", + "content": "To enable the modification animated with the facial expressions, we follow the architecture of 3DMM-based 3D GAN [50] to build our expression-aware modification generator. As shown in Fig. 2, our generator consists of a geometry generator " + }, + { + "bbox": [ + 46, + 443, + 287, + 563 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\Delta g}" + }, + { + "bbox": [ + 46, + 443, + 287, + 563 + ], + "type": "text", + "content": " and a texture generator " + }, + { + "bbox": [ + 46, + 443, + 287, + 563 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\Delta t}" + }, + { + "bbox": [ + 46, + 443, + 287, + 563 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 443, + 287, + 563 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\Delta g}" + }, + { + "bbox": [ + 46, + 443, + 287, + 563 + ], + "type": "text", + "content": " encodes the expression-dependent geometry modification by deforming the query points in the edited space to the original template space under each expression. " + }, + { + "bbox": [ + 46, + 443, + 287, + 563 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\Delta t}" + }, + { + "bbox": [ + 46, + 443, + 287, + 563 + ], + "type": "text", + "content": " encodes the expression-dependent modification color of query points under each expression:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 60, + 571, + 287, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 571, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 60, + 571, + 287, + 586 + ], + "type": "interline_equation", + "content": "\\mathbf {x} ^ {\\prime} = \\mathbf {G} _ {\\Delta g} (\\mathbf {x}, \\mathbf {z} _ {g}, \\mathbf {v}), \\left(\\mathbf {c} _ {\\Delta}, \\beta_ {\\Delta}\\right) = \\mathbf {G} _ {\\Delta t} (\\mathbf {x}, \\mathbf {z} _ {t}, \\mathbf {v}), \\tag {2}", + "image_path": "59d1bae86244496d15f32a59aecbda08322823829c0ec67e76ff3ee1be51d2ed.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " is the query points in the edited space, and " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}'" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " is the deformed point in the space of original avatar under current expression " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{e}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{\\Delta}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " is the modification color and " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\beta_{\\Delta}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " determines the blending weights with the original color. " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_g,\\mathbf{z}_t" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " are the geometry and texture modification latent code respectively, where " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_g,\\mathbf{z}_t\\in R^{1024}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": ". They control the generation of modification feature maps in the UV space. " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " is the 3DMM mesh vertices [26] that condition the current expression " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{e}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": ". Each mesh vertex has a neural feature that is retrieved in the modification feature map using pre-defined" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 271, + 545, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 271, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 304, + 271, + 545, + 380 + ], + "type": "text", + "content": "UV mapping. We rasterize the vertex features to the three axis-aligned planes to generate the tri-plane feature. The modification information of the query point " + }, + { + "bbox": [ + 304, + 271, + 545, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 271, + 545, + 380 + ], + "type": "text", + "content": " is first collected by bilinear interpolation on the tri-plane feature and then decoded by the neural feature decoder [8]. Since the modification is defined as decoupled fields without relying on the original field, our generated modification field can be integrated into various volumetric avatar representations and be animated following the facial expression." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 386, + 541, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 386, + 541, + 399 + ], + "spans": [ + { + "bbox": [ + 305, + 386, + 541, + 399 + ], + "type": "text", + "content": "3.3. Expression-dependent Modification Learning" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 401, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 401, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 545, + 533 + ], + "type": "text", + "content": "To learn the proposed expression-aware modification, we need extensive training data on avatars with a wide range of geometry and texture changes, which is hard to obtain in practice. Following the spirit of learning high-fidelity editing ability from the large-scale generative model [7, 18, 32, 74], we propose a novel expression-aware distillation scheme to deal with insufficient real training data. We leverage the ability of 3DMM-based 3D GAN [50] and 2D face texture editing tools to generate facial editing data, which encompasses a wide range of geometry and texture editing across various expression and viewpoints." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": "Geometry Distillation. We use the teacher 3DMM-based 3D GAN [50] " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_n" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": " to synthesize two volumetric avatars with different geometry (an original avatar " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": " and an edited avatar " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{\\prime}" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": ") by modifying the 3DMM shape parameter of the original avatar. This provides the paired editing data for our generator to learn how to modify the geometry of the avatar while maintaining consistency across various expression and viewpoints. Specifically, we randomly sample the latent code " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_n" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": " in the latent space of " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_n" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": " as well as 3DMM shape parameter " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": ", expression parameter " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": ", and pose parameter " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": " are sampled from a normal distribution whose absolute mean and standard deviation are within " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "[0,1]" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": " are a group of rotation vectors that have random directions within a unit sphere and magnitude within " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "[-6,6]" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": " degrees. Then, we sample an edit vector " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\beta_{\\Delta}" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": " from a uniform distribution" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8954" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "\\mathcal{U}(-3,3)" + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": " and apply it to the original shape parameter by " + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "\\beta' = \\beta_{\\Delta} + \\beta" + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": ". These hyperparameters w.r.t. 3DMM coefficients sampling are selected empirically to maintain the shape definition of the human head. Please refer to our supplementary Sec. B.2 for more details on 3DMM sampling. The original avatar " + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": " and paired edited avatar " + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "\\mathbf{F}'" + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": " are generated by " + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "\\mathbf{F} = \\mathbf{G}_n(\\mathbf{z}_n, \\boldsymbol{\\beta}, \\boldsymbol{\\psi}, \\boldsymbol{\\theta})" + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "\\mathbf{F}' = \\mathbf{G}(\\mathbf{z}_n, \\boldsymbol{\\beta}', \\boldsymbol{\\psi}, \\boldsymbol{\\theta})" + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": ". During training, we will apply our modification generator to modify the geometry of " + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "F'" + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": " render the face with the same geometry." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 195, + 289, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 195, + 289, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 195, + 289, + 460 + ], + "type": "text", + "content": "Texture Distillation. We distill the capabilities of fine-grained texture editing from 2D face editing algorithms by generating texture-modified avatar " + }, + { + "bbox": [ + 46, + 195, + 289, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{F}'" + }, + { + "bbox": [ + 46, + 195, + 289, + 460 + ], + "type": "text", + "content": " with the teacher 3DMM GAN [50] " + }, + { + "bbox": [ + 46, + 195, + 289, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_n" + }, + { + "bbox": [ + 46, + 195, + 289, + 460 + ], + "type": "text", + "content": ". Specifically, we sample an original avatar " + }, + { + "bbox": [ + 46, + 195, + 289, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{F} = \\mathbf{G}_n(\\mathbf{z}_n, \\beta, \\psi, \\theta)" + }, + { + "bbox": [ + 46, + 195, + 289, + 460 + ], + "type": "text", + "content": " from the teacher generator and render the image of its positive face. A segmentation-based 2D face texture editing algorithm (SBA) [77] and two makeup transfer algorithms (MTA) [22, 27, 36] are referred to in the distillation. We randomly choose one of them to edit the texture of the rendered face image. For SBA, we define several editable semantic regions of the face. A subset of these regions is selected randomly for texture painting using hues randomly sampled from the HSV color spectrum. For MTA, we randomly choose a makeup image as a reference from the open-sourced makeup dataset [22, 27, 36] and transfer the reference makeup to the rendered face image. The makeup dataset [36] contains complex makeup, such as blushes and makeup jewelry, which allow our generator to learn complicated texture editing patterns. Then, we perform the PTI inversion [43] on the texture-modified face image to lift the 2D texture editing to 3D space and obtain a texture-modified avatar " + }, + { + "bbox": [ + 46, + 195, + 289, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{F}'" + }, + { + "bbox": [ + 46, + 195, + 289, + 460 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "type": "text", + "content": "Modification Learning. Following the training style of StyleGAN [24], we sample a modification latent code " + }, + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{g / t} \\in R^{1024}" + }, + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "type": "text", + "content": " latent space for each paired editing data. We do not fully sample a 1024-dimensional modification code but sample a reduced code " + }, + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{z}}" + }, + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "type": "text", + "content": " from a standard normal distribution and concatenate it with the latent code " + }, + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_n" + }, + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "type": "text", + "content": " of the original avatar " + }, + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "type": "text", + "content": " that is sampled from the teacher model, i.e., " + }, + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{g / t} = (\\mathbf{z}_n, \\bar{\\mathbf{z}}), \\mathbf{z}_n, \\bar{\\mathbf{z}} \\in R^{512}" + }, + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "type": "text", + "content": ". This design is regarded as implicit code guidance that decently integrates knowledge from the teacher model to facilitate the model convergence. " + }, + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_n" + }, + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "type": "text", + "content": " encodes the facial appearances of avatar " + }, + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "type": "text", + "content": ", serving as a reference to the superimposition of the modification onto the avatar " + }, + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "type": "text", + "content": ". Note that during inference, we do not require the concatenation of latent code from the teacher model and directly optimize the full modification code from the edited image using an auto-decoding manner. Our generator generates the modification field following Eq. (2) where " + }, + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "type": "text", + "content": " is decoded from the 3DMM parameters of the avatar " + }, + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "type": "text", + "content": " using FLAME model [26] " + }, + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{v} = \\mathbf{E}(\\boldsymbol{\\beta}, \\boldsymbol{\\psi}, \\boldsymbol{\\theta})" + }, + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "type": "text", + "content": ". To apply the modification field, we feed the deformed query points " + }, + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}'" + }, + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "type": "text", + "content": " to the original avatar " + }, + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "type": "text", + "content": " to" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "obtain the density and color, and composite the color with modification color by:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 105, + 545, + 119 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 105, + 545, + 119 + ], + "spans": [ + { + "bbox": [ + 313, + 105, + 545, + 119 + ], + "type": "interline_equation", + "content": "\\mathbf {c} = \\left(1 - \\beta_ {\\Delta}\\right) * \\mathbf {c} _ {o} + \\beta_ {\\Delta} * \\mathbf {c} _ {\\Delta}, \\quad (\\sigma , \\mathbf {c} _ {o}) = \\mathbf {F} \\left(\\mathbf {x} ^ {\\prime}, \\mathbf {d}\\right). \\tag {3}", + "image_path": "ded75670220b3dc4313b376a957da6aa045dd3135116acb23e9f3275fc57e3de.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 127, + 547, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 127, + 547, + 187 + ], + "spans": [ + { + "bbox": [ + 305, + 127, + 547, + 187 + ], + "type": "text", + "content": "Then, we perform volume rendering on the density " + }, + { + "bbox": [ + 305, + 127, + 547, + 187 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 305, + 127, + 547, + 187 + ], + "type": "text", + "content": " and color " + }, + { + "bbox": [ + 305, + 127, + 547, + 187 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 305, + 127, + 547, + 187 + ], + "type": "text", + "content": " using Eq. (1) to render the modified image " + }, + { + "bbox": [ + 305, + 127, + 547, + 187 + ], + "type": "inline_equation", + "content": "\\hat{I}_e" + }, + { + "bbox": [ + 305, + 127, + 547, + 187 + ], + "type": "text", + "content": " of avatar " + }, + { + "bbox": [ + 305, + 127, + 547, + 187 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 305, + 127, + 547, + 187 + ], + "type": "text", + "content": ". We use the photometric loss to supervise the modified image with the rendered image " + }, + { + "bbox": [ + 305, + 127, + 547, + 187 + ], + "type": "inline_equation", + "content": "\\hat{I}'" + }, + { + "bbox": [ + 305, + 127, + 547, + 187 + ], + "type": "text", + "content": " from the edited avatar " + }, + { + "bbox": [ + 305, + 127, + 547, + 187 + ], + "type": "inline_equation", + "content": "\\mathbf{F}'" + }, + { + "bbox": [ + 305, + 127, + 547, + 187 + ], + "type": "text", + "content": " under the same camera parameters." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 391, + 194, + 545, + 209 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 194, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 391, + 194, + 545, + 209 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\left\\| \\hat {I} _ {e} - \\hat {I} ^ {\\prime} \\right\\| _ {2} ^ {2}. \\tag {4}", + "image_path": "da21bcf49775e05e63ddec63d7a766dcc73625631ee26729537b6011f345c401.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 217, + 545, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 217, + 545, + 253 + ], + "spans": [ + { + "bbox": [ + 305, + 217, + 545, + 253 + ], + "type": "text", + "content": "During training, we sample multiple viewpoints and 3DMM expression parameters " + }, + { + "bbox": [ + 305, + 217, + 545, + 253 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 305, + 217, + 545, + 253 + ], + "type": "text", + "content": " for each editing pair " + }, + { + "bbox": [ + 305, + 217, + 545, + 253 + ], + "type": "inline_equation", + "content": "(\\mathbf{F},\\mathbf{F}^{\\prime})" + }, + { + "bbox": [ + 305, + 217, + 545, + 253 + ], + "type": "text", + "content": " to enhance the spatial consistency under different expressions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 255, + 485, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 255, + 485, + 268 + ], + "spans": [ + { + "bbox": [ + 305, + 255, + 485, + 268 + ], + "type": "text", + "content": "3.4. Avatar Editing with Single Image" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 269, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 269, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 269, + 545, + 498 + ], + "type": "text", + "content": "In our task, users are allowed to edit a single image with various out-of-box face editing tools, such as Photoshop, drag-based editing [30, 39], text-driven editing [7]. For each editing input, we use the auto-decoding optimization on modification code to lift 2D edits into a 3D expression-aware modification field generated by our model. This field adapts to expression and viewpoint changes and is not tied to the specific avatar representation. The StyleGAN-based generator [24, 43, 55] features a latent space mapping from " + }, + { + "bbox": [ + 304, + 269, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{z} \\in R^{Z}" + }, + { + "bbox": [ + 304, + 269, + 545, + 498 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 304, + 269, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 304, + 269, + 545, + 498 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 269, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{w} \\in R^{W_{n} \\times W_{d}}" + }, + { + "bbox": [ + 304, + 269, + 545, + 498 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 304, + 269, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 304, + 269, + 545, + 498 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 269, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 304, + 269, + 545, + 498 + ], + "type": "text", + "content": " is more influential as it conditions the generator. Therefore, we perform code inversion in " + }, + { + "bbox": [ + 304, + 269, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 304, + 269, + 545, + 498 + ], + "type": "text", + "content": " space by randomly sampling a modification code " + }, + { + "bbox": [ + 304, + 269, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{g/t}" + }, + { + "bbox": [ + 304, + 269, + 545, + 498 + ], + "type": "text", + "content": " during editing. This code conditions a modification field " + }, + { + "bbox": [ + 304, + 269, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\Delta g/t}(\\mathbf{x}, \\mathbf{w}_{g/t}, \\mathbf{v})" + }, + { + "bbox": [ + 304, + 269, + 545, + 498 + ], + "type": "text", + "content": " following Eq. (2). We apply the modification field to the original avatar using Eq. (4). The modified image " + }, + { + "bbox": [ + 304, + 269, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\hat{I}_{e}" + }, + { + "bbox": [ + 304, + 269, + 545, + 498 + ], + "type": "text", + "content": " is rendered following the original avatar's rendering pipeline and is encouraged to match the user-edited image " + }, + { + "bbox": [ + 304, + 269, + 545, + 498 + ], + "type": "inline_equation", + "content": "I_{e}" + }, + { + "bbox": [ + 304, + 269, + 545, + 498 + ], + "type": "text", + "content": " by optimizing " + }, + { + "bbox": [ + 304, + 269, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{g/t}" + }, + { + "bbox": [ + 304, + 269, + 545, + 498 + ], + "type": "text", + "content": " with the following loss terms:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 309, + 506, + 545, + 531 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 506, + 545, + 531 + ], + "spans": [ + { + "bbox": [ + 309, + 506, + 545, + 531 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {i} = \\lambda_ {1} \\mathcal {L} _ {2} (\\hat {I} _ {e}, I _ {e}) + \\lambda_ {2} \\mathcal {L} _ {\\mathrm {l p i p s}} (\\hat {I} _ {e}, I _ {e}) + \\lambda_ {3} \\mathcal {L} _ {\\mathrm {r e g}} (\\mathbf {w}, \\mathbf {w} _ {\\mathrm {a v g}}), \\tag {5}", + "image_path": "b5361de40098f087d08eb6e3c885e164af5b512ba496e09756e755cc72060148.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 531, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 531, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 305, + 531, + 545, + 676 + ], + "type": "text", + "content": "The L2 loss term " + }, + { + "bbox": [ + 305, + 531, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_2" + }, + { + "bbox": [ + 305, + 531, + 545, + 676 + ], + "type": "text", + "content": " and LPIPS perceptual loss term " + }, + { + "bbox": [ + 305, + 531, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{lpips}}" + }, + { + "bbox": [ + 305, + 531, + 545, + 676 + ], + "type": "text", + "content": " encourage the rendered face close to the appearance and structure of the edited face. We set " + }, + { + "bbox": [ + 305, + 531, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\lambda_1 = 1000, \\lambda_2 = 1, \\lambda_3 = 1" + }, + { + "bbox": [ + 305, + 531, + 545, + 676 + ], + "type": "text", + "content": ". The regularization term " + }, + { + "bbox": [ + 305, + 531, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reg}}" + }, + { + "bbox": [ + 305, + 531, + 545, + 676 + ], + "type": "text", + "content": " applied to the latent code " + }, + { + "bbox": [ + 305, + 531, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 305, + 531, + 545, + 676 + ], + "type": "text", + "content": " enforces alignment with the distribution of the " + }, + { + "bbox": [ + 305, + 531, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 305, + 531, + 545, + 676 + ], + "type": "text", + "content": " space, with " + }, + { + "bbox": [ + 305, + 531, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{\\mathrm{avg}}" + }, + { + "bbox": [ + 305, + 531, + 545, + 676 + ], + "type": "text", + "content": " representing the mean latent code computed from 1000 random samples within the " + }, + { + "bbox": [ + 305, + 531, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 305, + 531, + 545, + 676 + ], + "type": "text", + "content": " space. Besides, we observe that the L2 loss on the whole face will give an underfitting result for the fine-grained makeup on the eyes, eyebrows and lips. Therefore, we reweight the L2 loss on the facial features by face segmentation mask " + }, + { + "bbox": [ + 305, + 531, + 545, + 676 + ], + "type": "inline_equation", + "content": "M = \\{N_i | i = 1, \\dots, m\\}" + }, + { + "bbox": [ + 305, + 531, + 545, + 676 + ], + "type": "text", + "content": " for texture editing:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 344, + 684, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 684, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 344, + 684, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {2} = \\sum_ {i = 1} ^ {m} \\sum_ {r \\in N _ {i}} \\frac {1}{| N _ {i} |} \\left\\| \\hat {C} _ {e} (r) - C _ {e} (r) \\right\\| _ {2} ^ {2}, \\tag {6}", + "image_path": "7cc1e001289d1036588f005afb33d66a8fc3066cd928d6e6d9a4d50e25817ef2.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8955" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 55, + 286, + 112 + ], + "blocks": [ + { + "bbox": [ + 49, + 55, + 286, + 112 + ], + "lines": [ + { + "bbox": [ + 49, + 55, + 286, + 112 + ], + "spans": [ + { + "bbox": [ + 49, + 55, + 286, + 112 + ], + "type": "table", + "html": "
MethodsGeometryTexture
RoopPVPNext3DOursRoopPVPNext3DOurs
Editing preservation ↑29.44%23.33%5.56%41.67%8.33%10.56%5.00%76.11%
Identity preservation ↑31.11%21.11%5.56%42.22%7.78%12.78%3.33%76.11%
Temporal consistency ↑30.00%23.89%4.44%41.67%5.56%12.78%1.67%80.00%
Overall ↑29.44%23.33%3.89%43.33%5.56%12.78%2.22%79.44%
image identity similarity ↑0.83730.87040.85470.88450.73200.84760.85000.9147
", + "image_path": "3c0cd1832132cae32f371e627d8cd9ead3046dd6070fc7d7093364b0b412071d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 117, + 287, + 139 + ], + "lines": [ + { + "bbox": [ + 47, + 117, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 47, + 117, + 287, + 139 + ], + "type": "text", + "content": "Table 1. We quantitatively compare with the PVP [28], Roop [12], Next3D [50] by user study and image identity similarity [13]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 150, + 289, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 150, + 289, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 150, + 289, + 319 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 150, + 289, + 319 + ], + "type": "inline_equation", + "content": "\\hat{C}_e(r), C_e(r)" + }, + { + "bbox": [ + 46, + 150, + 289, + 319 + ], + "type": "text", + "content": " are the rendered and target color respectively, " + }, + { + "bbox": [ + 46, + 150, + 289, + 319 + ], + "type": "inline_equation", + "content": "N_i" + }, + { + "bbox": [ + 46, + 150, + 289, + 319 + ], + "type": "text", + "content": " are the rays within the " + }, + { + "bbox": [ + 46, + 150, + 289, + 319 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 150, + 289, + 319 + ], + "type": "text", + "content": "-th semantic part of the face. Generally, we freeze the weight of the modification generator and only optimize the modification latent code w to reach a satisfied 3D modification result. When an intense makeup or complicated pattern is painted onto the human face, we will continuously fine-tune the weight of our generator and freeze the latent code w to achieve more accurate editing results. To animate the edited avatar, users can input the new 3DMM expression parameter to the original avatar and our modification generator simultaneously. In this way, the generated modification field tightly sticks to the original avatar and presents reasonable editing results under different expression and viewpoints." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 327, + 128, + 340 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 327, + 128, + 340 + ], + "spans": [ + { + "bbox": [ + 47, + 327, + 128, + 340 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 347, + 287, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 347, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 347, + 287, + 407 + ], + "type": "text", + "content": "In this section, we evaluate our avatar editing capability from a single perspective. One major difference with static NeRF editing is that we focus on showing how the edits are correctly lifted to 3D avatars under various expression and camera viewpoints." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 411, + 171, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 411, + 171, + 423 + ], + "spans": [ + { + "bbox": [ + 47, + 411, + 171, + 423 + ], + "type": "text", + "content": "4.1. Dataset and Baselines" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 426, + 287, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 557 + ], + "type": "text", + "content": "Datasets. We use a total of 19 neural implicit head avatars from three methods, i.e., 7 from INSTA [76], 8 from NeRF-BlendShape [16], 4 from Next3D [50], and show editing results on them. For INSTA [76] and NeRFBlendShape [16], we use the human head data (i.e., a monocular video of a head) provided by their methods to reconstruct the volumetric avatar using their respective representations. For Next3D [50], we random sample its latent space to generate volumetric avatars and perform editing on them. The evaluation datasets exhibit a substantial variation in identities, encompassing a diverse range of races, ages, and genders." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "text", + "content": "Baselines. We pick several baseline methods that can support single-view-based avatar editing. Roop [12] is a face-swapping method that can swap the human face in a video from a single reference view. To compare with Roop, we generate videos of the original avatar rendered in driving signals and single edited frames, and perform face swap. PVP [28] learns a personalized avatar image generator from a monocular video by fine-tuning the latent space of StyleGAN [24], and performs GAN-inversion style optimization [41] to edit the shape and appearance of the avatar. Next3D [50] is a 3DMM-based 3D GAN. To make a fair comparison on the same input data (i.e., a monocular video of avatar and an edited image), we perform GAN inver" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 310, + 51, + 351, + 98 + ], + "blocks": [ + { + "bbox": [ + 310, + 51, + 351, + 98 + ], + "lines": [ + { + "bbox": [ + 310, + 51, + 351, + 98 + ], + "spans": [ + { + "bbox": [ + 310, + 51, + 351, + 98 + ], + "type": "image", + "image_path": "43bbe264c875fe4dc42c506fb6c8a978b42bcd69db9f9f7e040985c836fe9a9f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 101, + 342, + 142 + ], + "blocks": [ + { + "bbox": [ + 310, + 101, + 342, + 142 + ], + "lines": [ + { + "bbox": [ + 310, + 101, + 342, + 142 + ], + "spans": [ + { + "bbox": [ + 310, + 101, + 342, + 142 + ], + "type": "image", + "image_path": "3b5f7267aed76fbc90f89380c0630a0aa93e792934409ac98b04c6464f4009fa.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 143, + 351, + 156 + ], + "lines": [ + { + "bbox": [ + 310, + 143, + 351, + 156 + ], + "spans": [ + { + "bbox": [ + 310, + 143, + 351, + 156 + ], + "type": "text", + "content": "Hair++, Eyes++, Mouse" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 353, + 50, + 404, + 95 + ], + "blocks": [ + { + "bbox": [ + 353, + 50, + 404, + 95 + ], + "lines": [ + { + "bbox": [ + 353, + 50, + 404, + 95 + ], + "spans": [ + { + "bbox": [ + 353, + 50, + 404, + 95 + ], + "type": "image", + "image_path": "786fd0f0a878b9e63eab588746a0ecf80e36ab8dd8c910fb3c704666189f0bb4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 353, + 95, + 424, + 111 + ], + "blocks": [ + { + "bbox": [ + 353, + 95, + 424, + 111 + ], + "lines": [ + { + "bbox": [ + 353, + 95, + 424, + 111 + ], + "spans": [ + { + "bbox": [ + 353, + 95, + 424, + 111 + ], + "type": "image", + "image_path": "cf11cf46baf36d4d082e737f1a829d7539e40a5435a39547ee4070fef0e1eeab.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 353, + 112, + 405, + 157 + ], + "blocks": [ + { + "bbox": [ + 353, + 112, + 405, + 157 + ], + "lines": [ + { + "bbox": [ + 353, + 112, + 405, + 157 + ], + "spans": [ + { + "bbox": [ + 353, + 112, + 405, + 157 + ], + "type": "image", + "image_path": "f1872ea41b792926f5f3629b6e6567d860f1aa245ae580afa212ad58d2f0dfc9.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 408, + 50, + 462, + 95 + ], + "blocks": [ + { + "bbox": [ + 408, + 50, + 462, + 95 + ], + "lines": [ + { + "bbox": [ + 408, + 50, + 462, + 95 + ], + "spans": [ + { + "bbox": [ + 408, + 50, + 462, + 95 + ], + "type": "image", + "image_path": "3428f2123db18035433243704f6d6d4d11708e2e0a46f9bf121a567250b8ce34.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 411, + 95, + 462, + 111 + ], + "blocks": [ + { + "bbox": [ + 411, + 95, + 462, + 111 + ], + "lines": [ + { + "bbox": [ + 411, + 95, + 462, + 111 + ], + "spans": [ + { + "bbox": [ + 411, + 95, + 462, + 111 + ], + "type": "image", + "image_path": "3582093e97877d4d5a7bd0dd3b667ba7d0db2b5279fd6ac022bcef59fcaacf12.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 467, + 51, + 498, + 95 + ], + "blocks": [ + { + "bbox": [ + 467, + 51, + 498, + 95 + ], + "lines": [ + { + "bbox": [ + 467, + 51, + 498, + 95 + ], + "spans": [ + { + "bbox": [ + 467, + 51, + 498, + 95 + ], + "type": "image", + "image_path": "b84f3ed2e81320d68371589cdebb5b812c4124e2d47711908666447b83e7b9e9.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 466, + 95, + 500, + 111 + ], + "blocks": [ + { + "bbox": [ + 466, + 95, + 500, + 111 + ], + "lines": [ + { + "bbox": [ + 466, + 95, + 500, + 111 + ], + "spans": [ + { + "bbox": [ + 466, + 95, + 500, + 111 + ], + "type": "image", + "image_path": "d6ee9dd647e17db201d638acd8edc83ebd77e3c24fe38d288159308f95b809ad.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 503, + 51, + 533, + 95 + ], + "blocks": [ + { + "bbox": [ + 503, + 51, + 533, + 95 + ], + "lines": [ + { + "bbox": [ + 503, + 51, + 533, + 95 + ], + "spans": [ + { + "bbox": [ + 503, + 51, + 533, + 95 + ], + "type": "image", + "image_path": "ddad9fc09aa57dee0b664813232dff2635803507a570c156c59cdcd5f45f23de.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 503, + 95, + 536, + 111 + ], + "blocks": [ + { + "bbox": [ + 503, + 95, + 536, + 111 + ], + "lines": [ + { + "bbox": [ + 503, + 95, + 536, + 111 + ], + "spans": [ + { + "bbox": [ + 503, + 95, + 536, + 111 + ], + "type": "image", + "image_path": "320f79daa0c6adc0b20bbd09f87a86ff090e84ee04597f21ff749b1f70595a75.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 311, + 158, + 343, + 208 + ], + "blocks": [ + { + "bbox": [ + 311, + 158, + 343, + 208 + ], + "lines": [ + { + "bbox": [ + 311, + 158, + 343, + 208 + ], + "spans": [ + { + "bbox": [ + 311, + 158, + 343, + 208 + ], + "type": "image", + "image_path": "9c4d61b461c9bc294ee5c35bae04ffb8722fccba24a2b9fd8ad8221c9c884b56.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 310, + 208, + 343, + 257 + ], + "blocks": [ + { + "bbox": [ + 310, + 208, + 343, + 257 + ], + "lines": [ + { + "bbox": [ + 310, + 208, + 343, + 257 + ], + "spans": [ + { + "bbox": [ + 310, + 208, + 343, + 257 + ], + "type": "image", + "image_path": "b172bd92bdd7480f5fb4f432d5372048e3e458268ff6519d6a3177af230f5ebc.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 355, + 158, + 387, + 204 + ], + "blocks": [ + { + "bbox": [ + 355, + 158, + 387, + 204 + ], + "lines": [ + { + "bbox": [ + 355, + 158, + 387, + 204 + ], + "spans": [ + { + "bbox": [ + 355, + 158, + 387, + 204 + ], + "type": "image", + "image_path": "cc37aae23a265628ef1e8214f5400399ac17452f086fc21d9978bf7e80bb2417.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 353, + 204, + 389, + 220 + ], + "blocks": [ + { + "bbox": [ + 353, + 204, + 389, + 220 + ], + "lines": [ + { + "bbox": [ + 353, + 204, + 389, + 220 + ], + "spans": [ + { + "bbox": [ + 353, + 204, + 389, + 220 + ], + "type": "image", + "image_path": "8d9b606642f75a4de68acafcbafaaf7bf581d2bfb4e9e6e0e6ea588cc816c69a.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 392, + 158, + 425, + 204 + ], + "blocks": [ + { + "bbox": [ + 392, + 158, + 425, + 204 + ], + "lines": [ + { + "bbox": [ + 392, + 158, + 425, + 204 + ], + "spans": [ + { + "bbox": [ + 392, + 158, + 425, + 204 + ], + "type": "image", + "image_path": "73bfa2b7e1dfb4edd5bcf8e0941fbe4fb555e89db671424ebb5b55170dd14b4d.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 392, + 204, + 424, + 220 + ], + "blocks": [ + { + "bbox": [ + 392, + 204, + 424, + 220 + ], + "lines": [ + { + "bbox": [ + 392, + 204, + 424, + 220 + ], + "spans": [ + { + "bbox": [ + 392, + 204, + 424, + 220 + ], + "type": "image", + "image_path": "5fa147eda84fe98bbdc0733ca29629021f620c9a2765335de70f117ab1eec5b2.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 428, + 158, + 459, + 204 + ], + "blocks": [ + { + "bbox": [ + 428, + 158, + 459, + 204 + ], + "lines": [ + { + "bbox": [ + 428, + 158, + 459, + 204 + ], + "spans": [ + { + "bbox": [ + 428, + 158, + 459, + 204 + ], + "type": "image", + "image_path": "15c450b3883804d63d2659e01ce3b8ca0d27e38fe587d9954da3a2efb8c61734.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 428, + 204, + 463, + 220 + ], + "blocks": [ + { + "bbox": [ + 428, + 204, + 463, + 220 + ], + "lines": [ + { + "bbox": [ + 428, + 204, + 463, + 220 + ], + "spans": [ + { + "bbox": [ + 428, + 204, + 463, + 220 + ], + "type": "image", + "image_path": "097146ef864b05f192edd0440c381450d72919bc7524c779c793d63f593cd816.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 485, + 158, + 535, + 204 + ], + "blocks": [ + { + "bbox": [ + 485, + 158, + 535, + 204 + ], + "lines": [ + { + "bbox": [ + 485, + 158, + 535, + 204 + ], + "spans": [ + { + "bbox": [ + 485, + 158, + 535, + 204 + ], + "type": "image", + "image_path": "95b827a23a7c086a6528271f59bfe83d88718c13b13dc0f9d65ed854fd8473f0.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 466, + 204, + 501, + 220 + ], + "blocks": [ + { + "bbox": [ + 466, + 204, + 501, + 220 + ], + "lines": [ + { + "bbox": [ + 466, + 204, + 501, + 220 + ], + "spans": [ + { + "bbox": [ + 466, + 204, + 501, + 220 + ], + "type": "image", + "image_path": "545ff9b4f7f404d803c96ff05aae0dcf6137497646caabd7d46427fe8295eb03.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 504, + 158, + 536, + 204 + ], + "blocks": [ + { + "bbox": [ + 504, + 158, + 536, + 204 + ], + "lines": [ + { + "bbox": [ + 504, + 158, + 536, + 204 + ], + "spans": [ + { + "bbox": [ + 504, + 158, + 536, + 204 + ], + "type": "image", + "image_path": "62424f6e28e96a5024f72d5b028da3da1241069b9eb2b2bcfb98c52f40a66356.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 504, + 204, + 539, + 220 + ], + "blocks": [ + { + "bbox": [ + 504, + 204, + 539, + 220 + ], + "lines": [ + { + "bbox": [ + 504, + 204, + 539, + 220 + ], + "spans": [ + { + "bbox": [ + 504, + 204, + 539, + 220 + ], + "type": "image", + "image_path": "a4276534daa3eb40310b271f3f47743fd7d664b00f58e13c3b74591fb8cd52f3.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 311, + 270, + 343, + 319 + ], + "blocks": [ + { + "bbox": [ + 319, + 258, + 338, + 270 + ], + "lines": [ + { + "bbox": [ + 319, + 258, + 338, + 270 + ], + "spans": [ + { + "bbox": [ + 319, + 258, + 338, + 270 + ], + "type": "text", + "content": "Nose-, Lip++" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 311, + 270, + 343, + 319 + ], + "lines": [ + { + "bbox": [ + 311, + 270, + 343, + 319 + ], + "spans": [ + { + "bbox": [ + 311, + 270, + 343, + 319 + ], + "type": "image", + "image_path": "77764d4e5ce0b53bbf8b92f29ce3a0c74862dd6247e63aa924ca21229a625b63.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 310, + 319, + 344, + 369 + ], + "blocks": [ + { + "bbox": [ + 310, + 319, + 344, + 369 + ], + "lines": [ + { + "bbox": [ + 310, + 319, + 344, + 369 + ], + "spans": [ + { + "bbox": [ + 310, + 319, + 344, + 369 + ], + "type": "image", + "image_path": "e343c5cfc35b8bc98da8cc93c4f712ed9a4f5ffb1e93fda5798231f45557638b.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "bbox": [ + 316, + 372, + 345, + 379 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 345, + 379 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 345, + 379 + ], + "type": "text", + "content": "ForeHead--" + } + ] + } + ], + "index": 35, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 356, + 221, + 388, + 269 + ], + "blocks": [ + { + "bbox": [ + 356, + 221, + 388, + 269 + ], + "lines": [ + { + "bbox": [ + 356, + 221, + 388, + 269 + ], + "spans": [ + { + "bbox": [ + 356, + 221, + 388, + 269 + ], + "type": "image", + "image_path": "e599c41ff5a5c6c4ffd7f8c148c18ba330cd054fe3a6ea520397f8cea250430b.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 356, + 270, + 389, + 316 + ], + "blocks": [ + { + "bbox": [ + 356, + 270, + 389, + 316 + ], + "lines": [ + { + "bbox": [ + 356, + 270, + 389, + 316 + ], + "spans": [ + { + "bbox": [ + 356, + 270, + 389, + 316 + ], + "type": "image", + "image_path": "433eee09d0af752d581b901ca537b87053ba59982834cb1afd541e8af7e1cc00.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 394, + 222, + 426, + 268 + ], + "blocks": [ + { + "bbox": [ + 394, + 222, + 426, + 268 + ], + "lines": [ + { + "bbox": [ + 394, + 222, + 426, + 268 + ], + "spans": [ + { + "bbox": [ + 394, + 222, + 426, + 268 + ], + "type": "image", + "image_path": "c7e387714b649051414e74ebb46e0f381cf237632e68a1711cd4ae104837647c.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 394, + 270, + 427, + 316 + ], + "blocks": [ + { + "bbox": [ + 394, + 270, + 427, + 316 + ], + "lines": [ + { + "bbox": [ + 394, + 270, + 427, + 316 + ], + "spans": [ + { + "bbox": [ + 394, + 270, + 427, + 316 + ], + "type": "image", + "image_path": "4c8496bd89347c4ac6a8187b189141f0ad25e9d0b0fc8af82257161c1df186a7.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 430, + 222, + 461, + 268 + ], + "blocks": [ + { + "bbox": [ + 430, + 222, + 461, + 268 + ], + "lines": [ + { + "bbox": [ + 430, + 222, + 461, + 268 + ], + "spans": [ + { + "bbox": [ + 430, + 222, + 461, + 268 + ], + "type": "image", + "image_path": "4bf9e90bb7e4ee5aefb9ad000f83880f90548dbe158956b074d4841bc95df597.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 430, + 270, + 463, + 316 + ], + "blocks": [ + { + "bbox": [ + 430, + 270, + 463, + 316 + ], + "lines": [ + { + "bbox": [ + 430, + 270, + 463, + 316 + ], + "spans": [ + { + "bbox": [ + 430, + 270, + 463, + 316 + ], + "type": "image", + "image_path": "0ef874137b0767838154229871914045dc291d7a89d872be87657e2574e41ccc.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 468, + 222, + 498, + 268 + ], + "blocks": [ + { + "bbox": [ + 468, + 222, + 498, + 268 + ], + "lines": [ + { + "bbox": [ + 468, + 222, + 498, + 268 + ], + "spans": [ + { + "bbox": [ + 468, + 222, + 498, + 268 + ], + "type": "image", + "image_path": "0a2152595ce87420e6934bee87e2b3eeb54938e03e4d92ba2230961217520011.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 467, + 270, + 500, + 315 + ], + "blocks": [ + { + "bbox": [ + 467, + 270, + 500, + 315 + ], + "lines": [ + { + "bbox": [ + 467, + 270, + 500, + 315 + ], + "spans": [ + { + "bbox": [ + 467, + 270, + 500, + 315 + ], + "type": "image", + "image_path": "303976365356cdcd402ee5509a97778b231b9f94a20ce7a7e573d45f1904ee43.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 504, + 222, + 536, + 268 + ], + "blocks": [ + { + "bbox": [ + 504, + 222, + 536, + 268 + ], + "lines": [ + { + "bbox": [ + 504, + 222, + 536, + 268 + ], + "spans": [ + { + "bbox": [ + 504, + 222, + 536, + 268 + ], + "type": "image", + "image_path": "5d3e0e2362c64a7f915cf5b6a9362e5f26153b12487181d30ce8818b25cf7728.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 504, + 270, + 535, + 315 + ], + "blocks": [ + { + "bbox": [ + 504, + 270, + 535, + 315 + ], + "lines": [ + { + "bbox": [ + 504, + 270, + 535, + 315 + ], + "spans": [ + { + "bbox": [ + 504, + 270, + 535, + 315 + ], + "type": "image", + "image_path": "d041762c5b6d3f7aede17747b09e0b42fae10dfc74891eb89698fc71a5641f4f.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 354, + 317, + 388, + 333 + ], + "blocks": [ + { + "bbox": [ + 354, + 317, + 388, + 333 + ], + "lines": [ + { + "bbox": [ + 354, + 317, + 388, + 333 + ], + "spans": [ + { + "bbox": [ + 354, + 317, + 388, + 333 + ], + "type": "image", + "image_path": "f1044c94474c383d609feead99c75b0dbc0c4f8e56c933326cbaa7c123f0d0cc.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 392, + 317, + 425, + 333 + ], + "blocks": [ + { + "bbox": [ + 392, + 317, + 425, + 333 + ], + "lines": [ + { + "bbox": [ + 392, + 317, + 425, + 333 + ], + "spans": [ + { + "bbox": [ + 392, + 317, + 425, + 333 + ], + "type": "image", + "image_path": "137be98d9d39f9494c4ff6735b8f9e96ccecc2f250901dc8e4c4baf5ab1faae9.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + } + ], + "index": 47 + }, + { + "type": "image", + "bbox": [ + 428, + 317, + 463, + 333 + ], + "blocks": [ + { + "bbox": [ + 428, + 317, + 463, + 333 + ], + "lines": [ + { + "bbox": [ + 428, + 317, + 463, + 333 + ], + "spans": [ + { + "bbox": [ + 428, + 317, + 463, + 333 + ], + "type": "image", + "image_path": "62b9861c4201e18303d031ef72fd9ca9664776cb3f80297a8879638ce7beac79.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 467, + 317, + 501, + 333 + ], + "blocks": [ + { + "bbox": [ + 467, + 317, + 501, + 333 + ], + "lines": [ + { + "bbox": [ + 467, + 317, + 501, + 333 + ], + "spans": [ + { + "bbox": [ + 467, + 317, + 501, + 333 + ], + "type": "image", + "image_path": "186b222fd37d76cabe5ba45a4f44f43d7ec07d595432c00d32b7db8325f97454.jpg" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_body" + } + ], + "index": 49 + }, + { + "type": "image", + "bbox": [ + 504, + 317, + 539, + 333 + ], + "blocks": [ + { + "bbox": [ + 504, + 317, + 539, + 333 + ], + "lines": [ + { + "bbox": [ + 504, + 317, + 539, + 333 + ], + "spans": [ + { + "bbox": [ + 504, + 317, + 539, + 333 + ], + "type": "image", + "image_path": "864e9deda6c8595f366462ba94874f501538d386551e6e8189945a824fb97f13.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + } + ], + "index": 50 + }, + { + "bbox": [ + 310, + 334, + 343, + 369 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 334, + 343, + 369 + ], + "spans": [ + { + "bbox": [ + 310, + 334, + 343, + 369 + ], + "type": "image", + "image_path": "edcc2c8108cba756a74b80af5d322535b22155b8084b3dce3959b873d0506ec3.jpg" + } + ] + } + ], + "index": 51, + "type": "text" + }, + { + "bbox": [ + 316, + 372, + 345, + 379 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 345, + 379 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 345, + 379 + ], + "type": "text", + "content": "ForeHead--" + } + ] + } + ], + "index": 52, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 356, + 335, + 389, + 380 + ], + "blocks": [ + { + "bbox": [ + 356, + 335, + 389, + 380 + ], + "lines": [ + { + "bbox": [ + 356, + 335, + 389, + 380 + ], + "spans": [ + { + "bbox": [ + 356, + 335, + 389, + 380 + ], + "type": "image", + "image_path": "e8466da5d17f01e6573589d016afa53107bbb8b7ddf94167d77cbdb6a141b79a.jpg" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_body" + } + ], + "index": 53 + }, + { + "type": "image", + "bbox": [ + 356, + 380, + 389, + 429 + ], + "blocks": [ + { + "bbox": [ + 356, + 380, + 389, + 429 + ], + "lines": [ + { + "bbox": [ + 356, + 380, + 389, + 429 + ], + "spans": [ + { + "bbox": [ + 356, + 380, + 389, + 429 + ], + "type": "image", + "image_path": "6119cb31199c31852d9ecddbb8275798e5b61c478bdb6d7e59bd2e6aad691475.jpg" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_body" + } + ], + "index": 54 + }, + { + "type": "image", + "bbox": [ + 394, + 335, + 427, + 380 + ], + "blocks": [ + { + "bbox": [ + 394, + 335, + 427, + 380 + ], + "lines": [ + { + "bbox": [ + 394, + 335, + 427, + 380 + ], + "spans": [ + { + "bbox": [ + 394, + 335, + 427, + 380 + ], + "type": "image", + "image_path": "78fb73bac479c0a577c1a4689bdd01c7b21ac0d1f1b61cdf075f338a26b8c244.jpg" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_body" + } + ], + "index": 55 + }, + { + "type": "image", + "bbox": [ + 394, + 380, + 426, + 429 + ], + "blocks": [ + { + "bbox": [ + 394, + 380, + 426, + 429 + ], + "lines": [ + { + "bbox": [ + 394, + 380, + 426, + 429 + ], + "spans": [ + { + "bbox": [ + 394, + 380, + 426, + 429 + ], + "type": "image", + "image_path": "e32b9abf2967bef31e638841f549e659f85dbcbe383dc863508f74edbc581560.jpg" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_body" + } + ], + "index": 56 + }, + { + "type": "image", + "bbox": [ + 431, + 335, + 464, + 380 + ], + "blocks": [ + { + "bbox": [ + 431, + 335, + 464, + 380 + ], + "lines": [ + { + "bbox": [ + 431, + 335, + 464, + 380 + ], + "spans": [ + { + "bbox": [ + 431, + 335, + 464, + 380 + ], + "type": "image", + "image_path": "b5190ed077d108b883734f4cb89a49043dfa789174838bd18057d1074e5fe878.jpg" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_body" + } + ], + "index": 57 + }, + { + "type": "image", + "bbox": [ + 432, + 382, + 463, + 429 + ], + "blocks": [ + { + "bbox": [ + 432, + 382, + 463, + 429 + ], + "lines": [ + { + "bbox": [ + 432, + 382, + 463, + 429 + ], + "spans": [ + { + "bbox": [ + 432, + 382, + 463, + 429 + ], + "type": "image", + "image_path": "e31bdf143f5a053b03cf5f5194beb4b2adceee34721a0b38251a7cb80d38bae5.jpg" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_body" + } + ], + "index": 58 + }, + { + "type": "image", + "bbox": [ + 468, + 336, + 500, + 380 + ], + "blocks": [ + { + "bbox": [ + 468, + 336, + 500, + 380 + ], + "lines": [ + { + "bbox": [ + 468, + 336, + 500, + 380 + ], + "spans": [ + { + "bbox": [ + 468, + 336, + 500, + 380 + ], + "type": "image", + "image_path": "bc3ecca9b0e87ddfb2ff4e88ba23117a2302de9ad39402887453bae38bd15010.jpg" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_body" + } + ], + "index": 59 + }, + { + "type": "image", + "bbox": [ + 503, + 336, + 536, + 380 + ], + "blocks": [ + { + "bbox": [ + 503, + 336, + 536, + 380 + ], + "lines": [ + { + "bbox": [ + 503, + 336, + 536, + 380 + ], + "spans": [ + { + "bbox": [ + 503, + 336, + 536, + 380 + ], + "type": "image", + "image_path": "0e020db4fc3fcec971db7d04b39a60adf2935d1d546d66f499214514822c3f21.jpg" + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_body" + } + ], + "index": 60 + }, + { + "type": "image", + "bbox": [ + 353, + 430, + 388, + 446 + ], + "blocks": [ + { + "bbox": [ + 353, + 430, + 388, + 446 + ], + "lines": [ + { + "bbox": [ + 353, + 430, + 388, + 446 + ], + "spans": [ + { + "bbox": [ + 353, + 430, + 388, + 446 + ], + "type": "image", + "image_path": "a4115f15bb39d540bed289fd1895e05c190a30d46de764b516f38a0731090895.jpg" + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "image_body" + } + ], + "index": 61 + }, + { + "type": "image", + "bbox": [ + 356, + 446, + 389, + 489 + ], + "blocks": [ + { + "bbox": [ + 356, + 446, + 389, + 489 + ], + "lines": [ + { + "bbox": [ + 356, + 446, + 389, + 489 + ], + "spans": [ + { + "bbox": [ + 356, + 446, + 389, + 489 + ], + "type": "image", + "image_path": "1d4e24d3f38814b6b71574a413b355b5264d28b43d137e35a4964c45b814c2b6.jpg" + } + ] + } + ], + "index": 62, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 311, + 492, + 348, + 503 + ], + "lines": [ + { + "bbox": [ + 311, + 492, + 348, + 503 + ], + "spans": [ + { + "bbox": [ + 311, + 492, + 348, + 503 + ], + "type": "text", + "content": "2D Editing" + } + ] + } + ], + "index": 71, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 357, + 490, + 391, + 507 + ], + "lines": [ + { + "bbox": [ + 357, + 490, + 391, + 507 + ], + "spans": [ + { + "bbox": [ + 357, + 490, + 391, + 507 + ], + "type": "text", + "content": "Reference Animation" + } + ] + } + ], + "index": 72, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 506, + 545, + 551 + ], + "lines": [ + { + "bbox": [ + 306, + 506, + 545, + 551 + ], + "spans": [ + { + "bbox": [ + 306, + 506, + 545, + 551 + ], + "type": "text", + "content": "Figure 3. We compare geometry editing with PVP [28], Roop [12], Next3D [50] on INSTA [76] and NeRFBlendshape [16] avatars. The \"Reference Animation\" denotes the image of the original avatar under the same expression with the rendered edited view." + } + ] + } + ], + "index": 77, + "angle": 0, + "type": "image_caption" + } + ], + "index": 62 + }, + { + "type": "image", + "bbox": [ + 392, + 430, + 425, + 446 + ], + "blocks": [ + { + "bbox": [ + 392, + 430, + 425, + 446 + ], + "lines": [ + { + "bbox": [ + 392, + 430, + 425, + 446 + ], + "spans": [ + { + "bbox": [ + 392, + 430, + 425, + 446 + ], + "type": "image", + "image_path": "e790a4536369185f73ebdebc5f0426bf2622c17a280a93ebb85fb9a7437dac1b.jpg" + } + ] + } + ], + "index": 63, + "angle": 0, + "type": "image_body" + } + ], + "index": 63 + }, + { + "type": "image", + "bbox": [ + 395, + 447, + 424, + 489 + ], + "blocks": [ + { + "bbox": [ + 395, + 447, + 424, + 489 + ], + "lines": [ + { + "bbox": [ + 395, + 447, + 424, + 489 + ], + "spans": [ + { + "bbox": [ + 395, + 447, + 424, + 489 + ], + "type": "image", + "image_path": "f8b85e621ee8c72b2701d193d5b3253e043a529ed520955369ddeb50d53f1031.jpg" + } + ] + } + ], + "index": 64, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 495, + 419, + 502 + ], + "lines": [ + { + "bbox": [ + 403, + 495, + 419, + 502 + ], + "spans": [ + { + "bbox": [ + 403, + 495, + 419, + 502 + ], + "type": "text", + "content": "PVP" + } + ] + } + ], + "index": 73, + "angle": 0, + "type": "image_caption" + } + ], + "index": 64 + }, + { + "type": "image", + "bbox": [ + 431, + 430, + 463, + 446 + ], + "blocks": [ + { + "bbox": [ + 431, + 430, + 463, + 446 + ], + "lines": [ + { + "bbox": [ + 431, + 430, + 463, + 446 + ], + "spans": [ + { + "bbox": [ + 431, + 430, + 463, + 446 + ], + "type": "image", + "image_path": "34ddfcb176d2747a0a68641541ce75f078962dc5082bcfb73be22f4d8dc03b61.jpg" + } + ] + } + ], + "index": 65, + "angle": 0, + "type": "image_body" + } + ], + "index": 65 + }, + { + "type": "image", + "bbox": [ + 430, + 447, + 461, + 489 + ], + "blocks": [ + { + "bbox": [ + 430, + 447, + 461, + 489 + ], + "lines": [ + { + "bbox": [ + 430, + 447, + 461, + 489 + ], + "spans": [ + { + "bbox": [ + 430, + 447, + 461, + 489 + ], + "type": "image", + "image_path": "a653d4bbd77d58931dcadf6deba500da701bfd11db45fc7eb314d141c5a641c6.jpg" + } + ] + } + ], + "index": 66, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 436, + 495, + 454, + 502 + ], + "lines": [ + { + "bbox": [ + 436, + 495, + 454, + 502 + ], + "spans": [ + { + "bbox": [ + 436, + 495, + 454, + 502 + ], + "type": "text", + "content": "Roop" + } + ] + } + ], + "index": 74, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 471, + 495, + 496, + 502 + ], + "lines": [ + { + "bbox": [ + 471, + 495, + 496, + 502 + ], + "spans": [ + { + "bbox": [ + 471, + 495, + 496, + 502 + ], + "type": "text", + "content": "Next3D" + } + ] + } + ], + "index": 75, + "angle": 0, + "type": "image_caption" + } + ], + "index": 66 + }, + { + "type": "image", + "bbox": [ + 470, + 382, + 500, + 429 + ], + "blocks": [ + { + "bbox": [ + 470, + 382, + 500, + 429 + ], + "lines": [ + { + "bbox": [ + 470, + 382, + 500, + 429 + ], + "spans": [ + { + "bbox": [ + 470, + 382, + 500, + 429 + ], + "type": "image", + "image_path": "d6697f8dbf19f3dfbb45a456787d167f8012a432afa93307f33601d781409d5e.jpg" + } + ] + } + ], + "index": 67, + "angle": 0, + "type": "image_body" + } + ], + "index": 67 + }, + { + "type": "image", + "bbox": [ + 467, + 430, + 500, + 446 + ], + "blocks": [ + { + "bbox": [ + 467, + 430, + 500, + 446 + ], + "lines": [ + { + "bbox": [ + 467, + 430, + 500, + 446 + ], + "spans": [ + { + "bbox": [ + 467, + 430, + 500, + 446 + ], + "type": "image", + "image_path": "f2455b140aa55f770aa54608605506466a3354d0e990ca92987987f53727990e.jpg" + } + ] + } + ], + "index": 68, + "angle": 0, + "type": "image_body" + } + ], + "index": 68 + }, + { + "type": "image", + "bbox": [ + 503, + 382, + 535, + 429 + ], + "blocks": [ + { + "bbox": [ + 503, + 382, + 535, + 429 + ], + "lines": [ + { + "bbox": [ + 503, + 382, + 535, + 429 + ], + "spans": [ + { + "bbox": [ + 503, + 382, + 535, + 429 + ], + "type": "image", + "image_path": "c7fa26a7185bbf1e5f23cdc73c2a0c3c826e12c5c067789d011820177a272215.jpg" + } + ] + } + ], + "index": 69, + "angle": 0, + "type": "image_body" + } + ], + "index": 69 + }, + { + "type": "image", + "bbox": [ + 504, + 430, + 539, + 446 + ], + "blocks": [ + { + "bbox": [ + 504, + 430, + 539, + 446 + ], + "lines": [ + { + "bbox": [ + 504, + 430, + 539, + 446 + ], + "spans": [ + { + "bbox": [ + 504, + 430, + 539, + 446 + ], + "type": "image", + "image_path": "5dbd9bc2784406fd5f6ce41d0d65018dadc64176a90a060983706032eabeb524.jpg" + } + ] + } + ], + "index": 70, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 516, + 495, + 531, + 502 + ], + "lines": [ + { + "bbox": [ + 516, + 495, + 531, + 502 + ], + "spans": [ + { + "bbox": [ + 516, + 495, + 531, + 502 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 76, + "angle": 0, + "type": "image_caption" + } + ], + "index": 70 + }, + { + "bbox": [ + 304, + 559, + 545, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 559, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 304, + 559, + 545, + 620 + ], + "type": "text", + "content": "sion [43] with Next3D twice. First, we fine-tune Next3D with the input video to make it learn the original geometry and texture of the avatar. Second, we fine-tune Next3D on the edited image based on the weights of the code and generator from the first fine-tuning." + } + ] + } + ], + "index": 78 + }, + { + "bbox": [ + 306, + 627, + 441, + 640 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 627, + 441, + 640 + ], + "spans": [ + { + "bbox": [ + 306, + 627, + 441, + 640 + ], + "type": "text", + "content": "4.2. Qualitative Comparison" + } + ] + } + ], + "index": 79 + }, + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "content": "Geometry Editing. We first compare our method with baselines on geometry editing, e.g., changing the size of the eyes or the contour of the cheek. We use 2D editing tools, e.g., Photoshop and DragGAN [39], to modify the shape of various facial features. Figure. 3 shows qualitative results on avatars from INSTA [76] (the top two) and" + } + ] + } + ], + "index": 80 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8956" + } + ] + } + ], + "index": 81 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 50, + 105, + 114 + ], + "blocks": [ + { + "bbox": [ + 60, + 50, + 105, + 114 + ], + "lines": [ + { + "bbox": [ + 60, + 50, + 105, + 114 + ], + "spans": [ + { + "bbox": [ + 60, + 50, + 105, + 114 + ], + "type": "image", + "image_path": "b90f233289b3424296c9a976ab097c0d45a1186695d1585ded262f1d4bb18c9e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 115, + 97, + 123 + ], + "lines": [ + { + "bbox": [ + 70, + 115, + 97, + 123 + ], + "spans": [ + { + "bbox": [ + 70, + 115, + 97, + 123 + ], + "type": "text", + "content": "Source" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 111, + 50, + 157, + 114 + ], + "blocks": [ + { + "bbox": [ + 111, + 50, + 157, + 114 + ], + "lines": [ + { + "bbox": [ + 111, + 50, + 157, + 114 + ], + "spans": [ + { + "bbox": [ + 111, + 50, + 157, + 114 + ], + "type": "image", + "image_path": "bbca215fca18c5cbbe146b5d53edc3779f4e7a5ea246ee16d1bacd9035d3b7a3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 125, + 115, + 149, + 123 + ], + "lines": [ + { + "bbox": [ + 125, + 115, + 149, + 123 + ], + "spans": [ + { + "bbox": [ + 125, + 115, + 149, + 123 + ], + "type": "text", + "content": "Eyes++" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 159, + 50, + 203, + 114 + ], + "blocks": [ + { + "bbox": [ + 159, + 50, + 203, + 114 + ], + "lines": [ + { + "bbox": [ + 159, + 50, + 203, + 114 + ], + "spans": [ + { + "bbox": [ + 159, + 50, + 203, + 114 + ], + "type": "image", + "image_path": "2a722f651cd02ed7b58d5fd25dd1fb0977a1a9439bb44fdb04ffbaec3540e725.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 168, + 115, + 194, + 123 + ], + "lines": [ + { + "bbox": [ + 168, + 115, + 194, + 123 + ], + "spans": [ + { + "bbox": [ + 168, + 115, + 194, + 123 + ], + "type": "text", + "content": "Nose++" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 205, + 52, + 249, + 115 + ], + "blocks": [ + { + "bbox": [ + 205, + 52, + 249, + 115 + ], + "lines": [ + { + "bbox": [ + 205, + 52, + 249, + 115 + ], + "spans": [ + { + "bbox": [ + 205, + 52, + 249, + 115 + ], + "type": "image", + "image_path": "11f2fdbb504c0b0d37bfcd4b4ace6260d93a1af5e31f55992afe19d72d1ee635.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 214, + 116, + 238, + 123 + ], + "lines": [ + { + "bbox": [ + 214, + 116, + 238, + 123 + ], + "spans": [ + { + "bbox": [ + 214, + 116, + 238, + 123 + ], + "type": "text", + "content": "Face++" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 60, + 125, + 108, + 171 + ], + "blocks": [ + { + "bbox": [ + 60, + 125, + 108, + 171 + ], + "lines": [ + { + "bbox": [ + 60, + 125, + 108, + 171 + ], + "spans": [ + { + "bbox": [ + 60, + 125, + 108, + 171 + ], + "type": "image", + "image_path": "f0f487863b3cfeabcb9e4247c4c445967e9b20aa062a96e850c72935064d5c18.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 112, + 125, + 157, + 171 + ], + "blocks": [ + { + "bbox": [ + 112, + 125, + 157, + 171 + ], + "lines": [ + { + "bbox": [ + 112, + 125, + 157, + 171 + ], + "spans": [ + { + "bbox": [ + 112, + 125, + 157, + 171 + ], + "type": "image", + "image_path": "1304beb1885bde3f4d9a7ca7dfec5b8fa96ff1c142769379cbf9ac54025d0380.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 159, + 125, + 205, + 171 + ], + "blocks": [ + { + "bbox": [ + 159, + 125, + 205, + 171 + ], + "lines": [ + { + "bbox": [ + 159, + 125, + 205, + 171 + ], + "spans": [ + { + "bbox": [ + 159, + 125, + 205, + 171 + ], + "type": "image", + "image_path": "efd6a7c719849eb243cf4f786f340f814a920aa7f733f736fbae3cdf577ad52d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 206, + 125, + 251, + 171 + ], + "blocks": [ + { + "bbox": [ + 206, + 125, + 251, + 171 + ], + "lines": [ + { + "bbox": [ + 206, + 125, + 251, + 171 + ], + "spans": [ + { + "bbox": [ + 206, + 125, + 251, + 171 + ], + "type": "image", + "image_path": "5eded29fc1b7998dc77200c879b3f7471ef3c48a939c288e0422dbe487731404.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 64, + 181, + 104, + 233 + ], + "blocks": [ + { + "bbox": [ + 70, + 173, + 97, + 181 + ], + "lines": [ + { + "bbox": [ + 70, + 173, + 97, + 181 + ], + "spans": [ + { + "bbox": [ + 70, + 173, + 97, + 181 + ], + "type": "text", + "content": "Source" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 64, + 181, + 104, + 233 + ], + "lines": [ + { + "bbox": [ + 64, + 181, + 104, + 233 + ], + "spans": [ + { + "bbox": [ + 64, + 181, + 104, + 233 + ], + "type": "image", + "image_path": "a3fc954fc4aed7d2eee24ebba4d068c2a33c7c309221cb7e0bf3cbdacafbb7ec.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 234, + 97, + 242 + ], + "lines": [ + { + "bbox": [ + 70, + 234, + 97, + 242 + ], + "spans": [ + { + "bbox": [ + 70, + 234, + 97, + 242 + ], + "type": "text", + "content": "Source" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 113, + 181, + 154, + 233 + ], + "blocks": [ + { + "bbox": [ + 115, + 173, + 149, + 181 + ], + "lines": [ + { + "bbox": [ + 115, + 173, + 149, + 181 + ], + "spans": [ + { + "bbox": [ + 115, + 173, + 149, + 181 + ], + "type": "text", + "content": "Eyes++" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 113, + 181, + 154, + 233 + ], + "lines": [ + { + "bbox": [ + 113, + 181, + 154, + 233 + ], + "spans": [ + { + "bbox": [ + 113, + 181, + 154, + 233 + ], + "type": "image", + "image_path": "b9e3dc8e03f02ff5d2f03e4fe9ef0942679dbcffe76ff0ccd9053d06e3c0214e.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 112, + 242, + 153, + 294 + ], + "blocks": [ + { + "bbox": [ + 125, + 234, + 146, + 242 + ], + "lines": [ + { + "bbox": [ + 125, + 234, + 146, + 242 + ], + "spans": [ + { + "bbox": [ + 125, + 234, + 146, + 242 + ], + "type": "text", + "content": "Eyes--" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 112, + 242, + 153, + 294 + ], + "lines": [ + { + "bbox": [ + 112, + 242, + 153, + 294 + ], + "spans": [ + { + "bbox": [ + 112, + 242, + 153, + 294 + ], + "type": "image", + "image_path": "7197b1c39086bb37705a0fbded774806a2513e754af84ad1b676c586137f2d1a.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 120, + 296, + 151, + 304 + ], + "lines": [ + { + "bbox": [ + 120, + 296, + 151, + 304 + ], + "spans": [ + { + "bbox": [ + 120, + 296, + 151, + 304 + ], + "type": "text", + "content": "Mouth++" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 160, + 181, + 200, + 233 + ], + "blocks": [ + { + "bbox": [ + 169, + 173, + 190, + 181 + ], + "lines": [ + { + "bbox": [ + 169, + 173, + 190, + 181 + ], + "spans": [ + { + "bbox": [ + 169, + 173, + 190, + 181 + ], + "type": "text", + "content": "Face—" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 160, + 181, + 200, + 233 + ], + "lines": [ + { + "bbox": [ + 160, + 181, + 200, + 233 + ], + "spans": [ + { + "bbox": [ + 160, + 181, + 200, + 233 + ], + "type": "image", + "image_path": "478bff7e3da760bd9cef87f5f41d37076622f78a5b19d90044bc6fc0ed0d8465.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 159, + 242, + 200, + 294 + ], + "blocks": [ + { + "bbox": [ + 169, + 234, + 194, + 242 + ], + "lines": [ + { + "bbox": [ + 169, + 234, + 194, + 242 + ], + "spans": [ + { + "bbox": [ + 169, + 234, + 194, + 242 + ], + "type": "text", + "content": "Nose++" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 159, + 242, + 200, + 294 + ], + "lines": [ + { + "bbox": [ + 159, + 242, + 200, + 294 + ], + "spans": [ + { + "bbox": [ + 159, + 242, + 200, + 294 + ], + "type": "image", + "image_path": "ca7a3b80bc3f0171f58e6a3ba841465a1b5301fa1c19326bb7ac360b64eedf67.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 168, + 296, + 194, + 304 + ], + "lines": [ + { + "bbox": [ + 168, + 296, + 194, + 304 + ], + "spans": [ + { + "bbox": [ + 168, + 296, + 194, + 304 + ], + "type": "text", + "content": "Nose++" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 160, + 306, + 201, + 318 + ], + "lines": [ + { + "bbox": [ + 160, + 306, + 201, + 318 + ], + "spans": [ + { + "bbox": [ + 160, + 306, + 201, + 318 + ], + "type": "text", + "content": "2D Editing" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 208, + 181, + 246, + 233 + ], + "blocks": [ + { + "bbox": [ + 214, + 173, + 238, + 181 + ], + "lines": [ + { + "bbox": [ + 214, + 173, + 238, + 181 + ], + "spans": [ + { + "bbox": [ + 214, + 173, + 238, + 181 + ], + "type": "text", + "content": "Face++" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 208, + 181, + 246, + 233 + ], + "lines": [ + { + "bbox": [ + 208, + 181, + 246, + 233 + ], + "spans": [ + { + "bbox": [ + 208, + 181, + 246, + 233 + ], + "type": "image", + "image_path": "e9e712a6349a7975b87c7cebe0d43030027302496e5ee87ba44d8446a222de07.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 206, + 242, + 244, + 294 + ], + "blocks": [ + { + "bbox": [ + 218, + 234, + 237, + 242 + ], + "lines": [ + { + "bbox": [ + 218, + 234, + 237, + 242 + ], + "spans": [ + { + "bbox": [ + 218, + 234, + 237, + 242 + ], + "type": "text", + "content": "Jaw--" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 206, + 242, + 244, + 294 + ], + "lines": [ + { + "bbox": [ + 206, + 242, + 244, + 294 + ], + "spans": [ + { + "bbox": [ + 206, + 242, + 244, + 294 + ], + "type": "image", + "image_path": "ef9338c97ff64aca72b723801fa053ab1152ba1065bf5d7e4e0bbb8b98739901.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 218, + 296, + 236, + 304 + ], + "lines": [ + { + "bbox": [ + 218, + 296, + 236, + 304 + ], + "spans": [ + { + "bbox": [ + 218, + 296, + 236, + 304 + ], + "type": "text", + "content": "Jaw--" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 256, + 52, + 299, + 115 + ], + "blocks": [ + { + "bbox": [ + 256, + 52, + 299, + 115 + ], + "lines": [ + { + "bbox": [ + 256, + 52, + 299, + 115 + ], + "spans": [ + { + "bbox": [ + 256, + 52, + 299, + 115 + ], + "type": "image", + "image_path": "bb9011ed37410fd69b9793bc984434283713980fd18fdca3222b95cfc78906dd.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 268, + 116, + 292, + 123 + ], + "lines": [ + { + "bbox": [ + 268, + 116, + 292, + 123 + ], + "spans": [ + { + "bbox": [ + 268, + 116, + 292, + 123 + ], + "type": "text", + "content": "Eyes++" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 253, + 125, + 300, + 171 + ], + "blocks": [ + { + "bbox": [ + 253, + 125, + 300, + 171 + ], + "lines": [ + { + "bbox": [ + 253, + 125, + 300, + 171 + ], + "spans": [ + { + "bbox": [ + 253, + 125, + 300, + 171 + ], + "type": "image", + "image_path": "bca0d99aa388cdd58bce78abebb3ea0cf1ecc971a08ef9c6adaba7d2d4f10c19.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 259, + 182, + 298, + 232 + ], + "blocks": [ + { + "bbox": [ + 268, + 173, + 292, + 181 + ], + "lines": [ + { + "bbox": [ + 268, + 173, + 292, + 181 + ], + "spans": [ + { + "bbox": [ + 268, + 173, + 292, + 181 + ], + "type": "text", + "content": "Eyes++" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 259, + 182, + 298, + 232 + ], + "lines": [ + { + "bbox": [ + 259, + 182, + 298, + 232 + ], + "spans": [ + { + "bbox": [ + 259, + 182, + 298, + 232 + ], + "type": "image", + "image_path": "ca681e99a62d0ac892a7186c8de3ee5f00e978a820fa68d74385290840c2e5f0.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 258, + 242, + 294, + 294 + ], + "blocks": [ + { + "bbox": [ + 268, + 234, + 288, + 242 + ], + "lines": [ + { + "bbox": [ + 268, + 234, + 288, + 242 + ], + "spans": [ + { + "bbox": [ + 268, + 234, + 288, + 242 + ], + "type": "text", + "content": "Eyes" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 258, + 242, + 294, + 294 + ], + "lines": [ + { + "bbox": [ + 258, + 242, + 294, + 294 + ], + "spans": [ + { + "bbox": [ + 258, + 242, + 294, + 294 + ], + "type": "image", + "image_path": "b8997f3637d63b68b8a7826e899b88bee251f410a2a7b40ae5734a480190a727.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 263, + 296, + 294, + 304 + ], + "lines": [ + { + "bbox": [ + 263, + 296, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 263, + 296, + 294, + 304 + ], + "type": "text", + "content": "Mouth++" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 302, + 52, + 345, + 115 + ], + "blocks": [ + { + "bbox": [ + 302, + 52, + 345, + 115 + ], + "lines": [ + { + "bbox": [ + 302, + 52, + 345, + 115 + ], + "spans": [ + { + "bbox": [ + 302, + 52, + 345, + 115 + ], + "type": "image", + "image_path": "230766ffce13445e1c592c48d7dd848e97c543427124fb798e4c5f1b60bc7850.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 116, + 337, + 123 + ], + "lines": [ + { + "bbox": [ + 312, + 116, + 337, + 123 + ], + "spans": [ + { + "bbox": [ + 312, + 116, + 337, + 123 + ], + "type": "text", + "content": "Nose++" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 301, + 125, + 347, + 171 + ], + "blocks": [ + { + "bbox": [ + 301, + 125, + 347, + 171 + ], + "lines": [ + { + "bbox": [ + 301, + 125, + 347, + 171 + ], + "spans": [ + { + "bbox": [ + 301, + 125, + 347, + 171 + ], + "type": "image", + "image_path": "7232a9d916ef95bdabd8f8a74053b12449a71836f43b80907fa46538b2f2d258.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 347, + 53, + 391, + 171 + ], + "blocks": [ + { + "bbox": [ + 347, + 53, + 391, + 171 + ], + "lines": [ + { + "bbox": [ + 347, + 53, + 391, + 171 + ], + "spans": [ + { + "bbox": [ + 347, + 53, + 391, + 171 + ], + "type": "image", + "image_path": "e79f0fd28e428d35226fb4860bcccc720a4a9a5dc583dd015027b6a37277b20d.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + } + ], + "index": 43 + }, + { + "bbox": [ + 358, + 173, + 382, + 181 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 173, + 382, + 181 + ], + "spans": [ + { + "bbox": [ + 358, + 173, + 382, + 181 + ], + "type": "text", + "content": "Face++" + } + ] + } + ], + "index": 44, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 306, + 182, + 343, + 233 + ], + "blocks": [ + { + "bbox": [ + 313, + 173, + 334, + 181 + ], + "lines": [ + { + "bbox": [ + 313, + 173, + 334, + 181 + ], + "spans": [ + { + "bbox": [ + 313, + 173, + 334, + 181 + ], + "type": "text", + "content": "Face--" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 182, + 343, + 233 + ], + "lines": [ + { + "bbox": [ + 306, + 182, + 343, + 233 + ], + "spans": [ + { + "bbox": [ + 306, + 182, + 343, + 233 + ], + "type": "image", + "image_path": "f26317241bbdd62fa66d7a8deba6a4c7acf22955db5a06069d7af6d1f28aebbb.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 304, + 242, + 340, + 294 + ], + "blocks": [ + { + "bbox": [ + 312, + 234, + 337, + 242 + ], + "lines": [ + { + "bbox": [ + 312, + 234, + 337, + 242 + ], + "spans": [ + { + "bbox": [ + 312, + 234, + 337, + 242 + ], + "type": "text", + "content": "Nose++" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 242, + 340, + 294 + ], + "lines": [ + { + "bbox": [ + 304, + 242, + 340, + 294 + ], + "spans": [ + { + "bbox": [ + 304, + 242, + 340, + 294 + ], + "type": "image", + "image_path": "23bae771951b2c1798856ab0617da82f1d23c4fec4095f38e7955573d719acb2.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 296, + 337, + 303 + ], + "lines": [ + { + "bbox": [ + 312, + 296, + 337, + 303 + ], + "spans": [ + { + "bbox": [ + 312, + 296, + 337, + 303 + ], + "type": "text", + "content": "Nose++" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_caption" + } + ], + "index": 47 + }, + { + "type": "image", + "bbox": [ + 351, + 182, + 389, + 233 + ], + "blocks": [ + { + "bbox": [ + 348, + 173, + 384, + 181 + ], + "lines": [ + { + "bbox": [ + 348, + 173, + 384, + 181 + ], + "spans": [ + { + "bbox": [ + 348, + 173, + 384, + 181 + ], + "type": "text", + "content": "Face++" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 351, + 182, + 389, + 233 + ], + "lines": [ + { + "bbox": [ + 351, + 182, + 389, + 233 + ], + "spans": [ + { + "bbox": [ + 351, + 182, + 389, + 233 + ], + "type": "image", + "image_path": "6e7ddb7187d0d19303cdf76649ce0ad12ce1878c3027f02d6bf5315636df2588.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 350, + 242, + 386, + 294 + ], + "blocks": [ + { + "bbox": [ + 360, + 234, + 378, + 242 + ], + "lines": [ + { + "bbox": [ + 360, + 234, + 378, + 242 + ], + "spans": [ + { + "bbox": [ + 360, + 234, + 378, + 242 + ], + "type": "text", + "content": "Jaw" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 350, + 242, + 386, + 294 + ], + "lines": [ + { + "bbox": [ + 350, + 242, + 386, + 294 + ], + "spans": [ + { + "bbox": [ + 350, + 242, + 386, + 294 + ], + "type": "image", + "image_path": "afb64723be416433deff85c3a4665c6bd1f542fb0ec4f2a9cc66fda6d329fd28.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 296, + 378, + 303 + ], + "lines": [ + { + "bbox": [ + 361, + 296, + 378, + 303 + ], + "spans": [ + { + "bbox": [ + 361, + 296, + 378, + 303 + ], + "type": "text", + "content": "Jaw" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_caption" + } + ], + "index": 52 + }, + { + "type": "image", + "bbox": [ + 402, + 53, + 443, + 115 + ], + "blocks": [ + { + "bbox": [ + 402, + 53, + 443, + 115 + ], + "lines": [ + { + "bbox": [ + 402, + 53, + 443, + 115 + ], + "spans": [ + { + "bbox": [ + 402, + 53, + 443, + 115 + ], + "type": "image", + "image_path": "6e24651350dee56a36acf41eed355e22a704a1be2906b303ae65b268ac9faebc.jpg" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 116, + 435, + 123 + ], + "lines": [ + { + "bbox": [ + 410, + 116, + 435, + 123 + ], + "spans": [ + { + "bbox": [ + 410, + 116, + 435, + 123 + ], + "type": "text", + "content": "Eyes++" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_caption" + } + ], + "index": 54 + }, + { + "type": "image", + "bbox": [ + 446, + 53, + 487, + 115 + ], + "blocks": [ + { + "bbox": [ + 446, + 53, + 487, + 115 + ], + "lines": [ + { + "bbox": [ + 446, + 53, + 487, + 115 + ], + "spans": [ + { + "bbox": [ + 446, + 53, + 487, + 115 + ], + "type": "image", + "image_path": "d0afbee0d8c01e1e0d6ba342c2c77f189db98cafe07124cd87a9fbf4c9c31bc8.jpg" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 453, + 116, + 479, + 123 + ], + "lines": [ + { + "bbox": [ + 453, + 116, + 479, + 123 + ], + "spans": [ + { + "bbox": [ + 453, + 116, + 479, + 123 + ], + "type": "text", + "content": "Nose++" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_caption" + } + ], + "index": 56 + }, + { + "type": "image", + "bbox": [ + 491, + 53, + 533, + 114 + ], + "blocks": [ + { + "bbox": [ + 491, + 53, + 533, + 114 + ], + "lines": [ + { + "bbox": [ + 491, + 53, + 533, + 114 + ], + "spans": [ + { + "bbox": [ + 491, + 53, + 533, + 114 + ], + "type": "image", + "image_path": "ce2b50b1fed1ee519436a36d8dc0e293053d0d6bd9fc898d83a09a9389e69746.jpg" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 501, + 116, + 525, + 123 + ], + "lines": [ + { + "bbox": [ + 501, + 116, + 525, + 123 + ], + "spans": [ + { + "bbox": [ + 501, + 116, + 525, + 123 + ], + "type": "text", + "content": "Face++" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_caption" + } + ], + "index": 58 + }, + { + "type": "image", + "bbox": [ + 399, + 125, + 444, + 171 + ], + "blocks": [ + { + "bbox": [ + 399, + 125, + 444, + 171 + ], + "lines": [ + { + "bbox": [ + 399, + 125, + 444, + 171 + ], + "spans": [ + { + "bbox": [ + 399, + 125, + 444, + 171 + ], + "type": "image", + "image_path": "741ece174c8853b3d206655524ff8cb1e47fec356b7f6b0d038a5f51f33faeec.jpg" + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_body" + } + ], + "index": 60 + }, + { + "type": "image", + "bbox": [ + 445, + 125, + 488, + 170 + ], + "blocks": [ + { + "bbox": [ + 445, + 125, + 488, + 170 + ], + "lines": [ + { + "bbox": [ + 445, + 125, + 488, + 170 + ], + "spans": [ + { + "bbox": [ + 445, + 125, + 488, + 170 + ], + "type": "image", + "image_path": "80870e6918668f9108fe218213f2bad7a3dc1773af0c1dad14816f2e4bfd4c1d.jpg" + } + ] + } + ], + "index": 62, + "angle": 0, + "type": "image_body" + } + ], + "index": 62 + }, + { + "type": "image", + "bbox": [ + 490, + 125, + 535, + 170 + ], + "blocks": [ + { + "bbox": [ + 490, + 125, + 535, + 170 + ], + "lines": [ + { + "bbox": [ + 490, + 125, + 535, + 170 + ], + "spans": [ + { + "bbox": [ + 490, + 125, + 535, + 170 + ], + "type": "image", + "image_path": "cd14ad760734a44e717880f6c6cc2f83cf676c3ab6ea6e374b5801c9eac75a20.jpg" + } + ] + } + ], + "index": 64, + "angle": 0, + "type": "image_body" + } + ], + "index": 64 + }, + { + "type": "image", + "bbox": [ + 402, + 182, + 439, + 233 + ], + "blocks": [ + { + "bbox": [ + 410, + 173, + 435, + 181 + ], + "lines": [ + { + "bbox": [ + 410, + 173, + 435, + 181 + ], + "spans": [ + { + "bbox": [ + 410, + 173, + 435, + 181 + ], + "type": "text", + "content": "Eyes++" + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 402, + 182, + 439, + 233 + ], + "lines": [ + { + "bbox": [ + 402, + 182, + 439, + 233 + ], + "spans": [ + { + "bbox": [ + 402, + 182, + 439, + 233 + ], + "type": "image", + "image_path": "d5da76eca8a595452a5c7e83a8c96eaf23afafb04714f8f932270492c44c3bb9.jpg" + } + ] + } + ], + "index": 66, + "angle": 0, + "type": "image_body" + } + ], + "index": 66 + }, + { + "type": "image", + "bbox": [ + 447, + 182, + 484, + 233 + ], + "blocks": [ + { + "bbox": [ + 456, + 173, + 479, + 181 + ], + "lines": [ + { + "bbox": [ + 456, + 173, + 479, + 181 + ], + "spans": [ + { + "bbox": [ + 456, + 173, + 479, + 181 + ], + "type": "text", + "content": "Face--" + } + ] + } + ], + "index": 63, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 447, + 182, + 484, + 233 + ], + "lines": [ + { + "bbox": [ + 447, + 182, + 484, + 233 + ], + "spans": [ + { + "bbox": [ + 447, + 182, + 484, + 233 + ], + "type": "image", + "image_path": "d3ced2f2959cbd833307ed98ca40876f76275fbb348555d27d6ba9ffb6ceed57.jpg" + } + ] + } + ], + "index": 68, + "angle": 0, + "type": "image_body" + } + ], + "index": 68 + }, + { + "type": "image", + "bbox": [ + 493, + 182, + 530, + 233 + ], + "blocks": [ + { + "bbox": [ + 501, + 173, + 525, + 181 + ], + "lines": [ + { + "bbox": [ + 501, + 173, + 525, + 181 + ], + "spans": [ + { + "bbox": [ + 501, + 173, + 525, + 181 + ], + "type": "text", + "content": "Face++" + } + ] + } + ], + "index": 65, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 493, + 182, + 530, + 233 + ], + "lines": [ + { + "bbox": [ + 493, + 182, + 530, + 233 + ], + "spans": [ + { + "bbox": [ + 493, + 182, + 530, + 233 + ], + "type": "image", + "image_path": "667ffc8d7c57b39661e47dd63b4c79a6078676d3934bd0c3663c52aafc19ec17.jpg" + } + ] + } + ], + "index": 70, + "angle": 0, + "type": "image_body" + } + ], + "index": 70 + }, + { + "type": "image", + "bbox": [ + 401, + 242, + 439, + 294 + ], + "blocks": [ + { + "bbox": [ + 410, + 234, + 432, + 242 + ], + "lines": [ + { + "bbox": [ + 410, + 234, + 432, + 242 + ], + "spans": [ + { + "bbox": [ + 410, + 234, + 432, + 242 + ], + "type": "text", + "content": "Eyes" + } + ] + } + ], + "index": 67, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 401, + 242, + 439, + 294 + ], + "lines": [ + { + "bbox": [ + 401, + 242, + 439, + 294 + ], + "spans": [ + { + "bbox": [ + 401, + 242, + 439, + 294 + ], + "type": "image", + "image_path": "38b33b46641b59352520a050b71f1b011a91f147b55303bcbc1bfe09afff9de9.jpg" + } + ] + } + ], + "index": 72, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 296, + 436, + 303 + ], + "lines": [ + { + "bbox": [ + 406, + 296, + 436, + 303 + ], + "spans": [ + { + "bbox": [ + 406, + 296, + 436, + 303 + ], + "type": "text", + "content": "Mouth++" + } + ] + } + ], + "index": 73, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 435, + 306, + 500, + 316 + ], + "lines": [ + { + "bbox": [ + 435, + 306, + 500, + 316 + ], + "spans": [ + { + "bbox": [ + 435, + 306, + 500, + 316 + ], + "type": "text", + "content": "Rendered View 2" + } + ] + } + ], + "index": 78, + "angle": 0, + "type": "image_caption" + } + ], + "index": 72 + }, + { + "type": "image", + "bbox": [ + 447, + 242, + 485, + 294 + ], + "blocks": [ + { + "bbox": [ + 453, + 234, + 479, + 242 + ], + "lines": [ + { + "bbox": [ + 453, + 234, + 479, + 242 + ], + "spans": [ + { + "bbox": [ + 453, + 234, + 479, + 242 + ], + "type": "text", + "content": "Nose++" + } + ] + } + ], + "index": 69, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 447, + 242, + 485, + 294 + ], + "lines": [ + { + "bbox": [ + 447, + 242, + 485, + 294 + ], + "spans": [ + { + "bbox": [ + 447, + 242, + 485, + 294 + ], + "type": "image", + "image_path": "f4e33dede41e3e56575e3524a713d8893388a0f60e988117d82079771b19903c.jpg" + } + ] + } + ], + "index": 74, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 453, + 296, + 479, + 303 + ], + "lines": [ + { + "bbox": [ + 453, + 296, + 479, + 303 + ], + "spans": [ + { + "bbox": [ + 453, + 296, + 479, + 303 + ], + "type": "text", + "content": "Nose++" + } + ] + } + ], + "index": 75, + "angle": 0, + "type": "image_caption" + } + ], + "index": 74 + }, + { + "type": "image", + "bbox": [ + 492, + 242, + 531, + 294 + ], + "blocks": [ + { + "bbox": [ + 503, + 234, + 522, + 242 + ], + "lines": [ + { + "bbox": [ + 503, + 234, + 522, + 242 + ], + "spans": [ + { + "bbox": [ + 503, + 234, + 522, + 242 + ], + "type": "text", + "content": "Jaw--" + } + ] + } + ], + "index": 71, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 492, + 242, + 531, + 294 + ], + "lines": [ + { + "bbox": [ + 492, + 242, + 531, + 294 + ], + "spans": [ + { + "bbox": [ + 492, + 242, + 531, + 294 + ], + "type": "image", + "image_path": "da76fc00a6ee01e36bef667b915c92dcaa254ed633381314c3a159a39cc5a236.jpg" + } + ] + } + ], + "index": 76, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 504, + 296, + 522, + 303 + ], + "lines": [ + { + "bbox": [ + 504, + 296, + 522, + 303 + ], + "spans": [ + { + "bbox": [ + 504, + 296, + 522, + 303 + ], + "type": "text", + "content": "Jaw" + } + ] + } + ], + "index": 77, + "angle": 0, + "type": "image_caption" + } + ], + "index": 76 + }, + { + "type": "image", + "bbox": [ + 58, + 331, + 114, + 379 + ], + "blocks": [ + { + "bbox": [ + 58, + 331, + 114, + 379 + ], + "lines": [ + { + "bbox": [ + 58, + 331, + 114, + 379 + ], + "spans": [ + { + "bbox": [ + 58, + 331, + 114, + 379 + ], + "type": "image", + "image_path": "02c7af284df2b8168629893e30005183707e1eafc19abef6404e32aa97f73b6c.jpg" + } + ] + } + ], + "index": 79, + "angle": 0, + "type": "image_body" + } + ], + "index": 79 + }, + { + "type": "image", + "bbox": [ + 56, + 388, + 110, + 449 + ], + "blocks": [ + { + "bbox": [ + 73, + 381, + 99, + 389 + ], + "lines": [ + { + "bbox": [ + 73, + 381, + 99, + 389 + ], + "spans": [ + { + "bbox": [ + 73, + 381, + 99, + 389 + ], + "type": "text", + "content": "Nose++" + } + ] + } + ], + "index": 80, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 388, + 110, + 449 + ], + "lines": [ + { + "bbox": [ + 56, + 388, + 110, + 449 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 110, + 449 + ], + "type": "image", + "image_path": "1a31f7924d28469fe607b650ed390b34495eec10a8db058716820f4926348675.jpg" + } + ] + } + ], + "index": 81, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 73, + 452, + 96, + 459 + ], + "lines": [ + { + "bbox": [ + 73, + 452, + 96, + 459 + ], + "spans": [ + { + "bbox": [ + 73, + 452, + 96, + 459 + ], + "type": "text", + "content": "Hair++" + } + ] + } + ], + "index": 82, + "angle": 0, + "type": "image_caption" + } + ], + "index": 81 + }, + { + "type": "image", + "bbox": [ + 120, + 331, + 162, + 379 + ], + "blocks": [ + { + "bbox": [ + 120, + 331, + 162, + 379 + ], + "lines": [ + { + "bbox": [ + 120, + 331, + 162, + 379 + ], + "spans": [ + { + "bbox": [ + 120, + 331, + 162, + 379 + ], + "type": "image", + "image_path": "1cf75bf3815a98bd2b46912a9e6cb75c9bc87a62adc7af0cd7df841903dbfec4.jpg" + } + ] + } + ], + "index": 83, + "angle": 0, + "type": "image_body" + } + ], + "index": 83 + }, + { + "type": "image", + "bbox": [ + 119, + 390, + 160, + 450 + ], + "blocks": [ + { + "bbox": [ + 143, + 380, + 183, + 390 + ], + "lines": [ + { + "bbox": [ + 143, + 380, + 183, + 390 + ], + "spans": [ + { + "bbox": [ + 143, + 380, + 183, + 390 + ], + "type": "text", + "content": "Rendering" + } + ] + } + ], + "index": 84, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 119, + 390, + 160, + 450 + ], + "lines": [ + { + "bbox": [ + 119, + 390, + 160, + 450 + ], + "spans": [ + { + "bbox": [ + 119, + 390, + 160, + 450 + ], + "type": "image", + "image_path": "a22ceb8d96f639e53a0ed1a2e1ccbb494d5bd0e8e143744c47000fdf8de9fd0b.jpg" + } + ] + } + ], + "index": 85, + "angle": 0, + "type": "image_body" + } + ], + "index": 85 + }, + { + "bbox": [ + 141, + 452, + 181, + 461 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 452, + 181, + 461 + ], + "spans": [ + { + "bbox": [ + 141, + 452, + 181, + 461 + ], + "type": "text", + "content": "Rendering" + } + ] + } + ], + "index": 86, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 167, + 331, + 206, + 378 + ], + "blocks": [ + { + "bbox": [ + 167, + 331, + 206, + 378 + ], + "lines": [ + { + "bbox": [ + 167, + 331, + 206, + 378 + ], + "spans": [ + { + "bbox": [ + 167, + 331, + 206, + 378 + ], + "type": "image", + "image_path": "58e8c916ead852d59a7ca5c2648b45bce0c0e46562e3b7e1be13ef4a1a6d5fd9.jpg" + } + ] + } + ], + "index": 87, + "angle": 0, + "type": "image_body" + } + ], + "index": 87 + }, + { + "bbox": [ + 167, + 380, + 183, + 389 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 380, + 183, + 389 + ], + "spans": [ + { + "bbox": [ + 167, + 380, + 183, + 389 + ], + "type": "text", + "content": "ering" + } + ] + } + ], + "index": 88, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 167, + 390, + 205, + 450 + ], + "blocks": [ + { + "bbox": [ + 167, + 390, + 205, + 450 + ], + "lines": [ + { + "bbox": [ + 167, + 390, + 205, + 450 + ], + "spans": [ + { + "bbox": [ + 167, + 390, + 205, + 450 + ], + "type": "image", + "image_path": "e5a34c470e88eeff2443a2a861e733e8b673ae20804126433159de6e552c46dc.jpg" + } + ] + } + ], + "index": 89, + "angle": 0, + "type": "image_body" + } + ], + "index": 89 + }, + { + "bbox": [ + 141, + 452, + 181, + 461 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 452, + 181, + 461 + ], + "spans": [ + { + "bbox": [ + 141, + 452, + 181, + 461 + ], + "type": "text", + "content": "Rendering" + } + ] + } + ], + "index": 90, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 210, + 331, + 238, + 371 + ], + "blocks": [ + { + "bbox": [ + 210, + 331, + 238, + 371 + ], + "lines": [ + { + "bbox": [ + 210, + 331, + 238, + 371 + ], + "spans": [ + { + "bbox": [ + 210, + 331, + 238, + 371 + ], + "type": "image", + "image_path": "3a5417ce0ca53138a5b8bf3815abeb18e0aa8b55d86e66a4daed3ffc8b0718c5.jpg" + } + ] + } + ], + "index": 91, + "angle": 0, + "type": "image_body" + } + ], + "index": 91 + }, + { + "type": "image", + "bbox": [ + 235, + 342, + 263, + 382 + ], + "blocks": [ + { + "bbox": [ + 235, + 342, + 263, + 382 + ], + "lines": [ + { + "bbox": [ + 235, + 342, + 263, + 382 + ], + "spans": [ + { + "bbox": [ + 235, + 342, + 263, + 382 + ], + "type": "image", + "image_path": "ef7ad31c9a02a5c9dab557d9fafb8503df357b9163df2da4225f2f63f2a34dc9.jpg" + } + ] + } + ], + "index": 92, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 229, + 381, + 254, + 389 + ], + "lines": [ + { + "bbox": [ + 229, + 381, + 254, + 389 + ], + "spans": [ + { + "bbox": [ + 229, + 381, + 254, + 389 + ], + "type": "text", + "content": "Eyes++" + } + ] + } + ], + "index": 93, + "angle": 0, + "type": "image_caption" + } + ], + "index": 92 + }, + { + "type": "image", + "bbox": [ + 210, + 390, + 238, + 430 + ], + "blocks": [ + { + "bbox": [ + 210, + 390, + 238, + 430 + ], + "lines": [ + { + "bbox": [ + 210, + 390, + 238, + 430 + ], + "spans": [ + { + "bbox": [ + 210, + 390, + 238, + 430 + ], + "type": "image", + "image_path": "bd5fed5b3876ebcd03d8e399050945554b3a5be9adfa561e16a59f9265ebeecf.jpg" + } + ] + } + ], + "index": 94, + "angle": 0, + "type": "image_body" + } + ], + "index": 94 + }, + { + "type": "image", + "bbox": [ + 239, + 390, + 266, + 450 + ], + "blocks": [ + { + "bbox": [ + 239, + 390, + 266, + 450 + ], + "lines": [ + { + "bbox": [ + 239, + 390, + 266, + 450 + ], + "spans": [ + { + "bbox": [ + 239, + 390, + 266, + 450 + ], + "type": "image", + "image_path": "cbfd92e8c331c4dd5056e115b36e6fe125d690618614ee1c97f8526b1a8d962f.jpg" + } + ] + } + ], + "index": 95, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 226, + 452, + 256, + 460 + ], + "lines": [ + { + "bbox": [ + 226, + 452, + 256, + 460 + ], + "spans": [ + { + "bbox": [ + 226, + 452, + 256, + 460 + ], + "type": "text", + "content": "Mouth++" + } + ] + } + ], + "index": 96, + "angle": 0, + "type": "image_caption" + } + ], + "index": 95 + }, + { + "type": "image", + "bbox": [ + 266, + 331, + 306, + 380 + ], + "blocks": [ + { + "bbox": [ + 266, + 331, + 306, + 380 + ], + "lines": [ + { + "bbox": [ + 266, + 331, + 306, + 380 + ], + "spans": [ + { + "bbox": [ + 266, + 331, + 306, + 380 + ], + "type": "image", + "image_path": "ea2a0146457762a08953182b199d9af2ddd7ef22a452e89e00f7871c1cd412de.jpg" + } + ] + } + ], + "index": 97, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 289, + 380, + 329, + 390 + ], + "lines": [ + { + "bbox": [ + 289, + 380, + 329, + 390 + ], + "spans": [ + { + "bbox": [ + 289, + 380, + 329, + 390 + ], + "type": "text", + "content": "Rendering" + } + ] + } + ], + "index": 98, + "angle": 0, + "type": "image_caption" + } + ], + "index": 97 + }, + { + "type": "image", + "bbox": [ + 269, + 390, + 309, + 450 + ], + "blocks": [ + { + "bbox": [ + 269, + 390, + 309, + 450 + ], + "lines": [ + { + "bbox": [ + 269, + 390, + 309, + 450 + ], + "spans": [ + { + "bbox": [ + 269, + 390, + 309, + 450 + ], + "type": "image", + "image_path": "2f329292d92306774e6b1cf0c0645a91ebf8f09b27728936ca1a1595ef11f7f5.jpg" + } + ] + } + ], + "index": 99, + "angle": 0, + "type": "image_body" + } + ], + "index": 99 + }, + { + "type": "image", + "bbox": [ + 306, + 331, + 351, + 379 + ], + "blocks": [ + { + "bbox": [ + 306, + 331, + 351, + 379 + ], + "lines": [ + { + "bbox": [ + 306, + 331, + 351, + 379 + ], + "spans": [ + { + "bbox": [ + 306, + 331, + 351, + 379 + ], + "type": "image", + "image_path": "35326647f75ba52cb62ca8d4d0e92ade8e1c8673062fb979a50454d465082e65.jpg" + } + ] + } + ], + "index": 101, + "angle": 0, + "type": "image_body" + } + ], + "index": 101 + }, + { + "bbox": [ + 312, + 380, + 329, + 389 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 380, + 329, + 389 + ], + "spans": [ + { + "bbox": [ + 312, + 380, + 329, + 389 + ], + "type": "text", + "content": "ering" + } + ] + } + ], + "index": 102, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 306, + 390, + 353, + 452 + ], + "blocks": [ + { + "bbox": [ + 287, + 452, + 327, + 461 + ], + "lines": [ + { + "bbox": [ + 287, + 452, + 327, + 461 + ], + "spans": [ + { + "bbox": [ + 287, + 452, + 327, + 461 + ], + "type": "text", + "content": "Rendering" + } + ] + } + ], + "index": 100, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 390, + 353, + 452 + ], + "lines": [ + { + "bbox": [ + 306, + 390, + 353, + 452 + ], + "spans": [ + { + "bbox": [ + 306, + 390, + 353, + 452 + ], + "type": "image", + "image_path": "173c0476fffde8cb7cca357ce4246bd91fb28e70e1aea04980b3bf0d8c7496f1.jpg" + } + ] + } + ], + "index": 103, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 354, + 462, + 534, + 472 + ], + "lines": [ + { + "bbox": [ + 354, + 462, + 534, + 472 + ], + "spans": [ + { + "bbox": [ + 354, + 462, + 534, + 472 + ], + "type": "text", + "content": "(c) INSTA & Next3D Geometry Editing Results" + } + ] + } + ], + "index": 120, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 52, + 473, + 539, + 484 + ], + "lines": [ + { + "bbox": [ + 52, + 473, + 539, + 484 + ], + "spans": [ + { + "bbox": [ + 52, + 473, + 539, + 484 + ], + "type": "text", + "content": "Figure 4. Our geometry editing results with the drag-style 2D editing on INSTA [76], NeRFBlendshape [16], and Next3D [50] avatars." + } + ] + } + ], + "index": 121, + "angle": 0, + "type": "image_caption" + } + ], + "index": 103 + }, + { + "type": "image", + "bbox": [ + 354, + 335, + 399, + 381 + ], + "blocks": [ + { + "bbox": [ + 354, + 335, + 399, + 381 + ], + "lines": [ + { + "bbox": [ + 354, + 335, + 399, + 381 + ], + "spans": [ + { + "bbox": [ + 354, + 335, + 399, + 381 + ], + "type": "image", + "image_path": "92381cd1262c3da770ddd8919d31a415d9ddad5a19ffbdf35a98a126574ca716.jpg" + } + ] + } + ], + "index": 104, + "angle": 0, + "type": "image_body" + } + ], + "index": 104 + }, + { + "type": "image", + "bbox": [ + 362, + 393, + 400, + 445 + ], + "blocks": [ + { + "bbox": [ + 365, + 383, + 392, + 392 + ], + "lines": [ + { + "bbox": [ + 365, + 383, + 392, + 392 + ], + "spans": [ + { + "bbox": [ + 365, + 383, + 392, + 392 + ], + "type": "text", + "content": "Source" + } + ] + } + ], + "index": 105, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 362, + 393, + 400, + 445 + ], + "lines": [ + { + "bbox": [ + 362, + 393, + 400, + 445 + ], + "spans": [ + { + "bbox": [ + 362, + 393, + 400, + 445 + ], + "type": "image", + "image_path": "9c9f40b2e66d389cf05f9cb2f47cb9f0f18b20ce821eadfbdc0914bd0b706701.jpg" + } + ] + } + ], + "index": 106, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 365, + 449, + 392, + 457 + ], + "lines": [ + { + "bbox": [ + 365, + 449, + 392, + 457 + ], + "spans": [ + { + "bbox": [ + 365, + 449, + 392, + 457 + ], + "type": "text", + "content": "Source" + } + ] + } + ], + "index": 107, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 399, + 383, + 451, + 392 + ], + "lines": [ + { + "bbox": [ + 399, + 383, + 451, + 392 + ], + "spans": [ + { + "bbox": [ + 399, + 383, + 451, + 392 + ], + "type": "text", + "content": "Mouse-,Face-" + } + ] + } + ], + "index": 109, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 400, + 447, + 452, + 463 + ], + "lines": [ + { + "bbox": [ + 400, + 447, + 452, + 463 + ], + "spans": [ + { + "bbox": [ + 400, + 447, + 452, + 463 + ], + "type": "text", + "content": "Eyes++, Nose-, Hair++" + } + ] + } + ], + "index": 111, + "angle": 0, + "type": "image_caption" + } + ], + "index": 106 + }, + { + "type": "image", + "bbox": [ + 401, + 335, + 445, + 381 + ], + "blocks": [ + { + "bbox": [ + 401, + 335, + 445, + 381 + ], + "lines": [ + { + "bbox": [ + 401, + 335, + 445, + 381 + ], + "spans": [ + { + "bbox": [ + 401, + 335, + 445, + 381 + ], + "type": "image", + "image_path": "e07791129d7142fcad928fc6c93713be12517e7fff1c476287590021ec7941fc.jpg" + } + ] + } + ], + "index": 108, + "angle": 0, + "type": "image_body" + } + ], + "index": 108 + }, + { + "type": "image", + "bbox": [ + 400, + 393, + 441, + 445 + ], + "blocks": [ + { + "bbox": [ + 400, + 393, + 441, + 445 + ], + "lines": [ + { + "bbox": [ + 400, + 393, + 441, + 445 + ], + "spans": [ + { + "bbox": [ + 400, + 393, + 441, + 445 + ], + "type": "image", + "image_path": "3f0b04bfad01a4bd5db9ef5d2be505e0c19cb96af43bac7fb2f46f855b9000a7.jpg" + } + ] + } + ], + "index": 110, + "angle": 0, + "type": "image_body" + } + ], + "index": 110 + }, + { + "type": "image", + "bbox": [ + 452, + 335, + 490, + 381 + ], + "blocks": [ + { + "bbox": [ + 452, + 335, + 490, + 381 + ], + "lines": [ + { + "bbox": [ + 452, + 335, + 490, + 381 + ], + "spans": [ + { + "bbox": [ + 452, + 335, + 490, + 381 + ], + "type": "image", + "image_path": "9f90c5f4153f2249ae183c84f8946c59068cfe25bb24c161b85ecef8953ed105.jpg" + } + ] + } + ], + "index": 112, + "angle": 0, + "type": "image_body" + } + ], + "index": 112 + }, + { + "type": "image", + "bbox": [ + 447, + 394, + 485, + 445 + ], + "blocks": [ + { + "bbox": [ + 470, + 383, + 510, + 393 + ], + "lines": [ + { + "bbox": [ + 470, + 383, + 510, + 393 + ], + "spans": [ + { + "bbox": [ + 470, + 383, + 510, + 393 + ], + "type": "text", + "content": "Rendering" + } + ] + } + ], + "index": 113, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 447, + 394, + 485, + 445 + ], + "lines": [ + { + "bbox": [ + 447, + 394, + 485, + 445 + ], + "spans": [ + { + "bbox": [ + 447, + 394, + 485, + 445 + ], + "type": "image", + "image_path": "3df63afd676cd555cd205036f5c1aab31fe77560826ac343268183e7485dfe5c.jpg" + } + ] + } + ], + "index": 114, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 471, + 449, + 511, + 459 + ], + "lines": [ + { + "bbox": [ + 471, + 449, + 511, + 459 + ], + "spans": [ + { + "bbox": [ + 471, + 449, + 511, + 459 + ], + "type": "text", + "content": "Rendering" + } + ] + } + ], + "index": 115, + "angle": 0, + "type": "image_caption" + } + ], + "index": 114 + }, + { + "type": "image", + "bbox": [ + 493, + 335, + 534, + 381 + ], + "blocks": [ + { + "bbox": [ + 493, + 335, + 534, + 381 + ], + "lines": [ + { + "bbox": [ + 493, + 335, + 534, + 381 + ], + "spans": [ + { + "bbox": [ + 493, + 335, + 534, + 381 + ], + "type": "image", + "image_path": "2c119aca7181f935b6918b36f113cef2e15bd6ac3695574a3494acf4c82d4c6b.jpg" + } + ] + } + ], + "index": 116, + "angle": 0, + "type": "image_body" + } + ], + "index": 116 + }, + { + "type": "image", + "bbox": [ + 494, + 394, + 532, + 445 + ], + "blocks": [ + { + "bbox": [ + 489, + 383, + 511, + 393 + ], + "lines": [ + { + "bbox": [ + 489, + 383, + 511, + 393 + ], + "spans": [ + { + "bbox": [ + 489, + 383, + 511, + 393 + ], + "type": "text", + "content": "ering" + } + ] + } + ], + "index": 117, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 494, + 394, + 532, + 445 + ], + "lines": [ + { + "bbox": [ + 494, + 394, + 532, + 445 + ], + "spans": [ + { + "bbox": [ + 494, + 394, + 532, + 445 + ], + "type": "image", + "image_path": "0684e4d46d041e57c89c0306991a29584c03dbcef7d86208bce1fe1ea9b1d28a.jpg" + } + ] + } + ], + "index": 118, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 489, + 449, + 528, + 459 + ], + "lines": [ + { + "bbox": [ + 489, + 449, + 528, + 459 + ], + "spans": [ + { + "bbox": [ + 489, + 449, + 528, + 459 + ], + "type": "text", + "content": "dering" + } + ] + } + ], + "index": 119, + "angle": 0, + "type": "image_caption" + } + ], + "index": 118 + }, + { + "bbox": [ + 46, + 490, + 287, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 490, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 46, + 490, + 287, + 610 + ], + "type": "text", + "content": "NeRFBlendshape [16] (the bottom two). Roop [12] fails to handle the find-grained geometry change, like hairlines, and lips. Nex3D [50] is able to successfully update the avatar based on the editing, however, changes the untouched part and causes an obvious identity shift. PVP [28] can make edits while preserving the identity, however, the magnitude of change tends to be smaller than the given image. In contrast, our method produces the desired editing effect from the edited image and preserves the multiview consistency and identity of the original face." + } + ] + } + ], + "index": 122 + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "content": "We further show more geometry editing results of our method on avatars in various representations in Fig. 4. Our method supports a convenient way to adjust the size of diverse facial features, such as eyes, mouths, jaw, etc., by editing a single rendered image from the avatars. The edits on 2D images are successfully lifted onto the avatar and rendered across different viewpoints and expression. Please refer to the supplementary Sec. C.3 for detailed visualization." + } + ] + } + ], + "index": 123 + }, + { + "bbox": [ + 305, + 490, + 440, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 490, + 440, + 502 + ], + "spans": [ + { + "bbox": [ + 305, + 490, + 440, + 502 + ], + "type": "text", + "content": "tions of modified head geometry." + } + ] + } + ], + "index": 124 + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "Texutre Editing. We then show our capability in texture editing. We utilize Photoshop, an online makeup app WebBeauty [1], and text-driven editing method Instructpix2pix [7] to modify the texture on 2D renderings. We show comparisons with PVP [28], Roop [12] and Next3D [50] on four distinct heads avatars (INSTA [76] for the upper three and NeRFBblendshape [16] for bottom one) in Fig. 5. Roop [12] is ineffective in transferring non-human-face-like texture, thus failing in all examples. PVP [28] only transfers partial or blurry textures, and also causes shifts across the expression and head poses. Next3D [50] successfully uplifts the texture editing in 2D images sharply. However, it still suffers from the identity shift issue in the lower two heads and a blurred pattern in the upper two heads in Fig. 5. In contrast, our method faithfully paints the complicated texture following the edited image and preserves the identity of the original avatar and consis" + } + ] + } + ], + "index": 125 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8957" + } + ] + } + ], + "index": 126 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 73, + 97, + 121 + ], + "blocks": [ + { + "bbox": [ + 60, + 73, + 97, + 121 + ], + "lines": [ + { + "bbox": [ + 60, + 73, + 97, + 121 + ], + "spans": [ + { + "bbox": [ + 60, + 73, + 97, + 121 + ], + "type": "image", + "image_path": "93c76c917b283731e84b3bf7e8db2622a160ae3e98a088cb642f83e46ad9ea4b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 99, + 74, + 135, + 121 + ], + "blocks": [ + { + "bbox": [ + 99, + 74, + 135, + 121 + ], + "lines": [ + { + "bbox": [ + 99, + 74, + 135, + 121 + ], + "spans": [ + { + "bbox": [ + 99, + 74, + 135, + 121 + ], + "type": "image", + "image_path": "3cf60b8772dd99d5e00a0e2b714a2e6c95bdc15d6ce8270373f5775c1eab50d6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 143, + 73, + 179, + 120 + ], + "blocks": [ + { + "bbox": [ + 143, + 73, + 179, + 120 + ], + "lines": [ + { + "bbox": [ + 143, + 73, + 179, + 120 + ], + "spans": [ + { + "bbox": [ + 143, + 73, + 179, + 120 + ], + "type": "image", + "image_path": "6a2b6efd56c79775607a700590ac5a5e1914bfbae6ffad9cde60200f89298f1c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 181, + 73, + 216, + 120 + ], + "blocks": [ + { + "bbox": [ + 181, + 73, + 216, + 120 + ], + "lines": [ + { + "bbox": [ + 181, + 73, + 216, + 120 + ], + "spans": [ + { + "bbox": [ + 181, + 73, + 216, + 120 + ], + "type": "image", + "image_path": "16507e58984ee4f84362d69add2b356d5cdbb7487ce1d23ba60acb5d13121634.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 219, + 73, + 255, + 120 + ], + "blocks": [ + { + "bbox": [ + 219, + 73, + 255, + 120 + ], + "lines": [ + { + "bbox": [ + 219, + 73, + 255, + 120 + ], + "spans": [ + { + "bbox": [ + 219, + 73, + 255, + 120 + ], + "type": "image", + "image_path": "5dbf690b0d43d237d0804d41963e940875a70cc1e71fec4a073fd96c4a630f02.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 257, + 73, + 291, + 120 + ], + "blocks": [ + { + "bbox": [ + 257, + 73, + 291, + 120 + ], + "lines": [ + { + "bbox": [ + 257, + 73, + 291, + 120 + ], + "spans": [ + { + "bbox": [ + 257, + 73, + 291, + 120 + ], + "type": "image", + "image_path": "5f407eed9fc984e8a369b44e7c274172a6118b7d79f465e0ba278f9ff6a26f94.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 293, + 73, + 330, + 120 + ], + "blocks": [ + { + "bbox": [ + 293, + 73, + 330, + 120 + ], + "lines": [ + { + "bbox": [ + 293, + 73, + 330, + 120 + ], + "spans": [ + { + "bbox": [ + 293, + 73, + 330, + 120 + ], + "type": "image", + "image_path": "5569949685e0a01f72da09cb42840b7d03ab727a447ea5faa8a1a8b8cc2313b5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 64, + 122, + 97, + 168 + ], + "blocks": [ + { + "bbox": [ + 64, + 122, + 97, + 168 + ], + "lines": [ + { + "bbox": [ + 64, + 122, + 97, + 168 + ], + "spans": [ + { + "bbox": [ + 64, + 122, + 97, + 168 + ], + "type": "image", + "image_path": "870a96d8475708e0b1fef9b511f58ac496c7d68d8a93e8b2657eb0b8327cfcf8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 102, + 122, + 135, + 168 + ], + "blocks": [ + { + "bbox": [ + 102, + 122, + 135, + 168 + ], + "lines": [ + { + "bbox": [ + 102, + 122, + 135, + 168 + ], + "spans": [ + { + "bbox": [ + 102, + 122, + 135, + 168 + ], + "type": "image", + "image_path": "b8a0ec6a0aae99d4ba145425c96ab06f38956b5e93d7ad767cd2b9b0ee68996c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 143, + 122, + 178, + 170 + ], + "blocks": [ + { + "bbox": [ + 143, + 122, + 178, + 170 + ], + "lines": [ + { + "bbox": [ + 143, + 122, + 178, + 170 + ], + "spans": [ + { + "bbox": [ + 143, + 122, + 178, + 170 + ], + "type": "image", + "image_path": "82b070bebc38f5049677adf569d8484ed060ce1daa225b9878342df6a445a877.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 182, + 122, + 216, + 170 + ], + "blocks": [ + { + "bbox": [ + 182, + 122, + 216, + 170 + ], + "lines": [ + { + "bbox": [ + 182, + 122, + 216, + 170 + ], + "spans": [ + { + "bbox": [ + 182, + 122, + 216, + 170 + ], + "type": "image", + "image_path": "274d3feccbfff7783ef952bb15d03656d93339cc11f338bd2f481e956786b59d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 219, + 122, + 253, + 170 + ], + "blocks": [ + { + "bbox": [ + 219, + 122, + 253, + 170 + ], + "lines": [ + { + "bbox": [ + 219, + 122, + 253, + 170 + ], + "spans": [ + { + "bbox": [ + 219, + 122, + 253, + 170 + ], + "type": "image", + "image_path": "3736910badfff332d0ef1a64ee9f5b5f07fe25971bc8ec3f5f269bba8a75c553.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 255, + 122, + 288, + 170 + ], + "blocks": [ + { + "bbox": [ + 255, + 122, + 288, + 170 + ], + "lines": [ + { + "bbox": [ + 255, + 122, + 288, + 170 + ], + "spans": [ + { + "bbox": [ + 255, + 122, + 288, + 170 + ], + "type": "image", + "image_path": "8f6b0fe5950c661de7baaf0854f5a3a591797b99c17d77fda0844f5b90a37997.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 291, + 122, + 328, + 170 + ], + "blocks": [ + { + "bbox": [ + 291, + 122, + 328, + 170 + ], + "lines": [ + { + "bbox": [ + 291, + 122, + 328, + 170 + ], + "spans": [ + { + "bbox": [ + 291, + 122, + 328, + 170 + ], + "type": "image", + "image_path": "00fa592bb222b620a7d2ef30e5e4a14dfb165ade6207ad3ab852bb5ecfbe1a17.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 63, + 171, + 97, + 217 + ], + "blocks": [ + { + "bbox": [ + 63, + 171, + 97, + 217 + ], + "lines": [ + { + "bbox": [ + 63, + 171, + 97, + 217 + ], + "spans": [ + { + "bbox": [ + 63, + 171, + 97, + 217 + ], + "type": "image", + "image_path": "da3856b52035ad07ab8f2f23f97db60d8a6be9c4db9061c041193480bda3bb4c.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 102, + 171, + 135, + 217 + ], + "blocks": [ + { + "bbox": [ + 102, + 171, + 135, + 217 + ], + "lines": [ + { + "bbox": [ + 102, + 171, + 135, + 217 + ], + "spans": [ + { + "bbox": [ + 102, + 171, + 135, + 217 + ], + "type": "image", + "image_path": "4c6537a3fa5d70e15d107dbb7c5808be7df9095fbf9cbb0153ece89c28f9a2dd.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 141, + 171, + 177, + 218 + ], + "blocks": [ + { + "bbox": [ + 141, + 171, + 177, + 218 + ], + "lines": [ + { + "bbox": [ + 141, + 171, + 177, + 218 + ], + "spans": [ + { + "bbox": [ + 141, + 171, + 177, + 218 + ], + "type": "image", + "image_path": "854c79f2f36cebfe8c56e6ac71d89066d0f131709d9de4633dda669c36b5551a.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 181, + 171, + 216, + 218 + ], + "blocks": [ + { + "bbox": [ + 181, + 171, + 216, + 218 + ], + "lines": [ + { + "bbox": [ + 181, + 171, + 216, + 218 + ], + "spans": [ + { + "bbox": [ + 181, + 171, + 216, + 218 + ], + "type": "image", + "image_path": "03a44820aecb534d32305107396b568443e567ebeccca15ef81a57eb4d5b2e6f.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 218, + 171, + 254, + 218 + ], + "blocks": [ + { + "bbox": [ + 218, + 171, + 254, + 218 + ], + "lines": [ + { + "bbox": [ + 218, + 171, + 254, + 218 + ], + "spans": [ + { + "bbox": [ + 218, + 171, + 254, + 218 + ], + "type": "image", + "image_path": "40a09dfe949365a35d78d66ec57f0a2eebe8505cdb1b3cedc543739c52e03b2a.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 257, + 171, + 290, + 218 + ], + "blocks": [ + { + "bbox": [ + 257, + 171, + 290, + 218 + ], + "lines": [ + { + "bbox": [ + 257, + 171, + 290, + 218 + ], + "spans": [ + { + "bbox": [ + 257, + 171, + 290, + 218 + ], + "type": "image", + "image_path": "3dc071161141e8dd00b1e90a53b2d64f9c0a7824bbf3126ced18bf205380c7f6.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 292, + 171, + 328, + 218 + ], + "blocks": [ + { + "bbox": [ + 292, + 171, + 328, + 218 + ], + "lines": [ + { + "bbox": [ + 292, + 171, + 328, + 218 + ], + "spans": [ + { + "bbox": [ + 292, + 171, + 328, + 218 + ], + "type": "image", + "image_path": "e2422af44a97c889265828ae743c16a0d75fd2383ef1191c3916bc33273ec7b1.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 64, + 224, + 94, + 265 + ], + "blocks": [ + { + "bbox": [ + 64, + 224, + 94, + 265 + ], + "lines": [ + { + "bbox": [ + 64, + 224, + 94, + 265 + ], + "spans": [ + { + "bbox": [ + 64, + 224, + 94, + 265 + ], + "type": "image", + "image_path": "e41fef444d7908c0cbc5cee7e576dfda7f4f6e6662fb00a3198502cc6b68eb0f.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 65, + 273, + 89, + 280 + ], + "lines": [ + { + "bbox": [ + 65, + 273, + 89, + 280 + ], + "spans": [ + { + "bbox": [ + 65, + 273, + 89, + 280 + ], + "type": "text", + "content": "Source" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 97, + 224, + 134, + 265 + ], + "blocks": [ + { + "bbox": [ + 97, + 224, + 134, + 265 + ], + "lines": [ + { + "bbox": [ + 97, + 224, + 134, + 265 + ], + "spans": [ + { + "bbox": [ + 97, + 224, + 134, + 265 + ], + "type": "image", + "image_path": "68b7b3ad5625c1d3e491ff00895803e4f226c18aed515158736b2f6d5eb7b40d.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 97, + 273, + 134, + 281 + ], + "lines": [ + { + "bbox": [ + 97, + 273, + 134, + 281 + ], + "spans": [ + { + "bbox": [ + 97, + 273, + 134, + 281 + ], + "type": "text", + "content": "2D Editing" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 142, + 220, + 177, + 267 + ], + "blocks": [ + { + "bbox": [ + 142, + 220, + 177, + 267 + ], + "lines": [ + { + "bbox": [ + 142, + 220, + 177, + 267 + ], + "spans": [ + { + "bbox": [ + 142, + 220, + 177, + 267 + ], + "type": "image", + "image_path": "443fe59e0873cc5e8c6833a9c528e78386bfaab33c93d6ed114aeff22796b102.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 269, + 181, + 286 + ], + "lines": [ + { + "bbox": [ + 146, + 269, + 181, + 286 + ], + "spans": [ + { + "bbox": [ + 146, + 269, + 181, + 286 + ], + "type": "text", + "content": "Reference Animation" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 180, + 220, + 216, + 267 + ], + "blocks": [ + { + "bbox": [ + 180, + 220, + 216, + 267 + ], + "lines": [ + { + "bbox": [ + 180, + 220, + 216, + 267 + ], + "spans": [ + { + "bbox": [ + 180, + 220, + 216, + 267 + ], + "type": "image", + "image_path": "f10a02b46c3033b3f461b036bc7bfe28904d20f589cd981a55db39059ffdc4a9.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 193, + 273, + 209, + 281 + ], + "lines": [ + { + "bbox": [ + 193, + 273, + 209, + 281 + ], + "spans": [ + { + "bbox": [ + 193, + 273, + 209, + 281 + ], + "type": "text", + "content": "PVP" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 218, + 220, + 253, + 267 + ], + "blocks": [ + { + "bbox": [ + 218, + 220, + 253, + 267 + ], + "lines": [ + { + "bbox": [ + 218, + 220, + 253, + 267 + ], + "spans": [ + { + "bbox": [ + 218, + 220, + 253, + 267 + ], + "type": "image", + "image_path": "6c01f410daf0aa7c524a0126b7aeb87c6582df5350e211cf675a62d16a7713fc.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 273, + 246, + 281 + ], + "lines": [ + { + "bbox": [ + 227, + 273, + 246, + 281 + ], + "spans": [ + { + "bbox": [ + 227, + 273, + 246, + 281 + ], + "type": "text", + "content": "Roop" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 257, + 224, + 290, + 267 + ], + "blocks": [ + { + "bbox": [ + 257, + 224, + 290, + 267 + ], + "lines": [ + { + "bbox": [ + 257, + 224, + 290, + 267 + ], + "spans": [ + { + "bbox": [ + 257, + 224, + 290, + 267 + ], + "type": "image", + "image_path": "ce6087c99512b74f1caf5b96fdcb54815366c619a23c28fb51135d334a08fbf6.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 260, + 273, + 287, + 281 + ], + "lines": [ + { + "bbox": [ + 260, + 273, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 260, + 273, + 287, + 281 + ], + "type": "text", + "content": "Next3D" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 292, + 221, + 328, + 267 + ], + "blocks": [ + { + "bbox": [ + 292, + 221, + 328, + 267 + ], + "lines": [ + { + "bbox": [ + 292, + 221, + 328, + 267 + ], + "spans": [ + { + "bbox": [ + 292, + 221, + 328, + 267 + ], + "type": "image", + "image_path": "94129cc35166d5abf1e6ca41942d34e597ce4f1a01432dddc92af159950105ec.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 274, + 319, + 284 + ], + "lines": [ + { + "bbox": [ + 302, + 274, + 319, + 284 + ], + "spans": [ + { + "bbox": [ + 302, + 274, + 319, + 284 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 335, + 72, + 372, + 119 + ], + "blocks": [ + { + "bbox": [ + 335, + 72, + 372, + 119 + ], + "lines": [ + { + "bbox": [ + 335, + 72, + 372, + 119 + ], + "spans": [ + { + "bbox": [ + 335, + 72, + 372, + 119 + ], + "type": "image", + "image_path": "69037e33481230528a1f2921e635a940c72097da1718e3eccdc3081d9fadacd4.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 375, + 72, + 411, + 119 + ], + "blocks": [ + { + "bbox": [ + 375, + 72, + 411, + 119 + ], + "lines": [ + { + "bbox": [ + 375, + 72, + 411, + 119 + ], + "spans": [ + { + "bbox": [ + 375, + 72, + 411, + 119 + ], + "type": "image", + "image_path": "1c4472994369a25d4b49ac36951319e3ec5c9d31a5fd1108fcaebc98a7ce2f21.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 413, + 72, + 449, + 119 + ], + "blocks": [ + { + "bbox": [ + 413, + 72, + 449, + 119 + ], + "lines": [ + { + "bbox": [ + 413, + 72, + 449, + 119 + ], + "spans": [ + { + "bbox": [ + 413, + 72, + 449, + 119 + ], + "type": "image", + "image_path": "419fdc01d70ab579a01c704d9c297dbfe7ecd98668dba16c3fa0a2e05f664304.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 451, + 72, + 488, + 119 + ], + "blocks": [ + { + "bbox": [ + 451, + 72, + 488, + 119 + ], + "lines": [ + { + "bbox": [ + 451, + 72, + 488, + 119 + ], + "spans": [ + { + "bbox": [ + 451, + 72, + 488, + 119 + ], + "type": "image", + "image_path": "d67de9d18779d38cfd6e51fdebe8467191d75ddef42b6c71a61eedc544cc590b.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 490, + 72, + 527, + 119 + ], + "blocks": [ + { + "bbox": [ + 490, + 72, + 527, + 119 + ], + "lines": [ + { + "bbox": [ + 490, + 72, + 527, + 119 + ], + "spans": [ + { + "bbox": [ + 490, + 72, + 527, + 119 + ], + "type": "image", + "image_path": "693849cfa7572135edd57630e46cb8c2527460ecc4b70ed938ba416dc67d9b77.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 337, + 122, + 372, + 171 + ], + "blocks": [ + { + "bbox": [ + 337, + 122, + 372, + 171 + ], + "lines": [ + { + "bbox": [ + 337, + 122, + 372, + 171 + ], + "spans": [ + { + "bbox": [ + 337, + 122, + 372, + 171 + ], + "type": "image", + "image_path": "270d379d3cf99c131a41185076bafd2e2e5b1d196cb43dea981f6778bdc7c310.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 374, + 122, + 410, + 171 + ], + "blocks": [ + { + "bbox": [ + 374, + 122, + 410, + 171 + ], + "lines": [ + { + "bbox": [ + 374, + 122, + 410, + 171 + ], + "spans": [ + { + "bbox": [ + 374, + 122, + 410, + 171 + ], + "type": "image", + "image_path": "eaf86ffe2902db415b71848c54e3d37fb0c321425fc3f59801d9da481ac9fa6a.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 414, + 122, + 449, + 171 + ], + "blocks": [ + { + "bbox": [ + 414, + 122, + 449, + 171 + ], + "lines": [ + { + "bbox": [ + 414, + 122, + 449, + 171 + ], + "spans": [ + { + "bbox": [ + 414, + 122, + 449, + 171 + ], + "type": "image", + "image_path": "bedeca8294a3af5e040ffe291d01a1871e6b37e268aa137bb455c56097712ff2.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 452, + 122, + 486, + 171 + ], + "blocks": [ + { + "bbox": [ + 452, + 122, + 486, + 171 + ], + "lines": [ + { + "bbox": [ + 452, + 122, + 486, + 171 + ], + "spans": [ + { + "bbox": [ + 452, + 122, + 486, + 171 + ], + "type": "image", + "image_path": "a998e56ea963608f1354780e84f24a4cb7fded5693d42fb2399b7ac4a6b23c78.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 490, + 122, + 525, + 171 + ], + "blocks": [ + { + "bbox": [ + 490, + 122, + 525, + 171 + ], + "lines": [ + { + "bbox": [ + 490, + 122, + 525, + 171 + ], + "spans": [ + { + "bbox": [ + 490, + 122, + 525, + 171 + ], + "type": "image", + "image_path": "3a3604e48aa3631cb242f6388bc15ba7f3f6ad7e3599665bb2c774f387807f59.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 338, + 171, + 375, + 219 + ], + "blocks": [ + { + "bbox": [ + 338, + 171, + 375, + 219 + ], + "lines": [ + { + "bbox": [ + 338, + 171, + 375, + 219 + ], + "spans": [ + { + "bbox": [ + 338, + 171, + 375, + 219 + ], + "type": "image", + "image_path": "1365f99133e2e83108312a66de1441267bdef4642a87768a8b91d09594682ff1.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 376, + 171, + 412, + 219 + ], + "blocks": [ + { + "bbox": [ + 376, + 171, + 412, + 219 + ], + "lines": [ + { + "bbox": [ + 376, + 171, + 412, + 219 + ], + "spans": [ + { + "bbox": [ + 376, + 171, + 412, + 219 + ], + "type": "image", + "image_path": "791fe9d51ead790c5ea66b7808dafa3129ac078a02feae920e5d23c49d749670.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 416, + 171, + 451, + 219 + ], + "blocks": [ + { + "bbox": [ + 416, + 171, + 451, + 219 + ], + "lines": [ + { + "bbox": [ + 416, + 171, + 451, + 219 + ], + "spans": [ + { + "bbox": [ + 416, + 171, + 451, + 219 + ], + "type": "image", + "image_path": "d0217c83307093b2e54ce669def02d63fe37926f31c14f5bcb307a1e28cc874d.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + } + ], + "index": 47 + }, + { + "type": "image", + "bbox": [ + 454, + 171, + 490, + 219 + ], + "blocks": [ + { + "bbox": [ + 454, + 171, + 490, + 219 + ], + "lines": [ + { + "bbox": [ + 454, + 171, + 490, + 219 + ], + "spans": [ + { + "bbox": [ + 454, + 171, + 490, + 219 + ], + "type": "image", + "image_path": "4f5e235dc0659090ee67172c561047d24efc656044a9d75829c5ca27dbea4765.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 492, + 171, + 530, + 219 + ], + "blocks": [ + { + "bbox": [ + 492, + 171, + 530, + 219 + ], + "lines": [ + { + "bbox": [ + 492, + 171, + 530, + 219 + ], + "spans": [ + { + "bbox": [ + 492, + 171, + 530, + 219 + ], + "type": "image", + "image_path": "2bb2d62cd2e78353554e4f9de7efa3d166f11675bc07c2a566f07eb1756ec4ca.jpg" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_body" + } + ], + "index": 49 + }, + { + "type": "image", + "bbox": [ + 340, + 221, + 373, + 268 + ], + "blocks": [ + { + "bbox": [ + 340, + 221, + 373, + 268 + ], + "lines": [ + { + "bbox": [ + 340, + 221, + 373, + 268 + ], + "spans": [ + { + "bbox": [ + 340, + 221, + 373, + 268 + ], + "type": "image", + "image_path": "5ce42b0aefb7fd52b678606a1827bd9e434285c0cc3b57b8cab40c7e46939a3e.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 270, + 375, + 286 + ], + "lines": [ + { + "bbox": [ + 340, + 270, + 375, + 286 + ], + "spans": [ + { + "bbox": [ + 340, + 270, + 375, + 286 + ], + "type": "text", + "content": "Reference Animation" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_caption" + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 376, + 221, + 411, + 268 + ], + "blocks": [ + { + "bbox": [ + 376, + 221, + 411, + 268 + ], + "lines": [ + { + "bbox": [ + 376, + 221, + 411, + 268 + ], + "spans": [ + { + "bbox": [ + 376, + 221, + 411, + 268 + ], + "type": "image", + "image_path": "449238b1e5b10ba421609285fd247a627168f4d10c21cb99ccc664e9dc47cf0d.jpg" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 388, + 273, + 404, + 281 + ], + "lines": [ + { + "bbox": [ + 388, + 273, + 404, + 281 + ], + "spans": [ + { + "bbox": [ + 388, + 273, + 404, + 281 + ], + "type": "text", + "content": "PVP" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_caption" + } + ], + "index": 51 + }, + { + "type": "image", + "bbox": [ + 419, + 221, + 450, + 268 + ], + "blocks": [ + { + "bbox": [ + 419, + 221, + 450, + 268 + ], + "lines": [ + { + "bbox": [ + 419, + 221, + 450, + 268 + ], + "spans": [ + { + "bbox": [ + 419, + 221, + 450, + 268 + ], + "type": "image", + "image_path": "828046c3acfee0c3ec5a73b6a232e84144d84d0b42149de9fc1bf4d86c720457.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 423, + 273, + 441, + 282 + ], + "lines": [ + { + "bbox": [ + 423, + 273, + 441, + 282 + ], + "spans": [ + { + "bbox": [ + 423, + 273, + 441, + 282 + ], + "type": "text", + "content": "Roop" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_caption" + } + ], + "index": 52 + }, + { + "type": "image", + "bbox": [ + 453, + 221, + 490, + 268 + ], + "blocks": [ + { + "bbox": [ + 453, + 221, + 490, + 268 + ], + "lines": [ + { + "bbox": [ + 453, + 221, + 490, + 268 + ], + "spans": [ + { + "bbox": [ + 453, + 221, + 490, + 268 + ], + "type": "image", + "image_path": "8ecc9323bb53efe8184ae53663e39c7df5791349fd916a49bba37924ca796313.jpg" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 458, + 273, + 485, + 281 + ], + "lines": [ + { + "bbox": [ + 458, + 273, + 485, + 281 + ], + "spans": [ + { + "bbox": [ + 458, + 273, + 485, + 281 + ], + "type": "text", + "content": "Next3D" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_caption" + } + ], + "index": 53 + }, + { + "type": "image", + "bbox": [ + 493, + 221, + 528, + 268 + ], + "blocks": [ + { + "bbox": [ + 493, + 221, + 528, + 268 + ], + "lines": [ + { + "bbox": [ + 493, + 221, + 528, + 268 + ], + "spans": [ + { + "bbox": [ + 493, + 221, + 528, + 268 + ], + "type": "image", + "image_path": "ba6727572b4d052774b7b5d34fca0e70fdc5b1184c972def0aa0b5f012398fc8.jpg" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 501, + 273, + 518, + 281 + ], + "lines": [ + { + "bbox": [ + 501, + 273, + 518, + 281 + ], + "spans": [ + { + "bbox": [ + 501, + 273, + 518, + 281 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_caption" + } + ], + "index": 54 + }, + { + "type": "image", + "bbox": [ + 52, + 319, + 281, + 432 + ], + "blocks": [ + { + "bbox": [ + 46, + 285, + 545, + 308 + ], + "lines": [ + { + "bbox": [ + 46, + 285, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 545, + 308 + ], + "type": "text", + "content": "Figure 5. We compare texture editing with PVP [28], Roop [12], Next3D [50] on INSTA [76] and NeRFBlendshape [16] avatars. The \"Reference Animation\" denotes the image of the original avatar under the same expression with the rendered edited view." + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 52, + 319, + 281, + 432 + ], + "lines": [ + { + "bbox": [ + 52, + 319, + 281, + 432 + ], + "spans": [ + { + "bbox": [ + 52, + 319, + 281, + 432 + ], + "type": "image", + "image_path": "1c072a5148be7a0fdee133d39d607944ffcc76662bc86df7feb7d07d50441b39.jpg" + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "image_body" + } + ], + "index": 61 + }, + { + "type": "image", + "bbox": [ + 52, + 433, + 282, + 492 + ], + "blocks": [ + { + "bbox": [ + 52, + 433, + 282, + 492 + ], + "lines": [ + { + "bbox": [ + 52, + 433, + 282, + 492 + ], + "spans": [ + { + "bbox": [ + 52, + 433, + 282, + 492 + ], + "type": "image", + "image_path": "2d144f5947784f4cdaaa08a753eb470a5db0c221e0ab5b7d24c1612c7250583e.jpg" + } + ] + } + ], + "index": 62, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 494, + 286, + 516 + ], + "lines": [ + { + "bbox": [ + 46, + 494, + 286, + 516 + ], + "spans": [ + { + "bbox": [ + 46, + 494, + 286, + 516 + ], + "type": "text", + "content": "Figure 6. Analysis of our effectiveness with the naive baselines that can accomplish the single-view avatar editing." + } + ] + } + ], + "index": 63, + "angle": 0, + "type": "image_caption" + } + ], + "index": 62 + }, + { + "bbox": [ + 46, + 529, + 249, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 529, + 249, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 529, + 249, + 540 + ], + "type": "text", + "content": "tency across multiple viewpoints and expressions." + } + ] + } + ], + "index": 64 + }, + { + "bbox": [ + 46, + 540, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 287, + 624 + ], + "type": "text", + "content": "We show extensive texture editing results in three avatar representations in Fig. 7. Our method supports a wide range of texture editing, including global style transfer (\"Add a clown makeup\" via text-driven editing), semantic-driven editing (changing the hair color), and free-form sketch (painting on the face). Please refer to the supplementary Sec. C.1/2 for hybrid editing and face reenactment results." + } + ] + } + ], + "index": 65 + }, + { + "bbox": [ + 47, + 628, + 189, + 640 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 628, + 189, + 640 + ], + "spans": [ + { + "bbox": [ + 47, + 628, + 189, + 640 + ], + "type": "text", + "content": "4.3.Quantitative Comparison" + } + ] + } + ], + "index": 66 + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": "User Study. We conducted a user study to validate our method further quantitatively. Following the evaluation protocol of Avatarstudio [38], users are required to watch the rendered videos of different methods side by side and answer each question by picking up one of the methods. For each group of editing results, we will ask four questions the" + } + ] + } + ], + "index": 67 + }, + { + "bbox": [ + 304, + 319, + 545, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 545, + 450 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 545, + 450 + ], + "type": "text", + "content": "same as Avatarstudio [38] on editing preservation, identity preservation, temporal consistency and overall performance in Tab. 1. We collected statistics from 30 participants in 12 groups of edit results. Results are reported in Tab. 1. Our method exhibits the best editing ability while keeping the best consistency across different facial expressions for geometry and texture editing. Moreover, our method performs the best in keeping non-edited parts untouched and making the human heads still recognizable after being edited e.g., identity preservation metric in Tab. 1. Please refer to the supplementary Sec. B.4 for more details." + } + ] + } + ], + "index": 68 + }, + { + "bbox": [ + 304, + 452, + 545, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 452, + 545, + 583 + ], + "spans": [ + { + "bbox": [ + 304, + 452, + 545, + 583 + ], + "type": "text", + "content": "Image Identity Similarity Evaluation. Following the VoLux-GAN [51], we further evaluate the cross-view identity consistency in our edit results. We take 7 groups of geometry editing results and 7 groups of texture editing results, and render the 350 images of edited avatars with different viewpoints and expressions. We calculate the cosine similarities between each rendered image and the single-view 2D edited image and average the similarities on all rendered images as metrics. As reported in Tab. 1, our method outperforms all baselines in recovering desired editing effect and retaining identity consistency." + } + ] + } + ], + "index": 69 + }, + { + "bbox": [ + 305, + 590, + 494, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 590, + 494, + 602 + ], + "spans": [ + { + "bbox": [ + 305, + 590, + 494, + 602 + ], + "type": "text", + "content": "4.4. Native Editing Capability of Avatar" + } + ] + } + ], + "index": 70 + }, + { + "bbox": [ + 304, + 605, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 628 + ], + "type": "text", + "content": "In this section, we analyze the editing capability from the original avatar model and show the necessity of our design." + } + ] + } + ], + "index": 71 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Comparison to 3DMM-based Geometry Editing. One intuitive way to support geometry editing is updating the underlying 3DMM geometry [14]. Here, we investigate this method and verify its capability. Specifically, we run state-of-the-art single image-based 3DMM reconstruction method [75], and update the 3DMM shape parameter of the avatar model with the estimated one. Note that such an ap-" + } + ] + } + ], + "index": 72 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8958" + } + ] + } + ], + "index": 73 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 69, + 532, + 502 + ], + "blocks": [ + { + "bbox": [ + 62, + 69, + 532, + 502 + ], + "lines": [ + { + "bbox": [ + 62, + 69, + 532, + 502 + ], + "spans": [ + { + "bbox": [ + 62, + 69, + 532, + 502 + ], + "type": "image", + "image_path": "b4d2729bd274985130b2a1b9d63c175f1cd732ccc0cc4c0272d23a83936acbf6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 506, + 547, + 529 + ], + "lines": [ + { + "bbox": [ + 46, + 506, + 547, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 547, + 529 + ], + "type": "text", + "content": "Figure 7. We show our texture editing results using the 2D editing method with text-prompt, pattern painting and makeup drawing on INSTA [76] and NeRFBlendshape [16], Next3D [50] avatars." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 539, + 289, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 539, + 289, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 539, + 289, + 635 + ], + "type": "text", + "content": "proach is only available for those avatars that use 3DMM explicitly, and we take INSTA [76] as an example and show the results in Fig. 6 (a). The 3DMM fitting tends to fail when the edited face is out-of-distribution, so the magnitude of editing may not be correct (e.g. the face shape). Changing the 3DMM parameter could also result in blurry rendering from the pre-trained avatar model, which cannot be trivially fixed even after fine-tuning the rendering decoder." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 641, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 289, + 715 + ], + "type": "text", + "content": "Comparison to Fine-tuning for Texture Editing. Texture editing could be done by fine-tuning the avatar with the one-shot edited image. We test this method on avatars from NeRFBlendShape [16] and show the results in Fig. 6 (b). While large structured changes, e.g., hair color, can be edited, detailed editing is largely ignored. Please refer to" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 539, + 523, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 539, + 523, + 552 + ], + "spans": [ + { + "bbox": [ + 305, + 539, + 523, + 552 + ], + "type": "text", + "content": "the supplementary Sec. C.4 for more ablation studies." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 555, + 379, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 555, + 379, + 567 + ], + "spans": [ + { + "bbox": [ + 306, + 555, + 379, + 567 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "text", + "content": "We have proposed a novel generic editing approach that allows users to edit various volumetric head avatar representations from a single image, where an expression-aware modification generator lifts the editing to the 3D avatar while maintaining consistency across multiple expression and viewpoints. As a limitation, we cannot add additional objects (e.g., hat) or modify the hairstyle as shown in our supplementary Sec. C.5, which may be improved by learning extra specialized geometry addition and hair modification generators." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 689, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 715 + ], + "type": "text", + "content": "Acknowledgment: This work was partially supported by the NSFC (No. 62102356) and Ant Group." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8959" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 168, + 102 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 168, + 102 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 168, + 102 + ], + "type": "text", + "content": "[1] Webbar/beauty demo app. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 103, + 287, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 103, + 287, + 169 + ], + "spans": [ + { + "bbox": [ + 53, + 103, + 287, + 169 + ], + "type": "text", + "content": "[2] Rameen Abdal, Hsin-Ying Lee, Peihao Zhu, Mengei Chai, Aliaksandr Siarohin, Peter Wonka, and Sergey Tulyakov. 3davatargan: Bridging domains for personalized editable avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4552-4562, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 171, + 288, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 288, + 225 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 288, + 225 + ], + "type": "text", + "content": "[3] ShahRukh Athar, Zexiang Xu, Kalyan Sunkavalli, Eli Shechtman, and Zhixin Shu. Rignerf: Fully controllable neural 3d portraits. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 20364-20373, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 228, + 288, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 228, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 53, + 228, + 288, + 304 + ], + "type": "text", + "content": "[4] Ziqian Bai, Feitong Tan, Zeng Huang, Kripasindhu Sarkar, Danhang Tang, Di Qiu, Abhimitra Meka, Ruofei Du, Mingsong Dou, Sergio Orts-Escalano, et al. Learning personalized high quality volumetric head avatars from monocular rgb videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16890-16900, 2023. 1, 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 306, + 288, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 306, + 288, + 372 + ], + "spans": [ + { + "bbox": [ + 53, + 306, + 288, + 372 + ], + "type": "text", + "content": "[5] Chong Bao, Yinda Zhang, Bangbang Yang, Tianxing Fan, Zesong Yang, Hujun Bao, Guofeng Zhang, and Zhaopeng Cui. Sine: Semantic-driven image-based nef editing with prior-guided editing field. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20919-20929, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 373, + 287, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 373, + 287, + 417 + ], + "spans": [ + { + "bbox": [ + 53, + 373, + 287, + 417 + ], + "type": "text", + "content": "[6] V Blanz and T Vetter. A morphable model for the synthesis of 3d faces. In 26th Annual Conference on Computer Graphics and Interactive Techniques (SIGGRAPH 1999), pages 187-194. ACM Press, 1999. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 419, + 288, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 419, + 288, + 474 + ], + "spans": [ + { + "bbox": [ + 53, + 419, + 288, + 474 + ], + "type": "text", + "content": "[7] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 2, 3, 4, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 475, + 288, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 475, + 288, + 552 + ], + "spans": [ + { + "bbox": [ + 53, + 475, + 288, + 552 + ], + "type": "text", + "content": "[8] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 554, + 288, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 554, + 288, + 598 + ], + "spans": [ + { + "bbox": [ + 53, + 554, + 288, + 598 + ], + "type": "text", + "content": "[9] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In Proceedings of the European Conference on Computer Vision, pages 333-350. Springer, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 600, + 287, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 600, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 48, + 600, + 287, + 654 + ], + "type": "text", + "content": "[10] Anton Cherepkov, Andrey Voynov, and Artem Babenko. Navigating the gan parameter space for semantic image editing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3671-3680, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 656, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 656, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 656, + 287, + 689 + ], + "type": "text", + "content": "[11] deepfakes. faceswap. https://github.com/deepfakes/faceswap, 2023. Accessed: 2023-10-10. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 690, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 690, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 690, + 287, + 713 + ], + "type": "text", + "content": "[12] deepfakes. roop. SomdevSangwan, 2023. Accessed: 2023-10-10. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[13] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690-4699, 2019. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 172 + ], + "type": "text", + "content": "[14] Yao Feng, Haiwen Feng, Michael J Black, and Timo Bolkart. Learning an animatable detailed 3d face model from in-the-wild images. ACM Transactions on Graphics (ToG), 40(4): 1-13, 2021. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 175, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 229 + ], + "type": "text", + "content": "[15] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 1, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 231, + 545, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 231, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 307, + 231, + 545, + 284 + ], + "type": "text", + "content": "[16] Xuan Gao, Chenglai Zhong, Jun Xiang, Yang Hong, Yudong Guo, and Juyong Zhang. Reconstructing personalized semantic facial nerf models from monocular video. ACM Transactions on Graphics (TOG), 41(6):1-12, 2022. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 286, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 545, + 330 + ], + "type": "text", + "content": "[17] Songwei Ge, Taesung Park, Jun-Yan Zhu, and Jia-Bin Huang. Expressive text-to-image generation with rich text. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7545-7556, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 332, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 332, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 307, + 332, + 545, + 374 + ], + "type": "text", + "content": "[18] Ayaan Haque, Matthew Tancik, Alexei A Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. arXiv preprint arXiv:2303.12789, 2023. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 376, + 545, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 419 + ], + "type": "text", + "content": "[19] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. Advances in neural information processing systems, 33:9841-9850, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 422, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 545, + 454 + ], + "type": "text", + "content": "[20] Amir Hertz, Kfir Aberman, and Daniel Cohen-Or. Delta denoising score. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2328–2337, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 456, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 498 + ], + "type": "text", + "content": "[21] Lianghua Huang, Di Chen, Yu Liu, Yujun Shen, Deli Zhao, and Jingren Zhou.Composer: Creative and controllable image synthesis with composable conditions. arXiv preprint arXiv:2302.09778, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 501, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 501, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 307, + 501, + 545, + 566 + ], + "type": "text", + "content": "[22] Wentao Jiang, Si Liu, Chen Gao, Jie Cao, Ran He, Jiashi Feng, and Shuicheng Yan. Psgan: Pose and expression robust spatial-aware gan for customizable makeup transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5194-5202, 2020. 2, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 568, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 568, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 307, + 568, + 545, + 612 + ], + "type": "text", + "content": "[23] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 613, + 547, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 547, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 547, + 668 + ], + "type": "text", + "content": "[24] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 2, 4, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "type": "text", + "content": "[25] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "8960" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[26] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194-1, 2017. 3, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 161 + ], + "type": "text", + "content": "[27] Tingting Li, Ruihe Qian, Chao Dong, Si Liu, Qiong Yan, Wenwu Zhu, and Liang Lin. Beautygan: Instance-level facial makeup transfer with deep generative adversarial network. In Proceedings of the 26th ACM international conference on Multimedia, pages 645-653, 2018. 2, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 161, + 288, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 161, + 288, + 225 + ], + "spans": [ + { + "bbox": [ + 48, + 161, + 288, + 225 + ], + "type": "text", + "content": "[28] K-E Lin, Alex Trevithick, Keli Cheng, Michel Sarkis, Mohsen Ghafoorian, Ning Bi, Gerhard Reitmayr, and Ravi Ramamoorthi. Pvp: Personalized video prior for editable dynamic portraits using stylegan. In Computer Graphics Forum, page e14890. Wiley Online Library, 2023. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 228, + 288, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 288, + 271 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 288, + 271 + ], + "type": "text", + "content": "[29] Huan Ling, Karsten Kreis, Daiqing Li, Seung Wook Kim, Antonio Torralba, and Sanja Fidler. Editgan: High-precision semantic image editing. Advances in Neural Information Processing Systems, 34:16331-16345, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 272, + 288, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 288, + 316 + ], + "type": "text", + "content": "[30] Pengyang Ling, Lin Chen, Pan Zhang, Huaian Chen, and Yi Jin. Freedrag: Point tracking is not you need for interactive point-based image editing. arXiv preprint arXiv:2307.04684, 2023. 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 316, + 288, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 288, + 358 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 288, + 358 + ], + "type": "text", + "content": "[31] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 360, + 288, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 288, + 414 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 288, + 414 + ], + "type": "text", + "content": "[32] Aryan Mikaeili, Or Perel, Mehdi Safaee, Daniel Cohen-Or, and Ali Mahdavi-Amiri. Sked: Sketch-guided text-based 3d editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14607–14619, 2023. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 415, + 288, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 288, + 469 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 288, + 469 + ], + "type": "text", + "content": "[33] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 471, + 288, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 288, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 288, + 514 + ], + "type": "text", + "content": "[34] Chong Mou, Xintao Wang, Jiechong Song, Ying Shan, and Jian Zhang. Dragondiffusion: Enabling drag-style manipulation on diffusion models. arXiv preprint arXiv:2307.02421, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 514, + 288, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 288, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 288, + 557 + ], + "type": "text", + "content": "[35] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. arXiv preprint arXiv:2201.05989, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 559, + 288, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 288, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 288, + 613 + ], + "type": "text", + "content": "[36] Thao Nguyen, Anh Tuan Tran, and Minh Hoai. Lipstick ain't enough: beyond color matching for in-the-wild makeup transfer. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 13305-13314, 2021. 2, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 614, + 288, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 288, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 288, + 658 + ], + "type": "text", + "content": "[37] Thu Nguyen-Phuoc, Gabriel Schwartz, Yuting Ye, Stephen Lombardi, and Lei Xiao. Alteredavatar: Stylizing dynamic 3d avatars with fast style adaptation. arXiv preprint arXiv:2305.19245, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "text", + "content": "[38] Mohit Mendiratta Pan, Mohamed Elgharib, Kartik Teotia, Ayush Tewari, Vladislav Golyanik, Adam Kortylewski, Christian Theobalt, et al. Avatarstudio: Text-driven editing of 3d dynamic human head avatars. arXiv preprint arXiv:2306.00547, 2023. 2, 7" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 715 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "text", + "content": "[39] Xingang Pan, Ayush Tewari, Thomas Leimkuhler, Lingjie Liu, Abhimitra Meka, and Christian Theobalt. Drag your gan: Interactive point-based manipulation on the generative image manifold. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2, 4, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 545, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 174 + ], + "type": "text", + "content": "[40] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 175, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 228 + ], + "type": "text", + "content": "[41] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021. 2, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 230, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 545, + 285 + ], + "type": "text", + "content": "[42] Ivan Perov, Daiheng Gao, Nikolay Chervoniy, Kunlin Liu, Sugasa Marangonda, Chris Umé, Mr Dpfks, Carl Shift Facenheim, Luis RP, Jian Jiang, et al. Deepfacelab: Integrated, flexible and extensible face-swapping framework. arXiv preprint arXiv:2005.05535, 2020. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 286, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 545, + 330 + ], + "type": "text", + "content": "[43] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on graphics (TOG), 42(1):1-13, 2022. 4, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 331, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 545, + 374 + ], + "type": "text", + "content": "[44] Sara Fridovich-Keil and Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance Fields without Neural Networks. In CVPR, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 376, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 431 + ], + "type": "text", + "content": "[45] Ruizhi Shao, Jingxiang Sun, Cheng Peng, Zerong Zheng, Boyao Zhou, Hongwen Zhang, and Yebin Liu. Control4d: Dynamic portrait editing by learning 4d gan from 2d diffusion-based editor. arXiv preprint arXiv:2305.20082, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 433, + 545, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 433, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 307, + 433, + 545, + 477 + ], + "type": "text", + "content": "[46] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9243-9252, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 478, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 478, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 478, + 545, + 521 + ], + "type": "text", + "content": "[47] Yujun Shi, Chuhui Xue, Jiachun Pan, Wenqing Zhang, Vincent YF Tan, and Song Bai. Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. arXiv preprint arXiv:2306.14435, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 522, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 522, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 522, + 545, + 578 + ], + "type": "text", + "content": "[48] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 579, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 545, + 624 + ], + "type": "text", + "content": "[49] Jingxiang Sun, Xuan Wang, Yichun Shi, Lizhen Wang, Jue Wang, and Yebin Liu. Ide-3d: Interactive disentangled editing for high-resolution 3d-aware portrait synthesis. arXiv preprint arXiv:2205.15517, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 624, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 689 + ], + "type": "text", + "content": "[50] Jingxiang Sun, Xuan Wang, Lizhen Wang, Xiaoyu Li, Yong Zhang, Hongwen Zhang, and Yebin Liu. Next3d: Generative neural texture rasterization for 3d-aware head avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20991-21002, 2023. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 691, + 545, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 715 + ], + "type": "text", + "content": "[51] Feitong Tan, Sean Fanello, Abhimitra Meka, Sergio Orts-Escolano, Danhang Tang, Rohit Pandey, Jonathan Taylor," + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8961" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "Ping Tan, and Yinda Zhang. Volux-gan: A generative model for 3d face synthesis with hdri relighting. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "text", + "content": "[52] Junshu Tang, Bo Zhang, Binxin Yang, Ting Zhang, Dong Chen, Lizhuang Ma, and Fang Wen. 3dfaceshop: Explicitly controllable 3d-aware portrait generation. IEEE Transactions on Visualization and Computer Graphics, 2023. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 228 + ], + "type": "text", + "content": "[53] Ziyan Wang, Timur Bagautdinov, Stephen Lombardi, Tomas Simon, Jason Saragih, Jessica Hodgins, and Michael Zollhofer. Learning compositional radiance fields of dynamic human heads. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5704-5713, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 230, + 287, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 287, + 284 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 287, + 284 + ], + "type": "text", + "content": "[54] Yue Wu, Yu Deng, Jiaolong Yang, Fangyun Wei, Qifeng Chen, and Xin Tong. Anifacegan: Animatable 3d-aware face image generation for video avatars. Advances in Neural Information Processing Systems, 35:36188-36201, 2022. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 286, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 287, + 330 + ], + "type": "text", + "content": "[55] Weihao Xia, Yulun Zhang, Yujiu Yang, Jing-Hao Xue, Bolei Zhou, and Ming-Hsuan Yang. Gan inversion: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3121-3138, 2022. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 332, + 287, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 332, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 332, + 287, + 387 + ], + "type": "text", + "content": "[56] Yinghao Xu, Sida Peng, Ceyuan Yang, Yujun Shen, and Bolei Zhou. 3d-aware image synthesis via learning structural and textural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18430-18439, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 388, + 287, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 442 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 442 + ], + "type": "text", + "content": "[57] Yuelang Xu, Lizhen Wang, Xiaochen Zhao, Hongwen Zhang, and Yebin Liu. Avatarmav: Fast 3d head avatar reconstruction using motion-aware neural voxels. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-10, 2023. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 445, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 287, + 510 + ], + "type": "text", + "content": "[58] Bangbang Yang, Yinda Zhang, Yinghao Xu, Yijin Li, Han Zhou, Hujun Bao, Guofeng Zhang, and Zhaopeng Cui. Learning object-compositional neural radiance field for ed-itable scene rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13779-13788, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 512, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 287, + 567 + ], + "type": "text", + "content": "[59] Bangbang Yang, Chong Bao, Junyi Zeng, Hujun Bao, Yinda Zhang, Zhaopeng Cui, and Guofeng Zhang. Neumesh: Learning disentangled neural mesh-based implicit field for geometry and texture editing. In European Conference on Computer Vision, pages 597-614. Springer, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 568, + 287, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 632 + ], + "type": "text", + "content": "[60] Bangbang Yang, Yinda Zhang, Yijin Li, Zhaopeng Cui, Sean Fanello, Hujun Bao, and Guofeng Zhang. Neural rendering in a room: amodal 3d understanding and free-viewpoint rendering for the closed scene composed of pre-captured objects. ACM Transactions on Graphics (TOG), 41(4):1-10, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "text", + "content": "[61] Chenyu Yang, Wanrong He, Yingqing Xu, and Yang Gao. Elegant: Exquisite and locally editable gan for makeup transfer. In European Conference on Computer Vision, pages 737-754. Springer, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[62] Xingrui Yang, Hai Li, Hongjia Zhai, Yuhang Ming, Yuqian Liu, and Guofeng Zhang. Vox-fusion: Dense tracking and mapping with voxel-based neural implicit representation. In" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "2022 IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 499-507. IEEE, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 97, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 97, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 97, + 545, + 161 + ], + "type": "text", + "content": "[63] Weicai Ye, Shuo Chen, Chong Bao, Hujun Bao, Marc Pollefeys, Zhaopeng Cui, and Guofeng Zhang. Intrinsicerf: Learning intrinsic neural radiance fields for editable novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 339-351, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "type": "text", + "content": "[64] Deheng Zhang, Clara Fernandez-Labrador, and Christopher Schroers. Coarf: Controllable 3d artistic style transfer for radiance fields. In 2024 International Conference on 3D Vision (3DV). IEEE, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 209, + 545, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 252 + ], + "type": "text", + "content": "[65] Hao Zhang, Yanbo Xu, Tianyuan Dai, Tai Chi-Keung Tang, et al. Fdnerf: Semantics-driven face reconstruction, prompt editing and relighting with diffusion models. arXiv preprint arXiv:2306.00783, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 254, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 254, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 254, + 545, + 308 + ], + "type": "text", + "content": "[66] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Nefactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (TOG), 40(6):1-18, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 310, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 310, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 307, + 310, + 545, + 376 + ], + "type": "text", + "content": "[67] Boming Zhao, Bangbang Yang, Zhenyang Li, Zuoyue Li, Guofeng Zhang, Jiashu Zhao, Dawei Yin, Zhaopeng Cui, and Hujun Bao. Factorized and controllable neural re-rendering of outdoor scene for photo extrapolation. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1455-1464, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 377, + 545, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 377, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 307, + 377, + 545, + 432 + ], + "type": "text", + "content": "[68] Yufeng Zheng, Victoria Fernández Abrevaya, Marcel C Bühler, Xu Chen, Michael J Black, and Otmar Hilliges. Im avatar: Implicit morphable head avatars from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13545-13555, 2022. 1, 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 434, + 545, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 434, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 307, + 434, + 545, + 488 + ], + "type": "text", + "content": "[69] Yufeng Zheng, Wang Yifan, Gordon Wetzstein, Michael J Black, and Otmar Hilliges. Pointavatar: Deformable point-based head avatars from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21057-21067, 2023. 1, 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 490, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 490, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 490, + 545, + 533 + ], + "type": "text", + "content": "[70] Jiapeng Zhu, Yujun Shen, Deli Zhao, and Bolei Zhou. Indomain gan inversion for real image editing. In European conference on computer vision, pages 592-608. Springer, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 535, + 545, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 535, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 307, + 535, + 545, + 590 + ], + "type": "text", + "content": "[71] Jiapeng Zhu, Ceyuan Yang, Yujun Shen, Zifan Shi, Bo Dai, Deli Zhao, and Qifeng Chen. Linkgan: Linking gan latents to pixels for controllable image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7656-7666, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 591, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 591, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 307, + 591, + 545, + 644 + ], + "type": "text", + "content": "[72] Peihao Zhu, Rameen Abdal, Yipeng Qin, and Peter Wonka. Sean: Image synthesis with semantic region-adaptive normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5104-5113, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "type": "text", + "content": "[73] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12786-12796, 2022. 2" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8962" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 229 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[74] Jingyu Zhuang, Chen Wang, Lingjie Liu, Liang Lin, and Guanbin Li. Dreameditor: Text-driven 3d scene editing with neural fields. arXiv preprint arXiv:2306.13455, 2023. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 148 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 148 + ], + "type": "text", + "content": "[75] Wojciech Zielonka, Timo Bolkart, and Justus Thies. Towards metrical reconstruction of human faces. In European Conference on Computer Vision, pages 250–269. Springer, 2022. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 196 + ], + "type": "text", + "content": "[76] Wojciech Zielonka, Timo Bolkart, and Justus Thies. Instant volumetric head avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4574-4584, 2023. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 198, + 287, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 287, + 229 + ], + "type": "text", + "content": "[77] zllrunning. face-makeup.pytorch. https://github.com/zllrunning/face-makeup.PyTorch, 2023. Accessed: 2023-10-10. 4" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8963" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/General Object Foundation Model for Images and Videos at Scale/0bc99999-be81-4577-bff4-eaa67b1ee51f_content_list.json b/2024/General Object Foundation Model for Images and Videos at Scale/0bc99999-be81-4577-bff4-eaa67b1ee51f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a948e8186b03d0a8b75b1c7383936a22c711eaca --- /dev/null +++ b/2024/General Object Foundation Model for Images and Videos at Scale/0bc99999-be81-4577-bff4-eaa67b1ee51f_content_list.json @@ -0,0 +1,1603 @@ +[ + { + "type": "text", + "text": "General Object Foundation Model for Images and Videos at Scale", + "text_level": 1, + "bbox": [ + 153, + 130, + 818, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Junfeng Wu $^{1*}$ , Yi Jiang $^{2*}$ , Qihao Liu $^{3}$ , Zehuan Yuan $^{2}$ , Xiang Bai $^{1\\dagger}$ , Song Bai $^{2\\dagger}$ $^{1}$ Huazhong University of Science and Technology, $^{2}$ ByteDance Inc., $^{3}$ Johns Hopkins University", + "bbox": [ + 153, + 179, + 815, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 251, + 313, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present GLEE in this work, an object-level foundation model for locating and identifying objects in images and videos. Through a unified framework, GLEE accomplishes detection, segmentation, tracking, grounding, and identification of arbitrary objects in the open world scenario for various object perception tasks. Adopting a cohesive learning strategy, GLEE acquires knowledge from diverse data sources with varying supervision levels to formulate general object representations, excelling in zero-shot transfer to new data and tasks. Specifically, we employ an image encoder, text encoder, and visual prompter to handle multi-modal inputs, enabling to simultaneously solve various object-centric downstream tasks while maintaining state-of-the-art performance. Demonstrated through extensive training on over five million images from diverse benchmarks, GLEE exhibits remarkable versatility and improved generalization performance, efficiently tackling downstream tasks without the need for task-specific adaptation. By integrating large volumes of automatically labeled data, we further enhance its zero-shot generalization capabilities. Additionally, GLEE is capable of being integrated into Large Language Models, serving as a foundational model to provide universal object-level information for multi-modal tasks. We hope that the versatility and universality of our method will mark a significant step in the development of efficient visual foundation models for AGI systems. The models and code are released at https://github.com/FoundationVision/GLEE.", + "bbox": [ + 76, + 282, + 473, + 705 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 734, + 209, + 752 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Foundation model [7] is an emerging paradigm for building artificial general intelligence (AGI) systems, signifying a model trained on broad data that is capable of being adapted to a wide range of downstream tasks in an general paradigm. Recently, NLP foundation models such as BERT [21], GPT-3 [9], T5 [70] developed with unified", + "bbox": [ + 75, + 761, + 468, + 853 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d775f818d6f3630be9dc6f16488fa0039086e3fee565a2cb231ebaa038f24436.jpg", + "image_caption": [ + "Figure 1. The performance of GLEE on a broad range of object-level tasks compared with existing models." + ], + "image_footnote": [], + "bbox": [ + 514, + 250, + 875, + 441 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "input-output paradigms and large-scale pre-training, have achieved remarkable generalization capabilities to address nearly all NLP tasks.", + "bbox": [ + 498, + 503, + 890, + 547 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In computer vision, the diversity of task types and the lack of a unified form make visual foundation models only serve specific subdomains, such as CLIP [69] for multimodal visual model, MAE [32] for visual representations model, SAM [39] for segmentation model. Despite being widely studied, current visual foundation models are still focusing on establishing correlations between global image features and language descriptions or learning image-level feature representations. However, locating and identifying objects constitute foundational capabilities in computer vision systems, serves as a basis for solving complex or high level vision tasks such as segmentation, scene understanding, object tracking, event detection, and activity recognition and support a wide range of applications.", + "bbox": [ + 496, + 550, + 892, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we advance the development of object-level foundation models within the visual domain. To address the aforementioned limitation, providing general and accurate object-level information, we introduce a general object visual foundation model, coined as GLEE, which simultaneously solve a wide range of object-centric tasks while ensuring SOTA performance, including object detection, instance segmentation, grounding, object tracking, interactive segmentation and tracking, etc., as shown in Figure 1. Through", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal Contribution. This work was done when Junfeng Wu and Qihao Liu were interns at ByteDance. †Correspondence to Xiang Bai and Song Bai .", + "bbox": [ + 75, + 862, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3783", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "a unified input and output paradigm definition, our model is capable of learning from a wide range of diverse data and predicting general object representations, which makes it generalize well to new data and tasks in a zero-shot manner and achieve amazing performance. In addition, thanks to the unified paradigm, the training data can be scaled up at low cost by introducing a large amount of automatically labeled data, and further improve the zero-shot generalization ability of the model.", + "bbox": [ + 75, + 90, + 470, + 224 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A general object foundation model framework. Our objective is to build an object visual foundation model capable of simultaneously addressing a wide range of object-centric tasks. Specifically, we employ an image encoder, a text encoder, and a visual prompter to encode multi-modal inputs. They are integrated into a detector to extract objects from images according to textual and visual input. This unified approach to handle multiple modalities enables us to concurrently solve various object-centric tasks, including detection [10, 52, 81, 119], instance segmentation [15, 31], referring expression comprehension [35, 55, 93, 118], interactive segmentation [1, 12, 122], multi-object tracking [20, 60, 99, 113, 116], video object segmentation [16, 17, 65, 98], video instance segmentation [34, 87, 90, 92, 102], and video referring segmentation [77, 91, 93], all while maintaining state-of-the-art performance.", + "bbox": [ + 75, + 229, + 470, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-granularity joint supervision and scalable training paradigm. The design of the unified framework capable of addressing multiple tasks enables joint training on over five million images from diverse benchmarks and varying levels of supervision. Existing datasets differ in annotation granularity: detection datasets like Objects365 [79] and OpenImages [42] offer bounding boxes and category names; COCO [52] and LVIS [29] provide finer-grained mask annotations; RefCOCO [64, 108] and Visual Genome [40] provide detailed object descriptions. Additionally, video data enhance the temporal consistency of model, while open-world data contribute class-agnostic object annotations. An intuitive display of the supervision types and data scales of the datasets employed is presented in Figure 2. The unified support for multi-source data in our approach greatly facilitates the incorporation of additional manually or automatically annotated data, enabling easy scaling up of the dataset. Furthermore, the alignment of model optimization across tasks means that joint training serves not only as a unifying strategy but also as a mechanism to boost performance across individual tasks.", + "bbox": [ + 75, + 474, + 470, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Strong zero-shot transferability to a wide range of object level image and video tasks. After joint training on data from diverse sources, GLEE demonstrates remarkable versatility and zero-shot generalization abilities. Extensive experiments demonstrate that GLEE achieves state-of-the-art performance compared to existing specialist and generalist models in object-level image tasks such as detec", + "bbox": [ + 75, + 795, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tion, referring expression comprehension, and open-world detection, all without requiring any task-specific designs or fine-tuning. Furthermore, we showcase the extraordinary generalization and zero-shot capabilities of GLEE in large-vocabulary open-world video tracking tasks, achieving significantly superior performance over existing models even in a zero-shot transfer manner. Additionally, by incorporating automatically annotated data like SA1B [39] and GRIT [67], we are able to scale up our training dataset to an impressive size of 10 million images at a low cost, which is typically challenging to achieve for object-level tasks and further enhances the generalization performance. Moreover, we replace the SAM [39] component with GLEE in a multimodal Large Language Model (mLLM) [43] and observe that it achieves comparable results. This demonstrates that GLEE is capable of supplying the visual object-level information that modern LLMs currently lack, thus laying a solid foundation for an object-centric mLLMs.", + "bbox": [ + 496, + 90, + 893, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 380, + 640, + 396 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Visual Foundation Model", + "text_level": 1, + "bbox": [ + 500, + 407, + 732, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As foundation models [9, 18, 21, 70, 82] in the NLP field have achieved remarkable success, the construction of visual foundation models attracts increasing attention. Unlike NLP tasks that are predominantly unified under a text-to-text paradigm, tasks in Computer Vision still exhibit significant differences in formulation. This disparity leads to the fact that visual models are typically trained in a single-task learning frameworks, limiting their applicability to tasks within certain sub-domains. For instance, multi-modal visual foundation models like CLIP [69], ALIGN [37], Florence [109], BEIT3 [86], Flamingo[2] make significant advancements in efficient transfer learning and demonstrate impressive zero-shot capabilities on vision-language tasks by employing contrastive learning and masked data modeling on large-scale image-text pairs. DALL-E [71, 72] and Stable Diffusion [74] are trained on massive pairs of images and captions, enabling them to generate detailed image content conditioned on textual instruction. DINO [11], MAE [32], EVA [25], ImageGPT [13] obtain strong visual representations through self-supervised training on largescale image data, which are then employed to adopt downstream tasks. These foundation models learn image-level features and are not directly applicable to object-level tasks. The recently proposed SAM [39], capable of segmenting any object of a given image based on visual prompt such as points and boxes, provides rich object-level information and demonstrates strong generalization capabilities. However, the object information lacks semantic context, limiting its application in object-level tasks. Unlike existing visual foundation models, we aim to develop an object foundation model that directly solve downstream tasks without the need", + "bbox": [ + 496, + 431, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3784", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ec3537dc2cf9d5bae65c56ad656596c2164510ac2cca85a3b8aa8f6dcbb89c8b.jpg", + "image_caption": [ + "Figure 2. An illustrative example showcasing annotations of varying granularities from different datasets, along with the scale of data we utilized. Training on datasets from multiple sources endows the model with more universal representations." + ], + "image_footnote": [], + "bbox": [ + 98, + 88, + 867, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "for additional parameters or fine-tuning.", + "bbox": [ + 76, + 513, + 344, + 529 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Unified and General Model", + "text_level": 1, + "bbox": [ + 76, + 542, + 321, + 558 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unified models share similarities with foundation models in the aspect of multi-task unification for their ability to handle multiple vision or multi-modal tasks within a single model. MuST [27] and Intern [78] propose to train across multiple vision tasks and solving them simultaneously. Inspired by the success of sequence-to-sequence NLP models [9, 70], models such as Uni-Perceiver [120], OFA [84], Unified-IO [58], Pix2Seq v2 [14], and UniTAB [103] propose modeling various tasks as sequence generation tasks within a unified paradigm. While these approaches have demonstrated promising cross-task generalization capabilities, they focus mainly on image-level understanding tasks. In addition, their auto-regressive generation of boxes and masks results in significantly slower inference speeds and the performance still falls short of state-of-the-art task-specific models. Building upon on detectors [45, 119], UniPerceiver v2 [46] and UNINEXT [100] utilize unified maximum likelihood estimation and object retrieval to support various tasks, effectively resolves the challenges of localization. Nonetheless, they are trained on closed-set data, thereby not exhibiting zero-shot generalization capabilities. X-decoder [121] and SEEM [122] construct a generalized", + "bbox": [ + 75, + 568, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "decoding model capable of predicting pixel-level segmentation and language tokens. Diverging from unified models, the proposed GLEE not only directly addresses object-level tasks in a unified manner but also provides universal object representations, which generalize well to new data and tasks, serving as a cornerstone for a broader range of tasks that require detailed object information.", + "bbox": [ + 496, + 513, + 893, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Vision-Language Understanding", + "text_level": 1, + "bbox": [ + 498, + 633, + 785, + 650 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Open-vocabulary detection (OVD) and Grounding models both necessitate the localization and recognition of as many objects as possible. With the recent advancements in vision-language pre-training [37, 69, 107, 109], a commonly employed strategy for OVD involves transferring the knowledge from pre-trained vision-language models (VLMs) to object detectors [28, 41, 63]. Another group of studies leverages extensive image-text pair datasets to broaden the detection vocabulary [26, 47, 51, 59, 101, 105, 110, 115]. However, these language-based detectors are inherently constrained by the capabilities and biases of language models, making it challenging to excel simultaneously in both localization and recognition. Our objective is to optimally utilize existing datasets to construct a general object-level foundation model, aims to not only detect and identify objects effectively but also to offer universal object representations.", + "bbox": [ + 496, + 657, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3785", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tations for a wide range of downstream tasks.", + "bbox": [ + 76, + 90, + 379, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 119, + 169, + 135 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Formulation", + "text_level": 1, + "bbox": [ + 76, + 145, + 210, + 159 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The proposed GLEE consists of an image encoder, a text encoder, a visual prompter, and an object decoder, as illustrated in Figure 3. The text encoder processes arbitrary descriptions related to the task, including object categories, names in any form, captions about objects, and referring expressions. The visual prompter encodes user inputs such as points, bounding boxes, or scribbles during interactive segmentation into corresponding visual representations of target objects. Then they are integrated into a detector for extracting objects from images according to textual and visual input. We build the object decoder upon MaskDINO [45] with a dynamic class head by compute similarity between object embedding from detector and text embedding from the text encoder. Given an input image $I \\in \\mathcal{R}^{3 \\times H \\times W}$ , we first extract multi-scale features $Z$ with backbones such as ResNet [30]. Then we feed them into the object decoder and adopt three prediction heads (classification, detection, and segmentation) on the output embedding $q_{d} \\in \\mathcal{R}^{N \\times C}$ from decoder. Following other object segmentation models [15, 45, 50], we construct a 1/4 resolution pixel embedding map $M_{p} \\in \\mathcal{R}^{C \\times \\frac{H}{4} \\times \\frac{W}{4}}$ which is obtained by upsampling and fusing multi-scale feature maps from the backbone and Transformer encoder. Finally, we obtain each binary mask prediction $m \\in \\mathcal{R}^{N \\times \\frac{H}{4} \\times \\frac{W}{4}}$ via a dot product between the N mask embeddings and pixel embedding map:", + "bbox": [ + 76, + 167, + 472, + 547 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nm = F F N \\left(q _ {d}\\right) \\otimes M _ {p}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 574, + 468, + 592 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where FFN is a 3-layer feed forward head with ReLU activation function and a linear projection layer.", + "bbox": [ + 76, + 601, + 468, + 632 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To support arbitrary vocabularies and object descriptions, we replace the FFN classifier with text embeddings following DetCLIP [104]. Specifically, we feed K category names as separate sentences into the text encoder $Enc_{L}$ and use the average of each sentence tokens as the output text embedding $e_t \\in \\mathcal{R}^{K \\times D}$ for each category or description. Then we compute the alignment scores $S_{\\text{align}} \\in \\mathcal{R}^{N \\times K}$ between object embedding and text embedding:", + "bbox": [ + 76, + 632, + 468, + 753 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS _ {\\text {a l i g n}} = q _ {d} \\cdot W _ {i 2 t} \\otimes e _ {t}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 766, + 468, + 782 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $W_{i2t} \\in \\mathcal{R}^{C \\times D}$ is image-to-text projection weights. We use logits $S_{\\text{align}}$ to replace traditional classification logits to compute Hungarian matching cost during training and assign categories to objects during inference. To make the original visual features prompt-aware, an early fusion module is adopted before Transformer encoder following UNNEXT [100], which takes image feature from backbone", + "bbox": [ + 76, + 794, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and prompt embedding as input and perform bi-directional cross-attention between them.", + "bbox": [ + 498, + 90, + 890, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Task Unification", + "text_level": 1, + "bbox": [ + 500, + 133, + 663, + 148 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the above designs, GLEE can be used to seamlessly unify a wide range of object perception tasks in images and videos, including object detection, instance segmentation, grounding, multi-target tracking (MOT), video instance segmentation (VIS), video object segmentation (VOS), interactive segmentation and tracking, and supports open-world/large-vocabulary image and video detection and segmentation tasks.", + "bbox": [ + 496, + 157, + 890, + 277 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Detection and Instance Segmentation. For detection task, a fixed-length category list is given and all objects in the category list are required to be detected. For a dataset with category list length $K$ , the text input can be formulated as $\\{p_k\\}_{k=1}^K$ where $p_k$ represents for the $k$ -th category name, e.g., $P = [\"person\", \"bicycle\", \"car\", ..., \"toothbrush\"]$ for COCO [52]. For datasets with large vocabulary, calculating the text embedding of all categories is very time-consuming and redundant. Therefore, for datasets with a category number greater than 100, such as objects365 [79] and LVIS [29], suppose there are $\\hat{K}$ positive categories in an image, we take the $\\hat{K}$ positive categories and then pad the category number to 100 by randomly sampling from the negative categories. For instance segmentation, we enable the mask branch and add mask matching cost with mask loss.", + "bbox": [ + 496, + 279, + 892, + 505 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Grounding and Referring Segmentation. These tasks provide reference textual expressions, where objects are described with attributes, for example, Referring Expression Comprehension (REC) [108, 118], Referring Expression Segmentation (RES) [55, 108], and Referring Video Object Segmentation (R-VOS) [77, 91] aim at finding objects matched with the given language expressions like \"The fourth person from the left\". For each image, we take the all the object expressions as text prompt and feed the them into the text encoder. For each expressions, we apply global average pooling along the sequence dimension to get text embedding $e_t$ . The text embeddings are feed into early fusion module and additionally interact with object queries through self-attention module in the decoder.", + "bbox": [ + 496, + 507, + 893, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "MOT and VIS. Both Multi-object Tracking (MOT)[4, 20, 60, 113, 116] and Video Instance Segmentation (VIS)[34, 68, 92, 102] need to detect and track all the objects in the predefined category list, and VIS requires additional mask for the objects. These two tasks can be considered as extended tasks of detection and instance segmentation on videos. We found that with sufficient image exposure, object embeddings from the decoder effectively distinguish objects in a video, showing strong discriminability and temporal consistency. As a result, they can be directly employed for tracking without the need for an additional tracking head. Training on image-level data can address", + "bbox": [ + 496, + 719, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3786", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "category: [person, car, motorcycle, dog ...]", + "bbox": [ + 127, + 95, + 325, + 108 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "arbitrary name: bollard, manhole cover", + "bbox": [ + 135, + 112, + 313, + 123 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "expression: motorcycle parked under the sign", + "bbox": [ + 135, + 128, + 336, + 138 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "object caption: air conditioner outdoor unit", + "bbox": [ + 135, + 142, + 326, + 151 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d77eac373c21a127a77e5f9ed1534ad09fedd8b60ecf6ae60d3aed317cd2d1d6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 184, + 152, + 287, + 210 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fd5bf6f7d382a23c2b209900ad9288d278edfa4a146ecd6445141003592729fa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 135, + 210, + 189, + 243 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6679f185efcc4c420e357c2b3a550885b414f1f11be178a3d04d42fd6cbf0546.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 200, + 212, + 243, + 242 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3504d38790bbdfc686d18d7c87b5544f69d5c75ca375cddaf498c1bd6cbde8b3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 254, + 212, + 310, + 243 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/825e83f34e9b6d73c922fd1650abbdafe1b6a28bec1a6d8ced44ae1560a4f802.jpg", + "image_caption": [ + "(a) Object foundation model framework" + ], + "image_footnote": [], + "bbox": [ + 344, + 97, + 816, + 243 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "category: person, car motorcycle,dog, cat.", + "bbox": [ + 117, + 266, + 220, + 287 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f4678e4883caeb60898c1376f7bf588a0ea878e5aead4d427a18537bbe48d4dc.jpg", + "image_caption": [ + "Figure 3. Framework of GLEE. The text encoder accepts textual descriptions in various forms from diverse data sources, including object categories, names, captions, and referring expressions. The visual prompt encodes points, bounding boxes, or scribbles into corresponding visual representations. The object decoder takes them and image features to predict objects in images. (b) illustrates the application of GLEE to image tasks tailored for different language descriptions and visual prompts. (c) demonstrates the application across various object-level video tasks." + ], + "image_footnote": [], + "bbox": [ + 107, + 289, + 174, + 347 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a13df688cdf1ec2d485edd19cd26c30d065014904c61b37d09462586d44b3d1c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 174, + 289, + 233, + 347 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "arbitrary name", + "bbox": [ + 241, + 267, + 313, + 279 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "bollard", + "bbox": [ + 241, + 280, + 279, + 291 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "manhole cover", + "bbox": [ + 300, + 280, + 341, + 295 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a9fa2576b35dd255726e944333ef5f7b8fff289815599f3faa3fda15a177bb4c.jpg", + "image_caption": [ + "(b) Applied to image tasks" + ], + "image_footnote": [], + "bbox": [ + 238, + 295, + 344, + 347 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "expression: motorcycle parked under the sign", + "bbox": [ + 357, + 266, + 452, + 295 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/09d3d65d6993c2444d0a26814dc06a76a9de79e59fc7d90e24780507f945bdd2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 295, + 431, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "interactive segment", + "bbox": [ + 457, + 266, + 511, + 295 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0cd06ba59cf65c2ef56bf0a031d4d5d0b7d037f876d0b73a70fae36756d742bd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 442, + 295, + 519, + 352 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "video tasks: VIS、MOT、VOS、RVOS、 open world/vocabulary tracking", + "bbox": [ + 527, + 267, + 730, + 291 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b0c0fbdb3d3df8d8b94d5418e57f67229ee00342958772a790951c1c1eec0230.jpg", + "image_caption": [ + "(c) Applied to video tasks" + ], + "image_footnote": [], + "bbox": [ + 529, + 292, + 723, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "interactive tracking", + "bbox": [ + 723, + 267, + 839, + 290 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7124886eb182e23855f7190caf02c1001c53d315a231d0c9af0091c2355a65dc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 727, + 292, + 864, + 343 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "straightforward tracking scenarios, but in cases of severe occlusion scenes, such as OVIS [68], image-level training cannot guarantee that the model exhibits strong temporal consistency under occlusion conditions. Therefore, for occlusion scenarios, it is essential to utilize video data for training. Following IDOL [92], we sample two frames from a video and introduce contrastive learning between frames to make the embedding of the same object instance closer in the embedding space, and the embedding of different object instances farther away. During Inference, the detected objects are tracked by simple bipartite matching of the corresponding object queries following MinVIS [36].", + "bbox": [ + 75, + 462, + 468, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Visual Prompted Segmentation. Interactive segmentation[8, 12, 56, 75, 80, 89, 97] takes various forms of visual prompt, such as points, boxes, or scribbles, to segment the specified objects within an image. On the other hand, VOS [22, 98] aims to segment the entire object throughout the entire video based on a mask provided in the first frame of the video. We extract visual prompt embeddings twice in the model. First, we crop the prompt square area from RGB image and send it to the backbone to obtain the visual prompt feature of the corresponding area, and send it to the early fusion module before the Transformer encoder. Second, we sample fine-grained visual embeddings from the pixel embedding map $M_{p}$ according to visual prompt and make them interact with object queries through self-attention module in the Transformer decoder layer, as the same with text embeddings.", + "bbox": [ + 75, + 659, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Training Unification", + "text_level": 1, + "bbox": [ + 500, + 462, + 691, + 478 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Tasks with Dynamic Loss. We jointly train GLEE in an end-to-end manner on over 5 million images from diverse benchmarks with various levels of supervision. Different loss functions are selected for training on various datasets. There are six types of losses for our GLEE: semantic loss, box loss, mask loss, confidence loss, contrastive tracking loss, and distillation loss. For all tasks with category list or object expressions, we apply Focal loss [53] as semantic loss on logits $S_{\\text{align}}$ to align the text concepts with object features. For box prediction, we use a combination of L1 loss and generalized IoU loss [73]. The mask loss is defined as a combination of the Dice loss [62] and Focal loss. For the Visual Prompt Segmentation tasks, we employ an additional FFN to predict the confidence score for each object queries supervised by Focal loss. For video tasks fine-tuning, we sample two frames and apply contrastive tracking loss on the object query from the last layer of decoder following IDOL [92]. For the text encoder, we distill the knowledge from the frozen teacher CLIP text encoder to ensure the text embedding in pre-trained vison-language embedding space. We apply an L1 loss between our text encoder and CLIP text encoder to minimize their distance:", + "bbox": [ + 496, + 484, + 890, + 816 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t e x t}} = \\frac {1}{K} \\sum_ {i = 0} ^ {K} \\| \\operatorname {E n c} _ {L} \\left(p _ {k}\\right) - \\operatorname {E n c} _ {C L I P} \\left(p _ {k}\\right) \\|, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 837, + 890, + 878 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\{p_k\\}$ is the category name from the category list $P$", + "bbox": [ + 500, + 885, + 883, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3787", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/91a55d0a477988555a9e3875244538343e178e414fb674fbc449ab854a9d49d3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTypeGeneric Detection & SegmentationReferring Detection & SegmentationOpenWorld
COCO-valCOCO-test-devLVISRefCOCORefCOCO+RefCOCOgUVO
APboxAPmaskAPboxAPmaskAPboxAPr-boxAPmaskAPr-maskP@0.5IoUP@0.5IoUP@0.5IoUARmask
MDETR [38]--------87.5-81.1-83.4--
SeqTR [118]--------87.071.778.763.082.764.7-
PolyFormer (L) [55]--------90.476.985.072.285.871.2-
ViTDet-L [50]Specialist Models57.649.8--51.2-46.034.3-------
ViTDet-H [50]58.750.9--53.4-48.136.9-------
EVA-02-L [24]64.255.064.555.865.2-57.3-
ODISE [95]--------------57.7
Mask2Former (L) [15]-50.1-50.5-----------
MaskDINO (L) [45]-54.5-54.7-----------
UniTAB (B) [103]--------88.6-81.0-84.6--
OFA (L) [84]--------90.1-85.8-85.9--
Pix2Seq v2 [14]46.538.2-------------
Uni-Perceiver-v2 (B) [46]58.650.6-------------
Uni-Perceiver-v2 (L) [46]61.953.6-------------
UNINEXT (R50) [100]Generalist Models51.344.9--36.4---89.777.979.866.284.070.0-
UNINEXT (L) [100]58.149.6------91.480.383.170.086.973.4-
UNINEXT (H) [100]60.651.8------92.682.285.272.588.774.7-
GLIPv2 (B) [111]--58.845.8-----------
GLIPv2 (H) [111]--60.648.9-----------
X-Decoder (B) [121]-45.8-45.8-----------
X-Decoder (L) [121]-46.7-47.1-----------
Florence-2 (L) [94]43.4-------93.4-88.3-91.2--
GLEE-LiteFoundation Models55.048.454.748.344.236.740.233.788.577.478.364.882.968.866.6
GLEE-Plus60.453.060.653.352.744.547.440.490.679.581.668.385.070.670.6
GLEE-Pro62.054.262.354.555.749.249.944.391.080.082.669.686.472.972.6
", + "bbox": [ + 80, + 88, + 890, + 371 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison of GLEE to recent specialist and generalist models on object-level image tasks. For REC and RES tasks, we report Precision@0.5 and overall IoU (oIoU). For open-world instance segmentation task, we reported the average recall of 100 mask proposals (AR@100) on the UVO [85].", + "bbox": [ + 75, + 376, + 892, + 419 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Data Scale Up. A visual foundation model should be able to easily scale up the training data and achieve better generalization performance. Thanks to the unified training paradigm, the training data can be scaled up at low cost by introducing a large amount of automatically labeled data from SA1B [39] and GRIT [67]. SA1B provides large and detailed mask annotations, which enhance the general object perception capabilities of model, while GRIT offers a more extensive collection of referring-expression-bounding-box pairs, improving the object identification abilities and the understanding capability of object descriptions. Ultimately, we introduced 2 million SA1B data and 5 million GRIT data into the training process, bringing the total training data to 10 million.", + "bbox": [ + 75, + 434, + 472, + 645 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 666, + 209, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 76, + 693, + 266, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conducted training in three stages. Initially, we performed pretraining for object detection task on Objects365 [79] and OpenImages [42], initializing the text encoder with pretrained CLIP [69] weights and keeping the parameters frozen. In the second training stage, we introduced additional instance segmentation datasets, including COCO [52], LVIS [29], and BDD [106]. Furthermore, we treat three VIS datasets: YTVIS19 [102], YTVIS21 [96], and OVIS [68], as independent image data to enrich the scenes. For datasets that provide descriptions of objects, we included RefCOCO [108], RefCOCO+ [108], RefCOCOg [64], Visual Genome [40], and RVOS [77]. Addi", + "bbox": [ + 73, + 719, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tionally, we introduced two open-world instance segmentation datasets, UVO [85] and a subset of SA1B [39]. Building upon this, we perform the third training stage by introducing more SA1B data and GRIT [67] data to scale up the training set, resulting in a model named GLEE-scale, which exhibited even stronger zero-shot performance on various downstream tasks. During the second and third stages, text encoder is unfrozen but supervised by distillation loss to ensure the predicted text embedding in CLIP embedding space. After the second step, GLEE demonstrated state-of-the-art performance on a range of downstream image and video tasks and exhibited strong zero-shot generalization capabilities, unless otherwise specified, all the experimental results presented below were obtained by the model from this stage. We developed GLEE-Lite, GLEE-Plus, and GLEE-Pro using ResNet-50 [30], SwinLarge [57], and EVA-02 Large [24] as the vision encoder respectively, and train GLEE on 64 A100 GPUs for 500,000 iterations in each stage. More detailed information on data usage, data sampling strategies, and model training can be found in the supplementary materials.", + "bbox": [ + 496, + 434, + 893, + 752 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with Generalist Models", + "text_level": 1, + "bbox": [ + 498, + 768, + 816, + 785 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We demonstrate the universality and effectiveness of our model as an object-level visual foundation model, directly applicable to various object-centric tasks while ensuring SOTA performance without needing fine-tuning. We report detection and instance segmentation results on both the COCO-2017 [52] and LVIS val v1.0 [29]. While sharing almost identical image sets, LVIS is distinguished by", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3788", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/06fe5ecbbd551d128d6ae3ce300b2909766a8b8aabbe295bb1d8f13c302fc0f6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTracking Any Object (TAO [19])BURST [3]LV-VIS [83]
TETALocAAssocAClsAALLCommonUncommonAP\\( AP_b \\)\\( AP_n \\)
HOTAmAPHOTAmAPHOTAmAP
Tracktor [5]24.247.413.012.1---------
DeepSORT [88]26.048.417.512.1---------
Tracktor++ [19]28.049.022.812.1---------
QDTrack [66]30.050.527.412.1---------
TETer [48]33.351.635.013.2---------
OVTrack† [49]34.749.336.718.1---------
STCN Tracker† [3]----5.50.917.50.72.50.6---
Box Tracker† [3]----8.21.427.03.03.60.9---
Detic [117]-SORT† [6]----------12.821.16.6
Detic [117]-XMem †[16]----------16.324.110.6
OV2Seg-R50† [83]-----3.7----14.217.211.9
OV2Seg-B† [83]-----4.9----21.127.516.3
UNINEXT (R50) [100]31.943.335.517.1---------
GLEE-Lite†40.156.339.924.122.612.636.418.919.111.019.622.117.7
GLEE-Plus†41.552.940.930.826.917.238.823.723.915.530.331.629.3
GLEE-Pro†47.266.246.229.131.219.248.724.826.917.723.924.623.3
", + "bbox": [ + 81, + 87, + 890, + 354 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Comparison of GLEE to recent specialist and generalist models on object-level video tasks in a zero-shot manner. Evaluation metrics of BURST are reported separately for 'common', 'uncommon' and 'all' classes. The mAP computes mask IoU at the track level, HOTA is a balance of per-frame detection accuracy (DetA) and temporal association accuracy (AssA), and TETA that deconstructs detection into localization and classification components. The AP, $\\mathrm{AP_b}$ , and $\\mathrm{AP_n}$ in LV-VIS mean the average precision of overall categories, base categories, and novel categories. $\\dagger$ does not use videos for training. The under-performance of Pro relative to Plus on LV-VIS is due to Pro employing larger training and inference resolutions, which prove to be sub-optimal for this specific dataset.", + "bbox": [ + 75, + 364, + 892, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "its annotations of over 1,200 object categories, showcasing a long-tail distribution. This distinction makes LVIS more representative of challenging real-world scenarios due to its broader category coverage. As indicated in Table 1, our model outperforms all generalist models on both COCO and LVIS benchmarks. Even when compared to SOTA specialist approaches, which are tailored with specific designs, our model remains highly competitive. This demonstrates that GLEE, while mastering universal and general object representations, concurrently maintains advanced performance. This characteristic is vitally important for adapting to a broad spectrum of downstream tasks requiring precise object localization. For the REC and RES tasks, we evaluated our model on RefCOCO [108], RefCOCO+ [108], and RefCOCOg [64], as show in Table 1, GLEE achieved comparable results with SOTA specialist methods PolyFormer [55], demonstrating strong capability to comprehend textual descriptions and showcasing potential to adapt to a broader range of multi-modal downstream tasks. In open-world instance segmentation tasks, GLEE outperforms previous arts ODISE [95] by 8.9 points, demonstrating the capability of identifying all plausible instances in an open-world scenario.", + "bbox": [ + 75, + 469, + 472, + 803 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Zero-shot Evaluation Across Tasks", + "text_level": 1, + "bbox": [ + 76, + 816, + 380, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Zero-shot Transfer to Video Tasks. The proposed GLEE is capable of adapting to new data and even new tasks in a zero-shot manner, without the need for additional finetuning. We evaluate its zero-shot capability on three large-", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "scale, large-vocabulary open-world video tracking datasets: TAO [19], BURST [3], and LV-VIS [83]. TAO comprises 2,907 high-resolution videos across 833 categories. BURST builds upon TAO, encompassing 425 base categories and 57 novel categories. LV-VIS offers 4,828 videos within 1,196 well-defined object categories. These three benchmarks require the model to detect, classify, and track all objects in videos, while BURST and LV-VIS additionally require segmentation results from the model. In Table 2, we compare the performance of our proposed model with existing specialist models. Notably, the GLEE here is from the second training stage, which has not been exposed to images from these three datasets nor trained on video-level data. Despite these constraints, GLEE achieves state-of-the-art performance that significantly exceeds existing methodologies. Specifically, GLEE surpasses the previous best method OVTrack by $36.0\\%$ in TAO, nearly triples the performance of the best baseline in BURST, and outperforms OV2Seg [83] by $43.6\\%$ in LV-VIS. This outstanding performance strongly validates the exceptional generalization and zero-shot capabilities of GLEE in handling object-level tasks across a range of benchmarks and tasks. It can be observed that GLEE yields more impressive results on video tasks, since the image tasks have progressed with plentiful data and models from lower costs, while video tasks have not due to higher costs. The models trained on extensive image data with strong general perception capabilities can effectively transfer to video tasks.", + "bbox": [ + 496, + 469, + 890, + 892 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3789", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7972a8ea907b92482910f2a3c2e9b17cc5ed39f6da494aa75ae9323577cbc139.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBackboneYTVIS 2019 val [102]OVIS val [68]
AP\\( AP_{50} \\)\\( AP_{75} \\)AP\\( AP_{50} \\)\\( AP_{75} \\)
SeqFormer [90]47.469.851.815.131.913.8
IDOL [92]49.574.052.930.251.330.0
VITA [33]49.872.654.519.641.217.4
GenVIS [34]ResNet-5051.372.057.834.559.435.0
DVIS [112]52.676.558.234.159.832.3
NOVIS [61]52.875.756.932.756.232.6
UNINEXT [100]53.075.259.134.055.535.6
GLEE-Lite53.174.059.327.1/32.345.4/52.226.3/33.7
SeqFormer [90]59.382.166.4---
VITA [33]63.086.967.927.751.924.9
IDOL [92]64.387.571.042.665.745.2
GenVIS [34]Swin-L63.885.768.545.469.247.8
DVIS [112]64.988.072.749.975.953.0
NOVIS [61]65.787.872.243.568.343.8
GLEE-Plus63.685.270.529.6/40.350.3/63.828.9/39.8
UNINEXT [100]ConvNeXt-L64.387.271.741.165.842.0
UNINEXT [100]ViT-H66.987.575.149.072.552.2
GLEE-ProEVA02-L67.487.174.138.7/50.459.4/71.439.7/55.5
", + "bbox": [ + 89, + 88, + 459, + 303 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/559aae4a2612fbcb495b02f9f6c6ba56a6027344ba4e572bf27105adb36e482d.jpg", + "table_caption": [ + "Table 3. Performance comparison on video instance segmentation tasks. $(\\_ /\\_)$ reports results from zero-shot and after fine-tuning." + ], + "table_footnote": [], + "table_body": "
Images MethodAPAP-categAP-descrAP-descr-posAP-descr-SAP-descr-MAP-descr-L
AllRegionCLIP [114]2.72.72.63.23.62.72.3
Detic [117]8.015.65.48.05.75.46.2
MDETR [38]--4.79.16.44.64.0
GLIP-T [47]19.323.616.425.829.414.88.2
GLIP-L [47]25.832.921.233.237.718.910.8
FIBER-B [23]25.730.322.334.838.619.512.4
GLEE-Lite20.337.514.019.123.012.710.0
GLEE-Lite-Scale22.735.516.722.333.714.310.2
GLEE-Plus25.446.717.523.928.416.312.5
GLEE-Plus-Scale27.044.519.425.936.017.212.4
", + "bbox": [ + 120, + 334, + 423, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Evaluation on the OmniLabel benchmark. The final AP value is the geometric mean of categories (AP-categ) and freeform descriptions (AP-descr).", + "bbox": [ + 76, + 508, + 468, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We additionally provide performance comparison on classic video instance segmentation tasks, whose data is incorporated as image-level data during the second stage of training. As shown in Table 3, on the YTVIS2019 [102] benchmark, our model achieves SOTA results across various model sizes, surpassing all specialist models with complex designs to enhance temporal capabilities and the video unified model UNINEXT [100]. On the OVIS [68] benchmark, which features lengthy videos with extensive object occlusions where temporal capabilities of object features are particularly crucial, our model does not directly reach SOTA. However, after a few hours of simple fine-tuning, it still achieves SOTA performance. More details on VOS, RVOS and demonstrations of interactive segmentation and tracking can be found in supplementary materials.", + "bbox": [ + 75, + 566, + 468, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Zero-shot Transfer to Real-world Downstream Tasks. To measure generalization on real-world tasks, we evaluate zero-shot performance on OmniLabel [76], which is a benchmark for evaluating language-based object detectors and encourages diverse free-form text descriptions of objects. As show in Table 4, compared to language-based detectors trained on large-scale caption data, GLEE signif", + "bbox": [ + 75, + 795, + 467, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c1c1e055e4ea5e7d63b61e43de0e25d2fadccf5f23d8359b4ae7924290d4f83d.jpg", + "image_caption": [ + "Figure 4. The performance comparison of replacing SAM with GLEE in LISA, GLEE achieves the same effectiveness as SAM in extracting objects." + ], + "image_footnote": [], + "bbox": [ + 519, + 88, + 867, + 242 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "icantly outperforms previous models in AP-categ. Due to the limited captions in our training dataset, it scores lower in AP-descr. By incorporating a more diverse set of box-tection data from the GRIT [67] to scale up our training set, the AP-descr can be elevated to a level comparable with existing models. A more comprehensive report on the zero-shot and few-shot performance on ODinW [44] and ablation studies are provided in the supplementary materials.", + "bbox": [ + 496, + 303, + 890, + 425 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Serve as Foundation Model", + "text_level": 1, + "bbox": [ + 500, + 434, + 745, + 449 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To explore whether GLEE can serve as a foundation model for other architectures, we selected LISA [43] for analysis, a mVLLM that combines LLAVA [54] with SAM [39] for reasoning segmentation. We substituted its vision backbone with a frozen, pretrained GLEE-Plus and fed the object queries from GLEE into LLAVA and remove decoder of LISA. We directly dot product the output SEG tokens with GLEE feature map to generate masks. As shown in Figure 4, after training for the same number of steps, our modified LISA-GLEE achieved comparable results to the original version, demonstrating the versatility of representations from GLEE and its effectiveness in serving other models.", + "bbox": [ + 496, + 457, + 890, + 640 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 654, + 617, + 669 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce GLEE, a cutting-edge object-level foundation model designed to be directly applicable to a wide range of object-level image and video tasks. Crafted with a unified learning paradigm, GLEE learns from diverse data sources with varying levels of supervisions. GLEE achieves state-of-the-art performance on numerous object-level tasks and excels in zero-shot transfer to new data and tasks, showing its exceptional versatility and generalization abilities. Additionally, GLEE provides general visual object-level information, which is currently missing in modern LLMs, establishing a robust foundation for object-centric mLLMs.", + "bbox": [ + 496, + 679, + 890, + 845 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 859, + 668, + 876 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This research is supported by NSFC (No.62225603).", + "bbox": [ + 500, + 885, + 846, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3790", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] David Acuna, Huan Ling, Amlan Kar, and Sanja Fidler. Efficient interactive annotation of segmentation datasets with polygon-rnn++. 2018. 2", + "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35: 23716-23736, 2022. 2", + "[3] Ali Athar, Jonathon Luiten, Paul Voigtlaender, Tarasha Khurana, Achal Dave, Bastian Leibe, and Deva Ramanan. Burst: A benchmark for unifying object recognition, segmentation and tracking in video. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1674-1683, 2023. 7", + "[4] Philipp Bergmann, Tim Meinhardt, and Laura Leal-Taixe. Tracking without bells and whistles. In ICCV, 2019. 4", + "[5] Philipp Bergmann, Tim Meinhardt, and Laura Leal-Taixe. Tracking without bells and whistles. In ICCV, 2019. 7", + "[6] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In 2016 IEEE international conference on image processing (ICIP), pages 3464-3468. IEEE, 2016. 7", + "[7] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. 1", + "[8] Yuri Y Boykov and M-P Jolly. Interactive graph cuts for optimal boundary & region segmentation of objects in nd images. In Proceedings eighth IEEE international conference on computer vision. ICCV 2001, pages 105-112. IEEE, 2001. 5", + "[9] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1, 2, 3", + "[10] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 2", + "[11] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2", + "[12] Lluis Castrejon, Kaustav Kundu, Raquel Urtasun, and Sanja Fidler. Annotating object instances with a polygon-rnn. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5230-5238, 2017. 2, 5", + "[13] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pre" + ], + "bbox": [ + 86, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "training from pixels. In International conference on machine learning, pages 1691-1703. PMLR, 2020. 2", + "[14] Ting Chen, Saurabh Saxena, Lala Li, Tsung-Yi Lin, David J Fleet, and Geoffrey E Hinton. A unified sequence interface for vision tasks. Advances in Neural Information Processing Systems, 35:31333-31346, 2022. 3, 6", + "[15] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 2, 4, 6", + "[16] Ho Kei Cheng and Alexander G Schwing. Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In European Conference on Computer Vision, pages 640-658. Springer, 2022. 2, 7", + "[17] Ho Kei Cheng, Yu-Wing Tai, and Chi-Keung Tang. Rethinking space-time networks with improved memory coverage for efficient video object segmentation. Advances in Neural Information Processing Systems, 34:11781-11794, 2021. 2", + "[18] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 2", + "[19] Achal Dave, Tarasha Khurana, Pavel Tokmakov, Cordelia Schmid, and Deva Ramanan. Tao: A large-scale benchmark for tracking any object. In ECCV, 2020. 7", + "[20] Patrick Dendorfer, Aljosa Osep, Anton Milan, Konrad Schindler, Daniel Cremers, Ian Reid, Stefan Roth, and Laura Leal-Taixe. Motchallenge: A benchmark for single-camera multiple target tracking. International Journal of Computer Vision, 129(4):845-881, 2021. 2, 4", + "[21] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 1, 2", + "[22] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, Philip HS Torr, and Song Bai. Mose: A new dataset for video object segmentation in complex scenes. arXiv preprint arXiv:2302.01872, 2023. 5", + "[23] Zi-Yi Dou, Aishwarya Kamath, Zhe Gan, Pengchuan Zhang, Jianfeng Wang, Linjie Li, Zicheng Liu, Ce Liu, Yann LeCun, Nanyun Peng, et al. Coarse-to-fine vision-language pre-training with fusion in the backbone. NeurIPS, 35:32942-32956, 2022. 8", + "[24] Yuxin Fang, Quan Sun, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva-02: A visual representation for neon genesis. arXiv preprint arXiv:2303.11331, 2023.6", + "[25] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19358-19369, 2023. 2", + "[26] Chengjian Feng, Yujie Zhong, Zequn Jie, Xiangxiang Chu, Haibing Ren, Xiaolin Wei, Weidi Xie, and Lin Ma. Promptdet: Towards open-vocabulary detection using uncurated" + ], + "bbox": [ + 509, + 93, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3791", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "images. In Computer Vision - ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IX, page 701-717, 2022. 3", + "[27] Golnaz Ghiasi, Barret Zoph, Ekin D Cubuk, Quoc V Le, and Tsung-Yi Lin. Multi-task self-training for learning general representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8856-8865, 2021. 3", + "[28] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In International Conference on Learning Representations, 2022. 3", + "[29] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 2, 4, 6", + "[30] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 4, 6", + "[31] Kaiming He, Georgia Gkioxari, Piotr Dollar, and Ross Girshick. Mask R-CNN. In ICCV, 2017. 2", + "[32] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 1, 2", + "[33] Miran Heo, Sukjun Hwang, Seoung Wug Oh, Joon-Young Lee, and Seon Joo Kim. Vita: Video instance segmentation via object token association. In Advances in Neural Information Processing Systems, 2022. 8", + "[34] Miran Heo, Sukjun Hwang, Jeongseok Hyun, Hanjung Kim, Seoung Wug Oh, Joon-Young Lee, and Seon Joo Kim. A generalized framework for video instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14623-14632, 2023. 2, 4, 8", + "[35] Ronghang Hu, Marcus Rohrbach, and Trevor Darrell. Segmentation from natural language expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 108-124. Springer, 2016. 2", + "[36] De-An Huang, Zhiding Yu, and Anima Anandkumar. Minvis: A minimal video instance segmentation framework without video-based training. In NeurIPS, 2022. 5", + "[37] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 2, 3", + "[38] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In ICCV, pages 1780-1790, 2021. 6, 8", + "[39] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, 2023. 1, 2, 6, 8" + ], + "bbox": [ + 86, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 2, 6", + "[41] Weicheng Kuo, Yin Cui, Xiuye Gu, AJ Piergiovanni, and Anelia Angelova. Open-vocabulary object detection upon frozen vision and language models. In The Eleventh International Conference on Learning Representations, 2023. 3", + "[42] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, Tom Duerig, and Vittorio Ferrari. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. IJCV, 2020. 2, 6", + "[43] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692, 2023. 2, 8", + "[44] Chunyuan Li, Haotian Liu, Lianian Li, Pengchuan Zhang, Jyoti Aneja, Jianwei Yang, Ping Jin, Houdong Hu, Zicheng Liu, Yong Jae Lee, et al. Elevater: A benchmark and toolkit for evaluating language-augmented visual models. In NeurlPS, 2022. 8", + "[45] Feng Li, Hao Zhang, Huaizhe Xu, Shilong Liu, Lei Zhang, Lionel M Ni, and Heung-Yeung Shum. Mask dino: Towards a unified transformer-based framework for object detection and segmentation. In CVPR, 2023. 3, 4, 6", + "[46] Hao Li, Jinguo Zhu, Xiaohu Jiang, Xizhou Zhu, Hongsheng Li, Chun Yuan, Xiaohua Wang, Yu Qiao, Xiaogang Wang, Wenhai Wang, et al. Uni-perceiver v2: A generalist model for large-scale vision and vision-language tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2691-2700, 2023. 3, 6", + "[47] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In CVPR, pages 10965-10975, 2022. 3, 8", + "[48] Siyuan Li, Martin Danelljan, Henghui Ding, Thomas E Huang, and Fisher Yu. Tracking every thing in the wild. In ECCV, 2022. 7", + "[49] Siyuan Li, Tobias Fischer, Lei Ke, Henghui Ding, Martin Danelljan, and Fisher Yu. Ovtrack: Open-vocabulary multiple object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5567-5577, 2023. 7", + "[50] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In European Conference on Computer Vision, pages 280-296. Springer, 2022. 4, 6", + "[51] Chuang Lin, Peize Sun, Yi Jiang, Ping Luo, Lizhen Qu, Gholamreza Haffari, Zehuan Yuan, and Jianfei Cai. Learning object-language alignments for open-vocabulary object detection. In The Eleventh International Conference on Learning Representations, 2023. 3" + ], + "bbox": [ + 509, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3792", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 2, 4, 6", + "[53] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In ICCV, 2017. 5", + "[54] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 8", + "[55] Jiang Liu, Hui Ding, Zhaowei Cai, Yuting Zhang, Ravi Kumar Satzoda, Vijay Mahadevan, and R Manmatha. Polyformer: Referring image segmentation as sequential polygon generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18653-18663, 2023. 2, 4, 6, 7", + "[56] Qin Liu, Zhenlin Xu, Gedas Bertasius, and Marc Niethammer. Simpleclick: Interactive image segmentation with simple vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22290-22300, 2023. 5", + "[57] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. arXiv preprint arXiv:2103.14030, 2021. 6", + "[58] Jiasen Lu, Christopher Clark, Rowan Zellers, Roozbeh Mottaghi, and Aniruddha Kembhavi. Unified-io: A unified model for vision, language, and multi-modal tasks. arXiv preprint arXiv:2206.08916, 2022.3", + "[59] Chuofan Ma, Yi Jiang, Xin Wen, Zehuan Yuan, and Xiaojuan Qi. Codet: Co-occurrence guided region-word alignment for open-vocabulary object detection. arXiv preprint arXiv:2310.16667, 2023. 3", + "[60] Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. TrackFormer: Multi-object tracking with transformers. arXiv preprint arXiv:2101.02702, 2021. 2, 4", + "[61] Tim Meinhardt, Matt Feiszli, Yuchen Fan, Laura Leal-Taixe, and Rakesh Ranjan. Novis: A case for end-to-end near-online video instance segmentation. arXiv preprint arXiv:2308.15266, 2023. 8", + "[62] Fausto Miletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), 2016. 5", + "[63] Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, Xiao Wang, Xiaohua Zhai, Thomas Kipf, and Neil Houlsby. Simple open-vocabulary object detection. In Computer Vision - ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part X, page 728-755, 2022. 3", + "[64] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In ECCV, 2016. 2, 6, 7", + "[65] Seoung Wug Oh, Joon-Young Lee, Ning Xu, and Seon Joo Kim. Video object segmentation using space-time memory" + ], + "bbox": [ + 86, + 92, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9226-9235, 2019. 2", + "[66] Jiangmiao Pang, Linlu Qiu, Xia Li, Haofeng Chen, Qi Li, Trevor Darrell, and Fisher Yu. Quasi-dense similarity learning for multiple object tracking. In CVPR, 2021. 7", + "[67] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 2, 6, 8", + "[68] Jiyang Qi, Yan Gao, Yao Hu, Xinggang Wang, Xiaoyu Liu, Xiang Bai, Serge Belongie, Alan Yuille, Philip HS Torr, and Song Bai. Occluded video instance segmentation: A benchmark. International Journal of Computer Vision, pages 1-18, 2022. 4, 5, 6, 8", + "[69] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 3, 6", + "[70] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485-5551, 2020. 1, 2, 3", + "[71] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2", + "[72] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2):3, 2022. 2", + "[73] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In CVPR, 2019. 5", + "[74] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2", + "[75] Carsten Rother, Vladimir Kolmogorov, and Andrew Blake. \"grabcut\" interactive foreground extraction using iterated graph cuts. ACM transactions on graphics (TOG), 23(3): 309-314, 2004. 5", + "[76] Samuel Schulter, Vijay Kumar B G, Yumin Suh, Konstantinos M. Dafnis, Zhixing Zhang, Shiyu Zhao, and Dimitris Metaxas. Omnilabel: A challenging benchmark for language-based object detection. In ICCV, 2023. 8", + "[77] Seonguk Seo, Joon-Young Lee, and Bohyung Han. Urvos: Unified referring video object segmentation network with a large-scale benchmark. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XV 16, pages 208–223. Springer, 2020. 2, 4, 6" + ], + "bbox": [ + 509, + 92, + 890, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "3793", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[78] Jing Shao, Siyu Chen, Yangguang Li, Kun Wang, Zhenfei Yin, Yinan He, Jianing Teng, Qinghong Sun, Mengya Gao, Jihao Liu, et al. Intern: A new learning paradigm towards general vision. arXiv preprint arXiv:2111.08687, 2021. 3", + "[79] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019. 2, 4, 6", + "[80] Konstantin Sofiuk, Ilya A Petrov, and Anton Konushin. Reviving iterative training with mask guidance for interactive segmentation. In 2022 IEEE International Conference on Image Processing (ICIP), pages 3141-3145. IEEE, 2022. 5", + "[81] Peize Sun, Rufeng Zhang, Yi Jiang, Tao Kong, Chenfeng Xu, Wei Zhan, Masayoshi Tomizuka, Lei Li, Zehuan Yuan, Changhu Wang, et al. Sparse r-cnn: End-to-end object detection with learnable proposals. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14454-14463, 2021. 2", + "[82] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2", + "[83] Haochen Wang, Cilin Yan, Shuai Wang, Xiaolong Jiang, Xu Tang, Yao Hu, Weidi Xie, and Efstratios Gavves. Towards open-vocabulary video instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4057-4066, 2023. 7", + "[84] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In International Conference on Machine Learning, pages 23318-23340. PMLR, 2022. 3, 6", + "[85] Weiyao Wang, Matt Feiszli, Heng Wang, and Du Tran. Unidentified video objects: A benchmark for dense, open-world segmentation. In ICCV, 2021. 6", + "[86] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, et al. Image as a foreign language: Beit pretraining for all vision and vision-language tasks. arXiv preprint arXiv:2208.10442, 2022. 2", + "[87] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In CVPR, 2021. 2", + "[88] Nicolai Wojke, Alex Bewley, and Dietrich Paulus. Simple online and realtime tracking with a deep association metric. In ICIP, 2017. 7", + "[89] Jiajun Wu, Yibiao Zhao, Jun-Yan Zhu, Siwei Luo, and Zhuowen Tu. Milcut: A sweeping line multiple instance learning paradigm for interactive image segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 256-263, 2014. 5" + ], + "bbox": [ + 86, + 90, + 470, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[90] Junfeng Wu, Yi Jiang, Song Bai, Wenqing Zhang, and Xiang Bai. Seqformer: Sequential transformer for video instance segmentation. In ECCV, 2022. 2, 8", + "[91] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4974-4984, 2022. 2, 4", + "[92] Junfeng Wu, Qihao Liu, Yi Jiang, Song Bai, Alan Yuille, and Xiang Bai. In defense of online models for video instance segmentation. In ECCV, pages 588-605. Springer, 2022. 2, 4, 5, 8", + "[93] Jiannan Wu, Yi Jiang, Bin Yan, Hutchuan Lu, Zehuan Yuan, and Ping Luo. Segment every reference object in spatial and temporal spaces. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2538-2550, 2023. 2", + "[94] Bin Xiao, Haiping Wu, Weijian Xu, Xiyang Dai, Houdong Hu, Yumao Lu, Michael Zeng, Ce Liu, and Lu Yuan. Florence-2: Advancing a unified representation for a variety of vision tasks. arXiv preprint arXiv:2311.06242, 2023. 6", + "[95] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2955-2966, 2023. 6, 7", + "[96] Ning Xu, Linjie Yang, Jianchao Yang, Dingcheng Yue, Yuchen Fan, Yuchen Liang, and Thomas S. Huang. Youtuvis dataset 2021 version. https://youtu-vos.org/dataset/vis/.6", + "[97] Ning Xu, Brian Price, Scott Cohen, Jimei Yang, and Thomas S Huang. Deep interactive object selection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 373-381, 2016. 5", + "[98] Ning Xu, Linjie Yang, Yuchen Fan, Dingcheng Yue, Yuchen Liang, Jianchao Yang, and Thomas Huang. Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327, 2018. 2, 5", + "[99] Bin Yan, Yi Jiang, Peize Sun, Dong Wang, Zehuan Yuan, Ping Luo, and Huchuan Lu. Towards grand unification of object tracking. In ECCV, 2022. 2", + "[100] Bin Yan, Yi Jiang, Jiannan Wu, Dong Wang, Ping Luo, Zehuan Yuan, and Huchuan Lu. Universal instance perception as object discovery and retrieval. In CVPR, 2023. 3, 4, 6, 7, 8", + "[101] Haosen Yang, Chuofan Ma, Bin Wen, Yi Jiang, Zehuan Yuan, and Xiatian Zhu. Recognize any regions. 2023. 3", + "[102] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV, 2019. 2, 4, 6, 8", + "[103] Zhengyuan Yang, Zhe Gan, Jianfeng Wang, Xiaowei Hu, Faisal Ahmed, Zicheng Liu, Yumao Lu, and Lijuan Wang. Unitab: Unifying text and box outputs for grounded vision-language modeling. In European Conference on Computer Vision, pages 521-539. Springer, 2022. 3, 6", + "[104] Lewei Yao, Jianhua Han, Youpeng Wen, Xiaodan Liang, Dan Xu, Wei Zhang, Zhenguo Li, Chunjing Xu, and Hang" + ], + "bbox": [ + 503, + 90, + 893, + 900 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "3794", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Xu. Detclip: Dictionary-enriched visual-concept paralleled pre-training for open-world detection. In NeurIPS, 2022. 4", + "[105] Lewei Yao, Jianhua Han, Xiaodan Liang, Dan Xu, Wei Zhang, Zhenguo Li, and Hang Xu. Detclipv2: Scalable open-vocabulary object detection pre-training via word-region alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23497-23506, 2023. 3", + "[106] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2636-2645, 2020. 6", + "[107] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. Transactions on Machine Learning Research, 2022. 3", + "[108] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In ECCV, 2016. 2, 4, 6, 7", + "[109] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 2, 3", + "[110] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-vocabulary object detection using captions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14393-14402, 2021. 3", + "[111] Haotian Zhang, Pengchuan Zhang, Xiaowei Hu, Yen-Chun Chen, Liunian Li, Xiyang Dai, Lijuan Wang, Lu Yuan, Jenq-Neng Hwang, and Jianfeng Gao. Glipv2: Unifying localization and vision-language understanding. In Advances in Neural Information Processing Systems, pages 36067-36080, 2022. 6", + "[112] Tao Zhang, Xingye Tian, Yu Wu, Shunping Ji, Xuebo Wang, Yuan Zhang, and Pengfei Wan. Dvis: Decoupled video instance segmentation framework. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1282-1291, 2023. 8", + "[113] Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. Byte-track: Multi-object tracking by associating every detection box. arXiv preprint arXiv:2110.06864, 2021. 2, 4", + "[114] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Lianian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In CVPR, pages 16793-16803, 2022. 8", + "[115] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Lianian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16793-16803, 2022. 3" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[116] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Tracking objects as points. In ECCV, 2020. 2, 4", + "[117] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In ECCV, pages 350-368. Springer, 2022. 7, 8", + "[118] Chaoyang Zhu, Yiyi Zhou, Yunhang Shen, Gen Luo, Xingjia Pan, Mingbao Lin, Chao Chen, Liujuan Cao, Xiaoshuai Sun, and Rongrong Ji. Seqtr: A simple yet universal network for visual grounding. In European Conference on Computer Vision, pages 598-615. Springer, 2022. 2, 4, 6", + "[119] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR, 2021. 2, 3", + "[120] Xizhou Zhu, Jinguo Zhu, Hao Li, Xiaoshi Wu, Hongsheng Li, Xiaohua Wang, and Jifeng Dai. Uni-perceiver: Pretraining unified architecture for generic perception for zero-shot and few-shot tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16804-16815, 2022. 3", + "[121] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15116-15127, 2023. 3, 6", + "[122] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. In NeurIPS, 2023. 2, 3" + ], + "bbox": [ + 501, + 92, + 890, + 515 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "3795", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/2024/General Object Foundation Model for Images and Videos at Scale/0bc99999-be81-4577-bff4-eaa67b1ee51f_model.json b/2024/General Object Foundation Model for Images and Videos at Scale/0bc99999-be81-4577-bff4-eaa67b1ee51f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e73972b0ae8f680aa305a82a45b0cc1e1d0dc95c --- /dev/null +++ b/2024/General Object Foundation Model for Images and Videos at Scale/0bc99999-be81-4577-bff4-eaa67b1ee51f_model.json @@ -0,0 +1,2877 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.154, + 0.131, + 0.819, + 0.154 + ], + "angle": 0, + "content": "General Object Foundation Model for Images and Videos at Scale" + }, + { + "type": "text", + "bbox": [ + 0.154, + 0.18, + 0.816, + 0.218 + ], + "angle": 0, + "content": "Junfeng Wu\\(^{1*}\\), Yi Jiang\\(^{2*}\\), Qihao Liu\\(^{3}\\), Zehuan Yuan\\(^{2}\\), Xiang Bai\\(^{1\\dagger}\\), Song Bai\\(^{2\\dagger}\\) \n\\(^{1}\\)Huazhong University of Science and Technology, \\(^{2}\\)ByteDance Inc., \\(^{3}\\)Johns Hopkins University" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.252, + 0.314, + 0.269 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.284, + 0.474, + 0.707 + ], + "angle": 0, + "content": "We present GLEE in this work, an object-level foundation model for locating and identifying objects in images and videos. Through a unified framework, GLEE accomplishes detection, segmentation, tracking, grounding, and identification of arbitrary objects in the open world scenario for various object perception tasks. Adopting a cohesive learning strategy, GLEE acquires knowledge from diverse data sources with varying supervision levels to formulate general object representations, excelling in zero-shot transfer to new data and tasks. Specifically, we employ an image encoder, text encoder, and visual prompter to handle multi-modal inputs, enabling to simultaneously solve various object-centric downstream tasks while maintaining state-of-the-art performance. Demonstrated through extensive training on over five million images from diverse benchmarks, GLEE exhibits remarkable versatility and improved generalization performance, efficiently tackling downstream tasks without the need for task-specific adaptation. By integrating large volumes of automatically labeled data, we further enhance its zero-shot generalization capabilities. Additionally, GLEE is capable of being integrated into Large Language Models, serving as a foundational model to provide universal object-level information for multi-modal tasks. We hope that the versatility and universality of our method will mark a significant step in the development of efficient visual foundation models for AGI systems. The models and code are released at https://github.com/FoundationVision/GLEE." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.736, + 0.21, + 0.753 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.762, + 0.47, + 0.854 + ], + "angle": 0, + "content": "Foundation model [7] is an emerging paradigm for building artificial general intelligence (AGI) systems, signifying a model trained on broad data that is capable of being adapted to a wide range of downstream tasks in an general paradigm. Recently, NLP foundation models such as BERT [21], GPT-3 [9], T5 [70] developed with unified" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.251, + 0.876, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.455, + 0.892, + 0.484 + ], + "angle": 0, + "content": "Figure 1. The performance of GLEE on a broad range of object-level tasks compared with existing models." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.504, + 0.892, + 0.548 + ], + "angle": 0, + "content": "input-output paradigms and large-scale pre-training, have achieved remarkable generalization capabilities to address nearly all NLP tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.551, + 0.893, + 0.763 + ], + "angle": 0, + "content": "In computer vision, the diversity of task types and the lack of a unified form make visual foundation models only serve specific subdomains, such as CLIP [69] for multimodal visual model, MAE [32] for visual representations model, SAM [39] for segmentation model. Despite being widely studied, current visual foundation models are still focusing on establishing correlations between global image features and language descriptions or learning image-level feature representations. However, locating and identifying objects constitute foundational capabilities in computer vision systems, serves as a basis for solving complex or high level vision tasks such as segmentation, scene understanding, object tracking, event detection, and activity recognition and support a wide range of applications." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In this work, we advance the development of object-level foundation models within the visual domain. To address the aforementioned limitation, providing general and accurate object-level information, we introduce a general object visual foundation model, coined as GLEE, which simultaneously solve a wide range of object-centric tasks while ensuring SOTA performance, including object detection, instance segmentation, grounding, object tracking, interactive segmentation and tracking, etc., as shown in Figure 1. Through" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.863, + 0.47, + 0.901 + ], + "angle": 0, + "content": "*Equal Contribution. This work was done when Junfeng Wu and Qihao Liu were interns at ByteDance. †Correspondence to Xiang Bai and Song Bai ." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3783" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.226 + ], + "angle": 0, + "content": "a unified input and output paradigm definition, our model is capable of learning from a wide range of diverse data and predicting general object representations, which makes it generalize well to new data and tasks in a zero-shot manner and achieve amazing performance. In addition, thanks to the unified paradigm, the training data can be scaled up at low cost by introducing a large amount of automatically labeled data, and further improve the zero-shot generalization ability of the model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.231, + 0.471, + 0.472 + ], + "angle": 0, + "content": "A general object foundation model framework. Our objective is to build an object visual foundation model capable of simultaneously addressing a wide range of object-centric tasks. Specifically, we employ an image encoder, a text encoder, and a visual prompter to encode multi-modal inputs. They are integrated into a detector to extract objects from images according to textual and visual input. This unified approach to handle multiple modalities enables us to concurrently solve various object-centric tasks, including detection [10, 52, 81, 119], instance segmentation [15, 31], referring expression comprehension [35, 55, 93, 118], interactive segmentation [1, 12, 122], multi-object tracking [20, 60, 99, 113, 116], video object segmentation [16, 17, 65, 98], video instance segmentation [34, 87, 90, 92, 102], and video referring segmentation [77, 91, 93], all while maintaining state-of-the-art performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.476, + 0.471, + 0.793 + ], + "angle": 0, + "content": "Multi-granularity joint supervision and scalable training paradigm. The design of the unified framework capable of addressing multiple tasks enables joint training on over five million images from diverse benchmarks and varying levels of supervision. Existing datasets differ in annotation granularity: detection datasets like Objects365 [79] and OpenImages [42] offer bounding boxes and category names; COCO [52] and LVIS [29] provide finer-grained mask annotations; RefCOCO [64, 108] and Visual Genome [40] provide detailed object descriptions. Additionally, video data enhance the temporal consistency of model, while open-world data contribute class-agnostic object annotations. An intuitive display of the supervision types and data scales of the datasets employed is presented in Figure 2. The unified support for multi-source data in our approach greatly facilitates the incorporation of additional manually or automatically annotated data, enabling easy scaling up of the dataset. Furthermore, the alignment of model optimization across tasks means that joint training serves not only as a unifying strategy but also as a mechanism to boost performance across individual tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Strong zero-shot transferability to a wide range of object level image and video tasks. After joint training on data from diverse sources, GLEE demonstrates remarkable versatility and zero-shot generalization abilities. Extensive experiments demonstrate that GLEE achieves state-of-the-art performance compared to existing specialist and generalist models in object-level image tasks such as detec" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.365 + ], + "angle": 0, + "content": "tion, referring expression comprehension, and open-world detection, all without requiring any task-specific designs or fine-tuning. Furthermore, we showcase the extraordinary generalization and zero-shot capabilities of GLEE in large-vocabulary open-world video tracking tasks, achieving significantly superior performance over existing models even in a zero-shot transfer manner. Additionally, by incorporating automatically annotated data like SA1B [39] and GRIT [67], we are able to scale up our training dataset to an impressive size of 10 million images at a low cost, which is typically challenging to achieve for object-level tasks and further enhances the generalization performance. Moreover, we replace the SAM [39] component with GLEE in a multimodal Large Language Model (mLLM) [43] and observe that it achieves comparable results. This demonstrates that GLEE is capable of supplying the visual object-level information that modern LLMs currently lack, thus laying a solid foundation for an object-centric mLLMs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.381, + 0.642, + 0.397 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.408, + 0.733, + 0.423 + ], + "angle": 0, + "content": "2.1. Visual Foundation Model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.433, + 0.895, + 0.903 + ], + "angle": 0, + "content": "As foundation models [9, 18, 21, 70, 82] in the NLP field have achieved remarkable success, the construction of visual foundation models attracts increasing attention. Unlike NLP tasks that are predominantly unified under a text-to-text paradigm, tasks in Computer Vision still exhibit significant differences in formulation. This disparity leads to the fact that visual models are typically trained in a single-task learning frameworks, limiting their applicability to tasks within certain sub-domains. For instance, multi-modal visual foundation models like CLIP [69], ALIGN [37], Florence [109], BEIT3 [86], Flamingo[2] make significant advancements in efficient transfer learning and demonstrate impressive zero-shot capabilities on vision-language tasks by employing contrastive learning and masked data modeling on large-scale image-text pairs. DALL-E [71, 72] and Stable Diffusion [74] are trained on massive pairs of images and captions, enabling them to generate detailed image content conditioned on textual instruction. DINO [11], MAE [32], EVA [25], ImageGPT [13] obtain strong visual representations through self-supervised training on largescale image data, which are then employed to adopt downstream tasks. These foundation models learn image-level features and are not directly applicable to object-level tasks. The recently proposed SAM [39], capable of segmenting any object of a given image based on visual prompt such as points and boxes, provides rich object-level information and demonstrates strong generalization capabilities. However, the object information lacks semantic context, limiting its application in object-level tasks. Unlike existing visual foundation models, we aim to develop an object foundation model that directly solve downstream tasks without the need" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3784" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.089, + 0.868, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.459, + 0.893, + 0.487 + ], + "angle": 0, + "content": "Figure 2. An illustrative example showcasing annotations of varying granularities from different datasets, along with the scale of data we utilized. Training on datasets from multiple sources endows the model with more universal representations." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.514, + 0.345, + 0.53 + ], + "angle": 0, + "content": "for additional parameters or fine-tuning." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.544, + 0.323, + 0.559 + ], + "angle": 0, + "content": "2.2. Unified and General Model" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Unified models share similarities with foundation models in the aspect of multi-task unification for their ability to handle multiple vision or multi-modal tasks within a single model. MuST [27] and Intern [78] propose to train across multiple vision tasks and solving them simultaneously. Inspired by the success of sequence-to-sequence NLP models [9, 70], models such as Uni-Perceiver [120], OFA [84], Unified-IO [58], Pix2Seq v2 [14], and UniTAB [103] propose modeling various tasks as sequence generation tasks within a unified paradigm. While these approaches have demonstrated promising cross-task generalization capabilities, they focus mainly on image-level understanding tasks. In addition, their auto-regressive generation of boxes and masks results in significantly slower inference speeds and the performance still falls short of state-of-the-art task-specific models. Building upon on detectors [45, 119], UniPerceiver v2 [46] and UNINEXT [100] utilize unified maximum likelihood estimation and object retrieval to support various tasks, effectively resolves the challenges of localization. Nonetheless, they are trained on closed-set data, thereby not exhibiting zero-shot generalization capabilities. X-decoder [121] and SEEM [122] construct a generalized" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.514, + 0.895, + 0.621 + ], + "angle": 0, + "content": "decoding model capable of predicting pixel-level segmentation and language tokens. Diverging from unified models, the proposed GLEE not only directly addresses object-level tasks in a unified manner but also provides universal object representations, which generalize well to new data and tasks, serving as a cornerstone for a broader range of tasks that require detailed object information." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.634, + 0.787, + 0.651 + ], + "angle": 0, + "content": "2.3. Vision-Language Understanding" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Open-vocabulary detection (OVD) and Grounding models both necessitate the localization and recognition of as many objects as possible. With the recent advancements in vision-language pre-training [37, 69, 107, 109], a commonly employed strategy for OVD involves transferring the knowledge from pre-trained vision-language models (VLMs) to object detectors [28, 41, 63]. Another group of studies leverages extensive image-text pair datasets to broaden the detection vocabulary [26, 47, 51, 59, 101, 105, 110, 115]. However, these language-based detectors are inherently constrained by the capabilities and biases of language models, making it challenging to excel simultaneously in both localization and recognition. Our objective is to optimally utilize existing datasets to construct a general object-level foundation model, aims to not only detect and identify objects effectively but also to offer universal object representations." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3785" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.38, + 0.108 + ], + "angle": 0, + "content": "tations for a wide range of downstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.121, + 0.17, + 0.136 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.146, + 0.212, + 0.16 + ], + "angle": 0, + "content": "3.1. Formulation" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.169, + 0.473, + 0.549 + ], + "angle": 0, + "content": "The proposed GLEE consists of an image encoder, a text encoder, a visual prompter, and an object decoder, as illustrated in Figure 3. The text encoder processes arbitrary descriptions related to the task, including object categories, names in any form, captions about objects, and referring expressions. The visual prompter encodes user inputs such as points, bounding boxes, or scribbles during interactive segmentation into corresponding visual representations of target objects. Then they are integrated into a detector for extracting objects from images according to textual and visual input. We build the object decoder upon MaskDINO [45] with a dynamic class head by compute similarity between object embedding from detector and text embedding from the text encoder. Given an input image \\( I \\in \\mathcal{R}^{3 \\times H \\times W} \\), we first extract multi-scale features \\( Z \\) with backbones such as ResNet [30]. Then we feed them into the object decoder and adopt three prediction heads (classification, detection, and segmentation) on the output embedding \\( q_{d} \\in \\mathcal{R}^{N \\times C} \\) from decoder. Following other object segmentation models [15, 45, 50], we construct a 1/4 resolution pixel embedding map \\( M_{p} \\in \\mathcal{R}^{C \\times \\frac{H}{4} \\times \\frac{W}{4}} \\) which is obtained by upsampling and fusing multi-scale feature maps from the backbone and Transformer encoder. Finally, we obtain each binary mask prediction \\( m \\in \\mathcal{R}^{N \\times \\frac{H}{4} \\times \\frac{W}{4}} \\) via a dot product between the N mask embeddings and pixel embedding map:" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.575, + 0.47, + 0.593 + ], + "angle": 0, + "content": "\\[\nm = F F N \\left(q _ {d}\\right) \\otimes M _ {p}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.602, + 0.47, + 0.633 + ], + "angle": 0, + "content": "where FFN is a 3-layer feed forward head with ReLU activation function and a linear projection layer." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.633, + 0.47, + 0.754 + ], + "angle": 0, + "content": "To support arbitrary vocabularies and object descriptions, we replace the FFN classifier with text embeddings following DetCLIP [104]. Specifically, we feed K category names as separate sentences into the text encoder \\( Enc_{L} \\) and use the average of each sentence tokens as the output text embedding \\( e_t \\in \\mathcal{R}^{K \\times D} \\) for each category or description. Then we compute the alignment scores \\( S_{\\text{align}} \\in \\mathcal{R}^{N \\times K} \\) between object embedding and text embedding:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.767, + 0.47, + 0.784 + ], + "angle": 0, + "content": "\\[\nS _ {\\text {a l i g n}} = q _ {d} \\cdot W _ {i 2 t} \\otimes e _ {t}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.795, + 0.471, + 0.902 + ], + "angle": 0, + "content": "where \\( W_{i2t} \\in \\mathcal{R}^{C \\times D} \\) is image-to-text projection weights. We use logits \\( S_{\\text{align}} \\) to replace traditional classification logits to compute Hungarian matching cost during training and assign categories to objects during inference. To make the original visual features prompt-aware, an early fusion module is adopted before Transformer encoder following UNNEXT [100], which takes image feature from backbone" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "and prompt embedding as input and perform bi-directional cross-attention between them." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.134, + 0.664, + 0.149 + ], + "angle": 0, + "content": "3.2. Task Unification" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.159, + 0.892, + 0.279 + ], + "angle": 0, + "content": "Based on the above designs, GLEE can be used to seamlessly unify a wide range of object perception tasks in images and videos, including object detection, instance segmentation, grounding, multi-target tracking (MOT), video instance segmentation (VIS), video object segmentation (VOS), interactive segmentation and tracking, and supports open-world/large-vocabulary image and video detection and segmentation tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.28, + 0.893, + 0.506 + ], + "angle": 0, + "content": "Detection and Instance Segmentation. For detection task, a fixed-length category list is given and all objects in the category list are required to be detected. For a dataset with category list length \\( K \\), the text input can be formulated as \\( \\{p_k\\}_{k=1}^K \\) where \\( p_k \\) represents for the \\( k \\)-th category name, e.g., \\( P = [\"person\", \"bicycle\", \"car\", ..., \"toothbrush\"] \\) for COCO [52]. For datasets with large vocabulary, calculating the text embedding of all categories is very time-consuming and redundant. Therefore, for datasets with a category number greater than 100, such as objects365 [79] and LVIS [29], suppose there are \\( \\hat{K} \\) positive categories in an image, we take the \\( \\hat{K} \\) positive categories and then pad the category number to 100 by randomly sampling from the negative categories. For instance segmentation, we enable the mask branch and add mask matching cost with mask loss." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.508, + 0.894, + 0.718 + ], + "angle": 0, + "content": "Grounding and Referring Segmentation. These tasks provide reference textual expressions, where objects are described with attributes, for example, Referring Expression Comprehension (REC) [108, 118], Referring Expression Segmentation (RES) [55, 108], and Referring Video Object Segmentation (R-VOS) [77, 91] aim at finding objects matched with the given language expressions like \"The fourth person from the left\". For each image, we take the all the object expressions as text prompt and feed the them into the text encoder. For each expressions, we apply global average pooling along the sequence dimension to get text embedding \\( e_t \\). The text embeddings are feed into early fusion module and additionally interact with object queries through self-attention module in the decoder." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.901 + ], + "angle": 0, + "content": "MOT and VIS. Both Multi-object Tracking (MOT)[4, 20, 60, 113, 116] and Video Instance Segmentation (VIS)[34, 68, 92, 102] need to detect and track all the objects in the predefined category list, and VIS requires additional mask for the objects. These two tasks can be considered as extended tasks of detection and instance segmentation on videos. We found that with sufficient image exposure, object embeddings from the decoder effectively distinguish objects in a video, showing strong discriminability and temporal consistency. As a result, they can be directly employed for tracking without the need for an additional tracking head. Training on image-level data can address" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3786" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.129, + 0.096, + 0.326, + 0.109 + ], + "angle": 0, + "content": "category: [person, car, motorcycle, dog ...]" + }, + { + "type": "text", + "bbox": [ + 0.136, + 0.113, + 0.314, + 0.124 + ], + "angle": 0, + "content": "arbitrary name: bollard, manhole cover" + }, + { + "type": "text", + "bbox": [ + 0.136, + 0.129, + 0.337, + 0.139 + ], + "angle": 0, + "content": "expression: motorcycle parked under the sign" + }, + { + "type": "text", + "bbox": [ + 0.136, + 0.143, + 0.327, + 0.152 + ], + "angle": 0, + "content": "object caption: air conditioner outdoor unit" + }, + { + "type": "image", + "bbox": [ + 0.185, + 0.154, + 0.288, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.137, + 0.212, + 0.19, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.201, + 0.213, + 0.245, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.255, + 0.213, + 0.311, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.346, + 0.098, + 0.818, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.368, + 0.246, + 0.583, + 0.259 + ], + "angle": 0, + "content": "(a) Object foundation model framework" + }, + { + "type": "text", + "bbox": [ + 0.119, + 0.267, + 0.222, + 0.288 + ], + "angle": 0, + "content": "category: person, car motorcycle,dog, cat." + }, + { + "type": "image", + "bbox": [ + 0.109, + 0.29, + 0.175, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.176, + 0.29, + 0.234, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.242, + 0.268, + 0.315, + 0.28 + ], + "angle": 0, + "content": "arbitrary name" + }, + { + "type": "text", + "bbox": [ + 0.243, + 0.281, + 0.28, + 0.292 + ], + "angle": 0, + "content": "bollard" + }, + { + "type": "text", + "bbox": [ + 0.301, + 0.281, + 0.342, + 0.296 + ], + "angle": 0, + "content": "manhole cover" + }, + { + "type": "image", + "bbox": [ + 0.24, + 0.296, + 0.346, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.251, + 0.353, + 0.394, + 0.366 + ], + "angle": 0, + "content": "(b) Applied to image tasks" + }, + { + "type": "text", + "bbox": [ + 0.359, + 0.267, + 0.453, + 0.296 + ], + "angle": 0, + "content": "expression: motorcycle parked under the sign" + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.296, + 0.433, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.459, + 0.267, + 0.512, + 0.296 + ], + "angle": 0, + "content": "interactive segment" + }, + { + "type": "image", + "bbox": [ + 0.444, + 0.296, + 0.521, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.528, + 0.268, + 0.732, + 0.292 + ], + "angle": 0, + "content": "video tasks: VIS、MOT、VOS、RVOS、 open world/vocabulary tracking" + }, + { + "type": "image", + "bbox": [ + 0.53, + 0.294, + 0.724, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.724, + 0.268, + 0.84, + 0.291 + ], + "angle": 0, + "content": "interactive tracking" + }, + { + "type": "image", + "bbox": [ + 0.728, + 0.294, + 0.865, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.634, + 0.353, + 0.774, + 0.366 + ], + "angle": 0, + "content": "(c) Applied to video tasks" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.378, + 0.892, + 0.448 + ], + "angle": 0, + "content": "Figure 3. Framework of GLEE. The text encoder accepts textual descriptions in various forms from diverse data sources, including object categories, names, captions, and referring expressions. The visual prompt encodes points, bounding boxes, or scribbles into corresponding visual representations. The object decoder takes them and image features to predict objects in images. (b) illustrates the application of GLEE to image tasks tailored for different language descriptions and visual prompts. (c) demonstrates the application across various object-level video tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.463, + 0.47, + 0.645 + ], + "angle": 0, + "content": "straightforward tracking scenarios, but in cases of severe occlusion scenes, such as OVIS [68], image-level training cannot guarantee that the model exhibits strong temporal consistency under occlusion conditions. Therefore, for occlusion scenarios, it is essential to utilize video data for training. Following IDOL [92], we sample two frames from a video and introduce contrastive learning between frames to make the embedding of the same object instance closer in the embedding space, and the embedding of different object instances farther away. During Inference, the detected objects are tracked by simple bipartite matching of the corresponding object queries following MinVIS [36]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Visual Prompted Segmentation. Interactive segmentation[8, 12, 56, 75, 80, 89, 97] takes various forms of visual prompt, such as points, boxes, or scribbles, to segment the specified objects within an image. On the other hand, VOS [22, 98] aims to segment the entire object throughout the entire video based on a mask provided in the first frame of the video. We extract visual prompt embeddings twice in the model. First, we crop the prompt square area from RGB image and send it to the backbone to obtain the visual prompt feature of the corresponding area, and send it to the early fusion module before the Transformer encoder. Second, we sample fine-grained visual embeddings from the pixel embedding map \\( M_{p} \\) according to visual prompt and make them interact with object queries through self-attention module in the Transformer decoder layer, as the same with text embeddings." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.463, + 0.692, + 0.479 + ], + "angle": 0, + "content": "3.3. Training Unification" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.486, + 0.892, + 0.818 + ], + "angle": 0, + "content": "Tasks with Dynamic Loss. We jointly train GLEE in an end-to-end manner on over 5 million images from diverse benchmarks with various levels of supervision. Different loss functions are selected for training on various datasets. There are six types of losses for our GLEE: semantic loss, box loss, mask loss, confidence loss, contrastive tracking loss, and distillation loss. For all tasks with category list or object expressions, we apply Focal loss [53] as semantic loss on logits \\( S_{\\text{align}} \\) to align the text concepts with object features. For box prediction, we use a combination of L1 loss and generalized IoU loss [73]. The mask loss is defined as a combination of the Dice loss [62] and Focal loss. For the Visual Prompt Segmentation tasks, we employ an additional FFN to predict the confidence score for each object queries supervised by Focal loss. For video tasks fine-tuning, we sample two frames and apply contrastive tracking loss on the object query from the last layer of decoder following IDOL [92]. For the text encoder, we distill the knowledge from the frozen teacher CLIP text encoder to ensure the text embedding in pre-trained vison-language embedding space. We apply an L1 loss between our text encoder and CLIP text encoder to minimize their distance:" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.838, + 0.892, + 0.879 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t e x t}} = \\frac {1}{K} \\sum_ {i = 0} ^ {K} \\| \\operatorname {E n c} _ {L} \\left(p _ {k}\\right) - \\operatorname {E n c} _ {C L I P} \\left(p _ {k}\\right) \\|, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.884, + 0.902 + ], + "angle": 0, + "content": "where \\(\\{p_k\\}\\) is the category name from the category list \\(P\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3787" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.089, + 0.891, + 0.372 + ], + "angle": 0, + "content": "
MethodTypeGeneric Detection & SegmentationReferring Detection & SegmentationOpenWorld
COCO-valCOCO-test-devLVISRefCOCORefCOCO+RefCOCOgUVO
APboxAPmaskAPboxAPmaskAPboxAPr-boxAPmaskAPr-maskP@0.5IoUP@0.5IoUP@0.5IoUARmask
MDETR [38]--------87.5-81.1-83.4--
SeqTR [118]--------87.071.778.763.082.764.7-
PolyFormer (L) [55]--------90.476.985.072.285.871.2-
ViTDet-L [50]Specialist Models57.649.8--51.2-46.034.3-------
ViTDet-H [50]58.750.9--53.4-48.136.9-------
EVA-02-L [24]64.255.064.555.865.2-57.3-
ODISE [95]--------------57.7
Mask2Former (L) [15]-50.1-50.5-----------
MaskDINO (L) [45]-54.5-54.7-----------
UniTAB (B) [103]--------88.6-81.0-84.6--
OFA (L) [84]--------90.1-85.8-85.9--
Pix2Seq v2 [14]46.538.2-------------
Uni-Perceiver-v2 (B) [46]58.650.6-------------
Uni-Perceiver-v2 (L) [46]61.953.6-------------
UNINEXT (R50) [100]Generalist Models51.344.9--36.4---89.777.979.866.284.070.0-
UNINEXT (L) [100]58.149.6------91.480.383.170.086.973.4-
UNINEXT (H) [100]60.651.8------92.682.285.272.588.774.7-
GLIPv2 (B) [111]--58.845.8-----------
GLIPv2 (H) [111]--60.648.9-----------
X-Decoder (B) [121]-45.8-45.8-----------
X-Decoder (L) [121]-46.7-47.1-----------
Florence-2 (L) [94]43.4-------93.4-88.3-91.2--
GLEE-LiteFoundation Models55.048.454.748.344.236.740.233.788.577.478.364.882.968.866.6
GLEE-Plus60.453.060.653.352.744.547.440.490.679.581.668.385.070.670.6
GLEE-Pro62.054.262.354.555.749.249.944.391.080.082.669.686.472.972.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.377, + 0.893, + 0.42 + ], + "angle": 0, + "content": "Table 1. Comparison of GLEE to recent specialist and generalist models on object-level image tasks. For REC and RES tasks, we report Precision@0.5 and overall IoU (oIoU). For open-world instance segmentation task, we reported the average recall of 100 mask proposals (AR@100) on the UVO [85]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.435, + 0.473, + 0.646 + ], + "angle": 0, + "content": "Data Scale Up. A visual foundation model should be able to easily scale up the training data and achieve better generalization performance. Thanks to the unified training paradigm, the training data can be scaled up at low cost by introducing a large amount of automatically labeled data from SA1B [39] and GRIT [67]. SA1B provides large and detailed mask annotations, which enhance the general object perception capabilities of model, while GRIT offers a more extensive collection of referring-expression-bounding-box pairs, improving the object identification abilities and the understanding capability of object descriptions. Ultimately, we introduced 2 million SA1B data and 5 million GRIT data into the training process, bringing the total training data to 10 million." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.667, + 0.21, + 0.684 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.694, + 0.267, + 0.71 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.72, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We conducted training in three stages. Initially, we performed pretraining for object detection task on Objects365 [79] and OpenImages [42], initializing the text encoder with pretrained CLIP [69] weights and keeping the parameters frozen. In the second training stage, we introduced additional instance segmentation datasets, including COCO [52], LVIS [29], and BDD [106]. Furthermore, we treat three VIS datasets: YTVIS19 [102], YTVIS21 [96], and OVIS [68], as independent image data to enrich the scenes. For datasets that provide descriptions of objects, we included RefCOCO [108], RefCOCO+ [108], RefCOCOg [64], Visual Genome [40], and RVOS [77]. Addi" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.435, + 0.895, + 0.753 + ], + "angle": 0, + "content": "tionally, we introduced two open-world instance segmentation datasets, UVO [85] and a subset of SA1B [39]. Building upon this, we perform the third training stage by introducing more SA1B data and GRIT [67] data to scale up the training set, resulting in a model named GLEE-scale, which exhibited even stronger zero-shot performance on various downstream tasks. During the second and third stages, text encoder is unfrozen but supervised by distillation loss to ensure the predicted text embedding in CLIP embedding space. After the second step, GLEE demonstrated state-of-the-art performance on a range of downstream image and video tasks and exhibited strong zero-shot generalization capabilities, unless otherwise specified, all the experimental results presented below were obtained by the model from this stage. We developed GLEE-Lite, GLEE-Plus, and GLEE-Pro using ResNet-50 [30], SwinLarge [57], and EVA-02 Large [24] as the vision encoder respectively, and train GLEE on 64 A100 GPUs for 500,000 iterations in each stage. More detailed information on data usage, data sampling strategies, and model training can be found in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.769, + 0.817, + 0.786 + ], + "angle": 0, + "content": "4.2. Comparison with Generalist Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We demonstrate the universality and effectiveness of our model as an object-level visual foundation model, directly applicable to various object-centric tasks while ensuring SOTA performance without needing fine-tuning. We report detection and instance segmentation results on both the COCO-2017 [52] and LVIS val v1.0 [29]. While sharing almost identical image sets, LVIS is distinguished by" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3788" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.088, + 0.891, + 0.356 + ], + "angle": 0, + "content": "
MethodTracking Any Object (TAO [19])BURST [3]LV-VIS [83]
TETALocAAssocAClsAALLCommonUncommonAP\\( AP_b \\)\\( AP_n \\)
HOTAmAPHOTAmAPHOTAmAP
Tracktor [5]24.247.413.012.1---------
DeepSORT [88]26.048.417.512.1---------
Tracktor++ [19]28.049.022.812.1---------
QDTrack [66]30.050.527.412.1---------
TETer [48]33.351.635.013.2---------
OVTrack† [49]34.749.336.718.1---------
STCN Tracker† [3]----5.50.917.50.72.50.6---
Box Tracker† [3]----8.21.427.03.03.60.9---
Detic [117]-SORT† [6]----------12.821.16.6
Detic [117]-XMem †[16]----------16.324.110.6
OV2Seg-R50† [83]-----3.7----14.217.211.9
OV2Seg-B† [83]-----4.9----21.127.516.3
UNINEXT (R50) [100]31.943.335.517.1---------
GLEE-Lite†40.156.339.924.122.612.636.418.919.111.019.622.117.7
GLEE-Plus†41.552.940.930.826.917.238.823.723.915.530.331.629.3
GLEE-Pro†47.266.246.229.131.219.248.724.826.917.723.924.623.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.366, + 0.893, + 0.451 + ], + "angle": 0, + "content": "Table 2. Comparison of GLEE to recent specialist and generalist models on object-level video tasks in a zero-shot manner. Evaluation metrics of BURST are reported separately for 'common', 'uncommon' and 'all' classes. The mAP computes mask IoU at the track level, HOTA is a balance of per-frame detection accuracy (DetA) and temporal association accuracy (AssA), and TETA that deconstructs detection into localization and classification components. The AP, \\(\\mathrm{AP_b}\\), and \\(\\mathrm{AP_n}\\) in LV-VIS mean the average precision of overall categories, base categories, and novel categories. \\(\\dagger\\) does not use videos for training. The under-performance of Pro relative to Plus on LV-VIS is due to Pro employing larger training and inference resolutions, which prove to be sub-optimal for this specific dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.47, + 0.473, + 0.804 + ], + "angle": 0, + "content": "its annotations of over 1,200 object categories, showcasing a long-tail distribution. This distinction makes LVIS more representative of challenging real-world scenarios due to its broader category coverage. As indicated in Table 1, our model outperforms all generalist models on both COCO and LVIS benchmarks. Even when compared to SOTA specialist approaches, which are tailored with specific designs, our model remains highly competitive. This demonstrates that GLEE, while mastering universal and general object representations, concurrently maintains advanced performance. This characteristic is vitally important for adapting to a broad spectrum of downstream tasks requiring precise object localization. For the REC and RES tasks, we evaluated our model on RefCOCO [108], RefCOCO+ [108], and RefCOCOg [64], as show in Table 1, GLEE achieved comparable results with SOTA specialist methods PolyFormer [55], demonstrating strong capability to comprehend textual descriptions and showcasing potential to adapt to a broader range of multi-modal downstream tasks. In open-world instance segmentation tasks, GLEE outperforms previous arts ODISE [95] by 8.9 points, demonstrating the capability of identifying all plausible instances in an open-world scenario." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.382, + 0.832 + ], + "angle": 0, + "content": "4.3. Zero-shot Evaluation Across Tasks" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Zero-shot Transfer to Video Tasks. The proposed GLEE is capable of adapting to new data and even new tasks in a zero-shot manner, without the need for additional finetuning. We evaluate its zero-shot capability on three large-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.47, + 0.892, + 0.893 + ], + "angle": 0, + "content": "scale, large-vocabulary open-world video tracking datasets: TAO [19], BURST [3], and LV-VIS [83]. TAO comprises 2,907 high-resolution videos across 833 categories. BURST builds upon TAO, encompassing 425 base categories and 57 novel categories. LV-VIS offers 4,828 videos within 1,196 well-defined object categories. These three benchmarks require the model to detect, classify, and track all objects in videos, while BURST and LV-VIS additionally require segmentation results from the model. In Table 2, we compare the performance of our proposed model with existing specialist models. Notably, the GLEE here is from the second training stage, which has not been exposed to images from these three datasets nor trained on video-level data. Despite these constraints, GLEE achieves state-of-the-art performance that significantly exceeds existing methodologies. Specifically, GLEE surpasses the previous best method OVTrack by \\(36.0\\%\\) in TAO, nearly triples the performance of the best baseline in BURST, and outperforms OV2Seg [83] by \\(43.6\\%\\) in LV-VIS. This outstanding performance strongly validates the exceptional generalization and zero-shot capabilities of GLEE in handling object-level tasks across a range of benchmarks and tasks. It can be observed that GLEE yields more impressive results on video tasks, since the image tasks have progressed with plentiful data and models from lower costs, while video tasks have not due to higher costs. The models trained on extensive image data with strong general perception capabilities can effectively transfer to video tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3789" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.091, + 0.089, + 0.46, + 0.304 + ], + "angle": 0, + "content": "
MethodBackboneYTVIS 2019 val [102]OVIS val [68]
AP\\( AP_{50} \\)\\( AP_{75} \\)AP\\( AP_{50} \\)\\( AP_{75} \\)
SeqFormer [90]47.469.851.815.131.913.8
IDOL [92]49.574.052.930.251.330.0
VITA [33]49.872.654.519.641.217.4
GenVIS [34]ResNet-5051.372.057.834.559.435.0
DVIS [112]52.676.558.234.159.832.3
NOVIS [61]52.875.756.932.756.232.6
UNINEXT [100]53.075.259.134.055.535.6
GLEE-Lite53.174.059.327.1/32.345.4/52.226.3/33.7
SeqFormer [90]59.382.166.4---
VITA [33]63.086.967.927.751.924.9
IDOL [92]64.387.571.042.665.745.2
GenVIS [34]Swin-L63.885.768.545.469.247.8
DVIS [112]64.988.072.749.975.953.0
NOVIS [61]65.787.872.243.568.343.8
GLEE-Plus63.685.270.529.6/40.350.3/63.828.9/39.8
UNINEXT [100]ConvNeXt-L64.387.271.741.165.842.0
UNINEXT [100]ViT-H66.987.575.149.072.552.2
GLEE-ProEVA02-L67.487.174.138.7/50.459.4/71.439.7/55.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.304, + 0.468, + 0.333 + ], + "angle": 0, + "content": "Table 3. Performance comparison on video instance segmentation tasks. \\((\\_ /\\_)\\) reports results from zero-shot and after fine-tuning." + }, + { + "type": "table", + "bbox": [ + 0.121, + 0.335, + 0.424, + 0.509 + ], + "angle": 0, + "content": "
Images MethodAPAP-categAP-descrAP-descr-posAP-descr-SAP-descr-MAP-descr-L
AllRegionCLIP [114]2.72.72.63.23.62.72.3
Detic [117]8.015.65.48.05.75.46.2
MDETR [38]--4.79.16.44.64.0
GLIP-T [47]19.323.616.425.829.414.88.2
GLIP-L [47]25.832.921.233.237.718.910.8
FIBER-B [23]25.730.322.334.838.619.512.4
GLEE-Lite20.337.514.019.123.012.710.0
GLEE-Lite-Scale22.735.516.722.333.714.310.2
GLEE-Plus25.446.717.523.928.416.312.5
GLEE-Plus-Scale27.044.519.425.936.017.212.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.51, + 0.47, + 0.553 + ], + "angle": 0, + "content": "Table 4. Evaluation on the OmniLabel benchmark. The final AP value is the geometric mean of categories (AP-categ) and freeform descriptions (AP-descr)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.567, + 0.47, + 0.793 + ], + "angle": 0, + "content": "We additionally provide performance comparison on classic video instance segmentation tasks, whose data is incorporated as image-level data during the second stage of training. As shown in Table 3, on the YTVIS2019 [102] benchmark, our model achieves SOTA results across various model sizes, surpassing all specialist models with complex designs to enhance temporal capabilities and the video unified model UNINEXT [100]. On the OVIS [68] benchmark, which features lengthy videos with extensive object occlusions where temporal capabilities of object features are particularly crucial, our model does not directly reach SOTA. However, after a few hours of simple fine-tuning, it still achieves SOTA performance. More details on VOS, RVOS and demonstrations of interactive segmentation and tracking can be found in supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.468, + 0.902 + ], + "angle": 0, + "content": "Zero-shot Transfer to Real-world Downstream Tasks. To measure generalization on real-world tasks, we evaluate zero-shot performance on OmniLabel [76], which is a benchmark for evaluating language-based object detectors and encourages diverse free-form text descriptions of objects. As show in Table 4, compared to language-based detectors trained on large-scale caption data, GLEE signif" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.089, + 0.868, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.252, + 0.892, + 0.294 + ], + "angle": 0, + "content": "Figure 4. The performance comparison of replacing SAM with GLEE in LISA, GLEE achieves the same effectiveness as SAM in extracting objects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.304, + 0.892, + 0.426 + ], + "angle": 0, + "content": "icantly outperforms previous models in AP-categ. Due to the limited captions in our training dataset, it scores lower in AP-descr. By incorporating a more diverse set of box-tection data from the GRIT [67] to scale up our training set, the AP-descr can be elevated to a level comparable with existing models. A more comprehensive report on the zero-shot and few-shot performance on ODinW [44] and ablation studies are provided in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.435, + 0.746, + 0.45 + ], + "angle": 0, + "content": "4.4. Serve as Foundation Model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.458, + 0.892, + 0.641 + ], + "angle": 0, + "content": "To explore whether GLEE can serve as a foundation model for other architectures, we selected LISA [43] for analysis, a mVLLM that combines LLAVA [54] with SAM [39] for reasoning segmentation. We substituted its vision backbone with a frozen, pretrained GLEE-Plus and fed the object queries from GLEE into LLAVA and remove decoder of LISA. We directly dot product the output SEG tokens with GLEE feature map to generate masks. As shown in Figure 4, after training for the same number of steps, our modified LISA-GLEE achieved comparable results to the original version, demonstrating the versatility of representations from GLEE and its effectiveness in serving other models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.655, + 0.619, + 0.67 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.68, + 0.892, + 0.846 + ], + "angle": 0, + "content": "We introduce GLEE, a cutting-edge object-level foundation model designed to be directly applicable to a wide range of object-level image and video tasks. Crafted with a unified learning paradigm, GLEE learns from diverse data sources with varying levels of supervisions. GLEE achieves state-of-the-art performance on numerous object-level tasks and excels in zero-shot transfer to new data and tasks, showing its exceptional versatility and generalization abilities. Additionally, GLEE provides general visual object-level information, which is currently missing in modern LLMs, establishing a robust foundation for object-centric mLLMs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.861, + 0.669, + 0.877 + ], + "angle": 0, + "content": "6. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.848, + 0.901 + ], + "angle": 0, + "content": "This research is supported by NSFC (No.62225603)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3790" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.115, + 0.469, + 0.156 + ], + "angle": 0, + "content": "[1] David Acuna, Huan Ling, Amlan Kar, and Sanja Fidler. Efficient interactive annotation of segmentation datasets with polygon-rnn++. 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.095, + 0.159, + 0.469, + 0.24 + ], + "angle": 0, + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35: 23716-23736, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.244, + 0.469, + 0.327 + ], + "angle": 0, + "content": "[3] Ali Athar, Jonathon Luiten, Paul Voigtlaender, Tarasha Khurana, Achal Dave, Bastian Leibe, and Deva Ramanan. Burst: A benchmark for unifying object recognition, segmentation and tracking in video. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1674-1683, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.329, + 0.468, + 0.357 + ], + "angle": 0, + "content": "[4] Philipp Bergmann, Tim Meinhardt, and Laura Leal-Taixe. Tracking without bells and whistles. In ICCV, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.359, + 0.468, + 0.386 + ], + "angle": 0, + "content": "[5] Philipp Bergmann, Tim Meinhardt, and Laura Leal-Taixe. Tracking without bells and whistles. In ICCV, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.389, + 0.469, + 0.444 + ], + "angle": 0, + "content": "[6] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In 2016 IEEE international conference on image processing (ICIP), pages 3464-3468. IEEE, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.446, + 0.469, + 0.514 + ], + "angle": 0, + "content": "[7] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.517, + 0.469, + 0.585 + ], + "angle": 0, + "content": "[8] Yuri Y Boykov and M-P Jolly. Interactive graph cuts for optimal boundary & region segmentation of objects in nd images. In Proceedings eighth IEEE international conference on computer vision. ICCV 2001, pages 105-112. IEEE, 2001. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.589, + 0.469, + 0.67 + ], + "angle": 0, + "content": "[9] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.673, + 0.469, + 0.728 + ], + "angle": 0, + "content": "[10] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.731, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[11] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.802, + 0.469, + 0.869 + ], + "angle": 0, + "content": "[12] Lluis Castrejon, Kaustav Kundu, Raquel Urtasun, and Sanja Fidler. Annotating object instances with a polygon-rnn. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5230-5238, 2017. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[13] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pre" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.115, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.094, + 0.892, + 0.12 + ], + "angle": 0, + "content": "training from pixels. In International conference on machine learning, pages 1691-1703. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.122, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[14] Ting Chen, Saurabh Saxena, Lala Li, Tsung-Yi Lin, David J Fleet, and Geoffrey E Hinton. A unified sequence interface for vision tasks. Advances in Neural Information Processing Systems, 35:31333-31346, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.178, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[15] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.234, + 0.892, + 0.288 + ], + "angle": 0, + "content": "[16] Ho Kei Cheng and Alexander G Schwing. Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In European Conference on Computer Vision, pages 640-658. Springer, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.289, + 0.892, + 0.355 + ], + "angle": 0, + "content": "[17] Ho Kei Cheng, Yu-Wing Tai, and Chi-Keung Tang. Rethinking space-time networks with improved memory coverage for efficient video object segmentation. Advances in Neural Information Processing Systems, 34:11781-11794, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.358, + 0.892, + 0.428 + ], + "angle": 0, + "content": "[18] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.429, + 0.892, + 0.469 + ], + "angle": 0, + "content": "[19] Achal Dave, Tarasha Khurana, Pavel Tokmakov, Cordelia Schmid, and Deva Ramanan. Tao: A large-scale benchmark for tracking any object. In ECCV, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.47, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[20] Patrick Dendorfer, Aljosa Osep, Anton Milan, Konrad Schindler, Daniel Cremers, Ian Reid, Stefan Roth, and Laura Leal-Taixe. Motchallenge: A benchmark for single-camera multiple target tracking. International Journal of Computer Vision, 129(4):845-881, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.539, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[21] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.596, + 0.892, + 0.65 + ], + "angle": 0, + "content": "[22] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, Philip HS Torr, and Song Bai. Mose: A new dataset for video object segmentation in complex scenes. arXiv preprint arXiv:2302.01872, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.651, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[23] Zi-Yi Dou, Aishwarya Kamath, Zhe Gan, Pengchuan Zhang, Jianfeng Wang, Linjie Li, Zicheng Liu, Ce Liu, Yann LeCun, Nanyun Peng, et al. Coarse-to-fine vision-language pre-training with fusion in the backbone. NeurIPS, 35:32942-32956, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.72, + 0.892, + 0.774 + ], + "angle": 0, + "content": "[24] Yuxin Fang, Quan Sun, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva-02: A visual representation for neon genesis. arXiv preprint arXiv:2303.11331, 2023.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.776, + 0.892, + 0.859 + ], + "angle": 0, + "content": "[25] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19358-19369, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[26] Chengjian Feng, Yujie Zhong, Zequn Jie, Xiangxiang Chu, Haibing Ren, Xiaolin Wei, Weidi Xie, and Lin Ma. Promptdet: Towards open-vocabulary detection using uncurated" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.094, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3791" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.092, + 0.468, + 0.134 + ], + "angle": 0, + "content": "images. In Computer Vision - ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IX, page 701-717, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.135, + 0.47, + 0.203 + ], + "angle": 0, + "content": "[27] Golnaz Ghiasi, Barret Zoph, Ekin D Cubuk, Quoc V Le, and Tsung-Yi Lin. Multi-task self-training for learning general representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8856-8865, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.205, + 0.469, + 0.259 + ], + "angle": 0, + "content": "[28] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In International Conference on Learning Representations, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.261, + 0.469, + 0.315 + ], + "angle": 0, + "content": "[29] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.317, + 0.468, + 0.356 + ], + "angle": 0, + "content": "[30] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.359, + 0.468, + 0.384 + ], + "angle": 0, + "content": "[31] Kaiming He, Georgia Gkioxari, Piotr Dollar, and Ross Girshick. Mask R-CNN. In ICCV, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.387, + 0.469, + 0.453 + ], + "angle": 0, + "content": "[32] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.456, + 0.469, + 0.51 + ], + "angle": 0, + "content": "[33] Miran Heo, Sukjun Hwang, Seoung Wug Oh, Joon-Young Lee, and Seon Joo Kim. Vita: Video instance segmentation via object token association. In Advances in Neural Information Processing Systems, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.512, + 0.469, + 0.593 + ], + "angle": 0, + "content": "[34] Miran Heo, Sukjun Hwang, Jeongseok Hyun, Hanjung Kim, Seoung Wug Oh, Joon-Young Lee, and Seon Joo Kim. A generalized framework for video instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14623-14632, 2023. 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.596, + 0.469, + 0.663 + ], + "angle": 0, + "content": "[35] Ronghang Hu, Marcus Rohrbach, and Trevor Darrell. Segmentation from natural language expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 108-124. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.665, + 0.468, + 0.705 + ], + "angle": 0, + "content": "[36] De-An Huang, Zhiding Yu, and Anima Anandkumar. Minvis: A minimal video instance segmentation framework without video-based training. In NeurIPS, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.707, + 0.469, + 0.788 + ], + "angle": 0, + "content": "[37] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.79, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[38] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In ICCV, pages 1780-1790, 2021. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[39] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, 2023. 1, 2, 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.175 + ], + "angle": 0, + "content": "[40] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.178, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[41] Weicheng Kuo, Yin Cui, Xiuye Gu, AJ Piergiovanni, and Anelia Angelova. Open-vocabulary object detection upon frozen vision and language models. In The Eleventh International Conference on Learning Representations, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.235, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[42] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, Tom Duerig, and Vittorio Ferrari. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. IJCV, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.32, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[43] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692, 2023. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.377, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[44] Chunyuan Li, Haotian Liu, Lianian Li, Pengchuan Zhang, Jyoti Aneja, Jianwei Yang, Ping Jin, Houdong Hu, Zicheng Liu, Yong Jae Lee, et al. Elevater: A benchmark and toolkit for evaluating language-augmented visual models. In NeurlPS, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.448, + 0.892, + 0.502 + ], + "angle": 0, + "content": "[45] Feng Li, Hao Zhang, Huaizhe Xu, Shilong Liu, Lei Zhang, Lionel M Ni, and Heung-Yeung Shum. Mask dino: Towards a unified transformer-based framework for object detection and segmentation. In CVPR, 2023. 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.505, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[46] Hao Li, Jinguo Zhu, Xiaohu Jiang, Xizhou Zhu, Hongsheng Li, Chun Yuan, Xiaohua Wang, Yu Qiao, Xiaogang Wang, Wenhai Wang, et al. Uni-perceiver v2: A generalist model for large-scale vision and vision-language tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2691-2700, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.59, + 0.892, + 0.657 + ], + "angle": 0, + "content": "[47] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In CVPR, pages 10965-10975, 2022. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.661, + 0.892, + 0.7 + ], + "angle": 0, + "content": "[48] Siyuan Li, Martin Danelljan, Henghui Ding, Thomas E Huang, and Fisher Yu. Tracking every thing in the wild. In ECCV, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.704, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[49] Siyuan Li, Tobias Fischer, Lei Ke, Henghui Ding, Martin Danelljan, and Fisher Yu. Ovtrack: Open-vocabulary multiple object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5567-5577, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.775, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[50] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In European Conference on Computer Vision, pages 280-296. Springer, 2022. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.832, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[51] Chuang Lin, Peize Sun, Yi Jiang, Ping Luo, Lizhen Qu, Gholamreza Haffari, Zehuan Yuan, and Jianfei Cai. Learning object-language alignments for open-vocabulary object detection. In The Eleventh International Conference on Learning Representations, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3792" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.093, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[52] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.149, + 0.47, + 0.189 + ], + "angle": 0, + "content": "[53] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In ICCV, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.191, + 0.471, + 0.232 + ], + "angle": 0, + "content": "[54] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.233, + 0.471, + 0.315 + ], + "angle": 0, + "content": "[55] Jiang Liu, Hui Ding, Zhaowei Cai, Yuting Zhang, Ravi Kumar Satzoda, Vijay Mahadevan, and R Manmatha. Polyformer: Referring image segmentation as sequential polygon generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18653-18663, 2023. 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.316, + 0.471, + 0.384 + ], + "angle": 0, + "content": "[56] Qin Liu, Zhenlin Xu, Gedas Bertasius, and Marc Niethammer. Simpleclick: Interactive image segmentation with simple vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22290-22300, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.385, + 0.471, + 0.44 + ], + "angle": 0, + "content": "[57] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. arXiv preprint arXiv:2103.14030, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.442, + 0.471, + 0.496 + ], + "angle": 0, + "content": "[58] Jiasen Lu, Christopher Clark, Rowan Zellers, Roozbeh Mottaghi, and Aniruddha Kembhavi. Unified-io: A unified model for vision, language, and multi-modal tasks. arXiv preprint arXiv:2206.08916, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.498, + 0.471, + 0.551 + ], + "angle": 0, + "content": "[59] Chuofan Ma, Yi Jiang, Xin Wen, Zehuan Yuan, and Xiaojuan Qi. Codet: Co-occurrence guided region-word alignment for open-vocabulary object detection. arXiv preprint arXiv:2310.16667, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.553, + 0.471, + 0.607 + ], + "angle": 0, + "content": "[60] Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. TrackFormer: Multi-object tracking with transformers. arXiv preprint arXiv:2101.02702, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.609, + 0.471, + 0.663 + ], + "angle": 0, + "content": "[61] Tim Meinhardt, Matt Feiszli, Yuchen Fan, Laura Leal-Taixe, and Rakesh Ranjan. Novis: A case for end-to-end near-online video instance segmentation. arXiv preprint arXiv:2308.15266, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.665, + 0.471, + 0.718 + ], + "angle": 0, + "content": "[62] Fausto Miletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.72, + 0.471, + 0.83 + ], + "angle": 0, + "content": "[63] Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, Xiao Wang, Xiaohua Zhai, Thomas Kipf, and Neil Houlsby. Simple open-vocabulary object detection. In Computer Vision - ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part X, page 728-755, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.832, + 0.471, + 0.872 + ], + "angle": 0, + "content": "[64] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In ECCV, 2016. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.874, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[65] Seoung Wug Oh, Joon-Young Lee, Ning Xu, and Seon Joo Kim. Video object segmentation using space-time memory" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.093, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9226-9235, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.135, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[66] Jiangmiao Pang, Linlu Qiu, Xia Li, Haofeng Chen, Qi Li, Trevor Darrell, and Fisher Yu. Quasi-dense similarity learning for multiple object tracking. In CVPR, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.178, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[67] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.233, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[68] Jiyang Qi, Yan Gao, Yao Hu, Xinggang Wang, Xiaoyu Liu, Xiang Bai, Serge Belongie, Alan Yuille, Philip HS Torr, and Song Bai. Occluded video instance segmentation: A benchmark. International Journal of Computer Vision, pages 1-18, 2022. 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.302, + 0.892, + 0.385 + ], + "angle": 0, + "content": "[69] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.386, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[70] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485-5551, 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.456, + 0.892, + 0.523 + ], + "angle": 0, + "content": "[71] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.525, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[72] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2):3, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.581, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[73] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In CVPR, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.637, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[74] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.707, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[75] Carsten Rother, Vladimir Kolmogorov, and Andrew Blake. \"grabcut\" interactive foreground extraction using iterated graph cuts. ACM transactions on graphics (TOG), 23(3): 309-314, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.763, + 0.892, + 0.817 + ], + "angle": 0, + "content": "[76] Samuel Schulter, Vijay Kumar B G, Yumin Suh, Konstantinos M. Dafnis, Zhixing Zhang, Shiyu Zhao, and Dimitris Metaxas. Omnilabel: A challenging benchmark for language-based object detection. In ICCV, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.818, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[77] Seonguk Seo, Joon-Young Lee, and Bohyung Han. Urvos: Unified referring video object segmentation network with a large-scale benchmark. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XV 16, pages 208–223. Springer, 2020. 2, 4, 6" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3793" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.092, + 0.47, + 0.148 + ], + "angle": 0, + "content": "[78] Jing Shao, Siyu Chen, Yangguang Li, Kun Wang, Zhenfei Yin, Yinan He, Jianing Teng, Qinghong Sun, Mengya Gao, Jihao Liu, et al. Intern: A new learning paradigm towards general vision. arXiv preprint arXiv:2111.08687, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.151, + 0.472, + 0.221 + ], + "angle": 0, + "content": "[79] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.223, + 0.471, + 0.292 + ], + "angle": 0, + "content": "[80] Konstantin Sofiuk, Ilya A Petrov, and Anton Konushin. Reviving iterative training with mask guidance for interactive segmentation. In 2022 IEEE International Conference on Image Processing (ICIP), pages 3141-3145. IEEE, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.295, + 0.471, + 0.378 + ], + "angle": 0, + "content": "[81] Peize Sun, Rufeng Zhang, Yi Jiang, Tao Kong, Chenfeng Xu, Wei Zhan, Masayoshi Tomizuka, Lei Li, Zehuan Yuan, Changhu Wang, et al. Sparse r-cnn: End-to-end object detection with learnable proposals. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14454-14463, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.381, + 0.471, + 0.45 + ], + "angle": 0, + "content": "[82] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.453, + 0.47, + 0.523 + ], + "angle": 0, + "content": "[83] Haochen Wang, Cilin Yan, Shuai Wang, Xiaolong Jiang, Xu Tang, Yao Hu, Weidi Xie, and Efstratios Gavves. Towards open-vocabulary video instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4057-4066, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.526, + 0.471, + 0.609 + ], + "angle": 0, + "content": "[84] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In International Conference on Machine Learning, pages 23318-23340. PMLR, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.612, + 0.471, + 0.653 + ], + "angle": 0, + "content": "[85] Weiyao Wang, Matt Feiszli, Heng Wang, and Du Tran. Unidentified video objects: A benchmark for dense, open-world segmentation. In ICCV, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.656, + 0.471, + 0.726 + ], + "angle": 0, + "content": "[86] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, et al. Image as a foreign language: Beit pretraining for all vision and vision-language tasks. arXiv preprint arXiv:2208.10442, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.729, + 0.471, + 0.783 + ], + "angle": 0, + "content": "[87] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.787, + 0.471, + 0.828 + ], + "angle": 0, + "content": "[88] Nicolai Wojke, Alex Bewley, and Dietrich Paulus. Simple online and realtime tracking with a deep association metric. In ICIP, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.831, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[89] Jiajun Wu, Yibiao Zhao, Jun-Yan Zhu, Siwei Luo, and Zhuowen Tu. Milcut: A sweeping line multiple instance learning paradigm for interactive image segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 256-263, 2014. 5" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.092, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[90] Junfeng Wu, Yi Jiang, Song Bai, Wenqing Zhang, and Xiang Bai. Seqformer: Sequential transformer for video instance segmentation. In ECCV, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.136, + 0.894, + 0.203 + ], + "angle": 0, + "content": "[91] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4974-4984, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.205, + 0.892, + 0.258 + ], + "angle": 0, + "content": "[92] Junfeng Wu, Qihao Liu, Yi Jiang, Song Bai, Alan Yuille, and Xiang Bai. In defense of online models for video instance segmentation. In ECCV, pages 588-605. Springer, 2022. 2, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.261, + 0.892, + 0.327 + ], + "angle": 0, + "content": "[93] Jiannan Wu, Yi Jiang, Bin Yan, Hutchuan Lu, Zehuan Yuan, and Ping Luo. Segment every reference object in spatial and temporal spaces. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2538-2550, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.33, + 0.892, + 0.397 + ], + "angle": 0, + "content": "[94] Bin Xiao, Haiping Wu, Weijian Xu, Xiyang Dai, Houdong Hu, Yumao Lu, Michael Zeng, Ce Liu, and Lu Yuan. Florence-2: Advancing a unified representation for a variety of vision tasks. arXiv preprint arXiv:2311.06242, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.4, + 0.892, + 0.48 + ], + "angle": 0, + "content": "[95] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2955-2966, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.483, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[96] Ning Xu, Linjie Yang, Jianchao Yang, Dingcheng Yue, Yuchen Fan, Yuchen Liang, and Thomas S. Huang. Youtuvis dataset 2021 version. https://youtu-vos.org/dataset/vis/.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.539, + 0.892, + 0.594 + ], + "angle": 0, + "content": "[97] Ning Xu, Brian Price, Scott Cohen, Jimei Yang, and Thomas S Huang. Deep interactive object selection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 373-381, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.596, + 0.892, + 0.65 + ], + "angle": 0, + "content": "[98] Ning Xu, Linjie Yang, Yuchen Fan, Dingcheng Yue, Yuchen Liang, Jianchao Yang, and Thomas Huang. Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327, 2018. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.651, + 0.892, + 0.692 + ], + "angle": 0, + "content": "[99] Bin Yan, Yi Jiang, Peize Sun, Dong Wang, Zehuan Yuan, Ping Luo, and Huchuan Lu. Towards grand unification of object tracking. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.693, + 0.892, + 0.746 + ], + "angle": 0, + "content": "[100] Bin Yan, Yi Jiang, Jiannan Wu, Dong Wang, Ping Luo, Zehuan Yuan, and Huchuan Lu. Universal instance perception as object discovery and retrieval. In CVPR, 2023. 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.748, + 0.892, + 0.775 + ], + "angle": 0, + "content": "[101] Haosen Yang, Chuofan Ma, Bin Wen, Yi Jiang, Zehuan Yuan, and Xiatian Zhu. Recognize any regions. 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.777, + 0.892, + 0.803 + ], + "angle": 0, + "content": "[102] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV, 2019. 2, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.805, + 0.892, + 0.873 + ], + "angle": 0, + "content": "[103] Zhengyuan Yang, Zhe Gan, Jianfeng Wang, Xiaowei Hu, Faisal Ahmed, Zicheng Liu, Yumao Lu, and Lijuan Wang. Unitab: Unifying text and box outputs for grounded vision-language modeling. In European Conference on Computer Vision, pages 521-539. Springer, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[104] Lewei Yao, Jianhua Han, Youpeng Wen, Xiaodan Liang, Dan Xu, Wei Zhang, Zhenguo Li, Chunjing Xu, and Hang" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.092, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3794" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "Xu. Detclip: Dictionary-enriched visual-concept paralleled pre-training for open-world detection. In NeurIPS, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.471, + 0.205 + ], + "angle": 0, + "content": "[105] Lewei Yao, Jianhua Han, Xiaodan Liang, Dan Xu, Wei Zhang, Zhenguo Li, and Hang Xu. Detclipv2: Scalable open-vocabulary object detection pre-training via word-region alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23497-23506, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.471, + 0.29 + ], + "angle": 0, + "content": "[106] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2636-2645, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.292, + 0.471, + 0.348 + ], + "angle": 0, + "content": "[107] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. Transactions on Machine Learning Research, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.471, + 0.39 + ], + "angle": 0, + "content": "[108] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In ECCV, 2016. 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.392, + 0.471, + 0.461 + ], + "angle": 0, + "content": "[109] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.463, + 0.471, + 0.532 + ], + "angle": 0, + "content": "[110] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-vocabulary object detection using captions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14393-14402, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.534, + 0.471, + 0.617 + ], + "angle": 0, + "content": "[111] Haotian Zhang, Pengchuan Zhang, Xiaowei Hu, Yen-Chun Chen, Liunian Li, Xiyang Dai, Lijuan Wang, Lu Yuan, Jenq-Neng Hwang, and Jianfeng Gao. Glipv2: Unifying localization and vision-language understanding. In Advances in Neural Information Processing Systems, pages 36067-36080, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.619, + 0.471, + 0.688 + ], + "angle": 0, + "content": "[112] Tao Zhang, Xingye Tian, Yu Wu, Shunping Ji, Xuebo Wang, Yuan Zhang, and Pengfei Wan. Dvis: Decoupled video instance segmentation framework. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1282-1291, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.69, + 0.471, + 0.745 + ], + "angle": 0, + "content": "[113] Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. Byte-track: Multi-object tracking by associating every detection box. arXiv preprint arXiv:2110.06864, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.471, + 0.815 + ], + "angle": 0, + "content": "[114] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Lianian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In CVPR, pages 16793-16803, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[115] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Lianian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16793-16803, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "[116] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Tracking objects as points. In ECCV, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[117] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In ECCV, pages 350-368. Springer, 2022. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[118] Chaoyang Zhu, Yiyi Zhou, Yunhang Shen, Gen Luo, Xingjia Pan, Mingbao Lin, Chao Chen, Liujuan Cao, Xiaoshuai Sun, and Rongrong Ji. Seqtr: A simple yet universal network for visual grounding. In European Conference on Computer Vision, pages 598-615. Springer, 2022. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[119] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[120] Xizhou Zhu, Jinguo Zhu, Hao Li, Xiaoshi Wu, Hongsheng Li, Xiaohua Wang, and Jifeng Dai. Uni-perceiver: Pretraining unified architecture for generic perception for zero-shot and few-shot tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16804-16815, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.39, + 0.892, + 0.472 + ], + "angle": 0, + "content": "[121] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15116-15127, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.474, + 0.892, + 0.516 + ], + "angle": 0, + "content": "[122] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. In NeurIPS, 2023. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3795" + } + ] +] \ No newline at end of file diff --git a/2024/General Object Foundation Model for Images and Videos at Scale/0bc99999-be81-4577-bff4-eaa67b1ee51f_origin.pdf b/2024/General Object Foundation Model for Images and Videos at Scale/0bc99999-be81-4577-bff4-eaa67b1ee51f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0593f1dbbeba75794beb6fbac2fe9457cda74994 --- /dev/null +++ b/2024/General Object Foundation Model for Images and Videos at Scale/0bc99999-be81-4577-bff4-eaa67b1ee51f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3f0e088bdb5f1ef98aa2d5afe1172b8f8e648ccda1fa0f3449f3daee98322395 +size 9764654 diff --git a/2024/General Object Foundation Model for Images and Videos at Scale/full.md b/2024/General Object Foundation Model for Images and Videos at Scale/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ad9c4a2de60e908436d5bd8d6a2add7a2f754399 --- /dev/null +++ b/2024/General Object Foundation Model for Images and Videos at Scale/full.md @@ -0,0 +1,348 @@ +# General Object Foundation Model for Images and Videos at Scale + +Junfeng Wu $^{1*}$ , Yi Jiang $^{2*}$ , Qihao Liu $^{3}$ , Zehuan Yuan $^{2}$ , Xiang Bai $^{1\dagger}$ , Song Bai $^{2\dagger}$ $^{1}$ Huazhong University of Science and Technology, $^{2}$ ByteDance Inc., $^{3}$ Johns Hopkins University + +# Abstract + +We present GLEE in this work, an object-level foundation model for locating and identifying objects in images and videos. Through a unified framework, GLEE accomplishes detection, segmentation, tracking, grounding, and identification of arbitrary objects in the open world scenario for various object perception tasks. Adopting a cohesive learning strategy, GLEE acquires knowledge from diverse data sources with varying supervision levels to formulate general object representations, excelling in zero-shot transfer to new data and tasks. Specifically, we employ an image encoder, text encoder, and visual prompter to handle multi-modal inputs, enabling to simultaneously solve various object-centric downstream tasks while maintaining state-of-the-art performance. Demonstrated through extensive training on over five million images from diverse benchmarks, GLEE exhibits remarkable versatility and improved generalization performance, efficiently tackling downstream tasks without the need for task-specific adaptation. By integrating large volumes of automatically labeled data, we further enhance its zero-shot generalization capabilities. Additionally, GLEE is capable of being integrated into Large Language Models, serving as a foundational model to provide universal object-level information for multi-modal tasks. We hope that the versatility and universality of our method will mark a significant step in the development of efficient visual foundation models for AGI systems. The models and code are released at https://github.com/FoundationVision/GLEE. + +# 1. Introduction + +Foundation model [7] is an emerging paradigm for building artificial general intelligence (AGI) systems, signifying a model trained on broad data that is capable of being adapted to a wide range of downstream tasks in an general paradigm. Recently, NLP foundation models such as BERT [21], GPT-3 [9], T5 [70] developed with unified + +![](images/d775f818d6f3630be9dc6f16488fa0039086e3fee565a2cb231ebaa038f24436.jpg) +Figure 1. The performance of GLEE on a broad range of object-level tasks compared with existing models. + +input-output paradigms and large-scale pre-training, have achieved remarkable generalization capabilities to address nearly all NLP tasks. + +In computer vision, the diversity of task types and the lack of a unified form make visual foundation models only serve specific subdomains, such as CLIP [69] for multimodal visual model, MAE [32] for visual representations model, SAM [39] for segmentation model. Despite being widely studied, current visual foundation models are still focusing on establishing correlations between global image features and language descriptions or learning image-level feature representations. However, locating and identifying objects constitute foundational capabilities in computer vision systems, serves as a basis for solving complex or high level vision tasks such as segmentation, scene understanding, object tracking, event detection, and activity recognition and support a wide range of applications. + +In this work, we advance the development of object-level foundation models within the visual domain. To address the aforementioned limitation, providing general and accurate object-level information, we introduce a general object visual foundation model, coined as GLEE, which simultaneously solve a wide range of object-centric tasks while ensuring SOTA performance, including object detection, instance segmentation, grounding, object tracking, interactive segmentation and tracking, etc., as shown in Figure 1. Through + +a unified input and output paradigm definition, our model is capable of learning from a wide range of diverse data and predicting general object representations, which makes it generalize well to new data and tasks in a zero-shot manner and achieve amazing performance. In addition, thanks to the unified paradigm, the training data can be scaled up at low cost by introducing a large amount of automatically labeled data, and further improve the zero-shot generalization ability of the model. + +A general object foundation model framework. Our objective is to build an object visual foundation model capable of simultaneously addressing a wide range of object-centric tasks. Specifically, we employ an image encoder, a text encoder, and a visual prompter to encode multi-modal inputs. They are integrated into a detector to extract objects from images according to textual and visual input. This unified approach to handle multiple modalities enables us to concurrently solve various object-centric tasks, including detection [10, 52, 81, 119], instance segmentation [15, 31], referring expression comprehension [35, 55, 93, 118], interactive segmentation [1, 12, 122], multi-object tracking [20, 60, 99, 113, 116], video object segmentation [16, 17, 65, 98], video instance segmentation [34, 87, 90, 92, 102], and video referring segmentation [77, 91, 93], all while maintaining state-of-the-art performance. + +Multi-granularity joint supervision and scalable training paradigm. The design of the unified framework capable of addressing multiple tasks enables joint training on over five million images from diverse benchmarks and varying levels of supervision. Existing datasets differ in annotation granularity: detection datasets like Objects365 [79] and OpenImages [42] offer bounding boxes and category names; COCO [52] and LVIS [29] provide finer-grained mask annotations; RefCOCO [64, 108] and Visual Genome [40] provide detailed object descriptions. Additionally, video data enhance the temporal consistency of model, while open-world data contribute class-agnostic object annotations. An intuitive display of the supervision types and data scales of the datasets employed is presented in Figure 2. The unified support for multi-source data in our approach greatly facilitates the incorporation of additional manually or automatically annotated data, enabling easy scaling up of the dataset. Furthermore, the alignment of model optimization across tasks means that joint training serves not only as a unifying strategy but also as a mechanism to boost performance across individual tasks. + +Strong zero-shot transferability to a wide range of object level image and video tasks. After joint training on data from diverse sources, GLEE demonstrates remarkable versatility and zero-shot generalization abilities. Extensive experiments demonstrate that GLEE achieves state-of-the-art performance compared to existing specialist and generalist models in object-level image tasks such as detec + +tion, referring expression comprehension, and open-world detection, all without requiring any task-specific designs or fine-tuning. Furthermore, we showcase the extraordinary generalization and zero-shot capabilities of GLEE in large-vocabulary open-world video tracking tasks, achieving significantly superior performance over existing models even in a zero-shot transfer manner. Additionally, by incorporating automatically annotated data like SA1B [39] and GRIT [67], we are able to scale up our training dataset to an impressive size of 10 million images at a low cost, which is typically challenging to achieve for object-level tasks and further enhances the generalization performance. Moreover, we replace the SAM [39] component with GLEE in a multimodal Large Language Model (mLLM) [43] and observe that it achieves comparable results. This demonstrates that GLEE is capable of supplying the visual object-level information that modern LLMs currently lack, thus laying a solid foundation for an object-centric mLLMs. + +# 2. Related Work + +# 2.1. Visual Foundation Model + +As foundation models [9, 18, 21, 70, 82] in the NLP field have achieved remarkable success, the construction of visual foundation models attracts increasing attention. Unlike NLP tasks that are predominantly unified under a text-to-text paradigm, tasks in Computer Vision still exhibit significant differences in formulation. This disparity leads to the fact that visual models are typically trained in a single-task learning frameworks, limiting their applicability to tasks within certain sub-domains. For instance, multi-modal visual foundation models like CLIP [69], ALIGN [37], Florence [109], BEIT3 [86], Flamingo[2] make significant advancements in efficient transfer learning and demonstrate impressive zero-shot capabilities on vision-language tasks by employing contrastive learning and masked data modeling on large-scale image-text pairs. DALL-E [71, 72] and Stable Diffusion [74] are trained on massive pairs of images and captions, enabling them to generate detailed image content conditioned on textual instruction. DINO [11], MAE [32], EVA [25], ImageGPT [13] obtain strong visual representations through self-supervised training on largescale image data, which are then employed to adopt downstream tasks. These foundation models learn image-level features and are not directly applicable to object-level tasks. The recently proposed SAM [39], capable of segmenting any object of a given image based on visual prompt such as points and boxes, provides rich object-level information and demonstrates strong generalization capabilities. However, the object information lacks semantic context, limiting its application in object-level tasks. Unlike existing visual foundation models, we aim to develop an object foundation model that directly solve downstream tasks without the need + +![](images/ec3537dc2cf9d5bae65c56ad656596c2164510ac2cca85a3b8aa8f6dcbb89c8b.jpg) +Figure 2. An illustrative example showcasing annotations of varying granularities from different datasets, along with the scale of data we utilized. Training on datasets from multiple sources endows the model with more universal representations. + +for additional parameters or fine-tuning. + +# 2.2. Unified and General Model + +Unified models share similarities with foundation models in the aspect of multi-task unification for their ability to handle multiple vision or multi-modal tasks within a single model. MuST [27] and Intern [78] propose to train across multiple vision tasks and solving them simultaneously. Inspired by the success of sequence-to-sequence NLP models [9, 70], models such as Uni-Perceiver [120], OFA [84], Unified-IO [58], Pix2Seq v2 [14], and UniTAB [103] propose modeling various tasks as sequence generation tasks within a unified paradigm. While these approaches have demonstrated promising cross-task generalization capabilities, they focus mainly on image-level understanding tasks. In addition, their auto-regressive generation of boxes and masks results in significantly slower inference speeds and the performance still falls short of state-of-the-art task-specific models. Building upon on detectors [45, 119], UniPerceiver v2 [46] and UNINEXT [100] utilize unified maximum likelihood estimation and object retrieval to support various tasks, effectively resolves the challenges of localization. Nonetheless, they are trained on closed-set data, thereby not exhibiting zero-shot generalization capabilities. X-decoder [121] and SEEM [122] construct a generalized + +decoding model capable of predicting pixel-level segmentation and language tokens. Diverging from unified models, the proposed GLEE not only directly addresses object-level tasks in a unified manner but also provides universal object representations, which generalize well to new data and tasks, serving as a cornerstone for a broader range of tasks that require detailed object information. + +# 2.3. Vision-Language Understanding + +Open-vocabulary detection (OVD) and Grounding models both necessitate the localization and recognition of as many objects as possible. With the recent advancements in vision-language pre-training [37, 69, 107, 109], a commonly employed strategy for OVD involves transferring the knowledge from pre-trained vision-language models (VLMs) to object detectors [28, 41, 63]. Another group of studies leverages extensive image-text pair datasets to broaden the detection vocabulary [26, 47, 51, 59, 101, 105, 110, 115]. However, these language-based detectors are inherently constrained by the capabilities and biases of language models, making it challenging to excel simultaneously in both localization and recognition. Our objective is to optimally utilize existing datasets to construct a general object-level foundation model, aims to not only detect and identify objects effectively but also to offer universal object representations. + +tations for a wide range of downstream tasks. + +# 3. Method + +# 3.1. Formulation + +The proposed GLEE consists of an image encoder, a text encoder, a visual prompter, and an object decoder, as illustrated in Figure 3. The text encoder processes arbitrary descriptions related to the task, including object categories, names in any form, captions about objects, and referring expressions. The visual prompter encodes user inputs such as points, bounding boxes, or scribbles during interactive segmentation into corresponding visual representations of target objects. Then they are integrated into a detector for extracting objects from images according to textual and visual input. We build the object decoder upon MaskDINO [45] with a dynamic class head by compute similarity between object embedding from detector and text embedding from the text encoder. Given an input image $I \in \mathcal{R}^{3 \times H \times W}$ , we first extract multi-scale features $Z$ with backbones such as ResNet [30]. Then we feed them into the object decoder and adopt three prediction heads (classification, detection, and segmentation) on the output embedding $q_{d} \in \mathcal{R}^{N \times C}$ from decoder. Following other object segmentation models [15, 45, 50], we construct a 1/4 resolution pixel embedding map $M_{p} \in \mathcal{R}^{C \times \frac{H}{4} \times \frac{W}{4}}$ which is obtained by upsampling and fusing multi-scale feature maps from the backbone and Transformer encoder. Finally, we obtain each binary mask prediction $m \in \mathcal{R}^{N \times \frac{H}{4} \times \frac{W}{4}}$ via a dot product between the N mask embeddings and pixel embedding map: + +$$ +m = F F N \left(q _ {d}\right) \otimes M _ {p}, \tag {1} +$$ + +where FFN is a 3-layer feed forward head with ReLU activation function and a linear projection layer. + +To support arbitrary vocabularies and object descriptions, we replace the FFN classifier with text embeddings following DetCLIP [104]. Specifically, we feed K category names as separate sentences into the text encoder $Enc_{L}$ and use the average of each sentence tokens as the output text embedding $e_t \in \mathcal{R}^{K \times D}$ for each category or description. Then we compute the alignment scores $S_{\text{align}} \in \mathcal{R}^{N \times K}$ between object embedding and text embedding: + +$$ +S _ {\text {a l i g n}} = q _ {d} \cdot W _ {i 2 t} \otimes e _ {t}, \tag {2} +$$ + +where $W_{i2t} \in \mathcal{R}^{C \times D}$ is image-to-text projection weights. We use logits $S_{\text{align}}$ to replace traditional classification logits to compute Hungarian matching cost during training and assign categories to objects during inference. To make the original visual features prompt-aware, an early fusion module is adopted before Transformer encoder following UNNEXT [100], which takes image feature from backbone + +and prompt embedding as input and perform bi-directional cross-attention between them. + +# 3.2. Task Unification + +Based on the above designs, GLEE can be used to seamlessly unify a wide range of object perception tasks in images and videos, including object detection, instance segmentation, grounding, multi-target tracking (MOT), video instance segmentation (VIS), video object segmentation (VOS), interactive segmentation and tracking, and supports open-world/large-vocabulary image and video detection and segmentation tasks. + +Detection and Instance Segmentation. For detection task, a fixed-length category list is given and all objects in the category list are required to be detected. For a dataset with category list length $K$ , the text input can be formulated as $\{p_k\}_{k=1}^K$ where $p_k$ represents for the $k$ -th category name, e.g., $P = ["person", "bicycle", "car", ..., "toothbrush"]$ for COCO [52]. For datasets with large vocabulary, calculating the text embedding of all categories is very time-consuming and redundant. Therefore, for datasets with a category number greater than 100, such as objects365 [79] and LVIS [29], suppose there are $\hat{K}$ positive categories in an image, we take the $\hat{K}$ positive categories and then pad the category number to 100 by randomly sampling from the negative categories. For instance segmentation, we enable the mask branch and add mask matching cost with mask loss. + +Grounding and Referring Segmentation. These tasks provide reference textual expressions, where objects are described with attributes, for example, Referring Expression Comprehension (REC) [108, 118], Referring Expression Segmentation (RES) [55, 108], and Referring Video Object Segmentation (R-VOS) [77, 91] aim at finding objects matched with the given language expressions like "The fourth person from the left". For each image, we take the all the object expressions as text prompt and feed the them into the text encoder. For each expressions, we apply global average pooling along the sequence dimension to get text embedding $e_t$ . The text embeddings are feed into early fusion module and additionally interact with object queries through self-attention module in the decoder. + +MOT and VIS. Both Multi-object Tracking (MOT)[4, 20, 60, 113, 116] and Video Instance Segmentation (VIS)[34, 68, 92, 102] need to detect and track all the objects in the predefined category list, and VIS requires additional mask for the objects. These two tasks can be considered as extended tasks of detection and instance segmentation on videos. We found that with sufficient image exposure, object embeddings from the decoder effectively distinguish objects in a video, showing strong discriminability and temporal consistency. As a result, they can be directly employed for tracking without the need for an additional tracking head. Training on image-level data can address + +category: [person, car, motorcycle, dog ...] + +arbitrary name: bollard, manhole cover + +expression: motorcycle parked under the sign + +object caption: air conditioner outdoor unit + +![](images/d77eac373c21a127a77e5f9ed1534ad09fedd8b60ecf6ae60d3aed317cd2d1d6.jpg) + +![](images/fd5bf6f7d382a23c2b209900ad9288d278edfa4a146ecd6445141003592729fa.jpg) + +![](images/6679f185efcc4c420e357c2b3a550885b414f1f11be178a3d04d42fd6cbf0546.jpg) + +![](images/3504d38790bbdfc686d18d7c87b5544f69d5c75ca375cddaf498c1bd6cbde8b3.jpg) + +![](images/825e83f34e9b6d73c922fd1650abbdafe1b6a28bec1a6d8ced44ae1560a4f802.jpg) +(a) Object foundation model framework + +category: person, car motorcycle,dog, cat. + +![](images/f4678e4883caeb60898c1376f7bf588a0ea878e5aead4d427a18537bbe48d4dc.jpg) +Figure 3. Framework of GLEE. The text encoder accepts textual descriptions in various forms from diverse data sources, including object categories, names, captions, and referring expressions. The visual prompt encodes points, bounding boxes, or scribbles into corresponding visual representations. The object decoder takes them and image features to predict objects in images. (b) illustrates the application of GLEE to image tasks tailored for different language descriptions and visual prompts. (c) demonstrates the application across various object-level video tasks. + +![](images/a13df688cdf1ec2d485edd19cd26c30d065014904c61b37d09462586d44b3d1c.jpg) + +arbitrary name + +bollard + +manhole cover + +![](images/a9fa2576b35dd255726e944333ef5f7b8fff289815599f3faa3fda15a177bb4c.jpg) +(b) Applied to image tasks + +expression: motorcycle parked under the sign + +![](images/09d3d65d6993c2444d0a26814dc06a76a9de79e59fc7d90e24780507f945bdd2.jpg) + +interactive segment + +![](images/0cd06ba59cf65c2ef56bf0a031d4d5d0b7d037f876d0b73a70fae36756d742bd.jpg) + +video tasks: VIS、MOT、VOS、RVOS、 open world/vocabulary tracking + +![](images/b0c0fbdb3d3df8d8b94d5418e57f67229ee00342958772a790951c1c1eec0230.jpg) +(c) Applied to video tasks + +interactive tracking + +![](images/7124886eb182e23855f7190caf02c1001c53d315a231d0c9af0091c2355a65dc.jpg) + +straightforward tracking scenarios, but in cases of severe occlusion scenes, such as OVIS [68], image-level training cannot guarantee that the model exhibits strong temporal consistency under occlusion conditions. Therefore, for occlusion scenarios, it is essential to utilize video data for training. Following IDOL [92], we sample two frames from a video and introduce contrastive learning between frames to make the embedding of the same object instance closer in the embedding space, and the embedding of different object instances farther away. During Inference, the detected objects are tracked by simple bipartite matching of the corresponding object queries following MinVIS [36]. + +Visual Prompted Segmentation. Interactive segmentation[8, 12, 56, 75, 80, 89, 97] takes various forms of visual prompt, such as points, boxes, or scribbles, to segment the specified objects within an image. On the other hand, VOS [22, 98] aims to segment the entire object throughout the entire video based on a mask provided in the first frame of the video. We extract visual prompt embeddings twice in the model. First, we crop the prompt square area from RGB image and send it to the backbone to obtain the visual prompt feature of the corresponding area, and send it to the early fusion module before the Transformer encoder. Second, we sample fine-grained visual embeddings from the pixel embedding map $M_{p}$ according to visual prompt and make them interact with object queries through self-attention module in the Transformer decoder layer, as the same with text embeddings. + +# 3.3. Training Unification + +Tasks with Dynamic Loss. We jointly train GLEE in an end-to-end manner on over 5 million images from diverse benchmarks with various levels of supervision. Different loss functions are selected for training on various datasets. There are six types of losses for our GLEE: semantic loss, box loss, mask loss, confidence loss, contrastive tracking loss, and distillation loss. For all tasks with category list or object expressions, we apply Focal loss [53] as semantic loss on logits $S_{\text{align}}$ to align the text concepts with object features. For box prediction, we use a combination of L1 loss and generalized IoU loss [73]. The mask loss is defined as a combination of the Dice loss [62] and Focal loss. For the Visual Prompt Segmentation tasks, we employ an additional FFN to predict the confidence score for each object queries supervised by Focal loss. For video tasks fine-tuning, we sample two frames and apply contrastive tracking loss on the object query from the last layer of decoder following IDOL [92]. For the text encoder, we distill the knowledge from the frozen teacher CLIP text encoder to ensure the text embedding in pre-trained vison-language embedding space. We apply an L1 loss between our text encoder and CLIP text encoder to minimize their distance: + +$$ +\mathcal {L} _ {\text {t e x t}} = \frac {1}{K} \sum_ {i = 0} ^ {K} \| \operatorname {E n c} _ {L} \left(p _ {k}\right) - \operatorname {E n c} _ {C L I P} \left(p _ {k}\right) \|, \tag {3} +$$ + +where $\{p_k\}$ is the category name from the category list $P$ + +
MethodTypeGeneric Detection & SegmentationReferring Detection & SegmentationOpenWorld
COCO-valCOCO-test-devLVISRefCOCORefCOCO+RefCOCOgUVO
APboxAPmaskAPboxAPmaskAPboxAPr-boxAPmaskAPr-maskP@0.5IoUP@0.5IoUP@0.5IoUARmask
MDETR [38]--------87.5-81.1-83.4--
SeqTR [118]--------87.071.778.763.082.764.7-
PolyFormer (L) [55]--------90.476.985.072.285.871.2-
ViTDet-L [50]Specialist Models57.649.8--51.2-46.034.3-------
ViTDet-H [50]58.750.9--53.4-48.136.9-------
EVA-02-L [24]64.255.064.555.865.2-57.3-
ODISE [95]--------------57.7
Mask2Former (L) [15]-50.1-50.5-----------
MaskDINO (L) [45]-54.5-54.7-----------
UniTAB (B) [103]--------88.6-81.0-84.6--
OFA (L) [84]--------90.1-85.8-85.9--
Pix2Seq v2 [14]46.538.2-------------
Uni-Perceiver-v2 (B) [46]58.650.6-------------
Uni-Perceiver-v2 (L) [46]61.953.6-------------
UNINEXT (R50) [100]Generalist Models51.344.9--36.4---89.777.979.866.284.070.0-
UNINEXT (L) [100]58.149.6------91.480.383.170.086.973.4-
UNINEXT (H) [100]60.651.8------92.682.285.272.588.774.7-
GLIPv2 (B) [111]--58.845.8-----------
GLIPv2 (H) [111]--60.648.9-----------
X-Decoder (B) [121]-45.8-45.8-----------
X-Decoder (L) [121]-46.7-47.1-----------
Florence-2 (L) [94]43.4-------93.4-88.3-91.2--
GLEE-LiteFoundation Models55.048.454.748.344.236.740.233.788.577.478.364.882.968.866.6
GLEE-Plus60.453.060.653.352.744.547.440.490.679.581.668.385.070.670.6
GLEE-Pro62.054.262.354.555.749.249.944.391.080.082.669.686.472.972.6
+ +Table 1. Comparison of GLEE to recent specialist and generalist models on object-level image tasks. For REC and RES tasks, we report Precision@0.5 and overall IoU (oIoU). For open-world instance segmentation task, we reported the average recall of 100 mask proposals (AR@100) on the UVO [85]. + +Data Scale Up. A visual foundation model should be able to easily scale up the training data and achieve better generalization performance. Thanks to the unified training paradigm, the training data can be scaled up at low cost by introducing a large amount of automatically labeled data from SA1B [39] and GRIT [67]. SA1B provides large and detailed mask annotations, which enhance the general object perception capabilities of model, while GRIT offers a more extensive collection of referring-expression-bounding-box pairs, improving the object identification abilities and the understanding capability of object descriptions. Ultimately, we introduced 2 million SA1B data and 5 million GRIT data into the training process, bringing the total training data to 10 million. + +# 4. Experiments + +# 4.1. Experimental Setup + +We conducted training in three stages. Initially, we performed pretraining for object detection task on Objects365 [79] and OpenImages [42], initializing the text encoder with pretrained CLIP [69] weights and keeping the parameters frozen. In the second training stage, we introduced additional instance segmentation datasets, including COCO [52], LVIS [29], and BDD [106]. Furthermore, we treat three VIS datasets: YTVIS19 [102], YTVIS21 [96], and OVIS [68], as independent image data to enrich the scenes. For datasets that provide descriptions of objects, we included RefCOCO [108], RefCOCO+ [108], RefCOCOg [64], Visual Genome [40], and RVOS [77]. Addi + +tionally, we introduced two open-world instance segmentation datasets, UVO [85] and a subset of SA1B [39]. Building upon this, we perform the third training stage by introducing more SA1B data and GRIT [67] data to scale up the training set, resulting in a model named GLEE-scale, which exhibited even stronger zero-shot performance on various downstream tasks. During the second and third stages, text encoder is unfrozen but supervised by distillation loss to ensure the predicted text embedding in CLIP embedding space. After the second step, GLEE demonstrated state-of-the-art performance on a range of downstream image and video tasks and exhibited strong zero-shot generalization capabilities, unless otherwise specified, all the experimental results presented below were obtained by the model from this stage. We developed GLEE-Lite, GLEE-Plus, and GLEE-Pro using ResNet-50 [30], SwinLarge [57], and EVA-02 Large [24] as the vision encoder respectively, and train GLEE on 64 A100 GPUs for 500,000 iterations in each stage. More detailed information on data usage, data sampling strategies, and model training can be found in the supplementary materials. + +# 4.2. Comparison with Generalist Models + +We demonstrate the universality and effectiveness of our model as an object-level visual foundation model, directly applicable to various object-centric tasks while ensuring SOTA performance without needing fine-tuning. We report detection and instance segmentation results on both the COCO-2017 [52] and LVIS val v1.0 [29]. While sharing almost identical image sets, LVIS is distinguished by + +
MethodTracking Any Object (TAO [19])BURST [3]LV-VIS [83]
TETALocAAssocAClsAALLCommonUncommonAP\( AP_b \)\( AP_n \)
HOTAmAPHOTAmAPHOTAmAP
Tracktor [5]24.247.413.012.1---------
DeepSORT [88]26.048.417.512.1---------
Tracktor++ [19]28.049.022.812.1---------
QDTrack [66]30.050.527.412.1---------
TETer [48]33.351.635.013.2---------
OVTrack† [49]34.749.336.718.1---------
STCN Tracker† [3]----5.50.917.50.72.50.6---
Box Tracker† [3]----8.21.427.03.03.60.9---
Detic [117]-SORT† [6]----------12.821.16.6
Detic [117]-XMem †[16]----------16.324.110.6
OV2Seg-R50† [83]-----3.7----14.217.211.9
OV2Seg-B† [83]-----4.9----21.127.516.3
UNINEXT (R50) [100]31.943.335.517.1---------
GLEE-Lite†40.156.339.924.122.612.636.418.919.111.019.622.117.7
GLEE-Plus†41.552.940.930.826.917.238.823.723.915.530.331.629.3
GLEE-Pro†47.266.246.229.131.219.248.724.826.917.723.924.623.3
+ +Table 2. Comparison of GLEE to recent specialist and generalist models on object-level video tasks in a zero-shot manner. Evaluation metrics of BURST are reported separately for 'common', 'uncommon' and 'all' classes. The mAP computes mask IoU at the track level, HOTA is a balance of per-frame detection accuracy (DetA) and temporal association accuracy (AssA), and TETA that deconstructs detection into localization and classification components. The AP, $\mathrm{AP_b}$ , and $\mathrm{AP_n}$ in LV-VIS mean the average precision of overall categories, base categories, and novel categories. $\dagger$ does not use videos for training. The under-performance of Pro relative to Plus on LV-VIS is due to Pro employing larger training and inference resolutions, which prove to be sub-optimal for this specific dataset. + +its annotations of over 1,200 object categories, showcasing a long-tail distribution. This distinction makes LVIS more representative of challenging real-world scenarios due to its broader category coverage. As indicated in Table 1, our model outperforms all generalist models on both COCO and LVIS benchmarks. Even when compared to SOTA specialist approaches, which are tailored with specific designs, our model remains highly competitive. This demonstrates that GLEE, while mastering universal and general object representations, concurrently maintains advanced performance. This characteristic is vitally important for adapting to a broad spectrum of downstream tasks requiring precise object localization. For the REC and RES tasks, we evaluated our model on RefCOCO [108], RefCOCO+ [108], and RefCOCOg [64], as show in Table 1, GLEE achieved comparable results with SOTA specialist methods PolyFormer [55], demonstrating strong capability to comprehend textual descriptions and showcasing potential to adapt to a broader range of multi-modal downstream tasks. In open-world instance segmentation tasks, GLEE outperforms previous arts ODISE [95] by 8.9 points, demonstrating the capability of identifying all plausible instances in an open-world scenario. + +# 4.3. Zero-shot Evaluation Across Tasks + +Zero-shot Transfer to Video Tasks. The proposed GLEE is capable of adapting to new data and even new tasks in a zero-shot manner, without the need for additional finetuning. We evaluate its zero-shot capability on three large- + +scale, large-vocabulary open-world video tracking datasets: TAO [19], BURST [3], and LV-VIS [83]. TAO comprises 2,907 high-resolution videos across 833 categories. BURST builds upon TAO, encompassing 425 base categories and 57 novel categories. LV-VIS offers 4,828 videos within 1,196 well-defined object categories. These three benchmarks require the model to detect, classify, and track all objects in videos, while BURST and LV-VIS additionally require segmentation results from the model. In Table 2, we compare the performance of our proposed model with existing specialist models. Notably, the GLEE here is from the second training stage, which has not been exposed to images from these three datasets nor trained on video-level data. Despite these constraints, GLEE achieves state-of-the-art performance that significantly exceeds existing methodologies. Specifically, GLEE surpasses the previous best method OVTrack by $36.0\%$ in TAO, nearly triples the performance of the best baseline in BURST, and outperforms OV2Seg [83] by $43.6\%$ in LV-VIS. This outstanding performance strongly validates the exceptional generalization and zero-shot capabilities of GLEE in handling object-level tasks across a range of benchmarks and tasks. It can be observed that GLEE yields more impressive results on video tasks, since the image tasks have progressed with plentiful data and models from lower costs, while video tasks have not due to higher costs. The models trained on extensive image data with strong general perception capabilities can effectively transfer to video tasks. + +
MethodBackboneYTVIS 2019 val [102]OVIS val [68]
AP\( AP_{50} \)\( AP_{75} \)AP\( AP_{50} \)\( AP_{75} \)
SeqFormer [90]47.469.851.815.131.913.8
IDOL [92]49.574.052.930.251.330.0
VITA [33]49.872.654.519.641.217.4
GenVIS [34]ResNet-5051.372.057.834.559.435.0
DVIS [112]52.676.558.234.159.832.3
NOVIS [61]52.875.756.932.756.232.6
UNINEXT [100]53.075.259.134.055.535.6
GLEE-Lite53.174.059.327.1/32.345.4/52.226.3/33.7
SeqFormer [90]59.382.166.4---
VITA [33]63.086.967.927.751.924.9
IDOL [92]64.387.571.042.665.745.2
GenVIS [34]Swin-L63.885.768.545.469.247.8
DVIS [112]64.988.072.749.975.953.0
NOVIS [61]65.787.872.243.568.343.8
GLEE-Plus63.685.270.529.6/40.350.3/63.828.9/39.8
UNINEXT [100]ConvNeXt-L64.387.271.741.165.842.0
UNINEXT [100]ViT-H66.987.575.149.072.552.2
GLEE-ProEVA02-L67.487.174.138.7/50.459.4/71.439.7/55.5
+ +Table 3. Performance comparison on video instance segmentation tasks. $(\_ /\_)$ reports results from zero-shot and after fine-tuning. + +
Images MethodAPAP-categAP-descrAP-descr-posAP-descr-SAP-descr-MAP-descr-L
AllRegionCLIP [114]2.72.72.63.23.62.72.3
Detic [117]8.015.65.48.05.75.46.2
MDETR [38]--4.79.16.44.64.0
GLIP-T [47]19.323.616.425.829.414.88.2
GLIP-L [47]25.832.921.233.237.718.910.8
FIBER-B [23]25.730.322.334.838.619.512.4
GLEE-Lite20.337.514.019.123.012.710.0
GLEE-Lite-Scale22.735.516.722.333.714.310.2
GLEE-Plus25.446.717.523.928.416.312.5
GLEE-Plus-Scale27.044.519.425.936.017.212.4
+ +Table 4. Evaluation on the OmniLabel benchmark. The final AP value is the geometric mean of categories (AP-categ) and freeform descriptions (AP-descr). + +We additionally provide performance comparison on classic video instance segmentation tasks, whose data is incorporated as image-level data during the second stage of training. As shown in Table 3, on the YTVIS2019 [102] benchmark, our model achieves SOTA results across various model sizes, surpassing all specialist models with complex designs to enhance temporal capabilities and the video unified model UNINEXT [100]. On the OVIS [68] benchmark, which features lengthy videos with extensive object occlusions where temporal capabilities of object features are particularly crucial, our model does not directly reach SOTA. However, after a few hours of simple fine-tuning, it still achieves SOTA performance. More details on VOS, RVOS and demonstrations of interactive segmentation and tracking can be found in supplementary materials. + +Zero-shot Transfer to Real-world Downstream Tasks. To measure generalization on real-world tasks, we evaluate zero-shot performance on OmniLabel [76], which is a benchmark for evaluating language-based object detectors and encourages diverse free-form text descriptions of objects. As show in Table 4, compared to language-based detectors trained on large-scale caption data, GLEE signif + +![](images/c1c1e055e4ea5e7d63b61e43de0e25d2fadccf5f23d8359b4ae7924290d4f83d.jpg) +Figure 4. The performance comparison of replacing SAM with GLEE in LISA, GLEE achieves the same effectiveness as SAM in extracting objects. + +icantly outperforms previous models in AP-categ. Due to the limited captions in our training dataset, it scores lower in AP-descr. By incorporating a more diverse set of box-tection data from the GRIT [67] to scale up our training set, the AP-descr can be elevated to a level comparable with existing models. A more comprehensive report on the zero-shot and few-shot performance on ODinW [44] and ablation studies are provided in the supplementary materials. + +# 4.4. Serve as Foundation Model + +To explore whether GLEE can serve as a foundation model for other architectures, we selected LISA [43] for analysis, a mVLLM that combines LLAVA [54] with SAM [39] for reasoning segmentation. We substituted its vision backbone with a frozen, pretrained GLEE-Plus and fed the object queries from GLEE into LLAVA and remove decoder of LISA. We directly dot product the output SEG tokens with GLEE feature map to generate masks. As shown in Figure 4, after training for the same number of steps, our modified LISA-GLEE achieved comparable results to the original version, demonstrating the versatility of representations from GLEE and its effectiveness in serving other models. + +# 5. Conclusion + +We introduce GLEE, a cutting-edge object-level foundation model designed to be directly applicable to a wide range of object-level image and video tasks. Crafted with a unified learning paradigm, GLEE learns from diverse data sources with varying levels of supervisions. GLEE achieves state-of-the-art performance on numerous object-level tasks and excels in zero-shot transfer to new data and tasks, showing its exceptional versatility and generalization abilities. Additionally, GLEE provides general visual object-level information, which is currently missing in modern LLMs, establishing a robust foundation for object-centric mLLMs. + +# 6. Acknowledgement + +This research is supported by NSFC (No.62225603). + +# References + +[1] David Acuna, Huan Ling, Amlan Kar, and Sanja Fidler. Efficient interactive annotation of segmentation datasets with polygon-rnn++. 2018. 2 +[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35: 23716-23736, 2022. 2 +[3] Ali Athar, Jonathon Luiten, Paul Voigtlaender, Tarasha Khurana, Achal Dave, Bastian Leibe, and Deva Ramanan. Burst: A benchmark for unifying object recognition, segmentation and tracking in video. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1674-1683, 2023. 7 +[4] Philipp Bergmann, Tim Meinhardt, and Laura Leal-Taixe. Tracking without bells and whistles. In ICCV, 2019. 4 +[5] Philipp Bergmann, Tim Meinhardt, and Laura Leal-Taixe. Tracking without bells and whistles. In ICCV, 2019. 7 +[6] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In 2016 IEEE international conference on image processing (ICIP), pages 3464-3468. IEEE, 2016. 7 +[7] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. 1 +[8] Yuri Y Boykov and M-P Jolly. Interactive graph cuts for optimal boundary & region segmentation of objects in nd images. In Proceedings eighth IEEE international conference on computer vision. ICCV 2001, pages 105-112. IEEE, 2001. 5 +[9] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1, 2, 3 +[10] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 2 +[11] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2 +[12] Lluis Castrejon, Kaustav Kundu, Raquel Urtasun, and Sanja Fidler. Annotating object instances with a polygon-rnn. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5230-5238, 2017. 2, 5 +[13] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pre + +training from pixels. In International conference on machine learning, pages 1691-1703. PMLR, 2020. 2 +[14] Ting Chen, Saurabh Saxena, Lala Li, Tsung-Yi Lin, David J Fleet, and Geoffrey E Hinton. A unified sequence interface for vision tasks. Advances in Neural Information Processing Systems, 35:31333-31346, 2022. 3, 6 +[15] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 2, 4, 6 +[16] Ho Kei Cheng and Alexander G Schwing. Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In European Conference on Computer Vision, pages 640-658. Springer, 2022. 2, 7 +[17] Ho Kei Cheng, Yu-Wing Tai, and Chi-Keung Tang. Rethinking space-time networks with improved memory coverage for efficient video object segmentation. Advances in Neural Information Processing Systems, 34:11781-11794, 2021. 2 +[18] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 2 +[19] Achal Dave, Tarasha Khurana, Pavel Tokmakov, Cordelia Schmid, and Deva Ramanan. Tao: A large-scale benchmark for tracking any object. In ECCV, 2020. 7 +[20] Patrick Dendorfer, Aljosa Osep, Anton Milan, Konrad Schindler, Daniel Cremers, Ian Reid, Stefan Roth, and Laura Leal-Taixe. Motchallenge: A benchmark for single-camera multiple target tracking. International Journal of Computer Vision, 129(4):845-881, 2021. 2, 4 +[21] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 1, 2 +[22] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, Philip HS Torr, and Song Bai. Mose: A new dataset for video object segmentation in complex scenes. arXiv preprint arXiv:2302.01872, 2023. 5 +[23] Zi-Yi Dou, Aishwarya Kamath, Zhe Gan, Pengchuan Zhang, Jianfeng Wang, Linjie Li, Zicheng Liu, Ce Liu, Yann LeCun, Nanyun Peng, et al. Coarse-to-fine vision-language pre-training with fusion in the backbone. NeurIPS, 35:32942-32956, 2022. 8 +[24] Yuxin Fang, Quan Sun, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva-02: A visual representation for neon genesis. arXiv preprint arXiv:2303.11331, 2023.6 +[25] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19358-19369, 2023. 2 +[26] Chengjian Feng, Yujie Zhong, Zequn Jie, Xiangxiang Chu, Haibing Ren, Xiaolin Wei, Weidi Xie, and Lin Ma. Promptdet: Towards open-vocabulary detection using uncurated + +images. In Computer Vision - ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IX, page 701-717, 2022. 3 +[27] Golnaz Ghiasi, Barret Zoph, Ekin D Cubuk, Quoc V Le, and Tsung-Yi Lin. Multi-task self-training for learning general representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8856-8865, 2021. 3 +[28] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In International Conference on Learning Representations, 2022. 3 +[29] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 2, 4, 6 +[30] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 4, 6 +[31] Kaiming He, Georgia Gkioxari, Piotr Dollar, and Ross Girshick. Mask R-CNN. In ICCV, 2017. 2 +[32] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 1, 2 +[33] Miran Heo, Sukjun Hwang, Seoung Wug Oh, Joon-Young Lee, and Seon Joo Kim. Vita: Video instance segmentation via object token association. In Advances in Neural Information Processing Systems, 2022. 8 +[34] Miran Heo, Sukjun Hwang, Jeongseok Hyun, Hanjung Kim, Seoung Wug Oh, Joon-Young Lee, and Seon Joo Kim. A generalized framework for video instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14623-14632, 2023. 2, 4, 8 +[35] Ronghang Hu, Marcus Rohrbach, and Trevor Darrell. Segmentation from natural language expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 108-124. Springer, 2016. 2 +[36] De-An Huang, Zhiding Yu, and Anima Anandkumar. Minvis: A minimal video instance segmentation framework without video-based training. In NeurIPS, 2022. 5 +[37] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 2, 3 +[38] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In ICCV, pages 1780-1790, 2021. 6, 8 +[39] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, 2023. 1, 2, 6, 8 + +[40] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 2, 6 +[41] Weicheng Kuo, Yin Cui, Xiuye Gu, AJ Piergiovanni, and Anelia Angelova. Open-vocabulary object detection upon frozen vision and language models. In The Eleventh International Conference on Learning Representations, 2023. 3 +[42] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, Tom Duerig, and Vittorio Ferrari. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. IJCV, 2020. 2, 6 +[43] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692, 2023. 2, 8 +[44] Chunyuan Li, Haotian Liu, Lianian Li, Pengchuan Zhang, Jyoti Aneja, Jianwei Yang, Ping Jin, Houdong Hu, Zicheng Liu, Yong Jae Lee, et al. Elevater: A benchmark and toolkit for evaluating language-augmented visual models. In NeurlPS, 2022. 8 +[45] Feng Li, Hao Zhang, Huaizhe Xu, Shilong Liu, Lei Zhang, Lionel M Ni, and Heung-Yeung Shum. Mask dino: Towards a unified transformer-based framework for object detection and segmentation. In CVPR, 2023. 3, 4, 6 +[46] Hao Li, Jinguo Zhu, Xiaohu Jiang, Xizhou Zhu, Hongsheng Li, Chun Yuan, Xiaohua Wang, Yu Qiao, Xiaogang Wang, Wenhai Wang, et al. Uni-perceiver v2: A generalist model for large-scale vision and vision-language tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2691-2700, 2023. 3, 6 +[47] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In CVPR, pages 10965-10975, 2022. 3, 8 +[48] Siyuan Li, Martin Danelljan, Henghui Ding, Thomas E Huang, and Fisher Yu. Tracking every thing in the wild. In ECCV, 2022. 7 +[49] Siyuan Li, Tobias Fischer, Lei Ke, Henghui Ding, Martin Danelljan, and Fisher Yu. Ovtrack: Open-vocabulary multiple object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5567-5577, 2023. 7 +[50] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In European Conference on Computer Vision, pages 280-296. Springer, 2022. 4, 6 +[51] Chuang Lin, Peize Sun, Yi Jiang, Ping Luo, Lizhen Qu, Gholamreza Haffari, Zehuan Yuan, and Jianfei Cai. Learning object-language alignments for open-vocabulary object detection. In The Eleventh International Conference on Learning Representations, 2023. 3 + +[52] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 2, 4, 6 +[53] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In ICCV, 2017. 5 +[54] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 8 +[55] Jiang Liu, Hui Ding, Zhaowei Cai, Yuting Zhang, Ravi Kumar Satzoda, Vijay Mahadevan, and R Manmatha. Polyformer: Referring image segmentation as sequential polygon generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18653-18663, 2023. 2, 4, 6, 7 +[56] Qin Liu, Zhenlin Xu, Gedas Bertasius, and Marc Niethammer. Simpleclick: Interactive image segmentation with simple vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22290-22300, 2023. 5 +[57] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. arXiv preprint arXiv:2103.14030, 2021. 6 +[58] Jiasen Lu, Christopher Clark, Rowan Zellers, Roozbeh Mottaghi, and Aniruddha Kembhavi. Unified-io: A unified model for vision, language, and multi-modal tasks. arXiv preprint arXiv:2206.08916, 2022.3 +[59] Chuofan Ma, Yi Jiang, Xin Wen, Zehuan Yuan, and Xiaojuan Qi. Codet: Co-occurrence guided region-word alignment for open-vocabulary object detection. arXiv preprint arXiv:2310.16667, 2023. 3 +[60] Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. TrackFormer: Multi-object tracking with transformers. arXiv preprint arXiv:2101.02702, 2021. 2, 4 +[61] Tim Meinhardt, Matt Feiszli, Yuchen Fan, Laura Leal-Taixe, and Rakesh Ranjan. Novis: A case for end-to-end near-online video instance segmentation. arXiv preprint arXiv:2308.15266, 2023. 8 +[62] Fausto Miletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), 2016. 5 +[63] Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, Xiao Wang, Xiaohua Zhai, Thomas Kipf, and Neil Houlsby. Simple open-vocabulary object detection. In Computer Vision - ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part X, page 728-755, 2022. 3 +[64] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In ECCV, 2016. 2, 6, 7 +[65] Seoung Wug Oh, Joon-Young Lee, Ning Xu, and Seon Joo Kim. Video object segmentation using space-time memory + +networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9226-9235, 2019. 2 +[66] Jiangmiao Pang, Linlu Qiu, Xia Li, Haofeng Chen, Qi Li, Trevor Darrell, and Fisher Yu. Quasi-dense similarity learning for multiple object tracking. In CVPR, 2021. 7 +[67] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 2, 6, 8 +[68] Jiyang Qi, Yan Gao, Yao Hu, Xinggang Wang, Xiaoyu Liu, Xiang Bai, Serge Belongie, Alan Yuille, Philip HS Torr, and Song Bai. Occluded video instance segmentation: A benchmark. International Journal of Computer Vision, pages 1-18, 2022. 4, 5, 6, 8 +[69] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 3, 6 +[70] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485-5551, 2020. 1, 2, 3 +[71] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2 +[72] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2):3, 2022. 2 +[73] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In CVPR, 2019. 5 +[74] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2 +[75] Carsten Rother, Vladimir Kolmogorov, and Andrew Blake. "grabcut" interactive foreground extraction using iterated graph cuts. ACM transactions on graphics (TOG), 23(3): 309-314, 2004. 5 +[76] Samuel Schulter, Vijay Kumar B G, Yumin Suh, Konstantinos M. Dafnis, Zhixing Zhang, Shiyu Zhao, and Dimitris Metaxas. Omnilabel: A challenging benchmark for language-based object detection. In ICCV, 2023. 8 +[77] Seonguk Seo, Joon-Young Lee, and Bohyung Han. Urvos: Unified referring video object segmentation network with a large-scale benchmark. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XV 16, pages 208–223. Springer, 2020. 2, 4, 6 + +[78] Jing Shao, Siyu Chen, Yangguang Li, Kun Wang, Zhenfei Yin, Yinan He, Jianing Teng, Qinghong Sun, Mengya Gao, Jihao Liu, et al. Intern: A new learning paradigm towards general vision. arXiv preprint arXiv:2111.08687, 2021. 3 +[79] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019. 2, 4, 6 +[80] Konstantin Sofiuk, Ilya A Petrov, and Anton Konushin. Reviving iterative training with mask guidance for interactive segmentation. In 2022 IEEE International Conference on Image Processing (ICIP), pages 3141-3145. IEEE, 2022. 5 +[81] Peize Sun, Rufeng Zhang, Yi Jiang, Tao Kong, Chenfeng Xu, Wei Zhan, Masayoshi Tomizuka, Lei Li, Zehuan Yuan, Changhu Wang, et al. Sparse r-cnn: End-to-end object detection with learnable proposals. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14454-14463, 2021. 2 +[82] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2 +[83] Haochen Wang, Cilin Yan, Shuai Wang, Xiaolong Jiang, Xu Tang, Yao Hu, Weidi Xie, and Efstratios Gavves. Towards open-vocabulary video instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4057-4066, 2023. 7 +[84] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In International Conference on Machine Learning, pages 23318-23340. PMLR, 2022. 3, 6 +[85] Weiyao Wang, Matt Feiszli, Heng Wang, and Du Tran. Unidentified video objects: A benchmark for dense, open-world segmentation. In ICCV, 2021. 6 +[86] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, et al. Image as a foreign language: Beit pretraining for all vision and vision-language tasks. arXiv preprint arXiv:2208.10442, 2022. 2 +[87] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In CVPR, 2021. 2 +[88] Nicolai Wojke, Alex Bewley, and Dietrich Paulus. Simple online and realtime tracking with a deep association metric. In ICIP, 2017. 7 +[89] Jiajun Wu, Yibiao Zhao, Jun-Yan Zhu, Siwei Luo, and Zhuowen Tu. Milcut: A sweeping line multiple instance learning paradigm for interactive image segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 256-263, 2014. 5 + +[90] Junfeng Wu, Yi Jiang, Song Bai, Wenqing Zhang, and Xiang Bai. Seqformer: Sequential transformer for video instance segmentation. In ECCV, 2022. 2, 8 +[91] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4974-4984, 2022. 2, 4 +[92] Junfeng Wu, Qihao Liu, Yi Jiang, Song Bai, Alan Yuille, and Xiang Bai. In defense of online models for video instance segmentation. In ECCV, pages 588-605. Springer, 2022. 2, 4, 5, 8 +[93] Jiannan Wu, Yi Jiang, Bin Yan, Hutchuan Lu, Zehuan Yuan, and Ping Luo. Segment every reference object in spatial and temporal spaces. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2538-2550, 2023. 2 +[94] Bin Xiao, Haiping Wu, Weijian Xu, Xiyang Dai, Houdong Hu, Yumao Lu, Michael Zeng, Ce Liu, and Lu Yuan. Florence-2: Advancing a unified representation for a variety of vision tasks. arXiv preprint arXiv:2311.06242, 2023. 6 +[95] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2955-2966, 2023. 6, 7 +[96] Ning Xu, Linjie Yang, Jianchao Yang, Dingcheng Yue, Yuchen Fan, Yuchen Liang, and Thomas S. Huang. Youtuvis dataset 2021 version. https://youtu-vos.org/dataset/vis/.6 +[97] Ning Xu, Brian Price, Scott Cohen, Jimei Yang, and Thomas S Huang. Deep interactive object selection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 373-381, 2016. 5 +[98] Ning Xu, Linjie Yang, Yuchen Fan, Dingcheng Yue, Yuchen Liang, Jianchao Yang, and Thomas Huang. Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327, 2018. 2, 5 +[99] Bin Yan, Yi Jiang, Peize Sun, Dong Wang, Zehuan Yuan, Ping Luo, and Huchuan Lu. Towards grand unification of object tracking. In ECCV, 2022. 2 +[100] Bin Yan, Yi Jiang, Jiannan Wu, Dong Wang, Ping Luo, Zehuan Yuan, and Huchuan Lu. Universal instance perception as object discovery and retrieval. In CVPR, 2023. 3, 4, 6, 7, 8 +[101] Haosen Yang, Chuofan Ma, Bin Wen, Yi Jiang, Zehuan Yuan, and Xiatian Zhu. Recognize any regions. 2023. 3 +[102] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV, 2019. 2, 4, 6, 8 +[103] Zhengyuan Yang, Zhe Gan, Jianfeng Wang, Xiaowei Hu, Faisal Ahmed, Zicheng Liu, Yumao Lu, and Lijuan Wang. Unitab: Unifying text and box outputs for grounded vision-language modeling. In European Conference on Computer Vision, pages 521-539. Springer, 2022. 3, 6 +[104] Lewei Yao, Jianhua Han, Youpeng Wen, Xiaodan Liang, Dan Xu, Wei Zhang, Zhenguo Li, Chunjing Xu, and Hang + +Xu. Detclip: Dictionary-enriched visual-concept paralleled pre-training for open-world detection. In NeurIPS, 2022. 4 +[105] Lewei Yao, Jianhua Han, Xiaodan Liang, Dan Xu, Wei Zhang, Zhenguo Li, and Hang Xu. Detclipv2: Scalable open-vocabulary object detection pre-training via word-region alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23497-23506, 2023. 3 +[106] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2636-2645, 2020. 6 +[107] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. Transactions on Machine Learning Research, 2022. 3 +[108] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In ECCV, 2016. 2, 4, 6, 7 +[109] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 2, 3 +[110] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-vocabulary object detection using captions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14393-14402, 2021. 3 +[111] Haotian Zhang, Pengchuan Zhang, Xiaowei Hu, Yen-Chun Chen, Liunian Li, Xiyang Dai, Lijuan Wang, Lu Yuan, Jenq-Neng Hwang, and Jianfeng Gao. Glipv2: Unifying localization and vision-language understanding. In Advances in Neural Information Processing Systems, pages 36067-36080, 2022. 6 +[112] Tao Zhang, Xingye Tian, Yu Wu, Shunping Ji, Xuebo Wang, Yuan Zhang, and Pengfei Wan. Dvis: Decoupled video instance segmentation framework. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1282-1291, 2023. 8 +[113] Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. Byte-track: Multi-object tracking by associating every detection box. arXiv preprint arXiv:2110.06864, 2021. 2, 4 +[114] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Lianian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In CVPR, pages 16793-16803, 2022. 8 +[115] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Lianian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16793-16803, 2022. 3 + +[116] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Tracking objects as points. In ECCV, 2020. 2, 4 +[117] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In ECCV, pages 350-368. Springer, 2022. 7, 8 +[118] Chaoyang Zhu, Yiyi Zhou, Yunhang Shen, Gen Luo, Xingjia Pan, Mingbao Lin, Chao Chen, Liujuan Cao, Xiaoshuai Sun, and Rongrong Ji. Seqtr: A simple yet universal network for visual grounding. In European Conference on Computer Vision, pages 598-615. Springer, 2022. 2, 4, 6 +[119] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR, 2021. 2, 3 +[120] Xizhou Zhu, Jinguo Zhu, Hao Li, Xiaoshi Wu, Hongsheng Li, Xiaohua Wang, and Jifeng Dai. Uni-perceiver: Pretraining unified architecture for generic perception for zero-shot and few-shot tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16804-16815, 2022. 3 +[121] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15116-15127, 2023. 3, 6 +[122] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. In NeurIPS, 2023. 2, 3 \ No newline at end of file diff --git a/2024/General Object Foundation Model for Images and Videos at Scale/images.zip b/2024/General Object Foundation Model for Images and Videos at Scale/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..61047bc44f8316fedbbd1d6c30817343a598abf4 --- /dev/null +++ b/2024/General Object Foundation Model for Images and Videos at Scale/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:327bc09d4dd774f0148a12e24da469d1fd972e9df48b5e24f7f6f8f879e41832 +size 644799 diff --git a/2024/General Object Foundation Model for Images and Videos at Scale/layout.json b/2024/General Object Foundation Model for Images and Videos at Scale/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0f874562e0d2b04654459c4ce84a15a20b5e8654 --- /dev/null +++ b/2024/General Object Foundation Model for Images and Videos at Scale/layout.json @@ -0,0 +1,9461 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 94, + 103, + 501, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 103, + 501, + 121 + ], + "spans": [ + { + "bbox": [ + 94, + 103, + 501, + 121 + ], + "type": "text", + "content": "General Object Foundation Model for Images and Videos at Scale" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 94, + 142, + 499, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 142, + 499, + 172 + ], + "spans": [ + { + "bbox": [ + 94, + 142, + 499, + 172 + ], + "type": "text", + "content": "Junfeng Wu" + }, + { + "bbox": [ + 94, + 142, + 499, + 172 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 94, + 142, + 499, + 172 + ], + "type": "text", + "content": ", Yi Jiang" + }, + { + "bbox": [ + 94, + 142, + 499, + 172 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 94, + 142, + 499, + 172 + ], + "type": "text", + "content": ", Qihao Liu" + }, + { + "bbox": [ + 94, + 142, + 499, + 172 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 94, + 142, + 499, + 172 + ], + "type": "text", + "content": ", Zehuan Yuan" + }, + { + "bbox": [ + 94, + 142, + 499, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 94, + 142, + 499, + 172 + ], + "type": "text", + "content": ", Xiang Bai" + }, + { + "bbox": [ + 94, + 142, + 499, + 172 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 94, + 142, + 499, + 172 + ], + "type": "text", + "content": ", Song Bai" + }, + { + "bbox": [ + 94, + 142, + 499, + 172 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 94, + 142, + 499, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 94, + 142, + 499, + 172 + ], + "type": "text", + "content": "Huazhong University of Science and Technology, " + }, + { + "bbox": [ + 94, + 142, + 499, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 94, + 142, + 499, + 172 + ], + "type": "text", + "content": "ByteDance Inc., " + }, + { + "bbox": [ + 94, + 142, + 499, + 172 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 94, + 142, + 499, + 172 + ], + "type": "text", + "content": "Johns Hopkins University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 199, + 192, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 199, + 192, + 213 + ], + "spans": [ + { + "bbox": [ + 143, + 199, + 192, + 213 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 224, + 290, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 224, + 290, + 559 + ], + "spans": [ + { + "bbox": [ + 47, + 224, + 290, + 559 + ], + "type": "text", + "content": "We present GLEE in this work, an object-level foundation model for locating and identifying objects in images and videos. Through a unified framework, GLEE accomplishes detection, segmentation, tracking, grounding, and identification of arbitrary objects in the open world scenario for various object perception tasks. Adopting a cohesive learning strategy, GLEE acquires knowledge from diverse data sources with varying supervision levels to formulate general object representations, excelling in zero-shot transfer to new data and tasks. Specifically, we employ an image encoder, text encoder, and visual prompter to handle multi-modal inputs, enabling to simultaneously solve various object-centric downstream tasks while maintaining state-of-the-art performance. Demonstrated through extensive training on over five million images from diverse benchmarks, GLEE exhibits remarkable versatility and improved generalization performance, efficiently tackling downstream tasks without the need for task-specific adaptation. By integrating large volumes of automatically labeled data, we further enhance its zero-shot generalization capabilities. Additionally, GLEE is capable of being integrated into Large Language Models, serving as a foundational model to provide universal object-level information for multi-modal tasks. We hope that the versatility and universality of our method will mark a significant step in the development of efficient visual foundation models for AGI systems. The models and code are released at https://github.com/FoundationVision/GLEE." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 582, + 128, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 582, + 128, + 596 + ], + "spans": [ + { + "bbox": [ + 47, + 582, + 128, + 596 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 603, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 603, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 603, + 287, + 676 + ], + "type": "text", + "content": "Foundation model [7] is an emerging paradigm for building artificial general intelligence (AGI) systems, signifying a model trained on broad data that is capable of being adapted to a wide range of downstream tasks in an general paradigm. Recently, NLP foundation models such as BERT [21], GPT-3 [9], T5 [70] developed with unified" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 315, + 198, + 536, + 350 + ], + "blocks": [ + { + "bbox": [ + 315, + 198, + 536, + 350 + ], + "lines": [ + { + "bbox": [ + 315, + 198, + 536, + 350 + ], + "spans": [ + { + "bbox": [ + 315, + 198, + 536, + 350 + ], + "type": "image", + "image_path": "d775f818d6f3630be9dc6f16488fa0039086e3fee565a2cb231ebaa038f24436.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 360, + 545, + 383 + ], + "lines": [ + { + "bbox": [ + 306, + 360, + 545, + 383 + ], + "spans": [ + { + "bbox": [ + 306, + 360, + 545, + 383 + ], + "type": "text", + "content": "Figure 1. The performance of GLEE on a broad range of object-level tasks compared with existing models." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 399, + 545, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 399, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 305, + 399, + 545, + 434 + ], + "type": "text", + "content": "input-output paradigms and large-scale pre-training, have achieved remarkable generalization capabilities to address nearly all NLP tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 436, + 546, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 436, + 546, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 436, + 546, + 604 + ], + "type": "text", + "content": "In computer vision, the diversity of task types and the lack of a unified form make visual foundation models only serve specific subdomains, such as CLIP [69] for multimodal visual model, MAE [32] for visual representations model, SAM [39] for segmentation model. Despite being widely studied, current visual foundation models are still focusing on establishing correlations between global image features and language descriptions or learning image-level feature representations. However, locating and identifying objects constitute foundational capabilities in computer vision systems, serves as a basis for solving complex or high level vision tasks such as segmentation, scene understanding, object tracking, event detection, and activity recognition and support a wide range of applications." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "In this work, we advance the development of object-level foundation models within the visual domain. To address the aforementioned limitation, providing general and accurate object-level information, we introduce a general object visual foundation model, coined as GLEE, which simultaneously solve a wide range of object-centric tasks while ensuring SOTA performance, including object detection, instance segmentation, grounding, object tracking, interactive segmentation and tracking, etc., as shown in Figure 1. Through" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 683, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 683, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 683, + 287, + 713 + ], + "type": "text", + "content": "*Equal Contribution. This work was done when Junfeng Wu and Qihao Liu were interns at ByteDance. †Correspondence to Xiang Bai and Song Bai ." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3783" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 178 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 178 + ], + "type": "text", + "content": "a unified input and output paradigm definition, our model is capable of learning from a wide range of diverse data and predicting general object representations, which makes it generalize well to new data and tasks in a zero-shot manner and achieve amazing performance. In addition, thanks to the unified paradigm, the training data can be scaled up at low cost by introducing a large amount of automatically labeled data, and further improve the zero-shot generalization ability of the model." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 182, + 288, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 182, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 288, + 373 + ], + "type": "text", + "content": "A general object foundation model framework. Our objective is to build an object visual foundation model capable of simultaneously addressing a wide range of object-centric tasks. Specifically, we employ an image encoder, a text encoder, and a visual prompter to encode multi-modal inputs. They are integrated into a detector to extract objects from images according to textual and visual input. This unified approach to handle multiple modalities enables us to concurrently solve various object-centric tasks, including detection [10, 52, 81, 119], instance segmentation [15, 31], referring expression comprehension [35, 55, 93, 118], interactive segmentation [1, 12, 122], multi-object tracking [20, 60, 99, 113, 116], video object segmentation [16, 17, 65, 98], video instance segmentation [34, 87, 90, 92, 102], and video referring segmentation [77, 91, 93], all while maintaining state-of-the-art performance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 376, + 288, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 288, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 288, + 628 + ], + "type": "text", + "content": "Multi-granularity joint supervision and scalable training paradigm. The design of the unified framework capable of addressing multiple tasks enables joint training on over five million images from diverse benchmarks and varying levels of supervision. Existing datasets differ in annotation granularity: detection datasets like Objects365 [79] and OpenImages [42] offer bounding boxes and category names; COCO [52] and LVIS [29] provide finer-grained mask annotations; RefCOCO [64, 108] and Visual Genome [40] provide detailed object descriptions. Additionally, video data enhance the temporal consistency of model, while open-world data contribute class-agnostic object annotations. An intuitive display of the supervision types and data scales of the datasets employed is presented in Figure 2. The unified support for multi-source data in our approach greatly facilitates the incorporation of additional manually or automatically annotated data, enabling easy scaling up of the dataset. Furthermore, the alignment of model optimization across tasks means that joint training serves not only as a unifying strategy but also as a mechanism to boost performance across individual tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": "Strong zero-shot transferability to a wide range of object level image and video tasks. After joint training on data from diverse sources, GLEE demonstrates remarkable versatility and zero-shot generalization abilities. Extensive experiments demonstrate that GLEE achieves state-of-the-art performance compared to existing specialist and generalist models in object-level image tasks such as detec" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 289 + ], + "type": "text", + "content": "tion, referring expression comprehension, and open-world detection, all without requiring any task-specific designs or fine-tuning. Furthermore, we showcase the extraordinary generalization and zero-shot capabilities of GLEE in large-vocabulary open-world video tracking tasks, achieving significantly superior performance over existing models even in a zero-shot transfer manner. Additionally, by incorporating automatically annotated data like SA1B [39] and GRIT [67], we are able to scale up our training dataset to an impressive size of 10 million images at a low cost, which is typically challenging to achieve for object-level tasks and further enhances the generalization performance. Moreover, we replace the SAM [39] component with GLEE in a multimodal Large Language Model (mLLM) [43] and observe that it achieves comparable results. This demonstrates that GLEE is capable of supplying the visual object-level information that modern LLMs currently lack, thus laying a solid foundation for an object-centric mLLMs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 301, + 392, + 314 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 301, + 392, + 314 + ], + "spans": [ + { + "bbox": [ + 306, + 301, + 392, + 314 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 323, + 448, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 323, + 448, + 335 + ], + "spans": [ + { + "bbox": [ + 306, + 323, + 448, + 335 + ], + "type": "text", + "content": "2.1. Visual Foundation Model" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 342, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 547, + 715 + ], + "type": "text", + "content": "As foundation models [9, 18, 21, 70, 82] in the NLP field have achieved remarkable success, the construction of visual foundation models attracts increasing attention. Unlike NLP tasks that are predominantly unified under a text-to-text paradigm, tasks in Computer Vision still exhibit significant differences in formulation. This disparity leads to the fact that visual models are typically trained in a single-task learning frameworks, limiting their applicability to tasks within certain sub-domains. For instance, multi-modal visual foundation models like CLIP [69], ALIGN [37], Florence [109], BEIT3 [86], Flamingo[2] make significant advancements in efficient transfer learning and demonstrate impressive zero-shot capabilities on vision-language tasks by employing contrastive learning and masked data modeling on large-scale image-text pairs. DALL-E [71, 72] and Stable Diffusion [74] are trained on massive pairs of images and captions, enabling them to generate detailed image content conditioned on textual instruction. DINO [11], MAE [32], EVA [25], ImageGPT [13] obtain strong visual representations through self-supervised training on largescale image data, which are then employed to adopt downstream tasks. These foundation models learn image-level features and are not directly applicable to object-level tasks. The recently proposed SAM [39], capable of segmenting any object of a given image based on visual prompt such as points and boxes, provides rich object-level information and demonstrates strong generalization capabilities. However, the object information lacks semantic context, limiting its application in object-level tasks. Unlike existing visual foundation models, we aim to develop an object foundation model that directly solve downstream tasks without the need" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3784" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 70, + 531, + 361 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 531, + 361 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 531, + 361 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 531, + 361 + ], + "type": "image", + "image_path": "ec3537dc2cf9d5bae65c56ad656596c2164510ac2cca85a3b8aa8f6dcbb89c8b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 363, + 546, + 385 + ], + "lines": [ + { + "bbox": [ + 46, + 363, + 546, + 385 + ], + "spans": [ + { + "bbox": [ + 46, + 363, + 546, + 385 + ], + "type": "text", + "content": "Figure 2. An illustrative example showcasing annotations of varying granularities from different datasets, along with the scale of data we utilized. Training on datasets from multiple sources endows the model with more universal representations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 407, + 211, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 211, + 419 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 211, + 419 + ], + "type": "text", + "content": "for additional parameters or fine-tuning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 430, + 197, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 430, + 197, + 442 + ], + "spans": [ + { + "bbox": [ + 47, + 430, + 197, + 442 + ], + "type": "text", + "content": "2.2. Unified and General Model" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "type": "text", + "content": "Unified models share similarities with foundation models in the aspect of multi-task unification for their ability to handle multiple vision or multi-modal tasks within a single model. MuST [27] and Intern [78] propose to train across multiple vision tasks and solving them simultaneously. Inspired by the success of sequence-to-sequence NLP models [9, 70], models such as Uni-Perceiver [120], OFA [84], Unified-IO [58], Pix2Seq v2 [14], and UniTAB [103] propose modeling various tasks as sequence generation tasks within a unified paradigm. While these approaches have demonstrated promising cross-task generalization capabilities, they focus mainly on image-level understanding tasks. In addition, their auto-regressive generation of boxes and masks results in significantly slower inference speeds and the performance still falls short of state-of-the-art task-specific models. Building upon on detectors [45, 119], UniPerceiver v2 [46] and UNINEXT [100] utilize unified maximum likelihood estimation and object retrieval to support various tasks, effectively resolves the challenges of localization. Nonetheless, they are trained on closed-set data, thereby not exhibiting zero-shot generalization capabilities. X-decoder [121] and SEEM [122] construct a generalized" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 407, + 547, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 407, + 547, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 407, + 547, + 491 + ], + "type": "text", + "content": "decoding model capable of predicting pixel-level segmentation and language tokens. Diverging from unified models, the proposed GLEE not only directly addresses object-level tasks in a unified manner but also provides universal object representations, which generalize well to new data and tasks, serving as a cornerstone for a broader range of tasks that require detailed object information." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 502, + 481, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 502, + 481, + 515 + ], + "spans": [ + { + "bbox": [ + 305, + 502, + 481, + 515 + ], + "type": "text", + "content": "2.3. Vision-Language Understanding" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "text", + "content": "Open-vocabulary detection (OVD) and Grounding models both necessitate the localization and recognition of as many objects as possible. With the recent advancements in vision-language pre-training [37, 69, 107, 109], a commonly employed strategy for OVD involves transferring the knowledge from pre-trained vision-language models (VLMs) to object detectors [28, 41, 63]. Another group of studies leverages extensive image-text pair datasets to broaden the detection vocabulary [26, 47, 51, 59, 101, 105, 110, 115]. However, these language-based detectors are inherently constrained by the capabilities and biases of language models, making it challenging to excel simultaneously in both localization and recognition. Our objective is to optimally utilize existing datasets to construct a general object-level foundation model, aims to not only detect and identify objects effectively but also to offer universal object representations." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3785" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 232, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 232, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 232, + 85 + ], + "type": "text", + "content": "tations for a wide range of downstream tasks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 95, + 104, + 107 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 95, + 104, + 107 + ], + "spans": [ + { + "bbox": [ + 47, + 95, + 104, + 107 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 115, + 129, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 115, + 129, + 126 + ], + "spans": [ + { + "bbox": [ + 47, + 115, + 129, + 126 + ], + "type": "text", + "content": "3.1. Formulation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 133, + 289, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 133, + 289, + 434 + ], + "spans": [ + { + "bbox": [ + 47, + 133, + 289, + 434 + ], + "type": "text", + "content": "The proposed GLEE consists of an image encoder, a text encoder, a visual prompter, and an object decoder, as illustrated in Figure 3. The text encoder processes arbitrary descriptions related to the task, including object categories, names in any form, captions about objects, and referring expressions. The visual prompter encodes user inputs such as points, bounding boxes, or scribbles during interactive segmentation into corresponding visual representations of target objects. Then they are integrated into a detector for extracting objects from images according to textual and visual input. We build the object decoder upon MaskDINO [45] with a dynamic class head by compute similarity between object embedding from detector and text embedding from the text encoder. Given an input image " + }, + { + "bbox": [ + 47, + 133, + 289, + 434 + ], + "type": "inline_equation", + "content": "I \\in \\mathcal{R}^{3 \\times H \\times W}" + }, + { + "bbox": [ + 47, + 133, + 289, + 434 + ], + "type": "text", + "content": ", we first extract multi-scale features " + }, + { + "bbox": [ + 47, + 133, + 289, + 434 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 47, + 133, + 289, + 434 + ], + "type": "text", + "content": " with backbones such as ResNet [30]. Then we feed them into the object decoder and adopt three prediction heads (classification, detection, and segmentation) on the output embedding " + }, + { + "bbox": [ + 47, + 133, + 289, + 434 + ], + "type": "inline_equation", + "content": "q_{d} \\in \\mathcal{R}^{N \\times C}" + }, + { + "bbox": [ + 47, + 133, + 289, + 434 + ], + "type": "text", + "content": " from decoder. Following other object segmentation models [15, 45, 50], we construct a 1/4 resolution pixel embedding map " + }, + { + "bbox": [ + 47, + 133, + 289, + 434 + ], + "type": "inline_equation", + "content": "M_{p} \\in \\mathcal{R}^{C \\times \\frac{H}{4} \\times \\frac{W}{4}}" + }, + { + "bbox": [ + 47, + 133, + 289, + 434 + ], + "type": "text", + "content": " which is obtained by upsampling and fusing multi-scale feature maps from the backbone and Transformer encoder. Finally, we obtain each binary mask prediction " + }, + { + "bbox": [ + 47, + 133, + 289, + 434 + ], + "type": "inline_equation", + "content": "m \\in \\mathcal{R}^{N \\times \\frac{H}{4} \\times \\frac{W}{4}}" + }, + { + "bbox": [ + 47, + 133, + 289, + 434 + ], + "type": "text", + "content": " via a dot product between the N mask embeddings and pixel embedding map:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 118, + 455, + 287, + 469 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 455, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 118, + 455, + 287, + 469 + ], + "type": "interline_equation", + "content": "m = F F N \\left(q _ {d}\\right) \\otimes M _ {p}, \\tag {1}", + "image_path": "bb461210beab2987efcecac77b38ffecfa66e9285321ed04b1a24e04a4ff2859.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 476, + 287, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 476, + 287, + 501 + ], + "spans": [ + { + "bbox": [ + 47, + 476, + 287, + 501 + ], + "type": "text", + "content": "where FFN is a 3-layer feed forward head with ReLU activation function and a linear projection layer." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 501, + 287, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 501, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 501, + 287, + 597 + ], + "type": "text", + "content": "To support arbitrary vocabularies and object descriptions, we replace the FFN classifier with text embeddings following DetCLIP [104]. Specifically, we feed K category names as separate sentences into the text encoder " + }, + { + "bbox": [ + 47, + 501, + 287, + 597 + ], + "type": "inline_equation", + "content": "Enc_{L}" + }, + { + "bbox": [ + 47, + 501, + 287, + 597 + ], + "type": "text", + "content": " and use the average of each sentence tokens as the output text embedding " + }, + { + "bbox": [ + 47, + 501, + 287, + 597 + ], + "type": "inline_equation", + "content": "e_t \\in \\mathcal{R}^{K \\times D}" + }, + { + "bbox": [ + 47, + 501, + 287, + 597 + ], + "type": "text", + "content": " for each category or description. Then we compute the alignment scores " + }, + { + "bbox": [ + 47, + 501, + 287, + 597 + ], + "type": "inline_equation", + "content": "S_{\\text{align}} \\in \\mathcal{R}^{N \\times K}" + }, + { + "bbox": [ + 47, + 501, + 287, + 597 + ], + "type": "text", + "content": " between object embedding and text embedding:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 607, + 287, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 607, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 116, + 607, + 287, + 620 + ], + "type": "interline_equation", + "content": "S _ {\\text {a l i g n}} = q _ {d} \\cdot W _ {i 2 t} \\otimes e _ {t}, \\tag {2}", + "image_path": "ad79b7bdafd042a42906ac0dd726a513914a03e1a017d1b20ef5e7746cad4c1c.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 629, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 629, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 629, + 288, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 629, + 288, + 714 + ], + "type": "inline_equation", + "content": "W_{i2t} \\in \\mathcal{R}^{C \\times D}" + }, + { + "bbox": [ + 47, + 629, + 288, + 714 + ], + "type": "text", + "content": " is image-to-text projection weights. We use logits " + }, + { + "bbox": [ + 47, + 629, + 288, + 714 + ], + "type": "inline_equation", + "content": "S_{\\text{align}}" + }, + { + "bbox": [ + 47, + 629, + 288, + 714 + ], + "type": "text", + "content": " to replace traditional classification logits to compute Hungarian matching cost during training and assign categories to objects during inference. To make the original visual features prompt-aware, an early fusion module is adopted before Transformer encoder following UNNEXT [100], which takes image feature from backbone" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": "and prompt embedding as input and perform bi-directional cross-attention between them." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 106, + 406, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 106, + 406, + 118 + ], + "spans": [ + { + "bbox": [ + 306, + 106, + 406, + 118 + ], + "type": "text", + "content": "3.2. Task Unification" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 125, + 545, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 125, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 304, + 125, + 545, + 220 + ], + "type": "text", + "content": "Based on the above designs, GLEE can be used to seamlessly unify a wide range of object perception tasks in images and videos, including object detection, instance segmentation, grounding, multi-target tracking (MOT), video instance segmentation (VIS), video object segmentation (VOS), interactive segmentation and tracking, and supports open-world/large-vocabulary image and video detection and segmentation tasks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 221, + 546, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 221, + 546, + 400 + ], + "spans": [ + { + "bbox": [ + 304, + 221, + 546, + 400 + ], + "type": "text", + "content": "Detection and Instance Segmentation. For detection task, a fixed-length category list is given and all objects in the category list are required to be detected. For a dataset with category list length " + }, + { + "bbox": [ + 304, + 221, + 546, + 400 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 221, + 546, + 400 + ], + "type": "text", + "content": ", the text input can be formulated as " + }, + { + "bbox": [ + 304, + 221, + 546, + 400 + ], + "type": "inline_equation", + "content": "\\{p_k\\}_{k=1}^K" + }, + { + "bbox": [ + 304, + 221, + 546, + 400 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 221, + 546, + 400 + ], + "type": "inline_equation", + "content": "p_k" + }, + { + "bbox": [ + 304, + 221, + 546, + 400 + ], + "type": "text", + "content": " represents for the " + }, + { + "bbox": [ + 304, + 221, + 546, + 400 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 221, + 546, + 400 + ], + "type": "text", + "content": "-th category name, e.g., " + }, + { + "bbox": [ + 304, + 221, + 546, + 400 + ], + "type": "inline_equation", + "content": "P = [\"person\", \"bicycle\", \"car\", ..., \"toothbrush\"]" + }, + { + "bbox": [ + 304, + 221, + 546, + 400 + ], + "type": "text", + "content": " for COCO [52]. For datasets with large vocabulary, calculating the text embedding of all categories is very time-consuming and redundant. Therefore, for datasets with a category number greater than 100, such as objects365 [79] and LVIS [29], suppose there are " + }, + { + "bbox": [ + 304, + 221, + 546, + 400 + ], + "type": "inline_equation", + "content": "\\hat{K}" + }, + { + "bbox": [ + 304, + 221, + 546, + 400 + ], + "type": "text", + "content": " positive categories in an image, we take the " + }, + { + "bbox": [ + 304, + 221, + 546, + 400 + ], + "type": "inline_equation", + "content": "\\hat{K}" + }, + { + "bbox": [ + 304, + 221, + 546, + 400 + ], + "type": "text", + "content": " positive categories and then pad the category number to 100 by randomly sampling from the negative categories. For instance segmentation, we enable the mask branch and add mask matching cost with mask loss." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 402, + 547, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 402, + 547, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 547, + 568 + ], + "type": "text", + "content": "Grounding and Referring Segmentation. These tasks provide reference textual expressions, where objects are described with attributes, for example, Referring Expression Comprehension (REC) [108, 118], Referring Expression Segmentation (RES) [55, 108], and Referring Video Object Segmentation (R-VOS) [77, 91] aim at finding objects matched with the given language expressions like \"The fourth person from the left\". For each image, we take the all the object expressions as text prompt and feed the them into the text encoder. For each expressions, we apply global average pooling along the sequence dimension to get text embedding " + }, + { + "bbox": [ + 304, + 402, + 547, + 568 + ], + "type": "inline_equation", + "content": "e_t" + }, + { + "bbox": [ + 304, + 402, + 547, + 568 + ], + "type": "text", + "content": ". The text embeddings are feed into early fusion module and additionally interact with object queries through self-attention module in the decoder." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": "MOT and VIS. Both Multi-object Tracking (MOT)[4, 20, 60, 113, 116] and Video Instance Segmentation (VIS)[34, 68, 92, 102] need to detect and track all the objects in the predefined category list, and VIS requires additional mask for the objects. These two tasks can be considered as extended tasks of detection and instance segmentation on videos. We found that with sufficient image exposure, object embeddings from the decoder effectively distinguish objects in a video, showing strong discriminability and temporal consistency. As a result, they can be directly employed for tracking without the need for an additional tracking head. Training on image-level data can address" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "3786" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 78, + 76, + 199, + 86 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 76, + 199, + 86 + ], + "spans": [ + { + "bbox": [ + 78, + 76, + 199, + 86 + ], + "type": "text", + "content": "category: [person, car, motorcycle, dog ...]" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 83, + 89, + 192, + 98 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 89, + 192, + 98 + ], + "spans": [ + { + "bbox": [ + 83, + 89, + 192, + 98 + ], + "type": "text", + "content": "arbitrary name: bollard, manhole cover" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 83, + 102, + 206, + 110 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 102, + 206, + 110 + ], + "spans": [ + { + "bbox": [ + 83, + 102, + 206, + 110 + ], + "type": "text", + "content": "expression: motorcycle parked under the sign" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 83, + 113, + 200, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 113, + 200, + 120 + ], + "spans": [ + { + "bbox": [ + 83, + 113, + 200, + 120 + ], + "type": "text", + "content": "object caption: air conditioner outdoor unit" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 113, + 121, + 176, + 167 + ], + "blocks": [ + { + "bbox": [ + 113, + 121, + 176, + 167 + ], + "lines": [ + { + "bbox": [ + 113, + 121, + 176, + 167 + ], + "spans": [ + { + "bbox": [ + 113, + 121, + 176, + 167 + ], + "type": "image", + "image_path": "d77eac373c21a127a77e5f9ed1534ad09fedd8b60ecf6ae60d3aed317cd2d1d6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 83, + 167, + 116, + 193 + ], + "blocks": [ + { + "bbox": [ + 83, + 167, + 116, + 193 + ], + "lines": [ + { + "bbox": [ + 83, + 167, + 116, + 193 + ], + "spans": [ + { + "bbox": [ + 83, + 167, + 116, + 193 + ], + "type": "image", + "image_path": "fd5bf6f7d382a23c2b209900ad9288d278edfa4a146ecd6445141003592729fa.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 123, + 168, + 149, + 192 + ], + "blocks": [ + { + "bbox": [ + 123, + 168, + 149, + 192 + ], + "lines": [ + { + "bbox": [ + 123, + 168, + 149, + 192 + ], + "spans": [ + { + "bbox": [ + 123, + 168, + 149, + 192 + ], + "type": "image", + "image_path": "6679f185efcc4c420e357c2b3a550885b414f1f11be178a3d04d42fd6cbf0546.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 156, + 168, + 190, + 193 + ], + "blocks": [ + { + "bbox": [ + 156, + 168, + 190, + 193 + ], + "lines": [ + { + "bbox": [ + 156, + 168, + 190, + 193 + ], + "spans": [ + { + "bbox": [ + 156, + 168, + 190, + 193 + ], + "type": "image", + "image_path": "3504d38790bbdfc686d18d7c87b5544f69d5c75ca375cddaf498c1bd6cbde8b3.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 211, + 77, + 500, + 193 + ], + "blocks": [ + { + "bbox": [ + 211, + 77, + 500, + 193 + ], + "lines": [ + { + "bbox": [ + 211, + 77, + 500, + 193 + ], + "spans": [ + { + "bbox": [ + 211, + 77, + 500, + 193 + ], + "type": "image", + "image_path": "825e83f34e9b6d73c922fd1650abbdafe1b6a28bec1a6d8ced44ae1560a4f802.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 225, + 194, + 356, + 205 + ], + "lines": [ + { + "bbox": [ + 225, + 194, + 356, + 205 + ], + "spans": [ + { + "bbox": [ + 225, + 194, + 356, + 205 + ], + "type": "text", + "content": "(a) Object foundation model framework" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 72, + 211, + 135, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 211, + 135, + 228 + ], + "spans": [ + { + "bbox": [ + 72, + 211, + 135, + 228 + ], + "type": "text", + "content": "category: person, car motorcycle,dog, cat." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 66, + 229, + 107, + 275 + ], + "blocks": [ + { + "bbox": [ + 66, + 229, + 107, + 275 + ], + "lines": [ + { + "bbox": [ + 66, + 229, + 107, + 275 + ], + "spans": [ + { + "bbox": [ + 66, + 229, + 107, + 275 + ], + "type": "image", + "image_path": "f4678e4883caeb60898c1376f7bf588a0ea878e5aead4d427a18537bbe48d4dc.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 299, + 545, + 354 + ], + "lines": [ + { + "bbox": [ + 46, + 299, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 299, + 545, + 354 + ], + "type": "text", + "content": "Figure 3. Framework of GLEE. The text encoder accepts textual descriptions in various forms from diverse data sources, including object categories, names, captions, and referring expressions. The visual prompt encodes points, bounding boxes, or scribbles into corresponding visual representations. The object decoder takes them and image features to predict objects in images. (b) illustrates the application of GLEE to image tasks tailored for different language descriptions and visual prompts. (c) demonstrates the application across various object-level video tasks." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 107, + 229, + 143, + 275 + ], + "blocks": [ + { + "bbox": [ + 107, + 229, + 143, + 275 + ], + "lines": [ + { + "bbox": [ + 107, + 229, + 143, + 275 + ], + "spans": [ + { + "bbox": [ + 107, + 229, + 143, + 275 + ], + "type": "image", + "image_path": "a13df688cdf1ec2d485edd19cd26c30d065014904c61b37d09462586d44b3d1c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 148, + 212, + 192, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 212, + 192, + 221 + ], + "spans": [ + { + "bbox": [ + 148, + 212, + 192, + 221 + ], + "type": "text", + "content": "arbitrary name" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 148, + 222, + 171, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 222, + 171, + 231 + ], + "spans": [ + { + "bbox": [ + 148, + 222, + 171, + 231 + ], + "type": "text", + "content": "bollard" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 184, + 222, + 209, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 222, + 209, + 234 + ], + "spans": [ + { + "bbox": [ + 184, + 222, + 209, + 234 + ], + "type": "text", + "content": "manhole cover" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 146, + 234, + 211, + 275 + ], + "blocks": [ + { + "bbox": [ + 146, + 234, + 211, + 275 + ], + "lines": [ + { + "bbox": [ + 146, + 234, + 211, + 275 + ], + "spans": [ + { + "bbox": [ + 146, + 234, + 211, + 275 + ], + "type": "image", + "image_path": "a9fa2576b35dd255726e944333ef5f7b8fff289815599f3faa3fda15a177bb4c.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 279, + 241, + 289 + ], + "lines": [ + { + "bbox": [ + 153, + 279, + 241, + 289 + ], + "spans": [ + { + "bbox": [ + 153, + 279, + 241, + 289 + ], + "type": "text", + "content": "(b) Applied to image tasks" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 219, + 211, + 277, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 211, + 277, + 234 + ], + "spans": [ + { + "bbox": [ + 219, + 211, + 277, + 234 + ], + "type": "text", + "content": "expression: motorcycle parked under the sign" + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 220, + 234, + 264, + 277 + ], + "blocks": [ + { + "bbox": [ + 220, + 234, + 264, + 277 + ], + "lines": [ + { + "bbox": [ + 220, + 234, + 264, + 277 + ], + "spans": [ + { + "bbox": [ + 220, + 234, + 264, + 277 + ], + "type": "image", + "image_path": "09d3d65d6993c2444d0a26814dc06a76a9de79e59fc7d90e24780507f945bdd2.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 280, + 211, + 313, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 280, + 211, + 313, + 234 + ], + "spans": [ + { + "bbox": [ + 280, + 211, + 313, + 234 + ], + "type": "text", + "content": "interactive segment" + } + ] + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 271, + 234, + 318, + 279 + ], + "blocks": [ + { + "bbox": [ + 271, + 234, + 318, + 279 + ], + "lines": [ + { + "bbox": [ + 271, + 234, + 318, + 279 + ], + "spans": [ + { + "bbox": [ + 271, + 234, + 318, + 279 + ], + "type": "image", + "image_path": "0cd06ba59cf65c2ef56bf0a031d4d5d0b7d037f876d0b73a70fae36756d742bd.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 323, + 212, + 447, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 212, + 447, + 231 + ], + "spans": [ + { + "bbox": [ + 323, + 212, + 447, + 231 + ], + "type": "text", + "content": "video tasks: VIS、MOT、VOS、RVOS、 open world/vocabulary tracking" + } + ] + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 324, + 232, + 443, + 270 + ], + "blocks": [ + { + "bbox": [ + 324, + 232, + 443, + 270 + ], + "lines": [ + { + "bbox": [ + 324, + 232, + 443, + 270 + ], + "spans": [ + { + "bbox": [ + 324, + 232, + 443, + 270 + ], + "type": "image", + "image_path": "b0c0fbdb3d3df8d8b94d5418e57f67229ee00342958772a790951c1c1eec0230.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 388, + 279, + 473, + 289 + ], + "lines": [ + { + "bbox": [ + 388, + 279, + 473, + 289 + ], + "spans": [ + { + "bbox": [ + 388, + 279, + 473, + 289 + ], + "type": "text", + "content": "(c) Applied to video tasks" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "bbox": [ + 443, + 212, + 514, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 443, + 212, + 514, + 230 + ], + "spans": [ + { + "bbox": [ + 443, + 212, + 514, + 230 + ], + "type": "text", + "content": "interactive tracking" + } + ] + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 445, + 232, + 529, + 272 + ], + "blocks": [ + { + "bbox": [ + 445, + 232, + 529, + 272 + ], + "lines": [ + { + "bbox": [ + 445, + 232, + 529, + 272 + ], + "spans": [ + { + "bbox": [ + 445, + 232, + 529, + 272 + ], + "type": "image", + "image_path": "7124886eb182e23855f7190caf02c1001c53d315a231d0c9af0091c2355a65dc.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "bbox": [ + 46, + 366, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 366, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 366, + 287, + 510 + ], + "type": "text", + "content": "straightforward tracking scenarios, but in cases of severe occlusion scenes, such as OVIS [68], image-level training cannot guarantee that the model exhibits strong temporal consistency under occlusion conditions. Therefore, for occlusion scenarios, it is essential to utilize video data for training. Following IDOL [92], we sample two frames from a video and introduce contrastive learning between frames to make the embedding of the same object instance closer in the embedding space, and the embedding of different object instances farther away. During Inference, the detected objects are tracked by simple bipartite matching of the corresponding object queries following MinVIS [36]." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 46, + 522, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 714 + ], + "type": "text", + "content": "Visual Prompted Segmentation. Interactive segmentation[8, 12, 56, 75, 80, 89, 97] takes various forms of visual prompt, such as points, boxes, or scribbles, to segment the specified objects within an image. On the other hand, VOS [22, 98] aims to segment the entire object throughout the entire video based on a mask provided in the first frame of the video. We extract visual prompt embeddings twice in the model. First, we crop the prompt square area from RGB image and send it to the backbone to obtain the visual prompt feature of the corresponding area, and send it to the early fusion module before the Transformer encoder. Second, we sample fine-grained visual embeddings from the pixel embedding map " + }, + { + "bbox": [ + 46, + 522, + 287, + 714 + ], + "type": "inline_equation", + "content": "M_{p}" + }, + { + "bbox": [ + 46, + 522, + 287, + 714 + ], + "type": "text", + "content": " according to visual prompt and make them interact with object queries through self-attention module in the Transformer decoder layer, as the same with text embeddings." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 306, + 366, + 423, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 366, + 423, + 379 + ], + "spans": [ + { + "bbox": [ + 306, + 366, + 423, + 379 + ], + "type": "text", + "content": "3.3. Training Unification" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 384, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 545, + 647 + ], + "type": "text", + "content": "Tasks with Dynamic Loss. We jointly train GLEE in an end-to-end manner on over 5 million images from diverse benchmarks with various levels of supervision. Different loss functions are selected for training on various datasets. There are six types of losses for our GLEE: semantic loss, box loss, mask loss, confidence loss, contrastive tracking loss, and distillation loss. For all tasks with category list or object expressions, we apply Focal loss [53] as semantic loss on logits " + }, + { + "bbox": [ + 304, + 384, + 545, + 647 + ], + "type": "inline_equation", + "content": "S_{\\text{align}}" + }, + { + "bbox": [ + 304, + 384, + 545, + 647 + ], + "type": "text", + "content": " to align the text concepts with object features. For box prediction, we use a combination of L1 loss and generalized IoU loss [73]. The mask loss is defined as a combination of the Dice loss [62] and Focal loss. For the Visual Prompt Segmentation tasks, we employ an additional FFN to predict the confidence score for each object queries supervised by Focal loss. For video tasks fine-tuning, we sample two frames and apply contrastive tracking loss on the object query from the last layer of decoder following IDOL [92]. For the text encoder, we distill the knowledge from the frozen teacher CLIP text encoder to ensure the text embedding in pre-trained vison-language embedding space. We apply an L1 loss between our text encoder and CLIP text encoder to minimize their distance:" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 322, + 663, + 545, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 663, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 322, + 663, + 545, + 696 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t e x t}} = \\frac {1}{K} \\sum_ {i = 0} ^ {K} \\| \\operatorname {E n c} _ {L} \\left(p _ {k}\\right) - \\operatorname {E n c} _ {C L I P} \\left(p _ {k}\\right) \\|, \\tag {3}", + "image_path": "7a74f0dbe213b8f3d4a130f88e634f956bbd767fccc0031dc3e704ccbeb73647.jpg" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 306, + 701, + 541, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 541, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 541, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 701, + 541, + 714 + ], + "type": "inline_equation", + "content": "\\{p_k\\}" + }, + { + "bbox": [ + 306, + 701, + 541, + 714 + ], + "type": "text", + "content": " is the category name from the category list " + }, + { + "bbox": [ + 306, + 701, + 541, + 714 + ], + "type": "inline_equation", + "content": "P" + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3787" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 70, + 545, + 294 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 545, + 294 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 545, + 294 + ], + "type": "table", + "html": "
MethodTypeGeneric Detection & SegmentationReferring Detection & SegmentationOpenWorld
COCO-valCOCO-test-devLVISRefCOCORefCOCO+RefCOCOgUVO
APboxAPmaskAPboxAPmaskAPboxAPr-boxAPmaskAPr-maskP@0.5IoUP@0.5IoUP@0.5IoUARmask
MDETR [38]--------87.5-81.1-83.4--
SeqTR [118]--------87.071.778.763.082.764.7-
PolyFormer (L) [55]--------90.476.985.072.285.871.2-
ViTDet-L [50]Specialist Models57.649.8--51.2-46.034.3-------
ViTDet-H [50]58.750.9--53.4-48.136.9-------
EVA-02-L [24]64.255.064.555.865.2-57.3-
ODISE [95]--------------57.7
Mask2Former (L) [15]-50.1-50.5-----------
MaskDINO (L) [45]-54.5-54.7-----------
UniTAB (B) [103]--------88.6-81.0-84.6--
OFA (L) [84]--------90.1-85.8-85.9--
Pix2Seq v2 [14]46.538.2-------------
Uni-Perceiver-v2 (B) [46]58.650.6-------------
Uni-Perceiver-v2 (L) [46]61.953.6-------------
UNINEXT (R50) [100]Generalist Models51.344.9--36.4---89.777.979.866.284.070.0-
UNINEXT (L) [100]58.149.6------91.480.383.170.086.973.4-
UNINEXT (H) [100]60.651.8------92.682.285.272.588.774.7-
GLIPv2 (B) [111]--58.845.8-----------
GLIPv2 (H) [111]--60.648.9-----------
X-Decoder (B) [121]-45.8-45.8-----------
X-Decoder (L) [121]-46.7-47.1-----------
Florence-2 (L) [94]43.4-------93.4-88.3-91.2--
GLEE-LiteFoundation Models55.048.454.748.344.236.740.233.788.577.478.364.882.968.866.6
GLEE-Plus60.453.060.653.352.744.547.440.490.679.581.668.385.070.670.6
GLEE-Pro62.054.262.354.555.749.249.944.391.080.082.669.686.472.972.6
", + "image_path": "91a55d0a477988555a9e3875244538343e178e414fb674fbc449ab854a9d49d3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 298, + 546, + 332 + ], + "lines": [ + { + "bbox": [ + 46, + 298, + 546, + 332 + ], + "spans": [ + { + "bbox": [ + 46, + 298, + 546, + 332 + ], + "type": "text", + "content": "Table 1. Comparison of GLEE to recent specialist and generalist models on object-level image tasks. For REC and RES tasks, we report Precision@0.5 and overall IoU (oIoU). For open-world instance segmentation task, we reported the average recall of 100 mask proposals (AR@100) on the UVO [85]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 344, + 289, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 344, + 289, + 511 + ], + "spans": [ + { + "bbox": [ + 46, + 344, + 289, + 511 + ], + "type": "text", + "content": "Data Scale Up. A visual foundation model should be able to easily scale up the training data and achieve better generalization performance. Thanks to the unified training paradigm, the training data can be scaled up at low cost by introducing a large amount of automatically labeled data from SA1B [39] and GRIT [67]. SA1B provides large and detailed mask annotations, which enhance the general object perception capabilities of model, while GRIT offers a more extensive collection of referring-expression-bounding-box pairs, improving the object identification abilities and the understanding capability of object descriptions. Ultimately, we introduced 2 million SA1B data and 5 million GRIT data into the training process, bringing the total training data to 10 million." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 528, + 128, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 528, + 128, + 541 + ], + "spans": [ + { + "bbox": [ + 47, + 528, + 128, + 541 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 549, + 163, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 549, + 163, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 549, + 163, + 562 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 570, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 570, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 45, + 570, + 288, + 715 + ], + "type": "text", + "content": "We conducted training in three stages. Initially, we performed pretraining for object detection task on Objects365 [79] and OpenImages [42], initializing the text encoder with pretrained CLIP [69] weights and keeping the parameters frozen. In the second training stage, we introduced additional instance segmentation datasets, including COCO [52], LVIS [29], and BDD [106]. Furthermore, we treat three VIS datasets: YTVIS19 [102], YTVIS21 [96], and OVIS [68], as independent image data to enrich the scenes. For datasets that provide descriptions of objects, we included RefCOCO [108], RefCOCO+ [108], RefCOCOg [64], Visual Genome [40], and RVOS [77]. Addi" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 344, + 547, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 344, + 547, + 596 + ], + "spans": [ + { + "bbox": [ + 304, + 344, + 547, + 596 + ], + "type": "text", + "content": "tionally, we introduced two open-world instance segmentation datasets, UVO [85] and a subset of SA1B [39]. Building upon this, we perform the third training stage by introducing more SA1B data and GRIT [67] data to scale up the training set, resulting in a model named GLEE-scale, which exhibited even stronger zero-shot performance on various downstream tasks. During the second and third stages, text encoder is unfrozen but supervised by distillation loss to ensure the predicted text embedding in CLIP embedding space. After the second step, GLEE demonstrated state-of-the-art performance on a range of downstream image and video tasks and exhibited strong zero-shot generalization capabilities, unless otherwise specified, all the experimental results presented below were obtained by the model from this stage. We developed GLEE-Lite, GLEE-Plus, and GLEE-Pro using ResNet-50 [30], SwinLarge [57], and EVA-02 Large [24] as the vision encoder respectively, and train GLEE on 64 A100 GPUs for 500,000 iterations in each stage. More detailed information on data usage, data sampling strategies, and model training can be found in the supplementary materials." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 609, + 500, + 622 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 609, + 500, + 622 + ], + "spans": [ + { + "bbox": [ + 305, + 609, + 500, + 622 + ], + "type": "text", + "content": "4.2. Comparison with Generalist Models" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "We demonstrate the universality and effectiveness of our model as an object-level visual foundation model, directly applicable to various object-centric tasks while ensuring SOTA performance without needing fine-tuning. We report detection and instance segmentation results on both the COCO-2017 [52] and LVIS val v1.0 [29]. While sharing almost identical image sets, LVIS is distinguished by" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3788" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 69, + 545, + 281 + ], + "blocks": [ + { + "bbox": [ + 50, + 69, + 545, + 281 + ], + "lines": [ + { + "bbox": [ + 50, + 69, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 50, + 69, + 545, + 281 + ], + "type": "table", + "html": "
MethodTracking Any Object (TAO [19])BURST [3]LV-VIS [83]
TETALocAAssocAClsAALLCommonUncommonAP\\( AP_b \\)\\( AP_n \\)
HOTAmAPHOTAmAPHOTAmAP
Tracktor [5]24.247.413.012.1---------
DeepSORT [88]26.048.417.512.1---------
Tracktor++ [19]28.049.022.812.1---------
QDTrack [66]30.050.527.412.1---------
TETer [48]33.351.635.013.2---------
OVTrack† [49]34.749.336.718.1---------
STCN Tracker† [3]----5.50.917.50.72.50.6---
Box Tracker† [3]----8.21.427.03.03.60.9---
Detic [117]-SORT† [6]----------12.821.16.6
Detic [117]-XMem †[16]----------16.324.110.6
OV2Seg-R50† [83]-----3.7----14.217.211.9
OV2Seg-B† [83]-----4.9----21.127.516.3
UNINEXT (R50) [100]31.943.335.517.1---------
GLEE-Lite†40.156.339.924.122.612.636.418.919.111.019.622.117.7
GLEE-Plus†41.552.940.930.826.917.238.823.723.915.530.331.629.3
GLEE-Pro†47.266.246.229.131.219.248.724.826.917.723.924.623.3
", + "image_path": "06fe5ecbbd551d128d6ae3ce300b2909766a8b8aabbe295bb1d8f13c302fc0f6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 289, + 546, + 357 + ], + "lines": [ + { + "bbox": [ + 46, + 289, + 546, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 546, + 357 + ], + "type": "text", + "content": "Table 2. Comparison of GLEE to recent specialist and generalist models on object-level video tasks in a zero-shot manner. Evaluation metrics of BURST are reported separately for 'common', 'uncommon' and 'all' classes. The mAP computes mask IoU at the track level, HOTA is a balance of per-frame detection accuracy (DetA) and temporal association accuracy (AssA), and TETA that deconstructs detection into localization and classification components. The AP, " + }, + { + "bbox": [ + 46, + 289, + 546, + 357 + ], + "type": "inline_equation", + "content": "\\mathrm{AP_b}" + }, + { + "bbox": [ + 46, + 289, + 546, + 357 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 289, + 546, + 357 + ], + "type": "inline_equation", + "content": "\\mathrm{AP_n}" + }, + { + "bbox": [ + 46, + 289, + 546, + 357 + ], + "type": "text", + "content": " in LV-VIS mean the average precision of overall categories, base categories, and novel categories. " + }, + { + "bbox": [ + 46, + 289, + 546, + 357 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 46, + 289, + 546, + 357 + ], + "type": "text", + "content": " does not use videos for training. The under-performance of Pro relative to Plus on LV-VIS is due to Pro employing larger training and inference resolutions, which prove to be sub-optimal for this specific dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 372, + 289, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 372, + 289, + 636 + ], + "spans": [ + { + "bbox": [ + 46, + 372, + 289, + 636 + ], + "type": "text", + "content": "its annotations of over 1,200 object categories, showcasing a long-tail distribution. This distinction makes LVIS more representative of challenging real-world scenarios due to its broader category coverage. As indicated in Table 1, our model outperforms all generalist models on both COCO and LVIS benchmarks. Even when compared to SOTA specialist approaches, which are tailored with specific designs, our model remains highly competitive. This demonstrates that GLEE, while mastering universal and general object representations, concurrently maintains advanced performance. This characteristic is vitally important for adapting to a broad spectrum of downstream tasks requiring precise object localization. For the REC and RES tasks, we evaluated our model on RefCOCO [108], RefCOCO+ [108], and RefCOCOg [64], as show in Table 1, GLEE achieved comparable results with SOTA specialist methods PolyFormer [55], demonstrating strong capability to comprehend textual descriptions and showcasing potential to adapt to a broader range of multi-modal downstream tasks. In open-world instance segmentation tasks, GLEE outperforms previous arts ODISE [95] by 8.9 points, demonstrating the capability of identifying all plausible instances in an open-world scenario." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 647, + 233, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 233, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 233, + 658 + ], + "type": "text", + "content": "4.3. Zero-shot Evaluation Across Tasks" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "Zero-shot Transfer to Video Tasks. The proposed GLEE is capable of adapting to new data and even new tasks in a zero-shot manner, without the need for additional finetuning. We evaluate its zero-shot capability on three large-" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 372, + 545, + 707 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 372, + 545, + 707 + ], + "spans": [ + { + "bbox": [ + 304, + 372, + 545, + 707 + ], + "type": "text", + "content": "scale, large-vocabulary open-world video tracking datasets: TAO [19], BURST [3], and LV-VIS [83]. TAO comprises 2,907 high-resolution videos across 833 categories. BURST builds upon TAO, encompassing 425 base categories and 57 novel categories. LV-VIS offers 4,828 videos within 1,196 well-defined object categories. These three benchmarks require the model to detect, classify, and track all objects in videos, while BURST and LV-VIS additionally require segmentation results from the model. In Table 2, we compare the performance of our proposed model with existing specialist models. Notably, the GLEE here is from the second training stage, which has not been exposed to images from these three datasets nor trained on video-level data. Despite these constraints, GLEE achieves state-of-the-art performance that significantly exceeds existing methodologies. Specifically, GLEE surpasses the previous best method OVTrack by " + }, + { + "bbox": [ + 304, + 372, + 545, + 707 + ], + "type": "inline_equation", + "content": "36.0\\%" + }, + { + "bbox": [ + 304, + 372, + 545, + 707 + ], + "type": "text", + "content": " in TAO, nearly triples the performance of the best baseline in BURST, and outperforms OV2Seg [83] by " + }, + { + "bbox": [ + 304, + 372, + 545, + 707 + ], + "type": "inline_equation", + "content": "43.6\\%" + }, + { + "bbox": [ + 304, + 372, + 545, + 707 + ], + "type": "text", + "content": " in LV-VIS. This outstanding performance strongly validates the exceptional generalization and zero-shot capabilities of GLEE in handling object-level tasks across a range of benchmarks and tasks. It can be observed that GLEE yields more impressive results on video tasks, since the image tasks have progressed with plentiful data and models from lower costs, while video tasks have not due to higher costs. The models trained on extensive image data with strong general perception capabilities can effectively transfer to video tasks." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3789" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 70, + 281, + 240 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 281, + 240 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 281, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 281, + 240 + ], + "type": "table", + "html": "
MethodBackboneYTVIS 2019 val [102]OVIS val [68]
AP\\( AP_{50} \\)\\( AP_{75} \\)AP\\( AP_{50} \\)\\( AP_{75} \\)
SeqFormer [90]47.469.851.815.131.913.8
IDOL [92]49.574.052.930.251.330.0
VITA [33]49.872.654.519.641.217.4
GenVIS [34]ResNet-5051.372.057.834.559.435.0
DVIS [112]52.676.558.234.159.832.3
NOVIS [61]52.875.756.932.756.232.6
UNINEXT [100]53.075.259.134.055.535.6
GLEE-Lite53.174.059.327.1/32.345.4/52.226.3/33.7
SeqFormer [90]59.382.166.4---
VITA [33]63.086.967.927.751.924.9
IDOL [92]64.387.571.042.665.745.2
GenVIS [34]Swin-L63.885.768.545.469.247.8
DVIS [112]64.988.072.749.975.953.0
NOVIS [61]65.787.872.243.568.343.8
GLEE-Plus63.685.270.529.6/40.350.3/63.828.9/39.8
UNINEXT [100]ConvNeXt-L64.387.271.741.165.842.0
UNINEXT [100]ViT-H66.987.575.149.072.552.2
GLEE-ProEVA02-L67.487.174.138.7/50.459.4/71.439.7/55.5
", + "image_path": "7972a8ea907b92482910f2a3c2e9b17cc5ed39f6da494aa75ae9323577cbc139.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 74, + 265, + 259, + 403 + ], + "blocks": [ + { + "bbox": [ + 47, + 240, + 286, + 263 + ], + "lines": [ + { + "bbox": [ + 47, + 240, + 286, + 263 + ], + "spans": [ + { + "bbox": [ + 47, + 240, + 286, + 263 + ], + "type": "text", + "content": "Table 3. Performance comparison on video instance segmentation tasks. " + }, + { + "bbox": [ + 47, + 240, + 286, + 263 + ], + "type": "inline_equation", + "content": "(\\_ /\\_)" + }, + { + "bbox": [ + 47, + 240, + 286, + 263 + ], + "type": "text", + "content": " reports results from zero-shot and after fine-tuning." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 74, + 265, + 259, + 403 + ], + "lines": [ + { + "bbox": [ + 74, + 265, + 259, + 403 + ], + "spans": [ + { + "bbox": [ + 74, + 265, + 259, + 403 + ], + "type": "table", + "html": "
Images MethodAPAP-categAP-descrAP-descr-posAP-descr-SAP-descr-MAP-descr-L
AllRegionCLIP [114]2.72.72.63.23.62.72.3
Detic [117]8.015.65.48.05.75.46.2
MDETR [38]--4.79.16.44.64.0
GLIP-T [47]19.323.616.425.829.414.88.2
GLIP-L [47]25.832.921.233.237.718.910.8
FIBER-B [23]25.730.322.334.838.619.512.4
GLEE-Lite20.337.514.019.123.012.710.0
GLEE-Lite-Scale22.735.516.722.333.714.310.2
GLEE-Plus25.446.717.523.928.416.312.5
GLEE-Plus-Scale27.044.519.425.936.017.212.4
", + "image_path": "559aae4a2612fbcb495b02f9f6c6ba56a6027344ba4e572bf27105adb36e482d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 403, + 287, + 437 + ], + "lines": [ + { + "bbox": [ + 47, + 403, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 47, + 403, + 287, + 437 + ], + "type": "text", + "content": "Table 4. Evaluation on the OmniLabel benchmark. The final AP value is the geometric mean of categories (AP-categ) and freeform descriptions (AP-descr)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 449, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 628 + ], + "type": "text", + "content": "We additionally provide performance comparison on classic video instance segmentation tasks, whose data is incorporated as image-level data during the second stage of training. As shown in Table 3, on the YTVIS2019 [102] benchmark, our model achieves SOTA results across various model sizes, surpassing all specialist models with complex designs to enhance temporal capabilities and the video unified model UNINEXT [100]. On the OVIS [68] benchmark, which features lengthy videos with extensive object occlusions where temporal capabilities of object features are particularly crucial, our model does not directly reach SOTA. However, after a few hours of simple fine-tuning, it still achieves SOTA performance. More details on VOS, RVOS and demonstrations of interactive segmentation and tracking can be found in supplementary materials." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 630, + 286, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 286, + 714 + ], + "type": "text", + "content": "Zero-shot Transfer to Real-world Downstream Tasks. To measure generalization on real-world tasks, we evaluate zero-shot performance on OmniLabel [76], which is a benchmark for evaluating language-based object detectors and encourages diverse free-form text descriptions of objects. As show in Table 4, compared to language-based detectors trained on large-scale caption data, GLEE signif" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 318, + 70, + 531, + 192 + ], + "blocks": [ + { + "bbox": [ + 318, + 70, + 531, + 192 + ], + "lines": [ + { + "bbox": [ + 318, + 70, + 531, + 192 + ], + "spans": [ + { + "bbox": [ + 318, + 70, + 531, + 192 + ], + "type": "image", + "image_path": "c1c1e055e4ea5e7d63b61e43de0e25d2fadccf5f23d8359b4ae7924290d4f83d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 199, + 545, + 232 + ], + "lines": [ + { + "bbox": [ + 305, + 199, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 305, + 199, + 545, + 232 + ], + "type": "text", + "content": "Figure 4. The performance comparison of replacing SAM with GLEE in LISA, GLEE achieves the same effectiveness as SAM in extracting objects." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 240, + 545, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 240, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 304, + 240, + 545, + 337 + ], + "type": "text", + "content": "icantly outperforms previous models in AP-categ. Due to the limited captions in our training dataset, it scores lower in AP-descr. By incorporating a more diverse set of box-tection data from the GRIT [67] to scale up our training set, the AP-descr can be elevated to a level comparable with existing models. A more comprehensive report on the zero-shot and few-shot performance on ODinW [44] and ablation studies are provided in the supplementary materials." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 344, + 456, + 356 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 344, + 456, + 356 + ], + "spans": [ + { + "bbox": [ + 306, + 344, + 456, + 356 + ], + "type": "text", + "content": "4.4. Serve as Foundation Model" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 362, + 545, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 362, + 545, + 507 + ], + "spans": [ + { + "bbox": [ + 304, + 362, + 545, + 507 + ], + "type": "text", + "content": "To explore whether GLEE can serve as a foundation model for other architectures, we selected LISA [43] for analysis, a mVLLM that combines LLAVA [54] with SAM [39] for reasoning segmentation. We substituted its vision backbone with a frozen, pretrained GLEE-Plus and fed the object queries from GLEE into LLAVA and remove decoder of LISA. We directly dot product the output SEG tokens with GLEE feature map to generate masks. As shown in Figure 4, after training for the same number of steps, our modified LISA-GLEE achieved comparable results to the original version, demonstrating the versatility of representations from GLEE and its effectiveness in serving other models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 518, + 378, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 518, + 378, + 530 + ], + "spans": [ + { + "bbox": [ + 306, + 518, + 378, + 530 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 538, + 545, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 538, + 545, + 670 + ], + "spans": [ + { + "bbox": [ + 304, + 538, + 545, + 670 + ], + "type": "text", + "content": "We introduce GLEE, a cutting-edge object-level foundation model designed to be directly applicable to a wide range of object-level image and video tasks. Crafted with a unified learning paradigm, GLEE learns from diverse data sources with varying levels of supervisions. GLEE achieves state-of-the-art performance on numerous object-level tasks and excels in zero-shot transfer to new data and tasks, showing its exceptional versatility and generalization abilities. Additionally, GLEE provides general visual object-level information, which is currently missing in modern LLMs, establishing a robust foundation for object-centric mLLMs." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 681, + 409, + 694 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 681, + 409, + 694 + ], + "spans": [ + { + "bbox": [ + 306, + 681, + 409, + 694 + ], + "type": "text", + "content": "6. Acknowledgement" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 701, + 518, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 518, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 518, + 713 + ], + "type": "text", + "content": "This research is supported by NSFC (No.62225603)." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3790" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 58, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 58, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] David Acuna, Huan Ling, Amlan Kar, and Sanja Fidler. Efficient interactive annotation of segmentation datasets with polygon-rnn++. 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 125, + 287, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 125, + 287, + 190 + ], + "spans": [ + { + "bbox": [ + 58, + 125, + 287, + 190 + ], + "type": "text", + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35: 23716-23736, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 193, + 287, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 193, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 57, + 193, + 287, + 258 + ], + "type": "text", + "content": "[3] Ali Athar, Jonathon Luiten, Paul Voigtlaender, Tarasha Khurana, Achal Dave, Bastian Leibe, and Deva Ramanan. Burst: A benchmark for unifying object recognition, segmentation and tracking in video. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1674-1683, 2023. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 260, + 286, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 260, + 286, + 282 + ], + "spans": [ + { + "bbox": [ + 57, + 260, + 286, + 282 + ], + "type": "text", + "content": "[4] Philipp Bergmann, Tim Meinhardt, and Laura Leal-Taixe. Tracking without bells and whistles. In ICCV, 2019. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 284, + 286, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 284, + 286, + 305 + ], + "spans": [ + { + "bbox": [ + 57, + 284, + 286, + 305 + ], + "type": "text", + "content": "[5] Philipp Bergmann, Tim Meinhardt, and Laura Leal-Taixe. Tracking without bells and whistles. In ICCV, 2019. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 308, + 287, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 308, + 287, + 351 + ], + "spans": [ + { + "bbox": [ + 57, + 308, + 287, + 351 + ], + "type": "text", + "content": "[6] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In 2016 IEEE international conference on image processing (ICIP), pages 3464-3468. IEEE, 2016. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 353, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 353, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 57, + 353, + 287, + 407 + ], + "type": "text", + "content": "[7] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 409, + 287, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 409, + 287, + 463 + ], + "spans": [ + { + "bbox": [ + 57, + 409, + 287, + 463 + ], + "type": "text", + "content": "[8] Yuri Y Boykov and M-P Jolly. Interactive graph cuts for optimal boundary & region segmentation of objects in nd images. In Proceedings eighth IEEE international conference on computer vision. ICCV 2001, pages 105-112. IEEE, 2001. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 466, + 287, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 466, + 287, + 530 + ], + "spans": [ + { + "bbox": [ + 57, + 466, + 287, + 530 + ], + "type": "text", + "content": "[9] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1, 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 533, + 287, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 533, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 53, + 533, + 287, + 576 + ], + "type": "text", + "content": "[10] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 578, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 578, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 53, + 578, + 287, + 633 + ], + "type": "text", + "content": "[11] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 635, + 287, + 688 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 635, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 53, + 635, + 287, + 688 + ], + "type": "text", + "content": "[12] Lluis Castrejon, Kaustav Kundu, Raquel Urtasun, and Sanja Fidler. Annotating object instances with a polygon-rnn. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5230-5238, 2017. 2, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 691, + 287, + 713 + ], + "type": "text", + "content": "[13] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pre" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 74, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 315, + 74, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 74, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 315, + 74, + 545, + 95 + ], + "type": "text", + "content": "training from pixels. In International conference on machine learning, pages 1691-1703. PMLR, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 96, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 96, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 312, + 96, + 545, + 139 + ], + "type": "text", + "content": "[14] Ting Chen, Saurabh Saxena, Lala Li, Tsung-Yi Lin, David J Fleet, and Geoffrey E Hinton. A unified sequence interface for vision tasks. Advances in Neural Information Processing Systems, 35:31333-31346, 2022. 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 140, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 140, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 312, + 140, + 545, + 183 + ], + "type": "text", + "content": "[15] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 2, 4, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 185, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 185, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 312, + 185, + 545, + 228 + ], + "type": "text", + "content": "[16] Ho Kei Cheng and Alexander G Schwing. Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In European Conference on Computer Vision, pages 640-658. Springer, 2022. 2, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 228, + 545, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 228, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 312, + 228, + 545, + 281 + ], + "type": "text", + "content": "[17] Ho Kei Cheng, Yu-Wing Tai, and Chi-Keung Tang. Rethinking space-time networks with improved memory coverage for efficient video object segmentation. Advances in Neural Information Processing Systems, 34:11781-11794, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 283, + 545, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 283, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 312, + 283, + 545, + 338 + ], + "type": "text", + "content": "[18] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 339, + 545, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 339, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 312, + 339, + 545, + 371 + ], + "type": "text", + "content": "[19] Achal Dave, Tarasha Khurana, Pavel Tokmakov, Cordelia Schmid, and Deva Ramanan. Tao: A large-scale benchmark for tracking any object. In ECCV, 2020. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 372, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 372, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 312, + 372, + 545, + 426 + ], + "type": "text", + "content": "[20] Patrick Dendorfer, Aljosa Osep, Anton Milan, Konrad Schindler, Daniel Cremers, Ian Reid, Stefan Roth, and Laura Leal-Taixe. Motchallenge: A benchmark for single-camera multiple target tracking. International Journal of Computer Vision, 129(4):845-881, 2021. 2, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 426, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 426, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 312, + 426, + 545, + 469 + ], + "type": "text", + "content": "[21] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 472, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 472, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 312, + 472, + 545, + 514 + ], + "type": "text", + "content": "[22] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, Philip HS Torr, and Song Bai. Mose: A new dataset for video object segmentation in complex scenes. arXiv preprint arXiv:2302.01872, 2023. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 515, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 515, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 312, + 515, + 545, + 568 + ], + "type": "text", + "content": "[23] Zi-Yi Dou, Aishwarya Kamath, Zhe Gan, Pengchuan Zhang, Jianfeng Wang, Linjie Li, Zicheng Liu, Ce Liu, Yann LeCun, Nanyun Peng, et al. Coarse-to-fine vision-language pre-training with fusion in the backbone. NeurIPS, 35:32942-32956, 2022. 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 570, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 570, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 312, + 570, + 545, + 613 + ], + "type": "text", + "content": "[24] Yuxin Fang, Quan Sun, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva-02: A visual representation for neon genesis. arXiv preprint arXiv:2303.11331, 2023.6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 614, + 545, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 614, + 545, + 680 + ], + "spans": [ + { + "bbox": [ + 312, + 614, + 545, + 680 + ], + "type": "text", + "content": "[25] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19358-19369, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 681, + 545, + 713 + ], + "type": "text", + "content": "[26] Chengjian Feng, Yujie Zhong, Zequn Jie, Xiangxiang Chu, Haibing Ren, Xiaolin Wei, Weidi Xie, and Lin Ma. Promptdet: Towards open-vocabulary detection using uncurated" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "3791" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 72, + 72, + 286, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 72, + 286, + 106 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 286, + 106 + ], + "type": "text", + "content": "images. In Computer Vision - ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IX, page 701-717, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 106, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 106, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 53, + 106, + 287, + 160 + ], + "type": "text", + "content": "[27] Golnaz Ghiasi, Barret Zoph, Ekin D Cubuk, Quoc V Le, and Tsung-Yi Lin. Multi-task self-training for learning general representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8856-8865, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 162, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 162, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 54, + 162, + 287, + 205 + ], + "type": "text", + "content": "[28] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In International Conference on Learning Representations, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 206, + 287, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 206, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 54, + 206, + 287, + 249 + ], + "type": "text", + "content": "[29] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 2, 4, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 251, + 286, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 251, + 286, + 281 + ], + "spans": [ + { + "bbox": [ + 54, + 251, + 286, + 281 + ], + "type": "text", + "content": "[30] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 4, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 284, + 286, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 284, + 286, + 304 + ], + "spans": [ + { + "bbox": [ + 53, + 284, + 286, + 304 + ], + "type": "text", + "content": "[31] Kaiming He, Georgia Gkioxari, Piotr Dollar, and Ross Girshick. Mask R-CNN. In ICCV, 2017. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 306, + 287, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 306, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 54, + 306, + 287, + 358 + ], + "type": "text", + "content": "[32] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 361, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 361, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 54, + 361, + 287, + 403 + ], + "type": "text", + "content": "[33] Miran Heo, Sukjun Hwang, Seoung Wug Oh, Joon-Young Lee, and Seon Joo Kim. Vita: Video instance segmentation via object token association. In Advances in Neural Information Processing Systems, 2022. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 405, + 287, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 405, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 54, + 405, + 287, + 469 + ], + "type": "text", + "content": "[34] Miran Heo, Sukjun Hwang, Jeongseok Hyun, Hanjung Kim, Seoung Wug Oh, Joon-Young Lee, and Seon Joo Kim. A generalized framework for video instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14623-14632, 2023. 2, 4, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 472, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 472, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 54, + 472, + 287, + 525 + ], + "type": "text", + "content": "[35] Ronghang Hu, Marcus Rohrbach, and Trevor Darrell. Segmentation from natural language expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 108-124. Springer, 2016. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 526, + 286, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 526, + 286, + 558 + ], + "spans": [ + { + "bbox": [ + 53, + 526, + 286, + 558 + ], + "type": "text", + "content": "[36] De-An Huang, Zhiding Yu, and Anima Anandkumar. Minvis: A minimal video instance segmentation framework without video-based training. In NeurIPS, 2022. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 559, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 559, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 53, + 559, + 287, + 624 + ], + "type": "text", + "content": "[37] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 625, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 625, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 53, + 625, + 287, + 668 + ], + "type": "text", + "content": "[38] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In ICCV, pages 1780-1790, 2021. 6, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 670, + 287, + 712 + ], + "type": "text", + "content": "[39] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, 2023. 1, 2, 6, 8" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 312, + 73, + 545, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 545, + 138 + ], + "type": "text", + "content": "[40] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 2, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 140, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 140, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 312, + 140, + 545, + 184 + ], + "type": "text", + "content": "[41] Weicheng Kuo, Yin Cui, Xiuye Gu, AJ Piergiovanni, and Anelia Angelova. Open-vocabulary object detection upon frozen vision and language models. In The Eleventh International Conference on Learning Representations, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 186, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 186, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 312, + 186, + 545, + 251 + ], + "type": "text", + "content": "[42] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, Tom Duerig, and Vittorio Ferrari. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. IJCV, 2020. 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 253, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 253, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 312, + 253, + 545, + 296 + ], + "type": "text", + "content": "[43] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692, 2023. 2, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 298, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 298, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 312, + 298, + 545, + 352 + ], + "type": "text", + "content": "[44] Chunyuan Li, Haotian Liu, Lianian Li, Pengchuan Zhang, Jyoti Aneja, Jianwei Yang, Ping Jin, Houdong Hu, Zicheng Liu, Yong Jae Lee, et al. Elevater: A benchmark and toolkit for evaluating language-augmented visual models. In NeurlPS, 2022. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 354, + 545, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 354, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 312, + 354, + 545, + 397 + ], + "type": "text", + "content": "[45] Feng Li, Hao Zhang, Huaizhe Xu, Shilong Liu, Lei Zhang, Lionel M Ni, and Heung-Yeung Shum. Mask dino: Towards a unified transformer-based framework for object detection and segmentation. In CVPR, 2023. 3, 4, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 399, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 399, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 312, + 399, + 545, + 464 + ], + "type": "text", + "content": "[46] Hao Li, Jinguo Zhu, Xiaohu Jiang, Xizhou Zhu, Hongsheng Li, Chun Yuan, Xiaohua Wang, Yu Qiao, Xiaogang Wang, Wenhai Wang, et al. Uni-perceiver v2: A generalist model for large-scale vision and vision-language tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2691-2700, 2023. 3, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 467, + 545, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 467, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 312, + 467, + 545, + 520 + ], + "type": "text", + "content": "[47] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In CVPR, pages 10965-10975, 2022. 3, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 523, + 545, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 523, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 312, + 523, + 545, + 554 + ], + "type": "text", + "content": "[48] Siyuan Li, Martin Danelljan, Henghui Ding, Thomas E Huang, and Fisher Yu. Tracking every thing in the wild. In ECCV, 2022. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 557, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 557, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 312, + 557, + 545, + 610 + ], + "type": "text", + "content": "[49] Siyuan Li, Tobias Fischer, Lei Ke, Henghui Ding, Martin Danelljan, and Fisher Yu. Ovtrack: Open-vocabulary multiple object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5567-5577, 2023. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 613, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 613, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 312, + 613, + 545, + 657 + ], + "type": "text", + "content": "[50] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In European Conference on Computer Vision, pages 280-296. Springer, 2022. 4, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 312, + 658, + 545, + 712 + ], + "type": "text", + "content": "[51] Chuang Lin, Peize Sun, Yi Jiang, Ping Luo, Lizhen Qu, Gholamreza Haffari, Zehuan Yuan, and Jianfei Cai. Learning object-language alignments for open-vocabulary object detection. In The Eleventh International Conference on Learning Representations, 2023. 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "3792" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 73, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 73, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 73, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 53, + 73, + 287, + 116 + ], + "type": "text", + "content": "[52] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 2, 4, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 118, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 118, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 53, + 118, + 287, + 149 + ], + "type": "text", + "content": "[53] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In ICCV, 2017. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 151, + 288, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 151, + 288, + 183 + ], + "spans": [ + { + "bbox": [ + 53, + 151, + 288, + 183 + ], + "type": "text", + "content": "[54] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 184, + 288, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 184, + 288, + 249 + ], + "spans": [ + { + "bbox": [ + 53, + 184, + 288, + 249 + ], + "type": "text", + "content": "[55] Jiang Liu, Hui Ding, Zhaowei Cai, Yuting Zhang, Ravi Kumar Satzoda, Vijay Mahadevan, and R Manmatha. Polyformer: Referring image segmentation as sequential polygon generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18653-18663, 2023. 2, 4, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 250, + 288, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 250, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 53, + 250, + 288, + 304 + ], + "type": "text", + "content": "[56] Qin Liu, Zhenlin Xu, Gedas Bertasius, and Marc Niethammer. Simpleclick: Interactive image segmentation with simple vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22290-22300, 2023. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 304, + 288, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 304, + 288, + 348 + ], + "spans": [ + { + "bbox": [ + 53, + 304, + 288, + 348 + ], + "type": "text", + "content": "[57] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. arXiv preprint arXiv:2103.14030, 2021. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 350, + 288, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 350, + 288, + 392 + ], + "spans": [ + { + "bbox": [ + 53, + 350, + 288, + 392 + ], + "type": "text", + "content": "[58] Jiasen Lu, Christopher Clark, Rowan Zellers, Roozbeh Mottaghi, and Aniruddha Kembhavi. Unified-io: A unified model for vision, language, and multi-modal tasks. arXiv preprint arXiv:2206.08916, 2022.3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 394, + 288, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 394, + 288, + 436 + ], + "spans": [ + { + "bbox": [ + 53, + 394, + 288, + 436 + ], + "type": "text", + "content": "[59] Chuofan Ma, Yi Jiang, Xin Wen, Zehuan Yuan, and Xiaojuan Qi. Codet: Co-occurrence guided region-word alignment for open-vocabulary object detection. arXiv preprint arXiv:2310.16667, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 437, + 288, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 437, + 288, + 480 + ], + "spans": [ + { + "bbox": [ + 53, + 437, + 288, + 480 + ], + "type": "text", + "content": "[60] Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. TrackFormer: Multi-object tracking with transformers. arXiv preprint arXiv:2101.02702, 2021. 2, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 482, + 288, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 482, + 288, + 525 + ], + "spans": [ + { + "bbox": [ + 53, + 482, + 288, + 525 + ], + "type": "text", + "content": "[61] Tim Meinhardt, Matt Feiszli, Yuchen Fan, Laura Leal-Taixe, and Rakesh Ranjan. Novis: A case for end-to-end near-online video instance segmentation. arXiv preprint arXiv:2308.15266, 2023. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 526, + 288, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 526, + 288, + 568 + ], + "spans": [ + { + "bbox": [ + 53, + 526, + 288, + 568 + ], + "type": "text", + "content": "[62] Fausto Miletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), 2016. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 570, + 288, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 570, + 288, + 657 + ], + "spans": [ + { + "bbox": [ + 53, + 570, + 288, + 657 + ], + "type": "text", + "content": "[63] Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, Xiao Wang, Xiaohua Zhai, Thomas Kipf, and Neil Houlsby. Simple open-vocabulary object detection. In Computer Vision - ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part X, page 728-755, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 658, + 288, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 658, + 288, + 690 + ], + "spans": [ + { + "bbox": [ + 53, + 658, + 288, + 690 + ], + "type": "text", + "content": "[64] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In ECCV, 2016. 2, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 692, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 692, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 692, + 288, + 713 + ], + "type": "text", + "content": "[65] Seoung Wug Oh, Joon-Young Lee, Ning Xu, and Seon Joo Kim. Video object segmentation using space-time memory" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 331, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 331, + 73, + 545, + 105 + ], + "type": "text", + "content": "networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9226-9235, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 106, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 106, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 312, + 106, + 545, + 139 + ], + "type": "text", + "content": "[66] Jiangmiao Pang, Linlu Qiu, Xia Li, Haofeng Chen, Qi Li, Trevor Darrell, and Fisher Yu. Quasi-dense similarity learning for multiple object tracking. In CVPR, 2021. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 140, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 140, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 312, + 140, + 545, + 183 + ], + "type": "text", + "content": "[67] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 2, 6, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 184, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 184, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 312, + 184, + 545, + 237 + ], + "type": "text", + "content": "[68] Jiyang Qi, Yan Gao, Yao Hu, Xinggang Wang, Xiaoyu Liu, Xiang Bai, Serge Belongie, Alan Yuille, Philip HS Torr, and Song Bai. Occluded video instance segmentation: A benchmark. International Journal of Computer Vision, pages 1-18, 2022. 4, 5, 6, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 239, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 239, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 312, + 239, + 545, + 304 + ], + "type": "text", + "content": "[69] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 3, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 305, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 305, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 312, + 305, + 545, + 360 + ], + "type": "text", + "content": "[70] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485-5551, 2020. 1, 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 361, + 545, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 361, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 312, + 361, + 545, + 414 + ], + "type": "text", + "content": "[71] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 415, + 545, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 415, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 312, + 415, + 545, + 458 + ], + "type": "text", + "content": "[72] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2):3, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 460, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 460, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 312, + 460, + 545, + 502 + ], + "type": "text", + "content": "[73] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In CVPR, 2019. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 504, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 504, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 312, + 504, + 545, + 558 + ], + "type": "text", + "content": "[74] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 559, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 559, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 312, + 559, + 545, + 602 + ], + "type": "text", + "content": "[75] Carsten Rother, Vladimir Kolmogorov, and Andrew Blake. \"grabcut\" interactive foreground extraction using iterated graph cuts. ACM transactions on graphics (TOG), 23(3): 309-314, 2004. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 604, + 545, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 604, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 312, + 604, + 545, + 647 + ], + "type": "text", + "content": "[76] Samuel Schulter, Vijay Kumar B G, Yumin Suh, Konstantinos M. Dafnis, Zhixing Zhang, Shiyu Zhao, and Dimitris Metaxas. Omnilabel: A challenging benchmark for language-based object detection. In ICCV, 2023. 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 647, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 647, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 312, + 647, + 545, + 712 + ], + "type": "text", + "content": "[77] Seonguk Seo, Joon-Young Lee, and Bohyung Han. Urvos: Unified referring video object segmentation network with a large-scale benchmark. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XV 16, pages 208–223. Springer, 2020. 2, 4, 6" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "3793" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 72, + 287, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 72, + 287, + 117 + ], + "spans": [ + { + "bbox": [ + 53, + 72, + 287, + 117 + ], + "type": "text", + "content": "[78] Jing Shao, Siyu Chen, Yangguang Li, Kun Wang, Zhenfei Yin, Yinan He, Jianing Teng, Qinghong Sun, Mengya Gao, Jihao Liu, et al. Intern: A new learning paradigm towards general vision. arXiv preprint arXiv:2111.08687, 2021. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 119, + 288, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 119, + 288, + 175 + ], + "spans": [ + { + "bbox": [ + 53, + 119, + 288, + 175 + ], + "type": "text", + "content": "[79] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019. 2, 4, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 176, + 288, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 176, + 288, + 231 + ], + "spans": [ + { + "bbox": [ + 53, + 176, + 288, + 231 + ], + "type": "text", + "content": "[80] Konstantin Sofiuk, Ilya A Petrov, and Anton Konushin. Reviving iterative training with mask guidance for interactive segmentation. In 2022 IEEE International Conference on Image Processing (ICIP), pages 3141-3145. IEEE, 2022. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 233, + 288, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 233, + 288, + 299 + ], + "spans": [ + { + "bbox": [ + 53, + 233, + 288, + 299 + ], + "type": "text", + "content": "[81] Peize Sun, Rufeng Zhang, Yi Jiang, Tao Kong, Chenfeng Xu, Wei Zhan, Masayoshi Tomizuka, Lei Li, Zehuan Yuan, Changhu Wang, et al. Sparse r-cnn: End-to-end object detection with learnable proposals. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14454-14463, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 301, + 288, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 301, + 288, + 356 + ], + "spans": [ + { + "bbox": [ + 53, + 301, + 288, + 356 + ], + "type": "text", + "content": "[82] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 358, + 287, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 358, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 53, + 358, + 287, + 414 + ], + "type": "text", + "content": "[83] Haochen Wang, Cilin Yan, Shuai Wang, Xiaolong Jiang, Xu Tang, Yao Hu, Weidi Xie, and Efstratios Gavves. Towards open-vocabulary video instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4057-4066, 2023. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 416, + 288, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 416, + 288, + 482 + ], + "spans": [ + { + "bbox": [ + 53, + 416, + 288, + 482 + ], + "type": "text", + "content": "[84] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In International Conference on Machine Learning, pages 23318-23340. PMLR, 2022. 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 484, + 288, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 484, + 288, + 517 + ], + "spans": [ + { + "bbox": [ + 53, + 484, + 288, + 517 + ], + "type": "text", + "content": "[85] Weiyao Wang, Matt Feiszli, Heng Wang, and Du Tran. Unidentified video objects: A benchmark for dense, open-world segmentation. In ICCV, 2021. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 519, + 288, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 519, + 288, + 574 + ], + "spans": [ + { + "bbox": [ + 53, + 519, + 288, + 574 + ], + "type": "text", + "content": "[86] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, et al. Image as a foreign language: Beit pretraining for all vision and vision-language tasks. arXiv preprint arXiv:2208.10442, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 577, + 288, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 577, + 288, + 620 + ], + "spans": [ + { + "bbox": [ + 53, + 577, + 288, + 620 + ], + "type": "text", + "content": "[87] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In CVPR, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 623, + 288, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 623, + 288, + 655 + ], + "spans": [ + { + "bbox": [ + 53, + 623, + 288, + 655 + ], + "type": "text", + "content": "[88] Nicolai Wojke, Alex Bewley, and Dietrich Paulus. Simple online and realtime tracking with a deep association metric. In ICIP, 2017. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 658, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 658, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 658, + 288, + 713 + ], + "type": "text", + "content": "[89] Jiajun Wu, Yibiao Zhao, Jun-Yan Zhu, Siwei Luo, and Zhuowen Tu. Milcut: A sweeping line multiple instance learning paradigm for interactive image segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 256-263, 2014. 5" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 72, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 312, + 72, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 72, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 312, + 72, + 545, + 106 + ], + "type": "text", + "content": "[90] Junfeng Wu, Yi Jiang, Song Bai, Wenqing Zhang, and Xiang Bai. Seqformer: Sequential transformer for video instance segmentation. In ECCV, 2022. 2, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 107, + 547, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 107, + 547, + 160 + ], + "spans": [ + { + "bbox": [ + 312, + 107, + 547, + 160 + ], + "type": "text", + "content": "[91] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4974-4984, 2022. 2, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 162, + 545, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 162, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 312, + 162, + 545, + 204 + ], + "type": "text", + "content": "[92] Junfeng Wu, Qihao Liu, Yi Jiang, Song Bai, Alan Yuille, and Xiang Bai. In defense of online models for video instance segmentation. In ECCV, pages 588-605. Springer, 2022. 2, 4, 5, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 206, + 545, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 206, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 312, + 206, + 545, + 258 + ], + "type": "text", + "content": "[93] Jiannan Wu, Yi Jiang, Bin Yan, Hutchuan Lu, Zehuan Yuan, and Ping Luo. Segment every reference object in spatial and temporal spaces. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2538-2550, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 261, + 545, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 261, + 545, + 314 + ], + "spans": [ + { + "bbox": [ + 312, + 261, + 545, + 314 + ], + "type": "text", + "content": "[94] Bin Xiao, Haiping Wu, Weijian Xu, Xiyang Dai, Houdong Hu, Yumao Lu, Michael Zeng, Ce Liu, and Lu Yuan. Florence-2: Advancing a unified representation for a variety of vision tasks. arXiv preprint arXiv:2311.06242, 2023. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 316, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 316, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 312, + 316, + 545, + 380 + ], + "type": "text", + "content": "[95] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2955-2966, 2023. 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 382, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 382, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 312, + 382, + 545, + 425 + ], + "type": "text", + "content": "[96] Ning Xu, Linjie Yang, Jianchao Yang, Dingcheng Yue, Yuchen Fan, Yuchen Liang, and Thomas S. Huang. Youtuvis dataset 2021 version. https://youtu-vos.org/dataset/vis/.6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 426, + 545, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 426, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 312, + 426, + 545, + 470 + ], + "type": "text", + "content": "[97] Ning Xu, Brian Price, Scott Cohen, Jimei Yang, and Thomas S Huang. Deep interactive object selection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 373-381, 2016. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 472, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 472, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 312, + 472, + 545, + 514 + ], + "type": "text", + "content": "[98] Ning Xu, Linjie Yang, Yuchen Fan, Dingcheng Yue, Yuchen Liang, Jianchao Yang, and Thomas Huang. Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327, 2018. 2, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 515, + 545, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 515, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 312, + 515, + 545, + 548 + ], + "type": "text", + "content": "[99] Bin Yan, Yi Jiang, Peize Sun, Dong Wang, Zehuan Yuan, Ping Luo, and Huchuan Lu. Towards grand unification of object tracking. In ECCV, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 548, + 545, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 590 + ], + "type": "text", + "content": "[100] Bin Yan, Yi Jiang, Jiannan Wu, Dong Wang, Ping Luo, Zehuan Yuan, and Huchuan Lu. Universal instance perception as object discovery and retrieval. In CVPR, 2023. 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 592, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 592, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 592, + 545, + 613 + ], + "type": "text", + "content": "[101] Haosen Yang, Chuofan Ma, Bin Wen, Yi Jiang, Zehuan Yuan, and Xiatian Zhu. Recognize any regions. 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 615, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 615, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 615, + 545, + 635 + ], + "type": "text", + "content": "[102] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV, 2019. 2, 4, 6, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 637, + 545, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 637, + 545, + 691 + ], + "spans": [ + { + "bbox": [ + 308, + 637, + 545, + 691 + ], + "type": "text", + "content": "[103] Zhengyuan Yang, Zhe Gan, Jianfeng Wang, Xiaowei Hu, Faisal Ahmed, Zicheng Liu, Yumao Lu, and Lijuan Wang. Unitab: Unifying text and box outputs for grounded vision-language modeling. In European Conference on Computer Vision, pages 521-539. Springer, 2022. 3, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "text", + "content": "[104] Lewei Yao, Jianhua Han, Youpeng Wen, Xiaodan Liang, Dan Xu, Wei Zhang, Zhenguo Li, Chunjing Xu, and Hang" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "3794" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 72, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 287, + 95 + ], + "type": "text", + "content": "Xu. Detclip: Dictionary-enriched visual-concept paralleled pre-training for open-world detection. In NeurIPS, 2022. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 288, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 288, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 288, + 162 + ], + "type": "text", + "content": "[105] Lewei Yao, Jianhua Han, Xiaodan Liang, Dan Xu, Wei Zhang, Zhenguo Li, and Hang Xu. Detclipv2: Scalable open-vocabulary object detection pre-training via word-region alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23497-23506, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 229 + ], + "type": "text", + "content": "[106] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2636-2645, 2020. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 231, + 288, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 231, + 288, + 275 + ], + "spans": [ + { + "bbox": [ + 48, + 231, + 288, + 275 + ], + "type": "text", + "content": "[107] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. Transactions on Machine Learning Research, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 276, + 288, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 288, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 288, + 308 + ], + "type": "text", + "content": "[108] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In ECCV, 2016. 2, 4, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 310, + 288, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 310, + 288, + 365 + ], + "spans": [ + { + "bbox": [ + 48, + 310, + 288, + 365 + ], + "type": "text", + "content": "[109] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 366, + 288, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 366, + 288, + 421 + ], + "spans": [ + { + "bbox": [ + 48, + 366, + 288, + 421 + ], + "type": "text", + "content": "[110] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-vocabulary object detection using captions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14393-14402, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 422, + 288, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 288, + 488 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 288, + 488 + ], + "type": "text", + "content": "[111] Haotian Zhang, Pengchuan Zhang, Xiaowei Hu, Yen-Chun Chen, Liunian Li, Xiyang Dai, Lijuan Wang, Lu Yuan, Jenq-Neng Hwang, and Jianfeng Gao. Glipv2: Unifying localization and vision-language understanding. In Advances in Neural Information Processing Systems, pages 36067-36080, 2022. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 490, + 288, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 490, + 288, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 490, + 288, + 544 + ], + "type": "text", + "content": "[112] Tao Zhang, Xingye Tian, Yu Wu, Shunping Ji, Xuebo Wang, Yuan Zhang, and Pengfei Wan. Dvis: Decoupled video instance segmentation framework. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1282-1291, 2023. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 546, + 288, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 546, + 288, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 546, + 288, + 590 + ], + "type": "text", + "content": "[113] Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. Byte-track: Multi-object tracking by associating every detection box. arXiv preprint arXiv:2110.06864, 2021. 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 591, + 288, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 288, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 288, + 645 + ], + "type": "text", + "content": "[114] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Lianian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In CVPR, pages 16793-16803, 2022. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 647, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 714 + ], + "type": "text", + "content": "[115] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Lianian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16793-16803, 2022. 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 408 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "text", + "content": "[116] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Tracking objects as points. In ECCV, 2020. 2, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "text", + "content": "[117] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In ECCV, pages 350-368. Springer, 2022. 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 141, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 206 + ], + "type": "text", + "content": "[118] Chaoyang Zhu, Yiyi Zhou, Yunhang Shen, Gen Luo, Xingjia Pan, Mingbao Lin, Chao Chen, Liujuan Cao, Xiaoshuai Sun, and Rongrong Ji. Seqtr: A simple yet universal network for visual grounding. In European Conference on Computer Vision, pages 598-615. Springer, 2022. 2, 4, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 208, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 545, + 240 + ], + "type": "text", + "content": "[119] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR, 2021. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 241, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 308 + ], + "type": "text", + "content": "[120] Xizhou Zhu, Jinguo Zhu, Hao Li, Xiaoshi Wu, Hongsheng Li, Xiaohua Wang, and Jifeng Dai. Uni-perceiver: Pretraining unified architecture for generic perception for zero-shot and few-shot tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16804-16815, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 308, + 545, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 545, + 373 + ], + "type": "text", + "content": "[121] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15116-15127, 2023. 3, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 375, + 545, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 375, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 308, + 375, + 545, + 408 + ], + "type": "text", + "content": "[122] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. In NeurIPS, 2023. 2, 3" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "3795" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/General Point Model Pretraining with Autoencoding and Autoregressive/a64a8093-c9f6-4338-aa51-464da1052ad9_content_list.json b/2024/General Point Model Pretraining with Autoencoding and Autoregressive/a64a8093-c9f6-4338-aa51-464da1052ad9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5bdf8f82ef37c2f20a630c487b584be651dc34f8 --- /dev/null +++ b/2024/General Point Model Pretraining with Autoencoding and Autoregressive/a64a8093-c9f6-4338-aa51-464da1052ad9_content_list.json @@ -0,0 +1,1517 @@ +[ + { + "type": "text", + "text": "General Point Model Pretraining with Autoencoding and Autoregressive", + "text_level": 1, + "bbox": [ + 120, + 130, + 849, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhe Li $^{1,\\dagger}$ , Zhangyang Gao $^{3,\\dagger}$ , Cheng Tan $^{3,\\dagger}$ , Bocheng Ren $^{1}$ , Laurence T. Yang $^{1,2,*}$ Stan Z. Li $^{3,*}$ $^{1}$ Huazhong University of Science and Technology \n $^{2}$ Zhengzhou University", + "bbox": [ + 102, + 178, + 885, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ AI Lab, Research Center for Industries of the Future, Westlake University", + "bbox": [ + 197, + 233, + 792, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "keycharon0122@gmail.com, {gaozhangyang,tancheng,stan.zq.li}@westlake.edu.cn, ltyang@ieee.org", + "bbox": [ + 86, + 253, + 898, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 304, + 313, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The pre-training architectures of large language models encompass various types, including autoencoding models, autoregressive models, and encoder-decoder models. We posit that any modality can potentially benefit from a large language model, as long as it undergoes vector quantization to become discrete tokens. Inspired by the General Language Model, we propose a General Point Model (GPM) that seamlessly integrates autoencoding and autoregressive tasks in a point cloud transformer. This model is versatile, allowing fine-tuning for downstream point cloud representation tasks, as well as unconditional and conditional generation tasks. GPM enhances masked prediction in autoencoding through various forms of mask padding tasks, leading to improved performance in point cloud understanding. Additionally, GPM demonstrates highly competitive results in unconditional point cloud generation tasks, even exhibiting the potential for conditional generation tasks by modifying the input's conditional information. Compared to models like Point-BERT, MaskPoint, and PointMAE, our GPM achieves superior performance in point cloud understanding tasks. Furthermore, the integration of autoregressive and autoencoding within the same transformer underscores its versatility across different downstream tasks. Codes are available at https://github.com/gentlefress/GPM", + "bbox": [ + 76, + 335, + 473, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 724, + 209, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, the natural language processing (NLP) [4, 19, 20, 29, 40] and computer vision (CV) [3, 7, 9, 49, 57, 66] realms have witnessed a proliferation of transformer-based pretrained models. Their primary advantage lies in the inclusion of massive parameters and data in the training process, overcoming inductive biases introduced by traditional Convolutional Neural Networks [21]. Point clouds serve as fundamental data structures in fields like autonomous driving and robotics, thus emphasizing the escalating significance of tasks related to point cloud representation learn", + "bbox": [ + 75, + 750, + 468, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/708b3f6f3a8b454058fec970b4754baa20f6e95bca55be7e90f4987b1946cbe3.jpg", + "image_caption": [ + "Figure 1. Comparison of pre-training frameworks between GPM and Point-BERT [63]. Point-BERT (left) conducts the masked prediction task in an autoencoding manner, while GPM (right) combines both autoencoding and autoregressive tasks, performing tasks for both masked prediction and masked region blank filling." + ], + "image_footnote": [], + "bbox": [ + 506, + 303, + 885, + 439 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing and generation. However, the realm of pretrained point cloud models based on transformers remains relatively scarce. Existing transformer-based point cloud models [13, 65] encounter inevitable inductive biases due to local feature aggregation [13], neighbor embedding [65], and the scarcity of annotated data [9]. These biases deviate from the mainstream of standard Transformers. Self-supervised models [12, 15, 20, 25, 26, 39, 40, 61] have emerged as the dominant methodology. They excel in learning high-quality representations across modalities without extensive labeled data, models like GPT [39] even dominate in text generation. In this end, there is a compelling need to design a transformer-based point cloud model that minimizes inductive biases, learns superior point cloud representations from limited data, and simultaneously possesses the capacities for conditional and unconditional point cloud generation. The transformer-based BERT [20] has marked a significant milestone in the field of natural language processing (NLP). Its pre-training phase, involving masked prediction tasks, grants the model the capacity to learn language representations, showcasing the prowess of autoencoding language models. Given BERT's proficiency in learning language representations, the question arises: can we extend BERT's", + "bbox": [ + 496, + 553, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20954", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "capabilities to the realm of point clouds? However, since point clouds lack a direct analog to the concept of 'words' in language, constructing a discrete vocabulary for point clouds, as done in language models, is infeasible. Point clouds are composed of individual points, and considering each point as a token would lead to a quadratic increase in computational cost as the number of tokens grows. Furthermore, each point in a point cloud contains limited semantic information, necessitating the maximization of geometric information utilization. Inspired by Point-BERT [63], we embarked on a journey to vector quantization, transforming point clouds into discrete tokens. Through the application of the farthest point sampling (FPS) [36] algorithm, we segment the point cloud into multiple patches, each patch encapsulates plentiful geometric information, akin to a unit. By constructing a vocabulary specific to the point cloud domain, we obtain labels for each unit, discretizing the entire point cloud. This paves the way for the completion of the autoencoding model's masked prediction tasks in the point cloud domain.", + "bbox": [ + 76, + 90, + 472, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Transformer-based GPT [39] has dominated the field of text generation, with its core being an autoregressive language model, predicting the next token based on preceding tokens. In recent years, researchers have progressively extended this paradigm into the realm of images [7, 41, 42], ushering in a new era of image generation. In this endeavor, we bring this paradigm to the domain of point clouds, aiming to empower it for autoregressive conditional and unconditional point cloud generation tasks. After discretization, each patch of the point cloud possesses distinct geometric information, enabling it to excel in autoregressive tasks.", + "bbox": [ + 76, + 395, + 472, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While these pre-training frameworks can be adapted to the point cloud domain, they lack the flexibility to satisfy all point cloud tasks. Therefore, we aim to integrate these two pre-training frameworks from the NLP domain and apply them to the point cloud domain, depicted in Figure 1.", + "bbox": [ + 76, + 563, + 472, + 641 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we introduce a novel approach termed General Point Model (GPM) with autoencoding and autoregressive, which integrates point cloud representation learning and point cloud generation (both conditional and unconditional) within the same transformer framework. Similar to Point-BERT, we employ a point cloud tokenizer designed through dVAE-based point cloud reconstruction, where the point cloud can be discretely labeled based on the learned vocabulary. Additionally, we utilize the Masked Point Modeling (MPM) task. This involves partitioning the point cloud into distinct patches and randomly masking some adjacent patches. This aims to predict the masked portion, with the goal of learning geometric information between neighboring blocks and capturing meaningful geometric features for understanding the point cloud.", + "bbox": [ + 76, + 642, + 472, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our primary contributions can be summarized as follows:", + "bbox": [ + 76, + 869, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce autoencoding + autoregressive as a novel point cloud transformer paradigm, unifying tasks in point cloud understanding and generation.", + "- We partition the input into two segments, the first part predicts masked points (autoencoding), the second part generates masked points (autoregressive), enhancing comprehension and generation.", + "- Our approach demonstrates notable competitiveness across various point cloud understanding tasks and unconditional generation tasks. Moreover, it exhibits the potential for conditional generation, such as text- or image-conditioned point clouds generation." + ], + "bbox": [ + 500, + 90, + 893, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 287, + 640, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Self-supervised Learning", + "text_level": 1, + "bbox": [ + 500, + 311, + 728, + 329 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Self-supervised learning (SSL) has garnered significant attention owing to the substantial labor required for acquiring annotated data in large quantities. SSL enables models to learn feature representations from unlabeled data, adapting to downstream tasks. The core of SSL lies in the design of proxy tasks to replace traditional classification tasks, thereby learning feature representations. ELMo [44] employs bidirectional LSTMs [16] and generates subsequent words from left to right given the representation of preceding content. ViP [25] employs a dynamically updated momentum encoder for contrastive learning and designs a text swapping task to enhance the model's sentence representation capability. In the computer vision domain, Image-GPT [7] trains a sequence transformer to autoregressively predict pixels, showing promising representation learning capabilities without incorporating specific knowledge about the 2D input structure after pretraining. Furthermore, SSL in the field of point clouds has garnered significant attention. ACT [8] employs cross-modal autoencoders as teacher models to acquire knowledge from other modalities. The generative approaches [1, 22, 33, 45, 63, 64] are the most pertinent research to our work. Point-MAE [34] extends MAE by randomly masking point patches and reconstructing the masked regions. Point-M2AE [64] further utilizes a hierarchical transformer architecture and designs corresponding masking strategies. However, mask-based point modeling methods still face the issue of shape leakage, limiting their effective generalization to downstream tasks. In this paper, we leverage autoregressive pretraining on point clouds and address the unique challenges associated with point cloud attributes. Our concise design avoids position information leakage, thereby enhancing generalization capabilities.", + "bbox": [ + 496, + 335, + 895, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. General Language Model", + "text_level": 1, + "bbox": [ + 500, + 830, + 732, + 847 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the field of natural language processing, GLM [10] integrates both autoregressive and autoencoding methods. This enables it to perform tasks related to sentence representa", + "bbox": [ + 496, + 854, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20955", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0fa6104b24bfd09b60eb3ba44183d81ac3af200dc0b164ea659119ea5aab469a.jpg", + "image_caption": [ + "Figure 2. The framework of our GPM. We divide the entire pipeline into two stages: 1) The input point cloud is divided into several subclouds. Then, we utilize Mini-PointNet [35] to obtain a sequence of point embeddings. These embeddings are transformed into a sequence of discrete point tokens using a pre-trained tokenizer; 2) During pre-training, for PartA, certain portions of the masked embeddings are masked and replaced with [M]. They are predicted in an autoencoding manner. For PartB, masks for PartA are generated in an autoregressive manner." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 500, + 296 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/811f0447454e5f41f7d95a3bd47d25a4b52e82cd4fe638ddd16f4d29061de711.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 90, + 890, + 294 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tion as well as conditioned and unconditioned generation through fine-tuning. To the best of our knowledge, we are the first to integrate autoregressive and autoencoding techniques on a single transformer in the domain of point clouds. In downstream tasks, we have the potential to fine-tune point cloud representations and perform unconditioned generation even in zero-shot scenarios, with the possibility of fine-tuning towards conditioned generation.", + "bbox": [ + 75, + 402, + 470, + 525 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 76, + 539, + 174, + 554 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this study, we endeavor to seamlessly integrate BERT-style and GPT-style pre-trained strategies, extending their applicability to point cloud transformers. Following the Point-BERT[63], our training process unfolds in two distinctive stages. Firstly, we embark on training a specialized Tokenizer, a critical step towards acquiring discrete point labels for each input point cloud. Concurrently, a dedicated discrete VAE (dVAE) is employed to masterfully reconstruct the discrete point cloud. The subsequent stage involves the splicing of two sequences of point cloud embeddings, partA for autoencoding and partB for autoregressive. The overall idea of our approach is illustrated in Figure 2.", + "bbox": [ + 75, + 564, + 468, + 746 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Stage 1: Discrete Varitional Autoencoder Pretraining", + "text_level": 1, + "bbox": [ + 76, + 755, + 468, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Point Cloud Partitioning. We posit that discrete tokens derived from points encapsulate crucial geometric information, allowing for a discrete representation of any point cloud. However, a naive strategy of assigning one token to each point in a point-wise reconstruction task poses a formidable computational challenge. This arises from the quadratic complexity of self-attention within transformers.", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Building upon the foundations laid by Point-BERT [63] and ViT[9], we adopt a strategy of partitioning each point cloud into distinct patches, each serving as a single token. Specifically, for an input point cloud $p \\in \\mathbb{R}^{\\mathcal{N} \\times 3}$ , we initiate the process by extracting $m$ center points $\\mathcal{C}_u \\in \\mathbb{R}^{m \\times 3}$ from the point cloud $p$ through farthest point sampling (FPS). Subsequently, we employ the k-nearest neighbor (kNN) algorithm to identify the $k$ nearest neighboring points for every center point. This forms $m$ localized patches or sub-clouds denoted as $\\mathcal{P}_u \\in \\mathbb{R}^{m \\times k \\times 3}$ . To ensure these patches are free of bias, we normalize them by subtracting their respective center coordinates. This operation effectively disentangles the structural patterns from the spatial coordinates inherent to each local patch.", + "bbox": [ + 496, + 402, + 893, + 616 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Point Cloud Embedding. Following Point-BERT [63], we employ a mini-PointNet [35] to embed the point patches as a sequence of point embedding $\\{h_i\\}_{i=1}^m$ . For the further vector quantization, we pre-define a learnable codebook $\\mathcal{V} := \\{(s, z_s)\\}_{s \\in S}^{d_z}$ , where $d_z$ is the dimension of codes, $S$ is the size of codebook and $s$ is the index of embedding in $\\mathcal{V}$ . We adopt DGCNN [53] as tokenizer $\\mathcal{Q}: h_i \\mapsto z_i$ , which maps $\\{h_i\\}_{i=1}^m$ into $\\{z_i\\}_{i=1}^m$ in $\\mathcal{V}$ . However, given the non-differentiable nature of the discrete tokens, applying reparameterization gradients for dVAE training becomes unfeasible. As suggested in [41], we resort to the Gumbel-softmax relaxation technique [18], coupled with a uniform prior, as a workaround during the dVAE training process.", + "bbox": [ + 496, + 636, + 893, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Point Cloud Reconstruction. Given $\\{z_i\\}_{i=1}^m$ as input, the decoder $\\mathcal{D}(\\cdot)$ is tasked with the reconstruction of the entire point cloud. Inspired by [63], to effectively capture inter", + "bbox": [ + 496, + 854, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20956", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "point relationships and bolster the representational capacity of discrete point tokens across various local structures, we adopt the DGCNN [53]-FoldingNet [60] architecture to reconstruct the whole point cloud.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The overarching reconstruction objective can be denoted as $\\mathbb{E}_{z\\sim \\mathcal{Q}(z|h)}\\log D(\\hat{p} |z)$ . Inspired by Point-BERT [63], the entire reconstruction process can be conceptualized as the maximization of the Evidence Lower Bound (ELB) for the log-likelihood, denoted as $\\mathcal{D}(\\hat{p} |p)$ :", + "bbox": [ + 75, + 151, + 468, + 227 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\sum_ {i = 1} ^ {\\mathcal {N}} \\log D (\\hat {p} _ {i} | p _ {i}) \\geq \\sum_ {i = 1} ^ {\\mathcal {N}} \\left(\\mathbb {E} _ {z _ {i}} \\sim \\mathcal {Q} (z _ {i} | p _ {i})\\right) [ \\log D (p _ {i} | z _ {i}) ] \\\\ - \\operatorname {K L} \\left[ \\mathcal {Q} \\left(z \\mid p _ {i}\\right), \\mathcal {D} \\left(z \\mid \\hat {p} _ {i}\\right) \\right]. \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 234, + 478, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Simultaneously, we account for an intuitive reconstruction loss, leveraging the $l1$ Chamfer Distance [11] to supervise prediction point cloud with the ground-truth point cloud:", + "bbox": [ + 75, + 309, + 468, + 354 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C D}} (\\mathcal {P}, \\mathcal {G}) = \\frac {1}{| \\mathcal {P} |} \\sum_ {p \\in \\mathcal {P}} \\min _ {g \\in \\mathcal {G}} \\| p - g \\| + \\frac {1}{| \\mathcal {G} |} \\sum_ {g \\in \\mathcal {G}} \\min _ {p \\in \\mathcal {P}} \\| g - p \\|, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 363, + 468, + 412 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{P}$ denotes the predicted point clouds and $\\mathcal{G}$ is the ground truth point clouds. Additional, we follow [41], optimize the KL-divergence $\\mathcal{L}_{\\mathrm{KL}}$ between the generated point cloud distribution and a uniform distribution:", + "bbox": [ + 75, + 414, + 468, + 473 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {d V A E}} = \\mathcal {L} _ {\\mathrm {C D}} + \\mathcal {L} _ {\\mathrm {K L}}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 199, + 486, + 468, + 501 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Stage 2: GPM Pre-training", + "text_level": 1, + "bbox": [ + 76, + 511, + 321, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To ensure GPM excel in both point cloud representation learning and point cloud generation task, we formulate a training framework that amalgamates autoencoding and autoregressive techniques motivated by [10].", + "bbox": [ + 75, + 534, + 468, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Autoencoding Masked Sequence Generation. We employ the standard Transformers [51], which encompass multi-headed self-attention layers and FFN blocks. Initially, we partition each input point cloud into $m$ local patches with designated center points. Subsequently, these local patches undergo discretization into a code sequence $\\{f_i\\}_{i=1}^m$ via our pre-trained dVAE in stage 1. Formally, the input embeddings $\\{e_i\\}_{i=1}^m$ are constructed as a combination of point embeddings $\\{f_i\\}_{i=1}^m$ and positional embeddings $\\{pos_i\\}_{i=1}^m$ . Adhering to language model [20], a class token [CLS] is appended to the input sequences, rendering the transformer input sequence as $I = \\{[\\mathrm{CLS}], e_1^0, e_2^0, \\dots, e_i^0\\}$ .", + "bbox": [ + 75, + 613, + 468, + 795 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Drawing inspiration from [63], we initiate by selecting a central point $\\mathcal{C}_i$ alongside its corresponding sub-cloud $\\mathcal{P}_i$ , integrate it with $m$ neighboring sub-clouds as a coherent local region. Within this region, we apply a masking operation to obscure all local patches, generating what we refer to as the \"masked point cloud\". Specifically, we substitute all the masked point embeddings with a universally learnable", + "bbox": [ + 75, + 795, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "pre-defined mask embeddings [M]. We mark the masked position as $\\mathcal{M} = \\{1,\\dots,b\\}$ , and the final input embeddings $E^{\\mathcal{M}} = \\{e_i,i\\notin \\mathcal{M}\\} \\cup \\{[\\mathbf{M}] + pos_i,i\\in \\mathcal{M}\\}$ are fed into the transformer.", + "bbox": [ + 498, + 90, + 890, + 150 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The objective of our Masked Point Modeling (MPM) task is to deduce the geometric arrangement of absent portions and reconstruct point tokens aligned with the masked positions using the available information. Geometric information refers to the relative positions and morphological attributes of points within a localized region. These crucial geometric details provide vital cues for understanding and processing the shape and structure of the local region, playing a pivotal role in delving deep into the intricacies of point cloud data. Formally, the pre-training objective can be expressed as maximizing the log-likelihood of the point tokens $e_i$ conditioned on the masked input embeddings $E^{\\mathcal{M}}$ :", + "bbox": [ + 496, + 151, + 892, + 332 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\max \\sum_ {E \\in W} \\mathbb {E} \\left[ \\sum_ {i \\in \\mathcal {M}} \\log P \\left(e _ {i} \\mid E ^ {\\mathcal {M}}\\right) \\right], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 342, + 890, + 380 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $W$ is the set of all input embeddings. Similar to [63], we also adopt Point Patch Mixing and contrastive learning to help the model to better learn high-level semantics. Point Patch Mixing involves mixing sub-clouds without complex alignment techniques. However, training solely with masked precition has limitations in understanding high-level semantics. To overcome this, we incorporate contrastive learning to enhance semantic understanding. Assuming the masking ratio is $\\beta$ , the loss function is defined as:", + "bbox": [ + 496, + 388, + 890, + 540 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n- \\beta \\log \\frac {\\exp \\left(q \\hat {k} _ {1} / \\tau\\right)}{\\sum_ {i = 0} ^ {K} \\exp \\left(q k _ {i}\\right)} - (1 - \\beta) \\log \\frac {\\exp \\left(q \\hat {k} _ {2} / \\tau\\right)}{\\sum_ {i = 0} ^ {K} \\exp \\left(q k _ {i}\\right)}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 547, + 890, + 587 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $q$ is the feature of a mixed sample that comes from two other samples, whose features are $\\hat{k_1}$ and $\\hat{k_2}$ . $k_{i}$ are extracted by the momentum feature encoder, $\\tau$ is the temperature and $K$ is the size of memory bank.", + "bbox": [ + 496, + 595, + 890, + 656 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Autoregressive Sequence Generation. While Point-BERT models the relationship between unmasked and masked regions using the Masked Point Modeling (MPM) task, it does not adequately model the interactions within masked regions. Furthermore, our aim extends beyond acquiring high-quality point cloud representations; we also aspire to proficiently execute point cloud generation tasks. Hence, for the latter segment, PartB, we engage in autoregressive mask generation tasks to enhance the MPM task.", + "bbox": [ + 496, + 672, + 890, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the case of PartB, the mask operation is ignored. Our focus lies in acquiring the generated tokens from the masked positions of PartA. We take the initial $n - 1$ tokens from PartA, appending the start token [S] at its forefront to get the PartB $E^{\\mathcal{B}} = \\{[\\mathrm{S}], e_1, \\dots, e_{n-1}\\}$ . Following the autoregressive approach, the prediction of the next token relies", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "footer", + "text": "20957", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "on the preceding token:", + "bbox": [ + 76, + 90, + 235, + 104 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {\\theta} \\mathbb {E} \\left[ \\sum_ {i = 1} ^ {n} \\log P _ {\\theta} \\left(e _ {i} | \\mathrm {P a r t A}, e _ {< i}, [ \\mathrm {S} ]\\right) \\right]. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 114, + 468, + 154 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although no masking operation is performed on PartB, since the mask positions of PartA are multiple consecutive sub-clouds instead of a single one, in PartB we are also equivalent to autoregressive generation of a continuous mask span. We extract the generated tokens from PartB, corresponding to the masked positions in PartA, and compare them with the discrete tokens produced by the Tokenizer:", + "bbox": [ + 75, + 160, + 468, + 280 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {A R} = \\operatorname {C r o s s E n t r o p y} \\left(e _ {j, j \\in \\mathcal {M}}, \\operatorname {a r g m a x} \\mathcal {Q} (h) _ {j, j \\in \\mathcal {M}}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 290, + 468, + 310 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this way, it allows our model to acquire proficiency in an autoencoding bidirectional encoding scheme for PartA, and an autoregressive unidirectional encoder for PartB.", + "bbox": [ + 76, + 314, + 468, + 359 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Multi-task Pretraining", + "text_level": 1, + "bbox": [ + 76, + 367, + 287, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The masking of $15\\%$ of tokens in BERT is tailored for downstream natural language understanding tasks, whereas Point-BERT masks $25\\% \\sim 45\\%$ to acquire enhanced point cloud representations. Our GPM focus on both point cloud representation learning and point cloud generation, we concatenate PartA and PartB as input to the transformer, aiming to simultaneously perform both autoencoding and autoregressive tasks. Therefore, we need to establish specific attention masks to facilitate the interaction of information between these two segments.", + "bbox": [ + 75, + 390, + 468, + 541 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Motivated by [10], the tokens in PartA can participate in MLM task in PartA and autoregressive generation in PartB. PartB tokens can not be observed by PartA, but those antecedents. PartA is treated as the conditioning for autoregressive generation in PartB, akin to prefix tuning [23] in NLP. The implement of GPM multi-task pretraining is depicted in Figure 3.", + "bbox": [ + 75, + 542, + 468, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The tokens output by the pretrained tokenizer $\\mathcal{Q}(\\cdot)$ are regarded as labels. PartA conducts a masked prediction task in an autoencoding manner, and the prediction loss between the predicted masked portion and the labels is computed as follows:", + "bbox": [ + 75, + 648, + 468, + 722 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {A E} = \\operatorname {C r o s s E n t r o p y} \\left(e _ {j, j \\in \\mathcal {M}}, \\operatorname {a r g m a x} \\mathcal {Q} (h) _ {j, j \\in \\mathcal {M}}\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 732, + 468, + 752 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although Equation 7 and 8 exhibit identical formal structures. However, one entails mask prediction through autoencoding, while the other involves mask generation in an autoregressive way, indicating a fundamental distinction between the two approaches.", + "bbox": [ + 75, + 757, + 468, + 832 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 845, + 207, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we commence by outlining the configurations of our pretraining scheme. Subsequently, we pro", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/dce3458aaa424daa59ba179a4a38dbf740e92d73722392201881b9ccfd2d985f.jpg", + "image_caption": [ + "Figure 3. Self-attention mask. The regions in grey are masked. Tokens in PartA can attend to themselves, but not to PartB. Tokens in PartB can attend to PartA and their antecedents in PartB." + ], + "image_footnote": [], + "bbox": [ + 578, + 88, + 815, + 281 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ceed to assess the proposed model through various downstream tasks, encompassing object classification, partial segmentation, few-shot learning, transfer learning (to validate the model's representation learning capabilities), and point cloud generation (to validate its generative capabilities). Additionally, we conduct ablation studies on GPM.", + "bbox": [ + 496, + 361, + 890, + 452 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Pre-training Setups and Implementation", + "text_level": 1, + "bbox": [ + 498, + 462, + 846, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pre-training Data. Following the dataset configuration similar to Point-BERT [63], we employ ShapeNet [5] as our pretraining dataset, encompassing over 50,000 unique 3D models spanning 55 common object categories. From each 3D model, we sample 1024 points, dividing them into 64 point patches (sub-clouds) and each sub-cloud contains 32 points. Utilizing a lightweight PointNet [35] with two MLP layers, we project each sub-cloud into a 64-point embedding, serving as input for both dVAE and Transformer [51].", + "bbox": [ + 496, + 486, + 890, + 637 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "dVAE Pretraining Setups. The purpose of pretraining the dVAE is to acquire a high-quality tokenizer, enabling the reconstruction of discretized features back into the original point cloud to the fullest extent. Our dVAE is composed of a tokenizer and a decoder. To be specific, the tokenizer consists of a 4-layer DGCNN [53], while the decoder encompasses a 4-layer DGCNN followed by a FoldingNet [60]. Additionally, the FoldingLayer establishes a connection between a 2D grid and the input, ultimately generating a 3D point cloud.", + "bbox": [ + 496, + 657, + 890, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During the training of dVAE [41], we set the number of words in the codebook $N$ to 8192. We employ the common $l1$ -style Chamfer Distance loss in the reconstruction process. Due to the small numerical value of this $l1$ loss, the weight of the KL loss in Equation 1 must be smaller than that in image tasks. In the initial 10,000 steps of training,", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "footer", + "text": "20958", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the weight of the KL loss is kept at 0 and gradually raised to 0.1 in the subsequent 100,000 steps. Our learning rate is set to 0.0005, and follows a cosine annealing schedule spanning 60,000 steps. Consistent with prior works [41] and Point-BERT [63], we decay the temperature in the Gumblesoftmax function from 1 to 0.0625 over 100,000 steps. The training of dVAE encompasses a total of 150,000 steps, with a batch size of 64. It's worth noting that our dVAE's network architecture closely mirrors that of Point-BERT.", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "GPM Pretraining Setups. In our experiments, we adhere to the standard Transformer architecture [9], which consists of a stack of Transformer blocks [51], each composed of multi-head self-attention layers and feed-forward networks. LayerNorm is applied in both layers. We set the depth of our Transformer to 12, feature dimension to 384, and the number of heads to 6. A random depth of 0.1 [17] is applied in our Transformer encoder. During GPM pretraining, the weights of the Tokenizer learned by dVAE are kept fixed. Input point embeddings from $25\\% \\sim 45\\%$ are randomly masked, and the model is trained to infer the expected point labels at these masked positions. Unlike Point-BERT [63], which does not have a generation task during pretraining, we incorporate both autoregressive and autoencoding tasks in our work. Thus, an attention mask is necessary, as illustrated in Figure 3. For MoCo, we set the same configuration as [63]. More Details can be found in the supplementary 6.1 and 6.2.", + "bbox": [ + 75, + 247, + 472, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Downstream Tasks", + "text_level": 1, + "bbox": [ + 76, + 530, + 261, + 544 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.1 Point Cloud Representation Evaluation", + "text_level": 1, + "bbox": [ + 76, + 553, + 405, + 569 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Object classification on ModelNet40. We evaluate our pre-trained model on the ModelNet40 dataset [55], comprising 12,311 clean 3D CAD models spanning 40 categories. Following the Point-BERT [63] setups, we employ a two-layer MLP with a dropout rate of 0.5 as the classification head for the task. We optimize the model using AdamW with a weight decay of 0.05 and a learning rate of 0.0005, while employing a batch size of 32 and a cosine annealing schedule. We conduct comparisons with various Transformer-based models with identical Point-BERT settings, denoting [ST] for the standard Transformer architecture and [T] for Transformer models with specific designs or inductive biases. The results in Table 1 demonstrate that our model not only outperforms Point-BERT in classification metrics on this dataset but also exhibits autoregressive generation capabilities.", + "bbox": [ + 75, + 578, + 470, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Object classification on ScanObjectNN. The practical applicability of our model on real datasets serves as a crucial metric. Therefore, the pre-trained models are transferred to the ScanObjectNN dataset [50], which comprises", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0e95fac268c980147db8531cd041fb7385e301819d814a8942c88f50f55d7015.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Methodsnumber of pointsAcc
Supervised Learning
PointNet[35]1k89.2
SO-Net[22]1k92.5
PointNet++ [36]1k90.5
PointCNN[24]1k92.2
DGCNN[53]1k92.9
MVTN[14]1k93.8
RSCNN [43]1k92.9
GBNet[38]1k93.8
PointMLP[32]1k94.5
DensePoint[28]1k92.8
PointNeXt[37]1k94.0
P2P-RN101[54]1k93.1
P2P-HorNet[54]1k94.0
KPConv[48]~6.8k92.9
Self-Supervised Representation Learning
[T]PCT [65]1k93.2
[T]PointTransformer [13]-93.7
[ST]NPCT [13]1k91.0
[ST]Transformer[51]1k91.4
[ST]Transformer-OcCo[51]1k92.1
[ST]Point-BERT[63]1k93.2
[ST]MaskPoint[27]1k93.8
[ST]Point-MAE[34]1k93.8
[ST]GPM1k94.0
Methods with Cross-modal Information and Teacher Models
ACT[8]1k93.7
[ST]Transformer[51]4k91.2
[ST]Transformer-OcCo[51]4k92.2
[ST]Point-BERT[63]4k93.4
[ST]GPM4k93.1
[ST]Point-BERT[63]8k93.8
[ST]MaskPoint[27]8k-
[ST]Point-MAE[34]8k94.0
[ST]GPM8k94.3
", + "bbox": [ + 504, + 88, + 893, + 460 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Classification results on ModelNet40. All results are expressed as percentages accuracy.", + "bbox": [ + 500, + 470, + 890, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "approximately 15,000 objects extracted from real indoor scans, encompassing 2902 point clouds from 15 categories. This dataset presents a greater challenge as it involves sampling from real-world scans with backgrounds and occlusions. We conducted experiments on three main variants, namely OBJ-BG, OBJ-ONLY, and PB-T50-RS, in line with prior work. The experimental results are summarized in Table 2. It is observed that GPM exhibits an improvement of approximately $2.77\\%$ , $1.88\\%$ , and $1.73\\%$ over the regular Point-BERT across the three variants.", + "bbox": [ + 496, + 527, + 893, + 679 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Few-shot learning. To demonstrate the ability to acquire knowledge for new tasks with limited training data, we evaluate our model under the setting of few-shot learning, following the methodology of previous work [47, 63]. In the typical 'W-way S-shot' setup, we initially randomly select $W$ classes and then sample $(S + 20)$ objects for each class [46]. The model is trained on $W \\times S$ samples (support set) and evaluated on the remaining $20W$ samples (query set). We conduct 10 independent experiments for each setting and report the average performance and standard deviation across the 10 runs. As shown in Table 3, our approach outperforms other methods in all tests, achieving absolute improvements of $0.4\\%$ , $0.4\\%$ , and $0.6\\%$ over Point-BERT.", + "bbox": [ + 496, + 704, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20959", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/761df663db1b236daae64593b469b68d8f1acbe88334348343741fbfa9afb3c9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsOBJ-BGOBJ-ONLYPB-T50-RS
Supervised Learning
PointNet[35]73.379.268.0
SpiderCNN[58]77.179.573.7
PointNet++ [36]82.384.377.9
PointCNN[24]86.185.578.5
DGCNN[53]82.886.278.1
MVTN[14]92.692.382.8
BGA-DGCNN[50]--79.7
BGA-PN++ [50]--80.2
GBNet[38]--81.0
PointMLP[32]--85.4
PointNeXt[37]--87.7
P2P-RN101[54]--87.4
P2P-HorNet[54]--89.3
Self-Supervised Representation Learning
Transformer[51]79.8680.5577.24
Transformer-OcCo[51]84.8585.5478.79
Point-BERT[63]87.4388.1283.07
MaskPoint[27]89.388.184.3
Point-MAE[34]90.088.285.2
GPM90.290.084.8
", + "bbox": [ + 81, + 88, + 472, + 333 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8425c7a0e45d0f18be1c684937dff7eae8baabf98e6a1aa0d7a743b4e34bdfd3.jpg", + "table_caption": [ + "Table 2. Classification results on ScanObjectNN. All results are expressed as percentages accuracy." + ], + "table_footnote": [], + "table_body": "
Methods5-way10-way
10-shot20-shot10-shot20-shot
Supervised Learning
DGCNN-rand [53]31.6 ± 2.840.8 ± 4.619.9 ± 2.116.9 ± 1.5
OcCo [52]90.6 ± 2.892.5 ± 1.982.9 ± 1.386.5 ± 2.2
Self-Supervised Representation Learning
Point-BERT [63]94.6 ± 3.196.3 ± 2.791.0 ± 5.492.7 ± 5.1
MaskPoint [27]95.0 ± 3.797.2 ± 1.791.4 ±4.093.4 ± 3.5
Piont-MAE [34]96.3 ± 2.597.8 ± 1.892.6 ± 4.195.0 ± 3.0
Piont-M2AE [64]96.8 ± 1.898.3 ± 1.492.3 ± 4.595.0 ± 3.0
GPM97.2 ± 2.698.7 ± 2.292.9 ± 4.295.0 ± 3.0
", + "bbox": [ + 81, + 385, + 468, + 511 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Part segmentation. We evaluate the representation learning capability of our approach on the ShapeNetPart dataset [62], aiming to predict more fine-grained class labels for each point. This dataset consists of 16 categories and comprises 16,881 objects. The point cloud is downsampled to 2048 points, and the segmentation head [34] connects features $\\mathcal{F}^4, \\mathcal{F}^8, \\mathcal{F}^{12}$ extracted from the 4-th, 6-th, and 12-th layers of the transformer blocks. Subsequently, average pooling, max pooling, and upsampling are employed to generate features for each point, followed by label prediction using MLP. In our experiments, we evaluate the performance under the settings of \"5 way 10 shot\", \"5 way 20 shot\", \"10 way 10 shot\", and \"10 way 20 shot\". The experimental results shown in Table 4 demonstrate the superior performance of our GPM compared to all other methods.", + "bbox": [ + 75, + 604, + 468, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.2 Point Cloud Generation Evaluation", + "text_level": 1, + "bbox": [ + 76, + 832, + 377, + 845 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the pre-training phase, we perform unconditional autoregressive generation on tokens in PartB, endowing the model with the capability of point cloud unconditional generation.", + "bbox": [ + 75, + 854, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/92a9fac1ce9381843b43a0eff820e045a35f33da624d9eae13b22fcfafc14f25.jpg", + "table_caption": [ + "Table 3. The results of few-shot classification on the ModelNet40 dataset. For each experimental setting, we conduct 10 independent experiments and report the average accuracy $(\\%)$ along with its standard deviation." + ], + "table_footnote": [], + "table_body": "
MethodsCls.mIoUInst.mIoU
Supervised Learning
PointNet[35]80.3983.7
PointNet++ [36]81.3585.1
DGCNN[53]82.3385.2
PointMLP [32]84.686.1
Self-Supervised Representation Learning
Transformer[51]83.4285.1
Transformer-OcCo[51]83.4285.1
PointContrast [56]-85.1
CrossPoint [2]-85.5
Point-BERT [63]84.1185.6
Point-MAE [34]-86.1
GPM84.2085.8
", + "bbox": [ + 504, + 88, + 893, + 287 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9f59f3d923e8796735367fd03869045c8afeb2263e1cb38fd16c5576d8e4a675.jpg", + "table_caption": [ + "Table 4. Part segmentation results on the ShapeNetPart dataset. We report the average intersection mIoU over the union of all classes (Cls.) and instances (Inst.)." + ], + "table_footnote": [], + "table_body": "
CategoryModelJSD (↓)MMD (↓)Cov (%, ↑)1-NNA % (%, ↓)
CDEMDCDEMDCDEMD
AirplanePoint-Flow4.920.2173.2446.9148.4075.6875.06
Point-BERT4.070.1892.7550.1753.4470.1661.07
GPM3.320.1762.4752.3456.4468.2260.30
ChairPoint-Flow1.742.427.8746.8346.9860.8859.89
Point-BERT1.231.827.1149.7350.0757.3056.61
GPM0.981.676.8251.0852.1454.0954.27
CarPoint-Flow0.870.915.2244.0346.5960.6562.36
Point-BERT0.730.824.9647.2752.1359.0960.48
GPM0.660.754.3849.6754.1758.8858.16
", + "bbox": [ + 503, + 354, + 890, + 454 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Generation results on ShapeNet dataset. $\\uparrow$ : the higher the better, $\\downarrow$ : the lower the better. MMD-CD is multiplied by $10^{3}$ ; MMD-EMD and JSD are multiplied by $10^{2}$ .", + "bbox": [ + 498, + 464, + 890, + 506 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Therefore, in downstream tasks, without fine-tuning, we conduct an autoregressive point cloud generation task on the ShapeNet dataset. Existing models do not possess both point cloud generation and point cloud classification tasks. Hence, our model is the first known point cloud transformer that integrates these two tasks into one. The specific metrics are shown in Table 5. We compare our results with generation model Point-Flow [59] and Point-BERT. More details about downstream tasks and conditional point cloud generation results can be found in the supplementary 6.3.", + "bbox": [ + 496, + 534, + 890, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 695, + 653, + 710 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To validate the fundamental design of the GPM model, we conducted a comprehensive set of ablation studies. In order to assess the impact of autoregressive mask generation on the autoencoding masked prediction task, we present results comparing fine-tuning scenarios on the ModelNet40 dataset: one with autoencoding only, and the other with a combination of autoencoding + autoregressive. Furthermore, to demonstrate that the order of autoencoding and autoregressive tasks in the sequence does not affect point cloud understanding, we conduct additional experiments, the results are summarized in the Table 6. More ablation results are shown in the supplementary 6.4.", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20960", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a8f2f5445a6e4bb216661e0a0b6ff10c748c5c72104235068ee35b6edc9d2f14.jpg", + "image_caption": [ + "(a) Point cloud generation results of dVAE." + ], + "image_footnote": [], + "bbox": [ + 81, + 68, + 444, + 198 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9d24243c1e5ac04edb415f0ce15298a8bba6eff659116e964425f2e82847a49a.jpg", + "image_caption": [ + "(b) Point cloud generation results of Point-BERT.", + "Figure 4. The reconstruction results visualization on dVAE and Point-BERT." + ], + "image_footnote": [], + "bbox": [ + 531, + 66, + 883, + 199 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e5f6a8ae075afcdec45b4ecef00e03f635052741a104ca2e7a73ce04d76e8a57.jpg", + "table_caption": [], + "table_footnote": [ + "Table 6. Results on autoencoding, autoencoding $\\Rightarrow$ autoregressive and autoregressive $\\Rightarrow$ autoencoding on the ModelNet dataset. It is clear that autoregressive can enhance the completion accuracy of autoencoding, the order of autoencoding and autoregressive does not have a significant impact on point cloud understanding tasks." + ], + "table_body": "
TasksnpointsAcc
autoencoding only1k92.9
autoregressive ⇒ autoencoding1k93.6
autoencoding ⇒ autoregressive1k93.8
autoencoding only4k93.4
autoregressive ⇒ autoencoding4k93.9
autoencoding ⇒ autoregressive4k93.9
", + "bbox": [ + 81, + 263, + 472, + 452 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Analysis", + "text_level": 1, + "bbox": [ + 76, + 561, + 179, + 577 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Compare with Point-BERT. Motivated by BERT, Point-BERT leverages MPM as a pretext task for training. Due to the assumption of independence between MLM and MPM, Point-BERT fails to capture interdependencies among masked tokens. In our approach, autoregressive generation explicitly models dependencies between masked regions, unlike autoencoding which relies solely on un-masked tokens to predict the masked ones. Autoregressive masked prediction allows previously predicted tokens to influence subsequent predictions, enhancing information exchange between them. This leads to improved performance in downstream tasks in the realm of representation learning. Simultaneously, tokens in PartA can attend to task in PartB, making PartA a generation condition for PartB, resulting in enhanced generation performance.", + "bbox": [ + 75, + 585, + 468, + 813 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Compare with PointGPT. PointGPT [6] adopts a two-stage transformer architecture for learning representations of point clouds, comprising a feature extractor and a generator. It first extracts features from the point cloud us", + "bbox": [ + 75, + 839, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ing the feature extractor and then directly generates with the obtained features in the generator. During training, it directly employs Chamfer Distance [11] between generated point clouds and ground truth points as the optimization objective, rather than operating on tokens. This is because it doesn't discretize points embeddings, thus failing to achieve genuine autoregressive generation tasks. Moreover, it doesn't integrate the two tasks for joint training in a single transformer; instead, it trains them separately on two different transformers, which increases the complexity.", + "bbox": [ + 496, + 265, + 893, + 417 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Visualization", + "text_level": 1, + "bbox": [ + 500, + 426, + 637, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We visualize the generation results of dVAE and Point-BERT, as shown in the Figure 4a and Figure 4b. Although Point-BERT is capable of masked tokens generation through autoencoding, our approach allows for autoregressive conditional generation tasks, such as text-conditioned point clouds and image-conditioned point clouds. As observed in Figure 4b, the generation results of Point-BERT exhibit some scattered points at the edges of the overall structure, appearing relatively sparse and disorganized. The generated outcome is not as satisfactory as that of GPM. More visualizations of generation results can refer to Unconditional and Conditional Point Generation in the supplementary 6.3.", + "bbox": [ + 496, + 449, + 893, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 659, + 617, + 674 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our study integrates both autoencoding and autoregressive tasks within a unified transformer to enhance the model's performance in representation learning and generation at the same time. By this framework, we address the limitation of current pretraining frameworks and make the proxy tasks non-trivial, effectively enhance the information interaction among masked tokens. Moreover, our framework endows the model with the potential for conditional generation.", + "bbox": [ + 496, + 685, + 893, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 819, + 660, + 837 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by National Science and Technology Major Project (No.2022ZD0115101), National Natural Science Foundation of China Project (No.U21A20427).", + "bbox": [ + 496, + 844, + 893, + 887 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20961", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In International conference on machine learning, pages 40-49. PMLR, 2018. 2", + "[2] Mohamed Afham, Isuru Dissanayake, Dinithi Dissanayake, Amaya Dharmasiri, Kanchana Thilakarathna, and Ranga Rodrigo. Crosspoint: Self-supervised cross-modal contrastive learning for 3d point cloud understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9902-9912, 2022. 7", + "[3] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. In International Conference on Learning Representations, 2021. 1", + "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1", + "[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 5", + "[6] Guangyan Chen, Meiling Wang, Yi Yang, Kai Yu, Li Yuan, and Yufeng Yue. Pointgpt: Auto-regressively generative pre-training from point clouds. arXiv preprint arXiv:2305.11487, 2023. 8", + "[7] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International conference on machine learning, pages 1691-1703. PMLR, 2020. 1, 2", + "[8] Runpei Dong, Zekun Qi, Linfeng Zhang, Junbo Zhang, Jianjian Sun, Zheng Ge, Li Yi, and Kaisheng Ma. Autoencoders as cross-modal teachers: Can pretrained 2d image transformers help 3d representation learning? In The Eleventh International Conference on Learning Representations, 2022. 2, 6", + "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 1, 3, 6", + "[10] Zhengxiao Du, Yujie Qian, Xiao Liu, Ming Ding, Jiezhong Qiu, Zhilin Yang, and Jie Tang. Glm: General language model pretraining with autoregressive blank infilling. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 320-335, 2022. 2, 4, 5", + "[11] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 605-613, 2017. 4, 8", + "[12] Tianyu Gao, Xingcheng Yao, and Danqi Chen. Simcse: Simple contrastive learning of sentence embeddings. In 2021" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Empirical Methods in Natural Language Processing, EMNLP 2021, pages 6894-6910. Association for Computational Linguistics (ACL), 2021. 1", + "[13] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. Pct: Point cloud transformer. Computational Visual Media, 7:187-199, 2021. 1, 6", + "[14] Abdullah Hamdi, Silvio Giancola, and Bernard Ghanem. Mvtn: Multi-view transformation network for 3d shape recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1-11, 2021. 6, 7", + "[15] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 1", + "[16] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997. 2", + "[17] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q Weinberger. Deep networks with stochastic depth. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 646-661. Springer, 2016. 6", + "[18] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. In International Conference on Learning Representations, 2016. 3", + "[19] Mandar Joshi, Danqi Chen, Yinhan Liu, Daniel S Weld, Luke Zettlemoyer, and Omer Levy. Spanbert: Improving pre-training by representing and predicting spans. Transactions of the association for computational linguistics, 8:64-77, 2020. 1", + "[20] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, pages 4171–4186, 2019. 1, 4", + "[21] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012. 1", + "[22] Jiaxin Li, Ben M Chen, and Gim Hee Lee. So-net: Self-organizing network for point cloud analysis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9397–9406, 2018. 2, 6", + "[23] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 4582-4597, 2021. 5", + "[24] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. Advances in neural information processing systems, 31, 2018. 6, 7", + "[25] Zhe Li, T. Yang Laurence, Xin Nie, BoCheng Ren, and Xianjun Deng. Enhancing sentence representation with visually-supervised multimodal pre-training. In Proceedings of the 31st ACM International Conference on Multimedia (ACM MM'23), 2023. 1, 2" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20962", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Zhe Li, Laurence T Yang, Bocheng Ren, Xin Nie, Zhangyang Gao, Cheng Tan, and Stan Z Li. Mlip: Enhancing medical visual representation with divergence encoder and knowledge-guided contrastive learning. arXiv preprint arXiv:2402.02045, 2024.1", + "[27] Haotian Liu, Mu Cai, and Yong Jae Lee. Masked discrimination for self-supervised learning on point clouds. In European Conference on Computer Vision, pages 657-675. Springer, 2022. 6, 7", + "[28] Yongcheng Liu, Bin Fan, Gaofeng Meng, Jiwen Lu, Shiming Xiang, and Chunhong Pan. Densepoint: Learning densely contextual representation for efficient point cloud processing. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5239-5248, 2019. 6", + "[29] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 1", + "[30] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. In International Conference on Learning Representations, 2016. 1", + "[31] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 1", + "[32] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. In International Conference on Learning Representations, 2021. 6, 7", + "[33] Chen Min, Dawei Zhao, Liang Xiao, Yiming Nie, and Bin Dai. Voxel-mae: Masked autoencoders for pre-training large-scale point clouds. arXiv preprint arXiv:2206.09900, 2022. 2", + "[34] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In European conference on computer vision, pages 604-621. Springer, 2022. 2, 6, 7", + "[35] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 3, 5, 6, 7", + "[36] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 2, 6, 7", + "[37] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. Advances in Neural Information Processing Systems, 35:23192-23204, 2022. 6, 7", + "[38] Shi Qiu, Saeed Anwar, and Nick Barnes. Geometric backprojection network for point cloud classification. IEEE Transactions on Multimedia, 24:1943-1955, 2021. 6, 7", + "[39] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 1, 2" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 1", + "[41] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2, 3, 4, 5, 6, 1", + "[42] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2", + "[43] Yongming Rao, Jiwen Lu, and Jie Zhou. Global-local bidirectional reasoning for unsupervised representation learning of 3d point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5376-5385, 2020. 6", + "[44] Justyna Sarzynska-Wawer, Aleksander Wawer, Aleksandra Pawlak, Julia Szymanowska, Izabela Stefaniak, Michal Jarkiewicz, and Lukasz Okruszek. Detecting formal thought disorder by deep contextualized word representations. *Psychiatry Research*, 304:114135, 2021. 2", + "[45] Jonathan Sauder and Bjarne Sievers. Self-supervised deep learning on point clouds by reconstructing space. Advances in Neural Information Processing Systems, 32, 2019. 2", + "[46] Charu Sharma and Manohar Kaul. Self-supervised few-shot learning on point clouds. Advances in Neural Information Processing Systems, 33:7212-7221, 2020. 6", + "[47] Chen Sun, Abhinav Shrivastava, Saurabh Singh, and Abhinav Gupta. Revisiting unreasonable effectiveness of data in deep learning era. In Proceedings of the IEEE international conference on computer vision, pages 843-852, 2017. 6", + "[48] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6411–6420, 2019. 6", + "[49] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 1, 2", + "[50] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1588–1597, 2019. 6, 7, 2", + "[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4, 5, 6, 7, 2", + "[52] Hanchen Wang, Qi Liu, Xiangyu Yue, Joan Lasenby, and Matt J Kusner. Unsupervised point cloud pre-training via occlusion completion. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9782-9792, 2021. 7" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20963", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (tog), 38(5):1-12, 2019. 3, 4, 5, 6, 7, 1", + "[54] Ziyi Wang, Xumin Yu, Yongming Rao, Jie Zhou, and Jiwen Lu. P2p: Tuning pre-trained image models for point cloud analysis with point-to-pixel prompting. Advances in neural information processing systems, 35:14388-14402, 2022. 6, 7", + "[55] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1912-1920, 2015. 6, 2", + "[56] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 574-591. Springer, 2020. 7", + "[57] Zhenda Xie, Yutong Lin, Zhuliang Yao, Zheng Zhang, Qi Dai, Yue Cao, and Han Hu. Self-supervised learning with swim transformers. arXiv preprint arXiv:2105.04553, 2021. 1", + "[58] Yifan Xu, Tianqi Fan, Mingye Xu, Long Zeng, and Yu Qiao. Spidercnn: Deep learning on point sets with parameterized convolutional filters. In Proceedings of the European conference on computer vision (ECCV), pages 87-102, 2018. 7", + "[59] Guandao Yang, Xun Huang, Zekun Hao, Ming-Yu Liu, Serge Belongie, and Bharath Hariharan. Pointflow: 3d point cloud generation with continuous normalizing flows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4541-4550, 2019. 7", + "[60] Yaoqing Yang, Chen Feng, Yiru Shen, and Dong Tian. Foldingnet: Point cloud auto-encoder via deep grid deformation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 206–215, 2018. 4, 5, 1", + "[61] Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Russ R Salakhutdinov, and Quoc V Le. Xlnet: Generalized autoregressive pretraining for language understanding. Advances in neural information processing systems, 32, 2019. 1", + "[62] Li Yi, Vladimir G Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas Guibas. A scalable active framework for region annotation in 3d shape collections. ACM Transactions on Graphics (ToG), 35(6):1-12, 2016. 7", + "[63] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19313-19322, 2022. 1, 2, 3, 4, 5, 6, 7", + "[64] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. Point-m2ae: multi-scale masked autoencoders for hierarchical point cloud pre-training. Advances in neural information processing systems, 35:27061-27074, 2022. 2, 7" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[65] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16259-16268, 2021. 1, 6", + "[66] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In International Conference on Learning Representations, 2020. 1" + ], + "bbox": [ + 501, + 90, + 893, + 204 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "20964", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/General Point Model Pretraining with Autoencoding and Autoregressive/a64a8093-c9f6-4338-aa51-464da1052ad9_model.json b/2024/General Point Model Pretraining with Autoencoding and Autoregressive/a64a8093-c9f6-4338-aa51-464da1052ad9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a27a832dcbe5f4f414a987acde89076ff96be6a1 --- /dev/null +++ b/2024/General Point Model Pretraining with Autoencoding and Autoregressive/a64a8093-c9f6-4338-aa51-464da1052ad9_model.json @@ -0,0 +1,2257 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.121, + 0.131, + 0.85, + 0.154 + ], + "angle": 0, + "content": "General Point Model Pretraining with Autoencoding and Autoregressive" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.179, + 0.886, + 0.234 + ], + "angle": 0, + "content": "Zhe Li \\(^{1,\\dagger}\\), Zhangyang Gao \\(^{3,\\dagger}\\), Cheng Tan \\(^{3,\\dagger}\\), Bocheng Ren \\(^{1}\\), Laurence T. Yang \\(^{1,2,*}\\) Stan Z. Li \\(^{3,*}\\) \n\\(^{1}\\) Huazhong University of Science and Technology \n\\(^{2}\\) Zhengzhou University" + }, + { + "type": "text", + "bbox": [ + 0.198, + 0.234, + 0.794, + 0.253 + ], + "angle": 0, + "content": "\\(^{3}\\) AI Lab, Research Center for Industries of the Future, Westlake University" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.255, + 0.9, + 0.27 + ], + "angle": 0, + "content": "keycharon0122@gmail.com, {gaozhangyang,tancheng,stan.zq.li}@westlake.edu.cn, ltyang@ieee.org" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.305, + 0.314, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.336, + 0.474, + 0.699 + ], + "angle": 0, + "content": "The pre-training architectures of large language models encompass various types, including autoencoding models, autoregressive models, and encoder-decoder models. We posit that any modality can potentially benefit from a large language model, as long as it undergoes vector quantization to become discrete tokens. Inspired by the General Language Model, we propose a General Point Model (GPM) that seamlessly integrates autoencoding and autoregressive tasks in a point cloud transformer. This model is versatile, allowing fine-tuning for downstream point cloud representation tasks, as well as unconditional and conditional generation tasks. GPM enhances masked prediction in autoencoding through various forms of mask padding tasks, leading to improved performance in point cloud understanding. Additionally, GPM demonstrates highly competitive results in unconditional point cloud generation tasks, even exhibiting the potential for conditional generation tasks by modifying the input's conditional information. Compared to models like Point-BERT, MaskPoint, and PointMAE, our GPM achieves superior performance in point cloud understanding tasks. Furthermore, the integration of autoregressive and autoencoding within the same transformer underscores its versatility across different downstream tasks. Codes are available at https://github.com/gentlefress/GPM" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.725, + 0.21, + 0.74 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.902 + ], + "angle": 0, + "content": "In recent years, the natural language processing (NLP) [4, 19, 20, 29, 40] and computer vision (CV) [3, 7, 9, 49, 57, 66] realms have witnessed a proliferation of transformer-based pretrained models. Their primary advantage lies in the inclusion of massive parameters and data in the training process, overcoming inductive biases introduced by traditional Convolutional Neural Networks [21]. Point clouds serve as fundamental data structures in fields like autonomous driving and robotics, thus emphasizing the escalating significance of tasks related to point cloud representation learn" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.304, + 0.886, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.449, + 0.893, + 0.52 + ], + "angle": 0, + "content": "Figure 1. Comparison of pre-training frameworks between GPM and Point-BERT [63]. Point-BERT (left) conducts the masked prediction task in an autoencoding manner, while GPM (right) combines both autoencoding and autoregressive tasks, performing tasks for both masked prediction and masked region blank filling." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.895, + 0.903 + ], + "angle": 0, + "content": "ing and generation. However, the realm of pretrained point cloud models based on transformers remains relatively scarce. Existing transformer-based point cloud models [13, 65] encounter inevitable inductive biases due to local feature aggregation [13], neighbor embedding [65], and the scarcity of annotated data [9]. These biases deviate from the mainstream of standard Transformers. Self-supervised models [12, 15, 20, 25, 26, 39, 40, 61] have emerged as the dominant methodology. They excel in learning high-quality representations across modalities without extensive labeled data, models like GPT [39] even dominate in text generation. In this end, there is a compelling need to design a transformer-based point cloud model that minimizes inductive biases, learns superior point cloud representations from limited data, and simultaneously possesses the capacities for conditional and unconditional point cloud generation. The transformer-based BERT [20] has marked a significant milestone in the field of natural language processing (NLP). Its pre-training phase, involving masked prediction tasks, grants the model the capacity to learn language representations, showcasing the prowess of autoencoding language models. Given BERT's proficiency in learning language representations, the question arises: can we extend BERT's" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20954" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.394 + ], + "angle": 0, + "content": "capabilities to the realm of point clouds? However, since point clouds lack a direct analog to the concept of 'words' in language, constructing a discrete vocabulary for point clouds, as done in language models, is infeasible. Point clouds are composed of individual points, and considering each point as a token would lead to a quadratic increase in computational cost as the number of tokens grows. Furthermore, each point in a point cloud contains limited semantic information, necessitating the maximization of geometric information utilization. Inspired by Point-BERT [63], we embarked on a journey to vector quantization, transforming point clouds into discrete tokens. Through the application of the farthest point sampling (FPS) [36] algorithm, we segment the point cloud into multiple patches, each patch encapsulates plentiful geometric information, akin to a unit. By constructing a vocabulary specific to the point cloud domain, we obtain labels for each unit, discretizing the entire point cloud. This paves the way for the completion of the autoencoding model's masked prediction tasks in the point cloud domain." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.396, + 0.473, + 0.562 + ], + "angle": 0, + "content": "Transformer-based GPT [39] has dominated the field of text generation, with its core being an autoregressive language model, predicting the next token based on preceding tokens. In recent years, researchers have progressively extended this paradigm into the realm of images [7, 41, 42], ushering in a new era of image generation. In this endeavor, we bring this paradigm to the domain of point clouds, aiming to empower it for autoregressive conditional and unconditional point cloud generation tasks. After discretization, each patch of the point cloud possesses distinct geometric information, enabling it to excel in autoregressive tasks." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.564, + 0.473, + 0.642 + ], + "angle": 0, + "content": "While these pre-training frameworks can be adapted to the point cloud domain, they lack the flexibility to satisfy all point cloud tasks. Therefore, we aim to integrate these two pre-training frameworks from the NLP domain and apply them to the point cloud domain, depicted in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.643, + 0.473, + 0.868 + ], + "angle": 0, + "content": "In this work, we introduce a novel approach termed General Point Model (GPM) with autoencoding and autoregressive, which integrates point cloud representation learning and point cloud generation (both conditional and unconditional) within the same transformer framework. Similar to Point-BERT, we employ a point cloud tokenizer designed through dVAE-based point cloud reconstruction, where the point cloud can be discretely labeled based on the learned vocabulary. Additionally, we utilize the Masked Point Modeling (MPM) task. This involves partitioning the point cloud into distinct patches and randomly masking some adjacent patches. This aims to predict the masked portion, with the goal of learning geometric information between neighboring blocks and capturing meaningful geometric features for understanding the point cloud." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.472, + 0.902 + ], + "angle": 0, + "content": "Our primary contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.894, + 0.137 + ], + "angle": 0, + "content": "- We introduce autoencoding + autoregressive as a novel point cloud transformer paradigm, unifying tasks in point cloud understanding and generation." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.137, + 0.894, + 0.197 + ], + "angle": 0, + "content": "- We partition the input into two segments, the first part predicts masked points (autoencoding), the second part generates masked points (autoregressive), enhancing comprehension and generation." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.198, + 0.894, + 0.273 + ], + "angle": 0, + "content": "- Our approach demonstrates notable competitiveness across various point cloud understanding tasks and unconditional generation tasks. Moreover, it exhibits the potential for conditional generation, such as text- or image-conditioned point clouds generation." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.092, + 0.894, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.288, + 0.642, + 0.304 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.313, + 0.729, + 0.33 + ], + "angle": 0, + "content": "2.1. Self-supervised Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.337, + 0.896, + 0.822 + ], + "angle": 0, + "content": "Self-supervised learning (SSL) has garnered significant attention owing to the substantial labor required for acquiring annotated data in large quantities. SSL enables models to learn feature representations from unlabeled data, adapting to downstream tasks. The core of SSL lies in the design of proxy tasks to replace traditional classification tasks, thereby learning feature representations. ELMo [44] employs bidirectional LSTMs [16] and generates subsequent words from left to right given the representation of preceding content. ViP [25] employs a dynamically updated momentum encoder for contrastive learning and designs a text swapping task to enhance the model's sentence representation capability. In the computer vision domain, Image-GPT [7] trains a sequence transformer to autoregressively predict pixels, showing promising representation learning capabilities without incorporating specific knowledge about the 2D input structure after pretraining. Furthermore, SSL in the field of point clouds has garnered significant attention. ACT [8] employs cross-modal autoencoders as teacher models to acquire knowledge from other modalities. The generative approaches [1, 22, 33, 45, 63, 64] are the most pertinent research to our work. Point-MAE [34] extends MAE by randomly masking point patches and reconstructing the masked regions. Point-M2AE [64] further utilizes a hierarchical transformer architecture and designs corresponding masking strategies. However, mask-based point modeling methods still face the issue of shape leakage, limiting their effective generalization to downstream tasks. In this paper, we leverage autoregressive pretraining on point clouds and address the unique challenges associated with point cloud attributes. Our concise design avoids position information leakage, thereby enhancing generalization capabilities." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.832, + 0.733, + 0.848 + ], + "angle": 0, + "content": "2.2. General Language Model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.894, + 0.903 + ], + "angle": 0, + "content": "In the field of natural language processing, GLM [10] integrates both autoregressive and autoencoding methods. This enables it to perform tasks related to sentence representa" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "20955" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.083, + 0.09, + 0.5, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.091, + 0.891, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.308, + 0.894, + 0.379 + ], + "angle": 0, + "content": "Figure 2. The framework of our GPM. We divide the entire pipeline into two stages: 1) The input point cloud is divided into several subclouds. Then, we utilize Mini-PointNet [35] to obtain a sequence of point embeddings. These embeddings are transformed into a sequence of discrete point tokens using a pre-trained tokenizer; 2) During pre-training, for PartA, certain portions of the masked embeddings are masked and replaced with [M]. They are predicted in an autoencoding manner. For PartB, masks for PartA are generated in an autoregressive manner." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.404, + 0.471, + 0.526 + ], + "angle": 0, + "content": "tion as well as conditioned and unconditioned generation through fine-tuning. To the best of our knowledge, we are the first to integrate autoregressive and autoencoding techniques on a single transformer in the domain of point clouds. In downstream tasks, we have the potential to fine-tune point cloud representations and perform unconditioned generation even in zero-shot scenarios, with the possibility of fine-tuning towards conditioned generation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.54, + 0.176, + 0.555 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.565, + 0.47, + 0.747 + ], + "angle": 0, + "content": "In this study, we endeavor to seamlessly integrate BERT-style and GPT-style pre-trained strategies, extending their applicability to point cloud transformers. Following the Point-BERT[63], our training process unfolds in two distinctive stages. Firstly, we embark on training a specialized Tokenizer, a critical step towards acquiring discrete point labels for each input point cloud. Concurrently, a dedicated discrete VAE (dVAE) is employed to masterfully reconstruct the discrete point cloud. The subsequent stage involves the splicing of two sequences of point cloud embeddings, partA for autoencoding and partB for autoregressive. The overall idea of our approach is illustrated in Figure 2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.756, + 0.47, + 0.789 + ], + "angle": 0, + "content": "3.1. Stage 1: Discrete Varitional Autoencoder Pretraining" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Point Cloud Partitioning. We posit that discrete tokens derived from points encapsulate crucial geometric information, allowing for a discrete representation of any point cloud. However, a naive strategy of assigning one token to each point in a point-wise reconstruction task poses a formidable computational challenge. This arises from the quadratic complexity of self-attention within transformers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.404, + 0.895, + 0.617 + ], + "angle": 0, + "content": "Building upon the foundations laid by Point-BERT [63] and ViT[9], we adopt a strategy of partitioning each point cloud into distinct patches, each serving as a single token. Specifically, for an input point cloud \\( p \\in \\mathbb{R}^{\\mathcal{N} \\times 3} \\), we initiate the process by extracting \\( m \\) center points \\( \\mathcal{C}_u \\in \\mathbb{R}^{m \\times 3} \\) from the point cloud \\( p \\) through farthest point sampling (FPS). Subsequently, we employ the k-nearest neighbor (kNN) algorithm to identify the \\( k \\) nearest neighboring points for every center point. This forms \\( m \\) localized patches or sub-clouds denoted as \\( \\mathcal{P}_u \\in \\mathbb{R}^{m \\times k \\times 3} \\). To ensure these patches are free of bias, we normalize them by subtracting their respective center coordinates. This operation effectively disentangles the structural patterns from the spatial coordinates inherent to each local patch." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.637, + 0.895, + 0.837 + ], + "angle": 0, + "content": "Point Cloud Embedding. Following Point-BERT [63], we employ a mini-PointNet [35] to embed the point patches as a sequence of point embedding \\(\\{h_i\\}_{i=1}^m\\). For the further vector quantization, we pre-define a learnable codebook \\(\\mathcal{V} := \\{(s, z_s)\\}_{s \\in S}^{d_z}\\), where \\(d_z\\) is the dimension of codes, \\(S\\) is the size of codebook and \\(s\\) is the index of embedding in \\(\\mathcal{V}\\). We adopt DGCNN [53] as tokenizer \\(\\mathcal{Q}: h_i \\mapsto z_i\\), which maps \\(\\{h_i\\}_{i=1}^m\\) into \\(\\{z_i\\}_{i=1}^m\\) in \\(\\mathcal{V}\\). However, given the non-differentiable nature of the discrete tokens, applying reparameterization gradients for dVAE training becomes unfeasible. As suggested in [41], we resort to the Gumbel-softmax relaxation technique [18], coupled with a uniform prior, as a workaround during the dVAE training process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Point Cloud Reconstruction. Given \\(\\{z_i\\}_{i=1}^m\\) as input, the decoder \\(\\mathcal{D}(\\cdot)\\) is tasked with the reconstruction of the entire point cloud. Inspired by [63], to effectively capture inter" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20956" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "point relationships and bolster the representational capacity of discrete point tokens across various local structures, we adopt the DGCNN [53]-FoldingNet [60] architecture to reconstruct the whole point cloud." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.152, + 0.47, + 0.228 + ], + "angle": 0, + "content": "The overarching reconstruction objective can be denoted as \\(\\mathbb{E}_{z\\sim \\mathcal{Q}(z|h)}\\log D(\\hat{p} |z)\\). Inspired by Point-BERT [63], the entire reconstruction process can be conceptualized as the maximization of the Evidence Lower Bound (ELB) for the log-likelihood, denoted as \\(\\mathcal{D}(\\hat{p} |p)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.236, + 0.48, + 0.299 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\sum_ {i = 1} ^ {\\mathcal {N}} \\log D (\\hat {p} _ {i} | p _ {i}) \\geq \\sum_ {i = 1} ^ {\\mathcal {N}} \\left(\\mathbb {E} _ {z _ {i}} \\sim \\mathcal {Q} (z _ {i} | p _ {i})\\right) [ \\log D (p _ {i} | z _ {i}) ] \\\\ - \\operatorname {K L} \\left[ \\mathcal {Q} \\left(z \\mid p _ {i}\\right), \\mathcal {D} \\left(z \\mid \\hat {p} _ {i}\\right) \\right]. \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.31, + 0.469, + 0.356 + ], + "angle": 0, + "content": "Simultaneously, we account for an intuitive reconstruction loss, leveraging the \\(l1\\) Chamfer Distance [11] to supervise prediction point cloud with the ground-truth point cloud:" + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.364, + 0.469, + 0.414 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {C D}} (\\mathcal {P}, \\mathcal {G}) = \\frac {1}{| \\mathcal {P} |} \\sum_ {p \\in \\mathcal {P}} \\min _ {g \\in \\mathcal {G}} \\| p - g \\| + \\frac {1}{| \\mathcal {G} |} \\sum_ {g \\in \\mathcal {G}} \\min _ {p \\in \\mathcal {P}} \\| g - p \\|, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.415, + 0.469, + 0.474 + ], + "angle": 0, + "content": "where \\(\\mathcal{P}\\) denotes the predicted point clouds and \\(\\mathcal{G}\\) is the ground truth point clouds. Additional, we follow [41], optimize the KL-divergence \\(\\mathcal{L}_{\\mathrm{KL}}\\) between the generated point cloud distribution and a uniform distribution:" + }, + { + "type": "equation", + "bbox": [ + 0.2, + 0.487, + 0.469, + 0.502 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {d V A E}} = \\mathcal {L} _ {\\mathrm {C D}} + \\mathcal {L} _ {\\mathrm {K L}}. \\tag {3}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.512, + 0.323, + 0.529 + ], + "angle": 0, + "content": "3.2. Stage 2: GPM Pre-training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.535, + 0.469, + 0.597 + ], + "angle": 0, + "content": "To ensure GPM excel in both point cloud representation learning and point cloud generation task, we formulate a training framework that amalgamates autoencoding and autoregressive techniques motivated by [10]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.469, + 0.796 + ], + "angle": 0, + "content": "Autoencoding Masked Sequence Generation. We employ the standard Transformers [51], which encompass multi-headed self-attention layers and FFN blocks. Initially, we partition each input point cloud into \\( m \\) local patches with designated center points. Subsequently, these local patches undergo discretization into a code sequence \\( \\{f_i\\}_{i=1}^m \\) via our pre-trained dVAE in stage 1. Formally, the input embeddings \\( \\{e_i\\}_{i=1}^m \\) are constructed as a combination of point embeddings \\( \\{f_i\\}_{i=1}^m \\) and positional embeddings \\( \\{pos_i\\}_{i=1}^m \\). Adhering to language model [20], a class token [CLS] is appended to the input sequences, rendering the transformer input sequence as \\( I = \\{[\\mathrm{CLS}], e_1^0, e_2^0, \\dots, e_i^0\\} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Drawing inspiration from [63], we initiate by selecting a central point \\(\\mathcal{C}_i\\) alongside its corresponding sub-cloud \\(\\mathcal{P}_i\\), integrate it with \\(m\\) neighboring sub-clouds as a coherent local region. Within this region, we apply a masking operation to obscure all local patches, generating what we refer to as the \"masked point cloud\". Specifically, we substitute all the masked point embeddings with a universally learnable" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.151 + ], + "angle": 0, + "content": "pre-defined mask embeddings [M]. We mark the masked position as \\(\\mathcal{M} = \\{1,\\dots,b\\}\\), and the final input embeddings \\(E^{\\mathcal{M}} = \\{e_i,i\\notin \\mathcal{M}\\} \\cup \\{[\\mathbf{M}] + pos_i,i\\in \\mathcal{M}\\}\\) are fed into the transformer." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.152, + 0.893, + 0.333 + ], + "angle": 0, + "content": "The objective of our Masked Point Modeling (MPM) task is to deduce the geometric arrangement of absent portions and reconstruct point tokens aligned with the masked positions using the available information. Geometric information refers to the relative positions and morphological attributes of points within a localized region. These crucial geometric details provide vital cues for understanding and processing the shape and structure of the local region, playing a pivotal role in delving deep into the intricacies of point cloud data. Formally, the pre-training objective can be expressed as maximizing the log-likelihood of the point tokens \\( e_i \\) conditioned on the masked input embeddings \\( E^{\\mathcal{M}} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.343, + 0.892, + 0.381 + ], + "angle": 0, + "content": "\\[\n\\max \\sum_ {E \\in W} \\mathbb {E} \\left[ \\sum_ {i \\in \\mathcal {M}} \\log P \\left(e _ {i} \\mid E ^ {\\mathcal {M}}\\right) \\right], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.39, + 0.892, + 0.541 + ], + "angle": 0, + "content": "where \\( W \\) is the set of all input embeddings. Similar to [63], we also adopt Point Patch Mixing and contrastive learning to help the model to better learn high-level semantics. Point Patch Mixing involves mixing sub-clouds without complex alignment techniques. However, training solely with masked precition has limitations in understanding high-level semantics. To overcome this, we incorporate contrastive learning to enhance semantic understanding. Assuming the masking ratio is \\( \\beta \\), the loss function is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.549, + 0.891, + 0.588 + ], + "angle": 0, + "content": "\\[\n- \\beta \\log \\frac {\\exp \\left(q \\hat {k} _ {1} / \\tau\\right)}{\\sum_ {i = 0} ^ {K} \\exp \\left(q k _ {i}\\right)} - (1 - \\beta) \\log \\frac {\\exp \\left(q \\hat {k} _ {2} / \\tau\\right)}{\\sum_ {i = 0} ^ {K} \\exp \\left(q k _ {i}\\right)}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.596, + 0.892, + 0.657 + ], + "angle": 0, + "content": "where \\(q\\) is the feature of a mixed sample that comes from two other samples, whose features are \\(\\hat{k_1}\\) and \\(\\hat{k_2}\\). \\(k_{i}\\) are extracted by the momentum feature encoder, \\(\\tau\\) is the temperature and \\(K\\) is the size of memory bank." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.892, + 0.81 + ], + "angle": 0, + "content": "Autoregressive Sequence Generation. While Point-BERT models the relationship between unmasked and masked regions using the Masked Point Modeling (MPM) task, it does not adequately model the interactions within masked regions. Furthermore, our aim extends beyond acquiring high-quality point cloud representations; we also aspire to proficiently execute point cloud generation tasks. Hence, for the latter segment, PartB, we engage in autoregressive mask generation tasks to enhance the MPM task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In the case of PartB, the mask operation is ignored. Our focus lies in acquiring the generated tokens from the masked positions of PartA. We take the initial \\( n - 1 \\) tokens from PartA, appending the start token [S] at its forefront to get the PartB \\( E^{\\mathcal{B}} = \\{[\\mathrm{S}], e_1, \\dots, e_{n-1}\\} \\). Following the autoregressive approach, the prediction of the next token relies" + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20957" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.236, + 0.106 + ], + "angle": 0, + "content": "on the preceding token:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.115, + 0.47, + 0.155 + ], + "angle": 0, + "content": "\\[\n\\max _ {\\theta} \\mathbb {E} \\left[ \\sum_ {i = 1} ^ {n} \\log P _ {\\theta} \\left(e _ {i} | \\mathrm {P a r t A}, e _ {< i}, [ \\mathrm {S} ]\\right) \\right]. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.161, + 0.47, + 0.281 + ], + "angle": 0, + "content": "Although no masking operation is performed on PartB, since the mask positions of PartA are multiple consecutive sub-clouds instead of a single one, in PartB we are also equivalent to autoregressive generation of a continuous mask span. We extract the generated tokens from PartB, corresponding to the masked positions in PartA, and compare them with the discrete tokens produced by the Tokenizer:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.291, + 0.469, + 0.311 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {A R} = \\operatorname {C r o s s E n t r o p y} \\left(e _ {j, j \\in \\mathcal {M}}, \\operatorname {a r g m a x} \\mathcal {Q} (h) _ {j, j \\in \\mathcal {M}}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.315, + 0.469, + 0.36 + ], + "angle": 0, + "content": "In this way, it allows our model to acquire proficiency in an autoencoding bidirectional encoding scheme for PartA, and an autoregressive unidirectional encoder for PartB." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.368, + 0.289, + 0.385 + ], + "angle": 0, + "content": "3.3. Multi-task Pretraining" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.391, + 0.469, + 0.542 + ], + "angle": 0, + "content": "The masking of \\(15\\%\\) of tokens in BERT is tailored for downstream natural language understanding tasks, whereas Point-BERT masks \\(25\\% \\sim 45\\%\\) to acquire enhanced point cloud representations. Our GPM focus on both point cloud representation learning and point cloud generation, we concatenate PartA and PartB as input to the transformer, aiming to simultaneously perform both autoencoding and autoregressive tasks. Therefore, we need to establish specific attention masks to facilitate the interaction of information between these two segments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.543, + 0.469, + 0.649 + ], + "angle": 0, + "content": "Motivated by [10], the tokens in PartA can participate in MLM task in PartA and autoregressive generation in PartB. PartB tokens can not be observed by PartA, but those antecedents. PartA is treated as the conditioning for autoregressive generation in PartB, akin to prefix tuning [23] in NLP. The implement of GPM multi-task pretraining is depicted in Figure 3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.649, + 0.469, + 0.723 + ], + "angle": 0, + "content": "The tokens output by the pretrained tokenizer \\(\\mathcal{Q}(\\cdot)\\) are regarded as labels. PartA conducts a masked prediction task in an autoencoding manner, and the prediction loss between the predicted masked portion and the labels is computed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.733, + 0.469, + 0.753 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {A E} = \\operatorname {C r o s s E n t r o p y} \\left(e _ {j, j \\in \\mathcal {M}}, \\operatorname {a r g m a x} \\mathcal {Q} (h) _ {j, j \\in \\mathcal {M}}\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.758, + 0.469, + 0.833 + ], + "angle": 0, + "content": "Although Equation 7 and 8 exhibit identical formal structures. However, one entails mask prediction through autoencoding, while the other involves mask generation in an autoregressive way, indicating a fundamental distinction between the two approaches." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.846, + 0.209, + 0.862 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "In this section, we commence by outlining the configurations of our pretraining scheme. Subsequently, we pro" + }, + { + "type": "image", + "bbox": [ + 0.579, + 0.089, + 0.816, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.291, + 0.892, + 0.334 + ], + "angle": 0, + "content": "Figure 3. Self-attention mask. The regions in grey are masked. Tokens in PartA can attend to themselves, but not to PartB. Tokens in PartB can attend to PartA and their antecedents in PartB." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.362, + 0.892, + 0.453 + ], + "angle": 0, + "content": "ceed to assess the proposed model through various downstream tasks, encompassing object classification, partial segmentation, few-shot learning, transfer learning (to validate the model's representation learning capabilities), and point cloud generation (to validate its generative capabilities). Additionally, we conduct ablation studies on GPM." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.463, + 0.847, + 0.48 + ], + "angle": 0, + "content": "4.1. Pre-training Setups and Implementation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.487, + 0.892, + 0.638 + ], + "angle": 0, + "content": "Pre-training Data. Following the dataset configuration similar to Point-BERT [63], we employ ShapeNet [5] as our pretraining dataset, encompassing over 50,000 unique 3D models spanning 55 common object categories. From each 3D model, we sample 1024 points, dividing them into 64 point patches (sub-clouds) and each sub-cloud contains 32 points. Utilizing a lightweight PointNet [35] with two MLP layers, we project each sub-cloud into a 64-point embedding, serving as input for both dVAE and Transformer [51]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.809 + ], + "angle": 0, + "content": "dVAE Pretraining Setups. The purpose of pretraining the dVAE is to acquire a high-quality tokenizer, enabling the reconstruction of discretized features back into the original point cloud to the fullest extent. Our dVAE is composed of a tokenizer and a decoder. To be specific, the tokenizer consists of a 4-layer DGCNN [53], while the decoder encompasses a 4-layer DGCNN followed by a FoldingNet [60]. Additionally, the FoldingLayer establishes a connection between a 2D grid and the input, ultimately generating a 3D point cloud." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "During the training of dVAE [41], we set the number of words in the codebook \\(N\\) to 8192. We employ the common \\(l1\\)-style Chamfer Distance loss in the reconstruction process. Due to the small numerical value of this \\(l1\\) loss, the weight of the KL loss in Equation 1 must be smaller than that in image tasks. In the initial 10,000 steps of training," + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20958" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "the weight of the KL loss is kept at 0 and gradually raised to 0.1 in the subsequent 100,000 steps. Our learning rate is set to 0.0005, and follows a cosine annealing schedule spanning 60,000 steps. Consistent with prior works [41] and Point-BERT [63], we decay the temperature in the Gumblesoftmax function from 1 to 0.0625 over 100,000 steps. The training of dVAE encompasses a total of 150,000 steps, with a batch size of 64. It's worth noting that our dVAE's network architecture closely mirrors that of Point-BERT." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.248, + 0.473, + 0.52 + ], + "angle": 0, + "content": "GPM Pretraining Setups. In our experiments, we adhere to the standard Transformer architecture [9], which consists of a stack of Transformer blocks [51], each composed of multi-head self-attention layers and feed-forward networks. LayerNorm is applied in both layers. We set the depth of our Transformer to 12, feature dimension to 384, and the number of heads to 6. A random depth of 0.1 [17] is applied in our Transformer encoder. During GPM pretraining, the weights of the Tokenizer learned by dVAE are kept fixed. Input point embeddings from \\(25\\% \\sim 45\\%\\) are randomly masked, and the model is trained to infer the expected point labels at these masked positions. Unlike Point-BERT [63], which does not have a generation task during pretraining, we incorporate both autoregressive and autoencoding tasks in our work. Thus, an attention mask is necessary, as illustrated in Figure 3. For MoCo, we set the same configuration as [63]. More Details can be found in the supplementary 6.1 and 6.2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.531, + 0.262, + 0.545 + ], + "angle": 0, + "content": "4.2. Downstream Tasks" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.554, + 0.406, + 0.57 + ], + "angle": 0, + "content": "4.2.1 Point Cloud Representation Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.579, + 0.471, + 0.822 + ], + "angle": 0, + "content": "Object classification on ModelNet40. We evaluate our pre-trained model on the ModelNet40 dataset [55], comprising 12,311 clean 3D CAD models spanning 40 categories. Following the Point-BERT [63] setups, we employ a two-layer MLP with a dropout rate of 0.5 as the classification head for the task. We optimize the model using AdamW with a weight decay of 0.05 and a learning rate of 0.0005, while employing a batch size of 32 and a cosine annealing schedule. We conduct comparisons with various Transformer-based models with identical Point-BERT settings, denoting [ST] for the standard Transformer architecture and [T] for Transformer models with specific designs or inductive biases. The results in Table 1 demonstrate that our model not only outperforms Point-BERT in classification metrics on this dataset but also exhibits autoregressive generation capabilities." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Object classification on ScanObjectNN. The practical applicability of our model on real datasets serves as a crucial metric. Therefore, the pre-trained models are transferred to the ScanObjectNN dataset [50], which comprises" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.089, + 0.895, + 0.462 + ], + "angle": 0, + "content": "
Methodsnumber of pointsAcc
Supervised Learning
PointNet[35]1k89.2
SO-Net[22]1k92.5
PointNet++ [36]1k90.5
PointCNN[24]1k92.2
DGCNN[53]1k92.9
MVTN[14]1k93.8
RSCNN [43]1k92.9
GBNet[38]1k93.8
PointMLP[32]1k94.5
DensePoint[28]1k92.8
PointNeXt[37]1k94.0
P2P-RN101[54]1k93.1
P2P-HorNet[54]1k94.0
KPConv[48]~6.8k92.9
Self-Supervised Representation Learning
[T]PCT [65]1k93.2
[T]PointTransformer [13]-93.7
[ST]NPCT [13]1k91.0
[ST]Transformer[51]1k91.4
[ST]Transformer-OcCo[51]1k92.1
[ST]Point-BERT[63]1k93.2
[ST]MaskPoint[27]1k93.8
[ST]Point-MAE[34]1k93.8
[ST]GPM1k94.0
Methods with Cross-modal Information and Teacher Models
ACT[8]1k93.7
[ST]Transformer[51]4k91.2
[ST]Transformer-OcCo[51]4k92.2
[ST]Point-BERT[63]4k93.4
[ST]GPM4k93.1
[ST]Point-BERT[63]8k93.8
[ST]MaskPoint[27]8k-
[ST]Point-MAE[34]8k94.0
[ST]GPM8k94.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.471, + 0.892, + 0.501 + ], + "angle": 0, + "content": "Table 1. Classification results on ModelNet40. All results are expressed as percentages accuracy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.529, + 0.895, + 0.68 + ], + "angle": 0, + "content": "approximately 15,000 objects extracted from real indoor scans, encompassing 2902 point clouds from 15 categories. This dataset presents a greater challenge as it involves sampling from real-world scans with backgrounds and occlusions. We conducted experiments on three main variants, namely OBJ-BG, OBJ-ONLY, and PB-T50-RS, in line with prior work. The experimental results are summarized in Table 2. It is observed that GPM exhibits an improvement of approximately \\(2.77\\%\\), \\(1.88\\%\\), and \\(1.73\\%\\) over the regular Point-BERT across the three variants." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Few-shot learning. To demonstrate the ability to acquire knowledge for new tasks with limited training data, we evaluate our model under the setting of few-shot learning, following the methodology of previous work [47, 63]. In the typical 'W-way S-shot' setup, we initially randomly select \\( W \\) classes and then sample \\( (S + 20) \\) objects for each class [46]. The model is trained on \\( W \\times S \\) samples (support set) and evaluated on the remaining \\( 20W \\) samples (query set). We conduct 10 independent experiments for each setting and report the average performance and standard deviation across the 10 runs. As shown in Table 3, our approach outperforms other methods in all tests, achieving absolute improvements of \\( 0.4\\% \\), \\( 0.4\\% \\), and \\( 0.6\\% \\) over Point-BERT." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "20959" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.473, + 0.334 + ], + "angle": 0, + "content": "
MethodsOBJ-BGOBJ-ONLYPB-T50-RS
Supervised Learning
PointNet[35]73.379.268.0
SpiderCNN[58]77.179.573.7
PointNet++ [36]82.384.377.9
PointCNN[24]86.185.578.5
DGCNN[53]82.886.278.1
MVTN[14]92.692.382.8
BGA-DGCNN[50]--79.7
BGA-PN++ [50]--80.2
GBNet[38]--81.0
PointMLP[32]--85.4
PointNeXt[37]--87.7
P2P-RN101[54]--87.4
P2P-HorNet[54]--89.3
Self-Supervised Representation Learning
Transformer[51]79.8680.5577.24
Transformer-OcCo[51]84.8585.5478.79
Point-BERT[63]87.4388.1283.07
MaskPoint[27]89.388.184.3
Point-MAE[34]90.088.285.2
GPM90.290.084.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.345, + 0.47, + 0.373 + ], + "angle": 0, + "content": "Table 2. Classification results on ScanObjectNN. All results are expressed as percentages accuracy." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.386, + 0.47, + 0.512 + ], + "angle": 0, + "content": "
Methods5-way10-way
10-shot20-shot10-shot20-shot
Supervised Learning
DGCNN-rand [53]31.6 ± 2.840.8 ± 4.619.9 ± 2.116.9 ± 1.5
OcCo [52]90.6 ± 2.892.5 ± 1.982.9 ± 1.386.5 ± 2.2
Self-Supervised Representation Learning
Point-BERT [63]94.6 ± 3.196.3 ± 2.791.0 ± 5.492.7 ± 5.1
MaskPoint [27]95.0 ± 3.797.2 ± 1.791.4 ±4.093.4 ± 3.5
Piont-MAE [34]96.3 ± 2.597.8 ± 1.892.6 ± 4.195.0 ± 3.0
Piont-M2AE [64]96.8 ± 1.898.3 ± 1.492.3 ± 4.595.0 ± 3.0
GPM97.2 ± 2.698.7 ± 2.292.9 ± 4.295.0 ± 3.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.522, + 0.47, + 0.579 + ], + "angle": 0, + "content": "Table 3. The results of few-shot classification on the ModelNet40 dataset. For each experimental setting, we conduct 10 independent experiments and report the average accuracy \\((\\%)\\) along with its standard deviation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.605, + 0.47, + 0.831 + ], + "angle": 0, + "content": "Part segmentation. We evaluate the representation learning capability of our approach on the ShapeNetPart dataset [62], aiming to predict more fine-grained class labels for each point. This dataset consists of 16 categories and comprises 16,881 objects. The point cloud is downsampled to 2048 points, and the segmentation head [34] connects features \\(\\mathcal{F}^4, \\mathcal{F}^8, \\mathcal{F}^{12}\\) extracted from the 4-th, 6-th, and 12-th layers of the transformer blocks. Subsequently, average pooling, max pooling, and upsampling are employed to generate features for each point, followed by label prediction using MLP. In our experiments, we evaluate the performance under the settings of \"5 way 10 shot\", \"5 way 20 shot\", \"10 way 10 shot\", and \"10 way 20 shot\". The experimental results shown in Table 4 demonstrate the superior performance of our GPM compared to all other methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.833, + 0.379, + 0.847 + ], + "angle": 0, + "content": "4.2.2 Point Cloud Generation Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.903 + ], + "angle": 0, + "content": "In the pre-training phase, we perform unconditional autoregressive generation on tokens in PartB, endowing the model with the capability of point cloud unconditional generation." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.089, + 0.895, + 0.288 + ], + "angle": 0, + "content": "
MethodsCls.mIoUInst.mIoU
Supervised Learning
PointNet[35]80.3983.7
PointNet++ [36]81.3585.1
DGCNN[53]82.3385.2
PointMLP [32]84.686.1
Self-Supervised Representation Learning
Transformer[51]83.4285.1
Transformer-OcCo[51]83.4285.1
PointContrast [56]-85.1
CrossPoint [2]-85.5
Point-BERT [63]84.1185.6
Point-MAE [34]-86.1
GPM84.2085.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.298, + 0.892, + 0.34 + ], + "angle": 0, + "content": "Table 4. Part segmentation results on the ShapeNetPart dataset. We report the average intersection mIoU over the union of all classes (Cls.) and instances (Inst.)." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.355, + 0.892, + 0.455 + ], + "angle": 0, + "content": "
CategoryModelJSD (↓)MMD (↓)Cov (%, ↑)1-NNA % (%, ↓)
CDEMDCDEMDCDEMD
AirplanePoint-Flow4.920.2173.2446.9148.4075.6875.06
Point-BERT4.070.1892.7550.1753.4470.1661.07
GPM3.320.1762.4752.3456.4468.2260.30
ChairPoint-Flow1.742.427.8746.8346.9860.8859.89
Point-BERT1.231.827.1149.7350.0757.3056.61
GPM0.981.676.8251.0852.1454.0954.27
CarPoint-Flow0.870.915.2244.0346.5960.6562.36
Point-BERT0.730.824.9647.2752.1359.0960.48
GPM0.660.754.3849.6754.1758.8858.16
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.465, + 0.892, + 0.507 + ], + "angle": 0, + "content": "Table 5. Generation results on ShapeNet dataset. \\(\\uparrow\\): the higher the better, \\(\\downarrow\\): the lower the better. MMD-CD is multiplied by \\(10^{3}\\); MMD-EMD and JSD are multiplied by \\(10^{2}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.892, + 0.687 + ], + "angle": 0, + "content": "Therefore, in downstream tasks, without fine-tuning, we conduct an autoregressive point cloud generation task on the ShapeNet dataset. Existing models do not possess both point cloud generation and point cloud classification tasks. Hence, our model is the first known point cloud transformer that integrates these two tasks into one. The specific metrics are shown in Table 5. We compare our results with generation model Point-Flow [59] and Point-BERT. More details about downstream tasks and conditional point cloud generation results can be found in the supplementary 6.3." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.696, + 0.655, + 0.712 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To validate the fundamental design of the GPM model, we conducted a comprehensive set of ablation studies. In order to assess the impact of autoregressive mask generation on the autoencoding masked prediction task, we present results comparing fine-tuning scenarios on the ModelNet40 dataset: one with autoencoding only, and the other with a combination of autoencoding + autoregressive. Furthermore, to demonstrate that the order of autoencoding and autoregressive tasks in the sequence does not affect point cloud understanding, we conduct additional experiments, the results are summarized in the Table 6. More ablation results are shown in the supplementary 6.4." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20960" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.083, + 0.069, + 0.445, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.149, + 0.201, + 0.376, + 0.213 + ], + "angle": 0, + "content": "(a) Point cloud generation results of dVAE." + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.068, + 0.885, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.577, + 0.201, + 0.838, + 0.213 + ], + "angle": 0, + "content": "(b) Point cloud generation results of Point-BERT." + }, + { + "type": "image_caption", + "bbox": [ + 0.257, + 0.225, + 0.712, + 0.239 + ], + "angle": 0, + "content": "Figure 4. The reconstruction results visualization on dVAE and Point-BERT." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.264, + 0.473, + 0.453 + ], + "angle": 0, + "content": "
TasksnpointsAcc
autoencoding only1k92.9
autoregressive ⇒ autoencoding1k93.6
autoencoding ⇒ autoregressive1k93.8
autoencoding only4k93.4
autoregressive ⇒ autoencoding4k93.9
autoencoding ⇒ autoregressive4k93.9
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.463, + 0.47, + 0.533 + ], + "angle": 0, + "content": "Table 6. Results on autoencoding, autoencoding \\(\\Rightarrow\\) autoregressive and autoregressive \\(\\Rightarrow\\) autoencoding on the ModelNet dataset. It is clear that autoregressive can enhance the completion accuracy of autoencoding, the order of autoencoding and autoregressive does not have a significant impact on point cloud understanding tasks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.562, + 0.18, + 0.578 + ], + "angle": 0, + "content": "4.4. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.587, + 0.47, + 0.814 + ], + "angle": 0, + "content": "Compare with Point-BERT. Motivated by BERT, Point-BERT leverages MPM as a pretext task for training. Due to the assumption of independence between MLM and MPM, Point-BERT fails to capture interdependencies among masked tokens. In our approach, autoregressive generation explicitly models dependencies between masked regions, unlike autoencoding which relies solely on un-masked tokens to predict the masked ones. Autoregressive masked prediction allows previously predicted tokens to influence subsequent predictions, enhancing information exchange between them. This leads to improved performance in downstream tasks in the realm of representation learning. Simultaneously, tokens in PartA can attend to task in PartB, making PartA a generation condition for PartB, resulting in enhanced generation performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Compare with PointGPT. PointGPT [6] adopts a two-stage transformer architecture for learning representations of point clouds, comprising a feature extractor and a generator. It first extracts features from the point cloud us" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.266, + 0.895, + 0.419 + ], + "angle": 0, + "content": "ing the feature extractor and then directly generates with the obtained features in the generator. During training, it directly employs Chamfer Distance [11] between generated point clouds and ground truth points as the optimization objective, rather than operating on tokens. This is because it doesn't discretize points embeddings, thus failing to achieve genuine autoregressive generation tasks. Moreover, it doesn't integrate the two tasks for joint training in a single transformer; instead, it trains them separately on two different transformers, which increases the complexity." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.427, + 0.638, + 0.442 + ], + "angle": 0, + "content": "4.5. Visualization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.45, + 0.895, + 0.648 + ], + "angle": 0, + "content": "We visualize the generation results of dVAE and Point-BERT, as shown in the Figure 4a and Figure 4b. Although Point-BERT is capable of masked tokens generation through autoencoding, our approach allows for autoregressive conditional generation tasks, such as text-conditioned point clouds and image-conditioned point clouds. As observed in Figure 4b, the generation results of Point-BERT exhibit some scattered points at the edges of the overall structure, appearing relatively sparse and disorganized. The generated outcome is not as satisfactory as that of GPM. More visualizations of generation results can refer to Unconditional and Conditional Point Generation in the supplementary 6.3." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.66, + 0.619, + 0.675 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.686, + 0.894, + 0.807 + ], + "angle": 0, + "content": "Our study integrates both autoencoding and autoregressive tasks within a unified transformer to enhance the model's performance in representation learning and generation at the same time. By this framework, we address the limitation of current pretraining frameworks and make the proxy tasks non-trivial, effectively enhance the information interaction among masked tokens. Moreover, our framework endows the model with the potential for conditional generation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.82, + 0.661, + 0.838 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.845, + 0.894, + 0.888 + ], + "angle": 0, + "content": "This work was supported by National Science and Technology Major Project (No.2022ZD0115101), National Natural Science Foundation of China Project (No.U21A20427)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20961" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In International conference on machine learning, pages 40-49. PMLR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.257 + ], + "angle": 0, + "content": "[2] Mohamed Afham, Isuru Dissanayake, Dinithi Dissanayake, Amaya Dharmasiri, Kanchana Thilakarathna, and Ranga Rodrigo. Crosspoint: Self-supervised cross-modal contrastive learning for 3d point cloud understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9902-9912, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.471, + 0.3 + ], + "angle": 0, + "content": "[3] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. In International Conference on Learning Representations, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.302, + 0.471, + 0.372 + ], + "angle": 0, + "content": "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.373, + 0.47, + 0.442 + ], + "angle": 0, + "content": "[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.445, + 0.47, + 0.5 + ], + "angle": 0, + "content": "[6] Guangyan Chen, Meiling Wang, Yi Yang, Kai Yu, Li Yuan, and Yufeng Yue. Pointgpt: Auto-regressively generative pre-training from point clouds. arXiv preprint arXiv:2305.11487, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.502, + 0.47, + 0.558 + ], + "angle": 0, + "content": "[7] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International conference on machine learning, pages 1691-1703. PMLR, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.56, + 0.47, + 0.643 + ], + "angle": 0, + "content": "[8] Runpei Dong, Zekun Qi, Linfeng Zhang, Junbo Zhang, Jianjian Sun, Zheng Ge, Li Yi, and Kaisheng Ma. Autoencoders as cross-modal teachers: Can pretrained 2d image transformers help 3d representation learning? In The Eleventh International Conference on Learning Representations, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.645, + 0.47, + 0.728 + ], + "angle": 0, + "content": "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.73, + 0.47, + 0.813 + ], + "angle": 0, + "content": "[10] Zhengxiao Du, Yujie Qian, Xiao Liu, Ming Ding, Jiezhong Qiu, Zhilin Yang, and Jie Tang. Glm: General language model pretraining with autoregressive blank infilling. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 320-335, 2022. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[11] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 605-613, 2017. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[12] Tianyu Gao, Xingcheng Yao, and Danqi Chen. Simcse: Simple contrastive learning of sentence embeddings. In 2021" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "Conference on Empirical Methods in Natural Language Processing, EMNLP 2021, pages 6894-6910. Association for Computational Linguistics (ACL), 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[13] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. Pct: Point cloud transformer. Computational Visual Media, 7:187-199, 2021. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.191, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[14] Abdullah Hamdi, Silvio Giancola, and Bernard Ghanem. Mvtn: Multi-view transformation network for 3d shape recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1-11, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.247, + 0.892, + 0.314 + ], + "angle": 0, + "content": "[15] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.316, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[16] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.344, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[17] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q Weinberger. Deep networks with stochastic depth. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 646-661. Springer, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[18] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. In International Conference on Learning Representations, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.456, + 0.892, + 0.523 + ], + "angle": 0, + "content": "[19] Mandar Joshi, Danqi Chen, Yinhan Liu, Daniel S Weld, Luke Zettlemoyer, and Omer Levy. Spanbert: Improving pre-training by representing and predicting spans. Transactions of the association for computational linguistics, 8:64-77, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.525, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[20] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, pages 4171–4186, 2019. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[21] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.892, + 0.692 + ], + "angle": 0, + "content": "[22] Jiaxin Li, Ben M Chen, and Gim Hee Lee. So-net: Self-organizing network for point cloud analysis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9397–9406, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.892, + 0.775 + ], + "angle": 0, + "content": "[23] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 4582-4597, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[24] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. Advances in neural information processing systems, 31, 2018. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] Zhe Li, T. Yang Laurence, Xin Nie, BoCheng Ren, and Xianjun Deng. Enhancing sentence representation with visually-supervised multimodal pre-training. In Proceedings of the 31st ACM International Conference on Multimedia (ACM MM'23), 2023. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20962" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[26] Zhe Li, Laurence T Yang, Bocheng Ren, Xin Nie, Zhangyang Gao, Cheng Tan, and Stan Z Li. Mlip: Enhancing medical visual representation with divergence encoder and knowledge-guided contrastive learning. arXiv preprint arXiv:2402.02045, 2024.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.165, + 0.47, + 0.22 + ], + "angle": 0, + "content": "[27] Haotian Liu, Mu Cai, and Yong Jae Lee. Masked discrimination for self-supervised learning on point clouds. In European Conference on Computer Vision, pages 657-675. Springer, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.223, + 0.47, + 0.292 + ], + "angle": 0, + "content": "[28] Yongcheng Liu, Bin Fan, Gaofeng Meng, Jiwen Lu, Shiming Xiang, and Chunhong Pan. Densepoint: Learning densely contextual representation for efficient point cloud processing. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5239-5248, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.295, + 0.47, + 0.362 + ], + "angle": 0, + "content": "[29] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.366, + 0.469, + 0.407 + ], + "angle": 0, + "content": "[30] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. In International Conference on Learning Representations, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.41, + 0.469, + 0.438 + ], + "angle": 0, + "content": "[31] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.441, + 0.469, + 0.495 + ], + "angle": 0, + "content": "[32] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. In International Conference on Learning Representations, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.498, + 0.469, + 0.552 + ], + "angle": 0, + "content": "[33] Chen Min, Dawei Zhao, Liang Xiao, Yiming Nie, and Bin Dai. Voxel-mae: Masked autoencoders for pre-training large-scale point clouds. arXiv preprint arXiv:2206.09900, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.556, + 0.469, + 0.612 + ], + "angle": 0, + "content": "[34] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In European conference on computer vision, pages 604-621. Springer, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.614, + 0.469, + 0.682 + ], + "angle": 0, + "content": "[35] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.686, + 0.469, + 0.741 + ], + "angle": 0, + "content": "[36] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.744, + 0.469, + 0.813 + ], + "angle": 0, + "content": "[37] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. Advances in Neural Information Processing Systems, 35:23192-23204, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.469, + 0.856 + ], + "angle": 0, + "content": "[38] Shi Qiu, Saeed Anwar, and Nick Barnes. Geometric backprojection network for point cloud classification. IEEE Transactions on Multimedia, 24:1943-1955, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[39] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[40] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[41] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2, 3, 4, 5, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[42] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.264, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[43] Yongming Rao, Jiwen Lu, and Jie Zhou. Global-local bidirectional reasoning for unsupervised representation learning of 3d point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5376-5385, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[44] Justyna Sarzynska-Wawer, Aleksander Wawer, Aleksandra Pawlak, Julia Szymanowska, Izabela Stefaniak, Michal Jarkiewicz, and Lukasz Okruszek. Detecting formal thought disorder by deep contextualized word representations. *Psychiatry Research*, 304:114135, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.406, + 0.892, + 0.447 + ], + "angle": 0, + "content": "[45] Jonathan Sauder and Bjarne Sievers. Self-supervised deep learning on point clouds by reconstructing space. Advances in Neural Information Processing Systems, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.449, + 0.892, + 0.49 + ], + "angle": 0, + "content": "[46] Charu Sharma and Manohar Kaul. Self-supervised few-shot learning on point clouds. Advances in Neural Information Processing Systems, 33:7212-7221, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.492, + 0.892, + 0.547 + ], + "angle": 0, + "content": "[47] Chen Sun, Abhinav Shrivastava, Saurabh Singh, and Abhinav Gupta. Revisiting unreasonable effectiveness of data in deep learning era. In Proceedings of the IEEE international conference on computer vision, pages 843-852, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.549, + 0.892, + 0.618 + ], + "angle": 0, + "content": "[48] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6411–6420, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.62, + 0.892, + 0.688 + ], + "angle": 0, + "content": "[49] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.69, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[50] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1588–1597, 2019. 6, 7, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.775, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4, 5, 6, 7, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[52] Hanchen Wang, Qi Liu, Xiangyu Yue, Joan Lasenby, and Matt J Kusner. Unsupervised point cloud pre-training via occlusion completion. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9782-9792, 2021. 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20963" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[53] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (tog), 38(5):1-12, 2019. 3, 4, 5, 6, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.472, + 0.219 + ], + "angle": 0, + "content": "[54] Ziyi Wang, Xumin Yu, Yongming Rao, Jie Zhou, and Jiwen Lu. P2p: Tuning pre-trained image models for point cloud analysis with point-to-pixel prompting. Advances in neural information processing systems, 35:14388-14402, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.472, + 0.29 + ], + "angle": 0, + "content": "[55] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1912-1920, 2015. 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.472, + 0.375 + ], + "angle": 0, + "content": "[56] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 574-591. Springer, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.472, + 0.431 + ], + "angle": 0, + "content": "[57] Zhenda Xie, Yutong Lin, Zhuliang Yao, Zheng Zhang, Qi Dai, Yue Cao, and Han Hu. Self-supervised learning with swim transformers. arXiv preprint arXiv:2105.04553, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.433, + 0.472, + 0.489 + ], + "angle": 0, + "content": "[58] Yifan Xu, Tianqi Fan, Mingye Xu, Long Zeng, and Yu Qiao. Spidercnn: Deep learning on point sets with parameterized convolutional filters. In Proceedings of the European conference on computer vision (ECCV), pages 87-102, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.472, + 0.56 + ], + "angle": 0, + "content": "[59] Guandao Yang, Xun Huang, Zekun Hao, Ming-Yu Liu, Serge Belongie, and Bharath Hariharan. Pointflow: 3d point cloud generation with continuous normalizing flows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4541-4550, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.562, + 0.472, + 0.618 + ], + "angle": 0, + "content": "[60] Yaoqing Yang, Chen Feng, Yiru Shen, and Dong Tian. Foldingnet: Point cloud auto-encoder via deep grid deformation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 206–215, 2018. 4, 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.619, + 0.472, + 0.687 + ], + "angle": 0, + "content": "[61] Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Russ R Salakhutdinov, and Quoc V Le. Xlnet: Generalized autoregressive pretraining for language understanding. Advances in neural information processing systems, 32, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.472, + 0.759 + ], + "angle": 0, + "content": "[62] Li Yi, Vladimir G Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas Guibas. A scalable active framework for region annotation in 3d shape collections. ACM Transactions on Graphics (ToG), 35(6):1-12, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.472, + 0.83 + ], + "angle": 0, + "content": "[63] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19313-19322, 2022. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.472, + 0.9 + ], + "angle": 0, + "content": "[64] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. Point-m2ae: multi-scale masked autoencoders for hierarchical point cloud pre-training. Advances in neural information processing systems, 35:27061-27074, 2022. 2, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.894, + 0.147 + ], + "angle": 0, + "content": "[65] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16259-16268, 2021. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.894, + 0.205 + ], + "angle": 0, + "content": "[66] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In International Conference on Learning Representations, 2020. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.894, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20964" + } + ] +] \ No newline at end of file diff --git a/2024/General Point Model Pretraining with Autoencoding and Autoregressive/a64a8093-c9f6-4338-aa51-464da1052ad9_origin.pdf b/2024/General Point Model Pretraining with Autoencoding and Autoregressive/a64a8093-c9f6-4338-aa51-464da1052ad9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d6871fc358608676c83c36edffbc7c6fe457a827 --- /dev/null +++ b/2024/General Point Model Pretraining with Autoencoding and Autoregressive/a64a8093-c9f6-4338-aa51-464da1052ad9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:23b42766770bf2af59ddff8955cfe7698d663adc21fe69829a5b1d9f5b09f0b9 +size 2042612 diff --git a/2024/General Point Model Pretraining with Autoencoding and Autoregressive/full.md b/2024/General Point Model Pretraining with Autoencoding and Autoregressive/full.md new file mode 100644 index 0000000000000000000000000000000000000000..399448843ddb0abce6a56378b5985f580ae50181 --- /dev/null +++ b/2024/General Point Model Pretraining with Autoencoding and Autoregressive/full.md @@ -0,0 +1,313 @@ +# General Point Model Pretraining with Autoencoding and Autoregressive + +Zhe Li $^{1,\dagger}$ , Zhangyang Gao $^{3,\dagger}$ , Cheng Tan $^{3,\dagger}$ , Bocheng Ren $^{1}$ , Laurence T. Yang $^{1,2,*}$ Stan Z. Li $^{3,*}$ $^{1}$ Huazhong University of Science and Technology + $^{2}$ Zhengzhou University + +$^{3}$ AI Lab, Research Center for Industries of the Future, Westlake University + +keycharon0122@gmail.com, {gaozhangyang,tancheng,stan.zq.li}@westlake.edu.cn, ltyang@ieee.org + +# Abstract + +The pre-training architectures of large language models encompass various types, including autoencoding models, autoregressive models, and encoder-decoder models. We posit that any modality can potentially benefit from a large language model, as long as it undergoes vector quantization to become discrete tokens. Inspired by the General Language Model, we propose a General Point Model (GPM) that seamlessly integrates autoencoding and autoregressive tasks in a point cloud transformer. This model is versatile, allowing fine-tuning for downstream point cloud representation tasks, as well as unconditional and conditional generation tasks. GPM enhances masked prediction in autoencoding through various forms of mask padding tasks, leading to improved performance in point cloud understanding. Additionally, GPM demonstrates highly competitive results in unconditional point cloud generation tasks, even exhibiting the potential for conditional generation tasks by modifying the input's conditional information. Compared to models like Point-BERT, MaskPoint, and PointMAE, our GPM achieves superior performance in point cloud understanding tasks. Furthermore, the integration of autoregressive and autoencoding within the same transformer underscores its versatility across different downstream tasks. Codes are available at https://github.com/gentlefress/GPM + +# 1. Introduction + +In recent years, the natural language processing (NLP) [4, 19, 20, 29, 40] and computer vision (CV) [3, 7, 9, 49, 57, 66] realms have witnessed a proliferation of transformer-based pretrained models. Their primary advantage lies in the inclusion of massive parameters and data in the training process, overcoming inductive biases introduced by traditional Convolutional Neural Networks [21]. Point clouds serve as fundamental data structures in fields like autonomous driving and robotics, thus emphasizing the escalating significance of tasks related to point cloud representation learn + +![](images/708b3f6f3a8b454058fec970b4754baa20f6e95bca55be7e90f4987b1946cbe3.jpg) +Figure 1. Comparison of pre-training frameworks between GPM and Point-BERT [63]. Point-BERT (left) conducts the masked prediction task in an autoencoding manner, while GPM (right) combines both autoencoding and autoregressive tasks, performing tasks for both masked prediction and masked region blank filling. + +ing and generation. However, the realm of pretrained point cloud models based on transformers remains relatively scarce. Existing transformer-based point cloud models [13, 65] encounter inevitable inductive biases due to local feature aggregation [13], neighbor embedding [65], and the scarcity of annotated data [9]. These biases deviate from the mainstream of standard Transformers. Self-supervised models [12, 15, 20, 25, 26, 39, 40, 61] have emerged as the dominant methodology. They excel in learning high-quality representations across modalities without extensive labeled data, models like GPT [39] even dominate in text generation. In this end, there is a compelling need to design a transformer-based point cloud model that minimizes inductive biases, learns superior point cloud representations from limited data, and simultaneously possesses the capacities for conditional and unconditional point cloud generation. The transformer-based BERT [20] has marked a significant milestone in the field of natural language processing (NLP). Its pre-training phase, involving masked prediction tasks, grants the model the capacity to learn language representations, showcasing the prowess of autoencoding language models. Given BERT's proficiency in learning language representations, the question arises: can we extend BERT's + +capabilities to the realm of point clouds? However, since point clouds lack a direct analog to the concept of 'words' in language, constructing a discrete vocabulary for point clouds, as done in language models, is infeasible. Point clouds are composed of individual points, and considering each point as a token would lead to a quadratic increase in computational cost as the number of tokens grows. Furthermore, each point in a point cloud contains limited semantic information, necessitating the maximization of geometric information utilization. Inspired by Point-BERT [63], we embarked on a journey to vector quantization, transforming point clouds into discrete tokens. Through the application of the farthest point sampling (FPS) [36] algorithm, we segment the point cloud into multiple patches, each patch encapsulates plentiful geometric information, akin to a unit. By constructing a vocabulary specific to the point cloud domain, we obtain labels for each unit, discretizing the entire point cloud. This paves the way for the completion of the autoencoding model's masked prediction tasks in the point cloud domain. + +Transformer-based GPT [39] has dominated the field of text generation, with its core being an autoregressive language model, predicting the next token based on preceding tokens. In recent years, researchers have progressively extended this paradigm into the realm of images [7, 41, 42], ushering in a new era of image generation. In this endeavor, we bring this paradigm to the domain of point clouds, aiming to empower it for autoregressive conditional and unconditional point cloud generation tasks. After discretization, each patch of the point cloud possesses distinct geometric information, enabling it to excel in autoregressive tasks. + +While these pre-training frameworks can be adapted to the point cloud domain, they lack the flexibility to satisfy all point cloud tasks. Therefore, we aim to integrate these two pre-training frameworks from the NLP domain and apply them to the point cloud domain, depicted in Figure 1. + +In this work, we introduce a novel approach termed General Point Model (GPM) with autoencoding and autoregressive, which integrates point cloud representation learning and point cloud generation (both conditional and unconditional) within the same transformer framework. Similar to Point-BERT, we employ a point cloud tokenizer designed through dVAE-based point cloud reconstruction, where the point cloud can be discretely labeled based on the learned vocabulary. Additionally, we utilize the Masked Point Modeling (MPM) task. This involves partitioning the point cloud into distinct patches and randomly masking some adjacent patches. This aims to predict the masked portion, with the goal of learning geometric information between neighboring blocks and capturing meaningful geometric features for understanding the point cloud. + +Our primary contributions can be summarized as follows: + +- We introduce autoencoding + autoregressive as a novel point cloud transformer paradigm, unifying tasks in point cloud understanding and generation. +- We partition the input into two segments, the first part predicts masked points (autoencoding), the second part generates masked points (autoregressive), enhancing comprehension and generation. +- Our approach demonstrates notable competitiveness across various point cloud understanding tasks and unconditional generation tasks. Moreover, it exhibits the potential for conditional generation, such as text- or image-conditioned point clouds generation. + +# 2. Related Work + +# 2.1. Self-supervised Learning + +Self-supervised learning (SSL) has garnered significant attention owing to the substantial labor required for acquiring annotated data in large quantities. SSL enables models to learn feature representations from unlabeled data, adapting to downstream tasks. The core of SSL lies in the design of proxy tasks to replace traditional classification tasks, thereby learning feature representations. ELMo [44] employs bidirectional LSTMs [16] and generates subsequent words from left to right given the representation of preceding content. ViP [25] employs a dynamically updated momentum encoder for contrastive learning and designs a text swapping task to enhance the model's sentence representation capability. In the computer vision domain, Image-GPT [7] trains a sequence transformer to autoregressively predict pixels, showing promising representation learning capabilities without incorporating specific knowledge about the 2D input structure after pretraining. Furthermore, SSL in the field of point clouds has garnered significant attention. ACT [8] employs cross-modal autoencoders as teacher models to acquire knowledge from other modalities. The generative approaches [1, 22, 33, 45, 63, 64] are the most pertinent research to our work. Point-MAE [34] extends MAE by randomly masking point patches and reconstructing the masked regions. Point-M2AE [64] further utilizes a hierarchical transformer architecture and designs corresponding masking strategies. However, mask-based point modeling methods still face the issue of shape leakage, limiting their effective generalization to downstream tasks. In this paper, we leverage autoregressive pretraining on point clouds and address the unique challenges associated with point cloud attributes. Our concise design avoids position information leakage, thereby enhancing generalization capabilities. + +# 2.2. General Language Model + +In the field of natural language processing, GLM [10] integrates both autoregressive and autoencoding methods. This enables it to perform tasks related to sentence representa + +![](images/0fa6104b24bfd09b60eb3ba44183d81ac3af200dc0b164ea659119ea5aab469a.jpg) +Figure 2. The framework of our GPM. We divide the entire pipeline into two stages: 1) The input point cloud is divided into several subclouds. Then, we utilize Mini-PointNet [35] to obtain a sequence of point embeddings. These embeddings are transformed into a sequence of discrete point tokens using a pre-trained tokenizer; 2) During pre-training, for PartA, certain portions of the masked embeddings are masked and replaced with [M]. They are predicted in an autoencoding manner. For PartB, masks for PartA are generated in an autoregressive manner. + +![](images/811f0447454e5f41f7d95a3bd47d25a4b52e82cd4fe638ddd16f4d29061de711.jpg) + +tion as well as conditioned and unconditioned generation through fine-tuning. To the best of our knowledge, we are the first to integrate autoregressive and autoencoding techniques on a single transformer in the domain of point clouds. In downstream tasks, we have the potential to fine-tune point cloud representations and perform unconditioned generation even in zero-shot scenarios, with the possibility of fine-tuning towards conditioned generation. + +# 3. Methods + +In this study, we endeavor to seamlessly integrate BERT-style and GPT-style pre-trained strategies, extending their applicability to point cloud transformers. Following the Point-BERT[63], our training process unfolds in two distinctive stages. Firstly, we embark on training a specialized Tokenizer, a critical step towards acquiring discrete point labels for each input point cloud. Concurrently, a dedicated discrete VAE (dVAE) is employed to masterfully reconstruct the discrete point cloud. The subsequent stage involves the splicing of two sequences of point cloud embeddings, partA for autoencoding and partB for autoregressive. The overall idea of our approach is illustrated in Figure 2. + +# 3.1. Stage 1: Discrete Varitional Autoencoder Pretraining + +Point Cloud Partitioning. We posit that discrete tokens derived from points encapsulate crucial geometric information, allowing for a discrete representation of any point cloud. However, a naive strategy of assigning one token to each point in a point-wise reconstruction task poses a formidable computational challenge. This arises from the quadratic complexity of self-attention within transformers. + +Building upon the foundations laid by Point-BERT [63] and ViT[9], we adopt a strategy of partitioning each point cloud into distinct patches, each serving as a single token. Specifically, for an input point cloud $p \in \mathbb{R}^{\mathcal{N} \times 3}$ , we initiate the process by extracting $m$ center points $\mathcal{C}_u \in \mathbb{R}^{m \times 3}$ from the point cloud $p$ through farthest point sampling (FPS). Subsequently, we employ the k-nearest neighbor (kNN) algorithm to identify the $k$ nearest neighboring points for every center point. This forms $m$ localized patches or sub-clouds denoted as $\mathcal{P}_u \in \mathbb{R}^{m \times k \times 3}$ . To ensure these patches are free of bias, we normalize them by subtracting their respective center coordinates. This operation effectively disentangles the structural patterns from the spatial coordinates inherent to each local patch. + +Point Cloud Embedding. Following Point-BERT [63], we employ a mini-PointNet [35] to embed the point patches as a sequence of point embedding $\{h_i\}_{i=1}^m$ . For the further vector quantization, we pre-define a learnable codebook $\mathcal{V} := \{(s, z_s)\}_{s \in S}^{d_z}$ , where $d_z$ is the dimension of codes, $S$ is the size of codebook and $s$ is the index of embedding in $\mathcal{V}$ . We adopt DGCNN [53] as tokenizer $\mathcal{Q}: h_i \mapsto z_i$ , which maps $\{h_i\}_{i=1}^m$ into $\{z_i\}_{i=1}^m$ in $\mathcal{V}$ . However, given the non-differentiable nature of the discrete tokens, applying reparameterization gradients for dVAE training becomes unfeasible. As suggested in [41], we resort to the Gumbel-softmax relaxation technique [18], coupled with a uniform prior, as a workaround during the dVAE training process. + +Point Cloud Reconstruction. Given $\{z_i\}_{i=1}^m$ as input, the decoder $\mathcal{D}(\cdot)$ is tasked with the reconstruction of the entire point cloud. Inspired by [63], to effectively capture inter + +point relationships and bolster the representational capacity of discrete point tokens across various local structures, we adopt the DGCNN [53]-FoldingNet [60] architecture to reconstruct the whole point cloud. + +The overarching reconstruction objective can be denoted as $\mathbb{E}_{z\sim \mathcal{Q}(z|h)}\log D(\hat{p} |z)$ . Inspired by Point-BERT [63], the entire reconstruction process can be conceptualized as the maximization of the Evidence Lower Bound (ELB) for the log-likelihood, denoted as $\mathcal{D}(\hat{p} |p)$ : + +$$ +\begin{array}{l} \sum_ {i = 1} ^ {\mathcal {N}} \log D (\hat {p} _ {i} | p _ {i}) \geq \sum_ {i = 1} ^ {\mathcal {N}} \left(\mathbb {E} _ {z _ {i}} \sim \mathcal {Q} (z _ {i} | p _ {i})\right) [ \log D (p _ {i} | z _ {i}) ] \\ - \operatorname {K L} \left[ \mathcal {Q} \left(z \mid p _ {i}\right), \mathcal {D} \left(z \mid \hat {p} _ {i}\right) \right]. \tag {1} \\ \end{array} +$$ + +Simultaneously, we account for an intuitive reconstruction loss, leveraging the $l1$ Chamfer Distance [11] to supervise prediction point cloud with the ground-truth point cloud: + +$$ +\mathcal {L} _ {\mathrm {C D}} (\mathcal {P}, \mathcal {G}) = \frac {1}{| \mathcal {P} |} \sum_ {p \in \mathcal {P}} \min _ {g \in \mathcal {G}} \| p - g \| + \frac {1}{| \mathcal {G} |} \sum_ {g \in \mathcal {G}} \min _ {p \in \mathcal {P}} \| g - p \|, \tag {2} +$$ + +where $\mathcal{P}$ denotes the predicted point clouds and $\mathcal{G}$ is the ground truth point clouds. Additional, we follow [41], optimize the KL-divergence $\mathcal{L}_{\mathrm{KL}}$ between the generated point cloud distribution and a uniform distribution: + +$$ +\mathcal {L} _ {\mathrm {d V A E}} = \mathcal {L} _ {\mathrm {C D}} + \mathcal {L} _ {\mathrm {K L}}. \tag {3} +$$ + +# 3.2. Stage 2: GPM Pre-training + +To ensure GPM excel in both point cloud representation learning and point cloud generation task, we formulate a training framework that amalgamates autoencoding and autoregressive techniques motivated by [10]. + +Autoencoding Masked Sequence Generation. We employ the standard Transformers [51], which encompass multi-headed self-attention layers and FFN blocks. Initially, we partition each input point cloud into $m$ local patches with designated center points. Subsequently, these local patches undergo discretization into a code sequence $\{f_i\}_{i=1}^m$ via our pre-trained dVAE in stage 1. Formally, the input embeddings $\{e_i\}_{i=1}^m$ are constructed as a combination of point embeddings $\{f_i\}_{i=1}^m$ and positional embeddings $\{pos_i\}_{i=1}^m$ . Adhering to language model [20], a class token [CLS] is appended to the input sequences, rendering the transformer input sequence as $I = \{[\mathrm{CLS}], e_1^0, e_2^0, \dots, e_i^0\}$ . + +Drawing inspiration from [63], we initiate by selecting a central point $\mathcal{C}_i$ alongside its corresponding sub-cloud $\mathcal{P}_i$ , integrate it with $m$ neighboring sub-clouds as a coherent local region. Within this region, we apply a masking operation to obscure all local patches, generating what we refer to as the "masked point cloud". Specifically, we substitute all the masked point embeddings with a universally learnable + +pre-defined mask embeddings [M]. We mark the masked position as $\mathcal{M} = \{1,\dots,b\}$ , and the final input embeddings $E^{\mathcal{M}} = \{e_i,i\notin \mathcal{M}\} \cup \{[\mathbf{M}] + pos_i,i\in \mathcal{M}\}$ are fed into the transformer. + +The objective of our Masked Point Modeling (MPM) task is to deduce the geometric arrangement of absent portions and reconstruct point tokens aligned with the masked positions using the available information. Geometric information refers to the relative positions and morphological attributes of points within a localized region. These crucial geometric details provide vital cues for understanding and processing the shape and structure of the local region, playing a pivotal role in delving deep into the intricacies of point cloud data. Formally, the pre-training objective can be expressed as maximizing the log-likelihood of the point tokens $e_i$ conditioned on the masked input embeddings $E^{\mathcal{M}}$ : + +$$ +\max \sum_ {E \in W} \mathbb {E} \left[ \sum_ {i \in \mathcal {M}} \log P \left(e _ {i} \mid E ^ {\mathcal {M}}\right) \right], \tag {4} +$$ + +where $W$ is the set of all input embeddings. Similar to [63], we also adopt Point Patch Mixing and contrastive learning to help the model to better learn high-level semantics. Point Patch Mixing involves mixing sub-clouds without complex alignment techniques. However, training solely with masked precition has limitations in understanding high-level semantics. To overcome this, we incorporate contrastive learning to enhance semantic understanding. Assuming the masking ratio is $\beta$ , the loss function is defined as: + +$$ +- \beta \log \frac {\exp \left(q \hat {k} _ {1} / \tau\right)}{\sum_ {i = 0} ^ {K} \exp \left(q k _ {i}\right)} - (1 - \beta) \log \frac {\exp \left(q \hat {k} _ {2} / \tau\right)}{\sum_ {i = 0} ^ {K} \exp \left(q k _ {i}\right)}, \tag {5} +$$ + +where $q$ is the feature of a mixed sample that comes from two other samples, whose features are $\hat{k_1}$ and $\hat{k_2}$ . $k_{i}$ are extracted by the momentum feature encoder, $\tau$ is the temperature and $K$ is the size of memory bank. + +Autoregressive Sequence Generation. While Point-BERT models the relationship between unmasked and masked regions using the Masked Point Modeling (MPM) task, it does not adequately model the interactions within masked regions. Furthermore, our aim extends beyond acquiring high-quality point cloud representations; we also aspire to proficiently execute point cloud generation tasks. Hence, for the latter segment, PartB, we engage in autoregressive mask generation tasks to enhance the MPM task. + +In the case of PartB, the mask operation is ignored. Our focus lies in acquiring the generated tokens from the masked positions of PartA. We take the initial $n - 1$ tokens from PartA, appending the start token [S] at its forefront to get the PartB $E^{\mathcal{B}} = \{[\mathrm{S}], e_1, \dots, e_{n-1}\}$ . Following the autoregressive approach, the prediction of the next token relies + +on the preceding token: + +$$ +\max _ {\theta} \mathbb {E} \left[ \sum_ {i = 1} ^ {n} \log P _ {\theta} \left(e _ {i} | \mathrm {P a r t A}, e _ {< i}, [ \mathrm {S} ]\right) \right]. \tag {6} +$$ + +Although no masking operation is performed on PartB, since the mask positions of PartA are multiple consecutive sub-clouds instead of a single one, in PartB we are also equivalent to autoregressive generation of a continuous mask span. We extract the generated tokens from PartB, corresponding to the masked positions in PartA, and compare them with the discrete tokens produced by the Tokenizer: + +$$ +\mathcal {L} _ {A R} = \operatorname {C r o s s E n t r o p y} \left(e _ {j, j \in \mathcal {M}}, \operatorname {a r g m a x} \mathcal {Q} (h) _ {j, j \in \mathcal {M}}\right). \tag {7} +$$ + +In this way, it allows our model to acquire proficiency in an autoencoding bidirectional encoding scheme for PartA, and an autoregressive unidirectional encoder for PartB. + +# 3.3. Multi-task Pretraining + +The masking of $15\%$ of tokens in BERT is tailored for downstream natural language understanding tasks, whereas Point-BERT masks $25\% \sim 45\%$ to acquire enhanced point cloud representations. Our GPM focus on both point cloud representation learning and point cloud generation, we concatenate PartA and PartB as input to the transformer, aiming to simultaneously perform both autoencoding and autoregressive tasks. Therefore, we need to establish specific attention masks to facilitate the interaction of information between these two segments. + +Motivated by [10], the tokens in PartA can participate in MLM task in PartA and autoregressive generation in PartB. PartB tokens can not be observed by PartA, but those antecedents. PartA is treated as the conditioning for autoregressive generation in PartB, akin to prefix tuning [23] in NLP. The implement of GPM multi-task pretraining is depicted in Figure 3. + +The tokens output by the pretrained tokenizer $\mathcal{Q}(\cdot)$ are regarded as labels. PartA conducts a masked prediction task in an autoencoding manner, and the prediction loss between the predicted masked portion and the labels is computed as follows: + +$$ +\mathcal {L} _ {A E} = \operatorname {C r o s s E n t r o p y} \left(e _ {j, j \in \mathcal {M}}, \operatorname {a r g m a x} \mathcal {Q} (h) _ {j, j \in \mathcal {M}}\right). \tag {8} +$$ + +Although Equation 7 and 8 exhibit identical formal structures. However, one entails mask prediction through autoencoding, while the other involves mask generation in an autoregressive way, indicating a fundamental distinction between the two approaches. + +# 4. Experiments + +In this section, we commence by outlining the configurations of our pretraining scheme. Subsequently, we pro + +![](images/dce3458aaa424daa59ba179a4a38dbf740e92d73722392201881b9ccfd2d985f.jpg) +Figure 3. Self-attention mask. The regions in grey are masked. Tokens in PartA can attend to themselves, but not to PartB. Tokens in PartB can attend to PartA and their antecedents in PartB. + +ceed to assess the proposed model through various downstream tasks, encompassing object classification, partial segmentation, few-shot learning, transfer learning (to validate the model's representation learning capabilities), and point cloud generation (to validate its generative capabilities). Additionally, we conduct ablation studies on GPM. + +# 4.1. Pre-training Setups and Implementation + +Pre-training Data. Following the dataset configuration similar to Point-BERT [63], we employ ShapeNet [5] as our pretraining dataset, encompassing over 50,000 unique 3D models spanning 55 common object categories. From each 3D model, we sample 1024 points, dividing them into 64 point patches (sub-clouds) and each sub-cloud contains 32 points. Utilizing a lightweight PointNet [35] with two MLP layers, we project each sub-cloud into a 64-point embedding, serving as input for both dVAE and Transformer [51]. + +dVAE Pretraining Setups. The purpose of pretraining the dVAE is to acquire a high-quality tokenizer, enabling the reconstruction of discretized features back into the original point cloud to the fullest extent. Our dVAE is composed of a tokenizer and a decoder. To be specific, the tokenizer consists of a 4-layer DGCNN [53], while the decoder encompasses a 4-layer DGCNN followed by a FoldingNet [60]. Additionally, the FoldingLayer establishes a connection between a 2D grid and the input, ultimately generating a 3D point cloud. + +During the training of dVAE [41], we set the number of words in the codebook $N$ to 8192. We employ the common $l1$ -style Chamfer Distance loss in the reconstruction process. Due to the small numerical value of this $l1$ loss, the weight of the KL loss in Equation 1 must be smaller than that in image tasks. In the initial 10,000 steps of training, + +the weight of the KL loss is kept at 0 and gradually raised to 0.1 in the subsequent 100,000 steps. Our learning rate is set to 0.0005, and follows a cosine annealing schedule spanning 60,000 steps. Consistent with prior works [41] and Point-BERT [63], we decay the temperature in the Gumblesoftmax function from 1 to 0.0625 over 100,000 steps. The training of dVAE encompasses a total of 150,000 steps, with a batch size of 64. It's worth noting that our dVAE's network architecture closely mirrors that of Point-BERT. + +GPM Pretraining Setups. In our experiments, we adhere to the standard Transformer architecture [9], which consists of a stack of Transformer blocks [51], each composed of multi-head self-attention layers and feed-forward networks. LayerNorm is applied in both layers. We set the depth of our Transformer to 12, feature dimension to 384, and the number of heads to 6. A random depth of 0.1 [17] is applied in our Transformer encoder. During GPM pretraining, the weights of the Tokenizer learned by dVAE are kept fixed. Input point embeddings from $25\% \sim 45\%$ are randomly masked, and the model is trained to infer the expected point labels at these masked positions. Unlike Point-BERT [63], which does not have a generation task during pretraining, we incorporate both autoregressive and autoencoding tasks in our work. Thus, an attention mask is necessary, as illustrated in Figure 3. For MoCo, we set the same configuration as [63]. More Details can be found in the supplementary 6.1 and 6.2. + +# 4.2. Downstream Tasks + +# 4.2.1 Point Cloud Representation Evaluation + +Object classification on ModelNet40. We evaluate our pre-trained model on the ModelNet40 dataset [55], comprising 12,311 clean 3D CAD models spanning 40 categories. Following the Point-BERT [63] setups, we employ a two-layer MLP with a dropout rate of 0.5 as the classification head for the task. We optimize the model using AdamW with a weight decay of 0.05 and a learning rate of 0.0005, while employing a batch size of 32 and a cosine annealing schedule. We conduct comparisons with various Transformer-based models with identical Point-BERT settings, denoting [ST] for the standard Transformer architecture and [T] for Transformer models with specific designs or inductive biases. The results in Table 1 demonstrate that our model not only outperforms Point-BERT in classification metrics on this dataset but also exhibits autoregressive generation capabilities. + +Object classification on ScanObjectNN. The practical applicability of our model on real datasets serves as a crucial metric. Therefore, the pre-trained models are transferred to the ScanObjectNN dataset [50], which comprises + +
Methodsnumber of pointsAcc
Supervised Learning
PointNet[35]1k89.2
SO-Net[22]1k92.5
PointNet++ [36]1k90.5
PointCNN[24]1k92.2
DGCNN[53]1k92.9
MVTN[14]1k93.8
RSCNN [43]1k92.9
GBNet[38]1k93.8
PointMLP[32]1k94.5
DensePoint[28]1k92.8
PointNeXt[37]1k94.0
P2P-RN101[54]1k93.1
P2P-HorNet[54]1k94.0
KPConv[48]~6.8k92.9
Self-Supervised Representation Learning
[T]PCT [65]1k93.2
[T]PointTransformer [13]-93.7
[ST]NPCT [13]1k91.0
[ST]Transformer[51]1k91.4
[ST]Transformer-OcCo[51]1k92.1
[ST]Point-BERT[63]1k93.2
[ST]MaskPoint[27]1k93.8
[ST]Point-MAE[34]1k93.8
[ST]GPM1k94.0
Methods with Cross-modal Information and Teacher Models
ACT[8]1k93.7
[ST]Transformer[51]4k91.2
[ST]Transformer-OcCo[51]4k92.2
[ST]Point-BERT[63]4k93.4
[ST]GPM4k93.1
[ST]Point-BERT[63]8k93.8
[ST]MaskPoint[27]8k-
[ST]Point-MAE[34]8k94.0
[ST]GPM8k94.3
+ +Table 1. Classification results on ModelNet40. All results are expressed as percentages accuracy. + +approximately 15,000 objects extracted from real indoor scans, encompassing 2902 point clouds from 15 categories. This dataset presents a greater challenge as it involves sampling from real-world scans with backgrounds and occlusions. We conducted experiments on three main variants, namely OBJ-BG, OBJ-ONLY, and PB-T50-RS, in line with prior work. The experimental results are summarized in Table 2. It is observed that GPM exhibits an improvement of approximately $2.77\%$ , $1.88\%$ , and $1.73\%$ over the regular Point-BERT across the three variants. + +Few-shot learning. To demonstrate the ability to acquire knowledge for new tasks with limited training data, we evaluate our model under the setting of few-shot learning, following the methodology of previous work [47, 63]. In the typical 'W-way S-shot' setup, we initially randomly select $W$ classes and then sample $(S + 20)$ objects for each class [46]. The model is trained on $W \times S$ samples (support set) and evaluated on the remaining $20W$ samples (query set). We conduct 10 independent experiments for each setting and report the average performance and standard deviation across the 10 runs. As shown in Table 3, our approach outperforms other methods in all tests, achieving absolute improvements of $0.4\%$ , $0.4\%$ , and $0.6\%$ over Point-BERT. + +
MethodsOBJ-BGOBJ-ONLYPB-T50-RS
Supervised Learning
PointNet[35]73.379.268.0
SpiderCNN[58]77.179.573.7
PointNet++ [36]82.384.377.9
PointCNN[24]86.185.578.5
DGCNN[53]82.886.278.1
MVTN[14]92.692.382.8
BGA-DGCNN[50]--79.7
BGA-PN++ [50]--80.2
GBNet[38]--81.0
PointMLP[32]--85.4
PointNeXt[37]--87.7
P2P-RN101[54]--87.4
P2P-HorNet[54]--89.3
Self-Supervised Representation Learning
Transformer[51]79.8680.5577.24
Transformer-OcCo[51]84.8585.5478.79
Point-BERT[63]87.4388.1283.07
MaskPoint[27]89.388.184.3
Point-MAE[34]90.088.285.2
GPM90.290.084.8
+ +Table 2. Classification results on ScanObjectNN. All results are expressed as percentages accuracy. + +
Methods5-way10-way
10-shot20-shot10-shot20-shot
Supervised Learning
DGCNN-rand [53]31.6 ± 2.840.8 ± 4.619.9 ± 2.116.9 ± 1.5
OcCo [52]90.6 ± 2.892.5 ± 1.982.9 ± 1.386.5 ± 2.2
Self-Supervised Representation Learning
Point-BERT [63]94.6 ± 3.196.3 ± 2.791.0 ± 5.492.7 ± 5.1
MaskPoint [27]95.0 ± 3.797.2 ± 1.791.4 ±4.093.4 ± 3.5
Piont-MAE [34]96.3 ± 2.597.8 ± 1.892.6 ± 4.195.0 ± 3.0
Piont-M2AE [64]96.8 ± 1.898.3 ± 1.492.3 ± 4.595.0 ± 3.0
GPM97.2 ± 2.698.7 ± 2.292.9 ± 4.295.0 ± 3.0
+ +Part segmentation. We evaluate the representation learning capability of our approach on the ShapeNetPart dataset [62], aiming to predict more fine-grained class labels for each point. This dataset consists of 16 categories and comprises 16,881 objects. The point cloud is downsampled to 2048 points, and the segmentation head [34] connects features $\mathcal{F}^4, \mathcal{F}^8, \mathcal{F}^{12}$ extracted from the 4-th, 6-th, and 12-th layers of the transformer blocks. Subsequently, average pooling, max pooling, and upsampling are employed to generate features for each point, followed by label prediction using MLP. In our experiments, we evaluate the performance under the settings of "5 way 10 shot", "5 way 20 shot", "10 way 10 shot", and "10 way 20 shot". The experimental results shown in Table 4 demonstrate the superior performance of our GPM compared to all other methods. + +# 4.2.2 Point Cloud Generation Evaluation + +In the pre-training phase, we perform unconditional autoregressive generation on tokens in PartB, endowing the model with the capability of point cloud unconditional generation. + +Table 3. The results of few-shot classification on the ModelNet40 dataset. For each experimental setting, we conduct 10 independent experiments and report the average accuracy $(\%)$ along with its standard deviation. + +
MethodsCls.mIoUInst.mIoU
Supervised Learning
PointNet[35]80.3983.7
PointNet++ [36]81.3585.1
DGCNN[53]82.3385.2
PointMLP [32]84.686.1
Self-Supervised Representation Learning
Transformer[51]83.4285.1
Transformer-OcCo[51]83.4285.1
PointContrast [56]-85.1
CrossPoint [2]-85.5
Point-BERT [63]84.1185.6
Point-MAE [34]-86.1
GPM84.2085.8
+ +Table 4. Part segmentation results on the ShapeNetPart dataset. We report the average intersection mIoU over the union of all classes (Cls.) and instances (Inst.). + +
CategoryModelJSD (↓)MMD (↓)Cov (%, ↑)1-NNA % (%, ↓)
CDEMDCDEMDCDEMD
AirplanePoint-Flow4.920.2173.2446.9148.4075.6875.06
Point-BERT4.070.1892.7550.1753.4470.1661.07
GPM3.320.1762.4752.3456.4468.2260.30
ChairPoint-Flow1.742.427.8746.8346.9860.8859.89
Point-BERT1.231.827.1149.7350.0757.3056.61
GPM0.981.676.8251.0852.1454.0954.27
CarPoint-Flow0.870.915.2244.0346.5960.6562.36
Point-BERT0.730.824.9647.2752.1359.0960.48
GPM0.660.754.3849.6754.1758.8858.16
+ +Table 5. Generation results on ShapeNet dataset. $\uparrow$ : the higher the better, $\downarrow$ : the lower the better. MMD-CD is multiplied by $10^{3}$ ; MMD-EMD and JSD are multiplied by $10^{2}$ . + +Therefore, in downstream tasks, without fine-tuning, we conduct an autoregressive point cloud generation task on the ShapeNet dataset. Existing models do not possess both point cloud generation and point cloud classification tasks. Hence, our model is the first known point cloud transformer that integrates these two tasks into one. The specific metrics are shown in Table 5. We compare our results with generation model Point-Flow [59] and Point-BERT. More details about downstream tasks and conditional point cloud generation results can be found in the supplementary 6.3. + +# 4.3. Ablation Study + +To validate the fundamental design of the GPM model, we conducted a comprehensive set of ablation studies. In order to assess the impact of autoregressive mask generation on the autoencoding masked prediction task, we present results comparing fine-tuning scenarios on the ModelNet40 dataset: one with autoencoding only, and the other with a combination of autoencoding + autoregressive. Furthermore, to demonstrate that the order of autoencoding and autoregressive tasks in the sequence does not affect point cloud understanding, we conduct additional experiments, the results are summarized in the Table 6. More ablation results are shown in the supplementary 6.4. + +![](images/a8f2f5445a6e4bb216661e0a0b6ff10c748c5c72104235068ee35b6edc9d2f14.jpg) +(a) Point cloud generation results of dVAE. + +![](images/9d24243c1e5ac04edb415f0ce15298a8bba6eff659116e964425f2e82847a49a.jpg) +(b) Point cloud generation results of Point-BERT. +Figure 4. The reconstruction results visualization on dVAE and Point-BERT. + +
TasksnpointsAcc
autoencoding only1k92.9
autoregressive ⇒ autoencoding1k93.6
autoencoding ⇒ autoregressive1k93.8
autoencoding only4k93.4
autoregressive ⇒ autoencoding4k93.9
autoencoding ⇒ autoregressive4k93.9
+ +Table 6. Results on autoencoding, autoencoding $\Rightarrow$ autoregressive and autoregressive $\Rightarrow$ autoencoding on the ModelNet dataset. It is clear that autoregressive can enhance the completion accuracy of autoencoding, the order of autoencoding and autoregressive does not have a significant impact on point cloud understanding tasks. + +# 4.4. Analysis + +Compare with Point-BERT. Motivated by BERT, Point-BERT leverages MPM as a pretext task for training. Due to the assumption of independence between MLM and MPM, Point-BERT fails to capture interdependencies among masked tokens. In our approach, autoregressive generation explicitly models dependencies between masked regions, unlike autoencoding which relies solely on un-masked tokens to predict the masked ones. Autoregressive masked prediction allows previously predicted tokens to influence subsequent predictions, enhancing information exchange between them. This leads to improved performance in downstream tasks in the realm of representation learning. Simultaneously, tokens in PartA can attend to task in PartB, making PartA a generation condition for PartB, resulting in enhanced generation performance. + +Compare with PointGPT. PointGPT [6] adopts a two-stage transformer architecture for learning representations of point clouds, comprising a feature extractor and a generator. It first extracts features from the point cloud us + +ing the feature extractor and then directly generates with the obtained features in the generator. During training, it directly employs Chamfer Distance [11] between generated point clouds and ground truth points as the optimization objective, rather than operating on tokens. This is because it doesn't discretize points embeddings, thus failing to achieve genuine autoregressive generation tasks. Moreover, it doesn't integrate the two tasks for joint training in a single transformer; instead, it trains them separately on two different transformers, which increases the complexity. + +# 4.5. Visualization + +We visualize the generation results of dVAE and Point-BERT, as shown in the Figure 4a and Figure 4b. Although Point-BERT is capable of masked tokens generation through autoencoding, our approach allows for autoregressive conditional generation tasks, such as text-conditioned point clouds and image-conditioned point clouds. As observed in Figure 4b, the generation results of Point-BERT exhibit some scattered points at the edges of the overall structure, appearing relatively sparse and disorganized. The generated outcome is not as satisfactory as that of GPM. More visualizations of generation results can refer to Unconditional and Conditional Point Generation in the supplementary 6.3. + +# 5. Conclusion + +Our study integrates both autoencoding and autoregressive tasks within a unified transformer to enhance the model's performance in representation learning and generation at the same time. By this framework, we address the limitation of current pretraining frameworks and make the proxy tasks non-trivial, effectively enhance the information interaction among masked tokens. Moreover, our framework endows the model with the potential for conditional generation. + +# Acknowledgement + +This work was supported by National Science and Technology Major Project (No.2022ZD0115101), National Natural Science Foundation of China Project (No.U21A20427). + +# References + +[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In International conference on machine learning, pages 40-49. PMLR, 2018. 2 +[2] Mohamed Afham, Isuru Dissanayake, Dinithi Dissanayake, Amaya Dharmasiri, Kanchana Thilakarathna, and Ranga Rodrigo. Crosspoint: Self-supervised cross-modal contrastive learning for 3d point cloud understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9902-9912, 2022. 7 +[3] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. In International Conference on Learning Representations, 2021. 1 +[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1 +[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 5 +[6] Guangyan Chen, Meiling Wang, Yi Yang, Kai Yu, Li Yuan, and Yufeng Yue. Pointgpt: Auto-regressively generative pre-training from point clouds. arXiv preprint arXiv:2305.11487, 2023. 8 +[7] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International conference on machine learning, pages 1691-1703. PMLR, 2020. 1, 2 +[8] Runpei Dong, Zekun Qi, Linfeng Zhang, Junbo Zhang, Jianjian Sun, Zheng Ge, Li Yi, and Kaisheng Ma. Autoencoders as cross-modal teachers: Can pretrained 2d image transformers help 3d representation learning? In The Eleventh International Conference on Learning Representations, 2022. 2, 6 +[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 1, 3, 6 +[10] Zhengxiao Du, Yujie Qian, Xiao Liu, Ming Ding, Jiezhong Qiu, Zhilin Yang, and Jie Tang. Glm: General language model pretraining with autoregressive blank infilling. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 320-335, 2022. 2, 4, 5 +[11] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 605-613, 2017. 4, 8 +[12] Tianyu Gao, Xingcheng Yao, and Danqi Chen. Simcse: Simple contrastive learning of sentence embeddings. In 2021 + +Conference on Empirical Methods in Natural Language Processing, EMNLP 2021, pages 6894-6910. Association for Computational Linguistics (ACL), 2021. 1 +[13] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. Pct: Point cloud transformer. Computational Visual Media, 7:187-199, 2021. 1, 6 +[14] Abdullah Hamdi, Silvio Giancola, and Bernard Ghanem. Mvtn: Multi-view transformation network for 3d shape recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1-11, 2021. 6, 7 +[15] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 1 +[16] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997. 2 +[17] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q Weinberger. Deep networks with stochastic depth. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 646-661. Springer, 2016. 6 +[18] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. In International Conference on Learning Representations, 2016. 3 +[19] Mandar Joshi, Danqi Chen, Yinhan Liu, Daniel S Weld, Luke Zettlemoyer, and Omer Levy. Spanbert: Improving pre-training by representing and predicting spans. Transactions of the association for computational linguistics, 8:64-77, 2020. 1 +[20] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, pages 4171–4186, 2019. 1, 4 +[21] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012. 1 +[22] Jiaxin Li, Ben M Chen, and Gim Hee Lee. So-net: Self-organizing network for point cloud analysis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9397–9406, 2018. 2, 6 +[23] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 4582-4597, 2021. 5 +[24] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. Advances in neural information processing systems, 31, 2018. 6, 7 +[25] Zhe Li, T. Yang Laurence, Xin Nie, BoCheng Ren, and Xianjun Deng. Enhancing sentence representation with visually-supervised multimodal pre-training. In Proceedings of the 31st ACM International Conference on Multimedia (ACM MM'23), 2023. 1, 2 + +[26] Zhe Li, Laurence T Yang, Bocheng Ren, Xin Nie, Zhangyang Gao, Cheng Tan, and Stan Z Li. Mlip: Enhancing medical visual representation with divergence encoder and knowledge-guided contrastive learning. arXiv preprint arXiv:2402.02045, 2024.1 +[27] Haotian Liu, Mu Cai, and Yong Jae Lee. Masked discrimination for self-supervised learning on point clouds. In European Conference on Computer Vision, pages 657-675. Springer, 2022. 6, 7 +[28] Yongcheng Liu, Bin Fan, Gaofeng Meng, Jiwen Lu, Shiming Xiang, and Chunhong Pan. Densepoint: Learning densely contextual representation for efficient point cloud processing. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5239-5248, 2019. 6 +[29] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 1 +[30] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. In International Conference on Learning Representations, 2016. 1 +[31] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 1 +[32] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. In International Conference on Learning Representations, 2021. 6, 7 +[33] Chen Min, Dawei Zhao, Liang Xiao, Yiming Nie, and Bin Dai. Voxel-mae: Masked autoencoders for pre-training large-scale point clouds. arXiv preprint arXiv:2206.09900, 2022. 2 +[34] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In European conference on computer vision, pages 604-621. Springer, 2022. 2, 6, 7 +[35] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 3, 5, 6, 7 +[36] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 2, 6, 7 +[37] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. Advances in Neural Information Processing Systems, 35:23192-23204, 2022. 6, 7 +[38] Shi Qiu, Saeed Anwar, and Nick Barnes. Geometric backprojection network for point cloud classification. IEEE Transactions on Multimedia, 24:1943-1955, 2021. 6, 7 +[39] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 1, 2 + +[40] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 1 +[41] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2, 3, 4, 5, 6, 1 +[42] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2 +[43] Yongming Rao, Jiwen Lu, and Jie Zhou. Global-local bidirectional reasoning for unsupervised representation learning of 3d point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5376-5385, 2020. 6 +[44] Justyna Sarzynska-Wawer, Aleksander Wawer, Aleksandra Pawlak, Julia Szymanowska, Izabela Stefaniak, Michal Jarkiewicz, and Lukasz Okruszek. Detecting formal thought disorder by deep contextualized word representations. *Psychiatry Research*, 304:114135, 2021. 2 +[45] Jonathan Sauder and Bjarne Sievers. Self-supervised deep learning on point clouds by reconstructing space. Advances in Neural Information Processing Systems, 32, 2019. 2 +[46] Charu Sharma and Manohar Kaul. Self-supervised few-shot learning on point clouds. Advances in Neural Information Processing Systems, 33:7212-7221, 2020. 6 +[47] Chen Sun, Abhinav Shrivastava, Saurabh Singh, and Abhinav Gupta. Revisiting unreasonable effectiveness of data in deep learning era. In Proceedings of the IEEE international conference on computer vision, pages 843-852, 2017. 6 +[48] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6411–6420, 2019. 6 +[49] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 1, 2 +[50] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1588–1597, 2019. 6, 7, 2 +[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4, 5, 6, 7, 2 +[52] Hanchen Wang, Qi Liu, Xiangyu Yue, Joan Lasenby, and Matt J Kusner. Unsupervised point cloud pre-training via occlusion completion. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9782-9792, 2021. 7 + +[53] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (tog), 38(5):1-12, 2019. 3, 4, 5, 6, 7, 1 +[54] Ziyi Wang, Xumin Yu, Yongming Rao, Jie Zhou, and Jiwen Lu. P2p: Tuning pre-trained image models for point cloud analysis with point-to-pixel prompting. Advances in neural information processing systems, 35:14388-14402, 2022. 6, 7 +[55] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1912-1920, 2015. 6, 2 +[56] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 574-591. Springer, 2020. 7 +[57] Zhenda Xie, Yutong Lin, Zhuliang Yao, Zheng Zhang, Qi Dai, Yue Cao, and Han Hu. Self-supervised learning with swim transformers. arXiv preprint arXiv:2105.04553, 2021. 1 +[58] Yifan Xu, Tianqi Fan, Mingye Xu, Long Zeng, and Yu Qiao. Spidercnn: Deep learning on point sets with parameterized convolutional filters. In Proceedings of the European conference on computer vision (ECCV), pages 87-102, 2018. 7 +[59] Guandao Yang, Xun Huang, Zekun Hao, Ming-Yu Liu, Serge Belongie, and Bharath Hariharan. Pointflow: 3d point cloud generation with continuous normalizing flows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4541-4550, 2019. 7 +[60] Yaoqing Yang, Chen Feng, Yiru Shen, and Dong Tian. Foldingnet: Point cloud auto-encoder via deep grid deformation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 206–215, 2018. 4, 5, 1 +[61] Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Russ R Salakhutdinov, and Quoc V Le. Xlnet: Generalized autoregressive pretraining for language understanding. Advances in neural information processing systems, 32, 2019. 1 +[62] Li Yi, Vladimir G Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas Guibas. A scalable active framework for region annotation in 3d shape collections. ACM Transactions on Graphics (ToG), 35(6):1-12, 2016. 7 +[63] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19313-19322, 2022. 1, 2, 3, 4, 5, 6, 7 +[64] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. Point-m2ae: multi-scale masked autoencoders for hierarchical point cloud pre-training. Advances in neural information processing systems, 35:27061-27074, 2022. 2, 7 + +[65] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16259-16268, 2021. 1, 6 +[66] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In International Conference on Learning Representations, 2020. 1 \ No newline at end of file diff --git a/2024/General Point Model Pretraining with Autoencoding and Autoregressive/images.zip b/2024/General Point Model Pretraining with Autoencoding and Autoregressive/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6de95a804f13ec9a15c3588fe827a0860734019d --- /dev/null +++ b/2024/General Point Model Pretraining with Autoencoding and Autoregressive/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:83850ed3c01c3895dfddab955e00e9f55a101606601673af23553d4b70440fd3 +size 468676 diff --git a/2024/General Point Model Pretraining with Autoencoding and Autoregressive/layout.json b/2024/General Point Model Pretraining with Autoencoding and Autoregressive/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..38a5c4735591fc47ddc50cf85ad9037ad1fa0cc0 --- /dev/null +++ b/2024/General Point Model Pretraining with Autoencoding and Autoregressive/layout.json @@ -0,0 +1,8471 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 74, + 103, + 520, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 103, + 520, + 121 + ], + "spans": [ + { + "bbox": [ + 74, + 103, + 520, + 121 + ], + "type": "text", + "content": "General Point Model Pretraining with Autoencoding and Autoregressive" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 141, + 542, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 141, + 542, + 185 + ], + "spans": [ + { + "bbox": [ + 63, + 141, + 542, + 185 + ], + "type": "text", + "content": "Zhe Li " + }, + { + "bbox": [ + 63, + 141, + 542, + 185 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + }, + { + "bbox": [ + 63, + 141, + 542, + 185 + ], + "type": "text", + "content": ", Zhangyang Gao " + }, + { + "bbox": [ + 63, + 141, + 542, + 185 + ], + "type": "inline_equation", + "content": "^{3,\\dagger}" + }, + { + "bbox": [ + 63, + 141, + 542, + 185 + ], + "type": "text", + "content": ", Cheng Tan " + }, + { + "bbox": [ + 63, + 141, + 542, + 185 + ], + "type": "inline_equation", + "content": "^{3,\\dagger}" + }, + { + "bbox": [ + 63, + 141, + 542, + 185 + ], + "type": "text", + "content": ", Bocheng Ren " + }, + { + "bbox": [ + 63, + 141, + 542, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 63, + 141, + 542, + 185 + ], + "type": "text", + "content": ", Laurence T. Yang " + }, + { + "bbox": [ + 63, + 141, + 542, + 185 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + }, + { + "bbox": [ + 63, + 141, + 542, + 185 + ], + "type": "text", + "content": " Stan Z. Li " + }, + { + "bbox": [ + 63, + 141, + 542, + 185 + ], + "type": "inline_equation", + "content": "^{3,*}" + }, + { + "bbox": [ + 63, + 141, + 542, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 63, + 141, + 542, + 185 + ], + "type": "text", + "content": " Huazhong University of Science and Technology \n" + }, + { + "bbox": [ + 63, + 141, + 542, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 63, + 141, + 542, + 185 + ], + "type": "text", + "content": " Zhengzhou University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 121, + 185, + 485, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 185, + 485, + 200 + ], + "spans": [ + { + "bbox": [ + 121, + 185, + 485, + 200 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 121, + 185, + 485, + 200 + ], + "type": "text", + "content": " AI Lab, Research Center for Industries of the Future, Westlake University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 201, + 550, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 201, + 550, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 201, + 550, + 213 + ], + "type": "text", + "content": "keycharon0122@gmail.com, {gaozhangyang,tancheng,stan.zq.li}@westlake.edu.cn, ltyang@ieee.org" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 241, + 192, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 241, + 192, + 253 + ], + "spans": [ + { + "bbox": [ + 143, + 241, + 192, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 266, + 290, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 266, + 290, + 553 + ], + "spans": [ + { + "bbox": [ + 47, + 266, + 290, + 553 + ], + "type": "text", + "content": "The pre-training architectures of large language models encompass various types, including autoencoding models, autoregressive models, and encoder-decoder models. We posit that any modality can potentially benefit from a large language model, as long as it undergoes vector quantization to become discrete tokens. Inspired by the General Language Model, we propose a General Point Model (GPM) that seamlessly integrates autoencoding and autoregressive tasks in a point cloud transformer. This model is versatile, allowing fine-tuning for downstream point cloud representation tasks, as well as unconditional and conditional generation tasks. GPM enhances masked prediction in autoencoding through various forms of mask padding tasks, leading to improved performance in point cloud understanding. Additionally, GPM demonstrates highly competitive results in unconditional point cloud generation tasks, even exhibiting the potential for conditional generation tasks by modifying the input's conditional information. Compared to models like Point-BERT, MaskPoint, and PointMAE, our GPM achieves superior performance in point cloud understanding tasks. Furthermore, the integration of autoregressive and autoencoding within the same transformer underscores its versatility across different downstream tasks. Codes are available at https://github.com/gentlefress/GPM" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 574, + 128, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 128, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 128, + 586 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": "In recent years, the natural language processing (NLP) [4, 19, 20, 29, 40] and computer vision (CV) [3, 7, 9, 49, 57, 66] realms have witnessed a proliferation of transformer-based pretrained models. Their primary advantage lies in the inclusion of massive parameters and data in the training process, overcoming inductive biases introduced by traditional Convolutional Neural Networks [21]. Point clouds serve as fundamental data structures in fields like autonomous driving and robotics, thus emphasizing the escalating significance of tasks related to point cloud representation learn" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 310, + 240, + 542, + 348 + ], + "blocks": [ + { + "bbox": [ + 310, + 240, + 542, + 348 + ], + "lines": [ + { + "bbox": [ + 310, + 240, + 542, + 348 + ], + "spans": [ + { + "bbox": [ + 310, + 240, + 542, + 348 + ], + "type": "image", + "image_path": "708b3f6f3a8b454058fec970b4754baa20f6e95bca55be7e90f4987b1946cbe3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 355, + 546, + 411 + ], + "lines": [ + { + "bbox": [ + 305, + 355, + 546, + 411 + ], + "spans": [ + { + "bbox": [ + 305, + 355, + 546, + 411 + ], + "type": "text", + "content": "Figure 1. Comparison of pre-training frameworks between GPM and Point-BERT [63]. Point-BERT (left) conducts the masked prediction task in an autoencoding manner, while GPM (right) combines both autoencoding and autoregressive tasks, performing tasks for both masked prediction and masked region blank filling." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 438, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 547, + 715 + ], + "type": "text", + "content": "ing and generation. However, the realm of pretrained point cloud models based on transformers remains relatively scarce. Existing transformer-based point cloud models [13, 65] encounter inevitable inductive biases due to local feature aggregation [13], neighbor embedding [65], and the scarcity of annotated data [9]. These biases deviate from the mainstream of standard Transformers. Self-supervised models [12, 15, 20, 25, 26, 39, 40, 61] have emerged as the dominant methodology. They excel in learning high-quality representations across modalities without extensive labeled data, models like GPT [39] even dominate in text generation. In this end, there is a compelling need to design a transformer-based point cloud model that minimizes inductive biases, learns superior point cloud representations from limited data, and simultaneously possesses the capacities for conditional and unconditional point cloud generation. The transformer-based BERT [20] has marked a significant milestone in the field of natural language processing (NLP). Its pre-training phase, involving masked prediction tasks, grants the model the capacity to learn language representations, showcasing the prowess of autoencoding language models. Given BERT's proficiency in learning language representations, the question arises: can we extend BERT's" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20954" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "type": "text", + "content": "capabilities to the realm of point clouds? However, since point clouds lack a direct analog to the concept of 'words' in language, constructing a discrete vocabulary for point clouds, as done in language models, is infeasible. Point clouds are composed of individual points, and considering each point as a token would lead to a quadratic increase in computational cost as the number of tokens grows. Furthermore, each point in a point cloud contains limited semantic information, necessitating the maximization of geometric information utilization. Inspired by Point-BERT [63], we embarked on a journey to vector quantization, transforming point clouds into discrete tokens. Through the application of the farthest point sampling (FPS) [36] algorithm, we segment the point cloud into multiple patches, each patch encapsulates plentiful geometric information, akin to a unit. By constructing a vocabulary specific to the point cloud domain, we obtain labels for each unit, discretizing the entire point cloud. This paves the way for the completion of the autoencoding model's masked prediction tasks in the point cloud domain." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 313, + 289, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 313, + 289, + 445 + ], + "spans": [ + { + "bbox": [ + 47, + 313, + 289, + 445 + ], + "type": "text", + "content": "Transformer-based GPT [39] has dominated the field of text generation, with its core being an autoregressive language model, predicting the next token based on preceding tokens. In recent years, researchers have progressively extended this paradigm into the realm of images [7, 41, 42], ushering in a new era of image generation. In this endeavor, we bring this paradigm to the domain of point clouds, aiming to empower it for autoregressive conditional and unconditional point cloud generation tasks. After discretization, each patch of the point cloud possesses distinct geometric information, enabling it to excel in autoregressive tasks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 446, + 289, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 446, + 289, + 508 + ], + "spans": [ + { + "bbox": [ + 47, + 446, + 289, + 508 + ], + "type": "text", + "content": "While these pre-training frameworks can be adapted to the point cloud domain, they lack the flexibility to satisfy all point cloud tasks. Therefore, we aim to integrate these two pre-training frameworks from the NLP domain and apply them to the point cloud domain, depicted in Figure 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 509, + 289, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 509, + 289, + 687 + ], + "spans": [ + { + "bbox": [ + 47, + 509, + 289, + 687 + ], + "type": "text", + "content": "In this work, we introduce a novel approach termed General Point Model (GPM) with autoencoding and autoregressive, which integrates point cloud representation learning and point cloud generation (both conditional and unconditional) within the same transformer framework. Similar to Point-BERT, we employ a point cloud tokenizer designed through dVAE-based point cloud reconstruction, where the point cloud can be discretely labeled based on the learned vocabulary. Additionally, we utilize the Masked Point Modeling (MPM) task. This involves partitioning the point cloud into distinct patches and randomly masking some adjacent patches. This aims to predict the masked portion, with the goal of learning geometric information between neighboring blocks and capturing meaningful geometric features for understanding the point cloud." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": "Our primary contributions can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 72, + 547, + 216 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 306, + 72, + 547, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 547, + 108 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 547, + 108 + ], + "type": "text", + "content": "- We introduce autoencoding + autoregressive as a novel point cloud transformer paradigm, unifying tasks in point cloud understanding and generation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 108, + 547, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 108, + 547, + 156 + ], + "spans": [ + { + "bbox": [ + 306, + 108, + 547, + 156 + ], + "type": "text", + "content": "- We partition the input into two segments, the first part predicts masked points (autoencoding), the second part generates masked points (autoregressive), enhancing comprehension and generation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 156, + 547, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 156, + 547, + 216 + ], + "spans": [ + { + "bbox": [ + 306, + 156, + 547, + 216 + ], + "type": "text", + "content": "- Our approach demonstrates notable competitiveness across various point cloud understanding tasks and unconditional generation tasks. Moreover, it exhibits the potential for conditional generation, such as text- or image-conditioned point clouds generation." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 228, + 392, + 240 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 228, + 392, + 240 + ], + "spans": [ + { + "bbox": [ + 306, + 228, + 392, + 240 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 247, + 446, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 247, + 446, + 261 + ], + "spans": [ + { + "bbox": [ + 306, + 247, + 446, + 261 + ], + "type": "text", + "content": "2.1. Self-supervised Learning" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 266, + 548, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 266, + 548, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 266, + 548, + 651 + ], + "type": "text", + "content": "Self-supervised learning (SSL) has garnered significant attention owing to the substantial labor required for acquiring annotated data in large quantities. SSL enables models to learn feature representations from unlabeled data, adapting to downstream tasks. The core of SSL lies in the design of proxy tasks to replace traditional classification tasks, thereby learning feature representations. ELMo [44] employs bidirectional LSTMs [16] and generates subsequent words from left to right given the representation of preceding content. ViP [25] employs a dynamically updated momentum encoder for contrastive learning and designs a text swapping task to enhance the model's sentence representation capability. In the computer vision domain, Image-GPT [7] trains a sequence transformer to autoregressively predict pixels, showing promising representation learning capabilities without incorporating specific knowledge about the 2D input structure after pretraining. Furthermore, SSL in the field of point clouds has garnered significant attention. ACT [8] employs cross-modal autoencoders as teacher models to acquire knowledge from other modalities. The generative approaches [1, 22, 33, 45, 63, 64] are the most pertinent research to our work. Point-MAE [34] extends MAE by randomly masking point patches and reconstructing the masked regions. Point-M2AE [64] further utilizes a hierarchical transformer architecture and designs corresponding masking strategies. However, mask-based point modeling methods still face the issue of shape leakage, limiting their effective generalization to downstream tasks. In this paper, we leverage autoregressive pretraining on point clouds and address the unique challenges associated with point cloud attributes. Our concise design avoids position information leakage, thereby enhancing generalization capabilities." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 658, + 448, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 448, + 671 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 448, + 671 + ], + "type": "text", + "content": "2.2. General Language Model" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "In the field of natural language processing, GLM [10] integrates both autoregressive and autoencoding methods. This enables it to perform tasks related to sentence representa" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "20955" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 306, + 235 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 306, + 235 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 306, + 235 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 306, + 235 + ], + "type": "image", + "image_path": "0fa6104b24bfd09b60eb3ba44183d81ac3af200dc0b164ea659119ea5aab469a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 243, + 547, + 300 + ], + "lines": [ + { + "bbox": [ + 46, + 243, + 547, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 547, + 300 + ], + "type": "text", + "content": "Figure 2. The framework of our GPM. We divide the entire pipeline into two stages: 1) The input point cloud is divided into several subclouds. Then, we utilize Mini-PointNet [35] to obtain a sequence of point embeddings. These embeddings are transformed into a sequence of discrete point tokens using a pre-trained tokenizer; 2) During pre-training, for PartA, certain portions of the masked embeddings are masked and replaced with [M]. They are predicted in an autoencoding manner. For PartB, masks for PartA are generated in an autoregressive manner." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 310, + 72, + 545, + 233 + ], + "blocks": [ + { + "bbox": [ + 310, + 72, + 545, + 233 + ], + "lines": [ + { + "bbox": [ + 310, + 72, + 545, + 233 + ], + "spans": [ + { + "bbox": [ + 310, + 72, + 545, + 233 + ], + "type": "image", + "image_path": "811f0447454e5f41f7d95a3bd47d25a4b52e82cd4fe638ddd16f4d29061de711.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 319, + 288, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 319, + 288, + 416 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 288, + 416 + ], + "type": "text", + "content": "tion as well as conditioned and unconditioned generation through fine-tuning. To the best of our knowledge, we are the first to integrate autoregressive and autoencoding techniques on a single transformer in the domain of point clouds. In downstream tasks, we have the potential to fine-tune point cloud representations and perform unconditioned generation even in zero-shot scenarios, with the possibility of fine-tuning towards conditioned generation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 427, + 107, + 439 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 427, + 107, + 439 + ], + "spans": [ + { + "bbox": [ + 47, + 427, + 107, + 439 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 447, + 287, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 447, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 46, + 447, + 287, + 591 + ], + "type": "text", + "content": "In this study, we endeavor to seamlessly integrate BERT-style and GPT-style pre-trained strategies, extending their applicability to point cloud transformers. Following the Point-BERT[63], our training process unfolds in two distinctive stages. Firstly, we embark on training a specialized Tokenizer, a critical step towards acquiring discrete point labels for each input point cloud. Concurrently, a dedicated discrete VAE (dVAE) is employed to masterfully reconstruct the discrete point cloud. The subsequent stage involves the splicing of two sequences of point cloud embeddings, partA for autoencoding and partB for autoregressive. The overall idea of our approach is illustrated in Figure 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 598, + 287, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 598, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 598, + 287, + 624 + ], + "type": "text", + "content": "3.1. Stage 1: Discrete Varitional Autoencoder Pretraining" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "Point Cloud Partitioning. We posit that discrete tokens derived from points encapsulate crucial geometric information, allowing for a discrete representation of any point cloud. However, a naive strategy of assigning one token to each point in a point-wise reconstruction task poses a formidable computational challenge. This arises from the quadratic complexity of self-attention within transformers." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 319, + 547, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 547, + 488 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 547, + 488 + ], + "type": "text", + "content": "Building upon the foundations laid by Point-BERT [63] and ViT[9], we adopt a strategy of partitioning each point cloud into distinct patches, each serving as a single token. Specifically, for an input point cloud " + }, + { + "bbox": [ + 304, + 319, + 547, + 488 + ], + "type": "inline_equation", + "content": "p \\in \\mathbb{R}^{\\mathcal{N} \\times 3}" + }, + { + "bbox": [ + 304, + 319, + 547, + 488 + ], + "type": "text", + "content": ", we initiate the process by extracting " + }, + { + "bbox": [ + 304, + 319, + 547, + 488 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 319, + 547, + 488 + ], + "type": "text", + "content": " center points " + }, + { + "bbox": [ + 304, + 319, + 547, + 488 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_u \\in \\mathbb{R}^{m \\times 3}" + }, + { + "bbox": [ + 304, + 319, + 547, + 488 + ], + "type": "text", + "content": " from the point cloud " + }, + { + "bbox": [ + 304, + 319, + 547, + 488 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 319, + 547, + 488 + ], + "type": "text", + "content": " through farthest point sampling (FPS). Subsequently, we employ the k-nearest neighbor (kNN) algorithm to identify the " + }, + { + "bbox": [ + 304, + 319, + 547, + 488 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 319, + 547, + 488 + ], + "type": "text", + "content": " nearest neighboring points for every center point. This forms " + }, + { + "bbox": [ + 304, + 319, + 547, + 488 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 319, + 547, + 488 + ], + "type": "text", + "content": " localized patches or sub-clouds denoted as " + }, + { + "bbox": [ + 304, + 319, + 547, + 488 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_u \\in \\mathbb{R}^{m \\times k \\times 3}" + }, + { + "bbox": [ + 304, + 319, + 547, + 488 + ], + "type": "text", + "content": ". To ensure these patches are free of bias, we normalize them by subtracting their respective center coordinates. This operation effectively disentangles the structural patterns from the spatial coordinates inherent to each local patch." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 504, + 547, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 504, + 547, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 504, + 547, + 662 + ], + "type": "text", + "content": "Point Cloud Embedding. Following Point-BERT [63], we employ a mini-PointNet [35] to embed the point patches as a sequence of point embedding " + }, + { + "bbox": [ + 304, + 504, + 547, + 662 + ], + "type": "inline_equation", + "content": "\\{h_i\\}_{i=1}^m" + }, + { + "bbox": [ + 304, + 504, + 547, + 662 + ], + "type": "text", + "content": ". For the further vector quantization, we pre-define a learnable codebook " + }, + { + "bbox": [ + 304, + 504, + 547, + 662 + ], + "type": "inline_equation", + "content": "\\mathcal{V} := \\{(s, z_s)\\}_{s \\in S}^{d_z}" + }, + { + "bbox": [ + 304, + 504, + 547, + 662 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 504, + 547, + 662 + ], + "type": "inline_equation", + "content": "d_z" + }, + { + "bbox": [ + 304, + 504, + 547, + 662 + ], + "type": "text", + "content": " is the dimension of codes, " + }, + { + "bbox": [ + 304, + 504, + 547, + 662 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 504, + 547, + 662 + ], + "type": "text", + "content": " is the size of codebook and " + }, + { + "bbox": [ + 304, + 504, + 547, + 662 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 504, + 547, + 662 + ], + "type": "text", + "content": " is the index of embedding in " + }, + { + "bbox": [ + 304, + 504, + 547, + 662 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 304, + 504, + 547, + 662 + ], + "type": "text", + "content": ". We adopt DGCNN [53] as tokenizer " + }, + { + "bbox": [ + 304, + 504, + 547, + 662 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}: h_i \\mapsto z_i" + }, + { + "bbox": [ + 304, + 504, + 547, + 662 + ], + "type": "text", + "content": ", which maps " + }, + { + "bbox": [ + 304, + 504, + 547, + 662 + ], + "type": "inline_equation", + "content": "\\{h_i\\}_{i=1}^m" + }, + { + "bbox": [ + 304, + 504, + 547, + 662 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 304, + 504, + 547, + 662 + ], + "type": "inline_equation", + "content": "\\{z_i\\}_{i=1}^m" + }, + { + "bbox": [ + 304, + 504, + 547, + 662 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 304, + 504, + 547, + 662 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 304, + 504, + 547, + 662 + ], + "type": "text", + "content": ". However, given the non-differentiable nature of the discrete tokens, applying reparameterization gradients for dVAE training becomes unfeasible. As suggested in [41], we resort to the Gumbel-softmax relaxation technique [18], coupled with a uniform prior, as a workaround during the dVAE training process." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "Point Cloud Reconstruction. Given " + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\{z_i\\}_{i=1}^m" + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": " as input, the decoder " + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(\\cdot)" + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": " is tasked with the reconstruction of the entire point cloud. Inspired by [63], to effectively capture inter" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20956" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "point relationships and bolster the representational capacity of discrete point tokens across various local structures, we adopt the DGCNN [53]-FoldingNet [60] architecture to reconstruct the whole point cloud." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 120, + 287, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 120, + 287, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 120, + 287, + 180 + ], + "type": "text", + "content": "The overarching reconstruction objective can be denoted as " + }, + { + "bbox": [ + 46, + 120, + 287, + 180 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{z\\sim \\mathcal{Q}(z|h)}\\log D(\\hat{p} |z)" + }, + { + "bbox": [ + 46, + 120, + 287, + 180 + ], + "type": "text", + "content": ". Inspired by Point-BERT [63], the entire reconstruction process can be conceptualized as the maximization of the Evidence Lower Bound (ELB) for the log-likelihood, denoted as " + }, + { + "bbox": [ + 46, + 120, + 287, + 180 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(\\hat{p} |p)" + }, + { + "bbox": [ + 46, + 120, + 287, + 180 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 186, + 293, + 236 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 186, + 293, + 236 + ], + "spans": [ + { + "bbox": [ + 47, + 186, + 293, + 236 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\sum_ {i = 1} ^ {\\mathcal {N}} \\log D (\\hat {p} _ {i} | p _ {i}) \\geq \\sum_ {i = 1} ^ {\\mathcal {N}} \\left(\\mathbb {E} _ {z _ {i}} \\sim \\mathcal {Q} (z _ {i} | p _ {i})\\right) [ \\log D (p _ {i} | z _ {i}) ] \\\\ - \\operatorname {K L} \\left[ \\mathcal {Q} \\left(z \\mid p _ {i}\\right), \\mathcal {D} \\left(z \\mid \\hat {p} _ {i}\\right) \\right]. \\tag {1} \\\\ \\end{array}", + "image_path": "d7b8d641442671435a67966b06315622b3888ccfc770a7de637e362217c2275c.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 245, + 287, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 245, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 245, + 287, + 281 + ], + "type": "text", + "content": "Simultaneously, we account for an intuitive reconstruction loss, leveraging the " + }, + { + "bbox": [ + 46, + 245, + 287, + 281 + ], + "type": "inline_equation", + "content": "l1" + }, + { + "bbox": [ + 46, + 245, + 287, + 281 + ], + "type": "text", + "content": " Chamfer Distance [11] to supervise prediction point cloud with the ground-truth point cloud:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 288, + 287, + 327 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 288, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 47, + 288, + 287, + 327 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {C D}} (\\mathcal {P}, \\mathcal {G}) = \\frac {1}{| \\mathcal {P} |} \\sum_ {p \\in \\mathcal {P}} \\min _ {g \\in \\mathcal {G}} \\| p - g \\| + \\frac {1}{| \\mathcal {G} |} \\sum_ {g \\in \\mathcal {G}} \\min _ {p \\in \\mathcal {P}} \\| g - p \\|, \\tag {2}", + "image_path": "74c605859c2f05256babb7f438292b4f4328c88c67557fad0da5ad32b15d491d.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 328, + 287, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 328, + 287, + 375 + ], + "spans": [ + { + "bbox": [ + 46, + 328, + 287, + 375 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 328, + 287, + 375 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 328, + 287, + 375 + ], + "type": "text", + "content": " denotes the predicted point clouds and " + }, + { + "bbox": [ + 46, + 328, + 287, + 375 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 46, + 328, + 287, + 375 + ], + "type": "text", + "content": " is the ground truth point clouds. Additional, we follow [41], optimize the KL-divergence " + }, + { + "bbox": [ + 46, + 328, + 287, + 375 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{KL}}" + }, + { + "bbox": [ + 46, + 328, + 287, + 375 + ], + "type": "text", + "content": " between the generated point cloud distribution and a uniform distribution:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 122, + 385, + 287, + 397 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 385, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 122, + 385, + 287, + 397 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {d V A E}} = \\mathcal {L} _ {\\mathrm {C D}} + \\mathcal {L} _ {\\mathrm {K L}}. \\tag {3}", + "image_path": "1d8905a90a94d9a1829d9009daa3174682472f97748ce223698a9c1e70b6e810.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 405, + 197, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 405, + 197, + 418 + ], + "spans": [ + { + "bbox": [ + 47, + 405, + 197, + 418 + ], + "type": "text", + "content": "3.2. Stage 2: GPM Pre-training" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 423, + 287, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 423, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 287, + 472 + ], + "type": "text", + "content": "To ensure GPM excel in both point cloud representation learning and point cloud generation task, we formulate a training framework that amalgamates autoencoding and autoregressive techniques motivated by [10]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 486, + 287, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 630 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 630 + ], + "type": "text", + "content": "Autoencoding Masked Sequence Generation. We employ the standard Transformers [51], which encompass multi-headed self-attention layers and FFN blocks. Initially, we partition each input point cloud into " + }, + { + "bbox": [ + 46, + 486, + 287, + 630 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 486, + 287, + 630 + ], + "type": "text", + "content": " local patches with designated center points. Subsequently, these local patches undergo discretization into a code sequence " + }, + { + "bbox": [ + 46, + 486, + 287, + 630 + ], + "type": "inline_equation", + "content": "\\{f_i\\}_{i=1}^m" + }, + { + "bbox": [ + 46, + 486, + 287, + 630 + ], + "type": "text", + "content": " via our pre-trained dVAE in stage 1. Formally, the input embeddings " + }, + { + "bbox": [ + 46, + 486, + 287, + 630 + ], + "type": "inline_equation", + "content": "\\{e_i\\}_{i=1}^m" + }, + { + "bbox": [ + 46, + 486, + 287, + 630 + ], + "type": "text", + "content": " are constructed as a combination of point embeddings " + }, + { + "bbox": [ + 46, + 486, + 287, + 630 + ], + "type": "inline_equation", + "content": "\\{f_i\\}_{i=1}^m" + }, + { + "bbox": [ + 46, + 486, + 287, + 630 + ], + "type": "text", + "content": " and positional embeddings " + }, + { + "bbox": [ + 46, + 486, + 287, + 630 + ], + "type": "inline_equation", + "content": "\\{pos_i\\}_{i=1}^m" + }, + { + "bbox": [ + 46, + 486, + 287, + 630 + ], + "type": "text", + "content": ". Adhering to language model [20], a class token [CLS] is appended to the input sequences, rendering the transformer input sequence as " + }, + { + "bbox": [ + 46, + 486, + 287, + 630 + ], + "type": "inline_equation", + "content": "I = \\{[\\mathrm{CLS}], e_1^0, e_2^0, \\dots, e_i^0\\}" + }, + { + "bbox": [ + 46, + 486, + 287, + 630 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": "Drawing inspiration from [63], we initiate by selecting a central point " + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_i" + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": " alongside its corresponding sub-cloud " + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": ", integrate it with " + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": " neighboring sub-clouds as a coherent local region. Within this region, we apply a masking operation to obscure all local patches, generating what we refer to as the \"masked point cloud\". Specifically, we substitute all the masked point embeddings with a universally learnable" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "type": "text", + "content": "pre-defined mask embeddings [M]. We mark the masked position as " + }, + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = \\{1,\\dots,b\\}" + }, + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "type": "text", + "content": ", and the final input embeddings " + }, + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "type": "inline_equation", + "content": "E^{\\mathcal{M}} = \\{e_i,i\\notin \\mathcal{M}\\} \\cup \\{[\\mathbf{M}] + pos_i,i\\in \\mathcal{M}\\}" + }, + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "type": "text", + "content": " are fed into the transformer." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 120, + 546, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 120, + 546, + 263 + ], + "spans": [ + { + "bbox": [ + 304, + 120, + 546, + 263 + ], + "type": "text", + "content": "The objective of our Masked Point Modeling (MPM) task is to deduce the geometric arrangement of absent portions and reconstruct point tokens aligned with the masked positions using the available information. Geometric information refers to the relative positions and morphological attributes of points within a localized region. These crucial geometric details provide vital cues for understanding and processing the shape and structure of the local region, playing a pivotal role in delving deep into the intricacies of point cloud data. Formally, the pre-training objective can be expressed as maximizing the log-likelihood of the point tokens " + }, + { + "bbox": [ + 304, + 120, + 546, + 263 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 304, + 120, + 546, + 263 + ], + "type": "text", + "content": " conditioned on the masked input embeddings " + }, + { + "bbox": [ + 304, + 120, + 546, + 263 + ], + "type": "inline_equation", + "content": "E^{\\mathcal{M}}" + }, + { + "bbox": [ + 304, + 120, + 546, + 263 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 353, + 271, + 545, + 301 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 271, + 545, + 301 + ], + "spans": [ + { + "bbox": [ + 353, + 271, + 545, + 301 + ], + "type": "interline_equation", + "content": "\\max \\sum_ {E \\in W} \\mathbb {E} \\left[ \\sum_ {i \\in \\mathcal {M}} \\log P \\left(e _ {i} \\mid E ^ {\\mathcal {M}}\\right) \\right], \\tag {4}", + "image_path": "a31dfe6e739d97b87902740125b1a0973a8c8e71e39e64e5f9784fadd5490ea7.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 308, + 545, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 308, + 545, + 428 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 545, + 428 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 308, + 545, + 428 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 308, + 545, + 428 + ], + "type": "text", + "content": " is the set of all input embeddings. Similar to [63], we also adopt Point Patch Mixing and contrastive learning to help the model to better learn high-level semantics. Point Patch Mixing involves mixing sub-clouds without complex alignment techniques. However, training solely with masked precition has limitations in understanding high-level semantics. To overcome this, we incorporate contrastive learning to enhance semantic understanding. Assuming the masking ratio is " + }, + { + "bbox": [ + 304, + 308, + 545, + 428 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 308, + 545, + 428 + ], + "type": "text", + "content": ", the loss function is defined as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 310, + 434, + 545, + 465 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 434, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 310, + 434, + 545, + 465 + ], + "type": "interline_equation", + "content": "- \\beta \\log \\frac {\\exp \\left(q \\hat {k} _ {1} / \\tau\\right)}{\\sum_ {i = 0} ^ {K} \\exp \\left(q k _ {i}\\right)} - (1 - \\beta) \\log \\frac {\\exp \\left(q \\hat {k} _ {2} / \\tau\\right)}{\\sum_ {i = 0} ^ {K} \\exp \\left(q k _ {i}\\right)}, \\tag {5}", + "image_path": "e015c2e08fe063b4ad28828c57bf928600b0a6ef1be8dbf12e537b35818ebee3.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 472, + 545, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 545, + 520 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 472, + 545, + 520 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 304, + 472, + 545, + 520 + ], + "type": "text", + "content": " is the feature of a mixed sample that comes from two other samples, whose features are " + }, + { + "bbox": [ + 304, + 472, + 545, + 520 + ], + "type": "inline_equation", + "content": "\\hat{k_1}" + }, + { + "bbox": [ + 304, + 472, + 545, + 520 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 472, + 545, + 520 + ], + "type": "inline_equation", + "content": "\\hat{k_2}" + }, + { + "bbox": [ + 304, + 472, + 545, + 520 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 472, + 545, + 520 + ], + "type": "inline_equation", + "content": "k_{i}" + }, + { + "bbox": [ + 304, + 472, + 545, + 520 + ], + "type": "text", + "content": " are extracted by the momentum feature encoder, " + }, + { + "bbox": [ + 304, + 472, + 545, + 520 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 472, + 545, + 520 + ], + "type": "text", + "content": " is the temperature and " + }, + { + "bbox": [ + 304, + 472, + 545, + 520 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 472, + 545, + 520 + ], + "type": "text", + "content": " is the size of memory bank." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 533, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 641 + ], + "type": "text", + "content": "Autoregressive Sequence Generation. While Point-BERT models the relationship between unmasked and masked regions using the Masked Point Modeling (MPM) task, it does not adequately model the interactions within masked regions. Furthermore, our aim extends beyond acquiring high-quality point cloud representations; we also aspire to proficiently execute point cloud generation tasks. Hence, for the latter segment, PartB, we engage in autoregressive mask generation tasks to enhance the MPM task." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "In the case of PartB, the mask operation is ignored. Our focus lies in acquiring the generated tokens from the masked positions of PartA. We take the initial " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "n - 1" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": " tokens from PartA, appending the start token [S] at its forefront to get the PartB " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "E^{\\mathcal{B}} = \\{[\\mathrm{S}], e_1, \\dots, e_{n-1}\\}" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": ". Following the autoregressive approach, the prediction of the next token relies" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20957" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 144, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 144, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 144, + 83 + ], + "type": "text", + "content": "on the preceding token:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 85, + 91, + 287, + 122 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 91, + 287, + 122 + ], + "spans": [ + { + "bbox": [ + 85, + 91, + 287, + 122 + ], + "type": "interline_equation", + "content": "\\max _ {\\theta} \\mathbb {E} \\left[ \\sum_ {i = 1} ^ {n} \\log P _ {\\theta} \\left(e _ {i} | \\mathrm {P a r t A}, e _ {< i}, [ \\mathrm {S} ]\\right) \\right]. \\tag {6}", + "image_path": "dbdae070b6c7b8a1489ffcd47c9f9ca76df87383cf449e8ed233f73d7e90c746.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 127, + 287, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 127, + 287, + 222 + ], + "spans": [ + { + "bbox": [ + 46, + 127, + 287, + 222 + ], + "type": "text", + "content": "Although no masking operation is performed on PartB, since the mask positions of PartA are multiple consecutive sub-clouds instead of a single one, in PartB we are also equivalent to autoregressive generation of a continuous mask span. We extract the generated tokens from PartB, corresponding to the masked positions in PartA, and compare them with the discrete tokens produced by the Tokenizer:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 230, + 287, + 246 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 230, + 287, + 246 + ], + "spans": [ + { + "bbox": [ + 55, + 230, + 287, + 246 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {A R} = \\operatorname {C r o s s E n t r o p y} \\left(e _ {j, j \\in \\mathcal {M}}, \\operatorname {a r g m a x} \\mathcal {Q} (h) _ {j, j \\in \\mathcal {M}}\\right). \\tag {7}", + "image_path": "876b277fc1f4b22ee4c606f04fe5842381224852abf0f02dff4b4c83f2551562.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 249, + 287, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 249, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 47, + 249, + 287, + 285 + ], + "type": "text", + "content": "In this way, it allows our model to acquire proficiency in an autoencoding bidirectional encoding scheme for PartA, and an autoregressive unidirectional encoder for PartB." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 291, + 176, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 291, + 176, + 304 + ], + "spans": [ + { + "bbox": [ + 47, + 291, + 176, + 304 + ], + "type": "text", + "content": "3.3. Multi-task Pretraining" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 309, + 287, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 309, + 287, + 429 + ], + "spans": [ + { + "bbox": [ + 46, + 309, + 287, + 429 + ], + "type": "text", + "content": "The masking of " + }, + { + "bbox": [ + 46, + 309, + 287, + 429 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 46, + 309, + 287, + 429 + ], + "type": "text", + "content": " of tokens in BERT is tailored for downstream natural language understanding tasks, whereas Point-BERT masks " + }, + { + "bbox": [ + 46, + 309, + 287, + 429 + ], + "type": "inline_equation", + "content": "25\\% \\sim 45\\%" + }, + { + "bbox": [ + 46, + 309, + 287, + 429 + ], + "type": "text", + "content": " to acquire enhanced point cloud representations. Our GPM focus on both point cloud representation learning and point cloud generation, we concatenate PartA and PartB as input to the transformer, aiming to simultaneously perform both autoencoding and autoregressive tasks. Therefore, we need to establish specific attention masks to facilitate the interaction of information between these two segments." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 430, + 287, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 430, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 287, + 514 + ], + "type": "text", + "content": "Motivated by [10], the tokens in PartA can participate in MLM task in PartA and autoregressive generation in PartB. PartB tokens can not be observed by PartA, but those antecedents. PartA is treated as the conditioning for autoregressive generation in PartB, akin to prefix tuning [23] in NLP. The implement of GPM multi-task pretraining is depicted in Figure 3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 514, + 287, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 514, + 287, + 572 + ], + "spans": [ + { + "bbox": [ + 46, + 514, + 287, + 572 + ], + "type": "text", + "content": "The tokens output by the pretrained tokenizer " + }, + { + "bbox": [ + 46, + 514, + 287, + 572 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}(\\cdot)" + }, + { + "bbox": [ + 46, + 514, + 287, + 572 + ], + "type": "text", + "content": " are regarded as labels. PartA conducts a masked prediction task in an autoencoding manner, and the prediction loss between the predicted masked portion and the labels is computed as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 580, + 287, + 596 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 580, + 287, + 596 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 287, + 596 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {A E} = \\operatorname {C r o s s E n t r o p y} \\left(e _ {j, j \\in \\mathcal {M}}, \\operatorname {a r g m a x} \\mathcal {Q} (h) _ {j, j \\in \\mathcal {M}}\\right). \\tag {8}", + "image_path": "7de4f65d14e546d30e87aa19aa9eb192c64d7c400294be055d741edcc62c364c.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 600, + 287, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 600, + 287, + 659 + ], + "spans": [ + { + "bbox": [ + 46, + 600, + 287, + 659 + ], + "type": "text", + "content": "Although Equation 7 and 8 exhibit identical formal structures. However, one entails mask prediction through autoencoding, while the other involves mask generation in an autoregressive way, indicating a fundamental distinction between the two approaches." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 670, + 127, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 127, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 127, + 682 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "In this section, we commence by outlining the configurations of our pretraining scheme. Subsequently, we pro" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 354, + 70, + 499, + 223 + ], + "blocks": [ + { + "bbox": [ + 354, + 70, + 499, + 223 + ], + "lines": [ + { + "bbox": [ + 354, + 70, + 499, + 223 + ], + "spans": [ + { + "bbox": [ + 354, + 70, + 499, + 223 + ], + "type": "image", + "image_path": "dce3458aaa424daa59ba179a4a38dbf740e92d73722392201881b9ccfd2d985f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 230, + 545, + 264 + ], + "lines": [ + { + "bbox": [ + 305, + 230, + 545, + 264 + ], + "spans": [ + { + "bbox": [ + 305, + 230, + 545, + 264 + ], + "type": "text", + "content": "Figure 3. Self-attention mask. The regions in grey are masked. Tokens in PartA can attend to themselves, but not to PartB. Tokens in PartB can attend to PartA and their antecedents in PartB." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 286, + 545, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 286, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 304, + 286, + 545, + 358 + ], + "type": "text", + "content": "ceed to assess the proposed model through various downstream tasks, encompassing object classification, partial segmentation, few-shot learning, transfer learning (to validate the model's representation learning capabilities), and point cloud generation (to validate its generative capabilities). Additionally, we conduct ablation studies on GPM." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 366, + 518, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 366, + 518, + 380 + ], + "spans": [ + { + "bbox": [ + 305, + 366, + 518, + 380 + ], + "type": "text", + "content": "4.1. Pre-training Setups and Implementation" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 385, + 545, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 385, + 545, + 505 + ], + "spans": [ + { + "bbox": [ + 304, + 385, + 545, + 505 + ], + "type": "text", + "content": "Pre-training Data. Following the dataset configuration similar to Point-BERT [63], we employ ShapeNet [5] as our pretraining dataset, encompassing over 50,000 unique 3D models spanning 55 common object categories. From each 3D model, we sample 1024 points, dividing them into 64 point patches (sub-clouds) and each sub-cloud contains 32 points. Utilizing a lightweight PointNet [35] with two MLP layers, we project each sub-cloud into a 64-point embedding, serving as input for both dVAE and Transformer [51]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 521, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 640 + ], + "type": "text", + "content": "dVAE Pretraining Setups. The purpose of pretraining the dVAE is to acquire a high-quality tokenizer, enabling the reconstruction of discretized features back into the original point cloud to the fullest extent. Our dVAE is composed of a tokenizer and a decoder. To be specific, the tokenizer consists of a 4-layer DGCNN [53], while the decoder encompasses a 4-layer DGCNN followed by a FoldingNet [60]. Additionally, the FoldingLayer establishes a connection between a 2D grid and the input, ultimately generating a 3D point cloud." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "During the training of dVAE [41], we set the number of words in the codebook " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": " to 8192. We employ the common " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "l1" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "-style Chamfer Distance loss in the reconstruction process. Due to the small numerical value of this " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "l1" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": " loss, the weight of the KL loss in Equation 1 must be smaller than that in image tasks. In the initial 10,000 steps of training," + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20958" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "the weight of the KL loss is kept at 0 and gradually raised to 0.1 in the subsequent 100,000 steps. Our learning rate is set to 0.0005, and follows a cosine annealing schedule spanning 60,000 steps. Consistent with prior works [41] and Point-BERT [63], we decay the temperature in the Gumblesoftmax function from 1 to 0.0625 over 100,000 steps. The training of dVAE encompasses a total of 150,000 steps, with a batch size of 64. It's worth noting that our dVAE's network architecture closely mirrors that of Point-BERT." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 196, + 289, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 196, + 289, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 196, + 289, + 411 + ], + "type": "text", + "content": "GPM Pretraining Setups. In our experiments, we adhere to the standard Transformer architecture [9], which consists of a stack of Transformer blocks [51], each composed of multi-head self-attention layers and feed-forward networks. LayerNorm is applied in both layers. We set the depth of our Transformer to 12, feature dimension to 384, and the number of heads to 6. A random depth of 0.1 [17] is applied in our Transformer encoder. During GPM pretraining, the weights of the Tokenizer learned by dVAE are kept fixed. Input point embeddings from " + }, + { + "bbox": [ + 46, + 196, + 289, + 411 + ], + "type": "inline_equation", + "content": "25\\% \\sim 45\\%" + }, + { + "bbox": [ + 46, + 196, + 289, + 411 + ], + "type": "text", + "content": " are randomly masked, and the model is trained to infer the expected point labels at these masked positions. Unlike Point-BERT [63], which does not have a generation task during pretraining, we incorporate both autoregressive and autoencoding tasks in our work. Thus, an attention mask is necessary, as illustrated in Figure 3. For MoCo, we set the same configuration as [63]. More Details can be found in the supplementary 6.1 and 6.2." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 420, + 160, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 420, + 160, + 431 + ], + "spans": [ + { + "bbox": [ + 47, + 420, + 160, + 431 + ], + "type": "text", + "content": "4.2. Downstream Tasks" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 438, + 248, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 438, + 248, + 451 + ], + "spans": [ + { + "bbox": [ + 47, + 438, + 248, + 451 + ], + "type": "text", + "content": "4.2.1 Point Cloud Representation Evaluation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 458, + 288, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 458, + 288, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 458, + 288, + 651 + ], + "type": "text", + "content": "Object classification on ModelNet40. We evaluate our pre-trained model on the ModelNet40 dataset [55], comprising 12,311 clean 3D CAD models spanning 40 categories. Following the Point-BERT [63] setups, we employ a two-layer MLP with a dropout rate of 0.5 as the classification head for the task. We optimize the model using AdamW with a weight decay of 0.05 and a learning rate of 0.0005, while employing a batch size of 32 and a cosine annealing schedule. We conduct comparisons with various Transformer-based models with identical Point-BERT settings, denoting [ST] for the standard Transformer architecture and [T] for Transformer models with specific designs or inductive biases. The results in Table 1 demonstrate that our model not only outperforms Point-BERT in classification metrics on this dataset but also exhibits autoregressive generation capabilities." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "Object classification on ScanObjectNN. The practical applicability of our model on real datasets serves as a crucial metric. Therefore, the pre-trained models are transferred to the ScanObjectNN dataset [50], which comprises" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 309, + 70, + 547, + 365 + ], + "blocks": [ + { + "bbox": [ + 309, + 70, + 547, + 365 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 547, + 365 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 547, + 365 + ], + "type": "table", + "html": "
Methodsnumber of pointsAcc
Supervised Learning
PointNet[35]1k89.2
SO-Net[22]1k92.5
PointNet++ [36]1k90.5
PointCNN[24]1k92.2
DGCNN[53]1k92.9
MVTN[14]1k93.8
RSCNN [43]1k92.9
GBNet[38]1k93.8
PointMLP[32]1k94.5
DensePoint[28]1k92.8
PointNeXt[37]1k94.0
P2P-RN101[54]1k93.1
P2P-HorNet[54]1k94.0
KPConv[48]~6.8k92.9
Self-Supervised Representation Learning
[T]PCT [65]1k93.2
[T]PointTransformer [13]-93.7
[ST]NPCT [13]1k91.0
[ST]Transformer[51]1k91.4
[ST]Transformer-OcCo[51]1k92.1
[ST]Point-BERT[63]1k93.2
[ST]MaskPoint[27]1k93.8
[ST]Point-MAE[34]1k93.8
[ST]GPM1k94.0
Methods with Cross-modal Information and Teacher Models
ACT[8]1k93.7
[ST]Transformer[51]4k91.2
[ST]Transformer-OcCo[51]4k92.2
[ST]Point-BERT[63]4k93.4
[ST]GPM4k93.1
[ST]Point-BERT[63]8k93.8
[ST]MaskPoint[27]8k-
[ST]Point-MAE[34]8k94.0
[ST]GPM8k94.3
", + "image_path": "0e95fac268c980147db8531cd041fb7385e301819d814a8942c88f50f55d7015.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 373, + 545, + 396 + ], + "lines": [ + { + "bbox": [ + 306, + 373, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 306, + 373, + 545, + 396 + ], + "type": "text", + "content": "Table 1. Classification results on ModelNet40. All results are expressed as percentages accuracy." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 418, + 547, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 547, + 538 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 547, + 538 + ], + "type": "text", + "content": "approximately 15,000 objects extracted from real indoor scans, encompassing 2902 point clouds from 15 categories. This dataset presents a greater challenge as it involves sampling from real-world scans with backgrounds and occlusions. We conducted experiments on three main variants, namely OBJ-BG, OBJ-ONLY, and PB-T50-RS, in line with prior work. The experimental results are summarized in Table 2. It is observed that GPM exhibits an improvement of approximately " + }, + { + "bbox": [ + 304, + 418, + 547, + 538 + ], + "type": "inline_equation", + "content": "2.77\\%" + }, + { + "bbox": [ + 304, + 418, + 547, + 538 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 418, + 547, + 538 + ], + "type": "inline_equation", + "content": "1.88\\%" + }, + { + "bbox": [ + 304, + 418, + 547, + 538 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 418, + 547, + 538 + ], + "type": "inline_equation", + "content": "1.73\\%" + }, + { + "bbox": [ + 304, + 418, + 547, + 538 + ], + "type": "text", + "content": " over the regular Point-BERT across the three variants." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "type": "text", + "content": "Few-shot learning. To demonstrate the ability to acquire knowledge for new tasks with limited training data, we evaluate our model under the setting of few-shot learning, following the methodology of previous work [47, 63]. In the typical 'W-way S-shot' setup, we initially randomly select " + }, + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "type": "text", + "content": " classes and then sample " + }, + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "type": "inline_equation", + "content": "(S + 20)" + }, + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "type": "text", + "content": " objects for each class [46]. The model is trained on " + }, + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "type": "inline_equation", + "content": "W \\times S" + }, + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "type": "text", + "content": " samples (support set) and evaluated on the remaining " + }, + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "type": "inline_equation", + "content": "20W" + }, + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "type": "text", + "content": " samples (query set). We conduct 10 independent experiments for each setting and report the average performance and standard deviation across the 10 runs. As shown in Table 3, our approach outperforms other methods in all tests, achieving absolute improvements of " + }, + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "type": "inline_equation", + "content": "0.6\\%" + }, + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "type": "text", + "content": " over Point-BERT." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20959" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 289, + 264 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 289, + 264 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 289, + 264 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 289, + 264 + ], + "type": "table", + "html": "
MethodsOBJ-BGOBJ-ONLYPB-T50-RS
Supervised Learning
PointNet[35]73.379.268.0
SpiderCNN[58]77.179.573.7
PointNet++ [36]82.384.377.9
PointCNN[24]86.185.578.5
DGCNN[53]82.886.278.1
MVTN[14]92.692.382.8
BGA-DGCNN[50]--79.7
BGA-PN++ [50]--80.2
GBNet[38]--81.0
PointMLP[32]--85.4
PointNeXt[37]--87.7
P2P-RN101[54]--87.4
P2P-HorNet[54]--89.3
Self-Supervised Representation Learning
Transformer[51]79.8680.5577.24
Transformer-OcCo[51]84.8585.5478.79
Point-BERT[63]87.4388.1283.07
MaskPoint[27]89.388.184.3
Point-MAE[34]90.088.285.2
GPM90.290.084.8
", + "image_path": "761df663db1b236daae64593b469b68d8f1acbe88334348343741fbfa9afb3c9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 305, + 287, + 405 + ], + "blocks": [ + { + "bbox": [ + 47, + 273, + 287, + 295 + ], + "lines": [ + { + "bbox": [ + 47, + 273, + 287, + 295 + ], + "spans": [ + { + "bbox": [ + 47, + 273, + 287, + 295 + ], + "type": "text", + "content": "Table 2. Classification results on ScanObjectNN. All results are expressed as percentages accuracy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 305, + 287, + 405 + ], + "lines": [ + { + "bbox": [ + 50, + 305, + 287, + 405 + ], + "spans": [ + { + "bbox": [ + 50, + 305, + 287, + 405 + ], + "type": "table", + "html": "
Methods5-way10-way
10-shot20-shot10-shot20-shot
Supervised Learning
DGCNN-rand [53]31.6 ± 2.840.8 ± 4.619.9 ± 2.116.9 ± 1.5
OcCo [52]90.6 ± 2.892.5 ± 1.982.9 ± 1.386.5 ± 2.2
Self-Supervised Representation Learning
Point-BERT [63]94.6 ± 3.196.3 ± 2.791.0 ± 5.492.7 ± 5.1
MaskPoint [27]95.0 ± 3.797.2 ± 1.791.4 ±4.093.4 ± 3.5
Piont-MAE [34]96.3 ± 2.597.8 ± 1.892.6 ± 4.195.0 ± 3.0
Piont-M2AE [64]96.8 ± 1.898.3 ± 1.492.3 ± 4.595.0 ± 3.0
GPM97.2 ± 2.698.7 ± 2.292.9 ± 4.295.0 ± 3.0
", + "image_path": "8425c7a0e45d0f18be1c684937dff7eae8baabf98e6a1aa0d7a743b4e34bdfd3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 479, + 287, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 287, + 658 + ], + "type": "text", + "content": "Part segmentation. We evaluate the representation learning capability of our approach on the ShapeNetPart dataset [62], aiming to predict more fine-grained class labels for each point. This dataset consists of 16 categories and comprises 16,881 objects. The point cloud is downsampled to 2048 points, and the segmentation head [34] connects features " + }, + { + "bbox": [ + 46, + 479, + 287, + 658 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^4, \\mathcal{F}^8, \\mathcal{F}^{12}" + }, + { + "bbox": [ + 46, + 479, + 287, + 658 + ], + "type": "text", + "content": " extracted from the 4-th, 6-th, and 12-th layers of the transformer blocks. Subsequently, average pooling, max pooling, and upsampling are employed to generate features for each point, followed by label prediction using MLP. In our experiments, we evaluate the performance under the settings of \"5 way 10 shot\", \"5 way 20 shot\", \"10 way 10 shot\", and \"10 way 20 shot\". The experimental results shown in Table 4 demonstrate the superior performance of our GPM compared to all other methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 659, + 231, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 659, + 231, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 659, + 231, + 670 + ], + "type": "text", + "content": "4.2.2 Point Cloud Generation Evaluation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": "In the pre-training phase, we perform unconditional autoregressive generation on tokens in PartB, endowing the model with the capability of point cloud unconditional generation." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 309, + 70, + 547, + 228 + ], + "blocks": [ + { + "bbox": [ + 46, + 413, + 287, + 458 + ], + "lines": [ + { + "bbox": [ + 46, + 413, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 287, + 458 + ], + "type": "text", + "content": "Table 3. The results of few-shot classification on the ModelNet40 dataset. For each experimental setting, we conduct 10 independent experiments and report the average accuracy " + }, + { + "bbox": [ + 46, + 413, + 287, + 458 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 46, + 413, + 287, + 458 + ], + "type": "text", + "content": " along with its standard deviation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 70, + 547, + 228 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 547, + 228 + ], + "type": "table", + "html": "
MethodsCls.mIoUInst.mIoU
Supervised Learning
PointNet[35]80.3983.7
PointNet++ [36]81.3585.1
DGCNN[53]82.3385.2
PointMLP [32]84.686.1
Self-Supervised Representation Learning
Transformer[51]83.4285.1
Transformer-OcCo[51]83.4285.1
PointContrast [56]-85.1
CrossPoint [2]-85.5
Point-BERT [63]84.1185.6
Point-MAE [34]-86.1
GPM84.2085.8
", + "image_path": "92a9fac1ce9381843b43a0eff820e045a35f33da624d9eae13b22fcfafc14f25.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 308, + 281, + 545, + 360 + ], + "blocks": [ + { + "bbox": [ + 305, + 236, + 545, + 269 + ], + "lines": [ + { + "bbox": [ + 305, + 236, + 545, + 269 + ], + "spans": [ + { + "bbox": [ + 305, + 236, + 545, + 269 + ], + "type": "text", + "content": "Table 4. Part segmentation results on the ShapeNetPart dataset. We report the average intersection mIoU over the union of all classes (Cls.) and instances (Inst.)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 281, + 545, + 360 + ], + "lines": [ + { + "bbox": [ + 308, + 281, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 308, + 281, + 545, + 360 + ], + "type": "table", + "html": "
CategoryModelJSD (↓)MMD (↓)Cov (%, ↑)1-NNA % (%, ↓)
CDEMDCDEMDCDEMD
AirplanePoint-Flow4.920.2173.2446.9148.4075.6875.06
Point-BERT4.070.1892.7550.1753.4470.1661.07
GPM3.320.1762.4752.3456.4468.2260.30
ChairPoint-Flow1.742.427.8746.8346.9860.8859.89
Point-BERT1.231.827.1149.7350.0757.3056.61
GPM0.981.676.8251.0852.1454.0954.27
CarPoint-Flow0.870.915.2244.0346.5960.6562.36
Point-BERT0.730.824.9647.2752.1359.0960.48
GPM0.660.754.3849.6754.1758.8858.16
", + "image_path": "9f59f3d923e8796735367fd03869045c8afeb2263e1cb38fd16c5576d8e4a675.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 368, + 545, + 401 + ], + "lines": [ + { + "bbox": [ + 305, + 368, + 545, + 401 + ], + "spans": [ + { + "bbox": [ + 305, + 368, + 545, + 401 + ], + "type": "text", + "content": "Table 5. Generation results on ShapeNet dataset. " + }, + { + "bbox": [ + 305, + 368, + 545, + 401 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 305, + 368, + 545, + 401 + ], + "type": "text", + "content": ": the higher the better, " + }, + { + "bbox": [ + 305, + 368, + 545, + 401 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 305, + 368, + 545, + 401 + ], + "type": "text", + "content": ": the lower the better. MMD-CD is multiplied by " + }, + { + "bbox": [ + 305, + 368, + 545, + 401 + ], + "type": "inline_equation", + "content": "10^{3}" + }, + { + "bbox": [ + 305, + 368, + 545, + 401 + ], + "type": "text", + "content": "; MMD-EMD and JSD are multiplied by " + }, + { + "bbox": [ + 305, + 368, + 545, + 401 + ], + "type": "inline_equation", + "content": "10^{2}" + }, + { + "bbox": [ + 305, + 368, + 545, + 401 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 423, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 545, + 544 + ], + "type": "text", + "content": "Therefore, in downstream tasks, without fine-tuning, we conduct an autoregressive point cloud generation task on the ShapeNet dataset. Existing models do not possess both point cloud generation and point cloud classification tasks. Hence, our model is the first known point cloud transformer that integrates these two tasks into one. The specific metrics are shown in Table 5. We compare our results with generation model Point-Flow [59] and Point-BERT. More details about downstream tasks and conditional point cloud generation results can be found in the supplementary 6.3." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 551, + 400, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 551, + 400, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 551, + 400, + 563 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "To validate the fundamental design of the GPM model, we conducted a comprehensive set of ablation studies. In order to assess the impact of autoregressive mask generation on the autoencoding masked prediction task, we present results comparing fine-tuning scenarios on the ModelNet40 dataset: one with autoencoding only, and the other with a combination of autoencoding + autoregressive. Furthermore, to demonstrate that the order of autoencoding and autoregressive tasks in the sequence does not affect point cloud understanding, we conduct additional experiments, the results are summarized in the Table 6. More ablation results are shown in the supplementary 6.4." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20960" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 54, + 272, + 157 + ], + "blocks": [ + { + "bbox": [ + 50, + 54, + 272, + 157 + ], + "lines": [ + { + "bbox": [ + 50, + 54, + 272, + 157 + ], + "spans": [ + { + "bbox": [ + 50, + 54, + 272, + 157 + ], + "type": "image", + "image_path": "a8f2f5445a6e4bb216661e0a0b6ff10c748c5c72104235068ee35b6edc9d2f14.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 91, + 159, + 230, + 168 + ], + "lines": [ + { + "bbox": [ + 91, + 159, + 230, + 168 + ], + "spans": [ + { + "bbox": [ + 91, + 159, + 230, + 168 + ], + "type": "text", + "content": "(a) Point cloud generation results of dVAE." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 325, + 53, + 541, + 158 + ], + "blocks": [ + { + "bbox": [ + 325, + 53, + 541, + 158 + ], + "lines": [ + { + "bbox": [ + 325, + 53, + 541, + 158 + ], + "spans": [ + { + "bbox": [ + 325, + 53, + 541, + 158 + ], + "type": "image", + "image_path": "9d24243c1e5ac04edb415f0ce15298a8bba6eff659116e964425f2e82847a49a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 353, + 159, + 512, + 168 + ], + "lines": [ + { + "bbox": [ + 353, + 159, + 512, + 168 + ], + "spans": [ + { + "bbox": [ + 353, + 159, + 512, + 168 + ], + "type": "text", + "content": "(b) Point cloud generation results of Point-BERT." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 157, + 178, + 435, + 189 + ], + "lines": [ + { + "bbox": [ + 157, + 178, + 435, + 189 + ], + "spans": [ + { + "bbox": [ + 157, + 178, + 435, + 189 + ], + "type": "text", + "content": "Figure 4. The reconstruction results visualization on dVAE and Point-BERT." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 50, + 209, + 289, + 358 + ], + "blocks": [ + { + "bbox": [ + 50, + 209, + 289, + 358 + ], + "lines": [ + { + "bbox": [ + 50, + 209, + 289, + 358 + ], + "spans": [ + { + "bbox": [ + 50, + 209, + 289, + 358 + ], + "type": "table", + "html": "
TasksnpointsAcc
autoencoding only1k92.9
autoregressive ⇒ autoencoding1k93.6
autoencoding ⇒ autoregressive1k93.8
autoencoding only4k93.4
autoregressive ⇒ autoencoding4k93.9
autoencoding ⇒ autoregressive4k93.9
", + "image_path": "e5f6a8ae075afcdec45b4ecef00e03f635052741a104ca2e7a73ce04d76e8a57.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 366, + 287, + 422 + ], + "lines": [ + { + "bbox": [ + 46, + 366, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 366, + 287, + 422 + ], + "type": "text", + "content": "Table 6. Results on autoencoding, autoencoding " + }, + { + "bbox": [ + 46, + 366, + 287, + 422 + ], + "type": "inline_equation", + "content": "\\Rightarrow" + }, + { + "bbox": [ + 46, + 366, + 287, + 422 + ], + "type": "text", + "content": " autoregressive and autoregressive " + }, + { + "bbox": [ + 46, + 366, + 287, + 422 + ], + "type": "inline_equation", + "content": "\\Rightarrow" + }, + { + "bbox": [ + 46, + 366, + 287, + 422 + ], + "type": "text", + "content": " autoencoding on the ModelNet dataset. It is clear that autoregressive can enhance the completion accuracy of autoencoding, the order of autoencoding and autoregressive does not have a significant impact on point cloud understanding tasks." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 445, + 110, + 457 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 445, + 110, + 457 + ], + "spans": [ + { + "bbox": [ + 47, + 445, + 110, + 457 + ], + "type": "text", + "content": "4.4. Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 464, + 287, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 464, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 46, + 464, + 287, + 644 + ], + "type": "text", + "content": "Compare with Point-BERT. Motivated by BERT, Point-BERT leverages MPM as a pretext task for training. Due to the assumption of independence between MLM and MPM, Point-BERT fails to capture interdependencies among masked tokens. In our approach, autoregressive generation explicitly models dependencies between masked regions, unlike autoencoding which relies solely on un-masked tokens to predict the masked ones. Autoregressive masked prediction allows previously predicted tokens to influence subsequent predictions, enhancing information exchange between them. This leads to improved performance in downstream tasks in the realm of representation learning. Simultaneously, tokens in PartA can attend to task in PartB, making PartA a generation condition for PartB, resulting in enhanced generation performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "content": "Compare with PointGPT. PointGPT [6] adopts a two-stage transformer architecture for learning representations of point clouds, comprising a feature extractor and a generator. It first extracts features from the point cloud us" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 210, + 547, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 210, + 547, + 331 + ], + "spans": [ + { + "bbox": [ + 304, + 210, + 547, + 331 + ], + "type": "text", + "content": "ing the feature extractor and then directly generates with the obtained features in the generator. During training, it directly employs Chamfer Distance [11] between generated point clouds and ground truth points as the optimization objective, rather than operating on tokens. This is because it doesn't discretize points embeddings, thus failing to achieve genuine autoregressive generation tasks. Moreover, it doesn't integrate the two tasks for joint training in a single transformer; instead, it trains them separately on two different transformers, which increases the complexity." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 338, + 390, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 338, + 390, + 350 + ], + "spans": [ + { + "bbox": [ + 306, + 338, + 390, + 350 + ], + "type": "text", + "content": "4.5. Visualization" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 356, + 547, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 356, + 547, + 513 + ], + "spans": [ + { + "bbox": [ + 304, + 356, + 547, + 513 + ], + "type": "text", + "content": "We visualize the generation results of dVAE and Point-BERT, as shown in the Figure 4a and Figure 4b. Although Point-BERT is capable of masked tokens generation through autoencoding, our approach allows for autoregressive conditional generation tasks, such as text-conditioned point clouds and image-conditioned point clouds. As observed in Figure 4b, the generation results of Point-BERT exhibit some scattered points at the edges of the overall structure, appearing relatively sparse and disorganized. The generated outcome is not as satisfactory as that of GPM. More visualizations of generation results can refer to Unconditional and Conditional Point Generation in the supplementary 6.3." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 522, + 378, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 522, + 378, + 534 + ], + "spans": [ + { + "bbox": [ + 306, + 522, + 378, + 534 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 543, + 547, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 543, + 547, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 543, + 547, + 639 + ], + "type": "text", + "content": "Our study integrates both autoencoding and autoregressive tasks within a unified transformer to enhance the model's performance in representation learning and generation at the same time. By this framework, we address the limitation of current pretraining frameworks and make the proxy tasks non-trivial, effectively enhance the information interaction among masked tokens. Moreover, our framework endows the model with the potential for conditional generation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 649, + 404, + 663 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 649, + 404, + 663 + ], + "spans": [ + { + "bbox": [ + 306, + 649, + 404, + 663 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 669, + 547, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 669, + 547, + 703 + ], + "spans": [ + { + "bbox": [ + 304, + 669, + 547, + 703 + ], + "type": "text", + "content": "This work was supported by National Science and Technology Major Project (No.2022ZD0115101), National Natural Science Foundation of China Project (No.U21A20427)." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20961" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In International conference on machine learning, pages 40-49. PMLR, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 203 + ], + "type": "text", + "content": "[2] Mohamed Afham, Isuru Dissanayake, Dinithi Dissanayake, Amaya Dharmasiri, Kanchana Thilakarathna, and Ranga Rodrigo. Crosspoint: Self-supervised cross-modal contrastive learning for 3d point cloud understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9902-9912, 2022. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 205, + 288, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 288, + 237 + ], + "type": "text", + "content": "[3] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. In International Conference on Learning Representations, 2021. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 239, + 288, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 288, + 294 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 288, + 294 + ], + "type": "text", + "content": "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 295, + 287, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 295, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 53, + 295, + 287, + 350 + ], + "type": "text", + "content": "[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 352, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 352, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 53, + 352, + 287, + 396 + ], + "type": "text", + "content": "[6] Guangyan Chen, Meiling Wang, Yi Yang, Kai Yu, Li Yuan, and Yufeng Yue. Pointgpt: Auto-regressively generative pre-training from point clouds. arXiv preprint arXiv:2305.11487, 2023. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 397, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 397, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 397, + 287, + 441 + ], + "type": "text", + "content": "[7] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International conference on machine learning, pages 1691-1703. PMLR, 2020. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 443, + 287, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 443, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 53, + 443, + 287, + 509 + ], + "type": "text", + "content": "[8] Runpei Dong, Zekun Qi, Linfeng Zhang, Junbo Zhang, Jianjian Sun, Zheng Ge, Li Yi, and Kaisheng Ma. Autoencoders as cross-modal teachers: Can pretrained 2d image transformers help 3d representation learning? In The Eleventh International Conference on Learning Representations, 2022. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 510, + 287, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 510, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 53, + 510, + 287, + 576 + ], + "type": "text", + "content": "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 1, 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 578, + 287, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 643 + ], + "type": "text", + "content": "[10] Zhengxiao Du, Yujie Qian, Xiao Liu, Ming Ding, Jiezhong Qiu, Zhilin Yang, and Jie Tang. Glm: General language model pretraining with autoregressive blank infilling. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 320-335, 2022. 2, 4, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "type": "text", + "content": "[11] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 605-613, 2017. 4, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "text", + "content": "[12] Tianyu Gao, Xingcheng Yao, and Danqi Chen. Simcse: Simple contrastive learning of sentence embeddings. In 2021" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "Conference on Empirical Methods in Natural Language Processing, EMNLP 2021, pages 6894-6910. Association for Computational Linguistics (ACL), 2021. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 107, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 149 + ], + "type": "text", + "content": "[13] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. Pct: Point cloud transformer. Computational Visual Media, 7:187-199, 2021. 1, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 151, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 151, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 151, + 545, + 194 + ], + "type": "text", + "content": "[14] Abdullah Hamdi, Silvio Giancola, and Bernard Ghanem. Mvtn: Multi-view transformation network for 3d shape recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1-11, 2021. 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 195, + 545, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 545, + 248 + ], + "type": "text", + "content": "[15] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 250, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 250, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 308, + 250, + 545, + 271 + ], + "type": "text", + "content": "[16] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 272, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 272, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 308, + 272, + 545, + 327 + ], + "type": "text", + "content": "[17] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q Weinberger. Deep networks with stochastic depth. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 646-661. Springer, 2016. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 327, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 360 + ], + "type": "text", + "content": "[18] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. In International Conference on Learning Representations, 2016. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 361, + 545, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 361, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 308, + 361, + 545, + 414 + ], + "type": "text", + "content": "[19] Mandar Joshi, Danqi Chen, Yinhan Liu, Daniel S Weld, Luke Zettlemoyer, and Omer Levy. Spanbert: Improving pre-training by representing and predicting spans. Transactions of the association for computational linguistics, 8:64-77, 2020. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 415, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 415, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 308, + 415, + 545, + 459 + ], + "type": "text", + "content": "[20] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, pages 4171–4186, 2019. 1, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 460, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 502 + ], + "type": "text", + "content": "[21] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 504, + 545, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 545, + 548 + ], + "type": "text", + "content": "[22] Jiaxin Li, Ben M Chen, and Gim Hee Lee. So-net: Self-organizing network for point cloud analysis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9397–9406, 2018. 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 548, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 613 + ], + "type": "text", + "content": "[23] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 4582-4597, 2021. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 614, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 657 + ], + "type": "text", + "content": "[24] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. Advances in neural information processing systems, 31, 2018. 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "text", + "content": "[25] Zhe Li, T. Yang Laurence, Xin Nie, BoCheng Ren, and Xianjun Deng. Enhancing sentence representation with visually-supervised multimodal pre-training. In Proceedings of the 31st ACM International Conference on Multimedia (ACM MM'23), 2023. 1, 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20962" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[26] Zhe Li, Laurence T Yang, Bocheng Ren, Xin Nie, Zhangyang Gao, Cheng Tan, and Stan Z Li. Mlip: Enhancing medical visual representation with divergence encoder and knowledge-guided contrastive learning. arXiv preprint arXiv:2402.02045, 2024.1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 130, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 130, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 130, + 287, + 174 + ], + "type": "text", + "content": "[27] Haotian Liu, Mu Cai, and Yong Jae Lee. Masked discrimination for self-supervised learning on point clouds. In European Conference on Computer Vision, pages 657-675. Springer, 2022. 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 176, + 287, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 176, + 287, + 231 + ], + "spans": [ + { + "bbox": [ + 48, + 176, + 287, + 231 + ], + "type": "text", + "content": "[28] Yongcheng Liu, Bin Fan, Gaofeng Meng, Jiwen Lu, Shiming Xiang, and Chunhong Pan. Densepoint: Learning densely contextual representation for efficient point cloud processing. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5239-5248, 2019. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 233, + 287, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 233, + 287, + 286 + ], + "spans": [ + { + "bbox": [ + 48, + 233, + 287, + 286 + ], + "type": "text", + "content": "[29] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 289, + 287, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 289, + 287, + 322 + ], + "spans": [ + { + "bbox": [ + 48, + 289, + 287, + 322 + ], + "type": "text", + "content": "[30] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. In International Conference on Learning Representations, 2016. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 324, + 287, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 324, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 48, + 324, + 287, + 346 + ], + "type": "text", + "content": "[31] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 349, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 349, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 349, + 287, + 392 + ], + "type": "text", + "content": "[32] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. In International Conference on Learning Representations, 2021. 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 394, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 394, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 394, + 287, + 437 + ], + "type": "text", + "content": "[33] Chen Min, Dawei Zhao, Liang Xiao, Yiming Nie, and Bin Dai. Voxel-mae: Masked autoencoders for pre-training large-scale point clouds. arXiv preprint arXiv:2206.09900, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 440, + 287, + 484 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 440, + 287, + 484 + ], + "spans": [ + { + "bbox": [ + 48, + 440, + 287, + 484 + ], + "type": "text", + "content": "[34] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In European conference on computer vision, pages 604-621. Springer, 2022. 2, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 486, + 287, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 486, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 48, + 486, + 287, + 540 + ], + "type": "text", + "content": "[35] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 3, 5, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 543, + 287, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 543, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 543, + 287, + 586 + ], + "type": "text", + "content": "[36] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 2, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 589, + 287, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 589, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 589, + 287, + 643 + ], + "type": "text", + "content": "[37] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. Advances in Neural Information Processing Systems, 35:23192-23204, 2022. 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 646, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 677 + ], + "type": "text", + "content": "[38] Shi Qiu, Saeed Anwar, and Nick Barnes. Geometric backprojection network for point cloud classification. IEEE Transactions on Multimedia, 24:1943-1955, 2021. 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[39] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 1, 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[40] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 108, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 161 + ], + "type": "text", + "content": "[41] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2, 3, 4, 5, 6, 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "type": "text", + "content": "[42] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 209, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 262 + ], + "type": "text", + "content": "[43] Yongming Rao, Jiwen Lu, and Jie Zhou. Global-local bidirectional reasoning for unsupervised representation learning of 3d point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5376-5385, 2020. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 264, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 319 + ], + "type": "text", + "content": "[44] Justyna Sarzynska-Wawer, Aleksander Wawer, Aleksandra Pawlak, Julia Szymanowska, Izabela Stefaniak, Michal Jarkiewicz, and Lukasz Okruszek. Detecting formal thought disorder by deep contextualized word representations. *Psychiatry Research*, 304:114135, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 321, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 321, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 308, + 321, + 545, + 354 + ], + "type": "text", + "content": "[45] Jonathan Sauder and Bjarne Sievers. Self-supervised deep learning on point clouds by reconstructing space. Advances in Neural Information Processing Systems, 32, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 355, + 545, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 355, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 308, + 355, + 545, + 388 + ], + "type": "text", + "content": "[46] Charu Sharma and Manohar Kaul. Self-supervised few-shot learning on point clouds. Advances in Neural Information Processing Systems, 33:7212-7221, 2020. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 389, + 545, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 389, + 545, + 433 + ], + "spans": [ + { + "bbox": [ + 308, + 389, + 545, + 433 + ], + "type": "text", + "content": "[47] Chen Sun, Abhinav Shrivastava, Saurabh Singh, and Abhinav Gupta. Revisiting unreasonable effectiveness of data in deep learning era. In Proceedings of the IEEE international conference on computer vision, pages 843-852, 2017. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 434, + 545, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 434, + 545, + 489 + ], + "spans": [ + { + "bbox": [ + 308, + 434, + 545, + 489 + ], + "type": "text", + "content": "[48] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6411–6420, 2019. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 491, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 491, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 491, + 545, + 544 + ], + "type": "text", + "content": "[49] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 546, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 546, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 308, + 546, + 545, + 611 + ], + "type": "text", + "content": "[50] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1588–1597, 2019. 6, 7, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 613, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 657 + ], + "type": "text", + "content": "[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4, 5, 6, 7, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "text", + "content": "[52] Hanchen Wang, Qi Liu, Xiangyu Yue, Joan Lasenby, and Matt J Kusner. Unsupervised point cloud pre-training via occlusion completion. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9782-9792, 2021. 7" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "20963" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[53] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (tog), 38(5):1-12, 2019. 3, 4, 5, 6, 7, 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 288, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 288, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 288, + 173 + ], + "type": "text", + "content": "[54] Ziyi Wang, Xumin Yu, Yongming Rao, Jie Zhou, and Jiwen Lu. P2p: Tuning pre-trained image models for point cloud analysis with point-to-pixel prompting. Advances in neural information processing systems, 35:14388-14402, 2022. 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 288, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 288, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 288, + 229 + ], + "type": "text", + "content": "[55] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1912-1920, 2015. 6, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 230, + 288, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 288, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 288, + 297 + ], + "type": "text", + "content": "[56] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 574-591. Springer, 2020. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 298, + 288, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 288, + 341 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 288, + 341 + ], + "type": "text", + "content": "[57] Zhenda Xie, Yutong Lin, Zhuliang Yao, Zheng Zhang, Qi Dai, Yue Cao, and Han Hu. Self-supervised learning with swim transformers. arXiv preprint arXiv:2105.04553, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 342, + 288, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 342, + 288, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 342, + 288, + 387 + ], + "type": "text", + "content": "[58] Yifan Xu, Tianqi Fan, Mingye Xu, Long Zeng, and Yu Qiao. Spidercnn: Deep learning on point sets with parameterized convolutional filters. In Proceedings of the European conference on computer vision (ECCV), pages 87-102, 2018. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 388, + 288, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 288, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 288, + 443 + ], + "type": "text", + "content": "[59] Guandao Yang, Xun Huang, Zekun Hao, Ming-Yu Liu, Serge Belongie, and Bharath Hariharan. Pointflow: 3d point cloud generation with continuous normalizing flows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4541-4550, 2019. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 445, + 288, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 288, + 489 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 288, + 489 + ], + "type": "text", + "content": "[60] Yaoqing Yang, Chen Feng, Yiru Shen, and Dong Tian. Foldingnet: Point cloud auto-encoder via deep grid deformation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 206–215, 2018. 4, 5, 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 490, + 288, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 490, + 288, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 490, + 288, + 544 + ], + "type": "text", + "content": "[61] Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Russ R Salakhutdinov, and Quoc V Le. Xlnet: Generalized autoregressive pretraining for language understanding. Advances in neural information processing systems, 32, 2019. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 545, + 288, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 288, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 288, + 601 + ], + "type": "text", + "content": "[62] Li Yi, Vladimir G Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas Guibas. A scalable active framework for region annotation in 3d shape collections. ACM Transactions on Graphics (ToG), 35(6):1-12, 2016. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 602, + 288, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 288, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 288, + 657 + ], + "type": "text", + "content": "[63] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19313-19322, 2022. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "text", + "content": "[64] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. Point-m2ae: multi-scale masked autoencoders for hierarchical point cloud pre-training. Advances in neural information processing systems, 35:27061-27074, 2022. 2, 7" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 547, + 162 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 307, + 72, + 547, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 547, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 547, + 116 + ], + "type": "text", + "content": "[65] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16259-16268, 2021. 1, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 118, + 547, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 162 + ], + "type": "text", + "content": "[66] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In International Conference on Learning Representations, 2020. 1" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "20964" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Generalizable Face Landmarking Guided by Conditional Face Warping/33b95235-e8cb-4b6b-a79e-5de2b0524965_content_list.json b/2024/Generalizable Face Landmarking Guided by Conditional Face Warping/33b95235-e8cb-4b6b-a79e-5de2b0524965_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..dc8ec13aae32d5cbfec2bf2b2dc1a6d84f8c8a36 --- /dev/null +++ b/2024/Generalizable Face Landmarking Guided by Conditional Face Warping/33b95235-e8cb-4b6b-a79e-5de2b0524965_content_list.json @@ -0,0 +1,1834 @@ +[ + { + "type": "text", + "text": "Generalizable Face Landmarking Guided by Conditional Face Warping", + "text_level": 1, + "bbox": [ + 122, + 130, + 846, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiayi Liang $^{1}$ Haotian Liu $^{1*}$ Hongteng Xu $^{2,3}$ Dixin Luo $^{1,4\\dagger}$", + "bbox": [ + 240, + 179, + 732, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Computer Science and Technology, Beijing Institute of Technology, Beijing", + "bbox": [ + 143, + 198, + 828, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Gaoling School of Artificial Intelligence, Renmin University of China, Beijing", + "bbox": [ + 171, + 215, + 797, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3Beijing Key Laboratory of Big Data Management and Analysis Methods, Beijing", + "bbox": [ + 161, + 233, + 807, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ Key Laboratory of Artificial Intelligence, Ministry of Education, Shanghai", + "bbox": [ + 187, + 251, + 779, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{jiayi.liang, haotianliu, dixin.luo}@bit.edu.cn, hongtengxu@ruc.edu.cn", + "bbox": [ + 173, + 271, + 789, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 321, + 313, + 337 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As a significant step for human face modeling, editing, and generation, face landmarking aims at extracting facial keypoints from images. A generalizable face landmarker is required in practice because real-world facial images, e.g., the avatars in animations and games, are often stylized in various ways. However, achieving generalizable face landmarking is challenging due to the diversity of facial styles and the scarcity of labeled stylized faces. In this study, we propose a simple but effective paradigm to learn a generalizable face landmarker based on labeled real human faces and unlabeled stylized faces. Our method learns the face landmarker as the key module of a conditional face warper. Given a pair of real and stylized facial images, the conditional face warper predicts a warping field from the real face to the stylized one, in which the face landmarker predicts the ending points of the warping field and provides us with high-quality pseudo landmarks for the corresponding stylized facial images. Applying an alternating optimization strategy, we learn the face landmarker to minimize i) the discrepancy between the stylized faces and the warped real ones and ii) the prediction errors of both real and pseudo landmarks. Experiments on various datasets show that our method outperforms existing state-of-the-art domain adaptation methods in face landmarking tasks, leading to a face landmarker with better generalizability. Code is available at https://plustwo0.github.io/project-face-landmarker.", + "bbox": [ + 75, + 354, + 473, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 776, + 209, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Face landmarking seeks to extract human facial keypoints (e.g., eyes, nose, facial contour, and so on) from facial images. This task is important for many applications in the field of computer vision and graphics, such as face recogni", + "bbox": [ + 75, + 801, + 468, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/66c55eec947896835ab6c5a3a888c16ac0860ff886075170bc03387086dab6c3.jpg", + "image_caption": [ + "Figure 1. Both commercial software like Face++ and open-source method like SLPT [52] work well on landmarking real faces (e.g., those in 300W [38]) while achieving suboptimal performance when landmarking stylized faces (e.g., those in CariFace [4] and ArtiFace [55]). While existing domain adaptation method does not improve the performance significantly, our method achieves a generalizable face landmarker for various facial images." + ], + "image_footnote": [], + "bbox": [ + 516, + 323, + 877, + 531 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tion [1, 28, 33, 43, 53], face stylization [5], and 3D face reconstruction [11, 26, 37]. Currently, many open-source and commercial face landmarks [19, 51, 58] have been developed and achieved encouraging performance in this task.", + "bbox": [ + 496, + 670, + 892, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Most existing face landmarks are designed and trained for landmarking real human faces, while the rapid development of AIGC applications, such as artistic character creation and cartoon generation [2, 15], leads to a massive increase in demands for landmarking stylized facial images. Unfortunately, as shown in Fig. 1, existing face landmarks often fail to landmark stylized facial images. Even if applying state-of-the-art domain adaptation strategies [12, 25, 27, 59, 61], the generalizability of the learned landmarks in the stylized facial image domain is still unsatisfactory. Essentially, traditional face landmarks work", + "bbox": [ + 496, + 734, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*The first two authors contributed equally to this work.", + "bbox": [ + 94, + 875, + 385, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 96, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2425", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/585c55b834d70529e61d8d9e43412b1b0772b11c4d293ddc9e534febc672ec3a.jpg", + "image_caption": [ + "Figure 2. The scheme of our proposed method for learning a generalizable face landmarker." + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 467, + 333 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "well on real human faces because of the relatively stable geometry of real human faces and the sufficient labeled facial images. These two conditions, however, become questionable when landmarking stylized faces — the stylized facial images often have various facial styles, and manually landmarking such stylized faces is much more time-consuming than landmarking real faces. As a result, learning a generalizable face landmarker becomes a challenging task.", + "bbox": [ + 75, + 401, + 468, + 521 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a simple but effective paradigm for learning a generalizable face landmarker, overcoming the challenges caused by the diversity of facial styles and the scarcity of labeled stylized faces. As illustrated in Fig. 2, given labeled real human facial images and unlabeled stylized facial images, we learn a face landmarker embedded in a conditional face warper. The face warper aims to deform real human faces according to stylized facial images, generating warped faces and corresponding warping fields. The face landmarker, as the key module of the warper, predicts the ending points of the warping fields and thus provides us with pseudo landmarks for the stylized facial images. The warping field is parametrized by a polyharmonic interpolation model. Under the guidance of the conditional face warping, we learn the face landmarker in an alternating optimization framework: The face landmarker is updated to $i$ minimize the discrepancy between the stylized faces and the corresponding warped real human faces and $ii$ minimize the prediction errors of both real and pseudo landmarks. In the first step, the face landmarker is learned associated with the warping field model, while in the second step, the face landmarker is updated with proximal regularization.", + "bbox": [ + 75, + 522, + 470, + 853 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The extensive experiments in various face landmarking tasks demonstrate the effectiveness of learning method. The impacts of different loss functions and data settings on", + "bbox": [ + 76, + 854, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "our learning method are analyzed through detailed ablation studies. Experimental results show that our learning method results in a face landmarker that is generalizable to the facial images with different styles, which outperforms representative domain adaptation methods consistently in various stylized face landmarking tasks.", + "bbox": [ + 496, + 90, + 890, + 183 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 196, + 640, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Face Landmarking", + "text_level": 1, + "bbox": [ + 500, + 222, + 684, + 239 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given facial images, the early face landmarking methods learn regression models to predict landmark coordinates directly [6, 32]. The models are often parameterized by neural networks like Transformer [46], capturing facial attributes that have been proven to be crucial for landmark prediction [24]. The coordinate regression is achieved in a coarse-to-fine framework [58], leading to a cascading landmarking pipeline. Recently, to make the landmark prediction robust to the variations in pose, scale, and occlusion, DAN [22] introduces the novel use of heatmaps and extracts features from the entire face rather than local patches around landmarks. SBR [9] utilizes the registration of synthesized images to provide supervisory signals for training. Adaptive Wing Loss [48] is proposed to address the imbalance between foreground pixels and background pixels by analysis of the main drawbacks of different loss functions. HRNet [47] produces high-resolution maps by connecting and exchanging information via merging multi-scale picture features across many branches.", + "bbox": [ + 496, + 244, + 890, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As aforementioned, most existing face landmarks are learned for real human faces, which can not be adapted directly to stylized faces (e.g., cartoon and artistic faces). Essentially, treating labeled real human faces as a source domain and unlabeled stylized faces as a target domain, we can learn a generalizable face landmark by solving a domain adaptation (DA) problem [30, 31]. Accordingly, many DA techniques [27, 49] have potential in our problem, including the classic metric learning-based methods (e.g., CORAL [41], contrastive domain discrepancy [17], and maximum mean discrepancy [16]) and the recent adversarial learning-based methods [12, 35, 56, 59]. However, in the following content, we will show that directly applying these DA techniques often fails to achieve generalizable face landmarking because of the significant gap between the source and target face domains.", + "bbox": [ + 496, + 534, + 890, + 776 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Face Warping", + "text_level": 1, + "bbox": [ + 500, + 786, + 645, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Face warping is a technique that involves geometrically deforming source facial images to specified target shapes. The key step of this task is predicting a warping field between the source and target images that captures the shifts of image pixels. To achieve this aim, DST [20] and FoA [55] find matching keypoints between source and target images", + "bbox": [ + 496, + 809, + 890, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2426", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and then generate a dense warping field through data interpolation [8]. Instead of matching keypoints, some methods learn neural networks to predict dense warping fields directly based on paired images, e.g., Flownet [10], AutoToon [15], RAFT [44], and their variants [29]. However, these methods require the paired images to be similar to each other, which cannot capture significant deformations between the faces in different domains.", + "bbox": [ + 75, + 90, + 468, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Compared to face stylization [13, 18, 39, 40], face warping is a relatively easier task because it only considers the deformation of shapes while ignoring the transfer of textures. However, it should be noted that this task is more relevant to face landmarking, in which the warping field provides us with strong evidence to shift face landmarks of source faces to target ones [50]. Inspired by such a strong correlation, we develop the proposed learning paradigm.", + "bbox": [ + 75, + 210, + 470, + 333 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 76, + 344, + 250, + 361 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Denote $\\mathcal{X}$ as the image space and $\\mathcal{V}$ as the landmark space, respectively. In this work, we observe a set of labeled real human faces, i.e., $\\mathcal{D}^{(L)} = \\{\\pmb{X}_i^{(L)},\\pmb{Y}_i^{(L)}\\}_{i = 1}^{N_L}\\subset \\mathcal{X}_R\\times \\mathcal{V}$ and a set of unlabeled stylized faces, i.e., $\\mathcal{D}^{(U)} = \\{\\pmb{X}_i^{(U)}\\}_{i = 1}^{N_U}\\subset$ $\\mathcal{X}_S$ , where $\\mathcal{X}_R,\\mathcal{X}_S\\subset \\mathcal{X}$ correspond to the real and stylized face domains, respectively. Each $\\pmb {X}_i\\in \\mathbb{R}^{H\\times W\\times 3}$ represents an image, and each $\\pmb {Y}_i = [\\pmb {y}_{i,k}]\\in \\mathbb{R}^{2\\times K}$ records $K$ face landmark coordinates, where $\\pmb {y}_{i,k}\\in \\mathbb{R}^2$ . We aim to learn a face landmarker, denoted as $f_{\\theta}:\\mathcal{X}\\mapsto \\mathcal{V}$ , where $\\theta$ is the model parameter. The model should be able to predict face landmarks from facial images and moreover, generalize to both $\\mathcal{X}_R$ and $\\mathcal{X}_S$ . To achieve this aim, we embed the face landmarker into a conditional face warper and learn it associated with a parametric warping field predictor in an alternating optimization framework, as illustrated in Fig. 2.", + "bbox": [ + 75, + 369, + 468, + 602 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Face Landmarking Guided by Face Warping", + "text_level": 1, + "bbox": [ + 76, + 609, + 457, + 626 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this study, we take the SLPT model [52] as the backbone of our face landmarker. Given a stylized face $\\mathbf{X}_i^{(U)}$ , the face landmarker predicts its landmarks as $\\widehat{\\mathbf{Y}}_i^{(U)} = [\\widehat{\\mathbf{y}}_{i,k}^{(U)}] = f_\\theta(\\mathbf{X}_i^{(U)})$ . At the same time, we can sample a labeled real face $(\\mathbf{X}_j^{(L)},\\mathbf{Y}_j^{(L)}) \\sim \\mathcal{D}^L$ . Treating the labeled and predicted landmarks as keypoints, we can model a warping field from the real face to the stylized one by the following polyharmonic interpolation model [14]:", + "bbox": [ + 75, + 632, + 467, + 766 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nw _ {i, \\gamma} (\\boldsymbol {y}) = \\sum_ {k = 1} ^ {K} \\omega_ {k} \\phi \\left(\\left\\| \\boldsymbol {y} - \\hat {\\boldsymbol {y}} _ {i, k} ^ {(U)} \\right\\| _ {2}\\right) + \\boldsymbol {V} \\boldsymbol {y} + \\boldsymbol {b}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 772, + 468, + 800 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\gamma = \\{\\{\\pmb{\\omega}_k\\in \\mathbb{R}^2\\}_{k = 1}^K,\\pmb {V}\\in \\mathbb{R}^{2\\times 2},\\pmb {b}\\in \\mathbb{R}^2\\}$ correspond to the parameters of the warping field. As shown in (1), the vector $\\pmb{y}$ denotes the u-v coordinate of a pixel in the stylized facial image, and $w_{i,\\gamma}(\\pmb {y})$ gives the inverse mapping from the pixel $\\pmb{y}$ to a coordinate in the real human facial image, conditioned on $\\pmb{X}_i^{(U)}$ . The first term", + "bbox": [ + 75, + 806, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\sum_{k=1}^{K} \\omega_k \\phi(\\| \\pmb{y} - \\hat{\\pmb{y}}_{i,k}^{(U)} \\|_2)$ achieves nonparametric regression for modeling nonrigid deformations, in which $\\phi(r)$ is a predefined thin-plate spline function. The second term $\\pmb{V}\\pmb{y} + \\pmb{b}$ is a linear parametric model capturing the rigid transformation of $\\pmb{y}$ .", + "bbox": [ + 498, + 88, + 890, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For each pixel coordinate $\\pmb{y} \\in \\{1, \\dots, H\\} \\times \\{1, \\dots, W\\}$ , we can trace it back to the real human facial image based on $w_{i,\\gamma}(\\pmb{y})$ and obtain the pixel color as $\\pmb{X}_j^{(L)}(w_{i,\\gamma}(\\pmb{y}))$ . Accordingly, with the grid sampler constructed via inverse mapping function $w_{i,\\gamma}$ , we obtain the warped real human facial image conditioned on $\\pmb{X}_i^{(U)}$ , denoted as $\\widehat{\\pmb{X}}_{j|i}^{(L)}$ . For $\\pmb{y} \\in \\{1, \\dots, H\\} \\times \\{1, \\dots, W\\}$ and $j = 1, \\dots, N_L$ , we have", + "bbox": [ + 496, + 167, + 890, + 281 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {\\boldsymbol {X}} _ {j \\mid i} ^ {(L)} (\\boldsymbol {y}) = \\boldsymbol {X} _ {j} ^ {(L)} \\left(w _ {i, \\gamma} (\\boldsymbol {y})\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 286, + 890, + 310 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unlike WarpGAN [40], which generates the warped face by predicting dense keypoints and their displacements by two fully-connected layers during training, we directly use the predicted and observed landmarks to define sparse displacements and estimate other pixels' displacements by spline-based interpolation, which improves computational efficiency significantly. Moreover, by applying the warping field model with limited degree-of-freedom (i.e., few learnable parameters), we can focus more on the learning of the face landmarker in the training phase.", + "bbox": [ + 496, + 315, + 890, + 465 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, the warped face together with the warping field provides a useful guidance for the learning of the face landmarker. In particular, we formulate the learning problem of the face landmarker as follows:", + "bbox": [ + 496, + 467, + 890, + 526 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min _ {\\theta , \\gamma} \\underbrace {\\sum_ {j = 1} ^ {N _ {L}} \\| f _ {\\theta} \\left(\\boldsymbol {X} _ {j} ^ {(L)}\\right) - \\boldsymbol {Y} _ {j} ^ {(L)} \\| _ {F} ^ {2}} _ {\\text {L a n d m a r k i n g e r r o r i n t h e s o u r c e d o m a i n}} \\\\ + \\underbrace {\\sum_ {i = 1} ^ {N _ {U}} \\sum_ {j = 1} ^ {N _ {L}} \\left\\| \\nabla \\widehat {\\boldsymbol {X}} _ {j | i} ^ {(L)} - \\nabla \\boldsymbol {X} _ {i} ^ {(U)} \\right\\| _ {F} ^ {2}} _ {\\text {D i s c r e p a n c y o f i m a g e g r a d i e n t}} \\tag {3} \\\\ + \\underbrace {\\sum_ {i = 1} ^ {N _ {U}} \\sum_ {j = 1} ^ {N _ {L}} \\| w _ {i , \\gamma} (\\widehat {\\mathbf {Y}} _ {i} ^ {(U)}) - \\mathbf {Y} _ {j} ^ {(L)} \\| _ {F} ^ {2}} _ {\\text {L a n d m a r k w a r p i n g e r r o r}}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 531, + 890, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\| \\cdot \\| _F$ represents the Frobenius norm of matrix. In (3), the first term is the landmarking error for real faces, which corresponds to the data fidelity loss in the source domain. The second term measures the discrepancy between the stylized face and the warped real face in the gradient field, in which the gradient operation $\\nabla$ is implemented by the Sobel operator. The third term is the landmark warping error. Both the second and third terms are determined jointly by the landmarker $f_{\\theta}$ and warping field model $w_{i,\\gamma}$ .", + "bbox": [ + 496, + 686, + 890, + 825 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Alternating Optimization Strategy", + "text_level": 1, + "bbox": [ + 500, + 832, + 800, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The optimization problem in (3) is non-convex because the landmarker is implemented by a neural network and is coupled to the warping field model. As a result, learning $\\theta$ and", + "bbox": [ + 496, + 854, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2427", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b2e949308641b6c3c5184263e60a6bee22e42be3fc42af04f0f2207e4b5845cf.jpg", + "image_caption": [ + "Warped Face with Warped Landmarks" + ], + "image_footnote": [], + "bbox": [ + 83, + 99, + 158, + 157 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d7db80611db195609997eb5cd87fdc0da8d47ff03d9995835ccc1b87ec45b688.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 160, + 101, + 233, + 157 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2799767401c9c88f2bcad3101714f5661bc61eb91b9976204df0e645bf76233f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 235, + 101, + 310, + 157 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9756a62a1b30f6c9a3900aace7257873629c25a0ab32b9a23c71b796fdd7aa8d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 310, + 101, + 385, + 157 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b4e305fe09aecd53102671f37ace164c08b8d0fc1adf6643c5bd69489be3a340.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 387, + 101, + 462, + 157 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3b2ec01642ccf3f00d47fadc7173d258c25a78116ad57632b544c60de13cfebb.jpg", + "image_caption": [ + "Figure 3. Illustrations of conditional face warping results. Taking a cartoon face as the target, our model warps real human faces accordingly. The red dots indicate real human face landmarks, and green dots indicate cartoon and warped face landmarks." + ], + "image_footnote": [], + "bbox": [ + 158, + 174, + 233, + 232 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c642614f178cfca3d97bdd1f4840cdf6807463e1fe737706398ea34431f6e0c4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 235, + 174, + 310, + 232 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/678ec486c5029a74aba8d95f0948a1fee8e4256f96951544bbb3be72a35f21d2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 310, + 174, + 385, + 232 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/213138cb5e4acc88168d0514a3dc3c26031dcc07841dd9a2e2615a3a432759fc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 387, + 174, + 462, + 232 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\gamma$ jointly often falls into an undesired local optimum even an unstable saddle point. To mitigate this issue, we propose an alternating optimization framework. In principle, we can decompose the optimization problem in (3) into the following two subproblems and solve them iteratively.", + "bbox": [ + 75, + 327, + 468, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Face Warper Optimization: The first subproblem corresponds to the optimization of the face warper, i.e.,", + "bbox": [ + 76, + 405, + 468, + 436 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\theta^ {(1)}, \\gamma^ {(1)} = \\arg \\min _ {\\theta , \\gamma} \\sum_ {i, j} \\| \\nabla \\widehat {\\boldsymbol {X}} _ {j | i} ^ {(L)} - \\nabla \\boldsymbol {X} _ {i} ^ {(U)} \\| _ {F} ^ {2} \\tag {4} \\\\ + \\sum_ {i, j} \\| w _ {i, \\gamma} (\\widehat {\\pmb {Y}} _ {i} ^ {(U)}) - \\pmb {Y} _ {j} ^ {(L)} \\| _ {F} ^ {2}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 444, + 477, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this subproblem, we only care about whether the real human faces can be warped as the stylized faces with high accuracy, so the term of landmarking error is ignored. We solve this problem by Adam [21]: in each step, we update $\\theta$ and $\\gamma$ based on a batch of randomly-sampled face pairs.", + "bbox": [ + 89, + 507, + 468, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Proximal Face Landmarker Optimization: Given $\\theta^{(1)}$ and the predicted landmarks (i.e., $\\widehat{\\pmb{Y}}_i^{(U)} = f_{\\theta^{(1)}}(\\pmb{X}_i^{(U)})$ for $i = 1, \\dots, N_L$ ), we can treat $\\theta^{(1)}$ as the initial variable and optimize it with a proximal regularizer:", + "bbox": [ + 76, + 583, + 468, + 646 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\theta^ {(2)} = \\arg \\min _ {\\theta} \\sum_ {j = 1} ^ {N _ {L}} \\| f _ {\\theta} (\\boldsymbol {X} _ {j} ^ {(L)}) - \\boldsymbol {Y} _ {j} ^ {(L)} \\| _ {F} ^ {2} \\\\ + \\underbrace {\\sum_ {i = 1} ^ {N _ {U}} \\| f _ {\\theta} \\left(\\boldsymbol {X} _ {i} ^ {(U)}\\right) - \\widehat {\\boldsymbol {Y}} _ {i} ^ {(U)} \\| _ {F} ^ {2}}. \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 655, + 468, + 719 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pseudo landmarking error in the target domain", + "bbox": [ + 184, + 720, + 401, + 732 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, the second term in (5) measures the estimation errors of the pseudo landmarks achieved in the previous step. Essentially, it works as a proximal regularizer, ensuring that the optimized landmarks $f_{\\theta^{(2)}}(\\mathbf{X}_i^{(U)})$ is not too far away from the previous estimation $\\widehat{\\mathbf{Y}}_i^{(U)}$ . Similarly, we can solve this problem by Adam [21] as well.", + "bbox": [ + 89, + 742, + 468, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fig. 3 shows the warping effect on real human faces achieved by solving (4). In Fig. 3, the first row shows the real human faces with landmarks and the target stylized face, and the second row shows the warping results,", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 Proposed learning scheme of face landmarker", + "text_level": 1, + "bbox": [ + 500, + 90, + 890, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Require: Labeled real faces $\\mathcal{D}^{(L)}$ and unlabeled stylized faces $\\mathcal{D}^{(U)}$ . The number of iterations (i.e., $M$ ). Epochs for the subproblems (i.e., $L_{1}$ and $L_{2}$ ).", + "bbox": [ + 501, + 112, + 890, + 157 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Initialize $\\{\\gamma^{(0)}\\}$ with a pretrained model on $\\mathcal{D}^{(L)}$ and $\\{\\theta^{(0)}\\}$ randomly.", + "2: for $m = 0, \\dots, M - 1$ do", + "3: Sample a batch $\\{\\pmb{X}_i^{(U)},\\pmb{X}_i^{(L)},\\pmb{Y}_i^{(L)}\\}_{i = 1}^N$", + "4: Face warper optimization:", + "5: Take $\\theta^{(2m)},\\gamma^{(m)}$ as the initialization, then solve (4) by Adam with $L_{1}$ epochs and obtain $\\theta^{(2m + 1)}$", + "6: Proximal face landmarker optimization:", + "7: Take $\\theta^{(2m + 1)}$ as the initialization, then solve (5) by Adam with $L_{2}$ epochs and obtain $\\theta^{(2m + 2)}$ .", + "8: end for", + "9: return Output a generalizable face landmarker $f_{\\theta^{(2M)}}$ ." + ], + "bbox": [ + 509, + 157, + 890, + 342 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "in which the green dots are predicted landmarks. These results empirically demonstrate the rationality of our alternating optimization framework. In particular, we can find that solving (4) leads to reasonable warping results, which are similar to the target stylized face on shape. The similarity on face shape indicates that the predicted landmarks can be treated as reliable pseudo labels of the stylized face, which can be used to construct the proximal regularizer that penalizing the peso do landmarking errors in the target domain. Repeating the above two steps till converge, we obtain the target face landmarker that is generalizable for both real and stylized faces. Algorithm 1 shows the learning scheme.", + "bbox": [ + 496, + 372, + 890, + 554 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 500, + 568, + 624, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We apply our learning method to learn a face landmarker and test it on landmarking faces with various styles. Extensive experiments, including comparisons with baselines and analytic ablation studies, demonstrate the effectiveness of our learning method and the generalizability of the corresponding face landmarker. All the experiments are conducted on a single NVIDIA 3090 GPU. Representative experimental results are shown below. More experimental results and implementation details are given in the supplementary material.", + "bbox": [ + 496, + 594, + 890, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Dataset", + "text_level": 1, + "bbox": [ + 500, + 755, + 594, + 770 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this study, we conduct experiments based on the following three commonly-used face datasets.", + "bbox": [ + 496, + 779, + 888, + 809 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- 300W Dataset. 300W [38] is comprised of five well-known real human face datasets including LFPW [3], AFW [60], HELEN [23], XM2VTS [34], and IBUG [38].", + "- CariFace Dataset. CariFace [4] is created by searching and selecting thousands of various caricatures from different celebrities on the Internet." + ], + "bbox": [ + 500, + 810, + 890, + 898 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Cartoon Face Real Human Face", + "bbox": [ + 84, + 88, + 243, + 98 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2428", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/0216e2426655312b416a8d63a1eb44d9c7615834e823ffb6796152cd4f9fe7c8.jpg", + "table_caption": [ + "Table 1. Data settings for the three learning paradigms." + ], + "table_footnote": [], + "table_body": "
Learning Paradigm#Training Images and Label Information#Testing Images
300WCariFaceArtiFaceCommon300W ChallengeFullCariFaceArtiFace
DA (300W→CariFace)3,148Labeled3,372Unlabeled——554135689800—
DA (300W→ArtiFace)3,148Labeled——128Unlabeled554135689—32
GZSL (Unseen ArtiFace)3,148Labeled3,372Unlabeled——554135689800160
GZSL (Unseen CariFace)3,148Labeled——128Unlabeled55413568980032
Oracle3,148Labeled3,372Labeled128Labeled55413568980032
", + "bbox": [ + 78, + 114, + 890, + 236 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5fa9e5a7fd20df897a92c2f02d3b3db7c70ae67afbc1f356308b48918f7dda41.jpg", + "image_caption": [ + "Figure 4. Illustrations of typical samples in the 300W, CariFace, and ArtiFace datasets, each of which is annotated with landmarks." + ], + "image_footnote": [], + "bbox": [ + 81, + 258, + 468, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- ArtiFace Dataset. ArtiFace [55] contains 160 artistic portraits of 16 artists, which covers diverse artwork styles ranging from Renaissance to Comics.", + "bbox": [ + 76, + 527, + 468, + 573 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Each face in the datasets is annotated with 68 landmarks. Typical faces in the datasets and their landmarks are shown in Fig. 4. We can find that the faces in the three datasets have distinguished styles, which correspond to three different domains. In particular, compared to 300W [38], CariFace [4] exhibits abstract and exaggerated patterns, leading to large representation variations. ArtiFace [55] not only has larger variations across different artistic categories but also differs greatly in terms of the aspect of facial scales, orientations, locations, and so on.", + "bbox": [ + 75, + 575, + 468, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Learning Paradigms and Baselines", + "text_level": 1, + "bbox": [ + 76, + 738, + 380, + 755 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the above datasets, we consider the following three learning paradigms:", + "bbox": [ + 75, + 763, + 468, + 792 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Domain Adaptation (DA). Given labeled 300W faces and unlabeled stylized faces from CariFace or ArtiFace, we learn a face landmarker based on various domain adaptation methods.", + "- Generalized Zero-shot Learning (GZSL). In the challenging GZSL setting, we learn a face landmarker based on the above DA-based methods and test it in an unseen" + ], + "bbox": [ + 76, + 795, + 468, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "face domain (e.g., learning the landmarker on labeled 300W and unlabeled CariFace and testing on ArtiFace).", + "bbox": [ + 509, + 262, + 890, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- Oracle. In this setting, the labeled faces of all three datasets are accessible, and we can learn the face landmark by classic supervised learning.", + "bbox": [ + 500, + 292, + 890, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For a fair comparison, in each learning paradigm, we set the architecture of the face landmark based on the SLPT in [52]. Ideally, we would like to learn landmarks in the DA and GZSL settings, making its performance comparable to the oracle. In the oracle setting, we can learn the face landmark directly via classic supervised learning (SL), i.e., $\\min_{\\theta}\\sum_{(\\mathbf{X},\\mathbf{Y})\\sim \\mathcal{D}}\\| f_{\\theta}(\\mathbf{X}) - \\mathbf{Y}\\| _F^2$ . In the DA and GZSL settings, besides minimizing the landmark estimation errors, we can apply various image style transfer and domain adaptation methods, e.g., RevGrad [12], CycleGAN [59], BDL [25], AdaptSegNet [45] and FDA [54], to impose domain adaptation regularization during training. These methods work as the baselines of our method.", + "bbox": [ + 496, + 337, + 890, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the landmarkers learned by various methods, we evaluate them with the standard metric, Normalized Mean Error (NME). In Tab. 1, we show the training and testing data settings in the above three learning paradigms. Following existing work [9, 42, 52], we further split the 689 testing faces in 300W into 554 faces in common scenarios and 135 faces in challenging scenarios. The NMEs for the common, challenge, and full scenarios are recorded.", + "bbox": [ + 496, + 534, + 890, + 655 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Numerical and Visual Comparisons", + "text_level": 1, + "bbox": [ + 498, + 665, + 808, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Tab. 2, we show the performance of various learning methods in different settings, demonstrating the effectiveness and superiority of our method. In particular, existing DA methods often fail to improve the generalization power of model a lot in face landmarking tasks — their performance in target and unseen domains is inferior to that in the oracle setting, with a significant gap on NME. A potential reason for this phenomenon is that these methods focus on the adaptation of image domain and the landmark-related loss is not dominant in their learning processes. As a result, instead of learning the face landmarker, they make more efforts to optimize the parameters of other modules (e.g., the neural network-based face stylization modules and discriminators) during training.", + "bbox": [ + 496, + 688, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2429", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/436e72c9fcd668ccad6ff1f955849c66922266cd48cf548f2e8bbca5054db73d.jpg", + "table_caption": [ + "Table 2. Comparisons for various methods on their NMEs. In DA and GZSL settings, the best results are bold." + ], + "table_footnote": [], + "table_body": "
Learning ParadigmLearning Method300WCariFaceArtiFaceAverage NME
CommonChallengeFull
DA (300W→CariFace) and GZSL (Unseen ArtiFace)SL+RevGrad [12]2.845.583.3812.195.167.83
SL+CycleGAN [59]2.745.433.2712.114.707.70
SL+BDL [25]3.286.173.8413.635.658.77
SL+ASN [45]2.925.263.3812.214.757.80
SL+FDA [54]2.895.183.3412.664.608.07
Ours2.794.913.207.703.955.46
DA (300W→ArtiFace) and GZSL (Unseen CariFace)SL+RevGrad [12]2.995.813.5512.464.748.26
SL+CycleGAN [59]3.005.653.5212.645.348.36
SL+BDL [25]2.995.323.4413.405.908.73
SL+ASN [45]2.895.813.4616.585.6510.31
SL+FDA [54]3.056.213.6812.335.878.27
Ours2.905.143.3410.933.937.34
OracleSL2.684.863.105.483.314.36
", + "bbox": [ + 120, + 114, + 846, + 333 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4dfc873a1ce4d4928e5caad88db58af52719b4d664410a90e434def2aca28bfd.jpg", + "table_caption": [ + "Table 3. The impacts of different model architectures on the NME performance of our method. In each setting, the best results are bold." + ], + "table_footnote": [], + "table_body": "
Learning ParadigmLearning Method300WCariFaceArtiFaceAverage NME
CommonChallengeFull
DA (300W→CariFace) and GZSL (Unseen ArtiFace)SBR [9]3.616.364.158.005.096.11
HRNet [47]2.935.293.407.734.335.60
SLPT [52] (Default)2.794.913.207.703.955.46
DA (300W→ArtiFace) and GZSL (Unseen CariFace)SBR [9]3.706.824.3211.355.328.04
HRNet [47]3.045.443.519.854.306.86
SLPT [52] (Default)2.905.143.3410.933.937.34
", + "bbox": [ + 120, + 373, + 846, + 494 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0f8cab360045e690a2ef3f18f61a48a52f72e1c04aecf4ba1c142ec126f49549.jpg", + "image_caption": [ + "Figure 5. Visual comparisons for various methods in the two DA settings. We only highlight points on the inner lips in the enlarged region of the mouth in (a), as well as the eyes and the sides of the cheeks, excluding points on the eyebrows in (b)." + ], + "image_footnote": [], + "bbox": [ + 78, + 515, + 470, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Different from the baselines, the performance of our model in the target domain (e.g., CariFace and ArtiFace) is improved consistently, while the performance in the source", + "bbox": [ + 75, + 854, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/be662788e90a37c6c76e7c6d4cdb09bcb6bed248a4a8d98d513228704ab22472.jpg", + "image_caption": [ + "Figure 6. Visual comparisons for various methods in the GZSL (Unseen ArtiFace) setting. The zoomed-in area in (a) highlights the noses and facial contours, while that in (b) concentrates on the upper part of faces." + ], + "image_footnote": [], + "bbox": [ + 500, + 515, + 893, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "domain does not degrade a lot. In both DA and GZSL settings, our method outperforms the baselines in most situations. Especially in the GZSL settings, the landmarks", + "bbox": [ + 496, + 854, + 892, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2430", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/68c259f5c1c68b871a15d646c741b149739eef3874cb2d78b720ccb00cd59f25.jpg", + "image_caption": [ + "(a) 300W (Common)" + ], + "image_footnote": [], + "bbox": [ + 98, + 98, + 246, + 215 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e704056cb9d7ab65b4a82d604e79c6df3f903bb4d7559b802ccfe96407c6b825.jpg", + "image_caption": [ + "(b) 300W (Challenge)" + ], + "image_footnote": [], + "bbox": [ + 256, + 98, + 403, + 215 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/616dba4ead884d1aedf03fb224f032c7ace69779ce82530ae5bc0863910a4685.jpg", + "image_caption": [ + "(c) 300W (Full)", + "Figure 7. Comparisons for different optimization strategies on NME performance in the GZSL (Unseen ArtiFace) setting." + ], + "image_footnote": [], + "bbox": [ + 413, + 98, + 560, + 215 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/623d17bf8a17fe90e75948eb5ddd406f1058539ba3643be74823c7e87dd174f7.jpg", + "image_caption": [ + "(d) CariFace (Target)" + ], + "image_footnote": [], + "bbox": [ + 566, + 98, + 717, + 215 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8af5e33f14e5a98d379b62e8b5a9a1aba8c3603882bb906513a3f8dd29f368f3.jpg", + "image_caption": [ + "(e) ArtiFace (Unseen)" + ], + "image_footnote": [], + "bbox": [ + 728, + 98, + 875, + 215 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c66f473e15a7922191763432053e1a6f6465f54508ed665838b36dd40fa694bc.jpg", + "table_caption": [ + "Table 4. Comparison with SOTA landmark detectors on NME." + ], + "table_footnote": [], + "table_body": "
Learning MethodTrain on 300WOur method
BackboneOPSPIGASTARSLPTSTARSLPT
DA on Cariface10.4611.2310.9711.057.627.70
GZSL on Artiface5.555.195.304.564.863.93
", + "bbox": [ + 78, + 306, + 475, + 373 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "obtained by our method show encouraging generalization power, which achieves lower NME than the baselines in the unseen domain (i.e., ArtiFace). Overall, the NME of our method in the source domain (i.e., 300W) is comparable to that achieved in the oracle setting. For the target even unseen domains, our method reduces the gap to the oracle.", + "bbox": [ + 75, + 401, + 468, + 491 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Besides the numerical comparisons, we provide some visualization results obtained by different methods in Figs. 5 and 6. For a complete comparison, the results of the SLPT trained only on $300\\mathrm{W}$ are shown in the figures as well, which corresponds to learning a face landmarker only based on the data in the source domain. We can find that by applying our learning method, the face landmarker obtains better landmarks, which aligns with the ground truth with smaller errors, particularly for the landmarks of face contour, nose, mouth, and eyebrow.", + "bbox": [ + 75, + 492, + 468, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further verify the generalization of our method, we select three more state-of-the-art landmarks, including OpenPose (OP) [7], SPIGA [36] and STAR [57] for numerical comparisons, and incorporate STAR as the backbone model of our method. Tab. 4 demonstrates that our learning method is applicable for various backbone models (e.g., STAR and SLPT), and integrating existing landmarkers with our learning method is an effective and competitive approach to enhance their generalizability.", + "bbox": [ + 75, + 643, + 470, + 781 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Analytic Experiments", + "text_level": 1, + "bbox": [ + 76, + 791, + 281, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.1 Joint v.s. Alternating Optimization", + "text_level": 1, + "bbox": [ + 76, + 815, + 374, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As aforementioned, learning the face landmarker and the warping field model jointly makes it much easier to fall into undesired local optimum or unstable saddle points. To verify this claim, we apply the joint optimization strategy, i.e.,", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "solving (3) by optimizing $\\theta$ and $\\gamma$ jointly in each gradient descent step and compare its performance with ours. In Fig. 7, we show the best NME achieved by this joint optimization strategy and the NME achieved by our alternating optimization method in each iteration. We can find that the joint optimization strategy tends to overfit the source domain (300W), leading to power generalizability in the target and unseen domains. On the contrary, with the increase of iterations, the NMEs of our method on the target and unseen domains decrease consistently and become lower than those of the joint optimization after several iterations.", + "bbox": [ + 496, + 282, + 893, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.2 Impacts of Model Architectures", + "text_level": 1, + "bbox": [ + 498, + 469, + 774, + 484 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Besides the optimization strategy, the model architecture also has an impact on the model performance. By default, we implement the face landmarker as the SLPT model. In this experiment, we further explore the performance of other model architectures, including SBR [9] and HRNet [47]. Tab. 3 shows the performance of different model architectures in DA and GZSL settings. We can find that the SLPT-based face landmarker works better than the two competitors in this experiment. A potential reason for this phenomenon is that both SBR and HRNet are heatmap-based face landmarks. Given an input facial image, they output a heatmap indicating the distribution of landmarks rather than a set of deterministic landmark coordinates. We have to first detect the landmarks from the heatmap and then pass them through the warping field model. Accordingly, in the face warper optimization step (i.e., solving (4)), the landmarker and the warping field model cannot be trained in an end-to-end way because the backpropagation of the gradient becomes inapplicable. As a result, we have to update $\\theta$ and $\\gamma$ alternatively, leading to suboptimal performance.", + "bbox": [ + 496, + 493, + 890, + 796 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.3 Effects on Loss Term", + "text_level": 1, + "bbox": [ + 500, + 815, + 700, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the context of our model's loss function, the image gradient field serves as its input. It has been observed through experiments that utilizing distinct gradient operators leads to the generation of varying image gradient fields, which in", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2431", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a9c9c173fdf6deaeb922089993299153750deb9a9fed51bc0848693c5a29d00b.jpg", + "table_caption": [ + "Table 5. Quantitative comparison on CariFace dataset." + ], + "table_footnote": [], + "table_body": "
OperatorGraySpatialLaplacianCannySobel
NME✓8.2827.8708.0057.695
×7.9587.9218.0117.863
", + "bbox": [ + 84, + 114, + 459, + 172 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2803b262f27e6126b63dbd287c4fce76ec888e312cb29274d9154d6bc7a6ed5b.jpg", + "table_caption": [ + "Table 6. Comparisons for different losses on NME." + ], + "table_footnote": [], + "table_body": "
Type of LossMSEPerceptualw.o. Grad MSEGrad MSE
300WCommon3.082.892.962.79
Challenge5.435.135.004.91
Full3.543.333.363.20
DA on Caricature8.268.239.487.70
GZSL on ArtiFace4.124.074.063.95
", + "bbox": [ + 78, + 214, + 473, + 306 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "turn influences the model's performance. Consequently, in this section, we assess and compare the training outcomes derived from image gradient fields obtained through the application of diverse gradient operators.", + "bbox": [ + 75, + 335, + 468, + 396 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Moreover, it is crucial to recognize that the choice of the input image, specifically whether it is a grayscale image or not, can also exert a certain influence on the results. As a result, we have incorporated a comparison of the outcomes when the input is a grayscale image. It is important to emphasize that all experiment settings are maintained consistently. The results are shown in Tab. 5. Utilizing grayscale images yields superior effects compared to not using grayscale images. This could be attributed to the fact that converting an image to grayscale eliminates the trivial effects of color and complex textures. The most optimal outcome is achieved by employing the Sobel operator. This may be due to the Sobel operator's emphasis on edge information within the image, enabling it to capture more geometric details. Additionally, the Sobel operator incorporates a mild smoothing effect during gradient computation, which mitigates the impact of noise.", + "bbox": [ + 75, + 397, + 468, + 654 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Furthermore, we consider $i)$ replacing the gradient MSE loss in (3) with the pixel MSE or perceptual loss and $ii)$ removing the gradient MSE loss. Tab. 6 shows the rationality of the gradient MSE for the reason that $i)$ landmarks are distributed on edges; and $ii)$ the gradient field filters out unnecessary color information, simplifying the task. These findings form the basis for our decision to adopt this particular method in our experimental process.", + "bbox": [ + 75, + 654, + 468, + 776 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4.4 Rationality of Proposed Warping Field Model", + "text_level": 1, + "bbox": [ + 76, + 799, + 449, + 814 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Besides the model architecture of the face landmarker, the warping field model impacts our face landmarker as well — when the predicted warping field is inaccurate, we cannot obtain reliable pseudo landmarks for stylized faces. Therefore, we investigate different face warping models and", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/359e98d3068b708fe1b84baf873d783d7b880e9df8b874149cb2d4625443f3b2.jpg", + "image_caption": [ + "(a) Source" + ], + "image_footnote": [], + "bbox": [ + 511, + 88, + 627, + 176 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/15b97daac31685b24a66ea760ece96cbd351e3af902952727e7c166893c970ee.jpg", + "image_caption": [ + "(b) Target" + ], + "image_footnote": [], + "bbox": [ + 638, + 88, + 754, + 176 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0a715dfe02da865a1f9186db1f0443eee7771f7f8d602fe6cefb31b33ec90609.jpg", + "image_caption": [ + "(c) AutoToon" + ], + "image_footnote": [], + "bbox": [ + 766, + 88, + 880, + 176 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/54496ec481a0673426a460b7ff20c23f36bb5240a5712310e4d7b56e83c2b078.jpg", + "image_caption": [ + "(d) CariGANs" + ], + "image_footnote": [], + "bbox": [ + 511, + 191, + 625, + 279 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cf322da2ddab6d762f1564a2e626fad3117d8c694b3b1cb1a5e04d48d543e9bc.jpg", + "image_caption": [ + "(e) RAFT" + ], + "image_footnote": [], + "bbox": [ + 638, + 191, + 753, + 279 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/494bc94405d2d2241ae6514047ff04b26435dba3e8593177a6c5a0a6ed87df98.jpg", + "image_caption": [ + "(f) Ours", + "Figure 8. Comparisons for different face warpers." + ], + "image_footnote": [], + "bbox": [ + 766, + 191, + 880, + 279 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "demonstrate the rationality of the warping field model implemented in our work. In particular, applying different warping field models in our training process, we visualize their warping results in Fig. 8. We can find that although the face landmarker with the polyharmonic interpolation model leads to a very simple face warper, it outperforms many existing neural network-based image warping methods, such as AutoToon [15] and CariGANs [5] for facial manipulation and the optical flow methods like RAFT [44]. These methods either require one-to-one correspondence information between source and target images or assume the deformation between the two images to be slight, making them unsuitable for our problem, especially for the stylized faces with significant nonrigid deformations.", + "bbox": [ + 496, + 348, + 890, + 560 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 575, + 619, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a simple but effective method for learning a generalizable face landmarker applicable to facial images with different styles. Given labeled real human faces and unlabeled stylized faces, our method learns the face landmarker under the guidance of conditional face warping, demonstrating the usefulness of the warping information. An alternating optimization framework is proposed to learn the face landmarker together with the warping field model. Experiments demonstrate the effectiveness of our method. Especially in the generalized zero-shot learning scenarios, our method achieves encouraging landmarking accuracy in unseen face domains.", + "bbox": [ + 496, + 601, + 890, + 782 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 799, + 668, + 815 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by the National Natural Science Foundation of China (62102031), the foundation of Key Laboratory of Artificial Intelligence, Ministry of Education, P.R. China, and the Young Scholar Program (XSQD-202107001) from Beijing Institute of Technology.", + "bbox": [ + 496, + 824, + 890, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2432", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Olalekan Agbolade, Azree Nazri, Razali Yaakob, Abdul Azim Abd Ghani, and Yoke Kqueen Cheah. Landmark-based homologous multi-point warping approach to 3d facial recognition using multiple datasets. PeerJ Computer Science, 6:e249, 2020. 1", + "[2] Sadegh Aliakbarian, Pashmina Cameron, Federica Bogo, Andrew Fitzgibbon, and Thomas J. Cashman. Flag: Flow-based 3d avatar generation from sparse observations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13253-13262, 2022. 1", + "[3] Peter N. Belhumeur, David W. Jacobs, David J. Kriegman, and Neeraj Kumar. Localizing parts of faces using a consensus of exemplars. In CVPR 2011, 2011. 4", + "[4] Hongrui Cai, Yudong Guo, Zhuang Peng, and Juyong Zhang. Landmark detection and 3d face reconstruction for caricature using a nonlinear parametric model, 2020. 1, 4, 5", + "[5] Kaidi Cao, Jing Liao, and Lu Yuan. Carigans: Unpaired photo-to-caricature translation. arXiv preprint arXiv:1811.00222, 2018. 1, 8", + "[6] Xudong Cao, Yichen Wei, Fang Wen, and Jian Sun. Face alignment by explicit shape regression. International journal of computer vision, 107:177-190, 2014. 2", + "[7] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7291-7299, 2017. 7", + "[8] Forrester Cole, David Belanger, Dilip Krishnan, Aaron Sarna, Inbar Mosseri, and William T Freeman. Synthesizing normalized faces from facial identity features. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3703-3712, 2017. 3", + "[9] Xuanyi Dong, Shouu-I Yu, Xinshuo Weng, Shih-En Wei, Yi Yang, and Yaser Sheikh. Supervision-by-registration: An unsupervised approach to improve the precision of facial landmark detectors. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 360–368, 2018. 2, 5, 6, 7", + "[10] Alexey Dosovitskiy, Philipp Fischer, Eddy Ilg, Philip Hausser, Caner Hazirbas, Vladimir Golkov, Patrick Van Der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 2758-2766, 2015. 3", + "[11] Pengfei Dou, Shishir K Shah, and Ioannis A Kakadiaris. End-to-end 3d face reconstruction with deep neural networks. In proceedings of the IEEE conference on computer vision and pattern recognition, pages 5908-5917, 2017. 1", + "[12] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In International conference on machine learning, pages 1180-1189. PMLR, 2015. 1, 2, 5, 6", + "[13] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. A neural algorithm of artistic style. arXiv preprint arXiv:1508.06576, 2015. 3" + ], + "bbox": [ + 76, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] C. A. Glasbey and K. V. Mardia. A review of image-warping methods. Journal of Applied Statistics, 25(2):155-171, 2002. 3", + "[15] Julia Gong, Yannick Hold-Geoffroy, and Jingwan Lu. Auto-toon: Automatic geometric warping for face cartoon generation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 360–369, 2020. 1, 3, 8", + "[16] Arthur Gretton, Karsten Borgwardt, Malte Rasch, Bernhard Schölkopf, and Alex Smola. A kernel method for the two-sample-problem. Advances in neural information processing systems, 19, 2006. 2", + "[17] Guoliang Kang, Lu Jiang, Yi Yang, and Alexander G Hauptmann. Contrastive adaptation network for unsupervised domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4893-4902, 2019. 2", + "[18] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 3", + "[19] Vahid Kazemi and Josephine Sullivan. One millisecond face alignment with an ensemble of regression trees. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1867-1874, 2014. 1", + "[20] SunnieS. Kim, NicholasI. Kolkin, Jason Salavon, and Gregory Shakhnarovich. Deformable style transfer, 2020. 2", + "[21] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 4", + "[22] Marek Kowalski, Jacek Naruniec, and Tomasz Trzcinski. Deep alignment network: A convolutional neural network for robust face alignment. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2017. 2", + "[23] Vuong Le, Jonathan Brandt, Zhe Lin, Lubomir Bourdev, and Thomas S. Huang. Interactive Facial Feature Localization, page 679-692. 2012. 4", + "[24] Hui Li, Zidong Guo, Seon-Min Rhee, Seungju Han, and Jae-Joon Han. Towards accurate facial landmark detection via cascaded transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4176-4185, 2022. 2", + "[25] Yunsheng Li, Lu Yuan, and Nuno Vasconcelos. Bidirectional learning for domain adaptation of semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6936-6945, 2019. 1, 5, 6", + "[26] Feng Liu, Dan Zeng, Qijun Zhao, and Xiaoming Liu. Joint face alignment and 3d face reconstruction. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14, pages 545-560. Springer, 2016. 1", + "[27] Shaolei Liu, Xiaoyuan Luo, Kexue Fu, Manning Wang, and Zhijian Song. A learnable self-supervised task for unsupervised domain adaptation on point cloud classification and segmentation. Frontiers of Computer Science, 17(6):176708, 2023. 1, 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2433", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Weiyang Liu, Yandong Wen, Zhiding Yu, Ming Li, Bhiksha Raj, and Le Song. Spherface: Deep hypersphere embedding for face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 212-220, 2017. 1", + "[29] Xiao-Chang Liu, Yong-Liang Yang, and Peter Hall. Learning to warp for style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3702-3711, 2021. 3", + "[30] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In International conference on machine learning, pages 97-105. PMLR, 2015. 2", + "[31] Mingsheng Long, Han Zhu, Jianmin Wang, and Michael I Jordan. Unsupervised domain adaptation with residual transfer networks. Advances in neural information processing systems, 29, 2016. 2", + "[32] Jiangjing Lv, Xiaohu Shao, Junliang Xing, Cheng Cheng, and Xi Zhou. A deep regression architecture with two-stage re-initialization for high performance facial landmark detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3317-3326, 2017. 2", + "[33] Iacopo Masi, Stephen Rawls, Gérard Medioni, and Prem Natarajan. Pose-aware face recognition in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4838-4846, 2016. 1", + "[34] Kieron Messer, Jiri Matas, Josef Kittler, Juergen Luettin, Gilbert Maitre, et al. Xm2vtsdb: The extended m2vts database. In Second international conference on audio and video-based biometric person authentication, pages 965-966. CiteSeer, 1999. 4", + "[35] Zhongyi Pei, Zhangjie Cao, Mingsheng Long, and Jianmin Wang. Multi-adversarial domain adaptation. In Thirty-second AAAI conference on artificial intelligence, 2018. 2", + "[36] Andrés Prados-Torreblanca, José M Buenaposada, and Luis Baumela. Shape preserving facial landmarks with graph attention networks. arXiv preprint arXiv:2210.07233, 2022. 7", + "[37] Joseph Roth, Yiying Tong, and Xiaoming Liu. Unconstrained 3d face reconstruction. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2606-2615, 2015. 1", + "[38] Christos Sagonas, Georgios Tzimiropoulos, Stefanos Zafeiriou, and Maja Pantic. 300 faces in-the-wild challenge: The first facial landmark localization challenge. In 2013 IEEE International Conference on Computer Vision Workshops, 2014. 1, 4, 5", + "[39] Ahmed Selim, Mohamed Elgharib, and Linda Doyle. Painting style transfer for head portraits using convolutional neural networks. ACM Transactions on Graphics (ToG), 35(4): 1-18, 2016. 3", + "[40] Yichun Shi, Debayan Deb, and Anil K Jain. Warpgan: Automatic caricature generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10762-10771, 2019. 3", + "[41] Baochen Sun, Jiashi Feng, and Kate Saenko. Correlation alignment for unsupervised domain adaptation. Domain" + ], + "bbox": [ + 78, + 90, + 468, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "adaptation in computer vision applications, pages 153-171, 2017. 2", + "[42] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 5", + "[43] Yaniv Taigman, Ming Yang, Marc'Aurelio Ranzato, and Lior Wolf. Deepface: Closing the gap to human-level performance in face verification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1701-1708, 2014. 1", + "[44] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 3, 8", + "[45] Yi-Hsuan Tsai, Wei-Chih Hung, Samuel Schulter, Kiyuh Yuk Sohn, Ming-Hsuan Yang, and Manmohan Chandraker. Learning to adapt structured output space for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7472-7481, 2018. 5, 6", + "[46] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2", + "[47] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 43(10):3349-3364, 2020. 2, 6, 7", + "[48] Xinyao Wang, Liefeng Bo, and Li Fuxin. Adaptive wing loss for robust face alignment via heatmap regression. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 2", + "[49] Kunhong Wu, Fan Jia, and Yahong Han. Domain-specific feature elimination: multi-source domain adaptation for image classification. Frontiers of Computer Science, 17(4): 174705, 2023. 2", + "[50] Ruizheng Wu, Xiaodong Gu, Xin Tao, Xiaoyong Shen, Yu-Wing Tai, et al. Landmark assisted cyclegan for cartoon face generation. arXiv preprint arXiv:1907.01424, 2019. 3", + "[51] Wayne Wu, Chen Qian, Shuo Yang, Quan Wang, Yici Cai, and Qiang Zhou. Look at boundary: A boundary-aware face alignment algorithm. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2129-2138, 2018. 1", + "[52] Jiahao Xia, Weiwei Qu, Wenjian Huang, Jianguo Zhang, Xi Wang, and Min Xu. Sparse local patch transformer for robust face alignment and landmarks inherent relation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4052-4061, 2022. 1, 3, 5, 6", + "[53] Jiaolong Yang, Peiran Ren, Dongqing Zhang, Dong Chen, Fang Wen, Hongdong Li, and Gang Hua. Neural aggregation network for video face recognition. In Proceedings of" + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2434", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE conference on computer vision and pattern recognition, pages 4362-4371, 2017. 1", + "[54] Yanchao Yang and Stefano Soatto. Fda: Fourier domain adaptation for semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4085-4095, 2020. 5, 6", + "[55] Jordan Yaniv, Yael Newman, and Ariel Shamir. The face of art: landmark detection and geometric style in portraits. ACM Transactions on graphics (TOG), 38(4):1-15, 2019. 1, 2, 5", + "[56] Jing Zhang, Zewei Ding, Wanqing Li, and Philip Ogunbona. Importance weighted adversarial nets for partial domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8156-8164, 2018. 2", + "[57] Zhenglin Zhou, Huaxia Li, Hong Liu, Nanyang Wang, Gang Yu, and Rongrong Ji. Star loss: Reducing semantic ambiguity in facial landmark detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15475-15484, 2023. 7", + "[58] Zhou, Erjin and Fan, Haoqiang and Cao, Zhimin and Jiang, Yuning and Yin, Qi. Extensive Facial Landmark Localization with Coarse-to-Fine Convolutional Network Cascade. In Proceedings of the IEEE International Conference on Computer Vision (ICCV) Workshops, 2013. 1, 2", + "[59] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 1, 2, 5, 6", + "[60] Xiangxin Zhu and Deva Ramanan. Face detection, pose estimation, and landmark localization in the wild. In 2012 IEEE conference on computer vision and pattern recognition, pages 2879-2886. IEEE, 2012. 4", + "[61] Yi Zhu, Xindong Wu, Jipeng Qiang, Yunhao Yuan, and Yun Li. Representation learning via an integrated autoencoder for unsupervised domain adaptation. Frontiers of Computer Science, 17(5):175334, 2023. 1" + ], + "bbox": [ + 78, + 92, + 470, + 627 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "2435", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Generalizable Face Landmarking Guided by Conditional Face Warping/33b95235-e8cb-4b6b-a79e-5de2b0524965_model.json b/2024/Generalizable Face Landmarking Guided by Conditional Face Warping/33b95235-e8cb-4b6b-a79e-5de2b0524965_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b14cd299de0ca2cd6e7c11db259edc21257a7f8b --- /dev/null +++ b/2024/Generalizable Face Landmarking Guided by Conditional Face Warping/33b95235-e8cb-4b6b-a79e-5de2b0524965_model.json @@ -0,0 +1,2752 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.124, + 0.131, + 0.848, + 0.154 + ], + "angle": 0, + "content": "Generalizable Face Landmarking Guided by Conditional Face Warping" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.18, + 0.733, + 0.199 + ], + "angle": 0, + "content": "Jiayi Liang\\(^{1}\\) Haotian Liu\\(^{1*}\\) Hongteng Xu\\(^{2,3}\\) Dixin Luo\\(^{1,4\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.144, + 0.199, + 0.829, + 0.218 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Computer Science and Technology, Beijing Institute of Technology, Beijing" + }, + { + "type": "text", + "bbox": [ + 0.173, + 0.217, + 0.798, + 0.235 + ], + "angle": 0, + "content": "\\(^{2}\\)Gaoling School of Artificial Intelligence, Renmin University of China, Beijing" + }, + { + "type": "text", + "bbox": [ + 0.163, + 0.234, + 0.808, + 0.252 + ], + "angle": 0, + "content": "3Beijing Key Laboratory of Big Data Management and Analysis Methods, Beijing" + }, + { + "type": "text", + "bbox": [ + 0.188, + 0.252, + 0.781, + 0.27 + ], + "angle": 0, + "content": "\\(^{4}\\)Key Laboratory of Artificial Intelligence, Ministry of Education, Shanghai" + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.272, + 0.79, + 0.288 + ], + "angle": 0, + "content": "{jiayi.liang, haotianliu, dixin.luo}@bit.edu.cn, hongtengxu@ruc.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.322, + 0.314, + 0.338 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.355, + 0.474, + 0.749 + ], + "angle": 0, + "content": "As a significant step for human face modeling, editing, and generation, face landmarking aims at extracting facial keypoints from images. A generalizable face landmarker is required in practice because real-world facial images, e.g., the avatars in animations and games, are often stylized in various ways. However, achieving generalizable face landmarking is challenging due to the diversity of facial styles and the scarcity of labeled stylized faces. In this study, we propose a simple but effective paradigm to learn a generalizable face landmarker based on labeled real human faces and unlabeled stylized faces. Our method learns the face landmarker as the key module of a conditional face warper. Given a pair of real and stylized facial images, the conditional face warper predicts a warping field from the real face to the stylized one, in which the face landmarker predicts the ending points of the warping field and provides us with high-quality pseudo landmarks for the corresponding stylized facial images. Applying an alternating optimization strategy, we learn the face landmarker to minimize i) the discrepancy between the stylized faces and the warped real ones and ii) the prediction errors of both real and pseudo landmarks. Experiments on various datasets show that our method outperforms existing state-of-the-art domain adaptation methods in face landmarking tasks, leading to a face landmarker with better generalizability. Code is available at https://plustwo0.github.io/project-face-landmarker." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.777, + 0.21, + 0.793 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.803, + 0.47, + 0.866 + ], + "angle": 0, + "content": "Face landmarking seeks to extract human facial keypoints (e.g., eyes, nose, facial contour, and so on) from facial images. This task is important for many applications in the field of computer vision and graphics, such as face recogni" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.324, + 0.878, + 0.532 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.54, + 0.895, + 0.64 + ], + "angle": 0, + "content": "Figure 1. Both commercial software like Face++ and open-source method like SLPT [52] work well on landmarking real faces (e.g., those in 300W [38]) while achieving suboptimal performance when landmarking stylized faces (e.g., those in CariFace [4] and ArtiFace [55]). While existing domain adaptation method does not improve the performance significantly, our method achieves a generalizable face landmarker for various facial images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.671, + 0.893, + 0.734 + ], + "angle": 0, + "content": "tion [1, 28, 33, 43, 53], face stylization [5], and 3D face reconstruction [11, 26, 37]. Currently, many open-source and commercial face landmarks [19, 51, 58] have been developed and achieved encouraging performance in this task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Most existing face landmarks are designed and trained for landmarking real human faces, while the rapid development of AIGC applications, such as artistic character creation and cartoon generation [2, 15], leads to a massive increase in demands for landmarking stylized facial images. Unfortunately, as shown in Fig. 1, existing face landmarks often fail to landmark stylized facial images. Even if applying state-of-the-art domain adaptation strategies [12, 25, 27, 59, 61], the generalizability of the learned landmarks in the stylized facial image domain is still unsatisfactory. Essentially, traditional face landmarks work" + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.875, + 0.386, + 0.888 + ], + "angle": 0, + "content": "*The first two authors contributed equally to this work." + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.888, + 0.222, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.095, + 0.875, + 0.386, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2425" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.092, + 0.468, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.345, + 0.47, + 0.375 + ], + "angle": 0, + "content": "Figure 2. The scheme of our proposed method for learning a generalizable face landmarker." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.402, + 0.47, + 0.522 + ], + "angle": 0, + "content": "well on real human faces because of the relatively stable geometry of real human faces and the sufficient labeled facial images. These two conditions, however, become questionable when landmarking stylized faces — the stylized facial images often have various facial styles, and manually landmarking such stylized faces is much more time-consuming than landmarking real faces. As a result, learning a generalizable face landmarker becomes a challenging task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.523, + 0.471, + 0.854 + ], + "angle": 0, + "content": "In this paper, we propose a simple but effective paradigm for learning a generalizable face landmarker, overcoming the challenges caused by the diversity of facial styles and the scarcity of labeled stylized faces. As illustrated in Fig. 2, given labeled real human facial images and unlabeled stylized facial images, we learn a face landmarker embedded in a conditional face warper. The face warper aims to deform real human faces according to stylized facial images, generating warped faces and corresponding warping fields. The face landmarker, as the key module of the warper, predicts the ending points of the warping fields and thus provides us with pseudo landmarks for the stylized facial images. The warping field is parametrized by a polyharmonic interpolation model. Under the guidance of the conditional face warping, we learn the face landmarker in an alternating optimization framework: The face landmarker is updated to \\( i \\) minimize the discrepancy between the stylized faces and the corresponding warped real human faces and \\( ii \\) minimize the prediction errors of both real and pseudo landmarks. In the first step, the face landmarker is learned associated with the warping field model, while in the second step, the face landmarker is updated with proximal regularization." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "The extensive experiments in various face landmarking tasks demonstrate the effectiveness of learning method. The impacts of different loss functions and data settings on" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.184 + ], + "angle": 0, + "content": "our learning method are analyzed through detailed ablation studies. Experimental results show that our learning method results in a face landmarker that is generalizable to the facial images with different styles, which outperforms representative domain adaptation methods consistently in various stylized face landmarking tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.198, + 0.642, + 0.213 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.223, + 0.685, + 0.24 + ], + "angle": 0, + "content": "2.1. Face Landmarking" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.246, + 0.892, + 0.533 + ], + "angle": 0, + "content": "Given facial images, the early face landmarking methods learn regression models to predict landmark coordinates directly [6, 32]. The models are often parameterized by neural networks like Transformer [46], capturing facial attributes that have been proven to be crucial for landmark prediction [24]. The coordinate regression is achieved in a coarse-to-fine framework [58], leading to a cascading landmarking pipeline. Recently, to make the landmark prediction robust to the variations in pose, scale, and occlusion, DAN [22] introduces the novel use of heatmaps and extracts features from the entire face rather than local patches around landmarks. SBR [9] utilizes the registration of synthesized images to provide supervisory signals for training. Adaptive Wing Loss [48] is proposed to address the imbalance between foreground pixels and background pixels by analysis of the main drawbacks of different loss functions. HRNet [47] produces high-resolution maps by connecting and exchanging information via merging multi-scale picture features across many branches." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.892, + 0.777 + ], + "angle": 0, + "content": "As aforementioned, most existing face landmarks are learned for real human faces, which can not be adapted directly to stylized faces (e.g., cartoon and artistic faces). Essentially, treating labeled real human faces as a source domain and unlabeled stylized faces as a target domain, we can learn a generalizable face landmark by solving a domain adaptation (DA) problem [30, 31]. Accordingly, many DA techniques [27, 49] have potential in our problem, including the classic metric learning-based methods (e.g., CORAL [41], contrastive domain discrepancy [17], and maximum mean discrepancy [16]) and the recent adversarial learning-based methods [12, 35, 56, 59]. However, in the following content, we will show that directly applying these DA techniques often fails to achieve generalizable face landmarking because of the significant gap between the source and target face domains." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.646, + 0.803 + ], + "angle": 0, + "content": "2.2. Face Warping" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Face warping is a technique that involves geometrically deforming source facial images to specified target shapes. The key step of this task is predicting a warping field between the source and target images that captures the shifts of image pixels. To achieve this aim, DST [20] and FoA [55] find matching keypoints between source and target images" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2426" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.211 + ], + "angle": 0, + "content": "and then generate a dense warping field through data interpolation [8]. Instead of matching keypoints, some methods learn neural networks to predict dense warping fields directly based on paired images, e.g., Flownet [10], AutoToon [15], RAFT [44], and their variants [29]. However, these methods require the paired images to be similar to each other, which cannot capture significant deformations between the faces in different domains." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.212, + 0.471, + 0.334 + ], + "angle": 0, + "content": "Compared to face stylization [13, 18, 39, 40], face warping is a relatively easier task because it only considers the deformation of shapes while ignoring the transfer of textures. However, it should be noted that this task is more relevant to face landmarking, in which the warping field provides us with strong evidence to shift face landmarks of source faces to target ones [50]. Inspired by such a strong correlation, we develop the proposed learning paradigm." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.345, + 0.251, + 0.362 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.371, + 0.47, + 0.603 + ], + "angle": 0, + "content": "Denote \\(\\mathcal{X}\\) as the image space and \\(\\mathcal{V}\\) as the landmark space, respectively. In this work, we observe a set of labeled real human faces, i.e., \\(\\mathcal{D}^{(L)} = \\{\\pmb{X}_i^{(L)},\\pmb{Y}_i^{(L)}\\}_{i = 1}^{N_L}\\subset \\mathcal{X}_R\\times \\mathcal{V}\\) and a set of unlabeled stylized faces, i.e., \\(\\mathcal{D}^{(U)} = \\{\\pmb{X}_i^{(U)}\\}_{i = 1}^{N_U}\\subset\\) \\(\\mathcal{X}_S\\), where \\(\\mathcal{X}_R,\\mathcal{X}_S\\subset \\mathcal{X}\\) correspond to the real and stylized face domains, respectively. Each \\(\\pmb {X}_i\\in \\mathbb{R}^{H\\times W\\times 3}\\) represents an image, and each \\(\\pmb {Y}_i = [\\pmb {y}_{i,k}]\\in \\mathbb{R}^{2\\times K}\\) records \\(K\\) face landmark coordinates, where \\(\\pmb {y}_{i,k}\\in \\mathbb{R}^2\\). We aim to learn a face landmarker, denoted as \\(f_{\\theta}:\\mathcal{X}\\mapsto \\mathcal{V}\\), where \\(\\theta\\) is the model parameter. The model should be able to predict face landmarks from facial images and moreover, generalize to both \\(\\mathcal{X}_R\\) and \\(\\mathcal{X}_S\\). To achieve this aim, we embed the face landmarker into a conditional face warper and learn it associated with a parametric warping field predictor in an alternating optimization framework, as illustrated in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.61, + 0.458, + 0.627 + ], + "angle": 0, + "content": "3.1. Face Landmarking Guided by Face Warping" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.633, + 0.468, + 0.767 + ], + "angle": 0, + "content": "In this study, we take the SLPT model [52] as the backbone of our face landmarker. Given a stylized face \\( \\mathbf{X}_i^{(U)} \\), the face landmarker predicts its landmarks as \\( \\widehat{\\mathbf{Y}}_i^{(U)} = [\\widehat{\\mathbf{y}}_{i,k}^{(U)}] = f_\\theta(\\mathbf{X}_i^{(U)}) \\). At the same time, we can sample a labeled real face \\( (\\mathbf{X}_j^{(L)},\\mathbf{Y}_j^{(L)}) \\sim \\mathcal{D}^L \\). Treating the labeled and predicted landmarks as keypoints, we can model a warping field from the real face to the stylized one by the following polyharmonic interpolation model [14]:" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.773, + 0.47, + 0.801 + ], + "angle": 0, + "content": "\\[\nw _ {i, \\gamma} (\\boldsymbol {y}) = \\sum_ {k = 1} ^ {K} \\omega_ {k} \\phi \\left(\\left\\| \\boldsymbol {y} - \\hat {\\boldsymbol {y}} _ {i, k} ^ {(U)} \\right\\| _ {2}\\right) + \\boldsymbol {V} \\boldsymbol {y} + \\boldsymbol {b}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.807, + 0.471, + 0.903 + ], + "angle": 0, + "content": "where \\(\\gamma = \\{\\{\\pmb{\\omega}_k\\in \\mathbb{R}^2\\}_{k = 1}^K,\\pmb {V}\\in \\mathbb{R}^{2\\times 2},\\pmb {b}\\in \\mathbb{R}^2\\}\\) correspond to the parameters of the warping field. As shown in (1), the vector \\(\\pmb{y}\\) denotes the u-v coordinate of a pixel in the stylized facial image, and \\(w_{i,\\gamma}(\\pmb {y})\\) gives the inverse mapping from the pixel \\(\\pmb{y}\\) to a coordinate in the real human facial image, conditioned on \\(\\pmb{X}_i^{(U)}\\). The first term" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.168 + ], + "angle": 0, + "content": "\\(\\sum_{k=1}^{K} \\omega_k \\phi(\\| \\pmb{y} - \\hat{\\pmb{y}}_{i,k}^{(U)} \\|_2)\\) achieves nonparametric regression for modeling nonrigid deformations, in which \\(\\phi(r)\\) is a predefined thin-plate spline function. The second term \\(\\pmb{V}\\pmb{y} + \\pmb{b}\\) is a linear parametric model capturing the rigid transformation of \\(\\pmb{y}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.168, + 0.892, + 0.282 + ], + "angle": 0, + "content": "For each pixel coordinate \\(\\pmb{y} \\in \\{1, \\dots, H\\} \\times \\{1, \\dots, W\\}\\), we can trace it back to the real human facial image based on \\(w_{i,\\gamma}(\\pmb{y})\\) and obtain the pixel color as \\(\\pmb{X}_j^{(L)}(w_{i,\\gamma}(\\pmb{y}))\\). Accordingly, with the grid sampler constructed via inverse mapping function \\(w_{i,\\gamma}\\), we obtain the warped real human facial image conditioned on \\(\\pmb{X}_i^{(U)}\\), denoted as \\(\\widehat{\\pmb{X}}_{j|i}^{(L)}\\). For \\(\\pmb{y} \\in \\{1, \\dots, H\\} \\times \\{1, \\dots, W\\}\\) and \\(j = 1, \\dots, N_L\\), we have" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.287, + 0.892, + 0.311 + ], + "angle": 0, + "content": "\\[\n\\widehat {\\boldsymbol {X}} _ {j \\mid i} ^ {(L)} (\\boldsymbol {y}) = \\boldsymbol {X} _ {j} ^ {(L)} \\left(w _ {i, \\gamma} (\\boldsymbol {y})\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.316, + 0.892, + 0.467 + ], + "angle": 0, + "content": "Unlike WarpGAN [40], which generates the warped face by predicting dense keypoints and their displacements by two fully-connected layers during training, we directly use the predicted and observed landmarks to define sparse displacements and estimate other pixels' displacements by spline-based interpolation, which improves computational efficiency significantly. Moreover, by applying the warping field model with limited degree-of-freedom (i.e., few learnable parameters), we can focus more on the learning of the face landmarker in the training phase." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.468, + 0.892, + 0.527 + ], + "angle": 0, + "content": "Specifically, the warped face together with the warping field provides a useful guidance for the learning of the face landmarker. In particular, we formulate the learning problem of the face landmarker as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.532, + 0.891, + 0.681 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min _ {\\theta , \\gamma} \\underbrace {\\sum_ {j = 1} ^ {N _ {L}} \\| f _ {\\theta} \\left(\\boldsymbol {X} _ {j} ^ {(L)}\\right) - \\boldsymbol {Y} _ {j} ^ {(L)} \\| _ {F} ^ {2}} _ {\\text {L a n d m a r k i n g e r r o r i n t h e s o u r c e d o m a i n}} \\\\ + \\underbrace {\\sum_ {i = 1} ^ {N _ {U}} \\sum_ {j = 1} ^ {N _ {L}} \\left\\| \\nabla \\widehat {\\boldsymbol {X}} _ {j | i} ^ {(L)} - \\nabla \\boldsymbol {X} _ {i} ^ {(U)} \\right\\| _ {F} ^ {2}} _ {\\text {D i s c r e p a n c y o f i m a g e g r a d i e n t}} \\tag {3} \\\\ + \\underbrace {\\sum_ {i = 1} ^ {N _ {U}} \\sum_ {j = 1} ^ {N _ {L}} \\| w _ {i , \\gamma} (\\widehat {\\mathbf {Y}} _ {i} ^ {(U)}) - \\mathbf {Y} _ {j} ^ {(L)} \\| _ {F} ^ {2}} _ {\\text {L a n d m a r k w a r p i n g e r r o r}}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.688, + 0.892, + 0.827 + ], + "angle": 0, + "content": "where \\(\\| \\cdot \\| _F\\) represents the Frobenius norm of matrix. In (3), the first term is the landmarking error for real faces, which corresponds to the data fidelity loss in the source domain. The second term measures the discrepancy between the stylized face and the warped real face in the gradient field, in which the gradient operation \\(\\nabla\\) is implemented by the Sobel operator. The third term is the landmark warping error. Both the second and third terms are determined jointly by the landmarker \\(f_{\\theta}\\) and warping field model \\(w_{i,\\gamma}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.833, + 0.802, + 0.849 + ], + "angle": 0, + "content": "3.2. Alternating Optimization Strategy" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.902 + ], + "angle": 0, + "content": "The optimization problem in (3) is non-convex because the landmarker is implemented by a neural network and is coupled to the warping field model. As a result, learning \\(\\theta\\) and" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2427" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.085, + 0.089, + 0.244, + 0.099 + ], + "angle": 0, + "content": "Cartoon Face Real Human Face" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.101, + 0.16, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.161, + 0.102, + 0.235, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.236, + 0.102, + 0.311, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.102, + 0.387, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.388, + 0.102, + 0.464, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.162, + 0.162, + 0.336, + 0.171 + ], + "angle": 0, + "content": "Warped Face with Warped Landmarks" + }, + { + "type": "image", + "bbox": [ + 0.16, + 0.175, + 0.235, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.236, + 0.175, + 0.311, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.175, + 0.387, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.388, + 0.175, + 0.464, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.245, + 0.47, + 0.303 + ], + "angle": 0, + "content": "Figure 3. Illustrations of conditional face warping results. Taking a cartoon face as the target, our model warps real human faces accordingly. The red dots indicate real human face landmarks, and green dots indicate cartoon and warped face landmarks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.328, + 0.47, + 0.404 + ], + "angle": 0, + "content": "\\(\\gamma\\) jointly often falls into an undesired local optimum even an unstable saddle point. To mitigate this issue, we propose an alternating optimization framework. In principle, we can decompose the optimization problem in (3) into the following two subproblems and solve them iteratively." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.406, + 0.469, + 0.437 + ], + "angle": 0, + "content": "- Face Warper Optimization: The first subproblem corresponds to the optimization of the face warper, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.445, + 0.478, + 0.5 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\theta^ {(1)}, \\gamma^ {(1)} = \\arg \\min _ {\\theta , \\gamma} \\sum_ {i, j} \\| \\nabla \\widehat {\\boldsymbol {X}} _ {j | i} ^ {(L)} - \\nabla \\boldsymbol {X} _ {i} ^ {(U)} \\| _ {F} ^ {2} \\tag {4} \\\\ + \\sum_ {i, j} \\| w _ {i, \\gamma} (\\widehat {\\pmb {Y}} _ {i} ^ {(U)}) - \\pmb {Y} _ {j} ^ {(L)} \\| _ {F} ^ {2}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.508, + 0.47, + 0.584 + ], + "angle": 0, + "content": "In this subproblem, we only care about whether the real human faces can be warped as the stylized faces with high accuracy, so the term of landmarking error is ignored. We solve this problem by Adam [21]: in each step, we update \\(\\theta\\) and \\(\\gamma\\) based on a batch of randomly-sampled face pairs." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.584, + 0.469, + 0.647 + ], + "angle": 0, + "content": "- Proximal Face Landmarker Optimization: Given \\(\\theta^{(1)}\\) and the predicted landmarks (i.e., \\(\\widehat{\\pmb{Y}}_i^{(U)} = f_{\\theta^{(1)}}(\\pmb{X}_i^{(U)})\\) for \\(i = 1, \\dots, N_L\\)), we can treat \\(\\theta^{(1)}\\) as the initial variable and optimize it with a proximal regularizer:" + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.656, + 0.469, + 0.72 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\theta^ {(2)} = \\arg \\min _ {\\theta} \\sum_ {j = 1} ^ {N _ {L}} \\| f _ {\\theta} (\\boldsymbol {X} _ {j} ^ {(L)}) - \\boldsymbol {Y} _ {j} ^ {(L)} \\| _ {F} ^ {2} \\\\ + \\underbrace {\\sum_ {i = 1} ^ {N _ {U}} \\| f _ {\\theta} \\left(\\boldsymbol {X} _ {i} ^ {(U)}\\right) - \\widehat {\\boldsymbol {Y}} _ {i} ^ {(U)} \\| _ {F} ^ {2}}. \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.186, + 0.722, + 0.403, + 0.733 + ], + "angle": 0, + "content": "Pseudo landmarking error in the target domain" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.743, + 0.47, + 0.839 + ], + "angle": 0, + "content": "Here, the second term in (5) measures the estimation errors of the pseudo landmarks achieved in the previous step. Essentially, it works as a proximal regularizer, ensuring that the optimized landmarks \\(f_{\\theta^{(2)}}(\\mathbf{X}_i^{(U)})\\) is not too far away from the previous estimation \\(\\widehat{\\mathbf{Y}}_i^{(U)}\\). Similarly, we can solve this problem by Adam [21] as well." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Fig. 3 shows the warping effect on real human faces achieved by solving (4). In Fig. 3, the first row shows the real human faces with landmarks and the target stylized face, and the second row shows the warping results," + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.091, + 0.892, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Proposed learning scheme of face landmarker" + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.113, + 0.892, + 0.158 + ], + "angle": 0, + "content": "Require: Labeled real faces \\(\\mathcal{D}^{(L)}\\) and unlabeled stylized faces \\(\\mathcal{D}^{(U)}\\). The number of iterations (i.e., \\(M\\)). Epochs for the subproblems (i.e., \\(L_{1}\\) and \\(L_{2}\\))." + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.158, + 0.892, + 0.189 + ], + "angle": 0, + "content": "1: Initialize \\(\\{\\gamma^{(0)}\\}\\) with a pretrained model on \\(\\mathcal{D}^{(L)}\\) and \\(\\{\\theta^{(0)}\\}\\) randomly." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.19, + 0.697, + 0.202 + ], + "angle": 0, + "content": "2: for \\( m = 0, \\dots, M - 1 \\) do" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.203, + 0.818, + 0.222 + ], + "angle": 0, + "content": "3: Sample a batch \\(\\{\\pmb{X}_i^{(U)},\\pmb{X}_i^{(L)},\\pmb{Y}_i^{(L)}\\}_{i = 1}^N\\)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.222, + 0.733, + 0.235 + ], + "angle": 0, + "content": "4: Face warper optimization:" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.236, + 0.892, + 0.266 + ], + "angle": 0, + "content": "5: Take \\(\\theta^{(2m)},\\gamma^{(m)}\\) as the initialization, then solve (4) by Adam with \\(L_{1}\\) epochs and obtain \\(\\theta^{(2m + 1)}\\)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.267, + 0.829, + 0.28 + ], + "angle": 0, + "content": "6: Proximal face landmarker optimization:" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.281, + 0.891, + 0.311 + ], + "angle": 0, + "content": "7: Take \\(\\theta^{(2m + 1)}\\) as the initialization, then solve (5) by Adam with \\(L_{2}\\) epochs and obtain \\(\\theta^{(2m + 2)}\\)." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.312, + 0.585, + 0.325 + ], + "angle": 0, + "content": "8: end for" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.327, + 0.891, + 0.343 + ], + "angle": 0, + "content": "9: return Output a generalizable face landmarker \\( f_{\\theta^{(2M)}} \\)." + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.158, + 0.892, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.373, + 0.892, + 0.555 + ], + "angle": 0, + "content": "in which the green dots are predicted landmarks. These results empirically demonstrate the rationality of our alternating optimization framework. In particular, we can find that solving (4) leads to reasonable warping results, which are similar to the target stylized face on shape. The similarity on face shape indicates that the predicted landmarks can be treated as reliable pseudo labels of the stylized face, which can be used to construct the proximal regularizer that penalizing the peso do landmarking errors in the target domain. Repeating the above two steps till converge, we obtain the target face landmarker that is generalizable for both real and stylized faces. Algorithm 1 shows the learning scheme." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.569, + 0.625, + 0.586 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.595, + 0.892, + 0.746 + ], + "angle": 0, + "content": "We apply our learning method to learn a face landmarker and test it on landmarking faces with various styles. Extensive experiments, including comparisons with baselines and analytic ablation studies, demonstrate the effectiveness of our learning method and the generalizability of the corresponding face landmarker. All the experiments are conducted on a single NVIDIA 3090 GPU. Representative experimental results are shown below. More experimental results and implementation details are given in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.756, + 0.596, + 0.771 + ], + "angle": 0, + "content": "4.1. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.89, + 0.81 + ], + "angle": 0, + "content": "In this study, we conduct experiments based on the following three commonly-used face datasets." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.811, + 0.892, + 0.856 + ], + "angle": 0, + "content": "- 300W Dataset. 300W [38] is comprised of five well-known real human face datasets including LFPW [3], AFW [60], HELEN [23], XM2VTS [34], and IBUG [38]." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.856, + 0.892, + 0.9 + ], + "angle": 0, + "content": "- CariFace Dataset. CariFace [4] is created by searching and selecting thousands of various caricatures from different celebrities on the Internet." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.811, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2428" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.321, + 0.09, + 0.647, + 0.104 + ], + "angle": 0, + "content": "Table 1. Data settings for the three learning paradigms." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.116, + 0.891, + 0.237 + ], + "angle": 0, + "content": "
Learning Paradigm#Training Images and Label Information#Testing Images
300WCariFaceArtiFaceCommon300W ChallengeFullCariFaceArtiFace
DA (300W→CariFace)3,148Labeled3,372Unlabeled——554135689800—
DA (300W→ArtiFace)3,148Labeled——128Unlabeled554135689—32
GZSL (Unseen ArtiFace)3,148Labeled3,372Unlabeled——554135689800160
GZSL (Unseen CariFace)3,148Labeled——128Unlabeled55413568980032
Oracle3,148Labeled3,372Labeled128Labeled55413568980032
" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.259, + 0.47, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.472, + 0.47, + 0.5 + ], + "angle": 0, + "content": "Figure 4. Illustrations of typical samples in the 300W, CariFace, and ArtiFace datasets, each of which is annotated with landmarks." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.529, + 0.47, + 0.574 + ], + "angle": 0, + "content": "- ArtiFace Dataset. ArtiFace [55] contains 160 artistic portraits of 16 artists, which covers diverse artwork styles ranging from Renaissance to Comics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.576, + 0.47, + 0.727 + ], + "angle": 0, + "content": "Each face in the datasets is annotated with 68 landmarks. Typical faces in the datasets and their landmarks are shown in Fig. 4. We can find that the faces in the three datasets have distinguished styles, which correspond to three different domains. In particular, compared to 300W [38], CariFace [4] exhibits abstract and exaggerated patterns, leading to large representation variations. ArtiFace [55] not only has larger variations across different artistic categories but also differs greatly in terms of the aspect of facial scales, orientations, locations, and so on." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.739, + 0.382, + 0.756 + ], + "angle": 0, + "content": "4.2. Learning Paradigms and Baselines" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.764, + 0.469, + 0.794 + ], + "angle": 0, + "content": "Given the above datasets, we consider the following three learning paradigms:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.796, + 0.469, + 0.854 + ], + "angle": 0, + "content": "- Domain Adaptation (DA). Given labeled 300W faces and unlabeled stylized faces from CariFace or ArtiFace, we learn a face landmarker based on various domain adaptation methods." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.856, + 0.469, + 0.9 + ], + "angle": 0, + "content": "- Generalized Zero-shot Learning (GZSL). In the challenging GZSL setting, we learn a face landmarker based on the above DA-based methods and test it in an unseen" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.796, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.263, + 0.892, + 0.293 + ], + "angle": 0, + "content": "face domain (e.g., learning the landmarker on labeled 300W and unlabeled CariFace and testing on ArtiFace)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.293, + 0.891, + 0.337 + ], + "angle": 0, + "content": "- Oracle. In this setting, the labeled faces of all three datasets are accessible, and we can learn the face landmark by classic supervised learning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.338, + 0.892, + 0.534 + ], + "angle": 0, + "content": "For a fair comparison, in each learning paradigm, we set the architecture of the face landmark based on the SLPT in [52]. Ideally, we would like to learn landmarks in the DA and GZSL settings, making its performance comparable to the oracle. In the oracle setting, we can learn the face landmark directly via classic supervised learning (SL), i.e., \\(\\min_{\\theta}\\sum_{(\\mathbf{X},\\mathbf{Y})\\sim \\mathcal{D}}\\| f_{\\theta}(\\mathbf{X}) - \\mathbf{Y}\\| _F^2\\). In the DA and GZSL settings, besides minimizing the landmark estimation errors, we can apply various image style transfer and domain adaptation methods, e.g., RevGrad [12], CycleGAN [59], BDL [25], AdaptSegNet [45] and FDA [54], to impose domain adaptation regularization during training. These methods work as the baselines of our method." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.892, + 0.656 + ], + "angle": 0, + "content": "Given the landmarkers learned by various methods, we evaluate them with the standard metric, Normalized Mean Error (NME). In Tab. 1, we show the training and testing data settings in the above three learning paradigms. Following existing work [9, 42, 52], we further split the 689 testing faces in 300W into 554 faces in common scenarios and 135 faces in challenging scenarios. The NMEs for the common, challenge, and full scenarios are recorded." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.666, + 0.809, + 0.683 + ], + "angle": 0, + "content": "4.3. Numerical and Visual Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In Tab. 2, we show the performance of various learning methods in different settings, demonstrating the effectiveness and superiority of our method. In particular, existing DA methods often fail to improve the generalization power of model a lot in face landmarking tasks — their performance in target and unseen domains is inferior to that in the oracle setting, with a significant gap on NME. A potential reason for this phenomenon is that these methods focus on the adaptation of image domain and the landmark-related loss is not dominant in their learning processes. As a result, instead of learning the face landmarker, they make more efforts to optimize the parameters of other modules (e.g., the neural network-based face stylization modules and discriminators) during training." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2429" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.158, + 0.089, + 0.812, + 0.104 + ], + "angle": 0, + "content": "Table 2. Comparisons for various methods on their NMEs. In DA and GZSL settings, the best results are bold." + }, + { + "type": "table", + "bbox": [ + 0.122, + 0.116, + 0.848, + 0.334 + ], + "angle": 0, + "content": "
Learning ParadigmLearning Method300WCariFaceArtiFaceAverage NME
CommonChallengeFull
DA (300W→CariFace) and GZSL (Unseen ArtiFace)SL+RevGrad [12]2.845.583.3812.195.167.83
SL+CycleGAN [59]2.745.433.2712.114.707.70
SL+BDL [25]3.286.173.8413.635.658.77
SL+ASN [45]2.925.263.3812.214.757.80
SL+FDA [54]2.895.183.3412.664.608.07
Ours2.794.913.207.703.955.46
DA (300W→ArtiFace) and GZSL (Unseen CariFace)SL+RevGrad [12]2.995.813.5512.464.748.26
SL+CycleGAN [59]3.005.653.5212.645.348.36
SL+BDL [25]2.995.323.4413.405.908.73
SL+ASN [45]2.895.813.4616.585.6510.31
SL+FDA [54]3.056.213.6812.335.878.27
Ours2.905.143.3410.933.937.34
OracleSL2.684.863.105.483.314.36
" + }, + { + "type": "table_caption", + "bbox": [ + 0.084, + 0.347, + 0.885, + 0.362 + ], + "angle": 0, + "content": "Table 3. The impacts of different model architectures on the NME performance of our method. In each setting, the best results are bold." + }, + { + "type": "table", + "bbox": [ + 0.122, + 0.374, + 0.848, + 0.495 + ], + "angle": 0, + "content": "
Learning ParadigmLearning Method300WCariFaceArtiFaceAverage NME
CommonChallengeFull
DA (300W→CariFace) and GZSL (Unseen ArtiFace)SBR [9]3.616.364.158.005.096.11
HRNet [47]2.935.293.407.734.335.60
SLPT [52] (Default)2.794.913.207.703.955.46
DA (300W→ArtiFace) and GZSL (Unseen CariFace)SBR [9]3.706.824.3211.355.328.04
HRNet [47]3.045.443.519.854.306.86
SLPT [52] (Default)2.905.143.3410.933.937.34
" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.516, + 0.472, + 0.761 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.771, + 0.471, + 0.827 + ], + "angle": 0, + "content": "Figure 5. Visual comparisons for various methods in the two DA settings. We only highlight points on the inner lips in the enlarged region of the mouth in (a), as well as the eyes and the sides of the cheeks, excluding points on the eyebrows in (b)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Different from the baselines, the performance of our model in the target domain (e.g., CariFace and ArtiFace) is improved consistently, while the performance in the source" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.516, + 0.895, + 0.761 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.772, + 0.894, + 0.829 + ], + "angle": 0, + "content": "Figure 6. Visual comparisons for various methods in the GZSL (Unseen ArtiFace) setting. The zoomed-in area in (a) highlights the noses and facial contours, while that in (b) concentrates on the upper part of faces." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.893, + 0.901 + ], + "angle": 0, + "content": "domain does not degrade a lot. In both DA and GZSL settings, our method outperforms the baselines in most situations. Especially in the GZSL settings, the landmarks" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2430" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.099, + 0.248, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.219, + 0.227, + 0.231 + ], + "angle": 0, + "content": "(a) 300W (Common)" + }, + { + "type": "image", + "bbox": [ + 0.257, + 0.099, + 0.404, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.269, + 0.219, + 0.386, + 0.231 + ], + "angle": 0, + "content": "(b) 300W (Challenge)" + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.099, + 0.561, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.443, + 0.219, + 0.528, + 0.231 + ], + "angle": 0, + "content": "(c) 300W (Full)" + }, + { + "type": "image", + "bbox": [ + 0.567, + 0.099, + 0.718, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.587, + 0.219, + 0.699, + 0.231 + ], + "angle": 0, + "content": "(d) CariFace (Target)" + }, + { + "type": "image", + "bbox": [ + 0.729, + 0.099, + 0.877, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.743, + 0.219, + 0.859, + 0.231 + ], + "angle": 0, + "content": "(e) ArtiFace (Unseen)" + }, + { + "type": "image_caption", + "bbox": [ + 0.124, + 0.243, + 0.843, + 0.258 + ], + "angle": 0, + "content": "Figure 7. Comparisons for different optimization strategies on NME performance in the GZSL (Unseen ArtiFace) setting." + }, + { + "type": "table_caption", + "bbox": [ + 0.086, + 0.282, + 0.459, + 0.296 + ], + "angle": 0, + "content": "Table 4. Comparison with SOTA landmark detectors on NME." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.308, + 0.477, + 0.374 + ], + "angle": 0, + "content": "
Learning MethodTrain on 300WOur method
BackboneOPSPIGASTARSLPTSTARSLPT
DA on Cariface10.4611.2310.9711.057.627.70
GZSL on Artiface5.555.195.304.564.863.93
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.402, + 0.47, + 0.492 + ], + "angle": 0, + "content": "obtained by our method show encouraging generalization power, which achieves lower NME than the baselines in the unseen domain (i.e., ArtiFace). Overall, the NME of our method in the source domain (i.e., 300W) is comparable to that achieved in the oracle setting. For the target even unseen domains, our method reduces the gap to the oracle." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.493, + 0.47, + 0.644 + ], + "angle": 0, + "content": "Besides the numerical comparisons, we provide some visualization results obtained by different methods in Figs. 5 and 6. For a complete comparison, the results of the SLPT trained only on \\(300\\mathrm{W}\\) are shown in the figures as well, which corresponds to learning a face landmarker only based on the data in the source domain. We can find that by applying our learning method, the face landmarker obtains better landmarks, which aligns with the ground truth with smaller errors, particularly for the landmarks of face contour, nose, mouth, and eyebrow." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.471, + 0.782 + ], + "angle": 0, + "content": "To further verify the generalization of our method, we select three more state-of-the-art landmarks, including OpenPose (OP) [7], SPIGA [36] and STAR [57] for numerical comparisons, and incorporate STAR as the backbone model of our method. Tab. 4 demonstrates that our learning method is applicable for various backbone models (e.g., STAR and SLPT), and integrating existing landmarkers with our learning method is an effective and competitive approach to enhance their generalizability." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.792, + 0.282, + 0.808 + ], + "angle": 0, + "content": "4.4. Analytic Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.816, + 0.375, + 0.831 + ], + "angle": 0, + "content": "4.4.1 Joint v.s. Alternating Optimization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "As aforementioned, learning the face landmarker and the warping field model jointly makes it much easier to fall into undesired local optimum or unstable saddle points. To verify this claim, we apply the joint optimization strategy, i.e.," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.284, + 0.895, + 0.451 + ], + "angle": 0, + "content": "solving (3) by optimizing \\(\\theta\\) and \\(\\gamma\\) jointly in each gradient descent step and compare its performance with ours. In Fig. 7, we show the best NME achieved by this joint optimization strategy and the NME achieved by our alternating optimization method in each iteration. We can find that the joint optimization strategy tends to overfit the source domain (300W), leading to power generalizability in the target and unseen domains. On the contrary, with the increase of iterations, the NMEs of our method on the target and unseen domains decrease consistently and become lower than those of the joint optimization after several iterations." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.47, + 0.776, + 0.485 + ], + "angle": 0, + "content": "4.4.2 Impacts of Model Architectures" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.494, + 0.892, + 0.797 + ], + "angle": 0, + "content": "Besides the optimization strategy, the model architecture also has an impact on the model performance. By default, we implement the face landmarker as the SLPT model. In this experiment, we further explore the performance of other model architectures, including SBR [9] and HRNet [47]. Tab. 3 shows the performance of different model architectures in DA and GZSL settings. We can find that the SLPT-based face landmarker works better than the two competitors in this experiment. A potential reason for this phenomenon is that both SBR and HRNet are heatmap-based face landmarks. Given an input facial image, they output a heatmap indicating the distribution of landmarks rather than a set of deterministic landmark coordinates. We have to first detect the landmarks from the heatmap and then pass them through the warping field model. Accordingly, in the face warper optimization step (i.e., solving (4)), the landmarker and the warping field model cannot be trained in an end-to-end way because the backpropagation of the gradient becomes inapplicable. As a result, we have to update \\(\\theta\\) and \\(\\gamma\\) alternatively, leading to suboptimal performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.816, + 0.702, + 0.831 + ], + "angle": 0, + "content": "4.4.3 Effects on Loss Term" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "In the context of our model's loss function, the image gradient field serves as its input. It has been observed through experiments that utilizing distinct gradient operators leads to the generation of varying image gradient fields, which in" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2431" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.112, + 0.09, + 0.435, + 0.104 + ], + "angle": 0, + "content": "Table 5. Quantitative comparison on CariFace dataset." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.116, + 0.46, + 0.173 + ], + "angle": 0, + "content": "
OperatorGraySpatialLaplacianCannySobel
NME✓8.2827.8708.0057.695
×7.9587.9218.0117.863
" + }, + { + "type": "table_caption", + "bbox": [ + 0.12, + 0.188, + 0.426, + 0.202 + ], + "angle": 0, + "content": "Table 6. Comparisons for different losses on NME." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.215, + 0.475, + 0.307 + ], + "angle": 0, + "content": "
Type of LossMSEPerceptualw.o. Grad MSEGrad MSE
300WCommon3.082.892.962.79
Challenge5.435.135.004.91
Full3.543.333.363.20
DA on Caricature8.268.239.487.70
GZSL on ArtiFace4.124.074.063.95
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.336, + 0.47, + 0.397 + ], + "angle": 0, + "content": "turn influences the model's performance. Consequently, in this section, we assess and compare the training outcomes derived from image gradient fields obtained through the application of diverse gradient operators." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.398, + 0.47, + 0.655 + ], + "angle": 0, + "content": "Moreover, it is crucial to recognize that the choice of the input image, specifically whether it is a grayscale image or not, can also exert a certain influence on the results. As a result, we have incorporated a comparison of the outcomes when the input is a grayscale image. It is important to emphasize that all experiment settings are maintained consistently. The results are shown in Tab. 5. Utilizing grayscale images yields superior effects compared to not using grayscale images. This could be attributed to the fact that converting an image to grayscale eliminates the trivial effects of color and complex textures. The most optimal outcome is achieved by employing the Sobel operator. This may be due to the Sobel operator's emphasis on edge information within the image, enabling it to capture more geometric details. Additionally, the Sobel operator incorporates a mild smoothing effect during gradient computation, which mitigates the impact of noise." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.655, + 0.47, + 0.777 + ], + "angle": 0, + "content": "Furthermore, we consider \\( i) \\) replacing the gradient MSE loss in (3) with the pixel MSE or perceptual loss and \\( ii) \\) removing the gradient MSE loss. Tab. 6 shows the rationality of the gradient MSE for the reason that \\( i) \\) landmarks are distributed on edges; and \\( ii) \\) the gradient field filters out unnecessary color information, simplifying the task. These findings form the basis for our decision to adopt this particular method in our experimental process." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.8, + 0.45, + 0.815 + ], + "angle": 0, + "content": "4.4.4 Rationality of Proposed Warping Field Model" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Besides the model architecture of the face landmarker, the warping field model impacts our face landmarker as well — when the predicted warping field is inaccurate, we cannot obtain reliable pseudo landmarks for stylized faces. Therefore, we investigate different face warping models and" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.089, + 0.628, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.18, + 0.598, + 0.191 + ], + "angle": 0, + "content": "(a) Source" + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.089, + 0.755, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.671, + 0.18, + 0.724, + 0.191 + ], + "angle": 0, + "content": "(b) Target" + }, + { + "type": "image", + "bbox": [ + 0.767, + 0.089, + 0.882, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.789, + 0.18, + 0.86, + 0.191 + ], + "angle": 0, + "content": "(c) AutoToon" + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.192, + 0.627, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.282, + 0.608, + 0.294 + ], + "angle": 0, + "content": "(d) CariGANs" + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.192, + 0.754, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.671, + 0.282, + 0.724, + 0.294 + ], + "angle": 0, + "content": "(e) RAFT" + }, + { + "type": "image", + "bbox": [ + 0.767, + 0.192, + 0.882, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.803, + 0.282, + 0.847, + 0.294 + ], + "angle": 0, + "content": "(f) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.547, + 0.307, + 0.843, + 0.321 + ], + "angle": 0, + "content": "Figure 8. Comparisons for different face warpers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.349, + 0.892, + 0.561 + ], + "angle": 0, + "content": "demonstrate the rationality of the warping field model implemented in our work. In particular, applying different warping field models in our training process, we visualize their warping results in Fig. 8. We can find that although the face landmarker with the polyharmonic interpolation model leads to a very simple face warper, it outperforms many existing neural network-based image warping methods, such as AutoToon [15] and CariGANs [5] for facial manipulation and the optical flow methods like RAFT [44]. These methods either require one-to-one correspondence information between source and target images or assume the deformation between the two images to be slight, making them unsuitable for our problem, especially for the stylized faces with significant nonrigid deformations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.576, + 0.62, + 0.592 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.602, + 0.892, + 0.784 + ], + "angle": 0, + "content": "In this paper, we propose a simple but effective method for learning a generalizable face landmarker applicable to facial images with different styles. Given labeled real human faces and unlabeled stylized faces, our method learns the face landmarker under the guidance of conditional face warping, demonstrating the usefulness of the warping information. An alternating optimization framework is proposed to learn the face landmarker together with the warping field model. Experiments demonstrate the effectiveness of our method. Especially in the generalized zero-shot learning scenarios, our method achieves encouraging landmarking accuracy in unseen face domains." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.8, + 0.669, + 0.816 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.902 + ], + "angle": 0, + "content": "This work was supported in part by the National Natural Science Foundation of China (62102031), the foundation of Key Laboratory of Artificial Intelligence, Ministry of Education, P.R. China, and the Young Scholar Program (XSQD-202107001) from Beijing Institute of Technology." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2432" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Olalekan Agbolade, Azree Nazri, Razali Yaakob, Abdul Azim Abd Ghani, and Yoke Kqueen Cheah. Landmark-based homologous multi-point warping approach to 3d facial recognition using multiple datasets. PeerJ Computer Science, 6:e249, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.268 + ], + "angle": 0, + "content": "[2] Sadegh Aliakbarian, Pashmina Cameron, Federica Bogo, Andrew Fitzgibbon, and Thomas J. Cashman. Flag: Flow-based 3d avatar generation from sparse observations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13253-13262, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.271, + 0.47, + 0.313 + ], + "angle": 0, + "content": "[3] Peter N. Belhumeur, David W. Jacobs, David J. Kriegman, and Neeraj Kumar. Localizing parts of faces using a consensus of exemplars. In CVPR 2011, 2011. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.315, + 0.47, + 0.357 + ], + "angle": 0, + "content": "[4] Hongrui Cai, Yudong Guo, Zhuang Peng, and Juyong Zhang. Landmark detection and 3d face reconstruction for caricature using a nonlinear parametric model, 2020. 1, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.359, + 0.47, + 0.4 + ], + "angle": 0, + "content": "[5] Kaidi Cao, Jing Liao, and Lu Yuan. Carigans: Unpaired photo-to-caricature translation. arXiv preprint arXiv:1811.00222, 2018. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.402, + 0.47, + 0.444 + ], + "angle": 0, + "content": "[6] Xudong Cao, Yichen Wei, Fang Wen, and Jian Sun. Face alignment by explicit shape regression. International journal of computer vision, 107:177-190, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.446, + 0.47, + 0.502 + ], + "angle": 0, + "content": "[7] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7291-7299, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.504, + 0.47, + 0.572 + ], + "angle": 0, + "content": "[8] Forrester Cole, David Belanger, Dilip Krishnan, Aaron Sarna, Inbar Mosseri, and William T Freeman. Synthesizing normalized faces from facial identity features. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3703-3712, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.574, + 0.47, + 0.657 + ], + "angle": 0, + "content": "[9] Xuanyi Dong, Shouu-I Yu, Xinshuo Weng, Shih-En Wei, Yi Yang, and Yaser Sheikh. Supervision-by-registration: An unsupervised approach to improve the precision of facial landmark detectors. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 360–368, 2018. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.66, + 0.47, + 0.743 + ], + "angle": 0, + "content": "[10] Alexey Dosovitskiy, Philipp Fischer, Eddy Ilg, Philip Hausser, Caner Hazirbas, Vladimir Golkov, Patrick Van Der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 2758-2766, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.745, + 0.47, + 0.801 + ], + "angle": 0, + "content": "[11] Pengfei Dou, Shishir K Shah, and Ioannis A Kakadiaris. End-to-end 3d face reconstruction with deep neural networks. In proceedings of the IEEE conference on computer vision and pattern recognition, pages 5908-5917, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.803, + 0.47, + 0.857 + ], + "angle": 0, + "content": "[12] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In International conference on machine learning, pages 1180-1189. PMLR, 2015. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.859, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[13] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. A neural algorithm of artistic style. arXiv preprint arXiv:1508.06576, 2015. 3" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "[14] C. A. Glasbey and K. V. Mardia. A review of image-warping methods. Journal of Applied Statistics, 25(2):155-171, 2002. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.134, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[15] Julia Gong, Yannick Hold-Geoffroy, and Jingwan Lu. Auto-toon: Automatic geometric warping for face cartoon generation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 360–369, 2020. 1, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.205, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[16] Arthur Gretton, Karsten Borgwardt, Malte Rasch, Bernhard Schölkopf, and Alex Smola. A kernel method for the two-sample-problem. Advances in neural information processing systems, 19, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.261, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[17] Guoliang Kang, Lu Jiang, Yi Yang, and Alexander G Hauptmann. Contrastive adaptation network for unsupervised domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4893-4902, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.33, + 0.892, + 0.385 + ], + "angle": 0, + "content": "[18] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.386, + 0.892, + 0.441 + ], + "angle": 0, + "content": "[19] Vahid Kazemi and Josephine Sullivan. One millisecond face alignment with an ensemble of regression trees. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1867-1874, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.442, + 0.892, + 0.469 + ], + "angle": 0, + "content": "[20] SunnieS. Kim, NicholasI. Kolkin, Jason Salavon, and Gregory Shakhnarovich. Deformable style transfer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.47, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[21] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.511, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[22] Marek Kowalski, Jacek Naruniec, and Tomasz Trzcinski. Deep alignment network: A convolutional neural network for robust face alignment. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.581, + 0.892, + 0.622 + ], + "angle": 0, + "content": "[23] Vuong Le, Jonathan Brandt, Zhe Lin, Lubomir Bourdev, and Thomas S. Huang. Interactive Facial Feature Localization, page 679-692. 2012. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.623, + 0.892, + 0.692 + ], + "angle": 0, + "content": "[24] Hui Li, Zidong Guo, Seon-Min Rhee, Seungju Han, and Jae-Joon Han. Towards accurate facial landmark detection via cascaded transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4176-4185, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.692, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[25] Yunsheng Li, Lu Yuan, and Nuno Vasconcelos. Bidirectional learning for domain adaptation of semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6936-6945, 2019. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.762, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[26] Feng Liu, Dan Zeng, Qijun Zhao, and Xiaoming Liu. Joint face alignment and 3d face reconstruction. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14, pages 545-560. Springer, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[27] Shaolei Liu, Xiaoyuan Luo, Kexue Fu, Manning Wang, and Zhijian Song. A learnable self-supervised task for unsupervised domain adaptation on point cloud classification and segmentation. Frontiers of Computer Science, 17(6):176708, 2023. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2433" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[28] Weiyang Liu, Yandong Wen, Zhiding Yu, Ming Li, Bhiksha Raj, and Le Song. Spherface: Deep hypersphere embedding for face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 212-220, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.219 + ], + "angle": 0, + "content": "[29] Xiao-Chang Liu, Yong-Liang Yang, and Peter Hall. Learning to warp for style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3702-3711, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.47, + 0.276 + ], + "angle": 0, + "content": "[30] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In International conference on machine learning, pages 97-105. PMLR, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.47, + 0.332 + ], + "angle": 0, + "content": "[31] Mingsheng Long, Han Zhu, Jianmin Wang, and Michael I Jordan. Unsupervised domain adaptation with residual transfer networks. Advances in neural information processing systems, 29, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.47, + 0.404 + ], + "angle": 0, + "content": "[32] Jiangjing Lv, Xiaohu Shao, Junliang Xing, Cheng Cheng, and Xi Zhou. A deep regression architecture with two-stage re-initialization for high performance facial landmark detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3317-3326, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.47, + 0.461 + ], + "angle": 0, + "content": "[33] Iacopo Masi, Stephen Rawls, Gérard Medioni, and Prem Natarajan. Pose-aware face recognition in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4838-4846, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.462, + 0.47, + 0.531 + ], + "angle": 0, + "content": "[34] Kieron Messer, Jiri Matas, Josef Kittler, Juergen Luettin, Gilbert Maitre, et al. Xm2vtsdb: The extended m2vts database. In Second international conference on audio and video-based biometric person authentication, pages 965-966. CiteSeer, 1999. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.532, + 0.47, + 0.574 + ], + "angle": 0, + "content": "[35] Zhongyi Pei, Zhangjie Cao, Mingsheng Long, and Jianmin Wang. Multi-adversarial domain adaptation. In Thirty-second AAAI conference on artificial intelligence, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.47, + 0.629 + ], + "angle": 0, + "content": "[36] Andrés Prados-Torreblanca, José M Buenaposada, and Luis Baumela. Shape preserving facial landmarks with graph attention networks. arXiv preprint arXiv:2210.07233, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.47, + 0.688 + ], + "angle": 0, + "content": "[37] Joseph Roth, Yiying Tong, and Xiaoming Liu. Unconstrained 3d face reconstruction. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2606-2615, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.47, + 0.758 + ], + "angle": 0, + "content": "[38] Christos Sagonas, Georgios Tzimiropoulos, Stefanos Zafeiriou, and Maja Pantic. 300 faces in-the-wild challenge: The first facial landmark localization challenge. In 2013 IEEE International Conference on Computer Vision Workshops, 2014. 1, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.47, + 0.814 + ], + "angle": 0, + "content": "[39] Ahmed Selim, Mohamed Elgharib, and Linda Doyle. Painting style transfer for head portraits using convolutional neural networks. ACM Transactions on Graphics (ToG), 35(4): 1-18, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.47, + 0.872 + ], + "angle": 0, + "content": "[40] Yichun Shi, Debayan Deb, and Anil K Jain. Warpgan: Automatic caricature generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10762-10771, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[41] Baochen Sun, Jiashi Feng, and Kate Saenko. Correlation alignment for unsupervised domain adaptation. Domain" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "adaptation in computer vision applications, pages 153-171, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[42] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[43] Yaniv Taigman, Ming Yang, Marc'Aurelio Ranzato, and Lior Wolf. Deepface: Closing the gap to human-level performance in face verification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1701-1708, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.25, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[44] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.321, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[45] Yi-Hsuan Tsai, Wei-Chih Hung, Samuel Schulter, Kiyuh Yuk Sohn, Ming-Hsuan Yang, and Manmohan Chandraker. Learning to adapt structured output space for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7472-7481, 2018. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.405, + 0.892, + 0.461 + ], + "angle": 0, + "content": "[46] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.462, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[47] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 43(10):3349-3364, 2020. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.547, + 0.892, + 0.603 + ], + "angle": 0, + "content": "[48] Xinyao Wang, Liefeng Bo, and Li Fuxin. Adaptive wing loss for robust face alignment via heatmap regression. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.604, + 0.892, + 0.658 + ], + "angle": 0, + "content": "[49] Kunhong Wu, Fan Jia, and Yahong Han. Domain-specific feature elimination: multi-source domain adaptation for image classification. Frontiers of Computer Science, 17(4): 174705, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.66, + 0.892, + 0.703 + ], + "angle": 0, + "content": "[50] Ruizheng Wu, Xiaodong Gu, Xin Tao, Xiaoyong Shen, Yu-Wing Tai, et al. Landmark assisted cyclegan for cartoon face generation. arXiv preprint arXiv:1907.01424, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.704, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[51] Wayne Wu, Chen Qian, Shuo Yang, Quan Wang, Yici Cai, and Qiang Zhou. Look at boundary: A boundary-aware face alignment algorithm. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2129-2138, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.775, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[52] Jiahao Xia, Weiwei Qu, Wenjian Huang, Jianguo Zhang, Xi Wang, and Min Xu. Sparse local patch transformer for robust face alignment and landmarks inherent relation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4052-4061, 2022. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[53] Jiaolong Yang, Peiran Ren, Dongqing Zhang, Dong Chen, Fang Wen, Hongdong Li, and Gang Hua. Neural aggregation network for video face recognition. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2434" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.468, + 0.12 + ], + "angle": 0, + "content": "the IEEE conference on computer vision and pattern recognition, pages 4362-4371, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.471, + 0.177 + ], + "angle": 0, + "content": "[54] Yanchao Yang and Stefano Soatto. Fda: Fourier domain adaptation for semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4085-4095, 2020. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.468, + 0.232 + ], + "angle": 0, + "content": "[55] Jordan Yaniv, Yael Newman, and Ariel Shamir. The face of art: landmark detection and geometric style in portraits. ACM Transactions on graphics (TOG), 38(4):1-15, 2019. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.468, + 0.303 + ], + "angle": 0, + "content": "[56] Jing Zhang, Zewei Ding, Wanqing Li, and Philip Ogunbona. Importance weighted adversarial nets for partial domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8156-8164, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.468, + 0.374 + ], + "angle": 0, + "content": "[57] Zhenglin Zhou, Huaxia Li, Hong Liu, Nanyang Wang, Gang Yu, and Rongrong Ji. Star loss: Reducing semantic ambiguity in facial landmark detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15475-15484, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.376, + 0.468, + 0.445 + ], + "angle": 0, + "content": "[58] Zhou, Erjin and Fan, Haoqiang and Cao, Zhimin and Jiang, Yuning and Yin, Qi. Extensive Facial Landmark Localization with Coarse-to-Fine Convolutional Network Cascade. In Proceedings of the IEEE International Conference on Computer Vision (ICCV) Workshops, 2013. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.447, + 0.468, + 0.514 + ], + "angle": 0, + "content": "[59] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.517, + 0.468, + 0.572 + ], + "angle": 0, + "content": "[60] Xiangxin Zhu and Deva Ramanan. Face detection, pose estimation, and landmark localization in the wild. In 2012 IEEE conference on computer vision and pattern recognition, pages 2879-2886. IEEE, 2012. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.574, + 0.468, + 0.628 + ], + "angle": 0, + "content": "[61] Yi Zhu, Xindong Wu, Jipeng Qiang, Yunhao Yuan, and Yun Li. Representation learning via an integrated autoencoder for unsupervised domain adaptation. Frontiers of Computer Science, 17(5):175334, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.471, + 0.628 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2435" + } + ] +] \ No newline at end of file diff --git a/2024/Generalizable Face Landmarking Guided by Conditional Face Warping/33b95235-e8cb-4b6b-a79e-5de2b0524965_origin.pdf b/2024/Generalizable Face Landmarking Guided by Conditional Face Warping/33b95235-e8cb-4b6b-a79e-5de2b0524965_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a5d3ae9f15498354a1f9cf3e1e1208cbec349d56 --- /dev/null +++ b/2024/Generalizable Face Landmarking Guided by Conditional Face Warping/33b95235-e8cb-4b6b-a79e-5de2b0524965_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b37614fb77cf765c13485f5dbbf2051d4459e33a5d9c1644cdea8db159917c1d +size 7174343 diff --git a/2024/Generalizable Face Landmarking Guided by Conditional Face Warping/full.md b/2024/Generalizable Face Landmarking Guided by Conditional Face Warping/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6c2f979b8fee4c100ae0f75abe0dbcc1b139dedf --- /dev/null +++ b/2024/Generalizable Face Landmarking Guided by Conditional Face Warping/full.md @@ -0,0 +1,367 @@ +# Generalizable Face Landmarking Guided by Conditional Face Warping + +Jiayi Liang $^{1}$ Haotian Liu $^{1*}$ Hongteng Xu $^{2,3}$ Dixin Luo $^{1,4\dagger}$ + +$^{1}$ School of Computer Science and Technology, Beijing Institute of Technology, Beijing + +$^{2}$ Gaoling School of Artificial Intelligence, Renmin University of China, Beijing + +3Beijing Key Laboratory of Big Data Management and Analysis Methods, Beijing + +$^{4}$ Key Laboratory of Artificial Intelligence, Ministry of Education, Shanghai + +{jiayi.liang, haotianliu, dixin.luo}@bit.edu.cn, hongtengxu@ruc.edu.cn + +# Abstract + +As a significant step for human face modeling, editing, and generation, face landmarking aims at extracting facial keypoints from images. A generalizable face landmarker is required in practice because real-world facial images, e.g., the avatars in animations and games, are often stylized in various ways. However, achieving generalizable face landmarking is challenging due to the diversity of facial styles and the scarcity of labeled stylized faces. In this study, we propose a simple but effective paradigm to learn a generalizable face landmarker based on labeled real human faces and unlabeled stylized faces. Our method learns the face landmarker as the key module of a conditional face warper. Given a pair of real and stylized facial images, the conditional face warper predicts a warping field from the real face to the stylized one, in which the face landmarker predicts the ending points of the warping field and provides us with high-quality pseudo landmarks for the corresponding stylized facial images. Applying an alternating optimization strategy, we learn the face landmarker to minimize i) the discrepancy between the stylized faces and the warped real ones and ii) the prediction errors of both real and pseudo landmarks. Experiments on various datasets show that our method outperforms existing state-of-the-art domain adaptation methods in face landmarking tasks, leading to a face landmarker with better generalizability. Code is available at https://plustwo0.github.io/project-face-landmarker. + +# 1. Introduction + +Face landmarking seeks to extract human facial keypoints (e.g., eyes, nose, facial contour, and so on) from facial images. This task is important for many applications in the field of computer vision and graphics, such as face recogni + +![](images/66c55eec947896835ab6c5a3a888c16ac0860ff886075170bc03387086dab6c3.jpg) +Figure 1. Both commercial software like Face++ and open-source method like SLPT [52] work well on landmarking real faces (e.g., those in 300W [38]) while achieving suboptimal performance when landmarking stylized faces (e.g., those in CariFace [4] and ArtiFace [55]). While existing domain adaptation method does not improve the performance significantly, our method achieves a generalizable face landmarker for various facial images. + +tion [1, 28, 33, 43, 53], face stylization [5], and 3D face reconstruction [11, 26, 37]. Currently, many open-source and commercial face landmarks [19, 51, 58] have been developed and achieved encouraging performance in this task. + +Most existing face landmarks are designed and trained for landmarking real human faces, while the rapid development of AIGC applications, such as artistic character creation and cartoon generation [2, 15], leads to a massive increase in demands for landmarking stylized facial images. Unfortunately, as shown in Fig. 1, existing face landmarks often fail to landmark stylized facial images. Even if applying state-of-the-art domain adaptation strategies [12, 25, 27, 59, 61], the generalizability of the learned landmarks in the stylized facial image domain is still unsatisfactory. Essentially, traditional face landmarks work + +![](images/585c55b834d70529e61d8d9e43412b1b0772b11c4d293ddc9e534febc672ec3a.jpg) +Figure 2. The scheme of our proposed method for learning a generalizable face landmarker. + +well on real human faces because of the relatively stable geometry of real human faces and the sufficient labeled facial images. These two conditions, however, become questionable when landmarking stylized faces — the stylized facial images often have various facial styles, and manually landmarking such stylized faces is much more time-consuming than landmarking real faces. As a result, learning a generalizable face landmarker becomes a challenging task. + +In this paper, we propose a simple but effective paradigm for learning a generalizable face landmarker, overcoming the challenges caused by the diversity of facial styles and the scarcity of labeled stylized faces. As illustrated in Fig. 2, given labeled real human facial images and unlabeled stylized facial images, we learn a face landmarker embedded in a conditional face warper. The face warper aims to deform real human faces according to stylized facial images, generating warped faces and corresponding warping fields. The face landmarker, as the key module of the warper, predicts the ending points of the warping fields and thus provides us with pseudo landmarks for the stylized facial images. The warping field is parametrized by a polyharmonic interpolation model. Under the guidance of the conditional face warping, we learn the face landmarker in an alternating optimization framework: The face landmarker is updated to $i$ minimize the discrepancy between the stylized faces and the corresponding warped real human faces and $ii$ minimize the prediction errors of both real and pseudo landmarks. In the first step, the face landmarker is learned associated with the warping field model, while in the second step, the face landmarker is updated with proximal regularization. + +The extensive experiments in various face landmarking tasks demonstrate the effectiveness of learning method. The impacts of different loss functions and data settings on + +our learning method are analyzed through detailed ablation studies. Experimental results show that our learning method results in a face landmarker that is generalizable to the facial images with different styles, which outperforms representative domain adaptation methods consistently in various stylized face landmarking tasks. + +# 2. Related Work + +# 2.1. Face Landmarking + +Given facial images, the early face landmarking methods learn regression models to predict landmark coordinates directly [6, 32]. The models are often parameterized by neural networks like Transformer [46], capturing facial attributes that have been proven to be crucial for landmark prediction [24]. The coordinate regression is achieved in a coarse-to-fine framework [58], leading to a cascading landmarking pipeline. Recently, to make the landmark prediction robust to the variations in pose, scale, and occlusion, DAN [22] introduces the novel use of heatmaps and extracts features from the entire face rather than local patches around landmarks. SBR [9] utilizes the registration of synthesized images to provide supervisory signals for training. Adaptive Wing Loss [48] is proposed to address the imbalance between foreground pixels and background pixels by analysis of the main drawbacks of different loss functions. HRNet [47] produces high-resolution maps by connecting and exchanging information via merging multi-scale picture features across many branches. + +As aforementioned, most existing face landmarks are learned for real human faces, which can not be adapted directly to stylized faces (e.g., cartoon and artistic faces). Essentially, treating labeled real human faces as a source domain and unlabeled stylized faces as a target domain, we can learn a generalizable face landmark by solving a domain adaptation (DA) problem [30, 31]. Accordingly, many DA techniques [27, 49] have potential in our problem, including the classic metric learning-based methods (e.g., CORAL [41], contrastive domain discrepancy [17], and maximum mean discrepancy [16]) and the recent adversarial learning-based methods [12, 35, 56, 59]. However, in the following content, we will show that directly applying these DA techniques often fails to achieve generalizable face landmarking because of the significant gap between the source and target face domains. + +# 2.2. Face Warping + +Face warping is a technique that involves geometrically deforming source facial images to specified target shapes. The key step of this task is predicting a warping field between the source and target images that captures the shifts of image pixels. To achieve this aim, DST [20] and FoA [55] find matching keypoints between source and target images + +and then generate a dense warping field through data interpolation [8]. Instead of matching keypoints, some methods learn neural networks to predict dense warping fields directly based on paired images, e.g., Flownet [10], AutoToon [15], RAFT [44], and their variants [29]. However, these methods require the paired images to be similar to each other, which cannot capture significant deformations between the faces in different domains. + +Compared to face stylization [13, 18, 39, 40], face warping is a relatively easier task because it only considers the deformation of shapes while ignoring the transfer of textures. However, it should be noted that this task is more relevant to face landmarking, in which the warping field provides us with strong evidence to shift face landmarks of source faces to target ones [50]. Inspired by such a strong correlation, we develop the proposed learning paradigm. + +# 3. Proposed Method + +Denote $\mathcal{X}$ as the image space and $\mathcal{V}$ as the landmark space, respectively. In this work, we observe a set of labeled real human faces, i.e., $\mathcal{D}^{(L)} = \{\pmb{X}_i^{(L)},\pmb{Y}_i^{(L)}\}_{i = 1}^{N_L}\subset \mathcal{X}_R\times \mathcal{V}$ and a set of unlabeled stylized faces, i.e., $\mathcal{D}^{(U)} = \{\pmb{X}_i^{(U)}\}_{i = 1}^{N_U}\subset$ $\mathcal{X}_S$ , where $\mathcal{X}_R,\mathcal{X}_S\subset \mathcal{X}$ correspond to the real and stylized face domains, respectively. Each $\pmb {X}_i\in \mathbb{R}^{H\times W\times 3}$ represents an image, and each $\pmb {Y}_i = [\pmb {y}_{i,k}]\in \mathbb{R}^{2\times K}$ records $K$ face landmark coordinates, where $\pmb {y}_{i,k}\in \mathbb{R}^2$ . We aim to learn a face landmarker, denoted as $f_{\theta}:\mathcal{X}\mapsto \mathcal{V}$ , where $\theta$ is the model parameter. The model should be able to predict face landmarks from facial images and moreover, generalize to both $\mathcal{X}_R$ and $\mathcal{X}_S$ . To achieve this aim, we embed the face landmarker into a conditional face warper and learn it associated with a parametric warping field predictor in an alternating optimization framework, as illustrated in Fig. 2. + +# 3.1. Face Landmarking Guided by Face Warping + +In this study, we take the SLPT model [52] as the backbone of our face landmarker. Given a stylized face $\mathbf{X}_i^{(U)}$ , the face landmarker predicts its landmarks as $\widehat{\mathbf{Y}}_i^{(U)} = [\widehat{\mathbf{y}}_{i,k}^{(U)}] = f_\theta(\mathbf{X}_i^{(U)})$ . At the same time, we can sample a labeled real face $(\mathbf{X}_j^{(L)},\mathbf{Y}_j^{(L)}) \sim \mathcal{D}^L$ . Treating the labeled and predicted landmarks as keypoints, we can model a warping field from the real face to the stylized one by the following polyharmonic interpolation model [14]: + +$$ +w _ {i, \gamma} (\boldsymbol {y}) = \sum_ {k = 1} ^ {K} \omega_ {k} \phi \left(\left\| \boldsymbol {y} - \hat {\boldsymbol {y}} _ {i, k} ^ {(U)} \right\| _ {2}\right) + \boldsymbol {V} \boldsymbol {y} + \boldsymbol {b}, \tag {1} +$$ + +where $\gamma = \{\{\pmb{\omega}_k\in \mathbb{R}^2\}_{k = 1}^K,\pmb {V}\in \mathbb{R}^{2\times 2},\pmb {b}\in \mathbb{R}^2\}$ correspond to the parameters of the warping field. As shown in (1), the vector $\pmb{y}$ denotes the u-v coordinate of a pixel in the stylized facial image, and $w_{i,\gamma}(\pmb {y})$ gives the inverse mapping from the pixel $\pmb{y}$ to a coordinate in the real human facial image, conditioned on $\pmb{X}_i^{(U)}$ . The first term + +$\sum_{k=1}^{K} \omega_k \phi(\| \pmb{y} - \hat{\pmb{y}}_{i,k}^{(U)} \|_2)$ achieves nonparametric regression for modeling nonrigid deformations, in which $\phi(r)$ is a predefined thin-plate spline function. The second term $\pmb{V}\pmb{y} + \pmb{b}$ is a linear parametric model capturing the rigid transformation of $\pmb{y}$ . + +For each pixel coordinate $\pmb{y} \in \{1, \dots, H\} \times \{1, \dots, W\}$ , we can trace it back to the real human facial image based on $w_{i,\gamma}(\pmb{y})$ and obtain the pixel color as $\pmb{X}_j^{(L)}(w_{i,\gamma}(\pmb{y}))$ . Accordingly, with the grid sampler constructed via inverse mapping function $w_{i,\gamma}$ , we obtain the warped real human facial image conditioned on $\pmb{X}_i^{(U)}$ , denoted as $\widehat{\pmb{X}}_{j|i}^{(L)}$ . For $\pmb{y} \in \{1, \dots, H\} \times \{1, \dots, W\}$ and $j = 1, \dots, N_L$ , we have + +$$ +\widehat {\boldsymbol {X}} _ {j \mid i} ^ {(L)} (\boldsymbol {y}) = \boldsymbol {X} _ {j} ^ {(L)} \left(w _ {i, \gamma} (\boldsymbol {y})\right). \tag {2} +$$ + +Unlike WarpGAN [40], which generates the warped face by predicting dense keypoints and their displacements by two fully-connected layers during training, we directly use the predicted and observed landmarks to define sparse displacements and estimate other pixels' displacements by spline-based interpolation, which improves computational efficiency significantly. Moreover, by applying the warping field model with limited degree-of-freedom (i.e., few learnable parameters), we can focus more on the learning of the face landmarker in the training phase. + +Specifically, the warped face together with the warping field provides a useful guidance for the learning of the face landmarker. In particular, we formulate the learning problem of the face landmarker as follows: + +$$ +\begin{array}{l} \min _ {\theta , \gamma} \underbrace {\sum_ {j = 1} ^ {N _ {L}} \| f _ {\theta} \left(\boldsymbol {X} _ {j} ^ {(L)}\right) - \boldsymbol {Y} _ {j} ^ {(L)} \| _ {F} ^ {2}} _ {\text {L a n d m a r k i n g e r r o r i n t h e s o u r c e d o m a i n}} \\ + \underbrace {\sum_ {i = 1} ^ {N _ {U}} \sum_ {j = 1} ^ {N _ {L}} \left\| \nabla \widehat {\boldsymbol {X}} _ {j | i} ^ {(L)} - \nabla \boldsymbol {X} _ {i} ^ {(U)} \right\| _ {F} ^ {2}} _ {\text {D i s c r e p a n c y o f i m a g e g r a d i e n t}} \tag {3} \\ + \underbrace {\sum_ {i = 1} ^ {N _ {U}} \sum_ {j = 1} ^ {N _ {L}} \| w _ {i , \gamma} (\widehat {\mathbf {Y}} _ {i} ^ {(U)}) - \mathbf {Y} _ {j} ^ {(L)} \| _ {F} ^ {2}} _ {\text {L a n d m a r k w a r p i n g e r r o r}}, \\ \end{array} +$$ + +where $\| \cdot \| _F$ represents the Frobenius norm of matrix. In (3), the first term is the landmarking error for real faces, which corresponds to the data fidelity loss in the source domain. The second term measures the discrepancy between the stylized face and the warped real face in the gradient field, in which the gradient operation $\nabla$ is implemented by the Sobel operator. The third term is the landmark warping error. Both the second and third terms are determined jointly by the landmarker $f_{\theta}$ and warping field model $w_{i,\gamma}$ . + +# 3.2. Alternating Optimization Strategy + +The optimization problem in (3) is non-convex because the landmarker is implemented by a neural network and is coupled to the warping field model. As a result, learning $\theta$ and + +![](images/b2e949308641b6c3c5184263e60a6bee22e42be3fc42af04f0f2207e4b5845cf.jpg) +Warped Face with Warped Landmarks + +![](images/d7db80611db195609997eb5cd87fdc0da8d47ff03d9995835ccc1b87ec45b688.jpg) + +![](images/2799767401c9c88f2bcad3101714f5661bc61eb91b9976204df0e645bf76233f.jpg) + +![](images/9756a62a1b30f6c9a3900aace7257873629c25a0ab32b9a23c71b796fdd7aa8d.jpg) + +![](images/b4e305fe09aecd53102671f37ace164c08b8d0fc1adf6643c5bd69489be3a340.jpg) + +![](images/3b2ec01642ccf3f00d47fadc7173d258c25a78116ad57632b544c60de13cfebb.jpg) +Figure 3. Illustrations of conditional face warping results. Taking a cartoon face as the target, our model warps real human faces accordingly. The red dots indicate real human face landmarks, and green dots indicate cartoon and warped face landmarks. + +![](images/c642614f178cfca3d97bdd1f4840cdf6807463e1fe737706398ea34431f6e0c4.jpg) + +![](images/678ec486c5029a74aba8d95f0948a1fee8e4256f96951544bbb3be72a35f21d2.jpg) + +![](images/213138cb5e4acc88168d0514a3dc3c26031dcc07841dd9a2e2615a3a432759fc.jpg) + +$\gamma$ jointly often falls into an undesired local optimum even an unstable saddle point. To mitigate this issue, we propose an alternating optimization framework. In principle, we can decompose the optimization problem in (3) into the following two subproblems and solve them iteratively. + +- Face Warper Optimization: The first subproblem corresponds to the optimization of the face warper, i.e., + +$$ +\begin{array}{l} \theta^ {(1)}, \gamma^ {(1)} = \arg \min _ {\theta , \gamma} \sum_ {i, j} \| \nabla \widehat {\boldsymbol {X}} _ {j | i} ^ {(L)} - \nabla \boldsymbol {X} _ {i} ^ {(U)} \| _ {F} ^ {2} \tag {4} \\ + \sum_ {i, j} \| w _ {i, \gamma} (\widehat {\pmb {Y}} _ {i} ^ {(U)}) - \pmb {Y} _ {j} ^ {(L)} \| _ {F} ^ {2}. \\ \end{array} +$$ + +In this subproblem, we only care about whether the real human faces can be warped as the stylized faces with high accuracy, so the term of landmarking error is ignored. We solve this problem by Adam [21]: in each step, we update $\theta$ and $\gamma$ based on a batch of randomly-sampled face pairs. + +- Proximal Face Landmarker Optimization: Given $\theta^{(1)}$ and the predicted landmarks (i.e., $\widehat{\pmb{Y}}_i^{(U)} = f_{\theta^{(1)}}(\pmb{X}_i^{(U)})$ for $i = 1, \dots, N_L$ ), we can treat $\theta^{(1)}$ as the initial variable and optimize it with a proximal regularizer: + +$$ +\begin{array}{l} \theta^ {(2)} = \arg \min _ {\theta} \sum_ {j = 1} ^ {N _ {L}} \| f _ {\theta} (\boldsymbol {X} _ {j} ^ {(L)}) - \boldsymbol {Y} _ {j} ^ {(L)} \| _ {F} ^ {2} \\ + \underbrace {\sum_ {i = 1} ^ {N _ {U}} \| f _ {\theta} \left(\boldsymbol {X} _ {i} ^ {(U)}\right) - \widehat {\boldsymbol {Y}} _ {i} ^ {(U)} \| _ {F} ^ {2}}. \tag {5} \\ \end{array} +$$ + +Pseudo landmarking error in the target domain + +Here, the second term in (5) measures the estimation errors of the pseudo landmarks achieved in the previous step. Essentially, it works as a proximal regularizer, ensuring that the optimized landmarks $f_{\theta^{(2)}}(\mathbf{X}_i^{(U)})$ is not too far away from the previous estimation $\widehat{\mathbf{Y}}_i^{(U)}$ . Similarly, we can solve this problem by Adam [21] as well. + +Fig. 3 shows the warping effect on real human faces achieved by solving (4). In Fig. 3, the first row shows the real human faces with landmarks and the target stylized face, and the second row shows the warping results, + +# Algorithm 1 Proposed learning scheme of face landmarker + +Require: Labeled real faces $\mathcal{D}^{(L)}$ and unlabeled stylized faces $\mathcal{D}^{(U)}$ . The number of iterations (i.e., $M$ ). Epochs for the subproblems (i.e., $L_{1}$ and $L_{2}$ ). + +1: Initialize $\{\gamma^{(0)}\}$ with a pretrained model on $\mathcal{D}^{(L)}$ and $\{\theta^{(0)}\}$ randomly. +2: for $m = 0, \dots, M - 1$ do +3: Sample a batch $\{\pmb{X}_i^{(U)},\pmb{X}_i^{(L)},\pmb{Y}_i^{(L)}\}_{i = 1}^N$ +4: Face warper optimization: +5: Take $\theta^{(2m)},\gamma^{(m)}$ as the initialization, then solve (4) by Adam with $L_{1}$ epochs and obtain $\theta^{(2m + 1)}$ +6: Proximal face landmarker optimization: +7: Take $\theta^{(2m + 1)}$ as the initialization, then solve (5) by Adam with $L_{2}$ epochs and obtain $\theta^{(2m + 2)}$ . +8: end for +9: return Output a generalizable face landmarker $f_{\theta^{(2M)}}$ . + +in which the green dots are predicted landmarks. These results empirically demonstrate the rationality of our alternating optimization framework. In particular, we can find that solving (4) leads to reasonable warping results, which are similar to the target stylized face on shape. The similarity on face shape indicates that the predicted landmarks can be treated as reliable pseudo labels of the stylized face, which can be used to construct the proximal regularizer that penalizing the peso do landmarking errors in the target domain. Repeating the above two steps till converge, we obtain the target face landmarker that is generalizable for both real and stylized faces. Algorithm 1 shows the learning scheme. + +# 4. Experiment + +We apply our learning method to learn a face landmarker and test it on landmarking faces with various styles. Extensive experiments, including comparisons with baselines and analytic ablation studies, demonstrate the effectiveness of our learning method and the generalizability of the corresponding face landmarker. All the experiments are conducted on a single NVIDIA 3090 GPU. Representative experimental results are shown below. More experimental results and implementation details are given in the supplementary material. + +# 4.1. Dataset + +In this study, we conduct experiments based on the following three commonly-used face datasets. + +- 300W Dataset. 300W [38] is comprised of five well-known real human face datasets including LFPW [3], AFW [60], HELEN [23], XM2VTS [34], and IBUG [38]. +- CariFace Dataset. CariFace [4] is created by searching and selecting thousands of various caricatures from different celebrities on the Internet. + +Table 1. Data settings for the three learning paradigms. + +
Learning Paradigm#Training Images and Label Information#Testing Images
300WCariFaceArtiFaceCommon300W ChallengeFullCariFaceArtiFace
DA (300W→CariFace)3,148Labeled3,372Unlabeled——554135689800—
DA (300W→ArtiFace)3,148Labeled——128Unlabeled554135689—32
GZSL (Unseen ArtiFace)3,148Labeled3,372Unlabeled——554135689800160
GZSL (Unseen CariFace)3,148Labeled——128Unlabeled55413568980032
Oracle3,148Labeled3,372Labeled128Labeled55413568980032
+ +![](images/5fa9e5a7fd20df897a92c2f02d3b3db7c70ae67afbc1f356308b48918f7dda41.jpg) +Figure 4. Illustrations of typical samples in the 300W, CariFace, and ArtiFace datasets, each of which is annotated with landmarks. + +- ArtiFace Dataset. ArtiFace [55] contains 160 artistic portraits of 16 artists, which covers diverse artwork styles ranging from Renaissance to Comics. + +Each face in the datasets is annotated with 68 landmarks. Typical faces in the datasets and their landmarks are shown in Fig. 4. We can find that the faces in the three datasets have distinguished styles, which correspond to three different domains. In particular, compared to 300W [38], CariFace [4] exhibits abstract and exaggerated patterns, leading to large representation variations. ArtiFace [55] not only has larger variations across different artistic categories but also differs greatly in terms of the aspect of facial scales, orientations, locations, and so on. + +# 4.2. Learning Paradigms and Baselines + +Given the above datasets, we consider the following three learning paradigms: + +- Domain Adaptation (DA). Given labeled 300W faces and unlabeled stylized faces from CariFace or ArtiFace, we learn a face landmarker based on various domain adaptation methods. +- Generalized Zero-shot Learning (GZSL). In the challenging GZSL setting, we learn a face landmarker based on the above DA-based methods and test it in an unseen + +face domain (e.g., learning the landmarker on labeled 300W and unlabeled CariFace and testing on ArtiFace). + +- Oracle. In this setting, the labeled faces of all three datasets are accessible, and we can learn the face landmark by classic supervised learning. + +For a fair comparison, in each learning paradigm, we set the architecture of the face landmark based on the SLPT in [52]. Ideally, we would like to learn landmarks in the DA and GZSL settings, making its performance comparable to the oracle. In the oracle setting, we can learn the face landmark directly via classic supervised learning (SL), i.e., $\min_{\theta}\sum_{(\mathbf{X},\mathbf{Y})\sim \mathcal{D}}\| f_{\theta}(\mathbf{X}) - \mathbf{Y}\| _F^2$ . In the DA and GZSL settings, besides minimizing the landmark estimation errors, we can apply various image style transfer and domain adaptation methods, e.g., RevGrad [12], CycleGAN [59], BDL [25], AdaptSegNet [45] and FDA [54], to impose domain adaptation regularization during training. These methods work as the baselines of our method. + +Given the landmarkers learned by various methods, we evaluate them with the standard metric, Normalized Mean Error (NME). In Tab. 1, we show the training and testing data settings in the above three learning paradigms. Following existing work [9, 42, 52], we further split the 689 testing faces in 300W into 554 faces in common scenarios and 135 faces in challenging scenarios. The NMEs for the common, challenge, and full scenarios are recorded. + +# 4.3. Numerical and Visual Comparisons + +In Tab. 2, we show the performance of various learning methods in different settings, demonstrating the effectiveness and superiority of our method. In particular, existing DA methods often fail to improve the generalization power of model a lot in face landmarking tasks — their performance in target and unseen domains is inferior to that in the oracle setting, with a significant gap on NME. A potential reason for this phenomenon is that these methods focus on the adaptation of image domain and the landmark-related loss is not dominant in their learning processes. As a result, instead of learning the face landmarker, they make more efforts to optimize the parameters of other modules (e.g., the neural network-based face stylization modules and discriminators) during training. + +Table 2. Comparisons for various methods on their NMEs. In DA and GZSL settings, the best results are bold. + +
Learning ParadigmLearning Method300WCariFaceArtiFaceAverage NME
CommonChallengeFull
DA (300W→CariFace) and GZSL (Unseen ArtiFace)SL+RevGrad [12]2.845.583.3812.195.167.83
SL+CycleGAN [59]2.745.433.2712.114.707.70
SL+BDL [25]3.286.173.8413.635.658.77
SL+ASN [45]2.925.263.3812.214.757.80
SL+FDA [54]2.895.183.3412.664.608.07
Ours2.794.913.207.703.955.46
DA (300W→ArtiFace) and GZSL (Unseen CariFace)SL+RevGrad [12]2.995.813.5512.464.748.26
SL+CycleGAN [59]3.005.653.5212.645.348.36
SL+BDL [25]2.995.323.4413.405.908.73
SL+ASN [45]2.895.813.4616.585.6510.31
SL+FDA [54]3.056.213.6812.335.878.27
Ours2.905.143.3410.933.937.34
OracleSL2.684.863.105.483.314.36
+ +Table 3. The impacts of different model architectures on the NME performance of our method. In each setting, the best results are bold. + +
Learning ParadigmLearning Method300WCariFaceArtiFaceAverage NME
CommonChallengeFull
DA (300W→CariFace) and GZSL (Unseen ArtiFace)SBR [9]3.616.364.158.005.096.11
HRNet [47]2.935.293.407.734.335.60
SLPT [52] (Default)2.794.913.207.703.955.46
DA (300W→ArtiFace) and GZSL (Unseen CariFace)SBR [9]3.706.824.3211.355.328.04
HRNet [47]3.045.443.519.854.306.86
SLPT [52] (Default)2.905.143.3410.933.937.34
+ +![](images/0f8cab360045e690a2ef3f18f61a48a52f72e1c04aecf4ba1c142ec126f49549.jpg) +Figure 5. Visual comparisons for various methods in the two DA settings. We only highlight points on the inner lips in the enlarged region of the mouth in (a), as well as the eyes and the sides of the cheeks, excluding points on the eyebrows in (b). + +Different from the baselines, the performance of our model in the target domain (e.g., CariFace and ArtiFace) is improved consistently, while the performance in the source + +![](images/be662788e90a37c6c76e7c6d4cdb09bcb6bed248a4a8d98d513228704ab22472.jpg) +Figure 6. Visual comparisons for various methods in the GZSL (Unseen ArtiFace) setting. The zoomed-in area in (a) highlights the noses and facial contours, while that in (b) concentrates on the upper part of faces. + +domain does not degrade a lot. In both DA and GZSL settings, our method outperforms the baselines in most situations. Especially in the GZSL settings, the landmarks + +![](images/68c259f5c1c68b871a15d646c741b149739eef3874cb2d78b720ccb00cd59f25.jpg) +(a) 300W (Common) + +![](images/e704056cb9d7ab65b4a82d604e79c6df3f903bb4d7559b802ccfe96407c6b825.jpg) +(b) 300W (Challenge) + +![](images/616dba4ead884d1aedf03fb224f032c7ace69779ce82530ae5bc0863910a4685.jpg) +(c) 300W (Full) +Figure 7. Comparisons for different optimization strategies on NME performance in the GZSL (Unseen ArtiFace) setting. + +![](images/623d17bf8a17fe90e75948eb5ddd406f1058539ba3643be74823c7e87dd174f7.jpg) +(d) CariFace (Target) + +![](images/8af5e33f14e5a98d379b62e8b5a9a1aba8c3603882bb906513a3f8dd29f368f3.jpg) +(e) ArtiFace (Unseen) + +Table 4. Comparison with SOTA landmark detectors on NME. + +
Learning MethodTrain on 300WOur method
BackboneOPSPIGASTARSLPTSTARSLPT
DA on Cariface10.4611.2310.9711.057.627.70
GZSL on Artiface5.555.195.304.564.863.93
+ +obtained by our method show encouraging generalization power, which achieves lower NME than the baselines in the unseen domain (i.e., ArtiFace). Overall, the NME of our method in the source domain (i.e., 300W) is comparable to that achieved in the oracle setting. For the target even unseen domains, our method reduces the gap to the oracle. + +Besides the numerical comparisons, we provide some visualization results obtained by different methods in Figs. 5 and 6. For a complete comparison, the results of the SLPT trained only on $300\mathrm{W}$ are shown in the figures as well, which corresponds to learning a face landmarker only based on the data in the source domain. We can find that by applying our learning method, the face landmarker obtains better landmarks, which aligns with the ground truth with smaller errors, particularly for the landmarks of face contour, nose, mouth, and eyebrow. + +To further verify the generalization of our method, we select three more state-of-the-art landmarks, including OpenPose (OP) [7], SPIGA [36] and STAR [57] for numerical comparisons, and incorporate STAR as the backbone model of our method. Tab. 4 demonstrates that our learning method is applicable for various backbone models (e.g., STAR and SLPT), and integrating existing landmarkers with our learning method is an effective and competitive approach to enhance their generalizability. + +# 4.4. Analytic Experiments + +# 4.4.1 Joint v.s. Alternating Optimization + +As aforementioned, learning the face landmarker and the warping field model jointly makes it much easier to fall into undesired local optimum or unstable saddle points. To verify this claim, we apply the joint optimization strategy, i.e., + +solving (3) by optimizing $\theta$ and $\gamma$ jointly in each gradient descent step and compare its performance with ours. In Fig. 7, we show the best NME achieved by this joint optimization strategy and the NME achieved by our alternating optimization method in each iteration. We can find that the joint optimization strategy tends to overfit the source domain (300W), leading to power generalizability in the target and unseen domains. On the contrary, with the increase of iterations, the NMEs of our method on the target and unseen domains decrease consistently and become lower than those of the joint optimization after several iterations. + +# 4.4.2 Impacts of Model Architectures + +Besides the optimization strategy, the model architecture also has an impact on the model performance. By default, we implement the face landmarker as the SLPT model. In this experiment, we further explore the performance of other model architectures, including SBR [9] and HRNet [47]. Tab. 3 shows the performance of different model architectures in DA and GZSL settings. We can find that the SLPT-based face landmarker works better than the two competitors in this experiment. A potential reason for this phenomenon is that both SBR and HRNet are heatmap-based face landmarks. Given an input facial image, they output a heatmap indicating the distribution of landmarks rather than a set of deterministic landmark coordinates. We have to first detect the landmarks from the heatmap and then pass them through the warping field model. Accordingly, in the face warper optimization step (i.e., solving (4)), the landmarker and the warping field model cannot be trained in an end-to-end way because the backpropagation of the gradient becomes inapplicable. As a result, we have to update $\theta$ and $\gamma$ alternatively, leading to suboptimal performance. + +# 4.4.3 Effects on Loss Term + +In the context of our model's loss function, the image gradient field serves as its input. It has been observed through experiments that utilizing distinct gradient operators leads to the generation of varying image gradient fields, which in + +Table 5. Quantitative comparison on CariFace dataset. + +
OperatorGraySpatialLaplacianCannySobel
NME✓8.2827.8708.0057.695
×7.9587.9218.0117.863
+ +Table 6. Comparisons for different losses on NME. + +
Type of LossMSEPerceptualw.o. Grad MSEGrad MSE
300WCommon3.082.892.962.79
Challenge5.435.135.004.91
Full3.543.333.363.20
DA on Caricature8.268.239.487.70
GZSL on ArtiFace4.124.074.063.95
+ +turn influences the model's performance. Consequently, in this section, we assess and compare the training outcomes derived from image gradient fields obtained through the application of diverse gradient operators. + +Moreover, it is crucial to recognize that the choice of the input image, specifically whether it is a grayscale image or not, can also exert a certain influence on the results. As a result, we have incorporated a comparison of the outcomes when the input is a grayscale image. It is important to emphasize that all experiment settings are maintained consistently. The results are shown in Tab. 5. Utilizing grayscale images yields superior effects compared to not using grayscale images. This could be attributed to the fact that converting an image to grayscale eliminates the trivial effects of color and complex textures. The most optimal outcome is achieved by employing the Sobel operator. This may be due to the Sobel operator's emphasis on edge information within the image, enabling it to capture more geometric details. Additionally, the Sobel operator incorporates a mild smoothing effect during gradient computation, which mitigates the impact of noise. + +Furthermore, we consider $i)$ replacing the gradient MSE loss in (3) with the pixel MSE or perceptual loss and $ii)$ removing the gradient MSE loss. Tab. 6 shows the rationality of the gradient MSE for the reason that $i)$ landmarks are distributed on edges; and $ii)$ the gradient field filters out unnecessary color information, simplifying the task. These findings form the basis for our decision to adopt this particular method in our experimental process. + +# 4.4.4 Rationality of Proposed Warping Field Model + +Besides the model architecture of the face landmarker, the warping field model impacts our face landmarker as well — when the predicted warping field is inaccurate, we cannot obtain reliable pseudo landmarks for stylized faces. Therefore, we investigate different face warping models and + +![](images/359e98d3068b708fe1b84baf873d783d7b880e9df8b874149cb2d4625443f3b2.jpg) +(a) Source + +![](images/15b97daac31685b24a66ea760ece96cbd351e3af902952727e7c166893c970ee.jpg) +(b) Target + +![](images/0a715dfe02da865a1f9186db1f0443eee7771f7f8d602fe6cefb31b33ec90609.jpg) +(c) AutoToon + +![](images/54496ec481a0673426a460b7ff20c23f36bb5240a5712310e4d7b56e83c2b078.jpg) +(d) CariGANs + +![](images/cf322da2ddab6d762f1564a2e626fad3117d8c694b3b1cb1a5e04d48d543e9bc.jpg) +(e) RAFT + +![](images/494bc94405d2d2241ae6514047ff04b26435dba3e8593177a6c5a0a6ed87df98.jpg) +(f) Ours +Figure 8. Comparisons for different face warpers. + +demonstrate the rationality of the warping field model implemented in our work. In particular, applying different warping field models in our training process, we visualize their warping results in Fig. 8. We can find that although the face landmarker with the polyharmonic interpolation model leads to a very simple face warper, it outperforms many existing neural network-based image warping methods, such as AutoToon [15] and CariGANs [5] for facial manipulation and the optical flow methods like RAFT [44]. These methods either require one-to-one correspondence information between source and target images or assume the deformation between the two images to be slight, making them unsuitable for our problem, especially for the stylized faces with significant nonrigid deformations. + +# 5. Conclusion + +In this paper, we propose a simple but effective method for learning a generalizable face landmarker applicable to facial images with different styles. Given labeled real human faces and unlabeled stylized faces, our method learns the face landmarker under the guidance of conditional face warping, demonstrating the usefulness of the warping information. An alternating optimization framework is proposed to learn the face landmarker together with the warping field model. Experiments demonstrate the effectiveness of our method. Especially in the generalized zero-shot learning scenarios, our method achieves encouraging landmarking accuracy in unseen face domains. + +# Acknowledgements + +This work was supported in part by the National Natural Science Foundation of China (62102031), the foundation of Key Laboratory of Artificial Intelligence, Ministry of Education, P.R. China, and the Young Scholar Program (XSQD-202107001) from Beijing Institute of Technology. + +# References + +[1] Olalekan Agbolade, Azree Nazri, Razali Yaakob, Abdul Azim Abd Ghani, and Yoke Kqueen Cheah. Landmark-based homologous multi-point warping approach to 3d facial recognition using multiple datasets. PeerJ Computer Science, 6:e249, 2020. 1 +[2] Sadegh Aliakbarian, Pashmina Cameron, Federica Bogo, Andrew Fitzgibbon, and Thomas J. Cashman. Flag: Flow-based 3d avatar generation from sparse observations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13253-13262, 2022. 1 +[3] Peter N. Belhumeur, David W. Jacobs, David J. Kriegman, and Neeraj Kumar. Localizing parts of faces using a consensus of exemplars. In CVPR 2011, 2011. 4 +[4] Hongrui Cai, Yudong Guo, Zhuang Peng, and Juyong Zhang. Landmark detection and 3d face reconstruction for caricature using a nonlinear parametric model, 2020. 1, 4, 5 +[5] Kaidi Cao, Jing Liao, and Lu Yuan. Carigans: Unpaired photo-to-caricature translation. arXiv preprint arXiv:1811.00222, 2018. 1, 8 +[6] Xudong Cao, Yichen Wei, Fang Wen, and Jian Sun. Face alignment by explicit shape regression. International journal of computer vision, 107:177-190, 2014. 2 +[7] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7291-7299, 2017. 7 +[8] Forrester Cole, David Belanger, Dilip Krishnan, Aaron Sarna, Inbar Mosseri, and William T Freeman. Synthesizing normalized faces from facial identity features. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3703-3712, 2017. 3 +[9] Xuanyi Dong, Shouu-I Yu, Xinshuo Weng, Shih-En Wei, Yi Yang, and Yaser Sheikh. Supervision-by-registration: An unsupervised approach to improve the precision of facial landmark detectors. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 360–368, 2018. 2, 5, 6, 7 +[10] Alexey Dosovitskiy, Philipp Fischer, Eddy Ilg, Philip Hausser, Caner Hazirbas, Vladimir Golkov, Patrick Van Der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 2758-2766, 2015. 3 +[11] Pengfei Dou, Shishir K Shah, and Ioannis A Kakadiaris. End-to-end 3d face reconstruction with deep neural networks. In proceedings of the IEEE conference on computer vision and pattern recognition, pages 5908-5917, 2017. 1 +[12] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In International conference on machine learning, pages 1180-1189. PMLR, 2015. 1, 2, 5, 6 +[13] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. A neural algorithm of artistic style. arXiv preprint arXiv:1508.06576, 2015. 3 + +[14] C. A. Glasbey and K. V. Mardia. A review of image-warping methods. Journal of Applied Statistics, 25(2):155-171, 2002. 3 +[15] Julia Gong, Yannick Hold-Geoffroy, and Jingwan Lu. Auto-toon: Automatic geometric warping for face cartoon generation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 360–369, 2020. 1, 3, 8 +[16] Arthur Gretton, Karsten Borgwardt, Malte Rasch, Bernhard Schölkopf, and Alex Smola. A kernel method for the two-sample-problem. Advances in neural information processing systems, 19, 2006. 2 +[17] Guoliang Kang, Lu Jiang, Yi Yang, and Alexander G Hauptmann. Contrastive adaptation network for unsupervised domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4893-4902, 2019. 2 +[18] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 3 +[19] Vahid Kazemi and Josephine Sullivan. One millisecond face alignment with an ensemble of regression trees. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1867-1874, 2014. 1 +[20] SunnieS. Kim, NicholasI. Kolkin, Jason Salavon, and Gregory Shakhnarovich. Deformable style transfer, 2020. 2 +[21] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 4 +[22] Marek Kowalski, Jacek Naruniec, and Tomasz Trzcinski. Deep alignment network: A convolutional neural network for robust face alignment. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2017. 2 +[23] Vuong Le, Jonathan Brandt, Zhe Lin, Lubomir Bourdev, and Thomas S. Huang. Interactive Facial Feature Localization, page 679-692. 2012. 4 +[24] Hui Li, Zidong Guo, Seon-Min Rhee, Seungju Han, and Jae-Joon Han. Towards accurate facial landmark detection via cascaded transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4176-4185, 2022. 2 +[25] Yunsheng Li, Lu Yuan, and Nuno Vasconcelos. Bidirectional learning for domain adaptation of semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6936-6945, 2019. 1, 5, 6 +[26] Feng Liu, Dan Zeng, Qijun Zhao, and Xiaoming Liu. Joint face alignment and 3d face reconstruction. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14, pages 545-560. Springer, 2016. 1 +[27] Shaolei Liu, Xiaoyuan Luo, Kexue Fu, Manning Wang, and Zhijian Song. A learnable self-supervised task for unsupervised domain adaptation on point cloud classification and segmentation. Frontiers of Computer Science, 17(6):176708, 2023. 1, 2 + +[28] Weiyang Liu, Yandong Wen, Zhiding Yu, Ming Li, Bhiksha Raj, and Le Song. Spherface: Deep hypersphere embedding for face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 212-220, 2017. 1 +[29] Xiao-Chang Liu, Yong-Liang Yang, and Peter Hall. Learning to warp for style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3702-3711, 2021. 3 +[30] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In International conference on machine learning, pages 97-105. PMLR, 2015. 2 +[31] Mingsheng Long, Han Zhu, Jianmin Wang, and Michael I Jordan. Unsupervised domain adaptation with residual transfer networks. Advances in neural information processing systems, 29, 2016. 2 +[32] Jiangjing Lv, Xiaohu Shao, Junliang Xing, Cheng Cheng, and Xi Zhou. A deep regression architecture with two-stage re-initialization for high performance facial landmark detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3317-3326, 2017. 2 +[33] Iacopo Masi, Stephen Rawls, Gérard Medioni, and Prem Natarajan. Pose-aware face recognition in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4838-4846, 2016. 1 +[34] Kieron Messer, Jiri Matas, Josef Kittler, Juergen Luettin, Gilbert Maitre, et al. Xm2vtsdb: The extended m2vts database. In Second international conference on audio and video-based biometric person authentication, pages 965-966. CiteSeer, 1999. 4 +[35] Zhongyi Pei, Zhangjie Cao, Mingsheng Long, and Jianmin Wang. Multi-adversarial domain adaptation. In Thirty-second AAAI conference on artificial intelligence, 2018. 2 +[36] Andrés Prados-Torreblanca, José M Buenaposada, and Luis Baumela. Shape preserving facial landmarks with graph attention networks. arXiv preprint arXiv:2210.07233, 2022. 7 +[37] Joseph Roth, Yiying Tong, and Xiaoming Liu. Unconstrained 3d face reconstruction. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2606-2615, 2015. 1 +[38] Christos Sagonas, Georgios Tzimiropoulos, Stefanos Zafeiriou, and Maja Pantic. 300 faces in-the-wild challenge: The first facial landmark localization challenge. In 2013 IEEE International Conference on Computer Vision Workshops, 2014. 1, 4, 5 +[39] Ahmed Selim, Mohamed Elgharib, and Linda Doyle. Painting style transfer for head portraits using convolutional neural networks. ACM Transactions on Graphics (ToG), 35(4): 1-18, 2016. 3 +[40] Yichun Shi, Debayan Deb, and Anil K Jain. Warpgan: Automatic caricature generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10762-10771, 2019. 3 +[41] Baochen Sun, Jiashi Feng, and Kate Saenko. Correlation alignment for unsupervised domain adaptation. Domain + +adaptation in computer vision applications, pages 153-171, 2017. 2 +[42] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 5 +[43] Yaniv Taigman, Ming Yang, Marc'Aurelio Ranzato, and Lior Wolf. Deepface: Closing the gap to human-level performance in face verification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1701-1708, 2014. 1 +[44] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 3, 8 +[45] Yi-Hsuan Tsai, Wei-Chih Hung, Samuel Schulter, Kiyuh Yuk Sohn, Ming-Hsuan Yang, and Manmohan Chandraker. Learning to adapt structured output space for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7472-7481, 2018. 5, 6 +[46] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2 +[47] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 43(10):3349-3364, 2020. 2, 6, 7 +[48] Xinyao Wang, Liefeng Bo, and Li Fuxin. Adaptive wing loss for robust face alignment via heatmap regression. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 2 +[49] Kunhong Wu, Fan Jia, and Yahong Han. Domain-specific feature elimination: multi-source domain adaptation for image classification. Frontiers of Computer Science, 17(4): 174705, 2023. 2 +[50] Ruizheng Wu, Xiaodong Gu, Xin Tao, Xiaoyong Shen, Yu-Wing Tai, et al. Landmark assisted cyclegan for cartoon face generation. arXiv preprint arXiv:1907.01424, 2019. 3 +[51] Wayne Wu, Chen Qian, Shuo Yang, Quan Wang, Yici Cai, and Qiang Zhou. Look at boundary: A boundary-aware face alignment algorithm. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2129-2138, 2018. 1 +[52] Jiahao Xia, Weiwei Qu, Wenjian Huang, Jianguo Zhang, Xi Wang, and Min Xu. Sparse local patch transformer for robust face alignment and landmarks inherent relation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4052-4061, 2022. 1, 3, 5, 6 +[53] Jiaolong Yang, Peiran Ren, Dongqing Zhang, Dong Chen, Fang Wen, Hongdong Li, and Gang Hua. Neural aggregation network for video face recognition. In Proceedings of + +the IEEE conference on computer vision and pattern recognition, pages 4362-4371, 2017. 1 +[54] Yanchao Yang and Stefano Soatto. Fda: Fourier domain adaptation for semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4085-4095, 2020. 5, 6 +[55] Jordan Yaniv, Yael Newman, and Ariel Shamir. The face of art: landmark detection and geometric style in portraits. ACM Transactions on graphics (TOG), 38(4):1-15, 2019. 1, 2, 5 +[56] Jing Zhang, Zewei Ding, Wanqing Li, and Philip Ogunbona. Importance weighted adversarial nets for partial domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8156-8164, 2018. 2 +[57] Zhenglin Zhou, Huaxia Li, Hong Liu, Nanyang Wang, Gang Yu, and Rongrong Ji. Star loss: Reducing semantic ambiguity in facial landmark detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15475-15484, 2023. 7 +[58] Zhou, Erjin and Fan, Haoqiang and Cao, Zhimin and Jiang, Yuning and Yin, Qi. Extensive Facial Landmark Localization with Coarse-to-Fine Convolutional Network Cascade. In Proceedings of the IEEE International Conference on Computer Vision (ICCV) Workshops, 2013. 1, 2 +[59] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 1, 2, 5, 6 +[60] Xiangxin Zhu and Deva Ramanan. Face detection, pose estimation, and landmark localization in the wild. In 2012 IEEE conference on computer vision and pattern recognition, pages 2879-2886. IEEE, 2012. 4 +[61] Yi Zhu, Xindong Wu, Jipeng Qiang, Yunhao Yuan, and Yun Li. Representation learning via an integrated autoencoder for unsupervised domain adaptation. Frontiers of Computer Science, 17(5):175334, 2023. 1 \ No newline at end of file diff --git a/2024/Generalizable Face Landmarking Guided by Conditional Face Warping/images.zip b/2024/Generalizable Face Landmarking Guided by Conditional Face Warping/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fb0f7a5e0da92744f770af6fc6e9ac79f432127e --- /dev/null +++ b/2024/Generalizable Face Landmarking Guided by Conditional Face Warping/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5c7187d40da63f854aec51a7f109b7187a7aa0fd0f101963d089ae9c580c5d2f +size 786613 diff --git a/2024/Generalizable Face Landmarking Guided by Conditional Face Warping/layout.json b/2024/Generalizable Face Landmarking Guided by Conditional Face Warping/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1b226f39a9ad4de42e69701ffb68b950928b66e2 --- /dev/null +++ b/2024/Generalizable Face Landmarking Guided by Conditional Face Warping/layout.json @@ -0,0 +1,10099 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 75, + 103, + 518, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 103, + 518, + 121 + ], + "spans": [ + { + "bbox": [ + 75, + 103, + 518, + 121 + ], + "type": "text", + "content": "Generalizable Face Landmarking Guided by Conditional Face Warping" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 147, + 142, + 448, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 142, + 448, + 157 + ], + "spans": [ + { + "bbox": [ + 147, + 142, + 448, + 157 + ], + "type": "text", + "content": "Jiayi Liang" + }, + { + "bbox": [ + 147, + 142, + 448, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 147, + 142, + 448, + 157 + ], + "type": "text", + "content": " Haotian Liu" + }, + { + "bbox": [ + 147, + 142, + 448, + 157 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 147, + 142, + 448, + 157 + ], + "type": "text", + "content": " Hongteng Xu" + }, + { + "bbox": [ + 147, + 142, + 448, + 157 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 147, + 142, + 448, + 157 + ], + "type": "text", + "content": " Dixin Luo" + }, + { + "bbox": [ + 147, + 142, + 448, + 157 + ], + "type": "inline_equation", + "content": "^{1,4\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 88, + 157, + 507, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 157, + 507, + 172 + ], + "spans": [ + { + "bbox": [ + 88, + 157, + 507, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 88, + 157, + 507, + 172 + ], + "type": "text", + "content": "School of Computer Science and Technology, Beijing Institute of Technology, Beijing" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 171, + 488, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 171, + 488, + 186 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 488, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 105, + 171, + 488, + 186 + ], + "type": "text", + "content": "Gaoling School of Artificial Intelligence, Renmin University of China, Beijing" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 99, + 185, + 494, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 185, + 494, + 199 + ], + "spans": [ + { + "bbox": [ + 99, + 185, + 494, + 199 + ], + "type": "text", + "content": "3Beijing Key Laboratory of Big Data Management and Analysis Methods, Beijing" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 199, + 477, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 199, + 477, + 213 + ], + "spans": [ + { + "bbox": [ + 115, + 199, + 477, + 213 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 115, + 199, + 477, + 213 + ], + "type": "text", + "content": "Key Laboratory of Artificial Intelligence, Ministry of Education, Shanghai" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 215, + 483, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 215, + 483, + 228 + ], + "spans": [ + { + "bbox": [ + 106, + 215, + 483, + 228 + ], + "type": "text", + "content": "{jiayi.liang, haotianliu, dixin.luo}@bit.edu.cn, hongtengxu@ruc.edu.cn" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 255, + 192, + 267 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 255, + 192, + 267 + ], + "spans": [ + { + "bbox": [ + 143, + 255, + 192, + 267 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 281, + 290, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 290, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 290, + 593 + ], + "type": "text", + "content": "As a significant step for human face modeling, editing, and generation, face landmarking aims at extracting facial keypoints from images. A generalizable face landmarker is required in practice because real-world facial images, e.g., the avatars in animations and games, are often stylized in various ways. However, achieving generalizable face landmarking is challenging due to the diversity of facial styles and the scarcity of labeled stylized faces. In this study, we propose a simple but effective paradigm to learn a generalizable face landmarker based on labeled real human faces and unlabeled stylized faces. Our method learns the face landmarker as the key module of a conditional face warper. Given a pair of real and stylized facial images, the conditional face warper predicts a warping field from the real face to the stylized one, in which the face landmarker predicts the ending points of the warping field and provides us with high-quality pseudo landmarks for the corresponding stylized facial images. Applying an alternating optimization strategy, we learn the face landmarker to minimize i) the discrepancy between the stylized faces and the warped real ones and ii) the prediction errors of both real and pseudo landmarks. Experiments on various datasets show that our method outperforms existing state-of-the-art domain adaptation methods in face landmarking tasks, leading to a face landmarker with better generalizability. Code is available at https://plustwo0.github.io/project-face-landmarker." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 615, + 128, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 615, + 128, + 628 + ], + "spans": [ + { + "bbox": [ + 47, + 615, + 128, + 628 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 635, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 635, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 635, + 287, + 685 + ], + "type": "text", + "content": "Face landmarking seeks to extract human facial keypoints (e.g., eyes, nose, facial contour, and so on) from facial images. This task is important for many applications in the field of computer vision and graphics, such as face recogni" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 316, + 256, + 537, + 421 + ], + "blocks": [ + { + "bbox": [ + 316, + 256, + 537, + 421 + ], + "lines": [ + { + "bbox": [ + 316, + 256, + 537, + 421 + ], + "spans": [ + { + "bbox": [ + 316, + 256, + 537, + 421 + ], + "type": "image", + "image_path": "66c55eec947896835ab6c5a3a888c16ac0860ff886075170bc03387086dab6c3.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 427, + 547, + 506 + ], + "lines": [ + { + "bbox": [ + 304, + 427, + 547, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 427, + 547, + 506 + ], + "type": "text", + "content": "Figure 1. Both commercial software like Face++ and open-source method like SLPT [52] work well on landmarking real faces (e.g., those in 300W [38]) while achieving suboptimal performance when landmarking stylized faces (e.g., those in CariFace [4] and ArtiFace [55]). While existing domain adaptation method does not improve the performance significantly, our method achieves a generalizable face landmarker for various facial images." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 531, + 546, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 531, + 546, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 546, + 581 + ], + "type": "text", + "content": "tion [1, 28, 33, 43, 53], face stylization [5], and 3D face reconstruction [11, 26, 37]. Currently, many open-source and commercial face landmarks [19, 51, 58] have been developed and achieved encouraging performance in this task." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "Most existing face landmarks are designed and trained for landmarking real human faces, while the rapid development of AIGC applications, such as artistic character creation and cartoon generation [2, 15], leads to a massive increase in demands for landmarking stylized facial images. Unfortunately, as shown in Fig. 1, existing face landmarks often fail to landmark stylized facial images. Even if applying state-of-the-art domain adaptation strategies [12, 25, 27, 59, 61], the generalizability of the learned landmarks in the stylized facial image domain is still unsatisfactory. Essentially, traditional face landmarks work" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 236, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 236, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 236, + 703 + ], + "type": "text", + "content": "*The first two authors contributed equally to this work." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2425" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 286, + 264 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 286, + 264 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 286, + 264 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 286, + 264 + ], + "type": "image", + "image_path": "585c55b834d70529e61d8d9e43412b1b0772b11c4d293ddc9e534febc672ec3a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 273, + 287, + 297 + ], + "lines": [ + { + "bbox": [ + 46, + 273, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 46, + 273, + 287, + 297 + ], + "type": "text", + "content": "Figure 2. The scheme of our proposed method for learning a generalizable face landmarker." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 318, + 287, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 318, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 318, + 287, + 413 + ], + "type": "text", + "content": "well on real human faces because of the relatively stable geometry of real human faces and the sufficient labeled facial images. These two conditions, however, become questionable when landmarking stylized faces — the stylized facial images often have various facial styles, and manually landmarking such stylized faces is much more time-consuming than landmarking real faces. As a result, learning a generalizable face landmarker becomes a challenging task." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 414, + 288, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 414, + 288, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 414, + 288, + 676 + ], + "type": "text", + "content": "In this paper, we propose a simple but effective paradigm for learning a generalizable face landmarker, overcoming the challenges caused by the diversity of facial styles and the scarcity of labeled stylized faces. As illustrated in Fig. 2, given labeled real human facial images and unlabeled stylized facial images, we learn a face landmarker embedded in a conditional face warper. The face warper aims to deform real human faces according to stylized facial images, generating warped faces and corresponding warping fields. The face landmarker, as the key module of the warper, predicts the ending points of the warping fields and thus provides us with pseudo landmarks for the stylized facial images. The warping field is parametrized by a polyharmonic interpolation model. Under the guidance of the conditional face warping, we learn the face landmarker in an alternating optimization framework: The face landmarker is updated to " + }, + { + "bbox": [ + 46, + 414, + 288, + 676 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 414, + 288, + 676 + ], + "type": "text", + "content": " minimize the discrepancy between the stylized faces and the corresponding warped real human faces and " + }, + { + "bbox": [ + 46, + 414, + 288, + 676 + ], + "type": "inline_equation", + "content": "ii" + }, + { + "bbox": [ + 46, + 414, + 288, + 676 + ], + "type": "text", + "content": " minimize the prediction errors of both real and pseudo landmarks. In the first step, the face landmarker is learned associated with the warping field model, while in the second step, the face landmarker is updated with proximal regularization." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "content": "The extensive experiments in various face landmarking tasks demonstrate the effectiveness of learning method. The impacts of different loss functions and data settings on" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": "our learning method are analyzed through detailed ablation studies. Experimental results show that our learning method results in a face landmarker that is generalizable to the facial images with different styles, which outperforms representative domain adaptation methods consistently in various stylized face landmarking tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 156, + 392, + 168 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 156, + 392, + 168 + ], + "spans": [ + { + "bbox": [ + 306, + 156, + 392, + 168 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 176, + 419, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 176, + 419, + 190 + ], + "spans": [ + { + "bbox": [ + 306, + 176, + 419, + 190 + ], + "type": "text", + "content": "2.1. Face Landmarking" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 194, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 194, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 194, + 545, + 422 + ], + "type": "text", + "content": "Given facial images, the early face landmarking methods learn regression models to predict landmark coordinates directly [6, 32]. The models are often parameterized by neural networks like Transformer [46], capturing facial attributes that have been proven to be crucial for landmark prediction [24]. The coordinate regression is achieved in a coarse-to-fine framework [58], leading to a cascading landmarking pipeline. Recently, to make the landmark prediction robust to the variations in pose, scale, and occlusion, DAN [22] introduces the novel use of heatmaps and extracts features from the entire face rather than local patches around landmarks. SBR [9] utilizes the registration of synthesized images to provide supervisory signals for training. Adaptive Wing Loss [48] is proposed to address the imbalance between foreground pixels and background pixels by analysis of the main drawbacks of different loss functions. HRNet [47] produces high-resolution maps by connecting and exchanging information via merging multi-scale picture features across many branches." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 423, + 545, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 545, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 545, + 615 + ], + "type": "text", + "content": "As aforementioned, most existing face landmarks are learned for real human faces, which can not be adapted directly to stylized faces (e.g., cartoon and artistic faces). Essentially, treating labeled real human faces as a source domain and unlabeled stylized faces as a target domain, we can learn a generalizable face landmark by solving a domain adaptation (DA) problem [30, 31]. Accordingly, many DA techniques [27, 49] have potential in our problem, including the classic metric learning-based methods (e.g., CORAL [41], contrastive domain discrepancy [17], and maximum mean discrepancy [16]) and the recent adversarial learning-based methods [12, 35, 56, 59]. However, in the following content, we will show that directly applying these DA techniques often fails to achieve generalizable face landmarking because of the significant gap between the source and target face domains." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 623, + 395, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 395, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 395, + 635 + ], + "type": "text", + "content": "2.2. Face Warping" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 641, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 715 + ], + "type": "text", + "content": "Face warping is a technique that involves geometrically deforming source facial images to specified target shapes. The key step of this task is predicting a warping field between the source and target images that captures the shifts of image pixels. To achieve this aim, DST [20] and FoA [55] find matching keypoints between source and target images" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2426" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "content": "and then generate a dense warping field through data interpolation [8]. Instead of matching keypoints, some methods learn neural networks to predict dense warping fields directly based on paired images, e.g., Flownet [10], AutoToon [15], RAFT [44], and their variants [29]. However, these methods require the paired images to be similar to each other, which cannot capture significant deformations between the faces in different domains." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 167, + 288, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 167, + 288, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 167, + 288, + 264 + ], + "type": "text", + "content": "Compared to face stylization [13, 18, 39, 40], face warping is a relatively easier task because it only considers the deformation of shapes while ignoring the transfer of textures. However, it should be noted that this task is more relevant to face landmarking, in which the warping field provides us with strong evidence to shift face landmarks of source faces to target ones [50]. Inspired by such a strong correlation, we develop the proposed learning paradigm." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 273, + 153, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 273, + 153, + 286 + ], + "spans": [ + { + "bbox": [ + 47, + 273, + 153, + 286 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 293, + 287, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 287, + 477 + ], + "type": "text", + "content": "Denote " + }, + { + "bbox": [ + 46, + 293, + 287, + 477 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 46, + 293, + 287, + 477 + ], + "type": "text", + "content": " as the image space and " + }, + { + "bbox": [ + 46, + 293, + 287, + 477 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 46, + 293, + 287, + 477 + ], + "type": "text", + "content": " as the landmark space, respectively. In this work, we observe a set of labeled real human faces, i.e., " + }, + { + "bbox": [ + 46, + 293, + 287, + 477 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{(L)} = \\{\\pmb{X}_i^{(L)},\\pmb{Y}_i^{(L)}\\}_{i = 1}^{N_L}\\subset \\mathcal{X}_R\\times \\mathcal{V}" + }, + { + "bbox": [ + 46, + 293, + 287, + 477 + ], + "type": "text", + "content": " and a set of unlabeled stylized faces, i.e., " + }, + { + "bbox": [ + 46, + 293, + 287, + 477 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{(U)} = \\{\\pmb{X}_i^{(U)}\\}_{i = 1}^{N_U}\\subset" + }, + { + "bbox": [ + 46, + 293, + 287, + 477 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_S" + }, + { + "bbox": [ + 46, + 293, + 287, + 477 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 293, + 287, + 477 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_R,\\mathcal{X}_S\\subset \\mathcal{X}" + }, + { + "bbox": [ + 46, + 293, + 287, + 477 + ], + "type": "text", + "content": " correspond to the real and stylized face domains, respectively. Each " + }, + { + "bbox": [ + 46, + 293, + 287, + 477 + ], + "type": "inline_equation", + "content": "\\pmb {X}_i\\in \\mathbb{R}^{H\\times W\\times 3}" + }, + { + "bbox": [ + 46, + 293, + 287, + 477 + ], + "type": "text", + "content": " represents an image, and each " + }, + { + "bbox": [ + 46, + 293, + 287, + 477 + ], + "type": "inline_equation", + "content": "\\pmb {Y}_i = [\\pmb {y}_{i,k}]\\in \\mathbb{R}^{2\\times K}" + }, + { + "bbox": [ + 46, + 293, + 287, + 477 + ], + "type": "text", + "content": " records " + }, + { + "bbox": [ + 46, + 293, + 287, + 477 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 293, + 287, + 477 + ], + "type": "text", + "content": " face landmark coordinates, where " + }, + { + "bbox": [ + 46, + 293, + 287, + 477 + ], + "type": "inline_equation", + "content": "\\pmb {y}_{i,k}\\in \\mathbb{R}^2" + }, + { + "bbox": [ + 46, + 293, + 287, + 477 + ], + "type": "text", + "content": ". We aim to learn a face landmarker, denoted as " + }, + { + "bbox": [ + 46, + 293, + 287, + 477 + ], + "type": "inline_equation", + "content": "f_{\\theta}:\\mathcal{X}\\mapsto \\mathcal{V}" + }, + { + "bbox": [ + 46, + 293, + 287, + 477 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 293, + 287, + 477 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 293, + 287, + 477 + ], + "type": "text", + "content": " is the model parameter. The model should be able to predict face landmarks from facial images and moreover, generalize to both " + }, + { + "bbox": [ + 46, + 293, + 287, + 477 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_R" + }, + { + "bbox": [ + 46, + 293, + 287, + 477 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 293, + 287, + 477 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_S" + }, + { + "bbox": [ + 46, + 293, + 287, + 477 + ], + "type": "text", + "content": ". To achieve this aim, we embed the face landmarker into a conditional face warper and learn it associated with a parametric warping field predictor in an alternating optimization framework, as illustrated in Fig. 2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 483, + 280, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 483, + 280, + 496 + ], + "spans": [ + { + "bbox": [ + 47, + 483, + 280, + 496 + ], + "type": "text", + "content": "3.1. Face Landmarking Guided by Face Warping" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 501, + 286, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 501, + 286, + 607 + ], + "spans": [ + { + "bbox": [ + 46, + 501, + 286, + 607 + ], + "type": "text", + "content": "In this study, we take the SLPT model [52] as the backbone of our face landmarker. Given a stylized face " + }, + { + "bbox": [ + 46, + 501, + 286, + 607 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_i^{(U)}" + }, + { + "bbox": [ + 46, + 501, + 286, + 607 + ], + "type": "text", + "content": ", the face landmarker predicts its landmarks as " + }, + { + "bbox": [ + 46, + 501, + 286, + 607 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{Y}}_i^{(U)} = [\\widehat{\\mathbf{y}}_{i,k}^{(U)}] = f_\\theta(\\mathbf{X}_i^{(U)})" + }, + { + "bbox": [ + 46, + 501, + 286, + 607 + ], + "type": "text", + "content": ". At the same time, we can sample a labeled real face " + }, + { + "bbox": [ + 46, + 501, + 286, + 607 + ], + "type": "inline_equation", + "content": "(\\mathbf{X}_j^{(L)},\\mathbf{Y}_j^{(L)}) \\sim \\mathcal{D}^L" + }, + { + "bbox": [ + 46, + 501, + 286, + 607 + ], + "type": "text", + "content": ". Treating the labeled and predicted landmarks as keypoints, we can model a warping field from the real face to the stylized one by the following polyharmonic interpolation model [14]:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 612, + 287, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 612, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 53, + 612, + 287, + 634 + ], + "type": "interline_equation", + "content": "w _ {i, \\gamma} (\\boldsymbol {y}) = \\sum_ {k = 1} ^ {K} \\omega_ {k} \\phi \\left(\\left\\| \\boldsymbol {y} - \\hat {\\boldsymbol {y}} _ {i, k} ^ {(U)} \\right\\| _ {2}\\right) + \\boldsymbol {V} \\boldsymbol {y} + \\boldsymbol {b}, \\tag {1}", + "image_path": "92b8c3cf1f54b6f10b62f69a82d4ca6f70c7fcbe1193af2cf01a77c1960d5123.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 639, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 639, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 639, + 288, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 639, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\gamma = \\{\\{\\pmb{\\omega}_k\\in \\mathbb{R}^2\\}_{k = 1}^K,\\pmb {V}\\in \\mathbb{R}^{2\\times 2},\\pmb {b}\\in \\mathbb{R}^2\\}" + }, + { + "bbox": [ + 46, + 639, + 288, + 715 + ], + "type": "text", + "content": " correspond to the parameters of the warping field. As shown in (1), the vector " + }, + { + "bbox": [ + 46, + 639, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 46, + 639, + 288, + 715 + ], + "type": "text", + "content": " denotes the u-v coordinate of a pixel in the stylized facial image, and " + }, + { + "bbox": [ + 46, + 639, + 288, + 715 + ], + "type": "inline_equation", + "content": "w_{i,\\gamma}(\\pmb {y})" + }, + { + "bbox": [ + 46, + 639, + 288, + 715 + ], + "type": "text", + "content": " gives the inverse mapping from the pixel " + }, + { + "bbox": [ + 46, + 639, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 46, + 639, + 288, + 715 + ], + "type": "text", + "content": " to a coordinate in the real human facial image, conditioned on " + }, + { + "bbox": [ + 46, + 639, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{X}_i^{(U)}" + }, + { + "bbox": [ + 46, + 639, + 288, + 715 + ], + "type": "text", + "content": ". The first term" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 70, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\sum_{k=1}^{K} \\omega_k \\phi(\\| \\pmb{y} - \\hat{\\pmb{y}}_{i,k}^{(U)} \\|_2)" + }, + { + "bbox": [ + 305, + 70, + 545, + 133 + ], + "type": "text", + "content": " achieves nonparametric regression for modeling nonrigid deformations, in which " + }, + { + "bbox": [ + 305, + 70, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\phi(r)" + }, + { + "bbox": [ + 305, + 70, + 545, + 133 + ], + "type": "text", + "content": " is a predefined thin-plate spline function. The second term " + }, + { + "bbox": [ + 305, + 70, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\pmb{V}\\pmb{y} + \\pmb{b}" + }, + { + "bbox": [ + 305, + 70, + 545, + 133 + ], + "type": "text", + "content": " is a linear parametric model capturing the rigid transformation of " + }, + { + "bbox": [ + 305, + 70, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 305, + 70, + 545, + 133 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 133, + 545, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 545, + 223 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 545, + 223 + ], + "type": "text", + "content": "For each pixel coordinate " + }, + { + "bbox": [ + 304, + 133, + 545, + 223 + ], + "type": "inline_equation", + "content": "\\pmb{y} \\in \\{1, \\dots, H\\} \\times \\{1, \\dots, W\\}" + }, + { + "bbox": [ + 304, + 133, + 545, + 223 + ], + "type": "text", + "content": ", we can trace it back to the real human facial image based on " + }, + { + "bbox": [ + 304, + 133, + 545, + 223 + ], + "type": "inline_equation", + "content": "w_{i,\\gamma}(\\pmb{y})" + }, + { + "bbox": [ + 304, + 133, + 545, + 223 + ], + "type": "text", + "content": " and obtain the pixel color as " + }, + { + "bbox": [ + 304, + 133, + 545, + 223 + ], + "type": "inline_equation", + "content": "\\pmb{X}_j^{(L)}(w_{i,\\gamma}(\\pmb{y}))" + }, + { + "bbox": [ + 304, + 133, + 545, + 223 + ], + "type": "text", + "content": ". Accordingly, with the grid sampler constructed via inverse mapping function " + }, + { + "bbox": [ + 304, + 133, + 545, + 223 + ], + "type": "inline_equation", + "content": "w_{i,\\gamma}" + }, + { + "bbox": [ + 304, + 133, + 545, + 223 + ], + "type": "text", + "content": ", we obtain the warped real human facial image conditioned on " + }, + { + "bbox": [ + 304, + 133, + 545, + 223 + ], + "type": "inline_equation", + "content": "\\pmb{X}_i^{(U)}" + }, + { + "bbox": [ + 304, + 133, + 545, + 223 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 304, + 133, + 545, + 223 + ], + "type": "inline_equation", + "content": "\\widehat{\\pmb{X}}_{j|i}^{(L)}" + }, + { + "bbox": [ + 304, + 133, + 545, + 223 + ], + "type": "text", + "content": ". For " + }, + { + "bbox": [ + 304, + 133, + 545, + 223 + ], + "type": "inline_equation", + "content": "\\pmb{y} \\in \\{1, \\dots, H\\} \\times \\{1, \\dots, W\\}" + }, + { + "bbox": [ + 304, + 133, + 545, + 223 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 133, + 545, + 223 + ], + "type": "inline_equation", + "content": "j = 1, \\dots, N_L" + }, + { + "bbox": [ + 304, + 133, + 545, + 223 + ], + "type": "text", + "content": ", we have" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 357, + 227, + 545, + 246 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 227, + 545, + 246 + ], + "spans": [ + { + "bbox": [ + 357, + 227, + 545, + 246 + ], + "type": "interline_equation", + "content": "\\widehat {\\boldsymbol {X}} _ {j \\mid i} ^ {(L)} (\\boldsymbol {y}) = \\boldsymbol {X} _ {j} ^ {(L)} \\left(w _ {i, \\gamma} (\\boldsymbol {y})\\right). \\tag {2}", + "image_path": "9ca99e18c1ea7c2e922f24eb3a95faf3cbc692838d48ffdbc63cdb6e9a8f9d25.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 250, + 545, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 250, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 304, + 250, + 545, + 369 + ], + "type": "text", + "content": "Unlike WarpGAN [40], which generates the warped face by predicting dense keypoints and their displacements by two fully-connected layers during training, we directly use the predicted and observed landmarks to define sparse displacements and estimate other pixels' displacements by spline-based interpolation, which improves computational efficiency significantly. Moreover, by applying the warping field model with limited degree-of-freedom (i.e., few learnable parameters), we can focus more on the learning of the face landmarker in the training phase." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 370, + 545, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 370, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 304, + 370, + 545, + 417 + ], + "type": "text", + "content": "Specifically, the warped face together with the warping field provides a useful guidance for the learning of the face landmarker. In particular, we formulate the learning problem of the face landmarker as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 322, + 421, + 545, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 421, + 545, + 539 + ], + "spans": [ + { + "bbox": [ + 322, + 421, + 545, + 539 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min _ {\\theta , \\gamma} \\underbrace {\\sum_ {j = 1} ^ {N _ {L}} \\| f _ {\\theta} \\left(\\boldsymbol {X} _ {j} ^ {(L)}\\right) - \\boldsymbol {Y} _ {j} ^ {(L)} \\| _ {F} ^ {2}} _ {\\text {L a n d m a r k i n g e r r o r i n t h e s o u r c e d o m a i n}} \\\\ + \\underbrace {\\sum_ {i = 1} ^ {N _ {U}} \\sum_ {j = 1} ^ {N _ {L}} \\left\\| \\nabla \\widehat {\\boldsymbol {X}} _ {j | i} ^ {(L)} - \\nabla \\boldsymbol {X} _ {i} ^ {(U)} \\right\\| _ {F} ^ {2}} _ {\\text {D i s c r e p a n c y o f i m a g e g r a d i e n t}} \\tag {3} \\\\ + \\underbrace {\\sum_ {i = 1} ^ {N _ {U}} \\sum_ {j = 1} ^ {N _ {L}} \\| w _ {i , \\gamma} (\\widehat {\\mathbf {Y}} _ {i} ^ {(U)}) - \\mathbf {Y} _ {j} ^ {(L)} \\| _ {F} ^ {2}} _ {\\text {L a n d m a r k w a r p i n g e r r o r}}, \\\\ \\end{array}", + "image_path": "b868857e9b5bf5cb73512c753162b8f5607af631986c2ee777adad555b9511c0.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 544, + 545, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 545, + 654 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 544, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\| _F" + }, + { + "bbox": [ + 304, + 544, + 545, + 654 + ], + "type": "text", + "content": " represents the Frobenius norm of matrix. In (3), the first term is the landmarking error for real faces, which corresponds to the data fidelity loss in the source domain. The second term measures the discrepancy between the stylized face and the warped real face in the gradient field, in which the gradient operation " + }, + { + "bbox": [ + 304, + 544, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\nabla" + }, + { + "bbox": [ + 304, + 544, + 545, + 654 + ], + "type": "text", + "content": " is implemented by the Sobel operator. The third term is the landmark warping error. Both the second and third terms are determined jointly by the landmarker " + }, + { + "bbox": [ + 304, + 544, + 545, + 654 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 304, + 544, + 545, + 654 + ], + "type": "text", + "content": " and warping field model " + }, + { + "bbox": [ + 304, + 544, + 545, + 654 + ], + "type": "inline_equation", + "content": "w_{i,\\gamma}" + }, + { + "bbox": [ + 304, + 544, + 545, + 654 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 659, + 490, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 659, + 490, + 672 + ], + "spans": [ + { + "bbox": [ + 306, + 659, + 490, + 672 + ], + "type": "text", + "content": "3.2. Alternating Optimization Strategy" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": "The optimization problem in (3) is non-convex because the landmarker is implemented by a neural network and is coupled to the warping field model. As a result, learning " + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "2427" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 79, + 97, + 125 + ], + "blocks": [ + { + "bbox": [ + 51, + 79, + 97, + 125 + ], + "lines": [ + { + "bbox": [ + 51, + 79, + 97, + 125 + ], + "spans": [ + { + "bbox": [ + 51, + 79, + 97, + 125 + ], + "type": "image", + "image_path": "b2e949308641b6c3c5184263e60a6bee22e42be3fc42af04f0f2207e4b5845cf.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 99, + 128, + 205, + 135 + ], + "lines": [ + { + "bbox": [ + 99, + 128, + 205, + 135 + ], + "spans": [ + { + "bbox": [ + 99, + 128, + 205, + 135 + ], + "type": "text", + "content": "Warped Face with Warped Landmarks" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 98, + 80, + 143, + 125 + ], + "blocks": [ + { + "bbox": [ + 98, + 80, + 143, + 125 + ], + "lines": [ + { + "bbox": [ + 98, + 80, + 143, + 125 + ], + "spans": [ + { + "bbox": [ + 98, + 80, + 143, + 125 + ], + "type": "image", + "image_path": "d7db80611db195609997eb5cd87fdc0da8d47ff03d9995835ccc1b87ec45b688.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 144, + 80, + 190, + 125 + ], + "blocks": [ + { + "bbox": [ + 144, + 80, + 190, + 125 + ], + "lines": [ + { + "bbox": [ + 144, + 80, + 190, + 125 + ], + "spans": [ + { + "bbox": [ + 144, + 80, + 190, + 125 + ], + "type": "image", + "image_path": "2799767401c9c88f2bcad3101714f5661bc61eb91b9976204df0e645bf76233f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 190, + 80, + 236, + 125 + ], + "blocks": [ + { + "bbox": [ + 190, + 80, + 236, + 125 + ], + "lines": [ + { + "bbox": [ + 190, + 80, + 236, + 125 + ], + "spans": [ + { + "bbox": [ + 190, + 80, + 236, + 125 + ], + "type": "image", + "image_path": "9756a62a1b30f6c9a3900aace7257873629c25a0ab32b9a23c71b796fdd7aa8d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 237, + 80, + 283, + 125 + ], + "blocks": [ + { + "bbox": [ + 237, + 80, + 283, + 125 + ], + "lines": [ + { + "bbox": [ + 237, + 80, + 283, + 125 + ], + "spans": [ + { + "bbox": [ + 237, + 80, + 283, + 125 + ], + "type": "image", + "image_path": "b4e305fe09aecd53102671f37ace164c08b8d0fc1adf6643c5bd69489be3a340.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 97, + 138, + 143, + 184 + ], + "blocks": [ + { + "bbox": [ + 97, + 138, + 143, + 184 + ], + "lines": [ + { + "bbox": [ + 97, + 138, + 143, + 184 + ], + "spans": [ + { + "bbox": [ + 97, + 138, + 143, + 184 + ], + "type": "image", + "image_path": "3b2ec01642ccf3f00d47fadc7173d258c25a78116ad57632b544c60de13cfebb.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 194, + 287, + 239 + ], + "lines": [ + { + "bbox": [ + 46, + 194, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 287, + 239 + ], + "type": "text", + "content": "Figure 3. Illustrations of conditional face warping results. Taking a cartoon face as the target, our model warps real human faces accordingly. The red dots indicate real human face landmarks, and green dots indicate cartoon and warped face landmarks." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 144, + 138, + 190, + 184 + ], + "blocks": [ + { + "bbox": [ + 144, + 138, + 190, + 184 + ], + "lines": [ + { + "bbox": [ + 144, + 138, + 190, + 184 + ], + "spans": [ + { + "bbox": [ + 144, + 138, + 190, + 184 + ], + "type": "image", + "image_path": "c642614f178cfca3d97bdd1f4840cdf6807463e1fe737706398ea34431f6e0c4.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 190, + 138, + 236, + 184 + ], + "blocks": [ + { + "bbox": [ + 190, + 138, + 236, + 184 + ], + "lines": [ + { + "bbox": [ + 190, + 138, + 236, + 184 + ], + "spans": [ + { + "bbox": [ + 190, + 138, + 236, + 184 + ], + "type": "image", + "image_path": "678ec486c5029a74aba8d95f0948a1fee8e4256f96951544bbb3be72a35f21d2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 237, + 138, + 283, + 184 + ], + "blocks": [ + { + "bbox": [ + 237, + 138, + 283, + 184 + ], + "lines": [ + { + "bbox": [ + 237, + 138, + 283, + 184 + ], + "spans": [ + { + "bbox": [ + 237, + 138, + 283, + 184 + ], + "type": "image", + "image_path": "213138cb5e4acc88168d0514a3dc3c26031dcc07841dd9a2e2615a3a432759fc.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 259, + 287, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 259, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 259, + 287, + 319 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 259, + 287, + 319 + ], + "type": "text", + "content": " jointly often falls into an undesired local optimum even an unstable saddle point. To mitigate this issue, we propose an alternating optimization framework. In principle, we can decompose the optimization problem in (3) into the following two subproblems and solve them iteratively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 321, + 287, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 321, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 47, + 321, + 287, + 346 + ], + "type": "text", + "content": "- Face Warper Optimization: The first subproblem corresponds to the optimization of the face warper, i.e.," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 352, + 292, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 352, + 292, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 352, + 292, + 396 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\theta^ {(1)}, \\gamma^ {(1)} = \\arg \\min _ {\\theta , \\gamma} \\sum_ {i, j} \\| \\nabla \\widehat {\\boldsymbol {X}} _ {j | i} ^ {(L)} - \\nabla \\boldsymbol {X} _ {i} ^ {(U)} \\| _ {F} ^ {2} \\tag {4} \\\\ + \\sum_ {i, j} \\| w _ {i, \\gamma} (\\widehat {\\pmb {Y}} _ {i} ^ {(U)}) - \\pmb {Y} _ {j} ^ {(L)} \\| _ {F} ^ {2}. \\\\ \\end{array}", + "image_path": "08f399d10b1d24bc76b08744b17aeb8d918b18c2ecfcf38de8dc539eddfd21c8.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 402, + 287, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 402, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 402, + 287, + 462 + ], + "type": "text", + "content": "In this subproblem, we only care about whether the real human faces can be warped as the stylized faces with high accuracy, so the term of landmarking error is ignored. We solve this problem by Adam [21]: in each step, we update " + }, + { + "bbox": [ + 55, + 402, + 287, + 462 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 402, + 287, + 462 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 402, + 287, + 462 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 402, + 287, + 462 + ], + "type": "text", + "content": " based on a batch of randomly-sampled face pairs." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 462, + 287, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 462, + 287, + 512 + ], + "spans": [ + { + "bbox": [ + 47, + 462, + 287, + 512 + ], + "type": "text", + "content": "- Proximal Face Landmarker Optimization: Given " + }, + { + "bbox": [ + 47, + 462, + 287, + 512 + ], + "type": "inline_equation", + "content": "\\theta^{(1)}" + }, + { + "bbox": [ + 47, + 462, + 287, + 512 + ], + "type": "text", + "content": " and the predicted landmarks (i.e., " + }, + { + "bbox": [ + 47, + 462, + 287, + 512 + ], + "type": "inline_equation", + "content": "\\widehat{\\pmb{Y}}_i^{(U)} = f_{\\theta^{(1)}}(\\pmb{X}_i^{(U)})" + }, + { + "bbox": [ + 47, + 462, + 287, + 512 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 47, + 462, + 287, + 512 + ], + "type": "inline_equation", + "content": "i = 1, \\dots, N_L" + }, + { + "bbox": [ + 47, + 462, + 287, + 512 + ], + "type": "text", + "content": "), we can treat " + }, + { + "bbox": [ + 47, + 462, + 287, + 512 + ], + "type": "inline_equation", + "content": "\\theta^{(1)}" + }, + { + "bbox": [ + 47, + 462, + 287, + 512 + ], + "type": "text", + "content": " as the initial variable and optimize it with a proximal regularizer:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 64, + 519, + 287, + 570 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 519, + 287, + 570 + ], + "spans": [ + { + "bbox": [ + 64, + 519, + 287, + 570 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\theta^ {(2)} = \\arg \\min _ {\\theta} \\sum_ {j = 1} ^ {N _ {L}} \\| f _ {\\theta} (\\boldsymbol {X} _ {j} ^ {(L)}) - \\boldsymbol {Y} _ {j} ^ {(L)} \\| _ {F} ^ {2} \\\\ + \\underbrace {\\sum_ {i = 1} ^ {N _ {U}} \\| f _ {\\theta} \\left(\\boldsymbol {X} _ {i} ^ {(U)}\\right) - \\widehat {\\boldsymbol {Y}} _ {i} ^ {(U)} \\| _ {F} ^ {2}}. \\tag {5} \\\\ \\end{array}", + "image_path": "30e84a3d6cccf8c64537058cbaca44e52151679b4a4085b1dd058009172d4afe.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 113, + 571, + 246, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 571, + 246, + 580 + ], + "spans": [ + { + "bbox": [ + 113, + 571, + 246, + 580 + ], + "type": "text", + "content": "Pseudo landmarking error in the target domain" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 588, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 287, + 664 + ], + "type": "text", + "content": "Here, the second term in (5) measures the estimation errors of the pseudo landmarks achieved in the previous step. Essentially, it works as a proximal regularizer, ensuring that the optimized landmarks " + }, + { + "bbox": [ + 55, + 588, + 287, + 664 + ], + "type": "inline_equation", + "content": "f_{\\theta^{(2)}}(\\mathbf{X}_i^{(U)})" + }, + { + "bbox": [ + 55, + 588, + 287, + 664 + ], + "type": "text", + "content": " is not too far away from the previous estimation " + }, + { + "bbox": [ + 55, + 588, + 287, + 664 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{Y}}_i^{(U)}" + }, + { + "bbox": [ + 55, + 588, + 287, + 664 + ], + "type": "text", + "content": ". Similarly, we can solve this problem by Adam [21] as well." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "Fig. 3 shows the warping effect on real human faces achieved by solving (4). In Fig. 3, the first row shows the real human faces with landmarks and the target stylized face, and the second row shows the warping results," + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 72, + 545, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 84 + ], + "type": "text", + "content": "Algorithm 1 Proposed learning scheme of face landmarker" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 89, + 545, + 125 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 89, + 545, + 125 + ], + "spans": [ + { + "bbox": [ + 307, + 89, + 545, + 125 + ], + "type": "text", + "content": "Require: Labeled real faces " + }, + { + "bbox": [ + 307, + 89, + 545, + 125 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{(L)}" + }, + { + "bbox": [ + 307, + 89, + 545, + 125 + ], + "type": "text", + "content": " and unlabeled stylized faces " + }, + { + "bbox": [ + 307, + 89, + 545, + 125 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{(U)}" + }, + { + "bbox": [ + 307, + 89, + 545, + 125 + ], + "type": "text", + "content": ". The number of iterations (i.e., " + }, + { + "bbox": [ + 307, + 89, + 545, + 125 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 307, + 89, + 545, + 125 + ], + "type": "text", + "content": "). Epochs for the subproblems (i.e., " + }, + { + "bbox": [ + 307, + 89, + 545, + 125 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 307, + 89, + 545, + 125 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 307, + 89, + 545, + 125 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 307, + 89, + 545, + 125 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 125, + 545, + 271 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 312, + 125, + 545, + 149 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 125, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 312, + 125, + 545, + 149 + ], + "type": "text", + "content": "1: Initialize " + }, + { + "bbox": [ + 312, + 125, + 545, + 149 + ], + "type": "inline_equation", + "content": "\\{\\gamma^{(0)}\\}" + }, + { + "bbox": [ + 312, + 125, + 545, + 149 + ], + "type": "text", + "content": " with a pretrained model on " + }, + { + "bbox": [ + 312, + 125, + 545, + 149 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{(L)}" + }, + { + "bbox": [ + 312, + 125, + 545, + 149 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 312, + 125, + 545, + 149 + ], + "type": "inline_equation", + "content": "\\{\\theta^{(0)}\\}" + }, + { + "bbox": [ + 312, + 125, + 545, + 149 + ], + "type": "text", + "content": " randomly." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 150, + 426, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 150, + 426, + 159 + ], + "spans": [ + { + "bbox": [ + 312, + 150, + 426, + 159 + ], + "type": "text", + "content": "2: for " + }, + { + "bbox": [ + 312, + 150, + 426, + 159 + ], + "type": "inline_equation", + "content": "m = 0, \\dots, M - 1" + }, + { + "bbox": [ + 312, + 150, + 426, + 159 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 160, + 500, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 160, + 500, + 175 + ], + "spans": [ + { + "bbox": [ + 312, + 160, + 500, + 175 + ], + "type": "text", + "content": "3: Sample a batch " + }, + { + "bbox": [ + 312, + 160, + 500, + 175 + ], + "type": "inline_equation", + "content": "\\{\\pmb{X}_i^{(U)},\\pmb{X}_i^{(L)},\\pmb{Y}_i^{(L)}\\}_{i = 1}^N" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 175, + 448, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 175, + 448, + 186 + ], + "spans": [ + { + "bbox": [ + 312, + 175, + 448, + 186 + ], + "type": "text", + "content": "4: Face warper optimization:" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 186, + 545, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 186, + 545, + 210 + ], + "spans": [ + { + "bbox": [ + 312, + 186, + 545, + 210 + ], + "type": "text", + "content": "5: Take " + }, + { + "bbox": [ + 312, + 186, + 545, + 210 + ], + "type": "inline_equation", + "content": "\\theta^{(2m)},\\gamma^{(m)}" + }, + { + "bbox": [ + 312, + 186, + 545, + 210 + ], + "type": "text", + "content": " as the initialization, then solve (4) by Adam with " + }, + { + "bbox": [ + 312, + 186, + 545, + 210 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 312, + 186, + 545, + 210 + ], + "type": "text", + "content": " epochs and obtain " + }, + { + "bbox": [ + 312, + 186, + 545, + 210 + ], + "type": "inline_equation", + "content": "\\theta^{(2m + 1)}" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 211, + 507, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 211, + 507, + 221 + ], + "spans": [ + { + "bbox": [ + 312, + 211, + 507, + 221 + ], + "type": "text", + "content": "6: Proximal face landmarker optimization:" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 222, + 545, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 222, + 545, + 246 + ], + "spans": [ + { + "bbox": [ + 312, + 222, + 545, + 246 + ], + "type": "text", + "content": "7: Take " + }, + { + "bbox": [ + 312, + 222, + 545, + 246 + ], + "type": "inline_equation", + "content": "\\theta^{(2m + 1)}" + }, + { + "bbox": [ + 312, + 222, + 545, + 246 + ], + "type": "text", + "content": " as the initialization, then solve (5) by Adam with " + }, + { + "bbox": [ + 312, + 222, + 545, + 246 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 312, + 222, + 545, + 246 + ], + "type": "text", + "content": " epochs and obtain " + }, + { + "bbox": [ + 312, + 222, + 545, + 246 + ], + "type": "inline_equation", + "content": "\\theta^{(2m + 2)}" + }, + { + "bbox": [ + 312, + 222, + 545, + 246 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 247, + 358, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 247, + 358, + 257 + ], + "spans": [ + { + "bbox": [ + 312, + 247, + 358, + 257 + ], + "type": "text", + "content": "8: end for" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 312, + 258, + 545, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 258, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 312, + 258, + 545, + 271 + ], + "type": "text", + "content": "9: return Output a generalizable face landmarker " + }, + { + "bbox": [ + 312, + 258, + 545, + 271 + ], + "type": "inline_equation", + "content": "f_{\\theta^{(2M)}}" + }, + { + "bbox": [ + 312, + 258, + 545, + 271 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 295, + 545, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 295, + 545, + 439 + ], + "spans": [ + { + "bbox": [ + 304, + 295, + 545, + 439 + ], + "type": "text", + "content": "in which the green dots are predicted landmarks. These results empirically demonstrate the rationality of our alternating optimization framework. In particular, we can find that solving (4) leads to reasonable warping results, which are similar to the target stylized face on shape. The similarity on face shape indicates that the predicted landmarks can be treated as reliable pseudo labels of the stylized face, which can be used to construct the proximal regularizer that penalizing the peso do landmarking errors in the target domain. Repeating the above two steps till converge, we obtain the target face landmarker that is generalizable for both real and stylized faces. Algorithm 1 shows the learning scheme." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 306, + 450, + 382, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 450, + 382, + 464 + ], + "spans": [ + { + "bbox": [ + 306, + 450, + 382, + 464 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 304, + 471, + 545, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 471, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 471, + 545, + 590 + ], + "type": "text", + "content": "We apply our learning method to learn a face landmarker and test it on landmarking faces with various styles. Extensive experiments, including comparisons with baselines and analytic ablation studies, demonstrate the effectiveness of our learning method and the generalizability of the corresponding face landmarker. All the experiments are conducted on a single NVIDIA 3090 GPU. Representative experimental results are shown below. More experimental results and implementation details are given in the supplementary material." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 306, + 598, + 364, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 598, + 364, + 610 + ], + "spans": [ + { + "bbox": [ + 306, + 598, + 364, + 610 + ], + "type": "text", + "content": "4.1. Dataset" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 304, + 617, + 544, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 544, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 544, + 641 + ], + "type": "text", + "content": "In this study, we conduct experiments based on the following three commonly-used face datasets." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 306, + 642, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 40, + "blocks": [ + { + "bbox": [ + 306, + 642, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 642, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 306, + 642, + 545, + 677 + ], + "type": "text", + "content": "- 300W Dataset. 300W [38] is comprised of five well-known real human face datasets including LFPW [3], AFW [60], HELEN [23], XM2VTS [34], and IBUG [38]." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 306, + 677, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 677, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 677, + 545, + 712 + ], + "type": "text", + "content": "- CariFace Dataset. CariFace [4] is created by searching and selecting thousands of various caricatures from different celebrities on the Internet." + } + ] + } + ], + "index": 39 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 52, + 70, + 149, + 78 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 70, + 149, + 78 + ], + "spans": [ + { + "bbox": [ + 52, + 70, + 149, + 78 + ], + "type": "text", + "content": "Cartoon Face Real Human Face" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "2428" + } + ] + } + ], + "index": 41 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 91, + 545, + 187 + ], + "blocks": [ + { + "bbox": [ + 196, + 71, + 395, + 82 + ], + "lines": [ + { + "bbox": [ + 196, + 71, + 395, + 82 + ], + "spans": [ + { + "bbox": [ + 196, + 71, + 395, + 82 + ], + "type": "text", + "content": "Table 1. Data settings for the three learning paradigms." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 91, + 545, + 187 + ], + "lines": [ + { + "bbox": [ + 48, + 91, + 545, + 187 + ], + "spans": [ + { + "bbox": [ + 48, + 91, + 545, + 187 + ], + "type": "table", + "html": "
Learning Paradigm#Training Images and Label Information#Testing Images
300WCariFaceArtiFaceCommon300W ChallengeFullCariFaceArtiFace
DA (300W→CariFace)3,148Labeled3,372Unlabeled——554135689800—
DA (300W→ArtiFace)3,148Labeled——128Unlabeled554135689—32
GZSL (Unseen ArtiFace)3,148Labeled3,372Unlabeled——554135689800160
GZSL (Unseen CariFace)3,148Labeled——128Unlabeled55413568980032
Oracle3,148Labeled3,372Labeled128Labeled55413568980032
", + "image_path": "0216e2426655312b416a8d63a1eb44d9c7615834e823ffb6796152cd4f9fe7c8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 50, + 205, + 287, + 365 + ], + "blocks": [ + { + "bbox": [ + 50, + 205, + 287, + 365 + ], + "lines": [ + { + "bbox": [ + 50, + 205, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 50, + 205, + 287, + 365 + ], + "type": "image", + "image_path": "5fa9e5a7fd20df897a92c2f02d3b3db7c70ae67afbc1f356308b48918f7dda41.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 373, + 287, + 396 + ], + "lines": [ + { + "bbox": [ + 46, + 373, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 287, + 396 + ], + "type": "text", + "content": "Figure 4. Illustrations of typical samples in the 300W, CariFace, and ArtiFace datasets, each of which is annotated with landmarks." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 418, + 287, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 418, + 287, + 454 + ], + "spans": [ + { + "bbox": [ + 47, + 418, + 287, + 454 + ], + "type": "text", + "content": "- ArtiFace Dataset. ArtiFace [55] contains 160 artistic portraits of 16 artists, which covers diverse artwork styles ranging from Renaissance to Comics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 456, + 287, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 287, + 575 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 287, + 575 + ], + "type": "text", + "content": "Each face in the datasets is annotated with 68 landmarks. Typical faces in the datasets and their landmarks are shown in Fig. 4. We can find that the faces in the three datasets have distinguished styles, which correspond to three different domains. In particular, compared to 300W [38], CariFace [4] exhibits abstract and exaggerated patterns, leading to large representation variations. ArtiFace [55] not only has larger variations across different artistic categories but also differs greatly in terms of the aspect of facial scales, orientations, locations, and so on." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 585, + 233, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 585, + 233, + 598 + ], + "spans": [ + { + "bbox": [ + 47, + 585, + 233, + 598 + ], + "type": "text", + "content": "4.2. Learning Paradigms and Baselines" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 628 + ], + "type": "text", + "content": "Given the above datasets, we consider the following three learning paradigms:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 630, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 47, + 630, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 630, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 47, + 630, + 287, + 676 + ], + "type": "text", + "content": "- Domain Adaptation (DA). Given labeled 300W faces and unlabeled stylized faces from CariFace or ArtiFace, we learn a face landmarker based on various domain adaptation methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 677, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 712 + ], + "type": "text", + "content": "- Generalized Zero-shot Learning (GZSL). In the challenging GZSL setting, we learn a face landmarker based on the above DA-based methods and test it in an unseen" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 312, + 208, + 545, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 208, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 312, + 208, + 545, + 232 + ], + "type": "text", + "content": "face domain (e.g., learning the landmarker on labeled 300W and unlabeled CariFace and testing on ArtiFace)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 232, + 545, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 232, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 306, + 232, + 545, + 266 + ], + "type": "text", + "content": "- Oracle. In this setting, the labeled faces of all three datasets are accessible, and we can learn the face landmark by classic supervised learning." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 267, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 267, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 267, + 545, + 422 + ], + "type": "text", + "content": "For a fair comparison, in each learning paradigm, we set the architecture of the face landmark based on the SLPT in [52]. Ideally, we would like to learn landmarks in the DA and GZSL settings, making its performance comparable to the oracle. In the oracle setting, we can learn the face landmark directly via classic supervised learning (SL), i.e., " + }, + { + "bbox": [ + 304, + 267, + 545, + 422 + ], + "type": "inline_equation", + "content": "\\min_{\\theta}\\sum_{(\\mathbf{X},\\mathbf{Y})\\sim \\mathcal{D}}\\| f_{\\theta}(\\mathbf{X}) - \\mathbf{Y}\\| _F^2" + }, + { + "bbox": [ + 304, + 267, + 545, + 422 + ], + "type": "text", + "content": ". In the DA and GZSL settings, besides minimizing the landmark estimation errors, we can apply various image style transfer and domain adaptation methods, e.g., RevGrad [12], CycleGAN [59], BDL [25], AdaptSegNet [45] and FDA [54], to impose domain adaptation regularization during training. These methods work as the baselines of our method." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 423, + 545, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 545, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 545, + 519 + ], + "type": "text", + "content": "Given the landmarkers learned by various methods, we evaluate them with the standard metric, Normalized Mean Error (NME). In Tab. 1, we show the training and testing data settings in the above three learning paradigms. Following existing work [9, 42, 52], we further split the 689 testing faces in 300W into 554 faces in common scenarios and 135 faces in challenging scenarios. The NMEs for the common, challenge, and full scenarios are recorded." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 527, + 495, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 527, + 495, + 540 + ], + "spans": [ + { + "bbox": [ + 305, + 527, + 495, + 540 + ], + "type": "text", + "content": "4.3. Numerical and Visual Comparisons" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": "In Tab. 2, we show the performance of various learning methods in different settings, demonstrating the effectiveness and superiority of our method. In particular, existing DA methods often fail to improve the generalization power of model a lot in face landmarking tasks — their performance in target and unseen domains is inferior to that in the oracle setting, with a significant gap on NME. A potential reason for this phenomenon is that these methods focus on the adaptation of image domain and the landmark-related loss is not dominant in their learning processes. As a result, instead of learning the face landmarker, they make more efforts to optimize the parameters of other modules (e.g., the neural network-based face stylization modules and discriminators) during training." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2429" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 74, + 91, + 518, + 264 + ], + "blocks": [ + { + "bbox": [ + 96, + 70, + 496, + 82 + ], + "lines": [ + { + "bbox": [ + 96, + 70, + 496, + 82 + ], + "spans": [ + { + "bbox": [ + 96, + 70, + 496, + 82 + ], + "type": "text", + "content": "Table 2. Comparisons for various methods on their NMEs. In DA and GZSL settings, the best results are bold." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 74, + 91, + 518, + 264 + ], + "lines": [ + { + "bbox": [ + 74, + 91, + 518, + 264 + ], + "spans": [ + { + "bbox": [ + 74, + 91, + 518, + 264 + ], + "type": "table", + "html": "
Learning ParadigmLearning Method300WCariFaceArtiFaceAverage NME
CommonChallengeFull
DA (300W→CariFace) and GZSL (Unseen ArtiFace)SL+RevGrad [12]2.845.583.3812.195.167.83
SL+CycleGAN [59]2.745.433.2712.114.707.70
SL+BDL [25]3.286.173.8413.635.658.77
SL+ASN [45]2.925.263.3812.214.757.80
SL+FDA [54]2.895.183.3412.664.608.07
Ours2.794.913.207.703.955.46
DA (300W→ArtiFace) and GZSL (Unseen CariFace)SL+RevGrad [12]2.995.813.5512.464.748.26
SL+CycleGAN [59]3.005.653.5212.645.348.36
SL+BDL [25]2.995.323.4413.405.908.73
SL+ASN [45]2.895.813.4616.585.6510.31
SL+FDA [54]3.056.213.6812.335.878.27
Ours2.905.143.3410.933.937.34
OracleSL2.684.863.105.483.314.36
", + "image_path": "436e72c9fcd668ccad6ff1f955849c66922266cd48cf548f2e8bbca5054db73d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 74, + 296, + 518, + 392 + ], + "blocks": [ + { + "bbox": [ + 51, + 274, + 541, + 286 + ], + "lines": [ + { + "bbox": [ + 51, + 274, + 541, + 286 + ], + "spans": [ + { + "bbox": [ + 51, + 274, + 541, + 286 + ], + "type": "text", + "content": "Table 3. The impacts of different model architectures on the NME performance of our method. In each setting, the best results are bold." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 74, + 296, + 518, + 392 + ], + "lines": [ + { + "bbox": [ + 74, + 296, + 518, + 392 + ], + "spans": [ + { + "bbox": [ + 74, + 296, + 518, + 392 + ], + "type": "table", + "html": "
Learning ParadigmLearning Method300WCariFaceArtiFaceAverage NME
CommonChallengeFull
DA (300W→CariFace) and GZSL (Unseen ArtiFace)SBR [9]3.616.364.158.005.096.11
HRNet [47]2.935.293.407.734.335.60
SLPT [52] (Default)2.794.913.207.703.955.46
DA (300W→ArtiFace) and GZSL (Unseen CariFace)SBR [9]3.706.824.3211.355.328.04
HRNet [47]3.045.443.519.854.306.86
SLPT [52] (Default)2.905.143.3410.933.937.34
", + "image_path": "4dfc873a1ce4d4928e5caad88db58af52719b4d664410a90e434def2aca28bfd.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 48, + 408, + 288, + 602 + ], + "blocks": [ + { + "bbox": [ + 48, + 408, + 288, + 602 + ], + "lines": [ + { + "bbox": [ + 48, + 408, + 288, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 408, + 288, + 602 + ], + "type": "image", + "image_path": "0f8cab360045e690a2ef3f18f61a48a52f72e1c04aecf4ba1c142ec126f49549.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 610, + 288, + 654 + ], + "lines": [ + { + "bbox": [ + 46, + 610, + 288, + 654 + ], + "spans": [ + { + "bbox": [ + 46, + 610, + 288, + 654 + ], + "type": "text", + "content": "Figure 5. Visual comparisons for various methods in the two DA settings. We only highlight points on the inner lips in the enlarged region of the mouth in (a), as well as the eyes and the sides of the cheeks, excluding points on the eyebrows in (b)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": "Different from the baselines, the performance of our model in the target domain (e.g., CariFace and ArtiFace) is improved consistently, while the performance in the source" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 306, + 408, + 547, + 602 + ], + "blocks": [ + { + "bbox": [ + 306, + 408, + 547, + 602 + ], + "lines": [ + { + "bbox": [ + 306, + 408, + 547, + 602 + ], + "spans": [ + { + "bbox": [ + 306, + 408, + 547, + 602 + ], + "type": "image", + "image_path": "be662788e90a37c6c76e7c6d4cdb09bcb6bed248a4a8d98d513228704ab22472.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 611, + 547, + 656 + ], + "lines": [ + { + "bbox": [ + 304, + 611, + 547, + 656 + ], + "spans": [ + { + "bbox": [ + 304, + 611, + 547, + 656 + ], + "type": "text", + "content": "Figure 6. Visual comparisons for various methods in the GZSL (Unseen ArtiFace) setting. The zoomed-in area in (a) highlights the noses and facial contours, while that in (b) concentrates on the upper part of faces." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "text", + "content": "domain does not degrade a lot. In both DA and GZSL settings, our method outperforms the baselines in most situations. Especially in the GZSL settings, the landmarks" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2430" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 78, + 151, + 171 + ], + "blocks": [ + { + "bbox": [ + 60, + 78, + 151, + 171 + ], + "lines": [ + { + "bbox": [ + 60, + 78, + 151, + 171 + ], + "spans": [ + { + "bbox": [ + 60, + 78, + 151, + 171 + ], + "type": "image", + "image_path": "68c259f5c1c68b871a15d646c741b149739eef3874cb2d78b720ccb00cd59f25.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 173, + 138, + 182 + ], + "lines": [ + { + "bbox": [ + 69, + 173, + 138, + 182 + ], + "spans": [ + { + "bbox": [ + 69, + 173, + 138, + 182 + ], + "type": "text", + "content": "(a) 300W (Common)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 157, + 78, + 247, + 171 + ], + "blocks": [ + { + "bbox": [ + 157, + 78, + 247, + 171 + ], + "lines": [ + { + "bbox": [ + 157, + 78, + 247, + 171 + ], + "spans": [ + { + "bbox": [ + 157, + 78, + 247, + 171 + ], + "type": "image", + "image_path": "e704056cb9d7ab65b4a82d604e79c6df3f903bb4d7559b802ccfe96407c6b825.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 164, + 173, + 236, + 182 + ], + "lines": [ + { + "bbox": [ + 164, + 173, + 236, + 182 + ], + "spans": [ + { + "bbox": [ + 164, + 173, + 236, + 182 + ], + "type": "text", + "content": "(b) 300W (Challenge)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 253, + 78, + 343, + 171 + ], + "blocks": [ + { + "bbox": [ + 253, + 78, + 343, + 171 + ], + "lines": [ + { + "bbox": [ + 253, + 78, + 343, + 171 + ], + "spans": [ + { + "bbox": [ + 253, + 78, + 343, + 171 + ], + "type": "image", + "image_path": "616dba4ead884d1aedf03fb224f032c7ace69779ce82530ae5bc0863910a4685.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 271, + 173, + 323, + 182 + ], + "lines": [ + { + "bbox": [ + 271, + 173, + 323, + 182 + ], + "spans": [ + { + "bbox": [ + 271, + 173, + 323, + 182 + ], + "type": "text", + "content": "(c) 300W (Full)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 75, + 192, + 515, + 204 + ], + "lines": [ + { + "bbox": [ + 75, + 192, + 515, + 204 + ], + "spans": [ + { + "bbox": [ + 75, + 192, + 515, + 204 + ], + "type": "text", + "content": "Figure 7. Comparisons for different optimization strategies on NME performance in the GZSL (Unseen ArtiFace) setting." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 347, + 78, + 439, + 171 + ], + "blocks": [ + { + "bbox": [ + 347, + 78, + 439, + 171 + ], + "lines": [ + { + "bbox": [ + 347, + 78, + 439, + 171 + ], + "spans": [ + { + "bbox": [ + 347, + 78, + 439, + 171 + ], + "type": "image", + "image_path": "623d17bf8a17fe90e75948eb5ddd406f1058539ba3643be74823c7e87dd174f7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 359, + 173, + 427, + 182 + ], + "lines": [ + { + "bbox": [ + 359, + 173, + 427, + 182 + ], + "spans": [ + { + "bbox": [ + 359, + 173, + 427, + 182 + ], + "type": "text", + "content": "(d) CariFace (Target)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 446, + 78, + 536, + 171 + ], + "blocks": [ + { + "bbox": [ + 446, + 78, + 536, + 171 + ], + "lines": [ + { + "bbox": [ + 446, + 78, + 536, + 171 + ], + "spans": [ + { + "bbox": [ + 446, + 78, + 536, + 171 + ], + "type": "image", + "image_path": "8af5e33f14e5a98d379b62e8b5a9a1aba8c3603882bb906513a3f8dd29f368f3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 454, + 173, + 525, + 182 + ], + "lines": [ + { + "bbox": [ + 454, + 173, + 525, + 182 + ], + "spans": [ + { + "bbox": [ + 454, + 173, + 525, + 182 + ], + "type": "text", + "content": "(e) ArtiFace (Unseen)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 48, + 243, + 291, + 296 + ], + "blocks": [ + { + "bbox": [ + 52, + 223, + 280, + 234 + ], + "lines": [ + { + "bbox": [ + 52, + 223, + 280, + 234 + ], + "spans": [ + { + "bbox": [ + 52, + 223, + 280, + 234 + ], + "type": "text", + "content": "Table 4. Comparison with SOTA landmark detectors on NME." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 243, + 291, + 296 + ], + "lines": [ + { + "bbox": [ + 48, + 243, + 291, + 296 + ], + "spans": [ + { + "bbox": [ + 48, + 243, + 291, + 296 + ], + "type": "table", + "html": "
Learning MethodTrain on 300WOur method
BackboneOPSPIGASTARSLPTSTARSLPT
DA on Cariface10.4611.2310.9711.057.627.70
GZSL on Artiface5.555.195.304.564.863.93
", + "image_path": "c66f473e15a7922191763432053e1a6f6465f54508ed665838b36dd40fa694bc.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 318, + 287, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 318, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 318, + 287, + 389 + ], + "type": "text", + "content": "obtained by our method show encouraging generalization power, which achieves lower NME than the baselines in the unseen domain (i.e., ArtiFace). Overall, the NME of our method in the source domain (i.e., 300W) is comparable to that achieved in the oracle setting. For the target even unseen domains, our method reduces the gap to the oracle." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 390, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 390, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 390, + 287, + 510 + ], + "type": "text", + "content": "Besides the numerical comparisons, we provide some visualization results obtained by different methods in Figs. 5 and 6. For a complete comparison, the results of the SLPT trained only on " + }, + { + "bbox": [ + 46, + 390, + 287, + 510 + ], + "type": "inline_equation", + "content": "300\\mathrm{W}" + }, + { + "bbox": [ + 46, + 390, + 287, + 510 + ], + "type": "text", + "content": " are shown in the figures as well, which corresponds to learning a face landmarker only based on the data in the source domain. We can find that by applying our learning method, the face landmarker obtains better landmarks, which aligns with the ground truth with smaller errors, particularly for the landmarks of face contour, nose, mouth, and eyebrow." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 510, + 288, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 288, + 619 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 288, + 619 + ], + "type": "text", + "content": "To further verify the generalization of our method, we select three more state-of-the-art landmarks, including OpenPose (OP) [7], SPIGA [36] and STAR [57] for numerical comparisons, and incorporate STAR as the backbone model of our method. Tab. 4 demonstrates that our learning method is applicable for various backbone models (e.g., STAR and SLPT), and integrating existing landmarkers with our learning method is an effective and competitive approach to enhance their generalizability." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 627, + 172, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 627, + 172, + 639 + ], + "spans": [ + { + "bbox": [ + 47, + 627, + 172, + 639 + ], + "type": "text", + "content": "4.4. Analytic Experiments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 47, + 646, + 229, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 646, + 229, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 646, + 229, + 658 + ], + "type": "text", + "content": "4.4.1 Joint v.s. Alternating Optimization" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "As aforementioned, learning the face landmarker and the warping field model jointly makes it much easier to fall into undesired local optimum or unstable saddle points. To verify this claim, we apply the joint optimization strategy, i.e.," + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 224, + 547, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 224, + 547, + 357 + ], + "spans": [ + { + "bbox": [ + 304, + 224, + 547, + 357 + ], + "type": "text", + "content": "solving (3) by optimizing " + }, + { + "bbox": [ + 304, + 224, + 547, + 357 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 224, + 547, + 357 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 224, + 547, + 357 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 224, + 547, + 357 + ], + "type": "text", + "content": " jointly in each gradient descent step and compare its performance with ours. In Fig. 7, we show the best NME achieved by this joint optimization strategy and the NME achieved by our alternating optimization method in each iteration. We can find that the joint optimization strategy tends to overfit the source domain (300W), leading to power generalizability in the target and unseen domains. On the contrary, with the increase of iterations, the NMEs of our method on the target and unseen domains decrease consistently and become lower than those of the joint optimization after several iterations." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 372, + 474, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 372, + 474, + 384 + ], + "spans": [ + { + "bbox": [ + 305, + 372, + 474, + 384 + ], + "type": "text", + "content": "4.4.2 Impacts of Model Architectures" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 391, + 545, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 391, + 545, + 631 + ], + "spans": [ + { + "bbox": [ + 304, + 391, + 545, + 631 + ], + "type": "text", + "content": "Besides the optimization strategy, the model architecture also has an impact on the model performance. By default, we implement the face landmarker as the SLPT model. In this experiment, we further explore the performance of other model architectures, including SBR [9] and HRNet [47]. Tab. 3 shows the performance of different model architectures in DA and GZSL settings. We can find that the SLPT-based face landmarker works better than the two competitors in this experiment. A potential reason for this phenomenon is that both SBR and HRNet are heatmap-based face landmarks. Given an input facial image, they output a heatmap indicating the distribution of landmarks rather than a set of deterministic landmark coordinates. We have to first detect the landmarks from the heatmap and then pass them through the warping field model. Accordingly, in the face warper optimization step (i.e., solving (4)), the landmarker and the warping field model cannot be trained in an end-to-end way because the backpropagation of the gradient becomes inapplicable. As a result, we have to update " + }, + { + "bbox": [ + 304, + 391, + 545, + 631 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 391, + 545, + 631 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 391, + 545, + 631 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 391, + 545, + 631 + ], + "type": "text", + "content": " alternatively, leading to suboptimal performance." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 646, + 429, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 646, + 429, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 646, + 429, + 658 + ], + "type": "text", + "content": "4.4.3 Effects on Loss Term" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "In the context of our model's loss function, the image gradient field serves as its input. It has been observed through experiments that utilizing distinct gradient operators leads to the generation of varying image gradient fields, which in" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2431" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 91, + 281, + 137 + ], + "blocks": [ + { + "bbox": [ + 68, + 71, + 266, + 82 + ], + "lines": [ + { + "bbox": [ + 68, + 71, + 266, + 82 + ], + "spans": [ + { + "bbox": [ + 68, + 71, + 266, + 82 + ], + "type": "text", + "content": "Table 5. Quantitative comparison on CariFace dataset." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 91, + 281, + 137 + ], + "lines": [ + { + "bbox": [ + 52, + 91, + 281, + 137 + ], + "spans": [ + { + "bbox": [ + 52, + 91, + 281, + 137 + ], + "type": "table", + "html": "
OperatorGraySpatialLaplacianCannySobel
NME✓8.2827.8708.0057.695
×7.9587.9218.0117.863
", + "image_path": "a9c9c173fdf6deaeb922089993299153750deb9a9fed51bc0848693c5a29d00b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 48, + 170, + 290, + 243 + ], + "blocks": [ + { + "bbox": [ + 73, + 148, + 260, + 159 + ], + "lines": [ + { + "bbox": [ + 73, + 148, + 260, + 159 + ], + "spans": [ + { + "bbox": [ + 73, + 148, + 260, + 159 + ], + "type": "text", + "content": "Table 6. Comparisons for different losses on NME." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 170, + 290, + 243 + ], + "lines": [ + { + "bbox": [ + 48, + 170, + 290, + 243 + ], + "spans": [ + { + "bbox": [ + 48, + 170, + 290, + 243 + ], + "type": "table", + "html": "
Type of LossMSEPerceptualw.o. Grad MSEGrad MSE
300WCommon3.082.892.962.79
Challenge5.435.135.004.91
Full3.543.333.363.20
DA on Caricature8.268.239.487.70
GZSL on ArtiFace4.124.074.063.95
", + "image_path": "2803b262f27e6126b63dbd287c4fce76ec888e312cb29274d9154d6bc7a6ed5b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 266, + 287, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 287, + 314 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 287, + 314 + ], + "type": "text", + "content": "turn influences the model's performance. Consequently, in this section, we assess and compare the training outcomes derived from image gradient fields obtained through the application of diverse gradient operators." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 315, + 287, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 315, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 315, + 287, + 518 + ], + "type": "text", + "content": "Moreover, it is crucial to recognize that the choice of the input image, specifically whether it is a grayscale image or not, can also exert a certain influence on the results. As a result, we have incorporated a comparison of the outcomes when the input is a grayscale image. It is important to emphasize that all experiment settings are maintained consistently. The results are shown in Tab. 5. Utilizing grayscale images yields superior effects compared to not using grayscale images. This could be attributed to the fact that converting an image to grayscale eliminates the trivial effects of color and complex textures. The most optimal outcome is achieved by employing the Sobel operator. This may be due to the Sobel operator's emphasis on edge information within the image, enabling it to capture more geometric details. Additionally, the Sobel operator incorporates a mild smoothing effect during gradient computation, which mitigates the impact of noise." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 518, + 287, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 518, + 287, + 615 + ], + "spans": [ + { + "bbox": [ + 46, + 518, + 287, + 615 + ], + "type": "text", + "content": "Furthermore, we consider " + }, + { + "bbox": [ + 46, + 518, + 287, + 615 + ], + "type": "inline_equation", + "content": "i)" + }, + { + "bbox": [ + 46, + 518, + 287, + 615 + ], + "type": "text", + "content": " replacing the gradient MSE loss in (3) with the pixel MSE or perceptual loss and " + }, + { + "bbox": [ + 46, + 518, + 287, + 615 + ], + "type": "inline_equation", + "content": "ii)" + }, + { + "bbox": [ + 46, + 518, + 287, + 615 + ], + "type": "text", + "content": " removing the gradient MSE loss. Tab. 6 shows the rationality of the gradient MSE for the reason that " + }, + { + "bbox": [ + 46, + 518, + 287, + 615 + ], + "type": "inline_equation", + "content": "i)" + }, + { + "bbox": [ + 46, + 518, + 287, + 615 + ], + "type": "text", + "content": " landmarks are distributed on edges; and " + }, + { + "bbox": [ + 46, + 518, + 287, + 615 + ], + "type": "inline_equation", + "content": "ii)" + }, + { + "bbox": [ + 46, + 518, + 287, + 615 + ], + "type": "text", + "content": " the gradient field filters out unnecessary color information, simplifying the task. These findings form the basis for our decision to adopt this particular method in our experimental process." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 633, + 275, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 633, + 275, + 645 + ], + "spans": [ + { + "bbox": [ + 47, + 633, + 275, + 645 + ], + "type": "text", + "content": "4.4.4 Rationality of Proposed Warping Field Model" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "Besides the model architecture of the face landmarker, the warping field model impacts our face landmarker as well — when the predicted warping field is inaccurate, we cannot obtain reliable pseudo landmarks for stylized faces. Therefore, we investigate different face warping models and" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 313, + 70, + 384, + 140 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 384, + 140 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 384, + 140 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 384, + 140 + ], + "type": "image", + "image_path": "359e98d3068b708fe1b84baf873d783d7b880e9df8b874149cb2d4625443f3b2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 142, + 365, + 151 + ], + "lines": [ + { + "bbox": [ + 331, + 142, + 365, + 151 + ], + "spans": [ + { + "bbox": [ + 331, + 142, + 365, + 151 + ], + "type": "text", + "content": "(a) Source" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 391, + 70, + 462, + 140 + ], + "blocks": [ + { + "bbox": [ + 391, + 70, + 462, + 140 + ], + "lines": [ + { + "bbox": [ + 391, + 70, + 462, + 140 + ], + "spans": [ + { + "bbox": [ + 391, + 70, + 462, + 140 + ], + "type": "image", + "image_path": "15b97daac31685b24a66ea760ece96cbd351e3af902952727e7c166893c970ee.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 142, + 443, + 151 + ], + "lines": [ + { + "bbox": [ + 410, + 142, + 443, + 151 + ], + "spans": [ + { + "bbox": [ + 410, + 142, + 443, + 151 + ], + "type": "text", + "content": "(b) Target" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 469, + 70, + 539, + 140 + ], + "blocks": [ + { + "bbox": [ + 469, + 70, + 539, + 140 + ], + "lines": [ + { + "bbox": [ + 469, + 70, + 539, + 140 + ], + "spans": [ + { + "bbox": [ + 469, + 70, + 539, + 140 + ], + "type": "image", + "image_path": "0a715dfe02da865a1f9186db1f0443eee7771f7f8d602fe6cefb31b33ec90609.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 482, + 142, + 526, + 151 + ], + "lines": [ + { + "bbox": [ + 482, + 142, + 526, + 151 + ], + "spans": [ + { + "bbox": [ + 482, + 142, + 526, + 151 + ], + "type": "text", + "content": "(c) AutoToon" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 313, + 152, + 383, + 221 + ], + "blocks": [ + { + "bbox": [ + 313, + 152, + 383, + 221 + ], + "lines": [ + { + "bbox": [ + 313, + 152, + 383, + 221 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 383, + 221 + ], + "type": "image", + "image_path": "54496ec481a0673426a460b7ff20c23f36bb5240a5712310e4d7b56e83c2b078.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 325, + 223, + 372, + 232 + ], + "lines": [ + { + "bbox": [ + 325, + 223, + 372, + 232 + ], + "spans": [ + { + "bbox": [ + 325, + 223, + 372, + 232 + ], + "type": "text", + "content": "(d) CariGANs" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 391, + 152, + 461, + 221 + ], + "blocks": [ + { + "bbox": [ + 391, + 152, + 461, + 221 + ], + "lines": [ + { + "bbox": [ + 391, + 152, + 461, + 221 + ], + "spans": [ + { + "bbox": [ + 391, + 152, + 461, + 221 + ], + "type": "image", + "image_path": "cf322da2ddab6d762f1564a2e626fad3117d8c694b3b1cb1a5e04d48d543e9bc.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 223, + 443, + 232 + ], + "lines": [ + { + "bbox": [ + 410, + 223, + 443, + 232 + ], + "spans": [ + { + "bbox": [ + 410, + 223, + 443, + 232 + ], + "type": "text", + "content": "(e) RAFT" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 469, + 152, + 539, + 221 + ], + "blocks": [ + { + "bbox": [ + 469, + 152, + 539, + 221 + ], + "lines": [ + { + "bbox": [ + 469, + 152, + 539, + 221 + ], + "spans": [ + { + "bbox": [ + 469, + 152, + 539, + 221 + ], + "type": "image", + "image_path": "494bc94405d2d2241ae6514047ff04b26435dba3e8593177a6c5a0a6ed87df98.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 491, + 223, + 518, + 232 + ], + "lines": [ + { + "bbox": [ + 491, + 223, + 518, + 232 + ], + "spans": [ + { + "bbox": [ + 491, + 223, + 518, + 232 + ], + "type": "text", + "content": "(f) Ours" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 334, + 243, + 515, + 254 + ], + "lines": [ + { + "bbox": [ + 334, + 243, + 515, + 254 + ], + "spans": [ + { + "bbox": [ + 334, + 243, + 515, + 254 + ], + "type": "text", + "content": "Figure 8. Comparisons for different face warpers." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 276, + 545, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 276, + 545, + 444 + ], + "spans": [ + { + "bbox": [ + 304, + 276, + 545, + 444 + ], + "type": "text", + "content": "demonstrate the rationality of the warping field model implemented in our work. In particular, applying different warping field models in our training process, we visualize their warping results in Fig. 8. We can find that although the face landmarker with the polyharmonic interpolation model leads to a very simple face warper, it outperforms many existing neural network-based image warping methods, such as AutoToon [15] and CariGANs [5] for facial manipulation and the optical flow methods like RAFT [44]. These methods either require one-to-one correspondence information between source and target images or assume the deformation between the two images to be slight, making them unsuitable for our problem, especially for the stylized faces with significant nonrigid deformations." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 456, + 379, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 456, + 379, + 468 + ], + "spans": [ + { + "bbox": [ + 306, + 456, + 379, + 468 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 476, + 545, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 476, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 304, + 476, + 545, + 620 + ], + "type": "text", + "content": "In this paper, we propose a simple but effective method for learning a generalizable face landmarker applicable to facial images with different styles. Given labeled real human faces and unlabeled stylized faces, our method learns the face landmarker under the guidance of conditional face warping, demonstrating the usefulness of the warping information. An alternating optimization framework is proposed to learn the face landmarker together with the warping field model. Experiments demonstrate the effectiveness of our method. Especially in the generalized zero-shot learning scenarios, our method achieves encouraging landmarking accuracy in unseen face domains." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 633, + 409, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 633, + 409, + 646 + ], + "spans": [ + { + "bbox": [ + 306, + 633, + 409, + 646 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 653, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 714 + ], + "type": "text", + "content": "This work was supported in part by the National Natural Science Foundation of China (62102031), the foundation of Key Laboratory of Artificial Intelligence, Ministry of Education, P.R. China, and the Young Scholar Program (XSQD-202107001) from Beijing Institute of Technology." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2432" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Olalekan Agbolade, Azree Nazri, Razali Yaakob, Abdul Azim Abd Ghani, and Yoke Kqueen Cheah. Landmark-based homologous multi-point warping approach to 3d facial recognition using multiple datasets. PeerJ Computer Science, 6:e249, 2020. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 212 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 212 + ], + "type": "text", + "content": "[2] Sadegh Aliakbarian, Pashmina Cameron, Federica Bogo, Andrew Fitzgibbon, and Thomas J. Cashman. Flag: Flow-based 3d avatar generation from sparse observations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13253-13262, 2022. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 214, + 287, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 214, + 287, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 214, + 287, + 247 + ], + "type": "text", + "content": "[3] Peter N. Belhumeur, David W. Jacobs, David J. Kriegman, and Neeraj Kumar. Localizing parts of faces using a consensus of exemplars. In CVPR 2011, 2011. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 249, + 287, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 249, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 249, + 287, + 282 + ], + "type": "text", + "content": "[4] Hongrui Cai, Yudong Guo, Zhuang Peng, and Juyong Zhang. Landmark detection and 3d face reconstruction for caricature using a nonlinear parametric model, 2020. 1, 4, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 284, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 284, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 284, + 287, + 316 + ], + "type": "text", + "content": "[5] Kaidi Cao, Jing Liao, and Lu Yuan. Carigans: Unpaired photo-to-caricature translation. arXiv preprint arXiv:1811.00222, 2018. 1, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 318, + 287, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 318, + 287, + 351 + ], + "spans": [ + { + "bbox": [ + 53, + 318, + 287, + 351 + ], + "type": "text", + "content": "[6] Xudong Cao, Yichen Wei, Fang Wen, and Jian Sun. Face alignment by explicit shape regression. International journal of computer vision, 107:177-190, 2014. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 353, + 287, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 353, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 53, + 353, + 287, + 397 + ], + "type": "text", + "content": "[7] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7291-7299, 2017. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 399, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 399, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 53, + 399, + 287, + 453 + ], + "type": "text", + "content": "[8] Forrester Cole, David Belanger, Dilip Krishnan, Aaron Sarna, Inbar Mosseri, and William T Freeman. Synthesizing normalized faces from facial identity features. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3703-3712, 2017. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 454, + 287, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 454, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 53, + 454, + 287, + 520 + ], + "type": "text", + "content": "[9] Xuanyi Dong, Shouu-I Yu, Xinshuo Weng, Shih-En Wei, Yi Yang, and Yaser Sheikh. Supervision-by-registration: An unsupervised approach to improve the precision of facial landmark detectors. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 360–368, 2018. 2, 5, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 522, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 522, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 522, + 287, + 588 + ], + "type": "text", + "content": "[10] Alexey Dosovitskiy, Philipp Fischer, Eddy Ilg, Philip Hausser, Caner Hazirbas, Vladimir Golkov, Patrick Van Der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 2758-2766, 2015. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 590, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 590, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 590, + 287, + 634 + ], + "type": "text", + "content": "[11] Pengfei Dou, Shishir K Shah, and Ioannis A Kakadiaris. End-to-end 3d face reconstruction with deep neural networks. In proceedings of the IEEE conference on computer vision and pattern recognition, pages 5908-5917, 2017. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 635, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 287, + 678 + ], + "type": "text", + "content": "[12] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In International conference on machine learning, pages 1180-1189. PMLR, 2015. 1, 2, 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 680, + 287, + 713 + ], + "type": "text", + "content": "[13] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. A neural algorithm of artistic style. arXiv preprint arXiv:1508.06576, 2015. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "text", + "content": "[14] C. A. Glasbey and K. V. Mardia. A review of image-warping methods. Journal of Applied Statistics, 25(2):155-171, 2002. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 106, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 106, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 106, + 545, + 160 + ], + "type": "text", + "content": "[15] Julia Gong, Yannick Hold-Geoffroy, and Jingwan Lu. Auto-toon: Automatic geometric warping for face cartoon generation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 360–369, 2020. 1, 3, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 162, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 162, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 162, + 545, + 205 + ], + "type": "text", + "content": "[16] Arthur Gretton, Karsten Borgwardt, Malte Rasch, Bernhard Schölkopf, and Alex Smola. A kernel method for the two-sample-problem. Advances in neural information processing systems, 19, 2006. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 206, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 206, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 307, + 206, + 545, + 259 + ], + "type": "text", + "content": "[17] Guoliang Kang, Lu Jiang, Yi Yang, and Alexander G Hauptmann. Contrastive adaptation network for unsupervised domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4893-4902, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 261, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 261, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 261, + 545, + 304 + ], + "type": "text", + "content": "[18] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 305, + 545, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 305, + 545, + 349 + ], + "spans": [ + { + "bbox": [ + 307, + 305, + 545, + 349 + ], + "type": "text", + "content": "[19] Vahid Kazemi and Josephine Sullivan. One millisecond face alignment with an ensemble of regression trees. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1867-1874, 2014. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 350, + 545, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 350, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 307, + 350, + 545, + 371 + ], + "type": "text", + "content": "[20] SunnieS. Kim, NicholasI. Kolkin, Jason Salavon, and Gregory Shakhnarovich. Deformable style transfer, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 372, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 372, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 372, + 545, + 403 + ], + "type": "text", + "content": "[21] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 404, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 404, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 307, + 404, + 545, + 459 + ], + "type": "text", + "content": "[22] Marek Kowalski, Jacek Naruniec, and Tomasz Trzcinski. Deep alignment network: A convolutional neural network for robust face alignment. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2017. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 460, + 545, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 460, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 307, + 460, + 545, + 492 + ], + "type": "text", + "content": "[23] Vuong Le, Jonathan Brandt, Zhe Lin, Lubomir Bourdev, and Thomas S. Huang. Interactive Facial Feature Localization, page 679-692. 2012. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 493, + 545, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 493, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 307, + 493, + 545, + 548 + ], + "type": "text", + "content": "[24] Hui Li, Zidong Guo, Seon-Min Rhee, Seungju Han, and Jae-Joon Han. Towards accurate facial landmark detection via cascaded transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4176-4185, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 548, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 548, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 307, + 548, + 545, + 602 + ], + "type": "text", + "content": "[25] Yunsheng Li, Lu Yuan, and Nuno Vasconcelos. Bidirectional learning for domain adaptation of semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6936-6945, 2019. 1, 5, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 603, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 603, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 307, + 603, + 545, + 658 + ], + "type": "text", + "content": "[26] Feng Liu, Dan Zeng, Qijun Zhao, and Xiaoming Liu. Joint face alignment and 3d face reconstruction. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14, pages 545-560. Springer, 2016. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "text", + "content": "[27] Shaolei Liu, Xiaoyuan Luo, Kexue Fu, Manning Wang, and Zhijian Song. A learnable self-supervised task for unsupervised domain adaptation on point cloud classification and segmentation. Frontiers of Computer Science, 17(6):176708, 2023. 1, 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2433" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[28] Weiyang Liu, Yandong Wen, Zhiding Yu, Ming Li, Bhiksha Raj, and Le Song. Spherface: Deep hypersphere embedding for face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 212-220, 2017. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "text", + "content": "[29] Xiao-Chang Liu, Yong-Liang Yang, and Peter Hall. Learning to warp for style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3702-3711, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 287, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 218 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 218 + ], + "type": "text", + "content": "[30] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In International conference on machine learning, pages 97-105. PMLR, 2015. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 220, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 262 + ], + "type": "text", + "content": "[31] Mingsheng Long, Han Zhu, Jianmin Wang, and Michael I Jordan. Unsupervised domain adaptation with residual transfer networks. Advances in neural information processing systems, 29, 2016. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 319 + ], + "type": "text", + "content": "[32] Jiangjing Lv, Xiaohu Shao, Junliang Xing, Cheng Cheng, and Xi Zhou. A deep regression architecture with two-stage re-initialization for high performance facial landmark detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3317-3326, 2017. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 320, + 287, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 287, + 365 + ], + "type": "text", + "content": "[33] Iacopo Masi, Stephen Rawls, Gérard Medioni, and Prem Natarajan. Pose-aware face recognition in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4838-4846, 2016. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 365, + 287, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 287, + 420 + ], + "type": "text", + "content": "[34] Kieron Messer, Jiri Matas, Josef Kittler, Juergen Luettin, Gilbert Maitre, et al. Xm2vtsdb: The extended m2vts database. In Second international conference on audio and video-based biometric person authentication, pages 965-966. CiteSeer, 1999. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 421, + 287, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 421, + 287, + 454 + ], + "spans": [ + { + "bbox": [ + 48, + 421, + 287, + 454 + ], + "type": "text", + "content": "[35] Zhongyi Pei, Zhangjie Cao, Mingsheng Long, and Jianmin Wang. Multi-adversarial domain adaptation. In Thirty-second AAAI conference on artificial intelligence, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 456, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 498 + ], + "type": "text", + "content": "[36] Andrés Prados-Torreblanca, José M Buenaposada, and Luis Baumela. Shape preserving facial landmarks with graph attention networks. arXiv preprint arXiv:2210.07233, 2022. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 500, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 287, + 544 + ], + "type": "text", + "content": "[37] Joseph Roth, Yiying Tong, and Xiaoming Liu. Unconstrained 3d face reconstruction. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2606-2615, 2015. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 545, + 287, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 600 + ], + "type": "text", + "content": "[38] Christos Sagonas, Georgios Tzimiropoulos, Stefanos Zafeiriou, and Maja Pantic. 300 faces in-the-wild challenge: The first facial landmark localization challenge. In 2013 IEEE International Conference on Computer Vision Workshops, 2014. 1, 4, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "type": "text", + "content": "[39] Ahmed Selim, Mohamed Elgharib, and Linda Doyle. Painting style transfer for head portraits using convolutional neural networks. ACM Transactions on Graphics (ToG), 35(4): 1-18, 2016. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "type": "text", + "content": "[40] Yichun Shi, Debayan Deb, and Anil K Jain. Warpgan: Automatic caricature generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10762-10771, 2019. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "text", + "content": "[41] Baochen Sun, Jiashi Feng, and Kate Saenko. Correlation alignment for unsupervised domain adaptation. Domain" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "adaptation in computer vision applications, pages 153-171, 2017. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "text", + "content": "[42] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 141, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 195 + ], + "type": "text", + "content": "[43] Yaniv Taigman, Ming Yang, Marc'Aurelio Ranzato, and Lior Wolf. Deepface: Closing the gap to human-level performance in face verification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1701-1708, 2014. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 198, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 198, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 198, + 545, + 251 + ], + "type": "text", + "content": "[44] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 3, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 254, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 254, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 254, + 545, + 319 + ], + "type": "text", + "content": "[45] Yi-Hsuan Tsai, Wei-Chih Hung, Samuel Schulter, Kiyuh Yuk Sohn, Ming-Hsuan Yang, and Manmohan Chandraker. Learning to adapt structured output space for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7472-7481, 2018. 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 320, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 320, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 308, + 320, + 545, + 365 + ], + "type": "text", + "content": "[46] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 365, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 545, + 430 + ], + "type": "text", + "content": "[47] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 43(10):3349-3364, 2020. 2, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 433, + 545, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 433, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 308, + 433, + 545, + 477 + ], + "type": "text", + "content": "[48] Xinyao Wang, Liefeng Bo, and Li Fuxin. Adaptive wing loss for robust face alignment via heatmap regression. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 478, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 478, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 478, + 545, + 521 + ], + "type": "text", + "content": "[49] Kunhong Wu, Fan Jia, and Yahong Han. Domain-specific feature elimination: multi-source domain adaptation for image classification. Frontiers of Computer Science, 17(4): 174705, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 522, + 545, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 522, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 308, + 522, + 545, + 556 + ], + "type": "text", + "content": "[50] Ruizheng Wu, Xiaodong Gu, Xin Tao, Xiaoyong Shen, Yu-Wing Tai, et al. Landmark assisted cyclegan for cartoon face generation. arXiv preprint arXiv:1907.01424, 2019. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 557, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 557, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 308, + 557, + 545, + 611 + ], + "type": "text", + "content": "[51] Wayne Wu, Chen Qian, Shuo Yang, Quan Wang, Yici Cai, and Qiang Zhou. Look at boundary: A boundary-aware face alignment algorithm. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2129-2138, 2018. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 613, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 678 + ], + "type": "text", + "content": "[52] Jiahao Xia, Weiwei Qu, Wenjian Huang, Jianguo Zhang, Xi Wang, and Min Xu. Sparse local patch transformer for robust face alignment and landmarks inherent relation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4052-4061, 2022. 1, 3, 5, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 680, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 714 + ], + "type": "text", + "content": "[53] Jiaolong Yang, Peiran Ren, Dongqing Zhang, Dong Chen, Fang Wen, Hongdong Li, and Gang Hua. Neural aggregation network for video face recognition. In Proceedings of" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2434" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 288, + 497 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 67, + 73, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 286, + 95 + ], + "type": "text", + "content": "the IEEE conference on computer vision and pattern recognition, pages 4362-4371, 2017. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 288, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 288, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 288, + 140 + ], + "type": "text", + "content": "[54] Yanchao Yang and Stefano Soatto. Fda: Fourier domain adaptation for semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4085-4095, 2020. 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 286, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 286, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 286, + 183 + ], + "type": "text", + "content": "[55] Jordan Yaniv, Yael Newman, and Ariel Shamir. The face of art: landmark detection and geometric style in portraits. ACM Transactions on graphics (TOG), 38(4):1-15, 2019. 1, 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 286, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 286, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 286, + 239 + ], + "type": "text", + "content": "[56] Jing Zhang, Zewei Ding, Wanqing Li, and Philip Ogunbona. Importance weighted adversarial nets for partial domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8156-8164, 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 241, + 286, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 286, + 296 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 286, + 296 + ], + "type": "text", + "content": "[57] Zhenglin Zhou, Huaxia Li, Hong Liu, Nanyang Wang, Gang Yu, and Rongrong Ji. Star loss: Reducing semantic ambiguity in facial landmark detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15475-15484, 2023. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 297, + 286, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 297, + 286, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 286, + 352 + ], + "type": "text", + "content": "[58] Zhou, Erjin and Fan, Haoqiang and Cao, Zhimin and Jiang, Yuning and Yin, Qi. Extensive Facial Landmark Localization with Coarse-to-Fine Convolutional Network Cascade. In Proceedings of the IEEE International Conference on Computer Vision (ICCV) Workshops, 2013. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 354, + 286, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 286, + 407 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 286, + 407 + ], + "type": "text", + "content": "[59] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 1, 2, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 409, + 286, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 409, + 286, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 409, + 286, + 453 + ], + "type": "text", + "content": "[60] Xiangxin Zhu and Deva Ramanan. Face detection, pose estimation, and landmark localization in the wild. In 2012 IEEE conference on computer vision and pattern recognition, pages 2879-2886. IEEE, 2012. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 454, + 286, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 454, + 286, + 497 + ], + "spans": [ + { + "bbox": [ + 48, + 454, + 286, + 497 + ], + "type": "text", + "content": "[61] Yi Zhu, Xindong Wu, Jipeng Qiang, Yunhao Yuan, and Yun Li. Representation learning via an integrated autoencoder for unsupervised domain adaptation. Frontiers of Computer Science, 17(5):175334, 2023. 1" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2435" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Generalizable Novel-View Synthesis using a Stereo Camera/516ec60a-018b-4106-92f4-813c127e6028_content_list.json b/2024/Generalizable Novel-View Synthesis using a Stereo Camera/516ec60a-018b-4106-92f4-813c127e6028_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3ffb2a24832239527ba843236020ac215d467092 --- /dev/null +++ b/2024/Generalizable Novel-View Synthesis using a Stereo Camera/516ec60a-018b-4106-92f4-813c127e6028_content_list.json @@ -0,0 +1,1577 @@ +[ + { + "type": "text", + "text": "Generalizable Novel-View Synthesis using a Stereo Camera", + "text_level": 1, + "bbox": [ + 186, + 130, + 784, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haechan Lee $^{1,*}$", + "bbox": [ + 138, + 179, + 264, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wonjoon Jin $^{2,*}$", + "bbox": [ + 308, + 180, + 431, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Seung-Hwan Baek $^{1,2}$", + "bbox": [ + 475, + 180, + 645, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sunghyun Cho $^{1,2}$", + "bbox": [ + 687, + 180, + 828, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "POSTECH $^{1}$ GSAI & $^{2}$ CSE", + "text_level": 1, + "bbox": [ + 374, + 205, + 594, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{gocks8, jinwj1996, shwbaek, s.cho}@postech.ac.kr", + "bbox": [ + 266, + 226, + 699, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 276, + 313, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose the first generalizable view synthesis approach that specifically targets multi-view stereocamera images. Since recent stereo matching has demonstrated accurate geometry prediction, we introduce stereo matching into novel-view synthesis for high-quality geometry reconstruction. To this end, this paper proposes a novel framework, dubbed StereoNeRF, which integrates stereo matching into a NeRF-based generalizable view synthesis approach. StereoNeRF is equipped with three key components to effectively exploit stereo matching in novel-view synthesis: a stereo feature extractor, a depth-guided plane-sweeping, and a stereo depth loss. Moreover, we propose the StereoNVS dataset, the first multi-view dataset of stereocamera images, encompassing a wide variety of both real and synthetic scenes. Our experimental results demonstrate that StereoNeRF surpasses previous approaches in generalizable view synthesis.", + "bbox": [ + 75, + 308, + 473, + 565 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 579, + 209, + 594 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Novel-view synthesis is a long-standing ill-posed problem in computer vision and graphics, which is inherently challenging due to the necessity of predicting both the geometry and texture from images of a target scene. Recently, Neural Radiance Fields (NeRF) [24] have achieved photorealistic results by jointly optimizing geometry and radiance fields with a coordinate-based network. However, the need for per-scene optimization, which has to learn representations for each target scene individually, restricts its applicability, as it requires additional training time for such optimization.", + "bbox": [ + 75, + 604, + 468, + 755 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent approaches [2, 4, 16, 21, 30, 32, 33, 38] have addressed this issue of synthesizing novel-view images on-the-fly for unseen scenes without per-scene optimization. Early studies [4, 33, 38] utilize an image encoder to train a generic view interpolation function, enabling the estimation of NeRF parameters from unseen input images in a feed-forward manner. However, this single feed-forward manner for estimating geometry and color exacerbates the ill", + "bbox": [ + 75, + 756, + 470, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/43f30b7c5765bfbaecae03499b724cdaea92243b9d92122ef04f9e8d9414441e.jpg", + "image_caption": [ + "(a) Baseline", + "Figure 1. Novel-view synthesis results of a baseline method [16] and ours. The baseline shows degraded performances, even trained using stereo-camera images (b). In contrast, fully exploiting stereo-camera images, our method shows superior results (c)." + ], + "image_footnote": [], + "bbox": [ + 501, + 276, + 625, + 398 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c37eff9227541f2a927ca8ed80391be85c840f40bb690cd779b4ccaa448dfc4f.jpg", + "image_caption": [ + "(b) Baseline", + "(using stereo images)" + ], + "image_footnote": [], + "bbox": [ + 627, + 276, + 751, + 398 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b505cd829b7e1639400ad26b07f9406d86fdcb5b9df89a2fa92d553cc92b96e4.jpg", + "image_caption": [ + "(c) Ours", + "(using stereo images)" + ], + "image_footnote": [], + "bbox": [ + 751, + 276, + 893, + 398 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "posedness, resulting in low-quality geometries and rendering results. For better geometry reasoning, MVSNeRF [2] and GeoNeRF [16] leverage the multi-view stereo (MVS) approach to handle occlusions in a 3D scene. Nonetheless, they struggle with inaccurate geometry prediction and limited synthesis accuracy, as shown in the synthesis result of GeoNeRF (Fig. 1 (a)).", + "bbox": [ + 496, + 489, + 893, + 595 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To tackle this challenge, we propose the first generalizable NeRF approach that leverages stereo-camera images, which are easily accessible thanks to the ubiquity of stereo cameras in most mobile devices. Recent advance in learning-based stereo estimation has demonstrated accurate geometry prediction, often even outperforming learning-based MVS methods as shown in Fig. 2. The superior performance of stereo estimation can be attributed to several key factors. First, unlike MVS that assumes arbitrary number of inputs with arbitrary camera positions and orientations, stereo matching assumes two stereo inputs with a fixed baseline. This constraint allows a more optimal network architecture that can effectively find dense correspondences between input images, and significantly eases the training difficulty. Moreover, larger scale of stereomatching datasets [12, 23, 27] compared to MVS datasets has facilitated remarkable generalization capabilities in stereo estimation network. Therefore, we aim to harness this accurate geometric information from stereo images to alleviate the ill-posedness of generalizable view synthesis.", + "bbox": [ + 496, + 598, + 895, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 101, + 887, + 212, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4939", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a5d0501def5d90ee2d3f6acd8e1baec8277930a78ebfea87967b841bc45298d6.jpg", + "image_caption": [ + "Figure 2. Usefulness of exploiting binocular stereo. Comparison on depth estimation between a learning-based MVS method [26] and a learning-based binocular stereo method [35]." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 467, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, since previous methods do not explicitly consider stereoscopic inputs, they cannot leverage the aforementioned benefits, resulting in degenerated performances shown in Fig. 1 (b) despite using stereo-camera images.", + "bbox": [ + 76, + 205, + 467, + 265 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This paper proposes StereoNeRF, a novel generalizable view synthesis framework leveraging stereo images. StereoNeRF integrates stereo matching into NeRF-based generalizable view synthesis approach, where the stereo matching provides vital geometric information. To this end, we first introduce a stereo feature extractor, which extracts geometry-aware features by correlating horizontal epipolar lines within stereo images. In addition, the stereo feature extractor takes advantage of stereo-correlated features from an off-the-shelf stereo estimation network, which can transfer rich geometric knowledge to our model. Furthermore, with the reliable depth estimated from the stereo estimation network, we aggregate multi-view features through a depth-guided plane-sweeping technique. This technique ensures correspondence matching around the geometry in cost volume construction. We also present a stereo depth loss utilizing the estimated stereo depth. These additional geometric cues from the stereo matching effectively mitigate the ill-posedness in generalizable view synthesis. Notably, our framework leveraging stereo images surpasses the previous approaches [16, 21], which rely on extra depth information.", + "bbox": [ + 76, + 266, + 467, + 583 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, we propose the StereoNVS dataset, which is the first dataset for training and evaluation of novel-view synthesis using stereo-camera images. Our StereoNVS dataset provides real-world and synthetic stereo images. Our extensive evaluation on the StereoNVS dataset shows that stereo-pair inputs can effectively enhance the quality of novel-view synthesis, and shows that StereoNeRF outperforms previous generalizable novel-view synthesis approaches in terms of image and shape qualities.", + "bbox": [ + 75, + 584, + 467, + 718 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are as follows:", + "bbox": [ + 96, + 719, + 315, + 733 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a generalized NeRF approach that leverages stereo-camera images for the first time to alleviate the ill-posedness in generalizable novel-view synthesis.", + "- We propose a novel framework, StereoNeRF, which exploits the benefits of stereo images by integrating stereo matching into generalizable view synthesis. To this end, we present a stereo feature extractor, a depth-guided plane-sweeping, and a stereo depth loss in our framework.", + "- We also present the StereoNVS dataset, the first dataset for training and evaluation of novel-view synthesis obtained by stereo cameras." + ], + "bbox": [ + 76, + 734, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 89, + 640, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Novel-View Synthesis", + "text_level": 1, + "bbox": [ + 500, + 114, + 699, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Novel-view synthesis aims to synthesize target-view images from reference-view images. Early approaches synthesize novel-view images by blending pixels from multiple input images [6, 13, 20]. Recent work adopting neural volume representations has shown remarkable novel-view synthesis results. Zhou et al. [43] propose multi-plane images (MPI) representation estimated from input images, but their methods produce valid novel-view images only for narrow ranges of camera poses. Mildenhall et al. [24] propose Neural Radiance Fields (NeRF) that can synthesize photo-realistic target-view images via neural implicit representations and volume rendering. Albeit its photo-realism, computation-heavy per-scene optimization is needed. Recent variants such as [9, 17, 25, 37] have remarkably reduced the optimization time, but they still require large memory and several minutes for training.", + "bbox": [ + 496, + 137, + 890, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generalizable View Synthesis. Recently, many novel-view synthesis approaches without per-scene optimization have been proposed. Several studies directly predict pixel colors by aggregating image features from aligned pixels, without 3D representations. Suhail et al. [30] and Varma et al. [32] adopt a transformer-based network [7] to compute features along the epipolar lines, which needs a large number of training images for high-quality view synthesis. Du et al. [8] propose a framework to synthesize target views from two images with small overlapped regions.", + "bbox": [ + 496, + 386, + 890, + 537 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another research direction predicts volumetric representation [24] from aggregated features from reference views and synthesizes images via volumetric rendering. PixelNeRF, SRF, and GRF [4, 31, 38] predict radiance fields from pixel-aligned features using an MLP. Among these methods, SRF [4] uses two sampled images from an image collection captured by a monocular camera as a stereo pair, but shows limited synthesis quality. IBRNet [33] proposes to learn generic view interpolation functions, but suffers from artifacts for challenging scenes with complex geometries. MVSNeRF, GeoNeRF and NeuRay [2, 16, 21] utilize the MVS approach using cost volume for better occlusion handling in generalizable view synthesis.", + "bbox": [ + 496, + 537, + 890, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, all the aforementioned methods often fail to capture accurate geometry, particularly in textureless regions, leading to limited synthesis results. Unlike these methods, our framework exploits stereo-camera images with a fixed baseline to effectively capture the geometry of complex scenes as well as textureless regions for high-quality novel-view synthesis.", + "bbox": [ + 496, + 733, + 890, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Stereo Matching and Multi-View Stereo", + "text_level": 1, + "bbox": [ + 500, + 847, + 841, + 863 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Geometry estimation is a long-standing problem in computer vision that has a variety of applications. Among them,", + "bbox": [ + 496, + 869, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4940", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "stereo matching is a task that takes rectified stereo images and computes stereo correspondence to estimate disparities [15, 27]. Recently, a huge number of data-driven approaches have been introduced with the emergence of a vast amount of stereo-matching datasets [12, 23, 27] and made significant progress [1, 23, 35, 39]. MVS approaches that use more than two views have been extensively studied as well [11, 19, 29]. Recently, learning-based MVS approaches have been proposed, e.g., Cheng et al. [3], Gu et al. [14], and Yang et al. [36] present efficient frameworks that cascade cost volumes in a coarse-to-fine manner to enable high-resolution depth estimation.", + "bbox": [ + 75, + 90, + 472, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our work, we have the best of both worlds in our generalizable view synthesis framework. We adopt the MVS approach to aggregate multi-view information for occlusion-aware geometry estimation, as done in GeoNerveRF [16]. Furthermore, our framework also assumes structured stereo images as inputs, and integrates the two-view stereo matching into our framework for robust and accurate geometry estimation in generalizable novel-view synthesis.", + "bbox": [ + 75, + 273, + 472, + 396 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. StereoNeRF", + "text_level": 1, + "bbox": [ + 76, + 412, + 205, + 429 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For novel-view synthesis, our framework takes a set of rectified stereo-camera images and estimates neural radiance fields [24] from which novel-view images are rendered. StereoNeRF integrates a well-designed stereo-matching algorithm into the existing generalizable view synthesis approach [16]. In this section, we first provide an overall pipeline that utilizes an pre-trained stereo estimation network [35]. Then, we explain each step of our framework and the training process in detail, highlighting how to integrate stereo matching into our framework.", + "bbox": [ + 75, + 441, + 472, + 592 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overall pipeline. Fig. 3 shows an overview of our pipeline. Our framework utilizes $N$ pairs of stereo images of a target scene $\\{(I_L^n,I_R^n)\\}_{n = 1}^N$ for novel-view synthesis. In the first step, the pre-trained stereo estimation network takes the $n$ -th image pair $(I_L^n,I_R^n)$ , and outputs stereo depths $(d_{s,L}^{n},d_{s,R}^{n})$ and stereo-correlated features $(t_L^n,t_R^n)$ . In the second step, a stereo feature extractor takes the stereo image pair and the stereo-correlated features, and outputs stereo image features $(f_L^n,f_R^n)$ , explained in Sec. 3.1. The third step aggregates stereo image features from all viewpoints $\\{(f_L^n,f_R^n)\\}_{n = 1}^N$ to build 3D feature volumes $\\{(\\phi_L^n,\\phi_R^n)\\}_{n = 1}^N$ using an MVS network (Sec. 3.2). For constructing feature volumes that faithfully reflect the 3D geometry of a target scene, the third step adopts a depth-guided plane-sweeping which utilizes the estimated stereo depths $(d_{s,L}^{n},d_{s,R}^{n})$ . Finally, a neural renderer predicts radiance fields from the stereo image features $\\{(f_L^n,f_R^n)\\}_{n = 1}^N$ and the feature volumes $\\{(\\phi_L^n,\\phi_R^n)\\}_{n = 1}^N$ from all viewpoints, and novel-view images are synthesized from the radiance fields (Sec. 3.3).", + "bbox": [ + 75, + 613, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a02480fe535e1d999f71da61d0d90036ebf182ec44e92387dbfb611677a36928.jpg", + "image_caption": [ + "Figure 3. Overview of StereoNeRF. StereoNeRF consists of a shared feature extractor and a neural renderer. We design the feature extractor with a stereo estimation network, a stereo feature extractor, and a MVS network. StereoNeRF takes $N$ pairs of stereo images and a camera pose as inputs, and synthesizes a novel-view image of the camera pose." + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 893, + 244 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Feature Extraction from Stereo Image Pairs", + "text_level": 1, + "bbox": [ + 498, + 348, + 872, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unlike previous approaches, which extract features of input images respectively, our stereo feature extractor takes rectified stereo-camera images and computes image feature maps by exploiting the epipolar geometry between them. To this end, our stereo feature extractor consists of three parts: CNN encoders, stereo attention modules (SAM) [5], and CNN decoders (Fig. 4). Moreover, we propose integrating the stereo-correlated features $(t_L^n,t_R^n)$ from the pretrained stereo estimation network into the SAM (green line in Fig. 4). This integration inherits geometric cues from the stereo estimation network. In the following, we provide the detailed description of the stereo feature extractor.", + "bbox": [ + 496, + 371, + 890, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "First, the weight-shared CNN encoders project each of $I_L^n$ and $I_R^n$ into the feature space. Then, we adopt the stereo attention module (SAM) [5] to fuse two feature maps on the horizontal epipolar lines. To this end, SAM estimates stereo correspondences between two feature maps and explicitly adds stereo-correspondent features, as shown in Fig. 5. In addition, we extend the SAM by aggregating the stereocorrelated features $(t_L^n, t_R^n)$ in this feature fusion process. Note that the stereo-correlated features are computed from the pre-trained stereo estimation network, which provides vital geometric cues. These stereo-correlated features enhance stereo correspondences between two feature maps, leading to higher-quality fused feature maps.", + "bbox": [ + 496, + 553, + 892, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specific explanation of the aforementioned feature fusion of the SAM is as follows (Fig. 5). For the left feature, a warping matrix $\\mathcal{W}_{R\\to L}$ is built by multiplying a query matrix and a key matrix followed by a softmax layer. The query matrix is computed from the left feature and the stereocorrelated features, and the key matrix is computed from the right feature and the stereo-correlated features. Then, a value matrix, computed from the right feature and the stereo-correlated feature, is warped along the epipolar lines by multiplying the warping matrix $\\mathcal{W}_{R\\to L}$ , then added to", + "bbox": [ + 496, + 750, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4941", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/85b68ed08f4f2f565d13c77166d36d7ac23cb54574c5916cbfd37d6202094e17.jpg", + "image_caption": [ + "Figure 4. Stereo feature extractor of StereoNeRF, where an pretrained stereo estimation network is incorporated." + ], + "image_footnote": [], + "bbox": [ + 76, + 87, + 472, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the left feature. The same feature fusion process is also performed for the right feature.", + "bbox": [ + 75, + 506, + 468, + 536 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For effective fusion of information from both stereo images and stereo-correlated features, the stereo feature extractor has three sequentially stacked SAM. Finally, the weight-shared CNN decoders take each of the fused features from the SAM, and generate stereo image features $(f_{L}^{n}, f_{R}^{n})$ , which will be used for building feature volumes.", + "bbox": [ + 75, + 536, + 468, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Depth-guided Feature Volume Construction", + "text_level": 1, + "bbox": [ + 76, + 635, + 449, + 651 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Once stereo image features are obtained, we aggregate these features from all viewpoints to create cost volumes. To this end, we adopt a plane-sweeping-based approach, which computes multi-view correspondences among these features. The plane-sweeping-based approach first defines a depth range within pre-computed near and far depths obtained from COLMAP [28]. Then, it hypothesizes depth planes in the entire depth range. On these depth planes, multi-view stereo image features are aggregated via plane-sweeping to build cost volumes. Then, these cost volumes are processed to yield feature volumes, which will be used later in predicting neural radiance fields.", + "bbox": [ + 75, + 659, + 468, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, this plane-sweeping-based approach often struggles with accurate geometry estimation. Since this approach searches the entire depth range of a scene, correspondence matching is inaccurate and prone to error, espe", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/bd3f7efda17ad332653d31cc2a8a05a2e9740ed4281c10f946bbc2e77355d969.jpg", + "image_caption": [ + "Figure 5. Stereo attention module used in the stereo feature extractor. We exploit the rich features from the pre-trained stereo estimation module." + ], + "image_footnote": [], + "bbox": [ + 504, + 90, + 893, + 318 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6b75aa4878ea606421023cecf99e7405733a829548f84266859484cf17cc1e76.jpg", + "image_caption": [ + "Figure 6. Feature volume construction using DGPS." + ], + "image_footnote": [], + "bbox": [ + 509, + 375, + 879, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "cially in textureless regions. To tackle this issue, we introduce a depth-guided plane-sweeping (DGPS), which constructs cost volumes around the stereo depth predicted from the stereo estimation network. Unlike the previous approach using the entire depth range, DGPS introduces a dynamic search range that varies based on the estimated stereo depth. This technique significantly constrains the search space and reduces outliers in correspondence matching.", + "bbox": [ + 496, + 561, + 890, + 684 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fig. 6 depicts the feature volume construction using DGPS. Instead of searching the entire depth range, we hypothesize depth planes around the stereo depths $(d_{s,L}^n,d_{s,R}^n)$ . Then, through DGPS, we aggregate the stereo image features from all viewpoints $\\{(f_L^n,f_R^n)\\}_{n = 1}^N$ on these depth planes to build cost volumes. These cost volumes are further processed via the MVS network to obtain feature volumes $(\\phi_L^n,\\phi_R^n)$ and depth maps $(d_{s,L}^n,d_{s,R}^n)$ . We repeat this process to obtain feature volumes and depth maps for every viewpoint. The resulting multi-view feature volumes and depth maps are denoted as $\\{(\\phi_L^L,\\phi_R^L)\\}_{n = 1}^N$ and $\\{(d_{m,L}^{n},d_{m,R}^{n})\\}_{n = 1}^{N}$ , respectively. Refer to Sec. B.3 in the supplementary document for more details about our feature volume construction.", + "bbox": [ + 496, + 686, + 892, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4942", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Rendering Novel Views", + "text_level": 1, + "bbox": [ + 76, + 90, + 292, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Once feature volumes are constructed, novel-view images can be rendered via neural rendering. For rendering novel-view images, our framework adopts the rendering procedure of GeoNeRF [16]. In the following, we briefly describe the rendering procedure. Given a target viewpoint, we cast a ray for each pixel and sample points along the ray in the 3D space. Then, for each sampled point, we sample features from all the feature volumes $\\{(\\phi_L^n,\\phi_R^n)\\}_{n = 1}^N$ . We also sample features from all the image features $\\{(f_L^n,f_R^n)\\}_{n = 1}^N$ by projecting the sampled point onto input images. The sampled volume and image features are aggregated by a neural renderer network, and the color and density are estimated. Finally, by integrating the estimated color values and depth values of sampled points with their densities along each ray, we obtain the color $\\hat{c}$ and the depth $d_{r}$ at each pixel of a novel view, respectively.", + "bbox": [ + 75, + 113, + 472, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Training StereoNeRF", + "text_level": 1, + "bbox": [ + 76, + 364, + 277, + 380 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We train StereoNeRF with ground-truth images and pseudoground-truth depths for high-quality novel-view synthesis. The pseudo-ground-truth depths $d_{gt}$ are obtained from the pre-trained stereo estimation network [35]. Our training loss $\\mathcal{L}$ is defined as follows:", + "bbox": [ + 75, + 387, + 468, + 463 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {c} + \\mathcal {L} _ {d}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 222, + 474, + 468, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_c$ is a color loss and $\\mathcal{L}_d$ is a depth loss. The color loss is defined as a mean-squared-error (MSE) between the rendered colors $\\hat{c}$ and ground-truth colors [24].", + "bbox": [ + 75, + 501, + 468, + 546 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To alleviate the shape-radiance ambiguity [40] in scene reconstruction using neural radiance fields, we introduce the depth loss $\\mathcal{L}_d$ , which is defined as:", + "bbox": [ + 75, + 547, + 468, + 592 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d} = \\lambda_ {d} ^ {s e l f} \\mathcal {L} _ {d} ^ {s e l f} + \\lambda_ {d} ^ {s t e r e o} \\mathcal {L} _ {d} ^ {s t e r e o}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 601, + 468, + 622 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_d^{self}$ is a self-supervised depth loss from GeoN-NeRF [16] and $\\mathcal{L}_d^{stereo}$ is our proposed stereo depth loss. $\\mathcal{L}_d^{self}$ penalizes the depths estimated from both the stereo estimation network $(d_s)$ and the MVS network $(d_m)$ by comparing them to the depth rendered from NeRF $(d_r)$ .", + "bbox": [ + 75, + 632, + 468, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our framework estimates depths from the stereo estimation network $(d_{s})$ , the MVS network $(d_{m})$ , and the neural renderer network $(d_{r})$ . The stereo depth loss $\\mathcal{L}_d^{stereo}$ guides these networks to predict more accurate depths using the pseudo-ground-truth depths $d_{gt}$ . This loss introduces a further quality improvement, especially in geometric details as will be shown in Sec. 5.3. $\\mathcal{L}_d^{stereo}$ is defined as:", + "bbox": [ + 75, + 710, + 468, + 818 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d} ^ {\\text {s t e r e o}} = \\lambda_ {1} \\mathcal {L} _ {d} ^ {s} + \\lambda_ {2} \\mathcal {L} _ {d} ^ {m} + \\lambda_ {3} \\mathcal {L} _ {d} ^ {r}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 827, + 468, + 844 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_d^s$ , $\\mathcal{L}_d^m$ , and $\\mathcal{L}_d^r$ penalizes depths $d_s$ , $d_m$ , and $d_r$ , respectively, by comparing them with $d_{gt}$ . Note that $d_{gt}$ is different from $d_s$ . We obtain $d_{gt}$ from the pre-trained stereo", + "bbox": [ + 75, + 854, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "estimation network with frozen parameters. On the other hand, we obtain $d_{s}$ from the stereo estimation network in our framework, which has trainable parameters. Refer to Sec. B.4 in the supplementary document for more details about our stereo depth loss.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Due to the different characteristics of datasets such as baseline length, $d_{s}$ may have estimation error, which leads to an error in the depth-guided plane-sweeping. To tackle this, we partially train the matching and propagation network in the stereo estimation network (Fig. 4) using $\\mathcal{L}_d^{self}$ . $\\mathcal{L}_d^{self}$ provides multi-view supervision to the stereo estimation network by leveraging $d_{r}$ estimated from multi-view images. However, since $d_{r}$ may also have estimation error, we further regularize the stereo estimation network using $\\mathcal{L}_d^s$ . This training scheme for the stereo estimation network ensures consistent depth estimation, while preserving the stereo matching capability. Sec. D.1 in the supplementary document further discusses the training scheme of the stereo estimation network.", + "bbox": [ + 496, + 167, + 892, + 380 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. StereoNVS Dataset", + "text_level": 1, + "bbox": [ + 498, + 396, + 684, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We propose the StereoNVS dataset, the first stereo-camera image dataset for training and evaluating novel-view synthesis using stereo-camera images. The StereoNVS dataset provides both real and synthetic datasets, each of which is dubbed StereoNVS-Real and StereoNVS-Synthetic. In the following, we present the details of each dataset.", + "bbox": [ + 496, + 422, + 890, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. StereoNVS-Real", + "text_level": 1, + "bbox": [ + 498, + 523, + 661, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "StereoNVS-Real provides real-world stereo-camera images for the training and evaluation of novel-view synthesis. The dataset provides stereo-camera images of 53 static scenes, and around 25 stereo-image pairs per scene. The images are undistorted and stereo-rectified, and have a resolution of $1792 \\times 896$ . The camera parameters such as the camera poses are provided as well. In our experiments, we divide the dataset into 45 and 8 scenes as training and test sets.", + "bbox": [ + 496, + 547, + 890, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To capture stereo images, we built a camera rig with two Basler machine vision cameras. We measured the camera parameters including the intrinsic and distortion parameters by camera calibration using multi-view images of a checkerboard [42]. Then, we collected stereo images from multiple viewpoints for various indoor and bounded scenes using our camera system. The captured images were then undistorted and stereo-rectified. Finally, we obtained the camera poses of the captured images using COLMAP [28].", + "bbox": [ + 496, + 669, + 890, + 805 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. StereoNVS-Synthetic", + "text_level": 1, + "bbox": [ + 498, + 816, + 699, + 832 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For quantitative evaluation of synthesized geometries, we also present the StereoNVS-Synthetic dataset, constructed by rendering synthetic 3D models using the 3D-Front dataset [10]. The images have a resolution of $864 \\times 448$ ,", + "bbox": [ + 496, + 840, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4943", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c1e53e4bba21ebda7abbf50d1f1c699b97a7f022312bf20229397601f804d6fc.jpg", + "image_caption": [ + "Figure 7. Qualitative comparison of novel-view synthesis on the StereoNVS dataset, showing rendering results for two real-world scenes (above) and one synthetic scene (below). All models are trained using stereo images. Our method outperforms the baseline methods [16, 21, 32, 33] on both real-world and synthetic scenes, especially in thin structures and textureless regions." + ], + "image_footnote": [], + "bbox": [ + 76, + 88, + 890, + 579 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "which is half of the real dataset. StereoNVS-Synthetic provides stereo-camera images of 50 scenes, and around 150 stereo image pairs per scene as well as the ground-truth camera parameters, camera poses, and depth maps. For more details about the StereoNVS dataset, refer to Sec. E in the supplementary document.", + "bbox": [ + 75, + 636, + 470, + 727 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 750, + 209, + 767 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct extensive validation of our method on the StereoNVS dataset. In the following, we will provide implementation details of our method (Sec. 5.1), compare our method with other baselines (Sec. 5.2), and conduct a comprehensive analysis of our proposed components, which proves the effectiveness of our framework including the stereo feature extractor, the depth-guided plane-sweeping, and the stereo depth loss (Sec. 5.3).", + "bbox": [ + 75, + 779, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/399e1227d9b7ac43944ee9d038f726b0341fe8f159873f9c2b34d0a3f90779c4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodStereoNVS-RealStereoNVS-Synthetic
PSNR(↑)SSIM(↑)LPIPS(↓)ABS(↓)PSNR(↑)SSIM(↑)LPIPS(↓)ABS(↓)
SRF21.120.69330.41642.917622.360.71620.42450.8125
IBRNet26.120.84210.20950.756630.830.88890.18230.2628
GeoNeRF28.010.89290.14600.506432.130.91790.14380.1577
GNT26.080.84340.22851.095926.170.84060.26500.4512
GeoNeRF+D25.650.81720.20820.869332.850.93210.11710.0782
NeuRay26.510.85380.18870.614732.260.91040.14780.1571
Ours28.440.90000.13960.486833.450.93360.12030.1056
", + "bbox": [ + 506, + 633, + 888, + 727 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparison between the baseline methods [4, 16, 21, 32, 33] and ours. All models are trained using stereo images. Our method shows superior performance on both StereoNVS-Real and StereoNVS-Synthetic datasets.", + "bbox": [ + 498, + 729, + 890, + 785 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 800, + 717, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We train our model on the training set of StereoNVS-Real, which has multi-view stereo image pairs of real scenes. Our model is trained for 250K iterations. For each iteration, we randomly select one scene and one target viewpoint of the scene. For both training and evaluation, image", + "bbox": [ + 496, + 824, + 890, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4944", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "features and feature volumes are extracted from three stereo image pairs (i.e., total six images) at three viewpoints nearest to the target viewpoint. During training, 512 rays are randomly selected for the training batch. We use the Adam optimizer [18] with learning rates of 0.0005 and the cosine annealing scheduling [22]. We employ UniMatch [35] for both the pre-trained stereo estimation network within the stereo feature extractor and the pseudo-ground-truth depth of the stereo depth loss. Refer to Sec. B in the supplementary document for additional implementation details.", + "bbox": [ + 75, + 90, + 472, + 243 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Comparison", + "text_level": 1, + "bbox": [ + 76, + 258, + 209, + 273 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare our method with recent generalizable novel view synthesis methods: SRF [4], IBRNet [33], GeoNeRF [16], GNT [32] and NeuRay [21]. Like our method, we use three stereo-camera image pairs to synthesize each target-view image for all the baseline methods. While the previous methods do not explicitly assume stereo-camera images as their inputs, we also train them using stereocamera images as they can handle stereo-camera images as independent inputs. All the baseline models are trained on the training set of StereoNVS-Real, as done for our method.", + "bbox": [ + 75, + 282, + 468, + 435 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate both image and depth qualities on the StereoNVS-Real and the StereoNVS-Synthetic datasets. For StereoNVS-Real, we utilize pseudo-ground-truth depths obtained from COLMAP [28] to assess depth quality. On the other hand, for StereoNVS-Synthetic, we use rendered depth maps as ground truths for depth quality assessment. We employ PSNR, SSIM [34], and LPIPS [41] as metrics for image quality and absolute error (ABS) as a metric for depth quality.", + "bbox": [ + 75, + 436, + 468, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Among the compared methods, for training and inference, NeuRay [21] requires depth maps and GeoNeRF [16] can use depth maps as additional inputs. We denote such variation of GeoNeRF as $\\mathrm{GeoNeRF}_{+D}$ . For their training and inference, we used the pseudo-ground-truth depth maps estimated by UniMatch [35] as done for our method.", + "bbox": [ + 75, + 575, + 470, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative comparison. Fig. 7 presents a qualitative comparison on novel view synthesis between our method and previous methods [16, 21, 32, 33] using the StereoNVS dataset. Previous methods estimate inaccurate geometry for scenes with thin structures or textureless regions, resulting in severe artifacts in the synthesized novel view images. In contrast, our method clearly outperforms the baseline methods in view synthesis results with significantly fewer artifacts even in textureless regions, thanks to more accurately estimated geometry. Moreover, our method shows better synthesis results compared to $\\mathrm{GeoNeRF}_{+D}$ [16] and Neu-Ray [21], even though they explicitly use depth maps in the inference time, demonstrating the robust utilization of depth maps in our framework.", + "bbox": [ + 75, + 688, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f121cc7082c298ad825d0effd0d2bc42fcf4993b9523d7c9d436d25dabd80d1f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
StereoNVS-RealStereoNVS-Synthetic
PSNR(↑)SSIM(↑)LPIPS(↓)ABS (↓)PSNR(↑)SSIM(↑)LPIPS(↓)ABS (↓)
Baseline (a)28.050.89530.13720.559832.220.91720.13860.1679
+ Stereo setting (b)28.010.89290.14600.506432.130.91790.14380.1577
+ SAM (c)28.210.89670.13980.493531.900.91670.13860.1531
+ Correlated feature (d)28.310.89880.13700.505732.350.92450.12820.1416
+ DGPS (e)28.420.89970.14030.509833.090.93040.12300.1246
+ Stereo depth loss (f)28.440.90000.13960.486833.450.93360.12030.1056
", + "bbox": [ + 504, + 89, + 890, + 165 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Quantitative ablation study.", + "bbox": [ + 586, + 167, + 803, + 180 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d74b88b8db2a0db5d67717e6819349684bd5e3eb740e0cdb82156df526d29f4e.jpg", + "image_caption": [ + "Tab. 2 (b)", + "Figure 8. Qualitative ablation study. Our full method enables accurate synthesis of a novel-view image and a depth map." + ], + "image_footnote": [], + "bbox": [ + 501, + 196, + 598, + 314 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/78f95b3819d8497d6f718295c0647b5e8e96701275ae00ad2ef409f567e776d7.jpg", + "image_caption": [ + "Tab. 2 (d)" + ], + "image_footnote": [], + "bbox": [ + 599, + 196, + 694, + 313 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2c15cedb309dbd6eff3912afe3f3d1e5c2cf8f2eecf67b8367c4dc6d4e975879.jpg", + "image_caption": [ + "Tab. 2 (e)" + ], + "image_footnote": [], + "bbox": [ + 696, + 196, + 792, + 313 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7f90db9a239d98b1ea3ec45d42e90b4b8594594817eb1c000e8ec7bf7cfd8c3b.jpg", + "image_caption": [ + "Tab. 2 (f)" + ], + "image_footnote": [], + "bbox": [ + 792, + 196, + 890, + 313 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative comparison. Tab. 1 shows that our method generally surpasses other baseline methods [4, 16, 21, 32, 33] in terms of image and depth qualities. GNT [32] exhibits degraded performances, likely due to its data-hungry transformer backbone. While performance differences with GeoNeRF are not substantial on the StereoNVS-Real dataset, our method shows the superior perceptual quality (Fig. 7) and significantly better performance on the StereoNVS-Synthetic dataset. While $\\mathrm{GeoNeRF}_{+D}$ [16] and NeuRay [21] achieve comparable performances on the StereoNVS-Synthetic dataset, they show considerably degenerated results on the StereoNVS-Real dataset. In contrast, our method demonstrates superior performance thanks to the robustness of our framework. Sec. D.1 in the supplementary document further discusses the sensitivity of $\\mathrm{GeoNeRF}_{+D}$ , NeuRay, and our method to depth errors in real-world images.", + "bbox": [ + 496, + 359, + 892, + 618 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Analysis and Discussion", + "text_level": 1, + "bbox": [ + 500, + 627, + 718, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3.1 Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 650, + 660, + 666 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To assess the impact of our proposed methods, we conduct an ablation study starting with our baseline model [16], which is trained on three views in a monocular setting (i.e., three images). Although equipped with a stereo camera setting, the baseline model trained on three stereo views (i.e., six images) shows similar image qualities as shown in Tab. 2 (b). This result indicates that sophisticated methods are needed to exploit the invaluable information from stereo-camera images.", + "bbox": [ + 496, + 674, + 890, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 2 shows that our model consistently demonstrates improvements in view synthesis results, as we introduce our proposed components. Our stereo feature extractor enables us to estimate better geometry, leading to fewer artifacts in novel view synthesis, as shown in Fig. 8 (d). The stereo attention module (c) and the stereo-correlated features (d)", + "bbox": [ + 496, + 810, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4945", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "help extract robust stereo image features, which are particularly effective for real scenes and synthetic scenes, respectively. Our DGPS is essential for better geometry estimation, leading to significantly improved results as shown in Fig. 8 (e), especially in textureless regions. This is further evident in Tab. 2 (e), with considerable performance gain on the StereoNVS-Synthetic dataset. Our final model shows the best performances (Tab. 2 (f)), with high-quality depth and view synthesis results (Fig. 8 (f)).", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3.2 Effectiveness of Depth-Guided Plane-Sweeping", + "text_level": 1, + "bbox": [ + 76, + 237, + 459, + 255 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct two experiments to demonstrate the effectiveness of our DGPS. In the first experiment, we compare our final model against a model without using DGPS. The \"Final model\" in Tab. 3 denotes our final model with DGPS. As shown in Tab. 3, the model without using DGPS shows degenerated results compared to our final model, highlighting the effectiveness of DGPS.", + "bbox": [ + 75, + 261, + 468, + 366 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the second experiment, we compare our model using DGPS against a model using more depth planes for cost volume construction. The term \"Base model\" in Tab. 4 refers to our model that is solely equipped with the stereo feature extractor, excluding DGPS and the stereo depth loss. We train an additional model that uses approximately 1.5 times the number of depth planes, compared to the base model. Although this additional model utilizes more depth planes, it shows similar results to the base model on StereoNeRF-Synthetic, as reported in Tab. 4. This is because increasing the number of planes does not guarantee accurate correspondence matching, especially in textureless regions. In contrast, our DGPS guarantees correspondence matching across multi-view image features near the geometry, resulting in better image and depth qualities. Refer to Sec. C.2.1 in the supplementary document for more details.", + "bbox": [ + 75, + 367, + 472, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3.3 Benefit of Stereo Estimation in Depth Loss", + "text_level": 1, + "bbox": [ + 76, + 619, + 426, + 635 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To show the effectiveness of using stereo estimation networks for depth supervision, we conduct an additional experiment as follows. First, we obtain two pseudo-GT depths: one from the pre-trained stereo network [35] ( $d_{gt}$ ) as stated in Sec. 3.4 and the other from the state-of-the-art learning-based MVS network ( $d_{gt}^{mvs}$ ) [26]. Then, we train two models with our methods using $d_{gt}$ and $d_{gt}^{mvs}$ , respectively. Then, we compare their synthesis results based on image and shape qualities.", + "bbox": [ + 75, + 643, + 468, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Tab. 5, our model trained with $d_{gt}$ surpasses the other model using $d_{gt}^{mvs}$ on StereoNVS-Synthetic. Note that the MVS network takes more images than the stereo network, seven images and two images, respectively. These results show that the stereo network provides a reliable depth signal for high-quality view synthesis thanks to its generalization ability, which came from standardized inputs of stereo-camera images and large-scale stereo datasets.", + "bbox": [ + 75, + 780, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ed7b766af2609f4b8bf2012478854bbd0bf62013eb7105299867e019f994a389.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PSNR(↑)SSIM(↑)LPIPS(↓)ABS(↓)
Final model (Tab. 2 (f))33.450.93360.12030.1056
- DGPS32.300.92400.12900.1300
", + "bbox": [ + 506, + 89, + 890, + 141 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/06185e3889705915fc254b143f91fd5e79bf2069a9dcfbd1cdb69005f51d5cd0.jpg", + "table_caption": [ + "Table 3. Effectiveness of our depth-guided plane-sweeping (DGPS) compared to the baseline models without using DGPS." + ], + "table_footnote": [], + "table_body": "
PSNR(↑)SSIM(↑)LPIPS(↓)ABS(↓)
Base model (Tab. 2 (d))32.350.92450.12820.1416
+ more depth planes32.330.92470.12990.1470
+ DGPS (Tab. 2 (e))33.090.93040.12300.1246
", + "bbox": [ + 506, + 179, + 890, + 244 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/adc8222181961adea9c874169718b55f7af5425759e1a8573c4a5665e4f3f027.jpg", + "table_caption": [ + "Table 4. Efficiency of our depth-guided plane-sweeping (DGPS) compared to the baseline models using more depth planes." + ], + "table_footnote": [], + "table_body": "
PSNR (↑)SSIM (↑)LPIPS (↓)ABS (↓)
Ours w/ MVS depth (dmvs)32.730.92770.12530.1216
Ours w/ Stereo depth (dtg)33.450.93360.12030.1056
", + "bbox": [ + 506, + 284, + 890, + 333 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Effectiveness of using stereo depths as pseudo-ground truth for depth loss compared to using MVS depths.", + "bbox": [ + 498, + 335, + 890, + 364 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 375, + 617, + 390 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper proposes StereoNeRF, a novel generalizable view synthesis framework leveraging stereo-camera images for high-quality novel-view synthesis. Due to the ill-posedness, previous methods struggle with accurate geometry estimation, which leads to severe artifacts in novel-view synthesis. Since the stereo matching provides vital information for accurate geometry reconstruction, our framework incorporates the stereo matching into NeRF-based generalizable view synthesis approach. To this end, we introduce a stereo feature extractor, a depth-guided plane-sweeping, and a stereo depth loss. We also present the StereoNVS dataset, the first stereo-camera image dataset for training and evaluating novel-view synthesis. Our extensive experiments show that StereoNeRF is effective in generalizable novel-view synthesis, particularly in scenes with complex structures or textureless regions.", + "bbox": [ + 496, + 400, + 890, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and Future Work. Our method is not free from limitations. In sparse view settings, our method produces blurry images and inaccurate geometry, issues also present in other methods. This limitation arises from the insufficient information for novel viewpoints in the sparse view settings. Our future work will involve additional geometric or generative prior, along with stereo prior, to compensate for the lack of information in such settings.", + "bbox": [ + 496, + 648, + 890, + 771 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. We thank Woohyeok Kim and Hyeongmin Lee for their assistance in acquisition of the StereoNVS-Real dataset. This work was supported by the NRF grant (No.2023R1A2C200494611, 2022R1A6A1A03052954, RS-2023-00211658) and IITP grant (No.2019-0-01906, Artificial Intelligence Graduate School Program (POSTECH)) funded by the Korea government (MSIT) and Samsung Electronics Co., Ltd.", + "bbox": [ + 496, + 773, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4946", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jia-Ren Chang and Yong-Sheng Chen. Pyramid stereo matching network. In CVPR, pages 5410-5418, 2018. 3", + "[2] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In ICCV, 2021. 1, 2", + "[3] Shuo Cheng, Zexiang Xu, Shilin Zhu, Zhuwen Li, Li Erran Li, Ravi Ramamoorthi, and Hao Su. Deep stereo using adaptive thin volume representation with uncertainty awareness. In CVPR, pages 2524-2534, 2020. 3", + "[4] Julian Chibane, Aayush Bansal, Verica Lazova, and Gerard Pons-Moll. Stereo radiance fields (srf): Learning view synthesis for sparse views of novel scenes. In CVPR. 1, 2, 6, 7", + "[5] Xiaojie Chu, Liangyu Chen, and Wenqing Yu. Nafssr: Stereo image super-resolution using nafnet. In CVPR, pages 1239-1248, 2022. 3", + "[6] Paul E Debevec, Camillo J Taylor, and Jitendra Malik. Modeling and rendering architecture from photographs: A hybrid geometry-and image-based approach. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 11-20, 1996. 2", + "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2", + "[8] Yilun Du, Cameron Smith, Ayush Tewari, and Vincent Sitzmann. Learning to render novel views from wide-baseline stereo pairs. arXiv preprint arXiv:2304.08463, 2023. 2", + "[9] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, pages 5501–5510, 2022. 2", + "[10] Huan Fu, Bowen Cai, Lin Gao, Ling-Xiao Zhang, Jiaming Wang, Cao Li, Qixun Zeng, Chengyue Sun, Rongfei Jia, Binqiang Zhao, et al. 3d-front: 3d furnished rooms with layouts and semantics. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10933-10942, 2021. 5", + "[11] Yasutaka Furukawa, Carlos Hernández, et al. Multi-view stereo: A tutorial. Foundations and Trends® in Computer Graphics and Vision, 9(1-2):1-148, 2015. 3", + "[12] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, pages 3354-3361. IEEE, 2012. 1, 3", + "[13] Steven J Gortler, Radek Grzeszczuk, Richard Szeliski, and Michael F Cohen. The lumigraph. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 43-54, 1996. 2", + "[14] Xiaodong Gu, Zhiwen Fan, Siyu Zhu, Zuozhuo Dai, Feitong Tan, and Ping Tan. Cascade cost volume for high-resolution multi-view stereo and stereo matching. In CVPR, pages 2495-2504, 2020. 3" + ], + "bbox": [ + 78, + 114, + 467, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 3", + "[16] Mohammad Mahdi Johari, Yann Lepoittevin, and François Fleuret. Geonerf: Generalizing nerf with geometry priors. In CVPR, pages 18365-18375, 2022. 1, 2, 3, 5, 6, 7", + "[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM TOG, 42(4):1-14, 2023. 2", + "[18] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 7", + "[19] Kiriakos N Kutulakos and Steven M Seitz. A theory of shape by space carving. International journal of computer vision, 38:199-218, 2000. 3", + "[20] Marc Levoy and Pat Hanrahan. Light field rendering. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 31-42, 1996. 2", + "[21] Yuan Liu, Sida Peng, Lingjie Liu, Qianqian Wang, Peng Wang, Christian Theobalt, Xiaowei Zhou, and Wenping Wang. Neural rays for occlusion-aware image-based rendering. In CVPR, pages 7824-7833, 2022. 1, 2, 6, 7", + "[22] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 7", + "[23] Nikolaus Mayer, Eddy Ilg, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In CVPR, pages 4040-4048, 2016. 1, 3", + "[24] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2, 3, 5", + "[25] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM TOG, 41(4):1-15, 2022. 2", + "[26] Rui Peng, Rongjie Wang, Zhenyu Wang, Yawen Lai, and Ronggang Wang. Rethinking depth estimation for multiview stereo: A unified representation. In CVPR, pages 8645-8654, 2022. 2, 8", + "[27] Daniel Scharstein and Richard Szeliski. A taxonomy and evaluation of dense two-frame stereo correspondence algorithms. IJCV, 47:7-42, 2002. 1, 3", + "[28] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, pages 4104-4113, 2016. 4, 5, 7", + "[29] Steven M Seitz and Charles R Dyer. Photorealistic scene reconstruction by voxel coloring. International Journal of Computer Vision, 35:151-173, 1999. 3", + "[30] Mohammed Suhail, Carlos Esteves, Leonid Sigal, and Ameesh Makadia. Generalizable patch-based neural rendering. In ECCV, pages 156-174. Springer, 2022. 1, 2", + "[31] Alex Trevithick and Bo Yang. Grf: Learning a general radiance field for 3d representation and rendering. In ICCV, pages 15182-15192, 2021. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4947", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Mukund Varma, Peihao Wang, Xuxi Chen, Tianlong Chen, Subhashini Venugopalan, and Zhangyang Wang. Is attention all that nef needs? In ICLR, 2022. 1, 2, 6, 7", + "[33] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In CVPR, pages 4690-4699, 2021. 1, 2, 6, 7", + "[34] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 13(4):600-612, 2004. 7", + "[35] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, Fisher Yu, Dacheng Tao, and Andreas Geiger. Unifying flow, stereo and depth estimation. IEEE TPAMI, 2023. 2, 3, 5, 7, 8", + "[36] Jiayu Yang, Wei Mao, Jose M Alvarez, and Miaomiao Liu. Cost volume pyramid based depth inference for multi-view stereo. In CVPR, pages 4877-4886, 2020. 3", + "[37] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoptrees for real-time rendering of neural radiance fields. In ICCV, pages 5752-5761, 2021. 2", + "[38] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In CVPR, pages 4578-4587, 2021. 1, 2", + "[39] Feihu Zhang, Victor Prisacariu, Ruigang Yang, and Philip HS Torr. Ga-net: Guided aggregation net for end-to-end stereo matching. In CVPR, pages 185-194, 2019. 3", + "[40] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020.5", + "[41] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 7", + "[42] Zhengyou Zhang. A flexible new technique for camera calibration. IEEE TPAMI, 22(11):1330-1334, 2000. 5", + "[43] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. ACM TOG, 2018. 2" + ], + "bbox": [ + 78, + 90, + 467, + 645 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4948", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Generalizable Novel-View Synthesis using a Stereo Camera/516ec60a-018b-4106-92f4-813c127e6028_model.json b/2024/Generalizable Novel-View Synthesis using a Stereo Camera/516ec60a-018b-4106-92f4-813c127e6028_model.json new file mode 100644 index 0000000000000000000000000000000000000000..270cec9a13013ed7799ca8837d4cec3ad6a8df47 --- /dev/null +++ b/2024/Generalizable Novel-View Synthesis using a Stereo Camera/516ec60a-018b-4106-92f4-813c127e6028_model.json @@ -0,0 +1,2156 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.187, + 0.131, + 0.785, + 0.154 + ], + "angle": 0, + "content": "Generalizable Novel-View Synthesis using a Stereo Camera" + }, + { + "type": "text", + "bbox": [ + 0.14, + 0.18, + 0.266, + 0.198 + ], + "angle": 0, + "content": "Haechan Lee\\(^{1,*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.31, + 0.181, + 0.432, + 0.199 + ], + "angle": 0, + "content": "Wonjoon Jin\\(^{2,*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.476, + 0.181, + 0.646, + 0.199 + ], + "angle": 0, + "content": "Seung-Hwan Baek\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.689, + 0.181, + 0.829, + 0.199 + ], + "angle": 0, + "content": "Sunghyun Cho\\(^{1,2}\\)" + }, + { + "type": "title", + "bbox": [ + 0.375, + 0.206, + 0.596, + 0.223 + ], + "angle": 0, + "content": "POSTECH \\(^{1}\\)GSAI & \\(^{2}\\)CSE" + }, + { + "type": "text", + "bbox": [ + 0.267, + 0.227, + 0.7, + 0.242 + ], + "angle": 0, + "content": "{gocks8, jinwj1996, shwbaek, s.cho}@postech.ac.kr" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.277, + 0.314, + 0.293 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.309, + 0.474, + 0.566 + ], + "angle": 0, + "content": "In this paper, we propose the first generalizable view synthesis approach that specifically targets multi-view stereocamera images. Since recent stereo matching has demonstrated accurate geometry prediction, we introduce stereo matching into novel-view synthesis for high-quality geometry reconstruction. To this end, this paper proposes a novel framework, dubbed StereoNeRF, which integrates stereo matching into a NeRF-based generalizable view synthesis approach. StereoNeRF is equipped with three key components to effectively exploit stereo matching in novel-view synthesis: a stereo feature extractor, a depth-guided plane-sweeping, and a stereo depth loss. Moreover, we propose the StereoNVS dataset, the first multi-view dataset of stereocamera images, encompassing a wide variety of both real and synthetic scenes. Our experimental results demonstrate that StereoNeRF surpasses previous approaches in generalizable view synthesis." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.58, + 0.21, + 0.595 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.605, + 0.47, + 0.756 + ], + "angle": 0, + "content": "Novel-view synthesis is a long-standing ill-posed problem in computer vision and graphics, which is inherently challenging due to the necessity of predicting both the geometry and texture from images of a target scene. Recently, Neural Radiance Fields (NeRF) [24] have achieved photorealistic results by jointly optimizing geometry and radiance fields with a coordinate-based network. However, the need for per-scene optimization, which has to learn representations for each target scene individually, restricts its applicability, as it requires additional training time for such optimization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.757, + 0.471, + 0.879 + ], + "angle": 0, + "content": "Recent approaches [2, 4, 16, 21, 30, 32, 33, 38] have addressed this issue of synthesizing novel-view images on-the-fly for unseen scenes without per-scene optimization. Early studies [4, 33, 38] utilize an image encoder to train a generic view interpolation function, enabling the estimation of NeRF parameters from unseen input images in a feed-forward manner. However, this single feed-forward manner for estimating geometry and color exacerbates the ill" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.277, + 0.627, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.531, + 0.4, + 0.603, + 0.41 + ], + "angle": 0, + "content": "(a) Baseline" + }, + { + "type": "image", + "bbox": [ + 0.628, + 0.277, + 0.753, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.656, + 0.4, + 0.728, + 0.41 + ], + "angle": 0, + "content": "(b) Baseline" + }, + { + "type": "image_caption", + "bbox": [ + 0.639, + 0.411, + 0.745, + 0.421 + ], + "angle": 0, + "content": "(using stereo images)" + }, + { + "type": "image", + "bbox": [ + 0.752, + 0.277, + 0.895, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.791, + 0.4, + 0.842, + 0.41 + ], + "angle": 0, + "content": "(c) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.764, + 0.411, + 0.87, + 0.421 + ], + "angle": 0, + "content": "(using stereo images)" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.425, + 0.895, + 0.482 + ], + "angle": 0, + "content": "Figure 1. Novel-view synthesis results of a baseline method [16] and ours. The baseline shows degraded performances, even trained using stereo-camera images (b). In contrast, fully exploiting stereo-camera images, our method shows superior results (c)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.491, + 0.895, + 0.597 + ], + "angle": 0, + "content": "posedness, resulting in low-quality geometries and rendering results. For better geometry reasoning, MVSNeRF [2] and GeoNeRF [16] leverage the multi-view stereo (MVS) approach to handle occlusions in a 3D scene. Nonetheless, they struggle with inaccurate geometry prediction and limited synthesis accuracy, as shown in the synthesis result of GeoNeRF (Fig. 1 (a))." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.896, + 0.902 + ], + "angle": 0, + "content": "To tackle this challenge, we propose the first generalizable NeRF approach that leverages stereo-camera images, which are easily accessible thanks to the ubiquity of stereo cameras in most mobile devices. Recent advance in learning-based stereo estimation has demonstrated accurate geometry prediction, often even outperforming learning-based MVS methods as shown in Fig. 2. The superior performance of stereo estimation can be attributed to several key factors. First, unlike MVS that assumes arbitrary number of inputs with arbitrary camera positions and orientations, stereo matching assumes two stereo inputs with a fixed baseline. This constraint allows a more optimal network architecture that can effectively find dense correspondences between input images, and significantly eases the training difficulty. Moreover, larger scale of stereomatching datasets [12, 23, 27] compared to MVS datasets has facilitated remarkable generalization capabilities in stereo estimation network. Therefore, we aim to harness this accurate geometric information from stereo images to alleviate the ill-posedness of generalizable view synthesis." + }, + { + "type": "page_footnote", + "bbox": [ + 0.102, + 0.888, + 0.214, + 0.9 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4939" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.468, + 0.152 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.156, + 0.468, + 0.198 + ], + "angle": 0, + "content": "Figure 2. Usefulness of exploiting binocular stereo. Comparison on depth estimation between a learning-based MVS method [26] and a learning-based binocular stereo method [35]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.206, + 0.468, + 0.266 + ], + "angle": 0, + "content": "However, since previous methods do not explicitly consider stereoscopic inputs, they cannot leverage the aforementioned benefits, resulting in degenerated performances shown in Fig. 1 (b) despite using stereo-camera images." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.267, + 0.468, + 0.584 + ], + "angle": 0, + "content": "This paper proposes StereoNeRF, a novel generalizable view synthesis framework leveraging stereo images. StereoNeRF integrates stereo matching into NeRF-based generalizable view synthesis approach, where the stereo matching provides vital geometric information. To this end, we first introduce a stereo feature extractor, which extracts geometry-aware features by correlating horizontal epipolar lines within stereo images. In addition, the stereo feature extractor takes advantage of stereo-correlated features from an off-the-shelf stereo estimation network, which can transfer rich geometric knowledge to our model. Furthermore, with the reliable depth estimated from the stereo estimation network, we aggregate multi-view features through a depth-guided plane-sweeping technique. This technique ensures correspondence matching around the geometry in cost volume construction. We also present a stereo depth loss utilizing the estimated stereo depth. These additional geometric cues from the stereo matching effectively mitigate the ill-posedness in generalizable view synthesis. Notably, our framework leveraging stereo images surpasses the previous approaches [16, 21], which rely on extra depth information." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.585, + 0.468, + 0.719 + ], + "angle": 0, + "content": "Furthermore, we propose the StereoNVS dataset, which is the first dataset for training and evaluation of novel-view synthesis using stereo-camera images. Our StereoNVS dataset provides real-world and synthetic stereo images. Our extensive evaluation on the StereoNVS dataset shows that stereo-pair inputs can effectively enhance the quality of novel-view synthesis, and shows that StereoNeRF outperforms previous generalizable novel-view synthesis approaches in terms of image and shape qualities." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.72, + 0.316, + 0.734 + ], + "angle": 0, + "content": "Our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.735, + 0.468, + 0.779 + ], + "angle": 0, + "content": "- We propose a generalized NeRF approach that leverages stereo-camera images for the first time to alleviate the ill-posedness in generalizable novel-view synthesis." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.781, + 0.468, + 0.855 + ], + "angle": 0, + "content": "- We propose a novel framework, StereoNeRF, which exploits the benefits of stereo images by integrating stereo matching into generalizable view synthesis. To this end, we present a stereo feature extractor, a depth-guided plane-sweeping, and a stereo depth loss in our framework." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.856, + 0.468, + 0.9 + ], + "angle": 0, + "content": "- We also present the StereoNVS dataset, the first dataset for training and evaluation of novel-view synthesis obtained by stereo cameras." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.735, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.642, + 0.105 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.115, + 0.7, + 0.13 + ], + "angle": 0, + "content": "2.1. Novel-View Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.138, + 0.892, + 0.38 + ], + "angle": 0, + "content": "Novel-view synthesis aims to synthesize target-view images from reference-view images. Early approaches synthesize novel-view images by blending pixels from multiple input images [6, 13, 20]. Recent work adopting neural volume representations has shown remarkable novel-view synthesis results. Zhou et al. [43] propose multi-plane images (MPI) representation estimated from input images, but their methods produce valid novel-view images only for narrow ranges of camera poses. Mildenhall et al. [24] propose Neural Radiance Fields (NeRF) that can synthesize photo-realistic target-view images via neural implicit representations and volume rendering. Albeit its photo-realism, computation-heavy per-scene optimization is needed. Recent variants such as [9, 17, 25, 37] have remarkably reduced the optimization time, but they still require large memory and several minutes for training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.387, + 0.892, + 0.538 + ], + "angle": 0, + "content": "Generalizable View Synthesis. Recently, many novel-view synthesis approaches without per-scene optimization have been proposed. Several studies directly predict pixel colors by aggregating image features from aligned pixels, without 3D representations. Suhail et al. [30] and Varma et al. [32] adopt a transformer-based network [7] to compute features along the epipolar lines, which needs a large number of training images for high-quality view synthesis. Du et al. [8] propose a framework to synthesize target views from two images with small overlapped regions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.892, + 0.734 + ], + "angle": 0, + "content": "Another research direction predicts volumetric representation [24] from aggregated features from reference views and synthesizes images via volumetric rendering. PixelNeRF, SRF, and GRF [4, 31, 38] predict radiance fields from pixel-aligned features using an MLP. Among these methods, SRF [4] uses two sampled images from an image collection captured by a monocular camera as a stereo pair, but shows limited synthesis quality. IBRNet [33] proposes to learn generic view interpolation functions, but suffers from artifacts for challenging scenes with complex geometries. MVSNeRF, GeoNeRF and NeuRay [2, 16, 21] utilize the MVS approach using cost volume for better occlusion handling in generalizable view synthesis." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.734, + 0.892, + 0.84 + ], + "angle": 0, + "content": "However, all the aforementioned methods often fail to capture accurate geometry, particularly in textureless regions, leading to limited synthesis results. Unlike these methods, our framework exploits stereo-camera images with a fixed baseline to effectively capture the geometry of complex scenes as well as textureless regions for high-quality novel-view synthesis." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.842, + 0.864 + ], + "angle": 0, + "content": "2.2. Stereo Matching and Multi-View Stereo" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Geometry estimation is a long-standing problem in computer vision that has a variety of applications. Among them," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4940" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.273 + ], + "angle": 0, + "content": "stereo matching is a task that takes rectified stereo images and computes stereo correspondence to estimate disparities [15, 27]. Recently, a huge number of data-driven approaches have been introduced with the emergence of a vast amount of stereo-matching datasets [12, 23, 27] and made significant progress [1, 23, 35, 39]. MVS approaches that use more than two views have been extensively studied as well [11, 19, 29]. Recently, learning-based MVS approaches have been proposed, e.g., Cheng et al. [3], Gu et al. [14], and Yang et al. [36] present efficient frameworks that cascade cost volumes in a coarse-to-fine manner to enable high-resolution depth estimation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.275, + 0.473, + 0.397 + ], + "angle": 0, + "content": "In our work, we have the best of both worlds in our generalizable view synthesis framework. We adopt the MVS approach to aggregate multi-view information for occlusion-aware geometry estimation, as done in GeoNerveRF [16]. Furthermore, our framework also assumes structured stereo images as inputs, and integrates the two-view stereo matching into our framework for robust and accurate geometry estimation in generalizable novel-view synthesis." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.414, + 0.206, + 0.43 + ], + "angle": 0, + "content": "3. StereoNeRF" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.442, + 0.473, + 0.593 + ], + "angle": 0, + "content": "For novel-view synthesis, our framework takes a set of rectified stereo-camera images and estimates neural radiance fields [24] from which novel-view images are rendered. StereoNeRF integrates a well-designed stereo-matching algorithm into the existing generalizable view synthesis approach [16]. In this section, we first provide an overall pipeline that utilizes an pre-trained stereo estimation network [35]. Then, we explain each step of our framework and the training process in detail, highlighting how to integrate stereo matching into our framework." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Overall pipeline. Fig. 3 shows an overview of our pipeline. Our framework utilizes \\(N\\) pairs of stereo images of a target scene \\(\\{(I_L^n,I_R^n)\\}_{n = 1}^N\\) for novel-view synthesis. In the first step, the pre-trained stereo estimation network takes the \\(n\\)-th image pair \\((I_L^n,I_R^n)\\), and outputs stereo depths \\((d_{s,L}^{n},d_{s,R}^{n})\\) and stereo-correlated features \\((t_L^n,t_R^n)\\). In the second step, a stereo feature extractor takes the stereo image pair and the stereo-correlated features, and outputs stereo image features \\((f_L^n,f_R^n)\\), explained in Sec. 3.1. The third step aggregates stereo image features from all viewpoints \\(\\{(f_L^n,f_R^n)\\}_{n = 1}^N\\) to build 3D feature volumes \\(\\{(\\phi_L^n,\\phi_R^n)\\}_{n = 1}^N\\) using an MVS network (Sec. 3.2). For constructing feature volumes that faithfully reflect the 3D geometry of a target scene, the third step adopts a depth-guided plane-sweeping which utilizes the estimated stereo depths \\((d_{s,L}^{n},d_{s,R}^{n})\\). Finally, a neural renderer predicts radiance fields from the stereo image features \\(\\{(f_L^n,f_R^n)\\}_{n = 1}^N\\) and the feature volumes \\(\\{(\\phi_L^n,\\phi_R^n)\\}_{n = 1}^N\\) from all viewpoints, and novel-view images are synthesized from the radiance fields (Sec. 3.3)." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.895, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.253, + 0.895, + 0.338 + ], + "angle": 0, + "content": "Figure 3. Overview of StereoNeRF. StereoNeRF consists of a shared feature extractor and a neural renderer. We design the feature extractor with a stereo estimation network, a stereo feature extractor, and a MVS network. StereoNeRF takes \\( N \\) pairs of stereo images and a camera pose as inputs, and synthesizes a novel-view image of the camera pose." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.349, + 0.874, + 0.366 + ], + "angle": 0, + "content": "3.1. Feature Extraction from Stereo Image Pairs" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.372, + 0.892, + 0.553 + ], + "angle": 0, + "content": "Unlike previous approaches, which extract features of input images respectively, our stereo feature extractor takes rectified stereo-camera images and computes image feature maps by exploiting the epipolar geometry between them. To this end, our stereo feature extractor consists of three parts: CNN encoders, stereo attention modules (SAM) [5], and CNN decoders (Fig. 4). Moreover, we propose integrating the stereo-correlated features \\((t_L^n,t_R^n)\\) from the pretrained stereo estimation network into the SAM (green line in Fig. 4). This integration inherits geometric cues from the stereo estimation network. In the following, we provide the detailed description of the stereo feature extractor." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.893, + 0.75 + ], + "angle": 0, + "content": "First, the weight-shared CNN encoders project each of \\( I_L^n \\) and \\( I_R^n \\) into the feature space. Then, we adopt the stereo attention module (SAM) [5] to fuse two feature maps on the horizontal epipolar lines. To this end, SAM estimates stereo correspondences between two feature maps and explicitly adds stereo-correspondent features, as shown in Fig. 5. In addition, we extend the SAM by aggregating the stereocorrelated features \\( (t_L^n, t_R^n) \\) in this feature fusion process. Note that the stereo-correlated features are computed from the pre-trained stereo estimation network, which provides vital geometric cues. These stereo-correlated features enhance stereo correspondences between two feature maps, leading to higher-quality fused feature maps." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Specific explanation of the aforementioned feature fusion of the SAM is as follows (Fig. 5). For the left feature, a warping matrix \\(\\mathcal{W}_{R\\to L}\\) is built by multiplying a query matrix and a key matrix followed by a softmax layer. The query matrix is computed from the left feature and the stereocorrelated features, and the key matrix is computed from the right feature and the stereo-correlated features. Then, a value matrix, computed from the right feature and the stereo-correlated feature, is warped along the epipolar lines by multiplying the warping matrix \\(\\mathcal{W}_{R\\to L}\\), then added to" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4941" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.088, + 0.473, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.467, + 0.47, + 0.495 + ], + "angle": 0, + "content": "Figure 4. Stereo feature extractor of StereoNeRF, where an pretrained stereo estimation network is incorporated." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.47, + 0.537 + ], + "angle": 0, + "content": "the left feature. The same feature fusion process is also performed for the right feature." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.537, + 0.47, + 0.629 + ], + "angle": 0, + "content": "For effective fusion of information from both stereo images and stereo-correlated features, the stereo feature extractor has three sequentially stacked SAM. Finally, the weight-shared CNN decoders take each of the fused features from the SAM, and generate stereo image features \\((f_{L}^{n}, f_{R}^{n})\\), which will be used for building feature volumes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.636, + 0.45, + 0.652 + ], + "angle": 0, + "content": "3.2. Depth-guided Feature Volume Construction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.47, + 0.84 + ], + "angle": 0, + "content": "Once stereo image features are obtained, we aggregate these features from all viewpoints to create cost volumes. To this end, we adopt a plane-sweeping-based approach, which computes multi-view correspondences among these features. The plane-sweeping-based approach first defines a depth range within pre-computed near and far depths obtained from COLMAP [28]. Then, it hypothesizes depth planes in the entire depth range. On these depth planes, multi-view stereo image features are aggregated via plane-sweeping to build cost volumes. Then, these cost volumes are processed to yield feature volumes, which will be used later in predicting neural radiance fields." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "However, this plane-sweeping-based approach often struggles with accurate geometry estimation. Since this approach searches the entire depth range of a scene, correspondence matching is inaccurate and prone to error, espe" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.091, + 0.895, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.327, + 0.892, + 0.368 + ], + "angle": 0, + "content": "Figure 5. Stereo attention module used in the stereo feature extractor. We exploit the rich features from the pre-trained stereo estimation module." + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.376, + 0.88, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.527, + 0.851, + 0.541 + ], + "angle": 0, + "content": "Figure 6. Feature volume construction using DGPS." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.563, + 0.892, + 0.685 + ], + "angle": 0, + "content": "cially in textureless regions. To tackle this issue, we introduce a depth-guided plane-sweeping (DGPS), which constructs cost volumes around the stereo depth predicted from the stereo estimation network. Unlike the previous approach using the entire depth range, DGPS introduces a dynamic search range that varies based on the estimated stereo depth. This technique significantly constrains the search space and reduces outliers in correspondence matching." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.688, + 0.893, + 0.9 + ], + "angle": 0, + "content": "Fig. 6 depicts the feature volume construction using DGPS. Instead of searching the entire depth range, we hypothesize depth planes around the stereo depths \\((d_{s,L}^n,d_{s,R}^n)\\). Then, through DGPS, we aggregate the stereo image features from all viewpoints \\(\\{(f_L^n,f_R^n)\\}_{n = 1}^N\\) on these depth planes to build cost volumes. These cost volumes are further processed via the MVS network to obtain feature volumes \\((\\phi_L^n,\\phi_R^n)\\) and depth maps \\((d_{s,L}^n,d_{s,R}^n)\\). We repeat this process to obtain feature volumes and depth maps for every viewpoint. The resulting multi-view feature volumes and depth maps are denoted as \\(\\{(\\phi_L^L,\\phi_R^L)\\}_{n = 1}^N\\) and \\(\\{(d_{m,L}^{n},d_{m,R}^{n})\\}_{n = 1}^{N}\\), respectively. Refer to Sec. B.3 in the supplementary document for more details about our feature volume construction." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4942" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.294, + 0.108 + ], + "angle": 0, + "content": "3.3. Rendering Novel Views" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.114, + 0.473, + 0.356 + ], + "angle": 0, + "content": "Once feature volumes are constructed, novel-view images can be rendered via neural rendering. For rendering novel-view images, our framework adopts the rendering procedure of GeoNeRF [16]. In the following, we briefly describe the rendering procedure. Given a target viewpoint, we cast a ray for each pixel and sample points along the ray in the 3D space. Then, for each sampled point, we sample features from all the feature volumes \\(\\{(\\phi_L^n,\\phi_R^n)\\}_{n = 1}^N\\). We also sample features from all the image features \\(\\{(f_L^n,f_R^n)\\}_{n = 1}^N\\) by projecting the sampled point onto input images. The sampled volume and image features are aggregated by a neural renderer network, and the color and density are estimated. Finally, by integrating the estimated color values and depth values of sampled points with their densities along each ray, we obtain the color \\(\\hat{c}\\) and the depth \\(d_{r}\\) at each pixel of a novel view, respectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.365, + 0.279, + 0.381 + ], + "angle": 0, + "content": "3.4. Training StereoNeRF" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.388, + 0.47, + 0.464 + ], + "angle": 0, + "content": "We train StereoNeRF with ground-truth images and pseudoground-truth depths for high-quality novel-view synthesis. The pseudo-ground-truth depths \\( d_{gt} \\) are obtained from the pre-trained stereo estimation network [35]. Our training loss \\( \\mathcal{L} \\) is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.223, + 0.475, + 0.47, + 0.492 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {c} + \\mathcal {L} _ {d}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.502, + 0.47, + 0.547 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_c\\) is a color loss and \\(\\mathcal{L}_d\\) is a depth loss. The color loss is defined as a mean-squared-error (MSE) between the rendered colors \\(\\hat{c}\\) and ground-truth colors [24]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.548, + 0.47, + 0.593 + ], + "angle": 0, + "content": "To alleviate the shape-radiance ambiguity [40] in scene reconstruction using neural radiance fields, we introduce the depth loss \\(\\mathcal{L}_d\\), which is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.602, + 0.47, + 0.623 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d} = \\lambda_ {d} ^ {s e l f} \\mathcal {L} _ {d} ^ {s e l f} + \\lambda_ {d} ^ {s t e r e o} \\mathcal {L} _ {d} ^ {s t e r e o}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.633, + 0.469, + 0.712 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_d^{self}\\) is a self-supervised depth loss from GeoN-NeRF [16] and \\(\\mathcal{L}_d^{stereo}\\) is our proposed stereo depth loss. \\(\\mathcal{L}_d^{self}\\) penalizes the depths estimated from both the stereo estimation network \\((d_s)\\) and the MVS network \\((d_m)\\) by comparing them to the depth rendered from NeRF \\((d_r)\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.712, + 0.469, + 0.819 + ], + "angle": 0, + "content": "Our framework estimates depths from the stereo estimation network \\((d_{s})\\), the MVS network \\((d_{m})\\), and the neural renderer network \\((d_{r})\\). The stereo depth loss \\(\\mathcal{L}_d^{stereo}\\) guides these networks to predict more accurate depths using the pseudo-ground-truth depths \\(d_{gt}\\). This loss introduces a further quality improvement, especially in geometric details as will be shown in Sec. 5.3. \\(\\mathcal{L}_d^{stereo}\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.828, + 0.47, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d} ^ {\\text {s t e r e o}} = \\lambda_ {1} \\mathcal {L} _ {d} ^ {s} + \\lambda_ {2} \\mathcal {L} _ {d} ^ {m} + \\lambda_ {3} \\mathcal {L} _ {d} ^ {r}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.903 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_d^s\\), \\(\\mathcal{L}_d^m\\), and \\(\\mathcal{L}_d^r\\) penalizes depths \\(d_s\\), \\(d_m\\), and \\(d_r\\), respectively, by comparing them with \\(d_{gt}\\). Note that \\(d_{gt}\\) is different from \\(d_s\\). We obtain \\(d_{gt}\\) from the pre-trained stereo" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "estimation network with frozen parameters. On the other hand, we obtain \\( d_{s} \\) from the stereo estimation network in our framework, which has trainable parameters. Refer to Sec. B.4 in the supplementary document for more details about our stereo depth loss." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.168, + 0.893, + 0.381 + ], + "angle": 0, + "content": "Due to the different characteristics of datasets such as baseline length, \\( d_{s} \\) may have estimation error, which leads to an error in the depth-guided plane-sweeping. To tackle this, we partially train the matching and propagation network in the stereo estimation network (Fig. 4) using \\( \\mathcal{L}_d^{self} \\). \\( \\mathcal{L}_d^{self} \\) provides multi-view supervision to the stereo estimation network by leveraging \\( d_{r} \\) estimated from multi-view images. However, since \\( d_{r} \\) may also have estimation error, we further regularize the stereo estimation network using \\( \\mathcal{L}_d^s \\). This training scheme for the stereo estimation network ensures consistent depth estimation, while preserving the stereo matching capability. Sec. D.1 in the supplementary document further discusses the training scheme of the stereo estimation network." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.397, + 0.686, + 0.413 + ], + "angle": 0, + "content": "4. StereoNVS Dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.423, + 0.892, + 0.514 + ], + "angle": 0, + "content": "We propose the StereoNVS dataset, the first stereo-camera image dataset for training and evaluating novel-view synthesis using stereo-camera images. The StereoNVS dataset provides both real and synthetic datasets, each of which is dubbed StereoNVS-Real and StereoNVS-Synthetic. In the following, we present the details of each dataset." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.524, + 0.663, + 0.539 + ], + "angle": 0, + "content": "4.1. StereoNVS-Real" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.548, + 0.892, + 0.669 + ], + "angle": 0, + "content": "StereoNVS-Real provides real-world stereo-camera images for the training and evaluation of novel-view synthesis. The dataset provides stereo-camera images of 53 static scenes, and around 25 stereo-image pairs per scene. The images are undistorted and stereo-rectified, and have a resolution of \\(1792 \\times 896\\). The camera parameters such as the camera poses are provided as well. In our experiments, we divide the dataset into 45 and 8 scenes as training and test sets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.67, + 0.892, + 0.806 + ], + "angle": 0, + "content": "To capture stereo images, we built a camera rig with two Basler machine vision cameras. We measured the camera parameters including the intrinsic and distortion parameters by camera calibration using multi-view images of a checkerboard [42]. Then, we collected stereo images from multiple viewpoints for various indoor and bounded scenes using our camera system. The captured images were then undistorted and stereo-rectified. Finally, we obtained the camera poses of the captured images using COLMAP [28]." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.817, + 0.7, + 0.833 + ], + "angle": 0, + "content": "4.2. StereoNVS-Synthetic" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.892, + 0.901 + ], + "angle": 0, + "content": "For quantitative evaluation of synthesized geometries, we also present the StereoNVS-Synthetic dataset, constructed by rendering synthetic 3D models using the 3D-Front dataset [10]. The images have a resolution of \\(864 \\times 448\\)," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4943" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.089, + 0.892, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.582, + 0.893, + 0.625 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparison of novel-view synthesis on the StereoNVS dataset, showing rendering results for two real-world scenes (above) and one synthetic scene (below). All models are trained using stereo images. Our method outperforms the baseline methods [16, 21, 32, 33] on both real-world and synthetic scenes, especially in thin structures and textureless regions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.637, + 0.471, + 0.728 + ], + "angle": 0, + "content": "which is half of the real dataset. StereoNVS-Synthetic provides stereo-camera images of 50 scenes, and around 150 stereo image pairs per scene as well as the ground-truth camera parameters, camera poses, and depth maps. For more details about the StereoNVS dataset, refer to Sec. E in the supplementary document." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.751, + 0.21, + 0.768 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.47, + 0.901 + ], + "angle": 0, + "content": "We conduct extensive validation of our method on the StereoNVS dataset. In the following, we will provide implementation details of our method (Sec. 5.1), compare our method with other baselines (Sec. 5.2), and conduct a comprehensive analysis of our proposed components, which proves the effectiveness of our framework including the stereo feature extractor, the depth-guided plane-sweeping, and the stereo depth loss (Sec. 5.3)." + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.635, + 0.89, + 0.728 + ], + "angle": 0, + "content": "
MethodStereoNVS-RealStereoNVS-Synthetic
PSNR(↑)SSIM(↑)LPIPS(↓)ABS(↓)PSNR(↑)SSIM(↑)LPIPS(↓)ABS(↓)
SRF21.120.69330.41642.917622.360.71620.42450.8125
IBRNet26.120.84210.20950.756630.830.88890.18230.2628
GeoNeRF28.010.89290.14600.506432.130.91790.14380.1577
GNT26.080.84340.22851.095926.170.84060.26500.4512
GeoNeRF+D25.650.81720.20820.869332.850.93210.11710.0782
NeuRay26.510.85380.18870.614732.260.91040.14780.1571
Ours28.440.90000.13960.486833.450.93360.12030.1056
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.73, + 0.892, + 0.786 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison between the baseline methods [4, 16, 21, 32, 33] and ours. All models are trained using stereo images. Our method shows superior performance on both StereoNVS-Real and StereoNVS-Synthetic datasets." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.801, + 0.718, + 0.817 + ], + "angle": 0, + "content": "5.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.903 + ], + "angle": 0, + "content": "We train our model on the training set of StereoNVS-Real, which has multi-view stereo image pairs of real scenes. Our model is trained for 250K iterations. For each iteration, we randomly select one scene and one target viewpoint of the scene. For both training and evaluation, image" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4944" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.244 + ], + "angle": 0, + "content": "features and feature volumes are extracted from three stereo image pairs (i.e., total six images) at three viewpoints nearest to the target viewpoint. During training, 512 rays are randomly selected for the training batch. We use the Adam optimizer [18] with learning rates of 0.0005 and the cosine annealing scheduling [22]. We employ UniMatch [35] for both the pre-trained stereo estimation network within the stereo feature extractor and the pseudo-ground-truth depth of the stereo depth loss. Refer to Sec. B in the supplementary document for additional implementation details." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.259, + 0.21, + 0.275 + ], + "angle": 0, + "content": "5.2. Comparison" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.284, + 0.47, + 0.436 + ], + "angle": 0, + "content": "We compare our method with recent generalizable novel view synthesis methods: SRF [4], IBRNet [33], GeoNeRF [16], GNT [32] and NeuRay [21]. Like our method, we use three stereo-camera image pairs to synthesize each target-view image for all the baseline methods. While the previous methods do not explicitly assume stereo-camera images as their inputs, we also train them using stereocamera images as they can handle stereo-camera images as independent inputs. All the baseline models are trained on the training set of StereoNVS-Real, as done for our method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.438, + 0.47, + 0.573 + ], + "angle": 0, + "content": "We evaluate both image and depth qualities on the StereoNVS-Real and the StereoNVS-Synthetic datasets. For StereoNVS-Real, we utilize pseudo-ground-truth depths obtained from COLMAP [28] to assess depth quality. On the other hand, for StereoNVS-Synthetic, we use rendered depth maps as ground truths for depth quality assessment. We employ PSNR, SSIM [34], and LPIPS [41] as metrics for image quality and absolute error (ABS) as a metric for depth quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.576, + 0.471, + 0.668 + ], + "angle": 0, + "content": "Among the compared methods, for training and inference, NeuRay [21] requires depth maps and GeoNeRF [16] can use depth maps as additional inputs. We denote such variation of GeoNeRF as \\(\\mathrm{GeoNeRF}_{+D}\\). For their training and inference, we used the pseudo-ground-truth depth maps estimated by UniMatch [35] as done for our method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Qualitative comparison. Fig. 7 presents a qualitative comparison on novel view synthesis between our method and previous methods [16, 21, 32, 33] using the StereoNVS dataset. Previous methods estimate inaccurate geometry for scenes with thin structures or textureless regions, resulting in severe artifacts in the synthesized novel view images. In contrast, our method clearly outperforms the baseline methods in view synthesis results with significantly fewer artifacts even in textureless regions, thanks to more accurately estimated geometry. Moreover, our method shows better synthesis results compared to \\(\\mathrm{GeoNeRF}_{+D}\\) [16] and Neu-Ray [21], even though they explicitly use depth maps in the inference time, demonstrating the robust utilization of depth maps in our framework." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.09, + 0.891, + 0.166 + ], + "angle": 0, + "content": "
StereoNVS-RealStereoNVS-Synthetic
PSNR(↑)SSIM(↑)LPIPS(↓)ABS (↓)PSNR(↑)SSIM(↑)LPIPS(↓)ABS (↓)
Baseline (a)28.050.89530.13720.559832.220.91720.13860.1679
+ Stereo setting (b)28.010.89290.14600.506432.130.91790.14380.1577
+ SAM (c)28.210.89670.13980.493531.900.91670.13860.1531
+ Correlated feature (d)28.310.89880.13700.505732.350.92450.12820.1416
+ DGPS (e)28.420.89970.14030.509833.090.93040.12300.1246
+ Stereo depth loss (f)28.440.90000.13960.486833.450.93360.12030.1056
" + }, + { + "type": "table_caption", + "bbox": [ + 0.588, + 0.168, + 0.804, + 0.181 + ], + "angle": 0, + "content": "Table 2. Quantitative ablation study." + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.184, + 0.583, + 0.198 + ], + "angle": 0, + "content": "Tab. 2 (b)" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.198, + 0.599, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.619, + 0.184, + 0.681, + 0.198 + ], + "angle": 0, + "content": "Tab. 2 (d)" + }, + { + "type": "image", + "bbox": [ + 0.6, + 0.198, + 0.696, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.715, + 0.184, + 0.777, + 0.198 + ], + "angle": 0, + "content": "Tab. 2 (e)" + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.198, + 0.793, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.813, + 0.184, + 0.871, + 0.197 + ], + "angle": 0, + "content": "Tab. 2 (f)" + }, + { + "type": "image", + "bbox": [ + 0.794, + 0.198, + 0.891, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.321, + 0.892, + 0.35 + ], + "angle": 0, + "content": "Figure 8. Qualitative ablation study. Our full method enables accurate synthesis of a novel-view image and a depth map." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.361, + 0.893, + 0.619 + ], + "angle": 0, + "content": "Quantitative comparison. Tab. 1 shows that our method generally surpasses other baseline methods [4, 16, 21, 32, 33] in terms of image and depth qualities. GNT [32] exhibits degraded performances, likely due to its data-hungry transformer backbone. While performance differences with GeoNeRF are not substantial on the StereoNVS-Real dataset, our method shows the superior perceptual quality (Fig. 7) and significantly better performance on the StereoNVS-Synthetic dataset. While \\(\\mathrm{GeoNeRF}_{+D}\\) [16] and NeuRay [21] achieve comparable performances on the StereoNVS-Synthetic dataset, they show considerably degenerated results on the StereoNVS-Real dataset. In contrast, our method demonstrates superior performance thanks to the robustness of our framework. Sec. D.1 in the supplementary document further discusses the sensitivity of \\(\\mathrm{GeoNeRF}_{+D}\\), NeuRay, and our method to depth errors in real-world images." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.628, + 0.72, + 0.643 + ], + "angle": 0, + "content": "5.3. Analysis and Discussion" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.651, + 0.661, + 0.667 + ], + "angle": 0, + "content": "5.3.1 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.81 + ], + "angle": 0, + "content": "To assess the impact of our proposed methods, we conduct an ablation study starting with our baseline model [16], which is trained on three views in a monocular setting (i.e., three images). Although equipped with a stereo camera setting, the baseline model trained on three stereo views (i.e., six images) shows similar image qualities as shown in Tab. 2 (b). This result indicates that sophisticated methods are needed to exploit the invaluable information from stereo-camera images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Tab. 2 shows that our model consistently demonstrates improvements in view synthesis results, as we introduce our proposed components. Our stereo feature extractor enables us to estimate better geometry, leading to fewer artifacts in novel view synthesis, as shown in Fig. 8 (d). The stereo attention module (c) and the stereo-correlated features (d)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4945" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "help extract robust stereo image features, which are particularly effective for real scenes and synthetic scenes, respectively. Our DGPS is essential for better geometry estimation, leading to significantly improved results as shown in Fig. 8 (e), especially in textureless regions. This is further evident in Tab. 2 (e), with considerable performance gain on the StereoNVS-Synthetic dataset. Our final model shows the best performances (Tab. 2 (f)), with high-quality depth and view synthesis results (Fig. 8 (f))." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.238, + 0.46, + 0.256 + ], + "angle": 0, + "content": "5.3.2 Effectiveness of Depth-Guided Plane-Sweeping" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.262, + 0.47, + 0.367 + ], + "angle": 0, + "content": "We conduct two experiments to demonstrate the effectiveness of our DGPS. In the first experiment, we compare our final model against a model without using DGPS. The \"Final model\" in Tab. 3 denotes our final model with DGPS. As shown in Tab. 3, the model without using DGPS shows degenerated results compared to our final model, highlighting the effectiveness of DGPS." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.368, + 0.473, + 0.611 + ], + "angle": 0, + "content": "In the second experiment, we compare our model using DGPS against a model using more depth planes for cost volume construction. The term \"Base model\" in Tab. 4 refers to our model that is solely equipped with the stereo feature extractor, excluding DGPS and the stereo depth loss. We train an additional model that uses approximately 1.5 times the number of depth planes, compared to the base model. Although this additional model utilizes more depth planes, it shows similar results to the base model on StereoNeRF-Synthetic, as reported in Tab. 4. This is because increasing the number of planes does not guarantee accurate correspondence matching, especially in textureless regions. In contrast, our DGPS guarantees correspondence matching across multi-view image features near the geometry, resulting in better image and depth qualities. Refer to Sec. C.2.1 in the supplementary document for more details." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.62, + 0.427, + 0.636 + ], + "angle": 0, + "content": "5.3.3 Benefit of Stereo Estimation in Depth Loss" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.47, + 0.78 + ], + "angle": 0, + "content": "To show the effectiveness of using stereo estimation networks for depth supervision, we conduct an additional experiment as follows. First, we obtain two pseudo-GT depths: one from the pre-trained stereo network [35] (\\(d_{gt}\\)) as stated in Sec. 3.4 and the other from the state-of-the-art learning-based MVS network (\\(d_{gt}^{mvs}\\)) [26]. Then, we train two models with our methods using \\(d_{gt}\\) and \\(d_{gt}^{mvs}\\), respectively. Then, we compare their synthesis results based on image and shape qualities." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.903 + ], + "angle": 0, + "content": "As shown in Tab. 5, our model trained with \\( d_{gt} \\) surpasses the other model using \\( d_{gt}^{mvs} \\) on StereoNVS-Synthetic. Note that the MVS network takes more images than the stereo network, seven images and two images, respectively. These results show that the stereo network provides a reliable depth signal for high-quality view synthesis thanks to its generalization ability, which came from standardized inputs of stereo-camera images and large-scale stereo datasets." + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.09, + 0.891, + 0.142 + ], + "angle": 0, + "content": "
PSNR(↑)SSIM(↑)LPIPS(↓)ABS(↓)
Final model (Tab. 2 (f))33.450.93360.12030.1056
- DGPS32.300.92400.12900.1300
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.145, + 0.892, + 0.174 + ], + "angle": 0, + "content": "Table 3. Effectiveness of our depth-guided plane-sweeping (DGPS) compared to the baseline models without using DGPS." + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.18, + 0.891, + 0.246 + ], + "angle": 0, + "content": "
PSNR(↑)SSIM(↑)LPIPS(↓)ABS(↓)
Base model (Tab. 2 (d))32.350.92450.12820.1416
+ more depth planes32.330.92470.12990.1470
+ DGPS (Tab. 2 (e))33.090.93040.12300.1246
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.25, + 0.892, + 0.279 + ], + "angle": 0, + "content": "Table 4. Efficiency of our depth-guided plane-sweeping (DGPS) compared to the baseline models using more depth planes." + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.285, + 0.891, + 0.334 + ], + "angle": 0, + "content": "
PSNR (↑)SSIM (↑)LPIPS (↓)ABS (↓)
Ours w/ MVS depth (dmvs)32.730.92770.12530.1216
Ours w/ Stereo depth (dtg)33.450.93360.12030.1056
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.337, + 0.892, + 0.366 + ], + "angle": 0, + "content": "Table 5. Effectiveness of using stereo depths as pseudo-ground truth for depth loss compared to using MVS depths." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.376, + 0.619, + 0.391 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.401, + 0.892, + 0.644 + ], + "angle": 0, + "content": "This paper proposes StereoNeRF, a novel generalizable view synthesis framework leveraging stereo-camera images for high-quality novel-view synthesis. Due to the ill-posedness, previous methods struggle with accurate geometry estimation, which leads to severe artifacts in novel-view synthesis. Since the stereo matching provides vital information for accurate geometry reconstruction, our framework incorporates the stereo matching into NeRF-based generalizable view synthesis approach. To this end, we introduce a stereo feature extractor, a depth-guided plane-sweeping, and a stereo depth loss. We also present the StereoNVS dataset, the first stereo-camera image dataset for training and evaluating novel-view synthesis. Our extensive experiments show that StereoNeRF is effective in generalizable novel-view synthesis, particularly in scenes with complex structures or textureless regions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.649, + 0.892, + 0.772 + ], + "angle": 0, + "content": "Limitations and Future Work. Our method is not free from limitations. In sparse view settings, our method produces blurry images and inaccurate geometry, issues also present in other methods. This limitation arises from the insufficient information for novel viewpoints in the sparse view settings. Our future work will involve additional geometric or generative prior, along with stereo prior, to compensate for the lack of information in such settings." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.775, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgement. We thank Woohyeok Kim and Hyeongmin Lee for their assistance in acquisition of the StereoNVS-Real dataset. This work was supported by the NRF grant (No.2023R1A2C200494611, 2022R1A6A1A03052954, RS-2023-00211658) and IITP grant (No.2019-0-01906, Artificial Intelligence Graduate School Program (POSTECH)) funded by the Korea government (MSIT) and Samsung Electronics Co., Ltd." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4946" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.143 + ], + "angle": 0, + "content": "[1] Jia-Ren Chang and Yong-Sheng Chen. Pyramid stereo matching network. In CVPR, pages 5410-5418, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.145, + 0.468, + 0.2 + ], + "angle": 0, + "content": "[2] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In ICCV, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.202, + 0.468, + 0.258 + ], + "angle": 0, + "content": "[3] Shuo Cheng, Zexiang Xu, Shilin Zhu, Zhuwen Li, Li Erran Li, Ravi Ramamoorthi, and Hao Su. Deep stereo using adaptive thin volume representation with uncertainty awareness. In CVPR, pages 2524-2534, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.26, + 0.468, + 0.313 + ], + "angle": 0, + "content": "[4] Julian Chibane, Aayush Bansal, Verica Lazova, and Gerard Pons-Moll. Stereo radiance fields (srf): Learning view synthesis for sparse views of novel scenes. In CVPR. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.316, + 0.468, + 0.357 + ], + "angle": 0, + "content": "[5] Xiaojie Chu, Liangyu Chen, and Wenqing Yu. Nafssr: Stereo image super-resolution using nafnet. In CVPR, pages 1239-1248, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.36, + 0.468, + 0.429 + ], + "angle": 0, + "content": "[6] Paul E Debevec, Camillo J Taylor, and Jitendra Malik. Modeling and rendering architecture from photographs: A hybrid geometry-and image-based approach. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 11-20, 1996. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.431, + 0.468, + 0.514 + ], + "angle": 0, + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.516, + 0.468, + 0.558 + ], + "angle": 0, + "content": "[8] Yilun Du, Cameron Smith, Ayush Tewari, and Vincent Sitzmann. Learning to render novel views from wide-baseline stereo pairs. arXiv preprint arXiv:2304.08463, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.56, + 0.468, + 0.613 + ], + "angle": 0, + "content": "[9] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, pages 5501–5510, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.616, + 0.468, + 0.698 + ], + "angle": 0, + "content": "[10] Huan Fu, Bowen Cai, Lin Gao, Ling-Xiao Zhang, Jiaming Wang, Cao Li, Qixun Zeng, Chengyue Sun, Rongfei Jia, Binqiang Zhao, et al. 3d-front: 3d furnished rooms with layouts and semantics. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10933-10942, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.701, + 0.468, + 0.743 + ], + "angle": 0, + "content": "[11] Yasutaka Furukawa, Carlos Hernández, et al. Multi-view stereo: A tutorial. Foundations and Trends® in Computer Graphics and Vision, 9(1-2):1-148, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.745, + 0.468, + 0.786 + ], + "angle": 0, + "content": "[12] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, pages 3354-3361. IEEE, 2012. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.468, + 0.843 + ], + "angle": 0, + "content": "[13] Steven J Gortler, Radek Grzeszczuk, Richard Szeliski, and Michael F Cohen. The lumigraph. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 43-54, 1996. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[14] Xiaodong Gu, Zhiwen Fan, Siyu Zhu, Zuozhuo Dai, Feitong Tan, and Ping Tan. Cascade cost volume for high-resolution multi-view stereo and stereo matching. In CVPR, pages 2495-2504, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "[15] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[16] Mohammad Mahdi Johari, Yann Lepoittevin, and François Fleuret. Geonerf: Generalizing nerf with geometry priors. In CVPR, pages 18365-18375, 2022. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.18, + 0.892, + 0.221 + ], + "angle": 0, + "content": "[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM TOG, 42(4):1-14, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.224, + 0.892, + 0.263 + ], + "angle": 0, + "content": "[18] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.267, + 0.892, + 0.307 + ], + "angle": 0, + "content": "[19] Kiriakos N Kutulakos and Steven M Seitz. A theory of shape by space carving. International journal of computer vision, 38:199-218, 2000. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.311, + 0.892, + 0.352 + ], + "angle": 0, + "content": "[20] Marc Levoy and Pat Hanrahan. Light field rendering. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 31-42, 1996. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.354, + 0.892, + 0.409 + ], + "angle": 0, + "content": "[21] Yuan Liu, Sida Peng, Lingjie Liu, Qianqian Wang, Peng Wang, Christian Theobalt, Xiaowei Zhou, and Wenping Wang. Neural rays for occlusion-aware image-based rendering. In CVPR, pages 7824-7833, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.412, + 0.892, + 0.452 + ], + "angle": 0, + "content": "[22] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.455, + 0.892, + 0.523 + ], + "angle": 0, + "content": "[23] Nikolaus Mayer, Eddy Ilg, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In CVPR, pages 4040-4048, 2016. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.526, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[24] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.584, + 0.892, + 0.625 + ], + "angle": 0, + "content": "[25] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM TOG, 41(4):1-15, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.628, + 0.892, + 0.682 + ], + "angle": 0, + "content": "[26] Rui Peng, Rongjie Wang, Zhenyu Wang, Yawen Lai, and Ronggang Wang. Rethinking depth estimation for multiview stereo: A unified representation. In CVPR, pages 8645-8654, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.686, + 0.892, + 0.726 + ], + "angle": 0, + "content": "[27] Daniel Scharstein and Richard Szeliski. A taxonomy and evaluation of dense two-frame stereo correspondence algorithms. IJCV, 47:7-42, 2002. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.729, + 0.892, + 0.769 + ], + "angle": 0, + "content": "[28] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, pages 4104-4113, 2016. 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.772, + 0.892, + 0.813 + ], + "angle": 0, + "content": "[29] Steven M Seitz and Charles R Dyer. Photorealistic scene reconstruction by voxel coloring. International Journal of Computer Vision, 35:151-173, 1999. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.816, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[30] Mohammed Suhail, Carlos Esteves, Leonid Sigal, and Ameesh Makadia. Generalizable patch-based neural rendering. In ECCV, pages 156-174. Springer, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[31] Alex Trevithick and Bo Yang. Grf: Learning a general radiance field for 3d representation and rendering. In ICCV, pages 15182-15192, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4947" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.133 + ], + "angle": 0, + "content": "[32] Mukund Varma, Peihao Wang, Xuxi Chen, Tianlong Chen, Subhashini Venugopalan, and Zhangyang Wang. Is attention all that nef needs? In ICLR, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.468, + 0.205 + ], + "angle": 0, + "content": "[33] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In CVPR, pages 4690-4699, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.207, + 0.468, + 0.247 + ], + "angle": 0, + "content": "[34] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 13(4):600-612, 2004. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.25, + 0.468, + 0.302 + ], + "angle": 0, + "content": "[35] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, Fisher Yu, Dacheng Tao, and Andreas Geiger. Unifying flow, stereo and depth estimation. IEEE TPAMI, 2023. 2, 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.306, + 0.468, + 0.346 + ], + "angle": 0, + "content": "[36] Jiayu Yang, Wei Mao, Jose M Alvarez, and Miaomiao Liu. Cost volume pyramid based depth inference for multi-view stereo. In CVPR, pages 4877-4886, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.349, + 0.468, + 0.389 + ], + "angle": 0, + "content": "[37] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoptrees for real-time rendering of neural radiance fields. In ICCV, pages 5752-5761, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.392, + 0.468, + 0.432 + ], + "angle": 0, + "content": "[38] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In CVPR, pages 4578-4587, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.435, + 0.468, + 0.475 + ], + "angle": 0, + "content": "[39] Feihu Zhang, Victor Prisacariu, Ruigang Yang, and Philip HS Torr. Ga-net: Guided aggregation net for end-to-end stereo matching. In CVPR, pages 185-194, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.478, + 0.468, + 0.517 + ], + "angle": 0, + "content": "[40] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.52, + 0.468, + 0.572 + ], + "angle": 0, + "content": "[41] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.576, + 0.468, + 0.603 + ], + "angle": 0, + "content": "[42] Zhengyou Zhang. A flexible new technique for camera calibration. IEEE TPAMI, 22(11):1330-1334, 2000. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.606, + 0.468, + 0.646 + ], + "angle": 0, + "content": "[43] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. ACM TOG, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.646 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4948" + } + ] +] \ No newline at end of file diff --git a/2024/Generalizable Novel-View Synthesis using a Stereo Camera/516ec60a-018b-4106-92f4-813c127e6028_origin.pdf b/2024/Generalizable Novel-View Synthesis using a Stereo Camera/516ec60a-018b-4106-92f4-813c127e6028_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a94b179249f38d6786dd35cb0c38632e3f585262 --- /dev/null +++ b/2024/Generalizable Novel-View Synthesis using a Stereo Camera/516ec60a-018b-4106-92f4-813c127e6028_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0cf95eb2bba3316e8ea68b431eca7ec0c8264e878c0764c5ad0649bf8c73dd31 +size 2513837 diff --git a/2024/Generalizable Novel-View Synthesis using a Stereo Camera/full.md b/2024/Generalizable Novel-View Synthesis using a Stereo Camera/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0853738c6e4f96c8ab4382dad967d3254a7ebb6a --- /dev/null +++ b/2024/Generalizable Novel-View Synthesis using a Stereo Camera/full.md @@ -0,0 +1,300 @@ +# Generalizable Novel-View Synthesis using a Stereo Camera + +Haechan Lee $^{1,*}$ + +Wonjoon Jin $^{2,*}$ + +Seung-Hwan Baek $^{1,2}$ + +Sunghyun Cho $^{1,2}$ + +# POSTECH $^{1}$ GSAI & $^{2}$ CSE + +{gocks8, jinwj1996, shwbaek, s.cho}@postech.ac.kr + +# Abstract + +In this paper, we propose the first generalizable view synthesis approach that specifically targets multi-view stereocamera images. Since recent stereo matching has demonstrated accurate geometry prediction, we introduce stereo matching into novel-view synthesis for high-quality geometry reconstruction. To this end, this paper proposes a novel framework, dubbed StereoNeRF, which integrates stereo matching into a NeRF-based generalizable view synthesis approach. StereoNeRF is equipped with three key components to effectively exploit stereo matching in novel-view synthesis: a stereo feature extractor, a depth-guided plane-sweeping, and a stereo depth loss. Moreover, we propose the StereoNVS dataset, the first multi-view dataset of stereocamera images, encompassing a wide variety of both real and synthetic scenes. Our experimental results demonstrate that StereoNeRF surpasses previous approaches in generalizable view synthesis. + +# 1. Introduction + +Novel-view synthesis is a long-standing ill-posed problem in computer vision and graphics, which is inherently challenging due to the necessity of predicting both the geometry and texture from images of a target scene. Recently, Neural Radiance Fields (NeRF) [24] have achieved photorealistic results by jointly optimizing geometry and radiance fields with a coordinate-based network. However, the need for per-scene optimization, which has to learn representations for each target scene individually, restricts its applicability, as it requires additional training time for such optimization. + +Recent approaches [2, 4, 16, 21, 30, 32, 33, 38] have addressed this issue of synthesizing novel-view images on-the-fly for unseen scenes without per-scene optimization. Early studies [4, 33, 38] utilize an image encoder to train a generic view interpolation function, enabling the estimation of NeRF parameters from unseen input images in a feed-forward manner. However, this single feed-forward manner for estimating geometry and color exacerbates the ill + +![](images/43f30b7c5765bfbaecae03499b724cdaea92243b9d92122ef04f9e8d9414441e.jpg) +(a) Baseline +Figure 1. Novel-view synthesis results of a baseline method [16] and ours. The baseline shows degraded performances, even trained using stereo-camera images (b). In contrast, fully exploiting stereo-camera images, our method shows superior results (c). + +![](images/c37eff9227541f2a927ca8ed80391be85c840f40bb690cd779b4ccaa448dfc4f.jpg) +(b) Baseline +(using stereo images) + +![](images/b505cd829b7e1639400ad26b07f9406d86fdcb5b9df89a2fa92d553cc92b96e4.jpg) +(c) Ours +(using stereo images) + +posedness, resulting in low-quality geometries and rendering results. For better geometry reasoning, MVSNeRF [2] and GeoNeRF [16] leverage the multi-view stereo (MVS) approach to handle occlusions in a 3D scene. Nonetheless, they struggle with inaccurate geometry prediction and limited synthesis accuracy, as shown in the synthesis result of GeoNeRF (Fig. 1 (a)). + +To tackle this challenge, we propose the first generalizable NeRF approach that leverages stereo-camera images, which are easily accessible thanks to the ubiquity of stereo cameras in most mobile devices. Recent advance in learning-based stereo estimation has demonstrated accurate geometry prediction, often even outperforming learning-based MVS methods as shown in Fig. 2. The superior performance of stereo estimation can be attributed to several key factors. First, unlike MVS that assumes arbitrary number of inputs with arbitrary camera positions and orientations, stereo matching assumes two stereo inputs with a fixed baseline. This constraint allows a more optimal network architecture that can effectively find dense correspondences between input images, and significantly eases the training difficulty. Moreover, larger scale of stereomatching datasets [12, 23, 27] compared to MVS datasets has facilitated remarkable generalization capabilities in stereo estimation network. Therefore, we aim to harness this accurate geometric information from stereo images to alleviate the ill-posedness of generalizable view synthesis. + +![](images/a5d0501def5d90ee2d3f6acd8e1baec8277930a78ebfea87967b841bc45298d6.jpg) +Figure 2. Usefulness of exploiting binocular stereo. Comparison on depth estimation between a learning-based MVS method [26] and a learning-based binocular stereo method [35]. + +However, since previous methods do not explicitly consider stereoscopic inputs, they cannot leverage the aforementioned benefits, resulting in degenerated performances shown in Fig. 1 (b) despite using stereo-camera images. + +This paper proposes StereoNeRF, a novel generalizable view synthesis framework leveraging stereo images. StereoNeRF integrates stereo matching into NeRF-based generalizable view synthesis approach, where the stereo matching provides vital geometric information. To this end, we first introduce a stereo feature extractor, which extracts geometry-aware features by correlating horizontal epipolar lines within stereo images. In addition, the stereo feature extractor takes advantage of stereo-correlated features from an off-the-shelf stereo estimation network, which can transfer rich geometric knowledge to our model. Furthermore, with the reliable depth estimated from the stereo estimation network, we aggregate multi-view features through a depth-guided plane-sweeping technique. This technique ensures correspondence matching around the geometry in cost volume construction. We also present a stereo depth loss utilizing the estimated stereo depth. These additional geometric cues from the stereo matching effectively mitigate the ill-posedness in generalizable view synthesis. Notably, our framework leveraging stereo images surpasses the previous approaches [16, 21], which rely on extra depth information. + +Furthermore, we propose the StereoNVS dataset, which is the first dataset for training and evaluation of novel-view synthesis using stereo-camera images. Our StereoNVS dataset provides real-world and synthetic stereo images. Our extensive evaluation on the StereoNVS dataset shows that stereo-pair inputs can effectively enhance the quality of novel-view synthesis, and shows that StereoNeRF outperforms previous generalizable novel-view synthesis approaches in terms of image and shape qualities. + +Our contributions are as follows: + +- We propose a generalized NeRF approach that leverages stereo-camera images for the first time to alleviate the ill-posedness in generalizable novel-view synthesis. +- We propose a novel framework, StereoNeRF, which exploits the benefits of stereo images by integrating stereo matching into generalizable view synthesis. To this end, we present a stereo feature extractor, a depth-guided plane-sweeping, and a stereo depth loss in our framework. +- We also present the StereoNVS dataset, the first dataset for training and evaluation of novel-view synthesis obtained by stereo cameras. + +# 2. Related Work + +# 2.1. Novel-View Synthesis + +Novel-view synthesis aims to synthesize target-view images from reference-view images. Early approaches synthesize novel-view images by blending pixels from multiple input images [6, 13, 20]. Recent work adopting neural volume representations has shown remarkable novel-view synthesis results. Zhou et al. [43] propose multi-plane images (MPI) representation estimated from input images, but their methods produce valid novel-view images only for narrow ranges of camera poses. Mildenhall et al. [24] propose Neural Radiance Fields (NeRF) that can synthesize photo-realistic target-view images via neural implicit representations and volume rendering. Albeit its photo-realism, computation-heavy per-scene optimization is needed. Recent variants such as [9, 17, 25, 37] have remarkably reduced the optimization time, but they still require large memory and several minutes for training. + +Generalizable View Synthesis. Recently, many novel-view synthesis approaches without per-scene optimization have been proposed. Several studies directly predict pixel colors by aggregating image features from aligned pixels, without 3D representations. Suhail et al. [30] and Varma et al. [32] adopt a transformer-based network [7] to compute features along the epipolar lines, which needs a large number of training images for high-quality view synthesis. Du et al. [8] propose a framework to synthesize target views from two images with small overlapped regions. + +Another research direction predicts volumetric representation [24] from aggregated features from reference views and synthesizes images via volumetric rendering. PixelNeRF, SRF, and GRF [4, 31, 38] predict radiance fields from pixel-aligned features using an MLP. Among these methods, SRF [4] uses two sampled images from an image collection captured by a monocular camera as a stereo pair, but shows limited synthesis quality. IBRNet [33] proposes to learn generic view interpolation functions, but suffers from artifacts for challenging scenes with complex geometries. MVSNeRF, GeoNeRF and NeuRay [2, 16, 21] utilize the MVS approach using cost volume for better occlusion handling in generalizable view synthesis. + +However, all the aforementioned methods often fail to capture accurate geometry, particularly in textureless regions, leading to limited synthesis results. Unlike these methods, our framework exploits stereo-camera images with a fixed baseline to effectively capture the geometry of complex scenes as well as textureless regions for high-quality novel-view synthesis. + +# 2.2. Stereo Matching and Multi-View Stereo + +Geometry estimation is a long-standing problem in computer vision that has a variety of applications. Among them, + +stereo matching is a task that takes rectified stereo images and computes stereo correspondence to estimate disparities [15, 27]. Recently, a huge number of data-driven approaches have been introduced with the emergence of a vast amount of stereo-matching datasets [12, 23, 27] and made significant progress [1, 23, 35, 39]. MVS approaches that use more than two views have been extensively studied as well [11, 19, 29]. Recently, learning-based MVS approaches have been proposed, e.g., Cheng et al. [3], Gu et al. [14], and Yang et al. [36] present efficient frameworks that cascade cost volumes in a coarse-to-fine manner to enable high-resolution depth estimation. + +In our work, we have the best of both worlds in our generalizable view synthesis framework. We adopt the MVS approach to aggregate multi-view information for occlusion-aware geometry estimation, as done in GeoNerveRF [16]. Furthermore, our framework also assumes structured stereo images as inputs, and integrates the two-view stereo matching into our framework for robust and accurate geometry estimation in generalizable novel-view synthesis. + +# 3. StereoNeRF + +For novel-view synthesis, our framework takes a set of rectified stereo-camera images and estimates neural radiance fields [24] from which novel-view images are rendered. StereoNeRF integrates a well-designed stereo-matching algorithm into the existing generalizable view synthesis approach [16]. In this section, we first provide an overall pipeline that utilizes an pre-trained stereo estimation network [35]. Then, we explain each step of our framework and the training process in detail, highlighting how to integrate stereo matching into our framework. + +Overall pipeline. Fig. 3 shows an overview of our pipeline. Our framework utilizes $N$ pairs of stereo images of a target scene $\{(I_L^n,I_R^n)\}_{n = 1}^N$ for novel-view synthesis. In the first step, the pre-trained stereo estimation network takes the $n$ -th image pair $(I_L^n,I_R^n)$ , and outputs stereo depths $(d_{s,L}^{n},d_{s,R}^{n})$ and stereo-correlated features $(t_L^n,t_R^n)$ . In the second step, a stereo feature extractor takes the stereo image pair and the stereo-correlated features, and outputs stereo image features $(f_L^n,f_R^n)$ , explained in Sec. 3.1. The third step aggregates stereo image features from all viewpoints $\{(f_L^n,f_R^n)\}_{n = 1}^N$ to build 3D feature volumes $\{(\phi_L^n,\phi_R^n)\}_{n = 1}^N$ using an MVS network (Sec. 3.2). For constructing feature volumes that faithfully reflect the 3D geometry of a target scene, the third step adopts a depth-guided plane-sweeping which utilizes the estimated stereo depths $(d_{s,L}^{n},d_{s,R}^{n})$ . Finally, a neural renderer predicts radiance fields from the stereo image features $\{(f_L^n,f_R^n)\}_{n = 1}^N$ and the feature volumes $\{(\phi_L^n,\phi_R^n)\}_{n = 1}^N$ from all viewpoints, and novel-view images are synthesized from the radiance fields (Sec. 3.3). + +![](images/a02480fe535e1d999f71da61d0d90036ebf182ec44e92387dbfb611677a36928.jpg) +Figure 3. Overview of StereoNeRF. StereoNeRF consists of a shared feature extractor and a neural renderer. We design the feature extractor with a stereo estimation network, a stereo feature extractor, and a MVS network. StereoNeRF takes $N$ pairs of stereo images and a camera pose as inputs, and synthesizes a novel-view image of the camera pose. + +# 3.1. Feature Extraction from Stereo Image Pairs + +Unlike previous approaches, which extract features of input images respectively, our stereo feature extractor takes rectified stereo-camera images and computes image feature maps by exploiting the epipolar geometry between them. To this end, our stereo feature extractor consists of three parts: CNN encoders, stereo attention modules (SAM) [5], and CNN decoders (Fig. 4). Moreover, we propose integrating the stereo-correlated features $(t_L^n,t_R^n)$ from the pretrained stereo estimation network into the SAM (green line in Fig. 4). This integration inherits geometric cues from the stereo estimation network. In the following, we provide the detailed description of the stereo feature extractor. + +First, the weight-shared CNN encoders project each of $I_L^n$ and $I_R^n$ into the feature space. Then, we adopt the stereo attention module (SAM) [5] to fuse two feature maps on the horizontal epipolar lines. To this end, SAM estimates stereo correspondences between two feature maps and explicitly adds stereo-correspondent features, as shown in Fig. 5. In addition, we extend the SAM by aggregating the stereocorrelated features $(t_L^n, t_R^n)$ in this feature fusion process. Note that the stereo-correlated features are computed from the pre-trained stereo estimation network, which provides vital geometric cues. These stereo-correlated features enhance stereo correspondences between two feature maps, leading to higher-quality fused feature maps. + +Specific explanation of the aforementioned feature fusion of the SAM is as follows (Fig. 5). For the left feature, a warping matrix $\mathcal{W}_{R\to L}$ is built by multiplying a query matrix and a key matrix followed by a softmax layer. The query matrix is computed from the left feature and the stereocorrelated features, and the key matrix is computed from the right feature and the stereo-correlated features. Then, a value matrix, computed from the right feature and the stereo-correlated feature, is warped along the epipolar lines by multiplying the warping matrix $\mathcal{W}_{R\to L}$ , then added to + +![](images/85b68ed08f4f2f565d13c77166d36d7ac23cb54574c5916cbfd37d6202094e17.jpg) +Figure 4. Stereo feature extractor of StereoNeRF, where an pretrained stereo estimation network is incorporated. + +the left feature. The same feature fusion process is also performed for the right feature. + +For effective fusion of information from both stereo images and stereo-correlated features, the stereo feature extractor has three sequentially stacked SAM. Finally, the weight-shared CNN decoders take each of the fused features from the SAM, and generate stereo image features $(f_{L}^{n}, f_{R}^{n})$ , which will be used for building feature volumes. + +# 3.2. Depth-guided Feature Volume Construction + +Once stereo image features are obtained, we aggregate these features from all viewpoints to create cost volumes. To this end, we adopt a plane-sweeping-based approach, which computes multi-view correspondences among these features. The plane-sweeping-based approach first defines a depth range within pre-computed near and far depths obtained from COLMAP [28]. Then, it hypothesizes depth planes in the entire depth range. On these depth planes, multi-view stereo image features are aggregated via plane-sweeping to build cost volumes. Then, these cost volumes are processed to yield feature volumes, which will be used later in predicting neural radiance fields. + +However, this plane-sweeping-based approach often struggles with accurate geometry estimation. Since this approach searches the entire depth range of a scene, correspondence matching is inaccurate and prone to error, espe + +![](images/bd3f7efda17ad332653d31cc2a8a05a2e9740ed4281c10f946bbc2e77355d969.jpg) +Figure 5. Stereo attention module used in the stereo feature extractor. We exploit the rich features from the pre-trained stereo estimation module. + +![](images/6b75aa4878ea606421023cecf99e7405733a829548f84266859484cf17cc1e76.jpg) +Figure 6. Feature volume construction using DGPS. + +cially in textureless regions. To tackle this issue, we introduce a depth-guided plane-sweeping (DGPS), which constructs cost volumes around the stereo depth predicted from the stereo estimation network. Unlike the previous approach using the entire depth range, DGPS introduces a dynamic search range that varies based on the estimated stereo depth. This technique significantly constrains the search space and reduces outliers in correspondence matching. + +Fig. 6 depicts the feature volume construction using DGPS. Instead of searching the entire depth range, we hypothesize depth planes around the stereo depths $(d_{s,L}^n,d_{s,R}^n)$ . Then, through DGPS, we aggregate the stereo image features from all viewpoints $\{(f_L^n,f_R^n)\}_{n = 1}^N$ on these depth planes to build cost volumes. These cost volumes are further processed via the MVS network to obtain feature volumes $(\phi_L^n,\phi_R^n)$ and depth maps $(d_{s,L}^n,d_{s,R}^n)$ . We repeat this process to obtain feature volumes and depth maps for every viewpoint. The resulting multi-view feature volumes and depth maps are denoted as $\{(\phi_L^L,\phi_R^L)\}_{n = 1}^N$ and $\{(d_{m,L}^{n},d_{m,R}^{n})\}_{n = 1}^{N}$ , respectively. Refer to Sec. B.3 in the supplementary document for more details about our feature volume construction. + +# 3.3. Rendering Novel Views + +Once feature volumes are constructed, novel-view images can be rendered via neural rendering. For rendering novel-view images, our framework adopts the rendering procedure of GeoNeRF [16]. In the following, we briefly describe the rendering procedure. Given a target viewpoint, we cast a ray for each pixel and sample points along the ray in the 3D space. Then, for each sampled point, we sample features from all the feature volumes $\{(\phi_L^n,\phi_R^n)\}_{n = 1}^N$ . We also sample features from all the image features $\{(f_L^n,f_R^n)\}_{n = 1}^N$ by projecting the sampled point onto input images. The sampled volume and image features are aggregated by a neural renderer network, and the color and density are estimated. Finally, by integrating the estimated color values and depth values of sampled points with their densities along each ray, we obtain the color $\hat{c}$ and the depth $d_{r}$ at each pixel of a novel view, respectively. + +# 3.4. Training StereoNeRF + +We train StereoNeRF with ground-truth images and pseudoground-truth depths for high-quality novel-view synthesis. The pseudo-ground-truth depths $d_{gt}$ are obtained from the pre-trained stereo estimation network [35]. Our training loss $\mathcal{L}$ is defined as follows: + +$$ +\mathcal {L} = \mathcal {L} _ {c} + \mathcal {L} _ {d}, \tag {1} +$$ + +where $\mathcal{L}_c$ is a color loss and $\mathcal{L}_d$ is a depth loss. The color loss is defined as a mean-squared-error (MSE) between the rendered colors $\hat{c}$ and ground-truth colors [24]. + +To alleviate the shape-radiance ambiguity [40] in scene reconstruction using neural radiance fields, we introduce the depth loss $\mathcal{L}_d$ , which is defined as: + +$$ +\mathcal {L} _ {d} = \lambda_ {d} ^ {s e l f} \mathcal {L} _ {d} ^ {s e l f} + \lambda_ {d} ^ {s t e r e o} \mathcal {L} _ {d} ^ {s t e r e o}, \tag {2} +$$ + +where $\mathcal{L}_d^{self}$ is a self-supervised depth loss from GeoN-NeRF [16] and $\mathcal{L}_d^{stereo}$ is our proposed stereo depth loss. $\mathcal{L}_d^{self}$ penalizes the depths estimated from both the stereo estimation network $(d_s)$ and the MVS network $(d_m)$ by comparing them to the depth rendered from NeRF $(d_r)$ . + +Our framework estimates depths from the stereo estimation network $(d_{s})$ , the MVS network $(d_{m})$ , and the neural renderer network $(d_{r})$ . The stereo depth loss $\mathcal{L}_d^{stereo}$ guides these networks to predict more accurate depths using the pseudo-ground-truth depths $d_{gt}$ . This loss introduces a further quality improvement, especially in geometric details as will be shown in Sec. 5.3. $\mathcal{L}_d^{stereo}$ is defined as: + +$$ +\mathcal {L} _ {d} ^ {\text {s t e r e o}} = \lambda_ {1} \mathcal {L} _ {d} ^ {s} + \lambda_ {2} \mathcal {L} _ {d} ^ {m} + \lambda_ {3} \mathcal {L} _ {d} ^ {r}, \tag {3} +$$ + +where $\mathcal{L}_d^s$ , $\mathcal{L}_d^m$ , and $\mathcal{L}_d^r$ penalizes depths $d_s$ , $d_m$ , and $d_r$ , respectively, by comparing them with $d_{gt}$ . Note that $d_{gt}$ is different from $d_s$ . We obtain $d_{gt}$ from the pre-trained stereo + +estimation network with frozen parameters. On the other hand, we obtain $d_{s}$ from the stereo estimation network in our framework, which has trainable parameters. Refer to Sec. B.4 in the supplementary document for more details about our stereo depth loss. + +Due to the different characteristics of datasets such as baseline length, $d_{s}$ may have estimation error, which leads to an error in the depth-guided plane-sweeping. To tackle this, we partially train the matching and propagation network in the stereo estimation network (Fig. 4) using $\mathcal{L}_d^{self}$ . $\mathcal{L}_d^{self}$ provides multi-view supervision to the stereo estimation network by leveraging $d_{r}$ estimated from multi-view images. However, since $d_{r}$ may also have estimation error, we further regularize the stereo estimation network using $\mathcal{L}_d^s$ . This training scheme for the stereo estimation network ensures consistent depth estimation, while preserving the stereo matching capability. Sec. D.1 in the supplementary document further discusses the training scheme of the stereo estimation network. + +# 4. StereoNVS Dataset + +We propose the StereoNVS dataset, the first stereo-camera image dataset for training and evaluating novel-view synthesis using stereo-camera images. The StereoNVS dataset provides both real and synthetic datasets, each of which is dubbed StereoNVS-Real and StereoNVS-Synthetic. In the following, we present the details of each dataset. + +# 4.1. StereoNVS-Real + +StereoNVS-Real provides real-world stereo-camera images for the training and evaluation of novel-view synthesis. The dataset provides stereo-camera images of 53 static scenes, and around 25 stereo-image pairs per scene. The images are undistorted and stereo-rectified, and have a resolution of $1792 \times 896$ . The camera parameters such as the camera poses are provided as well. In our experiments, we divide the dataset into 45 and 8 scenes as training and test sets. + +To capture stereo images, we built a camera rig with two Basler machine vision cameras. We measured the camera parameters including the intrinsic and distortion parameters by camera calibration using multi-view images of a checkerboard [42]. Then, we collected stereo images from multiple viewpoints for various indoor and bounded scenes using our camera system. The captured images were then undistorted and stereo-rectified. Finally, we obtained the camera poses of the captured images using COLMAP [28]. + +# 4.2. StereoNVS-Synthetic + +For quantitative evaluation of synthesized geometries, we also present the StereoNVS-Synthetic dataset, constructed by rendering synthetic 3D models using the 3D-Front dataset [10]. The images have a resolution of $864 \times 448$ , + +![](images/c1e53e4bba21ebda7abbf50d1f1c699b97a7f022312bf20229397601f804d6fc.jpg) +Figure 7. Qualitative comparison of novel-view synthesis on the StereoNVS dataset, showing rendering results for two real-world scenes (above) and one synthetic scene (below). All models are trained using stereo images. Our method outperforms the baseline methods [16, 21, 32, 33] on both real-world and synthetic scenes, especially in thin structures and textureless regions. + +which is half of the real dataset. StereoNVS-Synthetic provides stereo-camera images of 50 scenes, and around 150 stereo image pairs per scene as well as the ground-truth camera parameters, camera poses, and depth maps. For more details about the StereoNVS dataset, refer to Sec. E in the supplementary document. + +# 5. Experiments + +We conduct extensive validation of our method on the StereoNVS dataset. In the following, we will provide implementation details of our method (Sec. 5.1), compare our method with other baselines (Sec. 5.2), and conduct a comprehensive analysis of our proposed components, which proves the effectiveness of our framework including the stereo feature extractor, the depth-guided plane-sweeping, and the stereo depth loss (Sec. 5.3). + +
MethodStereoNVS-RealStereoNVS-Synthetic
PSNR(↑)SSIM(↑)LPIPS(↓)ABS(↓)PSNR(↑)SSIM(↑)LPIPS(↓)ABS(↓)
SRF21.120.69330.41642.917622.360.71620.42450.8125
IBRNet26.120.84210.20950.756630.830.88890.18230.2628
GeoNeRF28.010.89290.14600.506432.130.91790.14380.1577
GNT26.080.84340.22851.095926.170.84060.26500.4512
GeoNeRF+D25.650.81720.20820.869332.850.93210.11710.0782
NeuRay26.510.85380.18870.614732.260.91040.14780.1571
Ours28.440.90000.13960.486833.450.93360.12030.1056
+ +Table 1. Quantitative comparison between the baseline methods [4, 16, 21, 32, 33] and ours. All models are trained using stereo images. Our method shows superior performance on both StereoNVS-Real and StereoNVS-Synthetic datasets. + +# 5.1. Implementation Details + +We train our model on the training set of StereoNVS-Real, which has multi-view stereo image pairs of real scenes. Our model is trained for 250K iterations. For each iteration, we randomly select one scene and one target viewpoint of the scene. For both training and evaluation, image + +features and feature volumes are extracted from three stereo image pairs (i.e., total six images) at three viewpoints nearest to the target viewpoint. During training, 512 rays are randomly selected for the training batch. We use the Adam optimizer [18] with learning rates of 0.0005 and the cosine annealing scheduling [22]. We employ UniMatch [35] for both the pre-trained stereo estimation network within the stereo feature extractor and the pseudo-ground-truth depth of the stereo depth loss. Refer to Sec. B in the supplementary document for additional implementation details. + +# 5.2. Comparison + +We compare our method with recent generalizable novel view synthesis methods: SRF [4], IBRNet [33], GeoNeRF [16], GNT [32] and NeuRay [21]. Like our method, we use three stereo-camera image pairs to synthesize each target-view image for all the baseline methods. While the previous methods do not explicitly assume stereo-camera images as their inputs, we also train them using stereocamera images as they can handle stereo-camera images as independent inputs. All the baseline models are trained on the training set of StereoNVS-Real, as done for our method. + +We evaluate both image and depth qualities on the StereoNVS-Real and the StereoNVS-Synthetic datasets. For StereoNVS-Real, we utilize pseudo-ground-truth depths obtained from COLMAP [28] to assess depth quality. On the other hand, for StereoNVS-Synthetic, we use rendered depth maps as ground truths for depth quality assessment. We employ PSNR, SSIM [34], and LPIPS [41] as metrics for image quality and absolute error (ABS) as a metric for depth quality. + +Among the compared methods, for training and inference, NeuRay [21] requires depth maps and GeoNeRF [16] can use depth maps as additional inputs. We denote such variation of GeoNeRF as $\mathrm{GeoNeRF}_{+D}$ . For their training and inference, we used the pseudo-ground-truth depth maps estimated by UniMatch [35] as done for our method. + +Qualitative comparison. Fig. 7 presents a qualitative comparison on novel view synthesis between our method and previous methods [16, 21, 32, 33] using the StereoNVS dataset. Previous methods estimate inaccurate geometry for scenes with thin structures or textureless regions, resulting in severe artifacts in the synthesized novel view images. In contrast, our method clearly outperforms the baseline methods in view synthesis results with significantly fewer artifacts even in textureless regions, thanks to more accurately estimated geometry. Moreover, our method shows better synthesis results compared to $\mathrm{GeoNeRF}_{+D}$ [16] and Neu-Ray [21], even though they explicitly use depth maps in the inference time, demonstrating the robust utilization of depth maps in our framework. + +
StereoNVS-RealStereoNVS-Synthetic
PSNR(↑)SSIM(↑)LPIPS(↓)ABS (↓)PSNR(↑)SSIM(↑)LPIPS(↓)ABS (↓)
Baseline (a)28.050.89530.13720.559832.220.91720.13860.1679
+ Stereo setting (b)28.010.89290.14600.506432.130.91790.14380.1577
+ SAM (c)28.210.89670.13980.493531.900.91670.13860.1531
+ Correlated feature (d)28.310.89880.13700.505732.350.92450.12820.1416
+ DGPS (e)28.420.89970.14030.509833.090.93040.12300.1246
+ Stereo depth loss (f)28.440.90000.13960.486833.450.93360.12030.1056
+ +Table 2. Quantitative ablation study. + +![](images/d74b88b8db2a0db5d67717e6819349684bd5e3eb740e0cdb82156df526d29f4e.jpg) +Tab. 2 (b) +Figure 8. Qualitative ablation study. Our full method enables accurate synthesis of a novel-view image and a depth map. + +![](images/78f95b3819d8497d6f718295c0647b5e8e96701275ae00ad2ef409f567e776d7.jpg) +Tab. 2 (d) + +![](images/2c15cedb309dbd6eff3912afe3f3d1e5c2cf8f2eecf67b8367c4dc6d4e975879.jpg) +Tab. 2 (e) + +![](images/7f90db9a239d98b1ea3ec45d42e90b4b8594594817eb1c000e8ec7bf7cfd8c3b.jpg) +Tab. 2 (f) + +Quantitative comparison. Tab. 1 shows that our method generally surpasses other baseline methods [4, 16, 21, 32, 33] in terms of image and depth qualities. GNT [32] exhibits degraded performances, likely due to its data-hungry transformer backbone. While performance differences with GeoNeRF are not substantial on the StereoNVS-Real dataset, our method shows the superior perceptual quality (Fig. 7) and significantly better performance on the StereoNVS-Synthetic dataset. While $\mathrm{GeoNeRF}_{+D}$ [16] and NeuRay [21] achieve comparable performances on the StereoNVS-Synthetic dataset, they show considerably degenerated results on the StereoNVS-Real dataset. In contrast, our method demonstrates superior performance thanks to the robustness of our framework. Sec. D.1 in the supplementary document further discusses the sensitivity of $\mathrm{GeoNeRF}_{+D}$ , NeuRay, and our method to depth errors in real-world images. + +# 5.3. Analysis and Discussion + +# 5.3.1 Ablation Study + +To assess the impact of our proposed methods, we conduct an ablation study starting with our baseline model [16], which is trained on three views in a monocular setting (i.e., three images). Although equipped with a stereo camera setting, the baseline model trained on three stereo views (i.e., six images) shows similar image qualities as shown in Tab. 2 (b). This result indicates that sophisticated methods are needed to exploit the invaluable information from stereo-camera images. + +Tab. 2 shows that our model consistently demonstrates improvements in view synthesis results, as we introduce our proposed components. Our stereo feature extractor enables us to estimate better geometry, leading to fewer artifacts in novel view synthesis, as shown in Fig. 8 (d). The stereo attention module (c) and the stereo-correlated features (d) + +help extract robust stereo image features, which are particularly effective for real scenes and synthetic scenes, respectively. Our DGPS is essential for better geometry estimation, leading to significantly improved results as shown in Fig. 8 (e), especially in textureless regions. This is further evident in Tab. 2 (e), with considerable performance gain on the StereoNVS-Synthetic dataset. Our final model shows the best performances (Tab. 2 (f)), with high-quality depth and view synthesis results (Fig. 8 (f)). + +# 5.3.2 Effectiveness of Depth-Guided Plane-Sweeping + +We conduct two experiments to demonstrate the effectiveness of our DGPS. In the first experiment, we compare our final model against a model without using DGPS. The "Final model" in Tab. 3 denotes our final model with DGPS. As shown in Tab. 3, the model without using DGPS shows degenerated results compared to our final model, highlighting the effectiveness of DGPS. + +In the second experiment, we compare our model using DGPS against a model using more depth planes for cost volume construction. The term "Base model" in Tab. 4 refers to our model that is solely equipped with the stereo feature extractor, excluding DGPS and the stereo depth loss. We train an additional model that uses approximately 1.5 times the number of depth planes, compared to the base model. Although this additional model utilizes more depth planes, it shows similar results to the base model on StereoNeRF-Synthetic, as reported in Tab. 4. This is because increasing the number of planes does not guarantee accurate correspondence matching, especially in textureless regions. In contrast, our DGPS guarantees correspondence matching across multi-view image features near the geometry, resulting in better image and depth qualities. Refer to Sec. C.2.1 in the supplementary document for more details. + +# 5.3.3 Benefit of Stereo Estimation in Depth Loss + +To show the effectiveness of using stereo estimation networks for depth supervision, we conduct an additional experiment as follows. First, we obtain two pseudo-GT depths: one from the pre-trained stereo network [35] ( $d_{gt}$ ) as stated in Sec. 3.4 and the other from the state-of-the-art learning-based MVS network ( $d_{gt}^{mvs}$ ) [26]. Then, we train two models with our methods using $d_{gt}$ and $d_{gt}^{mvs}$ , respectively. Then, we compare their synthesis results based on image and shape qualities. + +As shown in Tab. 5, our model trained with $d_{gt}$ surpasses the other model using $d_{gt}^{mvs}$ on StereoNVS-Synthetic. Note that the MVS network takes more images than the stereo network, seven images and two images, respectively. These results show that the stereo network provides a reliable depth signal for high-quality view synthesis thanks to its generalization ability, which came from standardized inputs of stereo-camera images and large-scale stereo datasets. + +
PSNR(↑)SSIM(↑)LPIPS(↓)ABS(↓)
Final model (Tab. 2 (f))33.450.93360.12030.1056
- DGPS32.300.92400.12900.1300
+ +Table 3. Effectiveness of our depth-guided plane-sweeping (DGPS) compared to the baseline models without using DGPS. + +
PSNR(↑)SSIM(↑)LPIPS(↓)ABS(↓)
Base model (Tab. 2 (d))32.350.92450.12820.1416
+ more depth planes32.330.92470.12990.1470
+ DGPS (Tab. 2 (e))33.090.93040.12300.1246
+ +Table 4. Efficiency of our depth-guided plane-sweeping (DGPS) compared to the baseline models using more depth planes. + +
PSNR (↑)SSIM (↑)LPIPS (↓)ABS (↓)
Ours w/ MVS depth (dmvs)32.730.92770.12530.1216
Ours w/ Stereo depth (dtg)33.450.93360.12030.1056
+ +Table 5. Effectiveness of using stereo depths as pseudo-ground truth for depth loss compared to using MVS depths. + +# 6. Conclusion + +This paper proposes StereoNeRF, a novel generalizable view synthesis framework leveraging stereo-camera images for high-quality novel-view synthesis. Due to the ill-posedness, previous methods struggle with accurate geometry estimation, which leads to severe artifacts in novel-view synthesis. Since the stereo matching provides vital information for accurate geometry reconstruction, our framework incorporates the stereo matching into NeRF-based generalizable view synthesis approach. To this end, we introduce a stereo feature extractor, a depth-guided plane-sweeping, and a stereo depth loss. We also present the StereoNVS dataset, the first stereo-camera image dataset for training and evaluating novel-view synthesis. Our extensive experiments show that StereoNeRF is effective in generalizable novel-view synthesis, particularly in scenes with complex structures or textureless regions. + +Limitations and Future Work. Our method is not free from limitations. In sparse view settings, our method produces blurry images and inaccurate geometry, issues also present in other methods. This limitation arises from the insufficient information for novel viewpoints in the sparse view settings. Our future work will involve additional geometric or generative prior, along with stereo prior, to compensate for the lack of information in such settings. + +Acknowledgement. We thank Woohyeok Kim and Hyeongmin Lee for their assistance in acquisition of the StereoNVS-Real dataset. This work was supported by the NRF grant (No.2023R1A2C200494611, 2022R1A6A1A03052954, RS-2023-00211658) and IITP grant (No.2019-0-01906, Artificial Intelligence Graduate School Program (POSTECH)) funded by the Korea government (MSIT) and Samsung Electronics Co., Ltd. + +# References + +[1] Jia-Ren Chang and Yong-Sheng Chen. Pyramid stereo matching network. In CVPR, pages 5410-5418, 2018. 3 +[2] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In ICCV, 2021. 1, 2 +[3] Shuo Cheng, Zexiang Xu, Shilin Zhu, Zhuwen Li, Li Erran Li, Ravi Ramamoorthi, and Hao Su. Deep stereo using adaptive thin volume representation with uncertainty awareness. In CVPR, pages 2524-2534, 2020. 3 +[4] Julian Chibane, Aayush Bansal, Verica Lazova, and Gerard Pons-Moll. Stereo radiance fields (srf): Learning view synthesis for sparse views of novel scenes. In CVPR. 1, 2, 6, 7 +[5] Xiaojie Chu, Liangyu Chen, and Wenqing Yu. Nafssr: Stereo image super-resolution using nafnet. In CVPR, pages 1239-1248, 2022. 3 +[6] Paul E Debevec, Camillo J Taylor, and Jitendra Malik. Modeling and rendering architecture from photographs: A hybrid geometry-and image-based approach. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 11-20, 1996. 2 +[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2 +[8] Yilun Du, Cameron Smith, Ayush Tewari, and Vincent Sitzmann. Learning to render novel views from wide-baseline stereo pairs. arXiv preprint arXiv:2304.08463, 2023. 2 +[9] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, pages 5501–5510, 2022. 2 +[10] Huan Fu, Bowen Cai, Lin Gao, Ling-Xiao Zhang, Jiaming Wang, Cao Li, Qixun Zeng, Chengyue Sun, Rongfei Jia, Binqiang Zhao, et al. 3d-front: 3d furnished rooms with layouts and semantics. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10933-10942, 2021. 5 +[11] Yasutaka Furukawa, Carlos Hernández, et al. Multi-view stereo: A tutorial. Foundations and Trends® in Computer Graphics and Vision, 9(1-2):1-148, 2015. 3 +[12] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, pages 3354-3361. IEEE, 2012. 1, 3 +[13] Steven J Gortler, Radek Grzeszczuk, Richard Szeliski, and Michael F Cohen. The lumigraph. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 43-54, 1996. 2 +[14] Xiaodong Gu, Zhiwen Fan, Siyu Zhu, Zuozhuo Dai, Feitong Tan, and Ping Tan. Cascade cost volume for high-resolution multi-view stereo and stereo matching. In CVPR, pages 2495-2504, 2020. 3 + +[15] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 3 +[16] Mohammad Mahdi Johari, Yann Lepoittevin, and François Fleuret. Geonerf: Generalizing nerf with geometry priors. In CVPR, pages 18365-18375, 2022. 1, 2, 3, 5, 6, 7 +[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM TOG, 42(4):1-14, 2023. 2 +[18] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 7 +[19] Kiriakos N Kutulakos and Steven M Seitz. A theory of shape by space carving. International journal of computer vision, 38:199-218, 2000. 3 +[20] Marc Levoy and Pat Hanrahan. Light field rendering. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 31-42, 1996. 2 +[21] Yuan Liu, Sida Peng, Lingjie Liu, Qianqian Wang, Peng Wang, Christian Theobalt, Xiaowei Zhou, and Wenping Wang. Neural rays for occlusion-aware image-based rendering. In CVPR, pages 7824-7833, 2022. 1, 2, 6, 7 +[22] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 7 +[23] Nikolaus Mayer, Eddy Ilg, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In CVPR, pages 4040-4048, 2016. 1, 3 +[24] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2, 3, 5 +[25] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM TOG, 41(4):1-15, 2022. 2 +[26] Rui Peng, Rongjie Wang, Zhenyu Wang, Yawen Lai, and Ronggang Wang. Rethinking depth estimation for multiview stereo: A unified representation. In CVPR, pages 8645-8654, 2022. 2, 8 +[27] Daniel Scharstein and Richard Szeliski. A taxonomy and evaluation of dense two-frame stereo correspondence algorithms. IJCV, 47:7-42, 2002. 1, 3 +[28] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, pages 4104-4113, 2016. 4, 5, 7 +[29] Steven M Seitz and Charles R Dyer. Photorealistic scene reconstruction by voxel coloring. International Journal of Computer Vision, 35:151-173, 1999. 3 +[30] Mohammed Suhail, Carlos Esteves, Leonid Sigal, and Ameesh Makadia. Generalizable patch-based neural rendering. In ECCV, pages 156-174. Springer, 2022. 1, 2 +[31] Alex Trevithick and Bo Yang. Grf: Learning a general radiance field for 3d representation and rendering. In ICCV, pages 15182-15192, 2021. 2 + +[32] Mukund Varma, Peihao Wang, Xuxi Chen, Tianlong Chen, Subhashini Venugopalan, and Zhangyang Wang. Is attention all that nef needs? In ICLR, 2022. 1, 2, 6, 7 +[33] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In CVPR, pages 4690-4699, 2021. 1, 2, 6, 7 +[34] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 13(4):600-612, 2004. 7 +[35] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, Fisher Yu, Dacheng Tao, and Andreas Geiger. Unifying flow, stereo and depth estimation. IEEE TPAMI, 2023. 2, 3, 5, 7, 8 +[36] Jiayu Yang, Wei Mao, Jose M Alvarez, and Miaomiao Liu. Cost volume pyramid based depth inference for multi-view stereo. In CVPR, pages 4877-4886, 2020. 3 +[37] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoptrees for real-time rendering of neural radiance fields. In ICCV, pages 5752-5761, 2021. 2 +[38] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In CVPR, pages 4578-4587, 2021. 1, 2 +[39] Feihu Zhang, Victor Prisacariu, Ruigang Yang, and Philip HS Torr. Ga-net: Guided aggregation net for end-to-end stereo matching. In CVPR, pages 185-194, 2019. 3 +[40] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020.5 +[41] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 7 +[42] Zhengyou Zhang. A flexible new technique for camera calibration. IEEE TPAMI, 22(11):1330-1334, 2000. 5 +[43] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. ACM TOG, 2018. 2 \ No newline at end of file diff --git a/2024/Generalizable Novel-View Synthesis using a Stereo Camera/images.zip b/2024/Generalizable Novel-View Synthesis using a Stereo Camera/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f6709a6cd906cd67fd90810e81218d567eee3b49 --- /dev/null +++ b/2024/Generalizable Novel-View Synthesis using a Stereo Camera/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:412c41485444e0e3079fa5dac23b48854ce4f33309b2b59d6e2cda6dec4a0c48 +size 586360 diff --git a/2024/Generalizable Novel-View Synthesis using a Stereo Camera/layout.json b/2024/Generalizable Novel-View Synthesis using a Stereo Camera/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2755c920e9889f46f922dab5e1c758860839162b --- /dev/null +++ b/2024/Generalizable Novel-View Synthesis using a Stereo Camera/layout.json @@ -0,0 +1,8286 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 114, + 103, + 480, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 103, + 480, + 121 + ], + "spans": [ + { + "bbox": [ + 114, + 103, + 480, + 121 + ], + "type": "text", + "content": "Generalizable Novel-View Synthesis using a Stereo Camera" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 85, + 142, + 162, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 142, + 162, + 156 + ], + "spans": [ + { + "bbox": [ + 85, + 142, + 162, + 156 + ], + "type": "text", + "content": "Haechan Lee" + }, + { + "bbox": [ + 85, + 142, + 162, + 156 + ], + "type": "inline_equation", + "content": "^{1,*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 189, + 143, + 264, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 143, + 264, + 157 + ], + "spans": [ + { + "bbox": [ + 189, + 143, + 264, + 157 + ], + "type": "text", + "content": "Wonjoon Jin" + }, + { + "bbox": [ + 189, + 143, + 264, + 157 + ], + "type": "inline_equation", + "content": "^{2,*}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 291, + 143, + 395, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 291, + 143, + 395, + 157 + ], + "spans": [ + { + "bbox": [ + 291, + 143, + 395, + 157 + ], + "type": "text", + "content": "Seung-Hwan Baek" + }, + { + "bbox": [ + 291, + 143, + 395, + 157 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 421, + 143, + 507, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 421, + 143, + 507, + 157 + ], + "spans": [ + { + "bbox": [ + 421, + 143, + 507, + 157 + ], + "type": "text", + "content": "Sunghyun Cho" + }, + { + "bbox": [ + 421, + 143, + 507, + 157 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 229, + 163, + 364, + 176 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 163, + 364, + 176 + ], + "spans": [ + { + "bbox": [ + 229, + 163, + 364, + 176 + ], + "type": "text", + "content": "POSTECH " + }, + { + "bbox": [ + 229, + 163, + 364, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 229, + 163, + 364, + 176 + ], + "type": "text", + "content": "GSAI & " + }, + { + "bbox": [ + 229, + 163, + 364, + 176 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 229, + 163, + 364, + 176 + ], + "type": "text", + "content": "CSE" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 163, + 179, + 428, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 163, + 179, + 428, + 191 + ], + "spans": [ + { + "bbox": [ + 163, + 179, + 428, + 191 + ], + "type": "text", + "content": "{gocks8, jinwj1996, shwbaek, s.cho}@postech.ac.kr" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 219, + 192, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 219, + 192, + 232 + ], + "spans": [ + { + "bbox": [ + 143, + 219, + 192, + 232 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 244, + 290, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 244, + 290, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 244, + 290, + 448 + ], + "type": "text", + "content": "In this paper, we propose the first generalizable view synthesis approach that specifically targets multi-view stereocamera images. Since recent stereo matching has demonstrated accurate geometry prediction, we introduce stereo matching into novel-view synthesis for high-quality geometry reconstruction. To this end, this paper proposes a novel framework, dubbed StereoNeRF, which integrates stereo matching into a NeRF-based generalizable view synthesis approach. StereoNeRF is equipped with three key components to effectively exploit stereo matching in novel-view synthesis: a stereo feature extractor, a depth-guided plane-sweeping, and a stereo depth loss. Moreover, we propose the StereoNVS dataset, the first multi-view dataset of stereocamera images, encompassing a wide variety of both real and synthetic scenes. Our experimental results demonstrate that StereoNeRF surpasses previous approaches in generalizable view synthesis." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 459, + 128, + 471 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 459, + 128, + 471 + ], + "spans": [ + { + "bbox": [ + 47, + 459, + 128, + 471 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 479, + 287, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 287, + 598 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 287, + 598 + ], + "type": "text", + "content": "Novel-view synthesis is a long-standing ill-posed problem in computer vision and graphics, which is inherently challenging due to the necessity of predicting both the geometry and texture from images of a target scene. Recently, Neural Radiance Fields (NeRF) [24] have achieved photorealistic results by jointly optimizing geometry and radiance fields with a coordinate-based network. However, the need for per-scene optimization, which has to learn representations for each target scene individually, restricts its applicability, as it requires additional training time for such optimization." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 599, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 599, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 599, + 288, + 696 + ], + "type": "text", + "content": "Recent approaches [2, 4, 16, 21, 30, 32, 33, 38] have addressed this issue of synthesizing novel-view images on-the-fly for unseen scenes without per-scene optimization. Early studies [4, 33, 38] utilize an image encoder to train a generic view interpolation function, enabling the estimation of NeRF parameters from unseen input images in a feed-forward manner. However, this single feed-forward manner for estimating geometry and color exacerbates the ill" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 307, + 219, + 383, + 316 + ], + "blocks": [ + { + "bbox": [ + 307, + 219, + 383, + 316 + ], + "lines": [ + { + "bbox": [ + 307, + 219, + 383, + 316 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 383, + 316 + ], + "type": "image", + "image_path": "43f30b7c5765bfbaecae03499b724cdaea92243b9d92122ef04f9e8d9414441e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 316, + 369, + 324 + ], + "lines": [ + { + "bbox": [ + 324, + 316, + 369, + 324 + ], + "spans": [ + { + "bbox": [ + 324, + 316, + 369, + 324 + ], + "type": "text", + "content": "(a) Baseline" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 336, + 547, + 381 + ], + "lines": [ + { + "bbox": [ + 305, + 336, + 547, + 381 + ], + "spans": [ + { + "bbox": [ + 305, + 336, + 547, + 381 + ], + "type": "text", + "content": "Figure 1. Novel-view synthesis results of a baseline method [16] and ours. The baseline shows degraded performances, even trained using stereo-camera images (b). In contrast, fully exploiting stereo-camera images, our method shows superior results (c)." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 384, + 219, + 460, + 316 + ], + "blocks": [ + { + "bbox": [ + 384, + 219, + 460, + 316 + ], + "lines": [ + { + "bbox": [ + 384, + 219, + 460, + 316 + ], + "spans": [ + { + "bbox": [ + 384, + 219, + 460, + 316 + ], + "type": "image", + "image_path": "c37eff9227541f2a927ca8ed80391be85c840f40bb690cd779b4ccaa448dfc4f.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 401, + 316, + 445, + 324 + ], + "lines": [ + { + "bbox": [ + 401, + 316, + 445, + 324 + ], + "spans": [ + { + "bbox": [ + 401, + 316, + 445, + 324 + ], + "type": "text", + "content": "(b) Baseline" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 391, + 325, + 455, + 333 + ], + "lines": [ + { + "bbox": [ + 391, + 325, + 455, + 333 + ], + "spans": [ + { + "bbox": [ + 391, + 325, + 455, + 333 + ], + "type": "text", + "content": "(using stereo images)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 460, + 219, + 547, + 316 + ], + "blocks": [ + { + "bbox": [ + 460, + 219, + 547, + 316 + ], + "lines": [ + { + "bbox": [ + 460, + 219, + 547, + 316 + ], + "spans": [ + { + "bbox": [ + 460, + 219, + 547, + 316 + ], + "type": "image", + "image_path": "b505cd829b7e1639400ad26b07f9406d86fdcb5b9df89a2fa92d553cc92b96e4.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 484, + 316, + 515, + 324 + ], + "lines": [ + { + "bbox": [ + 484, + 316, + 515, + 324 + ], + "spans": [ + { + "bbox": [ + 484, + 316, + 515, + 324 + ], + "type": "text", + "content": "(c) Ours" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 467, + 325, + 532, + 333 + ], + "lines": [ + { + "bbox": [ + 467, + 325, + 532, + 333 + ], + "spans": [ + { + "bbox": [ + 467, + 325, + 532, + 333 + ], + "type": "text", + "content": "(using stereo images)" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 388, + 547, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 547, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 547, + 472 + ], + "type": "text", + "content": "posedness, resulting in low-quality geometries and rendering results. For better geometry reasoning, MVSNeRF [2] and GeoNeRF [16] leverage the multi-view stereo (MVS) approach to handle occlusions in a 3D scene. Nonetheless, they struggle with inaccurate geometry prediction and limited synthesis accuracy, as shown in the synthesis result of GeoNeRF (Fig. 1 (a))." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 474, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 548, + 714 + ], + "type": "text", + "content": "To tackle this challenge, we propose the first generalizable NeRF approach that leverages stereo-camera images, which are easily accessible thanks to the ubiquity of stereo cameras in most mobile devices. Recent advance in learning-based stereo estimation has demonstrated accurate geometry prediction, often even outperforming learning-based MVS methods as shown in Fig. 2. The superior performance of stereo estimation can be attributed to several key factors. First, unlike MVS that assumes arbitrary number of inputs with arbitrary camera positions and orientations, stereo matching assumes two stereo inputs with a fixed baseline. This constraint allows a more optimal network architecture that can effectively find dense correspondences between input images, and significantly eases the training difficulty. Moreover, larger scale of stereomatching datasets [12, 23, 27] compared to MVS datasets has facilitated remarkable generalization capabilities in stereo estimation network. Therefore, we aim to harness this accurate geometric information from stereo images to alleviate the ill-posedness of generalizable view synthesis." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 703, + 130, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 703, + 130, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 703, + 130, + 712 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4939" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 286, + 120 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 286, + 120 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 286, + 120 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 286, + 120 + ], + "type": "image", + "image_path": "a5d0501def5d90ee2d3f6acd8e1baec8277930a78ebfea87967b841bc45298d6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 123, + 286, + 156 + ], + "lines": [ + { + "bbox": [ + 47, + 123, + 286, + 156 + ], + "spans": [ + { + "bbox": [ + 47, + 123, + 286, + 156 + ], + "type": "text", + "content": "Figure 2. Usefulness of exploiting binocular stereo. Comparison on depth estimation between a learning-based MVS method [26] and a learning-based binocular stereo method [35]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 163, + 286, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 163, + 286, + 210 + ], + "spans": [ + { + "bbox": [ + 47, + 163, + 286, + 210 + ], + "type": "text", + "content": "However, since previous methods do not explicitly consider stereoscopic inputs, they cannot leverage the aforementioned benefits, resulting in degenerated performances shown in Fig. 1 (b) despite using stereo-camera images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 211, + 286, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 211, + 286, + 462 + ], + "spans": [ + { + "bbox": [ + 47, + 211, + 286, + 462 + ], + "type": "text", + "content": "This paper proposes StereoNeRF, a novel generalizable view synthesis framework leveraging stereo images. StereoNeRF integrates stereo matching into NeRF-based generalizable view synthesis approach, where the stereo matching provides vital geometric information. To this end, we first introduce a stereo feature extractor, which extracts geometry-aware features by correlating horizontal epipolar lines within stereo images. In addition, the stereo feature extractor takes advantage of stereo-correlated features from an off-the-shelf stereo estimation network, which can transfer rich geometric knowledge to our model. Furthermore, with the reliable depth estimated from the stereo estimation network, we aggregate multi-view features through a depth-guided plane-sweeping technique. This technique ensures correspondence matching around the geometry in cost volume construction. We also present a stereo depth loss utilizing the estimated stereo depth. These additional geometric cues from the stereo matching effectively mitigate the ill-posedness in generalizable view synthesis. Notably, our framework leveraging stereo images surpasses the previous approaches [16, 21], which rely on extra depth information." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 463, + 286, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 463, + 286, + 569 + ], + "spans": [ + { + "bbox": [ + 46, + 463, + 286, + 569 + ], + "type": "text", + "content": "Furthermore, we propose the StereoNVS dataset, which is the first dataset for training and evaluation of novel-view synthesis using stereo-camera images. Our StereoNVS dataset provides real-world and synthetic stereo images. Our extensive evaluation on the StereoNVS dataset shows that stereo-pair inputs can effectively enhance the quality of novel-view synthesis, and shows that StereoNeRF outperforms previous generalizable novel-view synthesis approaches in terms of image and shape qualities." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 570, + 193, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 570, + 193, + 581 + ], + "spans": [ + { + "bbox": [ + 59, + 570, + 193, + 581 + ], + "type": "text", + "content": "Our contributions are as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 582, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 47, + 582, + 286, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 582, + 286, + 616 + ], + "spans": [ + { + "bbox": [ + 47, + 582, + 286, + 616 + ], + "type": "text", + "content": "- We propose a generalized NeRF approach that leverages stereo-camera images for the first time to alleviate the ill-posedness in generalizable novel-view synthesis." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 618, + 286, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 618, + 286, + 677 + ], + "spans": [ + { + "bbox": [ + 47, + 618, + 286, + 677 + ], + "type": "text", + "content": "- We propose a novel framework, StereoNeRF, which exploits the benefits of stereo images by integrating stereo matching into generalizable view synthesis. To this end, we present a stereo feature extractor, a depth-guided plane-sweeping, and a stereo depth loss in our framework." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 677, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 286, + 712 + ], + "type": "text", + "content": "- We also present the StereoNVS dataset, the first dataset for training and evaluation of novel-view synthesis obtained by stereo cameras." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 71, + 392, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 392, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 392, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 91, + 428, + 102 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 91, + 428, + 102 + ], + "spans": [ + { + "bbox": [ + 306, + 91, + 428, + 102 + ], + "type": "text", + "content": "2.1. Novel-View Synthesis" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 109, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 109, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 109, + 545, + 300 + ], + "type": "text", + "content": "Novel-view synthesis aims to synthesize target-view images from reference-view images. Early approaches synthesize novel-view images by blending pixels from multiple input images [6, 13, 20]. Recent work adopting neural volume representations has shown remarkable novel-view synthesis results. Zhou et al. [43] propose multi-plane images (MPI) representation estimated from input images, but their methods produce valid novel-view images only for narrow ranges of camera poses. Mildenhall et al. [24] propose Neural Radiance Fields (NeRF) that can synthesize photo-realistic target-view images via neural implicit representations and volume rendering. Albeit its photo-realism, computation-heavy per-scene optimization is needed. Recent variants such as [9, 17, 25, 37] have remarkably reduced the optimization time, but they still require large memory and several minutes for training." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 306, + 545, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 306, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 304, + 306, + 545, + 426 + ], + "type": "text", + "content": "Generalizable View Synthesis. Recently, many novel-view synthesis approaches without per-scene optimization have been proposed. Several studies directly predict pixel colors by aggregating image features from aligned pixels, without 3D representations. Suhail et al. [30] and Varma et al. [32] adopt a transformer-based network [7] to compute features along the epipolar lines, which needs a large number of training images for high-quality view synthesis. Du et al. [8] propose a framework to synthesize target views from two images with small overlapped regions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "text", + "content": "Another research direction predicts volumetric representation [24] from aggregated features from reference views and synthesizes images via volumetric rendering. PixelNeRF, SRF, and GRF [4, 31, 38] predict radiance fields from pixel-aligned features using an MLP. Among these methods, SRF [4] uses two sampled images from an image collection captured by a monocular camera as a stereo pair, but shows limited synthesis quality. IBRNet [33] proposes to learn generic view interpolation functions, but suffers from artifacts for challenging scenes with complex geometries. MVSNeRF, GeoNeRF and NeuRay [2, 16, 21] utilize the MVS approach using cost volume for better occlusion handling in generalizable view synthesis." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 581, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 581, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 581, + 545, + 665 + ], + "type": "text", + "content": "However, all the aforementioned methods often fail to capture accurate geometry, particularly in textureless regions, leading to limited synthesis results. Unlike these methods, our framework exploits stereo-camera images with a fixed baseline to effectively capture the geometry of complex scenes as well as textureless regions for high-quality novel-view synthesis." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 671, + 515, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 515, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 515, + 684 + ], + "type": "text", + "content": "2.2. Stereo Matching and Multi-View Stereo" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": "Geometry estimation is a long-standing problem in computer vision that has a variety of applications. Among them," + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4940" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": "stereo matching is a task that takes rectified stereo images and computes stereo correspondence to estimate disparities [15, 27]. Recently, a huge number of data-driven approaches have been introduced with the emergence of a vast amount of stereo-matching datasets [12, 23, 27] and made significant progress [1, 23, 35, 39]. MVS approaches that use more than two views have been extensively studied as well [11, 19, 29]. Recently, learning-based MVS approaches have been proposed, e.g., Cheng et al. [3], Gu et al. [14], and Yang et al. [36] present efficient frameworks that cascade cost volumes in a coarse-to-fine manner to enable high-resolution depth estimation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 217, + 289, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 217, + 289, + 314 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 289, + 314 + ], + "type": "text", + "content": "In our work, we have the best of both worlds in our generalizable view synthesis framework. We adopt the MVS approach to aggregate multi-view information for occlusion-aware geometry estimation, as done in GeoNerveRF [16]. Furthermore, our framework also assumes structured stereo images as inputs, and integrates the two-view stereo matching into our framework for robust and accurate geometry estimation in generalizable novel-view synthesis." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 327, + 126, + 340 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 327, + 126, + 340 + ], + "spans": [ + { + "bbox": [ + 47, + 327, + 126, + 340 + ], + "type": "text", + "content": "3. StereoNeRF" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 350, + 289, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 350, + 289, + 469 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 289, + 469 + ], + "type": "text", + "content": "For novel-view synthesis, our framework takes a set of rectified stereo-camera images and estimates neural radiance fields [24] from which novel-view images are rendered. StereoNeRF integrates a well-designed stereo-matching algorithm into the existing generalizable view synthesis approach [16]. In this section, we first provide an overall pipeline that utilizes an pre-trained stereo estimation network [35]. Then, we explain each step of our framework and the training process in detail, highlighting how to integrate stereo matching into our framework." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "content": "Overall pipeline. Fig. 3 shows an overview of our pipeline. Our framework utilizes " + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "content": " pairs of stereo images of a target scene " + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\{(I_L^n,I_R^n)\\}_{n = 1}^N" + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "content": " for novel-view synthesis. In the first step, the pre-trained stereo estimation network takes the " + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "content": "-th image pair " + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "inline_equation", + "content": "(I_L^n,I_R^n)" + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "content": ", and outputs stereo depths " + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "inline_equation", + "content": "(d_{s,L}^{n},d_{s,R}^{n})" + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "content": " and stereo-correlated features " + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "inline_equation", + "content": "(t_L^n,t_R^n)" + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "content": ". In the second step, a stereo feature extractor takes the stereo image pair and the stereo-correlated features, and outputs stereo image features " + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "inline_equation", + "content": "(f_L^n,f_R^n)" + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "content": ", explained in Sec. 3.1. The third step aggregates stereo image features from all viewpoints " + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\{(f_L^n,f_R^n)\\}_{n = 1}^N" + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "content": " to build 3D feature volumes " + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\{(\\phi_L^n,\\phi_R^n)\\}_{n = 1}^N" + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "content": " using an MVS network (Sec. 3.2). For constructing feature volumes that faithfully reflect the 3D geometry of a target scene, the third step adopts a depth-guided plane-sweeping which utilizes the estimated stereo depths " + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "inline_equation", + "content": "(d_{s,L}^{n},d_{s,R}^{n})" + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "content": ". Finally, a neural renderer predicts radiance fields from the stereo image features " + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\{(f_L^n,f_R^n)\\}_{n = 1}^N" + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "content": " and the feature volumes " + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\{(\\phi_L^n,\\phi_R^n)\\}_{n = 1}^N" + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "content": " from all viewpoints, and novel-view images are synthesized from the radiance fields (Sec. 3.3)." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 547, + 194 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 547, + 194 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 547, + 194 + ], + "type": "image", + "image_path": "a02480fe535e1d999f71da61d0d90036ebf182ec44e92387dbfb611677a36928.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 200, + 547, + 267 + ], + "lines": [ + { + "bbox": [ + 304, + 200, + 547, + 267 + ], + "spans": [ + { + "bbox": [ + 304, + 200, + 547, + 267 + ], + "type": "text", + "content": "Figure 3. Overview of StereoNeRF. StereoNeRF consists of a shared feature extractor and a neural renderer. We design the feature extractor with a stereo estimation network, a stereo feature extractor, and a MVS network. StereoNeRF takes " + }, + { + "bbox": [ + 304, + 200, + 547, + 267 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 200, + 547, + 267 + ], + "type": "text", + "content": " pairs of stereo images and a camera pose as inputs, and synthesizes a novel-view image of the camera pose." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 276, + 534, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 276, + 534, + 289 + ], + "spans": [ + { + "bbox": [ + 305, + 276, + 534, + 289 + ], + "type": "text", + "content": "3.1. Feature Extraction from Stereo Image Pairs" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 294, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 294, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 294, + 545, + 437 + ], + "type": "text", + "content": "Unlike previous approaches, which extract features of input images respectively, our stereo feature extractor takes rectified stereo-camera images and computes image feature maps by exploiting the epipolar geometry between them. To this end, our stereo feature extractor consists of three parts: CNN encoders, stereo attention modules (SAM) [5], and CNN decoders (Fig. 4). Moreover, we propose integrating the stereo-correlated features " + }, + { + "bbox": [ + 304, + 294, + 545, + 437 + ], + "type": "inline_equation", + "content": "(t_L^n,t_R^n)" + }, + { + "bbox": [ + 304, + 294, + 545, + 437 + ], + "type": "text", + "content": " from the pretrained stereo estimation network into the SAM (green line in Fig. 4). This integration inherits geometric cues from the stereo estimation network. In the following, we provide the detailed description of the stereo feature extractor." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 438, + 546, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 546, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 546, + 594 + ], + "type": "text", + "content": "First, the weight-shared CNN encoders project each of " + }, + { + "bbox": [ + 304, + 438, + 546, + 594 + ], + "type": "inline_equation", + "content": "I_L^n" + }, + { + "bbox": [ + 304, + 438, + 546, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 438, + 546, + 594 + ], + "type": "inline_equation", + "content": "I_R^n" + }, + { + "bbox": [ + 304, + 438, + 546, + 594 + ], + "type": "text", + "content": " into the feature space. Then, we adopt the stereo attention module (SAM) [5] to fuse two feature maps on the horizontal epipolar lines. To this end, SAM estimates stereo correspondences between two feature maps and explicitly adds stereo-correspondent features, as shown in Fig. 5. In addition, we extend the SAM by aggregating the stereocorrelated features " + }, + { + "bbox": [ + 304, + 438, + 546, + 594 + ], + "type": "inline_equation", + "content": "(t_L^n, t_R^n)" + }, + { + "bbox": [ + 304, + 438, + 546, + 594 + ], + "type": "text", + "content": " in this feature fusion process. Note that the stereo-correlated features are computed from the pre-trained stereo estimation network, which provides vital geometric cues. These stereo-correlated features enhance stereo correspondences between two feature maps, leading to higher-quality fused feature maps." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": "Specific explanation of the aforementioned feature fusion of the SAM is as follows (Fig. 5). For the left feature, a warping matrix " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{R\\to L}" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " is built by multiplying a query matrix and a key matrix followed by a softmax layer. The query matrix is computed from the left feature and the stereocorrelated features, and the key matrix is computed from the right feature and the stereo-correlated features. Then, a value matrix, computed from the right feature and the stereo-correlated feature, is warped along the epipolar lines by multiplying the warping matrix " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{R\\to L}" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": ", then added to" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "4941" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 69, + 289, + 361 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 289, + 361 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 289, + 361 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 289, + 361 + ], + "type": "image", + "image_path": "85b68ed08f4f2f565d13c77166d36d7ac23cb54574c5916cbfd37d6202094e17.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 369, + 287, + 392 + ], + "lines": [ + { + "bbox": [ + 46, + 369, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 287, + 392 + ], + "type": "text", + "content": "Figure 4. Stereo feature extractor of StereoNeRF, where an pretrained stereo estimation network is incorporated." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 401, + 287, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 287, + 425 + ], + "type": "text", + "content": "the left feature. The same feature fusion process is also performed for the right feature." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 425, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 425, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 287, + 498 + ], + "type": "text", + "content": "For effective fusion of information from both stereo images and stereo-correlated features, the stereo feature extractor has three sequentially stacked SAM. Finally, the weight-shared CNN decoders take each of the fused features from the SAM, and generate stereo image features " + }, + { + "bbox": [ + 46, + 425, + 287, + 498 + ], + "type": "inline_equation", + "content": "(f_{L}^{n}, f_{R}^{n})" + }, + { + "bbox": [ + 46, + 425, + 287, + 498 + ], + "type": "text", + "content": ", which will be used for building feature volumes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 503, + 275, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 503, + 275, + 516 + ], + "spans": [ + { + "bbox": [ + 47, + 503, + 275, + 516 + ], + "type": "text", + "content": "3.2. Depth-guided Feature Volume Construction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 522, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 665 + ], + "type": "text", + "content": "Once stereo image features are obtained, we aggregate these features from all viewpoints to create cost volumes. To this end, we adopt a plane-sweeping-based approach, which computes multi-view correspondences among these features. The plane-sweeping-based approach first defines a depth range within pre-computed near and far depths obtained from COLMAP [28]. Then, it hypothesizes depth planes in the entire depth range. On these depth planes, multi-view stereo image features are aggregated via plane-sweeping to build cost volumes. Then, these cost volumes are processed to yield feature volumes, which will be used later in predicting neural radiance fields." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "However, this plane-sweeping-based approach often struggles with accurate geometry estimation. Since this approach searches the entire depth range of a scene, correspondence matching is inaccurate and prone to error, espe" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 547, + 252 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 547, + 252 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 547, + 252 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 547, + 252 + ], + "type": "image", + "image_path": "bd3f7efda17ad332653d31cc2a8a05a2e9740ed4281c10f946bbc2e77355d969.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 258, + 545, + 291 + ], + "lines": [ + { + "bbox": [ + 305, + 258, + 545, + 291 + ], + "spans": [ + { + "bbox": [ + 305, + 258, + 545, + 291 + ], + "type": "text", + "content": "Figure 5. Stereo attention module used in the stereo feature extractor. We exploit the rich features from the pre-trained stereo estimation module." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 312, + 297, + 538, + 415 + ], + "blocks": [ + { + "bbox": [ + 312, + 297, + 538, + 415 + ], + "lines": [ + { + "bbox": [ + 312, + 297, + 538, + 415 + ], + "spans": [ + { + "bbox": [ + 312, + 297, + 538, + 415 + ], + "type": "image", + "image_path": "6b75aa4878ea606421023cecf99e7405733a829548f84266859484cf17cc1e76.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 417, + 520, + 428 + ], + "lines": [ + { + "bbox": [ + 331, + 417, + 520, + 428 + ], + "spans": [ + { + "bbox": [ + 331, + 417, + 520, + 428 + ], + "type": "text", + "content": "Figure 6. Feature volume construction using DGPS." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 445, + 545, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 545, + 542 + ], + "type": "text", + "content": "cially in textureless regions. To tackle this issue, we introduce a depth-guided plane-sweeping (DGPS), which constructs cost volumes around the stereo depth predicted from the stereo estimation network. Unlike the previous approach using the entire depth range, DGPS introduces a dynamic search range that varies based on the estimated stereo depth. This technique significantly constrains the search space and reduces outliers in correspondence matching." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 544, + 546, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 546, + 712 + ], + "type": "text", + "content": "Fig. 6 depicts the feature volume construction using DGPS. Instead of searching the entire depth range, we hypothesize depth planes around the stereo depths " + }, + { + "bbox": [ + 304, + 544, + 546, + 712 + ], + "type": "inline_equation", + "content": "(d_{s,L}^n,d_{s,R}^n)" + }, + { + "bbox": [ + 304, + 544, + 546, + 712 + ], + "type": "text", + "content": ". Then, through DGPS, we aggregate the stereo image features from all viewpoints " + }, + { + "bbox": [ + 304, + 544, + 546, + 712 + ], + "type": "inline_equation", + "content": "\\{(f_L^n,f_R^n)\\}_{n = 1}^N" + }, + { + "bbox": [ + 304, + 544, + 546, + 712 + ], + "type": "text", + "content": " on these depth planes to build cost volumes. These cost volumes are further processed via the MVS network to obtain feature volumes " + }, + { + "bbox": [ + 304, + 544, + 546, + 712 + ], + "type": "inline_equation", + "content": "(\\phi_L^n,\\phi_R^n)" + }, + { + "bbox": [ + 304, + 544, + 546, + 712 + ], + "type": "text", + "content": " and depth maps " + }, + { + "bbox": [ + 304, + 544, + 546, + 712 + ], + "type": "inline_equation", + "content": "(d_{s,L}^n,d_{s,R}^n)" + }, + { + "bbox": [ + 304, + 544, + 546, + 712 + ], + "type": "text", + "content": ". We repeat this process to obtain feature volumes and depth maps for every viewpoint. The resulting multi-view feature volumes and depth maps are denoted as " + }, + { + "bbox": [ + 304, + 544, + 546, + 712 + ], + "type": "inline_equation", + "content": "\\{(\\phi_L^L,\\phi_R^L)\\}_{n = 1}^N" + }, + { + "bbox": [ + 304, + 544, + 546, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 544, + 546, + 712 + ], + "type": "inline_equation", + "content": "\\{(d_{m,L}^{n},d_{m,R}^{n})\\}_{n = 1}^{N}" + }, + { + "bbox": [ + 304, + 544, + 546, + 712 + ], + "type": "text", + "content": ", respectively. Refer to Sec. B.3 in the supplementary document for more details about our feature volume construction." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4942" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 179, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 179, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 179, + 85 + ], + "type": "text", + "content": "3.3. Rendering Novel Views" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 90, + 289, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 90, + 289, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 90, + 289, + 281 + ], + "type": "text", + "content": "Once feature volumes are constructed, novel-view images can be rendered via neural rendering. For rendering novel-view images, our framework adopts the rendering procedure of GeoNeRF [16]. In the following, we briefly describe the rendering procedure. Given a target viewpoint, we cast a ray for each pixel and sample points along the ray in the 3D space. Then, for each sampled point, we sample features from all the feature volumes " + }, + { + "bbox": [ + 46, + 90, + 289, + 281 + ], + "type": "inline_equation", + "content": "\\{(\\phi_L^n,\\phi_R^n)\\}_{n = 1}^N" + }, + { + "bbox": [ + 46, + 90, + 289, + 281 + ], + "type": "text", + "content": ". We also sample features from all the image features " + }, + { + "bbox": [ + 46, + 90, + 289, + 281 + ], + "type": "inline_equation", + "content": "\\{(f_L^n,f_R^n)\\}_{n = 1}^N" + }, + { + "bbox": [ + 46, + 90, + 289, + 281 + ], + "type": "text", + "content": " by projecting the sampled point onto input images. The sampled volume and image features are aggregated by a neural renderer network, and the color and density are estimated. Finally, by integrating the estimated color values and depth values of sampled points with their densities along each ray, we obtain the color " + }, + { + "bbox": [ + 46, + 90, + 289, + 281 + ], + "type": "inline_equation", + "content": "\\hat{c}" + }, + { + "bbox": [ + 46, + 90, + 289, + 281 + ], + "type": "text", + "content": " and the depth " + }, + { + "bbox": [ + 46, + 90, + 289, + 281 + ], + "type": "inline_equation", + "content": "d_{r}" + }, + { + "bbox": [ + 46, + 90, + 289, + 281 + ], + "type": "text", + "content": " at each pixel of a novel view, respectively." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 289, + 170, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 289, + 170, + 301 + ], + "spans": [ + { + "bbox": [ + 47, + 289, + 170, + 301 + ], + "type": "text", + "content": "3.4. Training StereoNeRF" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 307, + 287, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 307, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 46, + 307, + 287, + 367 + ], + "type": "text", + "content": "We train StereoNeRF with ground-truth images and pseudoground-truth depths for high-quality novel-view synthesis. The pseudo-ground-truth depths " + }, + { + "bbox": [ + 46, + 307, + 287, + 367 + ], + "type": "inline_equation", + "content": "d_{gt}" + }, + { + "bbox": [ + 46, + 307, + 287, + 367 + ], + "type": "text", + "content": " are obtained from the pre-trained stereo estimation network [35]. Our training loss " + }, + { + "bbox": [ + 46, + 307, + 287, + 367 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 46, + 307, + 287, + 367 + ], + "type": "text", + "content": " is defined as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 136, + 376, + 287, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 376, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 136, + 376, + 287, + 389 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {c} + \\mathcal {L} _ {d}, \\tag {1}", + "image_path": "85c058d9e14efcdb999a4075feb943340f08af691d3614a4e442a4f3264611ba.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 397, + 287, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 397, + 287, + 433 + ], + "spans": [ + { + "bbox": [ + 46, + 397, + 287, + 433 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 397, + 287, + 433 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 46, + 397, + 287, + 433 + ], + "type": "text", + "content": " is a color loss and " + }, + { + "bbox": [ + 46, + 397, + 287, + 433 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_d" + }, + { + "bbox": [ + 46, + 397, + 287, + 433 + ], + "type": "text", + "content": " is a depth loss. The color loss is defined as a mean-squared-error (MSE) between the rendered colors " + }, + { + "bbox": [ + 46, + 397, + 287, + 433 + ], + "type": "inline_equation", + "content": "\\hat{c}" + }, + { + "bbox": [ + 46, + 397, + 287, + 433 + ], + "type": "text", + "content": " and ground-truth colors [24]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 434, + 287, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 287, + 469 + ], + "type": "text", + "content": "To alleviate the shape-radiance ambiguity [40] in scene reconstruction using neural radiance fields, we introduce the depth loss " + }, + { + "bbox": [ + 46, + 434, + 287, + 469 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_d" + }, + { + "bbox": [ + 46, + 434, + 287, + 469 + ], + "type": "text", + "content": ", which is defined as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 94, + 476, + 287, + 493 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 476, + 287, + 493 + ], + "spans": [ + { + "bbox": [ + 94, + 476, + 287, + 493 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d} = \\lambda_ {d} ^ {s e l f} \\mathcal {L} _ {d} ^ {s e l f} + \\lambda_ {d} ^ {s t e r e o} \\mathcal {L} _ {d} ^ {s t e r e o}, \\tag {2}", + "image_path": "e6c5d2ea688f2c2b56214cc6acf70cc4b22770e480bcc1146c6b8b222ba62356.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 501, + 287, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 501, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 46, + 501, + 287, + 563 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 501, + 287, + 563 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_d^{self}" + }, + { + "bbox": [ + 46, + 501, + 287, + 563 + ], + "type": "text", + "content": " is a self-supervised depth loss from GeoN-NeRF [16] and " + }, + { + "bbox": [ + 46, + 501, + 287, + 563 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_d^{stereo}" + }, + { + "bbox": [ + 46, + 501, + 287, + 563 + ], + "type": "text", + "content": " is our proposed stereo depth loss. " + }, + { + "bbox": [ + 46, + 501, + 287, + 563 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_d^{self}" + }, + { + "bbox": [ + 46, + 501, + 287, + 563 + ], + "type": "text", + "content": " penalizes the depths estimated from both the stereo estimation network " + }, + { + "bbox": [ + 46, + 501, + 287, + 563 + ], + "type": "inline_equation", + "content": "(d_s)" + }, + { + "bbox": [ + 46, + 501, + 287, + 563 + ], + "type": "text", + "content": " and the MVS network " + }, + { + "bbox": [ + 46, + 501, + 287, + 563 + ], + "type": "inline_equation", + "content": "(d_m)" + }, + { + "bbox": [ + 46, + 501, + 287, + 563 + ], + "type": "text", + "content": " by comparing them to the depth rendered from NeRF " + }, + { + "bbox": [ + 46, + 501, + 287, + 563 + ], + "type": "inline_equation", + "content": "(d_r)" + }, + { + "bbox": [ + 46, + 501, + 287, + 563 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 563, + 287, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 563, + 287, + 648 + ], + "spans": [ + { + "bbox": [ + 46, + 563, + 287, + 648 + ], + "type": "text", + "content": "Our framework estimates depths from the stereo estimation network " + }, + { + "bbox": [ + 46, + 563, + 287, + 648 + ], + "type": "inline_equation", + "content": "(d_{s})" + }, + { + "bbox": [ + 46, + 563, + 287, + 648 + ], + "type": "text", + "content": ", the MVS network " + }, + { + "bbox": [ + 46, + 563, + 287, + 648 + ], + "type": "inline_equation", + "content": "(d_{m})" + }, + { + "bbox": [ + 46, + 563, + 287, + 648 + ], + "type": "text", + "content": ", and the neural renderer network " + }, + { + "bbox": [ + 46, + 563, + 287, + 648 + ], + "type": "inline_equation", + "content": "(d_{r})" + }, + { + "bbox": [ + 46, + 563, + 287, + 648 + ], + "type": "text", + "content": ". The stereo depth loss " + }, + { + "bbox": [ + 46, + 563, + 287, + 648 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_d^{stereo}" + }, + { + "bbox": [ + 46, + 563, + 287, + 648 + ], + "type": "text", + "content": " guides these networks to predict more accurate depths using the pseudo-ground-truth depths " + }, + { + "bbox": [ + 46, + 563, + 287, + 648 + ], + "type": "inline_equation", + "content": "d_{gt}" + }, + { + "bbox": [ + 46, + 563, + 287, + 648 + ], + "type": "text", + "content": ". This loss introduces a further quality improvement, especially in geometric details as will be shown in Sec. 5.3. " + }, + { + "bbox": [ + 46, + 563, + 287, + 648 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_d^{stereo}" + }, + { + "bbox": [ + 46, + 563, + 287, + 648 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 96, + 655, + 287, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 655, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 96, + 655, + 287, + 669 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d} ^ {\\text {s t e r e o}} = \\lambda_ {1} \\mathcal {L} _ {d} ^ {s} + \\lambda_ {2} \\mathcal {L} _ {d} ^ {m} + \\lambda_ {3} \\mathcal {L} _ {d} ^ {r}, \\tag {3}", + "image_path": "c74da43229a388bd2aef469126bde4431791895e970a27f605c8927e4492f141.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_d^s" + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_d^m" + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_d^r" + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": " penalizes depths " + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "inline_equation", + "content": "d_s" + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "inline_equation", + "content": "d_m" + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "inline_equation", + "content": "d_r" + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": ", respectively, by comparing them with " + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "inline_equation", + "content": "d_{gt}" + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "inline_equation", + "content": "d_{gt}" + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": " is different from " + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "inline_equation", + "content": "d_s" + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": ". We obtain " + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "inline_equation", + "content": "d_{gt}" + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": " from the pre-trained stereo" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "estimation network with frozen parameters. On the other hand, we obtain " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "d_{s}" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " from the stereo estimation network in our framework, which has trainable parameters. Refer to Sec. B.4 in the supplementary document for more details about our stereo depth loss." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 133, + 546, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 546, + 301 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 546, + 301 + ], + "type": "text", + "content": "Due to the different characteristics of datasets such as baseline length, " + }, + { + "bbox": [ + 304, + 133, + 546, + 301 + ], + "type": "inline_equation", + "content": "d_{s}" + }, + { + "bbox": [ + 304, + 133, + 546, + 301 + ], + "type": "text", + "content": " may have estimation error, which leads to an error in the depth-guided plane-sweeping. To tackle this, we partially train the matching and propagation network in the stereo estimation network (Fig. 4) using " + }, + { + "bbox": [ + 304, + 133, + 546, + 301 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_d^{self}" + }, + { + "bbox": [ + 304, + 133, + 546, + 301 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 133, + 546, + 301 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_d^{self}" + }, + { + "bbox": [ + 304, + 133, + 546, + 301 + ], + "type": "text", + "content": " provides multi-view supervision to the stereo estimation network by leveraging " + }, + { + "bbox": [ + 304, + 133, + 546, + 301 + ], + "type": "inline_equation", + "content": "d_{r}" + }, + { + "bbox": [ + 304, + 133, + 546, + 301 + ], + "type": "text", + "content": " estimated from multi-view images. However, since " + }, + { + "bbox": [ + 304, + 133, + 546, + 301 + ], + "type": "inline_equation", + "content": "d_{r}" + }, + { + "bbox": [ + 304, + 133, + 546, + 301 + ], + "type": "text", + "content": " may also have estimation error, we further regularize the stereo estimation network using " + }, + { + "bbox": [ + 304, + 133, + 546, + 301 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_d^s" + }, + { + "bbox": [ + 304, + 133, + 546, + 301 + ], + "type": "text", + "content": ". This training scheme for the stereo estimation network ensures consistent depth estimation, while preserving the stereo matching capability. Sec. D.1 in the supplementary document further discusses the training scheme of the stereo estimation network." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 314, + 419, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 314, + 419, + 327 + ], + "spans": [ + { + "bbox": [ + 305, + 314, + 419, + 327 + ], + "type": "text", + "content": "4. StereoNVS Dataset" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 335, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 545, + 407 + ], + "type": "text", + "content": "We propose the StereoNVS dataset, the first stereo-camera image dataset for training and evaluating novel-view synthesis using stereo-camera images. The StereoNVS dataset provides both real and synthetic datasets, each of which is dubbed StereoNVS-Real and StereoNVS-Synthetic. In the following, we present the details of each dataset." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 415, + 405, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 415, + 405, + 426 + ], + "spans": [ + { + "bbox": [ + 305, + 415, + 405, + 426 + ], + "type": "text", + "content": "4.1. StereoNVS-Real" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 434, + 545, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 545, + 529 + ], + "type": "text", + "content": "StereoNVS-Real provides real-world stereo-camera images for the training and evaluation of novel-view synthesis. The dataset provides stereo-camera images of 53 static scenes, and around 25 stereo-image pairs per scene. The images are undistorted and stereo-rectified, and have a resolution of " + }, + { + "bbox": [ + 304, + 434, + 545, + 529 + ], + "type": "inline_equation", + "content": "1792 \\times 896" + }, + { + "bbox": [ + 304, + 434, + 545, + 529 + ], + "type": "text", + "content": ". The camera parameters such as the camera poses are provided as well. In our experiments, we divide the dataset into 45 and 8 scenes as training and test sets." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 530, + 545, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 530, + 545, + 638 + ], + "spans": [ + { + "bbox": [ + 304, + 530, + 545, + 638 + ], + "type": "text", + "content": "To capture stereo images, we built a camera rig with two Basler machine vision cameras. We measured the camera parameters including the intrinsic and distortion parameters by camera calibration using multi-view images of a checkerboard [42]. Then, we collected stereo images from multiple viewpoints for various indoor and bounded scenes using our camera system. The captured images were then undistorted and stereo-rectified. Finally, we obtained the camera poses of the captured images using COLMAP [28]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 647, + 428, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 647, + 428, + 659 + ], + "spans": [ + { + "bbox": [ + 305, + 647, + 428, + 659 + ], + "type": "text", + "content": "4.2. StereoNVS-Synthetic" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "content": "For quantitative evaluation of synthesized geometries, we also present the StereoNVS-Synthetic dataset, constructed by rendering synthetic 3D models using the 3D-Front dataset [10]. The images have a resolution of " + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "inline_equation", + "content": "864 \\times 448" + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4943" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 70, + 545, + 459 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 459 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 459 + ], + "type": "image", + "image_path": "c1e53e4bba21ebda7abbf50d1f1c699b97a7f022312bf20229397601f804d6fc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 460, + 546, + 495 + ], + "lines": [ + { + "bbox": [ + 46, + 460, + 546, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 546, + 495 + ], + "type": "text", + "content": "Figure 7. Qualitative comparison of novel-view synthesis on the StereoNVS dataset, showing rendering results for two real-world scenes (above) and one synthetic scene (below). All models are trained using stereo images. Our method outperforms the baseline methods [16, 21, 32, 33] on both real-world and synthetic scenes, especially in thin structures and textureless regions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 504, + 288, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 504, + 288, + 576 + ], + "spans": [ + { + "bbox": [ + 46, + 504, + 288, + 576 + ], + "type": "text", + "content": "which is half of the real dataset. StereoNVS-Synthetic provides stereo-camera images of 50 scenes, and around 150 stereo image pairs per scene as well as the ground-truth camera parameters, camera poses, and depth maps. For more details about the StereoNVS dataset, refer to Sec. E in the supplementary document." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 594, + 128, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 594, + 128, + 608 + ], + "spans": [ + { + "bbox": [ + 47, + 594, + 128, + 608 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": "We conduct extensive validation of our method on the StereoNVS dataset. In the following, we will provide implementation details of our method (Sec. 5.1), compare our method with other baselines (Sec. 5.2), and conduct a comprehensive analysis of our proposed components, which proves the effectiveness of our framework including the stereo feature extractor, the depth-guided plane-sweeping, and the stereo depth loss (Sec. 5.3)." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 310, + 502, + 544, + 576 + ], + "blocks": [ + { + "bbox": [ + 310, + 502, + 544, + 576 + ], + "lines": [ + { + "bbox": [ + 310, + 502, + 544, + 576 + ], + "spans": [ + { + "bbox": [ + 310, + 502, + 544, + 576 + ], + "type": "table", + "html": "
MethodStereoNVS-RealStereoNVS-Synthetic
PSNR(↑)SSIM(↑)LPIPS(↓)ABS(↓)PSNR(↑)SSIM(↑)LPIPS(↓)ABS(↓)
SRF21.120.69330.41642.917622.360.71620.42450.8125
IBRNet26.120.84210.20950.756630.830.88890.18230.2628
GeoNeRF28.010.89290.14600.506432.130.91790.14380.1577
GNT26.080.84340.22851.095926.170.84060.26500.4512
GeoNeRF+D25.650.81720.20820.869332.850.93210.11710.0782
NeuRay26.510.85380.18870.614732.260.91040.14780.1571
Ours28.440.90000.13960.486833.450.93360.12030.1056
", + "image_path": "399e1227d9b7ac43944ee9d038f726b0341fe8f159873f9c2b34d0a3f90779c4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 578, + 545, + 622 + ], + "lines": [ + { + "bbox": [ + 305, + 578, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 305, + 578, + 545, + 622 + ], + "type": "text", + "content": "Table 1. Quantitative comparison between the baseline methods [4, 16, 21, 32, 33] and ours. All models are trained using stereo images. Our method shows superior performance on both StereoNVS-Real and StereoNVS-Synthetic datasets." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 634, + 439, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 634, + 439, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 634, + 439, + 647 + ], + "type": "text", + "content": "5.1. Implementation Details" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "type": "text", + "content": "We train our model on the training set of StereoNVS-Real, which has multi-view stereo image pairs of real scenes. Our model is trained for 250K iterations. For each iteration, we randomly select one scene and one target viewpoint of the scene. For both training and evaluation, image" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4944" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": "features and feature volumes are extracted from three stereo image pairs (i.e., total six images) at three viewpoints nearest to the target viewpoint. During training, 512 rays are randomly selected for the training batch. We use the Adam optimizer [18] with learning rates of 0.0005 and the cosine annealing scheduling [22]. We employ UniMatch [35] for both the pre-trained stereo estimation network within the stereo feature extractor and the pseudo-ground-truth depth of the stereo depth loss. Refer to Sec. B in the supplementary document for additional implementation details." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 205, + 128, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 205, + 128, + 217 + ], + "spans": [ + { + "bbox": [ + 47, + 205, + 128, + 217 + ], + "type": "text", + "content": "5.2. Comparison" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 224, + 287, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 224, + 287, + 345 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 287, + 345 + ], + "type": "text", + "content": "We compare our method with recent generalizable novel view synthesis methods: SRF [4], IBRNet [33], GeoNeRF [16], GNT [32] and NeuRay [21]. Like our method, we use three stereo-camera image pairs to synthesize each target-view image for all the baseline methods. While the previous methods do not explicitly assume stereo-camera images as their inputs, we also train them using stereocamera images as they can handle stereo-camera images as independent inputs. All the baseline models are trained on the training set of StereoNVS-Real, as done for our method." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 346, + 287, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 346, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 287, + 453 + ], + "type": "text", + "content": "We evaluate both image and depth qualities on the StereoNVS-Real and the StereoNVS-Synthetic datasets. For StereoNVS-Real, we utilize pseudo-ground-truth depths obtained from COLMAP [28] to assess depth quality. On the other hand, for StereoNVS-Synthetic, we use rendered depth maps as ground truths for depth quality assessment. We employ PSNR, SSIM [34], and LPIPS [41] as metrics for image quality and absolute error (ABS) as a metric for depth quality." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 456, + 288, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 288, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 288, + 529 + ], + "type": "text", + "content": "Among the compared methods, for training and inference, NeuRay [21] requires depth maps and GeoNeRF [16] can use depth maps as additional inputs. We denote such variation of GeoNeRF as " + }, + { + "bbox": [ + 46, + 456, + 288, + 529 + ], + "type": "inline_equation", + "content": "\\mathrm{GeoNeRF}_{+D}" + }, + { + "bbox": [ + 46, + 456, + 288, + 529 + ], + "type": "text", + "content": ". For their training and inference, we used the pseudo-ground-truth depth maps estimated by UniMatch [35] as done for our method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": "Qualitative comparison. Fig. 7 presents a qualitative comparison on novel view synthesis between our method and previous methods [16, 21, 32, 33] using the StereoNVS dataset. Previous methods estimate inaccurate geometry for scenes with thin structures or textureless regions, resulting in severe artifacts in the synthesized novel view images. In contrast, our method clearly outperforms the baseline methods in view synthesis results with significantly fewer artifacts even in textureless regions, thanks to more accurately estimated geometry. Moreover, our method shows better synthesis results compared to " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{GeoNeRF}_{+D}" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": " [16] and Neu-Ray [21], even though they explicitly use depth maps in the inference time, demonstrating the robust utilization of depth maps in our framework." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 309, + 71, + 545, + 131 + ], + "blocks": [ + { + "bbox": [ + 309, + 71, + 545, + 131 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 545, + 131 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 545, + 131 + ], + "type": "table", + "html": "
StereoNVS-RealStereoNVS-Synthetic
PSNR(↑)SSIM(↑)LPIPS(↓)ABS (↓)PSNR(↑)SSIM(↑)LPIPS(↓)ABS (↓)
Baseline (a)28.050.89530.13720.559832.220.91720.13860.1679
+ Stereo setting (b)28.010.89290.14600.506432.130.91790.14380.1577
+ SAM (c)28.210.89670.13980.493531.900.91670.13860.1531
+ Correlated feature (d)28.310.89880.13700.505732.350.92450.12820.1416
+ DGPS (e)28.420.89970.14030.509833.090.93040.12300.1246
+ Stereo depth loss (f)28.440.90000.13960.486833.450.93360.12030.1056
", + "image_path": "f121cc7082c298ad825d0effd0d2bc42fcf4993b9523d7c9d436d25dabd80d1f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 359, + 133, + 492, + 143 + ], + "lines": [ + { + "bbox": [ + 359, + 133, + 492, + 143 + ], + "spans": [ + { + "bbox": [ + 359, + 133, + 492, + 143 + ], + "type": "text", + "content": "Table 2. Quantitative ablation study." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 307, + 156, + 366, + 249 + ], + "blocks": [ + { + "bbox": [ + 318, + 145, + 356, + 156 + ], + "lines": [ + { + "bbox": [ + 318, + 145, + 356, + 156 + ], + "spans": [ + { + "bbox": [ + 318, + 145, + 356, + 156 + ], + "type": "text", + "content": "Tab. 2 (b)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 156, + 366, + 249 + ], + "lines": [ + { + "bbox": [ + 307, + 156, + 366, + 249 + ], + "spans": [ + { + "bbox": [ + 307, + 156, + 366, + 249 + ], + "type": "image", + "image_path": "d74b88b8db2a0db5d67717e6819349684bd5e3eb740e0cdb82156df526d29f4e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 254, + 545, + 277 + ], + "lines": [ + { + "bbox": [ + 305, + 254, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 305, + 254, + 545, + 277 + ], + "type": "text", + "content": "Figure 8. Qualitative ablation study. Our full method enables accurate synthesis of a novel-view image and a depth map." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 367, + 156, + 425, + 248 + ], + "blocks": [ + { + "bbox": [ + 378, + 145, + 416, + 156 + ], + "lines": [ + { + "bbox": [ + 378, + 145, + 416, + 156 + ], + "spans": [ + { + "bbox": [ + 378, + 145, + 416, + 156 + ], + "type": "text", + "content": "Tab. 2 (d)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 367, + 156, + 425, + 248 + ], + "lines": [ + { + "bbox": [ + 367, + 156, + 425, + 248 + ], + "spans": [ + { + "bbox": [ + 367, + 156, + 425, + 248 + ], + "type": "image", + "image_path": "78f95b3819d8497d6f718295c0647b5e8e96701275ae00ad2ef409f567e776d7.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 426, + 156, + 485, + 248 + ], + "blocks": [ + { + "bbox": [ + 437, + 145, + 475, + 156 + ], + "lines": [ + { + "bbox": [ + 437, + 145, + 475, + 156 + ], + "spans": [ + { + "bbox": [ + 437, + 145, + 475, + 156 + ], + "type": "text", + "content": "Tab. 2 (e)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 426, + 156, + 485, + 248 + ], + "lines": [ + { + "bbox": [ + 426, + 156, + 485, + 248 + ], + "spans": [ + { + "bbox": [ + 426, + 156, + 485, + 248 + ], + "type": "image", + "image_path": "2c15cedb309dbd6eff3912afe3f3d1e5c2cf8f2eecf67b8367c4dc6d4e975879.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 485, + 156, + 545, + 248 + ], + "blocks": [ + { + "bbox": [ + 497, + 145, + 533, + 156 + ], + "lines": [ + { + "bbox": [ + 497, + 145, + 533, + 156 + ], + "spans": [ + { + "bbox": [ + 497, + 145, + 533, + 156 + ], + "type": "text", + "content": "Tab. 2 (f)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 485, + 156, + 545, + 248 + ], + "lines": [ + { + "bbox": [ + 485, + 156, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 485, + 156, + 545, + 248 + ], + "type": "image", + "image_path": "7f90db9a239d98b1ea3ec45d42e90b4b8594594817eb1c000e8ec7bf7cfd8c3b.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 285, + 546, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 285, + 546, + 490 + ], + "spans": [ + { + "bbox": [ + 304, + 285, + 546, + 490 + ], + "type": "text", + "content": "Quantitative comparison. Tab. 1 shows that our method generally surpasses other baseline methods [4, 16, 21, 32, 33] in terms of image and depth qualities. GNT [32] exhibits degraded performances, likely due to its data-hungry transformer backbone. While performance differences with GeoNeRF are not substantial on the StereoNVS-Real dataset, our method shows the superior perceptual quality (Fig. 7) and significantly better performance on the StereoNVS-Synthetic dataset. While " + }, + { + "bbox": [ + 304, + 285, + 546, + 490 + ], + "type": "inline_equation", + "content": "\\mathrm{GeoNeRF}_{+D}" + }, + { + "bbox": [ + 304, + 285, + 546, + 490 + ], + "type": "text", + "content": " [16] and NeuRay [21] achieve comparable performances on the StereoNVS-Synthetic dataset, they show considerably degenerated results on the StereoNVS-Real dataset. In contrast, our method demonstrates superior performance thanks to the robustness of our framework. Sec. D.1 in the supplementary document further discusses the sensitivity of " + }, + { + "bbox": [ + 304, + 285, + 546, + 490 + ], + "type": "inline_equation", + "content": "\\mathrm{GeoNeRF}_{+D}" + }, + { + "bbox": [ + 304, + 285, + 546, + 490 + ], + "type": "text", + "content": ", NeuRay, and our method to depth errors in real-world images." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 497, + 440, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 497, + 440, + 509 + ], + "spans": [ + { + "bbox": [ + 306, + 497, + 440, + 509 + ], + "type": "text", + "content": "5.3. Analysis and Discussion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 515, + 404, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 515, + 404, + 528 + ], + "spans": [ + { + "bbox": [ + 306, + 515, + 404, + 528 + ], + "type": "text", + "content": "5.3.1 Ablation Study" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 534, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 641 + ], + "type": "text", + "content": "To assess the impact of our proposed methods, we conduct an ablation study starting with our baseline model [16], which is trained on three views in a monocular setting (i.e., three images). Although equipped with a stereo camera setting, the baseline model trained on three stereo views (i.e., six images) shows similar image qualities as shown in Tab. 2 (b). This result indicates that sophisticated methods are needed to exploit the invaluable information from stereo-camera images." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 642, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 546, + 713 + ], + "type": "text", + "content": "Tab. 2 shows that our model consistently demonstrates improvements in view synthesis results, as we introduce our proposed components. Our stereo feature extractor enables us to estimate better geometry, leading to fewer artifacts in novel view synthesis, as shown in Fig. 8 (d). The stereo attention module (c) and the stereo-correlated features (d)" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4945" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "help extract robust stereo image features, which are particularly effective for real scenes and synthetic scenes, respectively. Our DGPS is essential for better geometry estimation, leading to significantly improved results as shown in Fig. 8 (e), especially in textureless regions. This is further evident in Tab. 2 (e), with considerable performance gain on the StereoNVS-Synthetic dataset. Our final model shows the best performances (Tab. 2 (f)), with high-quality depth and view synthesis results (Fig. 8 (f))." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 188, + 281, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 188, + 281, + 202 + ], + "spans": [ + { + "bbox": [ + 47, + 188, + 281, + 202 + ], + "type": "text", + "content": "5.3.2 Effectiveness of Depth-Guided Plane-Sweeping" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 207, + 287, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 207, + 287, + 290 + ], + "spans": [ + { + "bbox": [ + 46, + 207, + 287, + 290 + ], + "type": "text", + "content": "We conduct two experiments to demonstrate the effectiveness of our DGPS. In the first experiment, we compare our final model against a model without using DGPS. The \"Final model\" in Tab. 3 denotes our final model with DGPS. As shown in Tab. 3, the model without using DGPS shows degenerated results compared to our final model, highlighting the effectiveness of DGPS." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 291, + 289, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 291, + 289, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 291, + 289, + 483 + ], + "type": "text", + "content": "In the second experiment, we compare our model using DGPS against a model using more depth planes for cost volume construction. The term \"Base model\" in Tab. 4 refers to our model that is solely equipped with the stereo feature extractor, excluding DGPS and the stereo depth loss. We train an additional model that uses approximately 1.5 times the number of depth planes, compared to the base model. Although this additional model utilizes more depth planes, it shows similar results to the base model on StereoNeRF-Synthetic, as reported in Tab. 4. This is because increasing the number of planes does not guarantee accurate correspondence matching, especially in textureless regions. In contrast, our DGPS guarantees correspondence matching across multi-view image features near the geometry, resulting in better image and depth qualities. Refer to Sec. C.2.1 in the supplementary document for more details." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 491, + 261, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 491, + 261, + 503 + ], + "spans": [ + { + "bbox": [ + 47, + 491, + 261, + 503 + ], + "type": "text", + "content": "5.3.3 Benefit of Stereo Estimation in Depth Loss" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 510, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 617 + ], + "type": "text", + "content": "To show the effectiveness of using stereo estimation networks for depth supervision, we conduct an additional experiment as follows. First, we obtain two pseudo-GT depths: one from the pre-trained stereo network [35] (" + }, + { + "bbox": [ + 46, + 510, + 287, + 617 + ], + "type": "inline_equation", + "content": "d_{gt}" + }, + { + "bbox": [ + 46, + 510, + 287, + 617 + ], + "type": "text", + "content": ") as stated in Sec. 3.4 and the other from the state-of-the-art learning-based MVS network (" + }, + { + "bbox": [ + 46, + 510, + 287, + 617 + ], + "type": "inline_equation", + "content": "d_{gt}^{mvs}" + }, + { + "bbox": [ + 46, + 510, + 287, + 617 + ], + "type": "text", + "content": ") [26]. Then, we train two models with our methods using " + }, + { + "bbox": [ + 46, + 510, + 287, + 617 + ], + "type": "inline_equation", + "content": "d_{gt}" + }, + { + "bbox": [ + 46, + 510, + 287, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 510, + 287, + 617 + ], + "type": "inline_equation", + "content": "d_{gt}^{mvs}" + }, + { + "bbox": [ + 46, + 510, + 287, + 617 + ], + "type": "text", + "content": ", respectively. Then, we compare their synthesis results based on image and shape qualities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "content": "As shown in Tab. 5, our model trained with " + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "inline_equation", + "content": "d_{gt}" + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "content": " surpasses the other model using " + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "inline_equation", + "content": "d_{gt}^{mvs}" + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "content": " on StereoNVS-Synthetic. Note that the MVS network takes more images than the stereo network, seven images and two images, respectively. These results show that the stereo network provides a reliable depth signal for high-quality view synthesis thanks to its generalization ability, which came from standardized inputs of stereo-camera images and large-scale stereo datasets." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 310, + 71, + 545, + 112 + ], + "blocks": [ + { + "bbox": [ + 310, + 71, + 545, + 112 + ], + "lines": [ + { + "bbox": [ + 310, + 71, + 545, + 112 + ], + "spans": [ + { + "bbox": [ + 310, + 71, + 545, + 112 + ], + "type": "table", + "html": "
PSNR(↑)SSIM(↑)LPIPS(↓)ABS(↓)
Final model (Tab. 2 (f))33.450.93360.12030.1056
- DGPS32.300.92400.12900.1300
", + "image_path": "ed7b766af2609f4b8bf2012478854bbd0bf62013eb7105299867e019f994a389.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 310, + 142, + 545, + 194 + ], + "blocks": [ + { + "bbox": [ + 306, + 114, + 545, + 137 + ], + "lines": [ + { + "bbox": [ + 306, + 114, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 306, + 114, + 545, + 137 + ], + "type": "text", + "content": "Table 3. Effectiveness of our depth-guided plane-sweeping (DGPS) compared to the baseline models without using DGPS." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 142, + 545, + 194 + ], + "lines": [ + { + "bbox": [ + 310, + 142, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 310, + 142, + 545, + 194 + ], + "type": "table", + "html": "
PSNR(↑)SSIM(↑)LPIPS(↓)ABS(↓)
Base model (Tab. 2 (d))32.350.92450.12820.1416
+ more depth planes32.330.92470.12990.1470
+ DGPS (Tab. 2 (e))33.090.93040.12300.1246
", + "image_path": "06185e3889705915fc254b143f91fd5e79bf2069a9dcfbd1cdb69005f51d5cd0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 310, + 225, + 545, + 264 + ], + "blocks": [ + { + "bbox": [ + 306, + 198, + 545, + 220 + ], + "lines": [ + { + "bbox": [ + 306, + 198, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 306, + 198, + 545, + 220 + ], + "type": "text", + "content": "Table 4. Efficiency of our depth-guided plane-sweeping (DGPS) compared to the baseline models using more depth planes." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 225, + 545, + 264 + ], + "lines": [ + { + "bbox": [ + 310, + 225, + 545, + 264 + ], + "spans": [ + { + "bbox": [ + 310, + 225, + 545, + 264 + ], + "type": "table", + "html": "
PSNR (↑)SSIM (↑)LPIPS (↓)ABS (↓)
Ours w/ MVS depth (dmvs)32.730.92770.12530.1216
Ours w/ Stereo depth (dtg)33.450.93360.12030.1056
", + "image_path": "adc8222181961adea9c874169718b55f7af5425759e1a8573c4a5665e4f3f027.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 266, + 545, + 289 + ], + "lines": [ + { + "bbox": [ + 305, + 266, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 305, + 266, + 545, + 289 + ], + "type": "text", + "content": "Table 5. Effectiveness of using stereo depths as pseudo-ground truth for depth loss compared to using MVS depths." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 297, + 378, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 297, + 378, + 309 + ], + "spans": [ + { + "bbox": [ + 306, + 297, + 378, + 309 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 317, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 317, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 317, + 545, + 510 + ], + "type": "text", + "content": "This paper proposes StereoNeRF, a novel generalizable view synthesis framework leveraging stereo-camera images for high-quality novel-view synthesis. Due to the ill-posedness, previous methods struggle with accurate geometry estimation, which leads to severe artifacts in novel-view synthesis. Since the stereo matching provides vital information for accurate geometry reconstruction, our framework incorporates the stereo matching into NeRF-based generalizable view synthesis approach. To this end, we introduce a stereo feature extractor, a depth-guided plane-sweeping, and a stereo depth loss. We also present the StereoNVS dataset, the first stereo-camera image dataset for training and evaluating novel-view synthesis. Our extensive experiments show that StereoNeRF is effective in generalizable novel-view synthesis, particularly in scenes with complex structures or textureless regions." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 514, + 545, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 514, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 304, + 514, + 545, + 611 + ], + "type": "text", + "content": "Limitations and Future Work. Our method is not free from limitations. In sparse view settings, our method produces blurry images and inaccurate geometry, issues also present in other methods. This limitation arises from the insufficient information for novel viewpoints in the sparse view settings. Our future work will involve additional geometric or generative prior, along with stereo prior, to compensate for the lack of information in such settings." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 613, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 613, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 613, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgement. We thank Woohyeok Kim and Hyeongmin Lee for their assistance in acquisition of the StereoNVS-Real dataset. This work was supported by the NRF grant (No.2023R1A2C200494611, 2022R1A6A1A03052954, RS-2023-00211658) and IITP grant (No.2019-0-01906, Artificial Intelligence Graduate School Program (POSTECH)) funded by the Korea government (MSIT) and Samsung Electronics Co., Ltd." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4946" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "type": "text", + "content": "[1] Jia-Ren Chang and Yong-Sheng Chen. Pyramid stereo matching network. In CVPR, pages 5410-5418, 2018. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 114, + 286, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 286, + 158 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 286, + 158 + ], + "type": "text", + "content": "[2] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In ICCV, 2021. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 159, + 286, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 159, + 286, + 204 + ], + "spans": [ + { + "bbox": [ + 53, + 159, + 286, + 204 + ], + "type": "text", + "content": "[3] Shuo Cheng, Zexiang Xu, Shilin Zhu, Zhuwen Li, Li Erran Li, Ravi Ramamoorthi, and Hao Su. Deep stereo using adaptive thin volume representation with uncertainty awareness. In CVPR, pages 2524-2534, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 205, + 286, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 286, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 286, + 247 + ], + "type": "text", + "content": "[4] Julian Chibane, Aayush Bansal, Verica Lazova, and Gerard Pons-Moll. Stereo radiance fields (srf): Learning view synthesis for sparse views of novel scenes. In CVPR. 1, 2, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 250, + 286, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 250, + 286, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 250, + 286, + 282 + ], + "type": "text", + "content": "[5] Xiaojie Chu, Liangyu Chen, and Wenqing Yu. Nafssr: Stereo image super-resolution using nafnet. In CVPR, pages 1239-1248, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 285, + 286, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 285, + 286, + 339 + ], + "spans": [ + { + "bbox": [ + 53, + 285, + 286, + 339 + ], + "type": "text", + "content": "[6] Paul E Debevec, Camillo J Taylor, and Jitendra Malik. Modeling and rendering architecture from photographs: A hybrid geometry-and image-based approach. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 11-20, 1996. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 341, + 286, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 341, + 286, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 341, + 286, + 407 + ], + "type": "text", + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 408, + 286, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 408, + 286, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 408, + 286, + 441 + ], + "type": "text", + "content": "[8] Yilun Du, Cameron Smith, Ayush Tewari, and Vincent Sitzmann. Learning to render novel views from wide-baseline stereo pairs. arXiv preprint arXiv:2304.08463, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 443, + 286, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 443, + 286, + 485 + ], + "spans": [ + { + "bbox": [ + 53, + 443, + 286, + 485 + ], + "type": "text", + "content": "[9] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, pages 5501–5510, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 487, + 286, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 487, + 286, + 552 + ], + "spans": [ + { + "bbox": [ + 48, + 487, + 286, + 552 + ], + "type": "text", + "content": "[10] Huan Fu, Bowen Cai, Lin Gao, Ling-Xiao Zhang, Jiaming Wang, Cao Li, Qixun Zeng, Chengyue Sun, Rongfei Jia, Binqiang Zhao, et al. 3d-front: 3d furnished rooms with layouts and semantics. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10933-10942, 2021. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 555, + 286, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 286, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 286, + 588 + ], + "type": "text", + "content": "[11] Yasutaka Furukawa, Carlos Hernández, et al. Multi-view stereo: A tutorial. Foundations and Trends® in Computer Graphics and Vision, 9(1-2):1-148, 2015. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 590, + 286, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 286, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 286, + 622 + ], + "type": "text", + "content": "[12] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, pages 3354-3361. IEEE, 2012. 1, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 624, + 286, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 286, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 286, + 667 + ], + "type": "text", + "content": "[13] Steven J Gortler, Radek Grzeszczuk, Richard Szeliski, and Michael F Cohen. The lumigraph. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 43-54, 1996. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "type": "text", + "content": "[14] Xiaodong Gu, Zhiwen Fan, Siyu Zhu, Zuozhuo Dai, Feitong Tan, and Ping Tan. Cascade cost volume for high-resolution multi-view stereo and stereo matching. In CVPR, pages 2495-2504, 2020. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "text", + "content": "[15] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "type": "text", + "content": "[16] Mohammad Mahdi Johari, Yann Lepoittevin, and François Fleuret. Geonerf: Generalizing nerf with geometry priors. In CVPR, pages 18365-18375, 2022. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 142, + 545, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 142, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 307, + 142, + 545, + 175 + ], + "type": "text", + "content": "[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM TOG, 42(4):1-14, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 177, + 545, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 177, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 307, + 177, + 545, + 208 + ], + "type": "text", + "content": "[18] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 211, + 545, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 211, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 307, + 211, + 545, + 243 + ], + "type": "text", + "content": "[19] Kiriakos N Kutulakos and Steven M Seitz. A theory of shape by space carving. International journal of computer vision, 38:199-218, 2000. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 246, + 545, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 246, + 545, + 278 + ], + "spans": [ + { + "bbox": [ + 307, + 246, + 545, + 278 + ], + "type": "text", + "content": "[20] Marc Levoy and Pat Hanrahan. Light field rendering. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 31-42, 1996. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 280, + 545, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 280, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 307, + 280, + 545, + 323 + ], + "type": "text", + "content": "[21] Yuan Liu, Sida Peng, Lingjie Liu, Qianqian Wang, Peng Wang, Christian Theobalt, Xiaowei Zhou, and Wenping Wang. Neural rays for occlusion-aware image-based rendering. In CVPR, pages 7824-7833, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 326, + 545, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 326, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 307, + 326, + 545, + 357 + ], + "type": "text", + "content": "[22] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 360, + 545, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 360, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 307, + 360, + 545, + 414 + ], + "type": "text", + "content": "[23] Nikolaus Mayer, Eddy Ilg, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In CVPR, pages 4040-4048, 2016. 1, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 416, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 416, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 307, + 416, + 545, + 459 + ], + "type": "text", + "content": "[24] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2, 3, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 462, + 545, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 462, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 307, + 462, + 545, + 495 + ], + "type": "text", + "content": "[25] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM TOG, 41(4):1-15, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 497, + 545, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 497, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 307, + 497, + 545, + 540 + ], + "type": "text", + "content": "[26] Rui Peng, Rongjie Wang, Zhenyu Wang, Yawen Lai, and Ronggang Wang. Rethinking depth estimation for multiview stereo: A unified representation. In CVPR, pages 8645-8654, 2022. 2, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 543, + 545, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 543, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 307, + 543, + 545, + 574 + ], + "type": "text", + "content": "[27] Daniel Scharstein and Richard Szeliski. A taxonomy and evaluation of dense two-frame stereo correspondence algorithms. IJCV, 47:7-42, 2002. 1, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 577, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 577, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 307, + 577, + 545, + 609 + ], + "type": "text", + "content": "[28] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, pages 4104-4113, 2016. 4, 5, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 611, + 545, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 611, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 307, + 611, + 545, + 643 + ], + "type": "text", + "content": "[29] Steven M Seitz and Charles R Dyer. Photorealistic scene reconstruction by voxel coloring. International Journal of Computer Vision, 35:151-173, 1999. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 646, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 646, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 646, + 545, + 678 + ], + "type": "text", + "content": "[30] Mohammed Suhail, Carlos Esteves, Leonid Sigal, and Ameesh Makadia. Generalizable patch-based neural rendering. In ECCV, pages 156-174. Springer, 2022. 1, 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 681, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 545, + 712 + ], + "type": "text", + "content": "[31] Alex Trevithick and Bo Yang. Grf: Learning a general radiance field for 3d representation and rendering. In ICCV, pages 15182-15192, 2021. 2" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4947" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 511 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 105 + ], + "type": "text", + "content": "[32] Mukund Varma, Peihao Wang, Xuxi Chen, Tianlong Chen, Subhashini Venugopalan, and Zhangyang Wang. Is attention all that nef needs? In ICLR, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 286, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 286, + 162 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 286, + 162 + ], + "type": "text", + "content": "[33] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In CVPR, pages 4690-4699, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 286, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 286, + 195 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 286, + 195 + ], + "type": "text", + "content": "[34] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 13(4):600-612, 2004. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 198, + 286, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 198, + 286, + 239 + ], + "spans": [ + { + "bbox": [ + 49, + 198, + 286, + 239 + ], + "type": "text", + "content": "[35] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, Fisher Yu, Dacheng Tao, and Andreas Geiger. Unifying flow, stereo and depth estimation. IEEE TPAMI, 2023. 2, 3, 5, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 242, + 286, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 242, + 286, + 274 + ], + "spans": [ + { + "bbox": [ + 49, + 242, + 286, + 274 + ], + "type": "text", + "content": "[36] Jiayu Yang, Wei Mao, Jose M Alvarez, and Miaomiao Liu. Cost volume pyramid based depth inference for multi-view stereo. In CVPR, pages 4877-4886, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 276, + 286, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 276, + 286, + 308 + ], + "spans": [ + { + "bbox": [ + 49, + 276, + 286, + 308 + ], + "type": "text", + "content": "[37] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoptrees for real-time rendering of neural radiance fields. In ICCV, pages 5752-5761, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 310, + 286, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 310, + 286, + 342 + ], + "spans": [ + { + "bbox": [ + 49, + 310, + 286, + 342 + ], + "type": "text", + "content": "[38] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In CVPR, pages 4578-4587, 2021. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 344, + 286, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 344, + 286, + 376 + ], + "spans": [ + { + "bbox": [ + 49, + 344, + 286, + 376 + ], + "type": "text", + "content": "[39] Feihu Zhang, Victor Prisacariu, Ruigang Yang, and Philip HS Torr. Ga-net: Guided aggregation net for end-to-end stereo matching. In CVPR, pages 185-194, 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 378, + 286, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 378, + 286, + 409 + ], + "spans": [ + { + "bbox": [ + 49, + 378, + 286, + 409 + ], + "type": "text", + "content": "[40] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020.5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 411, + 286, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 411, + 286, + 453 + ], + "spans": [ + { + "bbox": [ + 49, + 411, + 286, + 453 + ], + "type": "text", + "content": "[41] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 456, + 286, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 456, + 286, + 477 + ], + "spans": [ + { + "bbox": [ + 49, + 456, + 286, + 477 + ], + "type": "text", + "content": "[42] Zhengyou Zhang. A flexible new technique for camera calibration. IEEE TPAMI, 22(11):1330-1334, 2000. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 479, + 286, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 479, + 286, + 511 + ], + "spans": [ + { + "bbox": [ + 49, + 479, + 286, + 511 + ], + "type": "text", + "content": "[43] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. ACM TOG, 2018. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4948" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Generalizable Whole Slide Image Classification with Fine-Grained Visual-Semantic Interaction/c4c5691b-0982-48b4-b41b-bc14bcf8455f_content_list.json b/2024/Generalizable Whole Slide Image Classification with Fine-Grained Visual-Semantic Interaction/c4c5691b-0982-48b4-b41b-bc14bcf8455f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5b64b32f73c22807b963fe6747f7d57ac1f4d341 --- /dev/null +++ b/2024/Generalizable Whole Slide Image Classification with Fine-Grained Visual-Semantic Interaction/c4c5691b-0982-48b4-b41b-bc14bcf8455f_content_list.json @@ -0,0 +1,1659 @@ +[ + { + "type": "text", + "text": "Generalizable Whole Slide Image Classification with Fine-Grained Visual-Semantic Interaction", + "text_level": 1, + "bbox": [ + 148, + 130, + 818, + 172 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hao Li", + "bbox": [ + 243, + 202, + 313, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wenxian Yang3", + "bbox": [ + 196, + 220, + 328, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ying Chen", + "bbox": [ + 359, + 203, + 457, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Liansheng Wang", + "bbox": [ + 359, + 220, + 498, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yifei Chen", + "bbox": [ + 506, + 203, + 604, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bowen Ding", + "bbox": [ + 540, + 220, + 651, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rongshan $\\mathbf{Y}\\mathbf{u}^{1*}$", + "bbox": [ + 651, + 203, + 779, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuchen Han\\*", + "bbox": [ + 686, + 220, + 805, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Informatics, Xiamen University $^{2}$ Huawei $^{3}$ Aginome Scientific, Xiamen, China", + "bbox": [ + 112, + 239, + 883, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ Department of Pathology, Shanghai Chest Hospital, Shanghai Jiao Tong University School of Medicine", + "bbox": [ + 89, + 256, + 906, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{111ih, cying2023}@stu.xmu.edu.cn, {rsyu, 1swang}@xmu.edu.cn,", + "bbox": [ + 228, + 276, + 761, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "chenyifei14@huawei.com,wx@aginome.com，dingbowenmail@126.com，ychan@cmu.edu.cn", + "bbox": [ + 145, + 295, + 843, + 308 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 344, + 312, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Whole Slide Image (WSI) classification is often formulated as a Multiple Instance Learning (MIL) problem. Recently, Vision-Language Models (VLMs) have demonstrated remarkable performance in WSI classification. However, existing methods leverage coarse-grained pathogenetic descriptions for visual representation supervision, which are insufficient to capture the complex visual appearance of pathogenetic images, hindering the generalizability of models on diverse downstream tasks. Additionally, processing high-resolution WSIs can be computationally expensive. In this paper, we propose a novel \"Fine-grained Visual-Semantic Interaction\" (FiVE) framework for WSI classification. It is designed to enhance the model's generalizability by leveraging the interaction between localized visual patterns and fine-grained pathological semantics. Specifically, with meticulously designed queries, we start by utilizing a large language model to extract fine-grained pathological descriptions from various non-standardized raw reports. The output descriptions are then reconstructed into fine-grained labels used for training. By introducing a Task-specific Fine-grained Semantics (TFS) module, we enable prompts to capture crucial visual information in WSIs, which enhances representation learning and augments generalization capabilities significantly. Furthermore, given that pathological visual patterns are redundantly distributed across tissue slices, we sample a subset of visual instances during training. Our method demonstrates robust generalizability and strong transferability, dominantly outperforming the counterparts on the TCGA Lung Cancer dataset with at least $9.19\\%$ higher accuracy in few-shot experiments. The code is available at: https://github.com/ls1rius/WSI_FIVE.", + "bbox": [ + 75, + 376, + 472, + 859 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 344, + 630, + 359 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/28b2ad7e21f2112a208d419264e6b82df4c4aedcc96471ca2744a9ba059c2cb3.jpg", + "image_caption": [ + "Figure 1. Challenges in WSI-text contrastive learning. Most conventional VLM approaches categorize whole slide images using category-level text descriptions, overlooking intra-class differences, leading to a decline in performance and limitations in generalization capabilities. Instead, we extract fine-grained descriptions from pathology reports as slide-level labels to develop our model, exhibiting detailed variations in each WSI." + ], + "image_footnote": [], + "bbox": [ + 501, + 381, + 890, + 558 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Histological Whole Slide Image (WSI) classification plays a crucial role in computational pathology by automating disease diagnosis and subtyping. For high-resolution WSI analysis, Multiple Instance Learning (MIL) has become the dominant method. These methods treat each WSI as a \"Bag\" sample, and aggregate numerous patches within it as instances for thorough decision-making. Nevertheless, most existing methods [15, 16, 18, 31] focused on processing image data to conduct WSI classification, potentially not emphasizing critical pathological insights, particularly the expert textual annotations that accompany these slides.", + "bbox": [ + 496, + 686, + 890, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, Vision-Language Models (VLMs) [11, 25, 33] underscored the significance of integrating multimodal information for developing robust encoders. Zhang et al. [38]", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Co-corresponding authors. This work was supported by National Natural Science Foundation of China (Grant No. 62371409).", + "bbox": [ + 75, + 875, + 467, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11398", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "exploited disease-grade text labels and extracted text insights using pre-trained language models. Qu et al. [23] utilized GPT-4 in a question-and-answer mode to obtain language prior knowledge at both instance and bag levels for VLM training. However, the challenge lies in the uniqueness and variability of content in each WSI. Existing methods developed their models with coarse-grained descriptions (i.e., simplistic Category-Level text labels [38] or descriptive Category-Level text labels constructed by GPT-4 [23]), as shown in Fig. 1. They may have omitted crucial fine-grained pathological details, including differentiation level, vascular invasion, etc., which results in reduced model performance and limited generalization.", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "WSIs accompanied by their corresponding reports (i.e., WSI-report pairs) offer detailed descriptions and fine-grained information vital for WSI analysis. Furthermore, a substantial collection of these pairs is accessible in public databases, such as The Cancer Genome Atlas (TCGA) [9]. However, their full potential has not been adequately harnessed yet. The challenge in developing a Visual Language Model (VLM) using WSI-report pairs mainly lies in the diverse formats and standards of the raw reports from different hospitals, which increases the complexity of data preprocessing and standardization processes. Additionally, pathology reports often contain extraneous information, including report metadata, tissue processing descriptions, and repetitive elements, which can introduce noise to the textual data. How to extract useful information from raw pathology reports to construct WSI-report pairs is a key issue. Moreover, recent studies have demonstrated the efficacy of prompt engineering in enhancing VLMs. In contrast to natural images, WSI data encompasses extensive professional pathological information and intricate details. How to craft prompts to make full use of this semantic information to guide fine-grained feature learning is a challenging task. Besides, the high computational costs to train models with high-resolution WSIs also limits the promotion of the model, resulting in a certain resource threshold for WSI analysis.", + "bbox": [ + 75, + 292, + 472, + 685 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these issues, we propose a novel whole slide image classification method with Fine-grained VisualsEmantic interaction termed as FiVE, which shows robust generalizability and efficiency in computation. Firstly, we obtain WSIs with non-standardized raw pathology reports from a public database. Collaborating with professional pathologists, we craft a set of specialized prompts to standardize reports. Following this, we employ the large language model GPT-4 to automatically clean and standardize the raw report data. In addition, we propose the Task-specific Fine-grained Semantic (TFS) Module, which utilizes manual-designed prompts to direct visual attention to specific pathological areas while constructing Fine-Grained Guidance to enhance the semantic relevance of model fea", + "bbox": [ + 75, + 689, + 472, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tures. Considering the diffuse distribution of pathological diseases within tissue sections and the presence of numerous non-diagnostic regions in WSIs, we also incorporate a patch sampling strategy during the training phase to enhance training efficiency and reduce computational costs. The contributions of this paper are summarized as follows:", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We pioneer the utilization of the available WSI diagnostic reports with fine-grained guidance. The obtained fine-grained description labels lead to improved supervision by discriminating the visual appearances more precisely.", + "- We introduce a novel Task-specific Fine-grained Semantics (TFS) Module to offer fine-grained guidance, significantly enhancing the model's generalization capabilities.", + "- We implement a patch sampling strategy on visual instances during training to enhance computational efficiency without significantly compromising accuracy, thereby optimizing the model's training process." + ], + "bbox": [ + 496, + 185, + 890, + 351 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 498, + 364, + 640, + 380 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Whole Slide Image Analysis", + "text_level": 1, + "bbox": [ + 498, + 390, + 750, + 406 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contemporary methodologies for WSI analysis predominantly employ MIL methods where each WSI is treated as a \"Bag\" and its extracted patches as instances within this bag. MIL methods consist of instance-based methods [3, 14, 32] and embedding representation-based methods [15, 20, 26, 29, 36]. However, the majority of existing methods [5, 18, 24, 27] almost exclusively rely on image data, neglecting vital pathological details, notably the specialist text annotations that accompany the images. Recent works [23, 38] have taken note of this issue and started to utilize text information to improve pathological image classification. They used bag-level text labels or the descriptive labels generated by GPT. However, given the unique and varied descriptions of each WSI, their methods fall short of fully leveraging the detailed textual information present in the slides.", + "bbox": [ + 496, + 412, + 890, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Vision-Language Models", + "text_level": 1, + "bbox": [ + 498, + 665, + 725, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent researches have made efforts to develop Vision-Language Models (VLMs). CLIP [25] gathered 400 million image-text pairs and initiated training with synchronized vision and text encoders from the onset. LiT [35] developed a text encoder compatible with a pre-trained vision encoder. FLIP [33] integrated region-word alignment in contrastive learning to enhance detailed vision-language correlation. Coca [34] pre-trained an image-text encoder-decoder foundation model using contrastive and captioning loss. For pathological images, some research works adapted VLMs for training with pathological images and text. Lu et al. [21] built a VLM using over 1.17 million histopathology image-caption pairs based on a task-agnostic pre-training approach derived from Coca. Huang et al. [8] curated the OpenPath", + "bbox": [ + 496, + 688, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11399", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/25b31fb8e0c1ee118cf04665ee03203ae1feb6d33a687b1473fe7c98875cc92c.jpg", + "image_caption": [ + "Figure 2. Left: The structure of the FiVE framework. The model consists of a frozen image encoder, a text encoder, and the TFS module. Whole slide images are divided into instances for embedding extraction by the image encoder. Raw pathology reports are standardized by GPT-4 into fine-grained descriptions. The fine-grained descriptions and manual prompts are sampled, shuffled, and reconstructed in pairs. These prompts aggregate instances into bag-level features, subsequently aligned with the descriptions utilizing contrastive loss. Top Right: Fine-grained pathological descriptions. The fine-grained pathological descriptions are generated from multiple answers based on specific queries. These descriptions undergo a process of random sampling, shuffling, and reconstruction to form a unified sentence. Bottom Right: The Instance Aggregator module. The instance aggregator consists of a self-attention module and a cross-attention module, fusing image instance embeddings and prompt embeddings to create bag-level features." + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 877, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "dataset, consisting of 208,414 pathology images with natural language descriptions from public forums, and finetuned a pre-trained CLIP on OpenPath. Lai et al. [13] also explored the generalization of CLIP in pathology image classification. These methods typically employ instance-level images (small patches from WSIs) and descriptions, requiring significant human and material resources. Zhang et al. [38] injected meaningful medical domain knowledge to advance pathological image embedding and classification. Qu et al. [23] employed GPT-4 to supplement image labels to enrich the information for training. However, the texts employed in their training offer only rudimentary descriptions of the images, primarily categorizing the general type of pathological slides, such as the disease category. This approach significantly constrains the model's capacity to discern fine-grained features, including the degree of differentiation, spread, and other details. Consequently, this limitation substantially restricts the model's generalizability and applicability to more nuanced diagnostic tasks.", + "bbox": [ + 75, + 465, + 472, + 752 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Prompt Learning in Vision-Language Models", + "text_level": 1, + "bbox": [ + 76, + 768, + 464, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Drawing inspiration from prompt learning in natural language processing, some studies have proposed adapting Vision-Language models through end-to-end training of prompt tokens. CoOp [40] enhanced CLIP for few-shot transfer by optimizing a continuous array of prompt vectors within its language branch. CoCoOp [39] identified CoOp's suboptimal performance on new classes and tackled", + "bbox": [ + 75, + 794, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the generalization issue by conditioning prompts directly on image instances. Lu et al. [22] advocated for optimizing diverse sets of prompts by understanding their distribution. Bahng et al. [1] undertook visual prompt tuning on CLIP, focusing the prompting on the vision branch. MaPLe [12] investigated the effectiveness of multi-modal prompt learning in order to improve alignment between vision and language representations. Zhang et al. [37] adopted a set of learnable adaption prompts and presuppend them to the word tokens at higher transformer layers, efficiently fine-tuning LLaMA with less cost. Furthermore, in the context of WSI classification, prompts function as valuable adjuncts, enriching contextual information and semantic interpretation. The strategic utilization of prompts substantially improved model performance [23].", + "bbox": [ + 496, + 465, + 893, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 712, + 591, + 727 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 500, + 738, + 609, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fig. 2 shows the pipeline of our proposed FiVE method. To initiate the process, we collaborate with professional pathologists to establish a set of standards. Following this, we employ GPT-4 to automatically extract and standardize information based on these various standards. During the training phase, we construct Fine-Grained Guidance by intricately dividing and reconstructing the text description labels and manual prompts in pairs. The combination of manual prompts and learnable prompts forms the Diagnosis", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11400", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/f486abcdf1abee818562d564e4fd2e29c21c616114715b812bbbc6125fa8e3c4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Manual-Designed StandardsFine-Grained Text Description Label Examples
1. What is the differentiation of the lesion?TCGA-44-6774: Lesion differentiation is moderately to poorly differentiated; Unknown; No indication of vascular invasion by the lesion; No indication of pleural invasion by the lesion; Unknown; Margins of the excised tissue are clear of disease.
2. Is there any indication of spread through air spaces around the lesion?
3. Is there any indication of vascular invasion by the lesion?
4. Is there any indication of pleural invasion by the lesion?
5. Is there any evidence of the lesion invading adjacent tissues or organs?TCGA-49-4505: Lesion differentiation is well-differentiated; Unknown; Unknown; Pleural invasion by the lesion is present, as the carcinoma extends through the visceral pleura; The lesion invades adjacent tissues or organs; Margins of the excised tissue are clear of disease.
6. Are the margins of the excised tissue clear of disease?
", + "bbox": [ + 80, + 88, + 890, + 247 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. Manual-Designed Standards and Fine-Grained Text Description Label Examples. The answers on the right correspond to the standards on the left. \"Unknown\" is used as a placeholder when relevant information cannot be found.", + "bbox": [ + 76, + 257, + 892, + 286 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Prompts, which are utilized to enhance the semantic relevance of the features. Subsequently, the instance aggregator module fuses instance features with fine-grained prompts, generating bag-level features, subsequently align with the corresponding fine-grained text description labels. Additionally, to reduce computational costs, we implement the patch sampling strategy, optimizing the model's training efficiency while minimizing performance loss.", + "bbox": [ + 75, + 311, + 470, + 434 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Text Standardization via GPT-4", + "text_level": 1, + "bbox": [ + 76, + 467, + 357, + 483 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We utilize fine-grained text description labels extracted from pathology reports to align image bag-level features. Though pathology reports are readily accessible from public databases, their content exhibits significant variability depending on the source. Despite these format differences, pathology diagnoses consistently adhere to specific and well-established diagnostic standards. In our work, we develop fine-grained diagnostic criteria under the guidance of professional pathologists to standardize report data and extract fine-grained insights pertinent to pathological diagnosis, aiming to enhance its generalization capabilities substantially.", + "bbox": [ + 75, + 500, + 468, + 680 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The manual-designed standards aim to extract the morphological characteristics under the microscope, such as the degree of differentiation and lesion invasion, and filter out information irrelevant to the diagnosis. Subsequently, we employ GPT-4 to automatically extract answers from the original diagnosis reports based on prompts composed of these standards. If the information queried is absent in the pathology reports, \"Unknown\" is used as the answer. Tab. 1 shows manual-designed standards and two fine-grained text description label examples. Then, we recombine these extracted fine-grained information and integrate it into a complete description of the case image. More details about the prompts used for Text Standardization are provided in Supplementary Material.", + "bbox": [ + 75, + 689, + 470, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Task-specific Fine-grained Semantics Module", + "text_level": 1, + "bbox": [ + 498, + 311, + 883, + 328 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1 Fine-grained Guidance Construction", + "text_level": 1, + "bbox": [ + 500, + 337, + 810, + 353 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Due to the Text Standardization process, our data has achieved standardization. Utilizing these fine-grained text description directly as training labels can yield performance improvements. Additionally, leveraging them to generate more diverse and semantically enriched fine-grained guidance can further boost the model's performance.", + "bbox": [ + 496, + 366, + 890, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During the training process we utilize these manual-designed standards as our manual-designed prompts. We divide the original fine-grained text descriptions into several parts according to manual-designed prompts, followed by random sampling and eliminating \"Unknown\" tags from the initial labels. Given that the staged diagnostic reports in pathological descriptions are sequence-independent, we shuffle these preliminary labels and reconstruct them into a full-sentence description. During the training, we train reconstructed text description labels and reconstructed manual prompts in pairs. For example, consider description A: \"Lesion differentiation is moderately differentiated; Margins of the excised tissue are clear of disease.\" and description B: \"Lesion differentiation is moderately differentiated; Margins of the excised tissue are not clear of disease, as the tumor is within the bronchial margin and parenchymal margin.\" When only the first part of each description is sampled, they would be grouped into the same category. However, when sampling the entire sections, they are considered as distinct descriptions. Changes in granularity provide diverse perspectives on the visual image, aligning visual image with text descriptions of varying granularities.", + "bbox": [ + 496, + 460, + 892, + 805 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This strategy offers three key benefits: 1) Effectively alleviating the parent-child relationship in pathology categories. 2) Providing additional hierarchical semantic perspectives to enhance the text encoder's semantic comprehension ability. 3) Mitigating discrepancies in category annotation due to incomplete diagnostic information.", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11401", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.2 Diagnosis Prompts", + "text_level": 1, + "bbox": [ + 76, + 90, + 263, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We introduce Diagnosis Prompts to guide the aggregation of instance features into bag-level features. We compute the similarity between the instance features and the given manual prompts, utilizing the similarity scores as weights $W$ for feature aggregation to improve the task-specific relevance of the features. Here we utilize the identical manual prompts as those used to standardize the raw data, as shown in the left of the Tab. 1.", + "bbox": [ + 75, + 114, + 468, + 234 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition, manual-designed prompts may have some flaws, potentially failing to comprehensively capture the specific morphological characteristics of the lesion, and the model struggles to generalize towards unseen classes due to the late fusion through the transformer layers. Besides, fine-tuning the model may not always be feasible as it requires training a large number of parameters. Particularly in the case of low-data regimes, where the availability of training data like whole slide images is extremely limited. LLaMA-Adapter [37] and LLaMA-Adapter-v2 [6] explore the way to efficient fine-tuning of Language Models and Vision-Language Models respectively. These approaches introduced the Adaptation Prompt to gradually acquire instructional knowledge. They adopted zero-initialized attention with gating mechanisms to ensure stable training in the early stages. Inspired by these methods, we introduce learnable continuous diagnosis prompts to enrich the context information and enhance the model's transferability.", + "bbox": [ + 75, + 236, + 470, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, we get the manual text prompt tokens $Q_{h} = [q_{h1}, q_{h2}, \\dots, q_{hn}]$ , here $n$ represents the number of the manual prompts. We concatenate the learnable continuous prompt tokens $Q_{l} = [q_{l1}, q_{l2}, \\dots, q_{lm}]$ on it, here $m$ represents the number of the learnable prompt tokens. Finally we get the diagnosis prompts $Q = [q_{h1}, q_{h2}, \\dots, q_{hn}, q_{l1}, q_{l2}, \\dots, q_{lm}]$ . In the training phase, part of manual prompt tokens $Q_{h}$ will be sampled randomly paired with text description labels, while the whole learnable prompt tokens $Q_{l}$ will be consistently retained.", + "bbox": [ + 75, + 508, + 468, + 659 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Different with the traditional context learning prompts method, our approach pays attention to the acquisition of prior knowledge, similar to the methodology employed in Detection Transformer (DETR) [4]. We aim to acquire a set of appropriate query values to improve performance in subsequent feature screening processes. Additionally, it can also enable the model to quickly transfer to other tasks by fine-tuning this set of queries.", + "bbox": [ + 75, + 659, + 470, + 781 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.3 Instance Aggregator Module", + "text_level": 1, + "bbox": [ + 76, + 800, + 331, + 815 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Instance Aggregator (IA) module is used to aggregate the fine-grained diagnosis prompts and instance features. As shown in the right of the Fig. 2, IA consists of a self-attention module and a cross-attention module.", + "bbox": [ + 75, + 824, + 468, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We employ self-attention to enable feature interaction", + "bbox": [ + 96, + 885, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "among instance features $I_{i} = [e_{i1}, e_{i2}, \\dots, e_{ij}]$ , resulting in the feature $s_{i}$ . Subsequently, utilizing the diagnosis prompts $Q$ to aggregate the instance features and acquire the feature $z_{i}$ . Then we concatenate $s_{i}$ and $z_{i}$ , utilizing the learnable parameter $W$ to fuse these features, yielding the bag-level feature $v_{i}$ . The formulas are shown as follows:", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns _ {i} = \\text {S e l f A t t e n t i o n} \\left(I _ {i}, I _ {i}\\right) + I _ {i} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 191, + 890, + 209 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nz _ {i} = \\text {C r o s s A t t e n t i o n} (Q, s _ {i}) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 212, + 890, + 227 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nv _ {i} = \\operatorname {c o n c a t} \\left(\\operatorname {m e a n} \\left(s _ {i}\\right), \\operatorname {m e a n} \\left(z _ {i}\\right)\\right) \\cdot W \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 229, + 890, + 247 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Ultimately, we acquire the image bag-level features guided by the fine-grained diagnosis prompts, which are then employed to align the fine-grained text features.", + "bbox": [ + 496, + 255, + 890, + 300 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Patch Sample Strategy", + "text_level": 1, + "bbox": [ + 498, + 308, + 710, + 324 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Each Whole Slide Image (WSI) is partitioned into a variable number of instances, ranging from approximately 50 to 45,000. Handling such a wide range of instances markedly increases computational complexity and substantially extends the training duration. FLIP [33] reduced computation and reached higher accuracy than CLIP [25] counterpart by randomly removing a large portion of image patches during training [17]. In the case of whole slide images, pathological visual patterns are often redundantly distributed across a tissue slice. Therefore, it is feasible to sample only a subset of visual instances during training.", + "bbox": [ + 496, + 330, + 890, + 497 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the instances in each bag (i.e., slide), we select a sample amounting to $S_{m}$ percent of the total number $p$ of each group of instances (patches). The required number of instances $S_{n}$ is described by the following formula:", + "bbox": [ + 496, + 497, + 890, + 558 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS _ {n} = \\min \\left(p * S _ {m}, S _ {\\text {m a x n}}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 566, + 890, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here $S_{maxn}$ denotes the maximum number of sample instances. To achieve this, we evenly divide each group of instances into $S_{n}$ chunks and randomly select one instance from each chunk. Since the different whole slide images could sample different $S_{n}$ , we pad the rest of the space with the same padding.", + "bbox": [ + 496, + 590, + 890, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Encoder and Loss Function", + "text_level": 1, + "bbox": [ + 498, + 690, + 746, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We divide each WSI into instances $x_{k}$ and encode these instances into embeddings $e_{k} \\in R^{D}$ using pre-trained vision encoder $E_{img}$ , composed with ResNet structure following [15], here $D$ represents the dimension of the embeddings. Then we send the instance embeddings into the TFS Module to aggregate the instance features and prompts, and obtain the bag-level embeddings $v_{i} \\in R^{D}$ . The formulas are shown as follows:", + "bbox": [ + 496, + 712, + 890, + 833 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ne _ {k} = E _ {i m g} \\left(x _ {k}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 842, + 890, + 859 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI _ {i} = \\left[ e _ {i 1}, e _ {i 2}, \\dots , e _ {i k} \\right] \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 861, + 890, + 878 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nv _ {i} = I A \\left(I _ {i}, Q\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 880, + 890, + 897 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11402", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Besides, we generate pathologically meaningful text embeddings for each WSI, represented as $t_i \\in R^D$ , by leveraging the fine-tuned text encoder BioClinicalBERT [30].", + "bbox": [ + 76, + 90, + 468, + 137 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nt _ {i} ^ {c} = E _ {t x t} \\left(x _ {t x t} ^ {c}\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 148, + 468, + 165 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $E_{txt}$ denotes the text encoder, and $x_{txt}^{c}(c \\in [1, C])$ where $C$ denotes the number of categories. Here we use the same embedding dimension $D$ as the vision encoder, suitable for contrastive learning. For text encoder $E_{txt}$ , we adopt the Low-Rank Adaptation (LoRA) [7] approach for efficient fine-tuning.", + "bbox": [ + 76, + 175, + 468, + 265 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Subsequently, the bag-level embeddings $v_{i}$ are aligned with the text embeddings $t_{i}^{c}$ to complete the training process. In this case, prediction $\\hat{y}$ is obtained by applying softmax on scaled cosine similarities between the image embeddings and text embeddings:", + "bbox": [ + 76, + 266, + 468, + 340 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\np (\\hat {y} = c \\mid I) = \\frac {\\exp \\left(\\operatorname {s i m} \\left(t _ {i} ^ {c} , v _ {i}\\right) / \\tau\\right)}{\\sum_ {c ^ {\\prime} = 1} ^ {C} \\exp \\left(\\operatorname {s i m} \\left(t _ {i} ^ {c ^ {\\prime}} , v _ {i}\\right) / \\tau\\right)} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 351, + 468, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\text{sim}(\\cdot, \\cdot)$ refers to cosine similarity and $\\tau$ is the temperature parameter.", + "bbox": [ + 76, + 398, + 467, + 428 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The fine-grained training loss is computed as the cross-entropy between the logits and soft targets as:", + "bbox": [ + 76, + 429, + 467, + 458 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL ^ {v \\rightarrow t} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {N} y _ {i j} \\log \\left(p _ {i j}\\right) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 160, + 469, + 468, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "here $N$ corresponds to the batch size.", + "bbox": [ + 76, + 522, + 323, + 536 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Likewise, we can compute $L^{t \\to v}$ and serve $L$ as the final training objective.", + "bbox": [ + 76, + 537, + 468, + 568 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL = \\frac {L ^ {v \\rightarrow t} + L ^ {t \\rightarrow v}}{2} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 578, + 468, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments and Results", + "text_level": 1, + "bbox": [ + 76, + 618, + 310, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 76, + 643, + 179, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluated our method on public histopathology WSI datasets: The Cancer Genome Atlas Lung (TCGA Lung) Cancer1 and Camelyon16 [2].", + "bbox": [ + 76, + 666, + 468, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "TCGA Lung Cancer. The TCGA Lung Cancer dataset comprises two cancer subtypes: Lung Adenocarcinoma (LUAD) and Lung Squamous Cell Carcinoma (LUSC). It includes diagnostic slides with 541 slides from 478 LUAD cases and 512 slides from 478 LUSC cases. For WSI preprocessing, following [15], we cropped each WSI into $256 \\times 256$ non-overlapping patches and removed the background region. The dataset encompasses approximately 5.2 million patches at $20 \\times$ magnification, averaging about 5,000 patches per WSI. Following [28], we randomly split the dataset into training, validation, and testing sets with a ratio", + "bbox": [ + 75, + 710, + 468, + 877 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "of 65:10:25 on the patient level and adopted 4-fold cross-validation. We collected corresponding pathology reports ${}^{2}$ and cleaned them by GPT-4 to produce fine-grained text description labels. To ensure professionalism and accuracy, we invited professional pathologists to check and correct the textual labels. Additionally, to evaluate the generalizability of the model and perform zero-shot classification, we utilized a dataset of histological subtype labels for TCGALUAD from the cBioPortal database ${}^{3}$ ,More details about subtype labels are provided in Supplementary Material. Camelyon16. The Camelyon16 dataset [2] consists of 399 Hematoxylin and Eosin (H&E) stained slide images, utilized for metastasis detection in breast cancer. We preprocessed each WSI by segmenting it into ${256} \\times {256}$ nonoverlapping patches, excluding background regions. In total, this process yielded approximately 2.8 million patches at a ${20} \\times$ magnification level,with about 7,200 patches per bag. We adopted 3-times 3-fold cross-validation.", + "bbox": [ + 500, + 90, + 890, + 363 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 372, + 718, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In experiments, we employed ResNet following [15] as image encoder to extract image features, while pre-trained BioClinicalBERT from [30] as text encoder to generate text features. LoRA [7] was adopted for fine-tuning the text encoder, with an alpha value of 32 and a rank value of 8. We divided the whole slide image into patches with $256 \\times 256$ , then applied a random crop with size $224 \\times 224$ . We adopted AdamW [19] with beta (0.9, 0.98), eps 1e-8, learning rate 3e-6, warmup ratio 0.1, weight decay 1e-4 as our optimizer. Additionally, we used the batch size of 1 with an accumulation step of 8 and trained for 150 epochs. We utilized mixed-precision training on 4 NVIDIA-A800 GPUs.", + "bbox": [ + 498, + 396, + 890, + 577 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Zero-Shot Histological Subtype Classification", + "text_level": 1, + "bbox": [ + 500, + 587, + 883, + 603 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Prior researches [23, 28, 38] have predominantly concentrated on classifying primary cancer categories. Our approach extends beyond this by attempting to classify detailed histological subtypes. It is crucial to emphasize that this task poses a significant challenge, often proving difficult for even skilled pathologists to make direct judgments.", + "bbox": [ + 498, + 609, + 890, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since only the LUAD's histological subtype dataset was provided on the online database, we conducted zero-shot subtype classification evaluation on LUAD subtype datasets, with the model being pre-trained on TCGA-LUAD or TCGA-LUSC. Throughout the training phase, subtype label information was deliberately excluded, ensuring that all experiments are conducted solely with fine-grained labels. This aims to evaluate the model's ability to identify novel diagnostic categories without specific training on subtypes. We extended the morphological appearance text de", + "bbox": [ + 498, + 700, + 890, + 852 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1http://www.cancer.gov/ctga", + "bbox": [ + 94, + 886, + 308, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "2https://github.com/tatonetti-lab/tcga-path-reports", + "bbox": [ + 500, + 862, + 888, + 886 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "3https://www.cbiportal.org/", + "bbox": [ + 517, + 887, + 738, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11403", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "scription labels of the target data using GPT-4. After the text encoder encodes the labels, similarity calculation with image features achieved zero-shot classification.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Since existing zero-shot learning methods cannot be used in WSI subtype classification, we constructed three Linear-Probe method baselines: Mean pooling, Max pooling, and Attention pooling. As shown in Tab. 2, FiVE attains $65.23\\%$ top-1 accuracy and $95.18\\%$ top-5 accuracy when pre-trained with TCGA-LUAD fine-grained labels. When pre-trained with TCGA-LUSC fine-grained labels, it achieves $62.02\\%$ top-1 accuracy and $94.36\\%$ top-5 accuracy. Moreover, the zero-shot performance of FiVE notably exceeds that of the baseline. This capability in classifying LUAD subtypes is attributed to the focus on fine-grained pathological features during training.", + "bbox": [ + 75, + 138, + 470, + 319 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2bc9859bca4dffac32a9fa34e4cdcc7a922914a8979b4ac296d86900d85f4354.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTCGA-LUADTCGA-LUSC
Top-1Top-5Top-1Top-5
Mean-pooling40.8283.4631.8682.96
Max-pooling45.0588.7736.3686.09
Attention-pooling58.3693.5354.4192.50
FiVE (Ours)65.2395.1862.0294.36
+6.87+1.65+7.61+1.86
", + "bbox": [ + 86, + 335, + 460, + 470 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Few-Shot Classification", + "text_level": 1, + "bbox": [ + 76, + 570, + 294, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our model demonstrates adaptability to various tasks even in scenarios with limited data availability. Few-shot experiments were conducted to demonstrate its transferability to downstream tasks. We initialized the networks with pretrained weights derived from the model trained on TCGA image-report pairs, and subsequently fine-tuned the model on downstream datasets for few-shot image classification. We followed [23] and conducted experiments with 1, 2, 4, 8, 16, and additional 0 shot on the downstream dataset. The results are summarized in Tab. 3.", + "bbox": [ + 75, + 595, + 468, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our model exhibits remarkable performance in zero-shot classification, achieving an accuracy of $71.26\\%$ , even surpassing the SOTA method's one-shot experiment. Upon the introduction of training data, our models display exceptional transferability, outperforming the SOTA by $12.90\\%$ in the one-shot setting. The model's performance improves accordingly with the number of shot. Upon reaching 16-shot, our model reaches an impressive accuracy of $91.25\\%$ , showcasing a notable $9.19\\%$ improvement close to the fully supervised performance level.", + "bbox": [ + 75, + 750, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/44a4b61d9aebe11ccfffa069fbfbf1eec6de362c197ed3cc458347ee5e7c29fa.jpg", + "table_caption": [ + "Table 2. Zero-Shot performance on histological subtype classification. We pre-trained the model using fine-grained labels of TCGA-LUAD and TCGA-LUSC, then applied zero-shot classification to the histological subtypes of TCGA-LUAD." + ], + "table_footnote": [], + "table_body": "
Method16-shot8-shot4-shot2-shot1-shot0-shot
Mean-pool65.3353.8944.8552.9345.34\\
Max-pool48.4849.5544.2248.3949.03\\
Attn-pool72.5065.7962.4758.3656.23\\
CoOp [40]78.3567.9967.6067.5467.81\\
TOP [23]82.0680.5175.4172.3871.01\\
FiVE91.2590.8088.1085.5183.9171.26
(Ours)+9.19+10.29+12.69+13.13+12.90
", + "bbox": [ + 504, + 88, + 888, + 202 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Few-shot classification performance on TCGA Lung Cancer. Mean-pool, Max-pool, and Attn-pool correspond to Linear-Probe implementations with Mean-pooling, Max-pooling, and Attention-pooling, respectively.", + "bbox": [ + 496, + 212, + 893, + 270 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Performance Comparison with Existing Works", + "text_level": 1, + "bbox": [ + 498, + 297, + 890, + 314 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compared FiVE with ABMIL [10], DSMIL [15], CLAM-SB [20], CLAM-MB [20], TransMIL [26], DTFDMIL [36], and MHIM-MIL [28], all of which are attention-based MIL methods. In addition, we included two traditional MIL pooling operations, Max-pooling and Mean-pooling, for comparison. The results of all other methods are reproduced using the official code they provide under the same settings.", + "bbox": [ + 496, + 321, + 890, + 443 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate the performance of FiVE, we conducted fine-grained pre-training exclusively on the TCGA dataset, followed by fully supervised experiments on Camelyon16 and TCGA Lung Cancer datasets for WSI classification. Note that the test data is not used in pre-training. Results are shown in Tab. 4. It can be found that our method outperforms all the other baselines by a great margin, which fully demonstrates the significance of our fine-grained training scheme in improving performance on downstream tasks. Besides, since the pre-training data mainly comes from TCGA data, on the other hand, the task-specific prompt design is more suitable for TCGA data. This results in the performance improvement of our method for TCGA data being greater than Camelyon16 data.", + "bbox": [ + 496, + 444, + 892, + 655 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 667, + 663, + 684 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6.1 Effectiveness of TFS Module", + "text_level": 1, + "bbox": [ + 500, + 693, + 754, + 708 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We focused on evaluating the impact of the TFS module on the model's overall performance with the TCGA Lung Cancer. The results detailed in Tab. 5 reveal significant improvements at each stage of feature enhancement. Initially, the model with only Self Attention attained ACC of $89.77\\%$ , AUC of $92.85\\%$ , and F1-score of $89.95\\%$ . The incorporation of fine-grained labels led to increases of $1.44\\%$ in ACC, $1.51\\%$ in AUC, and $1.43\\%$ in F1-score. Subsequent integration of additional fine-grained guidance further improved performance. Ultimately, full framework with the TFS Module achieved the highest performance of $94.62\\%$ ACC, $96.33\\%$ AUC, and $93.89\\%$ F1-score. These ablation", + "bbox": [ + 496, + 719, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11404", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1bdf47fcf107708d170daee9c7880782f8f2761fd80356a34a6685cc04a78994.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCamelyon16TCGA Lung Cancer
ACCAUCF1-scoreACCAUCF1-score
Max-pooling78.95±2.2881.28±3.7471.06±2.5981.49±1.2486.45±0.7180.56±1.09
Mean-pooling76.69±0.2080.07±0.7870.41±0.1684.14±2.9790.13±2.4083.39±3.14
ABMIL [10]90.06±0.6094.00±0.8387.40±1.0588.03±2.1993.17±2.0587.41±2.42
DSMIL [15]90.17±1.0294.57±0.4087.65±1.1888.32±2.7093.71±1.8287.90±2.50
CLAM-SB [20]90.32±0.1294.65±0.3087.89±0.5987.74±2.2293.67±1.6487.36±2.24
CLAM-MB [20]90.14±0.8594.70±0.7688.10±0.6388.73±1.6293.69±0.5488.28±1.58
TransMIL [26]89.22±2.3293.51±2.1385.10±4.3387.08±1.9792.51±1.7686.40±2.08
DTFD-MIL [36]90.22±0.3695.15±0.1487.62±0.5988.23±2.1293.83±1.3987.71±2.04
MHIM-MIL [28]92.48±0.3596.49±0.6590.75±0.7389.93±3.3795.53±1.7489.71±2.92
FiVE (Ours)94.25±0.3397.56±0.5993.24±0.7294.62±2.1396.33±1.2193.89±1.90
", + "bbox": [ + 125, + 88, + 846, + 292 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/469b212bbaf2a191fedaec192c91e5037f792ec6cca093978d3efb47157a7ea4.jpg", + "table_caption": [ + "Table 4. Comparison Performance of slide classification on Camelyon16 and TCGA Lung Cancer." + ], + "table_footnote": [ + "Table 5. The ablation experiments of the pre-trained model finetuned on the TCGA Lung Cancer. SA, FGL, FGG, and LDP represent Self Attention, Fine-Grained Labels, Fine-Grained Guidance, and Learnable Diagnosis Prompts, respectively." + ], + "table_body": "
SAFGLFGGLDPACCAUCF1-score
✓89.7792.8589.58
✓✓91.2194.3691.01
✓✓✓93.5696.0193.17
✓✓✓✓94.6296.3393.89
", + "bbox": [ + 78, + 339, + 467, + 421 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "experiments highlight the benefits of integrating these methods in the WSI classification task.", + "bbox": [ + 75, + 502, + 468, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6.2 Effectiveness of Patch Sample Strategy", + "text_level": 1, + "bbox": [ + 76, + 553, + 401, + 569 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We verified the impact of different sampling strategies on model performance on TCGA data. Here, we assumed that there are two main indicators that affect model performance, sample ratio and max sample threshold (represented as MAXN). At the same time, in order to differentiate the experimental results, we used the unfrozen image encoder for experimental verification.", + "bbox": [ + 75, + 578, + 468, + 685 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/46a414cffb3970e915eb9e2bcc441f82f580a6c0d9aa1fc3bde5b1e90bd3a029.jpg", + "image_caption": [ + "Figure 3. Classification performance on TCGA Lung Cancer with diverse sampling strategies, presenting average and standard deviation (std) ACC values." + ], + "image_footnote": [], + "bbox": [ + 81, + 699, + 269, + 790 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dcefa4dcf6a6b3ec8e8923aa33e0d8e8909eb6510db76eb431b162b5e66f8c25.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 699, + 464, + 790 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Fig. 3, when MAXN is sufficiently large ( $\\geq$ 2048), the correlation between the model's performance and the sample ratio is not significant. Even with a small sample", + "bbox": [ + 76, + 854, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ratio $(\\leq 0.5)$ , the model can still effectively align with the original data distribution. Conversely, when MAXN is not large enough (<2048), the primary change in the model's performance depends on whether the magnitude of the sample ratio can match the original data distribution. When the sample ratio increases to 0.75, a steep performance improvement can be observed, suggesting that the sampled data at this point conforms to the original data distribution, after which the model's performance stabilizes.", + "bbox": [ + 496, + 343, + 890, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Based on comprehensive experimental results, we established the sample ratio and MAXN as 0.5 and 2048 as suitable hyperparameters for the unfrozen image encoder. As for the frozen image encoder experiment, taking into account the performance constraints imposed by the frozen image encoder on the model's performance upper limit [38], we recommend the sample ratio and MAXN to be set at 0.5 and 16384 based on our experiments.", + "bbox": [ + 496, + 479, + 892, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 617, + 617, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce FiVE, a novel framework that demonstrates robust generalization and strong transferability for WSI classification. Our work pioneers the utilization of non-standardized WSI-report pairs from public databases to develop a VLM. To capture the complexities and diversity within these reports, we introduce the Task-specific Fine-grained Semantics (TFS) module. This module reconstructs fine-grained labels and diagnosis prompts during training, enhancing the semantic relevance of its features by introducing diagnosis prompts. Furthermore, considering that pathological visual patterns are redundantly distributed across tissue slices, we employ a sampling strategy to reduce computational costs. Our experiments demonstrate the robust generalizability and computational efficiency of the proposed framework, which can also be easily adapted to other tasks with minimal fine-tuning. We aspire to provide empirical insights and contribute to AI pathology research.", + "bbox": [ + 496, + 643, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11405", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Visual prompting: Modifying pixel space to adapt pre-trained models. arXiv preprint arXiv:2203.17274, 3:11-12, 2022. 3", + "[2] Babak Ehteshami Bejnordi, Mitko Veta, Paul Johannes Van Diest, Bram Van Ginneken, Nico Karssemeijer, Geert Litjens, Jeroen AWM Van Der Laak, Meyke Hermsen, Quirine F Manson, Maschenka Balkenhol, et al. Diagnostic assessment of deep learning algorithms for detection of lymph node metastases in women with breast cancer. Jama, 318(22):2199-2210, 2017. 6", + "[3] Gabriele Campanella, Matthew G Hanna, Luke Geneslaw, Allen Miraflor, Vitor Werneck Krauss Silva, Klaus J Busam, Edi Brogi, Victor E Reuter, David S Klimstra, and Thomas J Fuchs. Clinical-grade computational pathology using weakly supervised deep learning on whole slide images. Nature medicine, 25(8):1301-1309, 2019. 2", + "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 5", + "[5] Tsai Hor Chan, Fernando Julio Cendra, Lan Ma, Guosheng Yin, and Lequan Yu. Histopathology whole slide image analysis with heterogeneous graph representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15661-15670, 2023. 2", + "[6] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xi-angyu Yue, et al. Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010, 2023. 5", + "[7] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6", + "[8] Zhi Huang, Federico Bianchi, Mert Yuksekgonul, Thomas J Montine, and James Zou. A visual-language foundation model for pathology image analysis using medical twitter. Nature Medicine, pages 1-10, 2023. 2", + "[9] Carolyn Hutter and Jean Claude Zenklusen. The cancer genome atlas: creating lasting value beyond its data. Cell, 173(2):283-285, 2018. 2", + "[10] Maximilian Ilse, Jakub Tomczak, and Max Welling. Attention-based deep multiple instance learning. In International conference on machine learning, pages 2127-2136. PMLR, 2018. 7, 8", + "[11] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 1", + "[12] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In Proceedings of the" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19113-19122, 2023. 3", + "[13] Zhengfeng Lai, Zhuoheng Li, Luca Cerny Oliveira, Joohi Chauhan, Brittany N Dugger, and Chen-Nee Chuah. Clipse: Fine-tune clip with visual feature fusion for pathology image analysis towards minimizing data collection efforts. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2374-2380, 2023. 3", + "[14] Marvin Lerousseau, Maria Vakalopoulou, Marion Classe, Julien Adam, Enzo Battistella, Alexandre Carré, Théo Estienne, Théophraste Henry, Eric Deutsch, and Nikos Paragios. Weakly supervised multiple instance learning histopathological tumor segmentation. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part V 23, pages 470-479. Springer, 2020. 2", + "[15] Bin Li, Yin Li, and Kevin W Eliceiri. Dual-stream multiple instance learning network for whole slide image classification with self-supervised contrastive learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14318-14328, 2021. 1, 2, 5, 6, 7, 8", + "[16] Honglin Li, Chenglu Zhu, Yunlong Zhang, Yuxuan Sun, Zhongyi Shui, Wenwei Kuang, Sunyi Zheng, and Lin Yang. Task-specific fine-tuning via variational information bottleneck for weakly-supervised pathology whole slide image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7454-7463, 2023. 1", + "[17] Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. Scaling language-image pre-training via masking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23390-23400, 2023. 5", + "[18] Tiancheng Lin, Zhimiao Yu, Hongyu Hu, Yi Xu, and Chang-Wen Chen. Interventional bag multi-instance learning on whole-slide pathological images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19830-19839, 2023. 1, 2", + "[19] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6", + "[20] Ming Y Lu, Drew FK Williamson, Tiffany Y Chen, Richard J Chen, Matteo Barbieri, and Faisal Mahmood. Data-efficient and weakly supervised computational pathology on whole-slide images. Nature biomedical engineering, 5(6):555-570, 2021. 2, 7, 8", + "[21] Ming Y Lu, Bowen Chen, Drew FK Williamson, Richard J Chen, Ivy Liang, Tong Ding, Guillaume Jaume, Igor Odintsov, Andrew Zhang, Long Phi Le, et al. Towards a visual-language foundation model for computational pathology. arXiv preprint arXiv:2307.12914, 2023. 2", + "[22] Yuning Lu, Jianzhuang Liu, Yonggang Zhang, Yajing Liu, and Xinmei Tian. Prompt distribution learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5206-5215, 2022. 3", + "[23] Linhao Qu, Xiaoyuan Luo, Kexue Fu, Manning Wang, and Zhijian Song. The rise of ai language pathologists:" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11406", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Exploring two-level prompt learning for few-shot weakly-supervised whole slide image classification. arXiv preprint arXiv:2305.17891, 2023. 2, 3, 6, 7", + "[24] Linhao Qu, Zhiwei Yang, Minghong Duan, Yingfan Ma, Shuo Wang, Manning Wang, and Zhijian Song. Boosting whole slide image classification from the perspectives of distribution, correlation and magnification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21463-21473, 2023. 2", + "[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 5", + "[26] Zhuchen Shao, Hao Bian, Yang Chen, Yifeng Wang, Jian Zhang, Xiangyang Ji, et al. Transmil: Transformer based correlated multiple instance learning for whole slide image classification. Advances in neural information processing systems, 34:2136-2147, 2021. 2, 7, 8", + "[27] Zhuchen Shao, Yifeng Wang, Yang Chen, Hao Bian, Shaohui Liu, Haoqian Wang, and Yongbing Zhang. Lnpl-mil: Learning from noisy pseudo labels for promoting multiple instance learning in whole slide image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21495-21505, 2023. 2", + "[28] Wenhao Tang, Sheng Huang, Xiaoxian Zhang, Fengtao Zhou, Yi Zhang, and Bo Liu. Multiple instance learning framework with masked hard instance mining for whole slide image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4078-4087, 2023. 6, 7, 8", + "[29] Xinggang Wang, Yongluan Yan, Peng Tang, Xiang Bai, and Wenyu Liu. Revisiting multiple instance neural networks. Pattern Recognition, 74:15-24, 2018. 2", + "[30] Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chau-mond, Clement Delangue, Anthony Moi, Pierrick Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, et al. Huggingface's transformers: State-of-the-art natural language processing. arXiv preprint arXiv:1910.03771, 2019. 6", + "[31] Jinxi Xiang and Jun Zhang. Exploring low-rank property in multiple instance learning for whole slide image classification. In The Eleventh International Conference on Learning Representations, 2022. 1", + "[32] Gang Xu, Zhigang Song, Zhuo Sun, Calvin Ku, Zhe Yang, Cancheng Liu, Shuhao Wang, Jianpeng Ma, and Wei Xu. Camel: A weakly supervised learning framework for histopathology image segmentation. In Proceedings of the IEEE/CVF International Conference on computer vision, pages 10682-10691, 2019. 2", + "[33] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. arXiv preprint arXiv:2111.07783, 2021. 1, 2, 5", + "[34] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 2" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18123-18133, 2022. 2", + "[36] Hongrun Zhang, Yanda Meng, Yitian Zhao, Yihong Qiao, Xiaoyun Yang, Sarah E Coupland, and Yalin Zheng. Dtdmil: Double-tier feature distillation multiple instance learning for histopathology whole slide image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18802-18812, 2022. 2, 7, 8", + "[37] Renrui Zhang, Jiaming Han, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, Peng Gao, and Yu Qiao. Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023. 3, 5", + "[38] Yunkun Zhang, Jin Gao, Mu Zhou, Xiaosong Wang, Yu Qiao, Shaoting Zhang, and Dequan Wang. Text-guided foundation model adaptation for pathological image classification. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 272-282. Springer, 2023. 1, 2, 3, 6, 8", + "[39] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 3", + "[40] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 3, 7" + ], + "bbox": [ + 501, + 90, + 890, + 540 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11407", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Generalizable Whole Slide Image Classification with Fine-Grained Visual-Semantic Interaction/c4c5691b-0982-48b4-b41b-bc14bcf8455f_model.json b/2024/Generalizable Whole Slide Image Classification with Fine-Grained Visual-Semantic Interaction/c4c5691b-0982-48b4-b41b-bc14bcf8455f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3088a4572506d69fbc8e983a0a6ddaad6b9b6bbc --- /dev/null +++ b/2024/Generalizable Whole Slide Image Classification with Fine-Grained Visual-Semantic Interaction/c4c5691b-0982-48b4-b41b-bc14bcf8455f_model.json @@ -0,0 +1,2123 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.15, + 0.131, + 0.82, + 0.174 + ], + "angle": 0, + "content": "Generalizable Whole Slide Image Classification with Fine-Grained Visual-Semantic Interaction" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.203, + 0.314, + 0.221 + ], + "angle": 0, + "content": "Hao Li" + }, + { + "type": "text", + "bbox": [ + 0.197, + 0.222, + 0.329, + 0.24 + ], + "angle": 0, + "content": "Wenxian Yang3" + }, + { + "type": "text", + "bbox": [ + 0.361, + 0.204, + 0.459, + 0.222 + ], + "angle": 0, + "content": "Ying Chen" + }, + { + "type": "text", + "bbox": [ + 0.361, + 0.222, + 0.499, + 0.24 + ], + "angle": 0, + "content": "Liansheng Wang" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.204, + 0.606, + 0.221 + ], + "angle": 0, + "content": "Yifei Chen" + }, + { + "type": "text", + "bbox": [ + 0.542, + 0.222, + 0.653, + 0.24 + ], + "angle": 0, + "content": "Bowen Ding" + }, + { + "type": "text", + "bbox": [ + 0.652, + 0.204, + 0.781, + 0.222 + ], + "angle": 0, + "content": "Rongshan \\(\\mathbf{Y}\\mathbf{u}^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.687, + 0.222, + 0.806, + 0.238 + ], + "angle": 0, + "content": "Yuchen Han\\*" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.24, + 0.885, + 0.257 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Informatics, Xiamen University \\(^{2}\\)Huawei \\(^{3}\\)Aginome Scientific, Xiamen, China" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.257, + 0.908, + 0.275 + ], + "angle": 0, + "content": "\\(^{4}\\)Department of Pathology, Shanghai Chest Hospital, Shanghai Jiao Tong University School of Medicine" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.277, + 0.762, + 0.293 + ], + "angle": 0, + "content": "{111ih, cying2023}@stu.xmu.edu.cn, {rsyu, 1swang}@xmu.edu.cn," + }, + { + "type": "text", + "bbox": [ + 0.146, + 0.296, + 0.844, + 0.309 + ], + "angle": 0, + "content": "chenyifei14@huawei.com,wx@aginome.com，dingbowenmail@126.com，ychan@cmu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.345, + 0.313, + 0.36 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.377, + 0.473, + 0.86 + ], + "angle": 0, + "content": "Whole Slide Image (WSI) classification is often formulated as a Multiple Instance Learning (MIL) problem. Recently, Vision-Language Models (VLMs) have demonstrated remarkable performance in WSI classification. However, existing methods leverage coarse-grained pathogenetic descriptions for visual representation supervision, which are insufficient to capture the complex visual appearance of pathogenetic images, hindering the generalizability of models on diverse downstream tasks. Additionally, processing high-resolution WSIs can be computationally expensive. In this paper, we propose a novel \"Fine-grained Visual-Semantic Interaction\" (FiVE) framework for WSI classification. It is designed to enhance the model's generalizability by leveraging the interaction between localized visual patterns and fine-grained pathological semantics. Specifically, with meticulously designed queries, we start by utilizing a large language model to extract fine-grained pathological descriptions from various non-standardized raw reports. The output descriptions are then reconstructed into fine-grained labels used for training. By introducing a Task-specific Fine-grained Semantics (TFS) module, we enable prompts to capture crucial visual information in WSIs, which enhances representation learning and augments generalization capabilities significantly. Furthermore, given that pathological visual patterns are redundantly distributed across tissue slices, we sample a subset of visual instances during training. Our method demonstrates robust generalizability and strong transferability, dominantly outperforming the counterparts on the TCGA Lung Cancer dataset with at least \\(9.19\\%\\) higher accuracy in few-shot experiments. The code is available at: https://github.com/ls1rius/WSI_FIVE." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.345, + 0.631, + 0.36 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.382, + 0.891, + 0.559 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.568, + 0.892, + 0.665 + ], + "angle": 0, + "content": "Figure 1. Challenges in WSI-text contrastive learning. Most conventional VLM approaches categorize whole slide images using category-level text descriptions, overlooking intra-class differences, leading to a decline in performance and limitations in generalization capabilities. Instead, we extract fine-grained descriptions from pathology reports as slide-level labels to develop our model, exhibiting detailed variations in each WSI." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.688, + 0.892, + 0.854 + ], + "angle": 0, + "content": "Histological Whole Slide Image (WSI) classification plays a crucial role in computational pathology by automating disease diagnosis and subtyping. For high-resolution WSI analysis, Multiple Instance Learning (MIL) has become the dominant method. These methods treat each WSI as a \"Bag\" sample, and aggregate numerous patches within it as instances for thorough decision-making. Nevertheless, most existing methods [15, 16, 18, 31] focused on processing image data to conduct WSI classification, potentially not emphasizing critical pathological insights, particularly the expert textual annotations that accompany these slides." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Recently, Vision-Language Models (VLMs) [11, 25, 33] underscored the significance of integrating multimodal information for developing robust encoders. Zhang et al. [38]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.875, + 0.468, + 0.9 + ], + "angle": 0, + "content": "*Co-corresponding authors. This work was supported by National Natural Science Foundation of China (Grant No. 62371409)." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11398" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.288 + ], + "angle": 0, + "content": "exploited disease-grade text labels and extracted text insights using pre-trained language models. Qu et al. [23] utilized GPT-4 in a question-and-answer mode to obtain language prior knowledge at both instance and bag levels for VLM training. However, the challenge lies in the uniqueness and variability of content in each WSI. Existing methods developed their models with coarse-grained descriptions (i.e., simplistic Category-Level text labels [38] or descriptive Category-Level text labels constructed by GPT-4 [23]), as shown in Fig. 1. They may have omitted crucial fine-grained pathological details, including differentiation level, vascular invasion, etc., which results in reduced model performance and limited generalization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.293, + 0.473, + 0.686 + ], + "angle": 0, + "content": "WSIs accompanied by their corresponding reports (i.e., WSI-report pairs) offer detailed descriptions and fine-grained information vital for WSI analysis. Furthermore, a substantial collection of these pairs is accessible in public databases, such as The Cancer Genome Atlas (TCGA) [9]. However, their full potential has not been adequately harnessed yet. The challenge in developing a Visual Language Model (VLM) using WSI-report pairs mainly lies in the diverse formats and standards of the raw reports from different hospitals, which increases the complexity of data preprocessing and standardization processes. Additionally, pathology reports often contain extraneous information, including report metadata, tissue processing descriptions, and repetitive elements, which can introduce noise to the textual data. How to extract useful information from raw pathology reports to construct WSI-report pairs is a key issue. Moreover, recent studies have demonstrated the efficacy of prompt engineering in enhancing VLMs. In contrast to natural images, WSI data encompasses extensive professional pathological information and intricate details. How to craft prompts to make full use of this semantic information to guide fine-grained feature learning is a challenging task. Besides, the high computational costs to train models with high-resolution WSIs also limits the promotion of the model, resulting in a certain resource threshold for WSI analysis." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.69, + 0.473, + 0.901 + ], + "angle": 0, + "content": "To address these issues, we propose a novel whole slide image classification method with Fine-grained VisualsEmantic interaction termed as FiVE, which shows robust generalizability and efficiency in computation. Firstly, we obtain WSIs with non-standardized raw pathology reports from a public database. Collaborating with professional pathologists, we craft a set of specialized prompts to standardize reports. Following this, we employ the large language model GPT-4 to automatically clean and standardize the raw report data. In addition, we propose the Task-specific Fine-grained Semantic (TFS) Module, which utilizes manual-designed prompts to direct visual attention to specific pathological areas while constructing Fine-Grained Guidance to enhance the semantic relevance of model fea" + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.092, + 0.892, + 0.183 + ], + "angle": 0, + "content": "tures. Considering the diffuse distribution of pathological diseases within tissue sections and the presence of numerous non-diagnostic regions in WSIs, we also incorporate a patch sampling strategy during the training phase to enhance training efficiency and reduce computational costs. The contributions of this paper are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.186, + 0.892, + 0.245 + ], + "angle": 0, + "content": "- We pioneer the utilization of the available WSI diagnostic reports with fine-grained guidance. The obtained fine-grained description labels lead to improved supervision by discriminating the visual appearances more precisely." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.247, + 0.891, + 0.291 + ], + "angle": 0, + "content": "- We introduce a novel Task-specific Fine-grained Semantics (TFS) Module to offer fine-grained guidance, significantly enhancing the model's generalization capabilities." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.292, + 0.892, + 0.352 + ], + "angle": 0, + "content": "- We implement a patch sampling strategy on visual instances during training to enhance computational efficiency without significantly compromising accuracy, thereby optimizing the model's training process." + }, + { + "type": "list", + "bbox": [ + 0.498, + 0.186, + 0.892, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.366, + 0.642, + 0.381 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.391, + 0.75, + 0.407 + ], + "angle": 0, + "content": "2.1. Whole Slide Image Analysis" + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.414, + 0.892, + 0.655 + ], + "angle": 0, + "content": "Contemporary methodologies for WSI analysis predominantly employ MIL methods where each WSI is treated as a \"Bag\" and its extracted patches as instances within this bag. MIL methods consist of instance-based methods [3, 14, 32] and embedding representation-based methods [15, 20, 26, 29, 36]. However, the majority of existing methods [5, 18, 24, 27] almost exclusively rely on image data, neglecting vital pathological details, notably the specialist text annotations that accompany the images. Recent works [23, 38] have taken note of this issue and started to utilize text information to improve pathological image classification. They used bag-level text labels or the descriptive labels generated by GPT. However, given the unique and varied descriptions of each WSI, their methods fall short of fully leveraging the detailed textual information present in the slides." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.666, + 0.727, + 0.682 + ], + "angle": 0, + "content": "2.2. Vision-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.689, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Recent researches have made efforts to develop Vision-Language Models (VLMs). CLIP [25] gathered 400 million image-text pairs and initiated training with synchronized vision and text encoders from the onset. LiT [35] developed a text encoder compatible with a pre-trained vision encoder. FLIP [33] integrated region-word alignment in contrastive learning to enhance detailed vision-language correlation. Coca [34] pre-trained an image-text encoder-decoder foundation model using contrastive and captioning loss. For pathological images, some research works adapted VLMs for training with pathological images and text. Lu et al. [21] built a VLM using over 1.17 million histopathology image-caption pairs based on a task-agnostic pre-training approach derived from Coca. Huang et al. [8] curated the OpenPath" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11399" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.878, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.328, + 0.896, + 0.442 + ], + "angle": 0, + "content": "Figure 2. Left: The structure of the FiVE framework. The model consists of a frozen image encoder, a text encoder, and the TFS module. Whole slide images are divided into instances for embedding extraction by the image encoder. Raw pathology reports are standardized by GPT-4 into fine-grained descriptions. The fine-grained descriptions and manual prompts are sampled, shuffled, and reconstructed in pairs. These prompts aggregate instances into bag-level features, subsequently aligned with the descriptions utilizing contrastive loss. Top Right: Fine-grained pathological descriptions. The fine-grained pathological descriptions are generated from multiple answers based on specific queries. These descriptions undergo a process of random sampling, shuffling, and reconstruction to form a unified sentence. Bottom Right: The Instance Aggregator module. The instance aggregator consists of a self-attention module and a cross-attention module, fusing image instance embeddings and prompt embeddings to create bag-level features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.466, + 0.473, + 0.753 + ], + "angle": 0, + "content": "dataset, consisting of 208,414 pathology images with natural language descriptions from public forums, and finetuned a pre-trained CLIP on OpenPath. Lai et al. [13] also explored the generalization of CLIP in pathology image classification. These methods typically employ instance-level images (small patches from WSIs) and descriptions, requiring significant human and material resources. Zhang et al. [38] injected meaningful medical domain knowledge to advance pathological image embedding and classification. Qu et al. [23] employed GPT-4 to supplement image labels to enrich the information for training. However, the texts employed in their training offer only rudimentary descriptions of the images, primarily categorizing the general type of pathological slides, such as the disease category. This approach significantly constrains the model's capacity to discern fine-grained features, including the degree of differentiation, spread, and other details. Consequently, this limitation substantially restricts the model's generalizability and applicability to more nuanced diagnostic tasks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.769, + 0.465, + 0.788 + ], + "angle": 0, + "content": "2.3. Prompt Learning in Vision-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.473, + 0.904 + ], + "angle": 0, + "content": "Drawing inspiration from prompt learning in natural language processing, some studies have proposed adapting Vision-Language models through end-to-end training of prompt tokens. CoOp [40] enhanced CLIP for few-shot transfer by optimizing a continuous array of prompt vectors within its language branch. CoCoOp [39] identified CoOp's suboptimal performance on new classes and tackled" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.466, + 0.895, + 0.693 + ], + "angle": 0, + "content": "the generalization issue by conditioning prompts directly on image instances. Lu et al. [22] advocated for optimizing diverse sets of prompts by understanding their distribution. Bahng et al. [1] undertook visual prompt tuning on CLIP, focusing the prompting on the vision branch. MaPLe [12] investigated the effectiveness of multi-modal prompt learning in order to improve alignment between vision and language representations. Zhang et al. [37] adopted a set of learnable adaption prompts and presuppend them to the word tokens at higher transformer layers, efficiently fine-tuning LLaMA with less cost. Furthermore, in the context of WSI classification, prompts function as valuable adjuncts, enriching contextual information and semantic interpretation. The strategic utilization of prompts substantially improved model performance [23]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.713, + 0.593, + 0.728 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.739, + 0.611, + 0.755 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Fig. 2 shows the pipeline of our proposed FiVE method. To initiate the process, we collaborate with professional pathologists to establish a set of standards. Following this, we employ GPT-4 to automatically extract and standardize information based on these various standards. During the training phase, we construct Fine-Grained Guidance by intricately dividing and reconstructing the text description labels and manual prompts in pairs. The combination of manual prompts and learnable prompts forms the Diagnosis" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "11400" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.089, + 0.891, + 0.248 + ], + "angle": 0, + "content": "
Manual-Designed StandardsFine-Grained Text Description Label Examples
1. What is the differentiation of the lesion?TCGA-44-6774: Lesion differentiation is moderately to poorly differentiated; Unknown; No indication of vascular invasion by the lesion; No indication of pleural invasion by the lesion; Unknown; Margins of the excised tissue are clear of disease.
2. Is there any indication of spread through air spaces around the lesion?
3. Is there any indication of vascular invasion by the lesion?
4. Is there any indication of pleural invasion by the lesion?
5. Is there any evidence of the lesion invading adjacent tissues or organs?TCGA-49-4505: Lesion differentiation is well-differentiated; Unknown; Unknown; Pleural invasion by the lesion is present, as the carcinoma extends through the visceral pleura; The lesion invades adjacent tissues or organs; Margins of the excised tissue are clear of disease.
6. Are the margins of the excised tissue clear of disease?
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.258, + 0.893, + 0.287 + ], + "angle": 0, + "content": "Table 1. Manual-Designed Standards and Fine-Grained Text Description Label Examples. The answers on the right correspond to the standards on the left. \"Unknown\" is used as a placeholder when relevant information cannot be found." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.313, + 0.471, + 0.435 + ], + "angle": 0, + "content": "Prompts, which are utilized to enhance the semantic relevance of the features. Subsequently, the instance aggregator module fuses instance features with fine-grained prompts, generating bag-level features, subsequently align with the corresponding fine-grained text description labels. Additionally, to reduce computational costs, we implement the patch sampling strategy, optimizing the model's training efficiency while minimizing performance loss." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.468, + 0.358, + 0.484 + ], + "angle": 0, + "content": "3.2. Text Standardization via GPT-4" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.5, + 0.47, + 0.681 + ], + "angle": 0, + "content": "We utilize fine-grained text description labels extracted from pathology reports to align image bag-level features. Though pathology reports are readily accessible from public databases, their content exhibits significant variability depending on the source. Despite these format differences, pathology diagnoses consistently adhere to specific and well-established diagnostic standards. In our work, we develop fine-grained diagnostic criteria under the guidance of professional pathologists to standardize report data and extract fine-grained insights pertinent to pathological diagnosis, aiming to enhance its generalization capabilities substantially." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.69, + 0.471, + 0.901 + ], + "angle": 0, + "content": "The manual-designed standards aim to extract the morphological characteristics under the microscope, such as the degree of differentiation and lesion invasion, and filter out information irrelevant to the diagnosis. Subsequently, we employ GPT-4 to automatically extract answers from the original diagnosis reports based on prompts composed of these standards. If the information queried is absent in the pathology reports, \"Unknown\" is used as the answer. Tab. 1 shows manual-designed standards and two fine-grained text description label examples. Then, we recombine these extracted fine-grained information and integrate it into a complete description of the case image. More details about the prompts used for Text Standardization are provided in Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.312, + 0.884, + 0.329 + ], + "angle": 0, + "content": "3.3. Task-specific Fine-grained Semantics Module" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.338, + 0.811, + 0.354 + ], + "angle": 0, + "content": "3.3.1 Fine-grained Guidance Construction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.367, + 0.892, + 0.457 + ], + "angle": 0, + "content": "Due to the Text Standardization process, our data has achieved standardization. Utilizing these fine-grained text description directly as training labels can yield performance improvements. Additionally, leveraging them to generate more diverse and semantically enriched fine-grained guidance can further boost the model's performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.461, + 0.893, + 0.806 + ], + "angle": 0, + "content": "During the training process we utilize these manual-designed standards as our manual-designed prompts. We divide the original fine-grained text descriptions into several parts according to manual-designed prompts, followed by random sampling and eliminating \"Unknown\" tags from the initial labels. Given that the staged diagnostic reports in pathological descriptions are sequence-independent, we shuffle these preliminary labels and reconstruct them into a full-sentence description. During the training, we train reconstructed text description labels and reconstructed manual prompts in pairs. For example, consider description A: \"Lesion differentiation is moderately differentiated; Margins of the excised tissue are clear of disease.\" and description B: \"Lesion differentiation is moderately differentiated; Margins of the excised tissue are not clear of disease, as the tumor is within the bronchial margin and parenchymal margin.\" When only the first part of each description is sampled, they would be grouped into the same category. However, when sampling the entire sections, they are considered as distinct descriptions. Changes in granularity provide diverse perspectives on the visual image, aligning visual image with text descriptions of varying granularities." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This strategy offers three key benefits: 1) Effectively alleviating the parent-child relationship in pathology categories. 2) Providing additional hierarchical semantic perspectives to enhance the text encoder's semantic comprehension ability. 3) Mitigating discrepancies in category annotation due to incomplete diagnostic information." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11401" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.092, + 0.264, + 0.108 + ], + "angle": 0, + "content": "3.3.2 Diagnosis Prompts" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.115, + 0.47, + 0.235 + ], + "angle": 0, + "content": "We introduce Diagnosis Prompts to guide the aggregation of instance features into bag-level features. We compute the similarity between the instance features and the given manual prompts, utilizing the similarity scores as weights \\( W \\) for feature aggregation to improve the task-specific relevance of the features. Here we utilize the identical manual prompts as those used to standardize the raw data, as shown in the left of the Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.237, + 0.471, + 0.508 + ], + "angle": 0, + "content": "In addition, manual-designed prompts may have some flaws, potentially failing to comprehensively capture the specific morphological characteristics of the lesion, and the model struggles to generalize towards unseen classes due to the late fusion through the transformer layers. Besides, fine-tuning the model may not always be feasible as it requires training a large number of parameters. Particularly in the case of low-data regimes, where the availability of training data like whole slide images is extremely limited. LLaMA-Adapter [37] and LLaMA-Adapter-v2 [6] explore the way to efficient fine-tuning of Language Models and Vision-Language Models respectively. These approaches introduced the Adaptation Prompt to gradually acquire instructional knowledge. They adopted zero-initialized attention with gating mechanisms to ensure stable training in the early stages. Inspired by these methods, we introduce learnable continuous diagnosis prompts to enrich the context information and enhance the model's transferability." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.509, + 0.47, + 0.66 + ], + "angle": 0, + "content": "Specifically, we get the manual text prompt tokens \\( Q_{h} = [q_{h1}, q_{h2}, \\dots, q_{hn}] \\), here \\( n \\) represents the number of the manual prompts. We concatenate the learnable continuous prompt tokens \\( Q_{l} = [q_{l1}, q_{l2}, \\dots, q_{lm}] \\) on it, here \\( m \\) represents the number of the learnable prompt tokens. Finally we get the diagnosis prompts \\( Q = [q_{h1}, q_{h2}, \\dots, q_{hn}, q_{l1}, q_{l2}, \\dots, q_{lm}] \\). In the training phase, part of manual prompt tokens \\( Q_{h} \\) will be sampled randomly paired with text description labels, while the whole learnable prompt tokens \\( Q_{l} \\) will be consistently retained." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.471, + 0.782 + ], + "angle": 0, + "content": "Different with the traditional context learning prompts method, our approach pays attention to the acquisition of prior knowledge, similar to the methodology employed in Detection Transformer (DETR) [4]. We aim to acquire a set of appropriate query values to improve performance in subsequent feature screening processes. Additionally, it can also enable the model to quickly transfer to other tasks by fine-tuning this set of queries." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.801, + 0.332, + 0.816 + ], + "angle": 0, + "content": "3.3.3 Instance Aggregator Module" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.884 + ], + "angle": 0, + "content": "The Instance Aggregator (IA) module is used to aggregate the fine-grained diagnosis prompts and instance features. As shown in the right of the Fig. 2, IA consists of a self-attention module and a cross-attention module." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We employ self-attention to enable feature interaction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.182 + ], + "angle": 0, + "content": "among instance features \\( I_{i} = [e_{i1}, e_{i2}, \\dots, e_{ij}] \\), resulting in the feature \\( s_{i} \\). Subsequently, utilizing the diagnosis prompts \\( Q \\) to aggregate the instance features and acquire the feature \\( z_{i} \\). Then we concatenate \\( s_{i} \\) and \\( z_{i} \\), utilizing the learnable parameter \\( W \\) to fuse these features, yielding the bag-level feature \\( v_{i} \\). The formulas are shown as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.193, + 0.892, + 0.21 + ], + "angle": 0, + "content": "\\[\ns _ {i} = \\text {S e l f A t t e n t i o n} \\left(I _ {i}, I _ {i}\\right) + I _ {i} \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.213, + 0.892, + 0.228 + ], + "angle": 0, + "content": "\\[\nz _ {i} = \\text {C r o s s A t t e n t i o n} (Q, s _ {i}) \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.562, + 0.231, + 0.892, + 0.248 + ], + "angle": 0, + "content": "\\[\nv _ {i} = \\operatorname {c o n c a t} \\left(\\operatorname {m e a n} \\left(s _ {i}\\right), \\operatorname {m e a n} \\left(z _ {i}\\right)\\right) \\cdot W \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.256, + 0.892, + 0.301 + ], + "angle": 0, + "content": "Ultimately, we acquire the image bag-level features guided by the fine-grained diagnosis prompts, which are then employed to align the fine-grained text features." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.309, + 0.712, + 0.325 + ], + "angle": 0, + "content": "3.4. Patch Sample Strategy" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.332, + 0.892, + 0.498 + ], + "angle": 0, + "content": "Each Whole Slide Image (WSI) is partitioned into a variable number of instances, ranging from approximately 50 to 45,000. Handling such a wide range of instances markedly increases computational complexity and substantially extends the training duration. FLIP [33] reduced computation and reached higher accuracy than CLIP [25] counterpart by randomly removing a large portion of image patches during training [17]. In the case of whole slide images, pathological visual patterns are often redundantly distributed across a tissue slice. Therefore, it is feasible to sample only a subset of visual instances during training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.498, + 0.892, + 0.559 + ], + "angle": 0, + "content": "For the instances in each bag (i.e., slide), we select a sample amounting to \\( S_{m} \\) percent of the total number \\( p \\) of each group of instances (patches). The required number of instances \\( S_{n} \\) is described by the following formula:" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.567, + 0.892, + 0.584 + ], + "angle": 0, + "content": "\\[\nS _ {n} = \\min \\left(p * S _ {m}, S _ {\\text {m a x n}}\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.592, + 0.892, + 0.682 + ], + "angle": 0, + "content": "Here \\( S_{maxn} \\) denotes the maximum number of sample instances. To achieve this, we evenly divide each group of instances into \\( S_{n} \\) chunks and randomly select one instance from each chunk. Since the different whole slide images could sample different \\( S_{n} \\), we pad the rest of the space with the same padding." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.691, + 0.748, + 0.705 + ], + "angle": 0, + "content": "3.5. Encoder and Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.713, + 0.892, + 0.834 + ], + "angle": 0, + "content": "We divide each WSI into instances \\( x_{k} \\) and encode these instances into embeddings \\( e_{k} \\in R^{D} \\) using pre-trained vision encoder \\( E_{img} \\), composed with ResNet structure following [15], here \\( D \\) represents the dimension of the embeddings. Then we send the instance embeddings into the TFS Module to aggregate the instance features and prompts, and obtain the bag-level embeddings \\( v_{i} \\in R^{D} \\). The formulas are shown as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.843, + 0.892, + 0.86 + ], + "angle": 0, + "content": "\\[\ne _ {k} = E _ {i m g} \\left(x _ {k}\\right) \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.862, + 0.892, + 0.879 + ], + "angle": 0, + "content": "\\[\nI _ {i} = \\left[ e _ {i 1}, e _ {i 2}, \\dots , e _ {i k} \\right] \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.881, + 0.892, + 0.898 + ], + "angle": 0, + "content": "\\[\nv _ {i} = I A \\left(I _ {i}, Q\\right) \\tag {7}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11402" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.138 + ], + "angle": 0, + "content": "Besides, we generate pathologically meaningful text embeddings for each WSI, represented as \\( t_i \\in R^D \\), by leveraging the fine-tuned text encoder BioClinicalBERT [30]." + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.149, + 0.469, + 0.166 + ], + "angle": 0, + "content": "\\[\nt _ {i} ^ {c} = E _ {t x t} \\left(x _ {t x t} ^ {c}\\right) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.176, + 0.469, + 0.266 + ], + "angle": 0, + "content": "where \\( E_{txt} \\) denotes the text encoder, and \\( x_{txt}^{c}(c \\in [1, C]) \\) where \\( C \\) denotes the number of categories. Here we use the same embedding dimension \\( D \\) as the vision encoder, suitable for contrastive learning. For text encoder \\( E_{txt} \\), we adopt the Low-Rank Adaptation (LoRA) [7] approach for efficient fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.267, + 0.469, + 0.342 + ], + "angle": 0, + "content": "Subsequently, the bag-level embeddings \\( v_{i} \\) are aligned with the text embeddings \\( t_{i}^{c} \\) to complete the training process. In this case, prediction \\( \\hat{y} \\) is obtained by applying softmax on scaled cosine similarities between the image embeddings and text embeddings:" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.352, + 0.469, + 0.39 + ], + "angle": 0, + "content": "\\[\np (\\hat {y} = c \\mid I) = \\frac {\\exp \\left(\\operatorname {s i m} \\left(t _ {i} ^ {c} , v _ {i}\\right) / \\tau\\right)}{\\sum_ {c ^ {\\prime} = 1} ^ {C} \\exp \\left(\\operatorname {s i m} \\left(t _ {i} ^ {c ^ {\\prime}} , v _ {i}\\right) / \\tau\\right)} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.399, + 0.468, + 0.429 + ], + "angle": 0, + "content": "where \\( \\text{sim}(\\cdot, \\cdot) \\) refers to cosine similarity and \\( \\tau \\) is the temperature parameter." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.43, + 0.468, + 0.459 + ], + "angle": 0, + "content": "The fine-grained training loss is computed as the cross-entropy between the logits and soft targets as:" + }, + { + "type": "equation", + "bbox": [ + 0.161, + 0.47, + 0.469, + 0.514 + ], + "angle": 0, + "content": "\\[\nL ^ {v \\rightarrow t} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {N} y _ {i j} \\log \\left(p _ {i j}\\right) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.523, + 0.325, + 0.537 + ], + "angle": 0, + "content": "here \\(N\\) corresponds to the batch size." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.538, + 0.469, + 0.569 + ], + "angle": 0, + "content": "Likewise, we can compute \\( L^{t \\to v} \\) and serve \\( L \\) as the final training objective." + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.579, + 0.469, + 0.61 + ], + "angle": 0, + "content": "\\[\nL = \\frac {L ^ {v \\rightarrow t} + L ^ {t \\rightarrow v}}{2} \\tag {11}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.619, + 0.312, + 0.635 + ], + "angle": 0, + "content": "4. Experiments and Results" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.644, + 0.18, + 0.658 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.667, + 0.469, + 0.711 + ], + "angle": 0, + "content": "We evaluated our method on public histopathology WSI datasets: The Cancer Genome Atlas Lung (TCGA Lung) Cancer1 and Camelyon16 [2]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.712, + 0.469, + 0.878 + ], + "angle": 0, + "content": "TCGA Lung Cancer. The TCGA Lung Cancer dataset comprises two cancer subtypes: Lung Adenocarcinoma (LUAD) and Lung Squamous Cell Carcinoma (LUSC). It includes diagnostic slides with 541 slides from 478 LUAD cases and 512 slides from 478 LUSC cases. For WSI preprocessing, following [15], we cropped each WSI into \\(256 \\times 256\\) non-overlapping patches and removed the background region. The dataset encompasses approximately 5.2 million patches at \\(20 \\times\\) magnification, averaging about 5,000 patches per WSI. Following [28], we randomly split the dataset into training, validation, and testing sets with a ratio" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.364 + ], + "angle": 0, + "content": "of 65:10:25 on the patient level and adopted 4-fold cross-validation. We collected corresponding pathology reports \\( {}^{2} \\) and cleaned them by GPT-4 to produce fine-grained text description labels. To ensure professionalism and accuracy, we invited professional pathologists to check and correct the textual labels. Additionally, to evaluate the generalizability of the model and perform zero-shot classification, we utilized a dataset of histological subtype labels for TCGALUAD from the cBioPortal database \\( {}^{3} \\) ,More details about subtype labels are provided in Supplementary Material. Camelyon16. The Camelyon16 dataset [2] consists of 399 Hematoxylin and Eosin (H&E) stained slide images, utilized for metastasis detection in breast cancer. We preprocessed each WSI by segmenting it into \\( {256} \\times {256} \\) nonoverlapping patches, excluding background regions. In total, this process yielded approximately 2.8 million patches at a \\( {20} \\times \\) magnification level,with about 7,200 patches per bag. We adopted 3-times 3-fold cross-validation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.373, + 0.719, + 0.389 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.397, + 0.892, + 0.578 + ], + "angle": 0, + "content": "In experiments, we employed ResNet following [15] as image encoder to extract image features, while pre-trained BioClinicalBERT from [30] as text encoder to generate text features. LoRA [7] was adopted for fine-tuning the text encoder, with an alpha value of 32 and a rank value of 8. We divided the whole slide image into patches with \\(256 \\times 256\\), then applied a random crop with size \\(224 \\times 224\\). We adopted AdamW [19] with beta (0.9, 0.98), eps 1e-8, learning rate 3e-6, warmup ratio 0.1, weight decay 1e-4 as our optimizer. Additionally, we used the batch size of 1 with an accumulation step of 8 and trained for 150 epochs. We utilized mixed-precision training on 4 NVIDIA-A800 GPUs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.588, + 0.884, + 0.604 + ], + "angle": 0, + "content": "4.3. Zero-Shot Histological Subtype Classification" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.611, + 0.891, + 0.702 + ], + "angle": 0, + "content": "Prior researches [23, 28, 38] have predominantly concentrated on classifying primary cancer categories. Our approach extends beyond this by attempting to classify detailed histological subtypes. It is crucial to emphasize that this task poses a significant challenge, often proving difficult for even skilled pathologists to make direct judgments." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.702, + 0.892, + 0.853 + ], + "angle": 0, + "content": "Since only the LUAD's histological subtype dataset was provided on the online database, we conducted zero-shot subtype classification evaluation on LUAD subtype datasets, with the model being pre-trained on TCGA-LUAD or TCGA-LUSC. Throughout the training phase, subtype label information was deliberately excluded, ensuring that all experiments are conducted solely with fine-grained labels. This aims to evaluate the model's ability to identify novel diagnostic categories without specific training on subtypes. We extended the morphological appearance text de" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.309, + 0.9 + ], + "angle": 0, + "content": "1http://www.cancer.gov/ctga" + }, + { + "type": "page_footnote", + "bbox": [ + 0.501, + 0.863, + 0.89, + 0.887 + ], + "angle": 0, + "content": "2https://github.com/tatonetti-lab/tcga-path-reports" + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.888, + 0.74, + 0.9 + ], + "angle": 0, + "content": "3https://www.cbiportal.org/" + }, + { + "type": "list", + "bbox": [ + 0.501, + 0.863, + 0.89, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11403" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "scription labels of the target data using GPT-4. After the text encoder encodes the labels, similarity calculation with image features achieved zero-shot classification." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.139, + 0.471, + 0.32 + ], + "angle": 0, + "content": "Since existing zero-shot learning methods cannot be used in WSI subtype classification, we constructed three Linear-Probe method baselines: Mean pooling, Max pooling, and Attention pooling. As shown in Tab. 2, FiVE attains \\(65.23\\%\\) top-1 accuracy and \\(95.18\\%\\) top-5 accuracy when pre-trained with TCGA-LUAD fine-grained labels. When pre-trained with TCGA-LUSC fine-grained labels, it achieves \\(62.02\\%\\) top-1 accuracy and \\(94.36\\%\\) top-5 accuracy. Moreover, the zero-shot performance of FiVE notably exceeds that of the baseline. This capability in classifying LUAD subtypes is attributed to the focus on fine-grained pathological features during training." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.336, + 0.462, + 0.472 + ], + "angle": 0, + "content": "
MethodTCGA-LUADTCGA-LUSC
Top-1Top-5Top-1Top-5
Mean-pooling40.8283.4631.8682.96
Max-pooling45.0588.7736.3686.09
Attention-pooling58.3693.5354.4192.50
FiVE (Ours)65.2395.1862.0294.36
+6.87+1.65+7.61+1.86
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.482, + 0.47, + 0.539 + ], + "angle": 0, + "content": "Table 2. Zero-Shot performance on histological subtype classification. We pre-trained the model using fine-grained labels of TCGA-LUAD and TCGA-LUSC, then applied zero-shot classification to the histological subtypes of TCGA-LUAD." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.571, + 0.295, + 0.587 + ], + "angle": 0, + "content": "4.4. Few-Shot Classification" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.597, + 0.47, + 0.748 + ], + "angle": 0, + "content": "Our model demonstrates adaptability to various tasks even in scenarios with limited data availability. Few-shot experiments were conducted to demonstrate its transferability to downstream tasks. We initialized the networks with pretrained weights derived from the model trained on TCGA image-report pairs, and subsequently fine-tuned the model on downstream datasets for few-shot image classification. We followed [23] and conducted experiments with 1, 2, 4, 8, 16, and additional 0 shot on the downstream dataset. The results are summarized in Tab. 3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Our model exhibits remarkable performance in zero-shot classification, achieving an accuracy of \\(71.26\\%\\), even surpassing the SOTA method's one-shot experiment. Upon the introduction of training data, our models display exceptional transferability, outperforming the SOTA by \\(12.90\\%\\) in the one-shot setting. The model's performance improves accordingly with the number of shot. Upon reaching 16-shot, our model reaches an impressive accuracy of \\(91.25\\%\\), showcasing a notable \\(9.19\\%\\) improvement close to the fully supervised performance level." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.089, + 0.89, + 0.203 + ], + "angle": 0, + "content": "
Method16-shot8-shot4-shot2-shot1-shot0-shot
Mean-pool65.3353.8944.8552.9345.34\\
Max-pool48.4849.5544.2248.3949.03\\
Attn-pool72.5065.7962.4758.3656.23\\
CoOp [40]78.3567.9967.6067.5467.81\\
TOP [23]82.0680.5175.4172.3871.01\\
FiVE91.2590.8088.1085.5183.9171.26
(Ours)+9.19+10.29+12.69+13.13+12.90
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.213, + 0.894, + 0.271 + ], + "angle": 0, + "content": "Table 3. Few-shot classification performance on TCGA Lung Cancer. Mean-pool, Max-pool, and Attn-pool correspond to Linear-Probe implementations with Mean-pooling, Max-pooling, and Attention-pooling, respectively." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.298, + 0.892, + 0.315 + ], + "angle": 0, + "content": "4.5. Performance Comparison with Existing Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.322, + 0.892, + 0.444 + ], + "angle": 0, + "content": "We compared FiVE with ABMIL [10], DSMIL [15], CLAM-SB [20], CLAM-MB [20], TransMIL [26], DTFDMIL [36], and MHIM-MIL [28], all of which are attention-based MIL methods. In addition, we included two traditional MIL pooling operations, Max-pooling and Mean-pooling, for comparison. The results of all other methods are reproduced using the official code they provide under the same settings." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.445, + 0.893, + 0.656 + ], + "angle": 0, + "content": "To evaluate the performance of FiVE, we conducted fine-grained pre-training exclusively on the TCGA dataset, followed by fully supervised experiments on Camelyon16 and TCGA Lung Cancer datasets for WSI classification. Note that the test data is not used in pre-training. Results are shown in Tab. 4. It can be found that our method outperforms all the other baselines by a great margin, which fully demonstrates the significance of our fine-grained training scheme in improving performance on downstream tasks. Besides, since the pre-training data mainly comes from TCGA data, on the other hand, the task-specific prompt design is more suitable for TCGA data. This results in the performance improvement of our method for TCGA data being greater than Camelyon16 data." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.669, + 0.665, + 0.685 + ], + "angle": 0, + "content": "4.6. Ablation Studies" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.694, + 0.755, + 0.709 + ], + "angle": 0, + "content": "4.6.1 Effectiveness of TFS Module" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.894, + 0.901 + ], + "angle": 0, + "content": "We focused on evaluating the impact of the TFS module on the model's overall performance with the TCGA Lung Cancer. The results detailed in Tab. 5 reveal significant improvements at each stage of feature enhancement. Initially, the model with only Self Attention attained ACC of \\(89.77\\%\\), AUC of \\(92.85\\%\\), and F1-score of \\(89.95\\%\\). The incorporation of fine-grained labels led to increases of \\(1.44\\%\\) in ACC, \\(1.51\\%\\) in AUC, and \\(1.43\\%\\) in F1-score. Subsequent integration of additional fine-grained guidance further improved performance. Ultimately, full framework with the TFS Module achieved the highest performance of \\(94.62\\%\\) ACC, \\(96.33\\%\\) AUC, and \\(93.89\\%\\) F1-score. These ablation" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "11404" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.127, + 0.089, + 0.847, + 0.293 + ], + "angle": 0, + "content": "
MethodCamelyon16TCGA Lung Cancer
ACCAUCF1-scoreACCAUCF1-score
Max-pooling78.95±2.2881.28±3.7471.06±2.5981.49±1.2486.45±0.7180.56±1.09
Mean-pooling76.69±0.2080.07±0.7870.41±0.1684.14±2.9790.13±2.4083.39±3.14
ABMIL [10]90.06±0.6094.00±0.8387.40±1.0588.03±2.1993.17±2.0587.41±2.42
DSMIL [15]90.17±1.0294.57±0.4087.65±1.1888.32±2.7093.71±1.8287.90±2.50
CLAM-SB [20]90.32±0.1294.65±0.3087.89±0.5987.74±2.2293.67±1.6487.36±2.24
CLAM-MB [20]90.14±0.8594.70±0.7688.10±0.6388.73±1.6293.69±0.5488.28±1.58
TransMIL [26]89.22±2.3293.51±2.1385.10±4.3387.08±1.9792.51±1.7686.40±2.08
DTFD-MIL [36]90.22±0.3695.15±0.1487.62±0.5988.23±2.1293.83±1.3987.71±2.04
MHIM-MIL [28]92.48±0.3596.49±0.6590.75±0.7389.93±3.3795.53±1.7489.71±2.92
FiVE (Ours)94.25±0.3397.56±0.5993.24±0.7294.62±2.1396.33±1.2193.89±1.90
" + }, + { + "type": "table_caption", + "bbox": [ + 0.193, + 0.303, + 0.776, + 0.318 + ], + "angle": 0, + "content": "Table 4. Comparison Performance of slide classification on Camelyon16 and TCGA Lung Cancer." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.34, + 0.468, + 0.422 + ], + "angle": 0, + "content": "
SAFGLFGGLDPACCAUCF1-score
✓89.7792.8589.58
✓✓91.2194.3691.01
✓✓✓93.5696.0193.17
✓✓✓✓94.6296.3393.89
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.431, + 0.47, + 0.488 + ], + "angle": 0, + "content": "Table 5. The ablation experiments of the pre-trained model finetuned on the TCGA Lung Cancer. SA, FGL, FGG, and LDP represent Self Attention, Fine-Grained Labels, Fine-Grained Guidance, and Learnable Diagnosis Prompts, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.503, + 0.47, + 0.534 + ], + "angle": 0, + "content": "experiments highlight the benefits of integrating these methods in the WSI classification task." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.554, + 0.403, + 0.57 + ], + "angle": 0, + "content": "4.6.2 Effectiveness of Patch Sample Strategy" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.579, + 0.47, + 0.686 + ], + "angle": 0, + "content": "We verified the impact of different sampling strategies on model performance on TCGA data. Here, we assumed that there are two main indicators that affect model performance, sample ratio and max sample threshold (represented as MAXN). At the same time, in order to differentiate the experimental results, we used the unfrozen image encoder for experimental verification." + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.7, + 0.27, + 0.791 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.7, + 0.465, + 0.791 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.796, + 0.468, + 0.838 + ], + "angle": 0, + "content": "Figure 3. Classification performance on TCGA Lung Cancer with diverse sampling strategies, presenting average and standard deviation (std) ACC values." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.903 + ], + "angle": 0, + "content": "As shown in Fig. 3, when MAXN is sufficiently large (\\(\\geq\\) 2048), the correlation between the model's performance and the sample ratio is not significant. Even with a small sample" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.344, + 0.892, + 0.48 + ], + "angle": 0, + "content": "ratio \\((\\leq 0.5)\\), the model can still effectively align with the original data distribution. Conversely, when MAXN is not large enough (<2048), the primary change in the model's performance depends on whether the magnitude of the sample ratio can match the original data distribution. When the sample ratio increases to 0.75, a steep performance improvement can be observed, suggesting that the sampled data at this point conforms to the original data distribution, after which the model's performance stabilizes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.481, + 0.893, + 0.603 + ], + "angle": 0, + "content": "Based on comprehensive experimental results, we established the sample ratio and MAXN as 0.5 and 2048 as suitable hyperparameters for the unfrozen image encoder. As for the frozen image encoder experiment, taking into account the performance constraints imposed by the frozen image encoder on the model's performance upper limit [38], we recommend the sample ratio and MAXN to be set at 0.5 and 16384 based on our experiments." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.618, + 0.619, + 0.634 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this paper, we introduce FiVE, a novel framework that demonstrates robust generalization and strong transferability for WSI classification. Our work pioneers the utilization of non-standardized WSI-report pairs from public databases to develop a VLM. To capture the complexities and diversity within these reports, we introduce the Task-specific Fine-grained Semantics (TFS) module. This module reconstructs fine-grained labels and diagnosis prompts during training, enhancing the semantic relevance of its features by introducing diagnosis prompts. Furthermore, considering that pathological visual patterns are redundantly distributed across tissue slices, we employ a sampling strategy to reduce computational costs. Our experiments demonstrate the robust generalizability and computational efficiency of the proposed framework, which can also be easily adapted to other tasks with minimal fine-tuning. We aspire to provide empirical insights and contribute to AI pathology research." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11405" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.169 + ], + "angle": 0, + "content": "[1] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Visual prompting: Modifying pixel space to adapt pre-trained models. arXiv preprint arXiv:2203.17274, 3:11-12, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.269 + ], + "angle": 0, + "content": "[2] Babak Ehteshami Bejnordi, Mitko Veta, Paul Johannes Van Diest, Bram Van Ginneken, Nico Karssemeijer, Geert Litjens, Jeroen AWM Van Der Laak, Meyke Hermsen, Quirine F Manson, Maschenka Balkenhol, et al. Diagnostic assessment of deep learning algorithms for detection of lymph node metastases in women with breast cancer. Jama, 318(22):2199-2210, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.272, + 0.47, + 0.354 + ], + "angle": 0, + "content": "[3] Gabriele Campanella, Matthew G Hanna, Luke Geneslaw, Allen Miraflor, Vitor Werneck Krauss Silva, Klaus J Busam, Edi Brogi, Victor E Reuter, David S Klimstra, and Thomas J Fuchs. Clinical-grade computational pathology using weakly supervised deep learning on whole slide images. Nature medicine, 25(8):1301-1309, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.357, + 0.47, + 0.413 + ], + "angle": 0, + "content": "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.415, + 0.47, + 0.485 + ], + "angle": 0, + "content": "[5] Tsai Hor Chan, Fernando Julio Cendra, Lan Ma, Guosheng Yin, and Lequan Yu. Histopathology whole slide image analysis with heterogeneous graph representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15661-15670, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.486, + 0.47, + 0.554 + ], + "angle": 0, + "content": "[6] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xi-angyu Yue, et al. Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.557, + 0.47, + 0.613 + ], + "angle": 0, + "content": "[7] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.615, + 0.47, + 0.671 + ], + "angle": 0, + "content": "[8] Zhi Huang, Federico Bianchi, Mert Yuksekgonul, Thomas J Montine, and James Zou. A visual-language foundation model for pathology image analysis using medical twitter. Nature Medicine, pages 1-10, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.673, + 0.47, + 0.713 + ], + "angle": 0, + "content": "[9] Carolyn Hutter and Jean Claude Zenklusen. The cancer genome atlas: creating lasting value beyond its data. Cell, 173(2):283-285, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.716, + 0.47, + 0.771 + ], + "angle": 0, + "content": "[10] Maximilian Ilse, Jakub Tomczak, and Max Welling. Attention-based deep multiple instance learning. In International conference on machine learning, pages 2127-2136. PMLR, 2018. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.47, + 0.856 + ], + "angle": 0, + "content": "[11] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[12] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19113-19122, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.206 + ], + "angle": 0, + "content": "[13] Zhengfeng Lai, Zhuoheng Li, Luca Cerny Oliveira, Joohi Chauhan, Brittany N Dugger, and Chen-Nee Chuah. Clipse: Fine-tune clip with visual feature fusion for pathology image analysis towards minimizing data collection efforts. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2374-2380, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[14] Marvin Lerousseau, Maria Vakalopoulou, Marion Classe, Julien Adam, Enzo Battistella, Alexandre Carré, Théo Estienne, Théophraste Henry, Eric Deutsch, and Nikos Paragios. Weakly supervised multiple instance learning histopathological tumor segmentation. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part V 23, pages 470-479. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.32, + 0.892, + 0.401 + ], + "angle": 0, + "content": "[15] Bin Li, Yin Li, and Kevin W Eliceiri. Dual-stream multiple instance learning network for whole slide image classification with self-supervised contrastive learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14318-14328, 2021. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.404, + 0.892, + 0.5 + ], + "angle": 0, + "content": "[16] Honglin Li, Chenglu Zhu, Yunlong Zhang, Yuxuan Sun, Zhongyi Shui, Wenwei Kuang, Sunyi Zheng, and Lin Yang. Task-specific fine-tuning via variational information bottleneck for weakly-supervised pathology whole slide image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7454-7463, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.502, + 0.892, + 0.571 + ], + "angle": 0, + "content": "[17] Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. Scaling language-image pre-training via masking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23390-23400, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.574, + 0.892, + 0.643 + ], + "angle": 0, + "content": "[18] Tiancheng Lin, Zhimiao Yu, Hongyu Hu, Yi Xu, and Chang-Wen Chen. Interventional bag multi-instance learning on whole-slide pathological images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19830-19839, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.645, + 0.892, + 0.672 + ], + "angle": 0, + "content": "[19] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.674, + 0.892, + 0.743 + ], + "angle": 0, + "content": "[20] Ming Y Lu, Drew FK Williamson, Tiffany Y Chen, Richard J Chen, Matteo Barbieri, and Faisal Mahmood. Data-efficient and weakly supervised computational pathology on whole-slide images. Nature biomedical engineering, 5(6):555-570, 2021. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.745, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[21] Ming Y Lu, Bowen Chen, Drew FK Williamson, Richard J Chen, Ivy Liang, Tong Ding, Guillaume Jaume, Igor Odintsov, Andrew Zhang, Long Phi Le, et al. Towards a visual-language foundation model for computational pathology. arXiv preprint arXiv:2307.12914, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.816, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[22] Yuning Lu, Jianzhuang Liu, Yonggang Zhang, Yajing Liu, and Xinmei Tian. Prompt distribution learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5206-5215, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[23] Linhao Qu, Xiaoyuan Luo, Kexue Fu, Manning Wang, and Zhijian Song. The rise of ai language pathologists:" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11406" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "Exploring two-level prompt learning for few-shot weakly-supervised whole slide image classification. arXiv preprint arXiv:2305.17891, 2023. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.135, + 0.472, + 0.218 + ], + "angle": 0, + "content": "[24] Linhao Qu, Zhiwei Yang, Minghong Duan, Yingfan Ma, Shuo Wang, Manning Wang, and Zhijian Song. Boosting whole slide image classification from the perspectives of distribution, correlation and magnification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21463-21473, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.471, + 0.3 + ], + "angle": 0, + "content": "[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.301, + 0.471, + 0.37 + ], + "angle": 0, + "content": "[26] Zhuchen Shao, Hao Bian, Yang Chen, Yifeng Wang, Jian Zhang, Xiangyang Ji, et al. Transmil: Transformer based correlated multiple instance learning for whole slide image classification. Advances in neural information processing systems, 34:2136-2147, 2021. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.371, + 0.471, + 0.455 + ], + "angle": 0, + "content": "[27] Zhuchen Shao, Yifeng Wang, Yang Chen, Hao Bian, Shaohui Liu, Haoqian Wang, and Yongbing Zhang. Lnpl-mil: Learning from noisy pseudo labels for promoting multiple instance learning in whole slide image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21495-21505, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.471, + 0.537 + ], + "angle": 0, + "content": "[28] Wenhao Tang, Sheng Huang, Xiaoxian Zhang, Fengtao Zhou, Yi Zhang, and Bo Liu. Multiple instance learning framework with masked hard instance mining for whole slide image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4078-4087, 2023. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.538, + 0.471, + 0.58 + ], + "angle": 0, + "content": "[29] Xinggang Wang, Yongluan Yan, Peng Tang, Xiang Bai, and Wenyu Liu. Revisiting multiple instance neural networks. Pattern Recognition, 74:15-24, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.58, + 0.471, + 0.65 + ], + "angle": 0, + "content": "[30] Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chau-mond, Clement Delangue, Anthony Moi, Pierrick Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, et al. Huggingface's transformers: State-of-the-art natural language processing. arXiv preprint arXiv:1910.03771, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.471, + 0.705 + ], + "angle": 0, + "content": "[31] Jinxi Xiang and Jun Zhang. Exploring low-rank property in multiple instance learning for whole slide image classification. In The Eleventh International Conference on Learning Representations, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.471, + 0.789 + ], + "angle": 0, + "content": "[32] Gang Xu, Zhigang Song, Zhuo Sun, Calvin Ku, Zhe Yang, Cancheng Liu, Shuhao Wang, Jianpeng Ma, and Wei Xu. Camel: A weakly supervised learning framework for histopathology image segmentation. In Proceedings of the IEEE/CVF International Conference on computer vision, pages 10682-10691, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.471, + 0.845 + ], + "angle": 0, + "content": "[33] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. arXiv preprint arXiv:2111.07783, 2021. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[34] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[35] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18123-18133, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[36] Hongrun Zhang, Yanda Meng, Yitian Zhao, Yihong Qiao, Xiaoyun Yang, Sarah E Coupland, and Yalin Zheng. Dtdmil: Double-tier feature distillation multiple instance learning for histopathology whole slide image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18802-18812, 2022. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.26, + 0.892, + 0.33 + ], + "angle": 0, + "content": "[37] Renrui Zhang, Jiaming Han, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, Peng Gao, and Yu Qiao. Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.33, + 0.892, + 0.415 + ], + "angle": 0, + "content": "[38] Yunkun Zhang, Jin Gao, Mu Zhou, Xiaosong Wang, Yu Qiao, Shaoting Zhang, and Dequan Wang. Text-guided foundation model adaptation for pathological image classification. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 272-282. Springer, 2023. 1, 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.415, + 0.892, + 0.485 + ], + "angle": 0, + "content": "[39] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.486, + 0.892, + 0.541 + ], + "angle": 0, + "content": "[40] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 3, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.541 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "11407" + } + ] +] \ No newline at end of file diff --git a/2024/Generalizable Whole Slide Image Classification with Fine-Grained Visual-Semantic Interaction/c4c5691b-0982-48b4-b41b-bc14bcf8455f_origin.pdf b/2024/Generalizable Whole Slide Image Classification with Fine-Grained Visual-Semantic Interaction/c4c5691b-0982-48b4-b41b-bc14bcf8455f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..29a4573e4e710238a7aff234516b6acadd449e68 --- /dev/null +++ b/2024/Generalizable Whole Slide Image Classification with Fine-Grained Visual-Semantic Interaction/c4c5691b-0982-48b4-b41b-bc14bcf8455f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a0747b3cd4d299db28c8bc550d665e3a65593cbab83f80bd713e02588602e5b8 +size 578923 diff --git a/2024/Generalizable Whole Slide Image Classification with Fine-Grained Visual-Semantic Interaction/full.md b/2024/Generalizable Whole Slide Image Classification with Fine-Grained Visual-Semantic Interaction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d28c1a50b2237f68f220adef7ae03cd5cabeeeea --- /dev/null +++ b/2024/Generalizable Whole Slide Image Classification with Fine-Grained Visual-Semantic Interaction/full.md @@ -0,0 +1,316 @@ +# Generalizable Whole Slide Image Classification with Fine-Grained Visual-Semantic Interaction + +Hao Li + +Wenxian Yang3 + +Ying Chen + +Liansheng Wang + +Yifei Chen + +Bowen Ding + +Rongshan $\mathbf{Y}\mathbf{u}^{1*}$ + +Yuchen Han\* + +$^{1}$ School of Informatics, Xiamen University $^{2}$ Huawei $^{3}$ Aginome Scientific, Xiamen, China + +$^{4}$ Department of Pathology, Shanghai Chest Hospital, Shanghai Jiao Tong University School of Medicine + +{111ih, cying2023}@stu.xmu.edu.cn, {rsyu, 1swang}@xmu.edu.cn, + +chenyifei14@huawei.com,wx@aginome.com，dingbowenmail@126.com，ychan@cmu.edu.cn + +# Abstract + +Whole Slide Image (WSI) classification is often formulated as a Multiple Instance Learning (MIL) problem. Recently, Vision-Language Models (VLMs) have demonstrated remarkable performance in WSI classification. However, existing methods leverage coarse-grained pathogenetic descriptions for visual representation supervision, which are insufficient to capture the complex visual appearance of pathogenetic images, hindering the generalizability of models on diverse downstream tasks. Additionally, processing high-resolution WSIs can be computationally expensive. In this paper, we propose a novel "Fine-grained Visual-Semantic Interaction" (FiVE) framework for WSI classification. It is designed to enhance the model's generalizability by leveraging the interaction between localized visual patterns and fine-grained pathological semantics. Specifically, with meticulously designed queries, we start by utilizing a large language model to extract fine-grained pathological descriptions from various non-standardized raw reports. The output descriptions are then reconstructed into fine-grained labels used for training. By introducing a Task-specific Fine-grained Semantics (TFS) module, we enable prompts to capture crucial visual information in WSIs, which enhances representation learning and augments generalization capabilities significantly. Furthermore, given that pathological visual patterns are redundantly distributed across tissue slices, we sample a subset of visual instances during training. Our method demonstrates robust generalizability and strong transferability, dominantly outperforming the counterparts on the TCGA Lung Cancer dataset with at least $9.19\%$ higher accuracy in few-shot experiments. The code is available at: https://github.com/ls1rius/WSI_FIVE. + +# 1. Introduction + +![](images/28b2ad7e21f2112a208d419264e6b82df4c4aedcc96471ca2744a9ba059c2cb3.jpg) +Figure 1. Challenges in WSI-text contrastive learning. Most conventional VLM approaches categorize whole slide images using category-level text descriptions, overlooking intra-class differences, leading to a decline in performance and limitations in generalization capabilities. Instead, we extract fine-grained descriptions from pathology reports as slide-level labels to develop our model, exhibiting detailed variations in each WSI. + +Histological Whole Slide Image (WSI) classification plays a crucial role in computational pathology by automating disease diagnosis and subtyping. For high-resolution WSI analysis, Multiple Instance Learning (MIL) has become the dominant method. These methods treat each WSI as a "Bag" sample, and aggregate numerous patches within it as instances for thorough decision-making. Nevertheless, most existing methods [15, 16, 18, 31] focused on processing image data to conduct WSI classification, potentially not emphasizing critical pathological insights, particularly the expert textual annotations that accompany these slides. + +Recently, Vision-Language Models (VLMs) [11, 25, 33] underscored the significance of integrating multimodal information for developing robust encoders. Zhang et al. [38] + +exploited disease-grade text labels and extracted text insights using pre-trained language models. Qu et al. [23] utilized GPT-4 in a question-and-answer mode to obtain language prior knowledge at both instance and bag levels for VLM training. However, the challenge lies in the uniqueness and variability of content in each WSI. Existing methods developed their models with coarse-grained descriptions (i.e., simplistic Category-Level text labels [38] or descriptive Category-Level text labels constructed by GPT-4 [23]), as shown in Fig. 1. They may have omitted crucial fine-grained pathological details, including differentiation level, vascular invasion, etc., which results in reduced model performance and limited generalization. + +WSIs accompanied by their corresponding reports (i.e., WSI-report pairs) offer detailed descriptions and fine-grained information vital for WSI analysis. Furthermore, a substantial collection of these pairs is accessible in public databases, such as The Cancer Genome Atlas (TCGA) [9]. However, their full potential has not been adequately harnessed yet. The challenge in developing a Visual Language Model (VLM) using WSI-report pairs mainly lies in the diverse formats and standards of the raw reports from different hospitals, which increases the complexity of data preprocessing and standardization processes. Additionally, pathology reports often contain extraneous information, including report metadata, tissue processing descriptions, and repetitive elements, which can introduce noise to the textual data. How to extract useful information from raw pathology reports to construct WSI-report pairs is a key issue. Moreover, recent studies have demonstrated the efficacy of prompt engineering in enhancing VLMs. In contrast to natural images, WSI data encompasses extensive professional pathological information and intricate details. How to craft prompts to make full use of this semantic information to guide fine-grained feature learning is a challenging task. Besides, the high computational costs to train models with high-resolution WSIs also limits the promotion of the model, resulting in a certain resource threshold for WSI analysis. + +To address these issues, we propose a novel whole slide image classification method with Fine-grained VisualsEmantic interaction termed as FiVE, which shows robust generalizability and efficiency in computation. Firstly, we obtain WSIs with non-standardized raw pathology reports from a public database. Collaborating with professional pathologists, we craft a set of specialized prompts to standardize reports. Following this, we employ the large language model GPT-4 to automatically clean and standardize the raw report data. In addition, we propose the Task-specific Fine-grained Semantic (TFS) Module, which utilizes manual-designed prompts to direct visual attention to specific pathological areas while constructing Fine-Grained Guidance to enhance the semantic relevance of model fea + +tures. Considering the diffuse distribution of pathological diseases within tissue sections and the presence of numerous non-diagnostic regions in WSIs, we also incorporate a patch sampling strategy during the training phase to enhance training efficiency and reduce computational costs. The contributions of this paper are summarized as follows: + +- We pioneer the utilization of the available WSI diagnostic reports with fine-grained guidance. The obtained fine-grained description labels lead to improved supervision by discriminating the visual appearances more precisely. +- We introduce a novel Task-specific Fine-grained Semantics (TFS) Module to offer fine-grained guidance, significantly enhancing the model's generalization capabilities. +- We implement a patch sampling strategy on visual instances during training to enhance computational efficiency without significantly compromising accuracy, thereby optimizing the model's training process. + +# 2. Related Work + +# 2.1. Whole Slide Image Analysis + +Contemporary methodologies for WSI analysis predominantly employ MIL methods where each WSI is treated as a "Bag" and its extracted patches as instances within this bag. MIL methods consist of instance-based methods [3, 14, 32] and embedding representation-based methods [15, 20, 26, 29, 36]. However, the majority of existing methods [5, 18, 24, 27] almost exclusively rely on image data, neglecting vital pathological details, notably the specialist text annotations that accompany the images. Recent works [23, 38] have taken note of this issue and started to utilize text information to improve pathological image classification. They used bag-level text labels or the descriptive labels generated by GPT. However, given the unique and varied descriptions of each WSI, their methods fall short of fully leveraging the detailed textual information present in the slides. + +# 2.2. Vision-Language Models + +Recent researches have made efforts to develop Vision-Language Models (VLMs). CLIP [25] gathered 400 million image-text pairs and initiated training with synchronized vision and text encoders from the onset. LiT [35] developed a text encoder compatible with a pre-trained vision encoder. FLIP [33] integrated region-word alignment in contrastive learning to enhance detailed vision-language correlation. Coca [34] pre-trained an image-text encoder-decoder foundation model using contrastive and captioning loss. For pathological images, some research works adapted VLMs for training with pathological images and text. Lu et al. [21] built a VLM using over 1.17 million histopathology image-caption pairs based on a task-agnostic pre-training approach derived from Coca. Huang et al. [8] curated the OpenPath + +![](images/25b31fb8e0c1ee118cf04665ee03203ae1feb6d33a687b1473fe7c98875cc92c.jpg) +Figure 2. Left: The structure of the FiVE framework. The model consists of a frozen image encoder, a text encoder, and the TFS module. Whole slide images are divided into instances for embedding extraction by the image encoder. Raw pathology reports are standardized by GPT-4 into fine-grained descriptions. The fine-grained descriptions and manual prompts are sampled, shuffled, and reconstructed in pairs. These prompts aggregate instances into bag-level features, subsequently aligned with the descriptions utilizing contrastive loss. Top Right: Fine-grained pathological descriptions. The fine-grained pathological descriptions are generated from multiple answers based on specific queries. These descriptions undergo a process of random sampling, shuffling, and reconstruction to form a unified sentence. Bottom Right: The Instance Aggregator module. The instance aggregator consists of a self-attention module and a cross-attention module, fusing image instance embeddings and prompt embeddings to create bag-level features. + +dataset, consisting of 208,414 pathology images with natural language descriptions from public forums, and finetuned a pre-trained CLIP on OpenPath. Lai et al. [13] also explored the generalization of CLIP in pathology image classification. These methods typically employ instance-level images (small patches from WSIs) and descriptions, requiring significant human and material resources. Zhang et al. [38] injected meaningful medical domain knowledge to advance pathological image embedding and classification. Qu et al. [23] employed GPT-4 to supplement image labels to enrich the information for training. However, the texts employed in their training offer only rudimentary descriptions of the images, primarily categorizing the general type of pathological slides, such as the disease category. This approach significantly constrains the model's capacity to discern fine-grained features, including the degree of differentiation, spread, and other details. Consequently, this limitation substantially restricts the model's generalizability and applicability to more nuanced diagnostic tasks. + +# 2.3. Prompt Learning in Vision-Language Models + +Drawing inspiration from prompt learning in natural language processing, some studies have proposed adapting Vision-Language models through end-to-end training of prompt tokens. CoOp [40] enhanced CLIP for few-shot transfer by optimizing a continuous array of prompt vectors within its language branch. CoCoOp [39] identified CoOp's suboptimal performance on new classes and tackled + +the generalization issue by conditioning prompts directly on image instances. Lu et al. [22] advocated for optimizing diverse sets of prompts by understanding their distribution. Bahng et al. [1] undertook visual prompt tuning on CLIP, focusing the prompting on the vision branch. MaPLe [12] investigated the effectiveness of multi-modal prompt learning in order to improve alignment between vision and language representations. Zhang et al. [37] adopted a set of learnable adaption prompts and presuppend them to the word tokens at higher transformer layers, efficiently fine-tuning LLaMA with less cost. Furthermore, in the context of WSI classification, prompts function as valuable adjuncts, enriching contextual information and semantic interpretation. The strategic utilization of prompts substantially improved model performance [23]. + +# 3. Method + +# 3.1. Overview + +Fig. 2 shows the pipeline of our proposed FiVE method. To initiate the process, we collaborate with professional pathologists to establish a set of standards. Following this, we employ GPT-4 to automatically extract and standardize information based on these various standards. During the training phase, we construct Fine-Grained Guidance by intricately dividing and reconstructing the text description labels and manual prompts in pairs. The combination of manual prompts and learnable prompts forms the Diagnosis + +
Manual-Designed StandardsFine-Grained Text Description Label Examples
1. What is the differentiation of the lesion?TCGA-44-6774: Lesion differentiation is moderately to poorly differentiated; Unknown; No indication of vascular invasion by the lesion; No indication of pleural invasion by the lesion; Unknown; Margins of the excised tissue are clear of disease.
2. Is there any indication of spread through air spaces around the lesion?
3. Is there any indication of vascular invasion by the lesion?
4. Is there any indication of pleural invasion by the lesion?
5. Is there any evidence of the lesion invading adjacent tissues or organs?TCGA-49-4505: Lesion differentiation is well-differentiated; Unknown; Unknown; Pleural invasion by the lesion is present, as the carcinoma extends through the visceral pleura; The lesion invades adjacent tissues or organs; Margins of the excised tissue are clear of disease.
6. Are the margins of the excised tissue clear of disease?
+ +Table 1. Manual-Designed Standards and Fine-Grained Text Description Label Examples. The answers on the right correspond to the standards on the left. "Unknown" is used as a placeholder when relevant information cannot be found. + +Prompts, which are utilized to enhance the semantic relevance of the features. Subsequently, the instance aggregator module fuses instance features with fine-grained prompts, generating bag-level features, subsequently align with the corresponding fine-grained text description labels. Additionally, to reduce computational costs, we implement the patch sampling strategy, optimizing the model's training efficiency while minimizing performance loss. + +# 3.2. Text Standardization via GPT-4 + +We utilize fine-grained text description labels extracted from pathology reports to align image bag-level features. Though pathology reports are readily accessible from public databases, their content exhibits significant variability depending on the source. Despite these format differences, pathology diagnoses consistently adhere to specific and well-established diagnostic standards. In our work, we develop fine-grained diagnostic criteria under the guidance of professional pathologists to standardize report data and extract fine-grained insights pertinent to pathological diagnosis, aiming to enhance its generalization capabilities substantially. + +The manual-designed standards aim to extract the morphological characteristics under the microscope, such as the degree of differentiation and lesion invasion, and filter out information irrelevant to the diagnosis. Subsequently, we employ GPT-4 to automatically extract answers from the original diagnosis reports based on prompts composed of these standards. If the information queried is absent in the pathology reports, "Unknown" is used as the answer. Tab. 1 shows manual-designed standards and two fine-grained text description label examples. Then, we recombine these extracted fine-grained information and integrate it into a complete description of the case image. More details about the prompts used for Text Standardization are provided in Supplementary Material. + +# 3.3. Task-specific Fine-grained Semantics Module + +# 3.3.1 Fine-grained Guidance Construction + +Due to the Text Standardization process, our data has achieved standardization. Utilizing these fine-grained text description directly as training labels can yield performance improvements. Additionally, leveraging them to generate more diverse and semantically enriched fine-grained guidance can further boost the model's performance. + +During the training process we utilize these manual-designed standards as our manual-designed prompts. We divide the original fine-grained text descriptions into several parts according to manual-designed prompts, followed by random sampling and eliminating "Unknown" tags from the initial labels. Given that the staged diagnostic reports in pathological descriptions are sequence-independent, we shuffle these preliminary labels and reconstruct them into a full-sentence description. During the training, we train reconstructed text description labels and reconstructed manual prompts in pairs. For example, consider description A: "Lesion differentiation is moderately differentiated; Margins of the excised tissue are clear of disease." and description B: "Lesion differentiation is moderately differentiated; Margins of the excised tissue are not clear of disease, as the tumor is within the bronchial margin and parenchymal margin." When only the first part of each description is sampled, they would be grouped into the same category. However, when sampling the entire sections, they are considered as distinct descriptions. Changes in granularity provide diverse perspectives on the visual image, aligning visual image with text descriptions of varying granularities. + +This strategy offers three key benefits: 1) Effectively alleviating the parent-child relationship in pathology categories. 2) Providing additional hierarchical semantic perspectives to enhance the text encoder's semantic comprehension ability. 3) Mitigating discrepancies in category annotation due to incomplete diagnostic information. + +# 3.3.2 Diagnosis Prompts + +We introduce Diagnosis Prompts to guide the aggregation of instance features into bag-level features. We compute the similarity between the instance features and the given manual prompts, utilizing the similarity scores as weights $W$ for feature aggregation to improve the task-specific relevance of the features. Here we utilize the identical manual prompts as those used to standardize the raw data, as shown in the left of the Tab. 1. + +In addition, manual-designed prompts may have some flaws, potentially failing to comprehensively capture the specific morphological characteristics of the lesion, and the model struggles to generalize towards unseen classes due to the late fusion through the transformer layers. Besides, fine-tuning the model may not always be feasible as it requires training a large number of parameters. Particularly in the case of low-data regimes, where the availability of training data like whole slide images is extremely limited. LLaMA-Adapter [37] and LLaMA-Adapter-v2 [6] explore the way to efficient fine-tuning of Language Models and Vision-Language Models respectively. These approaches introduced the Adaptation Prompt to gradually acquire instructional knowledge. They adopted zero-initialized attention with gating mechanisms to ensure stable training in the early stages. Inspired by these methods, we introduce learnable continuous diagnosis prompts to enrich the context information and enhance the model's transferability. + +Specifically, we get the manual text prompt tokens $Q_{h} = [q_{h1}, q_{h2}, \dots, q_{hn}]$ , here $n$ represents the number of the manual prompts. We concatenate the learnable continuous prompt tokens $Q_{l} = [q_{l1}, q_{l2}, \dots, q_{lm}]$ on it, here $m$ represents the number of the learnable prompt tokens. Finally we get the diagnosis prompts $Q = [q_{h1}, q_{h2}, \dots, q_{hn}, q_{l1}, q_{l2}, \dots, q_{lm}]$ . In the training phase, part of manual prompt tokens $Q_{h}$ will be sampled randomly paired with text description labels, while the whole learnable prompt tokens $Q_{l}$ will be consistently retained. + +Different with the traditional context learning prompts method, our approach pays attention to the acquisition of prior knowledge, similar to the methodology employed in Detection Transformer (DETR) [4]. We aim to acquire a set of appropriate query values to improve performance in subsequent feature screening processes. Additionally, it can also enable the model to quickly transfer to other tasks by fine-tuning this set of queries. + +# 3.3.3 Instance Aggregator Module + +The Instance Aggregator (IA) module is used to aggregate the fine-grained diagnosis prompts and instance features. As shown in the right of the Fig. 2, IA consists of a self-attention module and a cross-attention module. + +We employ self-attention to enable feature interaction + +among instance features $I_{i} = [e_{i1}, e_{i2}, \dots, e_{ij}]$ , resulting in the feature $s_{i}$ . Subsequently, utilizing the diagnosis prompts $Q$ to aggregate the instance features and acquire the feature $z_{i}$ . Then we concatenate $s_{i}$ and $z_{i}$ , utilizing the learnable parameter $W$ to fuse these features, yielding the bag-level feature $v_{i}$ . The formulas are shown as follows: + +$$ +s _ {i} = \text {S e l f A t t e n t i o n} \left(I _ {i}, I _ {i}\right) + I _ {i} \tag {1} +$$ + +$$ +z _ {i} = \text {C r o s s A t t e n t i o n} (Q, s _ {i}) \tag {2} +$$ + +$$ +v _ {i} = \operatorname {c o n c a t} \left(\operatorname {m e a n} \left(s _ {i}\right), \operatorname {m e a n} \left(z _ {i}\right)\right) \cdot W \tag {3} +$$ + +Ultimately, we acquire the image bag-level features guided by the fine-grained diagnosis prompts, which are then employed to align the fine-grained text features. + +# 3.4. Patch Sample Strategy + +Each Whole Slide Image (WSI) is partitioned into a variable number of instances, ranging from approximately 50 to 45,000. Handling such a wide range of instances markedly increases computational complexity and substantially extends the training duration. FLIP [33] reduced computation and reached higher accuracy than CLIP [25] counterpart by randomly removing a large portion of image patches during training [17]. In the case of whole slide images, pathological visual patterns are often redundantly distributed across a tissue slice. Therefore, it is feasible to sample only a subset of visual instances during training. + +For the instances in each bag (i.e., slide), we select a sample amounting to $S_{m}$ percent of the total number $p$ of each group of instances (patches). The required number of instances $S_{n}$ is described by the following formula: + +$$ +S _ {n} = \min \left(p * S _ {m}, S _ {\text {m a x n}}\right) \tag {4} +$$ + +Here $S_{maxn}$ denotes the maximum number of sample instances. To achieve this, we evenly divide each group of instances into $S_{n}$ chunks and randomly select one instance from each chunk. Since the different whole slide images could sample different $S_{n}$ , we pad the rest of the space with the same padding. + +# 3.5. Encoder and Loss Function + +We divide each WSI into instances $x_{k}$ and encode these instances into embeddings $e_{k} \in R^{D}$ using pre-trained vision encoder $E_{img}$ , composed with ResNet structure following [15], here $D$ represents the dimension of the embeddings. Then we send the instance embeddings into the TFS Module to aggregate the instance features and prompts, and obtain the bag-level embeddings $v_{i} \in R^{D}$ . The formulas are shown as follows: + +$$ +e _ {k} = E _ {i m g} \left(x _ {k}\right) \tag {5} +$$ + +$$ +I _ {i} = \left[ e _ {i 1}, e _ {i 2}, \dots , e _ {i k} \right] \tag {6} +$$ + +$$ +v _ {i} = I A \left(I _ {i}, Q\right) \tag {7} +$$ + +Besides, we generate pathologically meaningful text embeddings for each WSI, represented as $t_i \in R^D$ , by leveraging the fine-tuned text encoder BioClinicalBERT [30]. + +$$ +t _ {i} ^ {c} = E _ {t x t} \left(x _ {t x t} ^ {c}\right) \tag {8} +$$ + +where $E_{txt}$ denotes the text encoder, and $x_{txt}^{c}(c \in [1, C])$ where $C$ denotes the number of categories. Here we use the same embedding dimension $D$ as the vision encoder, suitable for contrastive learning. For text encoder $E_{txt}$ , we adopt the Low-Rank Adaptation (LoRA) [7] approach for efficient fine-tuning. + +Subsequently, the bag-level embeddings $v_{i}$ are aligned with the text embeddings $t_{i}^{c}$ to complete the training process. In this case, prediction $\hat{y}$ is obtained by applying softmax on scaled cosine similarities between the image embeddings and text embeddings: + +$$ +p (\hat {y} = c \mid I) = \frac {\exp \left(\operatorname {s i m} \left(t _ {i} ^ {c} , v _ {i}\right) / \tau\right)}{\sum_ {c ^ {\prime} = 1} ^ {C} \exp \left(\operatorname {s i m} \left(t _ {i} ^ {c ^ {\prime}} , v _ {i}\right) / \tau\right)} \tag {9} +$$ + +where $\text{sim}(\cdot, \cdot)$ refers to cosine similarity and $\tau$ is the temperature parameter. + +The fine-grained training loss is computed as the cross-entropy between the logits and soft targets as: + +$$ +L ^ {v \rightarrow t} = - \frac {1}{N} \sum_ {i = 1} ^ {N} \sum_ {j = 1} ^ {N} y _ {i j} \log \left(p _ {i j}\right) \tag {10} +$$ + +here $N$ corresponds to the batch size. + +Likewise, we can compute $L^{t \to v}$ and serve $L$ as the final training objective. + +$$ +L = \frac {L ^ {v \rightarrow t} + L ^ {t \rightarrow v}}{2} \tag {11} +$$ + +# 4. Experiments and Results + +# 4.1. Datasets + +We evaluated our method on public histopathology WSI datasets: The Cancer Genome Atlas Lung (TCGA Lung) Cancer1 and Camelyon16 [2]. + +TCGA Lung Cancer. The TCGA Lung Cancer dataset comprises two cancer subtypes: Lung Adenocarcinoma (LUAD) and Lung Squamous Cell Carcinoma (LUSC). It includes diagnostic slides with 541 slides from 478 LUAD cases and 512 slides from 478 LUSC cases. For WSI preprocessing, following [15], we cropped each WSI into $256 \times 256$ non-overlapping patches and removed the background region. The dataset encompasses approximately 5.2 million patches at $20 \times$ magnification, averaging about 5,000 patches per WSI. Following [28], we randomly split the dataset into training, validation, and testing sets with a ratio + +of 65:10:25 on the patient level and adopted 4-fold cross-validation. We collected corresponding pathology reports ${}^{2}$ and cleaned them by GPT-4 to produce fine-grained text description labels. To ensure professionalism and accuracy, we invited professional pathologists to check and correct the textual labels. Additionally, to evaluate the generalizability of the model and perform zero-shot classification, we utilized a dataset of histological subtype labels for TCGALUAD from the cBioPortal database ${}^{3}$ ,More details about subtype labels are provided in Supplementary Material. Camelyon16. The Camelyon16 dataset [2] consists of 399 Hematoxylin and Eosin (H&E) stained slide images, utilized for metastasis detection in breast cancer. We preprocessed each WSI by segmenting it into ${256} \times {256}$ nonoverlapping patches, excluding background regions. In total, this process yielded approximately 2.8 million patches at a ${20} \times$ magnification level,with about 7,200 patches per bag. We adopted 3-times 3-fold cross-validation. + +# 4.2. Implementation Details + +In experiments, we employed ResNet following [15] as image encoder to extract image features, while pre-trained BioClinicalBERT from [30] as text encoder to generate text features. LoRA [7] was adopted for fine-tuning the text encoder, with an alpha value of 32 and a rank value of 8. We divided the whole slide image into patches with $256 \times 256$ , then applied a random crop with size $224 \times 224$ . We adopted AdamW [19] with beta (0.9, 0.98), eps 1e-8, learning rate 3e-6, warmup ratio 0.1, weight decay 1e-4 as our optimizer. Additionally, we used the batch size of 1 with an accumulation step of 8 and trained for 150 epochs. We utilized mixed-precision training on 4 NVIDIA-A800 GPUs. + +# 4.3. Zero-Shot Histological Subtype Classification + +Prior researches [23, 28, 38] have predominantly concentrated on classifying primary cancer categories. Our approach extends beyond this by attempting to classify detailed histological subtypes. It is crucial to emphasize that this task poses a significant challenge, often proving difficult for even skilled pathologists to make direct judgments. + +Since only the LUAD's histological subtype dataset was provided on the online database, we conducted zero-shot subtype classification evaluation on LUAD subtype datasets, with the model being pre-trained on TCGA-LUAD or TCGA-LUSC. Throughout the training phase, subtype label information was deliberately excluded, ensuring that all experiments are conducted solely with fine-grained labels. This aims to evaluate the model's ability to identify novel diagnostic categories without specific training on subtypes. We extended the morphological appearance text de + +scription labels of the target data using GPT-4. After the text encoder encodes the labels, similarity calculation with image features achieved zero-shot classification. + +Since existing zero-shot learning methods cannot be used in WSI subtype classification, we constructed three Linear-Probe method baselines: Mean pooling, Max pooling, and Attention pooling. As shown in Tab. 2, FiVE attains $65.23\%$ top-1 accuracy and $95.18\%$ top-5 accuracy when pre-trained with TCGA-LUAD fine-grained labels. When pre-trained with TCGA-LUSC fine-grained labels, it achieves $62.02\%$ top-1 accuracy and $94.36\%$ top-5 accuracy. Moreover, the zero-shot performance of FiVE notably exceeds that of the baseline. This capability in classifying LUAD subtypes is attributed to the focus on fine-grained pathological features during training. + +
MethodTCGA-LUADTCGA-LUSC
Top-1Top-5Top-1Top-5
Mean-pooling40.8283.4631.8682.96
Max-pooling45.0588.7736.3686.09
Attention-pooling58.3693.5354.4192.50
FiVE (Ours)65.2395.1862.0294.36
+6.87+1.65+7.61+1.86
+ +# 4.4. Few-Shot Classification + +Our model demonstrates adaptability to various tasks even in scenarios with limited data availability. Few-shot experiments were conducted to demonstrate its transferability to downstream tasks. We initialized the networks with pretrained weights derived from the model trained on TCGA image-report pairs, and subsequently fine-tuned the model on downstream datasets for few-shot image classification. We followed [23] and conducted experiments with 1, 2, 4, 8, 16, and additional 0 shot on the downstream dataset. The results are summarized in Tab. 3. + +Our model exhibits remarkable performance in zero-shot classification, achieving an accuracy of $71.26\%$ , even surpassing the SOTA method's one-shot experiment. Upon the introduction of training data, our models display exceptional transferability, outperforming the SOTA by $12.90\%$ in the one-shot setting. The model's performance improves accordingly with the number of shot. Upon reaching 16-shot, our model reaches an impressive accuracy of $91.25\%$ , showcasing a notable $9.19\%$ improvement close to the fully supervised performance level. + +Table 2. Zero-Shot performance on histological subtype classification. We pre-trained the model using fine-grained labels of TCGA-LUAD and TCGA-LUSC, then applied zero-shot classification to the histological subtypes of TCGA-LUAD. + +
Method16-shot8-shot4-shot2-shot1-shot0-shot
Mean-pool65.3353.8944.8552.9345.34\
Max-pool48.4849.5544.2248.3949.03\
Attn-pool72.5065.7962.4758.3656.23\
CoOp [40]78.3567.9967.6067.5467.81\
TOP [23]82.0680.5175.4172.3871.01\
FiVE91.2590.8088.1085.5183.9171.26
(Ours)+9.19+10.29+12.69+13.13+12.90
+ +Table 3. Few-shot classification performance on TCGA Lung Cancer. Mean-pool, Max-pool, and Attn-pool correspond to Linear-Probe implementations with Mean-pooling, Max-pooling, and Attention-pooling, respectively. + +# 4.5. Performance Comparison with Existing Works + +We compared FiVE with ABMIL [10], DSMIL [15], CLAM-SB [20], CLAM-MB [20], TransMIL [26], DTFDMIL [36], and MHIM-MIL [28], all of which are attention-based MIL methods. In addition, we included two traditional MIL pooling operations, Max-pooling and Mean-pooling, for comparison. The results of all other methods are reproduced using the official code they provide under the same settings. + +To evaluate the performance of FiVE, we conducted fine-grained pre-training exclusively on the TCGA dataset, followed by fully supervised experiments on Camelyon16 and TCGA Lung Cancer datasets for WSI classification. Note that the test data is not used in pre-training. Results are shown in Tab. 4. It can be found that our method outperforms all the other baselines by a great margin, which fully demonstrates the significance of our fine-grained training scheme in improving performance on downstream tasks. Besides, since the pre-training data mainly comes from TCGA data, on the other hand, the task-specific prompt design is more suitable for TCGA data. This results in the performance improvement of our method for TCGA data being greater than Camelyon16 data. + +# 4.6. Ablation Studies + +# 4.6.1 Effectiveness of TFS Module + +We focused on evaluating the impact of the TFS module on the model's overall performance with the TCGA Lung Cancer. The results detailed in Tab. 5 reveal significant improvements at each stage of feature enhancement. Initially, the model with only Self Attention attained ACC of $89.77\%$ , AUC of $92.85\%$ , and F1-score of $89.95\%$ . The incorporation of fine-grained labels led to increases of $1.44\%$ in ACC, $1.51\%$ in AUC, and $1.43\%$ in F1-score. Subsequent integration of additional fine-grained guidance further improved performance. Ultimately, full framework with the TFS Module achieved the highest performance of $94.62\%$ ACC, $96.33\%$ AUC, and $93.89\%$ F1-score. These ablation + +
MethodCamelyon16TCGA Lung Cancer
ACCAUCF1-scoreACCAUCF1-score
Max-pooling78.95±2.2881.28±3.7471.06±2.5981.49±1.2486.45±0.7180.56±1.09
Mean-pooling76.69±0.2080.07±0.7870.41±0.1684.14±2.9790.13±2.4083.39±3.14
ABMIL [10]90.06±0.6094.00±0.8387.40±1.0588.03±2.1993.17±2.0587.41±2.42
DSMIL [15]90.17±1.0294.57±0.4087.65±1.1888.32±2.7093.71±1.8287.90±2.50
CLAM-SB [20]90.32±0.1294.65±0.3087.89±0.5987.74±2.2293.67±1.6487.36±2.24
CLAM-MB [20]90.14±0.8594.70±0.7688.10±0.6388.73±1.6293.69±0.5488.28±1.58
TransMIL [26]89.22±2.3293.51±2.1385.10±4.3387.08±1.9792.51±1.7686.40±2.08
DTFD-MIL [36]90.22±0.3695.15±0.1487.62±0.5988.23±2.1293.83±1.3987.71±2.04
MHIM-MIL [28]92.48±0.3596.49±0.6590.75±0.7389.93±3.3795.53±1.7489.71±2.92
FiVE (Ours)94.25±0.3397.56±0.5993.24±0.7294.62±2.1396.33±1.2193.89±1.90
+ +Table 4. Comparison Performance of slide classification on Camelyon16 and TCGA Lung Cancer. + +
SAFGLFGGLDPACCAUCF1-score
✓89.7792.8589.58
✓✓91.2194.3691.01
✓✓✓93.5696.0193.17
✓✓✓✓94.6296.3393.89
+ +Table 5. The ablation experiments of the pre-trained model finetuned on the TCGA Lung Cancer. SA, FGL, FGG, and LDP represent Self Attention, Fine-Grained Labels, Fine-Grained Guidance, and Learnable Diagnosis Prompts, respectively. + +experiments highlight the benefits of integrating these methods in the WSI classification task. + +# 4.6.2 Effectiveness of Patch Sample Strategy + +We verified the impact of different sampling strategies on model performance on TCGA data. Here, we assumed that there are two main indicators that affect model performance, sample ratio and max sample threshold (represented as MAXN). At the same time, in order to differentiate the experimental results, we used the unfrozen image encoder for experimental verification. + +![](images/46a414cffb3970e915eb9e2bcc441f82f580a6c0d9aa1fc3bde5b1e90bd3a029.jpg) +Figure 3. Classification performance on TCGA Lung Cancer with diverse sampling strategies, presenting average and standard deviation (std) ACC values. + +![](images/dcefa4dcf6a6b3ec8e8923aa33e0d8e8909eb6510db76eb431b162b5e66f8c25.jpg) + +As shown in Fig. 3, when MAXN is sufficiently large ( $\geq$ 2048), the correlation between the model's performance and the sample ratio is not significant. Even with a small sample + +ratio $(\leq 0.5)$ , the model can still effectively align with the original data distribution. Conversely, when MAXN is not large enough (<2048), the primary change in the model's performance depends on whether the magnitude of the sample ratio can match the original data distribution. When the sample ratio increases to 0.75, a steep performance improvement can be observed, suggesting that the sampled data at this point conforms to the original data distribution, after which the model's performance stabilizes. + +Based on comprehensive experimental results, we established the sample ratio and MAXN as 0.5 and 2048 as suitable hyperparameters for the unfrozen image encoder. As for the frozen image encoder experiment, taking into account the performance constraints imposed by the frozen image encoder on the model's performance upper limit [38], we recommend the sample ratio and MAXN to be set at 0.5 and 16384 based on our experiments. + +# 5. Conclusion + +In this paper, we introduce FiVE, a novel framework that demonstrates robust generalization and strong transferability for WSI classification. Our work pioneers the utilization of non-standardized WSI-report pairs from public databases to develop a VLM. To capture the complexities and diversity within these reports, we introduce the Task-specific Fine-grained Semantics (TFS) module. This module reconstructs fine-grained labels and diagnosis prompts during training, enhancing the semantic relevance of its features by introducing diagnosis prompts. Furthermore, considering that pathological visual patterns are redundantly distributed across tissue slices, we employ a sampling strategy to reduce computational costs. Our experiments demonstrate the robust generalizability and computational efficiency of the proposed framework, which can also be easily adapted to other tasks with minimal fine-tuning. We aspire to provide empirical insights and contribute to AI pathology research. + +# References + +[1] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Visual prompting: Modifying pixel space to adapt pre-trained models. arXiv preprint arXiv:2203.17274, 3:11-12, 2022. 3 +[2] Babak Ehteshami Bejnordi, Mitko Veta, Paul Johannes Van Diest, Bram Van Ginneken, Nico Karssemeijer, Geert Litjens, Jeroen AWM Van Der Laak, Meyke Hermsen, Quirine F Manson, Maschenka Balkenhol, et al. Diagnostic assessment of deep learning algorithms for detection of lymph node metastases in women with breast cancer. Jama, 318(22):2199-2210, 2017. 6 +[3] Gabriele Campanella, Matthew G Hanna, Luke Geneslaw, Allen Miraflor, Vitor Werneck Krauss Silva, Klaus J Busam, Edi Brogi, Victor E Reuter, David S Klimstra, and Thomas J Fuchs. Clinical-grade computational pathology using weakly supervised deep learning on whole slide images. Nature medicine, 25(8):1301-1309, 2019. 2 +[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 5 +[5] Tsai Hor Chan, Fernando Julio Cendra, Lan Ma, Guosheng Yin, and Lequan Yu. Histopathology whole slide image analysis with heterogeneous graph representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15661-15670, 2023. 2 +[6] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xi-angyu Yue, et al. Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010, 2023. 5 +[7] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6 +[8] Zhi Huang, Federico Bianchi, Mert Yuksekgonul, Thomas J Montine, and James Zou. A visual-language foundation model for pathology image analysis using medical twitter. Nature Medicine, pages 1-10, 2023. 2 +[9] Carolyn Hutter and Jean Claude Zenklusen. The cancer genome atlas: creating lasting value beyond its data. Cell, 173(2):283-285, 2018. 2 +[10] Maximilian Ilse, Jakub Tomczak, and Max Welling. Attention-based deep multiple instance learning. In International conference on machine learning, pages 2127-2136. PMLR, 2018. 7, 8 +[11] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 1 +[12] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In Proceedings of the + +IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19113-19122, 2023. 3 +[13] Zhengfeng Lai, Zhuoheng Li, Luca Cerny Oliveira, Joohi Chauhan, Brittany N Dugger, and Chen-Nee Chuah. Clipse: Fine-tune clip with visual feature fusion for pathology image analysis towards minimizing data collection efforts. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2374-2380, 2023. 3 +[14] Marvin Lerousseau, Maria Vakalopoulou, Marion Classe, Julien Adam, Enzo Battistella, Alexandre Carré, Théo Estienne, Théophraste Henry, Eric Deutsch, and Nikos Paragios. Weakly supervised multiple instance learning histopathological tumor segmentation. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part V 23, pages 470-479. Springer, 2020. 2 +[15] Bin Li, Yin Li, and Kevin W Eliceiri. Dual-stream multiple instance learning network for whole slide image classification with self-supervised contrastive learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14318-14328, 2021. 1, 2, 5, 6, 7, 8 +[16] Honglin Li, Chenglu Zhu, Yunlong Zhang, Yuxuan Sun, Zhongyi Shui, Wenwei Kuang, Sunyi Zheng, and Lin Yang. Task-specific fine-tuning via variational information bottleneck for weakly-supervised pathology whole slide image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7454-7463, 2023. 1 +[17] Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. Scaling language-image pre-training via masking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23390-23400, 2023. 5 +[18] Tiancheng Lin, Zhimiao Yu, Hongyu Hu, Yi Xu, and Chang-Wen Chen. Interventional bag multi-instance learning on whole-slide pathological images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19830-19839, 2023. 1, 2 +[19] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6 +[20] Ming Y Lu, Drew FK Williamson, Tiffany Y Chen, Richard J Chen, Matteo Barbieri, and Faisal Mahmood. Data-efficient and weakly supervised computational pathology on whole-slide images. Nature biomedical engineering, 5(6):555-570, 2021. 2, 7, 8 +[21] Ming Y Lu, Bowen Chen, Drew FK Williamson, Richard J Chen, Ivy Liang, Tong Ding, Guillaume Jaume, Igor Odintsov, Andrew Zhang, Long Phi Le, et al. Towards a visual-language foundation model for computational pathology. arXiv preprint arXiv:2307.12914, 2023. 2 +[22] Yuning Lu, Jianzhuang Liu, Yonggang Zhang, Yajing Liu, and Xinmei Tian. Prompt distribution learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5206-5215, 2022. 3 +[23] Linhao Qu, Xiaoyuan Luo, Kexue Fu, Manning Wang, and Zhijian Song. The rise of ai language pathologists: + +Exploring two-level prompt learning for few-shot weakly-supervised whole slide image classification. arXiv preprint arXiv:2305.17891, 2023. 2, 3, 6, 7 +[24] Linhao Qu, Zhiwei Yang, Minghong Duan, Yingfan Ma, Shuo Wang, Manning Wang, and Zhijian Song. Boosting whole slide image classification from the perspectives of distribution, correlation and magnification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21463-21473, 2023. 2 +[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 5 +[26] Zhuchen Shao, Hao Bian, Yang Chen, Yifeng Wang, Jian Zhang, Xiangyang Ji, et al. Transmil: Transformer based correlated multiple instance learning for whole slide image classification. Advances in neural information processing systems, 34:2136-2147, 2021. 2, 7, 8 +[27] Zhuchen Shao, Yifeng Wang, Yang Chen, Hao Bian, Shaohui Liu, Haoqian Wang, and Yongbing Zhang. Lnpl-mil: Learning from noisy pseudo labels for promoting multiple instance learning in whole slide image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21495-21505, 2023. 2 +[28] Wenhao Tang, Sheng Huang, Xiaoxian Zhang, Fengtao Zhou, Yi Zhang, and Bo Liu. Multiple instance learning framework with masked hard instance mining for whole slide image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4078-4087, 2023. 6, 7, 8 +[29] Xinggang Wang, Yongluan Yan, Peng Tang, Xiang Bai, and Wenyu Liu. Revisiting multiple instance neural networks. Pattern Recognition, 74:15-24, 2018. 2 +[30] Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chau-mond, Clement Delangue, Anthony Moi, Pierrick Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, et al. Huggingface's transformers: State-of-the-art natural language processing. arXiv preprint arXiv:1910.03771, 2019. 6 +[31] Jinxi Xiang and Jun Zhang. Exploring low-rank property in multiple instance learning for whole slide image classification. In The Eleventh International Conference on Learning Representations, 2022. 1 +[32] Gang Xu, Zhigang Song, Zhuo Sun, Calvin Ku, Zhe Yang, Cancheng Liu, Shuhao Wang, Jianpeng Ma, and Wei Xu. Camel: A weakly supervised learning framework for histopathology image segmentation. In Proceedings of the IEEE/CVF International Conference on computer vision, pages 10682-10691, 2019. 2 +[33] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. arXiv preprint arXiv:2111.07783, 2021. 1, 2, 5 +[34] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 2 + +[35] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18123-18133, 2022. 2 +[36] Hongrun Zhang, Yanda Meng, Yitian Zhao, Yihong Qiao, Xiaoyun Yang, Sarah E Coupland, and Yalin Zheng. Dtdmil: Double-tier feature distillation multiple instance learning for histopathology whole slide image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18802-18812, 2022. 2, 7, 8 +[37] Renrui Zhang, Jiaming Han, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, Peng Gao, and Yu Qiao. Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023. 3, 5 +[38] Yunkun Zhang, Jin Gao, Mu Zhou, Xiaosong Wang, Yu Qiao, Shaoting Zhang, and Dequan Wang. Text-guided foundation model adaptation for pathological image classification. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 272-282. Springer, 2023. 1, 2, 3, 6, 8 +[39] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 3 +[40] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 3, 7 \ No newline at end of file diff --git a/2024/Generalizable Whole Slide Image Classification with Fine-Grained Visual-Semantic Interaction/images.zip b/2024/Generalizable Whole Slide Image Classification with Fine-Grained Visual-Semantic Interaction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d1e72a0ad77a29d0f2c3753fd554abd1644ff9cf --- /dev/null +++ b/2024/Generalizable Whole Slide Image Classification with Fine-Grained Visual-Semantic Interaction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2c223b182c3604bfc7c4ce072f7e842ffc9ef83683402cc34a276daaac6e915b +size 494647 diff --git a/2024/Generalizable Whole Slide Image Classification with Fine-Grained Visual-Semantic Interaction/layout.json b/2024/Generalizable Whole Slide Image Classification with Fine-Grained Visual-Semantic Interaction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1e5b5c0586abcf787b291b00a471814b0232ea56 --- /dev/null +++ b/2024/Generalizable Whole Slide Image Classification with Fine-Grained Visual-Semantic Interaction/layout.json @@ -0,0 +1,7865 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 91, + 103, + 501, + 137 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 103, + 501, + 137 + ], + "spans": [ + { + "bbox": [ + 91, + 103, + 501, + 137 + ], + "type": "text", + "content": "Generalizable Whole Slide Image Classification with Fine-Grained Visual-Semantic Interaction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 149, + 160, + 192, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 160, + 192, + 175 + ], + "spans": [ + { + "bbox": [ + 149, + 160, + 192, + 175 + ], + "type": "text", + "content": "Hao Li" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 120, + 175, + 201, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 175, + 201, + 190 + ], + "spans": [ + { + "bbox": [ + 120, + 175, + 201, + 190 + ], + "type": "text", + "content": "Wenxian Yang3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 220, + 161, + 280, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 161, + 280, + 175 + ], + "spans": [ + { + "bbox": [ + 220, + 161, + 280, + 175 + ], + "type": "text", + "content": "Ying Chen" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 220, + 175, + 305, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 175, + 305, + 190 + ], + "spans": [ + { + "bbox": [ + 220, + 175, + 305, + 190 + ], + "type": "text", + "content": "Liansheng Wang" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 310, + 161, + 370, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 161, + 370, + 175 + ], + "spans": [ + { + "bbox": [ + 310, + 161, + 370, + 175 + ], + "type": "text", + "content": "Yifei Chen" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 331, + 175, + 399, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 175, + 399, + 190 + ], + "spans": [ + { + "bbox": [ + 331, + 175, + 399, + 190 + ], + "type": "text", + "content": "Bowen Ding" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 399, + 161, + 477, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 161, + 477, + 175 + ], + "spans": [ + { + "bbox": [ + 399, + 161, + 477, + 175 + ], + "type": "text", + "content": "Rongshan " + }, + { + "bbox": [ + 399, + 161, + 477, + 175 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}\\mathbf{u}^{1*}" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 420, + 175, + 493, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 420, + 175, + 493, + 188 + ], + "spans": [ + { + "bbox": [ + 420, + 175, + 493, + 188 + ], + "type": "text", + "content": "Yuchen Han\\*" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 69, + 190, + 541, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 190, + 541, + 203 + ], + "spans": [ + { + "bbox": [ + 69, + 190, + 541, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 69, + 190, + 541, + 203 + ], + "type": "text", + "content": "School of Informatics, Xiamen University " + }, + { + "bbox": [ + 69, + 190, + 541, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 69, + 190, + 541, + 203 + ], + "type": "text", + "content": "Huawei " + }, + { + "bbox": [ + 69, + 190, + 541, + 203 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 69, + 190, + 541, + 203 + ], + "type": "text", + "content": "Aginome Scientific, Xiamen, China" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 203, + 555, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 203, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 55, + 203, + 555, + 217 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 55, + 203, + 555, + 217 + ], + "type": "text", + "content": "Department of Pathology, Shanghai Chest Hospital, Shanghai Jiao Tong University School of Medicine" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 140, + 219, + 466, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 219, + 466, + 232 + ], + "spans": [ + { + "bbox": [ + 140, + 219, + 466, + 232 + ], + "type": "text", + "content": "{111ih, cying2023}@stu.xmu.edu.cn, {rsyu, 1swang}@xmu.edu.cn," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 89, + 234, + 516, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 234, + 516, + 244 + ], + "spans": [ + { + "bbox": [ + 89, + 234, + 516, + 244 + ], + "type": "text", + "content": "chenyifei14@huawei.com,wx@aginome.com，dingbowenmail@126.com，ychan@cmu.edu.cn" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 143, + 273, + 191, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 273, + 191, + 285 + ], + "spans": [ + { + "bbox": [ + 143, + 273, + 191, + 285 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 298, + 289, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 298, + 289, + 681 + ], + "spans": [ + { + "bbox": [ + 46, + 298, + 289, + 681 + ], + "type": "text", + "content": "Whole Slide Image (WSI) classification is often formulated as a Multiple Instance Learning (MIL) problem. Recently, Vision-Language Models (VLMs) have demonstrated remarkable performance in WSI classification. However, existing methods leverage coarse-grained pathogenetic descriptions for visual representation supervision, which are insufficient to capture the complex visual appearance of pathogenetic images, hindering the generalizability of models on diverse downstream tasks. Additionally, processing high-resolution WSIs can be computationally expensive. In this paper, we propose a novel \"Fine-grained Visual-Semantic Interaction\" (FiVE) framework for WSI classification. It is designed to enhance the model's generalizability by leveraging the interaction between localized visual patterns and fine-grained pathological semantics. Specifically, with meticulously designed queries, we start by utilizing a large language model to extract fine-grained pathological descriptions from various non-standardized raw reports. The output descriptions are then reconstructed into fine-grained labels used for training. By introducing a Task-specific Fine-grained Semantics (TFS) module, we enable prompts to capture crucial visual information in WSIs, which enhances representation learning and augments generalization capabilities significantly. Furthermore, given that pathological visual patterns are redundantly distributed across tissue slices, we sample a subset of visual instances during training. Our method demonstrates robust generalizability and strong transferability, dominantly outperforming the counterparts on the TCGA Lung Cancer dataset with at least " + }, + { + "bbox": [ + 46, + 298, + 289, + 681 + ], + "type": "inline_equation", + "content": "9.19\\%" + }, + { + "bbox": [ + 46, + 298, + 289, + 681 + ], + "type": "text", + "content": " higher accuracy in few-shot experiments. The code is available at: https://github.com/ls1rius/WSI_FIVE." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 273, + 386, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 273, + 386, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 273, + 386, + 285 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 307, + 302, + 545, + 442 + ], + "blocks": [ + { + "bbox": [ + 307, + 302, + 545, + 442 + ], + "lines": [ + { + "bbox": [ + 307, + 302, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 307, + 302, + 545, + 442 + ], + "type": "image", + "image_path": "28b2ad7e21f2112a208d419264e6b82df4c4aedcc96471ca2744a9ba059c2cb3.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 449, + 545, + 526 + ], + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 526 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 526 + ], + "type": "text", + "content": "Figure 1. Challenges in WSI-text contrastive learning. Most conventional VLM approaches categorize whole slide images using category-level text descriptions, overlooking intra-class differences, leading to a decline in performance and limitations in generalization capabilities. Instead, we extract fine-grained descriptions from pathology reports as slide-level labels to develop our model, exhibiting detailed variations in each WSI." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 544, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 545, + 676 + ], + "type": "text", + "content": "Histological Whole Slide Image (WSI) classification plays a crucial role in computational pathology by automating disease diagnosis and subtyping. For high-resolution WSI analysis, Multiple Instance Learning (MIL) has become the dominant method. These methods treat each WSI as a \"Bag\" sample, and aggregate numerous patches within it as instances for thorough decision-making. Nevertheless, most existing methods [15, 16, 18, 31] focused on processing image data to conduct WSI classification, potentially not emphasizing critical pathological insights, particularly the expert textual annotations that accompany these slides." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "Recently, Vision-Language Models (VLMs) [11, 25, 33] underscored the significance of integrating multimodal information for developing robust encoders. Zhang et al. [38]" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 693, + 286, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 693, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 693, + 286, + 712 + ], + "type": "text", + "content": "*Co-corresponding authors. This work was supported by National Natural Science Foundation of China (Grant No. 62371409)." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11398" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "exploited disease-grade text labels and extracted text insights using pre-trained language models. Qu et al. [23] utilized GPT-4 in a question-and-answer mode to obtain language prior knowledge at both instance and bag levels for VLM training. However, the challenge lies in the uniqueness and variability of content in each WSI. Existing methods developed their models with coarse-grained descriptions (i.e., simplistic Category-Level text labels [38] or descriptive Category-Level text labels constructed by GPT-4 [23]), as shown in Fig. 1. They may have omitted crucial fine-grained pathological details, including differentiation level, vascular invasion, etc., which results in reduced model performance and limited generalization." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 232, + 289, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 232, + 289, + 543 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 289, + 543 + ], + "type": "text", + "content": "WSIs accompanied by their corresponding reports (i.e., WSI-report pairs) offer detailed descriptions and fine-grained information vital for WSI analysis. Furthermore, a substantial collection of these pairs is accessible in public databases, such as The Cancer Genome Atlas (TCGA) [9]. However, their full potential has not been adequately harnessed yet. The challenge in developing a Visual Language Model (VLM) using WSI-report pairs mainly lies in the diverse formats and standards of the raw reports from different hospitals, which increases the complexity of data preprocessing and standardization processes. Additionally, pathology reports often contain extraneous information, including report metadata, tissue processing descriptions, and repetitive elements, which can introduce noise to the textual data. How to extract useful information from raw pathology reports to construct WSI-report pairs is a key issue. Moreover, recent studies have demonstrated the efficacy of prompt engineering in enhancing VLMs. In contrast to natural images, WSI data encompasses extensive professional pathological information and intricate details. How to craft prompts to make full use of this semantic information to guide fine-grained feature learning is a challenging task. Besides, the high computational costs to train models with high-resolution WSIs also limits the promotion of the model, resulting in a certain resource threshold for WSI analysis." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 546, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 546, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 546, + 289, + 713 + ], + "type": "text", + "content": "To address these issues, we propose a novel whole slide image classification method with Fine-grained VisualsEmantic interaction termed as FiVE, which shows robust generalizability and efficiency in computation. Firstly, we obtain WSIs with non-standardized raw pathology reports from a public database. Collaborating with professional pathologists, we craft a set of specialized prompts to standardize reports. Following this, we employ the large language model GPT-4 to automatically clean and standardize the raw report data. In addition, we propose the Task-specific Fine-grained Semantic (TFS) Module, which utilizes manual-designed prompts to direct visual attention to specific pathological areas while constructing Fine-Grained Guidance to enhance the semantic relevance of model fea" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "tures. Considering the diffuse distribution of pathological diseases within tissue sections and the presence of numerous non-diagnostic regions in WSIs, we also incorporate a patch sampling strategy during the training phase to enhance training efficiency and reduce computational costs. The contributions of this paper are summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 147, + 545, + 278 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 304, + 147, + 545, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 147, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 304, + 147, + 545, + 194 + ], + "type": "text", + "content": "- We pioneer the utilization of the available WSI diagnostic reports with fine-grained guidance. The obtained fine-grained description labels lead to improved supervision by discriminating the visual appearances more precisely." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 195, + 545, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 195, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 304, + 195, + 545, + 230 + ], + "type": "text", + "content": "- We introduce a novel Task-specific Fine-grained Semantics (TFS) Module to offer fine-grained guidance, significantly enhancing the model's generalization capabilities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 231, + 545, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 231, + 545, + 278 + ], + "spans": [ + { + "bbox": [ + 304, + 231, + 545, + 278 + ], + "type": "text", + "content": "- We implement a patch sampling strategy on visual instances during training to enhance computational efficiency without significantly compromising accuracy, thereby optimizing the model's training process." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 305, + 289, + 392, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 289, + 392, + 301 + ], + "spans": [ + { + "bbox": [ + 305, + 289, + 392, + 301 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 309, + 459, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 309, + 459, + 322 + ], + "spans": [ + { + "bbox": [ + 305, + 309, + 459, + 322 + ], + "type": "text", + "content": "2.1. Whole Slide Image Analysis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 327, + 545, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 545, + 518 + ], + "type": "text", + "content": "Contemporary methodologies for WSI analysis predominantly employ MIL methods where each WSI is treated as a \"Bag\" and its extracted patches as instances within this bag. MIL methods consist of instance-based methods [3, 14, 32] and embedding representation-based methods [15, 20, 26, 29, 36]. However, the majority of existing methods [5, 18, 24, 27] almost exclusively rely on image data, neglecting vital pathological details, notably the specialist text annotations that accompany the images. Recent works [23, 38] have taken note of this issue and started to utilize text information to improve pathological image classification. They used bag-level text labels or the descriptive labels generated by GPT. However, given the unique and varied descriptions of each WSI, their methods fall short of fully leveraging the detailed textual information present in the slides." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 527, + 444, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 527, + 444, + 540 + ], + "spans": [ + { + "bbox": [ + 305, + 527, + 444, + 540 + ], + "type": "text", + "content": "2.2. Vision-Language Models" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": "Recent researches have made efforts to develop Vision-Language Models (VLMs). CLIP [25] gathered 400 million image-text pairs and initiated training with synchronized vision and text encoders from the onset. LiT [35] developed a text encoder compatible with a pre-trained vision encoder. FLIP [33] integrated region-word alignment in contrastive learning to enhance detailed vision-language correlation. Coca [34] pre-trained an image-text encoder-decoder foundation model using contrastive and captioning loss. For pathological images, some research works adapted VLMs for training with pathological images and text. Lu et al. [21] built a VLM using over 1.17 million histopathology image-caption pairs based on a task-agnostic pre-training approach derived from Coca. Huang et al. [8] curated the OpenPath" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "11399" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 537, + 251 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 537, + 251 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 537, + 251 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 537, + 251 + ], + "type": "image", + "image_path": "25b31fb8e0c1ee118cf04665ee03203ae1feb6d33a687b1473fe7c98875cc92c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 259, + 548, + 350 + ], + "lines": [ + { + "bbox": [ + 46, + 259, + 548, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 259, + 548, + 350 + ], + "type": "text", + "content": "Figure 2. Left: The structure of the FiVE framework. The model consists of a frozen image encoder, a text encoder, and the TFS module. Whole slide images are divided into instances for embedding extraction by the image encoder. Raw pathology reports are standardized by GPT-4 into fine-grained descriptions. The fine-grained descriptions and manual prompts are sampled, shuffled, and reconstructed in pairs. These prompts aggregate instances into bag-level features, subsequently aligned with the descriptions utilizing contrastive loss. Top Right: Fine-grained pathological descriptions. The fine-grained pathological descriptions are generated from multiple answers based on specific queries. These descriptions undergo a process of random sampling, shuffling, and reconstruction to form a unified sentence. Bottom Right: The Instance Aggregator module. The instance aggregator consists of a self-attention module and a cross-attention module, fusing image instance embeddings and prompt embeddings to create bag-level features." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 369, + 289, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 289, + 596 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 289, + 596 + ], + "type": "text", + "content": "dataset, consisting of 208,414 pathology images with natural language descriptions from public forums, and finetuned a pre-trained CLIP on OpenPath. Lai et al. [13] also explored the generalization of CLIP in pathology image classification. These methods typically employ instance-level images (small patches from WSIs) and descriptions, requiring significant human and material resources. Zhang et al. [38] injected meaningful medical domain knowledge to advance pathological image embedding and classification. Qu et al. [23] employed GPT-4 to supplement image labels to enrich the information for training. However, the texts employed in their training offer only rudimentary descriptions of the images, primarily categorizing the general type of pathological slides, such as the disease category. This approach significantly constrains the model's capacity to discern fine-grained features, including the degree of differentiation, spread, and other details. Consequently, this limitation substantially restricts the model's generalizability and applicability to more nuanced diagnostic tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 609, + 284, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 284, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 284, + 624 + ], + "type": "text", + "content": "2.3. Prompt Learning in Vision-Language Models" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "type": "text", + "content": "Drawing inspiration from prompt learning in natural language processing, some studies have proposed adapting Vision-Language models through end-to-end training of prompt tokens. CoOp [40] enhanced CLIP for few-shot transfer by optimizing a continuous array of prompt vectors within its language branch. CoCoOp [39] identified CoOp's suboptimal performance on new classes and tackled" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 369, + 547, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 369, + 547, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 369, + 547, + 548 + ], + "type": "text", + "content": "the generalization issue by conditioning prompts directly on image instances. Lu et al. [22] advocated for optimizing diverse sets of prompts by understanding their distribution. Bahng et al. [1] undertook visual prompt tuning on CLIP, focusing the prompting on the vision branch. MaPLe [12] investigated the effectiveness of multi-modal prompt learning in order to improve alignment between vision and language representations. Zhang et al. [37] adopted a set of learnable adaption prompts and presuppend them to the word tokens at higher transformer layers, efficiently fine-tuning LLaMA with less cost. Furthermore, in the context of WSI classification, prompts function as valuable adjuncts, enriching contextual information and semantic interpretation. The strategic utilization of prompts substantially improved model performance [23]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 564, + 362, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 564, + 362, + 576 + ], + "spans": [ + { + "bbox": [ + 306, + 564, + 362, + 576 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 585, + 373, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 585, + 373, + 597 + ], + "spans": [ + { + "bbox": [ + 306, + 585, + 373, + 597 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "Fig. 2 shows the pipeline of our proposed FiVE method. To initiate the process, we collaborate with professional pathologists to establish a set of standards. Following this, we employ GPT-4 to automatically extract and standardize information based on these various standards. During the training phase, we construct Fine-Grained Guidance by intricately dividing and reconstructing the text description labels and manual prompts in pairs. The combination of manual prompts and learnable prompts forms the Diagnosis" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "11400" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 70, + 545, + 196 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 545, + 196 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 545, + 196 + ], + "type": "table", + "html": "
Manual-Designed StandardsFine-Grained Text Description Label Examples
1. What is the differentiation of the lesion?TCGA-44-6774: Lesion differentiation is moderately to poorly differentiated; Unknown; No indication of vascular invasion by the lesion; No indication of pleural invasion by the lesion; Unknown; Margins of the excised tissue are clear of disease.
2. Is there any indication of spread through air spaces around the lesion?
3. Is there any indication of vascular invasion by the lesion?
4. Is there any indication of pleural invasion by the lesion?
5. Is there any evidence of the lesion invading adjacent tissues or organs?TCGA-49-4505: Lesion differentiation is well-differentiated; Unknown; Unknown; Pleural invasion by the lesion is present, as the carcinoma extends through the visceral pleura; The lesion invades adjacent tissues or organs; Margins of the excised tissue are clear of disease.
6. Are the margins of the excised tissue clear of disease?
", + "image_path": "f486abcdf1abee818562d564e4fd2e29c21c616114715b812bbbc6125fa8e3c4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 204, + 546, + 227 + ], + "lines": [ + { + "bbox": [ + 47, + 204, + 546, + 227 + ], + "spans": [ + { + "bbox": [ + 47, + 204, + 546, + 227 + ], + "type": "text", + "content": "Table 1. Manual-Designed Standards and Fine-Grained Text Description Label Examples. The answers on the right correspond to the standards on the left. \"Unknown\" is used as a placeholder when relevant information cannot be found." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 247, + 288, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 247, + 288, + 344 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 288, + 344 + ], + "type": "text", + "content": "Prompts, which are utilized to enhance the semantic relevance of the features. Subsequently, the instance aggregator module fuses instance features with fine-grained prompts, generating bag-level features, subsequently align with the corresponding fine-grained text description labels. Additionally, to reduce computational costs, we implement the patch sampling strategy, optimizing the model's training efficiency while minimizing performance loss." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 370, + 219, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 370, + 219, + 383 + ], + "spans": [ + { + "bbox": [ + 47, + 370, + 219, + 383 + ], + "type": "text", + "content": "3.2. Text Standardization via GPT-4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 396, + 287, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 287, + 539 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 287, + 539 + ], + "type": "text", + "content": "We utilize fine-grained text description labels extracted from pathology reports to align image bag-level features. Though pathology reports are readily accessible from public databases, their content exhibits significant variability depending on the source. Despite these format differences, pathology diagnoses consistently adhere to specific and well-established diagnostic standards. In our work, we develop fine-grained diagnostic criteria under the guidance of professional pathologists to standardize report data and extract fine-grained insights pertinent to pathological diagnosis, aiming to enhance its generalization capabilities substantially." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "type": "text", + "content": "The manual-designed standards aim to extract the morphological characteristics under the microscope, such as the degree of differentiation and lesion invasion, and filter out information irrelevant to the diagnosis. Subsequently, we employ GPT-4 to automatically extract answers from the original diagnosis reports based on prompts composed of these standards. If the information queried is absent in the pathology reports, \"Unknown\" is used as the answer. Tab. 1 shows manual-designed standards and two fine-grained text description label examples. Then, we recombine these extracted fine-grained information and integrate it into a complete description of the case image. More details about the prompts used for Text Standardization are provided in Supplementary Material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 247, + 541, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 247, + 541, + 260 + ], + "spans": [ + { + "bbox": [ + 305, + 247, + 541, + 260 + ], + "type": "text", + "content": "3.3. Task-specific Fine-grained Semantics Module" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 267, + 496, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 267, + 496, + 280 + ], + "spans": [ + { + "bbox": [ + 306, + 267, + 496, + 280 + ], + "type": "text", + "content": "3.3.1 Fine-grained Guidance Construction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 290, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 290, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 290, + 545, + 361 + ], + "type": "text", + "content": "Due to the Text Standardization process, our data has achieved standardization. Utilizing these fine-grained text description directly as training labels can yield performance improvements. Additionally, leveraging them to generate more diverse and semantically enriched fine-grained guidance can further boost the model's performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 365, + 546, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 546, + 638 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 546, + 638 + ], + "type": "text", + "content": "During the training process we utilize these manual-designed standards as our manual-designed prompts. We divide the original fine-grained text descriptions into several parts according to manual-designed prompts, followed by random sampling and eliminating \"Unknown\" tags from the initial labels. Given that the staged diagnostic reports in pathological descriptions are sequence-independent, we shuffle these preliminary labels and reconstruct them into a full-sentence description. During the training, we train reconstructed text description labels and reconstructed manual prompts in pairs. For example, consider description A: \"Lesion differentiation is moderately differentiated; Margins of the excised tissue are clear of disease.\" and description B: \"Lesion differentiation is moderately differentiated; Margins of the excised tissue are not clear of disease, as the tumor is within the bronchial margin and parenchymal margin.\" When only the first part of each description is sampled, they would be grouped into the same category. However, when sampling the entire sections, they are considered as distinct descriptions. Changes in granularity provide diverse perspectives on the visual image, aligning visual image with text descriptions of varying granularities." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "This strategy offers three key benefits: 1) Effectively alleviating the parent-child relationship in pathology categories. 2) Providing additional hierarchical semantic perspectives to enhance the text encoder's semantic comprehension ability. 3) Mitigating discrepancies in category annotation due to incomplete diagnostic information." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11401" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 161, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 161, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 161, + 85 + ], + "type": "text", + "content": "3.3.2 Diagnosis Prompts" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 287, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 287, + 186 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 287, + 186 + ], + "type": "text", + "content": "We introduce Diagnosis Prompts to guide the aggregation of instance features into bag-level features. We compute the similarity between the instance features and the given manual prompts, utilizing the similarity scores as weights " + }, + { + "bbox": [ + 46, + 91, + 287, + 186 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 91, + 287, + 186 + ], + "type": "text", + "content": " for feature aggregation to improve the task-specific relevance of the features. Here we utilize the identical manual prompts as those used to standardize the raw data, as shown in the left of the Tab. 1." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 187, + 288, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 187, + 288, + 402 + ], + "spans": [ + { + "bbox": [ + 46, + 187, + 288, + 402 + ], + "type": "text", + "content": "In addition, manual-designed prompts may have some flaws, potentially failing to comprehensively capture the specific morphological characteristics of the lesion, and the model struggles to generalize towards unseen classes due to the late fusion through the transformer layers. Besides, fine-tuning the model may not always be feasible as it requires training a large number of parameters. Particularly in the case of low-data regimes, where the availability of training data like whole slide images is extremely limited. LLaMA-Adapter [37] and LLaMA-Adapter-v2 [6] explore the way to efficient fine-tuning of Language Models and Vision-Language Models respectively. These approaches introduced the Adaptation Prompt to gradually acquire instructional knowledge. They adopted zero-initialized attention with gating mechanisms to ensure stable training in the early stages. Inspired by these methods, we introduce learnable continuous diagnosis prompts to enrich the context information and enhance the model's transferability." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 403, + 287, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 287, + 522 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 287, + 522 + ], + "type": "text", + "content": "Specifically, we get the manual text prompt tokens " + }, + { + "bbox": [ + 46, + 403, + 287, + 522 + ], + "type": "inline_equation", + "content": "Q_{h} = [q_{h1}, q_{h2}, \\dots, q_{hn}]" + }, + { + "bbox": [ + 46, + 403, + 287, + 522 + ], + "type": "text", + "content": ", here " + }, + { + "bbox": [ + 46, + 403, + 287, + 522 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 403, + 287, + 522 + ], + "type": "text", + "content": " represents the number of the manual prompts. We concatenate the learnable continuous prompt tokens " + }, + { + "bbox": [ + 46, + 403, + 287, + 522 + ], + "type": "inline_equation", + "content": "Q_{l} = [q_{l1}, q_{l2}, \\dots, q_{lm}]" + }, + { + "bbox": [ + 46, + 403, + 287, + 522 + ], + "type": "text", + "content": " on it, here " + }, + { + "bbox": [ + 46, + 403, + 287, + 522 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 403, + 287, + 522 + ], + "type": "text", + "content": " represents the number of the learnable prompt tokens. Finally we get the diagnosis prompts " + }, + { + "bbox": [ + 46, + 403, + 287, + 522 + ], + "type": "inline_equation", + "content": "Q = [q_{h1}, q_{h2}, \\dots, q_{hn}, q_{l1}, q_{l2}, \\dots, q_{lm}]" + }, + { + "bbox": [ + 46, + 403, + 287, + 522 + ], + "type": "text", + "content": ". In the training phase, part of manual prompt tokens " + }, + { + "bbox": [ + 46, + 403, + 287, + 522 + ], + "type": "inline_equation", + "content": "Q_{h}" + }, + { + "bbox": [ + 46, + 403, + 287, + 522 + ], + "type": "text", + "content": " will be sampled randomly paired with text description labels, while the whole learnable prompt tokens " + }, + { + "bbox": [ + 46, + 403, + 287, + 522 + ], + "type": "inline_equation", + "content": "Q_{l}" + }, + { + "bbox": [ + 46, + 403, + 287, + 522 + ], + "type": "text", + "content": " will be consistently retained." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 522, + 288, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 288, + 619 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 288, + 619 + ], + "type": "text", + "content": "Different with the traditional context learning prompts method, our approach pays attention to the acquisition of prior knowledge, similar to the methodology employed in Detection Transformer (DETR) [4]. We aim to acquire a set of appropriate query values to improve performance in subsequent feature screening processes. Additionally, it can also enable the model to quickly transfer to other tasks by fine-tuning this set of queries." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 634, + 203, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 203, + 646 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 203, + 646 + ], + "type": "text", + "content": "3.3.3 Instance Aggregator Module" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 653, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 700 + ], + "type": "text", + "content": "The Instance Aggregator (IA) module is used to aggregate the fine-grained diagnosis prompts and instance features. As shown in the right of the Fig. 2, IA consists of a self-attention module and a cross-attention module." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "content": "We employ self-attention to enable feature interaction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "among instance features " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "I_{i} = [e_{i1}, e_{i2}, \\dots, e_{ij}]" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": ", resulting in the feature " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "s_{i}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": ". Subsequently, utilizing the diagnosis prompts " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " to aggregate the instance features and acquire the feature " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": ". Then we concatenate " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "s_{i}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": ", utilizing the learnable parameter " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " to fuse these features, yielding the bag-level feature " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": ". The formulas are shown as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 343, + 152, + 545, + 166 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 152, + 545, + 166 + ], + "spans": [ + { + "bbox": [ + 343, + 152, + 545, + 166 + ], + "type": "interline_equation", + "content": "s _ {i} = \\text {S e l f A t t e n t i o n} \\left(I _ {i}, I _ {i}\\right) + I _ {i} \\tag {1}", + "image_path": "b5d7eebf85e09a49db9e6ae18a043e6d8c52fe0bfdf47f8bee316d8536870297.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 344, + 168, + 545, + 180 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 168, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 344, + 168, + 545, + 180 + ], + "type": "interline_equation", + "content": "z _ {i} = \\text {C r o s s A t t e n t i o n} (Q, s _ {i}) \\tag {2}", + "image_path": "83bfad63d51908cf52c289773c7e842aa278337fdb7c580829f4ea18d519c0ca.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 343, + 182, + 545, + 196 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 182, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 343, + 182, + 545, + 196 + ], + "type": "interline_equation", + "content": "v _ {i} = \\operatorname {c o n c a t} \\left(\\operatorname {m e a n} \\left(s _ {i}\\right), \\operatorname {m e a n} \\left(z _ {i}\\right)\\right) \\cdot W \\tag {3}", + "image_path": "04fe0ea5522c19a54a3fa88882494031c1abd52f22b09ff6f513baee2b9c1996.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 202, + 545, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 202, + 545, + 238 + ], + "spans": [ + { + "bbox": [ + 304, + 202, + 545, + 238 + ], + "type": "text", + "content": "Ultimately, we acquire the image bag-level features guided by the fine-grained diagnosis prompts, which are then employed to align the fine-grained text features." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 244, + 435, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 244, + 435, + 257 + ], + "spans": [ + { + "bbox": [ + 305, + 244, + 435, + 257 + ], + "type": "text", + "content": "3.4. Patch Sample Strategy" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 262, + 545, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 262, + 545, + 394 + ], + "spans": [ + { + "bbox": [ + 304, + 262, + 545, + 394 + ], + "type": "text", + "content": "Each Whole Slide Image (WSI) is partitioned into a variable number of instances, ranging from approximately 50 to 45,000. Handling such a wide range of instances markedly increases computational complexity and substantially extends the training duration. FLIP [33] reduced computation and reached higher accuracy than CLIP [25] counterpart by randomly removing a large portion of image patches during training [17]. In the case of whole slide images, pathological visual patterns are often redundantly distributed across a tissue slice. Therefore, it is feasible to sample only a subset of visual instances during training." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 394, + 545, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 394, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 304, + 394, + 545, + 442 + ], + "type": "text", + "content": "For the instances in each bag (i.e., slide), we select a sample amounting to " + }, + { + "bbox": [ + 304, + 394, + 545, + 442 + ], + "type": "inline_equation", + "content": "S_{m}" + }, + { + "bbox": [ + 304, + 394, + 545, + 442 + ], + "type": "text", + "content": " percent of the total number " + }, + { + "bbox": [ + 304, + 394, + 545, + 442 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 394, + 545, + 442 + ], + "type": "text", + "content": " of each group of instances (patches). The required number of instances " + }, + { + "bbox": [ + 304, + 394, + 545, + 442 + ], + "type": "inline_equation", + "content": "S_{n}" + }, + { + "bbox": [ + 304, + 394, + 545, + 442 + ], + "type": "text", + "content": " is described by the following formula:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 368, + 449, + 545, + 462 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 449, + 545, + 462 + ], + "spans": [ + { + "bbox": [ + 368, + 449, + 545, + 462 + ], + "type": "interline_equation", + "content": "S _ {n} = \\min \\left(p * S _ {m}, S _ {\\text {m a x n}}\\right) \\tag {4}", + "image_path": "fe8d3dac35a4ebace6276d18712422badf238e12670b866d351f247420a1f6f7.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 468, + 545, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 468, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 468, + 545, + 540 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 304, + 468, + 545, + 540 + ], + "type": "inline_equation", + "content": "S_{maxn}" + }, + { + "bbox": [ + 304, + 468, + 545, + 540 + ], + "type": "text", + "content": " denotes the maximum number of sample instances. To achieve this, we evenly divide each group of instances into " + }, + { + "bbox": [ + 304, + 468, + 545, + 540 + ], + "type": "inline_equation", + "content": "S_{n}" + }, + { + "bbox": [ + 304, + 468, + 545, + 540 + ], + "type": "text", + "content": " chunks and randomly select one instance from each chunk. Since the different whole slide images could sample different " + }, + { + "bbox": [ + 304, + 468, + 545, + 540 + ], + "type": "inline_equation", + "content": "S_{n}" + }, + { + "bbox": [ + 304, + 468, + 545, + 540 + ], + "type": "text", + "content": ", we pad the rest of the space with the same padding." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 547, + 457, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 547, + 457, + 558 + ], + "spans": [ + { + "bbox": [ + 305, + 547, + 457, + 558 + ], + "type": "text", + "content": "3.5. Encoder and Loss Function" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 564, + 545, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 564, + 545, + 660 + ], + "spans": [ + { + "bbox": [ + 304, + 564, + 545, + 660 + ], + "type": "text", + "content": "We divide each WSI into instances " + }, + { + "bbox": [ + 304, + 564, + 545, + 660 + ], + "type": "inline_equation", + "content": "x_{k}" + }, + { + "bbox": [ + 304, + 564, + 545, + 660 + ], + "type": "text", + "content": " and encode these instances into embeddings " + }, + { + "bbox": [ + 304, + 564, + 545, + 660 + ], + "type": "inline_equation", + "content": "e_{k} \\in R^{D}" + }, + { + "bbox": [ + 304, + 564, + 545, + 660 + ], + "type": "text", + "content": " using pre-trained vision encoder " + }, + { + "bbox": [ + 304, + 564, + 545, + 660 + ], + "type": "inline_equation", + "content": "E_{img}" + }, + { + "bbox": [ + 304, + 564, + 545, + 660 + ], + "type": "text", + "content": ", composed with ResNet structure following [15], here " + }, + { + "bbox": [ + 304, + 564, + 545, + 660 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 564, + 545, + 660 + ], + "type": "text", + "content": " represents the dimension of the embeddings. Then we send the instance embeddings into the TFS Module to aggregate the instance features and prompts, and obtain the bag-level embeddings " + }, + { + "bbox": [ + 304, + 564, + 545, + 660 + ], + "type": "inline_equation", + "content": "v_{i} \\in R^{D}" + }, + { + "bbox": [ + 304, + 564, + 545, + 660 + ], + "type": "text", + "content": ". The formulas are shown as follows:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 377, + 667, + 545, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 667, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 377, + 667, + 545, + 681 + ], + "type": "interline_equation", + "content": "e _ {k} = E _ {i m g} \\left(x _ {k}\\right) \\tag {5}", + "image_path": "57d0d2fc36bf8d8e3983be306debdad8d72db6e4dafb2ba55c97eab6c97fe74c.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 380, + 682, + 545, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 682, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 380, + 682, + 545, + 696 + ], + "type": "interline_equation", + "content": "I _ {i} = \\left[ e _ {i 1}, e _ {i 2}, \\dots , e _ {i k} \\right] \\tag {6}", + "image_path": "34a1c3641f4e42627f717a72adec6fa95abcb0e25d692fa87b86fa27858bde5a.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 380, + 697, + 545, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 697, + 545, + 711 + ], + "spans": [ + { + "bbox": [ + 380, + 697, + 545, + 711 + ], + "type": "interline_equation", + "content": "v _ {i} = I A \\left(I _ {i}, Q\\right) \\tag {7}", + "image_path": "38188d860ff7d8aab1098a17c7c98672ef38f1956a9184d6e27a0835782ab8ec.jpg" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11402" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": "Besides, we generate pathologically meaningful text embeddings for each WSI, represented as " + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "inline_equation", + "content": "t_i \\in R^D" + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": ", by leveraging the fine-tuned text encoder BioClinicalBERT [30]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 133, + 118, + 287, + 131 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 118, + 287, + 131 + ], + "spans": [ + { + "bbox": [ + 133, + 118, + 287, + 131 + ], + "type": "interline_equation", + "content": "t _ {i} ^ {c} = E _ {t x t} \\left(x _ {t x t} ^ {c}\\right) \\tag {8}", + "image_path": "f1682e248559ac091d228f5ce2d6487471e65505ceb6753ae96c2e24e893cc1e.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 139, + 287, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 139, + 287, + 210 + ], + "spans": [ + { + "bbox": [ + 47, + 139, + 287, + 210 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 139, + 287, + 210 + ], + "type": "inline_equation", + "content": "E_{txt}" + }, + { + "bbox": [ + 47, + 139, + 287, + 210 + ], + "type": "text", + "content": " denotes the text encoder, and " + }, + { + "bbox": [ + 47, + 139, + 287, + 210 + ], + "type": "inline_equation", + "content": "x_{txt}^{c}(c \\in [1, C])" + }, + { + "bbox": [ + 47, + 139, + 287, + 210 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 47, + 139, + 287, + 210 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 139, + 287, + 210 + ], + "type": "text", + "content": " denotes the number of categories. Here we use the same embedding dimension " + }, + { + "bbox": [ + 47, + 139, + 287, + 210 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 47, + 139, + 287, + 210 + ], + "type": "text", + "content": " as the vision encoder, suitable for contrastive learning. For text encoder " + }, + { + "bbox": [ + 47, + 139, + 287, + 210 + ], + "type": "inline_equation", + "content": "E_{txt}" + }, + { + "bbox": [ + 47, + 139, + 287, + 210 + ], + "type": "text", + "content": ", we adopt the Low-Rank Adaptation (LoRA) [7] approach for efficient fine-tuning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 211, + 287, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 211, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 47, + 211, + 287, + 270 + ], + "type": "text", + "content": "Subsequently, the bag-level embeddings " + }, + { + "bbox": [ + 47, + 211, + 287, + 270 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 47, + 211, + 287, + 270 + ], + "type": "text", + "content": " are aligned with the text embeddings " + }, + { + "bbox": [ + 47, + 211, + 287, + 270 + ], + "type": "inline_equation", + "content": "t_{i}^{c}" + }, + { + "bbox": [ + 47, + 211, + 287, + 270 + ], + "type": "text", + "content": " to complete the training process. In this case, prediction " + }, + { + "bbox": [ + 47, + 211, + 287, + 270 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 47, + 211, + 287, + 270 + ], + "type": "text", + "content": " is obtained by applying softmax on scaled cosine similarities between the image embeddings and text embeddings:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 79, + 278, + 287, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 278, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 79, + 278, + 287, + 308 + ], + "type": "interline_equation", + "content": "p (\\hat {y} = c \\mid I) = \\frac {\\exp \\left(\\operatorname {s i m} \\left(t _ {i} ^ {c} , v _ {i}\\right) / \\tau\\right)}{\\sum_ {c ^ {\\prime} = 1} ^ {C} \\exp \\left(\\operatorname {s i m} \\left(t _ {i} ^ {c ^ {\\prime}} , v _ {i}\\right) / \\tau\\right)} \\tag {9}", + "image_path": "a361bd3028ae6237c65cea9ce712d534609c73c1c3c6835f93cecb55ba80be26.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 316, + 286, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 316, + 286, + 339 + ], + "spans": [ + { + "bbox": [ + 47, + 316, + 286, + 339 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 316, + 286, + 339 + ], + "type": "inline_equation", + "content": "\\text{sim}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 47, + 316, + 286, + 339 + ], + "type": "text", + "content": " refers to cosine similarity and " + }, + { + "bbox": [ + 47, + 316, + 286, + 339 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 47, + 316, + 286, + 339 + ], + "type": "text", + "content": " is the temperature parameter." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 340, + 286, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 340, + 286, + 363 + ], + "spans": [ + { + "bbox": [ + 47, + 340, + 286, + 363 + ], + "type": "text", + "content": "The fine-grained training loss is computed as the cross-entropy between the logits and soft targets as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 98, + 372, + 287, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 372, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 98, + 372, + 287, + 407 + ], + "type": "interline_equation", + "content": "L ^ {v \\rightarrow t} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {N} y _ {i j} \\log \\left(p _ {i j}\\right) \\tag {10}", + "image_path": "9f9e7ea55d64653fb49f7a3ac2d43fa3f9b75b13af8cf6acaee24696113b984b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 414, + 198, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 414, + 198, + 425 + ], + "spans": [ + { + "bbox": [ + 47, + 414, + 198, + 425 + ], + "type": "text", + "content": "here " + }, + { + "bbox": [ + 47, + 414, + 198, + 425 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 414, + 198, + 425 + ], + "type": "text", + "content": " corresponds to the batch size." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 426, + 287, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 426, + 287, + 450 + ], + "spans": [ + { + "bbox": [ + 47, + 426, + 287, + 450 + ], + "type": "text", + "content": "Likewise, we can compute " + }, + { + "bbox": [ + 47, + 426, + 287, + 450 + ], + "type": "inline_equation", + "content": "L^{t \\to v}" + }, + { + "bbox": [ + 47, + 426, + 287, + 450 + ], + "type": "text", + "content": " and serve " + }, + { + "bbox": [ + 47, + 426, + 287, + 450 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 47, + 426, + 287, + 450 + ], + "type": "text", + "content": " as the final training objective." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 125, + 458, + 287, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 458, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 125, + 458, + 287, + 483 + ], + "type": "interline_equation", + "content": "L = \\frac {L ^ {v \\rightarrow t} + L ^ {t \\rightarrow v}}{2} \\tag {11}", + "image_path": "f93d44804c60a6f0d7e9d8d0c9603334c2163a53109da240c1bd50464baa232e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 490, + 190, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 490, + 190, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 490, + 190, + 502 + ], + "type": "text", + "content": "4. Experiments and Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 510, + 110, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 510, + 110, + 521 + ], + "spans": [ + { + "bbox": [ + 47, + 510, + 110, + 521 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 528, + 287, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 528, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 528, + 287, + 563 + ], + "type": "text", + "content": "We evaluated our method on public histopathology WSI datasets: The Cancer Genome Atlas Lung (TCGA Lung) Cancer1 and Camelyon16 [2]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 563, + 287, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 563, + 287, + 695 + ], + "spans": [ + { + "bbox": [ + 46, + 563, + 287, + 695 + ], + "type": "text", + "content": "TCGA Lung Cancer. The TCGA Lung Cancer dataset comprises two cancer subtypes: Lung Adenocarcinoma (LUAD) and Lung Squamous Cell Carcinoma (LUSC). It includes diagnostic slides with 541 slides from 478 LUAD cases and 512 slides from 478 LUSC cases. For WSI preprocessing, following [15], we cropped each WSI into " + }, + { + "bbox": [ + 46, + 563, + 287, + 695 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 563, + 287, + 695 + ], + "type": "text", + "content": " non-overlapping patches and removed the background region. The dataset encompasses approximately 5.2 million patches at " + }, + { + "bbox": [ + 46, + 563, + 287, + 695 + ], + "type": "inline_equation", + "content": "20 \\times" + }, + { + "bbox": [ + 46, + 563, + 287, + 695 + ], + "type": "text", + "content": " magnification, averaging about 5,000 patches per WSI. Following [28], we randomly split the dataset into training, validation, and testing sets with a ratio" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 72, + 545, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 288 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 288 + ], + "type": "text", + "content": "of 65:10:25 on the patient level and adopted 4-fold cross-validation. We collected corresponding pathology reports " + }, + { + "bbox": [ + 306, + 72, + 545, + 288 + ], + "type": "inline_equation", + "content": "{}^{2}" + }, + { + "bbox": [ + 306, + 72, + 545, + 288 + ], + "type": "text", + "content": " and cleaned them by GPT-4 to produce fine-grained text description labels. To ensure professionalism and accuracy, we invited professional pathologists to check and correct the textual labels. Additionally, to evaluate the generalizability of the model and perform zero-shot classification, we utilized a dataset of histological subtype labels for TCGALUAD from the cBioPortal database " + }, + { + "bbox": [ + 306, + 72, + 545, + 288 + ], + "type": "inline_equation", + "content": "{}^{3}" + }, + { + "bbox": [ + 306, + 72, + 545, + 288 + ], + "type": "text", + "content": " ,More details about subtype labels are provided in Supplementary Material. Camelyon16. The Camelyon16 dataset [2] consists of 399 Hematoxylin and Eosin (H&E) stained slide images, utilized for metastasis detection in breast cancer. We preprocessed each WSI by segmenting it into " + }, + { + "bbox": [ + 306, + 72, + 545, + 288 + ], + "type": "inline_equation", + "content": "{256} \\times {256}" + }, + { + "bbox": [ + 306, + 72, + 545, + 288 + ], + "type": "text", + "content": " nonoverlapping patches, excluding background regions. In total, this process yielded approximately 2.8 million patches at a " + }, + { + "bbox": [ + 306, + 72, + 545, + 288 + ], + "type": "inline_equation", + "content": "{20} \\times" + }, + { + "bbox": [ + 306, + 72, + 545, + 288 + ], + "type": "text", + "content": " magnification level,with about 7,200 patches per bag. We adopted 3-times 3-fold cross-validation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 295, + 440, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 295, + 440, + 308 + ], + "spans": [ + { + "bbox": [ + 306, + 295, + 440, + 308 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 314, + 545, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 314, + 545, + 457 + ], + "spans": [ + { + "bbox": [ + 305, + 314, + 545, + 457 + ], + "type": "text", + "content": "In experiments, we employed ResNet following [15] as image encoder to extract image features, while pre-trained BioClinicalBERT from [30] as text encoder to generate text features. LoRA [7] was adopted for fine-tuning the text encoder, with an alpha value of 32 and a rank value of 8. We divided the whole slide image into patches with " + }, + { + "bbox": [ + 305, + 314, + 545, + 457 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 305, + 314, + 545, + 457 + ], + "type": "text", + "content": ", then applied a random crop with size " + }, + { + "bbox": [ + 305, + 314, + 545, + 457 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 305, + 314, + 545, + 457 + ], + "type": "text", + "content": ". We adopted AdamW [19] with beta (0.9, 0.98), eps 1e-8, learning rate 3e-6, warmup ratio 0.1, weight decay 1e-4 as our optimizer. Additionally, we used the batch size of 1 with an accumulation step of 8 and trained for 150 epochs. We utilized mixed-precision training on 4 NVIDIA-A800 GPUs." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 465, + 541, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 465, + 541, + 478 + ], + "spans": [ + { + "bbox": [ + 306, + 465, + 541, + 478 + ], + "type": "text", + "content": "4.3. Zero-Shot Histological Subtype Classification" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 483, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 483, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 305, + 483, + 545, + 555 + ], + "type": "text", + "content": "Prior researches [23, 28, 38] have predominantly concentrated on classifying primary cancer categories. Our approach extends beyond this by attempting to classify detailed histological subtypes. It is crucial to emphasize that this task poses a significant challenge, often proving difficult for even skilled pathologists to make direct judgments." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 555, + 545, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 555, + 545, + 675 + ], + "spans": [ + { + "bbox": [ + 305, + 555, + 545, + 675 + ], + "type": "text", + "content": "Since only the LUAD's histological subtype dataset was provided on the online database, we conducted zero-shot subtype classification evaluation on LUAD subtype datasets, with the model being pre-trained on TCGA-LUAD or TCGA-LUSC. Throughout the training phase, subtype label information was deliberately excluded, ensuring that all experiments are conducted solely with fine-grained labels. This aims to evaluate the model's ability to identify novel diagnostic categories without specific training on subtypes. We extended the morphological appearance text de" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 189, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 189, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 189, + 712 + ], + "type": "text", + "content": "1http://www.cancer.gov/ctga" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 683, + 544, + 702 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 683, + 544, + 702 + ], + "spans": [ + { + "bbox": [ + 306, + 683, + 544, + 702 + ], + "type": "text", + "content": "2https://github.com/tatonetti-lab/tcga-path-reports" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 703, + 452, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 703, + 452, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 703, + 452, + 712 + ], + "type": "text", + "content": "3https://www.cbiportal.org/" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11403" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "scription labels of the target data using GPT-4. After the text encoder encodes the labels, similarity calculation with image features achieved zero-shot classification." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 110, + 288, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 110, + 288, + 253 + ], + "spans": [ + { + "bbox": [ + 46, + 110, + 288, + 253 + ], + "type": "text", + "content": "Since existing zero-shot learning methods cannot be used in WSI subtype classification, we constructed three Linear-Probe method baselines: Mean pooling, Max pooling, and Attention pooling. As shown in Tab. 2, FiVE attains " + }, + { + "bbox": [ + 46, + 110, + 288, + 253 + ], + "type": "inline_equation", + "content": "65.23\\%" + }, + { + "bbox": [ + 46, + 110, + 288, + 253 + ], + "type": "text", + "content": " top-1 accuracy and " + }, + { + "bbox": [ + 46, + 110, + 288, + 253 + ], + "type": "inline_equation", + "content": "95.18\\%" + }, + { + "bbox": [ + 46, + 110, + 288, + 253 + ], + "type": "text", + "content": " top-5 accuracy when pre-trained with TCGA-LUAD fine-grained labels. When pre-trained with TCGA-LUSC fine-grained labels, it achieves " + }, + { + "bbox": [ + 46, + 110, + 288, + 253 + ], + "type": "inline_equation", + "content": "62.02\\%" + }, + { + "bbox": [ + 46, + 110, + 288, + 253 + ], + "type": "text", + "content": " top-1 accuracy and " + }, + { + "bbox": [ + 46, + 110, + 288, + 253 + ], + "type": "inline_equation", + "content": "94.36\\%" + }, + { + "bbox": [ + 46, + 110, + 288, + 253 + ], + "type": "text", + "content": " top-5 accuracy. Moreover, the zero-shot performance of FiVE notably exceeds that of the baseline. This capability in classifying LUAD subtypes is attributed to the focus on fine-grained pathological features during training." + } + ] + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 53, + 266, + 282, + 373 + ], + "blocks": [ + { + "bbox": [ + 53, + 266, + 282, + 373 + ], + "lines": [ + { + "bbox": [ + 53, + 266, + 282, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 266, + 282, + 373 + ], + "type": "table", + "html": "
MethodTCGA-LUADTCGA-LUSC
Top-1Top-5Top-1Top-5
Mean-pooling40.8283.4631.8682.96
Max-pooling45.0588.7736.3686.09
Attention-pooling58.3693.5354.4192.50
FiVE (Ours)65.2395.1862.0294.36
+6.87+1.65+7.61+1.86
", + "image_path": "2bc9859bca4dffac32a9fa34e4cdcc7a922914a8979b4ac296d86900d85f4354.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 452, + 180, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 452, + 180, + 464 + ], + "spans": [ + { + "bbox": [ + 47, + 452, + 180, + 464 + ], + "type": "text", + "content": "4.4. Few-Shot Classification" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 472, + 287, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 592 + ], + "type": "text", + "content": "Our model demonstrates adaptability to various tasks even in scenarios with limited data availability. Few-shot experiments were conducted to demonstrate its transferability to downstream tasks. We initialized the networks with pretrained weights derived from the model trained on TCGA image-report pairs, and subsequently fine-tuned the model on downstream datasets for few-shot image classification. We followed [23] and conducted experiments with 1, 2, 4, 8, 16, and additional 0 shot on the downstream dataset. The results are summarized in Tab. 3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": "Our model exhibits remarkable performance in zero-shot classification, achieving an accuracy of " + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "inline_equation", + "content": "71.26\\%" + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": ", even surpassing the SOTA method's one-shot experiment. Upon the introduction of training data, our models display exceptional transferability, outperforming the SOTA by " + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "inline_equation", + "content": "12.90\\%" + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": " in the one-shot setting. The model's performance improves accordingly with the number of shot. Upon reaching 16-shot, our model reaches an impressive accuracy of " + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "inline_equation", + "content": "91.25\\%" + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": ", showcasing a notable " + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "inline_equation", + "content": "9.19\\%" + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": " improvement close to the fully supervised performance level." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 309, + 70, + 544, + 160 + ], + "blocks": [ + { + "bbox": [ + 46, + 381, + 287, + 426 + ], + "lines": [ + { + "bbox": [ + 46, + 381, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 381, + 287, + 426 + ], + "type": "text", + "content": "Table 2. Zero-Shot performance on histological subtype classification. We pre-trained the model using fine-grained labels of TCGA-LUAD and TCGA-LUSC, then applied zero-shot classification to the histological subtypes of TCGA-LUAD." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 70, + 544, + 160 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 544, + 160 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 544, + 160 + ], + "type": "table", + "html": "
Method16-shot8-shot4-shot2-shot1-shot0-shot
Mean-pool65.3353.8944.8552.9345.34\\
Max-pool48.4849.5544.2248.3949.03\\
Attn-pool72.5065.7962.4758.3656.23\\
CoOp [40]78.3567.9967.6067.5467.81\\
TOP [23]82.0680.5175.4172.3871.01\\
FiVE91.2590.8088.1085.5183.9171.26
(Ours)+9.19+10.29+12.69+13.13+12.90
", + "image_path": "44a4b61d9aebe11ccfffa069fbfbf1eec6de362c197ed3cc458347ee5e7c29fa.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 168, + 547, + 214 + ], + "lines": [ + { + "bbox": [ + 304, + 168, + 547, + 214 + ], + "spans": [ + { + "bbox": [ + 304, + 168, + 547, + 214 + ], + "type": "text", + "content": "Table 3. Few-shot classification performance on TCGA Lung Cancer. Mean-pool, Max-pool, and Attn-pool correspond to Linear-Probe implementations with Mean-pooling, Max-pooling, and Attention-pooling, respectively." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 236, + 545, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 236, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 305, + 236, + 545, + 249 + ], + "type": "text", + "content": "4.5. Performance Comparison with Existing Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 255, + 545, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 545, + 351 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 545, + 351 + ], + "type": "text", + "content": "We compared FiVE with ABMIL [10], DSMIL [15], CLAM-SB [20], CLAM-MB [20], TransMIL [26], DTFDMIL [36], and MHIM-MIL [28], all of which are attention-based MIL methods. In addition, we included two traditional MIL pooling operations, Max-pooling and Mean-pooling, for comparison. The results of all other methods are reproduced using the official code they provide under the same settings." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 352, + 546, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 352, + 546, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 352, + 546, + 519 + ], + "type": "text", + "content": "To evaluate the performance of FiVE, we conducted fine-grained pre-training exclusively on the TCGA dataset, followed by fully supervised experiments on Camelyon16 and TCGA Lung Cancer datasets for WSI classification. Note that the test data is not used in pre-training. Results are shown in Tab. 4. It can be found that our method outperforms all the other baselines by a great margin, which fully demonstrates the significance of our fine-grained training scheme in improving performance on downstream tasks. Besides, since the pre-training data mainly comes from TCGA data, on the other hand, the task-specific prompt design is more suitable for TCGA data. This results in the performance improvement of our method for TCGA data being greater than Camelyon16 data." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 529, + 406, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 529, + 406, + 542 + ], + "spans": [ + { + "bbox": [ + 306, + 529, + 406, + 542 + ], + "type": "text", + "content": "4.6. Ablation Studies" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 549, + 462, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 549, + 462, + 561 + ], + "spans": [ + { + "bbox": [ + 306, + 549, + 462, + 561 + ], + "type": "text", + "content": "4.6.1 Effectiveness of TFS Module" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": "We focused on evaluating the impact of the TFS module on the model's overall performance with the TCGA Lung Cancer. The results detailed in Tab. 5 reveal significant improvements at each stage of feature enhancement. Initially, the model with only Self Attention attained ACC of " + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "inline_equation", + "content": "89.77\\%" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": ", AUC of " + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "inline_equation", + "content": "92.85\\%" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": ", and F1-score of " + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "inline_equation", + "content": "89.95\\%" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": ". The incorporation of fine-grained labels led to increases of " + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "inline_equation", + "content": "1.44\\%" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": " in ACC, " + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "inline_equation", + "content": "1.51\\%" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": " in AUC, and " + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "inline_equation", + "content": "1.43\\%" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": " in F1-score. Subsequent integration of additional fine-grained guidance further improved performance. Ultimately, full framework with the TFS Module achieved the highest performance of " + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "inline_equation", + "content": "94.62\\%" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": " ACC, " + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "inline_equation", + "content": "96.33\\%" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": " AUC, and " + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "inline_equation", + "content": "93.89\\%" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": " F1-score. These ablation" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "11404" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 77, + 70, + 518, + 232 + ], + "blocks": [ + { + "bbox": [ + 77, + 70, + 518, + 232 + ], + "lines": [ + { + "bbox": [ + 77, + 70, + 518, + 232 + ], + "spans": [ + { + "bbox": [ + 77, + 70, + 518, + 232 + ], + "type": "table", + "html": "
MethodCamelyon16TCGA Lung Cancer
ACCAUCF1-scoreACCAUCF1-score
Max-pooling78.95±2.2881.28±3.7471.06±2.5981.49±1.2486.45±0.7180.56±1.09
Mean-pooling76.69±0.2080.07±0.7870.41±0.1684.14±2.9790.13±2.4083.39±3.14
ABMIL [10]90.06±0.6094.00±0.8387.40±1.0588.03±2.1993.17±2.0587.41±2.42
DSMIL [15]90.17±1.0294.57±0.4087.65±1.1888.32±2.7093.71±1.8287.90±2.50
CLAM-SB [20]90.32±0.1294.65±0.3087.89±0.5987.74±2.2293.67±1.6487.36±2.24
CLAM-MB [20]90.14±0.8594.70±0.7688.10±0.6388.73±1.6293.69±0.5488.28±1.58
TransMIL [26]89.22±2.3293.51±2.1385.10±4.3387.08±1.9792.51±1.7686.40±2.08
DTFD-MIL [36]90.22±0.3695.15±0.1487.62±0.5988.23±2.1293.83±1.3987.71±2.04
MHIM-MIL [28]92.48±0.3596.49±0.6590.75±0.7389.93±3.3795.53±1.7489.71±2.92
FiVE (Ours)94.25±0.3397.56±0.5993.24±0.7294.62±2.1396.33±1.2193.89±1.90
", + "image_path": "1bdf47fcf107708d170daee9c7880782f8f2761fd80356a34a6685cc04a78994.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 269, + 286, + 334 + ], + "blocks": [ + { + "bbox": [ + 118, + 239, + 474, + 251 + ], + "lines": [ + { + "bbox": [ + 118, + 239, + 474, + 251 + ], + "spans": [ + { + "bbox": [ + 118, + 239, + 474, + 251 + ], + "type": "text", + "content": "Table 4. Comparison Performance of slide classification on Camelyon16 and TCGA Lung Cancer." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 269, + 286, + 334 + ], + "lines": [ + { + "bbox": [ + 48, + 269, + 286, + 334 + ], + "spans": [ + { + "bbox": [ + 48, + 269, + 286, + 334 + ], + "type": "table", + "html": "
SAFGLFGGLDPACCAUCF1-score
✓89.7792.8589.58
✓✓91.2194.3691.01
✓✓✓93.5696.0193.17
✓✓✓✓94.6296.3393.89
", + "image_path": "469b212bbaf2a191fedaec192c91e5037f792ec6cca093978d3efb47157a7ea4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 341, + 287, + 386 + ], + "lines": [ + { + "bbox": [ + 46, + 341, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 46, + 341, + 287, + 386 + ], + "type": "text", + "content": "Table 5. The ablation experiments of the pre-trained model finetuned on the TCGA Lung Cancer. SA, FGL, FGG, and LDP represent Self Attention, Fine-Grained Labels, Fine-Grained Guidance, and Learnable Diagnosis Prompts, respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 398, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 398, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 398, + 287, + 422 + ], + "type": "text", + "content": "experiments highlight the benefits of integrating these methods in the WSI classification task." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 438, + 246, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 438, + 246, + 451 + ], + "spans": [ + { + "bbox": [ + 47, + 438, + 246, + 451 + ], + "type": "text", + "content": "4.6.2 Effectiveness of Patch Sample Strategy" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 458, + 287, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 458, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 46, + 458, + 287, + 543 + ], + "type": "text", + "content": "We verified the impact of different sampling strategies on model performance on TCGA data. Here, we assumed that there are two main indicators that affect model performance, sample ratio and max sample threshold (represented as MAXN). At the same time, in order to differentiate the experimental results, we used the unfrozen image encoder for experimental verification." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 50, + 554, + 165, + 626 + ], + "blocks": [ + { + "bbox": [ + 50, + 554, + 165, + 626 + ], + "lines": [ + { + "bbox": [ + 50, + 554, + 165, + 626 + ], + "spans": [ + { + "bbox": [ + 50, + 554, + 165, + 626 + ], + "type": "image", + "image_path": "46a414cffb3970e915eb9e2bcc441f82f580a6c0d9aa1fc3bde5b1e90bd3a029.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 630, + 286, + 663 + ], + "lines": [ + { + "bbox": [ + 46, + 630, + 286, + 663 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 286, + 663 + ], + "type": "text", + "content": "Figure 3. Classification performance on TCGA Lung Cancer with diverse sampling strategies, presenting average and standard deviation (std) ACC values." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 168, + 554, + 284, + 626 + ], + "blocks": [ + { + "bbox": [ + 168, + 554, + 284, + 626 + ], + "lines": [ + { + "bbox": [ + 168, + 554, + 284, + 626 + ], + "spans": [ + { + "bbox": [ + 168, + 554, + 284, + 626 + ], + "type": "image", + "image_path": "dcefa4dcf6a6b3ec8e8923aa33e0d8e8909eb6510db76eb431b162b5e66f8c25.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "text", + "content": "As shown in Fig. 3, when MAXN is sufficiently large (" + }, + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\geq" + }, + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "text", + "content": " 2048), the correlation between the model's performance and the sample ratio is not significant. Even with a small sample" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 272, + 545, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 272, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 304, + 272, + 545, + 380 + ], + "type": "text", + "content": "ratio " + }, + { + "bbox": [ + 304, + 272, + 545, + 380 + ], + "type": "inline_equation", + "content": "(\\leq 0.5)" + }, + { + "bbox": [ + 304, + 272, + 545, + 380 + ], + "type": "text", + "content": ", the model can still effectively align with the original data distribution. Conversely, when MAXN is not large enough (<2048), the primary change in the model's performance depends on whether the magnitude of the sample ratio can match the original data distribution. When the sample ratio increases to 0.75, a steep performance improvement can be observed, suggesting that the sampled data at this point conforms to the original data distribution, after which the model's performance stabilizes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 380, + 546, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 380, + 546, + 477 + ], + "spans": [ + { + "bbox": [ + 304, + 380, + 546, + 477 + ], + "type": "text", + "content": "Based on comprehensive experimental results, we established the sample ratio and MAXN as 0.5 and 2048 as suitable hyperparameters for the unfrozen image encoder. As for the frozen image encoder experiment, taking into account the performance constraints imposed by the frozen image encoder on the model's performance upper limit [38], we recommend the sample ratio and MAXN to be set at 0.5 and 16384 based on our experiments." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 489, + 378, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 489, + 378, + 502 + ], + "spans": [ + { + "bbox": [ + 306, + 489, + 378, + 502 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "In this paper, we introduce FiVE, a novel framework that demonstrates robust generalization and strong transferability for WSI classification. Our work pioneers the utilization of non-standardized WSI-report pairs from public databases to develop a VLM. To capture the complexities and diversity within these reports, we introduce the Task-specific Fine-grained Semantics (TFS) module. This module reconstructs fine-grained labels and diagnosis prompts during training, enhancing the semantic relevance of its features by introducing diagnosis prompts. Furthermore, considering that pathological visual patterns are redundantly distributed across tissue slices, we employ a sampling strategy to reduce computational costs. Our experiments demonstrate the robust generalizability and computational efficiency of the proposed framework, which can also be easily adapted to other tasks with minimal fine-tuning. We aspire to provide empirical insights and contribute to AI pathology research." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11405" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "text", + "content": "[1] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Visual prompting: Modifying pixel space to adapt pre-trained models. arXiv preprint arXiv:2203.17274, 3:11-12, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 213 + ], + "type": "text", + "content": "[2] Babak Ehteshami Bejnordi, Mitko Veta, Paul Johannes Van Diest, Bram Van Ginneken, Nico Karssemeijer, Geert Litjens, Jeroen AWM Van Der Laak, Meyke Hermsen, Quirine F Manson, Maschenka Balkenhol, et al. Diagnostic assessment of deep learning algorithms for detection of lymph node metastases in women with breast cancer. Jama, 318(22):2199-2210, 2017. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 215, + 287, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 215, + 287, + 280 + ], + "spans": [ + { + "bbox": [ + 53, + 215, + 287, + 280 + ], + "type": "text", + "content": "[3] Gabriele Campanella, Matthew G Hanna, Luke Geneslaw, Allen Miraflor, Vitor Werneck Krauss Silva, Klaus J Busam, Edi Brogi, Victor E Reuter, David S Klimstra, and Thomas J Fuchs. Clinical-grade computational pathology using weakly supervised deep learning on whole slide images. Nature medicine, 25(8):1301-1309, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 282, + 287, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 282, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 282, + 287, + 327 + ], + "type": "text", + "content": "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 328, + 287, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 328, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 328, + 287, + 384 + ], + "type": "text", + "content": "[5] Tsai Hor Chan, Fernando Julio Cendra, Lan Ma, Guosheng Yin, and Lequan Yu. Histopathology whole slide image analysis with heterogeneous graph representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15661-15670, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 384, + 287, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 384, + 287, + 438 + ], + "spans": [ + { + "bbox": [ + 53, + 384, + 287, + 438 + ], + "type": "text", + "content": "[6] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xi-angyu Yue, et al. Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010, 2023. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 441, + 287, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 441, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 53, + 441, + 287, + 485 + ], + "type": "text", + "content": "[7] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 487, + 287, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 487, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 53, + 487, + 287, + 531 + ], + "type": "text", + "content": "[8] Zhi Huang, Federico Bianchi, Mert Yuksekgonul, Thomas J Montine, and James Zou. A visual-language foundation model for pathology image analysis using medical twitter. Nature Medicine, pages 1-10, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 533, + 287, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 533, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 53, + 533, + 287, + 564 + ], + "type": "text", + "content": "[9] Carolyn Hutter and Jean Claude Zenklusen. The cancer genome atlas: creating lasting value beyond its data. Cell, 173(2):283-285, 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 567, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 610 + ], + "type": "text", + "content": "[10] Maximilian Ilse, Jakub Tomczak, and Max Welling. Attention-based deep multiple instance learning. In International conference on machine learning, pages 2127-2136. PMLR, 2018. 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 677 + ], + "type": "text", + "content": "[11] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[12] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In Proceedings of the" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19113-19122, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 96, + 545, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 163 + ], + "type": "text", + "content": "[13] Zhengfeng Lai, Zhuoheng Li, Luca Cerny Oliveira, Joohi Chauhan, Brittany N Dugger, and Chen-Nee Chuah. Clipse: Fine-tune clip with visual feature fusion for pathology image analysis towards minimizing data collection efforts. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2374-2380, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 163, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 251 + ], + "type": "text", + "content": "[14] Marvin Lerousseau, Maria Vakalopoulou, Marion Classe, Julien Adam, Enzo Battistella, Alexandre Carré, Théo Estienne, Théophraste Henry, Eric Deutsch, and Nikos Paragios. Weakly supervised multiple instance learning histopathological tumor segmentation. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part V 23, pages 470-479. Springer, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 253, + 545, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 253, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 308, + 253, + 545, + 317 + ], + "type": "text", + "content": "[15] Bin Li, Yin Li, and Kevin W Eliceiri. Dual-stream multiple instance learning network for whole slide image classification with self-supervised contrastive learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14318-14328, 2021. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 319, + 545, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 319, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 308, + 319, + 545, + 396 + ], + "type": "text", + "content": "[16] Honglin Li, Chenglu Zhu, Yunlong Zhang, Yuxuan Sun, Zhongyi Shui, Wenwei Kuang, Sunyi Zheng, and Lin Yang. Task-specific fine-tuning via variational information bottleneck for weakly-supervised pathology whole slide image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7454-7463, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 397, + 545, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 397, + 545, + 452 + ], + "spans": [ + { + "bbox": [ + 308, + 397, + 545, + 452 + ], + "type": "text", + "content": "[17] Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. Scaling language-image pre-training via masking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23390-23400, 2023. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 454, + 545, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 454, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 308, + 454, + 545, + 509 + ], + "type": "text", + "content": "[18] Tiancheng Lin, Zhimiao Yu, Hongyu Hu, Yi Xu, and Chang-Wen Chen. Interventional bag multi-instance learning on whole-slide pathological images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19830-19839, 2023. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 510, + 545, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 510, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 308, + 510, + 545, + 532 + ], + "type": "text", + "content": "[19] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 533, + 545, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 533, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 308, + 533, + 545, + 588 + ], + "type": "text", + "content": "[20] Ming Y Lu, Drew FK Williamson, Tiffany Y Chen, Richard J Chen, Matteo Barbieri, and Faisal Mahmood. Data-efficient and weakly supervised computational pathology on whole-slide images. Nature biomedical engineering, 5(6):555-570, 2021. 2, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 590, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 545, + 644 + ], + "type": "text", + "content": "[21] Ming Y Lu, Bowen Chen, Drew FK Williamson, Richard J Chen, Ivy Liang, Tong Ding, Guillaume Jaume, Igor Odintsov, Andrew Zhang, Long Phi Le, et al. Towards a visual-language foundation model for computational pathology. arXiv preprint arXiv:2307.12914, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 646, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 646, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 646, + 545, + 689 + ], + "type": "text", + "content": "[22] Yuning Lu, Jianzhuang Liu, Yonggang Zhang, Yajing Liu, and Xinmei Tian. Prompt distribution learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5206-5215, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[23] Linhao Qu, Xiaoyuan Luo, Kexue Fu, Manning Wang, and Zhijian Song. The rise of ai language pathologists:" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "11406" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "Exploring two-level prompt learning for few-shot weakly-supervised whole slide image classification. arXiv preprint arXiv:2305.17891, 2023. 2, 3, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 288, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 288, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 288, + 172 + ], + "type": "text", + "content": "[24] Linhao Qu, Zhiwei Yang, Minghong Duan, Yingfan Ma, Shuo Wang, Manning Wang, and Zhijian Song. Boosting whole slide image classification from the perspectives of distribution, correlation and magnification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21463-21473, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 288, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 288, + 237 + ], + "type": "text", + "content": "[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 238, + 288, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 238, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 238, + 288, + 293 + ], + "type": "text", + "content": "[26] Zhuchen Shao, Hao Bian, Yang Chen, Yifeng Wang, Jian Zhang, Xiangyang Ji, et al. Transmil: Transformer based correlated multiple instance learning for whole slide image classification. Advances in neural information processing systems, 34:2136-2147, 2021. 2, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 293, + 288, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 293, + 288, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 293, + 288, + 360 + ], + "type": "text", + "content": "[27] Zhuchen Shao, Yifeng Wang, Yang Chen, Hao Bian, Shaohui Liu, Haoqian Wang, and Yongbing Zhang. Lnpl-mil: Learning from noisy pseudo labels for promoting multiple instance learning in whole slide image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21495-21505, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 360, + 288, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 288, + 425 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 288, + 425 + ], + "type": "text", + "content": "[28] Wenhao Tang, Sheng Huang, Xiaoxian Zhang, Fengtao Zhou, Yi Zhang, and Bo Liu. Multiple instance learning framework with masked hard instance mining for whole slide image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4078-4087, 2023. 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 426, + 288, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 288, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 288, + 459 + ], + "type": "text", + "content": "[29] Xinggang Wang, Yongluan Yan, Peng Tang, Xiang Bai, and Wenyu Liu. Revisiting multiple instance neural networks. Pattern Recognition, 74:15-24, 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 459, + 288, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 459, + 288, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 459, + 288, + 514 + ], + "type": "text", + "content": "[30] Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chau-mond, Clement Delangue, Anthony Moi, Pierrick Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, et al. Huggingface's transformers: State-of-the-art natural language processing. arXiv preprint arXiv:1910.03771, 2019. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 514, + 288, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 288, + 558 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 288, + 558 + ], + "type": "text", + "content": "[31] Jinxi Xiang and Jun Zhang. Exploring low-rank property in multiple instance learning for whole slide image classification. In The Eleventh International Conference on Learning Representations, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 559, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 288, + 624 + ], + "type": "text", + "content": "[32] Gang Xu, Zhigang Song, Zhuo Sun, Calvin Ku, Zhe Yang, Cancheng Liu, Shuhao Wang, Jianpeng Ma, and Wei Xu. Camel: A weakly supervised learning framework for histopathology image segmentation. In Proceedings of the IEEE/CVF International Conference on computer vision, pages 10682-10691, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 624, + 288, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 288, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 288, + 669 + ], + "type": "text", + "content": "[33] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. arXiv preprint arXiv:2111.07783, 2021. 1, 2, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 669, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 288, + 712 + ], + "type": "text", + "content": "[34] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 428 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 128 + ], + "type": "text", + "content": "[35] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18123-18133, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 129, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 205 + ], + "type": "text", + "content": "[36] Hongrun Zhang, Yanda Meng, Yitian Zhao, Yihong Qiao, Xiaoyun Yang, Sarah E Coupland, and Yalin Zheng. Dtdmil: Double-tier feature distillation multiple instance learning for histopathology whole slide image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18802-18812, 2022. 2, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 205, + 545, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 205, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 307, + 205, + 545, + 261 + ], + "type": "text", + "content": "[37] Renrui Zhang, Jiaming Han, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, Peng Gao, and Yu Qiao. Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023. 3, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 261, + 545, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 261, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 307, + 261, + 545, + 328 + ], + "type": "text", + "content": "[38] Yunkun Zhang, Jin Gao, Mu Zhou, Xiaosong Wang, Yu Qiao, Shaoting Zhang, and Dequan Wang. Text-guided foundation model adaptation for pathological image classification. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 272-282. Springer, 2023. 1, 2, 3, 6, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 328, + 545, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 328, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 307, + 328, + 545, + 384 + ], + "type": "text", + "content": "[39] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 384, + 545, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 384, + 545, + 428 + ], + "spans": [ + { + "bbox": [ + 307, + 384, + 545, + 428 + ], + "type": "text", + "content": "[40] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 3, 7" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11407" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Generalized Event Cameras/e213eaf6-87b7-49bd-b232-571683e80e75_content_list.json b/2024/Generalized Event Cameras/e213eaf6-87b7-49bd-b232-571683e80e75_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..376460778e50028521fe303c3d3443929a72ae9d --- /dev/null +++ b/2024/Generalized Event Cameras/e213eaf6-87b7-49bd-b232-571683e80e75_content_list.json @@ -0,0 +1,2402 @@ +[ + { + "type": "text", + "text": "Generalized Event Cameras", + "text_level": 1, + "bbox": [ + 348, + 131, + 620, + 148 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Varun Sundar†*", + "bbox": [ + 223, + 180, + 339, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "vsundar4@wisc.edu", + "bbox": [ + 205, + 202, + 352, + 214 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Matthew Dutson†*", + "bbox": [ + 408, + 180, + 553, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "dutson@wisc.edu", + "bbox": [ + 416, + 202, + 545, + 214 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Andrei Ardelean†", + "bbox": [ + 620, + 180, + 751, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "a.ardelean@epfl.ch", + "bbox": [ + 609, + 202, + 764, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Claudio Bruschini $^{\\S}$", + "bbox": [ + 225, + 222, + 370, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{claudio.bruschini, edoardo.charbon}@epfl.ch", + "bbox": [ + 194, + 244, + 566, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mohit Gupta†", + "bbox": [ + 645, + 223, + 750, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mohitg@cs.wisc.edu", + "bbox": [ + 620, + 244, + 774, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "†University of Wisconsin-Madison ‡NovoViz École Polytechnique Fédérale de Lausanne", + "bbox": [ + 133, + 266, + 833, + 285 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "wisionlab.com/project/generalized-event-cameras/", + "bbox": [ + 276, + 287, + 691, + 303 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 235, + 339, + 310, + 354 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Event cameras capture the world at high time resolution and with minimal bandwidth requirements. However, event streams, which only encode changes in brightness, do not contain sufficient scene information to support a wide variety of downstream tasks. In this work, we design generalized event cameras that inherently preserve scene intensity in a bandwidth-efficient manner. We generalize event cameras in terms of when an event is generated and what information is transmitted. To implement our designs, we turn to single-photon sensors that provide digital access to individual photon detections; this modality gives us the flexibility to realize a rich space of generalized event cameras. Our single-photon event cameras are capable of high-speed, high-fidelity imaging at low readout rates. Consequently, these event cameras can support plug-and-play downstream inference, without capturing new event datasets or designing specialized event-vision models. As a practical implication, our designs, which involve lightweight and near-sensor-compatible computations, provide a way to use single-photon sensors without exorbitant bandwidth costs.", + "bbox": [ + 76, + 369, + 473, + 656 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 666, + 202, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Event cameras [6, 34, 44] sense the world at high speeds, providing visual information with minimal bandwidth and power. They achieve this by transmitting only changes in scene brightness, when significant \"events\" occur. However, there is a cost. Raw event data, a sparse stream of binary values, does not hold sufficient information to be used directly with mainstream vision algorithms. Therefore, while event cameras have been successful at certain tasks (e.g., object tracking [15, 33], obstacle avoidance [7, 20, 49], and high-speed odometry [12, 21, 68]), they are not widely deployed as general-purpose vision sen", + "bbox": [ + 75, + 691, + 470, + 843 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sors, and often need to be supplemented with conventional cameras [14, 15, 21]. These limitations are holding back this otherwise powerful technology.", + "bbox": [ + 500, + 340, + 890, + 385 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Is it possible to realize the promise of event cameras, i.e., high temporal resolution at low bandwidth, while preserving rich scene intensity information? To realize these seemingly conflicting goals, we propose a novel family of generalized event cameras. We conceptualize a space of event cameras along two key axes (Fig. 1 (top)): (a) \"when to transmit information,\" formalized as a change detection procedure $\\Delta$ ; and (b) \"what information to transmit,\" characterized by an integrator $\\Sigma$ that encodes incident flux. Existing event cameras represent one operating point in this $(\\Sigma, \\Delta)$ space. Our key observation is that by exploring this space and considering new $(\\Sigma, \\Delta)$ combinations, we can design event cameras that preserve scene intensity. We propose more general integrators, e.g., that represent flux according to motion levels, that span spatial patches, or that employ temporal coding (Fig. 1 (middle)). We also introduce robust change detectors that better distinguish motion from noise, by considering increased spatiotemporal contexts and modeling noise in the sensor measurements.", + "bbox": [ + 496, + 386, + 893, + 656 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite their conceptual appeal, physically implementing generalized event cameras is a challenge. This is because the requisite computations must be performed at the sensor to achieve the desired bandwidth reductions. For example, existing event cameras perform simple integration and thresholding operations via analog in-pixel circuitry. However, more general $(\\Sigma ,\\Delta)$ combinations are not always amenable to analog implementations; even feasible designs might require years of hardware iteration and production scaling. To build physical realizations of generalized event cameras, we leverage an emerging sensor technology: single-photon avalanche diodes (SPADs) that provide digital access to photon detections at extremely high frame rates $(\\sim 100\\mathrm{kHz})$ . This allows us to compose arbitrary software-level signal transformations, such as those required by generalized event cameras. Further, we are not locked to a particular event camera design and can realize multiple configurations", + "bbox": [ + 496, + 659, + 895, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*denotes equal contribution. This research was supported in parts by NSF CAREER award 1943149, NSF award CNS-2107060, and the Swiss National Science Foundation grant 200021_166289. We thank Paul Mos for providing us access to SwissSPAD2 acquisition software.", + "bbox": [ + 75, + 849, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "25007", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Conventional event camera", + "text_level": 1, + "bbox": [ + 406, + 93, + 563, + 103 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/eb323521c6dfc72ea4b63c3af2a9ca83b4aedb0ab5742ac8e2be0aed01ad0c1c.jpg", + "image_caption": [ + "Integrator (accumulates flux, $\\Sigma$ )", + "Low-pass filtered brightness" + ], + "image_footnote": [], + "bbox": [ + 83, + 125, + 263, + 207 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/cb93cdc2af2460c9e6a311222e17c6b042980154ebefcdccaf0c3666cda7f75f.jpg", + "image_caption": [ + "Change detector (triggers events, $\\Delta$", + "Threshold-based detector Analog comparator" + ], + "image_footnote": [], + "bbox": [ + 287, + 125, + 500, + 207 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3a97b6196287fcae231ef2b60dc0a55be02d6d8d07bb57211323ba839ee4c0c4.jpg", + "image_caption": [ + "Sensor output", + "Stream of change polarities $(x,y,t,\\pm 1)$" + ], + "image_footnote": [], + "bbox": [ + 517, + 125, + 681, + 207 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8327e0f3b32b651a89e7934c762422219b2e46eef0dedcaca0058a77b8fca86b.jpg", + "image_caption": [ + "Intensity imaging", + "Reconstruction with E2VID+ 1000 FPS at 20 FPS bitrate" + ], + "image_footnote": [], + "bbox": [ + 700, + 127, + 885, + 207 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generalized event camera", + "text_level": 1, + "bbox": [ + 410, + 232, + 560, + 243 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/23f46768133bc5f79c9122299515b4f26b23a2361d6b17825fa94dc5caf2821c.jpg", + "image_caption": [ + "Motion adaptive", + "Spatial patches", + "Coded exposures" + ], + "image_footnote": [], + "bbox": [ + 84, + 250, + 264, + 332 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a39965934d4a9cfeb632084a849d16c3cc0970fd6aaa7eecd4191a33803b531f.jpg", + "image_caption": [ + "Enhanced spatio-temporal support Noise-aware thresholding" + ], + "image_footnote": [], + "bbox": [ + 287, + 251, + 501, + 330 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/91ef207541236e1bb5e060fc371fdc3e704e5ec4fe7d6b1fb2c3583f64a2450d.jpg", + "image_caption": [ + "Stream of integrator values $(x,y,t,\\Sigma (x,y,t))$" + ], + "image_footnote": [], + "bbox": [ + 519, + 250, + 681, + 330 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4ca0635e9f44a5a0f1241e046abe02096fe14bcc7dd3a3729e8621b868f11060.jpg", + "image_caption": [ + "Reconstruction 3025 FPS at 30 FPS bitrate" + ], + "image_footnote": [], + "bbox": [ + 699, + 252, + 885, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Plug-and-play event vision", + "text_level": 1, + "bbox": [ + 408, + 358, + 560, + 369 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2133d106cbe4c4917aeb158ba7c31d382cc7d2ca075c17a4d2995251bfb4a23d.jpg", + "image_caption": [ + "Contact point of a tennis shot Image Reconstruction", + "Figure 1. Generalized event cameras. (top) Event cameras generate outputs in response to abrupt changes in scene intensity. We describe this as a combination of a low-pass integrator and a threshold-based change detector. (middle) We generalize the space of event cameras by designing integrators that capture rich intensity information, and more reliable change detectors that utilize larger spatiotemporal contexts and noise-aware thresholding (Secs. 4.1 to 4.3). Unlike existing events, our generalized event streams inherently preserve scene intensity, e.g., this ping-pong ball slingshotted against a brick wall backdrop. (bottom) Generalized event cameras enable high-fidelity bandwidth-efficient imaging: providing 3025 FPS reconstructions with a readout equivalent to a 30 FPS camera. Consequently, generalized events facilitate plug-and-play inference on a multitude of tasks in challenging scenarios (insets depict the extent of motion over $30~\\mathrm{ms}$ )." + ], + "image_footnote": [], + "bbox": [ + 84, + 376, + 241, + 454 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b7c0e23152bc6b87266d6c12e39abceea82b4a8814836b4b1e1fbfaa0991e3f9.jpg", + "image_caption": [ + "Tracking points at 0.3 lux CoTracker [29]" + ], + "image_footnote": [], + "bbox": [ + 245, + 377, + 403, + 454 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/49f6658d4e310202bac994ccd084296f5c5bcbf1d185c35a77e4d318df22540b.jpg", + "image_caption": [ + "Nighttime object detection DETR [11]" + ], + "image_footnote": [], + "bbox": [ + 406, + 377, + 563, + 454 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/687e4267a552b02ea9648304fc4ea45c5331f352efe661270b4f3c9c93e0fc23.jpg", + "image_caption": [ + "OCR on a snapping tape Tesseract [30]" + ], + "image_footnote": [], + "bbox": [ + 568, + 377, + 725, + 454 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a3d46eedb599bc1167cd8f6a2f37ae73379ac2a6e2344e9caaa8c4d9f94724e9.jpg", + "image_caption": [ + "Segmenting a dart in midair Segment Anything [32]" + ], + "image_footnote": [], + "bbox": [ + 728, + 377, + 885, + 454 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "with the same sensor.", + "bbox": [ + 76, + 598, + 212, + 611 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Implications: extreme, bandwidth-efficient vision. Generalized event cameras support high-speed, high-quality image reconstruction, but at low bandwidths quintessential of current event cameras. For example, Fig. 1 (middle, bottom) shows reconstructions at 3025 FPS that have an effective readout of a 30 FPS frame-based camera. Further, our methods have strong low-light performance due to the SPAD's single-photon sensitivity. As we show in Fig. 1 (bottom), preserving scene intensity facilitates plug-and-play inference in challenging scenarios, with state-of-the-art vision algorithms. Critically, this does not require retraining vision models or curating dedicated datasets, which is a significant challenge for unconventional imagers. This plug-and-play capability is vital to realizing universal event vision that retains the benefits of current event cameras.", + "bbox": [ + 75, + 619, + 470, + 832 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Scope. We consider full-stack event perception: we conceptualize a novel space of event cameras, provide relevant single-photon algorithms, analyze their imaging capabilities and rate-distortion trade-offs, and show on-chip feasibility. We demon-", + "bbox": [ + 75, + 839, + 472, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "strate imaging capabilities in Secs. 5.1 and 5.2 using the SwissS-PAD2 array [62], and show viable implementations of our algorithms for UltraPhase [4], a recent single-photon compute platform. All of these are critical to unlocking the promise of event cameras. However, our objective is not to develop an integrated system that incorporates all these components; this paper merely takes the first steps toward that goal.", + "bbox": [ + 496, + 597, + 893, + 704 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 722, + 635, + 738 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Event camera designs. Perhaps most widespread is the DVS event camera [34], where each pixel generates an event in response to measured changes in (log) intensity. The DAVIS event camera [6, 9] couples DVS pixels with conventional CMOS pixels, providing access to image frames. However, the frames lack the dynamic range of DVS events. A recent design, CelexV [23], provides log-intensity frames using an external trigger. ATIS [44], a less prevalent design, features asynchronous intensity events, but its sophisticated circuitry reduces pixel fill factor. The above designs are based on analog processing; we instead", + "bbox": [ + 496, + 750, + 893, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "25008", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "design event cameras on digital photon detections.", + "bbox": [ + 76, + 90, + 392, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Intensity imaging with event cameras. Several approaches have been explored to obtain images from events, including Poisson solvers [5], manifold regularization [39], assuming knowledge of camera motion [13, 31] or optical flow [67], and learning-based methods [43, 46, 51, 58, 69]. However, because events often lack sufficient scene information, they are often supplemented by conventional frames [10, 41, 50, 53], either from sensors such as DAVIS or using a multi-camera setup. Fusing events and frames presents challenges due to potential spatiotemporal misalignment and discrepancies in imaging modalities. Even when these challenges are overcome, we show that fusion methods produce lower fidelity than our proposed generalized event cameras.", + "bbox": [ + 75, + 111, + 468, + 306 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Passive single-photon imaging. In the past few years, SPADs have found compelling passive imaging applications; this includes high-dynamic range imaging [26, 27, 35, 40], motion deblurring [28, 36, 37, 52], high-speed tracking [17], and ultra wide-band videography [65]. The fine granularity of passive single-photon acquisition makes it possible to emulate a diverse set of imaging modalities [60], including event cameras, via post-capture processing. In this work, we go beyond emulating existing event cameras and design alternate event cameras that preserve high-fidelity intensity information.", + "bbox": [ + 75, + 313, + 470, + 465 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. What is an Event Camera?", + "text_level": 1, + "bbox": [ + 76, + 479, + 313, + 496 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The defining characteristic of event cameras is that they transmit information selectively, in response to changes in scene content. This selectivity allows event cameras to encode scene information at high time resolutions required to capture scene dynamics, without proportionately high bandwidth. This is in contrast to frame-based cameras, where readout occurs at fixed intervals.", + "bbox": [ + 75, + 507, + 468, + 597 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We characterize event cameras in terms of two axes: what the camera transmits and when it transmits. As a concrete example, consider existing event cameras. They trigger events (\"when to transmit\") based on a fixed threshold:", + "bbox": [ + 75, + 598, + 468, + 657 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left| \\Phi (\\mathbf {x}, t) - \\Phi_ {\\text {r e f}} (\\mathbf {x}) \\right| \\geq \\tau , \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 666, + 468, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\Phi (\\mathbf{x},t)$ is a flux estimate at pixel $\\mathbf{x}$ and time $t$ , and $\\tau$ is the threshold. $\\Phi_{\\mathrm{ref}}(\\mathbf{x})$ is a previously-recorded reference, set to $\\Phi (\\mathbf{x},t)$ whenever an event is triggered. Each event consists of a packet", + "bbox": [ + 75, + 690, + 468, + 750 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n(\\mathbf {x}, t, \\operatorname {s i g n} (\\Phi (\\mathbf {x}, t) - \\Phi_ {\\mathrm {r e f}} (\\mathbf {x}))) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 750, + 468, + 766 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "that encodes the polarity of the change (\"what to transmit\").", + "bbox": [ + 75, + 771, + 450, + 786 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Event polarities, although adequate for some applications, do not retain sufficient information to support a general set of computer vision tasks. A stream of event polarities is an extremely lossy representation. Notably, it only defines scene intensity up to an initial unknown reference value, and it does", + "bbox": [ + 75, + 787, + 468, + 862 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3f686bfaec24cde85e96edda5ffa8a5848dd0072a989af09283ac91240a860b9.jpg", + "image_caption": [ + "(a) 1-bit events" + ], + "image_footnote": [], + "bbox": [ + 501, + 89, + 627, + 150 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/acd36c968fa6beb6f6c7691002d9cb411aa11b67edee335acad1777337357991.jpg", + "image_caption": [ + "(b) Flux levels", + "Figure 2. Altering \"what to transmit.\" (a) We sum the events generated by a jack-in-the-box toy as it springs up. This sum gives a lossy encoding of brightness changes in dynamic regions. (b) Transmitting levels instead of changes helps recover details in static regions. (c) Adaptive exposures, which accumulate flux between consecutive events, provide substantial noise reduction." + ], + "image_footnote": [], + "bbox": [ + 630, + 90, + 756, + 148 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/25bb2f6f106dc619d84d367dd05d04ce16e75cff28d1412ad034acafb0711644.jpg", + "image_caption": [ + "(c) Adaptive exposures" + ], + "image_footnote": [], + "bbox": [ + 761, + 90, + 885, + 148 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "not encode any information in regions not producing events, i.e., regions with little or no motion.", + "bbox": [ + 496, + 261, + 890, + 290 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our key observation is that existing event cameras represent just one operating point in a broader space of generalized event cameras, which is defined by two axes: \"what to transmit\" and \"when to transmit.\" By considering alternate points in this space, we can design event cameras that preserve high-fidelity scene intensity. This enables plug-and-play inference with a host of algorithms developed by the mainstream vision community.", + "bbox": [ + 496, + 292, + 892, + 397 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We begin with a conceptual exploration of this generalized space, before describing its physical implementation.", + "bbox": [ + 496, + 398, + 890, + 428 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generalizing \"what to transmit.\" As a first step, we can modify the event camera such that it transmits $n$ -bit values instead of one-bit change polarities. When an event is triggered, we send the current value of $\\Phi(\\mathbf{x}, t)$ ; if a pixel triggers no events, we transmit $\\Phi$ during the final readout. As we show in Fig. 2(b), this simple change allows us to recover scene intensity, even in static regions. It is important to note that, while the transmitted quantity differs from conventional events, we retain the defining feature of an event camera: selective transmission based on scene dynamics (where we transmit according to Eq. (1)). Thus, the readout remains decoupled from the time resolution.", + "bbox": [ + 496, + 431, + 892, + 598 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "If $\\Phi (\\mathbf{x},t)$ were a perfect estimate of scene intensity, then the changes thus far would suffice. However, $\\Phi$ is fundamentally noisy: to capture high-speed changes, $\\Phi$ must encompass a shorter duration, which leads to higher noise. This is a manifestation of the classical noise-blur tradeoff.", + "bbox": [ + 496, + 599, + 890, + 674 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address this problem, we introduce the abstraction of an integrator (or $\\Sigma$ ), that defines how we accumulate incident flux and, in turn, what we transmit. Ideally, we want the integrator to adapt to scene dynamics, i.e., accumulate over longer durations when there is less motion, and vice versa. We observe that event generation, which is based on scene dynamics, can be used to formulate an adaptive integrator. Specifically, we propose an integrator $\\Sigma_{\\mathrm{cum}}$ that computes the cumulative flux since the last event:", + "bbox": [ + 496, + 675, + 890, + 808 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Sigma_ {\\mathrm {c u m l}} (\\mathbf {x}, t) = \\int_ {T _ {0}} ^ {t} \\Phi (\\mathbf {x}, s) d s, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 808, + 890, + 840 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $T_{0}$ is the time of the last event. When an event is triggered at time $T_{1}$ , we communicate the value $^{2}$ of $\\Sigma_{\\mathrm{cum}}(\\mathbf{x}, T_{1})$ , which", + "bbox": [ + 496, + 844, + 890, + 875 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1Event cameras such as the DVS [9] measure a temporally low-pass filtered estimate of log-flux; we absorb this into $\\Phi (\\mathbf{x},t)$ for brevity.", + "bbox": [ + 75, + 875, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "2We can either transmit values of $\\Sigma_{\\mathrm{cum}}$ or differences (changes) to $\\Sigma_{\\mathrm{cum}}$ ;", + "bbox": [ + 514, + 886, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "25009", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/5e811514360da0238d51bba9e75f0ac508ceb4d98dc560a8f61f84c23abb8201.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Event cameraIntegrator (Σ)Change detector (Δ)Event packetsMin. latencyIntensity info.?Low-light perf.
Existing (DVS [34])logarithmiccomparatorbinary10-6to 10-5sXpoor
Sec. 4.1adaptive exposureBayesian change detector [2]scalar10-5s✓good
Sec. 4.2adaptive exposurevariance-aware differencespatches10-4s✓good
Sec. 4.3coded exposureBinomial confidence intervalvector10-3s✓good
", + "bbox": [ + 76, + 89, + 893, + 171 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. Summary of generalized event cameras. Our designs integrate photon detections $(\\Sigma)$ and detect scene-content changes $(\\Delta)$ in distinct ways. We compare our designs to existing DVS event cameras based on their event streams, latencies, and intensity-preserving nature. While providing a direct power comparison to DVS is difficult, we compare the power characteristics among our designs in Sec. 5.4.", + "bbox": [ + 75, + 174, + 893, + 219 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "we interpret as the intensity throughout $[T_0, T_1]$ . This approach yields a piece-wise constant time series, with segments delimited by events. Adaptive exposures significantly reduce noise, as we show in Fig. 2 (c).", + "bbox": [ + 75, + 229, + 470, + 291 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Generalizing \"when to transmit.\" The success of the adaptive integrator crucially depends on the reliability of events; for example, triggering false events in static regions causes unnecessary noise. We refer to the event-generation procedure as the change detector, denoted by $\\Delta$ . Current event cameras detect changes by applying a fixed threshold to measured intensity differences (Eq. (1)). This method has two key limitations: it only considers the value of $\\Phi$ at pixel location $\\mathbf{x}$ and time $t$ , and it is not attuned to the stochasticity in $\\Phi$ .", + "bbox": [ + 75, + 292, + 470, + 429 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We design more robust change detectors that (1) leverage enhanced spatiotemporal contexts, and (2) incorporate noise awareness, either explicitly by tuning thresholds, or implicitly by modulating the detector's behavior. Specifically, we improve reliability by using temporal forecasters (Sec. 4.1), by leveraging correlated changes in patches (Sec. 4.2), or by exploiting integrator statistics (Sec. 4.3).", + "bbox": [ + 75, + 429, + 470, + 535 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Realizing generalized event cameras. The critical detail remaining is how we implement our proposed designs in practice. We need direct access to flux estimates at a high time resolution. Conventional high-speed cameras can provide such access, however, they incur substantial per-frame read noise $(\\sim 20 - 40e^{-}$ [25]) that grows with frame rate [8].", + "bbox": [ + 75, + 537, + 470, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We turn to an emerging class of single-photon sensors, single-photon avalanche diodes (SPADs [48]), that has witnessed dramatic improvements in device practicality and key sensor characteristics (e.g., array sizes and fill factors) in recent years [38, 62]. SPADs can operate at extremely high speeds ( $\\sim 100\\mathrm{kHz}$ ) without incurring per-frame read noise. Each $\\Phi(\\mathbf{x},t)$ measured by a SPAD is limited only by the fundamental stochasticity of photon arrivals (shot noise). This allows SPADs to provide high timing resolution without a substantial noise penalty. In the next section, we describe the image formation model of SPADs and provide single-photon implementations of our designs.", + "bbox": [ + 75, + 628, + 470, + 796 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Single-Photon Generalized Event Cameras", + "text_level": 1, + "bbox": [ + 76, + 806, + 441, + 825 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A SPAD array can operate as a high-speed photon detector, producing binary frames as output. Each binary value indicates whether at least one photon was detected during an expo", + "bbox": [ + 75, + 833, + 472, + 878 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sure. The SPAD output response, $\\Phi (\\mathbf{x},t)$ , can be modeled as a Bernoulli random variable, with", + "bbox": [ + 500, + 229, + 890, + 258 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP (\\Phi (\\mathbf {x}, t) = 1) = 1 - e ^ {- N (\\mathbf {x}, t)}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 588, + 265, + 890, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $N(\\mathbf{x}, t)$ is the average number of photo-electrons during an exposure, including any spurious detections. The inherently digital SPAD response allows us to compute software-level transformations on the signal $\\Phi(\\mathbf{x}, t)$ , including operations that may be challenging to realize via analog processing. These transformations can be readily reconfigured, which permits a spectrum of event camera designs, not just one particular choice. However, there is one consideration: our designs should be lightweight and computable on chip. As we show in Sec. 5.4, this is vital to implementing generalized event cameras without the practical costs associated with reading off raw SPAD outputs.", + "bbox": [ + 496, + 291, + 893, + 458 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We now describe a set of SPAD-based event cameras (summarized in Tab. 1), beginning with the adaptive exposure method from the previous section.", + "bbox": [ + 496, + 459, + 893, + 505 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Adaptive-exposure event camera. We obtain a SPAD implementation of the adaptive exposure described in Eq. (3) by replacing the integral with a sum over photons:", + "bbox": [ + 496, + 513, + 893, + 559 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Sigma_ {\\mathrm {c u m l}} (\\mathbf {x}, t) = \\sum_ {s = T _ {0}} ^ {t} \\Phi (\\mathbf {x}, s). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 565, + 890, + 606 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To generate events, we can use a threshold-based change detector (Eq. (1)). Differences between individual binary values are not sufficiently informative; therefore, we apply Eq. (1) to an exponential moving average (EMA) computed on $\\Phi$ . We call this an \"adaptive-EMA\" event camera.", + "bbox": [ + 496, + 613, + 893, + 689 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Bayesian Change Detector", + "text_level": 1, + "bbox": [ + 500, + 705, + 738, + 722 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A fixed-threshold change detector such as Eq. (1) does not account for the SPAD's image formation model; it uses the same threshold irrespective of the underlying variance in photon detections. As a result, such a detector may fail to detect changes in low-contrast regions without producing a large number of false-positive detections (see Fig. 3 (left)).", + "bbox": [ + 496, + 731, + 890, + 823 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we consider a Bayesian change detector, BOCPD [1], that is tailored to the Bernoulli statistics of photon detections. BOCPD uses a series of forecasters to estimate the likelihood of an abrupt change. At each time step, a new forecaster $\\nu_{t}$ is initialized as a recurrence of previous forecasters,", + "bbox": [ + 496, + 825, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "we treat this as an implementation detail here.", + "bbox": [ + 76, + 887, + 307, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "25010", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b69d56f366f855315de174bff0eb73de6aa8b6b8681d4f513b537ae3fefee08e.jpg", + "image_caption": [ + "Figure 3. Bayesian- vs. EMA-based change detection. (left) A fixed-threshold change detector (used in adaptive-EMA) makes it difficult to segment low-contrast changes. (center) The Bayesian formulation attunes to the stochasticity in incident flux and can detect fine-grained changes such as the corners of the hole saw bit; (right) as a result, the integrator captures the rotational dynamics." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 462, + 172 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and existing forecasters are updated:", + "bbox": [ + 76, + 270, + 307, + 284 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\nu_ {t} = (1 - \\gamma) \\sum_ {s = 1} ^ {t - 1} l _ {s} \\nu_ {s}, \\quad \\nu_ {s} \\leftarrow \\gamma l _ {s} \\nu_ {s} \\forall s < t, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 291, + 468, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\gamma \\in [0,1]$ is the sensitivity of the change detector, with larger $\\gamma$ resulting in more frequent detections. $l_{s}$ is the predictive likelihood of each forecaster, which we compute by tracking two values per forecaster, $\\alpha_{s}$ and $\\beta_{s}$ , that correspond to the parameters of a Beta prior. For a new forecaster, these values are initialized to 1 each, reflecting a uniform prior. Existing $(\\alpha_{s},\\beta_{s}),\\forall s < t,$ are updated as", + "bbox": [ + 75, + 339, + 468, + 445 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {s} \\leftarrow \\alpha_ {s} + \\Phi (\\mathbf {x}, t), \\quad \\beta_ {s} \\leftarrow \\beta_ {s} + 1 - \\Phi (\\mathbf {x}, t). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 452, + 468, + 467 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$l_{s}$ is given by $\\alpha_{s} / (\\alpha_{s} + \\beta_{s})$ if $\\Phi (\\mathbf{x},t) = 1$ , and $\\beta_{s} / (\\alpha_{s} + \\beta_{s})$ otherwise. An event is triggered if the highest-value forecaster does not correspond to $T_{0}$ , the timestamp of the last event; mathematically, if $\\arg \\max_t\\nu_t\\neq T_0$ .", + "bbox": [ + 75, + 474, + 468, + 536 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To make BOCPD viable in memory-constrained scenarios, we apply extreme pruning by retaining only the three highest-value forecasters [64]. We also incorporate restarts, deleting previous forecasters when a change is detected [2].", + "bbox": [ + 75, + 536, + 468, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Compared to an EMA-based change detector, the Bayesian approach more reliably triggers events in response to scene changes while better filtering out stochastic variations caused by photon noise—which we show in Fig. 3.", + "bbox": [ + 75, + 595, + 468, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Spatiotemporal Chunk Events", + "text_level": 1, + "bbox": [ + 76, + 665, + 346, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sec. 4.1 leverages an expanded temporal context for change detection; however, it treats each pixel independently and does not exploit spatial information. In this section, we devise an event camera with enhanced spatial context that operates on small patches, e.g., of $4 \\times 4$ pixels. It is difficult to derive efficient Bayesian change detectors for multivariate time series; thus, we adopt a model-free approach that does not explicitly parameterize the patch distribution. To afford computational breathing room for more expensive patch-wise operations, we employ temporal chunking. That is, we average $\\Phi(\\mathbf{x}, t)$ over a small number of binary frames (e.g., 32 binary frames) instead of operating on individual binary frames; generally, this averaging does not induce perceptible blur.", + "bbox": [ + 75, + 688, + 468, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let vector $\\Phi_{\\mathrm{chunk}}(\\mathbf{y},t)$ represent the chunk-wise average of", + "bbox": [ + 94, + 885, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/884fb01b67138d44d512fcc08e7c7b6f5ee158be9ba9025679a5746b3073acba.jpg", + "image_caption": [ + "Figure 4. Spatiotemporal chunk events. We evaluate the difference between the current chunk and a stored reference in a learned linear-feature space. Unlike the L2 norm, which is permutation-invariant, the feature-space norm is sensitive to spatial structure. Randomly shuffling the pixel values reduces the transform-domain norm (the shuffled patch has a more \"noise-like\" structure)." + ], + "image_footnote": [], + "bbox": [ + 506, + 88, + 883, + 200 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "photon detections at patch location $\\mathbf{y}$ . Let vector $\\Sigma_{\\mathrm{patch}}(\\mathbf{y},t)$ be an integrator representing the cumulative mean since the last event, but excluding $\\Phi_{\\mathrm{chunk}}$ . We want to estimate whether $\\Phi_{\\mathrm{chunk}}$ belongs to the same distribution as $\\Sigma_{\\mathrm{patch}}$ . We do so with a lightweight approach, that computes the distance between $\\Phi_{\\mathrm{chunk}}$ and $\\Sigma_{\\mathrm{patch}}$ in the linear feature space of matrix $\\mathbf{P}$ . As we show in Fig. 4, linear features allow us to capture spatial structure within a patch. Geometrically, $\\mathbf{P}$ induces a hyperellipsoidal decision boundary, in contrast to the spherical boundary of the L2 norm.", + "bbox": [ + 496, + 300, + 892, + 449 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This method generates an event whenever", + "bbox": [ + 519, + 453, + 782, + 467 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\| \\mathbf {P} \\left(\\tilde {\\Phi} _ {\\text {c h u n k}} (\\mathbf {y}, t) - \\tilde {\\Sigma} _ {\\text {p a t c h}} (\\mathbf {y}, t)\\right) \\right\\| _ {2} \\geq \\tau , \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 473, + 890, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tau$ is the threshold. When there is no event, we extend the cumulative mean to include the current chunk. Before computing linear features, we normalize $\\Phi_{\\mathrm{chunk}}$ and $\\Sigma_{\\mathrm{patch}}$ element-wise according to the estimated variance in $\\Phi_{\\mathrm{chunk}} - \\Sigma_{\\mathrm{patch}}$ ; we annotate the normalized versions with a tilde. We estimate the variance based on the fact that, in a static patch, the elements of $\\Phi_{\\mathrm{chunk}}$ and $\\Sigma_{\\mathrm{patch}}$ are independent binomial random variables.", + "bbox": [ + 496, + 498, + 892, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We train the matrix $\\mathbf{P}$ on simulated SPAD data, generated from interpolated high-speed video. We apply backpropagation through time to minimize the MSE error of the transmitted patch values. To address the non-differentiability arising from the threshold, we employ surrogate gradients. Please see the supplementary material for complete details of this method.", + "bbox": [ + 496, + 604, + 890, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Coded-Exposure Events", + "text_level": 1, + "bbox": [ + 500, + 709, + 720, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we design a generalized event camera by applying change detection to coded exposures [22, 45, 47], which capture temporal variations by multiplexing photon detections over an integration window. This is interesting in two aspects. First, we are designing event streams based on an modality not typically associated with event cameras. Second, we show that high-speed information can be obtained even when the change detector operates at a coarser time granularity. Coded-exposure events provide somewhat lower fidelity than our designs in Secs. 4.1 and 4.2, but are more compute- and power-efficient, owing to less frequent execution of the change detector.", + "bbox": [ + 496, + 734, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "25011", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At each pixel, we multiplex a temporal chunk of $T_{\\mathrm{code}}$ ( $\\sim 256 - 512$ ) binary values with a set of $J$ ( $\\sim 2 - 6$ ) codes $C^j(\\mathbf{x}, t) \\forall 1 \\leq j \\leq J$ , producing $J$ coded exposures", + "bbox": [ + 76, + 90, + 472, + 137 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\Sigma_ {\\text {c o d e d}} ^ {j} (\\mathbf {x}, t) = \\sum_ {s = t - T _ {\\text {c o d e}}} ^ {t} \\Phi (\\mathbf {x}, s) C ^ {j} (\\mathbf {x}, s). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 143, + 470, + 185 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The codes $C^j$ are chosen to be random, mutually orthogonal binary masks, each containing $T_{\\mathrm{code}} / \\max (2,J)$ ones [60].", + "bbox": [ + 75, + 193, + 468, + 223 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We exploit the statistics of coded exposures to derive a change detector. Observe that in static regions, $\\Sigma_{\\mathrm{coded}}^{j}(\\mathbf{x},t)$ are independent and identically distributed (iid) binomial random variables. Thus, we can expect them to lie within a binomial confidence interval of one another. If not, we assume the pixel is dynamic and generate an event. We trigger an event if $\\Sigma_{\\mathrm{coded}}^{j}\\notin \\operatorname {conf}(n,\\hat{p})$ for any $j$ . Here, \"conf\" refers to a binomial confidence interval (e.g., Wilson's score), $n = T_{\\mathrm{code}} / J$ draws, and $\\hat{p} = \\sum_{s}\\Phi (\\mathbf{x},s) / T_{\\mathrm{code}}$ is the empirical success probability.", + "bbox": [ + 75, + 223, + 470, + 358 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "If a pixel is static, we store the sum of the $J$ coded exposures, which is a long exposure, denoted by $\\Sigma_{\\mathrm{long}}$ . If the pixel remains static across more than one temporal chunk, we extend $\\Sigma_{\\mathrm{long}}$ to include the entire duration. Whereas, if the pixel is dynamic, we transmit $\\{\\Sigma_{\\mathrm{coded}}^j\\}$ , as well as any previous static intensity encoded in $\\Sigma_{\\mathrm{long}}$ . Downstream, we can apply coded-exposure restoration techniques [54, 63, 66] to recover intensity frames from the coded measurements.", + "bbox": [ + 75, + 359, + 468, + 479 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experimental Results", + "text_level": 1, + "bbox": [ + 76, + 492, + 272, + 508 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We demonstrate the capabilities of generalized event cameras using a SwissSPAD2 array [62] with resolution $512 \\times 256$ , which we use to capture one-bit frames at $96.8\\mathrm{kHz}$ . We show the feasibility of our designs on UltraPhase [4], a recent single-photon computational platform (Sec. 5.4).", + "bbox": [ + 75, + 518, + 470, + 594 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Refinement model. For each of our event cameras, we train a refinement model that mitigates artifacts arising from the asynchronous nature of events. This model takes a periodic frame-based sampling of integrator values and outputs a video reconstruction. The sampling rate is configurable; in practice, we set it $\\sim 16 - 64\\times$ lower than the SPAD rate. We use a densely-connected residual architecture [63], trained on data generated by simulating photon detections on temporally interpolated [24] high-speed videos from the XVFI dataset [56]. See the supplement for training details.", + "bbox": [ + 75, + 595, + 470, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Extreme Bandwidth-Efficient Videography", + "text_level": 1, + "bbox": [ + 76, + 756, + 441, + 773 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "High-speed videography. In Fig. 5, we capture the dynamics of a deformable ball (a \"stress ball\") using a SPAD, a high-speed camera (Photron Infinicam) operated at 500 FPS, and a commercial event camera (Prophesee EVK4). The high-speed camera suffers from low SNR due to read noise, which manifests as prominent artifacts after on-camera compression. Meanwhile, conventional events captured by Prophesee, when processed by \"intensity-from-events\" methods such as E2VID+ [57] fail to", + "bbox": [ + 75, + 779, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/776b147388e4d73cbd0b218575acc3eaae318467ed507cf7feccbcc9c6245c19.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 90, + 633, + 181 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cc9ba3613e75143c8f80a2b01c65083eb77f6fca2198d5f2ab11ddefe2224e00.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 90, + 751, + 181 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/407a689e1086fe22a1e8ca633d5ee20a35b3b71c3e9653bd87edd3ccffc54290.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 90, + 872, + 181 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7f9824a62aaf42ed57bee61aa5475c8aa12c6203db21fee904cd78988a01e42d.jpg", + "image_caption": [ + "Spatiotemporal chunk (Sec 4.2, 431 bps/pixel, rendered at 3025 FPS)" + ], + "image_footnote": [], + "bbox": [ + 511, + 184, + 633, + 311 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4d054021013ee51bd8df9dc08f032793da90e2b8c7e8843282d7fac7c0d4517d.jpg", + "image_caption": [ + "Figure 5. High-speed videography of a stress ball hurled at a coffee mug. (top row) This indoor scene is challenging for existing imaging systems, including: high-speed cameras (SNR-related artifacts), event cameras (poor restoration quality), and even hybrid event + frame techniques (reconstruction artifacts). (bottom rows) In contrast, our generalized event cameras capture the stress ball's extensive deformations with high fidelity and an efficient readout." + ], + "image_footnote": [], + "bbox": [ + 633, + 184, + 754, + 311 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/006e3535b29621c81543a103cee50806de3f4ad1359892539877b12bc3006488.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 184, + 877, + 311 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "recover intensities reliably, especially in static regions. We also evaluate EDI [42], a hybrid event-frame method. We consider an idealized variant that operates on SPAD events (obtained via EMA thresholding), which gives perfect event-frame alignment and a precisely known event-generation model. We refine the outputs of EDI using the same model as for our methods. We refer to this idealized, refined version of EDI as \"EDI++. While EDI++ recovers more detail than other baselines, there are considerable artifacts in its outputs.", + "bbox": [ + 498, + 439, + 893, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our method achieves high-quality reconstructions at 3025 FPS (96800/32) that faithfully capture non-rigid deformations, with only 431 bits per second per pixel (bps/pixel) readout, which is a $227\\times$ compression (96800/431) of the raw SPAD capture. Viewed differently, for a 1 MPixel array, we would obtain a bitrate of 431 Mbps, implying that we can read off these 3025 FPS reconstructions over USB 2.0 (which supports transfer speeds of up to 480 Mbps).", + "bbox": [ + 498, + 577, + 893, + 698 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Event imaging in low light. Fig. 6 compares the low-light performance of frame-based, event-based, and a generalized event camera on an urban night-time scene at 7 lux (lux measured at the sensor). For frame-based cameras, a short exposure that preserves motion may be too noisy, while a long exposure can be severely blurred. The Propheesee's performance deteriorates in low light, resulting in blurred temporal gradients. EDI++, benefiting from the idealized SPAD-based implementation, can image this scene, but finer details like the motorcyclist are lost. Our generalized event cameras, on the other hand, provide reconstructions with minimal noise, blur, or artifacts—while retaining the bandwidth efficiency of event-based systems. The compression here is $307\\times$ with respect to raw SPAD outputs.", + "bbox": [ + 498, + 704, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "25012", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ae511bffccd971686007bd165f910cd3fcaced356b4d8dd4f75592ed7fee53f6.jpg", + "image_caption": [ + "Frame-based cameras" + ], + "image_footnote": [], + "bbox": [ + 99, + 102, + 248, + 162 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/29e24dbcbe0e27f115310fa80b4424dde132e011c8dd2271aabd1b9cb03914bb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 251, + 102, + 400, + 162 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a9c7ebc71b3f126014fa96329ad7621b888f565ed3cb845b516a7fa8f3f1c92c.jpg", + "image_caption": [ + "1 ms short exposure (120 bps/pixel at 10 FPS)" + ], + "image_footnote": [], + "bbox": [ + 101, + 164, + 248, + 223 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c4656beb468802d059a0573bf0fa988a272d561e7b8b4dcf933e260259ffb09b.jpg", + "image_caption": [ + "80 ms long exposure (120 bps/pixel at 10 FPS)" + ], + "image_footnote": [], + "bbox": [ + 251, + 162, + 400, + 223 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1245c45e0b42c4ffac7b5e892b4d0304db5c9d80b8861cbc315011d6e3d3d978.jpg", + "image_caption": [ + "Events only" + ], + "image_footnote": [], + "bbox": [ + 408, + 102, + 555, + 162 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/efc1e290b505e065d8166b8f6e6daf8dec466c4669205991e2393f347737630d.jpg", + "image_caption": [ + "E2VID+ (121 bps/pixel)" + ], + "image_footnote": [], + "bbox": [ + 408, + 164, + 555, + 223 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a6ea1fb6959e0ed56bca57eebc431c143d92a3ccc5b69e4243fecc7ce7dbb59a.jpg", + "image_caption": [ + "Event + frame imaging" + ], + "image_footnote": [], + "bbox": [ + 563, + 102, + 712, + 162 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/79748d1ed6a0e0dfb0ee7f4725e8a0f19d3c5ffdd7337754f7f21167e88b6f8e.jpg", + "image_caption": [ + "EDI++ (315 bps/pixel)" + ], + "image_footnote": [], + "bbox": [ + 565, + 162, + 710, + 223 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7c43d72feafcf5e7035d34c692dce0680bcbec997768dbfda9745d66ffa4a137.jpg", + "image_caption": [ + "Generalized events" + ], + "image_footnote": [], + "bbox": [ + 720, + 102, + 870, + 162 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/07b3f3bd9413bd0085d9ce0c60501e256d991927bff3747606ffb07c24f498ed.jpg", + "image_caption": [ + "Adaptive-Bayesian (Sec 4.1, 315 bps/pixel)" + ], + "image_footnote": [], + "bbox": [ + 720, + 162, + 867, + 223 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ce6116641a15931a23432b71376aaeaea42b5a8d5b4fd89d912b09914cb9a5c2.jpg", + "image_caption": [ + "Figure 6. Event imaging in urban nighttime (7 lux, sensor side). (left to right) Low-light conditions necessitate long exposures in frame-based cameras, resulting in unwanted motion blur. The Propheesee EVK4 suffers from severe degradation in low light, causing E2VID+ to fail. Running EDI++ on perfectly aligned SPAD-frames and -events improves overall restoration quality but still gives failures on fast-moving objects. Our generalized events recover significantly more detail in low light, as seen in the inset of the motorcyclist.", + "HRNet on E2VID+" + ], + "image_footnote": [], + "bbox": [ + 83, + 324, + 212, + 386 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e2c898c0b021dd00ba8ba03385d806335c27f678adbc5aea7a99ee06a33478c6.jpg", + "image_caption": [ + "Conventional events (Propheesee EVK4, 331 bps/pixel)", + "Arc* corners" + ], + "image_footnote": [], + "bbox": [ + 215, + 324, + 344, + 386 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3c68478644d38c4734ac641e3838138c3fa83b8cf7d7d03b205e7276ef735ad0.jpg", + "image_caption": [ + "E-RAFT flow" + ], + "image_footnote": [], + "bbox": [ + 348, + 324, + 475, + 386 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4da2f47605815447849cd4dce20013bab4102e5d1e6aeeb810dc8ae0a00ab7ee.jpg", + "image_caption": [ + "DETR on E2VID+" + ], + "image_footnote": [], + "bbox": [ + 480, + 324, + 607, + 386 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/29dad1bac0afc2e0edd276e253e8e6445027fd1adde091a2b5ef4b84b7ed3e02.jpg", + "image_caption": [ + "SAM on E2VID+" + ], + "image_footnote": [], + "bbox": [ + 611, + 324, + 738, + 386 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/37bff2fc404531272d5d078da8af8a496b88ada92f4ea3f8ff0678ab6b207d3b.jpg", + "image_caption": [ + "85 ms long exposure", + "120 bps/pixel at 10 FPS" + ], + "image_footnote": [], + "bbox": [ + 750, + 324, + 877, + 386 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1a5b976082f0a1f04b8835bc367145708c26a53df2474342b3bf0f433822e0d6.jpg", + "image_caption": [ + "HRNet pose" + ], + "image_footnote": [], + "bbox": [ + 84, + 415, + 212, + 477 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/635e3252d80a9ca2f675abb4ba90f20df65b0554d27600ffea6df6480e7f675f.jpg", + "image_caption": [ + "Generalized events (Spatiotemporal chunk, Sec 4.2, 520 bps/pixel)", + "Harris corners" + ], + "image_footnote": [], + "bbox": [ + 215, + 415, + 344, + 477 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/026cb2d61211598528c0c565282b3f74cfb6e3550e2278292cda54585f7fe207.jpg", + "image_caption": [ + "RAFT flow" + ], + "image_footnote": [], + "bbox": [ + 348, + 415, + 475, + 477 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/89f626bfab0ac5cdc1323856ca5f0ff544864b1df2ee1d6a4027585957df98f7.jpg", + "image_caption": [ + "DETR detection" + ], + "image_footnote": [], + "bbox": [ + 480, + 415, + 607, + 477 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/06b75834ea6b701472d74d09a713563ea54445a7588d21086efd89c02695ae3f.jpg", + "image_caption": [ + "SAM segmentation" + ], + "image_footnote": [], + "bbox": [ + 611, + 415, + 738, + 477 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/db5db6b7d05d380a522981f4079a4cce168982a9de7be8957cbcd5a5c102dff9.jpg", + "image_caption": [ + "0.3 ms burst exposures", + "15100 bps/pixel", + "Figure 7. Plug-and-play inference on a tennis scene. (top left) Conventional events encode temporal-gradient polarities; this lossy representation limits performance on downstream tasks. (bottom left) Generalized events encode rich scene-intensity information, with a readout comparable existing event cameras. They facilitate high-quality plug-and-play inference, without requiring dedicated algorithms. (right) Generalized event cameras give image quality comparable to burst photography techniques that have a much higher readout rate." + ], + "image_footnote": [], + "bbox": [ + 750, + 415, + 885, + 482 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Plug-and-Play Inference", + "text_level": 1, + "bbox": [ + 76, + 564, + 300, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Generalized event cameras preserve scene intensity, which enables plug-and-play event-based vision. We consider a tennis sequence (of 8196 binary frames) containing a range of object speeds. We evaluate a range of tasks: pose estimation (HRNet [59]), corner detection [18], optical flow (RAFT [61]), object detection (DETR [11]), and segmentation (SAM [32]). We compare against event-based methods applied to Prophesee events; we use $\\mathrm{Arc}^*$ [3] for corner detection and E-RAFT [16] for optical flow. For the remaining tasks, which do not have equivalent event methods, we run HRNet, DETR, and SAM on E2VID+ reconstructions.", + "bbox": [ + 75, + 593, + 470, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As Fig. 7 (top left) shows, traditional events are bandwidth efficient (331 bps/pixel), but do not provide sufficient information for successful inference. Generalized events (bottom left) have a modestly higher readout (520 bps/pixel), but support accurate inference without requiring dedicated algorithms. To provide context for these rates, we compare them against frame-based methods (right). A long exposure (120 bps/pixel) blurs out the ratchet. Burst methods [19] recover a sharp image from a stack of short exposures, but with a large readout of 15100 bps/pixel.", + "bbox": [ + 75, + 763, + 472, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Rate-Distortion Analysis", + "text_level": 1, + "bbox": [ + 500, + 564, + 725, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Each method in Sec. 4 features a sensitivity parameter that controls the sensor readout rate (event rate), which in turn influences image quality. In this subsection, we evaluate the impact of readout on image quality (PSNR) by performing a rate-distortion analysis. For ground truth, we use a set of YouTube-sourced high-speed videos captured by a Phantom Flex4k at 1000 FPS; see the supplement for thumbnails and links. We upsample these videos to the SPAD's frame rate and then simulate 4096 binary frames using the image formation model described in Eq. (4). When computing readout for our methods, we assume that events encode 10-bit values and account for the header bits of each event packet.", + "bbox": [ + 496, + 590, + 893, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As baselines, we consider EDI++, a long exposure, compressive sensing with 8-bucket masks, and burst denoising [19] using 32 short exposures. As Fig. 8 shows, generalized event cameras provide a pronounced 4-8 dB PSNR improvement over baseline methods. Further, our methods can compress the raw SPAD response by around $80\\times$ before a noticeable drop-off in PSNR is observed.", + "bbox": [ + 496, + 776, + 893, + 880 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Among our methods, the spatiotemporal chunk approach of", + "bbox": [ + 519, + 885, + 892, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "25013", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fd6c6fe236e8afa729df60d5beaf108f5aaa1cd28b3144be1b37cc8978f9e5d6.jpg", + "image_caption": [ + "Figure 8. Rate-distortion evaluation. Our techniques feature a tunable parameter that controls the output event rate. Generalized events offer a $4 - 8\\mathrm{dB}$ improvement in PSNR over $\\mathrm{EDI} + +$ (at the same readout), and can compress raw photon data by $80\\times$" + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 459, + 199 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Sec. 4.2 gives the best PSNR, followed by the Bayesian method (Sec. 4.1) and coded-exposure events (Sec. 4.3). That said, all methods are fairly similar in terms of rate-distortion (e.g., all three give comparable results for the scenes in Secs. 5.1 and 5.2). The methods are better distinguished by their practical characteristics. The Bayesian method gives single-photon temporal resolution; however, as we show in Sec. 5.4, it is the most expensive to compute on-chip. The chunk-based method occupies a middle ground in terms of latency and cost. Coded-exposure events have the highest latency—events are generated only every $\\sim 256 - 512$ binary frames—but the lowest on-chip cost. This provides an end user the flexibility to choose from the space of generalized event cameras based on the latency requirements and the compute constraints of the target application.", + "bbox": [ + 75, + 273, + 473, + 486 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. On-Chip Feasibility and Validation", + "text_level": 1, + "bbox": [ + 76, + 497, + 382, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A critical limitation of single-photon sensors is the exorbitant bandwidth and power costs involved in reading off raw photon detections. However, the lightweight nature of our event camera designs allows us to sidestep this limitation by performing computations on-chip. We demonstrate that our methods are feasible on UltraPhase [4], a SPAD compute platform. UltraPhase consists of $3 \\times 6$ compute cores, each of which is associated with $4 \\times 4$ pixels.", + "bbox": [ + 75, + 521, + 468, + 642 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We implement our methods for UltraPhase using custom assembly code. Some methods require minor modifications due to instruction-set limitations; see the supplement for details. We process 2500 SPAD frames from the tennis sequence used in Sec. 5.2, cropped to the UltraPhase array size of $12 \\times 24$ pixels. We determine the number of cycles required to execute the assembly code and estimate the chip's power consumption and readout bandwidth.", + "bbox": [ + 75, + 643, + 470, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "All our proposed methods run comfortably within the chip's compute budget of 4202 cycles per binary frame and its memory limit of 4 Kbit per core. As seen in Fig. 9, compared to raw photon-detection readout, our techniques reduce both bandwidth and power costs by over two orders of magnitude. The coded-exposure method is particularly efficient; on most binary frames, it only requires multiplying a binary code with incident photon detections. Our proof-of-concept evaluation may pave the way for future near-sensor implementations of generalized event", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c9725d942b2555ce16b8047a70f5fe65a0e5954f75454606192abb0289916fc5.jpg", + "image_caption": [ + "Figure 9. On-chip compatibility. We validate the feasibility of our approach on UltraPhase [4], a computational SPAD imager. Compared to reading out raw photon data, all of our approaches give marked reductions in both bandwidth (left) and power (right)." + ], + "image_footnote": [], + "bbox": [ + 511, + 90, + 877, + 185 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c57aa0525b99732632c97a11a8593ca21560f0f2f04143de57f04e9e16a9fb5c.jpg", + "image_caption": [ + "Rigid dynamics (200x compression)", + "Figure 10. Limitations and failure modes, (left) Our reconstructions (yellow inset) on dynamic scenes with rigid objects can be inferior to burst photography (green inset). (center) Modulated light sources, such as this phone screen, can trigger a deluge of events (change points shown in the inset). (right) Rapid camera motion can result in an event rate divergent from scene dynamics." + ], + "image_footnote": [], + "bbox": [ + 506, + 256, + 629, + 316 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/37e4fe3c13d231503f0a2500c081e897e30bd89ef6b87af2b8d9166a8d42479a.jpg", + "image_caption": [ + "Phone screen (11x compression)" + ], + "image_footnote": [], + "bbox": [ + 632, + 257, + 756, + 316 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e78523f46441ba8ae4f55eff71c3b62698bf81b86c0f70486a601ce9108e1be9.jpg", + "image_caption": [ + "Camera motion (130x compression)" + ], + "image_footnote": [], + "bbox": [ + 759, + 257, + 882, + 316 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "cameras, which with advances in chip-to-chip communication, could involve a dedicated \"photon processing unit\", similar to a camera image signal processor (ISP).", + "bbox": [ + 498, + 436, + 893, + 482 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Limitations and Discussion", + "text_level": 1, + "bbox": [ + 500, + 494, + 736, + 510 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Generalized events push the frontiers of event-based imaging; however, some scenarios lead to sub-optimal performance. As seen in Fig. 10 (left), if the scene dynamics is entirely comprised of rigid motion, burst photography [36] gives better image quality, albeit with much higher readout. (middle) Similar to current event cameras, modulated light sources trigger unwanted events that reduce bandwidth savings. However, it may be possible to ignore some of these events, perhaps by modeling the lighting variations [55].", + "bbox": [ + 498, + 520, + 893, + 656 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ego-motion events. Camera motion can trigger events in static regions, although our methods still yield substantial compression $(130 \\times$ over SPAD outputs, Fig. 10 right). We analyze the impact of ego-motion on bandwidth savings further in the supplement. However, single-photon cameras can emulate sensor motion by integrating flux along alternate spatiotemporal trajectories [60]. We can imagine a generalized event camera that is \"ego-motion compensated,\" by computing events along a suitable trajectory.", + "bbox": [ + 498, + 660, + 893, + 781 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Photon-stream compression. SPADs generate a torrent of data—e.g., 12.5 GBps for a MPixel array at $100\\mathrm{kHz}$ —that can easily overwhelm data interfaces. Generalized event cameras reduce readout by around two orders of magnitude, by decoupling readout from the SPAD's frame rate and instead basing it on scene dynamics. This could pave the way for practical, high-resolution single-photon sensors.", + "bbox": [ + 498, + 785, + 893, + 890 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "25014", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 169, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ryan Prescott Adams and David JC MacKay. Bayesian online changepoint detection. arXiv preprint arXiv:0710.3742, 2007. 4", + "[2] Reda Alami, Odalric Maillard, and Raphael Feraud. Restarted bayesian online change-point detector achieves optimal detection delay. In International conference on machine learning, pages 211-221. PMLR, 2020. 4, 5", + "[3] Ignacio Alzugaray and Margarita Chli. Asynchronous corner detection and tracking for event cameras in real time. IEEE Robotics and Automation Letters, 3(4):3177-3184, 2018. 7", + "[4] Andrei Ardelean. Computational Imaging SPAD Cameras. PhD thesis, École polytechnique fédérale de Lausanne, 2023. 2, 6, 8", + "[5] Souptik Barua, Yoshitaka Miyatani, and Ashok Veeraraghavan. Direct face detection and video reconstruction from event cameras. In WACV, pages 1-9. IEEE, 2016. 3", + "[6] Raphael Berner, Christian Brandli, Minhao Yang, S-C Liu, and Tobi Delbruck. A 240x180 120dB 10mW 12us-latency sparse output vision sensor for mobile applications. In Proceedings of the International Image Sensors Workshop, number CONF, pages 41-44, 2013. 1, 2", + "[7] Anthony Bisulco, Fernando Cladera Ojeda, Volkan Isler, and Daniel D Lee. Fast motion understanding with spatiotemporal neural networks and dynamic vision sensors. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 14098-14104. IEEE, 2021. 1", + "[8] Assim Boukhayma, Arnaud Peizerat, and Christian Enz. A sub-0.5 electron read noise VGA image sensor in a standard CMOS process. IEEE Journal of Solid-State Circuits, 2016. 4", + "[9] Christian Brandli, Raphael Berner, Minhao Yang, Shih-Chii Liu, and Tobi Delbruck. A $240 \\times 180$ dB $3\\mu s$ latency global shutter spatiotemporal vision sensor. IEEE Journal of Solid-State Circuits, 49(10):2333-2341, 2014. 2, 3", + "[10] Christian Brandli, Lorenz Muller, and Tobi Delbruck. Realtime, high-speed video decompression using a frame-and event-based DAVIS sensor. In 2014 IEEE International Symposium on Circuits and Systems (ISCAS), pages 686-689. IEEE, 2014. 3", + "[11] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Computer Vision – ECCV 2020, pages 213–229, Cham, 2020. Springer International Publishing. 7", + "[12] Andrea Censi and Davide Scaramuzza. Low-latency event-based visual odometry. In 2014 IEEE International Conference on Robotics and Automation (ICRA), pages 703-710. IEEE, 2014. 1", + "[13] Matthew Cook, Luca Gugelmann, Florian Jug, Christoph Krautz, and Angelika Steger. Interacting maps for fast visual interpretation. In The 2011 International Joint Conference on Neural Networks, pages 770-776, 2011. 3", + "[14] Yue Gao, Siqi Li, Yipeng Li, Yandong Guo, and Qionghai Dai. Superfast: $200 \\times$ video frame interpolation via event camera. IEEE TPAMI, 45(6):7764-7780, 2023. 1", + "[15] Daniel Gehrig, Henri Rebecq, Guillermo Gallego, and Davide Scaramuzza. EKLT: Asynchronous photometric feature tracking using events and frames. IJCV, 128(3):601-618, 2020. 1", + "[16] Mathias Gehrig, Mario Millhäuser, Daniel Gehrig, and Davide Scaramuzza. E-raft: Dense optical flow from event cameras. In 2021 International Conference on 3D Vision (3DV), pages" + ], + "bbox": [ + 78, + 114, + 470, + 890 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "197-206.IEEE,2021.7", + "[17] Istvan Gyongy, Neale AW Dutton, and Robert K Henderson. Single-photon tracking for high-speed vision. Sensors, 18(2):323, 2018. 3", + "[18] Chris Harris, Mike Stephens, et al. A combined corner and edge detector. In *Alvey vision conference*, pages 10-5244. CiteSeer, 1988. 7", + "[19] Samuel W. Hasinoff, Dillon Sharlet, Ryan Geiss, Andrew Adams, Jonathan T. Barron, Florian Kainz, Jiawen Chen, and Marc Levoy. Burst photography for high dynamic range and low-light imaging on mobile cameras. ACM TOG, 35(6):1-12, 2016. 7", + "[20] Botao He, Haojia Li, Siyuan Wu, Dong Wang, Zhiwei Zhang, Qianli Dong, Chao Xu, and Fei Gao. Fast-dynamic-vision: Detection and tracking dynamic objects with event and depth sensing. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3071-3078. IEEE, 2021. 1", + "[21] Javier Hidalgo-Carrio, Guillermo Gallego, and Davide Scaramuzza. Event-aided direct sparse odometry. In CVPR, pages 5781-5790, 2022. 1", + "[22] Yasunobu Hitomi, Jinwei Gu, Mohit Gupta, Tomoo Mitsunaga, and Shree K. Nayar. Video from a single coded exposure photograph using a learned over-complete dictionary. In 2011 International Conference on Computer Vision, pages 287-294, 2011. 5", + "[23] Jing Huang, Menghan Guo, and Shoushun Chen. A dynamic vision sensor with direct logarithmic output and full-frame picture-on-demand. In 2017 IEEE International Symposium on Circuits and Systems (ISCAS), pages 1-4, 2017. 2", + "[24] Zhewei Huang, Tianyuan Zhang, Wen Heng, Boxin Shi, and Shuchang Zhou. Real-time intermediate flow estimation for video frame interpolation. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 6", + "[25] Jorge Igual. Photographic noise performance measures based on raw files analysis of consumer cameras. *Electronics*, 8(11):1284, 2019. 4", + "[26] Atul Ingle, Andreas Velten, and Mohit Gupta. High Flux Passive Imaging With Single-Photon Sensors. In CVPR, 2019. 3", + "[27] Atul Ingle, Trevor Seets, Mauro Buttafava, Shantanu Gupta, Alberto Tosi, Mohit Gupta, and Andreas Velten. Passive inter-photon imaging. In CVPR, 2021. 3", + "[28] Kiyotaka Iwabuchi, Yusuke Kameda, and Takayuki Hamamoto. Image quality improvements based on motion-based deblurring for single-photon imaging. IEEE Access, 9:30080-30094, 2021. 3", + "[29] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. CoTracker: It is better to track together. 2023.", + "[30] Anthony Kay. Tesseract: an open-source optical character recognition engine. *Linux Journal*, 2007(159):2, 2007.", + "[31] Hanme Kim, Ankur Handa, Ryad Benosman, Sio-Hoi Ieng, and Andrew J Davison. Simultaneous mosaicing and tracking with an event camera. Proc. Brit. Mach. Vis. Conf, 43:566-576, 2014. 3", + "[32] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólár, and Ross Girshick. Segment anything, 2023. 7", + "[33] Haolong Li and Joerg Stueckler. Tracking 6-DoF object motion from events and frames. In 2021 IEEE International Conference" + ], + "bbox": [ + 503, + 92, + 893, + 895 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "25015", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Robotics and Automation (ICRA), pages 14171-14177. IEEE, 2021. 1", + "[34] P Lichtsteiner. 64x64 event-driven logarithmic temporal derivative silicon retina. In Program 2003 IEEE Workshop on CCD and AIS, 2003. 1, 2, 4", + "[35] Yuhao Liu, Felipe Gutierrez-Barragan, Atul Ingle, Mohit Gupta, and Andreas Velten. Single-photon camera guided extreme dynamic range imaging. In WACV, pages 1575-1585, 2022. 3", + "[36] Sizhuo Ma, Shantanu Gupta, Arin C. Ulku, Claudio Bruschini, Edoardo Charbon, and Mohit Gupta. Quanta burst photography. ACM TOG, 39(4):1-16, 2020. 3, 8", + "[37] Sizhuo Ma, Paul Mos, Edoardo Charbon, and Mohit Gupta. Burst vision using single-photon cameras. In WACV, pages 5375-5385, 2023. 3", + "[38] Kazuhiro Morimoto, Andrei Ardelean, Ming-Lo Wu, Arin Can Ulku, Ivan Michel Antolovic, Claudio Bruschini, and Edoardo Charbon. Megapixel time-gated SPAD image sensor for 2D and 3D imaging applications. Optica, 7(4):346-354, 2020. 4", + "[39] Gottfried Munda, Christian Reinbacher, and Thomas Pock. Realtime intensity-image reconstruction for event cameras using manifold regularisation. *IJCV*, 126:1381–1393, 2018. 3", + "[40] Shuichi Namiki, Shunichi Sato, Yusuke Kameda, and Takayuki Hamamoto. Imaging method using multi-threshold pattern for photon detection of quanta image sensor. In International Workshop on Advanced Imaging Technology (IWAIT) 2022, page 1217702. SPIE, 2022. 3", + "[41] Liyuan Pan, Cedric Scheerlinck, Xin Yu, Richard Hartley, Miao-miao Liu, and Yuchao Dai. Bringing a blurry frame alive at high frame-rate with an event camera. In CVPR, 2019. 3", + "[42] Liyuan Pan, Richard Hartley, Cedric Scheerlinck, Miaomiao Liu, Xin Yu, and Yuchao Dai. High frame rate video reconstruction based on an event camera. IEEE TPAMI, 44(5):2519-2533, 2022. 6", + "[43] Federico Paredes-Valles and Guido C. H. E. de Croon. Back to event basics: Self-supervised learning of image reconstruction for event cameras via photometric constancy. In CVPR, pages 3446–3455, 2021. 3", + "[44] Christoph Posch, Daniel Matolin, and Rainer Wohlgenannt. A QVGA 143 dB dynamic range frame-free pwm image sensor with lossless pixel-level video compression and time-domain cds. IEEE Journal of Solid-State Circuits, 46(1):259-275, 2011. 1, 2", + "[45] Ramesh Raskar, Amit Agrawal, and Jack Tumblin. Coded exposure photography: motion deblurring using fluttered shutter. In Acm Siggraph 2006 Papers, pages 795-804. 2006. 5", + "[46] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. High speed and high dynamic range video with an event camera. IEEE TPAMI, 2019. 3", + "[47] Dikpal Reddy, Ashok Veeraraghavan, and Rama Chellappa. P2C2: Programmable pixel compressive camera for high speed imaging. In CVPR 2011, pages 329-336, 2011. 5", + "[48] Alexis Rochas. Single photon avalanche diodes in cmos technology. Technical report, Citeseer, 2003. 4", + "[49] Nitin J Sanket, Chethan M Parameshwara, Chahat Deep Singh, Ashwin V Kuruttukulam, Cornelia Fermüller, Davide Scaramuzza, and Yiannis Aloimonos. Evdodgenet: Deep dynamic obstacle dodging with event cameras. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 10651-10657. IEEE, 2020. 1" + ], + "bbox": [ + 78, + 90, + 470, + 893 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Cedric Scheerlinck, Nick Barnes, and Robert Mahony. Continuous-time intensity estimation using event cameras. In ACCV, pages 308-324. Springer, 2018. 3", + "[51] Cedric Scheerlinck, Henri Rebecq, Daniel Gehrig, Nick Barnes, Robert Mahony, and Davide Scaramuzza. Fast image reconstruction with an event camera. In WACV, 2020. 3", + "[52] Trevor Seets, Atul Ingle, Martin Laurenzis, and Andreas Velten. Motion adaptive deblurring with single-photon cameras. In WACV, pages 1945-1954, 2021. 3", + "[53] Prasan Shedligeri and Kaushik Mitra. Photorealistic image reconstruction from hybrid intensity and event-based sensor. Journal of Electronic Imaging, 28(6):063012-063012, 2019. 3", + "[54] Prasan Shedligeri, Anupama S, and Kaushik Mitra. A unified framework for compressive video recovery from coded exposure techniques. In WACV, pages 1600-1609, 2021. 6", + "[55] Mark Sheinin, Yoav Y. Schechner, and Kiriakos N. Kutulakos. Computational imaging on the electric grid. In CVPR, 2017. 8", + "[56] Hyeunjun Sim, Jihyong Oh, and Munchurl Kim. XVFI: extreme video frame interpolation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14489-14498, 2021. 6", + "[57] Timo Stoffregen, Cedric Scheerlinck, Davide Scaramuzza, Tom Drummond, Nick Barnes, Lindsay Kleeman, and Robert Mahony. Reducing the sim-to-real gap for event cameras. In ECCV, pages 534-549. Springer, 2020. 6", + "[58] Binyi su, Lei Yu, and Wen Yang. Event-based high frame-rate video reconstruction with a novel cycle-event network. In ICIP, pages 86-90, 2020. 3", + "[59] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 7", + "[60] Varun Sundar, Andrei Ardelean, Tristan Swedish, Claudio Bruschini, Edoardo Charbon, and Mohit Gupta. Sodacam: Software-defined cameras via single-photon imaging. In ICCV, pages 8165-8176, 2023. 3, 6, 8", + "[61] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision – ECCV 2020, pages 402–419, Cham, 2020. Springer International Publishing. 7", + "[62] Arin Can Ulku, Claudio Bruschini, Ivan Michel Antolovic, Yung Kuo, Rinet Ankri, Shimon Weiss, Xavier Michalet, and Edoardo Charbon. A $512 \\times 512$ SPAD Image Sensor With Integrated Gating for Widefield FLIM. IEEE Journal of Selected Topics in Quantum Electronics, 25(1):1-12, 2019. 2, 4, 6", + "[63] Lishun Wang, Miao Cao, and Xin Yuan. Efficientsci: Densely connected network with space-time factorization for large-scale video snapshot compressive imaging. In CVPR, pages 18477-18486, 2023. 6", + "[64] Zhaohui Wang, Xiao Lin, Abhinav Mishra, and Ram Sriharsha. Online changepoint detection on a budget. In 2021 International Conference on Data Mining Workshops (ICDMW), pages 414-420. IEEE, 2021. 5", + "[65] Mian Wei, Sotiris Nousias, Rahul Gulve, David B. Lindell, and Kiriakos N. Kutulakos. Passive ultra-wideband single-photon imaging. In ICCV, pages 8135-8146, 2023. 3", + "[66] Xin Yuan, Yang Liu, Jinli Suo, Frédo Durand, and Qionghai Dai. Plug-and-play algorithms for video snapshot compressive imaging. IEEE Transactions on Pattern Analysis and Machine" + ], + "bbox": [ + 501, + 92, + 893, + 893 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "25016", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "ref_text", + "text": "Intelligence, 44(10):7093-7111, 2022. 6", + "bbox": [ + 107, + 90, + 339, + 104 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[67] Zelin Zhang, Anthony J Yezzi, and Guillermo Gallego. Formulating event-based image reconstruction as a linear inverse problem with deep regularization using optical flow. IEEE TPAMI, 45 (07):8372-8389, 2023. 3", + "[68] Alex Zihao Zhu, Nikolay Atanasov, and Kostas Daniilidis. Event-based visual inertial odometry. In CVPR, 2017. 1", + "[69] Yunhao Zou, Yinqiang Zheng, Tsuyoshi Takatani, and Ying Fu. Learning to reconstruct high speed and high dynamic range videos from events. In CVPR, pages 2024-2033, 2021. 3" + ], + "bbox": [ + 78, + 106, + 470, + 231 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "25017", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Generalized Event Cameras/e213eaf6-87b7-49bd-b232-571683e80e75_model.json b/2024/Generalized Event Cameras/e213eaf6-87b7-49bd-b232-571683e80e75_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d23219a0c46d63130d39c4e72f5c24bce54f90ba --- /dev/null +++ b/2024/Generalized Event Cameras/e213eaf6-87b7-49bd-b232-571683e80e75_model.json @@ -0,0 +1,3555 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.349, + 0.132, + 0.622, + 0.15 + ], + "angle": 0, + "content": "Generalized Event Cameras" + }, + { + "type": "text", + "bbox": [ + 0.224, + 0.181, + 0.341, + 0.198 + ], + "angle": 0, + "content": "Varun Sundar†*" + }, + { + "type": "text", + "bbox": [ + 0.206, + 0.203, + 0.353, + 0.215 + ], + "angle": 0, + "content": "vsundar4@wisc.edu" + }, + { + "type": "text", + "bbox": [ + 0.409, + 0.181, + 0.554, + 0.198 + ], + "angle": 0, + "content": "Matthew Dutson†*" + }, + { + "type": "text", + "bbox": [ + 0.417, + 0.203, + 0.546, + 0.215 + ], + "angle": 0, + "content": "dutson@wisc.edu" + }, + { + "type": "text", + "bbox": [ + 0.621, + 0.181, + 0.753, + 0.198 + ], + "angle": 0, + "content": "Andrei Ardelean†" + }, + { + "type": "text", + "bbox": [ + 0.61, + 0.203, + 0.765, + 0.216 + ], + "angle": 0, + "content": "a.ardelean@epfl.ch" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.223, + 0.372, + 0.24 + ], + "angle": 0, + "content": "Claudio Bruschini\\(^{\\S}\\)" + }, + { + "type": "text", + "bbox": [ + 0.195, + 0.245, + 0.568, + 0.26 + ], + "angle": 0, + "content": "{claudio.bruschini, edoardo.charbon}@epfl.ch" + }, + { + "type": "text", + "bbox": [ + 0.647, + 0.224, + 0.751, + 0.242 + ], + "angle": 0, + "content": "Mohit Gupta†" + }, + { + "type": "text", + "bbox": [ + 0.621, + 0.246, + 0.776, + 0.259 + ], + "angle": 0, + "content": "mohitg@cs.wisc.edu" + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.267, + 0.834, + 0.286 + ], + "angle": 0, + "content": "†University of Wisconsin-Madison ‡NovoViz École Polytechnique Fédérale de Lausanne" + }, + { + "type": "text", + "bbox": [ + 0.277, + 0.289, + 0.692, + 0.304 + ], + "angle": 0, + "content": "wisionlab.com/project/generalized-event-cameras/" + }, + { + "type": "title", + "bbox": [ + 0.236, + 0.34, + 0.312, + 0.355 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.371, + 0.474, + 0.657 + ], + "angle": 0, + "content": "Event cameras capture the world at high time resolution and with minimal bandwidth requirements. However, event streams, which only encode changes in brightness, do not contain sufficient scene information to support a wide variety of downstream tasks. In this work, we design generalized event cameras that inherently preserve scene intensity in a bandwidth-efficient manner. We generalize event cameras in terms of when an event is generated and what information is transmitted. To implement our designs, we turn to single-photon sensors that provide digital access to individual photon detections; this modality gives us the flexibility to realize a rich space of generalized event cameras. Our single-photon event cameras are capable of high-speed, high-fidelity imaging at low readout rates. Consequently, these event cameras can support plug-and-play downstream inference, without capturing new event datasets or designing specialized event-vision models. As a practical implication, our designs, which involve lightweight and near-sensor-compatible computations, provide a way to use single-photon sensors without exorbitant bandwidth costs." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.667, + 0.203, + 0.683 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.692, + 0.472, + 0.844 + ], + "angle": 0, + "content": "Event cameras [6, 34, 44] sense the world at high speeds, providing visual information with minimal bandwidth and power. They achieve this by transmitting only changes in scene brightness, when significant \"events\" occur. However, there is a cost. Raw event data, a sparse stream of binary values, does not hold sufficient information to be used directly with mainstream vision algorithms. Therefore, while event cameras have been successful at certain tasks (e.g., object tracking [15, 33], obstacle avoidance [7, 20, 49], and high-speed odometry [12, 21, 68]), they are not widely deployed as general-purpose vision sen" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.341, + 0.892, + 0.386 + ], + "angle": 0, + "content": "sors, and often need to be supplemented with conventional cameras [14, 15, 21]. These limitations are holding back this otherwise powerful technology." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.387, + 0.895, + 0.657 + ], + "angle": 0, + "content": "Is it possible to realize the promise of event cameras, i.e., high temporal resolution at low bandwidth, while preserving rich scene intensity information? To realize these seemingly conflicting goals, we propose a novel family of generalized event cameras. We conceptualize a space of event cameras along two key axes (Fig. 1 (top)): (a) \"when to transmit information,\" formalized as a change detection procedure \\(\\Delta\\); and (b) \"what information to transmit,\" characterized by an integrator \\(\\Sigma\\) that encodes incident flux. Existing event cameras represent one operating point in this \\((\\Sigma, \\Delta)\\) space. Our key observation is that by exploring this space and considering new \\((\\Sigma, \\Delta)\\) combinations, we can design event cameras that preserve scene intensity. We propose more general integrators, e.g., that represent flux according to motion levels, that span spatial patches, or that employ temporal coding (Fig. 1 (middle)). We also introduce robust change detectors that better distinguish motion from noise, by considering increased spatiotemporal contexts and modeling noise in the sensor measurements." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.896, + 0.902 + ], + "angle": 0, + "content": "Despite their conceptual appeal, physically implementing generalized event cameras is a challenge. This is because the requisite computations must be performed at the sensor to achieve the desired bandwidth reductions. For example, existing event cameras perform simple integration and thresholding operations via analog in-pixel circuitry. However, more general \\((\\Sigma ,\\Delta)\\) combinations are not always amenable to analog implementations; even feasible designs might require years of hardware iteration and production scaling. To build physical realizations of generalized event cameras, we leverage an emerging sensor technology: single-photon avalanche diodes (SPADs) that provide digital access to photon detections at extremely high frame rates \\((\\sim 100\\mathrm{kHz})\\). This allows us to compose arbitrary software-level signal transformations, such as those required by generalized event cameras. Further, we are not locked to a particular event camera design and can realize multiple configurations" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.851, + 0.47, + 0.901 + ], + "angle": 0, + "content": "*denotes equal contribution. This research was supported in parts by NSF CAREER award 1943149, NSF award CNS-2107060, and the Swiss National Science Foundation grant 200021_166289. We thank Paul Mos for providing us access to SwissSPAD2 acquisition software." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25007" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.407, + 0.094, + 0.565, + 0.104 + ], + "angle": 0, + "content": "Conventional event camera" + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.11, + 0.255, + 0.121 + ], + "angle": 0, + "content": "Integrator (accumulates flux, \\(\\Sigma\\))" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.125, + 0.264, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.21, + 0.213, + 0.23 + ], + "angle": 0, + "content": "Low-pass filtered brightness" + }, + { + "type": "image_caption", + "bbox": [ + 0.301, + 0.11, + 0.48, + 0.121 + ], + "angle": 0, + "content": "Change detector (triggers events, \\(\\Delta\\)" + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.125, + 0.501, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.209, + 0.465, + 0.23 + ], + "angle": 0, + "content": "Threshold-based detector Analog comparator" + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.111, + 0.633, + 0.121 + ], + "angle": 0, + "content": "Sensor output" + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.125, + 0.683, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.53, + 0.209, + 0.667, + 0.23 + ], + "angle": 0, + "content": "Stream of change polarities \\((x,y,t,\\pm 1)\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.751, + 0.111, + 0.837, + 0.121 + ], + "angle": 0, + "content": "Intensity imaging" + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.128, + 0.887, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.726, + 0.209, + 0.863, + 0.23 + ], + "angle": 0, + "content": "Reconstruction with E2VID+ 1000 FPS at 20 FPS bitrate" + }, + { + "type": "title", + "bbox": [ + 0.411, + 0.233, + 0.561, + 0.244 + ], + "angle": 0, + "content": "Generalized event camera" + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.25, + 0.265, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.096, + 0.335, + 0.139, + 0.354 + ], + "angle": 0, + "content": "Motion adaptive" + }, + { + "type": "image_caption", + "bbox": [ + 0.156, + 0.335, + 0.194, + 0.354 + ], + "angle": 0, + "content": "Spatial patches" + }, + { + "type": "image_caption", + "bbox": [ + 0.211, + 0.335, + 0.26, + 0.355 + ], + "angle": 0, + "content": "Coded exposures" + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.252, + 0.502, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.314, + 0.334, + 0.484, + 0.353 + ], + "angle": 0, + "content": "Enhanced spatio-temporal support Noise-aware thresholding" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.251, + 0.683, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.544, + 0.334, + 0.678, + 0.353 + ], + "angle": 0, + "content": "Stream of integrator values \\((x,y,t,\\Sigma (x,y,t))\\)" + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.253, + 0.887, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.731, + 0.335, + 0.863, + 0.354 + ], + "angle": 0, + "content": "Reconstruction 3025 FPS at 30 FPS bitrate" + }, + { + "type": "title", + "bbox": [ + 0.41, + 0.359, + 0.562, + 0.371 + ], + "angle": 0, + "content": "Plug-and-play event vision" + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.377, + 0.243, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.456, + 0.238, + 0.476 + ], + "angle": 0, + "content": "Contact point of a tennis shot Image Reconstruction" + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.378, + 0.405, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.26, + 0.456, + 0.385, + 0.476 + ], + "angle": 0, + "content": "Tracking points at 0.3 lux CoTracker [29]" + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.378, + 0.565, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.42, + 0.456, + 0.551, + 0.476 + ], + "angle": 0, + "content": "Nighttime object detection DETR [11]" + }, + { + "type": "image", + "bbox": [ + 0.569, + 0.378, + 0.726, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.588, + 0.456, + 0.706, + 0.476 + ], + "angle": 0, + "content": "OCR on a snapping tape Tesseract [30]" + }, + { + "type": "image", + "bbox": [ + 0.73, + 0.378, + 0.887, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.743, + 0.456, + 0.878, + 0.476 + ], + "angle": 0, + "content": "Segmenting a dart in midair Segment Anything [32]" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.483, + 0.892, + 0.581 + ], + "angle": 0, + "content": "Figure 1. Generalized event cameras. (top) Event cameras generate outputs in response to abrupt changes in scene intensity. We describe this as a combination of a low-pass integrator and a threshold-based change detector. (middle) We generalize the space of event cameras by designing integrators that capture rich intensity information, and more reliable change detectors that utilize larger spatiotemporal contexts and noise-aware thresholding (Secs. 4.1 to 4.3). Unlike existing events, our generalized event streams inherently preserve scene intensity, e.g., this ping-pong ball slingshotted against a brick wall backdrop. (bottom) Generalized event cameras enable high-fidelity bandwidth-efficient imaging: providing 3025 FPS reconstructions with a readout equivalent to a 30 FPS camera. Consequently, generalized events facilitate plug-and-play inference on a multitude of tasks in challenging scenarios (insets depict the extent of motion over \\(30~\\mathrm{ms}\\))." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.599, + 0.214, + 0.612 + ], + "angle": 0, + "content": "with the same sensor." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.621, + 0.471, + 0.833 + ], + "angle": 0, + "content": "Implications: extreme, bandwidth-efficient vision. Generalized event cameras support high-speed, high-quality image reconstruction, but at low bandwidths quintessential of current event cameras. For example, Fig. 1 (middle, bottom) shows reconstructions at 3025 FPS that have an effective readout of a 30 FPS frame-based camera. Further, our methods have strong low-light performance due to the SPAD's single-photon sensitivity. As we show in Fig. 1 (bottom), preserving scene intensity facilitates plug-and-play inference in challenging scenarios, with state-of-the-art vision algorithms. Critically, this does not require retraining vision models or curating dedicated datasets, which is a significant challenge for unconventional imagers. This plug-and-play capability is vital to realizing universal event vision that retains the benefits of current event cameras." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Scope. We consider full-stack event perception: we conceptualize a novel space of event cameras, provide relevant single-photon algorithms, analyze their imaging capabilities and rate-distortion trade-offs, and show on-chip feasibility. We demon-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.895, + 0.705 + ], + "angle": 0, + "content": "strate imaging capabilities in Secs. 5.1 and 5.2 using the SwissS-PAD2 array [62], and show viable implementations of our algorithms for UltraPhase [4], a recent single-photon compute platform. All of these are critical to unlocking the promise of event cameras. However, our objective is not to develop an integrated system that incorporates all these components; this paper merely takes the first steps toward that goal." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.723, + 0.636, + 0.739 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Event camera designs. Perhaps most widespread is the DVS event camera [34], where each pixel generates an event in response to measured changes in (log) intensity. The DAVIS event camera [6, 9] couples DVS pixels with conventional CMOS pixels, providing access to image frames. However, the frames lack the dynamic range of DVS events. A recent design, CelexV [23], provides log-intensity frames using an external trigger. ATIS [44], a less prevalent design, features asynchronous intensity events, but its sophisticated circuitry reduces pixel fill factor. The above designs are based on analog processing; we instead" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25008" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.393, + 0.106 + ], + "angle": 0, + "content": "design event cameras on digital photon detections." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.112, + 0.47, + 0.308 + ], + "angle": 0, + "content": "Intensity imaging with event cameras. Several approaches have been explored to obtain images from events, including Poisson solvers [5], manifold regularization [39], assuming knowledge of camera motion [13, 31] or optical flow [67], and learning-based methods [43, 46, 51, 58, 69]. However, because events often lack sufficient scene information, they are often supplemented by conventional frames [10, 41, 50, 53], either from sensors such as DAVIS or using a multi-camera setup. Fusing events and frames presents challenges due to potential spatiotemporal misalignment and discrepancies in imaging modalities. Even when these challenges are overcome, we show that fusion methods produce lower fidelity than our proposed generalized event cameras." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.314, + 0.471, + 0.466 + ], + "angle": 0, + "content": "Passive single-photon imaging. In the past few years, SPADs have found compelling passive imaging applications; this includes high-dynamic range imaging [26, 27, 35, 40], motion deblurring [28, 36, 37, 52], high-speed tracking [17], and ultra wide-band videography [65]. The fine granularity of passive single-photon acquisition makes it possible to emulate a diverse set of imaging modalities [60], including event cameras, via post-capture processing. In this work, we go beyond emulating existing event cameras and design alternate event cameras that preserve high-fidelity intensity information." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.481, + 0.315, + 0.497 + ], + "angle": 0, + "content": "3. What is an Event Camera?" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.47, + 0.598 + ], + "angle": 0, + "content": "The defining characteristic of event cameras is that they transmit information selectively, in response to changes in scene content. This selectivity allows event cameras to encode scene information at high time resolutions required to capture scene dynamics, without proportionately high bandwidth. This is in contrast to frame-based cameras, where readout occurs at fixed intervals." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.47, + 0.659 + ], + "angle": 0, + "content": "We characterize event cameras in terms of two axes: what the camera transmits and when it transmits. As a concrete example, consider existing event cameras. They trigger events (\"when to transmit\") based on a fixed threshold:" + }, + { + "type": "equation", + "bbox": [ + 0.195, + 0.667, + 0.47, + 0.683 + ], + "angle": 0, + "content": "\\[\n\\left| \\Phi (\\mathbf {x}, t) - \\Phi_ {\\text {r e f}} (\\mathbf {x}) \\right| \\geq \\tau , \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.691, + 0.469, + 0.75 + ], + "angle": 0, + "content": "where \\(\\Phi (\\mathbf{x},t)\\) is a flux estimate at pixel \\(\\mathbf{x}\\) and time \\(t\\), and \\(\\tau\\) is the threshold. \\(\\Phi_{\\mathrm{ref}}(\\mathbf{x})\\) is a previously-recorded reference, set to \\(\\Phi (\\mathbf{x},t)\\) whenever an event is triggered. Each event consists of a packet" + }, + { + "type": "equation", + "bbox": [ + 0.177, + 0.751, + 0.469, + 0.767 + ], + "angle": 0, + "content": "\\[\n(\\mathbf {x}, t, \\operatorname {s i g n} (\\Phi (\\mathbf {x}, t) - \\Phi_ {\\mathrm {r e f}} (\\mathbf {x}))) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.772, + 0.451, + 0.787 + ], + "angle": 0, + "content": "that encodes the polarity of the change (\"what to transmit\")." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.788, + 0.469, + 0.863 + ], + "angle": 0, + "content": "Event polarities, although adequate for some applications, do not retain sufficient information to support a general set of computer vision tasks. A stream of event polarities is an extremely lossy representation. Notably, it only defines scene intensity up to an initial unknown reference value, and it does" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.09, + 0.628, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.152, + 0.605, + 0.162 + ], + "angle": 0, + "content": "(a) 1-bit events" + }, + { + "type": "image", + "bbox": [ + 0.632, + 0.091, + 0.757, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.659, + 0.151, + 0.731, + 0.162 + ], + "angle": 0, + "content": "(b) Flux levels" + }, + { + "type": "image", + "bbox": [ + 0.762, + 0.091, + 0.886, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.765, + 0.151, + 0.883, + 0.163 + ], + "angle": 0, + "content": "(c) Adaptive exposures" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.168, + 0.894, + 0.251 + ], + "angle": 0, + "content": "Figure 2. Altering \"what to transmit.\" (a) We sum the events generated by a jack-in-the-box toy as it springs up. This sum gives a lossy encoding of brightness changes in dynamic regions. (b) Transmitting levels instead of changes helps recover details in static regions. (c) Adaptive exposures, which accumulate flux between consecutive events, provide substantial noise reduction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.262, + 0.892, + 0.291 + ], + "angle": 0, + "content": "not encode any information in regions not producing events, i.e., regions with little or no motion." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.293, + 0.893, + 0.398 + ], + "angle": 0, + "content": "Our key observation is that existing event cameras represent just one operating point in a broader space of generalized event cameras, which is defined by two axes: \"what to transmit\" and \"when to transmit.\" By considering alternate points in this space, we can design event cameras that preserve high-fidelity scene intensity. This enables plug-and-play inference with a host of algorithms developed by the mainstream vision community." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.399, + 0.892, + 0.429 + ], + "angle": 0, + "content": "We begin with a conceptual exploration of this generalized space, before describing its physical implementation." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.433, + 0.893, + 0.599 + ], + "angle": 0, + "content": "Generalizing \"what to transmit.\" As a first step, we can modify the event camera such that it transmits \\( n \\)-bit values instead of one-bit change polarities. When an event is triggered, we send the current value of \\( \\Phi(\\mathbf{x}, t) \\); if a pixel triggers no events, we transmit \\( \\Phi \\) during the final readout. As we show in Fig. 2(b), this simple change allows us to recover scene intensity, even in static regions. It is important to note that, while the transmitted quantity differs from conventional events, we retain the defining feature of an event camera: selective transmission based on scene dynamics (where we transmit according to Eq. (1)). Thus, the readout remains decoupled from the time resolution." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.6, + 0.892, + 0.675 + ], + "angle": 0, + "content": "If \\(\\Phi (\\mathbf{x},t)\\) were a perfect estimate of scene intensity, then the changes thus far would suffice. However, \\(\\Phi\\) is fundamentally noisy: to capture high-speed changes, \\(\\Phi\\) must encompass a shorter duration, which leads to higher noise. This is a manifestation of the classical noise-blur tradeoff." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.676, + 0.892, + 0.809 + ], + "angle": 0, + "content": "To address this problem, we introduce the abstraction of an integrator (or \\(\\Sigma\\)), that defines how we accumulate incident flux and, in turn, what we transmit. Ideally, we want the integrator to adapt to scene dynamics, i.e., accumulate over longer durations when there is less motion, and vice versa. We observe that event generation, which is based on scene dynamics, can be used to formulate an adaptive integrator. Specifically, we propose an integrator \\(\\Sigma_{\\mathrm{cum}}\\) that computes the cumulative flux since the last event:" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.809, + 0.892, + 0.842 + ], + "angle": 0, + "content": "\\[\n\\Sigma_ {\\mathrm {c u m l}} (\\mathbf {x}, t) = \\int_ {T _ {0}} ^ {t} \\Phi (\\mathbf {x}, s) d s, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.845, + 0.892, + 0.876 + ], + "angle": 0, + "content": "where \\(T_{0}\\) is the time of the last event. When an event is triggered at time \\(T_{1}\\), we communicate the value\\(^{2}\\) of \\(\\Sigma_{\\mathrm{cum}}(\\mathbf{x}, T_{1})\\), which" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.875, + 0.469, + 0.901 + ], + "angle": 0, + "content": "1Event cameras such as the DVS [9] measure a temporally low-pass filtered estimate of log-flux; we absorb this into \\(\\Phi (\\mathbf{x},t)\\) for brevity." + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.887, + 0.892, + 0.901 + ], + "angle": 0, + "content": "2We can either transmit values of \\(\\Sigma_{\\mathrm{cum}}\\) or differences (changes) to \\(\\Sigma_{\\mathrm{cum}}\\);" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25009" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.078, + 0.09, + 0.894, + 0.172 + ], + "angle": 0, + "content": "
Event cameraIntegrator (Σ)Change detector (Δ)Event packetsMin. latencyIntensity info.?Low-light perf.
Existing (DVS [34])logarithmiccomparatorbinary10-6to 10-5sXpoor
Sec. 4.1adaptive exposureBayesian change detector [2]scalar10-5s✓good
Sec. 4.2adaptive exposurevariance-aware differencespatches10-4s✓good
Sec. 4.3coded exposureBinomial confidence intervalvector10-3s✓good
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.175, + 0.895, + 0.22 + ], + "angle": 0, + "content": "Table 1. Summary of generalized event cameras. Our designs integrate photon detections \\((\\Sigma)\\) and detect scene-content changes \\((\\Delta)\\) in distinct ways. We compare our designs to existing DVS event cameras based on their event streams, latencies, and intensity-preserving nature. While providing a direct power comparison to DVS is difficult, we compare the power characteristics among our designs in Sec. 5.4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.23, + 0.471, + 0.292 + ], + "angle": 0, + "content": "we interpret as the intensity throughout \\([T_0, T_1]\\). This approach yields a piece-wise constant time series, with segments delimited by events. Adaptive exposures significantly reduce noise, as we show in Fig. 2 (c)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.294, + 0.471, + 0.43 + ], + "angle": 0, + "content": "Generalizing \"when to transmit.\" The success of the adaptive integrator crucially depends on the reliability of events; for example, triggering false events in static regions causes unnecessary noise. We refer to the event-generation procedure as the change detector, denoted by \\(\\Delta\\). Current event cameras detect changes by applying a fixed threshold to measured intensity differences (Eq. (1)). This method has two key limitations: it only considers the value of \\(\\Phi\\) at pixel location \\(\\mathbf{x}\\) and time \\(t\\), and it is not attuned to the stochasticity in \\(\\Phi\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.43, + 0.471, + 0.536 + ], + "angle": 0, + "content": "We design more robust change detectors that (1) leverage enhanced spatiotemporal contexts, and (2) incorporate noise awareness, either explicitly by tuning thresholds, or implicitly by modulating the detector's behavior. Specifically, we improve reliability by using temporal forecasters (Sec. 4.1), by leveraging correlated changes in patches (Sec. 4.2), or by exploiting integrator statistics (Sec. 4.3)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.471, + 0.63 + ], + "angle": 0, + "content": "Realizing generalized event cameras. The critical detail remaining is how we implement our proposed designs in practice. We need direct access to flux estimates at a high time resolution. Conventional high-speed cameras can provide such access, however, they incur substantial per-frame read noise \\((\\sim 20 - 40e^{-}\\) [25]) that grows with frame rate [8]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.63, + 0.471, + 0.797 + ], + "angle": 0, + "content": "We turn to an emerging class of single-photon sensors, single-photon avalanche diodes (SPADs [48]), that has witnessed dramatic improvements in device practicality and key sensor characteristics (e.g., array sizes and fill factors) in recent years [38, 62]. SPADs can operate at extremely high speeds (\\(\\sim 100\\mathrm{kHz}\\)) without incurring per-frame read noise. Each \\(\\Phi(\\mathbf{x},t)\\) measured by a SPAD is limited only by the fundamental stochasticity of photon arrivals (shot noise). This allows SPADs to provide high timing resolution without a substantial noise penalty. In the next section, we describe the image formation model of SPADs and provide single-photon implementations of our designs." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.808, + 0.442, + 0.826 + ], + "angle": 0, + "content": "4. Single-Photon Generalized Event Cameras" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.834, + 0.473, + 0.88 + ], + "angle": 0, + "content": "A SPAD array can operate as a high-speed photon detector, producing binary frames as output. Each binary value indicates whether at least one photon was detected during an expo" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.23, + 0.892, + 0.26 + ], + "angle": 0, + "content": "sure. The SPAD output response, \\(\\Phi (\\mathbf{x},t)\\), can be modeled as a Bernoulli random variable, with" + }, + { + "type": "equation", + "bbox": [ + 0.589, + 0.266, + 0.892, + 0.284 + ], + "angle": 0, + "content": "\\[\nP (\\Phi (\\mathbf {x}, t) = 1) = 1 - e ^ {- N (\\mathbf {x}, t)}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.292, + 0.895, + 0.459 + ], + "angle": 0, + "content": "where \\( N(\\mathbf{x}, t) \\) is the average number of photo-electrons during an exposure, including any spurious detections. The inherently digital SPAD response allows us to compute software-level transformations on the signal \\( \\Phi(\\mathbf{x}, t) \\), including operations that may be challenging to realize via analog processing. These transformations can be readily reconfigured, which permits a spectrum of event camera designs, not just one particular choice. However, there is one consideration: our designs should be lightweight and computable on chip. As we show in Sec. 5.4, this is vital to implementing generalized event cameras without the practical costs associated with reading off raw SPAD outputs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.46, + 0.895, + 0.506 + ], + "angle": 0, + "content": "We now describe a set of SPAD-based event cameras (summarized in Tab. 1), beginning with the adaptive exposure method from the previous section." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.515, + 0.894, + 0.56 + ], + "angle": 0, + "content": "Adaptive-exposure event camera. We obtain a SPAD implementation of the adaptive exposure described in Eq. (3) by replacing the integral with a sum over photons:" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.566, + 0.892, + 0.607 + ], + "angle": 0, + "content": "\\[\n\\Sigma_ {\\mathrm {c u m l}} (\\mathbf {x}, t) = \\sum_ {s = T _ {0}} ^ {t} \\Phi (\\mathbf {x}, s). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.894, + 0.69 + ], + "angle": 0, + "content": "To generate events, we can use a threshold-based change detector (Eq. (1)). Differences between individual binary values are not sufficiently informative; therefore, we apply Eq. (1) to an exponential moving average (EMA) computed on \\(\\Phi\\). We call this an \"adaptive-EMA\" event camera." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.707, + 0.74, + 0.723 + ], + "angle": 0, + "content": "4.1. Bayesian Change Detector" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.732, + 0.892, + 0.824 + ], + "angle": 0, + "content": "A fixed-threshold change detector such as Eq. (1) does not account for the SPAD's image formation model; it uses the same threshold irrespective of the underlying variance in photon detections. As a result, such a detector may fail to detect changes in low-contrast regions without producing a large number of false-positive detections (see Fig. 3 (left))." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.901 + ], + "angle": 0, + "content": "In this section, we consider a Bayesian change detector, BOCPD [1], that is tailored to the Bernoulli statistics of photon detections. BOCPD uses a series of forecasters to estimate the likelihood of an abrupt change. At each time step, a new forecaster \\(\\nu_{t}\\) is initialized as a recurrence of previous forecasters," + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.888, + 0.308, + 0.9 + ], + "angle": 0, + "content": "we treat this as an implementation detail here." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "25010" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.464, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.18, + 0.47, + 0.263 + ], + "angle": 0, + "content": "Figure 3. Bayesian- vs. EMA-based change detection. (left) A fixed-threshold change detector (used in adaptive-EMA) makes it difficult to segment low-contrast changes. (center) The Bayesian formulation attunes to the stochasticity in incident flux and can detect fine-grained changes such as the corners of the hole saw bit; (right) as a result, the integrator captures the rotational dynamics." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.271, + 0.308, + 0.285 + ], + "angle": 0, + "content": "and existing forecasters are updated:" + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.292, + 0.47, + 0.332 + ], + "angle": 0, + "content": "\\[\n\\nu_ {t} = (1 - \\gamma) \\sum_ {s = 1} ^ {t - 1} l _ {s} \\nu_ {s}, \\quad \\nu_ {s} \\leftarrow \\gamma l _ {s} \\nu_ {s} \\forall s < t, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.34, + 0.469, + 0.446 + ], + "angle": 0, + "content": "where \\(\\gamma \\in [0,1]\\) is the sensitivity of the change detector, with larger \\(\\gamma\\) resulting in more frequent detections. \\(l_{s}\\) is the predictive likelihood of each forecaster, which we compute by tracking two values per forecaster, \\(\\alpha_{s}\\) and \\(\\beta_{s}\\), that correspond to the parameters of a Beta prior. For a new forecaster, these values are initialized to 1 each, reflecting a uniform prior. Existing \\((\\alpha_{s},\\beta_{s}),\\forall s < t,\\) are updated as" + }, + { + "type": "equation", + "bbox": [ + 0.121, + 0.453, + 0.47, + 0.468 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {s} \\leftarrow \\alpha_ {s} + \\Phi (\\mathbf {x}, t), \\quad \\beta_ {s} \\leftarrow \\beta_ {s} + 1 - \\Phi (\\mathbf {x}, t). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.476, + 0.47, + 0.537 + ], + "angle": 0, + "content": "\\(l_{s}\\) is given by \\(\\alpha_{s} / (\\alpha_{s} + \\beta_{s})\\) if \\(\\Phi (\\mathbf{x},t) = 1\\), and \\(\\beta_{s} / (\\alpha_{s} + \\beta_{s})\\) otherwise. An event is triggered if the highest-value forecaster does not correspond to \\(T_{0}\\), the timestamp of the last event; mathematically, if \\(\\arg \\max_t\\nu_t\\neq T_0\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.537, + 0.47, + 0.597 + ], + "angle": 0, + "content": "To make BOCPD viable in memory-constrained scenarios, we apply extreme pruning by retaining only the three highest-value forecasters [64]. We also incorporate restarts, deleting previous forecasters when a change is detected [2]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.597, + 0.469, + 0.658 + ], + "angle": 0, + "content": "Compared to an EMA-based change detector, the Bayesian approach more reliably triggers events in response to scene changes while better filtering out stochastic variations caused by photon noise—which we show in Fig. 3." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.666, + 0.347, + 0.683 + ], + "angle": 0, + "content": "4.2. Spatiotemporal Chunk Events" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.469, + 0.885 + ], + "angle": 0, + "content": "Sec. 4.1 leverages an expanded temporal context for change detection; however, it treats each pixel independently and does not exploit spatial information. In this section, we devise an event camera with enhanced spatial context that operates on small patches, e.g., of \\(4 \\times 4\\) pixels. It is difficult to derive efficient Bayesian change detectors for multivariate time series; thus, we adopt a model-free approach that does not explicitly parameterize the patch distribution. To afford computational breathing room for more expensive patch-wise operations, we employ temporal chunking. That is, we average \\(\\Phi(\\mathbf{x}, t)\\) over a small number of binary frames (e.g., 32 binary frames) instead of operating on individual binary frames; generally, this averaging does not induce perceptible blur." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.886, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Let vector \\(\\Phi_{\\mathrm{chunk}}(\\mathbf{y},t)\\) represent the chunk-wise average of" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.089, + 0.885, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.207, + 0.894, + 0.29 + ], + "angle": 0, + "content": "Figure 4. Spatiotemporal chunk events. We evaluate the difference between the current chunk and a stored reference in a learned linear-feature space. Unlike the L2 norm, which is permutation-invariant, the feature-space norm is sensitive to spatial structure. Randomly shuffling the pixel values reduces the transform-domain norm (the shuffled patch has a more \"noise-like\" structure)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.301, + 0.893, + 0.45 + ], + "angle": 0, + "content": "photon detections at patch location \\(\\mathbf{y}\\). Let vector \\(\\Sigma_{\\mathrm{patch}}(\\mathbf{y},t)\\) be an integrator representing the cumulative mean since the last event, but excluding \\(\\Phi_{\\mathrm{chunk}}\\). We want to estimate whether \\(\\Phi_{\\mathrm{chunk}}\\) belongs to the same distribution as \\(\\Sigma_{\\mathrm{patch}}\\). We do so with a lightweight approach, that computes the distance between \\(\\Phi_{\\mathrm{chunk}}\\) and \\(\\Sigma_{\\mathrm{patch}}\\) in the linear feature space of matrix \\(\\mathbf{P}\\). As we show in Fig. 4, linear features allow us to capture spatial structure within a patch. Geometrically, \\(\\mathbf{P}\\) induces a hyperellipsoidal decision boundary, in contrast to the spherical boundary of the L2 norm." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.454, + 0.783, + 0.468 + ], + "angle": 0, + "content": "This method generates an event whenever" + }, + { + "type": "equation", + "bbox": [ + 0.572, + 0.474, + 0.892, + 0.492 + ], + "angle": 0, + "content": "\\[\n\\left\\| \\mathbf {P} \\left(\\tilde {\\Phi} _ {\\text {c h u n k}} (\\mathbf {y}, t) - \\tilde {\\Sigma} _ {\\text {p a t c h}} (\\mathbf {y}, t)\\right) \\right\\| _ {2} \\geq \\tau , \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.499, + 0.893, + 0.605 + ], + "angle": 0, + "content": "where \\(\\tau\\) is the threshold. When there is no event, we extend the cumulative mean to include the current chunk. Before computing linear features, we normalize \\(\\Phi_{\\mathrm{chunk}}\\) and \\(\\Sigma_{\\mathrm{patch}}\\) element-wise according to the estimated variance in \\(\\Phi_{\\mathrm{chunk}} - \\Sigma_{\\mathrm{patch}}\\); we annotate the normalized versions with a tilde. We estimate the variance based on the fact that, in a static patch, the elements of \\(\\Phi_{\\mathrm{chunk}}\\) and \\(\\Sigma_{\\mathrm{patch}}\\) are independent binomial random variables." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.606, + 0.892, + 0.697 + ], + "angle": 0, + "content": "We train the matrix \\(\\mathbf{P}\\) on simulated SPAD data, generated from interpolated high-speed video. We apply backpropagation through time to minimize the MSE error of the transmitted patch values. To address the non-differentiability arising from the threshold, we employ surrogate gradients. Please see the supplementary material for complete details of this method." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.71, + 0.722, + 0.726 + ], + "angle": 0, + "content": "4.3. Coded-Exposure Events" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.894, + 0.901 + ], + "angle": 0, + "content": "In this section, we design a generalized event camera by applying change detection to coded exposures [22, 45, 47], which capture temporal variations by multiplexing photon detections over an integration window. This is interesting in two aspects. First, we are designing event streams based on an modality not typically associated with event cameras. Second, we show that high-speed information can be obtained even when the change detector operates at a coarser time granularity. Coded-exposure events provide somewhat lower fidelity than our designs in Secs. 4.1 and 4.2, but are more compute- and power-efficient, owing to less frequent execution of the change detector." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25011" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.138 + ], + "angle": 0, + "content": "At each pixel, we multiplex a temporal chunk of \\( T_{\\mathrm{code}} \\) (\\( \\sim 256 - 512 \\)) binary values with a set of \\( J \\) (\\( \\sim 2 - 6 \\)) codes \\( C^j(\\mathbf{x}, t) \\forall 1 \\leq j \\leq J \\), producing \\( J \\) coded exposures" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.144, + 0.472, + 0.186 + ], + "angle": 0, + "content": "\\[\n\\Sigma_ {\\text {c o d e d}} ^ {j} (\\mathbf {x}, t) = \\sum_ {s = t - T _ {\\text {c o d e}}} ^ {t} \\Phi (\\mathbf {x}, s) C ^ {j} (\\mathbf {x}, s). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.194, + 0.47, + 0.224 + ], + "angle": 0, + "content": "The codes \\(C^j\\) are chosen to be random, mutually orthogonal binary masks, each containing \\(T_{\\mathrm{code}} / \\max (2,J)\\) ones [60]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.224, + 0.471, + 0.359 + ], + "angle": 0, + "content": "We exploit the statistics of coded exposures to derive a change detector. Observe that in static regions, \\(\\Sigma_{\\mathrm{coded}}^{j}(\\mathbf{x},t)\\) are independent and identically distributed (iid) binomial random variables. Thus, we can expect them to lie within a binomial confidence interval of one another. If not, we assume the pixel is dynamic and generate an event. We trigger an event if \\(\\Sigma_{\\mathrm{coded}}^{j}\\notin \\operatorname {conf}(n,\\hat{p})\\) for any \\(j\\). Here, \"conf\" refers to a binomial confidence interval (e.g., Wilson's score), \\(n = T_{\\mathrm{code}} / J\\) draws, and \\(\\hat{p} = \\sum_{s}\\Phi (\\mathbf{x},s) / T_{\\mathrm{code}}\\) is the empirical success probability." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.36, + 0.47, + 0.48 + ], + "angle": 0, + "content": "If a pixel is static, we store the sum of the \\( J \\) coded exposures, which is a long exposure, denoted by \\( \\Sigma_{\\mathrm{long}} \\). If the pixel remains static across more than one temporal chunk, we extend \\( \\Sigma_{\\mathrm{long}} \\) to include the entire duration. Whereas, if the pixel is dynamic, we transmit \\( \\{\\Sigma_{\\mathrm{coded}}^j\\} \\), as well as any previous static intensity encoded in \\( \\Sigma_{\\mathrm{long}} \\). Downstream, we can apply coded-exposure restoration techniques [54, 63, 66] to recover intensity frames from the coded measurements." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.493, + 0.273, + 0.51 + ], + "angle": 0, + "content": "5. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.519, + 0.471, + 0.595 + ], + "angle": 0, + "content": "We demonstrate the capabilities of generalized event cameras using a SwissSPAD2 array [62] with resolution \\(512 \\times 256\\), which we use to capture one-bit frames at \\(96.8\\mathrm{kHz}\\). We show the feasibility of our designs on UltraPhase [4], a recent single-photon computational platform (Sec. 5.4)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.597, + 0.471, + 0.749 + ], + "angle": 0, + "content": "Refinement model. For each of our event cameras, we train a refinement model that mitigates artifacts arising from the asynchronous nature of events. This model takes a periodic frame-based sampling of integrator values and outputs a video reconstruction. The sampling rate is configurable; in practice, we set it \\(\\sim 16 - 64\\times\\) lower than the SPAD rate. We use a densely-connected residual architecture [63], trained on data generated by simulating photon detections on temporally interpolated [24] high-speed videos from the XVFI dataset [56]. See the supplement for training details." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.757, + 0.442, + 0.774 + ], + "angle": 0, + "content": "5.1. Extreme Bandwidth-Efficient Videography" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.901 + ], + "angle": 0, + "content": "High-speed videography. In Fig. 5, we capture the dynamics of a deformable ball (a \"stress ball\") using a SPAD, a high-speed camera (Photron Infinicam) operated at 500 FPS, and a commercial event camera (Prophesee EVK4). The high-speed camera suffers from low SNR due to read noise, which manifests as prominent artifacts after on-camera compression. Meanwhile, conventional events captured by Prophesee, when processed by \"intensity-from-events\" methods such as E2VID+ [57] fail to" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.091, + 0.635, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.091, + 0.753, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.756, + 0.091, + 0.874, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.185, + 0.634, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.185, + 0.755, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.756, + 0.185, + 0.878, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.314, + 0.871, + 0.326 + ], + "angle": 0, + "content": "Spatiotemporal chunk (Sec 4.2, 431 bps/pixel, rendered at 3025 FPS)" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.332, + 0.894, + 0.43 + ], + "angle": 0, + "content": "Figure 5. High-speed videography of a stress ball hurled at a coffee mug. (top row) This indoor scene is challenging for existing imaging systems, including: high-speed cameras (SNR-related artifacts), event cameras (poor restoration quality), and even hybrid event + frame techniques (reconstruction artifacts). (bottom rows) In contrast, our generalized event cameras capture the stress ball's extensive deformations with high fidelity and an efficient readout." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.44, + 0.895, + 0.576 + ], + "angle": 0, + "content": "recover intensities reliably, especially in static regions. We also evaluate EDI [42], a hybrid event-frame method. We consider an idealized variant that operates on SPAD events (obtained via EMA thresholding), which gives perfect event-frame alignment and a precisely known event-generation model. We refine the outputs of EDI using the same model as for our methods. We refer to this idealized, refined version of EDI as \"EDI++. While EDI++ recovers more detail than other baselines, there are considerable artifacts in its outputs." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.578, + 0.894, + 0.699 + ], + "angle": 0, + "content": "Our method achieves high-quality reconstructions at 3025 FPS (96800/32) that faithfully capture non-rigid deformations, with only 431 bits per second per pixel (bps/pixel) readout, which is a \\(227\\times\\) compression (96800/431) of the raw SPAD capture. Viewed differently, for a 1 MPixel array, we would obtain a bitrate of 431 Mbps, implying that we can read off these 3025 FPS reconstructions over USB 2.0 (which supports transfer speeds of up to 480 Mbps)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.705, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Event imaging in low light. Fig. 6 compares the low-light performance of frame-based, event-based, and a generalized event camera on an urban night-time scene at 7 lux (lux measured at the sensor). For frame-based cameras, a short exposure that preserves motion may be too noisy, while a long exposure can be severely blurred. The Propheesee's performance deteriorates in low light, resulting in blurred temporal gradients. EDI++, benefiting from the idealized SPAD-based implementation, can image this scene, but finer details like the motorcyclist are lost. Our generalized event cameras, on the other hand, provide reconstructions with minimal noise, blur, or artifacts—while retaining the bandwidth efficiency of event-based systems. The compression here is \\(307\\times\\) with respect to raw SPAD outputs." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "25012" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.191, + 0.092, + 0.308, + 0.101 + ], + "angle": 0, + "content": "Frame-based cameras" + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.103, + 0.249, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.252, + 0.103, + 0.401, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.165, + 0.249, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.108, + 0.225, + 0.241, + 0.247 + ], + "angle": 0, + "content": "1 ms short exposure (120 bps/pixel at 10 FPS)" + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.164, + 0.401, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.26, + 0.225, + 0.392, + 0.247 + ], + "angle": 0, + "content": "80 ms long exposure (120 bps/pixel at 10 FPS)" + }, + { + "type": "image_caption", + "bbox": [ + 0.452, + 0.092, + 0.512, + 0.102 + ], + "angle": 0, + "content": "Events only" + }, + { + "type": "image", + "bbox": [ + 0.409, + 0.103, + 0.557, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.409, + 0.165, + 0.556, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.442, + 0.225, + 0.522, + 0.247 + ], + "angle": 0, + "content": "E2VID+ (121 bps/pixel)" + }, + { + "type": "image_caption", + "bbox": [ + 0.58, + 0.092, + 0.697, + 0.102 + ], + "angle": 0, + "content": "Event + frame imaging" + }, + { + "type": "image", + "bbox": [ + 0.565, + 0.103, + 0.713, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.566, + 0.164, + 0.712, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.599, + 0.225, + 0.678, + 0.247 + ], + "angle": 0, + "content": "EDI++ (315 bps/pixel)" + }, + { + "type": "image_caption", + "bbox": [ + 0.746, + 0.092, + 0.845, + 0.101 + ], + "angle": 0, + "content": "Generalized events" + }, + { + "type": "image", + "bbox": [ + 0.722, + 0.103, + 0.871, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.722, + 0.164, + 0.869, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.734, + 0.225, + 0.856, + 0.247 + ], + "angle": 0, + "content": "Adaptive-Bayesian (Sec 4.1, 315 bps/pixel)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.251, + 0.894, + 0.307 + ], + "angle": 0, + "content": "Figure 6. Event imaging in urban nighttime (7 lux, sensor side). (left to right) Low-light conditions necessitate long exposures in frame-based cameras, resulting in unwanted motion blur. The Propheesee EVK4 suffers from severe degradation in low light, causing E2VID+ to fail. Running EDI++ on perfectly aligned SPAD-frames and -events improves overall restoration quality but still gives failures on fast-moving objects. Our generalized events recover significantly more detail in low light, as seen in the inset of the motorcyclist." + }, + { + "type": "image_caption", + "bbox": [ + 0.274, + 0.312, + 0.55, + 0.324 + ], + "angle": 0, + "content": "Conventional events (Propheesee EVK4, 331 bps/pixel)" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.325, + 0.213, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.103, + 0.388, + 0.193, + 0.397 + ], + "angle": 0, + "content": "HRNet on E2VID+" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.325, + 0.346, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.388, + 0.314, + 0.397 + ], + "angle": 0, + "content": "Arc* corners" + }, + { + "type": "image", + "bbox": [ + 0.349, + 0.325, + 0.477, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.381, + 0.388, + 0.444, + 0.397 + ], + "angle": 0, + "content": "E-RAFT flow" + }, + { + "type": "image", + "bbox": [ + 0.481, + 0.325, + 0.608, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.501, + 0.388, + 0.587, + 0.397 + ], + "angle": 0, + "content": "DETR on E2VID+" + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.325, + 0.74, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.636, + 0.388, + 0.716, + 0.397 + ], + "angle": 0, + "content": "SAM on E2VID+" + }, + { + "type": "image_caption", + "bbox": [ + 0.76, + 0.313, + 0.869, + 0.323 + ], + "angle": 0, + "content": "85 ms long exposure" + }, + { + "type": "image", + "bbox": [ + 0.75, + 0.325, + 0.879, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.754, + 0.388, + 0.875, + 0.397 + ], + "angle": 0, + "content": "120 bps/pixel at 10 FPS" + }, + { + "type": "image_caption", + "bbox": [ + 0.243, + 0.404, + 0.583, + 0.415 + ], + "angle": 0, + "content": "Generalized events (Spatiotemporal chunk, Sec 4.2, 520 bps/pixel)" + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.416, + 0.213, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.479, + 0.18, + 0.489 + ], + "angle": 0, + "content": "HRNet pose" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.416, + 0.346, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.244, + 0.479, + 0.318, + 0.489 + ], + "angle": 0, + "content": "Harris corners" + }, + { + "type": "image", + "bbox": [ + 0.349, + 0.416, + 0.477, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.385, + 0.479, + 0.439, + 0.489 + ], + "angle": 0, + "content": "RAFT flow" + }, + { + "type": "image", + "bbox": [ + 0.481, + 0.416, + 0.608, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.504, + 0.479, + 0.585, + 0.489 + ], + "angle": 0, + "content": "DETR detection" + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.416, + 0.74, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.627, + 0.479, + 0.726, + 0.489 + ], + "angle": 0, + "content": "SAM segmentation" + }, + { + "type": "image_caption", + "bbox": [ + 0.753, + 0.404, + 0.875, + 0.415 + ], + "angle": 0, + "content": "0.3 ms burst exposures" + }, + { + "type": "image", + "bbox": [ + 0.751, + 0.416, + 0.886, + 0.483 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.773, + 0.484, + 0.857, + 0.495 + ], + "angle": 0, + "content": "15100 bps/pixel" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.497, + 0.894, + 0.553 + ], + "angle": 0, + "content": "Figure 7. Plug-and-play inference on a tennis scene. (top left) Conventional events encode temporal-gradient polarities; this lossy representation limits performance on downstream tasks. (bottom left) Generalized events encode rich scene-intensity information, with a readout comparable existing event cameras. They facilitate high-quality plug-and-play inference, without requiring dedicated algorithms. (right) Generalized event cameras give image quality comparable to burst photography techniques that have a much higher readout rate." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.565, + 0.301, + 0.581 + ], + "angle": 0, + "content": "5.2. Plug-and-Play Inference" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.594, + 0.471, + 0.759 + ], + "angle": 0, + "content": "Generalized event cameras preserve scene intensity, which enables plug-and-play event-based vision. We consider a tennis sequence (of 8196 binary frames) containing a range of object speeds. We evaluate a range of tasks: pose estimation (HRNet [59]), corner detection [18], optical flow (RAFT [61]), object detection (DETR [11]), and segmentation (SAM [32]). We compare against event-based methods applied to Prophesee events; we use \\(\\mathrm{Arc}^*\\) [3] for corner detection and E-RAFT [16] for optical flow. For the remaining tasks, which do not have equivalent event methods, we run HRNet, DETR, and SAM on E2VID+ reconstructions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.473, + 0.902 + ], + "angle": 0, + "content": "As Fig. 7 (top left) shows, traditional events are bandwidth efficient (331 bps/pixel), but do not provide sufficient information for successful inference. Generalized events (bottom left) have a modestly higher readout (520 bps/pixel), but support accurate inference without requiring dedicated algorithms. To provide context for these rates, we compare them against frame-based methods (right). A long exposure (120 bps/pixel) blurs out the ratchet. Burst methods [19] recover a sharp image from a stack of short exposures, but with a large readout of 15100 bps/pixel." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.565, + 0.726, + 0.581 + ], + "angle": 0, + "content": "5.3. Rate-Distortion Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.592, + 0.894, + 0.773 + ], + "angle": 0, + "content": "Each method in Sec. 4 features a sensitivity parameter that controls the sensor readout rate (event rate), which in turn influences image quality. In this subsection, we evaluate the impact of readout on image quality (PSNR) by performing a rate-distortion analysis. For ground truth, we use a set of YouTube-sourced high-speed videos captured by a Phantom Flex4k at 1000 FPS; see the supplement for thumbnails and links. We upsample these videos to the SPAD's frame rate and then simulate 4096 binary frames using the image formation model described in Eq. (4). When computing readout for our methods, we assume that events encode 10-bit values and account for the header bits of each event packet." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.777, + 0.894, + 0.881 + ], + "angle": 0, + "content": "As baselines, we consider EDI++, a long exposure, compressive sensing with 8-bucket masks, and burst denoising [19] using 32 short exposures. As Fig. 8 shows, generalized event cameras provide a pronounced 4-8 dB PSNR improvement over baseline methods. Further, our methods can compress the raw SPAD response by around \\(80\\times\\) before a noticeable drop-off in PSNR is observed." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Among our methods, the spatiotemporal chunk approach of" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25013" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.46, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.209, + 0.471, + 0.265 + ], + "angle": 0, + "content": "Figure 8. Rate-distortion evaluation. Our techniques feature a tunable parameter that controls the output event rate. Generalized events offer a \\(4 - 8\\mathrm{dB}\\) improvement in PSNR over \\(\\mathrm{EDI} + +\\) (at the same readout), and can compress raw photon data by \\(80\\times\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.274, + 0.474, + 0.487 + ], + "angle": 0, + "content": "Sec. 4.2 gives the best PSNR, followed by the Bayesian method (Sec. 4.1) and coded-exposure events (Sec. 4.3). That said, all methods are fairly similar in terms of rate-distortion (e.g., all three give comparable results for the scenes in Secs. 5.1 and 5.2). The methods are better distinguished by their practical characteristics. The Bayesian method gives single-photon temporal resolution; however, as we show in Sec. 5.4, it is the most expensive to compute on-chip. The chunk-based method occupies a middle ground in terms of latency and cost. Coded-exposure events have the highest latency—events are generated only every \\(\\sim 256 - 512\\) binary frames—but the lowest on-chip cost. This provides an end user the flexibility to choose from the space of generalized event cameras based on the latency requirements and the compute constraints of the target application." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.498, + 0.383, + 0.515 + ], + "angle": 0, + "content": "5.4. On-Chip Feasibility and Validation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.47, + 0.643 + ], + "angle": 0, + "content": "A critical limitation of single-photon sensors is the exorbitant bandwidth and power costs involved in reading off raw photon detections. However, the lightweight nature of our event camera designs allows us to sidestep this limitation by performing computations on-chip. We demonstrate that our methods are feasible on UltraPhase [4], a SPAD compute platform. UltraPhase consists of \\(3 \\times 6\\) compute cores, each of which is associated with \\(4 \\times 4\\) pixels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.471, + 0.763 + ], + "angle": 0, + "content": "We implement our methods for UltraPhase using custom assembly code. Some methods require minor modifications due to instruction-set limitations; see the supplement for details. We process 2500 SPAD frames from the tennis sequence used in Sec. 5.2, cropped to the UltraPhase array size of \\(12 \\times 24\\) pixels. We determine the number of cycles required to execute the assembly code and estimate the chip's power consumption and readout bandwidth." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "All our proposed methods run comfortably within the chip's compute budget of 4202 cycles per binary frame and its memory limit of 4 Kbit per core. As seen in Fig. 9, compared to raw photon-detection readout, our techniques reduce both bandwidth and power costs by over two orders of magnitude. The coded-exposure method is particularly efficient; on most binary frames, it only requires multiplying a binary code with incident photon detections. Our proof-of-concept evaluation may pave the way for future near-sensor implementations of generalized event" + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.091, + 0.878, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.191, + 0.893, + 0.248 + ], + "angle": 0, + "content": "Figure 9. On-chip compatibility. We validate the feasibility of our approach on UltraPhase [4], a computational SPAD imager. Compared to reading out raw photon data, all of our approaches give marked reductions in both bandwidth (left) and power (right)." + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.257, + 0.63, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.318, + 0.621, + 0.34 + ], + "angle": 0, + "content": "Rigid dynamics (200x compression)" + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.258, + 0.757, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.647, + 0.318, + 0.744, + 0.34 + ], + "angle": 0, + "content": "Phone screen (11x compression)" + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.258, + 0.883, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.77, + 0.318, + 0.873, + 0.34 + ], + "angle": 0, + "content": "Camera motion (130x compression)" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.345, + 0.894, + 0.429 + ], + "angle": 0, + "content": "Figure 10. Limitations and failure modes, (left) Our reconstructions (yellow inset) on dynamic scenes with rigid objects can be inferior to burst photography (green inset). (center) Modulated light sources, such as this phone screen, can trigger a deluge of events (change points shown in the inset). (right) Rapid camera motion can result in an event rate divergent from scene dynamics." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.437, + 0.894, + 0.483 + ], + "angle": 0, + "content": "cameras, which with advances in chip-to-chip communication, could involve a dedicated \"photon processing unit\", similar to a camera image signal processor (ISP)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.496, + 0.738, + 0.511 + ], + "angle": 0, + "content": "6. Limitations and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.521, + 0.894, + 0.657 + ], + "angle": 0, + "content": "Generalized events push the frontiers of event-based imaging; however, some scenarios lead to sub-optimal performance. As seen in Fig. 10 (left), if the scene dynamics is entirely comprised of rigid motion, burst photography [36] gives better image quality, albeit with much higher readout. (middle) Similar to current event cameras, modulated light sources trigger unwanted events that reduce bandwidth savings. However, it may be possible to ignore some of these events, perhaps by modeling the lighting variations [55]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.661, + 0.894, + 0.782 + ], + "angle": 0, + "content": "Ego-motion events. Camera motion can trigger events in static regions, although our methods still yield substantial compression \\((130 \\times\\) over SPAD outputs, Fig. 10 right). We analyze the impact of ego-motion on bandwidth savings further in the supplement. However, single-photon cameras can emulate sensor motion by integrating flux along alternate spatiotemporal trajectories [60]. We can imagine a generalized event camera that is \"ego-motion compensated,\" by computing events along a suitable trajectory." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.786, + 0.894, + 0.891 + ], + "angle": 0, + "content": "Photon-stream compression. SPADs generate a torrent of data—e.g., 12.5 GBps for a MPixel array at \\(100\\mathrm{kHz}\\)—that can easily overwhelm data interfaces. Generalized event cameras reduce readout by around two orders of magnitude, by decoupling readout from the SPAD's frame rate and instead basing it on scene dynamics. This could pave the way for practical, high-resolution single-photon sensors." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25014" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.17, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.143 + ], + "angle": 0, + "content": "[1] Ryan Prescott Adams and David JC MacKay. Bayesian online changepoint detection. arXiv preprint arXiv:0710.3742, 2007. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.143, + 0.47, + 0.197 + ], + "angle": 0, + "content": "[2] Reda Alami, Odalric Maillard, and Raphael Feraud. Restarted bayesian online change-point detector achieves optimal detection delay. In International conference on machine learning, pages 211-221. PMLR, 2020. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.199, + 0.471, + 0.239 + ], + "angle": 0, + "content": "[3] Ignacio Alzugaray and Margarita Chli. Asynchronous corner detection and tracking for event cameras in real time. IEEE Robotics and Automation Letters, 3(4):3177-3184, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.24, + 0.47, + 0.268 + ], + "angle": 0, + "content": "[4] Andrei Ardelean. Computational Imaging SPAD Cameras. PhD thesis, École polytechnique fédérale de Lausanne, 2023. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.268, + 0.471, + 0.309 + ], + "angle": 0, + "content": "[5] Souptik Barua, Yoshitaka Miyatani, and Ashok Veeraraghavan. Direct face detection and video reconstruction from event cameras. In WACV, pages 1-9. IEEE, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.31, + 0.471, + 0.377 + ], + "angle": 0, + "content": "[6] Raphael Berner, Christian Brandli, Minhao Yang, S-C Liu, and Tobi Delbruck. A 240x180 120dB 10mW 12us-latency sparse output vision sensor for mobile applications. In Proceedings of the International Image Sensors Workshop, number CONF, pages 41-44, 2013. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.379, + 0.471, + 0.447 + ], + "angle": 0, + "content": "[7] Anthony Bisulco, Fernando Cladera Ojeda, Volkan Isler, and Daniel D Lee. Fast motion understanding with spatiotemporal neural networks and dynamic vision sensors. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 14098-14104. IEEE, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.448, + 0.471, + 0.489 + ], + "angle": 0, + "content": "[8] Assim Boukhayma, Arnaud Peizerat, and Christian Enz. A sub-0.5 electron read noise VGA image sensor in a standard CMOS process. IEEE Journal of Solid-State Circuits, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.49, + 0.471, + 0.544 + ], + "angle": 0, + "content": "[9] Christian Brandli, Raphael Berner, Minhao Yang, Shih-Chii Liu, and Tobi Delbruck. A \\(240 \\times 180\\) dB \\(3\\mu s\\) latency global shutter spatiotemporal vision sensor. IEEE Journal of Solid-State Circuits, 49(10):2333-2341, 2014. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.545, + 0.471, + 0.599 + ], + "angle": 0, + "content": "[10] Christian Brandli, Lorenz Muller, and Tobi Delbruck. Realtime, high-speed video decompression using a frame-and event-based DAVIS sensor. In 2014 IEEE International Symposium on Circuits and Systems (ISCAS), pages 686-689. IEEE, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.6, + 0.471, + 0.669 + ], + "angle": 0, + "content": "[11] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Computer Vision – ECCV 2020, pages 213–229, Cham, 2020. Springer International Publishing. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.669, + 0.471, + 0.71 + ], + "angle": 0, + "content": "[12] Andrea Censi and Davide Scaramuzza. Low-latency event-based visual odometry. In 2014 IEEE International Conference on Robotics and Automation (ICRA), pages 703-710. IEEE, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.711, + 0.471, + 0.765 + ], + "angle": 0, + "content": "[13] Matthew Cook, Luca Gugelmann, Florian Jug, Christoph Krautz, and Angelika Steger. Interacting maps for fast visual interpretation. In The 2011 International Joint Conference on Neural Networks, pages 770-776, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.766, + 0.471, + 0.807 + ], + "angle": 0, + "content": "[14] Yue Gao, Siqi Li, Yipeng Li, Yandong Guo, and Qionghai Dai. Superfast: \\(200 \\times\\) video frame interpolation via event camera. IEEE TPAMI, 45(6):7764-7780, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.808, + 0.471, + 0.849 + ], + "angle": 0, + "content": "[15] Daniel Gehrig, Henri Rebecq, Guillermo Gallego, and Davide Scaramuzza. EKLT: Asynchronous photometric feature tracking using events and frames. IJCV, 128(3):601-618, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.849, + 0.471, + 0.891 + ], + "angle": 0, + "content": "[16] Mathias Gehrig, Mario Millhäuser, Daniel Gehrig, and Davide Scaramuzza. E-raft: Dense optical flow from event cameras. In 2021 International Conference on 3D Vision (3DV), pages" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.891 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.672, + 0.106 + ], + "angle": 0, + "content": "197-206.IEEE,2021.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.107, + 0.894, + 0.147 + ], + "angle": 0, + "content": "[17] Istvan Gyongy, Neale AW Dutton, and Robert K Henderson. Single-photon tracking for high-speed vision. Sensors, 18(2):323, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.149, + 0.894, + 0.188 + ], + "angle": 0, + "content": "[18] Chris Harris, Mike Stephens, et al. A combined corner and edge detector. In *Alvey vision conference*, pages 10-5244. CiteSeer, 1988. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.19, + 0.894, + 0.244 + ], + "angle": 0, + "content": "[19] Samuel W. Hasinoff, Dillon Sharlet, Ryan Geiss, Andrew Adams, Jonathan T. Barron, Florian Kainz, Jiawen Chen, and Marc Levoy. Burst photography for high dynamic range and low-light imaging on mobile cameras. ACM TOG, 35(6):1-12, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.245, + 0.894, + 0.314 + ], + "angle": 0, + "content": "[20] Botao He, Haojia Li, Siyuan Wu, Dong Wang, Zhiwei Zhang, Qianli Dong, Chao Xu, and Fei Gao. Fast-dynamic-vision: Detection and tracking dynamic objects with event and depth sensing. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3071-3078. IEEE, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.315, + 0.894, + 0.355 + ], + "angle": 0, + "content": "[21] Javier Hidalgo-Carrio, Guillermo Gallego, and Davide Scaramuzza. Event-aided direct sparse odometry. In CVPR, pages 5781-5790, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.356, + 0.894, + 0.423 + ], + "angle": 0, + "content": "[22] Yasunobu Hitomi, Jinwei Gu, Mohit Gupta, Tomoo Mitsunaga, and Shree K. Nayar. Video from a single coded exposure photograph using a learned over-complete dictionary. In 2011 International Conference on Computer Vision, pages 287-294, 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.424, + 0.894, + 0.481 + ], + "angle": 0, + "content": "[23] Jing Huang, Menghan Guo, and Shoushun Chen. A dynamic vision sensor with direct logarithmic output and full-frame picture-on-demand. In 2017 IEEE International Symposium on Circuits and Systems (ISCAS), pages 1-4, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.481, + 0.893, + 0.535 + ], + "angle": 0, + "content": "[24] Zhewei Huang, Tianyuan Zhang, Wen Heng, Boxin Shi, and Shuchang Zhou. Real-time intermediate flow estimation for video frame interpolation. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.536, + 0.893, + 0.576 + ], + "angle": 0, + "content": "[25] Jorge Igual. Photographic noise performance measures based on raw files analysis of consumer cameras. *Electronics*, 8(11):1284, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.577, + 0.892, + 0.605 + ], + "angle": 0, + "content": "[26] Atul Ingle, Andreas Velten, and Mohit Gupta. High Flux Passive Imaging With Single-Photon Sensors. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.606, + 0.894, + 0.646 + ], + "angle": 0, + "content": "[27] Atul Ingle, Trevor Seets, Mauro Buttafava, Shantanu Gupta, Alberto Tosi, Mohit Gupta, and Andreas Velten. Passive inter-photon imaging. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.647, + 0.894, + 0.7 + ], + "angle": 0, + "content": "[28] Kiyotaka Iwabuchi, Yusuke Kameda, and Takayuki Hamamoto. Image quality improvements based on motion-based deblurring for single-photon imaging. IEEE Access, 9:30080-30094, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.702, + 0.893, + 0.742 + ], + "angle": 0, + "content": "[29] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. CoTracker: It is better to track together. 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.743, + 0.893, + 0.771 + ], + "angle": 0, + "content": "[30] Anthony Kay. Tesseract: an open-source optical character recognition engine. *Linux Journal*, 2007(159):2, 2007." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.772, + 0.893, + 0.812 + ], + "angle": 0, + "content": "[31] Hanme Kim, Ankur Handa, Ryad Benosman, Sio-Hoi Ieng, and Andrew J Davison. Simultaneous mosaicing and tracking with an event camera. Proc. Brit. Mach. Vis. Conf, 43:566-576, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.813, + 0.893, + 0.868 + ], + "angle": 0, + "content": "[32] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólár, and Ross Girshick. Segment anything, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.869, + 0.893, + 0.896 + ], + "angle": 0, + "content": "[33] Haolong Li and Joerg Stueckler. Tracking 6-DoF object motion from events and frames. In 2021 IEEE International Conference" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.896 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25015" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.471, + 0.12 + ], + "angle": 0, + "content": "on Robotics and Automation (ICRA), pages 14171-14177. IEEE, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.472, + 0.16 + ], + "angle": 0, + "content": "[34] P Lichtsteiner. 64x64 event-driven logarithmic temporal derivative silicon retina. In Program 2003 IEEE Workshop on CCD and AIS, 2003. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.472, + 0.203 + ], + "angle": 0, + "content": "[35] Yuhao Liu, Felipe Gutierrez-Barragan, Atul Ingle, Mohit Gupta, and Andreas Velten. Single-photon camera guided extreme dynamic range imaging. In WACV, pages 1575-1585, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.472, + 0.244 + ], + "angle": 0, + "content": "[36] Sizhuo Ma, Shantanu Gupta, Arin C. Ulku, Claudio Bruschini, Edoardo Charbon, and Mohit Gupta. Quanta burst photography. ACM TOG, 39(4):1-16, 2020. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.246, + 0.472, + 0.285 + ], + "angle": 0, + "content": "[37] Sizhuo Ma, Paul Mos, Edoardo Charbon, and Mohit Gupta. Burst vision using single-photon cameras. In WACV, pages 5375-5385, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.287, + 0.472, + 0.342 + ], + "angle": 0, + "content": "[38] Kazuhiro Morimoto, Andrei Ardelean, Ming-Lo Wu, Arin Can Ulku, Ivan Michel Antolovic, Claudio Bruschini, and Edoardo Charbon. Megapixel time-gated SPAD image sensor for 2D and 3D imaging applications. Optica, 7(4):346-354, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.343, + 0.472, + 0.383 + ], + "angle": 0, + "content": "[39] Gottfried Munda, Christian Reinbacher, and Thomas Pock. Realtime intensity-image reconstruction for event cameras using manifold regularisation. *IJCV*, 126:1381–1393, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.384, + 0.472, + 0.451 + ], + "angle": 0, + "content": "[40] Shuichi Namiki, Shunichi Sato, Yusuke Kameda, and Takayuki Hamamoto. Imaging method using multi-threshold pattern for photon detection of quanta image sensor. In International Workshop on Advanced Imaging Technology (IWAIT) 2022, page 1217702. SPIE, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.453, + 0.472, + 0.493 + ], + "angle": 0, + "content": "[41] Liyuan Pan, Cedric Scheerlinck, Xin Yu, Richard Hartley, Miao-miao Liu, and Yuchao Dai. Bringing a blurry frame alive at high frame-rate with an event camera. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.495, + 0.472, + 0.548 + ], + "angle": 0, + "content": "[42] Liyuan Pan, Richard Hartley, Cedric Scheerlinck, Miaomiao Liu, Xin Yu, and Yuchao Dai. High frame rate video reconstruction based on an event camera. IEEE TPAMI, 44(5):2519-2533, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.55, + 0.472, + 0.604 + ], + "angle": 0, + "content": "[43] Federico Paredes-Valles and Guido C. H. E. de Croon. Back to event basics: Self-supervised learning of image reconstruction for event cameras via photometric constancy. In CVPR, pages 3446–3455, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.606, + 0.472, + 0.659 + ], + "angle": 0, + "content": "[44] Christoph Posch, Daniel Matolin, and Rainer Wohlgenannt. A QVGA 143 dB dynamic range frame-free pwm image sensor with lossless pixel-level video compression and time-domain cds. IEEE Journal of Solid-State Circuits, 46(1):259-275, 2011. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.472, + 0.702 + ], + "angle": 0, + "content": "[45] Ramesh Raskar, Amit Agrawal, and Jack Tumblin. Coded exposure photography: motion deblurring using fluttered shutter. In Acm Siggraph 2006 Papers, pages 795-804. 2006. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.703, + 0.472, + 0.742 + ], + "angle": 0, + "content": "[46] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. High speed and high dynamic range video with an event camera. IEEE TPAMI, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.744, + 0.472, + 0.785 + ], + "angle": 0, + "content": "[47] Dikpal Reddy, Ashok Veeraraghavan, and Rama Chellappa. P2C2: Programmable pixel compressive camera for high speed imaging. In CVPR 2011, pages 329-336, 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.786, + 0.472, + 0.812 + ], + "angle": 0, + "content": "[48] Alexis Rochas. Single photon avalanche diodes in cmos technology. Technical report, Citeseer, 2003. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.813, + 0.472, + 0.895 + ], + "angle": 0, + "content": "[49] Nitin J Sanket, Chethan M Parameshwara, Chahat Deep Singh, Ashwin V Kuruttukulam, Cornelia Fermüller, Davide Scaramuzza, and Yiannis Aloimonos. Evdodgenet: Deep dynamic obstacle dodging with event cameras. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 10651-10657. IEEE, 2020. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.895 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.135 + ], + "angle": 0, + "content": "[50] Cedric Scheerlinck, Nick Barnes, and Robert Mahony. Continuous-time intensity estimation using event cameras. In ACCV, pages 308-324. Springer, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.894, + 0.175 + ], + "angle": 0, + "content": "[51] Cedric Scheerlinck, Henri Rebecq, Daniel Gehrig, Nick Barnes, Robert Mahony, and Davide Scaramuzza. Fast image reconstruction with an event camera. In WACV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.177, + 0.894, + 0.217 + ], + "angle": 0, + "content": "[52] Trevor Seets, Atul Ingle, Martin Laurenzis, and Andreas Velten. Motion adaptive deblurring with single-photon cameras. In WACV, pages 1945-1954, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.218, + 0.894, + 0.259 + ], + "angle": 0, + "content": "[53] Prasan Shedligeri and Kaushik Mitra. Photorealistic image reconstruction from hybrid intensity and event-based sensor. Journal of Electronic Imaging, 28(6):063012-063012, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.26, + 0.893, + 0.3 + ], + "angle": 0, + "content": "[54] Prasan Shedligeri, Anupama S, and Kaushik Mitra. A unified framework for compressive video recovery from coded exposure techniques. In WACV, pages 1600-1609, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.301, + 0.893, + 0.328 + ], + "angle": 0, + "content": "[55] Mark Sheinin, Yoav Y. Schechner, and Kiriakos N. Kutulakos. Computational imaging on the electric grid. In CVPR, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.329, + 0.893, + 0.383 + ], + "angle": 0, + "content": "[56] Hyeunjun Sim, Jihyong Oh, and Munchurl Kim. XVFI: extreme video frame interpolation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14489-14498, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.384, + 0.893, + 0.438 + ], + "angle": 0, + "content": "[57] Timo Stoffregen, Cedric Scheerlinck, Davide Scaramuzza, Tom Drummond, Nick Barnes, Lindsay Kleeman, and Robert Mahony. Reducing the sim-to-real gap for event cameras. In ECCV, pages 534-549. Springer, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.439, + 0.893, + 0.48 + ], + "angle": 0, + "content": "[58] Binyi su, Lei Yu, and Wen Yang. Event-based high frame-rate video reconstruction with a novel cycle-event network. In ICIP, pages 86-90, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.481, + 0.893, + 0.535 + ], + "angle": 0, + "content": "[59] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.536, + 0.893, + 0.589 + ], + "angle": 0, + "content": "[60] Varun Sundar, Andrei Ardelean, Tristan Swedish, Claudio Bruschini, Edoardo Charbon, and Mohit Gupta. Sodacam: Software-defined cameras via single-photon imaging. In ICCV, pages 8165-8176, 2023. 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.591, + 0.893, + 0.632 + ], + "angle": 0, + "content": "[61] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision – ECCV 2020, pages 402–419, Cham, 2020. Springer International Publishing. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.633, + 0.893, + 0.702 + ], + "angle": 0, + "content": "[62] Arin Can Ulku, Claudio Bruschini, Ivan Michel Antolovic, Yung Kuo, Rinet Ankri, Shimon Weiss, Xavier Michalet, and Edoardo Charbon. A \\(512 \\times 512\\) SPAD Image Sensor With Integrated Gating for Widefield FLIM. IEEE Journal of Selected Topics in Quantum Electronics, 25(1):1-12, 2019. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.703, + 0.893, + 0.756 + ], + "angle": 0, + "content": "[63] Lishun Wang, Miao Cao, and Xin Yuan. Efficientsci: Densely connected network with space-time factorization for large-scale video snapshot compressive imaging. In CVPR, pages 18477-18486, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.758, + 0.893, + 0.811 + ], + "angle": 0, + "content": "[64] Zhaohui Wang, Xiao Lin, Abhinav Mishra, and Ram Sriharsha. Online changepoint detection on a budget. In 2021 International Conference on Data Mining Workshops (ICDMW), pages 414-420. IEEE, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.812, + 0.893, + 0.854 + ], + "angle": 0, + "content": "[65] Mian Wei, Sotiris Nousias, Rahul Gulve, David B. Lindell, and Kiriakos N. Kutulakos. Passive ultra-wideband single-photon imaging. In ICCV, pages 8135-8146, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.855, + 0.893, + 0.895 + ], + "angle": 0, + "content": "[66] Xin Yuan, Yang Liu, Jinli Suo, Frédo Durand, and Qionghai Dai. Plug-and-play algorithms for video snapshot compressive imaging. IEEE Transactions on Pattern Analysis and Machine" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.895 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "25016" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.341, + 0.106 + ], + "angle": 0, + "content": "Intelligence, 44(10):7093-7111, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.107, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[67] Zelin Zhang, Anthony J Yezzi, and Guillermo Gallego. Formulating event-based image reconstruction as a linear inverse problem with deep regularization using optical flow. IEEE TPAMI, 45 (07):8372-8389, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.472, + 0.19 + ], + "angle": 0, + "content": "[68] Alex Zihao Zhu, Nikolay Atanasov, and Kostas Daniilidis. Event-based visual inertial odometry. In CVPR, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.19, + 0.471, + 0.232 + ], + "angle": 0, + "content": "[69] Yunhao Zou, Yinqiang Zheng, Tsuyoshi Takatani, and Ying Fu. Learning to reconstruct high speed and high dynamic range videos from events. In CVPR, pages 2024-2033, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.107, + 0.472, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25017" + } + ] +] \ No newline at end of file diff --git a/2024/Generalized Event Cameras/e213eaf6-87b7-49bd-b232-571683e80e75_origin.pdf b/2024/Generalized Event Cameras/e213eaf6-87b7-49bd-b232-571683e80e75_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7dbdcac7e59f26ccd0c7f26c1403168ac1ddbf2e --- /dev/null +++ b/2024/Generalized Event Cameras/e213eaf6-87b7-49bd-b232-571683e80e75_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e51e02377ea9e5dc213c83247ca94d0c2331ccda6e376c072a29ca6b5680da68 +size 7800526 diff --git a/2024/Generalized Event Cameras/full.md b/2024/Generalized Event Cameras/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2ff0412e2c7ae93b456c78702afbc8d2c9d554ff --- /dev/null +++ b/2024/Generalized Event Cameras/full.md @@ -0,0 +1,488 @@ +# Generalized Event Cameras + +Varun Sundar†* + +vsundar4@wisc.edu + +Matthew Dutson†* + +dutson@wisc.edu + +Andrei Ardelean† + +a.ardelean@epfl.ch + +Claudio Bruschini $^{\S}$ + +{claudio.bruschini, edoardo.charbon}@epfl.ch + +Mohit Gupta† + +mohitg@cs.wisc.edu + +†University of Wisconsin-Madison ‡NovoViz École Polytechnique Fédérale de Lausanne + +wisionlab.com/project/generalized-event-cameras/ + +# Abstract + +Event cameras capture the world at high time resolution and with minimal bandwidth requirements. However, event streams, which only encode changes in brightness, do not contain sufficient scene information to support a wide variety of downstream tasks. In this work, we design generalized event cameras that inherently preserve scene intensity in a bandwidth-efficient manner. We generalize event cameras in terms of when an event is generated and what information is transmitted. To implement our designs, we turn to single-photon sensors that provide digital access to individual photon detections; this modality gives us the flexibility to realize a rich space of generalized event cameras. Our single-photon event cameras are capable of high-speed, high-fidelity imaging at low readout rates. Consequently, these event cameras can support plug-and-play downstream inference, without capturing new event datasets or designing specialized event-vision models. As a practical implication, our designs, which involve lightweight and near-sensor-compatible computations, provide a way to use single-photon sensors without exorbitant bandwidth costs. + +# 1. Introduction + +Event cameras [6, 34, 44] sense the world at high speeds, providing visual information with minimal bandwidth and power. They achieve this by transmitting only changes in scene brightness, when significant "events" occur. However, there is a cost. Raw event data, a sparse stream of binary values, does not hold sufficient information to be used directly with mainstream vision algorithms. Therefore, while event cameras have been successful at certain tasks (e.g., object tracking [15, 33], obstacle avoidance [7, 20, 49], and high-speed odometry [12, 21, 68]), they are not widely deployed as general-purpose vision sen + +sors, and often need to be supplemented with conventional cameras [14, 15, 21]. These limitations are holding back this otherwise powerful technology. + +Is it possible to realize the promise of event cameras, i.e., high temporal resolution at low bandwidth, while preserving rich scene intensity information? To realize these seemingly conflicting goals, we propose a novel family of generalized event cameras. We conceptualize a space of event cameras along two key axes (Fig. 1 (top)): (a) "when to transmit information," formalized as a change detection procedure $\Delta$ ; and (b) "what information to transmit," characterized by an integrator $\Sigma$ that encodes incident flux. Existing event cameras represent one operating point in this $(\Sigma, \Delta)$ space. Our key observation is that by exploring this space and considering new $(\Sigma, \Delta)$ combinations, we can design event cameras that preserve scene intensity. We propose more general integrators, e.g., that represent flux according to motion levels, that span spatial patches, or that employ temporal coding (Fig. 1 (middle)). We also introduce robust change detectors that better distinguish motion from noise, by considering increased spatiotemporal contexts and modeling noise in the sensor measurements. + +Despite their conceptual appeal, physically implementing generalized event cameras is a challenge. This is because the requisite computations must be performed at the sensor to achieve the desired bandwidth reductions. For example, existing event cameras perform simple integration and thresholding operations via analog in-pixel circuitry. However, more general $(\Sigma ,\Delta)$ combinations are not always amenable to analog implementations; even feasible designs might require years of hardware iteration and production scaling. To build physical realizations of generalized event cameras, we leverage an emerging sensor technology: single-photon avalanche diodes (SPADs) that provide digital access to photon detections at extremely high frame rates $(\sim 100\mathrm{kHz})$ . This allows us to compose arbitrary software-level signal transformations, such as those required by generalized event cameras. Further, we are not locked to a particular event camera design and can realize multiple configurations + +# Conventional event camera + +![](images/eb323521c6dfc72ea4b63c3af2a9ca83b4aedb0ab5742ac8e2be0aed01ad0c1c.jpg) +Integrator (accumulates flux, $\Sigma$ ) +Low-pass filtered brightness + +![](images/cb93cdc2af2460c9e6a311222e17c6b042980154ebefcdccaf0c3666cda7f75f.jpg) +Change detector (triggers events, $\Delta$ +Threshold-based detector Analog comparator + +![](images/3a97b6196287fcae231ef2b60dc0a55be02d6d8d07bb57211323ba839ee4c0c4.jpg) +Sensor output +Stream of change polarities $(x,y,t,\pm 1)$ + +![](images/8327e0f3b32b651a89e7934c762422219b2e46eef0dedcaca0058a77b8fca86b.jpg) +Intensity imaging +Reconstruction with E2VID+ 1000 FPS at 20 FPS bitrate + +# Generalized event camera + +![](images/23f46768133bc5f79c9122299515b4f26b23a2361d6b17825fa94dc5caf2821c.jpg) +Motion adaptive +Spatial patches +Coded exposures + +![](images/a39965934d4a9cfeb632084a849d16c3cc0970fd6aaa7eecd4191a33803b531f.jpg) +Enhanced spatio-temporal support Noise-aware thresholding + +![](images/91ef207541236e1bb5e060fc371fdc3e704e5ec4fe7d6b1fb2c3583f64a2450d.jpg) +Stream of integrator values $(x,y,t,\Sigma (x,y,t))$ + +![](images/4ca0635e9f44a5a0f1241e046abe02096fe14bcc7dd3a3729e8621b868f11060.jpg) +Reconstruction 3025 FPS at 30 FPS bitrate + +# Plug-and-play event vision + +![](images/2133d106cbe4c4917aeb158ba7c31d382cc7d2ca075c17a4d2995251bfb4a23d.jpg) +Contact point of a tennis shot Image Reconstruction +Figure 1. Generalized event cameras. (top) Event cameras generate outputs in response to abrupt changes in scene intensity. We describe this as a combination of a low-pass integrator and a threshold-based change detector. (middle) We generalize the space of event cameras by designing integrators that capture rich intensity information, and more reliable change detectors that utilize larger spatiotemporal contexts and noise-aware thresholding (Secs. 4.1 to 4.3). Unlike existing events, our generalized event streams inherently preserve scene intensity, e.g., this ping-pong ball slingshotted against a brick wall backdrop. (bottom) Generalized event cameras enable high-fidelity bandwidth-efficient imaging: providing 3025 FPS reconstructions with a readout equivalent to a 30 FPS camera. Consequently, generalized events facilitate plug-and-play inference on a multitude of tasks in challenging scenarios (insets depict the extent of motion over $30~\mathrm{ms}$ ). + +![](images/b7c0e23152bc6b87266d6c12e39abceea82b4a8814836b4b1e1fbfaa0991e3f9.jpg) +Tracking points at 0.3 lux CoTracker [29] + +![](images/49f6658d4e310202bac994ccd084296f5c5bcbf1d185c35a77e4d318df22540b.jpg) +Nighttime object detection DETR [11] + +![](images/687e4267a552b02ea9648304fc4ea45c5331f352efe661270b4f3c9c93e0fc23.jpg) +OCR on a snapping tape Tesseract [30] + +![](images/a3d46eedb599bc1167cd8f6a2f37ae73379ac2a6e2344e9caaa8c4d9f94724e9.jpg) +Segmenting a dart in midair Segment Anything [32] + +with the same sensor. + +Implications: extreme, bandwidth-efficient vision. Generalized event cameras support high-speed, high-quality image reconstruction, but at low bandwidths quintessential of current event cameras. For example, Fig. 1 (middle, bottom) shows reconstructions at 3025 FPS that have an effective readout of a 30 FPS frame-based camera. Further, our methods have strong low-light performance due to the SPAD's single-photon sensitivity. As we show in Fig. 1 (bottom), preserving scene intensity facilitates plug-and-play inference in challenging scenarios, with state-of-the-art vision algorithms. Critically, this does not require retraining vision models or curating dedicated datasets, which is a significant challenge for unconventional imagers. This plug-and-play capability is vital to realizing universal event vision that retains the benefits of current event cameras. + +Scope. We consider full-stack event perception: we conceptualize a novel space of event cameras, provide relevant single-photon algorithms, analyze their imaging capabilities and rate-distortion trade-offs, and show on-chip feasibility. We demon- + +strate imaging capabilities in Secs. 5.1 and 5.2 using the SwissS-PAD2 array [62], and show viable implementations of our algorithms for UltraPhase [4], a recent single-photon compute platform. All of these are critical to unlocking the promise of event cameras. However, our objective is not to develop an integrated system that incorporates all these components; this paper merely takes the first steps toward that goal. + +# 2. Related Work + +Event camera designs. Perhaps most widespread is the DVS event camera [34], where each pixel generates an event in response to measured changes in (log) intensity. The DAVIS event camera [6, 9] couples DVS pixels with conventional CMOS pixels, providing access to image frames. However, the frames lack the dynamic range of DVS events. A recent design, CelexV [23], provides log-intensity frames using an external trigger. ATIS [44], a less prevalent design, features asynchronous intensity events, but its sophisticated circuitry reduces pixel fill factor. The above designs are based on analog processing; we instead + +design event cameras on digital photon detections. + +Intensity imaging with event cameras. Several approaches have been explored to obtain images from events, including Poisson solvers [5], manifold regularization [39], assuming knowledge of camera motion [13, 31] or optical flow [67], and learning-based methods [43, 46, 51, 58, 69]. However, because events often lack sufficient scene information, they are often supplemented by conventional frames [10, 41, 50, 53], either from sensors such as DAVIS or using a multi-camera setup. Fusing events and frames presents challenges due to potential spatiotemporal misalignment and discrepancies in imaging modalities. Even when these challenges are overcome, we show that fusion methods produce lower fidelity than our proposed generalized event cameras. + +Passive single-photon imaging. In the past few years, SPADs have found compelling passive imaging applications; this includes high-dynamic range imaging [26, 27, 35, 40], motion deblurring [28, 36, 37, 52], high-speed tracking [17], and ultra wide-band videography [65]. The fine granularity of passive single-photon acquisition makes it possible to emulate a diverse set of imaging modalities [60], including event cameras, via post-capture processing. In this work, we go beyond emulating existing event cameras and design alternate event cameras that preserve high-fidelity intensity information. + +# 3. What is an Event Camera? + +The defining characteristic of event cameras is that they transmit information selectively, in response to changes in scene content. This selectivity allows event cameras to encode scene information at high time resolutions required to capture scene dynamics, without proportionately high bandwidth. This is in contrast to frame-based cameras, where readout occurs at fixed intervals. + +We characterize event cameras in terms of two axes: what the camera transmits and when it transmits. As a concrete example, consider existing event cameras. They trigger events ("when to transmit") based on a fixed threshold: + +$$ +\left| \Phi (\mathbf {x}, t) - \Phi_ {\text {r e f}} (\mathbf {x}) \right| \geq \tau , \tag {1} +$$ + +where $\Phi (\mathbf{x},t)$ is a flux estimate at pixel $\mathbf{x}$ and time $t$ , and $\tau$ is the threshold. $\Phi_{\mathrm{ref}}(\mathbf{x})$ is a previously-recorded reference, set to $\Phi (\mathbf{x},t)$ whenever an event is triggered. Each event consists of a packet + +$$ +(\mathbf {x}, t, \operatorname {s i g n} (\Phi (\mathbf {x}, t) - \Phi_ {\mathrm {r e f}} (\mathbf {x}))) \tag {2} +$$ + +that encodes the polarity of the change ("what to transmit"). + +Event polarities, although adequate for some applications, do not retain sufficient information to support a general set of computer vision tasks. A stream of event polarities is an extremely lossy representation. Notably, it only defines scene intensity up to an initial unknown reference value, and it does + +![](images/3f686bfaec24cde85e96edda5ffa8a5848dd0072a989af09283ac91240a860b9.jpg) +(a) 1-bit events + +![](images/acd36c968fa6beb6f6c7691002d9cb411aa11b67edee335acad1777337357991.jpg) +(b) Flux levels +Figure 2. Altering "what to transmit." (a) We sum the events generated by a jack-in-the-box toy as it springs up. This sum gives a lossy encoding of brightness changes in dynamic regions. (b) Transmitting levels instead of changes helps recover details in static regions. (c) Adaptive exposures, which accumulate flux between consecutive events, provide substantial noise reduction. + +![](images/25bb2f6f106dc619d84d367dd05d04ce16e75cff28d1412ad034acafb0711644.jpg) +(c) Adaptive exposures + +not encode any information in regions not producing events, i.e., regions with little or no motion. + +Our key observation is that existing event cameras represent just one operating point in a broader space of generalized event cameras, which is defined by two axes: "what to transmit" and "when to transmit." By considering alternate points in this space, we can design event cameras that preserve high-fidelity scene intensity. This enables plug-and-play inference with a host of algorithms developed by the mainstream vision community. + +We begin with a conceptual exploration of this generalized space, before describing its physical implementation. + +Generalizing "what to transmit." As a first step, we can modify the event camera such that it transmits $n$ -bit values instead of one-bit change polarities. When an event is triggered, we send the current value of $\Phi(\mathbf{x}, t)$ ; if a pixel triggers no events, we transmit $\Phi$ during the final readout. As we show in Fig. 2(b), this simple change allows us to recover scene intensity, even in static regions. It is important to note that, while the transmitted quantity differs from conventional events, we retain the defining feature of an event camera: selective transmission based on scene dynamics (where we transmit according to Eq. (1)). Thus, the readout remains decoupled from the time resolution. + +If $\Phi (\mathbf{x},t)$ were a perfect estimate of scene intensity, then the changes thus far would suffice. However, $\Phi$ is fundamentally noisy: to capture high-speed changes, $\Phi$ must encompass a shorter duration, which leads to higher noise. This is a manifestation of the classical noise-blur tradeoff. + +To address this problem, we introduce the abstraction of an integrator (or $\Sigma$ ), that defines how we accumulate incident flux and, in turn, what we transmit. Ideally, we want the integrator to adapt to scene dynamics, i.e., accumulate over longer durations when there is less motion, and vice versa. We observe that event generation, which is based on scene dynamics, can be used to formulate an adaptive integrator. Specifically, we propose an integrator $\Sigma_{\mathrm{cum}}$ that computes the cumulative flux since the last event: + +$$ +\Sigma_ {\mathrm {c u m l}} (\mathbf {x}, t) = \int_ {T _ {0}} ^ {t} \Phi (\mathbf {x}, s) d s, \tag {3} +$$ + +where $T_{0}$ is the time of the last event. When an event is triggered at time $T_{1}$ , we communicate the value $^{2}$ of $\Sigma_{\mathrm{cum}}(\mathbf{x}, T_{1})$ , which + +
Event cameraIntegrator (Σ)Change detector (Δ)Event packetsMin. latencyIntensity info.?Low-light perf.
Existing (DVS [34])logarithmiccomparatorbinary10-6to 10-5sXpoor
Sec. 4.1adaptive exposureBayesian change detector [2]scalar10-5s✓good
Sec. 4.2adaptive exposurevariance-aware differencespatches10-4s✓good
Sec. 4.3coded exposureBinomial confidence intervalvector10-3s✓good
+ +Table 1. Summary of generalized event cameras. Our designs integrate photon detections $(\Sigma)$ and detect scene-content changes $(\Delta)$ in distinct ways. We compare our designs to existing DVS event cameras based on their event streams, latencies, and intensity-preserving nature. While providing a direct power comparison to DVS is difficult, we compare the power characteristics among our designs in Sec. 5.4. + +we interpret as the intensity throughout $[T_0, T_1]$ . This approach yields a piece-wise constant time series, with segments delimited by events. Adaptive exposures significantly reduce noise, as we show in Fig. 2 (c). + +Generalizing "when to transmit." The success of the adaptive integrator crucially depends on the reliability of events; for example, triggering false events in static regions causes unnecessary noise. We refer to the event-generation procedure as the change detector, denoted by $\Delta$ . Current event cameras detect changes by applying a fixed threshold to measured intensity differences (Eq. (1)). This method has two key limitations: it only considers the value of $\Phi$ at pixel location $\mathbf{x}$ and time $t$ , and it is not attuned to the stochasticity in $\Phi$ . + +We design more robust change detectors that (1) leverage enhanced spatiotemporal contexts, and (2) incorporate noise awareness, either explicitly by tuning thresholds, or implicitly by modulating the detector's behavior. Specifically, we improve reliability by using temporal forecasters (Sec. 4.1), by leveraging correlated changes in patches (Sec. 4.2), or by exploiting integrator statistics (Sec. 4.3). + +Realizing generalized event cameras. The critical detail remaining is how we implement our proposed designs in practice. We need direct access to flux estimates at a high time resolution. Conventional high-speed cameras can provide such access, however, they incur substantial per-frame read noise $(\sim 20 - 40e^{-}$ [25]) that grows with frame rate [8]. + +We turn to an emerging class of single-photon sensors, single-photon avalanche diodes (SPADs [48]), that has witnessed dramatic improvements in device practicality and key sensor characteristics (e.g., array sizes and fill factors) in recent years [38, 62]. SPADs can operate at extremely high speeds ( $\sim 100\mathrm{kHz}$ ) without incurring per-frame read noise. Each $\Phi(\mathbf{x},t)$ measured by a SPAD is limited only by the fundamental stochasticity of photon arrivals (shot noise). This allows SPADs to provide high timing resolution without a substantial noise penalty. In the next section, we describe the image formation model of SPADs and provide single-photon implementations of our designs. + +# 4. Single-Photon Generalized Event Cameras + +A SPAD array can operate as a high-speed photon detector, producing binary frames as output. Each binary value indicates whether at least one photon was detected during an expo + +sure. The SPAD output response, $\Phi (\mathbf{x},t)$ , can be modeled as a Bernoulli random variable, with + +$$ +P (\Phi (\mathbf {x}, t) = 1) = 1 - e ^ {- N (\mathbf {x}, t)}, \tag {4} +$$ + +where $N(\mathbf{x}, t)$ is the average number of photo-electrons during an exposure, including any spurious detections. The inherently digital SPAD response allows us to compute software-level transformations on the signal $\Phi(\mathbf{x}, t)$ , including operations that may be challenging to realize via analog processing. These transformations can be readily reconfigured, which permits a spectrum of event camera designs, not just one particular choice. However, there is one consideration: our designs should be lightweight and computable on chip. As we show in Sec. 5.4, this is vital to implementing generalized event cameras without the practical costs associated with reading off raw SPAD outputs. + +We now describe a set of SPAD-based event cameras (summarized in Tab. 1), beginning with the adaptive exposure method from the previous section. + +Adaptive-exposure event camera. We obtain a SPAD implementation of the adaptive exposure described in Eq. (3) by replacing the integral with a sum over photons: + +$$ +\Sigma_ {\mathrm {c u m l}} (\mathbf {x}, t) = \sum_ {s = T _ {0}} ^ {t} \Phi (\mathbf {x}, s). \tag {5} +$$ + +To generate events, we can use a threshold-based change detector (Eq. (1)). Differences between individual binary values are not sufficiently informative; therefore, we apply Eq. (1) to an exponential moving average (EMA) computed on $\Phi$ . We call this an "adaptive-EMA" event camera. + +# 4.1. Bayesian Change Detector + +A fixed-threshold change detector such as Eq. (1) does not account for the SPAD's image formation model; it uses the same threshold irrespective of the underlying variance in photon detections. As a result, such a detector may fail to detect changes in low-contrast regions without producing a large number of false-positive detections (see Fig. 3 (left)). + +In this section, we consider a Bayesian change detector, BOCPD [1], that is tailored to the Bernoulli statistics of photon detections. BOCPD uses a series of forecasters to estimate the likelihood of an abrupt change. At each time step, a new forecaster $\nu_{t}$ is initialized as a recurrence of previous forecasters, + +![](images/b69d56f366f855315de174bff0eb73de6aa8b6b8681d4f513b537ae3fefee08e.jpg) +Figure 3. Bayesian- vs. EMA-based change detection. (left) A fixed-threshold change detector (used in adaptive-EMA) makes it difficult to segment low-contrast changes. (center) The Bayesian formulation attunes to the stochasticity in incident flux and can detect fine-grained changes such as the corners of the hole saw bit; (right) as a result, the integrator captures the rotational dynamics. + +and existing forecasters are updated: + +$$ +\nu_ {t} = (1 - \gamma) \sum_ {s = 1} ^ {t - 1} l _ {s} \nu_ {s}, \quad \nu_ {s} \leftarrow \gamma l _ {s} \nu_ {s} \forall s < t, \tag {6} +$$ + +where $\gamma \in [0,1]$ is the sensitivity of the change detector, with larger $\gamma$ resulting in more frequent detections. $l_{s}$ is the predictive likelihood of each forecaster, which we compute by tracking two values per forecaster, $\alpha_{s}$ and $\beta_{s}$ , that correspond to the parameters of a Beta prior. For a new forecaster, these values are initialized to 1 each, reflecting a uniform prior. Existing $(\alpha_{s},\beta_{s}),\forall s < t,$ are updated as + +$$ +\alpha_ {s} \leftarrow \alpha_ {s} + \Phi (\mathbf {x}, t), \quad \beta_ {s} \leftarrow \beta_ {s} + 1 - \Phi (\mathbf {x}, t). \tag {7} +$$ + +$l_{s}$ is given by $\alpha_{s} / (\alpha_{s} + \beta_{s})$ if $\Phi (\mathbf{x},t) = 1$ , and $\beta_{s} / (\alpha_{s} + \beta_{s})$ otherwise. An event is triggered if the highest-value forecaster does not correspond to $T_{0}$ , the timestamp of the last event; mathematically, if $\arg \max_t\nu_t\neq T_0$ . + +To make BOCPD viable in memory-constrained scenarios, we apply extreme pruning by retaining only the three highest-value forecasters [64]. We also incorporate restarts, deleting previous forecasters when a change is detected [2]. + +Compared to an EMA-based change detector, the Bayesian approach more reliably triggers events in response to scene changes while better filtering out stochastic variations caused by photon noise—which we show in Fig. 3. + +# 4.2. Spatiotemporal Chunk Events + +Sec. 4.1 leverages an expanded temporal context for change detection; however, it treats each pixel independently and does not exploit spatial information. In this section, we devise an event camera with enhanced spatial context that operates on small patches, e.g., of $4 \times 4$ pixels. It is difficult to derive efficient Bayesian change detectors for multivariate time series; thus, we adopt a model-free approach that does not explicitly parameterize the patch distribution. To afford computational breathing room for more expensive patch-wise operations, we employ temporal chunking. That is, we average $\Phi(\mathbf{x}, t)$ over a small number of binary frames (e.g., 32 binary frames) instead of operating on individual binary frames; generally, this averaging does not induce perceptible blur. + +Let vector $\Phi_{\mathrm{chunk}}(\mathbf{y},t)$ represent the chunk-wise average of + +![](images/884fb01b67138d44d512fcc08e7c7b6f5ee158be9ba9025679a5746b3073acba.jpg) +Figure 4. Spatiotemporal chunk events. We evaluate the difference between the current chunk and a stored reference in a learned linear-feature space. Unlike the L2 norm, which is permutation-invariant, the feature-space norm is sensitive to spatial structure. Randomly shuffling the pixel values reduces the transform-domain norm (the shuffled patch has a more "noise-like" structure). + +photon detections at patch location $\mathbf{y}$ . Let vector $\Sigma_{\mathrm{patch}}(\mathbf{y},t)$ be an integrator representing the cumulative mean since the last event, but excluding $\Phi_{\mathrm{chunk}}$ . We want to estimate whether $\Phi_{\mathrm{chunk}}$ belongs to the same distribution as $\Sigma_{\mathrm{patch}}$ . We do so with a lightweight approach, that computes the distance between $\Phi_{\mathrm{chunk}}$ and $\Sigma_{\mathrm{patch}}$ in the linear feature space of matrix $\mathbf{P}$ . As we show in Fig. 4, linear features allow us to capture spatial structure within a patch. Geometrically, $\mathbf{P}$ induces a hyperellipsoidal decision boundary, in contrast to the spherical boundary of the L2 norm. + +This method generates an event whenever + +$$ +\left\| \mathbf {P} \left(\tilde {\Phi} _ {\text {c h u n k}} (\mathbf {y}, t) - \tilde {\Sigma} _ {\text {p a t c h}} (\mathbf {y}, t)\right) \right\| _ {2} \geq \tau , \tag {8} +$$ + +where $\tau$ is the threshold. When there is no event, we extend the cumulative mean to include the current chunk. Before computing linear features, we normalize $\Phi_{\mathrm{chunk}}$ and $\Sigma_{\mathrm{patch}}$ element-wise according to the estimated variance in $\Phi_{\mathrm{chunk}} - \Sigma_{\mathrm{patch}}$ ; we annotate the normalized versions with a tilde. We estimate the variance based on the fact that, in a static patch, the elements of $\Phi_{\mathrm{chunk}}$ and $\Sigma_{\mathrm{patch}}$ are independent binomial random variables. + +We train the matrix $\mathbf{P}$ on simulated SPAD data, generated from interpolated high-speed video. We apply backpropagation through time to minimize the MSE error of the transmitted patch values. To address the non-differentiability arising from the threshold, we employ surrogate gradients. Please see the supplementary material for complete details of this method. + +# 4.3. Coded-Exposure Events + +In this section, we design a generalized event camera by applying change detection to coded exposures [22, 45, 47], which capture temporal variations by multiplexing photon detections over an integration window. This is interesting in two aspects. First, we are designing event streams based on an modality not typically associated with event cameras. Second, we show that high-speed information can be obtained even when the change detector operates at a coarser time granularity. Coded-exposure events provide somewhat lower fidelity than our designs in Secs. 4.1 and 4.2, but are more compute- and power-efficient, owing to less frequent execution of the change detector. + +At each pixel, we multiplex a temporal chunk of $T_{\mathrm{code}}$ ( $\sim 256 - 512$ ) binary values with a set of $J$ ( $\sim 2 - 6$ ) codes $C^j(\mathbf{x}, t) \forall 1 \leq j \leq J$ , producing $J$ coded exposures + +$$ +\Sigma_ {\text {c o d e d}} ^ {j} (\mathbf {x}, t) = \sum_ {s = t - T _ {\text {c o d e}}} ^ {t} \Phi (\mathbf {x}, s) C ^ {j} (\mathbf {x}, s). \tag {9} +$$ + +The codes $C^j$ are chosen to be random, mutually orthogonal binary masks, each containing $T_{\mathrm{code}} / \max (2,J)$ ones [60]. + +We exploit the statistics of coded exposures to derive a change detector. Observe that in static regions, $\Sigma_{\mathrm{coded}}^{j}(\mathbf{x},t)$ are independent and identically distributed (iid) binomial random variables. Thus, we can expect them to lie within a binomial confidence interval of one another. If not, we assume the pixel is dynamic and generate an event. We trigger an event if $\Sigma_{\mathrm{coded}}^{j}\notin \operatorname {conf}(n,\hat{p})$ for any $j$ . Here, "conf" refers to a binomial confidence interval (e.g., Wilson's score), $n = T_{\mathrm{code}} / J$ draws, and $\hat{p} = \sum_{s}\Phi (\mathbf{x},s) / T_{\mathrm{code}}$ is the empirical success probability. + +If a pixel is static, we store the sum of the $J$ coded exposures, which is a long exposure, denoted by $\Sigma_{\mathrm{long}}$ . If the pixel remains static across more than one temporal chunk, we extend $\Sigma_{\mathrm{long}}$ to include the entire duration. Whereas, if the pixel is dynamic, we transmit $\{\Sigma_{\mathrm{coded}}^j\}$ , as well as any previous static intensity encoded in $\Sigma_{\mathrm{long}}$ . Downstream, we can apply coded-exposure restoration techniques [54, 63, 66] to recover intensity frames from the coded measurements. + +# 5. Experimental Results + +We demonstrate the capabilities of generalized event cameras using a SwissSPAD2 array [62] with resolution $512 \times 256$ , which we use to capture one-bit frames at $96.8\mathrm{kHz}$ . We show the feasibility of our designs on UltraPhase [4], a recent single-photon computational platform (Sec. 5.4). + +Refinement model. For each of our event cameras, we train a refinement model that mitigates artifacts arising from the asynchronous nature of events. This model takes a periodic frame-based sampling of integrator values and outputs a video reconstruction. The sampling rate is configurable; in practice, we set it $\sim 16 - 64\times$ lower than the SPAD rate. We use a densely-connected residual architecture [63], trained on data generated by simulating photon detections on temporally interpolated [24] high-speed videos from the XVFI dataset [56]. See the supplement for training details. + +# 5.1. Extreme Bandwidth-Efficient Videography + +High-speed videography. In Fig. 5, we capture the dynamics of a deformable ball (a "stress ball") using a SPAD, a high-speed camera (Photron Infinicam) operated at 500 FPS, and a commercial event camera (Prophesee EVK4). The high-speed camera suffers from low SNR due to read noise, which manifests as prominent artifacts after on-camera compression. Meanwhile, conventional events captured by Prophesee, when processed by "intensity-from-events" methods such as E2VID+ [57] fail to + +![](images/776b147388e4d73cbd0b218575acc3eaae318467ed507cf7feccbcc9c6245c19.jpg) + +![](images/cc9ba3613e75143c8f80a2b01c65083eb77f6fca2198d5f2ab11ddefe2224e00.jpg) + +![](images/407a689e1086fe22a1e8ca633d5ee20a35b3b71c3e9653bd87edd3ccffc54290.jpg) + +![](images/7f9824a62aaf42ed57bee61aa5475c8aa12c6203db21fee904cd78988a01e42d.jpg) +Spatiotemporal chunk (Sec 4.2, 431 bps/pixel, rendered at 3025 FPS) + +![](images/4d054021013ee51bd8df9dc08f032793da90e2b8c7e8843282d7fac7c0d4517d.jpg) +Figure 5. High-speed videography of a stress ball hurled at a coffee mug. (top row) This indoor scene is challenging for existing imaging systems, including: high-speed cameras (SNR-related artifacts), event cameras (poor restoration quality), and even hybrid event + frame techniques (reconstruction artifacts). (bottom rows) In contrast, our generalized event cameras capture the stress ball's extensive deformations with high fidelity and an efficient readout. + +![](images/006e3535b29621c81543a103cee50806de3f4ad1359892539877b12bc3006488.jpg) + +recover intensities reliably, especially in static regions. We also evaluate EDI [42], a hybrid event-frame method. We consider an idealized variant that operates on SPAD events (obtained via EMA thresholding), which gives perfect event-frame alignment and a precisely known event-generation model. We refine the outputs of EDI using the same model as for our methods. We refer to this idealized, refined version of EDI as "EDI++. While EDI++ recovers more detail than other baselines, there are considerable artifacts in its outputs. + +Our method achieves high-quality reconstructions at 3025 FPS (96800/32) that faithfully capture non-rigid deformations, with only 431 bits per second per pixel (bps/pixel) readout, which is a $227\times$ compression (96800/431) of the raw SPAD capture. Viewed differently, for a 1 MPixel array, we would obtain a bitrate of 431 Mbps, implying that we can read off these 3025 FPS reconstructions over USB 2.0 (which supports transfer speeds of up to 480 Mbps). + +Event imaging in low light. Fig. 6 compares the low-light performance of frame-based, event-based, and a generalized event camera on an urban night-time scene at 7 lux (lux measured at the sensor). For frame-based cameras, a short exposure that preserves motion may be too noisy, while a long exposure can be severely blurred. The Propheesee's performance deteriorates in low light, resulting in blurred temporal gradients. EDI++, benefiting from the idealized SPAD-based implementation, can image this scene, but finer details like the motorcyclist are lost. Our generalized event cameras, on the other hand, provide reconstructions with minimal noise, blur, or artifacts—while retaining the bandwidth efficiency of event-based systems. The compression here is $307\times$ with respect to raw SPAD outputs. + +![](images/ae511bffccd971686007bd165f910cd3fcaced356b4d8dd4f75592ed7fee53f6.jpg) +Frame-based cameras + +![](images/29e24dbcbe0e27f115310fa80b4424dde132e011c8dd2271aabd1b9cb03914bb.jpg) + +![](images/a9c7ebc71b3f126014fa96329ad7621b888f565ed3cb845b516a7fa8f3f1c92c.jpg) +1 ms short exposure (120 bps/pixel at 10 FPS) + +![](images/c4656beb468802d059a0573bf0fa988a272d561e7b8b4dcf933e260259ffb09b.jpg) +80 ms long exposure (120 bps/pixel at 10 FPS) + +![](images/1245c45e0b42c4ffac7b5e892b4d0304db5c9d80b8861cbc315011d6e3d3d978.jpg) +Events only + +![](images/efc1e290b505e065d8166b8f6e6daf8dec466c4669205991e2393f347737630d.jpg) +E2VID+ (121 bps/pixel) + +![](images/a6ea1fb6959e0ed56bca57eebc431c143d92a3ccc5b69e4243fecc7ce7dbb59a.jpg) +Event + frame imaging + +![](images/79748d1ed6a0e0dfb0ee7f4725e8a0f19d3c5ffdd7337754f7f21167e88b6f8e.jpg) +EDI++ (315 bps/pixel) + +![](images/7c43d72feafcf5e7035d34c692dce0680bcbec997768dbfda9745d66ffa4a137.jpg) +Generalized events + +![](images/07b3f3bd9413bd0085d9ce0c60501e256d991927bff3747606ffb07c24f498ed.jpg) +Adaptive-Bayesian (Sec 4.1, 315 bps/pixel) + +![](images/ce6116641a15931a23432b71376aaeaea42b5a8d5b4fd89d912b09914cb9a5c2.jpg) +Figure 6. Event imaging in urban nighttime (7 lux, sensor side). (left to right) Low-light conditions necessitate long exposures in frame-based cameras, resulting in unwanted motion blur. The Propheesee EVK4 suffers from severe degradation in low light, causing E2VID+ to fail. Running EDI++ on perfectly aligned SPAD-frames and -events improves overall restoration quality but still gives failures on fast-moving objects. Our generalized events recover significantly more detail in low light, as seen in the inset of the motorcyclist. +HRNet on E2VID+ + +![](images/e2c898c0b021dd00ba8ba03385d806335c27f678adbc5aea7a99ee06a33478c6.jpg) +Conventional events (Propheesee EVK4, 331 bps/pixel) +Arc* corners + +![](images/3c68478644d38c4734ac641e3838138c3fa83b8cf7d7d03b205e7276ef735ad0.jpg) +E-RAFT flow + +![](images/4da2f47605815447849cd4dce20013bab4102e5d1e6aeeb810dc8ae0a00ab7ee.jpg) +DETR on E2VID+ + +![](images/29dad1bac0afc2e0edd276e253e8e6445027fd1adde091a2b5ef4b84b7ed3e02.jpg) +SAM on E2VID+ + +![](images/37bff2fc404531272d5d078da8af8a496b88ada92f4ea3f8ff0678ab6b207d3b.jpg) +85 ms long exposure +120 bps/pixel at 10 FPS + +![](images/1a5b976082f0a1f04b8835bc367145708c26a53df2474342b3bf0f433822e0d6.jpg) +HRNet pose + +![](images/635e3252d80a9ca2f675abb4ba90f20df65b0554d27600ffea6df6480e7f675f.jpg) +Generalized events (Spatiotemporal chunk, Sec 4.2, 520 bps/pixel) +Harris corners + +![](images/026cb2d61211598528c0c565282b3f74cfb6e3550e2278292cda54585f7fe207.jpg) +RAFT flow + +![](images/89f626bfab0ac5cdc1323856ca5f0ff544864b1df2ee1d6a4027585957df98f7.jpg) +DETR detection + +![](images/06b75834ea6b701472d74d09a713563ea54445a7588d21086efd89c02695ae3f.jpg) +SAM segmentation + +![](images/db5db6b7d05d380a522981f4079a4cce168982a9de7be8957cbcd5a5c102dff9.jpg) +0.3 ms burst exposures +15100 bps/pixel +Figure 7. Plug-and-play inference on a tennis scene. (top left) Conventional events encode temporal-gradient polarities; this lossy representation limits performance on downstream tasks. (bottom left) Generalized events encode rich scene-intensity information, with a readout comparable existing event cameras. They facilitate high-quality plug-and-play inference, without requiring dedicated algorithms. (right) Generalized event cameras give image quality comparable to burst photography techniques that have a much higher readout rate. + +# 5.2. Plug-and-Play Inference + +Generalized event cameras preserve scene intensity, which enables plug-and-play event-based vision. We consider a tennis sequence (of 8196 binary frames) containing a range of object speeds. We evaluate a range of tasks: pose estimation (HRNet [59]), corner detection [18], optical flow (RAFT [61]), object detection (DETR [11]), and segmentation (SAM [32]). We compare against event-based methods applied to Prophesee events; we use $\mathrm{Arc}^*$ [3] for corner detection and E-RAFT [16] for optical flow. For the remaining tasks, which do not have equivalent event methods, we run HRNet, DETR, and SAM on E2VID+ reconstructions. + +As Fig. 7 (top left) shows, traditional events are bandwidth efficient (331 bps/pixel), but do not provide sufficient information for successful inference. Generalized events (bottom left) have a modestly higher readout (520 bps/pixel), but support accurate inference without requiring dedicated algorithms. To provide context for these rates, we compare them against frame-based methods (right). A long exposure (120 bps/pixel) blurs out the ratchet. Burst methods [19] recover a sharp image from a stack of short exposures, but with a large readout of 15100 bps/pixel. + +# 5.3. Rate-Distortion Analysis + +Each method in Sec. 4 features a sensitivity parameter that controls the sensor readout rate (event rate), which in turn influences image quality. In this subsection, we evaluate the impact of readout on image quality (PSNR) by performing a rate-distortion analysis. For ground truth, we use a set of YouTube-sourced high-speed videos captured by a Phantom Flex4k at 1000 FPS; see the supplement for thumbnails and links. We upsample these videos to the SPAD's frame rate and then simulate 4096 binary frames using the image formation model described in Eq. (4). When computing readout for our methods, we assume that events encode 10-bit values and account for the header bits of each event packet. + +As baselines, we consider EDI++, a long exposure, compressive sensing with 8-bucket masks, and burst denoising [19] using 32 short exposures. As Fig. 8 shows, generalized event cameras provide a pronounced 4-8 dB PSNR improvement over baseline methods. Further, our methods can compress the raw SPAD response by around $80\times$ before a noticeable drop-off in PSNR is observed. + +Among our methods, the spatiotemporal chunk approach of + +![](images/fd6c6fe236e8afa729df60d5beaf108f5aaa1cd28b3144be1b37cc8978f9e5d6.jpg) +Figure 8. Rate-distortion evaluation. Our techniques feature a tunable parameter that controls the output event rate. Generalized events offer a $4 - 8\mathrm{dB}$ improvement in PSNR over $\mathrm{EDI} + +$ (at the same readout), and can compress raw photon data by $80\times$ + +Sec. 4.2 gives the best PSNR, followed by the Bayesian method (Sec. 4.1) and coded-exposure events (Sec. 4.3). That said, all methods are fairly similar in terms of rate-distortion (e.g., all three give comparable results for the scenes in Secs. 5.1 and 5.2). The methods are better distinguished by their practical characteristics. The Bayesian method gives single-photon temporal resolution; however, as we show in Sec. 5.4, it is the most expensive to compute on-chip. The chunk-based method occupies a middle ground in terms of latency and cost. Coded-exposure events have the highest latency—events are generated only every $\sim 256 - 512$ binary frames—but the lowest on-chip cost. This provides an end user the flexibility to choose from the space of generalized event cameras based on the latency requirements and the compute constraints of the target application. + +# 5.4. On-Chip Feasibility and Validation + +A critical limitation of single-photon sensors is the exorbitant bandwidth and power costs involved in reading off raw photon detections. However, the lightweight nature of our event camera designs allows us to sidestep this limitation by performing computations on-chip. We demonstrate that our methods are feasible on UltraPhase [4], a SPAD compute platform. UltraPhase consists of $3 \times 6$ compute cores, each of which is associated with $4 \times 4$ pixels. + +We implement our methods for UltraPhase using custom assembly code. Some methods require minor modifications due to instruction-set limitations; see the supplement for details. We process 2500 SPAD frames from the tennis sequence used in Sec. 5.2, cropped to the UltraPhase array size of $12 \times 24$ pixels. We determine the number of cycles required to execute the assembly code and estimate the chip's power consumption and readout bandwidth. + +All our proposed methods run comfortably within the chip's compute budget of 4202 cycles per binary frame and its memory limit of 4 Kbit per core. As seen in Fig. 9, compared to raw photon-detection readout, our techniques reduce both bandwidth and power costs by over two orders of magnitude. The coded-exposure method is particularly efficient; on most binary frames, it only requires multiplying a binary code with incident photon detections. Our proof-of-concept evaluation may pave the way for future near-sensor implementations of generalized event + +![](images/c9725d942b2555ce16b8047a70f5fe65a0e5954f75454606192abb0289916fc5.jpg) +Figure 9. On-chip compatibility. We validate the feasibility of our approach on UltraPhase [4], a computational SPAD imager. Compared to reading out raw photon data, all of our approaches give marked reductions in both bandwidth (left) and power (right). + +![](images/c57aa0525b99732632c97a11a8593ca21560f0f2f04143de57f04e9e16a9fb5c.jpg) +Rigid dynamics (200x compression) +Figure 10. Limitations and failure modes, (left) Our reconstructions (yellow inset) on dynamic scenes with rigid objects can be inferior to burst photography (green inset). (center) Modulated light sources, such as this phone screen, can trigger a deluge of events (change points shown in the inset). (right) Rapid camera motion can result in an event rate divergent from scene dynamics. + +![](images/37e4fe3c13d231503f0a2500c081e897e30bd89ef6b87af2b8d9166a8d42479a.jpg) +Phone screen (11x compression) + +![](images/e78523f46441ba8ae4f55eff71c3b62698bf81b86c0f70486a601ce9108e1be9.jpg) +Camera motion (130x compression) + +cameras, which with advances in chip-to-chip communication, could involve a dedicated "photon processing unit", similar to a camera image signal processor (ISP). + +# 6. Limitations and Discussion + +Generalized events push the frontiers of event-based imaging; however, some scenarios lead to sub-optimal performance. As seen in Fig. 10 (left), if the scene dynamics is entirely comprised of rigid motion, burst photography [36] gives better image quality, albeit with much higher readout. (middle) Similar to current event cameras, modulated light sources trigger unwanted events that reduce bandwidth savings. However, it may be possible to ignore some of these events, perhaps by modeling the lighting variations [55]. + +Ego-motion events. Camera motion can trigger events in static regions, although our methods still yield substantial compression $(130 \times$ over SPAD outputs, Fig. 10 right). We analyze the impact of ego-motion on bandwidth savings further in the supplement. However, single-photon cameras can emulate sensor motion by integrating flux along alternate spatiotemporal trajectories [60]. We can imagine a generalized event camera that is "ego-motion compensated," by computing events along a suitable trajectory. + +Photon-stream compression. SPADs generate a torrent of data—e.g., 12.5 GBps for a MPixel array at $100\mathrm{kHz}$ —that can easily overwhelm data interfaces. Generalized event cameras reduce readout by around two orders of magnitude, by decoupling readout from the SPAD's frame rate and instead basing it on scene dynamics. This could pave the way for practical, high-resolution single-photon sensors. + +# References + +[1] Ryan Prescott Adams and David JC MacKay. Bayesian online changepoint detection. arXiv preprint arXiv:0710.3742, 2007. 4 +[2] Reda Alami, Odalric Maillard, and Raphael Feraud. Restarted bayesian online change-point detector achieves optimal detection delay. In International conference on machine learning, pages 211-221. PMLR, 2020. 4, 5 +[3] Ignacio Alzugaray and Margarita Chli. Asynchronous corner detection and tracking for event cameras in real time. IEEE Robotics and Automation Letters, 3(4):3177-3184, 2018. 7 +[4] Andrei Ardelean. Computational Imaging SPAD Cameras. PhD thesis, École polytechnique fédérale de Lausanne, 2023. 2, 6, 8 +[5] Souptik Barua, Yoshitaka Miyatani, and Ashok Veeraraghavan. Direct face detection and video reconstruction from event cameras. In WACV, pages 1-9. IEEE, 2016. 3 +[6] Raphael Berner, Christian Brandli, Minhao Yang, S-C Liu, and Tobi Delbruck. A 240x180 120dB 10mW 12us-latency sparse output vision sensor for mobile applications. In Proceedings of the International Image Sensors Workshop, number CONF, pages 41-44, 2013. 1, 2 +[7] Anthony Bisulco, Fernando Cladera Ojeda, Volkan Isler, and Daniel D Lee. Fast motion understanding with spatiotemporal neural networks and dynamic vision sensors. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 14098-14104. IEEE, 2021. 1 +[8] Assim Boukhayma, Arnaud Peizerat, and Christian Enz. A sub-0.5 electron read noise VGA image sensor in a standard CMOS process. IEEE Journal of Solid-State Circuits, 2016. 4 +[9] Christian Brandli, Raphael Berner, Minhao Yang, Shih-Chii Liu, and Tobi Delbruck. A $240 \times 180$ dB $3\mu s$ latency global shutter spatiotemporal vision sensor. IEEE Journal of Solid-State Circuits, 49(10):2333-2341, 2014. 2, 3 +[10] Christian Brandli, Lorenz Muller, and Tobi Delbruck. Realtime, high-speed video decompression using a frame-and event-based DAVIS sensor. In 2014 IEEE International Symposium on Circuits and Systems (ISCAS), pages 686-689. IEEE, 2014. 3 +[11] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Computer Vision – ECCV 2020, pages 213–229, Cham, 2020. Springer International Publishing. 7 +[12] Andrea Censi and Davide Scaramuzza. Low-latency event-based visual odometry. In 2014 IEEE International Conference on Robotics and Automation (ICRA), pages 703-710. IEEE, 2014. 1 +[13] Matthew Cook, Luca Gugelmann, Florian Jug, Christoph Krautz, and Angelika Steger. Interacting maps for fast visual interpretation. In The 2011 International Joint Conference on Neural Networks, pages 770-776, 2011. 3 +[14] Yue Gao, Siqi Li, Yipeng Li, Yandong Guo, and Qionghai Dai. Superfast: $200 \times$ video frame interpolation via event camera. IEEE TPAMI, 45(6):7764-7780, 2023. 1 +[15] Daniel Gehrig, Henri Rebecq, Guillermo Gallego, and Davide Scaramuzza. EKLT: Asynchronous photometric feature tracking using events and frames. IJCV, 128(3):601-618, 2020. 1 +[16] Mathias Gehrig, Mario Millhäuser, Daniel Gehrig, and Davide Scaramuzza. E-raft: Dense optical flow from event cameras. In 2021 International Conference on 3D Vision (3DV), pages + +197-206.IEEE,2021.7 +[17] Istvan Gyongy, Neale AW Dutton, and Robert K Henderson. Single-photon tracking for high-speed vision. Sensors, 18(2):323, 2018. 3 +[18] Chris Harris, Mike Stephens, et al. A combined corner and edge detector. In *Alvey vision conference*, pages 10-5244. CiteSeer, 1988. 7 +[19] Samuel W. Hasinoff, Dillon Sharlet, Ryan Geiss, Andrew Adams, Jonathan T. Barron, Florian Kainz, Jiawen Chen, and Marc Levoy. Burst photography for high dynamic range and low-light imaging on mobile cameras. ACM TOG, 35(6):1-12, 2016. 7 +[20] Botao He, Haojia Li, Siyuan Wu, Dong Wang, Zhiwei Zhang, Qianli Dong, Chao Xu, and Fei Gao. Fast-dynamic-vision: Detection and tracking dynamic objects with event and depth sensing. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3071-3078. IEEE, 2021. 1 +[21] Javier Hidalgo-Carrio, Guillermo Gallego, and Davide Scaramuzza. Event-aided direct sparse odometry. In CVPR, pages 5781-5790, 2022. 1 +[22] Yasunobu Hitomi, Jinwei Gu, Mohit Gupta, Tomoo Mitsunaga, and Shree K. Nayar. Video from a single coded exposure photograph using a learned over-complete dictionary. In 2011 International Conference on Computer Vision, pages 287-294, 2011. 5 +[23] Jing Huang, Menghan Guo, and Shoushun Chen. A dynamic vision sensor with direct logarithmic output and full-frame picture-on-demand. In 2017 IEEE International Symposium on Circuits and Systems (ISCAS), pages 1-4, 2017. 2 +[24] Zhewei Huang, Tianyuan Zhang, Wen Heng, Boxin Shi, and Shuchang Zhou. Real-time intermediate flow estimation for video frame interpolation. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 6 +[25] Jorge Igual. Photographic noise performance measures based on raw files analysis of consumer cameras. *Electronics*, 8(11):1284, 2019. 4 +[26] Atul Ingle, Andreas Velten, and Mohit Gupta. High Flux Passive Imaging With Single-Photon Sensors. In CVPR, 2019. 3 +[27] Atul Ingle, Trevor Seets, Mauro Buttafava, Shantanu Gupta, Alberto Tosi, Mohit Gupta, and Andreas Velten. Passive inter-photon imaging. In CVPR, 2021. 3 +[28] Kiyotaka Iwabuchi, Yusuke Kameda, and Takayuki Hamamoto. Image quality improvements based on motion-based deblurring for single-photon imaging. IEEE Access, 9:30080-30094, 2021. 3 +[29] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. CoTracker: It is better to track together. 2023. +[30] Anthony Kay. Tesseract: an open-source optical character recognition engine. *Linux Journal*, 2007(159):2, 2007. +[31] Hanme Kim, Ankur Handa, Ryad Benosman, Sio-Hoi Ieng, and Andrew J Davison. Simultaneous mosaicing and tracking with an event camera. Proc. Brit. Mach. Vis. Conf, 43:566-576, 2014. 3 +[32] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólár, and Ross Girshick. Segment anything, 2023. 7 +[33] Haolong Li and Joerg Stueckler. Tracking 6-DoF object motion from events and frames. In 2021 IEEE International Conference + +on Robotics and Automation (ICRA), pages 14171-14177. IEEE, 2021. 1 +[34] P Lichtsteiner. 64x64 event-driven logarithmic temporal derivative silicon retina. In Program 2003 IEEE Workshop on CCD and AIS, 2003. 1, 2, 4 +[35] Yuhao Liu, Felipe Gutierrez-Barragan, Atul Ingle, Mohit Gupta, and Andreas Velten. Single-photon camera guided extreme dynamic range imaging. In WACV, pages 1575-1585, 2022. 3 +[36] Sizhuo Ma, Shantanu Gupta, Arin C. Ulku, Claudio Bruschini, Edoardo Charbon, and Mohit Gupta. Quanta burst photography. ACM TOG, 39(4):1-16, 2020. 3, 8 +[37] Sizhuo Ma, Paul Mos, Edoardo Charbon, and Mohit Gupta. Burst vision using single-photon cameras. In WACV, pages 5375-5385, 2023. 3 +[38] Kazuhiro Morimoto, Andrei Ardelean, Ming-Lo Wu, Arin Can Ulku, Ivan Michel Antolovic, Claudio Bruschini, and Edoardo Charbon. Megapixel time-gated SPAD image sensor for 2D and 3D imaging applications. Optica, 7(4):346-354, 2020. 4 +[39] Gottfried Munda, Christian Reinbacher, and Thomas Pock. Realtime intensity-image reconstruction for event cameras using manifold regularisation. *IJCV*, 126:1381–1393, 2018. 3 +[40] Shuichi Namiki, Shunichi Sato, Yusuke Kameda, and Takayuki Hamamoto. Imaging method using multi-threshold pattern for photon detection of quanta image sensor. In International Workshop on Advanced Imaging Technology (IWAIT) 2022, page 1217702. SPIE, 2022. 3 +[41] Liyuan Pan, Cedric Scheerlinck, Xin Yu, Richard Hartley, Miao-miao Liu, and Yuchao Dai. Bringing a blurry frame alive at high frame-rate with an event camera. In CVPR, 2019. 3 +[42] Liyuan Pan, Richard Hartley, Cedric Scheerlinck, Miaomiao Liu, Xin Yu, and Yuchao Dai. High frame rate video reconstruction based on an event camera. IEEE TPAMI, 44(5):2519-2533, 2022. 6 +[43] Federico Paredes-Valles and Guido C. H. E. de Croon. Back to event basics: Self-supervised learning of image reconstruction for event cameras via photometric constancy. In CVPR, pages 3446–3455, 2021. 3 +[44] Christoph Posch, Daniel Matolin, and Rainer Wohlgenannt. A QVGA 143 dB dynamic range frame-free pwm image sensor with lossless pixel-level video compression and time-domain cds. IEEE Journal of Solid-State Circuits, 46(1):259-275, 2011. 1, 2 +[45] Ramesh Raskar, Amit Agrawal, and Jack Tumblin. Coded exposure photography: motion deblurring using fluttered shutter. In Acm Siggraph 2006 Papers, pages 795-804. 2006. 5 +[46] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. High speed and high dynamic range video with an event camera. IEEE TPAMI, 2019. 3 +[47] Dikpal Reddy, Ashok Veeraraghavan, and Rama Chellappa. P2C2: Programmable pixel compressive camera for high speed imaging. In CVPR 2011, pages 329-336, 2011. 5 +[48] Alexis Rochas. Single photon avalanche diodes in cmos technology. Technical report, Citeseer, 2003. 4 +[49] Nitin J Sanket, Chethan M Parameshwara, Chahat Deep Singh, Ashwin V Kuruttukulam, Cornelia Fermüller, Davide Scaramuzza, and Yiannis Aloimonos. Evdodgenet: Deep dynamic obstacle dodging with event cameras. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 10651-10657. IEEE, 2020. 1 + +[50] Cedric Scheerlinck, Nick Barnes, and Robert Mahony. Continuous-time intensity estimation using event cameras. In ACCV, pages 308-324. Springer, 2018. 3 +[51] Cedric Scheerlinck, Henri Rebecq, Daniel Gehrig, Nick Barnes, Robert Mahony, and Davide Scaramuzza. Fast image reconstruction with an event camera. In WACV, 2020. 3 +[52] Trevor Seets, Atul Ingle, Martin Laurenzis, and Andreas Velten. Motion adaptive deblurring with single-photon cameras. In WACV, pages 1945-1954, 2021. 3 +[53] Prasan Shedligeri and Kaushik Mitra. Photorealistic image reconstruction from hybrid intensity and event-based sensor. Journal of Electronic Imaging, 28(6):063012-063012, 2019. 3 +[54] Prasan Shedligeri, Anupama S, and Kaushik Mitra. A unified framework for compressive video recovery from coded exposure techniques. In WACV, pages 1600-1609, 2021. 6 +[55] Mark Sheinin, Yoav Y. Schechner, and Kiriakos N. Kutulakos. Computational imaging on the electric grid. In CVPR, 2017. 8 +[56] Hyeunjun Sim, Jihyong Oh, and Munchurl Kim. XVFI: extreme video frame interpolation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14489-14498, 2021. 6 +[57] Timo Stoffregen, Cedric Scheerlinck, Davide Scaramuzza, Tom Drummond, Nick Barnes, Lindsay Kleeman, and Robert Mahony. Reducing the sim-to-real gap for event cameras. In ECCV, pages 534-549. Springer, 2020. 6 +[58] Binyi su, Lei Yu, and Wen Yang. Event-based high frame-rate video reconstruction with a novel cycle-event network. In ICIP, pages 86-90, 2020. 3 +[59] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 7 +[60] Varun Sundar, Andrei Ardelean, Tristan Swedish, Claudio Bruschini, Edoardo Charbon, and Mohit Gupta. Sodacam: Software-defined cameras via single-photon imaging. In ICCV, pages 8165-8176, 2023. 3, 6, 8 +[61] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision – ECCV 2020, pages 402–419, Cham, 2020. Springer International Publishing. 7 +[62] Arin Can Ulku, Claudio Bruschini, Ivan Michel Antolovic, Yung Kuo, Rinet Ankri, Shimon Weiss, Xavier Michalet, and Edoardo Charbon. A $512 \times 512$ SPAD Image Sensor With Integrated Gating for Widefield FLIM. IEEE Journal of Selected Topics in Quantum Electronics, 25(1):1-12, 2019. 2, 4, 6 +[63] Lishun Wang, Miao Cao, and Xin Yuan. Efficientsci: Densely connected network with space-time factorization for large-scale video snapshot compressive imaging. In CVPR, pages 18477-18486, 2023. 6 +[64] Zhaohui Wang, Xiao Lin, Abhinav Mishra, and Ram Sriharsha. Online changepoint detection on a budget. In 2021 International Conference on Data Mining Workshops (ICDMW), pages 414-420. IEEE, 2021. 5 +[65] Mian Wei, Sotiris Nousias, Rahul Gulve, David B. Lindell, and Kiriakos N. Kutulakos. Passive ultra-wideband single-photon imaging. In ICCV, pages 8135-8146, 2023. 3 +[66] Xin Yuan, Yang Liu, Jinli Suo, Frédo Durand, and Qionghai Dai. Plug-and-play algorithms for video snapshot compressive imaging. IEEE Transactions on Pattern Analysis and Machine + +Intelligence, 44(10):7093-7111, 2022. 6 + +[67] Zelin Zhang, Anthony J Yezzi, and Guillermo Gallego. Formulating event-based image reconstruction as a linear inverse problem with deep regularization using optical flow. IEEE TPAMI, 45 (07):8372-8389, 2023. 3 +[68] Alex Zihao Zhu, Nikolay Atanasov, and Kostas Daniilidis. Event-based visual inertial odometry. In CVPR, 2017. 1 +[69] Yunhao Zou, Yinqiang Zheng, Tsuyoshi Takatani, and Ying Fu. Learning to reconstruct high speed and high dynamic range videos from events. In CVPR, pages 2024-2033, 2021. 3 \ No newline at end of file diff --git a/2024/Generalized Event Cameras/images.zip b/2024/Generalized Event Cameras/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1da106af1b34c7af3efb08954cdcfd6e02352116 --- /dev/null +++ b/2024/Generalized Event Cameras/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:02181abfeadac2ec48a6961c7c3e65d640eedf5aaa9764ad1f8621f1ab0f1460 +size 546946 diff --git a/2024/Generalized Event Cameras/layout.json b/2024/Generalized Event Cameras/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..44dfdeb402cab412c242460bfe07dcc4e854d408 --- /dev/null +++ b/2024/Generalized Event Cameras/layout.json @@ -0,0 +1,13559 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 213, + 104, + 380, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 213, + 104, + 380, + 118 + ], + "spans": [ + { + "bbox": [ + 213, + 104, + 380, + 118 + ], + "type": "text", + "content": "Generalized Event Cameras" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 143, + 208, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 143, + 208, + 156 + ], + "spans": [ + { + "bbox": [ + 137, + 143, + 208, + 156 + ], + "type": "text", + "content": "Varun Sundar†*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 126, + 160, + 216, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 160, + 216, + 170 + ], + "spans": [ + { + "bbox": [ + 126, + 160, + 216, + 170 + ], + "type": "text", + "content": "vsundar4@wisc.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 250, + 143, + 339, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 143, + 339, + 156 + ], + "spans": [ + { + "bbox": [ + 250, + 143, + 339, + 156 + ], + "type": "text", + "content": "Matthew Dutson†*" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 255, + 160, + 334, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 160, + 334, + 170 + ], + "spans": [ + { + "bbox": [ + 255, + 160, + 334, + 170 + ], + "type": "text", + "content": "dutson@wisc.edu" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 380, + 143, + 460, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 143, + 460, + 156 + ], + "spans": [ + { + "bbox": [ + 380, + 143, + 460, + 156 + ], + "type": "text", + "content": "Andrei Ardelean†" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 373, + 160, + 468, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 160, + 468, + 171 + ], + "spans": [ + { + "bbox": [ + 373, + 160, + 468, + 171 + ], + "type": "text", + "content": "a.ardelean@epfl.ch" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 176, + 227, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 176, + 227, + 190 + ], + "spans": [ + { + "bbox": [ + 138, + 176, + 227, + 190 + ], + "type": "text", + "content": "Claudio Bruschini" + }, + { + "bbox": [ + 138, + 176, + 227, + 190 + ], + "type": "inline_equation", + "content": "^{\\S}" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 119, + 194, + 347, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 194, + 347, + 205 + ], + "spans": [ + { + "bbox": [ + 119, + 194, + 347, + 205 + ], + "type": "text", + "content": "{claudio.bruschini, edoardo.charbon}@epfl.ch" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 395, + 177, + 459, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 177, + 459, + 191 + ], + "spans": [ + { + "bbox": [ + 395, + 177, + 459, + 191 + ], + "type": "text", + "content": "Mohit Gupta†" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 380, + 194, + 474, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 194, + 474, + 205 + ], + "spans": [ + { + "bbox": [ + 380, + 194, + 474, + 205 + ], + "type": "text", + "content": "mohitg@cs.wisc.edu" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 82, + 211, + 510, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 211, + 510, + 226 + ], + "spans": [ + { + "bbox": [ + 82, + 211, + 510, + 226 + ], + "type": "text", + "content": "†University of Wisconsin-Madison ‡NovoViz École Polytechnique Fédérale de Lausanne" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 169, + 228, + 423, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 228, + 423, + 240 + ], + "spans": [ + { + "bbox": [ + 169, + 228, + 423, + 240 + ], + "type": "text", + "content": "wisionlab.com/project/generalized-event-cameras/" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 144, + 269, + 190, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 269, + 190, + 281 + ], + "spans": [ + { + "bbox": [ + 144, + 269, + 190, + 281 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 293, + 290, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 293, + 290, + 520 + ], + "spans": [ + { + "bbox": [ + 47, + 293, + 290, + 520 + ], + "type": "text", + "content": "Event cameras capture the world at high time resolution and with minimal bandwidth requirements. However, event streams, which only encode changes in brightness, do not contain sufficient scene information to support a wide variety of downstream tasks. In this work, we design generalized event cameras that inherently preserve scene intensity in a bandwidth-efficient manner. We generalize event cameras in terms of when an event is generated and what information is transmitted. To implement our designs, we turn to single-photon sensors that provide digital access to individual photon detections; this modality gives us the flexibility to realize a rich space of generalized event cameras. Our single-photon event cameras are capable of high-speed, high-fidelity imaging at low readout rates. Consequently, these event cameras can support plug-and-play downstream inference, without capturing new event datasets or designing specialized event-vision models. As a practical implication, our designs, which involve lightweight and near-sensor-compatible computations, provide a way to use single-photon sensors without exorbitant bandwidth costs." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 47, + 528, + 124, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 528, + 124, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 528, + 124, + 540 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 548, + 288, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 548, + 288, + 668 + ], + "spans": [ + { + "bbox": [ + 46, + 548, + 288, + 668 + ], + "type": "text", + "content": "Event cameras [6, 34, 44] sense the world at high speeds, providing visual information with minimal bandwidth and power. They achieve this by transmitting only changes in scene brightness, when significant \"events\" occur. However, there is a cost. Raw event data, a sparse stream of binary values, does not hold sufficient information to be used directly with mainstream vision algorithms. Therefore, while event cameras have been successful at certain tasks (e.g., object tracking [15, 33], obstacle avoidance [7, 20, 49], and high-speed odometry [12, 21, 68]), they are not widely deployed as general-purpose vision sen" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 270, + 545, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 270, + 545, + 305 + ], + "spans": [ + { + "bbox": [ + 306, + 270, + 545, + 305 + ], + "type": "text", + "content": "sors, and often need to be supplemented with conventional cameras [14, 15, 21]. These limitations are holding back this otherwise powerful technology." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 306, + 547, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 306, + 547, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 306, + 547, + 520 + ], + "type": "text", + "content": "Is it possible to realize the promise of event cameras, i.e., high temporal resolution at low bandwidth, while preserving rich scene intensity information? To realize these seemingly conflicting goals, we propose a novel family of generalized event cameras. We conceptualize a space of event cameras along two key axes (Fig. 1 (top)): (a) \"when to transmit information,\" formalized as a change detection procedure " + }, + { + "bbox": [ + 304, + 306, + 547, + 520 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 304, + 306, + 547, + 520 + ], + "type": "text", + "content": "; and (b) \"what information to transmit,\" characterized by an integrator " + }, + { + "bbox": [ + 304, + 306, + 547, + 520 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 304, + 306, + 547, + 520 + ], + "type": "text", + "content": " that encodes incident flux. Existing event cameras represent one operating point in this " + }, + { + "bbox": [ + 304, + 306, + 547, + 520 + ], + "type": "inline_equation", + "content": "(\\Sigma, \\Delta)" + }, + { + "bbox": [ + 304, + 306, + 547, + 520 + ], + "type": "text", + "content": " space. Our key observation is that by exploring this space and considering new " + }, + { + "bbox": [ + 304, + 306, + 547, + 520 + ], + "type": "inline_equation", + "content": "(\\Sigma, \\Delta)" + }, + { + "bbox": [ + 304, + 306, + 547, + 520 + ], + "type": "text", + "content": " combinations, we can design event cameras that preserve scene intensity. We propose more general integrators, e.g., that represent flux according to motion levels, that span spatial patches, or that employ temporal coding (Fig. 1 (middle)). We also introduce robust change detectors that better distinguish motion from noise, by considering increased spatiotemporal contexts and modeling noise in the sensor measurements." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 522, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 548, + 714 + ], + "type": "text", + "content": "Despite their conceptual appeal, physically implementing generalized event cameras is a challenge. This is because the requisite computations must be performed at the sensor to achieve the desired bandwidth reductions. For example, existing event cameras perform simple integration and thresholding operations via analog in-pixel circuitry. However, more general " + }, + { + "bbox": [ + 304, + 522, + 548, + 714 + ], + "type": "inline_equation", + "content": "(\\Sigma ,\\Delta)" + }, + { + "bbox": [ + 304, + 522, + 548, + 714 + ], + "type": "text", + "content": " combinations are not always amenable to analog implementations; even feasible designs might require years of hardware iteration and production scaling. To build physical realizations of generalized event cameras, we leverage an emerging sensor technology: single-photon avalanche diodes (SPADs) that provide digital access to photon detections at extremely high frame rates " + }, + { + "bbox": [ + 304, + 522, + 548, + 714 + ], + "type": "inline_equation", + "content": "(\\sim 100\\mathrm{kHz})" + }, + { + "bbox": [ + 304, + 522, + 548, + 714 + ], + "type": "text", + "content": ". This allows us to compose arbitrary software-level signal transformations, such as those required by generalized event cameras. Further, we are not locked to a particular event camera design and can realize multiple configurations" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 673, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 673, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 673, + 287, + 713 + ], + "type": "text", + "content": "*denotes equal contribution. This research was supported in parts by NSF CAREER award 1943149, NSF award CNS-2107060, and the Swiss National Science Foundation grant 200021_166289. We thank Paul Mos for providing us access to SwissSPAD2 acquisition software." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25007" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 249, + 74, + 345, + 82 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 74, + 345, + 82 + ], + "spans": [ + { + "bbox": [ + 249, + 74, + 345, + 82 + ], + "type": "text", + "content": "Conventional event camera" + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 51, + 99, + 161, + 164 + ], + "blocks": [ + { + "bbox": [ + 59, + 87, + 156, + 95 + ], + "lines": [ + { + "bbox": [ + 59, + 87, + 156, + 95 + ], + "spans": [ + { + "bbox": [ + 59, + 87, + 156, + 95 + ], + "type": "text", + "content": "Integrator (accumulates flux, " + }, + { + "bbox": [ + 59, + 87, + 156, + 95 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 59, + 87, + 156, + 95 + ], + "type": "text", + "content": ")" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 51, + 99, + 161, + 164 + ], + "lines": [ + { + "bbox": [ + 51, + 99, + 161, + 164 + ], + "spans": [ + { + "bbox": [ + 51, + 99, + 161, + 164 + ], + "type": "image", + "image_path": "eb323521c6dfc72ea4b63c3af2a9ca83b4aedb0ab5742ac8e2be0aed01ad0c1c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 166, + 130, + 182 + ], + "lines": [ + { + "bbox": [ + 78, + 166, + 130, + 182 + ], + "spans": [ + { + "bbox": [ + 78, + 166, + 130, + 182 + ], + "type": "text", + "content": "Low-pass filtered brightness" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 176, + 99, + 306, + 164 + ], + "blocks": [ + { + "bbox": [ + 184, + 87, + 293, + 95 + ], + "lines": [ + { + "bbox": [ + 184, + 87, + 293, + 95 + ], + "spans": [ + { + "bbox": [ + 184, + 87, + 293, + 95 + ], + "type": "text", + "content": "Change detector (triggers events, " + }, + { + "bbox": [ + 184, + 87, + 293, + 95 + ], + "type": "inline_equation", + "content": "\\Delta" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 176, + 99, + 306, + 164 + ], + "lines": [ + { + "bbox": [ + 176, + 99, + 306, + 164 + ], + "spans": [ + { + "bbox": [ + 176, + 99, + 306, + 164 + ], + "type": "image", + "image_path": "cb93cdc2af2460c9e6a311222e17c6b042980154ebefcdccaf0c3666cda7f75f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 165, + 284, + 182 + ], + "lines": [ + { + "bbox": [ + 206, + 165, + 284, + 182 + ], + "spans": [ + { + "bbox": [ + 206, + 165, + 284, + 182 + ], + "type": "text", + "content": "Threshold-based detector Analog comparator" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 317, + 99, + 417, + 164 + ], + "blocks": [ + { + "bbox": [ + 343, + 87, + 387, + 95 + ], + "lines": [ + { + "bbox": [ + 343, + 87, + 387, + 95 + ], + "spans": [ + { + "bbox": [ + 343, + 87, + 387, + 95 + ], + "type": "text", + "content": "Sensor output" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 317, + 99, + 417, + 164 + ], + "lines": [ + { + "bbox": [ + 317, + 99, + 417, + 164 + ], + "spans": [ + { + "bbox": [ + 317, + 99, + 417, + 164 + ], + "type": "image", + "image_path": "3a97b6196287fcae231ef2b60dc0a55be02d6d8d07bb57211323ba839ee4c0c4.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 165, + 408, + 182 + ], + "lines": [ + { + "bbox": [ + 324, + 165, + 408, + 182 + ], + "spans": [ + { + "bbox": [ + 324, + 165, + 408, + 182 + ], + "type": "text", + "content": "Stream of change polarities " + }, + { + "bbox": [ + 324, + 165, + 408, + 182 + ], + "type": "inline_equation", + "content": "(x,y,t,\\pm 1)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 429, + 101, + 542, + 164 + ], + "blocks": [ + { + "bbox": [ + 459, + 87, + 512, + 95 + ], + "lines": [ + { + "bbox": [ + 459, + 87, + 512, + 95 + ], + "spans": [ + { + "bbox": [ + 459, + 87, + 512, + 95 + ], + "type": "text", + "content": "Intensity imaging" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 429, + 101, + 542, + 164 + ], + "lines": [ + { + "bbox": [ + 429, + 101, + 542, + 164 + ], + "spans": [ + { + "bbox": [ + 429, + 101, + 542, + 164 + ], + "type": "image", + "image_path": "8327e0f3b32b651a89e7934c762422219b2e46eef0dedcaca0058a77b8fca86b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 444, + 165, + 528, + 182 + ], + "lines": [ + { + "bbox": [ + 444, + 165, + 528, + 182 + ], + "spans": [ + { + "bbox": [ + 444, + 165, + 528, + 182 + ], + "type": "text", + "content": "Reconstruction with E2VID+ 1000 FPS at 20 FPS bitrate" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 251, + 184, + 343, + 193 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 184, + 343, + 193 + ], + "spans": [ + { + "bbox": [ + 251, + 184, + 343, + 193 + ], + "type": "text", + "content": "Generalized event camera" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 52, + 198, + 162, + 263 + ], + "blocks": [ + { + "bbox": [ + 52, + 198, + 162, + 263 + ], + "lines": [ + { + "bbox": [ + 52, + 198, + 162, + 263 + ], + "spans": [ + { + "bbox": [ + 52, + 198, + 162, + 263 + ], + "type": "image", + "image_path": "23f46768133bc5f79c9122299515b4f26b23a2361d6b17825fa94dc5caf2821c.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 58, + 265, + 85, + 280 + ], + "lines": [ + { + "bbox": [ + 58, + 265, + 85, + 280 + ], + "spans": [ + { + "bbox": [ + 58, + 265, + 85, + 280 + ], + "type": "text", + "content": "Motion adaptive" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 95, + 265, + 118, + 280 + ], + "lines": [ + { + "bbox": [ + 95, + 265, + 118, + 280 + ], + "spans": [ + { + "bbox": [ + 95, + 265, + 118, + 280 + ], + "type": "text", + "content": "Spatial patches" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 129, + 265, + 159, + 281 + ], + "lines": [ + { + "bbox": [ + 129, + 265, + 159, + 281 + ], + "spans": [ + { + "bbox": [ + 129, + 265, + 159, + 281 + ], + "type": "text", + "content": "Coded exposures" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 176, + 199, + 307, + 262 + ], + "blocks": [ + { + "bbox": [ + 176, + 199, + 307, + 262 + ], + "lines": [ + { + "bbox": [ + 176, + 199, + 307, + 262 + ], + "spans": [ + { + "bbox": [ + 176, + 199, + 307, + 262 + ], + "type": "image", + "image_path": "a39965934d4a9cfeb632084a849d16c3cc0970fd6aaa7eecd4191a33803b531f.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 264, + 296, + 279 + ], + "lines": [ + { + "bbox": [ + 192, + 264, + 296, + 279 + ], + "spans": [ + { + "bbox": [ + 192, + 264, + 296, + 279 + ], + "type": "text", + "content": "Enhanced spatio-temporal support Noise-aware thresholding" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 318, + 198, + 417, + 262 + ], + "blocks": [ + { + "bbox": [ + 318, + 198, + 417, + 262 + ], + "lines": [ + { + "bbox": [ + 318, + 198, + 417, + 262 + ], + "spans": [ + { + "bbox": [ + 318, + 198, + 417, + 262 + ], + "type": "image", + "image_path": "91ef207541236e1bb5e060fc371fdc3e704e5ec4fe7d6b1fb2c3583f64a2450d.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 332, + 264, + 414, + 279 + ], + "lines": [ + { + "bbox": [ + 332, + 264, + 414, + 279 + ], + "spans": [ + { + "bbox": [ + 332, + 264, + 414, + 279 + ], + "type": "text", + "content": "Stream of integrator values " + }, + { + "bbox": [ + 332, + 264, + 414, + 279 + ], + "type": "inline_equation", + "content": "(x,y,t,\\Sigma (x,y,t))" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 428, + 200, + 542, + 263 + ], + "blocks": [ + { + "bbox": [ + 428, + 200, + 542, + 263 + ], + "lines": [ + { + "bbox": [ + 428, + 200, + 542, + 263 + ], + "spans": [ + { + "bbox": [ + 428, + 200, + 542, + 263 + ], + "type": "image", + "image_path": "4ca0635e9f44a5a0f1241e046abe02096fe14bcc7dd3a3729e8621b868f11060.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 447, + 265, + 528, + 280 + ], + "lines": [ + { + "bbox": [ + 447, + 265, + 528, + 280 + ], + "spans": [ + { + "bbox": [ + 447, + 265, + 528, + 280 + ], + "type": "text", + "content": "Reconstruction 3025 FPS at 30 FPS bitrate" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 250, + 284, + 343, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 284, + 343, + 293 + ], + "spans": [ + { + "bbox": [ + 250, + 284, + 343, + 293 + ], + "type": "text", + "content": "Plug-and-play event vision" + } + ] + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 52, + 298, + 148, + 360 + ], + "blocks": [ + { + "bbox": [ + 52, + 298, + 148, + 360 + ], + "lines": [ + { + "bbox": [ + 52, + 298, + 148, + 360 + ], + "spans": [ + { + "bbox": [ + 52, + 298, + 148, + 360 + ], + "type": "image", + "image_path": "2133d106cbe4c4917aeb158ba7c31d382cc7d2ca075c17a4d2995251bfb4a23d.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 361, + 145, + 376 + ], + "lines": [ + { + "bbox": [ + 55, + 361, + 145, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 145, + 376 + ], + "type": "text", + "content": "Contact point of a tennis shot Image Reconstruction" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 382, + 545, + 460 + ], + "lines": [ + { + "bbox": [ + 46, + 382, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 382, + 545, + 460 + ], + "type": "text", + "content": "Figure 1. Generalized event cameras. (top) Event cameras generate outputs in response to abrupt changes in scene intensity. We describe this as a combination of a low-pass integrator and a threshold-based change detector. (middle) We generalize the space of event cameras by designing integrators that capture rich intensity information, and more reliable change detectors that utilize larger spatiotemporal contexts and noise-aware thresholding (Secs. 4.1 to 4.3). Unlike existing events, our generalized event streams inherently preserve scene intensity, e.g., this ping-pong ball slingshotted against a brick wall backdrop. (bottom) Generalized event cameras enable high-fidelity bandwidth-efficient imaging: providing 3025 FPS reconstructions with a readout equivalent to a 30 FPS camera. Consequently, generalized events facilitate plug-and-play inference on a multitude of tasks in challenging scenarios (insets depict the extent of motion over " + }, + { + "bbox": [ + 46, + 382, + 545, + 460 + ], + "type": "inline_equation", + "content": "30~\\mathrm{ms}" + }, + { + "bbox": [ + 46, + 382, + 545, + 460 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 150, + 299, + 247, + 360 + ], + "blocks": [ + { + "bbox": [ + 150, + 299, + 247, + 360 + ], + "lines": [ + { + "bbox": [ + 150, + 299, + 247, + 360 + ], + "spans": [ + { + "bbox": [ + 150, + 299, + 247, + 360 + ], + "type": "image", + "image_path": "b7c0e23152bc6b87266d6c12e39abceea82b4a8814836b4b1e1fbfaa0991e3f9.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 159, + 361, + 235, + 376 + ], + "lines": [ + { + "bbox": [ + 159, + 361, + 235, + 376 + ], + "spans": [ + { + "bbox": [ + 159, + 361, + 235, + 376 + ], + "type": "text", + "content": "Tracking points at 0.3 lux CoTracker [29]" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 249, + 299, + 345, + 360 + ], + "blocks": [ + { + "bbox": [ + 249, + 299, + 345, + 360 + ], + "lines": [ + { + "bbox": [ + 249, + 299, + 345, + 360 + ], + "spans": [ + { + "bbox": [ + 249, + 299, + 345, + 360 + ], + "type": "image", + "image_path": "49f6658d4e310202bac994ccd084296f5c5bcbf1d185c35a77e4d318df22540b.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 257, + 361, + 337, + 376 + ], + "lines": [ + { + "bbox": [ + 257, + 361, + 337, + 376 + ], + "spans": [ + { + "bbox": [ + 257, + 361, + 337, + 376 + ], + "type": "text", + "content": "Nighttime object detection DETR [11]" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 348, + 299, + 444, + 360 + ], + "blocks": [ + { + "bbox": [ + 348, + 299, + 444, + 360 + ], + "lines": [ + { + "bbox": [ + 348, + 299, + 444, + 360 + ], + "spans": [ + { + "bbox": [ + 348, + 299, + 444, + 360 + ], + "type": "image", + "image_path": "687e4267a552b02ea9648304fc4ea45c5331f352efe661270b4f3c9c93e0fc23.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 359, + 361, + 432, + 376 + ], + "lines": [ + { + "bbox": [ + 359, + 361, + 432, + 376 + ], + "spans": [ + { + "bbox": [ + 359, + 361, + 432, + 376 + ], + "type": "text", + "content": "OCR on a snapping tape Tesseract [30]" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 446, + 299, + 542, + 360 + ], + "blocks": [ + { + "bbox": [ + 446, + 299, + 542, + 360 + ], + "lines": [ + { + "bbox": [ + 446, + 299, + 542, + 360 + ], + "spans": [ + { + "bbox": [ + 446, + 299, + 542, + 360 + ], + "type": "image", + "image_path": "a3d46eedb599bc1167cd8f6a2f37ae73379ac2a6e2344e9caaa8c4d9f94724e9.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 454, + 361, + 537, + 376 + ], + "lines": [ + { + "bbox": [ + 454, + 361, + 537, + 376 + ], + "spans": [ + { + "bbox": [ + 454, + 361, + 537, + 376 + ], + "type": "text", + "content": "Segmenting a dart in midair Segment Anything [32]" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "bbox": [ + 47, + 474, + 130, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 474, + 130, + 484 + ], + "spans": [ + { + "bbox": [ + 47, + 474, + 130, + 484 + ], + "type": "text", + "content": "with the same sensor." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 46, + 491, + 288, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 491, + 288, + 659 + ], + "spans": [ + { + "bbox": [ + 46, + 491, + 288, + 659 + ], + "type": "text", + "content": "Implications: extreme, bandwidth-efficient vision. Generalized event cameras support high-speed, high-quality image reconstruction, but at low bandwidths quintessential of current event cameras. For example, Fig. 1 (middle, bottom) shows reconstructions at 3025 FPS that have an effective readout of a 30 FPS frame-based camera. Further, our methods have strong low-light performance due to the SPAD's single-photon sensitivity. As we show in Fig. 1 (bottom), preserving scene intensity facilitates plug-and-play inference in challenging scenarios, with state-of-the-art vision algorithms. Critically, this does not require retraining vision models or curating dedicated datasets, which is a significant challenge for unconventional imagers. This plug-and-play capability is vital to realizing universal event vision that retains the benefits of current event cameras." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 46, + 665, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 289, + 714 + ], + "type": "text", + "content": "Scope. We consider full-stack event perception: we conceptualize a novel space of event cameras, provide relevant single-photon algorithms, analyze their imaging capabilities and rate-distortion trade-offs, and show on-chip feasibility. We demon-" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 304, + 473, + 547, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 547, + 558 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 547, + 558 + ], + "type": "text", + "content": "strate imaging capabilities in Secs. 5.1 and 5.2 using the SwissS-PAD2 array [62], and show viable implementations of our algorithms for UltraPhase [4], a recent single-photon compute platform. All of these are critical to unlocking the promise of event cameras. However, our objective is not to develop an integrated system that incorporates all these components; this paper merely takes the first steps toward that goal." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 306, + 572, + 389, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 572, + 389, + 585 + ], + "spans": [ + { + "bbox": [ + 306, + 572, + 389, + 585 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": "Event camera designs. Perhaps most widespread is the DVS event camera [34], where each pixel generates an event in response to measured changes in (log) intensity. The DAVIS event camera [6, 9] couples DVS pixels with conventional CMOS pixels, providing access to image frames. However, the frames lack the dynamic range of DVS events. A recent design, CelexV [23], provides log-intensity frames using an external trigger. ATIS [44], a less prevalent design, features asynchronous intensity events, but its sophisticated circuitry reduces pixel fill factor. The above designs are based on analog processing; we instead" + } + ] + } + ], + "index": 41 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25008" + } + ] + } + ], + "index": 42 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 240, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 240, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 240, + 83 + ], + "type": "text", + "content": "design event cameras on digital photon detections." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 88, + 287, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 88, + 287, + 243 + ], + "spans": [ + { + "bbox": [ + 46, + 88, + 287, + 243 + ], + "type": "text", + "content": "Intensity imaging with event cameras. Several approaches have been explored to obtain images from events, including Poisson solvers [5], manifold regularization [39], assuming knowledge of camera motion [13, 31] or optical flow [67], and learning-based methods [43, 46, 51, 58, 69]. However, because events often lack sufficient scene information, they are often supplemented by conventional frames [10, 41, 50, 53], either from sensors such as DAVIS or using a multi-camera setup. Fusing events and frames presents challenges due to potential spatiotemporal misalignment and discrepancies in imaging modalities. Even when these challenges are overcome, we show that fusion methods produce lower fidelity than our proposed generalized event cameras." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 248, + 288, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 248, + 288, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 248, + 288, + 369 + ], + "type": "text", + "content": "Passive single-photon imaging. In the past few years, SPADs have found compelling passive imaging applications; this includes high-dynamic range imaging [26, 27, 35, 40], motion deblurring [28, 36, 37, 52], high-speed tracking [17], and ultra wide-band videography [65]. The fine granularity of passive single-photon acquisition makes it possible to emulate a diverse set of imaging modalities [60], including event cameras, via post-capture processing. In this work, we go beyond emulating existing event cameras and design alternate event cameras that preserve high-fidelity intensity information." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 380, + 192, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 380, + 192, + 393 + ], + "spans": [ + { + "bbox": [ + 47, + 380, + 192, + 393 + ], + "type": "text", + "content": "3. What is an Event Camera?" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 402, + 287, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 287, + 473 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 287, + 473 + ], + "type": "text", + "content": "The defining characteristic of event cameras is that they transmit information selectively, in response to changes in scene content. This selectivity allows event cameras to encode scene information at high time resolutions required to capture scene dynamics, without proportionately high bandwidth. This is in contrast to frame-based cameras, where readout occurs at fixed intervals." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 474, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 287, + 521 + ], + "type": "text", + "content": "We characterize event cameras in terms of two axes: what the camera transmits and when it transmits. As a concrete example, consider existing event cameras. They trigger events (\"when to transmit\") based on a fixed threshold:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 119, + 528, + 287, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 528, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 119, + 528, + 287, + 540 + ], + "type": "interline_equation", + "content": "\\left| \\Phi (\\mathbf {x}, t) - \\Phi_ {\\text {r e f}} (\\mathbf {x}) \\right| \\geq \\tau , \\tag {1}", + "image_path": "80d243ab6fd4c7486bec444a01203e37c8d7f46dcf233b675470ffc50d235650.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 547, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 547, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 547, + 287, + 594 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 547, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\Phi (\\mathbf{x},t)" + }, + { + "bbox": [ + 46, + 547, + 287, + 594 + ], + "type": "text", + "content": " is a flux estimate at pixel " + }, + { + "bbox": [ + 46, + 547, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 547, + 287, + 594 + ], + "type": "text", + "content": " and time " + }, + { + "bbox": [ + 46, + 547, + 287, + 594 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 547, + 287, + 594 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 547, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 547, + 287, + 594 + ], + "type": "text", + "content": " is the threshold. " + }, + { + "bbox": [ + 46, + 547, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{ref}}(\\mathbf{x})" + }, + { + "bbox": [ + 46, + 547, + 287, + 594 + ], + "type": "text", + "content": " is a previously-recorded reference, set to " + }, + { + "bbox": [ + 46, + 547, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\Phi (\\mathbf{x},t)" + }, + { + "bbox": [ + 46, + 547, + 287, + 594 + ], + "type": "text", + "content": " whenever an event is triggered. Each event consists of a packet" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 108, + 594, + 287, + 607 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 594, + 287, + 607 + ], + "spans": [ + { + "bbox": [ + 108, + 594, + 287, + 607 + ], + "type": "interline_equation", + "content": "(\\mathbf {x}, t, \\operatorname {s i g n} (\\Phi (\\mathbf {x}, t) - \\Phi_ {\\mathrm {r e f}} (\\mathbf {x}))) \\tag {2}", + "image_path": "73b6dd48f9eb316b98b67d656afc3b68e38841839f91863017e7e36478133705.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 611, + 276, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 611, + 276, + 623 + ], + "spans": [ + { + "bbox": [ + 46, + 611, + 276, + 623 + ], + "type": "text", + "content": "that encodes the polarity of the change (\"what to transmit\")." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 624, + 287, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 624, + 287, + 683 + ], + "spans": [ + { + "bbox": [ + 46, + 624, + 287, + 683 + ], + "type": "text", + "content": "Event polarities, although adequate for some applications, do not retain sufficient information to support a general set of computer vision tasks. A stream of event polarities is an extremely lossy representation. Notably, it only defines scene intensity up to an initial unknown reference value, and it does" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 71, + 384, + 119 + ], + "blocks": [ + { + "bbox": [ + 307, + 71, + 384, + 119 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 384, + 119 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 384, + 119 + ], + "type": "image", + "image_path": "3f686bfaec24cde85e96edda5ffa8a5848dd0072a989af09283ac91240a860b9.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 322, + 120, + 370, + 128 + ], + "lines": [ + { + "bbox": [ + 322, + 120, + 370, + 128 + ], + "spans": [ + { + "bbox": [ + 322, + 120, + 370, + 128 + ], + "type": "text", + "content": "(a) 1-bit events" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 386, + 72, + 463, + 118 + ], + "blocks": [ + { + "bbox": [ + 386, + 72, + 463, + 118 + ], + "lines": [ + { + "bbox": [ + 386, + 72, + 463, + 118 + ], + "spans": [ + { + "bbox": [ + 386, + 72, + 463, + 118 + ], + "type": "image", + "image_path": "acd36c968fa6beb6f6c7691002d9cb411aa11b67edee335acad1777337357991.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 119, + 447, + 128 + ], + "lines": [ + { + "bbox": [ + 403, + 119, + 447, + 128 + ], + "spans": [ + { + "bbox": [ + 403, + 119, + 447, + 128 + ], + "type": "text", + "content": "(b) Flux levels" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 133, + 547, + 198 + ], + "lines": [ + { + "bbox": [ + 304, + 133, + 547, + 198 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 547, + 198 + ], + "type": "text", + "content": "Figure 2. Altering \"what to transmit.\" (a) We sum the events generated by a jack-in-the-box toy as it springs up. This sum gives a lossy encoding of brightness changes in dynamic regions. (b) Transmitting levels instead of changes helps recover details in static regions. (c) Adaptive exposures, which accumulate flux between consecutive events, provide substantial noise reduction." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 466, + 72, + 542, + 118 + ], + "blocks": [ + { + "bbox": [ + 466, + 72, + 542, + 118 + ], + "lines": [ + { + "bbox": [ + 466, + 72, + 542, + 118 + ], + "spans": [ + { + "bbox": [ + 466, + 72, + 542, + 118 + ], + "type": "image", + "image_path": "25bb2f6f106dc619d84d367dd05d04ce16e75cff28d1412ad034acafb0711644.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 468, + 119, + 540, + 129 + ], + "lines": [ + { + "bbox": [ + 468, + 119, + 540, + 129 + ], + "spans": [ + { + "bbox": [ + 468, + 119, + 540, + 129 + ], + "type": "text", + "content": "(c) Adaptive exposures" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 207, + 545, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 207, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 304, + 207, + 545, + 230 + ], + "type": "text", + "content": "not encode any information in regions not producing events, i.e., regions with little or no motion." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 232, + 546, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 232, + 546, + 315 + ], + "spans": [ + { + "bbox": [ + 304, + 232, + 546, + 315 + ], + "type": "text", + "content": "Our key observation is that existing event cameras represent just one operating point in a broader space of generalized event cameras, which is defined by two axes: \"what to transmit\" and \"when to transmit.\" By considering alternate points in this space, we can design event cameras that preserve high-fidelity scene intensity. This enables plug-and-play inference with a host of algorithms developed by the mainstream vision community." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 316, + 545, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 545, + 339 + ], + "type": "text", + "content": "We begin with a conceptual exploration of this generalized space, before describing its physical implementation." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 342, + 546, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 546, + 474 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 546, + 474 + ], + "type": "text", + "content": "Generalizing \"what to transmit.\" As a first step, we can modify the event camera such that it transmits " + }, + { + "bbox": [ + 304, + 342, + 546, + 474 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 342, + 546, + 474 + ], + "type": "text", + "content": "-bit values instead of one-bit change polarities. When an event is triggered, we send the current value of " + }, + { + "bbox": [ + 304, + 342, + 546, + 474 + ], + "type": "inline_equation", + "content": "\\Phi(\\mathbf{x}, t)" + }, + { + "bbox": [ + 304, + 342, + 546, + 474 + ], + "type": "text", + "content": "; if a pixel triggers no events, we transmit " + }, + { + "bbox": [ + 304, + 342, + 546, + 474 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 304, + 342, + 546, + 474 + ], + "type": "text", + "content": " during the final readout. As we show in Fig. 2(b), this simple change allows us to recover scene intensity, even in static regions. It is important to note that, while the transmitted quantity differs from conventional events, we retain the defining feature of an event camera: selective transmission based on scene dynamics (where we transmit according to Eq. (1)). Thus, the readout remains decoupled from the time resolution." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 475, + 545, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 475, + 545, + 534 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 545, + 534 + ], + "type": "text", + "content": "If " + }, + { + "bbox": [ + 304, + 475, + 545, + 534 + ], + "type": "inline_equation", + "content": "\\Phi (\\mathbf{x},t)" + }, + { + "bbox": [ + 304, + 475, + 545, + 534 + ], + "type": "text", + "content": " were a perfect estimate of scene intensity, then the changes thus far would suffice. However, " + }, + { + "bbox": [ + 304, + 475, + 545, + 534 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 304, + 475, + 545, + 534 + ], + "type": "text", + "content": " is fundamentally noisy: to capture high-speed changes, " + }, + { + "bbox": [ + 304, + 475, + 545, + 534 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 304, + 475, + 545, + 534 + ], + "type": "text", + "content": " must encompass a shorter duration, which leads to higher noise. This is a manifestation of the classical noise-blur tradeoff." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 535, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 535, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 535, + 545, + 640 + ], + "type": "text", + "content": "To address this problem, we introduce the abstraction of an integrator (or " + }, + { + "bbox": [ + 304, + 535, + 545, + 640 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 304, + 535, + 545, + 640 + ], + "type": "text", + "content": "), that defines how we accumulate incident flux and, in turn, what we transmit. Ideally, we want the integrator to adapt to scene dynamics, i.e., accumulate over longer durations when there is less motion, and vice versa. We observe that event generation, which is based on scene dynamics, can be used to formulate an adaptive integrator. Specifically, we propose an integrator " + }, + { + "bbox": [ + 304, + 535, + 545, + 640 + ], + "type": "inline_equation", + "content": "\\Sigma_{\\mathrm{cum}}" + }, + { + "bbox": [ + 304, + 535, + 545, + 640 + ], + "type": "text", + "content": " that computes the cumulative flux since the last event:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 368, + 640, + 545, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 640, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 368, + 640, + 545, + 666 + ], + "type": "interline_equation", + "content": "\\Sigma_ {\\mathrm {c u m l}} (\\mathbf {x}, t) = \\int_ {T _ {0}} ^ {t} \\Phi (\\mathbf {x}, s) d s, \\tag {3}", + "image_path": "161005dc9052789ae779259212587370e657554fda4bfeac8a243b4d10b647e4.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 669, + 545, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 669, + 545, + 693 + ], + "spans": [ + { + "bbox": [ + 304, + 669, + 545, + 693 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 669, + 545, + 693 + ], + "type": "inline_equation", + "content": "T_{0}" + }, + { + "bbox": [ + 304, + 669, + 545, + 693 + ], + "type": "text", + "content": " is the time of the last event. When an event is triggered at time " + }, + { + "bbox": [ + 304, + 669, + 545, + 693 + ], + "type": "inline_equation", + "content": "T_{1}" + }, + { + "bbox": [ + 304, + 669, + 545, + 693 + ], + "type": "text", + "content": ", we communicate the value" + }, + { + "bbox": [ + 304, + 669, + 545, + 693 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 304, + 669, + 545, + 693 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 669, + 545, + 693 + ], + "type": "inline_equation", + "content": "\\Sigma_{\\mathrm{cum}}(\\mathbf{x}, T_{1})" + }, + { + "bbox": [ + 304, + 669, + 545, + 693 + ], + "type": "text", + "content": ", which" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 693, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 693, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 693, + 287, + 713 + ], + "type": "text", + "content": "1Event cameras such as the DVS [9] measure a temporally low-pass filtered estimate of log-flux; we absorb this into " + }, + { + "bbox": [ + 46, + 693, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\Phi (\\mathbf{x},t)" + }, + { + "bbox": [ + 46, + 693, + 287, + 713 + ], + "type": "text", + "content": " for brevity." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 315, + 702, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 702, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 702, + 545, + 713 + ], + "type": "text", + "content": "2We can either transmit values of " + }, + { + "bbox": [ + 315, + 702, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\Sigma_{\\mathrm{cum}}" + }, + { + "bbox": [ + 315, + 702, + 545, + 713 + ], + "type": "text", + "content": " or differences (changes) to " + }, + { + "bbox": [ + 315, + 702, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\Sigma_{\\mathrm{cum}}" + }, + { + "bbox": [ + 315, + 702, + 545, + 713 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25009" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 71, + 547, + 136 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 547, + 136 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 547, + 136 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 547, + 136 + ], + "type": "table", + "html": "
Event cameraIntegrator (Σ)Change detector (Δ)Event packetsMin. latencyIntensity info.?Low-light perf.
Existing (DVS [34])logarithmiccomparatorbinary10-6to 10-5sXpoor
Sec. 4.1adaptive exposureBayesian change detector [2]scalar10-5s✓good
Sec. 4.2adaptive exposurevariance-aware differencespatches10-4s✓good
Sec. 4.3coded exposureBinomial confidence intervalvector10-3s✓good
", + "image_path": "5e811514360da0238d51bba9e75f0ac508ceb4d98dc560a8f61f84c23abb8201.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 138, + 547, + 174 + ], + "lines": [ + { + "bbox": [ + 46, + 138, + 547, + 174 + ], + "spans": [ + { + "bbox": [ + 46, + 138, + 547, + 174 + ], + "type": "text", + "content": "Table 1. Summary of generalized event cameras. Our designs integrate photon detections " + }, + { + "bbox": [ + 46, + 138, + 547, + 174 + ], + "type": "inline_equation", + "content": "(\\Sigma)" + }, + { + "bbox": [ + 46, + 138, + 547, + 174 + ], + "type": "text", + "content": " and detect scene-content changes " + }, + { + "bbox": [ + 46, + 138, + 547, + 174 + ], + "type": "inline_equation", + "content": "(\\Delta)" + }, + { + "bbox": [ + 46, + 138, + 547, + 174 + ], + "type": "text", + "content": " in distinct ways. We compare our designs to existing DVS event cameras based on their event streams, latencies, and intensity-preserving nature. While providing a direct power comparison to DVS is difficult, we compare the power characteristics among our designs in Sec. 5.4." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 182, + 288, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 182, + 288, + 231 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 288, + 231 + ], + "type": "text", + "content": "we interpret as the intensity throughout " + }, + { + "bbox": [ + 46, + 182, + 288, + 231 + ], + "type": "inline_equation", + "content": "[T_0, T_1]" + }, + { + "bbox": [ + 46, + 182, + 288, + 231 + ], + "type": "text", + "content": ". This approach yields a piece-wise constant time series, with segments delimited by events. Adaptive exposures significantly reduce noise, as we show in Fig. 2 (c)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 232, + 288, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 232, + 288, + 340 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 288, + 340 + ], + "type": "text", + "content": "Generalizing \"when to transmit.\" The success of the adaptive integrator crucially depends on the reliability of events; for example, triggering false events in static regions causes unnecessary noise. We refer to the event-generation procedure as the change detector, denoted by " + }, + { + "bbox": [ + 46, + 232, + 288, + 340 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 46, + 232, + 288, + 340 + ], + "type": "text", + "content": ". Current event cameras detect changes by applying a fixed threshold to measured intensity differences (Eq. (1)). This method has two key limitations: it only considers the value of " + }, + { + "bbox": [ + 46, + 232, + 288, + 340 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 46, + 232, + 288, + 340 + ], + "type": "text", + "content": " at pixel location " + }, + { + "bbox": [ + 46, + 232, + 288, + 340 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 232, + 288, + 340 + ], + "type": "text", + "content": " and time " + }, + { + "bbox": [ + 46, + 232, + 288, + 340 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 232, + 288, + 340 + ], + "type": "text", + "content": ", and it is not attuned to the stochasticity in " + }, + { + "bbox": [ + 46, + 232, + 288, + 340 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 46, + 232, + 288, + 340 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 340, + 288, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 340, + 288, + 424 + ], + "spans": [ + { + "bbox": [ + 46, + 340, + 288, + 424 + ], + "type": "text", + "content": "We design more robust change detectors that (1) leverage enhanced spatiotemporal contexts, and (2) incorporate noise awareness, either explicitly by tuning thresholds, or implicitly by modulating the detector's behavior. Specifically, we improve reliability by using temporal forecasters (Sec. 4.1), by leveraging correlated changes in patches (Sec. 4.2), or by exploiting integrator statistics (Sec. 4.3)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 426, + 288, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 288, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 288, + 498 + ], + "type": "text", + "content": "Realizing generalized event cameras. The critical detail remaining is how we implement our proposed designs in practice. We need direct access to flux estimates at a high time resolution. Conventional high-speed cameras can provide such access, however, they incur substantial per-frame read noise " + }, + { + "bbox": [ + 46, + 426, + 288, + 498 + ], + "type": "inline_equation", + "content": "(\\sim 20 - 40e^{-}" + }, + { + "bbox": [ + 46, + 426, + 288, + 498 + ], + "type": "text", + "content": " [25]) that grows with frame rate [8]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 498, + 288, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 288, + 631 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 288, + 631 + ], + "type": "text", + "content": "We turn to an emerging class of single-photon sensors, single-photon avalanche diodes (SPADs [48]), that has witnessed dramatic improvements in device practicality and key sensor characteristics (e.g., array sizes and fill factors) in recent years [38, 62]. SPADs can operate at extremely high speeds (" + }, + { + "bbox": [ + 46, + 498, + 288, + 631 + ], + "type": "inline_equation", + "content": "\\sim 100\\mathrm{kHz}" + }, + { + "bbox": [ + 46, + 498, + 288, + 631 + ], + "type": "text", + "content": ") without incurring per-frame read noise. Each " + }, + { + "bbox": [ + 46, + 498, + 288, + 631 + ], + "type": "inline_equation", + "content": "\\Phi(\\mathbf{x},t)" + }, + { + "bbox": [ + 46, + 498, + 288, + 631 + ], + "type": "text", + "content": " measured by a SPAD is limited only by the fundamental stochasticity of photon arrivals (shot noise). This allows SPADs to provide high timing resolution without a substantial noise penalty. In the next section, we describe the image formation model of SPADs and provide single-photon implementations of our designs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 639, + 270, + 654 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 639, + 270, + 654 + ], + "spans": [ + { + "bbox": [ + 47, + 639, + 270, + 654 + ], + "type": "text", + "content": "4. Single-Photon Generalized Event Cameras" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 660, + 289, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 660, + 289, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 660, + 289, + 696 + ], + "type": "text", + "content": "A SPAD array can operate as a high-speed photon detector, producing binary frames as output. Each binary value indicates whether at least one photon was detected during an expo" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 182, + 545, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 182, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 306, + 182, + 545, + 205 + ], + "type": "text", + "content": "sure. The SPAD output response, " + }, + { + "bbox": [ + 306, + 182, + 545, + 205 + ], + "type": "inline_equation", + "content": "\\Phi (\\mathbf{x},t)" + }, + { + "bbox": [ + 306, + 182, + 545, + 205 + ], + "type": "text", + "content": ", can be modeled as a Bernoulli random variable, with" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 360, + 210, + 545, + 224 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 360, + 210, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 360, + 210, + 545, + 224 + ], + "type": "interline_equation", + "content": "P (\\Phi (\\mathbf {x}, t) = 1) = 1 - e ^ {- N (\\mathbf {x}, t)}, \\tag {4}", + "image_path": "cad8370667680c8d57a28f572159268c2205a7c3d7c0cfd24cbf841885be6485.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 231, + 547, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 231, + 547, + 363 + ], + "spans": [ + { + "bbox": [ + 304, + 231, + 547, + 363 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 231, + 547, + 363 + ], + "type": "inline_equation", + "content": "N(\\mathbf{x}, t)" + }, + { + "bbox": [ + 304, + 231, + 547, + 363 + ], + "type": "text", + "content": " is the average number of photo-electrons during an exposure, including any spurious detections. The inherently digital SPAD response allows us to compute software-level transformations on the signal " + }, + { + "bbox": [ + 304, + 231, + 547, + 363 + ], + "type": "inline_equation", + "content": "\\Phi(\\mathbf{x}, t)" + }, + { + "bbox": [ + 304, + 231, + 547, + 363 + ], + "type": "text", + "content": ", including operations that may be challenging to realize via analog processing. These transformations can be readily reconfigured, which permits a spectrum of event camera designs, not just one particular choice. However, there is one consideration: our designs should be lightweight and computable on chip. As we show in Sec. 5.4, this is vital to implementing generalized event cameras without the practical costs associated with reading off raw SPAD outputs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 364, + 547, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 364, + 547, + 400 + ], + "spans": [ + { + "bbox": [ + 304, + 364, + 547, + 400 + ], + "type": "text", + "content": "We now describe a set of SPAD-based event cameras (summarized in Tab. 1), beginning with the adaptive exposure method from the previous section." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 407, + 547, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 407, + 547, + 443 + ], + "spans": [ + { + "bbox": [ + 304, + 407, + 547, + 443 + ], + "type": "text", + "content": "Adaptive-exposure event camera. We obtain a SPAD implementation of the adaptive exposure described in Eq. (3) by replacing the integral with a sum over photons:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 370, + 448, + 545, + 480 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 448, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 370, + 448, + 545, + 480 + ], + "type": "interline_equation", + "content": "\\Sigma_ {\\mathrm {c u m l}} (\\mathbf {x}, t) = \\sum_ {s = T _ {0}} ^ {t} \\Phi (\\mathbf {x}, s). \\tag {5}", + "image_path": "931a8b33ee44b2d0324ce7d3086d728cd861e1ea2742c36eb7f4476ecf24054a.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 486, + 547, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 547, + 546 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 547, + 546 + ], + "type": "text", + "content": "To generate events, we can use a threshold-based change detector (Eq. (1)). Differences between individual binary values are not sufficiently informative; therefore, we apply Eq. (1) to an exponential moving average (EMA) computed on " + }, + { + "bbox": [ + 304, + 486, + 547, + 546 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 304, + 486, + 547, + 546 + ], + "type": "text", + "content": ". We call this an \"adaptive-EMA\" event camera." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 559, + 452, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 559, + 452, + 572 + ], + "spans": [ + { + "bbox": [ + 306, + 559, + 452, + 572 + ], + "type": "text", + "content": "4.1. Bayesian Change Detector" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 579, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 579, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 579, + 545, + 652 + ], + "type": "text", + "content": "A fixed-threshold change detector such as Eq. (1) does not account for the SPAD's image formation model; it uses the same threshold irrespective of the underlying variance in photon detections. As a result, such a detector may fail to detect changes in low-contrast regions without producing a large number of false-positive detections (see Fig. 3 (left))." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": "In this section, we consider a Bayesian change detector, BOCPD [1], that is tailored to the Bernoulli statistics of photon detections. BOCPD uses a series of forecasters to estimate the likelihood of an abrupt change. At each time step, a new forecaster " + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\nu_{t}" + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": " is initialized as a recurrence of previous forecasters," + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 703, + 188, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 703, + 188, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 703, + 188, + 712 + ], + "type": "text", + "content": "we treat this as an implementation detail here." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25010" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 283, + 137 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 283, + 137 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 283, + 137 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 283, + 137 + ], + "type": "image", + "image_path": "b69d56f366f855315de174bff0eb73de6aa8b6b8681d4f513b537ae3fefee08e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 142, + 287, + 208 + ], + "lines": [ + { + "bbox": [ + 46, + 142, + 287, + 208 + ], + "spans": [ + { + "bbox": [ + 46, + 142, + 287, + 208 + ], + "type": "text", + "content": "Figure 3. Bayesian- vs. EMA-based change detection. (left) A fixed-threshold change detector (used in adaptive-EMA) makes it difficult to segment low-contrast changes. (center) The Bayesian formulation attunes to the stochasticity in incident flux and can detect fine-grained changes such as the corners of the hole saw bit; (right) as a result, the integrator captures the rotational dynamics." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 214, + 188, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 214, + 188, + 225 + ], + "spans": [ + { + "bbox": [ + 47, + 214, + 188, + 225 + ], + "type": "text", + "content": "and existing forecasters are updated:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 231, + 287, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 231, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 70, + 231, + 287, + 262 + ], + "type": "interline_equation", + "content": "\\nu_ {t} = (1 - \\gamma) \\sum_ {s = 1} ^ {t - 1} l _ {s} \\nu_ {s}, \\quad \\nu_ {s} \\leftarrow \\gamma l _ {s} \\nu_ {s} \\forall s < t, \\tag {6}", + "image_path": "d524217929bc0058eb1ccfcb97a09980678fd208cd055c954d954475ebbec398.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 269, + 287, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 269, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 46, + 269, + 287, + 353 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 269, + 287, + 353 + ], + "type": "inline_equation", + "content": "\\gamma \\in [0,1]" + }, + { + "bbox": [ + 46, + 269, + 287, + 353 + ], + "type": "text", + "content": " is the sensitivity of the change detector, with larger " + }, + { + "bbox": [ + 46, + 269, + 287, + 353 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 269, + 287, + 353 + ], + "type": "text", + "content": " resulting in more frequent detections. " + }, + { + "bbox": [ + 46, + 269, + 287, + 353 + ], + "type": "inline_equation", + "content": "l_{s}" + }, + { + "bbox": [ + 46, + 269, + 287, + 353 + ], + "type": "text", + "content": " is the predictive likelihood of each forecaster, which we compute by tracking two values per forecaster, " + }, + { + "bbox": [ + 46, + 269, + 287, + 353 + ], + "type": "inline_equation", + "content": "\\alpha_{s}" + }, + { + "bbox": [ + 46, + 269, + 287, + 353 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 269, + 287, + 353 + ], + "type": "inline_equation", + "content": "\\beta_{s}" + }, + { + "bbox": [ + 46, + 269, + 287, + 353 + ], + "type": "text", + "content": ", that correspond to the parameters of a Beta prior. For a new forecaster, these values are initialized to 1 each, reflecting a uniform prior. Existing " + }, + { + "bbox": [ + 46, + 269, + 287, + 353 + ], + "type": "inline_equation", + "content": "(\\alpha_{s},\\beta_{s}),\\forall s < t," + }, + { + "bbox": [ + 46, + 269, + 287, + 353 + ], + "type": "text", + "content": " are updated as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 74, + 358, + 287, + 370 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 358, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 74, + 358, + 287, + 370 + ], + "type": "interline_equation", + "content": "\\alpha_ {s} \\leftarrow \\alpha_ {s} + \\Phi (\\mathbf {x}, t), \\quad \\beta_ {s} \\leftarrow \\beta_ {s} + 1 - \\Phi (\\mathbf {x}, t). \\tag {7}", + "image_path": "1e0d37b1a4192e2f701c8d857e28f170f88e5c48a0213cd5a616fc4fc0ea0210.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 376, + 287, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 287, + 425 + ], + "type": "inline_equation", + "content": "l_{s}" + }, + { + "bbox": [ + 46, + 376, + 287, + 425 + ], + "type": "text", + "content": " is given by " + }, + { + "bbox": [ + 46, + 376, + 287, + 425 + ], + "type": "inline_equation", + "content": "\\alpha_{s} / (\\alpha_{s} + \\beta_{s})" + }, + { + "bbox": [ + 46, + 376, + 287, + 425 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 46, + 376, + 287, + 425 + ], + "type": "inline_equation", + "content": "\\Phi (\\mathbf{x},t) = 1" + }, + { + "bbox": [ + 46, + 376, + 287, + 425 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 376, + 287, + 425 + ], + "type": "inline_equation", + "content": "\\beta_{s} / (\\alpha_{s} + \\beta_{s})" + }, + { + "bbox": [ + 46, + 376, + 287, + 425 + ], + "type": "text", + "content": " otherwise. An event is triggered if the highest-value forecaster does not correspond to " + }, + { + "bbox": [ + 46, + 376, + 287, + 425 + ], + "type": "inline_equation", + "content": "T_{0}" + }, + { + "bbox": [ + 46, + 376, + 287, + 425 + ], + "type": "text", + "content": ", the timestamp of the last event; mathematically, if " + }, + { + "bbox": [ + 46, + 376, + 287, + 425 + ], + "type": "inline_equation", + "content": "\\arg \\max_t\\nu_t\\neq T_0" + }, + { + "bbox": [ + 46, + 376, + 287, + 425 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 425, + 287, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 425, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 287, + 472 + ], + "type": "text", + "content": "To make BOCPD viable in memory-constrained scenarios, we apply extreme pruning by retaining only the three highest-value forecasters [64]. We also incorporate restarts, deleting previous forecasters when a change is detected [2]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 472, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 521 + ], + "type": "text", + "content": "Compared to an EMA-based change detector, the Bayesian approach more reliably triggers events in response to scene changes while better filtering out stochastic variations caused by photon noise—which we show in Fig. 3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 527, + 212, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 527, + 212, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 527, + 212, + 540 + ], + "type": "text", + "content": "4.2. Spatiotemporal Chunk Events" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 545, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 700 + ], + "type": "text", + "content": "Sec. 4.1 leverages an expanded temporal context for change detection; however, it treats each pixel independently and does not exploit spatial information. In this section, we devise an event camera with enhanced spatial context that operates on small patches, e.g., of " + }, + { + "bbox": [ + 46, + 545, + 287, + 700 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 46, + 545, + 287, + 700 + ], + "type": "text", + "content": " pixels. It is difficult to derive efficient Bayesian change detectors for multivariate time series; thus, we adopt a model-free approach that does not explicitly parameterize the patch distribution. To afford computational breathing room for more expensive patch-wise operations, we employ temporal chunking. That is, we average " + }, + { + "bbox": [ + 46, + 545, + 287, + 700 + ], + "type": "inline_equation", + "content": "\\Phi(\\mathbf{x}, t)" + }, + { + "bbox": [ + 46, + 545, + 287, + 700 + ], + "type": "text", + "content": " over a small number of binary frames (e.g., 32 binary frames) instead of operating on individual binary frames; generally, this averaging does not induce perceptible blur." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 58, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 701, + 287, + 714 + ], + "type": "text", + "content": "Let vector " + }, + { + "bbox": [ + 58, + 701, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{chunk}}(\\mathbf{y},t)" + }, + { + "bbox": [ + 58, + 701, + 287, + 714 + ], + "type": "text", + "content": " represent the chunk-wise average of" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 310, + 70, + 541, + 159 + ], + "blocks": [ + { + "bbox": [ + 310, + 70, + 541, + 159 + ], + "lines": [ + { + "bbox": [ + 310, + 70, + 541, + 159 + ], + "spans": [ + { + "bbox": [ + 310, + 70, + 541, + 159 + ], + "type": "image", + "image_path": "884fb01b67138d44d512fcc08e7c7b6f5ee158be9ba9025679a5746b3073acba.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 163, + 547, + 229 + ], + "lines": [ + { + "bbox": [ + 304, + 163, + 547, + 229 + ], + "spans": [ + { + "bbox": [ + 304, + 163, + 547, + 229 + ], + "type": "text", + "content": "Figure 4. Spatiotemporal chunk events. We evaluate the difference between the current chunk and a stored reference in a learned linear-feature space. Unlike the L2 norm, which is permutation-invariant, the feature-space norm is sensitive to spatial structure. Randomly shuffling the pixel values reduces the transform-domain norm (the shuffled patch has a more \"noise-like\" structure)." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 238, + 546, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 238, + 546, + 356 + ], + "spans": [ + { + "bbox": [ + 304, + 238, + 546, + 356 + ], + "type": "text", + "content": "photon detections at patch location " + }, + { + "bbox": [ + 304, + 238, + 546, + 356 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 304, + 238, + 546, + 356 + ], + "type": "text", + "content": ". Let vector " + }, + { + "bbox": [ + 304, + 238, + 546, + 356 + ], + "type": "inline_equation", + "content": "\\Sigma_{\\mathrm{patch}}(\\mathbf{y},t)" + }, + { + "bbox": [ + 304, + 238, + 546, + 356 + ], + "type": "text", + "content": " be an integrator representing the cumulative mean since the last event, but excluding " + }, + { + "bbox": [ + 304, + 238, + 546, + 356 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{chunk}}" + }, + { + "bbox": [ + 304, + 238, + 546, + 356 + ], + "type": "text", + "content": ". We want to estimate whether " + }, + { + "bbox": [ + 304, + 238, + 546, + 356 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{chunk}}" + }, + { + "bbox": [ + 304, + 238, + 546, + 356 + ], + "type": "text", + "content": " belongs to the same distribution as " + }, + { + "bbox": [ + 304, + 238, + 546, + 356 + ], + "type": "inline_equation", + "content": "\\Sigma_{\\mathrm{patch}}" + }, + { + "bbox": [ + 304, + 238, + 546, + 356 + ], + "type": "text", + "content": ". We do so with a lightweight approach, that computes the distance between " + }, + { + "bbox": [ + 304, + 238, + 546, + 356 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{chunk}}" + }, + { + "bbox": [ + 304, + 238, + 546, + 356 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 238, + 546, + 356 + ], + "type": "inline_equation", + "content": "\\Sigma_{\\mathrm{patch}}" + }, + { + "bbox": [ + 304, + 238, + 546, + 356 + ], + "type": "text", + "content": " in the linear feature space of matrix " + }, + { + "bbox": [ + 304, + 238, + 546, + 356 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 304, + 238, + 546, + 356 + ], + "type": "text", + "content": ". As we show in Fig. 4, linear features allow us to capture spatial structure within a patch. Geometrically, " + }, + { + "bbox": [ + 304, + 238, + 546, + 356 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 304, + 238, + 546, + 356 + ], + "type": "text", + "content": " induces a hyperellipsoidal decision boundary, in contrast to the spherical boundary of the L2 norm." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 318, + 359, + 479, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 359, + 479, + 370 + ], + "spans": [ + { + "bbox": [ + 318, + 359, + 479, + 370 + ], + "type": "text", + "content": "This method generates an event whenever" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 350, + 375, + 545, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 375, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 350, + 375, + 545, + 389 + ], + "type": "interline_equation", + "content": "\\left\\| \\mathbf {P} \\left(\\tilde {\\Phi} _ {\\text {c h u n k}} (\\mathbf {y}, t) - \\tilde {\\Sigma} _ {\\text {p a t c h}} (\\mathbf {y}, t)\\right) \\right\\| _ {2} \\geq \\tau , \\tag {8}", + "image_path": "6308d58c26ca78ed56f880b606471b1f0a7864a980475153c7363d14cfb06b11.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 395, + 546, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 395, + 546, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 395, + 546, + 479 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 395, + 546, + 479 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 395, + 546, + 479 + ], + "type": "text", + "content": " is the threshold. When there is no event, we extend the cumulative mean to include the current chunk. Before computing linear features, we normalize " + }, + { + "bbox": [ + 304, + 395, + 546, + 479 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{chunk}}" + }, + { + "bbox": [ + 304, + 395, + 546, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 395, + 546, + 479 + ], + "type": "inline_equation", + "content": "\\Sigma_{\\mathrm{patch}}" + }, + { + "bbox": [ + 304, + 395, + 546, + 479 + ], + "type": "text", + "content": " element-wise according to the estimated variance in " + }, + { + "bbox": [ + 304, + 395, + 546, + 479 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{chunk}} - \\Sigma_{\\mathrm{patch}}" + }, + { + "bbox": [ + 304, + 395, + 546, + 479 + ], + "type": "text", + "content": "; we annotate the normalized versions with a tilde. We estimate the variance based on the fact that, in a static patch, the elements of " + }, + { + "bbox": [ + 304, + 395, + 546, + 479 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{chunk}}" + }, + { + "bbox": [ + 304, + 395, + 546, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 395, + 546, + 479 + ], + "type": "inline_equation", + "content": "\\Sigma_{\\mathrm{patch}}" + }, + { + "bbox": [ + 304, + 395, + 546, + 479 + ], + "type": "text", + "content": " are independent binomial random variables." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 479, + 545, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 545, + 552 + ], + "type": "text", + "content": "We train the matrix " + }, + { + "bbox": [ + 304, + 479, + 545, + 552 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 304, + 479, + 545, + 552 + ], + "type": "text", + "content": " on simulated SPAD data, generated from interpolated high-speed video. We apply backpropagation through time to minimize the MSE error of the transmitted patch values. To address the non-differentiability arising from the threshold, we employ surrogate gradients. Please see the supplementary material for complete details of this method." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 562, + 441, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 562, + 441, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 562, + 441, + 574 + ], + "type": "text", + "content": "4.3. Coded-Exposure Events" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "content": "In this section, we design a generalized event camera by applying change detection to coded exposures [22, 45, 47], which capture temporal variations by multiplexing photon detections over an integration window. This is interesting in two aspects. First, we are designing event streams based on an modality not typically associated with event cameras. Second, we show that high-speed information can be obtained even when the change detector operates at a coarser time granularity. Coded-exposure events provide somewhat lower fidelity than our designs in Secs. 4.1 and 4.2, but are more compute- and power-efficient, owing to less frequent execution of the change detector." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25011" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 109 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 109 + ], + "type": "text", + "content": "At each pixel, we multiplex a temporal chunk of " + }, + { + "bbox": [ + 47, + 72, + 289, + 109 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{code}}" + }, + { + "bbox": [ + 47, + 72, + 289, + 109 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 47, + 72, + 289, + 109 + ], + "type": "inline_equation", + "content": "\\sim 256 - 512" + }, + { + "bbox": [ + 47, + 72, + 289, + 109 + ], + "type": "text", + "content": ") binary values with a set of " + }, + { + "bbox": [ + 47, + 72, + 289, + 109 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 47, + 72, + 289, + 109 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 47, + 72, + 289, + 109 + ], + "type": "inline_equation", + "content": "\\sim 2 - 6" + }, + { + "bbox": [ + 47, + 72, + 289, + 109 + ], + "type": "text", + "content": ") codes " + }, + { + "bbox": [ + 47, + 72, + 289, + 109 + ], + "type": "inline_equation", + "content": "C^j(\\mathbf{x}, t) \\forall 1 \\leq j \\leq J" + }, + { + "bbox": [ + 47, + 72, + 289, + 109 + ], + "type": "text", + "content": ", producing " + }, + { + "bbox": [ + 47, + 72, + 289, + 109 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 47, + 72, + 289, + 109 + ], + "type": "text", + "content": " coded exposures" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 87, + 114, + 288, + 147 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 114, + 288, + 147 + ], + "spans": [ + { + "bbox": [ + 87, + 114, + 288, + 147 + ], + "type": "interline_equation", + "content": "\\Sigma_ {\\text {c o d e d}} ^ {j} (\\mathbf {x}, t) = \\sum_ {s = t - T _ {\\text {c o d e}}} ^ {t} \\Phi (\\mathbf {x}, s) C ^ {j} (\\mathbf {x}, s). \\tag {9}", + "image_path": "0db3b2db3d989f2ead8b6fa79b10ca15c988eba1da1af56812e5f62bed670a2b.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 153, + 287, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 153, + 287, + 177 + ], + "spans": [ + { + "bbox": [ + 46, + 153, + 287, + 177 + ], + "type": "text", + "content": "The codes " + }, + { + "bbox": [ + 46, + 153, + 287, + 177 + ], + "type": "inline_equation", + "content": "C^j" + }, + { + "bbox": [ + 46, + 153, + 287, + 177 + ], + "type": "text", + "content": " are chosen to be random, mutually orthogonal binary masks, each containing " + }, + { + "bbox": [ + 46, + 153, + 287, + 177 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{code}} / \\max (2,J)" + }, + { + "bbox": [ + 46, + 153, + 287, + 177 + ], + "type": "text", + "content": " ones [60]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 177, + 288, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 177, + 288, + 284 + ], + "spans": [ + { + "bbox": [ + 46, + 177, + 288, + 284 + ], + "type": "text", + "content": "We exploit the statistics of coded exposures to derive a change detector. Observe that in static regions, " + }, + { + "bbox": [ + 46, + 177, + 288, + 284 + ], + "type": "inline_equation", + "content": "\\Sigma_{\\mathrm{coded}}^{j}(\\mathbf{x},t)" + }, + { + "bbox": [ + 46, + 177, + 288, + 284 + ], + "type": "text", + "content": " are independent and identically distributed (iid) binomial random variables. Thus, we can expect them to lie within a binomial confidence interval of one another. If not, we assume the pixel is dynamic and generate an event. We trigger an event if " + }, + { + "bbox": [ + 46, + 177, + 288, + 284 + ], + "type": "inline_equation", + "content": "\\Sigma_{\\mathrm{coded}}^{j}\\notin \\operatorname {conf}(n,\\hat{p})" + }, + { + "bbox": [ + 46, + 177, + 288, + 284 + ], + "type": "text", + "content": " for any " + }, + { + "bbox": [ + 46, + 177, + 288, + 284 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 177, + 288, + 284 + ], + "type": "text", + "content": ". Here, \"conf\" refers to a binomial confidence interval (e.g., Wilson's score), " + }, + { + "bbox": [ + 46, + 177, + 288, + 284 + ], + "type": "inline_equation", + "content": "n = T_{\\mathrm{code}} / J" + }, + { + "bbox": [ + 46, + 177, + 288, + 284 + ], + "type": "text", + "content": " draws, and " + }, + { + "bbox": [ + 46, + 177, + 288, + 284 + ], + "type": "inline_equation", + "content": "\\hat{p} = \\sum_{s}\\Phi (\\mathbf{x},s) / T_{\\mathrm{code}}" + }, + { + "bbox": [ + 46, + 177, + 288, + 284 + ], + "type": "text", + "content": " is the empirical success probability." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 285, + 287, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 285, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 287, + 380 + ], + "type": "text", + "content": "If a pixel is static, we store the sum of the " + }, + { + "bbox": [ + 46, + 285, + 287, + 380 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 46, + 285, + 287, + 380 + ], + "type": "text", + "content": " coded exposures, which is a long exposure, denoted by " + }, + { + "bbox": [ + 46, + 285, + 287, + 380 + ], + "type": "inline_equation", + "content": "\\Sigma_{\\mathrm{long}}" + }, + { + "bbox": [ + 46, + 285, + 287, + 380 + ], + "type": "text", + "content": ". If the pixel remains static across more than one temporal chunk, we extend " + }, + { + "bbox": [ + 46, + 285, + 287, + 380 + ], + "type": "inline_equation", + "content": "\\Sigma_{\\mathrm{long}}" + }, + { + "bbox": [ + 46, + 285, + 287, + 380 + ], + "type": "text", + "content": " to include the entire duration. Whereas, if the pixel is dynamic, we transmit " + }, + { + "bbox": [ + 46, + 285, + 287, + 380 + ], + "type": "inline_equation", + "content": "\\{\\Sigma_{\\mathrm{coded}}^j\\}" + }, + { + "bbox": [ + 46, + 285, + 287, + 380 + ], + "type": "text", + "content": ", as well as any previous static intensity encoded in " + }, + { + "bbox": [ + 46, + 285, + 287, + 380 + ], + "type": "inline_equation", + "content": "\\Sigma_{\\mathrm{long}}" + }, + { + "bbox": [ + 46, + 285, + 287, + 380 + ], + "type": "text", + "content": ". Downstream, we can apply coded-exposure restoration techniques [54, 63, 66] to recover intensity frames from the coded measurements." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 390, + 167, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 390, + 167, + 403 + ], + "spans": [ + { + "bbox": [ + 47, + 390, + 167, + 403 + ], + "type": "text", + "content": "5. Experimental Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 411, + 288, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 288, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 288, + 471 + ], + "type": "text", + "content": "We demonstrate the capabilities of generalized event cameras using a SwissSPAD2 array [62] with resolution " + }, + { + "bbox": [ + 46, + 411, + 288, + 471 + ], + "type": "inline_equation", + "content": "512 \\times 256" + }, + { + "bbox": [ + 46, + 411, + 288, + 471 + ], + "type": "text", + "content": ", which we use to capture one-bit frames at " + }, + { + "bbox": [ + 46, + 411, + 288, + 471 + ], + "type": "inline_equation", + "content": "96.8\\mathrm{kHz}" + }, + { + "bbox": [ + 46, + 411, + 288, + 471 + ], + "type": "text", + "content": ". We show the feasibility of our designs on UltraPhase [4], a recent single-photon computational platform (Sec. 5.4)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 472, + 288, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 288, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 288, + 593 + ], + "type": "text", + "content": "Refinement model. For each of our event cameras, we train a refinement model that mitigates artifacts arising from the asynchronous nature of events. This model takes a periodic frame-based sampling of integrator values and outputs a video reconstruction. The sampling rate is configurable; in practice, we set it " + }, + { + "bbox": [ + 46, + 472, + 288, + 593 + ], + "type": "inline_equation", + "content": "\\sim 16 - 64\\times" + }, + { + "bbox": [ + 46, + 472, + 288, + 593 + ], + "type": "text", + "content": " lower than the SPAD rate. We use a densely-connected residual architecture [63], trained on data generated by simulating photon detections on temporally interpolated [24] high-speed videos from the XVFI dataset [56]. See the supplement for training details." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 599, + 270, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 270, + 613 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 270, + 613 + ], + "type": "text", + "content": "5.1. Extreme Bandwidth-Efficient Videography" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "content": "High-speed videography. In Fig. 5, we capture the dynamics of a deformable ball (a \"stress ball\") using a SPAD, a high-speed camera (Photron Infinicam) operated at 500 FPS, and a commercial event camera (Prophesee EVK4). The high-speed camera suffers from low SNR due to read noise, which manifests as prominent artifacts after on-camera compression. Meanwhile, conventional events captured by Prophesee, when processed by \"intensity-from-events\" methods such as E2VID+ [57] fail to" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 388, + 144 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 388, + 144 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 388, + 144 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 388, + 144 + ], + "type": "image", + "image_path": "776b147388e4d73cbd0b218575acc3eaae318467ed507cf7feccbcc9c6245c19.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 389, + 72, + 460, + 144 + ], + "blocks": [ + { + "bbox": [ + 389, + 72, + 460, + 144 + ], + "lines": [ + { + "bbox": [ + 389, + 72, + 460, + 144 + ], + "spans": [ + { + "bbox": [ + 389, + 72, + 460, + 144 + ], + "type": "image", + "image_path": "cc9ba3613e75143c8f80a2b01c65083eb77f6fca2198d5f2ab11ddefe2224e00.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 462, + 72, + 534, + 144 + ], + "blocks": [ + { + "bbox": [ + 462, + 72, + 534, + 144 + ], + "lines": [ + { + "bbox": [ + 462, + 72, + 534, + 144 + ], + "spans": [ + { + "bbox": [ + 462, + 72, + 534, + 144 + ], + "type": "image", + "image_path": "407a689e1086fe22a1e8ca633d5ee20a35b3b71c3e9653bd87edd3ccffc54290.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 313, + 146, + 388, + 247 + ], + "blocks": [ + { + "bbox": [ + 313, + 146, + 388, + 247 + ], + "lines": [ + { + "bbox": [ + 313, + 146, + 388, + 247 + ], + "spans": [ + { + "bbox": [ + 313, + 146, + 388, + 247 + ], + "type": "image", + "image_path": "7f9824a62aaf42ed57bee61aa5475c8aa12c6203db21fee904cd78988a01e42d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 248, + 533, + 258 + ], + "lines": [ + { + "bbox": [ + 317, + 248, + 533, + 258 + ], + "spans": [ + { + "bbox": [ + 317, + 248, + 533, + 258 + ], + "type": "text", + "content": "Spatiotemporal chunk (Sec 4.2, 431 bps/pixel, rendered at 3025 FPS)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 388, + 146, + 462, + 247 + ], + "blocks": [ + { + "bbox": [ + 388, + 146, + 462, + 247 + ], + "lines": [ + { + "bbox": [ + 388, + 146, + 462, + 247 + ], + "spans": [ + { + "bbox": [ + 388, + 146, + 462, + 247 + ], + "type": "image", + "image_path": "4d054021013ee51bd8df9dc08f032793da90e2b8c7e8843282d7fac7c0d4517d.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 262, + 547, + 340 + ], + "lines": [ + { + "bbox": [ + 305, + 262, + 547, + 340 + ], + "spans": [ + { + "bbox": [ + 305, + 262, + 547, + 340 + ], + "type": "text", + "content": "Figure 5. High-speed videography of a stress ball hurled at a coffee mug. (top row) This indoor scene is challenging for existing imaging systems, including: high-speed cameras (SNR-related artifacts), event cameras (poor restoration quality), and even hybrid event + frame techniques (reconstruction artifacts). (bottom rows) In contrast, our generalized event cameras capture the stress ball's extensive deformations with high fidelity and an efficient readout." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 462, + 146, + 537, + 247 + ], + "blocks": [ + { + "bbox": [ + 462, + 146, + 537, + 247 + ], + "lines": [ + { + "bbox": [ + 462, + 146, + 537, + 247 + ], + "spans": [ + { + "bbox": [ + 462, + 146, + 537, + 247 + ], + "type": "image", + "image_path": "006e3535b29621c81543a103cee50806de3f4ad1359892539877b12bc3006488.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 348, + 547, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 348, + 547, + 456 + ], + "spans": [ + { + "bbox": [ + 305, + 348, + 547, + 456 + ], + "type": "text", + "content": "recover intensities reliably, especially in static regions. We also evaluate EDI [42], a hybrid event-frame method. We consider an idealized variant that operates on SPAD events (obtained via EMA thresholding), which gives perfect event-frame alignment and a precisely known event-generation model. We refine the outputs of EDI using the same model as for our methods. We refer to this idealized, refined version of EDI as \"EDI++. While EDI++ recovers more detail than other baselines, there are considerable artifacts in its outputs." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 457, + 547, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 457, + 547, + 553 + ], + "spans": [ + { + "bbox": [ + 305, + 457, + 547, + 553 + ], + "type": "text", + "content": "Our method achieves high-quality reconstructions at 3025 FPS (96800/32) that faithfully capture non-rigid deformations, with only 431 bits per second per pixel (bps/pixel) readout, which is a " + }, + { + "bbox": [ + 305, + 457, + 547, + 553 + ], + "type": "inline_equation", + "content": "227\\times" + }, + { + "bbox": [ + 305, + 457, + 547, + 553 + ], + "type": "text", + "content": " compression (96800/431) of the raw SPAD capture. Viewed differently, for a 1 MPixel array, we would obtain a bitrate of 431 Mbps, implying that we can read off these 3025 FPS reconstructions over USB 2.0 (which supports transfer speeds of up to 480 Mbps)." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 558, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 558, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 558, + 547, + 713 + ], + "type": "text", + "content": "Event imaging in low light. Fig. 6 compares the low-light performance of frame-based, event-based, and a generalized event camera on an urban night-time scene at 7 lux (lux measured at the sensor). For frame-based cameras, a short exposure that preserves motion may be too noisy, while a long exposure can be severely blurred. The Propheesee's performance deteriorates in low light, resulting in blurred temporal gradients. EDI++, benefiting from the idealized SPAD-based implementation, can image this scene, but finer details like the motorcyclist are lost. Our generalized event cameras, on the other hand, provide reconstructions with minimal noise, blur, or artifacts—while retaining the bandwidth efficiency of event-based systems. The compression here is " + }, + { + "bbox": [ + 305, + 558, + 547, + 713 + ], + "type": "inline_equation", + "content": "307\\times" + }, + { + "bbox": [ + 305, + 558, + 547, + 713 + ], + "type": "text", + "content": " with respect to raw SPAD outputs." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25012" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 81, + 152, + 129 + ], + "blocks": [ + { + "bbox": [ + 116, + 72, + 188, + 79 + ], + "lines": [ + { + "bbox": [ + 116, + 72, + 188, + 79 + ], + "spans": [ + { + "bbox": [ + 116, + 72, + 188, + 79 + ], + "type": "text", + "content": "Frame-based cameras" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 61, + 81, + 152, + 129 + ], + "lines": [ + { + "bbox": [ + 61, + 81, + 152, + 129 + ], + "spans": [ + { + "bbox": [ + 61, + 81, + 152, + 129 + ], + "type": "image", + "image_path": "ae511bffccd971686007bd165f910cd3fcaced356b4d8dd4f75592ed7fee53f6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 154, + 81, + 245, + 129 + ], + "blocks": [ + { + "bbox": [ + 154, + 81, + 245, + 129 + ], + "lines": [ + { + "bbox": [ + 154, + 81, + 245, + 129 + ], + "spans": [ + { + "bbox": [ + 154, + 81, + 245, + 129 + ], + "type": "image", + "image_path": "29e24dbcbe0e27f115310fa80b4424dde132e011c8dd2271aabd1b9cb03914bb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 62, + 130, + 152, + 177 + ], + "blocks": [ + { + "bbox": [ + 62, + 130, + 152, + 177 + ], + "lines": [ + { + "bbox": [ + 62, + 130, + 152, + 177 + ], + "spans": [ + { + "bbox": [ + 62, + 130, + 152, + 177 + ], + "type": "image", + "image_path": "a9c7ebc71b3f126014fa96329ad7621b888f565ed3cb845b516a7fa8f3f1c92c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 178, + 147, + 195 + ], + "lines": [ + { + "bbox": [ + 66, + 178, + 147, + 195 + ], + "spans": [ + { + "bbox": [ + 66, + 178, + 147, + 195 + ], + "type": "text", + "content": "1 ms short exposure (120 bps/pixel at 10 FPS)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 154, + 129, + 245, + 177 + ], + "blocks": [ + { + "bbox": [ + 154, + 129, + 245, + 177 + ], + "lines": [ + { + "bbox": [ + 154, + 129, + 245, + 177 + ], + "spans": [ + { + "bbox": [ + 154, + 129, + 245, + 177 + ], + "type": "image", + "image_path": "c4656beb468802d059a0573bf0fa988a272d561e7b8b4dcf933e260259ffb09b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 159, + 178, + 239, + 195 + ], + "lines": [ + { + "bbox": [ + 159, + 178, + 239, + 195 + ], + "spans": [ + { + "bbox": [ + 159, + 178, + 239, + 195 + ], + "type": "text", + "content": "80 ms long exposure (120 bps/pixel at 10 FPS)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 250, + 81, + 340, + 129 + ], + "blocks": [ + { + "bbox": [ + 276, + 72, + 313, + 80 + ], + "lines": [ + { + "bbox": [ + 276, + 72, + 313, + 80 + ], + "spans": [ + { + "bbox": [ + 276, + 72, + 313, + 80 + ], + "type": "text", + "content": "Events only" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 250, + 81, + 340, + 129 + ], + "lines": [ + { + "bbox": [ + 250, + 81, + 340, + 129 + ], + "spans": [ + { + "bbox": [ + 250, + 81, + 340, + 129 + ], + "type": "image", + "image_path": "1245c45e0b42c4ffac7b5e892b4d0304db5c9d80b8861cbc315011d6e3d3d978.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 250, + 130, + 340, + 177 + ], + "blocks": [ + { + "bbox": [ + 250, + 130, + 340, + 177 + ], + "lines": [ + { + "bbox": [ + 250, + 130, + 340, + 177 + ], + "spans": [ + { + "bbox": [ + 250, + 130, + 340, + 177 + ], + "type": "image", + "image_path": "efc1e290b505e065d8166b8f6e6daf8dec466c4669205991e2393f347737630d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 270, + 178, + 319, + 195 + ], + "lines": [ + { + "bbox": [ + 270, + 178, + 319, + 195 + ], + "spans": [ + { + "bbox": [ + 270, + 178, + 319, + 195 + ], + "type": "text", + "content": "E2VID+ (121 bps/pixel)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 345, + 81, + 436, + 129 + ], + "blocks": [ + { + "bbox": [ + 354, + 72, + 426, + 80 + ], + "lines": [ + { + "bbox": [ + 354, + 72, + 426, + 80 + ], + "spans": [ + { + "bbox": [ + 354, + 72, + 426, + 80 + ], + "type": "text", + "content": "Event + frame imaging" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 345, + 81, + 436, + 129 + ], + "lines": [ + { + "bbox": [ + 345, + 81, + 436, + 129 + ], + "spans": [ + { + "bbox": [ + 345, + 81, + 436, + 129 + ], + "type": "image", + "image_path": "a6ea1fb6959e0ed56bca57eebc431c143d92a3ccc5b69e4243fecc7ce7dbb59a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 346, + 129, + 435, + 177 + ], + "blocks": [ + { + "bbox": [ + 346, + 129, + 435, + 177 + ], + "lines": [ + { + "bbox": [ + 346, + 129, + 435, + 177 + ], + "spans": [ + { + "bbox": [ + 346, + 129, + 435, + 177 + ], + "type": "image", + "image_path": "79748d1ed6a0e0dfb0ee7f4725e8a0f19d3c5ffdd7337754f7f21167e88b6f8e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 366, + 178, + 414, + 195 + ], + "lines": [ + { + "bbox": [ + 366, + 178, + 414, + 195 + ], + "spans": [ + { + "bbox": [ + 366, + 178, + 414, + 195 + ], + "type": "text", + "content": "EDI++ (315 bps/pixel)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 441, + 81, + 533, + 129 + ], + "blocks": [ + { + "bbox": [ + 456, + 72, + 517, + 79 + ], + "lines": [ + { + "bbox": [ + 456, + 72, + 517, + 79 + ], + "spans": [ + { + "bbox": [ + 456, + 72, + 517, + 79 + ], + "type": "text", + "content": "Generalized events" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 441, + 81, + 533, + 129 + ], + "lines": [ + { + "bbox": [ + 441, + 81, + 533, + 129 + ], + "spans": [ + { + "bbox": [ + 441, + 81, + 533, + 129 + ], + "type": "image", + "image_path": "7c43d72feafcf5e7035d34c692dce0680bcbec997768dbfda9745d66ffa4a137.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 441, + 129, + 531, + 177 + ], + "blocks": [ + { + "bbox": [ + 441, + 129, + 531, + 177 + ], + "lines": [ + { + "bbox": [ + 441, + 129, + 531, + 177 + ], + "spans": [ + { + "bbox": [ + 441, + 129, + 531, + 177 + ], + "type": "image", + "image_path": "07b3f3bd9413bd0085d9ce0c60501e256d991927bff3747606ffb07c24f498ed.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 449, + 178, + 523, + 195 + ], + "lines": [ + { + "bbox": [ + 449, + 178, + 523, + 195 + ], + "spans": [ + { + "bbox": [ + 449, + 178, + 523, + 195 + ], + "type": "text", + "content": "Adaptive-Bayesian (Sec 4.1, 315 bps/pixel)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 51, + 257, + 130, + 306 + ], + "blocks": [ + { + "bbox": [ + 46, + 198, + 547, + 243 + ], + "lines": [ + { + "bbox": [ + 46, + 198, + 547, + 243 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 547, + 243 + ], + "type": "text", + "content": "Figure 6. Event imaging in urban nighttime (7 lux, sensor side). (left to right) Low-light conditions necessitate long exposures in frame-based cameras, resulting in unwanted motion blur. The Propheesee EVK4 suffers from severe degradation in low light, causing E2VID+ to fail. Running EDI++ on perfectly aligned SPAD-frames and -events improves overall restoration quality but still gives failures on fast-moving objects. Our generalized events recover significantly more detail in low light, as seen in the inset of the motorcyclist." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 51, + 257, + 130, + 306 + ], + "lines": [ + { + "bbox": [ + 51, + 257, + 130, + 306 + ], + "spans": [ + { + "bbox": [ + 51, + 257, + 130, + 306 + ], + "type": "image", + "image_path": "ce6116641a15931a23432b71376aaeaea42b5a8d5b4fd89d912b09914cb9a5c2.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 307, + 118, + 314 + ], + "lines": [ + { + "bbox": [ + 63, + 307, + 118, + 314 + ], + "spans": [ + { + "bbox": [ + 63, + 307, + 118, + 314 + ], + "type": "text", + "content": "HRNet on E2VID+" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 132, + 257, + 211, + 306 + ], + "blocks": [ + { + "bbox": [ + 167, + 247, + 336, + 256 + ], + "lines": [ + { + "bbox": [ + 167, + 247, + 336, + 256 + ], + "spans": [ + { + "bbox": [ + 167, + 247, + 336, + 256 + ], + "type": "text", + "content": "Conventional events (Propheesee EVK4, 331 bps/pixel)" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 257, + 211, + 306 + ], + "lines": [ + { + "bbox": [ + 132, + 257, + 211, + 306 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 211, + 306 + ], + "type": "image", + "image_path": "e2c898c0b021dd00ba8ba03385d806335c27f678adbc5aea7a99ee06a33478c6.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 307, + 192, + 314 + ], + "lines": [ + { + "bbox": [ + 151, + 307, + 192, + 314 + ], + "spans": [ + { + "bbox": [ + 151, + 307, + 192, + 314 + ], + "type": "text", + "content": "Arc* corners" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 213, + 257, + 291, + 306 + ], + "blocks": [ + { + "bbox": [ + 213, + 257, + 291, + 306 + ], + "lines": [ + { + "bbox": [ + 213, + 257, + 291, + 306 + ], + "spans": [ + { + "bbox": [ + 213, + 257, + 291, + 306 + ], + "type": "image", + "image_path": "3c68478644d38c4734ac641e3838138c3fa83b8cf7d7d03b205e7276ef735ad0.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 233, + 307, + 271, + 314 + ], + "lines": [ + { + "bbox": [ + 233, + 307, + 271, + 314 + ], + "spans": [ + { + "bbox": [ + 233, + 307, + 271, + 314 + ], + "type": "text", + "content": "E-RAFT flow" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 294, + 257, + 372, + 306 + ], + "blocks": [ + { + "bbox": [ + 294, + 257, + 372, + 306 + ], + "lines": [ + { + "bbox": [ + 294, + 257, + 372, + 306 + ], + "spans": [ + { + "bbox": [ + 294, + 257, + 372, + 306 + ], + "type": "image", + "image_path": "4da2f47605815447849cd4dce20013bab4102e5d1e6aeeb810dc8ae0a00ab7ee.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 307, + 359, + 314 + ], + "lines": [ + { + "bbox": [ + 306, + 307, + 359, + 314 + ], + "spans": [ + { + "bbox": [ + 306, + 307, + 359, + 314 + ], + "type": "text", + "content": "DETR on E2VID+" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 374, + 257, + 452, + 306 + ], + "blocks": [ + { + "bbox": [ + 374, + 257, + 452, + 306 + ], + "lines": [ + { + "bbox": [ + 374, + 257, + 452, + 306 + ], + "spans": [ + { + "bbox": [ + 374, + 257, + 452, + 306 + ], + "type": "image", + "image_path": "29dad1bac0afc2e0edd276e253e8e6445027fd1adde091a2b5ef4b84b7ed3e02.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 389, + 307, + 438, + 314 + ], + "lines": [ + { + "bbox": [ + 389, + 307, + 438, + 314 + ], + "spans": [ + { + "bbox": [ + 389, + 307, + 438, + 314 + ], + "type": "text", + "content": "SAM on E2VID+" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 459, + 257, + 537, + 306 + ], + "blocks": [ + { + "bbox": [ + 465, + 247, + 531, + 255 + ], + "lines": [ + { + "bbox": [ + 465, + 247, + 531, + 255 + ], + "spans": [ + { + "bbox": [ + 465, + 247, + 531, + 255 + ], + "type": "text", + "content": "85 ms long exposure" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 459, + 257, + 537, + 306 + ], + "lines": [ + { + "bbox": [ + 459, + 257, + 537, + 306 + ], + "spans": [ + { + "bbox": [ + 459, + 257, + 537, + 306 + ], + "type": "image", + "image_path": "37bff2fc404531272d5d078da8af8a496b88ada92f4ea3f8ff0678ab6b207d3b.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 461, + 307, + 535, + 314 + ], + "lines": [ + { + "bbox": [ + 461, + 307, + 535, + 314 + ], + "spans": [ + { + "bbox": [ + 461, + 307, + 535, + 314 + ], + "type": "text", + "content": "120 bps/pixel at 10 FPS" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 52, + 329, + 130, + 378 + ], + "blocks": [ + { + "bbox": [ + 52, + 329, + 130, + 378 + ], + "lines": [ + { + "bbox": [ + 52, + 329, + 130, + 378 + ], + "spans": [ + { + "bbox": [ + 52, + 329, + 130, + 378 + ], + "type": "image", + "image_path": "1a5b976082f0a1f04b8835bc367145708c26a53df2474342b3bf0f433822e0d6.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 71, + 379, + 110, + 387 + ], + "lines": [ + { + "bbox": [ + 71, + 379, + 110, + 387 + ], + "spans": [ + { + "bbox": [ + 71, + 379, + 110, + 387 + ], + "type": "text", + "content": "HRNet pose" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 132, + 329, + 211, + 378 + ], + "blocks": [ + { + "bbox": [ + 148, + 319, + 356, + 328 + ], + "lines": [ + { + "bbox": [ + 148, + 319, + 356, + 328 + ], + "spans": [ + { + "bbox": [ + 148, + 319, + 356, + 328 + ], + "type": "text", + "content": "Generalized events (Spatiotemporal chunk, Sec 4.2, 520 bps/pixel)" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 329, + 211, + 378 + ], + "lines": [ + { + "bbox": [ + 132, + 329, + 211, + 378 + ], + "spans": [ + { + "bbox": [ + 132, + 329, + 211, + 378 + ], + "type": "image", + "image_path": "635e3252d80a9ca2f675abb4ba90f20df65b0554d27600ffea6df6480e7f675f.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 149, + 379, + 194, + 387 + ], + "lines": [ + { + "bbox": [ + 149, + 379, + 194, + 387 + ], + "spans": [ + { + "bbox": [ + 149, + 379, + 194, + 387 + ], + "type": "text", + "content": "Harris corners" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 213, + 329, + 291, + 378 + ], + "blocks": [ + { + "bbox": [ + 213, + 329, + 291, + 378 + ], + "lines": [ + { + "bbox": [ + 213, + 329, + 291, + 378 + ], + "spans": [ + { + "bbox": [ + 213, + 329, + 291, + 378 + ], + "type": "image", + "image_path": "026cb2d61211598528c0c565282b3f74cfb6e3550e2278292cda54585f7fe207.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 235, + 379, + 268, + 387 + ], + "lines": [ + { + "bbox": [ + 235, + 379, + 268, + 387 + ], + "spans": [ + { + "bbox": [ + 235, + 379, + 268, + 387 + ], + "type": "text", + "content": "RAFT flow" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 294, + 329, + 372, + 378 + ], + "blocks": [ + { + "bbox": [ + 294, + 329, + 372, + 378 + ], + "lines": [ + { + "bbox": [ + 294, + 329, + 372, + 378 + ], + "spans": [ + { + "bbox": [ + 294, + 329, + 372, + 378 + ], + "type": "image", + "image_path": "89f626bfab0ac5cdc1323856ca5f0ff544864b1df2ee1d6a4027585957df98f7.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 308, + 379, + 358, + 387 + ], + "lines": [ + { + "bbox": [ + 308, + 379, + 358, + 387 + ], + "spans": [ + { + "bbox": [ + 308, + 379, + 358, + 387 + ], + "type": "text", + "content": "DETR detection" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 374, + 329, + 452, + 378 + ], + "blocks": [ + { + "bbox": [ + 374, + 329, + 452, + 378 + ], + "lines": [ + { + "bbox": [ + 374, + 329, + 452, + 378 + ], + "spans": [ + { + "bbox": [ + 374, + 329, + 452, + 378 + ], + "type": "image", + "image_path": "06b75834ea6b701472d74d09a713563ea54445a7588d21086efd89c02695ae3f.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 383, + 379, + 444, + 387 + ], + "lines": [ + { + "bbox": [ + 383, + 379, + 444, + 387 + ], + "spans": [ + { + "bbox": [ + 383, + 379, + 444, + 387 + ], + "type": "text", + "content": "SAM segmentation" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 459, + 329, + 542, + 382 + ], + "blocks": [ + { + "bbox": [ + 460, + 319, + 535, + 328 + ], + "lines": [ + { + "bbox": [ + 460, + 319, + 535, + 328 + ], + "spans": [ + { + "bbox": [ + 460, + 319, + 535, + 328 + ], + "type": "text", + "content": "0.3 ms burst exposures" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 459, + 329, + 542, + 382 + ], + "lines": [ + { + "bbox": [ + 459, + 329, + 542, + 382 + ], + "spans": [ + { + "bbox": [ + 459, + 329, + 542, + 382 + ], + "type": "image", + "image_path": "db5db6b7d05d380a522981f4079a4cce168982a9de7be8957cbcd5a5c102dff9.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 473, + 383, + 524, + 392 + ], + "lines": [ + { + "bbox": [ + 473, + 383, + 524, + 392 + ], + "spans": [ + { + "bbox": [ + 473, + 383, + 524, + 392 + ], + "type": "text", + "content": "15100 bps/pixel" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 393, + 547, + 437 + ], + "lines": [ + { + "bbox": [ + 46, + 393, + 547, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 393, + 547, + 437 + ], + "type": "text", + "content": "Figure 7. Plug-and-play inference on a tennis scene. (top left) Conventional events encode temporal-gradient polarities; this lossy representation limits performance on downstream tasks. (bottom left) Generalized events encode rich scene-intensity information, with a readout comparable existing event cameras. They facilitate high-quality plug-and-play inference, without requiring dedicated algorithms. (right) Generalized event cameras give image quality comparable to burst photography techniques that have a much higher readout rate." + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_caption" + } + ], + "index": 46 + }, + { + "bbox": [ + 47, + 447, + 184, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 447, + 184, + 460 + ], + "spans": [ + { + "bbox": [ + 47, + 447, + 184, + 460 + ], + "type": "text", + "content": "5.2. Plug-and-Play Inference" + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 46, + 470, + 288, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 470, + 288, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 470, + 288, + 601 + ], + "type": "text", + "content": "Generalized event cameras preserve scene intensity, which enables plug-and-play event-based vision. We consider a tennis sequence (of 8196 binary frames) containing a range of object speeds. We evaluate a range of tasks: pose estimation (HRNet [59]), corner detection [18], optical flow (RAFT [61]), object detection (DETR [11]), and segmentation (SAM [32]). We compare against event-based methods applied to Prophesee events; we use " + }, + { + "bbox": [ + 46, + 470, + 288, + 601 + ], + "type": "inline_equation", + "content": "\\mathrm{Arc}^*" + }, + { + "bbox": [ + 46, + 470, + 288, + 601 + ], + "type": "text", + "content": " [3] for corner detection and E-RAFT [16] for optical flow. For the remaining tasks, which do not have equivalent event methods, we run HRNet, DETR, and SAM on E2VID+ reconstructions." + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 46, + 605, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 289, + 714 + ], + "type": "text", + "content": "As Fig. 7 (top left) shows, traditional events are bandwidth efficient (331 bps/pixel), but do not provide sufficient information for successful inference. Generalized events (bottom left) have a modestly higher readout (520 bps/pixel), but support accurate inference without requiring dedicated algorithms. To provide context for these rates, we compare them against frame-based methods (right). A long exposure (120 bps/pixel) blurs out the ratchet. Burst methods [19] recover a sharp image from a stack of short exposures, but with a large readout of 15100 bps/pixel." + } + ] + } + ], + "index": 51 + }, + { + "bbox": [ + 306, + 447, + 444, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 447, + 444, + 460 + ], + "spans": [ + { + "bbox": [ + 306, + 447, + 444, + 460 + ], + "type": "text", + "content": "5.3. Rate-Distortion Analysis" + } + ] + } + ], + "index": 52 + }, + { + "bbox": [ + 304, + 468, + 547, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 468, + 547, + 612 + ], + "spans": [ + { + "bbox": [ + 304, + 468, + 547, + 612 + ], + "type": "text", + "content": "Each method in Sec. 4 features a sensitivity parameter that controls the sensor readout rate (event rate), which in turn influences image quality. In this subsection, we evaluate the impact of readout on image quality (PSNR) by performing a rate-distortion analysis. For ground truth, we use a set of YouTube-sourced high-speed videos captured by a Phantom Flex4k at 1000 FPS; see the supplement for thumbnails and links. We upsample these videos to the SPAD's frame rate and then simulate 4096 binary frames using the image formation model described in Eq. (4). When computing readout for our methods, we assume that events encode 10-bit values and account for the header bits of each event packet." + } + ] + } + ], + "index": 53 + }, + { + "bbox": [ + 304, + 615, + 547, + 697 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 615, + 547, + 697 + ], + "spans": [ + { + "bbox": [ + 304, + 615, + 547, + 697 + ], + "type": "text", + "content": "As baselines, we consider EDI++, a long exposure, compressive sensing with 8-bucket masks, and burst denoising [19] using 32 short exposures. As Fig. 8 shows, generalized event cameras provide a pronounced 4-8 dB PSNR improvement over baseline methods. Further, our methods can compress the raw SPAD response by around " + }, + { + "bbox": [ + 304, + 615, + 547, + 697 + ], + "type": "inline_equation", + "content": "80\\times" + }, + { + "bbox": [ + 304, + 615, + 547, + 697 + ], + "type": "text", + "content": " before a noticeable drop-off in PSNR is observed." + } + ] + } + ], + "index": 54 + }, + { + "bbox": [ + 318, + 701, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 546, + 714 + ], + "type": "text", + "content": "Among our methods, the spatiotemporal chunk approach of" + } + ] + } + ], + "index": 55 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25013" + } + ] + } + ], + "index": 56 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 281, + 158 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 281, + 158 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 281, + 158 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 281, + 158 + ], + "type": "image", + "image_path": "fd6c6fe236e8afa729df60d5beaf108f5aaa1cd28b3144be1b37cc8978f9e5d6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 165, + 288, + 209 + ], + "lines": [ + { + "bbox": [ + 46, + 165, + 288, + 209 + ], + "spans": [ + { + "bbox": [ + 46, + 165, + 288, + 209 + ], + "type": "text", + "content": "Figure 8. Rate-distortion evaluation. Our techniques feature a tunable parameter that controls the output event rate. Generalized events offer a " + }, + { + "bbox": [ + 46, + 165, + 288, + 209 + ], + "type": "inline_equation", + "content": "4 - 8\\mathrm{dB}" + }, + { + "bbox": [ + 46, + 165, + 288, + 209 + ], + "type": "text", + "content": " improvement in PSNR over " + }, + { + "bbox": [ + 46, + 165, + 288, + 209 + ], + "type": "inline_equation", + "content": "\\mathrm{EDI} + +" + }, + { + "bbox": [ + 46, + 165, + 288, + 209 + ], + "type": "text", + "content": " (at the same readout), and can compress raw photon data by " + }, + { + "bbox": [ + 46, + 165, + 288, + 209 + ], + "type": "inline_equation", + "content": "80\\times" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 217, + 290, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 217, + 290, + 385 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 290, + 385 + ], + "type": "text", + "content": "Sec. 4.2 gives the best PSNR, followed by the Bayesian method (Sec. 4.1) and coded-exposure events (Sec. 4.3). That said, all methods are fairly similar in terms of rate-distortion (e.g., all three give comparable results for the scenes in Secs. 5.1 and 5.2). The methods are better distinguished by their practical characteristics. The Bayesian method gives single-photon temporal resolution; however, as we show in Sec. 5.4, it is the most expensive to compute on-chip. The chunk-based method occupies a middle ground in terms of latency and cost. Coded-exposure events have the highest latency—events are generated only every " + }, + { + "bbox": [ + 46, + 217, + 290, + 385 + ], + "type": "inline_equation", + "content": "\\sim 256 - 512" + }, + { + "bbox": [ + 46, + 217, + 290, + 385 + ], + "type": "text", + "content": " binary frames—but the lowest on-chip cost. This provides an end user the flexibility to choose from the space of generalized event cameras based on the latency requirements and the compute constraints of the target application." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 394, + 234, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 394, + 234, + 407 + ], + "spans": [ + { + "bbox": [ + 47, + 394, + 234, + 407 + ], + "type": "text", + "content": "5.4. On-Chip Feasibility and Validation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 413, + 287, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 413, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 287, + 509 + ], + "type": "text", + "content": "A critical limitation of single-photon sensors is the exorbitant bandwidth and power costs involved in reading off raw photon detections. However, the lightweight nature of our event camera designs allows us to sidestep this limitation by performing computations on-chip. We demonstrate that our methods are feasible on UltraPhase [4], a SPAD compute platform. UltraPhase consists of " + }, + { + "bbox": [ + 46, + 413, + 287, + 509 + ], + "type": "inline_equation", + "content": "3 \\times 6" + }, + { + "bbox": [ + 46, + 413, + 287, + 509 + ], + "type": "text", + "content": " compute cores, each of which is associated with " + }, + { + "bbox": [ + 46, + 413, + 287, + 509 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 46, + 413, + 287, + 509 + ], + "type": "text", + "content": " pixels." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 510, + 288, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 288, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 288, + 604 + ], + "type": "text", + "content": "We implement our methods for UltraPhase using custom assembly code. Some methods require minor modifications due to instruction-set limitations; see the supplement for details. We process 2500 SPAD frames from the tennis sequence used in Sec. 5.2, cropped to the UltraPhase array size of " + }, + { + "bbox": [ + 46, + 510, + 288, + 604 + ], + "type": "inline_equation", + "content": "12 \\times 24" + }, + { + "bbox": [ + 46, + 510, + 288, + 604 + ], + "type": "text", + "content": " pixels. We determine the number of cycles required to execute the assembly code and estimate the chip's power consumption and readout bandwidth." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "All our proposed methods run comfortably within the chip's compute budget of 4202 cycles per binary frame and its memory limit of 4 Kbit per core. As seen in Fig. 9, compared to raw photon-detection readout, our techniques reduce both bandwidth and power costs by over two orders of magnitude. The coded-exposure method is particularly efficient; on most binary frames, it only requires multiplying a binary code with incident photon detections. Our proof-of-concept evaluation may pave the way for future near-sensor implementations of generalized event" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 313, + 72, + 537, + 147 + ], + "blocks": [ + { + "bbox": [ + 313, + 72, + 537, + 147 + ], + "lines": [ + { + "bbox": [ + 313, + 72, + 537, + 147 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 537, + 147 + ], + "type": "image", + "image_path": "c9725d942b2555ce16b8047a70f5fe65a0e5954f75454606192abb0289916fc5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 151, + 546, + 196 + ], + "lines": [ + { + "bbox": [ + 305, + 151, + 546, + 196 + ], + "spans": [ + { + "bbox": [ + 305, + 151, + 546, + 196 + ], + "type": "text", + "content": "Figure 9. On-chip compatibility. We validate the feasibility of our approach on UltraPhase [4], a computational SPAD imager. Compared to reading out raw photon data, all of our approaches give marked reductions in both bandwidth (left) and power (right)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 310, + 203, + 385, + 251 + ], + "blocks": [ + { + "bbox": [ + 310, + 203, + 385, + 251 + ], + "lines": [ + { + "bbox": [ + 310, + 203, + 385, + 251 + ], + "spans": [ + { + "bbox": [ + 310, + 203, + 385, + 251 + ], + "type": "image", + "image_path": "c57aa0525b99732632c97a11a8593ca21560f0f2f04143de57f04e9e16a9fb5c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 316, + 251, + 380, + 269 + ], + "lines": [ + { + "bbox": [ + 316, + 251, + 380, + 269 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 380, + 269 + ], + "type": "text", + "content": "Rigid dynamics (200x compression)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 273, + 547, + 339 + ], + "lines": [ + { + "bbox": [ + 305, + 273, + 547, + 339 + ], + "spans": [ + { + "bbox": [ + 305, + 273, + 547, + 339 + ], + "type": "text", + "content": "Figure 10. Limitations and failure modes, (left) Our reconstructions (yellow inset) on dynamic scenes with rigid objects can be inferior to burst photography (green inset). (center) Modulated light sources, such as this phone screen, can trigger a deluge of events (change points shown in the inset). (right) Rapid camera motion can result in an event rate divergent from scene dynamics." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 387, + 204, + 463, + 251 + ], + "blocks": [ + { + "bbox": [ + 387, + 204, + 463, + 251 + ], + "lines": [ + { + "bbox": [ + 387, + 204, + 463, + 251 + ], + "spans": [ + { + "bbox": [ + 387, + 204, + 463, + 251 + ], + "type": "image", + "image_path": "37e4fe3c13d231503f0a2500c081e897e30bd89ef6b87af2b8d9166a8d42479a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 395, + 251, + 455, + 269 + ], + "lines": [ + { + "bbox": [ + 395, + 251, + 455, + 269 + ], + "spans": [ + { + "bbox": [ + 395, + 251, + 455, + 269 + ], + "type": "text", + "content": "Phone screen (11x compression)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 465, + 204, + 540, + 251 + ], + "blocks": [ + { + "bbox": [ + 465, + 204, + 540, + 251 + ], + "lines": [ + { + "bbox": [ + 465, + 204, + 540, + 251 + ], + "spans": [ + { + "bbox": [ + 465, + 204, + 540, + 251 + ], + "type": "image", + "image_path": "e78523f46441ba8ae4f55eff71c3b62698bf81b86c0f70486a601ce9108e1be9.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 471, + 251, + 534, + 269 + ], + "lines": [ + { + "bbox": [ + 471, + 251, + 534, + 269 + ], + "spans": [ + { + "bbox": [ + 471, + 251, + 534, + 269 + ], + "type": "text", + "content": "Camera motion (130x compression)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 346, + 547, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 346, + 547, + 382 + ], + "spans": [ + { + "bbox": [ + 305, + 346, + 547, + 382 + ], + "type": "text", + "content": "cameras, which with advances in chip-to-chip communication, could involve a dedicated \"photon processing unit\", similar to a camera image signal processor (ISP)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 392, + 451, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 392, + 451, + 404 + ], + "spans": [ + { + "bbox": [ + 306, + 392, + 451, + 404 + ], + "type": "text", + "content": "6. Limitations and Discussion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 412, + 547, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 412, + 547, + 520 + ], + "spans": [ + { + "bbox": [ + 305, + 412, + 547, + 520 + ], + "type": "text", + "content": "Generalized events push the frontiers of event-based imaging; however, some scenarios lead to sub-optimal performance. As seen in Fig. 10 (left), if the scene dynamics is entirely comprised of rigid motion, burst photography [36] gives better image quality, albeit with much higher readout. (middle) Similar to current event cameras, modulated light sources trigger unwanted events that reduce bandwidth savings. However, it may be possible to ignore some of these events, perhaps by modeling the lighting variations [55]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 523, + 547, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 523, + 547, + 619 + ], + "spans": [ + { + "bbox": [ + 305, + 523, + 547, + 619 + ], + "type": "text", + "content": "Ego-motion events. Camera motion can trigger events in static regions, although our methods still yield substantial compression " + }, + { + "bbox": [ + 305, + 523, + 547, + 619 + ], + "type": "inline_equation", + "content": "(130 \\times" + }, + { + "bbox": [ + 305, + 523, + 547, + 619 + ], + "type": "text", + "content": " over SPAD outputs, Fig. 10 right). We analyze the impact of ego-motion on bandwidth savings further in the supplement. However, single-photon cameras can emulate sensor motion by integrating flux along alternate spatiotemporal trajectories [60]. We can imagine a generalized event camera that is \"ego-motion compensated,\" by computing events along a suitable trajectory." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 622, + 547, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 622, + 547, + 705 + ], + "spans": [ + { + "bbox": [ + 305, + 622, + 547, + 705 + ], + "type": "text", + "content": "Photon-stream compression. SPADs generate a torrent of data—e.g., 12.5 GBps for a MPixel array at " + }, + { + "bbox": [ + 305, + 622, + 547, + 705 + ], + "type": "inline_equation", + "content": "100\\mathrm{kHz}" + }, + { + "bbox": [ + 305, + 622, + 547, + 705 + ], + "type": "text", + "content": "—that can easily overwhelm data interfaces. Generalized event cameras reduce readout by around two orders of magnitude, by decoupling readout from the SPAD's frame rate and instead basing it on scene dynamics. This could pave the way for practical, high-resolution single-photon sensors." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25014" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 104, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 104, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 104, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 705 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 113 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 113 + ], + "type": "text", + "content": "[1] Ryan Prescott Adams and David JC MacKay. Bayesian online changepoint detection. arXiv preprint arXiv:0710.3742, 2007. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 113, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 113, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 113, + 287, + 156 + ], + "type": "text", + "content": "[2] Reda Alami, Odalric Maillard, and Raphael Feraud. Restarted bayesian online change-point detector achieves optimal detection delay. In International conference on machine learning, pages 211-221. PMLR, 2020. 4, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 157, + 288, + 189 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 157, + 288, + 189 + ], + "spans": [ + { + "bbox": [ + 53, + 157, + 288, + 189 + ], + "type": "text", + "content": "[3] Ignacio Alzugaray and Margarita Chli. Asynchronous corner detection and tracking for event cameras in real time. IEEE Robotics and Automation Letters, 3(4):3177-3184, 2018. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 190, + 287, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 190, + 287, + 212 + ], + "spans": [ + { + "bbox": [ + 53, + 190, + 287, + 212 + ], + "type": "text", + "content": "[4] Andrei Ardelean. Computational Imaging SPAD Cameras. PhD thesis, École polytechnique fédérale de Lausanne, 2023. 2, 6, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 212, + 288, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 212, + 288, + 244 + ], + "spans": [ + { + "bbox": [ + 53, + 212, + 288, + 244 + ], + "type": "text", + "content": "[5] Souptik Barua, Yoshitaka Miyatani, and Ashok Veeraraghavan. Direct face detection and video reconstruction from event cameras. In WACV, pages 1-9. IEEE, 2016. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 245, + 288, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 245, + 288, + 298 + ], + "spans": [ + { + "bbox": [ + 53, + 245, + 288, + 298 + ], + "type": "text", + "content": "[6] Raphael Berner, Christian Brandli, Minhao Yang, S-C Liu, and Tobi Delbruck. A 240x180 120dB 10mW 12us-latency sparse output vision sensor for mobile applications. In Proceedings of the International Image Sensors Workshop, number CONF, pages 41-44, 2013. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 300, + 288, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 300, + 288, + 354 + ], + "spans": [ + { + "bbox": [ + 53, + 300, + 288, + 354 + ], + "type": "text", + "content": "[7] Anthony Bisulco, Fernando Cladera Ojeda, Volkan Isler, and Daniel D Lee. Fast motion understanding with spatiotemporal neural networks and dynamic vision sensors. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 14098-14104. IEEE, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 354, + 288, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 354, + 288, + 387 + ], + "spans": [ + { + "bbox": [ + 53, + 354, + 288, + 387 + ], + "type": "text", + "content": "[8] Assim Boukhayma, Arnaud Peizerat, and Christian Enz. A sub-0.5 electron read noise VGA image sensor in a standard CMOS process. IEEE Journal of Solid-State Circuits, 2016. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 388, + 288, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 388, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 388, + 288, + 430 + ], + "type": "text", + "content": "[9] Christian Brandli, Raphael Berner, Minhao Yang, Shih-Chii Liu, and Tobi Delbruck. A " + }, + { + "bbox": [ + 53, + 388, + 288, + 430 + ], + "type": "inline_equation", + "content": "240 \\times 180" + }, + { + "bbox": [ + 53, + 388, + 288, + 430 + ], + "type": "text", + "content": " dB " + }, + { + "bbox": [ + 53, + 388, + 288, + 430 + ], + "type": "inline_equation", + "content": "3\\mu s" + }, + { + "bbox": [ + 53, + 388, + 288, + 430 + ], + "type": "text", + "content": " latency global shutter spatiotemporal vision sensor. IEEE Journal of Solid-State Circuits, 49(10):2333-2341, 2014. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 431, + 288, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 431, + 288, + 474 + ], + "spans": [ + { + "bbox": [ + 48, + 431, + 288, + 474 + ], + "type": "text", + "content": "[10] Christian Brandli, Lorenz Muller, and Tobi Delbruck. Realtime, high-speed video decompression using a frame-and event-based DAVIS sensor. In 2014 IEEE International Symposium on Circuits and Systems (ISCAS), pages 686-689. IEEE, 2014. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 475, + 288, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 475, + 288, + 529 + ], + "spans": [ + { + "bbox": [ + 48, + 475, + 288, + 529 + ], + "type": "text", + "content": "[11] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Computer Vision – ECCV 2020, pages 213–229, Cham, 2020. Springer International Publishing. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 529, + 288, + 562 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 529, + 288, + 562 + ], + "spans": [ + { + "bbox": [ + 48, + 529, + 288, + 562 + ], + "type": "text", + "content": "[12] Andrea Censi and Davide Scaramuzza. Low-latency event-based visual odometry. In 2014 IEEE International Conference on Robotics and Automation (ICRA), pages 703-710. IEEE, 2014. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 563, + 288, + 605 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 563, + 288, + 605 + ], + "spans": [ + { + "bbox": [ + 48, + 563, + 288, + 605 + ], + "type": "text", + "content": "[13] Matthew Cook, Luca Gugelmann, Florian Jug, Christoph Krautz, and Angelika Steger. Interacting maps for fast visual interpretation. In The 2011 International Joint Conference on Neural Networks, pages 770-776, 2011. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 606, + 288, + 639 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 606, + 288, + 639 + ], + "spans": [ + { + "bbox": [ + 48, + 606, + 288, + 639 + ], + "type": "text", + "content": "[14] Yue Gao, Siqi Li, Yipeng Li, Yandong Guo, and Qionghai Dai. Superfast: " + }, + { + "bbox": [ + 48, + 606, + 288, + 639 + ], + "type": "inline_equation", + "content": "200 \\times" + }, + { + "bbox": [ + 48, + 606, + 288, + 639 + ], + "type": "text", + "content": " video frame interpolation via event camera. IEEE TPAMI, 45(6):7764-7780, 2023. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 639, + 288, + 672 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 639, + 288, + 672 + ], + "spans": [ + { + "bbox": [ + 48, + 639, + 288, + 672 + ], + "type": "text", + "content": "[15] Daniel Gehrig, Henri Rebecq, Guillermo Gallego, and Davide Scaramuzza. EKLT: Asynchronous photometric feature tracking using events and frames. IJCV, 128(3):601-618, 2020. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 672, + 288, + 705 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 672, + 288, + 705 + ], + "spans": [ + { + "bbox": [ + 48, + 672, + 288, + 705 + ], + "type": "text", + "content": "[16] Mathias Gehrig, Mario Millhäuser, Daniel Gehrig, and Davide Scaramuzza. E-raft: Dense optical flow from event cameras. In 2021 International Conference on 3D Vision (3DV), pages" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 709 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 326, + 73, + 411, + 83 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 411, + 83 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 411, + 83 + ], + "type": "text", + "content": "197-206.IEEE,2021.7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 84, + 547, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 84, + 547, + 116 + ], + "spans": [ + { + "bbox": [ + 308, + 84, + 547, + 116 + ], + "type": "text", + "content": "[17] Istvan Gyongy, Neale AW Dutton, and Robert K Henderson. Single-photon tracking for high-speed vision. Sensors, 18(2):323, 2018. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 118, + 547, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 547, + 148 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 547, + 148 + ], + "type": "text", + "content": "[18] Chris Harris, Mike Stephens, et al. A combined corner and edge detector. In *Alvey vision conference*, pages 10-5244. CiteSeer, 1988. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 150, + 547, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 150, + 547, + 193 + ], + "spans": [ + { + "bbox": [ + 308, + 150, + 547, + 193 + ], + "type": "text", + "content": "[19] Samuel W. Hasinoff, Dillon Sharlet, Ryan Geiss, Andrew Adams, Jonathan T. Barron, Florian Kainz, Jiawen Chen, and Marc Levoy. Burst photography for high dynamic range and low-light imaging on mobile cameras. ACM TOG, 35(6):1-12, 2016. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 194, + 547, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 194, + 547, + 248 + ], + "spans": [ + { + "bbox": [ + 308, + 194, + 547, + 248 + ], + "type": "text", + "content": "[20] Botao He, Haojia Li, Siyuan Wu, Dong Wang, Zhiwei Zhang, Qianli Dong, Chao Xu, and Fei Gao. Fast-dynamic-vision: Detection and tracking dynamic objects with event and depth sensing. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3071-3078. IEEE, 2021. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 249, + 547, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 249, + 547, + 281 + ], + "spans": [ + { + "bbox": [ + 308, + 249, + 547, + 281 + ], + "type": "text", + "content": "[21] Javier Hidalgo-Carrio, Guillermo Gallego, and Davide Scaramuzza. Event-aided direct sparse odometry. In CVPR, pages 5781-5790, 2022. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 281, + 547, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 281, + 547, + 335 + ], + "spans": [ + { + "bbox": [ + 308, + 281, + 547, + 335 + ], + "type": "text", + "content": "[22] Yasunobu Hitomi, Jinwei Gu, Mohit Gupta, Tomoo Mitsunaga, and Shree K. Nayar. Video from a single coded exposure photograph using a learned over-complete dictionary. In 2011 International Conference on Computer Vision, pages 287-294, 2011. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 335, + 547, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 335, + 547, + 380 + ], + "spans": [ + { + "bbox": [ + 308, + 335, + 547, + 380 + ], + "type": "text", + "content": "[23] Jing Huang, Menghan Guo, and Shoushun Chen. A dynamic vision sensor with direct logarithmic output and full-frame picture-on-demand. In 2017 IEEE International Symposium on Circuits and Systems (ISCAS), pages 1-4, 2017. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 380, + 546, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 380, + 546, + 423 + ], + "spans": [ + { + "bbox": [ + 308, + 380, + 546, + 423 + ], + "type": "text", + "content": "[24] Zhewei Huang, Tianyuan Zhang, Wen Heng, Boxin Shi, and Shuchang Zhou. Real-time intermediate flow estimation for video frame interpolation. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 424, + 546, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 424, + 546, + 456 + ], + "spans": [ + { + "bbox": [ + 308, + 424, + 546, + 456 + ], + "type": "text", + "content": "[25] Jorge Igual. Photographic noise performance measures based on raw files analysis of consumer cameras. *Electronics*, 8(11):1284, 2019. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 456, + 545, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 456, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 308, + 456, + 545, + 479 + ], + "type": "text", + "content": "[26] Atul Ingle, Andreas Velten, and Mohit Gupta. High Flux Passive Imaging With Single-Photon Sensors. In CVPR, 2019. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 479, + 547, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 479, + 547, + 511 + ], + "spans": [ + { + "bbox": [ + 308, + 479, + 547, + 511 + ], + "type": "text", + "content": "[27] Atul Ingle, Trevor Seets, Mauro Buttafava, Shantanu Gupta, Alberto Tosi, Mohit Gupta, and Andreas Velten. Passive inter-photon imaging. In CVPR, 2021. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 512, + 547, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 512, + 547, + 554 + ], + "spans": [ + { + "bbox": [ + 308, + 512, + 547, + 554 + ], + "type": "text", + "content": "[28] Kiyotaka Iwabuchi, Yusuke Kameda, and Takayuki Hamamoto. Image quality improvements based on motion-based deblurring for single-photon imaging. IEEE Access, 9:30080-30094, 2021. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 555, + 546, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 555, + 546, + 587 + ], + "spans": [ + { + "bbox": [ + 308, + 555, + 546, + 587 + ], + "type": "text", + "content": "[29] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. CoTracker: It is better to track together. 2023." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 588, + 546, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 588, + 546, + 610 + ], + "spans": [ + { + "bbox": [ + 308, + 588, + 546, + 610 + ], + "type": "text", + "content": "[30] Anthony Kay. Tesseract: an open-source optical character recognition engine. *Linux Journal*, 2007(159):2, 2007." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 611, + 546, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 611, + 546, + 643 + ], + "spans": [ + { + "bbox": [ + 308, + 611, + 546, + 643 + ], + "type": "text", + "content": "[31] Hanme Kim, Ankur Handa, Ryad Benosman, Sio-Hoi Ieng, and Andrew J Davison. Simultaneous mosaicing and tracking with an event camera. Proc. Brit. Mach. Vis. Conf, 43:566-576, 2014. 3" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 643, + 546, + 687 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 643, + 546, + 687 + ], + "spans": [ + { + "bbox": [ + 308, + 643, + 546, + 687 + ], + "type": "text", + "content": "[32] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólár, and Ross Girshick. Segment anything, 2023. 7" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 308, + 688, + 546, + 709 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 688, + 546, + 709 + ], + "spans": [ + { + "bbox": [ + 308, + 688, + 546, + 709 + ], + "type": "text", + "content": "[33] Haolong Li and Joerg Stueckler. Tracking 6-DoF object motion from events and frames. In 2021 IEEE International Conference" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25015" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 708 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 66, + 72, + 288, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 288, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 288, + 95 + ], + "type": "text", + "content": "on Robotics and Automation (ICRA), pages 14171-14177. IEEE, 2021. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 288, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 288, + 126 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 288, + 126 + ], + "type": "text", + "content": "[34] P Lichtsteiner. 64x64 event-driven logarithmic temporal derivative silicon retina. In Program 2003 IEEE Workshop on CCD and AIS, 2003. 1, 2, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 129, + 288, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 160 + ], + "type": "text", + "content": "[35] Yuhao Liu, Felipe Gutierrez-Barragan, Atul Ingle, Mohit Gupta, and Andreas Velten. Single-photon camera guided extreme dynamic range imaging. In WACV, pages 1575-1585, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 162, + 288, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 288, + 193 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 288, + 193 + ], + "type": "text", + "content": "[36] Sizhuo Ma, Shantanu Gupta, Arin C. Ulku, Claudio Bruschini, Edoardo Charbon, and Mohit Gupta. Quanta burst photography. ACM TOG, 39(4):1-16, 2020. 3, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 194, + 288, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 194, + 288, + 225 + ], + "spans": [ + { + "bbox": [ + 48, + 194, + 288, + 225 + ], + "type": "text", + "content": "[37] Sizhuo Ma, Paul Mos, Edoardo Charbon, and Mohit Gupta. Burst vision using single-photon cameras. In WACV, pages 5375-5385, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 227, + 288, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 227, + 288, + 270 + ], + "spans": [ + { + "bbox": [ + 48, + 227, + 288, + 270 + ], + "type": "text", + "content": "[38] Kazuhiro Morimoto, Andrei Ardelean, Ming-Lo Wu, Arin Can Ulku, Ivan Michel Antolovic, Claudio Bruschini, and Edoardo Charbon. Megapixel time-gated SPAD image sensor for 2D and 3D imaging applications. Optica, 7(4):346-354, 2020. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 271, + 288, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 271, + 288, + 303 + ], + "spans": [ + { + "bbox": [ + 48, + 271, + 288, + 303 + ], + "type": "text", + "content": "[39] Gottfried Munda, Christian Reinbacher, and Thomas Pock. Realtime intensity-image reconstruction for event cameras using manifold regularisation. *IJCV*, 126:1381–1393, 2018. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 304, + 288, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 304, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 288, + 357 + ], + "type": "text", + "content": "[40] Shuichi Namiki, Shunichi Sato, Yusuke Kameda, and Takayuki Hamamoto. Imaging method using multi-threshold pattern for photon detection of quanta image sensor. In International Workshop on Advanced Imaging Technology (IWAIT) 2022, page 1217702. SPIE, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 358, + 288, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 358, + 288, + 390 + ], + "spans": [ + { + "bbox": [ + 48, + 358, + 288, + 390 + ], + "type": "text", + "content": "[41] Liyuan Pan, Cedric Scheerlinck, Xin Yu, Richard Hartley, Miao-miao Liu, and Yuchao Dai. Bringing a blurry frame alive at high frame-rate with an event camera. In CVPR, 2019. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 392, + 288, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 392, + 288, + 434 + ], + "spans": [ + { + "bbox": [ + 48, + 392, + 288, + 434 + ], + "type": "text", + "content": "[42] Liyuan Pan, Richard Hartley, Cedric Scheerlinck, Miaomiao Liu, Xin Yu, and Yuchao Dai. High frame rate video reconstruction based on an event camera. IEEE TPAMI, 44(5):2519-2533, 2022. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 435, + 288, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 435, + 288, + 478 + ], + "spans": [ + { + "bbox": [ + 48, + 435, + 288, + 478 + ], + "type": "text", + "content": "[43] Federico Paredes-Valles and Guido C. H. E. de Croon. Back to event basics: Self-supervised learning of image reconstruction for event cameras via photometric constancy. In CVPR, pages 3446–3455, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 479, + 288, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 479, + 288, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 479, + 288, + 521 + ], + "type": "text", + "content": "[44] Christoph Posch, Daniel Matolin, and Rainer Wohlgenannt. A QVGA 143 dB dynamic range frame-free pwm image sensor with lossless pixel-level video compression and time-domain cds. IEEE Journal of Solid-State Circuits, 46(1):259-275, 2011. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 523, + 288, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 288, + 555 + ], + "type": "text", + "content": "[45] Ramesh Raskar, Amit Agrawal, and Jack Tumblin. Coded exposure photography: motion deblurring using fluttered shutter. In Acm Siggraph 2006 Papers, pages 795-804. 2006. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 556, + 288, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 556, + 288, + 587 + ], + "spans": [ + { + "bbox": [ + 48, + 556, + 288, + 587 + ], + "type": "text", + "content": "[46] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. High speed and high dynamic range video with an event camera. IEEE TPAMI, 2019. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 589, + 288, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 589, + 288, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 589, + 288, + 621 + ], + "type": "text", + "content": "[47] Dikpal Reddy, Ashok Veeraraghavan, and Rama Chellappa. P2C2: Programmable pixel compressive camera for high speed imaging. In CVPR 2011, pages 329-336, 2011. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 622, + 288, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 622, + 288, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 622, + 288, + 643 + ], + "type": "text", + "content": "[48] Alexis Rochas. Single photon avalanche diodes in cmos technology. Technical report, Citeseer, 2003. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 643, + 288, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 643, + 288, + 708 + ], + "spans": [ + { + "bbox": [ + 48, + 643, + 288, + 708 + ], + "type": "text", + "content": "[49] Nitin J Sanket, Chethan M Parameshwara, Chahat Deep Singh, Ashwin V Kuruttukulam, Cornelia Fermüller, Davide Scaramuzza, and Yiannis Aloimonos. Evdodgenet: Deep dynamic obstacle dodging with event cameras. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 10651-10657. IEEE, 2020. 1" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 708 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "type": "text", + "content": "[50] Cedric Scheerlinck, Nick Barnes, and Robert Mahony. Continuous-time intensity estimation using event cameras. In ACCV, pages 308-324. Springer, 2018. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 107, + 547, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 547, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 547, + 138 + ], + "type": "text", + "content": "[51] Cedric Scheerlinck, Henri Rebecq, Daniel Gehrig, Nick Barnes, Robert Mahony, and Davide Scaramuzza. Fast image reconstruction with an event camera. In WACV, 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 140, + 547, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 547, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 547, + 171 + ], + "type": "text", + "content": "[52] Trevor Seets, Atul Ingle, Martin Laurenzis, and Andreas Velten. Motion adaptive deblurring with single-photon cameras. In WACV, pages 1945-1954, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 172, + 547, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 172, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 172, + 547, + 205 + ], + "type": "text", + "content": "[53] Prasan Shedligeri and Kaushik Mitra. Photorealistic image reconstruction from hybrid intensity and event-based sensor. Journal of Electronic Imaging, 28(6):063012-063012, 2019. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 205, + 546, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 205, + 546, + 237 + ], + "spans": [ + { + "bbox": [ + 307, + 205, + 546, + 237 + ], + "type": "text", + "content": "[54] Prasan Shedligeri, Anupama S, and Kaushik Mitra. A unified framework for compressive video recovery from coded exposure techniques. In WACV, pages 1600-1609, 2021. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 238, + 546, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 238, + 546, + 259 + ], + "spans": [ + { + "bbox": [ + 307, + 238, + 546, + 259 + ], + "type": "text", + "content": "[55] Mark Sheinin, Yoav Y. Schechner, and Kiriakos N. Kutulakos. Computational imaging on the electric grid. In CVPR, 2017. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 260, + 546, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 260, + 546, + 303 + ], + "spans": [ + { + "bbox": [ + 307, + 260, + 546, + 303 + ], + "type": "text", + "content": "[56] Hyeunjun Sim, Jihyong Oh, and Munchurl Kim. XVFI: extreme video frame interpolation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14489-14498, 2021. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 304, + 546, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 304, + 546, + 346 + ], + "spans": [ + { + "bbox": [ + 307, + 304, + 546, + 346 + ], + "type": "text", + "content": "[57] Timo Stoffregen, Cedric Scheerlinck, Davide Scaramuzza, Tom Drummond, Nick Barnes, Lindsay Kleeman, and Robert Mahony. Reducing the sim-to-real gap for event cameras. In ECCV, pages 534-549. Springer, 2020. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 347, + 546, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 347, + 546, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 347, + 546, + 380 + ], + "type": "text", + "content": "[58] Binyi su, Lei Yu, and Wen Yang. Event-based high frame-rate video reconstruction with a novel cycle-event network. In ICIP, pages 86-90, 2020. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 380, + 546, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 380, + 546, + 423 + ], + "spans": [ + { + "bbox": [ + 307, + 380, + 546, + 423 + ], + "type": "text", + "content": "[59] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 424, + 546, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 424, + 546, + 466 + ], + "spans": [ + { + "bbox": [ + 307, + 424, + 546, + 466 + ], + "type": "text", + "content": "[60] Varun Sundar, Andrei Ardelean, Tristan Swedish, Claudio Bruschini, Edoardo Charbon, and Mohit Gupta. Sodacam: Software-defined cameras via single-photon imaging. In ICCV, pages 8165-8176, 2023. 3, 6, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 468, + 546, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 468, + 546, + 500 + ], + "spans": [ + { + "bbox": [ + 307, + 468, + 546, + 500 + ], + "type": "text", + "content": "[61] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision – ECCV 2020, pages 402–419, Cham, 2020. Springer International Publishing. 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 501, + 546, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 501, + 546, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 501, + 546, + 555 + ], + "type": "text", + "content": "[62] Arin Can Ulku, Claudio Bruschini, Ivan Michel Antolovic, Yung Kuo, Rinet Ankri, Shimon Weiss, Xavier Michalet, and Edoardo Charbon. A " + }, + { + "bbox": [ + 307, + 501, + 546, + 555 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 307, + 501, + 546, + 555 + ], + "type": "text", + "content": " SPAD Image Sensor With Integrated Gating for Widefield FLIM. IEEE Journal of Selected Topics in Quantum Electronics, 25(1):1-12, 2019. 2, 4, 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 556, + 546, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 556, + 546, + 598 + ], + "spans": [ + { + "bbox": [ + 307, + 556, + 546, + 598 + ], + "type": "text", + "content": "[63] Lishun Wang, Miao Cao, and Xin Yuan. Efficientsci: Densely connected network with space-time factorization for large-scale video snapshot compressive imaging. In CVPR, pages 18477-18486, 2023. 6" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 600, + 546, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 600, + 546, + 642 + ], + "spans": [ + { + "bbox": [ + 307, + 600, + 546, + 642 + ], + "type": "text", + "content": "[64] Zhaohui Wang, Xiao Lin, Abhinav Mishra, and Ram Sriharsha. Online changepoint detection on a budget. In 2021 International Conference on Data Mining Workshops (ICDMW), pages 414-420. IEEE, 2021. 5" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 643, + 546, + 676 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 643, + 546, + 676 + ], + "spans": [ + { + "bbox": [ + 307, + 643, + 546, + 676 + ], + "type": "text", + "content": "[65] Mian Wei, Sotiris Nousias, Rahul Gulve, David B. Lindell, and Kiriakos N. Kutulakos. Passive ultra-wideband single-photon imaging. In ICCV, pages 8135-8146, 2023. 3" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 677, + 546, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 677, + 546, + 708 + ], + "spans": [ + { + "bbox": [ + 307, + 677, + 546, + 708 + ], + "type": "text", + "content": "[66] Xin Yuan, Yang Liu, Jinli Suo, Frédo Durand, and Qionghai Dai. Plug-and-play algorithms for video snapshot compressive imaging. IEEE Transactions on Pattern Analysis and Machine" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "25016" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 66, + 72, + 208, + 83 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 208, + 83 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 208, + 83 + ], + "type": "text", + "content": "Intelligence, 44(10):7093-7111, 2022. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 84, + 288, + 183 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 48, + 84, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 84, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 84, + 287, + 127 + ], + "type": "text", + "content": "[67] Zelin Zhang, Anthony J Yezzi, and Guillermo Gallego. Formulating event-based image reconstruction as a linear inverse problem with deep regularization using optical flow. IEEE TPAMI, 45 (07):8372-8389, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 128, + 288, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 288, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 288, + 150 + ], + "type": "text", + "content": "[68] Alex Zihao Zhu, Nikolay Atanasov, and Kostas Daniilidis. Event-based visual inertial odometry. In CVPR, 2017. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 150, + 288, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 150, + 288, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 150, + 288, + 183 + ], + "type": "text", + "content": "[69] Yunhao Zou, Yinqiang Zheng, Tsuyoshi Takatani, and Ying Fu. Learning to reconstruct high speed and high dynamic range videos from events. In CVPR, pages 2024-2033, 2021. 3" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25017" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Generalized Large-Scale Data Condensation via Various Backbone and Statistical Matching/1c0e887f-135a-4bea-8bad-3d161e8c65c8_content_list.json b/2024/Generalized Large-Scale Data Condensation via Various Backbone and Statistical Matching/1c0e887f-135a-4bea-8bad-3d161e8c65c8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..88985930983f9bf899b162ae50f8ecb7aeee1908 --- /dev/null +++ b/2024/Generalized Large-Scale Data Condensation via Various Backbone and Statistical Matching/1c0e887f-135a-4bea-8bad-3d161e8c65c8_content_list.json @@ -0,0 +1,1459 @@ +[ + { + "type": "text", + "text": "Generalized Large-Scale Data Condensation via Various Backbone and Statistical Matching", + "text_level": 1, + "bbox": [ + 125, + 128, + 844, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shitong Shao $^{1,3}$ Zeyuan Yin $^{1}$ Muxin Zhou $^{1}$ Xindong Zhang $^{2,3}$ Zhiqiang Shen $^{1,*}$ $^{1}$ Mohamed bin Zayed University of AI $^{2}$ Hong Kong Polytechnic University $^{3}$ OPPO Research 1090784053sst@gmail.com, {zeyuan.yin,muxin.zhou,zhiqiang.shen}@mbzuai.ac.ae 17901410r@connect.polyu.hk, *:Corresponding author", + "bbox": [ + 91, + 200, + 875, + 276 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2a9b66f7538f9e2b546c67ba66bfc1ba57e38bbd4635080df03df5a6f7d951d4.jpg", + "image_caption": [ + "Figure 1. Left: Our proposed G-VBSM consists of three novel and effective modules, named DD, GSM and GBM. The richness and quality of information in the synthetic data have been significantly enhanced compared with the baseline SRe2L through the sequential merging of DD, GBM, and GSM. Right: G-VBSM prioritizes \"generalized matching\" to ensure consistency between distilled and complete datasets across various backbones, layers, and statistics, and achieves the highest accuracy $31.4\\%$ on ImageNet-1k under IPC 10." + ], + "image_footnote": [], + "bbox": [ + 78, + 309, + 330, + 520 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0ae6f95b9f1d1115b8711b7d4e83f15a64733b534be8adb54a71dbd53f5f3d61.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 309, + 531, + 520 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e9dc599c01779e3a0f1f98890a49d8ac3e9132403f402adeabdefcf87fc374b8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 532, + 309, + 735, + 520 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d73b32c7cfbae21d6376f5aecd8c891595761ee508a054a3a6fdf18df025c103.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 735, + 309, + 888, + 520 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 599, + 313, + 616 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The lightweight \"local-match-global\" matching introduced by SRe2L successfully creates a distilled dataset with comprehensive information on the full $224 \\times 224$ ImageNet-1k. However, this one-sided approach is limited to a particular backbone, layer, and statistics, which limits the improvement of the generalization of a distilled dataset. We suggest that sufficient and various \"local-match-global\" matching are more precise and effective than a single one and have the ability to create a distilled dataset with richer information and better generalization ability. We call this perspective \"generalized matching\" and propose Generalized Various Backbone and Statistical Matching (G-VBSM) in this work, which aims to create a synthetic dataset with densities, ensuring consistency with the complete dataset across various backbones, layers, and statistics. As experimentally demonstrated, G-VBSM is the first algorithm to obtain strong performance across both", + "bbox": [ + 75, + 643, + 472, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "small-scale and large-scale datasets. Specifically, G-VBSM achieves performances of $38.7\\%$ on CIFAR-100, $47.6\\%$ on Tiny-ImageNet, and $31.4\\%$ on the full $224 \\times 224$ ImageNet-1k, respectively1. These results surpass all SOTA methods by margins of $3.9\\%$ , $6.5\\%$ , and $10.1\\%$ , respectively.", + "bbox": [ + 498, + 601, + 893, + 679 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 686, + 632, + 702 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the development of deep learning, the number of model parameters and the quantity of training data have become increasingly large [4, 37]. Researchers have tried to minimize the training overhead while preventing a decline in the generalization ability. Data condensation (DC), also known as Dataset distillation, first introduced by Wang et al. [32], aims to alleviate the training burden by synthesizing a small yet informative distilled dataset derived from the complete training dataset, while ensuring that the behav", + "bbox": [ + 496, + 713, + 893, + 849 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1Settings: CIFAR-100 with 128-width ConvNet under 10 images per class (IPC), Tiny-ImageNet with ResNet18 under 50 IPC, and ImageNet1k with ResNet18 under 10 IPC.", + "bbox": [ + 500, + 862, + 893, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16709", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ior of the distilled dataset on the target task remains consistent with that of the complete dataset. The extremely compressed distilled dataset contains sufficiently valuable information and have the potential for fast model training, and have been become a popular choice for different downstream application, like federated learning [7, 22], continual learning [14, 22, 40], neural architecture search [24, 41, 42] and 3D point clouds [30].", + "bbox": [ + 75, + 90, + 470, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A persistent problem that researchers have been working to solve [2, 9, 12, 42] in DC is the substantial data synthesis overhead [19, 36], which hinders its application in real-world large-scale datasets (e.g., ImageNet-1k) training. Typical performance matching [15, 32, 44] and trajectory matching [1, 3] unroll recursive computation graphs, requiring substantial GPU memory and resulting in prohibitive training costs. Zhao et al. [42] proposed gradient matching to address this, synthesizing distilled datasets by matching gradients from synthetic and real data in a single step. However, gradient computation and matching remain time-consuming [9], leading to the proposal of distribution matching [41]. This method and its variants [31] employ a network-based feature extractor to embed both synthetic and real data into a high-dimensional Hilbert space, then perform distribution matching. The training load for this direct, single-step process stems only from one gradient update of the synthetic data and the feature extractor [36]. Unfortunately, all of the above mentioned improved methods still have extremely large training overheads on the full $224 \\times 224$ ImageNet-1k.", + "bbox": [ + 75, + 214, + 473, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, SRe2L [34] accomplished data condensation for the first time on the full $224 \\times 224$ ImageNet-1k [18], achieving Top-1 validation accuracy $21.3\\%$ with ResNet18 under IPC 10. This method outperformed the latest state-of-the-art TESLA [2], which conducted on a low-resolution version of ImageNet-1k, by being $16 \\times$ faster and improved performance by a margin of $13.6\\%$ . SRe2L is inspired by DeepInversion [33] and aims to match statistics in BatchNorm generated from synthetic and real data. We reevaluate the success of SRe2L through the lightweight \"local-match-global\" essentially. The \"local-match-global\" refers to utilizing more comprehensive information (e.g., statistics in BatchNorm), generated from the model using the complete dataset (global), to guide the parameter update of the distilled dataset (local).", + "bbox": [ + 75, + 535, + 470, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, such lightweight and effective matching in SRe2L is singular, depending on the particular layer (i.e., BatchNorm), model (i.e., ResNet18), and statistics (i.e., channel mean/variance). Intuitively, sufficient \"local-match-global\" matching can result in more accurate and rational supervision than a single one, further enhancing the generalization of the distilled dataset. We call this perspective \"generalized matching\" and suggest that the distilled dataset is likely to perform consistent with the complete", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "dataset on the evaluation model, provided that there is sufficient variety in backbones, layers, and statistics used for matching.", + "bbox": [ + 496, + 90, + 890, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by this, we propose Generalized Various Backbone and Statistical Matching (G-VBSM) to fulfill \"generalized matching\". G-VBSM comprises three integral and effective parts named data densification (DD), generalized statistical matching (GSM), and generalized backbone matching (GBM). DD is utilized to ensure that the images within each class are linearly independent, thereby enhancing the (intra-class) diversity of the distilled dataset. This ultimately guarantees that \"generalized matching\" preserves the rich and diverse information within the synthetic data. GBM and GSM are designed to implement \"generalized matching\". Among them, GSM traverses the complete dataset without computing and updating the gradient, to record the statistics of Convolution at both the patch and channel levels. These statistics are subsequently utilized for matching during the data synthesis phase, in conjunction with the channel-level statistics in BatchNorm. Furthermore, GBM aims to ensure consistency between distilled and complete datasets across various backbones, enhancing matching sufficiency and leading to strong generalization in the evaluation phase. In particular, G-VBSM also ensures the efficiency of dataset condensation through a series of strategies, as mentioned in Sec. 3.", + "bbox": [ + 496, + 138, + 892, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extensive experiments on CIFAR-10, CIFAR-100, TinyImageNet, and the full $224 \\times 224$ ImageNet-1k, demonstrating that G-VBSM is the first algorithm that performs well on both small-scale and large-scale datasets. Specifically, we not only verify through ablation studies that GSM, GBM and DD are consistently reliable, but also achieve the highest $38.7\\%$ , $47.6\\%$ and $31.4\\%$ on CIFAR-100 (128-width ConvNet), Tiny-ImageNet (ResNet18), and the full $224 \\times 224$ ImageNet-1k (ResNet18) under images per class (IPC) 10, 50 and 10, respectively, which outperforms all previous state-of-the-art (SOTA) methods by $3.9\\%$ , $6.5\\%$ and $10.1\\%$ , respectively.", + "bbox": [ + 496, + 487, + 890, + 670 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Background", + "text_level": 1, + "bbox": [ + 500, + 686, + 629, + 704 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dataset condensation (DC) represents a data synthesis procedure that aims to compress a complete, large dataset $\\mathcal{T} = \\{(X_i,y_i)\\}_{i = 1}^{|T|}$ into a smaller, distilled dataset $S = \\{(\\tilde{X}_i,\\tilde{y}_i)\\}_{i = 1}^{|S|}$ , subject to $|\\mathcal{S}|\\ll |\\mathcal{T}|$ , while ensuring that an arbitrary evaluation model $f_{\\mathrm{eval}}(\\cdot)$ trained on $S$ yields similar performance to one trained on $\\mathcal{T}$ . Classical data distillation algorithms invariably require the candidate model $f_{\\mathrm{cand}}(\\cdot)$ to execute one or more steps on $S$ to update its parameter $\\theta_{\\mathrm{cand}}$ , subsequently achieving matching in terms of performance [15, 32], gradient [42], trajectory [1, 2], or distribution [31, 41]. The process $\\theta_{\\mathrm{cand}} - \\alpha \\nabla_{\\theta_{\\mathrm{cand}}}\\ell (f_{\\mathrm{cand}}(\\tilde{X}),\\tilde{y})$ where $\\ell (\\cdot ,\\cdot)$ and $(\\tilde{X},\\tilde{y})$ represent the loss function and", + "bbox": [ + 496, + 714, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16710", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c1a58103c5aaca085ba63e123665764fe60b78a9ec1160efafb18d383dc7e70a.jpg", + "image_caption": [ + "Figure 2. The overview of G-VBSM on the full $224 \\times 224$ ImageNet-1k, which ensures the consistency between the distilled and the complete datasets across various backbones, layers and statistics via \"generalized matching\"." + ], + "image_footnote": [], + "bbox": [ + 130, + 90, + 844, + 345 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "a batch sampled from $S$ , respectively, is notably time-consuming. Consequently, even the relatively swifest distribution matching [20, 31] is slower than the recent proposed SRe2L [34]. In fact, SRe2L is the only workable way to achieve DC on the full $224 \\times 224$ ImageNet-1k, as it requires updating the parameters of the synthetic data only once per iteration.", + "bbox": [ + 75, + 398, + 470, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SRe2L [34] encompasses three incremental subprocesses: Squeeze, Recover, and Relabel. Squeeze is designed to train $f_{\\mathrm{cand}}(\\cdot)$ containing BatchNorm in a standard manner, aiming to record the global channel mean $\\mathbf{BN}_l^{\\mathrm{CM}}$ and channel variance $\\mathbf{BN}_l^{\\mathrm{CV}}$ ( $l$ refers to the index of the $l$ -th layer) via exponential moving average (EMA), extracted from $\\mathcal{T}$ , for subsequent matching in Recover. In Recover after that, given the channel mean $\\mu_l(\\tilde{X})$ and channel variance $\\sigma_l^2(\\tilde{X})$ in $l$ -th BatchNorm obtained from $S$ , the statistical matching loss function can be formulated as", + "bbox": [ + 75, + 505, + 473, + 656 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {B N}} (\\tilde {X}) = \\sum_ {l} \\left| \\left| \\mu_ {l} (\\tilde {X}) - \\mathbf {B} \\mathbf {N} _ {l} ^ {\\mathrm {C M}} \\right| \\right| _ {2} + \\left| \\left| \\sigma_ {l} ^ {2} (\\tilde {X}) - \\mathbf {B} \\mathbf {N} _ {l} ^ {\\mathrm {C V}} \\right| \\right| _ {2}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 661, + 468, + 689 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Based on this, we can give the entire optimization objective in Recover as", + "bbox": [ + 76, + 695, + 468, + 724 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\tilde {X}} {\\arg \\min } \\mathcal {L} _ {\\mathrm {B N}} (\\tilde {X}) + \\ell \\left(f _ {\\text {c a n d}} (\\tilde {X}), y\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 729, + 468, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $y$ denotes the ground truth label. Moreover, SRe2L assigns soft labels $\\tilde{y}$ to the synthetic data, utilizing the logit-based distillation [6, 8] to improve the generalization ability of the distilled dataset. This can be denoted as", + "bbox": [ + 76, + 758, + 470, + 819 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {y} = \\operatorname {s o f t m a x} \\left(f _ {\\text {c a n d}} (\\tilde {X}) / \\tau\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 828, + 468, + 845 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\tau$ denotes the temperature to regulate the smoothness of the soft labels, thereby enhancing the distilled dataset's potential for generalization to unseen evaluation", + "bbox": [ + 76, + 854, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "models. The generated soft label can be stored on disk using FKD [23] so as not to defeat the purpose of DC. A crucial point of SRe2L is that BatchNorm calculates the statistics of the entire dataset using EMA, thereby providing a comprehensive and representative matching for the distilled dataset. Encouraged by this, our research focuses on applying sufficient 'local-match-global' matching to achieve \"generalized matching\".", + "bbox": [ + 496, + 398, + 893, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 536, + 591, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The comprehensive and detailed framework of our proposed Generalized Various Backbone and Statistical Matching (G-VBSM) is illustrated in Fig. 2. In essence, G-VBSM employs the lightweight regularization strategy data densification (DD) to ensure both the diversity and density of the distilled dataset, ensuring that the potential of \"generalized matching\" can be fully exploited. Moreover, generalized backbone matching (GBM) and generalized statistical matching (GSM) are utilized to achieve \"generalized matching\" by performing \"local-match-global\" matching across various backbones, layers, and statistics. In particular, the efficiency and effectiveness of DD, GBM, GSM, SRe2L, and TESLA are illustrated in Fig. 5.", + "bbox": [ + 496, + 561, + 893, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Data Densification", + "text_level": 1, + "bbox": [ + 500, + 770, + 679, + 786 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As illustrated in Fig. 3, the synthetic data generated by SRe2L exhibit excessive similarity within the same class, leading to a lack of diversity. Consequently, even if \"generalized matching\" preserves sufficient valuable information within a single image, the aggregate information content across all images within the same class does not increase effectively, which ultimately prevents \"generalized match-", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16711", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8f58880b3cd8c4670f8e8cb6424c374c074981b2fb81338a8af8fca87549b6f7.jpg", + "image_caption": [ + "Figure 3. Visualization of the mean cosine similarity between pairwise samples within the same class on ImageNet-1k under IPC 10." + ], + "image_footnote": [], + "bbox": [ + 86, + 89, + 455, + 238 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ing\" from being sufficiently advantageous. Data densification (DD) is proposed to address this by ensuring the data $\\tilde{X}$ has full rank in the batch dimension, thereby guaranteeing that samples in each class are linearly independent, and ultimately ensuring that the data is diverse and abundant to fully exploit the potential of \"generalized matching\".", + "bbox": [ + 76, + 281, + 467, + 372 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To execute this pipeline, $\\tilde{X}$ first needs to be downsampled to reduce the computational cost of eigenvalue decomposition:", + "bbox": [ + 76, + 372, + 467, + 417 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {X} = \\operatorname {A v g P o o l 2 d} (\\tilde {X}), \\text {s . t .} \\hat {X} \\in \\mathbb {R} ^ {B \\times C \\times 3 2 \\times 3 2}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 424, + 468, + 441 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $B$ and $C$ represent the batch size and the number of channels, respectively. Afterward, we reshape $\\hat{X}$ from $B\\times C\\times 32\\times 32$ to $B\\times (1024C)$ and perform matrix multiplication in each class $y$ to obtain the set of the Gram matrix $\\{\\hat{X}_y\\hat{X}_y^T\\}_{y\\in \\mathcal{Y}}$ , where $\\mathcal{V}$ refers to a set of all classes existing in this batch. And $\\hat{X}_y$ is a subbatch with class $y$ . Note that the alternative form $\\{\\hat{X}_y^T\\hat{X}_y\\}_{y\\in \\mathcal{Y}}$ is not desirable, as it is applicable only for dimensionality reduction in feature dimensions, which is why we do not consider singular value decomposition (SVD). To render $\\hat{X}_y\\hat{X}_y^T$ as full-rank as possible, we introduce the data densification loss in Eq. 5.", + "bbox": [ + 76, + 450, + 468, + 621 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {D D}} (\\tilde {X}) = \\sum_ {y \\in \\mathcal {Y}} D _ {\\mathrm {K L}} (\\operatorname {s t o p} - \\operatorname {g r a d} (p \\left(\\Sigma_ {y} / \\tau\\right)) | | p \\left(\\Sigma_ {y}\\right)), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 628, + 468, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\Sigma_y$ , $\\tau$ , stop\\_grad(\\cdot) and $p(\\cdot)$ refer to the eigenvalues of $\\hat{X}_y\\hat{X}_y^T$ , the temperature, the stop gradient operator and the softmax function, respectively. And $D_{\\mathrm{KL}}(\\cdot ||\\cdot)$ denotes Kullback-Leibler divergence. As demonstrated in Sec. E, the diversity of the data is significantly enhanced by the employment of Eq. 5. In our experiment, $\\tau$ is set as 4 in default and we do not assign a deliberate weight (set to 1 by default) to $\\mathcal{L}_{\\mathrm{DD}}$ because $\\mathcal{L}_{\\mathrm{DD}} \\equiv 0$ at the early $10\\%$ of the iterations. In other words, DD is quite stable and the optimization objective $\\arg \\min_{\\{\\Sigma_y\\}_{y \\in \\mathcal{Y}}} \\mathcal{L}_{\\mathrm{DD}}$ is equivalent to $\\arg \\min_{\\{\\Sigma_y\\}_{y \\in \\mathcal{Y}}} \\sum_y \\sigma^2(\\Sigma_y)$ .", + "bbox": [ + 76, + 666, + 468, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Technical Detail. A problem that warrants attention is that in the SRe2L's implementation [34], having merely a single sample in each class of a batch indicates insufficient", + "bbox": [ + 76, + 854, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/53d0feb31231732b0ef278c1c9463a5864135201a3f84222449c2da59b8c46a7.jpg", + "image_caption": [ + "Figure 4. The illustration of the original loop and the reorder loop." + ], + "image_footnote": [], + "bbox": [ + 527, + 87, + 864, + 220 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "to execute DD under the order of the original loop, as depicted in Fig. 4 (Left). A simple solution is to translate the original loop to the reorder loop, as shown in Fig. 4 (Right). However, our experiment on ResNet50 (i.e., the evaluation model) substantiates that this solution suffers a $2.6\\%$ accuracy degradation (details can be found in Sec. 4.1) on ImageNet-1k under IPC 10. The reason is that the number of classes in each iteration within the reorder loop is insufficient, preventing a batch have the ability to match the global statistics in BatchNorm (i.e., $\\mathbf{BN}_l^{\\mathrm{CM}}$ and $\\mathbf{BN}_l^{\\mathrm{CV}}$ ). Motivated by score distillation sampling (SDS) [17], we update the statistics during data synthesis via EMA to solve this issue, so that the statistics of all past batches can assist the statistics of the current batch match $\\mathbf{BN}_l^{\\mathrm{CM}}$ and $\\mathbf{BN}_l^{\\mathrm{CV}}$ :", + "bbox": [ + 496, + 243, + 890, + 455 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mu_ {l} ^ {\\text {t o t a l}} = \\alpha \\mu_ {l} ^ {\\text {t o t a l}} + (1 - \\alpha) \\mu_ {l} (\\tilde {X}), \\sigma_ {l} ^ {2, \\text {t o t a l}} = \\alpha \\sigma_ {l} ^ {2, \\text {t o t a l}} + (1 - \\alpha) \\sigma_ {l} ^ {2} (\\tilde {X}), \\\\ \\mathcal {L} _ {\\mathrm {B N}} ^ {\\prime} (\\tilde {X}) = \\sum_ {l} | | \\mu_ {l} (\\tilde {X}) - \\mathbf {B N} _ {l} ^ {\\mathrm {C M}} - \\operatorname {s t o p \\_ g r a d} (\\mu_ {l} (\\tilde {X}) - \\mu_ {l} ^ {\\text {t o t a l}}) | | _ {2} \\\\ + \\left\\| \\sigma_ {l} ^ {2} (\\tilde {X}) - \\mathbf {B N} _ {l} ^ {\\mathrm {C V}} - \\operatorname {s t o p \\_ g r a d} \\left(\\sigma_ {l} ^ {2} (\\tilde {X}) - \\sigma_ {l} ^ {2, \\text {t o t a l}}\\right) \\right\\| _ {2}. \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 460, + 890, + 536 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The derivation of $\\mathcal{L}_{\\mathrm{BN}}^{\\prime}(X)$ can be found in Appendix B. We call this lightweight strategy as \"match in the form of score distillation sampling\" and have demonstrated its effectiveness and feasibility in our ablation studies.", + "bbox": [ + 496, + 537, + 890, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Generalized Backbone Matching", + "text_level": 1, + "bbox": [ + 500, + 604, + 787, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Performing data synthesis only on a single pre-trained model is not able to enjoy ensemble gains from various backbones. Meanwhile, classical DC algorithms such as MTT [1] and FTD [3] obtain performance improvements from multiple randomly initialized backbones. Therefore, introducing generalized backbone matching (GBM) to apply various backbones for data synthesis is a desirable choice for \"generalized matching\". It ensures a number of \"local-match-global\" matching nearly $N_{\\mathrm{b}} \\times$ higher compared to just depending on a single backbone, where $N_{\\mathrm{b}}$ denotes the number of backbones. Regrettably, unrolling various backbone computational graphs in parallel for data synthesis is extremely expensive at the computational cost level. A solution is to randomly sample a backbone from the candidates in per iteration. This simple yet effective strategy not only ensures computational efficiency but also allows the data synthesis process to benefit from the diverse and multifaceted matching that the various backbones pro", + "bbox": [ + 496, + 628, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16712", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "vide. We denote this pipeline as", + "bbox": [ + 76, + 90, + 292, + 107 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\text {c a n d}} (\\cdot) \\sim \\mathcal {U} (\\mathbb {S}), \\mathbb {S} = \\{\\text {R e s N e t 1 8}, \\dots , \\text {S h u f f l e N e t V 2 - 0 . 5} \\}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 112, + 468, + 127 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To maintain backbone consistency in both the data synthesis and soft label generation phases, we introduce a backbone ensemble during soft label generation:", + "bbox": [ + 76, + 133, + 468, + 181 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {z} = \\left\\{ \\begin{array}{c c} \\sum_ {f _ {\\text {c a n d}} \\in \\mathbb {S}} \\frac {\\sum_ {g \\in \\mathbb {S}} | | g (\\tilde {X}) | | _ {\\mathrm {F}}}{| \\mathbb {S} | | | f _ {\\text {c a n d}} (\\tilde {X}) | | _ {\\mathrm {F}}} \\frac {f _ {\\text {c a n d}} (\\tilde {X})}{| \\mathbb {S} |}, & \\mathrm {w / L N}, \\\\ \\sum_ {f _ {\\text {c a n d}} \\in \\mathbb {S}} \\frac {f _ {\\text {c a n d}} (\\tilde {X})}{| \\mathbb {S} |}, & \\mathrm {w / o L N}, \\end{array} \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 191, + 468, + 266 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $||\\cdot||_{\\mathrm{F}}$ and LN refer to the Frobenius norm and Logit Normalization, respectively. Thus, we can obtain the soft label by $\\tilde{y} = \\text{softmax}(\\tilde{z}/\\tau)$ . Particularly, the use of LN is optional, as demonstrated by our ablation studies; it is beneficial for ResNet18 but not for ResNet50 or ResNet101. Moreover, it's important to highlight that we apply a parallel mechanism for soft label generation since it provides a significantly lower computational cost, less than 1/30 of that required for data synthesis, thus making the computational overhead negligible.", + "bbox": [ + 75, + 276, + 472, + 429 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Generalized Statistical Matching", + "text_level": 1, + "bbox": [ + 76, + 438, + 366, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Only ensuring backbone diversity is insufficient to fully exploit the potential of \"generalized matching\". In this subsection, we aim to introduce additional statistics for matching during the data synthesis phase. Prior methods [33, 34] only utilize BatchNorm since the presence of global information statistics within the pre-trained model, with no apparent solution ready for other layers (e.g., Convolution). Retraining the model is the simplest way to address this problem, followed by updating the other layers' parameters through EMA. This approach is obviously impractical, as it defeats the purpose of using gradient descent. By contrast, we propose to allow the pre-trained model $f_{\\mathrm{eval}}(\\cdot)$ to loop through the training dataset $\\mathcal{T}$ once without calculating the gradient to obtain the global statistics of Convolution, thus serving \"local-match-global\":", + "bbox": [ + 75, + 462, + 472, + 690 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {C o n v} _ {l} ^ {\\mathbf {C M}} = \\frac {1}{| \\mathcal {T} |} \\sum_ {i} ^ {| \\mathcal {T} |} \\mathbf {C M} _ {i} ^ {l}, \\mathbf {C o n v} _ {l} ^ {\\mathbf {C V}} = \\frac {1}{| \\mathcal {T} |} \\sum_ {i} ^ {| \\mathcal {T} |} \\mathbf {C V} _ {i} ^ {l}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 696, + 468, + 739 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {C o n v} _ {l} ^ {\\mathbf {P M}} = \\frac {1}{| \\mathcal {T} |} \\sum_ {i} ^ {| \\mathcal {T} |} \\mathbf {P M} _ {i} ^ {l}, \\mathbf {C o n v} _ {l} ^ {\\mathbf {P V}} = \\frac {1}{| \\mathcal {T} |} \\sum_ {i} ^ {| \\mathcal {T} |} \\mathbf {P V} _ {i} ^ {l}.\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 736, + 411, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $\\mathbf{CM}_i^l\\in \\mathbb{R}^{C_l}$ $\\mathbf{CV}_i^l\\in \\mathbb{R}^{C_l}$ $\\mathbf{PM}_i^l\\in \\mathbb{R}^{\\left[\\frac{H}{N_l^p}\\right]\\times \\left[\\frac{W}{N_l^p}\\right]}$ and $\\mathbf{PV}_i^l\\in \\mathbb{R}^{\\left[\\frac{H}{N_l^p}\\right]\\times \\left[\\frac{W}{N_l^p}\\right]}$ refer to the channel mean, the channel variance, the patch mean and the patch variance, respectively, for the $l$ -th Convolution when the input to $f_{\\mathrm{cand}}(\\cdot)$ is the $i$ -th batch, where $C_l$ and $\\left[\\frac{H}{N_l^p}\\right]\\times \\left[\\frac{W}{N_l^p}\\right]$ denote the number of channels and patches of the $l$ -th Convolution, respectively. We define $N_{l}^{p}$ as 4, 4, 4 and 16 by default", + "bbox": [ + 76, + 777, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "on CIFAR-10, CIFAR-100, Tiny-ImageNet and ImageNet-1k, respectively. After obtaining the global channel mean $\\mathbf{Conv}_l^{\\mathbf{CM}}$ , the global channel variance $\\mathbf{Conv}_l^{\\mathbf{CV}}$ , the global patch mean $\\mathbf{Conv}_l^{\\mathbf{PM}}$ and the global patch variance $\\mathbf{Conv}_l^{\\mathbf{PV}}$ , we can store them in a disk thus avoiding secondary calculations. In the data synthesis phase, we introduce $\\mathcal{L}_{\\mathrm{Conv}}'(\\tilde{X})$ in Eq. 10 to accomplish joint matching with $\\mathcal{L}_{\\mathrm{BN}}'(\\tilde{X})$ .", + "bbox": [ + 496, + 90, + 890, + 200 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {C o n v}} ^ {\\prime} (\\tilde {X}) = \\sum_ {l} | | \\mu_ {l} ^ {c} (\\tilde {X}) - \\mathbf {C o n v} _ {l} ^ {\\mathrm {C M}} - \\operatorname {s t o p - g r a d} \\left(\\mu_ {l} ^ {c} (\\tilde {X}) - \\mu_ {l} ^ {c, \\text {t o t a l}}\\right) | | _ {2} \\\\ + \\left\\| \\sigma_ {l} ^ {c, 2} (\\tilde {X}) - \\mathbf {C o n v} _ {l} ^ {\\mathrm {C V}} - \\operatorname {s t o p - g r a d} \\left(\\sigma_ {l} ^ {c, 2} (\\tilde {X}) - \\sigma_ {l} ^ {c, 2, \\text {t o t a l}}\\right) \\right\\| _ {2}, \\\\ + \\left\\| \\mu_ {l} ^ {p} (\\tilde {X}) - \\mathbf {C o n v} _ {l} ^ {\\mathrm {P M}} - \\operatorname {s t o p - g r a d} \\left(\\mu_ {l} ^ {p} (\\tilde {X}) - \\mu_ {l} ^ {p, \\text {t o t a l}}\\right) \\right\\| _ {2} \\\\ + \\left\\| \\sigma_ {l} ^ {p, 2} (\\tilde {X}) - \\mathbf {C o n v} _ {l} ^ {\\mathrm {P V}} - \\operatorname {s t o p} - \\operatorname {g r a d} \\left(\\sigma_ {l} ^ {p, 2} (\\tilde {X}) - \\sigma_ {l} ^ {p, 2, \\text {t o t a l}}\\right) \\right\\| _ {2}, \\tag {10} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 208, + 890, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mu_l^{c,\\mathrm{total}}$ , $\\sigma_l^{c,2,\\mathrm{total}}$ , $\\mu_l^{p,\\mathrm{total}}$ , and $\\sigma_l^{p,2,\\mathrm{total}}$ are each updated from the channel mean $\\mu_l^c (\\tilde{X})$ , channel variance $\\sigma_l^{c,2}(\\tilde{X})$ , patch mean $\\mu_l^p (\\tilde{X})$ , and patch variance $\\sigma_l^{p,2}(\\tilde{X})$ respectively, all obtained via EMA from the current batch. In experiments, we discovered that Eq. 10 causes a slightly larger computational burden, so we randomly drop the matching of statistics with a probability of $\\beta_{\\mathrm{dr}}$ to ensure the efficiency of GSM.", + "bbox": [ + 496, + 301, + 890, + 428 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Loss Function in the Evaluation Phase. Unlike DD, GBM and GSM are designed to create a distilled dataset that is enriched with information. Here, we introduce an enhancement to the loss function tailored specifically for the evaluation phase. Essentially, the evaluation phase is a knowledge distillation framework for transferring knowledge from a pre-trained model to the evaluation model. SRe2L utilizes $D_{\\mathrm{KL}}(\\tilde{y} || \\mathrm{softmax}(f_{\\mathrm{eval}}(\\tilde{X}) / \\tau))$ as the loss function and experimentally illustrates that it improves performance by roughly $10\\%$ . As established in SRe2L, an increase in temperature $\\tau$ correlates with enhanced performance of the evaluation model. Inspired by this and $\\tau^2 D_{\\mathrm{KL}}(\\mathrm{softmax}(p) / \\tau ||\\mathrm{softmax}(q) / \\tau)$ is equivalent to $\\frac{1}{2C} ||p - q||_2^2$ when $\\tau \\rightarrow +\\infty$ [10], we introduce a novel loss function $\\mathrm{MSE} + \\gamma \\times \\mathrm{GT}$ to avoid numerical error caused by the large $\\tau$ and improve the generalization of the distilled dataset without any additional overhead (ignore the weights $\\tau^2$ and $\\frac{1}{2C}$ ):", + "bbox": [ + 496, + 431, + 892, + 705 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {e v a l}} (\\tilde {X}, \\tilde {y}, y) = | | f _ {\\text {e v a l}} (\\tilde {X}) - \\tilde {z} | | _ {2} ^ {2} - \\gamma \\mathbf {y} \\log \\left(\\operatorname {s o f t m a x} \\left(f _ {\\text {e v a l}} (\\tilde {X})\\right)\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 516, + 710, + 890, + 738 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $y$ represents the one-hot encoding (w.r.t. the ground truth label $y$ ). As illustrated in Fig. 5, simply replacing the loss function with $\\mathrm{MSE} + 0.1 \\times \\mathrm{GT}$ ( $\\gamma$ is set as 0.1) in SRe2L improves the performance of ResNet18 (i.e., the evaluation model) by a margin of $0.9\\%$ on ImageNet-1k under IPC 10.", + "bbox": [ + 496, + 739, + 890, + 814 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 500, + 829, + 624, + 847 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We perform comparison experiments on the large-scale dataset including the full $224 \\times 224$ ImageNet-1k [18] and the small-scale datasets including Tiny-ImageNet [26],", + "bbox": [ + 496, + 854, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16713", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/494301707fff98f4681695e00c6652ea30dde251ec38a3c20ef2b4bafcc8422c.jpg", + "image_caption": [ + "Figure 5. Comparison of the effectiveness and efficiency of G-VBSM components. Among them, \"DD+GBM+GSM $(\\beta_{\\mathrm{dr}} = 0.0) + (\\mathrm{MSE} + \\mathrm{GT})$ represents the comprehensive G-VBSM." + ], + "image_footnote": [], + "bbox": [ + 86, + 88, + 454, + 227 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CIFAR-10/100 [11]. To highlight that our proposed G-VBSM is designed for large-scale datasets, all ablation experiments are performed on ImageNet-1k.", + "bbox": [ + 75, + 292, + 467, + 339 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Hyperparameter Settings. We prioritize selecting various convolution-based backbone, ensuring the maximal difference in architecture, while also adhering to the criterion of minimizing the number of parameters, an approach empirically demonstrated to be superior [34, 38]. On ImageNet-1k, we skip the model pre-training phase by directly using Torchvision's open source pretraining weights [16] of {ResNet18 [5], MobileNetV2 [21], EfficientNet-B0 [25], ShuffleNetV2-0.5 [39]}. For the remaining dataset, we all train {128-width ConvNet [39], WRN-16-2 [37], ResNet18 [5], ShuffleNetV2-0.5 [39], MobileNetV2-0.5 [21]} from scratch with few epoch in the model pre-training phases. Gray cells in all tables represent the highest performance. Meanwhile, (R18), (R50), etc. in all tables, represent the evaluation models. More details about the remaining method hyperparameter settings and additional ablation studies can be found in Appendix A and Appendix C & D, respectively.", + "bbox": [ + 75, + 357, + 468, + 630 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 643, + 240, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The Trade-off between Efficiency and Effectiveness. The performance improvements and increased GPU latency for components of G-VBSM in comparison to the baseline SRe2L, are illustrated in Fig. 5. Clearly, DD, GBM, and GSM are highly effective in enhancing the generalization of distilled datasets. Compared with GSM, both DD and GBM are slightly lightweight and efficient. With DD and GBM alone - i.e., $\\mathrm{DD + GBM + GSM}$ $(\\beta_{\\mathrm{dr}} = 1.0) + (\\mathrm{MSE} + \\mathrm{GT})$ - the accuracy of this approach surpasses that of SRe2L by a significant margin of $6.5\\%$ . Additionally, the comprehensive G-VBSM, $\\mathrm{DD + GBM + GSM}$ $(\\beta_{\\mathrm{dr}} = 0.0) + (\\mathrm{MSE} + \\mathrm{GT})$ , further enhances the performance of $\\mathrm{DD + GBM + GSM}$ $(\\beta_{\\mathrm{dr}} = 1.0) + (\\mathrm{MSE} + \\mathrm{GT})$ by a notable margin of $3.6\\%$ . Note that $\\mathrm{MSE} + \\gamma \\times \\mathrm{GT}$ ( $\\gamma$ is set as 0.1) is also extremely critical, for SRe2L and $\\mathrm{DD + GBM}$ boosting $0.9\\%$ and $2.7\\%$ , respectively.", + "bbox": [ + 75, + 667, + 467, + 910 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/135f72913afabfb5185927415dfe8fe86c91c7599d6e29763608476c37d00384.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
α\\Evaluation ModelResNet18(MSE+0.1×GT)ResNet50(MSE+0.1×GT)
0.030.4%31.9%
0.431.4%34.5%
0.831.9%36.4%
", + "bbox": [ + 519, + 88, + 867, + 152 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Matching in the Form of SDS. Benefiting from this matching form, G-VBSM can achieve the best performance with small batch sizes (e.g., 40 in ImageNet-1k and 50 in CIFAR-100) in different settings, as demonstrated in our comparative experiments. Table 1 demonstrates the impact of the factor $\\alpha$ for performing the EMA update on the final performance achieved by G-VBSM, where $\\alpha = 0$ indicates that $\\mathcal{L}_{\\mathrm{BN}}^{\\prime}$ degenerates to $\\mathcal{L}_{\\mathrm{BN}}$ . We can conclude that the SDS matching form is feasible and the generalization of the distilled dataset improves with increasing $\\alpha$ .", + "bbox": [ + 498, + 191, + 890, + 343 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a654e3d0ca27600ca12fbbae6457ab5db7352055b3978d5017e2ae58b291461b.jpg", + "table_caption": [ + "Table 1. ImageNet-1k Top-1 Acc. about different $\\alpha$ under IPC 10." + ], + "table_footnote": [], + "table_body": "
MethodEvaluation Model
DeiT-TinyResNet18MobileNetV2Swin-Tiny
SRe2L15.41%46.79%36.59%39.23%
G-VBSM29.43%51.82%48.66%57.40%
", + "bbox": [ + 500, + 356, + 901, + 414 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Cross-Architecture Generalization. The evaluation of cross-architecture generalization is crucial in assessing the quality of distilled datasets. Unlike traditional methods [1, 15, 20] which focus on CIFAR-100, our approach evaluates the effectiveness of the distilled dataset on ImageNet1k, employing a suite of models with real-world applicability, including ResNet18 [5], MobileNetV2 [21], DeiTiny [28], and Swin-Tiny [13]. The experimental results are reported in Table 2. From Tables 7 and 2, it is evident that the G-VBSM-synthetic dataset can effectively generalize across ResNet {18, 50, 101}, MobileNetV2, DeiTiny, and Swin-Tiny architectures. Notably, DeiT-Tiny and Swin-Tiny, two architectures not encountered during the data synthesis phase, demonstrates significant proficiency with the accuracy $29.43\\%$ and $57.40\\%$ , outperforming the latest SOTA SRe2L by a margin of $14.01\\%$ and $18.17\\%$ , respectively.", + "bbox": [ + 498, + 459, + 890, + 715 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1efe11730878152534f4e8a44d0188f5d5b537c7c47be4cd50ee7c5426c11112.jpg", + "table_caption": [ + "Table 2. ImageNet-1k Top-1 Acc. on cross-architecture generalization under IPC 50." + ], + "table_footnote": [], + "table_body": "
Candidates (Backbone)Evaluation Model
ResNet18MobileNetV2EfficientNet-B0ShuffleNetV2-0.5ResNet18ResNet50
✓25.7%30.2%
✓✓27.2%31.2%
✓✓✓27.9%32.0%
✓✓✓✓31.4%34.5%
", + "bbox": [ + 500, + 729, + 901, + 797 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. ImageNet-1k Top-1 Acc. about the number of candidate backbone in the soft label generation pipeline under IPC 10.", + "bbox": [ + 498, + 808, + 888, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Ensemble in Soft Label Generation. In knowledge distillation, soft labels obtained through multiple teacher ensembles can effectively enhance the generalization of the student [35]. This observation is similarly corroborated in", + "bbox": [ + 498, + 839, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16714", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8f339e50e538ee9c08e709dd0e54907fcf5e2e0327ae27d36a7bb7963b97d5c0.jpg", + "image_caption": [ + "Figure 6. ImageNet-1k Top-1 Acc. of different loss function, $\\beta_{\\mathrm{dr}}$ and batch sizes under IPC 10." + ], + "image_footnote": [], + "bbox": [ + 200, + 88, + 367, + 227 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0aee72a94832890741a2abcdbdaa4367f85762f86a1795342ab35c3a0fd03085.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 393, + 88, + 566, + 227 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/16d36d0cbaa7310f4460a46b42cdd24c90c5548c1d09d61236e2af8ccf7a2187.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 602, + 89, + 767, + 224 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e63f6b6e087433123e662cb744ee629324b629f71a5cb41e2c6ba767b2fab48a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetIPCMTT [1] (CW128)DataDAM [20] (CW128)TESLA [2] (R18)SRe2L (R18)SRe2L (R50)SRe2L (R101)G-VBSM (R18)G-VBSM (R50)G-VBSM (R101)
Tiny-ImageNet5028.0±0.328.7±0.3-41.1±0.442.2±0.542.5±0.247.6±0.348.7±0.248.8±0.4
100---49.7±0.351.2±0.451.5±0.351.0±0.452.1±0.352.3±0.1
ImageNet-1k1064.0±1.3†6.3±0.07.7±0.121.3±0.628.4±0.130.9±0.131.4±0.535.4±0.838.2±0.4
50-15.5±0.2-46.8±0.255.6±0.360.8±0.551.8±0.458.7±0.361.0±0.4
100---52.8±0.361.0±0.462.8±0.255.7±0.462.2±0.363.7±0.2
", + "bbox": [ + 80, + 253, + 890, + 325 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/33335728a087809a15de3ad8ac0dde668bf21efb9773d0db8e8177afd7e5ba14.jpg", + "table_caption": [ + "Table 4. Comparison with baseline models in Tiny-ImageNet and ImageNet-1k. ${}^{ \\dagger }$ indicates the ImageNette dataset,which contains only 10 classes. DataDAM [20] and TESLA [2] use the downsampled ${64} \\times {64}$ ImageNet-1k. We cite the experimental results from SRe2L [34]." + ], + "table_footnote": [], + "table_body": "
DatasetIPCCoreset SelectionTraining Set Synthesis (CW128)Training Set Synthesis (R18)Whole Dataset
RandomHerdingK-CenterForgettingDC [42]DM [41]CAFE [31]KIP [15]MTT [1]DataDAM[20]G-VBSMSRe2LG-VBSM(CW128)
CIFAR-101026.0±1.231.6±0.714.7±0.923.3±1.044.9±0.548.9±0.650.9±0.546.1±0.765.3±0.754.2±0.846.5±0.727.2±0.553.5±0.684.8±0.1
5043.4±1.040.4±0.627.0±1.423.3±1.153.9±0.563.0±0.462.3±0.453.2±0.771.6±0.267.0±0.454.3±0.347.5±0.659.2±0.4
CIFAR-10014.2±0.38.3±0.38.4±0.34.5±0.212.8±0.311.4±0.314.0±0.312.0±0.224.3±0.314.5±0.516.4±0.72.0±0.225.9±0.556.2±0.3
1014.6±0.517.3±0.317.3±0.315.1±0.325.2±0.329.7±0.331.5±0.240.1±0.433.1±0.434.8±0.538.7±0.231.6±0.559.5±0.4
5030.0±0.433.7±0.530.5±0.3-30.6±0.643.6±0.447.7±0.2-42.9±0.349.4±0.345.7±0.449.5±0.365.0±0.5
", + "bbox": [ + 80, + 368, + 890, + 452 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "dataset distillation, as evidenced in Table 3. The greater the number of models, the stronger the generalization of the distilled dataset. It is particularly interesting to point out that ShuffleNetV2-0.5 improves further by more than $2\\%$ even though the other three models have already achieved modest ensemble gains. We attribute this enhancement to the channel shuffle mechanism within ShuffleNetV2-0.5, which imposes a beneficial regularization constraint for G-VBSM.", + "bbox": [ + 75, + 511, + 470, + 632 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b186609fc699993eeffa26fca44819c092b0533f171f9989dbe59551d6dc4316.jpg", + "table_caption": [ + "Table 5. Comparison with baseline models on CIFAR-10/100. All methods, except for SRe2L and G-VBSM, use a 128-width ConvNet (CW128) for data synthesis and evaluation. G-VBSM utilizes {CW128, WRN-16-2, ResNet18 (R18), ShuffleNetV2-0.5, MobileNetV2-0.5} for data synthesis and {CW128, R18} for evaluation. We cite the experimental results, except for SRe2L's, from DataDAM [20]." + ], + "table_footnote": [], + "table_body": "
Logit NormalizationEvaluation Model
ResNet18 (MSE+0.1×GT)ResNet18 (KL)ResNet50 (MSE+0.1×GT)ResNet50 (KL)
Yes31.4%25.4%34.5%28.6%
No31.0%25.1%35.4%31.9%
", + "bbox": [ + 78, + 645, + 478, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. ImageNet-1k Top-1 Acc. about the use of logit normalization under IPC 10.", + "bbox": [ + 75, + 710, + 468, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Logit Normalization in Soft Label Generation. The aim of this strategy is to maintain the consistency of the logits (can be viewed as vectors) magnitude across all models within a high-dimensional space, thereby ensuring equal contribution of these models to the ultimate soft label for ensemble. This approach is not universally effective, as shown in Table 6. When the evaluation model parameter count is low (e.g., ResNet18), it can enhance G-VBSM performance. Conversely, with a model like ResNet50, it may hinder performance, rendering G-VBSM uncompetitive. As", + "bbox": [ + 75, + 750, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "a result, in this work, ResNet $\\{50, 101\\}$ and ViT-based models do not employ logit normalization, while the remaining models do.", + "bbox": [ + 496, + 511, + 892, + 555 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Loss Function in the Evaluation Phase. As illustrated in Fig. 6, replacing $D_{\\mathrm{KL}}(\\cdot ||\\cdot)$ in SRe2L with $\\mathrm{MSE} + \\gamma \\times \\mathrm{GT}$ demonstrates to be extremely effective. For example, with $\\beta_{\\mathrm{dr}} = 0.0$ and batch size $= 40$ , adjusting $\\gamma$ to 0.01 and 0.1 enhances model performance during the evaluation phase by $1.4\\%$ and $6.0\\%$ , respectively. As a result, we employ a $\\gamma$ of 0.1 in all experiments conducted on ImageNet-1k. Furthermore, we observe that the distilled dataset's generalization is not significantly affected ( $\\leq 2\\%$ ) by variations in either batch size or $\\beta_{\\mathrm{dr}}$ . This ensured that G-VBSM attain SOTA results with a minimal batch size 40 and $\\beta_{\\mathrm{dr}} = 0.4$ (i.e., the default settings in ImageNet-1k).", + "bbox": [ + 496, + 561, + 893, + 743 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1f518e974b524b3100798be7098e0c0c470ae6fb5178780a279feb4d07290cee.jpg", + "image_caption": [ + "Figure 7. Synthetic data visualization on small-scale datasets." + ], + "image_footnote": [], + "bbox": [ + 506, + 756, + 879, + 883 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16715", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Synthetic Data Visualization. The visualization results for the large-scale (the full $224 \\times 224$ ImageNet-1k) and small-scale (CIFAR-10/100 and Tiny-ImageNet) datasets are displayed in Figs. 1 and 7, respectively. On ImageNet-1k, the images synthesized by G-VBSM are more informative and abstract compared with those from SRe2L. Furthermore, on small-scale datasets, the distilled images obtained by G-VBSM markedly differ from the images presented in MTT and DataDAM's papers [1, 20], illustrating that G-VBSM is an out-of-the ordinary algorithm. More synthetic data can be found in Appendix I.", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d0f0509b50203882be0e13b9935d8f09e5a2cba11aad947d1e4fa7ed3eabde19.jpg", + "image_caption": [ + "Figure 8. t-SNE visualization on ImageNet-1k." + ], + "image_footnote": [], + "bbox": [ + 101, + 276, + 450, + 426 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Logit Embedding Distributions Visualization. We feed distilled datasets into a pre-trained model (i.e., ResNet18) to obtain logit embeddings for t-SNE [29] visualization. Note that unlike previous methods that used feature embeddings before the classifier layer, we utilize logit embeddings after the classifier layer for presentation. Fig. 8 shows the distributions of G-VBSM and SRe2L logit embeddings. The logit embedding obtained by G-VBSM has less intra-class similarity than SRe2L, ensuring softer labels, preventing model overconfidence during the evaluation phase, and ultimately enhancing model generalization.", + "bbox": [ + 75, + 473, + 468, + 640 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2. Large-Scale Dataset Comparison", + "text_level": 1, + "bbox": [ + 76, + 650, + 366, + 666 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Full $224 \\times 224$ ImageNet-1k. As illustrated in the ImageNet-1k part of Table 4, G-VBSM consistently outperforms SRe2L across all IPC $\\{10, 50, 100\\}$ , with the evaluation model being ResNet $\\{18, 50, 101\\}$ , indicating that G-VBSM is highly competitive in the large-scale dataset ImageNet-1k. In particular, under extreme compression scenarios, such as IPC 10, G-VBSM achieves significant performance increases - $10.1\\%$ , $7.0\\%$ , and $7.3\\%$ for ResNet18, ResNet50, and ResNet101, respectively - compared with SRe2L. In comparison to the latest SOTA classical method, DataDAM, just experimenting on the full $64 \\times 64$ ImageNet-1k, we exceed it by a margin of $23.7\\%$ under IPC 10. These results are particularly impressive when considered in the context of the large-scale dataset ImageNet-1k.", + "bbox": [ + 75, + 672, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Small-Scale Dataset Comparison", + "text_level": 1, + "bbox": [ + 500, + 90, + 787, + 107 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Tiny-ImageNet. Similar to SRe2L, we evaluate the generalization of the distilled dataset on ResNet {18, 50, 101}. As presented in the Tiny-ImageNet part of Table 4, G-VBSM consistently surpasses the latest SOTA method SRe2L on different settings. Our method outperforms SRe2L, MTT and DataDAM by $6.5\\%$ , $19.6\\%$ and $18.9\\%$ , respectively, under IPC 50 applying 128-width ConvNet or ResNet18 as the evaluation model.", + "bbox": [ + 496, + 113, + 890, + 233 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "CIFAR-100. One of the disadvantages of SRe2L is its inability to compete with SOTA performance on this small-scale dataset benchmark (i.e., evaluated with 128-width ConvNet and ResNet18), such as the performance on CIFAR-100 shown in Table 5, as a result, it fails to demonstrate that \"local-match-global\" (e.g., statistical matching of BatchNorm) can rival or outperform traditional DC algorithms. In contrast, as shown in the CIFAR-100 part of Table 5, G-VBSM achieves accuracies $16.4\\%$ and $38.7\\%$ , the highest under IPC 1 and 10 on CIFAR-100 through \"generalized matching\", while maintaining the same number of epochs (i.e., 1000) and the model (i.e., 128-width ConvNet) for the evaluation phase as the traditional methods, outperforming the latest SOTA DataDAM by $1.9\\%$ and $3.9\\%$ , respectively. To the best of our knowledge, our proposed G-VBSM is the first algorithm that is strongly competitive on CIAR-100 as well as the full $224 \\times 224$ ImageNet-1k.", + "bbox": [ + 496, + 244, + 892, + 501 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "CIFAR-10. As illustrated in the CIFAR-10 part of Table 5, G-VBSM outperforms vanilla DC [32], KIP [15], and all coreset selection algorithms on the benchmark with 128-width ConvNet as the evaluation model. Meanwhile, on the benchmark with ResNet 18 as the evaluation model, G-VBSM outperforms the baseline SRe2L by $26.3\\%$ and $11.7\\%$ under IPC 10 and 50, respectively.", + "bbox": [ + 496, + 512, + 890, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 632, + 617, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce a novel perspective termed \"generalized matching\" for dataset condensation. This posits that an abundance of lightweight \"local-match-global\" matching are more effective than a single \"local-match-global\" matching, and even surpass the precise and costly matching used by traditional methods. Consequently, we present G-VBSM, designed to ensure data densification while performing matching based on sufficient and various backbones, layers, and statistics. Experiments conducted on CIFAR-10/100, Tiny-[ImageNet, and the full $224 \\times 224$ ImageNet-1k demonstrated that our method is the first algorithm to show strong performance across both small-scale and large-scale datasets. For future research, we plan to extend this approach's applicability to large-scale detection and segmentation datasets.", + "bbox": [ + 496, + 657, + 890, + 883 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16716", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] George Cazenavette, Tongzhou Wang, Antonio Torralba, Alexei A. Efros, and Jun-Yan Zhu. Dataset distillation by matching training trajectories. In Computer Vision and Pattern Recognition, New Orleans, LA, USA, 2022. IEEE. 2, 4, 6, 7, 8, 1, 3", + "[2] Justin Cui, Ruochen Wang, Si Si, and Cho-Jui Hsieh. Scaling up dataset distillation to imagenet-1k with constant memory. In International Conference on Machine Learning, pages 6565–6590, Honolulu, Hawaii, USA, 2023. PMLR. 2, 7", + "[3] Jiawei Du, Yidi Jiang, Vincent Y. F. Tan, Joey Tianyi Zhou, and Haizhou Li. Minimizing the accumulated trajectory error to improve dataset distillation. In Computer Vision and Pattern Recognition, pages 3749-3758, Vancouver, BC, Canada, 2023. IEEE. 2, 4", + "[4] Jianping Gou, Baosheng Yu, Stephen J Maybank, and Dacheng Tao. Knowledge distillation: A survey. International Journal of Computer Vision, 129(6):1789-1819, 2021. 1", + "[5] Kaiming He, Xiangyu Zhang, and Shaoqing Ren. Deep residual learning for image recognition. In Computer Vision and Pattern Recognition, pages 770-778, Las Vegas, NV, USA, 2016. IEEE. 6", + "[6] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network, 2015. 3, 1", + "[7] Shengyuan Hu, Jack Goetz, Kshitiz Malik, Hongyuan Zhan, Zhe Liu, and Yue Liu. Fedsynth: Gradient compression via synthetic data in federated learning. arXiv preprint arXiv:2204.01273, 2022. 2", + "[8] Tao Huang, Shan You, Fei Wang, Chen Qian, and Chang Xu. Knowledge distillation from a stronger teacher. In Neural Information Processing Systems, 2022. 3, 1", + "[9] Jang-Hyun Kim, Jinuk Kim, Seong Joon Oh, Sangdoo Yun, Hwanjun Song, Joonhyun Jeong, Jung-Woo Ha, and Hyun Oh Song. Dataset condensation via efficient synthetic-data parameterization. In International Conference on Machine Learning, pages 11102-11118, Baltimore, Maryland, USA, 2022. PMLR. 2", + "[10] Taehyeon Kim, Jaehoon Oh, NakYil Kim, Sangwook Cho, and Se-Young Yun. Comparing kullback-leibler divergence and mean squared error loss in knowledge distillation. In International Joint Conference on Artificial Intelligence, Virtual Event, 2021. Morgan Kaufmann. 5", + "[11] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6", + "[12] Yanqing Liu, Jianyang Gu, Kai Wang, Zheng Zhu, Wei Jiang, and Yang You. DREAM: efficient dataset distillation by representative matching. arXiv preprint arXiv:2302.14416, 2023. 2", + "[13] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In International Conference on Computer Vision, pages 10012-10022, 2021. 6", + "[14] Wojciech Masarczyk and Ivona Tautkute. Reducing catastrophic forgetting with learning on synthetic data. In Com" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "puter Vision and Pattern Recognition Workshops, pages 252-253, Virtual Event, 2020. IEEE. 2", + "[15] Timothy Nguyen, Zhourong Chen, and Jaehoon Lee. Dataset meta-learning from kernel ridge-regression. arXiv preprint arXiv:2011.00050, 2020. 2, 6, 7, 8", + "[16] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In Neural Information Processing Systems, Vancouver, BC, Canada, 2019. 6, 2", + "[17] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In International Conference on Learning Representations, 2023. 4, 2", + "[18] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International Journal of Computer Vision, 115(3):211-252, 2015. 2, 5", + "[19] Noveen Sachdeva and Julian McAuley. Data distillation: A survey. arXiv preprint arXiv:2301.04272, 2023. 2", + "[20] Ahmad Sajedi, Samir Khaki, Ehsan Amjadian, Lucy Z Liu, Yuri A Lawryshyn, and Konstantinos N Plataniotis. Datadam: Efficient dataset distillation with attention matching. In International Conference on Computer Vision, pages 17097-17107, Paris, France, 2023. IEEE. 3, 6, 7, 8, 1", + "[21] Mark Sandler, Andrew G. Howard, Menglong Zhu, Andrey Zhmoginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted residuals and linear bottlenecks*. In *Computer Vision and Pattern Recognition*, pages 4510-4520, Salt Lake City, UT, USA, 2018. IEEE. 6", + "[22] Mattia Sangermano, Antonio Carta, Andrea Cossu, and Davide Bacciu. Sample condensation in online continual learning. In International Joint Conference on Neural Networks, pages 1-8, Padua, Italy, 2022. IEEE. 2", + "[23] Zhiqiang Shen and Eric Xing. A fast knowledge distillation framework for visual recognition. In European Conference on Computer Vision, pages 673-690. Springer, 2022. 3", + "[24] Felipe Petroski Such, Aditya Rawal, Joel Lehman, Kenneth O. Stanley, and Jeffrey Clune. Generative teaching networks: Accelerating neural architecture search by learning to generate synthetic training data. In International Conference on Machine Learning, pages 9206-9216, Virtual Event, 2020. PMLR. 2", + "[25] Mingxing Tan and Quoc Le. Efficientnetv2: Smaller models and faster training. In International Conference on Machine Learning, pages 10096-10106, Virtual Event, 2021. PMLR. 6", + "[26] Amirhossein Tavanaei. Embedded encoder-decoder in convolutional networks towards explainable AI. 2020. 5", + "[27] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive representation distillation. In International Conference on Learning Representations, 2019. 1", + "[28] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16717", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357, Virtual Event, 2021. PMLR. 6", + "[29] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-SNE. Journal of Machine Learning Research, 9 (11), 2008. 8", + "[30] Chunwei Wang, Chao Ma, Ming Zhu, and Xiaokang Yang. Pointaugmenting: Cross-modal augmentation for 3d object detection. In Computer Vision and Pattern Recognition, pages 11794-11803, Virtual Event, 2021. 2", + "[31] Kai Wang, Bo Zhao, Xiangyu Peng, Zheng Zhu, Shuo Yang, Shuo Wang, Guan Huang, Hakan Bilen, Xinchao Wang, and Yang You. Cafe: Learning to condense dataset by aligning features. In Computer Vision and Pattern Recognition, pages 12196-12205, New Orleans, LA, USA, 2022. IEEE. 2, 3, 7", + "[32] Tongzhou Wang, Jun-Yan Zhu, Antonio Torralba, and Alexei A Efros. Dataset distillation. arXiv preprint arXiv:1811.10959, 2018. 1, 2, 8, 3", + "[33] Hongxu Yin, Pavlo Molchanov, Jose M Alvarez, Zhizhong Li, Arun Mallya, Derek Hoiem, Niraj K Jha, and Jan Kautz. Dreaming to distill: Data-free knowledge transfer via deep-inversion. In Computer Vision and Pattern Recognition, pages 8715-8724, Virtual Event, 2020. IEEE. 2, 5", + "[34] Zeyuan Yin, Eric P. Xing, and Zhiqiang Shen. Squeeze, recover and relabel: Dataset condensation at imagenet scale from A new perspective. In Neural Information Processing Systems. NIPS, 2023. 2, 3, 4, 5, 6, 7, 1", + "[35] Shan You, Chang Xu, Chao Xu, and Dacheng Tao. Learning from multiple teacher networks. In International Conference on Knowledge Discovery and Data Mining, pages 1285-1294, Halifax, NS, Canada, 2017. ACM. 6", + "[36] Ruonan Yu, Songhua Liu, and Xinchao Wang. Dataset distillation: A comprehensive review. arXiv preprint arXiv:2301.07014, 2023. 2", + "[37] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In British Machine Vision Conference, York, UK, 2016. BMVA Press. 1, 6", + "[38] Lei Zhang, Jie Zhang, Bowen Lei, Subhabrata Mukherjee, Xiang Pan, Bo Zhao, Caiwen Ding, Yao Li, and Dongkuan Xu. Accelerating dataset distillation via model augmentation. In Computer Vision and Pattern Recognition, Vancouver, BC, Canada, 2023. IEEE. 6, 5", + "[39] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In Computer Vision and Pattern Recognition, pages 6848-6856, 2018. 6", + "[40] Bo Zhao and Hakan Bilen. Dataset condensation with differentiable siamese augmentation. In International Conference on Machine Learning, pages 12674-12685, Virtual Event, 2021. PMLR. 2", + "[41] Bo Zhao and Hakan Bilen. Dataset condensation with distribution matching. In Winter Conference on Applications of Computer Vision, pages 6514-6523, Waikoloa, Hawaii, 2023. IEEE. 2, 7, 5", + "[42] Bo Zhao, Konda Reddy Mopuri, and Hakan Bilen. Dataset condensation with gradient matching. In International Conference on Learning Representations, Virtual Event, 2021. OpenReview.net. 2, 7" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Borui Zhao, Quan Cui, Renjie Song, Yiyu Qiu, and Jiajun Liang. Decoupled knowledge distillation. In Computer Vision and Pattern Recognition, pages 11953-11962, 2022. 1", + "[44] Yongchao Zhou, Ehsan Nezhadarya, and Jimmy Ba. Dataset distillation using neural feature regression. In Neural Information Processing Systems, New Orleans, LA, USA, 2022. NIPS. 2" + ], + "bbox": [ + 501, + 92, + 890, + 189 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16718", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Generalized Large-Scale Data Condensation via Various Backbone and Statistical Matching/1c0e887f-135a-4bea-8bad-3d161e8c65c8_model.json b/2024/Generalized Large-Scale Data Condensation via Various Backbone and Statistical Matching/1c0e887f-135a-4bea-8bad-3d161e8c65c8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..60c38ab6b216ddef487178c3a9adf90883006861 --- /dev/null +++ b/2024/Generalized Large-Scale Data Condensation via Various Backbone and Statistical Matching/1c0e887f-135a-4bea-8bad-3d161e8c65c8_model.json @@ -0,0 +1,1969 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.13, + 0.845, + 0.177 + ], + "angle": 0, + "content": "Generalized Large-Scale Data Condensation via Various Backbone and Statistical Matching" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.202, + 0.877, + 0.277 + ], + "angle": 0, + "content": "Shitong Shao\\(^{1,3}\\) Zeyuan Yin\\(^{1}\\) Muxin Zhou\\(^{1}\\) Xindong Zhang\\(^{2,3}\\) Zhiqiang Shen\\(^{1,*}\\) \n\\(^{1}\\)Mohamed bin Zayed University of AI \\(^{2}\\)Hong Kong Polytechnic University \\(^{3}\\)OPPO Research 1090784053sst@gmail.com, {zeyuan.yin,muxin.zhou,zhiqiang.shen}@mbzuai.ac.ae 17901410r@connect.polyu.hk, *:Corresponding author" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.31, + 0.331, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.332, + 0.31, + 0.532, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.31, + 0.736, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.736, + 0.31, + 0.889, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.532, + 0.894, + 0.59 + ], + "angle": 0, + "content": "Figure 1. Left: Our proposed G-VBSM consists of three novel and effective modules, named DD, GSM and GBM. The richness and quality of information in the synthetic data have been significantly enhanced compared with the baseline SRe2L through the sequential merging of DD, GBM, and GSM. Right: G-VBSM prioritizes \"generalized matching\" to ensure consistency between distilled and complete datasets across various backbones, layers, and statistics, and achieves the highest accuracy \\(31.4\\%\\) on ImageNet-1k under IPC 10." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.601, + 0.314, + 0.617 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.473, + 0.903 + ], + "angle": 0, + "content": "The lightweight \"local-match-global\" matching introduced by SRe2L successfully creates a distilled dataset with comprehensive information on the full \\(224 \\times 224\\) ImageNet-1k. However, this one-sided approach is limited to a particular backbone, layer, and statistics, which limits the improvement of the generalization of a distilled dataset. We suggest that sufficient and various \"local-match-global\" matching are more precise and effective than a single one and have the ability to create a distilled dataset with richer information and better generalization ability. We call this perspective \"generalized matching\" and propose Generalized Various Backbone and Statistical Matching (G-VBSM) in this work, which aims to create a synthetic dataset with densities, ensuring consistency with the complete dataset across various backbones, layers, and statistics. As experimentally demonstrated, G-VBSM is the first algorithm to obtain strong performance across both" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.602, + 0.895, + 0.68 + ], + "angle": 0, + "content": "small-scale and large-scale datasets. Specifically, G-VBSM achieves performances of \\(38.7\\%\\) on CIFAR-100, \\(47.6\\%\\) on Tiny-ImageNet, and \\(31.4\\%\\) on the full \\(224 \\times 224\\) ImageNet-1k, respectively1. These results surpass all SOTA methods by margins of \\(3.9\\%\\), \\(6.5\\%\\), and \\(10.1\\%\\), respectively." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.688, + 0.633, + 0.703 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.714, + 0.895, + 0.851 + ], + "angle": 0, + "content": "With the development of deep learning, the number of model parameters and the quantity of training data have become increasingly large [4, 37]. Researchers have tried to minimize the training overhead while preventing a decline in the generalization ability. Data condensation (DC), also known as Dataset distillation, first introduced by Wang et al. [32], aims to alleviate the training burden by synthesizing a small yet informative distilled dataset derived from the complete training dataset, while ensuring that the behav" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.863, + 0.895, + 0.9 + ], + "angle": 0, + "content": "1Settings: CIFAR-100 with 128-width ConvNet under 10 images per class (IPC), Tiny-ImageNet with ResNet18 under 50 IPC, and ImageNet1k with ResNet18 under 10 IPC." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16709" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.212 + ], + "angle": 0, + "content": "ior of the distilled dataset on the target task remains consistent with that of the complete dataset. The extremely compressed distilled dataset contains sufficiently valuable information and have the potential for fast model training, and have been become a popular choice for different downstream application, like federated learning [7, 22], continual learning [14, 22, 40], neural architecture search [24, 41, 42] and 3D point clouds [30]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.215, + 0.474, + 0.533 + ], + "angle": 0, + "content": "A persistent problem that researchers have been working to solve [2, 9, 12, 42] in DC is the substantial data synthesis overhead [19, 36], which hinders its application in real-world large-scale datasets (e.g., ImageNet-1k) training. Typical performance matching [15, 32, 44] and trajectory matching [1, 3] unroll recursive computation graphs, requiring substantial GPU memory and resulting in prohibitive training costs. Zhao et al. [42] proposed gradient matching to address this, synthesizing distilled datasets by matching gradients from synthetic and real data in a single step. However, gradient computation and matching remain time-consuming [9], leading to the proposal of distribution matching [41]. This method and its variants [31] employ a network-based feature extractor to embed both synthetic and real data into a high-dimensional Hilbert space, then perform distribution matching. The training load for this direct, single-step process stems only from one gradient update of the synthetic data and the feature extractor [36]. Unfortunately, all of the above mentioned improved methods still have extremely large training overheads on the full \\(224 \\times 224\\) ImageNet-1k." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.536, + 0.471, + 0.763 + ], + "angle": 0, + "content": "Recently, SRe2L [34] accomplished data condensation for the first time on the full \\(224 \\times 224\\) ImageNet-1k [18], achieving Top-1 validation accuracy \\(21.3\\%\\) with ResNet18 under IPC 10. This method outperformed the latest state-of-the-art TESLA [2], which conducted on a low-resolution version of ImageNet-1k, by being \\(16 \\times\\) faster and improved performance by a margin of \\(13.6\\%\\). SRe2L is inspired by DeepInversion [33] and aims to match statistics in BatchNorm generated from synthetic and real data. We reevaluate the success of SRe2L through the lightweight \"local-match-global\" essentially. The \"local-match-global\" refers to utilizing more comprehensive information (e.g., statistics in BatchNorm), generated from the model using the complete dataset (global), to guide the parameter update of the distilled dataset (local)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "However, such lightweight and effective matching in SRe2L is singular, depending on the particular layer (i.e., BatchNorm), model (i.e., ResNet18), and statistics (i.e., channel mean/variance). Intuitively, sufficient \"local-match-global\" matching can result in more accurate and rational supervision than a single one, further enhancing the generalization of the distilled dataset. We call this perspective \"generalized matching\" and suggest that the distilled dataset is likely to perform consistent with the complete" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "dataset on the evaluation model, provided that there is sufficient variety in backbones, layers, and statistics used for matching." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.139, + 0.893, + 0.486 + ], + "angle": 0, + "content": "Inspired by this, we propose Generalized Various Backbone and Statistical Matching (G-VBSM) to fulfill \"generalized matching\". G-VBSM comprises three integral and effective parts named data densification (DD), generalized statistical matching (GSM), and generalized backbone matching (GBM). DD is utilized to ensure that the images within each class are linearly independent, thereby enhancing the (intra-class) diversity of the distilled dataset. This ultimately guarantees that \"generalized matching\" preserves the rich and diverse information within the synthetic data. GBM and GSM are designed to implement \"generalized matching\". Among them, GSM traverses the complete dataset without computing and updating the gradient, to record the statistics of Convolution at both the patch and channel levels. These statistics are subsequently utilized for matching during the data synthesis phase, in conjunction with the channel-level statistics in BatchNorm. Furthermore, GBM aims to ensure consistency between distilled and complete datasets across various backbones, enhancing matching sufficiency and leading to strong generalization in the evaluation phase. In particular, G-VBSM also ensures the efficiency of dataset condensation through a series of strategies, as mentioned in Sec. 3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.488, + 0.892, + 0.671 + ], + "angle": 0, + "content": "Extensive experiments on CIFAR-10, CIFAR-100, TinyImageNet, and the full \\(224 \\times 224\\) ImageNet-1k, demonstrating that G-VBSM is the first algorithm that performs well on both small-scale and large-scale datasets. Specifically, we not only verify through ablation studies that GSM, GBM and DD are consistently reliable, but also achieve the highest \\(38.7\\%\\), \\(47.6\\%\\) and \\(31.4\\%\\) on CIFAR-100 (128-width ConvNet), Tiny-ImageNet (ResNet18), and the full \\(224 \\times 224\\) ImageNet-1k (ResNet18) under images per class (IPC) 10, 50 and 10, respectively, which outperforms all previous state-of-the-art (SOTA) methods by \\(3.9\\%\\), \\(6.5\\%\\) and \\(10.1\\%\\), respectively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.688, + 0.63, + 0.705 + ], + "angle": 0, + "content": "2. Background" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.715, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Dataset condensation (DC) represents a data synthesis procedure that aims to compress a complete, large dataset \\(\\mathcal{T} = \\{(X_i,y_i)\\}_{i = 1}^{|T|}\\) into a smaller, distilled dataset \\(S = \\{(\\tilde{X}_i,\\tilde{y}_i)\\}_{i = 1}^{|S|}\\), subject to \\(|\\mathcal{S}|\\ll |\\mathcal{T}|\\), while ensuring that an arbitrary evaluation model \\(f_{\\mathrm{eval}}(\\cdot)\\) trained on \\(S\\) yields similar performance to one trained on \\(\\mathcal{T}\\). Classical data distillation algorithms invariably require the candidate model \\(f_{\\mathrm{cand}}(\\cdot)\\) to execute one or more steps on \\(S\\) to update its parameter \\(\\theta_{\\mathrm{cand}}\\), subsequently achieving matching in terms of performance [15, 32], gradient [42], trajectory [1, 2], or distribution [31, 41]. The process \\(\\theta_{\\mathrm{cand}} - \\alpha \\nabla_{\\theta_{\\mathrm{cand}}}\\ell (f_{\\mathrm{cand}}(\\tilde{X}),\\tilde{y})\\) where \\(\\ell (\\cdot ,\\cdot)\\) and \\((\\tilde{X},\\tilde{y})\\) represent the loss function and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "16710" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.132, + 0.092, + 0.845, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.357, + 0.895, + 0.389 + ], + "angle": 0, + "content": "Figure 2. The overview of G-VBSM on the full \\(224 \\times 224\\) ImageNet-1k, which ensures the consistency between the distilled and the complete datasets across various backbones, layers and statistics via \"generalized matching\"." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.399, + 0.471, + 0.505 + ], + "angle": 0, + "content": "a batch sampled from \\( S \\), respectively, is notably time-consuming. Consequently, even the relatively swifest distribution matching [20, 31] is slower than the recent proposed SRe2L [34]. In fact, SRe2L is the only workable way to achieve DC on the full \\( 224 \\times 224 \\) ImageNet-1k, as it requires updating the parameters of the synthetic data only once per iteration." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.506, + 0.474, + 0.657 + ], + "angle": 0, + "content": "SRe2L [34] encompasses three incremental subprocesses: Squeeze, Recover, and Relabel. Squeeze is designed to train \\( f_{\\mathrm{cand}}(\\cdot) \\) containing BatchNorm in a standard manner, aiming to record the global channel mean \\( \\mathbf{BN}_l^{\\mathrm{CM}} \\) and channel variance \\( \\mathbf{BN}_l^{\\mathrm{CV}} \\) (\\( l \\) refers to the index of the \\( l \\)-th layer) via exponential moving average (EMA), extracted from \\( \\mathcal{T} \\), for subsequent matching in Recover. In Recover after that, given the channel mean \\( \\mu_l(\\tilde{X}) \\) and channel variance \\( \\sigma_l^2(\\tilde{X}) \\) in \\( l \\)-th BatchNorm obtained from \\( S \\), the statistical matching loss function can be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.662, + 0.47, + 0.69 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {B N}} (\\tilde {X}) = \\sum_ {l} \\left| \\left| \\mu_ {l} (\\tilde {X}) - \\mathbf {B} \\mathbf {N} _ {l} ^ {\\mathrm {C M}} \\right| \\right| _ {2} + \\left| \\left| \\sigma_ {l} ^ {2} (\\tilde {X}) - \\mathbf {B} \\mathbf {N} _ {l} ^ {\\mathrm {C V}} \\right| \\right| _ {2}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.696, + 0.47, + 0.725 + ], + "angle": 0, + "content": "Based on this, we can give the entire optimization objective in Recover as" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.73, + 0.47, + 0.754 + ], + "angle": 0, + "content": "\\[\n\\underset {\\tilde {X}} {\\arg \\min } \\mathcal {L} _ {\\mathrm {B N}} (\\tilde {X}) + \\ell \\left(f _ {\\text {c a n d}} (\\tilde {X}), y\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.759, + 0.471, + 0.82 + ], + "angle": 0, + "content": "where \\(y\\) denotes the ground truth label. Moreover, SRe2L assigns soft labels \\(\\tilde{y}\\) to the synthetic data, utilizing the logit-based distillation [6, 8] to improve the generalization ability of the distilled dataset. This can be denoted as" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.829, + 0.47, + 0.847 + ], + "angle": 0, + "content": "\\[\n\\tilde {y} = \\operatorname {s o f t m a x} \\left(f _ {\\text {c a n d}} (\\tilde {X}) / \\tau\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.473, + 0.903 + ], + "angle": 0, + "content": "where \\(\\tau\\) denotes the temperature to regulate the smoothness of the soft labels, thereby enhancing the distilled dataset's potential for generalization to unseen evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.399, + 0.895, + 0.522 + ], + "angle": 0, + "content": "models. The generated soft label can be stored on disk using FKD [23] so as not to defeat the purpose of DC. A crucial point of SRe2L is that BatchNorm calculates the statistics of the entire dataset using EMA, thereby providing a comprehensive and representative matching for the distilled dataset. Encouraged by this, our research focuses on applying sufficient 'local-match-global' matching to achieve \"generalized matching\"." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.537, + 0.593, + 0.553 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.563, + 0.895, + 0.76 + ], + "angle": 0, + "content": "The comprehensive and detailed framework of our proposed Generalized Various Backbone and Statistical Matching (G-VBSM) is illustrated in Fig. 2. In essence, G-VBSM employs the lightweight regularization strategy data densification (DD) to ensure both the diversity and density of the distilled dataset, ensuring that the potential of \"generalized matching\" can be fully exploited. Moreover, generalized backbone matching (GBM) and generalized statistical matching (GSM) are utilized to achieve \"generalized matching\" by performing \"local-match-global\" matching across various backbones, layers, and statistics. In particular, the efficiency and effectiveness of DD, GBM, GSM, SRe2L, and TESLA are illustrated in Fig. 5." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.771, + 0.681, + 0.787 + ], + "angle": 0, + "content": "3.1. Data Densification" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.903 + ], + "angle": 0, + "content": "As illustrated in Fig. 3, the synthetic data generated by SRe2L exhibit excessive similarity within the same class, leading to a lack of diversity. Consequently, even if \"generalized matching\" preserves sufficient valuable information within a single image, the aggregate information content across all images within the same class does not increase effectively, which ultimately prevents \"generalized match-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16711" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.09, + 0.457, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.244, + 0.47, + 0.273 + ], + "angle": 0, + "content": "Figure 3. Visualization of the mean cosine similarity between pairwise samples within the same class on ImageNet-1k under IPC 10." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.282, + 0.468, + 0.373 + ], + "angle": 0, + "content": "ing\" from being sufficiently advantageous. Data densification (DD) is proposed to address this by ensuring the data \\(\\tilde{X}\\) has full rank in the batch dimension, thereby guaranteeing that samples in each class are linearly independent, and ultimately ensuring that the data is diverse and abundant to fully exploit the potential of \"generalized matching\"." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.373, + 0.468, + 0.418 + ], + "angle": 0, + "content": "To execute this pipeline, \\( \\tilde{X} \\) first needs to be downsampled to reduce the computational cost of eigenvalue decomposition:" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.425, + 0.469, + 0.443 + ], + "angle": 0, + "content": "\\[\n\\hat {X} = \\operatorname {A v g P o o l 2 d} (\\tilde {X}), \\text {s . t .} \\hat {X} \\in \\mathbb {R} ^ {B \\times C \\times 3 2 \\times 3 2}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.451, + 0.469, + 0.622 + ], + "angle": 0, + "content": "where \\(B\\) and \\(C\\) represent the batch size and the number of channels, respectively. Afterward, we reshape \\(\\hat{X}\\) from \\(B\\times C\\times 32\\times 32\\) to \\(B\\times (1024C)\\) and perform matrix multiplication in each class \\(y\\) to obtain the set of the Gram matrix \\(\\{\\hat{X}_y\\hat{X}_y^T\\}_{y\\in \\mathcal{Y}}\\), where \\(\\mathcal{V}\\) refers to a set of all classes existing in this batch. And \\(\\hat{X}_y\\) is a subbatch with class \\(y\\). Note that the alternative form \\(\\{\\hat{X}_y^T\\hat{X}_y\\}_{y\\in \\mathcal{Y}}\\) is not desirable, as it is applicable only for dimensionality reduction in feature dimensions, which is why we do not consider singular value decomposition (SVD). To render \\(\\hat{X}_y\\hat{X}_y^T\\) as full-rank as possible, we introduce the data densification loss in Eq. 5." + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.629, + 0.469, + 0.659 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {D D}} (\\tilde {X}) = \\sum_ {y \\in \\mathcal {Y}} D _ {\\mathrm {K L}} (\\operatorname {s t o p} - \\operatorname {g r a d} (p \\left(\\Sigma_ {y} / \\tau\\right)) | | p \\left(\\Sigma_ {y}\\right)), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.667, + 0.469, + 0.842 + ], + "angle": 0, + "content": "where \\(\\Sigma_y\\), \\(\\tau\\), stop\\_grad(\\cdot) and \\(p(\\cdot)\\) refer to the eigenvalues of \\(\\hat{X}_y\\hat{X}_y^T\\), the temperature, the stop gradient operator and the softmax function, respectively. And \\(D_{\\mathrm{KL}}(\\cdot ||\\cdot)\\) denotes Kullback-Leibler divergence. As demonstrated in Sec. E, the diversity of the data is significantly enhanced by the employment of Eq. 5. In our experiment, \\(\\tau\\) is set as 4 in default and we do not assign a deliberate weight (set to 1 by default) to \\(\\mathcal{L}_{\\mathrm{DD}}\\) because \\(\\mathcal{L}_{\\mathrm{DD}} \\equiv 0\\) at the early \\(10\\%\\) of the iterations. In other words, DD is quite stable and the optimization objective \\(\\arg \\min_{\\{\\Sigma_y\\}_{y \\in \\mathcal{Y}}} \\mathcal{L}_{\\mathrm{DD}}\\) is equivalent to \\(\\arg \\min_{\\{\\Sigma_y\\}_{y \\in \\mathcal{Y}}} \\sum_y \\sigma^2(\\Sigma_y)\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Technical Detail. A problem that warrants attention is that in the SRe2L's implementation [34], having merely a single sample in each class of a batch indicates insufficient" + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.088, + 0.865, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.221, + 0.89, + 0.235 + ], + "angle": 0, + "content": "Figure 4. The illustration of the original loop and the reorder loop." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.244, + 0.892, + 0.457 + ], + "angle": 0, + "content": "to execute DD under the order of the original loop, as depicted in Fig. 4 (Left). A simple solution is to translate the original loop to the reorder loop, as shown in Fig. 4 (Right). However, our experiment on ResNet50 (i.e., the evaluation model) substantiates that this solution suffers a \\(2.6\\%\\) accuracy degradation (details can be found in Sec. 4.1) on ImageNet-1k under IPC 10. The reason is that the number of classes in each iteration within the reorder loop is insufficient, preventing a batch have the ability to match the global statistics in BatchNorm (i.e., \\(\\mathbf{BN}_l^{\\mathrm{CM}}\\) and \\(\\mathbf{BN}_l^{\\mathrm{CV}}\\)). Motivated by score distillation sampling (SDS) [17], we update the statistics during data synthesis via EMA to solve this issue, so that the statistics of all past batches can assist the statistics of the current batch match \\(\\mathbf{BN}_l^{\\mathrm{CM}}\\) and \\(\\mathbf{BN}_l^{\\mathrm{CV}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.461, + 0.891, + 0.537 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mu_ {l} ^ {\\text {t o t a l}} = \\alpha \\mu_ {l} ^ {\\text {t o t a l}} + (1 - \\alpha) \\mu_ {l} (\\tilde {X}), \\sigma_ {l} ^ {2, \\text {t o t a l}} = \\alpha \\sigma_ {l} ^ {2, \\text {t o t a l}} + (1 - \\alpha) \\sigma_ {l} ^ {2} (\\tilde {X}), \\\\ \\mathcal {L} _ {\\mathrm {B N}} ^ {\\prime} (\\tilde {X}) = \\sum_ {l} | | \\mu_ {l} (\\tilde {X}) - \\mathbf {B N} _ {l} ^ {\\mathrm {C M}} - \\operatorname {s t o p \\_ g r a d} (\\mu_ {l} (\\tilde {X}) - \\mu_ {l} ^ {\\text {t o t a l}}) | | _ {2} \\\\ + \\left\\| \\sigma_ {l} ^ {2} (\\tilde {X}) - \\mathbf {B N} _ {l} ^ {\\mathrm {C V}} - \\operatorname {s t o p \\_ g r a d} \\left(\\sigma_ {l} ^ {2} (\\tilde {X}) - \\sigma_ {l} ^ {2, \\text {t o t a l}}\\right) \\right\\| _ {2}. \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.892, + 0.597 + ], + "angle": 0, + "content": "The derivation of \\(\\mathcal{L}_{\\mathrm{BN}}^{\\prime}(X)\\) can be found in Appendix B. We call this lightweight strategy as \"match in the form of score distillation sampling\" and have demonstrated its effectiveness and feasibility in our ablation studies." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.606, + 0.788, + 0.623 + ], + "angle": 0, + "content": "3.2. Generalized Backbone Matching" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Performing data synthesis only on a single pre-trained model is not able to enjoy ensemble gains from various backbones. Meanwhile, classical DC algorithms such as MTT [1] and FTD [3] obtain performance improvements from multiple randomly initialized backbones. Therefore, introducing generalized backbone matching (GBM) to apply various backbones for data synthesis is a desirable choice for \"generalized matching\". It ensures a number of \"local-match-global\" matching nearly \\( N_{\\mathrm{b}} \\times \\) higher compared to just depending on a single backbone, where \\( N_{\\mathrm{b}} \\) denotes the number of backbones. Regrettably, unrolling various backbone computational graphs in parallel for data synthesis is extremely expensive at the computational cost level. A solution is to randomly sample a backbone from the candidates in per iteration. This simple yet effective strategy not only ensures computational efficiency but also allows the data synthesis process to benefit from the diverse and multifaceted matching that the various backbones pro" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16712" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.294, + 0.108 + ], + "angle": 0, + "content": "vide. We denote this pipeline as" + }, + { + "type": "equation", + "bbox": [ + 0.109, + 0.113, + 0.47, + 0.128 + ], + "angle": 0, + "content": "\\[\nf _ {\\text {c a n d}} (\\cdot) \\sim \\mathcal {U} (\\mathbb {S}), \\mathbb {S} = \\{\\text {R e s N e t 1 8}, \\dots , \\text {S h u f f l e N e t V 2 - 0 . 5} \\}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.135, + 0.47, + 0.182 + ], + "angle": 0, + "content": "To maintain backbone consistency in both the data synthesis and soft label generation phases, we introduce a backbone ensemble during soft label generation:" + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.192, + 0.47, + 0.267 + ], + "angle": 0, + "content": "\\[\n\\tilde {z} = \\left\\{ \\begin{array}{c c} \\sum_ {f _ {\\text {c a n d}} \\in \\mathbb {S}} \\frac {\\sum_ {g \\in \\mathbb {S}} | | g (\\tilde {X}) | | _ {\\mathrm {F}}}{| \\mathbb {S} | | | f _ {\\text {c a n d}} (\\tilde {X}) | | _ {\\mathrm {F}}} \\frac {f _ {\\text {c a n d}} (\\tilde {X})}{| \\mathbb {S} |}, & \\mathrm {w / L N}, \\\\ \\sum_ {f _ {\\text {c a n d}} \\in \\mathbb {S}} \\frac {f _ {\\text {c a n d}} (\\tilde {X})}{| \\mathbb {S} |}, & \\mathrm {w / o L N}, \\end{array} \\right. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.277, + 0.473, + 0.43 + ], + "angle": 0, + "content": "where \\( ||\\cdot||_{\\mathrm{F}} \\) and LN refer to the Frobenius norm and Logit Normalization, respectively. Thus, we can obtain the soft label by \\( \\tilde{y} = \\text{softmax}(\\tilde{z}/\\tau) \\). Particularly, the use of LN is optional, as demonstrated by our ablation studies; it is beneficial for ResNet18 but not for ResNet50 or ResNet101. Moreover, it's important to highlight that we apply a parallel mechanism for soft label generation since it provides a significantly lower computational cost, less than 1/30 of that required for data synthesis, thus making the computational overhead negligible." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.439, + 0.367, + 0.456 + ], + "angle": 0, + "content": "3.3. Generalized Statistical Matching" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.463, + 0.473, + 0.691 + ], + "angle": 0, + "content": "Only ensuring backbone diversity is insufficient to fully exploit the potential of \"generalized matching\". In this subsection, we aim to introduce additional statistics for matching during the data synthesis phase. Prior methods [33, 34] only utilize BatchNorm since the presence of global information statistics within the pre-trained model, with no apparent solution ready for other layers (e.g., Convolution). Retraining the model is the simplest way to address this problem, followed by updating the other layers' parameters through EMA. This approach is obviously impractical, as it defeats the purpose of using gradient descent. By contrast, we propose to allow the pre-trained model \\( f_{\\mathrm{eval}}(\\cdot) \\) to loop through the training dataset \\( \\mathcal{T} \\) once without calculating the gradient to obtain the global statistics of Convolution, thus serving \"local-match-global\":" + }, + { + "type": "equation", + "bbox": [ + 0.126, + 0.697, + 0.47, + 0.74 + ], + "angle": 0, + "content": "\\[\n\\mathbf {C o n v} _ {l} ^ {\\mathbf {C M}} = \\frac {1}{| \\mathcal {T} |} \\sum_ {i} ^ {| \\mathcal {T} |} \\mathbf {C M} _ {i} ^ {l}, \\mathbf {C o n v} _ {l} ^ {\\mathbf {C V}} = \\frac {1}{| \\mathcal {T} |} \\sum_ {i} ^ {| \\mathcal {T} |} \\mathbf {C V} _ {i} ^ {l}, \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.129, + 0.737, + 0.413, + 0.772 + ], + "angle": 0, + "content": "\\[\n\\mathbf {C o n v} _ {l} ^ {\\mathbf {P M}} = \\frac {1}{| \\mathcal {T} |} \\sum_ {i} ^ {| \\mathcal {T} |} \\mathbf {P M} _ {i} ^ {l}, \\mathbf {C o n v} _ {l} ^ {\\mathbf {P V}} = \\frac {1}{| \\mathcal {T} |} \\sum_ {i} ^ {| \\mathcal {T} |} \\mathbf {P V} _ {i} ^ {l}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.779, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Here, \\(\\mathbf{CM}_i^l\\in \\mathbb{R}^{C_l}\\) \\(\\mathbf{CV}_i^l\\in \\mathbb{R}^{C_l}\\) \\(\\mathbf{PM}_i^l\\in \\mathbb{R}^{\\left[\\frac{H}{N_l^p}\\right]\\times \\left[\\frac{W}{N_l^p}\\right]}\\) and \\(\\mathbf{PV}_i^l\\in \\mathbb{R}^{\\left[\\frac{H}{N_l^p}\\right]\\times \\left[\\frac{W}{N_l^p}\\right]}\\) refer to the channel mean, the channel variance, the patch mean and the patch variance, respectively, for the \\(l\\) -th Convolution when the input to \\(f_{\\mathrm{cand}}(\\cdot)\\) is the \\(i\\) -th batch, where \\(C_l\\) and \\(\\left[\\frac{H}{N_l^p}\\right]\\times \\left[\\frac{W}{N_l^p}\\right]\\) denote the number of channels and patches of the \\(l\\) -th Convolution, respectively. We define \\(N_{l}^{p}\\) as 4, 4, 4 and 16 by default" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.201 + ], + "angle": 0, + "content": "on CIFAR-10, CIFAR-100, Tiny-ImageNet and ImageNet-1k, respectively. After obtaining the global channel mean \\(\\mathbf{Conv}_l^{\\mathbf{CM}}\\), the global channel variance \\(\\mathbf{Conv}_l^{\\mathbf{CV}}\\), the global patch mean \\(\\mathbf{Conv}_l^{\\mathbf{PM}}\\) and the global patch variance \\(\\mathbf{Conv}_l^{\\mathbf{PV}}\\), we can store them in a disk thus avoiding secondary calculations. In the data synthesis phase, we introduce \\(\\mathcal{L}_{\\mathrm{Conv}}'(\\tilde{X})\\) in Eq. 10 to accomplish joint matching with \\(\\mathcal{L}_{\\mathrm{BN}}'(\\tilde{X})\\)." + }, + { + "type": "equation", + "bbox": [ + 0.501, + 0.209, + 0.891, + 0.303 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {C o n v}} ^ {\\prime} (\\tilde {X}) = \\sum_ {l} | | \\mu_ {l} ^ {c} (\\tilde {X}) - \\mathbf {C o n v} _ {l} ^ {\\mathrm {C M}} - \\operatorname {s t o p - g r a d} \\left(\\mu_ {l} ^ {c} (\\tilde {X}) - \\mu_ {l} ^ {c, \\text {t o t a l}}\\right) | | _ {2} \\\\ + \\left\\| \\sigma_ {l} ^ {c, 2} (\\tilde {X}) - \\mathbf {C o n v} _ {l} ^ {\\mathrm {C V}} - \\operatorname {s t o p - g r a d} \\left(\\sigma_ {l} ^ {c, 2} (\\tilde {X}) - \\sigma_ {l} ^ {c, 2, \\text {t o t a l}}\\right) \\right\\| _ {2}, \\\\ + \\left\\| \\mu_ {l} ^ {p} (\\tilde {X}) - \\mathbf {C o n v} _ {l} ^ {\\mathrm {P M}} - \\operatorname {s t o p - g r a d} \\left(\\mu_ {l} ^ {p} (\\tilde {X}) - \\mu_ {l} ^ {p, \\text {t o t a l}}\\right) \\right\\| _ {2} \\\\ + \\left\\| \\sigma_ {l} ^ {p, 2} (\\tilde {X}) - \\mathbf {C o n v} _ {l} ^ {\\mathrm {P V}} - \\operatorname {s t o p} - \\operatorname {g r a d} \\left(\\sigma_ {l} ^ {p, 2} (\\tilde {X}) - \\sigma_ {l} ^ {p, 2, \\text {t o t a l}}\\right) \\right\\| _ {2}, \\tag {10} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.302, + 0.892, + 0.429 + ], + "angle": 0, + "content": "where \\(\\mu_l^{c,\\mathrm{total}}\\), \\(\\sigma_l^{c,2,\\mathrm{total}}\\), \\(\\mu_l^{p,\\mathrm{total}}\\), and \\(\\sigma_l^{p,2,\\mathrm{total}}\\) are each updated from the channel mean \\(\\mu_l^c (\\tilde{X})\\), channel variance \\(\\sigma_l^{c,2}(\\tilde{X})\\), patch mean \\(\\mu_l^p (\\tilde{X})\\), and patch variance \\(\\sigma_l^{p,2}(\\tilde{X})\\) respectively, all obtained via EMA from the current batch. In experiments, we discovered that Eq. 10 causes a slightly larger computational burden, so we randomly drop the matching of statistics with a probability of \\(\\beta_{\\mathrm{dr}}\\) to ensure the efficiency of GSM." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.433, + 0.893, + 0.707 + ], + "angle": 0, + "content": "Loss Function in the Evaluation Phase. Unlike DD, GBM and GSM are designed to create a distilled dataset that is enriched with information. Here, we introduce an enhancement to the loss function tailored specifically for the evaluation phase. Essentially, the evaluation phase is a knowledge distillation framework for transferring knowledge from a pre-trained model to the evaluation model. SRe2L utilizes \\( D_{\\mathrm{KL}}(\\tilde{y} || \\mathrm{softmax}(f_{\\mathrm{eval}}(\\tilde{X}) / \\tau)) \\) as the loss function and experimentally illustrates that it improves performance by roughly \\( 10\\% \\). As established in SRe2L, an increase in temperature \\( \\tau \\) correlates with enhanced performance of the evaluation model. Inspired by this and \\( \\tau^2 D_{\\mathrm{KL}}(\\mathrm{softmax}(p) / \\tau ||\\mathrm{softmax}(q) / \\tau) \\) is equivalent to \\( \\frac{1}{2C} ||p - q||_2^2 \\) when \\( \\tau \\rightarrow +\\infty \\) [10], we introduce a novel loss function \\( \\mathrm{MSE} + \\gamma \\times \\mathrm{GT} \\) to avoid numerical error caused by the large \\( \\tau \\) and improve the generalization of the distilled dataset without any additional overhead (ignore the weights \\( \\tau^2 \\) and \\( \\frac{1}{2C} \\)):" + }, + { + "type": "equation", + "bbox": [ + 0.517, + 0.712, + 0.891, + 0.739 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {e v a l}} (\\tilde {X}, \\tilde {y}, y) = | | f _ {\\text {e v a l}} (\\tilde {X}) - \\tilde {z} | | _ {2} ^ {2} - \\gamma \\mathbf {y} \\log \\left(\\operatorname {s o f t m a x} \\left(f _ {\\text {e v a l}} (\\tilde {X})\\right)\\right), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.74, + 0.892, + 0.815 + ], + "angle": 0, + "content": "where \\( y \\) represents the one-hot encoding (w.r.t. the ground truth label \\( y \\)). As illustrated in Fig. 5, simply replacing the loss function with \\( \\mathrm{MSE} + 0.1 \\times \\mathrm{GT} \\) (\\( \\gamma \\) is set as 0.1) in SRe2L improves the performance of ResNet18 (i.e., the evaluation model) by a margin of \\( 0.9\\% \\) on ImageNet-1k under IPC 10." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.83, + 0.625, + 0.848 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.902 + ], + "angle": 0, + "content": "We perform comparison experiments on the large-scale dataset including the full \\(224 \\times 224\\) ImageNet-1k [18] and the small-scale datasets including Tiny-ImageNet [26]," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16713" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.089, + 0.455, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.238, + 0.47, + 0.281 + ], + "angle": 0, + "content": "Figure 5. Comparison of the effectiveness and efficiency of G-VBSM components. Among them, \"DD+GBM+GSM \\((\\beta_{\\mathrm{dr}} = 0.0) + (\\mathrm{MSE} + \\mathrm{GT})\\) represents the comprehensive G-VBSM." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.294, + 0.468, + 0.34 + ], + "angle": 0, + "content": "CIFAR-10/100 [11]. To highlight that our proposed G-VBSM is designed for large-scale datasets, all ablation experiments are performed on ImageNet-1k." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.358, + 0.47, + 0.631 + ], + "angle": 0, + "content": "Hyperparameter Settings. We prioritize selecting various convolution-based backbone, ensuring the maximal difference in architecture, while also adhering to the criterion of minimizing the number of parameters, an approach empirically demonstrated to be superior [34, 38]. On ImageNet-1k, we skip the model pre-training phase by directly using Torchvision's open source pretraining weights [16] of {ResNet18 [5], MobileNetV2 [21], EfficientNet-B0 [25], ShuffleNetV2-0.5 [39]}. For the remaining dataset, we all train {128-width ConvNet [39], WRN-16-2 [37], ResNet18 [5], ShuffleNetV2-0.5 [39], MobileNetV2-0.5 [21]} from scratch with few epoch in the model pre-training phases. Gray cells in all tables represent the highest performance. Meanwhile, (R18), (R50), etc. in all tables, represent the evaluation models. More details about the remaining method hyperparameter settings and additional ablation studies can be found in Appendix A and Appendix C & D, respectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.645, + 0.241, + 0.659 + ], + "angle": 0, + "content": "4.1. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.669, + 0.468, + 0.911 + ], + "angle": 0, + "content": "The Trade-off between Efficiency and Effectiveness. The performance improvements and increased GPU latency for components of G-VBSM in comparison to the baseline SRe2L, are illustrated in Fig. 5. Clearly, DD, GBM, and GSM are highly effective in enhancing the generalization of distilled datasets. Compared with GSM, both DD and GBM are slightly lightweight and efficient. With DD and GBM alone - i.e., \\(\\mathrm{DD + GBM + GSM}\\) \\((\\beta_{\\mathrm{dr}} = 1.0) + (\\mathrm{MSE} + \\mathrm{GT})\\) - the accuracy of this approach surpasses that of SRe2L by a significant margin of \\(6.5\\%\\). Additionally, the comprehensive G-VBSM, \\(\\mathrm{DD + GBM + GSM}\\) \\((\\beta_{\\mathrm{dr}} = 0.0) + (\\mathrm{MSE} + \\mathrm{GT})\\), further enhances the performance of \\(\\mathrm{DD + GBM + GSM}\\) \\((\\beta_{\\mathrm{dr}} = 1.0) + (\\mathrm{MSE} + \\mathrm{GT})\\) by a notable margin of \\(3.6\\%\\). Note that \\(\\mathrm{MSE} + \\gamma \\times \\mathrm{GT}\\) (\\(\\gamma\\) is set as 0.1) is also extremely critical, for SRe2L and \\(\\mathrm{DD + GBM}\\) boosting \\(0.9\\%\\) and \\(2.7\\%\\), respectively." + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.089, + 0.868, + 0.154 + ], + "angle": 0, + "content": "
α\\Evaluation ModelResNet18(MSE+0.1×GT)ResNet50(MSE+0.1×GT)
0.030.4%31.9%
0.431.4%34.5%
0.831.9%36.4%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.171, + 0.89, + 0.186 + ], + "angle": 0, + "content": "Table 1. ImageNet-1k Top-1 Acc. about different \\( \\alpha \\) under IPC 10." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.193, + 0.892, + 0.344 + ], + "angle": 0, + "content": "Matching in the Form of SDS. Benefiting from this matching form, G-VBSM can achieve the best performance with small batch sizes (e.g., 40 in ImageNet-1k and 50 in CIFAR-100) in different settings, as demonstrated in our comparative experiments. Table 1 demonstrates the impact of the factor \\(\\alpha\\) for performing the EMA update on the final performance achieved by G-VBSM, where \\(\\alpha = 0\\) indicates that \\(\\mathcal{L}_{\\mathrm{BN}}^{\\prime}\\) degenerates to \\(\\mathcal{L}_{\\mathrm{BN}}\\). We can conclude that the SDS matching form is feasible and the generalization of the distilled dataset improves with increasing \\(\\alpha\\)." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.357, + 0.902, + 0.415 + ], + "angle": 0, + "content": "
MethodEvaluation Model
DeiT-TinyResNet18MobileNetV2Swin-Tiny
SRe2L15.41%46.79%36.59%39.23%
G-VBSM29.43%51.82%48.66%57.40%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.425, + 0.89, + 0.452 + ], + "angle": 0, + "content": "Table 2. ImageNet-1k Top-1 Acc. on cross-architecture generalization under IPC 50." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.46, + 0.892, + 0.717 + ], + "angle": 0, + "content": "Cross-Architecture Generalization. The evaluation of cross-architecture generalization is crucial in assessing the quality of distilled datasets. Unlike traditional methods [1, 15, 20] which focus on CIFAR-100, our approach evaluates the effectiveness of the distilled dataset on ImageNet1k, employing a suite of models with real-world applicability, including ResNet18 [5], MobileNetV2 [21], DeiTiny [28], and Swin-Tiny [13]. The experimental results are reported in Table 2. From Tables 7 and 2, it is evident that the G-VBSM-synthetic dataset can effectively generalize across ResNet {18, 50, 101}, MobileNetV2, DeiTiny, and Swin-Tiny architectures. Notably, DeiT-Tiny and Swin-Tiny, two architectures not encountered during the data synthesis phase, demonstrates significant proficiency with the accuracy \\(29.43\\%\\) and \\(57.40\\%\\), outperforming the latest SOTA SRe2L by a margin of \\(14.01\\%\\) and \\(18.17\\%\\), respectively." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.731, + 0.902, + 0.798 + ], + "angle": 0, + "content": "
Candidates (Backbone)Evaluation Model
ResNet18MobileNetV2EfficientNet-B0ShuffleNetV2-0.5ResNet18ResNet50
✓25.7%30.2%
✓✓27.2%31.2%
✓✓✓27.9%32.0%
✓✓✓✓31.4%34.5%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.809, + 0.89, + 0.837 + ], + "angle": 0, + "content": "Table 3. ImageNet-1k Top-1 Acc. about the number of candidate backbone in the soft label generation pipeline under IPC 10." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Ensemble in Soft Label Generation. In knowledge distillation, soft labels obtained through multiple teacher ensembles can effectively enhance the generalization of the student [35]. This observation is similarly corroborated in" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16714" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.201, + 0.089, + 0.368, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.395, + 0.089, + 0.567, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.604, + 0.09, + 0.768, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.203, + 0.233, + 0.766, + 0.248 + ], + "angle": 0, + "content": "Figure 6. ImageNet-1k Top-1 Acc. of different loss function, \\(\\beta_{\\mathrm{dr}}\\) and batch sizes under IPC 10." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.254, + 0.891, + 0.326 + ], + "angle": 0, + "content": "
DatasetIPCMTT [1] (CW128)DataDAM [20] (CW128)TESLA [2] (R18)SRe2L (R18)SRe2L (R50)SRe2L (R101)G-VBSM (R18)G-VBSM (R50)G-VBSM (R101)
Tiny-ImageNet5028.0±0.328.7±0.3-41.1±0.442.2±0.542.5±0.247.6±0.348.7±0.248.8±0.4
100---49.7±0.351.2±0.451.5±0.351.0±0.452.1±0.352.3±0.1
ImageNet-1k1064.0±1.3†6.3±0.07.7±0.121.3±0.628.4±0.130.9±0.131.4±0.535.4±0.838.2±0.4
50-15.5±0.2-46.8±0.255.6±0.360.8±0.551.8±0.458.7±0.361.0±0.4
100---52.8±0.361.0±0.462.8±0.255.7±0.462.2±0.363.7±0.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.335, + 0.893, + 0.364 + ], + "angle": 0, + "content": "Table 4. Comparison with baseline models in Tiny-ImageNet and ImageNet-1k. \\( {}^{ \\dagger } \\) indicates the ImageNette dataset,which contains only 10 classes. DataDAM [20] and TESLA [2] use the downsampled \\( {64} \\times {64} \\) ImageNet-1k. We cite the experimental results from SRe2L [34]." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.369, + 0.892, + 0.453 + ], + "angle": 0, + "content": "
DatasetIPCCoreset SelectionTraining Set Synthesis (CW128)Training Set Synthesis (R18)Whole Dataset
RandomHerdingK-CenterForgettingDC [42]DM [41]CAFE [31]KIP [15]MTT [1]DataDAM[20]G-VBSMSRe2LG-VBSM(CW128)
CIFAR-101026.0±1.231.6±0.714.7±0.923.3±1.044.9±0.548.9±0.650.9±0.546.1±0.765.3±0.754.2±0.846.5±0.727.2±0.553.5±0.684.8±0.1
5043.4±1.040.4±0.627.0±1.423.3±1.153.9±0.563.0±0.462.3±0.453.2±0.771.6±0.267.0±0.454.3±0.347.5±0.659.2±0.4
CIFAR-10014.2±0.38.3±0.38.4±0.34.5±0.212.8±0.311.4±0.314.0±0.312.0±0.224.3±0.314.5±0.516.4±0.72.0±0.225.9±0.556.2±0.3
1014.6±0.517.3±0.317.3±0.315.1±0.325.2±0.329.7±0.331.5±0.240.1±0.433.1±0.434.8±0.538.7±0.231.6±0.559.5±0.4
5030.0±0.433.7±0.530.5±0.3-30.6±0.643.6±0.447.7±0.2-42.9±0.349.4±0.345.7±0.449.5±0.365.0±0.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.455, + 0.894, + 0.499 + ], + "angle": 0, + "content": "Table 5. Comparison with baseline models on CIFAR-10/100. All methods, except for SRe2L and G-VBSM, use a 128-width ConvNet (CW128) for data synthesis and evaluation. G-VBSM utilizes {CW128, WRN-16-2, ResNet18 (R18), ShuffleNetV2-0.5, MobileNetV2-0.5} for data synthesis and {CW128, R18} for evaluation. We cite the experimental results, except for SRe2L's, from DataDAM [20]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.512, + 0.471, + 0.633 + ], + "angle": 0, + "content": "dataset distillation, as evidenced in Table 3. The greater the number of models, the stronger the generalization of the distilled dataset. It is particularly interesting to point out that ShuffleNetV2-0.5 improves further by more than \\(2\\%\\) even though the other three models have already achieved modest ensemble gains. We attribute this enhancement to the channel shuffle mechanism within ShuffleNetV2-0.5, which imposes a beneficial regularization constraint for G-VBSM." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.646, + 0.48, + 0.702 + ], + "angle": 0, + "content": "
Logit NormalizationEvaluation Model
ResNet18 (MSE+0.1×GT)ResNet18 (KL)ResNet50 (MSE+0.1×GT)ResNet50 (KL)
Yes31.4%25.4%34.5%28.6%
No31.0%25.1%35.4%31.9%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.712, + 0.47, + 0.74 + ], + "angle": 0, + "content": "Table 6. ImageNet-1k Top-1 Acc. about the use of logit normalization under IPC 10." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Logit Normalization in Soft Label Generation. The aim of this strategy is to maintain the consistency of the logits (can be viewed as vectors) magnitude across all models within a high-dimensional space, thereby ensuring equal contribution of these models to the ultimate soft label for ensemble. This approach is not universally effective, as shown in Table 6. When the evaluation model parameter count is low (e.g., ResNet18), it can enhance G-VBSM performance. Conversely, with a model like ResNet50, it may hinder performance, rendering G-VBSM uncompetitive. As" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.512, + 0.893, + 0.556 + ], + "angle": 0, + "content": "a result, in this work, ResNet \\(\\{50, 101\\}\\) and ViT-based models do not employ logit normalization, while the remaining models do." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.562, + 0.895, + 0.744 + ], + "angle": 0, + "content": "Loss Function in the Evaluation Phase. As illustrated in Fig. 6, replacing \\(D_{\\mathrm{KL}}(\\cdot ||\\cdot)\\) in SRe2L with \\(\\mathrm{MSE} + \\gamma \\times \\mathrm{GT}\\) demonstrates to be extremely effective. For example, with \\(\\beta_{\\mathrm{dr}} = 0.0\\) and batch size \\(= 40\\), adjusting \\(\\gamma\\) to 0.01 and 0.1 enhances model performance during the evaluation phase by \\(1.4\\%\\) and \\(6.0\\%\\), respectively. As a result, we employ a \\(\\gamma\\) of 0.1 in all experiments conducted on ImageNet-1k. Furthermore, we observe that the distilled dataset's generalization is not significantly affected (\\(\\leq 2\\%\\)) by variations in either batch size or \\(\\beta_{\\mathrm{dr}}\\). This ensured that G-VBSM attain SOTA results with a minimal batch size 40 and \\(\\beta_{\\mathrm{dr}} = 0.4\\) (i.e., the default settings in ImageNet-1k)." + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.757, + 0.88, + 0.884 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.895, + 0.88, + 0.909 + ], + "angle": 0, + "content": "Figure 7. Synthetic data visualization on small-scale datasets." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16715" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "Synthetic Data Visualization. The visualization results for the large-scale (the full \\(224 \\times 224\\) ImageNet-1k) and small-scale (CIFAR-10/100 and Tiny-ImageNet) datasets are displayed in Figs. 1 and 7, respectively. On ImageNet-1k, the images synthesized by G-VBSM are more informative and abstract compared with those from SRe2L. Furthermore, on small-scale datasets, the distilled images obtained by G-VBSM markedly differ from the images presented in MTT and DataDAM's papers [1, 20], illustrating that G-VBSM is an out-of-the ordinary algorithm. More synthetic data can be found in Appendix I." + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.277, + 0.451, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.132, + 0.44, + 0.414, + 0.454 + ], + "angle": 0, + "content": "Figure 8. t-SNE visualization on ImageNet-1k." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.474, + 0.47, + 0.641 + ], + "angle": 0, + "content": "Logit Embedding Distributions Visualization. We feed distilled datasets into a pre-trained model (i.e., ResNet18) to obtain logit embeddings for t-SNE [29] visualization. Note that unlike previous methods that used feature embeddings before the classifier layer, we utilize logit embeddings after the classifier layer for presentation. Fig. 8 shows the distributions of G-VBSM and SRe2L logit embeddings. The logit embedding obtained by G-VBSM has less intra-class similarity than SRe2L, ensuring softer labels, preventing model overconfidence during the evaluation phase, and ultimately enhancing model generalization." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.651, + 0.367, + 0.667 + ], + "angle": 0, + "content": "4.2. Large-Scale Dataset Comparison" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Full \\(224 \\times 224\\) ImageNet-1k. As illustrated in the ImageNet-1k part of Table 4, G-VBSM consistently outperforms SRe2L across all IPC \\(\\{10, 50, 100\\}\\), with the evaluation model being ResNet \\(\\{18, 50, 101\\}\\), indicating that G-VBSM is highly competitive in the large-scale dataset ImageNet-1k. In particular, under extreme compression scenarios, such as IPC 10, G-VBSM achieves significant performance increases - \\(10.1\\%\\), \\(7.0\\%\\), and \\(7.3\\%\\) for ResNet18, ResNet50, and ResNet101, respectively - compared with SRe2L. In comparison to the latest SOTA classical method, DataDAM, just experimenting on the full \\(64 \\times 64\\) ImageNet-1k, we exceed it by a margin of \\(23.7\\%\\) under IPC 10. These results are particularly impressive when considered in the context of the large-scale dataset ImageNet-1k." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.789, + 0.108 + ], + "angle": 0, + "content": "4.3. Small-Scale Dataset Comparison" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.114, + 0.892, + 0.234 + ], + "angle": 0, + "content": "Tiny-ImageNet. Similar to SRe2L, we evaluate the generalization of the distilled dataset on ResNet {18, 50, 101}. As presented in the Tiny-ImageNet part of Table 4, G-VBSM consistently surpasses the latest SOTA method SRe2L on different settings. Our method outperforms SRe2L, MTT and DataDAM by \\(6.5\\%\\), \\(19.6\\%\\) and \\(18.9\\%\\), respectively, under IPC 50 applying 128-width ConvNet or ResNet18 as the evaluation model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.245, + 0.893, + 0.502 + ], + "angle": 0, + "content": "CIFAR-100. One of the disadvantages of SRe2L is its inability to compete with SOTA performance on this small-scale dataset benchmark (i.e., evaluated with 128-width ConvNet and ResNet18), such as the performance on CIFAR-100 shown in Table 5, as a result, it fails to demonstrate that \"local-match-global\" (e.g., statistical matching of BatchNorm) can rival or outperform traditional DC algorithms. In contrast, as shown in the CIFAR-100 part of Table 5, G-VBSM achieves accuracies \\(16.4\\%\\) and \\(38.7\\%\\), the highest under IPC 1 and 10 on CIFAR-100 through \"generalized matching\", while maintaining the same number of epochs (i.e., 1000) and the model (i.e., 128-width ConvNet) for the evaluation phase as the traditional methods, outperforming the latest SOTA DataDAM by \\(1.9\\%\\) and \\(3.9\\%\\), respectively. To the best of our knowledge, our proposed G-VBSM is the first algorithm that is strongly competitive on CIAR-100 as well as the full \\(224 \\times 224\\) ImageNet-1k." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.513, + 0.892, + 0.62 + ], + "angle": 0, + "content": "CIFAR-10. As illustrated in the CIFAR-10 part of Table 5, G-VBSM outperforms vanilla DC [32], KIP [15], and all coreset selection algorithms on the benchmark with 128-width ConvNet as the evaluation model. Meanwhile, on the benchmark with ResNet 18 as the evaluation model, G-VBSM outperforms the baseline SRe2L by \\(26.3\\%\\) and \\(11.7\\%\\) under IPC 10 and 50, respectively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.633, + 0.619, + 0.648 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.892, + 0.884 + ], + "angle": 0, + "content": "In this paper, we introduce a novel perspective termed \"generalized matching\" for dataset condensation. This posits that an abundance of lightweight \"local-match-global\" matching are more effective than a single \"local-match-global\" matching, and even surpass the precise and costly matching used by traditional methods. Consequently, we present G-VBSM, designed to ensure data densification while performing matching based on sufficient and various backbones, layers, and statistics. Experiments conducted on CIFAR-10/100, Tiny-[ImageNet, and the full \\(224 \\times 224\\) ImageNet-1k demonstrated that our method is the first algorithm to show strong performance across both small-scale and large-scale datasets. For future research, we plan to extend this approach's applicability to large-scale detection and segmentation datasets." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16716" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] George Cazenavette, Tongzhou Wang, Antonio Torralba, Alexei A. Efros, and Jun-Yan Zhu. Dataset distillation by matching training trajectories. In Computer Vision and Pattern Recognition, New Orleans, LA, USA, 2022. IEEE. 2, 4, 6, 7, 8, 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.472, + 0.243 + ], + "angle": 0, + "content": "[2] Justin Cui, Ruochen Wang, Si Si, and Cho-Jui Hsieh. Scaling up dataset distillation to imagenet-1k with constant memory. In International Conference on Machine Learning, pages 6565–6590, Honolulu, Hawaii, USA, 2023. PMLR. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.47, + 0.312 + ], + "angle": 0, + "content": "[3] Jiawei Du, Yidi Jiang, Vincent Y. F. Tan, Joey Tianyi Zhou, and Haizhou Li. Minimizing the accumulated trajectory error to improve dataset distillation. In Computer Vision and Pattern Recognition, pages 3749-3758, Vancouver, BC, Canada, 2023. IEEE. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.314, + 0.47, + 0.369 + ], + "angle": 0, + "content": "[4] Jianping Gou, Baosheng Yu, Stephen J Maybank, and Dacheng Tao. Knowledge distillation: A survey. International Journal of Computer Vision, 129(6):1789-1819, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.372, + 0.469, + 0.426 + ], + "angle": 0, + "content": "[5] Kaiming He, Xiangyu Zhang, and Shaoqing Ren. Deep residual learning for image recognition. In Computer Vision and Pattern Recognition, pages 770-778, Las Vegas, NV, USA, 2016. IEEE. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.429, + 0.469, + 0.457 + ], + "angle": 0, + "content": "[6] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network, 2015. 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.459, + 0.469, + 0.514 + ], + "angle": 0, + "content": "[7] Shengyuan Hu, Jack Goetz, Kshitiz Malik, Hongyuan Zhan, Zhe Liu, and Yue Liu. Fedsynth: Gradient compression via synthetic data in federated learning. arXiv preprint arXiv:2204.01273, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.516, + 0.469, + 0.558 + ], + "angle": 0, + "content": "[8] Tao Huang, Shan You, Fei Wang, Chen Qian, and Chang Xu. Knowledge distillation from a stronger teacher. In Neural Information Processing Systems, 2022. 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.56, + 0.469, + 0.642 + ], + "angle": 0, + "content": "[9] Jang-Hyun Kim, Jinuk Kim, Seong Joon Oh, Sangdoo Yun, Hwanjun Song, Joonhyun Jeong, Jung-Woo Ha, and Hyun Oh Song. Dataset condensation via efficient synthetic-data parameterization. In International Conference on Machine Learning, pages 11102-11118, Baltimore, Maryland, USA, 2022. PMLR. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.644, + 0.469, + 0.713 + ], + "angle": 0, + "content": "[10] Taehyeon Kim, Jaehoon Oh, NakYil Kim, Sangwook Cho, and Se-Young Yun. Comparing kullback-leibler divergence and mean squared error loss in knowledge distillation. In International Joint Conference on Artificial Intelligence, Virtual Event, 2021. Morgan Kaufmann. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.715, + 0.469, + 0.744 + ], + "angle": 0, + "content": "[11] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.745, + 0.469, + 0.799 + ], + "angle": 0, + "content": "[12] Yanqing Liu, Jianyang Gu, Kai Wang, Zheng Zhu, Wei Jiang, and Yang You. DREAM: efficient dataset distillation by representative matching. arXiv preprint arXiv:2302.14416, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[13] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In International Conference on Computer Vision, pages 10012-10022, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[14] Wojciech Masarczyk and Ivona Tautkute. Reducing catastrophic forgetting with learning on synthetic data. In Com" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "puter Vision and Pattern Recognition Workshops, pages 252-253, Virtual Event, 2020. IEEE. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.123, + 0.892, + 0.164 + ], + "angle": 0, + "content": "[15] Timothy Nguyen, Zhourong Chen, and Jaehoon Lee. Dataset meta-learning from kernel ridge-regression. arXiv preprint arXiv:2011.00050, 2020. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.166, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[16] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In Neural Information Processing Systems, Vancouver, BC, Canada, 2019. 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.252, + 0.892, + 0.306 + ], + "angle": 0, + "content": "[17] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In International Conference on Learning Representations, 2023. 4, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.31, + 0.892, + 0.379 + ], + "angle": 0, + "content": "[18] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International Journal of Computer Vision, 115(3):211-252, 2015. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.381, + 0.892, + 0.41 + ], + "angle": 0, + "content": "[19] Noveen Sachdeva and Julian McAuley. Data distillation: A survey. arXiv preprint arXiv:2301.04272, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.412, + 0.892, + 0.48 + ], + "angle": 0, + "content": "[20] Ahmad Sajedi, Samir Khaki, Ehsan Amjadian, Lucy Z Liu, Yuri A Lawryshyn, and Konstantinos N Plataniotis. Datadam: Efficient dataset distillation with attention matching. In International Conference on Computer Vision, pages 17097-17107, Paris, France, 2023. IEEE. 3, 6, 7, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.483, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[21] Mark Sandler, Andrew G. Howard, Menglong Zhu, Andrey Zhmoginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted residuals and linear bottlenecks*. In *Computer Vision and Pattern Recognition*, pages 4510-4520, Salt Lake City, UT, USA, 2018. IEEE. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.555, + 0.892, + 0.61 + ], + "angle": 0, + "content": "[22] Mattia Sangermano, Antonio Carta, Andrea Cossu, and Davide Bacciu. Sample condensation in online continual learning. In International Joint Conference on Neural Networks, pages 1-8, Padua, Italy, 2022. IEEE. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.612, + 0.892, + 0.655 + ], + "angle": 0, + "content": "[23] Zhiqiang Shen and Eric Xing. A fast knowledge distillation framework for visual recognition. In European Conference on Computer Vision, pages 673-690. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.656, + 0.892, + 0.738 + ], + "angle": 0, + "content": "[24] Felipe Petroski Such, Aditya Rawal, Joel Lehman, Kenneth O. Stanley, and Jeffrey Clune. Generative teaching networks: Accelerating neural architecture search by learning to generate synthetic training data. In International Conference on Machine Learning, pages 9206-9216, Virtual Event, 2020. PMLR. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.741, + 0.892, + 0.796 + ], + "angle": 0, + "content": "[25] Mingxing Tan and Quoc Le. Efficientnetv2: Smaller models and faster training. In International Conference on Machine Learning, pages 10096-10106, Virtual Event, 2021. PMLR. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.799, + 0.892, + 0.828 + ], + "angle": 0, + "content": "[26] Amirhossein Tavanaei. Embedded encoder-decoder in convolutional networks towards explainable AI. 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.83, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[27] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive representation distillation. In International Conference on Learning Representations, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[28] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16717" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357, Virtual Event, 2021. PMLR. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.469, + 0.175 + ], + "angle": 0, + "content": "[29] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-SNE. Journal of Machine Learning Research, 9 (11), 2008. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.178, + 0.469, + 0.232 + ], + "angle": 0, + "content": "[30] Chunwei Wang, Chao Ma, Ming Zhu, and Xiaokang Yang. Pointaugmenting: Cross-modal augmentation for 3d object detection. In Computer Vision and Pattern Recognition, pages 11794-11803, Virtual Event, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.469, + 0.3 + ], + "angle": 0, + "content": "[31] Kai Wang, Bo Zhao, Xiangyu Peng, Zheng Zhu, Shuo Yang, Shuo Wang, Guan Huang, Hakan Bilen, Xinchao Wang, and Yang You. Cafe: Learning to condense dataset by aligning features. In Computer Vision and Pattern Recognition, pages 12196-12205, New Orleans, LA, USA, 2022. IEEE. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.303, + 0.469, + 0.343 + ], + "angle": 0, + "content": "[32] Tongzhou Wang, Jun-Yan Zhu, Antonio Torralba, and Alexei A Efros. Dataset distillation. arXiv preprint arXiv:1811.10959, 2018. 1, 2, 8, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.345, + 0.469, + 0.413 + ], + "angle": 0, + "content": "[33] Hongxu Yin, Pavlo Molchanov, Jose M Alvarez, Zhizhong Li, Arun Mallya, Derek Hoiem, Niraj K Jha, and Jan Kautz. Dreaming to distill: Data-free knowledge transfer via deep-inversion. In Computer Vision and Pattern Recognition, pages 8715-8724, Virtual Event, 2020. IEEE. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.415, + 0.469, + 0.468 + ], + "angle": 0, + "content": "[34] Zeyuan Yin, Eric P. Xing, and Zhiqiang Shen. Squeeze, recover and relabel: Dataset condensation at imagenet scale from A new perspective. In Neural Information Processing Systems. NIPS, 2023. 2, 3, 4, 5, 6, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.471, + 0.469, + 0.523 + ], + "angle": 0, + "content": "[35] Shan You, Chang Xu, Chao Xu, and Dacheng Tao. Learning from multiple teacher networks. In International Conference on Knowledge Discovery and Data Mining, pages 1285-1294, Halifax, NS, Canada, 2017. ACM. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.526, + 0.469, + 0.565 + ], + "angle": 0, + "content": "[36] Ruonan Yu, Songhua Liu, and Xinchao Wang. Dataset distillation: A comprehensive review. arXiv preprint arXiv:2301.07014, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.568, + 0.469, + 0.607 + ], + "angle": 0, + "content": "[37] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In British Machine Vision Conference, York, UK, 2016. BMVA Press. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.61, + 0.469, + 0.676 + ], + "angle": 0, + "content": "[38] Lei Zhang, Jie Zhang, Bowen Lei, Subhabrata Mukherjee, Xiang Pan, Bo Zhao, Caiwen Ding, Yao Li, and Dongkuan Xu. Accelerating dataset distillation via model augmentation. In Computer Vision and Pattern Recognition, Vancouver, BC, Canada, 2023. IEEE. 6, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.469, + 0.734 + ], + "angle": 0, + "content": "[39] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In Computer Vision and Pattern Recognition, pages 6848-6856, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.735, + 0.469, + 0.788 + ], + "angle": 0, + "content": "[40] Bo Zhao and Hakan Bilen. Dataset condensation with differentiable siamese augmentation. In International Conference on Machine Learning, pages 12674-12685, Virtual Event, 2021. PMLR. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[41] Bo Zhao and Hakan Bilen. Dataset condensation with distribution matching. In Winter Conference on Applications of Computer Vision, pages 6514-6523, Waikoloa, Hawaii, 2023. IEEE. 2, 7, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[42] Bo Zhao, Konda Reddy Mopuri, and Hakan Bilen. Dataset condensation with gradient matching. In International Conference on Learning Representations, Virtual Event, 2021. OpenReview.net. 2, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[43] Borui Zhao, Quan Cui, Renjie Song, Yiyu Qiu, and Jiajun Liang. Decoupled knowledge distillation. In Computer Vision and Pattern Recognition, pages 11953-11962, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[44] Yongchao Zhou, Ehsan Nezhadarya, and Jimmy Ba. Dataset distillation using neural feature regression. In Neural Information Processing Systems, New Orleans, LA, USA, 2022. NIPS. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16718" + } + ] +] \ No newline at end of file diff --git a/2024/Generalized Large-Scale Data Condensation via Various Backbone and Statistical Matching/1c0e887f-135a-4bea-8bad-3d161e8c65c8_origin.pdf b/2024/Generalized Large-Scale Data Condensation via Various Backbone and Statistical Matching/1c0e887f-135a-4bea-8bad-3d161e8c65c8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2d096a3a63058c59a3dba7e1273d140168e5b55b --- /dev/null +++ b/2024/Generalized Large-Scale Data Condensation via Various Backbone and Statistical Matching/1c0e887f-135a-4bea-8bad-3d161e8c65c8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:69f60fda75265e86754fc6c51a22ca9ba74de08ab166606883c3586664c2172e +size 2251032 diff --git a/2024/Generalized Large-Scale Data Condensation via Various Backbone and Statistical Matching/full.md b/2024/Generalized Large-Scale Data Condensation via Various Backbone and Statistical Matching/full.md new file mode 100644 index 0000000000000000000000000000000000000000..869cb8f70235e2ed8c6f0602c87f5d7b8949992f --- /dev/null +++ b/2024/Generalized Large-Scale Data Condensation via Various Backbone and Statistical Matching/full.md @@ -0,0 +1,293 @@ +# Generalized Large-Scale Data Condensation via Various Backbone and Statistical Matching + +Shitong Shao $^{1,3}$ Zeyuan Yin $^{1}$ Muxin Zhou $^{1}$ Xindong Zhang $^{2,3}$ Zhiqiang Shen $^{1,*}$ $^{1}$ Mohamed bin Zayed University of AI $^{2}$ Hong Kong Polytechnic University $^{3}$ OPPO Research 1090784053sst@gmail.com, {zeyuan.yin,muxin.zhou,zhiqiang.shen}@mbzuai.ac.ae 17901410r@connect.polyu.hk, *:Corresponding author + +![](images/2a9b66f7538f9e2b546c67ba66bfc1ba57e38bbd4635080df03df5a6f7d951d4.jpg) +Figure 1. Left: Our proposed G-VBSM consists of three novel and effective modules, named DD, GSM and GBM. The richness and quality of information in the synthetic data have been significantly enhanced compared with the baseline SRe2L through the sequential merging of DD, GBM, and GSM. Right: G-VBSM prioritizes "generalized matching" to ensure consistency between distilled and complete datasets across various backbones, layers, and statistics, and achieves the highest accuracy $31.4\%$ on ImageNet-1k under IPC 10. + +![](images/0ae6f95b9f1d1115b8711b7d4e83f15a64733b534be8adb54a71dbd53f5f3d61.jpg) + +![](images/e9dc599c01779e3a0f1f98890a49d8ac3e9132403f402adeabdefcf87fc374b8.jpg) + +![](images/d73b32c7cfbae21d6376f5aecd8c891595761ee508a054a3a6fdf18df025c103.jpg) + +# Abstract + +The lightweight "local-match-global" matching introduced by SRe2L successfully creates a distilled dataset with comprehensive information on the full $224 \times 224$ ImageNet-1k. However, this one-sided approach is limited to a particular backbone, layer, and statistics, which limits the improvement of the generalization of a distilled dataset. We suggest that sufficient and various "local-match-global" matching are more precise and effective than a single one and have the ability to create a distilled dataset with richer information and better generalization ability. We call this perspective "generalized matching" and propose Generalized Various Backbone and Statistical Matching (G-VBSM) in this work, which aims to create a synthetic dataset with densities, ensuring consistency with the complete dataset across various backbones, layers, and statistics. As experimentally demonstrated, G-VBSM is the first algorithm to obtain strong performance across both + +small-scale and large-scale datasets. Specifically, G-VBSM achieves performances of $38.7\%$ on CIFAR-100, $47.6\%$ on Tiny-ImageNet, and $31.4\%$ on the full $224 \times 224$ ImageNet-1k, respectively1. These results surpass all SOTA methods by margins of $3.9\%$ , $6.5\%$ , and $10.1\%$ , respectively. + +# 1. Introduction + +With the development of deep learning, the number of model parameters and the quantity of training data have become increasingly large [4, 37]. Researchers have tried to minimize the training overhead while preventing a decline in the generalization ability. Data condensation (DC), also known as Dataset distillation, first introduced by Wang et al. [32], aims to alleviate the training burden by synthesizing a small yet informative distilled dataset derived from the complete training dataset, while ensuring that the behav + +ior of the distilled dataset on the target task remains consistent with that of the complete dataset. The extremely compressed distilled dataset contains sufficiently valuable information and have the potential for fast model training, and have been become a popular choice for different downstream application, like federated learning [7, 22], continual learning [14, 22, 40], neural architecture search [24, 41, 42] and 3D point clouds [30]. + +A persistent problem that researchers have been working to solve [2, 9, 12, 42] in DC is the substantial data synthesis overhead [19, 36], which hinders its application in real-world large-scale datasets (e.g., ImageNet-1k) training. Typical performance matching [15, 32, 44] and trajectory matching [1, 3] unroll recursive computation graphs, requiring substantial GPU memory and resulting in prohibitive training costs. Zhao et al. [42] proposed gradient matching to address this, synthesizing distilled datasets by matching gradients from synthetic and real data in a single step. However, gradient computation and matching remain time-consuming [9], leading to the proposal of distribution matching [41]. This method and its variants [31] employ a network-based feature extractor to embed both synthetic and real data into a high-dimensional Hilbert space, then perform distribution matching. The training load for this direct, single-step process stems only from one gradient update of the synthetic data and the feature extractor [36]. Unfortunately, all of the above mentioned improved methods still have extremely large training overheads on the full $224 \times 224$ ImageNet-1k. + +Recently, SRe2L [34] accomplished data condensation for the first time on the full $224 \times 224$ ImageNet-1k [18], achieving Top-1 validation accuracy $21.3\%$ with ResNet18 under IPC 10. This method outperformed the latest state-of-the-art TESLA [2], which conducted on a low-resolution version of ImageNet-1k, by being $16 \times$ faster and improved performance by a margin of $13.6\%$ . SRe2L is inspired by DeepInversion [33] and aims to match statistics in BatchNorm generated from synthetic and real data. We reevaluate the success of SRe2L through the lightweight "local-match-global" essentially. The "local-match-global" refers to utilizing more comprehensive information (e.g., statistics in BatchNorm), generated from the model using the complete dataset (global), to guide the parameter update of the distilled dataset (local). + +However, such lightweight and effective matching in SRe2L is singular, depending on the particular layer (i.e., BatchNorm), model (i.e., ResNet18), and statistics (i.e., channel mean/variance). Intuitively, sufficient "local-match-global" matching can result in more accurate and rational supervision than a single one, further enhancing the generalization of the distilled dataset. We call this perspective "generalized matching" and suggest that the distilled dataset is likely to perform consistent with the complete + +dataset on the evaluation model, provided that there is sufficient variety in backbones, layers, and statistics used for matching. + +Inspired by this, we propose Generalized Various Backbone and Statistical Matching (G-VBSM) to fulfill "generalized matching". G-VBSM comprises three integral and effective parts named data densification (DD), generalized statistical matching (GSM), and generalized backbone matching (GBM). DD is utilized to ensure that the images within each class are linearly independent, thereby enhancing the (intra-class) diversity of the distilled dataset. This ultimately guarantees that "generalized matching" preserves the rich and diverse information within the synthetic data. GBM and GSM are designed to implement "generalized matching". Among them, GSM traverses the complete dataset without computing and updating the gradient, to record the statistics of Convolution at both the patch and channel levels. These statistics are subsequently utilized for matching during the data synthesis phase, in conjunction with the channel-level statistics in BatchNorm. Furthermore, GBM aims to ensure consistency between distilled and complete datasets across various backbones, enhancing matching sufficiency and leading to strong generalization in the evaluation phase. In particular, G-VBSM also ensures the efficiency of dataset condensation through a series of strategies, as mentioned in Sec. 3. + +Extensive experiments on CIFAR-10, CIFAR-100, TinyImageNet, and the full $224 \times 224$ ImageNet-1k, demonstrating that G-VBSM is the first algorithm that performs well on both small-scale and large-scale datasets. Specifically, we not only verify through ablation studies that GSM, GBM and DD are consistently reliable, but also achieve the highest $38.7\%$ , $47.6\%$ and $31.4\%$ on CIFAR-100 (128-width ConvNet), Tiny-ImageNet (ResNet18), and the full $224 \times 224$ ImageNet-1k (ResNet18) under images per class (IPC) 10, 50 and 10, respectively, which outperforms all previous state-of-the-art (SOTA) methods by $3.9\%$ , $6.5\%$ and $10.1\%$ , respectively. + +# 2. Background + +Dataset condensation (DC) represents a data synthesis procedure that aims to compress a complete, large dataset $\mathcal{T} = \{(X_i,y_i)\}_{i = 1}^{|T|}$ into a smaller, distilled dataset $S = \{(\tilde{X}_i,\tilde{y}_i)\}_{i = 1}^{|S|}$ , subject to $|\mathcal{S}|\ll |\mathcal{T}|$ , while ensuring that an arbitrary evaluation model $f_{\mathrm{eval}}(\cdot)$ trained on $S$ yields similar performance to one trained on $\mathcal{T}$ . Classical data distillation algorithms invariably require the candidate model $f_{\mathrm{cand}}(\cdot)$ to execute one or more steps on $S$ to update its parameter $\theta_{\mathrm{cand}}$ , subsequently achieving matching in terms of performance [15, 32], gradient [42], trajectory [1, 2], or distribution [31, 41]. The process $\theta_{\mathrm{cand}} - \alpha \nabla_{\theta_{\mathrm{cand}}}\ell (f_{\mathrm{cand}}(\tilde{X}),\tilde{y})$ where $\ell (\cdot ,\cdot)$ and $(\tilde{X},\tilde{y})$ represent the loss function and + +![](images/c1a58103c5aaca085ba63e123665764fe60b78a9ec1160efafb18d383dc7e70a.jpg) +Figure 2. The overview of G-VBSM on the full $224 \times 224$ ImageNet-1k, which ensures the consistency between the distilled and the complete datasets across various backbones, layers and statistics via "generalized matching". + +a batch sampled from $S$ , respectively, is notably time-consuming. Consequently, even the relatively swifest distribution matching [20, 31] is slower than the recent proposed SRe2L [34]. In fact, SRe2L is the only workable way to achieve DC on the full $224 \times 224$ ImageNet-1k, as it requires updating the parameters of the synthetic data only once per iteration. + +SRe2L [34] encompasses three incremental subprocesses: Squeeze, Recover, and Relabel. Squeeze is designed to train $f_{\mathrm{cand}}(\cdot)$ containing BatchNorm in a standard manner, aiming to record the global channel mean $\mathbf{BN}_l^{\mathrm{CM}}$ and channel variance $\mathbf{BN}_l^{\mathrm{CV}}$ ( $l$ refers to the index of the $l$ -th layer) via exponential moving average (EMA), extracted from $\mathcal{T}$ , for subsequent matching in Recover. In Recover after that, given the channel mean $\mu_l(\tilde{X})$ and channel variance $\sigma_l^2(\tilde{X})$ in $l$ -th BatchNorm obtained from $S$ , the statistical matching loss function can be formulated as + +$$ +\mathcal {L} _ {\mathrm {B N}} (\tilde {X}) = \sum_ {l} \left| \left| \mu_ {l} (\tilde {X}) - \mathbf {B} \mathbf {N} _ {l} ^ {\mathrm {C M}} \right| \right| _ {2} + \left| \left| \sigma_ {l} ^ {2} (\tilde {X}) - \mathbf {B} \mathbf {N} _ {l} ^ {\mathrm {C V}} \right| \right| _ {2}. \tag {1} +$$ + +Based on this, we can give the entire optimization objective in Recover as + +$$ +\underset {\tilde {X}} {\arg \min } \mathcal {L} _ {\mathrm {B N}} (\tilde {X}) + \ell \left(f _ {\text {c a n d}} (\tilde {X}), y\right), \tag {2} +$$ + +where $y$ denotes the ground truth label. Moreover, SRe2L assigns soft labels $\tilde{y}$ to the synthetic data, utilizing the logit-based distillation [6, 8] to improve the generalization ability of the distilled dataset. This can be denoted as + +$$ +\tilde {y} = \operatorname {s o f t m a x} \left(f _ {\text {c a n d}} (\tilde {X}) / \tau\right), \tag {3} +$$ + +where $\tau$ denotes the temperature to regulate the smoothness of the soft labels, thereby enhancing the distilled dataset's potential for generalization to unseen evaluation + +models. The generated soft label can be stored on disk using FKD [23] so as not to defeat the purpose of DC. A crucial point of SRe2L is that BatchNorm calculates the statistics of the entire dataset using EMA, thereby providing a comprehensive and representative matching for the distilled dataset. Encouraged by this, our research focuses on applying sufficient 'local-match-global' matching to achieve "generalized matching". + +# 3. Method + +The comprehensive and detailed framework of our proposed Generalized Various Backbone and Statistical Matching (G-VBSM) is illustrated in Fig. 2. In essence, G-VBSM employs the lightweight regularization strategy data densification (DD) to ensure both the diversity and density of the distilled dataset, ensuring that the potential of "generalized matching" can be fully exploited. Moreover, generalized backbone matching (GBM) and generalized statistical matching (GSM) are utilized to achieve "generalized matching" by performing "local-match-global" matching across various backbones, layers, and statistics. In particular, the efficiency and effectiveness of DD, GBM, GSM, SRe2L, and TESLA are illustrated in Fig. 5. + +# 3.1. Data Densification + +As illustrated in Fig. 3, the synthetic data generated by SRe2L exhibit excessive similarity within the same class, leading to a lack of diversity. Consequently, even if "generalized matching" preserves sufficient valuable information within a single image, the aggregate information content across all images within the same class does not increase effectively, which ultimately prevents "generalized match- + +![](images/8f58880b3cd8c4670f8e8cb6424c374c074981b2fb81338a8af8fca87549b6f7.jpg) +Figure 3. Visualization of the mean cosine similarity between pairwise samples within the same class on ImageNet-1k under IPC 10. + +ing" from being sufficiently advantageous. Data densification (DD) is proposed to address this by ensuring the data $\tilde{X}$ has full rank in the batch dimension, thereby guaranteeing that samples in each class are linearly independent, and ultimately ensuring that the data is diverse and abundant to fully exploit the potential of "generalized matching". + +To execute this pipeline, $\tilde{X}$ first needs to be downsampled to reduce the computational cost of eigenvalue decomposition: + +$$ +\hat {X} = \operatorname {A v g P o o l 2 d} (\tilde {X}), \text {s . t .} \hat {X} \in \mathbb {R} ^ {B \times C \times 3 2 \times 3 2}, \tag {4} +$$ + +where $B$ and $C$ represent the batch size and the number of channels, respectively. Afterward, we reshape $\hat{X}$ from $B\times C\times 32\times 32$ to $B\times (1024C)$ and perform matrix multiplication in each class $y$ to obtain the set of the Gram matrix $\{\hat{X}_y\hat{X}_y^T\}_{y\in \mathcal{Y}}$ , where $\mathcal{V}$ refers to a set of all classes existing in this batch. And $\hat{X}_y$ is a subbatch with class $y$ . Note that the alternative form $\{\hat{X}_y^T\hat{X}_y\}_{y\in \mathcal{Y}}$ is not desirable, as it is applicable only for dimensionality reduction in feature dimensions, which is why we do not consider singular value decomposition (SVD). To render $\hat{X}_y\hat{X}_y^T$ as full-rank as possible, we introduce the data densification loss in Eq. 5. + +$$ +\mathcal {L} _ {\mathrm {D D}} (\tilde {X}) = \sum_ {y \in \mathcal {Y}} D _ {\mathrm {K L}} (\operatorname {s t o p} - \operatorname {g r a d} (p \left(\Sigma_ {y} / \tau\right)) | | p \left(\Sigma_ {y}\right)), \tag {5} +$$ + +where $\Sigma_y$ , $\tau$ , stop\_grad(\cdot) and $p(\cdot)$ refer to the eigenvalues of $\hat{X}_y\hat{X}_y^T$ , the temperature, the stop gradient operator and the softmax function, respectively. And $D_{\mathrm{KL}}(\cdot ||\cdot)$ denotes Kullback-Leibler divergence. As demonstrated in Sec. E, the diversity of the data is significantly enhanced by the employment of Eq. 5. In our experiment, $\tau$ is set as 4 in default and we do not assign a deliberate weight (set to 1 by default) to $\mathcal{L}_{\mathrm{DD}}$ because $\mathcal{L}_{\mathrm{DD}} \equiv 0$ at the early $10\%$ of the iterations. In other words, DD is quite stable and the optimization objective $\arg \min_{\{\Sigma_y\}_{y \in \mathcal{Y}}} \mathcal{L}_{\mathrm{DD}}$ is equivalent to $\arg \min_{\{\Sigma_y\}_{y \in \mathcal{Y}}} \sum_y \sigma^2(\Sigma_y)$ . + +Technical Detail. A problem that warrants attention is that in the SRe2L's implementation [34], having merely a single sample in each class of a batch indicates insufficient + +![](images/53d0feb31231732b0ef278c1c9463a5864135201a3f84222449c2da59b8c46a7.jpg) +Figure 4. The illustration of the original loop and the reorder loop. + +to execute DD under the order of the original loop, as depicted in Fig. 4 (Left). A simple solution is to translate the original loop to the reorder loop, as shown in Fig. 4 (Right). However, our experiment on ResNet50 (i.e., the evaluation model) substantiates that this solution suffers a $2.6\%$ accuracy degradation (details can be found in Sec. 4.1) on ImageNet-1k under IPC 10. The reason is that the number of classes in each iteration within the reorder loop is insufficient, preventing a batch have the ability to match the global statistics in BatchNorm (i.e., $\mathbf{BN}_l^{\mathrm{CM}}$ and $\mathbf{BN}_l^{\mathrm{CV}}$ ). Motivated by score distillation sampling (SDS) [17], we update the statistics during data synthesis via EMA to solve this issue, so that the statistics of all past batches can assist the statistics of the current batch match $\mathbf{BN}_l^{\mathrm{CM}}$ and $\mathbf{BN}_l^{\mathrm{CV}}$ : + +$$ +\begin{array}{l} \mu_ {l} ^ {\text {t o t a l}} = \alpha \mu_ {l} ^ {\text {t o t a l}} + (1 - \alpha) \mu_ {l} (\tilde {X}), \sigma_ {l} ^ {2, \text {t o t a l}} = \alpha \sigma_ {l} ^ {2, \text {t o t a l}} + (1 - \alpha) \sigma_ {l} ^ {2} (\tilde {X}), \\ \mathcal {L} _ {\mathrm {B N}} ^ {\prime} (\tilde {X}) = \sum_ {l} | | \mu_ {l} (\tilde {X}) - \mathbf {B N} _ {l} ^ {\mathrm {C M}} - \operatorname {s t o p \_ g r a d} (\mu_ {l} (\tilde {X}) - \mu_ {l} ^ {\text {t o t a l}}) | | _ {2} \\ + \left\| \sigma_ {l} ^ {2} (\tilde {X}) - \mathbf {B N} _ {l} ^ {\mathrm {C V}} - \operatorname {s t o p \_ g r a d} \left(\sigma_ {l} ^ {2} (\tilde {X}) - \sigma_ {l} ^ {2, \text {t o t a l}}\right) \right\| _ {2}. \tag {6} \\ \end{array} +$$ + +The derivation of $\mathcal{L}_{\mathrm{BN}}^{\prime}(X)$ can be found in Appendix B. We call this lightweight strategy as "match in the form of score distillation sampling" and have demonstrated its effectiveness and feasibility in our ablation studies. + +# 3.2. Generalized Backbone Matching + +Performing data synthesis only on a single pre-trained model is not able to enjoy ensemble gains from various backbones. Meanwhile, classical DC algorithms such as MTT [1] and FTD [3] obtain performance improvements from multiple randomly initialized backbones. Therefore, introducing generalized backbone matching (GBM) to apply various backbones for data synthesis is a desirable choice for "generalized matching". It ensures a number of "local-match-global" matching nearly $N_{\mathrm{b}} \times$ higher compared to just depending on a single backbone, where $N_{\mathrm{b}}$ denotes the number of backbones. Regrettably, unrolling various backbone computational graphs in parallel for data synthesis is extremely expensive at the computational cost level. A solution is to randomly sample a backbone from the candidates in per iteration. This simple yet effective strategy not only ensures computational efficiency but also allows the data synthesis process to benefit from the diverse and multifaceted matching that the various backbones pro + +vide. We denote this pipeline as + +$$ +f _ {\text {c a n d}} (\cdot) \sim \mathcal {U} (\mathbb {S}), \mathbb {S} = \{\text {R e s N e t 1 8}, \dots , \text {S h u f f l e N e t V 2 - 0 . 5} \}. \tag {7} +$$ + +To maintain backbone consistency in both the data synthesis and soft label generation phases, we introduce a backbone ensemble during soft label generation: + +$$ +\tilde {z} = \left\{ \begin{array}{c c} \sum_ {f _ {\text {c a n d}} \in \mathbb {S}} \frac {\sum_ {g \in \mathbb {S}} | | g (\tilde {X}) | | _ {\mathrm {F}}}{| \mathbb {S} | | | f _ {\text {c a n d}} (\tilde {X}) | | _ {\mathrm {F}}} \frac {f _ {\text {c a n d}} (\tilde {X})}{| \mathbb {S} |}, & \mathrm {w / L N}, \\ \sum_ {f _ {\text {c a n d}} \in \mathbb {S}} \frac {f _ {\text {c a n d}} (\tilde {X})}{| \mathbb {S} |}, & \mathrm {w / o L N}, \end{array} \right. \tag {8} +$$ + +where $||\cdot||_{\mathrm{F}}$ and LN refer to the Frobenius norm and Logit Normalization, respectively. Thus, we can obtain the soft label by $\tilde{y} = \text{softmax}(\tilde{z}/\tau)$ . Particularly, the use of LN is optional, as demonstrated by our ablation studies; it is beneficial for ResNet18 but not for ResNet50 or ResNet101. Moreover, it's important to highlight that we apply a parallel mechanism for soft label generation since it provides a significantly lower computational cost, less than 1/30 of that required for data synthesis, thus making the computational overhead negligible. + +# 3.3. Generalized Statistical Matching + +Only ensuring backbone diversity is insufficient to fully exploit the potential of "generalized matching". In this subsection, we aim to introduce additional statistics for matching during the data synthesis phase. Prior methods [33, 34] only utilize BatchNorm since the presence of global information statistics within the pre-trained model, with no apparent solution ready for other layers (e.g., Convolution). Retraining the model is the simplest way to address this problem, followed by updating the other layers' parameters through EMA. This approach is obviously impractical, as it defeats the purpose of using gradient descent. By contrast, we propose to allow the pre-trained model $f_{\mathrm{eval}}(\cdot)$ to loop through the training dataset $\mathcal{T}$ once without calculating the gradient to obtain the global statistics of Convolution, thus serving "local-match-global": + +$$ +\mathbf {C o n v} _ {l} ^ {\mathbf {C M}} = \frac {1}{| \mathcal {T} |} \sum_ {i} ^ {| \mathcal {T} |} \mathbf {C M} _ {i} ^ {l}, \mathbf {C o n v} _ {l} ^ {\mathbf {C V}} = \frac {1}{| \mathcal {T} |} \sum_ {i} ^ {| \mathcal {T} |} \mathbf {C V} _ {i} ^ {l}, \tag {9} +$$ + +$$ +\mathbf {C o n v} _ {l} ^ {\mathbf {P M}} = \frac {1}{| \mathcal {T} |} \sum_ {i} ^ {| \mathcal {T} |} \mathbf {P M} _ {i} ^ {l}, \mathbf {C o n v} _ {l} ^ {\mathbf {P V}} = \frac {1}{| \mathcal {T} |} \sum_ {i} ^ {| \mathcal {T} |} \mathbf {P V} _ {i} ^ {l}. +$$ + +Here, $\mathbf{CM}_i^l\in \mathbb{R}^{C_l}$ $\mathbf{CV}_i^l\in \mathbb{R}^{C_l}$ $\mathbf{PM}_i^l\in \mathbb{R}^{\left[\frac{H}{N_l^p}\right]\times \left[\frac{W}{N_l^p}\right]}$ and $\mathbf{PV}_i^l\in \mathbb{R}^{\left[\frac{H}{N_l^p}\right]\times \left[\frac{W}{N_l^p}\right]}$ refer to the channel mean, the channel variance, the patch mean and the patch variance, respectively, for the $l$ -th Convolution when the input to $f_{\mathrm{cand}}(\cdot)$ is the $i$ -th batch, where $C_l$ and $\left[\frac{H}{N_l^p}\right]\times \left[\frac{W}{N_l^p}\right]$ denote the number of channels and patches of the $l$ -th Convolution, respectively. We define $N_{l}^{p}$ as 4, 4, 4 and 16 by default + +on CIFAR-10, CIFAR-100, Tiny-ImageNet and ImageNet-1k, respectively. After obtaining the global channel mean $\mathbf{Conv}_l^{\mathbf{CM}}$ , the global channel variance $\mathbf{Conv}_l^{\mathbf{CV}}$ , the global patch mean $\mathbf{Conv}_l^{\mathbf{PM}}$ and the global patch variance $\mathbf{Conv}_l^{\mathbf{PV}}$ , we can store them in a disk thus avoiding secondary calculations. In the data synthesis phase, we introduce $\mathcal{L}_{\mathrm{Conv}}'(\tilde{X})$ in Eq. 10 to accomplish joint matching with $\mathcal{L}_{\mathrm{BN}}'(\tilde{X})$ . + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {C o n v}} ^ {\prime} (\tilde {X}) = \sum_ {l} | | \mu_ {l} ^ {c} (\tilde {X}) - \mathbf {C o n v} _ {l} ^ {\mathrm {C M}} - \operatorname {s t o p - g r a d} \left(\mu_ {l} ^ {c} (\tilde {X}) - \mu_ {l} ^ {c, \text {t o t a l}}\right) | | _ {2} \\ + \left\| \sigma_ {l} ^ {c, 2} (\tilde {X}) - \mathbf {C o n v} _ {l} ^ {\mathrm {C V}} - \operatorname {s t o p - g r a d} \left(\sigma_ {l} ^ {c, 2} (\tilde {X}) - \sigma_ {l} ^ {c, 2, \text {t o t a l}}\right) \right\| _ {2}, \\ + \left\| \mu_ {l} ^ {p} (\tilde {X}) - \mathbf {C o n v} _ {l} ^ {\mathrm {P M}} - \operatorname {s t o p - g r a d} \left(\mu_ {l} ^ {p} (\tilde {X}) - \mu_ {l} ^ {p, \text {t o t a l}}\right) \right\| _ {2} \\ + \left\| \sigma_ {l} ^ {p, 2} (\tilde {X}) - \mathbf {C o n v} _ {l} ^ {\mathrm {P V}} - \operatorname {s t o p} - \operatorname {g r a d} \left(\sigma_ {l} ^ {p, 2} (\tilde {X}) - \sigma_ {l} ^ {p, 2, \text {t o t a l}}\right) \right\| _ {2}, \tag {10} \\ \end{array} +$$ + +where $\mu_l^{c,\mathrm{total}}$ , $\sigma_l^{c,2,\mathrm{total}}$ , $\mu_l^{p,\mathrm{total}}$ , and $\sigma_l^{p,2,\mathrm{total}}$ are each updated from the channel mean $\mu_l^c (\tilde{X})$ , channel variance $\sigma_l^{c,2}(\tilde{X})$ , patch mean $\mu_l^p (\tilde{X})$ , and patch variance $\sigma_l^{p,2}(\tilde{X})$ respectively, all obtained via EMA from the current batch. In experiments, we discovered that Eq. 10 causes a slightly larger computational burden, so we randomly drop the matching of statistics with a probability of $\beta_{\mathrm{dr}}$ to ensure the efficiency of GSM. + +Loss Function in the Evaluation Phase. Unlike DD, GBM and GSM are designed to create a distilled dataset that is enriched with information. Here, we introduce an enhancement to the loss function tailored specifically for the evaluation phase. Essentially, the evaluation phase is a knowledge distillation framework for transferring knowledge from a pre-trained model to the evaluation model. SRe2L utilizes $D_{\mathrm{KL}}(\tilde{y} || \mathrm{softmax}(f_{\mathrm{eval}}(\tilde{X}) / \tau))$ as the loss function and experimentally illustrates that it improves performance by roughly $10\%$ . As established in SRe2L, an increase in temperature $\tau$ correlates with enhanced performance of the evaluation model. Inspired by this and $\tau^2 D_{\mathrm{KL}}(\mathrm{softmax}(p) / \tau ||\mathrm{softmax}(q) / \tau)$ is equivalent to $\frac{1}{2C} ||p - q||_2^2$ when $\tau \rightarrow +\infty$ [10], we introduce a novel loss function $\mathrm{MSE} + \gamma \times \mathrm{GT}$ to avoid numerical error caused by the large $\tau$ and improve the generalization of the distilled dataset without any additional overhead (ignore the weights $\tau^2$ and $\frac{1}{2C}$ ): + +$$ +\mathcal {L} _ {\text {e v a l}} (\tilde {X}, \tilde {y}, y) = | | f _ {\text {e v a l}} (\tilde {X}) - \tilde {z} | | _ {2} ^ {2} - \gamma \mathbf {y} \log \left(\operatorname {s o f t m a x} \left(f _ {\text {e v a l}} (\tilde {X})\right)\right), \tag {11} +$$ + +where $y$ represents the one-hot encoding (w.r.t. the ground truth label $y$ ). As illustrated in Fig. 5, simply replacing the loss function with $\mathrm{MSE} + 0.1 \times \mathrm{GT}$ ( $\gamma$ is set as 0.1) in SRe2L improves the performance of ResNet18 (i.e., the evaluation model) by a margin of $0.9\%$ on ImageNet-1k under IPC 10. + +# 4. Experiment + +We perform comparison experiments on the large-scale dataset including the full $224 \times 224$ ImageNet-1k [18] and the small-scale datasets including Tiny-ImageNet [26], + +![](images/494301707fff98f4681695e00c6652ea30dde251ec38a3c20ef2b4bafcc8422c.jpg) +Figure 5. Comparison of the effectiveness and efficiency of G-VBSM components. Among them, "DD+GBM+GSM $(\beta_{\mathrm{dr}} = 0.0) + (\mathrm{MSE} + \mathrm{GT})$ represents the comprehensive G-VBSM. + +CIFAR-10/100 [11]. To highlight that our proposed G-VBSM is designed for large-scale datasets, all ablation experiments are performed on ImageNet-1k. + +Hyperparameter Settings. We prioritize selecting various convolution-based backbone, ensuring the maximal difference in architecture, while also adhering to the criterion of minimizing the number of parameters, an approach empirically demonstrated to be superior [34, 38]. On ImageNet-1k, we skip the model pre-training phase by directly using Torchvision's open source pretraining weights [16] of {ResNet18 [5], MobileNetV2 [21], EfficientNet-B0 [25], ShuffleNetV2-0.5 [39]}. For the remaining dataset, we all train {128-width ConvNet [39], WRN-16-2 [37], ResNet18 [5], ShuffleNetV2-0.5 [39], MobileNetV2-0.5 [21]} from scratch with few epoch in the model pre-training phases. Gray cells in all tables represent the highest performance. Meanwhile, (R18), (R50), etc. in all tables, represent the evaluation models. More details about the remaining method hyperparameter settings and additional ablation studies can be found in Appendix A and Appendix C & D, respectively. + +# 4.1. Ablation Studies + +The Trade-off between Efficiency and Effectiveness. The performance improvements and increased GPU latency for components of G-VBSM in comparison to the baseline SRe2L, are illustrated in Fig. 5. Clearly, DD, GBM, and GSM are highly effective in enhancing the generalization of distilled datasets. Compared with GSM, both DD and GBM are slightly lightweight and efficient. With DD and GBM alone - i.e., $\mathrm{DD + GBM + GSM}$ $(\beta_{\mathrm{dr}} = 1.0) + (\mathrm{MSE} + \mathrm{GT})$ - the accuracy of this approach surpasses that of SRe2L by a significant margin of $6.5\%$ . Additionally, the comprehensive G-VBSM, $\mathrm{DD + GBM + GSM}$ $(\beta_{\mathrm{dr}} = 0.0) + (\mathrm{MSE} + \mathrm{GT})$ , further enhances the performance of $\mathrm{DD + GBM + GSM}$ $(\beta_{\mathrm{dr}} = 1.0) + (\mathrm{MSE} + \mathrm{GT})$ by a notable margin of $3.6\%$ . Note that $\mathrm{MSE} + \gamma \times \mathrm{GT}$ ( $\gamma$ is set as 0.1) is also extremely critical, for SRe2L and $\mathrm{DD + GBM}$ boosting $0.9\%$ and $2.7\%$ , respectively. + +
α\Evaluation ModelResNet18(MSE+0.1×GT)ResNet50(MSE+0.1×GT)
0.030.4%31.9%
0.431.4%34.5%
0.831.9%36.4%
+ +Matching in the Form of SDS. Benefiting from this matching form, G-VBSM can achieve the best performance with small batch sizes (e.g., 40 in ImageNet-1k and 50 in CIFAR-100) in different settings, as demonstrated in our comparative experiments. Table 1 demonstrates the impact of the factor $\alpha$ for performing the EMA update on the final performance achieved by G-VBSM, where $\alpha = 0$ indicates that $\mathcal{L}_{\mathrm{BN}}^{\prime}$ degenerates to $\mathcal{L}_{\mathrm{BN}}$ . We can conclude that the SDS matching form is feasible and the generalization of the distilled dataset improves with increasing $\alpha$ . + +Table 1. ImageNet-1k Top-1 Acc. about different $\alpha$ under IPC 10. + +
MethodEvaluation Model
DeiT-TinyResNet18MobileNetV2Swin-Tiny
SRe2L15.41%46.79%36.59%39.23%
G-VBSM29.43%51.82%48.66%57.40%
+ +Cross-Architecture Generalization. The evaluation of cross-architecture generalization is crucial in assessing the quality of distilled datasets. Unlike traditional methods [1, 15, 20] which focus on CIFAR-100, our approach evaluates the effectiveness of the distilled dataset on ImageNet1k, employing a suite of models with real-world applicability, including ResNet18 [5], MobileNetV2 [21], DeiTiny [28], and Swin-Tiny [13]. The experimental results are reported in Table 2. From Tables 7 and 2, it is evident that the G-VBSM-synthetic dataset can effectively generalize across ResNet {18, 50, 101}, MobileNetV2, DeiTiny, and Swin-Tiny architectures. Notably, DeiT-Tiny and Swin-Tiny, two architectures not encountered during the data synthesis phase, demonstrates significant proficiency with the accuracy $29.43\%$ and $57.40\%$ , outperforming the latest SOTA SRe2L by a margin of $14.01\%$ and $18.17\%$ , respectively. + +Table 2. ImageNet-1k Top-1 Acc. on cross-architecture generalization under IPC 50. + +
Candidates (Backbone)Evaluation Model
ResNet18MobileNetV2EfficientNet-B0ShuffleNetV2-0.5ResNet18ResNet50
✓25.7%30.2%
✓✓27.2%31.2%
✓✓✓27.9%32.0%
✓✓✓✓31.4%34.5%
+ +Table 3. ImageNet-1k Top-1 Acc. about the number of candidate backbone in the soft label generation pipeline under IPC 10. + +Ensemble in Soft Label Generation. In knowledge distillation, soft labels obtained through multiple teacher ensembles can effectively enhance the generalization of the student [35]. This observation is similarly corroborated in + +![](images/8f339e50e538ee9c08e709dd0e54907fcf5e2e0327ae27d36a7bb7963b97d5c0.jpg) +Figure 6. ImageNet-1k Top-1 Acc. of different loss function, $\beta_{\mathrm{dr}}$ and batch sizes under IPC 10. + +![](images/0aee72a94832890741a2abcdbdaa4367f85762f86a1795342ab35c3a0fd03085.jpg) + +![](images/16d36d0cbaa7310f4460a46b42cdd24c90c5548c1d09d61236e2af8ccf7a2187.jpg) + +
DatasetIPCMTT [1] (CW128)DataDAM [20] (CW128)TESLA [2] (R18)SRe2L (R18)SRe2L (R50)SRe2L (R101)G-VBSM (R18)G-VBSM (R50)G-VBSM (R101)
Tiny-ImageNet5028.0±0.328.7±0.3-41.1±0.442.2±0.542.5±0.247.6±0.348.7±0.248.8±0.4
100---49.7±0.351.2±0.451.5±0.351.0±0.452.1±0.352.3±0.1
ImageNet-1k1064.0±1.3†6.3±0.07.7±0.121.3±0.628.4±0.130.9±0.131.4±0.535.4±0.838.2±0.4
50-15.5±0.2-46.8±0.255.6±0.360.8±0.551.8±0.458.7±0.361.0±0.4
100---52.8±0.361.0±0.462.8±0.255.7±0.462.2±0.363.7±0.2
+ +Table 4. Comparison with baseline models in Tiny-ImageNet and ImageNet-1k. ${}^{ \dagger }$ indicates the ImageNette dataset,which contains only 10 classes. DataDAM [20] and TESLA [2] use the downsampled ${64} \times {64}$ ImageNet-1k. We cite the experimental results from SRe2L [34]. + +
DatasetIPCCoreset SelectionTraining Set Synthesis (CW128)Training Set Synthesis (R18)Whole Dataset
RandomHerdingK-CenterForgettingDC [42]DM [41]CAFE [31]KIP [15]MTT [1]DataDAM[20]G-VBSMSRe2LG-VBSM(CW128)
CIFAR-101026.0±1.231.6±0.714.7±0.923.3±1.044.9±0.548.9±0.650.9±0.546.1±0.765.3±0.754.2±0.846.5±0.727.2±0.553.5±0.684.8±0.1
5043.4±1.040.4±0.627.0±1.423.3±1.153.9±0.563.0±0.462.3±0.453.2±0.771.6±0.267.0±0.454.3±0.347.5±0.659.2±0.4
CIFAR-10014.2±0.38.3±0.38.4±0.34.5±0.212.8±0.311.4±0.314.0±0.312.0±0.224.3±0.314.5±0.516.4±0.72.0±0.225.9±0.556.2±0.3
1014.6±0.517.3±0.317.3±0.315.1±0.325.2±0.329.7±0.331.5±0.240.1±0.433.1±0.434.8±0.538.7±0.231.6±0.559.5±0.4
5030.0±0.433.7±0.530.5±0.3-30.6±0.643.6±0.447.7±0.2-42.9±0.349.4±0.345.7±0.449.5±0.365.0±0.5
+ +dataset distillation, as evidenced in Table 3. The greater the number of models, the stronger the generalization of the distilled dataset. It is particularly interesting to point out that ShuffleNetV2-0.5 improves further by more than $2\%$ even though the other three models have already achieved modest ensemble gains. We attribute this enhancement to the channel shuffle mechanism within ShuffleNetV2-0.5, which imposes a beneficial regularization constraint for G-VBSM. + +Table 5. Comparison with baseline models on CIFAR-10/100. All methods, except for SRe2L and G-VBSM, use a 128-width ConvNet (CW128) for data synthesis and evaluation. G-VBSM utilizes {CW128, WRN-16-2, ResNet18 (R18), ShuffleNetV2-0.5, MobileNetV2-0.5} for data synthesis and {CW128, R18} for evaluation. We cite the experimental results, except for SRe2L's, from DataDAM [20]. + +
Logit NormalizationEvaluation Model
ResNet18 (MSE+0.1×GT)ResNet18 (KL)ResNet50 (MSE+0.1×GT)ResNet50 (KL)
Yes31.4%25.4%34.5%28.6%
No31.0%25.1%35.4%31.9%
+ +Table 6. ImageNet-1k Top-1 Acc. about the use of logit normalization under IPC 10. + +Logit Normalization in Soft Label Generation. The aim of this strategy is to maintain the consistency of the logits (can be viewed as vectors) magnitude across all models within a high-dimensional space, thereby ensuring equal contribution of these models to the ultimate soft label for ensemble. This approach is not universally effective, as shown in Table 6. When the evaluation model parameter count is low (e.g., ResNet18), it can enhance G-VBSM performance. Conversely, with a model like ResNet50, it may hinder performance, rendering G-VBSM uncompetitive. As + +a result, in this work, ResNet $\{50, 101\}$ and ViT-based models do not employ logit normalization, while the remaining models do. + +Loss Function in the Evaluation Phase. As illustrated in Fig. 6, replacing $D_{\mathrm{KL}}(\cdot ||\cdot)$ in SRe2L with $\mathrm{MSE} + \gamma \times \mathrm{GT}$ demonstrates to be extremely effective. For example, with $\beta_{\mathrm{dr}} = 0.0$ and batch size $= 40$ , adjusting $\gamma$ to 0.01 and 0.1 enhances model performance during the evaluation phase by $1.4\%$ and $6.0\%$ , respectively. As a result, we employ a $\gamma$ of 0.1 in all experiments conducted on ImageNet-1k. Furthermore, we observe that the distilled dataset's generalization is not significantly affected ( $\leq 2\%$ ) by variations in either batch size or $\beta_{\mathrm{dr}}$ . This ensured that G-VBSM attain SOTA results with a minimal batch size 40 and $\beta_{\mathrm{dr}} = 0.4$ (i.e., the default settings in ImageNet-1k). + +![](images/1f518e974b524b3100798be7098e0c0c470ae6fb5178780a279feb4d07290cee.jpg) +Figure 7. Synthetic data visualization on small-scale datasets. + +Synthetic Data Visualization. The visualization results for the large-scale (the full $224 \times 224$ ImageNet-1k) and small-scale (CIFAR-10/100 and Tiny-ImageNet) datasets are displayed in Figs. 1 and 7, respectively. On ImageNet-1k, the images synthesized by G-VBSM are more informative and abstract compared with those from SRe2L. Furthermore, on small-scale datasets, the distilled images obtained by G-VBSM markedly differ from the images presented in MTT and DataDAM's papers [1, 20], illustrating that G-VBSM is an out-of-the ordinary algorithm. More synthetic data can be found in Appendix I. + +![](images/d0f0509b50203882be0e13b9935d8f09e5a2cba11aad947d1e4fa7ed3eabde19.jpg) +Figure 8. t-SNE visualization on ImageNet-1k. + +Logit Embedding Distributions Visualization. We feed distilled datasets into a pre-trained model (i.e., ResNet18) to obtain logit embeddings for t-SNE [29] visualization. Note that unlike previous methods that used feature embeddings before the classifier layer, we utilize logit embeddings after the classifier layer for presentation. Fig. 8 shows the distributions of G-VBSM and SRe2L logit embeddings. The logit embedding obtained by G-VBSM has less intra-class similarity than SRe2L, ensuring softer labels, preventing model overconfidence during the evaluation phase, and ultimately enhancing model generalization. + +# 4.2. Large-Scale Dataset Comparison + +Full $224 \times 224$ ImageNet-1k. As illustrated in the ImageNet-1k part of Table 4, G-VBSM consistently outperforms SRe2L across all IPC $\{10, 50, 100\}$ , with the evaluation model being ResNet $\{18, 50, 101\}$ , indicating that G-VBSM is highly competitive in the large-scale dataset ImageNet-1k. In particular, under extreme compression scenarios, such as IPC 10, G-VBSM achieves significant performance increases - $10.1\%$ , $7.0\%$ , and $7.3\%$ for ResNet18, ResNet50, and ResNet101, respectively - compared with SRe2L. In comparison to the latest SOTA classical method, DataDAM, just experimenting on the full $64 \times 64$ ImageNet-1k, we exceed it by a margin of $23.7\%$ under IPC 10. These results are particularly impressive when considered in the context of the large-scale dataset ImageNet-1k. + +# 4.3. Small-Scale Dataset Comparison + +Tiny-ImageNet. Similar to SRe2L, we evaluate the generalization of the distilled dataset on ResNet {18, 50, 101}. As presented in the Tiny-ImageNet part of Table 4, G-VBSM consistently surpasses the latest SOTA method SRe2L on different settings. Our method outperforms SRe2L, MTT and DataDAM by $6.5\%$ , $19.6\%$ and $18.9\%$ , respectively, under IPC 50 applying 128-width ConvNet or ResNet18 as the evaluation model. + +CIFAR-100. One of the disadvantages of SRe2L is its inability to compete with SOTA performance on this small-scale dataset benchmark (i.e., evaluated with 128-width ConvNet and ResNet18), such as the performance on CIFAR-100 shown in Table 5, as a result, it fails to demonstrate that "local-match-global" (e.g., statistical matching of BatchNorm) can rival or outperform traditional DC algorithms. In contrast, as shown in the CIFAR-100 part of Table 5, G-VBSM achieves accuracies $16.4\%$ and $38.7\%$ , the highest under IPC 1 and 10 on CIFAR-100 through "generalized matching", while maintaining the same number of epochs (i.e., 1000) and the model (i.e., 128-width ConvNet) for the evaluation phase as the traditional methods, outperforming the latest SOTA DataDAM by $1.9\%$ and $3.9\%$ , respectively. To the best of our knowledge, our proposed G-VBSM is the first algorithm that is strongly competitive on CIAR-100 as well as the full $224 \times 224$ ImageNet-1k. + +CIFAR-10. As illustrated in the CIFAR-10 part of Table 5, G-VBSM outperforms vanilla DC [32], KIP [15], and all coreset selection algorithms on the benchmark with 128-width ConvNet as the evaluation model. Meanwhile, on the benchmark with ResNet 18 as the evaluation model, G-VBSM outperforms the baseline SRe2L by $26.3\%$ and $11.7\%$ under IPC 10 and 50, respectively. + +# 5. Conclusion + +In this paper, we introduce a novel perspective termed "generalized matching" for dataset condensation. This posits that an abundance of lightweight "local-match-global" matching are more effective than a single "local-match-global" matching, and even surpass the precise and costly matching used by traditional methods. Consequently, we present G-VBSM, designed to ensure data densification while performing matching based on sufficient and various backbones, layers, and statistics. Experiments conducted on CIFAR-10/100, Tiny-[ImageNet, and the full $224 \times 224$ ImageNet-1k demonstrated that our method is the first algorithm to show strong performance across both small-scale and large-scale datasets. For future research, we plan to extend this approach's applicability to large-scale detection and segmentation datasets. + +# References + +[1] George Cazenavette, Tongzhou Wang, Antonio Torralba, Alexei A. Efros, and Jun-Yan Zhu. Dataset distillation by matching training trajectories. In Computer Vision and Pattern Recognition, New Orleans, LA, USA, 2022. IEEE. 2, 4, 6, 7, 8, 1, 3 +[2] Justin Cui, Ruochen Wang, Si Si, and Cho-Jui Hsieh. Scaling up dataset distillation to imagenet-1k with constant memory. In International Conference on Machine Learning, pages 6565–6590, Honolulu, Hawaii, USA, 2023. PMLR. 2, 7 +[3] Jiawei Du, Yidi Jiang, Vincent Y. F. Tan, Joey Tianyi Zhou, and Haizhou Li. Minimizing the accumulated trajectory error to improve dataset distillation. In Computer Vision and Pattern Recognition, pages 3749-3758, Vancouver, BC, Canada, 2023. IEEE. 2, 4 +[4] Jianping Gou, Baosheng Yu, Stephen J Maybank, and Dacheng Tao. Knowledge distillation: A survey. International Journal of Computer Vision, 129(6):1789-1819, 2021. 1 +[5] Kaiming He, Xiangyu Zhang, and Shaoqing Ren. Deep residual learning for image recognition. In Computer Vision and Pattern Recognition, pages 770-778, Las Vegas, NV, USA, 2016. IEEE. 6 +[6] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network, 2015. 3, 1 +[7] Shengyuan Hu, Jack Goetz, Kshitiz Malik, Hongyuan Zhan, Zhe Liu, and Yue Liu. Fedsynth: Gradient compression via synthetic data in federated learning. arXiv preprint arXiv:2204.01273, 2022. 2 +[8] Tao Huang, Shan You, Fei Wang, Chen Qian, and Chang Xu. Knowledge distillation from a stronger teacher. In Neural Information Processing Systems, 2022. 3, 1 +[9] Jang-Hyun Kim, Jinuk Kim, Seong Joon Oh, Sangdoo Yun, Hwanjun Song, Joonhyun Jeong, Jung-Woo Ha, and Hyun Oh Song. Dataset condensation via efficient synthetic-data parameterization. In International Conference on Machine Learning, pages 11102-11118, Baltimore, Maryland, USA, 2022. PMLR. 2 +[10] Taehyeon Kim, Jaehoon Oh, NakYil Kim, Sangwook Cho, and Se-Young Yun. Comparing kullback-leibler divergence and mean squared error loss in knowledge distillation. In International Joint Conference on Artificial Intelligence, Virtual Event, 2021. Morgan Kaufmann. 5 +[11] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6 +[12] Yanqing Liu, Jianyang Gu, Kai Wang, Zheng Zhu, Wei Jiang, and Yang You. DREAM: efficient dataset distillation by representative matching. arXiv preprint arXiv:2302.14416, 2023. 2 +[13] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In International Conference on Computer Vision, pages 10012-10022, 2021. 6 +[14] Wojciech Masarczyk and Ivona Tautkute. Reducing catastrophic forgetting with learning on synthetic data. In Com + +puter Vision and Pattern Recognition Workshops, pages 252-253, Virtual Event, 2020. IEEE. 2 +[15] Timothy Nguyen, Zhourong Chen, and Jaehoon Lee. Dataset meta-learning from kernel ridge-regression. arXiv preprint arXiv:2011.00050, 2020. 2, 6, 7, 8 +[16] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In Neural Information Processing Systems, Vancouver, BC, Canada, 2019. 6, 2 +[17] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In International Conference on Learning Representations, 2023. 4, 2 +[18] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International Journal of Computer Vision, 115(3):211-252, 2015. 2, 5 +[19] Noveen Sachdeva and Julian McAuley. Data distillation: A survey. arXiv preprint arXiv:2301.04272, 2023. 2 +[20] Ahmad Sajedi, Samir Khaki, Ehsan Amjadian, Lucy Z Liu, Yuri A Lawryshyn, and Konstantinos N Plataniotis. Datadam: Efficient dataset distillation with attention matching. In International Conference on Computer Vision, pages 17097-17107, Paris, France, 2023. IEEE. 3, 6, 7, 8, 1 +[21] Mark Sandler, Andrew G. Howard, Menglong Zhu, Andrey Zhmoginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted residuals and linear bottlenecks*. In *Computer Vision and Pattern Recognition*, pages 4510-4520, Salt Lake City, UT, USA, 2018. IEEE. 6 +[22] Mattia Sangermano, Antonio Carta, Andrea Cossu, and Davide Bacciu. Sample condensation in online continual learning. In International Joint Conference on Neural Networks, pages 1-8, Padua, Italy, 2022. IEEE. 2 +[23] Zhiqiang Shen and Eric Xing. A fast knowledge distillation framework for visual recognition. In European Conference on Computer Vision, pages 673-690. Springer, 2022. 3 +[24] Felipe Petroski Such, Aditya Rawal, Joel Lehman, Kenneth O. Stanley, and Jeffrey Clune. Generative teaching networks: Accelerating neural architecture search by learning to generate synthetic training data. In International Conference on Machine Learning, pages 9206-9216, Virtual Event, 2020. PMLR. 2 +[25] Mingxing Tan and Quoc Le. Efficientnetv2: Smaller models and faster training. In International Conference on Machine Learning, pages 10096-10106, Virtual Event, 2021. PMLR. 6 +[26] Amirhossein Tavanaei. Embedded encoder-decoder in convolutional networks towards explainable AI. 2020. 5 +[27] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive representation distillation. In International Conference on Learning Representations, 2019. 1 +[28] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training + +data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357, Virtual Event, 2021. PMLR. 6 +[29] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-SNE. Journal of Machine Learning Research, 9 (11), 2008. 8 +[30] Chunwei Wang, Chao Ma, Ming Zhu, and Xiaokang Yang. Pointaugmenting: Cross-modal augmentation for 3d object detection. In Computer Vision and Pattern Recognition, pages 11794-11803, Virtual Event, 2021. 2 +[31] Kai Wang, Bo Zhao, Xiangyu Peng, Zheng Zhu, Shuo Yang, Shuo Wang, Guan Huang, Hakan Bilen, Xinchao Wang, and Yang You. Cafe: Learning to condense dataset by aligning features. In Computer Vision and Pattern Recognition, pages 12196-12205, New Orleans, LA, USA, 2022. IEEE. 2, 3, 7 +[32] Tongzhou Wang, Jun-Yan Zhu, Antonio Torralba, and Alexei A Efros. Dataset distillation. arXiv preprint arXiv:1811.10959, 2018. 1, 2, 8, 3 +[33] Hongxu Yin, Pavlo Molchanov, Jose M Alvarez, Zhizhong Li, Arun Mallya, Derek Hoiem, Niraj K Jha, and Jan Kautz. Dreaming to distill: Data-free knowledge transfer via deep-inversion. In Computer Vision and Pattern Recognition, pages 8715-8724, Virtual Event, 2020. IEEE. 2, 5 +[34] Zeyuan Yin, Eric P. Xing, and Zhiqiang Shen. Squeeze, recover and relabel: Dataset condensation at imagenet scale from A new perspective. In Neural Information Processing Systems. NIPS, 2023. 2, 3, 4, 5, 6, 7, 1 +[35] Shan You, Chang Xu, Chao Xu, and Dacheng Tao. Learning from multiple teacher networks. In International Conference on Knowledge Discovery and Data Mining, pages 1285-1294, Halifax, NS, Canada, 2017. ACM. 6 +[36] Ruonan Yu, Songhua Liu, and Xinchao Wang. Dataset distillation: A comprehensive review. arXiv preprint arXiv:2301.07014, 2023. 2 +[37] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In British Machine Vision Conference, York, UK, 2016. BMVA Press. 1, 6 +[38] Lei Zhang, Jie Zhang, Bowen Lei, Subhabrata Mukherjee, Xiang Pan, Bo Zhao, Caiwen Ding, Yao Li, and Dongkuan Xu. Accelerating dataset distillation via model augmentation. In Computer Vision and Pattern Recognition, Vancouver, BC, Canada, 2023. IEEE. 6, 5 +[39] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In Computer Vision and Pattern Recognition, pages 6848-6856, 2018. 6 +[40] Bo Zhao and Hakan Bilen. Dataset condensation with differentiable siamese augmentation. In International Conference on Machine Learning, pages 12674-12685, Virtual Event, 2021. PMLR. 2 +[41] Bo Zhao and Hakan Bilen. Dataset condensation with distribution matching. In Winter Conference on Applications of Computer Vision, pages 6514-6523, Waikoloa, Hawaii, 2023. IEEE. 2, 7, 5 +[42] Bo Zhao, Konda Reddy Mopuri, and Hakan Bilen. Dataset condensation with gradient matching. In International Conference on Learning Representations, Virtual Event, 2021. OpenReview.net. 2, 7 + +[43] Borui Zhao, Quan Cui, Renjie Song, Yiyu Qiu, and Jiajun Liang. Decoupled knowledge distillation. In Computer Vision and Pattern Recognition, pages 11953-11962, 2022. 1 +[44] Yongchao Zhou, Ehsan Nezhadarya, and Jimmy Ba. Dataset distillation using neural feature regression. In Neural Information Processing Systems, New Orleans, LA, USA, 2022. NIPS. 2 \ No newline at end of file diff --git a/2024/Generalized Large-Scale Data Condensation via Various Backbone and Statistical Matching/images.zip b/2024/Generalized Large-Scale Data Condensation via Various Backbone and Statistical Matching/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e379399d0114cb064a2fcfd2e9081e36aeff31e6 --- /dev/null +++ b/2024/Generalized Large-Scale Data Condensation via Various Backbone and Statistical Matching/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:397d96acea2ffe7a97eab3e4359131b99acd6d54fefcad8b6d951267505f676e +size 688300 diff --git a/2024/Generalized Large-Scale Data Condensation via Various Backbone and Statistical Matching/layout.json b/2024/Generalized Large-Scale Data Condensation via Various Backbone and Statistical Matching/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..67762b917f06a4101512876d9daac5008049a106 --- /dev/null +++ b/2024/Generalized Large-Scale Data Condensation via Various Backbone and Statistical Matching/layout.json @@ -0,0 +1,10007 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 77, + 102, + 517, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 102, + 517, + 140 + ], + "spans": [ + { + "bbox": [ + 77, + 102, + 517, + 140 + ], + "type": "text", + "content": "Generalized Large-Scale Data Condensation via Various Backbone and Statistical Matching" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 159, + 536, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 159, + 536, + 219 + ], + "spans": [ + { + "bbox": [ + 56, + 159, + 536, + 219 + ], + "type": "text", + "content": "Shitong Shao" + }, + { + "bbox": [ + 56, + 159, + 536, + 219 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 56, + 159, + 536, + 219 + ], + "type": "text", + "content": " Zeyuan Yin" + }, + { + "bbox": [ + 56, + 159, + 536, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 56, + 159, + 536, + 219 + ], + "type": "text", + "content": " Muxin Zhou" + }, + { + "bbox": [ + 56, + 159, + 536, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 56, + 159, + 536, + 219 + ], + "type": "text", + "content": " Xindong Zhang" + }, + { + "bbox": [ + 56, + 159, + 536, + 219 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 56, + 159, + 536, + 219 + ], + "type": "text", + "content": " Zhiqiang Shen" + }, + { + "bbox": [ + 56, + 159, + 536, + 219 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 56, + 159, + 536, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 56, + 159, + 536, + 219 + ], + "type": "text", + "content": "Mohamed bin Zayed University of AI " + }, + { + "bbox": [ + 56, + 159, + 536, + 219 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 56, + 159, + 536, + 219 + ], + "type": "text", + "content": "Hong Kong Polytechnic University " + }, + { + "bbox": [ + 56, + 159, + 536, + 219 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 56, + 159, + 536, + 219 + ], + "type": "text", + "content": "OPPO Research 1090784053sst@gmail.com, {zeyuan.yin,muxin.zhou,zhiqiang.shen}@mbzuai.ac.ae 17901410r@connect.polyu.hk, *:Corresponding author" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 48, + 245, + 202, + 412 + ], + "blocks": [ + { + "bbox": [ + 48, + 245, + 202, + 412 + ], + "lines": [ + { + "bbox": [ + 48, + 245, + 202, + 412 + ], + "spans": [ + { + "bbox": [ + 48, + 245, + 202, + 412 + ], + "type": "image", + "image_path": "2a9b66f7538f9e2b546c67ba66bfc1ba57e38bbd4635080df03df5a6f7d951d4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 421, + 547, + 467 + ], + "lines": [ + { + "bbox": [ + 45, + 421, + 547, + 467 + ], + "spans": [ + { + "bbox": [ + 45, + 421, + 547, + 467 + ], + "type": "text", + "content": "Figure 1. Left: Our proposed G-VBSM consists of three novel and effective modules, named DD, GSM and GBM. The richness and quality of information in the synthetic data have been significantly enhanced compared with the baseline SRe2L through the sequential merging of DD, GBM, and GSM. Right: G-VBSM prioritizes \"generalized matching\" to ensure consistency between distilled and complete datasets across various backbones, layers, and statistics, and achieves the highest accuracy " + }, + { + "bbox": [ + 45, + 421, + 547, + 467 + ], + "type": "inline_equation", + "content": "31.4\\%" + }, + { + "bbox": [ + 45, + 421, + 547, + 467 + ], + "type": "text", + "content": " on ImageNet-1k under IPC 10." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 203, + 245, + 325, + 412 + ], + "blocks": [ + { + "bbox": [ + 203, + 245, + 325, + 412 + ], + "lines": [ + { + "bbox": [ + 203, + 245, + 325, + 412 + ], + "spans": [ + { + "bbox": [ + 203, + 245, + 325, + 412 + ], + "type": "image", + "image_path": "0ae6f95b9f1d1115b8711b7d4e83f15a64733b534be8adb54a71dbd53f5f3d61.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 326, + 245, + 450, + 412 + ], + "blocks": [ + { + "bbox": [ + 326, + 245, + 450, + 412 + ], + "lines": [ + { + "bbox": [ + 326, + 245, + 450, + 412 + ], + "spans": [ + { + "bbox": [ + 326, + 245, + 450, + 412 + ], + "type": "image", + "image_path": "e9dc599c01779e3a0f1f98890a49d8ac3e9132403f402adeabdefcf87fc374b8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 450, + 245, + 544, + 412 + ], + "blocks": [ + { + "bbox": [ + 450, + 245, + 544, + 412 + ], + "lines": [ + { + "bbox": [ + 450, + 245, + 544, + 412 + ], + "spans": [ + { + "bbox": [ + 450, + 245, + 544, + 412 + ], + "type": "image", + "image_path": "d73b32c7cfbae21d6376f5aecd8c891595761ee508a054a3a6fdf18df025c103.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 475, + 192, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 475, + 192, + 488 + ], + "spans": [ + { + "bbox": [ + 143, + 475, + 192, + 488 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": "The lightweight \"local-match-global\" matching introduced by SRe2L successfully creates a distilled dataset with comprehensive information on the full " + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": " ImageNet-1k. However, this one-sided approach is limited to a particular backbone, layer, and statistics, which limits the improvement of the generalization of a distilled dataset. We suggest that sufficient and various \"local-match-global\" matching are more precise and effective than a single one and have the ability to create a distilled dataset with richer information and better generalization ability. We call this perspective \"generalized matching\" and propose Generalized Various Backbone and Statistical Matching (G-VBSM) in this work, which aims to create a synthetic dataset with densities, ensuring consistency with the complete dataset across various backbones, layers, and statistics. As experimentally demonstrated, G-VBSM is the first algorithm to obtain strong performance across both" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 476, + 547, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 476, + 547, + 538 + ], + "spans": [ + { + "bbox": [ + 305, + 476, + 547, + 538 + ], + "type": "text", + "content": "small-scale and large-scale datasets. Specifically, G-VBSM achieves performances of " + }, + { + "bbox": [ + 305, + 476, + 547, + 538 + ], + "type": "inline_equation", + "content": "38.7\\%" + }, + { + "bbox": [ + 305, + 476, + 547, + 538 + ], + "type": "text", + "content": " on CIFAR-100, " + }, + { + "bbox": [ + 305, + 476, + 547, + 538 + ], + "type": "inline_equation", + "content": "47.6\\%" + }, + { + "bbox": [ + 305, + 476, + 547, + 538 + ], + "type": "text", + "content": " on Tiny-ImageNet, and " + }, + { + "bbox": [ + 305, + 476, + 547, + 538 + ], + "type": "inline_equation", + "content": "31.4\\%" + }, + { + "bbox": [ + 305, + 476, + 547, + 538 + ], + "type": "text", + "content": " on the full " + }, + { + "bbox": [ + 305, + 476, + 547, + 538 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 305, + 476, + 547, + 538 + ], + "type": "text", + "content": " ImageNet-1k, respectively1. These results surpass all SOTA methods by margins of " + }, + { + "bbox": [ + 305, + 476, + 547, + 538 + ], + "type": "inline_equation", + "content": "3.9\\%" + }, + { + "bbox": [ + 305, + 476, + 547, + 538 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 476, + 547, + 538 + ], + "type": "inline_equation", + "content": "6.5\\%" + }, + { + "bbox": [ + 305, + 476, + 547, + 538 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 476, + 547, + 538 + ], + "type": "inline_equation", + "content": "10.1\\%" + }, + { + "bbox": [ + 305, + 476, + 547, + 538 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 544, + 387, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 544, + 387, + 556 + ], + "spans": [ + { + "bbox": [ + 306, + 544, + 387, + 556 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 565, + 547, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 565, + 547, + 673 + ], + "spans": [ + { + "bbox": [ + 304, + 565, + 547, + 673 + ], + "type": "text", + "content": "With the development of deep learning, the number of model parameters and the quantity of training data have become increasingly large [4, 37]. Researchers have tried to minimize the training overhead while preventing a decline in the generalization ability. Data condensation (DC), also known as Dataset distillation, first introduced by Wang et al. [32], aims to alleviate the training burden by synthesizing a small yet informative distilled dataset derived from the complete training dataset, while ensuring that the behav" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 306, + 683, + 547, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 683, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 683, + 547, + 712 + ], + "type": "text", + "content": "1Settings: CIFAR-100 with 128-width ConvNet under 10 images per class (IPC), Tiny-ImageNet with ResNet18 under 50 IPC, and ImageNet1k with ResNet18 under 10 IPC." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16709" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 167 + ], + "type": "text", + "content": "ior of the distilled dataset on the target task remains consistent with that of the complete dataset. The extremely compressed distilled dataset contains sufficiently valuable information and have the potential for fast model training, and have been become a popular choice for different downstream application, like federated learning [7, 22], continual learning [14, 22, 40], neural architecture search [24, 41, 42] and 3D point clouds [30]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 170, + 290, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 170, + 290, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 170, + 290, + 422 + ], + "type": "text", + "content": "A persistent problem that researchers have been working to solve [2, 9, 12, 42] in DC is the substantial data synthesis overhead [19, 36], which hinders its application in real-world large-scale datasets (e.g., ImageNet-1k) training. Typical performance matching [15, 32, 44] and trajectory matching [1, 3] unroll recursive computation graphs, requiring substantial GPU memory and resulting in prohibitive training costs. Zhao et al. [42] proposed gradient matching to address this, synthesizing distilled datasets by matching gradients from synthetic and real data in a single step. However, gradient computation and matching remain time-consuming [9], leading to the proposal of distribution matching [41]. This method and its variants [31] employ a network-based feature extractor to embed both synthetic and real data into a high-dimensional Hilbert space, then perform distribution matching. The training load for this direct, single-step process stems only from one gradient update of the synthetic data and the feature extractor [36]. Unfortunately, all of the above mentioned improved methods still have extremely large training overheads on the full " + }, + { + "bbox": [ + 46, + 170, + 290, + 422 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 170, + 290, + 422 + ], + "type": "text", + "content": " ImageNet-1k." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 424, + 288, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 424, + 288, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 424, + 288, + 604 + ], + "type": "text", + "content": "Recently, SRe2L [34] accomplished data condensation for the first time on the full " + }, + { + "bbox": [ + 46, + 424, + 288, + 604 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 424, + 288, + 604 + ], + "type": "text", + "content": " ImageNet-1k [18], achieving Top-1 validation accuracy " + }, + { + "bbox": [ + 46, + 424, + 288, + 604 + ], + "type": "inline_equation", + "content": "21.3\\%" + }, + { + "bbox": [ + 46, + 424, + 288, + 604 + ], + "type": "text", + "content": " with ResNet18 under IPC 10. This method outperformed the latest state-of-the-art TESLA [2], which conducted on a low-resolution version of ImageNet-1k, by being " + }, + { + "bbox": [ + 46, + 424, + 288, + 604 + ], + "type": "inline_equation", + "content": "16 \\times" + }, + { + "bbox": [ + 46, + 424, + 288, + 604 + ], + "type": "text", + "content": " faster and improved performance by a margin of " + }, + { + "bbox": [ + 46, + 424, + 288, + 604 + ], + "type": "inline_equation", + "content": "13.6\\%" + }, + { + "bbox": [ + 46, + 424, + 288, + 604 + ], + "type": "text", + "content": ". SRe2L is inspired by DeepInversion [33] and aims to match statistics in BatchNorm generated from synthetic and real data. We reevaluate the success of SRe2L through the lightweight \"local-match-global\" essentially. The \"local-match-global\" refers to utilizing more comprehensive information (e.g., statistics in BatchNorm), generated from the model using the complete dataset (global), to guide the parameter update of the distilled dataset (local)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "However, such lightweight and effective matching in SRe2L is singular, depending on the particular layer (i.e., BatchNorm), model (i.e., ResNet18), and statistics (i.e., channel mean/variance). Intuitively, sufficient \"local-match-global\" matching can result in more accurate and rational supervision than a single one, further enhancing the generalization of the distilled dataset. We call this perspective \"generalized matching\" and suggest that the distilled dataset is likely to perform consistent with the complete" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": "dataset on the evaluation model, provided that there is sufficient variety in backbones, layers, and statistics used for matching." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 110, + 546, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 110, + 546, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 110, + 546, + 384 + ], + "type": "text", + "content": "Inspired by this, we propose Generalized Various Backbone and Statistical Matching (G-VBSM) to fulfill \"generalized matching\". G-VBSM comprises three integral and effective parts named data densification (DD), generalized statistical matching (GSM), and generalized backbone matching (GBM). DD is utilized to ensure that the images within each class are linearly independent, thereby enhancing the (intra-class) diversity of the distilled dataset. This ultimately guarantees that \"generalized matching\" preserves the rich and diverse information within the synthetic data. GBM and GSM are designed to implement \"generalized matching\". Among them, GSM traverses the complete dataset without computing and updating the gradient, to record the statistics of Convolution at both the patch and channel levels. These statistics are subsequently utilized for matching during the data synthesis phase, in conjunction with the channel-level statistics in BatchNorm. Furthermore, GBM aims to ensure consistency between distilled and complete datasets across various backbones, enhancing matching sufficiency and leading to strong generalization in the evaluation phase. In particular, G-VBSM also ensures the efficiency of dataset condensation through a series of strategies, as mentioned in Sec. 3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 386, + 545, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 386, + 545, + 531 + ], + "spans": [ + { + "bbox": [ + 304, + 386, + 545, + 531 + ], + "type": "text", + "content": "Extensive experiments on CIFAR-10, CIFAR-100, TinyImageNet, and the full " + }, + { + "bbox": [ + 304, + 386, + 545, + 531 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 304, + 386, + 545, + 531 + ], + "type": "text", + "content": " ImageNet-1k, demonstrating that G-VBSM is the first algorithm that performs well on both small-scale and large-scale datasets. Specifically, we not only verify through ablation studies that GSM, GBM and DD are consistently reliable, but also achieve the highest " + }, + { + "bbox": [ + 304, + 386, + 545, + 531 + ], + "type": "inline_equation", + "content": "38.7\\%" + }, + { + "bbox": [ + 304, + 386, + 545, + 531 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 386, + 545, + 531 + ], + "type": "inline_equation", + "content": "47.6\\%" + }, + { + "bbox": [ + 304, + 386, + 545, + 531 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 386, + 545, + 531 + ], + "type": "inline_equation", + "content": "31.4\\%" + }, + { + "bbox": [ + 304, + 386, + 545, + 531 + ], + "type": "text", + "content": " on CIFAR-100 (128-width ConvNet), Tiny-ImageNet (ResNet18), and the full " + }, + { + "bbox": [ + 304, + 386, + 545, + 531 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 304, + 386, + 545, + 531 + ], + "type": "text", + "content": " ImageNet-1k (ResNet18) under images per class (IPC) 10, 50 and 10, respectively, which outperforms all previous state-of-the-art (SOTA) methods by " + }, + { + "bbox": [ + 304, + 386, + 545, + 531 + ], + "type": "inline_equation", + "content": "3.9\\%" + }, + { + "bbox": [ + 304, + 386, + 545, + 531 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 386, + 545, + 531 + ], + "type": "inline_equation", + "content": "6.5\\%" + }, + { + "bbox": [ + 304, + 386, + 545, + 531 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 386, + 545, + 531 + ], + "type": "inline_equation", + "content": "10.1\\%" + }, + { + "bbox": [ + 304, + 386, + 545, + 531 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 544, + 385, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 544, + 385, + 558 + ], + "spans": [ + { + "bbox": [ + 306, + 544, + 385, + 558 + ], + "type": "text", + "content": "2. Background" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 566, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 566, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 566, + 547, + 715 + ], + "type": "text", + "content": "Dataset condensation (DC) represents a data synthesis procedure that aims to compress a complete, large dataset " + }, + { + "bbox": [ + 304, + 566, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = \\{(X_i,y_i)\\}_{i = 1}^{|T|}" + }, + { + "bbox": [ + 304, + 566, + 547, + 715 + ], + "type": "text", + "content": " into a smaller, distilled dataset " + }, + { + "bbox": [ + 304, + 566, + 547, + 715 + ], + "type": "inline_equation", + "content": "S = \\{(\\tilde{X}_i,\\tilde{y}_i)\\}_{i = 1}^{|S|}" + }, + { + "bbox": [ + 304, + 566, + 547, + 715 + ], + "type": "text", + "content": ", subject to " + }, + { + "bbox": [ + 304, + 566, + 547, + 715 + ], + "type": "inline_equation", + "content": "|\\mathcal{S}|\\ll |\\mathcal{T}|" + }, + { + "bbox": [ + 304, + 566, + 547, + 715 + ], + "type": "text", + "content": ", while ensuring that an arbitrary evaluation model " + }, + { + "bbox": [ + 304, + 566, + 547, + 715 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{eval}}(\\cdot)" + }, + { + "bbox": [ + 304, + 566, + 547, + 715 + ], + "type": "text", + "content": " trained on " + }, + { + "bbox": [ + 304, + 566, + 547, + 715 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 566, + 547, + 715 + ], + "type": "text", + "content": " yields similar performance to one trained on " + }, + { + "bbox": [ + 304, + 566, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 304, + 566, + 547, + 715 + ], + "type": "text", + "content": ". Classical data distillation algorithms invariably require the candidate model " + }, + { + "bbox": [ + 304, + 566, + 547, + 715 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{cand}}(\\cdot)" + }, + { + "bbox": [ + 304, + 566, + 547, + 715 + ], + "type": "text", + "content": " to execute one or more steps on " + }, + { + "bbox": [ + 304, + 566, + 547, + 715 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 566, + 547, + 715 + ], + "type": "text", + "content": " to update its parameter " + }, + { + "bbox": [ + 304, + 566, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{cand}}" + }, + { + "bbox": [ + 304, + 566, + 547, + 715 + ], + "type": "text", + "content": ", subsequently achieving matching in terms of performance [15, 32], gradient [42], trajectory [1, 2], or distribution [31, 41]. The process " + }, + { + "bbox": [ + 304, + 566, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{cand}} - \\alpha \\nabla_{\\theta_{\\mathrm{cand}}}\\ell (f_{\\mathrm{cand}}(\\tilde{X}),\\tilde{y})" + }, + { + "bbox": [ + 304, + 566, + 547, + 715 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 566, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\ell (\\cdot ,\\cdot)" + }, + { + "bbox": [ + 304, + 566, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 566, + 547, + 715 + ], + "type": "inline_equation", + "content": "(\\tilde{X},\\tilde{y})" + }, + { + "bbox": [ + 304, + 566, + 547, + 715 + ], + "type": "text", + "content": " represent the loss function and" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16710" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 72, + 517, + 274 + ], + "blocks": [ + { + "bbox": [ + 80, + 72, + 517, + 274 + ], + "lines": [ + { + "bbox": [ + 80, + 72, + 517, + 274 + ], + "spans": [ + { + "bbox": [ + 80, + 72, + 517, + 274 + ], + "type": "image", + "image_path": "c1a58103c5aaca085ba63e123665764fe60b78a9ec1160efafb18d383dc7e70a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 282, + 547, + 308 + ], + "lines": [ + { + "bbox": [ + 46, + 282, + 547, + 308 + ], + "spans": [ + { + "bbox": [ + 46, + 282, + 547, + 308 + ], + "type": "text", + "content": "Figure 2. The overview of G-VBSM on the full " + }, + { + "bbox": [ + 46, + 282, + 547, + 308 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 282, + 547, + 308 + ], + "type": "text", + "content": " ImageNet-1k, which ensures the consistency between the distilled and the complete datasets across various backbones, layers and statistics via \"generalized matching\"." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 316, + 288, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 288, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 288, + 399 + ], + "type": "text", + "content": "a batch sampled from " + }, + { + "bbox": [ + 46, + 316, + 288, + 399 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 316, + 288, + 399 + ], + "type": "text", + "content": ", respectively, is notably time-consuming. Consequently, even the relatively swifest distribution matching [20, 31] is slower than the recent proposed SRe2L [34]. In fact, SRe2L is the only workable way to achieve DC on the full " + }, + { + "bbox": [ + 46, + 316, + 288, + 399 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 316, + 288, + 399 + ], + "type": "text", + "content": " ImageNet-1k, as it requires updating the parameters of the synthetic data only once per iteration." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 400, + 290, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 400, + 290, + 520 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 290, + 520 + ], + "type": "text", + "content": "SRe2L [34] encompasses three incremental subprocesses: Squeeze, Recover, and Relabel. Squeeze is designed to train " + }, + { + "bbox": [ + 46, + 400, + 290, + 520 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{cand}}(\\cdot)" + }, + { + "bbox": [ + 46, + 400, + 290, + 520 + ], + "type": "text", + "content": " containing BatchNorm in a standard manner, aiming to record the global channel mean " + }, + { + "bbox": [ + 46, + 400, + 290, + 520 + ], + "type": "inline_equation", + "content": "\\mathbf{BN}_l^{\\mathrm{CM}}" + }, + { + "bbox": [ + 46, + 400, + 290, + 520 + ], + "type": "text", + "content": " and channel variance " + }, + { + "bbox": [ + 46, + 400, + 290, + 520 + ], + "type": "inline_equation", + "content": "\\mathbf{BN}_l^{\\mathrm{CV}}" + }, + { + "bbox": [ + 46, + 400, + 290, + 520 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 400, + 290, + 520 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 400, + 290, + 520 + ], + "type": "text", + "content": " refers to the index of the " + }, + { + "bbox": [ + 46, + 400, + 290, + 520 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 400, + 290, + 520 + ], + "type": "text", + "content": "-th layer) via exponential moving average (EMA), extracted from " + }, + { + "bbox": [ + 46, + 400, + 290, + 520 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 46, + 400, + 290, + 520 + ], + "type": "text", + "content": ", for subsequent matching in Recover. In Recover after that, given the channel mean " + }, + { + "bbox": [ + 46, + 400, + 290, + 520 + ], + "type": "inline_equation", + "content": "\\mu_l(\\tilde{X})" + }, + { + "bbox": [ + 46, + 400, + 290, + 520 + ], + "type": "text", + "content": " and channel variance " + }, + { + "bbox": [ + 46, + 400, + 290, + 520 + ], + "type": "inline_equation", + "content": "\\sigma_l^2(\\tilde{X})" + }, + { + "bbox": [ + 46, + 400, + 290, + 520 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 46, + 400, + 290, + 520 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 400, + 290, + 520 + ], + "type": "text", + "content": "-th BatchNorm obtained from " + }, + { + "bbox": [ + 46, + 400, + 290, + 520 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 400, + 290, + 520 + ], + "type": "text", + "content": ", the statistical matching loss function can be formulated as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 524, + 287, + 546 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 524, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 58, + 524, + 287, + 546 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {B N}} (\\tilde {X}) = \\sum_ {l} \\left| \\left| \\mu_ {l} (\\tilde {X}) - \\mathbf {B} \\mathbf {N} _ {l} ^ {\\mathrm {C M}} \\right| \\right| _ {2} + \\left| \\left| \\sigma_ {l} ^ {2} (\\tilde {X}) - \\mathbf {B} \\mathbf {N} _ {l} ^ {\\mathrm {C V}} \\right| \\right| _ {2}. \\tag {1}", + "image_path": "843d9f568ef9b26f635d28915e925e14b4e9919b40b779fe8523eeae38268156.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 551, + 287, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 551, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 551, + 287, + 574 + ], + "type": "text", + "content": "Based on this, we can give the entire optimization objective in Recover as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 578, + 287, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 578, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 106, + 578, + 287, + 597 + ], + "type": "interline_equation", + "content": "\\underset {\\tilde {X}} {\\arg \\min } \\mathcal {L} _ {\\mathrm {B N}} (\\tilde {X}) + \\ell \\left(f _ {\\text {c a n d}} (\\tilde {X}), y\\right), \\tag {2}", + "image_path": "86154afbbbc87dd6816a4384e0536cf0b84eddc0dfdb3872a9abe361422ade0b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 601, + 288, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 601, + 288, + 649 + ], + "spans": [ + { + "bbox": [ + 47, + 601, + 288, + 649 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 601, + 288, + 649 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 47, + 601, + 288, + 649 + ], + "type": "text", + "content": " denotes the ground truth label. Moreover, SRe2L assigns soft labels " + }, + { + "bbox": [ + 47, + 601, + 288, + 649 + ], + "type": "inline_equation", + "content": "\\tilde{y}" + }, + { + "bbox": [ + 47, + 601, + 288, + 649 + ], + "type": "text", + "content": " to the synthetic data, utilizing the logit-based distillation [6, 8] to improve the generalization ability of the distilled dataset. This can be denoted as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 656, + 287, + 670 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 656, + 287, + 670 + ], + "spans": [ + { + "bbox": [ + 116, + 656, + 287, + 670 + ], + "type": "interline_equation", + "content": "\\tilde {y} = \\operatorname {s o f t m a x} \\left(f _ {\\text {c a n d}} (\\tilde {X}) / \\tau\\right), \\tag {3}", + "image_path": "b5fcd117219f1e1ef1b6aba1d985a82dc1479fd0231cd7ac254e2322a783132d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "type": "text", + "content": " denotes the temperature to regulate the smoothness of the soft labels, thereby enhancing the distilled dataset's potential for generalization to unseen evaluation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 316, + 547, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 547, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 547, + 413 + ], + "type": "text", + "content": "models. The generated soft label can be stored on disk using FKD [23] so as not to defeat the purpose of DC. A crucial point of SRe2L is that BatchNorm calculates the statistics of the entire dataset using EMA, thereby providing a comprehensive and representative matching for the distilled dataset. Encouraged by this, our research focuses on applying sufficient 'local-match-global' matching to achieve \"generalized matching\"." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 425, + 362, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 425, + 362, + 437 + ], + "spans": [ + { + "bbox": [ + 306, + 425, + 362, + 437 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 445, + 547, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 547, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 547, + 601 + ], + "type": "text", + "content": "The comprehensive and detailed framework of our proposed Generalized Various Backbone and Statistical Matching (G-VBSM) is illustrated in Fig. 2. In essence, G-VBSM employs the lightweight regularization strategy data densification (DD) to ensure both the diversity and density of the distilled dataset, ensuring that the potential of \"generalized matching\" can be fully exploited. Moreover, generalized backbone matching (GBM) and generalized statistical matching (GSM) are utilized to achieve \"generalized matching\" by performing \"local-match-global\" matching across various backbones, layers, and statistics. In particular, the efficiency and effectiveness of DD, GBM, GSM, SRe2L, and TESLA are illustrated in Fig. 5." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 610, + 416, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 610, + 416, + 623 + ], + "spans": [ + { + "bbox": [ + 306, + 610, + 416, + 623 + ], + "type": "text", + "content": "3.1. Data Densification" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "As illustrated in Fig. 3, the synthetic data generated by SRe2L exhibit excessive similarity within the same class, leading to a lack of diversity. Consequently, even if \"generalized matching\" preserves sufficient valuable information within a single image, the aggregate information content across all images within the same class does not increase effectively, which ultimately prevents \"generalized match-" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16711" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 71, + 279, + 189 + ], + "blocks": [ + { + "bbox": [ + 53, + 71, + 279, + 189 + ], + "lines": [ + { + "bbox": [ + 53, + 71, + 279, + 189 + ], + "spans": [ + { + "bbox": [ + 53, + 71, + 279, + 189 + ], + "type": "image", + "image_path": "8f58880b3cd8c4670f8e8cb6424c374c074981b2fb81338a8af8fca87549b6f7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 193, + 287, + 216 + ], + "lines": [ + { + "bbox": [ + 47, + 193, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 47, + 193, + 287, + 216 + ], + "type": "text", + "content": "Figure 3. Visualization of the mean cosine similarity between pairwise samples within the same class on ImageNet-1k under IPC 10." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 223, + 286, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 223, + 286, + 295 + ], + "spans": [ + { + "bbox": [ + 47, + 223, + 286, + 295 + ], + "type": "text", + "content": "ing\" from being sufficiently advantageous. Data densification (DD) is proposed to address this by ensuring the data " + }, + { + "bbox": [ + 47, + 223, + 286, + 295 + ], + "type": "inline_equation", + "content": "\\tilde{X}" + }, + { + "bbox": [ + 47, + 223, + 286, + 295 + ], + "type": "text", + "content": " has full rank in the batch dimension, thereby guaranteeing that samples in each class are linearly independent, and ultimately ensuring that the data is diverse and abundant to fully exploit the potential of \"generalized matching\"." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 295, + 286, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 295, + 286, + 331 + ], + "spans": [ + { + "bbox": [ + 47, + 295, + 286, + 331 + ], + "type": "text", + "content": "To execute this pipeline, " + }, + { + "bbox": [ + 47, + 295, + 286, + 331 + ], + "type": "inline_equation", + "content": "\\tilde{X}" + }, + { + "bbox": [ + 47, + 295, + 286, + 331 + ], + "type": "text", + "content": " first needs to be downsampled to reduce the computational cost of eigenvalue decomposition:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 80, + 336, + 287, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 336, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 80, + 336, + 287, + 350 + ], + "type": "interline_equation", + "content": "\\hat {X} = \\operatorname {A v g P o o l 2 d} (\\tilde {X}), \\text {s . t .} \\hat {X} \\in \\mathbb {R} ^ {B \\times C \\times 3 2 \\times 3 2}, \\tag {4}", + "image_path": "5c073eef355d36e6a26e72dddcebe0ed99e63972afadfede388b133448529bcc.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 357, + 287, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 357, + 287, + 492 + ], + "spans": [ + { + "bbox": [ + 47, + 357, + 287, + 492 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 357, + 287, + 492 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 47, + 357, + 287, + 492 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 357, + 287, + 492 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 357, + 287, + 492 + ], + "type": "text", + "content": " represent the batch size and the number of channels, respectively. Afterward, we reshape " + }, + { + "bbox": [ + 47, + 357, + 287, + 492 + ], + "type": "inline_equation", + "content": "\\hat{X}" + }, + { + "bbox": [ + 47, + 357, + 287, + 492 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 47, + 357, + 287, + 492 + ], + "type": "inline_equation", + "content": "B\\times C\\times 32\\times 32" + }, + { + "bbox": [ + 47, + 357, + 287, + 492 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 47, + 357, + 287, + 492 + ], + "type": "inline_equation", + "content": "B\\times (1024C)" + }, + { + "bbox": [ + 47, + 357, + 287, + 492 + ], + "type": "text", + "content": " and perform matrix multiplication in each class " + }, + { + "bbox": [ + 47, + 357, + 287, + 492 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 47, + 357, + 287, + 492 + ], + "type": "text", + "content": " to obtain the set of the Gram matrix " + }, + { + "bbox": [ + 47, + 357, + 287, + 492 + ], + "type": "inline_equation", + "content": "\\{\\hat{X}_y\\hat{X}_y^T\\}_{y\\in \\mathcal{Y}}" + }, + { + "bbox": [ + 47, + 357, + 287, + 492 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 357, + 287, + 492 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 47, + 357, + 287, + 492 + ], + "type": "text", + "content": " refers to a set of all classes existing in this batch. And " + }, + { + "bbox": [ + 47, + 357, + 287, + 492 + ], + "type": "inline_equation", + "content": "\\hat{X}_y" + }, + { + "bbox": [ + 47, + 357, + 287, + 492 + ], + "type": "text", + "content": " is a subbatch with class " + }, + { + "bbox": [ + 47, + 357, + 287, + 492 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 47, + 357, + 287, + 492 + ], + "type": "text", + "content": ". Note that the alternative form " + }, + { + "bbox": [ + 47, + 357, + 287, + 492 + ], + "type": "inline_equation", + "content": "\\{\\hat{X}_y^T\\hat{X}_y\\}_{y\\in \\mathcal{Y}}" + }, + { + "bbox": [ + 47, + 357, + 287, + 492 + ], + "type": "text", + "content": " is not desirable, as it is applicable only for dimensionality reduction in feature dimensions, which is why we do not consider singular value decomposition (SVD). To render " + }, + { + "bbox": [ + 47, + 357, + 287, + 492 + ], + "type": "inline_equation", + "content": "\\hat{X}_y\\hat{X}_y^T" + }, + { + "bbox": [ + 47, + 357, + 287, + 492 + ], + "type": "text", + "content": " as full-rank as possible, we introduce the data densification loss in Eq. 5." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 71, + 498, + 287, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 498, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 71, + 498, + 287, + 521 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {D D}} (\\tilde {X}) = \\sum_ {y \\in \\mathcal {Y}} D _ {\\mathrm {K L}} (\\operatorname {s t o p} - \\operatorname {g r a d} (p \\left(\\Sigma_ {y} / \\tau\\right)) | | p \\left(\\Sigma_ {y}\\right)), \\tag {5}", + "image_path": "49331d719bcac82a8198c48a3ec9e747df2a595155e0c52cbdc799d63228f070.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 528, + 287, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 528, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 47, + 528, + 287, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 528, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\Sigma_y" + }, + { + "bbox": [ + 47, + 528, + 287, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 528, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 47, + 528, + 287, + 666 + ], + "type": "text", + "content": ", stop\\_grad(\\cdot) and " + }, + { + "bbox": [ + 47, + 528, + 287, + 666 + ], + "type": "inline_equation", + "content": "p(\\cdot)" + }, + { + "bbox": [ + 47, + 528, + 287, + 666 + ], + "type": "text", + "content": " refer to the eigenvalues of " + }, + { + "bbox": [ + 47, + 528, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\hat{X}_y\\hat{X}_y^T" + }, + { + "bbox": [ + 47, + 528, + 287, + 666 + ], + "type": "text", + "content": ", the temperature, the stop gradient operator and the softmax function, respectively. And " + }, + { + "bbox": [ + 47, + 528, + 287, + 666 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{KL}}(\\cdot ||\\cdot)" + }, + { + "bbox": [ + 47, + 528, + 287, + 666 + ], + "type": "text", + "content": " denotes Kullback-Leibler divergence. As demonstrated in Sec. E, the diversity of the data is significantly enhanced by the employment of Eq. 5. In our experiment, " + }, + { + "bbox": [ + 47, + 528, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 47, + 528, + 287, + 666 + ], + "type": "text", + "content": " is set as 4 in default and we do not assign a deliberate weight (set to 1 by default) to " + }, + { + "bbox": [ + 47, + 528, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{DD}}" + }, + { + "bbox": [ + 47, + 528, + 287, + 666 + ], + "type": "text", + "content": " because " + }, + { + "bbox": [ + 47, + 528, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{DD}} \\equiv 0" + }, + { + "bbox": [ + 47, + 528, + 287, + 666 + ], + "type": "text", + "content": " at the early " + }, + { + "bbox": [ + 47, + 528, + 287, + 666 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 47, + 528, + 287, + 666 + ], + "type": "text", + "content": " of the iterations. In other words, DD is quite stable and the optimization objective " + }, + { + "bbox": [ + 47, + 528, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\arg \\min_{\\{\\Sigma_y\\}_{y \\in \\mathcal{Y}}} \\mathcal{L}_{\\mathrm{DD}}" + }, + { + "bbox": [ + 47, + 528, + 287, + 666 + ], + "type": "text", + "content": " is equivalent to " + }, + { + "bbox": [ + 47, + 528, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\arg \\min_{\\{\\Sigma_y\\}_{y \\in \\mathcal{Y}}} \\sum_y \\sigma^2(\\Sigma_y)" + }, + { + "bbox": [ + 47, + 528, + 287, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "Technical Detail. A problem that warrants attention is that in the SRe2L's implementation [34], having merely a single sample in each class of a batch indicates insufficient" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 323, + 69, + 529, + 175 + ], + "blocks": [ + { + "bbox": [ + 323, + 69, + 529, + 175 + ], + "lines": [ + { + "bbox": [ + 323, + 69, + 529, + 175 + ], + "spans": [ + { + "bbox": [ + 323, + 69, + 529, + 175 + ], + "type": "image", + "image_path": "53d0feb31231732b0ef278c1c9463a5864135201a3f84222449c2da59b8c46a7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 175, + 544, + 186 + ], + "lines": [ + { + "bbox": [ + 306, + 175, + 544, + 186 + ], + "spans": [ + { + "bbox": [ + 306, + 175, + 544, + 186 + ], + "type": "text", + "content": "Figure 4. The illustration of the original loop and the reorder loop." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 193, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 193, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 193, + 545, + 361 + ], + "type": "text", + "content": "to execute DD under the order of the original loop, as depicted in Fig. 4 (Left). A simple solution is to translate the original loop to the reorder loop, as shown in Fig. 4 (Right). However, our experiment on ResNet50 (i.e., the evaluation model) substantiates that this solution suffers a " + }, + { + "bbox": [ + 304, + 193, + 545, + 361 + ], + "type": "inline_equation", + "content": "2.6\\%" + }, + { + "bbox": [ + 304, + 193, + 545, + 361 + ], + "type": "text", + "content": " accuracy degradation (details can be found in Sec. 4.1) on ImageNet-1k under IPC 10. The reason is that the number of classes in each iteration within the reorder loop is insufficient, preventing a batch have the ability to match the global statistics in BatchNorm (i.e., " + }, + { + "bbox": [ + 304, + 193, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{BN}_l^{\\mathrm{CM}}" + }, + { + "bbox": [ + 304, + 193, + 545, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 193, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{BN}_l^{\\mathrm{CV}}" + }, + { + "bbox": [ + 304, + 193, + 545, + 361 + ], + "type": "text", + "content": "). Motivated by score distillation sampling (SDS) [17], we update the statistics during data synthesis via EMA to solve this issue, so that the statistics of all past batches can assist the statistics of the current batch match " + }, + { + "bbox": [ + 304, + 193, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{BN}_l^{\\mathrm{CM}}" + }, + { + "bbox": [ + 304, + 193, + 545, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 193, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{BN}_l^{\\mathrm{CV}}" + }, + { + "bbox": [ + 304, + 193, + 545, + 361 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 365, + 545, + 425 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 365, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 306, + 365, + 545, + 425 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mu_ {l} ^ {\\text {t o t a l}} = \\alpha \\mu_ {l} ^ {\\text {t o t a l}} + (1 - \\alpha) \\mu_ {l} (\\tilde {X}), \\sigma_ {l} ^ {2, \\text {t o t a l}} = \\alpha \\sigma_ {l} ^ {2, \\text {t o t a l}} + (1 - \\alpha) \\sigma_ {l} ^ {2} (\\tilde {X}), \\\\ \\mathcal {L} _ {\\mathrm {B N}} ^ {\\prime} (\\tilde {X}) = \\sum_ {l} | | \\mu_ {l} (\\tilde {X}) - \\mathbf {B N} _ {l} ^ {\\mathrm {C M}} - \\operatorname {s t o p \\_ g r a d} (\\mu_ {l} (\\tilde {X}) - \\mu_ {l} ^ {\\text {t o t a l}}) | | _ {2} \\\\ + \\left\\| \\sigma_ {l} ^ {2} (\\tilde {X}) - \\mathbf {B N} _ {l} ^ {\\mathrm {C V}} - \\operatorname {s t o p \\_ g r a d} \\left(\\sigma_ {l} ^ {2} (\\tilde {X}) - \\sigma_ {l} ^ {2, \\text {t o t a l}}\\right) \\right\\| _ {2}. \\tag {6} \\\\ \\end{array}", + "image_path": "b7f58f65b8ccd3a8b208ccf8f90ffd986fc680aa4a16e080dd61157c57315c76.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 426, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 472 + ], + "type": "text", + "content": "The derivation of " + }, + { + "bbox": [ + 304, + 426, + 545, + 472 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{BN}}^{\\prime}(X)" + }, + { + "bbox": [ + 304, + 426, + 545, + 472 + ], + "type": "text", + "content": " can be found in Appendix B. We call this lightweight strategy as \"match in the form of score distillation sampling\" and have demonstrated its effectiveness and feasibility in our ablation studies." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 479, + 482, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 482, + 493 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 482, + 493 + ], + "type": "text", + "content": "3.2. Generalized Backbone Matching" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": "Performing data synthesis only on a single pre-trained model is not able to enjoy ensemble gains from various backbones. Meanwhile, classical DC algorithms such as MTT [1] and FTD [3] obtain performance improvements from multiple randomly initialized backbones. Therefore, introducing generalized backbone matching (GBM) to apply various backbones for data synthesis is a desirable choice for \"generalized matching\". It ensures a number of \"local-match-global\" matching nearly " + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{b}} \\times" + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": " higher compared to just depending on a single backbone, where " + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{b}}" + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": " denotes the number of backbones. Regrettably, unrolling various backbone computational graphs in parallel for data synthesis is extremely expensive at the computational cost level. A solution is to randomly sample a backbone from the candidates in per iteration. This simple yet effective strategy not only ensures computational efficiency but also allows the data synthesis process to benefit from the diverse and multifaceted matching that the various backbones pro" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16712" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 179, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 179, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 179, + 85 + ], + "type": "text", + "content": "vide. We denote this pipeline as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 66, + 89, + 287, + 101 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 89, + 287, + 101 + ], + "spans": [ + { + "bbox": [ + 66, + 89, + 287, + 101 + ], + "type": "interline_equation", + "content": "f _ {\\text {c a n d}} (\\cdot) \\sim \\mathcal {U} (\\mathbb {S}), \\mathbb {S} = \\{\\text {R e s N e t 1 8}, \\dots , \\text {S h u f f l e N e t V 2 - 0 . 5} \\}. \\tag {7}", + "image_path": "8d2855c46a42cfc162dc04e02c89960e83df91713cdb00e78e5993edd01d020a.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 106, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 106, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 47, + 106, + 287, + 144 + ], + "type": "text", + "content": "To maintain backbone consistency in both the data synthesis and soft label generation phases, we introduce a backbone ensemble during soft label generation:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 75, + 152, + 287, + 211 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 152, + 287, + 211 + ], + "spans": [ + { + "bbox": [ + 75, + 152, + 287, + 211 + ], + "type": "interline_equation", + "content": "\\tilde {z} = \\left\\{ \\begin{array}{c c} \\sum_ {f _ {\\text {c a n d}} \\in \\mathbb {S}} \\frac {\\sum_ {g \\in \\mathbb {S}} | | g (\\tilde {X}) | | _ {\\mathrm {F}}}{| \\mathbb {S} | | | f _ {\\text {c a n d}} (\\tilde {X}) | | _ {\\mathrm {F}}} \\frac {f _ {\\text {c a n d}} (\\tilde {X})}{| \\mathbb {S} |}, & \\mathrm {w / L N}, \\\\ \\sum_ {f _ {\\text {c a n d}} \\in \\mathbb {S}} \\frac {f _ {\\text {c a n d}} (\\tilde {X})}{| \\mathbb {S} |}, & \\mathrm {w / o L N}, \\end{array} \\right. \\tag {8}", + "image_path": "be53d5f7387a51e256ce9e921ebd4ddd2373cf2a6fef23dabd4a6cb6e8148270.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 219, + 289, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 219, + 289, + 340 + ], + "spans": [ + { + "bbox": [ + 46, + 219, + 289, + 340 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 219, + 289, + 340 + ], + "type": "inline_equation", + "content": "||\\cdot||_{\\mathrm{F}}" + }, + { + "bbox": [ + 46, + 219, + 289, + 340 + ], + "type": "text", + "content": " and LN refer to the Frobenius norm and Logit Normalization, respectively. Thus, we can obtain the soft label by " + }, + { + "bbox": [ + 46, + 219, + 289, + 340 + ], + "type": "inline_equation", + "content": "\\tilde{y} = \\text{softmax}(\\tilde{z}/\\tau)" + }, + { + "bbox": [ + 46, + 219, + 289, + 340 + ], + "type": "text", + "content": ". Particularly, the use of LN is optional, as demonstrated by our ablation studies; it is beneficial for ResNet18 but not for ResNet50 or ResNet101. Moreover, it's important to highlight that we apply a parallel mechanism for soft label generation since it provides a significantly lower computational cost, less than 1/30 of that required for data synthesis, thus making the computational overhead negligible." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 347, + 224, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 347, + 224, + 361 + ], + "spans": [ + { + "bbox": [ + 47, + 347, + 224, + 361 + ], + "type": "text", + "content": "3.3. Generalized Statistical Matching" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 366, + 289, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 366, + 289, + 547 + ], + "spans": [ + { + "bbox": [ + 46, + 366, + 289, + 547 + ], + "type": "text", + "content": "Only ensuring backbone diversity is insufficient to fully exploit the potential of \"generalized matching\". In this subsection, we aim to introduce additional statistics for matching during the data synthesis phase. Prior methods [33, 34] only utilize BatchNorm since the presence of global information statistics within the pre-trained model, with no apparent solution ready for other layers (e.g., Convolution). Retraining the model is the simplest way to address this problem, followed by updating the other layers' parameters through EMA. This approach is obviously impractical, as it defeats the purpose of using gradient descent. By contrast, we propose to allow the pre-trained model " + }, + { + "bbox": [ + 46, + 366, + 289, + 547 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{eval}}(\\cdot)" + }, + { + "bbox": [ + 46, + 366, + 289, + 547 + ], + "type": "text", + "content": " to loop through the training dataset " + }, + { + "bbox": [ + 46, + 366, + 289, + 547 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 46, + 366, + 289, + 547 + ], + "type": "text", + "content": " once without calculating the gradient to obtain the global statistics of Convolution, thus serving \"local-match-global\":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 77, + 552, + 287, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 552, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 77, + 552, + 287, + 586 + ], + "type": "interline_equation", + "content": "\\mathbf {C o n v} _ {l} ^ {\\mathbf {C M}} = \\frac {1}{| \\mathcal {T} |} \\sum_ {i} ^ {| \\mathcal {T} |} \\mathbf {C M} _ {i} ^ {l}, \\mathbf {C o n v} _ {l} ^ {\\mathbf {C V}} = \\frac {1}{| \\mathcal {T} |} \\sum_ {i} ^ {| \\mathcal {T} |} \\mathbf {C V} _ {i} ^ {l}, \\tag {9}", + "image_path": "25502165645d07341cf87aea0e2d4dd11673e4b5d624912e296448a998d584fb.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 78, + 583, + 252, + 611 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 583, + 252, + 611 + ], + "spans": [ + { + "bbox": [ + 78, + 583, + 252, + 611 + ], + "type": "interline_equation", + "content": "\\mathbf {C o n v} _ {l} ^ {\\mathbf {P M}} = \\frac {1}{| \\mathcal {T} |} \\sum_ {i} ^ {| \\mathcal {T} |} \\mathbf {P M} _ {i} ^ {l}, \\mathbf {C o n v} _ {l} ^ {\\mathbf {P V}} = \\frac {1}{| \\mathcal {T} |} \\sum_ {i} ^ {| \\mathcal {T} |} \\mathbf {P V} _ {i} ^ {l}.", + "image_path": "0b10a2d0e12a52f8f6d97ccdfc608ecd1daceb6572c00d8eca6619c14ca4f490.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 616, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 616, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 616, + 289, + 715 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 47, + 616, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{CM}_i^l\\in \\mathbb{R}^{C_l}" + }, + { + "bbox": [ + 47, + 616, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{CV}_i^l\\in \\mathbb{R}^{C_l}" + }, + { + "bbox": [ + 47, + 616, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{PM}_i^l\\in \\mathbb{R}^{\\left[\\frac{H}{N_l^p}\\right]\\times \\left[\\frac{W}{N_l^p}\\right]}" + }, + { + "bbox": [ + 47, + 616, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 616, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{PV}_i^l\\in \\mathbb{R}^{\\left[\\frac{H}{N_l^p}\\right]\\times \\left[\\frac{W}{N_l^p}\\right]}" + }, + { + "bbox": [ + 47, + 616, + 289, + 715 + ], + "type": "text", + "content": " refer to the channel mean, the channel variance, the patch mean and the patch variance, respectively, for the " + }, + { + "bbox": [ + 47, + 616, + 289, + 715 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 47, + 616, + 289, + 715 + ], + "type": "text", + "content": " -th Convolution when the input to " + }, + { + "bbox": [ + 47, + 616, + 289, + 715 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{cand}}(\\cdot)" + }, + { + "bbox": [ + 47, + 616, + 289, + 715 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 47, + 616, + 289, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 616, + 289, + 715 + ], + "type": "text", + "content": " -th batch, where " + }, + { + "bbox": [ + 47, + 616, + 289, + 715 + ], + "type": "inline_equation", + "content": "C_l" + }, + { + "bbox": [ + 47, + 616, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 616, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\left[\\frac{H}{N_l^p}\\right]\\times \\left[\\frac{W}{N_l^p}\\right]" + }, + { + "bbox": [ + 47, + 616, + 289, + 715 + ], + "type": "text", + "content": " denote the number of channels and patches of the " + }, + { + "bbox": [ + 47, + 616, + 289, + 715 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 47, + 616, + 289, + 715 + ], + "type": "text", + "content": " -th Convolution, respectively. We define " + }, + { + "bbox": [ + 47, + 616, + 289, + 715 + ], + "type": "inline_equation", + "content": "N_{l}^{p}" + }, + { + "bbox": [ + 47, + 616, + 289, + 715 + ], + "type": "text", + "content": " as 4, 4, 4 and 16 by default" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 159 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 159 + ], + "type": "text", + "content": "on CIFAR-10, CIFAR-100, Tiny-ImageNet and ImageNet-1k, respectively. After obtaining the global channel mean " + }, + { + "bbox": [ + 304, + 72, + 545, + 159 + ], + "type": "inline_equation", + "content": "\\mathbf{Conv}_l^{\\mathbf{CM}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 159 + ], + "type": "text", + "content": ", the global channel variance " + }, + { + "bbox": [ + 304, + 72, + 545, + 159 + ], + "type": "inline_equation", + "content": "\\mathbf{Conv}_l^{\\mathbf{CV}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 159 + ], + "type": "text", + "content": ", the global patch mean " + }, + { + "bbox": [ + 304, + 72, + 545, + 159 + ], + "type": "inline_equation", + "content": "\\mathbf{Conv}_l^{\\mathbf{PM}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 159 + ], + "type": "text", + "content": " and the global patch variance " + }, + { + "bbox": [ + 304, + 72, + 545, + 159 + ], + "type": "inline_equation", + "content": "\\mathbf{Conv}_l^{\\mathbf{PV}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 159 + ], + "type": "text", + "content": ", we can store them in a disk thus avoiding secondary calculations. In the data synthesis phase, we introduce " + }, + { + "bbox": [ + 304, + 72, + 545, + 159 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Conv}}'(\\tilde{X})" + }, + { + "bbox": [ + 304, + 72, + 545, + 159 + ], + "type": "text", + "content": " in Eq. 10 to accomplish joint matching with " + }, + { + "bbox": [ + 304, + 72, + 545, + 159 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{BN}}'(\\tilde{X})" + }, + { + "bbox": [ + 304, + 72, + 545, + 159 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 165, + 545, + 239 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 165, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 306, + 165, + 545, + 239 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {C o n v}} ^ {\\prime} (\\tilde {X}) = \\sum_ {l} | | \\mu_ {l} ^ {c} (\\tilde {X}) - \\mathbf {C o n v} _ {l} ^ {\\mathrm {C M}} - \\operatorname {s t o p - g r a d} \\left(\\mu_ {l} ^ {c} (\\tilde {X}) - \\mu_ {l} ^ {c, \\text {t o t a l}}\\right) | | _ {2} \\\\ + \\left\\| \\sigma_ {l} ^ {c, 2} (\\tilde {X}) - \\mathbf {C o n v} _ {l} ^ {\\mathrm {C V}} - \\operatorname {s t o p - g r a d} \\left(\\sigma_ {l} ^ {c, 2} (\\tilde {X}) - \\sigma_ {l} ^ {c, 2, \\text {t o t a l}}\\right) \\right\\| _ {2}, \\\\ + \\left\\| \\mu_ {l} ^ {p} (\\tilde {X}) - \\mathbf {C o n v} _ {l} ^ {\\mathrm {P M}} - \\operatorname {s t o p - g r a d} \\left(\\mu_ {l} ^ {p} (\\tilde {X}) - \\mu_ {l} ^ {p, \\text {t o t a l}}\\right) \\right\\| _ {2} \\\\ + \\left\\| \\sigma_ {l} ^ {p, 2} (\\tilde {X}) - \\mathbf {C o n v} _ {l} ^ {\\mathrm {P V}} - \\operatorname {s t o p} - \\operatorname {g r a d} \\left(\\sigma_ {l} ^ {p, 2} (\\tilde {X}) - \\sigma_ {l} ^ {p, 2, \\text {t o t a l}}\\right) \\right\\| _ {2}, \\tag {10} \\\\ \\end{array}", + "image_path": "d899f488d035c5ad60e84bf519be7ceb8f2231a0b52c887a1f2fc088b0df6f05.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 239, + 545, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 239, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 304, + 239, + 545, + 339 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 239, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\mu_l^{c,\\mathrm{total}}" + }, + { + "bbox": [ + 304, + 239, + 545, + 339 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 239, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\sigma_l^{c,2,\\mathrm{total}}" + }, + { + "bbox": [ + 304, + 239, + 545, + 339 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 239, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\mu_l^{p,\\mathrm{total}}" + }, + { + "bbox": [ + 304, + 239, + 545, + 339 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 239, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\sigma_l^{p,2,\\mathrm{total}}" + }, + { + "bbox": [ + 304, + 239, + 545, + 339 + ], + "type": "text", + "content": " are each updated from the channel mean " + }, + { + "bbox": [ + 304, + 239, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\mu_l^c (\\tilde{X})" + }, + { + "bbox": [ + 304, + 239, + 545, + 339 + ], + "type": "text", + "content": ", channel variance " + }, + { + "bbox": [ + 304, + 239, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\sigma_l^{c,2}(\\tilde{X})" + }, + { + "bbox": [ + 304, + 239, + 545, + 339 + ], + "type": "text", + "content": ", patch mean " + }, + { + "bbox": [ + 304, + 239, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\mu_l^p (\\tilde{X})" + }, + { + "bbox": [ + 304, + 239, + 545, + 339 + ], + "type": "text", + "content": ", and patch variance " + }, + { + "bbox": [ + 304, + 239, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\sigma_l^{p,2}(\\tilde{X})" + }, + { + "bbox": [ + 304, + 239, + 545, + 339 + ], + "type": "text", + "content": " respectively, all obtained via EMA from the current batch. In experiments, we discovered that Eq. 10 causes a slightly larger computational burden, so we randomly drop the matching of statistics with a probability of " + }, + { + "bbox": [ + 304, + 239, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\beta_{\\mathrm{dr}}" + }, + { + "bbox": [ + 304, + 239, + 545, + 339 + ], + "type": "text", + "content": " to ensure the efficiency of GSM." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 342, + 546, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 546, + 559 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 546, + 559 + ], + "type": "text", + "content": "Loss Function in the Evaluation Phase. Unlike DD, GBM and GSM are designed to create a distilled dataset that is enriched with information. Here, we introduce an enhancement to the loss function tailored specifically for the evaluation phase. Essentially, the evaluation phase is a knowledge distillation framework for transferring knowledge from a pre-trained model to the evaluation model. SRe2L utilizes " + }, + { + "bbox": [ + 304, + 342, + 546, + 559 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{KL}}(\\tilde{y} || \\mathrm{softmax}(f_{\\mathrm{eval}}(\\tilde{X}) / \\tau))" + }, + { + "bbox": [ + 304, + 342, + 546, + 559 + ], + "type": "text", + "content": " as the loss function and experimentally illustrates that it improves performance by roughly " + }, + { + "bbox": [ + 304, + 342, + 546, + 559 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 342, + 546, + 559 + ], + "type": "text", + "content": ". As established in SRe2L, an increase in temperature " + }, + { + "bbox": [ + 304, + 342, + 546, + 559 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 342, + 546, + 559 + ], + "type": "text", + "content": " correlates with enhanced performance of the evaluation model. Inspired by this and " + }, + { + "bbox": [ + 304, + 342, + 546, + 559 + ], + "type": "inline_equation", + "content": "\\tau^2 D_{\\mathrm{KL}}(\\mathrm{softmax}(p) / \\tau ||\\mathrm{softmax}(q) / \\tau)" + }, + { + "bbox": [ + 304, + 342, + 546, + 559 + ], + "type": "text", + "content": " is equivalent to " + }, + { + "bbox": [ + 304, + 342, + 546, + 559 + ], + "type": "inline_equation", + "content": "\\frac{1}{2C} ||p - q||_2^2" + }, + { + "bbox": [ + 304, + 342, + 546, + 559 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 304, + 342, + 546, + 559 + ], + "type": "inline_equation", + "content": "\\tau \\rightarrow +\\infty" + }, + { + "bbox": [ + 304, + 342, + 546, + 559 + ], + "type": "text", + "content": " [10], we introduce a novel loss function " + }, + { + "bbox": [ + 304, + 342, + 546, + 559 + ], + "type": "inline_equation", + "content": "\\mathrm{MSE} + \\gamma \\times \\mathrm{GT}" + }, + { + "bbox": [ + 304, + 342, + 546, + 559 + ], + "type": "text", + "content": " to avoid numerical error caused by the large " + }, + { + "bbox": [ + 304, + 342, + 546, + 559 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 342, + 546, + 559 + ], + "type": "text", + "content": " and improve the generalization of the distilled dataset without any additional overhead (ignore the weights " + }, + { + "bbox": [ + 304, + 342, + 546, + 559 + ], + "type": "inline_equation", + "content": "\\tau^2" + }, + { + "bbox": [ + 304, + 342, + 546, + 559 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 342, + 546, + 559 + ], + "type": "inline_equation", + "content": "\\frac{1}{2C}" + }, + { + "bbox": [ + 304, + 342, + 546, + 559 + ], + "type": "text", + "content": "):" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 563, + 545, + 585 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 563, + 545, + 585 + ], + "spans": [ + { + "bbox": [ + 316, + 563, + 545, + 585 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {e v a l}} (\\tilde {X}, \\tilde {y}, y) = | | f _ {\\text {e v a l}} (\\tilde {X}) - \\tilde {z} | | _ {2} ^ {2} - \\gamma \\mathbf {y} \\log \\left(\\operatorname {s o f t m a x} \\left(f _ {\\text {e v a l}} (\\tilde {X})\\right)\\right), \\tag {11}", + "image_path": "8ff1751682b36fe24c6b8232353141829be1249683770c12f9814de73812ae7c.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 586, + 545, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 586, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 304, + 586, + 545, + 645 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 586, + 545, + 645 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 586, + 545, + 645 + ], + "type": "text", + "content": " represents the one-hot encoding (w.r.t. the ground truth label " + }, + { + "bbox": [ + 304, + 586, + 545, + 645 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 586, + 545, + 645 + ], + "type": "text", + "content": "). As illustrated in Fig. 5, simply replacing the loss function with " + }, + { + "bbox": [ + 304, + 586, + 545, + 645 + ], + "type": "inline_equation", + "content": "\\mathrm{MSE} + 0.1 \\times \\mathrm{GT}" + }, + { + "bbox": [ + 304, + 586, + 545, + 645 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 586, + 545, + 645 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 586, + 545, + 645 + ], + "type": "text", + "content": " is set as 0.1) in SRe2L improves the performance of ResNet18 (i.e., the evaluation model) by a margin of " + }, + { + "bbox": [ + 304, + 586, + 545, + 645 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 304, + 586, + 545, + 645 + ], + "type": "text", + "content": " on ImageNet-1k under IPC 10." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 657, + 382, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 657, + 382, + 671 + ], + "spans": [ + { + "bbox": [ + 306, + 657, + 382, + 671 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": "We perform comparison experiments on the large-scale dataset including the full " + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": " ImageNet-1k [18] and the small-scale datasets including Tiny-ImageNet [26]," + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16713" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 70, + 278, + 180 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 278, + 180 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 278, + 180 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 278, + 180 + ], + "type": "image", + "image_path": "494301707fff98f4681695e00c6652ea30dde251ec38a3c20ef2b4bafcc8422c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 188, + 287, + 222 + ], + "lines": [ + { + "bbox": [ + 46, + 188, + 287, + 222 + ], + "spans": [ + { + "bbox": [ + 46, + 188, + 287, + 222 + ], + "type": "text", + "content": "Figure 5. Comparison of the effectiveness and efficiency of G-VBSM components. Among them, \"DD+GBM+GSM " + }, + { + "bbox": [ + 46, + 188, + 287, + 222 + ], + "type": "inline_equation", + "content": "(\\beta_{\\mathrm{dr}} = 0.0) + (\\mathrm{MSE} + \\mathrm{GT})" + }, + { + "bbox": [ + 46, + 188, + 287, + 222 + ], + "type": "text", + "content": " represents the comprehensive G-VBSM." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 232, + 286, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 232, + 286, + 269 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 286, + 269 + ], + "type": "text", + "content": "CIFAR-10/100 [11]. To highlight that our proposed G-VBSM is designed for large-scale datasets, all ablation experiments are performed on ImageNet-1k." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 283, + 287, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 283, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 46, + 283, + 287, + 499 + ], + "type": "text", + "content": "Hyperparameter Settings. We prioritize selecting various convolution-based backbone, ensuring the maximal difference in architecture, while also adhering to the criterion of minimizing the number of parameters, an approach empirically demonstrated to be superior [34, 38]. On ImageNet-1k, we skip the model pre-training phase by directly using Torchvision's open source pretraining weights [16] of {ResNet18 [5], MobileNetV2 [21], EfficientNet-B0 [25], ShuffleNetV2-0.5 [39]}. For the remaining dataset, we all train {128-width ConvNet [39], WRN-16-2 [37], ResNet18 [5], ShuffleNetV2-0.5 [39], MobileNetV2-0.5 [21]} from scratch with few epoch in the model pre-training phases. Gray cells in all tables represent the highest performance. Meanwhile, (R18), (R50), etc. in all tables, represent the evaluation models. More details about the remaining method hyperparameter settings and additional ablation studies can be found in Appendix A and Appendix C & D, respectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 510, + 147, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 510, + 147, + 521 + ], + "spans": [ + { + "bbox": [ + 47, + 510, + 147, + 521 + ], + "type": "text", + "content": "4.1. Ablation Studies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 529, + 286, + 721 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 529, + 286, + 721 + ], + "spans": [ + { + "bbox": [ + 46, + 529, + 286, + 721 + ], + "type": "text", + "content": "The Trade-off between Efficiency and Effectiveness. The performance improvements and increased GPU latency for components of G-VBSM in comparison to the baseline SRe2L, are illustrated in Fig. 5. Clearly, DD, GBM, and GSM are highly effective in enhancing the generalization of distilled datasets. Compared with GSM, both DD and GBM are slightly lightweight and efficient. With DD and GBM alone - i.e., " + }, + { + "bbox": [ + 46, + 529, + 286, + 721 + ], + "type": "inline_equation", + "content": "\\mathrm{DD + GBM + GSM}" + }, + { + "bbox": [ + 46, + 529, + 286, + 721 + ], + "type": "inline_equation", + "content": "(\\beta_{\\mathrm{dr}} = 1.0) + (\\mathrm{MSE} + \\mathrm{GT})" + }, + { + "bbox": [ + 46, + 529, + 286, + 721 + ], + "type": "text", + "content": " - the accuracy of this approach surpasses that of SRe2L by a significant margin of " + }, + { + "bbox": [ + 46, + 529, + 286, + 721 + ], + "type": "inline_equation", + "content": "6.5\\%" + }, + { + "bbox": [ + 46, + 529, + 286, + 721 + ], + "type": "text", + "content": ". Additionally, the comprehensive G-VBSM, " + }, + { + "bbox": [ + 46, + 529, + 286, + 721 + ], + "type": "inline_equation", + "content": "\\mathrm{DD + GBM + GSM}" + }, + { + "bbox": [ + 46, + 529, + 286, + 721 + ], + "type": "inline_equation", + "content": "(\\beta_{\\mathrm{dr}} = 0.0) + (\\mathrm{MSE} + \\mathrm{GT})" + }, + { + "bbox": [ + 46, + 529, + 286, + 721 + ], + "type": "text", + "content": ", further enhances the performance of " + }, + { + "bbox": [ + 46, + 529, + 286, + 721 + ], + "type": "inline_equation", + "content": "\\mathrm{DD + GBM + GSM}" + }, + { + "bbox": [ + 46, + 529, + 286, + 721 + ], + "type": "inline_equation", + "content": "(\\beta_{\\mathrm{dr}} = 1.0) + (\\mathrm{MSE} + \\mathrm{GT})" + }, + { + "bbox": [ + 46, + 529, + 286, + 721 + ], + "type": "text", + "content": " by a notable margin of " + }, + { + "bbox": [ + 46, + 529, + 286, + 721 + ], + "type": "inline_equation", + "content": "3.6\\%" + }, + { + "bbox": [ + 46, + 529, + 286, + 721 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 46, + 529, + 286, + 721 + ], + "type": "inline_equation", + "content": "\\mathrm{MSE} + \\gamma \\times \\mathrm{GT}" + }, + { + "bbox": [ + 46, + 529, + 286, + 721 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 529, + 286, + 721 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 529, + 286, + 721 + ], + "type": "text", + "content": " is set as 0.1) is also extremely critical, for SRe2L and " + }, + { + "bbox": [ + 46, + 529, + 286, + 721 + ], + "type": "inline_equation", + "content": "\\mathrm{DD + GBM}" + }, + { + "bbox": [ + 46, + 529, + 286, + 721 + ], + "type": "text", + "content": " boosting " + }, + { + "bbox": [ + 46, + 529, + 286, + 721 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 46, + 529, + 286, + 721 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 529, + 286, + 721 + ], + "type": "inline_equation", + "content": "2.7\\%" + }, + { + "bbox": [ + 46, + 529, + 286, + 721 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 318, + 70, + 531, + 121 + ], + "blocks": [ + { + "bbox": [ + 318, + 70, + 531, + 121 + ], + "lines": [ + { + "bbox": [ + 318, + 70, + 531, + 121 + ], + "spans": [ + { + "bbox": [ + 318, + 70, + 531, + 121 + ], + "type": "table", + "html": "
α\\Evaluation ModelResNet18(MSE+0.1×GT)ResNet50(MSE+0.1×GT)
0.030.4%31.9%
0.431.4%34.5%
0.831.9%36.4%
", + "image_path": "135f72913afabfb5185927415dfe8fe86c91c7599d6e29763608476c37d00384.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 152, + 545, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 152, + 545, + 272 + ], + "spans": [ + { + "bbox": [ + 305, + 152, + 545, + 272 + ], + "type": "text", + "content": "Matching in the Form of SDS. Benefiting from this matching form, G-VBSM can achieve the best performance with small batch sizes (e.g., 40 in ImageNet-1k and 50 in CIFAR-100) in different settings, as demonstrated in our comparative experiments. Table 1 demonstrates the impact of the factor " + }, + { + "bbox": [ + 305, + 152, + 545, + 272 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 305, + 152, + 545, + 272 + ], + "type": "text", + "content": " for performing the EMA update on the final performance achieved by G-VBSM, where " + }, + { + "bbox": [ + 305, + 152, + 545, + 272 + ], + "type": "inline_equation", + "content": "\\alpha = 0" + }, + { + "bbox": [ + 305, + 152, + 545, + 272 + ], + "type": "text", + "content": " indicates that " + }, + { + "bbox": [ + 305, + 152, + 545, + 272 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{BN}}^{\\prime}" + }, + { + "bbox": [ + 305, + 152, + 545, + 272 + ], + "type": "text", + "content": " degenerates to " + }, + { + "bbox": [ + 305, + 152, + 545, + 272 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{BN}}" + }, + { + "bbox": [ + 305, + 152, + 545, + 272 + ], + "type": "text", + "content": ". We can conclude that the SDS matching form is feasible and the generalization of the distilled dataset improves with increasing " + }, + { + "bbox": [ + 305, + 152, + 545, + 272 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 305, + 152, + 545, + 272 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 306, + 282, + 552, + 328 + ], + "blocks": [ + { + "bbox": [ + 306, + 135, + 544, + 147 + ], + "lines": [ + { + "bbox": [ + 306, + 135, + 544, + 147 + ], + "spans": [ + { + "bbox": [ + 306, + 135, + 544, + 147 + ], + "type": "text", + "content": "Table 1. ImageNet-1k Top-1 Acc. about different " + }, + { + "bbox": [ + 306, + 135, + 544, + 147 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 306, + 135, + 544, + 147 + ], + "type": "text", + "content": " under IPC 10." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 282, + 552, + 328 + ], + "lines": [ + { + "bbox": [ + 306, + 282, + 552, + 328 + ], + "spans": [ + { + "bbox": [ + 306, + 282, + 552, + 328 + ], + "type": "table", + "html": "
MethodEvaluation Model
DeiT-TinyResNet18MobileNetV2Swin-Tiny
SRe2L15.41%46.79%36.59%39.23%
G-VBSM29.43%51.82%48.66%57.40%
", + "image_path": "a654e3d0ca27600ca12fbbae6457ab5db7352055b3978d5017e2ae58b291461b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 364, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 364, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 305, + 364, + 545, + 567 + ], + "type": "text", + "content": "Cross-Architecture Generalization. The evaluation of cross-architecture generalization is crucial in assessing the quality of distilled datasets. Unlike traditional methods [1, 15, 20] which focus on CIFAR-100, our approach evaluates the effectiveness of the distilled dataset on ImageNet1k, employing a suite of models with real-world applicability, including ResNet18 [5], MobileNetV2 [21], DeiTiny [28], and Swin-Tiny [13]. The experimental results are reported in Table 2. From Tables 7 and 2, it is evident that the G-VBSM-synthetic dataset can effectively generalize across ResNet {18, 50, 101}, MobileNetV2, DeiTiny, and Swin-Tiny architectures. Notably, DeiT-Tiny and Swin-Tiny, two architectures not encountered during the data synthesis phase, demonstrates significant proficiency with the accuracy " + }, + { + "bbox": [ + 305, + 364, + 545, + 567 + ], + "type": "inline_equation", + "content": "29.43\\%" + }, + { + "bbox": [ + 305, + 364, + 545, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 364, + 545, + 567 + ], + "type": "inline_equation", + "content": "57.40\\%" + }, + { + "bbox": [ + 305, + 364, + 545, + 567 + ], + "type": "text", + "content": ", outperforming the latest SOTA SRe2L by a margin of " + }, + { + "bbox": [ + 305, + 364, + 545, + 567 + ], + "type": "inline_equation", + "content": "14.01\\%" + }, + { + "bbox": [ + 305, + 364, + 545, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 364, + 545, + 567 + ], + "type": "inline_equation", + "content": "18.17\\%" + }, + { + "bbox": [ + 305, + 364, + 545, + 567 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 306, + 578, + 552, + 632 + ], + "blocks": [ + { + "bbox": [ + 305, + 336, + 544, + 357 + ], + "lines": [ + { + "bbox": [ + 305, + 336, + 544, + 357 + ], + "spans": [ + { + "bbox": [ + 305, + 336, + 544, + 357 + ], + "type": "text", + "content": "Table 2. ImageNet-1k Top-1 Acc. on cross-architecture generalization under IPC 50." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 578, + 552, + 632 + ], + "lines": [ + { + "bbox": [ + 306, + 578, + 552, + 632 + ], + "spans": [ + { + "bbox": [ + 306, + 578, + 552, + 632 + ], + "type": "table", + "html": "
Candidates (Backbone)Evaluation Model
ResNet18MobileNetV2EfficientNet-B0ShuffleNetV2-0.5ResNet18ResNet50
✓25.7%30.2%
✓✓27.2%31.2%
✓✓✓27.9%32.0%
✓✓✓✓31.4%34.5%
", + "image_path": "1efe11730878152534f4e8a44d0188f5d5b537c7c47be4cd50ee7c5426c11112.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 640, + 544, + 662 + ], + "lines": [ + { + "bbox": [ + 305, + 640, + 544, + 662 + ], + "spans": [ + { + "bbox": [ + 305, + 640, + 544, + 662 + ], + "type": "text", + "content": "Table 3. ImageNet-1k Top-1 Acc. about the number of candidate backbone in the soft label generation pipeline under IPC 10." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "content": "Ensemble in Soft Label Generation. In knowledge distillation, soft labels obtained through multiple teacher ensembles can effectively enhance the generalization of the student [35]. This observation is similarly corroborated in" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16714" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 123, + 70, + 225, + 180 + ], + "blocks": [ + { + "bbox": [ + 123, + 70, + 225, + 180 + ], + "lines": [ + { + "bbox": [ + 123, + 70, + 225, + 180 + ], + "spans": [ + { + "bbox": [ + 123, + 70, + 225, + 180 + ], + "type": "image", + "image_path": "8f339e50e538ee9c08e709dd0e54907fcf5e2e0327ae27d36a7bb7963b97d5c0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 124, + 184, + 468, + 196 + ], + "lines": [ + { + "bbox": [ + 124, + 184, + 468, + 196 + ], + "spans": [ + { + "bbox": [ + 124, + 184, + 468, + 196 + ], + "type": "text", + "content": "Figure 6. ImageNet-1k Top-1 Acc. of different loss function, " + }, + { + "bbox": [ + 124, + 184, + 468, + 196 + ], + "type": "inline_equation", + "content": "\\beta_{\\mathrm{dr}}" + }, + { + "bbox": [ + 124, + 184, + 468, + 196 + ], + "type": "text", + "content": " and batch sizes under IPC 10." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 241, + 70, + 347, + 180 + ], + "blocks": [ + { + "bbox": [ + 241, + 70, + 347, + 180 + ], + "lines": [ + { + "bbox": [ + 241, + 70, + 347, + 180 + ], + "spans": [ + { + "bbox": [ + 241, + 70, + 347, + 180 + ], + "type": "image", + "image_path": "0aee72a94832890741a2abcdbdaa4367f85762f86a1795342ab35c3a0fd03085.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 369, + 71, + 470, + 178 + ], + "blocks": [ + { + "bbox": [ + 369, + 71, + 470, + 178 + ], + "lines": [ + { + "bbox": [ + 369, + 71, + 470, + 178 + ], + "spans": [ + { + "bbox": [ + 369, + 71, + 470, + 178 + ], + "type": "image", + "image_path": "16d36d0cbaa7310f4460a46b42cdd24c90c5548c1d09d61236e2af8ccf7a2187.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 49, + 201, + 545, + 258 + ], + "blocks": [ + { + "bbox": [ + 49, + 201, + 545, + 258 + ], + "lines": [ + { + "bbox": [ + 49, + 201, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 49, + 201, + 545, + 258 + ], + "type": "table", + "html": "
DatasetIPCMTT [1] (CW128)DataDAM [20] (CW128)TESLA [2] (R18)SRe2L (R18)SRe2L (R50)SRe2L (R101)G-VBSM (R18)G-VBSM (R50)G-VBSM (R101)
Tiny-ImageNet5028.0±0.328.7±0.3-41.1±0.442.2±0.542.5±0.247.6±0.348.7±0.248.8±0.4
100---49.7±0.351.2±0.451.5±0.351.0±0.452.1±0.352.3±0.1
ImageNet-1k1064.0±1.3†6.3±0.07.7±0.121.3±0.628.4±0.130.9±0.131.4±0.535.4±0.838.2±0.4
50-15.5±0.2-46.8±0.255.6±0.360.8±0.551.8±0.458.7±0.361.0±0.4
100---52.8±0.361.0±0.462.8±0.255.7±0.462.2±0.363.7±0.2
", + "image_path": "e63f6b6e087433123e662cb744ee629324b629f71a5cb41e2c6ba767b2fab48a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 49, + 292, + 545, + 358 + ], + "blocks": [ + { + "bbox": [ + 47, + 265, + 546, + 288 + ], + "lines": [ + { + "bbox": [ + 47, + 265, + 546, + 288 + ], + "spans": [ + { + "bbox": [ + 47, + 265, + 546, + 288 + ], + "type": "text", + "content": "Table 4. Comparison with baseline models in Tiny-ImageNet and ImageNet-1k. " + }, + { + "bbox": [ + 47, + 265, + 546, + 288 + ], + "type": "inline_equation", + "content": "{}^{ \\dagger }" + }, + { + "bbox": [ + 47, + 265, + 546, + 288 + ], + "type": "text", + "content": " indicates the ImageNette dataset,which contains only 10 classes. DataDAM [20] and TESLA [2] use the downsampled " + }, + { + "bbox": [ + 47, + 265, + 546, + 288 + ], + "type": "inline_equation", + "content": "{64} \\times {64}" + }, + { + "bbox": [ + 47, + 265, + 546, + 288 + ], + "type": "text", + "content": " ImageNet-1k. We cite the experimental results from SRe2L [34]." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 292, + 545, + 358 + ], + "lines": [ + { + "bbox": [ + 49, + 292, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 49, + 292, + 545, + 358 + ], + "type": "table", + "html": "
DatasetIPCCoreset SelectionTraining Set Synthesis (CW128)Training Set Synthesis (R18)Whole Dataset
RandomHerdingK-CenterForgettingDC [42]DM [41]CAFE [31]KIP [15]MTT [1]DataDAM[20]G-VBSMSRe2LG-VBSM(CW128)
CIFAR-101026.0±1.231.6±0.714.7±0.923.3±1.044.9±0.548.9±0.650.9±0.546.1±0.765.3±0.754.2±0.846.5±0.727.2±0.553.5±0.684.8±0.1
5043.4±1.040.4±0.627.0±1.423.3±1.153.9±0.563.0±0.462.3±0.453.2±0.771.6±0.267.0±0.454.3±0.347.5±0.659.2±0.4
CIFAR-10014.2±0.38.3±0.38.4±0.34.5±0.212.8±0.311.4±0.314.0±0.312.0±0.224.3±0.314.5±0.516.4±0.72.0±0.225.9±0.556.2±0.3
1014.6±0.517.3±0.317.3±0.315.1±0.325.2±0.329.7±0.331.5±0.240.1±0.433.1±0.434.8±0.538.7±0.231.6±0.559.5±0.4
5030.0±0.433.7±0.530.5±0.3-30.6±0.643.6±0.447.7±0.2-42.9±0.349.4±0.345.7±0.449.5±0.365.0±0.5
", + "image_path": "33335728a087809a15de3ad8ac0dde668bf21efb9773d0db8e8177afd7e5ba14.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 405, + 288, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 405, + 288, + 501 + ], + "spans": [ + { + "bbox": [ + 46, + 405, + 288, + 501 + ], + "type": "text", + "content": "dataset distillation, as evidenced in Table 3. The greater the number of models, the stronger the generalization of the distilled dataset. It is particularly interesting to point out that ShuffleNetV2-0.5 improves further by more than " + }, + { + "bbox": [ + 46, + 405, + 288, + 501 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 46, + 405, + 288, + 501 + ], + "type": "text", + "content": " even though the other three models have already achieved modest ensemble gains. We attribute this enhancement to the channel shuffle mechanism within ShuffleNetV2-0.5, which imposes a beneficial regularization constraint for G-VBSM." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 48, + 511, + 293, + 555 + ], + "blocks": [ + { + "bbox": [ + 46, + 360, + 547, + 395 + ], + "lines": [ + { + "bbox": [ + 46, + 360, + 547, + 395 + ], + "spans": [ + { + "bbox": [ + 46, + 360, + 547, + 395 + ], + "type": "text", + "content": "Table 5. Comparison with baseline models on CIFAR-10/100. All methods, except for SRe2L and G-VBSM, use a 128-width ConvNet (CW128) for data synthesis and evaluation. G-VBSM utilizes {CW128, WRN-16-2, ResNet18 (R18), ShuffleNetV2-0.5, MobileNetV2-0.5} for data synthesis and {CW128, R18} for evaluation. We cite the experimental results, except for SRe2L's, from DataDAM [20]." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 511, + 293, + 555 + ], + "lines": [ + { + "bbox": [ + 48, + 511, + 293, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 511, + 293, + 555 + ], + "type": "table", + "html": "
Logit NormalizationEvaluation Model
ResNet18 (MSE+0.1×GT)ResNet18 (KL)ResNet50 (MSE+0.1×GT)ResNet50 (KL)
Yes31.4%25.4%34.5%28.6%
No31.0%25.1%35.4%31.9%
", + "image_path": "b186609fc699993eeffa26fca44819c092b0533f171f9989dbe59551d6dc4316.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 563, + 287, + 586 + ], + "lines": [ + { + "bbox": [ + 46, + 563, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 563, + 287, + 586 + ], + "type": "text", + "content": "Table 6. ImageNet-1k Top-1 Acc. about the use of logit normalization under IPC 10." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "content": "Logit Normalization in Soft Label Generation. The aim of this strategy is to maintain the consistency of the logits (can be viewed as vectors) magnitude across all models within a high-dimensional space, thereby ensuring equal contribution of these models to the ultimate soft label for ensemble. This approach is not universally effective, as shown in Table 6. When the evaluation model parameter count is low (e.g., ResNet18), it can enhance G-VBSM performance. Conversely, with a model like ResNet50, it may hinder performance, rendering G-VBSM uncompetitive. As" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 405, + 546, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 405, + 546, + 440 + ], + "spans": [ + { + "bbox": [ + 304, + 405, + 546, + 440 + ], + "type": "text", + "content": "a result, in this work, ResNet " + }, + { + "bbox": [ + 304, + 405, + 546, + 440 + ], + "type": "inline_equation", + "content": "\\{50, 101\\}" + }, + { + "bbox": [ + 304, + 405, + 546, + 440 + ], + "type": "text", + "content": " and ViT-based models do not employ logit normalization, while the remaining models do." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 445, + 547, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 547, + 589 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 547, + 589 + ], + "type": "text", + "content": "Loss Function in the Evaluation Phase. As illustrated in Fig. 6, replacing " + }, + { + "bbox": [ + 304, + 445, + 547, + 589 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{KL}}(\\cdot ||\\cdot)" + }, + { + "bbox": [ + 304, + 445, + 547, + 589 + ], + "type": "text", + "content": " in SRe2L with " + }, + { + "bbox": [ + 304, + 445, + 547, + 589 + ], + "type": "inline_equation", + "content": "\\mathrm{MSE} + \\gamma \\times \\mathrm{GT}" + }, + { + "bbox": [ + 304, + 445, + 547, + 589 + ], + "type": "text", + "content": " demonstrates to be extremely effective. For example, with " + }, + { + "bbox": [ + 304, + 445, + 547, + 589 + ], + "type": "inline_equation", + "content": "\\beta_{\\mathrm{dr}} = 0.0" + }, + { + "bbox": [ + 304, + 445, + 547, + 589 + ], + "type": "text", + "content": " and batch size " + }, + { + "bbox": [ + 304, + 445, + 547, + 589 + ], + "type": "inline_equation", + "content": "= 40" + }, + { + "bbox": [ + 304, + 445, + 547, + 589 + ], + "type": "text", + "content": ", adjusting " + }, + { + "bbox": [ + 304, + 445, + 547, + 589 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 445, + 547, + 589 + ], + "type": "text", + "content": " to 0.01 and 0.1 enhances model performance during the evaluation phase by " + }, + { + "bbox": [ + 304, + 445, + 547, + 589 + ], + "type": "inline_equation", + "content": "1.4\\%" + }, + { + "bbox": [ + 304, + 445, + 547, + 589 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 445, + 547, + 589 + ], + "type": "inline_equation", + "content": "6.0\\%" + }, + { + "bbox": [ + 304, + 445, + 547, + 589 + ], + "type": "text", + "content": ", respectively. As a result, we employ a " + }, + { + "bbox": [ + 304, + 445, + 547, + 589 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 445, + 547, + 589 + ], + "type": "text", + "content": " of 0.1 in all experiments conducted on ImageNet-1k. Furthermore, we observe that the distilled dataset's generalization is not significantly affected (" + }, + { + "bbox": [ + 304, + 445, + 547, + 589 + ], + "type": "inline_equation", + "content": "\\leq 2\\%" + }, + { + "bbox": [ + 304, + 445, + 547, + 589 + ], + "type": "text", + "content": ") by variations in either batch size or " + }, + { + "bbox": [ + 304, + 445, + 547, + 589 + ], + "type": "inline_equation", + "content": "\\beta_{\\mathrm{dr}}" + }, + { + "bbox": [ + 304, + 445, + 547, + 589 + ], + "type": "text", + "content": ". This ensured that G-VBSM attain SOTA results with a minimal batch size 40 and " + }, + { + "bbox": [ + 304, + 445, + 547, + 589 + ], + "type": "inline_equation", + "content": "\\beta_{\\mathrm{dr}} = 0.4" + }, + { + "bbox": [ + 304, + 445, + 547, + 589 + ], + "type": "text", + "content": " (i.e., the default settings in ImageNet-1k)." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 310, + 599, + 538, + 700 + ], + "blocks": [ + { + "bbox": [ + 310, + 599, + 538, + 700 + ], + "lines": [ + { + "bbox": [ + 310, + 599, + 538, + 700 + ], + "spans": [ + { + "bbox": [ + 310, + 599, + 538, + 700 + ], + "type": "image", + "image_path": "1f518e974b524b3100798be7098e0c0c470ae6fb5178780a279feb4d07290cee.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 708, + 538, + 719 + ], + "lines": [ + { + "bbox": [ + 313, + 708, + 538, + 719 + ], + "spans": [ + { + "bbox": [ + 313, + 708, + 538, + 719 + ], + "type": "text", + "content": "Figure 7. Synthetic data visualization on small-scale datasets." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "16715" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "Synthetic Data Visualization. The visualization results for the large-scale (the full " + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": " ImageNet-1k) and small-scale (CIFAR-10/100 and Tiny-ImageNet) datasets are displayed in Figs. 1 and 7, respectively. On ImageNet-1k, the images synthesized by G-VBSM are more informative and abstract compared with those from SRe2L. Furthermore, on small-scale datasets, the distilled images obtained by G-VBSM markedly differ from the images presented in MTT and DataDAM's papers [1, 20], illustrating that G-VBSM is an out-of-the ordinary algorithm. More synthetic data can be found in Appendix I." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 62, + 219, + 276, + 338 + ], + "blocks": [ + { + "bbox": [ + 62, + 219, + 276, + 338 + ], + "lines": [ + { + "bbox": [ + 62, + 219, + 276, + 338 + ], + "spans": [ + { + "bbox": [ + 62, + 219, + 276, + 338 + ], + "type": "image", + "image_path": "d0f0509b50203882be0e13b9935d8f09e5a2cba11aad947d1e4fa7ed3eabde19.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 348, + 253, + 359 + ], + "lines": [ + { + "bbox": [ + 80, + 348, + 253, + 359 + ], + "spans": [ + { + "bbox": [ + 80, + 348, + 253, + 359 + ], + "type": "text", + "content": "Figure 8. t-SNE visualization on ImageNet-1k." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 375, + 287, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 375, + 287, + 507 + ], + "spans": [ + { + "bbox": [ + 46, + 375, + 287, + 507 + ], + "type": "text", + "content": "Logit Embedding Distributions Visualization. We feed distilled datasets into a pre-trained model (i.e., ResNet18) to obtain logit embeddings for t-SNE [29] visualization. Note that unlike previous methods that used feature embeddings before the classifier layer, we utilize logit embeddings after the classifier layer for presentation. Fig. 8 shows the distributions of G-VBSM and SRe2L logit embeddings. The logit embedding obtained by G-VBSM has less intra-class similarity than SRe2L, ensuring softer labels, preventing model overconfidence during the evaluation phase, and ultimately enhancing model generalization." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 515, + 224, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 515, + 224, + 528 + ], + "spans": [ + { + "bbox": [ + 47, + 515, + 224, + 528 + ], + "type": "text", + "content": "4.2. Large-Scale Dataset Comparison" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "text", + "content": "Full " + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "text", + "content": " ImageNet-1k. As illustrated in the ImageNet-1k part of Table 4, G-VBSM consistently outperforms SRe2L across all IPC " + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\{10, 50, 100\\}" + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "text", + "content": ", with the evaluation model being ResNet " + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\{18, 50, 101\\}" + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "text", + "content": ", indicating that G-VBSM is highly competitive in the large-scale dataset ImageNet-1k. In particular, under extreme compression scenarios, such as IPC 10, G-VBSM achieves significant performance increases - " + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "inline_equation", + "content": "10.1\\%" + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "inline_equation", + "content": "7.0\\%" + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "inline_equation", + "content": "7.3\\%" + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "text", + "content": " for ResNet18, ResNet50, and ResNet101, respectively - compared with SRe2L. In comparison to the latest SOTA classical method, DataDAM, just experimenting on the full " + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "text", + "content": " ImageNet-1k, we exceed it by a margin of " + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "inline_equation", + "content": "23.7\\%" + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "text", + "content": " under IPC 10. These results are particularly impressive when considered in the context of the large-scale dataset ImageNet-1k." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 72, + 482, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 482, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 482, + 85 + ], + "type": "text", + "content": "4.3. Small-Scale Dataset Comparison" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 90, + 545, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 90, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 304, + 90, + 545, + 185 + ], + "type": "text", + "content": "Tiny-ImageNet. Similar to SRe2L, we evaluate the generalization of the distilled dataset on ResNet {18, 50, 101}. As presented in the Tiny-ImageNet part of Table 4, G-VBSM consistently surpasses the latest SOTA method SRe2L on different settings. Our method outperforms SRe2L, MTT and DataDAM by " + }, + { + "bbox": [ + 304, + 90, + 545, + 185 + ], + "type": "inline_equation", + "content": "6.5\\%" + }, + { + "bbox": [ + 304, + 90, + 545, + 185 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 90, + 545, + 185 + ], + "type": "inline_equation", + "content": "19.6\\%" + }, + { + "bbox": [ + 304, + 90, + 545, + 185 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 90, + 545, + 185 + ], + "type": "inline_equation", + "content": "18.9\\%" + }, + { + "bbox": [ + 304, + 90, + 545, + 185 + ], + "type": "text", + "content": ", respectively, under IPC 50 applying 128-width ConvNet or ResNet18 as the evaluation model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 194, + 546, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 194, + 546, + 397 + ], + "spans": [ + { + "bbox": [ + 304, + 194, + 546, + 397 + ], + "type": "text", + "content": "CIFAR-100. One of the disadvantages of SRe2L is its inability to compete with SOTA performance on this small-scale dataset benchmark (i.e., evaluated with 128-width ConvNet and ResNet18), such as the performance on CIFAR-100 shown in Table 5, as a result, it fails to demonstrate that \"local-match-global\" (e.g., statistical matching of BatchNorm) can rival or outperform traditional DC algorithms. In contrast, as shown in the CIFAR-100 part of Table 5, G-VBSM achieves accuracies " + }, + { + "bbox": [ + 304, + 194, + 546, + 397 + ], + "type": "inline_equation", + "content": "16.4\\%" + }, + { + "bbox": [ + 304, + 194, + 546, + 397 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 194, + 546, + 397 + ], + "type": "inline_equation", + "content": "38.7\\%" + }, + { + "bbox": [ + 304, + 194, + 546, + 397 + ], + "type": "text", + "content": ", the highest under IPC 1 and 10 on CIFAR-100 through \"generalized matching\", while maintaining the same number of epochs (i.e., 1000) and the model (i.e., 128-width ConvNet) for the evaluation phase as the traditional methods, outperforming the latest SOTA DataDAM by " + }, + { + "bbox": [ + 304, + 194, + 546, + 397 + ], + "type": "inline_equation", + "content": "1.9\\%" + }, + { + "bbox": [ + 304, + 194, + 546, + 397 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 194, + 546, + 397 + ], + "type": "inline_equation", + "content": "3.9\\%" + }, + { + "bbox": [ + 304, + 194, + 546, + 397 + ], + "type": "text", + "content": ", respectively. To the best of our knowledge, our proposed G-VBSM is the first algorithm that is strongly competitive on CIAR-100 as well as the full " + }, + { + "bbox": [ + 304, + 194, + 546, + 397 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 304, + 194, + 546, + 397 + ], + "type": "text", + "content": " ImageNet-1k." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 406, + 545, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 406, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 406, + 545, + 491 + ], + "type": "text", + "content": "CIFAR-10. As illustrated in the CIFAR-10 part of Table 5, G-VBSM outperforms vanilla DC [32], KIP [15], and all coreset selection algorithms on the benchmark with 128-width ConvNet as the evaluation model. Meanwhile, on the benchmark with ResNet 18 as the evaluation model, G-VBSM outperforms the baseline SRe2L by " + }, + { + "bbox": [ + 304, + 406, + 545, + 491 + ], + "type": "inline_equation", + "content": "26.3\\%" + }, + { + "bbox": [ + 304, + 406, + 545, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 406, + 545, + 491 + ], + "type": "inline_equation", + "content": "11.7\\%" + }, + { + "bbox": [ + 304, + 406, + 545, + 491 + ], + "type": "text", + "content": " under IPC 10 and 50, respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 501, + 378, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 501, + 378, + 513 + ], + "spans": [ + { + "bbox": [ + 306, + 501, + 378, + 513 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "text", + "content": "In this paper, we introduce a novel perspective termed \"generalized matching\" for dataset condensation. This posits that an abundance of lightweight \"local-match-global\" matching are more effective than a single \"local-match-global\" matching, and even surpass the precise and costly matching used by traditional methods. Consequently, we present G-VBSM, designed to ensure data densification while performing matching based on sufficient and various backbones, layers, and statistics. Experiments conducted on CIFAR-10/100, Tiny-[ImageNet, and the full " + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "text", + "content": " ImageNet-1k demonstrated that our method is the first algorithm to show strong performance across both small-scale and large-scale datasets. For future research, we plan to extend this approach's applicability to large-scale detection and segmentation datasets." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16716" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] George Cazenavette, Tongzhou Wang, Antonio Torralba, Alexei A. Efros, and Jun-Yan Zhu. Dataset distillation by matching training trajectories. In Computer Vision and Pattern Recognition, New Orleans, LA, USA, 2022. IEEE. 2, 4, 6, 7, 8, 1, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 288, + 192 + ], + "type": "text", + "content": "[2] Justin Cui, Ruochen Wang, Si Si, and Cho-Jui Hsieh. Scaling up dataset distillation to imagenet-1k with constant memory. In International Conference on Machine Learning, pages 6565–6590, Honolulu, Hawaii, USA, 2023. PMLR. 2, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 287, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 287, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 287, + 247 + ], + "type": "text", + "content": "[3] Jiawei Du, Yidi Jiang, Vincent Y. F. Tan, Joey Tianyi Zhou, and Haizhou Li. Minimizing the accumulated trajectory error to improve dataset distillation. In Computer Vision and Pattern Recognition, pages 3749-3758, Vancouver, BC, Canada, 2023. IEEE. 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 248, + 287, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 248, + 287, + 292 + ], + "spans": [ + { + "bbox": [ + 53, + 248, + 287, + 292 + ], + "type": "text", + "content": "[4] Jianping Gou, Baosheng Yu, Stephen J Maybank, and Dacheng Tao. Knowledge distillation: A survey. International Journal of Computer Vision, 129(6):1789-1819, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 294, + 287, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 294, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 53, + 294, + 287, + 337 + ], + "type": "text", + "content": "[5] Kaiming He, Xiangyu Zhang, and Shaoqing Ren. Deep residual learning for image recognition. In Computer Vision and Pattern Recognition, pages 770-778, Las Vegas, NV, USA, 2016. IEEE. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 339, + 287, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 339, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 53, + 339, + 287, + 361 + ], + "type": "text", + "content": "[6] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network, 2015. 3, 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 363, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 363, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 363, + 287, + 407 + ], + "type": "text", + "content": "[7] Shengyuan Hu, Jack Goetz, Kshitiz Malik, Hongyuan Zhan, Zhe Liu, and Yue Liu. Fedsynth: Gradient compression via synthetic data in federated learning. arXiv preprint arXiv:2204.01273, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 408, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 408, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 408, + 287, + 441 + ], + "type": "text", + "content": "[8] Tao Huang, Shan You, Fei Wang, Chen Qian, and Chang Xu. Knowledge distillation from a stronger teacher. In Neural Information Processing Systems, 2022. 3, 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 443, + 287, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 443, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 53, + 443, + 287, + 508 + ], + "type": "text", + "content": "[9] Jang-Hyun Kim, Jinuk Kim, Seong Joon Oh, Sangdoo Yun, Hwanjun Song, Joonhyun Jeong, Jung-Woo Ha, and Hyun Oh Song. Dataset condensation via efficient synthetic-data parameterization. In International Conference on Machine Learning, pages 11102-11118, Baltimore, Maryland, USA, 2022. PMLR. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 510, + 287, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 287, + 564 + ], + "type": "text", + "content": "[10] Taehyeon Kim, Jaehoon Oh, NakYil Kim, Sangwook Cho, and Se-Young Yun. Comparing kullback-leibler divergence and mean squared error loss in knowledge distillation. In International Joint Conference on Artificial Intelligence, Virtual Event, 2021. Morgan Kaufmann. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 566, + 287, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 566, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 48, + 566, + 287, + 589 + ], + "type": "text", + "content": "[11] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 590, + 287, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 632 + ], + "type": "text", + "content": "[12] Yanqing Liu, Jianyang Gu, Kai Wang, Zheng Zhu, Wei Jiang, and Yang You. DREAM: efficient dataset distillation by representative matching. arXiv preprint arXiv:2302.14416, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "text", + "content": "[13] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In International Conference on Computer Vision, pages 10012-10022, 2021. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[14] Wojciech Masarczyk and Ivona Tautkute. Reducing catastrophic forgetting with learning on synthetic data. In Com" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "puter Vision and Pattern Recognition Workshops, pages 252-253, Virtual Event, 2020. IEEE. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 97, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 97, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 308, + 97, + 545, + 129 + ], + "type": "text", + "content": "[15] Timothy Nguyen, Zhourong Chen, and Jaehoon Lee. Dataset meta-learning from kernel ridge-regression. arXiv preprint arXiv:2011.00050, 2020. 2, 6, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 131, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 131, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 308, + 131, + 545, + 196 + ], + "type": "text", + "content": "[16] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In Neural Information Processing Systems, Vancouver, BC, Canada, 2019. 6, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 199, + 545, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 199, + 545, + 242 + ], + "spans": [ + { + "bbox": [ + 308, + 199, + 545, + 242 + ], + "type": "text", + "content": "[17] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In International Conference on Learning Representations, 2023. 4, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 245, + 545, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 245, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 308, + 245, + 545, + 300 + ], + "type": "text", + "content": "[18] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International Journal of Computer Vision, 115(3):211-252, 2015. 2, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 301, + 545, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 301, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 308, + 301, + 545, + 324 + ], + "type": "text", + "content": "[19] Noveen Sachdeva and Julian McAuley. Data distillation: A survey. arXiv preprint arXiv:2301.04272, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 326, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 326, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 308, + 326, + 545, + 380 + ], + "type": "text", + "content": "[20] Ahmad Sajedi, Samir Khaki, Ehsan Amjadian, Lucy Z Liu, Yuri A Lawryshyn, and Konstantinos N Plataniotis. Datadam: Efficient dataset distillation with attention matching. In International Conference on Computer Vision, pages 17097-17107, Paris, France, 2023. IEEE. 3, 6, 7, 8, 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 382, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 382, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 382, + 545, + 437 + ], + "type": "text", + "content": "[21] Mark Sandler, Andrew G. Howard, Menglong Zhu, Andrey Zhmoginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted residuals and linear bottlenecks*. In *Computer Vision and Pattern Recognition*, pages 4510-4520, Salt Lake City, UT, USA, 2018. IEEE. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 439, + 545, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 439, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 308, + 439, + 545, + 483 + ], + "type": "text", + "content": "[22] Mattia Sangermano, Antonio Carta, Andrea Cossu, and Davide Bacciu. Sample condensation in online continual learning. In International Joint Conference on Neural Networks, pages 1-8, Padua, Italy, 2022. IEEE. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 484, + 545, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 484, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 308, + 484, + 545, + 518 + ], + "type": "text", + "content": "[23] Zhiqiang Shen and Eric Xing. A fast knowledge distillation framework for visual recognition. In European Conference on Computer Vision, pages 673-690. Springer, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 519, + 545, + 584 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 519, + 545, + 584 + ], + "spans": [ + { + "bbox": [ + 308, + 519, + 545, + 584 + ], + "type": "text", + "content": "[24] Felipe Petroski Such, Aditya Rawal, Joel Lehman, Kenneth O. Stanley, and Jeffrey Clune. Generative teaching networks: Accelerating neural architecture search by learning to generate synthetic training data. In International Conference on Machine Learning, pages 9206-9216, Virtual Event, 2020. PMLR. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 586, + 545, + 630 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 586, + 545, + 630 + ], + "spans": [ + { + "bbox": [ + 308, + 586, + 545, + 630 + ], + "type": "text", + "content": "[25] Mingxing Tan and Quoc Le. Efficientnetv2: Smaller models and faster training. In International Conference on Machine Learning, pages 10096-10106, Virtual Event, 2021. PMLR. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 632, + 545, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 632, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 308, + 632, + 545, + 655 + ], + "type": "text", + "content": "[26] Amirhossein Tavanaei. Embedded encoder-decoder in convolutional networks towards explainable AI. 2020. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 657, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 657, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 657, + 545, + 689 + ], + "type": "text", + "content": "[27] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive representation distillation. In International Conference on Learning Representations, 2019. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[28] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16717" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357, Virtual Event, 2021. PMLR. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 138 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 138 + ], + "type": "text", + "content": "[29] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-SNE. Journal of Machine Learning Research, 9 (11), 2008. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "type": "text", + "content": "[30] Chunwei Wang, Chao Ma, Ming Zhu, and Xiaokang Yang. Pointaugmenting: Cross-modal augmentation for 3d object detection. In Computer Vision and Pattern Recognition, pages 11794-11803, Virtual Event, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 185, + 287, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 287, + 237 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 287, + 237 + ], + "type": "text", + "content": "[31] Kai Wang, Bo Zhao, Xiangyu Peng, Zheng Zhu, Shuo Yang, Shuo Wang, Guan Huang, Hakan Bilen, Xinchao Wang, and Yang You. Cafe: Learning to condense dataset by aligning features. In Computer Vision and Pattern Recognition, pages 12196-12205, New Orleans, LA, USA, 2022. IEEE. 2, 3, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 239, + 287, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 287, + 271 + ], + "type": "text", + "content": "[32] Tongzhou Wang, Jun-Yan Zhu, Antonio Torralba, and Alexei A Efros. Dataset distillation. arXiv preprint arXiv:1811.10959, 2018. 1, 2, 8, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 273, + 287, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 273, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 48, + 273, + 287, + 327 + ], + "type": "text", + "content": "[33] Hongxu Yin, Pavlo Molchanov, Jose M Alvarez, Zhizhong Li, Arun Mallya, Derek Hoiem, Niraj K Jha, and Jan Kautz. Dreaming to distill: Data-free knowledge transfer via deep-inversion. In Computer Vision and Pattern Recognition, pages 8715-8724, Virtual Event, 2020. IEEE. 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 328, + 287, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 328, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 328, + 287, + 370 + ], + "type": "text", + "content": "[34] Zeyuan Yin, Eric P. Xing, and Zhiqiang Shen. Squeeze, recover and relabel: Dataset condensation at imagenet scale from A new perspective. In Neural Information Processing Systems. NIPS, 2023. 2, 3, 4, 5, 6, 7, 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 373, + 287, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 373, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 48, + 373, + 287, + 414 + ], + "type": "text", + "content": "[35] Shan You, Chang Xu, Chao Xu, and Dacheng Tao. Learning from multiple teacher networks. In International Conference on Knowledge Discovery and Data Mining, pages 1285-1294, Halifax, NS, Canada, 2017. ACM. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 416, + 287, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 416, + 287, + 447 + ], + "spans": [ + { + "bbox": [ + 48, + 416, + 287, + 447 + ], + "type": "text", + "content": "[36] Ruonan Yu, Songhua Liu, and Xinchao Wang. Dataset distillation: A comprehensive review. arXiv preprint arXiv:2301.07014, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 449, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 480 + ], + "type": "text", + "content": "[37] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In British Machine Vision Conference, York, UK, 2016. BMVA Press. 1, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 483, + 287, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 483, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 48, + 483, + 287, + 535 + ], + "type": "text", + "content": "[38] Lei Zhang, Jie Zhang, Bowen Lei, Subhabrata Mukherjee, Xiang Pan, Bo Zhao, Caiwen Ding, Yao Li, and Dongkuan Xu. Accelerating dataset distillation via model augmentation. In Computer Vision and Pattern Recognition, Vancouver, BC, Canada, 2023. IEEE. 6, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 537, + 287, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 287, + 581 + ], + "type": "text", + "content": "[39] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In Computer Vision and Pattern Recognition, pages 6848-6856, 2018. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 582, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 582, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 582, + 287, + 624 + ], + "type": "text", + "content": "[40] Bo Zhao and Hakan Bilen. Dataset condensation with differentiable siamese augmentation. In International Conference on Machine Learning, pages 12674-12685, Virtual Event, 2021. PMLR. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 625, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 667 + ], + "type": "text", + "content": "[41] Bo Zhao and Hakan Bilen. Dataset condensation with distribution matching. In Winter Conference on Applications of Computer Vision, pages 6514-6523, Waikoloa, Hawaii, 2023. IEEE. 2, 7, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[42] Bo Zhao, Konda Reddy Mopuri, and Hakan Bilen. Dataset condensation with gradient matching. In International Conference on Learning Representations, Virtual Event, 2021. OpenReview.net. 2, 7" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 150 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[43] Borui Zhao, Quan Cui, Renjie Song, Yiyu Qiu, and Jiajun Liang. Decoupled knowledge distillation. In Computer Vision and Pattern Recognition, pages 11953-11962, 2022. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 108, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 150 + ], + "type": "text", + "content": "[44] Yongchao Zhou, Ehsan Nezhadarya, and Jimmy Ba. Dataset distillation using neural feature regression. In Neural Information Processing Systems, New Orleans, LA, USA, 2022. NIPS. 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16718" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Generalized Predictive Model for Autonomous Driving/554856ec-712f-446a-b09c-0cce647f2af6_content_list.json b/2024/Generalized Predictive Model for Autonomous Driving/554856ec-712f-446a-b09c-0cce647f2af6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1d45cfab766570484ff51fc8efaa8b55362ad0c1 --- /dev/null +++ b/2024/Generalized Predictive Model for Autonomous Driving/554856ec-712f-446a-b09c-0cce647f2af6_content_list.json @@ -0,0 +1,1517 @@ +[ + { + "type": "text", + "text": "Generalized Predictive Model for Autonomous Driving", + "text_level": 1, + "bbox": [ + 207, + 130, + 763, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiazhi Yang $^{1*}$ Shenyuan Gao $^{2,1*}$ Yihang Qiu $^{1*}$ Li Chen $^{3,1\\dagger}$ Tianyu Li $^{1}$ Bo Dai $^{1}$ Kashyap Chitta $^{4,5}$ Penghao Wu $^{1}$ Jia Zeng $^{1}$ Ping Luo $^{3}$ Jun Zhang $^{2\\sharp}$ Andreas Geiger $^{4,5\\sharp}$ Yu Qiao $^{1\\sharp}$ Hongyang Li $^{1\\dagger}$", + "bbox": [ + 125, + 178, + 841, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 OpenDriveLab and Shanghai AI Lab 2 Hong Kong University of Science and Technology 3 University of Hong Kong 4 University of Tübingen 5 Tübingen AI Center", + "bbox": [ + 116, + 239, + 854, + 277 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b80cd5b5f461d4fd4cdc6680a5a80f78390f39e532e73a9719c51b85443b4c30.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 145, + 315, + 171, + 329 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "OpenDV-2K", + "text_level": 1, + "bbox": [ + 174, + 316, + 251, + 329 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cb2a404a9278330cd862b1bc1b938fabc376187acc9fda4362a08d287d0385e6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 78, + 330, + 330, + 473 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6d108f23ef58e3208f38828182bba2975dc9d3a31d659c14a88aab89bfc6b2be.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 416, + 311, + 439, + 332 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "GenAD", + "text_level": 1, + "bbox": [ + 444, + 316, + 491, + 328 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e19f39be9e834ce2ba2f960545759988df3469bfad8ed6cbdae86f58e2f9fa22.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 341, + 330, + 568, + 473 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0cd856b96f349dfbbd8ff9ae0fc5fa0e9a187141fda5adc06125fa773e8432f4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 311, + 717, + 333 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tasks", + "text_level": 1, + "bbox": [ + 723, + 316, + 756, + 328 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/48201f52fac94aa848c29342b5689af78076b4daa7a33a69e107d21e971a9a26.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 578, + 337, + 890, + 468 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/57270d275699fef0f89f2a74725ff5f9947d6f9c8817e08dadcc34b4f19bb07a.jpg", + "image_caption": [ + "2. Language-conditioned Prediction" + ], + "image_footnote": [], + "bbox": [ + 81, + 492, + 370, + 593 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/51c6bdb6e97a0982e4bf55874a52e1dc0df0e46ab56367adff2a33661bd99bdf.jpg", + "image_caption": [ + "3. Action-conditioned Prediction" + ], + "image_footnote": [], + "bbox": [ + 383, + 492, + 674, + 590 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/57aaf11851c35fa0fdae05e906b224d79236dd3965619e6e637c61b7175cad81.jpg", + "image_caption": [ + "4. Planning", + "Figure 1. Overview of the GenAD paradigm. We aim to establish a generalized video prediction paradigm for autonomous driving by presenting the largest multimodal driving video dataset to date, OpenDV-2K, and a generative model that predicts the future given past visual and textual input, GenAD. The strong generalization and controllability of GenAD is validated spanning a diverse spectrum of tasks, including zero-shot domain transfer, language-conditioned prediction, action-conditioned prediction, and motion planning." + ], + "image_footnote": [], + "bbox": [ + 681, + 492, + 890, + 593 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 691, + 313, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we introduce the first large-scale video prediction model in the autonomous driving discipline. To eliminate the restriction of high-cost data collection and empower the generalization ability of our model, we acquire massive data from the web and pair it with diverse and high-quality text descriptions. The resultant dataset accumulates over 2000 hours of driving videos, spanning areas all over the world with diverse weather conditions and traffic scenarios. Inheriting the merits from recent la", + "bbox": [ + 75, + 724, + 470, + 862 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tent diffusion models, our model, dubbed GenAD, handles the challenging dynamics in driving scenes with novel temporal reasoning blocks. We showcase that it can generalize to various unseen driving datasets in a zero-shot manner, surpassing general or driving-specific video prediction counterparts. Furthermore, GenAD can be adapted into an action-conditioned prediction model or a motion planner, holding great potential for real-world driving applications.", + "bbox": [ + 500, + 679, + 890, + 800 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 844, + 630, + 859 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Autonomous driving agents, as a promising application of high-level artificial intelligence, perceive the surrounding", + "bbox": [ + 500, + 869, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution, ordered by coin toss. Equal co-advising.", + "bbox": [ + 99, + 875, + 460, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Project lead. Primary contact: yangjiazhi@opendrivelab.com", + "bbox": [ + 99, + 887, + 460, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14662", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "environment, build internal world model representations, make decisions, and take actions in response [9, 50]. However, despite dedicated efforts in academia and industry for decades, their deployment is still restricted to certain areas or scenarios, and they cannot be applied over the world seamlessly. One critical reason is the limited generalization ability of learned models in structured autonomous driving systems. Typically, perception models face challenges of generalizing to diverse environments with changes in geographical locations, sensor configurations, weather conditions, open-set objects, etc.; prediction and planning models fail to generalize to nondeterministic futures with rare scenarios and different driving intentions [2, 16, 54].", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by how humans learn to perceive and cognize the world [27, 28, 49], we advocate employing driving videos as the universal interface that generalizes to diverse environments with dynamic futures. Based on this, a driving video predictive model is preferred to fully capture the world knowledge about driving scenarios (Fig. 1). By predicting the future, the video predictor essentially learns two vital aspects of autonomous driving: how the world operates, and how to maneuver safely in the wild.", + "bbox": [ + 75, + 287, + 472, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, the community has begun to adopt video as the interface to represent observation behavior and action for various robot tasks [11]. For domains such as classical video prediction and robotics, the video backgrounds are mostly static, the movement of robots is slow, and the resolution of videos is low. In contrast, for the driving scenarios, it struggles with outdoor environments being highly dynamic, agents encompassing much larger motions, and the sensory resolution covering a large range of view. These distinctions lead to substantial challenges for autonomous driving applications. Fortunately, there are some preliminary attempts on developing a video predictive model in the driving domain [4, 15, 19, 23, 25, 33, 38, 45, 47]. Despite promising progress in terms of prediction quality, these attempts have not achieved desirable capability of generalization as in classical robot tasks (e.g., manipulation), being confined to either limited scenarios such as highways with low traffic density [4] and small-scale datasets [15, 23, 33, 45, 47], or restricted conditions that raises difficulties to generate diverse environments [38]. How to unveil the potential of video prediction models for driving remains seldom explored.", + "bbox": [ + 75, + 425, + 472, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by the discussions above, we target at building a video predictive model for autonomous driving, capable of generalizing to new conditions and environments. To this end, we have to answer the following questions: (1) What data can be obtained in a feasible and scalable manner? (2) How can we formulate a predictive model to capture the complex evolution of dynamic scenarios? (3) How can we apply the (foundation) model for downstream tasks?", + "bbox": [ + 75, + 758, + 472, + 881 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Scaled Data. To achieve powerful generalization ability, a", + "bbox": [ + 76, + 885, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "substantial and diverse corpus of data is necessary. Inspired by the success of learning from Internet-scale data in foundation models [1, 26, 39], we construct our driving dataset from both the web and publicly licensed datasets. Compared to existing options, which are limited in scale and diversity due to their regulated collection processes, online data owns great diversity in several aspects: geographic locations, terrains, weather conditions, safety-critical scenarios, sensor settings, traffic elements, etc. To guarantee the data is of high-quality and desirable for large-scale training, we exhaustively collect driving recordings on YouTube and remove unintended corruption frames via rigorous human verification. Furthermore, videos are paired with diverse text-level conditions, including descriptions generated and refined with the aid of existing foundation models [30, 35], and high-level instructions inferred by a video classifier. Through these steps, we construct OpenDV-2K, the largest public driving dataset to date, containing more than 2000 hours of driving videos and being 374 times larger than the widely used nuScenes counterpart. Our dataset is publicly available at https://github.com/OpenDriveLab/DriveAGI.", + "bbox": [ + 496, + 90, + 893, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generalized Predictive Model. Learning a generalized driving video predictor bears several key challenges: generation quality, training efficiency, causal reasoning, and drastic view shift. We address these aspects by presenting a novel temporal generative model with two-stage learning. To capture the environment details, enhance generation quality, and maintain training efficiency simultaneously, we build upon the recent success of latent diffusion models (LDMs) [37, 41]. In the first stage, we transfer the generation distribution of LDM from its pre-trained general vision domain to the driving domain by fine-tuning it on OpenDV2K images. In the second stage, we interleave the proposed temporal reasoning blocks into the original model and learn to predict the future given past frames and conditions. Contrary to conventional temporal modules [4, 18] that suffer from causal confusion and large motion, our solution consists of causal temporal attention and decoupled spatial attention to efficiently model the drastic spatiotemporal shift in highly dynamic driving scenes. After sufficient training, our Generative model for Autonomous Driving (GenAD)1 can generalize to various scenarios in a zero-shot fashion.", + "bbox": [ + 496, + 425, + 893, + 743 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extensions for Simulation and Planning. After large-scale pre-training of video prediction, GenAD essentially understands how the world evolves and how to drive. We show how to adapt its learned knowledge for real-world driving problems, i.e., simulation and planning. For simulation, we fine-tune the pre-trained model with future ego trajectories as additional conditions, to associate future imaginations with different ego actions. We also empower", + "bbox": [ + 496, + 746, + 893, + 868 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1Note that GenAD is abbreviated from both Generative models and Generalized capabilities.", + "bbox": [ + 500, + 875, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14663", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/86c336cd91636c4b4b48a616a3121fb6c246e5755f166ef8f8c8f211c939b43f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetDuration (hours)Front-view FramesGeographic Diversity CountriesCitiesSensor Setup
XKITTI [14]1.415k11fixed
XCityscapes [10]0.525k350fixed
XWaymo Open* [43]11390k13fixed
XArgoverse 2* [48]4.2300k16fixed
✓nuScenes [6]5.5241k22fixed
✓nuPlan* [7]1204.0M24fixed
✓Talk2Car [12]4.7-22fixed
✓ONCE [34]1447M1-fixed
✓Honda-HAD [24]321.2M1-fixed
✓Honda-HDD-Action [40]1041.1M1-fixed
✓Honda-HDD-Cause [40]32-1-fixed
✓OpenDV-YouTube (Ours)174760.2M≥40†≥244†uncalibrated
-OpenDV-2K (Ours)205965.1M≥40†≥244†uncalibrated
", + "bbox": [ + 76, + 89, + 632, + 299 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. OpenDV-2K comparison at a glance to existing counterparts in terms of scale and diversity. Note that datasets with $\\checkmark$ are included in OpenDV-2K (last row). \\*Perception subset in Waymo Open, Argoverse 2, and nuPlan. Estimated by GPT [36] from video titles.", + "bbox": [ + 75, + 303, + 625, + 347 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3415a03895c87d97d29527857e876efbd9de7ba7bae8be4262d02422b3c74959.jpg", + "image_caption": [ + "Figure 2. Geographic distribution of OpenDV-2K. Our dataset covers ample driving scenarios around the world." + ], + "image_footnote": [], + "bbox": [ + 635, + 88, + 885, + 299 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "GenAD to perform planning on challenging benchmarks by using a lightweight planner to translate latent features into the future trajectory of the ego vehicle. On account of its pre-trained ability to predict accurate future frames, our algorithm exhibits promising results in both simulation consistency and planning reliability.", + "bbox": [ + 75, + 372, + 472, + 464 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. OpenDV-2K Dataset", + "text_level": 1, + "bbox": [ + 76, + 476, + 274, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We introduce OpenDV-2K, a large-scale multimodal dataset for autonomous driving, to support the training of a generalized video prediction model. The main component is a vast corpus of high-quality YouTube driving videos, which are collected from all over the world, and are gathered into our dataset after a careful curation process. We automatically create language annotations for these videos using vision-language models. To further improve its diversity in sensor configurations and language expressions, we merge 7 publicly licensed datasets into our OpenDV-2K, as shown in Tab. 1. As a result, OpenDV-2K occupies a total of 2059 hours of videos paired with texts, including 1747 hours from YouTube and 312 hours from public datasets. We use OpenDV-YouTube and OpenDV-2K to specify the YouTube split and the overall dataset, respectively.", + "bbox": [ + 75, + 502, + 470, + 729 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Diversity over Prior Datasets", + "text_level": 1, + "bbox": [ + 76, + 737, + 336, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A brief comparison with other public datasets is provided in Tab. 1. Beyond its significant scale, the proposed OpenDV-2K represents diversity across various aspects as follows.", + "bbox": [ + 75, + 760, + 468, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Globe-wise Geographic Distribution. Due to the global nature of online videos, OpenDV-2K covers more than 40 countries and 244 cities worldwide. This is a tremendous improvement over previous public datasets, which are typically gathered in a small number of restricted areas. We plot the specific distribution of OpenDV-YouTube in Fig. 2.", + "bbox": [ + 75, + 809, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Open-world Driving Scenarios. Our dataset provides a huge amount of realistic driving experience in the open world, covering rare environments like forests, extreme weather conditions like heavy snow, and appropriate driving behaviors in response to interactive traffic situations. These data are crucial for diversity and generalization yet are seldom collected in existing public datasets.", + "bbox": [ + 496, + 372, + 893, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unrestricted Sensor Configurations. Current driving datasets are confined to specific sensor configurations, including intrinsic and extrinsic camera parameters, image, sensor type, optics, etc., which poses great challenges for deploying the learned models with different sensors [32]. In contrast, YouTube driving videos are recorded in various types of vehicles with flexible camera setups, which aids in the robustness of the trained model when deployed using a novel camera setting.", + "bbox": [ + 496, + 484, + 893, + 621 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Towards High-quality Multimodal Dataset", + "text_level": 1, + "bbox": [ + 498, + 633, + 862, + 651 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Driving Video Collection and Curation. Finding clean driving videos from the vast pool of the web is a tedious and costly task. To simplify the process, we start by selecting certain video uploaders, i.e., YouTubers. Judging from the average length and overall quality, we collect 43 YouTubers with 2139 high-quality front-view driving videos. To make sure there is no overlap between training and validation sets, we take all videos from 3 YouTubers for validation, with the remaining videos as the training set. To rule out non-driving frames like video introductions and subscription reminders, we discard a certain length of segments at the beginning and end of each video. Each frame is then described with language contexts using a VLM model, BLIP-2 [30]. We further remove the black frames and transition frames, which are not ideal for training, by manually checking if there are certain keywords in these contexts. We give an illustration", + "bbox": [ + 496, + 657, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14664", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of the dataset construction pipeline in Appendix C.1.1, and we introduce how to generate the contexts below.", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Language Annotation for YouTube Videos. To create a predictive model that can be controlled by natural language to simulate different futures accordingly, To make the predictive model controllable and improve the sample quality [3], it is crucial to pair the driving videos with meaningful and varied language annotations. We construct two types of texts for OpenDV-YouTube, i.e., driving commands for ego-vehicle and frame descriptions, namely \"command\" and \"context\", to help the model comprehend ego actions and open-world concepts, respectively. For commands, we train a video classifier on Honda-HDD-Action [40] for 14 types of actions to label ego behaviors in a 4s sequence. These categorical commands will be further mapped to multiple free-form expressions from a predefined dictionary. For contexts, we leverage an established vision-language model, BLIP-2 [30], to describe the main objects and scenarios for each frame. For more details on annotations, please refer to Appendix C.1.2.", + "bbox": [ + 75, + 125, + 470, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Enlarging Language Spectrum with Public Datasets. Considering that BLIP-2 annotations are generated for static frames without comprehension of dynamic driving scenarios such as the traffic light transitions, we exploit several public datasets that provide linguistic descriptions for driving scenarios [6, 7, 12, 24, 34, 40]. However, their metadata is relatively sparse with only a few words such as \"sunny road\". We further enhance their text quality using GPT [36] to form a descriptive \"context\" and generate a \"command\" by categorizing the logged trajectory for each video clip. Ultimately, we integrate these datasets with OpenDV-YouTube to establish OpenDV-2K dataset, as shown in the last row of Tab. 1.", + "bbox": [ + 75, + 402, + 468, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. GenAD Framework", + "text_level": 1, + "bbox": [ + 76, + 613, + 269, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we introduce the training and design of the GenAD model. As shown in Fig. 3, GenAD is trained in two stages, i.e., image domain transferring and video prediction pre-training. The first stage adapts the general text-to-image model to the driving domain (Sec. 3.1). The second stage lifts the text-to-image model to a video prediction model with our proposed temporal reasoning block and modified training schemes (Sec. 3.2). In Sec. 3.3, we explore how the predictive model can be extended to action-conditioned prediction and planning.", + "bbox": [ + 75, + 638, + 468, + 791 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Image Domain Transfer", + "text_level": 1, + "bbox": [ + 76, + 801, + 297, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "On-board cameras capture a large field of views with abundant visual contents, including the road, background buildings, surrounding vehicles, etc., which require strong and robust generation capability to produce continuous and realistic driving scenarios. To facilitate the learning process, we", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "start with independent image generation in the first stage. Concretely, we initialize our model with SDXL [37], which is a large-scale latent diffusion model (LDM) for text-to-image generation, to leverage its ability to synthesize high-quality images with plenty of visual details. It is implemented as a denoising UNet $\\mathbf{f}_{\\theta}$ with several stacked convolution and attention blocks, which learns to synthesize images by denoising the noisy latents [41]. Specifically, given a noisy input latent $\\mathbf{x}_t$ corrupted by the forward diffusion process, it is trained to predict the added noise $\\epsilon$ of $\\mathbf{x}_t$ via the following objective:", + "bbox": [ + 496, + 90, + 890, + 257 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {i m g}} := \\mathbb {E} _ {\\mathbf {x}, \\epsilon \\sim \\mathcal {N} (0, 1), \\mathbf {c}, t} \\left[ \\| \\epsilon - \\mathbf {f} _ {\\theta} (\\mathbf {x} _ {t}; \\mathbf {c}, t) \\| _ {2} ^ {2} \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 544, + 262, + 890, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{x}$ and $\\mathbf{x}_t$ are the clean and noisy latent, respectively, $t$ denotes the timestep for different noise scales, and $\\mathbf{c}$ is the text condition that guides the denoising process, which is a concatenation of context and command. For training efficiency, the learning process takes place in a compressed latent space [13, 37, 41] instead of pixel space. During sampling, the model generates images from standard Gaussian noise by denoising the last-step predictions iteratively.", + "bbox": [ + 496, + 292, + 890, + 412 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, the original SDXL is trained on data in the general domain, such as portraits and artistic paintings, which are not concerned with autonomy systems. To adapt the model to synthesize images for driving, we fine-tune it on text-to-image generation using image-text pairs in OpenDV-2K with the same objective as Eq. (1). Following the original training of SDXL, all parameters $\\theta$ of the UNet are fine-tuned at this stage, whereas the CLIP text encoders [39] and the autoencoder [13] remain frozen.", + "bbox": [ + 496, + 414, + 890, + 550 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Video Prediction Pre-training", + "text_level": 1, + "bbox": [ + 500, + 556, + 764, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the second stage, with a few frames of a consecutive video as past observations, GenAD is trained to reason about all visual observations and predict several future frames in plausible ways. Similar to the first stage stage, the prediction process can also be guided by text conditions. However, predicting the highly dynamic driving world temporally is challenging due to two fundamental barriers.", + "bbox": [ + 496, + 580, + 890, + 686 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Causal Reasoning: To predict plausible futures following the temporal causality of the driving world, the model needs to comprehend the intentions of all other agents together with the ego vehicle, and understand underlying traffic rules, e.g., how the traffic will change with the transition of traffic lights.", + "2. Drastic View Shift: Contrary to typical video generation benchmarks which mainly have a static background with slow motion of centered objects, the view of driving changes drastically over time. Each pixel in every frame may move to a distant location in the next frame." + ], + "bbox": [ + 500, + 686, + 890, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose temporal reasoning blocks to address these problems. As illustrated in Fig. 3(c), each block is composed of three successive attention layers, i.e., the causal", + "bbox": [ + 496, + 854, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14665", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/529d3d44fcae5674be5d92a354b7f27841dd548e493f20c4d81919d42a78b120.jpg", + "image_caption": [ + "Figure 3. Framework of GenAD. (a) The two-stage learning for GenAD is composed of transferring the image domain of an image diffusion model to the driving field (a.1 Stage one), and video prediction pre-training for modeling the temporal dependency of videos (a.2 Stage two). (b) One transformer block in GenAD for the second stage training has interleaved temporal reasoning blocks before each frozen layer to align spatiotemporal features. (c) The proposed Temporal Reasoning Block includes one causal temporal attention (TA) and two decoupled spatial attention (SA) layers to extract features in different axes. A query grid attends to itself as well as blue grids while the dark gray grid is masked out in causal attention. 'Zero init' is appended at the end of each attention block to stabilize training." + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 676, + 345 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/809d540c959eee6b5b0cab1933780f2c041de5aebf8a0957020f426af1a665fd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 683, + 88, + 893, + 345 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "temporal attention layer and two decoupled spatial attention layers, which are tailored for the causal reasoning and modeling large shifts in the driving scenes, respectively.", + "bbox": [ + 75, + 459, + 470, + 505 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Causal Temporal Attention. Since the model after the stage-one training can only process each frame independently, we leverage temporal attention to exchange information among different video frames. The attention takes place in the time axis and models the temporal dependency of each grid-wise feature. However, directly adapting bidirectional temporal attention here as [4, 18, 46, 51] can hardly acquire the ability of causal reasoning, since the predictions will be inevitably dependent on the subsequent frames instead of past conditions. Therefore, we restrict the attention direction by adding a causal attention mask, as shown in the last row of Fig. 3(c), to encourage the model to fully exploit knowledge from past observations and faithfully reason about the future as if in real-world driving. We empirically found that the causality constraint greatly regularizes the predicted frames to be coherent with past frames. Following common practice, we also add temporal bias implemented as relative position embeddings on the time axis [42] to distinguish different frames of a sequence for temporal attention.", + "bbox": [ + 75, + 512, + 468, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Decoupled Spatial Attention. As driving videos feature fast perspective changes, features in a specific grid could vary greatly in different timesteps and are hard to correlate and learn by temporal attention, which suffers from a limited receptive field. In light of this, we introduce spa", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tial attention to propagate each grid feature in spatial axes to aid in gathering information for temporal attention. We implement a decoupled variant of self-attention for its efficiency with linear computational complexity, compared to quadratic full self-attention. As shown in Fig. 3(c), the two decoupled attention layers propagate features in horizontal and vertical axes, respectively.", + "bbox": [ + 496, + 459, + 890, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Deep Interaction. Intuitively, the spatial blocks fine-tuned in stage one refine features of each frame independently towards photorealism, whereas the temporal blocks introduced in stage two align features of all video frames towards coherency and consistency. To further boost the spatiotemporal feature interaction, we interleave the proposed temporal reasoning blocks with the original Transformer blocks in SDXL, i.e., spatial attention, cross attention, and feedforward network, as shown in Fig. 3(b).", + "bbox": [ + 496, + 570, + 890, + 708 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Zero Initialization. Similar to the previous practices [1, 52], for each block that is newly introduced in stage two, we initialized all parameters of its final layer as zero. This avoids disrupting the prior knowledge of the well-trained image generation model in the beginning and stabilizes the training process.", + "bbox": [ + 496, + 712, + 890, + 804 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training. GenAD is trained to predict the future by jointly denoising from the noisy latents with the guidance of past frames and text conditions. We first project $T$ consecutive frames of a video clip into a batch of latents $\\mathbf{v} = \\{\\mathbf{v}^m, \\mathbf{v}^n\\}$ , where the leading $m$ frame latents $\\mathbf{v}^m$ are clean, representing historical observations, and other $n = T - m$ frame latents", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14666", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathbf{v}^n$ indicate the future to be predicted. $\\mathbf{v}^n$ are then corrupted to $\\mathbf{v}_t^n$ by the forward diffusion process, where $t$ indexes a randomly sampled noise scale. The model is trained to predict the noise of $\\mathbf{v}_t^n$ conditioned on observations $\\mathbf{v}^m$ and text c. The learning objective of the video prediction model is formulated as follows:", + "bbox": [ + 75, + 90, + 470, + 181 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {v i d}} := \\mathbb {E} _ {\\mathbf {v}, \\boldsymbol {\\epsilon} \\sim \\mathcal {N} (0, 1), \\mathbf {c}, t} \\left[ \\| \\boldsymbol {\\epsilon} - \\mathbf {f} _ {\\theta , \\phi} \\left(\\mathbf {v} _ {t} ^ {n}; \\mathbf {v} ^ {m}, \\mathbf {c}, t\\right) \\| _ {2} ^ {2} \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 188, + 468, + 215 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\theta$ denotes the inherited stage-one model and $\\phi$ represents the newly inserted temporal reasoning blocks. Following [4], we freeze $\\theta$ and only train the temporal reasoning blocks to avoid perturbing the generation ability of the image generation model and focus on learning temporal dependencies in videos. Notably, only the outputs from the corrupted frames $\\mathbf{v}_t^n$ contribute to the training loss while those from condition frames $\\mathbf{v}^m$ are ignored.", + "bbox": [ + 75, + 224, + 468, + 347 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3. Extensions", + "text_level": 1, + "bbox": [ + 76, + 354, + 197, + 369 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Relying on the well-trained video prediction capability in driving scenarios, we further exploit the potential of the pretrained model in action-controlled prediction and planning, which are important for real-world driving systems. Here, we explore the downstream tasks on nuScenes [6] which provides recorded poses.", + "bbox": [ + 75, + 378, + 468, + 469 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Action-conditioned Prediction. To make our predictive model controllable with exact ego actions and act as a simulator [25], we fine-tune the model with the paired future trajectory as an additional condition. Specifically, we map the raw trajectory to a high-dimensional feature with Fourier embeddings [44]. After further projection by a linear layer, it is added to the original conditions. Thus, the ego actions are injected into the network through the conditional cross-attention layer in Fig. 3(b).", + "bbox": [ + 75, + 472, + 468, + 608 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Planning. By learning to predict the future, GenAD acquires strong representations of complex driving scenes, which can be further exploited for planning. Specifically, we extract spatiotemporal features of two historical frames through the UNet encoder of the frozen GenAD, which is nearly half the size of the entire model, and feed them to a multi-layer perceptron (MLP) to predict future waypoints. With the frozen GenAD encoder and a learnable MLP layer, the training process of our planner can be sped up by 3400 times compared to an end-to-end planning model UniAD [22], validating the effectiveness of the learned spatiotemporal feature of GenAD.", + "bbox": [ + 75, + 612, + 468, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 806, + 209, + 823 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Setup and Protocols", + "text_level": 1, + "bbox": [ + 76, + 832, + 266, + 848 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "GenAD is learned in two stages on OpenDV-2K but with different learning objectives (in Sec. 3) and input formats. In stage one, the model takes input (image, text) pairs and", + "bbox": [ + 75, + 854, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/85d117eb19d7f9aa80f575b1ab6e5f3499a61101716e5671fd5b76edb8fab99a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTraining DatasetPred.nuScenes
FID (↓)FVD (↓)
DriveGAN [25]✓73.4502
DriveDreamer* [45]nuScenes✓52.6452
DrivingDiffusion* [31]✗15.8332
GenAD-nus (Ours)nuScenes✓15.4244
GenAD (Ours)OpenDV-2K✓15.4184
", + "bbox": [ + 501, + 93, + 898, + 199 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Video generation quality compared to state-of-the-arts trained on nuScenes. \"Pred.\": evaluation by future prediction. *: requiring 3D layout inputs.", + "bbox": [ + 498, + 209, + 890, + 252 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "is trained on text-to-image generation. We broadcast the command annotation, which is labeled for each 4s video sequence, to all frames included. The model is trained for $300\\mathrm{K}$ iterations on 32 NVIDIA Tesla A100 GPUs with a total batch size of 256. In the second stage, GenAD is trained to jointly denoise future latents conditioned on past latents and texts. Its inputs are (video clip, text) pairs where each video clip is 4s at $2\\mathrm{Hz}$ . The current version of GenAD is trained on 64 GPUs for $112.5\\mathrm{K}$ iterations with a total batch size of 64. The input frames are resized to $256 \\times 448$ for training in both stages, and the text condition c is dropped at a probability of $p = 0.1$ to enable classifier-free guidance [17] in sampling, which is commonly used in diffusion models to improve sample quality. More training and sampling details are in Appendix D.", + "bbox": [ + 496, + 270, + 890, + 497 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Results of Video Prediction Pre-training", + "text_level": 1, + "bbox": [ + 498, + 508, + 843, + 525 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison to Recent Video Generation Approaches. We compare GenAD to recent advances on an unseen set with geofencing from OpenDV-YouTube, Waymo [43], KITTI [14], and Cityscapes [10] in a zero-shot generation manner. Fig. 4 depicts the qualitative results. Image-to-video models I2VGen-XL [53] and VideoCrafter1 [8] can not strictly follow the given frames to make predictions, yielding poor consistency between the predicted frames and past frames. The video prediction model DMVFN [21] that is trained on Cityscapes suffers from the unfavorable shape distortions in its predictions, especially on the three unseen datasets. In contrast, GenAD exhibits remarkable zero-shot generalization ability and visual quality although none of these sets are included in the training.", + "bbox": [ + 496, + 532, + 890, + 743 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison to nuScenes Experts. We also compare GenAD with the most recent available driving video generation models which are exclusively trained for nuScenes. Tab. 2 shows that GenAD surpasses all previous methods in both image fidelity (FID) and video coherence (FVD). Specifically, GenAD significantly reduces FVD by $44.5\\%$ compared to DrivingDiffusion [31], without taking 3D future layouts as additional inputs. For fair comparisons, we train a model variant (GenAD-nus) on nuScenes dataset only. We find that although GenAD-nus performs on par", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14667", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/948e930580ca8d90ade38de77861e2b964f2a400dc91a3f38cbba7fac5625c41.jpg", + "image_caption": [ + "Figure 4. Task on zero-shot video prediction for unseen scenarios. We show the generation results (in blue boxes) of different models given the same starting frames. GenAD makes more robust, realistic, and reasonable future predictions on unseen datasets (scenarios). More comparisons and visualizations are shown in Appendix." + ], + "image_footnote": [], + "bbox": [ + 86, + 88, + 529, + 459 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f85e60b3a0cbcc1383e23dd21035f4caa820a272079bcec0e55dbd74c5155760.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 532, + 88, + 890, + 458 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/88f0cfdb8928df107f6dcc46f75535ce3398111076335205dd499c99fe188ca0.jpg", + "image_caption": [ + "Figure 5. Task on language-conditioned prediction. Given two frames of a rainy scenario in the intersection and three high-level text conditions, GenAD simulates reasonable futures accordingly." + ], + "image_footnote": [], + "bbox": [ + 78, + 525, + 468, + 705 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "with GenAD on nuScenes, it struggles to generalize to unseen datasets like Waymo, where the generation degrades to the nuScenes visual pattern. In contrast, GenAD trained on OpenDV-2K exhibits strong generalization ability across datasets as shown in Fig. 4.", + "bbox": [ + 75, + 791, + 468, + 867 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We provide language-conditioned prediction samples on nuScenes in Fig. 5, where GenAD simulates various futures", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "from the same start following different textual instructions. The impressive generation quality is exhibited in the intricate details of the environment, and the natural transition of ego motion.", + "bbox": [ + 496, + 529, + 892, + 589 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation Study. We perform ablations by training each variant on a subset of OpenDV-2K for 75K steps. Starting from the baseline with plain temporal attentions [4, 18], we gradually introduce our proposed components. Notably, by interleaving the temporal blocks with the spatial blocks, the FVD significantly improves $(-17\\%)$ due to more sufficient spatiotemporal interactions. Both temporal causality and decoupled spatial attention contribute to better CLIP-SIM, improving the temporal consistency between future predictions and the condition frames. To be clear, the slight increase in FID and FVD, shown in fourth and third rows of Tab. 3 respectively, does not faithfully reflect a decline in generation quality as discussed in [4, 5, 37]. The effectiveness of each design is shown in Fig. 6.", + "bbox": [ + 496, + 594, + 892, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Results of Extensions", + "text_level": 1, + "bbox": [ + 500, + 816, + 699, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Action-conditioned Prediction. We further showcase the performance of the action-conditioned model fine-tuned on nuScenes, GenAD-act, in Fig. 7 and Tab. 4. Given two starting frames and a trajectory w composed of 6 future way", + "bbox": [ + 496, + 839, + 892, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14668", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cac3305aa3f6014f26b6ac0c03e4935457a9eee25b4ededa50bcdad6299f2dbc.jpg", + "image_caption": [ + "Figure 6. Case study for model designs. All components help alleviate artifacts and improve the consistency of future predictions." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 467, + 348 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4f3206e2a2ff8cc73aa1195b8a1aea4422fe2f6a451d22e6e5f24dfdde1bb63b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodYouTube
FID (↓)FVD (↓)CLIPSIM (↑)
Baseline18.32244.440.8405
+ Deep Interaction17.96201.690.8409
+ Temporal Causality16.54207.450.8550
+ Decoupled Spatial Attn.17.67189.540.8652
", + "bbox": [ + 81, + 398, + 460, + 486 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5d35442a6044a59cb021274383942f5df5d40b5ba33b222d3c379694f86c97a1.jpg", + "table_caption": [ + "Table 3. Ablation on model designs in GenAD. All proposed designs contribute to the final performance." + ], + "table_footnote": [], + "table_body": "
MethodConditionnuScenes\nAction Prediction Error (↓)
Ground truth-0.90
GenADtext2.54
GenAD-acttext + traj.2.02
", + "bbox": [ + 107, + 532, + 434, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "points, GenAD-act imagines 6 future frames following the trajectory sequence. To evaluate the consistency between the input trajectory $\\mathbf{w}$ and predicted frames, we establish an inverse dynamics model (IDM) on nuScenes as the evaluator, which projects a video sequence into a corresponding ego trajectory. We leverage the IDM to translate predicted frames into the trajectory $\\hat{\\mathbf{w}}$ , and calculate the L2 distance between $\\mathbf{w}$ and $\\hat{\\mathbf{w}}$ as the Action Prediction Error. Specifically, GenAD-act substantially reduces the Action Prediction Error by $20.4\\%$ compared to GenAD with text condition, allowing for more accurate future simulations.", + "bbox": [ + 75, + 683, + 468, + 849 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Planning Results. Tab. 5 depicts the planning results on nuScenes where ground truth poses for the ego vehicle are available. By freezing GenAD encoder and only optimizing", + "bbox": [ + 76, + 854, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cd7f531bb6325f055a4422f1202577063f1d2274d692faaa9b962d82dc87b944.jpg", + "image_caption": [ + "Figure 7. Task on action-conditioned prediction (simulation). Given the same starting frames and different future trajectories (shown in yellow dots in the first column), GenAD-act can simulate diverse futures following different ego intentions. More visualizations are in Appendix." + ], + "image_footnote": [], + "bbox": [ + 500, + 92, + 890, + 277 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2ea0b54bed8889da117118c9ba874c32c3d41aefd46350aa20aa2a10863cb5ac.jpg", + "table_caption": [ + "Table 4. Task on action-conditioned prediction. Compared to GenAD with text conditions only, GenAD-act enables more precise future predictions that follow the action condition." + ], + "table_footnote": [], + "table_body": "
Method# Trainable Params.nuScenes
ADE (↓)FDE (↓)
ST-P3* [20]10.9M2.653.73
UniAD* [22]58.8M1.031.65
GenAD (Ours)0.8M1.232.31
", + "bbox": [ + 539, + 383, + 849, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Task on open-loop planning. A lightweight MLP with frozen GenAD gets competitive planning results with $73 \\times$ fewer trainable parameters and front-view image alone. *: multi-view inputs. Evaluation protocols are aligned with UniAD [22].", + "bbox": [ + 496, + 468, + 890, + 525 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "an additional MLP on top of it, the model can effectively learn to plan. Notably, by pre-extracting image features through the UNet encoder of GenAD, the entire learning process for planning adaptation takes only 10 minutes on a single NVIDIA Tesla V100 device, which is 3400 times more efficient than the training of the UniAD planner [22].", + "bbox": [ + 496, + 564, + 890, + 656 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Limitations and Discussion", + "text_level": 1, + "bbox": [ + 500, + 686, + 751, + 704 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We study the system-level development of GenAD, a large-scale generalized video predictive model for autonomous driving. We also validate the adaptation of the learned representation of GenAD to driving tasks, i.e., learning a \"world model\" and motion planning. Although we obtain improved generalization to open domains, the increased model capacity poses challenges in both training efficiency and real-time deployment. We envision the unified video prediction task will serve as a scalable objective for future research on representation learning and policy learning. Another interesting direction involves distilling the encoded knowledge for a wider range of downstream tasks [29].", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14669", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andy Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022. 2, 5", + "[2] Mohammadhossein Bahari, Saeed Saadatnejad, Ahmad Rahimi, Mohammad Shaverdikondori, Amir Hossein Shahidzadeh, Seyed-Mohsen Moosavi-Dezfooli, and Alexandre Alahi. Vehicle trajectory prediction works, but not everywhere. In CVPR, 2022. 2", + "[3] Fan Bao, Chongxuan Li, Jiacheng Sun, and Jun Zhu. Why are conditional generative models better than unconditional ones? arXiv preprint arXiv:2212.00362, 2022. 4", + "[4] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 2, 5, 6, 7", + "[5] Tim Brooks, Janne Hellsten, Miika Aittala, Ting-Chun Wang, Timo Aila, Jaakko Lehtinen, Ming-Yu Liu, Alexei Efros, and Tero Karras. Generating long videos of dynamic scenes. In NeurIPS, 2022. 7", + "[6] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yuxin Pan, Giancarlo Baldan, and Oscar Beijbom. nuScenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 3, 4, 6", + "[7] Holger Caesar, Juraj Kabzan, Kok Seang Tan, Whye Kit Fong, Eric M. Wolff, Alex Lang, Luke Fletcher, Oscar Bei-jbom, and Sammy Omari. nuPlan: A closed-loop ml-based planning benchmark for autonomous vehicles. In CVPR Workshops, 2021. 3, 4", + "[8] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, Chao Weng, and Ying Shan. VideoCrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 6", + "[9] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers. arXiv preprint arXiv:2306.16927, 2023. 2", + "[10] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The Cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 3, 6", + "[11] Yilun Dai, Mengjiao Yang, Bo Dai, Hanjun Dai, Ofir Nachum, Josh Tenenbaum, Dale Schuurmans, and Pieter Abbeel. Learning universal policies via text-guided video generation. In NeurIPS, 2023. 2" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Thierry Deruyttere, Simon Vandenhende, Dusan Grujicic, Luc Van Gool, and Marie-Francine Moens. Talk2Car: Taking control of your self-driving car. In EMNLP, 2019. 3, 4", + "[13] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 4", + "[14] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The KITTI dataset. *IJRR*, 2013. 3, 6", + "[15] Agrim Gupta, Stephen Tian, Yunzhi Zhang, Jiajun Wu, Roberto Martin-Martin, and Li Fei-Fei. MaskViT: Masked visual pre-training for video prediction. In ICLR, 2023. 2", + "[16] Niklas Hanselmann, Katrin Renz, Kashyap Chitta, Apratim Bhattacharyya, and Andreas Geiger. KING: Generating safety-critical driving scenarios for robust imitation via kinematics gradients. In ECCV, 2022. 2", + "[17] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS Workshops, 2021. 6", + "[18] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. In NeurIPS, 2022. 2, 5, 7", + "[19] Anthony Hu, Lloyd Russell, Hudson Yeo, Zak Murez, George Fedoseev, Alex Kendall, Jamie Shotton, and Gianluca Corrado. GAIA-1: A generative world model for autonomous driving. arXiv preprint arXiv:2309.17080, 2023. 2", + "[20] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. ST-P3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In ECCV, 2022. 8", + "[21] Xiaotao Hu, Zhewei Huang, Ailin Huang, Jun Xu, and Shuchang Zhou. A dynamic multi-scale voxel flow network for video prediction. In CVPR, 2023. 6", + "[22] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, Lewei Lu, Xiaosong Jia, Qiang Liu, Jifeng Dai, Yu Qiao, and Hongyang Li. Planning-oriented autonomous driving. In CVPR, 2023. 6, 8", + "[23] Fan Jia, Weixin Mao, Yingfei Liu, Yucheng Zhao, Yuqing Wen, Chi Zhang, Xiangyu Zhang, and Tiancai Wang. A driver-i: A general world model for autonomous driving. arXiv preprint arXiv:2311.13549, 2023. 2", + "[24] Jinkyu Kim, Teruhisa Miso, Yi-Ting Chen, Ashish Tawari, and John Canny. Grounding human-to-vehicle advice for self-driving vehicles. In CVPR, 2019. 3, 4", + "[25] Seung Wook Kim, Jonah Philion, Antonio Torralba, and Sanja Fidler. DriveGAN: Towards a controllable high-quality neural simulation. In CVPR, 2021. 2, 6", + "[26] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In ICCV, 2023. 2", + "[27] Yann LeCun. A path towards autonomous machine intelligence version 0.9. 2, 2022-06-27. Open Review, 62, 2022. 2", + "[28] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. Nature, 2015. 2" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14670", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Daiqing Li, Huan Ling, Amlan Kar, David Acuna, Seung Wook Kim, Karsten Kreis, Antonio Torralba, and Sanja Fidler. DreamTeacher: Pretraining image backbones with deep generative models. In ICCV, 2023. 8", + "[30] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 2, 3, 4", + "[31] Xiaofan Li, Yifu Zhang, and Xiaoqing Ye. DrivingDiffusion: Layout-guided multi-view driving scene video generation with latent diffusion model. arXiv preprint arXiv:2310.07771, 2023. 6", + "[32] Zhenyu Li, Zehui Chen, Ang Li, Liangji Fang, Qinhong Jiang, Xianming Liu, and Junjun Jiang. Unsupervised domain adaptation for monocular 3d object detection via self-training. In ECCV, 2022. 3", + "[33] Jiachen Lu, Ze Huang, Jiahui Zhang, Zeyu Yang, and Li Zhang. Wovogen: World volume-aware diffusion for controllable multi-camera driving scene generation. arXiv preprint arXiv:2312.02934, 2023. 2", + "[34] Jiageng Mao, Minzhe Niu, Chenhan Jiang, Hanxue Liang, Xiaodan Liang, Yamin Li, Chao Ye, Wei Zhang, Zhenguo Li, Jie Yu, Hang Xu, and Chunjing Xu. One million scenes for autonomous driving: ONCE dataset. In NeurIPS Datasets and Benchmarks, 2021. 3, 4", + "[35] OpenAI. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2", + "[36] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul F Christiano, Jan Leike, and Ryan Lowe. Training language models to follow instructions with human feedback. In NeurIPS, 2022. 3, 4", + "[37] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 4, 7", + "[38] Xiaojuan Qi, Zhengzhe Liu, Qifeng Chen, and Jiaya Jia. 3D motion decomposition for rgbd future dynamic scene synthesis. In CVPR, 2019. 2", + "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 4", + "[40] Vasili Ramanishka, Yi-Ting Chen, Teruhisa Mitsu, and Kate Saenko. Toward driving scene understanding: A dataset for learning driver behavior and causal reasoning. In CVPR, 2018. 3, 4", + "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 4", + "[42] Peter Shaw, Jakob Uszkoreit, and Ashish Vaswani. Self-attention with relative position representations. arXiv preprint arXiv:1803.02155, 2018. 5" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott M. Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, 2020. 3, 6", + "[44] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. In NeurIPS, 2020. 6", + "[45] Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, and Jiwen Lu. DriveDreamer: Towards real-world-driven world models for autonomous driving. arXiv preprint arXiv:2309.09777, 2023. 2, 6", + "[46] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, Yuwei Guo, Tianxing Wu, Chenyang Si, Yuming Jiang, Cunjian Chen, Chen Change Loy, Bo Dai, Dahua Lin, Yu Qiao, and Ziwei Liu. LAVIE: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 5", + "[47] Yuqi Wang, Jiawei He, Lue Fan, Hongxin Li, Yuntao Chen, and Zhaoxiang Zhang. Driving into the future: Multiview visual forecasting and planning with world model for autonomous driving. arXiv preprint arXiv:2311.17918, 2023. 2", + "[48] Benjamin Wilson, William Qi, Tanmay Agarwal, John Lambert, Jagjeet Singh, Siddhesh Khandelwal, Bowen Pan, Ratnesh Kumar, Andrew Hartnett, Jhony Kaesemodel Pontes, Deva Ramanan, Peter Carr, and James Hays. Argoverse 2: Next generation datasets for self-driving perception and forecasting. In NeurIPS Datasets and Benchmarks, 2021. 3", + "[49] Daniel M Wolpert, Zoubin Ghahramani, and Michael I Jordan. An internal model for sensorimotor integration. Science, 1995. 2", + "[50] Zhiheng Xi, Wenxiang Chen, Xin Guo, Wei He, Yiwen Ding, Boyang Hong, Ming Zhang, Junzhe Wang, Senjie Jin, Enyu Zhou, Rui Zheng, Xiaoran Fan, Xiao Wang, Liao Xiong, Yuhao Zhou, Weiran Wang, Changhao Jiang, Yicheng Zou, Xiangyang Liu, Zhangyue Yin, Shihan Dou, Rongxiang Weng, Wensen Cheng, Qi Zhang, Wenjuan Qin, Yongyan Zheng, Xipeng Qiu, Xuanjing Huang, and Tao Gui. The rise and potential of large language model based agents: A survey. arXiv preprint arXiv:2309.07864, 2023. 2", + "[51] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. arXiv preprint arXiv:2309.15818, 2023. 5", + "[52] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 5", + "[53] Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qin, Xiang Wang, Deli Zhao, and" + ], + "bbox": [ + 501, + 90, + 890, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14671", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Jingren Zhou. I2VGen-XL: High-quality image-to-video synthesis via cascaded diffusion models. arXiv preprint arXiv:2311.04145, 2023. 6", + "[54] Ruizhao Zhu, Peng Huang, Eshed Ohn-Bar, and Venkatesh Saligrama. Learning to drive anywhere. In CoRL, 2023. 2" + ], + "bbox": [ + 78, + 90, + 468, + 162 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "14672", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Generalized Predictive Model for Autonomous Driving/554856ec-712f-446a-b09c-0cce647f2af6_model.json b/2024/Generalized Predictive Model for Autonomous Driving/554856ec-712f-446a-b09c-0cce647f2af6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..750e86647eb4476595b7470f418034c6b2827ab4 --- /dev/null +++ b/2024/Generalized Predictive Model for Autonomous Driving/554856ec-712f-446a-b09c-0cce647f2af6_model.json @@ -0,0 +1,2147 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.208, + 0.131, + 0.764, + 0.154 + ], + "angle": 0, + "content": "Generalized Predictive Model for Autonomous Driving" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.179, + 0.842, + 0.236 + ], + "angle": 0, + "content": "Jiazhi Yang\\(^{1*}\\) Shenyuan Gao\\(^{2,1*}\\) Yihang Qiu\\(^{1*}\\) Li Chen\\(^{3,1\\dagger}\\) Tianyu Li\\(^{1}\\) Bo Dai\\(^{1}\\) Kashyap Chitta\\(^{4,5}\\) Penghao Wu\\(^{1}\\) Jia Zeng\\(^{1}\\) Ping Luo\\(^{3}\\) Jun Zhang\\(^{2\\sharp}\\) Andreas Geiger\\(^{4,5\\sharp}\\) Yu Qiao\\(^{1\\sharp}\\) Hongyang Li\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.241, + 0.855, + 0.278 + ], + "angle": 0, + "content": "1 OpenDriveLab and Shanghai AI Lab 2 Hong Kong University of Science and Technology 3 University of Hong Kong 4 University of Tübingen 5 Tübingen AI Center" + }, + { + "type": "image", + "bbox": [ + 0.147, + 0.316, + 0.172, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.176, + 0.317, + 0.252, + 0.33 + ], + "angle": 0, + "content": "OpenDV-2K" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.331, + 0.331, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.417, + 0.313, + 0.441, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.445, + 0.318, + 0.493, + 0.329 + ], + "angle": 0, + "content": "GenAD" + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.331, + 0.57, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.313, + 0.718, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.724, + 0.318, + 0.758, + 0.329 + ], + "angle": 0, + "content": "Tasks" + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.338, + 0.891, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.13, + 0.481, + 0.325, + 0.492 + ], + "angle": 0, + "content": "2. Language-conditioned Prediction" + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.493, + 0.372, + 0.594 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.44, + 0.481, + 0.614, + 0.492 + ], + "angle": 0, + "content": "3. Action-conditioned Prediction" + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.493, + 0.675, + 0.591 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.758, + 0.481, + 0.822, + 0.492 + ], + "angle": 0, + "content": "4. Planning" + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.493, + 0.891, + 0.594 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.609, + 0.894, + 0.667 + ], + "angle": 0, + "content": "Figure 1. Overview of the GenAD paradigm. We aim to establish a generalized video prediction paradigm for autonomous driving by presenting the largest multimodal driving video dataset to date, OpenDV-2K, and a generative model that predicts the future given past visual and textual input, GenAD. The strong generalization and controllability of GenAD is validated spanning a diverse spectrum of tasks, including zero-shot domain transfer, language-conditioned prediction, action-conditioned prediction, and motion planning." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.693, + 0.314, + 0.71 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.726, + 0.471, + 0.863 + ], + "angle": 0, + "content": "In this paper, we introduce the first large-scale video prediction model in the autonomous driving discipline. To eliminate the restriction of high-cost data collection and empower the generalization ability of our model, we acquire massive data from the web and pair it with diverse and high-quality text descriptions. The resultant dataset accumulates over 2000 hours of driving videos, spanning areas all over the world with diverse weather conditions and traffic scenarios. Inheriting the merits from recent la" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.68, + 0.892, + 0.801 + ], + "angle": 0, + "content": "tent diffusion models, our model, dubbed GenAD, handles the challenging dynamics in driving scenes with novel temporal reasoning blocks. We showcase that it can generalize to various unseen driving datasets in a zero-shot manner, surpassing general or driving-specific video prediction counterparts. Furthermore, GenAD can be adapted into an action-conditioned prediction model or a motion planner, holding great potential for real-world driving applications." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.845, + 0.632, + 0.861 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Autonomous driving agents, as a promising application of high-level artificial intelligence, perceive the surrounding" + }, + { + "type": "page_footnote", + "bbox": [ + 0.1, + 0.875, + 0.461, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution, ordered by coin toss. Equal co-advising." + }, + { + "type": "page_footnote", + "bbox": [ + 0.1, + 0.888, + 0.461, + 0.901 + ], + "angle": 0, + "content": "†Project lead. Primary contact: yangjiazhi@opendrivelab.com" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.875, + 0.461, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14662" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.288 + ], + "angle": 0, + "content": "environment, build internal world model representations, make decisions, and take actions in response [9, 50]. However, despite dedicated efforts in academia and industry for decades, their deployment is still restricted to certain areas or scenarios, and they cannot be applied over the world seamlessly. One critical reason is the limited generalization ability of learned models in structured autonomous driving systems. Typically, perception models face challenges of generalizing to diverse environments with changes in geographical locations, sensor configurations, weather conditions, open-set objects, etc.; prediction and planning models fail to generalize to nondeterministic futures with rare scenarios and different driving intentions [2, 16, 54]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.289, + 0.473, + 0.426 + ], + "angle": 0, + "content": "Motivated by how humans learn to perceive and cognize the world [27, 28, 49], we advocate employing driving videos as the universal interface that generalizes to diverse environments with dynamic futures. Based on this, a driving video predictive model is preferred to fully capture the world knowledge about driving scenarios (Fig. 1). By predicting the future, the video predictor essentially learns two vital aspects of autonomous driving: how the world operates, and how to maneuver safely in the wild." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.426, + 0.473, + 0.759 + ], + "angle": 0, + "content": "Recently, the community has begun to adopt video as the interface to represent observation behavior and action for various robot tasks [11]. For domains such as classical video prediction and robotics, the video backgrounds are mostly static, the movement of robots is slow, and the resolution of videos is low. In contrast, for the driving scenarios, it struggles with outdoor environments being highly dynamic, agents encompassing much larger motions, and the sensory resolution covering a large range of view. These distinctions lead to substantial challenges for autonomous driving applications. Fortunately, there are some preliminary attempts on developing a video predictive model in the driving domain [4, 15, 19, 23, 25, 33, 38, 45, 47]. Despite promising progress in terms of prediction quality, these attempts have not achieved desirable capability of generalization as in classical robot tasks (e.g., manipulation), being confined to either limited scenarios such as highways with low traffic density [4] and small-scale datasets [15, 23, 33, 45, 47], or restricted conditions that raises difficulties to generate diverse environments [38]. How to unveil the potential of video prediction models for driving remains seldom explored." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.759, + 0.473, + 0.882 + ], + "angle": 0, + "content": "Motivated by the discussions above, we target at building a video predictive model for autonomous driving, capable of generalizing to new conditions and environments. To this end, we have to answer the following questions: (1) What data can be obtained in a feasible and scalable manner? (2) How can we formulate a predictive model to capture the complex evolution of dynamic scenarios? (3) How can we apply the (foundation) model for downstream tasks?" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.472, + 0.902 + ], + "angle": 0, + "content": "Scaled Data. To achieve powerful generalization ability, a" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.424 + ], + "angle": 0, + "content": "substantial and diverse corpus of data is necessary. Inspired by the success of learning from Internet-scale data in foundation models [1, 26, 39], we construct our driving dataset from both the web and publicly licensed datasets. Compared to existing options, which are limited in scale and diversity due to their regulated collection processes, online data owns great diversity in several aspects: geographic locations, terrains, weather conditions, safety-critical scenarios, sensor settings, traffic elements, etc. To guarantee the data is of high-quality and desirable for large-scale training, we exhaustively collect driving recordings on YouTube and remove unintended corruption frames via rigorous human verification. Furthermore, videos are paired with diverse text-level conditions, including descriptions generated and refined with the aid of existing foundation models [30, 35], and high-level instructions inferred by a video classifier. Through these steps, we construct OpenDV-2K, the largest public driving dataset to date, containing more than 2000 hours of driving videos and being 374 times larger than the widely used nuScenes counterpart. Our dataset is publicly available at https://github.com/OpenDriveLab/DriveAGI." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.426, + 0.895, + 0.744 + ], + "angle": 0, + "content": "Generalized Predictive Model. Learning a generalized driving video predictor bears several key challenges: generation quality, training efficiency, causal reasoning, and drastic view shift. We address these aspects by presenting a novel temporal generative model with two-stage learning. To capture the environment details, enhance generation quality, and maintain training efficiency simultaneously, we build upon the recent success of latent diffusion models (LDMs) [37, 41]. In the first stage, we transfer the generation distribution of LDM from its pre-trained general vision domain to the driving domain by fine-tuning it on OpenDV2K images. In the second stage, we interleave the proposed temporal reasoning blocks into the original model and learn to predict the future given past frames and conditions. Contrary to conventional temporal modules [4, 18] that suffer from causal confusion and large motion, our solution consists of causal temporal attention and decoupled spatial attention to efficiently model the drastic spatiotemporal shift in highly dynamic driving scenes. After sufficient training, our Generative model for Autonomous Driving (GenAD)1 can generalize to various scenarios in a zero-shot fashion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.747, + 0.895, + 0.869 + ], + "angle": 0, + "content": "Extensions for Simulation and Planning. After large-scale pre-training of video prediction, GenAD essentially understands how the world evolves and how to drive. We show how to adapt its learned knowledge for real-world driving problems, i.e., simulation and planning. For simulation, we fine-tune the pre-trained model with future ego trajectories as additional conditions, to associate future imaginations with different ego actions. We also empower" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.892, + 0.901 + ], + "angle": 0, + "content": "1Note that GenAD is abbreviated from both Generative models and Generalized capabilities." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14663" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.078, + 0.09, + 0.633, + 0.3 + ], + "angle": 0, + "content": "
DatasetDuration (hours)Front-view FramesGeographic Diversity CountriesCitiesSensor Setup
XKITTI [14]1.415k11fixed
XCityscapes [10]0.525k350fixed
XWaymo Open* [43]11390k13fixed
XArgoverse 2* [48]4.2300k16fixed
✓nuScenes [6]5.5241k22fixed
✓nuPlan* [7]1204.0M24fixed
✓Talk2Car [12]4.7-22fixed
✓ONCE [34]1447M1-fixed
✓Honda-HAD [24]321.2M1-fixed
✓Honda-HDD-Action [40]1041.1M1-fixed
✓Honda-HDD-Cause [40]32-1-fixed
✓OpenDV-YouTube (Ours)174760.2M≥40†≥244†uncalibrated
-OpenDV-2K (Ours)205965.1M≥40†≥244†uncalibrated
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.304, + 0.627, + 0.348 + ], + "angle": 0, + "content": "Table 1. OpenDV-2K comparison at a glance to existing counterparts in terms of scale and diversity. Note that datasets with \\(\\checkmark\\) are included in OpenDV-2K (last row). \\*Perception subset in Waymo Open, Argoverse 2, and nuPlan. Estimated by GPT [36] from video titles." + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.089, + 0.887, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.635, + 0.304, + 0.895, + 0.348 + ], + "angle": 0, + "content": "Figure 2. Geographic distribution of OpenDV-2K. Our dataset covers ample driving scenarios around the world." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.373, + 0.473, + 0.465 + ], + "angle": 0, + "content": "GenAD to perform planning on challenging benchmarks by using a lightweight planner to translate latent features into the future trajectory of the ego vehicle. On account of its pre-trained ability to predict accurate future frames, our algorithm exhibits promising results in both simulation consistency and planning reliability." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.477, + 0.275, + 0.495 + ], + "angle": 0, + "content": "2. OpenDV-2K Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.503, + 0.471, + 0.73 + ], + "angle": 0, + "content": "We introduce OpenDV-2K, a large-scale multimodal dataset for autonomous driving, to support the training of a generalized video prediction model. The main component is a vast corpus of high-quality YouTube driving videos, which are collected from all over the world, and are gathered into our dataset after a careful curation process. We automatically create language annotations for these videos using vision-language models. To further improve its diversity in sensor configurations and language expressions, we merge 7 publicly licensed datasets into our OpenDV-2K, as shown in Tab. 1. As a result, OpenDV-2K occupies a total of 2059 hours of videos paired with texts, including 1747 hours from YouTube and 312 hours from public datasets. We use OpenDV-YouTube and OpenDV-2K to specify the YouTube split and the overall dataset, respectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.738, + 0.338, + 0.755 + ], + "angle": 0, + "content": "2.1. Diversity over Prior Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.761, + 0.47, + 0.807 + ], + "angle": 0, + "content": "A brief comparison with other public datasets is provided in Tab. 1. Beyond its significant scale, the proposed OpenDV-2K represents diversity across various aspects as follows." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Globe-wise Geographic Distribution. Due to the global nature of online videos, OpenDV-2K covers more than 40 countries and 244 cities worldwide. This is a tremendous improvement over previous public datasets, which are typically gathered in a small number of restricted areas. We plot the specific distribution of OpenDV-YouTube in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.373, + 0.895, + 0.48 + ], + "angle": 0, + "content": "Open-world Driving Scenarios. Our dataset provides a huge amount of realistic driving experience in the open world, covering rare environments like forests, extreme weather conditions like heavy snow, and appropriate driving behaviors in response to interactive traffic situations. These data are crucial for diversity and generalization yet are seldom collected in existing public datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.485, + 0.895, + 0.622 + ], + "angle": 0, + "content": "Unrestricted Sensor Configurations. Current driving datasets are confined to specific sensor configurations, including intrinsic and extrinsic camera parameters, image, sensor type, optics, etc., which poses great challenges for deploying the learned models with different sensors [32]. In contrast, YouTube driving videos are recorded in various types of vehicles with flexible camera setups, which aids in the robustness of the trained model when deployed using a novel camera setting." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.635, + 0.864, + 0.652 + ], + "angle": 0, + "content": "2.2. Towards High-quality Multimodal Dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Driving Video Collection and Curation. Finding clean driving videos from the vast pool of the web is a tedious and costly task. To simplify the process, we start by selecting certain video uploaders, i.e., YouTubers. Judging from the average length and overall quality, we collect 43 YouTubers with 2139 high-quality front-view driving videos. To make sure there is no overlap between training and validation sets, we take all videos from 3 YouTubers for validation, with the remaining videos as the training set. To rule out non-driving frames like video introductions and subscription reminders, we discard a certain length of segments at the beginning and end of each video. Each frame is then described with language contexts using a VLM model, BLIP-2 [30]. We further remove the black frames and transition frames, which are not ideal for training, by manually checking if there are certain keywords in these contexts. We give an illustration" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "14664" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "of the dataset construction pipeline in Appendix C.1.1, and we introduce how to generate the contexts below." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.126, + 0.471, + 0.399 + ], + "angle": 0, + "content": "Language Annotation for YouTube Videos. To create a predictive model that can be controlled by natural language to simulate different futures accordingly, To make the predictive model controllable and improve the sample quality [3], it is crucial to pair the driving videos with meaningful and varied language annotations. We construct two types of texts for OpenDV-YouTube, i.e., driving commands for ego-vehicle and frame descriptions, namely \"command\" and \"context\", to help the model comprehend ego actions and open-world concepts, respectively. For commands, we train a video classifier on Honda-HDD-Action [40] for 14 types of actions to label ego behaviors in a 4s sequence. These categorical commands will be further mapped to multiple free-form expressions from a predefined dictionary. For contexts, we leverage an established vision-language model, BLIP-2 [30], to describe the main objects and scenarios for each frame. For more details on annotations, please refer to Appendix C.1.2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.403, + 0.47, + 0.601 + ], + "angle": 0, + "content": "Enlarging Language Spectrum with Public Datasets. Considering that BLIP-2 annotations are generated for static frames without comprehension of dynamic driving scenarios such as the traffic light transitions, we exploit several public datasets that provide linguistic descriptions for driving scenarios [6, 7, 12, 24, 34, 40]. However, their metadata is relatively sparse with only a few words such as \"sunny road\". We further enhance their text quality using GPT [36] to form a descriptive \"context\" and generate a \"command\" by categorizing the logged trajectory for each video clip. Ultimately, we integrate these datasets with OpenDV-YouTube to establish OpenDV-2K dataset, as shown in the last row of Tab. 1." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.614, + 0.27, + 0.63 + ], + "angle": 0, + "content": "3. GenAD Framework" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.64, + 0.47, + 0.792 + ], + "angle": 0, + "content": "In this section, we introduce the training and design of the GenAD model. As shown in Fig. 3, GenAD is trained in two stages, i.e., image domain transferring and video prediction pre-training. The first stage adapts the general text-to-image model to the driving domain (Sec. 3.1). The second stage lifts the text-to-image model to a video prediction model with our proposed temporal reasoning block and modified training schemes (Sec. 3.2). In Sec. 3.3, we explore how the predictive model can be extended to action-conditioned prediction and planning." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.299, + 0.817 + ], + "angle": 0, + "content": "3.1. Image Domain Transfer" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.903 + ], + "angle": 0, + "content": "On-board cameras capture a large field of views with abundant visual contents, including the road, background buildings, surrounding vehicles, etc., which require strong and robust generation capability to produce continuous and realistic driving scenarios. To facilitate the learning process, we" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.258 + ], + "angle": 0, + "content": "start with independent image generation in the first stage. Concretely, we initialize our model with SDXL [37], which is a large-scale latent diffusion model (LDM) for text-to-image generation, to leverage its ability to synthesize high-quality images with plenty of visual details. It is implemented as a denoising UNet \\(\\mathbf{f}_{\\theta}\\) with several stacked convolution and attention blocks, which learns to synthesize images by denoising the noisy latents [41]. Specifically, given a noisy input latent \\(\\mathbf{x}_t\\) corrupted by the forward diffusion process, it is trained to predict the added noise \\(\\epsilon\\) of \\(\\mathbf{x}_t\\) via the following objective:" + }, + { + "type": "equation", + "bbox": [ + 0.545, + 0.263, + 0.892, + 0.289 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {i m g}} := \\mathbb {E} _ {\\mathbf {x}, \\epsilon \\sim \\mathcal {N} (0, 1), \\mathbf {c}, t} \\left[ \\| \\epsilon - \\mathbf {f} _ {\\theta} (\\mathbf {x} _ {t}; \\mathbf {c}, t) \\| _ {2} ^ {2} \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.294, + 0.892, + 0.414 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}\\) and \\(\\mathbf{x}_t\\) are the clean and noisy latent, respectively, \\(t\\) denotes the timestep for different noise scales, and \\(\\mathbf{c}\\) is the text condition that guides the denoising process, which is a concatenation of context and command. For training efficiency, the learning process takes place in a compressed latent space [13, 37, 41] instead of pixel space. During sampling, the model generates images from standard Gaussian noise by denoising the last-step predictions iteratively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.415, + 0.892, + 0.551 + ], + "angle": 0, + "content": "However, the original SDXL is trained on data in the general domain, such as portraits and artistic paintings, which are not concerned with autonomy systems. To adapt the model to synthesize images for driving, we fine-tune it on text-to-image generation using image-text pairs in OpenDV-2K with the same objective as Eq. (1). Following the original training of SDXL, all parameters \\(\\theta\\) of the UNet are fine-tuned at this stage, whereas the CLIP text encoders [39] and the autoencoder [13] remain frozen." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.558, + 0.766, + 0.573 + ], + "angle": 0, + "content": "3.2. Video Prediction Pre-training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.892, + 0.687 + ], + "angle": 0, + "content": "In the second stage, with a few frames of a consecutive video as past observations, GenAD is trained to reason about all visual observations and predict several future frames in plausible ways. Similar to the first stage stage, the prediction process can also be guided by text conditions. However, predicting the highly dynamic driving world temporally is challenging due to two fundamental barriers." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.688, + 0.892, + 0.779 + ], + "angle": 0, + "content": "1. Causal Reasoning: To predict plausible futures following the temporal causality of the driving world, the model needs to comprehend the intentions of all other agents together with the ego vehicle, and understand underlying traffic rules, e.g., how the traffic will change with the transition of traffic lights." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.779, + 0.892, + 0.854 + ], + "angle": 0, + "content": "2. Drastic View Shift: Contrary to typical video generation benchmarks which mainly have a static background with slow motion of centered objects, the view of driving changes drastically over time. Each pixel in every frame may move to a distant location in the next frame." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.688, + 0.892, + 0.854 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.902 + ], + "angle": 0, + "content": "We propose temporal reasoning blocks to address these problems. As illustrated in Fig. 3(c), each block is composed of three successive attention layers, i.e., the causal" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14665" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.677, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.684, + 0.089, + 0.895, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.349, + 0.895, + 0.435 + ], + "angle": 0, + "content": "Figure 3. Framework of GenAD. (a) The two-stage learning for GenAD is composed of transferring the image domain of an image diffusion model to the driving field (a.1 Stage one), and video prediction pre-training for modeling the temporal dependency of videos (a.2 Stage two). (b) One transformer block in GenAD for the second stage training has interleaved temporal reasoning blocks before each frozen layer to align spatiotemporal features. (c) The proposed Temporal Reasoning Block includes one causal temporal attention (TA) and two decoupled spatial attention (SA) layers to extract features in different axes. A query grid attends to itself as well as blue grids while the dark gray grid is masked out in causal attention. 'Zero init' is appended at the end of each attention block to stabilize training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.46, + 0.471, + 0.506 + ], + "angle": 0, + "content": "temporal attention layer and two decoupled spatial attention layers, which are tailored for the causal reasoning and modeling large shifts in the driving scenes, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.513, + 0.47, + 0.817 + ], + "angle": 0, + "content": "Causal Temporal Attention. Since the model after the stage-one training can only process each frame independently, we leverage temporal attention to exchange information among different video frames. The attention takes place in the time axis and models the temporal dependency of each grid-wise feature. However, directly adapting bidirectional temporal attention here as [4, 18, 46, 51] can hardly acquire the ability of causal reasoning, since the predictions will be inevitably dependent on the subsequent frames instead of past conditions. Therefore, we restrict the attention direction by adding a causal attention mask, as shown in the last row of Fig. 3(c), to encourage the model to fully exploit knowledge from past observations and faithfully reason about the future as if in real-world driving. We empirically found that the causality constraint greatly regularizes the predicted frames to be coherent with past frames. Following common practice, we also add temporal bias implemented as relative position embeddings on the time axis [42] to distinguish different frames of a sequence for temporal attention." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Decoupled Spatial Attention. As driving videos feature fast perspective changes, features in a specific grid could vary greatly in different timesteps and are hard to correlate and learn by temporal attention, which suffers from a limited receptive field. In light of this, we introduce spa" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.46, + 0.892, + 0.567 + ], + "angle": 0, + "content": "tial attention to propagate each grid feature in spatial axes to aid in gathering information for temporal attention. We implement a decoupled variant of self-attention for its efficiency with linear computational complexity, compared to quadratic full self-attention. As shown in Fig. 3(c), the two decoupled attention layers propagate features in horizontal and vertical axes, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.571, + 0.892, + 0.709 + ], + "angle": 0, + "content": "Deep Interaction. Intuitively, the spatial blocks fine-tuned in stage one refine features of each frame independently towards photorealism, whereas the temporal blocks introduced in stage two align features of all video frames towards coherency and consistency. To further boost the spatiotemporal feature interaction, we interleave the proposed temporal reasoning blocks with the original Transformer blocks in SDXL, i.e., spatial attention, cross attention, and feedforward network, as shown in Fig. 3(b)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.713, + 0.892, + 0.805 + ], + "angle": 0, + "content": "Zero Initialization. Similar to the previous practices [1, 52], for each block that is newly introduced in stage two, we initialized all parameters of its final layer as zero. This avoids disrupting the prior knowledge of the well-trained image generation model in the beginning and stabilizes the training process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Training. GenAD is trained to predict the future by jointly denoising from the noisy latents with the guidance of past frames and text conditions. We first project \\( T \\) consecutive frames of a video clip into a batch of latents \\( \\mathbf{v} = \\{\\mathbf{v}^m, \\mathbf{v}^n\\} \\), where the leading \\( m \\) frame latents \\( \\mathbf{v}^m \\) are clean, representing historical observations, and other \\( n = T - m \\) frame latents" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14666" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.182 + ], + "angle": 0, + "content": "\\(\\mathbf{v}^n\\) indicate the future to be predicted. \\(\\mathbf{v}^n\\) are then corrupted to \\(\\mathbf{v}_t^n\\) by the forward diffusion process, where \\(t\\) indexes a randomly sampled noise scale. The model is trained to predict the noise of \\(\\mathbf{v}_t^n\\) conditioned on observations \\(\\mathbf{v}^m\\) and text c. The learning objective of the video prediction model is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.189, + 0.47, + 0.217 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {v i d}} := \\mathbb {E} _ {\\mathbf {v}, \\boldsymbol {\\epsilon} \\sim \\mathcal {N} (0, 1), \\mathbf {c}, t} \\left[ \\| \\boldsymbol {\\epsilon} - \\mathbf {f} _ {\\theta , \\phi} \\left(\\mathbf {v} _ {t} ^ {n}; \\mathbf {v} ^ {m}, \\mathbf {c}, t\\right) \\| _ {2} ^ {2} \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.225, + 0.469, + 0.348 + ], + "angle": 0, + "content": "where \\(\\theta\\) denotes the inherited stage-one model and \\(\\phi\\) represents the newly inserted temporal reasoning blocks. Following [4], we freeze \\(\\theta\\) and only train the temporal reasoning blocks to avoid perturbing the generation ability of the image generation model and focus on learning temporal dependencies in videos. Notably, only the outputs from the corrupted frames \\(\\mathbf{v}_t^n\\) contribute to the training loss while those from condition frames \\(\\mathbf{v}^m\\) are ignored." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.356, + 0.198, + 0.37 + ], + "angle": 0, + "content": "3.3. Extensions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.379, + 0.469, + 0.47 + ], + "angle": 0, + "content": "Relying on the well-trained video prediction capability in driving scenarios, we further exploit the potential of the pretrained model in action-controlled prediction and planning, which are important for real-world driving systems. Here, we explore the downstream tasks on nuScenes [6] which provides recorded poses." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.473, + 0.469, + 0.609 + ], + "angle": 0, + "content": "Action-conditioned Prediction. To make our predictive model controllable with exact ego actions and act as a simulator [25], we fine-tune the model with the paired future trajectory as an additional condition. Specifically, we map the raw trajectory to a high-dimensional feature with Fourier embeddings [44]. After further projection by a linear layer, it is added to the original conditions. Thus, the ego actions are injected into the network through the conditional cross-attention layer in Fig. 3(b)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.613, + 0.469, + 0.794 + ], + "angle": 0, + "content": "Planning. By learning to predict the future, GenAD acquires strong representations of complex driving scenes, which can be further exploited for planning. Specifically, we extract spatiotemporal features of two historical frames through the UNet encoder of the frozen GenAD, which is nearly half the size of the entire model, and feed them to a multi-layer perceptron (MLP) to predict future waypoints. With the frozen GenAD encoder and a learnable MLP layer, the training process of our planner can be sped up by 3400 times compared to an end-to-end planning model UniAD [22], validating the effectiveness of the learned spatiotemporal feature of GenAD." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.808, + 0.21, + 0.824 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.833, + 0.267, + 0.849 + ], + "angle": 0, + "content": "4.1. Setup and Protocols" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.902 + ], + "angle": 0, + "content": "GenAD is learned in two stages on OpenDV-2K but with different learning objectives (in Sec. 3) and input formats. In stage one, the model takes input (image, text) pairs and" + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.094, + 0.899, + 0.2 + ], + "angle": 0, + "content": "
MethodTraining DatasetPred.nuScenes
FID (↓)FVD (↓)
DriveGAN [25]✓73.4502
DriveDreamer* [45]nuScenes✓52.6452
DrivingDiffusion* [31]✗15.8332
GenAD-nus (Ours)nuScenes✓15.4244
GenAD (Ours)OpenDV-2K✓15.4184
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.21, + 0.892, + 0.253 + ], + "angle": 0, + "content": "Table 2. Video generation quality compared to state-of-the-arts trained on nuScenes. \"Pred.\": evaluation by future prediction. *: requiring 3D layout inputs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.271, + 0.892, + 0.498 + ], + "angle": 0, + "content": "is trained on text-to-image generation. We broadcast the command annotation, which is labeled for each 4s video sequence, to all frames included. The model is trained for \\(300\\mathrm{K}\\) iterations on 32 NVIDIA Tesla A100 GPUs with a total batch size of 256. In the second stage, GenAD is trained to jointly denoise future latents conditioned on past latents and texts. Its inputs are (video clip, text) pairs where each video clip is 4s at \\(2\\mathrm{Hz}\\). The current version of GenAD is trained on 64 GPUs for \\(112.5\\mathrm{K}\\) iterations with a total batch size of 64. The input frames are resized to \\(256 \\times 448\\) for training in both stages, and the text condition c is dropped at a probability of \\(p = 0.1\\) to enable classifier-free guidance [17] in sampling, which is commonly used in diffusion models to improve sample quality. More training and sampling details are in Appendix D." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.509, + 0.844, + 0.526 + ], + "angle": 0, + "content": "4.2. Results of Video Prediction Pre-training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.533, + 0.892, + 0.744 + ], + "angle": 0, + "content": "Comparison to Recent Video Generation Approaches. We compare GenAD to recent advances on an unseen set with geofencing from OpenDV-YouTube, Waymo [43], KITTI [14], and Cityscapes [10] in a zero-shot generation manner. Fig. 4 depicts the qualitative results. Image-to-video models I2VGen-XL [53] and VideoCrafter1 [8] can not strictly follow the given frames to make predictions, yielding poor consistency between the predicted frames and past frames. The video prediction model DMVFN [21] that is trained on Cityscapes suffers from the unfavorable shape distortions in its predictions, especially on the three unseen datasets. In contrast, GenAD exhibits remarkable zero-shot generalization ability and visual quality although none of these sets are included in the training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Comparison to nuScenes Experts. We also compare GenAD with the most recent available driving video generation models which are exclusively trained for nuScenes. Tab. 2 shows that GenAD surpasses all previous methods in both image fidelity (FID) and video coherence (FVD). Specifically, GenAD significantly reduces FVD by \\(44.5\\%\\) compared to DrivingDiffusion [31], without taking 3D future layouts as additional inputs. For fair comparisons, we train a model variant (GenAD-nus) on nuScenes dataset only. We find that although GenAD-nus performs on par" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14667" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.089, + 0.53, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.089, + 0.891, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.462, + 0.893, + 0.503 + ], + "angle": 0, + "content": "Figure 4. Task on zero-shot video prediction for unseen scenarios. We show the generation results (in blue boxes) of different models given the same starting frames. GenAD makes more robust, realistic, and reasonable future predictions on unseen datasets (scenarios). More comparisons and visualizations are shown in Appendix." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.526, + 0.469, + 0.707 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.717, + 0.47, + 0.76 + ], + "angle": 0, + "content": "Figure 5. Task on language-conditioned prediction. Given two frames of a rainy scenario in the intersection and three high-level text conditions, GenAD simulates reasonable futures accordingly." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.792, + 0.469, + 0.868 + ], + "angle": 0, + "content": "with GenAD on nuScenes, it struggles to generalize to unseen datasets like Waymo, where the generation degrades to the nuScenes visual pattern. In contrast, GenAD trained on OpenDV-2K exhibits strong generalization ability across datasets as shown in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "We provide language-conditioned prediction samples on nuScenes in Fig. 5, where GenAD simulates various futures" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.53, + 0.893, + 0.59 + ], + "angle": 0, + "content": "from the same start following different textual instructions. The impressive generation quality is exhibited in the intricate details of the environment, and the natural transition of ego motion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.595, + 0.893, + 0.807 + ], + "angle": 0, + "content": "Ablation Study. We perform ablations by training each variant on a subset of OpenDV-2K for 75K steps. Starting from the baseline with plain temporal attentions [4, 18], we gradually introduce our proposed components. Notably, by interleaving the temporal blocks with the spatial blocks, the FVD significantly improves \\((-17\\%)\\) due to more sufficient spatiotemporal interactions. Both temporal causality and decoupled spatial attention contribute to better CLIP-SIM, improving the temporal consistency between future predictions and the condition frames. To be clear, the slight increase in FID and FVD, shown in fourth and third rows of Tab. 3 respectively, does not faithfully reflect a decline in generation quality as discussed in [4, 5, 37]. The effectiveness of each design is shown in Fig. 6." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.817, + 0.7, + 0.832 + ], + "angle": 0, + "content": "4.3. Results of Extensions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Action-conditioned Prediction. We further showcase the performance of the action-conditioned model fine-tuned on nuScenes, GenAD-act, in Fig. 7 and Tab. 4. Given two starting frames and a trajectory w composed of 6 future way" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "14668" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.468, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.355, + 0.47, + 0.385 + ], + "angle": 0, + "content": "Figure 6. Case study for model designs. All components help alleviate artifacts and improve the consistency of future predictions." + }, + { + "type": "table", + "bbox": [ + 0.083, + 0.399, + 0.462, + 0.487 + ], + "angle": 0, + "content": "
MethodYouTube
FID (↓)FVD (↓)CLIPSIM (↑)
Baseline18.32244.440.8405
+ Deep Interaction17.96201.690.8409
+ Temporal Causality16.54207.450.8550
+ Decoupled Spatial Attn.17.67189.540.8652
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.49, + 0.47, + 0.519 + ], + "angle": 0, + "content": "Table 3. Ablation on model designs in GenAD. All proposed designs contribute to the final performance." + }, + { + "type": "table", + "bbox": [ + 0.109, + 0.534, + 0.436, + 0.61 + ], + "angle": 0, + "content": "
MethodConditionnuScenes\nAction Prediction Error (↓)
Ground truth-0.90
GenADtext2.54
GenAD-acttext + traj.2.02
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.614, + 0.47, + 0.656 + ], + "angle": 0, + "content": "Table 4. Task on action-conditioned prediction. Compared to GenAD with text conditions only, GenAD-act enables more precise future predictions that follow the action condition." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.684, + 0.47, + 0.851 + ], + "angle": 0, + "content": "points, GenAD-act imagines 6 future frames following the trajectory sequence. To evaluate the consistency between the input trajectory \\(\\mathbf{w}\\) and predicted frames, we establish an inverse dynamics model (IDM) on nuScenes as the evaluator, which projects a video sequence into a corresponding ego trajectory. We leverage the IDM to translate predicted frames into the trajectory \\(\\hat{\\mathbf{w}}\\), and calculate the L2 distance between \\(\\mathbf{w}\\) and \\(\\hat{\\mathbf{w}}\\) as the Action Prediction Error. Specifically, GenAD-act substantially reduces the Action Prediction Error by \\(20.4\\%\\) compared to GenAD with text condition, allowing for more accurate future simulations." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Planning Results. Tab. 5 depicts the planning results on nuScenes where ground truth poses for the ego vehicle are available. By freezing GenAD encoder and only optimizing" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.093, + 0.892, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.287, + 0.892, + 0.358 + ], + "angle": 0, + "content": "Figure 7. Task on action-conditioned prediction (simulation). Given the same starting frames and different future trajectories (shown in yellow dots in the first column), GenAD-act can simulate diverse futures following different ego intentions. More visualizations are in Appendix." + }, + { + "type": "table", + "bbox": [ + 0.54, + 0.384, + 0.851, + 0.466 + ], + "angle": 0, + "content": "
Method# Trainable Params.nuScenes
ADE (↓)FDE (↓)
ST-P3* [20]10.9M2.653.73
UniAD* [22]58.8M1.031.65
GenAD (Ours)0.8M1.232.31
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.469, + 0.892, + 0.526 + ], + "angle": 0, + "content": "Table 5. Task on open-loop planning. A lightweight MLP with frozen GenAD gets competitive planning results with \\(73 \\times\\) fewer trainable parameters and front-view image alone. *: multi-view inputs. Evaluation protocols are aligned with UniAD [22]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.565, + 0.892, + 0.657 + ], + "angle": 0, + "content": "an additional MLP on top of it, the model can effectively learn to plan. Notably, by pre-extracting image features through the UNet encoder of GenAD, the entire learning process for planning adaptation takes only 10 minutes on a single NVIDIA Tesla V100 device, which is 3400 times more efficient than the training of the UniAD planner [22]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.688, + 0.753, + 0.705 + ], + "angle": 0, + "content": "5. Limitations and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We study the system-level development of GenAD, a large-scale generalized video predictive model for autonomous driving. We also validate the adaptation of the learned representation of GenAD to driving tasks, i.e., learning a \"world model\" and motion planning. Although we obtain improved generalization to open domains, the increased model capacity poses challenges in both training efficiency and real-time deployment. We envision the unified video prediction task will serve as a scalable objective for future research on representation learning and policy learning. Another interesting direction involves distilling the encoded knowledge for a wider range of downstream tasks [29]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14669" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.253 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andy Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.256, + 0.471, + 0.325 + ], + "angle": 0, + "content": "[2] Mohammadhossein Bahari, Saeed Saadatnejad, Ahmad Rahimi, Mohammad Shaverdikondori, Amir Hossein Shahidzadeh, Seyed-Mohsen Moosavi-Dezfooli, and Alexandre Alahi. Vehicle trajectory prediction works, but not everywhere. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.328, + 0.47, + 0.37 + ], + "angle": 0, + "content": "[3] Fan Bao, Chongxuan Li, Jiacheng Sun, and Jun Zhu. Why are conditional generative models better than unconditional ones? arXiv preprint arXiv:2212.00362, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.372, + 0.47, + 0.427 + ], + "angle": 0, + "content": "[4] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.429, + 0.47, + 0.485 + ], + "angle": 0, + "content": "[5] Tim Brooks, Janne Hellsten, Miika Aittala, Ting-Chun Wang, Timo Aila, Jaakko Lehtinen, Ming-Yu Liu, Alexei Efros, and Tero Karras. Generating long videos of dynamic scenes. In NeurIPS, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.487, + 0.47, + 0.556 + ], + "angle": 0, + "content": "[6] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yuxin Pan, Giancarlo Baldan, and Oscar Beijbom. nuScenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.558, + 0.47, + 0.628 + ], + "angle": 0, + "content": "[7] Holger Caesar, Juraj Kabzan, Kok Seang Tan, Whye Kit Fong, Eric M. Wolff, Alex Lang, Luke Fletcher, Oscar Bei-jbom, and Sammy Omari. nuPlan: A closed-loop ml-based planning benchmark for autonomous vehicles. In CVPR Workshops, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.631, + 0.47, + 0.712 + ], + "angle": 0, + "content": "[8] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, Chao Weng, and Ying Shan. VideoCrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.716, + 0.47, + 0.77 + ], + "angle": 0, + "content": "[9] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers. arXiv preprint arXiv:2306.16927, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.773, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[10] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The Cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.845, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[11] Yilun Dai, Mengjiao Yang, Bo Dai, Hanjun Dai, Ofir Nachum, Josh Tenenbaum, Dale Schuurmans, and Pieter Abbeel. Learning universal policies via text-guided video generation. In NeurIPS, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[12] Thierry Deruyttere, Simon Vandenhende, Dusan Grujicic, Luc Van Gool, and Marie-Francine Moens. Talk2Car: Taking control of your self-driving car. In EMNLP, 2019. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.188 + ], + "angle": 0, + "content": "[13] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.19, + 0.892, + 0.231 + ], + "angle": 0, + "content": "[14] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The KITTI dataset. *IJRR*, 2013. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.232, + 0.892, + 0.274 + ], + "angle": 0, + "content": "[15] Agrim Gupta, Stephen Tian, Yunzhi Zhang, Jiajun Wu, Roberto Martin-Martin, and Li Fei-Fei. MaskViT: Masked visual pre-training for video prediction. In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.275, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[16] Niklas Hanselmann, Katrin Renz, Kashyap Chitta, Apratim Bhattacharyya, and Andreas Geiger. KING: Generating safety-critical driving scenarios for robust imitation via kinematics gradients. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.33, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[17] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS Workshops, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.358, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[18] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. In NeurIPS, 2022. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.4, + 0.892, + 0.466 + ], + "angle": 0, + "content": "[19] Anthony Hu, Lloyd Russell, Hudson Yeo, Zak Murez, George Fedoseev, Alex Kendall, Jamie Shotton, and Gianluca Corrado. GAIA-1: A generative world model for autonomous driving. arXiv preprint arXiv:2309.17080, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.469, + 0.892, + 0.523 + ], + "angle": 0, + "content": "[20] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. ST-P3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In ECCV, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.525, + 0.892, + 0.566 + ], + "angle": 0, + "content": "[21] Xiaotao Hu, Zhewei Huang, Ailin Huang, Jun Xu, and Shuchang Zhou. A dynamic multi-scale voxel flow network for video prediction. In CVPR, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.567, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[22] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, Lewei Lu, Xiaosong Jia, Qiang Liu, Jifeng Dai, Yu Qiao, and Hongyang Li. Planning-oriented autonomous driving. In CVPR, 2023. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.637, + 0.892, + 0.692 + ], + "angle": 0, + "content": "[23] Fan Jia, Weixin Mao, Yingfei Liu, Yucheng Zhao, Yuqing Wen, Chi Zhang, Xiangyu Zhang, and Tiancai Wang. A driver-i: A general world model for autonomous driving. arXiv preprint arXiv:2311.13549, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.692, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[24] Jinkyu Kim, Teruhisa Miso, Yi-Ting Chen, Ashish Tawari, and John Canny. Grounding human-to-vehicle advice for self-driving vehicles. In CVPR, 2019. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.735, + 0.892, + 0.775 + ], + "angle": 0, + "content": "[25] Seung Wook Kim, Jonah Philion, Antonio Torralba, and Sanja Fidler. DriveGAN: Towards a controllable high-quality neural simulation. In CVPR, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.776, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[26] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[27] Yann LeCun. A path towards autonomous machine intelligence version 0.9. 2, 2022-06-27. Open Review, 62, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[28] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. Nature, 2015. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14670" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[29] Daiqing Li, Huan Ling, Amlan Kar, David Acuna, Seung Wook Kim, Karsten Kreis, Antonio Torralba, and Sanja Fidler. DreamTeacher: Pretraining image backbones with deep generative models. In ICCV, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.148, + 0.47, + 0.203 + ], + "angle": 0, + "content": "[30] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.204, + 0.471, + 0.259 + ], + "angle": 0, + "content": "[31] Xiaofan Li, Yifu Zhang, and Xiaoqing Ye. DrivingDiffusion: Layout-guided multi-view driving scene video generation with latent diffusion model. arXiv preprint arXiv:2310.07771, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.26, + 0.47, + 0.315 + ], + "angle": 0, + "content": "[32] Zhenyu Li, Zehui Chen, Ang Li, Liangji Fang, Qinhong Jiang, Xianming Liu, and Junjun Jiang. Unsupervised domain adaptation for monocular 3d object detection via self-training. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.316, + 0.47, + 0.372 + ], + "angle": 0, + "content": "[33] Jiachen Lu, Ze Huang, Jiahui Zhang, Zeyu Yang, and Li Zhang. Wovogen: World volume-aware diffusion for controllable multi-camera driving scene generation. arXiv preprint arXiv:2312.02934, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.372, + 0.47, + 0.44 + ], + "angle": 0, + "content": "[34] Jiageng Mao, Minzhe Niu, Chenhan Jiang, Hanxue Liang, Xiaodan Liang, Yamin Li, Chao Ye, Wei Zhang, Zhenguo Li, Jie Yu, Hang Xu, and Chunjing Xu. One million scenes for autonomous driving: ONCE dataset. In NeurIPS Datasets and Benchmarks, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.441, + 0.47, + 0.468 + ], + "angle": 0, + "content": "[35] OpenAI. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.47, + 0.566 + ], + "angle": 0, + "content": "[36] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul F Christiano, Jan Leike, and Ryan Lowe. Training language models to follow instructions with human feedback. In NeurIPS, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.47, + 0.635 + ], + "angle": 0, + "content": "[37] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.636, + 0.47, + 0.677 + ], + "angle": 0, + "content": "[38] Xiaojuan Qi, Zhengzhe Liu, Qifeng Chen, and Jiaya Jia. 3D motion decomposition for rgbd future dynamic scene synthesis. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.678, + 0.47, + 0.761 + ], + "angle": 0, + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.47, + 0.816 + ], + "angle": 0, + "content": "[40] Vasili Ramanishka, Yi-Ting Chen, Teruhisa Mitsu, and Kate Saenko. Toward driving scene understanding: A dataset for learning driver behavior and causal reasoning. In CVPR, 2018. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.47, + 0.859 + ], + "angle": 0, + "content": "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[42] Peter Shaw, Jakob Uszkoreit, and Ashish Vaswani. Self-attention with relative position representations. arXiv preprint arXiv:1803.02155, 2018. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[43] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott M. Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.204, + 0.892, + 0.274 + ], + "angle": 0, + "content": "[44] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. In NeurIPS, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.276, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[45] Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, and Jiwen Lu. DriveDreamer: Towards real-world-driven world models for autonomous driving. arXiv preprint arXiv:2309.09777, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.333, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[46] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, Yuwei Guo, Tianxing Wu, Chenyang Si, Yuming Jiang, Cunjian Chen, Chen Change Loy, Bo Dai, Dahua Lin, Yu Qiao, and Ziwei Liu. LAVIE: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.432, + 0.892, + 0.5 + ], + "angle": 0, + "content": "[47] Yuqi Wang, Jiawei He, Lue Fan, Hongxin Li, Yuntao Chen, and Zhaoxiang Zhang. Driving into the future: Multiview visual forecasting and planning with world model for autonomous driving. arXiv preprint arXiv:2311.17918, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.503, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[48] Benjamin Wilson, William Qi, Tanmay Agarwal, John Lambert, Jagjeet Singh, Siddhesh Khandelwal, Bowen Pan, Ratnesh Kumar, Andrew Hartnett, Jhony Kaesemodel Pontes, Deva Ramanan, Peter Carr, and James Hays. Argoverse 2: Next generation datasets for self-driving perception and forecasting. In NeurIPS Datasets and Benchmarks, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.589, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[49] Daniel M Wolpert, Zoubin Ghahramani, and Michael I Jordan. An internal model for sensorimotor integration. Science, 1995. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.632, + 0.892, + 0.757 + ], + "angle": 0, + "content": "[50] Zhiheng Xi, Wenxiang Chen, Xin Guo, Wei He, Yiwen Ding, Boyang Hong, Ming Zhang, Junzhe Wang, Senjie Jin, Enyu Zhou, Rui Zheng, Xiaoran Fan, Xiao Wang, Liao Xiong, Yuhao Zhou, Weiran Wang, Changhao Jiang, Yicheng Zou, Xiangyang Liu, Zhangyue Yin, Shihan Dou, Rongxiang Weng, Wensen Cheng, Qi Zhang, Wenjuan Qin, Yongyan Zheng, Xipeng Qiu, Xuanjing Huang, and Tao Gui. The rise and potential of large language model based agents: A survey. arXiv preprint arXiv:2309.07864, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.758, + 0.892, + 0.827 + ], + "angle": 0, + "content": "[51] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. arXiv preprint arXiv:2309.15818, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.829, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[52] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.872, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[53] Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qin, Xiang Wang, Deli Zhao, and" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14671" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "Jingren Zhou. I2VGen-XL: High-quality image-to-video synthesis via cascaded diffusion models. arXiv preprint arXiv:2311.04145, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.469, + 0.164 + ], + "angle": 0, + "content": "[54] Ruizhao Zhu, Peng Huang, Eshed Ohn-Bar, and Venkatesh Saligrama. Learning to drive anywhere. In CoRL, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "14672" + } + ] +] \ No newline at end of file diff --git a/2024/Generalized Predictive Model for Autonomous Driving/554856ec-712f-446a-b09c-0cce647f2af6_origin.pdf b/2024/Generalized Predictive Model for Autonomous Driving/554856ec-712f-446a-b09c-0cce647f2af6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d87b798518a9f9f2ce0fc32fc692da326182eccd --- /dev/null +++ b/2024/Generalized Predictive Model for Autonomous Driving/554856ec-712f-446a-b09c-0cce647f2af6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c61fb7b193069dd5f913a43ace257728eca87bf6ab89e774fa268ce66ee4310d +size 8674755 diff --git a/2024/Generalized Predictive Model for Autonomous Driving/full.md b/2024/Generalized Predictive Model for Autonomous Driving/full.md new file mode 100644 index 0000000000000000000000000000000000000000..421fc7bc5bf9a77e1af9472595d0b3caf7234515 --- /dev/null +++ b/2024/Generalized Predictive Model for Autonomous Driving/full.md @@ -0,0 +1,283 @@ +# Generalized Predictive Model for Autonomous Driving + +Jiazhi Yang $^{1*}$ Shenyuan Gao $^{2,1*}$ Yihang Qiu $^{1*}$ Li Chen $^{3,1\dagger}$ Tianyu Li $^{1}$ Bo Dai $^{1}$ Kashyap Chitta $^{4,5}$ Penghao Wu $^{1}$ Jia Zeng $^{1}$ Ping Luo $^{3}$ Jun Zhang $^{2\sharp}$ Andreas Geiger $^{4,5\sharp}$ Yu Qiao $^{1\sharp}$ Hongyang Li $^{1\dagger}$ + +1 OpenDriveLab and Shanghai AI Lab 2 Hong Kong University of Science and Technology 3 University of Hong Kong 4 University of Tübingen 5 Tübingen AI Center + +![](images/b80cd5b5f461d4fd4cdc6680a5a80f78390f39e532e73a9719c51b85443b4c30.jpg) + +# OpenDV-2K + +![](images/cb2a404a9278330cd862b1bc1b938fabc376187acc9fda4362a08d287d0385e6.jpg) + +![](images/6d108f23ef58e3208f38828182bba2975dc9d3a31d659c14a88aab89bfc6b2be.jpg) + +# GenAD + +![](images/e19f39be9e834ce2ba2f960545759988df3469bfad8ed6cbdae86f58e2f9fa22.jpg) + +![](images/0cd856b96f349dfbbd8ff9ae0fc5fa0e9a187141fda5adc06125fa773e8432f4.jpg) + +# Tasks + +![](images/48201f52fac94aa848c29342b5689af78076b4daa7a33a69e107d21e971a9a26.jpg) + +![](images/57270d275699fef0f89f2a74725ff5f9947d6f9c8817e08dadcc34b4f19bb07a.jpg) +2. Language-conditioned Prediction + +![](images/51c6bdb6e97a0982e4bf55874a52e1dc0df0e46ab56367adff2a33661bd99bdf.jpg) +3. Action-conditioned Prediction + +![](images/57aaf11851c35fa0fdae05e906b224d79236dd3965619e6e637c61b7175cad81.jpg) +4. Planning +Figure 1. Overview of the GenAD paradigm. We aim to establish a generalized video prediction paradigm for autonomous driving by presenting the largest multimodal driving video dataset to date, OpenDV-2K, and a generative model that predicts the future given past visual and textual input, GenAD. The strong generalization and controllability of GenAD is validated spanning a diverse spectrum of tasks, including zero-shot domain transfer, language-conditioned prediction, action-conditioned prediction, and motion planning. + +# Abstract + +In this paper, we introduce the first large-scale video prediction model in the autonomous driving discipline. To eliminate the restriction of high-cost data collection and empower the generalization ability of our model, we acquire massive data from the web and pair it with diverse and high-quality text descriptions. The resultant dataset accumulates over 2000 hours of driving videos, spanning areas all over the world with diverse weather conditions and traffic scenarios. Inheriting the merits from recent la + +tent diffusion models, our model, dubbed GenAD, handles the challenging dynamics in driving scenes with novel temporal reasoning blocks. We showcase that it can generalize to various unseen driving datasets in a zero-shot manner, surpassing general or driving-specific video prediction counterparts. Furthermore, GenAD can be adapted into an action-conditioned prediction model or a motion planner, holding great potential for real-world driving applications. + +# 1. Introduction + +Autonomous driving agents, as a promising application of high-level artificial intelligence, perceive the surrounding + +environment, build internal world model representations, make decisions, and take actions in response [9, 50]. However, despite dedicated efforts in academia and industry for decades, their deployment is still restricted to certain areas or scenarios, and they cannot be applied over the world seamlessly. One critical reason is the limited generalization ability of learned models in structured autonomous driving systems. Typically, perception models face challenges of generalizing to diverse environments with changes in geographical locations, sensor configurations, weather conditions, open-set objects, etc.; prediction and planning models fail to generalize to nondeterministic futures with rare scenarios and different driving intentions [2, 16, 54]. + +Motivated by how humans learn to perceive and cognize the world [27, 28, 49], we advocate employing driving videos as the universal interface that generalizes to diverse environments with dynamic futures. Based on this, a driving video predictive model is preferred to fully capture the world knowledge about driving scenarios (Fig. 1). By predicting the future, the video predictor essentially learns two vital aspects of autonomous driving: how the world operates, and how to maneuver safely in the wild. + +Recently, the community has begun to adopt video as the interface to represent observation behavior and action for various robot tasks [11]. For domains such as classical video prediction and robotics, the video backgrounds are mostly static, the movement of robots is slow, and the resolution of videos is low. In contrast, for the driving scenarios, it struggles with outdoor environments being highly dynamic, agents encompassing much larger motions, and the sensory resolution covering a large range of view. These distinctions lead to substantial challenges for autonomous driving applications. Fortunately, there are some preliminary attempts on developing a video predictive model in the driving domain [4, 15, 19, 23, 25, 33, 38, 45, 47]. Despite promising progress in terms of prediction quality, these attempts have not achieved desirable capability of generalization as in classical robot tasks (e.g., manipulation), being confined to either limited scenarios such as highways with low traffic density [4] and small-scale datasets [15, 23, 33, 45, 47], or restricted conditions that raises difficulties to generate diverse environments [38]. How to unveil the potential of video prediction models for driving remains seldom explored. + +Motivated by the discussions above, we target at building a video predictive model for autonomous driving, capable of generalizing to new conditions and environments. To this end, we have to answer the following questions: (1) What data can be obtained in a feasible and scalable manner? (2) How can we formulate a predictive model to capture the complex evolution of dynamic scenarios? (3) How can we apply the (foundation) model for downstream tasks? + +Scaled Data. To achieve powerful generalization ability, a + +substantial and diverse corpus of data is necessary. Inspired by the success of learning from Internet-scale data in foundation models [1, 26, 39], we construct our driving dataset from both the web and publicly licensed datasets. Compared to existing options, which are limited in scale and diversity due to their regulated collection processes, online data owns great diversity in several aspects: geographic locations, terrains, weather conditions, safety-critical scenarios, sensor settings, traffic elements, etc. To guarantee the data is of high-quality and desirable for large-scale training, we exhaustively collect driving recordings on YouTube and remove unintended corruption frames via rigorous human verification. Furthermore, videos are paired with diverse text-level conditions, including descriptions generated and refined with the aid of existing foundation models [30, 35], and high-level instructions inferred by a video classifier. Through these steps, we construct OpenDV-2K, the largest public driving dataset to date, containing more than 2000 hours of driving videos and being 374 times larger than the widely used nuScenes counterpart. Our dataset is publicly available at https://github.com/OpenDriveLab/DriveAGI. + +Generalized Predictive Model. Learning a generalized driving video predictor bears several key challenges: generation quality, training efficiency, causal reasoning, and drastic view shift. We address these aspects by presenting a novel temporal generative model with two-stage learning. To capture the environment details, enhance generation quality, and maintain training efficiency simultaneously, we build upon the recent success of latent diffusion models (LDMs) [37, 41]. In the first stage, we transfer the generation distribution of LDM from its pre-trained general vision domain to the driving domain by fine-tuning it on OpenDV2K images. In the second stage, we interleave the proposed temporal reasoning blocks into the original model and learn to predict the future given past frames and conditions. Contrary to conventional temporal modules [4, 18] that suffer from causal confusion and large motion, our solution consists of causal temporal attention and decoupled spatial attention to efficiently model the drastic spatiotemporal shift in highly dynamic driving scenes. After sufficient training, our Generative model for Autonomous Driving (GenAD)1 can generalize to various scenarios in a zero-shot fashion. + +Extensions for Simulation and Planning. After large-scale pre-training of video prediction, GenAD essentially understands how the world evolves and how to drive. We show how to adapt its learned knowledge for real-world driving problems, i.e., simulation and planning. For simulation, we fine-tune the pre-trained model with future ego trajectories as additional conditions, to associate future imaginations with different ego actions. We also empower + +
DatasetDuration (hours)Front-view FramesGeographic Diversity CountriesCitiesSensor Setup
XKITTI [14]1.415k11fixed
XCityscapes [10]0.525k350fixed
XWaymo Open* [43]11390k13fixed
XArgoverse 2* [48]4.2300k16fixed
✓nuScenes [6]5.5241k22fixed
✓nuPlan* [7]1204.0M24fixed
✓Talk2Car [12]4.7-22fixed
✓ONCE [34]1447M1-fixed
✓Honda-HAD [24]321.2M1-fixed
✓Honda-HDD-Action [40]1041.1M1-fixed
✓Honda-HDD-Cause [40]32-1-fixed
✓OpenDV-YouTube (Ours)174760.2M≥40†≥244†uncalibrated
-OpenDV-2K (Ours)205965.1M≥40†≥244†uncalibrated
+ +Table 1. OpenDV-2K comparison at a glance to existing counterparts in terms of scale and diversity. Note that datasets with $\checkmark$ are included in OpenDV-2K (last row). \*Perception subset in Waymo Open, Argoverse 2, and nuPlan. Estimated by GPT [36] from video titles. + +![](images/3415a03895c87d97d29527857e876efbd9de7ba7bae8be4262d02422b3c74959.jpg) +Figure 2. Geographic distribution of OpenDV-2K. Our dataset covers ample driving scenarios around the world. + +GenAD to perform planning on challenging benchmarks by using a lightweight planner to translate latent features into the future trajectory of the ego vehicle. On account of its pre-trained ability to predict accurate future frames, our algorithm exhibits promising results in both simulation consistency and planning reliability. + +# 2. OpenDV-2K Dataset + +We introduce OpenDV-2K, a large-scale multimodal dataset for autonomous driving, to support the training of a generalized video prediction model. The main component is a vast corpus of high-quality YouTube driving videos, which are collected from all over the world, and are gathered into our dataset after a careful curation process. We automatically create language annotations for these videos using vision-language models. To further improve its diversity in sensor configurations and language expressions, we merge 7 publicly licensed datasets into our OpenDV-2K, as shown in Tab. 1. As a result, OpenDV-2K occupies a total of 2059 hours of videos paired with texts, including 1747 hours from YouTube and 312 hours from public datasets. We use OpenDV-YouTube and OpenDV-2K to specify the YouTube split and the overall dataset, respectively. + +# 2.1. Diversity over Prior Datasets + +A brief comparison with other public datasets is provided in Tab. 1. Beyond its significant scale, the proposed OpenDV-2K represents diversity across various aspects as follows. + +Globe-wise Geographic Distribution. Due to the global nature of online videos, OpenDV-2K covers more than 40 countries and 244 cities worldwide. This is a tremendous improvement over previous public datasets, which are typically gathered in a small number of restricted areas. We plot the specific distribution of OpenDV-YouTube in Fig. 2. + +Open-world Driving Scenarios. Our dataset provides a huge amount of realistic driving experience in the open world, covering rare environments like forests, extreme weather conditions like heavy snow, and appropriate driving behaviors in response to interactive traffic situations. These data are crucial for diversity and generalization yet are seldom collected in existing public datasets. + +Unrestricted Sensor Configurations. Current driving datasets are confined to specific sensor configurations, including intrinsic and extrinsic camera parameters, image, sensor type, optics, etc., which poses great challenges for deploying the learned models with different sensors [32]. In contrast, YouTube driving videos are recorded in various types of vehicles with flexible camera setups, which aids in the robustness of the trained model when deployed using a novel camera setting. + +# 2.2. Towards High-quality Multimodal Dataset + +Driving Video Collection and Curation. Finding clean driving videos from the vast pool of the web is a tedious and costly task. To simplify the process, we start by selecting certain video uploaders, i.e., YouTubers. Judging from the average length and overall quality, we collect 43 YouTubers with 2139 high-quality front-view driving videos. To make sure there is no overlap between training and validation sets, we take all videos from 3 YouTubers for validation, with the remaining videos as the training set. To rule out non-driving frames like video introductions and subscription reminders, we discard a certain length of segments at the beginning and end of each video. Each frame is then described with language contexts using a VLM model, BLIP-2 [30]. We further remove the black frames and transition frames, which are not ideal for training, by manually checking if there are certain keywords in these contexts. We give an illustration + +of the dataset construction pipeline in Appendix C.1.1, and we introduce how to generate the contexts below. + +Language Annotation for YouTube Videos. To create a predictive model that can be controlled by natural language to simulate different futures accordingly, To make the predictive model controllable and improve the sample quality [3], it is crucial to pair the driving videos with meaningful and varied language annotations. We construct two types of texts for OpenDV-YouTube, i.e., driving commands for ego-vehicle and frame descriptions, namely "command" and "context", to help the model comprehend ego actions and open-world concepts, respectively. For commands, we train a video classifier on Honda-HDD-Action [40] for 14 types of actions to label ego behaviors in a 4s sequence. These categorical commands will be further mapped to multiple free-form expressions from a predefined dictionary. For contexts, we leverage an established vision-language model, BLIP-2 [30], to describe the main objects and scenarios for each frame. For more details on annotations, please refer to Appendix C.1.2. + +Enlarging Language Spectrum with Public Datasets. Considering that BLIP-2 annotations are generated for static frames without comprehension of dynamic driving scenarios such as the traffic light transitions, we exploit several public datasets that provide linguistic descriptions for driving scenarios [6, 7, 12, 24, 34, 40]. However, their metadata is relatively sparse with only a few words such as "sunny road". We further enhance their text quality using GPT [36] to form a descriptive "context" and generate a "command" by categorizing the logged trajectory for each video clip. Ultimately, we integrate these datasets with OpenDV-YouTube to establish OpenDV-2K dataset, as shown in the last row of Tab. 1. + +# 3. GenAD Framework + +In this section, we introduce the training and design of the GenAD model. As shown in Fig. 3, GenAD is trained in two stages, i.e., image domain transferring and video prediction pre-training. The first stage adapts the general text-to-image model to the driving domain (Sec. 3.1). The second stage lifts the text-to-image model to a video prediction model with our proposed temporal reasoning block and modified training schemes (Sec. 3.2). In Sec. 3.3, we explore how the predictive model can be extended to action-conditioned prediction and planning. + +# 3.1. Image Domain Transfer + +On-board cameras capture a large field of views with abundant visual contents, including the road, background buildings, surrounding vehicles, etc., which require strong and robust generation capability to produce continuous and realistic driving scenarios. To facilitate the learning process, we + +start with independent image generation in the first stage. Concretely, we initialize our model with SDXL [37], which is a large-scale latent diffusion model (LDM) for text-to-image generation, to leverage its ability to synthesize high-quality images with plenty of visual details. It is implemented as a denoising UNet $\mathbf{f}_{\theta}$ with several stacked convolution and attention blocks, which learns to synthesize images by denoising the noisy latents [41]. Specifically, given a noisy input latent $\mathbf{x}_t$ corrupted by the forward diffusion process, it is trained to predict the added noise $\epsilon$ of $\mathbf{x}_t$ via the following objective: + +$$ +\mathcal {L} _ {\mathrm {i m g}} := \mathbb {E} _ {\mathbf {x}, \epsilon \sim \mathcal {N} (0, 1), \mathbf {c}, t} \left[ \| \epsilon - \mathbf {f} _ {\theta} (\mathbf {x} _ {t}; \mathbf {c}, t) \| _ {2} ^ {2} \right], \tag {1} +$$ + +where $\mathbf{x}$ and $\mathbf{x}_t$ are the clean and noisy latent, respectively, $t$ denotes the timestep for different noise scales, and $\mathbf{c}$ is the text condition that guides the denoising process, which is a concatenation of context and command. For training efficiency, the learning process takes place in a compressed latent space [13, 37, 41] instead of pixel space. During sampling, the model generates images from standard Gaussian noise by denoising the last-step predictions iteratively. + +However, the original SDXL is trained on data in the general domain, such as portraits and artistic paintings, which are not concerned with autonomy systems. To adapt the model to synthesize images for driving, we fine-tune it on text-to-image generation using image-text pairs in OpenDV-2K with the same objective as Eq. (1). Following the original training of SDXL, all parameters $\theta$ of the UNet are fine-tuned at this stage, whereas the CLIP text encoders [39] and the autoencoder [13] remain frozen. + +# 3.2. Video Prediction Pre-training + +In the second stage, with a few frames of a consecutive video as past observations, GenAD is trained to reason about all visual observations and predict several future frames in plausible ways. Similar to the first stage stage, the prediction process can also be guided by text conditions. However, predicting the highly dynamic driving world temporally is challenging due to two fundamental barriers. + +1. Causal Reasoning: To predict plausible futures following the temporal causality of the driving world, the model needs to comprehend the intentions of all other agents together with the ego vehicle, and understand underlying traffic rules, e.g., how the traffic will change with the transition of traffic lights. +2. Drastic View Shift: Contrary to typical video generation benchmarks which mainly have a static background with slow motion of centered objects, the view of driving changes drastically over time. Each pixel in every frame may move to a distant location in the next frame. + +We propose temporal reasoning blocks to address these problems. As illustrated in Fig. 3(c), each block is composed of three successive attention layers, i.e., the causal + +![](images/529d3d44fcae5674be5d92a354b7f27841dd548e493f20c4d81919d42a78b120.jpg) +Figure 3. Framework of GenAD. (a) The two-stage learning for GenAD is composed of transferring the image domain of an image diffusion model to the driving field (a.1 Stage one), and video prediction pre-training for modeling the temporal dependency of videos (a.2 Stage two). (b) One transformer block in GenAD for the second stage training has interleaved temporal reasoning blocks before each frozen layer to align spatiotemporal features. (c) The proposed Temporal Reasoning Block includes one causal temporal attention (TA) and two decoupled spatial attention (SA) layers to extract features in different axes. A query grid attends to itself as well as blue grids while the dark gray grid is masked out in causal attention. 'Zero init' is appended at the end of each attention block to stabilize training. + +![](images/809d540c959eee6b5b0cab1933780f2c041de5aebf8a0957020f426af1a665fd.jpg) + +temporal attention layer and two decoupled spatial attention layers, which are tailored for the causal reasoning and modeling large shifts in the driving scenes, respectively. + +Causal Temporal Attention. Since the model after the stage-one training can only process each frame independently, we leverage temporal attention to exchange information among different video frames. The attention takes place in the time axis and models the temporal dependency of each grid-wise feature. However, directly adapting bidirectional temporal attention here as [4, 18, 46, 51] can hardly acquire the ability of causal reasoning, since the predictions will be inevitably dependent on the subsequent frames instead of past conditions. Therefore, we restrict the attention direction by adding a causal attention mask, as shown in the last row of Fig. 3(c), to encourage the model to fully exploit knowledge from past observations and faithfully reason about the future as if in real-world driving. We empirically found that the causality constraint greatly regularizes the predicted frames to be coherent with past frames. Following common practice, we also add temporal bias implemented as relative position embeddings on the time axis [42] to distinguish different frames of a sequence for temporal attention. + +Decoupled Spatial Attention. As driving videos feature fast perspective changes, features in a specific grid could vary greatly in different timesteps and are hard to correlate and learn by temporal attention, which suffers from a limited receptive field. In light of this, we introduce spa + +tial attention to propagate each grid feature in spatial axes to aid in gathering information for temporal attention. We implement a decoupled variant of self-attention for its efficiency with linear computational complexity, compared to quadratic full self-attention. As shown in Fig. 3(c), the two decoupled attention layers propagate features in horizontal and vertical axes, respectively. + +Deep Interaction. Intuitively, the spatial blocks fine-tuned in stage one refine features of each frame independently towards photorealism, whereas the temporal blocks introduced in stage two align features of all video frames towards coherency and consistency. To further boost the spatiotemporal feature interaction, we interleave the proposed temporal reasoning blocks with the original Transformer blocks in SDXL, i.e., spatial attention, cross attention, and feedforward network, as shown in Fig. 3(b). + +Zero Initialization. Similar to the previous practices [1, 52], for each block that is newly introduced in stage two, we initialized all parameters of its final layer as zero. This avoids disrupting the prior knowledge of the well-trained image generation model in the beginning and stabilizes the training process. + +Training. GenAD is trained to predict the future by jointly denoising from the noisy latents with the guidance of past frames and text conditions. We first project $T$ consecutive frames of a video clip into a batch of latents $\mathbf{v} = \{\mathbf{v}^m, \mathbf{v}^n\}$ , where the leading $m$ frame latents $\mathbf{v}^m$ are clean, representing historical observations, and other $n = T - m$ frame latents + +$\mathbf{v}^n$ indicate the future to be predicted. $\mathbf{v}^n$ are then corrupted to $\mathbf{v}_t^n$ by the forward diffusion process, where $t$ indexes a randomly sampled noise scale. The model is trained to predict the noise of $\mathbf{v}_t^n$ conditioned on observations $\mathbf{v}^m$ and text c. The learning objective of the video prediction model is formulated as follows: + +$$ +\mathcal {L} _ {\mathrm {v i d}} := \mathbb {E} _ {\mathbf {v}, \boldsymbol {\epsilon} \sim \mathcal {N} (0, 1), \mathbf {c}, t} \left[ \| \boldsymbol {\epsilon} - \mathbf {f} _ {\theta , \phi} \left(\mathbf {v} _ {t} ^ {n}; \mathbf {v} ^ {m}, \mathbf {c}, t\right) \| _ {2} ^ {2} \right], \tag {2} +$$ + +where $\theta$ denotes the inherited stage-one model and $\phi$ represents the newly inserted temporal reasoning blocks. Following [4], we freeze $\theta$ and only train the temporal reasoning blocks to avoid perturbing the generation ability of the image generation model and focus on learning temporal dependencies in videos. Notably, only the outputs from the corrupted frames $\mathbf{v}_t^n$ contribute to the training loss while those from condition frames $\mathbf{v}^m$ are ignored. + +# 3.3. Extensions + +Relying on the well-trained video prediction capability in driving scenarios, we further exploit the potential of the pretrained model in action-controlled prediction and planning, which are important for real-world driving systems. Here, we explore the downstream tasks on nuScenes [6] which provides recorded poses. + +Action-conditioned Prediction. To make our predictive model controllable with exact ego actions and act as a simulator [25], we fine-tune the model with the paired future trajectory as an additional condition. Specifically, we map the raw trajectory to a high-dimensional feature with Fourier embeddings [44]. After further projection by a linear layer, it is added to the original conditions. Thus, the ego actions are injected into the network through the conditional cross-attention layer in Fig. 3(b). + +Planning. By learning to predict the future, GenAD acquires strong representations of complex driving scenes, which can be further exploited for planning. Specifically, we extract spatiotemporal features of two historical frames through the UNet encoder of the frozen GenAD, which is nearly half the size of the entire model, and feed them to a multi-layer perceptron (MLP) to predict future waypoints. With the frozen GenAD encoder and a learnable MLP layer, the training process of our planner can be sped up by 3400 times compared to an end-to-end planning model UniAD [22], validating the effectiveness of the learned spatiotemporal feature of GenAD. + +# 4. Experiments + +# 4.1. Setup and Protocols + +GenAD is learned in two stages on OpenDV-2K but with different learning objectives (in Sec. 3) and input formats. In stage one, the model takes input (image, text) pairs and + +
MethodTraining DatasetPred.nuScenes
FID (↓)FVD (↓)
DriveGAN [25]✓73.4502
DriveDreamer* [45]nuScenes✓52.6452
DrivingDiffusion* [31]✗15.8332
GenAD-nus (Ours)nuScenes✓15.4244
GenAD (Ours)OpenDV-2K✓15.4184
+ +Table 2. Video generation quality compared to state-of-the-arts trained on nuScenes. "Pred.": evaluation by future prediction. *: requiring 3D layout inputs. + +is trained on text-to-image generation. We broadcast the command annotation, which is labeled for each 4s video sequence, to all frames included. The model is trained for $300\mathrm{K}$ iterations on 32 NVIDIA Tesla A100 GPUs with a total batch size of 256. In the second stage, GenAD is trained to jointly denoise future latents conditioned on past latents and texts. Its inputs are (video clip, text) pairs where each video clip is 4s at $2\mathrm{Hz}$ . The current version of GenAD is trained on 64 GPUs for $112.5\mathrm{K}$ iterations with a total batch size of 64. The input frames are resized to $256 \times 448$ for training in both stages, and the text condition c is dropped at a probability of $p = 0.1$ to enable classifier-free guidance [17] in sampling, which is commonly used in diffusion models to improve sample quality. More training and sampling details are in Appendix D. + +# 4.2. Results of Video Prediction Pre-training + +Comparison to Recent Video Generation Approaches. We compare GenAD to recent advances on an unseen set with geofencing from OpenDV-YouTube, Waymo [43], KITTI [14], and Cityscapes [10] in a zero-shot generation manner. Fig. 4 depicts the qualitative results. Image-to-video models I2VGen-XL [53] and VideoCrafter1 [8] can not strictly follow the given frames to make predictions, yielding poor consistency between the predicted frames and past frames. The video prediction model DMVFN [21] that is trained on Cityscapes suffers from the unfavorable shape distortions in its predictions, especially on the three unseen datasets. In contrast, GenAD exhibits remarkable zero-shot generalization ability and visual quality although none of these sets are included in the training. + +Comparison to nuScenes Experts. We also compare GenAD with the most recent available driving video generation models which are exclusively trained for nuScenes. Tab. 2 shows that GenAD surpasses all previous methods in both image fidelity (FID) and video coherence (FVD). Specifically, GenAD significantly reduces FVD by $44.5\%$ compared to DrivingDiffusion [31], without taking 3D future layouts as additional inputs. For fair comparisons, we train a model variant (GenAD-nus) on nuScenes dataset only. We find that although GenAD-nus performs on par + +![](images/948e930580ca8d90ade38de77861e2b964f2a400dc91a3f38cbba7fac5625c41.jpg) +Figure 4. Task on zero-shot video prediction for unseen scenarios. We show the generation results (in blue boxes) of different models given the same starting frames. GenAD makes more robust, realistic, and reasonable future predictions on unseen datasets (scenarios). More comparisons and visualizations are shown in Appendix. + +![](images/f85e60b3a0cbcc1383e23dd21035f4caa820a272079bcec0e55dbd74c5155760.jpg) + +![](images/88f0cfdb8928df107f6dcc46f75535ce3398111076335205dd499c99fe188ca0.jpg) +Figure 5. Task on language-conditioned prediction. Given two frames of a rainy scenario in the intersection and three high-level text conditions, GenAD simulates reasonable futures accordingly. + +with GenAD on nuScenes, it struggles to generalize to unseen datasets like Waymo, where the generation degrades to the nuScenes visual pattern. In contrast, GenAD trained on OpenDV-2K exhibits strong generalization ability across datasets as shown in Fig. 4. + +We provide language-conditioned prediction samples on nuScenes in Fig. 5, where GenAD simulates various futures + +from the same start following different textual instructions. The impressive generation quality is exhibited in the intricate details of the environment, and the natural transition of ego motion. + +Ablation Study. We perform ablations by training each variant on a subset of OpenDV-2K for 75K steps. Starting from the baseline with plain temporal attentions [4, 18], we gradually introduce our proposed components. Notably, by interleaving the temporal blocks with the spatial blocks, the FVD significantly improves $(-17\%)$ due to more sufficient spatiotemporal interactions. Both temporal causality and decoupled spatial attention contribute to better CLIP-SIM, improving the temporal consistency between future predictions and the condition frames. To be clear, the slight increase in FID and FVD, shown in fourth and third rows of Tab. 3 respectively, does not faithfully reflect a decline in generation quality as discussed in [4, 5, 37]. The effectiveness of each design is shown in Fig. 6. + +# 4.3. Results of Extensions + +Action-conditioned Prediction. We further showcase the performance of the action-conditioned model fine-tuned on nuScenes, GenAD-act, in Fig. 7 and Tab. 4. Given two starting frames and a trajectory w composed of 6 future way + +![](images/cac3305aa3f6014f26b6ac0c03e4935457a9eee25b4ededa50bcdad6299f2dbc.jpg) +Figure 6. Case study for model designs. All components help alleviate artifacts and improve the consistency of future predictions. + +
MethodYouTube
FID (↓)FVD (↓)CLIPSIM (↑)
Baseline18.32244.440.8405
+ Deep Interaction17.96201.690.8409
+ Temporal Causality16.54207.450.8550
+ Decoupled Spatial Attn.17.67189.540.8652
+ +Table 3. Ablation on model designs in GenAD. All proposed designs contribute to the final performance. + +
MethodConditionnuScenes +Action Prediction Error (↓)
Ground truth-0.90
GenADtext2.54
GenAD-acttext + traj.2.02
+ +points, GenAD-act imagines 6 future frames following the trajectory sequence. To evaluate the consistency between the input trajectory $\mathbf{w}$ and predicted frames, we establish an inverse dynamics model (IDM) on nuScenes as the evaluator, which projects a video sequence into a corresponding ego trajectory. We leverage the IDM to translate predicted frames into the trajectory $\hat{\mathbf{w}}$ , and calculate the L2 distance between $\mathbf{w}$ and $\hat{\mathbf{w}}$ as the Action Prediction Error. Specifically, GenAD-act substantially reduces the Action Prediction Error by $20.4\%$ compared to GenAD with text condition, allowing for more accurate future simulations. + +Planning Results. Tab. 5 depicts the planning results on nuScenes where ground truth poses for the ego vehicle are available. By freezing GenAD encoder and only optimizing + +![](images/cd7f531bb6325f055a4422f1202577063f1d2274d692faaa9b962d82dc87b944.jpg) +Figure 7. Task on action-conditioned prediction (simulation). Given the same starting frames and different future trajectories (shown in yellow dots in the first column), GenAD-act can simulate diverse futures following different ego intentions. More visualizations are in Appendix. + +Table 4. Task on action-conditioned prediction. Compared to GenAD with text conditions only, GenAD-act enables more precise future predictions that follow the action condition. + +
Method# Trainable Params.nuScenes
ADE (↓)FDE (↓)
ST-P3* [20]10.9M2.653.73
UniAD* [22]58.8M1.031.65
GenAD (Ours)0.8M1.232.31
+ +Table 5. Task on open-loop planning. A lightweight MLP with frozen GenAD gets competitive planning results with $73 \times$ fewer trainable parameters and front-view image alone. *: multi-view inputs. Evaluation protocols are aligned with UniAD [22]. + +an additional MLP on top of it, the model can effectively learn to plan. Notably, by pre-extracting image features through the UNet encoder of GenAD, the entire learning process for planning adaptation takes only 10 minutes on a single NVIDIA Tesla V100 device, which is 3400 times more efficient than the training of the UniAD planner [22]. + +# 5. Limitations and Discussion + +We study the system-level development of GenAD, a large-scale generalized video predictive model for autonomous driving. We also validate the adaptation of the learned representation of GenAD to driving tasks, i.e., learning a "world model" and motion planning. Although we obtain improved generalization to open domains, the increased model capacity poses challenges in both training efficiency and real-time deployment. We envision the unified video prediction task will serve as a scalable objective for future research on representation learning and policy learning. Another interesting direction involves distilling the encoded knowledge for a wider range of downstream tasks [29]. + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andy Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022. 2, 5 +[2] Mohammadhossein Bahari, Saeed Saadatnejad, Ahmad Rahimi, Mohammad Shaverdikondori, Amir Hossein Shahidzadeh, Seyed-Mohsen Moosavi-Dezfooli, and Alexandre Alahi. Vehicle trajectory prediction works, but not everywhere. In CVPR, 2022. 2 +[3] Fan Bao, Chongxuan Li, Jiacheng Sun, and Jun Zhu. Why are conditional generative models better than unconditional ones? arXiv preprint arXiv:2212.00362, 2022. 4 +[4] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 2, 5, 6, 7 +[5] Tim Brooks, Janne Hellsten, Miika Aittala, Ting-Chun Wang, Timo Aila, Jaakko Lehtinen, Ming-Yu Liu, Alexei Efros, and Tero Karras. Generating long videos of dynamic scenes. In NeurIPS, 2022. 7 +[6] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yuxin Pan, Giancarlo Baldan, and Oscar Beijbom. nuScenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 3, 4, 6 +[7] Holger Caesar, Juraj Kabzan, Kok Seang Tan, Whye Kit Fong, Eric M. Wolff, Alex Lang, Luke Fletcher, Oscar Bei-jbom, and Sammy Omari. nuPlan: A closed-loop ml-based planning benchmark for autonomous vehicles. In CVPR Workshops, 2021. 3, 4 +[8] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, Chao Weng, and Ying Shan. VideoCrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 6 +[9] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers. arXiv preprint arXiv:2306.16927, 2023. 2 +[10] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The Cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 3, 6 +[11] Yilun Dai, Mengjiao Yang, Bo Dai, Hanjun Dai, Ofir Nachum, Josh Tenenbaum, Dale Schuurmans, and Pieter Abbeel. Learning universal policies via text-guided video generation. In NeurIPS, 2023. 2 + +[12] Thierry Deruyttere, Simon Vandenhende, Dusan Grujicic, Luc Van Gool, and Marie-Francine Moens. Talk2Car: Taking control of your self-driving car. In EMNLP, 2019. 3, 4 +[13] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 4 +[14] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The KITTI dataset. *IJRR*, 2013. 3, 6 +[15] Agrim Gupta, Stephen Tian, Yunzhi Zhang, Jiajun Wu, Roberto Martin-Martin, and Li Fei-Fei. MaskViT: Masked visual pre-training for video prediction. In ICLR, 2023. 2 +[16] Niklas Hanselmann, Katrin Renz, Kashyap Chitta, Apratim Bhattacharyya, and Andreas Geiger. KING: Generating safety-critical driving scenarios for robust imitation via kinematics gradients. In ECCV, 2022. 2 +[17] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS Workshops, 2021. 6 +[18] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. In NeurIPS, 2022. 2, 5, 7 +[19] Anthony Hu, Lloyd Russell, Hudson Yeo, Zak Murez, George Fedoseev, Alex Kendall, Jamie Shotton, and Gianluca Corrado. GAIA-1: A generative world model for autonomous driving. arXiv preprint arXiv:2309.17080, 2023. 2 +[20] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. ST-P3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In ECCV, 2022. 8 +[21] Xiaotao Hu, Zhewei Huang, Ailin Huang, Jun Xu, and Shuchang Zhou. A dynamic multi-scale voxel flow network for video prediction. In CVPR, 2023. 6 +[22] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, Lewei Lu, Xiaosong Jia, Qiang Liu, Jifeng Dai, Yu Qiao, and Hongyang Li. Planning-oriented autonomous driving. In CVPR, 2023. 6, 8 +[23] Fan Jia, Weixin Mao, Yingfei Liu, Yucheng Zhao, Yuqing Wen, Chi Zhang, Xiangyu Zhang, and Tiancai Wang. A driver-i: A general world model for autonomous driving. arXiv preprint arXiv:2311.13549, 2023. 2 +[24] Jinkyu Kim, Teruhisa Miso, Yi-Ting Chen, Ashish Tawari, and John Canny. Grounding human-to-vehicle advice for self-driving vehicles. In CVPR, 2019. 3, 4 +[25] Seung Wook Kim, Jonah Philion, Antonio Torralba, and Sanja Fidler. DriveGAN: Towards a controllable high-quality neural simulation. In CVPR, 2021. 2, 6 +[26] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In ICCV, 2023. 2 +[27] Yann LeCun. A path towards autonomous machine intelligence version 0.9. 2, 2022-06-27. Open Review, 62, 2022. 2 +[28] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. Nature, 2015. 2 + +[29] Daiqing Li, Huan Ling, Amlan Kar, David Acuna, Seung Wook Kim, Karsten Kreis, Antonio Torralba, and Sanja Fidler. DreamTeacher: Pretraining image backbones with deep generative models. In ICCV, 2023. 8 +[30] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 2, 3, 4 +[31] Xiaofan Li, Yifu Zhang, and Xiaoqing Ye. DrivingDiffusion: Layout-guided multi-view driving scene video generation with latent diffusion model. arXiv preprint arXiv:2310.07771, 2023. 6 +[32] Zhenyu Li, Zehui Chen, Ang Li, Liangji Fang, Qinhong Jiang, Xianming Liu, and Junjun Jiang. Unsupervised domain adaptation for monocular 3d object detection via self-training. In ECCV, 2022. 3 +[33] Jiachen Lu, Ze Huang, Jiahui Zhang, Zeyu Yang, and Li Zhang. Wovogen: World volume-aware diffusion for controllable multi-camera driving scene generation. arXiv preprint arXiv:2312.02934, 2023. 2 +[34] Jiageng Mao, Minzhe Niu, Chenhan Jiang, Hanxue Liang, Xiaodan Liang, Yamin Li, Chao Ye, Wei Zhang, Zhenguo Li, Jie Yu, Hang Xu, and Chunjing Xu. One million scenes for autonomous driving: ONCE dataset. In NeurIPS Datasets and Benchmarks, 2021. 3, 4 +[35] OpenAI. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2 +[36] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul F Christiano, Jan Leike, and Ryan Lowe. Training language models to follow instructions with human feedback. In NeurIPS, 2022. 3, 4 +[37] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 4, 7 +[38] Xiaojuan Qi, Zhengzhe Liu, Qifeng Chen, and Jiaya Jia. 3D motion decomposition for rgbd future dynamic scene synthesis. In CVPR, 2019. 2 +[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 4 +[40] Vasili Ramanishka, Yi-Ting Chen, Teruhisa Mitsu, and Kate Saenko. Toward driving scene understanding: A dataset for learning driver behavior and causal reasoning. In CVPR, 2018. 3, 4 +[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 4 +[42] Peter Shaw, Jakob Uszkoreit, and Ashish Vaswani. Self-attention with relative position representations. arXiv preprint arXiv:1803.02155, 2018. 5 + +[43] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott M. Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, 2020. 3, 6 +[44] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. In NeurIPS, 2020. 6 +[45] Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, and Jiwen Lu. DriveDreamer: Towards real-world-driven world models for autonomous driving. arXiv preprint arXiv:2309.09777, 2023. 2, 6 +[46] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, Yuwei Guo, Tianxing Wu, Chenyang Si, Yuming Jiang, Cunjian Chen, Chen Change Loy, Bo Dai, Dahua Lin, Yu Qiao, and Ziwei Liu. LAVIE: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 5 +[47] Yuqi Wang, Jiawei He, Lue Fan, Hongxin Li, Yuntao Chen, and Zhaoxiang Zhang. Driving into the future: Multiview visual forecasting and planning with world model for autonomous driving. arXiv preprint arXiv:2311.17918, 2023. 2 +[48] Benjamin Wilson, William Qi, Tanmay Agarwal, John Lambert, Jagjeet Singh, Siddhesh Khandelwal, Bowen Pan, Ratnesh Kumar, Andrew Hartnett, Jhony Kaesemodel Pontes, Deva Ramanan, Peter Carr, and James Hays. Argoverse 2: Next generation datasets for self-driving perception and forecasting. In NeurIPS Datasets and Benchmarks, 2021. 3 +[49] Daniel M Wolpert, Zoubin Ghahramani, and Michael I Jordan. An internal model for sensorimotor integration. Science, 1995. 2 +[50] Zhiheng Xi, Wenxiang Chen, Xin Guo, Wei He, Yiwen Ding, Boyang Hong, Ming Zhang, Junzhe Wang, Senjie Jin, Enyu Zhou, Rui Zheng, Xiaoran Fan, Xiao Wang, Liao Xiong, Yuhao Zhou, Weiran Wang, Changhao Jiang, Yicheng Zou, Xiangyang Liu, Zhangyue Yin, Shihan Dou, Rongxiang Weng, Wensen Cheng, Qi Zhang, Wenjuan Qin, Yongyan Zheng, Xipeng Qiu, Xuanjing Huang, and Tao Gui. The rise and potential of large language model based agents: A survey. arXiv preprint arXiv:2309.07864, 2023. 2 +[51] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. arXiv preprint arXiv:2309.15818, 2023. 5 +[52] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 5 +[53] Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qin, Xiang Wang, Deli Zhao, and + +Jingren Zhou. I2VGen-XL: High-quality image-to-video synthesis via cascaded diffusion models. arXiv preprint arXiv:2311.04145, 2023. 6 +[54] Ruizhao Zhu, Peng Huang, Eshed Ohn-Bar, and Venkatesh Saligrama. Learning to drive anywhere. In CoRL, 2023. 2 \ No newline at end of file diff --git a/2024/Generalized Predictive Model for Autonomous Driving/images.zip b/2024/Generalized Predictive Model for Autonomous Driving/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e4c64957b44b5429531ba13a8b021cbc8001d8fb --- /dev/null +++ b/2024/Generalized Predictive Model for Autonomous Driving/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9bcc712aee52d41ed7aecd9fa85b0a6078e07913cfb035711e18763e93528e57 +size 769597 diff --git a/2024/Generalized Predictive Model for Autonomous Driving/layout.json b/2024/Generalized Predictive Model for Autonomous Driving/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c2faae5fd1f51239db49bfa31ac77ef2258aa269 --- /dev/null +++ b/2024/Generalized Predictive Model for Autonomous Driving/layout.json @@ -0,0 +1,7660 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 127, + 103, + 467, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 103, + 467, + 121 + ], + "spans": [ + { + "bbox": [ + 127, + 103, + 467, + 121 + ], + "type": "text", + "content": "Generalized Predictive Model for Autonomous Driving" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 141, + 515, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 141, + 515, + 186 + ], + "spans": [ + { + "bbox": [ + 77, + 141, + 515, + 186 + ], + "type": "text", + "content": "Jiazhi Yang" + }, + { + "bbox": [ + 77, + 141, + 515, + 186 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 77, + 141, + 515, + 186 + ], + "type": "text", + "content": " Shenyuan Gao" + }, + { + "bbox": [ + 77, + 141, + 515, + 186 + ], + "type": "inline_equation", + "content": "^{2,1*}" + }, + { + "bbox": [ + 77, + 141, + 515, + 186 + ], + "type": "text", + "content": " Yihang Qiu" + }, + { + "bbox": [ + 77, + 141, + 515, + 186 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 77, + 141, + 515, + 186 + ], + "type": "text", + "content": " Li Chen" + }, + { + "bbox": [ + 77, + 141, + 515, + 186 + ], + "type": "inline_equation", + "content": "^{3,1\\dagger}" + }, + { + "bbox": [ + 77, + 141, + 515, + 186 + ], + "type": "text", + "content": " Tianyu Li" + }, + { + "bbox": [ + 77, + 141, + 515, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 141, + 515, + 186 + ], + "type": "text", + "content": " Bo Dai" + }, + { + "bbox": [ + 77, + 141, + 515, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 141, + 515, + 186 + ], + "type": "text", + "content": " Kashyap Chitta" + }, + { + "bbox": [ + 77, + 141, + 515, + 186 + ], + "type": "inline_equation", + "content": "^{4,5}" + }, + { + "bbox": [ + 77, + 141, + 515, + 186 + ], + "type": "text", + "content": " Penghao Wu" + }, + { + "bbox": [ + 77, + 141, + 515, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 141, + 515, + 186 + ], + "type": "text", + "content": " Jia Zeng" + }, + { + "bbox": [ + 77, + 141, + 515, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 141, + 515, + 186 + ], + "type": "text", + "content": " Ping Luo" + }, + { + "bbox": [ + 77, + 141, + 515, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 77, + 141, + 515, + 186 + ], + "type": "text", + "content": " Jun Zhang" + }, + { + "bbox": [ + 77, + 141, + 515, + 186 + ], + "type": "inline_equation", + "content": "^{2\\sharp}" + }, + { + "bbox": [ + 77, + 141, + 515, + 186 + ], + "type": "text", + "content": " Andreas Geiger" + }, + { + "bbox": [ + 77, + 141, + 515, + 186 + ], + "type": "inline_equation", + "content": "^{4,5\\sharp}" + }, + { + "bbox": [ + 77, + 141, + 515, + 186 + ], + "type": "text", + "content": " Yu Qiao" + }, + { + "bbox": [ + 77, + 141, + 515, + 186 + ], + "type": "inline_equation", + "content": "^{1\\sharp}" + }, + { + "bbox": [ + 77, + 141, + 515, + 186 + ], + "type": "text", + "content": " Hongyang Li" + }, + { + "bbox": [ + 77, + 141, + 515, + 186 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 71, + 190, + 523, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 190, + 523, + 220 + ], + "spans": [ + { + "bbox": [ + 71, + 190, + 523, + 220 + ], + "type": "text", + "content": "1 OpenDriveLab and Shanghai AI Lab 2 Hong Kong University of Science and Technology 3 University of Hong Kong 4 University of Tübingen 5 Tübingen AI Center" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 89, + 250, + 105, + 261 + ], + "blocks": [ + { + "bbox": [ + 89, + 250, + 105, + 261 + ], + "lines": [ + { + "bbox": [ + 89, + 250, + 105, + 261 + ], + "spans": [ + { + "bbox": [ + 89, + 250, + 105, + 261 + ], + "type": "image", + "image_path": "b80cd5b5f461d4fd4cdc6680a5a80f78390f39e532e73a9719c51b85443b4c30.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 107, + 251, + 154, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 251, + 154, + 261 + ], + "spans": [ + { + "bbox": [ + 107, + 251, + 154, + 261 + ], + "type": "text", + "content": "OpenDV-2K" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 48, + 262, + 202, + 375 + ], + "blocks": [ + { + "bbox": [ + 48, + 262, + 202, + 375 + ], + "lines": [ + { + "bbox": [ + 48, + 262, + 202, + 375 + ], + "spans": [ + { + "bbox": [ + 48, + 262, + 202, + 375 + ], + "type": "image", + "image_path": "cb2a404a9278330cd862b1bc1b938fabc376187acc9fda4362a08d287d0385e6.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 255, + 247, + 269, + 263 + ], + "blocks": [ + { + "bbox": [ + 255, + 247, + 269, + 263 + ], + "lines": [ + { + "bbox": [ + 255, + 247, + 269, + 263 + ], + "spans": [ + { + "bbox": [ + 255, + 247, + 269, + 263 + ], + "type": "image", + "image_path": "6d108f23ef58e3208f38828182bba2975dc9d3a31d659c14a88aab89bfc6b2be.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 272, + 251, + 301, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 251, + 301, + 260 + ], + "spans": [ + { + "bbox": [ + 272, + 251, + 301, + 260 + ], + "type": "text", + "content": "GenAD" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 209, + 262, + 348, + 375 + ], + "blocks": [ + { + "bbox": [ + 209, + 262, + 348, + 375 + ], + "lines": [ + { + "bbox": [ + 209, + 262, + 348, + 375 + ], + "spans": [ + { + "bbox": [ + 209, + 262, + 348, + 375 + ], + "type": "image", + "image_path": "e19f39be9e834ce2ba2f960545759988df3469bfad8ed6cbdae86f58e2f9fa22.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 422, + 247, + 439, + 264 + ], + "blocks": [ + { + "bbox": [ + 422, + 247, + 439, + 264 + ], + "lines": [ + { + "bbox": [ + 422, + 247, + 439, + 264 + ], + "spans": [ + { + "bbox": [ + 422, + 247, + 439, + 264 + ], + "type": "image", + "image_path": "0cd856b96f349dfbbd8ff9ae0fc5fa0e9a187141fda5adc06125fa773e8432f4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 443, + 251, + 463, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 443, + 251, + 463, + 260 + ], + "spans": [ + { + "bbox": [ + 443, + 251, + 463, + 260 + ], + "type": "text", + "content": "Tasks" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 354, + 267, + 545, + 371 + ], + "blocks": [ + { + "bbox": [ + 354, + 267, + 545, + 371 + ], + "lines": [ + { + "bbox": [ + 354, + 267, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 354, + 267, + 545, + 371 + ], + "type": "image", + "image_path": "48201f52fac94aa848c29342b5689af78076b4daa7a33a69e107d21e971a9a26.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 50, + 390, + 227, + 470 + ], + "blocks": [ + { + "bbox": [ + 79, + 380, + 198, + 389 + ], + "lines": [ + { + "bbox": [ + 79, + 380, + 198, + 389 + ], + "spans": [ + { + "bbox": [ + 79, + 380, + 198, + 389 + ], + "type": "text", + "content": "2. Language-conditioned Prediction" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 390, + 227, + 470 + ], + "lines": [ + { + "bbox": [ + 50, + 390, + 227, + 470 + ], + "spans": [ + { + "bbox": [ + 50, + 390, + 227, + 470 + ], + "type": "image", + "image_path": "57270d275699fef0f89f2a74725ff5f9947d6f9c8817e08dadcc34b4f19bb07a.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 235, + 390, + 413, + 468 + ], + "blocks": [ + { + "bbox": [ + 269, + 380, + 375, + 389 + ], + "lines": [ + { + "bbox": [ + 269, + 380, + 375, + 389 + ], + "spans": [ + { + "bbox": [ + 269, + 380, + 375, + 389 + ], + "type": "text", + "content": "3. Action-conditioned Prediction" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 235, + 390, + 413, + 468 + ], + "lines": [ + { + "bbox": [ + 235, + 390, + 413, + 468 + ], + "spans": [ + { + "bbox": [ + 235, + 390, + 413, + 468 + ], + "type": "image", + "image_path": "51c6bdb6e97a0982e4bf55874a52e1dc0df0e46ab56367adff2a33661bd99bdf.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 417, + 390, + 545, + 470 + ], + "blocks": [ + { + "bbox": [ + 463, + 380, + 503, + 389 + ], + "lines": [ + { + "bbox": [ + 463, + 380, + 503, + 389 + ], + "spans": [ + { + "bbox": [ + 463, + 380, + 503, + 389 + ], + "type": "text", + "content": "4. Planning" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 417, + 390, + 545, + 470 + ], + "lines": [ + { + "bbox": [ + 417, + 390, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 417, + 390, + 545, + 470 + ], + "type": "image", + "image_path": "57aaf11851c35fa0fdae05e906b224d79236dd3965619e6e637c61b7175cad81.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 482, + 547, + 528 + ], + "lines": [ + { + "bbox": [ + 45, + 482, + 547, + 528 + ], + "spans": [ + { + "bbox": [ + 45, + 482, + 547, + 528 + ], + "type": "text", + "content": "Figure 1. Overview of the GenAD paradigm. We aim to establish a generalized video prediction paradigm for autonomous driving by presenting the largest multimodal driving video dataset to date, OpenDV-2K, and a generative model that predicts the future given past visual and textual input, GenAD. The strong generalization and controllability of GenAD is validated spanning a diverse spectrum of tasks, including zero-shot domain transfer, language-conditioned prediction, action-conditioned prediction, and motion planning." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 143, + 548, + 192, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 548, + 192, + 562 + ], + "spans": [ + { + "bbox": [ + 143, + 548, + 192, + 562 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 574, + 288, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 574, + 288, + 683 + ], + "spans": [ + { + "bbox": [ + 46, + 574, + 288, + 683 + ], + "type": "text", + "content": "In this paper, we introduce the first large-scale video prediction model in the autonomous driving discipline. To eliminate the restriction of high-cost data collection and empower the generalization ability of our model, we acquire massive data from the web and pair it with diverse and high-quality text descriptions. The resultant dataset accumulates over 2000 hours of driving videos, spanning areas all over the world with diverse weather conditions and traffic scenarios. Inheriting the merits from recent la" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 538, + 545, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 538, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 306, + 538, + 545, + 634 + ], + "type": "text", + "content": "tent diffusion models, our model, dubbed GenAD, handles the challenging dynamics in driving scenes with novel temporal reasoning blocks. We showcase that it can generalize to various unseen driving datasets in a zero-shot manner, surpassing general or driving-specific video prediction counterparts. Furthermore, GenAD can be adapted into an action-conditioned prediction model or a motion planner, holding great potential for real-world driving applications." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 669, + 386, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 669, + 386, + 681 + ], + "spans": [ + { + "bbox": [ + 306, + 669, + 386, + 681 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "content": "Autonomous driving agents, as a promising application of high-level artificial intelligence, perceive the surrounding" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 693, + 282, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 693, + 282, + 703 + ], + "spans": [ + { + "bbox": [ + 61, + 693, + 282, + 703 + ], + "type": "text", + "content": "*Equal contribution, ordered by coin toss. Equal co-advising." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 61, + 703, + 282, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 703, + 282, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 703, + 282, + 713 + ], + "type": "text", + "content": "†Project lead. Primary contact: yangjiazhi@opendrivelab.com" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14662" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "environment, build internal world model representations, make decisions, and take actions in response [9, 50]. However, despite dedicated efforts in academia and industry for decades, their deployment is still restricted to certain areas or scenarios, and they cannot be applied over the world seamlessly. One critical reason is the limited generalization ability of learned models in structured autonomous driving systems. Typically, perception models face challenges of generalizing to diverse environments with changes in geographical locations, sensor configurations, weather conditions, open-set objects, etc.; prediction and planning models fail to generalize to nondeterministic futures with rare scenarios and different driving intentions [2, 16, 54]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 228, + 289, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 289, + 337 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 289, + 337 + ], + "type": "text", + "content": "Motivated by how humans learn to perceive and cognize the world [27, 28, 49], we advocate employing driving videos as the universal interface that generalizes to diverse environments with dynamic futures. Based on this, a driving video predictive model is preferred to fully capture the world knowledge about driving scenarios (Fig. 1). By predicting the future, the video predictor essentially learns two vital aspects of autonomous driving: how the world operates, and how to maneuver safely in the wild." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 337, + 289, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 337, + 289, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 337, + 289, + 601 + ], + "type": "text", + "content": "Recently, the community has begun to adopt video as the interface to represent observation behavior and action for various robot tasks [11]. For domains such as classical video prediction and robotics, the video backgrounds are mostly static, the movement of robots is slow, and the resolution of videos is low. In contrast, for the driving scenarios, it struggles with outdoor environments being highly dynamic, agents encompassing much larger motions, and the sensory resolution covering a large range of view. These distinctions lead to substantial challenges for autonomous driving applications. Fortunately, there are some preliminary attempts on developing a video predictive model in the driving domain [4, 15, 19, 23, 25, 33, 38, 45, 47]. Despite promising progress in terms of prediction quality, these attempts have not achieved desirable capability of generalization as in classical robot tasks (e.g., manipulation), being confined to either limited scenarios such as highways with low traffic density [4] and small-scale datasets [15, 23, 33, 45, 47], or restricted conditions that raises difficulties to generate diverse environments [38]. How to unveil the potential of video prediction models for driving remains seldom explored." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 601, + 289, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 601, + 289, + 698 + ], + "spans": [ + { + "bbox": [ + 46, + 601, + 289, + 698 + ], + "type": "text", + "content": "Motivated by the discussions above, we target at building a video predictive model for autonomous driving, capable of generalizing to new conditions and environments. To this end, we have to answer the following questions: (1) What data can be obtained in a feasible and scalable manner? (2) How can we formulate a predictive model to capture the complex evolution of dynamic scenarios? (3) How can we apply the (foundation) model for downstream tasks?" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 701, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 288, + 714 + ], + "type": "text", + "content": "Scaled Data. To achieve powerful generalization ability, a" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 335 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 335 + ], + "type": "text", + "content": "substantial and diverse corpus of data is necessary. Inspired by the success of learning from Internet-scale data in foundation models [1, 26, 39], we construct our driving dataset from both the web and publicly licensed datasets. Compared to existing options, which are limited in scale and diversity due to their regulated collection processes, online data owns great diversity in several aspects: geographic locations, terrains, weather conditions, safety-critical scenarios, sensor settings, traffic elements, etc. To guarantee the data is of high-quality and desirable for large-scale training, we exhaustively collect driving recordings on YouTube and remove unintended corruption frames via rigorous human verification. Furthermore, videos are paired with diverse text-level conditions, including descriptions generated and refined with the aid of existing foundation models [30, 35], and high-level instructions inferred by a video classifier. Through these steps, we construct OpenDV-2K, the largest public driving dataset to date, containing more than 2000 hours of driving videos and being 374 times larger than the widely used nuScenes counterpart. Our dataset is publicly available at https://github.com/OpenDriveLab/DriveAGI." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 337, + 547, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 337, + 547, + 589 + ], + "spans": [ + { + "bbox": [ + 304, + 337, + 547, + 589 + ], + "type": "text", + "content": "Generalized Predictive Model. Learning a generalized driving video predictor bears several key challenges: generation quality, training efficiency, causal reasoning, and drastic view shift. We address these aspects by presenting a novel temporal generative model with two-stage learning. To capture the environment details, enhance generation quality, and maintain training efficiency simultaneously, we build upon the recent success of latent diffusion models (LDMs) [37, 41]. In the first stage, we transfer the generation distribution of LDM from its pre-trained general vision domain to the driving domain by fine-tuning it on OpenDV2K images. In the second stage, we interleave the proposed temporal reasoning blocks into the original model and learn to predict the future given past frames and conditions. Contrary to conventional temporal modules [4, 18] that suffer from causal confusion and large motion, our solution consists of causal temporal attention and decoupled spatial attention to efficiently model the drastic spatiotemporal shift in highly dynamic driving scenes. After sufficient training, our Generative model for Autonomous Driving (GenAD)1 can generalize to various scenarios in a zero-shot fashion." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 591, + 547, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 591, + 547, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 591, + 547, + 688 + ], + "type": "text", + "content": "Extensions for Simulation and Planning. After large-scale pre-training of video prediction, GenAD essentially understands how the world evolves and how to drive. We show how to adapt its learned knowledge for real-world driving problems, i.e., simulation and planning. For simulation, we fine-tune the pre-trained model with future ego trajectories as additional conditions, to associate future imaginations with different ego actions. We also empower" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "text", + "content": "1Note that GenAD is abbreviated from both Generative models and Generalized capabilities." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14663" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 71, + 387, + 237 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 387, + 237 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 387, + 237 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 387, + 237 + ], + "type": "table", + "html": "
DatasetDuration (hours)Front-view FramesGeographic Diversity CountriesCitiesSensor Setup
XKITTI [14]1.415k11fixed
XCityscapes [10]0.525k350fixed
XWaymo Open* [43]11390k13fixed
XArgoverse 2* [48]4.2300k16fixed
✓nuScenes [6]5.5241k22fixed
✓nuPlan* [7]1204.0M24fixed
✓Talk2Car [12]4.7-22fixed
✓ONCE [34]1447M1-fixed
✓Honda-HAD [24]321.2M1-fixed
✓Honda-HDD-Action [40]1041.1M1-fixed
✓Honda-HDD-Cause [40]32-1-fixed
✓OpenDV-YouTube (Ours)174760.2M≥40†≥244†uncalibrated
-OpenDV-2K (Ours)205965.1M≥40†≥244†uncalibrated
", + "image_path": "86c336cd91636c4b4b48a616a3121fb6c246e5755f166ef8f8c8f211c939b43f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 240, + 383, + 275 + ], + "lines": [ + { + "bbox": [ + 46, + 240, + 383, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 383, + 275 + ], + "type": "text", + "content": "Table 1. OpenDV-2K comparison at a glance to existing counterparts in terms of scale and diversity. Note that datasets with " + }, + { + "bbox": [ + 46, + 240, + 383, + 275 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 46, + 240, + 383, + 275 + ], + "type": "text", + "content": " are included in OpenDV-2K (last row). \\*Perception subset in Waymo Open, Argoverse 2, and nuPlan. Estimated by GPT [36] from video titles." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 389, + 70, + 542, + 237 + ], + "blocks": [ + { + "bbox": [ + 389, + 70, + 542, + 237 + ], + "lines": [ + { + "bbox": [ + 389, + 70, + 542, + 237 + ], + "spans": [ + { + "bbox": [ + 389, + 70, + 542, + 237 + ], + "type": "image", + "image_path": "3415a03895c87d97d29527857e876efbd9de7ba7bae8be4262d02422b3c74959.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 388, + 240, + 547, + 275 + ], + "lines": [ + { + "bbox": [ + 388, + 240, + 547, + 275 + ], + "spans": [ + { + "bbox": [ + 388, + 240, + 547, + 275 + ], + "type": "text", + "content": "Figure 2. Geographic distribution of OpenDV-2K. Our dataset covers ample driving scenarios around the world." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 295, + 289, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 295, + 289, + 368 + ], + "spans": [ + { + "bbox": [ + 46, + 295, + 289, + 368 + ], + "type": "text", + "content": "GenAD to perform planning on challenging benchmarks by using a lightweight planner to translate latent features into the future trajectory of the ego vehicle. On account of its pre-trained ability to predict accurate future frames, our algorithm exhibits promising results in both simulation consistency and planning reliability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 377, + 168, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 377, + 168, + 392 + ], + "spans": [ + { + "bbox": [ + 47, + 377, + 168, + 392 + ], + "type": "text", + "content": "2. OpenDV-2K Dataset" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 398, + 288, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 398, + 288, + 578 + ], + "spans": [ + { + "bbox": [ + 46, + 398, + 288, + 578 + ], + "type": "text", + "content": "We introduce OpenDV-2K, a large-scale multimodal dataset for autonomous driving, to support the training of a generalized video prediction model. The main component is a vast corpus of high-quality YouTube driving videos, which are collected from all over the world, and are gathered into our dataset after a careful curation process. We automatically create language annotations for these videos using vision-language models. To further improve its diversity in sensor configurations and language expressions, we merge 7 publicly licensed datasets into our OpenDV-2K, as shown in Tab. 1. As a result, OpenDV-2K occupies a total of 2059 hours of videos paired with texts, including 1747 hours from YouTube and 312 hours from public datasets. We use OpenDV-YouTube and OpenDV-2K to specify the YouTube split and the overall dataset, respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 584, + 206, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 584, + 206, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 584, + 206, + 597 + ], + "type": "text", + "content": "2.1. Diversity over Prior Datasets" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 602, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 602, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 602, + 287, + 639 + ], + "type": "text", + "content": "A brief comparison with other public datasets is provided in Tab. 1. Beyond its significant scale, the proposed OpenDV-2K represents diversity across various aspects as follows." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": "Globe-wise Geographic Distribution. Due to the global nature of online videos, OpenDV-2K covers more than 40 countries and 244 cities worldwide. This is a tremendous improvement over previous public datasets, which are typically gathered in a small number of restricted areas. We plot the specific distribution of OpenDV-YouTube in Fig. 2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 295, + 547, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 295, + 547, + 380 + ], + "spans": [ + { + "bbox": [ + 304, + 295, + 547, + 380 + ], + "type": "text", + "content": "Open-world Driving Scenarios. Our dataset provides a huge amount of realistic driving experience in the open world, covering rare environments like forests, extreme weather conditions like heavy snow, and appropriate driving behaviors in response to interactive traffic situations. These data are crucial for diversity and generalization yet are seldom collected in existing public datasets." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 384, + 547, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 547, + 492 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 547, + 492 + ], + "type": "text", + "content": "Unrestricted Sensor Configurations. Current driving datasets are confined to specific sensor configurations, including intrinsic and extrinsic camera parameters, image, sensor type, optics, etc., which poses great challenges for deploying the learned models with different sensors [32]. In contrast, YouTube driving videos are recorded in various types of vehicles with flexible camera setups, which aids in the robustness of the trained model when deployed using a novel camera setting." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 502, + 528, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 502, + 528, + 516 + ], + "spans": [ + { + "bbox": [ + 305, + 502, + 528, + 516 + ], + "type": "text", + "content": "2.2. Towards High-quality Multimodal Dataset" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "text", + "content": "Driving Video Collection and Curation. Finding clean driving videos from the vast pool of the web is a tedious and costly task. To simplify the process, we start by selecting certain video uploaders, i.e., YouTubers. Judging from the average length and overall quality, we collect 43 YouTubers with 2139 high-quality front-view driving videos. To make sure there is no overlap between training and validation sets, we take all videos from 3 YouTubers for validation, with the remaining videos as the training set. To rule out non-driving frames like video introductions and subscription reminders, we discard a certain length of segments at the beginning and end of each video. Each frame is then described with language contexts using a VLM model, BLIP-2 [30]. We further remove the black frames and transition frames, which are not ideal for training, by manually checking if there are certain keywords in these contexts. We give an illustration" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "14664" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "of the dataset construction pipeline in Appendix C.1.1, and we introduce how to generate the contexts below." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 99, + 288, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 99, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 99, + 288, + 316 + ], + "type": "text", + "content": "Language Annotation for YouTube Videos. To create a predictive model that can be controlled by natural language to simulate different futures accordingly, To make the predictive model controllable and improve the sample quality [3], it is crucial to pair the driving videos with meaningful and varied language annotations. We construct two types of texts for OpenDV-YouTube, i.e., driving commands for ego-vehicle and frame descriptions, namely \"command\" and \"context\", to help the model comprehend ego actions and open-world concepts, respectively. For commands, we train a video classifier on Honda-HDD-Action [40] for 14 types of actions to label ego behaviors in a 4s sequence. These categorical commands will be further mapped to multiple free-form expressions from a predefined dictionary. For contexts, we leverage an established vision-language model, BLIP-2 [30], to describe the main objects and scenarios for each frame. For more details on annotations, please refer to Appendix C.1.2." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 319, + 287, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 319, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 287, + 475 + ], + "type": "text", + "content": "Enlarging Language Spectrum with Public Datasets. Considering that BLIP-2 annotations are generated for static frames without comprehension of dynamic driving scenarios such as the traffic light transitions, we exploit several public datasets that provide linguistic descriptions for driving scenarios [6, 7, 12, 24, 34, 40]. However, their metadata is relatively sparse with only a few words such as \"sunny road\". We further enhance their text quality using GPT [36] to form a descriptive \"context\" and generate a \"command\" by categorizing the logged trajectory for each video clip. Ultimately, we integrate these datasets with OpenDV-YouTube to establish OpenDV-2K dataset, as shown in the last row of Tab. 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 486, + 165, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 486, + 165, + 498 + ], + "spans": [ + { + "bbox": [ + 47, + 486, + 165, + 498 + ], + "type": "text", + "content": "3. GenAD Framework" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 506, + 287, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 287, + 627 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 287, + 627 + ], + "type": "text", + "content": "In this section, we introduce the training and design of the GenAD model. As shown in Fig. 3, GenAD is trained in two stages, i.e., image domain transferring and video prediction pre-training. The first stage adapts the general text-to-image model to the driving domain (Sec. 3.1). The second stage lifts the text-to-image model to a video prediction model with our proposed temporal reasoning block and modified training schemes (Sec. 3.2). In Sec. 3.3, we explore how the predictive model can be extended to action-conditioned prediction and planning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 635, + 182, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 182, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 182, + 647 + ], + "type": "text", + "content": "3.1. Image Domain Transfer" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "On-board cameras capture a large field of views with abundant visual contents, including the road, background buildings, surrounding vehicles, etc., which require strong and robust generation capability to produce continuous and realistic driving scenarios. To facilitate the learning process, we" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": "start with independent image generation in the first stage. Concretely, we initialize our model with SDXL [37], which is a large-scale latent diffusion model (LDM) for text-to-image generation, to leverage its ability to synthesize high-quality images with plenty of visual details. It is implemented as a denoising UNet " + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\theta}" + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": " with several stacked convolution and attention blocks, which learns to synthesize images by denoising the noisy latents [41]. Specifically, given a noisy input latent " + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": " corrupted by the forward diffusion process, it is trained to predict the added noise " + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": " via the following objective:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 333, + 208, + 545, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 208, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 333, + 208, + 545, + 228 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {i m g}} := \\mathbb {E} _ {\\mathbf {x}, \\epsilon \\sim \\mathcal {N} (0, 1), \\mathbf {c}, t} \\left[ \\| \\epsilon - \\mathbf {f} _ {\\theta} (\\mathbf {x} _ {t}; \\mathbf {c}, t) \\| _ {2} ^ {2} \\right], \\tag {1}", + "image_path": "2b36c30433a526e4ab2cc24a530b08bfd05dafc3ee29b57664b613ccef80bba6.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 232, + 545, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 232, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 232, + 545, + 327 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 232, + 545, + 327 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 232, + 545, + 327 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 232, + 545, + 327 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 304, + 232, + 545, + 327 + ], + "type": "text", + "content": " are the clean and noisy latent, respectively, " + }, + { + "bbox": [ + 304, + 232, + 545, + 327 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 232, + 545, + 327 + ], + "type": "text", + "content": " denotes the timestep for different noise scales, and " + }, + { + "bbox": [ + 304, + 232, + 545, + 327 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 304, + 232, + 545, + 327 + ], + "type": "text", + "content": " is the text condition that guides the denoising process, which is a concatenation of context and command. For training efficiency, the learning process takes place in a compressed latent space [13, 37, 41] instead of pixel space. During sampling, the model generates images from standard Gaussian noise by denoising the last-step predictions iteratively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 328, + 545, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 328, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 304, + 328, + 545, + 436 + ], + "type": "text", + "content": "However, the original SDXL is trained on data in the general domain, such as portraits and artistic paintings, which are not concerned with autonomy systems. To adapt the model to synthesize images for driving, we fine-tune it on text-to-image generation using image-text pairs in OpenDV-2K with the same objective as Eq. (1). Following the original training of SDXL, all parameters " + }, + { + "bbox": [ + 304, + 328, + 545, + 436 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 328, + 545, + 436 + ], + "type": "text", + "content": " of the UNet are fine-tuned at this stage, whereas the CLIP text encoders [39] and the autoencoder [13] remain frozen." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 441, + 468, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 441, + 468, + 453 + ], + "spans": [ + { + "bbox": [ + 306, + 441, + 468, + 453 + ], + "type": "text", + "content": "3.2. Video Prediction Pre-training" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 460, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 544 + ], + "type": "text", + "content": "In the second stage, with a few frames of a consecutive video as past observations, GenAD is trained to reason about all visual observations and predict several future frames in plausible ways. Similar to the first stage stage, the prediction process can also be guided by text conditions. However, predicting the highly dynamic driving world temporally is challenging due to two fundamental barriers." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 544, + 545, + 676 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 306, + 544, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 544, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 306, + 544, + 545, + 616 + ], + "type": "text", + "content": "1. Causal Reasoning: To predict plausible futures following the temporal causality of the driving world, the model needs to comprehend the intentions of all other agents together with the ego vehicle, and understand underlying traffic rules, e.g., how the traffic will change with the transition of traffic lights." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 616, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 616, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 306, + 616, + 545, + 676 + ], + "type": "text", + "content": "2. Drastic View Shift: Contrary to typical video generation benchmarks which mainly have a static background with slow motion of centered objects, the view of driving changes drastically over time. Each pixel in every frame may move to a distant location in the next frame." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": "We propose temporal reasoning blocks to address these problems. As illustrated in Fig. 3(c), each block is composed of three successive attention layers, i.e., the causal" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14665" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 414, + 274 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 414, + 274 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 414, + 274 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 414, + 274 + ], + "type": "image", + "image_path": "529d3d44fcae5674be5d92a354b7f27841dd548e493f20c4d81919d42a78b120.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 276, + 547, + 344 + ], + "lines": [ + { + "bbox": [ + 46, + 276, + 547, + 344 + ], + "spans": [ + { + "bbox": [ + 46, + 276, + 547, + 344 + ], + "type": "text", + "content": "Figure 3. Framework of GenAD. (a) The two-stage learning for GenAD is composed of transferring the image domain of an image diffusion model to the driving field (a.1 Stage one), and video prediction pre-training for modeling the temporal dependency of videos (a.2 Stage two). (b) One transformer block in GenAD for the second stage training has interleaved temporal reasoning blocks before each frozen layer to align spatiotemporal features. (c) The proposed Temporal Reasoning Block includes one causal temporal attention (TA) and two decoupled spatial attention (SA) layers to extract features in different axes. A query grid attends to itself as well as blue grids while the dark gray grid is masked out in causal attention. 'Zero init' is appended at the end of each attention block to stabilize training." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 418, + 70, + 547, + 274 + ], + "blocks": [ + { + "bbox": [ + 418, + 70, + 547, + 274 + ], + "lines": [ + { + "bbox": [ + 418, + 70, + 547, + 274 + ], + "spans": [ + { + "bbox": [ + 418, + 70, + 547, + 274 + ], + "type": "image", + "image_path": "809d540c959eee6b5b0cab1933780f2c041de5aebf8a0957020f426af1a665fd.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 364, + 288, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 364, + 288, + 400 + ], + "spans": [ + { + "bbox": [ + 46, + 364, + 288, + 400 + ], + "type": "text", + "content": "temporal attention layer and two decoupled spatial attention layers, which are tailored for the causal reasoning and modeling large shifts in the driving scenes, respectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 406, + 287, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 406, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 406, + 287, + 647 + ], + "type": "text", + "content": "Causal Temporal Attention. Since the model after the stage-one training can only process each frame independently, we leverage temporal attention to exchange information among different video frames. The attention takes place in the time axis and models the temporal dependency of each grid-wise feature. However, directly adapting bidirectional temporal attention here as [4, 18, 46, 51] can hardly acquire the ability of causal reasoning, since the predictions will be inevitably dependent on the subsequent frames instead of past conditions. Therefore, we restrict the attention direction by adding a causal attention mask, as shown in the last row of Fig. 3(c), to encourage the model to fully exploit knowledge from past observations and faithfully reason about the future as if in real-world driving. We empirically found that the causality constraint greatly regularizes the predicted frames to be coherent with past frames. Following common practice, we also add temporal bias implemented as relative position embeddings on the time axis [42] to distinguish different frames of a sequence for temporal attention." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "Decoupled Spatial Attention. As driving videos feature fast perspective changes, features in a specific grid could vary greatly in different timesteps and are hard to correlate and learn by temporal attention, which suffers from a limited receptive field. In light of this, we introduce spa" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 364, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 364, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 364, + 545, + 449 + ], + "type": "text", + "content": "tial attention to propagate each grid feature in spatial axes to aid in gathering information for temporal attention. We implement a decoupled variant of self-attention for its efficiency with linear computational complexity, compared to quadratic full self-attention. As shown in Fig. 3(c), the two decoupled attention layers propagate features in horizontal and vertical axes, respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 452, + 545, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 452, + 545, + 561 + ], + "spans": [ + { + "bbox": [ + 304, + 452, + 545, + 561 + ], + "type": "text", + "content": "Deep Interaction. Intuitively, the spatial blocks fine-tuned in stage one refine features of each frame independently towards photorealism, whereas the temporal blocks introduced in stage two align features of all video frames towards coherency and consistency. To further boost the spatiotemporal feature interaction, we interleave the proposed temporal reasoning blocks with the original Transformer blocks in SDXL, i.e., spatial attention, cross attention, and feedforward network, as shown in Fig. 3(b)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 564, + 545, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 564, + 545, + 637 + ], + "spans": [ + { + "bbox": [ + 304, + 564, + 545, + 637 + ], + "type": "text", + "content": "Zero Initialization. Similar to the previous practices [1, 52], for each block that is newly introduced in stage two, we initialized all parameters of its final layer as zero. This avoids disrupting the prior knowledge of the well-trained image generation model in the beginning and stabilizes the training process." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "Training. GenAD is trained to predict the future by jointly denoising from the noisy latents with the guidance of past frames and text conditions. We first project " + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": " consecutive frames of a video clip into a batch of latents " + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{v} = \\{\\mathbf{v}^m, \\mathbf{v}^n\\}" + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": ", where the leading " + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": " frame latents " + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{v}^m" + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": " are clean, representing historical observations, and other " + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "inline_equation", + "content": "n = T - m" + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": " frame latents" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14666" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{v}^n" + }, + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": " indicate the future to be predicted. " + }, + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{v}^n" + }, + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": " are then corrupted to " + }, + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_t^n" + }, + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": " by the forward diffusion process, where " + }, + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": " indexes a randomly sampled noise scale. The model is trained to predict the noise of " + }, + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_t^n" + }, + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": " conditioned on observations " + }, + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{v}^m" + }, + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": " and text c. The learning objective of the video prediction model is formulated as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 149, + 287, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 149, + 287, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 149, + 287, + 171 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {v i d}} := \\mathbb {E} _ {\\mathbf {v}, \\boldsymbol {\\epsilon} \\sim \\mathcal {N} (0, 1), \\mathbf {c}, t} \\left[ \\| \\boldsymbol {\\epsilon} - \\mathbf {f} _ {\\theta , \\phi} \\left(\\mathbf {v} _ {t} ^ {n}; \\mathbf {v} ^ {m}, \\mathbf {c}, t\\right) \\| _ {2} ^ {2} \\right], \\tag {2}", + "image_path": "e3406fbaf0538a90ccaa72fcad9bc520cd13637feeabf2bf3738a87c00955e26.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 178, + 287, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 178, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 178, + 287, + 275 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 178, + 287, + 275 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 178, + 287, + 275 + ], + "type": "text", + "content": " denotes the inherited stage-one model and " + }, + { + "bbox": [ + 46, + 178, + 287, + 275 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 178, + 287, + 275 + ], + "type": "text", + "content": " represents the newly inserted temporal reasoning blocks. Following [4], we freeze " + }, + { + "bbox": [ + 46, + 178, + 287, + 275 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 178, + 287, + 275 + ], + "type": "text", + "content": " and only train the temporal reasoning blocks to avoid perturbing the generation ability of the image generation model and focus on learning temporal dependencies in videos. Notably, only the outputs from the corrupted frames " + }, + { + "bbox": [ + 46, + 178, + 287, + 275 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_t^n" + }, + { + "bbox": [ + 46, + 178, + 287, + 275 + ], + "type": "text", + "content": " contribute to the training loss while those from condition frames " + }, + { + "bbox": [ + 46, + 178, + 287, + 275 + ], + "type": "inline_equation", + "content": "\\mathbf{v}^m" + }, + { + "bbox": [ + 46, + 178, + 287, + 275 + ], + "type": "text", + "content": " are ignored." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 281, + 121, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 281, + 121, + 293 + ], + "spans": [ + { + "bbox": [ + 47, + 281, + 121, + 293 + ], + "type": "text", + "content": "3.3. Extensions" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 300, + 287, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 287, + 372 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 287, + 372 + ], + "type": "text", + "content": "Relying on the well-trained video prediction capability in driving scenarios, we further exploit the potential of the pretrained model in action-controlled prediction and planning, which are important for real-world driving systems. Here, we explore the downstream tasks on nuScenes [6] which provides recorded poses." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 374, + 287, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 374, + 287, + 482 + ], + "spans": [ + { + "bbox": [ + 46, + 374, + 287, + 482 + ], + "type": "text", + "content": "Action-conditioned Prediction. To make our predictive model controllable with exact ego actions and act as a simulator [25], we fine-tune the model with the paired future trajectory as an additional condition. Specifically, we map the raw trajectory to a high-dimensional feature with Fourier embeddings [44]. After further projection by a linear layer, it is added to the original conditions. Thus, the ego actions are injected into the network through the conditional cross-attention layer in Fig. 3(b)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 485, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 485, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 485, + 287, + 628 + ], + "type": "text", + "content": "Planning. By learning to predict the future, GenAD acquires strong representations of complex driving scenes, which can be further exploited for planning. Specifically, we extract spatiotemporal features of two historical frames through the UNet encoder of the frozen GenAD, which is nearly half the size of the entire model, and feed them to a multi-layer perceptron (MLP) to predict future waypoints. With the frozen GenAD encoder and a learnable MLP layer, the training process of our planner can be sped up by 3400 times compared to an end-to-end planning model UniAD [22], validating the effectiveness of the learned spatiotemporal feature of GenAD." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 639, + 128, + 652 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 639, + 128, + 652 + ], + "spans": [ + { + "bbox": [ + 47, + 639, + 128, + 652 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 659, + 163, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 659, + 163, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 659, + 163, + 672 + ], + "type": "text", + "content": "4.1. Setup and Protocols" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": "GenAD is learned in two stages on OpenDV-2K but with different learning objectives (in Sec. 3) and input formats. In stage one, the model takes input (image, text) pairs and" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 307, + 74, + 550, + 158 + ], + "blocks": [ + { + "bbox": [ + 307, + 74, + 550, + 158 + ], + "lines": [ + { + "bbox": [ + 307, + 74, + 550, + 158 + ], + "spans": [ + { + "bbox": [ + 307, + 74, + 550, + 158 + ], + "type": "table", + "html": "
MethodTraining DatasetPred.nuScenes
FID (↓)FVD (↓)
DriveGAN [25]✓73.4502
DriveDreamer* [45]nuScenes✓52.6452
DrivingDiffusion* [31]✗15.8332
GenAD-nus (Ours)nuScenes✓15.4244
GenAD (Ours)OpenDV-2K✓15.4184
", + "image_path": "85d117eb19d7f9aa80f575b1ab6e5f3499a61101716e5671fd5b76edb8fab99a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 166, + 545, + 200 + ], + "lines": [ + { + "bbox": [ + 305, + 166, + 545, + 200 + ], + "spans": [ + { + "bbox": [ + 305, + 166, + 545, + 200 + ], + "type": "text", + "content": "Table 2. Video generation quality compared to state-of-the-arts trained on nuScenes. \"Pred.\": evaluation by future prediction. *: requiring 3D layout inputs." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 214, + 545, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 214, + 545, + 394 + ], + "spans": [ + { + "bbox": [ + 304, + 214, + 545, + 394 + ], + "type": "text", + "content": "is trained on text-to-image generation. We broadcast the command annotation, which is labeled for each 4s video sequence, to all frames included. The model is trained for " + }, + { + "bbox": [ + 304, + 214, + 545, + 394 + ], + "type": "inline_equation", + "content": "300\\mathrm{K}" + }, + { + "bbox": [ + 304, + 214, + 545, + 394 + ], + "type": "text", + "content": " iterations on 32 NVIDIA Tesla A100 GPUs with a total batch size of 256. In the second stage, GenAD is trained to jointly denoise future latents conditioned on past latents and texts. Its inputs are (video clip, text) pairs where each video clip is 4s at " + }, + { + "bbox": [ + 304, + 214, + 545, + 394 + ], + "type": "inline_equation", + "content": "2\\mathrm{Hz}" + }, + { + "bbox": [ + 304, + 214, + 545, + 394 + ], + "type": "text", + "content": ". The current version of GenAD is trained on 64 GPUs for " + }, + { + "bbox": [ + 304, + 214, + 545, + 394 + ], + "type": "inline_equation", + "content": "112.5\\mathrm{K}" + }, + { + "bbox": [ + 304, + 214, + 545, + 394 + ], + "type": "text", + "content": " iterations with a total batch size of 64. The input frames are resized to " + }, + { + "bbox": [ + 304, + 214, + 545, + 394 + ], + "type": "inline_equation", + "content": "256 \\times 448" + }, + { + "bbox": [ + 304, + 214, + 545, + 394 + ], + "type": "text", + "content": " for training in both stages, and the text condition c is dropped at a probability of " + }, + { + "bbox": [ + 304, + 214, + 545, + 394 + ], + "type": "inline_equation", + "content": "p = 0.1" + }, + { + "bbox": [ + 304, + 214, + 545, + 394 + ], + "type": "text", + "content": " to enable classifier-free guidance [17] in sampling, which is commonly used in diffusion models to improve sample quality. More training and sampling details are in Appendix D." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 403, + 516, + 416 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 403, + 516, + 416 + ], + "spans": [ + { + "bbox": [ + 305, + 403, + 516, + 416 + ], + "type": "text", + "content": "4.2. Results of Video Prediction Pre-training" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 422, + 545, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 545, + 589 + ], + "type": "text", + "content": "Comparison to Recent Video Generation Approaches. We compare GenAD to recent advances on an unseen set with geofencing from OpenDV-YouTube, Waymo [43], KITTI [14], and Cityscapes [10] in a zero-shot generation manner. Fig. 4 depicts the qualitative results. Image-to-video models I2VGen-XL [53] and VideoCrafter1 [8] can not strictly follow the given frames to make predictions, yielding poor consistency between the predicted frames and past frames. The video prediction model DMVFN [21] that is trained on Cityscapes suffers from the unfavorable shape distortions in its predictions, especially on the three unseen datasets. In contrast, GenAD exhibits remarkable zero-shot generalization ability and visual quality although none of these sets are included in the training." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "Comparison to nuScenes Experts. We also compare GenAD with the most recent available driving video generation models which are exclusively trained for nuScenes. Tab. 2 shows that GenAD surpasses all previous methods in both image fidelity (FID) and video coherence (FVD). Specifically, GenAD significantly reduces FVD by " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "44.5\\%" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " compared to DrivingDiffusion [31], without taking 3D future layouts as additional inputs. For fair comparisons, we train a model variant (GenAD-nus) on nuScenes dataset only. We find that although GenAD-nus performs on par" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14667" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 70, + 324, + 364 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 324, + 364 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 324, + 364 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 324, + 364 + ], + "type": "image", + "image_path": "948e930580ca8d90ade38de77861e2b964f2a400dc91a3f38cbba7fac5625c41.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 365, + 546, + 398 + ], + "lines": [ + { + "bbox": [ + 46, + 365, + 546, + 398 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 546, + 398 + ], + "type": "text", + "content": "Figure 4. Task on zero-shot video prediction for unseen scenarios. We show the generation results (in blue boxes) of different models given the same starting frames. GenAD makes more robust, realistic, and reasonable future predictions on unseen datasets (scenarios). More comparisons and visualizations are shown in Appendix." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 326, + 70, + 545, + 363 + ], + "blocks": [ + { + "bbox": [ + 326, + 70, + 545, + 363 + ], + "lines": [ + { + "bbox": [ + 326, + 70, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 326, + 70, + 545, + 363 + ], + "type": "image", + "image_path": "f85e60b3a0cbcc1383e23dd21035f4caa820a272079bcec0e55dbd74c5155760.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 48, + 416, + 287, + 559 + ], + "blocks": [ + { + "bbox": [ + 48, + 416, + 287, + 559 + ], + "lines": [ + { + "bbox": [ + 48, + 416, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 48, + 416, + 287, + 559 + ], + "type": "image", + "image_path": "88f0cfdb8928df107f6dcc46f75535ce3398111076335205dd499c99fe188ca0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 567, + 287, + 601 + ], + "lines": [ + { + "bbox": [ + 46, + 567, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 567, + 287, + 601 + ], + "type": "text", + "content": "Figure 5. Task on language-conditioned prediction. Given two frames of a rainy scenario in the intersection and three high-level text conditions, GenAD simulates reasonable futures accordingly." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 627, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 627, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 627, + 287, + 687 + ], + "type": "text", + "content": "with GenAD on nuScenes, it struggles to generalize to unseen datasets like Waymo, where the generation degrades to the nuScenes visual pattern. In contrast, GenAD trained on OpenDV-2K exhibits strong generalization ability across datasets as shown in Fig. 4." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "We provide language-conditioned prediction samples on nuScenes in Fig. 5, where GenAD simulates various futures" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 419, + 546, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 419, + 546, + 467 + ], + "spans": [ + { + "bbox": [ + 304, + 419, + 546, + 467 + ], + "type": "text", + "content": "from the same start following different textual instructions. The impressive generation quality is exhibited in the intricate details of the environment, and the natural transition of ego motion." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 471, + 546, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 471, + 546, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 471, + 546, + 639 + ], + "type": "text", + "content": "Ablation Study. We perform ablations by training each variant on a subset of OpenDV-2K for 75K steps. Starting from the baseline with plain temporal attentions [4, 18], we gradually introduce our proposed components. Notably, by interleaving the temporal blocks with the spatial blocks, the FVD significantly improves " + }, + { + "bbox": [ + 304, + 471, + 546, + 639 + ], + "type": "inline_equation", + "content": "(-17\\%)" + }, + { + "bbox": [ + 304, + 471, + 546, + 639 + ], + "type": "text", + "content": " due to more sufficient spatiotemporal interactions. Both temporal causality and decoupled spatial attention contribute to better CLIP-SIM, improving the temporal consistency between future predictions and the condition frames. To be clear, the slight increase in FID and FVD, shown in fourth and third rows of Tab. 3 respectively, does not faithfully reflect a decline in generation quality as discussed in [4, 5, 37]. The effectiveness of each design is shown in Fig. 6." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 647, + 428, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 428, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 428, + 658 + ], + "type": "text", + "content": "4.3. Results of Extensions" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 665, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 546, + 715 + ], + "type": "text", + "content": "Action-conditioned Prediction. We further showcase the performance of the action-conditioned model fine-tuned on nuScenes, GenAD-act, in Fig. 7 and Tab. 4. Given two starting frames and a trajectory w composed of 6 future way" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14668" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 286, + 276 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 286, + 276 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 286, + 276 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 286, + 276 + ], + "type": "image", + "image_path": "cac3305aa3f6014f26b6ac0c03e4935457a9eee25b4ededa50bcdad6299f2dbc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 281, + 287, + 304 + ], + "lines": [ + { + "bbox": [ + 47, + 281, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 47, + 281, + 287, + 304 + ], + "type": "text", + "content": "Figure 6. Case study for model designs. All components help alleviate artifacts and improve the consistency of future predictions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 316, + 282, + 385 + ], + "blocks": [ + { + "bbox": [ + 50, + 316, + 282, + 385 + ], + "lines": [ + { + "bbox": [ + 50, + 316, + 282, + 385 + ], + "spans": [ + { + "bbox": [ + 50, + 316, + 282, + 385 + ], + "type": "table", + "html": "
MethodYouTube
FID (↓)FVD (↓)CLIPSIM (↑)
Baseline18.32244.440.8405
+ Deep Interaction17.96201.690.8409
+ Temporal Causality16.54207.450.8550
+ Decoupled Spatial Attn.17.67189.540.8652
", + "image_path": "4f3206e2a2ff8cc73aa1195b8a1aea4422fe2f6a451d22e6e5f24dfdde1bb63b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 66, + 422, + 266, + 483 + ], + "blocks": [ + { + "bbox": [ + 47, + 388, + 287, + 411 + ], + "lines": [ + { + "bbox": [ + 47, + 388, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 47, + 388, + 287, + 411 + ], + "type": "text", + "content": "Table 3. Ablation on model designs in GenAD. All proposed designs contribute to the final performance." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 422, + 266, + 483 + ], + "lines": [ + { + "bbox": [ + 66, + 422, + 266, + 483 + ], + "spans": [ + { + "bbox": [ + 66, + 422, + 266, + 483 + ], + "type": "table", + "html": "
MethodConditionnuScenes\nAction Prediction Error (↓)
Ground truth-0.90
GenADtext2.54
GenAD-acttext + traj.2.02
", + "image_path": "5d35442a6044a59cb021274383942f5df5d40b5ba33b222d3c379694f86c97a1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 541, + 287, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 541, + 287, + 673 + ], + "spans": [ + { + "bbox": [ + 46, + 541, + 287, + 673 + ], + "type": "text", + "content": "points, GenAD-act imagines 6 future frames following the trajectory sequence. To evaluate the consistency between the input trajectory " + }, + { + "bbox": [ + 46, + 541, + 287, + 673 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 46, + 541, + 287, + 673 + ], + "type": "text", + "content": " and predicted frames, we establish an inverse dynamics model (IDM) on nuScenes as the evaluator, which projects a video sequence into a corresponding ego trajectory. We leverage the IDM to translate predicted frames into the trajectory " + }, + { + "bbox": [ + 46, + 541, + 287, + 673 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{w}}" + }, + { + "bbox": [ + 46, + 541, + 287, + 673 + ], + "type": "text", + "content": ", and calculate the L2 distance between " + }, + { + "bbox": [ + 46, + 541, + 287, + 673 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 46, + 541, + 287, + 673 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 541, + 287, + 673 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{w}}" + }, + { + "bbox": [ + 46, + 541, + 287, + 673 + ], + "type": "text", + "content": " as the Action Prediction Error. Specifically, GenAD-act substantially reduces the Action Prediction Error by " + }, + { + "bbox": [ + 46, + 541, + 287, + 673 + ], + "type": "inline_equation", + "content": "20.4\\%" + }, + { + "bbox": [ + 46, + 541, + 287, + 673 + ], + "type": "text", + "content": " compared to GenAD with text condition, allowing for more accurate future simulations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "content": "Planning Results. Tab. 5 depicts the planning results on nuScenes where ground truth poses for the ego vehicle are available. By freezing GenAD encoder and only optimizing" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 306, + 73, + 545, + 220 + ], + "blocks": [ + { + "bbox": [ + 306, + 73, + 545, + 220 + ], + "lines": [ + { + "bbox": [ + 306, + 73, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 306, + 73, + 545, + 220 + ], + "type": "image", + "image_path": "cd7f531bb6325f055a4422f1202577063f1d2274d692faaa9b962d82dc87b944.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 227, + 545, + 283 + ], + "lines": [ + { + "bbox": [ + 305, + 227, + 545, + 283 + ], + "spans": [ + { + "bbox": [ + 305, + 227, + 545, + 283 + ], + "type": "text", + "content": "Figure 7. Task on action-conditioned prediction (simulation). Given the same starting frames and different future trajectories (shown in yellow dots in the first column), GenAD-act can simulate diverse futures following different ego intentions. More visualizations are in Appendix." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 330, + 304, + 520, + 369 + ], + "blocks": [ + { + "bbox": [ + 47, + 486, + 287, + 519 + ], + "lines": [ + { + "bbox": [ + 47, + 486, + 287, + 519 + ], + "spans": [ + { + "bbox": [ + 47, + 486, + 287, + 519 + ], + "type": "text", + "content": "Table 4. Task on action-conditioned prediction. Compared to GenAD with text conditions only, GenAD-act enables more precise future predictions that follow the action condition." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 330, + 304, + 520, + 369 + ], + "lines": [ + { + "bbox": [ + 330, + 304, + 520, + 369 + ], + "spans": [ + { + "bbox": [ + 330, + 304, + 520, + 369 + ], + "type": "table", + "html": "
Method# Trainable Params.nuScenes
ADE (↓)FDE (↓)
ST-P3* [20]10.9M2.653.73
UniAD* [22]58.8M1.031.65
GenAD (Ours)0.8M1.232.31
", + "image_path": "2ea0b54bed8889da117118c9ba874c32c3d41aefd46350aa20aa2a10863cb5ac.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 371, + 545, + 416 + ], + "lines": [ + { + "bbox": [ + 304, + 371, + 545, + 416 + ], + "spans": [ + { + "bbox": [ + 304, + 371, + 545, + 416 + ], + "type": "text", + "content": "Table 5. Task on open-loop planning. A lightweight MLP with frozen GenAD gets competitive planning results with " + }, + { + "bbox": [ + 304, + 371, + 545, + 416 + ], + "type": "inline_equation", + "content": "73 \\times" + }, + { + "bbox": [ + 304, + 371, + 545, + 416 + ], + "type": "text", + "content": " fewer trainable parameters and front-view image alone. *: multi-view inputs. Evaluation protocols are aligned with UniAD [22]." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 447, + 545, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 447, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 447, + 545, + 520 + ], + "type": "text", + "content": "an additional MLP on top of it, the model can effectively learn to plan. Notably, by pre-extracting image features through the UNet encoder of GenAD, the entire learning process for planning adaptation takes only 10 minutes on a single NVIDIA Tesla V100 device, which is 3400 times more efficient than the training of the UniAD planner [22]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 544, + 460, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 544, + 460, + 558 + ], + "spans": [ + { + "bbox": [ + 306, + 544, + 460, + 558 + ], + "type": "text", + "content": "5. Limitations and Discussion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "We study the system-level development of GenAD, a large-scale generalized video predictive model for autonomous driving. We also validate the adaptation of the learned representation of GenAD to driving tasks, i.e., learning a \"world model\" and motion planning. Although we obtain improved generalization to open domains, the increased model capacity poses challenges in both training efficiency and real-time deployment. We envision the unified video prediction task will serve as a scalable objective for future research on representation learning and policy learning. Another interesting direction involves distilling the encoded knowledge for a wider range of downstream tasks [29]." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14669" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 200 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 200 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 200 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andy Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022. 2, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 202, + 288, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 288, + 257 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 288, + 257 + ], + "type": "text", + "content": "[2] Mohammadhossein Bahari, Saeed Saadatnejad, Ahmad Rahimi, Mohammad Shaverdikondori, Amir Hossein Shahidzadeh, Seyed-Mohsen Moosavi-Dezfooli, and Alexandre Alahi. Vehicle trajectory prediction works, but not everywhere. In CVPR, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 259, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 259, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 259, + 287, + 293 + ], + "type": "text", + "content": "[3] Fan Bao, Chongxuan Li, Jiacheng Sun, and Jun Zhu. Why are conditional generative models better than unconditional ones? arXiv preprint arXiv:2212.00362, 2022. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 294, + 287, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 294, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 53, + 294, + 287, + 338 + ], + "type": "text", + "content": "[4] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 2, 5, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 339, + 287, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 339, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 339, + 287, + 384 + ], + "type": "text", + "content": "[5] Tim Brooks, Janne Hellsten, Miika Aittala, Ting-Chun Wang, Timo Aila, Jaakko Lehtinen, Ming-Yu Liu, Alexei Efros, and Tero Karras. Generating long videos of dynamic scenes. In NeurIPS, 2022. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 385, + 287, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 385, + 287, + 440 + ], + "spans": [ + { + "bbox": [ + 53, + 385, + 287, + 440 + ], + "type": "text", + "content": "[6] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yuxin Pan, Giancarlo Baldan, and Oscar Beijbom. nuScenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 3, 4, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 441, + 287, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 441, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 53, + 441, + 287, + 497 + ], + "type": "text", + "content": "[7] Holger Caesar, Juraj Kabzan, Kok Seang Tan, Whye Kit Fong, Eric M. Wolff, Alex Lang, Luke Fletcher, Oscar Bei-jbom, and Sammy Omari. nuPlan: A closed-loop ml-based planning benchmark for autonomous vehicles. In CVPR Workshops, 2021. 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 499, + 287, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 499, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 53, + 499, + 287, + 563 + ], + "type": "text", + "content": "[8] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, Chao Weng, and Ying Shan. VideoCrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 567, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 567, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 53, + 567, + 287, + 609 + ], + "type": "text", + "content": "[9] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers. arXiv preprint arXiv:2306.16927, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 612, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 612, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 612, + 287, + 667 + ], + "type": "text", + "content": "[10] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The Cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "text", + "content": "[11] Yilun Dai, Mengjiao Yang, Bo Dai, Hanjun Dai, Ofir Nachum, Josh Tenenbaum, Dale Schuurmans, and Pieter Abbeel. Learning universal policies via text-guided video generation. In NeurIPS, 2023. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[12] Thierry Deruyttere, Simon Vandenhende, Dusan Grujicic, Luc Van Gool, and Marie-Francine Moens. Talk2Car: Taking control of your self-driving car. In EMNLP, 2019. 3, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 118, + 545, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 148 + ], + "type": "text", + "content": "[13] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 150, + 545, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 150, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 307, + 150, + 545, + 182 + ], + "type": "text", + "content": "[14] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The KITTI dataset. *IJRR*, 2013. 3, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 183, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 183, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 183, + 545, + 217 + ], + "type": "text", + "content": "[15] Agrim Gupta, Stephen Tian, Yunzhi Zhang, Jiajun Wu, Roberto Martin-Martin, and Li Fei-Fei. MaskViT: Masked visual pre-training for video prediction. In ICLR, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 217, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 217, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 307, + 217, + 545, + 259 + ], + "type": "text", + "content": "[16] Niklas Hanselmann, Katrin Renz, Kashyap Chitta, Apratim Bhattacharyya, and Andreas Geiger. KING: Generating safety-critical driving scenarios for robust imitation via kinematics gradients. In ECCV, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 261, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 261, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 307, + 261, + 545, + 282 + ], + "type": "text", + "content": "[17] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS Workshops, 2021. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 283, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 283, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 307, + 283, + 545, + 316 + ], + "type": "text", + "content": "[18] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. In NeurIPS, 2022. 2, 5, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 316, + 545, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 316, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 307, + 316, + 545, + 369 + ], + "type": "text", + "content": "[19] Anthony Hu, Lloyd Russell, Hudson Yeo, Zak Murez, George Fedoseev, Alex Kendall, Jamie Shotton, and Gianluca Corrado. GAIA-1: A generative world model for autonomous driving. arXiv preprint arXiv:2309.17080, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 371, + 545, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 371, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 307, + 371, + 545, + 414 + ], + "type": "text", + "content": "[20] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. ST-P3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In ECCV, 2022. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 415, + 545, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 415, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 307, + 415, + 545, + 448 + ], + "type": "text", + "content": "[21] Xiaotao Hu, Zhewei Huang, Ailin Huang, Jun Xu, and Shuchang Zhou. A dynamic multi-scale voxel flow network for video prediction. In CVPR, 2023. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 449, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 545, + 502 + ], + "type": "text", + "content": "[22] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, Lewei Lu, Xiaosong Jia, Qiang Liu, Jifeng Dai, Yu Qiao, and Hongyang Li. Planning-oriented autonomous driving. In CVPR, 2023. 6, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 504, + 545, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 504, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 307, + 504, + 545, + 548 + ], + "type": "text", + "content": "[23] Fan Jia, Weixin Mao, Yingfei Liu, Yucheng Zhao, Yuqing Wen, Chi Zhang, Xiangyu Zhang, and Tiancai Wang. A driver-i: A general world model for autonomous driving. arXiv preprint arXiv:2311.13549, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 548, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 548, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 307, + 548, + 545, + 581 + ], + "type": "text", + "content": "[24] Jinkyu Kim, Teruhisa Miso, Yi-Ting Chen, Ashish Tawari, and John Canny. Grounding human-to-vehicle advice for self-driving vehicles. In CVPR, 2019. 3, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 582, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 582, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 307, + 582, + 545, + 613 + ], + "type": "text", + "content": "[25] Seung Wook Kim, Jonah Philion, Antonio Torralba, and Sanja Fidler. DriveGAN: Towards a controllable high-quality neural simulation. In CVPR, 2021. 2, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 614, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 614, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 307, + 614, + 545, + 658 + ], + "type": "text", + "content": "[26] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In ICCV, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 658, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 689 + ], + "type": "text", + "content": "[27] Yann LeCun. A path towards autonomous machine intelligence version 0.9. 2, 2022-06-27. Open Review, 62, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[28] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. Nature, 2015. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14670" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[29] Daiqing Li, Huan Ling, Amlan Kar, David Acuna, Seung Wook Kim, Karsten Kreis, Antonio Torralba, and Sanja Fidler. DreamTeacher: Pretraining image backbones with deep generative models. In ICCV, 2023. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 117, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 117, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 117, + 287, + 160 + ], + "type": "text", + "content": "[30] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 2, 3, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 161, + 288, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 161, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 161, + 288, + 205 + ], + "type": "text", + "content": "[31] Xiaofan Li, Yifu Zhang, and Xiaoqing Ye. DrivingDiffusion: Layout-guided multi-view driving scene video generation with latent diffusion model. arXiv preprint arXiv:2310.07771, 2023. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 205, + 287, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 205, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 48, + 205, + 287, + 249 + ], + "type": "text", + "content": "[32] Zhenyu Li, Zehui Chen, Ang Li, Liangji Fang, Qinhong Jiang, Xianming Liu, and Junjun Jiang. Unsupervised domain adaptation for monocular 3d object detection via self-training. In ECCV, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 250, + 287, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 287, + 294 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 287, + 294 + ], + "type": "text", + "content": "[33] Jiachen Lu, Ze Huang, Jiahui Zhang, Zeyu Yang, and Li Zhang. Wovogen: World volume-aware diffusion for controllable multi-camera driving scene generation. arXiv preprint arXiv:2312.02934, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 294, + 287, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 294, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 294, + 287, + 348 + ], + "type": "text", + "content": "[34] Jiageng Mao, Minzhe Niu, Chenhan Jiang, Hanxue Liang, Xiaodan Liang, Yamin Li, Chao Ye, Wei Zhang, Zhenguo Li, Jie Yu, Hang Xu, and Chunjing Xu. One million scenes for autonomous driving: ONCE dataset. In NeurIPS Datasets and Benchmarks, 2021. 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 349, + 287, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 349, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 349, + 287, + 370 + ], + "type": "text", + "content": "[35] OpenAI. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 371, + 287, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 448 + ], + "type": "text", + "content": "[36] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul F Christiano, Jan Leike, and Ryan Lowe. Training language models to follow instructions with human feedback. In NeurIPS, 2022. 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 449, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 502 + ], + "type": "text", + "content": "[37] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 4, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 503, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 503, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 503, + 287, + 536 + ], + "type": "text", + "content": "[38] Xiaojuan Qi, Zhengzhe Liu, Qifeng Chen, and Jiaya Jia. 3D motion decomposition for rgbd future dynamic scene synthesis. In CVPR, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 536, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 602 + ], + "type": "text", + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 602, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 646 + ], + "type": "text", + "content": "[40] Vasili Ramanishka, Yi-Ting Chen, Teruhisa Mitsu, and Kate Saenko. Toward driving scene understanding: A dataset for learning driver behavior and causal reasoning. In CVPR, 2018. 3, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 680 + ], + "type": "text", + "content": "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[42] Peter Shaw, Jakob Uszkoreit, and Ashish Vaswani. Self-attention with relative position representations. arXiv preprint arXiv:1803.02155, 2018. 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 160 + ], + "type": "text", + "content": "[43] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott M. Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, 2020. 3, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 161, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 161, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 161, + 545, + 217 + ], + "type": "text", + "content": "[44] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. In NeurIPS, 2020. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 218, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 218, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 218, + 545, + 262 + ], + "type": "text", + "content": "[45] Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, and Jiwen Lu. DriveDreamer: Towards real-world-driven world models for autonomous driving. arXiv preprint arXiv:2309.09777, 2023. 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 263, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 263, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 308, + 263, + 545, + 341 + ], + "type": "text", + "content": "[46] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, Yuwei Guo, Tianxing Wu, Chenyang Si, Yuming Jiang, Cunjian Chen, Chen Change Loy, Bo Dai, Dahua Lin, Yu Qiao, and Ziwei Liu. LAVIE: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 342, + 545, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 342, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 308, + 342, + 545, + 396 + ], + "type": "text", + "content": "[47] Yuqi Wang, Jiawei He, Lue Fan, Hongxin Li, Yuntao Chen, and Zhaoxiang Zhang. Driving into the future: Multiview visual forecasting and planning with world model for autonomous driving. arXiv preprint arXiv:2311.17918, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 398, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 398, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 308, + 398, + 545, + 464 + ], + "type": "text", + "content": "[48] Benjamin Wilson, William Qi, Tanmay Agarwal, John Lambert, Jagjeet Singh, Siddhesh Khandelwal, Bowen Pan, Ratnesh Kumar, Andrew Hartnett, Jhony Kaesemodel Pontes, Deva Ramanan, Peter Carr, and James Hays. Argoverse 2: Next generation datasets for self-driving perception and forecasting. In NeurIPS Datasets and Benchmarks, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 466, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 466, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 308, + 466, + 545, + 498 + ], + "type": "text", + "content": "[49] Daniel M Wolpert, Zoubin Ghahramani, and Michael I Jordan. An internal model for sensorimotor integration. Science, 1995. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 500, + 545, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 500, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 308, + 500, + 545, + 599 + ], + "type": "text", + "content": "[50] Zhiheng Xi, Wenxiang Chen, Xin Guo, Wei He, Yiwen Ding, Boyang Hong, Ming Zhang, Junzhe Wang, Senjie Jin, Enyu Zhou, Rui Zheng, Xiaoran Fan, Xiao Wang, Liao Xiong, Yuhao Zhou, Weiran Wang, Changhao Jiang, Yicheng Zou, Xiangyang Liu, Zhangyue Yin, Shihan Dou, Rongxiang Weng, Wensen Cheng, Qi Zhang, Wenjuan Qin, Yongyan Zheng, Xipeng Qiu, Xuanjing Huang, and Tao Gui. The rise and potential of large language model based agents: A survey. arXiv preprint arXiv:2309.07864, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 600, + 545, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 600, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 308, + 600, + 545, + 654 + ], + "type": "text", + "content": "[51] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. arXiv preprint arXiv:2309.15818, 2023. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 656, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 656, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 656, + 545, + 689 + ], + "type": "text", + "content": "[52] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 690, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 690, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 690, + 545, + 714 + ], + "type": "text", + "content": "[53] Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qin, Xiang Wang, Deli Zhao, and" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14671" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 129 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "Jingren Zhou. I2VGen-XL: High-quality image-to-video synthesis via cascaded diffusion models. arXiv preprint arXiv:2311.04145, 2023. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 129 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 129 + ], + "type": "text", + "content": "[54] Ruizhao Zhu, Peng Huang, Eshed Ohn-Bar, and Venkatesh Saligrama. Learning to drive anywhere. In CoRL, 2023. 2" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14672" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Generalizing 6-DoF Grasp Detection via Domain Prior Knowledge/65c67c8c-a659-4259-9daf-ad24fd3765f3_content_list.json b/2024/Generalizing 6-DoF Grasp Detection via Domain Prior Knowledge/65c67c8c-a659-4259-9daf-ad24fd3765f3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e40f0fc86b42227ede17a88d19c3fd10f0e6a39b --- /dev/null +++ b/2024/Generalizing 6-DoF Grasp Detection via Domain Prior Knowledge/65c67c8c-a659-4259-9daf-ad24fd3765f3_content_list.json @@ -0,0 +1,1494 @@ +[ + { + "type": "text", + "text": "Generalizing 6-DoF Grasp Detection via Domain Prior Knowledge", + "text_level": 1, + "bbox": [ + 148, + 130, + 821, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haoxiang Ma $^{1,2}$ Modi Shi $^{1,2}$ Boyang Gao $^{3,4}$ Di Huang $^{1,2*}$", + "bbox": [ + 236, + 179, + 738, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ State Key Laboratory of Software Development Environment, Beihang University, Beijing, China \n $^{2}$ School of Computer Science and Engineering, Beihang University, Beijing, China \n $^{3}$ School of Computer Science and Technology, Harbin Institute of Technology, Harbin, China \n $^{4}$ Geometry Robotics", + "bbox": [ + 98, + 199, + 872, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{mahaoxiang822, modishi, dhuang}@buaa.edu.cn, {boyang.gao}@geometryrobot.com", + "bbox": [ + 151, + 271, + 803, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 321, + 313, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We focus on the generalization ability of the 6-DoF grasp detection method in this paper. While learning-based grasp detection methods can predict grasp poses for unseen objects using the grasp distribution learned from the training set, they often exhibit a significant performance drop when encountering objects with diverse shapes and structures. To enhance the grasp detection methods' generalization ability, we incorporate domain prior knowledge of robotic grasping, enabling better adaptation to objects with significant shape and structure differences. More specifically, we employ the physical constraint regularization during the training phase to guide the model towards predicting grasps that comply with the physical rule on grasping. For the unstable grasp poses predicted on novel objects, we design a contact-score joint optimization using the projection contact map to refine these poses in cluttered scenarios. Extensive experiments conducted on the GraspNet1billion benchmark demonstrate a substantial performance gain on the novel object set and the real-world grasping experiments also demonstrate the effectiveness of our generalizing 6-DoF grasp detection method. Code is available at https://github.com/mahaoxiang822/Generalizing-Grasp.", + "bbox": [ + 75, + 353, + 473, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 729, + 209, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Given an object, robotic grasp detection aims to find suitable and sufficient gripper configurations for various manipulation tasks. Traditional methods [6, 27, 30] establish hand-crafted criteria to evaluate grasp samples according to 3D models of objects. Despite providing precise interaction between objects and grippers, they assume that object models are available in advance and also suffer a slow running speed, which limits their popularization. As the develop-", + "bbox": [ + 75, + 755, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2b83e1a29032bc4276eb2195c87004d7045fae79e5169984924e3d834b8b9787.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 532, + 321, + 864, + 474 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/406f56a7c2c9e2ba8a7008c28b895a9562dfae83e899c531e6d42950f5ff542f.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 531, + 479, + 864, + 540 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8967d8b099bbeea45c76cfe0c378003d1b6b48fcdfb700c63895f5f0a426ce40.jpg", + "image_caption": [ + "(b)", + "Figure 1. Illustration of (a) the performance and distribution gap between objects similar to training samples (similar objects) and those with largely varied shapes and structures (novel objects) and (b) the limitations of the object augmentation methods." + ], + "image_footnote": [], + "bbox": [ + 539, + 556, + 856, + 612 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ment of deep learning, data-driven methods [9, 20, 32] have been widely studied. These methods predict grasps without the need for pre-prepared 3D models and demonstrate the capability to handle unseen objects. However, they often struggle when encountering objects whose shapes and structures significantly deviate from those in the training set, making them difficult to adapt to more diverse applications. As illustrated in Fig. 1 (a), learning-based grasp detection methods are basically able to transfer to objects similar to those appearing in training, but they do not effectively generalize to novel objects with their distribution greatly changed.", + "bbox": [ + 496, + 703, + 893, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To facilitate the generalization of learning-based grasp", + "bbox": [ + 519, + 885, + 890, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18102", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "detection methods towards a variety of unseen objects, previous attempts usually apply data augmentation techniques to expand the distribution of the training set. More objects are typically produced by randomly concatenating a number of 3D primitives [1, 33] or by directly synthesizing through generative networks [13, 23]. They do enrich the diversity of objects for training, which leads to a gain in overall performance. However, since such augmentation is conducted only according to original data, it contributes more to similar objects whose distribution is largely overlapped and offers little help to new objects whose distribution is not well covered. As depicted in Fig. 1 (b), on the one hand, to the objects that are far from the center of training samples in the feature space, totally wrong grasps could be delivered. On the other hand, for the ones that are near the center of training samples in the feature space, although reasonable grasps may be obtained, the predictions tend to be unstable due to the distribution gap. Both cases make those methods problematic to generalize to unseen targets.", + "bbox": [ + 76, + 90, + 472, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the dilemma above, inspired by informed machine learning [15, 34], we introduce domain prior knowledge of robotic grasping. Compared to data augmentation, domain prior knowledge does not depend on the distribution of training data, allowing for easy adaptation to objects with significant shape and structure differences. Firstly, to enable the grasp detection network to generalize to novel objects, we incorporate certain physical rules on grasping, such as force conditions [26] and contact positions, which are typically used in the traditional analytical grasping methods. The physical rules provide valuable clues in terms of grasp stability regardless of object-specific properties, thus motivating us to integrate such rule-based priors into the grasp detection network. Secondly, to deal with the unstable results predicted by the network, we introduce another type of knowledge about the interaction of the gripper and object. The contact map [2] is able to denote the regions correlated with grasping as well as the kinematics of the gripper. By employing a neural network to encode the preferred object regions for optimal grasps, the learning-based contact map prior can be utilized to refine insufficiently accurate grasps.", + "bbox": [ + 76, + 382, + 472, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Concretely, in this paper, we propose a generalized 6-DoF grasp detection framework with domain prior knowledge, which consists of two components: Physical Constraint Regularization (PCR) and Contact-Score Joint Optimization (C-SJO). In PCR, we integrate the physical prior into the network as the regularization, thereby constraining the correlation between gripper poses and object models. To enable the back-propagation of PCR, we employ an end-to-end 6-DoF grasp detection network for grasp prediction and utilize the Signed Distance Field (SDF) to encode the object model, facilitating the differentiable computation of physical constraints. Compared to fitting grasp annotations directly, PCR guides the network to predict grasps follow", + "bbox": [ + 76, + 704, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ing object-independent physical rules, thereby enhancing the generalization capability for novel objects. To refine the unstable prediction for novel objects, we introduce the C-SJO at test time based on the contact map prior. The contact map represents the contact region on the object's surface by calculating the distance between the gripper model and the objects. A contact map prediction network is introduced to encode the contact pattern of good grasps by learning from grasp labels. By aligning the contact map of the current grasp and the prediction from the network, the grasp pose can be optimized within its neighborhood. However, as for 6-DoF grasp detection in clutter, using the euclidean distance for contact map calculation cannot adequately address inaccurate contact positions. Therefore, we introduce a projection contact map to solve this problem. Additionally, due to the noise from the depth sensor and the occlusion in cluttered scenarios, relying solely on the contact prior can lead to singular results. To mitigate this, we use score optimization to constrain the searching space of contact optimization, in which a grasp score network is utilized.", + "bbox": [ + 496, + 90, + 893, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contribution of this paper can be summarized as:", + "bbox": [ + 517, + 407, + 870, + 422 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a domain prior knowledge informed 6-DoF grasp detection framework to enhance the generalization ability for novel objects.", + "- We design the physical constraint regularization to represent the physical prior for 6-DoF grasping and integrate it into the network in a differential strategy.", + "- We employ a contact-score joint optimization with the projection contact map, which is suitable for refining the inaccurate prediction in cluttered scenarios." + ], + "bbox": [ + 500, + 425, + 890, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 500, + 573, + 635, + 588 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "6-DoF Grasp Detection To generate diverse and feasible grasping in cluttered scenes, 6-DoF grasp detection has been advanced recently. Compared to planar grasp detection where the grasp space is limited [9, 22, 28], 6-DoF grasp detection method can predict grasping in $SE(3)$ space, thereby supporting more complex downstream task. [32] pioneers a sampling and evaluation framework for 6-DoF grasp detection. Using point-cloud input, they heuristically sample grasp candidates and employ a Convolutional Neural Network (CNN) for scoring them. [16] build upon the framework, utilizing a neural network for point-cloud data to achieve improved grasp sample evaluation. More recently, a series of studies [3, 8, 10, 14, 19, 24, 29, 35] propose end-to-end strategy for grasp prediction. [24] employs a variational auto-encoder for grasp generation given the object point-clouds. [29] introduces a single-shot grasp proposal network for efficient grasp detection based on single view point-clouds, and [3] employs a Truncated Signed Distance Function (TSDF) to map multiple frames and predict grasps in voxel space. [10] provides a large benchmark", + "bbox": [ + 496, + 598, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18103", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "with densely annotated grasp labels and designs a baseline method for dense grasp prediction. Subsequently, [35] defines graspsness to represent grasp probability in searching space and [19] focuses on the scale imbalance problem for 6-DoF grasp detection. With the development of neural representations, [14] utilizes an occupancy network to learn a shared representation between 3D reconstruction and grasp detection. [8] introduces a generalizable neural radiance field for the grasp detection of transparent and specular objects. Despite these advancements in 6-DoF grasp detection, the performance for novel objects, particularly those with varied shapes and structures, remains sub-optimal. In this paper, we investigate this shortcoming.", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generalization on Grasp Detection Several strategies are proposed to enhance the generalization capability of grasp detection, which mainly focuses on enriching the distribution of training objects. [1] procedurally generates objects by attaching rectangular prisms at random locations and orientations. To enhance the shape diversity, [33] constructs a set of object primitives by decomposing everyday objects and generates diverse objects by randomly sampling these primitives and combing them. In addition to heuristically generating random objects, [23] provides the Evolved Grasping Analysis Dataset (EGAD), which includes objects of varying shape complexity and grasp difficulty generated by the 3D compositional pattern producing networks. [13] employs an AutoEncoder-Critic network to interpolate new shapes from two objects for augmentation. Besides augmenting current shapes to improve the performance of the learning-based grasp detection method, [36] explores the task of generating adversarial objects that are difficult to grasp. Different from previous augmentation methods for generalized grasp detection, we introduce the grasp domain knowledge to our framework explicitly. Therefore, our method doesn't rely on the distribution of training objects, resulting in an improvement for out-of-distribution objects.", + "bbox": [ + 75, + 292, + 472, + 641 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Usage of domain knowledge in Grasping Domain prior knowledge has been leveraged in various grasping applications. Early analytical methods [11, 21, 31] make use of the physical prior knowledge to examine the force dynamics between the gripper and the object. More recently, several methods incorporate some grasp prior knowledge into the hand-object grasp synthesis. [18] introduces a differentiable force closure algorithm designed to optimize hand configurations, which facilitates the fast generation of physically stable grasps. [17] designs a generalized Q1 metric that serves as a loss for a grasp planner, producing precise multifinger grasping for single object with watertight model or rendered depth images. While the aforementioned methods employ domain prior knowledge to assist grasp synthesis and optimization, they rely on the accurate object shapes and design complex calculations for prior integration. This makes it difficult to apply them in the 6-DoF grasp detection", + "bbox": [ + 75, + 643, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the generation of diverse grasps in complex scenarios is required and the object geometry from the depth sensor is inaccurate.", + "bbox": [ + 496, + 90, + 893, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 498, + 148, + 591, + 164 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Physical Constraint Regularization", + "text_level": 1, + "bbox": [ + 498, + 172, + 805, + 190 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In terms of physical prior knowledge, the force interaction between the object and the gripper is used to analyze grasp stability, among which the force closure [26] is widely adopted. A force closure grasp can resist any external wrenches with the contact force if the force direction lies in the friction cone. In scenarios where the friction coefficient of the object surface is unknown and a two-finger gripper is utilized, we employ the antipodal rule [5] as a simplification. Given two contact points $(u_{1}, u_{2})$ lying on the object surface, the antipodal rule is formulated as:", + "bbox": [ + 496, + 196, + 893, + 349 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ p \\left(u _ {1}\\right) - p \\left(u _ {2}\\right) \\right] \\cdot t \\left(u _ {1}\\right) = 0 \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 357, + 890, + 377 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n[ p (u _ {2}) - p (u _ {1}) ] \\cdot t (u _ {2}) = 0 \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 387, + 890, + 406 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nn \\left(u _ {1}\\right) + n \\left(u _ {2}\\right) = 0 \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 414, + 890, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $p(u)$ represents the contact position, $t(u)$ represents the unit tangent vector and $n(u)$ represents the unit outward normal vector. If the contact points $(u_1, u_2)$ between the gripper and the object comply with the antipodal rule, there is a higher probability of a successful grasp. Based on which, we introduce the PCR to constrain the output of the network with the rule.", + "bbox": [ + 496, + 436, + 890, + 540 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With the predicted grasp pose $g = [t, R, w]$ and the gripper kinematic model $K$ , the gripper contacts $(c_1, c_2)$ can be expressed as:", + "bbox": [ + 496, + 541, + 893, + 585 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left(c _ {1}, c _ {2}\\right) = K (t, R, w) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 595, + 890, + 614 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To enforce the gripper contacts to comply with the antipodal constraint, we propose an antipodal regularization term, which is calculated as follows:", + "bbox": [ + 496, + 622, + 890, + 667 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nR ^ {A} \\left(c _ {1}, c _ {2}\\right) = 1 - 0. 5 * \\left(\\cos \\left(\\overrightarrow {c _ {1} c _ {2}}, n \\left(c _ {2}\\right)\\right) + \\cos \\left(\\overrightarrow {c _ {2} c _ {1}}, n \\left(c _ {1}\\right)\\right)\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 676, + 890, + 708 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\overrightarrow{c_1c_2}$ is the vector connecting the gripper's contact points $c_{1}$ and $c_{2}$ , $n(c_{1})$ and $n(c_{2})$ are the normal vectors at the respective contact points and $\\cos$ represents the cosine similarity. Nevertheless, the predicted gripper contacts $c_{1}, c_{2}$ may not lie precisely on the object surface. To overcome this, we also introduce constraints on the distance between the gripper contacts and the object's surface. These constraints help to avoid collision and ensure that the contact points are sufficiently close to the object's surface. The collision constraint $R^C$ and surface constraint $R^S$ are formulated as:", + "bbox": [ + 496, + 709, + 893, + 875 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nR ^ {C} \\left(c _ {1}, c _ {2}\\right) = M a x (0, \\theta - d \\left(c _ {1}\\right)) + M a x (0, \\theta - d \\left(c _ {2}\\right)) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 883, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18104", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4b2c0601a788be80ddfa67e9424d92418051a31807ecdfd9e81f79906880d71a.jpg", + "image_caption": [ + "Figure 2. The pipeline of differential physical constraint integration. With the fusion point-cloud, the differentiable 6-DoF grasp network predicts the grasp configurations. The position of contacts $c_{1}, c_{2}$ are calculated from grasp configurations by the gripper model and the regularization $R$ is computed from the object SDF for back-propagation." + ], + "image_footnote": [], + "bbox": [ + 142, + 89, + 831, + 268 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nR ^ {S} \\left(c _ {1}, c _ {2}\\right) = M a x (0, d \\left(c _ {1}\\right) - \\mu) + M a x (0, d \\left(c _ {2}\\right) - \\mu) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 340, + 468, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the contact distance $d(c)$ is constrained between $[\\theta, \\mu]$ . Besides, we use the grasp score predicted by the grasp detection network to weight the regularization of different grasps, thereby ensuring the coherence between the grasp score and the physical constraints. With the grasp score of $i$ th grasp $s_i$ , the overall physical constraint $R_i$ can be represented as:", + "bbox": [ + 76, + 368, + 470, + 473 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nR _ {i} = \\frac {s _ {i} * \\left(R _ {i} ^ {A} + R _ {i} ^ {C} + R _ {i} ^ {S}\\right)}{\\frac {1}{M} \\sum_ {j} ^ {M} s _ {j}} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 479, + 468, + 536 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $M$ is the number of seed points. The overall loss $L$ can be formulated as:", + "bbox": [ + 76, + 542, + 468, + 571 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {\\text {g r a s p}} + \\phi * R \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 582, + 468, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For more details of the grasp configuration loss $L_{\\text{grasp}}$ , please refer to the supplementary.", + "bbox": [ + 75, + 607, + 468, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Differential Physical Constraint Integration", + "text_level": 1, + "bbox": [ + 76, + 646, + 449, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To integrate the PCR into the grasp detection network, the sub-gradients of normal vector $\\frac{\\partial n(c)}{\\partial c}$ and surface distance $\\frac{\\partial d(c)}{\\partial c}$ with respect to the contact point $c$ should be computed. Besides, the generation of grasp poses should be differential. As a result, we introduce an end-to-end 6-DoF grasp network and the differentiable Signed Distance Function (SDF) for the calculation of PCR. As shown in Fig. 2, the configurations of the gripper are regressed from the input fusion point-cloud in an end-to-end manner. With the configurations predicted by the grasp detection network, the contact point $c$ can be calculated by the gripper kinematics model. Utilizing SDF of the object, we can query the surface distance of the contact $c_{1}$ directly by $d(c_{1}) = SDF(c_{1})$ and the normal vector of $c_{1}$ can be calculated by $n(c_{1}) = \\nabla SDF(c_{1})$ . In practice, obtaining an", + "bbox": [ + 75, + 669, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "object's SDF directly is not feasible. Consequently, we employ a 3D grid to represent the object's SDF, where each grid point records the corresponding SDF value. During training, the SDF value at any arbitrary position is computed using tri-linear interpolation, which allows for the differentiable computation of the gradients required by the physical constraints.", + "bbox": [ + 496, + 332, + 893, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Contact-Score Joint Optimization", + "text_level": 1, + "bbox": [ + 498, + 446, + 795, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Due to the significant discrepancy in shape and structure between novel and training objects, the grasp prediction from the network can be unstable, which leads to failures. To solve this problem, inspired by dexterous grasp synthesis [12, 37], we introduce the contact map prior which encodes the preference contact region on the object to refine inaccurate grasps. Given the object point-cloud $X_{o}$ and gripper contacts $c_{1}, c_{2}$ , we define the contact map of each point $p_{i} \\in X_{o}$ for a 6-DoF grasp as:", + "bbox": [ + 496, + 470, + 893, + 607 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nD _ {i} = \\min _ {j} \\| p _ {i} - c _ {j} \\|, c _ {j} \\in \\left\\{c _ {1}, c _ {2} \\right\\} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 617, + 890, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Compared with optimization for a single object with accurate geometry in grasp synthesis, for the 6-DoF grasp detection task in cluttered scenes, there exist two issues that make the test-time optimization impractical and cause singular results: (1) the contact points don't always lie on the object surface; (2) the object point-clouds in real-world scenarios are inaccurate and incomplete. For the first issue, we propose the projection contact distance $PD$ illustrated in Fig. 3 (a) as a supplementary of original contact distance $D$ , which calculates the projection distance from the object point-cloud to the connecting line of two contacts.", + "bbox": [ + 496, + 650, + 893, + 816 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP D _ {i} = D _ {i} * \\sin \\left(\\theta_ {i}\\right) \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 828, + 890, + 845 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We visualize the original contact map and the proposed projection contact map in Fig. 3 (b). Rather than computing the contact distance directly, the projection contact", + "bbox": [ + 496, + 854, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18105", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/826c4048089a00cbed616167ff9ac253714145b1a46cda9df6bd412c611998ec.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 98, + 88, + 294, + 223 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/11383aa2771d4bb74fe7562a794ee768994b10d8d41338d00924358cd0acdc57.jpg", + "image_caption": [ + "(b)", + "Figure 3. (a) Calculation of the contact map and (b) comparison of the projection contact map and the original version." + ], + "image_footnote": [], + "bbox": [ + 295, + 88, + 450, + 223 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "map can highlight the accurate contact region on the object point-cloud for coarse grasp poses whose contact points may not lie on the surface. For contact map optimization, as shown in Fig. 4 (a), we employ a ContactNet for contact map prediction, which encodes the contact prior. Given the object point-cloud $X_{o}$ , the gripper point-cloud $X_{g}$ and the predicted grasp pose $g$ , the ContactNet can be formulated as:", + "bbox": [ + 75, + 295, + 468, + 412 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {D}, \\hat {P D} = \\operatorname {C o n t a c t N e t} \\left(g X _ {g}, X _ {o}\\right) \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 415, + 468, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and thus the contact optimization target $J_{c}$ is formulated as:", + "bbox": [ + 76, + 439, + 468, + 455 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nJ _ {c} = | D - \\hat {D} | + \\alpha * | P D - \\hat {P D} | \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 468, + 468, + 486 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the second issue, we incorporate an independent grasp score network ScoreNet in the optimization process as shown in Fig. 4 (b). Trained with the noisy object point-cloud and its corresponding grasp pose sampled from grasp labels, the ScoreNet predicts the grasp score $\\hat{S}$ . We sample both the good and bad quality grasps so the ScoreNet can give a low score when the failed grasp appears in the optimization. To suppress the decline in grasp score, the target of score optimization $J_{s}$ and the overall optimization target $J$ are formulated as:", + "bbox": [ + 75, + 493, + 468, + 643 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nJ _ {s} = T - \\min (T, \\hat {S}) \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 648, + 468, + 666 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nJ = J _ {c} + \\beta * J _ {s} + \\gamma * \\Delta t \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 679, + 468, + 695 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $T$ is the max grasp score, $\\beta$ is the weight of $J_{s}$ and $\\Delta t$ constrains the offset of position $t$ during optimization. During inference, we adopt an Adam optimizer to optimize the grasp pose iteratively by minimizing $J$ in a gradient descent strategy.", + "bbox": [ + 75, + 702, + 468, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 791, + 209, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 816, + 294, + 832 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Benchmark We conduct all the simulation experiments on the large-scale GraspNet-1billion benchmark [10]. The benchmark includes 190 cluttered scenes, in which 100 scenes for training, 90 for testing. The testing set is divided", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a6f203a872ee741fcaf0e40b5dfd4b5f52d7d00cf8bf4ad71d02c68815ec3769.jpg", + "image_caption": [ + "(a) Contact optimization" + ], + "image_footnote": [], + "bbox": [ + 526, + 94, + 870, + 308 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/30dfdef9513e5b802efa1489da34d042ed663d7a86c4a0e9ea03ae862774f43d.jpg", + "image_caption": [ + "(b) Score optimization", + "Figure 4. The pipeline of contact optimization and score optimization with ContactNet and ScoreNet." + ], + "image_footnote": [], + "bbox": [ + 540, + 333, + 849, + 481 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "into seen, similar and novel set based on the objects in the scenes. Each scene includes 256 RGB-D images captured from different views with RealSense/Kinect cameras. Most previous methods on the benchmark utilize single-view data for training and evaluation. However, the depth images captured from a single view suffer from significant noise. For novel object grasping, as the model has no prior about the objects, it is unable to infer the shape of occluded parts based solely on partial point-clouds, thereby hindering the grasp detection for novel objects. To bypass the interference caused by incomplete point clouds and enhance the capability of grasp detection on novel objects, we reconstruct the depth images from multiple views in a single scene into Truncated Signed Distance Function (TSDF) via KinectFusion [25]. The TSDFs are used for training and evaluation.", + "bbox": [ + 496, + 551, + 890, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metric We follow the metric used in GraspNet-1billion benchmark [10], in which the average precision of top- $k$ ranked grasps in a scene is considered. In the original metric, $k$ is set to 50, and a maximum of 10 grasps per-object are used for evaluation. Given that most scenes contain around 9 objects, this metric can overlook the grasping accuracy of some objects. Therefore, we employ an object-balanced metric for reconstructed scenes, where $k = N_{\\text{object}} * 5$ .", + "bbox": [ + 496, + 779, + 892, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18106", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/df1474e65f2b9b92b5786ef57bb6d65947fc089c19ac8d0a6a909cbe4fc9cd46.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelSeenSimilarNovel
AP\\(AP_{0.8}\\)\\(AP_{0.4}\\)AP\\(AP_{0.8}\\)\\(AP_{0.4}\\)AP\\(AP_{0.8}\\)\\(AP_{0.4}\\)
Baseline66.0975.5760.6364.8274.1061.5130.6137.6117.06
+ PCR66.6775.6761.8465.5574.4261.4735.5843.4919.70
w/o \\(R^A\\)65.6374.5961.1965.3874.2962.1032.3839.0418.88
w/o \\(R^C\\)65.2775.0558.6262.2571.6557.1132.3438.5819.27
w/o \\(R^S\\)63.5172.3760.2664.3173.0061.0635.5742.6020.45
", + "bbox": [ + 183, + 88, + 787, + 223 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$N_{\\text{object}}$ is the number of objects of the scene and a maximum of 5 grasps per-object are used for evaluation. This provides a more balanced consideration of the grasping accuracy across different objects compared to the original evaluation metrics. $\\mathbf{AP}_{\\mu}$ is employed as the metric, which represents the average Precision@m for $m$ ranges from 1 to $m$ with friction $\\mu$ . $\\mathbf{AP}$ is calculated by the average of $\\mathbf{AP}_{\\mu}$ , where $\\mu$ ranges from 0.2 to 1.2 with the interval $\\Delta \\mu = 0.2$ . Model implementations We first introduce a strong baseline model based on the scale balanced 6-DoF grasp detection network [19] and GSNet [35]. Following [35], we employ a sparse UNet based on Minkowski Engine [7] as the backbone and calculate the \"graspness\" for each point to generate grasp candidates. For local feature extraction, the multi-scale cylinder grouping proposed in [19] is employed to improve the performance of the baseline model. For the PCR during training, $\\theta$ and $\\mu$ which control the contact distance are set to 0.02 and 0.005 separately and the weight $\\phi$ for the regularization is set to 0.1. For the C-SJO, we set the hyper-parameter $\\alpha = 0.2$ , $\\beta = 0.01$ and $\\gamma = 5$ .", + "bbox": [ + 76, + 273, + 472, + 578 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Results on Physical Constraint Regularization", + "text_level": 1, + "bbox": [ + 76, + 585, + 465, + 603 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We show the results of PCR in Table 1. Compared to the baseline model, integrating the PCR delivers an improvement of $4.64\\%$ on the novel set, demonstrating its effectiveness for objects with diverse shapes and structures. Besides, the PCR also improves by $0.58\\%$ in seen set and $0.73\\%$ in similar set. This illustrates that for seen and similar objects in the training set, introducing physical prior can help the grasp detection network to fully utilize the grasp labels compared to fitting them directly. We also ablate the influence of different physical constraints and notice that the incomplete constraint conditions can attribute to the performance decline on seen and similar set. Only with partial constraints during training, not all the grasp poses which meet the constraints are correct, thus disturbing the learning from grasping label.", + "bbox": [ + 75, + 609, + 470, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Results on Contact-Score Joint Optimization", + "text_level": 1, + "bbox": [ + 76, + 845, + 455, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We employ the C-SJO in our 6-DoF grasp detection network trained with PCR and conduct ablations. The results", + "bbox": [ + 76, + 869, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/671cb5e04808572ea50d0a87c3e72f2a6a034de1833c745b791e9b1304bec377.jpg", + "table_caption": [ + "Table 1. Results of PCR on scenes captured by RealSense." + ], + "table_footnote": [], + "table_body": "
ModelSeenSimilarNovel
No Refine66.6765.5535.58
+ C-S (Predicted Mask)66.6165.3736.67
+ C-S (GT Mask)66.6965.5036.61
Original C-S66.4865.2936.27
w/o C66.6565.6435.86
w/o S66.5465.3736.23
", + "bbox": [ + 513, + 271, + 880, + 404 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Results of C-SJO on scenes captured by RealSense.", + "bbox": [ + 514, + 415, + 875, + 429 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "are shown in Table 2. We incorporate a recent proposed 3D segmentation method [38] to get object point-clouds from the clutter for optimization. With the ground-truth mask and segmented mask, C-SJO can improve by $1.03\\%$ and $1.00\\%$ on novel set separately. The C-SJO has almost no impact on the seen and similar set, primarily because the grasp poses predicted for seen and similar objects are sufficiently stable, making it challenging to find a better grasp than the original one. However, for the novel set, many unstable grasps exist that can potentially be optimized. We also conduct ablations about different designs of the C-SJO. Without the proposed projection contact map, only employing the original contact map (Original C-S) drops on seen, similar and novel set due to the singularity poses during optimization. Optimizing solely based on grasp score $(w / oC)$ overlooks the contact map prior, resulting in very limited improvement on the novel set. Using only the contact map for optimization $(w / oS)$ yields a $0.38\\%$ drop on the novel set, but the absence of grasp scores restricts the search space, making it challenging to locate the locally optimal grasp during the optimization process. The process of grasp pose optimization is shown in Fig. 5, where the optimal grasp poses are reached in the original grasps' neighbor $SE(3)$ space. For object (1)-(4), the optimization process refines the inaccurate initial grasp poses. For objects (5) and (6), the optimization process facilitates collision avoidance.", + "bbox": [ + 496, + 445, + 890, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Comparison with State-of-the-art", + "text_level": 1, + "bbox": [ + 500, + 847, + 794, + 863 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To make a fair comparison with the state-of-the-art methods of the GraspNet-1billion benchmark, we re-implement", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18107", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f943775637a3f63af03245e381fb94e1dcd4097a45233eac758ce49fcd7a553f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelSeenSimilarNovel
AP\\(AP_{0.8}\\)\\(AP_{0.4}\\)AP\\(AP_{0.8}\\)\\(AP_{0.4}\\)AP\\(AP_{0.8}\\)\\(AP_{0.4}\\)
GraspNet-baseline [10]28.1030.5326.5623.8726.9222.517.488.434.93
Scale-balanced Grasp [19]46.0551.0244.2737.7644.2734.7517.0921.0410.20
GSNet [35]60.4771.0552.0658.5569.3853.0828.0636.1914.09
Ours Baseline66.0975.5760.6364.8274.1061.5130.6137.6117.06
Ours66.6175.6761.5265.3774.3561.2836.6745.0820.90
", + "bbox": [ + 125, + 88, + 846, + 223 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5323dc25d95ebfc5c9c75e0c884d2b1cd27309bde9cf921971cccc02b44c8146.jpg", + "image_caption": [ + "Figure 5. Visualization of the process of C-SJO." + ], + "image_footnote": [], + "bbox": [ + 140, + 262, + 831, + 417 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "three representative 6-DoF grasp detection methods [10, 19, 35] with our multi-view reconstructed scene as input. As shown in Table 3, our method performs better than previous 6-DoF grasp detection methods on all test sets and achieves $36.67\\%$ on the novel set, which demonstrates the generalization capability of our method.", + "bbox": [ + 75, + 460, + 468, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Comparison with object augmentation", + "text_level": 1, + "bbox": [ + 76, + 560, + 408, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We give a comparison between our prior knowledge based method and the object augmentation which employed by previous methods to enhance the generalization capability. To migrate the procedural or learning-based object augmentation method on the GraspNet-1billion benchmark, since the similar objects and the training objects have similar distributions, we consider the similar set in the benchmark as objects obtained through the augmentation. We merge the two sets for joint training. In this way, we can conveniently validate the performance of data augmentation on this benchmark without the need to generate additional objects and annotate grasps. As shown in Table 4, introducing object augmentation can slightly improve the performance on seen and novel sets, benefiting from the richer object distribution. However, our method performs better on the novel set, improving by $4.37\\%$ compared to the object augmentation method, without introducing any additional data. Simultaneously, our method also exhibits an increase of $0.33\\%$ on novel set when paired with object augmentation, demonstrating that our method and object augmentation can be used together.", + "bbox": [ + 75, + 583, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6e4ac0502389fd58ba5375a3a7d31a89182027e9996097576357f341d74f84e1.jpg", + "table_caption": [ + "Table 3. Comparison with the state-of-the-art methods on RealSense scenes of GraspNet-billion benchmark." + ], + "table_footnote": [], + "table_body": "
ModelSeenSimilarNovel
Baseline66.0964.8230.61
Augmentation68.12-32.30
Ours66.6165.3736.67
Ours + Augmentation68.31-37.00
", + "bbox": [ + 519, + 458, + 874, + 555 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Comparison with object augmentation.", + "bbox": [ + 552, + 565, + 836, + 579 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6. Result Visualization", + "text_level": 1, + "bbox": [ + 500, + 594, + 689, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In figure 6, we visualized the results generated by the baseline method, object augmentation and our approach on the GraspNet-1billion benchmark. The gripper poses in red represent successful grasps, while those shown in purple and blue correspond to collision and bad grasps. By incorporating domain prior knowledge, our method can generate grasp poses that conform to the physical relationship between the gripper and the object. For objects (3)-(5), our method predicts some failed grasps, primarily due to poor grasp sampling locations or inaccurate perception of the object's geometry.", + "bbox": [ + 496, + 619, + 890, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.7. Real-world Evaluation", + "text_level": 1, + "bbox": [ + 500, + 800, + 709, + 814 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The effectiveness of the proposed method in real-world is validated by real robot in this section. As shown in Figure 7 (a), the robotic grasping system is built on a 6-DoF UR-10 robot arm and a RealSense D435i depth camera is employed for scene perception. The total of 20 objects used for grasp", + "bbox": [ + 496, + 824, + 890, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18108", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/866ec070b7c5e5f297b12d4dd96c75b2ce4b6cec7998958a1d105e384f2d11a9.jpg", + "image_caption": [ + "Figure 6. Visualization of the predicted grasp poses from baseline, object augmentation and the methods proposed in this paper." + ], + "image_footnote": [], + "bbox": [ + 140, + 92, + 836, + 417 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/937e7614cedf02197c3d3b380630379f10f8f4cdc9ab06cde4a742f13a403aaa.jpg", + "image_caption": [ + "Figure 7. (a) Robotic grasping system and (b) objects for real-world grasping." + ], + "image_footnote": [], + "bbox": [ + 89, + 458, + 259, + 575 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/457b9240a339d102dfc7fe2b1a327bd3e57fd19197a21ef7091c94e79d6ae3c2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 261, + 458, + 455, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ing (Fig. 7 (b)) are composed of two parts: 13 3D-printed objects from Dex-Net [20] and 7 objects chosen from the YCB dataset [4]. All of the objects are significantly different from the objects used for training in GraspNet-1billion. Before executing the proposed grasp detection method, we quickly reconstruct the scene using an arm-mounted depth camera based on the KinectFusion [25].", + "bbox": [ + 75, + 627, + 468, + 733 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a20f178ce436acbf0f9cb01dee167248e57ec7fa3ee283e1bb3cafc7320fd7f0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelIsolatedCluttered
SR (%)SR (%)SCR (%)
Baseline58.33 (35/60)48.21 (43/89)86.00 (43/50)
Ours68.33 (41/60)64.86 (48/74)96.00 (48/50)
", + "bbox": [ + 84, + 744, + 460, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Results of the real-world grasping experiments.", + "bbox": [ + 104, + 818, + 439, + 833 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We compare our model to the baseline in two settings: isolated object grasping and cluttered object grasping. For isolated object grasping, each object is placed in three different poses and Success Rate (SR) is used as the metric.", + "bbox": [ + 76, + 839, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For cluttered object grasping, we compose 5 objects into a scene and make the robot remove them all with a maximum number of operations at 10. SR and Scene Completion Rate (SCR) are employed as the metrics. As illustrated in Table 5, our model outperforms the baseline in both settings and achieves better SR and SCR, demonstrating its superiority for novel objects.", + "bbox": [ + 498, + 460, + 890, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 579, + 617, + 594 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we work for generalizing 6-DoF grasp detection with domain prior knowledge of robotic grasping. The physical constraint regularization based on physical rules is proposed to enable the generalization of objects with largely varied shapes and structures. To refine the unstable results predicted by network in cluttered scenarios, we specially design a contact-score joint optimization with the contact map prior, in which a projection contact map is utilized. Extensive experiments on both the benchmark and the real-world robot demonstrate the effectiveness of our method for novel objects.", + "bbox": [ + 496, + 604, + 890, + 771 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 500, + 784, + 651, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is partly supported by the National Key R&D Program of China (2022ZD0161902), the National Natural Science Foundation of China (62022011), the Research Program of State Key Laboratory of Software Development Environment (SKLSDE-2023ZX-14), and the Fundamental Research Funds for the Central Universities.", + "bbox": [ + 496, + 809, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18109", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Konstantinos Bousmalis, Alex Irpan, Paul Wohlhart, Yunfei Bai, Matthew Kelcey, Mrinal Kalakrishnan, Laura Downs, Julian Ibarz, Peter Pastor, Kurt Konolige, et al. Using simulation and domain adaptation to improve efficiency of deep robotic grasping. In IEEE International Conference on Robotics and Automation, 2018. 2, 3", + "[2] Samarth Brahmbhatt, Cusuh Ham, Charles C. Kemp, and James Hays. Contactdb: Analyzing and predicting grasp contact via thermal imaging. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2", + "[3] Michel Breyer, Jen Jen Chung, Lionel Ott, Roland Siegwart, and Juan I. Nieto. Volumetric grasping network: Real-time 6 DOF grasp detection in clutter. In Conference on Robot Learning, 2020. 2", + "[4] Berk Çallı, Arjun Singh, James Bruce, Aaron Walsman, Kurt Konolige, Siddhartha S. Srinivasa, Pieter Abbeel, and Aaron M. Dollar. Yale-cmu-berkeley dataset for robotic manipulation research. International Journal of Robotics Research, 36(3):261-268, 2017. 8", + "[5] I-Ming Chen and Joel W. Burdick. Finding antipodal point grasps on irregularly shaped objects. IEEE Transactions on Robotics and Automation, 9(4):507-512, 1993. 3", + "[6] I-Ming Chen and Joel W. Burdick. Finding antipodal point grasps on irregularly shaped objects. IEEE Transactions on Robotics and Automation, 9(4):507-512, 1993. 1", + "[7] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 6", + "[8] Qiyu Dai, Yan Zhu, Yiran Geng, Ciyu Ruan, Jiazhao Zhang, and He Wang. Graspnerf: Multiview-based 6-dof grasp detection for transparent and specular objects using generalizable nerf. In IEEE International Conference on Robotics and Automation, 2023. 2, 3", + "[9] Amaury Depierre, Emmanuel Dellandrea, and Liming Chen. Jacquard: A large scale dataset for robotic grasp detection. In IEEE/RSJ International Conference on Intelligent Robots and Systems, 2018. 1, 2", + "[10] Haoshu Fang, Chenxi Wang, Minghao Gou, and Cewu Lu. Graspnet-1billion: A large-scale benchmark for general object grasping. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2, 5, 7", + "[11] Carlo Ferrari and John F. Canny. Planning optimal grasps. In IEEE International Conference on Robotics and Automation, 1992. 3", + "[12] Hanwen Jiang, Shaowei Liu, Jiashun Wang, and Xiaolong Wang. Hand-object contact consistency reasoning for human grasps generation. In IEEE/CVF International Conference on Computer Vision, 2021. 4", + "[13] Junnan Jiang, Xiaohui Xiao, Fei Chen, and Miao Li. Learning grasp ability enhancement through deep shape generation. In International Conference on Intelligent Robotics and Applications, 2022. 2, 3", + "[14] Zhenyu Jiang, Yifeng Zhu, Maxwell Svetlik, Kuan Fang, and Yuke Zhu. Synergies between affordance and geometry: 6-" + ], + "bbox": [ + 78, + 114, + 470, + 902 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "dof grasp detection via implicit representations. In Robotics: Science and Systems, 2021. 2, 3", + "[15] George Em Karniadakis, Ioannis G Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, and Liu Yang. Physics-informed machine learning. Nature Reviews Physics, 3(6):422-440, 2021. 2", + "[16] Hongzhuo Liang, Xiaojian Ma, Shuang Li, Michael Gorner, Song Tang, Bin Fang, Fuchun Sun, and Jianwei Zhang. Pointnetgpd: Detecting grasp configurations from point sets. In IEEE International Conference on Robotics and Automation, 2019. 2", + "[17] Min Liu, Zherong Pan, Kai Xu, Kanishka Ganguly, and Dinesh Manocha. Deep differentiable grasp planner for high-dof grippers. In Robotics: Science and Systems, 2020. 3", + "[18] Tengyu Liu, Zeyu Liu, Ziyuan Jiao, Yixin Zhu, and Song-Chun Zhu. Synthesizing diverse and physically stable grasps with arbitrary hand structures using differentiable force closure estimator. IEEE Robotics and Automation Letters, 7(1): 470-477, 2022. 3", + "[19] Haoxiang Ma and Di Huang. Towards scale balanced 6-dof grasp detection in cluttered scenes. In Conference on Robot Learning, 2022. 2, 3, 6, 7", + "[20] Jeffrey Mahler, Jacky Liang, Sherdil Niyaz, Michael Laskey, Richard Doan, Xinyu Liu, Juan Aparicio Ojea, and Ken Goldberg. Dex-net 2.0: Deep learning to plan robust grasps with synthetic point clouds and analytic grasp metrics. In Robotics: Science and Systems, 2017. 1, 8", + "[21] Andrew T. Miller and Peter K. Allen. Graspit! A versatile simulator for robotic grasping. IEEE Robotics & Automation Magazine, 11(4):110-122, 2004. 3", + "[22] Douglas Morrison, Juxi Leitner, and Peter Corke. Closing the loop for robotic grasping: A real-time, generative grasp synthesis approach. In Robotics: Science and Systems, 2018. 2", + "[23] Douglas Morrison, Peter Corke, and Jürgen Leitner. Egad! an evolved grasping analysis dataset for diversity and reproducibility in robotic manipulation. IEEE Robotics and Automation Letters, 5(3):4368-4375, 2020. 2, 3", + "[24] Arsalan Mousavian, Clemens Eppner, and Dieter Fox. 6-dof grapnet: Variational grasp generation for object manipulation. In IEEE/CVF International Conference on Computer Vision, 2019. 2", + "[25] Richard A. Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J. Davison, Pushmeet Kohli, Jamie Shotton, Steve Hodges, and Andrew W. Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. In IEEE International Symposium on Mixed and Augmented Reality, 2011. 5, 8", + "[26] Van-Duc Nguyen. Constructing force-closure grasps. International Journal of Robotics Research, 7(3):3-16, 1988. 2, 3", + "[27] Nancy S. Pollard. Closure and quality equivalence for efficient synthesis of grasps from examples. International Journal of Robotics Research, 23(6):595-613, 2004. 1", + "[28] Ran Qin, Haoxiang Ma, Boyang Gao, and Di Huang. RGB-D grasp detection via depth guided learning with cross-modal attention. In IEEE International Conference on Robotics and Automation, 2023. 2" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18110", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Yuzhe Qin, Rui Chen, Hao Zhu, Meng Song, Jing Xu, and Hao Su. S4G: amodal single-view single-shot SE(3) grasp detection in cluttered scenes. In Conference on Robot Learning, 2019. 2", + "[30] Maximo A. Roa and Raul Suárez. Computation of independent contact regions for grasping 3-d objects. IEEE Transactions on Robotics, 25(4):839-850, 2009. 1", + "[31] Alberto Rodriguez, Matthew T. Mason, and Steve Ferry. From caging to grasping. International Journal of Robotics Research, 31(7):886-900, 2012. 3", + "[32] Andreas ten Pas, Marcus Gualtieri, Kate Saenko, and Robert Platt Jr. Grasp pose detection in point clouds. International Journal of Robotics Research, 36(13-14):1455-1473, 2017. 1, 2", + "[33] Josh Tobin, Lukas Biewald, Rocky Duan, Marcin Andrychowicz, Ankur Handa, Vikash Kumar, Bob McGrew, Alex Ray, Jonas Schneider, Peter Welinder, et al. Domain randomization and generative models for robotic grasping. In IEEE/RSJ International Conference on Intelligent Robots and Systems, 2018. 2, 3", + "[34] Laura Von Rueden, Sebastian Mayer, Katharina Beckh, Bogdan Georgiev, Sven Giesselbach, Raoul Heese, Birgit Kirsch, Julius Pfrommer, Annika Pick, Rajkumar Ramamurthy, et al. Informed machine learning-a taxonomy and survey of integrating prior knowledge into learning systems. IEEE Transactions on Knowledge and Data Engineering, 35(1):614-633, 2021. 2", + "[35] Chenxi Wang, Haoshu Fang, Minghao Gou, Hongjie Fang, Jin Gao, and Cewu Lu. Graspness discovery in clutters for fast and accurate grasp detection. In IEEE/CVF International Conference on Computer Vision, 2021. 2, 3, 6, 7", + "[36] David Wang, David Tseng, Pusong Li, Yiding Jiang, Menglong Guo, Michael Danielczuk, Jeffrey Mahler, Jeffrey Ichnowski, and Ken Goldberg. Adversarial grasp objects. In IEEE International Conference on Automation Science and Engineering, 2019. 3", + "[37] Yinzhen Xu, Weikang Wan, Jialiang Zhang, Haoran Liu, Zikang Shan, Hao Shen, Ruicheng Wang, Haoran Geng, Yi-jia Weng, Jiayi Chen, Tengyu Liu, Li Yi, and He Wang. Unidexgrasp: Universal robotic dexterous grasping via learning diverse proposal generation and goal-conditioned policy. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 4", + "[38] Yunhan Yang, Xiaoyang Wu, Tong He, Hengshuang Zhao, and Xihui Liu. Sam3d: Segment anything in 3d scenes, 2023. 6" + ], + "bbox": [ + 78, + 90, + 468, + 737 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18111", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Generalizing 6-DoF Grasp Detection via Domain Prior Knowledge/65c67c8c-a659-4259-9daf-ad24fd3765f3_model.json b/2024/Generalizing 6-DoF Grasp Detection via Domain Prior Knowledge/65c67c8c-a659-4259-9daf-ad24fd3765f3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..75c6e42924a556c4219ef4d7ea3c7fce63eba739 --- /dev/null +++ b/2024/Generalizing 6-DoF Grasp Detection via Domain Prior Knowledge/65c67c8c-a659-4259-9daf-ad24fd3765f3_model.json @@ -0,0 +1,1991 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.15, + 0.131, + 0.823, + 0.154 + ], + "angle": 0, + "content": "Generalizing 6-DoF Grasp Detection via Domain Prior Knowledge" + }, + { + "type": "text", + "bbox": [ + 0.238, + 0.18, + 0.74, + 0.2 + ], + "angle": 0, + "content": "Haoxiang Ma\\(^{1,2}\\) Modi Shi\\(^{1,2}\\) Boyang Gao\\(^{3,4}\\) Di Huang\\(^{1,2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.2, + 0.873, + 0.27 + ], + "angle": 0, + "content": "\\(^{1}\\)State Key Laboratory of Software Development Environment, Beihang University, Beijing, China \n\\(^{2}\\)School of Computer Science and Engineering, Beihang University, Beijing, China \n\\(^{3}\\)School of Computer Science and Technology, Harbin Institute of Technology, Harbin, China \n\\(^{4}\\)Geometry Robotics" + }, + { + "type": "text", + "bbox": [ + 0.153, + 0.272, + 0.805, + 0.288 + ], + "angle": 0, + "content": "{mahaoxiang822, modishi, dhuang}@buaa.edu.cn, {boyang.gao}@geometryrobot.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.322, + 0.314, + 0.339 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.354, + 0.474, + 0.702 + ], + "angle": 0, + "content": "We focus on the generalization ability of the 6-DoF grasp detection method in this paper. While learning-based grasp detection methods can predict grasp poses for unseen objects using the grasp distribution learned from the training set, they often exhibit a significant performance drop when encountering objects with diverse shapes and structures. To enhance the grasp detection methods' generalization ability, we incorporate domain prior knowledge of robotic grasping, enabling better adaptation to objects with significant shape and structure differences. More specifically, we employ the physical constraint regularization during the training phase to guide the model towards predicting grasps that comply with the physical rule on grasping. For the unstable grasp poses predicted on novel objects, we design a contact-score joint optimization using the projection contact map to refine these poses in cluttered scenarios. Extensive experiments conducted on the GraspNet1billion benchmark demonstrate a substantial performance gain on the novel object set and the real-world grasping experiments also demonstrate the effectiveness of our generalizing 6-DoF grasp detection method. Code is available at https://github.com/mahaoxiang822/Generalizing-Grasp." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.731, + 0.21, + 0.746 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.756, + 0.47, + 0.879 + ], + "angle": 0, + "content": "Given an object, robotic grasp detection aims to find suitable and sufficient gripper configurations for various manipulation tasks. Traditional methods [6, 27, 30] establish hand-crafted criteria to evaluate grasp samples according to 3D models of objects. Despite providing precise interaction between objects and grippers, they assume that object models are available in advance and also suffer a slow running speed, which limits their popularization. As the develop-" + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.323, + 0.865, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.48, + 0.865, + 0.541 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.687, + 0.544, + 0.706, + 0.556 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.54, + 0.557, + 0.857, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.688, + 0.615, + 0.706, + 0.627 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.639, + 0.895, + 0.696 + ], + "angle": 0, + "content": "Figure 1. Illustration of (a) the performance and distribution gap between objects similar to training samples (similar objects) and those with largely varied shapes and structures (novel objects) and (b) the limitations of the object augmentation methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.895, + 0.885 + ], + "angle": 0, + "content": "ment of deep learning, data-driven methods [9, 20, 32] have been widely studied. These methods predict grasps without the need for pre-prepared 3D models and demonstrate the capability to handle unseen objects. However, they often struggle when encountering objects whose shapes and structures significantly deviate from those in the training set, making them difficult to adapt to more diverse applications. As illustrated in Fig. 1 (a), learning-based grasp detection methods are basically able to transfer to objects similar to those appearing in training, but they do not effectively generalize to novel objects with their distribution greatly changed." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "To facilitate the generalization of learning-based grasp" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.222, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18102" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.38 + ], + "angle": 0, + "content": "detection methods towards a variety of unseen objects, previous attempts usually apply data augmentation techniques to expand the distribution of the training set. More objects are typically produced by randomly concatenating a number of 3D primitives [1, 33] or by directly synthesizing through generative networks [13, 23]. They do enrich the diversity of objects for training, which leads to a gain in overall performance. However, since such augmentation is conducted only according to original data, it contributes more to similar objects whose distribution is largely overlapped and offers little help to new objects whose distribution is not well covered. As depicted in Fig. 1 (b), on the one hand, to the objects that are far from the center of training samples in the feature space, totally wrong grasps could be delivered. On the other hand, for the ones that are near the center of training samples in the feature space, although reasonable grasps may be obtained, the predictions tend to be unstable due to the distribution gap. Both cases make those methods problematic to generalize to unseen targets." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.383, + 0.473, + 0.702 + ], + "angle": 0, + "content": "To address the dilemma above, inspired by informed machine learning [15, 34], we introduce domain prior knowledge of robotic grasping. Compared to data augmentation, domain prior knowledge does not depend on the distribution of training data, allowing for easy adaptation to objects with significant shape and structure differences. Firstly, to enable the grasp detection network to generalize to novel objects, we incorporate certain physical rules on grasping, such as force conditions [26] and contact positions, which are typically used in the traditional analytical grasping methods. The physical rules provide valuable clues in terms of grasp stability regardless of object-specific properties, thus motivating us to integrate such rule-based priors into the grasp detection network. Secondly, to deal with the unstable results predicted by the network, we introduce another type of knowledge about the interaction of the gripper and object. The contact map [2] is able to denote the regions correlated with grasping as well as the kinematics of the gripper. By employing a neural network to encode the preferred object regions for optimal grasps, the learning-based contact map prior can be utilized to refine insufficiently accurate grasps." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.705, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Concretely, in this paper, we propose a generalized 6-DoF grasp detection framework with domain prior knowledge, which consists of two components: Physical Constraint Regularization (PCR) and Contact-Score Joint Optimization (C-SJO). In PCR, we integrate the physical prior into the network as the regularization, thereby constraining the correlation between gripper poses and object models. To enable the back-propagation of PCR, we employ an end-to-end 6-DoF grasp detection network for grasp prediction and utilize the Signed Distance Field (SDF) to encode the object model, facilitating the differentiable computation of physical constraints. Compared to fitting grasp annotations directly, PCR guides the network to predict grasps follow" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.408 + ], + "angle": 0, + "content": "ing object-independent physical rules, thereby enhancing the generalization capability for novel objects. To refine the unstable prediction for novel objects, we introduce the C-SJO at test time based on the contact map prior. The contact map represents the contact region on the object's surface by calculating the distance between the gripper model and the objects. A contact map prediction network is introduced to encode the contact pattern of good grasps by learning from grasp labels. By aligning the contact map of the current grasp and the prediction from the network, the grasp pose can be optimized within its neighborhood. However, as for 6-DoF grasp detection in clutter, using the euclidean distance for contact map calculation cannot adequately address inaccurate contact positions. Therefore, we introduce a projection contact map to solve this problem. Additionally, due to the noise from the depth sensor and the occlusion in cluttered scenarios, relying solely on the contact prior can lead to singular results. To mitigate this, we use score optimization to constrain the searching space of contact optimization, in which a grasp score network is utilized." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.409, + 0.872, + 0.424 + ], + "angle": 0, + "content": "The contribution of this paper can be summarized as:" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.426, + 0.892, + 0.47 + ], + "angle": 0, + "content": "- We propose a domain prior knowledge informed 6-DoF grasp detection framework to enhance the generalization ability for novel objects." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.471, + 0.892, + 0.516 + ], + "angle": 0, + "content": "- We design the physical constraint regularization to represent the physical prior for 6-DoF grasping and integrate it into the network in a differential strategy." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.517, + 0.892, + 0.562 + ], + "angle": 0, + "content": "- We employ a contact-score joint optimization with the projection contact map, which is suitable for refining the inaccurate prediction in cluttered scenarios." + }, + { + "type": "list", + "bbox": [ + 0.501, + 0.426, + 0.892, + 0.562 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.574, + 0.637, + 0.589 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.895, + 0.903 + ], + "angle": 0, + "content": "6-DoF Grasp Detection To generate diverse and feasible grasping in cluttered scenes, 6-DoF grasp detection has been advanced recently. Compared to planar grasp detection where the grasp space is limited [9, 22, 28], 6-DoF grasp detection method can predict grasping in \\(SE(3)\\) space, thereby supporting more complex downstream task. [32] pioneers a sampling and evaluation framework for 6-DoF grasp detection. Using point-cloud input, they heuristically sample grasp candidates and employ a Convolutional Neural Network (CNN) for scoring them. [16] build upon the framework, utilizing a neural network for point-cloud data to achieve improved grasp sample evaluation. More recently, a series of studies [3, 8, 10, 14, 19, 24, 29, 35] propose end-to-end strategy for grasp prediction. [24] employs a variational auto-encoder for grasp generation given the object point-clouds. [29] introduces a single-shot grasp proposal network for efficient grasp detection based on single view point-clouds, and [3] employs a Truncated Signed Distance Function (TSDF) to map multiple frames and predict grasps in voxel space. [10] provides a large benchmark" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18103" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.289 + ], + "angle": 0, + "content": "with densely annotated grasp labels and designs a baseline method for dense grasp prediction. Subsequently, [35] defines graspsness to represent grasp probability in searching space and [19] focuses on the scale imbalance problem for 6-DoF grasp detection. With the development of neural representations, [14] utilizes an occupancy network to learn a shared representation between 3D reconstruction and grasp detection. [8] introduces a generalizable neural radiance field for the grasp detection of transparent and specular objects. Despite these advancements in 6-DoF grasp detection, the performance for novel objects, particularly those with varied shapes and structures, remains sub-optimal. In this paper, we investigate this shortcoming." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.293, + 0.473, + 0.642 + ], + "angle": 0, + "content": "Generalization on Grasp Detection Several strategies are proposed to enhance the generalization capability of grasp detection, which mainly focuses on enriching the distribution of training objects. [1] procedurally generates objects by attaching rectangular prisms at random locations and orientations. To enhance the shape diversity, [33] constructs a set of object primitives by decomposing everyday objects and generates diverse objects by randomly sampling these primitives and combing them. In addition to heuristically generating random objects, [23] provides the Evolved Grasping Analysis Dataset (EGAD), which includes objects of varying shape complexity and grasp difficulty generated by the 3D compositional pattern producing networks. [13] employs an AutoEncoder-Critic network to interpolate new shapes from two objects for augmentation. Besides augmenting current shapes to improve the performance of the learning-based grasp detection method, [36] explores the task of generating adversarial objects that are difficult to grasp. Different from previous augmentation methods for generalized grasp detection, we introduce the grasp domain knowledge to our framework explicitly. Therefore, our method doesn't rely on the distribution of training objects, resulting in an improvement for out-of-distribution objects." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Usage of domain knowledge in Grasping Domain prior knowledge has been leveraged in various grasping applications. Early analytical methods [11, 21, 31] make use of the physical prior knowledge to examine the force dynamics between the gripper and the object. More recently, several methods incorporate some grasp prior knowledge into the hand-object grasp synthesis. [18] introduces a differentiable force closure algorithm designed to optimize hand configurations, which facilitates the fast generation of physically stable grasps. [17] designs a generalized Q1 metric that serves as a loss for a grasp planner, producing precise multifinger grasping for single object with watertight model or rendered depth images. While the aforementioned methods employ domain prior knowledge to assist grasp synthesis and optimization, they rely on the accurate object shapes and design complex calculations for prior integration. This makes it difficult to apply them in the 6-DoF grasp detection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.138 + ], + "angle": 0, + "content": "where the generation of diverse grasps in complex scenarios is required and the object geometry from the depth sensor is inaccurate." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.15, + 0.593, + 0.165 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.174, + 0.807, + 0.191 + ], + "angle": 0, + "content": "3.1. Physical Constraint Regularization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.198, + 0.895, + 0.35 + ], + "angle": 0, + "content": "In terms of physical prior knowledge, the force interaction between the object and the gripper is used to analyze grasp stability, among which the force closure [26] is widely adopted. A force closure grasp can resist any external wrenches with the contact force if the force direction lies in the friction cone. In scenarios where the friction coefficient of the object surface is unknown and a two-finger gripper is utilized, we employ the antipodal rule [5] as a simplification. Given two contact points \\((u_{1}, u_{2})\\) lying on the object surface, the antipodal rule is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.358, + 0.892, + 0.378 + ], + "angle": 0, + "content": "\\[\n\\left[ p \\left(u _ {1}\\right) - p \\left(u _ {2}\\right) \\right] \\cdot t \\left(u _ {1}\\right) = 0 \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.603, + 0.388, + 0.892, + 0.407 + ], + "angle": 0, + "content": "\\[\n[ p (u _ {2}) - p (u _ {1}) ] \\cdot t (u _ {2}) = 0 \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.629, + 0.415, + 0.892, + 0.432 + ], + "angle": 0, + "content": "\\[\nn \\left(u _ {1}\\right) + n \\left(u _ {2}\\right) = 0 \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.437, + 0.892, + 0.541 + ], + "angle": 0, + "content": "where \\( p(u) \\) represents the contact position, \\( t(u) \\) represents the unit tangent vector and \\( n(u) \\) represents the unit outward normal vector. If the contact points \\( (u_1, u_2) \\) between the gripper and the object comply with the antipodal rule, there is a higher probability of a successful grasp. Based on which, we introduce the PCR to constrain the output of the network with the rule." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.542, + 0.895, + 0.587 + ], + "angle": 0, + "content": "With the predicted grasp pose \\( g = [t, R, w] \\) and the gripper kinematic model \\( K \\), the gripper contacts \\( (c_1, c_2) \\) can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.622, + 0.597, + 0.892, + 0.615 + ], + "angle": 0, + "content": "\\[\n\\left(c _ {1}, c _ {2}\\right) = K (t, R, w) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.623, + 0.892, + 0.669 + ], + "angle": 0, + "content": "To enforce the gripper contacts to comply with the antipodal constraint, we propose an antipodal regularization term, which is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.677, + 0.892, + 0.709 + ], + "angle": 0, + "content": "\\[\nR ^ {A} \\left(c _ {1}, c _ {2}\\right) = 1 - 0. 5 * \\left(\\cos \\left(\\overrightarrow {c _ {1} c _ {2}}, n \\left(c _ {2}\\right)\\right) + \\cos \\left(\\overrightarrow {c _ {2} c _ {1}}, n \\left(c _ {1}\\right)\\right)\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.71, + 0.895, + 0.875 + ], + "angle": 0, + "content": "where \\(\\overrightarrow{c_1c_2}\\) is the vector connecting the gripper's contact points \\(c_{1}\\) and \\(c_{2}\\), \\(n(c_{1})\\) and \\(n(c_{2})\\) are the normal vectors at the respective contact points and \\(\\cos\\) represents the cosine similarity. Nevertheless, the predicted gripper contacts \\(c_{1}, c_{2}\\) may not lie precisely on the object surface. To overcome this, we also introduce constraints on the distance between the gripper contacts and the object's surface. These constraints help to avoid collision and ensure that the contact points are sufficiently close to the object's surface. The collision constraint \\(R^C\\) and surface constraint \\(R^S\\) are formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.884, + 0.892, + 0.903 + ], + "angle": 0, + "content": "\\[\nR ^ {C} \\left(c _ {1}, c _ {2}\\right) = M a x (0, \\theta - d \\left(c _ {1}\\right)) + M a x (0, \\theta - d \\left(c _ {2}\\right)) \\tag {6}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "18104" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.143, + 0.09, + 0.832, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.279, + 0.895, + 0.322 + ], + "angle": 0, + "content": "Figure 2. The pipeline of differential physical constraint integration. With the fusion point-cloud, the differentiable 6-DoF grasp network predicts the grasp configurations. The position of contacts \\( c_{1}, c_{2} \\) are calculated from grasp configurations by the gripper model and the regularization \\( R \\) is computed from the object SDF for back-propagation." + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.342, + 0.47, + 0.361 + ], + "angle": 0, + "content": "\\[\nR ^ {S} \\left(c _ {1}, c _ {2}\\right) = M a x (0, d \\left(c _ {1}\\right) - \\mu) + M a x (0, d \\left(c _ {2}\\right) - \\mu) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.369, + 0.471, + 0.474 + ], + "angle": 0, + "content": "where the contact distance \\( d(c) \\) is constrained between \\( [\\theta, \\mu] \\). Besides, we use the grasp score predicted by the grasp detection network to weight the regularization of different grasps, thereby ensuring the coherence between the grasp score and the physical constraints. With the grasp score of \\( i \\)th grasp \\( s_i \\), the overall physical constraint \\( R_i \\) can be represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.481, + 0.47, + 0.537 + ], + "angle": 0, + "content": "\\[\nR _ {i} = \\frac {s _ {i} * \\left(R _ {i} ^ {A} + R _ {i} ^ {C} + R _ {i} ^ {S}\\right)}{\\frac {1}{M} \\sum_ {j} ^ {M} s _ {j}} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.543, + 0.47, + 0.573 + ], + "angle": 0, + "content": "where \\(M\\) is the number of seed points. The overall loss \\(L\\) can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.203, + 0.583, + 0.469, + 0.6 + ], + "angle": 0, + "content": "\\[\nL = L _ {\\text {g r a s p}} + \\phi * R \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.608, + 0.47, + 0.64 + ], + "angle": 0, + "content": "For more details of the grasp configuration loss \\(L_{\\text{grasp}}\\), please refer to the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.647, + 0.45, + 0.663 + ], + "angle": 0, + "content": "3.2. Differential Physical Constraint Integration" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.67, + 0.473, + 0.903 + ], + "angle": 0, + "content": "To integrate the PCR into the grasp detection network, the sub-gradients of normal vector \\(\\frac{\\partial n(c)}{\\partial c}\\) and surface distance \\(\\frac{\\partial d(c)}{\\partial c}\\) with respect to the contact point \\(c\\) should be computed. Besides, the generation of grasp poses should be differential. As a result, we introduce an end-to-end 6-DoF grasp network and the differentiable Signed Distance Function (SDF) for the calculation of PCR. As shown in Fig. 2, the configurations of the gripper are regressed from the input fusion point-cloud in an end-to-end manner. With the configurations predicted by the grasp detection network, the contact point \\(c\\) can be calculated by the gripper kinematics model. Utilizing SDF of the object, we can query the surface distance of the contact \\(c_{1}\\) directly by \\(d(c_{1}) = SDF(c_{1})\\) and the normal vector of \\(c_{1}\\) can be calculated by \\(n(c_{1}) = \\nabla SDF(c_{1})\\). In practice, obtaining an" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.333, + 0.895, + 0.438 + ], + "angle": 0, + "content": "object's SDF directly is not feasible. Consequently, we employ a 3D grid to represent the object's SDF, where each grid point records the corresponding SDF value. During training, the SDF value at any arbitrary position is computed using tri-linear interpolation, which allows for the differentiable computation of the gradients required by the physical constraints." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.448, + 0.796, + 0.466 + ], + "angle": 0, + "content": "3.3. Contact-Score Joint Optimization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.471, + 0.895, + 0.608 + ], + "angle": 0, + "content": "Due to the significant discrepancy in shape and structure between novel and training objects, the grasp prediction from the network can be unstable, which leads to failures. To solve this problem, inspired by dexterous grasp synthesis [12, 37], we introduce the contact map prior which encodes the preference contact region on the object to refine inaccurate grasps. Given the object point-cloud \\( X_{o} \\) and gripper contacts \\( c_{1}, c_{2} \\), we define the contact map of each point \\( p_{i} \\in X_{o} \\) for a 6-DoF grasp as:" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.618, + 0.892, + 0.643 + ], + "angle": 0, + "content": "\\[\nD _ {i} = \\min _ {j} \\| p _ {i} - c _ {j} \\|, c _ {j} \\in \\left\\{c _ {1}, c _ {2} \\right\\} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.651, + 0.895, + 0.818 + ], + "angle": 0, + "content": "Compared with optimization for a single object with accurate geometry in grasp synthesis, for the 6-DoF grasp detection task in cluttered scenes, there exist two issues that make the test-time optimization impractical and cause singular results: (1) the contact points don't always lie on the object surface; (2) the object point-clouds in real-world scenarios are inaccurate and incomplete. For the first issue, we propose the projection contact distance \\( PD \\) illustrated in Fig. 3 (a) as a supplementary of original contact distance \\( D \\), which calculates the projection distance from the object point-cloud to the connecting line of two contacts." + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.829, + 0.892, + 0.846 + ], + "angle": 0, + "content": "\\[\nP D _ {i} = D _ {i} * \\sin \\left(\\theta_ {i}\\right) \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We visualize the original contact map and the proposed projection contact map in Fig. 3 (b). Rather than computing the contact distance directly, the projection contact" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "18105" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.089, + 0.295, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.186, + 0.228, + 0.204, + 0.239 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.089, + 0.451, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.364, + 0.228, + 0.382, + 0.239 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.252, + 0.47, + 0.28 + ], + "angle": 0, + "content": "Figure 3. (a) Calculation of the contact map and (b) comparison of the projection contact map and the original version." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.296, + 0.469, + 0.414 + ], + "angle": 0, + "content": "map can highlight the accurate contact region on the object point-cloud for coarse grasp poses whose contact points may not lie on the surface. For contact map optimization, as shown in Fig. 4 (a), we employ a ContactNet for contact map prediction, which encodes the contact prior. Given the object point-cloud \\( X_{o} \\), the gripper point-cloud \\( X_{g} \\) and the predicted grasp pose \\( g \\), the ContactNet can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.416, + 0.469, + 0.435 + ], + "angle": 0, + "content": "\\[\n\\hat {D}, \\hat {P D} = \\operatorname {C o n t a c t N e t} \\left(g X _ {g}, X _ {o}\\right) \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.44, + 0.469, + 0.456 + ], + "angle": 0, + "content": "and thus the contact optimization target \\( J_{c} \\) is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.159, + 0.469, + 0.469, + 0.487 + ], + "angle": 0, + "content": "\\[\nJ _ {c} = | D - \\hat {D} | + \\alpha * | P D - \\hat {P D} | \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.494, + 0.47, + 0.644 + ], + "angle": 0, + "content": "For the second issue, we incorporate an independent grasp score network ScoreNet in the optimization process as shown in Fig. 4 (b). Trained with the noisy object point-cloud and its corresponding grasp pose sampled from grasp labels, the ScoreNet predicts the grasp score \\(\\hat{S}\\). We sample both the good and bad quality grasps so the ScoreNet can give a low score when the failed grasp appears in the optimization. To suppress the decline in grasp score, the target of score optimization \\(J_{s}\\) and the overall optimization target \\(J\\) are formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.201, + 0.649, + 0.469, + 0.667 + ], + "angle": 0, + "content": "\\[\nJ _ {s} = T - \\min (T, \\hat {S}) \\tag {14}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.68, + 0.469, + 0.696 + ], + "angle": 0, + "content": "\\[\nJ = J _ {c} + \\beta * J _ {s} + \\gamma * \\Delta t \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.703, + 0.47, + 0.779 + ], + "angle": 0, + "content": "where \\(T\\) is the max grasp score, \\(\\beta\\) is the weight of \\(J_{s}\\) and \\(\\Delta t\\) constrains the offset of position \\(t\\) during optimization. During inference, we adopt an Adam optimizer to optimize the grasp pose iteratively by minimizing \\(J\\) in a gradient descent strategy." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.792, + 0.21, + 0.809 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.295, + 0.833 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Benchmark We conduct all the simulation experiments on the large-scale GraspNet-1billion benchmark [10]. The benchmark includes 190 cluttered scenes, in which 100 scenes for training, 90 for testing. The testing set is divided" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.095, + 0.871, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.631, + 0.315, + 0.762, + 0.327 + ], + "angle": 0, + "content": "(a) Contact optimization" + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.334, + 0.85, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.637, + 0.487, + 0.757, + 0.499 + ], + "angle": 0, + "content": "(b) Score optimization" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.511, + 0.892, + 0.538 + ], + "angle": 0, + "content": "Figure 4. The pipeline of contact optimization and score optimization with ContactNet and ScoreNet." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.779 + ], + "angle": 0, + "content": "into seen, similar and novel set based on the objects in the scenes. Each scene includes 256 RGB-D images captured from different views with RealSense/Kinect cameras. Most previous methods on the benchmark utilize single-view data for training and evaluation. However, the depth images captured from a single view suffer from significant noise. For novel object grasping, as the model has no prior about the objects, it is unable to infer the shape of occluded parts based solely on partial point-clouds, thereby hindering the grasp detection for novel objects. To bypass the interference caused by incomplete point clouds and enhance the capability of grasp detection on novel objects, we reconstruct the depth images from multiple views in a single scene into Truncated Signed Distance Function (TSDF) via KinectFusion [25]. The TSDFs are used for training and evaluation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Metric We follow the metric used in GraspNet-1billion benchmark [10], in which the average precision of top-\\(k\\) ranked grasps in a scene is considered. In the original metric, \\(k\\) is set to 50, and a maximum of 10 grasps per-object are used for evaluation. Given that most scenes contain around 9 objects, this metric can overlook the grasping accuracy of some objects. Therefore, we employ an object-balanced metric for reconstructed scenes, where \\(k = N_{\\text{object}} * 5\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18106" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.184, + 0.089, + 0.788, + 0.224 + ], + "angle": 0, + "content": "
ModelSeenSimilarNovel
AP\\(AP_{0.8}\\)\\(AP_{0.4}\\)AP\\(AP_{0.8}\\)\\(AP_{0.4}\\)AP\\(AP_{0.8}\\)\\(AP_{0.4}\\)
Baseline66.0975.5760.6364.8274.1061.5130.6137.6117.06
+ PCR66.6775.6761.8465.5574.4261.4735.5843.4919.70
w/o \\(R^A\\)65.6374.5961.1965.3874.2962.1032.3839.0418.88
w/o \\(R^C\\)65.2775.0558.6262.2571.6557.1132.3438.5819.27
w/o \\(R^S\\)63.5172.3760.2664.3173.0061.0635.5742.6020.45
" + }, + { + "type": "table_caption", + "bbox": [ + 0.31, + 0.233, + 0.658, + 0.247 + ], + "angle": 0, + "content": "Table 1. Results of PCR on scenes captured by RealSense." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.274, + 0.473, + 0.579 + ], + "angle": 0, + "content": "\\(N_{\\text{object}}\\) is the number of objects of the scene and a maximum of 5 grasps per-object are used for evaluation. This provides a more balanced consideration of the grasping accuracy across different objects compared to the original evaluation metrics. \\(\\mathbf{AP}_{\\mu}\\) is employed as the metric, which represents the average Precision@m for \\(m\\) ranges from 1 to \\(m\\) with friction \\(\\mu\\). \\(\\mathbf{AP}\\) is calculated by the average of \\(\\mathbf{AP}_{\\mu}\\), where \\(\\mu\\) ranges from 0.2 to 1.2 with the interval \\(\\Delta \\mu = 0.2\\). Model implementations We first introduce a strong baseline model based on the scale balanced 6-DoF grasp detection network [19] and GSNet [35]. Following [35], we employ a sparse UNet based on Minkowski Engine [7] as the backbone and calculate the \"graspness\" for each point to generate grasp candidates. For local feature extraction, the multi-scale cylinder grouping proposed in [19] is employed to improve the performance of the baseline model. For the PCR during training, \\(\\theta\\) and \\(\\mu\\) which control the contact distance are set to 0.02 and 0.005 separately and the weight \\(\\phi\\) for the regularization is set to 0.1. For the C-SJO, we set the hyper-parameter \\(\\alpha = 0.2\\), \\(\\beta = 0.01\\) and \\(\\gamma = 5\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.587, + 0.467, + 0.604 + ], + "angle": 0, + "content": "4.2. Results on Physical Constraint Regularization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.61, + 0.471, + 0.839 + ], + "angle": 0, + "content": "We show the results of PCR in Table 1. Compared to the baseline model, integrating the PCR delivers an improvement of \\(4.64\\%\\) on the novel set, demonstrating its effectiveness for objects with diverse shapes and structures. Besides, the PCR also improves by \\(0.58\\%\\) in seen set and \\(0.73\\%\\) in similar set. This illustrates that for seen and similar objects in the training set, introducing physical prior can help the grasp detection network to fully utilize the grasp labels compared to fitting them directly. We also ablate the influence of different physical constraints and notice that the incomplete constraint conditions can attribute to the performance decline on seen and similar set. Only with partial constraints during training, not all the grasp poses which meet the constraints are correct, thus disturbing the learning from grasping label." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.847, + 0.456, + 0.865 + ], + "angle": 0, + "content": "4.3. Results on Contact-Score Joint Optimization" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.901 + ], + "angle": 0, + "content": "We employ the C-SJO in our 6-DoF grasp detection network trained with PCR and conduct ablations. The results" + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.272, + 0.882, + 0.405 + ], + "angle": 0, + "content": "
ModelSeenSimilarNovel
No Refine66.6765.5535.58
+ C-S (Predicted Mask)66.6165.3736.67
+ C-S (GT Mask)66.6965.5036.61
Original C-S66.4865.2936.27
w/o C66.6565.6435.86
w/o S66.5465.3736.23
" + }, + { + "type": "table_caption", + "bbox": [ + 0.515, + 0.416, + 0.877, + 0.43 + ], + "angle": 0, + "content": "Table 2. Results of C-SJO on scenes captured by RealSense." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.446, + 0.892, + 0.839 + ], + "angle": 0, + "content": "are shown in Table 2. We incorporate a recent proposed 3D segmentation method [38] to get object point-clouds from the clutter for optimization. With the ground-truth mask and segmented mask, C-SJO can improve by \\(1.03\\%\\) and \\(1.00\\%\\) on novel set separately. The C-SJO has almost no impact on the seen and similar set, primarily because the grasp poses predicted for seen and similar objects are sufficiently stable, making it challenging to find a better grasp than the original one. However, for the novel set, many unstable grasps exist that can potentially be optimized. We also conduct ablations about different designs of the C-SJO. Without the proposed projection contact map, only employing the original contact map (Original C-S) drops on seen, similar and novel set due to the singularity poses during optimization. Optimizing solely based on grasp score \\((w / oC)\\) overlooks the contact map prior, resulting in very limited improvement on the novel set. Using only the contact map for optimization \\((w / oS)\\) yields a \\(0.38\\%\\) drop on the novel set, but the absence of grasp scores restricts the search space, making it challenging to locate the locally optimal grasp during the optimization process. The process of grasp pose optimization is shown in Fig. 5, where the optimal grasp poses are reached in the original grasps' neighbor \\(SE(3)\\) space. For object (1)-(4), the optimization process refines the inaccurate initial grasp poses. For objects (5) and (6), the optimization process facilitates collision avoidance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.795, + 0.864 + ], + "angle": 0, + "content": "4.4. Comparison with State-of-the-art" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To make a fair comparison with the state-of-the-art methods of the GraspNet-1billion benchmark, we re-implement" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18107" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.126, + 0.089, + 0.847, + 0.224 + ], + "angle": 0, + "content": "
ModelSeenSimilarNovel
AP\\(AP_{0.8}\\)\\(AP_{0.4}\\)AP\\(AP_{0.8}\\)\\(AP_{0.4}\\)AP\\(AP_{0.8}\\)\\(AP_{0.4}\\)
GraspNet-baseline [10]28.1030.5326.5623.8726.9222.517.488.434.93
Scale-balanced Grasp [19]46.0551.0244.2737.7644.2734.7517.0921.0410.20
GSNet [35]60.4771.0552.0658.5569.3853.0828.0636.1914.09
Ours Baseline66.0975.5760.6364.8274.1061.5130.6137.6117.06
Ours66.6175.6761.5265.3774.3561.2836.6745.0820.90
" + }, + { + "type": "table_caption", + "bbox": [ + 0.164, + 0.233, + 0.804, + 0.248 + ], + "angle": 0, + "content": "Table 3. Comparison with the state-of-the-art methods on RealSense scenes of GraspNet-billion benchmark." + }, + { + "type": "image", + "bbox": [ + 0.141, + 0.263, + 0.832, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.34, + 0.427, + 0.629, + 0.442 + ], + "angle": 0, + "content": "Figure 5. Visualization of the process of C-SJO." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.461, + 0.47, + 0.553 + ], + "angle": 0, + "content": "three representative 6-DoF grasp detection methods [10, 19, 35] with our multi-view reconstructed scene as input. As shown in Table 3, our method performs better than previous 6-DoF grasp detection methods on all test sets and achieves \\(36.67\\%\\) on the novel set, which demonstrates the generalization capability of our method." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.561, + 0.409, + 0.577 + ], + "angle": 0, + "content": "4.5. Comparison with object augmentation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We give a comparison between our prior knowledge based method and the object augmentation which employed by previous methods to enhance the generalization capability. To migrate the procedural or learning-based object augmentation method on the GraspNet-1billion benchmark, since the similar objects and the training objects have similar distributions, we consider the similar set in the benchmark as objects obtained through the augmentation. We merge the two sets for joint training. In this way, we can conveniently validate the performance of data augmentation on this benchmark without the need to generate additional objects and annotate grasps. As shown in Table 4, introducing object augmentation can slightly improve the performance on seen and novel sets, benefiting from the richer object distribution. However, our method performs better on the novel set, improving by \\(4.37\\%\\) compared to the object augmentation method, without introducing any additional data. Simultaneously, our method also exhibits an increase of \\(0.33\\%\\) on novel set when paired with object augmentation, demonstrating that our method and object augmentation can be used together." + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.459, + 0.875, + 0.556 + ], + "angle": 0, + "content": "
ModelSeenSimilarNovel
Baseline66.0964.8230.61
Augmentation68.12-32.30
Ours66.6165.3736.67
Ours + Augmentation68.31-37.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.553, + 0.566, + 0.838, + 0.58 + ], + "angle": 0, + "content": "Table 4. Comparison with object augmentation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.595, + 0.691, + 0.61 + ], + "angle": 0, + "content": "4.6. Result Visualization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.62, + 0.892, + 0.788 + ], + "angle": 0, + "content": "In figure 6, we visualized the results generated by the baseline method, object augmentation and our approach on the GraspNet-1billion benchmark. The gripper poses in red represent successful grasps, while those shown in purple and blue correspond to collision and bad grasps. By incorporating domain prior knowledge, our method can generate grasp poses that conform to the physical relationship between the gripper and the object. For objects (3)-(5), our method predicts some failed grasps, primarily due to poor grasp sampling locations or inaccurate perception of the object's geometry." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.801, + 0.71, + 0.815 + ], + "angle": 0, + "content": "4.7. Real-world Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.903 + ], + "angle": 0, + "content": "The effectiveness of the proposed method in real-world is validated by real robot in this section. As shown in Figure 7 (a), the robotic grasping system is built on a 6-DoF UR-10 robot arm and a RealSense D435i depth camera is employed for scene perception. The total of 20 objects used for grasp" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18108" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.141, + 0.093, + 0.837, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.427, + 0.861, + 0.443 + ], + "angle": 0, + "content": "Figure 6. Visualization of the predicted grasp poses from baseline, object augmentation and the methods proposed in this paper." + }, + { + "type": "image", + "bbox": [ + 0.09, + 0.459, + 0.26, + 0.576 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.263, + 0.459, + 0.457, + 0.576 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.586, + 0.471, + 0.616 + ], + "angle": 0, + "content": "Figure 7. (a) Robotic grasping system and (b) objects for real-world grasping." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.628, + 0.47, + 0.734 + ], + "angle": 0, + "content": "ing (Fig. 7 (b)) are composed of two parts: 13 3D-printed objects from Dex-Net [20] and 7 objects chosen from the YCB dataset [4]. All of the objects are significantly different from the objects used for training in GraspNet-1billion. Before executing the proposed grasp detection method, we quickly reconstruct the scene using an arm-mounted depth camera based on the KinectFusion [25]." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.746, + 0.462, + 0.809 + ], + "angle": 0, + "content": "
ModelIsolatedCluttered
SR (%)SR (%)SCR (%)
Baseline58.33 (35/60)48.21 (43/89)86.00 (43/50)
Ours68.33 (41/60)64.86 (48/74)96.00 (48/50)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.105, + 0.819, + 0.441, + 0.834 + ], + "angle": 0, + "content": "Table 5. Results of the real-world grasping experiments." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.471, + 0.901 + ], + "angle": 0, + "content": "We compare our model to the baseline in two settings: isolated object grasping and cluttered object grasping. For isolated object grasping, each object is placed in three different poses and Success Rate (SR) is used as the metric." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.461, + 0.892, + 0.568 + ], + "angle": 0, + "content": "For cluttered object grasping, we compose 5 objects into a scene and make the robot remove them all with a maximum number of operations at 10. SR and Scene Completion Rate (SCR) are employed as the metrics. As illustrated in Table 5, our model outperforms the baseline in both settings and achieves better SR and SCR, demonstrating its superiority for novel objects." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.58, + 0.619, + 0.595 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.605, + 0.892, + 0.772 + ], + "angle": 0, + "content": "In this paper, we work for generalizing 6-DoF grasp detection with domain prior knowledge of robotic grasping. The physical constraint regularization based on physical rules is proposed to enable the generalization of objects with largely varied shapes and structures. To refine the unstable results predicted by network in cluttered scenarios, we specially design a contact-score joint optimization with the contact map prior, in which a projection contact map is utilized. Extensive experiments on both the benchmark and the real-world robot demonstrate the effectiveness of our method for novel objects." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.785, + 0.652, + 0.802 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.9 + ], + "angle": 0, + "content": "This work is partly supported by the National Key R&D Program of China (2022ZD0161902), the National Natural Science Foundation of China (62022011), the Research Program of State Key Laboratory of Software Development Environment (SKLSDE-2023ZX-14), and the Fundamental Research Funds for the Central Universities." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18109" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.198 + ], + "angle": 0, + "content": "[1] Konstantinos Bousmalis, Alex Irpan, Paul Wohlhart, Yunfei Bai, Matthew Kelcey, Mrinal Kalakrishnan, Laura Downs, Julian Ibarz, Peter Pastor, Kurt Konolige, et al. Using simulation and domain adaptation to improve efficiency of deep robotic grasping. In IEEE International Conference on Robotics and Automation, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.472, + 0.257 + ], + "angle": 0, + "content": "[2] Samarth Brahmbhatt, Cusuh Ham, Charles C. Kemp, and James Hays. Contactdb: Analyzing and predicting grasp contact via thermal imaging. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.258, + 0.471, + 0.313 + ], + "angle": 0, + "content": "[3] Michel Breyer, Jen Jen Chung, Lionel Ott, Roland Siegwart, and Juan I. Nieto. Volumetric grasping network: Real-time 6 DOF grasp detection in clutter. In Conference on Robot Learning, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.315, + 0.471, + 0.383 + ], + "angle": 0, + "content": "[4] Berk Çallı, Arjun Singh, James Bruce, Aaron Walsman, Kurt Konolige, Siddhartha S. Srinivasa, Pieter Abbeel, and Aaron M. Dollar. Yale-cmu-berkeley dataset for robotic manipulation research. International Journal of Robotics Research, 36(3):261-268, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.385, + 0.471, + 0.427 + ], + "angle": 0, + "content": "[5] I-Ming Chen and Joel W. Burdick. Finding antipodal point grasps on irregularly shaped objects. IEEE Transactions on Robotics and Automation, 9(4):507-512, 1993. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.429, + 0.471, + 0.47 + ], + "angle": 0, + "content": "[6] I-Ming Chen and Joel W. Burdick. Finding antipodal point grasps on irregularly shaped objects. IEEE Transactions on Robotics and Automation, 9(4):507-512, 1993. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.472, + 0.471, + 0.528 + ], + "angle": 0, + "content": "[7] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.53, + 0.471, + 0.599 + ], + "angle": 0, + "content": "[8] Qiyu Dai, Yan Zhu, Yiran Geng, Ciyu Ruan, Jiazhao Zhang, and He Wang. Graspnerf: Multiview-based 6-dof grasp detection for transparent and specular objects using generalizable nerf. In IEEE International Conference on Robotics and Automation, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.601, + 0.471, + 0.657 + ], + "angle": 0, + "content": "[9] Amaury Depierre, Emmanuel Dellandrea, and Liming Chen. Jacquard: A large scale dataset for robotic grasp detection. In IEEE/RSJ International Conference on Intelligent Robots and Systems, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.658, + 0.47, + 0.714 + ], + "angle": 0, + "content": "[10] Haoshu Fang, Chenxi Wang, Minghao Gou, and Cewu Lu. Graspnet-1billion: A large-scale benchmark for general object grasping. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.715, + 0.47, + 0.757 + ], + "angle": 0, + "content": "[11] Carlo Ferrari and John F. Canny. Planning optimal grasps. In IEEE International Conference on Robotics and Automation, 1992. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.758, + 0.47, + 0.814 + ], + "angle": 0, + "content": "[12] Hanwen Jiang, Shaowei Liu, Jiashun Wang, and Xiaolong Wang. Hand-object contact consistency reasoning for human grasps generation. In IEEE/CVF International Conference on Computer Vision, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[13] Junnan Jiang, Xiaohui Xiao, Fei Chen, and Miao Li. Learning grasp ability enhancement through deep shape generation. In International Conference on Intelligent Robotics and Applications, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.903 + ], + "angle": 0, + "content": "[14] Zhenyu Jiang, Yifeng Zhu, Maxwell Svetlik, Kuan Fang, and Yuke Zhu. Synergies between affordance and geometry: 6-" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "dof grasp detection via implicit representations. In Robotics: Science and Systems, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[15] George Em Karniadakis, Ioannis G Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, and Liu Yang. Physics-informed machine learning. Nature Reviews Physics, 3(6):422-440, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.177, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[16] Hongzhuo Liang, Xiaojian Ma, Shuang Li, Michael Gorner, Song Tang, Bin Fang, Fuchun Sun, and Jianwei Zhang. Pointnetgpd: Detecting grasp configurations from point sets. In IEEE International Conference on Robotics and Automation, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.247, + 0.892, + 0.288 + ], + "angle": 0, + "content": "[17] Min Liu, Zherong Pan, Kai Xu, Kanishka Ganguly, and Dinesh Manocha. Deep differentiable grasp planner for high-dof grippers. In Robotics: Science and Systems, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.289, + 0.892, + 0.356 + ], + "angle": 0, + "content": "[18] Tengyu Liu, Zeyu Liu, Ziyuan Jiao, Yixin Zhu, and Song-Chun Zhu. Synthesizing diverse and physically stable grasps with arbitrary hand structures using differentiable force closure estimator. IEEE Robotics and Automation Letters, 7(1): 470-477, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.358, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[19] Haoxiang Ma and Di Huang. Towards scale balanced 6-dof grasp detection in cluttered scenes. In Conference on Robot Learning, 2022. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[20] Jeffrey Mahler, Jacky Liang, Sherdil Niyaz, Michael Laskey, Richard Doan, Xinyu Liu, Juan Aparicio Ojea, and Ken Goldberg. Dex-net 2.0: Deep learning to plan robust grasps with synthetic point clouds and analytic grasp metrics. In Robotics: Science and Systems, 2017. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.469, + 0.892, + 0.511 + ], + "angle": 0, + "content": "[21] Andrew T. Miller and Peter K. Allen. Graspit! A versatile simulator for robotic grasping. IEEE Robotics & Automation Magazine, 11(4):110-122, 2004. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.565 + ], + "angle": 0, + "content": "[22] Douglas Morrison, Juxi Leitner, and Peter Corke. Closing the loop for robotic grasping: A real-time, generative grasp synthesis approach. In Robotics: Science and Systems, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.567, + 0.892, + 0.622 + ], + "angle": 0, + "content": "[23] Douglas Morrison, Peter Corke, and Jürgen Leitner. Egad! an evolved grasping analysis dataset for diversity and reproducibility in robotic manipulation. IEEE Robotics and Automation Letters, 5(3):4368-4375, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.622, + 0.892, + 0.677 + ], + "angle": 0, + "content": "[24] Arsalan Mousavian, Clemens Eppner, and Dieter Fox. 6-dof grapnet: Variational grasp generation for object manipulation. In IEEE/CVF International Conference on Computer Vision, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[25] Richard A. Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J. Davison, Pushmeet Kohli, Jamie Shotton, Steve Hodges, and Andrew W. Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. In IEEE International Symposium on Mixed and Augmented Reality, 2011. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.762, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[26] Van-Duc Nguyen. Constructing force-closure grasps. International Journal of Robotics Research, 7(3):3-16, 1988. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.803, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[27] Nancy S. Pollard. Closure and quality equivalence for efficient synthesis of grasps from examples. International Journal of Robotics Research, 23(6):595-613, 2004. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[28] Ran Qin, Haoxiang Ma, Boyang Gao, and Di Huang. RGB-D grasp detection via depth guided learning with cross-modal attention. In IEEE International Conference on Robotics and Automation, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18110" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[29] Yuzhe Qin, Rui Chen, Hao Zhu, Meng Song, Jing Xu, and Hao Su. S4G: amodal single-view single-shot SE(3) grasp detection in cluttered scenes. In Conference on Robot Learning, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.468, + 0.19 + ], + "angle": 0, + "content": "[30] Maximo A. Roa and Raul Suárez. Computation of independent contact regions for grasping 3-d objects. IEEE Transactions on Robotics, 25(4):839-850, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.192, + 0.468, + 0.233 + ], + "angle": 0, + "content": "[31] Alberto Rodriguez, Matthew T. Mason, and Steve Ferry. From caging to grasping. International Journal of Robotics Research, 31(7):886-900, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.235, + 0.468, + 0.289 + ], + "angle": 0, + "content": "[32] Andreas ten Pas, Marcus Gualtieri, Kate Saenko, and Robert Platt Jr. Grasp pose detection in point clouds. International Journal of Robotics Research, 36(13-14):1455-1473, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.291, + 0.468, + 0.373 + ], + "angle": 0, + "content": "[33] Josh Tobin, Lukas Biewald, Rocky Duan, Marcin Andrychowicz, Ankur Handa, Vikash Kumar, Bob McGrew, Alex Ray, Jonas Schneider, Peter Welinder, et al. Domain randomization and generative models for robotic grasping. In IEEE/RSJ International Conference on Intelligent Robots and Systems, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.375, + 0.468, + 0.47 + ], + "angle": 0, + "content": "[34] Laura Von Rueden, Sebastian Mayer, Katharina Beckh, Bogdan Georgiev, Sven Giesselbach, Raoul Heese, Birgit Kirsch, Julius Pfrommer, Annika Pick, Rajkumar Ramamurthy, et al. Informed machine learning-a taxonomy and survey of integrating prior knowledge into learning systems. IEEE Transactions on Knowledge and Data Engineering, 35(1):614-633, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.473, + 0.468, + 0.529 + ], + "angle": 0, + "content": "[35] Chenxi Wang, Haoshu Fang, Minghao Gou, Hongjie Fang, Jin Gao, and Cewu Lu. Graspness discovery in clutters for fast and accurate grasp detection. In IEEE/CVF International Conference on Computer Vision, 2021. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.531, + 0.468, + 0.599 + ], + "angle": 0, + "content": "[36] David Wang, David Tseng, Pusong Li, Yiding Jiang, Menglong Guo, Michael Danielczuk, Jeffrey Mahler, Jeffrey Ichnowski, and Ken Goldberg. Adversarial grasp objects. In IEEE International Conference on Automation Science and Engineering, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.601, + 0.468, + 0.697 + ], + "angle": 0, + "content": "[37] Yinzhen Xu, Weikang Wan, Jialiang Zhang, Haoran Liu, Zikang Shan, Hao Shen, Ruicheng Wang, Haoran Geng, Yi-jia Weng, Jiayi Chen, Tengyu Liu, Li Yi, and He Wang. Unidexgrasp: Universal robotic dexterous grasping via learning diverse proposal generation and goal-conditioned policy. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.7, + 0.468, + 0.738 + ], + "angle": 0, + "content": "[38] Yunhan Yang, Xiaoyang Wu, Tong He, Hengshuang Zhao, and Xihui Liu. Sam3d: Segment anything in 3d scenes, 2023. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.738 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "18111" + } + ] +] \ No newline at end of file diff --git a/2024/Generalizing 6-DoF Grasp Detection via Domain Prior Knowledge/65c67c8c-a659-4259-9daf-ad24fd3765f3_origin.pdf b/2024/Generalizing 6-DoF Grasp Detection via Domain Prior Knowledge/65c67c8c-a659-4259-9daf-ad24fd3765f3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..489f73c699e6a5a05dd4ebe1d17144d208d2d96d --- /dev/null +++ b/2024/Generalizing 6-DoF Grasp Detection via Domain Prior Knowledge/65c67c8c-a659-4259-9daf-ad24fd3765f3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:286ab593e35df8467b085ca874462efda5c73b0551b99c59afd2274d6daaae9a +size 8653397 diff --git a/2024/Generalizing 6-DoF Grasp Detection via Domain Prior Knowledge/full.md b/2024/Generalizing 6-DoF Grasp Detection via Domain Prior Knowledge/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ef69419f0e5e7f7aab5351a18e6ef4b1d84bae9e --- /dev/null +++ b/2024/Generalizing 6-DoF Grasp Detection via Domain Prior Knowledge/full.md @@ -0,0 +1,305 @@ +# Generalizing 6-DoF Grasp Detection via Domain Prior Knowledge + +Haoxiang Ma $^{1,2}$ Modi Shi $^{1,2}$ Boyang Gao $^{3,4}$ Di Huang $^{1,2*}$ + +$^{1}$ State Key Laboratory of Software Development Environment, Beihang University, Beijing, China + $^{2}$ School of Computer Science and Engineering, Beihang University, Beijing, China + $^{3}$ School of Computer Science and Technology, Harbin Institute of Technology, Harbin, China + $^{4}$ Geometry Robotics + +{mahaoxiang822, modishi, dhuang}@buaa.edu.cn, {boyang.gao}@geometryrobot.com + +# Abstract + +We focus on the generalization ability of the 6-DoF grasp detection method in this paper. While learning-based grasp detection methods can predict grasp poses for unseen objects using the grasp distribution learned from the training set, they often exhibit a significant performance drop when encountering objects with diverse shapes and structures. To enhance the grasp detection methods' generalization ability, we incorporate domain prior knowledge of robotic grasping, enabling better adaptation to objects with significant shape and structure differences. More specifically, we employ the physical constraint regularization during the training phase to guide the model towards predicting grasps that comply with the physical rule on grasping. For the unstable grasp poses predicted on novel objects, we design a contact-score joint optimization using the projection contact map to refine these poses in cluttered scenarios. Extensive experiments conducted on the GraspNet1billion benchmark demonstrate a substantial performance gain on the novel object set and the real-world grasping experiments also demonstrate the effectiveness of our generalizing 6-DoF grasp detection method. Code is available at https://github.com/mahaoxiang822/Generalizing-Grasp. + +# 1. Introduction + +Given an object, robotic grasp detection aims to find suitable and sufficient gripper configurations for various manipulation tasks. Traditional methods [6, 27, 30] establish hand-crafted criteria to evaluate grasp samples according to 3D models of objects. Despite providing precise interaction between objects and grippers, they assume that object models are available in advance and also suffer a slow running speed, which limits their popularization. As the develop- + +![](images/2b83e1a29032bc4276eb2195c87004d7045fae79e5169984924e3d834b8b9787.jpg) + +![](images/406f56a7c2c9e2ba8a7008c28b895a9562dfae83e899c531e6d42950f5ff542f.jpg) +(a) + +![](images/8967d8b099bbeea45c76cfe0c378003d1b6b48fcdfb700c63895f5f0a426ce40.jpg) +(b) +Figure 1. Illustration of (a) the performance and distribution gap between objects similar to training samples (similar objects) and those with largely varied shapes and structures (novel objects) and (b) the limitations of the object augmentation methods. + +ment of deep learning, data-driven methods [9, 20, 32] have been widely studied. These methods predict grasps without the need for pre-prepared 3D models and demonstrate the capability to handle unseen objects. However, they often struggle when encountering objects whose shapes and structures significantly deviate from those in the training set, making them difficult to adapt to more diverse applications. As illustrated in Fig. 1 (a), learning-based grasp detection methods are basically able to transfer to objects similar to those appearing in training, but they do not effectively generalize to novel objects with their distribution greatly changed. + +To facilitate the generalization of learning-based grasp + +detection methods towards a variety of unseen objects, previous attempts usually apply data augmentation techniques to expand the distribution of the training set. More objects are typically produced by randomly concatenating a number of 3D primitives [1, 33] or by directly synthesizing through generative networks [13, 23]. They do enrich the diversity of objects for training, which leads to a gain in overall performance. However, since such augmentation is conducted only according to original data, it contributes more to similar objects whose distribution is largely overlapped and offers little help to new objects whose distribution is not well covered. As depicted in Fig. 1 (b), on the one hand, to the objects that are far from the center of training samples in the feature space, totally wrong grasps could be delivered. On the other hand, for the ones that are near the center of training samples in the feature space, although reasonable grasps may be obtained, the predictions tend to be unstable due to the distribution gap. Both cases make those methods problematic to generalize to unseen targets. + +To address the dilemma above, inspired by informed machine learning [15, 34], we introduce domain prior knowledge of robotic grasping. Compared to data augmentation, domain prior knowledge does not depend on the distribution of training data, allowing for easy adaptation to objects with significant shape and structure differences. Firstly, to enable the grasp detection network to generalize to novel objects, we incorporate certain physical rules on grasping, such as force conditions [26] and contact positions, which are typically used in the traditional analytical grasping methods. The physical rules provide valuable clues in terms of grasp stability regardless of object-specific properties, thus motivating us to integrate such rule-based priors into the grasp detection network. Secondly, to deal with the unstable results predicted by the network, we introduce another type of knowledge about the interaction of the gripper and object. The contact map [2] is able to denote the regions correlated with grasping as well as the kinematics of the gripper. By employing a neural network to encode the preferred object regions for optimal grasps, the learning-based contact map prior can be utilized to refine insufficiently accurate grasps. + +Concretely, in this paper, we propose a generalized 6-DoF grasp detection framework with domain prior knowledge, which consists of two components: Physical Constraint Regularization (PCR) and Contact-Score Joint Optimization (C-SJO). In PCR, we integrate the physical prior into the network as the regularization, thereby constraining the correlation between gripper poses and object models. To enable the back-propagation of PCR, we employ an end-to-end 6-DoF grasp detection network for grasp prediction and utilize the Signed Distance Field (SDF) to encode the object model, facilitating the differentiable computation of physical constraints. Compared to fitting grasp annotations directly, PCR guides the network to predict grasps follow + +ing object-independent physical rules, thereby enhancing the generalization capability for novel objects. To refine the unstable prediction for novel objects, we introduce the C-SJO at test time based on the contact map prior. The contact map represents the contact region on the object's surface by calculating the distance between the gripper model and the objects. A contact map prediction network is introduced to encode the contact pattern of good grasps by learning from grasp labels. By aligning the contact map of the current grasp and the prediction from the network, the grasp pose can be optimized within its neighborhood. However, as for 6-DoF grasp detection in clutter, using the euclidean distance for contact map calculation cannot adequately address inaccurate contact positions. Therefore, we introduce a projection contact map to solve this problem. Additionally, due to the noise from the depth sensor and the occlusion in cluttered scenarios, relying solely on the contact prior can lead to singular results. To mitigate this, we use score optimization to constrain the searching space of contact optimization, in which a grasp score network is utilized. + +The contribution of this paper can be summarized as: + +- We propose a domain prior knowledge informed 6-DoF grasp detection framework to enhance the generalization ability for novel objects. +- We design the physical constraint regularization to represent the physical prior for 6-DoF grasping and integrate it into the network in a differential strategy. +- We employ a contact-score joint optimization with the projection contact map, which is suitable for refining the inaccurate prediction in cluttered scenarios. + +# 2. Related work + +6-DoF Grasp Detection To generate diverse and feasible grasping in cluttered scenes, 6-DoF grasp detection has been advanced recently. Compared to planar grasp detection where the grasp space is limited [9, 22, 28], 6-DoF grasp detection method can predict grasping in $SE(3)$ space, thereby supporting more complex downstream task. [32] pioneers a sampling and evaluation framework for 6-DoF grasp detection. Using point-cloud input, they heuristically sample grasp candidates and employ a Convolutional Neural Network (CNN) for scoring them. [16] build upon the framework, utilizing a neural network for point-cloud data to achieve improved grasp sample evaluation. More recently, a series of studies [3, 8, 10, 14, 19, 24, 29, 35] propose end-to-end strategy for grasp prediction. [24] employs a variational auto-encoder for grasp generation given the object point-clouds. [29] introduces a single-shot grasp proposal network for efficient grasp detection based on single view point-clouds, and [3] employs a Truncated Signed Distance Function (TSDF) to map multiple frames and predict grasps in voxel space. [10] provides a large benchmark + +with densely annotated grasp labels and designs a baseline method for dense grasp prediction. Subsequently, [35] defines graspsness to represent grasp probability in searching space and [19] focuses on the scale imbalance problem for 6-DoF grasp detection. With the development of neural representations, [14] utilizes an occupancy network to learn a shared representation between 3D reconstruction and grasp detection. [8] introduces a generalizable neural radiance field for the grasp detection of transparent and specular objects. Despite these advancements in 6-DoF grasp detection, the performance for novel objects, particularly those with varied shapes and structures, remains sub-optimal. In this paper, we investigate this shortcoming. + +Generalization on Grasp Detection Several strategies are proposed to enhance the generalization capability of grasp detection, which mainly focuses on enriching the distribution of training objects. [1] procedurally generates objects by attaching rectangular prisms at random locations and orientations. To enhance the shape diversity, [33] constructs a set of object primitives by decomposing everyday objects and generates diverse objects by randomly sampling these primitives and combing them. In addition to heuristically generating random objects, [23] provides the Evolved Grasping Analysis Dataset (EGAD), which includes objects of varying shape complexity and grasp difficulty generated by the 3D compositional pattern producing networks. [13] employs an AutoEncoder-Critic network to interpolate new shapes from two objects for augmentation. Besides augmenting current shapes to improve the performance of the learning-based grasp detection method, [36] explores the task of generating adversarial objects that are difficult to grasp. Different from previous augmentation methods for generalized grasp detection, we introduce the grasp domain knowledge to our framework explicitly. Therefore, our method doesn't rely on the distribution of training objects, resulting in an improvement for out-of-distribution objects. + +Usage of domain knowledge in Grasping Domain prior knowledge has been leveraged in various grasping applications. Early analytical methods [11, 21, 31] make use of the physical prior knowledge to examine the force dynamics between the gripper and the object. More recently, several methods incorporate some grasp prior knowledge into the hand-object grasp synthesis. [18] introduces a differentiable force closure algorithm designed to optimize hand configurations, which facilitates the fast generation of physically stable grasps. [17] designs a generalized Q1 metric that serves as a loss for a grasp planner, producing precise multifinger grasping for single object with watertight model or rendered depth images. While the aforementioned methods employ domain prior knowledge to assist grasp synthesis and optimization, they rely on the accurate object shapes and design complex calculations for prior integration. This makes it difficult to apply them in the 6-DoF grasp detection + +where the generation of diverse grasps in complex scenarios is required and the object geometry from the depth sensor is inaccurate. + +# 3. Method + +# 3.1. Physical Constraint Regularization + +In terms of physical prior knowledge, the force interaction between the object and the gripper is used to analyze grasp stability, among which the force closure [26] is widely adopted. A force closure grasp can resist any external wrenches with the contact force if the force direction lies in the friction cone. In scenarios where the friction coefficient of the object surface is unknown and a two-finger gripper is utilized, we employ the antipodal rule [5] as a simplification. Given two contact points $(u_{1}, u_{2})$ lying on the object surface, the antipodal rule is formulated as: + +$$ +\left[ p \left(u _ {1}\right) - p \left(u _ {2}\right) \right] \cdot t \left(u _ {1}\right) = 0 \tag {1} +$$ + +$$ +[ p (u _ {2}) - p (u _ {1}) ] \cdot t (u _ {2}) = 0 \tag {2} +$$ + +$$ +n \left(u _ {1}\right) + n \left(u _ {2}\right) = 0 \tag {3} +$$ + +where $p(u)$ represents the contact position, $t(u)$ represents the unit tangent vector and $n(u)$ represents the unit outward normal vector. If the contact points $(u_1, u_2)$ between the gripper and the object comply with the antipodal rule, there is a higher probability of a successful grasp. Based on which, we introduce the PCR to constrain the output of the network with the rule. + +With the predicted grasp pose $g = [t, R, w]$ and the gripper kinematic model $K$ , the gripper contacts $(c_1, c_2)$ can be expressed as: + +$$ +\left(c _ {1}, c _ {2}\right) = K (t, R, w) \tag {4} +$$ + +To enforce the gripper contacts to comply with the antipodal constraint, we propose an antipodal regularization term, which is calculated as follows: + +$$ +R ^ {A} \left(c _ {1}, c _ {2}\right) = 1 - 0. 5 * \left(\cos \left(\overrightarrow {c _ {1} c _ {2}}, n \left(c _ {2}\right)\right) + \cos \left(\overrightarrow {c _ {2} c _ {1}}, n \left(c _ {1}\right)\right)\right) \tag {5} +$$ + +where $\overrightarrow{c_1c_2}$ is the vector connecting the gripper's contact points $c_{1}$ and $c_{2}$ , $n(c_{1})$ and $n(c_{2})$ are the normal vectors at the respective contact points and $\cos$ represents the cosine similarity. Nevertheless, the predicted gripper contacts $c_{1}, c_{2}$ may not lie precisely on the object surface. To overcome this, we also introduce constraints on the distance between the gripper contacts and the object's surface. These constraints help to avoid collision and ensure that the contact points are sufficiently close to the object's surface. The collision constraint $R^C$ and surface constraint $R^S$ are formulated as: + +$$ +R ^ {C} \left(c _ {1}, c _ {2}\right) = M a x (0, \theta - d \left(c _ {1}\right)) + M a x (0, \theta - d \left(c _ {2}\right)) \tag {6} +$$ + +![](images/4b2c0601a788be80ddfa67e9424d92418051a31807ecdfd9e81f79906880d71a.jpg) +Figure 2. The pipeline of differential physical constraint integration. With the fusion point-cloud, the differentiable 6-DoF grasp network predicts the grasp configurations. The position of contacts $c_{1}, c_{2}$ are calculated from grasp configurations by the gripper model and the regularization $R$ is computed from the object SDF for back-propagation. + +$$ +R ^ {S} \left(c _ {1}, c _ {2}\right) = M a x (0, d \left(c _ {1}\right) - \mu) + M a x (0, d \left(c _ {2}\right) - \mu) \tag {7} +$$ + +where the contact distance $d(c)$ is constrained between $[\theta, \mu]$ . Besides, we use the grasp score predicted by the grasp detection network to weight the regularization of different grasps, thereby ensuring the coherence between the grasp score and the physical constraints. With the grasp score of $i$ th grasp $s_i$ , the overall physical constraint $R_i$ can be represented as: + +$$ +R _ {i} = \frac {s _ {i} * \left(R _ {i} ^ {A} + R _ {i} ^ {C} + R _ {i} ^ {S}\right)}{\frac {1}{M} \sum_ {j} ^ {M} s _ {j}} \tag {8} +$$ + +where $M$ is the number of seed points. The overall loss $L$ can be formulated as: + +$$ +L = L _ {\text {g r a s p}} + \phi * R \tag {9} +$$ + +For more details of the grasp configuration loss $L_{\text{grasp}}$ , please refer to the supplementary. + +# 3.2. Differential Physical Constraint Integration + +To integrate the PCR into the grasp detection network, the sub-gradients of normal vector $\frac{\partial n(c)}{\partial c}$ and surface distance $\frac{\partial d(c)}{\partial c}$ with respect to the contact point $c$ should be computed. Besides, the generation of grasp poses should be differential. As a result, we introduce an end-to-end 6-DoF grasp network and the differentiable Signed Distance Function (SDF) for the calculation of PCR. As shown in Fig. 2, the configurations of the gripper are regressed from the input fusion point-cloud in an end-to-end manner. With the configurations predicted by the grasp detection network, the contact point $c$ can be calculated by the gripper kinematics model. Utilizing SDF of the object, we can query the surface distance of the contact $c_{1}$ directly by $d(c_{1}) = SDF(c_{1})$ and the normal vector of $c_{1}$ can be calculated by $n(c_{1}) = \nabla SDF(c_{1})$ . In practice, obtaining an + +object's SDF directly is not feasible. Consequently, we employ a 3D grid to represent the object's SDF, where each grid point records the corresponding SDF value. During training, the SDF value at any arbitrary position is computed using tri-linear interpolation, which allows for the differentiable computation of the gradients required by the physical constraints. + +# 3.3. Contact-Score Joint Optimization + +Due to the significant discrepancy in shape and structure between novel and training objects, the grasp prediction from the network can be unstable, which leads to failures. To solve this problem, inspired by dexterous grasp synthesis [12, 37], we introduce the contact map prior which encodes the preference contact region on the object to refine inaccurate grasps. Given the object point-cloud $X_{o}$ and gripper contacts $c_{1}, c_{2}$ , we define the contact map of each point $p_{i} \in X_{o}$ for a 6-DoF grasp as: + +$$ +D _ {i} = \min _ {j} \| p _ {i} - c _ {j} \|, c _ {j} \in \left\{c _ {1}, c _ {2} \right\} \tag {10} +$$ + +Compared with optimization for a single object with accurate geometry in grasp synthesis, for the 6-DoF grasp detection task in cluttered scenes, there exist two issues that make the test-time optimization impractical and cause singular results: (1) the contact points don't always lie on the object surface; (2) the object point-clouds in real-world scenarios are inaccurate and incomplete. For the first issue, we propose the projection contact distance $PD$ illustrated in Fig. 3 (a) as a supplementary of original contact distance $D$ , which calculates the projection distance from the object point-cloud to the connecting line of two contacts. + +$$ +P D _ {i} = D _ {i} * \sin \left(\theta_ {i}\right) \tag {11} +$$ + +We visualize the original contact map and the proposed projection contact map in Fig. 3 (b). Rather than computing the contact distance directly, the projection contact + +![](images/826c4048089a00cbed616167ff9ac253714145b1a46cda9df6bd412c611998ec.jpg) +(a) + +![](images/11383aa2771d4bb74fe7562a794ee768994b10d8d41338d00924358cd0acdc57.jpg) +(b) +Figure 3. (a) Calculation of the contact map and (b) comparison of the projection contact map and the original version. + +map can highlight the accurate contact region on the object point-cloud for coarse grasp poses whose contact points may not lie on the surface. For contact map optimization, as shown in Fig. 4 (a), we employ a ContactNet for contact map prediction, which encodes the contact prior. Given the object point-cloud $X_{o}$ , the gripper point-cloud $X_{g}$ and the predicted grasp pose $g$ , the ContactNet can be formulated as: + +$$ +\hat {D}, \hat {P D} = \operatorname {C o n t a c t N e t} \left(g X _ {g}, X _ {o}\right) \tag {12} +$$ + +and thus the contact optimization target $J_{c}$ is formulated as: + +$$ +J _ {c} = | D - \hat {D} | + \alpha * | P D - \hat {P D} | \tag {13} +$$ + +For the second issue, we incorporate an independent grasp score network ScoreNet in the optimization process as shown in Fig. 4 (b). Trained with the noisy object point-cloud and its corresponding grasp pose sampled from grasp labels, the ScoreNet predicts the grasp score $\hat{S}$ . We sample both the good and bad quality grasps so the ScoreNet can give a low score when the failed grasp appears in the optimization. To suppress the decline in grasp score, the target of score optimization $J_{s}$ and the overall optimization target $J$ are formulated as: + +$$ +J _ {s} = T - \min (T, \hat {S}) \tag {14} +$$ + +$$ +J = J _ {c} + \beta * J _ {s} + \gamma * \Delta t \tag {15} +$$ + +where $T$ is the max grasp score, $\beta$ is the weight of $J_{s}$ and $\Delta t$ constrains the offset of position $t$ during optimization. During inference, we adopt an Adam optimizer to optimize the grasp pose iteratively by minimizing $J$ in a gradient descent strategy. + +# 4. Experiments + +# 4.1. Implementation Details + +Benchmark We conduct all the simulation experiments on the large-scale GraspNet-1billion benchmark [10]. The benchmark includes 190 cluttered scenes, in which 100 scenes for training, 90 for testing. The testing set is divided + +![](images/a6f203a872ee741fcaf0e40b5dfd4b5f52d7d00cf8bf4ad71d02c68815ec3769.jpg) +(a) Contact optimization + +![](images/30dfdef9513e5b802efa1489da34d042ed663d7a86c4a0e9ea03ae862774f43d.jpg) +(b) Score optimization +Figure 4. The pipeline of contact optimization and score optimization with ContactNet and ScoreNet. + +into seen, similar and novel set based on the objects in the scenes. Each scene includes 256 RGB-D images captured from different views with RealSense/Kinect cameras. Most previous methods on the benchmark utilize single-view data for training and evaluation. However, the depth images captured from a single view suffer from significant noise. For novel object grasping, as the model has no prior about the objects, it is unable to infer the shape of occluded parts based solely on partial point-clouds, thereby hindering the grasp detection for novel objects. To bypass the interference caused by incomplete point clouds and enhance the capability of grasp detection on novel objects, we reconstruct the depth images from multiple views in a single scene into Truncated Signed Distance Function (TSDF) via KinectFusion [25]. The TSDFs are used for training and evaluation. + +Metric We follow the metric used in GraspNet-1billion benchmark [10], in which the average precision of top- $k$ ranked grasps in a scene is considered. In the original metric, $k$ is set to 50, and a maximum of 10 grasps per-object are used for evaluation. Given that most scenes contain around 9 objects, this metric can overlook the grasping accuracy of some objects. Therefore, we employ an object-balanced metric for reconstructed scenes, where $k = N_{\text{object}} * 5$ . + +
ModelSeenSimilarNovel
AP\(AP_{0.8}\)\(AP_{0.4}\)AP\(AP_{0.8}\)\(AP_{0.4}\)AP\(AP_{0.8}\)\(AP_{0.4}\)
Baseline66.0975.5760.6364.8274.1061.5130.6137.6117.06
+ PCR66.6775.6761.8465.5574.4261.4735.5843.4919.70
w/o \(R^A\)65.6374.5961.1965.3874.2962.1032.3839.0418.88
w/o \(R^C\)65.2775.0558.6262.2571.6557.1132.3438.5819.27
w/o \(R^S\)63.5172.3760.2664.3173.0061.0635.5742.6020.45
+ +$N_{\text{object}}$ is the number of objects of the scene and a maximum of 5 grasps per-object are used for evaluation. This provides a more balanced consideration of the grasping accuracy across different objects compared to the original evaluation metrics. $\mathbf{AP}_{\mu}$ is employed as the metric, which represents the average Precision@m for $m$ ranges from 1 to $m$ with friction $\mu$ . $\mathbf{AP}$ is calculated by the average of $\mathbf{AP}_{\mu}$ , where $\mu$ ranges from 0.2 to 1.2 with the interval $\Delta \mu = 0.2$ . Model implementations We first introduce a strong baseline model based on the scale balanced 6-DoF grasp detection network [19] and GSNet [35]. Following [35], we employ a sparse UNet based on Minkowski Engine [7] as the backbone and calculate the "graspness" for each point to generate grasp candidates. For local feature extraction, the multi-scale cylinder grouping proposed in [19] is employed to improve the performance of the baseline model. For the PCR during training, $\theta$ and $\mu$ which control the contact distance are set to 0.02 and 0.005 separately and the weight $\phi$ for the regularization is set to 0.1. For the C-SJO, we set the hyper-parameter $\alpha = 0.2$ , $\beta = 0.01$ and $\gamma = 5$ . + +# 4.2. Results on Physical Constraint Regularization + +We show the results of PCR in Table 1. Compared to the baseline model, integrating the PCR delivers an improvement of $4.64\%$ on the novel set, demonstrating its effectiveness for objects with diverse shapes and structures. Besides, the PCR also improves by $0.58\%$ in seen set and $0.73\%$ in similar set. This illustrates that for seen and similar objects in the training set, introducing physical prior can help the grasp detection network to fully utilize the grasp labels compared to fitting them directly. We also ablate the influence of different physical constraints and notice that the incomplete constraint conditions can attribute to the performance decline on seen and similar set. Only with partial constraints during training, not all the grasp poses which meet the constraints are correct, thus disturbing the learning from grasping label. + +# 4.3. Results on Contact-Score Joint Optimization + +We employ the C-SJO in our 6-DoF grasp detection network trained with PCR and conduct ablations. The results + +Table 1. Results of PCR on scenes captured by RealSense. + +
ModelSeenSimilarNovel
No Refine66.6765.5535.58
+ C-S (Predicted Mask)66.6165.3736.67
+ C-S (GT Mask)66.6965.5036.61
Original C-S66.4865.2936.27
w/o C66.6565.6435.86
w/o S66.5465.3736.23
+ +Table 2. Results of C-SJO on scenes captured by RealSense. + +are shown in Table 2. We incorporate a recent proposed 3D segmentation method [38] to get object point-clouds from the clutter for optimization. With the ground-truth mask and segmented mask, C-SJO can improve by $1.03\%$ and $1.00\%$ on novel set separately. The C-SJO has almost no impact on the seen and similar set, primarily because the grasp poses predicted for seen and similar objects are sufficiently stable, making it challenging to find a better grasp than the original one. However, for the novel set, many unstable grasps exist that can potentially be optimized. We also conduct ablations about different designs of the C-SJO. Without the proposed projection contact map, only employing the original contact map (Original C-S) drops on seen, similar and novel set due to the singularity poses during optimization. Optimizing solely based on grasp score $(w / oC)$ overlooks the contact map prior, resulting in very limited improvement on the novel set. Using only the contact map for optimization $(w / oS)$ yields a $0.38\%$ drop on the novel set, but the absence of grasp scores restricts the search space, making it challenging to locate the locally optimal grasp during the optimization process. The process of grasp pose optimization is shown in Fig. 5, where the optimal grasp poses are reached in the original grasps' neighbor $SE(3)$ space. For object (1)-(4), the optimization process refines the inaccurate initial grasp poses. For objects (5) and (6), the optimization process facilitates collision avoidance. + +# 4.4. Comparison with State-of-the-art + +To make a fair comparison with the state-of-the-art methods of the GraspNet-1billion benchmark, we re-implement + +
ModelSeenSimilarNovel
AP\(AP_{0.8}\)\(AP_{0.4}\)AP\(AP_{0.8}\)\(AP_{0.4}\)AP\(AP_{0.8}\)\(AP_{0.4}\)
GraspNet-baseline [10]28.1030.5326.5623.8726.9222.517.488.434.93
Scale-balanced Grasp [19]46.0551.0244.2737.7644.2734.7517.0921.0410.20
GSNet [35]60.4771.0552.0658.5569.3853.0828.0636.1914.09
Ours Baseline66.0975.5760.6364.8274.1061.5130.6137.6117.06
Ours66.6175.6761.5265.3774.3561.2836.6745.0820.90
+ +![](images/5323dc25d95ebfc5c9c75e0c884d2b1cd27309bde9cf921971cccc02b44c8146.jpg) +Figure 5. Visualization of the process of C-SJO. + +three representative 6-DoF grasp detection methods [10, 19, 35] with our multi-view reconstructed scene as input. As shown in Table 3, our method performs better than previous 6-DoF grasp detection methods on all test sets and achieves $36.67\%$ on the novel set, which demonstrates the generalization capability of our method. + +# 4.5. Comparison with object augmentation + +We give a comparison between our prior knowledge based method and the object augmentation which employed by previous methods to enhance the generalization capability. To migrate the procedural or learning-based object augmentation method on the GraspNet-1billion benchmark, since the similar objects and the training objects have similar distributions, we consider the similar set in the benchmark as objects obtained through the augmentation. We merge the two sets for joint training. In this way, we can conveniently validate the performance of data augmentation on this benchmark without the need to generate additional objects and annotate grasps. As shown in Table 4, introducing object augmentation can slightly improve the performance on seen and novel sets, benefiting from the richer object distribution. However, our method performs better on the novel set, improving by $4.37\%$ compared to the object augmentation method, without introducing any additional data. Simultaneously, our method also exhibits an increase of $0.33\%$ on novel set when paired with object augmentation, demonstrating that our method and object augmentation can be used together. + +Table 3. Comparison with the state-of-the-art methods on RealSense scenes of GraspNet-billion benchmark. + +
ModelSeenSimilarNovel
Baseline66.0964.8230.61
Augmentation68.12-32.30
Ours66.6165.3736.67
Ours + Augmentation68.31-37.00
+ +Table 4. Comparison with object augmentation. + +# 4.6. Result Visualization + +In figure 6, we visualized the results generated by the baseline method, object augmentation and our approach on the GraspNet-1billion benchmark. The gripper poses in red represent successful grasps, while those shown in purple and blue correspond to collision and bad grasps. By incorporating domain prior knowledge, our method can generate grasp poses that conform to the physical relationship between the gripper and the object. For objects (3)-(5), our method predicts some failed grasps, primarily due to poor grasp sampling locations or inaccurate perception of the object's geometry. + +# 4.7. Real-world Evaluation + +The effectiveness of the proposed method in real-world is validated by real robot in this section. As shown in Figure 7 (a), the robotic grasping system is built on a 6-DoF UR-10 robot arm and a RealSense D435i depth camera is employed for scene perception. The total of 20 objects used for grasp + +![](images/866ec070b7c5e5f297b12d4dd96c75b2ce4b6cec7998958a1d105e384f2d11a9.jpg) +Figure 6. Visualization of the predicted grasp poses from baseline, object augmentation and the methods proposed in this paper. + +![](images/937e7614cedf02197c3d3b380630379f10f8f4cdc9ab06cde4a742f13a403aaa.jpg) +Figure 7. (a) Robotic grasping system and (b) objects for real-world grasping. + +![](images/457b9240a339d102dfc7fe2b1a327bd3e57fd19197a21ef7091c94e79d6ae3c2.jpg) + +ing (Fig. 7 (b)) are composed of two parts: 13 3D-printed objects from Dex-Net [20] and 7 objects chosen from the YCB dataset [4]. All of the objects are significantly different from the objects used for training in GraspNet-1billion. Before executing the proposed grasp detection method, we quickly reconstruct the scene using an arm-mounted depth camera based on the KinectFusion [25]. + +
ModelIsolatedCluttered
SR (%)SR (%)SCR (%)
Baseline58.33 (35/60)48.21 (43/89)86.00 (43/50)
Ours68.33 (41/60)64.86 (48/74)96.00 (48/50)
+ +Table 5. Results of the real-world grasping experiments. + +We compare our model to the baseline in two settings: isolated object grasping and cluttered object grasping. For isolated object grasping, each object is placed in three different poses and Success Rate (SR) is used as the metric. + +For cluttered object grasping, we compose 5 objects into a scene and make the robot remove them all with a maximum number of operations at 10. SR and Scene Completion Rate (SCR) are employed as the metrics. As illustrated in Table 5, our model outperforms the baseline in both settings and achieves better SR and SCR, demonstrating its superiority for novel objects. + +# 5. Conclusion + +In this paper, we work for generalizing 6-DoF grasp detection with domain prior knowledge of robotic grasping. The physical constraint regularization based on physical rules is proposed to enable the generalization of objects with largely varied shapes and structures. To refine the unstable results predicted by network in cluttered scenarios, we specially design a contact-score joint optimization with the contact map prior, in which a projection contact map is utilized. Extensive experiments on both the benchmark and the real-world robot demonstrate the effectiveness of our method for novel objects. + +# Acknowledgment + +This work is partly supported by the National Key R&D Program of China (2022ZD0161902), the National Natural Science Foundation of China (62022011), the Research Program of State Key Laboratory of Software Development Environment (SKLSDE-2023ZX-14), and the Fundamental Research Funds for the Central Universities. + +# References + +[1] Konstantinos Bousmalis, Alex Irpan, Paul Wohlhart, Yunfei Bai, Matthew Kelcey, Mrinal Kalakrishnan, Laura Downs, Julian Ibarz, Peter Pastor, Kurt Konolige, et al. Using simulation and domain adaptation to improve efficiency of deep robotic grasping. In IEEE International Conference on Robotics and Automation, 2018. 2, 3 +[2] Samarth Brahmbhatt, Cusuh Ham, Charles C. Kemp, and James Hays. Contactdb: Analyzing and predicting grasp contact via thermal imaging. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2 +[3] Michel Breyer, Jen Jen Chung, Lionel Ott, Roland Siegwart, and Juan I. Nieto. Volumetric grasping network: Real-time 6 DOF grasp detection in clutter. In Conference on Robot Learning, 2020. 2 +[4] Berk Çallı, Arjun Singh, James Bruce, Aaron Walsman, Kurt Konolige, Siddhartha S. Srinivasa, Pieter Abbeel, and Aaron M. Dollar. Yale-cmu-berkeley dataset for robotic manipulation research. International Journal of Robotics Research, 36(3):261-268, 2017. 8 +[5] I-Ming Chen and Joel W. Burdick. Finding antipodal point grasps on irregularly shaped objects. IEEE Transactions on Robotics and Automation, 9(4):507-512, 1993. 3 +[6] I-Ming Chen and Joel W. Burdick. Finding antipodal point grasps on irregularly shaped objects. IEEE Transactions on Robotics and Automation, 9(4):507-512, 1993. 1 +[7] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 6 +[8] Qiyu Dai, Yan Zhu, Yiran Geng, Ciyu Ruan, Jiazhao Zhang, and He Wang. Graspnerf: Multiview-based 6-dof grasp detection for transparent and specular objects using generalizable nerf. In IEEE International Conference on Robotics and Automation, 2023. 2, 3 +[9] Amaury Depierre, Emmanuel Dellandrea, and Liming Chen. Jacquard: A large scale dataset for robotic grasp detection. In IEEE/RSJ International Conference on Intelligent Robots and Systems, 2018. 1, 2 +[10] Haoshu Fang, Chenxi Wang, Minghao Gou, and Cewu Lu. Graspnet-1billion: A large-scale benchmark for general object grasping. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2, 5, 7 +[11] Carlo Ferrari and John F. Canny. Planning optimal grasps. In IEEE International Conference on Robotics and Automation, 1992. 3 +[12] Hanwen Jiang, Shaowei Liu, Jiashun Wang, and Xiaolong Wang. Hand-object contact consistency reasoning for human grasps generation. In IEEE/CVF International Conference on Computer Vision, 2021. 4 +[13] Junnan Jiang, Xiaohui Xiao, Fei Chen, and Miao Li. Learning grasp ability enhancement through deep shape generation. In International Conference on Intelligent Robotics and Applications, 2022. 2, 3 +[14] Zhenyu Jiang, Yifeng Zhu, Maxwell Svetlik, Kuan Fang, and Yuke Zhu. Synergies between affordance and geometry: 6- + +dof grasp detection via implicit representations. In Robotics: Science and Systems, 2021. 2, 3 +[15] George Em Karniadakis, Ioannis G Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, and Liu Yang. Physics-informed machine learning. Nature Reviews Physics, 3(6):422-440, 2021. 2 +[16] Hongzhuo Liang, Xiaojian Ma, Shuang Li, Michael Gorner, Song Tang, Bin Fang, Fuchun Sun, and Jianwei Zhang. Pointnetgpd: Detecting grasp configurations from point sets. In IEEE International Conference on Robotics and Automation, 2019. 2 +[17] Min Liu, Zherong Pan, Kai Xu, Kanishka Ganguly, and Dinesh Manocha. Deep differentiable grasp planner for high-dof grippers. In Robotics: Science and Systems, 2020. 3 +[18] Tengyu Liu, Zeyu Liu, Ziyuan Jiao, Yixin Zhu, and Song-Chun Zhu. Synthesizing diverse and physically stable grasps with arbitrary hand structures using differentiable force closure estimator. IEEE Robotics and Automation Letters, 7(1): 470-477, 2022. 3 +[19] Haoxiang Ma and Di Huang. Towards scale balanced 6-dof grasp detection in cluttered scenes. In Conference on Robot Learning, 2022. 2, 3, 6, 7 +[20] Jeffrey Mahler, Jacky Liang, Sherdil Niyaz, Michael Laskey, Richard Doan, Xinyu Liu, Juan Aparicio Ojea, and Ken Goldberg. Dex-net 2.0: Deep learning to plan robust grasps with synthetic point clouds and analytic grasp metrics. In Robotics: Science and Systems, 2017. 1, 8 +[21] Andrew T. Miller and Peter K. Allen. Graspit! A versatile simulator for robotic grasping. IEEE Robotics & Automation Magazine, 11(4):110-122, 2004. 3 +[22] Douglas Morrison, Juxi Leitner, and Peter Corke. Closing the loop for robotic grasping: A real-time, generative grasp synthesis approach. In Robotics: Science and Systems, 2018. 2 +[23] Douglas Morrison, Peter Corke, and Jürgen Leitner. Egad! an evolved grasping analysis dataset for diversity and reproducibility in robotic manipulation. IEEE Robotics and Automation Letters, 5(3):4368-4375, 2020. 2, 3 +[24] Arsalan Mousavian, Clemens Eppner, and Dieter Fox. 6-dof grapnet: Variational grasp generation for object manipulation. In IEEE/CVF International Conference on Computer Vision, 2019. 2 +[25] Richard A. Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J. Davison, Pushmeet Kohli, Jamie Shotton, Steve Hodges, and Andrew W. Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. In IEEE International Symposium on Mixed and Augmented Reality, 2011. 5, 8 +[26] Van-Duc Nguyen. Constructing force-closure grasps. International Journal of Robotics Research, 7(3):3-16, 1988. 2, 3 +[27] Nancy S. Pollard. Closure and quality equivalence for efficient synthesis of grasps from examples. International Journal of Robotics Research, 23(6):595-613, 2004. 1 +[28] Ran Qin, Haoxiang Ma, Boyang Gao, and Di Huang. RGB-D grasp detection via depth guided learning with cross-modal attention. In IEEE International Conference on Robotics and Automation, 2023. 2 + +[29] Yuzhe Qin, Rui Chen, Hao Zhu, Meng Song, Jing Xu, and Hao Su. S4G: amodal single-view single-shot SE(3) grasp detection in cluttered scenes. In Conference on Robot Learning, 2019. 2 +[30] Maximo A. Roa and Raul Suárez. Computation of independent contact regions for grasping 3-d objects. IEEE Transactions on Robotics, 25(4):839-850, 2009. 1 +[31] Alberto Rodriguez, Matthew T. Mason, and Steve Ferry. From caging to grasping. International Journal of Robotics Research, 31(7):886-900, 2012. 3 +[32] Andreas ten Pas, Marcus Gualtieri, Kate Saenko, and Robert Platt Jr. Grasp pose detection in point clouds. International Journal of Robotics Research, 36(13-14):1455-1473, 2017. 1, 2 +[33] Josh Tobin, Lukas Biewald, Rocky Duan, Marcin Andrychowicz, Ankur Handa, Vikash Kumar, Bob McGrew, Alex Ray, Jonas Schneider, Peter Welinder, et al. Domain randomization and generative models for robotic grasping. In IEEE/RSJ International Conference on Intelligent Robots and Systems, 2018. 2, 3 +[34] Laura Von Rueden, Sebastian Mayer, Katharina Beckh, Bogdan Georgiev, Sven Giesselbach, Raoul Heese, Birgit Kirsch, Julius Pfrommer, Annika Pick, Rajkumar Ramamurthy, et al. Informed machine learning-a taxonomy and survey of integrating prior knowledge into learning systems. IEEE Transactions on Knowledge and Data Engineering, 35(1):614-633, 2021. 2 +[35] Chenxi Wang, Haoshu Fang, Minghao Gou, Hongjie Fang, Jin Gao, and Cewu Lu. Graspness discovery in clutters for fast and accurate grasp detection. In IEEE/CVF International Conference on Computer Vision, 2021. 2, 3, 6, 7 +[36] David Wang, David Tseng, Pusong Li, Yiding Jiang, Menglong Guo, Michael Danielczuk, Jeffrey Mahler, Jeffrey Ichnowski, and Ken Goldberg. Adversarial grasp objects. In IEEE International Conference on Automation Science and Engineering, 2019. 3 +[37] Yinzhen Xu, Weikang Wan, Jialiang Zhang, Haoran Liu, Zikang Shan, Hao Shen, Ruicheng Wang, Haoran Geng, Yi-jia Weng, Jiayi Chen, Tengyu Liu, Li Yi, and He Wang. Unidexgrasp: Universal robotic dexterous grasping via learning diverse proposal generation and goal-conditioned policy. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 4 +[38] Yunhan Yang, Xiaoyang Wu, Tong He, Hengshuang Zhao, and Xihui Liu. Sam3d: Segment anything in 3d scenes, 2023. 6 \ No newline at end of file diff --git a/2024/Generalizing 6-DoF Grasp Detection via Domain Prior Knowledge/images.zip b/2024/Generalizing 6-DoF Grasp Detection via Domain Prior Knowledge/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6833c4f77b2d7f59dcd2ec2712549388a7abcc00 --- /dev/null +++ b/2024/Generalizing 6-DoF Grasp Detection via Domain Prior Knowledge/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8a410b77e7a38fd87a83037dad313d0fecd8a253097bc1c6fcbbf31a6e40a54f +size 651479 diff --git a/2024/Generalizing 6-DoF Grasp Detection via Domain Prior Knowledge/layout.json b/2024/Generalizing 6-DoF Grasp Detection via Domain Prior Knowledge/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a103677f696954d14f261ce25b91eb1e9a9575de --- /dev/null +++ b/2024/Generalizing 6-DoF Grasp Detection via Domain Prior Knowledge/layout.json @@ -0,0 +1,7981 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 91, + 103, + 503, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 103, + 503, + 121 + ], + "spans": [ + { + "bbox": [ + 91, + 103, + 503, + 121 + ], + "type": "text", + "content": "Generalizing 6-DoF Grasp Detection via Domain Prior Knowledge" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 145, + 142, + 452, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 142, + 452, + 158 + ], + "spans": [ + { + "bbox": [ + 145, + 142, + 452, + 158 + ], + "type": "text", + "content": "Haoxiang Ma" + }, + { + "bbox": [ + 145, + 142, + 452, + 158 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 145, + 142, + 452, + 158 + ], + "type": "text", + "content": " Modi Shi" + }, + { + "bbox": [ + 145, + 142, + 452, + 158 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 145, + 142, + 452, + 158 + ], + "type": "text", + "content": " Boyang Gao" + }, + { + "bbox": [ + 145, + 142, + 452, + 158 + ], + "type": "inline_equation", + "content": "^{3,4}" + }, + { + "bbox": [ + 145, + 142, + 452, + 158 + ], + "type": "text", + "content": " Di Huang" + }, + { + "bbox": [ + 145, + 142, + 452, + 158 + ], + "type": "inline_equation", + "content": "^{1,2*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 60, + 158, + 534, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 158, + 534, + 213 + ], + "spans": [ + { + "bbox": [ + 60, + 158, + 534, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 60, + 158, + 534, + 213 + ], + "type": "text", + "content": "State Key Laboratory of Software Development Environment, Beihang University, Beijing, China \n" + }, + { + "bbox": [ + 60, + 158, + 534, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 60, + 158, + 534, + 213 + ], + "type": "text", + "content": "School of Computer Science and Engineering, Beihang University, Beijing, China \n" + }, + { + "bbox": [ + 60, + 158, + 534, + 213 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 60, + 158, + 534, + 213 + ], + "type": "text", + "content": "School of Computer Science and Technology, Harbin Institute of Technology, Harbin, China \n" + }, + { + "bbox": [ + 60, + 158, + 534, + 213 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 60, + 158, + 534, + 213 + ], + "type": "text", + "content": "Geometry Robotics" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 93, + 215, + 492, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 215, + 492, + 228 + ], + "spans": [ + { + "bbox": [ + 93, + 215, + 492, + 228 + ], + "type": "text", + "content": "{mahaoxiang822, modishi, dhuang}@buaa.edu.cn, {boyang.gao}@geometryrobot.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "spans": [ + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 280, + 290, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 280, + 290, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 280, + 290, + 555 + ], + "type": "text", + "content": "We focus on the generalization ability of the 6-DoF grasp detection method in this paper. While learning-based grasp detection methods can predict grasp poses for unseen objects using the grasp distribution learned from the training set, they often exhibit a significant performance drop when encountering objects with diverse shapes and structures. To enhance the grasp detection methods' generalization ability, we incorporate domain prior knowledge of robotic grasping, enabling better adaptation to objects with significant shape and structure differences. More specifically, we employ the physical constraint regularization during the training phase to guide the model towards predicting grasps that comply with the physical rule on grasping. For the unstable grasp poses predicted on novel objects, we design a contact-score joint optimization using the projection contact map to refine these poses in cluttered scenarios. Extensive experiments conducted on the GraspNet1billion benchmark demonstrate a substantial performance gain on the novel object set and the real-world grasping experiments also demonstrate the effectiveness of our generalizing 6-DoF grasp detection method. Code is available at https://github.com/mahaoxiang822/Generalizing-Grasp." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 578, + 128, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 578, + 128, + 590 + ], + "spans": [ + { + "bbox": [ + 47, + 578, + 128, + 590 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 598, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 598, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 598, + 287, + 696 + ], + "type": "text", + "content": "Given an object, robotic grasp detection aims to find suitable and sufficient gripper configurations for various manipulation tasks. Traditional methods [6, 27, 30] establish hand-crafted criteria to evaluate grasp samples according to 3D models of objects. Despite providing precise interaction between objects and grippers, they assume that object models are available in advance and also suffer a slow running speed, which limits their popularization. As the develop-" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 326, + 255, + 529, + 376 + ], + "blocks": [ + { + "bbox": [ + 326, + 255, + 529, + 376 + ], + "lines": [ + { + "bbox": [ + 326, + 255, + 529, + 376 + ], + "spans": [ + { + "bbox": [ + 326, + 255, + 529, + 376 + ], + "type": "image", + "image_path": "2b83e1a29032bc4276eb2195c87004d7045fae79e5169984924e3d834b8b9787.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 325, + 380, + 529, + 428 + ], + "blocks": [ + { + "bbox": [ + 325, + 380, + 529, + 428 + ], + "lines": [ + { + "bbox": [ + 325, + 380, + 529, + 428 + ], + "spans": [ + { + "bbox": [ + 325, + 380, + 529, + 428 + ], + "type": "image", + "image_path": "406f56a7c2c9e2ba8a7008c28b895a9562dfae83e899c531e6d42950f5ff542f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 420, + 430, + 432, + 440 + ], + "lines": [ + { + "bbox": [ + 420, + 430, + 432, + 440 + ], + "spans": [ + { + "bbox": [ + 420, + 430, + 432, + 440 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 330, + 441, + 524, + 485 + ], + "blocks": [ + { + "bbox": [ + 330, + 441, + 524, + 485 + ], + "lines": [ + { + "bbox": [ + 330, + 441, + 524, + 485 + ], + "spans": [ + { + "bbox": [ + 330, + 441, + 524, + 485 + ], + "type": "image", + "image_path": "8967d8b099bbeea45c76cfe0c378003d1b6b48fcdfb700c63895f5f0a426ce40.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 421, + 487, + 432, + 496 + ], + "lines": [ + { + "bbox": [ + 421, + 487, + 432, + 496 + ], + "spans": [ + { + "bbox": [ + 421, + 487, + 432, + 496 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 506, + 547, + 551 + ], + "lines": [ + { + "bbox": [ + 304, + 506, + 547, + 551 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 547, + 551 + ], + "type": "text", + "content": "Figure 1. Illustration of (a) the performance and distribution gap between objects similar to training samples (similar objects) and those with largely varied shapes and structures (novel objects) and (b) the limitations of the object augmentation methods." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 557, + 547, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 700 + ], + "type": "text", + "content": "ment of deep learning, data-driven methods [9, 20, 32] have been widely studied. These methods predict grasps without the need for pre-prepared 3D models and demonstrate the capability to handle unseen objects. However, they often struggle when encountering objects whose shapes and structures significantly deviate from those in the training set, making them difficult to adapt to more diverse applications. As illustrated in Fig. 1 (a), learning-based grasp detection methods are basically able to transfer to objects similar to those appearing in training, but they do not effectively generalize to novel objects with their distribution greatly changed." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "text", + "content": "To facilitate the generalization of learning-based grasp" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18102" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 300 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 300 + ], + "type": "text", + "content": "detection methods towards a variety of unseen objects, previous attempts usually apply data augmentation techniques to expand the distribution of the training set. More objects are typically produced by randomly concatenating a number of 3D primitives [1, 33] or by directly synthesizing through generative networks [13, 23]. They do enrich the diversity of objects for training, which leads to a gain in overall performance. However, since such augmentation is conducted only according to original data, it contributes more to similar objects whose distribution is largely overlapped and offers little help to new objects whose distribution is not well covered. As depicted in Fig. 1 (b), on the one hand, to the objects that are far from the center of training samples in the feature space, totally wrong grasps could be delivered. On the other hand, for the ones that are near the center of training samples in the feature space, although reasonable grasps may be obtained, the predictions tend to be unstable due to the distribution gap. Both cases make those methods problematic to generalize to unseen targets." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 303, + 289, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 303, + 289, + 555 + ], + "spans": [ + { + "bbox": [ + 47, + 303, + 289, + 555 + ], + "type": "text", + "content": "To address the dilemma above, inspired by informed machine learning [15, 34], we introduce domain prior knowledge of robotic grasping. Compared to data augmentation, domain prior knowledge does not depend on the distribution of training data, allowing for easy adaptation to objects with significant shape and structure differences. Firstly, to enable the grasp detection network to generalize to novel objects, we incorporate certain physical rules on grasping, such as force conditions [26] and contact positions, which are typically used in the traditional analytical grasping methods. The physical rules provide valuable clues in terms of grasp stability regardless of object-specific properties, thus motivating us to integrate such rule-based priors into the grasp detection network. Secondly, to deal with the unstable results predicted by the network, we introduce another type of knowledge about the interaction of the gripper and object. The contact map [2] is able to denote the regions correlated with grasping as well as the kinematics of the gripper. By employing a neural network to encode the preferred object regions for optimal grasps, the learning-based contact map prior can be utilized to refine insufficiently accurate grasps." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 558, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 558, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 558, + 289, + 715 + ], + "type": "text", + "content": "Concretely, in this paper, we propose a generalized 6-DoF grasp detection framework with domain prior knowledge, which consists of two components: Physical Constraint Regularization (PCR) and Contact-Score Joint Optimization (C-SJO). In PCR, we integrate the physical prior into the network as the regularization, thereby constraining the correlation between gripper poses and object models. To enable the back-propagation of PCR, we employ an end-to-end 6-DoF grasp detection network for grasp prediction and utilize the Signed Distance Field (SDF) to encode the object model, facilitating the differentiable computation of physical constraints. Compared to fitting grasp annotations directly, PCR guides the network to predict grasps follow" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 547, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 323 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 323 + ], + "type": "text", + "content": "ing object-independent physical rules, thereby enhancing the generalization capability for novel objects. To refine the unstable prediction for novel objects, we introduce the C-SJO at test time based on the contact map prior. The contact map represents the contact region on the object's surface by calculating the distance between the gripper model and the objects. A contact map prediction network is introduced to encode the contact pattern of good grasps by learning from grasp labels. By aligning the contact map of the current grasp and the prediction from the network, the grasp pose can be optimized within its neighborhood. However, as for 6-DoF grasp detection in clutter, using the euclidean distance for contact map calculation cannot adequately address inaccurate contact positions. Therefore, we introduce a projection contact map to solve this problem. Additionally, due to the noise from the depth sensor and the occlusion in cluttered scenarios, relying solely on the contact prior can lead to singular results. To mitigate this, we use score optimization to constrain the searching space of contact optimization, in which a grasp score network is utilized." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 317, + 323, + 533, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 323, + 533, + 335 + ], + "spans": [ + { + "bbox": [ + 317, + 323, + 533, + 335 + ], + "type": "text", + "content": "The contribution of this paper can be summarized as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 337, + 545, + 445 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 306, + 337, + 545, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 337, + 545, + 372 + ], + "spans": [ + { + "bbox": [ + 306, + 337, + 545, + 372 + ], + "type": "text", + "content": "- We propose a domain prior knowledge informed 6-DoF grasp detection framework to enhance the generalization ability for novel objects." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 373, + 545, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 373, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 306, + 373, + 545, + 408 + ], + "type": "text", + "content": "- We design the physical constraint regularization to represent the physical prior for 6-DoF grasping and integrate it into the network in a differential strategy." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 409, + 545, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 409, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 306, + 409, + 545, + 445 + ], + "type": "text", + "content": "- We employ a contact-score joint optimization with the projection contact map, which is suitable for refining the inaccurate prediction in cluttered scenarios." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 454, + 389, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 454, + 389, + 466 + ], + "spans": [ + { + "bbox": [ + 306, + 454, + 389, + 466 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 474, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 547, + 715 + ], + "type": "text", + "content": "6-DoF Grasp Detection To generate diverse and feasible grasping in cluttered scenes, 6-DoF grasp detection has been advanced recently. Compared to planar grasp detection where the grasp space is limited [9, 22, 28], 6-DoF grasp detection method can predict grasping in " + }, + { + "bbox": [ + 304, + 474, + 547, + 715 + ], + "type": "inline_equation", + "content": "SE(3)" + }, + { + "bbox": [ + 304, + 474, + 547, + 715 + ], + "type": "text", + "content": " space, thereby supporting more complex downstream task. [32] pioneers a sampling and evaluation framework for 6-DoF grasp detection. Using point-cloud input, they heuristically sample grasp candidates and employ a Convolutional Neural Network (CNN) for scoring them. [16] build upon the framework, utilizing a neural network for point-cloud data to achieve improved grasp sample evaluation. More recently, a series of studies [3, 8, 10, 14, 19, 24, 29, 35] propose end-to-end strategy for grasp prediction. [24] employs a variational auto-encoder for grasp generation given the object point-clouds. [29] introduces a single-shot grasp proposal network for efficient grasp detection based on single view point-clouds, and [3] employs a Truncated Signed Distance Function (TSDF) to map multiple frames and predict grasps in voxel space. [10] provides a large benchmark" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18103" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "with densely annotated grasp labels and designs a baseline method for dense grasp prediction. Subsequently, [35] defines graspsness to represent grasp probability in searching space and [19] focuses on the scale imbalance problem for 6-DoF grasp detection. With the development of neural representations, [14] utilizes an occupancy network to learn a shared representation between 3D reconstruction and grasp detection. [8] introduces a generalizable neural radiance field for the grasp detection of transparent and specular objects. Despite these advancements in 6-DoF grasp detection, the performance for novel objects, particularly those with varied shapes and structures, remains sub-optimal. In this paper, we investigate this shortcoming." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 232, + 289, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 232, + 289, + 508 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 289, + 508 + ], + "type": "text", + "content": "Generalization on Grasp Detection Several strategies are proposed to enhance the generalization capability of grasp detection, which mainly focuses on enriching the distribution of training objects. [1] procedurally generates objects by attaching rectangular prisms at random locations and orientations. To enhance the shape diversity, [33] constructs a set of object primitives by decomposing everyday objects and generates diverse objects by randomly sampling these primitives and combing them. In addition to heuristically generating random objects, [23] provides the Evolved Grasping Analysis Dataset (EGAD), which includes objects of varying shape complexity and grasp difficulty generated by the 3D compositional pattern producing networks. [13] employs an AutoEncoder-Critic network to interpolate new shapes from two objects for augmentation. Besides augmenting current shapes to improve the performance of the learning-based grasp detection method, [36] explores the task of generating adversarial objects that are difficult to grasp. Different from previous augmentation methods for generalized grasp detection, we introduce the grasp domain knowledge to our framework explicitly. Therefore, our method doesn't rely on the distribution of training objects, resulting in an improvement for out-of-distribution objects." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": "Usage of domain knowledge in Grasping Domain prior knowledge has been leveraged in various grasping applications. Early analytical methods [11, 21, 31] make use of the physical prior knowledge to examine the force dynamics between the gripper and the object. More recently, several methods incorporate some grasp prior knowledge into the hand-object grasp synthesis. [18] introduces a differentiable force closure algorithm designed to optimize hand configurations, which facilitates the fast generation of physically stable grasps. [17] designs a generalized Q1 metric that serves as a loss for a grasp planner, producing precise multifinger grasping for single object with watertight model or rendered depth images. While the aforementioned methods employ domain prior knowledge to assist grasp synthesis and optimization, they rely on the accurate object shapes and design complex calculations for prior integration. This makes it difficult to apply them in the 6-DoF grasp detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 547, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 109 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 109 + ], + "type": "text", + "content": "where the generation of diverse grasps in complex scenarios is required and the object geometry from the depth sensor is inaccurate." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 118, + 362, + 130 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 118, + 362, + 130 + ], + "spans": [ + { + "bbox": [ + 305, + 118, + 362, + 130 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 137, + 493, + 151 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 137, + 493, + 151 + ], + "spans": [ + { + "bbox": [ + 305, + 137, + 493, + 151 + ], + "type": "text", + "content": "3.1. Physical Constraint Regularization" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 156, + 547, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 156, + 547, + 277 + ], + "spans": [ + { + "bbox": [ + 304, + 156, + 547, + 277 + ], + "type": "text", + "content": "In terms of physical prior knowledge, the force interaction between the object and the gripper is used to analyze grasp stability, among which the force closure [26] is widely adopted. A force closure grasp can resist any external wrenches with the contact force if the force direction lies in the friction cone. In scenarios where the friction coefficient of the object surface is unknown and a two-finger gripper is utilized, we employ the antipodal rule [5] as a simplification. Given two contact points " + }, + { + "bbox": [ + 304, + 156, + 547, + 277 + ], + "type": "inline_equation", + "content": "(u_{1}, u_{2})" + }, + { + "bbox": [ + 304, + 156, + 547, + 277 + ], + "type": "text", + "content": " lying on the object surface, the antipodal rule is formulated as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 368, + 283, + 545, + 299 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 283, + 545, + 299 + ], + "spans": [ + { + "bbox": [ + 368, + 283, + 545, + 299 + ], + "type": "interline_equation", + "content": "\\left[ p \\left(u _ {1}\\right) - p \\left(u _ {2}\\right) \\right] \\cdot t \\left(u _ {1}\\right) = 0 \\tag {1}", + "image_path": "d17c0f79a2941bf621525d7be27cd1228377ded460e7ff2952a0665cd8238f22.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 369, + 307, + 545, + 322 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 307, + 545, + 322 + ], + "spans": [ + { + "bbox": [ + 369, + 307, + 545, + 322 + ], + "type": "interline_equation", + "content": "[ p (u _ {2}) - p (u _ {1}) ] \\cdot t (u _ {2}) = 0 \\tag {2}", + "image_path": "10a538f8be64a81531a4f99b500141263e61835f609a8513a36f4a0a02a9982b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 384, + 328, + 545, + 342 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 328, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 384, + 328, + 545, + 342 + ], + "type": "interline_equation", + "content": "n \\left(u _ {1}\\right) + n \\left(u _ {2}\\right) = 0 \\tag {3}", + "image_path": "3d1b807c876040dc3ec0ef82d13011bdf02a5ccc8724692e4f99fce82a1adece.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 346, + 545, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 346, + 545, + 428 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 545, + 428 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 346, + 545, + 428 + ], + "type": "inline_equation", + "content": "p(u)" + }, + { + "bbox": [ + 304, + 346, + 545, + 428 + ], + "type": "text", + "content": " represents the contact position, " + }, + { + "bbox": [ + 304, + 346, + 545, + 428 + ], + "type": "inline_equation", + "content": "t(u)" + }, + { + "bbox": [ + 304, + 346, + 545, + 428 + ], + "type": "text", + "content": " represents the unit tangent vector and " + }, + { + "bbox": [ + 304, + 346, + 545, + 428 + ], + "type": "inline_equation", + "content": "n(u)" + }, + { + "bbox": [ + 304, + 346, + 545, + 428 + ], + "type": "text", + "content": " represents the unit outward normal vector. If the contact points " + }, + { + "bbox": [ + 304, + 346, + 545, + 428 + ], + "type": "inline_equation", + "content": "(u_1, u_2)" + }, + { + "bbox": [ + 304, + 346, + 545, + 428 + ], + "type": "text", + "content": " between the gripper and the object comply with the antipodal rule, there is a higher probability of a successful grasp. Based on which, we introduce the PCR to constrain the output of the network with the rule." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 429, + 547, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 429, + 547, + 464 + ], + "spans": [ + { + "bbox": [ + 304, + 429, + 547, + 464 + ], + "type": "text", + "content": "With the predicted grasp pose " + }, + { + "bbox": [ + 304, + 429, + 547, + 464 + ], + "type": "inline_equation", + "content": "g = [t, R, w]" + }, + { + "bbox": [ + 304, + 429, + 547, + 464 + ], + "type": "text", + "content": " and the gripper kinematic model " + }, + { + "bbox": [ + 304, + 429, + 547, + 464 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 429, + 547, + 464 + ], + "type": "text", + "content": ", the gripper contacts " + }, + { + "bbox": [ + 304, + 429, + 547, + 464 + ], + "type": "inline_equation", + "content": "(c_1, c_2)" + }, + { + "bbox": [ + 304, + 429, + 547, + 464 + ], + "type": "text", + "content": " can be expressed as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 380, + 472, + 545, + 487 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 472, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 380, + 472, + 545, + 487 + ], + "type": "interline_equation", + "content": "\\left(c _ {1}, c _ {2}\\right) = K (t, R, w) \\tag {4}", + "image_path": "4cf24b3f985c71cf185f82cd1ae1623375b9a8c88a18ef77783db4e5fa46d216.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 493, + 545, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 493, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 493, + 545, + 529 + ], + "type": "text", + "content": "To enforce the gripper contacts to comply with the antipodal constraint, we propose an antipodal regularization term, which is calculated as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 536, + 545, + 561 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 536, + 545, + 561 + ], + "spans": [ + { + "bbox": [ + 306, + 536, + 545, + 561 + ], + "type": "interline_equation", + "content": "R ^ {A} \\left(c _ {1}, c _ {2}\\right) = 1 - 0. 5 * \\left(\\cos \\left(\\overrightarrow {c _ {1} c _ {2}}, n \\left(c _ {2}\\right)\\right) + \\cos \\left(\\overrightarrow {c _ {2} c _ {1}}, n \\left(c _ {1}\\right)\\right)\\right) \\tag {5}", + "image_path": "1048607e9b7f68cdd20d19580b8089c944c5b42209870f7e8ae8f027729e2021.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 562, + 547, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 562, + 547, + 693 + ], + "spans": [ + { + "bbox": [ + 304, + 562, + 547, + 693 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 562, + 547, + 693 + ], + "type": "inline_equation", + "content": "\\overrightarrow{c_1c_2}" + }, + { + "bbox": [ + 304, + 562, + 547, + 693 + ], + "type": "text", + "content": " is the vector connecting the gripper's contact points " + }, + { + "bbox": [ + 304, + 562, + 547, + 693 + ], + "type": "inline_equation", + "content": "c_{1}" + }, + { + "bbox": [ + 304, + 562, + 547, + 693 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 562, + 547, + 693 + ], + "type": "inline_equation", + "content": "c_{2}" + }, + { + "bbox": [ + 304, + 562, + 547, + 693 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 562, + 547, + 693 + ], + "type": "inline_equation", + "content": "n(c_{1})" + }, + { + "bbox": [ + 304, + 562, + 547, + 693 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 562, + 547, + 693 + ], + "type": "inline_equation", + "content": "n(c_{2})" + }, + { + "bbox": [ + 304, + 562, + 547, + 693 + ], + "type": "text", + "content": " are the normal vectors at the respective contact points and " + }, + { + "bbox": [ + 304, + 562, + 547, + 693 + ], + "type": "inline_equation", + "content": "\\cos" + }, + { + "bbox": [ + 304, + 562, + 547, + 693 + ], + "type": "text", + "content": " represents the cosine similarity. Nevertheless, the predicted gripper contacts " + }, + { + "bbox": [ + 304, + 562, + 547, + 693 + ], + "type": "inline_equation", + "content": "c_{1}, c_{2}" + }, + { + "bbox": [ + 304, + 562, + 547, + 693 + ], + "type": "text", + "content": " may not lie precisely on the object surface. To overcome this, we also introduce constraints on the distance between the gripper contacts and the object's surface. These constraints help to avoid collision and ensure that the contact points are sufficiently close to the object's surface. The collision constraint " + }, + { + "bbox": [ + 304, + 562, + 547, + 693 + ], + "type": "inline_equation", + "content": "R^C" + }, + { + "bbox": [ + 304, + 562, + 547, + 693 + ], + "type": "text", + "content": " and surface constraint " + }, + { + "bbox": [ + 304, + 562, + 547, + 693 + ], + "type": "inline_equation", + "content": "R^S" + }, + { + "bbox": [ + 304, + 562, + 547, + 693 + ], + "type": "text", + "content": " are formulated as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 311, + 700, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 700, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 311, + 700, + 545, + 715 + ], + "type": "interline_equation", + "content": "R ^ {C} \\left(c _ {1}, c _ {2}\\right) = M a x (0, \\theta - d \\left(c _ {1}\\right)) + M a x (0, \\theta - d \\left(c _ {2}\\right)) \\tag {6}", + "image_path": "f6ef6522f265ae0c917d047946a22c341f294743780973213e94fe0e478e50e9.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "18104" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 87, + 71, + 509, + 213 + ], + "blocks": [ + { + "bbox": [ + 87, + 71, + 509, + 213 + ], + "lines": [ + { + "bbox": [ + 87, + 71, + 509, + 213 + ], + "spans": [ + { + "bbox": [ + 87, + 71, + 509, + 213 + ], + "type": "image", + "image_path": "4b2c0601a788be80ddfa67e9424d92418051a31807ecdfd9e81f79906880d71a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 220, + 547, + 255 + ], + "lines": [ + { + "bbox": [ + 46, + 220, + 547, + 255 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 547, + 255 + ], + "type": "text", + "content": "Figure 2. The pipeline of differential physical constraint integration. With the fusion point-cloud, the differentiable 6-DoF grasp network predicts the grasp configurations. The position of contacts " + }, + { + "bbox": [ + 46, + 220, + 547, + 255 + ], + "type": "inline_equation", + "content": "c_{1}, c_{2}" + }, + { + "bbox": [ + 46, + 220, + 547, + 255 + ], + "type": "text", + "content": " are calculated from grasp configurations by the gripper model and the regularization " + }, + { + "bbox": [ + 46, + 220, + 547, + 255 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 220, + 547, + 255 + ], + "type": "text", + "content": " is computed from the object SDF for back-propagation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 52, + 270, + 287, + 285 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 270, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 52, + 270, + 287, + 285 + ], + "type": "interline_equation", + "content": "R ^ {S} \\left(c _ {1}, c _ {2}\\right) = M a x (0, d \\left(c _ {1}\\right) - \\mu) + M a x (0, d \\left(c _ {2}\\right) - \\mu) \\tag {7}", + "image_path": "d9efcfa8eaf91b5d717c2e9729fce7c3aba19812bbb847ee7e8d3beef06c399f.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 292, + 288, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 292, + 288, + 375 + ], + "spans": [ + { + "bbox": [ + 47, + 292, + 288, + 375 + ], + "type": "text", + "content": "where the contact distance " + }, + { + "bbox": [ + 47, + 292, + 288, + 375 + ], + "type": "inline_equation", + "content": "d(c)" + }, + { + "bbox": [ + 47, + 292, + 288, + 375 + ], + "type": "text", + "content": " is constrained between " + }, + { + "bbox": [ + 47, + 292, + 288, + 375 + ], + "type": "inline_equation", + "content": "[\\theta, \\mu]" + }, + { + "bbox": [ + 47, + 292, + 288, + 375 + ], + "type": "text", + "content": ". Besides, we use the grasp score predicted by the grasp detection network to weight the regularization of different grasps, thereby ensuring the coherence between the grasp score and the physical constraints. With the grasp score of " + }, + { + "bbox": [ + 47, + 292, + 288, + 375 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 292, + 288, + 375 + ], + "type": "text", + "content": "th grasp " + }, + { + "bbox": [ + 47, + 292, + 288, + 375 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 47, + 292, + 288, + 375 + ], + "type": "text", + "content": ", the overall physical constraint " + }, + { + "bbox": [ + 47, + 292, + 288, + 375 + ], + "type": "inline_equation", + "content": "R_i" + }, + { + "bbox": [ + 47, + 292, + 288, + 375 + ], + "type": "text", + "content": " can be represented as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 380, + 287, + 425 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 380, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 106, + 380, + 287, + 425 + ], + "type": "interline_equation", + "content": "R _ {i} = \\frac {s _ {i} * \\left(R _ {i} ^ {A} + R _ {i} ^ {C} + R _ {i} ^ {S}\\right)}{\\frac {1}{M} \\sum_ {j} ^ {M} s _ {j}} \\tag {8}", + "image_path": "67905f4fd0fa0abc48bb4a77f9ef9671a718db615cbd251d399c7099bb2d585b.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 430, + 287, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 430, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 47, + 430, + 287, + 453 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 430, + 287, + 453 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 430, + 287, + 453 + ], + "type": "text", + "content": " is the number of seed points. The overall loss " + }, + { + "bbox": [ + 47, + 430, + 287, + 453 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 47, + 430, + 287, + 453 + ], + "type": "text", + "content": " can be formulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 124, + 461, + 287, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 461, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 124, + 461, + 287, + 475 + ], + "type": "interline_equation", + "content": "L = L _ {\\text {g r a s p}} + \\phi * R \\tag {9}", + "image_path": "c48c948a7329f1e03f7c5f19665eb886cdb5dec5d73781ef41e3659464051ff5.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 481, + 287, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 481, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 481, + 287, + 506 + ], + "type": "text", + "content": "For more details of the grasp configuration loss " + }, + { + "bbox": [ + 46, + 481, + 287, + 506 + ], + "type": "inline_equation", + "content": "L_{\\text{grasp}}" + }, + { + "bbox": [ + 46, + 481, + 287, + 506 + ], + "type": "text", + "content": ", please refer to the supplementary." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 512, + 275, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 512, + 275, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 512, + 275, + 525 + ], + "type": "text", + "content": "3.2. Differential Physical Constraint Integration" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 530, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 530, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 530, + 289, + 715 + ], + "type": "text", + "content": "To integrate the PCR into the grasp detection network, the sub-gradients of normal vector " + }, + { + "bbox": [ + 46, + 530, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\frac{\\partial n(c)}{\\partial c}" + }, + { + "bbox": [ + 46, + 530, + 289, + 715 + ], + "type": "text", + "content": " and surface distance " + }, + { + "bbox": [ + 46, + 530, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\frac{\\partial d(c)}{\\partial c}" + }, + { + "bbox": [ + 46, + 530, + 289, + 715 + ], + "type": "text", + "content": " with respect to the contact point " + }, + { + "bbox": [ + 46, + 530, + 289, + 715 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 530, + 289, + 715 + ], + "type": "text", + "content": " should be computed. Besides, the generation of grasp poses should be differential. As a result, we introduce an end-to-end 6-DoF grasp network and the differentiable Signed Distance Function (SDF) for the calculation of PCR. As shown in Fig. 2, the configurations of the gripper are regressed from the input fusion point-cloud in an end-to-end manner. With the configurations predicted by the grasp detection network, the contact point " + }, + { + "bbox": [ + 46, + 530, + 289, + 715 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 530, + 289, + 715 + ], + "type": "text", + "content": " can be calculated by the gripper kinematics model. Utilizing SDF of the object, we can query the surface distance of the contact " + }, + { + "bbox": [ + 46, + 530, + 289, + 715 + ], + "type": "inline_equation", + "content": "c_{1}" + }, + { + "bbox": [ + 46, + 530, + 289, + 715 + ], + "type": "text", + "content": " directly by " + }, + { + "bbox": [ + 46, + 530, + 289, + 715 + ], + "type": "inline_equation", + "content": "d(c_{1}) = SDF(c_{1})" + }, + { + "bbox": [ + 46, + 530, + 289, + 715 + ], + "type": "text", + "content": " and the normal vector of " + }, + { + "bbox": [ + 46, + 530, + 289, + 715 + ], + "type": "inline_equation", + "content": "c_{1}" + }, + { + "bbox": [ + 46, + 530, + 289, + 715 + ], + "type": "text", + "content": " can be calculated by " + }, + { + "bbox": [ + 46, + 530, + 289, + 715 + ], + "type": "inline_equation", + "content": "n(c_{1}) = \\nabla SDF(c_{1})" + }, + { + "bbox": [ + 46, + 530, + 289, + 715 + ], + "type": "text", + "content": ". In practice, obtaining an" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 263, + 547, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 263, + 547, + 346 + ], + "spans": [ + { + "bbox": [ + 304, + 263, + 547, + 346 + ], + "type": "text", + "content": "object's SDF directly is not feasible. Consequently, we employ a 3D grid to represent the object's SDF, where each grid point records the corresponding SDF value. During training, the SDF value at any arbitrary position is computed using tri-linear interpolation, which allows for the differentiable computation of the gradients required by the physical constraints." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 354, + 487, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 354, + 487, + 369 + ], + "spans": [ + { + "bbox": [ + 305, + 354, + 487, + 369 + ], + "type": "text", + "content": "3.3. Contact-Score Joint Optimization" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 373, + 547, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 547, + 481 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 547, + 481 + ], + "type": "text", + "content": "Due to the significant discrepancy in shape and structure between novel and training objects, the grasp prediction from the network can be unstable, which leads to failures. To solve this problem, inspired by dexterous grasp synthesis [12, 37], we introduce the contact map prior which encodes the preference contact region on the object to refine inaccurate grasps. Given the object point-cloud " + }, + { + "bbox": [ + 304, + 373, + 547, + 481 + ], + "type": "inline_equation", + "content": "X_{o}" + }, + { + "bbox": [ + 304, + 373, + 547, + 481 + ], + "type": "text", + "content": " and gripper contacts " + }, + { + "bbox": [ + 304, + 373, + 547, + 481 + ], + "type": "inline_equation", + "content": "c_{1}, c_{2}" + }, + { + "bbox": [ + 304, + 373, + 547, + 481 + ], + "type": "text", + "content": ", we define the contact map of each point " + }, + { + "bbox": [ + 304, + 373, + 547, + 481 + ], + "type": "inline_equation", + "content": "p_{i} \\in X_{o}" + }, + { + "bbox": [ + 304, + 373, + 547, + 481 + ], + "type": "text", + "content": " for a 6-DoF grasp as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 351, + 489, + 545, + 509 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 489, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 351, + 489, + 545, + 509 + ], + "type": "interline_equation", + "content": "D _ {i} = \\min _ {j} \\| p _ {i} - c _ {j} \\|, c _ {j} \\in \\left\\{c _ {1}, c _ {2} \\right\\} \\tag {10}", + "image_path": "034615b51ea3cefd4acbc040d9ffe47169421c57499282c4d577482cc71d1283.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 515, + 547, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 515, + 547, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 515, + 547, + 647 + ], + "type": "text", + "content": "Compared with optimization for a single object with accurate geometry in grasp synthesis, for the 6-DoF grasp detection task in cluttered scenes, there exist two issues that make the test-time optimization impractical and cause singular results: (1) the contact points don't always lie on the object surface; (2) the object point-clouds in real-world scenarios are inaccurate and incomplete. For the first issue, we propose the projection contact distance " + }, + { + "bbox": [ + 304, + 515, + 547, + 647 + ], + "type": "inline_equation", + "content": "PD" + }, + { + "bbox": [ + 304, + 515, + 547, + 647 + ], + "type": "text", + "content": " illustrated in Fig. 3 (a) as a supplementary of original contact distance " + }, + { + "bbox": [ + 304, + 515, + 547, + 647 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 515, + 547, + 647 + ], + "type": "text", + "content": ", which calculates the projection distance from the object point-cloud to the connecting line of two contacts." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 382, + 656, + 545, + 670 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 656, + 545, + 670 + ], + "spans": [ + { + "bbox": [ + 382, + 656, + 545, + 670 + ], + "type": "interline_equation", + "content": "P D _ {i} = D _ {i} * \\sin \\left(\\theta_ {i}\\right) \\tag {11}", + "image_path": "f1072328bd680c41eea20aa5565f8c843906fa58cb7042c75f069d2c1889e6db.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "We visualize the original contact map and the proposed projection contact map in Fig. 3 (b). Rather than computing the contact distance directly, the projection contact" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "18105" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 70, + 180, + 177 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 180, + 177 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 180, + 177 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 180, + 177 + ], + "type": "image", + "image_path": "826c4048089a00cbed616167ff9ac253714145b1a46cda9df6bd412c611998ec.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 113, + 180, + 124, + 189 + ], + "lines": [ + { + "bbox": [ + 113, + 180, + 124, + 189 + ], + "spans": [ + { + "bbox": [ + 113, + 180, + 124, + 189 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 181, + 70, + 276, + 177 + ], + "blocks": [ + { + "bbox": [ + 181, + 70, + 276, + 177 + ], + "lines": [ + { + "bbox": [ + 181, + 70, + 276, + 177 + ], + "spans": [ + { + "bbox": [ + 181, + 70, + 276, + 177 + ], + "type": "image", + "image_path": "11383aa2771d4bb74fe7562a794ee768994b10d8d41338d00924358cd0acdc57.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 222, + 180, + 233, + 189 + ], + "lines": [ + { + "bbox": [ + 222, + 180, + 233, + 189 + ], + "spans": [ + { + "bbox": [ + 222, + 180, + 233, + 189 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 199, + 287, + 221 + ], + "lines": [ + { + "bbox": [ + 46, + 199, + 287, + 221 + ], + "spans": [ + { + "bbox": [ + 46, + 199, + 287, + 221 + ], + "type": "text", + "content": "Figure 3. (a) Calculation of the contact map and (b) comparison of the projection contact map and the original version." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 234, + 287, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 234, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 234, + 287, + 327 + ], + "type": "text", + "content": "map can highlight the accurate contact region on the object point-cloud for coarse grasp poses whose contact points may not lie on the surface. For contact map optimization, as shown in Fig. 4 (a), we employ a ContactNet for contact map prediction, which encodes the contact prior. Given the object point-cloud " + }, + { + "bbox": [ + 46, + 234, + 287, + 327 + ], + "type": "inline_equation", + "content": "X_{o}" + }, + { + "bbox": [ + 46, + 234, + 287, + 327 + ], + "type": "text", + "content": ", the gripper point-cloud " + }, + { + "bbox": [ + 46, + 234, + 287, + 327 + ], + "type": "inline_equation", + "content": "X_{g}" + }, + { + "bbox": [ + 46, + 234, + 287, + 327 + ], + "type": "text", + "content": " and the predicted grasp pose " + }, + { + "bbox": [ + 46, + 234, + 287, + 327 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 234, + 287, + 327 + ], + "type": "text", + "content": ", the ContactNet can be formulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 96, + 329, + 287, + 344 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 329, + 287, + 344 + ], + "spans": [ + { + "bbox": [ + 96, + 329, + 287, + 344 + ], + "type": "interline_equation", + "content": "\\hat {D}, \\hat {P D} = \\operatorname {C o n t a c t N e t} \\left(g X _ {g}, X _ {o}\\right) \\tag {12}", + "image_path": "fa9a83d51c60ddb6a9b2981e8b54671a9d5c16a6cea0b47aed81b83fbf50316d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 348, + 287, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 348, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 47, + 348, + 287, + 361 + ], + "type": "text", + "content": "and thus the contact optimization target " + }, + { + "bbox": [ + 47, + 348, + 287, + 361 + ], + "type": "inline_equation", + "content": "J_{c}" + }, + { + "bbox": [ + 47, + 348, + 287, + 361 + ], + "type": "text", + "content": " is formulated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 97, + 371, + 287, + 385 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 371, + 287, + 385 + ], + "spans": [ + { + "bbox": [ + 97, + 371, + 287, + 385 + ], + "type": "interline_equation", + "content": "J _ {c} = | D - \\hat {D} | + \\alpha * | P D - \\hat {P D} | \\tag {13}", + "image_path": "59bb5358936274ad424ff4e8d58b409bab835d203e53b70ac1fcd9e962e599a2.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 391, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 391, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 391, + 287, + 510 + ], + "type": "text", + "content": "For the second issue, we incorporate an independent grasp score network ScoreNet in the optimization process as shown in Fig. 4 (b). Trained with the noisy object point-cloud and its corresponding grasp pose sampled from grasp labels, the ScoreNet predicts the grasp score " + }, + { + "bbox": [ + 46, + 391, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\hat{S}" + }, + { + "bbox": [ + 46, + 391, + 287, + 510 + ], + "type": "text", + "content": ". We sample both the good and bad quality grasps so the ScoreNet can give a low score when the failed grasp appears in the optimization. To suppress the decline in grasp score, the target of score optimization " + }, + { + "bbox": [ + 46, + 391, + 287, + 510 + ], + "type": "inline_equation", + "content": "J_{s}" + }, + { + "bbox": [ + 46, + 391, + 287, + 510 + ], + "type": "text", + "content": " and the overall optimization target " + }, + { + "bbox": [ + 46, + 391, + 287, + 510 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 46, + 391, + 287, + 510 + ], + "type": "text", + "content": " are formulated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 123, + 514, + 287, + 528 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 514, + 287, + 528 + ], + "spans": [ + { + "bbox": [ + 123, + 514, + 287, + 528 + ], + "type": "interline_equation", + "content": "J _ {s} = T - \\min (T, \\hat {S}) \\tag {14}", + "image_path": "31630b82ce76ce83c68dd0a80a5c7f125eaf593cde17b461d53b38afb31607ae.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 113, + 538, + 287, + 551 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 538, + 287, + 551 + ], + "spans": [ + { + "bbox": [ + 113, + 538, + 287, + 551 + ], + "type": "interline_equation", + "content": "J = J _ {c} + \\beta * J _ {s} + \\gamma * \\Delta t \\tag {15}", + "image_path": "b54eeb19d3ed3dca0d42410ac96008b9cf990704d144a5517b1bae7fb24c14c6.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 556, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 556, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 556, + 287, + 616 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 556, + 287, + 616 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 556, + 287, + 616 + ], + "type": "text", + "content": " is the max grasp score, " + }, + { + "bbox": [ + 46, + 556, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 556, + 287, + 616 + ], + "type": "text", + "content": " is the weight of " + }, + { + "bbox": [ + 46, + 556, + 287, + 616 + ], + "type": "inline_equation", + "content": "J_{s}" + }, + { + "bbox": [ + 46, + 556, + 287, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 556, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 46, + 556, + 287, + 616 + ], + "type": "text", + "content": " constrains the offset of position " + }, + { + "bbox": [ + 46, + 556, + 287, + 616 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 556, + 287, + 616 + ], + "type": "text", + "content": " during optimization. During inference, we adopt an Adam optimizer to optimize the grasp pose iteratively by minimizing " + }, + { + "bbox": [ + 46, + 556, + 287, + 616 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 46, + 556, + 287, + 616 + ], + "type": "text", + "content": " in a gradient descent strategy." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 627, + 128, + 640 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 627, + 128, + 640 + ], + "spans": [ + { + "bbox": [ + 47, + 627, + 128, + 640 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 647, + 180, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 180, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 180, + 659 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "Benchmark We conduct all the simulation experiments on the large-scale GraspNet-1billion benchmark [10]. The benchmark includes 190 cluttered scenes, in which 100 scenes for training, 90 for testing. The testing set is divided" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 322, + 75, + 533, + 244 + ], + "blocks": [ + { + "bbox": [ + 322, + 75, + 533, + 244 + ], + "lines": [ + { + "bbox": [ + 322, + 75, + 533, + 244 + ], + "spans": [ + { + "bbox": [ + 322, + 75, + 533, + 244 + ], + "type": "image", + "image_path": "a6f203a872ee741fcaf0e40b5dfd4b5f52d7d00cf8bf4ad71d02c68815ec3769.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 249, + 466, + 258 + ], + "lines": [ + { + "bbox": [ + 386, + 249, + 466, + 258 + ], + "spans": [ + { + "bbox": [ + 386, + 249, + 466, + 258 + ], + "type": "text", + "content": "(a) Contact optimization" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 331, + 264, + 520, + 381 + ], + "blocks": [ + { + "bbox": [ + 331, + 264, + 520, + 381 + ], + "lines": [ + { + "bbox": [ + 331, + 264, + 520, + 381 + ], + "spans": [ + { + "bbox": [ + 331, + 264, + 520, + 381 + ], + "type": "image", + "image_path": "30dfdef9513e5b802efa1489da34d042ed663d7a86c4a0e9ea03ae862774f43d.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 389, + 385, + 463, + 395 + ], + "lines": [ + { + "bbox": [ + 389, + 385, + 463, + 395 + ], + "spans": [ + { + "bbox": [ + 389, + 385, + 463, + 395 + ], + "type": "text", + "content": "(b) Score optimization" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 404, + 545, + 426 + ], + "lines": [ + { + "bbox": [ + 305, + 404, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 305, + 404, + 545, + 426 + ], + "type": "text", + "content": "Figure 4. The pipeline of contact optimization and score optimization with ContactNet and ScoreNet." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 437, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 616 + ], + "type": "text", + "content": "into seen, similar and novel set based on the objects in the scenes. Each scene includes 256 RGB-D images captured from different views with RealSense/Kinect cameras. Most previous methods on the benchmark utilize single-view data for training and evaluation. However, the depth images captured from a single view suffer from significant noise. For novel object grasping, as the model has no prior about the objects, it is unable to infer the shape of occluded parts based solely on partial point-clouds, thereby hindering the grasp detection for novel objects. To bypass the interference caused by incomplete point clouds and enhance the capability of grasp detection on novel objects, we reconstruct the depth images from multiple views in a single scene into Truncated Signed Distance Function (TSDF) via KinectFusion [25]. The TSDFs are used for training and evaluation." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "type": "text", + "content": "Metric We follow the metric used in GraspNet-1billion benchmark [10], in which the average precision of top-" + }, + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "type": "text", + "content": " ranked grasps in a scene is considered. In the original metric, " + }, + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "type": "text", + "content": " is set to 50, and a maximum of 10 grasps per-object are used for evaluation. Given that most scenes contain around 9 objects, this metric can overlook the grasping accuracy of some objects. Therefore, we employ an object-balanced metric for reconstructed scenes, where " + }, + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "type": "inline_equation", + "content": "k = N_{\\text{object}} * 5" + }, + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18106" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 112, + 70, + 482, + 177 + ], + "blocks": [ + { + "bbox": [ + 112, + 70, + 482, + 177 + ], + "lines": [ + { + "bbox": [ + 112, + 70, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 112, + 70, + 482, + 177 + ], + "type": "table", + "html": "
ModelSeenSimilarNovel
AP\\(AP_{0.8}\\)\\(AP_{0.4}\\)AP\\(AP_{0.8}\\)\\(AP_{0.4}\\)AP\\(AP_{0.8}\\)\\(AP_{0.4}\\)
Baseline66.0975.5760.6364.8274.1061.5130.6137.6117.06
+ PCR66.6775.6761.8465.5574.4261.4735.5843.4919.70
w/o \\(R^A\\)65.6374.5961.1965.3874.2962.1032.3839.0418.88
w/o \\(R^C\\)65.2775.0558.6262.2571.6557.1132.3438.5819.27
w/o \\(R^S\\)63.5172.3760.2664.3173.0061.0635.5742.6020.45
", + "image_path": "df1474e65f2b9b92b5786ef57bb6d65947fc089c19ac8d0a6a909cbe4fc9cd46.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 217, + 289, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 217, + 289, + 458 + ], + "spans": [ + { + "bbox": [ + 47, + 217, + 289, + 458 + ], + "type": "inline_equation", + "content": "N_{\\text{object}}" + }, + { + "bbox": [ + 47, + 217, + 289, + 458 + ], + "type": "text", + "content": " is the number of objects of the scene and a maximum of 5 grasps per-object are used for evaluation. This provides a more balanced consideration of the grasping accuracy across different objects compared to the original evaluation metrics. " + }, + { + "bbox": [ + 47, + 217, + 289, + 458 + ], + "type": "inline_equation", + "content": "\\mathbf{AP}_{\\mu}" + }, + { + "bbox": [ + 47, + 217, + 289, + 458 + ], + "type": "text", + "content": " is employed as the metric, which represents the average Precision@m for " + }, + { + "bbox": [ + 47, + 217, + 289, + 458 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 47, + 217, + 289, + 458 + ], + "type": "text", + "content": " ranges from 1 to " + }, + { + "bbox": [ + 47, + 217, + 289, + 458 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 47, + 217, + 289, + 458 + ], + "type": "text", + "content": " with friction " + }, + { + "bbox": [ + 47, + 217, + 289, + 458 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 47, + 217, + 289, + 458 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 217, + 289, + 458 + ], + "type": "inline_equation", + "content": "\\mathbf{AP}" + }, + { + "bbox": [ + 47, + 217, + 289, + 458 + ], + "type": "text", + "content": " is calculated by the average of " + }, + { + "bbox": [ + 47, + 217, + 289, + 458 + ], + "type": "inline_equation", + "content": "\\mathbf{AP}_{\\mu}" + }, + { + "bbox": [ + 47, + 217, + 289, + 458 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 217, + 289, + 458 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 47, + 217, + 289, + 458 + ], + "type": "text", + "content": " ranges from 0.2 to 1.2 with the interval " + }, + { + "bbox": [ + 47, + 217, + 289, + 458 + ], + "type": "inline_equation", + "content": "\\Delta \\mu = 0.2" + }, + { + "bbox": [ + 47, + 217, + 289, + 458 + ], + "type": "text", + "content": ". Model implementations We first introduce a strong baseline model based on the scale balanced 6-DoF grasp detection network [19] and GSNet [35]. Following [35], we employ a sparse UNet based on Minkowski Engine [7] as the backbone and calculate the \"graspness\" for each point to generate grasp candidates. For local feature extraction, the multi-scale cylinder grouping proposed in [19] is employed to improve the performance of the baseline model. For the PCR during training, " + }, + { + "bbox": [ + 47, + 217, + 289, + 458 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 47, + 217, + 289, + 458 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 217, + 289, + 458 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 47, + 217, + 289, + 458 + ], + "type": "text", + "content": " which control the contact distance are set to 0.02 and 0.005 separately and the weight " + }, + { + "bbox": [ + 47, + 217, + 289, + 458 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 47, + 217, + 289, + 458 + ], + "type": "text", + "content": " for the regularization is set to 0.1. For the C-SJO, we set the hyper-parameter " + }, + { + "bbox": [ + 47, + 217, + 289, + 458 + ], + "type": "inline_equation", + "content": "\\alpha = 0.2" + }, + { + "bbox": [ + 47, + 217, + 289, + 458 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 217, + 289, + 458 + ], + "type": "inline_equation", + "content": "\\beta = 0.01" + }, + { + "bbox": [ + 47, + 217, + 289, + 458 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 217, + 289, + 458 + ], + "type": "inline_equation", + "content": "\\gamma = 5" + }, + { + "bbox": [ + 47, + 217, + 289, + 458 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 464, + 285, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 464, + 285, + 478 + ], + "spans": [ + { + "bbox": [ + 47, + 464, + 285, + 478 + ], + "type": "text", + "content": "4.2. Results on Physical Constraint Regularization" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 483, + 288, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 288, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 288, + 664 + ], + "type": "text", + "content": "We show the results of PCR in Table 1. Compared to the baseline model, integrating the PCR delivers an improvement of " + }, + { + "bbox": [ + 46, + 483, + 288, + 664 + ], + "type": "inline_equation", + "content": "4.64\\%" + }, + { + "bbox": [ + 46, + 483, + 288, + 664 + ], + "type": "text", + "content": " on the novel set, demonstrating its effectiveness for objects with diverse shapes and structures. Besides, the PCR also improves by " + }, + { + "bbox": [ + 46, + 483, + 288, + 664 + ], + "type": "inline_equation", + "content": "0.58\\%" + }, + { + "bbox": [ + 46, + 483, + 288, + 664 + ], + "type": "text", + "content": " in seen set and " + }, + { + "bbox": [ + 46, + 483, + 288, + 664 + ], + "type": "inline_equation", + "content": "0.73\\%" + }, + { + "bbox": [ + 46, + 483, + 288, + 664 + ], + "type": "text", + "content": " in similar set. This illustrates that for seen and similar objects in the training set, introducing physical prior can help the grasp detection network to fully utilize the grasp labels compared to fitting them directly. We also ablate the influence of different physical constraints and notice that the incomplete constraint conditions can attribute to the performance decline on seen and similar set. Only with partial constraints during training, not all the grasp poses which meet the constraints are correct, thus disturbing the learning from grasping label." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 670, + 279, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 279, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 279, + 685 + ], + "type": "text", + "content": "4.3. Results on Contact-Score Joint Optimization" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "type": "text", + "content": "We employ the C-SJO in our 6-DoF grasp detection network trained with PCR and conduct ablations. The results" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 314, + 215, + 539, + 320 + ], + "blocks": [ + { + "bbox": [ + 189, + 184, + 402, + 195 + ], + "lines": [ + { + "bbox": [ + 189, + 184, + 402, + 195 + ], + "spans": [ + { + "bbox": [ + 189, + 184, + 402, + 195 + ], + "type": "text", + "content": "Table 1. Results of PCR on scenes captured by RealSense." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 314, + 215, + 539, + 320 + ], + "lines": [ + { + "bbox": [ + 314, + 215, + 539, + 320 + ], + "spans": [ + { + "bbox": [ + 314, + 215, + 539, + 320 + ], + "type": "table", + "html": "
ModelSeenSimilarNovel
No Refine66.6765.5535.58
+ C-S (Predicted Mask)66.6165.3736.67
+ C-S (GT Mask)66.6965.5036.61
Original C-S66.4865.2936.27
w/o C66.6565.6435.86
w/o S66.5465.3736.23
", + "image_path": "671cb5e04808572ea50d0a87c3e72f2a6a034de1833c745b791e9b1304bec377.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 315, + 329, + 536, + 340 + ], + "lines": [ + { + "bbox": [ + 315, + 329, + 536, + 340 + ], + "spans": [ + { + "bbox": [ + 315, + 329, + 536, + 340 + ], + "type": "text", + "content": "Table 2. Results of C-SJO on scenes captured by RealSense." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 353, + 545, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 353, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 353, + 545, + 664 + ], + "type": "text", + "content": "are shown in Table 2. We incorporate a recent proposed 3D segmentation method [38] to get object point-clouds from the clutter for optimization. With the ground-truth mask and segmented mask, C-SJO can improve by " + }, + { + "bbox": [ + 304, + 353, + 545, + 664 + ], + "type": "inline_equation", + "content": "1.03\\%" + }, + { + "bbox": [ + 304, + 353, + 545, + 664 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 353, + 545, + 664 + ], + "type": "inline_equation", + "content": "1.00\\%" + }, + { + "bbox": [ + 304, + 353, + 545, + 664 + ], + "type": "text", + "content": " on novel set separately. The C-SJO has almost no impact on the seen and similar set, primarily because the grasp poses predicted for seen and similar objects are sufficiently stable, making it challenging to find a better grasp than the original one. However, for the novel set, many unstable grasps exist that can potentially be optimized. We also conduct ablations about different designs of the C-SJO. Without the proposed projection contact map, only employing the original contact map (Original C-S) drops on seen, similar and novel set due to the singularity poses during optimization. Optimizing solely based on grasp score " + }, + { + "bbox": [ + 304, + 353, + 545, + 664 + ], + "type": "inline_equation", + "content": "(w / oC)" + }, + { + "bbox": [ + 304, + 353, + 545, + 664 + ], + "type": "text", + "content": " overlooks the contact map prior, resulting in very limited improvement on the novel set. Using only the contact map for optimization " + }, + { + "bbox": [ + 304, + 353, + 545, + 664 + ], + "type": "inline_equation", + "content": "(w / oS)" + }, + { + "bbox": [ + 304, + 353, + 545, + 664 + ], + "type": "text", + "content": " yields a " + }, + { + "bbox": [ + 304, + 353, + 545, + 664 + ], + "type": "inline_equation", + "content": "0.38\\%" + }, + { + "bbox": [ + 304, + 353, + 545, + 664 + ], + "type": "text", + "content": " drop on the novel set, but the absence of grasp scores restricts the search space, making it challenging to locate the locally optimal grasp during the optimization process. The process of grasp pose optimization is shown in Fig. 5, where the optimal grasp poses are reached in the original grasps' neighbor " + }, + { + "bbox": [ + 304, + 353, + 545, + 664 + ], + "type": "inline_equation", + "content": "SE(3)" + }, + { + "bbox": [ + 304, + 353, + 545, + 664 + ], + "type": "text", + "content": " space. For object (1)-(4), the optimization process refines the inaccurate initial grasp poses. For objects (5) and (6), the optimization process facilitates collision avoidance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 671, + 486, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 486, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 486, + 684 + ], + "type": "text", + "content": "4.4. Comparison with State-of-the-art" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "To make a fair comparison with the state-of-the-art methods of the GraspNet-1billion benchmark, we re-implement" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18107" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 77, + 70, + 518, + 177 + ], + "blocks": [ + { + "bbox": [ + 77, + 70, + 518, + 177 + ], + "lines": [ + { + "bbox": [ + 77, + 70, + 518, + 177 + ], + "spans": [ + { + "bbox": [ + 77, + 70, + 518, + 177 + ], + "type": "table", + "html": "
ModelSeenSimilarNovel
AP\\(AP_{0.8}\\)\\(AP_{0.4}\\)AP\\(AP_{0.8}\\)\\(AP_{0.4}\\)AP\\(AP_{0.8}\\)\\(AP_{0.4}\\)
GraspNet-baseline [10]28.1030.5326.5623.8726.9222.517.488.434.93
Scale-balanced Grasp [19]46.0551.0244.2737.7644.2734.7517.0921.0410.20
GSNet [35]60.4771.0552.0658.5569.3853.0828.0636.1914.09
Ours Baseline66.0975.5760.6364.8274.1061.5130.6137.6117.06
Ours66.6175.6761.5265.3774.3561.2836.6745.0820.90
", + "image_path": "f943775637a3f63af03245e381fb94e1dcd4097a45233eac758ce49fcd7a553f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 86, + 208, + 509, + 331 + ], + "blocks": [ + { + "bbox": [ + 86, + 208, + 509, + 331 + ], + "lines": [ + { + "bbox": [ + 86, + 208, + 509, + 331 + ], + "spans": [ + { + "bbox": [ + 86, + 208, + 509, + 331 + ], + "type": "image", + "image_path": "5323dc25d95ebfc5c9c75e0c884d2b1cd27309bde9cf921971cccc02b44c8146.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 208, + 338, + 384, + 350 + ], + "lines": [ + { + "bbox": [ + 208, + 338, + 384, + 350 + ], + "spans": [ + { + "bbox": [ + 208, + 338, + 384, + 350 + ], + "type": "text", + "content": "Figure 5. Visualization of the process of C-SJO." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 365, + 287, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 287, + 437 + ], + "type": "text", + "content": "three representative 6-DoF grasp detection methods [10, 19, 35] with our multi-view reconstructed scene as input. As shown in Table 3, our method performs better than previous 6-DoF grasp detection methods on all test sets and achieves " + }, + { + "bbox": [ + 46, + 365, + 287, + 437 + ], + "type": "inline_equation", + "content": "36.67\\%" + }, + { + "bbox": [ + 46, + 365, + 287, + 437 + ], + "type": "text", + "content": " on the novel set, which demonstrates the generalization capability of our method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 444, + 250, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 444, + 250, + 456 + ], + "spans": [ + { + "bbox": [ + 47, + 444, + 250, + 456 + ], + "type": "text", + "content": "4.5. Comparison with object augmentation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 462, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 288, + 714 + ], + "type": "text", + "content": "We give a comparison between our prior knowledge based method and the object augmentation which employed by previous methods to enhance the generalization capability. To migrate the procedural or learning-based object augmentation method on the GraspNet-1billion benchmark, since the similar objects and the training objects have similar distributions, we consider the similar set in the benchmark as objects obtained through the augmentation. We merge the two sets for joint training. In this way, we can conveniently validate the performance of data augmentation on this benchmark without the need to generate additional objects and annotate grasps. As shown in Table 4, introducing object augmentation can slightly improve the performance on seen and novel sets, benefiting from the richer object distribution. However, our method performs better on the novel set, improving by " + }, + { + "bbox": [ + 46, + 462, + 288, + 714 + ], + "type": "inline_equation", + "content": "4.37\\%" + }, + { + "bbox": [ + 46, + 462, + 288, + 714 + ], + "type": "text", + "content": " compared to the object augmentation method, without introducing any additional data. Simultaneously, our method also exhibits an increase of " + }, + { + "bbox": [ + 46, + 462, + 288, + 714 + ], + "type": "inline_equation", + "content": "0.33\\%" + }, + { + "bbox": [ + 46, + 462, + 288, + 714 + ], + "type": "text", + "content": " on novel set when paired with object augmentation, demonstrating that our method and object augmentation can be used together." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 318, + 363, + 535, + 440 + ], + "blocks": [ + { + "bbox": [ + 100, + 184, + 492, + 196 + ], + "lines": [ + { + "bbox": [ + 100, + 184, + 492, + 196 + ], + "spans": [ + { + "bbox": [ + 100, + 184, + 492, + 196 + ], + "type": "text", + "content": "Table 3. Comparison with the state-of-the-art methods on RealSense scenes of GraspNet-billion benchmark." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 363, + 535, + 440 + ], + "lines": [ + { + "bbox": [ + 318, + 363, + 535, + 440 + ], + "spans": [ + { + "bbox": [ + 318, + 363, + 535, + 440 + ], + "type": "table", + "html": "
ModelSeenSimilarNovel
Baseline66.0964.8230.61
Augmentation68.12-32.30
Ours66.6165.3736.67
Ours + Augmentation68.31-37.00
", + "image_path": "6e4ac0502389fd58ba5375a3a7d31a89182027e9996097576357f341d74f84e1.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 338, + 448, + 512, + 459 + ], + "lines": [ + { + "bbox": [ + 338, + 448, + 512, + 459 + ], + "spans": [ + { + "bbox": [ + 338, + 448, + 512, + 459 + ], + "type": "text", + "content": "Table 4. Comparison with object augmentation." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 471, + 422, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 471, + 422, + 483 + ], + "spans": [ + { + "bbox": [ + 306, + 471, + 422, + 483 + ], + "type": "text", + "content": "4.6. Result Visualization" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 491, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 491, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 491, + 545, + 624 + ], + "type": "text", + "content": "In figure 6, we visualized the results generated by the baseline method, object augmentation and our approach on the GraspNet-1billion benchmark. The gripper poses in red represent successful grasps, while those shown in purple and blue correspond to collision and bad grasps. By incorporating domain prior knowledge, our method can generate grasp poses that conform to the physical relationship between the gripper and the object. For objects (3)-(5), our method predicts some failed grasps, primarily due to poor grasp sampling locations or inaccurate perception of the object's geometry." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 634, + 434, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 634, + 434, + 645 + ], + "spans": [ + { + "bbox": [ + 306, + 634, + 434, + 645 + ], + "type": "text", + "content": "4.7. Real-world Evaluation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "type": "text", + "content": "The effectiveness of the proposed method in real-world is validated by real robot in this section. As shown in Figure 7 (a), the robotic grasping system is built on a 6-DoF UR-10 robot arm and a RealSense D435i depth camera is employed for scene perception. The total of 20 objects used for grasp" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18108" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 86, + 73, + 512, + 331 + ], + "blocks": [ + { + "bbox": [ + 86, + 73, + 512, + 331 + ], + "lines": [ + { + "bbox": [ + 86, + 73, + 512, + 331 + ], + "spans": [ + { + "bbox": [ + 86, + 73, + 512, + 331 + ], + "type": "image", + "image_path": "866ec070b7c5e5f297b12d4dd96c75b2ce4b6cec7998958a1d105e384f2d11a9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 65, + 338, + 526, + 350 + ], + "lines": [ + { + "bbox": [ + 65, + 338, + 526, + 350 + ], + "spans": [ + { + "bbox": [ + 65, + 338, + 526, + 350 + ], + "type": "text", + "content": "Figure 6. Visualization of the predicted grasp poses from baseline, object augmentation and the methods proposed in this paper." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 55, + 363, + 159, + 456 + ], + "blocks": [ + { + "bbox": [ + 55, + 363, + 159, + 456 + ], + "lines": [ + { + "bbox": [ + 55, + 363, + 159, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 363, + 159, + 456 + ], + "type": "image", + "image_path": "937e7614cedf02197c3d3b380630379f10f8f4cdc9ab06cde4a742f13a403aaa.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 464, + 288, + 487 + ], + "lines": [ + { + "bbox": [ + 47, + 464, + 288, + 487 + ], + "spans": [ + { + "bbox": [ + 47, + 464, + 288, + 487 + ], + "type": "text", + "content": "Figure 7. (a) Robotic grasping system and (b) objects for real-world grasping." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 160, + 363, + 279, + 456 + ], + "blocks": [ + { + "bbox": [ + 160, + 363, + 279, + 456 + ], + "lines": [ + { + "bbox": [ + 160, + 363, + 279, + 456 + ], + "spans": [ + { + "bbox": [ + 160, + 363, + 279, + 456 + ], + "type": "image", + "image_path": "457b9240a339d102dfc7fe2b1a327bd3e57fd19197a21ef7091c94e79d6ae3c2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 497, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 497, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 497, + 287, + 581 + ], + "type": "text", + "content": "ing (Fig. 7 (b)) are composed of two parts: 13 3D-printed objects from Dex-Net [20] and 7 objects chosen from the YCB dataset [4]. All of the objects are significantly different from the objects used for training in GraspNet-1billion. Before executing the proposed grasp detection method, we quickly reconstruct the scene using an arm-mounted depth camera based on the KinectFusion [25]." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 52, + 590, + 282, + 640 + ], + "blocks": [ + { + "bbox": [ + 52, + 590, + 282, + 640 + ], + "lines": [ + { + "bbox": [ + 52, + 590, + 282, + 640 + ], + "spans": [ + { + "bbox": [ + 52, + 590, + 282, + 640 + ], + "type": "table", + "html": "
ModelIsolatedCluttered
SR (%)SR (%)SCR (%)
Baseline58.33 (35/60)48.21 (43/89)86.00 (43/50)
Ours68.33 (41/60)64.86 (48/74)96.00 (48/50)
", + "image_path": "a20f178ce436acbf0f9cb01dee167248e57ec7fa3ee283e1bb3cafc7320fd7f0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 64, + 648, + 269, + 660 + ], + "lines": [ + { + "bbox": [ + 64, + 648, + 269, + 660 + ], + "spans": [ + { + "bbox": [ + 64, + 648, + 269, + 660 + ], + "type": "text", + "content": "Table 5. Results of the real-world grasping experiments." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 665, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 288, + 713 + ], + "type": "text", + "content": "We compare our model to the baseline in two settings: isolated object grasping and cluttered object grasping. For isolated object grasping, each object is placed in three different poses and Success Rate (SR) is used as the metric." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 365, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 365, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 305, + 365, + 545, + 449 + ], + "type": "text", + "content": "For cluttered object grasping, we compose 5 objects into a scene and make the robot remove them all with a maximum number of operations at 10. SR and Scene Completion Rate (SCR) are employed as the metrics. As illustrated in Table 5, our model outperforms the baseline in both settings and achieves better SR and SCR, demonstrating its superiority for novel objects." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 459, + 378, + 471 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 459, + 378, + 471 + ], + "spans": [ + { + "bbox": [ + 306, + 459, + 378, + 471 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 479, + 545, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 545, + 611 + ], + "type": "text", + "content": "In this paper, we work for generalizing 6-DoF grasp detection with domain prior knowledge of robotic grasping. The physical constraint regularization based on physical rules is proposed to enable the generalization of objects with largely varied shapes and structures. To refine the unstable results predicted by network in cluttered scenarios, we specially design a contact-score joint optimization with the contact map prior, in which a projection contact map is utilized. Extensive experiments on both the benchmark and the real-world robot demonstrate the effectiveness of our method for novel objects." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 621, + 399, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 621, + 399, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 621, + 399, + 635 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "type": "text", + "content": "This work is partly supported by the National Key R&D Program of China (2022ZD0161902), the National Natural Science Foundation of China (62022011), the Research Program of State Key Laboratory of Software Development Environment (SKLSDE-2023ZX-14), and the Fundamental Research Funds for the Central Universities." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18109" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "text", + "content": "[1] Konstantinos Bousmalis, Alex Irpan, Paul Wohlhart, Yunfei Bai, Matthew Kelcey, Mrinal Kalakrishnan, Laura Downs, Julian Ibarz, Peter Pastor, Kurt Konolige, et al. Using simulation and domain adaptation to improve efficiency of deep robotic grasping. In IEEE International Conference on Robotics and Automation, 2018. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 158, + 288, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 288, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 288, + 203 + ], + "type": "text", + "content": "[2] Samarth Brahmbhatt, Cusuh Ham, Charles C. Kemp, and James Hays. Contactdb: Analyzing and predicting grasp contact via thermal imaging. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 204, + 288, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 204, + 288, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 204, + 288, + 247 + ], + "type": "text", + "content": "[3] Michel Breyer, Jen Jen Chung, Lionel Ott, Roland Siegwart, and Juan I. Nieto. Volumetric grasping network: Real-time 6 DOF grasp detection in clutter. In Conference on Robot Learning, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 249, + 288, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 249, + 288, + 303 + ], + "spans": [ + { + "bbox": [ + 53, + 249, + 288, + 303 + ], + "type": "text", + "content": "[4] Berk Çallı, Arjun Singh, James Bruce, Aaron Walsman, Kurt Konolige, Siddhartha S. Srinivasa, Pieter Abbeel, and Aaron M. Dollar. Yale-cmu-berkeley dataset for robotic manipulation research. International Journal of Robotics Research, 36(3):261-268, 2017. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 304, + 288, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 304, + 288, + 338 + ], + "spans": [ + { + "bbox": [ + 53, + 304, + 288, + 338 + ], + "type": "text", + "content": "[5] I-Ming Chen and Joel W. Burdick. Finding antipodal point grasps on irregularly shaped objects. IEEE Transactions on Robotics and Automation, 9(4):507-512, 1993. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 339, + 288, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 339, + 288, + 372 + ], + "spans": [ + { + "bbox": [ + 53, + 339, + 288, + 372 + ], + "type": "text", + "content": "[6] I-Ming Chen and Joel W. Burdick. Finding antipodal point grasps on irregularly shaped objects. IEEE Transactions on Robotics and Automation, 9(4):507-512, 1993. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 373, + 288, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 373, + 288, + 418 + ], + "spans": [ + { + "bbox": [ + 53, + 373, + 288, + 418 + ], + "type": "text", + "content": "[7] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 419, + 288, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 419, + 288, + 474 + ], + "spans": [ + { + "bbox": [ + 53, + 419, + 288, + 474 + ], + "type": "text", + "content": "[8] Qiyu Dai, Yan Zhu, Yiran Geng, Ciyu Ruan, Jiazhao Zhang, and He Wang. Graspnerf: Multiview-based 6-dof grasp detection for transparent and specular objects using generalizable nerf. In IEEE International Conference on Robotics and Automation, 2023. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 475, + 288, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 475, + 288, + 520 + ], + "spans": [ + { + "bbox": [ + 53, + 475, + 288, + 520 + ], + "type": "text", + "content": "[9] Amaury Depierre, Emmanuel Dellandrea, and Liming Chen. Jacquard: A large scale dataset for robotic grasp detection. In IEEE/RSJ International Conference on Intelligent Robots and Systems, 2018. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 521, + 287, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 521, + 287, + 565 + ], + "spans": [ + { + "bbox": [ + 48, + 521, + 287, + 565 + ], + "type": "text", + "content": "[10] Haoshu Fang, Chenxi Wang, Minghao Gou, and Cewu Lu. Graspnet-1billion: A large-scale benchmark for general object grasping. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2, 5, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 566, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 566, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 566, + 287, + 599 + ], + "type": "text", + "content": "[11] Carlo Ferrari and John F. Canny. Planning optimal grasps. In IEEE International Conference on Robotics and Automation, 1992. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 600, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 600, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 600, + 287, + 644 + ], + "type": "text", + "content": "[12] Hanwen Jiang, Shaowei Liu, Jiashun Wang, and Xiaolong Wang. Hand-object contact consistency reasoning for human grasps generation. In IEEE/CVF International Conference on Computer Vision, 2021. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "type": "text", + "content": "[13] Junnan Jiang, Xiaohui Xiao, Fei Chen, and Miao Li. Learning grasp ability enhancement through deep shape generation. In International Conference on Intelligent Robotics and Applications, 2022. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "text", + "content": "[14] Zhenyu Jiang, Yifeng Zhu, Maxwell Svetlik, Kuan Fang, and Yuke Zhu. Synergies between affordance and geometry: 6-" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "dof grasp detection via implicit representations. In Robotics: Science and Systems, 2021. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 95, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 139 + ], + "type": "text", + "content": "[15] George Em Karniadakis, Ioannis G Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, and Liu Yang. Physics-informed machine learning. Nature Reviews Physics, 3(6):422-440, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 140, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 545, + 194 + ], + "type": "text", + "content": "[16] Hongzhuo Liang, Xiaojian Ma, Shuang Li, Michael Gorner, Song Tang, Bin Fang, Fuchun Sun, and Jianwei Zhang. Pointnetgpd: Detecting grasp configurations from point sets. In IEEE International Conference on Robotics and Automation, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 195, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 545, + 228 + ], + "type": "text", + "content": "[17] Min Liu, Zherong Pan, Kai Xu, Kanishka Ganguly, and Dinesh Manocha. Deep differentiable grasp planner for high-dof grippers. In Robotics: Science and Systems, 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 228, + 545, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 281 + ], + "type": "text", + "content": "[18] Tengyu Liu, Zeyu Liu, Ziyuan Jiao, Yixin Zhu, and Song-Chun Zhu. Synthesizing diverse and physically stable grasps with arbitrary hand structures using differentiable force closure estimator. IEEE Robotics and Automation Letters, 7(1): 470-477, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 283, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 283, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 308, + 283, + 545, + 316 + ], + "type": "text", + "content": "[19] Haoxiang Ma and Di Huang. Towards scale balanced 6-dof grasp detection in cluttered scenes. In Conference on Robot Learning, 2022. 2, 3, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 316, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 370 + ], + "type": "text", + "content": "[20] Jeffrey Mahler, Jacky Liang, Sherdil Niyaz, Michael Laskey, Richard Doan, Xinyu Liu, Juan Aparicio Ojea, and Ken Goldberg. Dex-net 2.0: Deep learning to plan robust grasps with synthetic point clouds and analytic grasp metrics. In Robotics: Science and Systems, 2017. 1, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 371, + 545, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 371, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 308, + 371, + 545, + 404 + ], + "type": "text", + "content": "[21] Andrew T. Miller and Peter K. Allen. Graspit! A versatile simulator for robotic grasping. IEEE Robotics & Automation Magazine, 11(4):110-122, 2004. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 404, + 545, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 447 + ], + "type": "text", + "content": "[22] Douglas Morrison, Juxi Leitner, and Peter Corke. Closing the loop for robotic grasping: A real-time, generative grasp synthesis approach. In Robotics: Science and Systems, 2018. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 449, + 545, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 492 + ], + "type": "text", + "content": "[23] Douglas Morrison, Peter Corke, and Jürgen Leitner. Egad! an evolved grasping analysis dataset for diversity and reproducibility in robotic manipulation. IEEE Robotics and Automation Letters, 5(3):4368-4375, 2020. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 492, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 492, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 492, + 545, + 536 + ], + "type": "text", + "content": "[24] Arsalan Mousavian, Clemens Eppner, and Dieter Fox. 6-dof grapnet: Variational grasp generation for object manipulation. In IEEE/CVF International Conference on Computer Vision, 2019. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 537, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 545, + 602 + ], + "type": "text", + "content": "[25] Richard A. Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J. Davison, Pushmeet Kohli, Jamie Shotton, Steve Hodges, and Andrew W. Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. In IEEE International Symposium on Mixed and Augmented Reality, 2011. 5, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 603, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 603, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 603, + 545, + 635 + ], + "type": "text", + "content": "[26] Van-Duc Nguyen. Constructing force-closure grasps. International Journal of Robotics Research, 7(3):3-16, 1988. 2, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 635, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 669 + ], + "type": "text", + "content": "[27] Nancy S. Pollard. Closure and quality equivalence for efficient synthesis of grasps from examples. International Journal of Robotics Research, 23(6):595-613, 2004. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "text", + "content": "[28] Ran Qin, Haoxiang Ma, Boyang Gao, and Di Huang. RGB-D grasp detection via depth guided learning with cross-modal attention. In IEEE International Conference on Robotics and Automation, 2023. 2" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18110" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 584 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[29] Yuzhe Qin, Rui Chen, Hao Zhu, Meng Song, Jing Xu, and Hao Su. S4G: amodal single-view single-shot SE(3) grasp detection in cluttered scenes. In Conference on Robot Learning, 2019. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 286, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 286, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 286, + 150 + ], + "type": "text", + "content": "[30] Maximo A. Roa and Raul Suárez. Computation of independent contact regions for grasping 3-d objects. IEEE Transactions on Robotics, 25(4):839-850, 2009. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 152, + 286, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 152, + 286, + 184 + ], + "spans": [ + { + "bbox": [ + 49, + 152, + 286, + 184 + ], + "type": "text", + "content": "[31] Alberto Rodriguez, Matthew T. Mason, and Steve Ferry. From caging to grasping. International Journal of Robotics Research, 31(7):886-900, 2012. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 186, + 286, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 186, + 286, + 228 + ], + "spans": [ + { + "bbox": [ + 49, + 186, + 286, + 228 + ], + "type": "text", + "content": "[32] Andreas ten Pas, Marcus Gualtieri, Kate Saenko, and Robert Platt Jr. Grasp pose detection in point clouds. International Journal of Robotics Research, 36(13-14):1455-1473, 2017. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 230, + 286, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 230, + 286, + 295 + ], + "spans": [ + { + "bbox": [ + 49, + 230, + 286, + 295 + ], + "type": "text", + "content": "[33] Josh Tobin, Lukas Biewald, Rocky Duan, Marcin Andrychowicz, Ankur Handa, Vikash Kumar, Bob McGrew, Alex Ray, Jonas Schneider, Peter Welinder, et al. Domain randomization and generative models for robotic grasping. In IEEE/RSJ International Conference on Intelligent Robots and Systems, 2018. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 297, + 286, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 297, + 286, + 372 + ], + "spans": [ + { + "bbox": [ + 49, + 297, + 286, + 372 + ], + "type": "text", + "content": "[34] Laura Von Rueden, Sebastian Mayer, Katharina Beckh, Bogdan Georgiev, Sven Giesselbach, Raoul Heese, Birgit Kirsch, Julius Pfrommer, Annika Pick, Rajkumar Ramamurthy, et al. Informed machine learning-a taxonomy and survey of integrating prior knowledge into learning systems. IEEE Transactions on Knowledge and Data Engineering, 35(1):614-633, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 374, + 286, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 374, + 286, + 418 + ], + "spans": [ + { + "bbox": [ + 49, + 374, + 286, + 418 + ], + "type": "text", + "content": "[35] Chenxi Wang, Haoshu Fang, Minghao Gou, Hongjie Fang, Jin Gao, and Cewu Lu. Graspness discovery in clutters for fast and accurate grasp detection. In IEEE/CVF International Conference on Computer Vision, 2021. 2, 3, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 420, + 286, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 420, + 286, + 474 + ], + "spans": [ + { + "bbox": [ + 49, + 420, + 286, + 474 + ], + "type": "text", + "content": "[36] David Wang, David Tseng, Pusong Li, Yiding Jiang, Menglong Guo, Michael Danielczuk, Jeffrey Mahler, Jeffrey Ichnowski, and Ken Goldberg. Adversarial grasp objects. In IEEE International Conference on Automation Science and Engineering, 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 475, + 286, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 475, + 286, + 552 + ], + "spans": [ + { + "bbox": [ + 49, + 475, + 286, + 552 + ], + "type": "text", + "content": "[37] Yinzhen Xu, Weikang Wan, Jialiang Zhang, Haoran Liu, Zikang Shan, Hao Shen, Ruicheng Wang, Haoran Geng, Yi-jia Weng, Jiayi Chen, Tengyu Liu, Li Yi, and He Wang. Unidexgrasp: Universal robotic dexterous grasping via learning diverse proposal generation and goal-conditioned policy. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 554, + 286, + 584 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 554, + 286, + 584 + ], + "spans": [ + { + "bbox": [ + 49, + 554, + 286, + 584 + ], + "type": "text", + "content": "[38] Yunhan Yang, Xiaoyang Wu, Tong He, Hengshuang Zhao, and Xihui Liu. Sam3d: Segment anything in 3d scenes, 2023. 6" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18111" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Generate Like Experts_ Multi-Stage Font Generation by Incorporating Font Transfer Process into Diffusion Models/bfb6c4d4-ce31-4c6d-ba23-750f3d411ecc_content_list.json b/2024/Generate Like Experts_ Multi-Stage Font Generation by Incorporating Font Transfer Process into Diffusion Models/bfb6c4d4-ce31-4c6d-ba23-750f3d411ecc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..186f4a1f00da001b42950bafea60d2cf437ad4d5 --- /dev/null +++ b/2024/Generate Like Experts_ Multi-Stage Font Generation by Incorporating Font Transfer Process into Diffusion Models/bfb6c4d4-ce31-4c6d-ba23-750f3d411ecc_content_list.json @@ -0,0 +1,1774 @@ +[ + { + "type": "text", + "text": "Generate Like Experts: Multi-Stage Font Generation by Incorporating Font Transfer Process into Diffusion Models", + "text_level": 1, + "bbox": [ + 127, + 128, + 843, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bin Fu $^{1}$ , Fanghua Yu $^{1}$ , Anran Liu $^{3}$ , Zixuan Wang $^{1}$ , Jie Wen $^{4}$ , Junjun He $^{2}$ , and Yu Qiao $^{1,2*}$", + "bbox": [ + 135, + 202, + 841, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ ShenZhen Key Lab of Computer Vision and Pattern Recognition, Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences $^{2}$ Shanghai Artificial Intelligence Laboratory $^{3}$ The University of Hong Kong $^{4}$ Harbin Institute of Technology, Shenzhen", + "bbox": [ + 174, + 238, + 792, + 308 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}\\{$ bin.fu, zx.wang8, yu.qiao\\}@siat.ac.cn, $^{2}\\{$ hejunjun, qiaoyu\\}@pjlab.org.cn $^{3}$ liuar616@connect.hku.hk, $^{4}$ jiewen-pr@126.com", + "bbox": [ + 165, + 310, + 803, + 344 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 378, + 313, + 395 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Few-shot font generation (FFG) produces stylized font images with a limited number of reference samples, which can significantly reduce labor costs in manual font designs. Most existing FFG methods follow the style-content disentanglement paradigm and employ the Generative Adversarial Network (GAN) to generate target fonts by combining the decoupled content and style representations. The complicated structure and detailed style are simultaneously generated in those methods, which may be the sub-optimal solutions for FFG task. Inspired by most manual font design processes of expert designers, in this paper, we model font generation as a multi-stage generative process. Specifically, as the injected noise and the data distribution in diffusion models can be well-separated into different sub-spaces, we are able to incorporate the font transfer process into these models. Based on this observation, we generalize diffusion methods to model font generative process by separating the reverse diffusion process into three stages with different functions: The structure construction stage first generates the structure information for the target character based on the source image, and the font transfer stage subsequently transforms the source font to the target font. Finally, the font refinement stage enhances the appearances and local details of the target font images. Based on the above multistage generative process, we construct our font generation framework, named MSD-Font, with a dual-network approach to generate font images. The superior performance demonstrates the effectiveness of our model. The code is available at: https://github.com/fubinfb/MSD-Font.", + "bbox": [ + 75, + 414, + 472, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 378, + 630, + 395 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Designing new fonts is time-consuming and labor-intensive work for humans, especially for some glyph-rich scripts (e.g., Chinese and Korean). The few-shot font generation (FFG) task aims to automatically generate new stylized fonts with a few reference images, and it has received significant interest due to its academic and commercial values. As the requirement of this task is to render the character images into the target font style, current FFG methods [17, 27, 28, 39] usually employ the Generative Adversarial Network (GAN) to perform few-shot font generation under the style-content disentanglement paradigm. These models usually simultaneously generate the complicated structure and detailed style of target font images by combining the content features from source characters and the style representations from reference font images. Although many methods have been proposed to obtain better content and style representations, the generated font images still contain artifacts, inconsistent styles, and broken structures.", + "bbox": [ + 496, + 410, + 893, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Inspired by the common practice of expert designers, we find the following characteristics that may be crucial for high-quality font generation: As font design requires preserving the global structure of the target characters, designers usually generate a new font by modifying from a similar existing font, which can be viewed as a transfer process from the source font to the target font. Thus, most manual font designs can be regarded as a multi-stage process, including template font generation for global structure, font transfer from the template font to the target font, and refinement of fine-grained details. In light of this, we think that the multi-stage generative process is an effective solution for high-quality font generation, since the problems of broken structures, inconsistent styles, and artifacts can be", + "bbox": [ + 496, + 689, + 895, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author: Yu Qiao", + "bbox": [ + 94, + 886, + 271, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6892", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/74738065b6cef9bb4a784b996d392061f1d7308b0a7e6fba13c357e257ad2913.jpg", + "image_caption": [ + "Figure 1. The forward process of our proposed model and the traditional diffusion model (DDPM). The first and third rows display the traditional forward process for the source image $I_{c}$ and target font image $I_{g}$ . The images with red boxes mark the path of our proposed diffusion process. In our model, we incorporate the font transfer process in time interval $(t_1, t_2]$ , and align the distribution with the traditional diffusion model at the timesteps $t_1$ and $t_2$ . With this forward process, the generative process of our model can be separated into three stages: the structure construction stage, the font transfer stage, and the font refinement stage, respectively. Note that the font images in this figure are visualized from the diffusion model in image space for better illustration1, while we implement our model in latent space to reduce computational consumption." + ], + "image_footnote": [], + "bbox": [ + 138, + 87, + 831, + 250 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "better handled by the above three stages.", + "bbox": [ + 76, + 359, + 346, + 373 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The diffusion model provides a powerful generative framework to achieve this multi-stage process, based on the following two features: (1). Since the injected noise and the data distribution in forward process can be well-separated into different sub-spaces, we are able to incorporate the font transfer process into diffusion models. (2). Recent studies [7, 18] have revealed that the different steps in the generative process have different functions for image synthesis. Therefore, modeling font generation as a transfer process and constructing a multi-stage generative process based on diffusion model is a promising direction for FFG task.", + "bbox": [ + 75, + 375, + 470, + 541 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by the above observations, in this paper, we incorporate the font transfer process into the diffusion model and formulate font generation as a multi-stage generative process. To alleviate computational consumption, we construct this diffusion process in latent space. Specifically, in latent space, we first separate the injected noise and the data distribution into different sub-spaces, and then incorporate the font transfer process into the data sub-space in a predefined time interval $(t_1,t_2]$ . Combining with the denoising process, the font generative process in our proposed diffusion model can be automatically separated into a multistage process (shown in Fig. 1): the structure construction stage, the font transfer stage, and the font refinement stage, respectively. The structure construction stage first generates the structure information for the target character based on the source image. Then, the font transfer stage gradually transforms the source font to the target font. Finally, the font refinement stage improves the appearances and local details of the generated font images. Under this scheme, the structure construction stage generates the global character structure from source images, while the font transfer stage", + "bbox": [ + 75, + 542, + 470, + 861 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "provides a long generative path to produce detailed styles. Moreover, as the prediction network has different behaviors in font transfer stage and font refinement stage, we further propose a dual-network approach to predict the target latent features for different stages. Experiments demonstrate that our model achieves the superior performance on FFG tasks, verifying the effectiveness of our proposed approach.", + "bbox": [ + 504, + 359, + 893, + 465 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions in this paper are summarized as follows:", + "bbox": [ + 500, + 465, + 890, + 494 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1). Instead of simultaneously generating font images with complicated structures and detailed styles in most traditional FFG methods, we reformulate the font generation as a multi-stage generative process by incorporating font transfer process into the diffusion models.", + "(2). We construct the multi-stage font generation framework, named MSD-Font, with a dual-network approach to better deal with different behaviors in different generative stages.", + "(3). Experimental results show that our method achieves superior performance on the FFG task, demonstrating the effectiveness of our proposed method." + ], + "bbox": [ + 500, + 496, + 890, + 678 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 694, + 650, + 709 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This section offers a concise review of recent development in font generation methods and diffusion models.", + "bbox": [ + 500, + 720, + 890, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Many-shot Font Generation", + "text_level": 1, + "bbox": [ + 500, + 755, + 750, + 771 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As font generation can be formulated as mapping font images from the source domain to the target domain, early methods [4, 40, 41] usually utilize the image-to-image (I2I) methods [8, 14, 20, 50, 51] to learn the mapping function between different fonts. The mapping function will be fine-tuned on hundreds of reference samples (many-shot) to generate unseen font images. Based on the pix2pix [14] framework, Zi2zi [40] and Rewrite [41] utilize one-hot style", + "bbox": [ + 500, + 780, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1Specifically, we replace $z_0^c$ and $z_0^g$ with $I_{c}$ and $I_{g}$ in Eq. (8)-(12) to visualize our proposed diffusion process in this figure.", + "bbox": [ + 76, + 875, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6893", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "labels to optimize font generation models in a supervised manner. HGAN [4] further extends Zi2zi model via a transfer network and the hierarchical adversarial discriminator. DC-font [15] utilizes a feature reconstruction network to embed style information for font synthesis. AGEN [23] employs the auto-encoder to transfer standard font to calligraphy font. Although these many-shot font generation methods can generate promising font images, collecting large numbers of reference samples to fine-tune the mapping function is still a difficult task for font generation.", + "bbox": [ + 75, + 90, + 472, + 243 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Few-shot Font Generation", + "text_level": 1, + "bbox": [ + 76, + 253, + 313, + 268 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To relieve the weakness of the above models, few-shot font generation (FFG) methods have become popular solutions in recent years, which generate font images with a few samples. As font styles have complex structures and fine-grained details, the common-used statistics-based style transfer techniques are not suitable for font generation. Instead, recent FFG models [2, 3, 10, 17, 21, 26-28, 39, 45-47, 49] mainly follow the style-content disentanglement paradigm, which transfers the target font styles by combining the content representations of source characters with the style codes of reference samples. SA-VAE [38], EMD [49], and AGISNet [10] extract the style and content representations as global features to generate target font images. DM-Font [3], LF-Font [27], and MX-Font [28] decompose characters into components and learn component-wise representations to improve local details. CG-GAN [17] supervises the font generator to decouple content and style at the component level via a component-wise discriminator. Diff-Font [11] proposes a stroke-wise diffusion model to generate font images under the guidance of three character attributes (content, style, and strokes). FsFont [39] utilizes a cross-attention mechanism to aggregate style features into the fine-grained style representation. XMP-Font [21] develops a self-supervised cross-modality pre-training strategy and a cross-modality transformer-based encoder to model the style representations at all levels. DS-Font [12] proposes a cluster-level contrastive style loss to learn better style representations. DG-Font [45] and its extended version DG-Font++ [5] propose a feature deformation skip connection to learn geometric displacement maps between different fonts. CF-Font [43] further extends DG-Font to construct a robust content feature by projecting the content feature into a linear space defined by the content features of basis fonts. Besides, NTF [9] models font generation as a continuous transformation process with a neural transformation field.", + "bbox": [ + 75, + 276, + 472, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Diffusion Probabilistic Models", + "text_level": 1, + "bbox": [ + 76, + 830, + 346, + 847 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion model (DM) provides a powerful probabilistic generative framework for high-quality image synthesis, which is first introduced in [33] and subsequently improved", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "by Denoising Diffusion Probabilistic Model (DDPM) [13], Denoising Diffusion Implicit Model (DDIM) [34], score-based diffusion [35], and other techniques [25, 36, 37]. Early diffusion models construct the diffusion process in RGB space, which requires huge computation powers to generate high-resolution images. Latent Diffusion Model (LDM) [30] builds the diffusion process in a lower-dimensional latent space, significantly reducing the computation cost. The diffusion model has become the dominant solution for various computer vision tasks, such as image synthesis [29, 30, 32], image editing [1, 6, 24], and image-to-image translation [31, 44]. Large-scale text-to-image (T2I) diffusion models (e.g. DALL-E 2 [29], Imagegen [32], Stable Diffusion [30]) have shown superior performance on high-quality image synthesis, which employ text embedding generated by large language models (LLM) as the condition to guide image generation. Recently, several works [19, 22, 52] modify the denoising diffusion process to model the residual between two given distributions. Different from these works, in this paper, we introduce the font transfer process into the diffusion model at a predefined region, forming a multi-stage diffusion process. Moreover, based on this process, a novel dual-network approach is proposed to predict latent features in different stages.", + "bbox": [ + 496, + 90, + 893, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 500, + 474, + 635, + 492 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we present our proposed method in detail, including a short background of the diffusion model (in Sec. 3.1), the proposed latent diffusion model with font transfer process (in Sec. 3.2), the multi-stage font generative process (in Sec. 3.3), and the overall framework (in Sec. 3.4).", + "bbox": [ + 496, + 502, + 892, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Background and Preliminary", + "text_level": 1, + "bbox": [ + 498, + 595, + 761, + 612 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The diffusion model is a probabilistic generative model aiming to learn the data distribution by gradually denoising a variable sampled from a Gaussian distribution, which includes a forward process and a reverse process.", + "bbox": [ + 496, + 619, + 892, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a data distribution $x_0 \\sim q(x_0)$ , the Forward Process (also named as the diffusion process) is fixed to a Markov chain that gradually injects Gaussian noise to the data at time $t$ with a predefined variance schedule $\\beta_t$ according to the equations:", + "bbox": [ + 496, + 684, + 893, + 760 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(x _ {1}, \\dots , x _ {T} \\mid x _ {0}\\right) = \\prod_ {t = 1} ^ {T} q \\left(x _ {t} \\mid x _ {t - 1}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 775, + 890, + 816 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(x _ {t} \\mid x _ {t - 1}\\right) = \\mathcal {N} \\left(x _ {t}; \\sqrt {1 - \\beta_ {t}} x _ {t - 1}, \\beta_ {t} I\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 820, + 890, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $x_{1},\\dots ,x_{T}$ are latent variables. With the sufficiently large $T$ , the distribution of $x_{T}$ will become an isotropic Gaussian distribution. With the reparameterization trick", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6894", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/627114c7efa60a8d931cbc1c36fa66cfa37905ce8f42190ff549551889b0e5fd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 90, + 504, + 220 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/afd3bb041299288f31c8b3840a4a19ba32fb1cc13a40b11fb4e16c2aac93c4b4.jpg", + "image_caption": [ + "Figure 2. (a). The forward diffusion process of our proposed model. (b). The reverse process of our proposed model. Our proposed diffusion process is marked as black circles. The $\\tilde{z}_t$ denotes the latent features in our diffusion process, while $z_t^c$ and $z_t^g$ denote the latent features of the source image and the target image in traditional diffusion process, respectively. As we have aligned our diffusion process with the traditional diffusion process of $z^g$ and $z^c$ at $t_1$ and $t_2$ , our model shares the same process with the traditional diffusion process in the region $(0, t_1]$ and $(t_2, T]$ . (c). Overall framework of our font generation model, including the VQ-VAE encoder $\\mathcal{E}$ and decoder $\\mathcal{D}$ , the dual prediction networks $\\tilde{z}_{\\theta_1}^{(g,1)}(\\tilde{z}_t, t, y_1)$ and $\\tilde{z}_{\\theta_2}^{(g,2)}(\\tilde{z}_t, t, y_2)$ , and the corresponding style and content encoders for each prediction network. Due to the same structure, in this figure, we only plot one prediction network $\\tilde{z}_{\\theta_i}^{(g,i)}$ ( $i \\in \\{1, 2\\}$ ) with its corresponding style encoder $E_s^i$ and content encoder $E_c^i$ , where $i$ is determined by the generative stage of timestep $t$ . We use $I_g$ and $I_G$ to denote the target image and generated image, respectively." + ], + "image_footnote": [ + "Generative Step in Structure Construction Stage", + "Generative Step in Font Transfer Stage" + ], + "bbox": [ + 83, + 220, + 506, + 354 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/674cf8f1cd9dd4dd7e6d9a0a6e23812a5cdbf368c3e6f1cc6091fc37e1c4d34f.jpg", + "image_caption": [], + "image_footnote": [ + "$\\tilde{z}_t$ Latent feature in our model \n $z_t^c$ Latent feature of source \nimage in traditional model", + "$\\widetilde{Z}_{\\theta_i}^{(g,i)}$ Prediction network $(i\\in \\{1,2\\})$ Latent feature of target image $z_{t}^{g}$ in traditional model", + "The Equivalent Process for Structure Construction Stage", + "Generative Step in Font Refinement Stage" + ], + "bbox": [ + 509, + 90, + 885, + 265 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "[16], the latent variable $x_{t}$ can be written as", + "bbox": [ + 76, + 506, + 370, + 522 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} x _ {t} = \\sqrt {\\alpha_ {t}} x _ {t - 1} + \\sqrt {1 - \\alpha_ {t}} \\epsilon_ {t - 1} ^ {*}, (3) \\\\ = \\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {0}, (4) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 530, + 468, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\epsilon_{t - 1}^{*},\\epsilon_{0}\\sim \\mathcal{N}(0,I),\\alpha_{t} = 1 - \\beta_{t}$ and $\\bar{\\alpha}_t = \\prod_{s = 1}^t\\alpha_s$ Thus sampling $x_{t}$ at arbitrary timestep has a closed form $q(x_{t}|x_{0}) = \\mathcal{N}(x_{t};\\sqrt{\\bar{\\alpha}_{t}} x_{0},(1 - \\bar{\\alpha}_{t})I)$", + "bbox": [ + 76, + 574, + 468, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Reverse Process generates data from isotropic Gaussian noise $x_{T}$ , and gradually recovers $x_{0}$ via the reverse distribution $p_{\\theta}(x_{t - 1}|x_t)$ :", + "bbox": [ + 75, + 619, + 468, + 666 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(x _ {0}\\right) = \\int p _ {\\theta} \\left(x _ {0: T}\\right) d x _ {1: T}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 672, + 468, + 704 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(x _ {0: T}\\right) = p _ {\\theta} \\left(x _ {T}\\right) \\prod_ {t = 1} ^ {T} p _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 707, + 468, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $p_{\\theta}(x_{t - 1}|x_t) = \\mathcal{N}(x_{t - 1}|\\mu_\\theta (x_t,t),\\Sigma_\\theta (x_t,t))$ is a parameterized Gaussian transition network, and defined as", + "bbox": [ + 76, + 753, + 468, + 785 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {\\theta} (x _ {t}, t) = \\frac {\\sqrt {\\alpha_ {t}} (1 - \\bar {\\alpha} _ {t - 1}) x _ {t} + \\sqrt {\\bar {\\alpha} _ {t - 1}} (1 - \\alpha_ {t}) \\hat {x} _ {\\theta} (x _ {t} , t)}{1 - \\bar {\\alpha} _ {t}},\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 791, + 468, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and $\\Sigma_{\\theta}(x_t,t) = (1 - \\bar{\\alpha}_{t - 1})\\beta_tI / (1 - \\bar{\\alpha}_t)$ . $\\hat{x}_{\\theta}(x_{t},t)$ is a function approximator for predicting $x_0$ from $x_{t}$ and $t$ . The corresponding objective function can be simplified to:", + "bbox": [ + 76, + 830, + 468, + 877 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL = \\mathbb {E} _ {x _ {0}, \\epsilon , t} \\left[ \\| \\hat {x} _ {\\theta} (x _ {t}, t) - x _ {0} \\| _ {2} ^ {2} \\right]. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 883, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Latent Diffusion Model (LDM) [30] further employs a VQ-VAE [42] encoder $\\mathcal{E}$ to project image $x_0$ into latent space $z_0$ , and constructs the diffusion model in this lower-dimensional latent space, significantly reducing computational consumption. Once $z_0$ is sampled via the reverse process, the VQ-VAE decoder $\\mathcal{D}$ will project it back to the image space.", + "bbox": [ + 496, + 506, + 890, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since LDM significantly reduces computation consumption and maintains high-quality image synthesis, in this paper, we construct our diffusion process in latent space.", + "bbox": [ + 496, + 612, + 890, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Latent Diffusion Model with Font Transfer", + "text_level": 1, + "bbox": [ + 500, + 665, + 864, + 680 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Eq. (4), in the forward process, the data distribution $x_0$ and the injected noise can be well-separated in arbitrary timesteps. This characteristic enables us to construct a multi-stage font generative framework in latent diffusion model, by incorporating the font transfer process into the data sub-space in a predefined time interval $t \\in (t_1, t_2]$ . Based on this observation, we develop our latent diffusion model in this section, then discuss the multi-stage font generative process in the next section.", + "bbox": [ + 496, + 688, + 890, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Forward Process. After projecting the source images $I_{c}$ and target images $I_{g}$ into the latent features $z_0^c$ and $z_0^g$ by the VQ-VAE encoder $\\mathcal{E}$ , we implement our model in latent space. As shown in Fig. 2(a), our proposed model (marked as black circles and denoted as $\\tilde{z}_t$ ) is constructed", + "bbox": [ + 496, + 825, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6895", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "by utilizing a font transfer process to connect the traditional diffusion processes (DDPM) of the latent features $z_{t}^{c}$ and $z_{t}^{g}$ . Therefore, our latent diffusion process $\\tilde{z}_{t}$ can be divided into three regions: (1). In the region $t \\in (0, t_1]$ , only latent feature $z_{t}^{g}$ evolved with noise. The forward latent $\\tilde{z}_{t}$ in this region can be expressed as", + "bbox": [ + 75, + 90, + 472, + 183 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {z} _ {t} = z _ {t} ^ {g} = \\sqrt {\\bar {\\alpha} _ {t}} z _ {0} ^ {g} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {0} t \\in (0, t _ {1} ]. \\qquad (8)\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 191, + 468, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(2). In the region $t \\in (t_1, t_2]$ , together with the continuously injected noise, the latent feature $z_t^g$ of the target font is gradually replaced by the latent feature $z_t^c$ of the source character, and only $z_t^c$ exists at $t_2$ . After careful derivation (provided in Supplementary Materials), the forward latent $\\tilde{z}_t$ in this region can be expressed as", + "bbox": [ + 75, + 218, + 468, + 309 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {z} _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} z _ {0} ^ {g} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {0} - \\left(z _ {0} ^ {g} - z _ {0} ^ {c}\\right) \\Psi \\mathcal {H} _ {t}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 319, + 468, + 338 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "for $t \\in (t_1, t_2]$ . The notations $\\mathcal{H}_t$ and $\\Psi$ are coefficients obtained by aligning our model with traditional diffusion processes (DDPM) of $z^g$ and $z^c$ at timesteps $t_1$ and $t_2$ , respectively. They are defined as", + "bbox": [ + 75, + 345, + 468, + 407 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {H} _ {t} = 1 + \\sqrt {\\alpha_ {t}} + \\sqrt {\\alpha_ {t} \\alpha_ {t - 1}} + \\dots + \\sqrt {\\alpha_ {t} \\cdots \\alpha_ {t + 2}}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 417, + 468, + 435 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Psi = \\frac {\\sqrt {\\tilde {\\alpha} (1 , t _ {2})}}{1 + \\sum_ {m = t _ {1} + 2} ^ {t _ {2}} \\sqrt {\\tilde {\\alpha} (m , t _ {2})}}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 438, + 468, + 476 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tilde{\\alpha}(t_1, t_2) = \\prod_{i=t_1}^{t_2} \\alpha_i$ . $\\mathcal{H}_t$ can be further reduced to $\\mathcal{H}_t = 1 + \\sqrt{\\alpha_t} \\mathcal{H}_{t-1}$ . (3). In the region $t \\in (t_2, T]$ , the injected noise finally disrupts the latent feature $z_t^c$ to Gaussian noise, and the forward latent $\\tilde{z}_t$ in this region can be expressed as", + "bbox": [ + 75, + 486, + 468, + 563 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {z} _ {t} = z _ {t} ^ {c} = \\sqrt {\\bar {\\alpha} _ {t}} z _ {0} ^ {c} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {0} \\quad t \\in (t _ {2}, T ]. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 573, + 468, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Reverse Process. Now we turn to discuss our choices for $p_{\\theta}(\\tilde{z}_{t - 1}|\\tilde{z}_t) = \\mathcal{N}(\\tilde{z}_{t - 1}|\\mu_\\theta (\\tilde{z}_t,t),\\Sigma_\\theta (\\tilde{z}_t,t))$ . As shown in Eq. (8) and Eq. (12), in the time region $(0,t_1]$ and $(t_2,T]$ , we have aligned the forward latent $\\tilde{z}_t$ with the traditional diffusion processes (DDPM [13]), thus they share the same parameterized mean functions: for $t\\in (0,t_1]$", + "bbox": [ + 75, + 599, + 468, + 691 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {\\theta} = \\frac {\\sqrt {\\alpha_ {t}} (1 - \\bar {\\alpha} _ {t - 1}) \\tilde {z} _ {t} + \\sqrt {\\bar {\\alpha} _ {t - 1}} (1 - \\alpha_ {t}) \\tilde {z} _ {\\theta} ^ {(g)} (\\tilde {z} _ {t} , t)}{1 - \\bar {\\alpha} _ {t}}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 700, + 468, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tilde{z}_{\\theta}^{(g)}(\\tilde{z}_t,t)$ is the prediction network for predicting $z_0^g$ from noisy latent feature $\\tilde{z}_t$ and time index $t$ , while for $t\\in (t_2,T]$", + "bbox": [ + 76, + 750, + 468, + 796 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {\\theta} = \\frac {\\sqrt {\\alpha_ {t}} (1 - \\bar {\\alpha} _ {t - 1}) \\tilde {z} _ {t} + \\sqrt {\\bar {\\alpha} _ {t - 1}} (1 - \\alpha_ {t}) \\tilde {z} _ {\\theta} ^ {(c)} (\\tilde {z} _ {t} , t)}{1 - \\bar {\\alpha} _ {t}}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 805, + 468, + 852 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tilde{z}_{\\theta}^{(c)}(z_t,t)$ is the prediction network for predicting $z_0^c$ from noisy latent feature $\\tilde{z}_t$ and time index $t$ . The parameterized mean function for the region $t\\in (t_1,t_2]$ can be", + "bbox": [ + 75, + 853, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "derivated via Bayes' theorem:", + "bbox": [ + 500, + 90, + 700, + 106 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mu_ {\\theta} = \\frac {\\sqrt {\\alpha_ {t}} (1 - \\bar {\\alpha} _ {t - 1})}{1 - \\bar {\\alpha} _ {t}} \\tilde {z} _ {t} + \\frac {\\sqrt {\\alpha_ {t}} (1 - \\bar {\\alpha} _ {t - 1})}{1 - \\bar {\\alpha} _ {t}} \\Psi \\tilde {z} _ {\\theta} ^ {(g)} (\\tilde {z} _ {t}, t) \\\\ + \\frac {\\left[ \\mathcal {H} _ {t - 1} (1 - \\alpha_ {t}) - \\sqrt {\\alpha_ {t}} (1 - \\bar {\\alpha} _ {t - 1}) \\right]}{1 - \\bar {\\alpha} _ {t}} \\Psi z _ {0} ^ {c} \\\\ + \\frac {(\\sqrt {\\bar {\\alpha} _ {t - 1}} - \\mathcal {H} _ {t - 1} \\Psi) (1 - \\alpha_ {t})}{1 - \\bar {\\alpha} _ {t}} \\tilde {z} _ {\\theta} ^ {(g)} (\\tilde {z} _ {t}, t). \\tag {15} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 112, + 892, + 213 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As we only incorporate the font transfer process in the data sub-space and keep the noise sub-space unchanged, the variance in the reverse process remains the same as DDPM:", + "bbox": [ + 498, + 218, + 890, + 263 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Sigma_ {\\theta} \\left(\\tilde {z} _ {t}, t\\right) = \\sigma_ {t} ^ {2} I = \\frac {(1 - \\alpha_ {t}) (1 - \\bar {\\alpha} _ {t - 1})}{1 - \\bar {\\alpha} _ {t}} I. \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 268, + 890, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Therefore, the latent $\\tilde{z}_{t-1}$ in reverse process can be obtained via the formula $p_{\\theta}(\\tilde{z}_{t-1}|\\tilde{z}_t) = \\mathcal{N}(\\tilde{z}_{t-1};\\mu_{\\theta}(\\tilde{z}_t,t),\\sigma_t^2 I)$ , where the $\\mu_{\\theta}(\\tilde{z}_t,t)$ is calculated from Eq. (13)-(15) based on different timesteps.", + "bbox": [ + 498, + 306, + 890, + 366 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Optimization Objective. As the forward behaviors keep the same as DDPM in the region $(0, t_1]$ and $(t_2, T]$ , the optimization objectives of $\\tilde{z}_{\\theta}^{(g)}(\\tilde{z}_t, t)$ and $\\hat{z}_{\\theta}^{(c)}(\\tilde{z}_t, t)$ have the same expressions as DDPM in these regions. For the optimization objective of $\\tilde{z}_{\\theta}^{(g)}(\\tilde{z}_t, t)$ in the region $(t_1, t_2]$ , it can be derived from the variational bound on negative log-likelihood [13], and only the coefficient is different with the DDPM's loss function. Moreover, as DDPM and LDM adopt the simplified training objectives in their implementations, we also simplify our loss functions and obtain the optimization objective as", + "bbox": [ + 498, + 367, + 892, + 537 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {g} + L _ {c}, \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 645, + 545, + 890, + 561 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {g} = \\mathbb {E} _ {\\mathcal {E} (x _ {0}), \\epsilon , t \\sim U (0, t _ {2} ]} \\left[ \\| \\tilde {z} _ {\\theta} ^ {(g)} (\\tilde {z} _ {t}, t) - z _ {0} ^ {g} \\| _ {2} ^ {2} \\right], \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 565, + 890, + 590 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {c} = \\mathbb {E} _ {\\mathcal {E} (x _ {0}), \\epsilon , t \\sim U (t _ {2}, T ]} \\left[ \\| \\tilde {z} _ {\\theta} ^ {(c)} (\\tilde {z} _ {t}, t) - z _ {0} ^ {c} \\| _ {2} ^ {2} \\right], \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 592, + 890, + 618 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $U(t_{1},t_{2}]$ denotes the uniform distribution in region $(t_1,t_2]$ .", + "bbox": [ + 500, + 623, + 890, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Conditioning Mechanism. Following the LDM [30], we generalize our method to model conditional distributions by implementing with the conditional prediction networks:", + "bbox": [ + 498, + 654, + 890, + 710 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {z} _ {\\theta} ^ {(i)} \\left(\\tilde {z} _ {t}, t, y\\right), i \\in \\{g, c \\}, \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 611, + 710, + 890, + 731 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $y$ is the condition for controlling the synthesis process.", + "bbox": [ + 498, + 734, + 890, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Extension to Conditional Font Generation. We further extend this conditional model to perform the font generation task. Since the few-shot font generation task aims to render the source character $I_{c}$ into the target font style by a few reference samples $I_{s}$ , we utilize the style encoder $E_{s}$ and the content encoder $E_{c}$ to extract the style code from the reference images $I_{s}$ and the content representations from the source images $I_{c}$ . The style and content features are flattened and concatenated to build condition $y$ .", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6896", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Multi-Stage Font Generative Process", + "text_level": 1, + "bbox": [ + 76, + 90, + 395, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Based on our proposed model, the font generation in the reverse process can be automatically separated into three different stages: the structure construction stage, the font transfer stage, and the font refinement stage.", + "bbox": [ + 75, + 121, + 468, + 181 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Fig. 2(b), starting with the sampled Gaussian noise $\\tilde{z}_T$ , in the region $t \\in (t_2, T]$ , Structure Construction Stage generates the structure information for the target character via predicting the latent feature $z_0^c$ of the source image $I_c$ . However, since we can access the source image $I_c$ during the inference, the iterative generative steps from $T$ to $t_2$ can be replaced by a single forward step to generate the noised latent variable $\\tilde{z}_{t_2}$ at time $t = t_2$ via Eq. (12). This will significantly reduce inference time and provide more accurate structure information. Font Transfer Stage performs font transformation from the timestep $t_2$ to $t_1$ , where the noised latent feature $\\tilde{z}_{t_2}$ is gradually transformed to the noised latent feature $\\tilde{z}_{t_1}$ of target images via the $\\tilde{z}_{\\theta}^{(g)}(\\tilde{z}_t, t, y)$ . To predict the latent feature $z_0^g$ , the prediction network $\\tilde{z}_{\\theta}^{(g)}(\\tilde{z}_t, t, y)$ in this stage is optimized to not only denoise the latent features but also perform font transfer under the condition $y$ . Finally, in the region $t \\in (0, t_1]$ , Font Refinement Stage improves appearances and local details for the generated font images via predicting $z_0^g$ from the noisy latent feature $\\tilde{z}_t$ . Based on the above three stages, we can implement the reverse process to generate high-quality font images.", + "bbox": [ + 75, + 188, + 472, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To better present this multi-stage font generative process, we summarize above steps and provide the pseudo-code in Supplementary Materials.", + "bbox": [ + 76, + 534, + 468, + 579 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Overall Framework for Font Generation", + "text_level": 1, + "bbox": [ + 76, + 609, + 423, + 626 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As discussed in Sec. 3.3, the network $\\tilde{z}_{\\theta}^{(c)}(\\tilde{z}_t,t,y)$ can be replaced by a single forward step. Moreover, instead of implementing a single prediction network $\\tilde{z}_{\\theta}^{(g)}(\\tilde{z}_t,t,y)$ to predict $z_0^g$ we adopt a dual-network approach in this paper, since the network $\\tilde{z}_{\\theta}^{(g)}(\\tilde{z}_t,t,y)$ has different behaviors in the font transfer stage and the font refinement stage. The dual networks $\\tilde{z}_{\\theta_1}^{(g,1)}(\\tilde{z}_t,t,y_1)$ and $\\tilde{z}_{\\theta_2}^{(g,2)}(\\tilde{z}_t,t,y_2)$ share the same structure, but are trained to predict the latent feature $z_0^g$ in different regions: one for font transfer stage and the other for font refinement stage. Therefore, as shown in Fig. 2(c), the overall framework of our font generation model contains the following parts: the VQ-VAE encoder $\\mathcal{E}$ , the VQ-VAE decoder $\\mathcal{D}$ , the dual prediction networks $\\tilde{z}_{\\theta_1}^{(g,1)}(\\tilde{z}_t,t,y_1)$ and $\\tilde{z}_{\\theta_2}^{(g,2)}(\\tilde{z}_t,t,y_2)$ , and the corresponding style and content encoders for each prediction network $\\tilde{z}_{\\theta_i}^{(g,i)}(\\tilde{z}_t,t,y_i)$ . Finally, from Eq. (17), the optimization objective of our font gener", + "bbox": [ + 75, + 638, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ative model can be further reduced to", + "bbox": [ + 500, + 90, + 750, + 104 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {g} ^ {1} + L _ {g} ^ {2}, \\tag {21}\n$$\n", + "text_format": "latex", + "bbox": [ + 645, + 111, + 890, + 131 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {g} ^ {1} = \\mathbb {E} _ {\\mathcal {E} (x _ {0}), \\epsilon , t \\sim U (t _ {1}, t _ {2} ]} \\left[ \\| \\tilde {z} _ {\\theta_ {1}} ^ {(g, 1)} (\\tilde {z} _ {t}, t, y _ {1}) - z _ {0} ^ {g} \\| _ {2} ^ {2} \\right], \\tag {22}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 133, + 890, + 157 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {g} ^ {2} = \\mathbb {E} _ {\\mathcal {E} (x _ {0}), \\epsilon , t \\sim U (0, t _ {1} ]} \\left[ \\| \\tilde {z} _ {\\theta_ {2}} ^ {(g, 2)} (\\tilde {z} _ {t}, t, y _ {2}) - z _ {0} ^ {g} \\| _ {2} ^ {2} \\right]. \\tag {23}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 161, + 890, + 186 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 191, + 633, + 209 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we conduct extensive experiments to evaluate our proposed model on the FFG task.", + "bbox": [ + 498, + 217, + 890, + 247 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implement Details", + "text_level": 1, + "bbox": [ + 498, + 255, + 678, + 271 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our models are implemented on the Stable Diffusion [30] platform. We utilize $T = 1000$ steps to generate $128 \\times 128 \\times 3$ font images with 8 reference samples. The VQ-VAE encoder $\\mathcal{E}$ and decoder $\\mathcal{D}$ , the dual-network $\\tilde{z}_{\\theta_1}^{(g,1)}(\\tilde{z}_t,t,y_1)$ and $\\tilde{z}_{\\theta_2}^{(g,2)}(\\tilde{z}_t,t,y_2)$ are initialized by the pre-trained Stable Diffusion models. The style encoder $E_{s}$ and content encoder $E_{c}$ are initialized by Kaiming initialization. We keep VQ-VAE encoder $\\mathcal{E}$ and decoder $\\mathcal{D}$ frozen and optimize the remaining networks.", + "bbox": [ + 498, + 277, + 890, + 421 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Dataset and Evaluation Metrics", + "text_level": 1, + "bbox": [ + 500, + 428, + 779, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets: As DM needs more training data to optimize, we collect 900 Chinese fonts to evaluate our method on FFG task. We randomly select 840 fonts as the training set (seen fonts) and the remaining 60 fonts as the testing fonts (unseen fonts). As each collected font covers different Chinese characters, we randomly sample 800 characters from the commonly-used Chinese characters (the first-level Chinese Character Table) to train our method. For the test set, we use 214 characters as the unseen contents and sample 200 characters in our training set as the seen contents.", + "bbox": [ + 496, + 450, + 890, + 601 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Therefore, the training set contains 840 fonts with 800 characters, while the test set includes two parts: one is the Unseen Fonts Unseen Contents (UFUC) test set, containing 60 unseen fonts with 214 unseen characters; the other is the Unseen Fonts Seen Contents (UFSC) test set, containing 60 unseen fonts with 200 seen characters.", + "bbox": [ + 496, + 602, + 890, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics: In this paper, we utilize four commonly used measurements to evaluate our model, including root mean squared error (RMSE), peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and learned perceptual image path similarity (LPIPS) [48].", + "bbox": [ + 496, + 691, + 890, + 768 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 775, + 653, + 791 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "With the above dataset and metrics, we conducted extensive experiments in different settings.", + "bbox": [ + 498, + 799, + 890, + 829 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.1 Ablation Study on the Overall Framework", + "text_level": 1, + "bbox": [ + 500, + 845, + 849, + 861 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first construct a Base Model with traditional diffusion process (DDPM) to verify the effectiveness of our proposed", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6897", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/036e1b68f8be3f20343b7849c20f43c7da10feb07e760b5a3ff191d7b37559c5.jpg", + "image_caption": [ + "Figure 3. Visualization results of the intermediate results $\\tilde{z}_t$ by projecting $\\tilde{z}_t$ back to image space via VQ-VAE decoder $\\mathcal{D}$ ." + ], + "image_footnote": [], + "bbox": [ + 81, + 87, + 890, + 229 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1bbb281e3f0dfe20f2db227bb746f82f4c7d078d53527a2eb62b3050abfd947d.jpg", + "image_caption": [ + "Figure 4. Visualization results of the predicted latent feature $z_0^g$ in the intermediate timesteps by projecting $z_0^g$ back to image space via VQ-VAE decoder $\\mathcal{D}$ ." + ], + "image_footnote": [], + "bbox": [ + 117, + 258, + 426, + 401 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c5bde9dce3f79063a994aec793d1d2fcb1e974141bacf212ce9b5406093d1d78.jpg", + "table_caption": [ + "Table 1. Ablation study for the overall framework on UFUC dataset. The bold number indicates the best." + ], + "table_footnote": [], + "table_body": "
MethodsRMSE↓PSNR↑SSIM↑LPIPS↓
Base Model0.267611.720.69610.1561
Single-Net0.260112.020.70960.1531
Dual-Net0.247512.450.71890.1527
", + "bbox": [ + 99, + 492, + 444, + 555 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "latent diffusion model. Moreover, as discussed in Sec. 3.4, there are two different selections to obtain $z_0^g$ in the reverse process: One utilizes a single network to predict $z_0^g$ at all timesteps. The other adopts the dual networks to predict the latent feature $z_0^g$ in font transfer stage and font refinement stage, respectively.", + "bbox": [ + 75, + 577, + 468, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Tab. 1, compared with the Base Model, our Single-Net model improves the Base Model with 0.0075, 0.30, 0.0135, and 0.0030 in terms of RMSE, PSNR, SSIM, and LPIPS, verifying the effectiveness of our proposed latent diffusion model. Moreover, the Dual-Net approach further improves the Single-Net model with 0.0126, 0.43, and 0.0093 in terms of RMSE, PSNR, and SSIM, illustrating that the use of separate models is beneficial for predicting $z_0^g$ in different stages.", + "bbox": [ + 75, + 669, + 468, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.2 Ablation Study on the Multi-Stage Font Generative Process", + "text_level": 1, + "bbox": [ + 76, + 829, + 468, + 858 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The selections of the region for different stages in the multistage generative process are hyper-parameters in our model,", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "which significantly impact image quality. Therefore, we study the influence on model performance by selecting different $t_1$ and $t_2$ . Due to the page limitation, the experimental results and the detailed discussions are provided in supplementary materials, and we briefly summarize the conclusion: The hyper-parameters $t_1$ and $t_2$ are trade-off selections, and the configuration $t_1 = 200$ and $t_2 = 800$ achieves the best record. Thus, we select this configuration for the following experiments.", + "bbox": [ + 496, + 261, + 890, + 398 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.3 Visualization Results", + "text_level": 1, + "bbox": [ + 498, + 404, + 699, + 417 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we provide the visualization results of our generative process to verify the effectiveness of our model.", + "bbox": [ + 496, + 428, + 890, + 458 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As our diffusion process is implemented in latent space, we visualize the intermediate generative results $\\tilde{z}_t$ by using the VQ-VAE decoder $\\mathcal{D}$ to project them back into the image space. As shown in Fig. 3, the images in the first and third rows are the intermediate results for source image $I_{c}$ and target image $I_{g}$ generated by the traditional diffusion process (DDPM), while the images in the second row are generated by our multi-stage font generative process. From this figure, we can clearly observe a font transfer process from the source font to the target font, and the font refinement stage slightly improves style consistency by modifying local details.", + "bbox": [ + 496, + 459, + 890, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further evaluate the font refinement stage, we also visualize the predicted $z_0^g$ from different timesteps in the font refinement stage. As shown in Fig. 4, from $t = 200$ to $t = 0$ , the font refinement stage can gradually enhance local details by modifying anomalous pixels, recovering unwanted structures, and improving visual appearances.", + "bbox": [ + 496, + 640, + 890, + 731 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Comparison with the State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 498, + 739, + 890, + 757 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we compare our model with several state-of-the-art works, including LF-Font [27], DG-Font [45], MX-Font [28], FsFont [39], NTF [9], and Diff-Font [11]. LF-Font and MX-Font are component-related methods, which decompose characters into components and learn component-wise representations to improve local details. FsFont focuses on fine-grained local styles, which utilizes cross attention to aggregate the reference styles into target styles. DG-Font and NTF are shape transformation mod", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6898", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e64a3ae2edfe1f303ede3df8a9d0d2652ee7ef870795a12ce8fd9b6d00eb009e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Content辛禁病安持理根英役破任始米保用李倍荷供唱血止在利打
LF-Font辛禁病安持理根英役破任始米保用李倍荷供唱血止在利打
DG-Font辛禁病安持理根英役破任始米保用李倍荷供唱血止在利打
MX-Font辛禁病安持理根英役破任始米保用李倍荷供唱血止在利打
FsFont辛禁病安持理根英役破任始米保用李倍荷供唱血止在利打
NTF辛禁病安持理根英役破任始米保用李倍荷供唱血止在利打
Ours辛禁病安持理根英役破任始米保用李倍荷供唱血止在利打
Ground Truth辛禁病安持理根英役破任始米保用李倍荷供唱血止在利打
", + "bbox": [ + 133, + 90, + 834, + 250 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e16ced0aa032da68223e9b80aab451e8ec60c2caa1b9192c6b4adff742c05d64.jpg", + "table_caption": [ + "Figure 5. Qualitative comparisons of our proposed model with other state-of-the-art methods on UFUC dataset.", + "Table 2. Quantitative comparison of our proposed model with other state-of-the-art methods on few-shot font generation task. The bold number indicates the best." + ], + "table_footnote": [], + "table_body": "
MethodsRMSE↓PSNR↑SSIM↑LPIPS↓
Unseen Fonts and Seen Contents
LF-Font [27]0.302510.660.64690.2787
DG-Font [45]0.283311.240.66510.1479
MX-Font [28]0.300710.750.64380.1651
FsFont [39]0.244312.560.71050.1468
NTF [9]0.265011.830.68960.1517
Diff-Font2[11]0.281110.950.68460.1557
MSD-Font0.243912.610.71100.1348
Unseen Fonts and Unseen Contents
LF-Font [27]0.291011.090.67620.2564
DG-Font [45]0.275011.540.68930.1534
MX-Font [28]0.284511.290.67890.1635
FsFont [39]0.250912.290.71850.1642
NTF [9]0.257412.110.71080.1553
MSD-Font0.247512.450.71890.1527
", + "bbox": [ + 78, + 324, + 475, + 573 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "els. DG-Font models font generation as a shape deformable problem via the deformation skip connection, while NTF models font generation as a continuous transformation process. We train above methods on our dataset and use the same reference images to generate fonts in inference.", + "bbox": [ + 75, + 587, + 468, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4.1 Quantitative Comparison", + "text_level": 1, + "bbox": [ + 76, + 667, + 310, + 684 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The quantitative results are shown in Tab. 2. From the table, we find that modeling font style in fine-grained manners by multiple localized encoders (MX-Font) or cross attention mechanism (FsFont) can improve the quality of generated font images. The shape transformation approaches (DG-Font and NTF) are also promising solutions for FFG task. Our MSD-Font achieves superior performance on both UFUC and UFSC datasets, which demonstrates the effectiveness of our proposed method. Specifically, compared with the NTF on the UFUC dataset, our method improves the generation performance with 0.0099, 0.34, 0.0081, and 0.0026 in terms of RMSE, PSNR, SSIM, and LPIPS.", + "bbox": [ + 75, + 694, + 468, + 875 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Diff-Font is one-shot font generation model, we randomly select one image from 8 reference samples to calculate style code.", + "bbox": [ + 76, + 875, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4.2 Qualitative Comparison", + "text_level": 1, + "bbox": [ + 498, + 279, + 723, + 295 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To further evaluate our methods, as shown in Fig. 5, we visualize the generated font images in UFUC setting. LF-Font sometimes generates font images with some broken structures. The DG-Font keeps the global structure for the target character, but the generated images may contain some unwanted local structures and artifacts. FsFont utilizes several carefully selected reference characters to generate font images, thus sometimes it will fail to produce accurate structures for target images with randomly selected reference characters. The MX-Font and NTF also contain some missing strokes and distorted local details. Our MSD-Font is able to generate high-quality font images, which achieves higher visual quality than other state-of-the-art methods in terms of style consistency and structure completeness.", + "bbox": [ + 496, + 313, + 893, + 526 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 498, + 550, + 619, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a multi-stage font generative process for few-shot font generation task, by incorporating font transfer process into the diffusion model. The proposed font generation model includes three stages: a structure construction stage to generate global structures from the source images, a font transfer stage to transform the source font into the target font, and a font refinement stage to improve the appearances and local details for the generated font images. Based on this multi-stage generative process, we construct our font generation framework, named MSD-Font, with a dual-network approach to generate high-quality font images using a few reference samples. The promising experimental results verify the effectiveness of our proposed model.", + "bbox": [ + 496, + 580, + 890, + 773 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations: As our MSD-Font based on diffusion models, there is an increase in the inference time and model size compared with previous GAN-based FFG methods. Many recent works have focused on improving the efficiency and model compactness of diffusion models, and most of these methods can be used in our model.", + "bbox": [ + 496, + 775, + 890, + 857 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements: This work is supported by the National Key R&D Program of China (NO. 2022ZD0160102).", + "bbox": [ + 498, + 858, + 890, + 887 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6899", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18208-18218, 2022. 3", + "[2] Samaneh Azadi, Matthew Fisher, Vladimir G Kim, Zhaowen Wang, Eli Shechtman, and Trevor Darrell. Multi-content gan for few-shot font style transfer. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7564-7573, 2018. 3", + "[3] Junbum Cha, Sanghyuk Chun, Gayoung Lee, Bado Lee, Seonghyeon Kim, and Hwalsuk Lee. Few-shot compositional font generation with dual memory. In European Conference on Computer Vision, pages 735-751. Springer, 2020. 3", + "[4] Jie Chang, Yujun Gu, Ya Zhang, Yan-Feng Wang, and CM Innovation. Chinese handwriting imitation with hierarchical generative adversarial network. In BMVC, page 290, 2018. 2, 3", + "[5] Xinyuan Chen, Yangchen Xie, Li Sun, and Yue Lu. Dgfont++: Robust deformable generative networks for unsupervised font generation. arXiv preprint arXiv:2212.14742, 2022.3", + "[6] Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. Ilvr: Conditioning method for denoising diffusion probabilistic models. arXiv preprint arXiv:2108.02938, 2021. 3", + "[7] Jooyoung Choi, Jungbeom Lee, Chaehun Shin, Sungwon Kim, Hyunwoo Kim, and Sungroh Yoon. Perception prioritized training of diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11472-11481, 2022. 2", + "[8] Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8789-8797, 2018. 2", + "[9] Bin Fu, Junjun He, Jianjun Wang, and Yu Qiao. Neural transformation fields for arbitrary-styled font generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22438-22447, 2023. 3, 7, 8", + "[10] Yue Gao, Yuan Guo, Zhouhui Lian, Yingmin Tang, and Jianguo Xiao. Artistic glyph image synthesis via one-stage few-shot learning. ACM Transactions on Graphics (TOG), 38(6): 1-12, 2019. 3", + "[11] Haibin He, Xinyuan Chen, Chaoyue Wang, Juhua Liu, Bo Du, Dacheng Tao, and Yu Qiao. Diff-font: Diffusion model for robust one-shot font generation. arXiv preprint arXiv:2212.05895, 2022. 3, 7, 8", + "[12] Xiao He, Mingrui Zhu, Nannan Wang, Xinbo Gao, and Heng Yang. Few-shot font generation by learning style difference and similarity. arXiv preprint arXiv:2301.10008, 2023. 3", + "[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3, 5" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 2", + "[15] Yue Jiang, Zhouhui Lian, Yingmin Tang, and Jianguo Xiao. Dcfont: an end-to-end deep chinese font generation system. In SIGGRAPH Asia 2017 Technical Briefs, pages 1-4. 2017. 3", + "[16] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 4", + "[17] Yuxin Kong, Canjie Luo, Weihong Ma, Qiyuan Zhu, Shenggao Zhu, Nicholas Yuan, and Lianwen Jin. Look closer to supervise better: One-shot font generation via component-based discriminator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13482-13491, 2022. 1, 3", + "[18] Mingi Kwon, Jaeseok Jeong, and Youngjung Uh. Diffusion models already have a semantic latent space. arXiv preprint arXiv:2210.10960, 2022. 2", + "[19] Guan-Horng Liu, Arash Vahdat, De-An Huang, Evangelos A Theodorou, Weili Nie, and Anima Anandkumar. I²sb: Image-to-image schrödinger bridge. arXiv preprint arXiv:2302.05872, 2023. 3", + "[20] Ming-Yu Liu, Xun Huang, Arun Mallya, Tero Karras, Timo Aila, Jaakko Lehtinen, and Jan Kautz. Few-shot unsupervised image-to-image translation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10551-10560, 2019. 2", + "[21] Wei Liu, Fangyue Liu, Fei Ding, Qian He, and Zili Yi. Xmpfont: Self-supervised cross-modality pre-training for few-shot font generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7905-7914, 2022. 3", + "[22] Ziwei Luo, Fredrik K Gustafsson, Zheng Zhao, Jens Sjolund, and Thomas B Schon. Image restoration with meanreverting stochastic differential equations. arXiv preprint arXiv:2301.11699, 2023. 3", + "[23] Pengyuan Lyu, Xiang Bai, Cong Yao, Zhen Zhu, Tengteng Huang, and Wenyu Liu. Auto-encoder guided gan for Chinese calligraphy synthesis. In 2017 14th IAPR International Conference on Document Analysis and Recognition (ICDAR), pages 1095–1100. IEEE, 2017. 3", + "[24] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 3", + "[25] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 3", + "[26] Wei Pan, Anna Zhu, Xinyu Zhou, Brian Kenji Iwana, and Shilin Li. Few shot font generation via transferring similarity guided global style and quantization local style. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 19506-19516, 2023. 3", + "[27] Song Park, Sanghyuk Chun, Junbum Cha, Bado Lee, and Hyunjung Shim. Few-shot font generation with localized" + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6900", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "style representations and factorization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2393-2402, 2021. 1, 3, 7, 8", + "[28] Song Park, Sanghyuk Chun, Junbum Cha, Bado Lee, and Hyunjung Shim. Multiple heads are better than one: Few-shot font generation with multiple localized experts. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13900-13909, 2021. 1, 3, 7, 8", + "[29] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 3", + "[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3, 4, 5, 6", + "[31] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 3", + "[32] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 3", + "[33] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 3", + "[34] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3", + "[35] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 3", + "[36] Yang Song and Stefano Ermon. Improved techniques for training score-based generative models. Advances in neural information processing systems, 33:12438-12448, 2020. 3", + "[37] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 3", + "[38] Danyang Sun, Tongzheng Ren, Chongxuan Li, Hang Su, and Jun Zhu. Learning to write stylized Chinese characters by reading a handful of examples. In Proceedings of the 27th International Joint Conference on Artificial Intelligence, pages 920–927, 2018. 3", + "[39] Licheng Tang, Yiyang Cai, Jiaming Liu, Zhibin Hong, Mingming Gong, Minhu Fan, Junyu Han, Jingtuo Liu, Errui Ding, and Jingdong Wang. Few-shot font generation by learning fine-grained local styles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7895-7904, 2022. 1, 3, 7, 8" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Yuchen Tian. zi2zi: Master chinese calligraphy with conditional adversarial networks. 2", + "[41] Yuchen Tian. Rewrite: Neural style transfer for chinese fonts. 2016. 2", + "[42] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 4", + "[43] Chi Wang, Min Zhou, Tiezheng Ge, Yuning Jiang, Hujun Bao, and Weiwei Xu. Cf-font: Content fusion for few-shot font generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1858-1867, 2023. 3", + "[44] Tengfei Wang, Ting Zhang, Bo Zhang, Hao Ouyang, Dong Chen, Qifeng Chen, and Fang Wen. Pretraining is all you need for image-to-image translation. arXiv preprint arXiv:2205.12952, 2022.3", + "[45] Yangchen Xie, Xinyuan Chen, Li Sun, and Yue Lu. Dgfont: Deformable generative networks for unsupervised font generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5130-5140, 2021. 3, 7, 8", + "[46] Zhenhua Yang, Dezhi Peng, Yuxin Kong, Yuyi Zhang, Cong Yao, and Lianwen Jin. Fontdiffuser: One-shot font generation via denoising diffusion with multi-scale content aggregation and style contrastive learning. In Proceedings of the AAAI conference on artificial intelligence, 2024.", + "[47] Mingshuai Yao, Yabo Zhang, Xianhui Lin, Xiaoming Li, and Wangmeng Zuo. Vq-font: Few-shot font generation with structure-aware enhancement and quantization. arXiv preprint arXiv:2308.14018, 2023. 3", + "[48] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6", + "[49] Yexun Zhang, Ya Zhang, and Wenbin Cai. Separating style and content for generalized style transfer. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8447-8455, 2018. 3", + "[50] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 2", + "[51] Mingrui Zhu, Xiao He, Nannan Wang, Xiaoyu Wang, and Xinbo Gao. All-to-key attention for arbitrary style transfer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23109-23119, 2023. 2", + "[52] Jianyi Wang Zongsheng Yue and Chen Change Loy. Reshift: Efficient diffusion model for image superresolution by residual shifting. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 3" + ], + "bbox": [ + 501, + 92, + 890, + 825 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6901", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Generate Like Experts_ Multi-Stage Font Generation by Incorporating Font Transfer Process into Diffusion Models/bfb6c4d4-ce31-4c6d-ba23-750f3d411ecc_model.json b/2024/Generate Like Experts_ Multi-Stage Font Generation by Incorporating Font Transfer Process into Diffusion Models/bfb6c4d4-ce31-4c6d-ba23-750f3d411ecc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1b962732f8b05644db79767a3186d9caf2328989 --- /dev/null +++ b/2024/Generate Like Experts_ Multi-Stage Font Generation by Incorporating Font Transfer Process into Diffusion Models/bfb6c4d4-ce31-4c6d-ba23-750f3d411ecc_model.json @@ -0,0 +1,2420 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.129, + 0.13, + 0.844, + 0.175 + ], + "angle": 0, + "content": "Generate Like Experts: Multi-Stage Font Generation by Incorporating Font Transfer Process into Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.136, + 0.203, + 0.842, + 0.223 + ], + "angle": 0, + "content": "Bin Fu\\(^{1}\\), Fanghua Yu\\(^{1}\\), Anran Liu\\(^{3}\\), Zixuan Wang\\(^{1}\\), Jie Wen\\(^{4}\\), Junjun He\\(^{2}\\), and Yu Qiao\\(^{1,2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.176, + 0.239, + 0.794, + 0.309 + ], + "angle": 0, + "content": "\\(^{1}\\)ShenZhen Key Lab of Computer Vision and Pattern Recognition, Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences \\(^{2}\\)Shanghai Artificial Intelligence Laboratory \\(^{3}\\) The University of Hong Kong \\(^{4}\\)Harbin Institute of Technology, Shenzhen" + }, + { + "type": "text", + "bbox": [ + 0.166, + 0.311, + 0.804, + 0.345 + ], + "angle": 0, + "content": "\\(^{1}\\{\\) bin.fu, zx.wang8, yu.qiao\\}@siat.ac.cn, \\(^{2}\\{\\) hejunjun, qiaoyu\\}@pjlab.org.cn \\(^{3}\\) liuar616@connect.hku.hk, \\(^{4}\\) jiewen-pr@126.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.38, + 0.314, + 0.396 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.415, + 0.473, + 0.854 + ], + "angle": 0, + "content": "Few-shot font generation (FFG) produces stylized font images with a limited number of reference samples, which can significantly reduce labor costs in manual font designs. Most existing FFG methods follow the style-content disentanglement paradigm and employ the Generative Adversarial Network (GAN) to generate target fonts by combining the decoupled content and style representations. The complicated structure and detailed style are simultaneously generated in those methods, which may be the sub-optimal solutions for FFG task. Inspired by most manual font design processes of expert designers, in this paper, we model font generation as a multi-stage generative process. Specifically, as the injected noise and the data distribution in diffusion models can be well-separated into different sub-spaces, we are able to incorporate the font transfer process into these models. Based on this observation, we generalize diffusion methods to model font generative process by separating the reverse diffusion process into three stages with different functions: The structure construction stage first generates the structure information for the target character based on the source image, and the font transfer stage subsequently transforms the source font to the target font. Finally, the font refinement stage enhances the appearances and local details of the target font images. Based on the above multistage generative process, we construct our font generation framework, named MSD-Font, with a dual-network approach to generate font images. The superior performance demonstrates the effectiveness of our model. The code is available at: https://github.com/fubinfb/MSD-Font." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.38, + 0.632, + 0.396 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.411, + 0.895, + 0.683 + ], + "angle": 0, + "content": "Designing new fonts is time-consuming and labor-intensive work for humans, especially for some glyph-rich scripts (e.g., Chinese and Korean). The few-shot font generation (FFG) task aims to automatically generate new stylized fonts with a few reference images, and it has received significant interest due to its academic and commercial values. As the requirement of this task is to render the character images into the target font style, current FFG methods [17, 27, 28, 39] usually employ the Generative Adversarial Network (GAN) to perform few-shot font generation under the style-content disentanglement paradigm. These models usually simultaneously generate the complicated structure and detailed style of target font images by combining the content features from source characters and the style representations from reference font images. Although many methods have been proposed to obtain better content and style representations, the generated font images still contain artifacts, inconsistent styles, and broken structures." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.896, + 0.902 + ], + "angle": 0, + "content": "Inspired by the common practice of expert designers, we find the following characteristics that may be crucial for high-quality font generation: As font design requires preserving the global structure of the target characters, designers usually generate a new font by modifying from a similar existing font, which can be viewed as a transfer process from the source font to the target font. Thus, most manual font designs can be regarded as a multi-stage process, including template font generation for global structure, font transfer from the template font to the target font, and refinement of fine-grained details. In light of this, we think that the multi-stage generative process is an effective solution for high-quality font generation, since the problems of broken structures, inconsistent styles, and artifacts can be" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.272, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author: Yu Qiao" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "6892" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.139, + 0.088, + 0.833, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.254, + 0.895, + 0.353 + ], + "angle": 0, + "content": "Figure 1. The forward process of our proposed model and the traditional diffusion model (DDPM). The first and third rows display the traditional forward process for the source image \\( I_{c} \\) and target font image \\( I_{g} \\). The images with red boxes mark the path of our proposed diffusion process. In our model, we incorporate the font transfer process in time interval \\( (t_1, t_2] \\), and align the distribution with the traditional diffusion model at the timesteps \\( t_1 \\) and \\( t_2 \\). With this forward process, the generative process of our model can be separated into three stages: the structure construction stage, the font transfer stage, and the font refinement stage, respectively. Note that the font images in this figure are visualized from the diffusion model in image space for better illustration1, while we implement our model in latent space to reduce computational consumption." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.36, + 0.347, + 0.374 + ], + "angle": 0, + "content": "better handled by the above three stages." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.376, + 0.471, + 0.542 + ], + "angle": 0, + "content": "The diffusion model provides a powerful generative framework to achieve this multi-stage process, based on the following two features: (1). Since the injected noise and the data distribution in forward process can be well-separated into different sub-spaces, we are able to incorporate the font transfer process into diffusion models. (2). Recent studies [7, 18] have revealed that the different steps in the generative process have different functions for image synthesis. Therefore, modeling font generation as a transfer process and constructing a multi-stage generative process based on diffusion model is a promising direction for FFG task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.544, + 0.471, + 0.862 + ], + "angle": 0, + "content": "Motivated by the above observations, in this paper, we incorporate the font transfer process into the diffusion model and formulate font generation as a multi-stage generative process. To alleviate computational consumption, we construct this diffusion process in latent space. Specifically, in latent space, we first separate the injected noise and the data distribution into different sub-spaces, and then incorporate the font transfer process into the data sub-space in a predefined time interval \\((t_1,t_2]\\). Combining with the denoising process, the font generative process in our proposed diffusion model can be automatically separated into a multistage process (shown in Fig. 1): the structure construction stage, the font transfer stage, and the font refinement stage, respectively. The structure construction stage first generates the structure information for the target character based on the source image. Then, the font transfer stage gradually transforms the source font to the target font. Finally, the font refinement stage improves the appearances and local details of the generated font images. Under this scheme, the structure construction stage generates the global character structure from source images, while the font transfer stage" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.36, + 0.895, + 0.466 + ], + "angle": 0, + "content": "provides a long generative path to produce detailed styles. Moreover, as the prediction network has different behaviors in font transfer stage and font refinement stage, we further propose a dual-network approach to predict the target latent features for different stages. Experiments demonstrate that our model achieves the superior performance on FFG tasks, verifying the effectiveness of our proposed approach." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.467, + 0.892, + 0.495 + ], + "angle": 0, + "content": "In summary, our contributions in this paper are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.497, + 0.892, + 0.571 + ], + "angle": 0, + "content": "(1). Instead of simultaneously generating font images with complicated structures and detailed styles in most traditional FFG methods, we reformulate the font generation as a multi-stage generative process by incorporating font transfer process into the diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.573, + 0.892, + 0.632 + ], + "angle": 0, + "content": "(2). We construct the multi-stage font generation framework, named MSD-Font, with a dual-network approach to better deal with different behaviors in different generative stages." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.634, + 0.892, + 0.679 + ], + "angle": 0, + "content": "(3). Experimental results show that our method achieves superior performance on the FFG task, demonstrating the effectiveness of our proposed method." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.497, + 0.892, + 0.679 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.695, + 0.651, + 0.71 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.721, + 0.892, + 0.751 + ], + "angle": 0, + "content": "This section offers a concise review of recent development in font generation methods and diffusion models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.756, + 0.75, + 0.772 + ], + "angle": 0, + "content": "2.1. Many-shot Font Generation" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.781, + 0.892, + 0.901 + ], + "angle": 0, + "content": "As font generation can be formulated as mapping font images from the source domain to the target domain, early methods [4, 40, 41] usually utilize the image-to-image (I2I) methods [8, 14, 20, 50, 51] to learn the mapping function between different fonts. The mapping function will be fine-tuned on hundreds of reference samples (many-shot) to generate unseen font images. Based on the pix2pix [14] framework, Zi2zi [40] and Rewrite [41] utilize one-hot style" + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.875, + 0.471, + 0.902 + ], + "angle": 0, + "content": "1Specifically, we replace \\( z_0^c \\) and \\( z_0^g \\) with \\( I_{c} \\) and \\( I_{g} \\) in Eq. (8)-(12) to visualize our proposed diffusion process in this figure." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6893" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.244 + ], + "angle": 0, + "content": "labels to optimize font generation models in a supervised manner. HGAN [4] further extends Zi2zi model via a transfer network and the hierarchical adversarial discriminator. DC-font [15] utilizes a feature reconstruction network to embed style information for font synthesis. AGEN [23] employs the auto-encoder to transfer standard font to calligraphy font. Although these many-shot font generation methods can generate promising font images, collecting large numbers of reference samples to fine-tune the mapping function is still a difficult task for font generation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.254, + 0.315, + 0.269 + ], + "angle": 0, + "content": "2.2. Few-shot Font Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.277, + 0.473, + 0.822 + ], + "angle": 0, + "content": "To relieve the weakness of the above models, few-shot font generation (FFG) methods have become popular solutions in recent years, which generate font images with a few samples. As font styles have complex structures and fine-grained details, the common-used statistics-based style transfer techniques are not suitable for font generation. Instead, recent FFG models [2, 3, 10, 17, 21, 26-28, 39, 45-47, 49] mainly follow the style-content disentanglement paradigm, which transfers the target font styles by combining the content representations of source characters with the style codes of reference samples. SA-VAE [38], EMD [49], and AGISNet [10] extract the style and content representations as global features to generate target font images. DM-Font [3], LF-Font [27], and MX-Font [28] decompose characters into components and learn component-wise representations to improve local details. CG-GAN [17] supervises the font generator to decouple content and style at the component level via a component-wise discriminator. Diff-Font [11] proposes a stroke-wise diffusion model to generate font images under the guidance of three character attributes (content, style, and strokes). FsFont [39] utilizes a cross-attention mechanism to aggregate style features into the fine-grained style representation. XMP-Font [21] develops a self-supervised cross-modality pre-training strategy and a cross-modality transformer-based encoder to model the style representations at all levels. DS-Font [12] proposes a cluster-level contrastive style loss to learn better style representations. DG-Font [45] and its extended version DG-Font++ [5] propose a feature deformation skip connection to learn geometric displacement maps between different fonts. CF-Font [43] further extends DG-Font to construct a robust content feature by projecting the content feature into a linear space defined by the content features of basis fonts. Besides, NTF [9] models font generation as a continuous transformation process with a neural transformation field." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.348, + 0.848 + ], + "angle": 0, + "content": "2.3. Diffusion Probabilistic Models" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Diffusion model (DM) provides a powerful probabilistic generative framework for high-quality image synthesis, which is first introduced in [33] and subsequently improved" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.456 + ], + "angle": 0, + "content": "by Denoising Diffusion Probabilistic Model (DDPM) [13], Denoising Diffusion Implicit Model (DDIM) [34], score-based diffusion [35], and other techniques [25, 36, 37]. Early diffusion models construct the diffusion process in RGB space, which requires huge computation powers to generate high-resolution images. Latent Diffusion Model (LDM) [30] builds the diffusion process in a lower-dimensional latent space, significantly reducing the computation cost. The diffusion model has become the dominant solution for various computer vision tasks, such as image synthesis [29, 30, 32], image editing [1, 6, 24], and image-to-image translation [31, 44]. Large-scale text-to-image (T2I) diffusion models (e.g. DALL-E 2 [29], Imagegen [32], Stable Diffusion [30]) have shown superior performance on high-quality image synthesis, which employ text embedding generated by large language models (LLM) as the condition to guide image generation. Recently, several works [19, 22, 52] modify the denoising diffusion process to model the residual between two given distributions. Different from these works, in this paper, we introduce the font transfer process into the diffusion model at a predefined region, forming a multi-stage diffusion process. Moreover, based on this process, a novel dual-network approach is proposed to predict latent features in different stages." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.475, + 0.637, + 0.493 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.503, + 0.893, + 0.581 + ], + "angle": 0, + "content": "In this section, we present our proposed method in detail, including a short background of the diffusion model (in Sec. 3.1), the proposed latent diffusion model with font transfer process (in Sec. 3.2), the multi-stage font generative process (in Sec. 3.3), and the overall framework (in Sec. 3.4)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.596, + 0.763, + 0.613 + ], + "angle": 0, + "content": "3.1. Background and Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.621, + 0.893, + 0.682 + ], + "angle": 0, + "content": "The diffusion model is a probabilistic generative model aiming to learn the data distribution by gradually denoising a variable sampled from a Gaussian distribution, which includes a forward process and a reverse process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.685, + 0.894, + 0.761 + ], + "angle": 0, + "content": "Given a data distribution \\( x_0 \\sim q(x_0) \\), the Forward Process (also named as the diffusion process) is fixed to a Markov chain that gradually injects Gaussian noise to the data at time \\( t \\) with a predefined variance schedule \\( \\beta_t \\) according to the equations:" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.776, + 0.892, + 0.817 + ], + "angle": 0, + "content": "\\[\nq \\left(x _ {1}, \\dots , x _ {T} \\mid x _ {0}\\right) = \\prod_ {t = 1} ^ {T} q \\left(x _ {t} \\mid x _ {t - 1}\\right), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.821, + 0.892, + 0.84 + ], + "angle": 0, + "content": "\\[\nq \\left(x _ {t} \\mid x _ {t - 1}\\right) = \\mathcal {N} \\left(x _ {t}; \\sqrt {1 - \\beta_ {t}} x _ {t - 1}, \\beta_ {t} I\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.901 + ], + "angle": 0, + "content": "where \\(x_{1},\\dots ,x_{T}\\) are latent variables. With the sufficiently large \\(T\\) , the distribution of \\(x_{T}\\) will become an isotropic Gaussian distribution. With the reparameterization trick" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6894" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.092, + 0.506, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.221, + 0.508, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.092, + 0.887, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.519, + 0.268, + 0.68, + 0.305 + ], + "angle": 0, + "content": "\\(\\tilde{z}_t\\) Latent feature in our model \n\\(z_t^c\\) Latent feature of source \nimage in traditional model" + }, + { + "type": "image_footnote", + "bbox": [ + 0.52, + 0.308, + 0.682, + 0.33 + ], + "angle": 0, + "content": "Generative Step in Structure Construction Stage" + }, + { + "type": "image_footnote", + "bbox": [ + 0.521, + 0.333, + 0.662, + 0.354 + ], + "angle": 0, + "content": "Generative Step in Font Transfer Stage" + }, + { + "type": "list", + "bbox": [ + 0.519, + 0.268, + 0.682, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.701, + 0.269, + 0.882, + 0.307 + ], + "angle": 0, + "content": "\\(\\widetilde{Z}_{\\theta_i}^{(g,i)}\\) Prediction network \\((i\\in \\{1,2\\})\\) Latent feature of target image \\(z_{t}^{g}\\) in traditional model" + }, + { + "type": "image_footnote", + "bbox": [ + 0.705, + 0.31, + 0.877, + 0.331 + ], + "angle": 0, + "content": "The Equivalent Process for Structure Construction Stage" + }, + { + "type": "image_footnote", + "bbox": [ + 0.705, + 0.333, + 0.849, + 0.354 + ], + "angle": 0, + "content": "Generative Step in Font Refinement Stage" + }, + { + "type": "list", + "bbox": [ + 0.701, + 0.269, + 0.882, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.36, + 0.895, + 0.493 + ], + "angle": 0, + "content": "Figure 2. (a). The forward diffusion process of our proposed model. (b). The reverse process of our proposed model. Our proposed diffusion process is marked as black circles. The \\(\\tilde{z}_t\\) denotes the latent features in our diffusion process, while \\(z_t^c\\) and \\(z_t^g\\) denote the latent features of the source image and the target image in traditional diffusion process, respectively. As we have aligned our diffusion process with the traditional diffusion process of \\(z^g\\) and \\(z^c\\) at \\(t_1\\) and \\(t_2\\), our model shares the same process with the traditional diffusion process in the region \\((0, t_1]\\) and \\((t_2, T]\\). (c). Overall framework of our font generation model, including the VQ-VAE encoder \\(\\mathcal{E}\\) and decoder \\(\\mathcal{D}\\), the dual prediction networks \\(\\tilde{z}_{\\theta_1}^{(g,1)}(\\tilde{z}_t, t, y_1)\\) and \\(\\tilde{z}_{\\theta_2}^{(g,2)}(\\tilde{z}_t, t, y_2)\\), and the corresponding style and content encoders for each prediction network. Due to the same structure, in this figure, we only plot one prediction network \\(\\tilde{z}_{\\theta_i}^{(g,i)}\\) (\\(i \\in \\{1, 2\\}\\)) with its corresponding style encoder \\(E_s^i\\) and content encoder \\(E_c^i\\), where \\(i\\) is determined by the generative stage of timestep \\(t\\). We use \\(I_g\\) and \\(I_G\\) to denote the target image and generated image, respectively." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.507, + 0.371, + 0.523 + ], + "angle": 0, + "content": "[16], the latent variable \\( x_{t} \\) can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.531, + 0.47, + 0.567 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} x _ {t} = \\sqrt {\\alpha_ {t}} x _ {t - 1} + \\sqrt {1 - \\alpha_ {t}} \\epsilon_ {t - 1} ^ {*}, (3) \\\\ = \\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {0}, (4) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.575, + 0.469, + 0.621 + ], + "angle": 0, + "content": "where \\(\\epsilon_{t - 1}^{*},\\epsilon_{0}\\sim \\mathcal{N}(0,I),\\alpha_{t} = 1 - \\beta_{t}\\) and \\(\\bar{\\alpha}_t = \\prod_{s = 1}^t\\alpha_s\\) Thus sampling \\(x_{t}\\) at arbitrary timestep has a closed form \\(q(x_{t}|x_{0}) = \\mathcal{N}(x_{t};\\sqrt{\\bar{\\alpha}_{t}} x_{0},(1 - \\bar{\\alpha}_{t})I)\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.621, + 0.47, + 0.667 + ], + "angle": 0, + "content": "The Reverse Process generates data from isotropic Gaussian noise \\( x_{T} \\), and gradually recovers \\( x_{0} \\) via the reverse distribution \\( p_{\\theta}(x_{t - 1}|x_t) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.674, + 0.47, + 0.705 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(x _ {0}\\right) = \\int p _ {\\theta} \\left(x _ {0: T}\\right) d x _ {1: T}, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.708, + 0.47, + 0.749 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(x _ {0: T}\\right) = p _ {\\theta} \\left(x _ {T}\\right) \\prod_ {t = 1} ^ {T} p _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.755, + 0.47, + 0.786 + ], + "angle": 0, + "content": "where \\(p_{\\theta}(x_{t - 1}|x_t) = \\mathcal{N}(x_{t - 1}|\\mu_\\theta (x_t,t),\\Sigma_\\theta (x_t,t))\\) is a parameterized Gaussian transition network, and defined as" + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.792, + 0.47, + 0.825 + ], + "angle": 0, + "content": "\\[\n\\mu_ {\\theta} (x _ {t}, t) = \\frac {\\sqrt {\\alpha_ {t}} (1 - \\bar {\\alpha} _ {t - 1}) x _ {t} + \\sqrt {\\bar {\\alpha} _ {t - 1}} (1 - \\alpha_ {t}) \\hat {x} _ {\\theta} (x _ {t} , t)}{1 - \\bar {\\alpha} _ {t}},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.831, + 0.47, + 0.878 + ], + "angle": 0, + "content": "and \\(\\Sigma_{\\theta}(x_t,t) = (1 - \\bar{\\alpha}_{t - 1})\\beta_tI / (1 - \\bar{\\alpha}_t)\\). \\(\\hat{x}_{\\theta}(x_{t},t)\\) is a function approximator for predicting \\(x_0\\) from \\(x_{t}\\) and \\(t\\). The corresponding objective function can be simplified to:" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.884, + 0.47, + 0.904 + ], + "angle": 0, + "content": "\\[\nL = \\mathbb {E} _ {x _ {0}, \\epsilon , t} \\left[ \\| \\hat {x} _ {\\theta} (x _ {t}, t) - x _ {0} \\| _ {2} ^ {2} \\right]. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.507, + 0.892, + 0.612 + ], + "angle": 0, + "content": "Latent Diffusion Model (LDM) [30] further employs a VQ-VAE [42] encoder \\(\\mathcal{E}\\) to project image \\(x_0\\) into latent space \\(z_0\\), and constructs the diffusion model in this lower-dimensional latent space, significantly reducing computational consumption. Once \\(z_0\\) is sampled via the reverse process, the VQ-VAE decoder \\(\\mathcal{D}\\) will project it back to the image space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.892, + 0.659 + ], + "angle": 0, + "content": "Since LDM significantly reduces computation consumption and maintains high-quality image synthesis, in this paper, we construct our diffusion process in latent space." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.666, + 0.866, + 0.681 + ], + "angle": 0, + "content": "3.2. Latent Diffusion Model with Font Transfer" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.825 + ], + "angle": 0, + "content": "As shown in Eq. (4), in the forward process, the data distribution \\( x_0 \\) and the injected noise can be well-separated in arbitrary timesteps. This characteristic enables us to construct a multi-stage font generative framework in latent diffusion model, by incorporating the font transfer process into the data sub-space in a predefined time interval \\( t \\in (t_1, t_2] \\). Based on this observation, we develop our latent diffusion model in this section, then discuss the multi-stage font generative process in the next section." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.901 + ], + "angle": 0, + "content": "The Forward Process. After projecting the source images \\(I_{c}\\) and target images \\(I_{g}\\) into the latent features \\(z_0^c\\) and \\(z_0^g\\) by the VQ-VAE encoder \\(\\mathcal{E}\\), we implement our model in latent space. As shown in Fig. 2(a), our proposed model (marked as black circles and denoted as \\(\\tilde{z}_t\\)) is constructed" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6895" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.184 + ], + "angle": 0, + "content": "by utilizing a font transfer process to connect the traditional diffusion processes (DDPM) of the latent features \\( z_{t}^{c} \\) and \\( z_{t}^{g} \\). Therefore, our latent diffusion process \\( \\tilde{z}_{t} \\) can be divided into three regions: (1). In the region \\( t \\in (0, t_1] \\), only latent feature \\( z_{t}^{g} \\) evolved with noise. The forward latent \\( \\tilde{z}_{t} \\) in this region can be expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.192, + 0.47, + 0.211 + ], + "angle": 0, + "content": "\\[\n\\tilde {z} _ {t} = z _ {t} ^ {g} = \\sqrt {\\bar {\\alpha} _ {t}} z _ {0} ^ {g} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {0} t \\in (0, t _ {1} ]. \\qquad (8)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.219, + 0.47, + 0.31 + ], + "angle": 0, + "content": "(2). In the region \\(t \\in (t_1, t_2]\\), together with the continuously injected noise, the latent feature \\(z_t^g\\) of the target font is gradually replaced by the latent feature \\(z_t^c\\) of the source character, and only \\(z_t^c\\) exists at \\(t_2\\). After careful derivation (provided in Supplementary Materials), the forward latent \\(\\tilde{z}_t\\) in this region can be expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.32, + 0.469, + 0.339 + ], + "angle": 0, + "content": "\\[\n\\tilde {z} _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} z _ {0} ^ {g} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {0} - \\left(z _ {0} ^ {g} - z _ {0} ^ {c}\\right) \\Psi \\mathcal {H} _ {t}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.347, + 0.47, + 0.409 + ], + "angle": 0, + "content": "for \\( t \\in (t_1, t_2] \\). The notations \\( \\mathcal{H}_t \\) and \\( \\Psi \\) are coefficients obtained by aligning our model with traditional diffusion processes (DDPM) of \\( z^g \\) and \\( z^c \\) at timesteps \\( t_1 \\) and \\( t_2 \\), respectively. They are defined as" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.418, + 0.469, + 0.436 + ], + "angle": 0, + "content": "\\[\n\\mathcal {H} _ {t} = 1 + \\sqrt {\\alpha_ {t}} + \\sqrt {\\alpha_ {t} \\alpha_ {t - 1}} + \\dots + \\sqrt {\\alpha_ {t} \\cdots \\alpha_ {t + 2}}, \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.439, + 0.469, + 0.477 + ], + "angle": 0, + "content": "\\[\n\\Psi = \\frac {\\sqrt {\\tilde {\\alpha} (1 , t _ {2})}}{1 + \\sum_ {m = t _ {1} + 2} ^ {t _ {2}} \\sqrt {\\tilde {\\alpha} (m , t _ {2})}}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.487, + 0.47, + 0.564 + ], + "angle": 0, + "content": "where \\(\\tilde{\\alpha}(t_1, t_2) = \\prod_{i=t_1}^{t_2} \\alpha_i\\). \\(\\mathcal{H}_t\\) can be further reduced to \\(\\mathcal{H}_t = 1 + \\sqrt{\\alpha_t} \\mathcal{H}_{t-1}\\). (3). In the region \\(t \\in (t_2, T]\\), the injected noise finally disrupts the latent feature \\(z_t^c\\) to Gaussian noise, and the forward latent \\(\\tilde{z}_t\\) in this region can be expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.574, + 0.469, + 0.592 + ], + "angle": 0, + "content": "\\[\n\\tilde {z} _ {t} = z _ {t} ^ {c} = \\sqrt {\\bar {\\alpha} _ {t}} z _ {0} ^ {c} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {0} \\quad t \\in (t _ {2}, T ]. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.601, + 0.47, + 0.692 + ], + "angle": 0, + "content": "The Reverse Process. Now we turn to discuss our choices for \\(p_{\\theta}(\\tilde{z}_{t - 1}|\\tilde{z}_t) = \\mathcal{N}(\\tilde{z}_{t - 1}|\\mu_\\theta (\\tilde{z}_t,t),\\Sigma_\\theta (\\tilde{z}_t,t))\\). As shown in Eq. (8) and Eq. (12), in the time region \\((0,t_1]\\) and \\((t_2,T]\\), we have aligned the forward latent \\(\\tilde{z}_t\\) with the traditional diffusion processes (DDPM [13]), thus they share the same parameterized mean functions: for \\(t\\in (0,t_1]\\)" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.702, + 0.469, + 0.748 + ], + "angle": 0, + "content": "\\[\n\\mu_ {\\theta} = \\frac {\\sqrt {\\alpha_ {t}} (1 - \\bar {\\alpha} _ {t - 1}) \\tilde {z} _ {t} + \\sqrt {\\bar {\\alpha} _ {t - 1}} (1 - \\alpha_ {t}) \\tilde {z} _ {\\theta} ^ {(g)} (\\tilde {z} _ {t} , t)}{1 - \\bar {\\alpha} _ {t}}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.75, + 0.469, + 0.797 + ], + "angle": 0, + "content": "where \\(\\tilde{z}_{\\theta}^{(g)}(\\tilde{z}_t,t)\\) is the prediction network for predicting \\(z_0^g\\) from noisy latent feature \\(\\tilde{z}_t\\) and time index \\(t\\), while for \\(t\\in (t_2,T]\\)" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.806, + 0.469, + 0.853 + ], + "angle": 0, + "content": "\\[\n\\mu_ {\\theta} = \\frac {\\sqrt {\\alpha_ {t}} (1 - \\bar {\\alpha} _ {t - 1}) \\tilde {z} _ {t} + \\sqrt {\\bar {\\alpha} _ {t - 1}} (1 - \\alpha_ {t}) \\tilde {z} _ {\\theta} ^ {(c)} (\\tilde {z} _ {t} , t)}{1 - \\bar {\\alpha} _ {t}}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.854, + 0.47, + 0.902 + ], + "angle": 0, + "content": "where \\(\\tilde{z}_{\\theta}^{(c)}(z_t,t)\\) is the prediction network for predicting \\(z_0^c\\) from noisy latent feature \\(\\tilde{z}_t\\) and time index \\(t\\). The parameterized mean function for the region \\(t\\in (t_1,t_2]\\) can be" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.701, + 0.107 + ], + "angle": 0, + "content": "derivated via Bayes' theorem:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.113, + 0.893, + 0.214 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mu_ {\\theta} = \\frac {\\sqrt {\\alpha_ {t}} (1 - \\bar {\\alpha} _ {t - 1})}{1 - \\bar {\\alpha} _ {t}} \\tilde {z} _ {t} + \\frac {\\sqrt {\\alpha_ {t}} (1 - \\bar {\\alpha} _ {t - 1})}{1 - \\bar {\\alpha} _ {t}} \\Psi \\tilde {z} _ {\\theta} ^ {(g)} (\\tilde {z} _ {t}, t) \\\\ + \\frac {\\left[ \\mathcal {H} _ {t - 1} (1 - \\alpha_ {t}) - \\sqrt {\\alpha_ {t}} (1 - \\bar {\\alpha} _ {t - 1}) \\right]}{1 - \\bar {\\alpha} _ {t}} \\Psi z _ {0} ^ {c} \\\\ + \\frac {(\\sqrt {\\bar {\\alpha} _ {t - 1}} - \\mathcal {H} _ {t - 1} \\Psi) (1 - \\alpha_ {t})}{1 - \\bar {\\alpha} _ {t}} \\tilde {z} _ {\\theta} ^ {(g)} (\\tilde {z} _ {t}, t). \\tag {15} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.219, + 0.892, + 0.265 + ], + "angle": 0, + "content": "As we only incorporate the font transfer process in the data sub-space and keep the noise sub-space unchanged, the variance in the reverse process remains the same as DDPM:" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.27, + 0.891, + 0.302 + ], + "angle": 0, + "content": "\\[\n\\Sigma_ {\\theta} \\left(\\tilde {z} _ {t}, t\\right) = \\sigma_ {t} ^ {2} I = \\frac {(1 - \\alpha_ {t}) (1 - \\bar {\\alpha} _ {t - 1})}{1 - \\bar {\\alpha} _ {t}} I. \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.307, + 0.892, + 0.367 + ], + "angle": 0, + "content": "Therefore, the latent \\(\\tilde{z}_{t-1}\\) in reverse process can be obtained via the formula \\(p_{\\theta}(\\tilde{z}_{t-1}|\\tilde{z}_t) = \\mathcal{N}(\\tilde{z}_{t-1};\\mu_{\\theta}(\\tilde{z}_t,t),\\sigma_t^2 I)\\), where the \\(\\mu_{\\theta}(\\tilde{z}_t,t)\\) is calculated from Eq. (13)-(15) based on different timesteps." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.368, + 0.893, + 0.538 + ], + "angle": 0, + "content": "The Optimization Objective. As the forward behaviors keep the same as DDPM in the region \\((0, t_1]\\) and \\((t_2, T]\\), the optimization objectives of \\(\\tilde{z}_{\\theta}^{(g)}(\\tilde{z}_t, t)\\) and \\(\\hat{z}_{\\theta}^{(c)}(\\tilde{z}_t, t)\\) have the same expressions as DDPM in these regions. For the optimization objective of \\(\\tilde{z}_{\\theta}^{(g)}(\\tilde{z}_t, t)\\) in the region \\((t_1, t_2]\\), it can be derived from the variational bound on negative log-likelihood [13], and only the coefficient is different with the DDPM's loss function. Moreover, as DDPM and LDM adopt the simplified training objectives in their implementations, we also simplify our loss functions and obtain the optimization objective as" + }, + { + "type": "equation", + "bbox": [ + 0.647, + 0.546, + 0.891, + 0.563 + ], + "angle": 0, + "content": "\\[\nL = L _ {g} + L _ {c}, \\tag {17}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.566, + 0.891, + 0.592 + ], + "angle": 0, + "content": "\\[\nL _ {g} = \\mathbb {E} _ {\\mathcal {E} (x _ {0}), \\epsilon , t \\sim U (0, t _ {2} ]} \\left[ \\| \\tilde {z} _ {\\theta} ^ {(g)} (\\tilde {z} _ {t}, t) - z _ {0} ^ {g} \\| _ {2} ^ {2} \\right], \\tag {18}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.593, + 0.891, + 0.619 + ], + "angle": 0, + "content": "\\[\nL _ {c} = \\mathbb {E} _ {\\mathcal {E} (x _ {0}), \\epsilon , t \\sim U (t _ {2}, T ]} \\left[ \\| \\tilde {z} _ {\\theta} ^ {(c)} (\\tilde {z} _ {t}, t) - z _ {0} ^ {c} \\| _ {2} ^ {2} \\right], \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.624, + 0.891, + 0.654 + ], + "angle": 0, + "content": "where \\( U(t_{1},t_{2}] \\) denotes the uniform distribution in region \\( (t_1,t_2] \\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.655, + 0.892, + 0.712 + ], + "angle": 0, + "content": "The Conditioning Mechanism. Following the LDM [30], we generalize our method to model conditional distributions by implementing with the conditional prediction networks:" + }, + { + "type": "equation", + "bbox": [ + 0.612, + 0.712, + 0.891, + 0.732 + ], + "angle": 0, + "content": "\\[\n\\tilde {z} _ {\\theta} ^ {(i)} \\left(\\tilde {z} _ {t}, t, y\\right), i \\in \\{g, c \\}, \\tag {20}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.735, + 0.891, + 0.763 + ], + "angle": 0, + "content": "where \\(y\\) is the condition for controlling the synthesis process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Extension to Conditional Font Generation. We further extend this conditional model to perform the font generation task. Since the few-shot font generation task aims to render the source character \\( I_{c} \\) into the target font style by a few reference samples \\( I_{s} \\), we utilize the style encoder \\( E_{s} \\) and the content encoder \\( E_{c} \\) to extract the style code from the reference images \\( I_{s} \\) and the content representations from the source images \\( I_{c} \\). The style and content features are flattened and concatenated to build condition \\( y \\)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6896" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.396, + 0.108 + ], + "angle": 0, + "content": "3.3. Multi-Stage Font Generative Process" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.122, + 0.47, + 0.182 + ], + "angle": 0, + "content": "Based on our proposed model, the font generation in the reverse process can be automatically separated into three different stages: the structure construction stage, the font transfer stage, and the font refinement stage." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.189, + 0.473, + 0.529 + ], + "angle": 0, + "content": "As shown in Fig. 2(b), starting with the sampled Gaussian noise \\(\\tilde{z}_T\\), in the region \\(t \\in (t_2, T]\\), Structure Construction Stage generates the structure information for the target character via predicting the latent feature \\(z_0^c\\) of the source image \\(I_c\\). However, since we can access the source image \\(I_c\\) during the inference, the iterative generative steps from \\(T\\) to \\(t_2\\) can be replaced by a single forward step to generate the noised latent variable \\(\\tilde{z}_{t_2}\\) at time \\(t = t_2\\) via Eq. (12). This will significantly reduce inference time and provide more accurate structure information. Font Transfer Stage performs font transformation from the timestep \\(t_2\\) to \\(t_1\\), where the noised latent feature \\(\\tilde{z}_{t_2}\\) is gradually transformed to the noised latent feature \\(\\tilde{z}_{t_1}\\) of target images via the \\(\\tilde{z}_{\\theta}^{(g)}(\\tilde{z}_t, t, y)\\). To predict the latent feature \\(z_0^g\\), the prediction network \\(\\tilde{z}_{\\theta}^{(g)}(\\tilde{z}_t, t, y)\\) in this stage is optimized to not only denoise the latent features but also perform font transfer under the condition \\(y\\). Finally, in the region \\(t \\in (0, t_1]\\), Font Refinement Stage improves appearances and local details for the generated font images via predicting \\(z_0^g\\) from the noisy latent feature \\(\\tilde{z}_t\\). Based on the above three stages, we can implement the reverse process to generate high-quality font images." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.535, + 0.47, + 0.58 + ], + "angle": 0, + "content": "To better present this multi-stage font generative process, we summarize above steps and provide the pseudo-code in Supplementary Materials." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.611, + 0.424, + 0.627 + ], + "angle": 0, + "content": "3.4. Overall Framework for Font Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.64, + 0.473, + 0.903 + ], + "angle": 0, + "content": "As discussed in Sec. 3.3, the network \\(\\tilde{z}_{\\theta}^{(c)}(\\tilde{z}_t,t,y)\\) can be replaced by a single forward step. Moreover, instead of implementing a single prediction network \\(\\tilde{z}_{\\theta}^{(g)}(\\tilde{z}_t,t,y)\\) to predict \\(z_0^g\\) we adopt a dual-network approach in this paper, since the network \\(\\tilde{z}_{\\theta}^{(g)}(\\tilde{z}_t,t,y)\\) has different behaviors in the font transfer stage and the font refinement stage. The dual networks \\(\\tilde{z}_{\\theta_1}^{(g,1)}(\\tilde{z}_t,t,y_1)\\) and \\(\\tilde{z}_{\\theta_2}^{(g,2)}(\\tilde{z}_t,t,y_2)\\) share the same structure, but are trained to predict the latent feature \\(z_0^g\\) in different regions: one for font transfer stage and the other for font refinement stage. Therefore, as shown in Fig. 2(c), the overall framework of our font generation model contains the following parts: the VQ-VAE encoder \\(\\mathcal{E}\\), the VQ-VAE decoder \\(\\mathcal{D}\\), the dual prediction networks \\(\\tilde{z}_{\\theta_1}^{(g,1)}(\\tilde{z}_t,t,y_1)\\) and \\(\\tilde{z}_{\\theta_2}^{(g,2)}(\\tilde{z}_t,t,y_2)\\), and the corresponding style and content encoders for each prediction network \\(\\tilde{z}_{\\theta_i}^{(g,i)}(\\tilde{z}_t,t,y_i)\\). Finally, from Eq. (17), the optimization objective of our font gener" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.75, + 0.106 + ], + "angle": 0, + "content": "ative model can be further reduced to" + }, + { + "type": "equation", + "bbox": [ + 0.646, + 0.112, + 0.892, + 0.132 + ], + "angle": 0, + "content": "\\[\nL = L _ {g} ^ {1} + L _ {g} ^ {2}, \\tag {21}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.51, + 0.134, + 0.892, + 0.159 + ], + "angle": 0, + "content": "\\[\nL _ {g} ^ {1} = \\mathbb {E} _ {\\mathcal {E} (x _ {0}), \\epsilon , t \\sim U (t _ {1}, t _ {2} ]} \\left[ \\| \\tilde {z} _ {\\theta_ {1}} ^ {(g, 1)} (\\tilde {z} _ {t}, t, y _ {1}) - z _ {0} ^ {g} \\| _ {2} ^ {2} \\right], \\tag {22}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.162, + 0.892, + 0.187 + ], + "angle": 0, + "content": "\\[\nL _ {g} ^ {2} = \\mathbb {E} _ {\\mathcal {E} (x _ {0}), \\epsilon , t \\sim U (0, t _ {1} ]} \\left[ \\| \\tilde {z} _ {\\theta_ {2}} ^ {(g, 2)} (\\tilde {z} _ {t}, t, y _ {2}) - z _ {0} ^ {g} \\| _ {2} ^ {2} \\right]. \\tag {23}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.193, + 0.634, + 0.21 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.218, + 0.892, + 0.248 + ], + "angle": 0, + "content": "In this section, we conduct extensive experiments to evaluate our proposed model on the FFG task." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.256, + 0.679, + 0.272 + ], + "angle": 0, + "content": "4.1. Implement Details" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.278, + 0.892, + 0.422 + ], + "angle": 0, + "content": "Our models are implemented on the Stable Diffusion [30] platform. We utilize \\( T = 1000 \\) steps to generate \\( 128 \\times 128 \\times 3 \\) font images with 8 reference samples. The VQ-VAE encoder \\( \\mathcal{E} \\) and decoder \\( \\mathcal{D} \\), the dual-network \\( \\tilde{z}_{\\theta_1}^{(g,1)}(\\tilde{z}_t,t,y_1) \\) and \\( \\tilde{z}_{\\theta_2}^{(g,2)}(\\tilde{z}_t,t,y_2) \\) are initialized by the pre-trained Stable Diffusion models. The style encoder \\( E_{s} \\) and content encoder \\( E_{c} \\) are initialized by Kaiming initialization. We keep VQ-VAE encoder \\( \\mathcal{E} \\) and decoder \\( \\mathcal{D} \\) frozen and optimize the remaining networks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.429, + 0.78, + 0.443 + ], + "angle": 0, + "content": "4.2. Dataset and Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.452, + 0.892, + 0.602 + ], + "angle": 0, + "content": "Datasets: As DM needs more training data to optimize, we collect 900 Chinese fonts to evaluate our method on FFG task. We randomly select 840 fonts as the training set (seen fonts) and the remaining 60 fonts as the testing fonts (unseen fonts). As each collected font covers different Chinese characters, we randomly sample 800 characters from the commonly-used Chinese characters (the first-level Chinese Character Table) to train our method. For the test set, we use 214 characters as the unseen contents and sample 200 characters in our training set as the seen contents." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.603, + 0.892, + 0.692 + ], + "angle": 0, + "content": "Therefore, the training set contains 840 fonts with 800 characters, while the test set includes two parts: one is the Unseen Fonts Unseen Contents (UFUC) test set, containing 60 unseen fonts with 214 unseen characters; the other is the Unseen Fonts Seen Contents (UFSC) test set, containing 60 unseen fonts with 200 seen characters." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.693, + 0.892, + 0.77 + ], + "angle": 0, + "content": "Evaluation Metrics: In this paper, we utilize four commonly used measurements to evaluate our model, including root mean squared error (RMSE), peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and learned perceptual image path similarity (LPIPS) [48]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.776, + 0.654, + 0.792 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.8, + 0.892, + 0.83 + ], + "angle": 0, + "content": "With the above dataset and metrics, we conducted extensive experiments in different settings." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.85, + 0.862 + ], + "angle": 0, + "content": "4.3.1 Ablation Study on the Overall Framework" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "We first construct a Base Model with traditional diffusion process (DDPM) to verify the effectiveness of our proposed" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6897" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.088, + 0.891, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.234, + 0.845, + 0.248 + ], + "angle": 0, + "content": "Figure 3. Visualization results of the intermediate results \\(\\tilde{z}_t\\) by projecting \\(\\tilde{z}_t\\) back to image space via VQ-VAE decoder \\(\\mathcal{D}\\)." + }, + { + "type": "image", + "bbox": [ + 0.118, + 0.26, + 0.428, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.406, + 0.47, + 0.448 + ], + "angle": 0, + "content": "Figure 4. Visualization results of the predicted latent feature \\( z_0^g \\) in the intermediate timesteps by projecting \\( z_0^g \\) back to image space via VQ-VAE decoder \\( \\mathcal{D} \\)." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.465, + 0.469, + 0.492 + ], + "angle": 0, + "content": "Table 1. Ablation study for the overall framework on UFUC dataset. The bold number indicates the best." + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.493, + 0.446, + 0.556 + ], + "angle": 0, + "content": "
MethodsRMSE↓PSNR↑SSIM↑LPIPS↓
Base Model0.267611.720.69610.1561
Single-Net0.260112.020.70960.1531
Dual-Net0.247512.450.71890.1527
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.578, + 0.469, + 0.669 + ], + "angle": 0, + "content": "latent diffusion model. Moreover, as discussed in Sec. 3.4, there are two different selections to obtain \\( z_0^g \\) in the reverse process: One utilizes a single network to predict \\( z_0^g \\) at all timesteps. The other adopts the dual networks to predict the latent feature \\( z_0^g \\) in font transfer stage and font refinement stage, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.67, + 0.469, + 0.807 + ], + "angle": 0, + "content": "As shown in Tab. 1, compared with the Base Model, our Single-Net model improves the Base Model with 0.0075, 0.30, 0.0135, and 0.0030 in terms of RMSE, PSNR, SSIM, and LPIPS, verifying the effectiveness of our proposed latent diffusion model. Moreover, the Dual-Net approach further improves the Single-Net model with 0.0126, 0.43, and 0.0093 in terms of RMSE, PSNR, and SSIM, illustrating that the use of separate models is beneficial for predicting \\( z_0^g \\) in different stages." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.83, + 0.469, + 0.859 + ], + "angle": 0, + "content": "4.3.2 Ablation Study on the Multi-Stage Font Generative Process" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "The selections of the region for different stages in the multistage generative process are hyper-parameters in our model," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.262, + 0.892, + 0.399 + ], + "angle": 0, + "content": "which significantly impact image quality. Therefore, we study the influence on model performance by selecting different \\( t_1 \\) and \\( t_2 \\). Due to the page limitation, the experimental results and the detailed discussions are provided in supplementary materials, and we briefly summarize the conclusion: The hyper-parameters \\( t_1 \\) and \\( t_2 \\) are trade-off selections, and the configuration \\( t_1 = 200 \\) and \\( t_2 = 800 \\) achieves the best record. Thus, we select this configuration for the following experiments." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.405, + 0.7, + 0.419 + ], + "angle": 0, + "content": "4.3.3 Visualization Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.429, + 0.892, + 0.459 + ], + "angle": 0, + "content": "In this section, we provide the visualization results of our generative process to verify the effectiveness of our model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.46, + 0.892, + 0.64 + ], + "angle": 0, + "content": "As our diffusion process is implemented in latent space, we visualize the intermediate generative results \\(\\tilde{z}_t\\) by using the VQ-VAE decoder \\(\\mathcal{D}\\) to project them back into the image space. As shown in Fig. 3, the images in the first and third rows are the intermediate results for source image \\(I_{c}\\) and target image \\(I_{g}\\) generated by the traditional diffusion process (DDPM), while the images in the second row are generated by our multi-stage font generative process. From this figure, we can clearly observe a font transfer process from the source font to the target font, and the font refinement stage slightly improves style consistency by modifying local details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.641, + 0.892, + 0.732 + ], + "angle": 0, + "content": "To further evaluate the font refinement stage, we also visualize the predicted \\( z_0^g \\) from different timesteps in the font refinement stage. As shown in Fig. 4, from \\( t = 200 \\) to \\( t = 0 \\), the font refinement stage can gradually enhance local details by modifying anomalous pixels, recovering unwanted structures, and improving visual appearances." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.741, + 0.892, + 0.758 + ], + "angle": 0, + "content": "4.4. Comparison with the State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this section, we compare our model with several state-of-the-art works, including LF-Font [27], DG-Font [45], MX-Font [28], FsFont [39], NTF [9], and Diff-Font [11]. LF-Font and MX-Font are component-related methods, which decompose characters into components and learn component-wise representations to improve local details. FsFont focuses on fine-grained local styles, which utilizes cross attention to aggregate the reference styles into target styles. DG-Font and NTF are shape transformation mod" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6898" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.134, + 0.091, + 0.835, + 0.251 + ], + "angle": 0, + "content": "
Content辛禁病安持理根英役破任始米保用李倍荷供唱血止在利打
LF-Font辛禁病安持理根英役破任始米保用李倍荷供唱血止在利打
DG-Font辛禁病安持理根英役破任始米保用李倍荷供唱血止在利打
MX-Font辛禁病安持理根英役破任始米保用李倍荷供唱血止在利打
FsFont辛禁病安持理根英役破任始米保用李倍荷供唱血止在利打
NTF辛禁病安持理根英役破任始米保用李倍荷供唱血止在利打
Ours辛禁病安持理根英役破任始米保用李倍荷供唱血止在利打
Ground Truth辛禁病安持理根英役破任始米保用李倍荷供唱血止在利打
" + }, + { + "type": "table_caption", + "bbox": [ + 0.154, + 0.252, + 0.815, + 0.266 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparisons of our proposed model with other state-of-the-art methods on UFUC dataset." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.278, + 0.471, + 0.319 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison of our proposed model with other state-of-the-art methods on few-shot font generation task. The bold number indicates the best." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.325, + 0.477, + 0.574 + ], + "angle": 0, + "content": "
MethodsRMSE↓PSNR↑SSIM↑LPIPS↓
Unseen Fonts and Seen Contents
LF-Font [27]0.302510.660.64690.2787
DG-Font [45]0.283311.240.66510.1479
MX-Font [28]0.300710.750.64380.1651
FsFont [39]0.244312.560.71050.1468
NTF [9]0.265011.830.68960.1517
Diff-Font2[11]0.281110.950.68460.1557
MSD-Font0.243912.610.71100.1348
Unseen Fonts and Unseen Contents
LF-Font [27]0.291011.090.67620.2564
DG-Font [45]0.275011.540.68930.1534
MX-Font [28]0.284511.290.67890.1635
FsFont [39]0.250912.290.71850.1642
NTF [9]0.257412.110.71080.1553
MSD-Font0.247512.450.71890.1527
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.588, + 0.47, + 0.664 + ], + "angle": 0, + "content": "els. DG-Font models font generation as a shape deformable problem via the deformation skip connection, while NTF models font generation as a continuous transformation process. We train above methods on our dataset and use the same reference images to generate fonts in inference." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.669, + 0.312, + 0.685 + ], + "angle": 0, + "content": "4.4.1 Quantitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.695, + 0.47, + 0.875 + ], + "angle": 0, + "content": "The quantitative results are shown in Tab. 2. From the table, we find that modeling font style in fine-grained manners by multiple localized encoders (MX-Font) or cross attention mechanism (FsFont) can improve the quality of generated font images. The shape transformation approaches (DG-Font and NTF) are also promising solutions for FFG task. Our MSD-Font achieves superior performance on both UFUC and UFSC datasets, which demonstrates the effectiveness of our proposed method. Specifically, compared with the NTF on the UFUC dataset, our method improves the generation performance with 0.0099, 0.34, 0.0081, and 0.0026 in terms of RMSE, PSNR, SSIM, and LPIPS." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.876, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Diff-Font is one-shot font generation model, we randomly select one image from 8 reference samples to calculate style code." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.28, + 0.724, + 0.296 + ], + "angle": 0, + "content": "4.4.2 Qualitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.314, + 0.895, + 0.527 + ], + "angle": 0, + "content": "To further evaluate our methods, as shown in Fig. 5, we visualize the generated font images in UFUC setting. LF-Font sometimes generates font images with some broken structures. The DG-Font keeps the global structure for the target character, but the generated images may contain some unwanted local structures and artifacts. FsFont utilizes several carefully selected reference characters to generate font images, thus sometimes it will fail to produce accurate structures for target images with randomly selected reference characters. The MX-Font and NTF also contain some missing strokes and distorted local details. Our MSD-Font is able to generate high-quality font images, which achieves higher visual quality than other state-of-the-art methods in terms of style consistency and structure completeness." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.551, + 0.62, + 0.567 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.892, + 0.775 + ], + "angle": 0, + "content": "In this paper, we propose a multi-stage font generative process for few-shot font generation task, by incorporating font transfer process into the diffusion model. The proposed font generation model includes three stages: a structure construction stage to generate global structures from the source images, a font transfer stage to transform the source font into the target font, and a font refinement stage to improve the appearances and local details for the generated font images. Based on this multi-stage generative process, we construct our font generation framework, named MSD-Font, with a dual-network approach to generate high-quality font images using a few reference samples. The promising experimental results verify the effectiveness of our proposed model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.776, + 0.892, + 0.858 + ], + "angle": 0, + "content": "Limitations: As our MSD-Font based on diffusion models, there is an increase in the inference time and model size compared with previous GAN-based FFG methods. Many recent works have focused on improving the efficiency and model compactness of diffusion models, and most of these methods can be used in our model." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.859, + 0.892, + 0.888 + ], + "angle": 0, + "content": "Acknowledgements: This work is supported by the National Key R&D Program of China (NO. 2022ZD0160102)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6899" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18208-18218, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.243 + ], + "angle": 0, + "content": "[2] Samaneh Azadi, Matthew Fisher, Vladimir G Kim, Zhaowen Wang, Eli Shechtman, and Trevor Darrell. Multi-content gan for few-shot font style transfer. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7564-7573, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.472, + 0.312 + ], + "angle": 0, + "content": "[3] Junbum Cha, Sanghyuk Chun, Gayoung Lee, Bado Lee, Seonghyeon Kim, and Hwalsuk Lee. Few-shot compositional font generation with dual memory. In European Conference on Computer Vision, pages 735-751. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.315, + 0.472, + 0.37 + ], + "angle": 0, + "content": "[4] Jie Chang, Yujun Gu, Ya Zhang, Yan-Feng Wang, and CM Innovation. Chinese handwriting imitation with hierarchical generative adversarial network. In BMVC, page 290, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.373, + 0.472, + 0.427 + ], + "angle": 0, + "content": "[5] Xinyuan Chen, Yangchen Xie, Li Sun, and Yue Lu. Dgfont++: Robust deformable generative networks for unsupervised font generation. arXiv preprint arXiv:2212.14742, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.43, + 0.472, + 0.485 + ], + "angle": 0, + "content": "[6] Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. Ilvr: Conditioning method for denoising diffusion probabilistic models. arXiv preprint arXiv:2108.02938, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.488, + 0.472, + 0.557 + ], + "angle": 0, + "content": "[7] Jooyoung Choi, Jungbeom Lee, Chaehun Shin, Sungwon Kim, Hyunwoo Kim, and Sungroh Yoon. Perception prioritized training of diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11472-11481, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.559, + 0.472, + 0.641 + ], + "angle": 0, + "content": "[8] Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8789-8797, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.644, + 0.472, + 0.699 + ], + "angle": 0, + "content": "[9] Bin Fu, Junjun He, Jianjun Wang, and Yu Qiao. Neural transformation fields for arbitrary-styled font generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22438-22447, 2023. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.702, + 0.47, + 0.756 + ], + "angle": 0, + "content": "[10] Yue Gao, Yuan Guo, Zhouhui Lian, Yingmin Tang, and Jianguo Xiao. Artistic glyph image synthesis via one-stage few-shot learning. ACM Transactions on Graphics (TOG), 38(6): 1-12, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.759, + 0.47, + 0.813 + ], + "angle": 0, + "content": "[11] Haibin He, Xinyuan Chen, Chaoyue Wang, Juhua Liu, Bo Du, Dacheng Tao, and Yu Qiao. Diff-font: Diffusion model for robust one-shot font generation. arXiv preprint arXiv:2212.05895, 2022. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.47, + 0.857 + ], + "angle": 0, + "content": "[12] Xiao He, Mingrui Zhu, Nannan Wang, Xinbo Gao, and Heng Yang. Few-shot font generation by learning style difference and similarity. arXiv preprint arXiv:2301.10008, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[14] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.215 + ], + "angle": 0, + "content": "[15] Yue Jiang, Zhouhui Lian, Yingmin Tang, and Jianguo Xiao. Dcfont: an end-to-end deep chinese font generation system. In SIGGRAPH Asia 2017 Technical Briefs, pages 1-4. 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.892, + 0.246 + ], + "angle": 0, + "content": "[16] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.247, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[17] Yuxin Kong, Canjie Luo, Weihong Ma, Qiyuan Zhu, Shenggao Zhu, Nicholas Yuan, and Lianwen Jin. Look closer to supervise better: One-shot font generation via component-based discriminator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13482-13491, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.33, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[18] Mingi Kwon, Jaeseok Jeong, and Youngjung Uh. Diffusion models already have a semantic latent space. arXiv preprint arXiv:2210.10960, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.372, + 0.892, + 0.426 + ], + "angle": 0, + "content": "[19] Guan-Horng Liu, Arash Vahdat, De-An Huang, Evangelos A Theodorou, Weili Nie, and Anima Anandkumar. I²sb: Image-to-image schrödinger bridge. arXiv preprint arXiv:2302.05872, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.428, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[20] Ming-Yu Liu, Xun Huang, Arun Mallya, Tero Karras, Timo Aila, Jaakko Lehtinen, and Jan Kautz. Few-shot unsupervised image-to-image translation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10551-10560, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.498, + 0.892, + 0.565 + ], + "angle": 0, + "content": "[21] Wei Liu, Fangyue Liu, Fei Ding, Qian He, and Zili Yi. Xmpfont: Self-supervised cross-modality pre-training for few-shot font generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7905-7914, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.567, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[22] Ziwei Luo, Fredrik K Gustafsson, Zheng Zhao, Jens Sjolund, and Thomas B Schon. Image restoration with meanreverting stochastic differential equations. arXiv preprint arXiv:2301.11699, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.623, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[23] Pengyuan Lyu, Xiang Bai, Cong Yao, Zhen Zhu, Tengteng Huang, and Wenyu Liu. Auto-encoder guided gan for Chinese calligraphy synthesis. In 2017 14th IAPR International Conference on Document Analysis and Recognition (ICDAR), pages 1095–1100. IEEE, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.692, + 0.892, + 0.748 + ], + "angle": 0, + "content": "[24] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.749, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[25] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.804, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[26] Wei Pan, Anna Zhu, Xinyu Zhou, Brian Kenji Iwana, and Shilin Li. Few shot font generation via transferring similarity guided global style and quantization local style. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 19506-19516, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.874, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[27] Song Park, Sanghyuk Chun, Junbum Cha, Bado Lee, and Hyunjung Shim. Few-shot font generation with localized" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6900" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.472, + 0.134 + ], + "angle": 0, + "content": "style representations and factorization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2393-2402, 2021. 1, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[28] Song Park, Sanghyuk Chun, Junbum Cha, Bado Lee, and Hyunjung Shim. Multiple heads are better than one: Few-shot font generation with multiple localized experts. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13900-13909, 2021. 1, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.47, + 0.262 + ], + "angle": 0, + "content": "[29] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.472, + 0.332 + ], + "angle": 0, + "content": "[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.47, + 0.403 + ], + "angle": 0, + "content": "[31] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.469, + 0.488 + ], + "angle": 0, + "content": "[32] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.469, + 0.558 + ], + "angle": 0, + "content": "[33] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.561, + 0.469, + 0.602 + ], + "angle": 0, + "content": "[34] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.604, + 0.469, + 0.646 + ], + "angle": 0, + "content": "[35] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.469, + 0.689 + ], + "angle": 0, + "content": "[36] Yang Song and Stefano Ermon. Improved techniques for training score-based generative models. Advances in neural information processing systems, 33:12438-12448, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.69, + 0.469, + 0.746 + ], + "angle": 0, + "content": "[37] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[38] Danyang Sun, Tongzheng Ren, Chongxuan Li, Hang Su, and Jun Zhu. Learning to write stylized Chinese characters by reading a handful of examples. In Proceedings of the 27th International Joint Conference on Artificial Intelligence, pages 920–927, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[39] Licheng Tang, Yiyang Cai, Jiaming Liu, Zhibin Hong, Mingming Gong, Minhu Fan, Junyu Han, Jingtuo Liu, Errui Ding, and Jingdong Wang. Few-shot font generation by learning fine-grained local styles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7895-7904, 2022. 1, 3, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "[40] Yuchen Tian. zi2zi: Master chinese calligraphy with conditional adversarial networks. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[41] Yuchen Tian. Rewrite: Neural style transfer for chinese fonts. 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.151, + 0.892, + 0.192 + ], + "angle": 0, + "content": "[42] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.194, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[43] Chi Wang, Min Zhou, Tiezheng Ge, Yuning Jiang, Hujun Bao, and Weiwei Xu. Cf-font: Content fusion for few-shot font generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1858-1867, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.264, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[44] Tengfei Wang, Ting Zhang, Bo Zhang, Hao Ouyang, Dong Chen, Qifeng Chen, and Fang Wen. Pretraining is all you need for image-to-image translation. arXiv preprint arXiv:2205.12952, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.321, + 0.892, + 0.388 + ], + "angle": 0, + "content": "[45] Yangchen Xie, Xinyuan Chen, Li Sun, and Yue Lu. Dgfont: Deformable generative networks for unsupervised font generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5130-5140, 2021. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.391, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[46] Zhenhua Yang, Dezhi Peng, Yuxin Kong, Yuyi Zhang, Cong Yao, and Lianwen Jin. Fontdiffuser: One-shot font generation via denoising diffusion with multi-scale content aggregation and style contrastive learning. In Proceedings of the AAAI conference on artificial intelligence, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.462, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[47] Mingshuai Yao, Yabo Zhang, Xianhui Lin, Xiaoming Li, and Wangmeng Zuo. Vq-font: Few-shot font generation with structure-aware enhancement and quantization. arXiv preprint arXiv:2308.14018, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.519, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[48] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.589, + 0.892, + 0.643 + ], + "angle": 0, + "content": "[49] Yexun Zhang, Ya Zhang, and Wenbin Cai. Separating style and content for generalized style transfer. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8447-8455, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.646, + 0.892, + 0.713 + ], + "angle": 0, + "content": "[50] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.716, + 0.892, + 0.77 + ], + "angle": 0, + "content": "[51] Mingrui Zhu, Xiao He, Nannan Wang, Xiaoyu Wang, and Xinbo Gao. All-to-key attention for arbitrary style transfer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23109-23119, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.772, + 0.892, + 0.827 + ], + "angle": 0, + "content": "[52] Jianyi Wang Zongsheng Yue and Chen Change Loy. Reshift: Efficient diffusion model for image superresolution by residual shifting. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.827 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "6901" + } + ] +] \ No newline at end of file diff --git a/2024/Generate Like Experts_ Multi-Stage Font Generation by Incorporating Font Transfer Process into Diffusion Models/bfb6c4d4-ce31-4c6d-ba23-750f3d411ecc_origin.pdf b/2024/Generate Like Experts_ Multi-Stage Font Generation by Incorporating Font Transfer Process into Diffusion Models/bfb6c4d4-ce31-4c6d-ba23-750f3d411ecc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c1067647552692c6fe9bbe1eab98234e5b6c8208 --- /dev/null +++ b/2024/Generate Like Experts_ Multi-Stage Font Generation by Incorporating Font Transfer Process into Diffusion Models/bfb6c4d4-ce31-4c6d-ba23-750f3d411ecc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5bc49e8988851ff271221d394d4490665ae173b7f11650ae0b071e688f965646 +size 4829408 diff --git a/2024/Generate Like Experts_ Multi-Stage Font Generation by Incorporating Font Transfer Process into Diffusion Models/full.md b/2024/Generate Like Experts_ Multi-Stage Font Generation by Incorporating Font Transfer Process into Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..897a1f025af4fa52bef78fc7b9a6a14af3c8eeaf --- /dev/null +++ b/2024/Generate Like Experts_ Multi-Stage Font Generation by Incorporating Font Transfer Process into Diffusion Models/full.md @@ -0,0 +1,379 @@ +# Generate Like Experts: Multi-Stage Font Generation by Incorporating Font Transfer Process into Diffusion Models + +Bin Fu $^{1}$ , Fanghua Yu $^{1}$ , Anran Liu $^{3}$ , Zixuan Wang $^{1}$ , Jie Wen $^{4}$ , Junjun He $^{2}$ , and Yu Qiao $^{1,2*}$ + +$^{1}$ ShenZhen Key Lab of Computer Vision and Pattern Recognition, Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences $^{2}$ Shanghai Artificial Intelligence Laboratory $^{3}$ The University of Hong Kong $^{4}$ Harbin Institute of Technology, Shenzhen + +$^{1}\{$ bin.fu, zx.wang8, yu.qiao\}@siat.ac.cn, $^{2}\{$ hejunjun, qiaoyu\}@pjlab.org.cn $^{3}$ liuar616@connect.hku.hk, $^{4}$ jiewen-pr@126.com + +# Abstract + +Few-shot font generation (FFG) produces stylized font images with a limited number of reference samples, which can significantly reduce labor costs in manual font designs. Most existing FFG methods follow the style-content disentanglement paradigm and employ the Generative Adversarial Network (GAN) to generate target fonts by combining the decoupled content and style representations. The complicated structure and detailed style are simultaneously generated in those methods, which may be the sub-optimal solutions for FFG task. Inspired by most manual font design processes of expert designers, in this paper, we model font generation as a multi-stage generative process. Specifically, as the injected noise and the data distribution in diffusion models can be well-separated into different sub-spaces, we are able to incorporate the font transfer process into these models. Based on this observation, we generalize diffusion methods to model font generative process by separating the reverse diffusion process into three stages with different functions: The structure construction stage first generates the structure information for the target character based on the source image, and the font transfer stage subsequently transforms the source font to the target font. Finally, the font refinement stage enhances the appearances and local details of the target font images. Based on the above multistage generative process, we construct our font generation framework, named MSD-Font, with a dual-network approach to generate font images. The superior performance demonstrates the effectiveness of our model. The code is available at: https://github.com/fubinfb/MSD-Font. + +# 1. Introduction + +Designing new fonts is time-consuming and labor-intensive work for humans, especially for some glyph-rich scripts (e.g., Chinese and Korean). The few-shot font generation (FFG) task aims to automatically generate new stylized fonts with a few reference images, and it has received significant interest due to its academic and commercial values. As the requirement of this task is to render the character images into the target font style, current FFG methods [17, 27, 28, 39] usually employ the Generative Adversarial Network (GAN) to perform few-shot font generation under the style-content disentanglement paradigm. These models usually simultaneously generate the complicated structure and detailed style of target font images by combining the content features from source characters and the style representations from reference font images. Although many methods have been proposed to obtain better content and style representations, the generated font images still contain artifacts, inconsistent styles, and broken structures. + +Inspired by the common practice of expert designers, we find the following characteristics that may be crucial for high-quality font generation: As font design requires preserving the global structure of the target characters, designers usually generate a new font by modifying from a similar existing font, which can be viewed as a transfer process from the source font to the target font. Thus, most manual font designs can be regarded as a multi-stage process, including template font generation for global structure, font transfer from the template font to the target font, and refinement of fine-grained details. In light of this, we think that the multi-stage generative process is an effective solution for high-quality font generation, since the problems of broken structures, inconsistent styles, and artifacts can be + +![](images/74738065b6cef9bb4a784b996d392061f1d7308b0a7e6fba13c357e257ad2913.jpg) +Figure 1. The forward process of our proposed model and the traditional diffusion model (DDPM). The first and third rows display the traditional forward process for the source image $I_{c}$ and target font image $I_{g}$ . The images with red boxes mark the path of our proposed diffusion process. In our model, we incorporate the font transfer process in time interval $(t_1, t_2]$ , and align the distribution with the traditional diffusion model at the timesteps $t_1$ and $t_2$ . With this forward process, the generative process of our model can be separated into three stages: the structure construction stage, the font transfer stage, and the font refinement stage, respectively. Note that the font images in this figure are visualized from the diffusion model in image space for better illustration1, while we implement our model in latent space to reduce computational consumption. + +better handled by the above three stages. + +The diffusion model provides a powerful generative framework to achieve this multi-stage process, based on the following two features: (1). Since the injected noise and the data distribution in forward process can be well-separated into different sub-spaces, we are able to incorporate the font transfer process into diffusion models. (2). Recent studies [7, 18] have revealed that the different steps in the generative process have different functions for image synthesis. Therefore, modeling font generation as a transfer process and constructing a multi-stage generative process based on diffusion model is a promising direction for FFG task. + +Motivated by the above observations, in this paper, we incorporate the font transfer process into the diffusion model and formulate font generation as a multi-stage generative process. To alleviate computational consumption, we construct this diffusion process in latent space. Specifically, in latent space, we first separate the injected noise and the data distribution into different sub-spaces, and then incorporate the font transfer process into the data sub-space in a predefined time interval $(t_1,t_2]$ . Combining with the denoising process, the font generative process in our proposed diffusion model can be automatically separated into a multistage process (shown in Fig. 1): the structure construction stage, the font transfer stage, and the font refinement stage, respectively. The structure construction stage first generates the structure information for the target character based on the source image. Then, the font transfer stage gradually transforms the source font to the target font. Finally, the font refinement stage improves the appearances and local details of the generated font images. Under this scheme, the structure construction stage generates the global character structure from source images, while the font transfer stage + +provides a long generative path to produce detailed styles. Moreover, as the prediction network has different behaviors in font transfer stage and font refinement stage, we further propose a dual-network approach to predict the target latent features for different stages. Experiments demonstrate that our model achieves the superior performance on FFG tasks, verifying the effectiveness of our proposed approach. + +In summary, our contributions in this paper are summarized as follows: + +(1). Instead of simultaneously generating font images with complicated structures and detailed styles in most traditional FFG methods, we reformulate the font generation as a multi-stage generative process by incorporating font transfer process into the diffusion models. +(2). We construct the multi-stage font generation framework, named MSD-Font, with a dual-network approach to better deal with different behaviors in different generative stages. +(3). Experimental results show that our method achieves superior performance on the FFG task, demonstrating the effectiveness of our proposed method. + +# 2. Related Works + +This section offers a concise review of recent development in font generation methods and diffusion models. + +# 2.1. Many-shot Font Generation + +As font generation can be formulated as mapping font images from the source domain to the target domain, early methods [4, 40, 41] usually utilize the image-to-image (I2I) methods [8, 14, 20, 50, 51] to learn the mapping function between different fonts. The mapping function will be fine-tuned on hundreds of reference samples (many-shot) to generate unseen font images. Based on the pix2pix [14] framework, Zi2zi [40] and Rewrite [41] utilize one-hot style + +labels to optimize font generation models in a supervised manner. HGAN [4] further extends Zi2zi model via a transfer network and the hierarchical adversarial discriminator. DC-font [15] utilizes a feature reconstruction network to embed style information for font synthesis. AGEN [23] employs the auto-encoder to transfer standard font to calligraphy font. Although these many-shot font generation methods can generate promising font images, collecting large numbers of reference samples to fine-tune the mapping function is still a difficult task for font generation. + +# 2.2. Few-shot Font Generation + +To relieve the weakness of the above models, few-shot font generation (FFG) methods have become popular solutions in recent years, which generate font images with a few samples. As font styles have complex structures and fine-grained details, the common-used statistics-based style transfer techniques are not suitable for font generation. Instead, recent FFG models [2, 3, 10, 17, 21, 26-28, 39, 45-47, 49] mainly follow the style-content disentanglement paradigm, which transfers the target font styles by combining the content representations of source characters with the style codes of reference samples. SA-VAE [38], EMD [49], and AGISNet [10] extract the style and content representations as global features to generate target font images. DM-Font [3], LF-Font [27], and MX-Font [28] decompose characters into components and learn component-wise representations to improve local details. CG-GAN [17] supervises the font generator to decouple content and style at the component level via a component-wise discriminator. Diff-Font [11] proposes a stroke-wise diffusion model to generate font images under the guidance of three character attributes (content, style, and strokes). FsFont [39] utilizes a cross-attention mechanism to aggregate style features into the fine-grained style representation. XMP-Font [21] develops a self-supervised cross-modality pre-training strategy and a cross-modality transformer-based encoder to model the style representations at all levels. DS-Font [12] proposes a cluster-level contrastive style loss to learn better style representations. DG-Font [45] and its extended version DG-Font++ [5] propose a feature deformation skip connection to learn geometric displacement maps between different fonts. CF-Font [43] further extends DG-Font to construct a robust content feature by projecting the content feature into a linear space defined by the content features of basis fonts. Besides, NTF [9] models font generation as a continuous transformation process with a neural transformation field. + +# 2.3. Diffusion Probabilistic Models + +Diffusion model (DM) provides a powerful probabilistic generative framework for high-quality image synthesis, which is first introduced in [33] and subsequently improved + +by Denoising Diffusion Probabilistic Model (DDPM) [13], Denoising Diffusion Implicit Model (DDIM) [34], score-based diffusion [35], and other techniques [25, 36, 37]. Early diffusion models construct the diffusion process in RGB space, which requires huge computation powers to generate high-resolution images. Latent Diffusion Model (LDM) [30] builds the diffusion process in a lower-dimensional latent space, significantly reducing the computation cost. The diffusion model has become the dominant solution for various computer vision tasks, such as image synthesis [29, 30, 32], image editing [1, 6, 24], and image-to-image translation [31, 44]. Large-scale text-to-image (T2I) diffusion models (e.g. DALL-E 2 [29], Imagegen [32], Stable Diffusion [30]) have shown superior performance on high-quality image synthesis, which employ text embedding generated by large language models (LLM) as the condition to guide image generation. Recently, several works [19, 22, 52] modify the denoising diffusion process to model the residual between two given distributions. Different from these works, in this paper, we introduce the font transfer process into the diffusion model at a predefined region, forming a multi-stage diffusion process. Moreover, based on this process, a novel dual-network approach is proposed to predict latent features in different stages. + +# 3. Methodology + +In this section, we present our proposed method in detail, including a short background of the diffusion model (in Sec. 3.1), the proposed latent diffusion model with font transfer process (in Sec. 3.2), the multi-stage font generative process (in Sec. 3.3), and the overall framework (in Sec. 3.4). + +# 3.1. Background and Preliminary + +The diffusion model is a probabilistic generative model aiming to learn the data distribution by gradually denoising a variable sampled from a Gaussian distribution, which includes a forward process and a reverse process. + +Given a data distribution $x_0 \sim q(x_0)$ , the Forward Process (also named as the diffusion process) is fixed to a Markov chain that gradually injects Gaussian noise to the data at time $t$ with a predefined variance schedule $\beta_t$ according to the equations: + +$$ +q \left(x _ {1}, \dots , x _ {T} \mid x _ {0}\right) = \prod_ {t = 1} ^ {T} q \left(x _ {t} \mid x _ {t - 1}\right), \tag {1} +$$ + +$$ +q \left(x _ {t} \mid x _ {t - 1}\right) = \mathcal {N} \left(x _ {t}; \sqrt {1 - \beta_ {t}} x _ {t - 1}, \beta_ {t} I\right), \tag {2} +$$ + +where $x_{1},\dots ,x_{T}$ are latent variables. With the sufficiently large $T$ , the distribution of $x_{T}$ will become an isotropic Gaussian distribution. With the reparameterization trick + +![](images/627114c7efa60a8d931cbc1c36fa66cfa37905ce8f42190ff549551889b0e5fd.jpg) + +Figure 2. (a). The forward diffusion process of our proposed model. (b). The reverse process of our proposed model. Our proposed diffusion process is marked as black circles. The $\tilde{z}_t$ denotes the latent features in our diffusion process, while $z_t^c$ and $z_t^g$ denote the latent features of the source image and the target image in traditional diffusion process, respectively. As we have aligned our diffusion process with the traditional diffusion process of $z^g$ and $z^c$ at $t_1$ and $t_2$ , our model shares the same process with the traditional diffusion process in the region $(0, t_1]$ and $(t_2, T]$ . (c). Overall framework of our font generation model, including the VQ-VAE encoder $\mathcal{E}$ and decoder $\mathcal{D}$ , the dual prediction networks $\tilde{z}_{\theta_1}^{(g,1)}(\tilde{z}_t, t, y_1)$ and $\tilde{z}_{\theta_2}^{(g,2)}(\tilde{z}_t, t, y_2)$ , and the corresponding style and content encoders for each prediction network. Due to the same structure, in this figure, we only plot one prediction network $\tilde{z}_{\theta_i}^{(g,i)}$ ( $i \in \{1, 2\}$ ) with its corresponding style encoder $E_s^i$ and content encoder $E_c^i$ , where $i$ is determined by the generative stage of timestep $t$ . We use $I_g$ and $I_G$ to denote the target image and generated image, respectively. +![](images/afd3bb041299288f31c8b3840a4a19ba32fb1cc13a40b11fb4e16c2aac93c4b4.jpg) +Generative Step in Structure Construction Stage +Generative Step in Font Transfer Stage + +![](images/674cf8f1cd9dd4dd7e6d9a0a6e23812a5cdbf368c3e6f1cc6091fc37e1c4d34f.jpg) +$\tilde{z}_t$ Latent feature in our model + $z_t^c$ Latent feature of source +image in traditional model +$\widetilde{Z}_{\theta_i}^{(g,i)}$ Prediction network $(i\in \{1,2\})$ Latent feature of target image $z_{t}^{g}$ in traditional model +The Equivalent Process for Structure Construction Stage +Generative Step in Font Refinement Stage + +[16], the latent variable $x_{t}$ can be written as + +$$ +\begin{array}{l} x _ {t} = \sqrt {\alpha_ {t}} x _ {t - 1} + \sqrt {1 - \alpha_ {t}} \epsilon_ {t - 1} ^ {*}, (3) \\ = \sqrt {\bar {\alpha} _ {t}} x _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon_ {0}, (4) \\ \end{array} +$$ + +where $\epsilon_{t - 1}^{*},\epsilon_{0}\sim \mathcal{N}(0,I),\alpha_{t} = 1 - \beta_{t}$ and $\bar{\alpha}_t = \prod_{s = 1}^t\alpha_s$ Thus sampling $x_{t}$ at arbitrary timestep has a closed form $q(x_{t}|x_{0}) = \mathcal{N}(x_{t};\sqrt{\bar{\alpha}_{t}} x_{0},(1 - \bar{\alpha}_{t})I)$ + +The Reverse Process generates data from isotropic Gaussian noise $x_{T}$ , and gradually recovers $x_{0}$ via the reverse distribution $p_{\theta}(x_{t - 1}|x_t)$ : + +$$ +p _ {\theta} \left(x _ {0}\right) = \int p _ {\theta} \left(x _ {0: T}\right) d x _ {1: T}, \tag {5} +$$ + +$$ +p _ {\theta} \left(x _ {0: T}\right) = p _ {\theta} \left(x _ {T}\right) \prod_ {t = 1} ^ {T} p _ {\theta} \left(x _ {t - 1} \mid x _ {t}\right), \tag {6} +$$ + +where $p_{\theta}(x_{t - 1}|x_t) = \mathcal{N}(x_{t - 1}|\mu_\theta (x_t,t),\Sigma_\theta (x_t,t))$ is a parameterized Gaussian transition network, and defined as + +$$ +\mu_ {\theta} (x _ {t}, t) = \frac {\sqrt {\alpha_ {t}} (1 - \bar {\alpha} _ {t - 1}) x _ {t} + \sqrt {\bar {\alpha} _ {t - 1}} (1 - \alpha_ {t}) \hat {x} _ {\theta} (x _ {t} , t)}{1 - \bar {\alpha} _ {t}}, +$$ + +and $\Sigma_{\theta}(x_t,t) = (1 - \bar{\alpha}_{t - 1})\beta_tI / (1 - \bar{\alpha}_t)$ . $\hat{x}_{\theta}(x_{t},t)$ is a function approximator for predicting $x_0$ from $x_{t}$ and $t$ . The corresponding objective function can be simplified to: + +$$ +L = \mathbb {E} _ {x _ {0}, \epsilon , t} \left[ \| \hat {x} _ {\theta} (x _ {t}, t) - x _ {0} \| _ {2} ^ {2} \right]. \tag {7} +$$ + +Latent Diffusion Model (LDM) [30] further employs a VQ-VAE [42] encoder $\mathcal{E}$ to project image $x_0$ into latent space $z_0$ , and constructs the diffusion model in this lower-dimensional latent space, significantly reducing computational consumption. Once $z_0$ is sampled via the reverse process, the VQ-VAE decoder $\mathcal{D}$ will project it back to the image space. + +Since LDM significantly reduces computation consumption and maintains high-quality image synthesis, in this paper, we construct our diffusion process in latent space. + +# 3.2. Latent Diffusion Model with Font Transfer + +As shown in Eq. (4), in the forward process, the data distribution $x_0$ and the injected noise can be well-separated in arbitrary timesteps. This characteristic enables us to construct a multi-stage font generative framework in latent diffusion model, by incorporating the font transfer process into the data sub-space in a predefined time interval $t \in (t_1, t_2]$ . Based on this observation, we develop our latent diffusion model in this section, then discuss the multi-stage font generative process in the next section. + +The Forward Process. After projecting the source images $I_{c}$ and target images $I_{g}$ into the latent features $z_0^c$ and $z_0^g$ by the VQ-VAE encoder $\mathcal{E}$ , we implement our model in latent space. As shown in Fig. 2(a), our proposed model (marked as black circles and denoted as $\tilde{z}_t$ ) is constructed + +by utilizing a font transfer process to connect the traditional diffusion processes (DDPM) of the latent features $z_{t}^{c}$ and $z_{t}^{g}$ . Therefore, our latent diffusion process $\tilde{z}_{t}$ can be divided into three regions: (1). In the region $t \in (0, t_1]$ , only latent feature $z_{t}^{g}$ evolved with noise. The forward latent $\tilde{z}_{t}$ in this region can be expressed as + +$$ +\tilde {z} _ {t} = z _ {t} ^ {g} = \sqrt {\bar {\alpha} _ {t}} z _ {0} ^ {g} + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon_ {0} t \in (0, t _ {1} ]. \qquad (8) +$$ + +(2). In the region $t \in (t_1, t_2]$ , together with the continuously injected noise, the latent feature $z_t^g$ of the target font is gradually replaced by the latent feature $z_t^c$ of the source character, and only $z_t^c$ exists at $t_2$ . After careful derivation (provided in Supplementary Materials), the forward latent $\tilde{z}_t$ in this region can be expressed as + +$$ +\tilde {z} _ {t} = \sqrt {\bar {\alpha} _ {t}} z _ {0} ^ {g} + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon_ {0} - \left(z _ {0} ^ {g} - z _ {0} ^ {c}\right) \Psi \mathcal {H} _ {t}, \tag {9} +$$ + +for $t \in (t_1, t_2]$ . The notations $\mathcal{H}_t$ and $\Psi$ are coefficients obtained by aligning our model with traditional diffusion processes (DDPM) of $z^g$ and $z^c$ at timesteps $t_1$ and $t_2$ , respectively. They are defined as + +$$ +\mathcal {H} _ {t} = 1 + \sqrt {\alpha_ {t}} + \sqrt {\alpha_ {t} \alpha_ {t - 1}} + \dots + \sqrt {\alpha_ {t} \cdots \alpha_ {t + 2}}, \tag {10} +$$ + +$$ +\Psi = \frac {\sqrt {\tilde {\alpha} (1 , t _ {2})}}{1 + \sum_ {m = t _ {1} + 2} ^ {t _ {2}} \sqrt {\tilde {\alpha} (m , t _ {2})}}, \tag {11} +$$ + +where $\tilde{\alpha}(t_1, t_2) = \prod_{i=t_1}^{t_2} \alpha_i$ . $\mathcal{H}_t$ can be further reduced to $\mathcal{H}_t = 1 + \sqrt{\alpha_t} \mathcal{H}_{t-1}$ . (3). In the region $t \in (t_2, T]$ , the injected noise finally disrupts the latent feature $z_t^c$ to Gaussian noise, and the forward latent $\tilde{z}_t$ in this region can be expressed as + +$$ +\tilde {z} _ {t} = z _ {t} ^ {c} = \sqrt {\bar {\alpha} _ {t}} z _ {0} ^ {c} + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon_ {0} \quad t \in (t _ {2}, T ]. \tag {12} +$$ + +The Reverse Process. Now we turn to discuss our choices for $p_{\theta}(\tilde{z}_{t - 1}|\tilde{z}_t) = \mathcal{N}(\tilde{z}_{t - 1}|\mu_\theta (\tilde{z}_t,t),\Sigma_\theta (\tilde{z}_t,t))$ . As shown in Eq. (8) and Eq. (12), in the time region $(0,t_1]$ and $(t_2,T]$ , we have aligned the forward latent $\tilde{z}_t$ with the traditional diffusion processes (DDPM [13]), thus they share the same parameterized mean functions: for $t\in (0,t_1]$ + +$$ +\mu_ {\theta} = \frac {\sqrt {\alpha_ {t}} (1 - \bar {\alpha} _ {t - 1}) \tilde {z} _ {t} + \sqrt {\bar {\alpha} _ {t - 1}} (1 - \alpha_ {t}) \tilde {z} _ {\theta} ^ {(g)} (\tilde {z} _ {t} , t)}{1 - \bar {\alpha} _ {t}}, \tag {13} +$$ + +where $\tilde{z}_{\theta}^{(g)}(\tilde{z}_t,t)$ is the prediction network for predicting $z_0^g$ from noisy latent feature $\tilde{z}_t$ and time index $t$ , while for $t\in (t_2,T]$ + +$$ +\mu_ {\theta} = \frac {\sqrt {\alpha_ {t}} (1 - \bar {\alpha} _ {t - 1}) \tilde {z} _ {t} + \sqrt {\bar {\alpha} _ {t - 1}} (1 - \alpha_ {t}) \tilde {z} _ {\theta} ^ {(c)} (\tilde {z} _ {t} , t)}{1 - \bar {\alpha} _ {t}}, \tag {14} +$$ + +where $\tilde{z}_{\theta}^{(c)}(z_t,t)$ is the prediction network for predicting $z_0^c$ from noisy latent feature $\tilde{z}_t$ and time index $t$ . The parameterized mean function for the region $t\in (t_1,t_2]$ can be + +derivated via Bayes' theorem: + +$$ +\begin{array}{l} \mu_ {\theta} = \frac {\sqrt {\alpha_ {t}} (1 - \bar {\alpha} _ {t - 1})}{1 - \bar {\alpha} _ {t}} \tilde {z} _ {t} + \frac {\sqrt {\alpha_ {t}} (1 - \bar {\alpha} _ {t - 1})}{1 - \bar {\alpha} _ {t}} \Psi \tilde {z} _ {\theta} ^ {(g)} (\tilde {z} _ {t}, t) \\ + \frac {\left[ \mathcal {H} _ {t - 1} (1 - \alpha_ {t}) - \sqrt {\alpha_ {t}} (1 - \bar {\alpha} _ {t - 1}) \right]}{1 - \bar {\alpha} _ {t}} \Psi z _ {0} ^ {c} \\ + \frac {(\sqrt {\bar {\alpha} _ {t - 1}} - \mathcal {H} _ {t - 1} \Psi) (1 - \alpha_ {t})}{1 - \bar {\alpha} _ {t}} \tilde {z} _ {\theta} ^ {(g)} (\tilde {z} _ {t}, t). \tag {15} \\ \end{array} +$$ + +As we only incorporate the font transfer process in the data sub-space and keep the noise sub-space unchanged, the variance in the reverse process remains the same as DDPM: + +$$ +\Sigma_ {\theta} \left(\tilde {z} _ {t}, t\right) = \sigma_ {t} ^ {2} I = \frac {(1 - \alpha_ {t}) (1 - \bar {\alpha} _ {t - 1})}{1 - \bar {\alpha} _ {t}} I. \tag {16} +$$ + +Therefore, the latent $\tilde{z}_{t-1}$ in reverse process can be obtained via the formula $p_{\theta}(\tilde{z}_{t-1}|\tilde{z}_t) = \mathcal{N}(\tilde{z}_{t-1};\mu_{\theta}(\tilde{z}_t,t),\sigma_t^2 I)$ , where the $\mu_{\theta}(\tilde{z}_t,t)$ is calculated from Eq. (13)-(15) based on different timesteps. + +The Optimization Objective. As the forward behaviors keep the same as DDPM in the region $(0, t_1]$ and $(t_2, T]$ , the optimization objectives of $\tilde{z}_{\theta}^{(g)}(\tilde{z}_t, t)$ and $\hat{z}_{\theta}^{(c)}(\tilde{z}_t, t)$ have the same expressions as DDPM in these regions. For the optimization objective of $\tilde{z}_{\theta}^{(g)}(\tilde{z}_t, t)$ in the region $(t_1, t_2]$ , it can be derived from the variational bound on negative log-likelihood [13], and only the coefficient is different with the DDPM's loss function. Moreover, as DDPM and LDM adopt the simplified training objectives in their implementations, we also simplify our loss functions and obtain the optimization objective as + +$$ +L = L _ {g} + L _ {c}, \tag {17} +$$ + +$$ +L _ {g} = \mathbb {E} _ {\mathcal {E} (x _ {0}), \epsilon , t \sim U (0, t _ {2} ]} \left[ \| \tilde {z} _ {\theta} ^ {(g)} (\tilde {z} _ {t}, t) - z _ {0} ^ {g} \| _ {2} ^ {2} \right], \tag {18} +$$ + +$$ +L _ {c} = \mathbb {E} _ {\mathcal {E} (x _ {0}), \epsilon , t \sim U (t _ {2}, T ]} \left[ \| \tilde {z} _ {\theta} ^ {(c)} (\tilde {z} _ {t}, t) - z _ {0} ^ {c} \| _ {2} ^ {2} \right], \tag {19} +$$ + +where $U(t_{1},t_{2}]$ denotes the uniform distribution in region $(t_1,t_2]$ . + +The Conditioning Mechanism. Following the LDM [30], we generalize our method to model conditional distributions by implementing with the conditional prediction networks: + +$$ +\tilde {z} _ {\theta} ^ {(i)} \left(\tilde {z} _ {t}, t, y\right), i \in \{g, c \}, \tag {20} +$$ + +where $y$ is the condition for controlling the synthesis process. + +Extension to Conditional Font Generation. We further extend this conditional model to perform the font generation task. Since the few-shot font generation task aims to render the source character $I_{c}$ into the target font style by a few reference samples $I_{s}$ , we utilize the style encoder $E_{s}$ and the content encoder $E_{c}$ to extract the style code from the reference images $I_{s}$ and the content representations from the source images $I_{c}$ . The style and content features are flattened and concatenated to build condition $y$ . + +# 3.3. Multi-Stage Font Generative Process + +Based on our proposed model, the font generation in the reverse process can be automatically separated into three different stages: the structure construction stage, the font transfer stage, and the font refinement stage. + +As shown in Fig. 2(b), starting with the sampled Gaussian noise $\tilde{z}_T$ , in the region $t \in (t_2, T]$ , Structure Construction Stage generates the structure information for the target character via predicting the latent feature $z_0^c$ of the source image $I_c$ . However, since we can access the source image $I_c$ during the inference, the iterative generative steps from $T$ to $t_2$ can be replaced by a single forward step to generate the noised latent variable $\tilde{z}_{t_2}$ at time $t = t_2$ via Eq. (12). This will significantly reduce inference time and provide more accurate structure information. Font Transfer Stage performs font transformation from the timestep $t_2$ to $t_1$ , where the noised latent feature $\tilde{z}_{t_2}$ is gradually transformed to the noised latent feature $\tilde{z}_{t_1}$ of target images via the $\tilde{z}_{\theta}^{(g)}(\tilde{z}_t, t, y)$ . To predict the latent feature $z_0^g$ , the prediction network $\tilde{z}_{\theta}^{(g)}(\tilde{z}_t, t, y)$ in this stage is optimized to not only denoise the latent features but also perform font transfer under the condition $y$ . Finally, in the region $t \in (0, t_1]$ , Font Refinement Stage improves appearances and local details for the generated font images via predicting $z_0^g$ from the noisy latent feature $\tilde{z}_t$ . Based on the above three stages, we can implement the reverse process to generate high-quality font images. + +To better present this multi-stage font generative process, we summarize above steps and provide the pseudo-code in Supplementary Materials. + +# 3.4. Overall Framework for Font Generation + +As discussed in Sec. 3.3, the network $\tilde{z}_{\theta}^{(c)}(\tilde{z}_t,t,y)$ can be replaced by a single forward step. Moreover, instead of implementing a single prediction network $\tilde{z}_{\theta}^{(g)}(\tilde{z}_t,t,y)$ to predict $z_0^g$ we adopt a dual-network approach in this paper, since the network $\tilde{z}_{\theta}^{(g)}(\tilde{z}_t,t,y)$ has different behaviors in the font transfer stage and the font refinement stage. The dual networks $\tilde{z}_{\theta_1}^{(g,1)}(\tilde{z}_t,t,y_1)$ and $\tilde{z}_{\theta_2}^{(g,2)}(\tilde{z}_t,t,y_2)$ share the same structure, but are trained to predict the latent feature $z_0^g$ in different regions: one for font transfer stage and the other for font refinement stage. Therefore, as shown in Fig. 2(c), the overall framework of our font generation model contains the following parts: the VQ-VAE encoder $\mathcal{E}$ , the VQ-VAE decoder $\mathcal{D}$ , the dual prediction networks $\tilde{z}_{\theta_1}^{(g,1)}(\tilde{z}_t,t,y_1)$ and $\tilde{z}_{\theta_2}^{(g,2)}(\tilde{z}_t,t,y_2)$ , and the corresponding style and content encoders for each prediction network $\tilde{z}_{\theta_i}^{(g,i)}(\tilde{z}_t,t,y_i)$ . Finally, from Eq. (17), the optimization objective of our font gener + +ative model can be further reduced to + +$$ +L = L _ {g} ^ {1} + L _ {g} ^ {2}, \tag {21} +$$ + +$$ +L _ {g} ^ {1} = \mathbb {E} _ {\mathcal {E} (x _ {0}), \epsilon , t \sim U (t _ {1}, t _ {2} ]} \left[ \| \tilde {z} _ {\theta_ {1}} ^ {(g, 1)} (\tilde {z} _ {t}, t, y _ {1}) - z _ {0} ^ {g} \| _ {2} ^ {2} \right], \tag {22} +$$ + +$$ +L _ {g} ^ {2} = \mathbb {E} _ {\mathcal {E} (x _ {0}), \epsilon , t \sim U (0, t _ {1} ]} \left[ \| \tilde {z} _ {\theta_ {2}} ^ {(g, 2)} (\tilde {z} _ {t}, t, y _ {2}) - z _ {0} ^ {g} \| _ {2} ^ {2} \right]. \tag {23} +$$ + +# 4. Experiments + +In this section, we conduct extensive experiments to evaluate our proposed model on the FFG task. + +# 4.1. Implement Details + +Our models are implemented on the Stable Diffusion [30] platform. We utilize $T = 1000$ steps to generate $128 \times 128 \times 3$ font images with 8 reference samples. The VQ-VAE encoder $\mathcal{E}$ and decoder $\mathcal{D}$ , the dual-network $\tilde{z}_{\theta_1}^{(g,1)}(\tilde{z}_t,t,y_1)$ and $\tilde{z}_{\theta_2}^{(g,2)}(\tilde{z}_t,t,y_2)$ are initialized by the pre-trained Stable Diffusion models. The style encoder $E_{s}$ and content encoder $E_{c}$ are initialized by Kaiming initialization. We keep VQ-VAE encoder $\mathcal{E}$ and decoder $\mathcal{D}$ frozen and optimize the remaining networks. + +# 4.2. Dataset and Evaluation Metrics + +Datasets: As DM needs more training data to optimize, we collect 900 Chinese fonts to evaluate our method on FFG task. We randomly select 840 fonts as the training set (seen fonts) and the remaining 60 fonts as the testing fonts (unseen fonts). As each collected font covers different Chinese characters, we randomly sample 800 characters from the commonly-used Chinese characters (the first-level Chinese Character Table) to train our method. For the test set, we use 214 characters as the unseen contents and sample 200 characters in our training set as the seen contents. + +Therefore, the training set contains 840 fonts with 800 characters, while the test set includes two parts: one is the Unseen Fonts Unseen Contents (UFUC) test set, containing 60 unseen fonts with 214 unseen characters; the other is the Unseen Fonts Seen Contents (UFSC) test set, containing 60 unseen fonts with 200 seen characters. + +Evaluation Metrics: In this paper, we utilize four commonly used measurements to evaluate our model, including root mean squared error (RMSE), peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and learned perceptual image path similarity (LPIPS) [48]. + +# 4.3. Ablation Study + +With the above dataset and metrics, we conducted extensive experiments in different settings. + +# 4.3.1 Ablation Study on the Overall Framework + +We first construct a Base Model with traditional diffusion process (DDPM) to verify the effectiveness of our proposed + +![](images/036e1b68f8be3f20343b7849c20f43c7da10feb07e760b5a3ff191d7b37559c5.jpg) +Figure 3. Visualization results of the intermediate results $\tilde{z}_t$ by projecting $\tilde{z}_t$ back to image space via VQ-VAE decoder $\mathcal{D}$ . + +![](images/1bbb281e3f0dfe20f2db227bb746f82f4c7d078d53527a2eb62b3050abfd947d.jpg) +Figure 4. Visualization results of the predicted latent feature $z_0^g$ in the intermediate timesteps by projecting $z_0^g$ back to image space via VQ-VAE decoder $\mathcal{D}$ . + +Table 1. Ablation study for the overall framework on UFUC dataset. The bold number indicates the best. + +
MethodsRMSE↓PSNR↑SSIM↑LPIPS↓
Base Model0.267611.720.69610.1561
Single-Net0.260112.020.70960.1531
Dual-Net0.247512.450.71890.1527
+ +latent diffusion model. Moreover, as discussed in Sec. 3.4, there are two different selections to obtain $z_0^g$ in the reverse process: One utilizes a single network to predict $z_0^g$ at all timesteps. The other adopts the dual networks to predict the latent feature $z_0^g$ in font transfer stage and font refinement stage, respectively. + +As shown in Tab. 1, compared with the Base Model, our Single-Net model improves the Base Model with 0.0075, 0.30, 0.0135, and 0.0030 in terms of RMSE, PSNR, SSIM, and LPIPS, verifying the effectiveness of our proposed latent diffusion model. Moreover, the Dual-Net approach further improves the Single-Net model with 0.0126, 0.43, and 0.0093 in terms of RMSE, PSNR, and SSIM, illustrating that the use of separate models is beneficial for predicting $z_0^g$ in different stages. + +# 4.3.2 Ablation Study on the Multi-Stage Font Generative Process + +The selections of the region for different stages in the multistage generative process are hyper-parameters in our model, + +which significantly impact image quality. Therefore, we study the influence on model performance by selecting different $t_1$ and $t_2$ . Due to the page limitation, the experimental results and the detailed discussions are provided in supplementary materials, and we briefly summarize the conclusion: The hyper-parameters $t_1$ and $t_2$ are trade-off selections, and the configuration $t_1 = 200$ and $t_2 = 800$ achieves the best record. Thus, we select this configuration for the following experiments. + +# 4.3.3 Visualization Results + +In this section, we provide the visualization results of our generative process to verify the effectiveness of our model. + +As our diffusion process is implemented in latent space, we visualize the intermediate generative results $\tilde{z}_t$ by using the VQ-VAE decoder $\mathcal{D}$ to project them back into the image space. As shown in Fig. 3, the images in the first and third rows are the intermediate results for source image $I_{c}$ and target image $I_{g}$ generated by the traditional diffusion process (DDPM), while the images in the second row are generated by our multi-stage font generative process. From this figure, we can clearly observe a font transfer process from the source font to the target font, and the font refinement stage slightly improves style consistency by modifying local details. + +To further evaluate the font refinement stage, we also visualize the predicted $z_0^g$ from different timesteps in the font refinement stage. As shown in Fig. 4, from $t = 200$ to $t = 0$ , the font refinement stage can gradually enhance local details by modifying anomalous pixels, recovering unwanted structures, and improving visual appearances. + +# 4.4. Comparison with the State-of-the-art Methods + +In this section, we compare our model with several state-of-the-art works, including LF-Font [27], DG-Font [45], MX-Font [28], FsFont [39], NTF [9], and Diff-Font [11]. LF-Font and MX-Font are component-related methods, which decompose characters into components and learn component-wise representations to improve local details. FsFont focuses on fine-grained local styles, which utilizes cross attention to aggregate the reference styles into target styles. DG-Font and NTF are shape transformation mod + +
Content辛禁病安持理根英役破任始米保用李倍荷供唱血止在利打
LF-Font辛禁病安持理根英役破任始米保用李倍荷供唱血止在利打
DG-Font辛禁病安持理根英役破任始米保用李倍荷供唱血止在利打
MX-Font辛禁病安持理根英役破任始米保用李倍荷供唱血止在利打
FsFont辛禁病安持理根英役破任始米保用李倍荷供唱血止在利打
NTF辛禁病安持理根英役破任始米保用李倍荷供唱血止在利打
Ours辛禁病安持理根英役破任始米保用李倍荷供唱血止在利打
Ground Truth辛禁病安持理根英役破任始米保用李倍荷供唱血止在利打
+ +Figure 5. Qualitative comparisons of our proposed model with other state-of-the-art methods on UFUC dataset. +Table 2. Quantitative comparison of our proposed model with other state-of-the-art methods on few-shot font generation task. The bold number indicates the best. + +
MethodsRMSE↓PSNR↑SSIM↑LPIPS↓
Unseen Fonts and Seen Contents
LF-Font [27]0.302510.660.64690.2787
DG-Font [45]0.283311.240.66510.1479
MX-Font [28]0.300710.750.64380.1651
FsFont [39]0.244312.560.71050.1468
NTF [9]0.265011.830.68960.1517
Diff-Font2[11]0.281110.950.68460.1557
MSD-Font0.243912.610.71100.1348
Unseen Fonts and Unseen Contents
LF-Font [27]0.291011.090.67620.2564
DG-Font [45]0.275011.540.68930.1534
MX-Font [28]0.284511.290.67890.1635
FsFont [39]0.250912.290.71850.1642
NTF [9]0.257412.110.71080.1553
MSD-Font0.247512.450.71890.1527
+ +els. DG-Font models font generation as a shape deformable problem via the deformation skip connection, while NTF models font generation as a continuous transformation process. We train above methods on our dataset and use the same reference images to generate fonts in inference. + +# 4.4.1 Quantitative Comparison + +The quantitative results are shown in Tab. 2. From the table, we find that modeling font style in fine-grained manners by multiple localized encoders (MX-Font) or cross attention mechanism (FsFont) can improve the quality of generated font images. The shape transformation approaches (DG-Font and NTF) are also promising solutions for FFG task. Our MSD-Font achieves superior performance on both UFUC and UFSC datasets, which demonstrates the effectiveness of our proposed method. Specifically, compared with the NTF on the UFUC dataset, our method improves the generation performance with 0.0099, 0.34, 0.0081, and 0.0026 in terms of RMSE, PSNR, SSIM, and LPIPS. + +Diff-Font is one-shot font generation model, we randomly select one image from 8 reference samples to calculate style code. + +# 4.4.2 Qualitative Comparison + +To further evaluate our methods, as shown in Fig. 5, we visualize the generated font images in UFUC setting. LF-Font sometimes generates font images with some broken structures. The DG-Font keeps the global structure for the target character, but the generated images may contain some unwanted local structures and artifacts. FsFont utilizes several carefully selected reference characters to generate font images, thus sometimes it will fail to produce accurate structures for target images with randomly selected reference characters. The MX-Font and NTF also contain some missing strokes and distorted local details. Our MSD-Font is able to generate high-quality font images, which achieves higher visual quality than other state-of-the-art methods in terms of style consistency and structure completeness. + +# 5. Conclusion + +In this paper, we propose a multi-stage font generative process for few-shot font generation task, by incorporating font transfer process into the diffusion model. The proposed font generation model includes three stages: a structure construction stage to generate global structures from the source images, a font transfer stage to transform the source font into the target font, and a font refinement stage to improve the appearances and local details for the generated font images. Based on this multi-stage generative process, we construct our font generation framework, named MSD-Font, with a dual-network approach to generate high-quality font images using a few reference samples. The promising experimental results verify the effectiveness of our proposed model. + +Limitations: As our MSD-Font based on diffusion models, there is an increase in the inference time and model size compared with previous GAN-based FFG methods. Many recent works have focused on improving the efficiency and model compactness of diffusion models, and most of these methods can be used in our model. + +Acknowledgements: This work is supported by the National Key R&D Program of China (NO. 2022ZD0160102). + +# References + +[1] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18208-18218, 2022. 3 +[2] Samaneh Azadi, Matthew Fisher, Vladimir G Kim, Zhaowen Wang, Eli Shechtman, and Trevor Darrell. Multi-content gan for few-shot font style transfer. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7564-7573, 2018. 3 +[3] Junbum Cha, Sanghyuk Chun, Gayoung Lee, Bado Lee, Seonghyeon Kim, and Hwalsuk Lee. Few-shot compositional font generation with dual memory. In European Conference on Computer Vision, pages 735-751. Springer, 2020. 3 +[4] Jie Chang, Yujun Gu, Ya Zhang, Yan-Feng Wang, and CM Innovation. Chinese handwriting imitation with hierarchical generative adversarial network. In BMVC, page 290, 2018. 2, 3 +[5] Xinyuan Chen, Yangchen Xie, Li Sun, and Yue Lu. Dgfont++: Robust deformable generative networks for unsupervised font generation. arXiv preprint arXiv:2212.14742, 2022.3 +[6] Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. Ilvr: Conditioning method for denoising diffusion probabilistic models. arXiv preprint arXiv:2108.02938, 2021. 3 +[7] Jooyoung Choi, Jungbeom Lee, Chaehun Shin, Sungwon Kim, Hyunwoo Kim, and Sungroh Yoon. Perception prioritized training of diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11472-11481, 2022. 2 +[8] Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8789-8797, 2018. 2 +[9] Bin Fu, Junjun He, Jianjun Wang, and Yu Qiao. Neural transformation fields for arbitrary-styled font generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22438-22447, 2023. 3, 7, 8 +[10] Yue Gao, Yuan Guo, Zhouhui Lian, Yingmin Tang, and Jianguo Xiao. Artistic glyph image synthesis via one-stage few-shot learning. ACM Transactions on Graphics (TOG), 38(6): 1-12, 2019. 3 +[11] Haibin He, Xinyuan Chen, Chaoyue Wang, Juhua Liu, Bo Du, Dacheng Tao, and Yu Qiao. Diff-font: Diffusion model for robust one-shot font generation. arXiv preprint arXiv:2212.05895, 2022. 3, 7, 8 +[12] Xiao He, Mingrui Zhu, Nannan Wang, Xinbo Gao, and Heng Yang. Few-shot font generation by learning style difference and similarity. arXiv preprint arXiv:2301.10008, 2023. 3 +[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3, 5 + +[14] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 2 +[15] Yue Jiang, Zhouhui Lian, Yingmin Tang, and Jianguo Xiao. Dcfont: an end-to-end deep chinese font generation system. In SIGGRAPH Asia 2017 Technical Briefs, pages 1-4. 2017. 3 +[16] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 4 +[17] Yuxin Kong, Canjie Luo, Weihong Ma, Qiyuan Zhu, Shenggao Zhu, Nicholas Yuan, and Lianwen Jin. Look closer to supervise better: One-shot font generation via component-based discriminator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13482-13491, 2022. 1, 3 +[18] Mingi Kwon, Jaeseok Jeong, and Youngjung Uh. Diffusion models already have a semantic latent space. arXiv preprint arXiv:2210.10960, 2022. 2 +[19] Guan-Horng Liu, Arash Vahdat, De-An Huang, Evangelos A Theodorou, Weili Nie, and Anima Anandkumar. I²sb: Image-to-image schrödinger bridge. arXiv preprint arXiv:2302.05872, 2023. 3 +[20] Ming-Yu Liu, Xun Huang, Arun Mallya, Tero Karras, Timo Aila, Jaakko Lehtinen, and Jan Kautz. Few-shot unsupervised image-to-image translation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10551-10560, 2019. 2 +[21] Wei Liu, Fangyue Liu, Fei Ding, Qian He, and Zili Yi. Xmpfont: Self-supervised cross-modality pre-training for few-shot font generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7905-7914, 2022. 3 +[22] Ziwei Luo, Fredrik K Gustafsson, Zheng Zhao, Jens Sjolund, and Thomas B Schon. Image restoration with meanreverting stochastic differential equations. arXiv preprint arXiv:2301.11699, 2023. 3 +[23] Pengyuan Lyu, Xiang Bai, Cong Yao, Zhen Zhu, Tengteng Huang, and Wenyu Liu. Auto-encoder guided gan for Chinese calligraphy synthesis. In 2017 14th IAPR International Conference on Document Analysis and Recognition (ICDAR), pages 1095–1100. IEEE, 2017. 3 +[24] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 3 +[25] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 3 +[26] Wei Pan, Anna Zhu, Xinyu Zhou, Brian Kenji Iwana, and Shilin Li. Few shot font generation via transferring similarity guided global style and quantization local style. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 19506-19516, 2023. 3 +[27] Song Park, Sanghyuk Chun, Junbum Cha, Bado Lee, and Hyunjung Shim. Few-shot font generation with localized + +style representations and factorization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2393-2402, 2021. 1, 3, 7, 8 +[28] Song Park, Sanghyuk Chun, Junbum Cha, Bado Lee, and Hyunjung Shim. Multiple heads are better than one: Few-shot font generation with multiple localized experts. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13900-13909, 2021. 1, 3, 7, 8 +[29] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 3 +[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3, 4, 5, 6 +[31] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 3 +[32] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 3 +[33] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 3 +[34] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3 +[35] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 3 +[36] Yang Song and Stefano Ermon. Improved techniques for training score-based generative models. Advances in neural information processing systems, 33:12438-12448, 2020. 3 +[37] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 3 +[38] Danyang Sun, Tongzheng Ren, Chongxuan Li, Hang Su, and Jun Zhu. Learning to write stylized Chinese characters by reading a handful of examples. In Proceedings of the 27th International Joint Conference on Artificial Intelligence, pages 920–927, 2018. 3 +[39] Licheng Tang, Yiyang Cai, Jiaming Liu, Zhibin Hong, Mingming Gong, Minhu Fan, Junyu Han, Jingtuo Liu, Errui Ding, and Jingdong Wang. Few-shot font generation by learning fine-grained local styles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7895-7904, 2022. 1, 3, 7, 8 + +[40] Yuchen Tian. zi2zi: Master chinese calligraphy with conditional adversarial networks. 2 +[41] Yuchen Tian. Rewrite: Neural style transfer for chinese fonts. 2016. 2 +[42] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 4 +[43] Chi Wang, Min Zhou, Tiezheng Ge, Yuning Jiang, Hujun Bao, and Weiwei Xu. Cf-font: Content fusion for few-shot font generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1858-1867, 2023. 3 +[44] Tengfei Wang, Ting Zhang, Bo Zhang, Hao Ouyang, Dong Chen, Qifeng Chen, and Fang Wen. Pretraining is all you need for image-to-image translation. arXiv preprint arXiv:2205.12952, 2022.3 +[45] Yangchen Xie, Xinyuan Chen, Li Sun, and Yue Lu. Dgfont: Deformable generative networks for unsupervised font generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5130-5140, 2021. 3, 7, 8 +[46] Zhenhua Yang, Dezhi Peng, Yuxin Kong, Yuyi Zhang, Cong Yao, and Lianwen Jin. Fontdiffuser: One-shot font generation via denoising diffusion with multi-scale content aggregation and style contrastive learning. In Proceedings of the AAAI conference on artificial intelligence, 2024. +[47] Mingshuai Yao, Yabo Zhang, Xianhui Lin, Xiaoming Li, and Wangmeng Zuo. Vq-font: Few-shot font generation with structure-aware enhancement and quantization. arXiv preprint arXiv:2308.14018, 2023. 3 +[48] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6 +[49] Yexun Zhang, Ya Zhang, and Wenbin Cai. Separating style and content for generalized style transfer. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8447-8455, 2018. 3 +[50] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 2 +[51] Mingrui Zhu, Xiao He, Nannan Wang, Xiaoyu Wang, and Xinbo Gao. All-to-key attention for arbitrary style transfer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23109-23119, 2023. 2 +[52] Jianyi Wang Zongsheng Yue and Chen Change Loy. Reshift: Efficient diffusion model for image superresolution by residual shifting. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 3 \ No newline at end of file diff --git a/2024/Generate Like Experts_ Multi-Stage Font Generation by Incorporating Font Transfer Process into Diffusion Models/images.zip b/2024/Generate Like Experts_ Multi-Stage Font Generation by Incorporating Font Transfer Process into Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9859c922924343b4703a33ca7b81b9e6292b86c7 --- /dev/null +++ b/2024/Generate Like Experts_ Multi-Stage Font Generation by Incorporating Font Transfer Process into Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6eb12423d98d281755e9226e20ceacee3b553c444214c9f06865c1f626a9a2ee +size 673060 diff --git a/2024/Generate Like Experts_ Multi-Stage Font Generation by Incorporating Font Transfer Process into Diffusion Models/layout.json b/2024/Generate Like Experts_ Multi-Stage Font Generation by Incorporating Font Transfer Process into Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c0ea1c022d107cc0921b1b3131ac6df8ce9165b2 --- /dev/null +++ b/2024/Generate Like Experts_ Multi-Stage Font Generation by Incorporating Font Transfer Process into Diffusion Models/layout.json @@ -0,0 +1,11363 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 78, + 102, + 516, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 102, + 516, + 138 + ], + "spans": [ + { + "bbox": [ + 78, + 102, + 516, + 138 + ], + "type": "text", + "content": "Generate Like Experts: Multi-Stage Font Generation by Incorporating Font Transfer Process into Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 83, + 160, + 515, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 160, + 515, + 176 + ], + "spans": [ + { + "bbox": [ + 83, + 160, + 515, + 176 + ], + "type": "text", + "content": "Bin Fu" + }, + { + "bbox": [ + 83, + 160, + 515, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 83, + 160, + 515, + 176 + ], + "type": "text", + "content": ", Fanghua Yu" + }, + { + "bbox": [ + 83, + 160, + 515, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 83, + 160, + 515, + 176 + ], + "type": "text", + "content": ", Anran Liu" + }, + { + "bbox": [ + 83, + 160, + 515, + 176 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 83, + 160, + 515, + 176 + ], + "type": "text", + "content": ", Zixuan Wang" + }, + { + "bbox": [ + 83, + 160, + 515, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 83, + 160, + 515, + 176 + ], + "type": "text", + "content": ", Jie Wen" + }, + { + "bbox": [ + 83, + 160, + 515, + 176 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 83, + 160, + 515, + 176 + ], + "type": "text", + "content": ", Junjun He" + }, + { + "bbox": [ + 83, + 160, + 515, + 176 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 83, + 160, + 515, + 176 + ], + "type": "text", + "content": ", and Yu Qiao" + }, + { + "bbox": [ + 83, + 160, + 515, + 176 + ], + "type": "inline_equation", + "content": "^{1,2*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 107, + 189, + 485, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 189, + 485, + 244 + ], + "spans": [ + { + "bbox": [ + 107, + 189, + 485, + 244 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 107, + 189, + 485, + 244 + ], + "type": "text", + "content": "ShenZhen Key Lab of Computer Vision and Pattern Recognition, Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences " + }, + { + "bbox": [ + 107, + 189, + 485, + 244 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 107, + 189, + 485, + 244 + ], + "type": "text", + "content": "Shanghai Artificial Intelligence Laboratory " + }, + { + "bbox": [ + 107, + 189, + 485, + 244 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 107, + 189, + 485, + 244 + ], + "type": "text", + "content": " The University of Hong Kong " + }, + { + "bbox": [ + 107, + 189, + 485, + 244 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 107, + 189, + 485, + 244 + ], + "type": "text", + "content": "Harbin Institute of Technology, Shenzhen" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 101, + 246, + 492, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 246, + 492, + 273 + ], + "spans": [ + { + "bbox": [ + 101, + 246, + 492, + 273 + ], + "type": "inline_equation", + "content": "^{1}\\{" + }, + { + "bbox": [ + 101, + 246, + 492, + 273 + ], + "type": "text", + "content": " bin.fu, zx.wang8, yu.qiao\\}@siat.ac.cn, " + }, + { + "bbox": [ + 101, + 246, + 492, + 273 + ], + "type": "inline_equation", + "content": "^{2}\\{" + }, + { + "bbox": [ + 101, + 246, + 492, + 273 + ], + "type": "text", + "content": " hejunjun, qiaoyu\\}@pjlab.org.cn " + }, + { + "bbox": [ + 101, + 246, + 492, + 273 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 101, + 246, + 492, + 273 + ], + "type": "text", + "content": " liuar616@connect.hku.hk, " + }, + { + "bbox": [ + 101, + 246, + 492, + 273 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 101, + 246, + 492, + 273 + ], + "type": "text", + "content": " jiewen-pr@126.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 300, + 192, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 300, + 192, + 313 + ], + "spans": [ + { + "bbox": [ + 143, + 300, + 192, + 313 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 328, + 289, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 328, + 289, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 328, + 289, + 676 + ], + "type": "text", + "content": "Few-shot font generation (FFG) produces stylized font images with a limited number of reference samples, which can significantly reduce labor costs in manual font designs. Most existing FFG methods follow the style-content disentanglement paradigm and employ the Generative Adversarial Network (GAN) to generate target fonts by combining the decoupled content and style representations. The complicated structure and detailed style are simultaneously generated in those methods, which may be the sub-optimal solutions for FFG task. Inspired by most manual font design processes of expert designers, in this paper, we model font generation as a multi-stage generative process. Specifically, as the injected noise and the data distribution in diffusion models can be well-separated into different sub-spaces, we are able to incorporate the font transfer process into these models. Based on this observation, we generalize diffusion methods to model font generative process by separating the reverse diffusion process into three stages with different functions: The structure construction stage first generates the structure information for the target character based on the source image, and the font transfer stage subsequently transforms the source font to the target font. Finally, the font refinement stage enhances the appearances and local details of the target font images. Based on the above multistage generative process, we construct our font generation framework, named MSD-Font, with a dual-network approach to generate font images. The superior performance demonstrates the effectiveness of our model. The code is available at: https://github.com/fubinfb/MSD-Font." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 307, + 300, + 386, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 300, + 386, + 313 + ], + "spans": [ + { + "bbox": [ + 307, + 300, + 386, + 313 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 325, + 547, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 325, + 547, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 325, + 547, + 540 + ], + "type": "text", + "content": "Designing new fonts is time-consuming and labor-intensive work for humans, especially for some glyph-rich scripts (e.g., Chinese and Korean). The few-shot font generation (FFG) task aims to automatically generate new stylized fonts with a few reference images, and it has received significant interest due to its academic and commercial values. As the requirement of this task is to render the character images into the target font style, current FFG methods [17, 27, 28, 39] usually employ the Generative Adversarial Network (GAN) to perform few-shot font generation under the style-content disentanglement paradigm. These models usually simultaneously generate the complicated structure and detailed style of target font images by combining the content features from source characters and the style representations from reference font images. Although many methods have been proposed to obtain better content and style representations, the generated font images still contain artifacts, inconsistent styles, and broken structures." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 546, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 548, + 714 + ], + "type": "text", + "content": "Inspired by the common practice of expert designers, we find the following characteristics that may be crucial for high-quality font generation: As font design requires preserving the global structure of the target characters, designers usually generate a new font by modifying from a similar existing font, which can be viewed as a transfer process from the source font to the target font. Thus, most manual font designs can be regarded as a multi-stage process, including template font generation for global structure, font transfer from the template font to the target font, and refinement of fine-grained details. In light of this, we think that the multi-stage generative process is an effective solution for high-quality font generation, since the problems of broken structures, inconsistent styles, and artifacts can be" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 166, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 166, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 166, + 713 + ], + "type": "text", + "content": "*Corresponding author: Yu Qiao" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "6892" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 69, + 509, + 198 + ], + "blocks": [ + { + "bbox": [ + 85, + 69, + 509, + 198 + ], + "lines": [ + { + "bbox": [ + 85, + 69, + 509, + 198 + ], + "spans": [ + { + "bbox": [ + 85, + 69, + 509, + 198 + ], + "type": "image", + "image_path": "74738065b6cef9bb4a784b996d392061f1d7308b0a7e6fba13c357e257ad2913.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 201, + 547, + 279 + ], + "lines": [ + { + "bbox": [ + 46, + 201, + 547, + 279 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 547, + 279 + ], + "type": "text", + "content": "Figure 1. The forward process of our proposed model and the traditional diffusion model (DDPM). The first and third rows display the traditional forward process for the source image " + }, + { + "bbox": [ + 46, + 201, + 547, + 279 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 46, + 201, + 547, + 279 + ], + "type": "text", + "content": " and target font image " + }, + { + "bbox": [ + 46, + 201, + 547, + 279 + ], + "type": "inline_equation", + "content": "I_{g}" + }, + { + "bbox": [ + 46, + 201, + 547, + 279 + ], + "type": "text", + "content": ". The images with red boxes mark the path of our proposed diffusion process. In our model, we incorporate the font transfer process in time interval " + }, + { + "bbox": [ + 46, + 201, + 547, + 279 + ], + "type": "inline_equation", + "content": "(t_1, t_2]" + }, + { + "bbox": [ + 46, + 201, + 547, + 279 + ], + "type": "text", + "content": ", and align the distribution with the traditional diffusion model at the timesteps " + }, + { + "bbox": [ + 46, + 201, + 547, + 279 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 46, + 201, + 547, + 279 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 201, + 547, + 279 + ], + "type": "inline_equation", + "content": "t_2" + }, + { + "bbox": [ + 46, + 201, + 547, + 279 + ], + "type": "text", + "content": ". With this forward process, the generative process of our model can be separated into three stages: the structure construction stage, the font transfer stage, and the font refinement stage, respectively. Note that the font images in this figure are visualized from the diffusion model in image space for better illustration1, while we implement our model in latent space to reduce computational consumption." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 285, + 212, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 285, + 212, + 296 + ], + "spans": [ + { + "bbox": [ + 47, + 285, + 212, + 296 + ], + "type": "text", + "content": "better handled by the above three stages." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 297, + 288, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 297, + 288, + 429 + ], + "spans": [ + { + "bbox": [ + 46, + 297, + 288, + 429 + ], + "type": "text", + "content": "The diffusion model provides a powerful generative framework to achieve this multi-stage process, based on the following two features: (1). Since the injected noise and the data distribution in forward process can be well-separated into different sub-spaces, we are able to incorporate the font transfer process into diffusion models. (2). Recent studies [7, 18] have revealed that the different steps in the generative process have different functions for image synthesis. Therefore, modeling font generation as a transfer process and constructing a multi-stage generative process based on diffusion model is a promising direction for FFG task." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 430, + 288, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 430, + 288, + 682 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 288, + 682 + ], + "type": "text", + "content": "Motivated by the above observations, in this paper, we incorporate the font transfer process into the diffusion model and formulate font generation as a multi-stage generative process. To alleviate computational consumption, we construct this diffusion process in latent space. Specifically, in latent space, we first separate the injected noise and the data distribution into different sub-spaces, and then incorporate the font transfer process into the data sub-space in a predefined time interval " + }, + { + "bbox": [ + 46, + 430, + 288, + 682 + ], + "type": "inline_equation", + "content": "(t_1,t_2]" + }, + { + "bbox": [ + 46, + 430, + 288, + 682 + ], + "type": "text", + "content": ". Combining with the denoising process, the font generative process in our proposed diffusion model can be automatically separated into a multistage process (shown in Fig. 1): the structure construction stage, the font transfer stage, and the font refinement stage, respectively. The structure construction stage first generates the structure information for the target character based on the source image. Then, the font transfer stage gradually transforms the source font to the target font. Finally, the font refinement stage improves the appearances and local details of the generated font images. Under this scheme, the structure construction stage generates the global character structure from source images, while the font transfer stage" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 309, + 285, + 547, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 285, + 547, + 369 + ], + "spans": [ + { + "bbox": [ + 309, + 285, + 547, + 369 + ], + "type": "text", + "content": "provides a long generative path to produce detailed styles. Moreover, as the prediction network has different behaviors in font transfer stage and font refinement stage, we further propose a dual-network approach to predict the target latent features for different stages. Experiments demonstrate that our model achieves the superior performance on FFG tasks, verifying the effectiveness of our proposed approach." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 369, + 545, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 369, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 306, + 369, + 545, + 392 + ], + "type": "text", + "content": "In summary, our contributions in this paper are summarized as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 393, + 545, + 537 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 306, + 393, + 545, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 393, + 545, + 452 + ], + "spans": [ + { + "bbox": [ + 306, + 393, + 545, + 452 + ], + "type": "text", + "content": "(1). Instead of simultaneously generating font images with complicated structures and detailed styles in most traditional FFG methods, we reformulate the font generation as a multi-stage generative process by incorporating font transfer process into the diffusion models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 453, + 545, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 453, + 545, + 500 + ], + "spans": [ + { + "bbox": [ + 306, + 453, + 545, + 500 + ], + "type": "text", + "content": "(2). We construct the multi-stage font generation framework, named MSD-Font, with a dual-network approach to better deal with different behaviors in different generative stages." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 502, + 545, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 502, + 545, + 537 + ], + "spans": [ + { + "bbox": [ + 306, + 502, + 545, + 537 + ], + "type": "text", + "content": "(3). Experimental results show that our method achieves superior performance on the FFG task, demonstrating the effectiveness of our proposed method." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 550, + 398, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 550, + 398, + 562 + ], + "spans": [ + { + "bbox": [ + 306, + 550, + 398, + 562 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 571, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 571, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 306, + 571, + 545, + 594 + ], + "type": "text", + "content": "This section offers a concise review of recent development in font generation methods and diffusion models." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 598, + 459, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 598, + 459, + 611 + ], + "spans": [ + { + "bbox": [ + 306, + 598, + 459, + 611 + ], + "type": "text", + "content": "2.1. Many-shot Font Generation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 618, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 618, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 618, + 545, + 713 + ], + "type": "text", + "content": "As font generation can be formulated as mapping font images from the source domain to the target domain, early methods [4, 40, 41] usually utilize the image-to-image (I2I) methods [8, 14, 20, 50, 51] to learn the mapping function between different fonts. The mapping function will be fine-tuned on hundreds of reference samples (many-shot) to generate unseen font images. Based on the pix2pix [14] framework, Zi2zi [40] and Rewrite [41] utilize one-hot style" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 288, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 288, + 714 + ], + "type": "text", + "content": "1Specifically, we replace " + }, + { + "bbox": [ + 47, + 693, + 288, + 714 + ], + "type": "inline_equation", + "content": "z_0^c" + }, + { + "bbox": [ + 47, + 693, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 693, + 288, + 714 + ], + "type": "inline_equation", + "content": "z_0^g" + }, + { + "bbox": [ + 47, + 693, + 288, + 714 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 47, + 693, + 288, + 714 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 47, + 693, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 693, + 288, + 714 + ], + "type": "inline_equation", + "content": "I_{g}" + }, + { + "bbox": [ + 47, + 693, + 288, + 714 + ], + "type": "text", + "content": " in Eq. (8)-(12) to visualize our proposed diffusion process in this figure." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6893" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": "labels to optimize font generation models in a supervised manner. HGAN [4] further extends Zi2zi model via a transfer network and the hierarchical adversarial discriminator. DC-font [15] utilizes a feature reconstruction network to embed style information for font synthesis. AGEN [23] employs the auto-encoder to transfer standard font to calligraphy font. Although these many-shot font generation methods can generate promising font images, collecting large numbers of reference samples to fine-tune the mapping function is still a difficult task for font generation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 201, + 192, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 201, + 192, + 213 + ], + "spans": [ + { + "bbox": [ + 47, + 201, + 192, + 213 + ], + "type": "text", + "content": "2.2. Few-shot Font Generation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 219, + 289, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 219, + 289, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 219, + 289, + 651 + ], + "type": "text", + "content": "To relieve the weakness of the above models, few-shot font generation (FFG) methods have become popular solutions in recent years, which generate font images with a few samples. As font styles have complex structures and fine-grained details, the common-used statistics-based style transfer techniques are not suitable for font generation. Instead, recent FFG models [2, 3, 10, 17, 21, 26-28, 39, 45-47, 49] mainly follow the style-content disentanglement paradigm, which transfers the target font styles by combining the content representations of source characters with the style codes of reference samples. SA-VAE [38], EMD [49], and AGISNet [10] extract the style and content representations as global features to generate target font images. DM-Font [3], LF-Font [27], and MX-Font [28] decompose characters into components and learn component-wise representations to improve local details. CG-GAN [17] supervises the font generator to decouple content and style at the component level via a component-wise discriminator. Diff-Font [11] proposes a stroke-wise diffusion model to generate font images under the guidance of three character attributes (content, style, and strokes). FsFont [39] utilizes a cross-attention mechanism to aggregate style features into the fine-grained style representation. XMP-Font [21] develops a self-supervised cross-modality pre-training strategy and a cross-modality transformer-based encoder to model the style representations at all levels. DS-Font [12] proposes a cluster-level contrastive style loss to learn better style representations. DG-Font [45] and its extended version DG-Font++ [5] propose a feature deformation skip connection to learn geometric displacement maps between different fonts. CF-Font [43] further extends DG-Font to construct a robust content feature by projecting the content feature into a linear space defined by the content features of basis fonts. Besides, NTF [9] models font generation as a continuous transformation process with a neural transformation field." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 658, + 212, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 212, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 212, + 671 + ], + "type": "text", + "content": "2.3. Diffusion Probabilistic Models" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "Diffusion model (DM) provides a powerful probabilistic generative framework for high-quality image synthesis, which is first introduced in [33] and subsequently improved" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 361 + ], + "type": "text", + "content": "by Denoising Diffusion Probabilistic Model (DDPM) [13], Denoising Diffusion Implicit Model (DDIM) [34], score-based diffusion [35], and other techniques [25, 36, 37]. Early diffusion models construct the diffusion process in RGB space, which requires huge computation powers to generate high-resolution images. Latent Diffusion Model (LDM) [30] builds the diffusion process in a lower-dimensional latent space, significantly reducing the computation cost. The diffusion model has become the dominant solution for various computer vision tasks, such as image synthesis [29, 30, 32], image editing [1, 6, 24], and image-to-image translation [31, 44]. Large-scale text-to-image (T2I) diffusion models (e.g. DALL-E 2 [29], Imagegen [32], Stable Diffusion [30]) have shown superior performance on high-quality image synthesis, which employ text embedding generated by large language models (LLM) as the condition to guide image generation. Recently, several works [19, 22, 52] modify the denoising diffusion process to model the residual between two given distributions. Different from these works, in this paper, we introduce the font transfer process into the diffusion model at a predefined region, forming a multi-stage diffusion process. Moreover, based on this process, a novel dual-network approach is proposed to predict latent features in different stages." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 376, + 389, + 390 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 376, + 389, + 390 + ], + "spans": [ + { + "bbox": [ + 306, + 376, + 389, + 390 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 398, + 546, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 398, + 546, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 398, + 546, + 460 + ], + "type": "text", + "content": "In this section, we present our proposed method in detail, including a short background of the diffusion model (in Sec. 3.1), the proposed latent diffusion model with font transfer process (in Sec. 3.2), the multi-stage font generative process (in Sec. 3.3), and the overall framework (in Sec. 3.4)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 472, + 466, + 485 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 472, + 466, + 485 + ], + "spans": [ + { + "bbox": [ + 305, + 472, + 466, + 485 + ], + "type": "text", + "content": "3.1. Background and Preliminary" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 491, + 546, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 491, + 546, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 491, + 546, + 540 + ], + "type": "text", + "content": "The diffusion model is a probabilistic generative model aiming to learn the data distribution by gradually denoising a variable sampled from a Gaussian distribution, which includes a forward process and a reverse process." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 542, + 547, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 542, + 547, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 542, + 547, + 602 + ], + "type": "text", + "content": "Given a data distribution " + }, + { + "bbox": [ + 304, + 542, + 547, + 602 + ], + "type": "inline_equation", + "content": "x_0 \\sim q(x_0)" + }, + { + "bbox": [ + 304, + 542, + 547, + 602 + ], + "type": "text", + "content": ", the Forward Process (also named as the diffusion process) is fixed to a Markov chain that gradually injects Gaussian noise to the data at time " + }, + { + "bbox": [ + 304, + 542, + 547, + 602 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 542, + 547, + 602 + ], + "type": "text", + "content": " with a predefined variance schedule " + }, + { + "bbox": [ + 304, + 542, + 547, + 602 + ], + "type": "inline_equation", + "content": "\\beta_t" + }, + { + "bbox": [ + 304, + 542, + 547, + 602 + ], + "type": "text", + "content": " according to the equations:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 351, + 614, + 545, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 614, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 351, + 614, + 545, + 647 + ], + "type": "interline_equation", + "content": "q \\left(x _ {1}, \\dots , x _ {T} \\mid x _ {0}\\right) = \\prod_ {t = 1} ^ {T} q \\left(x _ {t} \\mid x _ {t - 1}\\right), \\tag {1}", + "image_path": "7b0498e105ca815083723c0e826e0f53128819e5baeda77038f72268ca93af6e.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 341, + 650, + 545, + 665 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 650, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 341, + 650, + 545, + 665 + ], + "type": "interline_equation", + "content": "q \\left(x _ {t} \\mid x _ {t - 1}\\right) = \\mathcal {N} \\left(x _ {t}; \\sqrt {1 - \\beta_ {t}} x _ {t - 1}, \\beta_ {t} I\\right), \\tag {2}", + "image_path": "188a468fb7a914ae7a87d52ea3e4cd2304735f8df5908143f03d5b7a02dd0103.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "inline_equation", + "content": "x_{1},\\dots ,x_{T}" + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": " are latent variables. With the sufficiently large " + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": " , the distribution of " + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": " will become an isotropic Gaussian distribution. With the reparameterization trick" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6894" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 72, + 309, + 175 + ], + "blocks": [ + { + "bbox": [ + 51, + 72, + 309, + 175 + ], + "lines": [ + { + "bbox": [ + 51, + 72, + 309, + 175 + ], + "spans": [ + { + "bbox": [ + 51, + 72, + 309, + 175 + ], + "type": "image", + "image_path": "627114c7efa60a8d931cbc1c36fa66cfa37905ce8f42190ff549551889b0e5fd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 51, + 175, + 310, + 281 + ], + "blocks": [ + { + "bbox": [ + 51, + 175, + 310, + 281 + ], + "lines": [ + { + "bbox": [ + 51, + 175, + 310, + 281 + ], + "spans": [ + { + "bbox": [ + 51, + 175, + 310, + 281 + ], + "type": "image", + "image_path": "afd3bb041299288f31c8b3840a4a19ba32fb1cc13a40b11fb4e16c2aac93c4b4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 243, + 417, + 261 + ], + "lines": [ + { + "bbox": [ + 318, + 243, + 417, + 261 + ], + "spans": [ + { + "bbox": [ + 318, + 243, + 417, + 261 + ], + "type": "text", + "content": "Generative Step in Structure Construction Stage" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 318, + 263, + 405, + 280 + ], + "lines": [ + { + "bbox": [ + 318, + 263, + 405, + 280 + ], + "spans": [ + { + "bbox": [ + 318, + 263, + 405, + 280 + ], + "type": "text", + "content": "Generative Step in Font Transfer Stage" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 46, + 285, + 547, + 390 + ], + "lines": [ + { + "bbox": [ + 46, + 285, + 547, + 390 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 547, + 390 + ], + "type": "text", + "content": "Figure 2. (a). The forward diffusion process of our proposed model. (b). The reverse process of our proposed model. Our proposed diffusion process is marked as black circles. The " + }, + { + "bbox": [ + 46, + 285, + 547, + 390 + ], + "type": "inline_equation", + "content": "\\tilde{z}_t" + }, + { + "bbox": [ + 46, + 285, + 547, + 390 + ], + "type": "text", + "content": " denotes the latent features in our diffusion process, while " + }, + { + "bbox": [ + 46, + 285, + 547, + 390 + ], + "type": "inline_equation", + "content": "z_t^c" + }, + { + "bbox": [ + 46, + 285, + 547, + 390 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 285, + 547, + 390 + ], + "type": "inline_equation", + "content": "z_t^g" + }, + { + "bbox": [ + 46, + 285, + 547, + 390 + ], + "type": "text", + "content": " denote the latent features of the source image and the target image in traditional diffusion process, respectively. As we have aligned our diffusion process with the traditional diffusion process of " + }, + { + "bbox": [ + 46, + 285, + 547, + 390 + ], + "type": "inline_equation", + "content": "z^g" + }, + { + "bbox": [ + 46, + 285, + 547, + 390 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 285, + 547, + 390 + ], + "type": "inline_equation", + "content": "z^c" + }, + { + "bbox": [ + 46, + 285, + 547, + 390 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 46, + 285, + 547, + 390 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 46, + 285, + 547, + 390 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 285, + 547, + 390 + ], + "type": "inline_equation", + "content": "t_2" + }, + { + "bbox": [ + 46, + 285, + 547, + 390 + ], + "type": "text", + "content": ", our model shares the same process with the traditional diffusion process in the region " + }, + { + "bbox": [ + 46, + 285, + 547, + 390 + ], + "type": "inline_equation", + "content": "(0, t_1]" + }, + { + "bbox": [ + 46, + 285, + 547, + 390 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 285, + 547, + 390 + ], + "type": "inline_equation", + "content": "(t_2, T]" + }, + { + "bbox": [ + 46, + 285, + 547, + 390 + ], + "type": "text", + "content": ". (c). Overall framework of our font generation model, including the VQ-VAE encoder " + }, + { + "bbox": [ + 46, + 285, + 547, + 390 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 46, + 285, + 547, + 390 + ], + "type": "text", + "content": " and decoder " + }, + { + "bbox": [ + 46, + 285, + 547, + 390 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 285, + 547, + 390 + ], + "type": "text", + "content": ", the dual prediction networks " + }, + { + "bbox": [ + 46, + 285, + 547, + 390 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{\\theta_1}^{(g,1)}(\\tilde{z}_t, t, y_1)" + }, + { + "bbox": [ + 46, + 285, + 547, + 390 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 285, + 547, + 390 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{\\theta_2}^{(g,2)}(\\tilde{z}_t, t, y_2)" + }, + { + "bbox": [ + 46, + 285, + 547, + 390 + ], + "type": "text", + "content": ", and the corresponding style and content encoders for each prediction network. Due to the same structure, in this figure, we only plot one prediction network " + }, + { + "bbox": [ + 46, + 285, + 547, + 390 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{\\theta_i}^{(g,i)}" + }, + { + "bbox": [ + 46, + 285, + 547, + 390 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 285, + 547, + 390 + ], + "type": "inline_equation", + "content": "i \\in \\{1, 2\\}" + }, + { + "bbox": [ + 46, + 285, + 547, + 390 + ], + "type": "text", + "content": ") with its corresponding style encoder " + }, + { + "bbox": [ + 46, + 285, + 547, + 390 + ], + "type": "inline_equation", + "content": "E_s^i" + }, + { + "bbox": [ + 46, + 285, + 547, + 390 + ], + "type": "text", + "content": " and content encoder " + }, + { + "bbox": [ + 46, + 285, + 547, + 390 + ], + "type": "inline_equation", + "content": "E_c^i" + }, + { + "bbox": [ + 46, + 285, + 547, + 390 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 285, + 547, + 390 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 285, + 547, + 390 + ], + "type": "text", + "content": " is determined by the generative stage of timestep " + }, + { + "bbox": [ + 46, + 285, + 547, + 390 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 285, + 547, + 390 + ], + "type": "text", + "content": ". We use " + }, + { + "bbox": [ + 46, + 285, + 547, + 390 + ], + "type": "inline_equation", + "content": "I_g" + }, + { + "bbox": [ + 46, + 285, + 547, + 390 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 285, + 547, + 390 + ], + "type": "inline_equation", + "content": "I_G" + }, + { + "bbox": [ + 46, + 285, + 547, + 390 + ], + "type": "text", + "content": " to denote the target image and generated image, respectively." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 312, + 72, + 542, + 210 + ], + "blocks": [ + { + "bbox": [ + 312, + 72, + 542, + 210 + ], + "lines": [ + { + "bbox": [ + 312, + 72, + 542, + 210 + ], + "spans": [ + { + "bbox": [ + 312, + 72, + 542, + 210 + ], + "type": "image", + "image_path": "674cf8f1cd9dd4dd7e6d9a0a6e23812a5cdbf368c3e6f1cc6091fc37e1c4d34f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 212, + 416, + 241 + ], + "lines": [ + { + "bbox": [ + 317, + 212, + 416, + 241 + ], + "spans": [ + { + "bbox": [ + 317, + 212, + 416, + 241 + ], + "type": "inline_equation", + "content": "\\tilde{z}_t" + }, + { + "bbox": [ + 317, + 212, + 416, + 241 + ], + "type": "text", + "content": " Latent feature in our model \n" + }, + { + "bbox": [ + 317, + 212, + 416, + 241 + ], + "type": "inline_equation", + "content": "z_t^c" + }, + { + "bbox": [ + 317, + 212, + 416, + 241 + ], + "type": "text", + "content": " Latent feature of source \nimage in traditional model" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 429, + 213, + 539, + 243 + ], + "lines": [ + { + "bbox": [ + 429, + 213, + 539, + 243 + ], + "spans": [ + { + "bbox": [ + 429, + 213, + 539, + 243 + ], + "type": "inline_equation", + "content": "\\widetilde{Z}_{\\theta_i}^{(g,i)}" + }, + { + "bbox": [ + 429, + 213, + 539, + 243 + ], + "type": "text", + "content": " Prediction network " + }, + { + "bbox": [ + 429, + 213, + 539, + 243 + ], + "type": "inline_equation", + "content": "(i\\in \\{1,2\\})" + }, + { + "bbox": [ + 429, + 213, + 539, + 243 + ], + "type": "text", + "content": " Latent feature of target image " + }, + { + "bbox": [ + 429, + 213, + 539, + 243 + ], + "type": "inline_equation", + "content": "z_{t}^{g}" + }, + { + "bbox": [ + 429, + 213, + 539, + 243 + ], + "type": "text", + "content": " in traditional model" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 431, + 245, + 536, + 262 + ], + "lines": [ + { + "bbox": [ + 431, + 245, + 536, + 262 + ], + "spans": [ + { + "bbox": [ + 431, + 245, + 536, + 262 + ], + "type": "text", + "content": "The Equivalent Process for Structure Construction Stage" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 431, + 263, + 519, + 280 + ], + "lines": [ + { + "bbox": [ + 431, + 263, + 519, + 280 + ], + "spans": [ + { + "bbox": [ + 431, + 263, + 519, + 280 + ], + "type": "text", + "content": "Generative Step in Font Refinement Stage" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 401, + 227, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 401, + 227, + 414 + ], + "spans": [ + { + "bbox": [ + 47, + 401, + 227, + 414 + ], + "type": "text", + "content": "[16], the latent variable " + }, + { + "bbox": [ + 47, + 401, + 227, + 414 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 47, + 401, + 227, + 414 + ], + "type": "text", + "content": " can be written as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 93, + 420, + 287, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 420, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 93, + 420, + 287, + 449 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} x _ {t} = \\sqrt {\\alpha_ {t}} x _ {t - 1} + \\sqrt {1 - \\alpha_ {t}} \\epsilon_ {t - 1} ^ {*}, (3) \\\\ = \\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {0}, (4) \\\\ \\end{array}", + "image_path": "4420acefa60758240197848c60aeff1f8310c3dd184c90cb6eaec40ecc9ed291.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 455, + 287, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 455, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 455, + 287, + 491 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 455, + 287, + 491 + ], + "type": "inline_equation", + "content": "\\epsilon_{t - 1}^{*},\\epsilon_{0}\\sim \\mathcal{N}(0,I),\\alpha_{t} = 1 - \\beta_{t}" + }, + { + "bbox": [ + 47, + 455, + 287, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 455, + 287, + 491 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t = \\prod_{s = 1}^t\\alpha_s" + }, + { + "bbox": [ + 47, + 455, + 287, + 491 + ], + "type": "text", + "content": " Thus sampling " + }, + { + "bbox": [ + 47, + 455, + 287, + 491 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 47, + 455, + 287, + 491 + ], + "type": "text", + "content": " at arbitrary timestep has a closed form " + }, + { + "bbox": [ + 47, + 455, + 287, + 491 + ], + "type": "inline_equation", + "content": "q(x_{t}|x_{0}) = \\mathcal{N}(x_{t};\\sqrt{\\bar{\\alpha}_{t}} x_{0},(1 - \\bar{\\alpha}_{t})I)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 491, + 287, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 491, + 287, + 528 + ], + "spans": [ + { + "bbox": [ + 46, + 491, + 287, + 528 + ], + "type": "text", + "content": "The Reverse Process generates data from isotropic Gaussian noise " + }, + { + "bbox": [ + 46, + 491, + 287, + 528 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 46, + 491, + 287, + 528 + ], + "type": "text", + "content": ", and gradually recovers " + }, + { + "bbox": [ + 46, + 491, + 287, + 528 + ], + "type": "inline_equation", + "content": "x_{0}" + }, + { + "bbox": [ + 46, + 491, + 287, + 528 + ], + "type": "text", + "content": " via the reverse distribution " + }, + { + "bbox": [ + 46, + 491, + 287, + 528 + ], + "type": "inline_equation", + "content": "p_{\\theta}(x_{t - 1}|x_t)" + }, + { + "bbox": [ + 46, + 491, + 287, + 528 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 107, + 533, + 287, + 558 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 533, + 287, + 558 + ], + "spans": [ + { + "bbox": [ + 107, + 533, + 287, + 558 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(x _ {0}\\right) = \\int p _ {\\theta} \\left(x _ {0: T}\\right) d x _ {1: T}, \\tag {5}", + "image_path": "12a05e7c11eb66d2f03e64e3a89ef3a585731bad1eccb7ab581ad1e9d7257f21.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 92, + 560, + 287, + 593 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 560, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 92, + 560, + 287, + 593 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(x _ {0: T}\\right) = p _ {\\theta} \\left(x _ {T}\\right) \\prod_ {t = 1} ^ {T} p _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}\\right), \\tag {6}", + "image_path": "a1bd4011b2237d6046f29f6ef02a8ba4228ba7c18de5af742f105b1ca223b2f1.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 47, + 597, + 287, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 287, + 622 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 597, + 287, + 622 + ], + "type": "inline_equation", + "content": "p_{\\theta}(x_{t - 1}|x_t) = \\mathcal{N}(x_{t - 1}|\\mu_\\theta (x_t,t),\\Sigma_\\theta (x_t,t))" + }, + { + "bbox": [ + 47, + 597, + 287, + 622 + ], + "type": "text", + "content": " is a parameterized Gaussian transition network, and defined as" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 47, + 627, + 287, + 653 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 627, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 47, + 627, + 287, + 653 + ], + "type": "interline_equation", + "content": "\\mu_ {\\theta} (x _ {t}, t) = \\frac {\\sqrt {\\alpha_ {t}} (1 - \\bar {\\alpha} _ {t - 1}) x _ {t} + \\sqrt {\\bar {\\alpha} _ {t - 1}} (1 - \\alpha_ {t}) \\hat {x} _ {\\theta} (x _ {t} , t)}{1 - \\bar {\\alpha} _ {t}},", + "image_path": "115299674abf8ffd791c4b4ae796ee14184bb7f58a44a8c67b6f09734072271d.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 47, + 658, + 287, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 287, + 695 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 287, + 695 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 47, + 658, + 287, + 695 + ], + "type": "inline_equation", + "content": "\\Sigma_{\\theta}(x_t,t) = (1 - \\bar{\\alpha}_{t - 1})\\beta_tI / (1 - \\bar{\\alpha}_t)" + }, + { + "bbox": [ + 47, + 658, + 287, + 695 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 658, + 287, + 695 + ], + "type": "inline_equation", + "content": "\\hat{x}_{\\theta}(x_{t},t)" + }, + { + "bbox": [ + 47, + 658, + 287, + 695 + ], + "type": "text", + "content": " is a function approximator for predicting " + }, + { + "bbox": [ + 47, + 658, + 287, + 695 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 47, + 658, + 287, + 695 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 47, + 658, + 287, + 695 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 47, + 658, + 287, + 695 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 658, + 287, + 695 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 658, + 287, + 695 + ], + "type": "text", + "content": ". The corresponding objective function can be simplified to:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 99, + 700, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 700, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 99, + 700, + 287, + 715 + ], + "type": "interline_equation", + "content": "L = \\mathbb {E} _ {x _ {0}, \\epsilon , t} \\left[ \\| \\hat {x} _ {\\theta} (x _ {t}, t) - x _ {0} \\| _ {2} ^ {2} \\right]. \\tag {7}", + "image_path": "29d22478e40fcbd9a3c8491d03f56c86c20fb9e021f5e7e77a8adb8fe2bc757e.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 401, + 545, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 401, + 545, + 484 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 545, + 484 + ], + "type": "text", + "content": "Latent Diffusion Model (LDM) [30] further employs a VQ-VAE [42] encoder " + }, + { + "bbox": [ + 304, + 401, + 545, + 484 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 304, + 401, + 545, + 484 + ], + "type": "text", + "content": " to project image " + }, + { + "bbox": [ + 304, + 401, + 545, + 484 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 304, + 401, + 545, + 484 + ], + "type": "text", + "content": " into latent space " + }, + { + "bbox": [ + 304, + 401, + 545, + 484 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 304, + 401, + 545, + 484 + ], + "type": "text", + "content": ", and constructs the diffusion model in this lower-dimensional latent space, significantly reducing computational consumption. Once " + }, + { + "bbox": [ + 304, + 401, + 545, + 484 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 304, + 401, + 545, + 484 + ], + "type": "text", + "content": " is sampled via the reverse process, the VQ-VAE decoder " + }, + { + "bbox": [ + 304, + 401, + 545, + 484 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 401, + 545, + 484 + ], + "type": "text", + "content": " will project it back to the image space." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 485, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 545, + 521 + ], + "type": "text", + "content": "Since LDM significantly reduces computation consumption and maintains high-quality image synthesis, in this paper, we construct our diffusion process in latent space." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 527, + 529, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 527, + 529, + 539 + ], + "spans": [ + { + "bbox": [ + 306, + 527, + 529, + 539 + ], + "type": "text", + "content": "3.2. Latent Diffusion Model with Font Transfer" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 545, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 653 + ], + "type": "text", + "content": "As shown in Eq. (4), in the forward process, the data distribution " + }, + { + "bbox": [ + 304, + 545, + 545, + 653 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 304, + 545, + 545, + 653 + ], + "type": "text", + "content": " and the injected noise can be well-separated in arbitrary timesteps. This characteristic enables us to construct a multi-stage font generative framework in latent diffusion model, by incorporating the font transfer process into the data sub-space in a predefined time interval " + }, + { + "bbox": [ + 304, + 545, + 545, + 653 + ], + "type": "inline_equation", + "content": "t \\in (t_1, t_2]" + }, + { + "bbox": [ + 304, + 545, + 545, + 653 + ], + "type": "text", + "content": ". Based on this observation, we develop our latent diffusion model in this section, then discuss the multi-stage font generative process in the next section." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": "The Forward Process. After projecting the source images " + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": " and target images " + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "inline_equation", + "content": "I_{g}" + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": " into the latent features " + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "inline_equation", + "content": "z_0^c" + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "inline_equation", + "content": "z_0^g" + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": " by the VQ-VAE encoder " + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": ", we implement our model in latent space. As shown in Fig. 2(a), our proposed model (marked as black circles and denoted as " + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\tilde{z}_t" + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": ") is constructed" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6895" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": "by utilizing a font transfer process to connect the traditional diffusion processes (DDPM) of the latent features " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "z_{t}^{c}" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "z_{t}^{g}" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": ". Therefore, our latent diffusion process " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{t}" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": " can be divided into three regions: (1). In the region " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "t \\in (0, t_1]" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": ", only latent feature " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "z_{t}^{g}" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": " evolved with noise. The forward latent " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{t}" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": " in this region can be expressed as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 73, + 152, + 287, + 167 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 152, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 73, + 152, + 287, + 167 + ], + "type": "interline_equation", + "content": "\\tilde {z} _ {t} = z _ {t} ^ {g} = \\sqrt {\\bar {\\alpha} _ {t}} z _ {0} ^ {g} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {0} t \\in (0, t _ {1} ]. \\qquad (8)", + "image_path": "3e9958c010c0e4455f453e773aea89b617b9fd27467a8e8f14af1f23f66d6961.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 173, + 287, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 173, + 287, + 245 + ], + "spans": [ + { + "bbox": [ + 46, + 173, + 287, + 245 + ], + "type": "text", + "content": "(2). In the region " + }, + { + "bbox": [ + 46, + 173, + 287, + 245 + ], + "type": "inline_equation", + "content": "t \\in (t_1, t_2]" + }, + { + "bbox": [ + 46, + 173, + 287, + 245 + ], + "type": "text", + "content": ", together with the continuously injected noise, the latent feature " + }, + { + "bbox": [ + 46, + 173, + 287, + 245 + ], + "type": "inline_equation", + "content": "z_t^g" + }, + { + "bbox": [ + 46, + 173, + 287, + 245 + ], + "type": "text", + "content": " of the target font is gradually replaced by the latent feature " + }, + { + "bbox": [ + 46, + 173, + 287, + 245 + ], + "type": "inline_equation", + "content": "z_t^c" + }, + { + "bbox": [ + 46, + 173, + 287, + 245 + ], + "type": "text", + "content": " of the source character, and only " + }, + { + "bbox": [ + 46, + 173, + 287, + 245 + ], + "type": "inline_equation", + "content": "z_t^c" + }, + { + "bbox": [ + 46, + 173, + 287, + 245 + ], + "type": "text", + "content": " exists at " + }, + { + "bbox": [ + 46, + 173, + 287, + 245 + ], + "type": "inline_equation", + "content": "t_2" + }, + { + "bbox": [ + 46, + 173, + 287, + 245 + ], + "type": "text", + "content": ". After careful derivation (provided in Supplementary Materials), the forward latent " + }, + { + "bbox": [ + 46, + 173, + 287, + 245 + ], + "type": "inline_equation", + "content": "\\tilde{z}_t" + }, + { + "bbox": [ + 46, + 173, + 287, + 245 + ], + "type": "text", + "content": " in this region can be expressed as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 253, + 287, + 268 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 253, + 287, + 268 + ], + "spans": [ + { + "bbox": [ + 63, + 253, + 287, + 268 + ], + "type": "interline_equation", + "content": "\\tilde {z} _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} z _ {0} ^ {g} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {0} - \\left(z _ {0} ^ {g} - z _ {0} ^ {c}\\right) \\Psi \\mathcal {H} _ {t}, \\tag {9}", + "image_path": "ca97fed794f2503ff372fc4a2482f29993dcbbae9f76b2a74298ad2cede08c16.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 274, + 287, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 274, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 287, + 323 + ], + "type": "text", + "content": "for " + }, + { + "bbox": [ + 46, + 274, + 287, + 323 + ], + "type": "inline_equation", + "content": "t \\in (t_1, t_2]" + }, + { + "bbox": [ + 46, + 274, + 287, + 323 + ], + "type": "text", + "content": ". The notations " + }, + { + "bbox": [ + 46, + 274, + 287, + 323 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_t" + }, + { + "bbox": [ + 46, + 274, + 287, + 323 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 274, + 287, + 323 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 46, + 274, + 287, + 323 + ], + "type": "text", + "content": " are coefficients obtained by aligning our model with traditional diffusion processes (DDPM) of " + }, + { + "bbox": [ + 46, + 274, + 287, + 323 + ], + "type": "inline_equation", + "content": "z^g" + }, + { + "bbox": [ + 46, + 274, + 287, + 323 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 274, + 287, + 323 + ], + "type": "inline_equation", + "content": "z^c" + }, + { + "bbox": [ + 46, + 274, + 287, + 323 + ], + "type": "text", + "content": " at timesteps " + }, + { + "bbox": [ + 46, + 274, + 287, + 323 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 46, + 274, + 287, + 323 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 274, + 287, + 323 + ], + "type": "inline_equation", + "content": "t_2" + }, + { + "bbox": [ + 46, + 274, + 287, + 323 + ], + "type": "text", + "content": ", respectively. They are defined as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 52, + 331, + 287, + 345 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 331, + 287, + 345 + ], + "spans": [ + { + "bbox": [ + 52, + 331, + 287, + 345 + ], + "type": "interline_equation", + "content": "\\mathcal {H} _ {t} = 1 + \\sqrt {\\alpha_ {t}} + \\sqrt {\\alpha_ {t} \\alpha_ {t - 1}} + \\dots + \\sqrt {\\alpha_ {t} \\cdots \\alpha_ {t + 2}}, \\tag {10}", + "image_path": "a52518fc10b42c20ba2f63613672b22c5937e33f5923f7babb19c5bd086d4b9d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 99, + 347, + 287, + 377 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 347, + 287, + 377 + ], + "spans": [ + { + "bbox": [ + 99, + 347, + 287, + 377 + ], + "type": "interline_equation", + "content": "\\Psi = \\frac {\\sqrt {\\tilde {\\alpha} (1 , t _ {2})}}{1 + \\sum_ {m = t _ {1} + 2} ^ {t _ {2}} \\sqrt {\\tilde {\\alpha} (m , t _ {2})}}, \\tag {11}", + "image_path": "a3a0a424023803151b4134f4ac92b6ab2210ed073143d9f79d590aae07585967.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 385, + 287, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 385, + 287, + 446 + ], + "spans": [ + { + "bbox": [ + 46, + 385, + 287, + 446 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 385, + 287, + 446 + ], + "type": "inline_equation", + "content": "\\tilde{\\alpha}(t_1, t_2) = \\prod_{i=t_1}^{t_2} \\alpha_i" + }, + { + "bbox": [ + 46, + 385, + 287, + 446 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 385, + 287, + 446 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_t" + }, + { + "bbox": [ + 46, + 385, + 287, + 446 + ], + "type": "text", + "content": " can be further reduced to " + }, + { + "bbox": [ + 46, + 385, + 287, + 446 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_t = 1 + \\sqrt{\\alpha_t} \\mathcal{H}_{t-1}" + }, + { + "bbox": [ + 46, + 385, + 287, + 446 + ], + "type": "text", + "content": ". (3). In the region " + }, + { + "bbox": [ + 46, + 385, + 287, + 446 + ], + "type": "inline_equation", + "content": "t \\in (t_2, T]" + }, + { + "bbox": [ + 46, + 385, + 287, + 446 + ], + "type": "text", + "content": ", the injected noise finally disrupts the latent feature " + }, + { + "bbox": [ + 46, + 385, + 287, + 446 + ], + "type": "inline_equation", + "content": "z_t^c" + }, + { + "bbox": [ + 46, + 385, + 287, + 446 + ], + "type": "text", + "content": " to Gaussian noise, and the forward latent " + }, + { + "bbox": [ + 46, + 385, + 287, + 446 + ], + "type": "inline_equation", + "content": "\\tilde{z}_t" + }, + { + "bbox": [ + 46, + 385, + 287, + 446 + ], + "type": "text", + "content": " in this region can be expressed as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 64, + 454, + 287, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 454, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 64, + 454, + 287, + 468 + ], + "type": "interline_equation", + "content": "\\tilde {z} _ {t} = z _ {t} ^ {c} = \\sqrt {\\bar {\\alpha} _ {t}} z _ {0} ^ {c} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {0} \\quad t \\in (t _ {2}, T ]. \\tag {12}", + "image_path": "ea28a794bc53ea3918f4c244b9339b96e684f812127d2a09f753f837fd8b33cc.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 475, + 287, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 475, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 475, + 287, + 548 + ], + "type": "text", + "content": "The Reverse Process. Now we turn to discuss our choices for " + }, + { + "bbox": [ + 46, + 475, + 287, + 548 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\tilde{z}_{t - 1}|\\tilde{z}_t) = \\mathcal{N}(\\tilde{z}_{t - 1}|\\mu_\\theta (\\tilde{z}_t,t),\\Sigma_\\theta (\\tilde{z}_t,t))" + }, + { + "bbox": [ + 46, + 475, + 287, + 548 + ], + "type": "text", + "content": ". As shown in Eq. (8) and Eq. (12), in the time region " + }, + { + "bbox": [ + 46, + 475, + 287, + 548 + ], + "type": "inline_equation", + "content": "(0,t_1]" + }, + { + "bbox": [ + 46, + 475, + 287, + 548 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 475, + 287, + 548 + ], + "type": "inline_equation", + "content": "(t_2,T]" + }, + { + "bbox": [ + 46, + 475, + 287, + 548 + ], + "type": "text", + "content": ", we have aligned the forward latent " + }, + { + "bbox": [ + 46, + 475, + 287, + 548 + ], + "type": "inline_equation", + "content": "\\tilde{z}_t" + }, + { + "bbox": [ + 46, + 475, + 287, + 548 + ], + "type": "text", + "content": " with the traditional diffusion processes (DDPM [13]), thus they share the same parameterized mean functions: for " + }, + { + "bbox": [ + 46, + 475, + 287, + 548 + ], + "type": "inline_equation", + "content": "t\\in (0,t_1]" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 555, + 287, + 592 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 287, + 592 + ], + "type": "interline_equation", + "content": "\\mu_ {\\theta} = \\frac {\\sqrt {\\alpha_ {t}} (1 - \\bar {\\alpha} _ {t - 1}) \\tilde {z} _ {t} + \\sqrt {\\bar {\\alpha} _ {t - 1}} (1 - \\alpha_ {t}) \\tilde {z} _ {\\theta} ^ {(g)} (\\tilde {z} _ {t} , t)}{1 - \\bar {\\alpha} _ {t}}, \\tag {13}", + "image_path": "f93ccb6d164ae8f5d3227273fbc2cda6c4a3963331802c407440abce1b1e5706.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 594, + 287, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 594, + 287, + 631 + ], + "spans": [ + { + "bbox": [ + 47, + 594, + 287, + 631 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 594, + 287, + 631 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{\\theta}^{(g)}(\\tilde{z}_t,t)" + }, + { + "bbox": [ + 47, + 594, + 287, + 631 + ], + "type": "text", + "content": " is the prediction network for predicting " + }, + { + "bbox": [ + 47, + 594, + 287, + 631 + ], + "type": "inline_equation", + "content": "z_0^g" + }, + { + "bbox": [ + 47, + 594, + 287, + 631 + ], + "type": "text", + "content": " from noisy latent feature " + }, + { + "bbox": [ + 47, + 594, + 287, + 631 + ], + "type": "inline_equation", + "content": "\\tilde{z}_t" + }, + { + "bbox": [ + 47, + 594, + 287, + 631 + ], + "type": "text", + "content": " and time index " + }, + { + "bbox": [ + 47, + 594, + 287, + 631 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 594, + 287, + 631 + ], + "type": "text", + "content": ", while for " + }, + { + "bbox": [ + 47, + 594, + 287, + 631 + ], + "type": "inline_equation", + "content": "t\\in (t_2,T]" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 638, + 287, + 675 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 638, + 287, + 675 + ], + "spans": [ + { + "bbox": [ + 56, + 638, + 287, + 675 + ], + "type": "interline_equation", + "content": "\\mu_ {\\theta} = \\frac {\\sqrt {\\alpha_ {t}} (1 - \\bar {\\alpha} _ {t - 1}) \\tilde {z} _ {t} + \\sqrt {\\bar {\\alpha} _ {t - 1}} (1 - \\alpha_ {t}) \\tilde {z} _ {\\theta} ^ {(c)} (\\tilde {z} _ {t} , t)}{1 - \\bar {\\alpha} _ {t}}, \\tag {14}", + "image_path": "7702d4d0486153412d80ffe6ed06380b63e584f9e9cc6cb29a09a55226af5746.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 676, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 676, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 676, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 676, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{\\theta}^{(c)}(z_t,t)" + }, + { + "bbox": [ + 46, + 676, + 287, + 714 + ], + "type": "text", + "content": " is the prediction network for predicting " + }, + { + "bbox": [ + 46, + 676, + 287, + 714 + ], + "type": "inline_equation", + "content": "z_0^c" + }, + { + "bbox": [ + 46, + 676, + 287, + 714 + ], + "type": "text", + "content": " from noisy latent feature " + }, + { + "bbox": [ + 46, + 676, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\tilde{z}_t" + }, + { + "bbox": [ + 46, + 676, + 287, + 714 + ], + "type": "text", + "content": " and time index " + }, + { + "bbox": [ + 46, + 676, + 287, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 676, + 287, + 714 + ], + "type": "text", + "content": ". The parameterized mean function for the region " + }, + { + "bbox": [ + 46, + 676, + 287, + 714 + ], + "type": "inline_equation", + "content": "t\\in (t_1,t_2]" + }, + { + "bbox": [ + 46, + 676, + 287, + 714 + ], + "type": "text", + "content": " can be" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 72, + 429, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 429, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 429, + 84 + ], + "type": "text", + "content": "derivated via Bayes' theorem:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 89, + 546, + 169 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 89, + 546, + 169 + ], + "spans": [ + { + "bbox": [ + 306, + 89, + 546, + 169 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mu_ {\\theta} = \\frac {\\sqrt {\\alpha_ {t}} (1 - \\bar {\\alpha} _ {t - 1})}{1 - \\bar {\\alpha} _ {t}} \\tilde {z} _ {t} + \\frac {\\sqrt {\\alpha_ {t}} (1 - \\bar {\\alpha} _ {t - 1})}{1 - \\bar {\\alpha} _ {t}} \\Psi \\tilde {z} _ {\\theta} ^ {(g)} (\\tilde {z} _ {t}, t) \\\\ + \\frac {\\left[ \\mathcal {H} _ {t - 1} (1 - \\alpha_ {t}) - \\sqrt {\\alpha_ {t}} (1 - \\bar {\\alpha} _ {t - 1}) \\right]}{1 - \\bar {\\alpha} _ {t}} \\Psi z _ {0} ^ {c} \\\\ + \\frac {(\\sqrt {\\bar {\\alpha} _ {t - 1}} - \\mathcal {H} _ {t - 1} \\Psi) (1 - \\alpha_ {t})}{1 - \\bar {\\alpha} _ {t}} \\tilde {z} _ {\\theta} ^ {(g)} (\\tilde {z} _ {t}, t). \\tag {15} \\\\ \\end{array}", + "image_path": "264fdebfcf58b195e6f06a44ed1499cea5014b737c49a44e8dfca76e2ff9c268.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 173, + 545, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 173, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 305, + 173, + 545, + 209 + ], + "type": "text", + "content": "As we only incorporate the font transfer process in the data sub-space and keep the noise sub-space unchanged, the variance in the reverse process remains the same as DDPM:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 339, + 213, + 545, + 239 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 213, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 339, + 213, + 545, + 239 + ], + "type": "interline_equation", + "content": "\\Sigma_ {\\theta} \\left(\\tilde {z} _ {t}, t\\right) = \\sigma_ {t} ^ {2} I = \\frac {(1 - \\alpha_ {t}) (1 - \\bar {\\alpha} _ {t - 1})}{1 - \\bar {\\alpha} _ {t}} I. \\tag {16}", + "image_path": "9ce2a64a7edb6f1de15d0883ddf2451dedd6011e9c9272ae5b3e42e47b9ef158.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 243, + 545, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 243, + 545, + 290 + ], + "spans": [ + { + "bbox": [ + 305, + 243, + 545, + 290 + ], + "type": "text", + "content": "Therefore, the latent " + }, + { + "bbox": [ + 305, + 243, + 545, + 290 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{t-1}" + }, + { + "bbox": [ + 305, + 243, + 545, + 290 + ], + "type": "text", + "content": " in reverse process can be obtained via the formula " + }, + { + "bbox": [ + 305, + 243, + 545, + 290 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\tilde{z}_{t-1}|\\tilde{z}_t) = \\mathcal{N}(\\tilde{z}_{t-1};\\mu_{\\theta}(\\tilde{z}_t,t),\\sigma_t^2 I)" + }, + { + "bbox": [ + 305, + 243, + 545, + 290 + ], + "type": "text", + "content": ", where the " + }, + { + "bbox": [ + 305, + 243, + 545, + 290 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}(\\tilde{z}_t,t)" + }, + { + "bbox": [ + 305, + 243, + 545, + 290 + ], + "type": "text", + "content": " is calculated from Eq. (13)-(15) based on different timesteps." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 291, + 546, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 291, + 546, + 426 + ], + "spans": [ + { + "bbox": [ + 305, + 291, + 546, + 426 + ], + "type": "text", + "content": "The Optimization Objective. As the forward behaviors keep the same as DDPM in the region " + }, + { + "bbox": [ + 305, + 291, + 546, + 426 + ], + "type": "inline_equation", + "content": "(0, t_1]" + }, + { + "bbox": [ + 305, + 291, + 546, + 426 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 291, + 546, + 426 + ], + "type": "inline_equation", + "content": "(t_2, T]" + }, + { + "bbox": [ + 305, + 291, + 546, + 426 + ], + "type": "text", + "content": ", the optimization objectives of " + }, + { + "bbox": [ + 305, + 291, + 546, + 426 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{\\theta}^{(g)}(\\tilde{z}_t, t)" + }, + { + "bbox": [ + 305, + 291, + 546, + 426 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 291, + 546, + 426 + ], + "type": "inline_equation", + "content": "\\hat{z}_{\\theta}^{(c)}(\\tilde{z}_t, t)" + }, + { + "bbox": [ + 305, + 291, + 546, + 426 + ], + "type": "text", + "content": " have the same expressions as DDPM in these regions. For the optimization objective of " + }, + { + "bbox": [ + 305, + 291, + 546, + 426 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{\\theta}^{(g)}(\\tilde{z}_t, t)" + }, + { + "bbox": [ + 305, + 291, + 546, + 426 + ], + "type": "text", + "content": " in the region " + }, + { + "bbox": [ + 305, + 291, + 546, + 426 + ], + "type": "inline_equation", + "content": "(t_1, t_2]" + }, + { + "bbox": [ + 305, + 291, + 546, + 426 + ], + "type": "text", + "content": ", it can be derived from the variational bound on negative log-likelihood [13], and only the coefficient is different with the DDPM's loss function. Moreover, as DDPM and LDM adopt the simplified training objectives in their implementations, we also simplify our loss functions and obtain the optimization objective as" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 395, + 432, + 545, + 445 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 432, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 395, + 432, + 545, + 445 + ], + "type": "interline_equation", + "content": "L = L _ {g} + L _ {c}, \\tag {17}", + "image_path": "9a5199a46432b839da4fcf2aef7dd167d51a14219763ba9d5ee8c877a45fe176.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 323, + 448, + 545, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 448, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 323, + 448, + 545, + 468 + ], + "type": "interline_equation", + "content": "L _ {g} = \\mathbb {E} _ {\\mathcal {E} (x _ {0}), \\epsilon , t \\sim U (0, t _ {2} ]} \\left[ \\| \\tilde {z} _ {\\theta} ^ {(g)} (\\tilde {z} _ {t}, t) - z _ {0} ^ {g} \\| _ {2} ^ {2} \\right], \\tag {18}", + "image_path": "c0dbbdf593c4850d5273895c3832445a488e344da7661442c0b098a28355579b.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 323, + 469, + 545, + 490 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 469, + 545, + 490 + ], + "spans": [ + { + "bbox": [ + 323, + 469, + 545, + 490 + ], + "type": "interline_equation", + "content": "L _ {c} = \\mathbb {E} _ {\\mathcal {E} (x _ {0}), \\epsilon , t \\sim U (t _ {2}, T ]} \\left[ \\| \\tilde {z} _ {\\theta} ^ {(c)} (\\tilde {z} _ {t}, t) - z _ {0} ^ {c} \\| _ {2} ^ {2} \\right], \\tag {19}", + "image_path": "1325e343fe952df5a988282f86788ffecd483b28c7fc8f33c267bba762e42c10.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 494, + 545, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 494, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 306, + 494, + 545, + 517 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 494, + 545, + 517 + ], + "type": "inline_equation", + "content": "U(t_{1},t_{2}]" + }, + { + "bbox": [ + 306, + 494, + 545, + 517 + ], + "type": "text", + "content": " denotes the uniform distribution in region " + }, + { + "bbox": [ + 306, + 494, + 545, + 517 + ], + "type": "inline_equation", + "content": "(t_1,t_2]" + }, + { + "bbox": [ + 306, + 494, + 545, + 517 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 518, + 545, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 518, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 305, + 518, + 545, + 563 + ], + "type": "text", + "content": "The Conditioning Mechanism. Following the LDM [30], we generalize our method to model conditional distributions by implementing with the conditional prediction networks:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 374, + 563, + 545, + 579 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 563, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 374, + 563, + 545, + 579 + ], + "type": "interline_equation", + "content": "\\tilde {z} _ {\\theta} ^ {(i)} \\left(\\tilde {z} _ {t}, t, y\\right), i \\in \\{g, c \\}, \\tag {20}", + "image_path": "9f7af5032896eae9ba0cdf223d7bd1e63bb67758ceeccf9e57d73c0a20579037.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 305, + 582, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 582, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 305, + 582, + 545, + 604 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 582, + 545, + 604 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 305, + 582, + 545, + 604 + ], + "type": "text", + "content": " is the condition for controlling the synthesis process." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Extension to Conditional Font Generation. We further extend this conditional model to perform the font generation task. Since the few-shot font generation task aims to render the source character " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " into the target font style by a few reference samples " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": ", we utilize the style encoder " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "E_{s}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " and the content encoder " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "E_{c}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " to extract the style code from the reference images " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " and the content representations from the source images " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": ". The style and content features are flattened and concatenated to build condition " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6896" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 242, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 242, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 242, + 85 + ], + "type": "text", + "content": "3.3. Multi-Stage Font Generative Process" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 96, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 96, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 96, + 287, + 144 + ], + "type": "text", + "content": "Based on our proposed model, the font generation in the reverse process can be automatically separated into three different stages: the structure construction stage, the font transfer stage, and the font refinement stage." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 149, + 289, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 149, + 289, + 418 + ], + "spans": [ + { + "bbox": [ + 46, + 149, + 289, + 418 + ], + "type": "text", + "content": "As shown in Fig. 2(b), starting with the sampled Gaussian noise " + }, + { + "bbox": [ + 46, + 149, + 289, + 418 + ], + "type": "inline_equation", + "content": "\\tilde{z}_T" + }, + { + "bbox": [ + 46, + 149, + 289, + 418 + ], + "type": "text", + "content": ", in the region " + }, + { + "bbox": [ + 46, + 149, + 289, + 418 + ], + "type": "inline_equation", + "content": "t \\in (t_2, T]" + }, + { + "bbox": [ + 46, + 149, + 289, + 418 + ], + "type": "text", + "content": ", Structure Construction Stage generates the structure information for the target character via predicting the latent feature " + }, + { + "bbox": [ + 46, + 149, + 289, + 418 + ], + "type": "inline_equation", + "content": "z_0^c" + }, + { + "bbox": [ + 46, + 149, + 289, + 418 + ], + "type": "text", + "content": " of the source image " + }, + { + "bbox": [ + 46, + 149, + 289, + 418 + ], + "type": "inline_equation", + "content": "I_c" + }, + { + "bbox": [ + 46, + 149, + 289, + 418 + ], + "type": "text", + "content": ". However, since we can access the source image " + }, + { + "bbox": [ + 46, + 149, + 289, + 418 + ], + "type": "inline_equation", + "content": "I_c" + }, + { + "bbox": [ + 46, + 149, + 289, + 418 + ], + "type": "text", + "content": " during the inference, the iterative generative steps from " + }, + { + "bbox": [ + 46, + 149, + 289, + 418 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 149, + 289, + 418 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 149, + 289, + 418 + ], + "type": "inline_equation", + "content": "t_2" + }, + { + "bbox": [ + 46, + 149, + 289, + 418 + ], + "type": "text", + "content": " can be replaced by a single forward step to generate the noised latent variable " + }, + { + "bbox": [ + 46, + 149, + 289, + 418 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{t_2}" + }, + { + "bbox": [ + 46, + 149, + 289, + 418 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 46, + 149, + 289, + 418 + ], + "type": "inline_equation", + "content": "t = t_2" + }, + { + "bbox": [ + 46, + 149, + 289, + 418 + ], + "type": "text", + "content": " via Eq. (12). This will significantly reduce inference time and provide more accurate structure information. Font Transfer Stage performs font transformation from the timestep " + }, + { + "bbox": [ + 46, + 149, + 289, + 418 + ], + "type": "inline_equation", + "content": "t_2" + }, + { + "bbox": [ + 46, + 149, + 289, + 418 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 149, + 289, + 418 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 46, + 149, + 289, + 418 + ], + "type": "text", + "content": ", where the noised latent feature " + }, + { + "bbox": [ + 46, + 149, + 289, + 418 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{t_2}" + }, + { + "bbox": [ + 46, + 149, + 289, + 418 + ], + "type": "text", + "content": " is gradually transformed to the noised latent feature " + }, + { + "bbox": [ + 46, + 149, + 289, + 418 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{t_1}" + }, + { + "bbox": [ + 46, + 149, + 289, + 418 + ], + "type": "text", + "content": " of target images via the " + }, + { + "bbox": [ + 46, + 149, + 289, + 418 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{\\theta}^{(g)}(\\tilde{z}_t, t, y)" + }, + { + "bbox": [ + 46, + 149, + 289, + 418 + ], + "type": "text", + "content": ". To predict the latent feature " + }, + { + "bbox": [ + 46, + 149, + 289, + 418 + ], + "type": "inline_equation", + "content": "z_0^g" + }, + { + "bbox": [ + 46, + 149, + 289, + 418 + ], + "type": "text", + "content": ", the prediction network " + }, + { + "bbox": [ + 46, + 149, + 289, + 418 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{\\theta}^{(g)}(\\tilde{z}_t, t, y)" + }, + { + "bbox": [ + 46, + 149, + 289, + 418 + ], + "type": "text", + "content": " in this stage is optimized to not only denoise the latent features but also perform font transfer under the condition " + }, + { + "bbox": [ + 46, + 149, + 289, + 418 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 149, + 289, + 418 + ], + "type": "text", + "content": ". Finally, in the region " + }, + { + "bbox": [ + 46, + 149, + 289, + 418 + ], + "type": "inline_equation", + "content": "t \\in (0, t_1]" + }, + { + "bbox": [ + 46, + 149, + 289, + 418 + ], + "type": "text", + "content": ", Font Refinement Stage improves appearances and local details for the generated font images via predicting " + }, + { + "bbox": [ + 46, + 149, + 289, + 418 + ], + "type": "inline_equation", + "content": "z_0^g" + }, + { + "bbox": [ + 46, + 149, + 289, + 418 + ], + "type": "text", + "content": " from the noisy latent feature " + }, + { + "bbox": [ + 46, + 149, + 289, + 418 + ], + "type": "inline_equation", + "content": "\\tilde{z}_t" + }, + { + "bbox": [ + 46, + 149, + 289, + 418 + ], + "type": "text", + "content": ". Based on the above three stages, we can implement the reverse process to generate high-quality font images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 423, + 287, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 423, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 47, + 423, + 287, + 459 + ], + "type": "text", + "content": "To better present this multi-stage font generative process, we summarize above steps and provide the pseudo-code in Supplementary Materials." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 483, + 259, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 483, + 259, + 496 + ], + "spans": [ + { + "bbox": [ + 47, + 483, + 259, + 496 + ], + "type": "text", + "content": "3.4. Overall Framework for Font Generation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "text", + "content": "As discussed in Sec. 3.3, the network " + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{\\theta}^{(c)}(\\tilde{z}_t,t,y)" + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "text", + "content": " can be replaced by a single forward step. Moreover, instead of implementing a single prediction network " + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{\\theta}^{(g)}(\\tilde{z}_t,t,y)" + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "text", + "content": " to predict " + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "inline_equation", + "content": "z_0^g" + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "text", + "content": " we adopt a dual-network approach in this paper, since the network " + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{\\theta}^{(g)}(\\tilde{z}_t,t,y)" + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "text", + "content": " has different behaviors in the font transfer stage and the font refinement stage. The dual networks " + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{\\theta_1}^{(g,1)}(\\tilde{z}_t,t,y_1)" + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{\\theta_2}^{(g,2)}(\\tilde{z}_t,t,y_2)" + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "text", + "content": " share the same structure, but are trained to predict the latent feature " + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "inline_equation", + "content": "z_0^g" + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "text", + "content": " in different regions: one for font transfer stage and the other for font refinement stage. Therefore, as shown in Fig. 2(c), the overall framework of our font generation model contains the following parts: the VQ-VAE encoder " + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "text", + "content": ", the VQ-VAE decoder " + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "text", + "content": ", the dual prediction networks " + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{\\theta_1}^{(g,1)}(\\tilde{z}_t,t,y_1)" + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{\\theta_2}^{(g,2)}(\\tilde{z}_t,t,y_2)" + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "text", + "content": ", and the corresponding style and content encoders for each prediction network " + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{\\theta_i}^{(g,i)}(\\tilde{z}_t,t,y_i)" + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "text", + "content": ". Finally, from Eq. (17), the optimization objective of our font gener" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 72, + 459, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 459, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 459, + 83 + ], + "type": "text", + "content": "ative model can be further reduced to" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 395, + 88, + 545, + 104 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 88, + 545, + 104 + ], + "spans": [ + { + "bbox": [ + 395, + 88, + 545, + 104 + ], + "type": "interline_equation", + "content": "L = L _ {g} ^ {1} + L _ {g} ^ {2}, \\tag {21}", + "image_path": "aa15b33178a64a059738b8f371f81270f175743b634030fbda00d1f52ea65402.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 312, + 106, + 545, + 125 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 106, + 545, + 125 + ], + "spans": [ + { + "bbox": [ + 312, + 106, + 545, + 125 + ], + "type": "interline_equation", + "content": "L _ {g} ^ {1} = \\mathbb {E} _ {\\mathcal {E} (x _ {0}), \\epsilon , t \\sim U (t _ {1}, t _ {2} ]} \\left[ \\| \\tilde {z} _ {\\theta_ {1}} ^ {(g, 1)} (\\tilde {z} _ {t}, t, y _ {1}) - z _ {0} ^ {g} \\| _ {2} ^ {2} \\right], \\tag {22}", + "image_path": "6a13c3f6ad7310f988685e4426fec0c0c24530c148de152acf4f5521034ccf5a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 128, + 545, + 148 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 128, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 313, + 128, + 545, + 148 + ], + "type": "interline_equation", + "content": "L _ {g} ^ {2} = \\mathbb {E} _ {\\mathcal {E} (x _ {0}), \\epsilon , t \\sim U (0, t _ {1} ]} \\left[ \\| \\tilde {z} _ {\\theta_ {2}} ^ {(g, 2)} (\\tilde {z} _ {t}, t, y _ {2}) - z _ {0} ^ {g} \\| _ {2} ^ {2} \\right]. \\tag {23}", + "image_path": "ac502ca37559fe40efdcc368e03f9367c19a4e728a4ff0cc272797fc292e9994.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 152, + 388, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 152, + 388, + 166 + ], + "spans": [ + { + "bbox": [ + 306, + 152, + 388, + 166 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 172, + 545, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 172, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 305, + 172, + 545, + 196 + ], + "type": "text", + "content": "In this section, we conduct extensive experiments to evaluate our proposed model on the FFG task." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 202, + 415, + 215 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 202, + 415, + 215 + ], + "spans": [ + { + "bbox": [ + 305, + 202, + 415, + 215 + ], + "type": "text", + "content": "4.1. Implement Details" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 220, + 545, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 220, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 305, + 220, + 545, + 334 + ], + "type": "text", + "content": "Our models are implemented on the Stable Diffusion [30] platform. We utilize " + }, + { + "bbox": [ + 305, + 220, + 545, + 334 + ], + "type": "inline_equation", + "content": "T = 1000" + }, + { + "bbox": [ + 305, + 220, + 545, + 334 + ], + "type": "text", + "content": " steps to generate " + }, + { + "bbox": [ + 305, + 220, + 545, + 334 + ], + "type": "inline_equation", + "content": "128 \\times 128 \\times 3" + }, + { + "bbox": [ + 305, + 220, + 545, + 334 + ], + "type": "text", + "content": " font images with 8 reference samples. The VQ-VAE encoder " + }, + { + "bbox": [ + 305, + 220, + 545, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 305, + 220, + 545, + 334 + ], + "type": "text", + "content": " and decoder " + }, + { + "bbox": [ + 305, + 220, + 545, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 305, + 220, + 545, + 334 + ], + "type": "text", + "content": ", the dual-network " + }, + { + "bbox": [ + 305, + 220, + 545, + 334 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{\\theta_1}^{(g,1)}(\\tilde{z}_t,t,y_1)" + }, + { + "bbox": [ + 305, + 220, + 545, + 334 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 220, + 545, + 334 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{\\theta_2}^{(g,2)}(\\tilde{z}_t,t,y_2)" + }, + { + "bbox": [ + 305, + 220, + 545, + 334 + ], + "type": "text", + "content": " are initialized by the pre-trained Stable Diffusion models. The style encoder " + }, + { + "bbox": [ + 305, + 220, + 545, + 334 + ], + "type": "inline_equation", + "content": "E_{s}" + }, + { + "bbox": [ + 305, + 220, + 545, + 334 + ], + "type": "text", + "content": " and content encoder " + }, + { + "bbox": [ + 305, + 220, + 545, + 334 + ], + "type": "inline_equation", + "content": "E_{c}" + }, + { + "bbox": [ + 305, + 220, + 545, + 334 + ], + "type": "text", + "content": " are initialized by Kaiming initialization. We keep VQ-VAE encoder " + }, + { + "bbox": [ + 305, + 220, + 545, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 305, + 220, + 545, + 334 + ], + "type": "text", + "content": " and decoder " + }, + { + "bbox": [ + 305, + 220, + 545, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 305, + 220, + 545, + 334 + ], + "type": "text", + "content": " frozen and optimize the remaining networks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 339, + 477, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 339, + 477, + 350 + ], + "spans": [ + { + "bbox": [ + 306, + 339, + 477, + 350 + ], + "type": "text", + "content": "4.2. Dataset and Evaluation Metrics" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 357, + 545, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 357, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 304, + 357, + 545, + 476 + ], + "type": "text", + "content": "Datasets: As DM needs more training data to optimize, we collect 900 Chinese fonts to evaluate our method on FFG task. We randomly select 840 fonts as the training set (seen fonts) and the remaining 60 fonts as the testing fonts (unseen fonts). As each collected font covers different Chinese characters, we randomly sample 800 characters from the commonly-used Chinese characters (the first-level Chinese Character Table) to train our method. For the test set, we use 214 characters as the unseen contents and sample 200 characters in our training set as the seen contents." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 477, + 545, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 477, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 477, + 545, + 548 + ], + "type": "text", + "content": "Therefore, the training set contains 840 fonts with 800 characters, while the test set includes two parts: one is the Unseen Fonts Unseen Contents (UFUC) test set, containing 60 unseen fonts with 214 unseen characters; the other is the Unseen Fonts Seen Contents (UFSC) test set, containing 60 unseen fonts with 200 seen characters." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 548, + 545, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 548, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 304, + 548, + 545, + 609 + ], + "type": "text", + "content": "Evaluation Metrics: In this paper, we utilize four commonly used measurements to evaluate our model, including root mean squared error (RMSE), peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and learned perceptual image path similarity (LPIPS) [48]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 614, + 400, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 614, + 400, + 627 + ], + "spans": [ + { + "bbox": [ + 306, + 614, + 400, + 627 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 633, + 545, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 633, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 305, + 633, + 545, + 657 + ], + "type": "text", + "content": "With the above dataset and metrics, we conducted extensive experiments in different settings." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 670, + 520, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 520, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 520, + 682 + ], + "type": "text", + "content": "4.3.1 Ablation Study on the Overall Framework" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "We first construct a Base Model with traditional diffusion process (DDPM) to verify the effectiveness of our proposed" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6897" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 69, + 545, + 182 + ], + "blocks": [ + { + "bbox": [ + 50, + 69, + 545, + 182 + ], + "lines": [ + { + "bbox": [ + 50, + 69, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 50, + 69, + 545, + 182 + ], + "type": "image", + "image_path": "036e1b68f8be3f20343b7849c20f43c7da10feb07e760b5a3ff191d7b37559c5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 185, + 517, + 196 + ], + "lines": [ + { + "bbox": [ + 74, + 185, + 517, + 196 + ], + "spans": [ + { + "bbox": [ + 74, + 185, + 517, + 196 + ], + "type": "text", + "content": "Figure 3. Visualization results of the intermediate results " + }, + { + "bbox": [ + 74, + 185, + 517, + 196 + ], + "type": "inline_equation", + "content": "\\tilde{z}_t" + }, + { + "bbox": [ + 74, + 185, + 517, + 196 + ], + "type": "text", + "content": " by projecting " + }, + { + "bbox": [ + 74, + 185, + 517, + 196 + ], + "type": "inline_equation", + "content": "\\tilde{z}_t" + }, + { + "bbox": [ + 74, + 185, + 517, + 196 + ], + "type": "text", + "content": " back to image space via VQ-VAE decoder " + }, + { + "bbox": [ + 74, + 185, + 517, + 196 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 74, + 185, + 517, + 196 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 72, + 205, + 261, + 318 + ], + "blocks": [ + { + "bbox": [ + 72, + 205, + 261, + 318 + ], + "lines": [ + { + "bbox": [ + 72, + 205, + 261, + 318 + ], + "spans": [ + { + "bbox": [ + 72, + 205, + 261, + 318 + ], + "type": "image", + "image_path": "1bbb281e3f0dfe20f2db227bb746f82f4c7d078d53527a2eb62b3050abfd947d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 321, + 287, + 354 + ], + "lines": [ + { + "bbox": [ + 46, + 321, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 321, + 287, + 354 + ], + "type": "text", + "content": "Figure 4. Visualization results of the predicted latent feature " + }, + { + "bbox": [ + 46, + 321, + 287, + 354 + ], + "type": "inline_equation", + "content": "z_0^g" + }, + { + "bbox": [ + 46, + 321, + 287, + 354 + ], + "type": "text", + "content": " in the intermediate timesteps by projecting " + }, + { + "bbox": [ + 46, + 321, + 287, + 354 + ], + "type": "inline_equation", + "content": "z_0^g" + }, + { + "bbox": [ + 46, + 321, + 287, + 354 + ], + "type": "text", + "content": " back to image space via VQ-VAE decoder " + }, + { + "bbox": [ + 46, + 321, + 287, + 354 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 321, + 287, + 354 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 61, + 390, + 272, + 440 + ], + "blocks": [ + { + "bbox": [ + 47, + 368, + 287, + 389 + ], + "lines": [ + { + "bbox": [ + 47, + 368, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 47, + 368, + 287, + 389 + ], + "type": "text", + "content": "Table 1. Ablation study for the overall framework on UFUC dataset. The bold number indicates the best." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 390, + 272, + 440 + ], + "lines": [ + { + "bbox": [ + 61, + 390, + 272, + 440 + ], + "spans": [ + { + "bbox": [ + 61, + 390, + 272, + 440 + ], + "type": "table", + "html": "
MethodsRMSE↓PSNR↑SSIM↑LPIPS↓
Base Model0.267611.720.69610.1561
Single-Net0.260112.020.70960.1531
Dual-Net0.247512.450.71890.1527
", + "image_path": "c5bde9dce3f79063a994aec793d1d2fcb1e974141bacf212ce9b5406093d1d78.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 457, + 287, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 457, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 457, + 287, + 529 + ], + "type": "text", + "content": "latent diffusion model. Moreover, as discussed in Sec. 3.4, there are two different selections to obtain " + }, + { + "bbox": [ + 46, + 457, + 287, + 529 + ], + "type": "inline_equation", + "content": "z_0^g" + }, + { + "bbox": [ + 46, + 457, + 287, + 529 + ], + "type": "text", + "content": " in the reverse process: One utilizes a single network to predict " + }, + { + "bbox": [ + 46, + 457, + 287, + 529 + ], + "type": "inline_equation", + "content": "z_0^g" + }, + { + "bbox": [ + 46, + 457, + 287, + 529 + ], + "type": "text", + "content": " at all timesteps. The other adopts the dual networks to predict the latent feature " + }, + { + "bbox": [ + 46, + 457, + 287, + 529 + ], + "type": "inline_equation", + "content": "z_0^g" + }, + { + "bbox": [ + 46, + 457, + 287, + 529 + ], + "type": "text", + "content": " in font transfer stage and font refinement stage, respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 530, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 530, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 530, + 287, + 639 + ], + "type": "text", + "content": "As shown in Tab. 1, compared with the Base Model, our Single-Net model improves the Base Model with 0.0075, 0.30, 0.0135, and 0.0030 in terms of RMSE, PSNR, SSIM, and LPIPS, verifying the effectiveness of our proposed latent diffusion model. Moreover, the Dual-Net approach further improves the Single-Net model with 0.0126, 0.43, and 0.0093 in terms of RMSE, PSNR, and SSIM, illustrating that the use of separate models is beneficial for predicting " + }, + { + "bbox": [ + 46, + 530, + 287, + 639 + ], + "type": "inline_equation", + "content": "z_0^g" + }, + { + "bbox": [ + 46, + 530, + 287, + 639 + ], + "type": "text", + "content": " in different stages." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 657, + 287, + 680 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 657, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 47, + 657, + 287, + 680 + ], + "type": "text", + "content": "4.3.2 Ablation Study on the Multi-Stage Font Generative Process" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "The selections of the region for different stages in the multistage generative process are hyper-parameters in our model," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 207, + 545, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 207, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 304, + 207, + 545, + 316 + ], + "type": "text", + "content": "which significantly impact image quality. Therefore, we study the influence on model performance by selecting different " + }, + { + "bbox": [ + 304, + 207, + 545, + 316 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 304, + 207, + 545, + 316 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 207, + 545, + 316 + ], + "type": "inline_equation", + "content": "t_2" + }, + { + "bbox": [ + 304, + 207, + 545, + 316 + ], + "type": "text", + "content": ". Due to the page limitation, the experimental results and the detailed discussions are provided in supplementary materials, and we briefly summarize the conclusion: The hyper-parameters " + }, + { + "bbox": [ + 304, + 207, + 545, + 316 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 304, + 207, + 545, + 316 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 207, + 545, + 316 + ], + "type": "inline_equation", + "content": "t_2" + }, + { + "bbox": [ + 304, + 207, + 545, + 316 + ], + "type": "text", + "content": " are trade-off selections, and the configuration " + }, + { + "bbox": [ + 304, + 207, + 545, + 316 + ], + "type": "inline_equation", + "content": "t_1 = 200" + }, + { + "bbox": [ + 304, + 207, + 545, + 316 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 207, + 545, + 316 + ], + "type": "inline_equation", + "content": "t_2 = 800" + }, + { + "bbox": [ + 304, + 207, + 545, + 316 + ], + "type": "text", + "content": " achieves the best record. Thus, we select this configuration for the following experiments." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 320, + 428, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 320, + 428, + 331 + ], + "spans": [ + { + "bbox": [ + 305, + 320, + 428, + 331 + ], + "type": "text", + "content": "4.3.3 Visualization Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 339, + 545, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 339, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 304, + 339, + 545, + 363 + ], + "type": "text", + "content": "In this section, we provide the visualization results of our generative process to verify the effectiveness of our model." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 364, + 545, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 364, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 364, + 545, + 506 + ], + "type": "text", + "content": "As our diffusion process is implemented in latent space, we visualize the intermediate generative results " + }, + { + "bbox": [ + 304, + 364, + 545, + 506 + ], + "type": "inline_equation", + "content": "\\tilde{z}_t" + }, + { + "bbox": [ + 304, + 364, + 545, + 506 + ], + "type": "text", + "content": " by using the VQ-VAE decoder " + }, + { + "bbox": [ + 304, + 364, + 545, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 364, + 545, + 506 + ], + "type": "text", + "content": " to project them back into the image space. As shown in Fig. 3, the images in the first and third rows are the intermediate results for source image " + }, + { + "bbox": [ + 304, + 364, + 545, + 506 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 304, + 364, + 545, + 506 + ], + "type": "text", + "content": " and target image " + }, + { + "bbox": [ + 304, + 364, + 545, + 506 + ], + "type": "inline_equation", + "content": "I_{g}" + }, + { + "bbox": [ + 304, + 364, + 545, + 506 + ], + "type": "text", + "content": " generated by the traditional diffusion process (DDPM), while the images in the second row are generated by our multi-stage font generative process. From this figure, we can clearly observe a font transfer process from the source font to the target font, and the font refinement stage slightly improves style consistency by modifying local details." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 507, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 507, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 507, + 545, + 579 + ], + "type": "text", + "content": "To further evaluate the font refinement stage, we also visualize the predicted " + }, + { + "bbox": [ + 304, + 507, + 545, + 579 + ], + "type": "inline_equation", + "content": "z_0^g" + }, + { + "bbox": [ + 304, + 507, + 545, + 579 + ], + "type": "text", + "content": " from different timesteps in the font refinement stage. As shown in Fig. 4, from " + }, + { + "bbox": [ + 304, + 507, + 545, + 579 + ], + "type": "inline_equation", + "content": "t = 200" + }, + { + "bbox": [ + 304, + 507, + 545, + 579 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 507, + 545, + 579 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 304, + 507, + 545, + 579 + ], + "type": "text", + "content": ", the font refinement stage can gradually enhance local details by modifying anomalous pixels, recovering unwanted structures, and improving visual appearances." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 586, + 545, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 586, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 305, + 586, + 545, + 600 + ], + "type": "text", + "content": "4.4. Comparison with the State-of-the-art Methods" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "In this section, we compare our model with several state-of-the-art works, including LF-Font [27], DG-Font [45], MX-Font [28], FsFont [39], NTF [9], and Diff-Font [11]. LF-Font and MX-Font are component-related methods, which decompose characters into components and learn component-wise representations to improve local details. FsFont focuses on fine-grained local styles, which utilizes cross attention to aggregate the reference styles into target styles. DG-Font and NTF are shape transformation mod" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6898" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 82, + 72, + 511, + 198 + ], + "blocks": [ + { + "bbox": [ + 82, + 72, + 511, + 198 + ], + "lines": [ + { + "bbox": [ + 82, + 72, + 511, + 198 + ], + "spans": [ + { + "bbox": [ + 82, + 72, + 511, + 198 + ], + "type": "table", + "html": "
Content辛禁病安持理根英役破任始米保用李倍荷供唱血止在利打
LF-Font辛禁病安持理根英役破任始米保用李倍荷供唱血止在利打
DG-Font辛禁病安持理根英役破任始米保用李倍荷供唱血止在利打
MX-Font辛禁病安持理根英役破任始米保用李倍荷供唱血止在利打
FsFont辛禁病安持理根英役破任始米保用李倍荷供唱血止在利打
NTF辛禁病安持理根英役破任始米保用李倍荷供唱血止在利打
Ours辛禁病安持理根英役破任始米保用李倍荷供唱血止在利打
Ground Truth辛禁病安持理根英役破任始米保用李倍荷供唱血止在利打
", + "image_path": "e64a3ae2edfe1f303ede3df8a9d0d2652ee7ef870795a12ce8fd9b6d00eb009e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 257, + 291, + 454 + ], + "blocks": [ + { + "bbox": [ + 94, + 199, + 498, + 210 + ], + "lines": [ + { + "bbox": [ + 94, + 199, + 498, + 210 + ], + "spans": [ + { + "bbox": [ + 94, + 199, + 498, + 210 + ], + "type": "text", + "content": "Figure 5. Qualitative comparisons of our proposed model with other state-of-the-art methods on UFUC dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 46, + 220, + 288, + 252 + ], + "lines": [ + { + "bbox": [ + 46, + 220, + 288, + 252 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 288, + 252 + ], + "type": "text", + "content": "Table 2. Quantitative comparison of our proposed model with other state-of-the-art methods on few-shot font generation task. The bold number indicates the best." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 257, + 291, + 454 + ], + "lines": [ + { + "bbox": [ + 48, + 257, + 291, + 454 + ], + "spans": [ + { + "bbox": [ + 48, + 257, + 291, + 454 + ], + "type": "table", + "html": "
MethodsRMSE↓PSNR↑SSIM↑LPIPS↓
Unseen Fonts and Seen Contents
LF-Font [27]0.302510.660.64690.2787
DG-Font [45]0.283311.240.66510.1479
MX-Font [28]0.300710.750.64380.1651
FsFont [39]0.244312.560.71050.1468
NTF [9]0.265011.830.68960.1517
Diff-Font2[11]0.281110.950.68460.1557
MSD-Font0.243912.610.71100.1348
Unseen Fonts and Unseen Contents
LF-Font [27]0.291011.090.67620.2564
DG-Font [45]0.275011.540.68930.1534
MX-Font [28]0.284511.290.67890.1635
FsFont [39]0.250912.290.71850.1642
NTF [9]0.257412.110.71080.1553
MSD-Font0.247512.450.71890.1527
", + "image_path": "e16ced0aa032da68223e9b80aab451e8ec60c2caa1b9192c6b4adff742c05d64.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 465, + 287, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 465, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 46, + 465, + 287, + 525 + ], + "type": "text", + "content": "els. DG-Font models font generation as a shape deformable problem via the deformation skip connection, while NTF models font generation as a continuous transformation process. We train above methods on our dataset and use the same reference images to generate fonts in inference." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 529, + 190, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 529, + 190, + 542 + ], + "spans": [ + { + "bbox": [ + 47, + 529, + 190, + 542 + ], + "type": "text", + "content": "4.4.1 Quantitative Comparison" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 550, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 550, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 550, + 287, + 693 + ], + "type": "text", + "content": "The quantitative results are shown in Tab. 2. From the table, we find that modeling font style in fine-grained manners by multiple localized encoders (MX-Font) or cross attention mechanism (FsFont) can improve the quality of generated font images. The shape transformation approaches (DG-Font and NTF) are also promising solutions for FFG task. Our MSD-Font achieves superior performance on both UFUC and UFSC datasets, which demonstrates the effectiveness of our proposed method. Specifically, compared with the NTF on the UFUC dataset, our method improves the generation performance with 0.0099, 0.34, 0.0081, and 0.0026 in terms of RMSE, PSNR, SSIM, and LPIPS." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "text", + "content": "Diff-Font is one-shot font generation model, we randomly select one image from 8 reference samples to calculate style code." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 221, + 443, + 234 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 221, + 443, + 234 + ], + "spans": [ + { + "bbox": [ + 305, + 221, + 443, + 234 + ], + "type": "text", + "content": "4.4.2 Qualitative Comparison" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 248, + 547, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 248, + 547, + 417 + ], + "spans": [ + { + "bbox": [ + 304, + 248, + 547, + 417 + ], + "type": "text", + "content": "To further evaluate our methods, as shown in Fig. 5, we visualize the generated font images in UFUC setting. LF-Font sometimes generates font images with some broken structures. The DG-Font keeps the global structure for the target character, but the generated images may contain some unwanted local structures and artifacts. FsFont utilizes several carefully selected reference characters to generate font images, thus sometimes it will fail to produce accurate structures for target images with randomly selected reference characters. The MX-Font and NTF also contain some missing strokes and distorted local details. Our MSD-Font is able to generate high-quality font images, which achieves higher visual quality than other state-of-the-art methods in terms of style consistency and structure completeness." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 436, + 379, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 436, + 379, + 449 + ], + "spans": [ + { + "bbox": [ + 305, + 436, + 379, + 449 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 460, + 545, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 613 + ], + "type": "text", + "content": "In this paper, we propose a multi-stage font generative process for few-shot font generation task, by incorporating font transfer process into the diffusion model. The proposed font generation model includes three stages: a structure construction stage to generate global structures from the source images, a font transfer stage to transform the source font into the target font, and a font refinement stage to improve the appearances and local details for the generated font images. Based on this multi-stage generative process, we construct our font generation framework, named MSD-Font, with a dual-network approach to generate high-quality font images using a few reference samples. The promising experimental results verify the effectiveness of our proposed model." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 614, + 545, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 614, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 304, + 614, + 545, + 679 + ], + "type": "text", + "content": "Limitations: As our MSD-Font based on diffusion models, there is an increase in the inference time and model size compared with previous GAN-based FFG methods. Many recent works have focused on improving the efficiency and model compactness of diffusion models, and most of these methods can be used in our model." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 680, + 545, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 680, + 545, + 703 + ], + "spans": [ + { + "bbox": [ + 305, + 680, + 545, + 703 + ], + "type": "text", + "content": "Acknowledgements: This work is supported by the National Key R&D Program of China (NO. 2022ZD0160102)." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6899" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18208-18218, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "type": "text", + "content": "[2] Samaneh Azadi, Matthew Fisher, Vladimir G Kim, Zhaowen Wang, Eli Shechtman, and Trevor Darrell. Multi-content gan for few-shot font style transfer. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7564-7573, 2018. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 288, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 288, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 288, + 247 + ], + "type": "text", + "content": "[3] Junbum Cha, Sanghyuk Chun, Gayoung Lee, Bado Lee, Seonghyeon Kim, and Hwalsuk Lee. Few-shot compositional font generation with dual memory. In European Conference on Computer Vision, pages 735-751. Springer, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 249, + 288, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 249, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 249, + 288, + 293 + ], + "type": "text", + "content": "[4] Jie Chang, Yujun Gu, Ya Zhang, Yan-Feng Wang, and CM Innovation. Chinese handwriting imitation with hierarchical generative adversarial network. In BMVC, page 290, 2018. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 295, + 288, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 295, + 288, + 338 + ], + "spans": [ + { + "bbox": [ + 53, + 295, + 288, + 338 + ], + "type": "text", + "content": "[5] Xinyuan Chen, Yangchen Xie, Li Sun, and Yue Lu. Dgfont++: Robust deformable generative networks for unsupervised font generation. arXiv preprint arXiv:2212.14742, 2022.3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 340, + 288, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 340, + 288, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 340, + 288, + 384 + ], + "type": "text", + "content": "[6] Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. Ilvr: Conditioning method for denoising diffusion probabilistic models. arXiv preprint arXiv:2108.02938, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 386, + 288, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 386, + 288, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 386, + 288, + 441 + ], + "type": "text", + "content": "[7] Jooyoung Choi, Jungbeom Lee, Chaehun Shin, Sungwon Kim, Hyunwoo Kim, and Sungroh Yoon. Perception prioritized training of diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11472-11481, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 442, + 288, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 442, + 288, + 507 + ], + "spans": [ + { + "bbox": [ + 53, + 442, + 288, + 507 + ], + "type": "text", + "content": "[8] Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8789-8797, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 510, + 288, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 510, + 288, + 553 + ], + "spans": [ + { + "bbox": [ + 53, + 510, + 288, + 553 + ], + "type": "text", + "content": "[9] Bin Fu, Junjun He, Jianjun Wang, and Yu Qiao. Neural transformation fields for arbitrary-styled font generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22438-22447, 2023. 3, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 555, + 287, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 598 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 598 + ], + "type": "text", + "content": "[10] Yue Gao, Yuan Guo, Zhouhui Lian, Yingmin Tang, and Jianguo Xiao. Artistic glyph image synthesis via one-stage few-shot learning. ACM Transactions on Graphics (TOG), 38(6): 1-12, 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 601, + 287, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 643 + ], + "type": "text", + "content": "[11] Haibin He, Xinyuan Chen, Chaoyue Wang, Juhua Liu, Bo Du, Dacheng Tao, and Yu Qiao. Diff-font: Diffusion model for robust one-shot font generation. arXiv preprint arXiv:2212.05895, 2022. 3, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 646, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 678 + ], + "type": "text", + "content": "[12] Xiao He, Mingrui Zhu, Nannan Wang, Xinbo Gao, and Heng Yang. Few-shot font generation by learning style difference and similarity. arXiv preprint arXiv:2301.10008, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "text", + "content": "[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3, 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[14] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 545, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 170 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 170 + ], + "type": "text", + "content": "[15] Yue Jiang, Zhouhui Lian, Yingmin Tang, and Jianguo Xiao. Dcfont: an end-to-end deep chinese font generation system. In SIGGRAPH Asia 2017 Technical Briefs, pages 1-4. 2017. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 173, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 545, + 194 + ], + "type": "text", + "content": "[16] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 195, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 195, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 307, + 195, + 545, + 259 + ], + "type": "text", + "content": "[17] Yuxin Kong, Canjie Luo, Weihong Ma, Qiyuan Zhu, Shenggao Zhu, Nicholas Yuan, and Lianwen Jin. Look closer to supervise better: One-shot font generation via component-based discriminator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13482-13491, 2022. 1, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 261, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 261, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 307, + 261, + 545, + 293 + ], + "type": "text", + "content": "[18] Mingi Kwon, Jaeseok Jeong, and Youngjung Uh. Diffusion models already have a semantic latent space. arXiv preprint arXiv:2210.10960, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 294, + 545, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 294, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 307, + 294, + 545, + 337 + ], + "type": "text", + "content": "[19] Guan-Horng Liu, Arash Vahdat, De-An Huang, Evangelos A Theodorou, Weili Nie, and Anima Anandkumar. I²sb: Image-to-image schrödinger bridge. arXiv preprint arXiv:2302.05872, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 338, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 338, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 338, + 545, + 392 + ], + "type": "text", + "content": "[20] Ming-Yu Liu, Xun Huang, Arun Mallya, Tero Karras, Timo Aila, Jaakko Lehtinen, and Jan Kautz. Few-shot unsupervised image-to-image translation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10551-10560, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 394, + 545, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 394, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 307, + 394, + 545, + 447 + ], + "type": "text", + "content": "[21] Wei Liu, Fangyue Liu, Fei Ding, Qian He, and Zili Yi. Xmpfont: Self-supervised cross-modality pre-training for few-shot font generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7905-7914, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 449, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 545, + 491 + ], + "type": "text", + "content": "[22] Ziwei Luo, Fredrik K Gustafsson, Zheng Zhao, Jens Sjolund, and Thomas B Schon. Image restoration with meanreverting stochastic differential equations. arXiv preprint arXiv:2301.11699, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 493, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 493, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 307, + 493, + 545, + 547 + ], + "type": "text", + "content": "[23] Pengyuan Lyu, Xiang Bai, Cong Yao, Zhen Zhu, Tengteng Huang, and Wenyu Liu. Auto-encoder guided gan for Chinese calligraphy synthesis. In 2017 14th IAPR International Conference on Document Analysis and Recognition (ICDAR), pages 1095–1100. IEEE, 2017. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 548, + 545, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 548, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 307, + 548, + 545, + 592 + ], + "type": "text", + "content": "[24] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 593, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 593, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 593, + 545, + 635 + ], + "type": "text", + "content": "[25] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 636, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 636, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 636, + 545, + 690 + ], + "type": "text", + "content": "[26] Wei Pan, Anna Zhu, Xinyu Zhou, Brian Kenji Iwana, and Shilin Li. Few shot font generation via transferring similarity guided global style and quantization local style. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 19506-19516, 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 692, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 692, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 692, + 545, + 714 + ], + "type": "text", + "content": "[27] Song Park, Sanghyuk Chun, Junbum Cha, Bado Lee, and Hyunjung Shim. Few-shot font generation with localized" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6900" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 288, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 288, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 288, + 106 + ], + "type": "text", + "content": "style representations and factorization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2393-2402, 2021. 1, 3, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "text", + "content": "[28] Song Park, Sanghyuk Chun, Junbum Cha, Bado Lee, and Hyunjung Shim. Multiple heads are better than one: Few-shot font generation with multiple localized experts. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13900-13909, 2021. 1, 3, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 207 + ], + "type": "text", + "content": "[29] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 288, + 262 + ], + "type": "text", + "content": "[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3, 4, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 319 + ], + "type": "text", + "content": "[31] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 320, + 287, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 287, + 386 + ], + "type": "text", + "content": "[32] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 388, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 441 + ], + "type": "text", + "content": "[33] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 444, + 287, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 444, + 287, + 476 + ], + "spans": [ + { + "bbox": [ + 48, + 444, + 287, + 476 + ], + "type": "text", + "content": "[34] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 478, + 287, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 478, + 287, + 511 + ], + "spans": [ + { + "bbox": [ + 48, + 478, + 287, + 511 + ], + "type": "text", + "content": "[35] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 512, + 287, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 287, + 545 + ], + "type": "text", + "content": "[36] Yang Song and Stefano Ermon. Improved techniques for training score-based generative models. Advances in neural information processing systems, 33:12438-12448, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 546, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 546, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 546, + 287, + 590 + ], + "type": "text", + "content": "[37] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 591, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 287, + 645 + ], + "type": "text", + "content": "[38] Danyang Sun, Tongzheng Ren, Chongxuan Li, Hang Su, and Jun Zhu. Learning to write stylized Chinese characters by reading a handful of examples. In Proceedings of the 27th International Joint Conference on Artificial Intelligence, pages 920–927, 2018. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 713 + ], + "type": "text", + "content": "[39] Licheng Tang, Yiyang Cai, Jiaming Liu, Zhibin Hong, Mingming Gong, Minhu Fan, Junyu Han, Jingtuo Liu, Errui Ding, and Jingdong Wang. Few-shot font generation by learning fine-grained local styles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7895-7904, 2022. 1, 3, 7, 8" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 654 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "text", + "content": "[40] Yuchen Tian. zi2zi: Master chinese calligraphy with conditional adversarial networks. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 117 + ], + "type": "text", + "content": "[41] Yuchen Tian. Rewrite: Neural style transfer for chinese fonts. 2016. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 119, + 545, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 119, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 308, + 119, + 545, + 152 + ], + "type": "text", + "content": "[42] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 153, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 153, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 308, + 153, + 545, + 206 + ], + "type": "text", + "content": "[43] Chi Wang, Min Zhou, Tiezheng Ge, Yuning Jiang, Hujun Bao, and Weiwei Xu. Cf-font: Content fusion for few-shot font generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1858-1867, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 209, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 251 + ], + "type": "text", + "content": "[44] Tengfei Wang, Ting Zhang, Bo Zhang, Hao Ouyang, Dong Chen, Qifeng Chen, and Fang Wen. Pretraining is all you need for image-to-image translation. arXiv preprint arXiv:2205.12952, 2022.3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 254, + 545, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 254, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 308, + 254, + 545, + 307 + ], + "type": "text", + "content": "[45] Yangchen Xie, Xinyuan Chen, Li Sun, and Yue Lu. Dgfont: Deformable generative networks for unsupervised font generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5130-5140, 2021. 3, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 309, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 309, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 308, + 309, + 545, + 364 + ], + "type": "text", + "content": "[46] Zhenhua Yang, Dezhi Peng, Yuxin Kong, Yuyi Zhang, Cong Yao, and Lianwen Jin. Fontdiffuser: One-shot font generation via denoising diffusion with multi-scale content aggregation and style contrastive learning. In Proceedings of the AAAI conference on artificial intelligence, 2024." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 365, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 545, + 409 + ], + "type": "text", + "content": "[47] Mingshuai Yao, Yabo Zhang, Xianhui Lin, Xiaoming Li, and Wangmeng Zuo. Vq-font: Few-shot font generation with structure-aware enhancement and quantization. arXiv preprint arXiv:2308.14018, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 411, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 464 + ], + "type": "text", + "content": "[48] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 466, + 545, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 466, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 308, + 466, + 545, + 509 + ], + "type": "text", + "content": "[49] Yexun Zhang, Ya Zhang, and Wenbin Cai. Separating style and content for generalized style transfer. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8447-8455, 2018. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 511, + 545, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 511, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 308, + 511, + 545, + 564 + ], + "type": "text", + "content": "[50] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 567, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 567, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 308, + 567, + 545, + 609 + ], + "type": "text", + "content": "[51] Mingrui Zhu, Xiao He, Nannan Wang, Xiaoyu Wang, and Xinbo Gao. All-to-key attention for arbitrary style transfer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23109-23119, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 611, + 545, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 611, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 308, + 611, + 545, + 654 + ], + "type": "text", + "content": "[52] Jianyi Wang Zongsheng Yue and Chen Change Loy. Reshift: Efficient diffusion model for image superresolution by residual shifting. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "6901" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Generate Subgoal Images before Act_ Unlocking the Chain-of-Thought Reasoning in Diffusion Model for Robot Manipulation with Multimodal Prompts/d0cc6546-2994-4121-92a3-5e7ff79756f6_content_list.json b/2024/Generate Subgoal Images before Act_ Unlocking the Chain-of-Thought Reasoning in Diffusion Model for Robot Manipulation with Multimodal Prompts/d0cc6546-2994-4121-92a3-5e7ff79756f6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..520a25fc0b59f36fe1c58a60ee88a2009aa20574 --- /dev/null +++ b/2024/Generate Subgoal Images before Act_ Unlocking the Chain-of-Thought Reasoning in Diffusion Model for Robot Manipulation with Multimodal Prompts/d0cc6546-2994-4121-92a3-5e7ff79756f6_content_list.json @@ -0,0 +1,1171 @@ +[ + { + "type": "text", + "text": "Generate Subgoal Images before Act: Unlocking the Chain-of-Thought Reasoning in Diffusion Model for Robot Manipulation with Multimodal Prompts", + "text_level": 1, + "bbox": [ + 78, + 130, + 893, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fei Ni $^{1}$ Jianye Hao $^{1*}$ Shiguang Wu $^{2}$ Longxin Kou $^{1}$ Jiashun Liu $^{1}$ Yan Zheng $^{1}$ Bin Wang $^{2}$ Yuzheng Zhuang $^{2}$ Tianjin University, China ${ }^{2} \\mathrm{Huawei}$ Noah's Ark Lab, China", + "bbox": [ + 135, + 200, + 826, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Robotics agents often struggle to understand and follow the multi-modal prompts in complex manipulation scenes which are challenging to be sufficiently and accurately described by text alone. Moreover, for long-horizon manipulation tasks, the deviation from general instruction tends to accumulate if lack of intermediate guidance from high-level subgoals. For this, we consider can we generate subgoal images before act to enhance the instruction following in long-horizon manipulation with multi-modal prompts? Inspired by the great success of diffusion model in image generation tasks, we propose a novel hierarchical framework named as CoTDiffusion that incorporates diffusion model as a high-level planner to convert the general and multimodal prompts into coherent visual subgoal plans, which further guide the low-level policy model before action execution. We design a semantic alignment module that can anchor the progress of generated keyframes along a coherent generation chain, unlocking the chain-of-thought reasoning ability of diffusion model. Additionally, we propose bi-directional generation and frame concat mechanism to further enhance the fidelity of generated subgoal images and the accuracy of instruction following. The experiments cover various robotics manipulation scenarios including visual reasoning, visual rearrange, and visual constraints. CoTDiffusion achieves outstanding performance gain compared to the baselines without explicit subgoal generation, which proves that a subgoal image is worth a thousand words of instruction. The details and visualizations are available at https://cotdiffusion.github.io.", + "bbox": [ + 73, + 323, + 473, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 790, + 209, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Embodied manipulation focuses on creating generalists that can perceive, reason, and act within complex environments, which sits at the intersection of robotics control, computer vision, and natural language processing [11]. Recent ad", + "bbox": [ + 75, + 815, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/80efe11818f03782148d8dabaaef5b81279ec6b189539900031851b55d44a916.jpg", + "image_caption": [ + "Figure 1. A motivation example of robotics manipulation tasks in multi-modal instructions. The subgoal images are worth a thousand words, inspiring us to propose a novel framework CoTDiffusion to generate goal images step-by-step before act." + ], + "image_footnote": [], + "bbox": [ + 501, + 290, + 890, + 513 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "vances in vision-language model (VLM) such as CLIP [38], BLIP [22], enable open-vocabulary visual recognition and show promising potential of robotics closed-loop control by endowing them with the ability to predict robot actions in embodied robotics [5, 10, 18, 51]. However, robotics agents still face significant challenges in following instructions for long-horizon manipulation tasks, especially when the given general instructions are not progressive step-wise prompts, but implicitly contain several subtasks to accomplish. Moreover, complex manipulation scenarios with rich visual contexts are often challenging to be sufficiently and accurately described through text-only prompts, requiring multi-modal prompts to convey the instructions and intentions accurately [21], further increasing the difficulty of instruction following and task completion.", + "bbox": [ + 496, + 580, + 892, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "For example in Fig. 1, now given a short multi-modal prompt, the robot arm needs to follow the instructions and accomplish the task. The multi-modal instruction is brief but implicitly encapsulates long-horizon multi-step manipulation steps. We have two pathways to solve this task according to the recent embodied manipulation works. The", + "bbox": [ + 496, + 809, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 236, + 0, + 810, + 18 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 323, + 16, + 722, + 30 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 292, + 31, + 754, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author: jianye.hao@tju.edu.cn", + "bbox": [ + 94, + 886, + 346, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13991", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "first paradigm is directly leveraging the powerful VLM to encode prompts and observations into tokens and predict the entire action sequence with the subsequent decoder in an end-to-end manner, like RT-2 [51] or VIMA [21]. However, the compounding small errors over long horizons will lead to catastrophic deviations from the original task instructions due to the lack of intermediate guidance from coherent subgoals. The second paradigm is to incorporate the large language models (LLMs) to explicitly decompose the general prompts into several orderly subgoals with step-wise text instructions, like SayCan [1]. Though LLMs like ChatGPT [37] struggle to support the multi-modal prompts beyond text-only modality, we can seek help from the latest multi-modal LLMs such as LLaVA [25], EmbodiedGPT [33]. Even if we idealistically assume (M)LLM-based decomposition produces well-grounded and rational plans, generated text-only prompts tend to be extremely redundant and complex to accurately specify each subgoal in complex manipulation scene, may even be more confusing than the original general prompt and impose a heavy burden on the robot agents to re-parse and re-understand again.", + "bbox": [ + 76, + 90, + 472, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Considering the phrase 'a picture is worth a thousand words', subgoal images could provide higher expressive capabilities for conveying subtasks compared to sub-prompts with complex text. This motivates us to generate subgoal images step-by-step before act to enhance long-horizon instruction following. Fortunately, tremendous success in text-to-image generation sheds light on the possibility of directly translating multi-modal prompts into coherent grounded visual subgoals step by step. Our key insight is to propose a hierarchical framework CoTDiffusion that integrates diffusion model as the high-level visual planner that can understand multi-modal prompts and progressively generate chained subgoal images based on the current visual observations in a chain-of-thought manner. The chained subgoal images serve as visual milestones to anchor the execution of task, further fed into a low-level foundation model for the specific step-wise action planning to achieve the provided goal scene. The subgoal images act as a unified interface bridging the high-level visual planning and low-level action planning, decoupling the instruction understanding and action execution. The foundation module allows the high-level diffusion model to focus solely on instruction understanding and subgoal visualization without overwhelming joint action prediction training. Conversely, the high-level visual planner frees the low-level foundation model from ambiguous prompt understanding and long-horizon planning. By providing coherent subgoal images as visual landmarks, the requirements for the foundation model are reduced to basic single-object manipulation primitives.", + "bbox": [ + 76, + 412, + 472, + 851 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The key challenge to enabling CoTDiffusion to progressively generate subgoal images in a chain-of-thought manner lies in tracking the generated subgoal's progress on task", + "bbox": [ + 76, + 854, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "prompts. In other words, for coherent subgoal generation, diffusion model needs to know 'which step this subgoal image has reached' and 'which step needs to reach in the next subgoal image' based on task prompts. For this, we design a triple alignment module comparing the cross-frame visual contrasts between the generated subgoal image and initial observation, and align the progress back to prompts to anchor the stage of generation chains. Specifically, we use a coarse-to-fine training pipeline. First, the aligned module is coarsely pre-trained to predict residual mask patches between subgoal images for aligning spatial semantics, focusing on salient differences rather than pixel details in textures or colors. The semantic alignment module is then integrated into the diffusion model for step-wise image generation with fine-grained pixel reconstruction. Additionally, bi-directional generation and frame concatenation mechanism further enhance subgoal image fidelity and instruction following. The contributions of this work are as follows:", + "bbox": [ + 496, + 90, + 893, + 364 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a hierarchical framework CoTDiffusion that the high-level diffusion model translates the multi-modal prompts into coherent subgoal images in a chain-of-thought manner as visual milestones to anchor the low-level foundation model executing, enhance the instruction following on long-horizon manipulation tasks.", + "- We design a semantic alignment module to capture the semantic relevance between visual subgoals and prompt, progressively tracking the progress of generated images along the coherent generation chain to unlock the chain-of-thought reasoning capability for multi-modal prompt.", + "- The experiments on various long-horizon manipulation tasks empirically demonstrate that CoTDiffusion enjoys outstanding performance gain than prior methods by explicitly generating goal images step by step before act." + ], + "bbox": [ + 500, + 366, + 890, + 594 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 609, + 640, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Diffusion Models for Text-to-Image Generation", + "text_level": 1, + "bbox": [ + 500, + 633, + 890, + 650 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, diffusion models have emerged as a powerful paradigm for high-fidelity text-to-image synthesis [9, 17, 35, 42, 45]. Models such as DALL-E 2 [39], Imagen [44] and Stable Diffusion [43] have demonstrated impressive success in generating realistic images from textual descriptions. The text instructions are tokenized as specific conditions and injected into the denoising process for controllable image generation. Moreover, some recent works have begun to explore image generation with multi-modal prompts, incorporating diverse inputs beyond just text. UniDiffuser [3] treats both image and text as sequential token streams for diffusion generations and Versatile diffusion [49] employs a multi-flow design to tackle multi-modal prompts. However, directly applying existing text-to-image approaches to complex robotics manipulation poses challenges in instruction understanding and following. Prompts", + "bbox": [ + 496, + 657, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13992", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "in manipulation tasks are often too high-level and general to effectively ground and reason, making the generated images struggle to accurately follow instructions and ground to the given scenarios, not sufficient to serve as goal images to offer fine-grained guidance for manipulation. Our work employs a semantic alignment module for multi-stage cross-modal alignment to achieve better grounding of generated images and unlock the chain-of-thought reasoning abilities of diffusion model for long-horizon tasks.", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Robotics Manipulation and Control", + "text_level": 1, + "bbox": [ + 76, + 243, + 387, + 258 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With the development of VLMs for robotics, such as R3M [34], VIP [29], and LIV [30] pre-trained general visual representations for robotic perception, RT-2 [51], Gato [41] and PaLM-E [10] can convert instruction and observation into tokens and apply large transformers to predict action tokens in an end-to-end manner for robot manipulation tasks. While these works mostly focus on language-conditioned manipulation, more complex manipulation scenarios with multi-modal prompts introduced by VIMA [21] are also gaining increasing attention. Accurately instruction following becomes more challenging with general multimodal prompts, especially in long-horizon manipulation tasks. It is not trivial to directly borrow the recent works like SayCan [1] and EmbodiedGPT [33] to utilize MLLMs to decompose multi-modal instructions into step-wise text prompts. Even if we assume step-wise instructions can be rationally decomposed, they often require complex and redundant language to accurately specify subtasks and may be more confusing to understand. Inspired by the motto 'show, don't tell', our work aims to translate the general multi-modal prompts into coherent subgoal images instead of text step by step, serving as visual landmarks to enhance the instruction following in long horizon.", + "bbox": [ + 75, + 268, + 472, + 617 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Chain-of-Thought Reasoning", + "text_level": 1, + "bbox": [ + 76, + 633, + 338, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Chain-of-Thought [48] reasoning refers to the general strategy of solving multi-step problems by decomposing them into a sequence of intermediate steps, showing great success in guiding the model to think step-by-step for enhancing instructing following and reasoning accuracy. It has recently been applied extensively in a variety of problems such as mathematical reasoning [8, 24], program execution [36, 40], commonsense or general reasoning [23, 48]. However, chain-of-thought visual reasoning has not been well explored for robotics manipulation. Inspired by prior explorations into CoT reasoning, we introduce the novel concept of chaining conditional image generations step by step to guide long-horizon manipulation. Our work employs a semantic alignment module to anchor the generated subgoal image in the entire reasoning chain, unlocking the chain-of-thought reasoning capabilities for the diffusion model.", + "bbox": [ + 75, + 659, + 472, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 89, + 589, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose CoTDiffusion, a hierarchical framework that integrates the diffusion model as the high-level module to decompose multi-modal prompts in a chain-of-thought manner and progressively generate chained subgoals images step by step to guide the underlying foundation model for long-horizon tasks. Specifically, we first discuss the overview pipeline of CoTDiffusion in Sec. 3.1, and then introduce the pretraining of the coarse semantic alignment to capture the step-wise spatial information in Sec. 3.2. Then we discuss how to further fine-grain the CoT reasoning and grounded generation capabilities of the diffusion model in Sec. 3.3 and introduce the details of goal image conditioned foundation model for action planning in Sec. 3.4.", + "bbox": [ + 496, + 114, + 893, + 313 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Pipeline Overview", + "text_level": 1, + "bbox": [ + 498, + 325, + 674, + 342 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall pipeline of CoTDiffusion is presented in Fig. 2, decoupling the visual planning and action planning. Given the initial observation $x_0$ and a multi-modal prompt $\\mathcal{P}$ as task instruction potentially needs to be reached by $N$ subgoal steps, robots are required to learn a policy conditioned on the prompt and accomplish the specified long-horizon task. Our model is capable of understanding these general instructions and leverages diffusion model implicitly to decompose them into subgoal images chain $\\tau_x = \\{x_i\\}_{i=1}^N$ . The precisely generated goal images $x_i$ then guide the foundation model's planning for inferring the action sequences $\\tau_a^i = \\{a_{i,t}\\}_{t=1}^T$ with the subgoal horizon length $T$ . Let $p_\\Theta$ model this hierarchical decision-making framework. With bi-level modules, $p_\\Theta$ can be factorized into visual planning $p_\\phi$ , and action planning $p_\\psi$ . Under the Markovian assumption, the overall framework can be formulated as:", + "bbox": [ + 496, + 349, + 893, + 592 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\Theta} \\left(\\left\\{\\tau_ {a} ^ {i} \\right\\} _ {i = 1} ^ {N} \\mid \\mathcal {P}, x _ {0}\\right) = \\underbrace {\\left(\\prod_ {i = 1} ^ {N} p _ {\\phi} \\left(x _ {i} \\mid \\mathcal {P} , x _ {0}\\right)\\right)} _ {\\text {v i s u a l p l a n n i n g}} \\underbrace {\\left(\\prod_ {i = 1} ^ {N} \\prod_ {t = 1} ^ {T} p _ {\\psi} \\left(a _ {i , t} \\mid x _ {i}\\right)\\right)} _ {\\text {a c t i o n p l a n n i n g}} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 498, + 604, + 906, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The visual planning module consists of a multi-modal encoder to understand the complex task instructions and conditioned diffusion model which implicitly decomposes the task into progressive chained subgoals images. The standardized subgoal images serve as a unified interface bridging the visual planning and action planning modules. With the decoupled foundation model for action prediction, high-level visual planning focuses solely on comprehending instructions and visualizing subgoals without confusion from joint action prediction training. Conversely, the visual planning module frees the action planning module from ambiguous prompt parsing and long-horizon planning by providing coherent keyframes as visual landmarks, further reducing the capabilities requirements of the foundation model to basic single object manipulation skills.", + "bbox": [ + 496, + 674, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13993", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b1d80ffbef7691a905dac1c6b133e9089580b366c4dc70b16ac5f1359107611b.jpg", + "image_caption": [ + "Figure 2. Method overview: CoTDiffusion consists of a multi-modal encoder and vision encoder $\\mathcal{V}$ , semantic alignment module $\\mathcal{S}$ , conditional diffusion model $\\mathcal{E}$ and foundation model $\\mathcal{F}$ for action planning. The prompt and observation tokens are combined and fed into the semantic alignment module to identify the current reasoning chain step, providing progressive guidance for the diffusion model to generate the next subgoal image. The generated keyframes are further fed to the foundation model which predicts action sequences to achieve the imagined goal scene and this recursive process repeats in a receding horizon control loop until the task is finished." + ], + "image_footnote": [], + "bbox": [ + 101, + 85, + 869, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Pre-training Coarse Semantic Alignment", + "text_level": 1, + "bbox": [ + 76, + 385, + 426, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the challenge of explicit decomposing step-wise sub-prompts from multi-modal prompts, the diffusion model struggles to generate sequentially coherent keyframes that incrementally advance the prompt instructions and execution progress. The diffusion model fails to identify the stage of current generated keyframes onto the whole chain-of-thought progress and the generated sub-goals images may exhibit repetitions, skipping, or even backtracking to visited subgoals. To alleviate this dilemma, we design a semantic alignment module to empower diffusion model to track the progress of generated keyframes into the entire chain.", + "bbox": [ + 75, + 409, + 470, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Triple Alignment Architecture Since explicit decomposition for the multi-modal prompt is intractable, the initial observation $x_0$ and prompt $\\mathcal{P}$ remain fixed across the progressive generation of diffusion model. Thus, tracking the progress critically relies on extracting semantic information from the generated subgoal $x_i$ . To this end, we design a triple alignment module denoted as $S(x_0, x_i, \\mathcal{P})$ to compare current generated subgoal $x_i$ against initial observation $x_0$ and prompts $\\mathcal{P}$ , serving as an information bottleneck to capture the key semantic cues about the subgoal-prompt matching to locate the current chain stage and deliver to the diffusion model for next subgoal generation. Specifically, we first capture visual representation from both the initial and current observation using a shared vision encoder $\\mathcal{V}$ and concatenate with prompt tokens $\\mathcal{P}$ . Then they are refined through fusion module which consists of several self-attention blocks separately to obtain attention tokens $z_0$ and $z_i$ aligned to the prompts. The cross-attention between $z_0$ and $z_i$ aggregates a contrast representation $z_{\\mathrm{contrast}}^i$ and then concatenated with prompt tokens into another fusion module to capture the progress tokens $z_{\\mathrm{progress}}^i$ , including", + "bbox": [ + 75, + 583, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "rich context about the advancement and completion of the prompted task between the two subgoals. Finally, $z_{\\text{progress}}^i$ and the current visual tokens $x_i$ and then fed into another cross-attention pass, which visual tokens $x_i$ serve as keys and progress tokens $z_{\\text{progress}}^i$ as queries to infer the final aligned tokens $z_{\\text{align}}^i$ . The aligned token encapsulates cues about 'which subgoal has reached now' and 'which subgoal should reach next', further injected into the diffusion model as precise semantic guidance to steer the generation of next subgoal. Through multi-stage cross-frame attention mechanisms, the triple alignment module can capture the cross-modal semantic correlation and achieve dynamic grounding of generated subgoal image into the multi-modal prompts to identify progress along the CoT reasoning chain.", + "bbox": [ + 496, + 386, + 893, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Masked Patch Prediction It is not trivial to directly integrate the designed triple alignment module without pretraining into the diffusion model for joint training from scratch. Simultaneously optimizing semantic alignment and pixel-level reconstruction such as texture, color, shape, and spatial transformations overwhelms joint learning. Thus, we propose a two-stage coarse-to-fine approach decoupling semantic alignment pretraining from diffusion model finetuning, illustrated in Fig. 3. First, we pre-train coarse semantic alignment focused only on spatial correlations, without pixel generation. Considering the adjacent keyframe contrasts reveal object manipulations as mask residuals patch $\\hat{m}_i = x_{i + 1} - x_i$ , we can utilize these cross-frame spatial semantics to provide coarse alignment supervision. Regions with no visual changes likely correspond to stationary background elements or objects not involved in the current subtask operation. Contrastively, masked areas indicate spatial layout modifications concluding the rich spatial semantic information between keyframes. The sequence", + "bbox": [ + 496, + 613, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13994", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/01d450c565851e62fe9c738834e999cbcd32251a5114e7b78c416b654f995163.jpg", + "image_caption": [ + "Figure 3. Two phase of coarse-to-fine alignment pipeline." + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 457, + 223 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of mask residuals patch $\\{\\hat{m}_i\\}_{i=1}^{N-1}$ can be viewed as implicitly decomposing the prompt into spatial manipulation steps across visual frames that can be used for alignment. So we fed the aligned tokens into a mask predictor $\\mathcal{D}$ to decode the mask region and set modeling mask prediction as a pretext task. Mask prediction encourages the model to infer these salient spatial semantics and changes to incrementally track prompt completion and focus regions. The process can be formulated as: $\\{m_i\\}_{i=1}^N = \\sum_{i=1}^N \\mathcal{D}(m_i | z_{\\mathrm{align}}^i) S(z_{\\mathrm{align}}^i | x_0, x_i, \\mathcal{P})$ .", + "bbox": [ + 75, + 258, + 468, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By factorizing the alignment and generation objectives, the alignment module can develop basic visual reasoning skills through patch-level mask prediction before integrated into the diffusion model for controllable image synthesis.", + "bbox": [ + 75, + 412, + 468, + 473 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Fine-grained Diffusion Training", + "text_level": 1, + "bbox": [ + 76, + 482, + 359, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the second stage, we integrate the pretrained coarse alignment module into the diffusion model training for fine-grained image generation. To further enhance instruction following, we design a bi-directional generation that reconstructs the current frame from the aligned token, providing an additional learning constraint. We also employ a frame concatenation mechanism to enhance the consistency across generated keyframes and the fidelity of generation.", + "bbox": [ + 75, + 506, + 468, + 627 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Bi-directional Alignment Generation Building upon coarse pretraining, we integrate the alignment module into the diffusion model for fine-grained training with the pixel-level supervision of ground truth keyframes, illustrated in Fig. 3. Here we propose bi-directional aligned generation, where the aligned token $z_{\\mathrm{align}}^i$ not only guides forward prediction but also reconstructs the current frame through backward passing. In the forward pass, based on current subgoal $x_i$ , $z_{\\mathrm{align}}^i$ provides semantic conditioning to generate the next subgoal image $x_{i+1}$ . In the backward pass, the same aligned token can guide diffusion model to reconstruct the current subgoal $x_i$ from the next subgoal image $x_{i+1}$ , acting as an additional constraint. The training objective can be formulated as:", + "bbox": [ + 75, + 633, + 468, + 844 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL = \\mathbb {E} _ {x _ {i} \\in D} [ \\underbrace {\\| \\hat {x} _ {i} - \\mathcal {E} (x _ {i - 1} , z _ {\\text {a l i g n}} ^ {i} , \\mathcal {P})} _ {\\text {F o r w a r d G e n e r a t i o n}} \\| + \\underbrace {\\| \\hat {x} _ {i - 1} - \\mathcal {E} (x _ {i} , z _ {\\text {a l i g n}} ^ {i} , \\mathcal {P})} _ {\\text {B a c k w a r d G e n e r a t i o n}} \\| ] \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 849, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Bi-directional generation compels tighter instruction grounding, as accuracy is required for both progressive and reverse keyframe prediction. This dual-direction linkage enhances visual coherence and sequence controllability compared to forward-only training, improving the chain-of-thought capabilities of diffusion model.", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Frame Concat Mechanism Generating consistent and coherent subgoal images requires considering both semantic guidance to match prompts and visual grounding into observations. The aligned tokens $z_{\\mathrm{align}}^i$ and image tokens $x_i$ are both fused into the denoising process in a classifier-free manner, as shown in Eq. (2). However, merely injecting current observation tokens as condition into noise model struggles to provide sufficient guidance to prevent distortions conflicting with the current observation, which is critical for precisely grounded robotics manipulations. Inspired by VDT[28], we employ frame concat mechanism to directly take current observation frame as the component of input rather than only condition for the next subgoal image generation, illustrated in Fig. 2. The current observation frame concatenated with the standard Gaussian noise gives a strong visual continuity prior, then fed into diffusion model to denoise altogether. For the training, we split the corresponding frame from the denoising frame and supervised it to the ground truth subgoal image. The rich visual context from the concatenated current frame enables coherent denoising in diffusion model further guarantees consistency with grounded observation and enforces coherence and smoothness across chain-of-thought visual planning.", + "bbox": [ + 496, + 195, + 892, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Goal-conditioned Policy Model", + "text_level": 1, + "bbox": [ + 500, + 556, + 772, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The final component in our framework is the low-level policy model for action planning, generating an action trajectory $\\tau_{a}^{i}$ when given observation trajectory $\\tau_{x}^{i}$ from visual planning. The policy model can be parameterized as an image-conditioned planner that infers the action $a_{i,t}$ given the current observation $x_{i,t}$ and the generated subgoal image $g_{i}$ : $\\tau_{a}^{i} = \\{a_{i,t}\\}_{t=1}^{T} \\sim \\prod_{t=1}^{T} p_{\\psi}(a_{i,t}|x_{i,t}, g_{i})$ . The policy model is trained by imitation learning in an end-to-end manner, using paired subgoal images and expert trajectories. Thanks to the explicit subgoal generation from high-level visual planner, the low-level policy model does not need to master complex multi-step manipulation skills in long-horizon. Therefore, we simplify the policy model architecture and implement it as a simple MLP to only possess basic single-object manipulation primitives, which greatly reduces the burden of policy model training. Although we do not seek improvement through a more powerful policy model in this paper, CoTDiffusion is flexible to incorporate any goal-conditioned design like transformers or diffusion policies. For more details of low-level policy model, please refer to Appendix E.3.", + "bbox": [ + 496, + 580, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13995", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/611df21c38c8862e3cd76d06c27075011ff9873800b5b0b7a14ff0d38ec12229.jpg", + "image_caption": [ + "Figure 4. Visualization of CoTDiffusion in three typical long-horizon tasks with multi-modal prompts in VIMA-BENCH." + ], + "image_footnote": [], + "bbox": [ + 99, + 85, + 872, + 329 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 357, + 209, + 373 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 76, + 382, + 253, + 398 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Benchmark & Tasks We conduct evaluation on VIMABENCH, a benchmark suite for multimodal robot learning, which is built on the Ravens robot simulator [50]. VIMABENCH supports extensible collections of objects and textures to compose multi-modal prompts and to procedurally generate a large number of tasks. VIMA-BENCH can generate large quantities of expert trajectories via scripted oracle agents, and the ground truth keyframes that can leveraged for the training of CoTDiffusion. VIMA-BENCH contains various tasks ranging from simple object manipulation to multi-object manipulation and we select three representative long-horizon manipulation tasks in VIMABENCH-visual rearrangement, visual constraints, and visual reasoning, which represent different levels of complexity and requirements for embodied manipulation. These tasks represent different levels of complexity and requirements for embodied manipulation and can test the ability to understand spatial relationships and perform precise object manipulations. For more details regarding the simulation benchmark and tasks setting, please refer to Appendix A.", + "bbox": [ + 75, + 406, + 473, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metric The evaluation metrics we use in the experiments cover three aspects as following:", + "bbox": [ + 75, + 720, + 468, + 750 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Image Fidelity - measures the quality and realism of the generated keyframe images. We choose Fréchet Inception Distance (FID) [15] as a useful metric.", + "- Instruction Following - evaluates how well the step-by-step keyframes cover the complex instructions. We use CLIP-based text-image similarity scores [14] between the prompt and keyframes to quantify instruction alignment.", + "- Task Completion - tests the end-to-end utility by executing the embodied manipulation task. The final success rate on long-horizon tasks can be a fair metric." + ], + "bbox": [ + 75, + 750, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Baselines", + "text_level": 1, + "bbox": [ + 500, + 357, + 607, + 372 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "There are several existing methods for constructing robot manipulation policies conditioned on complex prompts, which we use as baselines in our experiments:", + "bbox": [ + 498, + 382, + 890, + 426 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Gato[41], a generalist agent that can solve tasks from multiple domains where tasks are specified by prompting the model with the observation and action subsequence. For fair comparisons, we provide the same conditioning manner with multi-modal prompts.", + "- Flamingo[2], a vision-language model that learns to generate textual completion in response to multimodal prompts. We borrow the architecture in [21] which adapts it to support decision-masking by replacing the output layer with robot action heads.", + "- VIMA[21], the first transformer-based generalist robot agent that processes manipulation tasks with multimodal prompts. VIMA adpots an object-centric approach to flatten all the observation and prompts into object tokens sequence and predicts motor actions autoregressively and demonstrates SOTA performance on VIMA-BENCH.", + "- SuSIE[4], the concurrent work for visual planning in manipulation tasks on the CALVIN [31], incorporating pretrained image-editing model and the low-level goal-conditioned policy. For fair comparison, we finetune the model with the same oracle data with CoTDiffusion." + ], + "bbox": [ + 500, + 428, + 890, + 743 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Quantitative Results of Success Rate", + "text_level": 1, + "bbox": [ + 500, + 755, + 815, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We begin by comparing the performance of CoTDiffusion and baselines to solve long-horizon tasks in visual rearrange, visual reasoning, and visual constraints. The performance comparisons shown in Tab. 1 demonstrate CoT-Diffusion significantly outperforms other baselines in success rate. The baselines can be divided into two kinds of planners, including abstract planner and visual planner. Abstract planner like Gato, Flamingo and VIMA directly", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13996", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "map general prompts to subsequent actions in an end-to-end manner. Gato and Flamingo gets low success rates on long-horizon tasks without explicit subgoal generation to correct the accumulative deviation errors from the instructions. In contrast, Visual planner like SuSIE and CoTDiffusion can generate intermediate goal images to guide the action planning, which can enhance instruction following for long-horizon tasks via visual planning. However, SuSIE lacks intrinsic chain-of-thought reasoning capabilities, struggling to generate logically progressing subgoal sequences as coherent plans from general prompts. To enable comparison, we grant SuSIE the manually decomposed prompts into privileged step-wise sub-prompts which are unavailable in fair settings, denoted as 'SuSIE + sub-prompts'. Nonetheless, its performance still lags CoTDiffusion, which can implicitly align generated images with multi-modal prompts through learned correspondence between visual signals and language semantics. One more potential reason to restrict the performance of SuSIE is that provided sub-prompts are challenging to perfectly encapsulate complex instructions. In contrast, CoTDiffusion develops intrinsic chain-of-thought reasoning and alignment for generated subgoal images for flexible visual planning directly from the raw multi-modal prompts, without the need for pre-decomposed sub-prompts. The comparisons highlight the importance of chain-of-thought capabilities in visual planning, especially the role of implicit semantic alignment for instruction following long-horizon manipulation tasks.", + "bbox": [ + 76, + 90, + 472, + 513 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f7024d7639341b2f27681311e6db79ea6a0c66e29f0540826e7b30261ec497b8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodologyRearrangeReasoningConstraintsOverall
Gato6.4 ± 1.32.5 ± 0.425.2 ± 3.111.4 ± 1.6
Flamingo17.5 ± 1.63.0 ± 0.536.1 ± 4.218.9 ± 2.1
VIMA43.1 ± 3.338.2 ± 4.467.2 ± 5.249.5 ± 4.3
SuSIE2.7 ± 0.33.1 ± 0.624.3 ± 5.910.0 ± 2.3
+sub-prompts37.7 ± 6.239.0 ± 4.552.3 ± 7.043.0 ± 5.9
CoTDiffusion59.0 ± 1.751.7 ± 2.683.1 ± 4.764.6 ± 3.0
", + "bbox": [ + 78, + 518, + 467, + 618 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Further Analysis", + "text_level": 1, + "bbox": [ + 76, + 665, + 243, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Robustness to Insufficient Perception Rich visual observations from diverse views are crucial for complex robot manipulation tasks. Restricted perspectives limit rich representations used for action planning, degrading performance on long-horizon multi-object manipulation. We evaluate different methods in single-view including the top and front views provided from VIMA-BENCH. As shown in Tab. 2, visual planning approach demonstrates greater robustness to limited observations, with a smaller performance drop in single-view compared to abstract methods. We attribute this to two potential reasons: First, accurate and grounded subgoal images generated in visual planners provide supplemental visual context, which can partly compensate for the insufficient perception to aid robustness under single-", + "bbox": [ + 75, + 688, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ba4a0e539b809ac93ed103fc3f9204ff3bd92eb99e3e38a35b56a6deffb9a275.jpg", + "table_caption": [ + "Table 1. The evaluations of success rates on three typical long-horizon tasks with multi-modal prompts." + ], + "table_footnote": [], + "table_body": "
MethodologyMulti-ViewSingle-ViewPerformance Drop
Gato11.4 ± 1.66.5 ± 1.64.9(42% ↓)
Flamingo18.9 ± 2.112.0 ± 2.46.9(36.2% ↓)
VIMA49.5 ± 4.334.9 ± 3.414.6(29.4% ↓)
SuSIE10.0 ± 2.37.9 ± 2.02.1(21.0% ↓)
+sub-prompts43.0 ± 5.935.6 ± 6.17.4(18.8% ↓)
CoTDiffusion64.6 ± 3.056.0 ± 2.48.6(13.2% ↓)
", + "bbox": [ + 504, + 88, + 888, + 195 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9535e3e4fefb47a881a7b0b39167476d51ef10790a355ee8ff43451ce0ba1114.jpg", + "table_caption": [ + "Table 2. The evaluations of performance drop of different methods on single-view and multi-view from VIMA-BENCH.", + "Table 3. Quantitative comparisons of FID between methods on all three tasks, including visual rearrange, reasoning and constraints." + ], + "table_footnote": [], + "table_body": "
MethodologyRearrangeReasoningConstraints
SuSIE [4]19.718.323.2
- w/o finetune38.4 ↑33.0 ↑52.7 ↑
- w/o stepwise prompt32.6 ↑28.1 ↑34.7 ↑
CoTDiffusion (ours)11.310.88.6
- w/o coarse pre-train23.1 ↑16.2 ↑18.9 ↑
- w/o bi-direction align18.7 ↑15.0 ↑15.8 ↑
- w/o frame concat15.4 ↑13.7 ↑12.1 ↑
", + "bbox": [ + 504, + 273, + 916, + 390 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "view. Second, by providing coherent subgoal images as visual landmarks, the requirements for the low-level action planner are reduced to basic single-object manipulation primitives in short horizon, with less reliance on rich visual perceptions. The experiments demonstrate that CoTDiffusion enjoys better robustness to restricted perception than abstract planners, highlighting the benefits of hierarchical framework decoupled visual planning and action planning.", + "bbox": [ + 496, + 398, + 890, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fidelity of Image Generation Here we conduct further analysis of visual planners by comparing goal image quality against ground truth keyframes with FID as evaluation metrics. As Tab. 3 shows, CoTDiffusion still achieves much better fidelity than SuSIE even though SuSIE has got improved after fine-tuning on the same datasets in VIMABENCH. Moreover, SuSIE can only understand short instructions of single-object manipulation tasks, struggling to handle long-horizon instructions due to the lack of chain-of-thought reasoning capabilities to handle multi-modal prompts. With step-wise sub-prompts decomposed in advance, the performance of SuSIE gets largely raised but still underperforms CoTDiffusion, which has no need to explicitly decompose the general prompts and can generate subgoal images in an implicit chain-of-thought manner. Additionally, ablating coarse pretraining and bi-directional generation degrades performance, validating their benefits. The coarse-to-fine semantic alignment training training allows developing spatial reasoning prior to synthesis. Frame concatenation further guides coherent denoising by providing rich context information as visual priors to ground the current observation and enhance the fidelity of generation.", + "bbox": [ + 496, + 527, + 892, + 861 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Accuracy of Instruction Following We evaluate instruction following accuracy via CLIP similarity between gen", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13997", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "erated keyframes and general prompts, normalized by the CLIP score between ground truth ultimate goal image and prompts. The results in Fig. 5 demonstrate that CoT-Diffusion enjoys better instruction following capabilities. Without chain-of-thought reasoning abilities, SuSIE struggles to follow instructions when given general multi-modal prompts, let alone generate subgoal images with smooth progressions. Progressively providing privileged prompts step by step improves SuSIE in instruction following and gradual advancement towards ultimate goals. However, text prompts decomposed by rules are not always perfect to convey the original multi-modal prompt, so the reliance on these prompts restricts the performance of SuSIE. In contrast, the coarse alignment pretraining and bi-directional generation can assist the diffusion model in tracking the progress of generated keyframes throughout the entire chain and generate sequenced keyframes incrementally advancing prompt instructions. Moreover, the implicit decomposition maintains semantic coherence between subgoals without overly large jumps and the smoothness and continuity between visual milestones make them more reachable by downstream foundation models with just single-object manipulation capabilities. Additionally, we observe that the bi-directional generation may impedes the diffusion model training if without coarse semantic pretraining. As the initial observation and prompt remain fixed across different generation steps, the align tokens at various stage tend to so similar that may confuse simultaneous progressive generation and reconstruction training.", + "bbox": [ + 76, + 90, + 472, + 530 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Generalization across Tasks We evaluate the generalization ability in three levels with increasing difficulty: placement generalization which randomizes the novel placement of objects (L1), object generalization which provides the objects with novel attributes (L2), and task combinatorial generalization which complexes the prompts with extra novel instruction (L3). As Fig. 6 shows, CoTDiffusion exhibits strong zero-shot generalization on unfamiliar objects, colors, and shapes at different placements by explicitly visualizing them into coherent subgoals grounded on new concepts. This avoids re-grounding concepts from multi-modal prompts in the low-level foundation model for", + "bbox": [ + 76, + 547, + 472, + 729 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2323b9c507967d8d3a6d98f7b39ab2117fbcdf9b5a73208aca9be77b68fe387b.jpg", + "image_caption": [ + "Figure 5. The normalized CLIP scores for each generation step, reflecting the step-wise accuracy of instruction following." + ], + "image_footnote": [], + "bbox": [ + 81, + 739, + 467, + 856 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b2bc2a91b3604bd403e5ea396ca0a9ddd8a3fc4f8af91b2eabfbec501460e426.jpg", + "image_caption": [ + "Figure 6. The evaluation on various generalization levels." + ], + "image_footnote": [], + "bbox": [ + 501, + 92, + 883, + 244 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the placement and object. For L3 generalization, taking an example, we complex the prompt such as 'rearrange ... then rotate/twist/stack ...' to demand extra tasks such as rotating, twisting degrees, or stacking objects. CoTDiffusion achieves outstanding gain in the zero-shot performance of combinatorial tasks. When CoTDiffusion visualizes novel concepts into goal images, the foundation model can still accomplish the stack by simply achieving the provided subgoal, with no need to inherently understand novel skills like stack. This demonstrates the power of 'a image is worth a thousand words' - the subgoal images facilitate the generalization of the foundation model to unseen objects and novel combinatorial task while the foundation model simply achieves them with reuse of simple known primitives.", + "bbox": [ + 496, + 276, + 890, + 488 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 508, + 617, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We presented CoTDiffusion, a hierarchical framework that integrates diffusion model as high-level module to translate the general multi-modal prompts into coherent subgoal images, serves as the visual milestones to anchor the low-level foundation model to plan action sequences, termed as 'generate subgoal images before act'. With the coarse-to-fine training for semantic alignment module, CoTDiffusion can identify the progress of generated subgoals images along reasoning chains, unlocking the chain-of-thought reasoning capabilities of diffusion model for long-horizon manipulation tasks. The experiments cover various long-horizon manipulation scenarios in VIMA-BENCH, and CoTDiffusion show the strong instruction following and outstanding performance gain compared to existed methods without visual planning. Incorporating commonsense knowledge from pre-trained MLLM like GPT-4V provides an avenue for more generalizable and promising reasoning in CoTDiffusion, which leaves us our future work.", + "bbox": [ + 496, + 535, + 890, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 828, + 679, + 845 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is supported by the National Natural Science Foundation of China (Grant Nos. 92370132) and the Xiaomi Young Talents Program of Xiaomi Foundation.", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13998", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, et al. Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691, 2022. 2, 3", + "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: A visual language model for few-shot learning. NeurIPS, 2022. 6, 5", + "[3] Fan Bao, Shen Nie, Kaiwen Xue, Chongxuan Li, Shi Pu, Yaole Wang, Gang Yue, Yue Cao, Hang Su, and Jun Zhu. One transformer fits all distributions in multi-modal diffusion at scale. arXiv preprint arXiv:2303.06555, 2023. 2", + "[4] Kevin Black, Mitsuhiko Nakamoto, Pranav Atreya, Homer Walke, Chelsea Finn, Aviral Kumar, and Sergey Levine. Zero-shot robotic manipulation with pretrained image-editing diffusion models. arXiv preprint arXiv:2310.10639, 2023. 6, 7, 5", + "[5] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817, 2022. 1", + "[6] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. arXiv preprint arXiv:2211.09800, 2022. 5", + "[7] Cheng Chi, Siyuan Feng, Yilun Du, Zhenjia Xu, Eric Cousineau, Benjamin Burchfiel, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. arXiv preprint arXiv:2303.04137. 8", + "[8] Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168, 2021. 3", + "[9] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In Advances in Neural Information Processing Systems, pages 8780-8794. Curran Associates, Inc., 2021. 2", + "[10] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palm- e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 1, 3", + "[11] Jiafei Duan, Samson Yu, Hui Li Tan, Hongyuan Zhu, and Cheston Tan. A survey of embodied ai: From simulators to research tasks. IEEE Transactions on Emerging Topics in Computational Intelligence, 6(2):230-244, 2022. 1", + "[12] Theophile Gervet, Zhou Xian, Nikolaos Gkanatsios, and Katerina Fragkiadaki. Act3d: Infinite resolution action detection transformer for robotic manipulation. arXiv preprint arXiv:2306.17817, 2023. 8", + "[13] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 4" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 6", + "[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6", + "[16] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 6", + "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2", + "[18] Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. Voxposer: Composable 3d value maps for robotic manipulation with language models. arXiv preprint arXiv:2307.05973, 2023. 1", + "[19] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General perception with iterative attention. In International conference on machine learning, pages 4651-4664. PMLR, 2021. 5", + "[20] Stephen James, Zicong Ma, David Rovick Arrojo, and Andrew J Davison. Rlbench: The robot learning benchmark & learning environment. IEEE Robotics and Automation Letters, 5(2):3019-3026, 2020. 8", + "[21] Yunfan Jiang, Agrim Gupta, Zichen Zhang, Guanzhi Wang, Yongqiang Dou, Yanjun Chen, Li Fei-Fei, Anima Anandkumar, Yuke Zhu, and Linxi Fan. Vima: General robot manipulation with multimodal prompts. arXiv preprint arXiv:2210.03094, 2022. 1, 2, 3, 6, 5", + "[22] Dongxu Li, Junnan Li, and Steven CH Hoi. Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. arXiv preprint arXiv:2305.14720, 2023. 1", + "[23] Zhengzhong Liang, Steven Bethard, and Mihai Surdeanu. Explainable multi-hop verbal reasoning through internal monologue. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, 2021. 3", + "[24] Wang Ling, Dani Yogatama, Chris Dyer, and Phil Blunson. Program induction by rationale generation: Learning to solve and explain algebraic word problems. arXiv preprint arXiv:1705.04146, 2017. 3", + "[25] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 2", + "[26] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6", + "[27] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 12", + "[28] Haoyu Lu, Guoxing Yang, Nanyi Fei, Yuqi Huo, Zhiwu Lu, Ping Luo, and Mingyu Ding. Vdt: An empirical study on video diffusion with transformers. arXiv preprint arXiv:2305.13311, 2023. 5" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13999", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Yecheng Jason Ma, Shagun Sodhani, Dinesh Jayaraman, Obert Bastani, Vikash Kumar, and Amy Zhang. Vip: Towards universal visual reward and representation via value-implicit pre-training. arXiv preprint arXiv:2210.00030, 2022. 3", + "[30] Yecheng Jason Ma, William Liang, Vaidehi Som, Vikash Kumar, Amy Zhang, Osbert Bastani, and Dinesh Jayaraman. Liv: Language-image representations and rewards for robotic control. arXiv preprint arXiv:2306.00958, 2023. 3", + "[31] Oier Mees, Lukas Hermann, Erick Rosete-Beas, and Wolfram Burgard. Calvin: A benchmark for language-conditioned policy learning for long-horizon robot manipulation tasks. IEEE Robotics and Automation Letters (RA-L), 7(3):7327-7334, 2022. 6, 5", + "[32] Oier Mees, Lukas Hermann, Erick Rosete-Beas, and Wolfram Burgard. Calvin: A benchmark for language-conditioned policy learning for long-horizon robot manipulation tasks. IEEE Robotics and Automation Letters, 7(3): 7327-7334, 2022. 8, 11", + "[33] Yao Mu, Qinglong Zhang, Mengkang Hu, Wenhai Wang, Mingyu Ding, Jun Jin, Bin Wang, Jifeng Dai, Yu Qiao, and Ping Luo. Embodiedgpt: Vision-language pretraining via embodied chain of thought. arXiv preprint arXiv:2305.15021, 2023. 2, 3", + "[34] Suraj Nair, Aravind Rajeswaran, Vikash Kumar, Chelsea Finn, and Abhinav Gupta. R3m: A universal visual representation for robot manipulation. arXiv preprint arXiv:2203.12601, 2022. 3", + "[35] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, pages 16784-16804. PMLR, 2022. 2", + "[36] Maxwell Nye, Anders Johan Andreassen, Guy Gur-Ari, Henryk Michalewski, Jacob Austin, David Bieber, David Dohan, Aitor Lewkowycz, Maarten Bosma, David Luan, et al. Show your work: Scratchpads for intermediate computation with language models. arXiv preprint arXiv:2112.00114, 2021. 3", + "[37] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35: 27730-27744, 2022. 2", + "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1", + "[39] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2", + "[40] Scott Reed and Nando De Freitas. Neural programmer-interpreters. arXiv preprint arXiv:1511.06279, 2015. 3" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Scott Reed, Konrad Zolna, Emilio Parisotto, Sergio Gomez Colmenarejo, Alexander Novikov, Gabriel Barth-Maron, Mai Gimenez, Yury Sulsky, Jackie Kay, Jost Tobias Springenberg, et al. A generalist agent. arXiv preprint arXiv:2205.06175, 2022. 3, 6, 5", + "[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2", + "[43] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 6", + "[44] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2", + "[45] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 2", + "[46] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 6, 12", + "[47] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, 2017. 5", + "[48] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Ed Chi, Quoc Le, and Denny Zhou. Chain of thought prompting elicits reasoning in large language models. arXiv preprint arXiv:2201.11903, 2022. 3", + "[49] Xingqian Xu, Zhangyang Wang, Gong Zhang, Kai Wang, and Humphrey Shi. Versatile diffusion: Text, images and variations all in one diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7754-7765, 2023. 2", + "[50] Andy Zeng, Pete Florence, Jonathan Tompson, Stefan Welker, Jonathan Chien, Maria Attarian, Travis Armstrong, Ivan Krasin, Dan Duong, Vikas Sindhwani, et al. Transporter networks: Rearranging the visual world for robotic manipulation. In Conference on Robot Learning, pages 726-747. PMLR, 2021. 6, 1", + "[51] Brianna Zitkovich, Tianhe Yu, Sichun Xu, Peng Xu, Ted Xiao, Fei Xia, Jialin Wu, Paul Wohlhart, Stefan Welker, Ayzaan Wahid, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. In 7th Annual Conference on Robot Learning, 2023. 1, 2, 3" + ], + "bbox": [ + 501, + 90, + 893, + 810 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14000", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Generate Subgoal Images before Act_ Unlocking the Chain-of-Thought Reasoning in Diffusion Model for Robot Manipulation with Multimodal Prompts/d0cc6546-2994-4121-92a3-5e7ff79756f6_model.json b/2024/Generate Subgoal Images before Act_ Unlocking the Chain-of-Thought Reasoning in Diffusion Model for Robot Manipulation with Multimodal Prompts/d0cc6546-2994-4121-92a3-5e7ff79756f6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4f823ad755daa41d5241ee3e8d27c6177d700c1b --- /dev/null +++ b/2024/Generate Subgoal Images before Act_ Unlocking the Chain-of-Thought Reasoning in Diffusion Model for Robot Manipulation with Multimodal Prompts/d0cc6546-2994-4121-92a3-5e7ff79756f6_model.json @@ -0,0 +1,1826 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.02 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.325, + 0.017, + 0.723, + 0.031 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.294, + 0.032, + 0.755, + 0.047 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.131, + 0.895, + 0.178 + ], + "angle": 0, + "content": "Generate Subgoal Images before Act: Unlocking the Chain-of-Thought Reasoning in Diffusion Model for Robot Manipulation with Multimodal Prompts" + }, + { + "type": "text", + "bbox": [ + 0.136, + 0.202, + 0.828, + 0.258 + ], + "angle": 0, + "content": "Fei Ni\\(^{1}\\) Jianye Hao\\(^{1*}\\) Shiguang Wu\\(^{2}\\) Longxin Kou\\(^{1}\\) Jiashun Liu\\(^{1}\\) Yan Zheng\\(^{1}\\) Bin Wang\\(^{2}\\) Yuzheng Zhuang\\(^{2}\\) Tianjin University, China \\({ }^{2} \\mathrm{Huawei}\\) Noah's Ark Lab, China" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.324, + 0.474, + 0.765 + ], + "angle": 0, + "content": "Robotics agents often struggle to understand and follow the multi-modal prompts in complex manipulation scenes which are challenging to be sufficiently and accurately described by text alone. Moreover, for long-horizon manipulation tasks, the deviation from general instruction tends to accumulate if lack of intermediate guidance from high-level subgoals. For this, we consider can we generate subgoal images before act to enhance the instruction following in long-horizon manipulation with multi-modal prompts? Inspired by the great success of diffusion model in image generation tasks, we propose a novel hierarchical framework named as CoTDiffusion that incorporates diffusion model as a high-level planner to convert the general and multimodal prompts into coherent visual subgoal plans, which further guide the low-level policy model before action execution. We design a semantic alignment module that can anchor the progress of generated keyframes along a coherent generation chain, unlocking the chain-of-thought reasoning ability of diffusion model. Additionally, we propose bi-directional generation and frame concat mechanism to further enhance the fidelity of generated subgoal images and the accuracy of instruction following. The experiments cover various robotics manipulation scenarios including visual reasoning, visual rearrange, and visual constraints. CoTDiffusion achieves outstanding performance gain compared to the baselines without explicit subgoal generation, which proves that a subgoal image is worth a thousand words of instruction. The details and visualizations are available at https://cotdiffusion.github.io." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.791, + 0.21, + 0.807 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.816, + 0.47, + 0.879 + ], + "angle": 0, + "content": "Embodied manipulation focuses on creating generalists that can perceive, reason, and act within complex environments, which sits at the intersection of robotics control, computer vision, and natural language processing [11]. Recent ad" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.291, + 0.892, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.518, + 0.895, + 0.574 + ], + "angle": 0, + "content": "Figure 1. A motivation example of robotics manipulation tasks in multi-modal instructions. The subgoal images are worth a thousand words, inspiring us to propose a novel framework CoTDiffusion to generate goal images step-by-step before act." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.893, + 0.808 + ], + "angle": 0, + "content": "vances in vision-language model (VLM) such as CLIP [38], BLIP [22], enable open-vocabulary visual recognition and show promising potential of robotics closed-loop control by endowing them with the ability to predict robot actions in embodied robotics [5, 10, 18, 51]. However, robotics agents still face significant challenges in following instructions for long-horizon manipulation tasks, especially when the given general instructions are not progressive step-wise prompts, but implicitly contain several subtasks to accomplish. Moreover, complex manipulation scenarios with rich visual contexts are often challenging to be sufficiently and accurately described through text-only prompts, requiring multi-modal prompts to convey the instructions and intentions accurately [21], further increasing the difficulty of instruction following and task completion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.902 + ], + "angle": 0, + "content": "For example in Fig. 1, now given a short multi-modal prompt, the robot arm needs to follow the instructions and accomplish the task. The multi-modal instruction is brief but implicitly encapsulates long-horizon multi-step manipulation steps. We have two pathways to solve this task according to the recent embodied manipulation works. The" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.347, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author: jianye.hao@tju.edu.cn" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13991" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.41 + ], + "angle": 0, + "content": "first paradigm is directly leveraging the powerful VLM to encode prompts and observations into tokens and predict the entire action sequence with the subsequent decoder in an end-to-end manner, like RT-2 [51] or VIMA [21]. However, the compounding small errors over long horizons will lead to catastrophic deviations from the original task instructions due to the lack of intermediate guidance from coherent subgoals. The second paradigm is to incorporate the large language models (LLMs) to explicitly decompose the general prompts into several orderly subgoals with step-wise text instructions, like SayCan [1]. Though LLMs like ChatGPT [37] struggle to support the multi-modal prompts beyond text-only modality, we can seek help from the latest multi-modal LLMs such as LLaVA [25], EmbodiedGPT [33]. Even if we idealistically assume (M)LLM-based decomposition produces well-grounded and rational plans, generated text-only prompts tend to be extremely redundant and complex to accurately specify each subgoal in complex manipulation scene, may even be more confusing than the original general prompt and impose a heavy burden on the robot agents to re-parse and re-understand again." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.413, + 0.473, + 0.852 + ], + "angle": 0, + "content": "Considering the phrase 'a picture is worth a thousand words', subgoal images could provide higher expressive capabilities for conveying subtasks compared to sub-prompts with complex text. This motivates us to generate subgoal images step-by-step before act to enhance long-horizon instruction following. Fortunately, tremendous success in text-to-image generation sheds light on the possibility of directly translating multi-modal prompts into coherent grounded visual subgoals step by step. Our key insight is to propose a hierarchical framework CoTDiffusion that integrates diffusion model as the high-level visual planner that can understand multi-modal prompts and progressively generate chained subgoal images based on the current visual observations in a chain-of-thought manner. The chained subgoal images serve as visual milestones to anchor the execution of task, further fed into a low-level foundation model for the specific step-wise action planning to achieve the provided goal scene. The subgoal images act as a unified interface bridging the high-level visual planning and low-level action planning, decoupling the instruction understanding and action execution. The foundation module allows the high-level diffusion model to focus solely on instruction understanding and subgoal visualization without overwhelming joint action prediction training. Conversely, the high-level visual planner frees the low-level foundation model from ambiguous prompt understanding and long-horizon planning. By providing coherent subgoal images as visual landmarks, the requirements for the foundation model are reduced to basic single-object manipulation primitives." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.472, + 0.903 + ], + "angle": 0, + "content": "The key challenge to enabling CoTDiffusion to progressively generate subgoal images in a chain-of-thought manner lies in tracking the generated subgoal's progress on task" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.365 + ], + "angle": 0, + "content": "prompts. In other words, for coherent subgoal generation, diffusion model needs to know 'which step this subgoal image has reached' and 'which step needs to reach in the next subgoal image' based on task prompts. For this, we design a triple alignment module comparing the cross-frame visual contrasts between the generated subgoal image and initial observation, and align the progress back to prompts to anchor the stage of generation chains. Specifically, we use a coarse-to-fine training pipeline. First, the aligned module is coarsely pre-trained to predict residual mask patches between subgoal images for aligning spatial semantics, focusing on salient differences rather than pixel details in textures or colors. The semantic alignment module is then integrated into the diffusion model for step-wise image generation with fine-grained pixel reconstruction. Additionally, bi-directional generation and frame concatenation mechanism further enhance subgoal image fidelity and instruction following. The contributions of this work are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.367, + 0.892, + 0.457 + ], + "angle": 0, + "content": "- We propose a hierarchical framework CoTDiffusion that the high-level diffusion model translates the multi-modal prompts into coherent subgoal images in a chain-of-thought manner as visual milestones to anchor the low-level foundation model executing, enhance the instruction following on long-horizon manipulation tasks." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.459, + 0.892, + 0.534 + ], + "angle": 0, + "content": "- We design a semantic alignment module to capture the semantic relevance between visual subgoals and prompt, progressively tracking the progress of generated images along the coherent generation chain to unlock the chain-of-thought reasoning capability for multi-modal prompt." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.535, + 0.892, + 0.595 + ], + "angle": 0, + "content": "- The experiments on various long-horizon manipulation tasks empirically demonstrate that CoTDiffusion enjoys outstanding performance gain than prior methods by explicitly generating goal images step by step before act." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.367, + 0.892, + 0.595 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.61, + 0.642, + 0.626 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.635, + 0.892, + 0.651 + ], + "angle": 0, + "content": "2.1. Diffusion Models for Text-to-Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Recently, diffusion models have emerged as a powerful paradigm for high-fidelity text-to-image synthesis [9, 17, 35, 42, 45]. Models such as DALL-E 2 [39], Imagen [44] and Stable Diffusion [43] have demonstrated impressive success in generating realistic images from textual descriptions. The text instructions are tokenized as specific conditions and injected into the denoising process for controllable image generation. Moreover, some recent works have begun to explore image generation with multi-modal prompts, incorporating diverse inputs beyond just text. UniDiffuser [3] treats both image and text as sequential token streams for diffusion generations and Versatile diffusion [49] employs a multi-flow design to tackle multi-modal prompts. However, directly applying existing text-to-image approaches to complex robotics manipulation poses challenges in instruction understanding and following. Prompts" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13992" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "in manipulation tasks are often too high-level and general to effectively ground and reason, making the generated images struggle to accurately follow instructions and ground to the given scenarios, not sufficient to serve as goal images to offer fine-grained guidance for manipulation. Our work employs a semantic alignment module for multi-stage cross-modal alignment to achieve better grounding of generated images and unlock the chain-of-thought reasoning abilities of diffusion model for long-horizon tasks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.244, + 0.388, + 0.26 + ], + "angle": 0, + "content": "2.2. Robotics Manipulation and Control" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.27, + 0.473, + 0.618 + ], + "angle": 0, + "content": "With the development of VLMs for robotics, such as R3M [34], VIP [29], and LIV [30] pre-trained general visual representations for robotic perception, RT-2 [51], Gato [41] and PaLM-E [10] can convert instruction and observation into tokens and apply large transformers to predict action tokens in an end-to-end manner for robot manipulation tasks. While these works mostly focus on language-conditioned manipulation, more complex manipulation scenarios with multi-modal prompts introduced by VIMA [21] are also gaining increasing attention. Accurately instruction following becomes more challenging with general multimodal prompts, especially in long-horizon manipulation tasks. It is not trivial to directly borrow the recent works like SayCan [1] and EmbodiedGPT [33] to utilize MLLMs to decompose multi-modal instructions into step-wise text prompts. Even if we assume step-wise instructions can be rationally decomposed, they often require complex and redundant language to accurately specify subtasks and may be more confusing to understand. Inspired by the motto 'show, don't tell', our work aims to translate the general multi-modal prompts into coherent subgoal images instead of text step by step, serving as visual landmarks to enhance the instruction following in long horizon." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.634, + 0.339, + 0.65 + ], + "angle": 0, + "content": "2.3. Chain-of-Thought Reasoning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.473, + 0.901 + ], + "angle": 0, + "content": "Chain-of-Thought [48] reasoning refers to the general strategy of solving multi-step problems by decomposing them into a sequence of intermediate steps, showing great success in guiding the model to think step-by-step for enhancing instructing following and reasoning accuracy. It has recently been applied extensively in a variety of problems such as mathematical reasoning [8, 24], program execution [36, 40], commonsense or general reasoning [23, 48]. However, chain-of-thought visual reasoning has not been well explored for robotics manipulation. Inspired by prior explorations into CoT reasoning, we introduce the novel concept of chaining conditional image generations step by step to guide long-horizon manipulation. Our work employs a semantic alignment module to anchor the generated subgoal image in the entire reasoning chain, unlocking the chain-of-thought reasoning capabilities for the diffusion model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.591, + 0.105 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.116, + 0.895, + 0.314 + ], + "angle": 0, + "content": "We propose CoTDiffusion, a hierarchical framework that integrates the diffusion model as the high-level module to decompose multi-modal prompts in a chain-of-thought manner and progressively generate chained subgoals images step by step to guide the underlying foundation model for long-horizon tasks. Specifically, we first discuss the overview pipeline of CoTDiffusion in Sec. 3.1, and then introduce the pretraining of the coarse semantic alignment to capture the step-wise spatial information in Sec. 3.2. Then we discuss how to further fine-grain the CoT reasoning and grounded generation capabilities of the diffusion model in Sec. 3.3 and introduce the details of goal image conditioned foundation model for action planning in Sec. 3.4." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.327, + 0.676, + 0.343 + ], + "angle": 0, + "content": "3.1. Pipeline Overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.351, + 0.895, + 0.593 + ], + "angle": 0, + "content": "The overall pipeline of CoTDiffusion is presented in Fig. 2, decoupling the visual planning and action planning. Given the initial observation \\( x_0 \\) and a multi-modal prompt \\( \\mathcal{P} \\) as task instruction potentially needs to be reached by \\( N \\) subgoal steps, robots are required to learn a policy conditioned on the prompt and accomplish the specified long-horizon task. Our model is capable of understanding these general instructions and leverages diffusion model implicitly to decompose them into subgoal images chain \\( \\tau_x = \\{x_i\\}_{i=1}^N \\). The precisely generated goal images \\( x_i \\) then guide the foundation model's planning for inferring the action sequences \\( \\tau_a^i = \\{a_{i,t}\\}_{t=1}^T \\) with the subgoal horizon length \\( T \\). Let \\( p_\\Theta \\) model this hierarchical decision-making framework. With bi-level modules, \\( p_\\Theta \\) can be factorized into visual planning \\( p_\\phi \\), and action planning \\( p_\\psi \\). Under the Markovian assumption, the overall framework can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.499, + 0.605, + 0.907, + 0.674 + ], + "angle": 0, + "content": "\\[\np _ {\\Theta} \\left(\\left\\{\\tau_ {a} ^ {i} \\right\\} _ {i = 1} ^ {N} \\mid \\mathcal {P}, x _ {0}\\right) = \\underbrace {\\left(\\prod_ {i = 1} ^ {N} p _ {\\phi} \\left(x _ {i} \\mid \\mathcal {P} , x _ {0}\\right)\\right)} _ {\\text {v i s u a l p l a n n i n g}} \\underbrace {\\left(\\prod_ {i = 1} ^ {N} \\prod_ {t = 1} ^ {T} p _ {\\psi} \\left(a _ {i , t} \\mid x _ {i}\\right)\\right)} _ {\\text {a c t i o n p l a n n i n g}} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.895, + 0.901 + ], + "angle": 0, + "content": "The visual planning module consists of a multi-modal encoder to understand the complex task instructions and conditioned diffusion model which implicitly decomposes the task into progressive chained subgoals images. The standardized subgoal images serve as a unified interface bridging the visual planning and action planning modules. With the decoupled foundation model for action prediction, high-level visual planning focuses solely on comprehending instructions and visualizing subgoals without confusion from joint action prediction training. Conversely, the visual planning module frees the action planning module from ambiguous prompt parsing and long-horizon planning by providing coherent keyframes as visual landmarks, further reducing the capabilities requirements of the foundation model to basic single object manipulation skills." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13993" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.087, + 0.87, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.309, + 0.896, + 0.381 + ], + "angle": 0, + "content": "Figure 2. Method overview: CoTDiffusion consists of a multi-modal encoder and vision encoder \\(\\mathcal{V}\\), semantic alignment module \\(\\mathcal{S}\\), conditional diffusion model \\(\\mathcal{E}\\) and foundation model \\(\\mathcal{F}\\) for action planning. The prompt and observation tokens are combined and fed into the semantic alignment module to identify the current reasoning chain step, providing progressive guidance for the diffusion model to generate the next subgoal image. The generated keyframes are further fed to the foundation model which predicts action sequences to achieve the imagined goal scene and this recursive process repeats in a receding horizon control loop until the task is finished." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.386, + 0.427, + 0.403 + ], + "angle": 0, + "content": "3.2. Pre-training Coarse Semantic Alignment" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.41, + 0.471, + 0.576 + ], + "angle": 0, + "content": "For the challenge of explicit decomposing step-wise sub-prompts from multi-modal prompts, the diffusion model struggles to generate sequentially coherent keyframes that incrementally advance the prompt instructions and execution progress. The diffusion model fails to identify the stage of current generated keyframes onto the whole chain-of-thought progress and the generated sub-goals images may exhibit repetitions, skipping, or even backtracking to visited subgoals. To alleviate this dilemma, we design a semantic alignment module to empower diffusion model to track the progress of generated keyframes into the entire chain." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Triple Alignment Architecture Since explicit decomposition for the multi-modal prompt is intractable, the initial observation \\( x_0 \\) and prompt \\( \\mathcal{P} \\) remain fixed across the progressive generation of diffusion model. Thus, tracking the progress critically relies on extracting semantic information from the generated subgoal \\( x_i \\). To this end, we design a triple alignment module denoted as \\( S(x_0, x_i, \\mathcal{P}) \\) to compare current generated subgoal \\( x_i \\) against initial observation \\( x_0 \\) and prompts \\( \\mathcal{P} \\), serving as an information bottleneck to capture the key semantic cues about the subgoal-prompt matching to locate the current chain stage and deliver to the diffusion model for next subgoal generation. Specifically, we first capture visual representation from both the initial and current observation using a shared vision encoder \\( \\mathcal{V} \\) and concatenate with prompt tokens \\( \\mathcal{P} \\). Then they are refined through fusion module which consists of several self-attention blocks separately to obtain attention tokens \\( z_0 \\) and \\( z_i \\) aligned to the prompts. The cross-attention between \\( z_0 \\) and \\( z_i \\) aggregates a contrast representation \\( z_{\\mathrm{contrast}}^i \\) and then concatenated with prompt tokens into another fusion module to capture the progress tokens \\( z_{\\mathrm{progress}}^i \\), including" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.387, + 0.895, + 0.6 + ], + "angle": 0, + "content": "rich context about the advancement and completion of the prompted task between the two subgoals. Finally, \\( z_{\\text{progress}}^i \\) and the current visual tokens \\( x_i \\) and then fed into another cross-attention pass, which visual tokens \\( x_i \\) serve as keys and progress tokens \\( z_{\\text{progress}}^i \\) as queries to infer the final aligned tokens \\( z_{\\text{align}}^i \\). The aligned token encapsulates cues about 'which subgoal has reached now' and 'which subgoal should reach next', further injected into the diffusion model as precise semantic guidance to steer the generation of next subgoal. Through multi-stage cross-frame attention mechanisms, the triple alignment module can capture the cross-modal semantic correlation and achieve dynamic grounding of generated subgoal image into the multi-modal prompts to identify progress along the CoT reasoning chain." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Masked Patch Prediction It is not trivial to directly integrate the designed triple alignment module without pretraining into the diffusion model for joint training from scratch. Simultaneously optimizing semantic alignment and pixel-level reconstruction such as texture, color, shape, and spatial transformations overwhelms joint learning. Thus, we propose a two-stage coarse-to-fine approach decoupling semantic alignment pretraining from diffusion model finetuning, illustrated in Fig. 3. First, we pre-train coarse semantic alignment focused only on spatial correlations, without pixel generation. Considering the adjacent keyframe contrasts reveal object manipulations as mask residuals patch \\(\\hat{m}_i = x_{i + 1} - x_i\\), we can utilize these cross-frame spatial semantics to provide coarse alignment supervision. Regions with no visual changes likely correspond to stationary background elements or objects not involved in the current subtask operation. Contrastively, masked areas indicate spatial layout modifications concluding the rich spatial semantic information between keyframes. The sequence" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13994" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.458, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.232, + 0.444, + 0.246 + ], + "angle": 0, + "content": "Figure 3. Two phase of coarse-to-fine alignment pipeline." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.259, + 0.47, + 0.414 + ], + "angle": 0, + "content": "of mask residuals patch \\(\\{\\hat{m}_i\\}_{i=1}^{N-1}\\) can be viewed as implicitly decomposing the prompt into spatial manipulation steps across visual frames that can be used for alignment. So we fed the aligned tokens into a mask predictor \\(\\mathcal{D}\\) to decode the mask region and set modeling mask prediction as a pretext task. Mask prediction encourages the model to infer these salient spatial semantics and changes to incrementally track prompt completion and focus regions. The process can be formulated as: \\(\\{m_i\\}_{i=1}^N = \\sum_{i=1}^N \\mathcal{D}(m_i | z_{\\mathrm{align}}^i) S(z_{\\mathrm{align}}^i | x_0, x_i, \\mathcal{P})\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.414, + 0.469, + 0.474 + ], + "angle": 0, + "content": "By factorizing the alignment and generation objectives, the alignment module can develop basic visual reasoning skills through patch-level mask prediction before integrated into the diffusion model for controllable image synthesis." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.483, + 0.36, + 0.5 + ], + "angle": 0, + "content": "3.3. Fine-grained Diffusion Training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.469, + 0.628 + ], + "angle": 0, + "content": "In the second stage, we integrate the pretrained coarse alignment module into the diffusion model training for fine-grained image generation. To further enhance instruction following, we design a bi-directional generation that reconstructs the current frame from the aligned token, providing an additional learning constraint. We also employ a frame concatenation mechanism to enhance the consistency across generated keyframes and the fidelity of generation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.634, + 0.47, + 0.845 + ], + "angle": 0, + "content": "Bi-directional Alignment Generation Building upon coarse pretraining, we integrate the alignment module into the diffusion model for fine-grained training with the pixel-level supervision of ground truth keyframes, illustrated in Fig. 3. Here we propose bi-directional aligned generation, where the aligned token \\( z_{\\mathrm{align}}^i \\) not only guides forward prediction but also reconstructs the current frame through backward passing. In the forward pass, based on current subgoal \\( x_i \\), \\( z_{\\mathrm{align}}^i \\) provides semantic conditioning to generate the next subgoal image \\( x_{i+1} \\). In the backward pass, the same aligned token can guide diffusion model to reconstruct the current subgoal \\( x_i \\) from the next subgoal image \\( x_{i+1} \\), acting as an additional constraint. The training objective can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.851, + 0.483, + 0.902 + ], + "angle": 0, + "content": "\\[\nL = \\mathbb {E} _ {x _ {i} \\in D} [ \\underbrace {\\| \\hat {x} _ {i} - \\mathcal {E} (x _ {i - 1} , z _ {\\text {a l i g n}} ^ {i} , \\mathcal {P})} _ {\\text {F o r w a r d G e n e r a t i o n}} \\| + \\underbrace {\\| \\hat {x} _ {i - 1} - \\mathcal {E} (x _ {i} , z _ {\\text {a l i g n}} ^ {i} , \\mathcal {P})} _ {\\text {B a c k w a r d G e n e r a t i o n}} \\| ] \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.183 + ], + "angle": 0, + "content": "Bi-directional generation compels tighter instruction grounding, as accuracy is required for both progressive and reverse keyframe prediction. This dual-direction linkage enhances visual coherence and sequence controllability compared to forward-only training, improving the chain-of-thought capabilities of diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.196, + 0.893, + 0.543 + ], + "angle": 0, + "content": "Frame Concat Mechanism Generating consistent and coherent subgoal images requires considering both semantic guidance to match prompts and visual grounding into observations. The aligned tokens \\( z_{\\mathrm{align}}^i \\) and image tokens \\( x_i \\) are both fused into the denoising process in a classifier-free manner, as shown in Eq. (2). However, merely injecting current observation tokens as condition into noise model struggles to provide sufficient guidance to prevent distortions conflicting with the current observation, which is critical for precisely grounded robotics manipulations. Inspired by VDT[28], we employ frame concat mechanism to directly take current observation frame as the component of input rather than only condition for the next subgoal image generation, illustrated in Fig. 2. The current observation frame concatenated with the standard Gaussian noise gives a strong visual continuity prior, then fed into diffusion model to denoise altogether. For the training, we split the corresponding frame from the denoising frame and supervised it to the ground truth subgoal image. The rich visual context from the concatenated current frame enables coherent denoising in diffusion model further guarantees consistency with grounded observation and enforces coherence and smoothness across chain-of-thought visual planning." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.557, + 0.774, + 0.573 + ], + "angle": 0, + "content": "3.4. Goal-conditioned Policy Model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The final component in our framework is the low-level policy model for action planning, generating an action trajectory \\(\\tau_{a}^{i}\\) when given observation trajectory \\(\\tau_{x}^{i}\\) from visual planning. The policy model can be parameterized as an image-conditioned planner that infers the action \\(a_{i,t}\\) given the current observation \\(x_{i,t}\\) and the generated subgoal image \\(g_{i}\\): \\(\\tau_{a}^{i} = \\{a_{i,t}\\}_{t=1}^{T} \\sim \\prod_{t=1}^{T} p_{\\psi}(a_{i,t}|x_{i,t}, g_{i})\\). The policy model is trained by imitation learning in an end-to-end manner, using paired subgoal images and expert trajectories. Thanks to the explicit subgoal generation from high-level visual planner, the low-level policy model does not need to master complex multi-step manipulation skills in long-horizon. Therefore, we simplify the policy model architecture and implement it as a simple MLP to only possess basic single-object manipulation primitives, which greatly reduces the burden of policy model training. Although we do not seek improvement through a more powerful policy model in this paper, CoTDiffusion is flexible to incorporate any goal-conditioned design like transformers or diffusion policies. For more details of low-level policy model, please refer to Appendix E.3." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13995" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.086, + 0.873, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.331, + 0.843, + 0.345 + ], + "angle": 0, + "content": "Figure 4. Visualization of CoTDiffusion in three typical long-horizon tasks with multi-modal prompts in VIMA-BENCH." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.358, + 0.21, + 0.374 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.383, + 0.254, + 0.399 + ], + "angle": 0, + "content": "4.1. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.407, + 0.474, + 0.71 + ], + "angle": 0, + "content": "Benchmark & Tasks We conduct evaluation on VIMABENCH, a benchmark suite for multimodal robot learning, which is built on the Ravens robot simulator [50]. VIMABENCH supports extensible collections of objects and textures to compose multi-modal prompts and to procedurally generate a large number of tasks. VIMA-BENCH can generate large quantities of expert trajectories via scripted oracle agents, and the ground truth keyframes that can leveraged for the training of CoTDiffusion. VIMA-BENCH contains various tasks ranging from simple object manipulation to multi-object manipulation and we select three representative long-horizon manipulation tasks in VIMABENCH-visual rearrangement, visual constraints, and visual reasoning, which represent different levels of complexity and requirements for embodied manipulation. These tasks represent different levels of complexity and requirements for embodied manipulation and can test the ability to understand spatial relationships and perform precise object manipulations. For more details regarding the simulation benchmark and tasks setting, please refer to Appendix A." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.721, + 0.47, + 0.751 + ], + "angle": 0, + "content": "Evaluation Metric The evaluation metrics we use in the experiments cover three aspects as following:" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.751, + 0.47, + 0.796 + ], + "angle": 0, + "content": "- Image Fidelity - measures the quality and realism of the generated keyframe images. We choose Fréchet Inception Distance (FID) [15] as a useful metric." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.469, + 0.856 + ], + "angle": 0, + "content": "- Instruction Following - evaluates how well the step-by-step keyframes cover the complex instructions. We use CLIP-based text-image similarity scores [14] between the prompt and keyframes to quantify instruction alignment." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "- Task Completion - tests the end-to-end utility by executing the embodied manipulation task. The final success rate on long-horizon tasks can be a fair metric." + }, + { + "type": "list", + "bbox": [ + 0.076, + 0.751, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.358, + 0.608, + 0.373 + ], + "angle": 0, + "content": "4.2. Baselines" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.383, + 0.892, + 0.427 + ], + "angle": 0, + "content": "There are several existing methods for constructing robot manipulation policies conditioned on complex prompts, which we use as baselines in our experiments:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.429, + 0.892, + 0.504 + ], + "angle": 0, + "content": "- Gato[41], a generalist agent that can solve tasks from multiple domains where tasks are specified by prompting the model with the observation and action subsequence. For fair comparisons, we provide the same conditioning manner with multi-modal prompts." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.504, + 0.892, + 0.579 + ], + "angle": 0, + "content": "- Flamingo[2], a vision-language model that learns to generate textual completion in response to multimodal prompts. We borrow the architecture in [21] which adapts it to support decision-masking by replacing the output layer with robot action heads." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.58, + 0.892, + 0.669 + ], + "angle": 0, + "content": "- VIMA[21], the first transformer-based generalist robot agent that processes manipulation tasks with multimodal prompts. VIMA adpots an object-centric approach to flatten all the observation and prompts into object tokens sequence and predicts motor actions autoregressively and demonstrates SOTA performance on VIMA-BENCH." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.67, + 0.892, + 0.744 + ], + "angle": 0, + "content": "- SuSIE[4], the concurrent work for visual planning in manipulation tasks on the CALVIN [31], incorporating pretrained image-editing model and the low-level goal-conditioned policy. For fair comparison, we finetune the model with the same oracle data with CoTDiffusion." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.429, + 0.892, + 0.744 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.756, + 0.816, + 0.772 + ], + "angle": 0, + "content": "4.3. Quantitative Results of Success Rate" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We begin by comparing the performance of CoTDiffusion and baselines to solve long-horizon tasks in visual rearrange, visual reasoning, and visual constraints. The performance comparisons shown in Tab. 1 demonstrate CoT-Diffusion significantly outperforms other baselines in success rate. The baselines can be divided into two kinds of planners, including abstract planner and visual planner. Abstract planner like Gato, Flamingo and VIMA directly" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13996" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.514 + ], + "angle": 0, + "content": "map general prompts to subsequent actions in an end-to-end manner. Gato and Flamingo gets low success rates on long-horizon tasks without explicit subgoal generation to correct the accumulative deviation errors from the instructions. In contrast, Visual planner like SuSIE and CoTDiffusion can generate intermediate goal images to guide the action planning, which can enhance instruction following for long-horizon tasks via visual planning. However, SuSIE lacks intrinsic chain-of-thought reasoning capabilities, struggling to generate logically progressing subgoal sequences as coherent plans from general prompts. To enable comparison, we grant SuSIE the manually decomposed prompts into privileged step-wise sub-prompts which are unavailable in fair settings, denoted as 'SuSIE + sub-prompts'. Nonetheless, its performance still lags CoTDiffusion, which can implicitly align generated images with multi-modal prompts through learned correspondence between visual signals and language semantics. One more potential reason to restrict the performance of SuSIE is that provided sub-prompts are challenging to perfectly encapsulate complex instructions. In contrast, CoTDiffusion develops intrinsic chain-of-thought reasoning and alignment for generated subgoal images for flexible visual planning directly from the raw multi-modal prompts, without the need for pre-decomposed sub-prompts. The comparisons highlight the importance of chain-of-thought capabilities in visual planning, especially the role of implicit semantic alignment for instruction following long-horizon manipulation tasks." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.52, + 0.468, + 0.619 + ], + "angle": 0, + "content": "
MethodologyRearrangeReasoningConstraintsOverall
Gato6.4 ± 1.32.5 ± 0.425.2 ± 3.111.4 ± 1.6
Flamingo17.5 ± 1.63.0 ± 0.536.1 ± 4.218.9 ± 2.1
VIMA43.1 ± 3.338.2 ± 4.467.2 ± 5.249.5 ± 4.3
SuSIE2.7 ± 0.33.1 ± 0.624.3 ± 5.910.0 ± 2.3
+sub-prompts37.7 ± 6.239.0 ± 4.552.3 ± 7.043.0 ± 5.9
CoTDiffusion59.0 ± 1.751.7 ± 2.683.1 ± 4.764.6 ± 3.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.625, + 0.47, + 0.655 + ], + "angle": 0, + "content": "Table 1. The evaluations of success rates on three typical long-horizon tasks with multi-modal prompts." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.666, + 0.245, + 0.683 + ], + "angle": 0, + "content": "4.4. Further Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Robustness to Insufficient Perception Rich visual observations from diverse views are crucial for complex robot manipulation tasks. Restricted perspectives limit rich representations used for action planning, degrading performance on long-horizon multi-object manipulation. We evaluate different methods in single-view including the top and front views provided from VIMA-BENCH. As shown in Tab. 2, visual planning approach demonstrates greater robustness to limited observations, with a smaller performance drop in single-view compared to abstract methods. We attribute this to two potential reasons: First, accurate and grounded subgoal images generated in visual planners provide supplemental visual context, which can partly compensate for the insufficient perception to aid robustness under single-" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.089, + 0.889, + 0.196 + ], + "angle": 0, + "content": "
MethodologyMulti-ViewSingle-ViewPerformance Drop
Gato11.4 ± 1.66.5 ± 1.64.9(42% ↓)
Flamingo18.9 ± 2.112.0 ± 2.46.9(36.2% ↓)
VIMA49.5 ± 4.334.9 ± 3.414.6(29.4% ↓)
SuSIE10.0 ± 2.37.9 ± 2.02.1(21.0% ↓)
+sub-prompts43.0 ± 5.935.6 ± 6.17.4(18.8% ↓)
CoTDiffusion64.6 ± 3.056.0 ± 2.48.6(13.2% ↓)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.201, + 0.892, + 0.23 + ], + "angle": 0, + "content": "Table 2. The evaluations of performance drop of different methods on single-view and multi-view from VIMA-BENCH." + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.246, + 0.892, + 0.273 + ], + "angle": 0, + "content": "Table 3. Quantitative comparisons of FID between methods on all three tasks, including visual rearrange, reasoning and constraints." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.274, + 0.918, + 0.391 + ], + "angle": 0, + "content": "
MethodologyRearrangeReasoningConstraints
SuSIE [4]19.718.323.2
- w/o finetune38.4 ↑33.0 ↑52.7 ↑
- w/o stepwise prompt32.6 ↑28.1 ↑34.7 ↑
CoTDiffusion (ours)11.310.88.6
- w/o coarse pre-train23.1 ↑16.2 ↑18.9 ↑
- w/o bi-direction align18.7 ↑15.0 ↑15.8 ↑
- w/o frame concat15.4 ↑13.7 ↑12.1 ↑
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.399, + 0.892, + 0.52 + ], + "angle": 0, + "content": "view. Second, by providing coherent subgoal images as visual landmarks, the requirements for the low-level action planner are reduced to basic single-object manipulation primitives in short horizon, with less reliance on rich visual perceptions. The experiments demonstrate that CoTDiffusion enjoys better robustness to restricted perception than abstract planners, highlighting the benefits of hierarchical framework decoupled visual planning and action planning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.529, + 0.893, + 0.862 + ], + "angle": 0, + "content": "Fidelity of Image Generation Here we conduct further analysis of visual planners by comparing goal image quality against ground truth keyframes with FID as evaluation metrics. As Tab. 3 shows, CoTDiffusion still achieves much better fidelity than SuSIE even though SuSIE has got improved after fine-tuning on the same datasets in VIMABENCH. Moreover, SuSIE can only understand short instructions of single-object manipulation tasks, struggling to handle long-horizon instructions due to the lack of chain-of-thought reasoning capabilities to handle multi-modal prompts. With step-wise sub-prompts decomposed in advance, the performance of SuSIE gets largely raised but still underperforms CoTDiffusion, which has no need to explicitly decompose the general prompts and can generate subgoal images in an implicit chain-of-thought manner. Additionally, ablating coarse pretraining and bi-directional generation degrades performance, validating their benefits. The coarse-to-fine semantic alignment training training allows developing spatial reasoning prior to synthesis. Frame concatenation further guides coherent denoising by providing rich context information as visual priors to ground the current observation and enhance the fidelity of generation." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Accuracy of Instruction Following We evaluate instruction following accuracy via CLIP similarity between gen" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13997" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.531 + ], + "angle": 0, + "content": "erated keyframes and general prompts, normalized by the CLIP score between ground truth ultimate goal image and prompts. The results in Fig. 5 demonstrate that CoT-Diffusion enjoys better instruction following capabilities. Without chain-of-thought reasoning abilities, SuSIE struggles to follow instructions when given general multi-modal prompts, let alone generate subgoal images with smooth progressions. Progressively providing privileged prompts step by step improves SuSIE in instruction following and gradual advancement towards ultimate goals. However, text prompts decomposed by rules are not always perfect to convey the original multi-modal prompt, so the reliance on these prompts restricts the performance of SuSIE. In contrast, the coarse alignment pretraining and bi-directional generation can assist the diffusion model in tracking the progress of generated keyframes throughout the entire chain and generate sequenced keyframes incrementally advancing prompt instructions. Moreover, the implicit decomposition maintains semantic coherence between subgoals without overly large jumps and the smoothness and continuity between visual milestones make them more reachable by downstream foundation models with just single-object manipulation capabilities. Additionally, we observe that the bi-directional generation may impedes the diffusion model training if without coarse semantic pretraining. As the initial observation and prompt remain fixed across different generation steps, the align tokens at various stage tend to so similar that may confuse simultaneous progressive generation and reconstruction training." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.548, + 0.473, + 0.73 + ], + "angle": 0, + "content": "Generalization across Tasks We evaluate the generalization ability in three levels with increasing difficulty: placement generalization which randomizes the novel placement of objects (L1), object generalization which provides the objects with novel attributes (L2), and task combinatorial generalization which complexes the prompts with extra novel instruction (L3). As Fig. 6 shows, CoTDiffusion exhibits strong zero-shot generalization on unfamiliar objects, colors, and shapes at different placements by explicitly visualizing them into coherent subgoals grounded on new concepts. This avoids re-grounding concepts from multi-modal prompts in the low-level foundation model for" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.74, + 0.468, + 0.857 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.868, + 0.468, + 0.897 + ], + "angle": 0, + "content": "Figure 5. The normalized CLIP scores for each generation step, reflecting the step-wise accuracy of instruction following." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.093, + 0.885, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.525, + 0.251, + 0.866, + 0.264 + ], + "angle": 0, + "content": "Figure 6. The evaluation on various generalization levels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.277, + 0.892, + 0.489 + ], + "angle": 0, + "content": "the placement and object. For L3 generalization, taking an example, we complex the prompt such as 'rearrange ... then rotate/twist/stack ...' to demand extra tasks such as rotating, twisting degrees, or stacking objects. CoTDiffusion achieves outstanding gain in the zero-shot performance of combinatorial tasks. When CoTDiffusion visualizes novel concepts into goal images, the foundation model can still accomplish the stack by simply achieving the provided subgoal, with no need to inherently understand novel skills like stack. This demonstrates the power of 'a image is worth a thousand words' - the subgoal images facilitate the generalization of the foundation model to unseen objects and novel combinatorial task while the foundation model simply achieves them with reuse of simple known primitives." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.509, + 0.619, + 0.525 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.536, + 0.892, + 0.807 + ], + "angle": 0, + "content": "We presented CoTDiffusion, a hierarchical framework that integrates diffusion model as high-level module to translate the general multi-modal prompts into coherent subgoal images, serves as the visual milestones to anchor the low-level foundation model to plan action sequences, termed as 'generate subgoal images before act'. With the coarse-to-fine training for semantic alignment module, CoTDiffusion can identify the progress of generated subgoals images along reasoning chains, unlocking the chain-of-thought reasoning capabilities of diffusion model for long-horizon manipulation tasks. The experiments cover various long-horizon manipulation scenarios in VIMA-BENCH, and CoTDiffusion show the strong instruction following and outstanding performance gain compared to existed methods without visual planning. Incorporating commonsense knowledge from pre-trained MLLM like GPT-4V provides an avenue for more generalizable and promising reasoning in CoTDiffusion, which leaves us our future work." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.829, + 0.681, + 0.846 + ], + "angle": 0, + "content": "6. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This work is supported by the National Natural Science Foundation of China (Grant Nos. 92370132) and the Xiaomi Young Talents Program of Xiaomi Foundation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13998" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, et al. Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.243 + ], + "angle": 0, + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: A visual language model for few-shot learning. NeurIPS, 2022. 6, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.47, + 0.3 + ], + "angle": 0, + "content": "[3] Fan Bao, Shen Nie, Kaiwen Xue, Chongxuan Li, Shi Pu, Yaole Wang, Gang Yue, Yue Cao, Hang Su, and Jun Zhu. One transformer fits all distributions in multi-modal diffusion at scale. arXiv preprint arXiv:2303.06555, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.302, + 0.47, + 0.37 + ], + "angle": 0, + "content": "[4] Kevin Black, Mitsuhiko Nakamoto, Pranav Atreya, Homer Walke, Chelsea Finn, Aviral Kumar, and Sergey Levine. Zero-shot robotic manipulation with pretrained image-editing diffusion models. arXiv preprint arXiv:2310.10639, 2023. 6, 7, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.373, + 0.469, + 0.442 + ], + "angle": 0, + "content": "[5] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.444, + 0.469, + 0.486 + ], + "angle": 0, + "content": "[6] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. arXiv preprint arXiv:2211.09800, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.488, + 0.469, + 0.543 + ], + "angle": 0, + "content": "[7] Cheng Chi, Siyuan Feng, Yilun Du, Zhenjia Xu, Eric Cousineau, Benjamin Burchfiel, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. arXiv preprint arXiv:2303.04137. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.545, + 0.469, + 0.613 + ], + "angle": 0, + "content": "[8] Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.616, + 0.469, + 0.671 + ], + "angle": 0, + "content": "[9] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In Advances in Neural Information Processing Systems, pages 8780-8794. Curran Associates, Inc., 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.673, + 0.469, + 0.743 + ], + "angle": 0, + "content": "[10] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palm- e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.745, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[11] Jiafei Duan, Samson Yu, Hui Li Tan, Hongyuan Zhu, and Cheston Tan. A survey of embodied ai: From simulators to research tasks. IEEE Transactions on Emerging Topics in Computational Intelligence, 6(2):230-244, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[12] Theophile Gervet, Zhou Xian, Nikolaos Gkanatsios, and Katerina Fragkiadaki. Act3d: Infinite resolution action detection transformer for robotic manipulation. arXiv preprint arXiv:2306.17817, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[13] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[14] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.216 + ], + "angle": 0, + "content": "[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.892, + 0.246 + ], + "angle": 0, + "content": "[16] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.247, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.289, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[18] Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. Voxposer: Composable 3d value maps for robotic manipulation with language models. arXiv preprint arXiv:2307.05973, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.345, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[19] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General perception with iterative attention. In International conference on machine learning, pages 4651-4664. PMLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.414, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[20] Stephen James, Zicong Ma, David Rovick Arrojo, and Andrew J Davison. Rlbench: The robot learning benchmark & learning environment. IEEE Robotics and Automation Letters, 5(2):3019-3026, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.47, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[21] Yunfan Jiang, Agrim Gupta, Zichen Zhang, Guanzhi Wang, Yongqiang Dou, Yanjun Chen, Li Fei-Fei, Anima Anandkumar, Yuke Zhu, and Linxi Fan. Vima: General robot manipulation with multimodal prompts. arXiv preprint arXiv:2210.03094, 2022. 1, 2, 3, 6, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.539, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[22] Dongxu Li, Junnan Li, and Steven CH Hoi. Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. arXiv preprint arXiv:2305.14720, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.595, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[23] Zhengzhong Liang, Steven Bethard, and Mihai Surdeanu. Explainable multi-hop verbal reasoning through internal monologue. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.665, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[24] Wang Ling, Dani Yogatama, Chris Dyer, and Phil Blunson. Program induction by rationale generation: Learning to solve and explain algebraic word problems. arXiv preprint arXiv:1705.04146, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.72, + 0.892, + 0.76 + ], + "angle": 0, + "content": "[25] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.762, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[26] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.79, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[27] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[28] Haoyu Lu, Guoxing Yang, Nanyi Fei, Yuqi Huo, Zhiwu Lu, Ping Luo, and Mingyu Ding. Vdt: An empirical study on video diffusion with transformers. arXiv preprint arXiv:2305.13311, 2023. 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13999" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.148 + ], + "angle": 0, + "content": "[29] Yecheng Jason Ma, Shagun Sodhani, Dinesh Jayaraman, Obert Bastani, Vikash Kumar, and Amy Zhang. Vip: Towards universal visual reward and representation via value-implicit pre-training. arXiv preprint arXiv:2210.00030, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.471, + 0.207 + ], + "angle": 0, + "content": "[30] Yecheng Jason Ma, William Liang, Vaidehi Som, Vikash Kumar, Amy Zhang, Osbert Bastani, and Dinesh Jayaraman. Liv: Language-image representations and rewards for robotic control. arXiv preprint arXiv:2306.00958, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.471, + 0.276 + ], + "angle": 0, + "content": "[31] Oier Mees, Lukas Hermann, Erick Rosete-Beas, and Wolfram Burgard. Calvin: A benchmark for language-conditioned policy learning for long-horizon robot manipulation tasks. IEEE Robotics and Automation Letters (RA-L), 7(3):7327-7334, 2022. 6, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.471, + 0.348 + ], + "angle": 0, + "content": "[32] Oier Mees, Lukas Hermann, Erick Rosete-Beas, and Wolfram Burgard. Calvin: A benchmark for language-conditioned policy learning for long-horizon robot manipulation tasks. IEEE Robotics and Automation Letters, 7(3): 7327-7334, 2022. 8, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.348, + 0.471, + 0.419 + ], + "angle": 0, + "content": "[33] Yao Mu, Qinglong Zhang, Mengkang Hu, Wenhai Wang, Mingyu Ding, Jun Jin, Bin Wang, Jifeng Dai, Yu Qiao, and Ping Luo. Embodiedgpt: Vision-language pretraining via embodied chain of thought. arXiv preprint arXiv:2305.15021, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.42, + 0.471, + 0.475 + ], + "angle": 0, + "content": "[34] Suraj Nair, Aravind Rajeswaran, Vikash Kumar, Chelsea Finn, and Abhinav Gupta. R3m: A universal visual representation for robot manipulation. arXiv preprint arXiv:2203.12601, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.477, + 0.471, + 0.56 + ], + "angle": 0, + "content": "[35] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, pages 16784-16804. PMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.561, + 0.471, + 0.645 + ], + "angle": 0, + "content": "[36] Maxwell Nye, Anders Johan Andreassen, Guy Gur-Ari, Henryk Michalewski, Jacob Austin, David Bieber, David Dohan, Aitor Lewkowycz, Maarten Bosma, David Luan, et al. Show your work: Scratchpads for intermediate computation with language models. arXiv preprint arXiv:2112.00114, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.646, + 0.471, + 0.73 + ], + "angle": 0, + "content": "[37] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35: 27730-27744, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.73, + 0.471, + 0.815 + ], + "angle": 0, + "content": "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.471, + 0.871 + ], + "angle": 0, + "content": "[39] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.872, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[40] Scott Reed and Nando De Freitas. Neural programmer-interpreters. arXiv preprint arXiv:1511.06279, 2015. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.162 + ], + "angle": 0, + "content": "[41] Scott Reed, Konrad Zolna, Emilio Parisotto, Sergio Gomez Colmenarejo, Alexander Novikov, Gabriel Barth-Maron, Mai Gimenez, Yury Sulsky, Jackie Kay, Jost Tobias Springenberg, et al. A generalist agent. arXiv preprint arXiv:2205.06175, 2022. 3, 6, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.894, + 0.233 + ], + "angle": 0, + "content": "[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.894, + 0.303 + ], + "angle": 0, + "content": "[43] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.304, + 0.893, + 0.387 + ], + "angle": 0, + "content": "[44] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.388, + 0.893, + 0.43 + ], + "angle": 0, + "content": "[45] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.431, + 0.893, + 0.473 + ], + "angle": 0, + "content": "[46] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 6, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.474, + 0.893, + 0.53 + ], + "angle": 0, + "content": "[47] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.53, + 0.893, + 0.586 + ], + "angle": 0, + "content": "[48] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Ed Chi, Quoc Le, and Denny Zhou. Chain of thought prompting elicits reasoning in large language models. arXiv preprint arXiv:2201.11903, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.587, + 0.893, + 0.657 + ], + "angle": 0, + "content": "[49] Xingqian Xu, Zhangyang Wang, Gong Zhang, Kai Wang, and Humphrey Shi. Versatile diffusion: Text, images and variations all in one diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7754-7765, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.657, + 0.893, + 0.74 + ], + "angle": 0, + "content": "[50] Andy Zeng, Pete Florence, Jonathan Tompson, Stefan Welker, Jonathan Chien, Maria Attarian, Travis Armstrong, Ivan Krasin, Dan Duong, Vikas Sindhwani, et al. Transporter networks: Rearranging the visual world for robotic manipulation. In Conference on Robot Learning, pages 726-747. PMLR, 2021. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.741, + 0.893, + 0.811 + ], + "angle": 0, + "content": "[51] Brianna Zitkovich, Tianhe Yu, Sichun Xu, Peng Xu, Ted Xiao, Fei Xia, Jialin Wu, Paul Wohlhart, Stefan Welker, Ayzaan Wahid, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. In 7th Annual Conference on Robot Learning, 2023. 1, 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.894, + 0.811 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14000" + } + ] +] \ No newline at end of file diff --git a/2024/Generate Subgoal Images before Act_ Unlocking the Chain-of-Thought Reasoning in Diffusion Model for Robot Manipulation with Multimodal Prompts/d0cc6546-2994-4121-92a3-5e7ff79756f6_origin.pdf b/2024/Generate Subgoal Images before Act_ Unlocking the Chain-of-Thought Reasoning in Diffusion Model for Robot Manipulation with Multimodal Prompts/d0cc6546-2994-4121-92a3-5e7ff79756f6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..90101b0cc5147f00f931befac6a3ab24d26f5b7d --- /dev/null +++ b/2024/Generate Subgoal Images before Act_ Unlocking the Chain-of-Thought Reasoning in Diffusion Model for Robot Manipulation with Multimodal Prompts/d0cc6546-2994-4121-92a3-5e7ff79756f6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e46a4fcc0c2ff1a7cea1de1a70be7beec05aa7075dd18f2873d60792d869c866 +size 1543214 diff --git a/2024/Generate Subgoal Images before Act_ Unlocking the Chain-of-Thought Reasoning in Diffusion Model for Robot Manipulation with Multimodal Prompts/full.md b/2024/Generate Subgoal Images before Act_ Unlocking the Chain-of-Thought Reasoning in Diffusion Model for Robot Manipulation with Multimodal Prompts/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5af56c59d0bdbb4c23a9c75b9fc1a9dae60cf96c --- /dev/null +++ b/2024/Generate Subgoal Images before Act_ Unlocking the Chain-of-Thought Reasoning in Diffusion Model for Robot Manipulation with Multimodal Prompts/full.md @@ -0,0 +1,226 @@ +# Generate Subgoal Images before Act: Unlocking the Chain-of-Thought Reasoning in Diffusion Model for Robot Manipulation with Multimodal Prompts + +Fei Ni $^{1}$ Jianye Hao $^{1*}$ Shiguang Wu $^{2}$ Longxin Kou $^{1}$ Jiashun Liu $^{1}$ Yan Zheng $^{1}$ Bin Wang $^{2}$ Yuzheng Zhuang $^{2}$ Tianjin University, China ${ }^{2} \mathrm{Huawei}$ Noah's Ark Lab, China + +# Abstract + +Robotics agents often struggle to understand and follow the multi-modal prompts in complex manipulation scenes which are challenging to be sufficiently and accurately described by text alone. Moreover, for long-horizon manipulation tasks, the deviation from general instruction tends to accumulate if lack of intermediate guidance from high-level subgoals. For this, we consider can we generate subgoal images before act to enhance the instruction following in long-horizon manipulation with multi-modal prompts? Inspired by the great success of diffusion model in image generation tasks, we propose a novel hierarchical framework named as CoTDiffusion that incorporates diffusion model as a high-level planner to convert the general and multimodal prompts into coherent visual subgoal plans, which further guide the low-level policy model before action execution. We design a semantic alignment module that can anchor the progress of generated keyframes along a coherent generation chain, unlocking the chain-of-thought reasoning ability of diffusion model. Additionally, we propose bi-directional generation and frame concat mechanism to further enhance the fidelity of generated subgoal images and the accuracy of instruction following. The experiments cover various robotics manipulation scenarios including visual reasoning, visual rearrange, and visual constraints. CoTDiffusion achieves outstanding performance gain compared to the baselines without explicit subgoal generation, which proves that a subgoal image is worth a thousand words of instruction. The details and visualizations are available at https://cotdiffusion.github.io. + +# 1. Introduction + +Embodied manipulation focuses on creating generalists that can perceive, reason, and act within complex environments, which sits at the intersection of robotics control, computer vision, and natural language processing [11]. Recent ad + +![](images/80efe11818f03782148d8dabaaef5b81279ec6b189539900031851b55d44a916.jpg) +Figure 1. A motivation example of robotics manipulation tasks in multi-modal instructions. The subgoal images are worth a thousand words, inspiring us to propose a novel framework CoTDiffusion to generate goal images step-by-step before act. + +vances in vision-language model (VLM) such as CLIP [38], BLIP [22], enable open-vocabulary visual recognition and show promising potential of robotics closed-loop control by endowing them with the ability to predict robot actions in embodied robotics [5, 10, 18, 51]. However, robotics agents still face significant challenges in following instructions for long-horizon manipulation tasks, especially when the given general instructions are not progressive step-wise prompts, but implicitly contain several subtasks to accomplish. Moreover, complex manipulation scenarios with rich visual contexts are often challenging to be sufficiently and accurately described through text-only prompts, requiring multi-modal prompts to convey the instructions and intentions accurately [21], further increasing the difficulty of instruction following and task completion. + +For example in Fig. 1, now given a short multi-modal prompt, the robot arm needs to follow the instructions and accomplish the task. The multi-modal instruction is brief but implicitly encapsulates long-horizon multi-step manipulation steps. We have two pathways to solve this task according to the recent embodied manipulation works. The + +first paradigm is directly leveraging the powerful VLM to encode prompts and observations into tokens and predict the entire action sequence with the subsequent decoder in an end-to-end manner, like RT-2 [51] or VIMA [21]. However, the compounding small errors over long horizons will lead to catastrophic deviations from the original task instructions due to the lack of intermediate guidance from coherent subgoals. The second paradigm is to incorporate the large language models (LLMs) to explicitly decompose the general prompts into several orderly subgoals with step-wise text instructions, like SayCan [1]. Though LLMs like ChatGPT [37] struggle to support the multi-modal prompts beyond text-only modality, we can seek help from the latest multi-modal LLMs such as LLaVA [25], EmbodiedGPT [33]. Even if we idealistically assume (M)LLM-based decomposition produces well-grounded and rational plans, generated text-only prompts tend to be extremely redundant and complex to accurately specify each subgoal in complex manipulation scene, may even be more confusing than the original general prompt and impose a heavy burden on the robot agents to re-parse and re-understand again. + +Considering the phrase 'a picture is worth a thousand words', subgoal images could provide higher expressive capabilities for conveying subtasks compared to sub-prompts with complex text. This motivates us to generate subgoal images step-by-step before act to enhance long-horizon instruction following. Fortunately, tremendous success in text-to-image generation sheds light on the possibility of directly translating multi-modal prompts into coherent grounded visual subgoals step by step. Our key insight is to propose a hierarchical framework CoTDiffusion that integrates diffusion model as the high-level visual planner that can understand multi-modal prompts and progressively generate chained subgoal images based on the current visual observations in a chain-of-thought manner. The chained subgoal images serve as visual milestones to anchor the execution of task, further fed into a low-level foundation model for the specific step-wise action planning to achieve the provided goal scene. The subgoal images act as a unified interface bridging the high-level visual planning and low-level action planning, decoupling the instruction understanding and action execution. The foundation module allows the high-level diffusion model to focus solely on instruction understanding and subgoal visualization without overwhelming joint action prediction training. Conversely, the high-level visual planner frees the low-level foundation model from ambiguous prompt understanding and long-horizon planning. By providing coherent subgoal images as visual landmarks, the requirements for the foundation model are reduced to basic single-object manipulation primitives. + +The key challenge to enabling CoTDiffusion to progressively generate subgoal images in a chain-of-thought manner lies in tracking the generated subgoal's progress on task + +prompts. In other words, for coherent subgoal generation, diffusion model needs to know 'which step this subgoal image has reached' and 'which step needs to reach in the next subgoal image' based on task prompts. For this, we design a triple alignment module comparing the cross-frame visual contrasts between the generated subgoal image and initial observation, and align the progress back to prompts to anchor the stage of generation chains. Specifically, we use a coarse-to-fine training pipeline. First, the aligned module is coarsely pre-trained to predict residual mask patches between subgoal images for aligning spatial semantics, focusing on salient differences rather than pixel details in textures or colors. The semantic alignment module is then integrated into the diffusion model for step-wise image generation with fine-grained pixel reconstruction. Additionally, bi-directional generation and frame concatenation mechanism further enhance subgoal image fidelity and instruction following. The contributions of this work are as follows: + +- We propose a hierarchical framework CoTDiffusion that the high-level diffusion model translates the multi-modal prompts into coherent subgoal images in a chain-of-thought manner as visual milestones to anchor the low-level foundation model executing, enhance the instruction following on long-horizon manipulation tasks. +- We design a semantic alignment module to capture the semantic relevance between visual subgoals and prompt, progressively tracking the progress of generated images along the coherent generation chain to unlock the chain-of-thought reasoning capability for multi-modal prompt. +- The experiments on various long-horizon manipulation tasks empirically demonstrate that CoTDiffusion enjoys outstanding performance gain than prior methods by explicitly generating goal images step by step before act. + +# 2. Related Work + +# 2.1. Diffusion Models for Text-to-Image Generation + +Recently, diffusion models have emerged as a powerful paradigm for high-fidelity text-to-image synthesis [9, 17, 35, 42, 45]. Models such as DALL-E 2 [39], Imagen [44] and Stable Diffusion [43] have demonstrated impressive success in generating realistic images from textual descriptions. The text instructions are tokenized as specific conditions and injected into the denoising process for controllable image generation. Moreover, some recent works have begun to explore image generation with multi-modal prompts, incorporating diverse inputs beyond just text. UniDiffuser [3] treats both image and text as sequential token streams for diffusion generations and Versatile diffusion [49] employs a multi-flow design to tackle multi-modal prompts. However, directly applying existing text-to-image approaches to complex robotics manipulation poses challenges in instruction understanding and following. Prompts + +in manipulation tasks are often too high-level and general to effectively ground and reason, making the generated images struggle to accurately follow instructions and ground to the given scenarios, not sufficient to serve as goal images to offer fine-grained guidance for manipulation. Our work employs a semantic alignment module for multi-stage cross-modal alignment to achieve better grounding of generated images and unlock the chain-of-thought reasoning abilities of diffusion model for long-horizon tasks. + +# 2.2. Robotics Manipulation and Control + +With the development of VLMs for robotics, such as R3M [34], VIP [29], and LIV [30] pre-trained general visual representations for robotic perception, RT-2 [51], Gato [41] and PaLM-E [10] can convert instruction and observation into tokens and apply large transformers to predict action tokens in an end-to-end manner for robot manipulation tasks. While these works mostly focus on language-conditioned manipulation, more complex manipulation scenarios with multi-modal prompts introduced by VIMA [21] are also gaining increasing attention. Accurately instruction following becomes more challenging with general multimodal prompts, especially in long-horizon manipulation tasks. It is not trivial to directly borrow the recent works like SayCan [1] and EmbodiedGPT [33] to utilize MLLMs to decompose multi-modal instructions into step-wise text prompts. Even if we assume step-wise instructions can be rationally decomposed, they often require complex and redundant language to accurately specify subtasks and may be more confusing to understand. Inspired by the motto 'show, don't tell', our work aims to translate the general multi-modal prompts into coherent subgoal images instead of text step by step, serving as visual landmarks to enhance the instruction following in long horizon. + +# 2.3. Chain-of-Thought Reasoning + +Chain-of-Thought [48] reasoning refers to the general strategy of solving multi-step problems by decomposing them into a sequence of intermediate steps, showing great success in guiding the model to think step-by-step for enhancing instructing following and reasoning accuracy. It has recently been applied extensively in a variety of problems such as mathematical reasoning [8, 24], program execution [36, 40], commonsense or general reasoning [23, 48]. However, chain-of-thought visual reasoning has not been well explored for robotics manipulation. Inspired by prior explorations into CoT reasoning, we introduce the novel concept of chaining conditional image generations step by step to guide long-horizon manipulation. Our work employs a semantic alignment module to anchor the generated subgoal image in the entire reasoning chain, unlocking the chain-of-thought reasoning capabilities for the diffusion model. + +# 3. Method + +We propose CoTDiffusion, a hierarchical framework that integrates the diffusion model as the high-level module to decompose multi-modal prompts in a chain-of-thought manner and progressively generate chained subgoals images step by step to guide the underlying foundation model for long-horizon tasks. Specifically, we first discuss the overview pipeline of CoTDiffusion in Sec. 3.1, and then introduce the pretraining of the coarse semantic alignment to capture the step-wise spatial information in Sec. 3.2. Then we discuss how to further fine-grain the CoT reasoning and grounded generation capabilities of the diffusion model in Sec. 3.3 and introduce the details of goal image conditioned foundation model for action planning in Sec. 3.4. + +# 3.1. Pipeline Overview + +The overall pipeline of CoTDiffusion is presented in Fig. 2, decoupling the visual planning and action planning. Given the initial observation $x_0$ and a multi-modal prompt $\mathcal{P}$ as task instruction potentially needs to be reached by $N$ subgoal steps, robots are required to learn a policy conditioned on the prompt and accomplish the specified long-horizon task. Our model is capable of understanding these general instructions and leverages diffusion model implicitly to decompose them into subgoal images chain $\tau_x = \{x_i\}_{i=1}^N$ . The precisely generated goal images $x_i$ then guide the foundation model's planning for inferring the action sequences $\tau_a^i = \{a_{i,t}\}_{t=1}^T$ with the subgoal horizon length $T$ . Let $p_\Theta$ model this hierarchical decision-making framework. With bi-level modules, $p_\Theta$ can be factorized into visual planning $p_\phi$ , and action planning $p_\psi$ . Under the Markovian assumption, the overall framework can be formulated as: + +$$ +p _ {\Theta} \left(\left\{\tau_ {a} ^ {i} \right\} _ {i = 1} ^ {N} \mid \mathcal {P}, x _ {0}\right) = \underbrace {\left(\prod_ {i = 1} ^ {N} p _ {\phi} \left(x _ {i} \mid \mathcal {P} , x _ {0}\right)\right)} _ {\text {v i s u a l p l a n n i n g}} \underbrace {\left(\prod_ {i = 1} ^ {N} \prod_ {t = 1} ^ {T} p _ {\psi} \left(a _ {i , t} \mid x _ {i}\right)\right)} _ {\text {a c t i o n p l a n n i n g}} \tag {1} +$$ + +The visual planning module consists of a multi-modal encoder to understand the complex task instructions and conditioned diffusion model which implicitly decomposes the task into progressive chained subgoals images. The standardized subgoal images serve as a unified interface bridging the visual planning and action planning modules. With the decoupled foundation model for action prediction, high-level visual planning focuses solely on comprehending instructions and visualizing subgoals without confusion from joint action prediction training. Conversely, the visual planning module frees the action planning module from ambiguous prompt parsing and long-horizon planning by providing coherent keyframes as visual landmarks, further reducing the capabilities requirements of the foundation model to basic single object manipulation skills. + +![](images/b1d80ffbef7691a905dac1c6b133e9089580b366c4dc70b16ac5f1359107611b.jpg) +Figure 2. Method overview: CoTDiffusion consists of a multi-modal encoder and vision encoder $\mathcal{V}$ , semantic alignment module $\mathcal{S}$ , conditional diffusion model $\mathcal{E}$ and foundation model $\mathcal{F}$ for action planning. The prompt and observation tokens are combined and fed into the semantic alignment module to identify the current reasoning chain step, providing progressive guidance for the diffusion model to generate the next subgoal image. The generated keyframes are further fed to the foundation model which predicts action sequences to achieve the imagined goal scene and this recursive process repeats in a receding horizon control loop until the task is finished. + +# 3.2. Pre-training Coarse Semantic Alignment + +For the challenge of explicit decomposing step-wise sub-prompts from multi-modal prompts, the diffusion model struggles to generate sequentially coherent keyframes that incrementally advance the prompt instructions and execution progress. The diffusion model fails to identify the stage of current generated keyframes onto the whole chain-of-thought progress and the generated sub-goals images may exhibit repetitions, skipping, or even backtracking to visited subgoals. To alleviate this dilemma, we design a semantic alignment module to empower diffusion model to track the progress of generated keyframes into the entire chain. + +Triple Alignment Architecture Since explicit decomposition for the multi-modal prompt is intractable, the initial observation $x_0$ and prompt $\mathcal{P}$ remain fixed across the progressive generation of diffusion model. Thus, tracking the progress critically relies on extracting semantic information from the generated subgoal $x_i$ . To this end, we design a triple alignment module denoted as $S(x_0, x_i, \mathcal{P})$ to compare current generated subgoal $x_i$ against initial observation $x_0$ and prompts $\mathcal{P}$ , serving as an information bottleneck to capture the key semantic cues about the subgoal-prompt matching to locate the current chain stage and deliver to the diffusion model for next subgoal generation. Specifically, we first capture visual representation from both the initial and current observation using a shared vision encoder $\mathcal{V}$ and concatenate with prompt tokens $\mathcal{P}$ . Then they are refined through fusion module which consists of several self-attention blocks separately to obtain attention tokens $z_0$ and $z_i$ aligned to the prompts. The cross-attention between $z_0$ and $z_i$ aggregates a contrast representation $z_{\mathrm{contrast}}^i$ and then concatenated with prompt tokens into another fusion module to capture the progress tokens $z_{\mathrm{progress}}^i$ , including + +rich context about the advancement and completion of the prompted task between the two subgoals. Finally, $z_{\text{progress}}^i$ and the current visual tokens $x_i$ and then fed into another cross-attention pass, which visual tokens $x_i$ serve as keys and progress tokens $z_{\text{progress}}^i$ as queries to infer the final aligned tokens $z_{\text{align}}^i$ . The aligned token encapsulates cues about 'which subgoal has reached now' and 'which subgoal should reach next', further injected into the diffusion model as precise semantic guidance to steer the generation of next subgoal. Through multi-stage cross-frame attention mechanisms, the triple alignment module can capture the cross-modal semantic correlation and achieve dynamic grounding of generated subgoal image into the multi-modal prompts to identify progress along the CoT reasoning chain. + +Masked Patch Prediction It is not trivial to directly integrate the designed triple alignment module without pretraining into the diffusion model for joint training from scratch. Simultaneously optimizing semantic alignment and pixel-level reconstruction such as texture, color, shape, and spatial transformations overwhelms joint learning. Thus, we propose a two-stage coarse-to-fine approach decoupling semantic alignment pretraining from diffusion model finetuning, illustrated in Fig. 3. First, we pre-train coarse semantic alignment focused only on spatial correlations, without pixel generation. Considering the adjacent keyframe contrasts reveal object manipulations as mask residuals patch $\hat{m}_i = x_{i + 1} - x_i$ , we can utilize these cross-frame spatial semantics to provide coarse alignment supervision. Regions with no visual changes likely correspond to stationary background elements or objects not involved in the current subtask operation. Contrastively, masked areas indicate spatial layout modifications concluding the rich spatial semantic information between keyframes. The sequence + +![](images/01d450c565851e62fe9c738834e999cbcd32251a5114e7b78c416b654f995163.jpg) +Figure 3. Two phase of coarse-to-fine alignment pipeline. + +of mask residuals patch $\{\hat{m}_i\}_{i=1}^{N-1}$ can be viewed as implicitly decomposing the prompt into spatial manipulation steps across visual frames that can be used for alignment. So we fed the aligned tokens into a mask predictor $\mathcal{D}$ to decode the mask region and set modeling mask prediction as a pretext task. Mask prediction encourages the model to infer these salient spatial semantics and changes to incrementally track prompt completion and focus regions. The process can be formulated as: $\{m_i\}_{i=1}^N = \sum_{i=1}^N \mathcal{D}(m_i | z_{\mathrm{align}}^i) S(z_{\mathrm{align}}^i | x_0, x_i, \mathcal{P})$ . + +By factorizing the alignment and generation objectives, the alignment module can develop basic visual reasoning skills through patch-level mask prediction before integrated into the diffusion model for controllable image synthesis. + +# 3.3. Fine-grained Diffusion Training + +In the second stage, we integrate the pretrained coarse alignment module into the diffusion model training for fine-grained image generation. To further enhance instruction following, we design a bi-directional generation that reconstructs the current frame from the aligned token, providing an additional learning constraint. We also employ a frame concatenation mechanism to enhance the consistency across generated keyframes and the fidelity of generation. + +Bi-directional Alignment Generation Building upon coarse pretraining, we integrate the alignment module into the diffusion model for fine-grained training with the pixel-level supervision of ground truth keyframes, illustrated in Fig. 3. Here we propose bi-directional aligned generation, where the aligned token $z_{\mathrm{align}}^i$ not only guides forward prediction but also reconstructs the current frame through backward passing. In the forward pass, based on current subgoal $x_i$ , $z_{\mathrm{align}}^i$ provides semantic conditioning to generate the next subgoal image $x_{i+1}$ . In the backward pass, the same aligned token can guide diffusion model to reconstruct the current subgoal $x_i$ from the next subgoal image $x_{i+1}$ , acting as an additional constraint. The training objective can be formulated as: + +$$ +L = \mathbb {E} _ {x _ {i} \in D} [ \underbrace {\| \hat {x} _ {i} - \mathcal {E} (x _ {i - 1} , z _ {\text {a l i g n}} ^ {i} , \mathcal {P})} _ {\text {F o r w a r d G e n e r a t i o n}} \| + \underbrace {\| \hat {x} _ {i - 1} - \mathcal {E} (x _ {i} , z _ {\text {a l i g n}} ^ {i} , \mathcal {P})} _ {\text {B a c k w a r d G e n e r a t i o n}} \| ] \tag {2} +$$ + +Bi-directional generation compels tighter instruction grounding, as accuracy is required for both progressive and reverse keyframe prediction. This dual-direction linkage enhances visual coherence and sequence controllability compared to forward-only training, improving the chain-of-thought capabilities of diffusion model. + +Frame Concat Mechanism Generating consistent and coherent subgoal images requires considering both semantic guidance to match prompts and visual grounding into observations. The aligned tokens $z_{\mathrm{align}}^i$ and image tokens $x_i$ are both fused into the denoising process in a classifier-free manner, as shown in Eq. (2). However, merely injecting current observation tokens as condition into noise model struggles to provide sufficient guidance to prevent distortions conflicting with the current observation, which is critical for precisely grounded robotics manipulations. Inspired by VDT[28], we employ frame concat mechanism to directly take current observation frame as the component of input rather than only condition for the next subgoal image generation, illustrated in Fig. 2. The current observation frame concatenated with the standard Gaussian noise gives a strong visual continuity prior, then fed into diffusion model to denoise altogether. For the training, we split the corresponding frame from the denoising frame and supervised it to the ground truth subgoal image. The rich visual context from the concatenated current frame enables coherent denoising in diffusion model further guarantees consistency with grounded observation and enforces coherence and smoothness across chain-of-thought visual planning. + +# 3.4. Goal-conditioned Policy Model + +The final component in our framework is the low-level policy model for action planning, generating an action trajectory $\tau_{a}^{i}$ when given observation trajectory $\tau_{x}^{i}$ from visual planning. The policy model can be parameterized as an image-conditioned planner that infers the action $a_{i,t}$ given the current observation $x_{i,t}$ and the generated subgoal image $g_{i}$ : $\tau_{a}^{i} = \{a_{i,t}\}_{t=1}^{T} \sim \prod_{t=1}^{T} p_{\psi}(a_{i,t}|x_{i,t}, g_{i})$ . The policy model is trained by imitation learning in an end-to-end manner, using paired subgoal images and expert trajectories. Thanks to the explicit subgoal generation from high-level visual planner, the low-level policy model does not need to master complex multi-step manipulation skills in long-horizon. Therefore, we simplify the policy model architecture and implement it as a simple MLP to only possess basic single-object manipulation primitives, which greatly reduces the burden of policy model training. Although we do not seek improvement through a more powerful policy model in this paper, CoTDiffusion is flexible to incorporate any goal-conditioned design like transformers or diffusion policies. For more details of low-level policy model, please refer to Appendix E.3. + +![](images/611df21c38c8862e3cd76d06c27075011ff9873800b5b0b7a14ff0d38ec12229.jpg) +Figure 4. Visualization of CoTDiffusion in three typical long-horizon tasks with multi-modal prompts in VIMA-BENCH. + +# 4. Experiments + +# 4.1. Experiment Setup + +Benchmark & Tasks We conduct evaluation on VIMABENCH, a benchmark suite for multimodal robot learning, which is built on the Ravens robot simulator [50]. VIMABENCH supports extensible collections of objects and textures to compose multi-modal prompts and to procedurally generate a large number of tasks. VIMA-BENCH can generate large quantities of expert trajectories via scripted oracle agents, and the ground truth keyframes that can leveraged for the training of CoTDiffusion. VIMA-BENCH contains various tasks ranging from simple object manipulation to multi-object manipulation and we select three representative long-horizon manipulation tasks in VIMABENCH-visual rearrangement, visual constraints, and visual reasoning, which represent different levels of complexity and requirements for embodied manipulation. These tasks represent different levels of complexity and requirements for embodied manipulation and can test the ability to understand spatial relationships and perform precise object manipulations. For more details regarding the simulation benchmark and tasks setting, please refer to Appendix A. + +Evaluation Metric The evaluation metrics we use in the experiments cover three aspects as following: + +- Image Fidelity - measures the quality and realism of the generated keyframe images. We choose Fréchet Inception Distance (FID) [15] as a useful metric. +- Instruction Following - evaluates how well the step-by-step keyframes cover the complex instructions. We use CLIP-based text-image similarity scores [14] between the prompt and keyframes to quantify instruction alignment. +- Task Completion - tests the end-to-end utility by executing the embodied manipulation task. The final success rate on long-horizon tasks can be a fair metric. + +# 4.2. Baselines + +There are several existing methods for constructing robot manipulation policies conditioned on complex prompts, which we use as baselines in our experiments: + +- Gato[41], a generalist agent that can solve tasks from multiple domains where tasks are specified by prompting the model with the observation and action subsequence. For fair comparisons, we provide the same conditioning manner with multi-modal prompts. +- Flamingo[2], a vision-language model that learns to generate textual completion in response to multimodal prompts. We borrow the architecture in [21] which adapts it to support decision-masking by replacing the output layer with robot action heads. +- VIMA[21], the first transformer-based generalist robot agent that processes manipulation tasks with multimodal prompts. VIMA adpots an object-centric approach to flatten all the observation and prompts into object tokens sequence and predicts motor actions autoregressively and demonstrates SOTA performance on VIMA-BENCH. +- SuSIE[4], the concurrent work for visual planning in manipulation tasks on the CALVIN [31], incorporating pretrained image-editing model and the low-level goal-conditioned policy. For fair comparison, we finetune the model with the same oracle data with CoTDiffusion. + +# 4.3. Quantitative Results of Success Rate + +We begin by comparing the performance of CoTDiffusion and baselines to solve long-horizon tasks in visual rearrange, visual reasoning, and visual constraints. The performance comparisons shown in Tab. 1 demonstrate CoT-Diffusion significantly outperforms other baselines in success rate. The baselines can be divided into two kinds of planners, including abstract planner and visual planner. Abstract planner like Gato, Flamingo and VIMA directly + +map general prompts to subsequent actions in an end-to-end manner. Gato and Flamingo gets low success rates on long-horizon tasks without explicit subgoal generation to correct the accumulative deviation errors from the instructions. In contrast, Visual planner like SuSIE and CoTDiffusion can generate intermediate goal images to guide the action planning, which can enhance instruction following for long-horizon tasks via visual planning. However, SuSIE lacks intrinsic chain-of-thought reasoning capabilities, struggling to generate logically progressing subgoal sequences as coherent plans from general prompts. To enable comparison, we grant SuSIE the manually decomposed prompts into privileged step-wise sub-prompts which are unavailable in fair settings, denoted as 'SuSIE + sub-prompts'. Nonetheless, its performance still lags CoTDiffusion, which can implicitly align generated images with multi-modal prompts through learned correspondence between visual signals and language semantics. One more potential reason to restrict the performance of SuSIE is that provided sub-prompts are challenging to perfectly encapsulate complex instructions. In contrast, CoTDiffusion develops intrinsic chain-of-thought reasoning and alignment for generated subgoal images for flexible visual planning directly from the raw multi-modal prompts, without the need for pre-decomposed sub-prompts. The comparisons highlight the importance of chain-of-thought capabilities in visual planning, especially the role of implicit semantic alignment for instruction following long-horizon manipulation tasks. + +
MethodologyRearrangeReasoningConstraintsOverall
Gato6.4 ± 1.32.5 ± 0.425.2 ± 3.111.4 ± 1.6
Flamingo17.5 ± 1.63.0 ± 0.536.1 ± 4.218.9 ± 2.1
VIMA43.1 ± 3.338.2 ± 4.467.2 ± 5.249.5 ± 4.3
SuSIE2.7 ± 0.33.1 ± 0.624.3 ± 5.910.0 ± 2.3
+sub-prompts37.7 ± 6.239.0 ± 4.552.3 ± 7.043.0 ± 5.9
CoTDiffusion59.0 ± 1.751.7 ± 2.683.1 ± 4.764.6 ± 3.0
+ +# 4.4. Further Analysis + +Robustness to Insufficient Perception Rich visual observations from diverse views are crucial for complex robot manipulation tasks. Restricted perspectives limit rich representations used for action planning, degrading performance on long-horizon multi-object manipulation. We evaluate different methods in single-view including the top and front views provided from VIMA-BENCH. As shown in Tab. 2, visual planning approach demonstrates greater robustness to limited observations, with a smaller performance drop in single-view compared to abstract methods. We attribute this to two potential reasons: First, accurate and grounded subgoal images generated in visual planners provide supplemental visual context, which can partly compensate for the insufficient perception to aid robustness under single- + +Table 1. The evaluations of success rates on three typical long-horizon tasks with multi-modal prompts. + +
MethodologyMulti-ViewSingle-ViewPerformance Drop
Gato11.4 ± 1.66.5 ± 1.64.9(42% ↓)
Flamingo18.9 ± 2.112.0 ± 2.46.9(36.2% ↓)
VIMA49.5 ± 4.334.9 ± 3.414.6(29.4% ↓)
SuSIE10.0 ± 2.37.9 ± 2.02.1(21.0% ↓)
+sub-prompts43.0 ± 5.935.6 ± 6.17.4(18.8% ↓)
CoTDiffusion64.6 ± 3.056.0 ± 2.48.6(13.2% ↓)
+ +Table 2. The evaluations of performance drop of different methods on single-view and multi-view from VIMA-BENCH. +Table 3. Quantitative comparisons of FID between methods on all three tasks, including visual rearrange, reasoning and constraints. + +
MethodologyRearrangeReasoningConstraints
SuSIE [4]19.718.323.2
- w/o finetune38.4 ↑33.0 ↑52.7 ↑
- w/o stepwise prompt32.6 ↑28.1 ↑34.7 ↑
CoTDiffusion (ours)11.310.88.6
- w/o coarse pre-train23.1 ↑16.2 ↑18.9 ↑
- w/o bi-direction align18.7 ↑15.0 ↑15.8 ↑
- w/o frame concat15.4 ↑13.7 ↑12.1 ↑
+ +view. Second, by providing coherent subgoal images as visual landmarks, the requirements for the low-level action planner are reduced to basic single-object manipulation primitives in short horizon, with less reliance on rich visual perceptions. The experiments demonstrate that CoTDiffusion enjoys better robustness to restricted perception than abstract planners, highlighting the benefits of hierarchical framework decoupled visual planning and action planning. + +Fidelity of Image Generation Here we conduct further analysis of visual planners by comparing goal image quality against ground truth keyframes with FID as evaluation metrics. As Tab. 3 shows, CoTDiffusion still achieves much better fidelity than SuSIE even though SuSIE has got improved after fine-tuning on the same datasets in VIMABENCH. Moreover, SuSIE can only understand short instructions of single-object manipulation tasks, struggling to handle long-horizon instructions due to the lack of chain-of-thought reasoning capabilities to handle multi-modal prompts. With step-wise sub-prompts decomposed in advance, the performance of SuSIE gets largely raised but still underperforms CoTDiffusion, which has no need to explicitly decompose the general prompts and can generate subgoal images in an implicit chain-of-thought manner. Additionally, ablating coarse pretraining and bi-directional generation degrades performance, validating their benefits. The coarse-to-fine semantic alignment training training allows developing spatial reasoning prior to synthesis. Frame concatenation further guides coherent denoising by providing rich context information as visual priors to ground the current observation and enhance the fidelity of generation. + +Accuracy of Instruction Following We evaluate instruction following accuracy via CLIP similarity between gen + +erated keyframes and general prompts, normalized by the CLIP score between ground truth ultimate goal image and prompts. The results in Fig. 5 demonstrate that CoT-Diffusion enjoys better instruction following capabilities. Without chain-of-thought reasoning abilities, SuSIE struggles to follow instructions when given general multi-modal prompts, let alone generate subgoal images with smooth progressions. Progressively providing privileged prompts step by step improves SuSIE in instruction following and gradual advancement towards ultimate goals. However, text prompts decomposed by rules are not always perfect to convey the original multi-modal prompt, so the reliance on these prompts restricts the performance of SuSIE. In contrast, the coarse alignment pretraining and bi-directional generation can assist the diffusion model in tracking the progress of generated keyframes throughout the entire chain and generate sequenced keyframes incrementally advancing prompt instructions. Moreover, the implicit decomposition maintains semantic coherence between subgoals without overly large jumps and the smoothness and continuity between visual milestones make them more reachable by downstream foundation models with just single-object manipulation capabilities. Additionally, we observe that the bi-directional generation may impedes the diffusion model training if without coarse semantic pretraining. As the initial observation and prompt remain fixed across different generation steps, the align tokens at various stage tend to so similar that may confuse simultaneous progressive generation and reconstruction training. + +Generalization across Tasks We evaluate the generalization ability in three levels with increasing difficulty: placement generalization which randomizes the novel placement of objects (L1), object generalization which provides the objects with novel attributes (L2), and task combinatorial generalization which complexes the prompts with extra novel instruction (L3). As Fig. 6 shows, CoTDiffusion exhibits strong zero-shot generalization on unfamiliar objects, colors, and shapes at different placements by explicitly visualizing them into coherent subgoals grounded on new concepts. This avoids re-grounding concepts from multi-modal prompts in the low-level foundation model for + +![](images/2323b9c507967d8d3a6d98f7b39ab2117fbcdf9b5a73208aca9be77b68fe387b.jpg) +Figure 5. The normalized CLIP scores for each generation step, reflecting the step-wise accuracy of instruction following. + +![](images/b2bc2a91b3604bd403e5ea396ca0a9ddd8a3fc4f8af91b2eabfbec501460e426.jpg) +Figure 6. The evaluation on various generalization levels. + +the placement and object. For L3 generalization, taking an example, we complex the prompt such as 'rearrange ... then rotate/twist/stack ...' to demand extra tasks such as rotating, twisting degrees, or stacking objects. CoTDiffusion achieves outstanding gain in the zero-shot performance of combinatorial tasks. When CoTDiffusion visualizes novel concepts into goal images, the foundation model can still accomplish the stack by simply achieving the provided subgoal, with no need to inherently understand novel skills like stack. This demonstrates the power of 'a image is worth a thousand words' - the subgoal images facilitate the generalization of the foundation model to unseen objects and novel combinatorial task while the foundation model simply achieves them with reuse of simple known primitives. + +# 5. Conclusion + +We presented CoTDiffusion, a hierarchical framework that integrates diffusion model as high-level module to translate the general multi-modal prompts into coherent subgoal images, serves as the visual milestones to anchor the low-level foundation model to plan action sequences, termed as 'generate subgoal images before act'. With the coarse-to-fine training for semantic alignment module, CoTDiffusion can identify the progress of generated subgoals images along reasoning chains, unlocking the chain-of-thought reasoning capabilities of diffusion model for long-horizon manipulation tasks. The experiments cover various long-horizon manipulation scenarios in VIMA-BENCH, and CoTDiffusion show the strong instruction following and outstanding performance gain compared to existed methods without visual planning. Incorporating commonsense knowledge from pre-trained MLLM like GPT-4V provides an avenue for more generalizable and promising reasoning in CoTDiffusion, which leaves us our future work. + +# 6. Acknowledgement + +This work is supported by the National Natural Science Foundation of China (Grant Nos. 92370132) and the Xiaomi Young Talents Program of Xiaomi Foundation. + +# References + +[1] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, et al. Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691, 2022. 2, 3 +[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: A visual language model for few-shot learning. NeurIPS, 2022. 6, 5 +[3] Fan Bao, Shen Nie, Kaiwen Xue, Chongxuan Li, Shi Pu, Yaole Wang, Gang Yue, Yue Cao, Hang Su, and Jun Zhu. One transformer fits all distributions in multi-modal diffusion at scale. arXiv preprint arXiv:2303.06555, 2023. 2 +[4] Kevin Black, Mitsuhiko Nakamoto, Pranav Atreya, Homer Walke, Chelsea Finn, Aviral Kumar, and Sergey Levine. Zero-shot robotic manipulation with pretrained image-editing diffusion models. arXiv preprint arXiv:2310.10639, 2023. 6, 7, 5 +[5] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817, 2022. 1 +[6] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. arXiv preprint arXiv:2211.09800, 2022. 5 +[7] Cheng Chi, Siyuan Feng, Yilun Du, Zhenjia Xu, Eric Cousineau, Benjamin Burchfiel, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. arXiv preprint arXiv:2303.04137. 8 +[8] Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168, 2021. 3 +[9] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In Advances in Neural Information Processing Systems, pages 8780-8794. Curran Associates, Inc., 2021. 2 +[10] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palm- e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 1, 3 +[11] Jiafei Duan, Samson Yu, Hui Li Tan, Hongyuan Zhu, and Cheston Tan. A survey of embodied ai: From simulators to research tasks. IEEE Transactions on Emerging Topics in Computational Intelligence, 6(2):230-244, 2022. 1 +[12] Theophile Gervet, Zhou Xian, Nikolaos Gkanatsios, and Katerina Fragkiadaki. Act3d: Infinite resolution action detection transformer for robotic manipulation. arXiv preprint arXiv:2306.17817, 2023. 8 +[13] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 4 + +[14] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 6 +[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6 +[16] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 6 +[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2 +[18] Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. Voxposer: Composable 3d value maps for robotic manipulation with language models. arXiv preprint arXiv:2307.05973, 2023. 1 +[19] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General perception with iterative attention. In International conference on machine learning, pages 4651-4664. PMLR, 2021. 5 +[20] Stephen James, Zicong Ma, David Rovick Arrojo, and Andrew J Davison. Rlbench: The robot learning benchmark & learning environment. IEEE Robotics and Automation Letters, 5(2):3019-3026, 2020. 8 +[21] Yunfan Jiang, Agrim Gupta, Zichen Zhang, Guanzhi Wang, Yongqiang Dou, Yanjun Chen, Li Fei-Fei, Anima Anandkumar, Yuke Zhu, and Linxi Fan. Vima: General robot manipulation with multimodal prompts. arXiv preprint arXiv:2210.03094, 2022. 1, 2, 3, 6, 5 +[22] Dongxu Li, Junnan Li, and Steven CH Hoi. Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. arXiv preprint arXiv:2305.14720, 2023. 1 +[23] Zhengzhong Liang, Steven Bethard, and Mihai Surdeanu. Explainable multi-hop verbal reasoning through internal monologue. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, 2021. 3 +[24] Wang Ling, Dani Yogatama, Chris Dyer, and Phil Blunson. Program induction by rationale generation: Learning to solve and explain algebraic word problems. arXiv preprint arXiv:1705.04146, 2017. 3 +[25] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 2 +[26] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6 +[27] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 12 +[28] Haoyu Lu, Guoxing Yang, Nanyi Fei, Yuqi Huo, Zhiwu Lu, Ping Luo, and Mingyu Ding. Vdt: An empirical study on video diffusion with transformers. arXiv preprint arXiv:2305.13311, 2023. 5 + +[29] Yecheng Jason Ma, Shagun Sodhani, Dinesh Jayaraman, Obert Bastani, Vikash Kumar, and Amy Zhang. Vip: Towards universal visual reward and representation via value-implicit pre-training. arXiv preprint arXiv:2210.00030, 2022. 3 +[30] Yecheng Jason Ma, William Liang, Vaidehi Som, Vikash Kumar, Amy Zhang, Osbert Bastani, and Dinesh Jayaraman. Liv: Language-image representations and rewards for robotic control. arXiv preprint arXiv:2306.00958, 2023. 3 +[31] Oier Mees, Lukas Hermann, Erick Rosete-Beas, and Wolfram Burgard. Calvin: A benchmark for language-conditioned policy learning for long-horizon robot manipulation tasks. IEEE Robotics and Automation Letters (RA-L), 7(3):7327-7334, 2022. 6, 5 +[32] Oier Mees, Lukas Hermann, Erick Rosete-Beas, and Wolfram Burgard. Calvin: A benchmark for language-conditioned policy learning for long-horizon robot manipulation tasks. IEEE Robotics and Automation Letters, 7(3): 7327-7334, 2022. 8, 11 +[33] Yao Mu, Qinglong Zhang, Mengkang Hu, Wenhai Wang, Mingyu Ding, Jun Jin, Bin Wang, Jifeng Dai, Yu Qiao, and Ping Luo. Embodiedgpt: Vision-language pretraining via embodied chain of thought. arXiv preprint arXiv:2305.15021, 2023. 2, 3 +[34] Suraj Nair, Aravind Rajeswaran, Vikash Kumar, Chelsea Finn, and Abhinav Gupta. R3m: A universal visual representation for robot manipulation. arXiv preprint arXiv:2203.12601, 2022. 3 +[35] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, pages 16784-16804. PMLR, 2022. 2 +[36] Maxwell Nye, Anders Johan Andreassen, Guy Gur-Ari, Henryk Michalewski, Jacob Austin, David Bieber, David Dohan, Aitor Lewkowycz, Maarten Bosma, David Luan, et al. Show your work: Scratchpads for intermediate computation with language models. arXiv preprint arXiv:2112.00114, 2021. 3 +[37] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35: 27730-27744, 2022. 2 +[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1 +[39] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2 +[40] Scott Reed and Nando De Freitas. Neural programmer-interpreters. arXiv preprint arXiv:1511.06279, 2015. 3 + +[41] Scott Reed, Konrad Zolna, Emilio Parisotto, Sergio Gomez Colmenarejo, Alexander Novikov, Gabriel Barth-Maron, Mai Gimenez, Yury Sulsky, Jackie Kay, Jost Tobias Springenberg, et al. A generalist agent. arXiv preprint arXiv:2205.06175, 2022. 3, 6, 5 +[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2 +[43] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 6 +[44] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2 +[45] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 2 +[46] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 6, 12 +[47] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, 2017. 5 +[48] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Ed Chi, Quoc Le, and Denny Zhou. Chain of thought prompting elicits reasoning in large language models. arXiv preprint arXiv:2201.11903, 2022. 3 +[49] Xingqian Xu, Zhangyang Wang, Gong Zhang, Kai Wang, and Humphrey Shi. Versatile diffusion: Text, images and variations all in one diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7754-7765, 2023. 2 +[50] Andy Zeng, Pete Florence, Jonathan Tompson, Stefan Welker, Jonathan Chien, Maria Attarian, Travis Armstrong, Ivan Krasin, Dan Duong, Vikas Sindhwani, et al. Transporter networks: Rearranging the visual world for robotic manipulation. In Conference on Robot Learning, pages 726-747. PMLR, 2021. 6, 1 +[51] Brianna Zitkovich, Tianhe Yu, Sichun Xu, Peng Xu, Ted Xiao, Fei Xia, Jialin Wu, Paul Wohlhart, Stefan Welker, Ayzaan Wahid, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. In 7th Annual Conference on Robot Learning, 2023. 1, 2, 3 \ No newline at end of file diff --git a/2024/Generate Subgoal Images before Act_ Unlocking the Chain-of-Thought Reasoning in Diffusion Model for Robot Manipulation with Multimodal Prompts/images.zip b/2024/Generate Subgoal Images before Act_ Unlocking the Chain-of-Thought Reasoning in Diffusion Model for Robot Manipulation with Multimodal Prompts/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..86ad8fb6e0d0749dcf472ec9d790ff93a1863a45 --- /dev/null +++ b/2024/Generate Subgoal Images before Act_ Unlocking the Chain-of-Thought Reasoning in Diffusion Model for Robot Manipulation with Multimodal Prompts/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6fd9d10b1b6f5033f71556222bad83d74d5a7b0f2d61b30e70e08a5749c34281 +size 385580 diff --git a/2024/Generate Subgoal Images before Act_ Unlocking the Chain-of-Thought Reasoning in Diffusion Model for Robot Manipulation with Multimodal Prompts/layout.json b/2024/Generate Subgoal Images before Act_ Unlocking the Chain-of-Thought Reasoning in Diffusion Model for Robot Manipulation with Multimodal Prompts/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..766eb4ae294db0e34c80f337a4099af32eecce55 --- /dev/null +++ b/2024/Generate Subgoal Images before Act_ Unlocking the Chain-of-Thought Reasoning in Diffusion Model for Robot Manipulation with Multimodal Prompts/layout.json @@ -0,0 +1,6748 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 48, + 103, + 547, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 103, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 103, + 547, + 140 + ], + "type": "text", + "content": "Generate Subgoal Images before Act: Unlocking the Chain-of-Thought Reasoning in Diffusion Model for Robot Manipulation with Multimodal Prompts" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 83, + 159, + 506, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 159, + 506, + 204 + ], + "spans": [ + { + "bbox": [ + 83, + 159, + 506, + 204 + ], + "type": "text", + "content": "Fei Ni" + }, + { + "bbox": [ + 83, + 159, + 506, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 83, + 159, + 506, + 204 + ], + "type": "text", + "content": " Jianye Hao" + }, + { + "bbox": [ + 83, + 159, + 506, + 204 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 83, + 159, + 506, + 204 + ], + "type": "text", + "content": " Shiguang Wu" + }, + { + "bbox": [ + 83, + 159, + 506, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 83, + 159, + 506, + 204 + ], + "type": "text", + "content": " Longxin Kou" + }, + { + "bbox": [ + 83, + 159, + 506, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 83, + 159, + 506, + 204 + ], + "type": "text", + "content": " Jiashun Liu" + }, + { + "bbox": [ + 83, + 159, + 506, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 83, + 159, + 506, + 204 + ], + "type": "text", + "content": " Yan Zheng" + }, + { + "bbox": [ + 83, + 159, + 506, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 83, + 159, + 506, + 204 + ], + "type": "text", + "content": " Bin Wang" + }, + { + "bbox": [ + 83, + 159, + 506, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 83, + 159, + 506, + 204 + ], + "type": "text", + "content": " Yuzheng Zhuang" + }, + { + "bbox": [ + 83, + 159, + 506, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 83, + 159, + 506, + 204 + ], + "type": "text", + "content": " Tianjin University, China " + }, + { + "bbox": [ + 83, + 159, + 506, + 204 + ], + "type": "inline_equation", + "content": "{ }^{2} \\mathrm{Huawei}" + }, + { + "bbox": [ + 83, + 159, + 506, + 204 + ], + "type": "text", + "content": " Noah's Ark Lab, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 256, + 290, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 256, + 290, + 605 + ], + "spans": [ + { + "bbox": [ + 45, + 256, + 290, + 605 + ], + "type": "text", + "content": "Robotics agents often struggle to understand and follow the multi-modal prompts in complex manipulation scenes which are challenging to be sufficiently and accurately described by text alone. Moreover, for long-horizon manipulation tasks, the deviation from general instruction tends to accumulate if lack of intermediate guidance from high-level subgoals. For this, we consider can we generate subgoal images before act to enhance the instruction following in long-horizon manipulation with multi-modal prompts? Inspired by the great success of diffusion model in image generation tasks, we propose a novel hierarchical framework named as CoTDiffusion that incorporates diffusion model as a high-level planner to convert the general and multimodal prompts into coherent visual subgoal plans, which further guide the low-level policy model before action execution. We design a semantic alignment module that can anchor the progress of generated keyframes along a coherent generation chain, unlocking the chain-of-thought reasoning ability of diffusion model. Additionally, we propose bi-directional generation and frame concat mechanism to further enhance the fidelity of generated subgoal images and the accuracy of instruction following. The experiments cover various robotics manipulation scenarios including visual reasoning, visual rearrange, and visual constraints. CoTDiffusion achieves outstanding performance gain compared to the baselines without explicit subgoal generation, which proves that a subgoal image is worth a thousand words of instruction. The details and visualizations are available at https://cotdiffusion.github.io." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 626, + 128, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 626, + 128, + 639 + ], + "spans": [ + { + "bbox": [ + 47, + 626, + 128, + 639 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 646, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 646, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 646, + 287, + 696 + ], + "type": "text", + "content": "Embodied manipulation focuses on creating generalists that can perceive, reason, and act within complex environments, which sits at the intersection of robotics control, computer vision, and natural language processing [11]. Recent ad" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 230, + 545, + 407 + ], + "blocks": [ + { + "bbox": [ + 307, + 230, + 545, + 407 + ], + "lines": [ + { + "bbox": [ + 307, + 230, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 545, + 407 + ], + "type": "image", + "image_path": "80efe11818f03782148d8dabaaef5b81279ec6b189539900031851b55d44a916.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 410, + 547, + 454 + ], + "lines": [ + { + "bbox": [ + 305, + 410, + 547, + 454 + ], + "spans": [ + { + "bbox": [ + 305, + 410, + 547, + 454 + ], + "type": "text", + "content": "Figure 1. A motivation example of robotics manipulation tasks in multi-modal instructions. The subgoal images are worth a thousand words, inspiring us to propose a novel framework CoTDiffusion to generate goal images step-by-step before act." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 460, + 546, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 546, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 546, + 639 + ], + "type": "text", + "content": "vances in vision-language model (VLM) such as CLIP [38], BLIP [22], enable open-vocabulary visual recognition and show promising potential of robotics closed-loop control by endowing them with the ability to predict robot actions in embodied robotics [5, 10, 18, 51]. However, robotics agents still face significant challenges in following instructions for long-horizon manipulation tasks, especially when the given general instructions are not progressive step-wise prompts, but implicitly contain several subtasks to accomplish. Moreover, complex manipulation scenarios with rich visual contexts are often challenging to be sufficiently and accurately described through text-only prompts, requiring multi-modal prompts to convey the instructions and intentions accurately [21], further increasing the difficulty of instruction following and task completion." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "content": "For example in Fig. 1, now given a short multi-modal prompt, the robot arm needs to follow the instructions and accomplish the task. The multi-modal instruction is brief but implicitly encapsulates long-horizon multi-step manipulation steps. We have two pathways to solve this task according to the recent embodied manipulation works. The" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 15 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 15 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 15 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "spans": [ + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "spans": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 702, + 212, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 212, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 212, + 713 + ], + "type": "text", + "content": "*Corresponding author: jianye.hao@tju.edu.cn" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13991" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 324 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 324 + ], + "type": "text", + "content": "first paradigm is directly leveraging the powerful VLM to encode prompts and observations into tokens and predict the entire action sequence with the subsequent decoder in an end-to-end manner, like RT-2 [51] or VIMA [21]. However, the compounding small errors over long horizons will lead to catastrophic deviations from the original task instructions due to the lack of intermediate guidance from coherent subgoals. The second paradigm is to incorporate the large language models (LLMs) to explicitly decompose the general prompts into several orderly subgoals with step-wise text instructions, like SayCan [1]. Though LLMs like ChatGPT [37] struggle to support the multi-modal prompts beyond text-only modality, we can seek help from the latest multi-modal LLMs such as LLaVA [25], EmbodiedGPT [33]. Even if we idealistically assume (M)LLM-based decomposition produces well-grounded and rational plans, generated text-only prompts tend to be extremely redundant and complex to accurately specify each subgoal in complex manipulation scene, may even be more confusing than the original general prompt and impose a heavy burden on the robot agents to re-parse and re-understand again." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 327, + 289, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 327, + 289, + 674 + ], + "spans": [ + { + "bbox": [ + 47, + 327, + 289, + 674 + ], + "type": "text", + "content": "Considering the phrase 'a picture is worth a thousand words', subgoal images could provide higher expressive capabilities for conveying subtasks compared to sub-prompts with complex text. This motivates us to generate subgoal images step-by-step before act to enhance long-horizon instruction following. Fortunately, tremendous success in text-to-image generation sheds light on the possibility of directly translating multi-modal prompts into coherent grounded visual subgoals step by step. Our key insight is to propose a hierarchical framework CoTDiffusion that integrates diffusion model as the high-level visual planner that can understand multi-modal prompts and progressively generate chained subgoal images based on the current visual observations in a chain-of-thought manner. The chained subgoal images serve as visual milestones to anchor the execution of task, further fed into a low-level foundation model for the specific step-wise action planning to achieve the provided goal scene. The subgoal images act as a unified interface bridging the high-level visual planning and low-level action planning, decoupling the instruction understanding and action execution. The foundation module allows the high-level diffusion model to focus solely on instruction understanding and subgoal visualization without overwhelming joint action prediction training. Conversely, the high-level visual planner frees the low-level foundation model from ambiguous prompt understanding and long-horizon planning. By providing coherent subgoal images as visual landmarks, the requirements for the foundation model are reduced to basic single-object manipulation primitives." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "content": "The key challenge to enabling CoTDiffusion to progressively generate subgoal images in a chain-of-thought manner lies in tracking the generated subgoal's progress on task" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 547, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 289 + ], + "type": "text", + "content": "prompts. In other words, for coherent subgoal generation, diffusion model needs to know 'which step this subgoal image has reached' and 'which step needs to reach in the next subgoal image' based on task prompts. For this, we design a triple alignment module comparing the cross-frame visual contrasts between the generated subgoal image and initial observation, and align the progress back to prompts to anchor the stage of generation chains. Specifically, we use a coarse-to-fine training pipeline. First, the aligned module is coarsely pre-trained to predict residual mask patches between subgoal images for aligning spatial semantics, focusing on salient differences rather than pixel details in textures or colors. The semantic alignment module is then integrated into the diffusion model for step-wise image generation with fine-grained pixel reconstruction. Additionally, bi-directional generation and frame concatenation mechanism further enhance subgoal image fidelity and instruction following. The contributions of this work are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 290, + 545, + 471 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 306, + 290, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 290, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 306, + 290, + 545, + 361 + ], + "type": "text", + "content": "- We propose a hierarchical framework CoTDiffusion that the high-level diffusion model translates the multi-modal prompts into coherent subgoal images in a chain-of-thought manner as visual milestones to anchor the low-level foundation model executing, enhance the instruction following on long-horizon manipulation tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 363, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 363, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 306, + 363, + 545, + 422 + ], + "type": "text", + "content": "- We design a semantic alignment module to capture the semantic relevance between visual subgoals and prompt, progressively tracking the progress of generated images along the coherent generation chain to unlock the chain-of-thought reasoning capability for multi-modal prompt." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 423, + 545, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 423, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 306, + 423, + 545, + 471 + ], + "type": "text", + "content": "- The experiments on various long-horizon manipulation tasks empirically demonstrate that CoTDiffusion enjoys outstanding performance gain than prior methods by explicitly generating goal images step by step before act." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 483, + 392, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 483, + 392, + 495 + ], + "spans": [ + { + "bbox": [ + 306, + 483, + 392, + 495 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 502, + 545, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 502, + 545, + 515 + ], + "spans": [ + { + "bbox": [ + 306, + 502, + 545, + 515 + ], + "type": "text", + "content": "2.1. Diffusion Models for Text-to-Image Generation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "text", + "content": "Recently, diffusion models have emerged as a powerful paradigm for high-fidelity text-to-image synthesis [9, 17, 35, 42, 45]. Models such as DALL-E 2 [39], Imagen [44] and Stable Diffusion [43] have demonstrated impressive success in generating realistic images from textual descriptions. The text instructions are tokenized as specific conditions and injected into the denoising process for controllable image generation. Moreover, some recent works have begun to explore image generation with multi-modal prompts, incorporating diverse inputs beyond just text. UniDiffuser [3] treats both image and text as sequential token streams for diffusion generations and Versatile diffusion [49] employs a multi-flow design to tackle multi-modal prompts. However, directly applying existing text-to-image approaches to complex robotics manipulation poses challenges in instruction understanding and following. Prompts" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13992" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "in manipulation tasks are often too high-level and general to effectively ground and reason, making the generated images struggle to accurately follow instructions and ground to the given scenarios, not sufficient to serve as goal images to offer fine-grained guidance for manipulation. Our work employs a semantic alignment module for multi-stage cross-modal alignment to achieve better grounding of generated images and unlock the chain-of-thought reasoning abilities of diffusion model for long-horizon tasks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 193, + 237, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 193, + 237, + 205 + ], + "spans": [ + { + "bbox": [ + 47, + 193, + 237, + 205 + ], + "type": "text", + "content": "2.2. Robotics Manipulation and Control" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 213, + 289, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 213, + 289, + 489 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 289, + 489 + ], + "type": "text", + "content": "With the development of VLMs for robotics, such as R3M [34], VIP [29], and LIV [30] pre-trained general visual representations for robotic perception, RT-2 [51], Gato [41] and PaLM-E [10] can convert instruction and observation into tokens and apply large transformers to predict action tokens in an end-to-end manner for robot manipulation tasks. While these works mostly focus on language-conditioned manipulation, more complex manipulation scenarios with multi-modal prompts introduced by VIMA [21] are also gaining increasing attention. Accurately instruction following becomes more challenging with general multimodal prompts, especially in long-horizon manipulation tasks. It is not trivial to directly borrow the recent works like SayCan [1] and EmbodiedGPT [33] to utilize MLLMs to decompose multi-modal instructions into step-wise text prompts. Even if we assume step-wise instructions can be rationally decomposed, they often require complex and redundant language to accurately specify subtasks and may be more confusing to understand. Inspired by the motto 'show, don't tell', our work aims to translate the general multi-modal prompts into coherent subgoal images instead of text step by step, serving as visual landmarks to enhance the instruction following in long horizon." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 502, + 207, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 502, + 207, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 502, + 207, + 514 + ], + "type": "text", + "content": "2.3. Chain-of-Thought Reasoning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 522, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 289, + 713 + ], + "type": "text", + "content": "Chain-of-Thought [48] reasoning refers to the general strategy of solving multi-step problems by decomposing them into a sequence of intermediate steps, showing great success in guiding the model to think step-by-step for enhancing instructing following and reasoning accuracy. It has recently been applied extensively in a variety of problems such as mathematical reasoning [8, 24], program execution [36, 40], commonsense or general reasoning [23, 48]. However, chain-of-thought visual reasoning has not been well explored for robotics manipulation. Inspired by prior explorations into CoT reasoning, we introduce the novel concept of chaining conditional image generations step by step to guide long-horizon manipulation. Our work employs a semantic alignment module to anchor the generated subgoal image in the entire reasoning chain, unlocking the chain-of-thought reasoning capabilities for the diffusion model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 71, + 361, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 361, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 361, + 83 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 91, + 547, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 547, + 248 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 547, + 248 + ], + "type": "text", + "content": "We propose CoTDiffusion, a hierarchical framework that integrates the diffusion model as the high-level module to decompose multi-modal prompts in a chain-of-thought manner and progressively generate chained subgoals images step by step to guide the underlying foundation model for long-horizon tasks. Specifically, we first discuss the overview pipeline of CoTDiffusion in Sec. 3.1, and then introduce the pretraining of the coarse semantic alignment to capture the step-wise spatial information in Sec. 3.2. Then we discuss how to further fine-grain the CoT reasoning and grounded generation capabilities of the diffusion model in Sec. 3.3 and introduce the details of goal image conditioned foundation model for action planning in Sec. 3.4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 258, + 413, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 258, + 413, + 271 + ], + "spans": [ + { + "bbox": [ + 305, + 258, + 413, + 271 + ], + "type": "text", + "content": "3.1. Pipeline Overview" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 277, + 547, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 547, + 469 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 547, + 469 + ], + "type": "text", + "content": "The overall pipeline of CoTDiffusion is presented in Fig. 2, decoupling the visual planning and action planning. Given the initial observation " + }, + { + "bbox": [ + 304, + 277, + 547, + 469 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 304, + 277, + 547, + 469 + ], + "type": "text", + "content": " and a multi-modal prompt " + }, + { + "bbox": [ + 304, + 277, + 547, + 469 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 304, + 277, + 547, + 469 + ], + "type": "text", + "content": " as task instruction potentially needs to be reached by " + }, + { + "bbox": [ + 304, + 277, + 547, + 469 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 277, + 547, + 469 + ], + "type": "text", + "content": " subgoal steps, robots are required to learn a policy conditioned on the prompt and accomplish the specified long-horizon task. Our model is capable of understanding these general instructions and leverages diffusion model implicitly to decompose them into subgoal images chain " + }, + { + "bbox": [ + 304, + 277, + 547, + 469 + ], + "type": "inline_equation", + "content": "\\tau_x = \\{x_i\\}_{i=1}^N" + }, + { + "bbox": [ + 304, + 277, + 547, + 469 + ], + "type": "text", + "content": ". The precisely generated goal images " + }, + { + "bbox": [ + 304, + 277, + 547, + 469 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 304, + 277, + 547, + 469 + ], + "type": "text", + "content": " then guide the foundation model's planning for inferring the action sequences " + }, + { + "bbox": [ + 304, + 277, + 547, + 469 + ], + "type": "inline_equation", + "content": "\\tau_a^i = \\{a_{i,t}\\}_{t=1}^T" + }, + { + "bbox": [ + 304, + 277, + 547, + 469 + ], + "type": "text", + "content": " with the subgoal horizon length " + }, + { + "bbox": [ + 304, + 277, + 547, + 469 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 277, + 547, + 469 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 304, + 277, + 547, + 469 + ], + "type": "inline_equation", + "content": "p_\\Theta" + }, + { + "bbox": [ + 304, + 277, + 547, + 469 + ], + "type": "text", + "content": " model this hierarchical decision-making framework. With bi-level modules, " + }, + { + "bbox": [ + 304, + 277, + 547, + 469 + ], + "type": "inline_equation", + "content": "p_\\Theta" + }, + { + "bbox": [ + 304, + 277, + 547, + 469 + ], + "type": "text", + "content": " can be factorized into visual planning " + }, + { + "bbox": [ + 304, + 277, + 547, + 469 + ], + "type": "inline_equation", + "content": "p_\\phi" + }, + { + "bbox": [ + 304, + 277, + 547, + 469 + ], + "type": "text", + "content": ", and action planning " + }, + { + "bbox": [ + 304, + 277, + 547, + 469 + ], + "type": "inline_equation", + "content": "p_\\psi" + }, + { + "bbox": [ + 304, + 277, + 547, + 469 + ], + "type": "text", + "content": ". Under the Markovian assumption, the overall framework can be formulated as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 479, + 555, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 479, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 305, + 479, + 555, + 533 + ], + "type": "interline_equation", + "content": "p _ {\\Theta} \\left(\\left\\{\\tau_ {a} ^ {i} \\right\\} _ {i = 1} ^ {N} \\mid \\mathcal {P}, x _ {0}\\right) = \\underbrace {\\left(\\prod_ {i = 1} ^ {N} p _ {\\phi} \\left(x _ {i} \\mid \\mathcal {P} , x _ {0}\\right)\\right)} _ {\\text {v i s u a l p l a n n i n g}} \\underbrace {\\left(\\prod_ {i = 1} ^ {N} \\prod_ {t = 1} ^ {T} p _ {\\psi} \\left(a _ {i , t} \\mid x _ {i}\\right)\\right)} _ {\\text {a c t i o n p l a n n i n g}} \\tag {1}", + "image_path": "36930f947a1264b3404057f82cd7fd2d2f24e904dad7e2baaefa08ce5974675f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "type": "text", + "content": "The visual planning module consists of a multi-modal encoder to understand the complex task instructions and conditioned diffusion model which implicitly decomposes the task into progressive chained subgoals images. The standardized subgoal images serve as a unified interface bridging the visual planning and action planning modules. With the decoupled foundation model for action prediction, high-level visual planning focuses solely on comprehending instructions and visualizing subgoals without confusion from joint action prediction training. Conversely, the visual planning module frees the action planning module from ambiguous prompt parsing and long-horizon planning by providing coherent keyframes as visual landmarks, further reducing the capabilities requirements of the foundation model to basic single object manipulation skills." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13993" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 68, + 532, + 240 + ], + "blocks": [ + { + "bbox": [ + 62, + 68, + 532, + 240 + ], + "lines": [ + { + "bbox": [ + 62, + 68, + 532, + 240 + ], + "spans": [ + { + "bbox": [ + 62, + 68, + 532, + 240 + ], + "type": "image", + "image_path": "b1d80ffbef7691a905dac1c6b133e9089580b366c4dc70b16ac5f1359107611b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 244, + 548, + 301 + ], + "lines": [ + { + "bbox": [ + 46, + 244, + 548, + 301 + ], + "spans": [ + { + "bbox": [ + 46, + 244, + 548, + 301 + ], + "type": "text", + "content": "Figure 2. Method overview: CoTDiffusion consists of a multi-modal encoder and vision encoder " + }, + { + "bbox": [ + 46, + 244, + 548, + 301 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 46, + 244, + 548, + 301 + ], + "type": "text", + "content": ", semantic alignment module " + }, + { + "bbox": [ + 46, + 244, + 548, + 301 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 46, + 244, + 548, + 301 + ], + "type": "text", + "content": ", conditional diffusion model " + }, + { + "bbox": [ + 46, + 244, + 548, + 301 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 46, + 244, + 548, + 301 + ], + "type": "text", + "content": " and foundation model " + }, + { + "bbox": [ + 46, + 244, + 548, + 301 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 46, + 244, + 548, + 301 + ], + "type": "text", + "content": " for action planning. The prompt and observation tokens are combined and fed into the semantic alignment module to identify the current reasoning chain step, providing progressive guidance for the diffusion model to generate the next subgoal image. The generated keyframes are further fed to the foundation model which predicts action sequences to achieve the imagined goal scene and this recursive process repeats in a receding horizon control loop until the task is finished." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 305, + 261, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 305, + 261, + 319 + ], + "spans": [ + { + "bbox": [ + 47, + 305, + 261, + 319 + ], + "type": "text", + "content": "3.2. Pre-training Coarse Semantic Alignment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 324, + 288, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 324, + 288, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 324, + 288, + 456 + ], + "type": "text", + "content": "For the challenge of explicit decomposing step-wise sub-prompts from multi-modal prompts, the diffusion model struggles to generate sequentially coherent keyframes that incrementally advance the prompt instructions and execution progress. The diffusion model fails to identify the stage of current generated keyframes onto the whole chain-of-thought progress and the generated sub-goals images may exhibit repetitions, skipping, or even backtracking to visited subgoals. To alleviate this dilemma, we design a semantic alignment module to empower diffusion model to track the progress of generated keyframes into the entire chain." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "text", + "content": "Triple Alignment Architecture Since explicit decomposition for the multi-modal prompt is intractable, the initial observation " + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "text", + "content": " and prompt " + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "text", + "content": " remain fixed across the progressive generation of diffusion model. Thus, tracking the progress critically relies on extracting semantic information from the generated subgoal " + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "text", + "content": ". To this end, we design a triple alignment module denoted as " + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "inline_equation", + "content": "S(x_0, x_i, \\mathcal{P})" + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "text", + "content": " to compare current generated subgoal " + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "text", + "content": " against initial observation " + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "text", + "content": " and prompts " + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "text", + "content": ", serving as an information bottleneck to capture the key semantic cues about the subgoal-prompt matching to locate the current chain stage and deliver to the diffusion model for next subgoal generation. Specifically, we first capture visual representation from both the initial and current observation using a shared vision encoder " + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "text", + "content": " and concatenate with prompt tokens " + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "text", + "content": ". Then they are refined through fusion module which consists of several self-attention blocks separately to obtain attention tokens " + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "inline_equation", + "content": "z_i" + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "text", + "content": " aligned to the prompts. The cross-attention between " + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "inline_equation", + "content": "z_i" + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "text", + "content": " aggregates a contrast representation " + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{contrast}}^i" + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "text", + "content": " and then concatenated with prompt tokens into another fusion module to capture the progress tokens " + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{progress}}^i" + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "text", + "content": ", including" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 306, + 547, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 306, + 547, + 475 + ], + "spans": [ + { + "bbox": [ + 304, + 306, + 547, + 475 + ], + "type": "text", + "content": "rich context about the advancement and completion of the prompted task between the two subgoals. Finally, " + }, + { + "bbox": [ + 304, + 306, + 547, + 475 + ], + "type": "inline_equation", + "content": "z_{\\text{progress}}^i" + }, + { + "bbox": [ + 304, + 306, + 547, + 475 + ], + "type": "text", + "content": " and the current visual tokens " + }, + { + "bbox": [ + 304, + 306, + 547, + 475 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 304, + 306, + 547, + 475 + ], + "type": "text", + "content": " and then fed into another cross-attention pass, which visual tokens " + }, + { + "bbox": [ + 304, + 306, + 547, + 475 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 304, + 306, + 547, + 475 + ], + "type": "text", + "content": " serve as keys and progress tokens " + }, + { + "bbox": [ + 304, + 306, + 547, + 475 + ], + "type": "inline_equation", + "content": "z_{\\text{progress}}^i" + }, + { + "bbox": [ + 304, + 306, + 547, + 475 + ], + "type": "text", + "content": " as queries to infer the final aligned tokens " + }, + { + "bbox": [ + 304, + 306, + 547, + 475 + ], + "type": "inline_equation", + "content": "z_{\\text{align}}^i" + }, + { + "bbox": [ + 304, + 306, + 547, + 475 + ], + "type": "text", + "content": ". The aligned token encapsulates cues about 'which subgoal has reached now' and 'which subgoal should reach next', further injected into the diffusion model as precise semantic guidance to steer the generation of next subgoal. Through multi-stage cross-frame attention mechanisms, the triple alignment module can capture the cross-modal semantic correlation and achieve dynamic grounding of generated subgoal image into the multi-modal prompts to identify progress along the CoT reasoning chain." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 486, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 547, + 713 + ], + "type": "text", + "content": "Masked Patch Prediction It is not trivial to directly integrate the designed triple alignment module without pretraining into the diffusion model for joint training from scratch. Simultaneously optimizing semantic alignment and pixel-level reconstruction such as texture, color, shape, and spatial transformations overwhelms joint learning. Thus, we propose a two-stage coarse-to-fine approach decoupling semantic alignment pretraining from diffusion model finetuning, illustrated in Fig. 3. First, we pre-train coarse semantic alignment focused only on spatial correlations, without pixel generation. Considering the adjacent keyframe contrasts reveal object manipulations as mask residuals patch " + }, + { + "bbox": [ + 304, + 486, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\hat{m}_i = x_{i + 1} - x_i" + }, + { + "bbox": [ + 304, + 486, + 547, + 713 + ], + "type": "text", + "content": ", we can utilize these cross-frame spatial semantics to provide coarse alignment supervision. Regions with no visual changes likely correspond to stationary background elements or objects not involved in the current subtask operation. Contrastively, masked areas indicate spatial layout modifications concluding the rich spatial semantic information between keyframes. The sequence" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13994" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 280, + 177 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 280, + 177 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 280, + 177 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 280, + 177 + ], + "type": "image", + "image_path": "01d450c565851e62fe9c738834e999cbcd32251a5114e7b78c416b654f995163.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 62, + 183, + 271, + 194 + ], + "lines": [ + { + "bbox": [ + 62, + 183, + 271, + 194 + ], + "spans": [ + { + "bbox": [ + 62, + 183, + 271, + 194 + ], + "type": "text", + "content": "Figure 3. Two phase of coarse-to-fine alignment pipeline." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 205, + 287, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 205, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 287, + 327 + ], + "type": "text", + "content": "of mask residuals patch " + }, + { + "bbox": [ + 46, + 205, + 287, + 327 + ], + "type": "inline_equation", + "content": "\\{\\hat{m}_i\\}_{i=1}^{N-1}" + }, + { + "bbox": [ + 46, + 205, + 287, + 327 + ], + "type": "text", + "content": " can be viewed as implicitly decomposing the prompt into spatial manipulation steps across visual frames that can be used for alignment. So we fed the aligned tokens into a mask predictor " + }, + { + "bbox": [ + 46, + 205, + 287, + 327 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 205, + 287, + 327 + ], + "type": "text", + "content": " to decode the mask region and set modeling mask prediction as a pretext task. Mask prediction encourages the model to infer these salient spatial semantics and changes to incrementally track prompt completion and focus regions. The process can be formulated as: " + }, + { + "bbox": [ + 46, + 205, + 287, + 327 + ], + "type": "inline_equation", + "content": "\\{m_i\\}_{i=1}^N = \\sum_{i=1}^N \\mathcal{D}(m_i | z_{\\mathrm{align}}^i) S(z_{\\mathrm{align}}^i | x_0, x_i, \\mathcal{P})" + }, + { + "bbox": [ + 46, + 205, + 287, + 327 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 327, + 287, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 287, + 375 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 287, + 375 + ], + "type": "text", + "content": "By factorizing the alignment and generation objectives, the alignment module can develop basic visual reasoning skills through patch-level mask prediction before integrated into the diffusion model for controllable image synthesis." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 382, + 220, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 382, + 220, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 382, + 220, + 396 + ], + "type": "text", + "content": "3.3. Fine-grained Diffusion Training" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 401, + 287, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 287, + 497 + ], + "type": "text", + "content": "In the second stage, we integrate the pretrained coarse alignment module into the diffusion model training for fine-grained image generation. To further enhance instruction following, we design a bi-directional generation that reconstructs the current frame from the aligned token, providing an additional learning constraint. We also employ a frame concatenation mechanism to enhance the consistency across generated keyframes and the fidelity of generation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 502, + 287, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 502, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 46, + 502, + 287, + 669 + ], + "type": "text", + "content": "Bi-directional Alignment Generation Building upon coarse pretraining, we integrate the alignment module into the diffusion model for fine-grained training with the pixel-level supervision of ground truth keyframes, illustrated in Fig. 3. Here we propose bi-directional aligned generation, where the aligned token " + }, + { + "bbox": [ + 46, + 502, + 287, + 669 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{align}}^i" + }, + { + "bbox": [ + 46, + 502, + 287, + 669 + ], + "type": "text", + "content": " not only guides forward prediction but also reconstructs the current frame through backward passing. In the forward pass, based on current subgoal " + }, + { + "bbox": [ + 46, + 502, + 287, + 669 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 46, + 502, + 287, + 669 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 502, + 287, + 669 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{align}}^i" + }, + { + "bbox": [ + 46, + 502, + 287, + 669 + ], + "type": "text", + "content": " provides semantic conditioning to generate the next subgoal image " + }, + { + "bbox": [ + 46, + 502, + 287, + 669 + ], + "type": "inline_equation", + "content": "x_{i+1}" + }, + { + "bbox": [ + 46, + 502, + 287, + 669 + ], + "type": "text", + "content": ". In the backward pass, the same aligned token can guide diffusion model to reconstruct the current subgoal " + }, + { + "bbox": [ + 46, + 502, + 287, + 669 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 46, + 502, + 287, + 669 + ], + "type": "text", + "content": " from the next subgoal image " + }, + { + "bbox": [ + 46, + 502, + 287, + 669 + ], + "type": "inline_equation", + "content": "x_{i+1}" + }, + { + "bbox": [ + 46, + 502, + 287, + 669 + ], + "type": "text", + "content": ", acting as an additional constraint. The training objective can be formulated as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 673, + 295, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 673, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 673, + 295, + 714 + ], + "type": "interline_equation", + "content": "L = \\mathbb {E} _ {x _ {i} \\in D} [ \\underbrace {\\| \\hat {x} _ {i} - \\mathcal {E} (x _ {i - 1} , z _ {\\text {a l i g n}} ^ {i} , \\mathcal {P})} _ {\\text {F o r w a r d G e n e r a t i o n}} \\| + \\underbrace {\\| \\hat {x} _ {i - 1} - \\mathcal {E} (x _ {i} , z _ {\\text {a l i g n}} ^ {i} , \\mathcal {P})} _ {\\text {B a c k w a r d G e n e r a t i o n}} \\| ] \\tag {2}", + "image_path": "cdbad1e3e9778328b8d99dd5f45119d7d5d67fdede4b469f8c4047e314b23640.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "Bi-directional generation compels tighter instruction grounding, as accuracy is required for both progressive and reverse keyframe prediction. This dual-direction linkage enhances visual coherence and sequence controllability compared to forward-only training, improving the chain-of-thought capabilities of diffusion model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 155, + 546, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 155, + 546, + 430 + ], + "spans": [ + { + "bbox": [ + 304, + 155, + 546, + 430 + ], + "type": "text", + "content": "Frame Concat Mechanism Generating consistent and coherent subgoal images requires considering both semantic guidance to match prompts and visual grounding into observations. The aligned tokens " + }, + { + "bbox": [ + 304, + 155, + 546, + 430 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{align}}^i" + }, + { + "bbox": [ + 304, + 155, + 546, + 430 + ], + "type": "text", + "content": " and image tokens " + }, + { + "bbox": [ + 304, + 155, + 546, + 430 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 304, + 155, + 546, + 430 + ], + "type": "text", + "content": " are both fused into the denoising process in a classifier-free manner, as shown in Eq. (2). However, merely injecting current observation tokens as condition into noise model struggles to provide sufficient guidance to prevent distortions conflicting with the current observation, which is critical for precisely grounded robotics manipulations. Inspired by VDT[28], we employ frame concat mechanism to directly take current observation frame as the component of input rather than only condition for the next subgoal image generation, illustrated in Fig. 2. The current observation frame concatenated with the standard Gaussian noise gives a strong visual continuity prior, then fed into diffusion model to denoise altogether. For the training, we split the corresponding frame from the denoising frame and supervised it to the ground truth subgoal image. The rich visual context from the concatenated current frame enables coherent denoising in diffusion model further guarantees consistency with grounded observation and enforces coherence and smoothness across chain-of-thought visual planning." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 441, + 473, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 441, + 473, + 453 + ], + "spans": [ + { + "bbox": [ + 306, + 441, + 473, + 453 + ], + "type": "text", + "content": "3.4. Goal-conditioned Policy Model" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 460, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 713 + ], + "type": "text", + "content": "The final component in our framework is the low-level policy model for action planning, generating an action trajectory " + }, + { + "bbox": [ + 304, + 460, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\tau_{a}^{i}" + }, + { + "bbox": [ + 304, + 460, + 545, + 713 + ], + "type": "text", + "content": " when given observation trajectory " + }, + { + "bbox": [ + 304, + 460, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\tau_{x}^{i}" + }, + { + "bbox": [ + 304, + 460, + 545, + 713 + ], + "type": "text", + "content": " from visual planning. The policy model can be parameterized as an image-conditioned planner that infers the action " + }, + { + "bbox": [ + 304, + 460, + 545, + 713 + ], + "type": "inline_equation", + "content": "a_{i,t}" + }, + { + "bbox": [ + 304, + 460, + 545, + 713 + ], + "type": "text", + "content": " given the current observation " + }, + { + "bbox": [ + 304, + 460, + 545, + 713 + ], + "type": "inline_equation", + "content": "x_{i,t}" + }, + { + "bbox": [ + 304, + 460, + 545, + 713 + ], + "type": "text", + "content": " and the generated subgoal image " + }, + { + "bbox": [ + 304, + 460, + 545, + 713 + ], + "type": "inline_equation", + "content": "g_{i}" + }, + { + "bbox": [ + 304, + 460, + 545, + 713 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 304, + 460, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\tau_{a}^{i} = \\{a_{i,t}\\}_{t=1}^{T} \\sim \\prod_{t=1}^{T} p_{\\psi}(a_{i,t}|x_{i,t}, g_{i})" + }, + { + "bbox": [ + 304, + 460, + 545, + 713 + ], + "type": "text", + "content": ". The policy model is trained by imitation learning in an end-to-end manner, using paired subgoal images and expert trajectories. Thanks to the explicit subgoal generation from high-level visual planner, the low-level policy model does not need to master complex multi-step manipulation skills in long-horizon. Therefore, we simplify the policy model architecture and implement it as a simple MLP to only possess basic single-object manipulation primitives, which greatly reduces the burden of policy model training. Although we do not seek improvement through a more powerful policy model in this paper, CoTDiffusion is flexible to incorporate any goal-conditioned design like transformers or diffusion policies. For more details of low-level policy model, please refer to Appendix E.3." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13995" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 68, + 534, + 261 + ], + "blocks": [ + { + "bbox": [ + 61, + 68, + 534, + 261 + ], + "lines": [ + { + "bbox": [ + 61, + 68, + 534, + 261 + ], + "spans": [ + { + "bbox": [ + 61, + 68, + 534, + 261 + ], + "type": "image", + "image_path": "611df21c38c8862e3cd76d06c27075011ff9873800b5b0b7a14ff0d38ec12229.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 76, + 262, + 515, + 273 + ], + "lines": [ + { + "bbox": [ + 76, + 262, + 515, + 273 + ], + "spans": [ + { + "bbox": [ + 76, + 262, + 515, + 273 + ], + "type": "text", + "content": "Figure 4. Visualization of CoTDiffusion in three typical long-horizon tasks with multi-modal prompts in VIMA-BENCH." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 283, + 128, + 296 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 283, + 128, + 296 + ], + "spans": [ + { + "bbox": [ + 47, + 283, + 128, + 296 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 303, + 155, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 303, + 155, + 316 + ], + "spans": [ + { + "bbox": [ + 47, + 303, + 155, + 316 + ], + "type": "text", + "content": "4.1. Experiment Setup" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 322, + 290, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 322, + 290, + 562 + ], + "spans": [ + { + "bbox": [ + 46, + 322, + 290, + 562 + ], + "type": "text", + "content": "Benchmark & Tasks We conduct evaluation on VIMABENCH, a benchmark suite for multimodal robot learning, which is built on the Ravens robot simulator [50]. VIMABENCH supports extensible collections of objects and textures to compose multi-modal prompts and to procedurally generate a large number of tasks. VIMA-BENCH can generate large quantities of expert trajectories via scripted oracle agents, and the ground truth keyframes that can leveraged for the training of CoTDiffusion. VIMA-BENCH contains various tasks ranging from simple object manipulation to multi-object manipulation and we select three representative long-horizon manipulation tasks in VIMABENCH-visual rearrangement, visual constraints, and visual reasoning, which represent different levels of complexity and requirements for embodied manipulation. These tasks represent different levels of complexity and requirements for embodied manipulation and can test the ability to understand spatial relationships and perform precise object manipulations. For more details regarding the simulation benchmark and tasks setting, please refer to Appendix A." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 571, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 571, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 571, + 287, + 594 + ], + "type": "text", + "content": "Evaluation Metric The evaluation metrics we use in the experiments cover three aspects as following:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 47, + 594, + 287, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 594, + 287, + 630 + ], + "spans": [ + { + "bbox": [ + 47, + 594, + 287, + 630 + ], + "type": "text", + "content": "- Image Fidelity - measures the quality and realism of the generated keyframe images. We choose Fréchet Inception Distance (FID) [15] as a useful metric." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 630, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 677 + ], + "type": "text", + "content": "- Instruction Following - evaluates how well the step-by-step keyframes cover the complex instructions. We use CLIP-based text-image similarity scores [14] between the prompt and keyframes to quantify instruction alignment." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "- Task Completion - tests the end-to-end utility by executing the embodied manipulation task. The final success rate on long-horizon tasks can be a fair metric." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 283, + 372, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 283, + 372, + 295 + ], + "spans": [ + { + "bbox": [ + 306, + 283, + 372, + 295 + ], + "type": "text", + "content": "4.2. Baselines" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 303, + 545, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 303, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 305, + 303, + 545, + 338 + ], + "type": "text", + "content": "There are several existing methods for constructing robot manipulation policies conditioned on complex prompts, which we use as baselines in our experiments:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 339, + 545, + 589 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 306, + 339, + 545, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 339, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 306, + 339, + 545, + 399 + ], + "type": "text", + "content": "- Gato[41], a generalist agent that can solve tasks from multiple domains where tasks are specified by prompting the model with the observation and action subsequence. For fair comparisons, we provide the same conditioning manner with multi-modal prompts." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 399, + 545, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 399, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 306, + 399, + 545, + 458 + ], + "type": "text", + "content": "- Flamingo[2], a vision-language model that learns to generate textual completion in response to multimodal prompts. We borrow the architecture in [21] which adapts it to support decision-masking by replacing the output layer with robot action heads." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 459, + 545, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 459, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 306, + 459, + 545, + 529 + ], + "type": "text", + "content": "- VIMA[21], the first transformer-based generalist robot agent that processes manipulation tasks with multimodal prompts. VIMA adpots an object-centric approach to flatten all the observation and prompts into object tokens sequence and predicts motor actions autoregressively and demonstrates SOTA performance on VIMA-BENCH." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 530, + 545, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 530, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 306, + 530, + 545, + 589 + ], + "type": "text", + "content": "- SuSIE[4], the concurrent work for visual planning in manipulation tasks on the CALVIN [31], incorporating pretrained image-editing model and the low-level goal-conditioned policy. For fair comparison, we finetune the model with the same oracle data with CoTDiffusion." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 598, + 499, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 598, + 499, + 611 + ], + "spans": [ + { + "bbox": [ + 306, + 598, + 499, + 611 + ], + "type": "text", + "content": "4.3. Quantitative Results of Success Rate" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "We begin by comparing the performance of CoTDiffusion and baselines to solve long-horizon tasks in visual rearrange, visual reasoning, and visual constraints. The performance comparisons shown in Tab. 1 demonstrate CoT-Diffusion significantly outperforms other baselines in success rate. The baselines can be divided into two kinds of planners, including abstract planner and visual planner. Abstract planner like Gato, Flamingo and VIMA directly" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "13996" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "text", + "content": "map general prompts to subsequent actions in an end-to-end manner. Gato and Flamingo gets low success rates on long-horizon tasks without explicit subgoal generation to correct the accumulative deviation errors from the instructions. In contrast, Visual planner like SuSIE and CoTDiffusion can generate intermediate goal images to guide the action planning, which can enhance instruction following for long-horizon tasks via visual planning. However, SuSIE lacks intrinsic chain-of-thought reasoning capabilities, struggling to generate logically progressing subgoal sequences as coherent plans from general prompts. To enable comparison, we grant SuSIE the manually decomposed prompts into privileged step-wise sub-prompts which are unavailable in fair settings, denoted as 'SuSIE + sub-prompts'. Nonetheless, its performance still lags CoTDiffusion, which can implicitly align generated images with multi-modal prompts through learned correspondence between visual signals and language semantics. One more potential reason to restrict the performance of SuSIE is that provided sub-prompts are challenging to perfectly encapsulate complex instructions. In contrast, CoTDiffusion develops intrinsic chain-of-thought reasoning and alignment for generated subgoal images for flexible visual planning directly from the raw multi-modal prompts, without the need for pre-decomposed sub-prompts. The comparisons highlight the importance of chain-of-thought capabilities in visual planning, especially the role of implicit semantic alignment for instruction following long-horizon manipulation tasks." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 411, + 286, + 490 + ], + "blocks": [ + { + "bbox": [ + 48, + 411, + 286, + 490 + ], + "lines": [ + { + "bbox": [ + 48, + 411, + 286, + 490 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 286, + 490 + ], + "type": "table", + "html": "
MethodologyRearrangeReasoningConstraintsOverall
Gato6.4 ± 1.32.5 ± 0.425.2 ± 3.111.4 ± 1.6
Flamingo17.5 ± 1.63.0 ± 0.536.1 ± 4.218.9 ± 2.1
VIMA43.1 ± 3.338.2 ± 4.467.2 ± 5.249.5 ± 4.3
SuSIE2.7 ± 0.33.1 ± 0.624.3 ± 5.910.0 ± 2.3
+sub-prompts37.7 ± 6.239.0 ± 4.552.3 ± 7.043.0 ± 5.9
CoTDiffusion59.0 ± 1.751.7 ± 2.683.1 ± 4.764.6 ± 3.0
", + "image_path": "f7024d7639341b2f27681311e6db79ea6a0c66e29f0540826e7b30261ec497b8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 527, + 149, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 527, + 149, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 527, + 149, + 540 + ], + "type": "text", + "content": "4.4. Further Analysis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": "Robustness to Insufficient Perception Rich visual observations from diverse views are crucial for complex robot manipulation tasks. Restricted perspectives limit rich representations used for action planning, degrading performance on long-horizon multi-object manipulation. We evaluate different methods in single-view including the top and front views provided from VIMA-BENCH. As shown in Tab. 2, visual planning approach demonstrates greater robustness to limited observations, with a smaller performance drop in single-view compared to abstract methods. We attribute this to two potential reasons: First, accurate and grounded subgoal images generated in visual planners provide supplemental visual context, which can partly compensate for the insufficient perception to aid robustness under single-" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 309, + 70, + 544, + 155 + ], + "blocks": [ + { + "bbox": [ + 47, + 495, + 287, + 518 + ], + "lines": [ + { + "bbox": [ + 47, + 495, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 47, + 495, + 287, + 518 + ], + "type": "text", + "content": "Table 1. The evaluations of success rates on three typical long-horizon tasks with multi-modal prompts." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 70, + 544, + 155 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 544, + 155 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 544, + 155 + ], + "type": "table", + "html": "
MethodologyMulti-ViewSingle-ViewPerformance Drop
Gato11.4 ± 1.66.5 ± 1.64.9(42% ↓)
Flamingo18.9 ± 2.112.0 ± 2.46.9(36.2% ↓)
VIMA49.5 ± 4.334.9 ± 3.414.6(29.4% ↓)
SuSIE10.0 ± 2.37.9 ± 2.02.1(21.0% ↓)
+sub-prompts43.0 ± 5.935.6 ± 6.17.4(18.8% ↓)
CoTDiffusion64.6 ± 3.056.0 ± 2.48.6(13.2% ↓)
", + "image_path": "ba4a0e539b809ac93ed103fc3f9204ff3bd92eb99e3e38a35b56a6deffb9a275.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 309, + 217, + 561, + 309 + ], + "blocks": [ + { + "bbox": [ + 306, + 159, + 545, + 182 + ], + "lines": [ + { + "bbox": [ + 306, + 159, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 306, + 159, + 545, + 182 + ], + "type": "text", + "content": "Table 2. The evaluations of performance drop of different methods on single-view and multi-view from VIMA-BENCH." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 194, + 545, + 216 + ], + "lines": [ + { + "bbox": [ + 306, + 194, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 306, + 194, + 545, + 216 + ], + "type": "text", + "content": "Table 3. Quantitative comparisons of FID between methods on all three tasks, including visual rearrange, reasoning and constraints." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 217, + 561, + 309 + ], + "lines": [ + { + "bbox": [ + 309, + 217, + 561, + 309 + ], + "spans": [ + { + "bbox": [ + 309, + 217, + 561, + 309 + ], + "type": "table", + "html": "
MethodologyRearrangeReasoningConstraints
SuSIE [4]19.718.323.2
- w/o finetune38.4 ↑33.0 ↑52.7 ↑
- w/o stepwise prompt32.6 ↑28.1 ↑34.7 ↑
CoTDiffusion (ours)11.310.88.6
- w/o coarse pre-train23.1 ↑16.2 ↑18.9 ↑
- w/o bi-direction align18.7 ↑15.0 ↑15.8 ↑
- w/o frame concat15.4 ↑13.7 ↑12.1 ↑
", + "image_path": "9535e3e4fefb47a881a7b0b39167476d51ef10790a355ee8ff43451ce0ba1114.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 316, + 545, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 545, + 411 + ], + "type": "text", + "content": "view. Second, by providing coherent subgoal images as visual landmarks, the requirements for the low-level action planner are reduced to basic single-object manipulation primitives in short horizon, with less reliance on rich visual perceptions. The experiments demonstrate that CoTDiffusion enjoys better robustness to restricted perception than abstract planners, highlighting the benefits of hierarchical framework decoupled visual planning and action planning." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 418, + 546, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 546, + 682 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 546, + 682 + ], + "type": "text", + "content": "Fidelity of Image Generation Here we conduct further analysis of visual planners by comparing goal image quality against ground truth keyframes with FID as evaluation metrics. As Tab. 3 shows, CoTDiffusion still achieves much better fidelity than SuSIE even though SuSIE has got improved after fine-tuning on the same datasets in VIMABENCH. Moreover, SuSIE can only understand short instructions of single-object manipulation tasks, struggling to handle long-horizon instructions due to the lack of chain-of-thought reasoning capabilities to handle multi-modal prompts. With step-wise sub-prompts decomposed in advance, the performance of SuSIE gets largely raised but still underperforms CoTDiffusion, which has no need to explicitly decompose the general prompts and can generate subgoal images in an implicit chain-of-thought manner. Additionally, ablating coarse pretraining and bi-directional generation degrades performance, validating their benefits. The coarse-to-fine semantic alignment training training allows developing spatial reasoning prior to synthesis. Frame concatenation further guides coherent denoising by providing rich context information as visual priors to ground the current observation and enhance the fidelity of generation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "Accuracy of Instruction Following We evaluate instruction following accuracy via CLIP similarity between gen" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13997" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 420 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 420 + ], + "type": "text", + "content": "erated keyframes and general prompts, normalized by the CLIP score between ground truth ultimate goal image and prompts. The results in Fig. 5 demonstrate that CoT-Diffusion enjoys better instruction following capabilities. Without chain-of-thought reasoning abilities, SuSIE struggles to follow instructions when given general multi-modal prompts, let alone generate subgoal images with smooth progressions. Progressively providing privileged prompts step by step improves SuSIE in instruction following and gradual advancement towards ultimate goals. However, text prompts decomposed by rules are not always perfect to convey the original multi-modal prompt, so the reliance on these prompts restricts the performance of SuSIE. In contrast, the coarse alignment pretraining and bi-directional generation can assist the diffusion model in tracking the progress of generated keyframes throughout the entire chain and generate sequenced keyframes incrementally advancing prompt instructions. Moreover, the implicit decomposition maintains semantic coherence between subgoals without overly large jumps and the smoothness and continuity between visual milestones make them more reachable by downstream foundation models with just single-object manipulation capabilities. Additionally, we observe that the bi-directional generation may impedes the diffusion model training if without coarse semantic pretraining. As the initial observation and prompt remain fixed across different generation steps, the align tokens at various stage tend to so similar that may confuse simultaneous progressive generation and reconstruction training." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 434, + 289, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 434, + 289, + 578 + ], + "spans": [ + { + "bbox": [ + 47, + 434, + 289, + 578 + ], + "type": "text", + "content": "Generalization across Tasks We evaluate the generalization ability in three levels with increasing difficulty: placement generalization which randomizes the novel placement of objects (L1), object generalization which provides the objects with novel attributes (L2), and task combinatorial generalization which complexes the prompts with extra novel instruction (L3). As Fig. 6 shows, CoTDiffusion exhibits strong zero-shot generalization on unfamiliar objects, colors, and shapes at different placements by explicitly visualizing them into coherent subgoals grounded on new concepts. This avoids re-grounding concepts from multi-modal prompts in the low-level foundation model for" + } + ] + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 50, + 586, + 286, + 678 + ], + "blocks": [ + { + "bbox": [ + 50, + 586, + 286, + 678 + ], + "lines": [ + { + "bbox": [ + 50, + 586, + 286, + 678 + ], + "spans": [ + { + "bbox": [ + 50, + 586, + 286, + 678 + ], + "type": "image", + "image_path": "2323b9c507967d8d3a6d98f7b39ab2117fbcdf9b5a73208aca9be77b68fe387b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 687, + 286, + 710 + ], + "lines": [ + { + "bbox": [ + 46, + 687, + 286, + 710 + ], + "spans": [ + { + "bbox": [ + 46, + 687, + 286, + 710 + ], + "type": "text", + "content": "Figure 5. The normalized CLIP scores for each generation step, reflecting the step-wise accuracy of instruction following." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 307, + 73, + 541, + 194 + ], + "blocks": [ + { + "bbox": [ + 307, + 73, + 541, + 194 + ], + "lines": [ + { + "bbox": [ + 307, + 73, + 541, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 541, + 194 + ], + "type": "image", + "image_path": "b2bc2a91b3604bd403e5ea396ca0a9ddd8a3fc4f8af91b2eabfbec501460e426.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 198, + 529, + 209 + ], + "lines": [ + { + "bbox": [ + 321, + 198, + 529, + 209 + ], + "spans": [ + { + "bbox": [ + 321, + 198, + 529, + 209 + ], + "type": "text", + "content": "Figure 6. The evaluation on various generalization levels." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 219, + 545, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 219, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 304, + 219, + 545, + 387 + ], + "type": "text", + "content": "the placement and object. For L3 generalization, taking an example, we complex the prompt such as 'rearrange ... then rotate/twist/stack ...' to demand extra tasks such as rotating, twisting degrees, or stacking objects. CoTDiffusion achieves outstanding gain in the zero-shot performance of combinatorial tasks. When CoTDiffusion visualizes novel concepts into goal images, the foundation model can still accomplish the stack by simply achieving the provided subgoal, with no need to inherently understand novel skills like stack. This demonstrates the power of 'a image is worth a thousand words' - the subgoal images facilitate the generalization of the foundation model to unseen objects and novel combinatorial task while the foundation model simply achieves them with reuse of simple known primitives." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 403, + 378, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 403, + 378, + 415 + ], + "spans": [ + { + "bbox": [ + 306, + 403, + 378, + 415 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 424, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 424, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 424, + 545, + 639 + ], + "type": "text", + "content": "We presented CoTDiffusion, a hierarchical framework that integrates diffusion model as high-level module to translate the general multi-modal prompts into coherent subgoal images, serves as the visual milestones to anchor the low-level foundation model to plan action sequences, termed as 'generate subgoal images before act'. With the coarse-to-fine training for semantic alignment module, CoTDiffusion can identify the progress of generated subgoals images along reasoning chains, unlocking the chain-of-thought reasoning capabilities of diffusion model for long-horizon manipulation tasks. The experiments cover various long-horizon manipulation scenarios in VIMA-BENCH, and CoTDiffusion show the strong instruction following and outstanding performance gain compared to existed methods without visual planning. Incorporating commonsense knowledge from pre-trained MLLM like GPT-4V provides an avenue for more generalizable and promising reasoning in CoTDiffusion, which leaves us our future work." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 656, + 416, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 656, + 416, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 656, + 416, + 670 + ], + "type": "text", + "content": "6. Acknowledgement" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "This work is supported by the National Natural Science Foundation of China (Grant Nos. 92370132) and the Xiaomi Young Talents Program of Xiaomi Foundation." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13998" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, et al. Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691, 2022. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "type": "text", + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: A visual language model for few-shot learning. NeurIPS, 2022. 6, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 287, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 287, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 287, + 237 + ], + "type": "text", + "content": "[3] Fan Bao, Shen Nie, Kaiwen Xue, Chongxuan Li, Shi Pu, Yaole Wang, Gang Yue, Yue Cao, Hang Su, and Jun Zhu. One transformer fits all distributions in multi-modal diffusion at scale. arXiv preprint arXiv:2303.06555, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 239, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 287, + 293 + ], + "type": "text", + "content": "[4] Kevin Black, Mitsuhiko Nakamoto, Pranav Atreya, Homer Walke, Chelsea Finn, Aviral Kumar, and Sergey Levine. Zero-shot robotic manipulation with pretrained image-editing diffusion models. arXiv preprint arXiv:2310.10639, 2023. 6, 7, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 295, + 287, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 295, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 53, + 295, + 287, + 350 + ], + "type": "text", + "content": "[5] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 351, + 287, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 351, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 351, + 287, + 384 + ], + "type": "text", + "content": "[6] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. arXiv preprint arXiv:2211.09800, 2022. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 386, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 386, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 386, + 287, + 430 + ], + "type": "text", + "content": "[7] Cheng Chi, Siyuan Feng, Yilun Du, Zhenjia Xu, Eric Cousineau, Benjamin Burchfiel, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. arXiv preprint arXiv:2303.04137. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 431, + 287, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 431, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 53, + 431, + 287, + 485 + ], + "type": "text", + "content": "[8] Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 487, + 287, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 487, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 53, + 487, + 287, + 531 + ], + "type": "text", + "content": "[9] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In Advances in Neural Information Processing Systems, pages 8780-8794. Curran Associates, Inc., 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 533, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 287, + 588 + ], + "type": "text", + "content": "[10] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palm- e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "type": "text", + "content": "[11] Jiafei Duan, Samson Yu, Hui Li Tan, Hongyuan Zhu, and Cheston Tan. A survey of embodied ai: From simulators to research tasks. IEEE Transactions on Emerging Topics in Computational Intelligence, 6(2):230-244, 2022. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "text", + "content": "[12] Theophile Gervet, Zhou Xian, Nikolaos Gkanatsios, and Katerina Fragkiadaki. Act3d: Infinite resolution action detection transformer for robotic manipulation. arXiv preprint arXiv:2306.17817, 2023. 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "text", + "content": "[13] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 4" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[14] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "type": "text", + "content": "[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 173, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 545, + 194 + ], + "type": "text", + "content": "[16] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 195, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 195, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 307, + 195, + 545, + 227 + ], + "type": "text", + "content": "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 228, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 545, + 271 + ], + "type": "text", + "content": "[18] Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. Voxposer: Composable 3d value maps for robotic manipulation with language models. arXiv preprint arXiv:2307.05973, 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 273, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 273, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 307, + 273, + 545, + 326 + ], + "type": "text", + "content": "[19] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General perception with iterative attention. In International conference on machine learning, pages 4651-4664. PMLR, 2021. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 327, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 545, + 370 + ], + "type": "text", + "content": "[20] Stephen James, Zicong Ma, David Rovick Arrojo, and Andrew J Davison. Rlbench: The robot learning benchmark & learning environment. IEEE Robotics and Automation Letters, 5(2):3019-3026, 2020. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 372, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 372, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 307, + 372, + 545, + 425 + ], + "type": "text", + "content": "[21] Yunfan Jiang, Agrim Gupta, Zichen Zhang, Guanzhi Wang, Yongqiang Dou, Yanjun Chen, Li Fei-Fei, Anima Anandkumar, Yuke Zhu, and Linxi Fan. Vima: General robot manipulation with multimodal prompts. arXiv preprint arXiv:2210.03094, 2022. 1, 2, 3, 6, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 426, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 426, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 307, + 426, + 545, + 469 + ], + "type": "text", + "content": "[22] Dongxu Li, Junnan Li, and Steven CH Hoi. Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. arXiv preprint arXiv:2305.14720, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 471, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 471, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 307, + 471, + 545, + 525 + ], + "type": "text", + "content": "[23] Zhengzhong Liang, Steven Bethard, and Mihai Surdeanu. Explainable multi-hop verbal reasoning through internal monologue. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 526, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 526, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 307, + 526, + 545, + 568 + ], + "type": "text", + "content": "[24] Wang Ling, Dani Yogatama, Chris Dyer, and Phil Blunson. Program induction by rationale generation: Learning to solve and explain algebraic word problems. arXiv preprint arXiv:1705.04146, 2017. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 570, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 570, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 307, + 570, + 545, + 601 + ], + "type": "text", + "content": "[25] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 603, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 603, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 603, + 545, + 624 + ], + "type": "text", + "content": "[26] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 625, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 625, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 545, + 668 + ], + "type": "text", + "content": "[27] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 12" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[28] Haoyu Lu, Guoxing Yang, Nanyi Fei, Yuqi Huo, Zhiwu Lu, Ping Luo, and Mingyu Ding. Vdt: An empirical study on video diffusion with transformers. arXiv preprint arXiv:2305.13311, 2023. 5" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "13999" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "text", + "content": "[29] Yecheng Jason Ma, Shagun Sodhani, Dinesh Jayaraman, Obert Bastani, Vikash Kumar, and Amy Zhang. Vip: Towards universal visual reward and representation via value-implicit pre-training. arXiv preprint arXiv:2210.00030, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 288, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 288, + 163 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 288, + 163 + ], + "type": "text", + "content": "[30] Yecheng Jason Ma, William Liang, Vaidehi Som, Vikash Kumar, Amy Zhang, Osbert Bastani, and Dinesh Jayaraman. Liv: Language-image representations and rewards for robotic control. arXiv preprint arXiv:2306.00958, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 218 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 218 + ], + "type": "text", + "content": "[31] Oier Mees, Lukas Hermann, Erick Rosete-Beas, and Wolfram Burgard. Calvin: A benchmark for language-conditioned policy learning for long-horizon robot manipulation tasks. IEEE Robotics and Automation Letters (RA-L), 7(3):7327-7334, 2022. 6, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 220, + 288, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 288, + 275 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 288, + 275 + ], + "type": "text", + "content": "[32] Oier Mees, Lukas Hermann, Erick Rosete-Beas, and Wolfram Burgard. Calvin: A benchmark for language-conditioned policy learning for long-horizon robot manipulation tasks. IEEE Robotics and Automation Letters, 7(3): 7327-7334, 2022. 8, 11" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 275, + 288, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 275, + 288, + 331 + ], + "spans": [ + { + "bbox": [ + 48, + 275, + 288, + 331 + ], + "type": "text", + "content": "[33] Yao Mu, Qinglong Zhang, Mengkang Hu, Wenhai Wang, Mingyu Ding, Jun Jin, Bin Wang, Jifeng Dai, Yu Qiao, and Ping Luo. Embodiedgpt: Vision-language pretraining via embodied chain of thought. arXiv preprint arXiv:2305.15021, 2023. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 332, + 288, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 332, + 288, + 376 + ], + "spans": [ + { + "bbox": [ + 48, + 332, + 288, + 376 + ], + "type": "text", + "content": "[34] Suraj Nair, Aravind Rajeswaran, Vikash Kumar, Chelsea Finn, and Abhinav Gupta. R3m: A universal visual representation for robot manipulation. arXiv preprint arXiv:2203.12601, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 377, + 288, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 377, + 288, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 377, + 288, + 443 + ], + "type": "text", + "content": "[35] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, pages 16784-16804. PMLR, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 444, + 288, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 444, + 288, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 444, + 288, + 510 + ], + "type": "text", + "content": "[36] Maxwell Nye, Anders Johan Andreassen, Guy Gur-Ari, Henryk Michalewski, Jacob Austin, David Bieber, David Dohan, Aitor Lewkowycz, Maarten Bosma, David Luan, et al. Show your work: Scratchpads for intermediate computation with language models. arXiv preprint arXiv:2112.00114, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 511, + 288, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 511, + 288, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 511, + 288, + 578 + ], + "type": "text", + "content": "[37] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35: 27730-27744, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 578, + 288, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 288, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 288, + 645 + ], + "type": "text", + "content": "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "type": "text", + "content": "[39] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 690, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 690, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 690, + 288, + 714 + ], + "type": "text", + "content": "[40] Scott Reed and Nando De Freitas. Neural programmer-interpreters. arXiv preprint arXiv:1511.06279, 2015. 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 547, + 642 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 307, + 72, + 546, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 128 + ], + "type": "text", + "content": "[41] Scott Reed, Konrad Zolna, Emilio Parisotto, Sergio Gomez Colmenarejo, Alexander Novikov, Gabriel Barth-Maron, Mai Gimenez, Yury Sulsky, Jackie Kay, Jost Tobias Springenberg, et al. A generalist agent. arXiv preprint arXiv:2205.06175, 2022. 3, 6, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 129, + 547, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 184 + ], + "type": "text", + "content": "[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 185, + 547, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 547, + 239 + ], + "type": "text", + "content": "[43] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 240, + 546, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 240, + 546, + 306 + ], + "spans": [ + { + "bbox": [ + 307, + 240, + 546, + 306 + ], + "type": "text", + "content": "[44] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 307, + 546, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 307, + 546, + 340 + ], + "spans": [ + { + "bbox": [ + 307, + 307, + 546, + 340 + ], + "type": "text", + "content": "[45] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 341, + 546, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 341, + 546, + 374 + ], + "spans": [ + { + "bbox": [ + 307, + 341, + 546, + 374 + ], + "type": "text", + "content": "[46] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 6, 12" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 375, + 546, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 375, + 546, + 419 + ], + "spans": [ + { + "bbox": [ + 307, + 375, + 546, + 419 + ], + "type": "text", + "content": "[47] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, 2017. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 419, + 546, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 419, + 546, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 419, + 546, + 464 + ], + "type": "text", + "content": "[48] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Ed Chi, Quoc Le, and Denny Zhou. Chain of thought prompting elicits reasoning in large language models. arXiv preprint arXiv:2201.11903, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 464, + 546, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 464, + 546, + 520 + ], + "spans": [ + { + "bbox": [ + 307, + 464, + 546, + 520 + ], + "type": "text", + "content": "[49] Xingqian Xu, Zhangyang Wang, Gong Zhang, Kai Wang, and Humphrey Shi. Versatile diffusion: Text, images and variations all in one diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7754-7765, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 520, + 546, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 520, + 546, + 586 + ], + "spans": [ + { + "bbox": [ + 307, + 520, + 546, + 586 + ], + "type": "text", + "content": "[50] Andy Zeng, Pete Florence, Jonathan Tompson, Stefan Welker, Jonathan Chien, Maria Attarian, Travis Armstrong, Ivan Krasin, Dan Duong, Vikas Sindhwani, et al. Transporter networks: Rearranging the visual world for robotic manipulation. In Conference on Robot Learning, pages 726-747. PMLR, 2021. 6, 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 586, + 546, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 586, + 546, + 642 + ], + "spans": [ + { + "bbox": [ + 307, + 586, + 546, + 642 + ], + "type": "text", + "content": "[51] Brianna Zitkovich, Tianhe Yu, Sichun Xu, Peng Xu, Ted Xiao, Fei Xia, Jialin Wu, Paul Wohlhart, Stefan Welker, Ayzaan Wahid, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. In 7th Annual Conference on Robot Learning, 2023. 1, 2, 3" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "14000" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Generating Content for HDR Deghosting from Frequency View/2b6f8409-576d-47cf-ab70-c1b4b9e76809_content_list.json b/2024/Generating Content for HDR Deghosting from Frequency View/2b6f8409-576d-47cf-ab70-c1b4b9e76809_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a66b7c09eed2a520969a9ce45e9a9444b623b9fe --- /dev/null +++ b/2024/Generating Content for HDR Deghosting from Frequency View/2b6f8409-576d-47cf-ab70-c1b4b9e76809_content_list.json @@ -0,0 +1,2461 @@ +[ + { + "type": "text", + "text": "Generating Content for HDR Deghosting from Frequency View", + "text_level": 1, + "bbox": [ + 165, + 130, + 803, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tao Hu $^{1\\dagger}$ Qingsen Yan $^{1\\dagger}$ * Yuankai Qi $^{2}$ Yanning Zhang $^{1}$", + "bbox": [ + 236, + 179, + 728, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Northwestern Polytechnical University $^{2}$ Macquarie University", + "bbox": [ + 235, + 198, + 733, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "hutaoxauat@gmail.com, $\\mathbb{1}\\{$ qingsenyan,ynzhang $\\} \\text{=} \\mathrm{nwpu.edu.cn},$ yuankai.qi@mq.edu.au", + "bbox": [ + 147, + 218, + 813, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recovering ghost-free High Dynamic Range (HDR) images from multiple Low Dynamic Range (LDR) images becomes challenging when the LDR images exhibit saturation and significant motion. Recent Diffusion Models (DMs) have been introduced in HDR imaging field, demonstrating promising performance, particularly in achieving visually perceptible results compared to previous DNN-based methods. However, DMs require extensive iterations with large models to estimate entire images, resulting in inefficiency that hinders their practical application. To address this challenge, we propose the Low-Frequency aware Diffusion (LF-Diff) model for ghost-free HDR imaging. The key idea of LF-Diff is implementing the DMs in a highly compacted latent space and integrating it into a regression-based model to enhance the details of reconstructed images. Specifically, as low-frequency information is closely related to human visual perception we propose to utilize DMs to create compact low-frequency priors for the reconstruction process. In addition, to take full advantage of the above low-frequency priors, the Dynamic HDR Reconstruction Network (DHRNet) is carried out in a regression-based manner to obtain final HDR images. Extensive experiments conducted on synthetic and real-world benchmark datasets demonstrate that our LF-Diff performs favorably against several state-of-the-art methods and is $10\\times$ faster than previous DM-based methods.", + "bbox": [ + 75, + 300, + 472, + 691 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 723, + 207, + 738 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multiple exposure High Dynamic Range (HDR) imaging aims to restore missing details from exposure-varied Low Dynamic Range (LDR) images. However, in dynamic scenes, it often causes ghosting artifacts due to object or camera movement, limiting practical applications. Researchers are actively exploring ghost-free image recon", + "bbox": [ + 75, + 750, + 468, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6195452a8473ddc281ee196038780c732efb56a605e8f53e9e275bb1b69c42a1.jpg", + "image_caption": [ + "Figure 1. Average PSNR vs. inference time on the Kalantari's dataset [11]. Our method performs favorably and is $10\\times$ faster than the previous diffusion-model-based method DiffHDR [44]." + ], + "image_footnote": [], + "bbox": [ + 500, + 268, + 890, + 425 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "struction methods to enable seamless capture of dynamic scenes with high dynamic ranges.", + "bbox": [ + 498, + 491, + 890, + 521 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, the rise of Deep Neural Networks (DNNs) has brought significant advancements to HDR imaging field. Numerous DNN-based methods have emerged, employing Convolutional Neural Networks (CNNs) [11, 41, 47] or Vision Transformers (ViTs) [16, 28] for HDR image reconstruction. Despite their advancements, DNN-based approaches face challenges when essential information (e.g., content, details) for overexposed areas is missing due to object or camera movement. More recently, Diffusion Models (DMs) [8, 26] have exhibited impressive performance in image synthesis [4, 29] and image reconstruction tasks (including HDR imaging) [24, 44, 46]. DMs achieve this by iteratively denoising Gaussian noise to generate high-fidelity images. Yan et al. [44] reconstruct complete HDR images from pure Gaussian noise, yielding impressive results. Note that, compared to previous Deep generative models [14, 21], DMs generate a more accurate target distribution without facing issues like optimization instability or mode collapse, making them a promising choice for enhancing HDR image quality. Nonetheless, DMs demand a substantial number of iteration steps on denoising models to capture intricate data details, which is time-consuming even with a high-end GPU card. For example, as shown in Fig. 1, DiffHDR [44] takes about 7.5 seconds to generate an HDR image on a single A100 GPU.", + "bbox": [ + 496, + 523, + 892, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author. † The first two authors contributed equally to this work. This work was partially supported by NSFC (62301432,62306240), NSBRPS (2023-JC-QN-0685, QCYRCXM-2023-057). Y. Qi is not supported by the above mentioned fundings.", + "bbox": [ + 75, + 851, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "25732", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, our goal is to devise a DM-based method that effectively harnesses the robust distribution mapping capabilities of DMs for HDR image reconstruction. We note that it is unnecessary to reconstruct a complete HDR image from pure noise, because the LDR reference images already provide most of the necessary content information for HDR images. As low-frequency information is closely related to human visual perception [34], we propose to utilize DMs to create compact low-frequency priors in the latent space. These priors are then integrated into the regression-based model to predict the low-frequency content of the reconstructed HDR images.", + "bbox": [ + 75, + 90, + 472, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To achieve our goal, we propose the Low-Frequency Aware Diffusion Model (LF-Diff), comprising a Low-Frequency Prior Extraction Network (LPENet), a denoising network, and a Dynamic HDR Reconstruction Network (DHRNet). Following established methods [3, 23, 39], we employ a two-stage training strategy for pre-train LF-Diff and DM training. In the first stage, LPENet learns to extract a compact Low-frequency Prior Representation (LPR) from ground-truth images, guiding the DHRNet. DHRNet consists of two modules: a Prior Integration Module (PIM) that fuses the LPR with intermediate features of DHRNet, and a Feature Refinement Module (FRM) that further processes the fused features to HDR image. Notably, LPENet and DHRNet are optimized together in this phase. In the second stage, we train DM to learn the compact LPR directly from LDR images. Since the LPR is lightweight and only used to low-frequency content for HDR imaging, our DM can estimate the LPR with extremely low computational cost, ensuring stable visual results after multiple iterations.", + "bbox": [ + 75, + 272, + 472, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions can be summarized as follows:", + "bbox": [ + 94, + 559, + 457, + 573 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce LF-Diff, a straightforward and efficient DM-based method for HDR imaging. LF-Diff leverages the capabilities of diffusion models to generate informative low-frequency priors, which are then integrated into the imaging process to enhance results.", + "- We propose PIM and FRM in the DHRNet to fully exploit the LPR. While PIM efficiently fuses the LPR with intermediate features of DHRNet, FRM further processes the fused features to reconstruct high-quality HDR images.", + "- Extensive experiments demonstrate that the proposed LF-Diff method achieves SOTA performance in HDR imaging tasks, and produces visually appealing results that align with human visual perception. But consuming significantly fewer computational resources compared to other DM-based methods." + ], + "bbox": [ + 76, + 574, + 468, + 800 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 814, + 218, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Ghost-free HDR Reconstruction. Traditional methods often employ motion rejection [7, 40], motion registration [31, 36], and patch matching [9, 25] approaches to align LDR images and reconstruct high-quality results. However,", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "these methods heavily rely on the performance of preprocessing techniques and often face challenges when dealing with motions across large spatial extents. Subsequently, DNN-based methods have emerged as the mainstream approach, thanks to their superior nonlinear expression capability. Researchers have explored various applications of DNNs in HDR imaging, devising sophisticated network architectures and model optimization schemes, such as attention [1, 41], transformer [16, 28, 30], optical flow [11, 22], GAN [6, 14, 21], and others [5, 15, 42, 43, 47]. Nevertheless, when LDR images lack sufficient information due to motion or saturation, they often manifest artifacts not aligned with human visual perception, commonly referred to as ghosting.", + "bbox": [ + 496, + 90, + 890, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion Models. Diffusion models have recently demonstrated promising results in various low-level vision tasks, including super-resolution [13, 24], HDR reconstruction [44, 46], colorization [12, 35], and deblurring [37]. DiffHDR [44] applied the conventional diffusion model paradigm to reconstruct HDR images, yielding remarkable outcomes. However, due to the high computational cost of diffusion models along with the requirement for numerous iterative steps, inference speed is limited. Despite efforts to mitigate this constraint [2, 17, 27], the overall complexity remains high, especially for high-resolution images commonly encountered in HDR scenarios. Recently, several methods [3, 23, 39] have addressed this issue by conducting diffusion modeling in the latent space. Rombach et al. [23] employed an autoencoder model to compress images into features equivalent to the image space. HI-Diff [3] leverages diffusion-based vector features for assisting image deblurring. However, given the unique characteristics of HDR imaging, further exploration is needed to select an appropriate latent space for HDR reconstruction tasks.", + "bbox": [ + 496, + 301, + 892, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 500, + 617, + 638, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The diffusion model introduced by Sohl-Dickstein et al. [26] is inspired by nonequilibrium thermodynamics. Here, we provide a brief overview of the \"variance-preserving\" diffusion model from [8], which encompasses diffusion processes and reverse processes.", + "bbox": [ + 496, + 642, + 890, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion Process. Given a clean image distribution, the diffusion process gradually injects isotropic Gaussian noise to generate $x_{t}$ according to a variance schedule", + "bbox": [ + 496, + 718, + 890, + 762 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\beta_ {1}, \\dots , \\beta_ {T} \\in (0, 1). \\text {L e t} \\alpha_ {t} = 1 - \\beta_ {t}, \\bar {\\alpha} _ {t} = \\prod_ {i = 1} ^ {T} \\alpha_ {i}:\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 763, + 857, + 780 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(x _ {t} \\mid x _ {0}\\right) = \\mathcal {N} \\left(x _ {t}; \\sqrt {\\bar {\\alpha} _ {t}} x _ {0}, (1 - \\bar {\\alpha} _ {t}) \\mathbf {I}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 789, + 890, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The closed-form for this process can be represented as:", + "bbox": [ + 500, + 816, + 864, + 832 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t} \\left(x _ {0}, \\epsilon_ {t}\\right) = \\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t}, \\epsilon \\in (0, \\mathbf {I}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 544, + 842, + 890, + 859 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $x_{t}$ represents the output at time $t$ , $x_0$ is the initial clean image, and $\\epsilon_t$ denotes isotropic Gaussian noise.", + "bbox": [ + 496, + 869, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "25733", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f5af3edc5e26d8545e8d3d2375df16679a33e8746d850afd5e159841d0002b6c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 89, + 521, + 324 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9d3c16e98df4a6cc282eca153975e7f37b222b385af8565e2503f392c9b0e05d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 527, + 90, + 888, + 324 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/763c7d3917bd418d786c515129ce3dd3e733dcc7876e9ae5c25797e8bca0da87.jpg", + "image_caption": [ + "Figure 2. The proposed LF-Diff comprises DHRNet, LPRNet, and a denoising network. LF-Diff undergoes two training stages: Pretrain LF-Diff (Sec. 4.1) and DM training (Sec. 4.2). Notably, during the inference stage, we do not input the ground-truth image into $\\mathrm{LPONet}_{DM}$ and the denoising networks. Instead, we solely utilize the reverse process of DMs." + ], + "image_footnote": [], + "bbox": [ + 80, + 327, + 893, + 352 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Reverse Process. The reverse process attempts to approximate the data distribution $q(x_0)$ by a Markovian process, starting with random Gaussian noise $x_{T} = \\mathcal{N}(x_{T};\\mathbf{0},\\mathbf{I})$ :", + "bbox": [ + 76, + 433, + 468, + 479 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(x _ {t - 1} \\mid x _ {t}, x _ {0}\\right) = \\mathcal {N} \\left(x _ {t - 1}; \\tilde {\\mu} _ {t} \\left(x _ {t}, x _ {0}\\right), \\tilde {\\beta} _ {t} \\mathbf {I}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 489, + 468, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\tilde{\\mu}_t$ and $\\tilde{\\beta}_t$ are distribution parameters. Since the true reverse process Eq.(3) relies on $q(x_0)$ and is intractable, the neural network $f_{\\theta}$ serves as the denoiser to estimate $p_{\\theta}(x_{t - 1}|x_t,t)$ instead of $q(x_{t - 1}|x_t,x_0)$ :", + "bbox": [ + 76, + 520, + 468, + 583 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} p _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}, t\\right) = \\mathcal {N} \\left(x _ {t - 1}; \\mu_ {\\theta} \\left(x _ {t}, t\\right), \\Sigma_ {\\theta} \\left(x _ {t}, t\\right)\\right) \\tag {4} \\\\ = q \\left(x _ {t - 1} \\mid f _ {\\theta} (x _ {t}, t)\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 592, + 468, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mu_{\\theta}$ and $\\Sigma_{\\theta}$ are the parameters to be estimated in the reverse process. Let fixed variances $(\\Sigma_{\\theta}(x_t,t) = \\tilde{\\beta}_t\\mathbf{I})$ , the optimization objective can be defined as:", + "bbox": [ + 76, + 638, + 468, + 684 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {t, x _ {0}, \\epsilon_ {t}} [ \\| \\epsilon_ {t} - f _ {\\theta} (\\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t}, t) \\| ^ {2} ]. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 694, + 468, + 712 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Methodology", + "text_level": 1, + "bbox": [ + 76, + 724, + 210, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As depicted in Fig. 2, we introduce the proposed LF-Diff framework, which consists of two training stages. During the first stage, we pre-train LF-Diff to extract accurate low-frequency prior representations from ground truths using LPenet, while DHRNet learns how to utilize the above priors to reconstruct HDR images. During the second stage, we train the DM to accurately predict the low-frequency prior representations directly from LDR images, and jointly optimize it with DHRNet to generate final high-quality HDR images.", + "bbox": [ + 75, + 750, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Stage One: Pretrain LF-Diff", + "text_level": 1, + "bbox": [ + 500, + 431, + 756, + 448 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the first stage, our objective is to train the LPERNet how to extract accurate low-frequency prior features from ground truth and embed this representation into the DHRNet to guide the reconstruction process. As illustrated in Fig. 2 (a), we utilize the Prior Integration Module (PIM) to incorporate these prior features into the DHRNet via the cross-attention mechanism. Within DHRNet, multiple Feature Refinement Modules (FRM) are stacked to directly learn the mapping relationship from input features to the target image distribution, while PIM leverages prior features to assist in recovering the low-frequency content of the reconstructed images. Next, we elaborate on above descriptions.", + "bbox": [ + 496, + 455, + 890, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Low-frequency Prior Extraction Network. The LPENet has a straightforward structure, as shown in Fig. 2 (a), comprising several residual blocks designed to extract the Low-frequency Prior Representation (LPR). Starting with the ground truth image $I_{gt} \\in \\mathbb{R}^{H \\times W \\times 3}$ , we initially perform a tonemapping operation $\\mathcal{T}(\\cdot)$ to obtain the LDR domain ground truth image $\\mathcal{T}(I_{gt})$ . We concatenate these images along the channel dimension and employ the PixelUnshuffle operation to downsample, generating the input for LPENet. Subsequently, LPENet extracts the LPR $z \\in \\mathbb{R}^{\\frac{H}{k} \\times \\frac{W}{k} \\times 3}$ as follows:", + "bbox": [ + 496, + 637, + 892, + 801 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nz = \\operatorname {L P E N e t} (\\operatorname {P i x i l e U n s h u r f l e} (\\operatorname {C o n c a t} (I _ {g t}, \\mathcal {T} (I _ {g t})))) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 811, + 890, + 828 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{T}(x) = \\frac{\\log(1 + \\mu x)}{\\log(1 + \\mu)}$ denotes the tonemapping operator, $\\mu = 5000$ is a predefined constant parameter that controls the degree of compression applied to the input signal, $k$ denotes the sampling multiple of PixelUnshuffle.", + "bbox": [ + 496, + 835, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "25734", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dynamic HDR Reconstruction Network. As illustrated in Fig. 2 (a), DHRNet consists of multiple stacked Reconstruction Blocks, with each block comprising one Prior Integration Module (PIM) and $N$ Feature Refinement Module (FRM). PIM fuses the LPR with intermediate features of DHRNet, we will give a detailed description in the following part. FRM is designed based on the understanding that natural images encompass diverse frequencies, since high and low frequencies serve distinct roles in image encoding.", + "bbox": [ + 75, + 90, + 472, + 241 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Fig. 3 (a), the structure of FRM involves several blocks. Specifically, we begin by processing the features through a residual block to obtain $F_{n-1}$ . Subsequently, we decouple these features into two components: high-frequency features $F_{n-1}^{\\mathrm{high}}$ , which remain resolution-invariant, and low-frequency features $F_{n-1}^{\\mathrm{low}}$ with reduced resolution, following the approach described in [18]:", + "bbox": [ + 75, + 243, + 472, + 349 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {n - 1} ^ {\\text {l o w}} = \\operatorname {A v g p o o l} \\left(F _ {n - 1}, k\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 362, + 344, + 378 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {n - 1} ^ {\\text {h i g h}} = F _ {n - 1} - \\text {U p s a m p l e} \\left(F _ {n - 1} ^ {\\text {l o w}}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 377, + 468, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $k$ denotes the kernel size of the avgpooling layer, Upsample denotes the upsampling operator. In essence, low-frequency features capture global dependencies within the input (image/features) and do not necessitate a high-resolution feature map, but require global attention. On the other hand, high-frequency features capture detailed local information, requiring a high-resolution feature map, and only can be achieved through local operators. Therefore, we employ the residual block (RB) to process $F_{n - 1}^{\\mathrm{high}}$ and utilize the self-attention (SA) from [45] to handle $F_{n - 1}^{\\mathrm{low}}$ . Subsequently, we fuse the low- and high-frequency features to preserve the initial details, resulting in the feature $F_{n - 1}^{\\prime}$ . This operation can be represented as follows:", + "bbox": [ + 75, + 414, + 472, + 612 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {n - 1} ^ {\\prime} = \\operatorname {C o n c a t} \\left[ R B \\left(F _ {n - 1} ^ {\\text {h i g h}}\\right), \\operatorname {U p s a m p l e} \\left(S A \\left(F _ {n - 1} ^ {\\text {l o w}}\\right)\\right) \\right]. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 83, + 622, + 468, + 656 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Considering that $F_{n-1}^{\\prime}$ is a concatenation of two features, a $1 \\times 1$ Conv is employed to effectuate a reduction in the number of channels. This downsizing step is followed by a channel attention module, designed to accentuate those channels exhibiting heightened activation values.", + "bbox": [ + 75, + 657, + 468, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In order to effectively merge the prior feature derived from the LPERNet with the intermediate feature produced by DHRNet, we replaced the self-attention mechanism in the low-frequency branch of FRM with a carefully designed cross-attention mechanism, resulting in the Prior Integration Module (PIM). As depicted in Fig. 2 (a), a PIM is positioned before each reconstruction block. Within each PIM, cross-attention computations are performed between the prior and intermediate features to facilitate feature fusion. This module enables the aggregation of information from the prior feature into the features of DHRNet.", + "bbox": [ + 75, + 734, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/126d2748d194f331ae8b4872766c520c57b2da0b4a8fee7f036be1a223e945bd.jpg", + "image_caption": [ + "(a) Feature Refinement Module" + ], + "image_footnote": [], + "bbox": [ + 504, + 89, + 890, + 207 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6dcd4f8b73cc7dc8103466d130b2ae4c17a8181ea480aa1e990d7930357d194e.jpg", + "image_caption": [ + "(b)Cross Self-attention module from PIM", + "Figure 3. The architecture of (a) Feature Refinement Module (FRM) and (b) Cross-attention mechanism." + ], + "image_footnote": [], + "bbox": [ + 506, + 220, + 874, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, as illustrated in Fig. 3 (b), our cross-attention module receives two types of input: the intermediate feature $\\mathbf{F} \\in \\mathbb{R}^{\\hat{H} \\times \\hat{W} \\times \\hat{C}}$ and the prior feature $z \\in \\mathbb{R}^{\\hat{H} \\times \\hat{W} \\times \\hat{N}}$ . We project $\\mathbf{F}$ into a query vector $\\mathbf{Q} = W_d^Q W_c^Q\\mathbf{F}$ using point-wise $1 \\times 1$ convolution and depth-wise $3 \\times 3$ convolution with weights $W_d^Q$ and $W_c^Q$ . Similarly, $z$ is transformed into key $\\mathbf{K}$ and value $\\mathbf{V}$ through analogous operations. Next, we reshape these projections into matrices suitable for attention computation: $\\hat{\\mathbf{Q}} \\in \\mathbb{R}^{\\hat{H}\\hat{W} \\times \\hat{C}}$ , $\\hat{\\mathbf{K}} \\in \\mathbb{R}^{\\hat{N} \\times \\hat{H}\\hat{W}}$ , and $\\hat{\\mathbf{V}} \\in \\mathbb{R}^{\\hat{H}\\hat{W} \\times \\hat{N}}$ . We then calculate a more computationally efficient attention map $A \\in \\mathbb{R}^{\\hat{N} \\times \\hat{C}}$ by performing dot product between $\\hat{\\mathbf{Q}}$ and $\\hat{\\mathbf{K}}$ . The process can be described as follows:", + "bbox": [ + 498, + 412, + 893, + 619 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {F}} = W _ {c} \\hat {\\mathbf {V}} \\cdot \\operatorname {S o f t m a x} (\\hat {\\mathbf {K}} \\cdot \\hat {\\mathbf {Q}} / \\gamma) + \\mathbf {F}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 626, + 890, + 646 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\gamma$ is a trainable parameter. Notably, we do not implement multi-head attention in PIM.", + "bbox": [ + 498, + 654, + 890, + 683 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training Strategy. As proposed in [11], we initially transform the provided set of input LDR images $\\{L_1, L_2, \\ldots, L_N\\}$ into their corresponding HDR versions $H_i$ using gamma correction. Subsequently, we concatenate each $L_i$ with its corresponding $H_i$ along the channel axis to generate six-channel input tensors $X_i = [L_i, H_i]$ . Similar to [16, 41, 44], we employ an alignment module [41] to process the input LDR images, align the features implicitly, and feed them into DHRNet. Then, the LPR feature $z$ extracted from LPENet is injected into DHRNet via PIM. The reconstructed HDR image $\\hat{H}$ is generated by:", + "bbox": [ + 498, + 684, + 890, + 851 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {H} = \\operatorname {C o n v} _ {3 \\times 3} \\left(\\mathrm {D H R N e t} \\left(X _ {i}, z\\right)\\right). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 858, + 890, + 877 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following previous approaches [16, 41], we utilize both", + "bbox": [ + 500, + 885, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "25735", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tonemapped per-pixel loss and perceptual loss as the image reconstruction loss function $\\mathcal{L}_r$ . This dual-loss strategy optimizes both pixel-level accuracy and high-level feature representations in the generated HDR results:", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r} = \\left\\| \\mathcal {T} (H) - \\mathcal {T} (\\hat {H}) \\right\\| _ {1} + \\lambda \\| \\phi_ {i, j} (\\mathcal {T} (H)) - \\phi_ {i, j} (\\mathcal {T} (\\hat {H})) \\| _ {1}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 162, + 468, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\phi_{i,j}(\\cdot)$ signifies the $j$ -th convolutional feature extracted from the VGG19 network after the $i$ -th maxpooling operation, and $\\lambda = 1e - 2$ is a hyperparameter that balances the contribution of the each component.", + "bbox": [ + 75, + 186, + 468, + 247 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Stage Two: Diffusion Models for HDR Imaging", + "text_level": 1, + "bbox": [ + 76, + 255, + 468, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After the above-mentioned learning procedure, we now possess LDR images and their corresponding LPR $z$ . In the second stage (Fig. 2 (b, c)), our objective is to efficiently harness the powerful distribution estimation capability of the DMs. Specifically, we utilize the LPENet from stage one to capture the LPR as the denoising target of the DM. The DM will learn how to extract accurate LPR from the LDR inputs and perform joint optimization with DHRNet.", + "bbox": [ + 75, + 277, + 468, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Diffusion Model for LRP Learning. After capturing the LPR $z \\in \\mathbb{R}^{\\frac{H}{k} \\times \\frac{W}{k} \\times N}$ from the pretrained LPENet, we transform the clean LPR feature $z$ into a noisy version $z_T$ using Eq.(2), which can be described as:", + "bbox": [ + 75, + 398, + 468, + 459 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(z _ {T} \\mid z\\right) = \\mathcal {N} \\left(z _ {T}; \\sqrt {\\bar {\\alpha} _ {T}} z, (1 - \\bar {\\alpha} _ {T}) \\mathbf {I}\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 468, + 468, + 486 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $T$ represents the total number of iterations, while $\\bar{\\alpha}_t$ and $\\alpha$ denote the pre-defined variance schedule.", + "bbox": [ + 75, + 494, + 467, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the reverse process, we iteratively generate the LPR from a pure Gaussian distribution, starting from $z_{T}$ and moving backward to $z_{0}$ , which utilizes the posterior distribution described in Eq. (3). Following previous work [24, 44], we utilize a neural network to estimate $p_{\\theta}(x_{t - 1}|x_t,t)$ instead of $q(x_{t - 1}|x_t,x_0)$ for each step. Specifically, we first use LPENetDM to obtain a conditional feature $D\\in \\mathbb{R}^{\\frac{H}{k}\\times \\frac{W}{k}\\times N}$ from aligned LDR feature:", + "bbox": [ + 75, + 523, + 468, + 645 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {D} = \\operatorname {L P E N e t} _ {D M} (A M (\\text {P i x e l U n s h u f f l e} (X _ {i}))). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 654, + 468, + 671 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{LPENet}_{DM}$ maintains a similar architecture to $\\mathrm{LPENet}$ but with a modified input dimensionality for its first convolution layer, $AM(\\cdot)$ denotes alignment module from [41]. The denoising neural network predicts the noise based on both $z_{t}$ and the derived conditional feature $D$ , i.e., $\\epsilon_{\\theta}(Concat(z_{t},\\mathbf{D}),t)$ . Upon substituting this estimated noise term into the equation governing the reverse process Eq. (4), we arrive at the following sampling formula:", + "bbox": [ + 75, + 679, + 468, + 801 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {z} _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\boldsymbol {z} _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\boldsymbol {z} _ {t}, \\mathbf {D}, t\\right)\\right) + \\sqrt {1 - \\alpha_ {t}} \\boldsymbol {\\epsilon} _ {t}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 809, + 468, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\epsilon_t \\sim \\mathcal{N}(0, \\mathbf{I})$ . By iteratively sampling $z_t$ using Eq. (14) $T$ times, we progressively reconstruct the predicted LPR representation $z_0$ (i.e. $\\hat{z}) \\in \\mathbb{R}^{\\frac{H}{k} \\times \\frac{W}{k} \\times N}$ .", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Strategy. Traditional DMs solely learn the probability distribution by optimizing the weighted variational bound (Eq. (5)), resulting in a slight deviation between the predicted prior and the actual prior. Integrating the DM directly with DHRNet might lead to a misalignment issue, thereby hampering the overall image processing performance. To address this, we joint training the DM and DHRNet. During each training iteration, we first sample the noise sample $z_{t}$ through the diffusion process (Eq. (2)). Given that our denoising neural network is lightweight, we use a reverse process for $S$ iterations based on the DDIM strategy [27] to generate the predicted prior feature $\\hat{z}$ . This $\\hat{z}$ guides DHRNet via PIM. The loss function in the second stage consists of the reconstruction loss $\\mathcal{L}_r$ (Eq. (11)) and the diffusion loss $\\mathcal{L}_{diff}$ :", + "bbox": [ + 496, + 90, + 890, + 319 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {d i f f}} = E _ {z _ {0}, t, \\epsilon_ {t} \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I})} \\left[ \\| \\epsilon_ {t} - \\epsilon_ {\\theta} (z _ {t}, D, t) \\| ^ {2} \\right] \\tag {15} \\\\ + \\left\\| \\hat {z} - z \\right\\| _ {1}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 329, + 890, + 376 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Inference", + "text_level": 1, + "bbox": [ + 500, + 388, + 609, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the inference stage, $L P E N t_{D M}$ extracts a conditional feature $\\mathbf{D}$ from aligned LDR feature (Eq. (13)), and we randomly sample a Gaussian noise $\\hat{z}_T$ . Then, denoising neural network utilizes the $\\hat{z}_T$ and $\\mathbf{D}$ to estimate LPR after $S$ iterations (Eq. (14)) based on DDIM [27]. After that, DHRNet exploits the LPR to restore HDR images as Eq. (10). More details about the DM training and the model inference can be found in the supplementary material.", + "bbox": [ + 496, + 412, + 890, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 547, + 632, + 565 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 500, + 574, + 705, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. All methods are trained using two publicly available datasets, employing identical training settings: Kalantari's dataset [11] and Hu's dataset [10]. Kalantari's dataset consists of 74 samples for training and 15 for testing, all captured under authentic environmental conditions. Each sample comprises three LDR images with exposure variations of $\\{-2,0, +2\\}$ or $\\{-3,0, +3\\}$ . In contrast, Hu's dataset is a synthetic dataset designed to emulate sensor realism, generated through a game engine. This dataset contains images captured at three distinct exposure levels $\\{-2,0, +2\\}$ , with our primary focus on dynamic scene images within this collection. Following the settings outlined in [10], we allocate the initial 85 samples for training, reserving the remaining 15 for testing. Moreover, to further validate the model's generalizability, we incorporate Sen et al.'s dataset [25] and Tursun et al.'s dataset [32] exclusively for qualitative assessment, as these datasets lack ground truths.", + "bbox": [ + 496, + 598, + 890, + 853 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. To conduct a quantitative comparison, we utilize five objective metrics: PSNR- $\\mu$ SSIM- $\\mu$ PSNR-L,SSIM-L,and HDR-VDP-2 [20].Here, the sub", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "25736", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7b7857606bd0387f1c721bafcfef56959bf9e71469029765249756ee6eb6e7fa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 89, + 155, + 207 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2ddb4897fdb8a274acf484ecc6b8847b275f9fd2e985b618a01f06da060e9562.jpg", + "image_caption": [ + "Tonemapped Result" + ], + "image_footnote": [], + "bbox": [ + 156, + 90, + 382, + 207 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/889b8682a37653b8985fd7d0b20754ddc71e43659c9b840ba5702687bf22c87e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 90, + 433, + 205 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7b7de5c855ab8fb37681575c80f9ca03faed921ff5a5333320e26564c373e28b.jpg", + "image_caption": [ + "LDR-Patches" + ], + "image_footnote": [], + "bbox": [ + 434, + 90, + 483, + 207 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 81, + 217, + 130, + 255 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e5705c9ab45703d500c27cd44c24c2757207eb7d859ee6ffbf259d5deefe8a8a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 133, + 217, + 183, + 255 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 184, + 217, + 233, + 255 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/088accdbb9e4ef077f43ecdb11edb2ebcc7cb42b5980ea1ad085efb213a8c184.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 217, + 284, + 255 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3177ca47349bee3ca18d8dbbdd0855f0f3e8616b808cc2f477c234b9742e5025.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 285, + 217, + 334, + 255 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 334, + 217, + 383, + 255 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 385, + 217, + 434, + 255 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 434, + 217, + 483, + 255 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 83, + 257, + 130, + 294 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 133, + 257, + 181, + 294 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 183, + 257, + 233, + 294 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 233, + 257, + 282, + 294 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 284, + 257, + 333, + 294 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 334, + 257, + 383, + 294 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 385, + 257, + 434, + 294 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 434, + 257, + 483, + 294 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fef82b101c723bff618b0870e9921d1bba6362a8ba934fe5ff5e3bf33bb18e48.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 90, + 562, + 205 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/131aaf67b61e8a48f7b2cce5d3631d67b178bc367dfad6df2026b7346ba180f8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 563, + 90, + 637, + 205 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7b3395093c25bafede3c26ce245f8d8462e6f28c91f593463313c77a07a18799.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 90, + 787, + 205 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/46fc7cf1113732f5c64d2766ba8f62317b0f85f3da4600ee1419c56e5fb416b2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 790, + 90, + 839, + 205 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/56688c7157a7627407afdcc270e999c9dde76315e5d91451537382a9655bfdc0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 839, + 90, + 888, + 205 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/691d89c830fe024b545b00633e37f4b33cd43524b16ff0ee6e208c778715c351.jpg", + "image_caption": [ + "Tonemapped Result" + ], + "image_footnote": [], + "bbox": [ + 718, + 207, + 787, + 215 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 81, + 217, + 130, + 255 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 133, + 217, + 181, + 255 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9dc4e1ba591d95bff701a7ad0835c8f7573887aa6b15cbc9841b145facef45be.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 183, + 217, + 233, + 255 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 233, + 217, + 282, + 255 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7161664df37b8d64e4b817c41e42fc2a31e91a5c2a2dd850c9fef3b569f1c093.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 284, + 217, + 333, + 255 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 334, + 217, + 383, + 255 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c5bc93054f887c1c909db2dd0fac183d5cedd6d78bc4571b862847fab8db8b2a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 217, + 434, + 255 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 434, + 217, + 483, + 255 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 83, + 257, + 130, + 294 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 133, + 257, + 181, + 294 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bbcda58e6437ff10fdc6e3ac16378bed102a425af2882c90586e8f5ff0286ef1.jpg", + "image_caption": [ + "(a) Example from Kalantari et al.'s dataset [11]", + "Figure 4. Visual comparisons are conducted on testing data from various datasets, focusing on zoomed-in local areas of the HDR images estimated by our method and the compared techniques. Our model demonstrates the ability to generate HDR images of superior quality." + ], + "image_footnote": [], + "bbox": [ + 183, + 257, + 233, + 295 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 233, + 257, + 282, + 294 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 284, + 257, + 333, + 294 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 334, + 257, + 383, + 294 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1f80ff558967179b3390143af807153da5a5a44b924e3c870b710b695f0f5780.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 257, + 434, + 294 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 434, + 257, + 483, + 294 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c49a2a7497c5df59e526d4b5999d4bc1b08176031adeb3801527a00983eed83b.jpg", + "image_caption": [ + "LDRs" + ], + "image_footnote": [], + "bbox": [ + 485, + 217, + 535, + 294 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/89028f331197adeb1d139b87b4ba5240d9a7ff4f99e753d1da267601d3bdf03a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 535, + 217, + 586, + 292 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/18d5893b6e67f0908cf482fc1ab24ea5a37441c4c16b66e47659524d874ced7a.jpg", + "image_caption": [ + "AHDR HDR-GAN CA-VIT HyHDR Ours" + ], + "image_footnote": [], + "bbox": [ + 588, + 217, + 637, + 292 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d087c4d450c606f3d187eeb38e7284ac40aabc114f7eb8e8027546e9d4bb2253.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 217, + 687, + 292 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4e08ab3f33793659f9087c10abba19ba314bb40aa1b2d46e1e4b593769eb53f3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 217, + 738, + 292 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d0d16a55a034f2984ff6cbc6251a7cbd754894bd5e9b20c54ec087c79548b88d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 738, + 217, + 787, + 292 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d216f66dd070b38e234aebae39952d2ea1f72aa8374811f5fa8ecc61a78dfe8d.jpg", + "image_caption": [ + "LDR-", + "GT" + ], + "image_footnote": [], + "bbox": [ + 789, + 217, + 836, + 292 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/73ff7f8951f798bf64fa368fa3f80ecd74113a5823cacba4897d48e5d7de1e76.jpg", + "image_caption": [ + "Patches" + ], + "image_footnote": [], + "bbox": [ + 838, + 217, + 888, + 292 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "scripts $\\mu$ and L denote that these metrics are computed in the tonemapped and linear domains, respectively.", + "bbox": [ + 75, + 383, + 467, + 412 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. Our implementation is conducted using PyTorch, and each training stage converges after 300 epochs on four NVIDIA A100 GPUs. We employ the Adam optimizer with an initial learning rate of 1e-4, which decays by a factor of 0.1 after every 50 epochs. The training dataset is processed by cropping $128 \\times 128$ patches with a stride of 64, and the batch size is set to 64. For LF-Diff, we adopt a variant of the commonly used U-Net architecture from [19] as the denoising network, with 3 levels of blocks $\\{2, 2, 2\\}$ . During training, the time step $T$ is set to 200, and the implicit sampling step $S$ is set to 10 for both training and inference phases to achieve efficient restoration. In DHRNet, the parameters $N_{i} \\in \\{L_{1}, L_{2}, L_{3}\\}$ are set to $\\{3, 3, 3\\}$ , and the channel $C$ is set to 60. The number of attention heads for the self-attention branch at the same level is set to $\\{6, 6, 6\\}$ , and the channel expansion factor in FFN is 2.66. The kernels for avgpool in PIM and FRM are set to 4 and 2, respectively. For LPENet, it comprises 4 residual blocks, with a Pixelunshuffle downsampling factor of 4 and the output channel of 3.", + "bbox": [ + 75, + 412, + 468, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Comparison with the State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 76, + 726, + 468, + 741 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we evaluate the performance of the proposed LF-Diff method and present experimental results to validate its reconstruction performance compared to state-of-the-art techniques. Specifically, we evaluate LF-Diff against two patch-based methods [9, 25], a optical flow-based method [11], five CNN-based methods [1, 15, 21, 38, 41], two ViT-based methods [16, 43], along with a diffusion-based method [44]. It's worth noting that all deep learning methods employ the same training dataset and settings for consistency.", + "bbox": [ + 75, + 750, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets w/ Ground Truth. The quantitative outcomes for LF-Diff on two datasets are presented in Tab. 1. Our method is compared with various state-of-the-art approaches using the testing data from [11] and [10], which includes challenging samples characterized by saturated backgrounds and foreground motions. All quantitative results are averaged across testing images. Notably, LF-Diff exhibits a significant improvement over other leading methods, exceeding the performance of the DM-based method DiffHDR [44], by $0.65\\mathrm{dB}$ and $0.86\\mathrm{dB}$ in PSNR- $\\mu$ and PSNR-L metrics, respectively, based on Kalantari's dataset [11]. Moreover, LF-Diff showcases superior performance compared to the runner-up method, HyHDR [43], with improvements of $0.28\\mathrm{dB}$ and $0.19\\mathrm{dB}$ in PSNR- $\\mu$ and PSNR-L metrics, respectively, on Hu's dataset [10].", + "bbox": [ + 496, + 382, + 890, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Fig. 4 (a) and (b), the datasets present some challenging samples due to information loss in the LDR images. Most existing approaches struggle in these areas, producing ghosting artifacts due to large motions and occlusions. Kalantari's method [11] and DHDR [38] struggle with background motion due to error-prone alignments (e.g., optical flow), resulting in undesirable ghosting. While HDR-GAN [21] exhibits noticeable ghosting artifacts around the balcony area and introduces erroneous color information, AHDR [41] aligns LDR images using convolutional spatial attention. However, it unintentionally suppresses valuable contextual information and encounters difficulties with significant motions in over/underexposed scenes. CA-ViT [16], relying on patch-based sampling, produces noticeable blocky ghosting. In contrast, aided by the DM, both Diff-HDR [44] and our LF-Diff generate HDR images aligned with human perception. Notably, our approach outperforms Diff-HDR with faster inference speeds and a smaller computational overhead.", + "bbox": [ + 496, + 613, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "25737", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5d31bc04308f765093082c38c427f59d11e4e137a5bf02d46c7d039ce4b749a4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 89, + 158, + 209 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6a2573a58dbcf7e471412254c872a0162cb38a7aa750162fc101fb43b9e4b1a8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 158, + 90, + 383, + 209 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/16c62efe2965acc99aa264c8d8da1a239cf5b08f6d6fc64808404db4e23b8cde.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 385, + 90, + 431, + 209 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bf4603127f35737ccdb178cb012c6ae5d264b0caa86f767c2a8ebb899f6d8545.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 431, + 90, + 482, + 209 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/04d1f1950ced272c3d5f51529738b1d2b5d364fc85bd739d80e6deca2f98ba7a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 90, + 560, + 209 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/396a71f5a929f18bf1a324b9b1df57f78866a3856be2ff19d6a5be690e50bb8f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 560, + 90, + 785, + 209 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dd9893aaef9588271472189f5e18d5bb86322584c41bf05f20d68de3fb338e0e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 787, + 90, + 862, + 209 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/17afcbe2cb497284b3f30b74acac5519bb6463ec40e7cf9a1c06b430ca8286cf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 862, + 90, + 888, + 209 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8b7847e3449307c3911ae8b411c03c9ac7c7f756f25c3c08295bae7d0065a8b0.jpg", + "image_caption": [ + "Kalantari", + "DHDR" + ], + "image_footnote": [], + "bbox": [ + 83, + 210, + 130, + 292 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a8d44f76469c82af20462c1db28f657f1c46f503205e43b633d1013798371fc4.jpg", + "image_caption": [ + "AHDR" + ], + "image_footnote": [], + "bbox": [ + 130, + 210, + 181, + 292 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f8008a2662f63ce3e703b8104efb48464fc62204ff690fc843590e6f5f8a8730.jpg", + "image_caption": [ + "HDR-GAN" + ], + "image_footnote": [], + "bbox": [ + 181, + 210, + 230, + 292 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9a9d5f0f8a02a48f5ee01093fc4c0d1ffe20c2aab2e4e0f4f8037a69fca5eb72.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 230, + 210, + 282, + 292 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/75f759117a9c60caa16f831e75a98005ca75150ede5b07bd238cbd26c49fc396.jpg", + "image_caption": [ + "APNET", + "Figure 5. Visual comparisons on the datasets without ground truth." + ], + "image_footnote": [], + "bbox": [ + 284, + 210, + 331, + 292 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bfbeddcc4ec3f11838d3a83d5468d17a3ff7d5aafa61ff021cd0a3eb5bd06430.jpg", + "image_caption": [ + "CA-VIT" + ], + "image_footnote": [], + "bbox": [ + 333, + 210, + 383, + 292 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f74c6febd42f3236418fe6f283cfd1bba2c55ecdf26d8de4d319f34b0a1dde8e.jpg", + "image_caption": [ + "DiffHDR" + ], + "image_footnote": [], + "bbox": [ + 385, + 210, + 433, + 292 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/612eb62b9ef537b2388e0ffc764c96efd631edab9ad5f07a4a10c94ec3a65b0e.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 434, + 210, + 483, + 292 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7e87e3b30d5113fb1c123b6f793b4f0d024b07715c50e9145cf9c61330602014.jpg", + "image_caption": [ + "Kalantari" + ], + "image_footnote": [], + "bbox": [ + 483, + 210, + 532, + 292 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8b10266a3e0467f31acaa4a557a5f26717be9448174efcdfd44675d9fd2d4867.jpg", + "image_caption": [ + "DHDR" + ], + "image_footnote": [], + "bbox": [ + 534, + 210, + 584, + 292 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/91c8c49197b7d0e472c0aa5f21e7b23fff14ba29385331eb347305d0135e7a1b.jpg", + "image_caption": [ + "AHDR HDR-GAN" + ], + "image_footnote": [], + "bbox": [ + 586, + 210, + 635, + 292 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3274dbf0b58470e46abb7c1349f0e3f8667c94f11a9ed0b9b320b3cb01a7d372.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 210, + 684, + 292 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "HDR-GAN", + "bbox": [ + 638, + 292, + 689, + 300 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0b8682997f38e13946d1d6a902a895178623bfbb8bcfeabebf8e7b7d89d105f5.jpg", + "image_caption": [ + "NET " + ], + "image_footnote": [], + "bbox": [ + 689, + 210, + 736, + 292 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9f2f0efdf7068ec2a68408ca3044c9bea2d24ca535156cd1a80795533d3767cc.jpg", + "image_caption": [ + "CA-VIT" + ], + "image_footnote": [], + "bbox": [ + 738, + 210, + 787, + 292 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a3266d1c8cc0965083a3d701884ed52827db4888a09b02b75b18d1c29b092b87.jpg", + "image_caption": [ + "DiffHDR" + ], + "image_footnote": [], + "bbox": [ + 787, + 210, + 836, + 292 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5b10815ac919c9c7b3527fea620e87bc7bb3d206dc75870d1f22338e70ba09e3.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 839, + 210, + 872, + 292 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ce1d27fb9bf3a718f3b109d723bd7dcef4a5b9e99a3a48cc8e8f1a1cd544f276.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetsModelsSen[25]Hu[9]Kalantari[11]DHDR[38]AHDR[41]NHDRR[42]HDRGAN[21]APNT[1]CA-ViT[16]HyHDR[43]DiffH[44]Ours
KalantariPSNR-μ40.9532.1942.7441.6443.6242.4143.9243.9444.3244.6444.1144.76
PSNR-L38.3130.8441.2240.9141.0341.0841.5741.6142.1842.4741.7342.59
SSIM-μ0.98050.97160.98770.98690.99000.98870.99050.98980.99160.99150.99110.9919
SSIM-L0.97260.95060.98480.98580.98620.98610.98650.98790.98840.98940.98850.9906
HDR-VDP-255.7255.2560.5160.5062.3061.2165.4564.0566.0366.0565.5266.54
DatasetsModelsSen[25]Hu[9]Kalantari[11]DHDR[38]AHDR[41]NHDRR[42]HDRGAN[21]APNT[1]ADNet[15]CA-ViT[16]HyHDR[43]Ours
HuPSNR-μ31.4836.5641.6041.1345.7645.1545.8646.4146.7948.1048.4648.74
PSNR-L33.5836.9443.7641.2049.2248.7549.1447.9750.3851.1751.9152.10
SSIM-μ0.95310.98240.99140.98700.99560.99560.99450.99530.99080.99470.99590.9968
SSIM-L0.96340.98770.99380.99410.99800.99810.99890.99860.99870.99890.99910.9993
HDR-VDP-266.3967.5872.9470.8275.0474.8675.1973.0676.2177.1277.2477.35
", + "bbox": [ + 83, + 356, + 890, + 511 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. The evaluation results on Kalantari's [11] and Hu's [10] datasets. The best results are highlighted in Bold.", + "bbox": [ + 143, + 521, + 821, + 536 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Datasets w/o Ground Truth. To assess the generalization capability of our proposed HDR imaging method, we evaluated the performance of the model trained on Kalantari et al.'s Dataset [11] by testing it on datasets from Sen et al. [25] and Tursun et al. [33], which lack ground truths. In Fig. 5 (a), numerous current methods encounter difficulties in restoring recover the large saturated region and large motion. Conversely, in Fig. 5 (b), while complete elimination of ghosting remains elusive for any method, our approach notably enhances image sharpness and detail compared to preceding techniques.", + "bbox": [ + 75, + 559, + 468, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Computational Budgets. We also conducted comparisons regarding model parameters and inference times with previous approaches. As illustrated in Table 2, patch match-based methods [9, 25] exhibit significantly longer inference times due to their CPU-based computation. Kalantari [11] requires considerable time for initial optical flow alignment. NHDRRNet [42], employing a U-shaped network architecture, demonstrates shorter inference times but substantially higher parameter counts compared to other methods. CAViT [16] has a large number of standard transformer blocks leading to high computational cost despite fewer parame", + "bbox": [ + 75, + 734, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ters. DiffHDR [44] has high inference time and parameter count due to reconstructing HDR from pure noise. In comparison, our method effectively leverages the powerful distribution estimation capability of DM with orders of magnitude fewer parameters and computations.", + "bbox": [ + 496, + 559, + 890, + 635 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 645, + 653, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we investigate the impact of various designs within our proposed method. All experiments are conducted using the Kalantari's dataset [11].", + "bbox": [ + 496, + 670, + 890, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effects of Diffusion Prior. As shown in Tab. 3, we establish a regression-based baseline model without LPR generation. In this configuration, the PIM in DHRNet is replaced with FRM, and HDR images are reconstructed through a regression-based approach. LF-Diff $_{s1}$ is a pretrained model in the first stage utilizing ground-truth images to provide LPR. Compared to the baseline, this brings a 8.7dB PSNR-L improvement, demonstrating that an accurate compact prior can greatly enhance results. Variant LF-Diff $_{s2}$ w/o DM does not use the DM, but instead directly learns the LPR using LPENet. This achieves a 0.16dB PSNR gain over the baseline. When learning the LPR based", + "bbox": [ + 496, + 719, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "25738", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c12eed3086646bc4e2439f4d9f3ece2ec4f8f98124d00b61555cec998eca5498.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Methods\nEnvironmentSen [25]\n(CPU)Hu [9]\n(CPU)Kalantari [11]\n(CPU+GPU)AHDR [41]\n(GPU)NHDRR [42]\n(GPU)HDRGAN [21]\n(GPU)APNT [1]\n(GPU)CA-ViT [16]\n(GPU)DiffHDR [44]\n(GPU)Ours\n(GPU)
Times(s)61.81s79.77s29.14s0.30s0.31s0.29s2.62s5.34s7.53s0.72s
Para.(M)--0.30M1.24M38.10M2.56M37.98M1.22M74.99M7.48M
", + "bbox": [ + 99, + 88, + 874, + 150 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "on the DM, LF-Diff $_{s2}$ w/ DM further improves PSNR by 0.24dB compared to variant LF-Diff $_{s2}$ w/o DM. This indicates that the DM demonstrates superior capability in accurately estimating the distribution for LPR prediction. Additionally, compared to DiffHDR, which requires over 70M Params for estimating complete images, LF-Diff only adds 2.39M Params over the baseline. Moreover, we present visual comparisons of the baseline (without the prior feature) and LF-Diff variants in Fig. 6. It can be observed that performing DM in a compact latent space to predict LPR effectively mitigates ghosting issues.", + "bbox": [ + 75, + 196, + 472, + 363 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effects of Joint Training Strategy. We conducted an ablation study on the joint training strategy. Under this strategy, only the DM is optimized in the second stage, referred to as Split-Training. Specifically, we first utilize the pretrained LPERNet to generate the prior feature $z$ from the ground truth, and then apply the training objective defined in Eq. (15) to independently train the DM. Subsequently, the trained DM is directly integrated with the DHRNet for evaluation. It's worth noting that the DHRNet in the second stage utilizes pre-trained weights from the stage one without additional training. It is evident that LF-Diff significantly outperforms Split-Training by 1.51 dB in terms of PSNR value. This comparison underscores the effectiveness of joint training strategy over Split-Training.", + "bbox": [ + 75, + 382, + 470, + 594 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Sampling Step. In addition to the lightweight DM structure and compact feature dimensions, our method achieves greater computational efficiency by employing a smaller sampling step, denoted as S, in the reverse process based on DDIM [27]. The performance of our method is demonstrated with different values of S ranging from 5 to 100, as presented in Tab. 4. Variations in sampling steps, when S is set to a value smaller than the training configuration, a noticeable performance degradation occurs. On the other hand, when S is larger than the training setting, the impact on performance is minimal, particularly with the SSIM metric remaining consistently unchanged. While a larger sampling step is known to enhance the visual quality of images in diffusion-based methods [24, 44], in our case, it primarily increases inference time. This observation suggests that larger sampling steps, such as $\\mathrm{S} = 1000$ in DiffHDR [44] and SR3 [24], may not be essential for diffusion-based HDR reconstruction and other related low-level vision tasks when using our proposed framework.", + "bbox": [ + 75, + 613, + 472, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d0e6deb6f3b38e8a9f489dc2ef3f1b59f095a15081fa9b1af298e526ee50b2be.jpg", + "table_caption": [ + "Table 2. Average runtime performance of various methods on the testing set with dimensions of ${1000} \\times {1500}$ ." + ], + "table_footnote": [], + "table_body": "
MethodGTDMJoint TrainingPara.PSNR-μPSNR-L
BaselineXXX5.09M44.2342.19
LF-Diffs1✓XX5.24M50.4250.89
Split trainingX✓X7.48M41.9339.30
LF-Diffs2 w/o DMXX✓5.85M44.4542.35
LF-Diffs2 w/ DMX✓✓7.48M44.7642.59
", + "bbox": [ + 506, + 193, + 890, + 281 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Ablation studies of Diffusion Prior and Joint Training.", + "bbox": [ + 506, + 292, + 883, + 306 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3e88852fff0d036ef21d9874ce29f2276d4fa615db71bab1299e89c90e0e315a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 319, + 888, + 441 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/315f3feef2171286c2baf8b46a40e8add4e2d02ad17b7af508a7bd0839780d88.jpg", + "table_caption": [ + "Figure 6. Qualitative results of our ablation study." + ], + "table_footnote": [], + "table_body": "
Sampling stepPSNR-μPSNR-LSSIM-μSSIM-LTime(s)
S=539.7641.330.98190.98780.049
S=1044.7642.590.99190.99060.106
S=2044.7342.540.99190.99060.208
S=5044.6642.450.99190.99060.479
S=10044.6542.440.99190.99060.971
", + "bbox": [ + 511, + 482, + 883, + 563 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablation studies of various settings on the sampling step during the reverse process. Time(s) only denotes the time expenditure of the DM for the corresponding setting.", + "bbox": [ + 498, + 573, + 890, + 616 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 633, + 619, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The potential of diffusion models in HDR deghosting has shown promising results, particularly in achieving visually perceptible outcomes. Different from image synthesis which generates each pixel from scratch, HDR imaging provides several LDR images as reference. Thus, it is inefficient to reconstruct the full HDR image starting from pure Gaussian noise. In this paper, we propose an efficient diffusion model called LF-Diff, consisting of LPENet, DHRNet, and a denoising network. Specifically, we apply the DM in a compact latent space to predict low-frequency priors of HDR images. These prior features provide explicit guidance for the image reconstruction process, thereby enhancing the details of the reconstructed HDR images. Compared to traditional DM-based methods, LF-Diff achieves accurate estimations and reduces artifacts in reconstructed images with much lower computational cost.", + "bbox": [ + 496, + 657, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "25739", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jie Chen, Zaifeng Yang, Tsz Nam Chan, Hui Li, Junhui Hou, and Lap-Pui Chau. Attention-guided progressive neural texture fusion for high dynamic range image restoration. IEEE Transactions on Image Processing, 31:2661-2672, 2022. 2, 6, 7, 8", + "[2] Nanxin Chen, Yu Zhang, Heiga Zen, Ron J Weiss, Mohammad Norouzi, and William Chan. Wavegrad: Estimating gradients for waveform generation. In International Conference on Learning Representations, 2020. 2", + "[3] Zheng Chen, Yulun Zhang, Ding Liu, Jinjin Gu, Linghe Kong, Xin Yuan, et al. Hierarchical integration diffusion model for realistic image deblurring. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[4] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In Advances in Neural Information Processing Systems, pages 8780-8794. Curran Associates, Inc., 2021. 1", + "[5] Li Fang, Qian Wang, and Long Ye. Glgnet: light field angular superresolution with arbitrary interpolation rates. Visual Intelligence, 2(1):6, 2024. 2", + "[6] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 2", + "[7] YongSeok Heo, KyoungMu Lee, SangUk Lee, Youngsu Moon, and Joonhyuk Cha. Ghost-free high dynamic range imaging. In IEEE Conference on Asian Conference on Computer Vision (ACCV), pages 486-500, 2011. 2", + "[8] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, pages 6840-6851. Curran Associates, Inc., 2020. 1, 2", + "[9] Jun Hu, O. Gallo, K. Pulli, and Xiaobai Sun. HDR deghosting: How to deal with saturation? In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1163-1170, 2013. 2, 6, 7, 8", + "[10] Jinhan Hu, Gyeongmin Choe, Zeeshan Nadir, Osama Nabil, Seok-Jun Lee, Hamid Sheikh, Youngjun Yoo, and Michael Polley. Sensor-realistic synthetic data engine for multiframe high dynamic range photography. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 516-517, 2020. 5, 6, 7", + "[11] Nima Khademi Kalantari and Ravi Ramamoorthi. Deep high dynamic range imaging of dynamic scenes. ACM Transactions on Graphics, 36(4):1-12, 2017. 1, 2, 4, 5, 6, 7, 8", + "[12] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. Advances in Neural Information Processing Systems, 35:23593-23606, 2022. 2", + "[13] Haoying Li, Yifan Yang, Meng Chang, Shiqi Chen, Huajun Feng, Zhihai Xu, Qi Li, and Yueting Chen. Srdiff: Single image super-resolution with diffusion probabilistic models. Neurocomputing, 479:47-59, 2022. 2", + "[14] Ru Li, Chuan Wang, Jue Wang, Guanghui Liu, Heng-Yu Zhang, Bing Zeng, and Shuaicheng Liu. Uphdr-gan: Generative adversarial network for high dynamic range imaging" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "with unpaired data. IEEE Transactions on Circuits and Systems for Video Technology, 32(11):7532-7546, 2022. 1, 2", + "[15] Zhen Liu, Wenjie Lin, Xinpeng Li, Qing Rao, Ting Jiang, Mingyan Han, Haoqiang Fan, Jian Sun, and Shuaicheng Liu. Adnet: Attention-guided deformable convolutional network for high dynamic range imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 463-470, 2021. 2, 6, 7", + "[16] Zhen Liu, Yinglong Wang, Bing Zeng, and Shuaicheng Liu. Ghost-free high dynamic range imaging with context-aware transformer. pages 344-360, 2022. 1, 2, 4, 6, 7, 8", + "[17] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 2", + "[18] Zhisheng Lu, Juncheng Li, Hong Liu, Chaoyan Huang, Linlin Zhang, and Tieyong Zeng. Transformer for single image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 457-466, 2022. 4", + "[19] Ziwei Luo, Fredrik K Gustafsson, Zheng Zhao, Jens Sjolund, and Thomas B Schön. Refusion: Enabling large-size realistic image restoration with latent-space diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1680-1691, 2023. 6", + "[20] Rafat Mantiuk, Kil Joong Kim, Allan G. Rempel, and Wolfgang Heidrich. HDR-VDP-2:a calibrated visual metric for visibility and quality predictions in all luminance conditions. In ACM Siggraph, pages 1–14, 2011. 5", + "[21] Yuzhen Niu, Jianbin Wu, Wenxi Liu, Wenzhong Guo, and Rynson WH Lau. Hdr-gan: Hdr image reconstruction from multi-exposed ldr images with large motions. IEEE Transactions on Image Processing, 30:3885-3896, 2021. 1, 2, 6, 7, 8", + "[22] Feiyue Peng, Maojun Zhang, Shiming Lai, Hanlin Tan, and Shen Yan. Deep hdr reconstruction of dynamic scenes. In 2018 IEEE 3rd International Conference on Image, Vision and Computing (ICIVC), pages 347-351. IEEE, 2018. 2", + "[23] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2", + "[24] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1, 2, 5, 8", + "[25] Pradeep Sen, Nima Khademi Kalantari, Maziar Yaesoubi, Soheil Darabi, Dan B Goldman, and Eli Shechtman. Robust patch-based hdr reconstruction of dynamic scenes. ACM Trans. Graph., 31(6):203-1, 2012. 2, 5, 6, 7, 8", + "[26] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In Proceedings of the 32nd International Conference on Machine Learning, pages 2256-2265, 2015. 1, 2" + ], + "bbox": [ + 503, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "25740", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 2, 5, 8", + "[28] Jou Won Song, Ye-In Park, Kyeongbo Kong, Jaeho Kwak, and Suk-Ju Kang. Selective transhdr: Transformer-based selective hdr imaging using ghost region mask. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVII, pages 288-304. Springer, 2022. 1, 2", + "[29] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 1", + "[30] Steven Tel, Zongwei Wu, Yulun Zhang, Barthélémy Heyrman, Cédric Demonceaux, Radu Timofte, and Dominique Gingac. Alignment-free hdr deghosting with semantics consistent transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12836-12845, 2023. 2", + "[31] Anna Tomaszewska and Radoslaw Mantiuk. Image registration for multi-exposure high dynamic range image acquisition. In International Conference in Central Europe on Computer Graphics and Visualization, WSCG'07, 2007. 2", + "[32] Okan Tarhan Tursun, Ahmet Oğuz Akyüz, Aykut Erdem, and Erkut Erdem. An objective dehosting quality metric for HDR images. Comput. Graph. Forum, 35(2):139-152, 2016. 5, 7", + "[33] Okan Tarhan Tursun, Ahmet Oğuz Akyüz, Aykut Erdem, and Erkut Erdem. An objective deghosting quality metric for hdr images. In Computer Graphics Forum, pages 139-152. Wiley Online Library, 2016. 7", + "[34] Haohan Wang, Xindi Wu, Zeyi Huang, and Eric P Xing. High-frequency component helps explain the generalization of convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8684-8694, 2020. 2", + "[35] Yinhuai Wang, Jiwen Yu, and Jian Zhang. Zero-shot image restoration using denoising diffusion null-space model. In The Eleventh International Conference on Learning Representations, 2022. 2", + "[36] Greg Ward. Fast, robust image registration for compositing high dynamic range photographs from hand-held exposures. Journal of Graphics Tools, 8, 2012. 2", + "[37] Jay Whang, Maurizio Delbracio, Hossein Talebi, Chitwan Sahara, Alexandros G Dimakis, and Peyman Milanfar. Deblurring via stochastic refinement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16293-16303, 2022. 2", + "[38] Shangzhe Wu, Jiarui Xu, Yu-Wing Tai, and Chi-Keung Tang. Deep high dynamic range imaging with large foreground motions. In European Conference on Computer Vision (ECCV), 2018. 6, 7", + "[39] Bin Xia, Yulun Zhang, Shiyin Wang, Yitong Wang, Xinglong Wu, Yapeng Tian, Wenming Yang, and Luc Van Gool. Diffir: Efficient diffusion model for image restoration. In" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13095-13105, 2023. 2", + "[40] Qingsen Yan, Jinqiu Sun, Haisen Li, Yu Zhu, and Yanning Zhang. High dynamic range imaging by sparse representation. Neurocomputing, 269:160-169, 2017. 2", + "[41] Qingsen Yan, Dong Gong, Qinfeng Shi, Anton van den Hengel, Chunhua Shen, Ian Reid, and Yanning Zhang. Attention-guided network for ghost-free high dynamic range imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1751-1760, 2019. 1, 2, 4, 5, 6, 7, 8", + "[42] Qingsen Yan, Lei Zhang, Yu Liu, Yu Zhu, Jinqiu Sun, Qinfeng Shi, and Yanning Zhang. Deep hdr imaging via a non-local network. IEEE Transactions on Image Processing, 29: 4308-4322, 2020. 2, 7, 8", + "[43] Qingsen Yan, Weiye Chen, Song Zhang, Yu Zhu, Jinqiu Sun, and Yanning Zhang. A unified hdr imaging method with pixel and patch level. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22211-22220, 2023. 2, 6, 7", + "[44] Qingsen Yan, Tao Hu, Yuan Sun, Hao Tang, Yu Zhu, Wei Dong, Luc Van Gool, and Yanning Zhang. Towards high-quality hdr deghosting with conditional diffusion models. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2023. 1, 2, 4, 5, 6, 7, 8", + "[45] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728-5739, 2022. 4", + "[46] Xiang Zhang, Tao Hu, Jiashuang He, and Qingsen Yan. Efficient content reconstruction for high dynamic range imaging. In ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7660-7664, 2024. 1, 2", + "[47] Xiang Zhang, Qiang Zhu, Tao Hu, and Qingsen Yan. Eiffhdr: An efficient network for multi-exposure high dynamic range imaging. In ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6560-6564, 2024. 1, 2" + ], + "bbox": [ + 501, + 92, + 890, + 670 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "25741", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Generating Content for HDR Deghosting from Frequency View/2b6f8409-576d-47cf-ab70-c1b4b9e76809_model.json b/2024/Generating Content for HDR Deghosting from Frequency View/2b6f8409-576d-47cf-ab70-c1b4b9e76809_model.json new file mode 100644 index 0000000000000000000000000000000000000000..89913e4524ae7131fe7fdec39cdab37f9c7d9789 --- /dev/null +++ b/2024/Generating Content for HDR Deghosting from Frequency View/2b6f8409-576d-47cf-ab70-c1b4b9e76809_model.json @@ -0,0 +1,3135 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.166, + 0.131, + 0.804, + 0.152 + ], + "angle": 0, + "content": "Generating Content for HDR Deghosting from Frequency View" + }, + { + "type": "text", + "bbox": [ + 0.238, + 0.18, + 0.73, + 0.199 + ], + "angle": 0, + "content": "Tao Hu\\(^{1\\dagger}\\) Qingsen Yan\\(^{1\\dagger}\\)* Yuankai Qi\\(^{2}\\) Yanning Zhang\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.236, + 0.199, + 0.734, + 0.218 + ], + "angle": 0, + "content": "\\(^{1}\\)Northwestern Polytechnical University \\(^{2}\\)Macquarie University" + }, + { + "type": "text", + "bbox": [ + 0.148, + 0.219, + 0.815, + 0.234 + ], + "angle": 0, + "content": "hutaoxauat@gmail.com, \\(\\mathbb{1}\\{\\) qingsenyan,ynzhang \\(\\} \\text{=} \\mathrm{nwpu.edu.cn},\\) yuankai.qi@mq.edu.au" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.284 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.473, + 0.692 + ], + "angle": 0, + "content": "Recovering ghost-free High Dynamic Range (HDR) images from multiple Low Dynamic Range (LDR) images becomes challenging when the LDR images exhibit saturation and significant motion. Recent Diffusion Models (DMs) have been introduced in HDR imaging field, demonstrating promising performance, particularly in achieving visually perceptible results compared to previous DNN-based methods. However, DMs require extensive iterations with large models to estimate entire images, resulting in inefficiency that hinders their practical application. To address this challenge, we propose the Low-Frequency aware Diffusion (LF-Diff) model for ghost-free HDR imaging. The key idea of LF-Diff is implementing the DMs in a highly compacted latent space and integrating it into a regression-based model to enhance the details of reconstructed images. Specifically, as low-frequency information is closely related to human visual perception we propose to utilize DMs to create compact low-frequency priors for the reconstruction process. In addition, to take full advantage of the above low-frequency priors, the Dynamic HDR Reconstruction Network (DHRNet) is carried out in a regression-based manner to obtain final HDR images. Extensive experiments conducted on synthetic and real-world benchmark datasets demonstrate that our LF-Diff performs favorably against several state-of-the-art methods and is \\(10\\times\\) faster than previous DM-based methods." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.724, + 0.208, + 0.739 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.841 + ], + "angle": 0, + "content": "Multiple exposure High Dynamic Range (HDR) imaging aims to restore missing details from exposure-varied Low Dynamic Range (LDR) images. However, in dynamic scenes, it often causes ghosting artifacts due to object or camera movement, limiting practical applications. Researchers are actively exploring ghost-free image recon" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.269, + 0.891, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.435, + 0.893, + 0.477 + ], + "angle": 0, + "content": "Figure 1. Average PSNR vs. inference time on the Kalantari's dataset [11]. Our method performs favorably and is \\(10\\times\\) faster than the previous diffusion-model-based method DiffHDR [44]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.492, + 0.892, + 0.522 + ], + "angle": 0, + "content": "struction methods to enable seamless capture of dynamic scenes with high dynamic ranges." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.893, + 0.903 + ], + "angle": 0, + "content": "In recent years, the rise of Deep Neural Networks (DNNs) has brought significant advancements to HDR imaging field. Numerous DNN-based methods have emerged, employing Convolutional Neural Networks (CNNs) [11, 41, 47] or Vision Transformers (ViTs) [16, 28] for HDR image reconstruction. Despite their advancements, DNN-based approaches face challenges when essential information (e.g., content, details) for overexposed areas is missing due to object or camera movement. More recently, Diffusion Models (DMs) [8, 26] have exhibited impressive performance in image synthesis [4, 29] and image reconstruction tasks (including HDR imaging) [24, 44, 46]. DMs achieve this by iteratively denoising Gaussian noise to generate high-fidelity images. Yan et al. [44] reconstruct complete HDR images from pure Gaussian noise, yielding impressive results. Note that, compared to previous Deep generative models [14, 21], DMs generate a more accurate target distribution without facing issues like optimization instability or mode collapse, making them a promising choice for enhancing HDR image quality. Nonetheless, DMs demand a substantial number of iteration steps on denoising models to capture intricate data details, which is time-consuming even with a high-end GPU card. For example, as shown in Fig. 1, DiffHDR [44] takes about 7.5 seconds to generate an HDR image on a single A100 GPU." + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.852, + 0.47, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author. † The first two authors contributed equally to this work. This work was partially supported by NSFC (62301432,62306240), NSBRPS (2023-JC-QN-0685, QCYRCXM-2023-057). Y. Qi is not supported by the above mentioned fundings." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25732" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.272 + ], + "angle": 0, + "content": "In this paper, our goal is to devise a DM-based method that effectively harnesses the robust distribution mapping capabilities of DMs for HDR image reconstruction. We note that it is unnecessary to reconstruct a complete HDR image from pure noise, because the LDR reference images already provide most of the necessary content information for HDR images. As low-frequency information is closely related to human visual perception [34], we propose to utilize DMs to create compact low-frequency priors in the latent space. These priors are then integrated into the regression-based model to predict the low-frequency content of the reconstructed HDR images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.273, + 0.473, + 0.56 + ], + "angle": 0, + "content": "To achieve our goal, we propose the Low-Frequency Aware Diffusion Model (LF-Diff), comprising a Low-Frequency Prior Extraction Network (LPENet), a denoising network, and a Dynamic HDR Reconstruction Network (DHRNet). Following established methods [3, 23, 39], we employ a two-stage training strategy for pre-train LF-Diff and DM training. In the first stage, LPENet learns to extract a compact Low-frequency Prior Representation (LPR) from ground-truth images, guiding the DHRNet. DHRNet consists of two modules: a Prior Integration Module (PIM) that fuses the LPR with intermediate features of DHRNet, and a Feature Refinement Module (FRM) that further processes the fused features to HDR image. Notably, LPENet and DHRNet are optimized together in this phase. In the second stage, we train DM to learn the compact LPR directly from LDR images. Since the LPR is lightweight and only used to low-frequency content for HDR imaging, our DM can estimate the LPR with extremely low computational cost, ensuring stable visual results after multiple iterations." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.56, + 0.458, + 0.574 + ], + "angle": 0, + "content": "The main contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.575, + 0.47, + 0.65 + ], + "angle": 0, + "content": "- We introduce LF-Diff, a straightforward and efficient DM-based method for HDR imaging. LF-Diff leverages the capabilities of diffusion models to generate informative low-frequency priors, which are then integrated into the imaging process to enhance results." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.651, + 0.47, + 0.711 + ], + "angle": 0, + "content": "- We propose PIM and FRM in the DHRNet to fully exploit the LPR. While PIM efficiently fuses the LPR with intermediate features of DHRNet, FRM further processes the fused features to reconstruct high-quality HDR images." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.712, + 0.47, + 0.801 + ], + "angle": 0, + "content": "- Extensive experiments demonstrate that the proposed LF-Diff method achieves SOTA performance in HDR imaging tasks, and produces visually appealing results that align with human visual perception. But consuming significantly fewer computational resources compared to other DM-based methods." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.575, + 0.47, + 0.801 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.815, + 0.22, + 0.831 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Ghost-free HDR Reconstruction. Traditional methods often employ motion rejection [7, 40], motion registration [31, 36], and patch matching [9, 25] approaches to align LDR images and reconstruct high-quality results. However," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.302 + ], + "angle": 0, + "content": "these methods heavily rely on the performance of preprocessing techniques and often face challenges when dealing with motions across large spatial extents. Subsequently, DNN-based methods have emerged as the mainstream approach, thanks to their superior nonlinear expression capability. Researchers have explored various applications of DNNs in HDR imaging, devising sophisticated network architectures and model optimization schemes, such as attention [1, 41], transformer [16, 28, 30], optical flow [11, 22], GAN [6, 14, 21], and others [5, 15, 42, 43, 47]. Nevertheless, when LDR images lack sufficient information due to motion or saturation, they often manifest artifacts not aligned with human visual perception, commonly referred to as ghosting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.303, + 0.893, + 0.605 + ], + "angle": 0, + "content": "Diffusion Models. Diffusion models have recently demonstrated promising results in various low-level vision tasks, including super-resolution [13, 24], HDR reconstruction [44, 46], colorization [12, 35], and deblurring [37]. DiffHDR [44] applied the conventional diffusion model paradigm to reconstruct HDR images, yielding remarkable outcomes. However, due to the high computational cost of diffusion models along with the requirement for numerous iterative steps, inference speed is limited. Despite efforts to mitigate this constraint [2, 17, 27], the overall complexity remains high, especially for high-resolution images commonly encountered in HDR scenarios. Recently, several methods [3, 23, 39] have addressed this issue by conducting diffusion modeling in the latent space. Rombach et al. [23] employed an autoencoder model to compress images into features equivalent to the image space. HI-Diff [3] leverages diffusion-based vector features for assisting image deblurring. However, given the unique characteristics of HDR imaging, further exploration is needed to select an appropriate latent space for HDR reconstruction tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.618, + 0.64, + 0.634 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.892, + 0.718 + ], + "angle": 0, + "content": "The diffusion model introduced by Sohl-Dickstein et al. [26] is inspired by nonequilibrium thermodynamics. Here, we provide a brief overview of the \"variance-preserving\" diffusion model from [8], which encompasses diffusion processes and reverse processes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.719, + 0.892, + 0.763 + ], + "angle": 0, + "content": "Diffusion Process. Given a clean image distribution, the diffusion process gradually injects isotropic Gaussian noise to generate \\( x_{t} \\) according to a variance schedule" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.764, + 0.859, + 0.781 + ], + "angle": 0, + "content": "\\[\n\\beta_ {1}, \\dots , \\beta_ {T} \\in (0, 1). \\text {L e t} \\alpha_ {t} = 1 - \\beta_ {t}, \\bar {\\alpha} _ {t} = \\prod_ {i = 1} ^ {T} \\alpha_ {i}:\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.79, + 0.892, + 0.807 + ], + "angle": 0, + "content": "\\[\nq \\left(x _ {t} \\mid x _ {0}\\right) = \\mathcal {N} \\left(x _ {t}; \\sqrt {\\bar {\\alpha} _ {t}} x _ {0}, (1 - \\bar {\\alpha} _ {t}) \\mathbf {I}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.818, + 0.866, + 0.833 + ], + "angle": 0, + "content": "The closed-form for this process can be represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.545, + 0.843, + 0.892, + 0.861 + ], + "angle": 0, + "content": "\\[\nx _ {t} \\left(x _ {0}, \\epsilon_ {t}\\right) = \\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t}, \\epsilon \\in (0, \\mathbf {I}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\( x_{t} \\) represents the output at time \\( t \\), \\( x_0 \\) is the initial clean image, and \\( \\epsilon_t \\) denotes isotropic Gaussian noise." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25733" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.522, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.091, + 0.889, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.328, + 0.895, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.364, + 0.893, + 0.407 + ], + "angle": 0, + "content": "Figure 2. The proposed LF-Diff comprises DHRNet, LPRNet, and a denoising network. LF-Diff undergoes two training stages: Pretrain LF-Diff (Sec. 4.1) and DM training (Sec. 4.2). Notably, during the inference stage, we do not input the ground-truth image into \\(\\mathrm{LPONet}_{DM}\\) and the denoising networks. Instead, we solely utilize the reverse process of DMs." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.434, + 0.47, + 0.48 + ], + "angle": 0, + "content": "Reverse Process. The reverse process attempts to approximate the data distribution \\( q(x_0) \\) by a Markovian process, starting with random Gaussian noise \\( x_{T} = \\mathcal{N}(x_{T};\\mathbf{0},\\mathbf{I}) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.127, + 0.491, + 0.47, + 0.51 + ], + "angle": 0, + "content": "\\[\nq \\left(x _ {t - 1} \\mid x _ {t}, x _ {0}\\right) = \\mathcal {N} \\left(x _ {t - 1}; \\tilde {\\mu} _ {t} \\left(x _ {t}, x _ {0}\\right), \\tilde {\\beta} _ {t} \\mathbf {I}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.521, + 0.469, + 0.584 + ], + "angle": 0, + "content": "where \\(\\tilde{\\mu}_t\\) and \\(\\tilde{\\beta}_t\\) are distribution parameters. Since the true reverse process Eq.(3) relies on \\(q(x_0)\\) and is intractable, the neural network \\(f_{\\theta}\\) serves as the denoiser to estimate \\(p_{\\theta}(x_{t - 1}|x_t,t)\\) instead of \\(q(x_{t - 1}|x_t,x_0)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.593, + 0.47, + 0.629 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} p _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}, t\\right) = \\mathcal {N} \\left(x _ {t - 1}; \\mu_ {\\theta} \\left(x _ {t}, t\\right), \\Sigma_ {\\theta} \\left(x _ {t}, t\\right)\\right) \\tag {4} \\\\ = q \\left(x _ {t - 1} \\mid f _ {\\theta} (x _ {t}, t)\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.64, + 0.47, + 0.685 + ], + "angle": 0, + "content": "where \\(\\mu_{\\theta}\\) and \\(\\Sigma_{\\theta}\\) are the parameters to be estimated in the reverse process. Let fixed variances \\((\\Sigma_{\\theta}(x_t,t) = \\tilde{\\beta}_t\\mathbf{I})\\) , the optimization objective can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.122, + 0.695, + 0.47, + 0.713 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {t, x _ {0}, \\epsilon_ {t}} [ \\| \\epsilon_ {t} - f _ {\\theta} (\\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t}, t) \\| ^ {2} ]. \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.725, + 0.212, + 0.743 + ], + "angle": 0, + "content": "4. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.903 + ], + "angle": 0, + "content": "As depicted in Fig. 2, we introduce the proposed LF-Diff framework, which consists of two training stages. During the first stage, we pre-train LF-Diff to extract accurate low-frequency prior representations from ground truths using LPenet, while DHRNet learns how to utilize the above priors to reconstruct HDR images. During the second stage, we train the DM to accurately predict the low-frequency prior representations directly from LDR images, and jointly optimize it with DHRNet to generate final high-quality HDR images." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.433, + 0.757, + 0.449 + ], + "angle": 0, + "content": "4.1. Stage One: Pretrain LF-Diff" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.456, + 0.892, + 0.637 + ], + "angle": 0, + "content": "In the first stage, our objective is to train the LPERNet how to extract accurate low-frequency prior features from ground truth and embed this representation into the DHRNet to guide the reconstruction process. As illustrated in Fig. 2 (a), we utilize the Prior Integration Module (PIM) to incorporate these prior features into the DHRNet via the cross-attention mechanism. Within DHRNet, multiple Feature Refinement Modules (FRM) are stacked to directly learn the mapping relationship from input features to the target image distribution, while PIM leverages prior features to assist in recovering the low-frequency content of the reconstructed images. Next, we elaborate on above descriptions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.638, + 0.893, + 0.803 + ], + "angle": 0, + "content": "Low-frequency Prior Extraction Network. The LPENet has a straightforward structure, as shown in Fig. 2 (a), comprising several residual blocks designed to extract the Low-frequency Prior Representation (LPR). Starting with the ground truth image \\( I_{gt} \\in \\mathbb{R}^{H \\times W \\times 3} \\), we initially perform a tonemapping operation \\( \\mathcal{T}(\\cdot) \\) to obtain the LDR domain ground truth image \\( \\mathcal{T}(I_{gt}) \\). We concatenate these images along the channel dimension and employ the PixelUnshuffle operation to downsample, generating the input for LPENet. Subsequently, LPENet extracts the LPR \\( z \\in \\mathbb{R}^{\\frac{H}{k} \\times \\frac{W}{k} \\times 3} \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.813, + 0.892, + 0.829 + ], + "angle": 0, + "content": "\\[\nz = \\operatorname {L P E N e t} (\\operatorname {P i x i l e U n s h u r f l e} (\\operatorname {C o n c a t} (I _ {g t}, \\mathcal {T} (I _ {g t})))) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.837, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathcal{T}(x) = \\frac{\\log(1 + \\mu x)}{\\log(1 + \\mu)}\\) denotes the tonemapping operator, \\(\\mu = 5000\\) is a predefined constant parameter that controls the degree of compression applied to the input signal, \\(k\\) denotes the sampling multiple of PixelUnshuffle." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "25734" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.242 + ], + "angle": 0, + "content": "Dynamic HDR Reconstruction Network. As illustrated in Fig. 2 (a), DHRNet consists of multiple stacked Reconstruction Blocks, with each block comprising one Prior Integration Module (PIM) and \\(N\\) Feature Refinement Module (FRM). PIM fuses the LPR with intermediate features of DHRNet, we will give a detailed description in the following part. FRM is designed based on the understanding that natural images encompass diverse frequencies, since high and low frequencies serve distinct roles in image encoding." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.244, + 0.473, + 0.351 + ], + "angle": 0, + "content": "In Fig. 3 (a), the structure of FRM involves several blocks. Specifically, we begin by processing the features through a residual block to obtain \\( F_{n-1} \\). Subsequently, we decouple these features into two components: high-frequency features \\( F_{n-1}^{\\mathrm{high}} \\), which remain resolution-invariant, and low-frequency features \\( F_{n-1}^{\\mathrm{low}} \\) with reduced resolution, following the approach described in [18]:" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.363, + 0.345, + 0.38 + ], + "angle": 0, + "content": "\\[\nF _ {n - 1} ^ {\\text {l o w}} = \\operatorname {A v g p o o l} \\left(F _ {n - 1}, k\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.378, + 0.469, + 0.404 + ], + "angle": 0, + "content": "\\[\nF _ {n - 1} ^ {\\text {h i g h}} = F _ {n - 1} - \\text {U p s a m p l e} \\left(F _ {n - 1} ^ {\\text {l o w}}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.415, + 0.473, + 0.613 + ], + "angle": 0, + "content": "where \\( k \\) denotes the kernel size of the avgpooling layer, Upsample denotes the upsampling operator. In essence, low-frequency features capture global dependencies within the input (image/features) and do not necessitate a high-resolution feature map, but require global attention. On the other hand, high-frequency features capture detailed local information, requiring a high-resolution feature map, and only can be achieved through local operators. Therefore, we employ the residual block (RB) to process \\( F_{n - 1}^{\\mathrm{high}} \\) and utilize the self-attention (SA) from [45] to handle \\( F_{n - 1}^{\\mathrm{low}} \\). Subsequently, we fuse the low- and high-frequency features to preserve the initial details, resulting in the feature \\( F_{n - 1}^{\\prime} \\). This operation can be represented as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.084, + 0.623, + 0.47, + 0.657 + ], + "angle": 0, + "content": "\\[\nF _ {n - 1} ^ {\\prime} = \\operatorname {C o n c a t} \\left[ R B \\left(F _ {n - 1} ^ {\\text {h i g h}}\\right), \\operatorname {U p s a m p l e} \\left(S A \\left(F _ {n - 1} ^ {\\text {l o w}}\\right)\\right) \\right]. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.47, + 0.735 + ], + "angle": 0, + "content": "Considering that \\( F_{n-1}^{\\prime} \\) is a concatenation of two features, a \\( 1 \\times 1 \\) Conv is employed to effectuate a reduction in the number of channels. This downsizing step is followed by a channel attention module, designed to accentuate those channels exhibiting heightened activation values." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In order to effectively merge the prior feature derived from the LPERNet with the intermediate feature produced by DHRNet, we replaced the self-attention mechanism in the low-frequency branch of FRM with a carefully designed cross-attention mechanism, resulting in the Prior Integration Module (PIM). As depicted in Fig. 2 (a), a PIM is positioned before each reconstruction block. Within each PIM, cross-attention computations are performed between the prior and intermediate features to facilitate feature fusion. This module enables the aggregation of information from the prior feature into the features of DHRNet." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.09, + 0.891, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.594, + 0.208, + 0.793, + 0.22 + ], + "angle": 0, + "content": "(a) Feature Refinement Module" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.221, + 0.875, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.572, + 0.35, + 0.83, + 0.362 + ], + "angle": 0, + "content": "(b)Cross Self-attention module from PIM" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.376, + 0.893, + 0.404 + ], + "angle": 0, + "content": "Figure 3. The architecture of (a) Feature Refinement Module (FRM) and (b) Cross-attention mechanism." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.414, + 0.895, + 0.62 + ], + "angle": 0, + "content": "Specifically, as illustrated in Fig. 3 (b), our cross-attention module receives two types of input: the intermediate feature \\(\\mathbf{F} \\in \\mathbb{R}^{\\hat{H} \\times \\hat{W} \\times \\hat{C}}\\) and the prior feature \\(z \\in \\mathbb{R}^{\\hat{H} \\times \\hat{W} \\times \\hat{N}}\\). We project \\(\\mathbf{F}\\) into a query vector \\(\\mathbf{Q} = W_d^Q W_c^Q\\mathbf{F}\\) using point-wise \\(1 \\times 1\\) convolution and depth-wise \\(3 \\times 3\\) convolution with weights \\(W_d^Q\\) and \\(W_c^Q\\). Similarly, \\(z\\) is transformed into key \\(\\mathbf{K}\\) and value \\(\\mathbf{V}\\) through analogous operations. Next, we reshape these projections into matrices suitable for attention computation: \\(\\hat{\\mathbf{Q}} \\in \\mathbb{R}^{\\hat{H}\\hat{W} \\times \\hat{C}}\\), \\(\\hat{\\mathbf{K}} \\in \\mathbb{R}^{\\hat{N} \\times \\hat{H}\\hat{W}}\\), and \\(\\hat{\\mathbf{V}} \\in \\mathbb{R}^{\\hat{H}\\hat{W} \\times \\hat{N}}\\). We then calculate a more computationally efficient attention map \\(A \\in \\mathbb{R}^{\\hat{N} \\times \\hat{C}}\\) by performing dot product between \\(\\hat{\\mathbf{Q}}\\) and \\(\\hat{\\mathbf{K}}\\). The process can be described as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.627, + 0.892, + 0.647 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {F}} = W _ {c} \\hat {\\mathbf {V}} \\cdot \\operatorname {S o f t m a x} (\\hat {\\mathbf {K}} \\cdot \\hat {\\mathbf {Q}} / \\gamma) + \\mathbf {F}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.655, + 0.892, + 0.684 + ], + "angle": 0, + "content": "where \\(\\gamma\\) is a trainable parameter. Notably, we do not implement multi-head attention in PIM." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.685, + 0.892, + 0.852 + ], + "angle": 0, + "content": "Training Strategy. As proposed in [11], we initially transform the provided set of input LDR images \\(\\{L_1, L_2, \\ldots, L_N\\}\\) into their corresponding HDR versions \\(H_i\\) using gamma correction. Subsequently, we concatenate each \\(L_i\\) with its corresponding \\(H_i\\) along the channel axis to generate six-channel input tensors \\(X_i = [L_i, H_i]\\). Similar to [16, 41, 44], we employ an alignment module [41] to process the input LDR images, align the features implicitly, and feed them into DHRNet. Then, the LPR feature \\(z\\) extracted from LPENet is injected into DHRNet via PIM. The reconstructed HDR image \\(\\hat{H}\\) is generated by:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.859, + 0.892, + 0.878 + ], + "angle": 0, + "content": "\\[\n\\hat {H} = \\operatorname {C o n v} _ {3 \\times 3} \\left(\\mathrm {D H R N e t} \\left(X _ {i}, z\\right)\\right). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Following previous approaches [16, 41], we utilize both" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "25735" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "tonemapped per-pixel loss and perceptual loss as the image reconstruction loss function \\(\\mathcal{L}_r\\). This dual-loss strategy optimizes both pixel-level accuracy and high-level feature representations in the generated HDR results:" + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.163, + 0.469, + 0.178 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r} = \\left\\| \\mathcal {T} (H) - \\mathcal {T} (\\hat {H}) \\right\\| _ {1} + \\lambda \\| \\phi_ {i, j} (\\mathcal {T} (H)) - \\phi_ {i, j} (\\mathcal {T} (\\hat {H})) \\| _ {1}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.187, + 0.469, + 0.248 + ], + "angle": 0, + "content": "where \\(\\phi_{i,j}(\\cdot)\\) signifies the \\(j\\)-th convolutional feature extracted from the VGG19 network after the \\(i\\)-th maxpooling operation, and \\(\\lambda = 1e - 2\\) is a hyperparameter that balances the contribution of the each component." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.256, + 0.469, + 0.273 + ], + "angle": 0, + "content": "4.2. Stage Two: Diffusion Models for HDR Imaging" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.279, + 0.469, + 0.399 + ], + "angle": 0, + "content": "After the above-mentioned learning procedure, we now possess LDR images and their corresponding LPR \\( z \\). In the second stage (Fig. 2 (b, c)), our objective is to efficiently harness the powerful distribution estimation capability of the DMs. Specifically, we utilize the LPENet from stage one to capture the LPR as the denoising target of the DM. The DM will learn how to extract accurate LPR from the LDR inputs and perform joint optimization with DHRNet." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.4, + 0.469, + 0.46 + ], + "angle": 0, + "content": "Diffusion Model for LRP Learning. After capturing the LPR \\( z \\in \\mathbb{R}^{\\frac{H}{k} \\times \\frac{W}{k} \\times N} \\) from the pretrained LPENet, we transform the clean LPR feature \\( z \\) into a noisy version \\( z_T \\) using Eq.(2), which can be described as:" + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.469, + 0.469, + 0.487 + ], + "angle": 0, + "content": "\\[\nq \\left(z _ {T} \\mid z\\right) = \\mathcal {N} \\left(z _ {T}; \\sqrt {\\bar {\\alpha} _ {T}} z, (1 - \\bar {\\alpha} _ {T}) \\mathbf {I}\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.495, + 0.468, + 0.525 + ], + "angle": 0, + "content": "where \\(T\\) represents the total number of iterations, while \\(\\bar{\\alpha}_t\\) and \\(\\alpha\\) denote the pre-defined variance schedule." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.525, + 0.469, + 0.646 + ], + "angle": 0, + "content": "In the reverse process, we iteratively generate the LPR from a pure Gaussian distribution, starting from \\( z_{T} \\) and moving backward to \\( z_{0} \\), which utilizes the posterior distribution described in Eq. (3). Following previous work [24, 44], we utilize a neural network to estimate \\( p_{\\theta}(x_{t - 1}|x_t,t) \\) instead of \\( q(x_{t - 1}|x_t,x_0) \\) for each step. Specifically, we first use LPENetDM to obtain a conditional feature \\( D\\in \\mathbb{R}^{\\frac{H}{k}\\times \\frac{W}{k}\\times N} \\) from aligned LDR feature:" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.655, + 0.469, + 0.672 + ], + "angle": 0, + "content": "\\[\n\\mathbf {D} = \\operatorname {L P E N e t} _ {D M} (A M (\\text {P i x e l U n s h u f f l e} (X _ {i}))). \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.68, + 0.469, + 0.803 + ], + "angle": 0, + "content": "where \\(\\mathrm{LPENet}_{DM}\\) maintains a similar architecture to \\(\\mathrm{LPENet}\\) but with a modified input dimensionality for its first convolution layer, \\(AM(\\cdot)\\) denotes alignment module from [41]. The denoising neural network predicts the noise based on both \\(z_{t}\\) and the derived conditional feature \\(D\\), i.e., \\(\\epsilon_{\\theta}(Concat(z_{t},\\mathbf{D}),t)\\). Upon substituting this estimated noise term into the equation governing the reverse process Eq. (4), we arrive at the following sampling formula:" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.81, + 0.469, + 0.855 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {z} _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\boldsymbol {z} _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\boldsymbol {z} _ {t}, \\mathbf {D}, t\\right)\\right) + \\sqrt {1 - \\alpha_ {t}} \\boldsymbol {\\epsilon} _ {t}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "where \\(\\epsilon_t \\sim \\mathcal{N}(0, \\mathbf{I})\\). By iteratively sampling \\(z_t\\) using Eq. (14) \\(T\\) times, we progressively reconstruct the predicted LPR representation \\(z_0\\) (i.e. \\(\\hat{z}) \\in \\mathbb{R}^{\\frac{H}{k} \\times \\frac{W}{k} \\times N}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.32 + ], + "angle": 0, + "content": "Training Strategy. Traditional DMs solely learn the probability distribution by optimizing the weighted variational bound (Eq. (5)), resulting in a slight deviation between the predicted prior and the actual prior. Integrating the DM directly with DHRNet might lead to a misalignment issue, thereby hampering the overall image processing performance. To address this, we joint training the DM and DHRNet. During each training iteration, we first sample the noise sample \\( z_{t} \\) through the diffusion process (Eq. (2)). Given that our denoising neural network is lightweight, we use a reverse process for \\( S \\) iterations based on the DDIM strategy [27] to generate the predicted prior feature \\( \\hat{z} \\). This \\( \\hat{z} \\) guides DHRNet via PIM. The loss function in the second stage consists of the reconstruction loss \\( \\mathcal{L}_r \\) (Eq. (11)) and the diffusion loss \\( \\mathcal{L}_{diff} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.33, + 0.892, + 0.377 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {d i f f}} = E _ {z _ {0}, t, \\epsilon_ {t} \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I})} \\left[ \\| \\epsilon_ {t} - \\epsilon_ {\\theta} (z _ {t}, D, t) \\| ^ {2} \\right] \\tag {15} \\\\ + \\left\\| \\hat {z} - z \\right\\| _ {1}. \\\\ \\end{array}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.389, + 0.61, + 0.403 + ], + "angle": 0, + "content": "4.3. Inference" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.413, + 0.892, + 0.534 + ], + "angle": 0, + "content": "In the inference stage, \\(L P E N t_{D M}\\) extracts a conditional feature \\(\\mathbf{D}\\) from aligned LDR feature (Eq. (13)), and we randomly sample a Gaussian noise \\(\\hat{z}_T\\). Then, denoising neural network utilizes the \\(\\hat{z}_T\\) and \\(\\mathbf{D}\\) to estimate LPR after \\(S\\) iterations (Eq. (14)) based on DDIM [27]. After that, DHRNet exploits the LPR to restore HDR images as Eq. (10). More details about the DM training and the model inference can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.549, + 0.633, + 0.566 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.575, + 0.707, + 0.59 + ], + "angle": 0, + "content": "5.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.854 + ], + "angle": 0, + "content": "Datasets. All methods are trained using two publicly available datasets, employing identical training settings: Kalantari's dataset [11] and Hu's dataset [10]. Kalantari's dataset consists of 74 samples for training and 15 for testing, all captured under authentic environmental conditions. Each sample comprises three LDR images with exposure variations of \\(\\{-2,0, +2\\}\\) or \\(\\{-3,0, +3\\}\\). In contrast, Hu's dataset is a synthetic dataset designed to emulate sensor realism, generated through a game engine. This dataset contains images captured at three distinct exposure levels \\(\\{-2,0, +2\\}\\), with our primary focus on dynamic scene images within this collection. Following the settings outlined in [10], we allocate the initial 85 samples for training, reserving the remaining 15 for testing. Moreover, to further validate the model's generalizability, we incorporate Sen et al.'s dataset [25] and Tursun et al.'s dataset [32] exclusively for qualitative assessment, as these datasets lack ground truths." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Evaluation Metrics. To conduct a quantitative comparison, we utilize five objective metrics: PSNR- \\(\\mu\\) SSIM- \\(\\mu\\) PSNR-L,SSIM-L,and HDR-VDP-2 [20].Here, the sub" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "25736" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.156, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.158, + 0.091, + 0.383, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.221, + 0.208, + 0.314, + 0.217 + ], + "angle": 0, + "content": "Tonemapped Result" + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.091, + 0.434, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.436, + 0.092, + 0.484, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.409, + 0.209, + 0.471, + 0.217 + ], + "angle": 0, + "content": "LDR-Patches" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.218, + 0.132, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.218, + 0.184, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.185, + 0.218, + 0.234, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.218, + 0.285, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.286, + 0.218, + 0.335, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.218, + 0.385, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.386, + 0.218, + 0.435, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.436, + 0.218, + 0.485, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.258, + 0.132, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.258, + 0.183, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.184, + 0.258, + 0.234, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.258, + 0.284, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.285, + 0.258, + 0.334, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.258, + 0.385, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.386, + 0.258, + 0.435, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.436, + 0.258, + 0.485, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.146, + 0.304, + 0.397, + 0.319 + ], + "angle": 0, + "content": "(a) Example from Kalantari et al.'s dataset [11]" + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.091, + 0.563, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.565, + 0.091, + 0.638, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.091, + 0.789, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.791, + 0.091, + 0.84, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.841, + 0.092, + 0.889, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.209, + 0.542, + 0.217 + ], + "angle": 0, + "content": "LDRs" + }, + { + "type": "image_caption", + "bbox": [ + 0.627, + 0.208, + 0.719, + 0.217 + ], + "angle": 0, + "content": "Tonemapped Result" + }, + { + "type": "image", + "bbox": [ + 0.72, + 0.208, + 0.789, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.791, + 0.209, + 0.84, + 0.217 + ], + "angle": 0, + "content": "LDR-" + }, + { + "type": "image_caption", + "bbox": [ + 0.841, + 0.209, + 0.889, + 0.217 + ], + "angle": 0, + "content": "Patches" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.218, + 0.132, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.218, + 0.183, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.184, + 0.218, + 0.234, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.218, + 0.284, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.285, + 0.218, + 0.334, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.218, + 0.384, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.385, + 0.218, + 0.435, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.436, + 0.218, + 0.485, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.258, + 0.132, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.258, + 0.183, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.184, + 0.258, + 0.234, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.258, + 0.284, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.285, + 0.258, + 0.334, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.258, + 0.384, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.385, + 0.258, + 0.435, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.436, + 0.258, + 0.485, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.486, + 0.218, + 0.536, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.218, + 0.588, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.218, + 0.638, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.218, + 0.689, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.69, + 0.218, + 0.739, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.74, + 0.218, + 0.789, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.79, + 0.218, + 0.838, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.839, + 0.218, + 0.889, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.606, + 0.295, + 0.825, + 0.303 + ], + "angle": 0, + "content": "AHDR HDR-GAN CA-VIT HyHDR Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.83, + 0.295, + 0.873, + 0.303 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.329, + 0.892, + 0.357 + ], + "angle": 0, + "content": "Figure 4. Visual comparisons are conducted on testing data from various datasets, focusing on zoomed-in local areas of the HDR images estimated by our method and the compared techniques. Our model demonstrates the ability to generate HDR images of superior quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.384, + 0.468, + 0.413 + ], + "angle": 0, + "content": "scripts \\(\\mu\\) and L denote that these metrics are computed in the tonemapped and linear domains, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.414, + 0.469, + 0.716 + ], + "angle": 0, + "content": "Implementation Details. Our implementation is conducted using PyTorch, and each training stage converges after 300 epochs on four NVIDIA A100 GPUs. We employ the Adam optimizer with an initial learning rate of 1e-4, which decays by a factor of 0.1 after every 50 epochs. The training dataset is processed by cropping \\(128 \\times 128\\) patches with a stride of 64, and the batch size is set to 64. For LF-Diff, we adopt a variant of the commonly used U-Net architecture from [19] as the denoising network, with 3 levels of blocks \\(\\{2, 2, 2\\}\\). During training, the time step \\(T\\) is set to 200, and the implicit sampling step \\(S\\) is set to 10 for both training and inference phases to achieve efficient restoration. In DHRNet, the parameters \\(N_{i} \\in \\{L_{1}, L_{2}, L_{3}\\}\\) are set to \\(\\{3, 3, 3\\}\\), and the channel \\(C\\) is set to 60. The number of attention heads for the self-attention branch at the same level is set to \\(\\{6, 6, 6\\}\\), and the channel expansion factor in FFN is 2.66. The kernels for avgpool in PIM and FRM are set to 4 and 2, respectively. For LPENet, it comprises 4 residual blocks, with a Pixelunshuffle downsampling factor of 4 and the output channel of 3." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.727, + 0.469, + 0.742 + ], + "angle": 0, + "content": "5.2. Comparison with the State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.469, + 0.901 + ], + "angle": 0, + "content": "In this section, we evaluate the performance of the proposed LF-Diff method and present experimental results to validate its reconstruction performance compared to state-of-the-art techniques. Specifically, we evaluate LF-Diff against two patch-based methods [9, 25], a optical flow-based method [11], five CNN-based methods [1, 15, 21, 38, 41], two ViT-based methods [16, 43], along with a diffusion-based method [44]. It's worth noting that all deep learning methods employ the same training dataset and settings for consistency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.383, + 0.892, + 0.61 + ], + "angle": 0, + "content": "Datasets w/ Ground Truth. The quantitative outcomes for LF-Diff on two datasets are presented in Tab. 1. Our method is compared with various state-of-the-art approaches using the testing data from [11] and [10], which includes challenging samples characterized by saturated backgrounds and foreground motions. All quantitative results are averaged across testing images. Notably, LF-Diff exhibits a significant improvement over other leading methods, exceeding the performance of the DM-based method DiffHDR [44], by \\(0.65\\mathrm{dB}\\) and \\(0.86\\mathrm{dB}\\) in PSNR- \\(\\mu\\) and PSNR-L metrics, respectively, based on Kalantari's dataset [11]. Moreover, LF-Diff showcases superior performance compared to the runner-up method, HyHDR [43], with improvements of \\(0.28\\mathrm{dB}\\) and \\(0.19\\mathrm{dB}\\) in PSNR- \\(\\mu\\) and PSNR-L metrics, respectively, on Hu's dataset [10]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.901 + ], + "angle": 0, + "content": "As shown in Fig. 4 (a) and (b), the datasets present some challenging samples due to information loss in the LDR images. Most existing approaches struggle in these areas, producing ghosting artifacts due to large motions and occlusions. Kalantari's method [11] and DHDR [38] struggle with background motion due to error-prone alignments (e.g., optical flow), resulting in undesirable ghosting. While HDR-GAN [21] exhibits noticeable ghosting artifacts around the balcony area and introduces erroneous color information, AHDR [41] aligns LDR images using convolutional spatial attention. However, it unintentionally suppresses valuable contextual information and encounters difficulties with significant motions in over/underexposed scenes. CA-ViT [16], relying on patch-based sampling, produces noticeable blocky ghosting. In contrast, aided by the DM, both Diff-HDR [44] and our LF-Diff generate HDR images aligned with human perception. Notably, our approach outperforms Diff-HDR with faster inference speeds and a smaller computational overhead." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25737" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.09, + 0.159, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.16, + 0.091, + 0.385, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.386, + 0.091, + 0.432, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.433, + 0.091, + 0.483, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.091, + 0.561, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.562, + 0.092, + 0.787, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.788, + 0.092, + 0.863, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.864, + 0.092, + 0.889, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.211, + 0.131, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.083, + 0.293, + 0.128, + 0.301 + ], + "angle": 0, + "content": "Kalantari" + }, + { + "type": "image", + "bbox": [ + 0.132, + 0.211, + 0.182, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.132, + 0.293, + 0.173, + 0.301 + ], + "angle": 0, + "content": "DHDR" + }, + { + "type": "image", + "bbox": [ + 0.183, + 0.211, + 0.231, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.183, + 0.293, + 0.223, + 0.301 + ], + "angle": 0, + "content": "AHDR" + }, + { + "type": "image", + "bbox": [ + 0.232, + 0.211, + 0.284, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.232, + 0.293, + 0.284, + 0.301 + ], + "angle": 0, + "content": "HDR-GAN" + }, + { + "type": "image", + "bbox": [ + 0.285, + 0.211, + 0.333, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.285, + 0.293, + 0.333, + 0.301 + ], + "angle": 0, + "content": "APNET" + }, + { + "type": "image", + "bbox": [ + 0.334, + 0.211, + 0.385, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.334, + 0.293, + 0.379, + 0.301 + ], + "angle": 0, + "content": "CA-VIT" + }, + { + "type": "image", + "bbox": [ + 0.386, + 0.211, + 0.434, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.386, + 0.293, + 0.434, + 0.301 + ], + "angle": 0, + "content": "DiffHDR" + }, + { + "type": "image", + "bbox": [ + 0.435, + 0.211, + 0.484, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.451, + 0.293, + 0.475, + 0.301 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.485, + 0.211, + 0.534, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.487, + 0.293, + 0.533, + 0.301 + ], + "angle": 0, + "content": "Kalantari" + }, + { + "type": "image", + "bbox": [ + 0.535, + 0.211, + 0.585, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.544, + 0.293, + 0.578, + 0.301 + ], + "angle": 0, + "content": "DHDR" + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.211, + 0.636, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.596, + 0.293, + 0.69, + 0.301 + ], + "angle": 0, + "content": "AHDR HDR-GAN" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.211, + 0.686, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.64, + 0.293, + 0.69, + 0.301 + ], + "angle": 0, + "content": "HDR-GAN" + }, + { + "type": "image", + "bbox": [ + 0.69, + 0.211, + 0.738, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.717, + 0.293, + 0.733, + 0.301 + ], + "angle": 0, + "content": "NET " + }, + { + "type": "image", + "bbox": [ + 0.739, + 0.211, + 0.788, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.745, + 0.293, + 0.782, + 0.301 + ], + "angle": 0, + "content": "CA-VIT" + }, + { + "type": "image", + "bbox": [ + 0.789, + 0.211, + 0.838, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.797, + 0.293, + 0.838, + 0.301 + ], + "angle": 0, + "content": "DiffHDR" + }, + { + "type": "image", + "bbox": [ + 0.84, + 0.211, + 0.874, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.856, + 0.293, + 0.88, + 0.301 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.286, + 0.33, + 0.682, + 0.344 + ], + "angle": 0, + "content": "Figure 5. Visual comparisons on the datasets without ground truth." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.357, + 0.891, + 0.512 + ], + "angle": 0, + "content": "
DatasetsModelsSen[25]Hu[9]Kalantari[11]DHDR[38]AHDR[41]NHDRR[42]HDRGAN[21]APNT[1]CA-ViT[16]HyHDR[43]DiffH[44]Ours
KalantariPSNR-μ40.9532.1942.7441.6443.6242.4143.9243.9444.3244.6444.1144.76
PSNR-L38.3130.8441.2240.9141.0341.0841.5741.6142.1842.4741.7342.59
SSIM-μ0.98050.97160.98770.98690.99000.98870.99050.98980.99160.99150.99110.9919
SSIM-L0.97260.95060.98480.98580.98620.98610.98650.98790.98840.98940.98850.9906
HDR-VDP-255.7255.2560.5160.5062.3061.2165.4564.0566.0366.0565.5266.54
DatasetsModelsSen[25]Hu[9]Kalantari[11]DHDR[38]AHDR[41]NHDRR[42]HDRGAN[21]APNT[1]ADNet[15]CA-ViT[16]HyHDR[43]Ours
HuPSNR-μ31.4836.5641.6041.1345.7645.1545.8646.4146.7948.1048.4648.74
PSNR-L33.5836.9443.7641.2049.2248.7549.1447.9750.3851.1751.9152.10
SSIM-μ0.95310.98240.99140.98700.99560.99560.99450.99530.99080.99470.99590.9968
SSIM-L0.96340.98770.99380.99410.99800.99810.99890.99860.99870.99890.99910.9993
HDR-VDP-266.3967.5872.9470.8275.0474.8675.1973.0676.2177.1277.2477.35
" + }, + { + "type": "table_caption", + "bbox": [ + 0.145, + 0.522, + 0.822, + 0.537 + ], + "angle": 0, + "content": "Table 1. The evaluation results on Kalantari's [11] and Hu's [10] datasets. The best results are highlighted in Bold." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.56, + 0.47, + 0.726 + ], + "angle": 0, + "content": "Datasets w/o Ground Truth. To assess the generalization capability of our proposed HDR imaging method, we evaluated the performance of the model trained on Kalantari et al.'s Dataset [11] by testing it on datasets from Sen et al. [25] and Tursun et al. [33], which lack ground truths. In Fig. 5 (a), numerous current methods encounter difficulties in restoring recover the large saturated region and large motion. Conversely, in Fig. 5 (b), while complete elimination of ghosting remains elusive for any method, our approach notably enhances image sharpness and detail compared to preceding techniques." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Computational Budgets. We also conducted comparisons regarding model parameters and inference times with previous approaches. As illustrated in Table 2, patch match-based methods [9, 25] exhibit significantly longer inference times due to their CPU-based computation. Kalantari [11] requires considerable time for initial optical flow alignment. NHDRRNet [42], employing a U-shaped network architecture, demonstrates shorter inference times but substantially higher parameter counts compared to other methods. CAViT [16] has a large number of standard transformer blocks leading to high computational cost despite fewer parame" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.56, + 0.892, + 0.636 + ], + "angle": 0, + "content": "ters. DiffHDR [44] has high inference time and parameter count due to reconstructing HDR from pure noise. In comparison, our method effectively leverages the powerful distribution estimation capability of DM with orders of magnitude fewer parameters and computations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.646, + 0.654, + 0.663 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.671, + 0.891, + 0.716 + ], + "angle": 0, + "content": "In this section, we investigate the impact of various designs within our proposed method. All experiments are conducted using the Kalantari's dataset [11]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Effects of Diffusion Prior. As shown in Tab. 3, we establish a regression-based baseline model without LPR generation. In this configuration, the PIM in DHRNet is replaced with FRM, and HDR images are reconstructed through a regression-based approach. LF-Diff\\(_{s1}\\) is a pretrained model in the first stage utilizing ground-truth images to provide LPR. Compared to the baseline, this brings a 8.7dB PSNR-L improvement, demonstrating that an accurate compact prior can greatly enhance results. Variant LF-Diff\\(_{s2}\\) w/o DM does not use the DM, but instead directly learns the LPR using LPENet. This achieves a 0.16dB PSNR gain over the baseline. When learning the LPR based" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25738" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.101, + 0.089, + 0.875, + 0.151 + ], + "angle": 0, + "content": "
Methods\nEnvironmentSen [25]\n(CPU)Hu [9]\n(CPU)Kalantari [11]\n(CPU+GPU)AHDR [41]\n(GPU)NHDRR [42]\n(GPU)HDRGAN [21]\n(GPU)APNT [1]\n(GPU)CA-ViT [16]\n(GPU)DiffHDR [44]\n(GPU)Ours\n(GPU)
Times(s)61.81s79.77s29.14s0.30s0.31s0.29s2.62s5.34s7.53s0.72s
Para.(M)--0.30M1.24M38.10M2.56M37.98M1.22M74.99M7.48M
" + }, + { + "type": "table_caption", + "bbox": [ + 0.158, + 0.161, + 0.812, + 0.175 + ], + "angle": 0, + "content": "Table 2. Average runtime performance of various methods on the testing set with dimensions of \\( {1000} \\times {1500} \\) ." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.197, + 0.473, + 0.364 + ], + "angle": 0, + "content": "on the DM, LF-Diff\\(_{s2}\\) w/ DM further improves PSNR by 0.24dB compared to variant LF-Diff\\(_{s2}\\) w/o DM. This indicates that the DM demonstrates superior capability in accurately estimating the distribution for LPR prediction. Additionally, compared to DiffHDR, which requires over 70M Params for estimating complete images, LF-Diff only adds 2.39M Params over the baseline. Moreover, we present visual comparisons of the baseline (without the prior feature) and LF-Diff variants in Fig. 6. It can be observed that performing DM in a compact latent space to predict LPR effectively mitigates ghosting issues." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.383, + 0.472, + 0.595 + ], + "angle": 0, + "content": "Effects of Joint Training Strategy. We conducted an ablation study on the joint training strategy. Under this strategy, only the DM is optimized in the second stage, referred to as Split-Training. Specifically, we first utilize the pretrained LPERNet to generate the prior feature \\( z \\) from the ground truth, and then apply the training objective defined in Eq. (15) to independently train the DM. Subsequently, the trained DM is directly integrated with the DHRNet for evaluation. It's worth noting that the DHRNet in the second stage utilizes pre-trained weights from the stage one without additional training. It is evident that LF-Diff significantly outperforms Split-Training by 1.51 dB in terms of PSNR value. This comparison underscores the effectiveness of joint training strategy over Split-Training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Sampling Step. In addition to the lightweight DM structure and compact feature dimensions, our method achieves greater computational efficiency by employing a smaller sampling step, denoted as S, in the reverse process based on DDIM [27]. The performance of our method is demonstrated with different values of S ranging from 5 to 100, as presented in Tab. 4. Variations in sampling steps, when S is set to a value smaller than the training configuration, a noticeable performance degradation occurs. On the other hand, when S is larger than the training setting, the impact on performance is minimal, particularly with the SSIM metric remaining consistently unchanged. While a larger sampling step is known to enhance the visual quality of images in diffusion-based methods [24, 44], in our case, it primarily increases inference time. This observation suggests that larger sampling steps, such as \\( \\mathrm{S} = 1000 \\) in DiffHDR [44] and SR3 [24], may not be essential for diffusion-based HDR reconstruction and other related low-level vision tasks when using our proposed framework." + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.194, + 0.891, + 0.282 + ], + "angle": 0, + "content": "
MethodGTDMJoint TrainingPara.PSNR-μPSNR-L
BaselineXXX5.09M44.2342.19
LF-Diffs1✓XX5.24M50.4250.89
Split trainingX✓X7.48M41.9339.30
LF-Diffs2 w/o DMXX✓5.85M44.4542.35
LF-Diffs2 w/ DMX✓✓7.48M44.7642.59
" + }, + { + "type": "table_caption", + "bbox": [ + 0.508, + 0.293, + 0.884, + 0.307 + ], + "angle": 0, + "content": "Table 3. Ablation studies of Diffusion Prior and Joint Training." + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.32, + 0.889, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.547, + 0.455, + 0.846, + 0.47 + ], + "angle": 0, + "content": "Figure 6. Qualitative results of our ablation study." + }, + { + "type": "table", + "bbox": [ + 0.513, + 0.483, + 0.885, + 0.564 + ], + "angle": 0, + "content": "
Sampling stepPSNR-μPSNR-LSSIM-μSSIM-LTime(s)
S=539.7641.330.98190.98780.049
S=1044.7642.590.99190.99060.106
S=2044.7342.540.99190.99060.208
S=5044.6642.450.99190.99060.479
S=10044.6542.440.99190.99060.971
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.574, + 0.892, + 0.617 + ], + "angle": 0, + "content": "Table 4. Ablation studies of various settings on the sampling step during the reverse process. Time(s) only denotes the time expenditure of the DM for the corresponding setting." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.634, + 0.62, + 0.65 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.895, + 0.903 + ], + "angle": 0, + "content": "The potential of diffusion models in HDR deghosting has shown promising results, particularly in achieving visually perceptible outcomes. Different from image synthesis which generates each pixel from scratch, HDR imaging provides several LDR images as reference. Thus, it is inefficient to reconstruct the full HDR image starting from pure Gaussian noise. In this paper, we propose an efficient diffusion model called LF-Diff, consisting of LPENet, DHRNet, and a denoising network. Specifically, we apply the DM in a compact latent space to predict low-frequency priors of HDR images. These prior features provide explicit guidance for the image reconstruction process, thereby enhancing the details of the reconstructed HDR images. Compared to traditional DM-based methods, LF-Diff achieves accurate estimations and reduces artifacts in reconstructed images with much lower computational cost." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25739" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Jie Chen, Zaifeng Yang, Tsz Nam Chan, Hui Li, Junhui Hou, and Lap-Pui Chau. Attention-guided progressive neural texture fusion for high dynamic range image restoration. IEEE Transactions on Image Processing, 31:2661-2672, 2022. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.185, + 0.472, + 0.242 + ], + "angle": 0, + "content": "[2] Nanxin Chen, Yu Zhang, Heiga Zen, Ron J Weiss, Mohammad Norouzi, and William Chan. Wavegrad: Estimating gradients for waveform generation. In International Conference on Learning Representations, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.471, + 0.297 + ], + "angle": 0, + "content": "[3] Zheng Chen, Yulun Zhang, Ding Liu, Jinjin Gu, Linghe Kong, Xin Yuan, et al. Hierarchical integration diffusion model for realistic image deblurring. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.471, + 0.354 + ], + "angle": 0, + "content": "[4] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In Advances in Neural Information Processing Systems, pages 8780-8794. Curran Associates, Inc., 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.354, + 0.471, + 0.396 + ], + "angle": 0, + "content": "[5] Li Fang, Qian Wang, and Long Ye. Glgnet: light field angular superresolution with arbitrary interpolation rates. Visual Intelligence, 2(1):6, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.396, + 0.471, + 0.452 + ], + "angle": 0, + "content": "[6] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.453, + 0.471, + 0.508 + ], + "angle": 0, + "content": "[7] YongSeok Heo, KyoungMu Lee, SangUk Lee, Youngsu Moon, and Joonhyuk Cha. Ghost-free high dynamic range imaging. In IEEE Conference on Asian Conference on Computer Vision (ACCV), pages 486-500, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.509, + 0.471, + 0.564 + ], + "angle": 0, + "content": "[8] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, pages 6840-6851. Curran Associates, Inc., 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.471, + 0.62 + ], + "angle": 0, + "content": "[9] Jun Hu, O. Gallo, K. Pulli, and Xiaobai Sun. HDR deghosting: How to deal with saturation? In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1163-1170, 2013. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.621, + 0.472, + 0.704 + ], + "angle": 0, + "content": "[10] Jinhan Hu, Gyeongmin Choe, Zeeshan Nadir, Osama Nabil, Seok-Jun Lee, Hamid Sheikh, Youngjun Yoo, and Michael Polley. Sensor-realistic synthetic data engine for multiframe high dynamic range photography. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 516-517, 2020. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.705, + 0.471, + 0.746 + ], + "angle": 0, + "content": "[11] Nima Khademi Kalantari and Ravi Ramamoorthi. Deep high dynamic range imaging of dynamic scenes. ACM Transactions on Graphics, 36(4):1-12, 2017. 1, 2, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.747, + 0.471, + 0.801 + ], + "angle": 0, + "content": "[12] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. Advances in Neural Information Processing Systems, 35:23593-23606, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.802, + 0.471, + 0.859 + ], + "angle": 0, + "content": "[13] Haoying Li, Yifan Yang, Meng Chang, Shiqi Chen, Huajun Feng, Zhihai Xu, Qi Li, and Yueting Chen. Srdiff: Single image super-resolution with diffusion probabilistic models. Neurocomputing, 479:47-59, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.86, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[14] Ru Li, Chuan Wang, Jue Wang, Guanghui Liu, Heng-Yu Zhang, Bing Zeng, and Shuaicheng Liu. Uphdr-gan: Generative adversarial network for high dynamic range imaging" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "with unpaired data. IEEE Transactions on Circuits and Systems for Video Technology, 32(11):7532-7546, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.894, + 0.206 + ], + "angle": 0, + "content": "[15] Zhen Liu, Wenjie Lin, Xinpeng Li, Qing Rao, Ting Jiang, Mingyan Han, Haoqiang Fan, Jian Sun, and Shuaicheng Liu. Adnet: Attention-guided deformable convolutional network for high dynamic range imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 463-470, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[16] Zhen Liu, Yinglong Wang, Bing Zeng, and Shuaicheng Liu. Ghost-free high dynamic range imaging with context-aware transformer. pages 344-360, 2022. 1, 2, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.25, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[17] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.321, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[18] Zhisheng Lu, Juncheng Li, Hong Liu, Chaoyan Huang, Linlin Zhang, and Tieyong Zeng. Transformer for single image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 457-466, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.391, + 0.892, + 0.461 + ], + "angle": 0, + "content": "[19] Ziwei Luo, Fredrik K Gustafsson, Zheng Zhao, Jens Sjolund, and Thomas B Schön. Refusion: Enabling large-size realistic image restoration with latent-space diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1680-1691, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.462, + 0.892, + 0.518 + ], + "angle": 0, + "content": "[20] Rafat Mantiuk, Kil Joong Kim, Allan G. Rempel, and Wolfgang Heidrich. HDR-VDP-2:a calibrated visual metric for visibility and quality predictions in all luminance conditions. In ACM Siggraph, pages 1–14, 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.519, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[21] Yuzhen Niu, Jianbin Wu, Wenxi Liu, Wenzhong Guo, and Rynson WH Lau. Hdr-gan: Hdr image reconstruction from multi-exposed ldr images with large motions. IEEE Transactions on Image Processing, 30:3885-3896, 2021. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.59, + 0.892, + 0.646 + ], + "angle": 0, + "content": "[22] Feiyue Peng, Maojun Zhang, Shiming Lai, Hanlin Tan, and Shen Yan. Deep hdr reconstruction of dynamic scenes. In 2018 IEEE 3rd International Conference on Image, Vision and Computing (ICIVC), pages 347-351. IEEE, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.647, + 0.894, + 0.716 + ], + "angle": 0, + "content": "[23] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[24] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1, 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.775, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[25] Pradeep Sen, Nima Khademi Kalantari, Maziar Yaesoubi, Soheil Darabi, Dan B Goldman, and Eli Shechtman. Robust patch-based hdr reconstruction of dynamic scenes. ACM Trans. Graph., 31(6):203-1, 2012. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[26] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In Proceedings of the 32nd International Conference on Machine Learning, pages 2256-2265, 2015. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "25740" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.16 + ], + "angle": 0, + "content": "[27] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.472, + 0.247 + ], + "angle": 0, + "content": "[28] Jou Won Song, Ye-In Park, Kyeongbo Kong, Jaeho Kwak, and Suk-Ju Kang. Selective transhdr: Transformer-based selective hdr imaging using ghost region mask. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVII, pages 288-304. Springer, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.47, + 0.303 + ], + "angle": 0, + "content": "[29] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.47, + 0.387 + ], + "angle": 0, + "content": "[30] Steven Tel, Zongwei Wu, Yulun Zhang, Barthélémy Heyrman, Cédric Demonceaux, Radu Timofte, and Dominique Gingac. Alignment-free hdr deghosting with semantics consistent transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12836-12845, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.39, + 0.469, + 0.445 + ], + "angle": 0, + "content": "[31] Anna Tomaszewska and Radoslaw Mantiuk. Image registration for multi-exposure high dynamic range image acquisition. In International Conference in Central Europe on Computer Graphics and Visualization, WSCG'07, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.447, + 0.469, + 0.5 + ], + "angle": 0, + "content": "[32] Okan Tarhan Tursun, Ahmet Oğuz Akyüz, Aykut Erdem, and Erkut Erdem. An objective dehosting quality metric for HDR images. Comput. Graph. Forum, 35(2):139-152, 2016. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.469, + 0.559 + ], + "angle": 0, + "content": "[33] Okan Tarhan Tursun, Ahmet Oğuz Akyüz, Aykut Erdem, and Erkut Erdem. An objective deghosting quality metric for hdr images. In Computer Graphics Forum, pages 139-152. Wiley Online Library, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.561, + 0.47, + 0.63 + ], + "angle": 0, + "content": "[34] Haohan Wang, Xindi Wu, Zeyi Huang, and Eric P Xing. High-frequency component helps explain the generalization of convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8684-8694, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.469, + 0.686 + ], + "angle": 0, + "content": "[35] Yinhuai Wang, Jiwen Yu, and Jian Zhang. Zero-shot image restoration using denoising diffusion null-space model. In The Eleventh International Conference on Learning Representations, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.469, + 0.73 + ], + "angle": 0, + "content": "[36] Greg Ward. Fast, robust image registration for compositing high dynamic range photographs from hand-held exposures. Journal of Graphics Tools, 8, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.469, + 0.801 + ], + "angle": 0, + "content": "[37] Jay Whang, Maurizio Delbracio, Hossein Talebi, Chitwan Sahara, Alexandros G Dimakis, and Peyman Milanfar. Deblurring via stochastic refinement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16293-16303, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[38] Shangzhe Wu, Jiarui Xu, Yu-Wing Tai, and Chi-Keung Tang. Deep high dynamic range imaging with large foreground motions. In European Conference on Computer Vision (ECCV), 2018. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[39] Bin Xia, Yulun Zhang, Shiyin Wang, Yitong Wang, Xinglong Wu, Yapeng Tian, Wenming Yang, and Luc Van Gool. Diffir: Efficient diffusion model for image restoration. In" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13095-13105, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.163 + ], + "angle": 0, + "content": "[40] Qingsen Yan, Jinqiu Sun, Haisen Li, Yu Zhu, and Yanning Zhang. High dynamic range imaging by sparse representation. Neurocomputing, 269:160-169, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.165, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[41] Qingsen Yan, Dong Gong, Qinfeng Shi, Anton van den Hengel, Chunhua Shen, Ian Reid, and Yanning Zhang. Attention-guided network for ghost-free high dynamic range imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1751-1760, 2019. 1, 2, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[42] Qingsen Yan, Lei Zhang, Yu Liu, Yu Zhu, Jinqiu Sun, Qinfeng Shi, and Yanning Zhang. Deep hdr imaging via a non-local network. IEEE Transactions on Image Processing, 29: 4308-4322, 2020. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[43] Qingsen Yan, Weiye Chen, Song Zhang, Yu Zhu, Jinqiu Sun, and Yanning Zhang. A unified hdr imaging method with pixel and patch level. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22211-22220, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.376, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[44] Qingsen Yan, Tao Hu, Yuan Sun, Hao Tang, Yu Zhu, Wei Dong, Luc Van Gool, and Yanning Zhang. Towards high-quality hdr deghosting with conditional diffusion models. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2023. 1, 2, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.447, + 0.892, + 0.528 + ], + "angle": 0, + "content": "[45] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728-5739, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.531, + 0.892, + 0.599 + ], + "angle": 0, + "content": "[46] Xiang Zhang, Tao Hu, Jiashuang He, and Qingsen Yan. Efficient content reconstruction for high dynamic range imaging. In ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7660-7664, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.601, + 0.892, + 0.671 + ], + "angle": 0, + "content": "[47] Xiang Zhang, Qiang Zhu, Tao Hu, and Qingsen Yan. Eiffhdr: An efficient network for multi-exposure high dynamic range imaging. In ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6560-6564, 2024. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.671 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25741" + } + ] +] \ No newline at end of file diff --git a/2024/Generating Content for HDR Deghosting from Frequency View/2b6f8409-576d-47cf-ab70-c1b4b9e76809_origin.pdf b/2024/Generating Content for HDR Deghosting from Frequency View/2b6f8409-576d-47cf-ab70-c1b4b9e76809_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..69f4bea251510cfd45f33616b92778cf79138684 --- /dev/null +++ b/2024/Generating Content for HDR Deghosting from Frequency View/2b6f8409-576d-47cf-ab70-c1b4b9e76809_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:83af11444cc017c2d91f4fbf1ca219ba79112e085774532da7fa44efc18b823b +size 4886674 diff --git a/2024/Generating Content for HDR Deghosting from Frequency View/full.md b/2024/Generating Content for HDR Deghosting from Frequency View/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d51b6cd6a6edbbaff9c8848f3979cab77d7c7981 --- /dev/null +++ b/2024/Generating Content for HDR Deghosting from Frequency View/full.md @@ -0,0 +1,448 @@ +# Generating Content for HDR Deghosting from Frequency View + +Tao Hu $^{1\dagger}$ Qingsen Yan $^{1\dagger}$ * Yuankai Qi $^{2}$ Yanning Zhang $^{1}$ + +$^{1}$ Northwestern Polytechnical University $^{2}$ Macquarie University + +hutaoxauat@gmail.com, $\mathbb{1}\{$ qingsenyan,ynzhang $\} \text{=} \mathrm{nwpu.edu.cn},$ yuankai.qi@mq.edu.au + +# Abstract + +Recovering ghost-free High Dynamic Range (HDR) images from multiple Low Dynamic Range (LDR) images becomes challenging when the LDR images exhibit saturation and significant motion. Recent Diffusion Models (DMs) have been introduced in HDR imaging field, demonstrating promising performance, particularly in achieving visually perceptible results compared to previous DNN-based methods. However, DMs require extensive iterations with large models to estimate entire images, resulting in inefficiency that hinders their practical application. To address this challenge, we propose the Low-Frequency aware Diffusion (LF-Diff) model for ghost-free HDR imaging. The key idea of LF-Diff is implementing the DMs in a highly compacted latent space and integrating it into a regression-based model to enhance the details of reconstructed images. Specifically, as low-frequency information is closely related to human visual perception we propose to utilize DMs to create compact low-frequency priors for the reconstruction process. In addition, to take full advantage of the above low-frequency priors, the Dynamic HDR Reconstruction Network (DHRNet) is carried out in a regression-based manner to obtain final HDR images. Extensive experiments conducted on synthetic and real-world benchmark datasets demonstrate that our LF-Diff performs favorably against several state-of-the-art methods and is $10\times$ faster than previous DM-based methods. + +# 1. Introduction + +Multiple exposure High Dynamic Range (HDR) imaging aims to restore missing details from exposure-varied Low Dynamic Range (LDR) images. However, in dynamic scenes, it often causes ghosting artifacts due to object or camera movement, limiting practical applications. Researchers are actively exploring ghost-free image recon + +![](images/6195452a8473ddc281ee196038780c732efb56a605e8f53e9e275bb1b69c42a1.jpg) +Figure 1. Average PSNR vs. inference time on the Kalantari's dataset [11]. Our method performs favorably and is $10\times$ faster than the previous diffusion-model-based method DiffHDR [44]. + +struction methods to enable seamless capture of dynamic scenes with high dynamic ranges. + +In recent years, the rise of Deep Neural Networks (DNNs) has brought significant advancements to HDR imaging field. Numerous DNN-based methods have emerged, employing Convolutional Neural Networks (CNNs) [11, 41, 47] or Vision Transformers (ViTs) [16, 28] for HDR image reconstruction. Despite their advancements, DNN-based approaches face challenges when essential information (e.g., content, details) for overexposed areas is missing due to object or camera movement. More recently, Diffusion Models (DMs) [8, 26] have exhibited impressive performance in image synthesis [4, 29] and image reconstruction tasks (including HDR imaging) [24, 44, 46]. DMs achieve this by iteratively denoising Gaussian noise to generate high-fidelity images. Yan et al. [44] reconstruct complete HDR images from pure Gaussian noise, yielding impressive results. Note that, compared to previous Deep generative models [14, 21], DMs generate a more accurate target distribution without facing issues like optimization instability or mode collapse, making them a promising choice for enhancing HDR image quality. Nonetheless, DMs demand a substantial number of iteration steps on denoising models to capture intricate data details, which is time-consuming even with a high-end GPU card. For example, as shown in Fig. 1, DiffHDR [44] takes about 7.5 seconds to generate an HDR image on a single A100 GPU. + +In this paper, our goal is to devise a DM-based method that effectively harnesses the robust distribution mapping capabilities of DMs for HDR image reconstruction. We note that it is unnecessary to reconstruct a complete HDR image from pure noise, because the LDR reference images already provide most of the necessary content information for HDR images. As low-frequency information is closely related to human visual perception [34], we propose to utilize DMs to create compact low-frequency priors in the latent space. These priors are then integrated into the regression-based model to predict the low-frequency content of the reconstructed HDR images. + +To achieve our goal, we propose the Low-Frequency Aware Diffusion Model (LF-Diff), comprising a Low-Frequency Prior Extraction Network (LPENet), a denoising network, and a Dynamic HDR Reconstruction Network (DHRNet). Following established methods [3, 23, 39], we employ a two-stage training strategy for pre-train LF-Diff and DM training. In the first stage, LPENet learns to extract a compact Low-frequency Prior Representation (LPR) from ground-truth images, guiding the DHRNet. DHRNet consists of two modules: a Prior Integration Module (PIM) that fuses the LPR with intermediate features of DHRNet, and a Feature Refinement Module (FRM) that further processes the fused features to HDR image. Notably, LPENet and DHRNet are optimized together in this phase. In the second stage, we train DM to learn the compact LPR directly from LDR images. Since the LPR is lightweight and only used to low-frequency content for HDR imaging, our DM can estimate the LPR with extremely low computational cost, ensuring stable visual results after multiple iterations. + +The main contributions can be summarized as follows: + +- We introduce LF-Diff, a straightforward and efficient DM-based method for HDR imaging. LF-Diff leverages the capabilities of diffusion models to generate informative low-frequency priors, which are then integrated into the imaging process to enhance results. +- We propose PIM and FRM in the DHRNet to fully exploit the LPR. While PIM efficiently fuses the LPR with intermediate features of DHRNet, FRM further processes the fused features to reconstruct high-quality HDR images. +- Extensive experiments demonstrate that the proposed LF-Diff method achieves SOTA performance in HDR imaging tasks, and produces visually appealing results that align with human visual perception. But consuming significantly fewer computational resources compared to other DM-based methods. + +# 2. Related Work + +Ghost-free HDR Reconstruction. Traditional methods often employ motion rejection [7, 40], motion registration [31, 36], and patch matching [9, 25] approaches to align LDR images and reconstruct high-quality results. However, + +these methods heavily rely on the performance of preprocessing techniques and often face challenges when dealing with motions across large spatial extents. Subsequently, DNN-based methods have emerged as the mainstream approach, thanks to their superior nonlinear expression capability. Researchers have explored various applications of DNNs in HDR imaging, devising sophisticated network architectures and model optimization schemes, such as attention [1, 41], transformer [16, 28, 30], optical flow [11, 22], GAN [6, 14, 21], and others [5, 15, 42, 43, 47]. Nevertheless, when LDR images lack sufficient information due to motion or saturation, they often manifest artifacts not aligned with human visual perception, commonly referred to as ghosting. + +Diffusion Models. Diffusion models have recently demonstrated promising results in various low-level vision tasks, including super-resolution [13, 24], HDR reconstruction [44, 46], colorization [12, 35], and deblurring [37]. DiffHDR [44] applied the conventional diffusion model paradigm to reconstruct HDR images, yielding remarkable outcomes. However, due to the high computational cost of diffusion models along with the requirement for numerous iterative steps, inference speed is limited. Despite efforts to mitigate this constraint [2, 17, 27], the overall complexity remains high, especially for high-resolution images commonly encountered in HDR scenarios. Recently, several methods [3, 23, 39] have addressed this issue by conducting diffusion modeling in the latent space. Rombach et al. [23] employed an autoencoder model to compress images into features equivalent to the image space. HI-Diff [3] leverages diffusion-based vector features for assisting image deblurring. However, given the unique characteristics of HDR imaging, further exploration is needed to select an appropriate latent space for HDR reconstruction tasks. + +# 3. Preliminaries + +The diffusion model introduced by Sohl-Dickstein et al. [26] is inspired by nonequilibrium thermodynamics. Here, we provide a brief overview of the "variance-preserving" diffusion model from [8], which encompasses diffusion processes and reverse processes. + +Diffusion Process. Given a clean image distribution, the diffusion process gradually injects isotropic Gaussian noise to generate $x_{t}$ according to a variance schedule + +$$ +\beta_ {1}, \dots , \beta_ {T} \in (0, 1). \text {L e t} \alpha_ {t} = 1 - \beta_ {t}, \bar {\alpha} _ {t} = \prod_ {i = 1} ^ {T} \alpha_ {i}: +$$ + +$$ +q \left(x _ {t} \mid x _ {0}\right) = \mathcal {N} \left(x _ {t}; \sqrt {\bar {\alpha} _ {t}} x _ {0}, (1 - \bar {\alpha} _ {t}) \mathbf {I}\right). \tag {1} +$$ + +The closed-form for this process can be represented as: + +$$ +x _ {t} \left(x _ {0}, \epsilon_ {t}\right) = \sqrt {\bar {\alpha} _ {t}} x _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon_ {t}, \epsilon \in (0, \mathbf {I}), \tag {2} +$$ + +where $x_{t}$ represents the output at time $t$ , $x_0$ is the initial clean image, and $\epsilon_t$ denotes isotropic Gaussian noise. + +![](images/f5af3edc5e26d8545e8d3d2375df16679a33e8746d850afd5e159841d0002b6c.jpg) + +![](images/9d3c16e98df4a6cc282eca153975e7f37b222b385af8565e2503f392c9b0e05d.jpg) + +![](images/763c7d3917bd418d786c515129ce3dd3e733dcc7876e9ae5c25797e8bca0da87.jpg) +Figure 2. The proposed LF-Diff comprises DHRNet, LPRNet, and a denoising network. LF-Diff undergoes two training stages: Pretrain LF-Diff (Sec. 4.1) and DM training (Sec. 4.2). Notably, during the inference stage, we do not input the ground-truth image into $\mathrm{LPONet}_{DM}$ and the denoising networks. Instead, we solely utilize the reverse process of DMs. + +Reverse Process. The reverse process attempts to approximate the data distribution $q(x_0)$ by a Markovian process, starting with random Gaussian noise $x_{T} = \mathcal{N}(x_{T};\mathbf{0},\mathbf{I})$ : + +$$ +q \left(x _ {t - 1} \mid x _ {t}, x _ {0}\right) = \mathcal {N} \left(x _ {t - 1}; \tilde {\mu} _ {t} \left(x _ {t}, x _ {0}\right), \tilde {\beta} _ {t} \mathbf {I}\right), \tag {3} +$$ + +where $\tilde{\mu}_t$ and $\tilde{\beta}_t$ are distribution parameters. Since the true reverse process Eq.(3) relies on $q(x_0)$ and is intractable, the neural network $f_{\theta}$ serves as the denoiser to estimate $p_{\theta}(x_{t - 1}|x_t,t)$ instead of $q(x_{t - 1}|x_t,x_0)$ : + +$$ +\begin{array}{l} p _ {\theta} \left(x _ {t - 1} \mid x _ {t}, t\right) = \mathcal {N} \left(x _ {t - 1}; \mu_ {\theta} \left(x _ {t}, t\right), \Sigma_ {\theta} \left(x _ {t}, t\right)\right) \tag {4} \\ = q \left(x _ {t - 1} \mid f _ {\theta} (x _ {t}, t)\right), \\ \end{array} +$$ + +where $\mu_{\theta}$ and $\Sigma_{\theta}$ are the parameters to be estimated in the reverse process. Let fixed variances $(\Sigma_{\theta}(x_t,t) = \tilde{\beta}_t\mathbf{I})$ , the optimization objective can be defined as: + +$$ +\mathbb {E} _ {t, x _ {0}, \epsilon_ {t}} [ \| \epsilon_ {t} - f _ {\theta} (\sqrt {\bar {\alpha} _ {t}} x _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon_ {t}, t) \| ^ {2} ]. \tag {5} +$$ + +# 4. Methodology + +As depicted in Fig. 2, we introduce the proposed LF-Diff framework, which consists of two training stages. During the first stage, we pre-train LF-Diff to extract accurate low-frequency prior representations from ground truths using LPenet, while DHRNet learns how to utilize the above priors to reconstruct HDR images. During the second stage, we train the DM to accurately predict the low-frequency prior representations directly from LDR images, and jointly optimize it with DHRNet to generate final high-quality HDR images. + +# 4.1. Stage One: Pretrain LF-Diff + +In the first stage, our objective is to train the LPERNet how to extract accurate low-frequency prior features from ground truth and embed this representation into the DHRNet to guide the reconstruction process. As illustrated in Fig. 2 (a), we utilize the Prior Integration Module (PIM) to incorporate these prior features into the DHRNet via the cross-attention mechanism. Within DHRNet, multiple Feature Refinement Modules (FRM) are stacked to directly learn the mapping relationship from input features to the target image distribution, while PIM leverages prior features to assist in recovering the low-frequency content of the reconstructed images. Next, we elaborate on above descriptions. + +Low-frequency Prior Extraction Network. The LPENet has a straightforward structure, as shown in Fig. 2 (a), comprising several residual blocks designed to extract the Low-frequency Prior Representation (LPR). Starting with the ground truth image $I_{gt} \in \mathbb{R}^{H \times W \times 3}$ , we initially perform a tonemapping operation $\mathcal{T}(\cdot)$ to obtain the LDR domain ground truth image $\mathcal{T}(I_{gt})$ . We concatenate these images along the channel dimension and employ the PixelUnshuffle operation to downsample, generating the input for LPENet. Subsequently, LPENet extracts the LPR $z \in \mathbb{R}^{\frac{H}{k} \times \frac{W}{k} \times 3}$ as follows: + +$$ +z = \operatorname {L P E N e t} (\operatorname {P i x i l e U n s h u r f l e} (\operatorname {C o n c a t} (I _ {g t}, \mathcal {T} (I _ {g t})))) \tag {6} +$$ + +where $\mathcal{T}(x) = \frac{\log(1 + \mu x)}{\log(1 + \mu)}$ denotes the tonemapping operator, $\mu = 5000$ is a predefined constant parameter that controls the degree of compression applied to the input signal, $k$ denotes the sampling multiple of PixelUnshuffle. + +Dynamic HDR Reconstruction Network. As illustrated in Fig. 2 (a), DHRNet consists of multiple stacked Reconstruction Blocks, with each block comprising one Prior Integration Module (PIM) and $N$ Feature Refinement Module (FRM). PIM fuses the LPR with intermediate features of DHRNet, we will give a detailed description in the following part. FRM is designed based on the understanding that natural images encompass diverse frequencies, since high and low frequencies serve distinct roles in image encoding. + +In Fig. 3 (a), the structure of FRM involves several blocks. Specifically, we begin by processing the features through a residual block to obtain $F_{n-1}$ . Subsequently, we decouple these features into two components: high-frequency features $F_{n-1}^{\mathrm{high}}$ , which remain resolution-invariant, and low-frequency features $F_{n-1}^{\mathrm{low}}$ with reduced resolution, following the approach described in [18]: + +$$ +F _ {n - 1} ^ {\text {l o w}} = \operatorname {A v g p o o l} \left(F _ {n - 1}, k\right), +$$ + +$$ +F _ {n - 1} ^ {\text {h i g h}} = F _ {n - 1} - \text {U p s a m p l e} \left(F _ {n - 1} ^ {\text {l o w}}\right), \tag {7} +$$ + +where $k$ denotes the kernel size of the avgpooling layer, Upsample denotes the upsampling operator. In essence, low-frequency features capture global dependencies within the input (image/features) and do not necessitate a high-resolution feature map, but require global attention. On the other hand, high-frequency features capture detailed local information, requiring a high-resolution feature map, and only can be achieved through local operators. Therefore, we employ the residual block (RB) to process $F_{n - 1}^{\mathrm{high}}$ and utilize the self-attention (SA) from [45] to handle $F_{n - 1}^{\mathrm{low}}$ . Subsequently, we fuse the low- and high-frequency features to preserve the initial details, resulting in the feature $F_{n - 1}^{\prime}$ . This operation can be represented as follows: + +$$ +F _ {n - 1} ^ {\prime} = \operatorname {C o n c a t} \left[ R B \left(F _ {n - 1} ^ {\text {h i g h}}\right), \operatorname {U p s a m p l e} \left(S A \left(F _ {n - 1} ^ {\text {l o w}}\right)\right) \right]. \tag {8} +$$ + +Considering that $F_{n-1}^{\prime}$ is a concatenation of two features, a $1 \times 1$ Conv is employed to effectuate a reduction in the number of channels. This downsizing step is followed by a channel attention module, designed to accentuate those channels exhibiting heightened activation values. + +In order to effectively merge the prior feature derived from the LPERNet with the intermediate feature produced by DHRNet, we replaced the self-attention mechanism in the low-frequency branch of FRM with a carefully designed cross-attention mechanism, resulting in the Prior Integration Module (PIM). As depicted in Fig. 2 (a), a PIM is positioned before each reconstruction block. Within each PIM, cross-attention computations are performed between the prior and intermediate features to facilitate feature fusion. This module enables the aggregation of information from the prior feature into the features of DHRNet. + +![](images/126d2748d194f331ae8b4872766c520c57b2da0b4a8fee7f036be1a223e945bd.jpg) +(a) Feature Refinement Module + +![](images/6dcd4f8b73cc7dc8103466d130b2ae4c17a8181ea480aa1e990d7930357d194e.jpg) +(b)Cross Self-attention module from PIM +Figure 3. The architecture of (a) Feature Refinement Module (FRM) and (b) Cross-attention mechanism. + +Specifically, as illustrated in Fig. 3 (b), our cross-attention module receives two types of input: the intermediate feature $\mathbf{F} \in \mathbb{R}^{\hat{H} \times \hat{W} \times \hat{C}}$ and the prior feature $z \in \mathbb{R}^{\hat{H} \times \hat{W} \times \hat{N}}$ . We project $\mathbf{F}$ into a query vector $\mathbf{Q} = W_d^Q W_c^Q\mathbf{F}$ using point-wise $1 \times 1$ convolution and depth-wise $3 \times 3$ convolution with weights $W_d^Q$ and $W_c^Q$ . Similarly, $z$ is transformed into key $\mathbf{K}$ and value $\mathbf{V}$ through analogous operations. Next, we reshape these projections into matrices suitable for attention computation: $\hat{\mathbf{Q}} \in \mathbb{R}^{\hat{H}\hat{W} \times \hat{C}}$ , $\hat{\mathbf{K}} \in \mathbb{R}^{\hat{N} \times \hat{H}\hat{W}}$ , and $\hat{\mathbf{V}} \in \mathbb{R}^{\hat{H}\hat{W} \times \hat{N}}$ . We then calculate a more computationally efficient attention map $A \in \mathbb{R}^{\hat{N} \times \hat{C}}$ by performing dot product between $\hat{\mathbf{Q}}$ and $\hat{\mathbf{K}}$ . The process can be described as follows: + +$$ +\hat {\mathbf {F}} = W _ {c} \hat {\mathbf {V}} \cdot \operatorname {S o f t m a x} (\hat {\mathbf {K}} \cdot \hat {\mathbf {Q}} / \gamma) + \mathbf {F}, \tag {9} +$$ + +where $\gamma$ is a trainable parameter. Notably, we do not implement multi-head attention in PIM. + +Training Strategy. As proposed in [11], we initially transform the provided set of input LDR images $\{L_1, L_2, \ldots, L_N\}$ into their corresponding HDR versions $H_i$ using gamma correction. Subsequently, we concatenate each $L_i$ with its corresponding $H_i$ along the channel axis to generate six-channel input tensors $X_i = [L_i, H_i]$ . Similar to [16, 41, 44], we employ an alignment module [41] to process the input LDR images, align the features implicitly, and feed them into DHRNet. Then, the LPR feature $z$ extracted from LPENet is injected into DHRNet via PIM. The reconstructed HDR image $\hat{H}$ is generated by: + +$$ +\hat {H} = \operatorname {C o n v} _ {3 \times 3} \left(\mathrm {D H R N e t} \left(X _ {i}, z\right)\right). \tag {10} +$$ + +Following previous approaches [16, 41], we utilize both + +tonemapped per-pixel loss and perceptual loss as the image reconstruction loss function $\mathcal{L}_r$ . This dual-loss strategy optimizes both pixel-level accuracy and high-level feature representations in the generated HDR results: + +$$ +\mathcal {L} _ {r} = \left\| \mathcal {T} (H) - \mathcal {T} (\hat {H}) \right\| _ {1} + \lambda \| \phi_ {i, j} (\mathcal {T} (H)) - \phi_ {i, j} (\mathcal {T} (\hat {H})) \| _ {1}, \tag {11} +$$ + +where $\phi_{i,j}(\cdot)$ signifies the $j$ -th convolutional feature extracted from the VGG19 network after the $i$ -th maxpooling operation, and $\lambda = 1e - 2$ is a hyperparameter that balances the contribution of the each component. + +# 4.2. Stage Two: Diffusion Models for HDR Imaging + +After the above-mentioned learning procedure, we now possess LDR images and their corresponding LPR $z$ . In the second stage (Fig. 2 (b, c)), our objective is to efficiently harness the powerful distribution estimation capability of the DMs. Specifically, we utilize the LPENet from stage one to capture the LPR as the denoising target of the DM. The DM will learn how to extract accurate LPR from the LDR inputs and perform joint optimization with DHRNet. + +Diffusion Model for LRP Learning. After capturing the LPR $z \in \mathbb{R}^{\frac{H}{k} \times \frac{W}{k} \times N}$ from the pretrained LPENet, we transform the clean LPR feature $z$ into a noisy version $z_T$ using Eq.(2), which can be described as: + +$$ +q \left(z _ {T} \mid z\right) = \mathcal {N} \left(z _ {T}; \sqrt {\bar {\alpha} _ {T}} z, (1 - \bar {\alpha} _ {T}) \mathbf {I}\right), \tag {12} +$$ + +where $T$ represents the total number of iterations, while $\bar{\alpha}_t$ and $\alpha$ denote the pre-defined variance schedule. + +In the reverse process, we iteratively generate the LPR from a pure Gaussian distribution, starting from $z_{T}$ and moving backward to $z_{0}$ , which utilizes the posterior distribution described in Eq. (3). Following previous work [24, 44], we utilize a neural network to estimate $p_{\theta}(x_{t - 1}|x_t,t)$ instead of $q(x_{t - 1}|x_t,x_0)$ for each step. Specifically, we first use LPENetDM to obtain a conditional feature $D\in \mathbb{R}^{\frac{H}{k}\times \frac{W}{k}\times N}$ from aligned LDR feature: + +$$ +\mathbf {D} = \operatorname {L P E N e t} _ {D M} (A M (\text {P i x e l U n s h u f f l e} (X _ {i}))). \tag {13} +$$ + +where $\mathrm{LPENet}_{DM}$ maintains a similar architecture to $\mathrm{LPENet}$ but with a modified input dimensionality for its first convolution layer, $AM(\cdot)$ denotes alignment module from [41]. The denoising neural network predicts the noise based on both $z_{t}$ and the derived conditional feature $D$ , i.e., $\epsilon_{\theta}(Concat(z_{t},\mathbf{D}),t)$ . Upon substituting this estimated noise term into the equation governing the reverse process Eq. (4), we arrive at the following sampling formula: + +$$ +\boldsymbol {z} _ {t - 1} = \frac {1}{\sqrt {\alpha_ {t}}} \left(\boldsymbol {z} _ {t} - \frac {1 - \alpha_ {t}}{\sqrt {1 - \bar {\alpha} _ {t}}} \boldsymbol {\epsilon} _ {\theta} \left(\boldsymbol {z} _ {t}, \mathbf {D}, t\right)\right) + \sqrt {1 - \alpha_ {t}} \boldsymbol {\epsilon} _ {t}, \tag {14} +$$ + +where $\epsilon_t \sim \mathcal{N}(0, \mathbf{I})$ . By iteratively sampling $z_t$ using Eq. (14) $T$ times, we progressively reconstruct the predicted LPR representation $z_0$ (i.e. $\hat{z}) \in \mathbb{R}^{\frac{H}{k} \times \frac{W}{k} \times N}$ . + +Training Strategy. Traditional DMs solely learn the probability distribution by optimizing the weighted variational bound (Eq. (5)), resulting in a slight deviation between the predicted prior and the actual prior. Integrating the DM directly with DHRNet might lead to a misalignment issue, thereby hampering the overall image processing performance. To address this, we joint training the DM and DHRNet. During each training iteration, we first sample the noise sample $z_{t}$ through the diffusion process (Eq. (2)). Given that our denoising neural network is lightweight, we use a reverse process for $S$ iterations based on the DDIM strategy [27] to generate the predicted prior feature $\hat{z}$ . This $\hat{z}$ guides DHRNet via PIM. The loss function in the second stage consists of the reconstruction loss $\mathcal{L}_r$ (Eq. (11)) and the diffusion loss $\mathcal{L}_{diff}$ : + +$$ +\begin{array}{l} \mathcal {L} _ {\text {d i f f}} = E _ {z _ {0}, t, \epsilon_ {t} \sim \mathcal {N} (\mathbf {0}, \mathbf {I})} \left[ \| \epsilon_ {t} - \epsilon_ {\theta} (z _ {t}, D, t) \| ^ {2} \right] \tag {15} \\ + \left\| \hat {z} - z \right\| _ {1}. \\ \end{array} +$$ + +# 4.3. Inference + +In the inference stage, $L P E N t_{D M}$ extracts a conditional feature $\mathbf{D}$ from aligned LDR feature (Eq. (13)), and we randomly sample a Gaussian noise $\hat{z}_T$ . Then, denoising neural network utilizes the $\hat{z}_T$ and $\mathbf{D}$ to estimate LPR after $S$ iterations (Eq. (14)) based on DDIM [27]. After that, DHRNet exploits the LPR to restore HDR images as Eq. (10). More details about the DM training and the model inference can be found in the supplementary material. + +# 5. Experiments + +# 5.1. Experimental Settings + +Datasets. All methods are trained using two publicly available datasets, employing identical training settings: Kalantari's dataset [11] and Hu's dataset [10]. Kalantari's dataset consists of 74 samples for training and 15 for testing, all captured under authentic environmental conditions. Each sample comprises three LDR images with exposure variations of $\{-2,0, +2\}$ or $\{-3,0, +3\}$ . In contrast, Hu's dataset is a synthetic dataset designed to emulate sensor realism, generated through a game engine. This dataset contains images captured at three distinct exposure levels $\{-2,0, +2\}$ , with our primary focus on dynamic scene images within this collection. Following the settings outlined in [10], we allocate the initial 85 samples for training, reserving the remaining 15 for testing. Moreover, to further validate the model's generalizability, we incorporate Sen et al.'s dataset [25] and Tursun et al.'s dataset [32] exclusively for qualitative assessment, as these datasets lack ground truths. + +Evaluation Metrics. To conduct a quantitative comparison, we utilize five objective metrics: PSNR- $\mu$ SSIM- $\mu$ PSNR-L,SSIM-L,and HDR-VDP-2 [20].Here, the sub + +![](images/7b7857606bd0387f1c721bafcfef56959bf9e71469029765249756ee6eb6e7fa.jpg) + +![](images/2ddb4897fdb8a274acf484ecc6b8847b275f9fd2e985b618a01f06da060e9562.jpg) +Tonemapped Result + +![](images/889b8682a37653b8985fd7d0b20754ddc71e43659c9b840ba5702687bf22c87e.jpg) + +![](images/7b7de5c855ab8fb37681575c80f9ca03faed921ff5a5333320e26564c373e28b.jpg) +LDR-Patches + +![](images/e5705c9ab45703d500c27cd44c24c2757207eb7d859ee6ffbf259d5deefe8a8a.jpg) + +![](images/088accdbb9e4ef077f43ecdb11edb2ebcc7cb42b5980ea1ad085efb213a8c184.jpg) + +![](images/3177ca47349bee3ca18d8dbbdd0855f0f3e8616b808cc2f477c234b9742e5025.jpg) + +![](images/fef82b101c723bff618b0870e9921d1bba6362a8ba934fe5ff5e3bf33bb18e48.jpg) + +![](images/131aaf67b61e8a48f7b2cce5d3631d67b178bc367dfad6df2026b7346ba180f8.jpg) + +![](images/7b3395093c25bafede3c26ce245f8d8462e6f28c91f593463313c77a07a18799.jpg) + +![](images/46fc7cf1113732f5c64d2766ba8f62317b0f85f3da4600ee1419c56e5fb416b2.jpg) + +![](images/56688c7157a7627407afdcc270e999c9dde76315e5d91451537382a9655bfdc0.jpg) + +![](images/691d89c830fe024b545b00633e37f4b33cd43524b16ff0ee6e208c778715c351.jpg) +Tonemapped Result + +![](images/9dc4e1ba591d95bff701a7ad0835c8f7573887aa6b15cbc9841b145facef45be.jpg) + +![](images/7161664df37b8d64e4b817c41e42fc2a31e91a5c2a2dd850c9fef3b569f1c093.jpg) + +![](images/c5bc93054f887c1c909db2dd0fac183d5cedd6d78bc4571b862847fab8db8b2a.jpg) + +![](images/bbcda58e6437ff10fdc6e3ac16378bed102a425af2882c90586e8f5ff0286ef1.jpg) +(a) Example from Kalantari et al.'s dataset [11] +Figure 4. Visual comparisons are conducted on testing data from various datasets, focusing on zoomed-in local areas of the HDR images estimated by our method and the compared techniques. Our model demonstrates the ability to generate HDR images of superior quality. + +![](images/1f80ff558967179b3390143af807153da5a5a44b924e3c870b710b695f0f5780.jpg) + +![](images/c49a2a7497c5df59e526d4b5999d4bc1b08176031adeb3801527a00983eed83b.jpg) +LDRs + +![](images/89028f331197adeb1d139b87b4ba5240d9a7ff4f99e753d1da267601d3bdf03a.jpg) + +![](images/18d5893b6e67f0908cf482fc1ab24ea5a37441c4c16b66e47659524d874ced7a.jpg) +AHDR HDR-GAN CA-VIT HyHDR Ours + +![](images/d087c4d450c606f3d187eeb38e7284ac40aabc114f7eb8e8027546e9d4bb2253.jpg) + +![](images/4e08ab3f33793659f9087c10abba19ba314bb40aa1b2d46e1e4b593769eb53f3.jpg) + +![](images/d0d16a55a034f2984ff6cbc6251a7cbd754894bd5e9b20c54ec087c79548b88d.jpg) + +![](images/d216f66dd070b38e234aebae39952d2ea1f72aa8374811f5fa8ecc61a78dfe8d.jpg) +LDR- +GT + +![](images/73ff7f8951f798bf64fa368fa3f80ecd74113a5823cacba4897d48e5d7de1e76.jpg) +Patches + +scripts $\mu$ and L denote that these metrics are computed in the tonemapped and linear domains, respectively. + +Implementation Details. Our implementation is conducted using PyTorch, and each training stage converges after 300 epochs on four NVIDIA A100 GPUs. We employ the Adam optimizer with an initial learning rate of 1e-4, which decays by a factor of 0.1 after every 50 epochs. The training dataset is processed by cropping $128 \times 128$ patches with a stride of 64, and the batch size is set to 64. For LF-Diff, we adopt a variant of the commonly used U-Net architecture from [19] as the denoising network, with 3 levels of blocks $\{2, 2, 2\}$ . During training, the time step $T$ is set to 200, and the implicit sampling step $S$ is set to 10 for both training and inference phases to achieve efficient restoration. In DHRNet, the parameters $N_{i} \in \{L_{1}, L_{2}, L_{3}\}$ are set to $\{3, 3, 3\}$ , and the channel $C$ is set to 60. The number of attention heads for the self-attention branch at the same level is set to $\{6, 6, 6\}$ , and the channel expansion factor in FFN is 2.66. The kernels for avgpool in PIM and FRM are set to 4 and 2, respectively. For LPENet, it comprises 4 residual blocks, with a Pixelunshuffle downsampling factor of 4 and the output channel of 3. + +# 5.2. Comparison with the State-of-the-art Methods + +In this section, we evaluate the performance of the proposed LF-Diff method and present experimental results to validate its reconstruction performance compared to state-of-the-art techniques. Specifically, we evaluate LF-Diff against two patch-based methods [9, 25], a optical flow-based method [11], five CNN-based methods [1, 15, 21, 38, 41], two ViT-based methods [16, 43], along with a diffusion-based method [44]. It's worth noting that all deep learning methods employ the same training dataset and settings for consistency. + +Datasets w/ Ground Truth. The quantitative outcomes for LF-Diff on two datasets are presented in Tab. 1. Our method is compared with various state-of-the-art approaches using the testing data from [11] and [10], which includes challenging samples characterized by saturated backgrounds and foreground motions. All quantitative results are averaged across testing images. Notably, LF-Diff exhibits a significant improvement over other leading methods, exceeding the performance of the DM-based method DiffHDR [44], by $0.65\mathrm{dB}$ and $0.86\mathrm{dB}$ in PSNR- $\mu$ and PSNR-L metrics, respectively, based on Kalantari's dataset [11]. Moreover, LF-Diff showcases superior performance compared to the runner-up method, HyHDR [43], with improvements of $0.28\mathrm{dB}$ and $0.19\mathrm{dB}$ in PSNR- $\mu$ and PSNR-L metrics, respectively, on Hu's dataset [10]. + +As shown in Fig. 4 (a) and (b), the datasets present some challenging samples due to information loss in the LDR images. Most existing approaches struggle in these areas, producing ghosting artifacts due to large motions and occlusions. Kalantari's method [11] and DHDR [38] struggle with background motion due to error-prone alignments (e.g., optical flow), resulting in undesirable ghosting. While HDR-GAN [21] exhibits noticeable ghosting artifacts around the balcony area and introduces erroneous color information, AHDR [41] aligns LDR images using convolutional spatial attention. However, it unintentionally suppresses valuable contextual information and encounters difficulties with significant motions in over/underexposed scenes. CA-ViT [16], relying on patch-based sampling, produces noticeable blocky ghosting. In contrast, aided by the DM, both Diff-HDR [44] and our LF-Diff generate HDR images aligned with human perception. Notably, our approach outperforms Diff-HDR with faster inference speeds and a smaller computational overhead. + +![](images/5d31bc04308f765093082c38c427f59d11e4e137a5bf02d46c7d039ce4b749a4.jpg) + +![](images/6a2573a58dbcf7e471412254c872a0162cb38a7aa750162fc101fb43b9e4b1a8.jpg) + +![](images/16c62efe2965acc99aa264c8d8da1a239cf5b08f6d6fc64808404db4e23b8cde.jpg) + +![](images/bf4603127f35737ccdb178cb012c6ae5d264b0caa86f767c2a8ebb899f6d8545.jpg) + +![](images/04d1f1950ced272c3d5f51529738b1d2b5d364fc85bd739d80e6deca2f98ba7a.jpg) + +![](images/396a71f5a929f18bf1a324b9b1df57f78866a3856be2ff19d6a5be690e50bb8f.jpg) + +![](images/dd9893aaef9588271472189f5e18d5bb86322584c41bf05f20d68de3fb338e0e.jpg) + +![](images/17afcbe2cb497284b3f30b74acac5519bb6463ec40e7cf9a1c06b430ca8286cf.jpg) + +![](images/8b7847e3449307c3911ae8b411c03c9ac7c7f756f25c3c08295bae7d0065a8b0.jpg) +Kalantari +DHDR + +![](images/a8d44f76469c82af20462c1db28f657f1c46f503205e43b633d1013798371fc4.jpg) +AHDR + +![](images/f8008a2662f63ce3e703b8104efb48464fc62204ff690fc843590e6f5f8a8730.jpg) +HDR-GAN + +![](images/9a9d5f0f8a02a48f5ee01093fc4c0d1ffe20c2aab2e4e0f4f8037a69fca5eb72.jpg) + +![](images/75f759117a9c60caa16f831e75a98005ca75150ede5b07bd238cbd26c49fc396.jpg) +APNET +Figure 5. Visual comparisons on the datasets without ground truth. + +![](images/bfbeddcc4ec3f11838d3a83d5468d17a3ff7d5aafa61ff021cd0a3eb5bd06430.jpg) +CA-VIT + +![](images/f74c6febd42f3236418fe6f283cfd1bba2c55ecdf26d8de4d319f34b0a1dde8e.jpg) +DiffHDR + +![](images/612eb62b9ef537b2388e0ffc764c96efd631edab9ad5f07a4a10c94ec3a65b0e.jpg) +Ours + +![](images/7e87e3b30d5113fb1c123b6f793b4f0d024b07715c50e9145cf9c61330602014.jpg) +Kalantari + +![](images/8b10266a3e0467f31acaa4a557a5f26717be9448174efcdfd44675d9fd2d4867.jpg) +DHDR + +![](images/91c8c49197b7d0e472c0aa5f21e7b23fff14ba29385331eb347305d0135e7a1b.jpg) +AHDR HDR-GAN + +![](images/3274dbf0b58470e46abb7c1349f0e3f8667c94f11a9ed0b9b320b3cb01a7d372.jpg) + +HDR-GAN + +![](images/0b8682997f38e13946d1d6a902a895178623bfbb8bcfeabebf8e7b7d89d105f5.jpg) +NET + +![](images/9f2f0efdf7068ec2a68408ca3044c9bea2d24ca535156cd1a80795533d3767cc.jpg) +CA-VIT + +![](images/a3266d1c8cc0965083a3d701884ed52827db4888a09b02b75b18d1c29b092b87.jpg) +DiffHDR + +![](images/5b10815ac919c9c7b3527fea620e87bc7bb3d206dc75870d1f22338e70ba09e3.jpg) +Ours + +
DatasetsModelsSen[25]Hu[9]Kalantari[11]DHDR[38]AHDR[41]NHDRR[42]HDRGAN[21]APNT[1]CA-ViT[16]HyHDR[43]DiffH[44]Ours
KalantariPSNR-μ40.9532.1942.7441.6443.6242.4143.9243.9444.3244.6444.1144.76
PSNR-L38.3130.8441.2240.9141.0341.0841.5741.6142.1842.4741.7342.59
SSIM-μ0.98050.97160.98770.98690.99000.98870.99050.98980.99160.99150.99110.9919
SSIM-L0.97260.95060.98480.98580.98620.98610.98650.98790.98840.98940.98850.9906
HDR-VDP-255.7255.2560.5160.5062.3061.2165.4564.0566.0366.0565.5266.54
DatasetsModelsSen[25]Hu[9]Kalantari[11]DHDR[38]AHDR[41]NHDRR[42]HDRGAN[21]APNT[1]ADNet[15]CA-ViT[16]HyHDR[43]Ours
HuPSNR-μ31.4836.5641.6041.1345.7645.1545.8646.4146.7948.1048.4648.74
PSNR-L33.5836.9443.7641.2049.2248.7549.1447.9750.3851.1751.9152.10
SSIM-μ0.95310.98240.99140.98700.99560.99560.99450.99530.99080.99470.99590.9968
SSIM-L0.96340.98770.99380.99410.99800.99810.99890.99860.99870.99890.99910.9993
HDR-VDP-266.3967.5872.9470.8275.0474.8675.1973.0676.2177.1277.2477.35
+ +Table 1. The evaluation results on Kalantari's [11] and Hu's [10] datasets. The best results are highlighted in Bold. + +Datasets w/o Ground Truth. To assess the generalization capability of our proposed HDR imaging method, we evaluated the performance of the model trained on Kalantari et al.'s Dataset [11] by testing it on datasets from Sen et al. [25] and Tursun et al. [33], which lack ground truths. In Fig. 5 (a), numerous current methods encounter difficulties in restoring recover the large saturated region and large motion. Conversely, in Fig. 5 (b), while complete elimination of ghosting remains elusive for any method, our approach notably enhances image sharpness and detail compared to preceding techniques. + +Computational Budgets. We also conducted comparisons regarding model parameters and inference times with previous approaches. As illustrated in Table 2, patch match-based methods [9, 25] exhibit significantly longer inference times due to their CPU-based computation. Kalantari [11] requires considerable time for initial optical flow alignment. NHDRRNet [42], employing a U-shaped network architecture, demonstrates shorter inference times but substantially higher parameter counts compared to other methods. CAViT [16] has a large number of standard transformer blocks leading to high computational cost despite fewer parame + +ters. DiffHDR [44] has high inference time and parameter count due to reconstructing HDR from pure noise. In comparison, our method effectively leverages the powerful distribution estimation capability of DM with orders of magnitude fewer parameters and computations. + +# 5.3. Ablation Study + +In this section, we investigate the impact of various designs within our proposed method. All experiments are conducted using the Kalantari's dataset [11]. + +Effects of Diffusion Prior. As shown in Tab. 3, we establish a regression-based baseline model without LPR generation. In this configuration, the PIM in DHRNet is replaced with FRM, and HDR images are reconstructed through a regression-based approach. LF-Diff $_{s1}$ is a pretrained model in the first stage utilizing ground-truth images to provide LPR. Compared to the baseline, this brings a 8.7dB PSNR-L improvement, demonstrating that an accurate compact prior can greatly enhance results. Variant LF-Diff $_{s2}$ w/o DM does not use the DM, but instead directly learns the LPR using LPENet. This achieves a 0.16dB PSNR gain over the baseline. When learning the LPR based + +
Methods +EnvironmentSen [25] +(CPU)Hu [9] +(CPU)Kalantari [11] +(CPU+GPU)AHDR [41] +(GPU)NHDRR [42] +(GPU)HDRGAN [21] +(GPU)APNT [1] +(GPU)CA-ViT [16] +(GPU)DiffHDR [44] +(GPU)Ours +(GPU)
Times(s)61.81s79.77s29.14s0.30s0.31s0.29s2.62s5.34s7.53s0.72s
Para.(M)--0.30M1.24M38.10M2.56M37.98M1.22M74.99M7.48M
+ +on the DM, LF-Diff $_{s2}$ w/ DM further improves PSNR by 0.24dB compared to variant LF-Diff $_{s2}$ w/o DM. This indicates that the DM demonstrates superior capability in accurately estimating the distribution for LPR prediction. Additionally, compared to DiffHDR, which requires over 70M Params for estimating complete images, LF-Diff only adds 2.39M Params over the baseline. Moreover, we present visual comparisons of the baseline (without the prior feature) and LF-Diff variants in Fig. 6. It can be observed that performing DM in a compact latent space to predict LPR effectively mitigates ghosting issues. + +Effects of Joint Training Strategy. We conducted an ablation study on the joint training strategy. Under this strategy, only the DM is optimized in the second stage, referred to as Split-Training. Specifically, we first utilize the pretrained LPERNet to generate the prior feature $z$ from the ground truth, and then apply the training objective defined in Eq. (15) to independently train the DM. Subsequently, the trained DM is directly integrated with the DHRNet for evaluation. It's worth noting that the DHRNet in the second stage utilizes pre-trained weights from the stage one without additional training. It is evident that LF-Diff significantly outperforms Split-Training by 1.51 dB in terms of PSNR value. This comparison underscores the effectiveness of joint training strategy over Split-Training. + +Sampling Step. In addition to the lightweight DM structure and compact feature dimensions, our method achieves greater computational efficiency by employing a smaller sampling step, denoted as S, in the reverse process based on DDIM [27]. The performance of our method is demonstrated with different values of S ranging from 5 to 100, as presented in Tab. 4. Variations in sampling steps, when S is set to a value smaller than the training configuration, a noticeable performance degradation occurs. On the other hand, when S is larger than the training setting, the impact on performance is minimal, particularly with the SSIM metric remaining consistently unchanged. While a larger sampling step is known to enhance the visual quality of images in diffusion-based methods [24, 44], in our case, it primarily increases inference time. This observation suggests that larger sampling steps, such as $\mathrm{S} = 1000$ in DiffHDR [44] and SR3 [24], may not be essential for diffusion-based HDR reconstruction and other related low-level vision tasks when using our proposed framework. + +Table 2. Average runtime performance of various methods on the testing set with dimensions of ${1000} \times {1500}$ . + +
MethodGTDMJoint TrainingPara.PSNR-μPSNR-L
BaselineXXX5.09M44.2342.19
LF-Diffs1✓XX5.24M50.4250.89
Split trainingX✓X7.48M41.9339.30
LF-Diffs2 w/o DMXX✓5.85M44.4542.35
LF-Diffs2 w/ DMX✓✓7.48M44.7642.59
+ +Table 3. Ablation studies of Diffusion Prior and Joint Training. + +![](images/3e88852fff0d036ef21d9874ce29f2276d4fa615db71bab1299e89c90e0e315a.jpg) + +Figure 6. Qualitative results of our ablation study. + +
Sampling stepPSNR-μPSNR-LSSIM-μSSIM-LTime(s)
S=539.7641.330.98190.98780.049
S=1044.7642.590.99190.99060.106
S=2044.7342.540.99190.99060.208
S=5044.6642.450.99190.99060.479
S=10044.6542.440.99190.99060.971
+ +Table 4. Ablation studies of various settings on the sampling step during the reverse process. Time(s) only denotes the time expenditure of the DM for the corresponding setting. + +# 6. Conclusion + +The potential of diffusion models in HDR deghosting has shown promising results, particularly in achieving visually perceptible outcomes. Different from image synthesis which generates each pixel from scratch, HDR imaging provides several LDR images as reference. Thus, it is inefficient to reconstruct the full HDR image starting from pure Gaussian noise. In this paper, we propose an efficient diffusion model called LF-Diff, consisting of LPENet, DHRNet, and a denoising network. Specifically, we apply the DM in a compact latent space to predict low-frequency priors of HDR images. These prior features provide explicit guidance for the image reconstruction process, thereby enhancing the details of the reconstructed HDR images. Compared to traditional DM-based methods, LF-Diff achieves accurate estimations and reduces artifacts in reconstructed images with much lower computational cost. + +# References + +[1] Jie Chen, Zaifeng Yang, Tsz Nam Chan, Hui Li, Junhui Hou, and Lap-Pui Chau. Attention-guided progressive neural texture fusion for high dynamic range image restoration. IEEE Transactions on Image Processing, 31:2661-2672, 2022. 2, 6, 7, 8 +[2] Nanxin Chen, Yu Zhang, Heiga Zen, Ron J Weiss, Mohammad Norouzi, and William Chan. Wavegrad: Estimating gradients for waveform generation. In International Conference on Learning Representations, 2020. 2 +[3] Zheng Chen, Yulun Zhang, Ding Liu, Jinjin Gu, Linghe Kong, Xin Yuan, et al. Hierarchical integration diffusion model for realistic image deblurring. Advances in Neural Information Processing Systems, 36, 2024. 2 +[4] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In Advances in Neural Information Processing Systems, pages 8780-8794. Curran Associates, Inc., 2021. 1 +[5] Li Fang, Qian Wang, and Long Ye. Glgnet: light field angular superresolution with arbitrary interpolation rates. Visual Intelligence, 2(1):6, 2024. 2 +[6] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 2 +[7] YongSeok Heo, KyoungMu Lee, SangUk Lee, Youngsu Moon, and Joonhyuk Cha. Ghost-free high dynamic range imaging. In IEEE Conference on Asian Conference on Computer Vision (ACCV), pages 486-500, 2011. 2 +[8] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, pages 6840-6851. Curran Associates, Inc., 2020. 1, 2 +[9] Jun Hu, O. Gallo, K. Pulli, and Xiaobai Sun. HDR deghosting: How to deal with saturation? In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1163-1170, 2013. 2, 6, 7, 8 +[10] Jinhan Hu, Gyeongmin Choe, Zeeshan Nadir, Osama Nabil, Seok-Jun Lee, Hamid Sheikh, Youngjun Yoo, and Michael Polley. Sensor-realistic synthetic data engine for multiframe high dynamic range photography. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 516-517, 2020. 5, 6, 7 +[11] Nima Khademi Kalantari and Ravi Ramamoorthi. Deep high dynamic range imaging of dynamic scenes. ACM Transactions on Graphics, 36(4):1-12, 2017. 1, 2, 4, 5, 6, 7, 8 +[12] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. Advances in Neural Information Processing Systems, 35:23593-23606, 2022. 2 +[13] Haoying Li, Yifan Yang, Meng Chang, Shiqi Chen, Huajun Feng, Zhihai Xu, Qi Li, and Yueting Chen. Srdiff: Single image super-resolution with diffusion probabilistic models. Neurocomputing, 479:47-59, 2022. 2 +[14] Ru Li, Chuan Wang, Jue Wang, Guanghui Liu, Heng-Yu Zhang, Bing Zeng, and Shuaicheng Liu. Uphdr-gan: Generative adversarial network for high dynamic range imaging + +with unpaired data. IEEE Transactions on Circuits and Systems for Video Technology, 32(11):7532-7546, 2022. 1, 2 +[15] Zhen Liu, Wenjie Lin, Xinpeng Li, Qing Rao, Ting Jiang, Mingyan Han, Haoqiang Fan, Jian Sun, and Shuaicheng Liu. Adnet: Attention-guided deformable convolutional network for high dynamic range imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 463-470, 2021. 2, 6, 7 +[16] Zhen Liu, Yinglong Wang, Bing Zeng, and Shuaicheng Liu. Ghost-free high dynamic range imaging with context-aware transformer. pages 344-360, 2022. 1, 2, 4, 6, 7, 8 +[17] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 2 +[18] Zhisheng Lu, Juncheng Li, Hong Liu, Chaoyan Huang, Linlin Zhang, and Tieyong Zeng. Transformer for single image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 457-466, 2022. 4 +[19] Ziwei Luo, Fredrik K Gustafsson, Zheng Zhao, Jens Sjolund, and Thomas B Schön. Refusion: Enabling large-size realistic image restoration with latent-space diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1680-1691, 2023. 6 +[20] Rafat Mantiuk, Kil Joong Kim, Allan G. Rempel, and Wolfgang Heidrich. HDR-VDP-2:a calibrated visual metric for visibility and quality predictions in all luminance conditions. In ACM Siggraph, pages 1–14, 2011. 5 +[21] Yuzhen Niu, Jianbin Wu, Wenxi Liu, Wenzhong Guo, and Rynson WH Lau. Hdr-gan: Hdr image reconstruction from multi-exposed ldr images with large motions. IEEE Transactions on Image Processing, 30:3885-3896, 2021. 1, 2, 6, 7, 8 +[22] Feiyue Peng, Maojun Zhang, Shiming Lai, Hanlin Tan, and Shen Yan. Deep hdr reconstruction of dynamic scenes. In 2018 IEEE 3rd International Conference on Image, Vision and Computing (ICIVC), pages 347-351. IEEE, 2018. 2 +[23] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2 +[24] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1, 2, 5, 8 +[25] Pradeep Sen, Nima Khademi Kalantari, Maziar Yaesoubi, Soheil Darabi, Dan B Goldman, and Eli Shechtman. Robust patch-based hdr reconstruction of dynamic scenes. ACM Trans. Graph., 31(6):203-1, 2012. 2, 5, 6, 7, 8 +[26] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In Proceedings of the 32nd International Conference on Machine Learning, pages 2256-2265, 2015. 1, 2 + +[27] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 2, 5, 8 +[28] Jou Won Song, Ye-In Park, Kyeongbo Kong, Jaeho Kwak, and Suk-Ju Kang. Selective transhdr: Transformer-based selective hdr imaging using ghost region mask. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVII, pages 288-304. Springer, 2022. 1, 2 +[29] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 1 +[30] Steven Tel, Zongwei Wu, Yulun Zhang, Barthélémy Heyrman, Cédric Demonceaux, Radu Timofte, and Dominique Gingac. Alignment-free hdr deghosting with semantics consistent transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12836-12845, 2023. 2 +[31] Anna Tomaszewska and Radoslaw Mantiuk. Image registration for multi-exposure high dynamic range image acquisition. In International Conference in Central Europe on Computer Graphics and Visualization, WSCG'07, 2007. 2 +[32] Okan Tarhan Tursun, Ahmet Oğuz Akyüz, Aykut Erdem, and Erkut Erdem. An objective dehosting quality metric for HDR images. Comput. Graph. Forum, 35(2):139-152, 2016. 5, 7 +[33] Okan Tarhan Tursun, Ahmet Oğuz Akyüz, Aykut Erdem, and Erkut Erdem. An objective deghosting quality metric for hdr images. In Computer Graphics Forum, pages 139-152. Wiley Online Library, 2016. 7 +[34] Haohan Wang, Xindi Wu, Zeyi Huang, and Eric P Xing. High-frequency component helps explain the generalization of convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8684-8694, 2020. 2 +[35] Yinhuai Wang, Jiwen Yu, and Jian Zhang. Zero-shot image restoration using denoising diffusion null-space model. In The Eleventh International Conference on Learning Representations, 2022. 2 +[36] Greg Ward. Fast, robust image registration for compositing high dynamic range photographs from hand-held exposures. Journal of Graphics Tools, 8, 2012. 2 +[37] Jay Whang, Maurizio Delbracio, Hossein Talebi, Chitwan Sahara, Alexandros G Dimakis, and Peyman Milanfar. Deblurring via stochastic refinement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16293-16303, 2022. 2 +[38] Shangzhe Wu, Jiarui Xu, Yu-Wing Tai, and Chi-Keung Tang. Deep high dynamic range imaging with large foreground motions. In European Conference on Computer Vision (ECCV), 2018. 6, 7 +[39] Bin Xia, Yulun Zhang, Shiyin Wang, Yitong Wang, Xinglong Wu, Yapeng Tian, Wenming Yang, and Luc Van Gool. Diffir: Efficient diffusion model for image restoration. In + +Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13095-13105, 2023. 2 +[40] Qingsen Yan, Jinqiu Sun, Haisen Li, Yu Zhu, and Yanning Zhang. High dynamic range imaging by sparse representation. Neurocomputing, 269:160-169, 2017. 2 +[41] Qingsen Yan, Dong Gong, Qinfeng Shi, Anton van den Hengel, Chunhua Shen, Ian Reid, and Yanning Zhang. Attention-guided network for ghost-free high dynamic range imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1751-1760, 2019. 1, 2, 4, 5, 6, 7, 8 +[42] Qingsen Yan, Lei Zhang, Yu Liu, Yu Zhu, Jinqiu Sun, Qinfeng Shi, and Yanning Zhang. Deep hdr imaging via a non-local network. IEEE Transactions on Image Processing, 29: 4308-4322, 2020. 2, 7, 8 +[43] Qingsen Yan, Weiye Chen, Song Zhang, Yu Zhu, Jinqiu Sun, and Yanning Zhang. A unified hdr imaging method with pixel and patch level. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22211-22220, 2023. 2, 6, 7 +[44] Qingsen Yan, Tao Hu, Yuan Sun, Hao Tang, Yu Zhu, Wei Dong, Luc Van Gool, and Yanning Zhang. Towards high-quality hdr deghosting with conditional diffusion models. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2023. 1, 2, 4, 5, 6, 7, 8 +[45] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728-5739, 2022. 4 +[46] Xiang Zhang, Tao Hu, Jiashuang He, and Qingsen Yan. Efficient content reconstruction for high dynamic range imaging. In ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7660-7664, 2024. 1, 2 +[47] Xiang Zhang, Qiang Zhu, Tao Hu, and Qingsen Yan. Eiffhdr: An efficient network for multi-exposure high dynamic range imaging. In ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6560-6564, 2024. 1, 2 \ No newline at end of file diff --git a/2024/Generating Content for HDR Deghosting from Frequency View/images.zip b/2024/Generating Content for HDR Deghosting from Frequency View/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..379c9dd1f9152e6b7a203dc113e16417cfe9562b --- /dev/null +++ b/2024/Generating Content for HDR Deghosting from Frequency View/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eb93089ec8fbf4331d3fc9616fc2d67abb74ba7e86cdbf3cfef1c89f788a2f73 +size 815013 diff --git a/2024/Generating Content for HDR Deghosting from Frequency View/layout.json b/2024/Generating Content for HDR Deghosting from Frequency View/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..33079e285cd1f184d852137703c088b51ec37f8d --- /dev/null +++ b/2024/Generating Content for HDR Deghosting from Frequency View/layout.json @@ -0,0 +1,12813 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 101, + 103, + 492, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 103, + 492, + 120 + ], + "spans": [ + { + "bbox": [ + 101, + 103, + 492, + 120 + ], + "type": "text", + "content": "Generating Content for HDR Deghosting from Frequency View" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 145, + 142, + 446, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 142, + 446, + 157 + ], + "spans": [ + { + "bbox": [ + 145, + 142, + 446, + 157 + ], + "type": "text", + "content": "Tao Hu" + }, + { + "bbox": [ + 145, + 142, + 446, + 157 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 145, + 142, + 446, + 157 + ], + "type": "text", + "content": " Qingsen Yan" + }, + { + "bbox": [ + 145, + 142, + 446, + 157 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 145, + 142, + 446, + 157 + ], + "type": "text", + "content": "* Yuankai Qi" + }, + { + "bbox": [ + 145, + 142, + 446, + 157 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 145, + 142, + 446, + 157 + ], + "type": "text", + "content": " Yanning Zhang" + }, + { + "bbox": [ + 145, + 142, + 446, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 144, + 157, + 449, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 157, + 449, + 172 + ], + "spans": [ + { + "bbox": [ + 144, + 157, + 449, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 144, + 157, + 449, + 172 + ], + "type": "text", + "content": "Northwestern Polytechnical University " + }, + { + "bbox": [ + 144, + 157, + 449, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 144, + 157, + 449, + 172 + ], + "type": "text", + "content": "Macquarie University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 90, + 173, + 498, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 173, + 498, + 185 + ], + "spans": [ + { + "bbox": [ + 90, + 173, + 498, + 185 + ], + "type": "text", + "content": "hutaoxauat@gmail.com, " + }, + { + "bbox": [ + 90, + 173, + 498, + 185 + ], + "type": "inline_equation", + "content": "\\mathbb{1}\\{" + }, + { + "bbox": [ + 90, + 173, + 498, + 185 + ], + "type": "text", + "content": " qingsenyan,ynzhang " + }, + { + "bbox": [ + 90, + 173, + 498, + 185 + ], + "type": "inline_equation", + "content": "\\} \\text{=} \\mathrm{nwpu.edu.cn}," + }, + { + "bbox": [ + 90, + 173, + 498, + 185 + ], + "type": "text", + "content": " yuankai.qi@mq.edu.au" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 213, + 192, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 224 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 224 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 238, + 289, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 289, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 289, + 548 + ], + "type": "text", + "content": "Recovering ghost-free High Dynamic Range (HDR) images from multiple Low Dynamic Range (LDR) images becomes challenging when the LDR images exhibit saturation and significant motion. Recent Diffusion Models (DMs) have been introduced in HDR imaging field, demonstrating promising performance, particularly in achieving visually perceptible results compared to previous DNN-based methods. However, DMs require extensive iterations with large models to estimate entire images, resulting in inefficiency that hinders their practical application. To address this challenge, we propose the Low-Frequency aware Diffusion (LF-Diff) model for ghost-free HDR imaging. The key idea of LF-Diff is implementing the DMs in a highly compacted latent space and integrating it into a regression-based model to enhance the details of reconstructed images. Specifically, as low-frequency information is closely related to human visual perception we propose to utilize DMs to create compact low-frequency priors for the reconstruction process. In addition, to take full advantage of the above low-frequency priors, the Dynamic HDR Reconstruction Network (DHRNet) is carried out in a regression-based manner to obtain final HDR images. Extensive experiments conducted on synthetic and real-world benchmark datasets demonstrate that our LF-Diff performs favorably against several state-of-the-art methods and is " + }, + { + "bbox": [ + 46, + 238, + 289, + 548 + ], + "type": "inline_equation", + "content": "10\\times" + }, + { + "bbox": [ + 46, + 238, + 289, + 548 + ], + "type": "text", + "content": " faster than previous DM-based methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 573, + 127, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 573, + 127, + 585 + ], + "spans": [ + { + "bbox": [ + 47, + 573, + 127, + 585 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 594, + 287, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 666 + ], + "type": "text", + "content": "Multiple exposure High Dynamic Range (HDR) imaging aims to restore missing details from exposure-varied Low Dynamic Range (LDR) images. However, in dynamic scenes, it often causes ghosting artifacts due to object or camera movement, limiting practical applications. Researchers are actively exploring ghost-free image recon" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 306, + 213, + 545, + 337 + ], + "blocks": [ + { + "bbox": [ + 306, + 213, + 545, + 337 + ], + "lines": [ + { + "bbox": [ + 306, + 213, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 306, + 213, + 545, + 337 + ], + "type": "image", + "image_path": "6195452a8473ddc281ee196038780c732efb56a605e8f53e9e275bb1b69c42a1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 344, + 546, + 377 + ], + "lines": [ + { + "bbox": [ + 305, + 344, + 546, + 377 + ], + "spans": [ + { + "bbox": [ + 305, + 344, + 546, + 377 + ], + "type": "text", + "content": "Figure 1. Average PSNR vs. inference time on the Kalantari's dataset [11]. Our method performs favorably and is " + }, + { + "bbox": [ + 305, + 344, + 546, + 377 + ], + "type": "inline_equation", + "content": "10\\times" + }, + { + "bbox": [ + 305, + 344, + 546, + 377 + ], + "type": "text", + "content": " faster than the previous diffusion-model-based method DiffHDR [44]." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 389, + 545, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 389, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 305, + 389, + 545, + 413 + ], + "type": "text", + "content": "struction methods to enable seamless capture of dynamic scenes with high dynamic ranges." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 415, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 546, + 715 + ], + "type": "text", + "content": "In recent years, the rise of Deep Neural Networks (DNNs) has brought significant advancements to HDR imaging field. Numerous DNN-based methods have emerged, employing Convolutional Neural Networks (CNNs) [11, 41, 47] or Vision Transformers (ViTs) [16, 28] for HDR image reconstruction. Despite their advancements, DNN-based approaches face challenges when essential information (e.g., content, details) for overexposed areas is missing due to object or camera movement. More recently, Diffusion Models (DMs) [8, 26] have exhibited impressive performance in image synthesis [4, 29] and image reconstruction tasks (including HDR imaging) [24, 44, 46]. DMs achieve this by iteratively denoising Gaussian noise to generate high-fidelity images. Yan et al. [44] reconstruct complete HDR images from pure Gaussian noise, yielding impressive results. Note that, compared to previous Deep generative models [14, 21], DMs generate a more accurate target distribution without facing issues like optimization instability or mode collapse, making them a promising choice for enhancing HDR image quality. Nonetheless, DMs demand a substantial number of iteration steps on denoising models to capture intricate data details, which is time-consuming even with a high-end GPU card. For example, as shown in Fig. 1, DiffHDR [44] takes about 7.5 seconds to generate an HDR image on a single A100 GPU." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 674, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 674, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 674, + 287, + 713 + ], + "type": "text", + "content": "*Corresponding author. † The first two authors contributed equally to this work. This work was partially supported by NSFC (62301432,62306240), NSBRPS (2023-JC-QN-0685, QCYRCXM-2023-057). Y. Qi is not supported by the above mentioned fundings." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25732" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "content": "In this paper, our goal is to devise a DM-based method that effectively harnesses the robust distribution mapping capabilities of DMs for HDR image reconstruction. We note that it is unnecessary to reconstruct a complete HDR image from pure noise, because the LDR reference images already provide most of the necessary content information for HDR images. As low-frequency information is closely related to human visual perception [34], we propose to utilize DMs to create compact low-frequency priors in the latent space. These priors are then integrated into the regression-based model to predict the low-frequency content of the reconstructed HDR images." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 216, + 289, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 216, + 289, + 443 + ], + "spans": [ + { + "bbox": [ + 46, + 216, + 289, + 443 + ], + "type": "text", + "content": "To achieve our goal, we propose the Low-Frequency Aware Diffusion Model (LF-Diff), comprising a Low-Frequency Prior Extraction Network (LPENet), a denoising network, and a Dynamic HDR Reconstruction Network (DHRNet). Following established methods [3, 23, 39], we employ a two-stage training strategy for pre-train LF-Diff and DM training. In the first stage, LPENet learns to extract a compact Low-frequency Prior Representation (LPR) from ground-truth images, guiding the DHRNet. DHRNet consists of two modules: a Prior Integration Module (PIM) that fuses the LPR with intermediate features of DHRNet, and a Feature Refinement Module (FRM) that further processes the fused features to HDR image. Notably, LPENet and DHRNet are optimized together in this phase. In the second stage, we train DM to learn the compact LPR directly from LDR images. Since the LPR is lightweight and only used to low-frequency content for HDR imaging, our DM can estimate the LPR with extremely low computational cost, ensuring stable visual results after multiple iterations." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 443, + 280, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 443, + 280, + 454 + ], + "spans": [ + { + "bbox": [ + 58, + 443, + 280, + 454 + ], + "type": "text", + "content": "The main contributions can be summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 455, + 287, + 634 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 455, + 287, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 455, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 455, + 287, + 514 + ], + "type": "text", + "content": "- We introduce LF-Diff, a straightforward and efficient DM-based method for HDR imaging. LF-Diff leverages the capabilities of diffusion models to generate informative low-frequency priors, which are then integrated into the imaging process to enhance results." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 515, + 287, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 515, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 515, + 287, + 563 + ], + "type": "text", + "content": "- We propose PIM and FRM in the DHRNet to fully exploit the LPR. While PIM efficiently fuses the LPR with intermediate features of DHRNet, FRM further processes the fused features to reconstruct high-quality HDR images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 563, + 287, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 287, + 634 + ], + "type": "text", + "content": "- Extensive experiments demonstrate that the proposed LF-Diff method achieves SOTA performance in HDR imaging tasks, and produces visually appealing results that align with human visual perception. But consuming significantly fewer computational resources compared to other DM-based methods." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 645, + 134, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 645, + 134, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 645, + 134, + 658 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "Ghost-free HDR Reconstruction. Traditional methods often employ motion rejection [7, 40], motion registration [31, 36], and patch matching [9, 25] approaches to align LDR images and reconstruct high-quality results. However," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 239 + ], + "type": "text", + "content": "these methods heavily rely on the performance of preprocessing techniques and often face challenges when dealing with motions across large spatial extents. Subsequently, DNN-based methods have emerged as the mainstream approach, thanks to their superior nonlinear expression capability. Researchers have explored various applications of DNNs in HDR imaging, devising sophisticated network architectures and model optimization schemes, such as attention [1, 41], transformer [16, 28, 30], optical flow [11, 22], GAN [6, 14, 21], and others [5, 15, 42, 43, 47]. Nevertheless, when LDR images lack sufficient information due to motion or saturation, they often manifest artifacts not aligned with human visual perception, commonly referred to as ghosting." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 239, + 546, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 239, + 546, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 239, + 546, + 479 + ], + "type": "text", + "content": "Diffusion Models. Diffusion models have recently demonstrated promising results in various low-level vision tasks, including super-resolution [13, 24], HDR reconstruction [44, 46], colorization [12, 35], and deblurring [37]. DiffHDR [44] applied the conventional diffusion model paradigm to reconstruct HDR images, yielding remarkable outcomes. However, due to the high computational cost of diffusion models along with the requirement for numerous iterative steps, inference speed is limited. Despite efforts to mitigate this constraint [2, 17, 27], the overall complexity remains high, especially for high-resolution images commonly encountered in HDR scenarios. Recently, several methods [3, 23, 39] have addressed this issue by conducting diffusion modeling in the latent space. Rombach et al. [23] employed an autoencoder model to compress images into features equivalent to the image space. HI-Diff [3] leverages diffusion-based vector features for assisting image deblurring. However, given the unique characteristics of HDR imaging, further exploration is needed to select an appropriate latent space for HDR reconstruction tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 489, + 391, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 489, + 391, + 502 + ], + "spans": [ + { + "bbox": [ + 306, + 489, + 391, + 502 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 509, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 545, + 568 + ], + "type": "text", + "content": "The diffusion model introduced by Sohl-Dickstein et al. [26] is inspired by nonequilibrium thermodynamics. Here, we provide a brief overview of the \"variance-preserving\" diffusion model from [8], which encompasses diffusion processes and reverse processes." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 569, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 569, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 545, + 604 + ], + "type": "text", + "content": "Diffusion Process. Given a clean image distribution, the diffusion process gradually injects isotropic Gaussian noise to generate " + }, + { + "bbox": [ + 304, + 569, + 545, + 604 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 304, + 569, + 545, + 604 + ], + "type": "text", + "content": " according to a variance schedule" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 605, + 525, + 618 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 605, + 525, + 618 + ], + "spans": [ + { + "bbox": [ + 306, + 605, + 525, + 618 + ], + "type": "interline_equation", + "content": "\\beta_ {1}, \\dots , \\beta_ {T} \\in (0, 1). \\text {L e t} \\alpha_ {t} = 1 - \\beta_ {t}, \\bar {\\alpha} _ {t} = \\prod_ {i = 1} ^ {T} \\alpha_ {i}:", + "image_path": "286f53246659f5e3a7694957ade51e41e8cc3efbd8b7513aa59274dc1d24ecf0.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 347, + 625, + 545, + 639 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 625, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 347, + 625, + 545, + 639 + ], + "type": "interline_equation", + "content": "q \\left(x _ {t} \\mid x _ {0}\\right) = \\mathcal {N} \\left(x _ {t}; \\sqrt {\\bar {\\alpha} _ {t}} x _ {0}, (1 - \\bar {\\alpha} _ {t}) \\mathbf {I}\\right). \\tag {1}", + "image_path": "338e3b4ffe84a7fedf366685415699c71d3aa064cbe2099c3a5e3a548db4cb4d.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 647, + 529, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 529, + 659 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 529, + 659 + ], + "type": "text", + "content": "The closed-form for this process can be represented as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 333, + 667, + 545, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 667, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 333, + 667, + 545, + 681 + ], + "type": "interline_equation", + "content": "x _ {t} \\left(x _ {0}, \\epsilon_ {t}\\right) = \\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t}, \\epsilon \\in (0, \\mathbf {I}), \\tag {2}", + "image_path": "f8733c6fa4a868bbda5700c8a3b4476604cbb0aec2fa0020ac4227fca2f3a6ca.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": " represents the output at time " + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": " is the initial clean image, and " + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon_t" + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": " denotes isotropic Gaussian noise." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25733" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 319, + 257 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 319, + 257 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 319, + 257 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 319, + 257 + ], + "type": "image", + "image_path": "f5af3edc5e26d8545e8d3d2375df16679a33e8746d850afd5e159841d0002b6c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 323, + 72, + 544, + 257 + ], + "blocks": [ + { + "bbox": [ + 323, + 72, + 544, + 257 + ], + "lines": [ + { + "bbox": [ + 323, + 72, + 544, + 257 + ], + "spans": [ + { + "bbox": [ + 323, + 72, + 544, + 257 + ], + "type": "image", + "image_path": "9d3c16e98df4a6cc282eca153975e7f37b222b385af8565e2503f392c9b0e05d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 49, + 259, + 547, + 279 + ], + "blocks": [ + { + "bbox": [ + 49, + 259, + 547, + 279 + ], + "lines": [ + { + "bbox": [ + 49, + 259, + 547, + 279 + ], + "spans": [ + { + "bbox": [ + 49, + 259, + 547, + 279 + ], + "type": "image", + "image_path": "763c7d3917bd418d786c515129ce3dd3e733dcc7876e9ae5c25797e8bca0da87.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 288, + 546, + 322 + ], + "lines": [ + { + "bbox": [ + 46, + 288, + 546, + 322 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 546, + 322 + ], + "type": "text", + "content": "Figure 2. The proposed LF-Diff comprises DHRNet, LPRNet, and a denoising network. LF-Diff undergoes two training stages: Pretrain LF-Diff (Sec. 4.1) and DM training (Sec. 4.2). Notably, during the inference stage, we do not input the ground-truth image into " + }, + { + "bbox": [ + 46, + 288, + 546, + 322 + ], + "type": "inline_equation", + "content": "\\mathrm{LPONet}_{DM}" + }, + { + "bbox": [ + 46, + 288, + 546, + 322 + ], + "type": "text", + "content": " and the denoising networks. Instead, we solely utilize the reverse process of DMs." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 343, + 287, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 343, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 47, + 343, + 287, + 380 + ], + "type": "text", + "content": "Reverse Process. The reverse process attempts to approximate the data distribution " + }, + { + "bbox": [ + 47, + 343, + 287, + 380 + ], + "type": "inline_equation", + "content": "q(x_0)" + }, + { + "bbox": [ + 47, + 343, + 287, + 380 + ], + "type": "text", + "content": " by a Markovian process, starting with random Gaussian noise " + }, + { + "bbox": [ + 47, + 343, + 287, + 380 + ], + "type": "inline_equation", + "content": "x_{T} = \\mathcal{N}(x_{T};\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 47, + 343, + 287, + 380 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 77, + 388, + 287, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 388, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 77, + 388, + 287, + 403 + ], + "type": "interline_equation", + "content": "q \\left(x _ {t - 1} \\mid x _ {t}, x _ {0}\\right) = \\mathcal {N} \\left(x _ {t - 1}; \\tilde {\\mu} _ {t} \\left(x _ {t}, x _ {0}\\right), \\tilde {\\beta} _ {t} \\mathbf {I}\\right), \\tag {3}", + "image_path": "8ae099c9157f69087ed54457115743f7e1be8450083a86187166d66132660ccd.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 412, + 287, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 412, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 47, + 412, + 287, + 462 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 412, + 287, + 462 + ], + "type": "inline_equation", + "content": "\\tilde{\\mu}_t" + }, + { + "bbox": [ + 47, + 412, + 287, + 462 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 412, + 287, + 462 + ], + "type": "inline_equation", + "content": "\\tilde{\\beta}_t" + }, + { + "bbox": [ + 47, + 412, + 287, + 462 + ], + "type": "text", + "content": " are distribution parameters. Since the true reverse process Eq.(3) relies on " + }, + { + "bbox": [ + 47, + 412, + 287, + 462 + ], + "type": "inline_equation", + "content": "q(x_0)" + }, + { + "bbox": [ + 47, + 412, + 287, + 462 + ], + "type": "text", + "content": " and is intractable, the neural network " + }, + { + "bbox": [ + 47, + 412, + 287, + 462 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 47, + 412, + 287, + 462 + ], + "type": "text", + "content": " serves as the denoiser to estimate " + }, + { + "bbox": [ + 47, + 412, + 287, + 462 + ], + "type": "inline_equation", + "content": "p_{\\theta}(x_{t - 1}|x_t,t)" + }, + { + "bbox": [ + 47, + 412, + 287, + 462 + ], + "type": "text", + "content": " instead of " + }, + { + "bbox": [ + 47, + 412, + 287, + 462 + ], + "type": "inline_equation", + "content": "q(x_{t - 1}|x_t,x_0)" + }, + { + "bbox": [ + 47, + 412, + 287, + 462 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 70, + 469, + 287, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 469, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 70, + 469, + 287, + 498 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} p _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}, t\\right) = \\mathcal {N} \\left(x _ {t - 1}; \\mu_ {\\theta} \\left(x _ {t}, t\\right), \\Sigma_ {\\theta} \\left(x _ {t}, t\\right)\\right) \\tag {4} \\\\ = q \\left(x _ {t - 1} \\mid f _ {\\theta} (x _ {t}, t)\\right), \\\\ \\end{array}", + "image_path": "45ba35217d66bc9607cf38dda5ea64d6e6fb28ae727adc145c998f189c539ba1.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 506, + 287, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 506, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 47, + 506, + 287, + 542 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 506, + 287, + 542 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}" + }, + { + "bbox": [ + 47, + 506, + 287, + 542 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 506, + 287, + 542 + ], + "type": "inline_equation", + "content": "\\Sigma_{\\theta}" + }, + { + "bbox": [ + 47, + 506, + 287, + 542 + ], + "type": "text", + "content": " are the parameters to be estimated in the reverse process. Let fixed variances " + }, + { + "bbox": [ + 47, + 506, + 287, + 542 + ], + "type": "inline_equation", + "content": "(\\Sigma_{\\theta}(x_t,t) = \\tilde{\\beta}_t\\mathbf{I})" + }, + { + "bbox": [ + 47, + 506, + 287, + 542 + ], + "type": "text", + "content": " , the optimization objective can be defined as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 74, + 550, + 287, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 550, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 74, + 550, + 287, + 564 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {t, x _ {0}, \\epsilon_ {t}} [ \\| \\epsilon_ {t} - f _ {\\theta} (\\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t}, t) \\| ^ {2} ]. \\tag {5}", + "image_path": "19e3bade453243d3e686d0baa72cb10cd9682dedf0bcfd5658e5f458b6a89186.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 574, + 129, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 129, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 129, + 588 + ], + "type": "text", + "content": "4. Methodology" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 594, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 715 + ], + "type": "text", + "content": "As depicted in Fig. 2, we introduce the proposed LF-Diff framework, which consists of two training stages. During the first stage, we pre-train LF-Diff to extract accurate low-frequency prior representations from ground truths using LPenet, while DHRNet learns how to utilize the above priors to reconstruct HDR images. During the second stage, we train the DM to accurately predict the low-frequency prior representations directly from LDR images, and jointly optimize it with DHRNet to generate final high-quality HDR images." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 342, + 463, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 342, + 463, + 355 + ], + "spans": [ + { + "bbox": [ + 306, + 342, + 463, + 355 + ], + "type": "text", + "content": "4.1. Stage One: Pretrain LF-Diff" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 361, + 545, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 545, + 504 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 545, + 504 + ], + "type": "text", + "content": "In the first stage, our objective is to train the LPERNet how to extract accurate low-frequency prior features from ground truth and embed this representation into the DHRNet to guide the reconstruction process. As illustrated in Fig. 2 (a), we utilize the Prior Integration Module (PIM) to incorporate these prior features into the DHRNet via the cross-attention mechanism. Within DHRNet, multiple Feature Refinement Modules (FRM) are stacked to directly learn the mapping relationship from input features to the target image distribution, while PIM leverages prior features to assist in recovering the low-frequency content of the reconstructed images. Next, we elaborate on above descriptions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 505, + 546, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 505, + 546, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 505, + 546, + 635 + ], + "type": "text", + "content": "Low-frequency Prior Extraction Network. The LPENet has a straightforward structure, as shown in Fig. 2 (a), comprising several residual blocks designed to extract the Low-frequency Prior Representation (LPR). Starting with the ground truth image " + }, + { + "bbox": [ + 304, + 505, + 546, + 635 + ], + "type": "inline_equation", + "content": "I_{gt} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 304, + 505, + 546, + 635 + ], + "type": "text", + "content": ", we initially perform a tonemapping operation " + }, + { + "bbox": [ + 304, + 505, + 546, + 635 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(\\cdot)" + }, + { + "bbox": [ + 304, + 505, + 546, + 635 + ], + "type": "text", + "content": " to obtain the LDR domain ground truth image " + }, + { + "bbox": [ + 304, + 505, + 546, + 635 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(I_{gt})" + }, + { + "bbox": [ + 304, + 505, + 546, + 635 + ], + "type": "text", + "content": ". We concatenate these images along the channel dimension and employ the PixelUnshuffle operation to downsample, generating the input for LPENet. Subsequently, LPENet extracts the LPR " + }, + { + "bbox": [ + 304, + 505, + 546, + 635 + ], + "type": "inline_equation", + "content": "z \\in \\mathbb{R}^{\\frac{H}{k} \\times \\frac{W}{k} \\times 3}" + }, + { + "bbox": [ + 304, + 505, + 546, + 635 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 643, + 545, + 656 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 643, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 315, + 643, + 545, + 656 + ], + "type": "interline_equation", + "content": "z = \\operatorname {L P E N e t} (\\operatorname {P i x i l e U n s h u r f l e} (\\operatorname {C o n c a t} (I _ {g t}, \\mathcal {T} (I _ {g t})))) \\tag {6}", + "image_path": "6e839d8cba619d908f6e8719f31ad52d4c4120576b19e11bfa979c26e63a88a3.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 662, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 662, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 662, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 662, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(x) = \\frac{\\log(1 + \\mu x)}{\\log(1 + \\mu)}" + }, + { + "bbox": [ + 304, + 662, + 545, + 713 + ], + "type": "text", + "content": " denotes the tonemapping operator, " + }, + { + "bbox": [ + 304, + 662, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mu = 5000" + }, + { + "bbox": [ + 304, + 662, + 545, + 713 + ], + "type": "text", + "content": " is a predefined constant parameter that controls the degree of compression applied to the input signal, " + }, + { + "bbox": [ + 304, + 662, + 545, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 662, + 545, + 713 + ], + "type": "text", + "content": " denotes the sampling multiple of PixelUnshuffle." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25734" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": "Dynamic HDR Reconstruction Network. As illustrated in Fig. 2 (a), DHRNet consists of multiple stacked Reconstruction Blocks, with each block comprising one Prior Integration Module (PIM) and " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": " Feature Refinement Module (FRM). PIM fuses the LPR with intermediate features of DHRNet, we will give a detailed description in the following part. FRM is designed based on the understanding that natural images encompass diverse frequencies, since high and low frequencies serve distinct roles in image encoding." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 193, + 289, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 193, + 289, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 193, + 289, + 277 + ], + "type": "text", + "content": "In Fig. 3 (a), the structure of FRM involves several blocks. Specifically, we begin by processing the features through a residual block to obtain " + }, + { + "bbox": [ + 46, + 193, + 289, + 277 + ], + "type": "inline_equation", + "content": "F_{n-1}" + }, + { + "bbox": [ + 46, + 193, + 289, + 277 + ], + "type": "text", + "content": ". Subsequently, we decouple these features into two components: high-frequency features " + }, + { + "bbox": [ + 46, + 193, + 289, + 277 + ], + "type": "inline_equation", + "content": "F_{n-1}^{\\mathrm{high}}" + }, + { + "bbox": [ + 46, + 193, + 289, + 277 + ], + "type": "text", + "content": ", which remain resolution-invariant, and low-frequency features " + }, + { + "bbox": [ + 46, + 193, + 289, + 277 + ], + "type": "inline_equation", + "content": "F_{n-1}^{\\mathrm{low}}" + }, + { + "bbox": [ + 46, + 193, + 289, + 277 + ], + "type": "text", + "content": " with reduced resolution, following the approach described in [18]:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 88, + 287, + 211, + 300 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 287, + 211, + 300 + ], + "spans": [ + { + "bbox": [ + 88, + 287, + 211, + 300 + ], + "type": "interline_equation", + "content": "F _ {n - 1} ^ {\\text {l o w}} = \\operatorname {A v g p o o l} \\left(F _ {n - 1}, k\\right),", + "image_path": "13677d8551b8a2efffef35d438e7952e0d22a528f7f58bf066f27b43d79f0c6c.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 88, + 299, + 287, + 319 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 299, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 88, + 299, + 287, + 319 + ], + "type": "interline_equation", + "content": "F _ {n - 1} ^ {\\text {h i g h}} = F _ {n - 1} - \\text {U p s a m p l e} \\left(F _ {n - 1} ^ {\\text {l o w}}\\right), \\tag {7}", + "image_path": "6bc4a7ea5af880959af9bffe3f5594cb053e955441eb03bdab11f16468554a71.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 328, + 289, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 328, + 289, + 485 + ], + "spans": [ + { + "bbox": [ + 46, + 328, + 289, + 485 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 328, + 289, + 485 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 328, + 289, + 485 + ], + "type": "text", + "content": " denotes the kernel size of the avgpooling layer, Upsample denotes the upsampling operator. In essence, low-frequency features capture global dependencies within the input (image/features) and do not necessitate a high-resolution feature map, but require global attention. On the other hand, high-frequency features capture detailed local information, requiring a high-resolution feature map, and only can be achieved through local operators. Therefore, we employ the residual block (RB) to process " + }, + { + "bbox": [ + 46, + 328, + 289, + 485 + ], + "type": "inline_equation", + "content": "F_{n - 1}^{\\mathrm{high}}" + }, + { + "bbox": [ + 46, + 328, + 289, + 485 + ], + "type": "text", + "content": " and utilize the self-attention (SA) from [45] to handle " + }, + { + "bbox": [ + 46, + 328, + 289, + 485 + ], + "type": "inline_equation", + "content": "F_{n - 1}^{\\mathrm{low}}" + }, + { + "bbox": [ + 46, + 328, + 289, + 485 + ], + "type": "text", + "content": ". Subsequently, we fuse the low- and high-frequency features to preserve the initial details, resulting in the feature " + }, + { + "bbox": [ + 46, + 328, + 289, + 485 + ], + "type": "inline_equation", + "content": "F_{n - 1}^{\\prime}" + }, + { + "bbox": [ + 46, + 328, + 289, + 485 + ], + "type": "text", + "content": ". This operation can be represented as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 51, + 493, + 287, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 493, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 51, + 493, + 287, + 520 + ], + "type": "interline_equation", + "content": "F _ {n - 1} ^ {\\prime} = \\operatorname {C o n c a t} \\left[ R B \\left(F _ {n - 1} ^ {\\text {h i g h}}\\right), \\operatorname {U p s a m p l e} \\left(S A \\left(F _ {n - 1} ^ {\\text {l o w}}\\right)\\right) \\right]. \\tag {8}", + "image_path": "33a5e15fd63307e76918ee0bfc408e38de784b32a7c96a8f1b214834284e4da2.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 521, + 287, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 582 + ], + "type": "text", + "content": "Considering that " + }, + { + "bbox": [ + 46, + 521, + 287, + 582 + ], + "type": "inline_equation", + "content": "F_{n-1}^{\\prime}" + }, + { + "bbox": [ + 46, + 521, + 287, + 582 + ], + "type": "text", + "content": " is a concatenation of two features, a " + }, + { + "bbox": [ + 46, + 521, + 287, + 582 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 46, + 521, + 287, + 582 + ], + "type": "text", + "content": " Conv is employed to effectuate a reduction in the number of channels. This downsizing step is followed by a channel attention module, designed to accentuate those channels exhibiting heightened activation values." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": "In order to effectively merge the prior feature derived from the LPERNet with the intermediate feature produced by DHRNet, we replaced the self-attention mechanism in the low-frequency branch of FRM with a carefully designed cross-attention mechanism, resulting in the Prior Integration Module (PIM). As depicted in Fig. 2 (a), a PIM is positioned before each reconstruction block. Within each PIM, cross-attention computations are performed between the prior and intermediate features to facilitate feature fusion. This module enables the aggregation of information from the prior feature into the features of DHRNet." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 309, + 71, + 545, + 164 + ], + "blocks": [ + { + "bbox": [ + 309, + 71, + 545, + 164 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 545, + 164 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 545, + 164 + ], + "type": "image", + "image_path": "126d2748d194f331ae8b4872766c520c57b2da0b4a8fee7f036be1a223e945bd.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 363, + 164, + 485, + 174 + ], + "lines": [ + { + "bbox": [ + 363, + 164, + 485, + 174 + ], + "spans": [ + { + "bbox": [ + 363, + 164, + 485, + 174 + ], + "type": "text", + "content": "(a) Feature Refinement Module" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 175, + 535, + 277 + ], + "blocks": [ + { + "bbox": [ + 310, + 175, + 535, + 277 + ], + "lines": [ + { + "bbox": [ + 310, + 175, + 535, + 277 + ], + "spans": [ + { + "bbox": [ + 310, + 175, + 535, + 277 + ], + "type": "image", + "image_path": "6dcd4f8b73cc7dc8103466d130b2ae4c17a8181ea480aa1e990d7930357d194e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 350, + 277, + 507, + 286 + ], + "lines": [ + { + "bbox": [ + 350, + 277, + 507, + 286 + ], + "spans": [ + { + "bbox": [ + 350, + 277, + 507, + 286 + ], + "type": "text", + "content": "(b)Cross Self-attention module from PIM" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 297, + 546, + 319 + ], + "lines": [ + { + "bbox": [ + 305, + 297, + 546, + 319 + ], + "spans": [ + { + "bbox": [ + 305, + 297, + 546, + 319 + ], + "type": "text", + "content": "Figure 3. The architecture of (a) Feature Refinement Module (FRM) and (b) Cross-attention mechanism." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 327, + 547, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 327, + 547, + 491 + ], + "spans": [ + { + "bbox": [ + 305, + 327, + 547, + 491 + ], + "type": "text", + "content": "Specifically, as illustrated in Fig. 3 (b), our cross-attention module receives two types of input: the intermediate feature " + }, + { + "bbox": [ + 305, + 327, + 547, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{F} \\in \\mathbb{R}^{\\hat{H} \\times \\hat{W} \\times \\hat{C}}" + }, + { + "bbox": [ + 305, + 327, + 547, + 491 + ], + "type": "text", + "content": " and the prior feature " + }, + { + "bbox": [ + 305, + 327, + 547, + 491 + ], + "type": "inline_equation", + "content": "z \\in \\mathbb{R}^{\\hat{H} \\times \\hat{W} \\times \\hat{N}}" + }, + { + "bbox": [ + 305, + 327, + 547, + 491 + ], + "type": "text", + "content": ". We project " + }, + { + "bbox": [ + 305, + 327, + 547, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 305, + 327, + 547, + 491 + ], + "type": "text", + "content": " into a query vector " + }, + { + "bbox": [ + 305, + 327, + 547, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{Q} = W_d^Q W_c^Q\\mathbf{F}" + }, + { + "bbox": [ + 305, + 327, + 547, + 491 + ], + "type": "text", + "content": " using point-wise " + }, + { + "bbox": [ + 305, + 327, + 547, + 491 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 305, + 327, + 547, + 491 + ], + "type": "text", + "content": " convolution and depth-wise " + }, + { + "bbox": [ + 305, + 327, + 547, + 491 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 305, + 327, + 547, + 491 + ], + "type": "text", + "content": " convolution with weights " + }, + { + "bbox": [ + 305, + 327, + 547, + 491 + ], + "type": "inline_equation", + "content": "W_d^Q" + }, + { + "bbox": [ + 305, + 327, + 547, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 327, + 547, + 491 + ], + "type": "inline_equation", + "content": "W_c^Q" + }, + { + "bbox": [ + 305, + 327, + 547, + 491 + ], + "type": "text", + "content": ". Similarly, " + }, + { + "bbox": [ + 305, + 327, + 547, + 491 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 305, + 327, + 547, + 491 + ], + "type": "text", + "content": " is transformed into key " + }, + { + "bbox": [ + 305, + 327, + 547, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 305, + 327, + 547, + 491 + ], + "type": "text", + "content": " and value " + }, + { + "bbox": [ + 305, + 327, + 547, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 305, + 327, + 547, + 491 + ], + "type": "text", + "content": " through analogous operations. Next, we reshape these projections into matrices suitable for attention computation: " + }, + { + "bbox": [ + 305, + 327, + 547, + 491 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Q}} \\in \\mathbb{R}^{\\hat{H}\\hat{W} \\times \\hat{C}}" + }, + { + "bbox": [ + 305, + 327, + 547, + 491 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 327, + 547, + 491 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{K}} \\in \\mathbb{R}^{\\hat{N} \\times \\hat{H}\\hat{W}}" + }, + { + "bbox": [ + 305, + 327, + 547, + 491 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 327, + 547, + 491 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{V}} \\in \\mathbb{R}^{\\hat{H}\\hat{W} \\times \\hat{N}}" + }, + { + "bbox": [ + 305, + 327, + 547, + 491 + ], + "type": "text", + "content": ". We then calculate a more computationally efficient attention map " + }, + { + "bbox": [ + 305, + 327, + 547, + 491 + ], + "type": "inline_equation", + "content": "A \\in \\mathbb{R}^{\\hat{N} \\times \\hat{C}}" + }, + { + "bbox": [ + 305, + 327, + 547, + 491 + ], + "type": "text", + "content": " by performing dot product between " + }, + { + "bbox": [ + 305, + 327, + 547, + 491 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Q}}" + }, + { + "bbox": [ + 305, + 327, + 547, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 327, + 547, + 491 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{K}}" + }, + { + "bbox": [ + 305, + 327, + 547, + 491 + ], + "type": "text", + "content": ". The process can be described as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 348, + 496, + 545, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 496, + 545, + 512 + ], + "spans": [ + { + "bbox": [ + 348, + 496, + 545, + 512 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {F}} = W _ {c} \\hat {\\mathbf {V}} \\cdot \\operatorname {S o f t m a x} (\\hat {\\mathbf {K}} \\cdot \\hat {\\mathbf {Q}} / \\gamma) + \\mathbf {F}, \\tag {9}", + "image_path": "b24fbd5eb3bbe3f172d47e7a6dc6797a7b41458dd63b5d448d466e5919d976e0.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 518, + 545, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 518, + 545, + 541 + ], + "spans": [ + { + "bbox": [ + 305, + 518, + 545, + 541 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 518, + 545, + 541 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 305, + 518, + 545, + 541 + ], + "type": "text", + "content": " is a trainable parameter. Notably, we do not implement multi-head attention in PIM." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 542, + 545, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 542, + 545, + 674 + ], + "spans": [ + { + "bbox": [ + 305, + 542, + 545, + 674 + ], + "type": "text", + "content": "Training Strategy. As proposed in [11], we initially transform the provided set of input LDR images " + }, + { + "bbox": [ + 305, + 542, + 545, + 674 + ], + "type": "inline_equation", + "content": "\\{L_1, L_2, \\ldots, L_N\\}" + }, + { + "bbox": [ + 305, + 542, + 545, + 674 + ], + "type": "text", + "content": " into their corresponding HDR versions " + }, + { + "bbox": [ + 305, + 542, + 545, + 674 + ], + "type": "inline_equation", + "content": "H_i" + }, + { + "bbox": [ + 305, + 542, + 545, + 674 + ], + "type": "text", + "content": " using gamma correction. Subsequently, we concatenate each " + }, + { + "bbox": [ + 305, + 542, + 545, + 674 + ], + "type": "inline_equation", + "content": "L_i" + }, + { + "bbox": [ + 305, + 542, + 545, + 674 + ], + "type": "text", + "content": " with its corresponding " + }, + { + "bbox": [ + 305, + 542, + 545, + 674 + ], + "type": "inline_equation", + "content": "H_i" + }, + { + "bbox": [ + 305, + 542, + 545, + 674 + ], + "type": "text", + "content": " along the channel axis to generate six-channel input tensors " + }, + { + "bbox": [ + 305, + 542, + 545, + 674 + ], + "type": "inline_equation", + "content": "X_i = [L_i, H_i]" + }, + { + "bbox": [ + 305, + 542, + 545, + 674 + ], + "type": "text", + "content": ". Similar to [16, 41, 44], we employ an alignment module [41] to process the input LDR images, align the features implicitly, and feed them into DHRNet. Then, the LPR feature " + }, + { + "bbox": [ + 305, + 542, + 545, + 674 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 305, + 542, + 545, + 674 + ], + "type": "text", + "content": " extracted from LPENet is injected into DHRNet via PIM. The reconstructed HDR image " + }, + { + "bbox": [ + 305, + 542, + 545, + 674 + ], + "type": "inline_equation", + "content": "\\hat{H}" + }, + { + "bbox": [ + 305, + 542, + 545, + 674 + ], + "type": "text", + "content": " is generated by:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 353, + 680, + 545, + 695 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 680, + 545, + 695 + ], + "spans": [ + { + "bbox": [ + 353, + 680, + 545, + 695 + ], + "type": "interline_equation", + "content": "\\hat {H} = \\operatorname {C o n v} _ {3 \\times 3} \\left(\\mathrm {D H R N e t} \\left(X _ {i}, z\\right)\\right). \\tag {10}", + "image_path": "155498bf23aa501f15242bf60b7cebebfb9bbd851a2bf9c6067b36fd4850a36d.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "content": "Following previous approaches [16, 41], we utilize both" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "25735" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "tonemapped per-pixel loss and perceptual loss as the image reconstruction loss function " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_r" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": ". This dual-loss strategy optimizes both pixel-level accuracy and high-level feature representations in the generated HDR results:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 64, + 129, + 287, + 140 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 129, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 64, + 129, + 287, + 140 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r} = \\left\\| \\mathcal {T} (H) - \\mathcal {T} (\\hat {H}) \\right\\| _ {1} + \\lambda \\| \\phi_ {i, j} (\\mathcal {T} (H)) - \\phi_ {i, j} (\\mathcal {T} (\\hat {H})) \\| _ {1}, \\tag {11}", + "image_path": "16714dd0dbe058ac607572268e070acd69ab0eb8aa4b40939f3d4332133feac0.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 148, + 287, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 148, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 46, + 148, + 287, + 196 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 148, + 287, + 196 + ], + "type": "inline_equation", + "content": "\\phi_{i,j}(\\cdot)" + }, + { + "bbox": [ + 46, + 148, + 287, + 196 + ], + "type": "text", + "content": " signifies the " + }, + { + "bbox": [ + 46, + 148, + 287, + 196 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 148, + 287, + 196 + ], + "type": "text", + "content": "-th convolutional feature extracted from the VGG19 network after the " + }, + { + "bbox": [ + 46, + 148, + 287, + 196 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 148, + 287, + 196 + ], + "type": "text", + "content": "-th maxpooling operation, and " + }, + { + "bbox": [ + 46, + 148, + 287, + 196 + ], + "type": "inline_equation", + "content": "\\lambda = 1e - 2" + }, + { + "bbox": [ + 46, + 148, + 287, + 196 + ], + "type": "text", + "content": " is a hyperparameter that balances the contribution of the each component." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 202, + 287, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 202, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 47, + 202, + 287, + 216 + ], + "type": "text", + "content": "4.2. Stage Two: Diffusion Models for HDR Imaging" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 220, + 287, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 220, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 287, + 316 + ], + "type": "text", + "content": "After the above-mentioned learning procedure, we now possess LDR images and their corresponding LPR " + }, + { + "bbox": [ + 46, + 220, + 287, + 316 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 220, + 287, + 316 + ], + "type": "text", + "content": ". In the second stage (Fig. 2 (b, c)), our objective is to efficiently harness the powerful distribution estimation capability of the DMs. Specifically, we utilize the LPENet from stage one to capture the LPR as the denoising target of the DM. The DM will learn how to extract accurate LPR from the LDR inputs and perform joint optimization with DHRNet." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 316, + 287, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 287, + 364 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 287, + 364 + ], + "type": "text", + "content": "Diffusion Model for LRP Learning. After capturing the LPR " + }, + { + "bbox": [ + 46, + 316, + 287, + 364 + ], + "type": "inline_equation", + "content": "z \\in \\mathbb{R}^{\\frac{H}{k} \\times \\frac{W}{k} \\times N}" + }, + { + "bbox": [ + 46, + 316, + 287, + 364 + ], + "type": "text", + "content": " from the pretrained LPENet, we transform the clean LPR feature " + }, + { + "bbox": [ + 46, + 316, + 287, + 364 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 316, + 287, + 364 + ], + "type": "text", + "content": " into a noisy version " + }, + { + "bbox": [ + 46, + 316, + 287, + 364 + ], + "type": "inline_equation", + "content": "z_T" + }, + { + "bbox": [ + 46, + 316, + 287, + 364 + ], + "type": "text", + "content": " using Eq.(2), which can be described as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 82, + 371, + 287, + 385 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 371, + 287, + 385 + ], + "spans": [ + { + "bbox": [ + 82, + 371, + 287, + 385 + ], + "type": "interline_equation", + "content": "q \\left(z _ {T} \\mid z\\right) = \\mathcal {N} \\left(z _ {T}; \\sqrt {\\bar {\\alpha} _ {T}} z, (1 - \\bar {\\alpha} _ {T}) \\mathbf {I}\\right), \\tag {12}", + "image_path": "77611c5ddcfc6049ad45a8368639b9695e9793eca8bf5c93d22ba7b60271cfe7.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 392, + 286, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 392, + 286, + 415 + ], + "spans": [ + { + "bbox": [ + 46, + 392, + 286, + 415 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 392, + 286, + 415 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 392, + 286, + 415 + ], + "type": "text", + "content": " represents the total number of iterations, while " + }, + { + "bbox": [ + 46, + 392, + 286, + 415 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t" + }, + { + "bbox": [ + 46, + 392, + 286, + 415 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 392, + 286, + 415 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 392, + 286, + 415 + ], + "type": "text", + "content": " denote the pre-defined variance schedule." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 415, + 287, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 287, + 511 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 287, + 511 + ], + "type": "text", + "content": "In the reverse process, we iteratively generate the LPR from a pure Gaussian distribution, starting from " + }, + { + "bbox": [ + 46, + 415, + 287, + 511 + ], + "type": "inline_equation", + "content": "z_{T}" + }, + { + "bbox": [ + 46, + 415, + 287, + 511 + ], + "type": "text", + "content": " and moving backward to " + }, + { + "bbox": [ + 46, + 415, + 287, + 511 + ], + "type": "inline_equation", + "content": "z_{0}" + }, + { + "bbox": [ + 46, + 415, + 287, + 511 + ], + "type": "text", + "content": ", which utilizes the posterior distribution described in Eq. (3). Following previous work [24, 44], we utilize a neural network to estimate " + }, + { + "bbox": [ + 46, + 415, + 287, + 511 + ], + "type": "inline_equation", + "content": "p_{\\theta}(x_{t - 1}|x_t,t)" + }, + { + "bbox": [ + 46, + 415, + 287, + 511 + ], + "type": "text", + "content": " instead of " + }, + { + "bbox": [ + 46, + 415, + 287, + 511 + ], + "type": "inline_equation", + "content": "q(x_{t - 1}|x_t,x_0)" + }, + { + "bbox": [ + 46, + 415, + 287, + 511 + ], + "type": "text", + "content": " for each step. Specifically, we first use LPENetDM to obtain a conditional feature " + }, + { + "bbox": [ + 46, + 415, + 287, + 511 + ], + "type": "inline_equation", + "content": "D\\in \\mathbb{R}^{\\frac{H}{k}\\times \\frac{W}{k}\\times N}" + }, + { + "bbox": [ + 46, + 415, + 287, + 511 + ], + "type": "text", + "content": " from aligned LDR feature:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 59, + 518, + 287, + 532 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 518, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 59, + 518, + 287, + 532 + ], + "type": "interline_equation", + "content": "\\mathbf {D} = \\operatorname {L P E N e t} _ {D M} (A M (\\text {P i x e l U n s h u f f l e} (X _ {i}))). \\tag {13}", + "image_path": "34de6e5a6e1c0dc0276242fb4a3c50fc6b4cfff1772423c9300cd006f385fc6c.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 538, + 287, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 538, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 538, + 287, + 635 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 538, + 287, + 635 + ], + "type": "inline_equation", + "content": "\\mathrm{LPENet}_{DM}" + }, + { + "bbox": [ + 46, + 538, + 287, + 635 + ], + "type": "text", + "content": " maintains a similar architecture to " + }, + { + "bbox": [ + 46, + 538, + 287, + 635 + ], + "type": "inline_equation", + "content": "\\mathrm{LPENet}" + }, + { + "bbox": [ + 46, + 538, + 287, + 635 + ], + "type": "text", + "content": " but with a modified input dimensionality for its first convolution layer, " + }, + { + "bbox": [ + 46, + 538, + 287, + 635 + ], + "type": "inline_equation", + "content": "AM(\\cdot)" + }, + { + "bbox": [ + 46, + 538, + 287, + 635 + ], + "type": "text", + "content": " denotes alignment module from [41]. The denoising neural network predicts the noise based on both " + }, + { + "bbox": [ + 46, + 538, + 287, + 635 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 46, + 538, + 287, + 635 + ], + "type": "text", + "content": " and the derived conditional feature " + }, + { + "bbox": [ + 46, + 538, + 287, + 635 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 538, + 287, + 635 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 46, + 538, + 287, + 635 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(Concat(z_{t},\\mathbf{D}),t)" + }, + { + "bbox": [ + 46, + 538, + 287, + 635 + ], + "type": "text", + "content": ". Upon substituting this estimated noise term into the equation governing the reverse process Eq. (4), we arrive at the following sampling formula:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 641, + 287, + 677 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 641, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 47, + 641, + 287, + 677 + ], + "type": "interline_equation", + "content": "\\boldsymbol {z} _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\boldsymbol {z} _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\boldsymbol {z} _ {t}, \\mathbf {D}, t\\right)\\right) + \\sqrt {1 - \\alpha_ {t}} \\boldsymbol {\\epsilon} _ {t}, \\tag {14}", + "image_path": "feeef9d22723866792fe574a394425252e89d638717ef1860b0789f70855fd4e.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon_t \\sim \\mathcal{N}(0, \\mathbf{I})" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": ". By iteratively sampling " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": " using Eq. (14) " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": " times, we progressively reconstruct the predicted LPR representation " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": " (i.e. " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\hat{z}) \\in \\mathbb{R}^{\\frac{H}{k} \\times \\frac{W}{k} \\times N}" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 72, + 545, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 253 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 253 + ], + "type": "text", + "content": "Training Strategy. Traditional DMs solely learn the probability distribution by optimizing the weighted variational bound (Eq. (5)), resulting in a slight deviation between the predicted prior and the actual prior. Integrating the DM directly with DHRNet might lead to a misalignment issue, thereby hampering the overall image processing performance. To address this, we joint training the DM and DHRNet. During each training iteration, we first sample the noise sample " + }, + { + "bbox": [ + 304, + 72, + 545, + 253 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 304, + 72, + 545, + 253 + ], + "type": "text", + "content": " through the diffusion process (Eq. (2)). Given that our denoising neural network is lightweight, we use a reverse process for " + }, + { + "bbox": [ + 304, + 72, + 545, + 253 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 72, + 545, + 253 + ], + "type": "text", + "content": " iterations based on the DDIM strategy [27] to generate the predicted prior feature " + }, + { + "bbox": [ + 304, + 72, + 545, + 253 + ], + "type": "inline_equation", + "content": "\\hat{z}" + }, + { + "bbox": [ + 304, + 72, + 545, + 253 + ], + "type": "text", + "content": ". This " + }, + { + "bbox": [ + 304, + 72, + 545, + 253 + ], + "type": "inline_equation", + "content": "\\hat{z}" + }, + { + "bbox": [ + 304, + 72, + 545, + 253 + ], + "type": "text", + "content": " guides DHRNet via PIM. The loss function in the second stage consists of the reconstruction loss " + }, + { + "bbox": [ + 304, + 72, + 545, + 253 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_r" + }, + { + "bbox": [ + 304, + 72, + 545, + 253 + ], + "type": "text", + "content": " (Eq. (11)) and the diffusion loss " + }, + { + "bbox": [ + 304, + 72, + 545, + 253 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{diff}" + }, + { + "bbox": [ + 304, + 72, + 545, + 253 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 323, + 261, + 545, + 298 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 261, + 545, + 298 + ], + "spans": [ + { + "bbox": [ + 323, + 261, + 545, + 298 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {d i f f}} = E _ {z _ {0}, t, \\epsilon_ {t} \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I})} \\left[ \\| \\epsilon_ {t} - \\epsilon_ {\\theta} (z _ {t}, D, t) \\| ^ {2} \\right] \\tag {15} \\\\ + \\left\\| \\hat {z} - z \\right\\| _ {1}. \\\\ \\end{array}", + "image_path": "7bb56e0798bdf2297e12d04e446e5fd35e709acbdf769f125cc0b64d7c2a0c34.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 308, + 373, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 308, + 373, + 319 + ], + "spans": [ + { + "bbox": [ + 306, + 308, + 373, + 319 + ], + "type": "text", + "content": "4.3. Inference" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 327, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 545, + 422 + ], + "type": "text", + "content": "In the inference stage, " + }, + { + "bbox": [ + 304, + 327, + 545, + 422 + ], + "type": "inline_equation", + "content": "L P E N t_{D M}" + }, + { + "bbox": [ + 304, + 327, + 545, + 422 + ], + "type": "text", + "content": " extracts a conditional feature " + }, + { + "bbox": [ + 304, + 327, + 545, + 422 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 304, + 327, + 545, + 422 + ], + "type": "text", + "content": " from aligned LDR feature (Eq. (13)), and we randomly sample a Gaussian noise " + }, + { + "bbox": [ + 304, + 327, + 545, + 422 + ], + "type": "inline_equation", + "content": "\\hat{z}_T" + }, + { + "bbox": [ + 304, + 327, + 545, + 422 + ], + "type": "text", + "content": ". Then, denoising neural network utilizes the " + }, + { + "bbox": [ + 304, + 327, + 545, + 422 + ], + "type": "inline_equation", + "content": "\\hat{z}_T" + }, + { + "bbox": [ + 304, + 327, + 545, + 422 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 327, + 545, + 422 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 304, + 327, + 545, + 422 + ], + "type": "text", + "content": " to estimate LPR after " + }, + { + "bbox": [ + 304, + 327, + 545, + 422 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 327, + 545, + 422 + ], + "type": "text", + "content": " iterations (Eq. (14)) based on DDIM [27]. After that, DHRNet exploits the LPR to restore HDR images as Eq. (10). More details about the DM training and the model inference can be found in the supplementary material." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 434, + 387, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 434, + 387, + 448 + ], + "spans": [ + { + "bbox": [ + 306, + 434, + 387, + 448 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 455, + 432, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 455, + 432, + 467 + ], + "spans": [ + { + "bbox": [ + 306, + 455, + 432, + 467 + ], + "type": "text", + "content": "5.1. Experimental Settings" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 474, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 676 + ], + "type": "text", + "content": "Datasets. All methods are trained using two publicly available datasets, employing identical training settings: Kalantari's dataset [11] and Hu's dataset [10]. Kalantari's dataset consists of 74 samples for training and 15 for testing, all captured under authentic environmental conditions. Each sample comprises three LDR images with exposure variations of " + }, + { + "bbox": [ + 304, + 474, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\{-2,0, +2\\}" + }, + { + "bbox": [ + 304, + 474, + 545, + 676 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 304, + 474, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\{-3,0, +3\\}" + }, + { + "bbox": [ + 304, + 474, + 545, + 676 + ], + "type": "text", + "content": ". In contrast, Hu's dataset is a synthetic dataset designed to emulate sensor realism, generated through a game engine. This dataset contains images captured at three distinct exposure levels " + }, + { + "bbox": [ + 304, + 474, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\{-2,0, +2\\}" + }, + { + "bbox": [ + 304, + 474, + 545, + 676 + ], + "type": "text", + "content": ", with our primary focus on dynamic scene images within this collection. Following the settings outlined in [10], we allocate the initial 85 samples for training, reserving the remaining 15 for testing. Moreover, to further validate the model's generalizability, we incorporate Sen et al.'s dataset [25] and Tursun et al.'s dataset [32] exclusively for qualitative assessment, as these datasets lack ground truths." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Evaluation Metrics. To conduct a quantitative comparison, we utilize five objective metrics: PSNR- " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " SSIM- " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " PSNR-L,SSIM-L,and HDR-VDP-2 [20].Here, the sub" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25736" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 95, + 164 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 95, + 164 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 95, + 164 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 95, + 164 + ], + "type": "image", + "image_path": "7b7857606bd0387f1c721bafcfef56959bf9e71469029765249756ee6eb6e7fa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 96, + 72, + 234, + 164 + ], + "blocks": [ + { + "bbox": [ + 96, + 72, + 234, + 164 + ], + "lines": [ + { + "bbox": [ + 96, + 72, + 234, + 164 + ], + "spans": [ + { + "bbox": [ + 96, + 72, + 234, + 164 + ], + "type": "image", + "image_path": "2ddb4897fdb8a274acf484ecc6b8847b275f9fd2e985b618a01f06da060e9562.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 135, + 164, + 192, + 171 + ], + "lines": [ + { + "bbox": [ + 135, + 164, + 192, + 171 + ], + "spans": [ + { + "bbox": [ + 135, + 164, + 192, + 171 + ], + "type": "text", + "content": "Tonemapped Result" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 235, + 72, + 265, + 163 + ], + "blocks": [ + { + "bbox": [ + 235, + 72, + 265, + 163 + ], + "lines": [ + { + "bbox": [ + 235, + 72, + 265, + 163 + ], + "spans": [ + { + "bbox": [ + 235, + 72, + 265, + 163 + ], + "type": "image", + "image_path": "889b8682a37653b8985fd7d0b20754ddc71e43659c9b840ba5702687bf22c87e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 266, + 72, + 296, + 164 + ], + "blocks": [ + { + "bbox": [ + 266, + 72, + 296, + 164 + ], + "lines": [ + { + "bbox": [ + 266, + 72, + 296, + 164 + ], + "spans": [ + { + "bbox": [ + 266, + 72, + 296, + 164 + ], + "type": "image", + "image_path": "7b7de5c855ab8fb37681575c80f9ca03faed921ff5a5333320e26564c373e28b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 250, + 165, + 288, + 171 + ], + "lines": [ + { + "bbox": [ + 250, + 165, + 288, + 171 + ], + "spans": [ + { + "bbox": [ + 250, + 165, + 288, + 171 + ], + "type": "text", + "content": "LDR-Patches" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 50, + 172, + 80, + 202 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 172, + 80, + 202 + ], + "spans": [ + { + "bbox": [ + 50, + 172, + 80, + 202 + ], + "type": "image", + "image_path": "3f3d08262f82ad6305d33750adb8f0664d926764279175640ec9f19d2183706c.jpg" + } + ] + } + ], + "index": 6, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 82, + 172, + 112, + 202 + ], + "blocks": [ + { + "bbox": [ + 82, + 172, + 112, + 202 + ], + "lines": [ + { + "bbox": [ + 82, + 172, + 112, + 202 + ], + "spans": [ + { + "bbox": [ + 82, + 172, + 112, + 202 + ], + "type": "image", + "image_path": "e5705c9ab45703d500c27cd44c24c2757207eb7d859ee6ffbf259d5deefe8a8a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 113, + 172, + 143, + 202 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 172, + 143, + 202 + ], + "spans": [ + { + "bbox": [ + 113, + 172, + 143, + 202 + ], + "type": "image", + "image_path": "e3abb9b8f291d21a5f7efdd5bad6af03656efabcf185eb53521af0dd9c70c2ce.jpg" + } + ] + } + ], + "index": 8, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 143, + 172, + 174, + 202 + ], + "blocks": [ + { + "bbox": [ + 143, + 172, + 174, + 202 + ], + "lines": [ + { + "bbox": [ + 143, + 172, + 174, + 202 + ], + "spans": [ + { + "bbox": [ + 143, + 172, + 174, + 202 + ], + "type": "image", + "image_path": "088accdbb9e4ef077f43ecdb11edb2ebcc7cb42b5980ea1ad085efb213a8c184.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 175, + 172, + 205, + 202 + ], + "blocks": [ + { + "bbox": [ + 175, + 172, + 205, + 202 + ], + "lines": [ + { + "bbox": [ + 175, + 172, + 205, + 202 + ], + "spans": [ + { + "bbox": [ + 175, + 172, + 205, + 202 + ], + "type": "image", + "image_path": "3177ca47349bee3ca18d8dbbdd0855f0f3e8616b808cc2f477c234b9742e5025.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 205, + 172, + 235, + 202 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 172, + 235, + 202 + ], + "spans": [ + { + "bbox": [ + 205, + 172, + 235, + 202 + ], + "type": "image", + "image_path": "3689309cb1a44d9353487e69fbd2950f6813413cf21fae4f0a906d05fcdd4040.jpg" + } + ] + } + ], + "index": 11, + "type": "text" + }, + { + "bbox": [ + 236, + 172, + 266, + 202 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 172, + 266, + 202 + ], + "spans": [ + { + "bbox": [ + 236, + 172, + 266, + 202 + ], + "type": "image", + "image_path": "c48f6b01ce6de261254af8200b522409bca41f2395c36257c42d1d8415ff5dd2.jpg" + } + ] + } + ], + "index": 12, + "type": "text" + }, + { + "bbox": [ + 266, + 172, + 296, + 202 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 172, + 296, + 202 + ], + "spans": [ + { + "bbox": [ + 266, + 172, + 296, + 202 + ], + "type": "image", + "image_path": "dcb0fc204763044f181c4e0f6dff715be17193d71b3feaee0752c61c37851248.jpg" + } + ] + } + ], + "index": 13, + "type": "text" + }, + { + "bbox": [ + 51, + 204, + 80, + 233 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 204, + 80, + 233 + ], + "spans": [ + { + "bbox": [ + 51, + 204, + 80, + 233 + ], + "type": "image", + "image_path": "9e28ec3224ab491fd459f9897dbb2d078e1f3616bf77a6fd35d5c59f7014ac3a.jpg" + } + ] + } + ], + "index": 14, + "type": "text" + }, + { + "bbox": [ + 82, + 204, + 111, + 233 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 204, + 111, + 233 + ], + "spans": [ + { + "bbox": [ + 82, + 204, + 111, + 233 + ], + "type": "image", + "image_path": "323cca5c1db996ea3de09181580296dacc16917789f7c09aad46f4c43042a6fd.jpg" + } + ] + } + ], + "index": 15, + "type": "text" + }, + { + "bbox": [ + 112, + 204, + 143, + 233 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 204, + 143, + 233 + ], + "spans": [ + { + "bbox": [ + 112, + 204, + 143, + 233 + ], + "type": "image", + "image_path": "466da06c29315b79a7eafc63857e2756eca97602aebb64d571d6dabca0f8679e.jpg" + } + ] + } + ], + "index": 16, + "type": "text" + }, + { + "bbox": [ + 143, + 204, + 173, + 233 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 204, + 173, + 233 + ], + "spans": [ + { + "bbox": [ + 143, + 204, + 173, + 233 + ], + "type": "image", + "image_path": "13d6d5cc3665182e24fceb91425cc6bc77d9afedc5ee1ea11b863ba80792d0b9.jpg" + } + ] + } + ], + "index": 17, + "type": "text" + }, + { + "bbox": [ + 174, + 204, + 204, + 233 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 204, + 204, + 233 + ], + "spans": [ + { + "bbox": [ + 174, + 204, + 204, + 233 + ], + "type": "image", + "image_path": "8457e77d407c6697d6bac7fd2e4c6af22a173de5994f21ed7bf9cf037f50356e.jpg" + } + ] + } + ], + "index": 18, + "type": "text" + }, + { + "bbox": [ + 205, + 204, + 235, + 233 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 204, + 235, + 233 + ], + "spans": [ + { + "bbox": [ + 205, + 204, + 235, + 233 + ], + "type": "image", + "image_path": "530059aee11198d177008a6618d019f30025eb7a11d081363014f3ba6aa5f250.jpg" + } + ] + } + ], + "index": 19, + "type": "text" + }, + { + "bbox": [ + 236, + 204, + 266, + 233 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 204, + 266, + 233 + ], + "spans": [ + { + "bbox": [ + 236, + 204, + 266, + 233 + ], + "type": "image", + "image_path": "743bd30be3beecd0c648801fb3e49eb8cae6a826efcec5a447d3b38de433ec37.jpg" + } + ] + } + ], + "index": 20, + "type": "text" + }, + { + "bbox": [ + 266, + 204, + 296, + 233 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 204, + 296, + 233 + ], + "spans": [ + { + "bbox": [ + 266, + 204, + 296, + 233 + ], + "type": "image", + "image_path": "9a1167415af1bbee947ac1f95bc12ad72bce00f2f858cecce81a76b1dd61c4bc.jpg" + } + ] + } + ], + "index": 21, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 299, + 72, + 344, + 163 + ], + "blocks": [ + { + "bbox": [ + 299, + 72, + 344, + 163 + ], + "lines": [ + { + "bbox": [ + 299, + 72, + 344, + 163 + ], + "spans": [ + { + "bbox": [ + 299, + 72, + 344, + 163 + ], + "type": "image", + "image_path": "fef82b101c723bff618b0870e9921d1bba6362a8ba934fe5ff5e3bf33bb18e48.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 345, + 72, + 390, + 163 + ], + "blocks": [ + { + "bbox": [ + 345, + 72, + 390, + 163 + ], + "lines": [ + { + "bbox": [ + 345, + 72, + 390, + 163 + ], + "spans": [ + { + "bbox": [ + 345, + 72, + 390, + 163 + ], + "type": "image", + "image_path": "131aaf67b61e8a48f7b2cce5d3631d67b178bc367dfad6df2026b7346ba180f8.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 391, + 72, + 482, + 163 + ], + "blocks": [ + { + "bbox": [ + 391, + 72, + 482, + 163 + ], + "lines": [ + { + "bbox": [ + 391, + 72, + 482, + 163 + ], + "spans": [ + { + "bbox": [ + 391, + 72, + 482, + 163 + ], + "type": "image", + "image_path": "7b3395093c25bafede3c26ce245f8d8462e6f28c91f593463313c77a07a18799.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 484, + 72, + 514, + 163 + ], + "blocks": [ + { + "bbox": [ + 484, + 72, + 514, + 163 + ], + "lines": [ + { + "bbox": [ + 484, + 72, + 514, + 163 + ], + "spans": [ + { + "bbox": [ + 484, + 72, + 514, + 163 + ], + "type": "image", + "image_path": "46fc7cf1113732f5c64d2766ba8f62317b0f85f3da4600ee1419c56e5fb416b2.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 514, + 72, + 544, + 163 + ], + "blocks": [ + { + "bbox": [ + 514, + 72, + 544, + 163 + ], + "lines": [ + { + "bbox": [ + 514, + 72, + 544, + 163 + ], + "spans": [ + { + "bbox": [ + 514, + 72, + 544, + 163 + ], + "type": "image", + "image_path": "56688c7157a7627407afdcc270e999c9dde76315e5d91451537382a9655bfdc0.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 440, + 164, + 482, + 171 + ], + "blocks": [ + { + "bbox": [ + 383, + 164, + 440, + 171 + ], + "lines": [ + { + "bbox": [ + 383, + 164, + 440, + 171 + ], + "spans": [ + { + "bbox": [ + 383, + 164, + 440, + 171 + ], + "type": "text", + "content": "Tonemapped Result" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 440, + 164, + 482, + 171 + ], + "lines": [ + { + "bbox": [ + 440, + 164, + 482, + 171 + ], + "spans": [ + { + "bbox": [ + 440, + 164, + 482, + 171 + ], + "type": "image", + "image_path": "691d89c830fe024b545b00633e37f4b33cd43524b16ff0ee6e208c778715c351.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "bbox": [ + 50, + 172, + 80, + 202 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 172, + 80, + 202 + ], + "spans": [ + { + "bbox": [ + 50, + 172, + 80, + 202 + ], + "type": "image", + "image_path": "3f3d08262f82ad6305d33750adb8f0664d926764279175640ec9f19d2183706c.jpg" + } + ] + } + ], + "index": 33, + "type": "text" + }, + { + "bbox": [ + 82, + 172, + 111, + 202 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 172, + 111, + 202 + ], + "spans": [ + { + "bbox": [ + 82, + 172, + 111, + 202 + ], + "type": "image", + "image_path": "fc00b80a0aad338fb23c5b197b231a9b8c950c0d15f4d02839dae4995b41ca45.jpg" + } + ] + } + ], + "index": 34, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 112, + 172, + 143, + 202 + ], + "blocks": [ + { + "bbox": [ + 112, + 172, + 143, + 202 + ], + "lines": [ + { + "bbox": [ + 112, + 172, + 143, + 202 + ], + "spans": [ + { + "bbox": [ + 112, + 172, + 143, + 202 + ], + "type": "image", + "image_path": "9dc4e1ba591d95bff701a7ad0835c8f7573887aa6b15cbc9841b145facef45be.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "bbox": [ + 143, + 172, + 173, + 202 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 172, + 173, + 202 + ], + "spans": [ + { + "bbox": [ + 143, + 172, + 173, + 202 + ], + "type": "image", + "image_path": "3c4bdb2f279a8bec7012156bedc8e64bc7d4d7051e8b03751777272a8f4d9c7a.jpg" + } + ] + } + ], + "index": 36, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 174, + 172, + 204, + 202 + ], + "blocks": [ + { + "bbox": [ + 174, + 172, + 204, + 202 + ], + "lines": [ + { + "bbox": [ + 174, + 172, + 204, + 202 + ], + "spans": [ + { + "bbox": [ + 174, + 172, + 204, + 202 + ], + "type": "image", + "image_path": "7161664df37b8d64e4b817c41e42fc2a31e91a5c2a2dd850c9fef3b569f1c093.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "bbox": [ + 205, + 172, + 235, + 202 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 172, + 235, + 202 + ], + "spans": [ + { + "bbox": [ + 205, + 172, + 235, + 202 + ], + "type": "image", + "image_path": "3689309cb1a44d9353487e69fbd2950f6813413cf21fae4f0a906d05fcdd4040.jpg" + } + ] + } + ], + "index": 38, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 235, + 172, + 266, + 202 + ], + "blocks": [ + { + "bbox": [ + 235, + 172, + 266, + 202 + ], + "lines": [ + { + "bbox": [ + 235, + 172, + 266, + 202 + ], + "spans": [ + { + "bbox": [ + 235, + 172, + 266, + 202 + ], + "type": "image", + "image_path": "c5bc93054f887c1c909db2dd0fac183d5cedd6d78bc4571b862847fab8db8b2a.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "bbox": [ + 266, + 172, + 296, + 202 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 172, + 296, + 202 + ], + "spans": [ + { + "bbox": [ + 266, + 172, + 296, + 202 + ], + "type": "image", + "image_path": "dcb0fc204763044f181c4e0f6dff715be17193d71b3feaee0752c61c37851248.jpg" + } + ] + } + ], + "index": 40, + "type": "text" + }, + { + "bbox": [ + 51, + 204, + 80, + 233 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 204, + 80, + 233 + ], + "spans": [ + { + "bbox": [ + 51, + 204, + 80, + 233 + ], + "type": "image", + "image_path": "9e28ec3224ab491fd459f9897dbb2d078e1f3616bf77a6fd35d5c59f7014ac3a.jpg" + } + ] + } + ], + "index": 41, + "type": "text" + }, + { + "bbox": [ + 82, + 204, + 111, + 233 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 204, + 111, + 233 + ], + "spans": [ + { + "bbox": [ + 82, + 204, + 111, + 233 + ], + "type": "image", + "image_path": "323cca5c1db996ea3de09181580296dacc16917789f7c09aad46f4c43042a6fd.jpg" + } + ] + } + ], + "index": 42, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 112, + 204, + 143, + 234 + ], + "blocks": [ + { + "bbox": [ + 89, + 240, + 242, + 252 + ], + "lines": [ + { + "bbox": [ + 89, + 240, + 242, + 252 + ], + "spans": [ + { + "bbox": [ + 89, + 240, + 242, + 252 + ], + "type": "text", + "content": "(a) Example from Kalantari et al.'s dataset [11]" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 112, + 204, + 143, + 234 + ], + "lines": [ + { + "bbox": [ + 112, + 204, + 143, + 234 + ], + "spans": [ + { + "bbox": [ + 112, + 204, + 143, + 234 + ], + "type": "image", + "image_path": "bbcda58e6437ff10fdc6e3ac16378bed102a425af2882c90586e8f5ff0286ef1.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 260, + 545, + 282 + ], + "lines": [ + { + "bbox": [ + 46, + 260, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 46, + 260, + 545, + 282 + ], + "type": "text", + "content": "Figure 4. Visual comparisons are conducted on testing data from various datasets, focusing on zoomed-in local areas of the HDR images estimated by our method and the compared techniques. Our model demonstrates the ability to generate HDR images of superior quality." + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_caption" + } + ], + "index": 43 + }, + { + "bbox": [ + 143, + 204, + 173, + 233 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 204, + 173, + 233 + ], + "spans": [ + { + "bbox": [ + 143, + 204, + 173, + 233 + ], + "type": "image", + "image_path": "13d6d5cc3665182e24fceb91425cc6bc77d9afedc5ee1ea11b863ba80792d0b9.jpg" + } + ] + } + ], + "index": 44, + "type": "text" + }, + { + "bbox": [ + 174, + 204, + 204, + 233 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 204, + 204, + 233 + ], + "spans": [ + { + "bbox": [ + 174, + 204, + 204, + 233 + ], + "type": "image", + "image_path": "8457e77d407c6697d6bac7fd2e4c6af22a173de5994f21ed7bf9cf037f50356e.jpg" + } + ] + } + ], + "index": 45, + "type": "text" + }, + { + "bbox": [ + 205, + 204, + 235, + 233 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 204, + 235, + 233 + ], + "spans": [ + { + "bbox": [ + 205, + 204, + 235, + 233 + ], + "type": "image", + "image_path": "530059aee11198d177008a6618d019f30025eb7a11d081363014f3ba6aa5f250.jpg" + } + ] + } + ], + "index": 46, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 235, + 204, + 266, + 233 + ], + "blocks": [ + { + "bbox": [ + 235, + 204, + 266, + 233 + ], + "lines": [ + { + "bbox": [ + 235, + 204, + 266, + 233 + ], + "spans": [ + { + "bbox": [ + 235, + 204, + 266, + 233 + ], + "type": "image", + "image_path": "1f80ff558967179b3390143af807153da5a5a44b924e3c870b710b695f0f5780.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + } + ], + "index": 47 + }, + { + "bbox": [ + 266, + 204, + 296, + 233 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 204, + 296, + 233 + ], + "spans": [ + { + "bbox": [ + 266, + 204, + 296, + 233 + ], + "type": "image", + "image_path": "9a1167415af1bbee947ac1f95bc12ad72bce00f2f858cecce81a76b1dd61c4bc.jpg" + } + ] + } + ], + "index": 48, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 297, + 172, + 328, + 233 + ], + "blocks": [ + { + "bbox": [ + 314, + 165, + 331, + 171 + ], + "lines": [ + { + "bbox": [ + 314, + 165, + 331, + 171 + ], + "spans": [ + { + "bbox": [ + 314, + 165, + 331, + 171 + ], + "type": "text", + "content": "LDRs" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 297, + 172, + 328, + 233 + ], + "lines": [ + { + "bbox": [ + 297, + 172, + 328, + 233 + ], + "spans": [ + { + "bbox": [ + 297, + 172, + 328, + 233 + ], + "type": "image", + "image_path": "c49a2a7497c5df59e526d4b5999d4bc1b08176031adeb3801527a00983eed83b.jpg" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_body" + } + ], + "index": 49 + }, + { + "type": "image", + "bbox": [ + 328, + 172, + 359, + 232 + ], + "blocks": [ + { + "bbox": [ + 328, + 172, + 359, + 232 + ], + "lines": [ + { + "bbox": [ + 328, + 172, + 359, + 232 + ], + "spans": [ + { + "bbox": [ + 328, + 172, + 359, + 232 + ], + "type": "image", + "image_path": "89028f331197adeb1d139b87b4ba5240d9a7ff4f99e753d1da267601d3bdf03a.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 360, + 172, + 390, + 232 + ], + "blocks": [ + { + "bbox": [ + 360, + 172, + 390, + 232 + ], + "lines": [ + { + "bbox": [ + 360, + 172, + 390, + 232 + ], + "spans": [ + { + "bbox": [ + 360, + 172, + 390, + 232 + ], + "type": "image", + "image_path": "18d5893b6e67f0908cf482fc1ab24ea5a37441c4c16b66e47659524d874ced7a.jpg" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 370, + 233, + 504, + 239 + ], + "lines": [ + { + "bbox": [ + 370, + 233, + 504, + 239 + ], + "spans": [ + { + "bbox": [ + 370, + 233, + 504, + 239 + ], + "type": "text", + "content": "AHDR HDR-GAN CA-VIT HyHDR Ours" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_caption" + } + ], + "index": 51 + }, + { + "type": "image", + "bbox": [ + 391, + 172, + 421, + 232 + ], + "blocks": [ + { + "bbox": [ + 391, + 172, + 421, + 232 + ], + "lines": [ + { + "bbox": [ + 391, + 172, + 421, + 232 + ], + "spans": [ + { + "bbox": [ + 391, + 172, + 421, + 232 + ], + "type": "image", + "image_path": "d087c4d450c606f3d187eeb38e7284ac40aabc114f7eb8e8027546e9d4bb2253.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + } + ], + "index": 52 + }, + { + "type": "image", + "bbox": [ + 422, + 172, + 452, + 232 + ], + "blocks": [ + { + "bbox": [ + 422, + 172, + 452, + 232 + ], + "lines": [ + { + "bbox": [ + 422, + 172, + 452, + 232 + ], + "spans": [ + { + "bbox": [ + 422, + 172, + 452, + 232 + ], + "type": "image", + "image_path": "4e08ab3f33793659f9087c10abba19ba314bb40aa1b2d46e1e4b593769eb53f3.jpg" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_body" + } + ], + "index": 53 + }, + { + "type": "image", + "bbox": [ + 452, + 172, + 482, + 232 + ], + "blocks": [ + { + "bbox": [ + 452, + 172, + 482, + 232 + ], + "lines": [ + { + "bbox": [ + 452, + 172, + 482, + 232 + ], + "spans": [ + { + "bbox": [ + 452, + 172, + 482, + 232 + ], + "type": "image", + "image_path": "d0d16a55a034f2984ff6cbc6251a7cbd754894bd5e9b20c54ec087c79548b88d.jpg" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_body" + } + ], + "index": 54 + }, + { + "type": "image", + "bbox": [ + 483, + 172, + 512, + 232 + ], + "blocks": [ + { + "bbox": [ + 484, + 165, + 514, + 171 + ], + "lines": [ + { + "bbox": [ + 484, + 165, + 514, + 171 + ], + "spans": [ + { + "bbox": [ + 484, + 165, + 514, + 171 + ], + "type": "text", + "content": "LDR-" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 483, + 172, + 512, + 232 + ], + "lines": [ + { + "bbox": [ + 483, + 172, + 512, + 232 + ], + "spans": [ + { + "bbox": [ + 483, + 172, + 512, + 232 + ], + "type": "image", + "image_path": "d216f66dd070b38e234aebae39952d2ea1f72aa8374811f5fa8ecc61a78dfe8d.jpg" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 507, + 233, + 534, + 239 + ], + "lines": [ + { + "bbox": [ + 507, + 233, + 534, + 239 + ], + "spans": [ + { + "bbox": [ + 507, + 233, + 534, + 239 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_caption" + } + ], + "index": 55 + }, + { + "type": "image", + "bbox": [ + 513, + 172, + 544, + 232 + ], + "blocks": [ + { + "bbox": [ + 514, + 165, + 544, + 171 + ], + "lines": [ + { + "bbox": [ + 514, + 165, + 544, + 171 + ], + "spans": [ + { + "bbox": [ + 514, + 165, + 544, + 171 + ], + "type": "text", + "content": "Patches" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 513, + 172, + 544, + 232 + ], + "lines": [ + { + "bbox": [ + 513, + 172, + 544, + 232 + ], + "spans": [ + { + "bbox": [ + 513, + 172, + 544, + 232 + ], + "type": "image", + "image_path": "73ff7f8951f798bf64fa368fa3f80ecd74113a5823cacba4897d48e5d7de1e76.jpg" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_body" + } + ], + "index": 56 + }, + { + "bbox": [ + 46, + 304, + 286, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 286, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 286, + 327 + ], + "type": "text", + "content": "scripts " + }, + { + "bbox": [ + 46, + 304, + 286, + 327 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 304, + 286, + 327 + ], + "type": "text", + "content": " and L denote that these metrics are computed in the tonemapped and linear domains, respectively." + } + ] + } + ], + "index": 60 + }, + { + "bbox": [ + 46, + 327, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 287, + 567 + ], + "type": "text", + "content": "Implementation Details. Our implementation is conducted using PyTorch, and each training stage converges after 300 epochs on four NVIDIA A100 GPUs. We employ the Adam optimizer with an initial learning rate of 1e-4, which decays by a factor of 0.1 after every 50 epochs. The training dataset is processed by cropping " + }, + { + "bbox": [ + 46, + 327, + 287, + 567 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 46, + 327, + 287, + 567 + ], + "type": "text", + "content": " patches with a stride of 64, and the batch size is set to 64. For LF-Diff, we adopt a variant of the commonly used U-Net architecture from [19] as the denoising network, with 3 levels of blocks " + }, + { + "bbox": [ + 46, + 327, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\{2, 2, 2\\}" + }, + { + "bbox": [ + 46, + 327, + 287, + 567 + ], + "type": "text", + "content": ". During training, the time step " + }, + { + "bbox": [ + 46, + 327, + 287, + 567 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 327, + 287, + 567 + ], + "type": "text", + "content": " is set to 200, and the implicit sampling step " + }, + { + "bbox": [ + 46, + 327, + 287, + 567 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 327, + 287, + 567 + ], + "type": "text", + "content": " is set to 10 for both training and inference phases to achieve efficient restoration. In DHRNet, the parameters " + }, + { + "bbox": [ + 46, + 327, + 287, + 567 + ], + "type": "inline_equation", + "content": "N_{i} \\in \\{L_{1}, L_{2}, L_{3}\\}" + }, + { + "bbox": [ + 46, + 327, + 287, + 567 + ], + "type": "text", + "content": " are set to " + }, + { + "bbox": [ + 46, + 327, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\{3, 3, 3\\}" + }, + { + "bbox": [ + 46, + 327, + 287, + 567 + ], + "type": "text", + "content": ", and the channel " + }, + { + "bbox": [ + 46, + 327, + 287, + 567 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 327, + 287, + 567 + ], + "type": "text", + "content": " is set to 60. The number of attention heads for the self-attention branch at the same level is set to " + }, + { + "bbox": [ + 46, + 327, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\{6, 6, 6\\}" + }, + { + "bbox": [ + 46, + 327, + 287, + 567 + ], + "type": "text", + "content": ", and the channel expansion factor in FFN is 2.66. The kernels for avgpool in PIM and FRM are set to 4 and 2, respectively. For LPENet, it comprises 4 residual blocks, with a Pixelunshuffle downsampling factor of 4 and the output channel of 3." + } + ] + } + ], + "index": 61 + }, + { + "bbox": [ + 47, + 575, + 287, + 587 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 287, + 587 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 287, + 587 + ], + "type": "text", + "content": "5.2. Comparison with the State-of-the-art Methods" + } + ] + } + ], + "index": 62 + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": "In this section, we evaluate the performance of the proposed LF-Diff method and present experimental results to validate its reconstruction performance compared to state-of-the-art techniques. Specifically, we evaluate LF-Diff against two patch-based methods [9, 25], a optical flow-based method [11], five CNN-based methods [1, 15, 21, 38, 41], two ViT-based methods [16, 43], along with a diffusion-based method [44]. It's worth noting that all deep learning methods employ the same training dataset and settings for consistency." + } + ] + } + ], + "index": 63 + }, + { + "bbox": [ + 304, + 303, + 545, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 303, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 303, + 545, + 483 + ], + "type": "text", + "content": "Datasets w/ Ground Truth. The quantitative outcomes for LF-Diff on two datasets are presented in Tab. 1. Our method is compared with various state-of-the-art approaches using the testing data from [11] and [10], which includes challenging samples characterized by saturated backgrounds and foreground motions. All quantitative results are averaged across testing images. Notably, LF-Diff exhibits a significant improvement over other leading methods, exceeding the performance of the DM-based method DiffHDR [44], by " + }, + { + "bbox": [ + 304, + 303, + 545, + 483 + ], + "type": "inline_equation", + "content": "0.65\\mathrm{dB}" + }, + { + "bbox": [ + 304, + 303, + 545, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 303, + 545, + 483 + ], + "type": "inline_equation", + "content": "0.86\\mathrm{dB}" + }, + { + "bbox": [ + 304, + 303, + 545, + 483 + ], + "type": "text", + "content": " in PSNR- " + }, + { + "bbox": [ + 304, + 303, + 545, + 483 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 303, + 545, + 483 + ], + "type": "text", + "content": " and PSNR-L metrics, respectively, based on Kalantari's dataset [11]. Moreover, LF-Diff showcases superior performance compared to the runner-up method, HyHDR [43], with improvements of " + }, + { + "bbox": [ + 304, + 303, + 545, + 483 + ], + "type": "inline_equation", + "content": "0.28\\mathrm{dB}" + }, + { + "bbox": [ + 304, + 303, + 545, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 303, + 545, + 483 + ], + "type": "inline_equation", + "content": "0.19\\mathrm{dB}" + }, + { + "bbox": [ + 304, + 303, + 545, + 483 + ], + "type": "text", + "content": " in PSNR- " + }, + { + "bbox": [ + 304, + 303, + 545, + 483 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 303, + 545, + 483 + ], + "type": "text", + "content": " and PSNR-L metrics, respectively, on Hu's dataset [10]." + } + ] + } + ], + "index": 64 + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "content": "As shown in Fig. 4 (a) and (b), the datasets present some challenging samples due to information loss in the LDR images. Most existing approaches struggle in these areas, producing ghosting artifacts due to large motions and occlusions. Kalantari's method [11] and DHDR [38] struggle with background motion due to error-prone alignments (e.g., optical flow), resulting in undesirable ghosting. While HDR-GAN [21] exhibits noticeable ghosting artifacts around the balcony area and introduces erroneous color information, AHDR [41] aligns LDR images using convolutional spatial attention. However, it unintentionally suppresses valuable contextual information and encounters difficulties with significant motions in over/underexposed scenes. CA-ViT [16], relying on patch-based sampling, produces noticeable blocky ghosting. In contrast, aided by the DM, both Diff-HDR [44] and our LF-Diff generate HDR images aligned with human perception. Notably, our approach outperforms Diff-HDR with faster inference speeds and a smaller computational overhead." + } + ] + } + ], + "index": 65 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25737" + } + ] + } + ], + "index": 66 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 71, + 97, + 166 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 97, + 166 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 97, + 166 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 97, + 166 + ], + "type": "image", + "image_path": "5d31bc04308f765093082c38c427f59d11e4e137a5bf02d46c7d039ce4b749a4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 97, + 72, + 235, + 166 + ], + "blocks": [ + { + "bbox": [ + 97, + 72, + 235, + 166 + ], + "lines": [ + { + "bbox": [ + 97, + 72, + 235, + 166 + ], + "spans": [ + { + "bbox": [ + 97, + 72, + 235, + 166 + ], + "type": "image", + "image_path": "6a2573a58dbcf7e471412254c872a0162cb38a7aa750162fc101fb43b9e4b1a8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 236, + 72, + 264, + 166 + ], + "blocks": [ + { + "bbox": [ + 236, + 72, + 264, + 166 + ], + "lines": [ + { + "bbox": [ + 236, + 72, + 264, + 166 + ], + "spans": [ + { + "bbox": [ + 236, + 72, + 264, + 166 + ], + "type": "image", + "image_path": "16c62efe2965acc99aa264c8d8da1a239cf5b08f6d6fc64808404db4e23b8cde.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 264, + 72, + 295, + 166 + ], + "blocks": [ + { + "bbox": [ + 264, + 72, + 295, + 166 + ], + "lines": [ + { + "bbox": [ + 264, + 72, + 295, + 166 + ], + "spans": [ + { + "bbox": [ + 264, + 72, + 295, + 166 + ], + "type": "image", + "image_path": "bf4603127f35737ccdb178cb012c6ae5d264b0caa86f767c2a8ebb899f6d8545.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 296, + 72, + 343, + 166 + ], + "blocks": [ + { + "bbox": [ + 296, + 72, + 343, + 166 + ], + "lines": [ + { + "bbox": [ + 296, + 72, + 343, + 166 + ], + "spans": [ + { + "bbox": [ + 296, + 72, + 343, + 166 + ], + "type": "image", + "image_path": "04d1f1950ced272c3d5f51529738b1d2b5d364fc85bd739d80e6deca2f98ba7a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 343, + 72, + 481, + 166 + ], + "blocks": [ + { + "bbox": [ + 343, + 72, + 481, + 166 + ], + "lines": [ + { + "bbox": [ + 343, + 72, + 481, + 166 + ], + "spans": [ + { + "bbox": [ + 343, + 72, + 481, + 166 + ], + "type": "image", + "image_path": "396a71f5a929f18bf1a324b9b1df57f78866a3856be2ff19d6a5be690e50bb8f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 482, + 72, + 528, + 166 + ], + "blocks": [ + { + "bbox": [ + 482, + 72, + 528, + 166 + ], + "lines": [ + { + "bbox": [ + 482, + 72, + 528, + 166 + ], + "spans": [ + { + "bbox": [ + 482, + 72, + 528, + 166 + ], + "type": "image", + "image_path": "dd9893aaef9588271472189f5e18d5bb86322584c41bf05f20d68de3fb338e0e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 528, + 72, + 544, + 166 + ], + "blocks": [ + { + "bbox": [ + 528, + 72, + 544, + 166 + ], + "lines": [ + { + "bbox": [ + 528, + 72, + 544, + 166 + ], + "spans": [ + { + "bbox": [ + 528, + 72, + 544, + 166 + ], + "type": "image", + "image_path": "17afcbe2cb497284b3f30b74acac5519bb6463ec40e7cf9a1c06b430ca8286cf.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 51, + 167, + 80, + 232 + ], + "blocks": [ + { + "bbox": [ + 51, + 167, + 80, + 232 + ], + "lines": [ + { + "bbox": [ + 51, + 167, + 80, + 232 + ], + "spans": [ + { + "bbox": [ + 51, + 167, + 80, + 232 + ], + "type": "image", + "image_path": "8b7847e3449307c3911ae8b411c03c9ac7c7f756f25c3c08295bae7d0065a8b0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 50, + 232, + 78, + 238 + ], + "lines": [ + { + "bbox": [ + 50, + 232, + 78, + 238 + ], + "spans": [ + { + "bbox": [ + 50, + 232, + 78, + 238 + ], + "type": "text", + "content": "Kalantari" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 80, + 232, + 105, + 238 + ], + "lines": [ + { + "bbox": [ + 80, + 232, + 105, + 238 + ], + "spans": [ + { + "bbox": [ + 80, + 232, + 105, + 238 + ], + "type": "text", + "content": "DHDR" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 80, + 167, + 111, + 232 + ], + "blocks": [ + { + "bbox": [ + 80, + 167, + 111, + 232 + ], + "lines": [ + { + "bbox": [ + 80, + 167, + 111, + 232 + ], + "spans": [ + { + "bbox": [ + 80, + 167, + 111, + 232 + ], + "type": "image", + "image_path": "a8d44f76469c82af20462c1db28f657f1c46f503205e43b633d1013798371fc4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 111, + 232, + 136, + 238 + ], + "lines": [ + { + "bbox": [ + 111, + 232, + 136, + 238 + ], + "spans": [ + { + "bbox": [ + 111, + 232, + 136, + 238 + ], + "type": "text", + "content": "AHDR" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 111, + 167, + 141, + 232 + ], + "blocks": [ + { + "bbox": [ + 111, + 167, + 141, + 232 + ], + "lines": [ + { + "bbox": [ + 111, + 167, + 141, + 232 + ], + "spans": [ + { + "bbox": [ + 111, + 167, + 141, + 232 + ], + "type": "image", + "image_path": "f8008a2662f63ce3e703b8104efb48464fc62204ff690fc843590e6f5f8a8730.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 141, + 232, + 173, + 238 + ], + "lines": [ + { + "bbox": [ + 141, + 232, + 173, + 238 + ], + "spans": [ + { + "bbox": [ + 141, + 232, + 173, + 238 + ], + "type": "text", + "content": "HDR-GAN" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 141, + 167, + 173, + 232 + ], + "blocks": [ + { + "bbox": [ + 141, + 167, + 173, + 232 + ], + "lines": [ + { + "bbox": [ + 141, + 167, + 173, + 232 + ], + "spans": [ + { + "bbox": [ + 141, + 167, + 173, + 232 + ], + "type": "image", + "image_path": "9a9d5f0f8a02a48f5ee01093fc4c0d1ffe20c2aab2e4e0f4f8037a69fca5eb72.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 174, + 167, + 203, + 232 + ], + "blocks": [ + { + "bbox": [ + 174, + 167, + 203, + 232 + ], + "lines": [ + { + "bbox": [ + 174, + 167, + 203, + 232 + ], + "spans": [ + { + "bbox": [ + 174, + 167, + 203, + 232 + ], + "type": "image", + "image_path": "75f759117a9c60caa16f831e75a98005ca75150ede5b07bd238cbd26c49fc396.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 174, + 232, + 203, + 238 + ], + "lines": [ + { + "bbox": [ + 174, + 232, + 203, + 238 + ], + "spans": [ + { + "bbox": [ + 174, + 232, + 203, + 238 + ], + "type": "text", + "content": "APNET" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 175, + 261, + 417, + 272 + ], + "lines": [ + { + "bbox": [ + 175, + 261, + 417, + 272 + ], + "spans": [ + { + "bbox": [ + 175, + 261, + 417, + 272 + ], + "type": "text", + "content": "Figure 5. Visual comparisons on the datasets without ground truth." + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 204, + 167, + 235, + 232 + ], + "blocks": [ + { + "bbox": [ + 204, + 167, + 235, + 232 + ], + "lines": [ + { + "bbox": [ + 204, + 167, + 235, + 232 + ], + "spans": [ + { + "bbox": [ + 204, + 167, + 235, + 232 + ], + "type": "image", + "image_path": "bfbeddcc4ec3f11838d3a83d5468d17a3ff7d5aafa61ff021cd0a3eb5bd06430.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 204, + 232, + 231, + 238 + ], + "lines": [ + { + "bbox": [ + 204, + 232, + 231, + 238 + ], + "spans": [ + { + "bbox": [ + 204, + 232, + 231, + 238 + ], + "type": "text", + "content": "CA-VIT" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 236, + 167, + 265, + 232 + ], + "blocks": [ + { + "bbox": [ + 236, + 167, + 265, + 232 + ], + "lines": [ + { + "bbox": [ + 236, + 167, + 265, + 232 + ], + "spans": [ + { + "bbox": [ + 236, + 167, + 265, + 232 + ], + "type": "image", + "image_path": "f74c6febd42f3236418fe6f283cfd1bba2c55ecdf26d8de4d319f34b0a1dde8e.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 236, + 232, + 265, + 238 + ], + "lines": [ + { + "bbox": [ + 236, + 232, + 265, + 238 + ], + "spans": [ + { + "bbox": [ + 236, + 232, + 265, + 238 + ], + "type": "text", + "content": "DiffHDR" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 266, + 167, + 296, + 232 + ], + "blocks": [ + { + "bbox": [ + 266, + 167, + 296, + 232 + ], + "lines": [ + { + "bbox": [ + 266, + 167, + 296, + 232 + ], + "spans": [ + { + "bbox": [ + 266, + 167, + 296, + 232 + ], + "type": "image", + "image_path": "612eb62b9ef537b2388e0ffc764c96efd631edab9ad5f07a4a10c94ec3a65b0e.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 276, + 232, + 290, + 238 + ], + "lines": [ + { + "bbox": [ + 276, + 232, + 290, + 238 + ], + "spans": [ + { + "bbox": [ + 276, + 232, + 290, + 238 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 296, + 167, + 326, + 232 + ], + "blocks": [ + { + "bbox": [ + 296, + 167, + 326, + 232 + ], + "lines": [ + { + "bbox": [ + 296, + 167, + 326, + 232 + ], + "spans": [ + { + "bbox": [ + 296, + 167, + 326, + 232 + ], + "type": "image", + "image_path": "7e87e3b30d5113fb1c123b6f793b4f0d024b07715c50e9145cf9c61330602014.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 298, + 232, + 326, + 238 + ], + "lines": [ + { + "bbox": [ + 298, + 232, + 326, + 238 + ], + "spans": [ + { + "bbox": [ + 298, + 232, + 326, + 238 + ], + "type": "text", + "content": "Kalantari" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 327, + 167, + 358, + 232 + ], + "blocks": [ + { + "bbox": [ + 327, + 167, + 358, + 232 + ], + "lines": [ + { + "bbox": [ + 327, + 167, + 358, + 232 + ], + "spans": [ + { + "bbox": [ + 327, + 167, + 358, + 232 + ], + "type": "image", + "image_path": "8b10266a3e0467f31acaa4a557a5f26717be9448174efcdfd44675d9fd2d4867.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 332, + 232, + 353, + 238 + ], + "lines": [ + { + "bbox": [ + 332, + 232, + 353, + 238 + ], + "spans": [ + { + "bbox": [ + 332, + 232, + 353, + 238 + ], + "type": "text", + "content": "DHDR" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 359, + 167, + 389, + 232 + ], + "blocks": [ + { + "bbox": [ + 359, + 167, + 389, + 232 + ], + "lines": [ + { + "bbox": [ + 359, + 167, + 389, + 232 + ], + "spans": [ + { + "bbox": [ + 359, + 167, + 389, + 232 + ], + "type": "image", + "image_path": "91c8c49197b7d0e472c0aa5f21e7b23fff14ba29385331eb347305d0135e7a1b.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 364, + 232, + 422, + 238 + ], + "lines": [ + { + "bbox": [ + 364, + 232, + 422, + 238 + ], + "spans": [ + { + "bbox": [ + 364, + 232, + 422, + 238 + ], + "type": "text", + "content": "AHDR HDR-GAN" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 389, + 167, + 419, + 232 + ], + "blocks": [ + { + "bbox": [ + 389, + 167, + 419, + 232 + ], + "lines": [ + { + "bbox": [ + 389, + 167, + 419, + 232 + ], + "spans": [ + { + "bbox": [ + 389, + 167, + 419, + 232 + ], + "type": "image", + "image_path": "3274dbf0b58470e46abb7c1349f0e3f8667c94f11a9ed0b9b320b3cb01a7d372.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "bbox": [ + 391, + 232, + 422, + 238 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 232, + 422, + 238 + ], + "spans": [ + { + "bbox": [ + 391, + 232, + 422, + 238 + ], + "type": "text", + "content": "HDR-GAN" + } + ] + } + ], + "index": 31, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 422, + 167, + 451, + 232 + ], + "blocks": [ + { + "bbox": [ + 422, + 167, + 451, + 232 + ], + "lines": [ + { + "bbox": [ + 422, + 167, + 451, + 232 + ], + "spans": [ + { + "bbox": [ + 422, + 167, + 451, + 232 + ], + "type": "image", + "image_path": "0b8682997f38e13946d1d6a902a895178623bfbb8bcfeabebf8e7b7d89d105f5.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 438, + 232, + 448, + 238 + ], + "lines": [ + { + "bbox": [ + 438, + 232, + 448, + 238 + ], + "spans": [ + { + "bbox": [ + 438, + 232, + 448, + 238 + ], + "type": "text", + "content": "NET " + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 452, + 167, + 482, + 232 + ], + "blocks": [ + { + "bbox": [ + 452, + 167, + 482, + 232 + ], + "lines": [ + { + "bbox": [ + 452, + 167, + 482, + 232 + ], + "spans": [ + { + "bbox": [ + 452, + 167, + 482, + 232 + ], + "type": "image", + "image_path": "9f2f0efdf7068ec2a68408ca3044c9bea2d24ca535156cd1a80795533d3767cc.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 455, + 232, + 478, + 238 + ], + "lines": [ + { + "bbox": [ + 455, + 232, + 478, + 238 + ], + "spans": [ + { + "bbox": [ + 455, + 232, + 478, + 238 + ], + "type": "text", + "content": "CA-VIT" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 482, + 167, + 512, + 232 + ], + "blocks": [ + { + "bbox": [ + 482, + 167, + 512, + 232 + ], + "lines": [ + { + "bbox": [ + 482, + 167, + 512, + 232 + ], + "spans": [ + { + "bbox": [ + 482, + 167, + 512, + 232 + ], + "type": "image", + "image_path": "a3266d1c8cc0965083a3d701884ed52827db4888a09b02b75b18d1c29b092b87.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 487, + 232, + 512, + 238 + ], + "lines": [ + { + "bbox": [ + 487, + 232, + 512, + 238 + ], + "spans": [ + { + "bbox": [ + 487, + 232, + 512, + 238 + ], + "type": "text", + "content": "DiffHDR" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 514, + 167, + 534, + 232 + ], + "blocks": [ + { + "bbox": [ + 514, + 167, + 534, + 232 + ], + "lines": [ + { + "bbox": [ + 514, + 167, + 534, + 232 + ], + "spans": [ + { + "bbox": [ + 514, + 167, + 534, + 232 + ], + "type": "image", + "image_path": "5b10815ac919c9c7b3527fea620e87bc7bb3d206dc75870d1f22338e70ba09e3.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 523, + 232, + 538, + 238 + ], + "lines": [ + { + "bbox": [ + 523, + 232, + 538, + 238 + ], + "spans": [ + { + "bbox": [ + 523, + 232, + 538, + 238 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "type": "table", + "bbox": [ + 51, + 282, + 545, + 405 + ], + "blocks": [ + { + "bbox": [ + 51, + 282, + 545, + 405 + ], + "lines": [ + { + "bbox": [ + 51, + 282, + 545, + 405 + ], + "spans": [ + { + "bbox": [ + 51, + 282, + 545, + 405 + ], + "type": "table", + "html": "
DatasetsModelsSen[25]Hu[9]Kalantari[11]DHDR[38]AHDR[41]NHDRR[42]HDRGAN[21]APNT[1]CA-ViT[16]HyHDR[43]DiffH[44]Ours
KalantariPSNR-μ40.9532.1942.7441.6443.6242.4143.9243.9444.3244.6444.1144.76
PSNR-L38.3130.8441.2240.9141.0341.0841.5741.6142.1842.4741.7342.59
SSIM-μ0.98050.97160.98770.98690.99000.98870.99050.98980.99160.99150.99110.9919
SSIM-L0.97260.95060.98480.98580.98620.98610.98650.98790.98840.98940.98850.9906
HDR-VDP-255.7255.2560.5160.5062.3061.2165.4564.0566.0366.0565.5266.54
DatasetsModelsSen[25]Hu[9]Kalantari[11]DHDR[38]AHDR[41]NHDRR[42]HDRGAN[21]APNT[1]ADNet[15]CA-ViT[16]HyHDR[43]Ours
HuPSNR-μ31.4836.5641.6041.1345.7645.1545.8646.4146.7948.1048.4648.74
PSNR-L33.5836.9443.7641.2049.2248.7549.1447.9750.3851.1751.9152.10
SSIM-μ0.95310.98240.99140.98700.99560.99560.99450.99530.99080.99470.99590.9968
SSIM-L0.96340.98770.99380.99410.99800.99810.99890.99860.99870.99890.99910.9993
HDR-VDP-266.3967.5872.9470.8275.0474.8675.1973.0676.2177.1277.2477.35
", + "image_path": "ce1d27fb9bf3a718f3b109d723bd7dcef4a5b9e99a3a48cc8e8f1a1cd544f276.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "table_body" + } + ], + "index": 41 + }, + { + "bbox": [ + 88, + 413, + 503, + 425 + ], + "lines": [ + { + "bbox": [ + 88, + 413, + 503, + 425 + ], + "spans": [ + { + "bbox": [ + 88, + 413, + 503, + 425 + ], + "type": "text", + "content": "Table 1. The evaluation results on Kalantari's [11] and Hu's [10] datasets. The best results are highlighted in Bold." + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 443, + 287, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 443, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 443, + 287, + 574 + ], + "type": "text", + "content": "Datasets w/o Ground Truth. To assess the generalization capability of our proposed HDR imaging method, we evaluated the performance of the model trained on Kalantari et al.'s Dataset [11] by testing it on datasets from Sen et al. [25] and Tursun et al. [33], which lack ground truths. In Fig. 5 (a), numerous current methods encounter difficulties in restoring recover the large saturated region and large motion. Conversely, in Fig. 5 (b), while complete elimination of ghosting remains elusive for any method, our approach notably enhances image sharpness and detail compared to preceding techniques." + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": "Computational Budgets. We also conducted comparisons regarding model parameters and inference times with previous approaches. As illustrated in Table 2, patch match-based methods [9, 25] exhibit significantly longer inference times due to their CPU-based computation. Kalantari [11] requires considerable time for initial optical flow alignment. NHDRRNet [42], employing a U-shaped network architecture, demonstrates shorter inference times but substantially higher parameter counts compared to other methods. CAViT [16] has a large number of standard transformer blocks leading to high computational cost despite fewer parame" + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 304, + 443, + 545, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 443, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 304, + 443, + 545, + 503 + ], + "type": "text", + "content": "ters. DiffHDR [44] has high inference time and parameter count due to reconstructing HDR from pure noise. In comparison, our method effectively leverages the powerful distribution estimation capability of DM with orders of magnitude fewer parameters and computations." + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 306, + 511, + 400, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 511, + 400, + 525 + ], + "spans": [ + { + "bbox": [ + 306, + 511, + 400, + 525 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 304, + 531, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 531, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 545, + 567 + ], + "type": "text", + "content": "In this section, we investigate the impact of various designs within our proposed method. All experiments are conducted using the Kalantari's dataset [11]." + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": "Effects of Diffusion Prior. As shown in Tab. 3, we establish a regression-based baseline model without LPR generation. In this configuration, the PIM in DHRNet is replaced with FRM, and HDR images are reconstructed through a regression-based approach. LF-Diff" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "_{s1}" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": " is a pretrained model in the first stage utilizing ground-truth images to provide LPR. Compared to the baseline, this brings a 8.7dB PSNR-L improvement, demonstrating that an accurate compact prior can greatly enhance results. Variant LF-Diff" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "_{s2}" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": " w/o DM does not use the DM, but instead directly learns the LPR using LPENet. This achieves a 0.16dB PSNR gain over the baseline. When learning the LPR based" + } + ] + } + ], + "index": 48 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25738" + } + ] + } + ], + "index": 49 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 70, + 535, + 119 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 535, + 119 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 535, + 119 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 535, + 119 + ], + "type": "table", + "html": "
Methods\nEnvironmentSen [25]\n(CPU)Hu [9]\n(CPU)Kalantari [11]\n(CPU+GPU)AHDR [41]\n(GPU)NHDRR [42]\n(GPU)HDRGAN [21]\n(GPU)APNT [1]\n(GPU)CA-ViT [16]\n(GPU)DiffHDR [44]\n(GPU)Ours\n(GPU)
Times(s)61.81s79.77s29.14s0.30s0.31s0.29s2.62s5.34s7.53s0.72s
Para.(M)--0.30M1.24M38.10M2.56M37.98M1.22M74.99M7.48M
", + "image_path": "c12eed3086646bc4e2439f4d9f3ece2ec4f8f98124d00b61555cec998eca5498.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 156, + 289, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 156, + 289, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 156, + 289, + 288 + ], + "type": "text", + "content": "on the DM, LF-Diff" + }, + { + "bbox": [ + 46, + 156, + 289, + 288 + ], + "type": "inline_equation", + "content": "_{s2}" + }, + { + "bbox": [ + 46, + 156, + 289, + 288 + ], + "type": "text", + "content": " w/ DM further improves PSNR by 0.24dB compared to variant LF-Diff" + }, + { + "bbox": [ + 46, + 156, + 289, + 288 + ], + "type": "inline_equation", + "content": "_{s2}" + }, + { + "bbox": [ + 46, + 156, + 289, + 288 + ], + "type": "text", + "content": " w/o DM. This indicates that the DM demonstrates superior capability in accurately estimating the distribution for LPR prediction. Additionally, compared to DiffHDR, which requires over 70M Params for estimating complete images, LF-Diff only adds 2.39M Params over the baseline. Moreover, we present visual comparisons of the baseline (without the prior feature) and LF-Diff variants in Fig. 6. It can be observed that performing DM in a compact latent space to predict LPR effectively mitigates ghosting issues." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 303, + 288, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 303, + 288, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 303, + 288, + 471 + ], + "type": "text", + "content": "Effects of Joint Training Strategy. We conducted an ablation study on the joint training strategy. Under this strategy, only the DM is optimized in the second stage, referred to as Split-Training. Specifically, we first utilize the pretrained LPERNet to generate the prior feature " + }, + { + "bbox": [ + 46, + 303, + 288, + 471 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 303, + 288, + 471 + ], + "type": "text", + "content": " from the ground truth, and then apply the training objective defined in Eq. (15) to independently train the DM. Subsequently, the trained DM is directly integrated with the DHRNet for evaluation. It's worth noting that the DHRNet in the second stage utilizes pre-trained weights from the stage one without additional training. It is evident that LF-Diff significantly outperforms Split-Training by 1.51 dB in terms of PSNR value. This comparison underscores the effectiveness of joint training strategy over Split-Training." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 486, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 289, + 714 + ], + "type": "text", + "content": "Sampling Step. In addition to the lightweight DM structure and compact feature dimensions, our method achieves greater computational efficiency by employing a smaller sampling step, denoted as S, in the reverse process based on DDIM [27]. The performance of our method is demonstrated with different values of S ranging from 5 to 100, as presented in Tab. 4. Variations in sampling steps, when S is set to a value smaller than the training configuration, a noticeable performance degradation occurs. On the other hand, when S is larger than the training setting, the impact on performance is minimal, particularly with the SSIM metric remaining consistently unchanged. While a larger sampling step is known to enhance the visual quality of images in diffusion-based methods [24, 44], in our case, it primarily increases inference time. This observation suggests that larger sampling steps, such as " + }, + { + "bbox": [ + 46, + 486, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{S} = 1000" + }, + { + "bbox": [ + 46, + 486, + 289, + 714 + ], + "type": "text", + "content": " in DiffHDR [44] and SR3 [24], may not be essential for diffusion-based HDR reconstruction and other related low-level vision tasks when using our proposed framework." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 310, + 153, + 545, + 223 + ], + "blocks": [ + { + "bbox": [ + 96, + 127, + 496, + 138 + ], + "lines": [ + { + "bbox": [ + 96, + 127, + 496, + 138 + ], + "spans": [ + { + "bbox": [ + 96, + 127, + 496, + 138 + ], + "type": "text", + "content": "Table 2. Average runtime performance of various methods on the testing set with dimensions of " + }, + { + "bbox": [ + 96, + 127, + 496, + 138 + ], + "type": "inline_equation", + "content": "{1000} \\times {1500}" + }, + { + "bbox": [ + 96, + 127, + 496, + 138 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 153, + 545, + 223 + ], + "lines": [ + { + "bbox": [ + 310, + 153, + 545, + 223 + ], + "spans": [ + { + "bbox": [ + 310, + 153, + 545, + 223 + ], + "type": "table", + "html": "
MethodGTDMJoint TrainingPara.PSNR-μPSNR-L
BaselineXXX5.09M44.2342.19
LF-Diffs1✓XX5.24M50.4250.89
Split trainingX✓X7.48M41.9339.30
LF-Diffs2 w/o DMXX✓5.85M44.4542.35
LF-Diffs2 w/ DMX✓✓7.48M44.7642.59
", + "image_path": "d0e6deb6f3b38e8a9f489dc2ef3f1b59f095a15081fa9b1af298e526ee50b2be.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 310, + 232, + 541, + 243 + ], + "lines": [ + { + "bbox": [ + 310, + 232, + 541, + 243 + ], + "spans": [ + { + "bbox": [ + 310, + 232, + 541, + 243 + ], + "type": "text", + "content": "Table 3. Ablation studies of Diffusion Prior and Joint Training." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 307, + 253, + 544, + 350 + ], + "blocks": [ + { + "bbox": [ + 307, + 253, + 544, + 350 + ], + "lines": [ + { + "bbox": [ + 307, + 253, + 544, + 350 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 544, + 350 + ], + "type": "image", + "image_path": "3e88852fff0d036ef21d9874ce29f2276d4fa615db71bab1299e89c90e0e315a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 313, + 382, + 541, + 446 + ], + "blocks": [ + { + "bbox": [ + 334, + 360, + 517, + 372 + ], + "lines": [ + { + "bbox": [ + 334, + 360, + 517, + 372 + ], + "spans": [ + { + "bbox": [ + 334, + 360, + 517, + 372 + ], + "type": "text", + "content": "Figure 6. Qualitative results of our ablation study." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 382, + 541, + 446 + ], + "lines": [ + { + "bbox": [ + 313, + 382, + 541, + 446 + ], + "spans": [ + { + "bbox": [ + 313, + 382, + 541, + 446 + ], + "type": "table", + "html": "
Sampling stepPSNR-μPSNR-LSSIM-μSSIM-LTime(s)
S=539.7641.330.98190.98780.049
S=1044.7642.590.99190.99060.106
S=2044.7342.540.99190.99060.208
S=5044.6642.450.99190.99060.479
S=10044.6542.440.99190.99060.971
", + "image_path": "315f3feef2171286c2baf8b46a40e8add4e2d02ad17b7af508a7bd0839780d88.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 454, + 545, + 488 + ], + "lines": [ + { + "bbox": [ + 305, + 454, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 305, + 454, + 545, + 488 + ], + "type": "text", + "content": "Table 4. Ablation studies of various settings on the sampling step during the reverse process. Time(s) only denotes the time expenditure of the DM for the corresponding setting." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 502, + 379, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 502, + 379, + 514 + ], + "spans": [ + { + "bbox": [ + 306, + 502, + 379, + 514 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "text", + "content": "The potential of diffusion models in HDR deghosting has shown promising results, particularly in achieving visually perceptible outcomes. Different from image synthesis which generates each pixel from scratch, HDR imaging provides several LDR images as reference. Thus, it is inefficient to reconstruct the full HDR image starting from pure Gaussian noise. In this paper, we propose an efficient diffusion model called LF-Diff, consisting of LPENet, DHRNet, and a denoising network. Specifically, we apply the DM in a compact latent space to predict low-frequency priors of HDR images. These prior features provide explicit guidance for the image reconstruction process, thereby enhancing the details of the reconstructed HDR images. Compared to traditional DM-based methods, LF-Diff achieves accurate estimations and reduces artifacts in reconstructed images with much lower computational cost." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25739" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Jie Chen, Zaifeng Yang, Tsz Nam Chan, Hui Li, Junhui Hou, and Lap-Pui Chau. Attention-guided progressive neural texture fusion for high dynamic range image restoration. IEEE Transactions on Image Processing, 31:2661-2672, 2022. 2, 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 146, + 288, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 146, + 288, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 146, + 288, + 191 + ], + "type": "text", + "content": "[2] Nanxin Chen, Yu Zhang, Heiga Zen, Ron J Weiss, Mohammad Norouzi, and William Chan. Wavegrad: Estimating gradients for waveform generation. In International Conference on Learning Representations, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 288, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 288, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 288, + 235 + ], + "type": "text", + "content": "[3] Zheng Chen, Yulun Zhang, Ding Liu, Jinjin Gu, Linghe Kong, Xin Yuan, et al. Hierarchical integration diffusion model for realistic image deblurring. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 288, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 288, + 280 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 288, + 280 + ], + "type": "text", + "content": "[4] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In Advances in Neural Information Processing Systems, pages 8780-8794. Curran Associates, Inc., 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 280, + 288, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 280, + 288, + 313 + ], + "spans": [ + { + "bbox": [ + 53, + 280, + 288, + 313 + ], + "type": "text", + "content": "[5] Li Fang, Qian Wang, and Long Ye. Glgnet: light field angular superresolution with arbitrary interpolation rates. Visual Intelligence, 2(1):6, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 313, + 288, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 313, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 313, + 288, + 357 + ], + "type": "text", + "content": "[6] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 358, + 288, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 358, + 288, + 402 + ], + "spans": [ + { + "bbox": [ + 53, + 358, + 288, + 402 + ], + "type": "text", + "content": "[7] YongSeok Heo, KyoungMu Lee, SangUk Lee, Youngsu Moon, and Joonhyuk Cha. Ghost-free high dynamic range imaging. In IEEE Conference on Asian Conference on Computer Vision (ACCV), pages 486-500, 2011. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 403, + 288, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 288, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 288, + 446 + ], + "type": "text", + "content": "[8] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, pages 6840-6851. Curran Associates, Inc., 2020. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 447, + 288, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 288, + 491 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 288, + 491 + ], + "type": "text", + "content": "[9] Jun Hu, O. Gallo, K. Pulli, and Xiaobai Sun. HDR deghosting: How to deal with saturation? In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1163-1170, 2013. 2, 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 491, + 288, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 288, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 288, + 557 + ], + "type": "text", + "content": "[10] Jinhan Hu, Gyeongmin Choe, Zeeshan Nadir, Osama Nabil, Seok-Jun Lee, Hamid Sheikh, Youngjun Yoo, and Michael Polley. Sensor-realistic synthetic data engine for multiframe high dynamic range photography. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 516-517, 2020. 5, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 558, + 288, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 288, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 288, + 590 + ], + "type": "text", + "content": "[11] Nima Khademi Kalantari and Ravi Ramamoorthi. Deep high dynamic range imaging of dynamic scenes. ACM Transactions on Graphics, 36(4):1-12, 2017. 1, 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 591, + 288, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 288, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 288, + 634 + ], + "type": "text", + "content": "[12] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. Advances in Neural Information Processing Systems, 35:23593-23606, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 288, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 680 + ], + "type": "text", + "content": "[13] Haoying Li, Yifan Yang, Meng Chang, Shiqi Chen, Huajun Feng, Zhihai Xu, Qi Li, and Yueting Chen. Srdiff: Single image super-resolution with diffusion probabilistic models. Neurocomputing, 479:47-59, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 681, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 288, + 714 + ], + "type": "text", + "content": "[14] Ru Li, Chuan Wang, Jue Wang, Guanghui Liu, Heng-Yu Zhang, Bing Zeng, and Shuaicheng Liu. Uphdr-gan: Generative adversarial network for high dynamic range imaging" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "with unpaired data. IEEE Transactions on Circuits and Systems for Video Technology, 32(11):7532-7546, 2022. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 96, + 547, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 547, + 163 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 547, + 163 + ], + "type": "text", + "content": "[15] Zhen Liu, Wenjie Lin, Xinpeng Li, Qing Rao, Ting Jiang, Mingyan Han, Haoqiang Fan, Jian Sun, and Shuaicheng Liu. Adnet: Attention-guided deformable convolutional network for high dynamic range imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 463-470, 2021. 2, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 163, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 196 + ], + "type": "text", + "content": "[16] Zhen Liu, Yinglong Wang, Bing Zeng, and Shuaicheng Liu. Ghost-free high dynamic range imaging with context-aware transformer. pages 344-360, 2022. 1, 2, 4, 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 198, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 198, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 308, + 198, + 545, + 251 + ], + "type": "text", + "content": "[17] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 254, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 254, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 254, + 545, + 308 + ], + "type": "text", + "content": "[18] Zhisheng Lu, Juncheng Li, Hong Liu, Chaoyan Huang, Linlin Zhang, and Tieyong Zeng. Transformer for single image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 457-466, 2022. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 309, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 309, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 308, + 309, + 545, + 365 + ], + "type": "text", + "content": "[19] Ziwei Luo, Fredrik K Gustafsson, Zheng Zhao, Jens Sjolund, and Thomas B Schön. Refusion: Enabling large-size realistic image restoration with latent-space diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1680-1691, 2023. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 365, + 545, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 545, + 410 + ], + "type": "text", + "content": "[20] Rafat Mantiuk, Kil Joong Kim, Allan G. Rempel, and Wolfgang Heidrich. HDR-VDP-2:a calibrated visual metric for visibility and quality predictions in all luminance conditions. In ACM Siggraph, pages 1–14, 2011. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 411, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 464 + ], + "type": "text", + "content": "[21] Yuzhen Niu, Jianbin Wu, Wenxi Liu, Wenzhong Guo, and Rynson WH Lau. Hdr-gan: Hdr image reconstruction from multi-exposed ldr images with large motions. IEEE Transactions on Image Processing, 30:3885-3896, 2021. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 467, + 545, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 467, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 308, + 467, + 545, + 511 + ], + "type": "text", + "content": "[22] Feiyue Peng, Maojun Zhang, Shiming Lai, Hanlin Tan, and Shen Yan. Deep hdr reconstruction of dynamic scenes. In 2018 IEEE 3rd International Conference on Image, Vision and Computing (ICIVC), pages 347-351. IEEE, 2018. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 512, + 547, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 512, + 547, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 512, + 547, + 567 + ], + "type": "text", + "content": "[23] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 568, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 545, + 612 + ], + "type": "text", + "content": "[24] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1, 2, 5, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 613, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 657 + ], + "type": "text", + "content": "[25] Pradeep Sen, Nima Khademi Kalantari, Maziar Yaesoubi, Soheil Darabi, Dan B Goldman, and Eli Shechtman. Robust patch-based hdr reconstruction of dynamic scenes. ACM Trans. Graph., 31(6):203-1, 2012. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "text", + "content": "[26] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In Proceedings of the 32nd International Conference on Machine Learning, pages 2256-2265, 2015. 1, 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25740" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "text", + "content": "[27] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 2, 5, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 288, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 195 + ], + "type": "text", + "content": "[28] Jou Won Song, Ye-In Park, Kyeongbo Kong, Jaeho Kwak, and Suk-Ju Kang. Selective transhdr: Transformer-based selective hdr imaging using ghost region mask. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVII, pages 288-304. Springer, 2022. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 197, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 287, + 239 + ], + "type": "text", + "content": "[29] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 241, + 287, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 287, + 306 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 287, + 306 + ], + "type": "text", + "content": "[30] Steven Tel, Zongwei Wu, Yulun Zhang, Barthélémy Heyrman, Cédric Demonceaux, Radu Timofte, and Dominique Gingac. Alignment-free hdr deghosting with semantics consistent transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12836-12845, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 308, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 308, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 287, + 352 + ], + "type": "text", + "content": "[31] Anna Tomaszewska and Radoslaw Mantiuk. Image registration for multi-exposure high dynamic range image acquisition. In International Conference in Central Europe on Computer Graphics and Visualization, WSCG'07, 2007. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 354, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 396 + ], + "type": "text", + "content": "[32] Okan Tarhan Tursun, Ahmet Oğuz Akyüz, Aykut Erdem, and Erkut Erdem. An objective dehosting quality metric for HDR images. Comput. Graph. Forum, 35(2):139-152, 2016. 5, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 399, + 287, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 442 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 442 + ], + "type": "text", + "content": "[33] Okan Tarhan Tursun, Ahmet Oğuz Akyüz, Aykut Erdem, and Erkut Erdem. An objective deghosting quality metric for hdr images. In Computer Graphics Forum, pages 139-152. Wiley Online Library, 2016. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 444, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 444, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 444, + 287, + 498 + ], + "type": "text", + "content": "[34] Haohan Wang, Xindi Wu, Zeyi Huang, and Eric P Xing. High-frequency component helps explain the generalization of convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8684-8694, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 500, + 287, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 287, + 543 + ], + "type": "text", + "content": "[35] Yinhuai Wang, Jiwen Yu, and Jian Zhang. Zero-shot image restoration using denoising diffusion null-space model. In The Eleventh International Conference on Learning Representations, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 544, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 578 + ], + "type": "text", + "content": "[36] Greg Ward. Fast, robust image registration for compositing high dynamic range photographs from hand-held exposures. Journal of Graphics Tools, 8, 2012. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 579, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 287, + 634 + ], + "type": "text", + "content": "[37] Jay Whang, Maurizio Delbracio, Hossein Talebi, Chitwan Sahara, Alexandros G Dimakis, and Peyman Milanfar. Deblurring via stochastic refinement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16293-16303, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "text", + "content": "[38] Shangzhe Wu, Jiarui Xu, Yu-Wing Tai, and Chi-Keung Tang. Deep high dynamic range imaging with large foreground motions. In European Conference on Computer Vision (ECCV), 2018. 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[39] Bin Xia, Yulun Zhang, Shiyin Wang, Yitong Wang, Xinglong Wu, Yapeng Tian, Wenming Yang, and Luc Van Gool. Diffir: Efficient diffusion model for image restoration. In" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 531 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13095-13105, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "type": "text", + "content": "[40] Qingsen Yan, Jinqiu Sun, Haisen Li, Yu Zhu, and Yanning Zhang. High dynamic range imaging by sparse representation. Neurocomputing, 269:160-169, 2017. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 130, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 130, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 130, + 545, + 195 + ], + "type": "text", + "content": "[41] Qingsen Yan, Dong Gong, Qinfeng Shi, Anton van den Hengel, Chunhua Shen, Ian Reid, and Yanning Zhang. Attention-guided network for ghost-free high dynamic range imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1751-1760, 2019. 1, 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "type": "text", + "content": "[42] Qingsen Yan, Lei Zhang, Yu Liu, Yu Zhu, Jinqiu Sun, Qinfeng Shi, and Yanning Zhang. Deep hdr imaging via a non-local network. IEEE Transactions on Image Processing, 29: 4308-4322, 2020. 2, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 241, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 296 + ], + "type": "text", + "content": "[43] Qingsen Yan, Weiye Chen, Song Zhang, Yu Zhu, Jinqiu Sun, and Yanning Zhang. A unified hdr imaging method with pixel and patch level. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22211-22220, 2023. 2, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 297, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 297, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 308, + 297, + 545, + 352 + ], + "type": "text", + "content": "[44] Qingsen Yan, Tao Hu, Yuan Sun, Hao Tang, Yu Zhu, Wei Dong, Luc Van Gool, and Yanning Zhang. Towards high-quality hdr deghosting with conditional diffusion models. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2023. 1, 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 354, + 545, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 354, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 308, + 354, + 545, + 418 + ], + "type": "text", + "content": "[45] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728-5739, 2022. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 420, + 545, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 420, + 545, + 474 + ], + "spans": [ + { + "bbox": [ + 308, + 420, + 545, + 474 + ], + "type": "text", + "content": "[46] Xiang Zhang, Tao Hu, Jiashuang He, and Qingsen Yan. Efficient content reconstruction for high dynamic range imaging. In ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7660-7664, 2024. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 475, + 545, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 475, + 545, + 531 + ], + "spans": [ + { + "bbox": [ + 308, + 475, + 545, + 531 + ], + "type": "text", + "content": "[47] Xiang Zhang, Qiang Zhu, Tao Hu, and Qingsen Yan. Eiffhdr: An efficient network for multi-exposure high dynamic range imaging. In ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6560-6564, 2024. 1, 2" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25741" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Generating Enhanced Negatives for Training Language-Based Object Detectors/2970baed-7bcf-45c6-83b6-e79faa1f9b77_content_list.json b/2024/Generating Enhanced Negatives for Training Language-Based Object Detectors/2970baed-7bcf-45c6-83b6-e79faa1f9b77_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6a81a841062c9b633b2d5a6a2649d512baa12224 --- /dev/null +++ b/2024/Generating Enhanced Negatives for Training Language-Based Object Detectors/2970baed-7bcf-45c6-83b6-e79faa1f9b77_content_list.json @@ -0,0 +1,1589 @@ +[ + { + "type": "text", + "text": "Generating Enhanced Negatives for Training Language-Based Object Detectors", + "text_level": 1, + "bbox": [ + 84, + 130, + 883, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shiyu Zhao $^{1,*}$", + "bbox": [ + 215, + 179, + 328, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Long Zhao3", + "bbox": [ + 354, + 180, + 452, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vijay Kumar B G²", + "bbox": [ + 480, + 181, + 627, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yumin Suh2", + "bbox": [ + 656, + 181, + 751, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dimitris N. Metaxas1", + "bbox": [ + 200, + 199, + 369, + 214 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Manmohan Chandraker2,4", + "bbox": [ + 397, + 198, + 602, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Samuel Schulter2", + "bbox": [ + 630, + 198, + 769, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Rutgers University", + "bbox": [ + 104, + 232, + 269, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ NEC Laboratories America", + "bbox": [ + 295, + 233, + 527, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Google Research", + "bbox": [ + 552, + 233, + 702, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 UC San Diego", + "bbox": [ + 728, + 233, + 857, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 285, + 312, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The recent progress in language-based open-vocabulary object detection can be largely attributed to finding better ways of leveraging large-scale data with free-form text annotations. Training such models with a discriminative objective function has proven successful, but requires good positive and negative samples. However, the free-form nature and the open vocabulary of object descriptions make the space of negatives extremely large. Prior works randomly sample negatives or use rule-based techniques to build them. In contrast, we propose to leverage the vast knowledge built into modern generative models to automatically build negatives that are more relevant to the original data. Specifically, we use large-language-models to generate negative text descriptions, and text-to-image diffusion models to also generate corresponding negative images. Our experimental analysis confirms the relevance of the generated negative data, and its use in language-based detectors improves performance on two complex benchmarks. Code is available at https://github.com/xiaofeng94/Gen-Enhanced-Negs.", + "bbox": [ + 75, + 318, + 473, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 640, + 207, + 655 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Using natural language in object detection to describe semantics bears the potential to significantly increase the size of the detector's label space and enable novel applications. While standard detectors operate on a fixed label space [23, 38, 42], natural language allows for a broad spectrum of object descriptions, ranging from generic terms like \"vehicle\" to specific expressions like \"the red sports car parked on the left side\" [12, 17, 25, 30, 41, 53, 54]. Several works advanced language-based object detection over the past few years with novel training strategies [3, 5, 19, 22, 32, 34, 57] and model architectures [11, 15, 33, 45].", + "bbox": [ + 75, + 665, + 468, + 830 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Referring expression or visual grounding datasets [14, 30, 36, 50, 54] provide the natural language object descrip", + "bbox": [ + 76, + 833, + 468, + 863 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f8eef1088602cc2133a650ab43cb86096fa93ebe24f8a1447afcce1c4b32bf07.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 277, + 625, + 347 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5fbc2afc988111bfc60b08b82abf9de8d5b339cae4a76cde954b770724e371cb.jpg", + "image_caption": [ + "Figure 1. The key contribution of our work is to leverage large-language-models and text-to-image diffusion models to automatically generate negative object descriptions and images for training language-based object detectors. In contrast to prior work, our generated negatives are more relevant to the original data and provide a better training signal for detectors." + ], + "image_footnote": [], + "bbox": [ + 627, + 280, + 888, + 460 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tions along with bounding box annotations needed for training. However, this data only describes what is present in the images, but misses to describe what is not. Yet, the notion of negatives is crucial for training discriminative models like language-based detectors [7, 24, 39].", + "bbox": [ + 496, + 582, + 892, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Detection datasets with a fixed label space provide negative classes implicitly or explicitly, with exhaustive [23, 42] or federated [6, 16] annotations, respectively. Any part of an image that does not overlap significantly with a bounding box of category $c$ is verified to not be of that category (for exhaustive annotation). On the other hand, the space of negatives for a free-form text description of an object is extremely large. While some existing datasets provide negative samples in free-form text [43, 46], they were not annotated with bounding boxes. Hence, existing language-based detectors often define the negatives for one object as the descriptions of all other objects in the same image or descriptions of other random samples [3, 11, 19]. However, such negatives may not be directly related to the original positive description and define a weaker training signal (see Fig. 1). By explicitly evaluating on human-curated negatives, a re", + "bbox": [ + 496, + 659, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Part of this work was done during an internship at NEC Laboratories America. Correspondence to: Shiyu Zhao sz553@rutgers.edu", + "bbox": [ + 76, + 875, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13592", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cent benchmark [41] identified a bias of existing language-based detectors to perform clearly better on positive rather than negative descriptions. However, creating a dataset with high-quality human-curated negatives for large-scale training is labor-intensive and costly.", + "bbox": [ + 75, + 90, + 468, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose to explicitly and automatically generate negative data in the form of free-form texts as well as images. Prior works [4, 7, 29, 43, 46, 55] rely on rule-based approaches with knowledge graphs and focus only on the language domain or the classification task. In contrast, we leverage generative large-language-models (LLMs) [35, 47] and text-to-image diffusion models [21, 40] to automatically create relevant but contradicting object descriptions along with the corresponding images for language-based object detection, see Fig. 1.", + "bbox": [ + 75, + 169, + 467, + 319 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given an object description of a dataset, we first use LLMs to generate a semantically contradicting description as the negative. Besides changing individual words (foils) based on explicit knowledge graphs or LLMs, like in prior work [4, 7, 20], we demonstrate improved detection performance with two alternative approaches. (Re-combination): An LLM first identifies all objects in a sentence, and then creates a contradicting one by re-arranging, ignoring or adding objects. (In-context summaries): We prompt an LLM to summarize the differences of a few (less than 100) positive-negative pairs collected from an existing image-level dataset [46]. This summary is then used as context to generate more such examples. Note that we do not need visual input for this step, allowing us to leverage powerful LLMs for semantic and textual reasoning. Moreover, while prior work only focused on the text [4, 7, 29, 43, 55], we also leverage text-to-image diffusion models like GLIGEN [21] to create images that match the generated negative descriptions of objects, which serves as additional training signal. While the direct output of such image-generation models is often noisy and even wrong (not matching the input description), we propose two filtering steps to reduce noise considerably (from $53\\%$ to $16\\%$ according to an empirical study). Having both negative object descriptions and the corresponding image, allows us to improve the discriminative loss for training language-based object detectors.", + "bbox": [ + 75, + 321, + 467, + 714 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our experiments demonstrate clear accuracy gains on two challenging benchmarks, +2.9AP on OmniLabel [41] and +3.3AP on $\\mathrm{D}^3$ [53], when adding our automatically-generated negative data into the training of baseline models like GLIP or FIBER. Moreover, we provide an in-depth analysis of the generated data (text and images) and how they contribute to better language-based detection.", + "bbox": [ + 75, + 717, + 467, + 823 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Summary of contributions: (1) Automatic generation of semantically relevant but contradicting negative text and images with large-scale generative models. (2) Recipes to integrate such negative data into language-based detection models like FIBER [3] and GLIP [19] (3) Clear improve", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ments on language-based detection benchmarks [41, 53] including a thorough analysis of the generated data.", + "bbox": [ + 498, + 90, + 890, + 122 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 500, + 137, + 637, + 152 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision & language localization tasks: Open-vocabulary detection (OVD) requires a model to localize object category names without having seen explicit bounding box annotation for them [5, 7, 15, 52, 56, 59]. In contrast, we focus on the more general language-based object detection task [41, 53], which goes beyond simple category names. Referring expression comprehension (REC) aims at localizing the subject of a free-form text expression. However, REC benchmarks [30, 50, 54] fall short in evaluating all aspects of the more general language-based detection task [41, 53]. In visual grounding (VG) [36], the task is to localize noun phrases of a caption in the image. Although being a task on its own, VG datasets have recently been used mostly as training data for OVD. Our work focuses on general language-based object detection, which subsumes and generalizes standard detection, OVD and REC [12, 25, 41, 53].", + "bbox": [ + 496, + 162, + 890, + 419 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Language-based object detectors: Two critical abilities of language-based detection are accurate localization and tight text-image fusion. Works like [1, 9, 28] use language-models like BERT [2, 27] to align regions extracted from (pre-trained) detectors with captions. The outstanding zero-shot classification accuracy of large-scale pre-trained models like CLIP [37] or [10, 13, 18] then sparked interest in extensions for localization, with different approaches like distillation [5], fine-tuning [15, 33], pseudo-labeling [34, 57, 58], or combinations thereof [3, 19]. We use such models as test bed, but explore the underlying training data with respect to negative samples.", + "bbox": [ + 496, + 425, + 890, + 607 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Negative samples for object detection: The notion of negatives is crucial for training discriminative models [24, 39]. Also for object detection, hard negative mining [44] has proven beneficial for model training. However, these prior works aim to find hard negative training examples rather than negatives in the label space, because the label space is fixed in standard detection. For language-based datasets, the space of potential negatives is extremely large because object descriptions are free-form text. Prior works [4, 29, 43, 46, 55] investigate negative texts for general vision & language models with different strategies, including changing individual words (foil) with rules based on knowledge graphs [31] or with LLMs. Sugar-Crepe [8] shares a similar idea as us to get negative texts with in-context learning but for image-text level pretraining. For language-based detection, [7] explores such rule-based foils, while [20] uses LLMs with specific templates to replace object names with alternative descriptions. In contrast, our work (1) focuses on the localization task, (2) ex", + "bbox": [ + 496, + 613, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13593", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/676445f31ce25fdf5841a8cb2598a6422f82f742803a1a4a1da667cbc67d81fd.jpg", + "image_caption": [ + "Figure 2. In language-based object detection, a detector receives as input an image and a (variable-length) list of free-form text descriptions of objects. For each description, the model predicts bounding boxes for objects that match the description." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 470, + 162 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "plies more comprehensive strategies to generate negatives with LLMs, and (3) proposes to also generate corresponding negative images with text-to-image diffusion models.", + "bbox": [ + 75, + 256, + 468, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 316, + 166, + 330 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Language-based object detection", + "text_level": 1, + "bbox": [ + 76, + 340, + 364, + 357 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Task definition: Given an image and a list of object descriptions, the task is to output bounding boxes along with confidence scores for each description, as shown in Fig. 2. Note the multi-label setting where one object instance can be referred to by multiple descriptions, like \"person\" and \"person looking at book\". Also note that an object description might refer to zero objects in the image and the desired output is an empty set of boxes.", + "bbox": [ + 75, + 364, + 468, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Training data: Many language-based detection models [3, 19] use a combination of object detection [23, 42] and visual grounding datasets [14, 36] for training their models. Both types of datasets provide images $I$ and bounding boxes $b_{l}$ to localize individual objects. Object detection data assigns each bounding box $b_{l}$ a unique category $c$ out of fixed label space $\\mathcal{C}$ . The exhaustive labeling of the fixed label space in detection datasets implies the space of negatives. An object of category $c$ is not any of the categories $\\mathcal{C} \\setminus c$ . On the other hand, grounding data provides an image caption $t$ in free-form text, where subsets of words $m_{l}$ (defined as indices of starting and ending characters in $t$ ) are linked to bounding box $b_{l}$ . For grounding data, the space of negatives is extremely large because one can find as many textual descriptions that do not match $t$ as desired due to the compositionality of free-form text. Many language-based detectors [3, 19] only use the words in $t$ that are not referred to by $m_{l}$ as negatives for the bounding box $b_{l}$ . We argue that this choice is sub-optimal because these words may refer to entirely different objects and are easy to discriminate. In the following section, we explain how we can automatically generate negative samples that semantically related to the original text $t$ and hence provide a better training signal.", + "bbox": [ + 75, + 489, + 470, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Generating negative samples", + "text_level": 1, + "bbox": [ + 76, + 847, + 333, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to automatically and explicitly generate negative samples based on the original text descriptions $t$ to im", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "prove the training signal for language-based detectors. A key observation of our work is to leverage the vast knowledge encoded into large-language models (LLMs) [35, 47] and text-to-image diffusion models (T2I) [21, 40]. Besides proposing novel ways to instruct LLMs for generating negative text descriptions (Sec. 3.2.1), we also propose to generate negative images (Sec. 3.2.2).", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.1 Generating negative descriptions with LLMs", + "text_level": 1, + "bbox": [ + 498, + 220, + 864, + 234 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given an object description $t$ that matches the visual content inside a bounding box $b_{l}$ , we define a \"negative\" description $t'$ as any text that is semantically different to the original text. Furthermore, our intuition is that good negative descriptions are still semantically related to the original description, but not the same. An example is: \"Person in red shirt\" as the original description and \"Person in blue shirt\" as a contradicting negative one.", + "bbox": [ + 496, + 247, + 890, + 367 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Prior work [4, 7, 43] explored rule-based approaches to generate negative text. However, such rules are typically limited to simple knowledge graphs and are limited to replacing only individual words, often just nouns, or swapping words. In contrast, we explore more powerful LLMs to automatically generate relevant negatives. To make the negative text generation efficient and economic, in all cases, we first leverage a strong instruction-tuned LLM [35] to generate $50\\mathrm{k}$ positive-negative pairs, and then finetune a LLaMA-7B [47] model with those pairs to then generate negative captions on large grounding datasets. In the following, we describe three ways to instruct an LLM for generating positive-negative pairs of object descriptions:", + "bbox": [ + 496, + 369, + 892, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LLM-based foils: We first prompt an instruction-tuned LLM [35] to find concepts (i.e., objects, attributes and relationships) in object descriptions. Compared to rule-based parsers [51], LLMs can provide richer information. For example, for the caption \"A transportation vehicle is carrying a crowd of people who are sitting and standing,\" the parser ignores \"sitting\" and \"standing\", while LLMs regard them as attributes. Then, we pick one concept from the first step sequentially and prompt LLMs again to generate a negative caption by changing the concept. For both steps, the prompts are manually curated with the task definition and step-by-step instructions for the generation. Please find the exact prompt for the LLM in the supplement.", + "bbox": [ + 496, + 574, + 890, + 771 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Re-combination: Next, we give the LLM more freedom in generating negative descriptions. We first prompt the LLM to identify all objects in the original caption, and then to recombine them to create a new sentence different from the original one. We allow the LLM to ignore, change or add new objects. For example, given the caption “A boy is playing with his dog” and two objects “boy” and “dog”, the LLM can output “The girl and her dog are playing fetch in", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13594", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the park\". Detailed prompts for both identifying objects and re-combination are in the supplement.", + "bbox": [ + 76, + 90, + 470, + 121 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In-context summary: Third, we enable LLMs to learn how to generate negative descriptions by providing human-annotated positive-negative pairs as in-context samples. We randomly sample 80 pairs of positive and negative texts from the Winoground dataset [46] and prompt the instruction-tuned LLM [35] to summarize the difference of those pairs in plain text. Then, instead of manually creating prompts to generate positive-negative pairs, we leverage the summary together with three randomly sampled Winoground pairs as prompts to the LLM, and generate several positive-negative pairs to finetune LLaMA. After finetuning, the LLaMA model is used to generate negative texts for given descriptions. This pipeline does not require handcrafted prompts to LLMs as the explanation of the concept of negatives and how to create them. The supplement contains full prompts for generating a summary, and generating positive-negative pairs for finetuning.", + "bbox": [ + 75, + 130, + 472, + 387 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2 Generating negative images with T2I models", + "text_level": 1, + "bbox": [ + 76, + 411, + 441, + 426 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given an original image $I$ , a bounding box $b$ and a corresponding object description $t$ , we define a negative image $I'$ as any image that has a different semantic content inside $b$ . The rest of the image can be equivalent to $I$ . To obtain such imagery, we start with visual grounding data that provide bounding boxes, positive captions with text phrases, and alignment between them. We propose a two-step process: First, we turn the positive caption into a negative one. Second, we use conditioned image generation tools to alter the visual content inside the bounding box $b$ .", + "bbox": [ + 75, + 436, + 468, + 588 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Negative text for negative images: Although we have already described an approach to generate negative descriptions in Sec. 3.2.1, doing so to generate a negative image requires a different approach. In this case, the generated negative text needs to preserve the alignment $m_{l}$ to the ground truth bounding box $b_{l}$ in order to instruct the generative image model GLIGEN [21]. Hence, we first select a bounding box $b_{l}$ and mask out the corresponding words (known via $m_{l}$ ) in the text $t$ . For example, \"A boy is playing with his dog\" turns into \"A boy is playing with [Mask]\" if the selected bounding box refers to \"his dog\". Again, we leverage LLMs [35] to fill in text for \"[Mask]\" to generate a negative text without reusing the original text. Please refer to Fig. 3 for illustrations.", + "bbox": [ + 75, + 597, + 468, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We finetune a LLaMA-7B for the mask filling task with triplets of positive texts, masked texts, and negative texts. To reduce manual efforts, we follow the approach of in-context summary to get triplet samples. We apply this process twice: We start with only 5 manually created triplets to build a summary and generate 100 samples from the", + "bbox": [ + 75, + 810, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/03da1533848a52141bebfc703c084b2abdf51c14c1354ec0fbb83913137aeac8.jpg", + "image_caption": [ + "Figure 3. Overview of using LLMs [35, 47] and text-to-image diffusion models [21, 40] to generate negative images." + ], + "image_footnote": [], + "bbox": [ + 500, + 90, + 890, + 251 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "LLM [35] with human checks. We then repeat the process to generate 50k examples without human checks from a summary of the 100 generated examples. This increases diversity in the generated data.", + "bbox": [ + 496, + 316, + 890, + 377 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Conditional image generation: Given an image $I$ , a bounding box $b$ and the altered text $t'$ , we generate a negative image $I'$ that is equal to $I$ except inside $b$ , where the visual content is altered to match the text $t'$ . To do so, we use the inpainting and conditioning abilities of GLIGEN [21], a T2I model [40]. Refer to [21] and our supplemental material for more details, and to Fig. 3 for an illustration of the process.", + "bbox": [ + 496, + 380, + 892, + 502 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Mitigating noise in image generation: We found that the generated images are often noisy for any of the following reasons: (1) The altered text refers to a big bounding box that covers other smaller boxes. Large portions of the image are then generated and often do not match the concepts those smaller boxes originally covered. (2) The generated negative text does not match the bounding box that is either too small, too large or in an inappropriate position. (3) The T2I model fails to understand the negative text and generates wrong content. We propose two steps to filter such noisy images. First, we simply ignore ground truth boxes $b_{l}$ for image generation if the box covers more than $75\\%$ of any other boxes in the image. Second, we adopt CLIP [37] to verify the semantic similarity of the generated image regions and the corresponding text. Specifically, we compute the similarity with CLIP between the generated image region (visual input) and the original and generated negative texts (text input). We filter out generated images that have a similarity score to the generated negative text lower than a user-defined threshold. Details on the filtering steps are given in the supplemental.", + "bbox": [ + 496, + 505, + 893, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Learning from negative samples", + "text_level": 1, + "bbox": [ + 500, + 830, + 782, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Detector design and training objective: The generated data does not prescribe any specific architecture for the detector. A common choice, which we also use for our exper", + "bbox": [ + 496, + 854, + 892, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13595", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/81aba8dd6003f4b8376745bf454bc4710d45a6a7309deccd0ccb3d7788986279.jpg", + "image_caption": [ + "Figure 4. Illustration of the grounding loss used for training. Predictions that are matched with ground truth receive a positive signal from the associated text (tall rectangles). All other words receive a negative signal (short rectangles). The top left quarter shows the original loss used in [3, 19]. The other three quarters are related to our proposed generated negative data and provide additional training signals." + ], + "image_footnote": [], + "bbox": [ + 76, + 85, + 472, + 234 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "iments, is [3, 19]. The inputs are image $I$ and text $t$ , and the output is a set of bounding boxes $\\hat{b}_i$ with corresponding logits $\\hat{p}_i \\in \\mathbb{R}^T$ . Here, $T$ is the number of tokens required to represent text $t$ . The ground truth can be represented by a binary assignment matrix $\\mathbf{A} \\in \\mathbb{B}^{L \\times T}$ . Rows refer to ground truth boxes $l$ and columns to tokens in $t$ . Each element indicates if a token corresponds to a box $l$ , which is given by the ground truth indices $m_l$ . To define a loss, bipartite graph matching associates predictions with ground truth. For matched predictions, the target vector $g_i \\in \\mathbb{B}^T$ is the corresponding row from $\\mathbf{A}$ , while it is an all-zero vector for unmatched targets. The loss is then computed as $\\mathcal{L} = \\sum_{i} \\ell_{\\mathrm{FL}}(\\hat{p}_i, g_i)$ where FL refers to a focal loss. Fig. 4 illustrates the loss.", + "bbox": [ + 75, + 369, + 472, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Integrating negative text: When sampling an image $I$ , along with text $t$ , boxes $b_{l}$ and indices $m_{l}$ , we also randomly sample $K > 1$ negative descriptions from $\\{t_j^{\\prime}\\}$ that defines the pool of negatives generated for text $t$ . We randomly shuffle the order of all texts to avoid any biases on the location of the one positive description, and then concatenate them into one text string.", + "bbox": [ + 75, + 589, + 468, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Integrating negative images: We explore two options: (1) Simply add the generated images $I'$ along with their generated (but semantically matching) captions $t'$ as additional visual grounding data. The original caption $t$ , which was the starting point to generate the negative image $I'$ , is now used as the negative caption. In this way, both the original image $I$ and the generated one $I'$ have positive and negative descriptions. This option is illustrated in Fig. 4. (2) To better leverage the relation between the original and generated data, the second option is to pack them into a single training sample. We simply concatenate the images $I$ and $I'$ , as well as the texts $t$ and $t'$ . The ground truth information $m_{l}$ is updated accordingly. See supplement for details.", + "bbox": [ + 75, + 704, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/d83d0267871e005ff781cc4b01aad084b9aac46e644fcb844c7d97e9acd7e45f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
OmniLabelOmniLabel-Negative
APAPcAPdAPdPAPAPcAPdAPdP
Detic [60]8.015.65.48.0----
MDETR [11]--4.79.1----
GLIP-T [19]19.323.616.425.813.924.89.626.1
+ Ours22.227.218.829.016.528.611.630.2
FIBER-B [3]25.730.322.334.818.731.213.336.3
+ Ours28.132.125.136.522.333.316.738.3
", + "bbox": [ + 500, + 88, + 888, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Evaluation on the OmniLabel [41] benchmark.", + "bbox": [ + 529, + 229, + 859, + 243 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 277, + 632, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental design", + "text_level": 1, + "bbox": [ + 500, + 305, + 694, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training procedure: We choose two recent methods, GLIP-T [19] and FIBER-B [3], to demonstrate the effect of our automatically generated negatives. We use the official code and publicly available checkpoints as a starting point. The Flickr30k dataset [36] serves as our grounding dataset to generate the negative data. We then fine-tune GLIP-T and FIBER-B with both positive and negative data, along with the Objects365 detection dataset [42] for 1 epoch. Note that both Objects365 and Flickr30k are part of the original training set. We do not introduce any extra data except our generated negatives. Most hyper-parameters are equal to the original settings of GLIP and FIBER. Any exceptions are described in the supplement.", + "bbox": [ + 496, + 333, + 890, + 529 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation benchmarks: We choose two recently proposed benchmarks, OmniLabel [41] and $\\mathrm{D}^3$ [53], as our test beds. These benchmarks evaluate more aspects of language-based detection than existing referring expressions [30, 50, 54] or open-vocabulary detection [6, 17] benchmarks. Specifically, both benchmarks contain complex object descriptions that go beyond simple category names from open-vocabulary detection benchmarks. Moreover, the descriptions can refer to zero, one or multiple instances in the image, in contrast to standard referring expression benchmarks. These properties enable a more stringent evaluation metric as in object detection, which is based on average precision (AP) in both OmniLabel [41] and $\\mathrm{D}^3$ [53]. Both benchmarks provide more fine-grained metrics. OmniLabel evaluates separately for categories, descriptions, and descriptions referring to at least one object, with APc, APd and APd-P, respectively. $\\mathrm{D}^3$ differentiates descriptions on absence (\"Abs\") and presence (\"Pres\") that indicate whether or not they contain any form of negation (e.g., \"without\"), as well as on text lengths. Finally, we create a specific split for OmniLabel, \"OmniLabel-Negative\", to evaluate the model only on images that contain at least one negative description (i.e., not referring to any object).", + "bbox": [ + 496, + 553, + 892, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13596", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/520bd7f1b99379a7c6cef0974e7fc205dbc17ecb7225e11ba1b949f0adc66108.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
D3(default)D3(by length of texts)
FullPresAbsSMLXL
OFA-L [49]4.24.14.64.95.43.02.1
OWL-ViT-L [33]9.610.76.420.79.46.05.3
G-DINO-B [26]20.720.122.522.622.518.916.5
OFA-DOD [53]21.623.715.423.622.620.518.4
GLIP-T [19]19.118.321.522.422.016.610.6
+ Ours21.420.623.728.124.517.411.5
FIBER-B [3]22.721.526.030.125.917.913.1
+ Ours26.025.228.135.529.720.514.2
", + "bbox": [ + 80, + 88, + 467, + 244 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Evaluation on the ${\\mathrm{D}}^{3}$ [53] benchmarks.", + "bbox": [ + 127, + 253, + 415, + 268 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Benchmark comparisons", + "text_level": 1, + "bbox": [ + 76, + 294, + 305, + 311 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tabs. 1 and 2 evaluate the impact of our generated negative training data on the OmniLabel [41] and $\\mathrm{D}^3$ [53] benchmarks. In both tables, the first set of rows are baselines provided by the benchmarks. The following rows show the main comparisons for GLIP-T [19] and FIBER-B [3] with and without adding our generated negative training data. First, we can see that adding negative data improves results across all metrics for both models and both benchmarks. On OmniLabel, we can see a $+2.9\\%$ and $+2.4\\%$ increase in AP for GLIP-T and FIBER-B, respectively. Similarly, we observe a $+2.3\\%$ and $+3.3\\%$ increase in the main metric of $\\mathrm{D}^3$ (AP on full descriptions) for GLIP-T and FIBER-B.", + "bbox": [ + 75, + 318, + 472, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Analysis on negative texts", + "text_level": 1, + "bbox": [ + 76, + 510, + 312, + 526 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effectiveness of different negative texts: We finetune FIBER-B without and with different kinds of negative texts mentioned in Sect. 3.2.1, i.e., Rule-based foils, LLM-based foils, Re-combination with LLMs, In-context summary with LLMs, and present results in Table 3. We find all kinds of negatives improve the original FIBER-B on both OmniLabel and $\\mathrm{D}^3$ benchmarks. Negative texts from LLMs generally achieve better results compared to LLM-based foils, which indicates that LLMs are powerful tools for negative text generation. Moreover, both recombination and in-context summary with LLMs outperform LLM-based foils in all metrics except APd-P. Note that APd-P refers to evaluations without negative label spaces, which is a task weaker than language-based detection. Based on such results, we argue that although word foils provide promising results in traditional studies [7, 43], it is sub-optimal to LLMs. We need to explore varied ways to unlock the ability of LLMs. We believe that our two solutions, i.e., Re-combination and In-context summary, provide a good starting point for future studies. Besides using only one kind of negative texts, we also explore the combinations of different kinds of negative texts in the supplement.", + "bbox": [ + 75, + 532, + 472, + 866 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Diversity of rule-based and LLM-based negatives: In this part, we investigate the diversity of different negative", + "bbox": [ + 76, + 869, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c0fb8d5c75d1f665d231f42d5e6db5374b87fda6c384f5ba7d1e78e28c3e962a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 92, + 700, + 170 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f19c8c37a648ad25643a81bb76ce8a85202db15430132966d7e379191708f518.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 704, + 93, + 883, + 170 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/32a8a4eb4a89f6e7d559893c1f8676b7156b60b9ce9aab9af5aad8c81b042070.jpg", + "image_caption": [ + "Figure 5. Percentage of negative texts with the numbers of words. Four negative generation methods are compared." + ], + "image_footnote": [], + "bbox": [ + 508, + 172, + 700, + 251 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c98dec11641bcb54e462f9f9cd49fc721e1e2963baeff1ca563ac65680933579.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 704, + 172, + 883, + 251 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/22fdd28b65468b385aad3e33355b3b02e56efc81147ed3a405912ccb3340ede2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 316, + 699, + 393 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/40bc52d2ee23dc901a4dd97c95be0116dec585d402ef605139164ea487e7217f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 704, + 316, + 883, + 393 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/aada24c5ca7f4a935dfd3425784770fda8e248bdfdb3d8180c83505f116591bf.jpg", + "image_caption": [ + "Figure 6. Percentage of negative texts with the numbers of words that are different from the original caption. Four negative generation methods are compared." + ], + "image_footnote": [], + "bbox": [ + 508, + 398, + 699, + 474 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a0f704dafcd969b164daf0ee64cb64dc4f62c227f26e5f8adbf6aa4dd4d2cfc0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 704, + 398, + 883, + 474 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9aa5fb75fda56cd538964bc03f29b59fb6de06347c09a149c787da01948a6bdf.jpg", + "image_caption": [ + "Figure 7. Average numbers of extra unique words per thousand generated negative texts, which are not included in the original dataset. We group words by their part-of-speech." + ], + "image_footnote": [], + "bbox": [ + 555, + 556, + 831, + 656 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "texts. First, we count number of words for each negative text and provide the distribution for negatives of different sources in Fig. 5. As shown, all four distributions have a peak around 10 words, but the one of rule-based foils is higher than others. That means rule-based foils provide more negative texts with similar lengths.", + "bbox": [ + 496, + 732, + 890, + 823 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Second, we count the number of different words between the original positive caption and the negative caption, and present the distributions in Fig. 6. We find that LLM-based methods usually changes more words than rule-based foils, which increases the diversity. Moreover, in-context sum", + "bbox": [ + 496, + 825, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13597", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/fec34ce9ad9ad02f1cdaf9129f6ca4984aaaa254993f9701b3078956c1451a4d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Whole OmniLabelOmniLabel-NegativeD3
APAPcAPdAPd-PAPAPcAPdAPd-PFullPresAbs
Original FIBER-B25.730.322.334.818.731.213.336.322.721.526.0
+ Rule-based foils26.431.722.634.919.232.613.636.424.123.226.9
+ LLM-based foils26.530.723.335.920.832.115.438.024.624.026.5
+ Re-combination26.930.823.935.921.132.315.637.625.324.627.3
+ In-context summary26.630.823.434.221.132.215.736.425.725.227.5
", + "bbox": [ + 80, + 89, + 566, + 200 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/557e9294ef85422db01b54e4baed5c237eb0a08245ea5b40555191b296d139e0.jpg", + "table_caption": [ + "Table 3. Performance of FIBER-B trained with negative texts from four negative generation methods." + ], + "table_footnote": [], + "table_body": "
OmniLabelD3
APcAPdAPd-P
Original FIBER30.322.334.822.7
FIBER w/ neg. texts30.723.935.525.9
+ W/ neg. img. directly30.122.433.723.0
+ Box filter31.023.835.423.6
+ Box&CLIP filters31.124.235.924.1
+ Above + concat. img.31.724.835.924.8
+ Above + weight ensemble32.125.236.526.0
", + "bbox": [ + 586, + 90, + 888, + 229 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. FIBER trained with negative images.", + "bbox": [ + 599, + 239, + 872, + 253 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/245a9dd89cbd49cca57c28c1b3f7df453277347f71c5dc4dd4831db6e614e365.jpg", + "image_caption": [ + "Figure 8. Noisy generated images. The orange box contains the red box, and editing the orange changes the red unexpectedly." + ], + "image_footnote": [], + "bbox": [ + 127, + 277, + 421, + 439 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "mary has a more flat distribution compared to others. Probably, in-context summary learns how to generate negatives automatically from data and has less restrictions. Besides, in-context summary has more cases with no word changed where negative texts are generated by just shuffling words or concepts in the original text. Such shuffling is a common pattern of Winoground [46], and our in-context summary can learn such data specific patterns.", + "bbox": [ + 75, + 503, + 468, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Third, we count how many extra words that does not exist in the original Flickr30k dataset are introduced in different negative generation methods. Fig. 7 shows the average number of extra words per 1000 negative texts. We group words into four part-of-speech categories, i.e., VERB, NOUN, ADP/ADJ, and others. As shown, LLMs introduce more extra words on average than rule-based foils probably because rule-based foils are limited in a predefined set of words. However, LLMs are open to any concepts and have great potentials of generating diverse texts. In-context summary introduces the most extra words for all categories, which is likely a benefit of learning negative generation from data. The above statistics indicate a clear view that LLMs generate more diverse data than rule-based foils.", + "bbox": [ + 75, + 626, + 468, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Analysis on negative images", + "text_level": 1, + "bbox": [ + 76, + 847, + 328, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Noise in generated images: As mentioned in the last paragraph of Sect. 3.2.2, the raw generated images are noisy in", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "several ways. First, the editing of a large box will override the context of smaller boxes that are covered by the large box. As shown in Fig. 8, GLIGEN did follow the instruction to generate a boy in the orange box, but the black jacket in the red box is missing. As a remedy, we apply our first de-noise step \"Box Filter\". That is, we ignore boxes that contain any other boxes when generating negative images. Second, GLIGEN may generate contents with wrong attributes or objects, as shown in Fig. 9 (Left). Moreover, our generation pipeline includes some cases where the edited text and the bounding box does not match. As shown in Fig. 9 (Right), the box for \"his lap\" cannot be modified as \"his knees\". Thus, GLIGEN generates wrong contents. As described in Sect. 3.2.2, we adopt a pretrained CLIP model to judge if generated contents are correct, which mitigates the noise to some extent. As shown in Fig. 9, both negative images get low CLIP scores and can be filtered out with a threshold. We call such thresholding \"CLIP Filter\".", + "bbox": [ + 496, + 280, + 892, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Subject studies on Box and CLIP filters: We employ human experts to check the amount of noisy generated images. First, for negative images w/o filter, w/ Box filter, and w/ Box&CLIP filters, we separately and randomly select 100 samples. Then, we ask two experts to check if a negative image is not noisy by comparing it with its caption and the original positive image. We regard an image as not noisy when both experts agree. As shown in Fig. 10, both filters reduce the noise. The Box filter improves from $47\\%$ to $63\\%$ , and the CLIP filter improves to $84\\%$ .", + "bbox": [ + 496, + 560, + 890, + 710 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of generated negative images: To show the effectiveness of generated images themselves, we take captions of generated images as additional negative texts to in-context summary, and finetune a FIBER model as baseline. Then, we compare the baseline with variants of adding generated negative images in Table 4. As shown, the performance drops if we directly take raw negative images as new visual grounding data without any filters (i.e., W/ neg. img. directly). Probably, there are too much noise in raw negative images as shown in Fig. 10. When applying both Box and CLIP filters on negative images, we can achieve slight improvement on OmniLabel compared to using neg-", + "bbox": [ + 496, + 719, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13598", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/23d9566b268ef8858588220de2689f124b5bc29f26ad20625525994e21a72dc5.jpg", + "image_caption": [ + "Original", + "Positive: A woman in a black mesh skirt plays acoustic guitar", + "Negative: A woman in a white lace dress plays acoustic guitar" + ], + "image_footnote": [], + "bbox": [ + 119, + 114, + 297, + 219 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/853c0d3bb4871d4e8bb3999eac24b1ba30064c504d1b38b65e7ea4b55255f775.jpg", + "image_caption": [ + "Generating wrong attributes", + "Generated", + "Positive: A man is sitting on a chair with his hands above his lap", + "Negative: A man is sitting on a chair with his hands above his knees" + ], + "image_footnote": [], + "bbox": [ + 300, + 114, + 478, + 219 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d6c9a61e13381799fe4271559890e473fe0136d9313c276f643b6b34624b3e78.jpg", + "image_caption": [ + "Generation with inappropriate boxes or texts", + "Original" + ], + "image_footnote": [], + "bbox": [ + 491, + 114, + 669, + 219 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d4ced8cb2308daaebd82cdfa84be84dd766c36a1cbfccdf827327ecddd3d0564.jpg", + "image_caption": [ + "Generated" + ], + "image_footnote": [], + "bbox": [ + 671, + 114, + 851, + 219 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/173612794985e0903a31e2b35b3de7826df42b76b2af7e84e7b62e54a5b38c97.jpg", + "image_caption": [ + "Figure 9. Left: Noisy negative images due to wrong attributes or objects generated by text-to-image models. Right: Noisy negative images caused by inappropriate bounding boxes or negative texts from LLMs. CLIP scores of generated images refer to the similarity between the box and the negative text compared to the positive text. Thresholding on CLIP scores remove those noisy images.", + "Figure 10. Percentage of good generated negative images." + ], + "image_footnote": [], + "bbox": [ + 130, + 333, + 410, + 402 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7676b5cf758c51a2742d1dc7942bb7ce5cc5a26580ce107b1fd93fd31331f902.jpg", + "image_caption": [ + "Figure 11. Distributions of image regions from Omnilabel, $\\mathbf{D}^3$ , and our generated images. Visualization with t-SNE." + ], + "image_footnote": [], + "bbox": [ + 94, + 439, + 468, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ative texts only.", + "bbox": [ + 76, + 664, + 183, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Concatenating images during training: Following the idea of concatenating the positive and negative captions as the text input, we concatenate the positive and negative images as one input image during training. See supplement for an example. In this way, models are forced to tell the difference between the positive and negative images within one training iteration, which helps detectors to learn better about the negative. As shown in Table 4, such a simple technique improves upon “+ Box&CLIP filters” both on OmniLabel and $\\mathbf{D}^{3}$ . Furthermore, we ensemble the weights of two FIBER models, one finetuned with negative texts only, and the other finetuned with both negative texts and images. Finally, compared to using negative texts only, we gain 1.3 APd on OmniLabel and no performance drop on $\\mathbf{D}^{3}$ .", + "bbox": [ + 75, + 688, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Looking into generated images and benchmarks: Table 4 shows that negative images help on OmniLabel but not much on $\\mathrm{D}^3$ . We explore this on a data basis. We first crop image regions for generated images, OmniLabel images, $\\mathrm{D}^3$ images, and Flickr30k images based on the bounding boxes. Then, we randomly select 1000 image regions and feed them into a CLIP image encoder to get CLIP embeddings. Later, we input those embeddings to t-SNE [48] to illustrate the similarities between different image regions. As shown in Fig. 11, $\\mathrm{D}^3$ 's regions are grouped into several clusters, while OmniLabel and our generated regions are scattered in the center. This indicates that there is a clear domain gap between $\\mathrm{D}^3$ and the others. Thus, it is plausible that our generated images only helps on OmniLabel. In our view, the gap comes from that $\\mathrm{D}^3$ collect data in groups based on categories. In contrast, OmniLabel collects data randomly.", + "bbox": [ + 496, + 330, + 890, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 602, + 617, + 617 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Language-based detection requires localization of objects by a referring free-form text descriptions. To train accurate models in a discriminative way, the training data must contain good negative samples. Starting with an existing dataset, we propose (1) novel ways to prompt LLMs for generating additional negative texts, and (2) generating negative images to complement the training signal. Based on our experimental evaluations, we conclude that such additional negative training data indeed translates into improved detection accuracy on standard benchmarks. Our analysis demonstrates the importance of diversity in the generated text, which is higher with our approach than with prior works, and the quality of the generated images, which our proposed filtering steps can significantly increase.", + "bbox": [ + 496, + 628, + 890, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments: This research project has been partially funded by research grants to Dimitris N. Metaxas through NSF: 2310966, 2235405, 2212301, 2003874, and FA9550-23-1-0417.", + "bbox": [ + 496, + 840, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13599", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. UNITER: UNiversal Image-TExt Representation Learning. In ECCV, 2020. 2", + "[2] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019. 2", + "[3] Zi-Yi Dou, Aishwarya Kamath, Zhe Gan, Pengchuan Zhang, Jianfeng Wang, Linjie Li, Zicheng Liu, Ce Liu, Yann LeCun, Nanyun Peng, et al. Coarse-to-Fine Vision-Language Pretraining with Fusion in the Backbone. In NeurIPS, 2022. 1, 2, 3, 5, 6", + "[4] Sivan Doveh, Assaf Arbelle, Sivan Harary, Eli Schwartz, Roei Herzig, Raja Giryes, Rogerio Feris, Rameswar Panda, Shimon Ullman, and Leonid Karlinsky. Teaching Structured Vision & Language Concepts to Vision & Language Models. In CVPR, 2023. 2, 3", + "[5] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary Object Detection via Vision and Language Knowledge Distillation. In ICLR, 2022. 1, 2", + "[6] Agrim Gupta, Piotr Dólár, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In CVPR, 2019. 1, 5", + "[7] Ryota Hinami and Shin'ichi Satoh. Discriminative Learning of Open-Vocabulary Object Retrieval and Localization by Negative Phrase Augmentation. In EMNLP, 2018. 1, 2, 3, 6", + "[8] Cheng-Yu Hsieh, Jieyu Zhang, Zixian Ma, Aniruddha Kembhavi, and Ranjay Krishna. Sugarcrepe: Fixing hackable benchmarks for vision-language compositionality. In NeurIPS, 2023. 2", + "[9] Ronghang Hu and Amanpreet Singh. UniT: Multimodal Multitask Learning with a Unified Transformer. In ICCV, 2021. 2", + "[10] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, and Tom Duerig. Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision. In ICML, 2021. 2", + "[11] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. MDETR - Modulated Detection for End-to-End Multi-Modal Understanding. In ICCV, 2021. 1, 5", + "[12] Aishwarya Kamath, Sara Price, Jonas Pfeiffer, Yann LeCun, and Nicolas Carion. TRICD: Testing Robust Image Understanding Through Contextual Phrase Detection. Technical report, NYU, 2023. 1, 2", + "[13] Zaid Khan, Vijay Kumar B.G., Xiang Yu, Samuel Schulter, Manmohan Chandraker, and Yun Fu. Single-Stream Multi-Level Alignment for Vision-Language Pretraining. In ECCV, 2022. 2", + "[14] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV, 123:32-73, 2017. 1, 3" + ], + "bbox": [ + 78, + 114, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Weicheng Kuo, Yin Cui, Xiuye Gu, AJ Piergiovanni, and Anelia Angelova. F-vlm: Open-vocabulary object detection upon frozen vision and language models. In ICLR, 2023. 1, 2", + "[16] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, Tom Duerig, and Vittorio Ferrari. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. IJCV, 2020. 1", + "[17] Chunyuan Li, Haotian Liu, Liusan Harold Li, Pengchuan Zhang, Jyoti Aneja, Jianwei Yang, Ping Jin, Houdong Hu, Zicheng Liu, Yong Jae Lee, and Jianfeng Gao. ELEVATER: A Benchmark and Toolkit for Evaluating Language-Augmented Visual Models. In NeurIPS, 2022. 1, 5", + "[18] Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Deepak Gotmare, Shafiq Joty, Caiming Xiong, and Steven Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In NeurIPS, 2021. 2", + "[19] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In CVPR, pages 10965-10975, 2022. 1, 2, 3, 5, 6", + "[20] Liunian Harold Li, Zi-Yi Dou, Nanyun Peng, and Kai-Wei Chang. DesCo: Learning Object Recognition with Rich Language Descriptions. In NeurIPS, 2023. 2", + "[21] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In CVPR, 2023. 2, 3, 4", + "[22] Chuang Lin, Peize Sun, Yi Jiang, Ping Luo, Lizhen Qu, Gholamreza Haffari, Zehuan Yuan, and Jianfei Cai. Learning object-language alignments for open-vocabulary object detection. In ICLR, 2023. 1", + "[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C. Lawrence Zitnick. Microsoft COCO: Common Objects in Context. In ECCV, 2014. 1, 3", + "[24] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal Loss for Dense Object Detection. In ICCV, 2017. 1, 2", + "[25] Chang Liu, Henghui Ding, and Xudong Jiang. GRES: Generalized Referring Expression Segmentation. In CVPR, 2023. 1, 2", + "[26] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 6", + "[27] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach, 2019. 2", + "[28] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. ViL-BERT: Pretraining Task-Agnostic Visiolinguistic Representations for Vision-and-Language Tasks. In NeurIPS, 2019. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13600", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Zixian Ma, Jerry Hong, Mustafa Omer Gul, Mona Gandhi, Irena Gao, and Ranjay Krishna. CREPE: Can Vision-Language Foundation Models Reason Compositionally? In CVPR, 2023. 2", + "[30] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L. Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In CVPR, 2016. 1, 2, 5", + "[31] George A. Miller. Wordnet: A lexical database for english. Commun. ACM, 38(11):39-41, 1995. 2", + "[32] Zhixiang Min, Bingbing Zhuang, Samuel Schulter, Buyu Liu, Enrique Dunn, and Manmohan Chandraker. Neurocs: Neural nocs supervision for monocular 3d object localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21404-21414, 2023. 1", + "[33] Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, Xiao Wang, Xiaohua Zhai, Thomas Kipf, and Neil Houlsby. Simple Open-Vocabulary Object Detection with Vision Transformers. In ECCV, 2022. 1, 2, 6", + "[34] Matthias Minderer, Alexey Gritsenko, and Neil Houlsby. Scaling Open-Vocabulary Object Detection. In NeurIPS, 2023. 1, 2", + "[35] OpenAI. Gpt-3.5. 2022. 2, 3, 4", + "[36] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k Entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In ICCV, pages 2641–2649, 2015. 1, 2, 3, 5", + "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 4", + "[38] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks. In NeurIPS, 2015. 1", + "[39] Joshua Robinson, Ching-Yao Chuang, Suvrit Sra, and Stefanie Jegelka. Contrastive Learning with Hard Negative Samples. In ICLR, 2021. 1, 2", + "[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3, 4", + "[41] Samuel Schulter, Vijay Kumar BG, Yumin Suh, Konstantinos M Dafnis, Zhixing Zhang, Shiyu Zhao, Dimitris Metaxas, et al. OmniLabel: A Challenging Benchmark for Language-Based Object Detection. In ICCV, 2023. 1, 2, 5, 6", + "[42] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Jing Li, Xiangyu Zhang, and Jian Sun. Objects365: A Large-scale, High-quality Dataset for Object Detection. In ICCV, 2019. 1, 3, 5", + "[43] Ravi Shekhar, Sandro Pezzelle, Yauhen Klimovich, Aurelie Herbelot, Moin Nabi, Enver Sangineto, and Raffaella" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Bernardi. FOIL it! Find One mismatch between Image and Language caption. In ACL, 2017. 1, 2, 3, 6", + "[44] Abhinav Shrivastava, Abhinav Gupta, and Ross Girshick. Training Region-based Object Detectors with Online Hard Example Mining. In CVPR, 2016. 2", + "[45] Sanjay Subramanian, William Merrill, Trevor Darrell, Matt Gardner, Sameer Singh, and Anna Rohrbach. ReCLIP: A Strong Zero-Shot Baseline for Referring Expression Comprehension. In ACL, 2022. 1", + "[46] Tristan Thrush, Ryan Jiang, Max Bartolo, Amanpreet Singh, Adina Williams, Douwe Kiela, and Candace Ross. Winoground: Probing Vision and Language Models for Visio-Linguistic Compositionality. In CVPR, 2022. 1, 2, 4, 7", + "[47] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models, 2023. 2, 3, 4", + "[48] Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9 (86):2579-2605, 2008. 8", + "[49] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. OFA: Unifying Architectures, Tasks, and Modalities Through a Simple Sequence-to-Sequence Learning Framework. In ICML, 2022. 6", + "[50] Chenyun Wu, Zhe Lin, Scott Cohen, Trung Bui, and Subhransu Maji. PhraseCut: Language-based Image Segmentation in the Wild. In CVPR, 2020. 1, 2, 5", + "[51] Hao Wu, Jiayuan Mao, Yufeng Zhang, Yuning Jiang, Lei Li, Weiwei Sun, and Wei-Ying Ma. Unified Visual-Semantic Embeddings: Bridging Vision and Language with Structured Meaning Representations. In CVPR, 2019. 3", + "[52] Xiaoshi Wu, Feng Zhu, Rui Zhao, and Hongsheng Li. CORA: Adapting CLIP for Open-Vocabulary Detection with Region Prompting and Anchor Pre-Matching. In CVPR, 2023. 2", + "[53] Chi Xie, Zhao Zhang, Yixuan Wu, Feng Zhu, Rui Zhao, and Shuang Liang. Described Object Detection: Liberating Object Detection with Flexible Expressions. In NeurIPS, 2023. 1, 2, 5, 6", + "[54] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C. Berg, and Tamara L. Berg. Modeling Context in Referring Expressions. In ECCV, 2016. 1, 2, 5", + "[55] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language models behave like bags-of-words, and what to do about it? In ICLR, 2023. 2", + "[56] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-Vocabulary Object Detection Using Captions. In CVPR, 2021. 2", + "[57] Shiyu Zhao, Zhixing Zhang, Samuel Schulter, Long Zhao, BG Vijay Kumar, Anastasis Stathopoulos, Manmohan Chandraker, and Dimitris N Metaxas. Exploiting unlabeled data with vision and language models for object detection. In ECCV, pages 159-175. Springer, 2022. 1, 2" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13601", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[58] Shiyu Zhao, Samuel Schulter, Long Zhao, Zhixing Zhang, Vijay Kumar B. G, Yumin Suh, Manmohan Chandraker, and Dimitris N. Metaxas. Improving pseudo labels for open-vocabulary object detection, 2023. 2", + "[59] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, and Jianfeng Gao. RegionCLIP: Region-based Language-Image Pretraining. In CVPR, 2022. 2", + "[60] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbihl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In ECCV, pages 350-368. Springer, 2022. 5" + ], + "bbox": [ + 78, + 90, + 468, + 273 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "13602", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Generating Enhanced Negatives for Training Language-Based Object Detectors/2970baed-7bcf-45c6-83b6-e79faa1f9b77_model.json b/2024/Generating Enhanced Negatives for Training Language-Based Object Detectors/2970baed-7bcf-45c6-83b6-e79faa1f9b77_model.json new file mode 100644 index 0000000000000000000000000000000000000000..159df5c0358ef64daed5fefe83a1fe2c81247664 --- /dev/null +++ b/2024/Generating Enhanced Negatives for Training Language-Based Object Detectors/2970baed-7bcf-45c6-83b6-e79faa1f9b77_model.json @@ -0,0 +1,2345 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.085, + 0.131, + 0.885, + 0.153 + ], + "angle": 0, + "content": "Generating Enhanced Negatives for Training Language-Based Object Detectors" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.18, + 0.329, + 0.198 + ], + "angle": 0, + "content": "Shiyu Zhao\\(^{1,*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.356, + 0.181, + 0.453, + 0.198 + ], + "angle": 0, + "content": "Long Zhao3" + }, + { + "type": "text", + "bbox": [ + 0.481, + 0.182, + 0.628, + 0.198 + ], + "angle": 0, + "content": "Vijay Kumar B G²" + }, + { + "type": "text", + "bbox": [ + 0.657, + 0.182, + 0.753, + 0.197 + ], + "angle": 0, + "content": "Yumin Suh2" + }, + { + "type": "text", + "bbox": [ + 0.202, + 0.2, + 0.37, + 0.215 + ], + "angle": 0, + "content": "Dimitris N. Metaxas1" + }, + { + "type": "text", + "bbox": [ + 0.398, + 0.199, + 0.604, + 0.216 + ], + "angle": 0, + "content": "Manmohan Chandraker2,4" + }, + { + "type": "text", + "bbox": [ + 0.631, + 0.199, + 0.77, + 0.216 + ], + "angle": 0, + "content": "Samuel Schulter2" + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.233, + 0.271, + 0.252 + ], + "angle": 0, + "content": "\\(^{1}\\) Rutgers University" + }, + { + "type": "text", + "bbox": [ + 0.296, + 0.234, + 0.528, + 0.251 + ], + "angle": 0, + "content": "\\(^{2}\\) NEC Laboratories America" + }, + { + "type": "text", + "bbox": [ + 0.553, + 0.234, + 0.704, + 0.251 + ], + "angle": 0, + "content": "3 Google Research" + }, + { + "type": "text", + "bbox": [ + 0.729, + 0.234, + 0.858, + 0.251 + ], + "angle": 0, + "content": "4 UC San Diego" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.286, + 0.313, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.474, + 0.621 + ], + "angle": 0, + "content": "The recent progress in language-based open-vocabulary object detection can be largely attributed to finding better ways of leveraging large-scale data with free-form text annotations. Training such models with a discriminative objective function has proven successful, but requires good positive and negative samples. However, the free-form nature and the open vocabulary of object descriptions make the space of negatives extremely large. Prior works randomly sample negatives or use rule-based techniques to build them. In contrast, we propose to leverage the vast knowledge built into modern generative models to automatically build negatives that are more relevant to the original data. Specifically, we use large-language-models to generate negative text descriptions, and text-to-image diffusion models to also generate corresponding negative images. Our experimental analysis confirms the relevance of the generated negative data, and its use in language-based detectors improves performance on two complex benchmarks. Code is available at https://github.com/xiaofeng94/Gen-Enhanced-Negs." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.641, + 0.208, + 0.656 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.666, + 0.47, + 0.832 + ], + "angle": 0, + "content": "Using natural language in object detection to describe semantics bears the potential to significantly increase the size of the detector's label space and enable novel applications. While standard detectors operate on a fixed label space [23, 38, 42], natural language allows for a broad spectrum of object descriptions, ranging from generic terms like \"vehicle\" to specific expressions like \"the red sports car parked on the left side\" [12, 17, 25, 30, 41, 53, 54]. Several works advanced language-based object detection over the past few years with novel training strategies [3, 5, 19, 22, 32, 34, 57] and model architectures [11, 15, 33, 45]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.834, + 0.47, + 0.864 + ], + "angle": 0, + "content": "Referring expression or visual grounding datasets [14, 30, 36, 50, 54] provide the natural language object descrip" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.279, + 0.626, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.628, + 0.281, + 0.89, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.471, + 0.893, + 0.554 + ], + "angle": 0, + "content": "Figure 1. The key contribution of our work is to leverage large-language-models and text-to-image diffusion models to automatically generate negative object descriptions and images for training language-based object detectors. In contrast to prior work, our generated negatives are more relevant to the original data and provide a better training signal for detectors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.893, + 0.659 + ], + "angle": 0, + "content": "tions along with bounding box annotations needed for training. However, this data only describes what is present in the images, but misses to describe what is not. Yet, the notion of negatives is crucial for training discriminative models like language-based detectors [7, 24, 39]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Detection datasets with a fixed label space provide negative classes implicitly or explicitly, with exhaustive [23, 42] or federated [6, 16] annotations, respectively. Any part of an image that does not overlap significantly with a bounding box of category \\( c \\) is verified to not be of that category (for exhaustive annotation). On the other hand, the space of negatives for a free-form text description of an object is extremely large. While some existing datasets provide negative samples in free-form text [43, 46], they were not annotated with bounding boxes. Hence, existing language-based detectors often define the negatives for one object as the descriptions of all other objects in the same image or descriptions of other random samples [3, 11, 19]. However, such negatives may not be directly related to the original positive description and define a weaker training signal (see Fig. 1). By explicitly evaluating on human-curated negatives, a re" + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.875, + 0.469, + 0.901 + ], + "angle": 0, + "content": "* Part of this work was done during an internship at NEC Laboratories America. Correspondence to: Shiyu Zhao sz553@rutgers.edu" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13592" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.168 + ], + "angle": 0, + "content": "cent benchmark [41] identified a bias of existing language-based detectors to perform clearly better on positive rather than negative descriptions. However, creating a dataset with high-quality human-curated negatives for large-scale training is labor-intensive and costly." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.17, + 0.468, + 0.32 + ], + "angle": 0, + "content": "In this work, we propose to explicitly and automatically generate negative data in the form of free-form texts as well as images. Prior works [4, 7, 29, 43, 46, 55] rely on rule-based approaches with knowledge graphs and focus only on the language domain or the classification task. In contrast, we leverage generative large-language-models (LLMs) [35, 47] and text-to-image diffusion models [21, 40] to automatically create relevant but contradicting object descriptions along with the corresponding images for language-based object detection, see Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.323, + 0.468, + 0.715 + ], + "angle": 0, + "content": "Given an object description of a dataset, we first use LLMs to generate a semantically contradicting description as the negative. Besides changing individual words (foils) based on explicit knowledge graphs or LLMs, like in prior work [4, 7, 20], we demonstrate improved detection performance with two alternative approaches. (Re-combination): An LLM first identifies all objects in a sentence, and then creates a contradicting one by re-arranging, ignoring or adding objects. (In-context summaries): We prompt an LLM to summarize the differences of a few (less than 100) positive-negative pairs collected from an existing image-level dataset [46]. This summary is then used as context to generate more such examples. Note that we do not need visual input for this step, allowing us to leverage powerful LLMs for semantic and textual reasoning. Moreover, while prior work only focused on the text [4, 7, 29, 43, 55], we also leverage text-to-image diffusion models like GLIGEN [21] to create images that match the generated negative descriptions of objects, which serves as additional training signal. While the direct output of such image-generation models is often noisy and even wrong (not matching the input description), we propose two filtering steps to reduce noise considerably (from \\(53\\%\\) to \\(16\\%\\) according to an empirical study). Having both negative object descriptions and the corresponding image, allows us to improve the discriminative loss for training language-based object detectors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.718, + 0.468, + 0.824 + ], + "angle": 0, + "content": "Our experiments demonstrate clear accuracy gains on two challenging benchmarks, +2.9AP on OmniLabel [41] and +3.3AP on \\(\\mathrm{D}^3\\) [53], when adding our automatically-generated negative data into the training of baseline models like GLIP or FIBER. Moreover, we provide an in-depth analysis of the generated data (text and images) and how they contribute to better language-based detection." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Summary of contributions: (1) Automatic generation of semantically relevant but contradicting negative text and images with large-scale generative models. (2) Recipes to integrate such negative data into language-based detection models like FIBER [3] and GLIP [19] (3) Clear improve" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.123 + ], + "angle": 0, + "content": "ments on language-based detection benchmarks [41, 53] including a thorough analysis of the generated data." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.138, + 0.638, + 0.154 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.163, + 0.892, + 0.42 + ], + "angle": 0, + "content": "Vision & language localization tasks: Open-vocabulary detection (OVD) requires a model to localize object category names without having seen explicit bounding box annotation for them [5, 7, 15, 52, 56, 59]. In contrast, we focus on the more general language-based object detection task [41, 53], which goes beyond simple category names. Referring expression comprehension (REC) aims at localizing the subject of a free-form text expression. However, REC benchmarks [30, 50, 54] fall short in evaluating all aspects of the more general language-based detection task [41, 53]. In visual grounding (VG) [36], the task is to localize noun phrases of a caption in the image. Although being a task on its own, VG datasets have recently been used mostly as training data for OVD. Our work focuses on general language-based object detection, which subsumes and generalizes standard detection, OVD and REC [12, 25, 41, 53]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.426, + 0.892, + 0.608 + ], + "angle": 0, + "content": "Language-based object detectors: Two critical abilities of language-based detection are accurate localization and tight text-image fusion. Works like [1, 9, 28] use language-models like BERT [2, 27] to align regions extracted from (pre-trained) detectors with captions. The outstanding zero-shot classification accuracy of large-scale pre-trained models like CLIP [37] or [10, 13, 18] then sparked interest in extensions for localization, with different approaches like distillation [5], fine-tuning [15, 33], pseudo-labeling [34, 57, 58], or combinations thereof [3, 19]. We use such models as test bed, but explore the underlying training data with respect to negative samples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Negative samples for object detection: The notion of negatives is crucial for training discriminative models [24, 39]. Also for object detection, hard negative mining [44] has proven beneficial for model training. However, these prior works aim to find hard negative training examples rather than negatives in the label space, because the label space is fixed in standard detection. For language-based datasets, the space of potential negatives is extremely large because object descriptions are free-form text. Prior works [4, 29, 43, 46, 55] investigate negative texts for general vision & language models with different strategies, including changing individual words (foil) with rules based on knowledge graphs [31] or with LLMs. Sugar-Crepe [8] shares a similar idea as us to get negative texts with in-context learning but for image-text level pretraining. For language-based detection, [7] explores such rule-based foils, while [20] uses LLMs with specific templates to replace object names with alternative descriptions. In contrast, our work (1) focuses on the localization task, (2) ex" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13593" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.472, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.174, + 0.47, + 0.23 + ], + "angle": 0, + "content": "Figure 2. In language-based object detection, a detector receives as input an image and a (variable-length) list of free-form text descriptions of objects. For each description, the model predicts bounding boxes for objects that match the description." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.257, + 0.469, + 0.303 + ], + "angle": 0, + "content": "plies more comprehensive strategies to generate negatives with LLMs, and (3) proposes to also generate corresponding negative images with text-to-image diffusion models." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.317, + 0.168, + 0.332 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.342, + 0.365, + 0.358 + ], + "angle": 0, + "content": "3.1. Language-based object detection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.365, + 0.47, + 0.486 + ], + "angle": 0, + "content": "Task definition: Given an image and a list of object descriptions, the task is to output bounding boxes along with confidence scores for each description, as shown in Fig. 2. Note the multi-label setting where one object instance can be referred to by multiple descriptions, like \"person\" and \"person looking at book\". Also note that an object description might refer to zero objects in the image and the desired output is an empty set of boxes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.491, + 0.471, + 0.839 + ], + "angle": 0, + "content": "Training data: Many language-based detection models [3, 19] use a combination of object detection [23, 42] and visual grounding datasets [14, 36] for training their models. Both types of datasets provide images \\(I\\) and bounding boxes \\(b_{l}\\) to localize individual objects. Object detection data assigns each bounding box \\(b_{l}\\) a unique category \\(c\\) out of fixed label space \\(\\mathcal{C}\\). The exhaustive labeling of the fixed label space in detection datasets implies the space of negatives. An object of category \\(c\\) is not any of the categories \\(\\mathcal{C} \\setminus c\\). On the other hand, grounding data provides an image caption \\(t\\) in free-form text, where subsets of words \\(m_{l}\\) (defined as indices of starting and ending characters in \\(t\\)) are linked to bounding box \\(b_{l}\\). For grounding data, the space of negatives is extremely large because one can find as many textual descriptions that do not match \\(t\\) as desired due to the compositionality of free-form text. Many language-based detectors [3, 19] only use the words in \\(t\\) that are not referred to by \\(m_{l}\\) as negatives for the bounding box \\(b_{l}\\). We argue that this choice is sub-optimal because these words may refer to entirely different objects and are easy to discriminate. In the following section, we explain how we can automatically generate negative samples that semantically related to the original text \\(t\\) and hence provide a better training signal." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.334, + 0.864 + ], + "angle": 0, + "content": "3.2. Generating negative samples" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Our goal is to automatically and explicitly generate negative samples based on the original text descriptions \\( t \\) to im" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "prove the training signal for language-based detectors. A key observation of our work is to leverage the vast knowledge encoded into large-language models (LLMs) [35, 47] and text-to-image diffusion models (T2I) [21, 40]. Besides proposing novel ways to instruct LLMs for generating negative text descriptions (Sec. 3.2.1), we also propose to generate negative images (Sec. 3.2.2)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.222, + 0.866, + 0.236 + ], + "angle": 0, + "content": "3.2.1 Generating negative descriptions with LLMs" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.248, + 0.892, + 0.368 + ], + "angle": 0, + "content": "Given an object description \\( t \\) that matches the visual content inside a bounding box \\( b_{l} \\), we define a \"negative\" description \\( t' \\) as any text that is semantically different to the original text. Furthermore, our intuition is that good negative descriptions are still semantically related to the original description, but not the same. An example is: \"Person in red shirt\" as the original description and \"Person in blue shirt\" as a contradicting negative one." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.37, + 0.893, + 0.567 + ], + "angle": 0, + "content": "Prior work [4, 7, 43] explored rule-based approaches to generate negative text. However, such rules are typically limited to simple knowledge graphs and are limited to replacing only individual words, often just nouns, or swapping words. In contrast, we explore more powerful LLMs to automatically generate relevant negatives. To make the negative text generation efficient and economic, in all cases, we first leverage a strong instruction-tuned LLM [35] to generate \\(50\\mathrm{k}\\) positive-negative pairs, and then finetune a LLaMA-7B [47] model with those pairs to then generate negative captions on large grounding datasets. In the following, we describe three ways to instruct an LLM for generating positive-negative pairs of object descriptions:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.575, + 0.892, + 0.772 + ], + "angle": 0, + "content": "LLM-based foils: We first prompt an instruction-tuned LLM [35] to find concepts (i.e., objects, attributes and relationships) in object descriptions. Compared to rule-based parsers [51], LLMs can provide richer information. For example, for the caption \"A transportation vehicle is carrying a crowd of people who are sitting and standing,\" the parser ignores \"sitting\" and \"standing\", while LLMs regard them as attributes. Then, we pick one concept from the first step sequentially and prompt LLMs again to generate a negative caption by changing the concept. For both steps, the prompts are manually curated with the task definition and step-by-step instructions for the generation. Please find the exact prompt for the LLM in the supplement." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Re-combination: Next, we give the LLM more freedom in generating negative descriptions. We first prompt the LLM to identify all objects in the original caption, and then to recombine them to create a new sentence different from the original one. We allow the LLM to ignore, change or add new objects. For example, given the caption “A boy is playing with his dog” and two objects “boy” and “dog”, the LLM can output “The girl and her dog are playing fetch in" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13594" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.471, + 0.122 + ], + "angle": 0, + "content": "the park\". Detailed prompts for both identifying objects and re-combination are in the supplement." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.131, + 0.473, + 0.388 + ], + "angle": 0, + "content": "In-context summary: Third, we enable LLMs to learn how to generate negative descriptions by providing human-annotated positive-negative pairs as in-context samples. We randomly sample 80 pairs of positive and negative texts from the Winoground dataset [46] and prompt the instruction-tuned LLM [35] to summarize the difference of those pairs in plain text. Then, instead of manually creating prompts to generate positive-negative pairs, we leverage the summary together with three randomly sampled Winoground pairs as prompts to the LLM, and generate several positive-negative pairs to finetune LLaMA. After finetuning, the LLaMA model is used to generate negative texts for given descriptions. This pipeline does not require handcrafted prompts to LLMs as the explanation of the concept of negatives and how to create them. The supplement contains full prompts for generating a summary, and generating positive-negative pairs for finetuning." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.412, + 0.442, + 0.427 + ], + "angle": 0, + "content": "3.2.2 Generating negative images with T2I models" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.438, + 0.47, + 0.589 + ], + "angle": 0, + "content": "Given an original image \\( I \\), a bounding box \\( b \\) and a corresponding object description \\( t \\), we define a negative image \\( I' \\) as any image that has a different semantic content inside \\( b \\). The rest of the image can be equivalent to \\( I \\). To obtain such imagery, we start with visual grounding data that provide bounding boxes, positive captions with text phrases, and alignment between them. We propose a two-step process: First, we turn the positive caption into a negative one. Second, we use conditioned image generation tools to alter the visual content inside the bounding box \\( b \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.47, + 0.809 + ], + "angle": 0, + "content": "Negative text for negative images: Although we have already described an approach to generate negative descriptions in Sec. 3.2.1, doing so to generate a negative image requires a different approach. In this case, the generated negative text needs to preserve the alignment \\( m_{l} \\) to the ground truth bounding box \\( b_{l} \\) in order to instruct the generative image model GLIGEN [21]. Hence, we first select a bounding box \\( b_{l} \\) and mask out the corresponding words (known via \\( m_{l} \\)) in the text \\( t \\). For example, \"A boy is playing with his dog\" turns into \"A boy is playing with [Mask]\" if the selected bounding box refers to \"his dog\". Again, we leverage LLMs [35] to fill in text for \"[Mask]\" to generate a negative text without reusing the original text. Please refer to Fig. 3 for illustrations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We finetune a LLaMA-7B for the mask filling task with triplets of positive texts, masked texts, and negative texts. To reduce manual efforts, we follow the approach of in-context summary to get triplet samples. We apply this process twice: We start with only 5 manually created triplets to build a summary and generate 100 samples from the" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.091, + 0.891, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.262, + 0.893, + 0.291 + ], + "angle": 0, + "content": "Figure 3. Overview of using LLMs [35, 47] and text-to-image diffusion models [21, 40] to generate negative images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.317, + 0.892, + 0.378 + ], + "angle": 0, + "content": "LLM [35] with human checks. We then repeat the process to generate 50k examples without human checks from a summary of the 100 generated examples. This increases diversity in the generated data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.381, + 0.893, + 0.503 + ], + "angle": 0, + "content": "Conditional image generation: Given an image \\( I \\), a bounding box \\( b \\) and the altered text \\( t' \\), we generate a negative image \\( I' \\) that is equal to \\( I \\) except inside \\( b \\), where the visual content is altered to match the text \\( t' \\). To do so, we use the inpainting and conditioning abilities of GLIGEN [21], a T2I model [40]. Refer to [21] and our supplemental material for more details, and to Fig. 3 for an illustration of the process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.506, + 0.895, + 0.823 + ], + "angle": 0, + "content": "Mitigating noise in image generation: We found that the generated images are often noisy for any of the following reasons: (1) The altered text refers to a big bounding box that covers other smaller boxes. Large portions of the image are then generated and often do not match the concepts those smaller boxes originally covered. (2) The generated negative text does not match the bounding box that is either too small, too large or in an inappropriate position. (3) The T2I model fails to understand the negative text and generates wrong content. We propose two steps to filter such noisy images. First, we simply ignore ground truth boxes \\( b_{l} \\) for image generation if the box covers more than \\( 75\\% \\) of any other boxes in the image. Second, we adopt CLIP [37] to verify the semantic similarity of the generated image regions and the corresponding text. Specifically, we compute the similarity with CLIP between the generated image region (visual input) and the original and generated negative texts (text input). We filter out generated images that have a similarity score to the generated negative text lower than a user-defined threshold. Details on the filtering steps are given in the supplemental." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.832, + 0.783, + 0.849 + ], + "angle": 0, + "content": "3.3. Learning from negative samples" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Detector design and training objective: The generated data does not prescribe any specific architecture for the detector. A common choice, which we also use for our exper" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13595" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.087, + 0.473, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.244, + 0.47, + 0.342 + ], + "angle": 0, + "content": "Figure 4. Illustration of the grounding loss used for training. Predictions that are matched with ground truth receive a positive signal from the associated text (tall rectangles). All other words receive a negative signal (short rectangles). The top left quarter shows the original loss used in [3, 19]. The other three quarters are related to our proposed generated negative data and provide additional training signals." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.37, + 0.473, + 0.581 + ], + "angle": 0, + "content": "iments, is [3, 19]. The inputs are image \\(I\\) and text \\(t\\), and the output is a set of bounding boxes \\(\\hat{b}_i\\) with corresponding logits \\(\\hat{p}_i \\in \\mathbb{R}^T\\). Here, \\(T\\) is the number of tokens required to represent text \\(t\\). The ground truth can be represented by a binary assignment matrix \\(\\mathbf{A} \\in \\mathbb{B}^{L \\times T}\\). Rows refer to ground truth boxes \\(l\\) and columns to tokens in \\(t\\). Each element indicates if a token corresponds to a box \\(l\\), which is given by the ground truth indices \\(m_l\\). To define a loss, bipartite graph matching associates predictions with ground truth. For matched predictions, the target vector \\(g_i \\in \\mathbb{B}^T\\) is the corresponding row from \\(\\mathbf{A}\\), while it is an all-zero vector for unmatched targets. The loss is then computed as \\(\\mathcal{L} = \\sum_{i} \\ell_{\\mathrm{FL}}(\\hat{p}_i, g_i)\\) where FL refers to a focal loss. Fig. 4 illustrates the loss." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.59, + 0.47, + 0.698 + ], + "angle": 0, + "content": "Integrating negative text: When sampling an image \\( I \\), along with text \\( t \\), boxes \\( b_{l} \\) and indices \\( m_{l} \\), we also randomly sample \\( K > 1 \\) negative descriptions from \\( \\{t_j^{\\prime}\\} \\) that defines the pool of negatives generated for text \\( t \\). We randomly shuffle the order of all texts to avoid any biases on the location of the one positive description, and then concatenate them into one text string." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Integrating negative images: We explore two options: (1) Simply add the generated images \\( I' \\) along with their generated (but semantically matching) captions \\( t' \\) as additional visual grounding data. The original caption \\( t \\), which was the starting point to generate the negative image \\( I' \\), is now used as the negative caption. In this way, both the original image \\( I \\) and the generated one \\( I' \\) have positive and negative descriptions. This option is illustrated in Fig. 4. (2) To better leverage the relation between the original and generated data, the second option is to pack them into a single training sample. We simply concatenate the images \\( I \\) and \\( I' \\), as well as the texts \\( t \\) and \\( t' \\). The ground truth information \\( m_{l} \\) is updated accordingly. See supplement for details." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.089, + 0.89, + 0.221 + ], + "angle": 0, + "content": "
OmniLabelOmniLabel-Negative
APAPcAPdAPdPAPAPcAPdAPdP
Detic [60]8.015.65.48.0----
MDETR [11]--4.79.1----
GLIP-T [19]19.323.616.425.813.924.89.626.1
+ Ours22.227.218.829.016.528.611.630.2
FIBER-B [3]25.730.322.334.818.731.213.336.3
+ Ours28.132.125.136.522.333.316.738.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.53, + 0.23, + 0.86, + 0.244 + ], + "angle": 0, + "content": "Table 1. Evaluation on the OmniLabel [41] benchmark." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.278, + 0.633, + 0.294 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.306, + 0.695, + 0.323 + ], + "angle": 0, + "content": "4.1. Experimental design" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.334, + 0.892, + 0.53 + ], + "angle": 0, + "content": "Training procedure: We choose two recent methods, GLIP-T [19] and FIBER-B [3], to demonstrate the effect of our automatically generated negatives. We use the official code and publicly available checkpoints as a starting point. The Flickr30k dataset [36] serves as our grounding dataset to generate the negative data. We then fine-tune GLIP-T and FIBER-B with both positive and negative data, along with the Objects365 detection dataset [42] for 1 epoch. Note that both Objects365 and Flickr30k are part of the original training set. We do not introduce any extra data except our generated negatives. Most hyper-parameters are equal to the original settings of GLIP and FIBER. Any exceptions are described in the supplement." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Evaluation benchmarks: We choose two recently proposed benchmarks, OmniLabel [41] and \\(\\mathrm{D}^3\\) [53], as our test beds. These benchmarks evaluate more aspects of language-based detection than existing referring expressions [30, 50, 54] or open-vocabulary detection [6, 17] benchmarks. Specifically, both benchmarks contain complex object descriptions that go beyond simple category names from open-vocabulary detection benchmarks. Moreover, the descriptions can refer to zero, one or multiple instances in the image, in contrast to standard referring expression benchmarks. These properties enable a more stringent evaluation metric as in object detection, which is based on average precision (AP) in both OmniLabel [41] and \\(\\mathrm{D}^3\\) [53]. Both benchmarks provide more fine-grained metrics. OmniLabel evaluates separately for categories, descriptions, and descriptions referring to at least one object, with APc, APd and APd-P, respectively. \\(\\mathrm{D}^3\\) differentiates descriptions on absence (\"Abs\") and presence (\"Pres\") that indicate whether or not they contain any form of negation (e.g., \"without\"), as well as on text lengths. Finally, we create a specific split for OmniLabel, \"OmniLabel-Negative\", to evaluate the model only on images that contain at least one negative description (i.e., not referring to any object)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13596" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.089, + 0.468, + 0.246 + ], + "angle": 0, + "content": "
D3(default)D3(by length of texts)
FullPresAbsSMLXL
OFA-L [49]4.24.14.64.95.43.02.1
OWL-ViT-L [33]9.610.76.420.79.46.05.3
G-DINO-B [26]20.720.122.522.622.518.916.5
OFA-DOD [53]21.623.715.423.622.620.518.4
GLIP-T [19]19.118.321.522.422.016.610.6
+ Ours21.420.623.728.124.517.411.5
FIBER-B [3]22.721.526.030.125.917.913.1
+ Ours26.025.228.135.529.720.514.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.129, + 0.254, + 0.416, + 0.269 + ], + "angle": 0, + "content": "Table 2. Evaluation on the \\( {\\mathrm{D}}^{3} \\) [53] benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.295, + 0.307, + 0.312 + ], + "angle": 0, + "content": "4.2. Benchmark comparisons" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.473, + 0.501 + ], + "angle": 0, + "content": "Tabs. 1 and 2 evaluate the impact of our generated negative training data on the OmniLabel [41] and \\(\\mathrm{D}^3\\) [53] benchmarks. In both tables, the first set of rows are baselines provided by the benchmarks. The following rows show the main comparisons for GLIP-T [19] and FIBER-B [3] with and without adding our generated negative training data. First, we can see that adding negative data improves results across all metrics for both models and both benchmarks. On OmniLabel, we can see a \\(+2.9\\%\\) and \\(+2.4\\%\\) increase in AP for GLIP-T and FIBER-B, respectively. Similarly, we observe a \\(+2.3\\%\\) and \\(+3.3\\%\\) increase in the main metric of \\(\\mathrm{D}^3\\) (AP on full descriptions) for GLIP-T and FIBER-B." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.511, + 0.313, + 0.527 + ], + "angle": 0, + "content": "4.3. Analysis on negative texts" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.534, + 0.473, + 0.867 + ], + "angle": 0, + "content": "Effectiveness of different negative texts: We finetune FIBER-B without and with different kinds of negative texts mentioned in Sect. 3.2.1, i.e., Rule-based foils, LLM-based foils, Re-combination with LLMs, In-context summary with LLMs, and present results in Table 3. We find all kinds of negatives improve the original FIBER-B on both OmniLabel and \\(\\mathrm{D}^3\\) benchmarks. Negative texts from LLMs generally achieve better results compared to LLM-based foils, which indicates that LLMs are powerful tools for negative text generation. Moreover, both recombination and in-context summary with LLMs outperform LLM-based foils in all metrics except APd-P. Note that APd-P refers to evaluations without negative label spaces, which is a task weaker than language-based detection. Based on such results, we argue that although word foils provide promising results in traditional studies [7, 43], it is sub-optimal to LLMs. We need to explore varied ways to unlock the ability of LLMs. We believe that our two solutions, i.e., Re-combination and In-context summary, provide a good starting point for future studies. Besides using only one kind of negative texts, we also explore the combinations of different kinds of negative texts in the supplement." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Diversity of rule-based and LLM-based negatives: In this part, we investigate the diversity of different negative" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.093, + 0.701, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.094, + 0.885, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.174, + 0.702, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.174, + 0.885, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.267, + 0.892, + 0.296 + ], + "angle": 0, + "content": "Figure 5. Percentage of negative texts with the numbers of words. Four negative generation methods are compared." + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.317, + 0.7, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.318, + 0.884, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.399, + 0.7, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.399, + 0.884, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.492, + 0.892, + 0.534 + ], + "angle": 0, + "content": "Figure 6. Percentage of negative texts with the numbers of words that are different from the original caption. Four negative generation methods are compared." + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.557, + 0.833, + 0.657 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.66, + 0.892, + 0.703 + ], + "angle": 0, + "content": "Figure 7. Average numbers of extra unique words per thousand generated negative texts, which are not included in the original dataset. We group words by their part-of-speech." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.733, + 0.892, + 0.824 + ], + "angle": 0, + "content": "texts. First, we count number of words for each negative text and provide the distribution for negatives of different sources in Fig. 5. As shown, all four distributions have a peak around 10 words, but the one of rule-based foils is higher than others. That means rule-based foils provide more negative texts with similar lengths." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Second, we count the number of different words between the original positive caption and the negative caption, and present the distributions in Fig. 6. We find that LLM-based methods usually changes more words than rule-based foils, which increases the diversity. Moreover, in-context sum" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13597" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.09, + 0.568, + 0.202 + ], + "angle": 0, + "content": "
Whole OmniLabelOmniLabel-NegativeD3
APAPcAPdAPd-PAPAPcAPdAPd-PFullPresAbs
Original FIBER-B25.730.322.334.818.731.213.336.322.721.526.0
+ Rule-based foils26.431.722.634.919.232.613.636.424.123.226.9
+ LLM-based foils26.530.723.335.920.832.115.438.024.624.026.5
+ Re-combination26.930.823.935.921.132.315.637.625.324.627.3
+ In-context summary26.630.823.434.221.132.215.736.425.725.227.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.212, + 0.57, + 0.24 + ], + "angle": 0, + "content": "Table 3. Performance of FIBER-B trained with negative texts from four negative generation methods." + }, + { + "type": "table", + "bbox": [ + 0.587, + 0.091, + 0.89, + 0.23 + ], + "angle": 0, + "content": "
OmniLabelD3
APcAPdAPd-P
Original FIBER30.322.334.822.7
FIBER w/ neg. texts30.723.935.525.9
+ W/ neg. img. directly30.122.433.723.0
+ Box filter31.023.835.423.6
+ Box&CLIP filters31.124.235.924.1
+ Above + concat. img.31.724.835.924.8
+ Above + weight ensemble32.125.236.526.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.6, + 0.24, + 0.874, + 0.255 + ], + "angle": 0, + "content": "Table 4. FIBER trained with negative images." + }, + { + "type": "image", + "bbox": [ + 0.128, + 0.278, + 0.422, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.45, + 0.47, + 0.478 + ], + "angle": 0, + "content": "Figure 8. Noisy generated images. The orange box contains the red box, and editing the orange changes the red unexpectedly." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.505, + 0.469, + 0.626 + ], + "angle": 0, + "content": "mary has a more flat distribution compared to others. Probably, in-context summary learns how to generate negatives automatically from data and has less restrictions. Besides, in-context summary has more cases with no word changed where negative texts are generated by just shuffling words or concepts in the original text. Such shuffling is a common pattern of Winoground [46], and our in-context summary can learn such data specific patterns." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.627, + 0.47, + 0.837 + ], + "angle": 0, + "content": "Third, we count how many extra words that does not exist in the original Flickr30k dataset are introduced in different negative generation methods. Fig. 7 shows the average number of extra words per 1000 negative texts. We group words into four part-of-speech categories, i.e., VERB, NOUN, ADP/ADJ, and others. As shown, LLMs introduce more extra words on average than rule-based foils probably because rule-based foils are limited in a predefined set of words. However, LLMs are open to any concepts and have great potentials of generating diverse texts. In-context summary introduces the most extra words for all categories, which is likely a benefit of learning negative generation from data. The above statistics indicate a clear view that LLMs generate more diverse data than rule-based foils." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.329, + 0.865 + ], + "angle": 0, + "content": "4.4. Analysis on negative images" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Noise in generated images: As mentioned in the last paragraph of Sect. 3.2.2, the raw generated images are noisy in" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.281, + 0.893, + 0.553 + ], + "angle": 0, + "content": "several ways. First, the editing of a large box will override the context of smaller boxes that are covered by the large box. As shown in Fig. 8, GLIGEN did follow the instruction to generate a boy in the orange box, but the black jacket in the red box is missing. As a remedy, we apply our first de-noise step \"Box Filter\". That is, we ignore boxes that contain any other boxes when generating negative images. Second, GLIGEN may generate contents with wrong attributes or objects, as shown in Fig. 9 (Left). Moreover, our generation pipeline includes some cases where the edited text and the bounding box does not match. As shown in Fig. 9 (Right), the box for \"his lap\" cannot be modified as \"his knees\". Thus, GLIGEN generates wrong contents. As described in Sect. 3.2.2, we adopt a pretrained CLIP model to judge if generated contents are correct, which mitigates the noise to some extent. As shown in Fig. 9, both negative images get low CLIP scores and can be filtered out with a threshold. We call such thresholding \"CLIP Filter\"." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.561, + 0.892, + 0.712 + ], + "angle": 0, + "content": "Subject studies on Box and CLIP filters: We employ human experts to check the amount of noisy generated images. First, for negative images w/o filter, w/ Box filter, and w/ Box&CLIP filters, we separately and randomly select 100 samples. Then, we ask two experts to check if a negative image is not noisy by comparing it with its caption and the original positive image. We regard an image as not noisy when both experts agree. As shown in Fig. 10, both filters reduce the noise. The Box filter improves from \\(47\\%\\) to \\(63\\%\\), and the CLIP filter improves to \\(84\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Effectiveness of generated negative images: To show the effectiveness of generated images themselves, we take captions of generated images as additional negative texts to in-context summary, and finetune a FIBER model as baseline. Then, we compare the baseline with variants of adding generated negative images in Table 4. As shown, the performance drops if we directly take raw negative images as new visual grounding data without any filters (i.e., W/ neg. img. directly). Probably, there are too much noise in raw negative images as shown in Fig. 10. When applying both Box and CLIP filters on negative images, we can achieve slight improvement on OmniLabel compared to using neg-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13598" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.206, + 0.089, + 0.377, + 0.101 + ], + "angle": 0, + "content": "Generating wrong attributes" + }, + { + "type": "image_caption", + "bbox": [ + 0.185, + 0.103, + 0.233, + 0.114 + ], + "angle": 0, + "content": "Original" + }, + { + "type": "image_caption", + "bbox": [ + 0.36, + 0.103, + 0.419, + 0.114 + ], + "angle": 0, + "content": "Generated" + }, + { + "type": "image", + "bbox": [ + 0.12, + 0.116, + 0.298, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.116, + 0.48, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.124, + 0.223, + 0.466, + 0.235 + ], + "angle": 0, + "content": "Positive: A woman in a black mesh skirt plays acoustic guitar" + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.239, + 0.468, + 0.252 + ], + "angle": 0, + "content": "Negative: A woman in a white lace dress plays acoustic guitar" + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.089, + 0.807, + 0.101 + ], + "angle": 0, + "content": "Generation with inappropriate boxes or texts" + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.103, + 0.605, + 0.115 + ], + "angle": 0, + "content": "Original" + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.116, + 0.671, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.731, + 0.102, + 0.79, + 0.113 + ], + "angle": 0, + "content": "Generated" + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.116, + 0.852, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.474, + 0.223, + 0.83, + 0.236 + ], + "angle": 0, + "content": "Positive: A man is sitting on a chair with his hands above his lap" + }, + { + "type": "image_caption", + "bbox": [ + 0.474, + 0.239, + 0.85, + 0.251 + ], + "angle": 0, + "content": "Negative: A man is sitting on a chair with his hands above his knees" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.263, + 0.895, + 0.306 + ], + "angle": 0, + "content": "Figure 9. Left: Noisy negative images due to wrong attributes or objects generated by text-to-image models. Right: Noisy negative images caused by inappropriate bounding boxes or negative texts from LLMs. CLIP scores of generated images refer to the similarity between the box and the negative text compared to the positive text. Thresholding on CLIP scores remove those noisy images." + }, + { + "type": "image", + "bbox": [ + 0.132, + 0.334, + 0.411, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.41, + 0.445, + 0.424 + ], + "angle": 0, + "content": "Figure 10. Percentage of good generated negative images." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.44, + 0.47, + 0.581 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.607, + 0.47, + 0.636 + ], + "angle": 0, + "content": "Figure 11. Distributions of image regions from Omnilabel, \\(\\mathbf{D}^3\\), and our generated images. Visualization with t-SNE." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.665, + 0.184, + 0.68 + ], + "angle": 0, + "content": "ative texts only." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Concatenating images during training: Following the idea of concatenating the positive and negative captions as the text input, we concatenate the positive and negative images as one input image during training. See supplement for an example. In this way, models are forced to tell the difference between the positive and negative images within one training iteration, which helps detectors to learn better about the negative. As shown in Table 4, such a simple technique improves upon “+ Box&CLIP filters” both on OmniLabel and \\(\\mathbf{D}^{3}\\). Furthermore, we ensemble the weights of two FIBER models, one finetuned with negative texts only, and the other finetuned with both negative texts and images. Finally, compared to using negative texts only, we gain 1.3 APd on OmniLabel and no performance drop on \\(\\mathbf{D}^{3}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.331, + 0.892, + 0.589 + ], + "angle": 0, + "content": "Looking into generated images and benchmarks: Table 4 shows that negative images help on OmniLabel but not much on \\(\\mathrm{D}^3\\). We explore this on a data basis. We first crop image regions for generated images, OmniLabel images, \\(\\mathrm{D}^3\\) images, and Flickr30k images based on the bounding boxes. Then, we randomly select 1000 image regions and feed them into a CLIP image encoder to get CLIP embeddings. Later, we input those embeddings to t-SNE [48] to illustrate the similarities between different image regions. As shown in Fig. 11, \\(\\mathrm{D}^3\\)'s regions are grouped into several clusters, while OmniLabel and our generated regions are scattered in the center. This indicates that there is a clear domain gap between \\(\\mathrm{D}^3\\) and the others. Thus, it is plausible that our generated images only helps on OmniLabel. In our view, the gap comes from that \\(\\mathrm{D}^3\\) collect data in groups based on categories. In contrast, OmniLabel collects data randomly." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.603, + 0.619, + 0.618 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.84 + ], + "angle": 0, + "content": "Language-based detection requires localization of objects by a referring free-form text descriptions. To train accurate models in a discriminative way, the training data must contain good negative samples. Starting with an existing dataset, we propose (1) novel ways to prompt LLMs for generating additional negative texts, and (2) generating negative images to complement the training signal. Based on our experimental evaluations, we conclude that such additional negative training data indeed translates into improved detection accuracy on standard benchmarks. Our analysis demonstrates the importance of diversity in the generated text, which is higher with our approach than with prior works, and the quality of the generated images, which our proposed filtering steps can significantly increase." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgments: This research project has been partially funded by research grants to Dimitris N. Metaxas through NSF: 2310966, 2235405, 2212301, 2003874, and FA9550-23-1-0417." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13599" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.469, + 0.169 + ], + "angle": 0, + "content": "[1] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. UNITER: UNiversal Image-TExt Representation Learning. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.469, + 0.213 + ], + "angle": 0, + "content": "[2] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.215, + 0.469, + 0.282 + ], + "angle": 0, + "content": "[3] Zi-Yi Dou, Aishwarya Kamath, Zhe Gan, Pengchuan Zhang, Jianfeng Wang, Linjie Li, Zicheng Liu, Ce Liu, Yann LeCun, Nanyun Peng, et al. Coarse-to-Fine Vision-Language Pretraining with Fusion in the Backbone. In NeurIPS, 2022. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.284, + 0.469, + 0.352 + ], + "angle": 0, + "content": "[4] Sivan Doveh, Assaf Arbelle, Sivan Harary, Eli Schwartz, Roei Herzig, Raja Giryes, Rogerio Feris, Rameswar Panda, Shimon Ullman, and Leonid Karlinsky. Teaching Structured Vision & Language Concepts to Vision & Language Models. In CVPR, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.354, + 0.469, + 0.395 + ], + "angle": 0, + "content": "[5] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary Object Detection via Vision and Language Knowledge Distillation. In ICLR, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.397, + 0.469, + 0.436 + ], + "angle": 0, + "content": "[6] Agrim Gupta, Piotr Dólár, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In CVPR, 2019. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.439, + 0.469, + 0.492 + ], + "angle": 0, + "content": "[7] Ryota Hinami and Shin'ichi Satoh. Discriminative Learning of Open-Vocabulary Object Retrieval and Localization by Negative Phrase Augmentation. In EMNLP, 2018. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.495, + 0.469, + 0.549 + ], + "angle": 0, + "content": "[8] Cheng-Yu Hsieh, Jieyu Zhang, Zixian Ma, Aniruddha Kembhavi, and Ranjay Krishna. Sugarcrepe: Fixing hackable benchmarks for vision-language compositionality. In NeurIPS, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.551, + 0.469, + 0.59 + ], + "angle": 0, + "content": "[9] Ronghang Hu and Amanpreet Singh. UniT: Multimodal Multitask Learning with a Unified Transformer. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.594, + 0.469, + 0.66 + ], + "angle": 0, + "content": "[10] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, and Tom Duerig. Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision. In ICML, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.469, + 0.718 + ], + "angle": 0, + "content": "[11] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. MDETR - Modulated Detection for End-to-End Multi-Modal Understanding. In ICCV, 2021. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.469, + 0.774 + ], + "angle": 0, + "content": "[12] Aishwarya Kamath, Sara Price, Jonas Pfeiffer, Yann LeCun, and Nicolas Carion. TRICD: Testing Robust Image Understanding Through Contextual Phrase Detection. Technical report, NYU, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.469, + 0.829 + ], + "angle": 0, + "content": "[13] Zaid Khan, Vijay Kumar B.G., Xiang Yu, Samuel Schulter, Manmohan Chandraker, and Yun Fu. Single-Stream Multi-Level Alignment for Vision-Language Pretraining. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[14] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV, 123:32-73, 2017. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.146 + ], + "angle": 0, + "content": "[15] Weicheng Kuo, Yin Cui, Xiuye Gu, AJ Piergiovanni, and Anelia Angelova. F-vlm: Open-vocabulary object detection upon frozen vision and language models. In ICLR, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.231 + ], + "angle": 0, + "content": "[16] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, Tom Duerig, and Vittorio Ferrari. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. IJCV, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.233, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[17] Chunyuan Li, Haotian Liu, Liusan Harold Li, Pengchuan Zhang, Jyoti Aneja, Jianwei Yang, Ping Jin, Houdong Hu, Zicheng Liu, Yong Jae Lee, and Jianfeng Gao. ELEVATER: A Benchmark and Toolkit for Evaluating Language-Augmented Visual Models. In NeurIPS, 2022. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.302, + 0.892, + 0.356 + ], + "angle": 0, + "content": "[18] Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Deepak Gotmare, Shafiq Joty, Caiming Xiong, and Steven Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.358, + 0.892, + 0.425 + ], + "angle": 0, + "content": "[19] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In CVPR, pages 10965-10975, 2022. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.428, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[20] Liunian Harold Li, Zi-Yi Dou, Nanyun Peng, and Kai-Wei Chang. DesCo: Learning Object Recognition with Rich Language Descriptions. In NeurIPS, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.47, + 0.892, + 0.523 + ], + "angle": 0, + "content": "[21] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In CVPR, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.525, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[22] Chuang Lin, Peize Sun, Yi Jiang, Ping Luo, Lizhen Qu, Gholamreza Haffari, Zehuan Yuan, and Jianfei Cai. Learning object-language alignments for open-vocabulary object detection. In ICLR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.581, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C. Lawrence Zitnick. Microsoft COCO: Common Objects in Context. In ECCV, 2014. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.637, + 0.892, + 0.676 + ], + "angle": 0, + "content": "[24] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal Loss for Dense Object Detection. In ICCV, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.679, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[25] Chang Liu, Henghui Ding, and Xudong Jiang. GRES: Generalized Referring Expression Segmentation. In CVPR, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.72, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[26] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.79, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[27] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.847, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[28] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. ViL-BERT: Pretraining Task-Agnostic Visiolinguistic Representations for Vision-and-Language Tasks. In NeurIPS, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13600" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[29] Zixian Ma, Jerry Hong, Mustafa Omer Gul, Mona Gandhi, Irena Gao, and Ranjay Krishna. CREPE: Can Vision-Language Foundation Models Reason Compositionally? In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.148, + 0.47, + 0.204 + ], + "angle": 0, + "content": "[30] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L. Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In CVPR, 2016. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.469, + 0.232 + ], + "angle": 0, + "content": "[31] George A. Miller. Wordnet: A lexical database for english. Commun. ACM, 38(11):39-41, 1995. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.469, + 0.315 + ], + "angle": 0, + "content": "[32] Zhixiang Min, Bingbing Zhuang, Samuel Schulter, Buyu Liu, Enrique Dunn, and Manmohan Chandraker. Neurocs: Neural nocs supervision for monocular 3d object localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21404-21414, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.316, + 0.469, + 0.398 + ], + "angle": 0, + "content": "[33] Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, Xiao Wang, Xiaohua Zhai, Thomas Kipf, and Neil Houlsby. Simple Open-Vocabulary Object Detection with Vision Transformers. In ECCV, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.469, + 0.44 + ], + "angle": 0, + "content": "[34] Matthias Minderer, Alexey Gritsenko, and Neil Houlsby. Scaling Open-Vocabulary Object Detection. In NeurIPS, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.441, + 0.301, + 0.455 + ], + "angle": 0, + "content": "[35] OpenAI. Gpt-3.5. 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.456, + 0.469, + 0.525 + ], + "angle": 0, + "content": "[36] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k Entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In ICCV, pages 2641–2649, 2015. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.526, + 0.469, + 0.607 + ], + "angle": 0, + "content": "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.469, + 0.65 + ], + "angle": 0, + "content": "[38] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks. In NeurIPS, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.651, + 0.469, + 0.692 + ], + "angle": 0, + "content": "[39] Joshua Robinson, Ching-Yao Chuang, Suvrit Sra, and Stefanie Jegelka. Contrastive Learning with Hard Negative Samples. In ICLR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.469, + 0.762 + ], + "angle": 0, + "content": "[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.469, + 0.817 + ], + "angle": 0, + "content": "[41] Samuel Schulter, Vijay Kumar BG, Yumin Suh, Konstantinos M Dafnis, Zhixing Zhang, Shiyu Zhao, Dimitris Metaxas, et al. OmniLabel: A Challenging Benchmark for Language-Based Object Detection. In ICCV, 2023. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.872 + ], + "angle": 0, + "content": "[42] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Jing Li, Xiangyu Zhang, and Jian Sun. Objects365: A Large-scale, High-quality Dataset for Object Detection. In ICCV, 2019. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[43] Ravi Shekhar, Sandro Pezzelle, Yauhen Klimovich, Aurelie Herbelot, Moin Nabi, Enver Sangineto, and Raffaella" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Bernardi. FOIL it! Find One mismatch between Image and Language caption. In ACL, 2017. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[44] Abhinav Shrivastava, Abhinav Gupta, and Ross Girshick. Training Region-based Object Detectors with Online Hard Example Mining. In CVPR, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.163, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[45] Sanjay Subramanian, William Merrill, Trevor Darrell, Matt Gardner, Sameer Singh, and Anna Rohrbach. ReCLIP: A Strong Zero-Shot Baseline for Referring Expression Comprehension. In ACL, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.219, + 0.892, + 0.286 + ], + "angle": 0, + "content": "[46] Tristan Thrush, Ryan Jiang, Max Bartolo, Amanpreet Singh, Adina Williams, Douwe Kiela, and Candace Ross. Winoground: Probing Vision and Language Models for Visio-Linguistic Compositionality. In CVPR, 2022. 1, 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.288, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[47] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.371, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[48] Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9 (86):2579-2605, 2008. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.413, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[49] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. OFA: Unifying Architectures, Tasks, and Modalities Through a Simple Sequence-to-Sequence Learning Framework. In ICML, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.483, + 0.892, + 0.524 + ], + "angle": 0, + "content": "[50] Chenyun Wu, Zhe Lin, Scott Cohen, Trung Bui, and Subhransu Maji. PhraseCut: Language-based Image Segmentation in the Wild. In CVPR, 2020. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.525, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[51] Hao Wu, Jiayuan Mao, Yufeng Zhang, Yuning Jiang, Lei Li, Weiwei Sun, and Wei-Ying Ma. Unified Visual-Semantic Embeddings: Bridging Vision and Language with Structured Meaning Representations. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.58, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[52] Xiaoshi Wu, Feng Zhu, Rui Zhao, and Hongsheng Li. CORA: Adapting CLIP for Open-Vocabulary Detection with Region Prompting and Anchor Pre-Matching. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.636, + 0.892, + 0.692 + ], + "angle": 0, + "content": "[53] Chi Xie, Zhao Zhang, Yixuan Wu, Feng Zhu, Rui Zhao, and Shuang Liang. Described Object Detection: Liberating Object Detection with Flexible Expressions. In NeurIPS, 2023. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[54] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C. Berg, and Tamara L. Berg. Modeling Context in Referring Expressions. In ECCV, 2016. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[55] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language models behave like bags-of-words, and what to do about it? In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[56] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-Vocabulary Object Detection Using Captions. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[57] Shiyu Zhao, Zhixing Zhang, Samuel Schulter, Long Zhao, BG Vijay Kumar, Anastasis Stathopoulos, Manmohan Chandraker, and Dimitris N Metaxas. Exploiting unlabeled data with vision and language models for object detection. In ECCV, pages 159-175. Springer, 2022. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13601" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[58] Shiyu Zhao, Samuel Schulter, Long Zhao, Zhixing Zhang, Vijay Kumar B. G, Yumin Suh, Manmohan Chandraker, and Dimitris N. Metaxas. Improving pseudo labels for open-vocabulary object detection, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.217 + ], + "angle": 0, + "content": "[59] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, and Jianfeng Gao. RegionCLIP: Region-based Language-Image Pretraining. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.22, + 0.469, + 0.275 + ], + "angle": 0, + "content": "[60] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbihl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In ECCV, pages 350-368. Springer, 2022. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "13602" + } + ] +] \ No newline at end of file diff --git a/2024/Generating Enhanced Negatives for Training Language-Based Object Detectors/2970baed-7bcf-45c6-83b6-e79faa1f9b77_origin.pdf b/2024/Generating Enhanced Negatives for Training Language-Based Object Detectors/2970baed-7bcf-45c6-83b6-e79faa1f9b77_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..dab311d8a3335be7015028ea1c7d4056366841e4 --- /dev/null +++ b/2024/Generating Enhanced Negatives for Training Language-Based Object Detectors/2970baed-7bcf-45c6-83b6-e79faa1f9b77_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2f5ef806ef48dd4ba554e50383491aba15ac3add23f503a3d4ee4137e3bfe88f +size 2016487 diff --git a/2024/Generating Enhanced Negatives for Training Language-Based Object Detectors/full.md b/2024/Generating Enhanced Negatives for Training Language-Based Object Detectors/full.md new file mode 100644 index 0000000000000000000000000000000000000000..59c3f7a40d7a327bb578a5b6c417c3996b109091 --- /dev/null +++ b/2024/Generating Enhanced Negatives for Training Language-Based Object Detectors/full.md @@ -0,0 +1,306 @@ +# Generating Enhanced Negatives for Training Language-Based Object Detectors + +Shiyu Zhao $^{1,*}$ + +Long Zhao3 + +Vijay Kumar B G² + +Yumin Suh2 + +Dimitris N. Metaxas1 + +Manmohan Chandraker2,4 + +Samuel Schulter2 + +$^{1}$ Rutgers University + +$^{2}$ NEC Laboratories America + +3 Google Research + +4 UC San Diego + +# Abstract + +The recent progress in language-based open-vocabulary object detection can be largely attributed to finding better ways of leveraging large-scale data with free-form text annotations. Training such models with a discriminative objective function has proven successful, but requires good positive and negative samples. However, the free-form nature and the open vocabulary of object descriptions make the space of negatives extremely large. Prior works randomly sample negatives or use rule-based techniques to build them. In contrast, we propose to leverage the vast knowledge built into modern generative models to automatically build negatives that are more relevant to the original data. Specifically, we use large-language-models to generate negative text descriptions, and text-to-image diffusion models to also generate corresponding negative images. Our experimental analysis confirms the relevance of the generated negative data, and its use in language-based detectors improves performance on two complex benchmarks. Code is available at https://github.com/xiaofeng94/Gen-Enhanced-Negs. + +# 1. Introduction + +Using natural language in object detection to describe semantics bears the potential to significantly increase the size of the detector's label space and enable novel applications. While standard detectors operate on a fixed label space [23, 38, 42], natural language allows for a broad spectrum of object descriptions, ranging from generic terms like "vehicle" to specific expressions like "the red sports car parked on the left side" [12, 17, 25, 30, 41, 53, 54]. Several works advanced language-based object detection over the past few years with novel training strategies [3, 5, 19, 22, 32, 34, 57] and model architectures [11, 15, 33, 45]. + +Referring expression or visual grounding datasets [14, 30, 36, 50, 54] provide the natural language object descrip + +![](images/f8eef1088602cc2133a650ab43cb86096fa93ebe24f8a1447afcce1c4b32bf07.jpg) + +![](images/5fbc2afc988111bfc60b08b82abf9de8d5b339cae4a76cde954b770724e371cb.jpg) +Figure 1. The key contribution of our work is to leverage large-language-models and text-to-image diffusion models to automatically generate negative object descriptions and images for training language-based object detectors. In contrast to prior work, our generated negatives are more relevant to the original data and provide a better training signal for detectors. + +tions along with bounding box annotations needed for training. However, this data only describes what is present in the images, but misses to describe what is not. Yet, the notion of negatives is crucial for training discriminative models like language-based detectors [7, 24, 39]. + +Detection datasets with a fixed label space provide negative classes implicitly or explicitly, with exhaustive [23, 42] or federated [6, 16] annotations, respectively. Any part of an image that does not overlap significantly with a bounding box of category $c$ is verified to not be of that category (for exhaustive annotation). On the other hand, the space of negatives for a free-form text description of an object is extremely large. While some existing datasets provide negative samples in free-form text [43, 46], they were not annotated with bounding boxes. Hence, existing language-based detectors often define the negatives for one object as the descriptions of all other objects in the same image or descriptions of other random samples [3, 11, 19]. However, such negatives may not be directly related to the original positive description and define a weaker training signal (see Fig. 1). By explicitly evaluating on human-curated negatives, a re + +cent benchmark [41] identified a bias of existing language-based detectors to perform clearly better on positive rather than negative descriptions. However, creating a dataset with high-quality human-curated negatives for large-scale training is labor-intensive and costly. + +In this work, we propose to explicitly and automatically generate negative data in the form of free-form texts as well as images. Prior works [4, 7, 29, 43, 46, 55] rely on rule-based approaches with knowledge graphs and focus only on the language domain or the classification task. In contrast, we leverage generative large-language-models (LLMs) [35, 47] and text-to-image diffusion models [21, 40] to automatically create relevant but contradicting object descriptions along with the corresponding images for language-based object detection, see Fig. 1. + +Given an object description of a dataset, we first use LLMs to generate a semantically contradicting description as the negative. Besides changing individual words (foils) based on explicit knowledge graphs or LLMs, like in prior work [4, 7, 20], we demonstrate improved detection performance with two alternative approaches. (Re-combination): An LLM first identifies all objects in a sentence, and then creates a contradicting one by re-arranging, ignoring or adding objects. (In-context summaries): We prompt an LLM to summarize the differences of a few (less than 100) positive-negative pairs collected from an existing image-level dataset [46]. This summary is then used as context to generate more such examples. Note that we do not need visual input for this step, allowing us to leverage powerful LLMs for semantic and textual reasoning. Moreover, while prior work only focused on the text [4, 7, 29, 43, 55], we also leverage text-to-image diffusion models like GLIGEN [21] to create images that match the generated negative descriptions of objects, which serves as additional training signal. While the direct output of such image-generation models is often noisy and even wrong (not matching the input description), we propose two filtering steps to reduce noise considerably (from $53\%$ to $16\%$ according to an empirical study). Having both negative object descriptions and the corresponding image, allows us to improve the discriminative loss for training language-based object detectors. + +Our experiments demonstrate clear accuracy gains on two challenging benchmarks, +2.9AP on OmniLabel [41] and +3.3AP on $\mathrm{D}^3$ [53], when adding our automatically-generated negative data into the training of baseline models like GLIP or FIBER. Moreover, we provide an in-depth analysis of the generated data (text and images) and how they contribute to better language-based detection. + +Summary of contributions: (1) Automatic generation of semantically relevant but contradicting negative text and images with large-scale generative models. (2) Recipes to integrate such negative data into language-based detection models like FIBER [3] and GLIP [19] (3) Clear improve + +ments on language-based detection benchmarks [41, 53] including a thorough analysis of the generated data. + +# 2. Related work + +Vision & language localization tasks: Open-vocabulary detection (OVD) requires a model to localize object category names without having seen explicit bounding box annotation for them [5, 7, 15, 52, 56, 59]. In contrast, we focus on the more general language-based object detection task [41, 53], which goes beyond simple category names. Referring expression comprehension (REC) aims at localizing the subject of a free-form text expression. However, REC benchmarks [30, 50, 54] fall short in evaluating all aspects of the more general language-based detection task [41, 53]. In visual grounding (VG) [36], the task is to localize noun phrases of a caption in the image. Although being a task on its own, VG datasets have recently been used mostly as training data for OVD. Our work focuses on general language-based object detection, which subsumes and generalizes standard detection, OVD and REC [12, 25, 41, 53]. + +Language-based object detectors: Two critical abilities of language-based detection are accurate localization and tight text-image fusion. Works like [1, 9, 28] use language-models like BERT [2, 27] to align regions extracted from (pre-trained) detectors with captions. The outstanding zero-shot classification accuracy of large-scale pre-trained models like CLIP [37] or [10, 13, 18] then sparked interest in extensions for localization, with different approaches like distillation [5], fine-tuning [15, 33], pseudo-labeling [34, 57, 58], or combinations thereof [3, 19]. We use such models as test bed, but explore the underlying training data with respect to negative samples. + +Negative samples for object detection: The notion of negatives is crucial for training discriminative models [24, 39]. Also for object detection, hard negative mining [44] has proven beneficial for model training. However, these prior works aim to find hard negative training examples rather than negatives in the label space, because the label space is fixed in standard detection. For language-based datasets, the space of potential negatives is extremely large because object descriptions are free-form text. Prior works [4, 29, 43, 46, 55] investigate negative texts for general vision & language models with different strategies, including changing individual words (foil) with rules based on knowledge graphs [31] or with LLMs. Sugar-Crepe [8] shares a similar idea as us to get negative texts with in-context learning but for image-text level pretraining. For language-based detection, [7] explores such rule-based foils, while [20] uses LLMs with specific templates to replace object names with alternative descriptions. In contrast, our work (1) focuses on the localization task, (2) ex + +![](images/676445f31ce25fdf5841a8cb2598a6422f82f742803a1a4a1da667cbc67d81fd.jpg) +Figure 2. In language-based object detection, a detector receives as input an image and a (variable-length) list of free-form text descriptions of objects. For each description, the model predicts bounding boxes for objects that match the description. + +plies more comprehensive strategies to generate negatives with LLMs, and (3) proposes to also generate corresponding negative images with text-to-image diffusion models. + +# 3. Method + +# 3.1. Language-based object detection + +Task definition: Given an image and a list of object descriptions, the task is to output bounding boxes along with confidence scores for each description, as shown in Fig. 2. Note the multi-label setting where one object instance can be referred to by multiple descriptions, like "person" and "person looking at book". Also note that an object description might refer to zero objects in the image and the desired output is an empty set of boxes. + +Training data: Many language-based detection models [3, 19] use a combination of object detection [23, 42] and visual grounding datasets [14, 36] for training their models. Both types of datasets provide images $I$ and bounding boxes $b_{l}$ to localize individual objects. Object detection data assigns each bounding box $b_{l}$ a unique category $c$ out of fixed label space $\mathcal{C}$ . The exhaustive labeling of the fixed label space in detection datasets implies the space of negatives. An object of category $c$ is not any of the categories $\mathcal{C} \setminus c$ . On the other hand, grounding data provides an image caption $t$ in free-form text, where subsets of words $m_{l}$ (defined as indices of starting and ending characters in $t$ ) are linked to bounding box $b_{l}$ . For grounding data, the space of negatives is extremely large because one can find as many textual descriptions that do not match $t$ as desired due to the compositionality of free-form text. Many language-based detectors [3, 19] only use the words in $t$ that are not referred to by $m_{l}$ as negatives for the bounding box $b_{l}$ . We argue that this choice is sub-optimal because these words may refer to entirely different objects and are easy to discriminate. In the following section, we explain how we can automatically generate negative samples that semantically related to the original text $t$ and hence provide a better training signal. + +# 3.2. Generating negative samples + +Our goal is to automatically and explicitly generate negative samples based on the original text descriptions $t$ to im + +prove the training signal for language-based detectors. A key observation of our work is to leverage the vast knowledge encoded into large-language models (LLMs) [35, 47] and text-to-image diffusion models (T2I) [21, 40]. Besides proposing novel ways to instruct LLMs for generating negative text descriptions (Sec. 3.2.1), we also propose to generate negative images (Sec. 3.2.2). + +# 3.2.1 Generating negative descriptions with LLMs + +Given an object description $t$ that matches the visual content inside a bounding box $b_{l}$ , we define a "negative" description $t'$ as any text that is semantically different to the original text. Furthermore, our intuition is that good negative descriptions are still semantically related to the original description, but not the same. An example is: "Person in red shirt" as the original description and "Person in blue shirt" as a contradicting negative one. + +Prior work [4, 7, 43] explored rule-based approaches to generate negative text. However, such rules are typically limited to simple knowledge graphs and are limited to replacing only individual words, often just nouns, or swapping words. In contrast, we explore more powerful LLMs to automatically generate relevant negatives. To make the negative text generation efficient and economic, in all cases, we first leverage a strong instruction-tuned LLM [35] to generate $50\mathrm{k}$ positive-negative pairs, and then finetune a LLaMA-7B [47] model with those pairs to then generate negative captions on large grounding datasets. In the following, we describe three ways to instruct an LLM for generating positive-negative pairs of object descriptions: + +LLM-based foils: We first prompt an instruction-tuned LLM [35] to find concepts (i.e., objects, attributes and relationships) in object descriptions. Compared to rule-based parsers [51], LLMs can provide richer information. For example, for the caption "A transportation vehicle is carrying a crowd of people who are sitting and standing," the parser ignores "sitting" and "standing", while LLMs regard them as attributes. Then, we pick one concept from the first step sequentially and prompt LLMs again to generate a negative caption by changing the concept. For both steps, the prompts are manually curated with the task definition and step-by-step instructions for the generation. Please find the exact prompt for the LLM in the supplement. + +Re-combination: Next, we give the LLM more freedom in generating negative descriptions. We first prompt the LLM to identify all objects in the original caption, and then to recombine them to create a new sentence different from the original one. We allow the LLM to ignore, change or add new objects. For example, given the caption “A boy is playing with his dog” and two objects “boy” and “dog”, the LLM can output “The girl and her dog are playing fetch in + +the park". Detailed prompts for both identifying objects and re-combination are in the supplement. + +In-context summary: Third, we enable LLMs to learn how to generate negative descriptions by providing human-annotated positive-negative pairs as in-context samples. We randomly sample 80 pairs of positive and negative texts from the Winoground dataset [46] and prompt the instruction-tuned LLM [35] to summarize the difference of those pairs in plain text. Then, instead of manually creating prompts to generate positive-negative pairs, we leverage the summary together with three randomly sampled Winoground pairs as prompts to the LLM, and generate several positive-negative pairs to finetune LLaMA. After finetuning, the LLaMA model is used to generate negative texts for given descriptions. This pipeline does not require handcrafted prompts to LLMs as the explanation of the concept of negatives and how to create them. The supplement contains full prompts for generating a summary, and generating positive-negative pairs for finetuning. + +# 3.2.2 Generating negative images with T2I models + +Given an original image $I$ , a bounding box $b$ and a corresponding object description $t$ , we define a negative image $I'$ as any image that has a different semantic content inside $b$ . The rest of the image can be equivalent to $I$ . To obtain such imagery, we start with visual grounding data that provide bounding boxes, positive captions with text phrases, and alignment between them. We propose a two-step process: First, we turn the positive caption into a negative one. Second, we use conditioned image generation tools to alter the visual content inside the bounding box $b$ . + +Negative text for negative images: Although we have already described an approach to generate negative descriptions in Sec. 3.2.1, doing so to generate a negative image requires a different approach. In this case, the generated negative text needs to preserve the alignment $m_{l}$ to the ground truth bounding box $b_{l}$ in order to instruct the generative image model GLIGEN [21]. Hence, we first select a bounding box $b_{l}$ and mask out the corresponding words (known via $m_{l}$ ) in the text $t$ . For example, "A boy is playing with his dog" turns into "A boy is playing with [Mask]" if the selected bounding box refers to "his dog". Again, we leverage LLMs [35] to fill in text for "[Mask]" to generate a negative text without reusing the original text. Please refer to Fig. 3 for illustrations. + +We finetune a LLaMA-7B for the mask filling task with triplets of positive texts, masked texts, and negative texts. To reduce manual efforts, we follow the approach of in-context summary to get triplet samples. We apply this process twice: We start with only 5 manually created triplets to build a summary and generate 100 samples from the + +![](images/03da1533848a52141bebfc703c084b2abdf51c14c1354ec0fbb83913137aeac8.jpg) +Figure 3. Overview of using LLMs [35, 47] and text-to-image diffusion models [21, 40] to generate negative images. + +LLM [35] with human checks. We then repeat the process to generate 50k examples without human checks from a summary of the 100 generated examples. This increases diversity in the generated data. + +Conditional image generation: Given an image $I$ , a bounding box $b$ and the altered text $t'$ , we generate a negative image $I'$ that is equal to $I$ except inside $b$ , where the visual content is altered to match the text $t'$ . To do so, we use the inpainting and conditioning abilities of GLIGEN [21], a T2I model [40]. Refer to [21] and our supplemental material for more details, and to Fig. 3 for an illustration of the process. + +Mitigating noise in image generation: We found that the generated images are often noisy for any of the following reasons: (1) The altered text refers to a big bounding box that covers other smaller boxes. Large portions of the image are then generated and often do not match the concepts those smaller boxes originally covered. (2) The generated negative text does not match the bounding box that is either too small, too large or in an inappropriate position. (3) The T2I model fails to understand the negative text and generates wrong content. We propose two steps to filter such noisy images. First, we simply ignore ground truth boxes $b_{l}$ for image generation if the box covers more than $75\%$ of any other boxes in the image. Second, we adopt CLIP [37] to verify the semantic similarity of the generated image regions and the corresponding text. Specifically, we compute the similarity with CLIP between the generated image region (visual input) and the original and generated negative texts (text input). We filter out generated images that have a similarity score to the generated negative text lower than a user-defined threshold. Details on the filtering steps are given in the supplemental. + +# 3.3. Learning from negative samples + +Detector design and training objective: The generated data does not prescribe any specific architecture for the detector. A common choice, which we also use for our exper + +![](images/81aba8dd6003f4b8376745bf454bc4710d45a6a7309deccd0ccb3d7788986279.jpg) +Figure 4. Illustration of the grounding loss used for training. Predictions that are matched with ground truth receive a positive signal from the associated text (tall rectangles). All other words receive a negative signal (short rectangles). The top left quarter shows the original loss used in [3, 19]. The other three quarters are related to our proposed generated negative data and provide additional training signals. + +iments, is [3, 19]. The inputs are image $I$ and text $t$ , and the output is a set of bounding boxes $\hat{b}_i$ with corresponding logits $\hat{p}_i \in \mathbb{R}^T$ . Here, $T$ is the number of tokens required to represent text $t$ . The ground truth can be represented by a binary assignment matrix $\mathbf{A} \in \mathbb{B}^{L \times T}$ . Rows refer to ground truth boxes $l$ and columns to tokens in $t$ . Each element indicates if a token corresponds to a box $l$ , which is given by the ground truth indices $m_l$ . To define a loss, bipartite graph matching associates predictions with ground truth. For matched predictions, the target vector $g_i \in \mathbb{B}^T$ is the corresponding row from $\mathbf{A}$ , while it is an all-zero vector for unmatched targets. The loss is then computed as $\mathcal{L} = \sum_{i} \ell_{\mathrm{FL}}(\hat{p}_i, g_i)$ where FL refers to a focal loss. Fig. 4 illustrates the loss. + +Integrating negative text: When sampling an image $I$ , along with text $t$ , boxes $b_{l}$ and indices $m_{l}$ , we also randomly sample $K > 1$ negative descriptions from $\{t_j^{\prime}\}$ that defines the pool of negatives generated for text $t$ . We randomly shuffle the order of all texts to avoid any biases on the location of the one positive description, and then concatenate them into one text string. + +Integrating negative images: We explore two options: (1) Simply add the generated images $I'$ along with their generated (but semantically matching) captions $t'$ as additional visual grounding data. The original caption $t$ , which was the starting point to generate the negative image $I'$ , is now used as the negative caption. In this way, both the original image $I$ and the generated one $I'$ have positive and negative descriptions. This option is illustrated in Fig. 4. (2) To better leverage the relation between the original and generated data, the second option is to pack them into a single training sample. We simply concatenate the images $I$ and $I'$ , as well as the texts $t$ and $t'$ . The ground truth information $m_{l}$ is updated accordingly. See supplement for details. + +
OmniLabelOmniLabel-Negative
APAPcAPdAPdPAPAPcAPdAPdP
Detic [60]8.015.65.48.0----
MDETR [11]--4.79.1----
GLIP-T [19]19.323.616.425.813.924.89.626.1
+ Ours22.227.218.829.016.528.611.630.2
FIBER-B [3]25.730.322.334.818.731.213.336.3
+ Ours28.132.125.136.522.333.316.738.3
+ +Table 1. Evaluation on the OmniLabel [41] benchmark. + +# 4. Experiments + +# 4.1. Experimental design + +Training procedure: We choose two recent methods, GLIP-T [19] and FIBER-B [3], to demonstrate the effect of our automatically generated negatives. We use the official code and publicly available checkpoints as a starting point. The Flickr30k dataset [36] serves as our grounding dataset to generate the negative data. We then fine-tune GLIP-T and FIBER-B with both positive and negative data, along with the Objects365 detection dataset [42] for 1 epoch. Note that both Objects365 and Flickr30k are part of the original training set. We do not introduce any extra data except our generated negatives. Most hyper-parameters are equal to the original settings of GLIP and FIBER. Any exceptions are described in the supplement. + +Evaluation benchmarks: We choose two recently proposed benchmarks, OmniLabel [41] and $\mathrm{D}^3$ [53], as our test beds. These benchmarks evaluate more aspects of language-based detection than existing referring expressions [30, 50, 54] or open-vocabulary detection [6, 17] benchmarks. Specifically, both benchmarks contain complex object descriptions that go beyond simple category names from open-vocabulary detection benchmarks. Moreover, the descriptions can refer to zero, one or multiple instances in the image, in contrast to standard referring expression benchmarks. These properties enable a more stringent evaluation metric as in object detection, which is based on average precision (AP) in both OmniLabel [41] and $\mathrm{D}^3$ [53]. Both benchmarks provide more fine-grained metrics. OmniLabel evaluates separately for categories, descriptions, and descriptions referring to at least one object, with APc, APd and APd-P, respectively. $\mathrm{D}^3$ differentiates descriptions on absence ("Abs") and presence ("Pres") that indicate whether or not they contain any form of negation (e.g., "without"), as well as on text lengths. Finally, we create a specific split for OmniLabel, "OmniLabel-Negative", to evaluate the model only on images that contain at least one negative description (i.e., not referring to any object). + +
D3(default)D3(by length of texts)
FullPresAbsSMLXL
OFA-L [49]4.24.14.64.95.43.02.1
OWL-ViT-L [33]9.610.76.420.79.46.05.3
G-DINO-B [26]20.720.122.522.622.518.916.5
OFA-DOD [53]21.623.715.423.622.620.518.4
GLIP-T [19]19.118.321.522.422.016.610.6
+ Ours21.420.623.728.124.517.411.5
FIBER-B [3]22.721.526.030.125.917.913.1
+ Ours26.025.228.135.529.720.514.2
+ +Table 2. Evaluation on the ${\mathrm{D}}^{3}$ [53] benchmarks. + +# 4.2. Benchmark comparisons + +Tabs. 1 and 2 evaluate the impact of our generated negative training data on the OmniLabel [41] and $\mathrm{D}^3$ [53] benchmarks. In both tables, the first set of rows are baselines provided by the benchmarks. The following rows show the main comparisons for GLIP-T [19] and FIBER-B [3] with and without adding our generated negative training data. First, we can see that adding negative data improves results across all metrics for both models and both benchmarks. On OmniLabel, we can see a $+2.9\%$ and $+2.4\%$ increase in AP for GLIP-T and FIBER-B, respectively. Similarly, we observe a $+2.3\%$ and $+3.3\%$ increase in the main metric of $\mathrm{D}^3$ (AP on full descriptions) for GLIP-T and FIBER-B. + +# 4.3. Analysis on negative texts + +Effectiveness of different negative texts: We finetune FIBER-B without and with different kinds of negative texts mentioned in Sect. 3.2.1, i.e., Rule-based foils, LLM-based foils, Re-combination with LLMs, In-context summary with LLMs, and present results in Table 3. We find all kinds of negatives improve the original FIBER-B on both OmniLabel and $\mathrm{D}^3$ benchmarks. Negative texts from LLMs generally achieve better results compared to LLM-based foils, which indicates that LLMs are powerful tools for negative text generation. Moreover, both recombination and in-context summary with LLMs outperform LLM-based foils in all metrics except APd-P. Note that APd-P refers to evaluations without negative label spaces, which is a task weaker than language-based detection. Based on such results, we argue that although word foils provide promising results in traditional studies [7, 43], it is sub-optimal to LLMs. We need to explore varied ways to unlock the ability of LLMs. We believe that our two solutions, i.e., Re-combination and In-context summary, provide a good starting point for future studies. Besides using only one kind of negative texts, we also explore the combinations of different kinds of negative texts in the supplement. + +Diversity of rule-based and LLM-based negatives: In this part, we investigate the diversity of different negative + +![](images/c0fb8d5c75d1f665d231f42d5e6db5374b87fda6c384f5ba7d1e78e28c3e962a.jpg) + +![](images/f19c8c37a648ad25643a81bb76ce8a85202db15430132966d7e379191708f518.jpg) + +![](images/32a8a4eb4a89f6e7d559893c1f8676b7156b60b9ce9aab9af5aad8c81b042070.jpg) +Figure 5. Percentage of negative texts with the numbers of words. Four negative generation methods are compared. + +![](images/c98dec11641bcb54e462f9f9cd49fc721e1e2963baeff1ca563ac65680933579.jpg) + +![](images/22fdd28b65468b385aad3e33355b3b02e56efc81147ed3a405912ccb3340ede2.jpg) + +![](images/40bc52d2ee23dc901a4dd97c95be0116dec585d402ef605139164ea487e7217f.jpg) + +![](images/aada24c5ca7f4a935dfd3425784770fda8e248bdfdb3d8180c83505f116591bf.jpg) +Figure 6. Percentage of negative texts with the numbers of words that are different from the original caption. Four negative generation methods are compared. + +![](images/a0f704dafcd969b164daf0ee64cb64dc4f62c227f26e5f8adbf6aa4dd4d2cfc0.jpg) + +![](images/9aa5fb75fda56cd538964bc03f29b59fb6de06347c09a149c787da01948a6bdf.jpg) +Figure 7. Average numbers of extra unique words per thousand generated negative texts, which are not included in the original dataset. We group words by their part-of-speech. + +texts. First, we count number of words for each negative text and provide the distribution for negatives of different sources in Fig. 5. As shown, all four distributions have a peak around 10 words, but the one of rule-based foils is higher than others. That means rule-based foils provide more negative texts with similar lengths. + +Second, we count the number of different words between the original positive caption and the negative caption, and present the distributions in Fig. 6. We find that LLM-based methods usually changes more words than rule-based foils, which increases the diversity. Moreover, in-context sum + +
Whole OmniLabelOmniLabel-NegativeD3
APAPcAPdAPd-PAPAPcAPdAPd-PFullPresAbs
Original FIBER-B25.730.322.334.818.731.213.336.322.721.526.0
+ Rule-based foils26.431.722.634.919.232.613.636.424.123.226.9
+ LLM-based foils26.530.723.335.920.832.115.438.024.624.026.5
+ Re-combination26.930.823.935.921.132.315.637.625.324.627.3
+ In-context summary26.630.823.434.221.132.215.736.425.725.227.5
+ +Table 3. Performance of FIBER-B trained with negative texts from four negative generation methods. + +
OmniLabelD3
APcAPdAPd-P
Original FIBER30.322.334.822.7
FIBER w/ neg. texts30.723.935.525.9
+ W/ neg. img. directly30.122.433.723.0
+ Box filter31.023.835.423.6
+ Box&CLIP filters31.124.235.924.1
+ Above + concat. img.31.724.835.924.8
+ Above + weight ensemble32.125.236.526.0
+ +Table 4. FIBER trained with negative images. + +![](images/245a9dd89cbd49cca57c28c1b3f7df453277347f71c5dc4dd4831db6e614e365.jpg) +Figure 8. Noisy generated images. The orange box contains the red box, and editing the orange changes the red unexpectedly. + +mary has a more flat distribution compared to others. Probably, in-context summary learns how to generate negatives automatically from data and has less restrictions. Besides, in-context summary has more cases with no word changed where negative texts are generated by just shuffling words or concepts in the original text. Such shuffling is a common pattern of Winoground [46], and our in-context summary can learn such data specific patterns. + +Third, we count how many extra words that does not exist in the original Flickr30k dataset are introduced in different negative generation methods. Fig. 7 shows the average number of extra words per 1000 negative texts. We group words into four part-of-speech categories, i.e., VERB, NOUN, ADP/ADJ, and others. As shown, LLMs introduce more extra words on average than rule-based foils probably because rule-based foils are limited in a predefined set of words. However, LLMs are open to any concepts and have great potentials of generating diverse texts. In-context summary introduces the most extra words for all categories, which is likely a benefit of learning negative generation from data. The above statistics indicate a clear view that LLMs generate more diverse data than rule-based foils. + +# 4.4. Analysis on negative images + +Noise in generated images: As mentioned in the last paragraph of Sect. 3.2.2, the raw generated images are noisy in + +several ways. First, the editing of a large box will override the context of smaller boxes that are covered by the large box. As shown in Fig. 8, GLIGEN did follow the instruction to generate a boy in the orange box, but the black jacket in the red box is missing. As a remedy, we apply our first de-noise step "Box Filter". That is, we ignore boxes that contain any other boxes when generating negative images. Second, GLIGEN may generate contents with wrong attributes or objects, as shown in Fig. 9 (Left). Moreover, our generation pipeline includes some cases where the edited text and the bounding box does not match. As shown in Fig. 9 (Right), the box for "his lap" cannot be modified as "his knees". Thus, GLIGEN generates wrong contents. As described in Sect. 3.2.2, we adopt a pretrained CLIP model to judge if generated contents are correct, which mitigates the noise to some extent. As shown in Fig. 9, both negative images get low CLIP scores and can be filtered out with a threshold. We call such thresholding "CLIP Filter". + +Subject studies on Box and CLIP filters: We employ human experts to check the amount of noisy generated images. First, for negative images w/o filter, w/ Box filter, and w/ Box&CLIP filters, we separately and randomly select 100 samples. Then, we ask two experts to check if a negative image is not noisy by comparing it with its caption and the original positive image. We regard an image as not noisy when both experts agree. As shown in Fig. 10, both filters reduce the noise. The Box filter improves from $47\%$ to $63\%$ , and the CLIP filter improves to $84\%$ . + +Effectiveness of generated negative images: To show the effectiveness of generated images themselves, we take captions of generated images as additional negative texts to in-context summary, and finetune a FIBER model as baseline. Then, we compare the baseline with variants of adding generated negative images in Table 4. As shown, the performance drops if we directly take raw negative images as new visual grounding data without any filters (i.e., W/ neg. img. directly). Probably, there are too much noise in raw negative images as shown in Fig. 10. When applying both Box and CLIP filters on negative images, we can achieve slight improvement on OmniLabel compared to using neg- + +![](images/23d9566b268ef8858588220de2689f124b5bc29f26ad20625525994e21a72dc5.jpg) +Original +Positive: A woman in a black mesh skirt plays acoustic guitar +Negative: A woman in a white lace dress plays acoustic guitar + +![](images/853c0d3bb4871d4e8bb3999eac24b1ba30064c504d1b38b65e7ea4b55255f775.jpg) +Generating wrong attributes +Generated +Positive: A man is sitting on a chair with his hands above his lap +Negative: A man is sitting on a chair with his hands above his knees + +![](images/d6c9a61e13381799fe4271559890e473fe0136d9313c276f643b6b34624b3e78.jpg) +Generation with inappropriate boxes or texts +Original + +![](images/d4ced8cb2308daaebd82cdfa84be84dd766c36a1cbfccdf827327ecddd3d0564.jpg) +Generated + +![](images/173612794985e0903a31e2b35b3de7826df42b76b2af7e84e7b62e54a5b38c97.jpg) +Figure 9. Left: Noisy negative images due to wrong attributes or objects generated by text-to-image models. Right: Noisy negative images caused by inappropriate bounding boxes or negative texts from LLMs. CLIP scores of generated images refer to the similarity between the box and the negative text compared to the positive text. Thresholding on CLIP scores remove those noisy images. +Figure 10. Percentage of good generated negative images. + +![](images/7676b5cf758c51a2742d1dc7942bb7ce5cc5a26580ce107b1fd93fd31331f902.jpg) +Figure 11. Distributions of image regions from Omnilabel, $\mathbf{D}^3$ , and our generated images. Visualization with t-SNE. + +ative texts only. + +Concatenating images during training: Following the idea of concatenating the positive and negative captions as the text input, we concatenate the positive and negative images as one input image during training. See supplement for an example. In this way, models are forced to tell the difference between the positive and negative images within one training iteration, which helps detectors to learn better about the negative. As shown in Table 4, such a simple technique improves upon “+ Box&CLIP filters” both on OmniLabel and $\mathbf{D}^{3}$ . Furthermore, we ensemble the weights of two FIBER models, one finetuned with negative texts only, and the other finetuned with both negative texts and images. Finally, compared to using negative texts only, we gain 1.3 APd on OmniLabel and no performance drop on $\mathbf{D}^{3}$ . + +Looking into generated images and benchmarks: Table 4 shows that negative images help on OmniLabel but not much on $\mathrm{D}^3$ . We explore this on a data basis. We first crop image regions for generated images, OmniLabel images, $\mathrm{D}^3$ images, and Flickr30k images based on the bounding boxes. Then, we randomly select 1000 image regions and feed them into a CLIP image encoder to get CLIP embeddings. Later, we input those embeddings to t-SNE [48] to illustrate the similarities between different image regions. As shown in Fig. 11, $\mathrm{D}^3$ 's regions are grouped into several clusters, while OmniLabel and our generated regions are scattered in the center. This indicates that there is a clear domain gap between $\mathrm{D}^3$ and the others. Thus, it is plausible that our generated images only helps on OmniLabel. In our view, the gap comes from that $\mathrm{D}^3$ collect data in groups based on categories. In contrast, OmniLabel collects data randomly. + +# 5. Conclusion + +Language-based detection requires localization of objects by a referring free-form text descriptions. To train accurate models in a discriminative way, the training data must contain good negative samples. Starting with an existing dataset, we propose (1) novel ways to prompt LLMs for generating additional negative texts, and (2) generating negative images to complement the training signal. Based on our experimental evaluations, we conclude that such additional negative training data indeed translates into improved detection accuracy on standard benchmarks. Our analysis demonstrates the importance of diversity in the generated text, which is higher with our approach than with prior works, and the quality of the generated images, which our proposed filtering steps can significantly increase. + +Acknowledgments: This research project has been partially funded by research grants to Dimitris N. Metaxas through NSF: 2310966, 2235405, 2212301, 2003874, and FA9550-23-1-0417. + +# References + +[1] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. UNITER: UNiversal Image-TExt Representation Learning. In ECCV, 2020. 2 +[2] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019. 2 +[3] Zi-Yi Dou, Aishwarya Kamath, Zhe Gan, Pengchuan Zhang, Jianfeng Wang, Linjie Li, Zicheng Liu, Ce Liu, Yann LeCun, Nanyun Peng, et al. Coarse-to-Fine Vision-Language Pretraining with Fusion in the Backbone. In NeurIPS, 2022. 1, 2, 3, 5, 6 +[4] Sivan Doveh, Assaf Arbelle, Sivan Harary, Eli Schwartz, Roei Herzig, Raja Giryes, Rogerio Feris, Rameswar Panda, Shimon Ullman, and Leonid Karlinsky. Teaching Structured Vision & Language Concepts to Vision & Language Models. In CVPR, 2023. 2, 3 +[5] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary Object Detection via Vision and Language Knowledge Distillation. In ICLR, 2022. 1, 2 +[6] Agrim Gupta, Piotr Dólár, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In CVPR, 2019. 1, 5 +[7] Ryota Hinami and Shin'ichi Satoh. Discriminative Learning of Open-Vocabulary Object Retrieval and Localization by Negative Phrase Augmentation. In EMNLP, 2018. 1, 2, 3, 6 +[8] Cheng-Yu Hsieh, Jieyu Zhang, Zixian Ma, Aniruddha Kembhavi, and Ranjay Krishna. Sugarcrepe: Fixing hackable benchmarks for vision-language compositionality. In NeurIPS, 2023. 2 +[9] Ronghang Hu and Amanpreet Singh. UniT: Multimodal Multitask Learning with a Unified Transformer. In ICCV, 2021. 2 +[10] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, and Tom Duerig. Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision. In ICML, 2021. 2 +[11] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. MDETR - Modulated Detection for End-to-End Multi-Modal Understanding. In ICCV, 2021. 1, 5 +[12] Aishwarya Kamath, Sara Price, Jonas Pfeiffer, Yann LeCun, and Nicolas Carion. TRICD: Testing Robust Image Understanding Through Contextual Phrase Detection. Technical report, NYU, 2023. 1, 2 +[13] Zaid Khan, Vijay Kumar B.G., Xiang Yu, Samuel Schulter, Manmohan Chandraker, and Yun Fu. Single-Stream Multi-Level Alignment for Vision-Language Pretraining. In ECCV, 2022. 2 +[14] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV, 123:32-73, 2017. 1, 3 + +[15] Weicheng Kuo, Yin Cui, Xiuye Gu, AJ Piergiovanni, and Anelia Angelova. F-vlm: Open-vocabulary object detection upon frozen vision and language models. In ICLR, 2023. 1, 2 +[16] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, Tom Duerig, and Vittorio Ferrari. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. IJCV, 2020. 1 +[17] Chunyuan Li, Haotian Liu, Liusan Harold Li, Pengchuan Zhang, Jyoti Aneja, Jianwei Yang, Ping Jin, Houdong Hu, Zicheng Liu, Yong Jae Lee, and Jianfeng Gao. ELEVATER: A Benchmark and Toolkit for Evaluating Language-Augmented Visual Models. In NeurIPS, 2022. 1, 5 +[18] Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Deepak Gotmare, Shafiq Joty, Caiming Xiong, and Steven Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In NeurIPS, 2021. 2 +[19] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In CVPR, pages 10965-10975, 2022. 1, 2, 3, 5, 6 +[20] Liunian Harold Li, Zi-Yi Dou, Nanyun Peng, and Kai-Wei Chang. DesCo: Learning Object Recognition with Rich Language Descriptions. In NeurIPS, 2023. 2 +[21] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In CVPR, 2023. 2, 3, 4 +[22] Chuang Lin, Peize Sun, Yi Jiang, Ping Luo, Lizhen Qu, Gholamreza Haffari, Zehuan Yuan, and Jianfei Cai. Learning object-language alignments for open-vocabulary object detection. In ICLR, 2023. 1 +[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C. Lawrence Zitnick. Microsoft COCO: Common Objects in Context. In ECCV, 2014. 1, 3 +[24] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal Loss for Dense Object Detection. In ICCV, 2017. 1, 2 +[25] Chang Liu, Henghui Ding, and Xudong Jiang. GRES: Generalized Referring Expression Segmentation. In CVPR, 2023. 1, 2 +[26] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 6 +[27] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach, 2019. 2 +[28] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. ViL-BERT: Pretraining Task-Agnostic Visiolinguistic Representations for Vision-and-Language Tasks. In NeurIPS, 2019. 2 + +[29] Zixian Ma, Jerry Hong, Mustafa Omer Gul, Mona Gandhi, Irena Gao, and Ranjay Krishna. CREPE: Can Vision-Language Foundation Models Reason Compositionally? In CVPR, 2023. 2 +[30] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L. Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In CVPR, 2016. 1, 2, 5 +[31] George A. Miller. Wordnet: A lexical database for english. Commun. ACM, 38(11):39-41, 1995. 2 +[32] Zhixiang Min, Bingbing Zhuang, Samuel Schulter, Buyu Liu, Enrique Dunn, and Manmohan Chandraker. Neurocs: Neural nocs supervision for monocular 3d object localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21404-21414, 2023. 1 +[33] Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, Xiao Wang, Xiaohua Zhai, Thomas Kipf, and Neil Houlsby. Simple Open-Vocabulary Object Detection with Vision Transformers. In ECCV, 2022. 1, 2, 6 +[34] Matthias Minderer, Alexey Gritsenko, and Neil Houlsby. Scaling Open-Vocabulary Object Detection. In NeurIPS, 2023. 1, 2 +[35] OpenAI. Gpt-3.5. 2022. 2, 3, 4 +[36] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k Entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In ICCV, pages 2641–2649, 2015. 1, 2, 3, 5 +[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 4 +[38] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks. In NeurIPS, 2015. 1 +[39] Joshua Robinson, Ching-Yao Chuang, Suvrit Sra, and Stefanie Jegelka. Contrastive Learning with Hard Negative Samples. In ICLR, 2021. 1, 2 +[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3, 4 +[41] Samuel Schulter, Vijay Kumar BG, Yumin Suh, Konstantinos M Dafnis, Zhixing Zhang, Shiyu Zhao, Dimitris Metaxas, et al. OmniLabel: A Challenging Benchmark for Language-Based Object Detection. In ICCV, 2023. 1, 2, 5, 6 +[42] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Jing Li, Xiangyu Zhang, and Jian Sun. Objects365: A Large-scale, High-quality Dataset for Object Detection. In ICCV, 2019. 1, 3, 5 +[43] Ravi Shekhar, Sandro Pezzelle, Yauhen Klimovich, Aurelie Herbelot, Moin Nabi, Enver Sangineto, and Raffaella + +Bernardi. FOIL it! Find One mismatch between Image and Language caption. In ACL, 2017. 1, 2, 3, 6 +[44] Abhinav Shrivastava, Abhinav Gupta, and Ross Girshick. Training Region-based Object Detectors with Online Hard Example Mining. In CVPR, 2016. 2 +[45] Sanjay Subramanian, William Merrill, Trevor Darrell, Matt Gardner, Sameer Singh, and Anna Rohrbach. ReCLIP: A Strong Zero-Shot Baseline for Referring Expression Comprehension. In ACL, 2022. 1 +[46] Tristan Thrush, Ryan Jiang, Max Bartolo, Amanpreet Singh, Adina Williams, Douwe Kiela, and Candace Ross. Winoground: Probing Vision and Language Models for Visio-Linguistic Compositionality. In CVPR, 2022. 1, 2, 4, 7 +[47] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models, 2023. 2, 3, 4 +[48] Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9 (86):2579-2605, 2008. 8 +[49] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. OFA: Unifying Architectures, Tasks, and Modalities Through a Simple Sequence-to-Sequence Learning Framework. In ICML, 2022. 6 +[50] Chenyun Wu, Zhe Lin, Scott Cohen, Trung Bui, and Subhransu Maji. PhraseCut: Language-based Image Segmentation in the Wild. In CVPR, 2020. 1, 2, 5 +[51] Hao Wu, Jiayuan Mao, Yufeng Zhang, Yuning Jiang, Lei Li, Weiwei Sun, and Wei-Ying Ma. Unified Visual-Semantic Embeddings: Bridging Vision and Language with Structured Meaning Representations. In CVPR, 2019. 3 +[52] Xiaoshi Wu, Feng Zhu, Rui Zhao, and Hongsheng Li. CORA: Adapting CLIP for Open-Vocabulary Detection with Region Prompting and Anchor Pre-Matching. In CVPR, 2023. 2 +[53] Chi Xie, Zhao Zhang, Yixuan Wu, Feng Zhu, Rui Zhao, and Shuang Liang. Described Object Detection: Liberating Object Detection with Flexible Expressions. In NeurIPS, 2023. 1, 2, 5, 6 +[54] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C. Berg, and Tamara L. Berg. Modeling Context in Referring Expressions. In ECCV, 2016. 1, 2, 5 +[55] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language models behave like bags-of-words, and what to do about it? In ICLR, 2023. 2 +[56] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-Vocabulary Object Detection Using Captions. In CVPR, 2021. 2 +[57] Shiyu Zhao, Zhixing Zhang, Samuel Schulter, Long Zhao, BG Vijay Kumar, Anastasis Stathopoulos, Manmohan Chandraker, and Dimitris N Metaxas. Exploiting unlabeled data with vision and language models for object detection. In ECCV, pages 159-175. Springer, 2022. 1, 2 + +[58] Shiyu Zhao, Samuel Schulter, Long Zhao, Zhixing Zhang, Vijay Kumar B. G, Yumin Suh, Manmohan Chandraker, and Dimitris N. Metaxas. Improving pseudo labels for open-vocabulary object detection, 2023. 2 +[59] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, and Jianfeng Gao. RegionCLIP: Region-based Language-Image Pretraining. In CVPR, 2022. 2 +[60] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbihl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In ECCV, pages 350-368. Springer, 2022. 5 \ No newline at end of file diff --git a/2024/Generating Enhanced Negatives for Training Language-Based Object Detectors/images.zip b/2024/Generating Enhanced Negatives for Training Language-Based Object Detectors/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c996aa21008acc1cd9f9877bc05aa24883442aee --- /dev/null +++ b/2024/Generating Enhanced Negatives for Training Language-Based Object Detectors/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e0cb184e520ea85f6e12a523a741c2bb1d69feb1e475b416c2aa231bd7c60b76 +size 537410 diff --git a/2024/Generating Enhanced Negatives for Training Language-Based Object Detectors/layout.json b/2024/Generating Enhanced Negatives for Training Language-Based Object Detectors/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..660aee0db5eea2401c53f3d3a49aaccd44f1d755 --- /dev/null +++ b/2024/Generating Enhanced Negatives for Training Language-Based Object Detectors/layout.json @@ -0,0 +1,9329 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 52, + 103, + 541, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 103, + 541, + 121 + ], + "spans": [ + { + "bbox": [ + 52, + 103, + 541, + 121 + ], + "type": "text", + "content": "Generating Enhanced Negatives for Training Language-Based Object Detectors" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 142, + 201, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 142, + 201, + 156 + ], + "spans": [ + { + "bbox": [ + 132, + 142, + 201, + 156 + ], + "type": "text", + "content": "Shiyu Zhao" + }, + { + "bbox": [ + 132, + 142, + 201, + 156 + ], + "type": "inline_equation", + "content": "^{1,*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 217, + 143, + 277, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 143, + 277, + 156 + ], + "spans": [ + { + "bbox": [ + 217, + 143, + 277, + 156 + ], + "type": "text", + "content": "Long Zhao3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 294, + 144, + 384, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 144, + 384, + 156 + ], + "spans": [ + { + "bbox": [ + 294, + 144, + 384, + 156 + ], + "type": "text", + "content": "Vijay Kumar B G²" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 402, + 144, + 460, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 144, + 460, + 156 + ], + "spans": [ + { + "bbox": [ + 402, + 144, + 460, + 156 + ], + "type": "text", + "content": "Yumin Suh2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 123, + 158, + 226, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 158, + 226, + 170 + ], + "spans": [ + { + "bbox": [ + 123, + 158, + 226, + 170 + ], + "type": "text", + "content": "Dimitris N. Metaxas1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 243, + 157, + 369, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 157, + 369, + 171 + ], + "spans": [ + { + "bbox": [ + 243, + 157, + 369, + 171 + ], + "type": "text", + "content": "Manmohan Chandraker2,4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 386, + 157, + 471, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 157, + 471, + 171 + ], + "spans": [ + { + "bbox": [ + 386, + 157, + 471, + 171 + ], + "type": "text", + "content": "Samuel Schulter2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 64, + 184, + 165, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 184, + 165, + 199 + ], + "spans": [ + { + "bbox": [ + 64, + 184, + 165, + 199 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 184, + 165, + 199 + ], + "type": "text", + "content": " Rutgers University" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 181, + 185, + 323, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 185, + 323, + 198 + ], + "spans": [ + { + "bbox": [ + 181, + 185, + 323, + 198 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 181, + 185, + 323, + 198 + ], + "type": "text", + "content": " NEC Laboratories America" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 338, + 185, + 430, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 185, + 430, + 198 + ], + "spans": [ + { + "bbox": [ + 338, + 185, + 430, + 198 + ], + "type": "text", + "content": "3 Google Research" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 446, + 185, + 525, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 446, + 185, + 525, + 198 + ], + "spans": [ + { + "bbox": [ + 446, + 185, + 525, + 198 + ], + "type": "text", + "content": "4 UC San Diego" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 143, + 226, + 191, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 226, + 191, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 226, + 191, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 252, + 290, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 290, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 290, + 491 + ], + "type": "text", + "content": "The recent progress in language-based open-vocabulary object detection can be largely attributed to finding better ways of leveraging large-scale data with free-form text annotations. Training such models with a discriminative objective function has proven successful, but requires good positive and negative samples. However, the free-form nature and the open vocabulary of object descriptions make the space of negatives extremely large. Prior works randomly sample negatives or use rule-based techniques to build them. In contrast, we propose to leverage the vast knowledge built into modern generative models to automatically build negatives that are more relevant to the original data. Specifically, we use large-language-models to generate negative text descriptions, and text-to-image diffusion models to also generate corresponding negative images. Our experimental analysis confirms the relevance of the generated negative data, and its use in language-based detectors improves performance on two complex benchmarks. Code is available at https://github.com/xiaofeng94/Gen-Enhanced-Negs." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 507, + 127, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 507, + 127, + 519 + ], + "spans": [ + { + "bbox": [ + 47, + 507, + 127, + 519 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 527, + 287, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 527, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 46, + 527, + 287, + 658 + ], + "type": "text", + "content": "Using natural language in object detection to describe semantics bears the potential to significantly increase the size of the detector's label space and enable novel applications. While standard detectors operate on a fixed label space [23, 38, 42], natural language allows for a broad spectrum of object descriptions, ranging from generic terms like \"vehicle\" to specific expressions like \"the red sports car parked on the left side\" [12, 17, 25, 30, 41, 53, 54]. Several works advanced language-based object detection over the past few years with novel training strategies [3, 5, 19, 22, 32, 34, 57] and model architectures [11, 15, 33, 45]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 47, + 660, + 287, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 660, + 287, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 660, + 287, + 684 + ], + "type": "text", + "content": "Referring expression or visual grounding datasets [14, 30, 36, 50, 54] provide the natural language object descrip" + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 307, + 220, + 383, + 275 + ], + "blocks": [ + { + "bbox": [ + 307, + 220, + 383, + 275 + ], + "lines": [ + { + "bbox": [ + 307, + 220, + 383, + 275 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 383, + 275 + ], + "type": "image", + "image_path": "f8eef1088602cc2133a650ab43cb86096fa93ebe24f8a1447afcce1c4b32bf07.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 384, + 222, + 544, + 365 + ], + "blocks": [ + { + "bbox": [ + 384, + 222, + 544, + 365 + ], + "lines": [ + { + "bbox": [ + 384, + 222, + 544, + 365 + ], + "spans": [ + { + "bbox": [ + 384, + 222, + 544, + 365 + ], + "type": "image", + "image_path": "5fbc2afc988111bfc60b08b82abf9de8d5b339cae4a76cde954b770724e371cb.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 373, + 546, + 438 + ], + "lines": [ + { + "bbox": [ + 304, + 373, + 546, + 438 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 546, + 438 + ], + "type": "text", + "content": "Figure 1. The key contribution of our work is to leverage large-language-models and text-to-image diffusion models to automatically generate negative object descriptions and images for training language-based object detectors. In contrast to prior work, our generated negatives are more relevant to the original data and provide a better training signal for detectors." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 461, + 546, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 546, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 546, + 521 + ], + "type": "text", + "content": "tions along with bounding box annotations needed for training. However, this data only describes what is present in the images, but misses to describe what is not. Yet, the notion of negatives is crucial for training discriminative models like language-based detectors [7, 24, 39]." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 522, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 547, + 714 + ], + "type": "text", + "content": "Detection datasets with a fixed label space provide negative classes implicitly or explicitly, with exhaustive [23, 42] or federated [6, 16] annotations, respectively. Any part of an image that does not overlap significantly with a bounding box of category " + }, + { + "bbox": [ + 304, + 522, + 547, + 714 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 522, + 547, + 714 + ], + "type": "text", + "content": " is verified to not be of that category (for exhaustive annotation). On the other hand, the space of negatives for a free-form text description of an object is extremely large. While some existing datasets provide negative samples in free-form text [43, 46], they were not annotated with bounding boxes. Hence, existing language-based detectors often define the negatives for one object as the descriptions of all other objects in the same image or descriptions of other random samples [3, 11, 19]. However, such negatives may not be directly related to the original positive description and define a weaker training signal (see Fig. 1). By explicitly evaluating on human-curated negatives, a re" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "text", + "content": "* Part of this work was done during an internship at NEC Laboratories America. Correspondence to: Shiyu Zhao sz553@rutgers.edu" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13592" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": "cent benchmark [41] identified a bias of existing language-based detectors to perform clearly better on positive rather than negative descriptions. However, creating a dataset with high-quality human-curated negatives for large-scale training is labor-intensive and costly." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 134, + 286, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 134, + 286, + 253 + ], + "spans": [ + { + "bbox": [ + 46, + 134, + 286, + 253 + ], + "type": "text", + "content": "In this work, we propose to explicitly and automatically generate negative data in the form of free-form texts as well as images. Prior works [4, 7, 29, 43, 46, 55] rely on rule-based approaches with knowledge graphs and focus only on the language domain or the classification task. In contrast, we leverage generative large-language-models (LLMs) [35, 47] and text-to-image diffusion models [21, 40] to automatically create relevant but contradicting object descriptions along with the corresponding images for language-based object detection, see Fig. 1." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 255, + 286, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 286, + 566 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 286, + 566 + ], + "type": "text", + "content": "Given an object description of a dataset, we first use LLMs to generate a semantically contradicting description as the negative. Besides changing individual words (foils) based on explicit knowledge graphs or LLMs, like in prior work [4, 7, 20], we demonstrate improved detection performance with two alternative approaches. (Re-combination): An LLM first identifies all objects in a sentence, and then creates a contradicting one by re-arranging, ignoring or adding objects. (In-context summaries): We prompt an LLM to summarize the differences of a few (less than 100) positive-negative pairs collected from an existing image-level dataset [46]. This summary is then used as context to generate more such examples. Note that we do not need visual input for this step, allowing us to leverage powerful LLMs for semantic and textual reasoning. Moreover, while prior work only focused on the text [4, 7, 29, 43, 55], we also leverage text-to-image diffusion models like GLIGEN [21] to create images that match the generated negative descriptions of objects, which serves as additional training signal. While the direct output of such image-generation models is often noisy and even wrong (not matching the input description), we propose two filtering steps to reduce noise considerably (from " + }, + { + "bbox": [ + 46, + 255, + 286, + 566 + ], + "type": "inline_equation", + "content": "53\\%" + }, + { + "bbox": [ + 46, + 255, + 286, + 566 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 255, + 286, + 566 + ], + "type": "inline_equation", + "content": "16\\%" + }, + { + "bbox": [ + 46, + 255, + 286, + 566 + ], + "type": "text", + "content": " according to an empirical study). Having both negative object descriptions and the corresponding image, allows us to improve the discriminative loss for training language-based object detectors." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 568, + 286, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 568, + 286, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 286, + 652 + ], + "type": "text", + "content": "Our experiments demonstrate clear accuracy gains on two challenging benchmarks, +2.9AP on OmniLabel [41] and +3.3AP on " + }, + { + "bbox": [ + 46, + 568, + 286, + 652 + ], + "type": "inline_equation", + "content": "\\mathrm{D}^3" + }, + { + "bbox": [ + 46, + 568, + 286, + 652 + ], + "type": "text", + "content": " [53], when adding our automatically-generated negative data into the training of baseline models like GLIP or FIBER. Moreover, we provide an in-depth analysis of the generated data (text and images) and how they contribute to better language-based detection." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "Summary of contributions: (1) Automatic generation of semantically relevant but contradicting negative text and images with large-scale generative models. (2) Recipes to integrate such negative data into language-based detection models like FIBER [3] and GLIP [19] (3) Clear improve" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "content": "ments on language-based detection benchmarks [41, 53] including a thorough analysis of the generated data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 109, + 390, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 109, + 390, + 121 + ], + "spans": [ + { + "bbox": [ + 306, + 109, + 390, + 121 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 129, + 545, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 129, + 545, + 332 + ], + "spans": [ + { + "bbox": [ + 304, + 129, + 545, + 332 + ], + "type": "text", + "content": "Vision & language localization tasks: Open-vocabulary detection (OVD) requires a model to localize object category names without having seen explicit bounding box annotation for them [5, 7, 15, 52, 56, 59]. In contrast, we focus on the more general language-based object detection task [41, 53], which goes beyond simple category names. Referring expression comprehension (REC) aims at localizing the subject of a free-form text expression. However, REC benchmarks [30, 50, 54] fall short in evaluating all aspects of the more general language-based detection task [41, 53]. In visual grounding (VG) [36], the task is to localize noun phrases of a caption in the image. Although being a task on its own, VG datasets have recently been used mostly as training data for OVD. Our work focuses on general language-based object detection, which subsumes and generalizes standard detection, OVD and REC [12, 25, 41, 53]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 337, + 545, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 337, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 304, + 337, + 545, + 481 + ], + "type": "text", + "content": "Language-based object detectors: Two critical abilities of language-based detection are accurate localization and tight text-image fusion. Works like [1, 9, 28] use language-models like BERT [2, 27] to align regions extracted from (pre-trained) detectors with captions. The outstanding zero-shot classification accuracy of large-scale pre-trained models like CLIP [37] or [10, 13, 18] then sparked interest in extensions for localization, with different approaches like distillation [5], fine-tuning [15, 33], pseudo-labeling [34, 57, 58], or combinations thereof [3, 19]. We use such models as test bed, but explore the underlying training data with respect to negative samples." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "content": "Negative samples for object detection: The notion of negatives is crucial for training discriminative models [24, 39]. Also for object detection, hard negative mining [44] has proven beneficial for model training. However, these prior works aim to find hard negative training examples rather than negatives in the label space, because the label space is fixed in standard detection. For language-based datasets, the space of potential negatives is extremely large because object descriptions are free-form text. Prior works [4, 29, 43, 46, 55] investigate negative texts for general vision & language models with different strategies, including changing individual words (foil) with rules based on knowledge graphs [31] or with LLMs. Sugar-Crepe [8] shares a similar idea as us to get negative texts with in-context learning but for image-text level pretraining. For language-based detection, [7] explores such rule-based foils, while [20] uses LLMs with specific templates to replace object names with alternative descriptions. In contrast, our work (1) focuses on the localization task, (2) ex" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13593" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 288, + 129 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 288, + 129 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 288, + 129 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 288, + 129 + ], + "type": "image", + "image_path": "676445f31ce25fdf5841a8cb2598a6422f82f742803a1a4a1da667cbc67d81fd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 137, + 287, + 182 + ], + "lines": [ + { + "bbox": [ + 46, + 137, + 287, + 182 + ], + "spans": [ + { + "bbox": [ + 46, + 137, + 287, + 182 + ], + "type": "text", + "content": "Figure 2. In language-based object detection, a detector receives as input an image and a (variable-length) list of free-form text descriptions of objects. For each description, the model predicts bounding boxes for objects that match the description." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 203, + 287, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 203, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 203, + 287, + 239 + ], + "type": "text", + "content": "plies more comprehensive strategies to generate negatives with LLMs, and (3) proposes to also generate corresponding negative images with text-to-image diffusion models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 251, + 102, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 251, + 102, + 262 + ], + "spans": [ + { + "bbox": [ + 47, + 251, + 102, + 262 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 270, + 223, + 283 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 270, + 223, + 283 + ], + "spans": [ + { + "bbox": [ + 47, + 270, + 223, + 283 + ], + "type": "text", + "content": "3.1. Language-based object detection" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 289, + 287, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 287, + 384 + ], + "type": "text", + "content": "Task definition: Given an image and a list of object descriptions, the task is to output bounding boxes along with confidence scores for each description, as shown in Fig. 2. Note the multi-label setting where one object instance can be referred to by multiple descriptions, like \"person\" and \"person looking at book\". Also note that an object description might refer to zero objects in the image and the desired output is an empty set of boxes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 388, + 288, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 288, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 288, + 664 + ], + "type": "text", + "content": "Training data: Many language-based detection models [3, 19] use a combination of object detection [23, 42] and visual grounding datasets [14, 36] for training their models. Both types of datasets provide images " + }, + { + "bbox": [ + 46, + 388, + 288, + 664 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 388, + 288, + 664 + ], + "type": "text", + "content": " and bounding boxes " + }, + { + "bbox": [ + 46, + 388, + 288, + 664 + ], + "type": "inline_equation", + "content": "b_{l}" + }, + { + "bbox": [ + 46, + 388, + 288, + 664 + ], + "type": "text", + "content": " to localize individual objects. Object detection data assigns each bounding box " + }, + { + "bbox": [ + 46, + 388, + 288, + 664 + ], + "type": "inline_equation", + "content": "b_{l}" + }, + { + "bbox": [ + 46, + 388, + 288, + 664 + ], + "type": "text", + "content": " a unique category " + }, + { + "bbox": [ + 46, + 388, + 288, + 664 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 388, + 288, + 664 + ], + "type": "text", + "content": " out of fixed label space " + }, + { + "bbox": [ + 46, + 388, + 288, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 46, + 388, + 288, + 664 + ], + "type": "text", + "content": ". The exhaustive labeling of the fixed label space in detection datasets implies the space of negatives. An object of category " + }, + { + "bbox": [ + 46, + 388, + 288, + 664 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 388, + 288, + 664 + ], + "type": "text", + "content": " is not any of the categories " + }, + { + "bbox": [ + 46, + 388, + 288, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{C} \\setminus c" + }, + { + "bbox": [ + 46, + 388, + 288, + 664 + ], + "type": "text", + "content": ". On the other hand, grounding data provides an image caption " + }, + { + "bbox": [ + 46, + 388, + 288, + 664 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 388, + 288, + 664 + ], + "type": "text", + "content": " in free-form text, where subsets of words " + }, + { + "bbox": [ + 46, + 388, + 288, + 664 + ], + "type": "inline_equation", + "content": "m_{l}" + }, + { + "bbox": [ + 46, + 388, + 288, + 664 + ], + "type": "text", + "content": " (defined as indices of starting and ending characters in " + }, + { + "bbox": [ + 46, + 388, + 288, + 664 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 388, + 288, + 664 + ], + "type": "text", + "content": ") are linked to bounding box " + }, + { + "bbox": [ + 46, + 388, + 288, + 664 + ], + "type": "inline_equation", + "content": "b_{l}" + }, + { + "bbox": [ + 46, + 388, + 288, + 664 + ], + "type": "text", + "content": ". For grounding data, the space of negatives is extremely large because one can find as many textual descriptions that do not match " + }, + { + "bbox": [ + 46, + 388, + 288, + 664 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 388, + 288, + 664 + ], + "type": "text", + "content": " as desired due to the compositionality of free-form text. Many language-based detectors [3, 19] only use the words in " + }, + { + "bbox": [ + 46, + 388, + 288, + 664 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 388, + 288, + 664 + ], + "type": "text", + "content": " that are not referred to by " + }, + { + "bbox": [ + 46, + 388, + 288, + 664 + ], + "type": "inline_equation", + "content": "m_{l}" + }, + { + "bbox": [ + 46, + 388, + 288, + 664 + ], + "type": "text", + "content": " as negatives for the bounding box " + }, + { + "bbox": [ + 46, + 388, + 288, + 664 + ], + "type": "inline_equation", + "content": "b_{l}" + }, + { + "bbox": [ + 46, + 388, + 288, + 664 + ], + "type": "text", + "content": ". We argue that this choice is sub-optimal because these words may refer to entirely different objects and are easy to discriminate. In the following section, we explain how we can automatically generate negative samples that semantically related to the original text " + }, + { + "bbox": [ + 46, + 388, + 288, + 664 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 388, + 288, + 664 + ], + "type": "text", + "content": " and hence provide a better training signal." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 671, + 204, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 204, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 204, + 684 + ], + "type": "text", + "content": "3.2. Generating negative samples" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "Our goal is to automatically and explicitly generate negative samples based on the original text descriptions " + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": " to im" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "prove the training signal for language-based detectors. A key observation of our work is to leverage the vast knowledge encoded into large-language models (LLMs) [35, 47] and text-to-image diffusion models (T2I) [21, 40]. Besides proposing novel ways to instruct LLMs for generating negative text descriptions (Sec. 3.2.1), we also propose to generate negative images (Sec. 3.2.2)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 175, + 529, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 175, + 529, + 186 + ], + "spans": [ + { + "bbox": [ + 305, + 175, + 529, + 186 + ], + "type": "text", + "content": "3.2.1 Generating negative descriptions with LLMs" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 196, + 545, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 196, + 545, + 291 + ], + "spans": [ + { + "bbox": [ + 304, + 196, + 545, + 291 + ], + "type": "text", + "content": "Given an object description " + }, + { + "bbox": [ + 304, + 196, + 545, + 291 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 196, + 545, + 291 + ], + "type": "text", + "content": " that matches the visual content inside a bounding box " + }, + { + "bbox": [ + 304, + 196, + 545, + 291 + ], + "type": "inline_equation", + "content": "b_{l}" + }, + { + "bbox": [ + 304, + 196, + 545, + 291 + ], + "type": "text", + "content": ", we define a \"negative\" description " + }, + { + "bbox": [ + 304, + 196, + 545, + 291 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 304, + 196, + 545, + 291 + ], + "type": "text", + "content": " as any text that is semantically different to the original text. Furthermore, our intuition is that good negative descriptions are still semantically related to the original description, but not the same. An example is: \"Person in red shirt\" as the original description and \"Person in blue shirt\" as a contradicting negative one." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 293, + 546, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 546, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 546, + 449 + ], + "type": "text", + "content": "Prior work [4, 7, 43] explored rule-based approaches to generate negative text. However, such rules are typically limited to simple knowledge graphs and are limited to replacing only individual words, often just nouns, or swapping words. In contrast, we explore more powerful LLMs to automatically generate relevant negatives. To make the negative text generation efficient and economic, in all cases, we first leverage a strong instruction-tuned LLM [35] to generate " + }, + { + "bbox": [ + 304, + 293, + 546, + 449 + ], + "type": "inline_equation", + "content": "50\\mathrm{k}" + }, + { + "bbox": [ + 304, + 293, + 546, + 449 + ], + "type": "text", + "content": " positive-negative pairs, and then finetune a LLaMA-7B [47] model with those pairs to then generate negative captions on large grounding datasets. In the following, we describe three ways to instruct an LLM for generating positive-negative pairs of object descriptions:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 455, + 545, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 455, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 304, + 455, + 545, + 611 + ], + "type": "text", + "content": "LLM-based foils: We first prompt an instruction-tuned LLM [35] to find concepts (i.e., objects, attributes and relationships) in object descriptions. Compared to rule-based parsers [51], LLMs can provide richer information. For example, for the caption \"A transportation vehicle is carrying a crowd of people who are sitting and standing,\" the parser ignores \"sitting\" and \"standing\", while LLMs regard them as attributes. Then, we pick one concept from the first step sequentially and prompt LLMs again to generate a negative caption by changing the concept. For both steps, the prompts are manually curated with the task definition and step-by-step instructions for the generation. Please find the exact prompt for the LLM in the supplement." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Re-combination: Next, we give the LLM more freedom in generating negative descriptions. We first prompt the LLM to identify all objects in the original caption, and then to recombine them to create a new sentence different from the original one. We allow the LLM to ignore, change or add new objects. For example, given the caption “A boy is playing with his dog” and two objects “boy” and “dog”, the LLM can output “The girl and her dog are playing fetch in" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13594" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 288, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 288, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 288, + 96 + ], + "type": "text", + "content": "the park\". Detailed prompts for both identifying objects and re-combination are in the supplement." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 103, + 289, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 103, + 289, + 307 + ], + "spans": [ + { + "bbox": [ + 46, + 103, + 289, + 307 + ], + "type": "text", + "content": "In-context summary: Third, we enable LLMs to learn how to generate negative descriptions by providing human-annotated positive-negative pairs as in-context samples. We randomly sample 80 pairs of positive and negative texts from the Winoground dataset [46] and prompt the instruction-tuned LLM [35] to summarize the difference of those pairs in plain text. Then, instead of manually creating prompts to generate positive-negative pairs, we leverage the summary together with three randomly sampled Winoground pairs as prompts to the LLM, and generate several positive-negative pairs to finetune LLaMA. After finetuning, the LLaMA model is used to generate negative texts for given descriptions. This pipeline does not require handcrafted prompts to LLMs as the explanation of the concept of negatives and how to create them. The supplement contains full prompts for generating a summary, and generating positive-negative pairs for finetuning." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 326, + 270, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 326, + 270, + 338 + ], + "spans": [ + { + "bbox": [ + 47, + 326, + 270, + 338 + ], + "type": "text", + "content": "3.2.2 Generating negative images with T2I models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 346, + 287, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 346, + 287, + 466 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 287, + 466 + ], + "type": "text", + "content": "Given an original image " + }, + { + "bbox": [ + 46, + 346, + 287, + 466 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 346, + 287, + 466 + ], + "type": "text", + "content": ", a bounding box " + }, + { + "bbox": [ + 46, + 346, + 287, + 466 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 46, + 346, + 287, + 466 + ], + "type": "text", + "content": " and a corresponding object description " + }, + { + "bbox": [ + 46, + 346, + 287, + 466 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 346, + 287, + 466 + ], + "type": "text", + "content": ", we define a negative image " + }, + { + "bbox": [ + 46, + 346, + 287, + 466 + ], + "type": "inline_equation", + "content": "I'" + }, + { + "bbox": [ + 46, + 346, + 287, + 466 + ], + "type": "text", + "content": " as any image that has a different semantic content inside " + }, + { + "bbox": [ + 46, + 346, + 287, + 466 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 46, + 346, + 287, + 466 + ], + "type": "text", + "content": ". The rest of the image can be equivalent to " + }, + { + "bbox": [ + 46, + 346, + 287, + 466 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 346, + 287, + 466 + ], + "type": "text", + "content": ". To obtain such imagery, we start with visual grounding data that provide bounding boxes, positive captions with text phrases, and alignment between them. We propose a two-step process: First, we turn the positive caption into a negative one. Second, we use conditioned image generation tools to alter the visual content inside the bounding box " + }, + { + "bbox": [ + 46, + 346, + 287, + 466 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 46, + 346, + 287, + 466 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 473, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 287, + 640 + ], + "type": "text", + "content": "Negative text for negative images: Although we have already described an approach to generate negative descriptions in Sec. 3.2.1, doing so to generate a negative image requires a different approach. In this case, the generated negative text needs to preserve the alignment " + }, + { + "bbox": [ + 46, + 473, + 287, + 640 + ], + "type": "inline_equation", + "content": "m_{l}" + }, + { + "bbox": [ + 46, + 473, + 287, + 640 + ], + "type": "text", + "content": " to the ground truth bounding box " + }, + { + "bbox": [ + 46, + 473, + 287, + 640 + ], + "type": "inline_equation", + "content": "b_{l}" + }, + { + "bbox": [ + 46, + 473, + 287, + 640 + ], + "type": "text", + "content": " in order to instruct the generative image model GLIGEN [21]. Hence, we first select a bounding box " + }, + { + "bbox": [ + 46, + 473, + 287, + 640 + ], + "type": "inline_equation", + "content": "b_{l}" + }, + { + "bbox": [ + 46, + 473, + 287, + 640 + ], + "type": "text", + "content": " and mask out the corresponding words (known via " + }, + { + "bbox": [ + 46, + 473, + 287, + 640 + ], + "type": "inline_equation", + "content": "m_{l}" + }, + { + "bbox": [ + 46, + 473, + 287, + 640 + ], + "type": "text", + "content": ") in the text " + }, + { + "bbox": [ + 46, + 473, + 287, + 640 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 473, + 287, + 640 + ], + "type": "text", + "content": ". For example, \"A boy is playing with his dog\" turns into \"A boy is playing with [Mask]\" if the selected bounding box refers to \"his dog\". Again, we leverage LLMs [35] to fill in text for \"[Mask]\" to generate a negative text without reusing the original text. Please refer to Fig. 3 for illustrations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": "We finetune a LLaMA-7B for the mask filling task with triplets of positive texts, masked texts, and negative texts. To reduce manual efforts, we follow the approach of in-context summary to get triplet samples. We apply this process twice: We start with only 5 manually created triplets to build a summary and generate 100 samples from the" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 306, + 72, + 545, + 199 + ], + "blocks": [ + { + "bbox": [ + 306, + 72, + 545, + 199 + ], + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 199 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 199 + ], + "type": "image", + "image_path": "03da1533848a52141bebfc703c084b2abdf51c14c1354ec0fbb83913137aeac8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 207, + 546, + 230 + ], + "lines": [ + { + "bbox": [ + 305, + 207, + 546, + 230 + ], + "spans": [ + { + "bbox": [ + 305, + 207, + 546, + 230 + ], + "type": "text", + "content": "Figure 3. Overview of using LLMs [35, 47] and text-to-image diffusion models [21, 40] to generate negative images." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 251, + 545, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 251, + 545, + 299 + ], + "spans": [ + { + "bbox": [ + 304, + 251, + 545, + 299 + ], + "type": "text", + "content": "LLM [35] with human checks. We then repeat the process to generate 50k examples without human checks from a summary of the 100 generated examples. This increases diversity in the generated data." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 301, + 546, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 301, + 546, + 398 + ], + "spans": [ + { + "bbox": [ + 304, + 301, + 546, + 398 + ], + "type": "text", + "content": "Conditional image generation: Given an image " + }, + { + "bbox": [ + 304, + 301, + 546, + 398 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 301, + 546, + 398 + ], + "type": "text", + "content": ", a bounding box " + }, + { + "bbox": [ + 304, + 301, + 546, + 398 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 304, + 301, + 546, + 398 + ], + "type": "text", + "content": " and the altered text " + }, + { + "bbox": [ + 304, + 301, + 546, + 398 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 304, + 301, + 546, + 398 + ], + "type": "text", + "content": ", we generate a negative image " + }, + { + "bbox": [ + 304, + 301, + 546, + 398 + ], + "type": "inline_equation", + "content": "I'" + }, + { + "bbox": [ + 304, + 301, + 546, + 398 + ], + "type": "text", + "content": " that is equal to " + }, + { + "bbox": [ + 304, + 301, + 546, + 398 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 301, + 546, + 398 + ], + "type": "text", + "content": " except inside " + }, + { + "bbox": [ + 304, + 301, + 546, + 398 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 304, + 301, + 546, + 398 + ], + "type": "text", + "content": ", where the visual content is altered to match the text " + }, + { + "bbox": [ + 304, + 301, + 546, + 398 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 304, + 301, + 546, + 398 + ], + "type": "text", + "content": ". To do so, we use the inpainting and conditioning abilities of GLIGEN [21], a T2I model [40]. Refer to [21] and our supplemental material for more details, and to Fig. 3 for an illustration of the process." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 400, + 547, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 400, + 547, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 400, + 547, + 651 + ], + "type": "text", + "content": "Mitigating noise in image generation: We found that the generated images are often noisy for any of the following reasons: (1) The altered text refers to a big bounding box that covers other smaller boxes. Large portions of the image are then generated and often do not match the concepts those smaller boxes originally covered. (2) The generated negative text does not match the bounding box that is either too small, too large or in an inappropriate position. (3) The T2I model fails to understand the negative text and generates wrong content. We propose two steps to filter such noisy images. First, we simply ignore ground truth boxes " + }, + { + "bbox": [ + 304, + 400, + 547, + 651 + ], + "type": "inline_equation", + "content": "b_{l}" + }, + { + "bbox": [ + 304, + 400, + 547, + 651 + ], + "type": "text", + "content": " for image generation if the box covers more than " + }, + { + "bbox": [ + 304, + 400, + 547, + 651 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 304, + 400, + 547, + 651 + ], + "type": "text", + "content": " of any other boxes in the image. Second, we adopt CLIP [37] to verify the semantic similarity of the generated image regions and the corresponding text. Specifically, we compute the similarity with CLIP between the generated image region (visual input) and the original and generated negative texts (text input). We filter out generated images that have a similarity score to the generated negative text lower than a user-defined threshold. Details on the filtering steps are given in the supplemental." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 658, + 479, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 479, + 672 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 479, + 672 + ], + "type": "text", + "content": "3.3. Learning from negative samples" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "text", + "content": "Detector design and training objective: The generated data does not prescribe any specific architecture for the detector. A common choice, which we also use for our exper" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13595" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 289, + 186 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 289, + 186 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 289, + 186 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 289, + 186 + ], + "type": "image", + "image_path": "81aba8dd6003f4b8376745bf454bc4710d45a6a7309deccd0ccb3d7788986279.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 193, + 287, + 270 + ], + "lines": [ + { + "bbox": [ + 46, + 193, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 46, + 193, + 287, + 270 + ], + "type": "text", + "content": "Figure 4. Illustration of the grounding loss used for training. Predictions that are matched with ground truth receive a positive signal from the associated text (tall rectangles). All other words receive a negative signal (short rectangles). The top left quarter shows the original loss used in [3, 19]. The other three quarters are related to our proposed generated negative data and provide additional training signals." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 293, + 289, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 289, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 289, + 460 + ], + "type": "text", + "content": "iments, is [3, 19]. The inputs are image " + }, + { + "bbox": [ + 46, + 293, + 289, + 460 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 293, + 289, + 460 + ], + "type": "text", + "content": " and text " + }, + { + "bbox": [ + 46, + 293, + 289, + 460 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 293, + 289, + 460 + ], + "type": "text", + "content": ", and the output is a set of bounding boxes " + }, + { + "bbox": [ + 46, + 293, + 289, + 460 + ], + "type": "inline_equation", + "content": "\\hat{b}_i" + }, + { + "bbox": [ + 46, + 293, + 289, + 460 + ], + "type": "text", + "content": " with corresponding logits " + }, + { + "bbox": [ + 46, + 293, + 289, + 460 + ], + "type": "inline_equation", + "content": "\\hat{p}_i \\in \\mathbb{R}^T" + }, + { + "bbox": [ + 46, + 293, + 289, + 460 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 46, + 293, + 289, + 460 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 293, + 289, + 460 + ], + "type": "text", + "content": " is the number of tokens required to represent text " + }, + { + "bbox": [ + 46, + 293, + 289, + 460 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 293, + 289, + 460 + ], + "type": "text", + "content": ". The ground truth can be represented by a binary assignment matrix " + }, + { + "bbox": [ + 46, + 293, + 289, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\in \\mathbb{B}^{L \\times T}" + }, + { + "bbox": [ + 46, + 293, + 289, + 460 + ], + "type": "text", + "content": ". Rows refer to ground truth boxes " + }, + { + "bbox": [ + 46, + 293, + 289, + 460 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 293, + 289, + 460 + ], + "type": "text", + "content": " and columns to tokens in " + }, + { + "bbox": [ + 46, + 293, + 289, + 460 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 293, + 289, + 460 + ], + "type": "text", + "content": ". Each element indicates if a token corresponds to a box " + }, + { + "bbox": [ + 46, + 293, + 289, + 460 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 293, + 289, + 460 + ], + "type": "text", + "content": ", which is given by the ground truth indices " + }, + { + "bbox": [ + 46, + 293, + 289, + 460 + ], + "type": "inline_equation", + "content": "m_l" + }, + { + "bbox": [ + 46, + 293, + 289, + 460 + ], + "type": "text", + "content": ". To define a loss, bipartite graph matching associates predictions with ground truth. For matched predictions, the target vector " + }, + { + "bbox": [ + 46, + 293, + 289, + 460 + ], + "type": "inline_equation", + "content": "g_i \\in \\mathbb{B}^T" + }, + { + "bbox": [ + 46, + 293, + 289, + 460 + ], + "type": "text", + "content": " is the corresponding row from " + }, + { + "bbox": [ + 46, + 293, + 289, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 46, + 293, + 289, + 460 + ], + "type": "text", + "content": ", while it is an all-zero vector for unmatched targets. The loss is then computed as " + }, + { + "bbox": [ + 46, + 293, + 289, + 460 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\sum_{i} \\ell_{\\mathrm{FL}}(\\hat{p}_i, g_i)" + }, + { + "bbox": [ + 46, + 293, + 289, + 460 + ], + "type": "text", + "content": " where FL refers to a focal loss. Fig. 4 illustrates the loss." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 467, + 287, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 467, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 467, + 287, + 552 + ], + "type": "text", + "content": "Integrating negative text: When sampling an image " + }, + { + "bbox": [ + 46, + 467, + 287, + 552 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 467, + 287, + 552 + ], + "type": "text", + "content": ", along with text " + }, + { + "bbox": [ + 46, + 467, + 287, + 552 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 467, + 287, + 552 + ], + "type": "text", + "content": ", boxes " + }, + { + "bbox": [ + 46, + 467, + 287, + 552 + ], + "type": "inline_equation", + "content": "b_{l}" + }, + { + "bbox": [ + 46, + 467, + 287, + 552 + ], + "type": "text", + "content": " and indices " + }, + { + "bbox": [ + 46, + 467, + 287, + 552 + ], + "type": "inline_equation", + "content": "m_{l}" + }, + { + "bbox": [ + 46, + 467, + 287, + 552 + ], + "type": "text", + "content": ", we also randomly sample " + }, + { + "bbox": [ + 46, + 467, + 287, + 552 + ], + "type": "inline_equation", + "content": "K > 1" + }, + { + "bbox": [ + 46, + 467, + 287, + 552 + ], + "type": "text", + "content": " negative descriptions from " + }, + { + "bbox": [ + 46, + 467, + 287, + 552 + ], + "type": "inline_equation", + "content": "\\{t_j^{\\prime}\\}" + }, + { + "bbox": [ + 46, + 467, + 287, + 552 + ], + "type": "text", + "content": " that defines the pool of negatives generated for text " + }, + { + "bbox": [ + 46, + 467, + 287, + 552 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 467, + 287, + 552 + ], + "type": "text", + "content": ". We randomly shuffle the order of all texts to avoid any biases on the location of the one positive description, and then concatenate them into one text string." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": "Integrating negative images: We explore two options: (1) Simply add the generated images " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "I'" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": " along with their generated (but semantically matching) captions " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": " as additional visual grounding data. The original caption " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": ", which was the starting point to generate the negative image " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "I'" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": ", is now used as the negative caption. In this way, both the original image " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": " and the generated one " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "I'" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": " have positive and negative descriptions. This option is illustrated in Fig. 4. (2) To better leverage the relation between the original and generated data, the second option is to pack them into a single training sample. We simply concatenate the images " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "I'" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": ", as well as the texts " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": ". The ground truth information " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "m_{l}" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": " is updated accordingly. See supplement for details." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 306, + 70, + 544, + 175 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 544, + 175 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 544, + 175 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 544, + 175 + ], + "type": "table", + "html": "
OmniLabelOmniLabel-Negative
APAPcAPdAPdPAPAPcAPdAPdP
Detic [60]8.015.65.48.0----
MDETR [11]--4.79.1----
GLIP-T [19]19.323.616.425.813.924.89.626.1
+ Ours22.227.218.829.016.528.611.630.2
FIBER-B [3]25.730.322.334.818.731.213.336.3
+ Ours28.132.125.136.522.333.316.738.3
", + "image_path": "d83d0267871e005ff781cc4b01aad084b9aac46e644fcb844c7d97e9acd7e45f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 324, + 182, + 526, + 193 + ], + "lines": [ + { + "bbox": [ + 324, + 182, + 526, + 193 + ], + "spans": [ + { + "bbox": [ + 324, + 182, + 526, + 193 + ], + "type": "text", + "content": "Table 1. Evaluation on the OmniLabel [41] benchmark." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 220, + 387, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 220, + 387, + 232 + ], + "spans": [ + { + "bbox": [ + 306, + 220, + 387, + 232 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 242, + 425, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 242, + 425, + 255 + ], + "spans": [ + { + "bbox": [ + 306, + 242, + 425, + 255 + ], + "type": "text", + "content": "4.1. Experimental design" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 264, + 545, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 264, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 304, + 264, + 545, + 419 + ], + "type": "text", + "content": "Training procedure: We choose two recent methods, GLIP-T [19] and FIBER-B [3], to demonstrate the effect of our automatically generated negatives. We use the official code and publicly available checkpoints as a starting point. The Flickr30k dataset [36] serves as our grounding dataset to generate the negative data. We then fine-tune GLIP-T and FIBER-B with both positive and negative data, along with the Objects365 detection dataset [42] for 1 epoch. Note that both Objects365 and Flickr30k are part of the original training set. We do not introduce any extra data except our generated negatives. Most hyper-parameters are equal to the original settings of GLIP and FIBER. Any exceptions are described in the supplement." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 438, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 546, + 715 + ], + "type": "text", + "content": "Evaluation benchmarks: We choose two recently proposed benchmarks, OmniLabel [41] and " + }, + { + "bbox": [ + 304, + 438, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{D}^3" + }, + { + "bbox": [ + 304, + 438, + 546, + 715 + ], + "type": "text", + "content": " [53], as our test beds. These benchmarks evaluate more aspects of language-based detection than existing referring expressions [30, 50, 54] or open-vocabulary detection [6, 17] benchmarks. Specifically, both benchmarks contain complex object descriptions that go beyond simple category names from open-vocabulary detection benchmarks. Moreover, the descriptions can refer to zero, one or multiple instances in the image, in contrast to standard referring expression benchmarks. These properties enable a more stringent evaluation metric as in object detection, which is based on average precision (AP) in both OmniLabel [41] and " + }, + { + "bbox": [ + 304, + 438, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{D}^3" + }, + { + "bbox": [ + 304, + 438, + 546, + 715 + ], + "type": "text", + "content": " [53]. Both benchmarks provide more fine-grained metrics. OmniLabel evaluates separately for categories, descriptions, and descriptions referring to at least one object, with APc, APd and APd-P, respectively. " + }, + { + "bbox": [ + 304, + 438, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{D}^3" + }, + { + "bbox": [ + 304, + 438, + 546, + 715 + ], + "type": "text", + "content": " differentiates descriptions on absence (\"Abs\") and presence (\"Pres\") that indicate whether or not they contain any form of negation (e.g., \"without\"), as well as on text lengths. Finally, we create a specific split for OmniLabel, \"OmniLabel-Negative\", to evaluate the model only on images that contain at least one negative description (i.e., not referring to any object)." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13596" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 70, + 286, + 194 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 286, + 194 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 286, + 194 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 286, + 194 + ], + "type": "table", + "html": "
D3(default)D3(by length of texts)
FullPresAbsSMLXL
OFA-L [49]4.24.14.64.95.43.02.1
OWL-ViT-L [33]9.610.76.420.79.46.05.3
G-DINO-B [26]20.720.122.522.622.518.916.5
OFA-DOD [53]21.623.715.423.622.620.518.4
GLIP-T [19]19.118.321.522.422.016.610.6
+ Ours21.420.623.728.124.517.411.5
FIBER-B [3]22.721.526.030.125.917.913.1
+ Ours26.025.228.135.529.720.514.2
", + "image_path": "520bd7f1b99379a7c6cef0974e7fc205dbc17ecb7225e11ba1b949f0adc66108.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 78, + 201, + 254, + 213 + ], + "lines": [ + { + "bbox": [ + 78, + 201, + 254, + 213 + ], + "spans": [ + { + "bbox": [ + 78, + 201, + 254, + 213 + ], + "type": "text", + "content": "Table 2. Evaluation on the " + }, + { + "bbox": [ + 78, + 201, + 254, + 213 + ], + "type": "inline_equation", + "content": "{\\mathrm{D}}^{3}" + }, + { + "bbox": [ + 78, + 201, + 254, + 213 + ], + "type": "text", + "content": " [53] benchmarks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 233, + 187, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 233, + 187, + 247 + ], + "spans": [ + { + "bbox": [ + 47, + 233, + 187, + 247 + ], + "type": "text", + "content": "4.2. Benchmark comparisons" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 252, + 289, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 289, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 289, + 396 + ], + "type": "text", + "content": "Tabs. 1 and 2 evaluate the impact of our generated negative training data on the OmniLabel [41] and " + }, + { + "bbox": [ + 46, + 252, + 289, + 396 + ], + "type": "inline_equation", + "content": "\\mathrm{D}^3" + }, + { + "bbox": [ + 46, + 252, + 289, + 396 + ], + "type": "text", + "content": " [53] benchmarks. In both tables, the first set of rows are baselines provided by the benchmarks. The following rows show the main comparisons for GLIP-T [19] and FIBER-B [3] with and without adding our generated negative training data. First, we can see that adding negative data improves results across all metrics for both models and both benchmarks. On OmniLabel, we can see a " + }, + { + "bbox": [ + 46, + 252, + 289, + 396 + ], + "type": "inline_equation", + "content": "+2.9\\%" + }, + { + "bbox": [ + 46, + 252, + 289, + 396 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 252, + 289, + 396 + ], + "type": "inline_equation", + "content": "+2.4\\%" + }, + { + "bbox": [ + 46, + 252, + 289, + 396 + ], + "type": "text", + "content": " increase in AP for GLIP-T and FIBER-B, respectively. Similarly, we observe a " + }, + { + "bbox": [ + 46, + 252, + 289, + 396 + ], + "type": "inline_equation", + "content": "+2.3\\%" + }, + { + "bbox": [ + 46, + 252, + 289, + 396 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 252, + 289, + 396 + ], + "type": "inline_equation", + "content": "+3.3\\%" + }, + { + "bbox": [ + 46, + 252, + 289, + 396 + ], + "type": "text", + "content": " increase in the main metric of " + }, + { + "bbox": [ + 46, + 252, + 289, + 396 + ], + "type": "inline_equation", + "content": "\\mathrm{D}^3" + }, + { + "bbox": [ + 46, + 252, + 289, + 396 + ], + "type": "text", + "content": " (AP on full descriptions) for GLIP-T and FIBER-B." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 404, + 191, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 404, + 191, + 417 + ], + "spans": [ + { + "bbox": [ + 47, + 404, + 191, + 417 + ], + "type": "text", + "content": "4.3. Analysis on negative texts" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 422, + 289, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 289, + 686 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 289, + 686 + ], + "type": "text", + "content": "Effectiveness of different negative texts: We finetune FIBER-B without and with different kinds of negative texts mentioned in Sect. 3.2.1, i.e., Rule-based foils, LLM-based foils, Re-combination with LLMs, In-context summary with LLMs, and present results in Table 3. We find all kinds of negatives improve the original FIBER-B on both OmniLabel and " + }, + { + "bbox": [ + 46, + 422, + 289, + 686 + ], + "type": "inline_equation", + "content": "\\mathrm{D}^3" + }, + { + "bbox": [ + 46, + 422, + 289, + 686 + ], + "type": "text", + "content": " benchmarks. Negative texts from LLMs generally achieve better results compared to LLM-based foils, which indicates that LLMs are powerful tools for negative text generation. Moreover, both recombination and in-context summary with LLMs outperform LLM-based foils in all metrics except APd-P. Note that APd-P refers to evaluations without negative label spaces, which is a task weaker than language-based detection. Based on such results, we argue that although word foils provide promising results in traditional studies [7, 43], it is sub-optimal to LLMs. We need to explore varied ways to unlock the ability of LLMs. We believe that our two solutions, i.e., Re-combination and In-context summary, provide a good starting point for future studies. Besides using only one kind of negative texts, we also explore the combinations of different kinds of negative texts in the supplement." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": "Diversity of rule-based and LLM-based negatives: In this part, we investigate the diversity of different negative" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 310, + 73, + 429, + 135 + ], + "blocks": [ + { + "bbox": [ + 310, + 73, + 429, + 135 + ], + "lines": [ + { + "bbox": [ + 310, + 73, + 429, + 135 + ], + "spans": [ + { + "bbox": [ + 310, + 73, + 429, + 135 + ], + "type": "image", + "image_path": "c0fb8d5c75d1f665d231f42d5e6db5374b87fda6c384f5ba7d1e78e28c3e962a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 431, + 74, + 541, + 135 + ], + "blocks": [ + { + "bbox": [ + 431, + 74, + 541, + 135 + ], + "lines": [ + { + "bbox": [ + 431, + 74, + 541, + 135 + ], + "spans": [ + { + "bbox": [ + 431, + 74, + 541, + 135 + ], + "type": "image", + "image_path": "f19c8c37a648ad25643a81bb76ce8a85202db15430132966d7e379191708f518.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 311, + 137, + 429, + 199 + ], + "blocks": [ + { + "bbox": [ + 311, + 137, + 429, + 199 + ], + "lines": [ + { + "bbox": [ + 311, + 137, + 429, + 199 + ], + "spans": [ + { + "bbox": [ + 311, + 137, + 429, + 199 + ], + "type": "image", + "image_path": "32a8a4eb4a89f6e7d559893c1f8676b7156b60b9ce9aab9af5aad8c81b042070.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 211, + 545, + 234 + ], + "lines": [ + { + "bbox": [ + 305, + 211, + 545, + 234 + ], + "spans": [ + { + "bbox": [ + 305, + 211, + 545, + 234 + ], + "type": "text", + "content": "Figure 5. Percentage of negative texts with the numbers of words. Four negative generation methods are compared." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 431, + 137, + 541, + 199 + ], + "blocks": [ + { + "bbox": [ + 431, + 137, + 541, + 199 + ], + "lines": [ + { + "bbox": [ + 431, + 137, + 541, + 199 + ], + "spans": [ + { + "bbox": [ + 431, + 137, + 541, + 199 + ], + "type": "image", + "image_path": "c98dec11641bcb54e462f9f9cd49fc721e1e2963baeff1ca563ac65680933579.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 310, + 251, + 428, + 312 + ], + "blocks": [ + { + "bbox": [ + 310, + 251, + 428, + 312 + ], + "lines": [ + { + "bbox": [ + 310, + 251, + 428, + 312 + ], + "spans": [ + { + "bbox": [ + 310, + 251, + 428, + 312 + ], + "type": "image", + "image_path": "22fdd28b65468b385aad3e33355b3b02e56efc81147ed3a405912ccb3340ede2.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 431, + 251, + 541, + 312 + ], + "blocks": [ + { + "bbox": [ + 431, + 251, + 541, + 312 + ], + "lines": [ + { + "bbox": [ + 431, + 251, + 541, + 312 + ], + "spans": [ + { + "bbox": [ + 431, + 251, + 541, + 312 + ], + "type": "image", + "image_path": "40bc52d2ee23dc901a4dd97c95be0116dec585d402ef605139164ea487e7217f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 311, + 316, + 428, + 376 + ], + "blocks": [ + { + "bbox": [ + 311, + 316, + 428, + 376 + ], + "lines": [ + { + "bbox": [ + 311, + 316, + 428, + 376 + ], + "spans": [ + { + "bbox": [ + 311, + 316, + 428, + 376 + ], + "type": "image", + "image_path": "aada24c5ca7f4a935dfd3425784770fda8e248bdfdb3d8180c83505f116591bf.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 389, + 545, + 422 + ], + "lines": [ + { + "bbox": [ + 305, + 389, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 305, + 389, + 545, + 422 + ], + "type": "text", + "content": "Figure 6. Percentage of negative texts with the numbers of words that are different from the original caption. Four negative generation methods are compared." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 431, + 316, + 541, + 376 + ], + "blocks": [ + { + "bbox": [ + 431, + 316, + 541, + 376 + ], + "lines": [ + { + "bbox": [ + 431, + 316, + 541, + 376 + ], + "spans": [ + { + "bbox": [ + 431, + 316, + 541, + 376 + ], + "type": "image", + "image_path": "a0f704dafcd969b164daf0ee64cb64dc4f62c227f26e5f8adbf6aa4dd4d2cfc0.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 340, + 441, + 509, + 520 + ], + "blocks": [ + { + "bbox": [ + 340, + 441, + 509, + 520 + ], + "lines": [ + { + "bbox": [ + 340, + 441, + 509, + 520 + ], + "spans": [ + { + "bbox": [ + 340, + 441, + 509, + 520 + ], + "type": "image", + "image_path": "9aa5fb75fda56cd538964bc03f29b59fb6de06347c09a149c787da01948a6bdf.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 522, + 545, + 556 + ], + "lines": [ + { + "bbox": [ + 305, + 522, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 305, + 522, + 545, + 556 + ], + "type": "text", + "content": "Figure 7. Average numbers of extra unique words per thousand generated negative texts, which are not included in the original dataset. We group words by their part-of-speech." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 580, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 580, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 580, + 545, + 652 + ], + "type": "text", + "content": "texts. First, we count number of words for each negative text and provide the distribution for negatives of different sources in Fig. 5. As shown, all four distributions have a peak around 10 words, but the one of rule-based foils is higher than others. That means rule-based foils provide more negative texts with similar lengths." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": "Second, we count the number of different words between the original positive caption and the negative caption, and present the distributions in Fig. 6. We find that LLM-based methods usually changes more words than rule-based foils, which increases the diversity. Moreover, in-context sum" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13597" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 71, + 347, + 159 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 347, + 159 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 347, + 159 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 347, + 159 + ], + "type": "table", + "html": "
Whole OmniLabelOmniLabel-NegativeD3
APAPcAPdAPd-PAPAPcAPdAPd-PFullPresAbs
Original FIBER-B25.730.322.334.818.731.213.336.322.721.526.0
+ Rule-based foils26.431.722.634.919.232.613.636.424.123.226.9
+ LLM-based foils26.530.723.335.920.832.115.438.024.624.026.5
+ Re-combination26.930.823.935.921.132.315.637.625.324.627.3
+ In-context summary26.630.823.434.221.132.215.736.425.725.227.5
", + "image_path": "fec34ce9ad9ad02f1cdaf9129f6ca4984aaaa254993f9701b3078956c1451a4d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 359, + 72, + 544, + 182 + ], + "blocks": [ + { + "bbox": [ + 47, + 167, + 348, + 190 + ], + "lines": [ + { + "bbox": [ + 47, + 167, + 348, + 190 + ], + "spans": [ + { + "bbox": [ + 47, + 167, + 348, + 190 + ], + "type": "text", + "content": "Table 3. Performance of FIBER-B trained with negative texts from four negative generation methods." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 359, + 72, + 544, + 182 + ], + "lines": [ + { + "bbox": [ + 359, + 72, + 544, + 182 + ], + "spans": [ + { + "bbox": [ + 359, + 72, + 544, + 182 + ], + "type": "table", + "html": "
OmniLabelD3
APcAPdAPd-P
Original FIBER30.322.334.822.7
FIBER w/ neg. texts30.723.935.525.9
+ W/ neg. img. directly30.122.433.723.0
+ Box filter31.023.835.423.6
+ Box&CLIP filters31.124.235.924.1
+ Above + concat. img.31.724.835.924.8
+ Above + weight ensemble32.125.236.526.0
", + "image_path": "557e9294ef85422db01b54e4baed5c237eb0a08245ea5b40555191b296d139e0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 367, + 190, + 534, + 201 + ], + "lines": [ + { + "bbox": [ + 367, + 190, + 534, + 201 + ], + "spans": [ + { + "bbox": [ + 367, + 190, + 534, + 201 + ], + "type": "text", + "content": "Table 4. FIBER trained with negative images." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 78, + 220, + 258, + 348 + ], + "blocks": [ + { + "bbox": [ + 78, + 220, + 258, + 348 + ], + "lines": [ + { + "bbox": [ + 78, + 220, + 258, + 348 + ], + "spans": [ + { + "bbox": [ + 78, + 220, + 258, + 348 + ], + "type": "image", + "image_path": "245a9dd89cbd49cca57c28c1b3f7df453277347f71c5dc4dd4831db6e614e365.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 356, + 287, + 378 + ], + "lines": [ + { + "bbox": [ + 46, + 356, + 287, + 378 + ], + "spans": [ + { + "bbox": [ + 46, + 356, + 287, + 378 + ], + "type": "text", + "content": "Figure 8. Noisy generated images. The orange box contains the red box, and editing the orange changes the red unexpectedly." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 399, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 287, + 495 + ], + "type": "text", + "content": "mary has a more flat distribution compared to others. Probably, in-context summary learns how to generate negatives automatically from data and has less restrictions. Besides, in-context summary has more cases with no word changed where negative texts are generated by just shuffling words or concepts in the original text. Such shuffling is a common pattern of Winoground [46], and our in-context summary can learn such data specific patterns." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 496, + 287, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 496, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 496, + 287, + 662 + ], + "type": "text", + "content": "Third, we count how many extra words that does not exist in the original Flickr30k dataset are introduced in different negative generation methods. Fig. 7 shows the average number of extra words per 1000 negative texts. We group words into four part-of-speech categories, i.e., VERB, NOUN, ADP/ADJ, and others. As shown, LLMs introduce more extra words on average than rule-based foils probably because rule-based foils are limited in a predefined set of words. However, LLMs are open to any concepts and have great potentials of generating diverse texts. In-context summary introduces the most extra words for all categories, which is likely a benefit of learning negative generation from data. The above statistics indicate a clear view that LLMs generate more diverse data than rule-based foils." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 671, + 201, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 201, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 201, + 685 + ], + "type": "text", + "content": "4.4. Analysis on negative images" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "Noise in generated images: As mentioned in the last paragraph of Sect. 3.2.2, the raw generated images are noisy in" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 222, + 546, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 222, + 546, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 222, + 546, + 437 + ], + "type": "text", + "content": "several ways. First, the editing of a large box will override the context of smaller boxes that are covered by the large box. As shown in Fig. 8, GLIGEN did follow the instruction to generate a boy in the orange box, but the black jacket in the red box is missing. As a remedy, we apply our first de-noise step \"Box Filter\". That is, we ignore boxes that contain any other boxes when generating negative images. Second, GLIGEN may generate contents with wrong attributes or objects, as shown in Fig. 9 (Left). Moreover, our generation pipeline includes some cases where the edited text and the bounding box does not match. As shown in Fig. 9 (Right), the box for \"his lap\" cannot be modified as \"his knees\". Thus, GLIGEN generates wrong contents. As described in Sect. 3.2.2, we adopt a pretrained CLIP model to judge if generated contents are correct, which mitigates the noise to some extent. As shown in Fig. 9, both negative images get low CLIP scores and can be filtered out with a threshold. We call such thresholding \"CLIP Filter\"." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 444, + 545, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 444, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 304, + 444, + 545, + 563 + ], + "type": "text", + "content": "Subject studies on Box and CLIP filters: We employ human experts to check the amount of noisy generated images. First, for negative images w/o filter, w/ Box filter, and w/ Box&CLIP filters, we separately and randomly select 100 samples. Then, we ask two experts to check if a negative image is not noisy by comparing it with its caption and the original positive image. We regard an image as not noisy when both experts agree. As shown in Fig. 10, both filters reduce the noise. The Box filter improves from " + }, + { + "bbox": [ + 304, + 444, + 545, + 563 + ], + "type": "inline_equation", + "content": "47\\%" + }, + { + "bbox": [ + 304, + 444, + 545, + 563 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 444, + 545, + 563 + ], + "type": "inline_equation", + "content": "63\\%" + }, + { + "bbox": [ + 304, + 444, + 545, + 563 + ], + "type": "text", + "content": ", and the CLIP filter improves to " + }, + { + "bbox": [ + 304, + 444, + 545, + 563 + ], + "type": "inline_equation", + "content": "84\\%" + }, + { + "bbox": [ + 304, + 444, + 545, + 563 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": "Effectiveness of generated negative images: To show the effectiveness of generated images themselves, we take captions of generated images as additional negative texts to in-context summary, and finetune a FIBER model as baseline. Then, we compare the baseline with variants of adding generated negative images in Table 4. As shown, the performance drops if we directly take raw negative images as new visual grounding data without any filters (i.e., W/ neg. img. directly). Probably, there are too much noise in raw negative images as shown in Fig. 10. When applying both Box and CLIP filters on negative images, we can achieve slight improvement on OmniLabel compared to using neg-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13598" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 73, + 91, + 182, + 174 + ], + "blocks": [ + { + "bbox": [ + 113, + 81, + 142, + 90 + ], + "lines": [ + { + "bbox": [ + 113, + 81, + 142, + 90 + ], + "spans": [ + { + "bbox": [ + 113, + 81, + 142, + 90 + ], + "type": "text", + "content": "Original" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 73, + 91, + 182, + 174 + ], + "lines": [ + { + "bbox": [ + 73, + 91, + 182, + 174 + ], + "spans": [ + { + "bbox": [ + 73, + 91, + 182, + 174 + ], + "type": "image", + "image_path": "23d9566b268ef8858588220de2689f124b5bc29f26ad20625525994e21a72dc5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 176, + 285, + 186 + ], + "lines": [ + { + "bbox": [ + 75, + 176, + 285, + 186 + ], + "spans": [ + { + "bbox": [ + 75, + 176, + 285, + 186 + ], + "type": "text", + "content": "Positive: A woman in a black mesh skirt plays acoustic guitar" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 76, + 189, + 286, + 199 + ], + "lines": [ + { + "bbox": [ + 76, + 189, + 286, + 199 + ], + "spans": [ + { + "bbox": [ + 76, + 189, + 286, + 199 + ], + "type": "text", + "content": "Negative: A woman in a white lace dress plays acoustic guitar" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 184, + 91, + 293, + 174 + ], + "blocks": [ + { + "bbox": [ + 126, + 70, + 230, + 79 + ], + "lines": [ + { + "bbox": [ + 126, + 70, + 230, + 79 + ], + "spans": [ + { + "bbox": [ + 126, + 70, + 230, + 79 + ], + "type": "text", + "content": "Generating wrong attributes" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 220, + 81, + 256, + 90 + ], + "lines": [ + { + "bbox": [ + 220, + 81, + 256, + 90 + ], + "spans": [ + { + "bbox": [ + 220, + 81, + 256, + 90 + ], + "type": "text", + "content": "Generated" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 184, + 91, + 293, + 174 + ], + "lines": [ + { + "bbox": [ + 184, + 91, + 293, + 174 + ], + "spans": [ + { + "bbox": [ + 184, + 91, + 293, + 174 + ], + "type": "image", + "image_path": "853c0d3bb4871d4e8bb3999eac24b1ba30064c504d1b38b65e7ea4b55255f775.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 290, + 176, + 507, + 186 + ], + "lines": [ + { + "bbox": [ + 290, + 176, + 507, + 186 + ], + "spans": [ + { + "bbox": [ + 290, + 176, + 507, + 186 + ], + "type": "text", + "content": "Positive: A man is sitting on a chair with his hands above his lap" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 290, + 189, + 520, + 198 + ], + "lines": [ + { + "bbox": [ + 290, + 189, + 520, + 198 + ], + "spans": [ + { + "bbox": [ + 290, + 189, + 520, + 198 + ], + "type": "text", + "content": "Negative: A man is sitting on a chair with his hands above his knees" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 301, + 91, + 410, + 174 + ], + "blocks": [ + { + "bbox": [ + 331, + 70, + 493, + 79 + ], + "lines": [ + { + "bbox": [ + 331, + 70, + 493, + 79 + ], + "spans": [ + { + "bbox": [ + 331, + 70, + 493, + 79 + ], + "type": "text", + "content": "Generation with inappropriate boxes or texts" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 340, + 81, + 370, + 91 + ], + "lines": [ + { + "bbox": [ + 340, + 81, + 370, + 91 + ], + "spans": [ + { + "bbox": [ + 340, + 81, + 370, + 91 + ], + "type": "text", + "content": "Original" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 301, + 91, + 410, + 174 + ], + "lines": [ + { + "bbox": [ + 301, + 91, + 410, + 174 + ], + "spans": [ + { + "bbox": [ + 301, + 91, + 410, + 174 + ], + "type": "image", + "image_path": "d6c9a61e13381799fe4271559890e473fe0136d9313c276f643b6b34624b3e78.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 411, + 91, + 521, + 174 + ], + "blocks": [ + { + "bbox": [ + 447, + 80, + 483, + 89 + ], + "lines": [ + { + "bbox": [ + 447, + 80, + 483, + 89 + ], + "spans": [ + { + "bbox": [ + 447, + 80, + 483, + 89 + ], + "type": "text", + "content": "Generated" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 411, + 91, + 521, + 174 + ], + "lines": [ + { + "bbox": [ + 411, + 91, + 521, + 174 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 521, + 174 + ], + "type": "image", + "image_path": "d4ced8cb2308daaebd82cdfa84be84dd766c36a1cbfccdf827327ecddd3d0564.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 80, + 264, + 251, + 319 + ], + "blocks": [ + { + "bbox": [ + 46, + 208, + 547, + 242 + ], + "lines": [ + { + "bbox": [ + 46, + 208, + 547, + 242 + ], + "spans": [ + { + "bbox": [ + 46, + 208, + 547, + 242 + ], + "type": "text", + "content": "Figure 9. Left: Noisy negative images due to wrong attributes or objects generated by text-to-image models. Right: Noisy negative images caused by inappropriate bounding boxes or negative texts from LLMs. CLIP scores of generated images refer to the similarity between the box and the negative text compared to the positive text. Thresholding on CLIP scores remove those noisy images." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 80, + 264, + 251, + 319 + ], + "lines": [ + { + "bbox": [ + 80, + 264, + 251, + 319 + ], + "spans": [ + { + "bbox": [ + 80, + 264, + 251, + 319 + ], + "type": "image", + "image_path": "173612794985e0903a31e2b35b3de7826df42b76b2af7e84e7b62e54a5b38c97.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 324, + 272, + 335 + ], + "lines": [ + { + "bbox": [ + 61, + 324, + 272, + 335 + ], + "spans": [ + { + "bbox": [ + 61, + 324, + 272, + 335 + ], + "type": "text", + "content": "Figure 10. Percentage of good generated negative images." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 58, + 348, + 287, + 460 + ], + "blocks": [ + { + "bbox": [ + 58, + 348, + 287, + 460 + ], + "lines": [ + { + "bbox": [ + 58, + 348, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 58, + 348, + 287, + 460 + ], + "type": "image", + "image_path": "7676b5cf758c51a2742d1dc7942bb7ce5cc5a26580ce107b1fd93fd31331f902.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 480, + 287, + 503 + ], + "lines": [ + { + "bbox": [ + 46, + 480, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 46, + 480, + 287, + 503 + ], + "type": "text", + "content": "Figure 11. Distributions of image regions from Omnilabel, " + }, + { + "bbox": [ + 46, + 480, + 287, + 503 + ], + "type": "inline_equation", + "content": "\\mathbf{D}^3" + }, + { + "bbox": [ + 46, + 480, + 287, + 503 + ], + "type": "text", + "content": ", and our generated images. Visualization with t-SNE." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 47, + 526, + 112, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 526, + 112, + 538 + ], + "spans": [ + { + "bbox": [ + 47, + 526, + 112, + 538 + ], + "type": "text", + "content": "ative texts only." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": "Concatenating images during training: Following the idea of concatenating the positive and negative captions as the text input, we concatenate the positive and negative images as one input image during training. See supplement for an example. In this way, models are forced to tell the difference between the positive and negative images within one training iteration, which helps detectors to learn better about the negative. As shown in Table 4, such a simple technique improves upon “+ Box&CLIP filters” both on OmniLabel and " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{D}^{3}" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": ". Furthermore, we ensemble the weights of two FIBER models, one finetuned with negative texts only, and the other finetuned with both negative texts and images. Finally, compared to using negative texts only, we gain 1.3 APd on OmniLabel and no performance drop on " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{D}^{3}" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 262, + 545, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 262, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 304, + 262, + 545, + 466 + ], + "type": "text", + "content": "Looking into generated images and benchmarks: Table 4 shows that negative images help on OmniLabel but not much on " + }, + { + "bbox": [ + 304, + 262, + 545, + 466 + ], + "type": "inline_equation", + "content": "\\mathrm{D}^3" + }, + { + "bbox": [ + 304, + 262, + 545, + 466 + ], + "type": "text", + "content": ". We explore this on a data basis. We first crop image regions for generated images, OmniLabel images, " + }, + { + "bbox": [ + 304, + 262, + 545, + 466 + ], + "type": "inline_equation", + "content": "\\mathrm{D}^3" + }, + { + "bbox": [ + 304, + 262, + 545, + 466 + ], + "type": "text", + "content": " images, and Flickr30k images based on the bounding boxes. Then, we randomly select 1000 image regions and feed them into a CLIP image encoder to get CLIP embeddings. Later, we input those embeddings to t-SNE [48] to illustrate the similarities between different image regions. As shown in Fig. 11, " + }, + { + "bbox": [ + 304, + 262, + 545, + 466 + ], + "type": "inline_equation", + "content": "\\mathrm{D}^3" + }, + { + "bbox": [ + 304, + 262, + 545, + 466 + ], + "type": "text", + "content": "'s regions are grouped into several clusters, while OmniLabel and our generated regions are scattered in the center. This indicates that there is a clear domain gap between " + }, + { + "bbox": [ + 304, + 262, + 545, + 466 + ], + "type": "inline_equation", + "content": "\\mathrm{D}^3" + }, + { + "bbox": [ + 304, + 262, + 545, + 466 + ], + "type": "text", + "content": " and the others. Thus, it is plausible that our generated images only helps on OmniLabel. In our view, the gap comes from that " + }, + { + "bbox": [ + 304, + 262, + 545, + 466 + ], + "type": "inline_equation", + "content": "\\mathrm{D}^3" + }, + { + "bbox": [ + 304, + 262, + 545, + 466 + ], + "type": "text", + "content": " collect data in groups based on categories. In contrast, OmniLabel collects data randomly." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 477, + 378, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 477, + 378, + 489 + ], + "spans": [ + { + "bbox": [ + 306, + 477, + 378, + 489 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 498, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 665 + ], + "type": "text", + "content": "Language-based detection requires localization of objects by a referring free-form text descriptions. To train accurate models in a discriminative way, the training data must contain good negative samples. Starting with an existing dataset, we propose (1) novel ways to prompt LLMs for generating additional negative texts, and (2) generating negative images to complement the training signal. Based on our experimental evaluations, we conclude that such additional negative training data indeed translates into improved detection accuracy on standard benchmarks. Our analysis demonstrates the importance of diversity in the generated text, which is higher with our approach than with prior works, and the quality of the generated images, which our proposed filtering steps can significantly increase." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 666, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgments: This research project has been partially funded by research grants to Dimitris N. Metaxas through NSF: 2310966, 2235405, 2212301, 2003874, and FA9550-23-1-0417." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13599" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "text", + "content": "[1] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. UNITER: UNiversal Image-TExt Representation Learning. In ECCV, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 287, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 287, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 287, + 168 + ], + "type": "text", + "content": "[2] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 170, + 287, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 170, + 287, + 223 + ], + "spans": [ + { + "bbox": [ + 53, + 170, + 287, + 223 + ], + "type": "text", + "content": "[3] Zi-Yi Dou, Aishwarya Kamath, Zhe Gan, Pengchuan Zhang, Jianfeng Wang, Linjie Li, Zicheng Liu, Ce Liu, Yann LeCun, Nanyun Peng, et al. Coarse-to-Fine Vision-Language Pretraining with Fusion in the Backbone. In NeurIPS, 2022. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 224, + 287, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 287, + 278 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 287, + 278 + ], + "type": "text", + "content": "[4] Sivan Doveh, Assaf Arbelle, Sivan Harary, Eli Schwartz, Roei Herzig, Raja Giryes, Rogerio Feris, Rameswar Panda, Shimon Ullman, and Leonid Karlinsky. Teaching Structured Vision & Language Concepts to Vision & Language Models. In CVPR, 2023. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 280, + 287, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 280, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 280, + 287, + 312 + ], + "type": "text", + "content": "[5] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary Object Detection via Vision and Language Knowledge Distillation. In ICLR, 2022. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 314, + 287, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 314, + 287, + 345 + ], + "spans": [ + { + "bbox": [ + 53, + 314, + 287, + 345 + ], + "type": "text", + "content": "[6] Agrim Gupta, Piotr Dólár, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In CVPR, 2019. 1, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 347, + 287, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 347, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 53, + 347, + 287, + 389 + ], + "type": "text", + "content": "[7] Ryota Hinami and Shin'ichi Satoh. Discriminative Learning of Open-Vocabulary Object Retrieval and Localization by Negative Phrase Augmentation. In EMNLP, 2018. 1, 2, 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 392, + 287, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 392, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 392, + 287, + 434 + ], + "type": "text", + "content": "[8] Cheng-Yu Hsieh, Jieyu Zhang, Zixian Ma, Aniruddha Kembhavi, and Ranjay Krishna. Sugarcrepe: Fixing hackable benchmarks for vision-language compositionality. In NeurIPS, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 436, + 287, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 436, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 53, + 436, + 287, + 467 + ], + "type": "text", + "content": "[9] Ronghang Hu and Amanpreet Singh. UniT: Multimodal Multitask Learning with a Unified Transformer. In ICCV, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 470, + 287, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 470, + 287, + 522 + ], + "spans": [ + { + "bbox": [ + 48, + 470, + 287, + 522 + ], + "type": "text", + "content": "[10] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, and Tom Duerig. Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision. In ICML, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 525, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 568 + ], + "type": "text", + "content": "[11] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. MDETR - Modulated Detection for End-to-End Multi-Modal Understanding. In ICCV, 2021. 1, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 570, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 613 + ], + "type": "text", + "content": "[12] Aishwarya Kamath, Sara Price, Jonas Pfeiffer, Yann LeCun, and Nicolas Carion. TRICD: Testing Robust Image Understanding Through Contextual Phrase Detection. Technical report, NYU, 2023. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 614, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 656 + ], + "type": "text", + "content": "[13] Zaid Khan, Vijay Kumar B.G., Xiang Yu, Samuel Schulter, Manmohan Chandraker, and Yun Fu. Single-Stream Multi-Level Alignment for Vision-Language Pretraining. In ECCV, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[14] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV, 123:32-73, 2017. 1, 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "type": "text", + "content": "[15] Weicheng Kuo, Yin Cui, Xiuye Gu, AJ Piergiovanni, and Anelia Angelova. F-vlm: Open-vocabulary object detection upon frozen vision and language models. In ICLR, 2023. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 545, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 182 + ], + "type": "text", + "content": "[16] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, Tom Duerig, and Vittorio Ferrari. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. IJCV, 2020. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 184, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 184, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 307, + 184, + 545, + 237 + ], + "type": "text", + "content": "[17] Chunyuan Li, Haotian Liu, Liusan Harold Li, Pengchuan Zhang, Jyoti Aneja, Jianwei Yang, Ping Jin, Houdong Hu, Zicheng Liu, Yong Jae Lee, and Jianfeng Gao. ELEVATER: A Benchmark and Toolkit for Evaluating Language-Augmented Visual Models. In NeurIPS, 2022. 1, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 239, + 545, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 239, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 307, + 239, + 545, + 281 + ], + "type": "text", + "content": "[18] Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Deepak Gotmare, Shafiq Joty, Caiming Xiong, and Steven Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 283, + 545, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 283, + 545, + 336 + ], + "spans": [ + { + "bbox": [ + 307, + 283, + 545, + 336 + ], + "type": "text", + "content": "[19] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In CVPR, pages 10965-10975, 2022. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 338, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 338, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 307, + 338, + 545, + 370 + ], + "type": "text", + "content": "[20] Liunian Harold Li, Zi-Yi Dou, Nanyun Peng, and Kai-Wei Chang. DesCo: Learning Object Recognition with Rich Language Descriptions. In NeurIPS, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 372, + 545, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 372, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 307, + 372, + 545, + 414 + ], + "type": "text", + "content": "[21] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In CVPR, 2023. 2, 3, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 415, + 545, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 415, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 307, + 415, + 545, + 458 + ], + "type": "text", + "content": "[22] Chuang Lin, Peize Sun, Yi Jiang, Ping Luo, Lizhen Qu, Gholamreza Haffari, Zehuan Yuan, and Jianfei Cai. Learning object-language alignments for open-vocabulary object detection. In ICLR, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 460, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 460, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 307, + 460, + 545, + 502 + ], + "type": "text", + "content": "[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C. Lawrence Zitnick. Microsoft COCO: Common Objects in Context. In ECCV, 2014. 1, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 504, + 545, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 504, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 307, + 504, + 545, + 535 + ], + "type": "text", + "content": "[24] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal Loss for Dense Object Detection. In ICCV, 2017. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 537, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 537, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 307, + 537, + 545, + 568 + ], + "type": "text", + "content": "[25] Chang Liu, Henghui Ding, and Xudong Jiang. GRES: Generalized Referring Expression Segmentation. In CVPR, 2023. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 570, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 570, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 570, + 545, + 624 + ], + "type": "text", + "content": "[26] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 625, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 625, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 545, + 669 + ], + "type": "text", + "content": "[27] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach, 2019. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[28] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. ViL-BERT: Pretraining Task-Agnostic Visiolinguistic Representations for Vision-and-Language Tasks. In NeurIPS, 2019. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13600" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[29] Zixian Ma, Jerry Hong, Mustafa Omer Gul, Mona Gandhi, Irena Gao, and Ranjay Krishna. CREPE: Can Vision-Language Foundation Models Reason Compositionally? In CVPR, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 117, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 117, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 117, + 287, + 161 + ], + "type": "text", + "content": "[30] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L. Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In CVPR, 2016. 1, 2, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 162, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 287, + 183 + ], + "type": "text", + "content": "[31] George A. Miller. Wordnet: A lexical database for english. Commun. ACM, 38(11):39-41, 1995. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 184, + 287, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 287, + 249 + ], + "type": "text", + "content": "[32] Zhixiang Min, Bingbing Zhuang, Samuel Schulter, Buyu Liu, Enrique Dunn, and Manmohan Chandraker. Neurocs: Neural nocs supervision for monocular 3d object localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21404-21414, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 250, + 287, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 287, + 315 + ], + "type": "text", + "content": "[33] Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, Xiao Wang, Xiaohua Zhai, Thomas Kipf, and Neil Houlsby. Simple Open-Vocabulary Object Detection with Vision Transformers. In ECCV, 2022. 1, 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 316, + 287, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 287, + 348 + ], + "type": "text", + "content": "[34] Matthias Minderer, Alexey Gritsenko, and Neil Houlsby. Scaling Open-Vocabulary Object Detection. In NeurIPS, 2023. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 349, + 184, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 349, + 184, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 349, + 184, + 360 + ], + "type": "text", + "content": "[35] OpenAI. Gpt-3.5. 2022. 2, 3, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 361, + 287, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 361, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 361, + 287, + 415 + ], + "type": "text", + "content": "[36] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k Entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In ICCV, pages 2641–2649, 2015. 1, 2, 3, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 416, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 416, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 416, + 287, + 480 + ], + "type": "text", + "content": "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 481, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 287, + 514 + ], + "type": "text", + "content": "[38] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks. In NeurIPS, 2015. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 515, + 287, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 515, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 48, + 515, + 287, + 548 + ], + "type": "text", + "content": "[39] Joshua Robinson, Ching-Yao Chuang, Suvrit Sra, and Stefanie Jegelka. Contrastive Learning with Hard Negative Samples. In ICLR, 2021. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 548, + 287, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 603 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 603 + ], + "type": "text", + "content": "[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 603, + 287, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 287, + 647 + ], + "type": "text", + "content": "[41] Samuel Schulter, Vijay Kumar BG, Yumin Suh, Konstantinos M Dafnis, Zhixing Zhang, Shiyu Zhao, Dimitris Metaxas, et al. OmniLabel: A Challenging Benchmark for Language-Based Object Detection. In ICCV, 2023. 1, 2, 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "type": "text", + "content": "[42] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Jing Li, Xiangyu Zhang, and Jian Sun. Objects365: A Large-scale, High-quality Dataset for Object Detection. In ICCV, 2019. 1, 3, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[43] Ravi Shekhar, Sandro Pezzelle, Yauhen Klimovich, Aurelie Herbelot, Moin Nabi, Enver Sangineto, and Raffaella" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "Bernardi. FOIL it! Find One mismatch between Image and Language caption. In ACL, 2017. 1, 2, 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 95, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 128 + ], + "type": "text", + "content": "[44] Abhinav Shrivastava, Abhinav Gupta, and Ross Girshick. Training Region-based Object Detectors with Online Hard Example Mining. In CVPR, 2016. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 129, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 129, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 308, + 129, + 545, + 172 + ], + "type": "text", + "content": "[45] Sanjay Subramanian, William Merrill, Trevor Darrell, Matt Gardner, Sameer Singh, and Anna Rohrbach. ReCLIP: A Strong Zero-Shot Baseline for Referring Expression Comprehension. In ACL, 2022. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 173, + 545, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 173, + 545, + 226 + ], + "spans": [ + { + "bbox": [ + 308, + 173, + 545, + 226 + ], + "type": "text", + "content": "[46] Tristan Thrush, Ryan Jiang, Max Bartolo, Amanpreet Singh, Adina Williams, Douwe Kiela, and Candace Ross. Winoground: Probing Vision and Language Models for Visio-Linguistic Compositionality. In CVPR, 2022. 1, 2, 4, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 228, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 293 + ], + "type": "text", + "content": "[47] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models, 2023. 2, 3, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 293, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 293, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 308, + 293, + 545, + 326 + ], + "type": "text", + "content": "[48] Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9 (86):2579-2605, 2008. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 327, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 381 + ], + "type": "text", + "content": "[49] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. OFA: Unifying Architectures, Tasks, and Modalities Through a Simple Sequence-to-Sequence Learning Framework. In ICML, 2022. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 382, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 382, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 382, + 545, + 415 + ], + "type": "text", + "content": "[50] Chenyun Wu, Zhe Lin, Scott Cohen, Trung Bui, and Subhransu Maji. PhraseCut: Language-based Image Segmentation in the Wild. In CVPR, 2020. 1, 2, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 415, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 415, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 308, + 415, + 545, + 459 + ], + "type": "text", + "content": "[51] Hao Wu, Jiayuan Mao, Yufeng Zhang, Yuning Jiang, Lei Li, Weiwei Sun, and Wei-Ying Ma. Unified Visual-Semantic Embeddings: Bridging Vision and Language with Structured Meaning Representations. In CVPR, 2019. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 459, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 459, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 308, + 459, + 545, + 502 + ], + "type": "text", + "content": "[52] Xiaoshi Wu, Feng Zhu, Rui Zhao, and Hongsheng Li. CORA: Adapting CLIP for Open-Vocabulary Detection with Region Prompting and Anchor Pre-Matching. In CVPR, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 503, + 545, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 503, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 308, + 503, + 545, + 548 + ], + "type": "text", + "content": "[53] Chi Xie, Zhao Zhang, Yixuan Wu, Feng Zhu, Rui Zhao, and Shuang Liang. Described Object Detection: Liberating Object Detection with Flexible Expressions. In NeurIPS, 2023. 1, 2, 5, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 548, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 581 + ], + "type": "text", + "content": "[54] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C. Berg, and Tamara L. Berg. Modeling Context in Referring Expressions. In ECCV, 2016. 1, 2, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "type": "text", + "content": "[55] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language models behave like bags-of-words, and what to do about it? In ICLR, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 625, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 657 + ], + "type": "text", + "content": "[56] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-Vocabulary Object Detection Using Captions. In CVPR, 2021. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "text", + "content": "[57] Shiyu Zhao, Zhixing Zhang, Samuel Schulter, Long Zhao, BG Vijay Kumar, Anastasis Stathopoulos, Manmohan Chandraker, and Dimitris N Metaxas. Exploiting unlabeled data with vision and language models for object detection. In ECCV, pages 159-175. Springer, 2022. 1, 2" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "13601" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 217 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[58] Shiyu Zhao, Samuel Schulter, Long Zhao, Zhixing Zhang, Vijay Kumar B. G, Yumin Suh, Manmohan Chandraker, and Dimitris N. Metaxas. Improving pseudo labels for open-vocabulary object detection, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 171 + ], + "type": "text", + "content": "[59] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, and Jianfeng Gao. RegionCLIP: Region-based Language-Image Pretraining. In CVPR, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 174, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 174, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 174, + 287, + 217 + ], + "type": "text", + "content": "[60] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbihl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In ECCV, pages 350-368. Springer, 2022. 5" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "13602" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Generating Handwritten Mathematical Expressions From Symbol Graphs_ An End-to-End Pipeline/f14645bb-d996-46da-9c4b-81a14daf5714_content_list.json b/2024/Generating Handwritten Mathematical Expressions From Symbol Graphs_ An End-to-End Pipeline/f14645bb-d996-46da-9c4b-81a14daf5714_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..684e348cfc87f472e84b225ef4efda9731aceba9 --- /dev/null +++ b/2024/Generating Handwritten Mathematical Expressions From Symbol Graphs_ An End-to-End Pipeline/f14645bb-d996-46da-9c4b-81a14daf5714_content_list.json @@ -0,0 +1,1602 @@ +[ + { + "type": "text", + "text": "Generating Handwritten Mathematical Expressions From Symbol Graphs: An End-to-End Pipeline", + "text_level": 1, + "bbox": [ + 107, + 128, + 864, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yu Chen $^{1,*}$ , Fei Gao $^{2,*}$ , Yanguang Zhang $^{3}$ , Maoying Qiao $^{4}$ , Nannan Wang $^{5,\\dagger}$", + "bbox": [ + 181, + 202, + 785, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Beijing Waiyan Online Digital Technology; 2 Hangzhou Institute of Technology, Xidian University; 3 Hangzhou Dianzi University; 4 University of Technology, Sydney (UTS); 5 Xidian University", + "bbox": [ + 88, + 220, + 880, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{fgao, nwwang}@xidian.edu.cn, ppakl991@outlook.com, maoying.qiao@uts.edu.au", + "bbox": [ + 150, + 258, + 813, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 308, + 313, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we explore a novel challenging generation task, i.e. Handwritten Mathematical Expression Generation (HMEG) from symbolic sequences. Since symbolic sequences are naturally graph-structured data, we formulate HMEG as a graph-to-image (G2I) generation problem. Unlike the generation of natural images, HMEG requires critic layout clarity for synthesizing correct and recognizable formulas, but has no real masks available to supervise the learning process. To alleviate this challenge, we propose a novel end-to-end G2I generation pipeline (i.e. graph $\\rightarrow$ layout $\\rightarrow$ mask $\\rightarrow$ image), which requires no real masks or nondifferentiable alignment between layouts and masks. Technically, to boost the capacity of predicting detailed relations among adjacent symbols, we propose a Less-is-More (LiM) learning strategy. In addition, we design a differentiable layout refinement module, which maps bounding boxes to pixel-level soft masks, so as to further alleviate ambiguous layout areas. Our whole model, including layout prediction, mask refinement, and image generation, can be jointly optimized in an end-to-end manner. Experimental results show that, our model can generate high-quality HME images, and outperforms previous generative methods. Besides, a series of ablations study demonstrate effectiveness of the proposed techniques. Finally, we validate that our generated images promisingly boosts the performance of HME recognition models, through data augmentation. Our code and results are available at: https://github.com/AiArt-HDU/HMEG.", + "bbox": [ + 76, + 340, + 473, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 791, + 209, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Handwritten Mathematical Expressions (HMEs) are common and play significant roles in our daily life, especially in the research and education areas. HMEs generally present complex structures, serious deformations, and diverse writ", + "bbox": [ + 76, + 816, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fcec784f4caba580b21f7fef86fdc9591e1206d1c2a25c48a2fbe88fb62c1301.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 308, + 890, + 395 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/00a7884c7514022f7c0d291821a5b25e0030444fd17480240baa6d776ccf0ea7.jpg", + "image_caption": [ + "Figure 1. Differences between (a) typical two-stage graph-to-image generation pipeline and (b) our end-to-end pipeline. In previous methods, the real masks are available as the input. In contrast, we propose a novel end-to-end pipeline of $\\text{graph} \\to \\text{layout} \\to \\text{mask} \\to \\text{image}$ , and requires no real masks or non-differentiable alignment between layouts and masks." + ], + "image_footnote": [], + "bbox": [ + 503, + 398, + 890, + 489 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing styles. Such characteristics, along with data scarcity, make HME Recognition (HMER) a grand challenge in the OCR community. Despite the tremendous efforts that have been made to this task, the HMER performance is still unsatisfactory [14, 27, 62].", + "bbox": [ + 496, + 612, + 893, + 688 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, synthetic data augmentation has shown inspiring performance in various recognition tasks [23, 51]. This is mainly achieved by the remarkable progress of conditional image generation [21] in the past few years [58]. The advances allow for creating realistic images through texts [39, 57, 66], scene graphs [24, 60], layouts [28, 69], semantic masks [37, 72], sketches [61, 67], and more [12, 65]. However, there has been no effective generative model for generating high-quality HME images.", + "bbox": [ + 496, + 689, + 893, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Some previous works synthesize HMEs by recomposing real online HMEs [46, 59]. Besides, FormulaGAN [41] generates HMEs from rendered images, via an Image-to-Image (I2I) Translation model. However, their generated images are limited in either diversity or realism. Be", + "bbox": [ + 496, + 825, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal Contributions. † Corresponding Author.", + "bbox": [ + 99, + 886, + 357, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15675", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sides, Mathematical Expressions (MEs) are highly structured data. Given an HME, its meaning and handwritten style, depend not only on the category and shape of each symbol, but also on the positional relations between symbols. It's unlikely to design an elaborate I2I translation model that can handle all these issues.", + "bbox": [ + 75, + 90, + 470, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To combat this challenge, in this paper, we explore a new generation task, i.e. Handwritten Mathematical Expression Generation (HMEG) from symbol graphs of symbolic sequences. It's natural and easy to use a graph, specifying symbols and positional relations between symbols, to represent an ME [63]. Besides, previous works have demonstrated the significance of fully mining the structural information in boosting generalization [52]. By converting symbolic sequences to symbol graphs, we propose a novel HMEG method, inspired by the classic graph-to-image (G2I) generation work, $\\mathrm{Sg2im}$ [24]. Similar to existing G2I generation works, our method consists of three stages: graph-to-query prediction, layout-to-mask transformation, and mask-to-image generation.", + "bbox": [ + 75, + 184, + 468, + 396 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Compared to the generation of natural images, the key challenge of HMEG is to generate unambiguous structural layouts, for accurately presenting meanings of formulas. From one hand, the requirement of layout clarity is much critical for generating formulas. Not only the size of each symbol, but also the positional relations between symbols, should be precisely predicted. Otherwise, mathematical meanings might change or become unrecognisable. For example, an input formula $x^{2}$ might become $x_{2}$ or $x2$ in a generated image. From the other hand, MEs have infinite possible layouts and structures. Namely, a formula could comprise of arbitrary number of symbols and arbitrary relations between symbols. It's impossible to collect completed training data for learning the layout predictor.", + "bbox": [ + 75, + 397, + 468, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To alleviate these difficulties, we first propose a Less-is-More (LiM) training strategy for learning an effective layout predictor. Specially, our layout predictor follows the structure of Graph Convolutional Networks (GCNs) [30]. During training, we only use minimal graphs with 1-degree connections, in stead of whole graphs with complex structures, to optimize the layout predictor. Note that the learned predictor can be applied to arbitrary symbol graphs, in the testing stage. By using LiM, the predictor emphasizes on local structures between adjacent symbols. Besides, the learning process is eased. Namely, we could learn an effective layout predictor by using moderate amount of training samples. Such benefits will be verified in the ablation study part. Besides, we add a layout discriminator to further boosts the realism of predicted layouts.", + "bbox": [ + 75, + 611, + 468, + 838 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition, we propose a Sequential BBox-to-Mask Transformation (B2M) module, for refining layouts under weak supervision. In the original layout, bounding boxes of symbols tend to overlap with each other. Such overlapped", + "bbox": [ + 76, + 839, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "areas may lead to overlapping symbols or chaotic strokes, in generated images. It's critical to refine coarse layouts to pixel-level soft masks, to alleviate the overlapping problems [22]. Most previous G2I generation works require real segmentation masks for learning refined layout masks [13, 22]. However, it's difficult to define the mask of a symbol, because white pixels around strokes are essential for representing a symbol. We have to refine the layout under weak supervision. To this end, we propose to learn a sequential mapping of layout $\\rightarrow$ grid $\\rightarrow$ mask, and use an subsequent image decoder to generate an HME image. During training, we use the image-related losses to train both B2M and the image decoder jointly.", + "bbox": [ + 496, + 90, + 890, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, we propose a novel end-to-end pipeline of graph $\\rightarrow$ layout $\\rightarrow$ mask $\\rightarrow$ image for G2I generation tasks. Our model requires no supervision of real masks, or interactive alignment operations between layouts and masks. The major differences between our method and previous typical G2I generation works, are illustrated in Fig. 1. To verify the effectiveness of the proposed techniques, we conduct experiments on CROHME2014/2016/2019 [34]. Results show that, our model can generate high-quality HME images with clear layouts and recognisable symbols. Besides, our model significantly outperforms a number of existing generative methods, both qualitatively and quantitatively. A series of ablation study demonstrate the effectiveness of the proposed techniques, including the LiM, layout discriminator, and B2M. Finally, our generated HME images prove to boost the HMER performance by $5 - 11\\%$ absolutely, through synthetic data augmentation.", + "bbox": [ + 496, + 287, + 892, + 544 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 558, + 648, + 573 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "HME Recognition (HMER). Most advanced methods treat HMER as an image-to-markup problem [53]. They typically use Convolutional Neural Networks (CNNs) or Recurrent Neural Networks (RNNs) [18] to encode input images; and use attention-based models, such as Gated Recurrent Units (GRU) [5, 14, 68], or Transformers [2, 47, 70], to generate LaTeX sequences. Besides, some methods have been proposed [41, 46, 59]. Some recent works propose to use additional information, e.g. symbol-level counting [27], or relation-level counting [14], or emphasis on visually similar symbols [33], or synthetic data augmentation, to boost the HMER performance. Another branch of HMER are graph-based methods. These methods typically transform LaTeX sequences to Symbol Label Trees (SLTs) [63], and use tree-structured decoders to obtain the outputs [14, 52, 54]. These works validate the capacity of graphs in representing the structural information of formulas. In this paper, we therefore convert symbolic sequences to graphs, and use a GCN to encode them for generating HME images.", + "bbox": [ + 496, + 583, + 890, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Graph-to-Image Generation. Image generation from scene graphs is first proposed in Sg2im [24]. The pipeline", + "bbox": [ + 500, + 869, + 890, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15676", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8c4dbbfe45917d3b66d51ac53881035d8b16b457cb7320288bfb3881619c192d.jpg", + "image_caption": [ + "Figure 2. Overview of our handwritten mathematical expression generator (HMEG)." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 893, + 250 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of Sg2im includes two stages: graph-to layout mapping and layout-to-image generation. Various subsequent methods have been proposed to boost the quality of generated images, following the same pipeline [9, 13, 16, 40]. Besides, some semi-parametric methods use crops from real images to boost the quality of generated images [31]. There are also some extensions to image manipulation [9, 42].", + "bbox": [ + 75, + 299, + 470, + 405 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Most of these methods use a image discriminator and a object discriminator during training. Besides, they use real layouts, segmentation masks, and images, to supervise the learning process [13, 22]. COLoR [22] additionally uses a mask discriminator to boost the precision of generated masks. Recently, KCGM [55] uses knowledge consensus to alleviate the dependency on real layouts or masks. Besides, several diffusion model based methods [12, 60] can generate high-quality images directly from input scene graphs. However, diffusion based methods still suffer the high cost during both training and inference. Besides, generating recognisable texts is still challenging for most diffusion models.", + "bbox": [ + 75, + 409, + 468, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Layout-to-Image Generation. Most layout-to-image generation methods separately predict the mask of each object based on object embeddings, and require frequent resizing interaction between bounding boxes and masks [43, 44, 69]. To improve the quality of generated masks, several methods use masks discriminator [1, 28], and requires real masks during training. Besides, OC-GAN [45] uses both global and local similarities between scene graphs and images are used to improve the realism of generated layouts. Similar to our method, LAMA [32] tries to alleviate the overlapping between objects. However, in all these methods, the real bounding boxes are available as the input. They at most implement the end-to-end computing of layout $\\rightarrow$ mask $\\rightarrow$ image. In contrast, we propose a novel end-to-end computing of graph $\\rightarrow$ layout $\\rightarrow$ mask $\\rightarrow$ image, and requires no interactive alignment operations between layouts and masks.", + "bbox": [ + 75, + 609, + 468, + 866 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Handwritten Font Generation. Another related task is Handwritten Font Generation (HFG), which aims at", + "bbox": [ + 76, + 869, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "transferring images of fonts to the target handwritten style [48, 50]. Most methods follow the I2I translation pipeline, and adopt encoder-decoder architectures [6]. Besides, great efforts have been made to solve the few-shot learning challenge [48, 64]. These methods focus on the style of strokes. In contrast, our work aims to generate rational layouts of multiple symbols, with complex relations, at the same time. It's interesting to explore generating HME images in personal styles in the future.", + "bbox": [ + 496, + 299, + 893, + 435 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 452, + 589, + 467 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 500, + 477, + 609, + 491 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to generate HME images conditioned on LaTeX sequences. To this end, we represent LaTeX sequences by symbol graphs [63] and design a novel G2I generation method to synthesize HME images. Our full model follows the idea of Generative Adversarial Networks (GANs) [21, 24]. It includes a generator, and several discriminators in the training stage. Fig. 2 shows the pipeline of our generator. It mainly include four parts: (i) First, the input LaTeX sequence is converted to a symbol graph, and embedded into high-dimensional feature vectors. (ii) A GCN based layout predictor estimates the coarse layout, i.e. bounding box (BBox) of each symbol. (iii) A sequential layout refinement module, including BBox-to-Grid mapping and Grid-to-Mask projection, generate a pixel-level soft mask for each symbol. (iv) The image decoder, a Cascaded Refinement Network (CRN) [4], generates an image conditioned on the predicted mask.", + "bbox": [ + 496, + 501, + 890, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Symbol Graph Construction", + "text_level": 1, + "bbox": [ + 500, + 770, + 754, + 786 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a set of mathematical symbol categories $\\mathcal{C}$ and a set of relation categories $\\mathcal{R}$ , a symbol graph $(S,E)$ can be constructed based on $\\mathcal{C}$ and $\\mathcal{R}$ : each node $i$ in $S = \\{s_1,\\dots ,s_n\\}$ is a mathematical symbol, with $s_i\\in \\mathcal{C}$ ; and each directed edge $(i,j)$ in $E = \\{e_{ij}\\}$ represents the positional relation of node $j$ w.r.t. node $i$ , with $e_{ij} = (s_i,r,s_j)$ and $r\\in \\mathcal{R}$ . $n$ is the number of symbols in a mathematical", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15677", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4f44b67ea813ce3720724f157f613465b666c53d36143b5da3335430baac276c.jpg", + "image_caption": [ + "Figure 3. Architecture of our layout predictor and the Graph Embedding Model (GEM) [30]." + ], + "image_footnote": [], + "bbox": [ + 223, + 87, + 336, + 131 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "expression. Let $m$ denote the number of edges in $E$ . For example, given an expression $x^2$ , it can be represented by a tuple $(x, \\sup, 2)$ , where $\\sup$ is short for superscript. Besides, we add an additional edge type None to indicate that the corresponding pair of symbols are unrelated.", + "bbox": [ + 75, + 244, + 468, + 320 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Afterwards, we map the nodes and edges $(S,E)$ to high-dimensional feature vectors $\\{\\mathbf{V},\\mathbf{U}\\}$ by using a learnable embedding layer, respectively [30]. The learned embedding feature vectors are denoted by:", + "bbox": [ + 75, + 320, + 468, + 381 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {V} = \\left\\{\\mathbf {v} _ {i}, \\forall i \\in S \\right\\}, \\mathbf {U} = \\left\\{\\mathbf {u} _ {i j}, \\forall (i, j) \\in E \\right\\}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 393, + 468, + 411 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{v}_i, \\mathbf{u}_{ij} \\in \\mathbb{R}^d$ associate with $s_i, e_{ij}$ , respectively; $d$ is the dimension of embedding vectors.", + "bbox": [ + 75, + 421, + 468, + 452 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Adversarial Layout Prediction", + "text_level": 1, + "bbox": [ + 76, + 460, + 346, + 477 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired by GAN [21], we train our layout prediction module in an adversarial learning manner. Specially, we use an layout predictor $G_{box}$ to generate a layout in the format of bounding boxes, conditioned on a given symbol graph; and use an additional layout discriminator $D_{box}$ to judge whether a layout is real or fake. Using the adversarial learning strategy like GAN, the layout predictor would finally predict high-quality bounding boxes.", + "bbox": [ + 75, + 484, + 468, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Layout Predictor. We first use the Graph Embedding Model (GEM) [30] to update features of nodes and edges, and then use a MLP to predict the BBox indices of each symbol. First, original embedding features are encoded by: $\\mathbf{h}_i^{(0)} = f_{\\mathrm{node}}(\\mathbf{v}_i), \\forall i \\in S$ and $\\mathbf{e}_{i,j} = f_{\\mathrm{edge}}(\\mathbf{u}_{ij}), \\forall (i,j) \\in E$ . Afterward, node features are updated through a stack of propagation layers, by accumulating information in local neighborhood. Features of node $i$ at the $t$ -th layer are updated by:", + "bbox": [ + 75, + 604, + 468, + 743 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {m} _ {j \\rightarrow i} = f _ {\\text {m e s s a g e}} \\left(\\mathbf {h} _ {i} ^ {(t)}, \\mathbf {h} _ {j} ^ {(t)}, \\mathbf {e} _ {i j}\\right), \\forall (i, j) \\in E,\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 753, + 426, + 775 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {h} _ {i} ^ {(t + 1)} = f _ {\\text {n o d e}} \\left(\\mathbf {h} _ {i} ^ {(t)}, \\sum_ {j} \\mathbf {m} _ {j \\rightarrow i}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 776, + 468, + 810 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After $T$ rounds of propagations, the model produces an aggregated graph level representation by:", + "bbox": [ + 76, + 821, + 468, + 853 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {h} _ {G} = f _ {G} \\left(\\sum_ {i} \\sigma \\left(f _ {\\text {g a t e}} \\left(\\mathbf {h} _ {i} ^ {(T)}\\right)\\right) \\odot f _ {\\text {u p d a t e}} \\left(\\mathbf {h} _ {i} ^ {(T)}\\right)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 863, + 468, + 904 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\sigma$ denotes the Sigmoid operation. Finally, we integrate the updated embedding features and the graph level features for predicting the BBox indices of each symbol. Besides, we concatenate a Gaussian noise vector $\\mathbf{z}_i\\in \\mathbb{R}^k,\\forall i\\in S$ to these features, to improve the diversity in predicted layouts. The BBox prediction is formulated by:", + "bbox": [ + 498, + 90, + 890, + 181 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {b}} _ {i} = f _ {\\mathrm {b b o x}} \\left(\\mathbf {h} _ {i} ^ {(T)}, \\mathbf {h} _ {G}, \\mathbf {z} _ {i}\\right), \\forall i \\in S. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 193, + 890, + 212 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Each of the mapping functions $f_{*}$ above is an MLP on the inputs [29, 30]. For each symbol, we obtain a 4-dimensional rectangular box $[x_0, y_0, x_1, y_1]$ in the $[0,1]$ coordinate space, denoting the coordinates of the upper left and lower right corners of the BBox border.", + "bbox": [ + 496, + 223, + 890, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Layout Discriminator. The layout discriminator $D_{box}$ is also composed of GEM [30], followed by an MLP to predict whether an input layout as real or fake. $D_{box}$ takes concatenations of the symbol embedding matrix $\\mathbf{V} \\in \\mathbf{R}^{n \\times d}$ and the layout indices $B \\in \\mathbf{R}^{n \\times 4}$ as input. In the implementation, we use the least square version of adversarial loss, i.e.", + "bbox": [ + 496, + 299, + 890, + 404 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {a d v} ^ {b o x} = \\left\\| D _ {b o x} (\\mathbf {V}, B) \\right\\| _ {2} ^ {2} + \\left\\| 1 - D _ {b o x} (\\mathbf {V}, \\hat {B}) \\right\\| _ {2} ^ {2}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 415, + 890, + 434 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Besides, we use the L2 distance between the predicted BBox indices $\\hat{B}$ and the target (real) indices $B$ as the layout reconstruction loss, i.e. $\\mathcal{L}_{box} = ||B - \\hat{B}||_1$ . We use a weighted sum of $\\mathcal{L}_{box}$ and $\\mathcal{L}_{adv}^{box}$ as the objective of $G_{box}$ . During training, $G_{box}$ and $D_{box}$ are alternatively optimized.", + "bbox": [ + 496, + 444, + 890, + 520 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Less-is-More Learning (LiM) Strategy. A mathematical expression could include arbitrary number of symbols or relations between symbols. In other words, there are infinite possible structures of symbol graphs. It's difficult to collect a complete training set. Besides, using complicated graphs during training may add the difficulty in learning an effective model. We therefore propose a Less-is-More (LiM) learning strategy [3]. Specially, we propose to use symbol graphs with merely 1-degree connections during training. In this way, the layout predictor is expected to emphasize on local structures, i.e. generating rational layout for locally connected symbols. While in the testing stage, the learned model can be applied to arbitrary symbol graphs, with arbitrary degrees of connections, and predict the corresponding layout.", + "bbox": [ + 496, + 521, + 890, + 747 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Sequential BBox-to-Mask Transformation", + "text_level": 1, + "bbox": [ + 500, + 756, + 857, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We next predict the pixel-level soft mask of an expression based on the predicted bounding boxes. This procedure is illustrated in Fig. 4, and consists of the following two steps:", + "bbox": [ + 496, + 780, + 890, + 825 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(1) BBox-to-Grid Mapping. We first transform the BBox indices to layout grids, as illustrated in Fig. 4. To this end, we subscribe each index value from the coordinate matrices $\\mathbf{X}$ and $\\mathbf{Y}$ in the [0, 1] coordinate space. For each symbol, we obtain a $4 \\times w \\times w$ tensor of grid.", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15678", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/028bf7dcbd64b165c184a53c78cc2a1f474979fb3652d96f64e40062449abbb4.jpg", + "image_caption": [ + "Figure 4. Illustration of the sequential BBox-to-Mask transformation (B2M) module." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 467, + 204 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(2) Grid-to-Mask Projection. Afterwards, we use a FCN with a Sigmoid output layer (denoted by $\\phi_{\\mathrm{mask}}$ ) to estimate a one-channel mask for each symbol, i.e.", + "bbox": [ + 76, + 268, + 468, + 314 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {i} = \\phi_ {\\text {m a s k}} (\\mathbf {X} - x _ {i, 0}, \\mathbf {X} - x _ {i, 1}, \\mathbf {Y} - y _ {i, 0}, \\mathbf {Y} - y _ {i, 1}). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 325, + 468, + 342 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Each value in $\\mathbf{M}_i$ indicates the probability of a pixel belonging to symbol $s_i$ . Specially, we first use two Convolutional layers to predict a low-resolution mask of dimension $w \\times w$ , and then use two upsampling Convolutional layers to sequentially upsampling the mask by a factor of 2. These masks are used in the image decoder for generating an image of the same size, respectively.", + "bbox": [ + 76, + 352, + 468, + 458 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note: Previous G2I generation works [9] also consists of BBox-to-Grid mapping, but they merely use the grids to select patches of objects. Besides, they predict the mask from the BBox indices $\\hat{B}$ or original embedding features $(\\mathbf{V},\\mathbf{U})$ . Differently, we use the transformed grids for layout mask prediction. Moreover, since the whole BBox-to-Mask prediction module is differentiable, our whole model enables training in an end-to-end manner.", + "bbox": [ + 75, + 458, + 468, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Image Generation", + "text_level": 1, + "bbox": [ + 76, + 588, + 253, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Image Decoder. In the implementation, we adopt CRN [4] as the image decoder. The input is a multiplication of the whole layout mask $\\mathbf{M} \\in \\mathbb{R}^{n \\times w \\times w}$ with the embedding vectors $\\mathbf{V} \\in \\mathbb{R}^{n \\times d}$ . Besides, we add Gaussian noise to the feature maps to improve the diversity in image generation. CRN includes three upsampling Convolutional layers, and outputs a formula image with resolutions of $w \\times w$ . Recall that we actually predicts three masks at different scales. In the implementation, we have three generated images, at the resolution of $64 \\times 64$ , $128 \\times 128$ , and $256 \\times 256$ , respectively. All these images are used to calculate the losses for optimizing our model.", + "bbox": [ + 75, + 612, + 468, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Image and Symbol Discriminators Finally, we use an image discriminator $D_{img}$ and a symbol discriminator $D_{sym}$ during training, to boost the realism of the generated whole image $\\hat{I}$ and patches of symbols, respectively. We use the grids to select patches of symbols, and input them into $D_{sym}$ . Both discriminators are fully CNNs with the cross-entropy (CE) version of adversarial loss. In addition,", + "bbox": [ + 75, + 792, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "following AC-GAN [36], we have the symbol discriminator additionally predict category labels of symbols. In this way, the generated symbols would be more recognizable and real.", + "bbox": [ + 496, + 90, + 890, + 150 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Loss Functions. Similar to previous G2I generation works [24], we use a series of loss functions to optimize the image generation. The losses are briefly introduced below.", + "bbox": [ + 496, + 151, + 890, + 196 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Pixel loss $\\mathcal{L}_{pix}$ : First, we use the L1 distance between a generated image and the target formula image and as the pixel loss: $\\mathcal{L}_{pix} = ||I - \\hat{I}||_1$ .", + "- Image adversarial loss $\\mathcal{L}_{adv}^{img}$ : Second, we use the global image discriminant loss from $D_{img}$ , to encourage the generated HME images in the real handwritten style.", + "- Symbol adversarial loss $\\mathcal{L}_{adv}^{sym}$ : Third, we use the symbol discriminative loss from $D_{sym}$ , to improve the fidelity of generated symbols.", + "- Auxiliary Symbol Recognition Loss $\\mathcal{L}_{aux}^{sym}$ : Finally, we use the auxiliary symbol recognition loss from $D_{sym}$ , to enforce the generator producing recognizable symbols." + ], + "bbox": [ + 500, + 200, + 890, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use a combination of these loss functions above, to train the B2M module and the image generation modules together. The total loss is formulated by:", + "bbox": [ + 496, + 387, + 890, + 431 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {i m g} = \\lambda_ {1} \\mathcal {L} _ {p i x} + \\lambda_ {2} \\mathcal {L} _ {a d v} ^ {i m g} + \\lambda_ {3} \\mathcal {L} _ {a d v} ^ {s y m} + \\lambda_ {4} \\mathcal {L} _ {a u x} ^ {s y m}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 444, + 890, + 462 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{i}, i = 1, \\ldots, 4$ are weighting factors and are set as 1, 0.01, 0.01, 0.001, respectively.", + "bbox": [ + 496, + 473, + 890, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 518, + 633, + 536 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 500, + 544, + 705, + 560 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data. We use CROHME2014/2016/2019 [34] to validate our HMEG method in terms of both generalization and recognition. The dataset collects pen trajectories, and labels them with symbols and symbol categories. The dataset contains 126 different symbol categories, and 8 positional relations, i.e. start, left superscript, superscript, subscript, below, above, right and end. By restoring the pen trajectory information, we can easily obtain the HME image and the BBox of each symbol. We use the standard partition of each dataset through all the experiments.", + "bbox": [ + 496, + 566, + 890, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. In the implementation, due to the limitation of our computing resources, we train our model in two stages: we first train the layout prediction part in the first stage, and then train the rest parts by fixing it. In the first stage, we train the model with a batch size 64 and a learning rate 5e5, for 40,000 iterations. In each iteration, we random sample a batch of 1-degree sub-graphs from the training set for learning the layout predictor. In the second stage, we use a batch size 8, a learning rate 1e4, and train for 600,000 iterations. Our codes are implemented by using Pytorch and a NVIDIA TITAN XP GPU. We use Adam [26] as the optimizer for all networks.", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15679", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/613af9732d1422f27ff311e08036d68b141f2d2ec5ce7038b53bc23708f78a0c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Print|ax0+by0+c/√a2+b2|p=√a2+b2-2abcosA1/2(a+1/b) = 2ab/a+b-b+√b2-4ac/2a∫0πcos(θ/2)dθx/a+x/b-x/c∫bdIπr2h/3sin(π/3)=1/2
CycleGANaxn+byn+c/√n2+b2|p=√n2+b2-2abcosA1/2(a+1/b) = 2ab/a+b-b+√b2-4ac/2a∫0πcos(θ/2)dθx/a+x/b-x/c∫bdIπr2h/3sin(π/3)=1/2
FormulaGAN|ax0+by0+c/√a2+b2|p=√a2+b2-2abcosA1/2(a+1/b) = 2ab/a+b-b+√b2-4ac/2a∫0πcos(θ/2)dθx/a+x/b-x/c∫bdIπr2h/3sin(π/3)=1/2
Sg2imaxn+byn+c/√a2+b2|p=√a2+b2-2abcosA1/2(a+1/b) = 2ab/a+b-b+√b2-4ac/2a∫0πcos(θ/2)dθx/a+x/b-x/c∫bdIπr2h/3sin(π/3)=1/2
Ours|ax0+by0+c/√a2+b2|p=√a2+b2-2abcosA1/2(a+1/b) = 2ab/a+b-b+√b2-4ac/2a∫0πcos(θ/2)dθx/a+x/b-x/c∫bdIπr2h/3sin(π/3)=1/2
", + "bbox": [ + 80, + 89, + 890, + 404 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 5. Handwritten mathematical expressions generated by CycleGAN [71], FormulaGAN [41], Sg2im [24], and our method.", + "bbox": [ + 102, + 416, + 864, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Criteria. We choose two image quality assessment indices as the criteria. The former includes the Fréchet Inception distance (FID) [17] and Structural Similarity Measure (SSIM) [49]. Lower values of FID indicate better realism of synthesised images; while greater values of SSIM generally indicate higher similarity between a synthesized image and the corresponding real image. We here report the average SSIM value of all the test samples. In addition, we use two HMER indices, including the Word Error Rate (WER) and the Expression Recognition Rate (ExpRate). Here we use an open-source HMER method [19], to recognize HME images and compute the corresponding indices. Lower WER and higher ExpRate values indicate better performance.", + "bbox": [ + 75, + 457, + 472, + 654 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with SOTAs", + "text_level": 1, + "bbox": [ + 76, + 664, + 305, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first compare with the existing GAN based HMEG method, FormulaGAN [41], our baseline, Sg2im [24], and the widely used unsupervised I2I translation method, CycleGAN [71], on CHROME2019. We train and test these methods following the same experimental settings as ours.", + "bbox": [ + 75, + 688, + 468, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Comparison. Fig. 5 shows HME images generated by different models. CycleGAN fails to transform input formulas to the handwritten style. FormulaGAN produces handwritten-style images, but with blurring and unnatural strokes. In contrast, both Sg2im and our method generate high quality images, with realistic styles and clear strokes. In addition, the images generated by our method present better layouts than those by Sg2im. Sg2im sometimes produce very small symbols, e.g. $y_{0}$ in the first col", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/715defda5b99a43211912c1a82eacb91b5b14817598596f721e218481c882fa5.jpg", + "table_caption": [ + "Table 1. Comparison with existing methods on CHROME2019." + ], + "table_footnote": [], + "table_body": "
SSIM↑FID ↓WER ↓ExpRate ↑
CycleGAN [71]0.75784.140.6710.026
FormulaGAN [41]0.72474.680.6010.066
Sg2im [24]0.78710.020.3930.219
Ours0.79310.980.3260.316
", + "bbox": [ + 501, + 479, + 898, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "unn and $p$ in the second column. Besides, Sg2im may produce overlapped symbols, e.g. $\\frac{1}{2} \\left( \\frac{1}{a} + \\frac{1}{b} \\right)$ in the third column. Finally, Sg2im produces more unrecognisable symbols than ours, e.g. Sin and $\\pi$ in last column. In contrast, the symbols generated by our method present reasonable sizes and clear structures. Besides, symbols are reasonably placed in relation to each other, neither overlapping nor far apart. This is similar to real formulas written by humans. Such observations demonstrate the superiority of our method in generating both high-quality layouts and HME images.", + "bbox": [ + 496, + 595, + 890, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Comparison. Table 1 shows the quantitative performance indices. Both Sg2im and our method achieves significant better indices than CycleGAN or FormulaGAN. This is consistent with the qualitative comparison results. According to SSIM and FID, Sg2im and our method almost perform equality in term of image quality. However, our method decreases the WER by 6.5 percent and improves the ExpRate by 10 percent, absolutely. Such distinct superiority in recognition implies that, the handwritten expressions generated by our method are signifi", + "bbox": [ + 496, + 750, + 892, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15680", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b6559d3ca4776bbde12b0e6e0f99be7f8555c3f5b379b2742eb80994e2df6aff.jpg", + "image_caption": [ + "Figure 6. Illustration of the layouts and formula images generated by model variants, in the ablation study. The input formula is $\\frac{1}{25} y^2 -\\frac{8}{25} y$" + ], + "image_footnote": [], + "bbox": [ + 76, + 87, + 893, + 280 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8206013411e259c7a821c90375dc5cb86609c37dc2c8a76253dafe39a0ceaac2.jpg", + "image_caption": [ + "Figure 7. Comparison with diffusion models. DiffSketcher and DiffAE generate HMEs conditioned on print formulas; while ControlNet and SD+LoRA take both the print formula and a textual prompt (\"a handwritten mathematical expression of latex code\") as input. We use official diffusion models, and fine-tune DiffAE and SD+LoRA using HMEs." + ], + "image_footnote": [], + "bbox": [ + 81, + 324, + 468, + 435 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "cantly better than $\\mathrm{Sg2im}$ , in terms of clarity and structure.", + "bbox": [ + 75, + 558, + 459, + 573 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with Diffusion Models. We additionally compare with ControlNet [67], Stable Diffusion (SD) [39] + LoRA [20], DiffSketcher [56], and DiffAE [38]. As shown in Fig. 7, these diffusion models cannot generate high-quality HMEs with realistic handwritten styles or recognizable symbols. Besides, their computational complexity is much heavier than ours. Some recent methods, e.g. ChiroDiff [8] and SDT [7], can only generate handwritten images of one single character. Such comparison results demonstrate the advantage of our method in HMEG.", + "bbox": [ + 75, + 574, + 468, + 726 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 738, + 230, + 753 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further analyse the impact of (i) the LiM training strategy, (ii) the layout discriminator $(D_{box})$ , and (iii) the B2M module. To this end, we build model variants by:", + "bbox": [ + 75, + 762, + 468, + 808 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Using original training graphs, instead of sampled minimal graphs, to train the layout predictor $(w / o LiM)$ ;", + "- Removing B2M from our full model, but using grids for embedding selection (w/o B2M);", + "- Removing both LiM and B2M from our full model ( $w / o$ LiM/B2M);" + ], + "bbox": [ + 76, + 810, + 467, + 898 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ae84d0b94014a264ed9d711098575ba8fdcb0f93aa8fdf84c9c65b8f141c6d72.jpg", + "table_caption": [ + "Table 2. Performance indices w.r.t. the ablation study on CHROME19. The model variant in the last row is our baseline." + ], + "table_footnote": [], + "table_body": "
SSIM↑FID↓WER↓ExpRate↑mIOU↑
Ours (full)0.79310.980.3260.3160.364
w/o LiM0.79011.550.3270.3150.335
w/o B2M0.79011.460.3320.3080.354
w/o LiM/B2M0.78613.300.3650.2790.338
Sg2im (base)0.78710.080.3930.2190.324
", + "bbox": [ + 501, + 366, + 898, + 467 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "- Training the whole model following Sg2im, without using LiM, $D_{box}$ , and B2M (Sg2im (base)).", + "bbox": [ + 498, + 491, + 890, + 521 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We additionally report the mean Intersection over Union (mIoU) between the predicted layouts and the target ones.", + "bbox": [ + 498, + 521, + 890, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Table 2, removing LiM or B2M slightly changes the indices about image quality, decreases the IoU. Specially, removing LiM decreases the mIoU by about 3 percent absolutely. This implies the significance of LiM in generating good layouts. In addition, if we removing $D_{box}$ or B2M, along with LiM, the performance seriously decreases, almost in terms of all performance indices. This implies that our $D_{box}$ and B2M are significant for boosting the quality of generated images from all aspects, including the fidelity (SSIM), the usability for recognition (WER & ExpRate), and the layout (mIoU). Fig. 6 illustrates the layouts and images generated by different model variants. All the model variants, except the full model, produce overlapped symbols or symbols in unnatural sizes. The image generated by our model presents no obvious defects and are easy to recognize. Such observations further demonstrate the significance of LiM, B2M, and layout discriminator in layout prediction and image generation.", + "bbox": [ + 496, + 551, + 890, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Applications", + "text_level": 1, + "bbox": [ + 500, + 832, + 633, + 848 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Mathematical Expression Manipulation. In addition, we apply the previous learned model to expression manipulation using symbol graphs. Given a source symbol graph,", + "bbox": [ + 498, + 854, + 890, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15681", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b1a5bcfe6d8a2bc9cf3378b11c009bf0a75cf7bb6983b0e6e0509e2ec6690653.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Source1/3πr2hx/a + x/b - x/cax + bx + c/2a + √b + c/2∫_0^1 x^2y dxdyT_1^2/T_2^2 = a_1^3/a_2^3
Add1/3πr^2hzx + d/a + x/b - x/cax + b^x / 2a + √b + c/2 + 7∫_0^1 x^2y^2/2 dxdyT_1^2/T_2^2 = a_1^3 + 2/a_2^3
Delete1/3πr^2x/a + x/bax + b^x + ca + √b / 2∫_0^1 x^2y^2 dxdyT_1^2/T_2^2 = a/a
Change1/7πr^2hx/a + x/e - x/cax + b^y + c/2a + √b + c/4∫_0^1 x^2y dxdyT_1^2/T_2^2 = e^3/a^3
", + "bbox": [ + 80, + 89, + 465, + 272 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 8. Illustration of expression manipulation using symbol graphs (LaTeX sequences).", + "bbox": [ + 75, + 284, + 468, + 314 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "we edit it by adding, deleting, or changing symbol nodes in it, and apply the previously learned generator to produce the corresponding HME image. As shown in Fig. 8, the generated image changes consistently with the editing operation. Besides, the manipulated images present natural layouts, recognizable symbols, and clear strokes.", + "bbox": [ + 75, + 340, + 468, + 431 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Data Augmentation for HMER. Finally, we validate the possible application of HMEG in augmenting HMER models, since the generated data could improve the amount and diversity of training data. To this end, we conduct experiments on CROHME2014/2016/2019 [34], and use CAN [27] as the HMER model. Here, we train CAN with (1) the standard training data (denoted by CAN); or (2) with previous data augmentation, e.g. rotation, (denoted by $\\mathrm{CAN^{*}}$ ); or (3) with our synthetic data augmentation, using 6,000 additional generated samples (denoted by $\\mathrm{CAN^{\\dagger}}$ ); or (4) both previous and our data augmentation (denoted by $\\mathrm{CAN^{*\\dagger}}$ ). The model is trained for 120 epochs under each setting, using the official training parameters. Afterward, we apply the learned HMER model to both generated and real data.", + "bbox": [ + 75, + 431, + 468, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Table 3, our method achieves $5 - 11\\%$ absolute improvement over the baseline, and performs comparably to previous data augmentation, across all the datasets. Besides, our method can be jointly used with previous augmentation to further boost the performance (i.e. $\\mathrm{CAN}^{*\\dagger}$ ). We additionally report the HMER performance, as well as the SSIM and StruRate, of our generated samples. The SSIM and StruRate values are high, and the ExpRate values (by $\\mathrm{CAN}^{*} / \\mathrm{CAN}^{\\dagger}$ ) approach those of real data, on all datasets. These results demonstrate the robustness of our method, in generating high-quality HMEs, with correct layouts and recognizable symbols.", + "bbox": [ + 75, + 643, + 468, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Note: Previous HME synthesis methods [10, 25] synthesize HMEs by shuffling the symbols and layouts contained in original HME datasets. Neither of them is a generative method. And they may cause overfitting to existing datasets. In contrast, our pipeline allows excellent flexibil", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ab7ebd64649ed58309c86373d9b15e931cbd0ba763d65d698246da6611886d7c.jpg", + "table_caption": [ + "Table 3. Results on CROHME14/16/19 testing sets and our generated images. *,†, *† denote using previous data augmentation, our synthetic data augmentation, and both, respectively." + ], + "table_footnote": [], + "table_body": "
HMEG (generated)HMER (real)
SSIMCAN*CAN†StruRateCANCAN*CAN†CAN*†
CROHME140.78952.154.198.144.752.950.255.4
CROHME160.79851.455.896.542.852.453.957.6
CROHME190.79355.456.497.739.348.449.658.5
", + "bbox": [ + 500, + 142, + 893, + 219 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ity in generating diverse new samples. The experimental results also demonstrate the advantages of our approach in both HMEG and HMER.", + "bbox": [ + 498, + 250, + 890, + 292 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 313, + 625, + 328 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a novel method to solve the challenging task of generating HMEs from LaTeX sequences. We formulate it as a graph-to-image generation task, and focus on boosting the layout prediction without mask supervision. Experimental results demonstrate the effectiveness of our method, and its possible applications in expression manipulation and the challenging inverse task, i.e. HMER. Inspired by the remarkable process of generative augmentation in visual understanding [35], we believe that: the novel HMEG task we explore here has great potential to significantly boost HMER.", + "bbox": [ + 496, + 339, + 890, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Due to computational limitations, we merely use naive networks in the implementation. The generated symbols are occasionally unrecognizable, especially in complex formulas. It's promising to boost the generation quality by using advanced networks (e.g. VQ-GAN [11] or diffusion models [15]). Besides, we will extend the proposed techniques to natural/artistic image generation by exploring the inspiring capacity of diffusion models, and other tasks of structured graph data. Finally, to further boost the HMER performance, it's promising to use the triples of $\\{LaTeX, predicted layout, generated image\\}$ in HMEG as pseudo labeled data, or to use the layout discriminator for enhancing the layout detection module in an HMER pipeline.", + "bbox": [ + 496, + 507, + 892, + 704 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 722, + 666, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We greatly appreciate Biao Ma's and Chang Jiang's help in the implementation of HMER. This work was supported in part by the National Natural Science Foundation of China under Grants U22A2096 and 61971172; in part by the Technology Innovation Leading Program of Shaanxi under Grant 2022QFY01-15; in part by the Fundamental Research Funds for the Central Universities under Grant QTZX23042 and ZYTS24012; and in part by the Proof of Concept Foundation of Xidian University Hangzhou Institute of Technology under Grant No. GNYZ2023YL0301.", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15682", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Oron Ashual and Lior Wolf. Specifying object attributes and relations in interactive scene generation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4561-4569, 2019. 3", + "[2] Xiaohang Bian, Bo Qin, Xiaozhe Xin, Jianwu Li, Xuefeng Su, and Yanfeng Wang. Handwritten mathematical expression recognition via attention aggregation based bidirectional mutual learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 113-121, 2022. 2", + "[3] Stefan Bornholdt. Less is more in modeling large genetic networks. Science, 310(5747):449-451, 2005. 4", + "[4] Qifeng Chen and Vladlen Koltun. Photographic image synthesis with cascaded refinement networks. In Proceedings of the IEEE international conference on computer vision, pages 1511-1520, 2017. 3, 5", + "[5] Kyunghyun Cho, Bart Van Merrienboer, Dzmitry Bahdanau, and Yoshua Bengio. On the properties of neural machine translation: Encoder-decoder approaches. arXiv preprint arXiv:1409.1259, 2014. 2", + "[6] Gang Dai, Yifan Zhang, Qingfeng Wang, Qing Du, Zhuliang Yu, Zhuoman Liu, and Shuangping Huang. Disentangling writer and character styles for handwriting generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5977-5986, 2023. 3", + "[7] Gang Dai, Yifan Zhang, Qingfeng Wang, Qing Du, Zhuliang Yu, Zhuoman Liu, and Shuangping Huang. Disentangling writer and character styles for handwriting generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5977-5986, 2023. 7", + "[8] Ayan Das, Yongxin Yang, Timothy M Hospedales, Tao Xiang, and Yi-Zhe Song. Chirodiff: Modelling chirographic data with diffusion models. In The Eleventh International Conference on Learning Representations, 2023. 7", + "[9] Helisa Dhamo, Azade Farshad, Iro Laina, Nassir Navab, Gregory D Hager, Federico Tombari, and Christian Rupprecht. Semantic image manipulation using scene graphs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5213-5222, 2020. 3, 5", + "[10] Lan-Fang Dong, Han-Chao Liu, and Xin-Ming Zhang. Synthetic data generation and shuffled multi-round training based offline handwritten mathematical expression recognition. Journal of Computer Science and Technology, 37(6): 1427-1443, 2022. 8", + "[11] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 8", + "[12] Wan-Cyuan Fan, Yen-Chun Chen, DongDong Chen, Yu Cheng, Lu Yuan, and Yu-Chiang Frank Wang. Frido: Feature pyramid diffusion for complex scene image synthesis. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 579-587, 2023. 1, 3", + "[13] Azade Farshad, Sabrina Musatian, Helisa Dhamo, and Nassir Navab. Migs: Meta image generation from scene graphs. arXiv preprint arXiv:2110.11918, 2021. 2, 3" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Yingnan Fu, Wenyuan Cai, Ming Gao, and Aoying Zhou. Symbol location-aware network for improving handwritten mathematical expression recognition. In Proceedings of the 2023 ACM International Conference on Multimedia Retrieval, pages 516-524, 2023. 1, 2", + "[15] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10696-10706, 2022. 8", + "[16] Roei Herzig, Amir Bar, Huijuan Xu, Gal Chechik, Trevor Darrell, and Amir Globerson. Learning canonical representations for scene graph to image generation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVI 16, pages 210-227. Springer, 2020. 3", + "[17] M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In NIPS, pages 6626-6637. 2017. 6", + "[18] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997. 2", + "[19] Hongyu Wang and Jianshu Zhang. Handwritten mathematical expression recognition (pytorch), 2019. https://github.com/whywhs/Pytorch-Handwritten-Mathematical-Expression-Recognition.6", + "[20] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2021. 7", + "[21] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 1, 3, 4", + "[22] Maor Ivgi, Yaniv Benny, Avichai Ben-David, Jonathan Berant, and Lior Wolf. Scene graph to image generation with contextualized object layout refinement. In 2021 IEEE International Conference on Image Processing (ICIP), pages 2428-2432. IEEE, 2021. 2, 3", + "[23] Nikita Jaipuria, Xianling Zhang, Rohan Bhasin, Mayar Arafa, Punarjay Chakravarty, Shubham Shrivastava, Sagar Manglani, and Vidya N Murali. Deflating dataset bias using synthetic data augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 772-773, 2020. 1", + "[24] Justin Johnson, Agrim Gupta, and Li Fei-Fei. Image generation from scene graphs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1219-1228, 2018. 1, 2, 3, 5, 6", + "[25] Vu Tran Minh Khuong, Ung Quang Huy, Nakagawa Masaki, and Minh Khanh Phan. Generating synthetic handwritten mathematical expressions from a latex sequence or a mathml script. In 2019 International Conference on Document Analysis and Recognition (ICDAR), pages 922-927. IEEE, 2019. 8", + "[26] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In 3rd International Conference on" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15683", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 5", + "[27] Bohan Li, Ye Yuan, Dingkang Liang, Xiao Liu, Zhilong Ji, Jinfeng Bai, Wenyu Liu, and Xiang Bai. When counting meets hmer: counting-aware network for handwritten mathematical expression recognition. In European Conference on Computer Vision, pages 197-214. Springer, 2022. 1, 2, 8", + "[28] Wenbo Li, Pengchuan Zhang, Lei Zhang, Qiuyuan Huang, Xiaodong He, Siwei Lyu, and Jianfeng Gao. Object-driven text-to-image synthesis via adversarial training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12174-12182, 2019. 1, 3", + "[29] Yujia Li, Daniel Tarlow, Marc Brockschmidt, and Richard Zemel. Gated graph sequence neural networks. arXiv preprint arXiv:1511.05493, 2015. 4", + "[30] Yujia Li, Chenjie Gu, Thomas Dullien, Oriol Vinyals, and Pushmeet Kohli. Graph matching networks for learning the similarity of graph structured objects. In International conference on machine learning, pages 3835-3845. PMLR, 2019. 2, 4", + "[31] Yikang Li, Tao Ma, Yeqi Bai, Nan Duan, Sining Wei, and Xiaogang Wang. Pastegan: A semi-parametric method to generate image from scene graph. Advances in Neural Information Processing Systems, 32, 2019. 3", + "[32] Zejian Li, Jingyu Wu, Immanuel Koh, Yongchuan Tang, and Lingyun Sun. Image synthesis from layout with locality-aware mask adaption. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13819-13828, 2021. 3", + "[33] Zhe Li, Xinyu Wang, Yuliang Liu, Lianwen Jin, Yichao Huang, and Kai Ding. Improving handwritten mathematical expression recognition via similar symbol distinguishing. IEEE Transactions on Multimedia, 2023. 2", + "[34] Mahshad Mahdavi, Richard Zanibbi, Harold Mouchere, Christian Viard-Gaudin, and Utpal Garain. Icdar 2019 crohme+ tfd: Competition on recognition of handwritten mathematical expressions and typeset formula detection. In 2019 International Conference on Document Analysis and Recognition (ICDAR), pages 1533-1538. IEEE, 2019. 2, 5, 8", + "[35] Quang Nguyen, Truong Vu, Anh Tran, and Khoi Nguyen. Dataset diffusion: Diffusion-based synthetic data generation for pixel-level semantic segmentation. Advances in Neural Information Processing Systems, 36, 2024. 8", + "[36] Augustus Odena, Christopher Olah, and Jonathon Shlens. Conditional image synthesis with auxiliary classifier gans. In International conference on machine learning, pages 2642-2651. PMLR, 2017. 5", + "[37] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2337-2346, 2019. 1", + "[38] Konpat Preechakul, Nattanat Chathee, Suttisak Wizadwongsa, and Supasorn Suwajanakorn. Diffusion autoencoders: Toward a meaningful and decodable representation. In Proceedings of the IEEE/CVF Conference on Computer" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Vision and Pattern Recognition, pages 10619-10629, 2022. 7", + "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 7", + "[40] Renato Sortino, Simone Palazzo, Francesco Rundo, and Concetto Spampinato. Transformer-based image generation from scene graphs. Computer Vision and Image Understanding, 233:103721, 2023. 3", + "[41] Matthias Springstein, Eric Müller-Budack, and Ralph Ewert. Unsupervised training data generation of handwritten formulas using generative adversarial networks with self-attention. In Proceedings of the 2021 Workshop on MultiModal Pre-Training for Multimedia Understanding, pages 46–54, 2021. 1, 2, 6", + "[42] Sitong Su, Lianli Gao, Junchen Zhu, Jie Shao, and Jingkuan Song. Fully functional image manipulation using scene graphs in a bounding-box free way. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1784-1792, 2021. 3", + "[43] Wei Sun and Tianfu Wu. Image synthesis from reconfigurable layout and style. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10531-10540, 2019. 3", + "[44] Wei Sun and Tianfu Wu. Learning layout and style reconfigurable gans for controllable image synthesis. IEEE transactions on pattern analysis and machine intelligence, 44(9): 5070-5087, 2021. 3", + "[45] Tristan Sylvain, Pengchuan Zhang, Yoshua Bengio, R Devon Hjelm, and Shikhar Sharma. Object-centric image generation from layouts. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2647-2655, 2021. 3", + "[46] Thanh-Nghia Truong, Cuong Tuan Nguyen, and Masaki Nakagawa. Syntactic data generation for handwritten mathematical expression recognition. Pattern Recognition Letters, 153:83-91, 2022. 1, 2", + "[47] Ashish Vaswani, Noam Shazeer, Niki Parmar, Llion Jones, Jakob Uszkoreit, Aidan N Gomez, and Lukasz Kaiser. Attention is all you need. In Advances in Neural Information Processing Systems, pages 5998-6008. 2017. 2", + "[48] Chi Wang, Min Zhou, Tiezheng Ge, Yuning Jiang, Hujun Bao, and Weiwei Xu. Cf-font: Content fusion for few-shot font generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1858-1867, 2023. 3", + "[49] Zhou Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Process, 13(4):600-612, 2004. 6", + "[50] Chuan Wen, Yujie Pan, Jie Chang, Ya Zhang, Siheng Chen, Yanfeng Wang, Mei Han, and Qi Tian. Handwritten chinese font generation with collaborative stroke refinement. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3882-3891, 2021. 3" + ], + "bbox": [ + 501, + 90, + 893, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15684", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Chenxi Whitehouse, Monojit Choudhury, and Alham Fikri Aji. Llm-powered data augmentation for enhanced crosslingual performance. arXiv preprint arXiv:2305.14288, 2023. 1", + "[52] Changjie Wu, Jun Du, Yunqing Li, Jianshu Zhang, Chen Yang, Bo Ren, and Yiqing Hu. Tdv2: A novel tree-structured decoder for offline mathematical expression recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2694–2702, 2022. 2", + "[53] Jin-Wen Wu, Fei Yin, Yan-Ming Zhang, Xu-Yao Zhang, and Cheng-Lin Liu. Handwritten mathematical expression recognition via paired adversarial learning. International Journal of Computer Vision, 128:2386-2401, 2020. 2", + "[54] Jin-Wen Wu, Fei Yin, Yan-Ming Zhang, Xu-Yao Zhang, and Cheng-Lin Liu. Graph-to-graph: towards accurate and interpretable online handwritten mathematical expression recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2925-2933, 2021. 2", + "[55] Yang Wu, Pengxu Wei, and Liang Lin. Scene graph to image synthesis via knowledge consensus. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2856-2865, 2023. 3", + "[56] Ximing Xing, Chuang Wang, Haitao Zhou, Jing Zhang, Qian Yu, and Dong Xu. Diffsketcher: Text guided vector sketch synthesis through latent diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 7", + "[57] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Fine-grained text to image generation with attentional generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1316-1324, 2018. 1", + "[58] Yuan Xue, Yuan-Chen Guo, Han Zhang, Tao Xu, Song-Hai Zhang, and Xiaolei Huang. Deep image synthesis from intuitive user input: A review and perspectives. Computational Visual Media, 8:3-31, 2022. 1", + "[59] Chen Yang, Jun Du, Jianshu Zhang, Changjie Wu, Mingjun Chen, and JiaJia Wu. Tree-based data augmentation and mutual learning for offline handwritten mathematical expression recognition. Pattern Recognition, 132:108910, 2022. 1, 2", + "[60] Ling Yang, Zhilin Huang, Yang Song, Shenda Hong, Guohao Li, Wentao Zhang, Bin Cui, Bernard Ghanem, and Ming-Hsuan Yang. Diffusion-based scene graph to image generation with masked contrastive pre-training. arXiv preprint arXiv:2211.11138, 2022. 1, 3", + "[61] Jun Yu, Xingxin Xu, Fei Gao, Shengjie Shi, Meng Wang, Dacheng Tao, and Qingming Huang. Toward realistic face photo-sketch synthesis via composition-aided gans. IEEE transactions on cybernetics, 51(9):4350-4362, 2020. 1", + "[62] Ye Yuan, Xiao Liu, Wondimu Dikubab, Hui Liu, Zhilong Ji, Zhongqin Wu, and Xiang Bai. Syntax-aware network for handwritten mathematical expression recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4553-4562, 2022. 1", + "[63] Richard Zanibbi and Dorothea Blostein. Recognition and retrieval of mathematical expressions. International Journal on Document Analysis and Recognition (IJDAR), 15:331-357, 2012. 2, 3" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[64] Jinshan Zeng, Qi Chen, Yunxin Liu, Mingwen Wang, and Yuan Yao. Strokegan: Reducing mode collapse in Chinese font generation via stroke encoding. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3270-3277, 2021. 3", + "[65] Yu Zeng, Zhe Lin, Jianming Zhang, Qing Liu, John Collo-mosse, Jason Kuen, and Vishal M Patel. Scenecomposer: Any-level semantic image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22468-22478, 2023. 1", + "[66] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris Metaxas. StackGAN++: Realistic image synthesis with stacked generative adversarial networks. IEEE TPAMI, PP(99), 2017. 1", + "[67] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1, 7", + "[68] Yaping Zhang, Shuai Nie, Wenju Liu, Xing Xu, Dongxiang Zhang, and Heng Tao Shen. Sequence-to-sequence domain adaptation network for robust text image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2", + "[69] Bo Zhao, Weidong Yin, Lili Meng, and Leonid Sigal. Layout2image: Image generation from layout. International Journal of Computer Vision, 128:2418-2435, 2020. 1, 3", + "[70] Wenqi Zhao and Liangcai Gao. Comer: Modeling coverage for transformer-based handwritten mathematical expression recognition. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVIII, pages 392-408. Springer, 2022. 2", + "[71] J. Zhu, T. Park, P. Isola, and A. A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In ICCV, pages 2242-2251, 2017. 6", + "[72] Peihao Zhu, Rameen Abdal, Yipeng Qin, and Peter Wonka. Sean: Image synthesis with semantic region-adaptive normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5104-5113, 2020. 1" + ], + "bbox": [ + 501, + 92, + 893, + 641 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "15685", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Generating Handwritten Mathematical Expressions From Symbol Graphs_ An End-to-End Pipeline/f14645bb-d996-46da-9c4b-81a14daf5714_model.json b/2024/Generating Handwritten Mathematical Expressions From Symbol Graphs_ An End-to-End Pipeline/f14645bb-d996-46da-9c4b-81a14daf5714_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ecde185b9fc5e4caef5f8b5a318b951b7a441842 --- /dev/null +++ b/2024/Generating Handwritten Mathematical Expressions From Symbol Graphs_ An End-to-End Pipeline/f14645bb-d996-46da-9c4b-81a14daf5714_model.json @@ -0,0 +1,2444 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.108, + 0.13, + 0.865, + 0.178 + ], + "angle": 0, + "content": "Generating Handwritten Mathematical Expressions From Symbol Graphs: An End-to-End Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.182, + 0.203, + 0.786, + 0.221 + ], + "angle": 0, + "content": "Yu Chen\\(^{1,*}\\), Fei Gao\\(^{2,*}\\), Yanguang Zhang\\(^{3}\\), Maoying Qiao\\(^{4}\\), Nannan Wang\\(^{5,\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.221, + 0.882, + 0.258 + ], + "angle": 0, + "content": "1 Beijing Waiyan Online Digital Technology; 2 Hangzhou Institute of Technology, Xidian University; 3 Hangzhou Dianzi University; 4 University of Technology, Sydney (UTS); 5 Xidian University" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.26, + 0.815, + 0.275 + ], + "angle": 0, + "content": "{fgao, nwwang}@xidian.edu.cn, ppakl991@outlook.com, maoying.qiao@uts.edu.au" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.309, + 0.314, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.341, + 0.474, + 0.765 + ], + "angle": 0, + "content": "In this paper, we explore a novel challenging generation task, i.e. Handwritten Mathematical Expression Generation (HMEG) from symbolic sequences. Since symbolic sequences are naturally graph-structured data, we formulate HMEG as a graph-to-image (G2I) generation problem. Unlike the generation of natural images, HMEG requires critic layout clarity for synthesizing correct and recognizable formulas, but has no real masks available to supervise the learning process. To alleviate this challenge, we propose a novel end-to-end G2I generation pipeline (i.e. graph \\(\\rightarrow\\) layout \\(\\rightarrow\\) mask \\(\\rightarrow\\) image), which requires no real masks or nondifferentiable alignment between layouts and masks. Technically, to boost the capacity of predicting detailed relations among adjacent symbols, we propose a Less-is-More (LiM) learning strategy. In addition, we design a differentiable layout refinement module, which maps bounding boxes to pixel-level soft masks, so as to further alleviate ambiguous layout areas. Our whole model, including layout prediction, mask refinement, and image generation, can be jointly optimized in an end-to-end manner. Experimental results show that, our model can generate high-quality HME images, and outperforms previous generative methods. Besides, a series of ablations study demonstrate effectiveness of the proposed techniques. Finally, we validate that our generated images promisingly boosts the performance of HME recognition models, through data augmentation. Our code and results are available at: https://github.com/AiArt-HDU/HMEG." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.792, + 0.21, + 0.808 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.817, + 0.47, + 0.879 + ], + "angle": 0, + "content": "Handwritten Mathematical Expressions (HMEs) are common and play significant roles in our daily life, especially in the research and education areas. HMEs generally present complex structures, serious deformations, and diverse writ" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.309, + 0.891, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.399, + 0.891, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.502, + 0.895, + 0.587 + ], + "angle": 0, + "content": "Figure 1. Differences between (a) typical two-stage graph-to-image generation pipeline and (b) our end-to-end pipeline. In previous methods, the real masks are available as the input. In contrast, we propose a novel end-to-end pipeline of \\( \\text{graph} \\to \\text{layout} \\to \\text{mask} \\to \\text{image} \\), and requires no real masks or non-differentiable alignment between layouts and masks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.894, + 0.689 + ], + "angle": 0, + "content": "ing styles. Such characteristics, along with data scarcity, make HME Recognition (HMER) a grand challenge in the OCR community. Despite the tremendous efforts that have been made to this task, the HMER performance is still unsatisfactory [14, 27, 62]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.895, + 0.825 + ], + "angle": 0, + "content": "Recently, synthetic data augmentation has shown inspiring performance in various recognition tasks [23, 51]. This is mainly achieved by the remarkable progress of conditional image generation [21] in the past few years [58]. The advances allow for creating realistic images through texts [39, 57, 66], scene graphs [24, 60], layouts [28, 69], semantic masks [37, 72], sketches [61, 67], and more [12, 65]. However, there has been no effective generative model for generating high-quality HME images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Some previous works synthesize HMEs by recomposing real online HMEs [46, 59]. Besides, FormulaGAN [41] generates HMEs from rendered images, via an Image-to-Image (I2I) Translation model. However, their generated images are limited in either diversity or realism. Be" + }, + { + "type": "page_footnote", + "bbox": [ + 0.101, + 0.887, + 0.359, + 0.901 + ], + "angle": 0, + "content": "* Equal Contributions. † Corresponding Author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15675" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.182 + ], + "angle": 0, + "content": "sides, Mathematical Expressions (MEs) are highly structured data. Given an HME, its meaning and handwritten style, depend not only on the category and shape of each symbol, but also on the positional relations between symbols. It's unlikely to design an elaborate I2I translation model that can handle all these issues." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.185, + 0.47, + 0.397 + ], + "angle": 0, + "content": "To combat this challenge, in this paper, we explore a new generation task, i.e. Handwritten Mathematical Expression Generation (HMEG) from symbol graphs of symbolic sequences. It's natural and easy to use a graph, specifying symbols and positional relations between symbols, to represent an ME [63]. Besides, previous works have demonstrated the significance of fully mining the structural information in boosting generalization [52]. By converting symbolic sequences to symbol graphs, we propose a novel HMEG method, inspired by the classic graph-to-image (G2I) generation work, \\(\\mathrm{Sg2im}\\) [24]. Similar to existing G2I generation works, our method consists of three stages: graph-to-query prediction, layout-to-mask transformation, and mask-to-image generation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.398, + 0.47, + 0.61 + ], + "angle": 0, + "content": "Compared to the generation of natural images, the key challenge of HMEG is to generate unambiguous structural layouts, for accurately presenting meanings of formulas. From one hand, the requirement of layout clarity is much critical for generating formulas. Not only the size of each symbol, but also the positional relations between symbols, should be precisely predicted. Otherwise, mathematical meanings might change or become unrecognisable. For example, an input formula \\( x^{2} \\) might become \\( x_{2} \\) or \\( x2 \\) in a generated image. From the other hand, MEs have infinite possible layouts and structures. Namely, a formula could comprise of arbitrary number of symbols and arbitrary relations between symbols. It's impossible to collect completed training data for learning the layout predictor." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.612, + 0.47, + 0.839 + ], + "angle": 0, + "content": "To alleviate these difficulties, we first propose a Less-is-More (LiM) training strategy for learning an effective layout predictor. Specially, our layout predictor follows the structure of Graph Convolutional Networks (GCNs) [30]. During training, we only use minimal graphs with 1-degree connections, in stead of whole graphs with complex structures, to optimize the layout predictor. Note that the learned predictor can be applied to arbitrary symbol graphs, in the testing stage. By using LiM, the predictor emphasizes on local structures between adjacent symbols. Besides, the learning process is eased. Namely, we could learn an effective layout predictor by using moderate amount of training samples. Such benefits will be verified in the ablation study part. Besides, we add a layout discriminator to further boosts the realism of predicted layouts." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In addition, we propose a Sequential BBox-to-Mask Transformation (B2M) module, for refining layouts under weak supervision. In the original layout, bounding boxes of symbols tend to overlap with each other. Such overlapped" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.288 + ], + "angle": 0, + "content": "areas may lead to overlapping symbols or chaotic strokes, in generated images. It's critical to refine coarse layouts to pixel-level soft masks, to alleviate the overlapping problems [22]. Most previous G2I generation works require real segmentation masks for learning refined layout masks [13, 22]. However, it's difficult to define the mask of a symbol, because white pixels around strokes are essential for representing a symbol. We have to refine the layout under weak supervision. To this end, we propose to learn a sequential mapping of layout \\(\\rightarrow\\) grid \\(\\rightarrow\\) mask, and use an subsequent image decoder to generate an HME image. During training, we use the image-related losses to train both B2M and the image decoder jointly." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.289, + 0.893, + 0.545 + ], + "angle": 0, + "content": "In summary, we propose a novel end-to-end pipeline of graph \\(\\rightarrow\\) layout \\(\\rightarrow\\) mask \\(\\rightarrow\\) image for G2I generation tasks. Our model requires no supervision of real masks, or interactive alignment operations between layouts and masks. The major differences between our method and previous typical G2I generation works, are illustrated in Fig. 1. To verify the effectiveness of the proposed techniques, we conduct experiments on CROHME2014/2016/2019 [34]. Results show that, our model can generate high-quality HME images with clear layouts and recognisable symbols. Besides, our model significantly outperforms a number of existing generative methods, both qualitatively and quantitatively. A series of ablation study demonstrate the effectiveness of the proposed techniques, including the LiM, layout discriminator, and B2M. Finally, our generated HME images prove to boost the HMER performance by \\(5 - 11\\%\\) absolutely, through synthetic data augmentation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.559, + 0.65, + 0.574 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.87 + ], + "angle": 0, + "content": "HME Recognition (HMER). Most advanced methods treat HMER as an image-to-markup problem [53]. They typically use Convolutional Neural Networks (CNNs) or Recurrent Neural Networks (RNNs) [18] to encode input images; and use attention-based models, such as Gated Recurrent Units (GRU) [5, 14, 68], or Transformers [2, 47, 70], to generate LaTeX sequences. Besides, some methods have been proposed [41, 46, 59]. Some recent works propose to use additional information, e.g. symbol-level counting [27], or relation-level counting [14], or emphasis on visually similar symbols [33], or synthetic data augmentation, to boost the HMER performance. Another branch of HMER are graph-based methods. These methods typically transform LaTeX sequences to Symbol Label Trees (SLTs) [63], and use tree-structured decoders to obtain the outputs [14, 52, 54]. These works validate the capacity of graphs in representing the structural information of formulas. In this paper, we therefore convert symbolic sequences to graphs, and use a GCN to encode them for generating HME images." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Graph-to-Image Generation. Image generation from scene graphs is first proposed in Sg2im [24]. The pipeline" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15676" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.895, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.233, + 0.259, + 0.735, + 0.273 + ], + "angle": 0, + "content": "Figure 2. Overview of our handwritten mathematical expression generator (HMEG)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.3, + 0.471, + 0.406 + ], + "angle": 0, + "content": "of Sg2im includes two stages: graph-to layout mapping and layout-to-image generation. Various subsequent methods have been proposed to boost the quality of generated images, following the same pipeline [9, 13, 16, 40]. Besides, some semi-parametric methods use crops from real images to boost the quality of generated images [31]. There are also some extensions to image manipulation [9, 42]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.41, + 0.47, + 0.606 + ], + "angle": 0, + "content": "Most of these methods use a image discriminator and a object discriminator during training. Besides, they use real layouts, segmentation masks, and images, to supervise the learning process [13, 22]. COLoR [22] additionally uses a mask discriminator to boost the precision of generated masks. Recently, KCGM [55] uses knowledge consensus to alleviate the dependency on real layouts or masks. Besides, several diffusion model based methods [12, 60] can generate high-quality images directly from input scene graphs. However, diffusion based methods still suffer the high cost during both training and inference. Besides, generating recognisable texts is still challenging for most diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.61, + 0.47, + 0.867 + ], + "angle": 0, + "content": "Layout-to-Image Generation. Most layout-to-image generation methods separately predict the mask of each object based on object embeddings, and require frequent resizing interaction between bounding boxes and masks [43, 44, 69]. To improve the quality of generated masks, several methods use masks discriminator [1, 28], and requires real masks during training. Besides, OC-GAN [45] uses both global and local similarities between scene graphs and images are used to improve the realism of generated layouts. Similar to our method, LAMA [32] tries to alleviate the overlapping between objects. However, in all these methods, the real bounding boxes are available as the input. They at most implement the end-to-end computing of layout \\(\\rightarrow\\) mask \\(\\rightarrow\\) image. In contrast, we propose a novel end-to-end computing of graph \\(\\rightarrow\\) layout \\(\\rightarrow\\) mask \\(\\rightarrow\\) image, and requires no interactive alignment operations between layouts and masks." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Handwritten Font Generation. Another related task is Handwritten Font Generation (HFG), which aims at" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.3, + 0.895, + 0.436 + ], + "angle": 0, + "content": "transferring images of fonts to the target handwritten style [48, 50]. Most methods follow the I2I translation pipeline, and adopt encoder-decoder architectures [6]. Besides, great efforts have been made to solve the few-shot learning challenge [48, 64]. These methods focus on the style of strokes. In contrast, our work aims to generate rational layouts of multiple symbols, with complex relations, at the same time. It's interesting to explore generating HME images in personal styles in the future." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.453, + 0.591, + 0.468 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.478, + 0.61, + 0.492 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.502, + 0.892, + 0.76 + ], + "angle": 0, + "content": "Our goal is to generate HME images conditioned on LaTeX sequences. To this end, we represent LaTeX sequences by symbol graphs [63] and design a novel G2I generation method to synthesize HME images. Our full model follows the idea of Generative Adversarial Networks (GANs) [21, 24]. It includes a generator, and several discriminators in the training stage. Fig. 2 shows the pipeline of our generator. It mainly include four parts: (i) First, the input LaTeX sequence is converted to a symbol graph, and embedded into high-dimensional feature vectors. (ii) A GCN based layout predictor estimates the coarse layout, i.e. bounding box (BBox) of each symbol. (iii) A sequential layout refinement module, including BBox-to-Grid mapping and Grid-to-Mask projection, generate a pixel-level soft mask for each symbol. (iv) The image decoder, a Cascaded Refinement Network (CRN) [4], generates an image conditioned on the predicted mask." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.771, + 0.755, + 0.787 + ], + "angle": 0, + "content": "3.2. Symbol Graph Construction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Given a set of mathematical symbol categories \\(\\mathcal{C}\\) and a set of relation categories \\(\\mathcal{R}\\), a symbol graph \\((S,E)\\) can be constructed based on \\(\\mathcal{C}\\) and \\(\\mathcal{R}\\): each node \\(i\\) in \\(S = \\{s_1,\\dots ,s_n\\}\\) is a mathematical symbol, with \\(s_i\\in \\mathcal{C}\\); and each directed edge \\((i,j)\\) in \\(E = \\{e_{ij}\\}\\) represents the positional relation of node \\(j\\) w.r.t. node \\(i\\), with \\(e_{ij} = (s_i,r,s_j)\\) and \\(r\\in \\mathcal{R}\\). \\(n\\) is the number of symbols in a mathematical" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15677" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.225, + 0.088, + 0.337, + 0.132 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.19, + 0.47, + 0.22 + ], + "angle": 0, + "content": "Figure 3. Architecture of our layout predictor and the Graph Embedding Model (GEM) [30]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.246, + 0.469, + 0.321 + ], + "angle": 0, + "content": "expression. Let \\( m \\) denote the number of edges in \\( E \\). For example, given an expression \\( x^2 \\), it can be represented by a tuple \\( (x, \\sup, 2) \\), where \\( \\sup \\) is short for superscript. Besides, we add an additional edge type None to indicate that the corresponding pair of symbols are unrelated." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.321, + 0.469, + 0.382 + ], + "angle": 0, + "content": "Afterwards, we map the nodes and edges \\((S,E)\\) to high-dimensional feature vectors \\(\\{\\mathbf{V},\\mathbf{U}\\}\\) by using a learnable embedding layer, respectively [30]. The learned embedding feature vectors are denoted by:" + }, + { + "type": "equation", + "bbox": [ + 0.126, + 0.394, + 0.469, + 0.412 + ], + "angle": 0, + "content": "\\[\n\\mathbf {V} = \\left\\{\\mathbf {v} _ {i}, \\forall i \\in S \\right\\}, \\mathbf {U} = \\left\\{\\mathbf {u} _ {i j}, \\forall (i, j) \\in E \\right\\}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.422, + 0.469, + 0.453 + ], + "angle": 0, + "content": "where \\(\\mathbf{v}_i, \\mathbf{u}_{ij} \\in \\mathbb{R}^d\\) associate with \\(s_i, e_{ij}\\), respectively; \\(d\\) is the dimension of embedding vectors." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.462, + 0.347, + 0.478 + ], + "angle": 0, + "content": "3.3. Adversarial Layout Prediction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.485, + 0.469, + 0.606 + ], + "angle": 0, + "content": "Inspired by GAN [21], we train our layout prediction module in an adversarial learning manner. Specially, we use an layout predictor \\(G_{box}\\) to generate a layout in the format of bounding boxes, conditioned on a given symbol graph; and use an additional layout discriminator \\(D_{box}\\) to judge whether a layout is real or fake. Using the adversarial learning strategy like GAN, the layout predictor would finally predict high-quality bounding boxes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.606, + 0.469, + 0.744 + ], + "angle": 0, + "content": "Layout Predictor. We first use the Graph Embedding Model (GEM) [30] to update features of nodes and edges, and then use a MLP to predict the BBox indices of each symbol. First, original embedding features are encoded by: \\(\\mathbf{h}_i^{(0)} = f_{\\mathrm{node}}(\\mathbf{v}_i), \\forall i \\in S\\) and \\(\\mathbf{e}_{i,j} = f_{\\mathrm{edge}}(\\mathbf{u}_{ij}), \\forall (i,j) \\in E\\). Afterward, node features are updated through a stack of propagation layers, by accumulating information in local neighborhood. Features of node \\(i\\) at the \\(t\\)-th layer are updated by:" + }, + { + "type": "equation", + "bbox": [ + 0.123, + 0.755, + 0.428, + 0.776 + ], + "angle": 0, + "content": "\\[\n\\mathbf {m} _ {j \\rightarrow i} = f _ {\\text {m e s s a g e}} \\left(\\mathbf {h} _ {i} ^ {(t)}, \\mathbf {h} _ {j} ^ {(t)}, \\mathbf {e} _ {i j}\\right), \\forall (i, j) \\in E,\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.777, + 0.469, + 0.811 + ], + "angle": 0, + "content": "\\[\n\\mathbf {h} _ {i} ^ {(t + 1)} = f _ {\\text {n o d e}} \\left(\\mathbf {h} _ {i} ^ {(t)}, \\sum_ {j} \\mathbf {m} _ {j \\rightarrow i}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.822, + 0.469, + 0.854 + ], + "angle": 0, + "content": "After \\(T\\) rounds of propagations, the model produces an aggregated graph level representation by:" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.864, + 0.469, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\mathbf {h} _ {G} = f _ {G} \\left(\\sum_ {i} \\sigma \\left(f _ {\\text {g a t e}} \\left(\\mathbf {h} _ {i} ^ {(T)}\\right)\\right) \\odot f _ {\\text {u p d a t e}} \\left(\\mathbf {h} _ {i} ^ {(T)}\\right)\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.183 + ], + "angle": 0, + "content": "where \\(\\sigma\\) denotes the Sigmoid operation. Finally, we integrate the updated embedding features and the graph level features for predicting the BBox indices of each symbol. Besides, we concatenate a Gaussian noise vector \\(\\mathbf{z}_i\\in \\mathbb{R}^k,\\forall i\\in S\\) to these features, to improve the diversity in predicted layouts. The BBox prediction is formulated by:" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.194, + 0.892, + 0.213 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {b}} _ {i} = f _ {\\mathrm {b b o x}} \\left(\\mathbf {h} _ {i} ^ {(T)}, \\mathbf {h} _ {G}, \\mathbf {z} _ {i}\\right), \\forall i \\in S. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.224, + 0.892, + 0.299 + ], + "angle": 0, + "content": "Each of the mapping functions \\( f_{*} \\) above is an MLP on the inputs [29, 30]. For each symbol, we obtain a 4-dimensional rectangular box \\( [x_0, y_0, x_1, y_1] \\) in the \\( [0,1] \\) coordinate space, denoting the coordinates of the upper left and lower right corners of the BBox border." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.3, + 0.892, + 0.405 + ], + "angle": 0, + "content": "Layout Discriminator. The layout discriminator \\( D_{box} \\) is also composed of GEM [30], followed by an MLP to predict whether an input layout as real or fake. \\( D_{box} \\) takes concatenations of the symbol embedding matrix \\( \\mathbf{V} \\in \\mathbf{R}^{n \\times d} \\) and the layout indices \\( B \\in \\mathbf{R}^{n \\times 4} \\) as input. In the implementation, we use the least square version of adversarial loss, i.e." + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.416, + 0.892, + 0.435 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {a d v} ^ {b o x} = \\left\\| D _ {b o x} (\\mathbf {V}, B) \\right\\| _ {2} ^ {2} + \\left\\| 1 - D _ {b o x} (\\mathbf {V}, \\hat {B}) \\right\\| _ {2} ^ {2}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.445, + 0.892, + 0.521 + ], + "angle": 0, + "content": "Besides, we use the L2 distance between the predicted BBox indices \\(\\hat{B}\\) and the target (real) indices \\(B\\) as the layout reconstruction loss, i.e. \\(\\mathcal{L}_{box} = ||B - \\hat{B}||_1\\). We use a weighted sum of \\(\\mathcal{L}_{box}\\) and \\(\\mathcal{L}_{adv}^{box}\\) as the objective of \\(G_{box}\\). During training, \\(G_{box}\\) and \\(D_{box}\\) are alternatively optimized." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.892, + 0.748 + ], + "angle": 0, + "content": "Less-is-More Learning (LiM) Strategy. A mathematical expression could include arbitrary number of symbols or relations between symbols. In other words, there are infinite possible structures of symbol graphs. It's difficult to collect a complete training set. Besides, using complicated graphs during training may add the difficulty in learning an effective model. We therefore propose a Less-is-More (LiM) learning strategy [3]. Specially, we propose to use symbol graphs with merely 1-degree connections during training. In this way, the layout predictor is expected to emphasize on local structures, i.e. generating rational layout for locally connected symbols. While in the testing stage, the learned model can be applied to arbitrary symbol graphs, with arbitrary degrees of connections, and predict the corresponding layout." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.757, + 0.858, + 0.773 + ], + "angle": 0, + "content": "3.4. Sequential BBox-to-Mask Transformation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.826 + ], + "angle": 0, + "content": "We next predict the pixel-level soft mask of an expression based on the predicted bounding boxes. This procedure is illustrated in Fig. 4, and consists of the following two steps:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "(1) BBox-to-Grid Mapping. We first transform the BBox indices to layout grids, as illustrated in Fig. 4. To this end, we subscribe each index value from the coordinate matrices \\(\\mathbf{X}\\) and \\(\\mathbf{Y}\\) in the [0, 1] coordinate space. For each symbol, we obtain a \\(4 \\times w \\times w\\) tensor of grid." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15678" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.468, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.214, + 0.47, + 0.242 + ], + "angle": 0, + "content": "Figure 4. Illustration of the sequential BBox-to-Mask transformation (B2M) module." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.269, + 0.469, + 0.315 + ], + "angle": 0, + "content": "(2) Grid-to-Mask Projection. Afterwards, we use a FCN with a Sigmoid output layer (denoted by \\(\\phi_{\\mathrm{mask}}\\)) to estimate a one-channel mask for each symbol, i.e." + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.326, + 0.469, + 0.343 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {i} = \\phi_ {\\text {m a s k}} (\\mathbf {X} - x _ {i, 0}, \\mathbf {X} - x _ {i, 1}, \\mathbf {Y} - y _ {i, 0}, \\mathbf {Y} - y _ {i, 1}). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.353, + 0.469, + 0.459 + ], + "angle": 0, + "content": "Each value in \\(\\mathbf{M}_i\\) indicates the probability of a pixel belonging to symbol \\(s_i\\). Specially, we first use two Convolutional layers to predict a low-resolution mask of dimension \\(w \\times w\\), and then use two upsampling Convolutional layers to sequentially upsampling the mask by a factor of 2. These masks are used in the image decoder for generating an image of the same size, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.459, + 0.47, + 0.581 + ], + "angle": 0, + "content": "Note: Previous G2I generation works [9] also consists of BBox-to-Grid mapping, but they merely use the grids to select patches of objects. Besides, they predict the mask from the BBox indices \\(\\hat{B}\\) or original embedding features \\((\\mathbf{V},\\mathbf{U})\\). Differently, we use the transformed grids for layout mask prediction. Moreover, since the whole BBox-to-Mask prediction module is differentiable, our whole model enables training in an end-to-end manner." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.589, + 0.254, + 0.605 + ], + "angle": 0, + "content": "3.5. Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.613, + 0.469, + 0.793 + ], + "angle": 0, + "content": "Image Decoder. In the implementation, we adopt CRN [4] as the image decoder. The input is a multiplication of the whole layout mask \\(\\mathbf{M} \\in \\mathbb{R}^{n \\times w \\times w}\\) with the embedding vectors \\(\\mathbf{V} \\in \\mathbb{R}^{n \\times d}\\). Besides, we add Gaussian noise to the feature maps to improve the diversity in image generation. CRN includes three upsampling Convolutional layers, and outputs a formula image with resolutions of \\(w \\times w\\). Recall that we actually predicts three masks at different scales. In the implementation, we have three generated images, at the resolution of \\(64 \\times 64\\), \\(128 \\times 128\\), and \\(256 \\times 256\\), respectively. All these images are used to calculate the losses for optimizing our model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.794, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Image and Symbol Discriminators Finally, we use an image discriminator \\( D_{img} \\) and a symbol discriminator \\( D_{sym} \\) during training, to boost the realism of the generated whole image \\( \\hat{I} \\) and patches of symbols, respectively. We use the grids to select patches of symbols, and input them into \\( D_{sym} \\). Both discriminators are fully CNNs with the cross-entropy (CE) version of adversarial loss. In addition," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.151 + ], + "angle": 0, + "content": "following AC-GAN [36], we have the symbol discriminator additionally predict category labels of symbols. In this way, the generated symbols would be more recognizable and real." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.152, + 0.892, + 0.198 + ], + "angle": 0, + "content": "Loss Functions. Similar to previous G2I generation works [24], we use a series of loss functions to optimize the image generation. The losses are briefly introduced below." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.201, + 0.892, + 0.247 + ], + "angle": 0, + "content": "- Pixel loss \\(\\mathcal{L}_{pix}\\): First, we use the L1 distance between a generated image and the target formula image and as the pixel loss: \\(\\mathcal{L}_{pix} = ||I - \\hat{I}||_1\\)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.248, + 0.892, + 0.293 + ], + "angle": 0, + "content": "- Image adversarial loss \\(\\mathcal{L}_{adv}^{img}\\): Second, we use the global image discriminant loss from \\(D_{img}\\), to encourage the generated HME images in the real handwritten style." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.293, + 0.892, + 0.338 + ], + "angle": 0, + "content": "- Symbol adversarial loss \\(\\mathcal{L}_{adv}^{sym}\\): Third, we use the symbol discriminative loss from \\(D_{sym}\\), to improve the fidelity of generated symbols." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.339, + 0.892, + 0.384 + ], + "angle": 0, + "content": "- Auxiliary Symbol Recognition Loss \\(\\mathcal{L}_{aux}^{sym}\\): Finally, we use the auxiliary symbol recognition loss from \\(D_{sym}\\), to enforce the generator producing recognizable symbols." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.201, + 0.892, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.388, + 0.892, + 0.432 + ], + "angle": 0, + "content": "We use a combination of these loss functions above, to train the B2M module and the image generation modules together. The total loss is formulated by:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.445, + 0.892, + 0.463 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {i m g} = \\lambda_ {1} \\mathcal {L} _ {p i x} + \\lambda_ {2} \\mathcal {L} _ {a d v} ^ {i m g} + \\lambda_ {3} \\mathcal {L} _ {a d v} ^ {s y m} + \\lambda_ {4} \\mathcal {L} _ {a u x} ^ {s y m}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.474, + 0.892, + 0.505 + ], + "angle": 0, + "content": "where \\(\\lambda_{i}, i = 1, \\ldots, 4\\) are weighting factors and are set as 1, 0.01, 0.01, 0.001, respectively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.52, + 0.634, + 0.537 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.545, + 0.707, + 0.561 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.568, + 0.892, + 0.719 + ], + "angle": 0, + "content": "Data. We use CROHME2014/2016/2019 [34] to validate our HMEG method in terms of both generalization and recognition. The dataset collects pen trajectories, and labels them with symbols and symbol categories. The dataset contains 126 different symbol categories, and 8 positional relations, i.e. start, left superscript, superscript, subscript, below, above, right and end. By restoring the pen trajectory information, we can easily obtain the HME image and the BBox of each symbol. We use the standard partition of each dataset through all the experiments." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. In the implementation, due to the limitation of our computing resources, we train our model in two stages: we first train the layout prediction part in the first stage, and then train the rest parts by fixing it. In the first stage, we train the model with a batch size 64 and a learning rate 5e5, for 40,000 iterations. In each iteration, we random sample a batch of 1-degree sub-graphs from the training set for learning the layout predictor. In the second stage, we use a batch size 8, a learning rate 1e4, and train for 600,000 iterations. Our codes are implemented by using Pytorch and a NVIDIA TITAN XP GPU. We use Adam [26] as the optimizer for all networks." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15679" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.09, + 0.891, + 0.405 + ], + "angle": 0, + "content": "
Print|ax0+by0+c/√a2+b2|p=√a2+b2-2abcosA1/2(a+1/b) = 2ab/a+b-b+√b2-4ac/2a∫0πcos(θ/2)dθx/a+x/b-x/c∫bdIπr2h/3sin(π/3)=1/2
CycleGANaxn+byn+c/√n2+b2|p=√n2+b2-2abcosA1/2(a+1/b) = 2ab/a+b-b+√b2-4ac/2a∫0πcos(θ/2)dθx/a+x/b-x/c∫bdIπr2h/3sin(π/3)=1/2
FormulaGAN|ax0+by0+c/√a2+b2|p=√a2+b2-2abcosA1/2(a+1/b) = 2ab/a+b-b+√b2-4ac/2a∫0πcos(θ/2)dθx/a+x/b-x/c∫bdIπr2h/3sin(π/3)=1/2
Sg2imaxn+byn+c/√a2+b2|p=√a2+b2-2abcosA1/2(a+1/b) = 2ab/a+b-b+√b2-4ac/2a∫0πcos(θ/2)dθx/a+x/b-x/c∫bdIπr2h/3sin(π/3)=1/2
Ours|ax0+by0+c/√a2+b2|p=√a2+b2-2abcosA1/2(a+1/b) = 2ab/a+b-b+√b2-4ac/2a∫0πcos(θ/2)dθx/a+x/b-x/c∫bdIπr2h/3sin(π/3)=1/2
" + }, + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.417, + 0.866, + 0.433 + ], + "angle": 0, + "content": "Figure 5. Handwritten mathematical expressions generated by CycleGAN [71], FormulaGAN [41], Sg2im [24], and our method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.458, + 0.473, + 0.655 + ], + "angle": 0, + "content": "Criteria. We choose two image quality assessment indices as the criteria. The former includes the Fréchet Inception distance (FID) [17] and Structural Similarity Measure (SSIM) [49]. Lower values of FID indicate better realism of synthesised images; while greater values of SSIM generally indicate higher similarity between a synthesized image and the corresponding real image. We here report the average SSIM value of all the test samples. In addition, we use two HMER indices, including the Word Error Rate (WER) and the Expression Recognition Rate (ExpRate). Here we use an open-source HMER method [19], to recognize HME images and compute the corresponding indices. Lower WER and higher ExpRate values indicate better performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.665, + 0.307, + 0.682 + ], + "angle": 0, + "content": "4.2. Comparison with SOTAs" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.47, + 0.765 + ], + "angle": 0, + "content": "We first compare with the existing GAN based HMEG method, FormulaGAN [41], our baseline, Sg2im [24], and the widely used unsupervised I2I translation method, CycleGAN [71], on CHROME2019. We train and test these methods following the same experimental settings as ours." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Qualitative Comparison. Fig. 5 shows HME images generated by different models. CycleGAN fails to transform input formulas to the handwritten style. FormulaGAN produces handwritten-style images, but with blurring and unnatural strokes. In contrast, both Sg2im and our method generate high quality images, with realistic styles and clear strokes. In addition, the images generated by our method present better layouts than those by Sg2im. Sg2im sometimes produce very small symbols, e.g. \\( y_{0} \\) in the first col" + }, + { + "type": "table_caption", + "bbox": [ + 0.506, + 0.456, + 0.885, + 0.47 + ], + "angle": 0, + "content": "Table 1. Comparison with existing methods on CHROME2019." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.481, + 0.9, + 0.568 + ], + "angle": 0, + "content": "
SSIM↑FID ↓WER ↓ExpRate ↑
CycleGAN [71]0.75784.140.6710.026
FormulaGAN [41]0.72474.680.6010.066
Sg2im [24]0.78710.020.3930.219
Ours0.79310.980.3260.316
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.892, + 0.749 + ], + "angle": 0, + "content": "unn and \\( p \\) in the second column. Besides, Sg2im may produce overlapped symbols, e.g. \\( \\frac{1}{2} \\left( \\frac{1}{a} + \\frac{1}{b} \\right) \\) in the third column. Finally, Sg2im produces more unrecognisable symbols than ours, e.g. Sin and \\( \\pi \\) in last column. In contrast, the symbols generated by our method present reasonable sizes and clear structures. Besides, symbols are reasonably placed in relation to each other, neither overlapping nor far apart. This is similar to real formulas written by humans. Such observations demonstrate the superiority of our method in generating both high-quality layouts and HME images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Quantitative Comparison. Table 1 shows the quantitative performance indices. Both Sg2im and our method achieves significant better indices than CycleGAN or FormulaGAN. This is consistent with the qualitative comparison results. According to SSIM and FID, Sg2im and our method almost perform equality in term of image quality. However, our method decreases the WER by 6.5 percent and improves the ExpRate by 10 percent, absolutely. Such distinct superiority in recognition implies that, the handwritten expressions generated by our method are signifi" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15680" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.088, + 0.895, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.287, + 0.893, + 0.308 + ], + "angle": 0, + "content": "Figure 6. Illustration of the layouts and formula images generated by model variants, in the ablation study. The input formula is \\(\\frac{1}{25} y^2 -\\frac{8}{25} y\\)" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.325, + 0.47, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.446, + 0.471, + 0.53 + ], + "angle": 0, + "content": "Figure 7. Comparison with diffusion models. DiffSketcher and DiffAE generate HMEs conditioned on print formulas; while ControlNet and SD+LoRA take both the print formula and a textual prompt (\"a handwritten mathematical expression of latex code\") as input. We use official diffusion models, and fine-tune DiffAE and SD+LoRA using HMEs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.559, + 0.46, + 0.574 + ], + "angle": 0, + "content": "cantly better than \\(\\mathrm{Sg2im}\\), in terms of clarity and structure." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.575, + 0.47, + 0.727 + ], + "angle": 0, + "content": "Comparison with Diffusion Models. We additionally compare with ControlNet [67], Stable Diffusion (SD) [39] + LoRA [20], DiffSketcher [56], and DiffAE [38]. As shown in Fig. 7, these diffusion models cannot generate high-quality HMEs with realistic handwritten styles or recognizable symbols. Besides, their computational complexity is much heavier than ours. Some recent methods, e.g. ChiroDiff [8] and SDT [7], can only generate handwritten images of one single character. Such comparison results demonstrate the advantage of our method in HMEG." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.739, + 0.231, + 0.755 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.763, + 0.47, + 0.809 + ], + "angle": 0, + "content": "We further analyse the impact of (i) the LiM training strategy, (ii) the layout discriminator \\((D_{box})\\), and (iii) the B2M module. To this end, we build model variants by:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.811, + 0.468, + 0.841 + ], + "angle": 0, + "content": "- Using original training graphs, instead of sampled minimal graphs, to train the layout predictor \\((w / o LiM)\\);" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.841, + 0.468, + 0.87 + ], + "angle": 0, + "content": "- Removing B2M from our full model, but using grids for embedding selection (w/o B2M);" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.871, + 0.468, + 0.9 + ], + "angle": 0, + "content": "- Removing both LiM and B2M from our full model (\\(w / o\\) LiM/B2M);" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.811, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.327, + 0.892, + 0.355 + ], + "angle": 0, + "content": "Table 2. Performance indices w.r.t. the ablation study on CHROME19. The model variant in the last row is our baseline." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.367, + 0.9, + 0.468 + ], + "angle": 0, + "content": "
SSIM↑FID↓WER↓ExpRate↑mIOU↑
Ours (full)0.79310.980.3260.3160.364
w/o LiM0.79011.550.3270.3150.335
w/o B2M0.79011.460.3320.3080.354
w/o LiM/B2M0.78613.300.3650.2790.338
Sg2im (base)0.78710.080.3930.2190.324
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.492, + 0.891, + 0.522 + ], + "angle": 0, + "content": "- Training the whole model following Sg2im, without using LiM, \\(D_{box}\\), and B2M (Sg2im (base))." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.522, + 0.891, + 0.552 + ], + "angle": 0, + "content": "We additionally report the mean Intersection over Union (mIoU) between the predicted layouts and the target ones." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.825 + ], + "angle": 0, + "content": "As shown in Table 2, removing LiM or B2M slightly changes the indices about image quality, decreases the IoU. Specially, removing LiM decreases the mIoU by about 3 percent absolutely. This implies the significance of LiM in generating good layouts. In addition, if we removing \\( D_{box} \\) or B2M, along with LiM, the performance seriously decreases, almost in terms of all performance indices. This implies that our \\( D_{box} \\) and B2M are significant for boosting the quality of generated images from all aspects, including the fidelity (SSIM), the usability for recognition (WER & ExpRate), and the layout (mIoU). Fig. 6 illustrates the layouts and images generated by different model variants. All the model variants, except the full model, produce overlapped symbols or symbols in unnatural sizes. The image generated by our model presents no obvious defects and are easy to recognize. Such observations further demonstrate the significance of LiM, B2M, and layout discriminator in layout prediction and image generation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.833, + 0.634, + 0.849 + ], + "angle": 0, + "content": "4.4. Applications" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Mathematical Expression Manipulation. In addition, we apply the previous learned model to expression manipulation using symbol graphs. Given a source symbol graph," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15681" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.09, + 0.466, + 0.273 + ], + "angle": 0, + "content": "
Source1/3πr2hx/a + x/b - x/cax + bx + c/2a + √b + c/2∫_0^1 x^2y dxdyT_1^2/T_2^2 = a_1^3/a_2^3
Add1/3πr^2hzx + d/a + x/b - x/cax + b^x / 2a + √b + c/2 + 7∫_0^1 x^2y^2/2 dxdyT_1^2/T_2^2 = a_1^3 + 2/a_2^3
Delete1/3πr^2x/a + x/bax + b^x + ca + √b / 2∫_0^1 x^2y^2 dxdyT_1^2/T_2^2 = a/a
Change1/7πr^2hx/a + x/e - x/cax + b^y + c/2a + √b + c/4∫_0^1 x^2y dxdyT_1^2/T_2^2 = e^3/a^3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.285, + 0.47, + 0.315 + ], + "angle": 0, + "content": "Figure 8. Illustration of expression manipulation using symbol graphs (LaTeX sequences)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.341, + 0.469, + 0.432 + ], + "angle": 0, + "content": "we edit it by adding, deleting, or changing symbol nodes in it, and apply the previously learned generator to produce the corresponding HME image. As shown in Fig. 8, the generated image changes consistently with the editing operation. Besides, the manipulated images present natural layouts, recognizable symbols, and clear strokes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.432, + 0.469, + 0.644 + ], + "angle": 0, + "content": "Data Augmentation for HMER. Finally, we validate the possible application of HMEG in augmenting HMER models, since the generated data could improve the amount and diversity of training data. To this end, we conduct experiments on CROHME2014/2016/2019 [34], and use CAN [27] as the HMER model. Here, we train CAN with (1) the standard training data (denoted by CAN); or (2) with previous data augmentation, e.g. rotation, (denoted by \\(\\mathrm{CAN^{*}}\\)); or (3) with our synthetic data augmentation, using 6,000 additional generated samples (denoted by \\(\\mathrm{CAN^{\\dagger}}\\)); or (4) both previous and our data augmentation (denoted by \\(\\mathrm{CAN^{*\\dagger}}\\)). The model is trained for 120 epochs under each setting, using the official training parameters. Afterward, we apply the learned HMER model to both generated and real data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.469, + 0.825 + ], + "angle": 0, + "content": "As shown in Table 3, our method achieves \\(5 - 11\\%\\) absolute improvement over the baseline, and performs comparably to previous data augmentation, across all the datasets. Besides, our method can be jointly used with previous augmentation to further boost the performance (i.e. \\(\\mathrm{CAN}^{*\\dagger}\\)). We additionally report the HMER performance, as well as the SSIM and StruRate, of our generated samples. The SSIM and StruRate values are high, and the ExpRate values (by \\(\\mathrm{CAN}^{*} / \\mathrm{CAN}^{\\dagger}\\)) approach those of real data, on all datasets. These results demonstrate the robustness of our method, in generating high-quality HMEs, with correct layouts and recognizable symbols." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Note: Previous HME synthesis methods [10, 25] synthesize HMEs by shuffling the symbols and layouts contained in original HME datasets. Neither of them is a generative method. And they may cause overfitting to existing datasets. In contrast, our pipeline allows excellent flexibil" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.132 + ], + "angle": 0, + "content": "Table 3. Results on CROHME14/16/19 testing sets and our generated images. *,†, *† denote using previous data augmentation, our synthetic data augmentation, and both, respectively." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.143, + 0.895, + 0.22 + ], + "angle": 0, + "content": "
HMEG (generated)HMER (real)
SSIMCAN*CAN†StruRateCANCAN*CAN†CAN*†
CROHME140.78952.154.198.144.752.950.255.4
CROHME160.79851.455.896.542.852.453.957.6
CROHME190.79355.456.497.739.348.449.658.5
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.25, + 0.892, + 0.294 + ], + "angle": 0, + "content": "ity in generating diverse new samples. The experimental results also demonstrate the advantages of our approach in both HMEG and HMER." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.314, + 0.627, + 0.329 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.34, + 0.892, + 0.505 + ], + "angle": 0, + "content": "In this paper, we propose a novel method to solve the challenging task of generating HMEs from LaTeX sequences. We formulate it as a graph-to-image generation task, and focus on boosting the layout prediction without mask supervision. Experimental results demonstrate the effectiveness of our method, and its possible applications in expression manipulation and the challenging inverse task, i.e. HMER. Inspired by the remarkable process of generative augmentation in visual understanding [35], we believe that: the novel HMEG task we explore here has great potential to significantly boost HMER." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.508, + 0.893, + 0.705 + ], + "angle": 0, + "content": "Due to computational limitations, we merely use naive networks in the implementation. The generated symbols are occasionally unrecognizable, especially in complex formulas. It's promising to boost the generation quality by using advanced networks (e.g. VQ-GAN [11] or diffusion models [15]). Besides, we will extend the proposed techniques to natural/artistic image generation by exploring the inspiring capacity of diffusion models, and other tasks of structured graph data. Finally, to further boost the HMER performance, it's promising to use the triples of \\(\\{LaTeX, predicted layout, generated image\\}\\) in HMEG as pseudo labeled data, or to use the layout discriminator for enhancing the layout detection module in an HMER pipeline." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.723, + 0.667, + 0.74 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We greatly appreciate Biao Ma's and Chang Jiang's help in the implementation of HMER. This work was supported in part by the National Natural Science Foundation of China under Grants U22A2096 and 61971172; in part by the Technology Innovation Leading Program of Shaanxi under Grant 2022QFY01-15; in part by the Fundamental Research Funds for the Central Universities under Grant QTZX23042 and ZYTS24012; and in part by the Proof of Concept Foundation of Xidian University Hangzhou Institute of Technology under Grant No. GNYZ2023YL0301." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15682" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Oron Ashual and Lior Wolf. Specifying object attributes and relations in interactive scene generation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4561-4569, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.243 + ], + "angle": 0, + "content": "[2] Xiaohang Bian, Bo Qin, Xiaozhe Xin, Jianwu Li, Xuefeng Su, and Yanfeng Wang. Handwritten mathematical expression recognition via attention aggregation based bidirectional mutual learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 113-121, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.471, + 0.271 + ], + "angle": 0, + "content": "[3] Stefan Bornholdt. Less is more in modeling large genetic networks. Science, 310(5747):449-451, 2005. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.274, + 0.472, + 0.328 + ], + "angle": 0, + "content": "[4] Qifeng Chen and Vladlen Koltun. Photographic image synthesis with cascaded refinement networks. In Proceedings of the IEEE international conference on computer vision, pages 1511-1520, 2017. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.331, + 0.47, + 0.386 + ], + "angle": 0, + "content": "[5] Kyunghyun Cho, Bart Van Merrienboer, Dzmitry Bahdanau, and Yoshua Bengio. On the properties of neural machine translation: Encoder-decoder approaches. arXiv preprint arXiv:1409.1259, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.388, + 0.47, + 0.457 + ], + "angle": 0, + "content": "[6] Gang Dai, Yifan Zhang, Qingfeng Wang, Qing Du, Zhuliang Yu, Zhuoman Liu, and Shuangping Huang. Disentangling writer and character styles for handwriting generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5977-5986, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.46, + 0.469, + 0.529 + ], + "angle": 0, + "content": "[7] Gang Dai, Yifan Zhang, Qingfeng Wang, Qing Du, Zhuliang Yu, Zhuoman Liu, and Shuangping Huang. Disentangling writer and character styles for handwriting generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5977-5986, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.532, + 0.47, + 0.587 + ], + "angle": 0, + "content": "[8] Ayan Das, Yongxin Yang, Timothy M Hospedales, Tao Xiang, and Yi-Zhe Song. Chirodiff: Modelling chirographic data with diffusion models. In The Eleventh International Conference on Learning Representations, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.589, + 0.47, + 0.658 + ], + "angle": 0, + "content": "[9] Helisa Dhamo, Azade Farshad, Iro Laina, Nassir Navab, Gregory D Hager, Federico Tombari, and Christian Rupprecht. Semantic image manipulation using scene graphs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5213-5222, 2020. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.469, + 0.728 + ], + "angle": 0, + "content": "[10] Lan-Fang Dong, Han-Chao Liu, and Xin-Ming Zhang. Synthetic data generation and shuffled multi-round training based offline handwritten mathematical expression recognition. Journal of Computer Science and Technology, 37(6): 1427-1443, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.731, + 0.469, + 0.786 + ], + "angle": 0, + "content": "[11] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[12] Wan-Cyuan Fan, Yen-Chun Chen, DongDong Chen, Yu Cheng, Lu Yuan, and Yu-Chiang Frank Wang. Frido: Feature pyramid diffusion for complex scene image synthesis. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 579-587, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[13] Azade Farshad, Sabrina Musatian, Helisa Dhamo, and Nassir Navab. Migs: Meta image generation from scene graphs. arXiv preprint arXiv:2110.11918, 2021. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.161 + ], + "angle": 0, + "content": "[14] Yingnan Fu, Wenyuan Cai, Ming Gao, and Aoying Zhou. Symbol location-aware network for improving handwritten mathematical expression recognition. In Proceedings of the 2023 ACM International Conference on Multimedia Retrieval, pages 516-524, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.893, + 0.232 + ], + "angle": 0, + "content": "[15] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10696-10706, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.894, + 0.315 + ], + "angle": 0, + "content": "[16] Roei Herzig, Amir Bar, Huijuan Xu, Gal Chechik, Trevor Darrell, and Amir Globerson. Learning canonical representations for scene graph to image generation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVI 16, pages 210-227. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.316, + 0.893, + 0.37 + ], + "angle": 0, + "content": "[17] M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In NIPS, pages 6626-6637. 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.893, + 0.399 + ], + "angle": 0, + "content": "[18] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.893, + 0.455 + ], + "angle": 0, + "content": "[19] Hongyu Wang and Jianshu Zhang. Handwritten mathematical expression recognition (pytorch), 2019. https://github.com/whywhs/Pytorch-Handwritten-Mathematical-Expression-Recognition.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.456, + 0.893, + 0.511 + ], + "angle": 0, + "content": "[20] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.512, + 0.893, + 0.58 + ], + "angle": 0, + "content": "[21] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.893, + 0.649 + ], + "angle": 0, + "content": "[22] Maor Ivgi, Yaniv Benny, Avichai Ben-David, Jonathan Berant, and Lior Wolf. Scene graph to image generation with contextualized object layout refinement. In 2021 IEEE International Conference on Image Processing (ICIP), pages 2428-2432. IEEE, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.65, + 0.893, + 0.734 + ], + "angle": 0, + "content": "[23] Nikita Jaipuria, Xianling Zhang, Rohan Bhasin, Mayar Arafa, Punarjay Chakravarty, Shubham Shrivastava, Sagar Manglani, and Vidya N Murali. Deflating dataset bias using synthetic data augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 772-773, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.893, + 0.788 + ], + "angle": 0, + "content": "[24] Justin Johnson, Agrim Gupta, and Li Fei-Fei. Image generation from scene graphs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1219-1228, 2018. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.893, + 0.871 + ], + "angle": 0, + "content": "[25] Vu Tran Minh Khuong, Ung Quang Huy, Nakagawa Masaki, and Minh Khanh Phan. Generating synthetic handwritten mathematical expressions from a latex sequence or a mathml script. In 2019 International Conference on Document Analysis and Recognition (ICDAR), pages 922-927. IEEE, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[26] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In 3rd International Conference on" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15683" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.192 + ], + "angle": 0, + "content": "[27] Bohan Li, Ye Yuan, Dingkang Liang, Xiao Liu, Zhilong Ji, Jinfeng Bai, Wenyu Liu, and Xiang Bai. When counting meets hmer: counting-aware network for handwritten mathematical expression recognition. In European Conference on Computer Vision, pages 197-214. Springer, 2022. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.471, + 0.263 + ], + "angle": 0, + "content": "[28] Wenbo Li, Pengchuan Zhang, Lei Zhang, Qiuyuan Huang, Xiaodong He, Siwei Lyu, and Jianfeng Gao. Object-driven text-to-image synthesis via adversarial training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12174-12182, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.471, + 0.305 + ], + "angle": 0, + "content": "[29] Yujia Li, Daniel Tarlow, Marc Brockschmidt, and Richard Zemel. Gated graph sequence neural networks. arXiv preprint arXiv:1511.05493, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.307, + 0.471, + 0.375 + ], + "angle": 0, + "content": "[30] Yujia Li, Chenjie Gu, Thomas Dullien, Oriol Vinyals, and Pushmeet Kohli. Graph matching networks for learning the similarity of graph structured objects. In International conference on machine learning, pages 3835-3845. PMLR, 2019. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.471, + 0.434 + ], + "angle": 0, + "content": "[31] Yikang Li, Tao Ma, Yeqi Bai, Nan Duan, Sining Wei, and Xiaogang Wang. Pastegan: A semi-parametric method to generate image from scene graph. Advances in Neural Information Processing Systems, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.435, + 0.471, + 0.503 + ], + "angle": 0, + "content": "[32] Zejian Li, Jingyu Wu, Immanuel Koh, Yongchuan Tang, and Lingyun Sun. Image synthesis from layout with locality-aware mask adaption. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13819-13828, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.471, + 0.56 + ], + "angle": 0, + "content": "[33] Zhe Li, Xinyu Wang, Yuliang Liu, Lianwen Jin, Yichao Huang, and Kai Ding. Improving handwritten mathematical expression recognition via similar symbol distinguishing. IEEE Transactions on Multimedia, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.562, + 0.471, + 0.658 + ], + "angle": 0, + "content": "[34] Mahshad Mahdavi, Richard Zanibbi, Harold Mouchere, Christian Viard-Gaudin, and Utpal Garain. Icdar 2019 crohme+ tfd: Competition on recognition of handwritten mathematical expressions and typeset formula detection. In 2019 International Conference on Document Analysis and Recognition (ICDAR), pages 1533-1538. IEEE, 2019. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.471, + 0.716 + ], + "angle": 0, + "content": "[35] Quang Nguyen, Truong Vu, Anh Tran, and Khoi Nguyen. Dataset diffusion: Diffusion-based synthetic data generation for pixel-level semantic segmentation. Advances in Neural Information Processing Systems, 36, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.471, + 0.773 + ], + "angle": 0, + "content": "[36] Augustus Odena, Christopher Olah, and Jonathon Shlens. Conditional image synthesis with auxiliary classifier gans. In International conference on machine learning, pages 2642-2651. PMLR, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.471, + 0.843 + ], + "angle": 0, + "content": "[37] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2337-2346, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[38] Konpat Preechakul, Nattanat Chathee, Suttisak Wizadwongsa, and Supasorn Suwajanakorn. Diffusion autoencoders: Toward a meaningful and decodable representation. In Proceedings of the IEEE/CVF Conference on Computer" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.092, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Vision and Pattern Recognition, pages 10619-10629, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.192 + ], + "angle": 0, + "content": "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.194, + 0.892, + 0.25 + ], + "angle": 0, + "content": "[40] Renato Sortino, Simone Palazzo, Francesco Rundo, and Concetto Spampinato. Transformer-based image generation from scene graphs. Computer Vision and Image Understanding, 233:103721, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.253, + 0.892, + 0.336 + ], + "angle": 0, + "content": "[41] Matthias Springstein, Eric Müller-Budack, and Ralph Ewert. Unsupervised training data generation of handwritten formulas using generative adversarial networks with self-attention. In Proceedings of the 2021 Workshop on MultiModal Pre-Training for Multimedia Understanding, pages 46–54, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.339, + 0.892, + 0.407 + ], + "angle": 0, + "content": "[42] Sitong Su, Lianli Gao, Junchen Zhu, Jie Shao, and Jingkuan Song. Fully functional image manipulation using scene graphs in a bounding-box free way. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1784-1792, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.41, + 0.892, + 0.466 + ], + "angle": 0, + "content": "[43] Wei Sun and Tianfu Wu. Image synthesis from reconfigurable layout and style. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10531-10540, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.469, + 0.892, + 0.524 + ], + "angle": 0, + "content": "[44] Wei Sun and Tianfu Wu. Learning layout and style reconfigurable gans for controllable image synthesis. IEEE transactions on pattern analysis and machine intelligence, 44(9): 5070-5087, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.527, + 0.892, + 0.582 + ], + "angle": 0, + "content": "[45] Tristan Sylvain, Pengchuan Zhang, Yoshua Bengio, R Devon Hjelm, and Shikhar Sharma. Object-centric image generation from layouts. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2647-2655, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.584, + 0.892, + 0.641 + ], + "angle": 0, + "content": "[46] Thanh-Nghia Truong, Cuong Tuan Nguyen, and Masaki Nakagawa. Syntactic data generation for handwritten mathematical expression recognition. Pattern Recognition Letters, 153:83-91, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.643, + 0.892, + 0.699 + ], + "angle": 0, + "content": "[47] Ashish Vaswani, Noam Shazeer, Niki Parmar, Llion Jones, Jakob Uszkoreit, Aidan N Gomez, and Lukasz Kaiser. Attention is all you need. In Advances in Neural Information Processing Systems, pages 5998-6008. 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.701, + 0.892, + 0.77 + ], + "angle": 0, + "content": "[48] Chi Wang, Min Zhou, Tiezheng Ge, Yuning Jiang, Hujun Bao, and Weiwei Xu. Cf-font: Content fusion for few-shot font generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1858-1867, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.772, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[49] Zhou Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Process, 13(4):600-612, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.831, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[50] Chuan Wen, Yujie Pan, Jie Chang, Ya Zhang, Siheng Chen, Yanfeng Wang, Mei Han, and Qi Tian. Handwritten chinese font generation with collaborative stroke refinement. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3882-3891, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15684" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.146 + ], + "angle": 0, + "content": "[51] Chenxi Whitehouse, Monojit Choudhury, and Alham Fikri Aji. Llm-powered data augmentation for enhanced crosslingual performance. arXiv preprint arXiv:2305.14288, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.148, + 0.47, + 0.218 + ], + "angle": 0, + "content": "[52] Changjie Wu, Jun Du, Yunqing Li, Jianshu Zhang, Chen Yang, Bo Ren, and Yiqing Hu. Tdv2: A novel tree-structured decoder for offline mathematical expression recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2694–2702, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.471, + 0.273 + ], + "angle": 0, + "content": "[53] Jin-Wen Wu, Fei Yin, Yan-Ming Zhang, Xu-Yao Zhang, and Cheng-Lin Liu. Handwritten mathematical expression recognition via paired adversarial learning. International Journal of Computer Vision, 128:2386-2401, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.471, + 0.343 + ], + "angle": 0, + "content": "[54] Jin-Wen Wu, Fei Yin, Yan-Ming Zhang, Xu-Yao Zhang, and Cheng-Lin Liu. Graph-to-graph: towards accurate and interpretable online handwritten mathematical expression recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2925-2933, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.472, + 0.398 + ], + "angle": 0, + "content": "[55] Yang Wu, Pengxu Wei, and Liang Lin. Scene graph to image synthesis via knowledge consensus. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2856-2865, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.47, + 0.455 + ], + "angle": 0, + "content": "[56] Ximing Xing, Chuang Wang, Haitao Zhou, Jing Zhang, Qian Yu, and Dong Xu. Diffsketcher: Text guided vector sketch synthesis through latent diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.456, + 0.47, + 0.537 + ], + "angle": 0, + "content": "[57] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Fine-grained text to image generation with attentional generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1316-1324, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.47, + 0.593 + ], + "angle": 0, + "content": "[58] Yuan Xue, Yuan-Chen Guo, Han Zhang, Tao Xu, Song-Hai Zhang, and Xiaolei Huang. Deep image synthesis from intuitive user input: A review and perspectives. Computational Visual Media, 8:3-31, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.47, + 0.65 + ], + "angle": 0, + "content": "[59] Chen Yang, Jun Du, Jianshu Zhang, Changjie Wu, Mingjun Chen, and JiaJia Wu. Tree-based data augmentation and mutual learning for offline handwritten mathematical expression recognition. Pattern Recognition, 132:108910, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.47, + 0.718 + ], + "angle": 0, + "content": "[60] Ling Yang, Zhilin Huang, Yang Song, Shenda Hong, Guohao Li, Wentao Zhang, Bin Cui, Bernard Ghanem, and Ming-Hsuan Yang. Diffusion-based scene graph to image generation with masked contrastive pre-training. arXiv preprint arXiv:2211.11138, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.47, + 0.774 + ], + "angle": 0, + "content": "[61] Jun Yu, Xingxin Xu, Fei Gao, Shengjie Shi, Meng Wang, Dacheng Tao, and Qingming Huang. Toward realistic face photo-sketch synthesis via composition-aided gans. IEEE transactions on cybernetics, 51(9):4350-4362, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.47, + 0.845 + ], + "angle": 0, + "content": "[62] Ye Yuan, Xiao Liu, Wondimu Dikubab, Hui Liu, Zhilong Ji, Zhongqin Wu, and Xiang Bai. Syntax-aware network for handwritten mathematical expression recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4553-4562, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[63] Richard Zanibbi and Dorothea Blostein. Recognition and retrieval of mathematical expressions. International Journal on Document Analysis and Recognition (IJDAR), 15:331-357, 2012. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.161 + ], + "angle": 0, + "content": "[64] Jinshan Zeng, Qi Chen, Yunxin Liu, Mingwen Wang, and Yuan Yao. Strokegan: Reducing mode collapse in Chinese font generation via stroke encoding. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3270-3277, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.894, + 0.233 + ], + "angle": 0, + "content": "[65] Yu Zeng, Zhe Lin, Jianming Zhang, Qing Liu, John Collo-mosse, Jason Kuen, and Vishal M Patel. Scenecomposer: Any-level semantic image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22468-22478, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.234, + 0.893, + 0.289 + ], + "angle": 0, + "content": "[66] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris Metaxas. StackGAN++: Realistic image synthesis with stacked generative adversarial networks. IEEE TPAMI, PP(99), 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.29, + 0.893, + 0.346 + ], + "angle": 0, + "content": "[67] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.346, + 0.893, + 0.416 + ], + "angle": 0, + "content": "[68] Yaping Zhang, Shuai Nie, Wenju Liu, Xing Xu, Dongxiang Zhang, and Heng Tao Shen. Sequence-to-sequence domain adaptation network for robust text image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.418, + 0.893, + 0.459 + ], + "angle": 0, + "content": "[69] Bo Zhao, Weidong Yin, Lili Meng, and Leonid Sigal. Layout2image: Image generation from layout. International Journal of Computer Vision, 128:2418-2435, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.46, + 0.893, + 0.53 + ], + "angle": 0, + "content": "[70] Wenqi Zhao and Liangcai Gao. Comer: Modeling coverage for transformer-based handwritten mathematical expression recognition. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVIII, pages 392-408. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.531, + 0.893, + 0.572 + ], + "angle": 0, + "content": "[71] J. Zhu, T. Park, P. Isola, and A. A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In ICCV, pages 2242-2251, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.573, + 0.893, + 0.642 + ], + "angle": 0, + "content": "[72] Peihao Zhu, Rameen Abdal, Yipeng Qin, and Peter Wonka. Sean: Image synthesis with semantic region-adaptive normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5104-5113, 2020. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.642 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15685" + } + ] +] \ No newline at end of file diff --git a/2024/Generating Handwritten Mathematical Expressions From Symbol Graphs_ An End-to-End Pipeline/f14645bb-d996-46da-9c4b-81a14daf5714_origin.pdf b/2024/Generating Handwritten Mathematical Expressions From Symbol Graphs_ An End-to-End Pipeline/f14645bb-d996-46da-9c4b-81a14daf5714_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7ec3b47f304a3cf2ffc2a5a29d8b26ba8e30f2e9 --- /dev/null +++ b/2024/Generating Handwritten Mathematical Expressions From Symbol Graphs_ An End-to-End Pipeline/f14645bb-d996-46da-9c4b-81a14daf5714_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f0e3850f32fe9e8c2b0d6fad4b293defa8d5f2030a90a003d4357070f9f85647 +size 4052545 diff --git a/2024/Generating Handwritten Mathematical Expressions From Symbol Graphs_ An End-to-End Pipeline/full.md b/2024/Generating Handwritten Mathematical Expressions From Symbol Graphs_ An End-to-End Pipeline/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c4716e094f474792eb991fb0da3162512e01f671 --- /dev/null +++ b/2024/Generating Handwritten Mathematical Expressions From Symbol Graphs_ An End-to-End Pipeline/full.md @@ -0,0 +1,332 @@ +# Generating Handwritten Mathematical Expressions From Symbol Graphs: An End-to-End Pipeline + +Yu Chen $^{1,*}$ , Fei Gao $^{2,*}$ , Yanguang Zhang $^{3}$ , Maoying Qiao $^{4}$ , Nannan Wang $^{5,\dagger}$ + +1 Beijing Waiyan Online Digital Technology; 2 Hangzhou Institute of Technology, Xidian University; 3 Hangzhou Dianzi University; 4 University of Technology, Sydney (UTS); 5 Xidian University + +{fgao, nwwang}@xidian.edu.cn, ppakl991@outlook.com, maoying.qiao@uts.edu.au + +# Abstract + +In this paper, we explore a novel challenging generation task, i.e. Handwritten Mathematical Expression Generation (HMEG) from symbolic sequences. Since symbolic sequences are naturally graph-structured data, we formulate HMEG as a graph-to-image (G2I) generation problem. Unlike the generation of natural images, HMEG requires critic layout clarity for synthesizing correct and recognizable formulas, but has no real masks available to supervise the learning process. To alleviate this challenge, we propose a novel end-to-end G2I generation pipeline (i.e. graph $\rightarrow$ layout $\rightarrow$ mask $\rightarrow$ image), which requires no real masks or nondifferentiable alignment between layouts and masks. Technically, to boost the capacity of predicting detailed relations among adjacent symbols, we propose a Less-is-More (LiM) learning strategy. In addition, we design a differentiable layout refinement module, which maps bounding boxes to pixel-level soft masks, so as to further alleviate ambiguous layout areas. Our whole model, including layout prediction, mask refinement, and image generation, can be jointly optimized in an end-to-end manner. Experimental results show that, our model can generate high-quality HME images, and outperforms previous generative methods. Besides, a series of ablations study demonstrate effectiveness of the proposed techniques. Finally, we validate that our generated images promisingly boosts the performance of HME recognition models, through data augmentation. Our code and results are available at: https://github.com/AiArt-HDU/HMEG. + +# 1. Introduction + +Handwritten Mathematical Expressions (HMEs) are common and play significant roles in our daily life, especially in the research and education areas. HMEs generally present complex structures, serious deformations, and diverse writ + +![](images/fcec784f4caba580b21f7fef86fdc9591e1206d1c2a25c48a2fbe88fb62c1301.jpg) + +![](images/00a7884c7514022f7c0d291821a5b25e0030444fd17480240baa6d776ccf0ea7.jpg) +Figure 1. Differences between (a) typical two-stage graph-to-image generation pipeline and (b) our end-to-end pipeline. In previous methods, the real masks are available as the input. In contrast, we propose a novel end-to-end pipeline of $\text{graph} \to \text{layout} \to \text{mask} \to \text{image}$ , and requires no real masks or non-differentiable alignment between layouts and masks. + +ing styles. Such characteristics, along with data scarcity, make HME Recognition (HMER) a grand challenge in the OCR community. Despite the tremendous efforts that have been made to this task, the HMER performance is still unsatisfactory [14, 27, 62]. + +Recently, synthetic data augmentation has shown inspiring performance in various recognition tasks [23, 51]. This is mainly achieved by the remarkable progress of conditional image generation [21] in the past few years [58]. The advances allow for creating realistic images through texts [39, 57, 66], scene graphs [24, 60], layouts [28, 69], semantic masks [37, 72], sketches [61, 67], and more [12, 65]. However, there has been no effective generative model for generating high-quality HME images. + +Some previous works synthesize HMEs by recomposing real online HMEs [46, 59]. Besides, FormulaGAN [41] generates HMEs from rendered images, via an Image-to-Image (I2I) Translation model. However, their generated images are limited in either diversity or realism. Be + +sides, Mathematical Expressions (MEs) are highly structured data. Given an HME, its meaning and handwritten style, depend not only on the category and shape of each symbol, but also on the positional relations between symbols. It's unlikely to design an elaborate I2I translation model that can handle all these issues. + +To combat this challenge, in this paper, we explore a new generation task, i.e. Handwritten Mathematical Expression Generation (HMEG) from symbol graphs of symbolic sequences. It's natural and easy to use a graph, specifying symbols and positional relations between symbols, to represent an ME [63]. Besides, previous works have demonstrated the significance of fully mining the structural information in boosting generalization [52]. By converting symbolic sequences to symbol graphs, we propose a novel HMEG method, inspired by the classic graph-to-image (G2I) generation work, $\mathrm{Sg2im}$ [24]. Similar to existing G2I generation works, our method consists of three stages: graph-to-query prediction, layout-to-mask transformation, and mask-to-image generation. + +Compared to the generation of natural images, the key challenge of HMEG is to generate unambiguous structural layouts, for accurately presenting meanings of formulas. From one hand, the requirement of layout clarity is much critical for generating formulas. Not only the size of each symbol, but also the positional relations between symbols, should be precisely predicted. Otherwise, mathematical meanings might change or become unrecognisable. For example, an input formula $x^{2}$ might become $x_{2}$ or $x2$ in a generated image. From the other hand, MEs have infinite possible layouts and structures. Namely, a formula could comprise of arbitrary number of symbols and arbitrary relations between symbols. It's impossible to collect completed training data for learning the layout predictor. + +To alleviate these difficulties, we first propose a Less-is-More (LiM) training strategy for learning an effective layout predictor. Specially, our layout predictor follows the structure of Graph Convolutional Networks (GCNs) [30]. During training, we only use minimal graphs with 1-degree connections, in stead of whole graphs with complex structures, to optimize the layout predictor. Note that the learned predictor can be applied to arbitrary symbol graphs, in the testing stage. By using LiM, the predictor emphasizes on local structures between adjacent symbols. Besides, the learning process is eased. Namely, we could learn an effective layout predictor by using moderate amount of training samples. Such benefits will be verified in the ablation study part. Besides, we add a layout discriminator to further boosts the realism of predicted layouts. + +In addition, we propose a Sequential BBox-to-Mask Transformation (B2M) module, for refining layouts under weak supervision. In the original layout, bounding boxes of symbols tend to overlap with each other. Such overlapped + +areas may lead to overlapping symbols or chaotic strokes, in generated images. It's critical to refine coarse layouts to pixel-level soft masks, to alleviate the overlapping problems [22]. Most previous G2I generation works require real segmentation masks for learning refined layout masks [13, 22]. However, it's difficult to define the mask of a symbol, because white pixels around strokes are essential for representing a symbol. We have to refine the layout under weak supervision. To this end, we propose to learn a sequential mapping of layout $\rightarrow$ grid $\rightarrow$ mask, and use an subsequent image decoder to generate an HME image. During training, we use the image-related losses to train both B2M and the image decoder jointly. + +In summary, we propose a novel end-to-end pipeline of graph $\rightarrow$ layout $\rightarrow$ mask $\rightarrow$ image for G2I generation tasks. Our model requires no supervision of real masks, or interactive alignment operations between layouts and masks. The major differences between our method and previous typical G2I generation works, are illustrated in Fig. 1. To verify the effectiveness of the proposed techniques, we conduct experiments on CROHME2014/2016/2019 [34]. Results show that, our model can generate high-quality HME images with clear layouts and recognisable symbols. Besides, our model significantly outperforms a number of existing generative methods, both qualitatively and quantitatively. A series of ablation study demonstrate the effectiveness of the proposed techniques, including the LiM, layout discriminator, and B2M. Finally, our generated HME images prove to boost the HMER performance by $5 - 11\%$ absolutely, through synthetic data augmentation. + +# 2. Related Works + +HME Recognition (HMER). Most advanced methods treat HMER as an image-to-markup problem [53]. They typically use Convolutional Neural Networks (CNNs) or Recurrent Neural Networks (RNNs) [18] to encode input images; and use attention-based models, such as Gated Recurrent Units (GRU) [5, 14, 68], or Transformers [2, 47, 70], to generate LaTeX sequences. Besides, some methods have been proposed [41, 46, 59]. Some recent works propose to use additional information, e.g. symbol-level counting [27], or relation-level counting [14], or emphasis on visually similar symbols [33], or synthetic data augmentation, to boost the HMER performance. Another branch of HMER are graph-based methods. These methods typically transform LaTeX sequences to Symbol Label Trees (SLTs) [63], and use tree-structured decoders to obtain the outputs [14, 52, 54]. These works validate the capacity of graphs in representing the structural information of formulas. In this paper, we therefore convert symbolic sequences to graphs, and use a GCN to encode them for generating HME images. + +Graph-to-Image Generation. Image generation from scene graphs is first proposed in Sg2im [24]. The pipeline + +![](images/8c4dbbfe45917d3b66d51ac53881035d8b16b457cb7320288bfb3881619c192d.jpg) +Figure 2. Overview of our handwritten mathematical expression generator (HMEG). + +of Sg2im includes two stages: graph-to layout mapping and layout-to-image generation. Various subsequent methods have been proposed to boost the quality of generated images, following the same pipeline [9, 13, 16, 40]. Besides, some semi-parametric methods use crops from real images to boost the quality of generated images [31]. There are also some extensions to image manipulation [9, 42]. + +Most of these methods use a image discriminator and a object discriminator during training. Besides, they use real layouts, segmentation masks, and images, to supervise the learning process [13, 22]. COLoR [22] additionally uses a mask discriminator to boost the precision of generated masks. Recently, KCGM [55] uses knowledge consensus to alleviate the dependency on real layouts or masks. Besides, several diffusion model based methods [12, 60] can generate high-quality images directly from input scene graphs. However, diffusion based methods still suffer the high cost during both training and inference. Besides, generating recognisable texts is still challenging for most diffusion models. + +Layout-to-Image Generation. Most layout-to-image generation methods separately predict the mask of each object based on object embeddings, and require frequent resizing interaction between bounding boxes and masks [43, 44, 69]. To improve the quality of generated masks, several methods use masks discriminator [1, 28], and requires real masks during training. Besides, OC-GAN [45] uses both global and local similarities between scene graphs and images are used to improve the realism of generated layouts. Similar to our method, LAMA [32] tries to alleviate the overlapping between objects. However, in all these methods, the real bounding boxes are available as the input. They at most implement the end-to-end computing of layout $\rightarrow$ mask $\rightarrow$ image. In contrast, we propose a novel end-to-end computing of graph $\rightarrow$ layout $\rightarrow$ mask $\rightarrow$ image, and requires no interactive alignment operations between layouts and masks. + +Handwritten Font Generation. Another related task is Handwritten Font Generation (HFG), which aims at + +transferring images of fonts to the target handwritten style [48, 50]. Most methods follow the I2I translation pipeline, and adopt encoder-decoder architectures [6]. Besides, great efforts have been made to solve the few-shot learning challenge [48, 64]. These methods focus on the style of strokes. In contrast, our work aims to generate rational layouts of multiple symbols, with complex relations, at the same time. It's interesting to explore generating HME images in personal styles in the future. + +# 3. Method + +# 3.1. Overview + +Our goal is to generate HME images conditioned on LaTeX sequences. To this end, we represent LaTeX sequences by symbol graphs [63] and design a novel G2I generation method to synthesize HME images. Our full model follows the idea of Generative Adversarial Networks (GANs) [21, 24]. It includes a generator, and several discriminators in the training stage. Fig. 2 shows the pipeline of our generator. It mainly include four parts: (i) First, the input LaTeX sequence is converted to a symbol graph, and embedded into high-dimensional feature vectors. (ii) A GCN based layout predictor estimates the coarse layout, i.e. bounding box (BBox) of each symbol. (iii) A sequential layout refinement module, including BBox-to-Grid mapping and Grid-to-Mask projection, generate a pixel-level soft mask for each symbol. (iv) The image decoder, a Cascaded Refinement Network (CRN) [4], generates an image conditioned on the predicted mask. + +# 3.2. Symbol Graph Construction + +Given a set of mathematical symbol categories $\mathcal{C}$ and a set of relation categories $\mathcal{R}$ , a symbol graph $(S,E)$ can be constructed based on $\mathcal{C}$ and $\mathcal{R}$ : each node $i$ in $S = \{s_1,\dots ,s_n\}$ is a mathematical symbol, with $s_i\in \mathcal{C}$ ; and each directed edge $(i,j)$ in $E = \{e_{ij}\}$ represents the positional relation of node $j$ w.r.t. node $i$ , with $e_{ij} = (s_i,r,s_j)$ and $r\in \mathcal{R}$ . $n$ is the number of symbols in a mathematical + +![](images/4f44b67ea813ce3720724f157f613465b666c53d36143b5da3335430baac276c.jpg) +Figure 3. Architecture of our layout predictor and the Graph Embedding Model (GEM) [30]. + +expression. Let $m$ denote the number of edges in $E$ . For example, given an expression $x^2$ , it can be represented by a tuple $(x, \sup, 2)$ , where $\sup$ is short for superscript. Besides, we add an additional edge type None to indicate that the corresponding pair of symbols are unrelated. + +Afterwards, we map the nodes and edges $(S,E)$ to high-dimensional feature vectors $\{\mathbf{V},\mathbf{U}\}$ by using a learnable embedding layer, respectively [30]. The learned embedding feature vectors are denoted by: + +$$ +\mathbf {V} = \left\{\mathbf {v} _ {i}, \forall i \in S \right\}, \mathbf {U} = \left\{\mathbf {u} _ {i j}, \forall (i, j) \in E \right\}, \tag {1} +$$ + +where $\mathbf{v}_i, \mathbf{u}_{ij} \in \mathbb{R}^d$ associate with $s_i, e_{ij}$ , respectively; $d$ is the dimension of embedding vectors. + +# 3.3. Adversarial Layout Prediction + +Inspired by GAN [21], we train our layout prediction module in an adversarial learning manner. Specially, we use an layout predictor $G_{box}$ to generate a layout in the format of bounding boxes, conditioned on a given symbol graph; and use an additional layout discriminator $D_{box}$ to judge whether a layout is real or fake. Using the adversarial learning strategy like GAN, the layout predictor would finally predict high-quality bounding boxes. + +Layout Predictor. We first use the Graph Embedding Model (GEM) [30] to update features of nodes and edges, and then use a MLP to predict the BBox indices of each symbol. First, original embedding features are encoded by: $\mathbf{h}_i^{(0)} = f_{\mathrm{node}}(\mathbf{v}_i), \forall i \in S$ and $\mathbf{e}_{i,j} = f_{\mathrm{edge}}(\mathbf{u}_{ij}), \forall (i,j) \in E$ . Afterward, node features are updated through a stack of propagation layers, by accumulating information in local neighborhood. Features of node $i$ at the $t$ -th layer are updated by: + +$$ +\mathbf {m} _ {j \rightarrow i} = f _ {\text {m e s s a g e}} \left(\mathbf {h} _ {i} ^ {(t)}, \mathbf {h} _ {j} ^ {(t)}, \mathbf {e} _ {i j}\right), \forall (i, j) \in E, +$$ + +$$ +\mathbf {h} _ {i} ^ {(t + 1)} = f _ {\text {n o d e}} \left(\mathbf {h} _ {i} ^ {(t)}, \sum_ {j} \mathbf {m} _ {j \rightarrow i}\right). \tag {2} +$$ + +After $T$ rounds of propagations, the model produces an aggregated graph level representation by: + +$$ +\mathbf {h} _ {G} = f _ {G} \left(\sum_ {i} \sigma \left(f _ {\text {g a t e}} \left(\mathbf {h} _ {i} ^ {(T)}\right)\right) \odot f _ {\text {u p d a t e}} \left(\mathbf {h} _ {i} ^ {(T)}\right)\right), \tag {3} +$$ + +where $\sigma$ denotes the Sigmoid operation. Finally, we integrate the updated embedding features and the graph level features for predicting the BBox indices of each symbol. Besides, we concatenate a Gaussian noise vector $\mathbf{z}_i\in \mathbb{R}^k,\forall i\in S$ to these features, to improve the diversity in predicted layouts. The BBox prediction is formulated by: + +$$ +\hat {\mathbf {b}} _ {i} = f _ {\mathrm {b b o x}} \left(\mathbf {h} _ {i} ^ {(T)}, \mathbf {h} _ {G}, \mathbf {z} _ {i}\right), \forall i \in S. \tag {4} +$$ + +Each of the mapping functions $f_{*}$ above is an MLP on the inputs [29, 30]. For each symbol, we obtain a 4-dimensional rectangular box $[x_0, y_0, x_1, y_1]$ in the $[0,1]$ coordinate space, denoting the coordinates of the upper left and lower right corners of the BBox border. + +Layout Discriminator. The layout discriminator $D_{box}$ is also composed of GEM [30], followed by an MLP to predict whether an input layout as real or fake. $D_{box}$ takes concatenations of the symbol embedding matrix $\mathbf{V} \in \mathbf{R}^{n \times d}$ and the layout indices $B \in \mathbf{R}^{n \times 4}$ as input. In the implementation, we use the least square version of adversarial loss, i.e. + +$$ +\mathcal {L} _ {a d v} ^ {b o x} = \left\| D _ {b o x} (\mathbf {V}, B) \right\| _ {2} ^ {2} + \left\| 1 - D _ {b o x} (\mathbf {V}, \hat {B}) \right\| _ {2} ^ {2}. \tag {5} +$$ + +Besides, we use the L2 distance between the predicted BBox indices $\hat{B}$ and the target (real) indices $B$ as the layout reconstruction loss, i.e. $\mathcal{L}_{box} = ||B - \hat{B}||_1$ . We use a weighted sum of $\mathcal{L}_{box}$ and $\mathcal{L}_{adv}^{box}$ as the objective of $G_{box}$ . During training, $G_{box}$ and $D_{box}$ are alternatively optimized. + +Less-is-More Learning (LiM) Strategy. A mathematical expression could include arbitrary number of symbols or relations between symbols. In other words, there are infinite possible structures of symbol graphs. It's difficult to collect a complete training set. Besides, using complicated graphs during training may add the difficulty in learning an effective model. We therefore propose a Less-is-More (LiM) learning strategy [3]. Specially, we propose to use symbol graphs with merely 1-degree connections during training. In this way, the layout predictor is expected to emphasize on local structures, i.e. generating rational layout for locally connected symbols. While in the testing stage, the learned model can be applied to arbitrary symbol graphs, with arbitrary degrees of connections, and predict the corresponding layout. + +# 3.4. Sequential BBox-to-Mask Transformation + +We next predict the pixel-level soft mask of an expression based on the predicted bounding boxes. This procedure is illustrated in Fig. 4, and consists of the following two steps: + +(1) BBox-to-Grid Mapping. We first transform the BBox indices to layout grids, as illustrated in Fig. 4. To this end, we subscribe each index value from the coordinate matrices $\mathbf{X}$ and $\mathbf{Y}$ in the [0, 1] coordinate space. For each symbol, we obtain a $4 \times w \times w$ tensor of grid. + +![](images/028bf7dcbd64b165c184a53c78cc2a1f474979fb3652d96f64e40062449abbb4.jpg) +Figure 4. Illustration of the sequential BBox-to-Mask transformation (B2M) module. + +(2) Grid-to-Mask Projection. Afterwards, we use a FCN with a Sigmoid output layer (denoted by $\phi_{\mathrm{mask}}$ ) to estimate a one-channel mask for each symbol, i.e. + +$$ +\mathbf {M} _ {i} = \phi_ {\text {m a s k}} (\mathbf {X} - x _ {i, 0}, \mathbf {X} - x _ {i, 1}, \mathbf {Y} - y _ {i, 0}, \mathbf {Y} - y _ {i, 1}). \tag {6} +$$ + +Each value in $\mathbf{M}_i$ indicates the probability of a pixel belonging to symbol $s_i$ . Specially, we first use two Convolutional layers to predict a low-resolution mask of dimension $w \times w$ , and then use two upsampling Convolutional layers to sequentially upsampling the mask by a factor of 2. These masks are used in the image decoder for generating an image of the same size, respectively. + +Note: Previous G2I generation works [9] also consists of BBox-to-Grid mapping, but they merely use the grids to select patches of objects. Besides, they predict the mask from the BBox indices $\hat{B}$ or original embedding features $(\mathbf{V},\mathbf{U})$ . Differently, we use the transformed grids for layout mask prediction. Moreover, since the whole BBox-to-Mask prediction module is differentiable, our whole model enables training in an end-to-end manner. + +# 3.5. Image Generation + +Image Decoder. In the implementation, we adopt CRN [4] as the image decoder. The input is a multiplication of the whole layout mask $\mathbf{M} \in \mathbb{R}^{n \times w \times w}$ with the embedding vectors $\mathbf{V} \in \mathbb{R}^{n \times d}$ . Besides, we add Gaussian noise to the feature maps to improve the diversity in image generation. CRN includes three upsampling Convolutional layers, and outputs a formula image with resolutions of $w \times w$ . Recall that we actually predicts three masks at different scales. In the implementation, we have three generated images, at the resolution of $64 \times 64$ , $128 \times 128$ , and $256 \times 256$ , respectively. All these images are used to calculate the losses for optimizing our model. + +Image and Symbol Discriminators Finally, we use an image discriminator $D_{img}$ and a symbol discriminator $D_{sym}$ during training, to boost the realism of the generated whole image $\hat{I}$ and patches of symbols, respectively. We use the grids to select patches of symbols, and input them into $D_{sym}$ . Both discriminators are fully CNNs with the cross-entropy (CE) version of adversarial loss. In addition, + +following AC-GAN [36], we have the symbol discriminator additionally predict category labels of symbols. In this way, the generated symbols would be more recognizable and real. + +Loss Functions. Similar to previous G2I generation works [24], we use a series of loss functions to optimize the image generation. The losses are briefly introduced below. + +- Pixel loss $\mathcal{L}_{pix}$ : First, we use the L1 distance between a generated image and the target formula image and as the pixel loss: $\mathcal{L}_{pix} = ||I - \hat{I}||_1$ . +- Image adversarial loss $\mathcal{L}_{adv}^{img}$ : Second, we use the global image discriminant loss from $D_{img}$ , to encourage the generated HME images in the real handwritten style. +- Symbol adversarial loss $\mathcal{L}_{adv}^{sym}$ : Third, we use the symbol discriminative loss from $D_{sym}$ , to improve the fidelity of generated symbols. +- Auxiliary Symbol Recognition Loss $\mathcal{L}_{aux}^{sym}$ : Finally, we use the auxiliary symbol recognition loss from $D_{sym}$ , to enforce the generator producing recognizable symbols. + +We use a combination of these loss functions above, to train the B2M module and the image generation modules together. The total loss is formulated by: + +$$ +\mathcal {L} _ {i m g} = \lambda_ {1} \mathcal {L} _ {p i x} + \lambda_ {2} \mathcal {L} _ {a d v} ^ {i m g} + \lambda_ {3} \mathcal {L} _ {a d v} ^ {s y m} + \lambda_ {4} \mathcal {L} _ {a u x} ^ {s y m}, \tag {7} +$$ + +where $\lambda_{i}, i = 1, \ldots, 4$ are weighting factors and are set as 1, 0.01, 0.01, 0.001, respectively. + +# 4. Experiments + +# 4.1. Experimental Settings + +Data. We use CROHME2014/2016/2019 [34] to validate our HMEG method in terms of both generalization and recognition. The dataset collects pen trajectories, and labels them with symbols and symbol categories. The dataset contains 126 different symbol categories, and 8 positional relations, i.e. start, left superscript, superscript, subscript, below, above, right and end. By restoring the pen trajectory information, we can easily obtain the HME image and the BBox of each symbol. We use the standard partition of each dataset through all the experiments. + +Implementation Details. In the implementation, due to the limitation of our computing resources, we train our model in two stages: we first train the layout prediction part in the first stage, and then train the rest parts by fixing it. In the first stage, we train the model with a batch size 64 and a learning rate 5e5, for 40,000 iterations. In each iteration, we random sample a batch of 1-degree sub-graphs from the training set for learning the layout predictor. In the second stage, we use a batch size 8, a learning rate 1e4, and train for 600,000 iterations. Our codes are implemented by using Pytorch and a NVIDIA TITAN XP GPU. We use Adam [26] as the optimizer for all networks. + +
Print|ax0+by0+c/√a2+b2|p=√a2+b2-2abcosA1/2(a+1/b) = 2ab/a+b-b+√b2-4ac/2a∫0πcos(θ/2)dθx/a+x/b-x/c∫bdIπr2h/3sin(π/3)=1/2
CycleGANaxn+byn+c/√n2+b2|p=√n2+b2-2abcosA1/2(a+1/b) = 2ab/a+b-b+√b2-4ac/2a∫0πcos(θ/2)dθx/a+x/b-x/c∫bdIπr2h/3sin(π/3)=1/2
FormulaGAN|ax0+by0+c/√a2+b2|p=√a2+b2-2abcosA1/2(a+1/b) = 2ab/a+b-b+√b2-4ac/2a∫0πcos(θ/2)dθx/a+x/b-x/c∫bdIπr2h/3sin(π/3)=1/2
Sg2imaxn+byn+c/√a2+b2|p=√a2+b2-2abcosA1/2(a+1/b) = 2ab/a+b-b+√b2-4ac/2a∫0πcos(θ/2)dθx/a+x/b-x/c∫bdIπr2h/3sin(π/3)=1/2
Ours|ax0+by0+c/√a2+b2|p=√a2+b2-2abcosA1/2(a+1/b) = 2ab/a+b-b+√b2-4ac/2a∫0πcos(θ/2)dθx/a+x/b-x/c∫bdIπr2h/3sin(π/3)=1/2
+ +Figure 5. Handwritten mathematical expressions generated by CycleGAN [71], FormulaGAN [41], Sg2im [24], and our method. + +Criteria. We choose two image quality assessment indices as the criteria. The former includes the Fréchet Inception distance (FID) [17] and Structural Similarity Measure (SSIM) [49]. Lower values of FID indicate better realism of synthesised images; while greater values of SSIM generally indicate higher similarity between a synthesized image and the corresponding real image. We here report the average SSIM value of all the test samples. In addition, we use two HMER indices, including the Word Error Rate (WER) and the Expression Recognition Rate (ExpRate). Here we use an open-source HMER method [19], to recognize HME images and compute the corresponding indices. Lower WER and higher ExpRate values indicate better performance. + +# 4.2. Comparison with SOTAs + +We first compare with the existing GAN based HMEG method, FormulaGAN [41], our baseline, Sg2im [24], and the widely used unsupervised I2I translation method, CycleGAN [71], on CHROME2019. We train and test these methods following the same experimental settings as ours. + +Qualitative Comparison. Fig. 5 shows HME images generated by different models. CycleGAN fails to transform input formulas to the handwritten style. FormulaGAN produces handwritten-style images, but with blurring and unnatural strokes. In contrast, both Sg2im and our method generate high quality images, with realistic styles and clear strokes. In addition, the images generated by our method present better layouts than those by Sg2im. Sg2im sometimes produce very small symbols, e.g. $y_{0}$ in the first col + +Table 1. Comparison with existing methods on CHROME2019. + +
SSIM↑FID ↓WER ↓ExpRate ↑
CycleGAN [71]0.75784.140.6710.026
FormulaGAN [41]0.72474.680.6010.066
Sg2im [24]0.78710.020.3930.219
Ours0.79310.980.3260.316
+ +unn and $p$ in the second column. Besides, Sg2im may produce overlapped symbols, e.g. $\frac{1}{2} \left( \frac{1}{a} + \frac{1}{b} \right)$ in the third column. Finally, Sg2im produces more unrecognisable symbols than ours, e.g. Sin and $\pi$ in last column. In contrast, the symbols generated by our method present reasonable sizes and clear structures. Besides, symbols are reasonably placed in relation to each other, neither overlapping nor far apart. This is similar to real formulas written by humans. Such observations demonstrate the superiority of our method in generating both high-quality layouts and HME images. + +Quantitative Comparison. Table 1 shows the quantitative performance indices. Both Sg2im and our method achieves significant better indices than CycleGAN or FormulaGAN. This is consistent with the qualitative comparison results. According to SSIM and FID, Sg2im and our method almost perform equality in term of image quality. However, our method decreases the WER by 6.5 percent and improves the ExpRate by 10 percent, absolutely. Such distinct superiority in recognition implies that, the handwritten expressions generated by our method are signifi + +![](images/b6559d3ca4776bbde12b0e6e0f99be7f8555c3f5b379b2742eb80994e2df6aff.jpg) +Figure 6. Illustration of the layouts and formula images generated by model variants, in the ablation study. The input formula is $\frac{1}{25} y^2 -\frac{8}{25} y$ + +![](images/8206013411e259c7a821c90375dc5cb86609c37dc2c8a76253dafe39a0ceaac2.jpg) +Figure 7. Comparison with diffusion models. DiffSketcher and DiffAE generate HMEs conditioned on print formulas; while ControlNet and SD+LoRA take both the print formula and a textual prompt ("a handwritten mathematical expression of latex code") as input. We use official diffusion models, and fine-tune DiffAE and SD+LoRA using HMEs. + +cantly better than $\mathrm{Sg2im}$ , in terms of clarity and structure. + +Comparison with Diffusion Models. We additionally compare with ControlNet [67], Stable Diffusion (SD) [39] + LoRA [20], DiffSketcher [56], and DiffAE [38]. As shown in Fig. 7, these diffusion models cannot generate high-quality HMEs with realistic handwritten styles or recognizable symbols. Besides, their computational complexity is much heavier than ours. Some recent methods, e.g. ChiroDiff [8] and SDT [7], can only generate handwritten images of one single character. Such comparison results demonstrate the advantage of our method in HMEG. + +# 4.3. Ablation Study + +We further analyse the impact of (i) the LiM training strategy, (ii) the layout discriminator $(D_{box})$ , and (iii) the B2M module. To this end, we build model variants by: + +- Using original training graphs, instead of sampled minimal graphs, to train the layout predictor $(w / o LiM)$ ; +- Removing B2M from our full model, but using grids for embedding selection (w/o B2M); +- Removing both LiM and B2M from our full model ( $w / o$ LiM/B2M); + +Table 2. Performance indices w.r.t. the ablation study on CHROME19. The model variant in the last row is our baseline. + +
SSIM↑FID↓WER↓ExpRate↑mIOU↑
Ours (full)0.79310.980.3260.3160.364
w/o LiM0.79011.550.3270.3150.335
w/o B2M0.79011.460.3320.3080.354
w/o LiM/B2M0.78613.300.3650.2790.338
Sg2im (base)0.78710.080.3930.2190.324
+ +- Training the whole model following Sg2im, without using LiM, $D_{box}$ , and B2M (Sg2im (base)). + +We additionally report the mean Intersection over Union (mIoU) between the predicted layouts and the target ones. + +As shown in Table 2, removing LiM or B2M slightly changes the indices about image quality, decreases the IoU. Specially, removing LiM decreases the mIoU by about 3 percent absolutely. This implies the significance of LiM in generating good layouts. In addition, if we removing $D_{box}$ or B2M, along with LiM, the performance seriously decreases, almost in terms of all performance indices. This implies that our $D_{box}$ and B2M are significant for boosting the quality of generated images from all aspects, including the fidelity (SSIM), the usability for recognition (WER & ExpRate), and the layout (mIoU). Fig. 6 illustrates the layouts and images generated by different model variants. All the model variants, except the full model, produce overlapped symbols or symbols in unnatural sizes. The image generated by our model presents no obvious defects and are easy to recognize. Such observations further demonstrate the significance of LiM, B2M, and layout discriminator in layout prediction and image generation. + +# 4.4. Applications + +Mathematical Expression Manipulation. In addition, we apply the previous learned model to expression manipulation using symbol graphs. Given a source symbol graph, + +
Source1/3πr2hx/a + x/b - x/cax + bx + c/2a + √b + c/2∫_0^1 x^2y dxdyT_1^2/T_2^2 = a_1^3/a_2^3
Add1/3πr^2hzx + d/a + x/b - x/cax + b^x / 2a + √b + c/2 + 7∫_0^1 x^2y^2/2 dxdyT_1^2/T_2^2 = a_1^3 + 2/a_2^3
Delete1/3πr^2x/a + x/bax + b^x + ca + √b / 2∫_0^1 x^2y^2 dxdyT_1^2/T_2^2 = a/a
Change1/7πr^2hx/a + x/e - x/cax + b^y + c/2a + √b + c/4∫_0^1 x^2y dxdyT_1^2/T_2^2 = e^3/a^3
+ +Figure 8. Illustration of expression manipulation using symbol graphs (LaTeX sequences). + +we edit it by adding, deleting, or changing symbol nodes in it, and apply the previously learned generator to produce the corresponding HME image. As shown in Fig. 8, the generated image changes consistently with the editing operation. Besides, the manipulated images present natural layouts, recognizable symbols, and clear strokes. + +Data Augmentation for HMER. Finally, we validate the possible application of HMEG in augmenting HMER models, since the generated data could improve the amount and diversity of training data. To this end, we conduct experiments on CROHME2014/2016/2019 [34], and use CAN [27] as the HMER model. Here, we train CAN with (1) the standard training data (denoted by CAN); or (2) with previous data augmentation, e.g. rotation, (denoted by $\mathrm{CAN^{*}}$ ); or (3) with our synthetic data augmentation, using 6,000 additional generated samples (denoted by $\mathrm{CAN^{\dagger}}$ ); or (4) both previous and our data augmentation (denoted by $\mathrm{CAN^{*\dagger}}$ ). The model is trained for 120 epochs under each setting, using the official training parameters. Afterward, we apply the learned HMER model to both generated and real data. + +As shown in Table 3, our method achieves $5 - 11\%$ absolute improvement over the baseline, and performs comparably to previous data augmentation, across all the datasets. Besides, our method can be jointly used with previous augmentation to further boost the performance (i.e. $\mathrm{CAN}^{*\dagger}$ ). We additionally report the HMER performance, as well as the SSIM and StruRate, of our generated samples. The SSIM and StruRate values are high, and the ExpRate values (by $\mathrm{CAN}^{*} / \mathrm{CAN}^{\dagger}$ ) approach those of real data, on all datasets. These results demonstrate the robustness of our method, in generating high-quality HMEs, with correct layouts and recognizable symbols. + +Note: Previous HME synthesis methods [10, 25] synthesize HMEs by shuffling the symbols and layouts contained in original HME datasets. Neither of them is a generative method. And they may cause overfitting to existing datasets. In contrast, our pipeline allows excellent flexibil + +Table 3. Results on CROHME14/16/19 testing sets and our generated images. *,†, *† denote using previous data augmentation, our synthetic data augmentation, and both, respectively. + +
HMEG (generated)HMER (real)
SSIMCAN*CAN†StruRateCANCAN*CAN†CAN*†
CROHME140.78952.154.198.144.752.950.255.4
CROHME160.79851.455.896.542.852.453.957.6
CROHME190.79355.456.497.739.348.449.658.5
+ +ity in generating diverse new samples. The experimental results also demonstrate the advantages of our approach in both HMEG and HMER. + +# 5. Conclusions + +In this paper, we propose a novel method to solve the challenging task of generating HMEs from LaTeX sequences. We formulate it as a graph-to-image generation task, and focus on boosting the layout prediction without mask supervision. Experimental results demonstrate the effectiveness of our method, and its possible applications in expression manipulation and the challenging inverse task, i.e. HMER. Inspired by the remarkable process of generative augmentation in visual understanding [35], we believe that: the novel HMEG task we explore here has great potential to significantly boost HMER. + +Due to computational limitations, we merely use naive networks in the implementation. The generated symbols are occasionally unrecognizable, especially in complex formulas. It's promising to boost the generation quality by using advanced networks (e.g. VQ-GAN [11] or diffusion models [15]). Besides, we will extend the proposed techniques to natural/artistic image generation by exploring the inspiring capacity of diffusion models, and other tasks of structured graph data. Finally, to further boost the HMER performance, it's promising to use the triples of $\{LaTeX, predicted layout, generated image\}$ in HMEG as pseudo labeled data, or to use the layout discriminator for enhancing the layout detection module in an HMER pipeline. + +# Acknowledgements + +We greatly appreciate Biao Ma's and Chang Jiang's help in the implementation of HMER. This work was supported in part by the National Natural Science Foundation of China under Grants U22A2096 and 61971172; in part by the Technology Innovation Leading Program of Shaanxi under Grant 2022QFY01-15; in part by the Fundamental Research Funds for the Central Universities under Grant QTZX23042 and ZYTS24012; and in part by the Proof of Concept Foundation of Xidian University Hangzhou Institute of Technology under Grant No. GNYZ2023YL0301. + +# References + +[1] Oron Ashual and Lior Wolf. Specifying object attributes and relations in interactive scene generation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4561-4569, 2019. 3 +[2] Xiaohang Bian, Bo Qin, Xiaozhe Xin, Jianwu Li, Xuefeng Su, and Yanfeng Wang. Handwritten mathematical expression recognition via attention aggregation based bidirectional mutual learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 113-121, 2022. 2 +[3] Stefan Bornholdt. Less is more in modeling large genetic networks. Science, 310(5747):449-451, 2005. 4 +[4] Qifeng Chen and Vladlen Koltun. Photographic image synthesis with cascaded refinement networks. In Proceedings of the IEEE international conference on computer vision, pages 1511-1520, 2017. 3, 5 +[5] Kyunghyun Cho, Bart Van Merrienboer, Dzmitry Bahdanau, and Yoshua Bengio. On the properties of neural machine translation: Encoder-decoder approaches. arXiv preprint arXiv:1409.1259, 2014. 2 +[6] Gang Dai, Yifan Zhang, Qingfeng Wang, Qing Du, Zhuliang Yu, Zhuoman Liu, and Shuangping Huang. Disentangling writer and character styles for handwriting generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5977-5986, 2023. 3 +[7] Gang Dai, Yifan Zhang, Qingfeng Wang, Qing Du, Zhuliang Yu, Zhuoman Liu, and Shuangping Huang. Disentangling writer and character styles for handwriting generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5977-5986, 2023. 7 +[8] Ayan Das, Yongxin Yang, Timothy M Hospedales, Tao Xiang, and Yi-Zhe Song. Chirodiff: Modelling chirographic data with diffusion models. In The Eleventh International Conference on Learning Representations, 2023. 7 +[9] Helisa Dhamo, Azade Farshad, Iro Laina, Nassir Navab, Gregory D Hager, Federico Tombari, and Christian Rupprecht. Semantic image manipulation using scene graphs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5213-5222, 2020. 3, 5 +[10] Lan-Fang Dong, Han-Chao Liu, and Xin-Ming Zhang. Synthetic data generation and shuffled multi-round training based offline handwritten mathematical expression recognition. Journal of Computer Science and Technology, 37(6): 1427-1443, 2022. 8 +[11] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 8 +[12] Wan-Cyuan Fan, Yen-Chun Chen, DongDong Chen, Yu Cheng, Lu Yuan, and Yu-Chiang Frank Wang. Frido: Feature pyramid diffusion for complex scene image synthesis. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 579-587, 2023. 1, 3 +[13] Azade Farshad, Sabrina Musatian, Helisa Dhamo, and Nassir Navab. Migs: Meta image generation from scene graphs. arXiv preprint arXiv:2110.11918, 2021. 2, 3 + +[14] Yingnan Fu, Wenyuan Cai, Ming Gao, and Aoying Zhou. Symbol location-aware network for improving handwritten mathematical expression recognition. In Proceedings of the 2023 ACM International Conference on Multimedia Retrieval, pages 516-524, 2023. 1, 2 +[15] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10696-10706, 2022. 8 +[16] Roei Herzig, Amir Bar, Huijuan Xu, Gal Chechik, Trevor Darrell, and Amir Globerson. Learning canonical representations for scene graph to image generation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVI 16, pages 210-227. Springer, 2020. 3 +[17] M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In NIPS, pages 6626-6637. 2017. 6 +[18] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997. 2 +[19] Hongyu Wang and Jianshu Zhang. Handwritten mathematical expression recognition (pytorch), 2019. https://github.com/whywhs/Pytorch-Handwritten-Mathematical-Expression-Recognition.6 +[20] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2021. 7 +[21] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 1, 3, 4 +[22] Maor Ivgi, Yaniv Benny, Avichai Ben-David, Jonathan Berant, and Lior Wolf. Scene graph to image generation with contextualized object layout refinement. In 2021 IEEE International Conference on Image Processing (ICIP), pages 2428-2432. IEEE, 2021. 2, 3 +[23] Nikita Jaipuria, Xianling Zhang, Rohan Bhasin, Mayar Arafa, Punarjay Chakravarty, Shubham Shrivastava, Sagar Manglani, and Vidya N Murali. Deflating dataset bias using synthetic data augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 772-773, 2020. 1 +[24] Justin Johnson, Agrim Gupta, and Li Fei-Fei. Image generation from scene graphs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1219-1228, 2018. 1, 2, 3, 5, 6 +[25] Vu Tran Minh Khuong, Ung Quang Huy, Nakagawa Masaki, and Minh Khanh Phan. Generating synthetic handwritten mathematical expressions from a latex sequence or a mathml script. In 2019 International Conference on Document Analysis and Recognition (ICDAR), pages 922-927. IEEE, 2019. 8 +[26] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In 3rd International Conference on + +Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 5 +[27] Bohan Li, Ye Yuan, Dingkang Liang, Xiao Liu, Zhilong Ji, Jinfeng Bai, Wenyu Liu, and Xiang Bai. When counting meets hmer: counting-aware network for handwritten mathematical expression recognition. In European Conference on Computer Vision, pages 197-214. Springer, 2022. 1, 2, 8 +[28] Wenbo Li, Pengchuan Zhang, Lei Zhang, Qiuyuan Huang, Xiaodong He, Siwei Lyu, and Jianfeng Gao. Object-driven text-to-image synthesis via adversarial training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12174-12182, 2019. 1, 3 +[29] Yujia Li, Daniel Tarlow, Marc Brockschmidt, and Richard Zemel. Gated graph sequence neural networks. arXiv preprint arXiv:1511.05493, 2015. 4 +[30] Yujia Li, Chenjie Gu, Thomas Dullien, Oriol Vinyals, and Pushmeet Kohli. Graph matching networks for learning the similarity of graph structured objects. In International conference on machine learning, pages 3835-3845. PMLR, 2019. 2, 4 +[31] Yikang Li, Tao Ma, Yeqi Bai, Nan Duan, Sining Wei, and Xiaogang Wang. Pastegan: A semi-parametric method to generate image from scene graph. Advances in Neural Information Processing Systems, 32, 2019. 3 +[32] Zejian Li, Jingyu Wu, Immanuel Koh, Yongchuan Tang, and Lingyun Sun. Image synthesis from layout with locality-aware mask adaption. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13819-13828, 2021. 3 +[33] Zhe Li, Xinyu Wang, Yuliang Liu, Lianwen Jin, Yichao Huang, and Kai Ding. Improving handwritten mathematical expression recognition via similar symbol distinguishing. IEEE Transactions on Multimedia, 2023. 2 +[34] Mahshad Mahdavi, Richard Zanibbi, Harold Mouchere, Christian Viard-Gaudin, and Utpal Garain. Icdar 2019 crohme+ tfd: Competition on recognition of handwritten mathematical expressions and typeset formula detection. In 2019 International Conference on Document Analysis and Recognition (ICDAR), pages 1533-1538. IEEE, 2019. 2, 5, 8 +[35] Quang Nguyen, Truong Vu, Anh Tran, and Khoi Nguyen. Dataset diffusion: Diffusion-based synthetic data generation for pixel-level semantic segmentation. Advances in Neural Information Processing Systems, 36, 2024. 8 +[36] Augustus Odena, Christopher Olah, and Jonathon Shlens. Conditional image synthesis with auxiliary classifier gans. In International conference on machine learning, pages 2642-2651. PMLR, 2017. 5 +[37] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2337-2346, 2019. 1 +[38] Konpat Preechakul, Nattanat Chathee, Suttisak Wizadwongsa, and Supasorn Suwajanakorn. Diffusion autoencoders: Toward a meaningful and decodable representation. In Proceedings of the IEEE/CVF Conference on Computer + +Vision and Pattern Recognition, pages 10619-10629, 2022. 7 +[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 7 +[40] Renato Sortino, Simone Palazzo, Francesco Rundo, and Concetto Spampinato. Transformer-based image generation from scene graphs. Computer Vision and Image Understanding, 233:103721, 2023. 3 +[41] Matthias Springstein, Eric Müller-Budack, and Ralph Ewert. Unsupervised training data generation of handwritten formulas using generative adversarial networks with self-attention. In Proceedings of the 2021 Workshop on MultiModal Pre-Training for Multimedia Understanding, pages 46–54, 2021. 1, 2, 6 +[42] Sitong Su, Lianli Gao, Junchen Zhu, Jie Shao, and Jingkuan Song. Fully functional image manipulation using scene graphs in a bounding-box free way. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1784-1792, 2021. 3 +[43] Wei Sun and Tianfu Wu. Image synthesis from reconfigurable layout and style. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10531-10540, 2019. 3 +[44] Wei Sun and Tianfu Wu. Learning layout and style reconfigurable gans for controllable image synthesis. IEEE transactions on pattern analysis and machine intelligence, 44(9): 5070-5087, 2021. 3 +[45] Tristan Sylvain, Pengchuan Zhang, Yoshua Bengio, R Devon Hjelm, and Shikhar Sharma. Object-centric image generation from layouts. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2647-2655, 2021. 3 +[46] Thanh-Nghia Truong, Cuong Tuan Nguyen, and Masaki Nakagawa. Syntactic data generation for handwritten mathematical expression recognition. Pattern Recognition Letters, 153:83-91, 2022. 1, 2 +[47] Ashish Vaswani, Noam Shazeer, Niki Parmar, Llion Jones, Jakob Uszkoreit, Aidan N Gomez, and Lukasz Kaiser. Attention is all you need. In Advances in Neural Information Processing Systems, pages 5998-6008. 2017. 2 +[48] Chi Wang, Min Zhou, Tiezheng Ge, Yuning Jiang, Hujun Bao, and Weiwei Xu. Cf-font: Content fusion for few-shot font generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1858-1867, 2023. 3 +[49] Zhou Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Process, 13(4):600-612, 2004. 6 +[50] Chuan Wen, Yujie Pan, Jie Chang, Ya Zhang, Siheng Chen, Yanfeng Wang, Mei Han, and Qi Tian. Handwritten chinese font generation with collaborative stroke refinement. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3882-3891, 2021. 3 + +[51] Chenxi Whitehouse, Monojit Choudhury, and Alham Fikri Aji. Llm-powered data augmentation for enhanced crosslingual performance. arXiv preprint arXiv:2305.14288, 2023. 1 +[52] Changjie Wu, Jun Du, Yunqing Li, Jianshu Zhang, Chen Yang, Bo Ren, and Yiqing Hu. Tdv2: A novel tree-structured decoder for offline mathematical expression recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2694–2702, 2022. 2 +[53] Jin-Wen Wu, Fei Yin, Yan-Ming Zhang, Xu-Yao Zhang, and Cheng-Lin Liu. Handwritten mathematical expression recognition via paired adversarial learning. International Journal of Computer Vision, 128:2386-2401, 2020. 2 +[54] Jin-Wen Wu, Fei Yin, Yan-Ming Zhang, Xu-Yao Zhang, and Cheng-Lin Liu. Graph-to-graph: towards accurate and interpretable online handwritten mathematical expression recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2925-2933, 2021. 2 +[55] Yang Wu, Pengxu Wei, and Liang Lin. Scene graph to image synthesis via knowledge consensus. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2856-2865, 2023. 3 +[56] Ximing Xing, Chuang Wang, Haitao Zhou, Jing Zhang, Qian Yu, and Dong Xu. Diffsketcher: Text guided vector sketch synthesis through latent diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 7 +[57] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Fine-grained text to image generation with attentional generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1316-1324, 2018. 1 +[58] Yuan Xue, Yuan-Chen Guo, Han Zhang, Tao Xu, Song-Hai Zhang, and Xiaolei Huang. Deep image synthesis from intuitive user input: A review and perspectives. Computational Visual Media, 8:3-31, 2022. 1 +[59] Chen Yang, Jun Du, Jianshu Zhang, Changjie Wu, Mingjun Chen, and JiaJia Wu. Tree-based data augmentation and mutual learning for offline handwritten mathematical expression recognition. Pattern Recognition, 132:108910, 2022. 1, 2 +[60] Ling Yang, Zhilin Huang, Yang Song, Shenda Hong, Guohao Li, Wentao Zhang, Bin Cui, Bernard Ghanem, and Ming-Hsuan Yang. Diffusion-based scene graph to image generation with masked contrastive pre-training. arXiv preprint arXiv:2211.11138, 2022. 1, 3 +[61] Jun Yu, Xingxin Xu, Fei Gao, Shengjie Shi, Meng Wang, Dacheng Tao, and Qingming Huang. Toward realistic face photo-sketch synthesis via composition-aided gans. IEEE transactions on cybernetics, 51(9):4350-4362, 2020. 1 +[62] Ye Yuan, Xiao Liu, Wondimu Dikubab, Hui Liu, Zhilong Ji, Zhongqin Wu, and Xiang Bai. Syntax-aware network for handwritten mathematical expression recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4553-4562, 2022. 1 +[63] Richard Zanibbi and Dorothea Blostein. Recognition and retrieval of mathematical expressions. International Journal on Document Analysis and Recognition (IJDAR), 15:331-357, 2012. 2, 3 + +[64] Jinshan Zeng, Qi Chen, Yunxin Liu, Mingwen Wang, and Yuan Yao. Strokegan: Reducing mode collapse in Chinese font generation via stroke encoding. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3270-3277, 2021. 3 +[65] Yu Zeng, Zhe Lin, Jianming Zhang, Qing Liu, John Collo-mosse, Jason Kuen, and Vishal M Patel. Scenecomposer: Any-level semantic image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22468-22478, 2023. 1 +[66] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris Metaxas. StackGAN++: Realistic image synthesis with stacked generative adversarial networks. IEEE TPAMI, PP(99), 2017. 1 +[67] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1, 7 +[68] Yaping Zhang, Shuai Nie, Wenju Liu, Xing Xu, Dongxiang Zhang, and Heng Tao Shen. Sequence-to-sequence domain adaptation network for robust text image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2 +[69] Bo Zhao, Weidong Yin, Lili Meng, and Leonid Sigal. Layout2image: Image generation from layout. International Journal of Computer Vision, 128:2418-2435, 2020. 1, 3 +[70] Wenqi Zhao and Liangcai Gao. Comer: Modeling coverage for transformer-based handwritten mathematical expression recognition. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVIII, pages 392-408. Springer, 2022. 2 +[71] J. Zhu, T. Park, P. Isola, and A. A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In ICCV, pages 2242-2251, 2017. 6 +[72] Peihao Zhu, Rameen Abdal, Yipeng Qin, and Peter Wonka. Sean: Image synthesis with semantic region-adaptive normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5104-5113, 2020. 1 \ No newline at end of file diff --git a/2024/Generating Handwritten Mathematical Expressions From Symbol Graphs_ An End-to-End Pipeline/images.zip b/2024/Generating Handwritten Mathematical Expressions From Symbol Graphs_ An End-to-End Pipeline/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9c69bcec0c94f6aada9571d251b4db94c769636f --- /dev/null +++ b/2024/Generating Handwritten Mathematical Expressions From Symbol Graphs_ An End-to-End Pipeline/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:78b9e665e39f2d00d18bee5768a81ce3b5d5578a7b1650d1332455951e51dcf2 +size 449340 diff --git a/2024/Generating Handwritten Mathematical Expressions From Symbol Graphs_ An End-to-End Pipeline/layout.json b/2024/Generating Handwritten Mathematical Expressions From Symbol Graphs_ An End-to-End Pipeline/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..347611e1b32f2b05a5ff3dba33978cdad3281f36 --- /dev/null +++ b/2024/Generating Handwritten Mathematical Expressions From Symbol Graphs_ An End-to-End Pipeline/layout.json @@ -0,0 +1,9772 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 66, + 102, + 529, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 102, + 529, + 140 + ], + "spans": [ + { + "bbox": [ + 66, + 102, + 529, + 140 + ], + "type": "text", + "content": "Generating Handwritten Mathematical Expressions From Symbol Graphs: An End-to-End Pipeline" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 160, + 481, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 160, + 481, + 175 + ], + "spans": [ + { + "bbox": [ + 111, + 160, + 481, + 175 + ], + "type": "text", + "content": "Yu Chen" + }, + { + "bbox": [ + 111, + 160, + 481, + 175 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 111, + 160, + 481, + 175 + ], + "type": "text", + "content": ", Fei Gao" + }, + { + "bbox": [ + 111, + 160, + 481, + 175 + ], + "type": "inline_equation", + "content": "^{2,*}" + }, + { + "bbox": [ + 111, + 160, + 481, + 175 + ], + "type": "text", + "content": ", Yanguang Zhang" + }, + { + "bbox": [ + 111, + 160, + 481, + 175 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 111, + 160, + 481, + 175 + ], + "type": "text", + "content": ", Maoying Qiao" + }, + { + "bbox": [ + 111, + 160, + 481, + 175 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 111, + 160, + 481, + 175 + ], + "type": "text", + "content": ", Nannan Wang" + }, + { + "bbox": [ + 111, + 160, + 481, + 175 + ], + "type": "inline_equation", + "content": "^{5,\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 175, + 539, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 175, + 539, + 204 + ], + "spans": [ + { + "bbox": [ + 54, + 175, + 539, + 204 + ], + "type": "text", + "content": "1 Beijing Waiyan Online Digital Technology; 2 Hangzhou Institute of Technology, Xidian University; 3 Hangzhou Dianzi University; 4 University of Technology, Sydney (UTS); 5 Xidian University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 92, + 205, + 498, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 205, + 498, + 217 + ], + "spans": [ + { + "bbox": [ + 92, + 205, + 498, + 217 + ], + "type": "text", + "content": "{fgao, nwwang}@xidian.edu.cn, ppakl991@outlook.com, maoying.qiao@uts.edu.au" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "spans": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 270, + 290, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 270, + 290, + 605 + ], + "spans": [ + { + "bbox": [ + 47, + 270, + 290, + 605 + ], + "type": "text", + "content": "In this paper, we explore a novel challenging generation task, i.e. Handwritten Mathematical Expression Generation (HMEG) from symbolic sequences. Since symbolic sequences are naturally graph-structured data, we formulate HMEG as a graph-to-image (G2I) generation problem. Unlike the generation of natural images, HMEG requires critic layout clarity for synthesizing correct and recognizable formulas, but has no real masks available to supervise the learning process. To alleviate this challenge, we propose a novel end-to-end G2I generation pipeline (i.e. graph " + }, + { + "bbox": [ + 47, + 270, + 290, + 605 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 47, + 270, + 290, + 605 + ], + "type": "text", + "content": " layout " + }, + { + "bbox": [ + 47, + 270, + 290, + 605 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 47, + 270, + 290, + 605 + ], + "type": "text", + "content": " mask " + }, + { + "bbox": [ + 47, + 270, + 290, + 605 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 47, + 270, + 290, + 605 + ], + "type": "text", + "content": " image), which requires no real masks or nondifferentiable alignment between layouts and masks. Technically, to boost the capacity of predicting detailed relations among adjacent symbols, we propose a Less-is-More (LiM) learning strategy. In addition, we design a differentiable layout refinement module, which maps bounding boxes to pixel-level soft masks, so as to further alleviate ambiguous layout areas. Our whole model, including layout prediction, mask refinement, and image generation, can be jointly optimized in an end-to-end manner. Experimental results show that, our model can generate high-quality HME images, and outperforms previous generative methods. Besides, a series of ablations study demonstrate effectiveness of the proposed techniques. Finally, we validate that our generated images promisingly boosts the performance of HME recognition models, through data augmentation. Our code and results are available at: https://github.com/AiArt-HDU/HMEG." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 627, + 128, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 627, + 128, + 639 + ], + "spans": [ + { + "bbox": [ + 47, + 627, + 128, + 639 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 647, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 287, + 696 + ], + "type": "text", + "content": "Handwritten Mathematical Expressions (HMEs) are common and play significant roles in our daily life, especially in the research and education areas. HMEs generally present complex structures, serious deformations, and diverse writ" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 309, + 244, + 545, + 313 + ], + "blocks": [ + { + "bbox": [ + 309, + 244, + 545, + 313 + ], + "lines": [ + { + "bbox": [ + 309, + 244, + 545, + 313 + ], + "spans": [ + { + "bbox": [ + 309, + 244, + 545, + 313 + ], + "type": "image", + "image_path": "fcec784f4caba580b21f7fef86fdc9591e1206d1c2a25c48a2fbe88fb62c1301.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 308, + 316, + 545, + 388 + ], + "blocks": [ + { + "bbox": [ + 308, + 316, + 545, + 388 + ], + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 388 + ], + "type": "image", + "image_path": "00a7884c7514022f7c0d291821a5b25e0030444fd17480240baa6d776ccf0ea7.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 397, + 547, + 464 + ], + "lines": [ + { + "bbox": [ + 305, + 397, + 547, + 464 + ], + "spans": [ + { + "bbox": [ + 305, + 397, + 547, + 464 + ], + "type": "text", + "content": "Figure 1. Differences between (a) typical two-stage graph-to-image generation pipeline and (b) our end-to-end pipeline. In previous methods, the real masks are available as the input. In contrast, we propose a novel end-to-end pipeline of " + }, + { + "bbox": [ + 305, + 397, + 547, + 464 + ], + "type": "inline_equation", + "content": "\\text{graph} \\to \\text{layout} \\to \\text{mask} \\to \\text{image}" + }, + { + "bbox": [ + 305, + 397, + 547, + 464 + ], + "type": "text", + "content": ", and requires no real masks or non-differentiable alignment between layouts and masks." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 485, + 547, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 547, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 547, + 545 + ], + "type": "text", + "content": "ing styles. Such characteristics, along with data scarcity, make HME Recognition (HMER) a grand challenge in the OCR community. Despite the tremendous efforts that have been made to this task, the HMER performance is still unsatisfactory [14, 27, 62]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 546, + 547, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 547, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 547, + 653 + ], + "type": "text", + "content": "Recently, synthetic data augmentation has shown inspiring performance in various recognition tasks [23, 51]. This is mainly achieved by the remarkable progress of conditional image generation [21] in the past few years [58]. The advances allow for creating realistic images through texts [39, 57, 66], scene graphs [24, 60], layouts [28, 69], semantic masks [37, 72], sketches [61, 67], and more [12, 65]. However, there has been no effective generative model for generating high-quality HME images." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": "Some previous works synthesize HMEs by recomposing real online HMEs [46, 59]. Besides, FormulaGAN [41] generates HMEs from rendered images, via an Image-to-Image (I2I) Translation model. However, their generated images are limited in either diversity or realism. Be" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 702, + 219, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 702, + 219, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 702, + 219, + 713 + ], + "type": "text", + "content": "* Equal Contributions. † Corresponding Author." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15675" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": "sides, Mathematical Expressions (MEs) are highly structured data. Given an HME, its meaning and handwritten style, depend not only on the category and shape of each symbol, but also on the positional relations between symbols. It's unlikely to design an elaborate I2I translation model that can handle all these issues." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 146, + 287, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 146, + 287, + 314 + ], + "spans": [ + { + "bbox": [ + 46, + 146, + 287, + 314 + ], + "type": "text", + "content": "To combat this challenge, in this paper, we explore a new generation task, i.e. Handwritten Mathematical Expression Generation (HMEG) from symbol graphs of symbolic sequences. It's natural and easy to use a graph, specifying symbols and positional relations between symbols, to represent an ME [63]. Besides, previous works have demonstrated the significance of fully mining the structural information in boosting generalization [52]. By converting symbolic sequences to symbol graphs, we propose a novel HMEG method, inspired by the classic graph-to-image (G2I) generation work, " + }, + { + "bbox": [ + 46, + 146, + 287, + 314 + ], + "type": "inline_equation", + "content": "\\mathrm{Sg2im}" + }, + { + "bbox": [ + 46, + 146, + 287, + 314 + ], + "type": "text", + "content": " [24]. Similar to existing G2I generation works, our method consists of three stages: graph-to-query prediction, layout-to-mask transformation, and mask-to-image generation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 315, + 287, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 315, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 315, + 287, + 483 + ], + "type": "text", + "content": "Compared to the generation of natural images, the key challenge of HMEG is to generate unambiguous structural layouts, for accurately presenting meanings of formulas. From one hand, the requirement of layout clarity is much critical for generating formulas. Not only the size of each symbol, but also the positional relations between symbols, should be precisely predicted. Otherwise, mathematical meanings might change or become unrecognisable. For example, an input formula " + }, + { + "bbox": [ + 46, + 315, + 287, + 483 + ], + "type": "inline_equation", + "content": "x^{2}" + }, + { + "bbox": [ + 46, + 315, + 287, + 483 + ], + "type": "text", + "content": " might become " + }, + { + "bbox": [ + 46, + 315, + 287, + 483 + ], + "type": "inline_equation", + "content": "x_{2}" + }, + { + "bbox": [ + 46, + 315, + 287, + 483 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 315, + 287, + 483 + ], + "type": "inline_equation", + "content": "x2" + }, + { + "bbox": [ + 46, + 315, + 287, + 483 + ], + "type": "text", + "content": " in a generated image. From the other hand, MEs have infinite possible layouts and structures. Namely, a formula could comprise of arbitrary number of symbols and arbitrary relations between symbols. It's impossible to collect completed training data for learning the layout predictor." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 484, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 484, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 484, + 287, + 664 + ], + "type": "text", + "content": "To alleviate these difficulties, we first propose a Less-is-More (LiM) training strategy for learning an effective layout predictor. Specially, our layout predictor follows the structure of Graph Convolutional Networks (GCNs) [30]. During training, we only use minimal graphs with 1-degree connections, in stead of whole graphs with complex structures, to optimize the layout predictor. Note that the learned predictor can be applied to arbitrary symbol graphs, in the testing stage. By using LiM, the predictor emphasizes on local structures between adjacent symbols. Besides, the learning process is eased. Namely, we could learn an effective layout predictor by using moderate amount of training samples. Such benefits will be verified in the ablation study part. Besides, we add a layout discriminator to further boosts the realism of predicted layouts." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 288, + 714 + ], + "type": "text", + "content": "In addition, we propose a Sequential BBox-to-Mask Transformation (B2M) module, for refining layouts under weak supervision. In the original layout, bounding boxes of symbols tend to overlap with each other. Such overlapped" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": "areas may lead to overlapping symbols or chaotic strokes, in generated images. It's critical to refine coarse layouts to pixel-level soft masks, to alleviate the overlapping problems [22]. Most previous G2I generation works require real segmentation masks for learning refined layout masks [13, 22]. However, it's difficult to define the mask of a symbol, because white pixels around strokes are essential for representing a symbol. We have to refine the layout under weak supervision. To this end, we propose to learn a sequential mapping of layout " + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": " grid " + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": " mask, and use an subsequent image decoder to generate an HME image. During training, we use the image-related losses to train both B2M and the image decoder jointly." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 228, + 546, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 546, + 431 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 546, + 431 + ], + "type": "text", + "content": "In summary, we propose a novel end-to-end pipeline of graph " + }, + { + "bbox": [ + 304, + 228, + 546, + 431 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 228, + 546, + 431 + ], + "type": "text", + "content": " layout " + }, + { + "bbox": [ + 304, + 228, + 546, + 431 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 228, + 546, + 431 + ], + "type": "text", + "content": " mask " + }, + { + "bbox": [ + 304, + 228, + 546, + 431 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 228, + 546, + 431 + ], + "type": "text", + "content": " image for G2I generation tasks. Our model requires no supervision of real masks, or interactive alignment operations between layouts and masks. The major differences between our method and previous typical G2I generation works, are illustrated in Fig. 1. To verify the effectiveness of the proposed techniques, we conduct experiments on CROHME2014/2016/2019 [34]. Results show that, our model can generate high-quality HME images with clear layouts and recognisable symbols. Besides, our model significantly outperforms a number of existing generative methods, both qualitatively and quantitatively. A series of ablation study demonstrate the effectiveness of the proposed techniques, including the LiM, layout discriminator, and B2M. Finally, our generated HME images prove to boost the HMER performance by " + }, + { + "bbox": [ + 304, + 228, + 546, + 431 + ], + "type": "inline_equation", + "content": "5 - 11\\%" + }, + { + "bbox": [ + 304, + 228, + 546, + 431 + ], + "type": "text", + "content": " absolutely, through synthetic data augmentation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 442, + 397, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 442, + 397, + 454 + ], + "spans": [ + { + "bbox": [ + 306, + 442, + 397, + 454 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 462, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 689 + ], + "type": "text", + "content": "HME Recognition (HMER). Most advanced methods treat HMER as an image-to-markup problem [53]. They typically use Convolutional Neural Networks (CNNs) or Recurrent Neural Networks (RNNs) [18] to encode input images; and use attention-based models, such as Gated Recurrent Units (GRU) [5, 14, 68], or Transformers [2, 47, 70], to generate LaTeX sequences. Besides, some methods have been proposed [41, 46, 59]. Some recent works propose to use additional information, e.g. symbol-level counting [27], or relation-level counting [14], or emphasis on visually similar symbols [33], or synthetic data augmentation, to boost the HMER performance. Another branch of HMER are graph-based methods. These methods typically transform LaTeX sequences to Symbol Label Trees (SLTs) [63], and use tree-structured decoders to obtain the outputs [14, 52, 54]. These works validate the capacity of graphs in representing the structural information of formulas. In this paper, we therefore convert symbolic sequences to graphs, and use a GCN to encode them for generating HME images." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": "Graph-to-Image Generation. Image generation from scene graphs is first proposed in Sg2im [24]. The pipeline" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15676" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 547, + 198 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 547, + 198 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 547, + 198 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 547, + 198 + ], + "type": "image", + "image_path": "8c4dbbfe45917d3b66d51ac53881035d8b16b457cb7320288bfb3881619c192d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 142, + 205, + 449, + 216 + ], + "lines": [ + { + "bbox": [ + 142, + 205, + 449, + 216 + ], + "spans": [ + { + "bbox": [ + 142, + 205, + 449, + 216 + ], + "type": "text", + "content": "Figure 2. Overview of our handwritten mathematical expression generator (HMEG)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 237, + 288, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 237, + 288, + 321 + ], + "spans": [ + { + "bbox": [ + 46, + 237, + 288, + 321 + ], + "type": "text", + "content": "of Sg2im includes two stages: graph-to layout mapping and layout-to-image generation. Various subsequent methods have been proposed to boost the quality of generated images, following the same pipeline [9, 13, 16, 40]. Besides, some semi-parametric methods use crops from real images to boost the quality of generated images [31]. There are also some extensions to image manipulation [9, 42]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 324, + 287, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 324, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 324, + 287, + 479 + ], + "type": "text", + "content": "Most of these methods use a image discriminator and a object discriminator during training. Besides, they use real layouts, segmentation masks, and images, to supervise the learning process [13, 22]. COLoR [22] additionally uses a mask discriminator to boost the precision of generated masks. Recently, KCGM [55] uses knowledge consensus to alleviate the dependency on real layouts or masks. Besides, several diffusion model based methods [12, 60] can generate high-quality images directly from input scene graphs. However, diffusion based methods still suffer the high cost during both training and inference. Besides, generating recognisable texts is still challenging for most diffusion models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 483, + 287, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 287, + 686 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 287, + 686 + ], + "type": "text", + "content": "Layout-to-Image Generation. Most layout-to-image generation methods separately predict the mask of each object based on object embeddings, and require frequent resizing interaction between bounding boxes and masks [43, 44, 69]. To improve the quality of generated masks, several methods use masks discriminator [1, 28], and requires real masks during training. Besides, OC-GAN [45] uses both global and local similarities between scene graphs and images are used to improve the realism of generated layouts. Similar to our method, LAMA [32] tries to alleviate the overlapping between objects. However, in all these methods, the real bounding boxes are available as the input. They at most implement the end-to-end computing of layout " + }, + { + "bbox": [ + 46, + 483, + 287, + 686 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 483, + 287, + 686 + ], + "type": "text", + "content": " mask " + }, + { + "bbox": [ + 46, + 483, + 287, + 686 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 483, + 287, + 686 + ], + "type": "text", + "content": " image. In contrast, we propose a novel end-to-end computing of graph " + }, + { + "bbox": [ + 46, + 483, + 287, + 686 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 483, + 287, + 686 + ], + "type": "text", + "content": " layout " + }, + { + "bbox": [ + 46, + 483, + 287, + 686 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 483, + 287, + 686 + ], + "type": "text", + "content": " mask " + }, + { + "bbox": [ + 46, + 483, + 287, + 686 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 483, + 287, + 686 + ], + "type": "text", + "content": " image, and requires no interactive alignment operations between layouts and masks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "type": "text", + "content": "Handwritten Font Generation. Another related task is Handwritten Font Generation (HFG), which aims at" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 237, + 547, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 237, + 547, + 345 + ], + "spans": [ + { + "bbox": [ + 304, + 237, + 547, + 345 + ], + "type": "text", + "content": "transferring images of fonts to the target handwritten style [48, 50]. Most methods follow the I2I translation pipeline, and adopt encoder-decoder architectures [6]. Besides, great efforts have been made to solve the few-shot learning challenge [48, 64]. These methods focus on the style of strokes. In contrast, our work aims to generate rational layouts of multiple symbols, with complex relations, at the same time. It's interesting to explore generating HME images in personal styles in the future." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 358, + 361, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 358, + 361, + 370 + ], + "spans": [ + { + "bbox": [ + 306, + 358, + 361, + 370 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 378, + 373, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 378, + 373, + 389 + ], + "spans": [ + { + "bbox": [ + 306, + 378, + 373, + 389 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 397, + 545, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 397, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 397, + 545, + 601 + ], + "type": "text", + "content": "Our goal is to generate HME images conditioned on LaTeX sequences. To this end, we represent LaTeX sequences by symbol graphs [63] and design a novel G2I generation method to synthesize HME images. Our full model follows the idea of Generative Adversarial Networks (GANs) [21, 24]. It includes a generator, and several discriminators in the training stage. Fig. 2 shows the pipeline of our generator. It mainly include four parts: (i) First, the input LaTeX sequence is converted to a symbol graph, and embedded into high-dimensional feature vectors. (ii) A GCN based layout predictor estimates the coarse layout, i.e. bounding box (BBox) of each symbol. (iii) A sequential layout refinement module, including BBox-to-Grid mapping and Grid-to-Mask projection, generate a pixel-level soft mask for each symbol. (iv) The image decoder, a Cascaded Refinement Network (CRN) [4], generates an image conditioned on the predicted mask." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 610, + 462, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 610, + 462, + 623 + ], + "spans": [ + { + "bbox": [ + 306, + 610, + 462, + 623 + ], + "type": "text", + "content": "3.2. Symbol Graph Construction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "Given a set of mathematical symbol categories " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " and a set of relation categories " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": ", a symbol graph " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "(S,E)" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " can be constructed based on " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": ": each node " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "S = \\{s_1,\\dots ,s_n\\}" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " is a mathematical symbol, with " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "s_i\\in \\mathcal{C}" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "; and each directed edge " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "E = \\{e_{ij}\\}" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " represents the positional relation of node " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " w.r.t. node " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "e_{ij} = (s_i,r,s_j)" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "r\\in \\mathcal{R}" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " is the number of symbols in a mathematical" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15677" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 69, + 206, + 104 + ], + "blocks": [ + { + "bbox": [ + 137, + 69, + 206, + 104 + ], + "lines": [ + { + "bbox": [ + 137, + 69, + 206, + 104 + ], + "spans": [ + { + "bbox": [ + 137, + 69, + 206, + 104 + ], + "type": "image", + "image_path": "4f44b67ea813ce3720724f157f613465b666c53d36143b5da3335430baac276c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 150, + 287, + 174 + ], + "lines": [ + { + "bbox": [ + 46, + 150, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 46, + 150, + 287, + 174 + ], + "type": "text", + "content": "Figure 3. Architecture of our layout predictor and the Graph Embedding Model (GEM) [30]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 194, + 287, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 194, + 287, + 254 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 287, + 254 + ], + "type": "text", + "content": "expression. Let " + }, + { + "bbox": [ + 46, + 194, + 287, + 254 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 194, + 287, + 254 + ], + "type": "text", + "content": " denote the number of edges in " + }, + { + "bbox": [ + 46, + 194, + 287, + 254 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 46, + 194, + 287, + 254 + ], + "type": "text", + "content": ". For example, given an expression " + }, + { + "bbox": [ + 46, + 194, + 287, + 254 + ], + "type": "inline_equation", + "content": "x^2" + }, + { + "bbox": [ + 46, + 194, + 287, + 254 + ], + "type": "text", + "content": ", it can be represented by a tuple " + }, + { + "bbox": [ + 46, + 194, + 287, + 254 + ], + "type": "inline_equation", + "content": "(x, \\sup, 2)" + }, + { + "bbox": [ + 46, + 194, + 287, + 254 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 194, + 287, + 254 + ], + "type": "inline_equation", + "content": "\\sup" + }, + { + "bbox": [ + 46, + 194, + 287, + 254 + ], + "type": "text", + "content": " is short for superscript. Besides, we add an additional edge type None to indicate that the corresponding pair of symbols are unrelated." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 254, + 287, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 254, + 287, + 302 + ], + "spans": [ + { + "bbox": [ + 46, + 254, + 287, + 302 + ], + "type": "text", + "content": "Afterwards, we map the nodes and edges " + }, + { + "bbox": [ + 46, + 254, + 287, + 302 + ], + "type": "inline_equation", + "content": "(S,E)" + }, + { + "bbox": [ + 46, + 254, + 287, + 302 + ], + "type": "text", + "content": " to high-dimensional feature vectors " + }, + { + "bbox": [ + 46, + 254, + 287, + 302 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{V},\\mathbf{U}\\}" + }, + { + "bbox": [ + 46, + 254, + 287, + 302 + ], + "type": "text", + "content": " by using a learnable embedding layer, respectively [30]. The learned embedding feature vectors are denoted by:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 312, + 287, + 326 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 312, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 77, + 312, + 287, + 326 + ], + "type": "interline_equation", + "content": "\\mathbf {V} = \\left\\{\\mathbf {v} _ {i}, \\forall i \\in S \\right\\}, \\mathbf {U} = \\left\\{\\mathbf {u} _ {i j}, \\forall (i, j) \\in E \\right\\}, \\tag {1}", + "image_path": "ff46c2562916c145b9c0addb59aa79ef339a800596ce31ab11db40f06c869fb3.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 334, + 287, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 334, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 46, + 334, + 287, + 358 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 334, + 287, + 358 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i, \\mathbf{u}_{ij} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 46, + 334, + 287, + 358 + ], + "type": "text", + "content": " associate with " + }, + { + "bbox": [ + 46, + 334, + 287, + 358 + ], + "type": "inline_equation", + "content": "s_i, e_{ij}" + }, + { + "bbox": [ + 46, + 334, + 287, + 358 + ], + "type": "text", + "content": ", respectively; " + }, + { + "bbox": [ + 46, + 334, + 287, + 358 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 334, + 287, + 358 + ], + "type": "text", + "content": " is the dimension of embedding vectors." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 365, + 212, + 378 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 365, + 212, + 378 + ], + "spans": [ + { + "bbox": [ + 47, + 365, + 212, + 378 + ], + "type": "text", + "content": "3.3. Adversarial Layout Prediction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 384, + 287, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 287, + 479 + ], + "type": "text", + "content": "Inspired by GAN [21], we train our layout prediction module in an adversarial learning manner. Specially, we use an layout predictor " + }, + { + "bbox": [ + 46, + 384, + 287, + 479 + ], + "type": "inline_equation", + "content": "G_{box}" + }, + { + "bbox": [ + 46, + 384, + 287, + 479 + ], + "type": "text", + "content": " to generate a layout in the format of bounding boxes, conditioned on a given symbol graph; and use an additional layout discriminator " + }, + { + "bbox": [ + 46, + 384, + 287, + 479 + ], + "type": "inline_equation", + "content": "D_{box}" + }, + { + "bbox": [ + 46, + 384, + 287, + 479 + ], + "type": "text", + "content": " to judge whether a layout is real or fake. Using the adversarial learning strategy like GAN, the layout predictor would finally predict high-quality bounding boxes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 479, + 287, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 287, + 589 + ], + "type": "text", + "content": "Layout Predictor. We first use the Graph Embedding Model (GEM) [30] to update features of nodes and edges, and then use a MLP to predict the BBox indices of each symbol. First, original embedding features are encoded by: " + }, + { + "bbox": [ + 46, + 479, + 287, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_i^{(0)} = f_{\\mathrm{node}}(\\mathbf{v}_i), \\forall i \\in S" + }, + { + "bbox": [ + 46, + 479, + 287, + 589 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 479, + 287, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_{i,j} = f_{\\mathrm{edge}}(\\mathbf{u}_{ij}), \\forall (i,j) \\in E" + }, + { + "bbox": [ + 46, + 479, + 287, + 589 + ], + "type": "text", + "content": ". Afterward, node features are updated through a stack of propagation layers, by accumulating information in local neighborhood. Features of node " + }, + { + "bbox": [ + 46, + 479, + 287, + 589 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 479, + 287, + 589 + ], + "type": "text", + "content": " at the " + }, + { + "bbox": [ + 46, + 479, + 287, + 589 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 479, + 287, + 589 + ], + "type": "text", + "content": "-th layer are updated by:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 75, + 597, + 261, + 614 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 597, + 261, + 614 + ], + "spans": [ + { + "bbox": [ + 75, + 597, + 261, + 614 + ], + "type": "interline_equation", + "content": "\\mathbf {m} _ {j \\rightarrow i} = f _ {\\text {m e s s a g e}} \\left(\\mathbf {h} _ {i} ^ {(t)}, \\mathbf {h} _ {j} ^ {(t)}, \\mathbf {e} _ {i j}\\right), \\forall (i, j) \\in E,", + "image_path": "c8b307ff26ae72f0642f495e558cd4eebe60da0e1e6eeed545d3cc5326e76de1.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 73, + 615, + 287, + 642 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 615, + 287, + 642 + ], + "spans": [ + { + "bbox": [ + 73, + 615, + 287, + 642 + ], + "type": "interline_equation", + "content": "\\mathbf {h} _ {i} ^ {(t + 1)} = f _ {\\text {n o d e}} \\left(\\mathbf {h} _ {i} ^ {(t)}, \\sum_ {j} \\mathbf {m} _ {j \\rightarrow i}\\right). \\tag {2}", + "image_path": "406721dbdead40907ee7b0f8c57c74c5ab2cdba4a6f412352da2c36a3f1eba9d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 651, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 651, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 47, + 651, + 287, + 676 + ], + "type": "text", + "content": "After " + }, + { + "bbox": [ + 47, + 651, + 287, + 676 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 47, + 651, + 287, + 676 + ], + "type": "text", + "content": " rounds of propagations, the model produces an aggregated graph level representation by:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 684, + 287, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 684, + 287, + 716 + ], + "spans": [ + { + "bbox": [ + 56, + 684, + 287, + 716 + ], + "type": "interline_equation", + "content": "\\mathbf {h} _ {G} = f _ {G} \\left(\\sum_ {i} \\sigma \\left(f _ {\\text {g a t e}} \\left(\\mathbf {h} _ {i} ^ {(T)}\\right)\\right) \\odot f _ {\\text {u p d a t e}} \\left(\\mathbf {h} _ {i} ^ {(T)}\\right)\\right), \\tag {3}", + "image_path": "d9a3a90584fb11118edce1a7785913510d953d2953482cb7bf7e5eb337173842.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "text", + "content": " denotes the Sigmoid operation. Finally, we integrate the updated embedding features and the graph level features for predicting the BBox indices of each symbol. Besides, we concatenate a Gaussian noise vector " + }, + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i\\in \\mathbb{R}^k,\\forall i\\in S" + }, + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "text", + "content": " to these features, to improve the diversity in predicted layouts. The BBox prediction is formulated by:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 354, + 153, + 545, + 168 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 153, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 354, + 153, + 545, + 168 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {b}} _ {i} = f _ {\\mathrm {b b o x}} \\left(\\mathbf {h} _ {i} ^ {(T)}, \\mathbf {h} _ {G}, \\mathbf {z} _ {i}\\right), \\forall i \\in S. \\tag {4}", + "image_path": "64e11f492c14ef95efaa4cf0fa696fe49e0e3584a323e0e027e2e7253a51e990.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 177, + 545, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 177, + 545, + 236 + ], + "spans": [ + { + "bbox": [ + 304, + 177, + 545, + 236 + ], + "type": "text", + "content": "Each of the mapping functions " + }, + { + "bbox": [ + 304, + 177, + 545, + 236 + ], + "type": "inline_equation", + "content": "f_{*}" + }, + { + "bbox": [ + 304, + 177, + 545, + 236 + ], + "type": "text", + "content": " above is an MLP on the inputs [29, 30]. For each symbol, we obtain a 4-dimensional rectangular box " + }, + { + "bbox": [ + 304, + 177, + 545, + 236 + ], + "type": "inline_equation", + "content": "[x_0, y_0, x_1, y_1]" + }, + { + "bbox": [ + 304, + 177, + 545, + 236 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 304, + 177, + 545, + 236 + ], + "type": "inline_equation", + "content": "[0,1]" + }, + { + "bbox": [ + 304, + 177, + 545, + 236 + ], + "type": "text", + "content": " coordinate space, denoting the coordinates of the upper left and lower right corners of the BBox border." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 237, + 545, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 237, + 545, + 320 + ], + "spans": [ + { + "bbox": [ + 304, + 237, + 545, + 320 + ], + "type": "text", + "content": "Layout Discriminator. The layout discriminator " + }, + { + "bbox": [ + 304, + 237, + 545, + 320 + ], + "type": "inline_equation", + "content": "D_{box}" + }, + { + "bbox": [ + 304, + 237, + 545, + 320 + ], + "type": "text", + "content": " is also composed of GEM [30], followed by an MLP to predict whether an input layout as real or fake. " + }, + { + "bbox": [ + 304, + 237, + 545, + 320 + ], + "type": "inline_equation", + "content": "D_{box}" + }, + { + "bbox": [ + 304, + 237, + 545, + 320 + ], + "type": "text", + "content": " takes concatenations of the symbol embedding matrix " + }, + { + "bbox": [ + 304, + 237, + 545, + 320 + ], + "type": "inline_equation", + "content": "\\mathbf{V} \\in \\mathbf{R}^{n \\times d}" + }, + { + "bbox": [ + 304, + 237, + 545, + 320 + ], + "type": "text", + "content": " and the layout indices " + }, + { + "bbox": [ + 304, + 237, + 545, + 320 + ], + "type": "inline_equation", + "content": "B \\in \\mathbf{R}^{n \\times 4}" + }, + { + "bbox": [ + 304, + 237, + 545, + 320 + ], + "type": "text", + "content": " as input. In the implementation, we use the least square version of adversarial loss, i.e." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 319, + 329, + 545, + 344 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 329, + 545, + 344 + ], + "spans": [ + { + "bbox": [ + 319, + 329, + 545, + 344 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {a d v} ^ {b o x} = \\left\\| D _ {b o x} (\\mathbf {V}, B) \\right\\| _ {2} ^ {2} + \\left\\| 1 - D _ {b o x} (\\mathbf {V}, \\hat {B}) \\right\\| _ {2} ^ {2}. \\tag {5}", + "image_path": "46ddd8b42040b1f22ab2127506037202625bbcc1c3bbdee0b37baa3be0810f96.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 352, + 545, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 352, + 545, + 412 + ], + "spans": [ + { + "bbox": [ + 304, + 352, + 545, + 412 + ], + "type": "text", + "content": "Besides, we use the L2 distance between the predicted BBox indices " + }, + { + "bbox": [ + 304, + 352, + 545, + 412 + ], + "type": "inline_equation", + "content": "\\hat{B}" + }, + { + "bbox": [ + 304, + 352, + 545, + 412 + ], + "type": "text", + "content": " and the target (real) indices " + }, + { + "bbox": [ + 304, + 352, + 545, + 412 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 352, + 545, + 412 + ], + "type": "text", + "content": " as the layout reconstruction loss, i.e. " + }, + { + "bbox": [ + 304, + 352, + 545, + 412 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{box} = ||B - \\hat{B}||_1" + }, + { + "bbox": [ + 304, + 352, + 545, + 412 + ], + "type": "text", + "content": ". We use a weighted sum of " + }, + { + "bbox": [ + 304, + 352, + 545, + 412 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{box}" + }, + { + "bbox": [ + 304, + 352, + 545, + 412 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 352, + 545, + 412 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{adv}^{box}" + }, + { + "bbox": [ + 304, + 352, + 545, + 412 + ], + "type": "text", + "content": " as the objective of " + }, + { + "bbox": [ + 304, + 352, + 545, + 412 + ], + "type": "inline_equation", + "content": "G_{box}" + }, + { + "bbox": [ + 304, + 352, + 545, + 412 + ], + "type": "text", + "content": ". During training, " + }, + { + "bbox": [ + 304, + 352, + 545, + 412 + ], + "type": "inline_equation", + "content": "G_{box}" + }, + { + "bbox": [ + 304, + 352, + 545, + 412 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 352, + 545, + 412 + ], + "type": "inline_equation", + "content": "D_{box}" + }, + { + "bbox": [ + 304, + 352, + 545, + 412 + ], + "type": "text", + "content": " are alternatively optimized." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 413, + 545, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 545, + 592 + ], + "type": "text", + "content": "Less-is-More Learning (LiM) Strategy. A mathematical expression could include arbitrary number of symbols or relations between symbols. In other words, there are infinite possible structures of symbol graphs. It's difficult to collect a complete training set. Besides, using complicated graphs during training may add the difficulty in learning an effective model. We therefore propose a Less-is-More (LiM) learning strategy [3]. Specially, we propose to use symbol graphs with merely 1-degree connections during training. In this way, the layout predictor is expected to emphasize on local structures, i.e. generating rational layout for locally connected symbols. While in the testing stage, the learned model can be applied to arbitrary symbol graphs, with arbitrary degrees of connections, and predict the corresponding layout." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 599, + 525, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 599, + 525, + 612 + ], + "spans": [ + { + "bbox": [ + 306, + 599, + 525, + 612 + ], + "type": "text", + "content": "3.4. Sequential BBox-to-Mask Transformation" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 618, + 545, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 654 + ], + "type": "text", + "content": "We next predict the pixel-level soft mask of an expression based on the predicted bounding boxes. This procedure is illustrated in Fig. 4, and consists of the following two steps:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "(1) BBox-to-Grid Mapping. We first transform the BBox indices to layout grids, as illustrated in Fig. 4. To this end, we subscribe each index value from the coordinate matrices " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " in the [0, 1] coordinate space. For each symbol, we obtain a " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "4 \\times w \\times w" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " tensor of grid." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15678" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 286, + 162 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 286, + 162 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 286, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 286, + 162 + ], + "type": "image", + "image_path": "028bf7dcbd64b165c184a53c78cc2a1f474979fb3652d96f64e40062449abbb4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 169, + 287, + 191 + ], + "lines": [ + { + "bbox": [ + 47, + 169, + 287, + 191 + ], + "spans": [ + { + "bbox": [ + 47, + 169, + 287, + 191 + ], + "type": "text", + "content": "Figure 4. Illustration of the sequential BBox-to-Mask transformation (B2M) module." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 213, + 287, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 213, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 47, + 213, + 287, + 249 + ], + "type": "text", + "content": "(2) Grid-to-Mask Projection. Afterwards, we use a FCN with a Sigmoid output layer (denoted by " + }, + { + "bbox": [ + 47, + 213, + 287, + 249 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{mask}}" + }, + { + "bbox": [ + 47, + 213, + 287, + 249 + ], + "type": "text", + "content": ") to estimate a one-channel mask for each symbol, i.e." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 52, + 258, + 287, + 271 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 258, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 52, + 258, + 287, + 271 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {i} = \\phi_ {\\text {m a s k}} (\\mathbf {X} - x _ {i, 0}, \\mathbf {X} - x _ {i, 1}, \\mathbf {Y} - y _ {i, 0}, \\mathbf {Y} - y _ {i, 1}). \\tag {6}", + "image_path": "68b3dfb3d79b07e88a537badebb9af5ce2034041dd3b00cbdce5a90a16659942.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 279, + 287, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 279, + 287, + 363 + ], + "spans": [ + { + "bbox": [ + 47, + 279, + 287, + 363 + ], + "type": "text", + "content": "Each value in " + }, + { + "bbox": [ + 47, + 279, + 287, + 363 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_i" + }, + { + "bbox": [ + 47, + 279, + 287, + 363 + ], + "type": "text", + "content": " indicates the probability of a pixel belonging to symbol " + }, + { + "bbox": [ + 47, + 279, + 287, + 363 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 47, + 279, + 287, + 363 + ], + "type": "text", + "content": ". Specially, we first use two Convolutional layers to predict a low-resolution mask of dimension " + }, + { + "bbox": [ + 47, + 279, + 287, + 363 + ], + "type": "inline_equation", + "content": "w \\times w" + }, + { + "bbox": [ + 47, + 279, + 287, + 363 + ], + "type": "text", + "content": ", and then use two upsampling Convolutional layers to sequentially upsampling the mask by a factor of 2. These masks are used in the image decoder for generating an image of the same size, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 363, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 363, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 363, + 287, + 460 + ], + "type": "text", + "content": "Note: Previous G2I generation works [9] also consists of BBox-to-Grid mapping, but they merely use the grids to select patches of objects. Besides, they predict the mask from the BBox indices " + }, + { + "bbox": [ + 46, + 363, + 287, + 460 + ], + "type": "inline_equation", + "content": "\\hat{B}" + }, + { + "bbox": [ + 46, + 363, + 287, + 460 + ], + "type": "text", + "content": " or original embedding features " + }, + { + "bbox": [ + 46, + 363, + 287, + 460 + ], + "type": "inline_equation", + "content": "(\\mathbf{V},\\mathbf{U})" + }, + { + "bbox": [ + 46, + 363, + 287, + 460 + ], + "type": "text", + "content": ". Differently, we use the transformed grids for layout mask prediction. Moreover, since the whole BBox-to-Mask prediction module is differentiable, our whole model enables training in an end-to-end manner." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 466, + 155, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 466, + 155, + 479 + ], + "spans": [ + { + "bbox": [ + 47, + 466, + 155, + 479 + ], + "type": "text", + "content": "3.5. Image Generation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 485, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 485, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 485, + 287, + 628 + ], + "type": "text", + "content": "Image Decoder. In the implementation, we adopt CRN [4] as the image decoder. The input is a multiplication of the whole layout mask " + }, + { + "bbox": [ + 46, + 485, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{M} \\in \\mathbb{R}^{n \\times w \\times w}" + }, + { + "bbox": [ + 46, + 485, + 287, + 628 + ], + "type": "text", + "content": " with the embedding vectors " + }, + { + "bbox": [ + 46, + 485, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{V} \\in \\mathbb{R}^{n \\times d}" + }, + { + "bbox": [ + 46, + 485, + 287, + 628 + ], + "type": "text", + "content": ". Besides, we add Gaussian noise to the feature maps to improve the diversity in image generation. CRN includes three upsampling Convolutional layers, and outputs a formula image with resolutions of " + }, + { + "bbox": [ + 46, + 485, + 287, + 628 + ], + "type": "inline_equation", + "content": "w \\times w" + }, + { + "bbox": [ + 46, + 485, + 287, + 628 + ], + "type": "text", + "content": ". Recall that we actually predicts three masks at different scales. In the implementation, we have three generated images, at the resolution of " + }, + { + "bbox": [ + 46, + 485, + 287, + 628 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 46, + 485, + 287, + 628 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 485, + 287, + 628 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 46, + 485, + 287, + 628 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 485, + 287, + 628 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 485, + 287, + 628 + ], + "type": "text", + "content": ", respectively. All these images are used to calculate the losses for optimizing our model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 628, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 628, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 628, + 287, + 713 + ], + "type": "text", + "content": "Image and Symbol Discriminators Finally, we use an image discriminator " + }, + { + "bbox": [ + 46, + 628, + 287, + 713 + ], + "type": "inline_equation", + "content": "D_{img}" + }, + { + "bbox": [ + 46, + 628, + 287, + 713 + ], + "type": "text", + "content": " and a symbol discriminator " + }, + { + "bbox": [ + 46, + 628, + 287, + 713 + ], + "type": "inline_equation", + "content": "D_{sym}" + }, + { + "bbox": [ + 46, + 628, + 287, + 713 + ], + "type": "text", + "content": " during training, to boost the realism of the generated whole image " + }, + { + "bbox": [ + 46, + 628, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\hat{I}" + }, + { + "bbox": [ + 46, + 628, + 287, + 713 + ], + "type": "text", + "content": " and patches of symbols, respectively. We use the grids to select patches of symbols, and input them into " + }, + { + "bbox": [ + 46, + 628, + 287, + 713 + ], + "type": "inline_equation", + "content": "D_{sym}" + }, + { + "bbox": [ + 46, + 628, + 287, + 713 + ], + "type": "text", + "content": ". Both discriminators are fully CNNs with the cross-entropy (CE) version of adversarial loss. In addition," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "text", + "content": "following AC-GAN [36], we have the symbol discriminator additionally predict category labels of symbols. In this way, the generated symbols would be more recognizable and real." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 120, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 120, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 120, + 545, + 156 + ], + "type": "text", + "content": "Loss Functions. Similar to previous G2I generation works [24], we use a series of loss functions to optimize the image generation. The losses are briefly introduced below." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 159, + 545, + 304 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 306, + 159, + 545, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 159, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 306, + 159, + 545, + 195 + ], + "type": "text", + "content": "- Pixel loss " + }, + { + "bbox": [ + 306, + 159, + 545, + 195 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pix}" + }, + { + "bbox": [ + 306, + 159, + 545, + 195 + ], + "type": "text", + "content": ": First, we use the L1 distance between a generated image and the target formula image and as the pixel loss: " + }, + { + "bbox": [ + 306, + 159, + 545, + 195 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pix} = ||I - \\hat{I}||_1" + }, + { + "bbox": [ + 306, + 159, + 545, + 195 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 196, + 545, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 196, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 306, + 196, + 545, + 232 + ], + "type": "text", + "content": "- Image adversarial loss " + }, + { + "bbox": [ + 306, + 196, + 545, + 232 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{adv}^{img}" + }, + { + "bbox": [ + 306, + 196, + 545, + 232 + ], + "type": "text", + "content": ": Second, we use the global image discriminant loss from " + }, + { + "bbox": [ + 306, + 196, + 545, + 232 + ], + "type": "inline_equation", + "content": "D_{img}" + }, + { + "bbox": [ + 306, + 196, + 545, + 232 + ], + "type": "text", + "content": ", to encourage the generated HME images in the real handwritten style." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 232, + 545, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 232, + 545, + 267 + ], + "spans": [ + { + "bbox": [ + 306, + 232, + 545, + 267 + ], + "type": "text", + "content": "- Symbol adversarial loss " + }, + { + "bbox": [ + 306, + 232, + 545, + 267 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{adv}^{sym}" + }, + { + "bbox": [ + 306, + 232, + 545, + 267 + ], + "type": "text", + "content": ": Third, we use the symbol discriminative loss from " + }, + { + "bbox": [ + 306, + 232, + 545, + 267 + ], + "type": "inline_equation", + "content": "D_{sym}" + }, + { + "bbox": [ + 306, + 232, + 545, + 267 + ], + "type": "text", + "content": ", to improve the fidelity of generated symbols." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 268, + 545, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 268, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 306, + 268, + 545, + 304 + ], + "type": "text", + "content": "- Auxiliary Symbol Recognition Loss " + }, + { + "bbox": [ + 306, + 268, + 545, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{aux}^{sym}" + }, + { + "bbox": [ + 306, + 268, + 545, + 304 + ], + "type": "text", + "content": ": Finally, we use the auxiliary symbol recognition loss from " + }, + { + "bbox": [ + 306, + 268, + 545, + 304 + ], + "type": "inline_equation", + "content": "D_{sym}" + }, + { + "bbox": [ + 306, + 268, + 545, + 304 + ], + "type": "text", + "content": ", to enforce the generator producing recognizable symbols." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 307, + 545, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 307, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 307, + 545, + 342 + ], + "type": "text", + "content": "We use a combination of these loss functions above, to train the B2M module and the image generation modules together. The total loss is formulated by:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 318, + 352, + 545, + 366 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 352, + 545, + 366 + ], + "spans": [ + { + "bbox": [ + 318, + 352, + 545, + 366 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {i m g} = \\lambda_ {1} \\mathcal {L} _ {p i x} + \\lambda_ {2} \\mathcal {L} _ {a d v} ^ {i m g} + \\lambda_ {3} \\mathcal {L} _ {a d v} ^ {s y m} + \\lambda_ {4} \\mathcal {L} _ {a u x} ^ {s y m}, \\tag {7}", + "image_path": "20b8f3e9410276a37550cbf07effc88d20a06dde6f2930018d47020ba0a11805.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 375, + 545, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 375, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 375, + 545, + 399 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 375, + 545, + 399 + ], + "type": "inline_equation", + "content": "\\lambda_{i}, i = 1, \\ldots, 4" + }, + { + "bbox": [ + 304, + 375, + 545, + 399 + ], + "type": "text", + "content": " are weighting factors and are set as 1, 0.01, 0.01, 0.001, respectively." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 411, + 388, + 425 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 411, + 388, + 425 + ], + "spans": [ + { + "bbox": [ + 306, + 411, + 388, + 425 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 431, + 432, + 444 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 431, + 432, + 444 + ], + "spans": [ + { + "bbox": [ + 306, + 431, + 432, + 444 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 449, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 569 + ], + "type": "text", + "content": "Data. We use CROHME2014/2016/2019 [34] to validate our HMEG method in terms of both generalization and recognition. The dataset collects pen trajectories, and labels them with symbols and symbol categories. The dataset contains 126 different symbol categories, and 8 positional relations, i.e. start, left superscript, superscript, subscript, below, above, right and end. By restoring the pen trajectory information, we can easily obtain the HME image and the BBox of each symbol. We use the standard partition of each dataset through all the experiments." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "Implementation Details. In the implementation, due to the limitation of our computing resources, we train our model in two stages: we first train the layout prediction part in the first stage, and then train the rest parts by fixing it. In the first stage, we train the model with a batch size 64 and a learning rate 5e5, for 40,000 iterations. In each iteration, we random sample a batch of 1-degree sub-graphs from the training set for learning the layout predictor. In the second stage, we use a batch size 8, a learning rate 1e4, and train for 600,000 iterations. Our codes are implemented by using Pytorch and a NVIDIA TITAN XP GPU. We use Adam [26] as the optimizer for all networks." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15679" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 71, + 545, + 320 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 545, + 320 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 545, + 320 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 545, + 320 + ], + "type": "table", + "html": "
Print|ax0+by0+c/√a2+b2|p=√a2+b2-2abcosA1/2(a+1/b) = 2ab/a+b-b+√b2-4ac/2a∫0πcos(θ/2)dθx/a+x/b-x/c∫bdIπr2h/3sin(π/3)=1/2
CycleGANaxn+byn+c/√n2+b2|p=√n2+b2-2abcosA1/2(a+1/b) = 2ab/a+b-b+√b2-4ac/2a∫0πcos(θ/2)dθx/a+x/b-x/c∫bdIπr2h/3sin(π/3)=1/2
FormulaGAN|ax0+by0+c/√a2+b2|p=√a2+b2-2abcosA1/2(a+1/b) = 2ab/a+b-b+√b2-4ac/2a∫0πcos(θ/2)dθx/a+x/b-x/c∫bdIπr2h/3sin(π/3)=1/2
Sg2imaxn+byn+c/√a2+b2|p=√a2+b2-2abcosA1/2(a+1/b) = 2ab/a+b-b+√b2-4ac/2a∫0πcos(θ/2)dθx/a+x/b-x/c∫bdIπr2h/3sin(π/3)=1/2
Ours|ax0+by0+c/√a2+b2|p=√a2+b2-2abcosA1/2(a+1/b) = 2ab/a+b-b+√b2-4ac/2a∫0πcos(θ/2)dθx/a+x/b-x/c∫bdIπr2h/3sin(π/3)=1/2
", + "image_path": "613af9732d1422f27ff311e08036d68b141f2d2ec5ce7038b53bc23708f78a0c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 63, + 330, + 529, + 342 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 330, + 529, + 342 + ], + "spans": [ + { + "bbox": [ + 63, + 330, + 529, + 342 + ], + "type": "text", + "content": "Figure 5. Handwritten mathematical expressions generated by CycleGAN [71], FormulaGAN [41], Sg2im [24], and our method." + } + ] + } + ], + "index": 1, + "type": "text" + }, + { + "bbox": [ + 46, + 362, + 289, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 362, + 289, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 362, + 289, + 518 + ], + "type": "text", + "content": "Criteria. We choose two image quality assessment indices as the criteria. The former includes the Fréchet Inception distance (FID) [17] and Structural Similarity Measure (SSIM) [49]. Lower values of FID indicate better realism of synthesised images; while greater values of SSIM generally indicate higher similarity between a synthesized image and the corresponding real image. We here report the average SSIM value of all the test samples. In addition, we use two HMER indices, including the Word Error Rate (WER) and the Expression Recognition Rate (ExpRate). Here we use an open-source HMER method [19], to recognize HME images and compute the corresponding indices. Lower WER and higher ExpRate values indicate better performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 526, + 187, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 526, + 187, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 526, + 187, + 540 + ], + "type": "text", + "content": "4.2. Comparison with SOTAs" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 545, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 605 + ], + "type": "text", + "content": "We first compare with the existing GAN based HMEG method, FormulaGAN [41], our baseline, Sg2im [24], and the widely used unsupervised I2I translation method, CycleGAN [71], on CHROME2019. We train and test these methods following the same experimental settings as ours." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "Qualitative Comparison. Fig. 5 shows HME images generated by different models. CycleGAN fails to transform input formulas to the handwritten style. FormulaGAN produces handwritten-style images, but with blurring and unnatural strokes. In contrast, both Sg2im and our method generate high quality images, with realistic styles and clear strokes. In addition, the images generated by our method present better layouts than those by Sg2im. Sg2im sometimes produce very small symbols, e.g. " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "y_{0}" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " in the first col" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 307, + 380, + 550, + 449 + ], + "blocks": [ + { + "bbox": [ + 309, + 361, + 541, + 372 + ], + "lines": [ + { + "bbox": [ + 309, + 361, + 541, + 372 + ], + "spans": [ + { + "bbox": [ + 309, + 361, + 541, + 372 + ], + "type": "text", + "content": "Table 1. Comparison with existing methods on CHROME2019." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 380, + 550, + 449 + ], + "lines": [ + { + "bbox": [ + 307, + 380, + 550, + 449 + ], + "spans": [ + { + "bbox": [ + 307, + 380, + 550, + 449 + ], + "type": "table", + "html": "
SSIM↑FID ↓WER ↓ExpRate ↑
CycleGAN [71]0.75784.140.6710.026
FormulaGAN [41]0.72474.680.6010.066
Sg2im [24]0.78710.020.3930.219
Ours0.79310.980.3260.316
", + "image_path": "715defda5b99a43211912c1a82eacb91b5b14817598596f721e218481c882fa5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 472, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 545, + 593 + ], + "type": "text", + "content": "unn and " + }, + { + "bbox": [ + 304, + 472, + 545, + 593 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 472, + 545, + 593 + ], + "type": "text", + "content": " in the second column. Besides, Sg2im may produce overlapped symbols, e.g. " + }, + { + "bbox": [ + 304, + 472, + 545, + 593 + ], + "type": "inline_equation", + "content": "\\frac{1}{2} \\left( \\frac{1}{a} + \\frac{1}{b} \\right)" + }, + { + "bbox": [ + 304, + 472, + 545, + 593 + ], + "type": "text", + "content": " in the third column. Finally, Sg2im produces more unrecognisable symbols than ours, e.g. Sin and " + }, + { + "bbox": [ + 304, + 472, + 545, + 593 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 304, + 472, + 545, + 593 + ], + "type": "text", + "content": " in last column. In contrast, the symbols generated by our method present reasonable sizes and clear structures. Besides, symbols are reasonably placed in relation to each other, neither overlapping nor far apart. This is similar to real formulas written by humans. Such observations demonstrate the superiority of our method in generating both high-quality layouts and HME images." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "content": "Quantitative Comparison. Table 1 shows the quantitative performance indices. Both Sg2im and our method achieves significant better indices than CycleGAN or FormulaGAN. This is consistent with the qualitative comparison results. According to SSIM and FID, Sg2im and our method almost perform equality in term of image quality. However, our method decreases the WER by 6.5 percent and improves the ExpRate by 10 percent, absolutely. Such distinct superiority in recognition implies that, the handwritten expressions generated by our method are signifi" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15680" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 69, + 547, + 222 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 547, + 222 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 547, + 222 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 547, + 222 + ], + "type": "image", + "image_path": "b6559d3ca4776bbde12b0e6e0f99be7f8555c3f5b379b2742eb80994e2df6aff.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 227, + 546, + 243 + ], + "lines": [ + { + "bbox": [ + 46, + 227, + 546, + 243 + ], + "spans": [ + { + "bbox": [ + 46, + 227, + 546, + 243 + ], + "type": "text", + "content": "Figure 6. Illustration of the layouts and formula images generated by model variants, in the ablation study. The input formula is " + }, + { + "bbox": [ + 46, + 227, + 546, + 243 + ], + "type": "inline_equation", + "content": "\\frac{1}{25} y^2 -\\frac{8}{25} y" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 257, + 287, + 345 + ], + "blocks": [ + { + "bbox": [ + 50, + 257, + 287, + 345 + ], + "lines": [ + { + "bbox": [ + 50, + 257, + 287, + 345 + ], + "spans": [ + { + "bbox": [ + 50, + 257, + 287, + 345 + ], + "type": "image", + "image_path": "8206013411e259c7a821c90375dc5cb86609c37dc2c8a76253dafe39a0ceaac2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 353, + 288, + 419 + ], + "lines": [ + { + "bbox": [ + 46, + 353, + 288, + 419 + ], + "spans": [ + { + "bbox": [ + 46, + 353, + 288, + 419 + ], + "type": "text", + "content": "Figure 7. Comparison with diffusion models. DiffSketcher and DiffAE generate HMEs conditioned on print formulas; while ControlNet and SD+LoRA take both the print formula and a textual prompt (\"a handwritten mathematical expression of latex code\") as input. We use official diffusion models, and fine-tune DiffAE and SD+LoRA using HMEs." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 442, + 281, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 442, + 281, + 454 + ], + "spans": [ + { + "bbox": [ + 46, + 442, + 281, + 454 + ], + "type": "text", + "content": "cantly better than " + }, + { + "bbox": [ + 46, + 442, + 281, + 454 + ], + "type": "inline_equation", + "content": "\\mathrm{Sg2im}" + }, + { + "bbox": [ + 46, + 442, + 281, + 454 + ], + "type": "text", + "content": ", in terms of clarity and structure." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 455, + 287, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 455, + 287, + 575 + ], + "spans": [ + { + "bbox": [ + 46, + 455, + 287, + 575 + ], + "type": "text", + "content": "Comparison with Diffusion Models. We additionally compare with ControlNet [67], Stable Diffusion (SD) [39] + LoRA [20], DiffSketcher [56], and DiffAE [38]. As shown in Fig. 7, these diffusion models cannot generate high-quality HMEs with realistic handwritten styles or recognizable symbols. Besides, their computational complexity is much heavier than ours. Some recent methods, e.g. ChiroDiff [8] and SDT [7], can only generate handwritten images of one single character. Such comparison results demonstrate the advantage of our method in HMEG." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 585, + 141, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 585, + 141, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 585, + 141, + 597 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 604, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 604, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 604, + 287, + 640 + ], + "type": "text", + "content": "We further analyse the impact of (i) the LiM training strategy, (ii) the layout discriminator " + }, + { + "bbox": [ + 46, + 604, + 287, + 640 + ], + "type": "inline_equation", + "content": "(D_{box})" + }, + { + "bbox": [ + 46, + 604, + 287, + 640 + ], + "type": "text", + "content": ", and (iii) the B2M module. To this end, we build model variants by:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 642, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 47, + 642, + 286, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 642, + 286, + 666 + ], + "spans": [ + { + "bbox": [ + 47, + 642, + 286, + 666 + ], + "type": "text", + "content": "- Using original training graphs, instead of sampled minimal graphs, to train the layout predictor " + }, + { + "bbox": [ + 47, + 642, + 286, + 666 + ], + "type": "inline_equation", + "content": "(w / o LiM)" + }, + { + "bbox": [ + 47, + 642, + 286, + 666 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 666, + 286, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 666, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 47, + 666, + 286, + 689 + ], + "type": "text", + "content": "- Removing B2M from our full model, but using grids for embedding selection (w/o B2M);" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 689, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 286, + 712 + ], + "type": "text", + "content": "- Removing both LiM and B2M from our full model (" + }, + { + "bbox": [ + 47, + 689, + 286, + 712 + ], + "type": "inline_equation", + "content": "w / o" + }, + { + "bbox": [ + 47, + 689, + 286, + 712 + ], + "type": "text", + "content": " LiM/B2M);" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "type": "table", + "bbox": [ + 307, + 290, + 550, + 370 + ], + "blocks": [ + { + "bbox": [ + 305, + 258, + 545, + 281 + ], + "lines": [ + { + "bbox": [ + 305, + 258, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 305, + 258, + 545, + 281 + ], + "type": "text", + "content": "Table 2. Performance indices w.r.t. the ablation study on CHROME19. The model variant in the last row is our baseline." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 290, + 550, + 370 + ], + "lines": [ + { + "bbox": [ + 307, + 290, + 550, + 370 + ], + "spans": [ + { + "bbox": [ + 307, + 290, + 550, + 370 + ], + "type": "table", + "html": "
SSIM↑FID↓WER↓ExpRate↑mIOU↑
Ours (full)0.79310.980.3260.3160.364
w/o LiM0.79011.550.3270.3150.335
w/o B2M0.79011.460.3320.3080.354
w/o LiM/B2M0.78613.300.3650.2790.338
Sg2im (base)0.78710.080.3930.2190.324
", + "image_path": "ae84d0b94014a264ed9d711098575ba8fdcb0f93aa8fdf84c9c65b8f141c6d72.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 389, + 545, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 389, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 305, + 389, + 545, + 413 + ], + "type": "text", + "content": "- Training the whole model following Sg2im, without using LiM, " + }, + { + "bbox": [ + 305, + 389, + 545, + 413 + ], + "type": "inline_equation", + "content": "D_{box}" + }, + { + "bbox": [ + 305, + 389, + 545, + 413 + ], + "type": "text", + "content": ", and B2M (Sg2im (base))." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 413, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 413, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 305, + 413, + 545, + 437 + ], + "type": "text", + "content": "We additionally report the mean Intersection over Union (mIoU) between the predicted layouts and the target ones." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 437, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 653 + ], + "type": "text", + "content": "As shown in Table 2, removing LiM or B2M slightly changes the indices about image quality, decreases the IoU. Specially, removing LiM decreases the mIoU by about 3 percent absolutely. This implies the significance of LiM in generating good layouts. In addition, if we removing " + }, + { + "bbox": [ + 304, + 437, + 545, + 653 + ], + "type": "inline_equation", + "content": "D_{box}" + }, + { + "bbox": [ + 304, + 437, + 545, + 653 + ], + "type": "text", + "content": " or B2M, along with LiM, the performance seriously decreases, almost in terms of all performance indices. This implies that our " + }, + { + "bbox": [ + 304, + 437, + 545, + 653 + ], + "type": "inline_equation", + "content": "D_{box}" + }, + { + "bbox": [ + 304, + 437, + 545, + 653 + ], + "type": "text", + "content": " and B2M are significant for boosting the quality of generated images from all aspects, including the fidelity (SSIM), the usability for recognition (WER & ExpRate), and the layout (mIoU). Fig. 6 illustrates the layouts and images generated by different model variants. All the model variants, except the full model, produce overlapped symbols or symbols in unnatural sizes. The image generated by our model presents no obvious defects and are easy to recognize. Such observations further demonstrate the significance of LiM, B2M, and layout discriminator in layout prediction and image generation." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 659, + 388, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 659, + 388, + 672 + ], + "spans": [ + { + "bbox": [ + 306, + 659, + 388, + 672 + ], + "type": "text", + "content": "4.4. Applications" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "text", + "content": "Mathematical Expression Manipulation. In addition, we apply the previous learned model to expression manipulation using symbol graphs. Given a source symbol graph," + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15681" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 71, + 285, + 216 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 285, + 216 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 285, + 216 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 285, + 216 + ], + "type": "table", + "html": "
Source1/3πr2hx/a + x/b - x/cax + bx + c/2a + √b + c/2∫_0^1 x^2y dxdyT_1^2/T_2^2 = a_1^3/a_2^3
Add1/3πr^2hzx + d/a + x/b - x/cax + b^x / 2a + √b + c/2 + 7∫_0^1 x^2y^2/2 dxdyT_1^2/T_2^2 = a_1^3 + 2/a_2^3
Delete1/3πr^2x/a + x/bax + b^x + ca + √b / 2∫_0^1 x^2y^2 dxdyT_1^2/T_2^2 = a/a
Change1/7πr^2hx/a + x/e - x/cax + b^y + c/2a + √b + c/4∫_0^1 x^2y dxdyT_1^2/T_2^2 = e^3/a^3
", + "image_path": "b1a5bcfe6d8a2bc9cf3378b11c009bf0a75cf7bb6983b0e6e0509e2ec6690653.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 225, + 287, + 249 + ], + "lines": [ + { + "bbox": [ + 46, + 225, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 46, + 225, + 287, + 249 + ], + "type": "text", + "content": "Figure 8. Illustration of expression manipulation using symbol graphs (LaTeX sequences)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 270, + 287, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 287, + 342 + ], + "type": "text", + "content": "we edit it by adding, deleting, or changing symbol nodes in it, and apply the previously learned generator to produce the corresponding HME image. As shown in Fig. 8, the generated image changes consistently with the editing operation. Besides, the manipulated images present natural layouts, recognizable symbols, and clear strokes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 342, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 287, + 510 + ], + "type": "text", + "content": "Data Augmentation for HMER. Finally, we validate the possible application of HMEG in augmenting HMER models, since the generated data could improve the amount and diversity of training data. To this end, we conduct experiments on CROHME2014/2016/2019 [34], and use CAN [27] as the HMER model. Here, we train CAN with (1) the standard training data (denoted by CAN); or (2) with previous data augmentation, e.g. rotation, (denoted by " + }, + { + "bbox": [ + 46, + 342, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\mathrm{CAN^{*}}" + }, + { + "bbox": [ + 46, + 342, + 287, + 510 + ], + "type": "text", + "content": "); or (3) with our synthetic data augmentation, using 6,000 additional generated samples (denoted by " + }, + { + "bbox": [ + 46, + 342, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\mathrm{CAN^{\\dagger}}" + }, + { + "bbox": [ + 46, + 342, + 287, + 510 + ], + "type": "text", + "content": "); or (4) both previous and our data augmentation (denoted by " + }, + { + "bbox": [ + 46, + 342, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\mathrm{CAN^{*\\dagger}}" + }, + { + "bbox": [ + 46, + 342, + 287, + 510 + ], + "type": "text", + "content": "). The model is trained for 120 epochs under each setting, using the official training parameters. Afterward, we apply the learned HMER model to both generated and real data." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 510, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 653 + ], + "type": "text", + "content": "As shown in Table 3, our method achieves " + }, + { + "bbox": [ + 46, + 510, + 287, + 653 + ], + "type": "inline_equation", + "content": "5 - 11\\%" + }, + { + "bbox": [ + 46, + 510, + 287, + 653 + ], + "type": "text", + "content": " absolute improvement over the baseline, and performs comparably to previous data augmentation, across all the datasets. Besides, our method can be jointly used with previous augmentation to further boost the performance (i.e. " + }, + { + "bbox": [ + 46, + 510, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\mathrm{CAN}^{*\\dagger}" + }, + { + "bbox": [ + 46, + 510, + 287, + 653 + ], + "type": "text", + "content": "). We additionally report the HMER performance, as well as the SSIM and StruRate, of our generated samples. The SSIM and StruRate values are high, and the ExpRate values (by " + }, + { + "bbox": [ + 46, + 510, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\mathrm{CAN}^{*} / \\mathrm{CAN}^{\\dagger}" + }, + { + "bbox": [ + 46, + 510, + 287, + 653 + ], + "type": "text", + "content": ") approach those of real data, on all datasets. These results demonstrate the robustness of our method, in generating high-quality HMEs, with correct layouts and recognizable symbols." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "Note: Previous HME synthesis methods [10, 25] synthesize HMEs by shuffling the symbols and layouts contained in original HME datasets. Neither of them is a generative method. And they may cause overfitting to existing datasets. In contrast, our pipeline allows excellent flexibil" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 306, + 113, + 547, + 174 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 545, + 104 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 104 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 104 + ], + "type": "text", + "content": "Table 3. Results on CROHME14/16/19 testing sets and our generated images. *,†, *† denote using previous data augmentation, our synthetic data augmentation, and both, respectively." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 113, + 547, + 174 + ], + "lines": [ + { + "bbox": [ + 306, + 113, + 547, + 174 + ], + "spans": [ + { + "bbox": [ + 306, + 113, + 547, + 174 + ], + "type": "table", + "html": "
HMEG (generated)HMER (real)
SSIMCAN*CAN†StruRateCANCAN*CAN†CAN*†
CROHME140.78952.154.198.144.752.950.255.4
CROHME160.79851.455.896.542.852.453.957.6
CROHME190.79355.456.497.739.348.449.658.5
", + "image_path": "ab7ebd64649ed58309c86373d9b15e931cbd0ba763d65d698246da6611886d7c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 198, + 545, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 198, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 305, + 198, + 545, + 232 + ], + "type": "text", + "content": "ity in generating diverse new samples. The experimental results also demonstrate the advantages of our approach in both HMEG and HMER." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 248, + 383, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 248, + 383, + 260 + ], + "spans": [ + { + "bbox": [ + 306, + 248, + 383, + 260 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 269, + 545, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 269, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 269, + 545, + 399 + ], + "type": "text", + "content": "In this paper, we propose a novel method to solve the challenging task of generating HMEs from LaTeX sequences. We formulate it as a graph-to-image generation task, and focus on boosting the layout prediction without mask supervision. Experimental results demonstrate the effectiveness of our method, and its possible applications in expression manipulation and the challenging inverse task, i.e. HMER. Inspired by the remarkable process of generative augmentation in visual understanding [35], we believe that: the novel HMEG task we explore here has great potential to significantly boost HMER." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 402, + 546, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 402, + 546, + 558 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 546, + 558 + ], + "type": "text", + "content": "Due to computational limitations, we merely use naive networks in the implementation. The generated symbols are occasionally unrecognizable, especially in complex formulas. It's promising to boost the generation quality by using advanced networks (e.g. VQ-GAN [11] or diffusion models [15]). Besides, we will extend the proposed techniques to natural/artistic image generation by exploring the inspiring capacity of diffusion models, and other tasks of structured graph data. Finally, to further boost the HMER performance, it's promising to use the triples of " + }, + { + "bbox": [ + 304, + 402, + 546, + 558 + ], + "type": "inline_equation", + "content": "\\{LaTeX, predicted layout, generated image\\}" + }, + { + "bbox": [ + 304, + 402, + 546, + 558 + ], + "type": "text", + "content": " in HMEG as pseudo labeled data, or to use the layout discriminator for enhancing the layout detection module in an HMER pipeline." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 572, + 408, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 572, + 408, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 572, + 408, + 586 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "We greatly appreciate Biao Ma's and Chang Jiang's help in the implementation of HMER. This work was supported in part by the National Natural Science Foundation of China under Grants U22A2096 and 61971172; in part by the Technology Innovation Leading Program of Shaanxi under Grant 2022QFY01-15; in part by the Fundamental Research Funds for the Central Universities under Grant QTZX23042 and ZYTS24012; and in part by the Proof of Concept Foundation of Xidian University Hangzhou Institute of Technology under Grant No. GNYZ2023YL0301." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15682" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Oron Ashual and Lior Wolf. Specifying object attributes and relations in interactive scene generation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4561-4569, 2019. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "type": "text", + "content": "[2] Xiaohang Bian, Bo Qin, Xiaozhe Xin, Jianwu Li, Xuefeng Su, and Yanfeng Wang. Handwritten mathematical expression recognition via attention aggregation based bidirectional mutual learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 113-121, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 288, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 288, + 214 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 288, + 214 + ], + "type": "text", + "content": "[3] Stefan Bornholdt. Less is more in modeling large genetic networks. Science, 310(5747):449-451, 2005. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 217, + 288, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 217, + 288, + 259 + ], + "spans": [ + { + "bbox": [ + 53, + 217, + 288, + 259 + ], + "type": "text", + "content": "[4] Qifeng Chen and Vladlen Koltun. Photographic image synthesis with cascaded refinement networks. In Proceedings of the IEEE international conference on computer vision, pages 1511-1520, 2017. 3, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 262, + 287, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 262, + 287, + 305 + ], + "spans": [ + { + "bbox": [ + 53, + 262, + 287, + 305 + ], + "type": "text", + "content": "[5] Kyunghyun Cho, Bart Van Merrienboer, Dzmitry Bahdanau, and Yoshua Bengio. On the properties of neural machine translation: Encoder-decoder approaches. arXiv preprint arXiv:1409.1259, 2014. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 307, + 287, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 307, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 53, + 307, + 287, + 361 + ], + "type": "text", + "content": "[6] Gang Dai, Yifan Zhang, Qingfeng Wang, Qing Du, Zhuliang Yu, Zhuoman Liu, and Shuangping Huang. Disentangling writer and character styles for handwriting generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5977-5986, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 364, + 287, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 364, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 53, + 364, + 287, + 418 + ], + "type": "text", + "content": "[7] Gang Dai, Yifan Zhang, Qingfeng Wang, Qing Du, Zhuliang Yu, Zhuoman Liu, and Shuangping Huang. Disentangling writer and character styles for handwriting generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5977-5986, 2023. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 421, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 421, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 53, + 421, + 287, + 464 + ], + "type": "text", + "content": "[8] Ayan Das, Yongxin Yang, Timothy M Hospedales, Tao Xiang, and Yi-Zhe Song. Chirodiff: Modelling chirographic data with diffusion models. In The Eleventh International Conference on Learning Representations, 2023. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 466, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 466, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 53, + 466, + 287, + 521 + ], + "type": "text", + "content": "[9] Helisa Dhamo, Azade Farshad, Iro Laina, Nassir Navab, Gregory D Hager, Federico Tombari, and Christian Rupprecht. Semantic image manipulation using scene graphs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5213-5222, 2020. 3, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 522, + 287, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 287, + 576 + ], + "type": "text", + "content": "[10] Lan-Fang Dong, Han-Chao Liu, and Xin-Ming Zhang. Synthetic data generation and shuffled multi-round training based offline handwritten mathematical expression recognition. Journal of Computer Science and Technology, 37(6): 1427-1443, 2022. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 578, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 622 + ], + "type": "text", + "content": "[11] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 678 + ], + "type": "text", + "content": "[12] Wan-Cyuan Fan, Yen-Chun Chen, DongDong Chen, Yu Cheng, Lu Yuan, and Yu-Chiang Frank Wang. Frido: Feature pyramid diffusion for complex scene image synthesis. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 579-587, 2023. 1, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[13] Azade Farshad, Sabrina Musatian, Helisa Dhamo, and Nassir Navab. Migs: Meta image generation from scene graphs. arXiv preprint arXiv:2110.11918, 2021. 2, 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "type": "text", + "content": "[14] Yingnan Fu, Wenyuan Cai, Ming Gao, and Aoying Zhou. Symbol location-aware network for improving handwritten mathematical expression recognition. In Proceedings of the 2023 ACM International Conference on Multimedia Retrieval, pages 516-524, 2023. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 546, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 546, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 546, + 183 + ], + "type": "text", + "content": "[15] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10696-10706, 2022. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 184, + 547, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 547, + 249 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 547, + 249 + ], + "type": "text", + "content": "[16] Roei Herzig, Amir Bar, Huijuan Xu, Gal Chechik, Trevor Darrell, and Amir Globerson. Learning canonical representations for scene graph to image generation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVI 16, pages 210-227. Springer, 2020. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 250, + 546, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 250, + 546, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 250, + 546, + 293 + ], + "type": "text", + "content": "[17] M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In NIPS, pages 6626-6637. 2017. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 294, + 546, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 546, + 316 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 546, + 316 + ], + "type": "text", + "content": "[18] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 316, + 546, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 546, + 360 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 546, + 360 + ], + "type": "text", + "content": "[19] Hongyu Wang and Jianshu Zhang. Handwritten mathematical expression recognition (pytorch), 2019. https://github.com/whywhs/Pytorch-Handwritten-Mathematical-Expression-Recognition.6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 361, + 546, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 361, + 546, + 404 + ], + "spans": [ + { + "bbox": [ + 308, + 361, + 546, + 404 + ], + "type": "text", + "content": "[20] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2021. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 405, + 546, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 405, + 546, + 459 + ], + "spans": [ + { + "bbox": [ + 308, + 405, + 546, + 459 + ], + "type": "text", + "content": "[21] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 1, 3, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 460, + 546, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 546, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 546, + 514 + ], + "type": "text", + "content": "[22] Maor Ivgi, Yaniv Benny, Avichai Ben-David, Jonathan Berant, and Lior Wolf. Scene graph to image generation with contextualized object layout refinement. In 2021 IEEE International Conference on Image Processing (ICIP), pages 2428-2432. IEEE, 2021. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 514, + 546, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 514, + 546, + 581 + ], + "spans": [ + { + "bbox": [ + 308, + 514, + 546, + 581 + ], + "type": "text", + "content": "[23] Nikita Jaipuria, Xianling Zhang, Rohan Bhasin, Mayar Arafa, Punarjay Chakravarty, Shubham Shrivastava, Sagar Manglani, and Vidya N Murali. Deflating dataset bias using synthetic data augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 772-773, 2020. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 582, + 546, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 546, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 546, + 624 + ], + "type": "text", + "content": "[24] Justin Johnson, Agrim Gupta, and Li Fei-Fei. Image generation from scene graphs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1219-1228, 2018. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 625, + 546, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 546, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 546, + 689 + ], + "type": "text", + "content": "[25] Vu Tran Minh Khuong, Ung Quang Huy, Nakagawa Masaki, and Minh Khanh Phan. Generating synthetic handwritten mathematical expressions from a latex sequence or a mathml script. In 2019 International Conference on Document Analysis and Recognition (ICDAR), pages 922-927. IEEE, 2019. 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 691, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 546, + 713 + ], + "type": "text", + "content": "[26] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In 3rd International Conference on" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15683" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "type": "text", + "content": "Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 152 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 152 + ], + "type": "text", + "content": "[27] Bohan Li, Ye Yuan, Dingkang Liang, Xiao Liu, Zhilong Ji, Jinfeng Bai, Wenyu Liu, and Xiang Bai. When counting meets hmer: counting-aware network for handwritten mathematical expression recognition. In European Conference on Computer Vision, pages 197-214. Springer, 2022. 1, 2, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 288, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 288, + 208 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 288, + 208 + ], + "type": "text", + "content": "[28] Wenbo Li, Pengchuan Zhang, Lei Zhang, Qiuyuan Huang, Xiaodong He, Siwei Lyu, and Jianfeng Gao. Object-driven text-to-image synthesis via adversarial training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12174-12182, 2019. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 288, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 288, + 241 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 288, + 241 + ], + "type": "text", + "content": "[29] Yujia Li, Daniel Tarlow, Marc Brockschmidt, and Richard Zemel. Gated graph sequence neural networks. arXiv preprint arXiv:1511.05493, 2015. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 243, + 288, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 243, + 288, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 243, + 288, + 297 + ], + "type": "text", + "content": "[30] Yujia Li, Chenjie Gu, Thomas Dullien, Oriol Vinyals, and Pushmeet Kohli. Graph matching networks for learning the similarity of graph structured objects. In International conference on machine learning, pages 3835-3845. PMLR, 2019. 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 298, + 288, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 288, + 343 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 288, + 343 + ], + "type": "text", + "content": "[31] Yikang Li, Tao Ma, Yeqi Bai, Nan Duan, Sining Wei, and Xiaogang Wang. Pastegan: A semi-parametric method to generate image from scene graph. Advances in Neural Information Processing Systems, 32, 2019. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 344, + 288, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 344, + 288, + 398 + ], + "spans": [ + { + "bbox": [ + 48, + 344, + 288, + 398 + ], + "type": "text", + "content": "[32] Zejian Li, Jingyu Wu, Immanuel Koh, Yongchuan Tang, and Lingyun Sun. Image synthesis from layout with locality-aware mask adaption. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13819-13828, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 399, + 288, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 288, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 288, + 443 + ], + "type": "text", + "content": "[33] Zhe Li, Xinyu Wang, Yuliang Liu, Lianwen Jin, Yichao Huang, and Kai Ding. Improving handwritten mathematical expression recognition via similar symbol distinguishing. IEEE Transactions on Multimedia, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 445, + 288, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 288, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 288, + 521 + ], + "type": "text", + "content": "[34] Mahshad Mahdavi, Richard Zanibbi, Harold Mouchere, Christian Viard-Gaudin, and Utpal Garain. Icdar 2019 crohme+ tfd: Competition on recognition of handwritten mathematical expressions and typeset formula detection. In 2019 International Conference on Document Analysis and Recognition (ICDAR), pages 1533-1538. IEEE, 2019. 2, 5, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 522, + 288, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 288, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 288, + 567 + ], + "type": "text", + "content": "[35] Quang Nguyen, Truong Vu, Anh Tran, and Khoi Nguyen. Dataset diffusion: Diffusion-based synthetic data generation for pixel-level semantic segmentation. Advances in Neural Information Processing Systems, 36, 2024. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 568, + 288, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 288, + 612 + ], + "type": "text", + "content": "[36] Augustus Odena, Christopher Olah, and Jonathon Shlens. Conditional image synthesis with auxiliary classifier gans. In International conference on machine learning, pages 2642-2651. PMLR, 2017. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 288, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 288, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 288, + 667 + ], + "type": "text", + "content": "[37] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2337-2346, 2019. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 669, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 288, + 714 + ], + "type": "text", + "content": "[38] Konpat Preechakul, Nattanat Chathee, Suttisak Wizadwongsa, and Supasorn Suwajanakorn. Diffusion autoencoders: Toward a meaningful and decodable representation. In Proceedings of the IEEE/CVF Conference on Computer" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 327, + 72, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 72, + 545, + 95 + ], + "type": "text", + "content": "Vision and Pattern Recognition, pages 10619-10629, 2022. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 547, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 152 + ], + "type": "text", + "content": "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 153, + 545, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 153, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 307, + 153, + 545, + 198 + ], + "type": "text", + "content": "[40] Renato Sortino, Simone Palazzo, Francesco Rundo, and Concetto Spampinato. Transformer-based image generation from scene graphs. Computer Vision and Image Understanding, 233:103721, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 200, + 545, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 200, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 307, + 200, + 545, + 266 + ], + "type": "text", + "content": "[41] Matthias Springstein, Eric Müller-Budack, and Ralph Ewert. Unsupervised training data generation of handwritten formulas using generative adversarial networks with self-attention. In Proceedings of the 2021 Workshop on MultiModal Pre-Training for Multimedia Understanding, pages 46–54, 2021. 1, 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 268, + 545, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 268, + 545, + 322 + ], + "spans": [ + { + "bbox": [ + 307, + 268, + 545, + 322 + ], + "type": "text", + "content": "[42] Sitong Su, Lianli Gao, Junchen Zhu, Jie Shao, and Jingkuan Song. Fully functional image manipulation using scene graphs in a bounding-box free way. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1784-1792, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 324, + 545, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 324, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 307, + 324, + 545, + 369 + ], + "type": "text", + "content": "[43] Wei Sun and Tianfu Wu. Image synthesis from reconfigurable layout and style. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10531-10540, 2019. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 371, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 371, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 371, + 545, + 415 + ], + "type": "text", + "content": "[44] Wei Sun and Tianfu Wu. Learning layout and style reconfigurable gans for controllable image synthesis. IEEE transactions on pattern analysis and machine intelligence, 44(9): 5070-5087, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 417, + 545, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 417, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 307, + 417, + 545, + 460 + ], + "type": "text", + "content": "[45] Tristan Sylvain, Pengchuan Zhang, Yoshua Bengio, R Devon Hjelm, and Shikhar Sharma. Object-centric image generation from layouts. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2647-2655, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 462, + 545, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 462, + 545, + 507 + ], + "spans": [ + { + "bbox": [ + 307, + 462, + 545, + 507 + ], + "type": "text", + "content": "[46] Thanh-Nghia Truong, Cuong Tuan Nguyen, and Masaki Nakagawa. Syntactic data generation for handwritten mathematical expression recognition. Pattern Recognition Letters, 153:83-91, 2022. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 509, + 545, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 509, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 307, + 509, + 545, + 553 + ], + "type": "text", + "content": "[47] Ashish Vaswani, Noam Shazeer, Niki Parmar, Llion Jones, Jakob Uszkoreit, Aidan N Gomez, and Lukasz Kaiser. Attention is all you need. In Advances in Neural Information Processing Systems, pages 5998-6008. 2017. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 555, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 555, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 307, + 555, + 545, + 609 + ], + "type": "text", + "content": "[48] Chi Wang, Min Zhou, Tiezheng Ge, Yuning Jiang, Hujun Bao, and Weiwei Xu. Cf-font: Content fusion for few-shot font generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1858-1867, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 611, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 611, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 307, + 611, + 545, + 656 + ], + "type": "text", + "content": "[49] Zhou Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Process, 13(4):600-612, 2004. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 658, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 714 + ], + "type": "text", + "content": "[50] Chuan Wen, Yujie Pan, Jie Chang, Ya Zhang, Siheng Chen, Yanfeng Wang, Mei Han, and Qi Tian. Handwritten chinese font generation with collaborative stroke refinement. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3882-3891, 2021. 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15684" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "text", + "content": "[51] Chenxi Whitehouse, Monojit Choudhury, and Alham Fikri Aji. Llm-powered data augmentation for enhanced crosslingual performance. arXiv preprint arXiv:2305.14288, 2023. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 117, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 117, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 117, + 287, + 172 + ], + "type": "text", + "content": "[52] Changjie Wu, Jun Du, Yunqing Li, Jianshu Zhang, Chen Yang, Bo Ren, and Yiqing Hu. Tdv2: A novel tree-structured decoder for offline mathematical expression recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2694–2702, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 288, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 288, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 288, + 216 + ], + "type": "text", + "content": "[53] Jin-Wen Wu, Fei Yin, Yan-Ming Zhang, Xu-Yao Zhang, and Cheng-Lin Liu. Handwritten mathematical expression recognition via paired adversarial learning. International Journal of Computer Vision, 128:2386-2401, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 217, + 288, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 288, + 271 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 288, + 271 + ], + "type": "text", + "content": "[54] Jin-Wen Wu, Fei Yin, Yan-Ming Zhang, Xu-Yao Zhang, and Cheng-Lin Liu. Graph-to-graph: towards accurate and interpretable online handwritten mathematical expression recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2925-2933, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 272, + 288, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 288, + 315 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 288, + 315 + ], + "type": "text", + "content": "[55] Yang Wu, Pengxu Wei, and Liang Lin. Scene graph to image synthesis via knowledge consensus. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2856-2865, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 316, + 287, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 287, + 360 + ], + "type": "text", + "content": "[56] Ximing Xing, Chuang Wang, Haitao Zhou, Jing Zhang, Qian Yu, and Dong Xu. Diffsketcher: Text guided vector sketch synthesis through latent diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 361, + 287, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 361, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 48, + 361, + 287, + 425 + ], + "type": "text", + "content": "[57] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Fine-grained text to image generation with attentional generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1316-1324, 2018. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 426, + 287, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 287, + 469 + ], + "type": "text", + "content": "[58] Yuan Xue, Yuan-Chen Guo, Han Zhang, Tao Xu, Song-Hai Zhang, and Xiaolei Huang. Deep image synthesis from intuitive user input: A review and perspectives. Computational Visual Media, 8:3-31, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 471, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 287, + 514 + ], + "type": "text", + "content": "[59] Chen Yang, Jun Du, Jianshu Zhang, Changjie Wu, Mingjun Chen, and JiaJia Wu. Tree-based data augmentation and mutual learning for offline handwritten mathematical expression recognition. Pattern Recognition, 132:108910, 2022. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 514, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 568 + ], + "type": "text", + "content": "[60] Ling Yang, Zhilin Huang, Yang Song, Shenda Hong, Guohao Li, Wentao Zhang, Bin Cui, Bernard Ghanem, and Ming-Hsuan Yang. Diffusion-based scene graph to image generation with masked contrastive pre-training. arXiv preprint arXiv:2211.11138, 2022. 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 570, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 613 + ], + "type": "text", + "content": "[61] Jun Yu, Xingxin Xu, Fei Gao, Shengjie Shi, Meng Wang, Dacheng Tao, and Qingming Huang. Toward realistic face photo-sketch synthesis via composition-aided gans. IEEE transactions on cybernetics, 51(9):4350-4362, 2020. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 614, + 287, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 669 + ], + "type": "text", + "content": "[62] Ye Yuan, Xiao Liu, Wondimu Dikubab, Hui Liu, Zhilong Ji, Zhongqin Wu, and Xiang Bai. Syntax-aware network for handwritten mathematical expression recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4553-4562, 2022. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[63] Richard Zanibbi and Dorothea Blostein. Recognition and retrieval of mathematical expressions. International Journal on Document Analysis and Recognition (IJDAR), 15:331-357, 2012. 2, 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 508 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "type": "text", + "content": "[64] Jinshan Zeng, Qi Chen, Yunxin Liu, Mingwen Wang, and Yuan Yao. Strokegan: Reducing mode collapse in Chinese font generation via stroke encoding. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3270-3277, 2021. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 547, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 184 + ], + "type": "text", + "content": "[65] Yu Zeng, Zhe Lin, Jianming Zhang, Qing Liu, John Collo-mosse, Jason Kuen, and Vishal M Patel. Scenecomposer: Any-level semantic image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22468-22478, 2023. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 185, + 546, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 185, + 546, + 228 + ], + "spans": [ + { + "bbox": [ + 308, + 185, + 546, + 228 + ], + "type": "text", + "content": "[66] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris Metaxas. StackGAN++: Realistic image synthesis with stacked generative adversarial networks. IEEE TPAMI, PP(99), 2017. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 229, + 546, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 229, + 546, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 229, + 546, + 274 + ], + "type": "text", + "content": "[67] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 274, + 546, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 274, + 546, + 329 + ], + "spans": [ + { + "bbox": [ + 308, + 274, + 546, + 329 + ], + "type": "text", + "content": "[68] Yaping Zhang, Shuai Nie, Wenju Liu, Xing Xu, Dongxiang Zhang, and Heng Tao Shen. Sequence-to-sequence domain adaptation network for robust text image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 331, + 546, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 546, + 363 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 546, + 363 + ], + "type": "text", + "content": "[69] Bo Zhao, Weidong Yin, Lili Meng, and Leonid Sigal. Layout2image: Image generation from layout. International Journal of Computer Vision, 128:2418-2435, 2020. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 364, + 546, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 364, + 546, + 419 + ], + "spans": [ + { + "bbox": [ + 308, + 364, + 546, + 419 + ], + "type": "text", + "content": "[70] Wenqi Zhao and Liangcai Gao. Comer: Modeling coverage for transformer-based handwritten mathematical expression recognition. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVIII, pages 392-408. Springer, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 420, + 546, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 420, + 546, + 453 + ], + "spans": [ + { + "bbox": [ + 308, + 420, + 546, + 453 + ], + "type": "text", + "content": "[71] J. Zhu, T. Park, P. Isola, and A. A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In ICCV, pages 2242-2251, 2017. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 453, + 546, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 453, + 546, + 508 + ], + "spans": [ + { + "bbox": [ + 308, + 453, + 546, + 508 + ], + "type": "text", + "content": "[72] Peihao Zhu, Rameen Abdal, Yipeng Qin, and Peter Wonka. Sean: Image synthesis with semantic region-adaptive normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5104-5113, 2020. 1" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15685" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Generating Human Motion in 3D Scenes from Text Descriptions/507f8570-ad9c-455e-838f-8fa1bda75b31_content_list.json b/2024/Generating Human Motion in 3D Scenes from Text Descriptions/507f8570-ad9c-455e-838f-8fa1bda75b31_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..aca3c5039dff85c36df9955f51915e3c90e34590 --- /dev/null +++ b/2024/Generating Human Motion in 3D Scenes from Text Descriptions/507f8570-ad9c-455e-838f-8fa1bda75b31_content_list.json @@ -0,0 +1,1650 @@ +[ + { + "type": "text", + "text": "Generating Human Motion in 3D Scenes from Text Descriptions", + "text_level": 1, + "bbox": [ + 161, + 130, + 810, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhi Cen1 Huaijin Pi1 Sida Peng1† Zehong Shen1 Minghui Yang2 Shuai Zhu2 Hujun Bao1 Xiaowei Zhou1 1Zhejiang University 2Ant Group", + "bbox": [ + 243, + 179, + 743, + 247 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3c2faf1e4e6f52949b7ce3ddc3dedf11722d5a9cbb2298a876ba3a259a10ce09.jpg", + "image_caption": [ + "Test on the HUAMNISE" + ], + "image_footnote": [], + "bbox": [ + 84, + 282, + 282, + 406 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/99d0d5a473d70f190d9a6114d0a2c4f4be9a882e72e49bd5e9eaac6acaa55af2.jpg", + "image_caption": [ + "Figure 1. Generating human motions in 3D scenes from text descriptions. Our method can generate human motions containing accurate human-object interactions in 3D scenes based on textural descriptions. Although our method is trained and tested on the HUMANISE dataset, it can generalize to other scenes, e.g., the scenes in the PROX dataset. Left: test results on the HUMANISE dataset. Right: generalization results on the PROX scenes." + ], + "image_footnote": [], + "bbox": [ + 285, + 282, + 482, + 406 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8388d847c4db90f8c836b8462a4fa13a7484f0ba97ae2a0a8ec189249e1fe094.jpg", + "image_caption": [ + "Generalize to the PROX scenes" + ], + "image_footnote": [], + "bbox": [ + 488, + 282, + 687, + 405 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3fd1225ec84ef23dc5e91c1bec639a253143045e3118f2216cb8f3151dd1b3da.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 691, + 282, + 887, + 405 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 505, + 313, + 520 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generating human motions from textual descriptions has gained growing research interest due to its wide range of applications. However, only a few works consider human-scene interactions together with text conditions, which is crucial for visual and physical realism. This paper focuses on the task of generating human motions in 3D indoor scenes given text descriptions of the human-scene interactions. This task presents challenges due to the multimodality nature of text, scene, and motion, as well as the need for spatial reasoning. To address these challenges, we propose a new approach that decomposes the complex problem into two more manageable sub-problems: (1) language grounding of the target object and (2) object-centric motion generation. For language grounding of the target object, we leverage the power of large language models. For motion generation, we design an object-centric scene representation for the generative model to focus on the target object, thereby reducing the scene complexity and facilitating the modeling of the relationship between human motions and the object. Experiments demonstrate the better motion quality of our approach compared to baselines and validate our design choices. Code will be available at link.", + "bbox": [ + 75, + 534, + 473, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 505, + 632, + 520 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Human motion generation has been a long-standing problem due to its broad range of applications such as game development, virtual reality, and movie production. Recently, this area witnessed a paradigm shift from avatar animation given rich user input [29] to learning-based motion generation from high-level language prompts, e.g. text descriptions about the desired motion [2, 3, 15, 18, 19, 39, 56, 74, 75]. However, most prior works on text-driven motion synthesis do not consider human-scene interactions [39, 56, 74, 75] while the scene context and physical constraints of the environment largely define the fidelity of the generated human motions.", + "bbox": [ + 496, + 532, + 890, + 714 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we focus on generating motions from text descriptions in 3D indoor scenes. Specifically, given a 3D scan of the target scene and a text description of a human action that interacts with the scene, we aim to generate natural human motions that are consistent with the text description.", + "bbox": [ + 496, + 715, + 893, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This problem presents several challenges, primarily due to the multi-modality nature of text, scene, and human motion. In contrast to previous methods [2, 3, 56, 74, 75] focusing solely on textual descriptions of how human moves, our task also includes texts that additionally describe the spatial details in the given scene (e.g., sit on the armchair near the desk). Therefore, this task requires spatial reason-", + "bbox": [ + 496, + 795, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "The authors from Zhejiang University are affiliated with the State Key Lab of CAD&CG. † Corresponding author.", + "bbox": [ + 75, + 875, + 468, + 902 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1855", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing skills [1], where the model should build a text-object mapping to locate a specific object in 3D scenes aligned with a natural language description. In addition, the generated motions should also be coherent with scene contexts.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As a pioneer work, HUMANISE [83] builds a conditional variational autoencoder (cVAE) [41, 69] with separate encoders of scenes and texts for multi-modality understanding. To enable spatial reasoning ability, they introduce auxiliary tasks of directly regressing object centers to learn 3D visual grounding in an implicit manner. But they do not explicitly utilize the predicted centers and thus the inductive bias of visual grounding cannot be fully incorporated. In addition, HUMANISE [83] encodes the entire scene with a single point transformer [99]. Directly generating motions with such a model is challenging, as 3D point clouds are inherently noisy and complex [48, 88], leading to the inability to locate the target object. As suggested by [51], not every point of the scene is relevant to the final human motions. Therefore, it is also necessary to develop a more targeted approach that focuses on the relevant parts of the scene to improve the quality of motion generation.", + "bbox": [ + 75, + 154, + 470, + 411 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle the problems mentioned above, we propose a novel approach that exploits the power of the large language models (LLMs) [6, 42, 52]. Our key idea is to address the challenging task of generating motion in a scene based on textual cues by breaking it down into two smaller problems: (1) language grounding of objects in 3D scenes and (2) generating motions with a focus on the target object. For language grounding, rather than directly learning text-object mapping, we propose to formulate it as question answering and utilize the large prior knowledge of LLMs. Specifically, we first construct scene graphs of 3D scenes and generate their textual descriptions. Then we employ ChatGPT [52] to analyze the relationship between scene descriptions and input instructions and respond with 3D visual grounding answers. Experiments prove the effectiveness of this strategy.", + "bbox": [ + 75, + 412, + 468, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For motion generation, we design an object-centric representation to help the generative model focus on the target object. Specifically, we convert point clouds around the target object into volumetric sensors [70] to build object-centric representation. Then we employ diffusion models [57, 75] to synthesize human motions given object-centric representation and texts. Compared with original scene point clouds which might have various scales, object-centric representation is more compact and robust to scales as objects in the same category are of similar size. Therefore, this representation reduces scene complexity and facilitates the modeling of the relationship between motions and objects.", + "bbox": [ + 75, + 642, + 468, + 823 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conducted thorough comparative and ablation experiments on the HUMANISE dataset. The results demonstrate that our method outperforms the baseline, reflected in more accurate object grounding results and better motions that align with the textual descriptions and scenes. We further", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "show that our approach can generalize to the PROX dataset [22] without any fine-tuning.", + "bbox": [ + 500, + 90, + 890, + 122 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 500, + 138, + 638, + 154 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Motion synthesis", + "text_level": 1, + "bbox": [ + 500, + 164, + 666, + 180 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Deep learning for motion synthesis. Deep learning methods for motion synthesis have attracted increasing attention in recent years [14, 21, 25, 28, 29, 50, 71, 91]. Various techniques, including MLP [29], mixture of experts (MoE) [72, 91], and recurrent neural network (RNN) [21], are employed to tackle this task. To generate diverse motion results, previous works explore cVAE [46, 96], GANs [44, 61], normalizing flow [26], and diffusion models [75].", + "bbox": [ + 498, + 193, + 890, + 315 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-driven motion generation. Recently, there has been a growing interest in text-driven motion generation [2, 56, 74]. This task takes natural language as input and synthesizes human motions that align with instructions. KIT-ML [58] is the first benchmark of this task. Some works further annotate the AMASS dataset [49] with action labels [59] and text [18]. To tackle this task, [2, 15, 74] propose to learn a shared latent space for text and motion. [56] employs transformer VAE [55] to generate diverse results. [19, 93] achieve better performance by discrete representation with VQ-VAE [65, 78]. [12, 39, 75, 94] successfully apply diffusion models in this direction. [5, 43] further explore latent diffusion models [67]. Based on [75], [38, 85] introduce sparse spatial control. [3, 60, 68] also investigate long sequence motion generation. Other existing works [4, 37, 82, 98] leverage Large Language Models (LLM) [6, 42, 53, 77] in the motion domain. [4] construct compositional motions by combining different body parts with LLM. [36, 98] regard motions as a kind of language and finetune LLMs [63] with motion tokens [19, 93].", + "bbox": [ + 498, + 320, + 892, + 623 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Scene-aware motion generation This direction is to generate human motions in a 3D scene [70, 80]. To synthesize walking and sitting motions, [70] construct volumetric sensors to encode object information and the surroundings of the character. [23] extends it to synthesize motions with diverse sitting and lying styles. Furthermore, [95] proposes to control sitting styles by hand contacts. To improve the performance, [57] employs a hierarchical framework that generates goal poses, milestones, and motion sequentially. [73, 84] synthesize whole-body (body and hand fingers) grasping motions with small objects. [16, 45, 71, 86] explore interactions with dynamic objects including manipulation [16, 45] and carrying [86]. [8, 24] explore human-object interactions with physically simulated characters. These works mainly focus on one or two objects while others [80, 81] consider a more complex scene input (e.g., point clouds of the scene including walls and floors.). [80, 81] adopt hierarchical frameworks and generate trajectory and", + "bbox": [ + 498, + 628, + 892, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1856", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/908da739b3d8d4a510d67112f70e53a8bb30397b1805dbfc1c3a8aa426e04417.jpg", + "image_caption": [ + "Figure 2. Overview of our two-stage pipeline. In the first stage, given an input scene and a text description (a), we use ChatGPT to locate the target object (b). In the second stage, human motions are synthesized by first producing human trajectories (c) and then generating local poses (d)." + ], + "image_footnote": [], + "bbox": [ + 122, + 88, + 843, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "motion separately. [103] introduces gaze to help generation. [33] further employs diffusion models and [51] synthesizes very long-term motions in scenes by dividing long sequences into several short sequences. [101] designs a reinforcement learning pipeline to enable navigation in a complex scene and interaction style control.", + "bbox": [ + 75, + 309, + 472, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text-driven scene-aware motion generation Only a few works consider text and scenes simultaneously [38, 83, 101]. Although [101] enables text control of sitting styles by [100], the text descriptions in our setting are used to select an object in a cluttered scene. [38] could avoid obstacles during walking while our task needs to handle various actions. The most relevant work to us is HUMANISE [83], which employs a transformer VAE architecture with a two-stream condition module for text and scenes. To accurately localize target objects, they design auxiliary tasks like directly regressing object centers. In contrast to HUMANISE, we propose a two-stage pipeline where we first localize the target object with the help of ChatGPT [52] and then generate human motion using the object-centric representation.", + "bbox": [ + 75, + 407, + 468, + 618 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. 3D visual grounding.", + "text_level": 1, + "bbox": [ + 76, + 633, + 272, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In recent years, visual grounding in 3D scenes has been explored [1, 11, 76] and also tackled in 3D question answering [13, 89]. Given the point cloud of 3D scenes, this task [1] requires models to locate the target object according to text instructions. Most works follow a two-stage scheme where multiple bounding boxes [9, 66, 102] or segmentation results [31, 90] are first predicted and then selecting the object according to language descriptions. [32] introduces multiview inputs, and [88] employs 2D semantics. [34] combines bottom-up [47] and top-down [7] detection methods. [48] designs a single-stage pipeline by progressively selecting key points. [20] extends [32] with the help of GPT [6] to generate multi-view text inputs. More recently, [30] proposes a neuro-symbolic framework with large language-tocode models [10]. Most works only localize a single object and [97] could localize a flexible number of objects. Differ", + "bbox": [ + 75, + 659, + 472, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ent from previous works which directly handle point clouds or multi-view images, we convert the scene into textual descriptions and leverage large language models to infer the target object. Like [30], in this work, we also leverage large language models for object localization.", + "bbox": [ + 496, + 309, + 890, + 383 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Problem setup and preliminaries", + "text_level": 1, + "bbox": [ + 498, + 405, + 799, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we discuss the definition of the task and preliminaries. We aim to generate human motion that is consistent with both the text description and the given scene.", + "bbox": [ + 496, + 433, + 890, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text descriptions. The text description follows the template in Sr3D [1] (e.g., “ [ ]”). There are four actions (walk, sit, stand up, and lie). The target represents the object that the agents need to interact with, and the anchor objects help to determine the target. A certain type of target furniture category usually has many instances in one scene, while the anchor furniture should be unique. To specify the exact target object, there are five types of spatial relations [1] between the target and the anchor: horizontal proximity, vertical proximity, between, allocentric, and support.", + "bbox": [ + 496, + 486, + 892, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Scene representations. The scene is denoted as a point cloud of $N$ points: $S \\in \\mathbf{R}^{N \\times 6}$ , containing the position and normal direction information of each point.", + "bbox": [ + 496, + 659, + 890, + 705 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Motion representations. The output motion sequence is represented as a sequence of SMPL-X [54] body meshes $M$ . SMPL-X is a parametric human body model. In this work, body parameters include body shape parameters $\\beta \\in \\mathbf{R}^{10}$ , global translation $r \\in \\mathbf{R}^3$ , 6D global orientation $\\gamma \\in \\mathbf{R}^6$ , and 6D pose parameters of $J$ joints $\\theta \\in \\mathbf{R}^{J \\times 6}$ [104]. Following [83], $\\beta$ is treated as a condition to model the effect of body shape, and we omit it for ease of notation. Note that start position and pose are also generated by the model.", + "bbox": [ + 496, + 710, + 890, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models. The diffusion [27] is defined as a Markov noising process $\\{\\mathbf{x}_t\\}_{t=0}^T$ that follows $q(\\mathbf{x}_t \\mid \\mathbf{x}_0)$ , where $\\mathbf{x}_0 \\sim q(\\mathbf{x}_0)$ is the data and $\\mathbf{x}_t$ is the noised data at", + "bbox": [ + 496, + 854, + 892, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1857", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5383c1e92cfa5b5fcccbdc841b5a757e4db03d6e018bee428aa39de50b0ffe44.jpg", + "image_caption": [ + "Figure 3. Pipeline of localizing the target object. In stage 1, given the input text description and detected object bounding boxes (bbx), we construct the first prompt asking ChatGPT the categories of target objects and anchor objects. Based on the response, the scene graph can be simplified. In stage 2, we construct the second prompt with inputs and results from stage 1, including object relations derived from the simplified scene graph. The second prompt is designed for asking ChatGPT to infer the target object. Finally, we can get the target object bounding box from the response of ChatGPT." + ], + "image_footnote": [], + "bbox": [ + 140, + 89, + 831, + 468 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "noising step $t$ . The formal definition is:", + "bbox": [ + 76, + 575, + 339, + 590 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {0}\\right) = \\mathcal {N} \\left(\\mathbf {x} _ {t}; \\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {x} _ {0}, (1 - \\bar {\\alpha} _ {t}) \\mathbf {I}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 603, + 468, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\bar{\\alpha}_t$ are constants with monotonically decreasing schedule. When $\\bar{\\alpha}_t$ is small enough, we can approximate $\\mathbf{x}_T\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ . In our context, we use conditional diffusion models like [64, 75]. The training loss is defined as:", + "bbox": [ + 76, + 633, + 468, + 695 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathrm {E} _ {t \\in [ 1, T ], \\mathbf {x} _ {0} \\sim q (\\mathbf {x} _ {0})} \\left[ \\| \\mathbf {x} _ {0} - G (\\mathbf {x} _ {t}, t, \\mathbf {C}) \\| \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 124, + 708, + 468, + 726 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $G$ is the generative model, and $\\mathbf{C}$ is the condition.", + "bbox": [ + 76, + 737, + 452, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 76, + 768, + 168, + 782 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The overview of our method is shown in Fig. 2. In Sec. 4.1, we leverage ChatGPT to localize the target object given a 3D scene and a textual description. Based on the accurate localization, we can focus on the target object and employ an object-centric generation pipeline to separately synthesize trajectories (Sec. 4.2) and motions (Sec. 4.3). Implementation details are discussed in Sec. 4.4.", + "bbox": [ + 75, + 794, + 470, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Language grounding of objects in 3D scenes", + "text_level": 1, + "bbox": [ + 500, + 574, + 872, + 592 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Scene-aware motion generation from textual descriptions requires the scene-understanding ability and build the relationship between scenes and texts. Since the instructions describe how the character moves and interacts with a single target object, the majority of the scene might bear negligible relevance to the final motions [51]. Motivated by this, we propose to locate the target object to identify the most pertinent information. Inspired by the recent progress of LLM [42, 52, 53], ChatGPT [52] is utilized to find the specific objects in the given text. We first obtain textual descriptions of scenes by building scene graphs. Then we feed them with text instructions to ChatGPT with specially designed prompts and parse the response to get target objects.", + "bbox": [ + 496, + 603, + 893, + 800 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Spatial scene graph extraction. To utilize LLM, the initial step involves converting a 3D scene into text. This is achieved by building a spatial scene graph. We utilize a pre-trained 3D object detection model in [47] to provide 3D box proposals. Subsequently, we follow the approach of [1] to obtain the relationships between objects. Specifically, for", + "bbox": [ + 496, + 809, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1858", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c41b916797d7a3209a895dd23392fcf0cc621003b8abd6c000d825dfa04cb7aa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 127, + 93, + 254, + 191 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d0b86e10e8804b71caabf3e792b51b9afa8c09f6f9d3abeb8812bc199a089d9b.jpg", + "image_caption": [ + "(b) Target sensor" + ], + "image_footnote": [], + "bbox": [ + 302, + 94, + 426, + 191 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0e4333041ce6956879ecff3d96021dceeebb273bfd1c9d4af66bf4f8c9aba3c5.jpg", + "image_caption": [ + "(a) Target object bbx", + "(c) Environment sensor", + "Figure 4. The visualization of the environment sensor, target sensor, and trajectory sensor. The target sensor (b) gives detailed geometry of the target object. The environment sensor (c) gives coarse spatial information around the target object. The trajectory sensor (d) is located around the human." + ], + "image_footnote": [], + "bbox": [ + 94, + 210, + 251, + 325 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/213367d8f1ff555c0fc4f75d1e3512092ba8160ed76b9d3897bbe6c2066926c4.jpg", + "image_caption": [ + "(d) Traj sensor" + ], + "image_footnote": [], + "bbox": [ + 303, + 220, + 426, + 318 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "every set of two objects, we can infer their relationships in three types from their bounding boxes: horizontal proximity, vertical proximity, and support as mentioned in Sec. 3; for every set of three objects, we can infer if one of them is in between of the other two objects. As detection results do not contain object poses, we do not construct allocentric relations (e.g., \"the shelf that is behind the sofa\"). Then, we build a scene graph, where each object is assigned as a node in the graph, and edges between nodes represent the relationships between objects. By converting 3D scenes into text, we can apply ChatGPT to extract meaningful insights from the data.", + "bbox": [ + 75, + 452, + 468, + 632 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Leveraging ChatGPT to localize the target object. A simple approach is to directly input the entire scene graph into ChatGPT and ask it to select the target objects. However, scenes may contain many objects, resulting in extremely long textural descriptions. We observe that ChatGPT is often confused in such settings and fails to respond with the right answer. To narrow the search space, we first employ ChatGPT to recognize objects that correlate to the provided text. We then exclude the unrelated objects from the scene graph, focusing solely on those with pertinent information, enabling us to pinpoint the target object effectively. This approach has the advantage of reducing the number of objects that need to be considered, making it easier for ChatGPT to identify the target object.", + "bbox": [ + 75, + 641, + 467, + 852 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 3, we construct two prompts in sequence. Take \"sit on the chair that is in the middle of the board and the end table\" as an example, we first need to nar", + "bbox": [ + 75, + 854, + 467, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "row down the object category search space. In order to find which type of objects we care about, we construct the first prompt that asks ChatGPT to find out target objects and anchor objects. Target object is defined as the final object that we want the agent to interact with, which in this case, is the \"table\". Anchor objects are the objects that help with determining the target object, for there might be many chairs in one scene. Based on the target object \"chair\" and anchor objects \"board\" and \"end table\", we can filter out all the unrelated objects in the scene, only keeping chairs, the board, and the end table in our scene graph. Next, according to the simplified scene graph, we can describe object relations in text: every edge in the scene graph could be converted to an edge sentence like \"chair 4 is far from the end table 0\". Converting all edges to such edge sentences gives a full description of the current scene. Finally, we construct a second prompt by asking ChatGPT to infer the target object from the accumulated edge sentences.", + "bbox": [ + 496, + 90, + 890, + 363 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Diffusion-based trajectory generation", + "text_level": 1, + "bbox": [ + 500, + 369, + 823, + 386 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the localized object from ChatGPT, we first generate the trajectory based on the instructions and then synthesize local human poses. Trajectory is defined as a sequence of characters' translations and orientations. As suggested by [51], not every point of the scene is relevant to the final human motions. Inspired by NSM [70] and ManipNet [92], we employ volumetric sensors (as shown in Fig. 4) around the target object to represent the scene.", + "bbox": [ + 496, + 393, + 890, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Object-centric scene representation. Denote the target object center location as $c_{o} = (c_{x},c_{y},c_{z})$ . We transform the point cloud of the scene $S$ to an object-centered coordinate axis centered at $c_{o}$ . To recognize the surrounding geometry of the target object, we create volumetric sensors called Environment Sensor and Target Sensor.", + "bbox": [ + 496, + 518, + 890, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Environment sensor. The environment sensor is centered at $c_{o}$ in a cubic shape with a volume of $4 \\times 4 \\times 4 \\, m^{3}$ , containing $8 \\times 8 \\times 8$ cubic voxels as shown in Fig. 4 (c). It is constructed by collecting all occupancies $o_{s}$ , center positions $c_{v}$ , and normal directions $n_{v}$ of each voxel to form a feature vector $E$ . Like [70], the scene occupancy $o_{s} \\in R^{1}$ in each voxel is defined based on the given scene mesh:", + "bbox": [ + 496, + 614, + 890, + 720 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\no _ {s} = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} d _ {s} < 0, \\\\ 0 & \\text {i f} d _ {s} > a _ {s}, \\\\ 1 - \\frac {d _ {s}}{a _ {s}} & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 728, + 890, + 785 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $d_{s}$ is the signed distance between the scene mesh and the voxel center, and $a_{s}$ is the voxel edge length. $n_v$ is the normal direction of the closest scene point to the voxel center. The environment sensor provides coarse scene geometry around the target object.", + "bbox": [ + 496, + 789, + 890, + 866 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Target sensor. To capture the detailed geometry of the target object, we further build a target sensor. As we already", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1859", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "obtain the 3D bounding box of the target object in Sec. 4.1, we crop the point clouds according to the bounding box and construct an $8 \\times 8 \\times 8$ cubic volumes that cover the bounding box, as shown in Fig. 4 (b). Target sensor $T$ is in the same form as the environment sensor $E$ , except the target sensor has a different voxel size, as visualized in Fig. 4 (b) and (c).", + "bbox": [ + 75, + 90, + 467, + 181 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given the constructed object-centric representations, we follow [57] to employ a transformer decoder architecture [79], which enables arbitrary length motions. As for the text input, we use CLIP [62] text encoder to encode input text to the text feature $L$ . The time-step $t$ is injected into the decoder in sinusoidal position embeddings form [79]. In summary, the condition for this generation model is", + "bbox": [ + 75, + 181, + 467, + 286 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {C} _ {\\mathbf {t}} = \\{L, E, T \\}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 297, + 467, + 314 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $L$ is the text feature, $E$ is the environment sensor, and $T$ is the target sensor. All the conditions are projected to the same dimension $D = 512$ by MLPs and summed with positional embeddings to form tokens. We use the simple objective described in Eq. 2 to train the trajectory generation model $G_{r}$ to generate the trajectory $\\mathbf{r}_{1:N}$ with length $N$ .", + "bbox": [ + 75, + 325, + 467, + 415 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Diffusion-based motion completion", + "text_level": 1, + "bbox": [ + 76, + 424, + 380, + 440 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given the trajectory from Sec. 4.2, the next step is to complete the whole motion. Based on the generated trajectories, we construct a Trajectory Sensor $O$ to explicitly reason about the interaction between the character and scenes.", + "bbox": [ + 75, + 448, + 467, + 506 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Trajectory sensor. The trajectory sensor is also a volumetric sensor which is similar to the form of the environment sensor but is designed for ego-centric perception [95] as shown in Fig. 4 (d). Specifically, trajectory sensors are positioned around characters in each frame. This sensor $O_{i}$ is centered at the predicted root position of the $i$ -th frame and faces to the $i$ -th frame's predicted root orientation, containing $8 \\times 8 \\times 8$ cubic voxels that store scene occupancy. The occupancy calculation is the same as Eq. 3.", + "bbox": [ + 75, + 513, + 467, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Another transformer-based conditional diffusion model is used to synthesize local poses along trajectories. The condition is defined as:", + "bbox": [ + 75, + 648, + 467, + 691 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {C} _ {\\mathbf {m}} = \\{L, E, T, O _ {1}, \\dots , O _ {N} \\}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 705, + 467, + 720 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $L, E,$ and $T$ have the same meanings in Eq. 4. Simple objective described in Eq. 2 is used to train the motion generation model $G_{m}$ to generate global orientation $\\gamma_{1:N}$ and local poses $\\theta_{1:N}$ with length $N$ .", + "bbox": [ + 75, + 731, + 467, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Implementation details", + "text_level": 1, + "bbox": [ + 76, + 801, + 290, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following [38, 57, 80, 81], we use separate diffusion models for the trajectory generation and the motion generation. Because the HUMANISE dataset [83] contains a relatively small number (51 minutes) of pure motion samples from the AMASS dataset (3772 minutes) [49], we first pre-train the", + "bbox": [ + 75, + 824, + 467, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "models on the whole AMASS dataset for 200 epochs and then fine-tune on the HUMANISE dataset for 200 epochs to improve the motion quality. During pretraining, the text feature $L$ is set to all zeros. Both models are trained with the AdamW optimizer [40], using a learning rate of 0.0001 on a single Nvidia RTX 3090 GPU. The batch size is set to 128. The version of ChatGPT is gpt-3.5-turbo. More details can be found in the supplementary material.", + "bbox": [ + 496, + 90, + 890, + 212 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 227, + 630, + 243 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Evaluation metrics", + "text_level": 1, + "bbox": [ + 500, + 252, + 679, + 266 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate the generated motions in three aspects: scene-conditional, action-conditional, and pure motion quality.", + "bbox": [ + 496, + 276, + 888, + 306 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Scene-conditional motion quality. To evaluate how the generated motion is aligned with the scene, we calculate the body-to-goal distance (goal dist.) [83] to measure how accurately the character interacts with target objects. However, goal dist. does not consider the consistency of the entire motion and scene (e.g., sitting on the sofa with incorrect orientation). To compensate for goal dist., a human perceptual study (indicated by scene score) is performed by randomly sampling 20 scenarios for each model, where ten workers are required to score each sample.", + "bbox": [ + 496, + 311, + 890, + 462 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Action-conditional motion quality. To measure how the generated motion is aligned with the text, we follow [17] to evaluate action recognition accuracy (accuracy), diversity, and multimodality of the results. Calculating these metrics relies on a pre-trained action recognition model [87] and we train the recognition model on the HUMANISE dataset.", + "bbox": [ + 496, + 468, + 890, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pure motion quality. We evaluate the realism of generated motions using the metric suggested by [17], namely Frechet Inception Distance (FID). A lower FID indicates that generated motions are closer to the groundtruth motions. We also perform a human perceptual study (indicated by quality score) to measure pure motion quality.", + "bbox": [ + 496, + 564, + 890, + 654 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Generating human motion from text and scene", + "text_level": 1, + "bbox": [ + 500, + 665, + 890, + 680 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our experiments are conducted on the HUMANISE dataset [83]. Following the previous setting [83], there are $16.5k$ motion sequences in 543 scenes for training and $3.1k$ motion sequences in 100 scenes for testing. We compare our method with four baselines: (1) $\\mathbf{MDM}^*$ [75]: a diffusion-based motion generator. (2) $\\mathbf{GMD}^*$ [38]: GMD proposes various techniques for enhancing the control and quality of MDM. We employ the single-stage setting of $\\mathbf{GMD}^*$ . Similar to HUMANISE, we use a point-transformer to provide scene features for $\\mathbf{MDM}^*$ and $\\mathbf{GMD}^*$ . (3) $\\mathbf{GMD}^{\\mathbf{HC}}$ : we provide object centers predicted by HUMANISE (denoted by HC) to guide the motion generation process in GMD by adding a proximity loss to encourage the motion to be close to the predicted object center. The proximity loss is defined", + "bbox": [ + 496, + 688, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1860", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/73248fd43b5552d911fa93068a4ac3022915e6ec820bca60160882b34b132b98.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsScene-conditionalAction-conditionalPure motion quality
goal dist.↓scene score↑accuracy↑diversity→multimodality→FID↓quality score↑
Real0.014-99.1%4.822.280.00-
MDM*2.0481.9996.0%4.832.570.162.45
GMD*1.2292.5096.6%4.912.280.242.64
GMDHC1.1302.4996.4%4.932.430.232.76
HUMANISE0.9952.3389.7%4.252.661.112.81
Ours0.3843.5497.1%4.782.310.123.50
", + "bbox": [ + 142, + 88, + 831, + 220 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Quantitative results on the HUMANISE dataset. We compare our method with four baselines (please refer to Sec. 5.2) and the real data. $\\mathrm{GMD}^{\\mathrm{HC}}$ means using HUMANISE's predicted center to guide motion generation in $\\mathrm{GMD^{*}}$ . Among the metrics, scene score and quality score are perceptual studies. $\\uparrow$ means higher is better and $\\downarrow$ means lower is better. $\\rightarrow$ means closer to the real data is better. Bold indicates the best results.underline indicates the second best.", + "bbox": [ + 75, + 229, + 893, + 287 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/24ed49104be155df0172295846b1e03d3e6fe6329a55c141d2c34fae0ea45a7e.jpg", + "image_caption": [ + "Figure 5. Qualitative results. We compare our method with groundtruth and four baselines (please refer to Sec. 5.2) given the same text descriptions. Our method synthesizes motions that interact with the object precisely as the groundtruth data while the baselines fail." + ], + "image_footnote": [], + "bbox": [ + 81, + 299, + 890, + 491 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "as the distance from HC to the predicted human pelvis at the interacting frame. (4) HUMANISE: we directly use their released models.", + "bbox": [ + 75, + 555, + 470, + 599 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The quantitative results are shown in Tab. 1. Our method outperforms the baseline in terms of goal dist., scene score, accuracy, FID, and quality score and achieves competitive results in diversity and multimodality. The quantitative results show that our approach can generate better-quality motions and are more consistent with all the conditions. The qualitative results are demonstrated in Fig. 5. Please refer to the supplementary material for more visualizations.", + "bbox": [ + 75, + 601, + 468, + 723 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation study", + "text_level": 1, + "bbox": [ + 76, + 734, + 227, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation of main components. Four variants are constructed to explore the effect of our design choice. (1) \"w/o localization\": the localization module is removed and trajectories and motions are generated directly using scene point clouds. Scene features are given as the same form of MDM* and GMD*. (2) \"w/o object-centric\": we remove our object-centric representation and predict motions in the scene coordinate. The scene features of our sensors are still employed. (3) \"w/o two-stage\": trajectories and", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "motions are generated together. (4) \"w/o diffusion\": we employ cVAE instead of the diffusion model as the architecture. (5) \"w/o pretrain\": the models are not pre-trained on the AMASS dataset. As shown in Tab. 2, the localization enhances the precise interaction ability and the object-centric representation improves motion quality. Although the (\"w/o two-stage\") is more efficient and has competitive performance metrics comparing to our method, it is prone to collapse directly into the target position, lacking gradual transition and realism.", + "bbox": [ + 496, + 555, + 892, + 707 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The design choice of object localization. For object localization, we compare our method with two baselines and five variants. (1) HUMANISE: predicts the object center in the auxiliary task. (2) BUTD-DETR [34]: is a 3D visual grounding method. (3) “ours w/o two-stage”: we employ a one-stage question-answering process. (4) “ours w/o few-shot”: the few-shot examples are not provided. (5) “ours using text matching”: we compute CLIP similarities instead of using ChatGPT. (6) “ours using LLaMA”: ChatGPT is replaced by LLaMA 2 7B model. (7) “ours using Mistral”: ChatGPT is replaced by Mistral 7B [35], an open-source LLM. The prompts for LLaMA 2 and Mistral", + "bbox": [ + 496, + 719, + 893, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1861", + "bbox": [ + 483, + 944, + 513, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f12a4e427bdfb63e92e30a435b669bb51ab543226d4a2883839550df5783b546.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
VariantsSceneActionPure
goal dist.↓accuracy↑diversity→mm→FID↓
Real0.01499.1%4.822.280.00
w/o localization0.59274.5%3.712.753.14
w/o object-centric0.41390.6%4.132.631.01
w/o two-stage0.38597.2%4.932.330.14
w/o diffusion0.39040.0%3.693.863.61
w/o pretrain0.39282.6%3.882.372.50
Ours0.38497.1%4.782.310.12
", + "bbox": [ + 81, + 88, + 467, + 210 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7B are slightly adjusted. We evaluate these variants under two scenarios: predicted detection [47] and groundtruth detection. The metrics include the accuracy (acc.) and the distance from predicted centers to the object center (center dist.). \"acc.\" is the percentage of times with an IoU (Intersection over Union) higher than the threshold (0.25) following [34]. The results are shown in Tab. 3. Since the results of Mistral are only slightly behind of ChatGPT, ChatGPT can be replaced by Mistral for better reproducibility, but cannot be replaced by the text matching method. Moreover, if groundtruth detection is provided, our method can achieve even better results.", + "bbox": [ + 75, + 304, + 468, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The design choice of sensor density. Please refer to the supplementary material.", + "bbox": [ + 76, + 491, + 468, + 522 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Generalization", + "text_level": 1, + "bbox": [ + 76, + 532, + 227, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To validate the generalization ability of our method, we run our method directly on the unseen PROX dataset [22] without fine-tuning. We provide illustrations in Fig. 6. With our localization method based on ChatGPT and motion generation method based on volumetric sensors, our pipeline can easily generalize to other datasets. Please refer to the supplementary material for more results.", + "bbox": [ + 75, + 556, + 468, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Discussion", + "text_level": 1, + "bbox": [ + 76, + 678, + 191, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have demonstrated that our approach could synthesize motions with better quality and more precise interactions. However, restricted by the dataset, the duration of motions is relatively short (60% are around 1-3s), and most texts follow the template [1] without detailed descriptions. We only tackle static scenes and extending our method to interact with moving objects is a future direction. We acknowledge that leveraging LLMs has several problems. ChatGPT may fail when the detector misses the objects and the behavior of ChatGPT is regulated by prompts [37]. Despite we have shown in Tab. 3 that ChatGPT can be replaced by an open source LLM Mistral 7B, using LLMs instead of classical parsing approaches is less efficient in the inference stage.", + "bbox": [ + 75, + 703, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b609ce663ef4d30c04f4a988daf3c7807791a05064d63499bd468c7a400dd92d.jpg", + "table_caption": [ + "Table 2. Ablation of main components. We compare our method with five variants (please refer to Sec. 5.3). Among them, $mm$ indicates multimodality. Bold indicates the best results. Underline indicates the second best." + ], + "table_footnote": [], + "table_body": "
VariantsPredicted detectionGT detection
acc. ↑center dist.↓acc. ↑center dist.↓
HUMANISE-1.48--
BUTD-DETR62.51.3363.11.23
Ours w/o two-stage49.01.3458.51.10
Ours w/o few-shot69.30.8086.30.45
Ours using text matching67.00.8979.70.52
Ours using LLaMA 240.31.5542.51.52
Ours using Mistral 7B72.40.7487.20.33
Ours using ChatGPT-3.575.60.6190.50.37
", + "bbox": [ + 504, + 88, + 890, + 231 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Design choices of object localization. We compare our method with two baselines and five variants (please refer to Sec. 5.3). Since HUMANISE directly regresses the coordinate of centers without utilizing groundtruth detection, we do not calculate acc. and only include their results under predicted detection. Bold indicates the best results.", + "bbox": [ + 496, + 241, + 893, + 324 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9cb3d200c0b4f617df9ff313c8771453e220e0b467a3816bf5906c45bfb74fb4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 521, + 349, + 694, + 457 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4e37dd5c6339ff292e5995d045529d6ddf1c2f70df48c51cffe23d69863af90e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 349, + 870, + 455 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3ad9ea70bd24639ce0169e49bf9d1b94fda76fffa27572b243d415862b60859b.jpg", + "image_caption": [ + "Figure 6. Qualitative results of our method on the PROX dataset. We run our method on the scenes from the PROX dataset without fine-tuning. Results show that our method is capable to generalize to unseen scenes and objects." + ], + "image_footnote": [], + "bbox": [ + 519, + 458, + 694, + 568 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0bb381f1c3adbf42f2abb0023b608d75552d245318cc5c6225472174caf46fea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 458, + 870, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 669, + 617, + 684 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we introduce a novel method for generating motion from text in a scene. To tackle this problem, we propose a two-step approach. The first step involves 3D visual grounding, where we identify the target object. In the second step, concentrating on the target object, we build a diffusion-based motion generation method. Our approach offers several advantages, including improved 3D visual grounding accuracy and motion quality.", + "bbox": [ + 496, + 699, + 890, + 820 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 830, + 660, + 848 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was partially supported by the NSFC (No. 62172364), Ant Group and Information Technology Center and State Key Lab of CAD&CG, Zhejiang University.", + "bbox": [ + 496, + 854, + 890, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1862", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Panos Achlioptas, Ahmed Abdelreheem, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In ECCV, 2020. 2, 3, 4, 8", + "[2] Chaitanya Ahuja and Louis-Philippe Morency. Language2pose: Natural language grounded pose forecasting. In 3DV, 2019. 1, 2", + "[3] Nikos Athanasiou, Mathis Petrovich, Michael J. Black, and Gül Varol. TEACH: Temporal Action Compositions for 3D Humans. In 3DV, 2022. 1, 2", + "[4] Nikos Athanasiou, Mathis Petrovich, Michael J. Black, and Gül Varol. SINC: Spatial composition of 3D human motions for simultaneous action generation. ICCV, 2023. 2", + "[5] German Barquero, Sergio Escalera, and Cristina Palmero. Belfusion: Latent diffusion for behavior-driven human motion prediction. In ICCV, 2023. 2", + "[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In NeurIPS, 2020. 2, 3", + "[7] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 3", + "[8] Yu-Wei Chao, Jimei Yang, Weifeng Chen, and Jia Deng. Learning to sit: Synthesizing human-chair interactions via hierarchical control. In AAAI, 2021. 2", + "[9] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In ECCV, 2020. 3", + "[10] Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, et al. Evaluating large language models trained on code. arXiv preprint, 2021. 3", + "[11] Yixin Chen, Qing Li, Deqian Kong, Yik Lun Kei, Song-Chun Zhu, Tao Gao, Yixin Zhu, and Siyuan Huang. Yourfit: Embodied reference understanding with language and gesture. In ICCV, 2021. 3", + "[12] Rishabh Dabral, Muhammad Hamza Mughal, Vladislav Golyanik, and Christian Theobalt. Mofusion: A framework for denoising-diffusion-based motion synthesis. In CVPR, 2023. 2", + "[13] Abhishek Das, Georgia Gkioxari, Stefan Lee, Devi Parikh, and Dhruv Batra. Neural Modular Control for Embodied Question Answering. In CoRL, 2018. 3", + "[14] Katerina Fragkiadaki, Sergey Levine, Panna Felsen, and Jitendra Malik. Recurrent network models for human dynamics. In ICCV, 2015. 2" + ], + "bbox": [ + 86, + 114, + 467, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Anindita Ghosh, Noshaba Cheema, Cennet Oguz, Christian Theobalt, and Philipp Slusallek. Synthesis of compositional animations from textual descriptions. In ICCV, 2021. 1, 2", + "[16] Anindita Ghosh, Rishabh Dabral, Vladislav Golyanik, Christian Theobalt, and Philipp Slusallek. Imos: Intent-driven full-body motion synthesis for human-object interactions. In Computer Graphics Forum, 2023. 2", + "[17] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3d human motions. In ACM MM, 2020. 6", + "[18] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In CVPR, 2022. 1, 2", + "[19] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In ECCV, 2022. 1, 2", + "[20] Zoey Guo, Yiwen Tang, Ray Zhang, Dong Wang, Zhigang Wang, Bin Zhao, and Xuelong Li. Viewrefer: Grasp the multi-view knowledge for 3d visual grounding. In ICCV, 2023. 3", + "[21] Félix G Harvey, Mike Yurick, Derek Nowrouzezahrai, and Christopher Pal. Robust motion in-betweening. ACM Trans. Graph., 2020. 2", + "[22] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J. Black. Resolving 3d human pose ambiguities with 3d scene constraints. In ICCV, 2019. 2, 8", + "[23] Mohamed Hassan, Duygu Ceylan, Ruben Villegas, Jun Saito, Jimei Yang, Yi Zhou, and Michael J. Black. Stochastic scene-aware motion prediction. In ICCV, 2021. 2", + "[24] Mohamed Hassan, Yunrong Guo, Tingwu Wang, Michael Black, Sanja Fidler, and Xue Bin Peng. Synthesizing physical character-scene interactions. In SIGGRAPH, 2023. 2", + "[25] Chengan He, Jun Saito, James Zachary, Holly Rushmeier, and Yi Zhou. Nemf: Neural motion fields for kinematic animation. In NeurIPS, 2022. 2", + "[26] Gustav Eje Henter, Simon Alexanderson, and Jonas Beskow. Moglow: Probabilistic and controllable motion synthesis using normalising flows. ACM Trans. Graph., 2020. 2", + "[27] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3", + "[28] Daniel Holden, Jun Saito, and Taku Komura. A deep learning framework for character motion synthesis and editing. ACM Trans. Graph., 2016. 2", + "[29] Daniel Holden, Taku Komura, and Jun Saito. Phase-functional neural networks for character control. ACM Trans. Graph., 2017. 1, 2", + "[30] Joy Hsu, Jiayuan Mao, and Jiajun Wu. Ns3d: Neurosymbolic grounding of 3d objects and relations. In CVPR, 2023. 3", + "[31] Pin-Hao Huang, Han-Hung Lee, Hwann-Tzong Chen, and Tyng-Luh Liu. Text-guided graph neural networks for referring 3d instance segmentation. AAAI, 2021. 3", + "[32] Shijia Huang, Yilun Chen, Jiaya Jia, and Liwei Wang. Multi-view transformer for 3d visual grounding. In CVPR, 2022. 3" + ], + "bbox": [ + 509, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1863", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Siyuan Huang, Zan Wang, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, and Song-Chun Zhu. Diffusion-based generation, optimization, and planning in 3d scenes. In CVPR, 2023. 3", + "[34] Ayush Jain, Nikolaos Gkanatsios, Ishita Mediratta, and Katerina Fragkiadaki. Bottom up top down detection transformers for language grounding in images and point clouds. In ECCV, 2022. 3, 7, 8", + "[35] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 7", + "[36] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. arXiv preprint arXiv:2306.14795, 2023. 2", + "[37] Sai Shashank Kalakonda, Shubh Maheshwari, and Ravi Kiran Sarvadevabhatla. Action-gpt: Leveraging large-scale language models for improved and generalized action generation. In ICME, 2023. 2, 8", + "[38] Korrawe Karunratanakul, Konpat Preechakul, Supasorn Suwajanakorn, and Siyu Tang. Guided motion diffusion for controllable human motion synthesis. In ICCV, 2023. 2, 3, 6", + "[39] Jihoon Kim, Jiseob Kim, and Sungjoon Choi. Flame: Freeform language-based motion synthesis & editing. In AAAI, 2023. 1, 2", + "[40] Diederik P. Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization. arXiv e-prints, 2014. 6", + "[41] Diederik P Kingma and Max Welling. Auto-Encoding Variational Bayes. arXiv e-prints, 2013. 2", + "[42] Takeshi Kojima, Shixiang (Shane) Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners. In NeurIPS, 2022. 2, 4", + "[43] Hanyang Kong, Kehong Gong, Dongze Lian, Michael Bi Mi, and Xinchao Wang. Priority-centric human motion generation in discrete latent space. In ICCV, 2023. 2", + "[44] Peizhuo Li, Kfir Aberman, Zihan Zhang, Rana Hanocka, and Olga Sorkine-Hornung. Ganimator: Neural motion synthesis from a single sequence. ACM Trans. Graph., 2022. 2", + "[45] Quanzhou Li, Jingbo Wang, Chen Change Loy, and Bo Dai. Task-oriented human-object interactions generation with implicit neural representations. arXiv preprint, 2023. 2", + "[46] Hung Yu Ling, Fabio Zinno, George Cheng, and Michiel Van De Panne. Character controllers using motion vaes. ACM Trans. Graph., 2020. 2", + "[47] Ze Liu, Zheng Zhang, Yue Cao, Han Hu, and Xin Tong. Group-free 3d object detection via transformers. In ICCV, 2021. 3, 4, 8", + "[48] Junyu Luo, Jiahui Fu, Xianghao Kong, Chen Gao, Haibing Ren, Hao Shen, Huaxia Xia, and Si Liu. 3d-sps: Single-stage 3d visual grounding via referred point progressive selection. In CVPR, 2022. 2, 3", + "[49] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In ICCV, 2019. 2, 6" + ], + "bbox": [ + 86, + 92, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Julieta Martinez, Michael J. Black, and Javier Romero. On human motion prediction using recurrent neural networks. In CVPR, 2017. 2", + "[51] Aymen Mir, Xavier Puig, Angjoo Kanazawa, and Gerard Pons-Moll. Generating continual human motion in diverse 3d scenes. arXiv preprint, 2023. 2, 3, 4, 5", + "[52] OpenAI. Openai: Introducing chatgpt. https://openai.com/blog/chatgpt, 2022. 2, 3, 4", + "[53] OpenAI. Gpt-4 technical report, 2023. 2, 4", + "[54] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3d hands, face, and body from a single image. In CVPR, 2019. 3", + "[55] Mathis Petrovich, Michael J. Black, and Gül Varol. Action-conditioned 3D human motion synthesis with transformer VAE. In ICCV, 2021. 2", + "[56] Mathis Petrovich, Michael J. Black, and Gül Varol. TEMOS: Generating diverse human motions from textual descriptions. In ECCV, 2022. 1, 2", + "[57] Huaijin Pi, Sida Peng, Minghui Yang, Xiaowei Zhou, and Hujun Bao. Hierarchical generation of human-object interactions with diffusion probabilistic models. In ICCV, 2023. 2, 6", + "[58] Matthias Plappert, Christian Mandery, and Tamim Asfour. The kit motion-language dataset. *Big data*, 2016. 2", + "[59] Abhinanda R. Punnakkal, Arjun Chandrasekaran, Nikos Athanasiou, Alejandra Quiros-Ramirez, and Michael J. Black. Babel: Bodies, action and behavior with english labels. In CVPR, 2021. 2", + "[60] Yijun Qian, Jack Urbanek, Alexander G Hauptmann, and Jungdam Won. Breaking the limits of text-conditioned 3d motion synthesis with elaborative descriptions. In ICCV, 2023. 2", + "[61] Sigal Raab, Inbal Leibovitch, Peizhuo Li, Kfir Aberman, Olga Sorkine-Hornung, and Daniel Cohen-Or. Modi: Unconditional motion synthesis from diverse data. In CVPR, 2023. 2", + "[62] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 6", + "[63] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485-5551, 2020. 2", + "[64] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical Text-Conditional Image Generation with CLIP Latents. arXiv e-prints, 2022. 4", + "[65] Ali Razavi, Aaron van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. In NeurIPS, 2019. 2", + "[66] Junha Roh, Karthik Desingh, Ali Farhadi, and Dieter Fox. Languagerefer: Spatial-language model for 3d visual grounding. In CoRL, 2022. 3" + ], + "bbox": [ + 509, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1864", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[67] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2", + "[68] Yonatan Shafir, Guy Tevet, Roy Kapon, and Amit H Bermano. Human motion diffusion as a generative prior. arXiv preprint, 2023. 2", + "[69] Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. In NeurIPS, 2015. 2", + "[70] Sebastian Starke, He Zhang, Taku Komura, and Jun Saito. Neural state machine for character-scene interactions. ACM Trans. Graph., 2019. 2, 5", + "[71] Sebastian Starke, Yiwei Zhao, Taku Komura, and Kazi Zaman. Local motion phases for learning multi-contact character movements. ACM Trans. Graph., 2020. 2", + "[72] Sebastian Starke, Ian Mason, and Taku Komura. Deepphase: Periodic autoencoders for learning motion phase manifolds. ACM Trans. Graph., 2022. 2", + "[73] Omid Taheri, Vasileios Choutas, Michael J. Black, and Dimitrios Tzionas. Goal: Generating 4d whole-body motion for hand-object grasping. In CVPR, 2022. 2", + "[74] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. In ECCV, 2022. 1, 2", + "[75] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In ICLR, 2023. 1, 2, 4, 6", + "[76] Jesse Thomason, Mohit Shridhar, Yonatan Bisk, Chris Paxton, and Luke Zettlemoyer. Language grounding with 3d objects. In CoRL, 2022. 3", + "[77] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models. arXiv preprint, 2023. 2", + "[78] Aaron van den Oord, Oriol Vinyals, and koray kavukcuoglu. Neural discrete representation learning. In NeurIPS, 2017. 2", + "[79] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 6", + "[80] Jiashun Wang, Huazhe Xu, Jingwei Xu, Sifei Liu, and Xiaolong Wang. Synthesizing long-term 3d human motion and interaction in 3d scenes. In CVPR, 2021. 2, 6", + "[81] Jingbo Wang, Yu Rong, Jingyuan Liu, Sijie Yan, Dahua Lin, and Bo Dai. Towards diverse and natural scene-aware 3d human motion synthesis. In CVPR, 2022. 2, 6", + "[82] Xi Wang, Gen Li, Yen-Ling Kuo, Muhammed Kocabas, Emre Aksan, and Otmar Hilliges. Reconstructing action-conditioned human-object interactions using commonsense knowledge priors. In 3DV, 2022. 2", + "[83] Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, Wei Liang, and Siyuan Huang. Humanise: Language-conditioned human motion generation in 3d scenes. In NeurIPS, 2022. 2, 3, 6" + ], + "bbox": [ + 86, + 92, + 468, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[84] Yan Wu, Jiahao Wang, Yan Zhang, Siwei Zhang, Otmar Hilliges, Fisher Yu, and Siyu Tang. Saga: Stochastic whole-body grasping with contact. In ECCV, 2022. 2", + "[85] Yiming Xie, Varun Jampani, Lei Zhong, Deqing Sun, and Huaizu Jiang. Omnicontrol: Control any joint at any time for human motion generation. arXiv preprint, 2023. 2", + "[86] Sirui Xu, Zhengyuan Li, Yu-Xiong Wang, and Liang-Yan Gui. Interdiff: Generating 3d human-object interactions with physics-informed diffusion. In ICCV, 2023. 2", + "[87] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In AAAI, 2018. 6", + "[88] Zhengyuan Yang, Songyang Zhang, Liwei Wang, and Jiebo Luo. Sat: 2d semantics assisted training for 3d visual grounding. In ICCV, pages 1856-1866, 2021. 2, 3", + "[89] Shuquan Ye, Dongdong Chen, Songfang Han, and Jing Liao. 3d question answering. IEEE Transactions on Visualization and Computer Graphics, 2022. 3", + "[90] Zhihao Yuan, Xu Yan, Yinghong Liao, Ruimao Zhang, Sheng Wang, Zhen Li, and Shuguang Cui. Instancerefer: Cooperative holistic understanding for visual grounding on point clouds through instance multi-level contextual referring. In ICCV, 2021. 3", + "[91] He Zhang, Sebastian Starke, Taku Komura, and Jun Saito. Mode-adaptive neural networks for quadruped motion control. ACM Trans. Graph., 2018. 2", + "[92] He Zhang, Yuting Ye, Takaaki Shiratori, and Taku Komura. Manipnet: Neural manipulation synthesis with a hand-object spatial representation. ACM Trans. Graph., 2021. 5", + "[93] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Yong Zhang, Hongwei Zhao, Hongtao Lu, Xi Shen, and Ying Shan. Generating human motion from textual descriptions with discrete representations. In CVPR, 2023. 2", + "[94] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint, 2022. 2", + "[95] Xiaohan Zhang, Bharat Lal Bhatnagar, Sebastian Starke, Vladimir Guzov, and Gerard Pons-Moll. Couch: Towards controllable human-chair interactions. In ECCV, 2022. 2, 6", + "[96] Yan Zhang and Siyu Tang. The wanderings of odysseus in 3d scenes. In CVPR, 2022. 2", + "[97] Yiming Zhang, ZeMing Gong, and Angel X Chang. Multi3drefer: Grounding text description to multiple 3d objects. In ICCV, 2023. 3", + "[98] Yaqi Zhang, Di Huang, Bin Liu, Shixiang Tang, Yan Lu, Lu Chen, Lei Bai, Qi Chu, Nenghai Yu, and Wanli Ouyang. Motiongpt: Finetuned llms are general-purpose motion generators. arXiv preprint, 2023. 2", + "[99] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In ICCV, 2021. 2", + "[100] Kaifeng Zhao, Shaofei Wang, Yan Zhang, Thabo Beeler, and Siyu Tang. Compositional human-scene interaction synthesis with semantic control. In ECCV, 2022. 3", + "[101] Kaifeng Zhao, Yan Zhang, Shaofei Wang, Thabo Beeler, and Siyu Tang. Synthesizing diverse human motions in 3d indoor scenes. In ICCV, 2023. 3" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1865", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[102] Lichen Zhao, Daigang Cai, Lu Sheng, and Dong Xu. 3dvg-transformer: Relation modeling for visual grounding on point clouds. In ICCV, 2021. 3", + "[103] Yang Zheng, Yanchao Yang, Kaichun Mo, Jiaman Li, Tao Yu, Yebin Liu, C Karen Liu, and Leonidas J Guibas. Gimo: Gaze-informed human motion prediction in context. In ECCV, 2022. 3", + "[104] Yi Zhou, Connelly Barnes, Lu Jingwan, Yang Jimei, and Li Hao. On the continuity of rotation representations in neural networks. In CVPR, 2019. 3" + ], + "bbox": [ + 78, + 90, + 468, + 233 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "1866", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2024/Generating Human Motion in 3D Scenes from Text Descriptions/507f8570-ad9c-455e-838f-8fa1bda75b31_model.json b/2024/Generating Human Motion in 3D Scenes from Text Descriptions/507f8570-ad9c-455e-838f-8fa1bda75b31_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4cc7803f37f6b58bc3d7b9d8bd3f371d56ee9538 --- /dev/null +++ b/2024/Generating Human Motion in 3D Scenes from Text Descriptions/507f8570-ad9c-455e-838f-8fa1bda75b31_model.json @@ -0,0 +1,2754 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.163, + 0.131, + 0.812, + 0.154 + ], + "angle": 0, + "content": "Generating Human Motion in 3D Scenes from Text Descriptions" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.18, + 0.744, + 0.248 + ], + "angle": 0, + "content": "Zhi Cen1 Huaijin Pi1 Sida Peng1† Zehong Shen1 Minghui Yang2 Shuai Zhu2 Hujun Bao1 Xiaowei Zhou1 1Zhejiang University 2Ant Group" + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.283, + 0.283, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.283, + 0.483, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.212, + 0.41, + 0.356, + 0.422 + ], + "angle": 0, + "content": "Test on the HUAMNISE" + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.283, + 0.688, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.692, + 0.283, + 0.888, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.594, + 0.41, + 0.783, + 0.422 + ], + "angle": 0, + "content": "Generalize to the PROX scenes" + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.434, + 0.895, + 0.492 + ], + "angle": 0, + "content": "Figure 1. Generating human motions in 3D scenes from text descriptions. Our method can generate human motions containing accurate human-object interactions in 3D scenes based on textural descriptions. Although our method is trained and tested on the HUMANISE dataset, it can generalize to other scenes, e.g., the scenes in the PROX dataset. Left: test results on the HUMANISE dataset. Right: generalization results on the PROX scenes." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.506, + 0.314, + 0.521 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.535, + 0.474, + 0.868 + ], + "angle": 0, + "content": "Generating human motions from textual descriptions has gained growing research interest due to its wide range of applications. However, only a few works consider human-scene interactions together with text conditions, which is crucial for visual and physical realism. This paper focuses on the task of generating human motions in 3D indoor scenes given text descriptions of the human-scene interactions. This task presents challenges due to the multimodality nature of text, scene, and motion, as well as the need for spatial reasoning. To address these challenges, we propose a new approach that decomposes the complex problem into two more manageable sub-problems: (1) language grounding of the target object and (2) object-centric motion generation. For language grounding of the target object, we leverage the power of large language models. For motion generation, we design an object-centric scene representation for the generative model to focus on the target object, thereby reducing the scene complexity and facilitating the modeling of the relationship between human motions and the object. Experiments demonstrate the better motion quality of our approach compared to baselines and validate our design choices. Code will be available at link." + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.875, + 0.47, + 0.903 + ], + "angle": 0, + "content": "The authors from Zhejiang University are affiliated with the State Key Lab of CAD&CG. † Corresponding author." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.506, + 0.633, + 0.521 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.534, + 0.892, + 0.715 + ], + "angle": 0, + "content": "Human motion generation has been a long-standing problem due to its broad range of applications such as game development, virtual reality, and movie production. Recently, this area witnessed a paradigm shift from avatar animation given rich user input [29] to learning-based motion generation from high-level language prompts, e.g. text descriptions about the desired motion [2, 3, 15, 18, 19, 39, 56, 74, 75]. However, most prior works on text-driven motion synthesis do not consider human-scene interactions [39, 56, 74, 75] while the scene context and physical constraints of the environment largely define the fidelity of the generated human motions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.717, + 0.894, + 0.793 + ], + "angle": 0, + "content": "In this paper, we focus on generating motions from text descriptions in 3D indoor scenes. Specifically, given a 3D scan of the target scene and a text description of a human action that interacts with the scene, we aim to generate natural human motions that are consistent with the text description." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.903 + ], + "angle": 0, + "content": "This problem presents several challenges, primarily due to the multi-modality nature of text, scene, and human motion. In contrast to previous methods [2, 3, 56, 74, 75] focusing solely on textual descriptions of how human moves, our task also includes texts that additionally describe the spatial details in the given scene (e.g., sit on the armchair near the desk). Therefore, this task requires spatial reason-" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1855" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "ing skills [1], where the model should build a text-object mapping to locate a specific object in 3D scenes aligned with a natural language description. In addition, the generated motions should also be coherent with scene contexts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.155, + 0.471, + 0.412 + ], + "angle": 0, + "content": "As a pioneer work, HUMANISE [83] builds a conditional variational autoencoder (cVAE) [41, 69] with separate encoders of scenes and texts for multi-modality understanding. To enable spatial reasoning ability, they introduce auxiliary tasks of directly regressing object centers to learn 3D visual grounding in an implicit manner. But they do not explicitly utilize the predicted centers and thus the inductive bias of visual grounding cannot be fully incorporated. In addition, HUMANISE [83] encodes the entire scene with a single point transformer [99]. Directly generating motions with such a model is challenging, as 3D point clouds are inherently noisy and complex [48, 88], leading to the inability to locate the target object. As suggested by [51], not every point of the scene is relevant to the final human motions. Therefore, it is also necessary to develop a more targeted approach that focuses on the relevant parts of the scene to improve the quality of motion generation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.413, + 0.47, + 0.64 + ], + "angle": 0, + "content": "To tackle the problems mentioned above, we propose a novel approach that exploits the power of the large language models (LLMs) [6, 42, 52]. Our key idea is to address the challenging task of generating motion in a scene based on textual cues by breaking it down into two smaller problems: (1) language grounding of objects in 3D scenes and (2) generating motions with a focus on the target object. For language grounding, rather than directly learning text-object mapping, we propose to formulate it as question answering and utilize the large prior knowledge of LLMs. Specifically, we first construct scene graphs of 3D scenes and generate their textual descriptions. Then we employ ChatGPT [52] to analyze the relationship between scene descriptions and input instructions and respond with 3D visual grounding answers. Experiments prove the effectiveness of this strategy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.643, + 0.47, + 0.824 + ], + "angle": 0, + "content": "For motion generation, we design an object-centric representation to help the generative model focus on the target object. Specifically, we convert point clouds around the target object into volumetric sensors [70] to build object-centric representation. Then we employ diffusion models [57, 75] to synthesize human motions given object-centric representation and texts. Compared with original scene point clouds which might have various scales, object-centric representation is more compact and robust to scales as objects in the same category are of similar size. Therefore, this representation reduces scene complexity and facilitates the modeling of the relationship between motions and objects." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "We conducted thorough comparative and ablation experiments on the HUMANISE dataset. The results demonstrate that our method outperforms the baseline, reflected in more accurate object grounding results and better motions that align with the textual descriptions and scenes. We further" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.123 + ], + "angle": 0, + "content": "show that our approach can generalize to the PROX dataset [22] without any fine-tuning." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.139, + 0.64, + 0.155 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.165, + 0.668, + 0.181 + ], + "angle": 0, + "content": "2.1. Motion synthesis" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.194, + 0.892, + 0.316 + ], + "angle": 0, + "content": "Deep learning for motion synthesis. Deep learning methods for motion synthesis have attracted increasing attention in recent years [14, 21, 25, 28, 29, 50, 71, 91]. Various techniques, including MLP [29], mixture of experts (MoE) [72, 91], and recurrent neural network (RNN) [21], are employed to tackle this task. To generate diverse motion results, previous works explore cVAE [46, 96], GANs [44, 61], normalizing flow [26], and diffusion models [75]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.321, + 0.893, + 0.624 + ], + "angle": 0, + "content": "Text-driven motion generation. Recently, there has been a growing interest in text-driven motion generation [2, 56, 74]. This task takes natural language as input and synthesizes human motions that align with instructions. KIT-ML [58] is the first benchmark of this task. Some works further annotate the AMASS dataset [49] with action labels [59] and text [18]. To tackle this task, [2, 15, 74] propose to learn a shared latent space for text and motion. [56] employs transformer VAE [55] to generate diverse results. [19, 93] achieve better performance by discrete representation with VQ-VAE [65, 78]. [12, 39, 75, 94] successfully apply diffusion models in this direction. [5, 43] further explore latent diffusion models [67]. Based on [75], [38, 85] introduce sparse spatial control. [3, 60, 68] also investigate long sequence motion generation. Other existing works [4, 37, 82, 98] leverage Large Language Models (LLM) [6, 42, 53, 77] in the motion domain. [4] construct compositional motions by combining different body parts with LLM. [36, 98] regard motions as a kind of language and finetune LLMs [63] with motion tokens [19, 93]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.629, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Scene-aware motion generation This direction is to generate human motions in a 3D scene [70, 80]. To synthesize walking and sitting motions, [70] construct volumetric sensors to encode object information and the surroundings of the character. [23] extends it to synthesize motions with diverse sitting and lying styles. Furthermore, [95] proposes to control sitting styles by hand contacts. To improve the performance, [57] employs a hierarchical framework that generates goal poses, milestones, and motion sequentially. [73, 84] synthesize whole-body (body and hand fingers) grasping motions with small objects. [16, 45, 71, 86] explore interactions with dynamic objects including manipulation [16, 45] and carrying [86]. [8, 24] explore human-object interactions with physically simulated characters. These works mainly focus on one or two objects while others [80, 81] consider a more complex scene input (e.g., point clouds of the scene including walls and floors.). [80, 81] adopt hierarchical frameworks and generate trajectory and" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1856" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.123, + 0.089, + 0.844, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.241, + 0.895, + 0.286 + ], + "angle": 0, + "content": "Figure 2. Overview of our two-stage pipeline. In the first stage, given an input scene and a text description (a), we use ChatGPT to locate the target object (b). In the second stage, human motions are synthesized by first producing human trajectories (c) and then generating local poses (d)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.31, + 0.473, + 0.4 + ], + "angle": 0, + "content": "motion separately. [103] introduces gaze to help generation. [33] further employs diffusion models and [51] synthesizes very long-term motions in scenes by dividing long sequences into several short sequences. [101] designs a reinforcement learning pipeline to enable navigation in a complex scene and interaction style control." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.408, + 0.47, + 0.619 + ], + "angle": 0, + "content": "Text-driven scene-aware motion generation Only a few works consider text and scenes simultaneously [38, 83, 101]. Although [101] enables text control of sitting styles by [100], the text descriptions in our setting are used to select an object in a cluttered scene. [38] could avoid obstacles during walking while our task needs to handle various actions. The most relevant work to us is HUMANISE [83], which employs a transformer VAE architecture with a two-stream condition module for text and scenes. To accurately localize target objects, they design auxiliary tasks like directly regressing object centers. In contrast to HUMANISE, we propose a two-stage pipeline where we first localize the target object with the help of ChatGPT [52] and then generate human motion using the object-centric representation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.634, + 0.273, + 0.65 + ], + "angle": 0, + "content": "2.2. 3D visual grounding." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.473, + 0.901 + ], + "angle": 0, + "content": "In recent years, visual grounding in 3D scenes has been explored [1, 11, 76] and also tackled in 3D question answering [13, 89]. Given the point cloud of 3D scenes, this task [1] requires models to locate the target object according to text instructions. Most works follow a two-stage scheme where multiple bounding boxes [9, 66, 102] or segmentation results [31, 90] are first predicted and then selecting the object according to language descriptions. [32] introduces multiview inputs, and [88] employs 2D semantics. [34] combines bottom-up [47] and top-down [7] detection methods. [48] designs a single-stage pipeline by progressively selecting key points. [20] extends [32] with the help of GPT [6] to generate multi-view text inputs. More recently, [30] proposes a neuro-symbolic framework with large language-tocode models [10]. Most works only localize a single object and [97] could localize a flexible number of objects. Differ" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.31, + 0.892, + 0.385 + ], + "angle": 0, + "content": "ent from previous works which directly handle point clouds or multi-view images, we convert the scene into textual descriptions and leverage large language models to infer the target object. Like [30], in this work, we also leverage large language models for object localization." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.406, + 0.8, + 0.423 + ], + "angle": 0, + "content": "3. Problem setup and preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.434, + 0.892, + 0.48 + ], + "angle": 0, + "content": "In this section, we discuss the definition of the task and preliminaries. We aim to generate human motion that is consistent with both the text description and the given scene." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.487, + 0.893, + 0.653 + ], + "angle": 0, + "content": "Text descriptions. The text description follows the template in Sr3D [1] (e.g., “ [ ]”). There are four actions (walk, sit, stand up, and lie). The target represents the object that the agents need to interact with, and the anchor objects help to determine the target. A certain type of target furniture category usually has many instances in one scene, while the anchor furniture should be unique. To specify the exact target object, there are five types of spatial relations [1] between the target and the anchor: horizontal proximity, vertical proximity, between, allocentric, and support." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.892, + 0.706 + ], + "angle": 0, + "content": "Scene representations. The scene is denoted as a point cloud of \\(N\\) points: \\(S \\in \\mathbf{R}^{N \\times 6}\\), containing the position and normal direction information of each point." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.712, + 0.892, + 0.849 + ], + "angle": 0, + "content": "Motion representations. The output motion sequence is represented as a sequence of SMPL-X [54] body meshes \\(M\\). SMPL-X is a parametric human body model. In this work, body parameters include body shape parameters \\(\\beta \\in \\mathbf{R}^{10}\\), global translation \\(r \\in \\mathbf{R}^3\\), 6D global orientation \\(\\gamma \\in \\mathbf{R}^6\\), and 6D pose parameters of \\(J\\) joints \\(\\theta \\in \\mathbf{R}^{J \\times 6}\\) [104]. Following [83], \\(\\beta\\) is treated as a condition to model the effect of body shape, and we omit it for ease of notation. Note that start position and pose are also generated by the model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Diffusion models. The diffusion [27] is defined as a Markov noising process \\(\\{\\mathbf{x}_t\\}_{t=0}^T\\) that follows \\(q(\\mathbf{x}_t \\mid \\mathbf{x}_0)\\), where \\(\\mathbf{x}_0 \\sim q(\\mathbf{x}_0)\\) is the data and \\(\\mathbf{x}_t\\) is the noised data at" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1857" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.141, + 0.09, + 0.832, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.479, + 0.896, + 0.551 + ], + "angle": 0, + "content": "Figure 3. Pipeline of localizing the target object. In stage 1, given the input text description and detected object bounding boxes (bbx), we construct the first prompt asking ChatGPT the categories of target objects and anchor objects. Based on the response, the scene graph can be simplified. In stage 2, we construct the second prompt with inputs and results from stage 1, including object relations derived from the simplified scene graph. The second prompt is designed for asking ChatGPT to infer the target object. Finally, we can get the target object bounding box from the response of ChatGPT." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.576, + 0.341, + 0.591 + ], + "angle": 0, + "content": "noising step \\(t\\). The formal definition is:" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.604, + 0.47, + 0.623 + ], + "angle": 0, + "content": "\\[\nq \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {0}\\right) = \\mathcal {N} \\left(\\mathbf {x} _ {t}; \\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {x} _ {0}, (1 - \\bar {\\alpha} _ {t}) \\mathbf {I}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.635, + 0.47, + 0.696 + ], + "angle": 0, + "content": "where \\(\\bar{\\alpha}_t\\) are constants with monotonically decreasing schedule. When \\(\\bar{\\alpha}_t\\) is small enough, we can approximate \\(\\mathbf{x}_T\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\). In our context, we use conditional diffusion models like [64, 75]. The training loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.125, + 0.709, + 0.47, + 0.727 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathrm {E} _ {t \\in [ 1, T ], \\mathbf {x} _ {0} \\sim q (\\mathbf {x} _ {0})} \\left[ \\| \\mathbf {x} _ {0} - G (\\mathbf {x} _ {t}, t, \\mathbf {C}) \\| \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.738, + 0.454, + 0.754 + ], + "angle": 0, + "content": "where \\(G\\) is the generative model, and \\(\\mathbf{C}\\) is the condition." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.769, + 0.169, + 0.784 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.9 + ], + "angle": 0, + "content": "The overview of our method is shown in Fig. 2. In Sec. 4.1, we leverage ChatGPT to localize the target object given a 3D scene and a textual description. Based on the accurate localization, we can focus on the target object and employ an object-centric generation pipeline to separately synthesize trajectories (Sec. 4.2) and motions (Sec. 4.3). Implementation details are discussed in Sec. 4.4." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.575, + 0.874, + 0.593 + ], + "angle": 0, + "content": "4.1. Language grounding of objects in 3D scenes" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.604, + 0.894, + 0.801 + ], + "angle": 0, + "content": "Scene-aware motion generation from textual descriptions requires the scene-understanding ability and build the relationship between scenes and texts. Since the instructions describe how the character moves and interacts with a single target object, the majority of the scene might bear negligible relevance to the final motions [51]. Motivated by this, we propose to locate the target object to identify the most pertinent information. Inspired by the recent progress of LLM [42, 52, 53], ChatGPT [52] is utilized to find the specific objects in the given text. We first obtain textual descriptions of scenes by building scene graphs. Then we feed them with text instructions to ChatGPT with specially designed prompts and parse the response to get target objects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Spatial scene graph extraction. To utilize LLM, the initial step involves converting a 3D scene into text. This is achieved by building a spatial scene graph. We utilize a pre-trained 3D object detection model in [47] to provide 3D box proposals. Subsequently, we follow the approach of [1] to obtain the relationships between objects. Specifically, for" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1858" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.129, + 0.094, + 0.255, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.196, + 0.262, + 0.21 + ], + "angle": 0, + "content": "(a) Target object bbx" + }, + { + "type": "image", + "bbox": [ + 0.303, + 0.095, + 0.428, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.304, + 0.197, + 0.426, + 0.21 + ], + "angle": 0, + "content": "(b) Target sensor" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.212, + 0.252, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.328, + 0.273, + 0.34 + ], + "angle": 0, + "content": "(c) Environment sensor" + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.221, + 0.428, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.316, + 0.327, + 0.421, + 0.34 + ], + "angle": 0, + "content": "(d) Traj sensor" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.352, + 0.47, + 0.421 + ], + "angle": 0, + "content": "Figure 4. The visualization of the environment sensor, target sensor, and trajectory sensor. The target sensor (b) gives detailed geometry of the target object. The environment sensor (c) gives coarse spatial information around the target object. The trajectory sensor (d) is located around the human." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.453, + 0.47, + 0.633 + ], + "angle": 0, + "content": "every set of two objects, we can infer their relationships in three types from their bounding boxes: horizontal proximity, vertical proximity, and support as mentioned in Sec. 3; for every set of three objects, we can infer if one of them is in between of the other two objects. As detection results do not contain object poses, we do not construct allocentric relations (e.g., \"the shelf that is behind the sofa\"). Then, we build a scene graph, where each object is assigned as a node in the graph, and edges between nodes represent the relationships between objects. By converting 3D scenes into text, we can apply ChatGPT to extract meaningful insights from the data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.642, + 0.468, + 0.853 + ], + "angle": 0, + "content": "Leveraging ChatGPT to localize the target object. A simple approach is to directly input the entire scene graph into ChatGPT and ask it to select the target objects. However, scenes may contain many objects, resulting in extremely long textural descriptions. We observe that ChatGPT is often confused in such settings and fails to respond with the right answer. To narrow the search space, we first employ ChatGPT to recognize objects that correlate to the provided text. We then exclude the unrelated objects from the scene graph, focusing solely on those with pertinent information, enabling us to pinpoint the target object effectively. This approach has the advantage of reducing the number of objects that need to be considered, making it easier for ChatGPT to identify the target object." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "As shown in Fig. 3, we construct two prompts in sequence. Take \"sit on the chair that is in the middle of the board and the end table\" as an example, we first need to nar" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.364 + ], + "angle": 0, + "content": "row down the object category search space. In order to find which type of objects we care about, we construct the first prompt that asks ChatGPT to find out target objects and anchor objects. Target object is defined as the final object that we want the agent to interact with, which in this case, is the \"table\". Anchor objects are the objects that help with determining the target object, for there might be many chairs in one scene. Based on the target object \"chair\" and anchor objects \"board\" and \"end table\", we can filter out all the unrelated objects in the scene, only keeping chairs, the board, and the end table in our scene graph. Next, according to the simplified scene graph, we can describe object relations in text: every edge in the scene graph could be converted to an edge sentence like \"chair 4 is far from the end table 0\". Converting all edges to such edge sentences gives a full description of the current scene. Finally, we construct a second prompt by asking ChatGPT to infer the target object from the accumulated edge sentences." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.371, + 0.825, + 0.387 + ], + "angle": 0, + "content": "4.2. Diffusion-based trajectory generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.394, + 0.892, + 0.515 + ], + "angle": 0, + "content": "Given the localized object from ChatGPT, we first generate the trajectory based on the instructions and then synthesize local human poses. Trajectory is defined as a sequence of characters' translations and orientations. As suggested by [51], not every point of the scene is relevant to the final human motions. Inspired by NSM [70] and ManipNet [92], we employ volumetric sensors (as shown in Fig. 4) around the target object to represent the scene." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.52, + 0.892, + 0.61 + ], + "angle": 0, + "content": "Object-centric scene representation. Denote the target object center location as \\( c_{o} = (c_{x},c_{y},c_{z}) \\). We transform the point cloud of the scene \\( S \\) to an object-centered coordinate axis centered at \\( c_{o} \\). To recognize the surrounding geometry of the target object, we create volumetric sensors called Environment Sensor and Target Sensor." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.616, + 0.892, + 0.722 + ], + "angle": 0, + "content": "Environment sensor. The environment sensor is centered at \\( c_{o} \\) in a cubic shape with a volume of \\( 4 \\times 4 \\times 4 \\, m^{3} \\), containing \\( 8 \\times 8 \\times 8 \\) cubic voxels as shown in Fig. 4 (c). It is constructed by collecting all occupancies \\( o_{s} \\), center positions \\( c_{v} \\), and normal directions \\( n_{v} \\) of each voxel to form a feature vector \\( E \\). Like [70], the scene occupancy \\( o_{s} \\in R^{1} \\) in each voxel is defined based on the given scene mesh:" + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.729, + 0.891, + 0.786 + ], + "angle": 0, + "content": "\\[\no _ {s} = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} d _ {s} < 0, \\\\ 0 & \\text {i f} d _ {s} > a _ {s}, \\\\ 1 - \\frac {d _ {s}}{a _ {s}} & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.79, + 0.892, + 0.867 + ], + "angle": 0, + "content": "where \\( d_{s} \\) is the signed distance between the scene mesh and the voxel center, and \\( a_{s} \\) is the voxel edge length. \\( n_v \\) is the normal direction of the closest scene point to the voxel center. The environment sensor provides coarse scene geometry around the target object." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Target sensor. To capture the detailed geometry of the target object, we further build a target sensor. As we already" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1859" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.182 + ], + "angle": 0, + "content": "obtain the 3D bounding box of the target object in Sec. 4.1, we crop the point clouds according to the bounding box and construct an \\(8 \\times 8 \\times 8\\) cubic volumes that cover the bounding box, as shown in Fig. 4 (b). Target sensor \\(T\\) is in the same form as the environment sensor \\(E\\), except the target sensor has a different voxel size, as visualized in Fig. 4 (b) and (c)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.183, + 0.468, + 0.287 + ], + "angle": 0, + "content": "Given the constructed object-centric representations, we follow [57] to employ a transformer decoder architecture [79], which enables arbitrary length motions. As for the text input, we use CLIP [62] text encoder to encode input text to the text feature \\( L \\). The time-step \\( t \\) is injected into the decoder in sinusoidal position embeddings form [79]. In summary, the condition for this generation model is" + }, + { + "type": "equation", + "bbox": [ + 0.214, + 0.299, + 0.468, + 0.315 + ], + "angle": 0, + "content": "\\[\n\\mathbf {C} _ {\\mathbf {t}} = \\{L, E, T \\}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.326, + 0.468, + 0.416 + ], + "angle": 0, + "content": "where \\( L \\) is the text feature, \\( E \\) is the environment sensor, and \\( T \\) is the target sensor. All the conditions are projected to the same dimension \\( D = 512 \\) by MLPs and summed with positional embeddings to form tokens. We use the simple objective described in Eq. 2 to train the trajectory generation model \\( G_{r} \\) to generate the trajectory \\( \\mathbf{r}_{1:N} \\) with length \\( N \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.425, + 0.382, + 0.441 + ], + "angle": 0, + "content": "4.3. Diffusion-based motion completion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.449, + 0.468, + 0.507 + ], + "angle": 0, + "content": "Given the trajectory from Sec. 4.2, the next step is to complete the whole motion. Based on the generated trajectories, we construct a Trajectory Sensor \\( O \\) to explicitly reason about the interaction between the character and scenes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.514, + 0.468, + 0.649 + ], + "angle": 0, + "content": "Trajectory sensor. The trajectory sensor is also a volumetric sensor which is similar to the form of the environment sensor but is designed for ego-centric perception [95] as shown in Fig. 4 (d). Specifically, trajectory sensors are positioned around characters in each frame. This sensor \\( O_{i} \\) is centered at the predicted root position of the \\( i \\)-th frame and faces to the \\( i \\)-th frame's predicted root orientation, containing \\( 8 \\times 8 \\times 8 \\) cubic voxels that store scene occupancy. The occupancy calculation is the same as Eq. 3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.65, + 0.468, + 0.693 + ], + "angle": 0, + "content": "Another transformer-based conditional diffusion model is used to synthesize local poses along trajectories. The condition is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.706, + 0.468, + 0.722 + ], + "angle": 0, + "content": "\\[\n\\mathbf {C} _ {\\mathbf {m}} = \\{L, E, T, O _ {1}, \\dots , O _ {N} \\}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.732, + 0.468, + 0.793 + ], + "angle": 0, + "content": "where \\(L, E,\\) and \\(T\\) have the same meanings in Eq. 4. Simple objective described in Eq. 2 is used to train the motion generation model \\(G_{m}\\) to generate global orientation \\(\\gamma_{1:N}\\) and local poses \\(\\theta_{1:N}\\) with length \\(N\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.291, + 0.817 + ], + "angle": 0, + "content": "4.4. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Following [38, 57, 80, 81], we use separate diffusion models for the trajectory generation and the motion generation. Because the HUMANISE dataset [83] contains a relatively small number (51 minutes) of pure motion samples from the AMASS dataset (3772 minutes) [49], we first pre-train the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.213 + ], + "angle": 0, + "content": "models on the whole AMASS dataset for 200 epochs and then fine-tune on the HUMANISE dataset for 200 epochs to improve the motion quality. During pretraining, the text feature \\( L \\) is set to all zeros. Both models are trained with the AdamW optimizer [40], using a learning rate of 0.0001 on a single Nvidia RTX 3090 GPU. The batch size is set to 128. The version of ChatGPT is gpt-3.5-turbo. More details can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.228, + 0.632, + 0.244 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.253, + 0.681, + 0.267 + ], + "angle": 0, + "content": "5.1. Evaluation metrics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.277, + 0.89, + 0.307 + ], + "angle": 0, + "content": "We evaluate the generated motions in three aspects: scene-conditional, action-conditional, and pure motion quality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.312, + 0.892, + 0.463 + ], + "angle": 0, + "content": "Scene-conditional motion quality. To evaluate how the generated motion is aligned with the scene, we calculate the body-to-goal distance (goal dist.) [83] to measure how accurately the character interacts with target objects. However, goal dist. does not consider the consistency of the entire motion and scene (e.g., sitting on the sofa with incorrect orientation). To compensate for goal dist., a human perceptual study (indicated by scene score) is performed by randomly sampling 20 scenarios for each model, where ten workers are required to score each sample." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.469, + 0.892, + 0.558 + ], + "angle": 0, + "content": "Action-conditional motion quality. To measure how the generated motion is aligned with the text, we follow [17] to evaluate action recognition accuracy (accuracy), diversity, and multimodality of the results. Calculating these metrics relies on a pre-trained action recognition model [87] and we train the recognition model on the HUMANISE dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.565, + 0.892, + 0.655 + ], + "angle": 0, + "content": "Pure motion quality. We evaluate the realism of generated motions using the metric suggested by [17], namely Frechet Inception Distance (FID). A lower FID indicates that generated motions are closer to the groundtruth motions. We also perform a human perceptual study (indicated by quality score) to measure pure motion quality." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.666, + 0.891, + 0.681 + ], + "angle": 0, + "content": "5.2. Generating human motion from text and scene" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Our experiments are conducted on the HUMANISE dataset [83]. Following the previous setting [83], there are \\(16.5k\\) motion sequences in 543 scenes for training and \\(3.1k\\) motion sequences in 100 scenes for testing. We compare our method with four baselines: (1) \\(\\mathbf{MDM}^*\\) [75]: a diffusion-based motion generator. (2) \\(\\mathbf{GMD}^*\\) [38]: GMD proposes various techniques for enhancing the control and quality of MDM. We employ the single-stage setting of \\(\\mathbf{GMD}^*\\). Similar to HUMANISE, we use a point-transformer to provide scene features for \\(\\mathbf{MDM}^*\\) and \\(\\mathbf{GMD}^*\\). (3) \\(\\mathbf{GMD}^{\\mathbf{HC}}\\): we provide object centers predicted by HUMANISE (denoted by HC) to guide the motion generation process in GMD by adding a proximity loss to encourage the motion to be close to the predicted object center. The proximity loss is defined" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1860" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.143, + 0.089, + 0.832, + 0.221 + ], + "angle": 0, + "content": "
MethodsScene-conditionalAction-conditionalPure motion quality
goal dist.↓scene score↑accuracy↑diversity→multimodality→FID↓quality score↑
Real0.014-99.1%4.822.280.00-
MDM*2.0481.9996.0%4.832.570.162.45
GMD*1.2292.5096.6%4.912.280.242.64
GMDHC1.1302.4996.4%4.932.430.232.76
HUMANISE0.9952.3389.7%4.252.661.112.81
Ours0.3843.5497.1%4.782.310.123.50
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.23, + 0.895, + 0.288 + ], + "angle": 0, + "content": "Table 1. Quantitative results on the HUMANISE dataset. We compare our method with four baselines (please refer to Sec. 5.2) and the real data. \\(\\mathrm{GMD}^{\\mathrm{HC}}\\) means using HUMANISE's predicted center to guide motion generation in \\(\\mathrm{GMD^{*}}\\). Among the metrics, scene score and quality score are perceptual studies. \\(\\uparrow\\) means higher is better and \\(\\downarrow\\) means lower is better. \\(\\rightarrow\\) means closer to the real data is better. Bold indicates the best results.underline indicates the second best." + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.3, + 0.891, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.502, + 0.893, + 0.531 + ], + "angle": 0, + "content": "Figure 5. Qualitative results. We compare our method with groundtruth and four baselines (please refer to Sec. 5.2) given the same text descriptions. Our method synthesizes motions that interact with the object precisely as the groundtruth data while the baselines fail." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.556, + 0.471, + 0.601 + ], + "angle": 0, + "content": "as the distance from HC to the predicted human pelvis at the interacting frame. (4) HUMANISE: we directly use their released models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.602, + 0.47, + 0.724 + ], + "angle": 0, + "content": "The quantitative results are shown in Tab. 1. Our method outperforms the baseline in terms of goal dist., scene score, accuracy, FID, and quality score and achieves competitive results in diversity and multimodality. The quantitative results show that our approach can generate better-quality motions and are more consistent with all the conditions. The qualitative results are demonstrated in Fig. 5. Please refer to the supplementary material for more visualizations." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.736, + 0.228, + 0.751 + ], + "angle": 0, + "content": "5.3. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Ablation of main components. Four variants are constructed to explore the effect of our design choice. (1) \"w/o localization\": the localization module is removed and trajectories and motions are generated directly using scene point clouds. Scene features are given as the same form of MDM* and GMD*. (2) \"w/o object-centric\": we remove our object-centric representation and predict motions in the scene coordinate. The scene features of our sensors are still employed. (3) \"w/o two-stage\": trajectories and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.556, + 0.893, + 0.708 + ], + "angle": 0, + "content": "motions are generated together. (4) \"w/o diffusion\": we employ cVAE instead of the diffusion model as the architecture. (5) \"w/o pretrain\": the models are not pre-trained on the AMASS dataset. As shown in Tab. 2, the localization enhances the precise interaction ability and the object-centric representation improves motion quality. Although the (\"w/o two-stage\") is more efficient and has competitive performance metrics comparing to our method, it is prone to collapse directly into the target position, lacking gradual transition and realism." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.902 + ], + "angle": 0, + "content": "The design choice of object localization. For object localization, we compare our method with two baselines and five variants. (1) HUMANISE: predicts the object center in the auxiliary task. (2) BUTD-DETR [34]: is a 3D visual grounding method. (3) “ours w/o two-stage”: we employ a one-stage question-answering process. (4) “ours w/o few-shot”: the few-shot examples are not provided. (5) “ours using text matching”: we compute CLIP similarities instead of using ChatGPT. (6) “ours using LLaMA”: ChatGPT is replaced by LLaMA 2 7B model. (7) “ours using Mistral”: ChatGPT is replaced by Mistral 7B [35], an open-source LLM. The prompts for LLaMA 2 and Mistral" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1861" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.468, + 0.212 + ], + "angle": 0, + "content": "
VariantsSceneActionPure
goal dist.↓accuracy↑diversity→mm→FID↓
Real0.01499.1%4.822.280.00
w/o localization0.59274.5%3.712.753.14
w/o object-centric0.41390.6%4.132.631.01
w/o two-stage0.38597.2%4.932.330.14
w/o diffusion0.39040.0%3.693.863.61
w/o pretrain0.39282.6%3.882.372.50
Ours0.38497.1%4.782.310.12
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.221, + 0.472, + 0.276 + ], + "angle": 0, + "content": "Table 2. Ablation of main components. We compare our method with five variants (please refer to Sec. 5.3). Among them, \\( mm \\) indicates multimodality. Bold indicates the best results. Underline indicates the second best." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.305, + 0.47, + 0.486 + ], + "angle": 0, + "content": "7B are slightly adjusted. We evaluate these variants under two scenarios: predicted detection [47] and groundtruth detection. The metrics include the accuracy (acc.) and the distance from predicted centers to the object center (center dist.). \"acc.\" is the percentage of times with an IoU (Intersection over Union) higher than the threshold (0.25) following [34]. The results are shown in Tab. 3. Since the results of Mistral are only slightly behind of ChatGPT, ChatGPT can be replaced by Mistral for better reproducibility, but cannot be replaced by the text matching method. Moreover, if groundtruth detection is provided, our method can achieve even better results." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.492, + 0.47, + 0.523 + ], + "angle": 0, + "content": "The design choice of sensor density. Please refer to the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.533, + 0.228, + 0.548 + ], + "angle": 0, + "content": "5.4. Generalization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.557, + 0.47, + 0.664 + ], + "angle": 0, + "content": "To validate the generalization ability of our method, we run our method directly on the unseen PROX dataset [22] without fine-tuning. We provide illustrations in Fig. 6. With our localization method based on ChatGPT and motion generation method based on volumetric sensors, our pipeline can easily generalize to other datasets. Please refer to the supplementary material for more results." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.679, + 0.192, + 0.695 + ], + "angle": 0, + "content": "6. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We have demonstrated that our approach could synthesize motions with better quality and more precise interactions. However, restricted by the dataset, the duration of motions is relatively short (60% are around 1-3s), and most texts follow the template [1] without detailed descriptions. We only tackle static scenes and extending our method to interact with moving objects is a future direction. We acknowledge that leveraging LLMs has several problems. ChatGPT may fail when the detector misses the objects and the behavior of ChatGPT is regulated by prompts [37]. Despite we have shown in Tab. 3 that ChatGPT can be replaced by an open source LLM Mistral 7B, using LLMs instead of classical parsing approaches is less efficient in the inference stage." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.089, + 0.891, + 0.232 + ], + "angle": 0, + "content": "
VariantsPredicted detectionGT detection
acc. ↑center dist.↓acc. ↑center dist.↓
HUMANISE-1.48--
BUTD-DETR62.51.3363.11.23
Ours w/o two-stage49.01.3458.51.10
Ours w/o few-shot69.30.8086.30.45
Ours using text matching67.00.8979.70.52
Ours using LLaMA 240.31.5542.51.52
Ours using Mistral 7B72.40.7487.20.33
Ours using ChatGPT-3.575.60.6190.50.37
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.242, + 0.894, + 0.325 + ], + "angle": 0, + "content": "Table 3. Design choices of object localization. We compare our method with two baselines and five variants (please refer to Sec. 5.3). Since HUMANISE directly regresses the coordinate of centers without utilizing groundtruth detection, we do not calculate acc. and only include their results under predicted detection. Bold indicates the best results." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.35, + 0.695, + 0.458 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.35, + 0.871, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.459, + 0.695, + 0.569 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.459, + 0.871, + 0.568 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.579, + 0.892, + 0.636 + ], + "angle": 0, + "content": "Figure 6. Qualitative results of our method on the PROX dataset. We run our method on the scenes from the PROX dataset without fine-tuning. Results show that our method is capable to generalize to unseen scenes and objects." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.67, + 0.619, + 0.685 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.7, + 0.892, + 0.821 + ], + "angle": 0, + "content": "In this work, we introduce a novel method for generating motion from text in a scene. To tackle this problem, we propose a two-step approach. The first step involves 3D visual grounding, where we identify the target object. In the second step, concentrating on the target object, we build a diffusion-based motion generation method. Our approach offers several advantages, including improved 3D visual grounding accuracy and motion quality." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.832, + 0.661, + 0.849 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.902 + ], + "angle": 0, + "content": "This work was partially supported by the NSFC (No. 62172364), Ant Group and Information Technology Center and State Key Lab of CAD&CG, Zhejiang University." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1862" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.116, + 0.468, + 0.169 + ], + "angle": 0, + "content": "[1] Panos Achlioptas, Ahmed Abdelreheem, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In ECCV, 2020. 2, 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.173, + 0.468, + 0.213 + ], + "angle": 0, + "content": "[2] Chaitanya Ahuja and Louis-Philippe Morency. Language2pose: Natural language grounded pose forecasting. In 3DV, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.217, + 0.468, + 0.257 + ], + "angle": 0, + "content": "[3] Nikos Athanasiou, Mathis Petrovich, Michael J. Black, and Gül Varol. TEACH: Temporal Action Compositions for 3D Humans. In 3DV, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.26, + 0.468, + 0.3 + ], + "angle": 0, + "content": "[4] Nikos Athanasiou, Mathis Petrovich, Michael J. Black, and Gül Varol. SINC: Spatial composition of 3D human motions for simultaneous action generation. ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.304, + 0.468, + 0.344 + ], + "angle": 0, + "content": "[5] German Barquero, Sergio Escalera, and Cristina Palmero. Belfusion: Latent diffusion for behavior-driven human motion prediction. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.347, + 0.468, + 0.483 + ], + "angle": 0, + "content": "[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In NeurIPS, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.487, + 0.468, + 0.54 + ], + "angle": 0, + "content": "[7] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.544, + 0.468, + 0.584 + ], + "angle": 0, + "content": "[8] Yu-Wei Chao, Jimei Yang, Weifeng Chen, and Jia Deng. Learning to sit: Synthesizing human-chair interactions via hierarchical control. In AAAI, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.587, + 0.468, + 0.628 + ], + "angle": 0, + "content": "[9] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.631, + 0.468, + 0.699 + ], + "angle": 0, + "content": "[10] Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, et al. Evaluating large language models trained on code. arXiv preprint, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.702, + 0.468, + 0.756 + ], + "angle": 0, + "content": "[11] Yixin Chen, Qing Li, Deqian Kong, Yik Lun Kei, Song-Chun Zhu, Tao Gao, Yixin Zhu, and Siyuan Huang. Yourfit: Embodied reference understanding with language and gesture. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.759, + 0.468, + 0.813 + ], + "angle": 0, + "content": "[12] Rishabh Dabral, Muhammad Hamza Mughal, Vladislav Golyanik, and Christian Theobalt. Mofusion: A framework for denoising-diffusion-based motion synthesis. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.817, + 0.468, + 0.857 + ], + "angle": 0, + "content": "[13] Abhishek Das, Georgia Gkioxari, Stefan Lee, Devi Parikh, and Dhruv Batra. Neural Modular Control for Embodied Question Answering. In CoRL, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.86, + 0.468, + 0.899 + ], + "angle": 0, + "content": "[14] Katerina Fragkiadaki, Sergey Levine, Panna Felsen, and Jitendra Malik. Recurrent network models for human dynamics. In ICCV, 2015. 2" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.116, + 0.468, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[15] Anindita Ghosh, Noshaba Cheema, Cennet Oguz, Christian Theobalt, and Philipp Slusallek. Synthesis of compositional animations from textual descriptions. In ICCV, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.136, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[16] Anindita Ghosh, Rishabh Dabral, Vladislav Golyanik, Christian Theobalt, and Philipp Slusallek. Imos: Intent-driven full-body motion synthesis for human-object interactions. In Computer Graphics Forum, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.191, + 0.892, + 0.244 + ], + "angle": 0, + "content": "[17] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3d human motions. In ACM MM, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.247, + 0.892, + 0.286 + ], + "angle": 0, + "content": "[18] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In CVPR, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.289, + 0.892, + 0.341 + ], + "angle": 0, + "content": "[19] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In ECCV, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.344, + 0.892, + 0.397 + ], + "angle": 0, + "content": "[20] Zoey Guo, Yiwen Tang, Ray Zhang, Dong Wang, Zhigang Wang, Bin Zhao, and Xuelong Li. Viewrefer: Grasp the multi-view knowledge for 3d visual grounding. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.4, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[21] Félix G Harvey, Mike Yurick, Derek Nowrouzezahrai, and Christopher Pal. Robust motion in-betweening. ACM Trans. Graph., 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.442, + 0.892, + 0.481 + ], + "angle": 0, + "content": "[22] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J. Black. Resolving 3d human pose ambiguities with 3d scene constraints. In ICCV, 2019. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.484, + 0.892, + 0.524 + ], + "angle": 0, + "content": "[23] Mohamed Hassan, Duygu Ceylan, Ruben Villegas, Jun Saito, Jimei Yang, Yi Zhou, and Michael J. Black. Stochastic scene-aware motion prediction. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.526, + 0.892, + 0.565 + ], + "angle": 0, + "content": "[24] Mohamed Hassan, Yunrong Guo, Tingwu Wang, Michael Black, Sanja Fidler, and Xue Bin Peng. Synthesizing physical character-scene interactions. In SIGGRAPH, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.568, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[25] Chengan He, Jun Saito, James Zachary, Holly Rushmeier, and Yi Zhou. Nemf: Neural motion fields for kinematic animation. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.609, + 0.892, + 0.661 + ], + "angle": 0, + "content": "[26] Gustav Eje Henter, Simon Alexanderson, and Jonas Beskow. Moglow: Probabilistic and controllable motion synthesis using normalising flows. ACM Trans. Graph., 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.665, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[27] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.693, + 0.892, + 0.733 + ], + "angle": 0, + "content": "[28] Daniel Holden, Jun Saito, and Taku Komura. A deep learning framework for character motion synthesis and editing. ACM Trans. Graph., 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.735, + 0.892, + 0.774 + ], + "angle": 0, + "content": "[29] Daniel Holden, Taku Komura, and Jun Saito. Phase-functional neural networks for character control. ACM Trans. Graph., 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.777, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[30] Joy Hsu, Jiayuan Mao, and Jiajun Wu. Ns3d: Neurosymbolic grounding of 3d objects and relations. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.818, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[31] Pin-Hao Huang, Han-Hung Lee, Hwann-Tzong Chen, and Tyng-Luh Liu. Text-guided graph neural networks for referring 3d instance segmentation. AAAI, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.86, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[32] Shijia Huang, Yilun Chen, Jiaya Jia, and Liwei Wang. Multi-view transformer for 3d visual grounding. In CVPR, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1863" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.093, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[33] Siyuan Huang, Zan Wang, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, and Song-Chun Zhu. Diffusion-based generation, optimization, and planning in 3d scenes. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.149, + 0.47, + 0.203 + ], + "angle": 0, + "content": "[34] Ayush Jain, Nikolaos Gkanatsios, Ishita Mediratta, and Katerina Fragkiadaki. Bottom up top down detection transformers for language grounding in images and point clouds. In ECCV, 2022. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.205, + 0.471, + 0.272 + ], + "angle": 0, + "content": "[35] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.274, + 0.471, + 0.316 + ], + "angle": 0, + "content": "[36] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. arXiv preprint arXiv:2306.14795, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.317, + 0.471, + 0.37 + ], + "angle": 0, + "content": "[37] Sai Shashank Kalakonda, Shubh Maheshwari, and Ravi Kiran Sarvadevabhatla. Action-gpt: Leveraging large-scale language models for improved and generalized action generation. In ICME, 2023. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.372, + 0.471, + 0.425 + ], + "angle": 0, + "content": "[38] Korrawe Karunratanakul, Konpat Preechakul, Supasorn Suwajanakorn, and Siyu Tang. Guided motion diffusion for controllable human motion synthesis. In ICCV, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.427, + 0.471, + 0.468 + ], + "angle": 0, + "content": "[39] Jihoon Kim, Jiseob Kim, and Sungjoon Choi. Flame: Freeform language-based motion synthesis & editing. In AAAI, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.469, + 0.471, + 0.497 + ], + "angle": 0, + "content": "[40] Diederik P. Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization. arXiv e-prints, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.498, + 0.471, + 0.525 + ], + "angle": 0, + "content": "[41] Diederik P Kingma and Max Welling. Auto-Encoding Variational Bayes. arXiv e-prints, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.526, + 0.471, + 0.566 + ], + "angle": 0, + "content": "[42] Takeshi Kojima, Shixiang (Shane) Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners. In NeurIPS, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.567, + 0.471, + 0.608 + ], + "angle": 0, + "content": "[43] Hanyang Kong, Kehong Gong, Dongze Lian, Michael Bi Mi, and Xinchao Wang. Priority-centric human motion generation in discrete latent space. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.609, + 0.471, + 0.663 + ], + "angle": 0, + "content": "[44] Peizhuo Li, Kfir Aberman, Zihan Zhang, Rana Hanocka, and Olga Sorkine-Hornung. Ganimator: Neural motion synthesis from a single sequence. ACM Trans. Graph., 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.665, + 0.471, + 0.718 + ], + "angle": 0, + "content": "[45] Quanzhou Li, Jingbo Wang, Chen Change Loy, and Bo Dai. Task-oriented human-object interactions generation with implicit neural representations. arXiv preprint, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.72, + 0.471, + 0.761 + ], + "angle": 0, + "content": "[46] Hung Yu Ling, Fabio Zinno, George Cheng, and Michiel Van De Panne. Character controllers using motion vaes. ACM Trans. Graph., 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.763, + 0.471, + 0.803 + ], + "angle": 0, + "content": "[47] Ze Liu, Zheng Zhang, Yue Cao, Han Hu, and Xin Tong. Group-free 3d object detection via transformers. In ICCV, 2021. 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.805, + 0.471, + 0.858 + ], + "angle": 0, + "content": "[48] Junyu Luo, Jiahui Fu, Xianghao Kong, Chen Gao, Haibing Ren, Hao Shen, Huaxia Xia, and Si Liu. 3d-sps: Single-stage 3d visual grounding via referred point progressive selection. In CVPR, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.86, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[49] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In ICCV, 2019. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.093, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[50] Julieta Martinez, Michael J. Black, and Javier Romero. On human motion prediction using recurrent neural networks. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.136, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[51] Aymen Mir, Xavier Puig, Angjoo Kanazawa, and Gerard Pons-Moll. Generating continual human motion in diverse 3d scenes. arXiv preprint, 2023. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.179, + 0.892, + 0.206 + ], + "angle": 0, + "content": "[52] OpenAI. Openai: Introducing chatgpt. https://openai.com/blog/chatgpt, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.208, + 0.8, + 0.221 + ], + "angle": 0, + "content": "[53] OpenAI. Gpt-4 technical report, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.223, + 0.892, + 0.277 + ], + "angle": 0, + "content": "[54] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3d hands, face, and body from a single image. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.279, + 0.892, + 0.32 + ], + "angle": 0, + "content": "[55] Mathis Petrovich, Michael J. Black, and Gül Varol. Action-conditioned 3D human motion synthesis with transformer VAE. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.322, + 0.892, + 0.363 + ], + "angle": 0, + "content": "[56] Mathis Petrovich, Michael J. Black, and Gül Varol. TEMOS: Generating diverse human motions from textual descriptions. In ECCV, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.365, + 0.892, + 0.419 + ], + "angle": 0, + "content": "[57] Huaijin Pi, Sida Peng, Minghui Yang, Xiaowei Zhou, and Hujun Bao. Hierarchical generation of human-object interactions with diffusion probabilistic models. In ICCV, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.421, + 0.892, + 0.449 + ], + "angle": 0, + "content": "[58] Matthias Plappert, Christian Mandery, and Tamim Asfour. The kit motion-language dataset. *Big data*, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.451, + 0.892, + 0.504 + ], + "angle": 0, + "content": "[59] Abhinanda R. Punnakkal, Arjun Chandrasekaran, Nikos Athanasiou, Alejandra Quiros-Ramirez, and Michael J. Black. Babel: Bodies, action and behavior with english labels. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.506, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[60] Yijun Qian, Jack Urbanek, Alexander G Hauptmann, and Jungdam Won. Breaking the limits of text-conditioned 3d motion synthesis with elaborative descriptions. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.563, + 0.892, + 0.617 + ], + "angle": 0, + "content": "[61] Sigal Raab, Inbal Leibovitch, Peizhuo Li, Kfir Aberman, Olga Sorkine-Hornung, and Daniel Cohen-Or. Modi: Unconditional motion synthesis from diverse data. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.62, + 0.892, + 0.688 + ], + "angle": 0, + "content": "[62] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.69, + 0.892, + 0.759 + ], + "angle": 0, + "content": "[63] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485-5551, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.761, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[64] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical Text-Conditional Image Generation with CLIP Latents. arXiv e-prints, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.817, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[65] Ali Razavi, Aaron van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. In NeurIPS, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[66] Junha Roh, Karthik Desingh, Ali Farhadi, and Dieter Fox. Languagerefer: Spatial-language model for 3d visual grounding. In CoRL, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1864" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.093, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[67] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.136, + 0.47, + 0.177 + ], + "angle": 0, + "content": "[68] Yonatan Shafir, Guy Tevet, Roy Kapon, and Amit H Bermano. Human motion diffusion as a generative prior. arXiv preprint, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.179, + 0.47, + 0.22 + ], + "angle": 0, + "content": "[69] Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. In NeurIPS, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.222, + 0.47, + 0.262 + ], + "angle": 0, + "content": "[70] Sebastian Starke, He Zhang, Taku Komura, and Jun Saito. Neural state machine for character-scene interactions. ACM Trans. Graph., 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.264, + 0.47, + 0.305 + ], + "angle": 0, + "content": "[71] Sebastian Starke, Yiwei Zhao, Taku Komura, and Kazi Zaman. Local motion phases for learning multi-contact character movements. ACM Trans. Graph., 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.307, + 0.47, + 0.348 + ], + "angle": 0, + "content": "[72] Sebastian Starke, Ian Mason, and Taku Komura. Deepphase: Periodic autoencoders for learning motion phase manifolds. ACM Trans. Graph., 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.35, + 0.47, + 0.39 + ], + "angle": 0, + "content": "[73] Omid Taheri, Vasileios Choutas, Michael J. Black, and Dimitrios Tzionas. Goal: Generating 4d whole-body motion for hand-object grasping. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.392, + 0.47, + 0.433 + ], + "angle": 0, + "content": "[74] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. In ECCV, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.435, + 0.47, + 0.475 + ], + "angle": 0, + "content": "[75] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In ICLR, 2023. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.478, + 0.47, + 0.518 + ], + "angle": 0, + "content": "[76] Jesse Thomason, Mohit Shridhar, Yonatan Bisk, Chris Paxton, and Luke Zettlemoyer. Language grounding with 3d objects. In CoRL, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.521, + 0.47, + 0.603 + ], + "angle": 0, + "content": "[77] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models. arXiv preprint, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.605, + 0.47, + 0.645 + ], + "angle": 0, + "content": "[78] Aaron van den Oord, Oriol Vinyals, and koray kavukcuoglu. Neural discrete representation learning. In NeurIPS, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.647, + 0.47, + 0.7 + ], + "angle": 0, + "content": "[79] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.704, + 0.47, + 0.744 + ], + "angle": 0, + "content": "[80] Jiashun Wang, Huazhe Xu, Jingwei Xu, Sifei Liu, and Xiaolong Wang. Synthesizing long-term 3d human motion and interaction in 3d scenes. In CVPR, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.747, + 0.47, + 0.787 + ], + "angle": 0, + "content": "[81] Jingbo Wang, Yu Rong, Jingyuan Liu, Sijie Yan, Dahua Lin, and Bo Dai. Towards diverse and natural scene-aware 3d human motion synthesis. In CVPR, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.789, + 0.47, + 0.844 + ], + "angle": 0, + "content": "[82] Xi Wang, Gen Li, Yen-Ling Kuo, Muhammed Kocabas, Emre Aksan, and Otmar Hilliges. Reconstructing action-conditioned human-object interactions using commonsense knowledge priors. In 3DV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.846, + 0.47, + 0.899 + ], + "angle": 0, + "content": "[83] Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, Wei Liang, and Siyuan Huang. Humanise: Language-conditioned human motion generation in 3d scenes. In NeurIPS, 2022. 2, 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.093, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[84] Yan Wu, Jiahao Wang, Yan Zhang, Siwei Zhang, Otmar Hilliges, Fisher Yu, and Siyu Tang. Saga: Stochastic whole-body grasping with contact. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.136, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[85] Yiming Xie, Varun Jampani, Lei Zhong, Deqing Sun, and Huaizu Jiang. Omnicontrol: Control any joint at any time for human motion generation. arXiv preprint, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.178, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[86] Sirui Xu, Zhengyuan Li, Yu-Xiong Wang, and Liang-Yan Gui. Interdiff: Generating 3d human-object interactions with physics-informed diffusion. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.22, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[87] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In AAAI, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.261, + 0.892, + 0.302 + ], + "angle": 0, + "content": "[88] Zhengyuan Yang, Songyang Zhang, Liwei Wang, and Jiebo Luo. Sat: 2d semantics assisted training for 3d visual grounding. In ICCV, pages 1856-1866, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.304, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[89] Shuquan Ye, Dongdong Chen, Songfang Han, and Jing Liao. 3d question answering. IEEE Transactions on Visualization and Computer Graphics, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.345, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[90] Zhihao Yuan, Xu Yan, Yinghong Liao, Ruimao Zhang, Sheng Wang, Zhen Li, and Shuguang Cui. Instancerefer: Cooperative holistic understanding for visual grounding on point clouds through instance multi-level contextual referring. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.415, + 0.892, + 0.454 + ], + "angle": 0, + "content": "[91] He Zhang, Sebastian Starke, Taku Komura, and Jun Saito. Mode-adaptive neural networks for quadruped motion control. ACM Trans. Graph., 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.456, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[92] He Zhang, Yuting Ye, Takaaki Shiratori, and Taku Komura. Manipnet: Neural manipulation synthesis with a hand-object spatial representation. ACM Trans. Graph., 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.511, + 0.892, + 0.566 + ], + "angle": 0, + "content": "[93] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Yong Zhang, Hongwei Zhao, Hongtao Lu, Xi Shen, and Ying Shan. Generating human motion from textual descriptions with discrete representations. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.568, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[94] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.623, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[95] Xiaohan Zhang, Bharat Lal Bhatnagar, Sebastian Starke, Vladimir Guzov, and Gerard Pons-Moll. Couch: Towards controllable human-chair interactions. In ECCV, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.665, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[96] Yan Zhang and Siyu Tang. The wanderings of odysseus in 3d scenes. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.693, + 0.892, + 0.733 + ], + "angle": 0, + "content": "[97] Yiming Zhang, ZeMing Gong, and Angel X Chang. Multi3drefer: Grounding text description to multiple 3d objects. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.735, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[98] Yaqi Zhang, Di Huang, Bin Liu, Shixiang Tang, Yan Lu, Lu Chen, Lei Bai, Qi Chu, Nenghai Yu, and Wanli Ouyang. Motiongpt: Finetuned llms are general-purpose motion generators. arXiv preprint, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.79, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[99] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[100] Kaifeng Zhao, Shaofei Wang, Yan Zhang, Thabo Beeler, and Siyu Tang. Compositional human-scene interaction synthesis with semantic control. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[101] Kaifeng Zhao, Yan Zhang, Shaofei Wang, Thabo Beeler, and Siyu Tang. Synthesizing diverse human motions in 3d indoor scenes. In ICCV, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1865" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[102] Lichen Zhao, Daigang Cai, Lu Sheng, and Dong Xu. 3dvg-transformer: Relation modeling for visual grounding on point clouds. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[103] Yang Zheng, Yanchao Yang, Kaichun Mo, Jiaman Li, Tao Yu, Yebin Liu, C Karen Liu, and Leonidas J Guibas. Gimo: Gaze-informed human motion prediction in context. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.234 + ], + "angle": 0, + "content": "[104] Yi Zhou, Connelly Barnes, Lu Jingwan, Yang Jimei, and Li Hao. On the continuity of rotation representations in neural networks. In CVPR, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.485, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "1866" + } + ] +] \ No newline at end of file diff --git a/2024/Generating Human Motion in 3D Scenes from Text Descriptions/507f8570-ad9c-455e-838f-8fa1bda75b31_origin.pdf b/2024/Generating Human Motion in 3D Scenes from Text Descriptions/507f8570-ad9c-455e-838f-8fa1bda75b31_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..07bf4fdfc5f44f0f19d306d6e990fba9703996fa --- /dev/null +++ b/2024/Generating Human Motion in 3D Scenes from Text Descriptions/507f8570-ad9c-455e-838f-8fa1bda75b31_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ac76ca6fd7f3471116ef33f8ad3ab682da200c36036c70022e6ca13cce26b18 +size 7995829 diff --git a/2024/Generating Human Motion in 3D Scenes from Text Descriptions/full.md b/2024/Generating Human Motion in 3D Scenes from Text Descriptions/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c32008e4fa007cc61ea75463bc48f332dd1af4f3 --- /dev/null +++ b/2024/Generating Human Motion in 3D Scenes from Text Descriptions/full.md @@ -0,0 +1,352 @@ +# Generating Human Motion in 3D Scenes from Text Descriptions + +Zhi Cen1 Huaijin Pi1 Sida Peng1† Zehong Shen1 Minghui Yang2 Shuai Zhu2 Hujun Bao1 Xiaowei Zhou1 1Zhejiang University 2Ant Group + +![](images/3c2faf1e4e6f52949b7ce3ddc3dedf11722d5a9cbb2298a876ba3a259a10ce09.jpg) +Test on the HUAMNISE + +![](images/99d0d5a473d70f190d9a6114d0a2c4f4be9a882e72e49bd5e9eaac6acaa55af2.jpg) +Figure 1. Generating human motions in 3D scenes from text descriptions. Our method can generate human motions containing accurate human-object interactions in 3D scenes based on textural descriptions. Although our method is trained and tested on the HUMANISE dataset, it can generalize to other scenes, e.g., the scenes in the PROX dataset. Left: test results on the HUMANISE dataset. Right: generalization results on the PROX scenes. + +![](images/8388d847c4db90f8c836b8462a4fa13a7484f0ba97ae2a0a8ec189249e1fe094.jpg) +Generalize to the PROX scenes + +![](images/3fd1225ec84ef23dc5e91c1bec639a253143045e3118f2216cb8f3151dd1b3da.jpg) + +# Abstract + +Generating human motions from textual descriptions has gained growing research interest due to its wide range of applications. However, only a few works consider human-scene interactions together with text conditions, which is crucial for visual and physical realism. This paper focuses on the task of generating human motions in 3D indoor scenes given text descriptions of the human-scene interactions. This task presents challenges due to the multimodality nature of text, scene, and motion, as well as the need for spatial reasoning. To address these challenges, we propose a new approach that decomposes the complex problem into two more manageable sub-problems: (1) language grounding of the target object and (2) object-centric motion generation. For language grounding of the target object, we leverage the power of large language models. For motion generation, we design an object-centric scene representation for the generative model to focus on the target object, thereby reducing the scene complexity and facilitating the modeling of the relationship between human motions and the object. Experiments demonstrate the better motion quality of our approach compared to baselines and validate our design choices. Code will be available at link. + +# 1. Introduction + +Human motion generation has been a long-standing problem due to its broad range of applications such as game development, virtual reality, and movie production. Recently, this area witnessed a paradigm shift from avatar animation given rich user input [29] to learning-based motion generation from high-level language prompts, e.g. text descriptions about the desired motion [2, 3, 15, 18, 19, 39, 56, 74, 75]. However, most prior works on text-driven motion synthesis do not consider human-scene interactions [39, 56, 74, 75] while the scene context and physical constraints of the environment largely define the fidelity of the generated human motions. + +In this paper, we focus on generating motions from text descriptions in 3D indoor scenes. Specifically, given a 3D scan of the target scene and a text description of a human action that interacts with the scene, we aim to generate natural human motions that are consistent with the text description. + +This problem presents several challenges, primarily due to the multi-modality nature of text, scene, and human motion. In contrast to previous methods [2, 3, 56, 74, 75] focusing solely on textual descriptions of how human moves, our task also includes texts that additionally describe the spatial details in the given scene (e.g., sit on the armchair near the desk). Therefore, this task requires spatial reason- + +ing skills [1], where the model should build a text-object mapping to locate a specific object in 3D scenes aligned with a natural language description. In addition, the generated motions should also be coherent with scene contexts. + +As a pioneer work, HUMANISE [83] builds a conditional variational autoencoder (cVAE) [41, 69] with separate encoders of scenes and texts for multi-modality understanding. To enable spatial reasoning ability, they introduce auxiliary tasks of directly regressing object centers to learn 3D visual grounding in an implicit manner. But they do not explicitly utilize the predicted centers and thus the inductive bias of visual grounding cannot be fully incorporated. In addition, HUMANISE [83] encodes the entire scene with a single point transformer [99]. Directly generating motions with such a model is challenging, as 3D point clouds are inherently noisy and complex [48, 88], leading to the inability to locate the target object. As suggested by [51], not every point of the scene is relevant to the final human motions. Therefore, it is also necessary to develop a more targeted approach that focuses on the relevant parts of the scene to improve the quality of motion generation. + +To tackle the problems mentioned above, we propose a novel approach that exploits the power of the large language models (LLMs) [6, 42, 52]. Our key idea is to address the challenging task of generating motion in a scene based on textual cues by breaking it down into two smaller problems: (1) language grounding of objects in 3D scenes and (2) generating motions with a focus on the target object. For language grounding, rather than directly learning text-object mapping, we propose to formulate it as question answering and utilize the large prior knowledge of LLMs. Specifically, we first construct scene graphs of 3D scenes and generate their textual descriptions. Then we employ ChatGPT [52] to analyze the relationship between scene descriptions and input instructions and respond with 3D visual grounding answers. Experiments prove the effectiveness of this strategy. + +For motion generation, we design an object-centric representation to help the generative model focus on the target object. Specifically, we convert point clouds around the target object into volumetric sensors [70] to build object-centric representation. Then we employ diffusion models [57, 75] to synthesize human motions given object-centric representation and texts. Compared with original scene point clouds which might have various scales, object-centric representation is more compact and robust to scales as objects in the same category are of similar size. Therefore, this representation reduces scene complexity and facilitates the modeling of the relationship between motions and objects. + +We conducted thorough comparative and ablation experiments on the HUMANISE dataset. The results demonstrate that our method outperforms the baseline, reflected in more accurate object grounding results and better motions that align with the textual descriptions and scenes. We further + +show that our approach can generalize to the PROX dataset [22] without any fine-tuning. + +# 2. Related work + +# 2.1. Motion synthesis + +Deep learning for motion synthesis. Deep learning methods for motion synthesis have attracted increasing attention in recent years [14, 21, 25, 28, 29, 50, 71, 91]. Various techniques, including MLP [29], mixture of experts (MoE) [72, 91], and recurrent neural network (RNN) [21], are employed to tackle this task. To generate diverse motion results, previous works explore cVAE [46, 96], GANs [44, 61], normalizing flow [26], and diffusion models [75]. + +Text-driven motion generation. Recently, there has been a growing interest in text-driven motion generation [2, 56, 74]. This task takes natural language as input and synthesizes human motions that align with instructions. KIT-ML [58] is the first benchmark of this task. Some works further annotate the AMASS dataset [49] with action labels [59] and text [18]. To tackle this task, [2, 15, 74] propose to learn a shared latent space for text and motion. [56] employs transformer VAE [55] to generate diverse results. [19, 93] achieve better performance by discrete representation with VQ-VAE [65, 78]. [12, 39, 75, 94] successfully apply diffusion models in this direction. [5, 43] further explore latent diffusion models [67]. Based on [75], [38, 85] introduce sparse spatial control. [3, 60, 68] also investigate long sequence motion generation. Other existing works [4, 37, 82, 98] leverage Large Language Models (LLM) [6, 42, 53, 77] in the motion domain. [4] construct compositional motions by combining different body parts with LLM. [36, 98] regard motions as a kind of language and finetune LLMs [63] with motion tokens [19, 93]. + +Scene-aware motion generation This direction is to generate human motions in a 3D scene [70, 80]. To synthesize walking and sitting motions, [70] construct volumetric sensors to encode object information and the surroundings of the character. [23] extends it to synthesize motions with diverse sitting and lying styles. Furthermore, [95] proposes to control sitting styles by hand contacts. To improve the performance, [57] employs a hierarchical framework that generates goal poses, milestones, and motion sequentially. [73, 84] synthesize whole-body (body and hand fingers) grasping motions with small objects. [16, 45, 71, 86] explore interactions with dynamic objects including manipulation [16, 45] and carrying [86]. [8, 24] explore human-object interactions with physically simulated characters. These works mainly focus on one or two objects while others [80, 81] consider a more complex scene input (e.g., point clouds of the scene including walls and floors.). [80, 81] adopt hierarchical frameworks and generate trajectory and + +![](images/908da739b3d8d4a510d67112f70e53a8bb30397b1805dbfc1c3a8aa426e04417.jpg) +Figure 2. Overview of our two-stage pipeline. In the first stage, given an input scene and a text description (a), we use ChatGPT to locate the target object (b). In the second stage, human motions are synthesized by first producing human trajectories (c) and then generating local poses (d). + +motion separately. [103] introduces gaze to help generation. [33] further employs diffusion models and [51] synthesizes very long-term motions in scenes by dividing long sequences into several short sequences. [101] designs a reinforcement learning pipeline to enable navigation in a complex scene and interaction style control. + +Text-driven scene-aware motion generation Only a few works consider text and scenes simultaneously [38, 83, 101]. Although [101] enables text control of sitting styles by [100], the text descriptions in our setting are used to select an object in a cluttered scene. [38] could avoid obstacles during walking while our task needs to handle various actions. The most relevant work to us is HUMANISE [83], which employs a transformer VAE architecture with a two-stream condition module for text and scenes. To accurately localize target objects, they design auxiliary tasks like directly regressing object centers. In contrast to HUMANISE, we propose a two-stage pipeline where we first localize the target object with the help of ChatGPT [52] and then generate human motion using the object-centric representation. + +# 2.2. 3D visual grounding. + +In recent years, visual grounding in 3D scenes has been explored [1, 11, 76] and also tackled in 3D question answering [13, 89]. Given the point cloud of 3D scenes, this task [1] requires models to locate the target object according to text instructions. Most works follow a two-stage scheme where multiple bounding boxes [9, 66, 102] or segmentation results [31, 90] are first predicted and then selecting the object according to language descriptions. [32] introduces multiview inputs, and [88] employs 2D semantics. [34] combines bottom-up [47] and top-down [7] detection methods. [48] designs a single-stage pipeline by progressively selecting key points. [20] extends [32] with the help of GPT [6] to generate multi-view text inputs. More recently, [30] proposes a neuro-symbolic framework with large language-tocode models [10]. Most works only localize a single object and [97] could localize a flexible number of objects. Differ + +ent from previous works which directly handle point clouds or multi-view images, we convert the scene into textual descriptions and leverage large language models to infer the target object. Like [30], in this work, we also leverage large language models for object localization. + +# 3. Problem setup and preliminaries + +In this section, we discuss the definition of the task and preliminaries. We aim to generate human motion that is consistent with both the text description and the given scene. + +Text descriptions. The text description follows the template in Sr3D [1] (e.g., “ [ ]”). There are four actions (walk, sit, stand up, and lie). The target represents the object that the agents need to interact with, and the anchor objects help to determine the target. A certain type of target furniture category usually has many instances in one scene, while the anchor furniture should be unique. To specify the exact target object, there are five types of spatial relations [1] between the target and the anchor: horizontal proximity, vertical proximity, between, allocentric, and support. + +Scene representations. The scene is denoted as a point cloud of $N$ points: $S \in \mathbf{R}^{N \times 6}$ , containing the position and normal direction information of each point. + +Motion representations. The output motion sequence is represented as a sequence of SMPL-X [54] body meshes $M$ . SMPL-X is a parametric human body model. In this work, body parameters include body shape parameters $\beta \in \mathbf{R}^{10}$ , global translation $r \in \mathbf{R}^3$ , 6D global orientation $\gamma \in \mathbf{R}^6$ , and 6D pose parameters of $J$ joints $\theta \in \mathbf{R}^{J \times 6}$ [104]. Following [83], $\beta$ is treated as a condition to model the effect of body shape, and we omit it for ease of notation. Note that start position and pose are also generated by the model. + +Diffusion models. The diffusion [27] is defined as a Markov noising process $\{\mathbf{x}_t\}_{t=0}^T$ that follows $q(\mathbf{x}_t \mid \mathbf{x}_0)$ , where $\mathbf{x}_0 \sim q(\mathbf{x}_0)$ is the data and $\mathbf{x}_t$ is the noised data at + +![](images/5383c1e92cfa5b5fcccbdc841b5a757e4db03d6e018bee428aa39de50b0ffe44.jpg) +Figure 3. Pipeline of localizing the target object. In stage 1, given the input text description and detected object bounding boxes (bbx), we construct the first prompt asking ChatGPT the categories of target objects and anchor objects. Based on the response, the scene graph can be simplified. In stage 2, we construct the second prompt with inputs and results from stage 1, including object relations derived from the simplified scene graph. The second prompt is designed for asking ChatGPT to infer the target object. Finally, we can get the target object bounding box from the response of ChatGPT. + +noising step $t$ . The formal definition is: + +$$ +q \left(\mathbf {x} _ {t} \mid \mathbf {x} _ {0}\right) = \mathcal {N} \left(\mathbf {x} _ {t}; \sqrt {\bar {\alpha} _ {t}} \mathbf {x} _ {0}, (1 - \bar {\alpha} _ {t}) \mathbf {I}\right), \tag {1} +$$ + +where $\bar{\alpha}_t$ are constants with monotonically decreasing schedule. When $\bar{\alpha}_t$ is small enough, we can approximate $\mathbf{x}_T\sim \mathcal{N}(\mathbf{0},\mathbf{I})$ . In our context, we use conditional diffusion models like [64, 75]. The training loss is defined as: + +$$ +\mathcal {L} = \mathrm {E} _ {t \in [ 1, T ], \mathbf {x} _ {0} \sim q (\mathbf {x} _ {0})} \left[ \| \mathbf {x} _ {0} - G (\mathbf {x} _ {t}, t, \mathbf {C}) \| \right], \tag {2} +$$ + +where $G$ is the generative model, and $\mathbf{C}$ is the condition. + +# 4. Method + +The overview of our method is shown in Fig. 2. In Sec. 4.1, we leverage ChatGPT to localize the target object given a 3D scene and a textual description. Based on the accurate localization, we can focus on the target object and employ an object-centric generation pipeline to separately synthesize trajectories (Sec. 4.2) and motions (Sec. 4.3). Implementation details are discussed in Sec. 4.4. + +# 4.1. Language grounding of objects in 3D scenes + +Scene-aware motion generation from textual descriptions requires the scene-understanding ability and build the relationship between scenes and texts. Since the instructions describe how the character moves and interacts with a single target object, the majority of the scene might bear negligible relevance to the final motions [51]. Motivated by this, we propose to locate the target object to identify the most pertinent information. Inspired by the recent progress of LLM [42, 52, 53], ChatGPT [52] is utilized to find the specific objects in the given text. We first obtain textual descriptions of scenes by building scene graphs. Then we feed them with text instructions to ChatGPT with specially designed prompts and parse the response to get target objects. + +Spatial scene graph extraction. To utilize LLM, the initial step involves converting a 3D scene into text. This is achieved by building a spatial scene graph. We utilize a pre-trained 3D object detection model in [47] to provide 3D box proposals. Subsequently, we follow the approach of [1] to obtain the relationships between objects. Specifically, for + +![](images/c41b916797d7a3209a895dd23392fcf0cc621003b8abd6c000d825dfa04cb7aa.jpg) + +![](images/d0b86e10e8804b71caabf3e792b51b9afa8c09f6f9d3abeb8812bc199a089d9b.jpg) +(b) Target sensor + +![](images/0e4333041ce6956879ecff3d96021dceeebb273bfd1c9d4af66bf4f8c9aba3c5.jpg) +(a) Target object bbx +(c) Environment sensor +Figure 4. The visualization of the environment sensor, target sensor, and trajectory sensor. The target sensor (b) gives detailed geometry of the target object. The environment sensor (c) gives coarse spatial information around the target object. The trajectory sensor (d) is located around the human. + +![](images/213367d8f1ff555c0fc4f75d1e3512092ba8160ed76b9d3897bbe6c2066926c4.jpg) +(d) Traj sensor + +every set of two objects, we can infer their relationships in three types from their bounding boxes: horizontal proximity, vertical proximity, and support as mentioned in Sec. 3; for every set of three objects, we can infer if one of them is in between of the other two objects. As detection results do not contain object poses, we do not construct allocentric relations (e.g., "the shelf that is behind the sofa"). Then, we build a scene graph, where each object is assigned as a node in the graph, and edges between nodes represent the relationships between objects. By converting 3D scenes into text, we can apply ChatGPT to extract meaningful insights from the data. + +Leveraging ChatGPT to localize the target object. A simple approach is to directly input the entire scene graph into ChatGPT and ask it to select the target objects. However, scenes may contain many objects, resulting in extremely long textural descriptions. We observe that ChatGPT is often confused in such settings and fails to respond with the right answer. To narrow the search space, we first employ ChatGPT to recognize objects that correlate to the provided text. We then exclude the unrelated objects from the scene graph, focusing solely on those with pertinent information, enabling us to pinpoint the target object effectively. This approach has the advantage of reducing the number of objects that need to be considered, making it easier for ChatGPT to identify the target object. + +As shown in Fig. 3, we construct two prompts in sequence. Take "sit on the chair that is in the middle of the board and the end table" as an example, we first need to nar + +row down the object category search space. In order to find which type of objects we care about, we construct the first prompt that asks ChatGPT to find out target objects and anchor objects. Target object is defined as the final object that we want the agent to interact with, which in this case, is the "table". Anchor objects are the objects that help with determining the target object, for there might be many chairs in one scene. Based on the target object "chair" and anchor objects "board" and "end table", we can filter out all the unrelated objects in the scene, only keeping chairs, the board, and the end table in our scene graph. Next, according to the simplified scene graph, we can describe object relations in text: every edge in the scene graph could be converted to an edge sentence like "chair 4 is far from the end table 0". Converting all edges to such edge sentences gives a full description of the current scene. Finally, we construct a second prompt by asking ChatGPT to infer the target object from the accumulated edge sentences. + +# 4.2. Diffusion-based trajectory generation + +Given the localized object from ChatGPT, we first generate the trajectory based on the instructions and then synthesize local human poses. Trajectory is defined as a sequence of characters' translations and orientations. As suggested by [51], not every point of the scene is relevant to the final human motions. Inspired by NSM [70] and ManipNet [92], we employ volumetric sensors (as shown in Fig. 4) around the target object to represent the scene. + +Object-centric scene representation. Denote the target object center location as $c_{o} = (c_{x},c_{y},c_{z})$ . We transform the point cloud of the scene $S$ to an object-centered coordinate axis centered at $c_{o}$ . To recognize the surrounding geometry of the target object, we create volumetric sensors called Environment Sensor and Target Sensor. + +Environment sensor. The environment sensor is centered at $c_{o}$ in a cubic shape with a volume of $4 \times 4 \times 4 \, m^{3}$ , containing $8 \times 8 \times 8$ cubic voxels as shown in Fig. 4 (c). It is constructed by collecting all occupancies $o_{s}$ , center positions $c_{v}$ , and normal directions $n_{v}$ of each voxel to form a feature vector $E$ . Like [70], the scene occupancy $o_{s} \in R^{1}$ in each voxel is defined based on the given scene mesh: + +$$ +o _ {s} = \left\{ \begin{array}{l l} 1 & \text {i f} d _ {s} < 0, \\ 0 & \text {i f} d _ {s} > a _ {s}, \\ 1 - \frac {d _ {s}}{a _ {s}} & \text {o t h e r w i s e ,} \end{array} \right. \tag {3} +$$ + +where $d_{s}$ is the signed distance between the scene mesh and the voxel center, and $a_{s}$ is the voxel edge length. $n_v$ is the normal direction of the closest scene point to the voxel center. The environment sensor provides coarse scene geometry around the target object. + +Target sensor. To capture the detailed geometry of the target object, we further build a target sensor. As we already + +obtain the 3D bounding box of the target object in Sec. 4.1, we crop the point clouds according to the bounding box and construct an $8 \times 8 \times 8$ cubic volumes that cover the bounding box, as shown in Fig. 4 (b). Target sensor $T$ is in the same form as the environment sensor $E$ , except the target sensor has a different voxel size, as visualized in Fig. 4 (b) and (c). + +Given the constructed object-centric representations, we follow [57] to employ a transformer decoder architecture [79], which enables arbitrary length motions. As for the text input, we use CLIP [62] text encoder to encode input text to the text feature $L$ . The time-step $t$ is injected into the decoder in sinusoidal position embeddings form [79]. In summary, the condition for this generation model is + +$$ +\mathbf {C} _ {\mathbf {t}} = \{L, E, T \}, \tag {4} +$$ + +where $L$ is the text feature, $E$ is the environment sensor, and $T$ is the target sensor. All the conditions are projected to the same dimension $D = 512$ by MLPs and summed with positional embeddings to form tokens. We use the simple objective described in Eq. 2 to train the trajectory generation model $G_{r}$ to generate the trajectory $\mathbf{r}_{1:N}$ with length $N$ . + +# 4.3. Diffusion-based motion completion + +Given the trajectory from Sec. 4.2, the next step is to complete the whole motion. Based on the generated trajectories, we construct a Trajectory Sensor $O$ to explicitly reason about the interaction between the character and scenes. + +Trajectory sensor. The trajectory sensor is also a volumetric sensor which is similar to the form of the environment sensor but is designed for ego-centric perception [95] as shown in Fig. 4 (d). Specifically, trajectory sensors are positioned around characters in each frame. This sensor $O_{i}$ is centered at the predicted root position of the $i$ -th frame and faces to the $i$ -th frame's predicted root orientation, containing $8 \times 8 \times 8$ cubic voxels that store scene occupancy. The occupancy calculation is the same as Eq. 3. + +Another transformer-based conditional diffusion model is used to synthesize local poses along trajectories. The condition is defined as: + +$$ +\mathbf {C} _ {\mathbf {m}} = \{L, E, T, O _ {1}, \dots , O _ {N} \}, \tag {5} +$$ + +where $L, E,$ and $T$ have the same meanings in Eq. 4. Simple objective described in Eq. 2 is used to train the motion generation model $G_{m}$ to generate global orientation $\gamma_{1:N}$ and local poses $\theta_{1:N}$ with length $N$ . + +# 4.4. Implementation details + +Following [38, 57, 80, 81], we use separate diffusion models for the trajectory generation and the motion generation. Because the HUMANISE dataset [83] contains a relatively small number (51 minutes) of pure motion samples from the AMASS dataset (3772 minutes) [49], we first pre-train the + +models on the whole AMASS dataset for 200 epochs and then fine-tune on the HUMANISE dataset for 200 epochs to improve the motion quality. During pretraining, the text feature $L$ is set to all zeros. Both models are trained with the AdamW optimizer [40], using a learning rate of 0.0001 on a single Nvidia RTX 3090 GPU. The batch size is set to 128. The version of ChatGPT is gpt-3.5-turbo. More details can be found in the supplementary material. + +# 5. Experiments + +# 5.1. Evaluation metrics + +We evaluate the generated motions in three aspects: scene-conditional, action-conditional, and pure motion quality. + +Scene-conditional motion quality. To evaluate how the generated motion is aligned with the scene, we calculate the body-to-goal distance (goal dist.) [83] to measure how accurately the character interacts with target objects. However, goal dist. does not consider the consistency of the entire motion and scene (e.g., sitting on the sofa with incorrect orientation). To compensate for goal dist., a human perceptual study (indicated by scene score) is performed by randomly sampling 20 scenarios for each model, where ten workers are required to score each sample. + +Action-conditional motion quality. To measure how the generated motion is aligned with the text, we follow [17] to evaluate action recognition accuracy (accuracy), diversity, and multimodality of the results. Calculating these metrics relies on a pre-trained action recognition model [87] and we train the recognition model on the HUMANISE dataset. + +Pure motion quality. We evaluate the realism of generated motions using the metric suggested by [17], namely Frechet Inception Distance (FID). A lower FID indicates that generated motions are closer to the groundtruth motions. We also perform a human perceptual study (indicated by quality score) to measure pure motion quality. + +# 5.2. Generating human motion from text and scene + +Our experiments are conducted on the HUMANISE dataset [83]. Following the previous setting [83], there are $16.5k$ motion sequences in 543 scenes for training and $3.1k$ motion sequences in 100 scenes for testing. We compare our method with four baselines: (1) $\mathbf{MDM}^*$ [75]: a diffusion-based motion generator. (2) $\mathbf{GMD}^*$ [38]: GMD proposes various techniques for enhancing the control and quality of MDM. We employ the single-stage setting of $\mathbf{GMD}^*$ . Similar to HUMANISE, we use a point-transformer to provide scene features for $\mathbf{MDM}^*$ and $\mathbf{GMD}^*$ . (3) $\mathbf{GMD}^{\mathbf{HC}}$ : we provide object centers predicted by HUMANISE (denoted by HC) to guide the motion generation process in GMD by adding a proximity loss to encourage the motion to be close to the predicted object center. The proximity loss is defined + +
MethodsScene-conditionalAction-conditionalPure motion quality
goal dist.↓scene score↑accuracy↑diversity→multimodality→FID↓quality score↑
Real0.014-99.1%4.822.280.00-
MDM*2.0481.9996.0%4.832.570.162.45
GMD*1.2292.5096.6%4.912.280.242.64
GMDHC1.1302.4996.4%4.932.430.232.76
HUMANISE0.9952.3389.7%4.252.661.112.81
Ours0.3843.5497.1%4.782.310.123.50
+ +Table 1. Quantitative results on the HUMANISE dataset. We compare our method with four baselines (please refer to Sec. 5.2) and the real data. $\mathrm{GMD}^{\mathrm{HC}}$ means using HUMANISE's predicted center to guide motion generation in $\mathrm{GMD^{*}}$ . Among the metrics, scene score and quality score are perceptual studies. $\uparrow$ means higher is better and $\downarrow$ means lower is better. $\rightarrow$ means closer to the real data is better. Bold indicates the best results.underline indicates the second best. + +![](images/24ed49104be155df0172295846b1e03d3e6fe6329a55c141d2c34fae0ea45a7e.jpg) +Figure 5. Qualitative results. We compare our method with groundtruth and four baselines (please refer to Sec. 5.2) given the same text descriptions. Our method synthesizes motions that interact with the object precisely as the groundtruth data while the baselines fail. + +as the distance from HC to the predicted human pelvis at the interacting frame. (4) HUMANISE: we directly use their released models. + +The quantitative results are shown in Tab. 1. Our method outperforms the baseline in terms of goal dist., scene score, accuracy, FID, and quality score and achieves competitive results in diversity and multimodality. The quantitative results show that our approach can generate better-quality motions and are more consistent with all the conditions. The qualitative results are demonstrated in Fig. 5. Please refer to the supplementary material for more visualizations. + +# 5.3. Ablation study + +Ablation of main components. Four variants are constructed to explore the effect of our design choice. (1) "w/o localization": the localization module is removed and trajectories and motions are generated directly using scene point clouds. Scene features are given as the same form of MDM* and GMD*. (2) "w/o object-centric": we remove our object-centric representation and predict motions in the scene coordinate. The scene features of our sensors are still employed. (3) "w/o two-stage": trajectories and + +motions are generated together. (4) "w/o diffusion": we employ cVAE instead of the diffusion model as the architecture. (5) "w/o pretrain": the models are not pre-trained on the AMASS dataset. As shown in Tab. 2, the localization enhances the precise interaction ability and the object-centric representation improves motion quality. Although the ("w/o two-stage") is more efficient and has competitive performance metrics comparing to our method, it is prone to collapse directly into the target position, lacking gradual transition and realism. + +The design choice of object localization. For object localization, we compare our method with two baselines and five variants. (1) HUMANISE: predicts the object center in the auxiliary task. (2) BUTD-DETR [34]: is a 3D visual grounding method. (3) “ours w/o two-stage”: we employ a one-stage question-answering process. (4) “ours w/o few-shot”: the few-shot examples are not provided. (5) “ours using text matching”: we compute CLIP similarities instead of using ChatGPT. (6) “ours using LLaMA”: ChatGPT is replaced by LLaMA 2 7B model. (7) “ours using Mistral”: ChatGPT is replaced by Mistral 7B [35], an open-source LLM. The prompts for LLaMA 2 and Mistral + +
VariantsSceneActionPure
goal dist.↓accuracy↑diversity→mm→FID↓
Real0.01499.1%4.822.280.00
w/o localization0.59274.5%3.712.753.14
w/o object-centric0.41390.6%4.132.631.01
w/o two-stage0.38597.2%4.932.330.14
w/o diffusion0.39040.0%3.693.863.61
w/o pretrain0.39282.6%3.882.372.50
Ours0.38497.1%4.782.310.12
+ +7B are slightly adjusted. We evaluate these variants under two scenarios: predicted detection [47] and groundtruth detection. The metrics include the accuracy (acc.) and the distance from predicted centers to the object center (center dist.). "acc." is the percentage of times with an IoU (Intersection over Union) higher than the threshold (0.25) following [34]. The results are shown in Tab. 3. Since the results of Mistral are only slightly behind of ChatGPT, ChatGPT can be replaced by Mistral for better reproducibility, but cannot be replaced by the text matching method. Moreover, if groundtruth detection is provided, our method can achieve even better results. + +The design choice of sensor density. Please refer to the supplementary material. + +# 5.4. Generalization + +To validate the generalization ability of our method, we run our method directly on the unseen PROX dataset [22] without fine-tuning. We provide illustrations in Fig. 6. With our localization method based on ChatGPT and motion generation method based on volumetric sensors, our pipeline can easily generalize to other datasets. Please refer to the supplementary material for more results. + +# 6. Discussion + +We have demonstrated that our approach could synthesize motions with better quality and more precise interactions. However, restricted by the dataset, the duration of motions is relatively short (60% are around 1-3s), and most texts follow the template [1] without detailed descriptions. We only tackle static scenes and extending our method to interact with moving objects is a future direction. We acknowledge that leveraging LLMs has several problems. ChatGPT may fail when the detector misses the objects and the behavior of ChatGPT is regulated by prompts [37]. Despite we have shown in Tab. 3 that ChatGPT can be replaced by an open source LLM Mistral 7B, using LLMs instead of classical parsing approaches is less efficient in the inference stage. + +Table 2. Ablation of main components. We compare our method with five variants (please refer to Sec. 5.3). Among them, $mm$ indicates multimodality. Bold indicates the best results. Underline indicates the second best. + +
VariantsPredicted detectionGT detection
acc. ↑center dist.↓acc. ↑center dist.↓
HUMANISE-1.48--
BUTD-DETR62.51.3363.11.23
Ours w/o two-stage49.01.3458.51.10
Ours w/o few-shot69.30.8086.30.45
Ours using text matching67.00.8979.70.52
Ours using LLaMA 240.31.5542.51.52
Ours using Mistral 7B72.40.7487.20.33
Ours using ChatGPT-3.575.60.6190.50.37
+ +Table 3. Design choices of object localization. We compare our method with two baselines and five variants (please refer to Sec. 5.3). Since HUMANISE directly regresses the coordinate of centers without utilizing groundtruth detection, we do not calculate acc. and only include their results under predicted detection. Bold indicates the best results. + +![](images/9cb3d200c0b4f617df9ff313c8771453e220e0b467a3816bf5906c45bfb74fb4.jpg) + +![](images/4e37dd5c6339ff292e5995d045529d6ddf1c2f70df48c51cffe23d69863af90e.jpg) + +![](images/3ad9ea70bd24639ce0169e49bf9d1b94fda76fffa27572b243d415862b60859b.jpg) +Figure 6. Qualitative results of our method on the PROX dataset. We run our method on the scenes from the PROX dataset without fine-tuning. Results show that our method is capable to generalize to unseen scenes and objects. + +![](images/0bb381f1c3adbf42f2abb0023b608d75552d245318cc5c6225472174caf46fea.jpg) + +# 7. Conclusion + +In this work, we introduce a novel method for generating motion from text in a scene. To tackle this problem, we propose a two-step approach. The first step involves 3D visual grounding, where we identify the target object. In the second step, concentrating on the target object, we build a diffusion-based motion generation method. Our approach offers several advantages, including improved 3D visual grounding accuracy and motion quality. + +# Acknowledgement + +This work was partially supported by the NSFC (No. 62172364), Ant Group and Information Technology Center and State Key Lab of CAD&CG, Zhejiang University. + +# References + +[1] Panos Achlioptas, Ahmed Abdelreheem, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In ECCV, 2020. 2, 3, 4, 8 +[2] Chaitanya Ahuja and Louis-Philippe Morency. Language2pose: Natural language grounded pose forecasting. In 3DV, 2019. 1, 2 +[3] Nikos Athanasiou, Mathis Petrovich, Michael J. Black, and Gül Varol. TEACH: Temporal Action Compositions for 3D Humans. In 3DV, 2022. 1, 2 +[4] Nikos Athanasiou, Mathis Petrovich, Michael J. Black, and Gül Varol. SINC: Spatial composition of 3D human motions for simultaneous action generation. ICCV, 2023. 2 +[5] German Barquero, Sergio Escalera, and Cristina Palmero. Belfusion: Latent diffusion for behavior-driven human motion prediction. In ICCV, 2023. 2 +[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In NeurIPS, 2020. 2, 3 +[7] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 3 +[8] Yu-Wei Chao, Jimei Yang, Weifeng Chen, and Jia Deng. Learning to sit: Synthesizing human-chair interactions via hierarchical control. In AAAI, 2021. 2 +[9] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In ECCV, 2020. 3 +[10] Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, et al. Evaluating large language models trained on code. arXiv preprint, 2021. 3 +[11] Yixin Chen, Qing Li, Deqian Kong, Yik Lun Kei, Song-Chun Zhu, Tao Gao, Yixin Zhu, and Siyuan Huang. Yourfit: Embodied reference understanding with language and gesture. In ICCV, 2021. 3 +[12] Rishabh Dabral, Muhammad Hamza Mughal, Vladislav Golyanik, and Christian Theobalt. Mofusion: A framework for denoising-diffusion-based motion synthesis. In CVPR, 2023. 2 +[13] Abhishek Das, Georgia Gkioxari, Stefan Lee, Devi Parikh, and Dhruv Batra. Neural Modular Control for Embodied Question Answering. In CoRL, 2018. 3 +[14] Katerina Fragkiadaki, Sergey Levine, Panna Felsen, and Jitendra Malik. Recurrent network models for human dynamics. In ICCV, 2015. 2 + +[15] Anindita Ghosh, Noshaba Cheema, Cennet Oguz, Christian Theobalt, and Philipp Slusallek. Synthesis of compositional animations from textual descriptions. In ICCV, 2021. 1, 2 +[16] Anindita Ghosh, Rishabh Dabral, Vladislav Golyanik, Christian Theobalt, and Philipp Slusallek. Imos: Intent-driven full-body motion synthesis for human-object interactions. In Computer Graphics Forum, 2023. 2 +[17] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3d human motions. In ACM MM, 2020. 6 +[18] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In CVPR, 2022. 1, 2 +[19] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In ECCV, 2022. 1, 2 +[20] Zoey Guo, Yiwen Tang, Ray Zhang, Dong Wang, Zhigang Wang, Bin Zhao, and Xuelong Li. Viewrefer: Grasp the multi-view knowledge for 3d visual grounding. In ICCV, 2023. 3 +[21] Félix G Harvey, Mike Yurick, Derek Nowrouzezahrai, and Christopher Pal. Robust motion in-betweening. ACM Trans. Graph., 2020. 2 +[22] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J. Black. Resolving 3d human pose ambiguities with 3d scene constraints. In ICCV, 2019. 2, 8 +[23] Mohamed Hassan, Duygu Ceylan, Ruben Villegas, Jun Saito, Jimei Yang, Yi Zhou, and Michael J. Black. Stochastic scene-aware motion prediction. In ICCV, 2021. 2 +[24] Mohamed Hassan, Yunrong Guo, Tingwu Wang, Michael Black, Sanja Fidler, and Xue Bin Peng. Synthesizing physical character-scene interactions. In SIGGRAPH, 2023. 2 +[25] Chengan He, Jun Saito, James Zachary, Holly Rushmeier, and Yi Zhou. Nemf: Neural motion fields for kinematic animation. In NeurIPS, 2022. 2 +[26] Gustav Eje Henter, Simon Alexanderson, and Jonas Beskow. Moglow: Probabilistic and controllable motion synthesis using normalising flows. ACM Trans. Graph., 2020. 2 +[27] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3 +[28] Daniel Holden, Jun Saito, and Taku Komura. A deep learning framework for character motion synthesis and editing. ACM Trans. Graph., 2016. 2 +[29] Daniel Holden, Taku Komura, and Jun Saito. Phase-functional neural networks for character control. ACM Trans. Graph., 2017. 1, 2 +[30] Joy Hsu, Jiayuan Mao, and Jiajun Wu. Ns3d: Neurosymbolic grounding of 3d objects and relations. In CVPR, 2023. 3 +[31] Pin-Hao Huang, Han-Hung Lee, Hwann-Tzong Chen, and Tyng-Luh Liu. Text-guided graph neural networks for referring 3d instance segmentation. AAAI, 2021. 3 +[32] Shijia Huang, Yilun Chen, Jiaya Jia, and Liwei Wang. Multi-view transformer for 3d visual grounding. In CVPR, 2022. 3 + +[33] Siyuan Huang, Zan Wang, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, and Song-Chun Zhu. Diffusion-based generation, optimization, and planning in 3d scenes. In CVPR, 2023. 3 +[34] Ayush Jain, Nikolaos Gkanatsios, Ishita Mediratta, and Katerina Fragkiadaki. Bottom up top down detection transformers for language grounding in images and point clouds. In ECCV, 2022. 3, 7, 8 +[35] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 7 +[36] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. arXiv preprint arXiv:2306.14795, 2023. 2 +[37] Sai Shashank Kalakonda, Shubh Maheshwari, and Ravi Kiran Sarvadevabhatla. Action-gpt: Leveraging large-scale language models for improved and generalized action generation. In ICME, 2023. 2, 8 +[38] Korrawe Karunratanakul, Konpat Preechakul, Supasorn Suwajanakorn, and Siyu Tang. Guided motion diffusion for controllable human motion synthesis. In ICCV, 2023. 2, 3, 6 +[39] Jihoon Kim, Jiseob Kim, and Sungjoon Choi. Flame: Freeform language-based motion synthesis & editing. In AAAI, 2023. 1, 2 +[40] Diederik P. Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization. arXiv e-prints, 2014. 6 +[41] Diederik P Kingma and Max Welling. Auto-Encoding Variational Bayes. arXiv e-prints, 2013. 2 +[42] Takeshi Kojima, Shixiang (Shane) Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners. In NeurIPS, 2022. 2, 4 +[43] Hanyang Kong, Kehong Gong, Dongze Lian, Michael Bi Mi, and Xinchao Wang. Priority-centric human motion generation in discrete latent space. In ICCV, 2023. 2 +[44] Peizhuo Li, Kfir Aberman, Zihan Zhang, Rana Hanocka, and Olga Sorkine-Hornung. Ganimator: Neural motion synthesis from a single sequence. ACM Trans. Graph., 2022. 2 +[45] Quanzhou Li, Jingbo Wang, Chen Change Loy, and Bo Dai. Task-oriented human-object interactions generation with implicit neural representations. arXiv preprint, 2023. 2 +[46] Hung Yu Ling, Fabio Zinno, George Cheng, and Michiel Van De Panne. Character controllers using motion vaes. ACM Trans. Graph., 2020. 2 +[47] Ze Liu, Zheng Zhang, Yue Cao, Han Hu, and Xin Tong. Group-free 3d object detection via transformers. In ICCV, 2021. 3, 4, 8 +[48] Junyu Luo, Jiahui Fu, Xianghao Kong, Chen Gao, Haibing Ren, Hao Shen, Huaxia Xia, and Si Liu. 3d-sps: Single-stage 3d visual grounding via referred point progressive selection. In CVPR, 2022. 2, 3 +[49] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In ICCV, 2019. 2, 6 + +[50] Julieta Martinez, Michael J. Black, and Javier Romero. On human motion prediction using recurrent neural networks. In CVPR, 2017. 2 +[51] Aymen Mir, Xavier Puig, Angjoo Kanazawa, and Gerard Pons-Moll. Generating continual human motion in diverse 3d scenes. arXiv preprint, 2023. 2, 3, 4, 5 +[52] OpenAI. Openai: Introducing chatgpt. https://openai.com/blog/chatgpt, 2022. 2, 3, 4 +[53] OpenAI. Gpt-4 technical report, 2023. 2, 4 +[54] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3d hands, face, and body from a single image. In CVPR, 2019. 3 +[55] Mathis Petrovich, Michael J. Black, and Gül Varol. Action-conditioned 3D human motion synthesis with transformer VAE. In ICCV, 2021. 2 +[56] Mathis Petrovich, Michael J. Black, and Gül Varol. TEMOS: Generating diverse human motions from textual descriptions. In ECCV, 2022. 1, 2 +[57] Huaijin Pi, Sida Peng, Minghui Yang, Xiaowei Zhou, and Hujun Bao. Hierarchical generation of human-object interactions with diffusion probabilistic models. In ICCV, 2023. 2, 6 +[58] Matthias Plappert, Christian Mandery, and Tamim Asfour. The kit motion-language dataset. *Big data*, 2016. 2 +[59] Abhinanda R. Punnakkal, Arjun Chandrasekaran, Nikos Athanasiou, Alejandra Quiros-Ramirez, and Michael J. Black. Babel: Bodies, action and behavior with english labels. In CVPR, 2021. 2 +[60] Yijun Qian, Jack Urbanek, Alexander G Hauptmann, and Jungdam Won. Breaking the limits of text-conditioned 3d motion synthesis with elaborative descriptions. In ICCV, 2023. 2 +[61] Sigal Raab, Inbal Leibovitch, Peizhuo Li, Kfir Aberman, Olga Sorkine-Hornung, and Daniel Cohen-Or. Modi: Unconditional motion synthesis from diverse data. In CVPR, 2023. 2 +[62] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 6 +[63] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485-5551, 2020. 2 +[64] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical Text-Conditional Image Generation with CLIP Latents. arXiv e-prints, 2022. 4 +[65] Ali Razavi, Aaron van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. In NeurIPS, 2019. 2 +[66] Junha Roh, Karthik Desingh, Ali Farhadi, and Dieter Fox. Languagerefer: Spatial-language model for 3d visual grounding. In CoRL, 2022. 3 + +[67] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2 +[68] Yonatan Shafir, Guy Tevet, Roy Kapon, and Amit H Bermano. Human motion diffusion as a generative prior. arXiv preprint, 2023. 2 +[69] Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. In NeurIPS, 2015. 2 +[70] Sebastian Starke, He Zhang, Taku Komura, and Jun Saito. Neural state machine for character-scene interactions. ACM Trans. Graph., 2019. 2, 5 +[71] Sebastian Starke, Yiwei Zhao, Taku Komura, and Kazi Zaman. Local motion phases for learning multi-contact character movements. ACM Trans. Graph., 2020. 2 +[72] Sebastian Starke, Ian Mason, and Taku Komura. Deepphase: Periodic autoencoders for learning motion phase manifolds. ACM Trans. Graph., 2022. 2 +[73] Omid Taheri, Vasileios Choutas, Michael J. Black, and Dimitrios Tzionas. Goal: Generating 4d whole-body motion for hand-object grasping. In CVPR, 2022. 2 +[74] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. In ECCV, 2022. 1, 2 +[75] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In ICLR, 2023. 1, 2, 4, 6 +[76] Jesse Thomason, Mohit Shridhar, Yonatan Bisk, Chris Paxton, and Luke Zettlemoyer. Language grounding with 3d objects. In CoRL, 2022. 3 +[77] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models. arXiv preprint, 2023. 2 +[78] Aaron van den Oord, Oriol Vinyals, and koray kavukcuoglu. Neural discrete representation learning. In NeurIPS, 2017. 2 +[79] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 6 +[80] Jiashun Wang, Huazhe Xu, Jingwei Xu, Sifei Liu, and Xiaolong Wang. Synthesizing long-term 3d human motion and interaction in 3d scenes. In CVPR, 2021. 2, 6 +[81] Jingbo Wang, Yu Rong, Jingyuan Liu, Sijie Yan, Dahua Lin, and Bo Dai. Towards diverse and natural scene-aware 3d human motion synthesis. In CVPR, 2022. 2, 6 +[82] Xi Wang, Gen Li, Yen-Ling Kuo, Muhammed Kocabas, Emre Aksan, and Otmar Hilliges. Reconstructing action-conditioned human-object interactions using commonsense knowledge priors. In 3DV, 2022. 2 +[83] Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, Wei Liang, and Siyuan Huang. Humanise: Language-conditioned human motion generation in 3d scenes. In NeurIPS, 2022. 2, 3, 6 + +[84] Yan Wu, Jiahao Wang, Yan Zhang, Siwei Zhang, Otmar Hilliges, Fisher Yu, and Siyu Tang. Saga: Stochastic whole-body grasping with contact. In ECCV, 2022. 2 +[85] Yiming Xie, Varun Jampani, Lei Zhong, Deqing Sun, and Huaizu Jiang. Omnicontrol: Control any joint at any time for human motion generation. arXiv preprint, 2023. 2 +[86] Sirui Xu, Zhengyuan Li, Yu-Xiong Wang, and Liang-Yan Gui. Interdiff: Generating 3d human-object interactions with physics-informed diffusion. In ICCV, 2023. 2 +[87] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In AAAI, 2018. 6 +[88] Zhengyuan Yang, Songyang Zhang, Liwei Wang, and Jiebo Luo. Sat: 2d semantics assisted training for 3d visual grounding. In ICCV, pages 1856-1866, 2021. 2, 3 +[89] Shuquan Ye, Dongdong Chen, Songfang Han, and Jing Liao. 3d question answering. IEEE Transactions on Visualization and Computer Graphics, 2022. 3 +[90] Zhihao Yuan, Xu Yan, Yinghong Liao, Ruimao Zhang, Sheng Wang, Zhen Li, and Shuguang Cui. Instancerefer: Cooperative holistic understanding for visual grounding on point clouds through instance multi-level contextual referring. In ICCV, 2021. 3 +[91] He Zhang, Sebastian Starke, Taku Komura, and Jun Saito. Mode-adaptive neural networks for quadruped motion control. ACM Trans. Graph., 2018. 2 +[92] He Zhang, Yuting Ye, Takaaki Shiratori, and Taku Komura. Manipnet: Neural manipulation synthesis with a hand-object spatial representation. ACM Trans. Graph., 2021. 5 +[93] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Yong Zhang, Hongwei Zhao, Hongtao Lu, Xi Shen, and Ying Shan. Generating human motion from textual descriptions with discrete representations. In CVPR, 2023. 2 +[94] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint, 2022. 2 +[95] Xiaohan Zhang, Bharat Lal Bhatnagar, Sebastian Starke, Vladimir Guzov, and Gerard Pons-Moll. Couch: Towards controllable human-chair interactions. In ECCV, 2022. 2, 6 +[96] Yan Zhang and Siyu Tang. The wanderings of odysseus in 3d scenes. In CVPR, 2022. 2 +[97] Yiming Zhang, ZeMing Gong, and Angel X Chang. Multi3drefer: Grounding text description to multiple 3d objects. In ICCV, 2023. 3 +[98] Yaqi Zhang, Di Huang, Bin Liu, Shixiang Tang, Yan Lu, Lu Chen, Lei Bai, Qi Chu, Nenghai Yu, and Wanli Ouyang. Motiongpt: Finetuned llms are general-purpose motion generators. arXiv preprint, 2023. 2 +[99] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In ICCV, 2021. 2 +[100] Kaifeng Zhao, Shaofei Wang, Yan Zhang, Thabo Beeler, and Siyu Tang. Compositional human-scene interaction synthesis with semantic control. In ECCV, 2022. 3 +[101] Kaifeng Zhao, Yan Zhang, Shaofei Wang, Thabo Beeler, and Siyu Tang. Synthesizing diverse human motions in 3d indoor scenes. In ICCV, 2023. 3 + +[102] Lichen Zhao, Daigang Cai, Lu Sheng, and Dong Xu. 3dvg-transformer: Relation modeling for visual grounding on point clouds. In ICCV, 2021. 3 +[103] Yang Zheng, Yanchao Yang, Kaichun Mo, Jiaman Li, Tao Yu, Yebin Liu, C Karen Liu, and Leonidas J Guibas. Gimo: Gaze-informed human motion prediction in context. In ECCV, 2022. 3 +[104] Yi Zhou, Connelly Barnes, Lu Jingwan, Yang Jimei, and Li Hao. On the continuity of rotation representations in neural networks. In CVPR, 2019. 3 \ No newline at end of file diff --git a/2024/Generating Human Motion in 3D Scenes from Text Descriptions/images.zip b/2024/Generating Human Motion in 3D Scenes from Text Descriptions/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..11fb2542b1f52304a08c17c32a46cf4efba9dd03 --- /dev/null +++ b/2024/Generating Human Motion in 3D Scenes from Text Descriptions/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:40e263a091137294cd5cae5b529192c6643be73ed0310225d3c9dd5efbc3e22f +size 553434 diff --git a/2024/Generating Human Motion in 3D Scenes from Text Descriptions/layout.json b/2024/Generating Human Motion in 3D Scenes from Text Descriptions/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..35da9d7bf89c02e6a16c61232823bb6abeb175bc --- /dev/null +++ b/2024/Generating Human Motion in 3D Scenes from Text Descriptions/layout.json @@ -0,0 +1,9833 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 99, + 103, + 496, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 103, + 496, + 121 + ], + "spans": [ + { + "bbox": [ + 99, + 103, + 496, + 121 + ], + "type": "text", + "content": "Generating Human Motion in 3D Scenes from Text Descriptions" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 149, + 142, + 455, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 142, + 455, + 196 + ], + "spans": [ + { + "bbox": [ + 149, + 142, + 455, + 196 + ], + "type": "text", + "content": "Zhi Cen1 Huaijin Pi1 Sida Peng1† Zehong Shen1 Minghui Yang2 Shuai Zhu2 Hujun Bao1 Xiaowei Zhou1 1Zhejiang University 2Ant Group" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 52, + 224, + 173, + 322 + ], + "blocks": [ + { + "bbox": [ + 52, + 224, + 173, + 322 + ], + "lines": [ + { + "bbox": [ + 52, + 224, + 173, + 322 + ], + "spans": [ + { + "bbox": [ + 52, + 224, + 173, + 322 + ], + "type": "image", + "image_path": "3c2faf1e4e6f52949b7ce3ddc3dedf11722d5a9cbb2298a876ba3a259a10ce09.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 129, + 324, + 217, + 334 + ], + "lines": [ + { + "bbox": [ + 129, + 324, + 217, + 334 + ], + "spans": [ + { + "bbox": [ + 129, + 324, + 217, + 334 + ], + "type": "text", + "content": "Test on the HUAMNISE" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 175, + 224, + 295, + 322 + ], + "blocks": [ + { + "bbox": [ + 175, + 224, + 295, + 322 + ], + "lines": [ + { + "bbox": [ + 175, + 224, + 295, + 322 + ], + "spans": [ + { + "bbox": [ + 175, + 224, + 295, + 322 + ], + "type": "image", + "image_path": "99d0d5a473d70f190d9a6114d0a2c4f4be9a882e72e49bd5e9eaac6acaa55af2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 343, + 547, + 389 + ], + "lines": [ + { + "bbox": [ + 45, + 343, + 547, + 389 + ], + "spans": [ + { + "bbox": [ + 45, + 343, + 547, + 389 + ], + "type": "text", + "content": "Figure 1. Generating human motions in 3D scenes from text descriptions. Our method can generate human motions containing accurate human-object interactions in 3D scenes based on textural descriptions. Although our method is trained and tested on the HUMANISE dataset, it can generalize to other scenes, e.g., the scenes in the PROX dataset. Left: test results on the HUMANISE dataset. Right: generalization results on the PROX scenes." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 299, + 224, + 421, + 321 + ], + "blocks": [ + { + "bbox": [ + 299, + 224, + 421, + 321 + ], + "lines": [ + { + "bbox": [ + 299, + 224, + 421, + 321 + ], + "spans": [ + { + "bbox": [ + 299, + 224, + 421, + 321 + ], + "type": "image", + "image_path": "8388d847c4db90f8c836b8462a4fa13a7484f0ba97ae2a0a8ec189249e1fe094.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 363, + 324, + 479, + 334 + ], + "lines": [ + { + "bbox": [ + 363, + 324, + 479, + 334 + ], + "spans": [ + { + "bbox": [ + 363, + 324, + 479, + 334 + ], + "type": "text", + "content": "Generalize to the PROX scenes" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 423, + 224, + 543, + 321 + ], + "blocks": [ + { + "bbox": [ + 423, + 224, + 543, + 321 + ], + "lines": [ + { + "bbox": [ + 423, + 224, + 543, + 321 + ], + "spans": [ + { + "bbox": [ + 423, + 224, + 543, + 321 + ], + "type": "image", + "image_path": "3fd1225ec84ef23dc5e91c1bec639a253143045e3118f2216cb8f3151dd1b3da.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 400, + 192, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 400, + 192, + 412 + ], + "spans": [ + { + "bbox": [ + 143, + 400, + 192, + 412 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 423, + 290, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 423, + 290, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 290, + 687 + ], + "type": "text", + "content": "Generating human motions from textual descriptions has gained growing research interest due to its wide range of applications. However, only a few works consider human-scene interactions together with text conditions, which is crucial for visual and physical realism. This paper focuses on the task of generating human motions in 3D indoor scenes given text descriptions of the human-scene interactions. This task presents challenges due to the multimodality nature of text, scene, and motion, as well as the need for spatial reasoning. To address these challenges, we propose a new approach that decomposes the complex problem into two more manageable sub-problems: (1) language grounding of the target object and (2) object-centric motion generation. For language grounding of the target object, we leverage the power of large language models. For motion generation, we design an object-centric scene representation for the generative model to focus on the target object, thereby reducing the scene complexity and facilitating the modeling of the relationship between human motions and the object. Experiments demonstrate the better motion quality of our approach compared to baselines and validate our design choices. Code will be available at link." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 400, + 387, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 400, + 387, + 412 + ], + "spans": [ + { + "bbox": [ + 307, + 400, + 387, + 412 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 422, + 545, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 545, + 566 + ], + "type": "text", + "content": "Human motion generation has been a long-standing problem due to its broad range of applications such as game development, virtual reality, and movie production. Recently, this area witnessed a paradigm shift from avatar animation given rich user input [29] to learning-based motion generation from high-level language prompts, e.g. text descriptions about the desired motion [2, 3, 15, 18, 19, 39, 56, 74, 75]. However, most prior works on text-driven motion synthesis do not consider human-scene interactions [39, 56, 74, 75] while the scene context and physical constraints of the environment largely define the fidelity of the generated human motions." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 567, + 547, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 567, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 567, + 547, + 628 + ], + "type": "text", + "content": "In this paper, we focus on generating motions from text descriptions in 3D indoor scenes. Specifically, given a 3D scan of the target scene and a text description of a human action that interacts with the scene, we aim to generate natural human motions that are consistent with the text description." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": "This problem presents several challenges, primarily due to the multi-modality nature of text, scene, and human motion. In contrast to previous methods [2, 3, 56, 74, 75] focusing solely on textual descriptions of how human moves, our task also includes texts that additionally describe the spatial details in the given scene (e.g., sit on the armchair near the desk). Therefore, this task requires spatial reason-" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 693, + 287, + 715 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 693, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 693, + 287, + 715 + ], + "type": "text", + "content": "The authors from Zhejiang University are affiliated with the State Key Lab of CAD&CG. † Corresponding author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1855" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "ing skills [1], where the model should build a text-object mapping to locate a specific object in 3D scenes aligned with a natural language description. In addition, the generated motions should also be coherent with scene contexts." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 122, + 288, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 122, + 288, + 326 + ], + "spans": [ + { + "bbox": [ + 46, + 122, + 288, + 326 + ], + "type": "text", + "content": "As a pioneer work, HUMANISE [83] builds a conditional variational autoencoder (cVAE) [41, 69] with separate encoders of scenes and texts for multi-modality understanding. To enable spatial reasoning ability, they introduce auxiliary tasks of directly regressing object centers to learn 3D visual grounding in an implicit manner. But they do not explicitly utilize the predicted centers and thus the inductive bias of visual grounding cannot be fully incorporated. In addition, HUMANISE [83] encodes the entire scene with a single point transformer [99]. Directly generating motions with such a model is challenging, as 3D point clouds are inherently noisy and complex [48, 88], leading to the inability to locate the target object. As suggested by [51], not every point of the scene is relevant to the final human motions. Therefore, it is also necessary to develop a more targeted approach that focuses on the relevant parts of the scene to improve the quality of motion generation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 327, + 287, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 287, + 506 + ], + "type": "text", + "content": "To tackle the problems mentioned above, we propose a novel approach that exploits the power of the large language models (LLMs) [6, 42, 52]. Our key idea is to address the challenging task of generating motion in a scene based on textual cues by breaking it down into two smaller problems: (1) language grounding of objects in 3D scenes and (2) generating motions with a focus on the target object. For language grounding, rather than directly learning text-object mapping, we propose to formulate it as question answering and utilize the large prior knowledge of LLMs. Specifically, we first construct scene graphs of 3D scenes and generate their textual descriptions. Then we employ ChatGPT [52] to analyze the relationship between scene descriptions and input instructions and respond with 3D visual grounding answers. Experiments prove the effectiveness of this strategy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 509, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 509, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 509, + 287, + 652 + ], + "type": "text", + "content": "For motion generation, we design an object-centric representation to help the generative model focus on the target object. Specifically, we convert point clouds around the target object into volumetric sensors [70] to build object-centric representation. Then we employ diffusion models [57, 75] to synthesize human motions given object-centric representation and texts. Compared with original scene point clouds which might have various scales, object-centric representation is more compact and robust to scales as objects in the same category are of similar size. Therefore, this representation reduces scene complexity and facilitates the modeling of the relationship between motions and objects." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "We conducted thorough comparative and ablation experiments on the HUMANISE dataset. The results demonstrate that our method outperforms the baseline, reflected in more accurate object grounding results and better motions that align with the textual descriptions and scenes. We further" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 72, + 545, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 97 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 97 + ], + "type": "text", + "content": "show that our approach can generalize to the PROX dataset [22] without any fine-tuning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 110, + 391, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 110, + 391, + 122 + ], + "spans": [ + { + "bbox": [ + 306, + 110, + 391, + 122 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 130, + 408, + 143 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 130, + 408, + 143 + ], + "spans": [ + { + "bbox": [ + 306, + 130, + 408, + 143 + ], + "type": "text", + "content": "2.1. Motion synthesis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 153, + 545, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 153, + 545, + 250 + ], + "spans": [ + { + "bbox": [ + 305, + 153, + 545, + 250 + ], + "type": "text", + "content": "Deep learning for motion synthesis. Deep learning methods for motion synthesis have attracted increasing attention in recent years [14, 21, 25, 28, 29, 50, 71, 91]. Various techniques, including MLP [29], mixture of experts (MoE) [72, 91], and recurrent neural network (RNN) [21], are employed to tackle this task. To generate diverse motion results, previous works explore cVAE [46, 96], GANs [44, 61], normalizing flow [26], and diffusion models [75]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 254, + 546, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 254, + 546, + 494 + ], + "spans": [ + { + "bbox": [ + 305, + 254, + 546, + 494 + ], + "type": "text", + "content": "Text-driven motion generation. Recently, there has been a growing interest in text-driven motion generation [2, 56, 74]. This task takes natural language as input and synthesizes human motions that align with instructions. KIT-ML [58] is the first benchmark of this task. Some works further annotate the AMASS dataset [49] with action labels [59] and text [18]. To tackle this task, [2, 15, 74] propose to learn a shared latent space for text and motion. [56] employs transformer VAE [55] to generate diverse results. [19, 93] achieve better performance by discrete representation with VQ-VAE [65, 78]. [12, 39, 75, 94] successfully apply diffusion models in this direction. [5, 43] further explore latent diffusion models [67]. Based on [75], [38, 85] introduce sparse spatial control. [3, 60, 68] also investigate long sequence motion generation. Other existing works [4, 37, 82, 98] leverage Large Language Models (LLM) [6, 42, 53, 77] in the motion domain. [4] construct compositional motions by combining different body parts with LLM. [36, 98] regard motions as a kind of language and finetune LLMs [63] with motion tokens [19, 93]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 498, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 498, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 498, + 546, + 713 + ], + "type": "text", + "content": "Scene-aware motion generation This direction is to generate human motions in a 3D scene [70, 80]. To synthesize walking and sitting motions, [70] construct volumetric sensors to encode object information and the surroundings of the character. [23] extends it to synthesize motions with diverse sitting and lying styles. Furthermore, [95] proposes to control sitting styles by hand contacts. To improve the performance, [57] employs a hierarchical framework that generates goal poses, milestones, and motion sequentially. [73, 84] synthesize whole-body (body and hand fingers) grasping motions with small objects. [16, 45, 71, 86] explore interactions with dynamic objects including manipulation [16, 45] and carrying [86]. [8, 24] explore human-object interactions with physically simulated characters. These works mainly focus on one or two objects while others [80, 81] consider a more complex scene input (e.g., point clouds of the scene including walls and floors.). [80, 81] adopt hierarchical frameworks and generate trajectory and" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1856" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 70, + 516, + 180 + ], + "blocks": [ + { + "bbox": [ + 75, + 70, + 516, + 180 + ], + "lines": [ + { + "bbox": [ + 75, + 70, + 516, + 180 + ], + "spans": [ + { + "bbox": [ + 75, + 70, + 516, + 180 + ], + "type": "image", + "image_path": "908da739b3d8d4a510d67112f70e53a8bb30397b1805dbfc1c3a8aa426e04417.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 190, + 547, + 226 + ], + "lines": [ + { + "bbox": [ + 46, + 190, + 547, + 226 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 547, + 226 + ], + "type": "text", + "content": "Figure 2. Overview of our two-stage pipeline. In the first stage, given an input scene and a text description (a), we use ChatGPT to locate the target object (b). In the second stage, human motions are synthesized by first producing human trajectories (c) and then generating local poses (d)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 245, + 289, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 245, + 289, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 245, + 289, + 316 + ], + "type": "text", + "content": "motion separately. [103] introduces gaze to help generation. [33] further employs diffusion models and [51] synthesizes very long-term motions in scenes by dividing long sequences into several short sequences. [101] designs a reinforcement learning pipeline to enable navigation in a complex scene and interaction style control." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 323, + 287, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 323, + 287, + 490 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 287, + 490 + ], + "type": "text", + "content": "Text-driven scene-aware motion generation Only a few works consider text and scenes simultaneously [38, 83, 101]. Although [101] enables text control of sitting styles by [100], the text descriptions in our setting are used to select an object in a cluttered scene. [38] could avoid obstacles during walking while our task needs to handle various actions. The most relevant work to us is HUMANISE [83], which employs a transformer VAE architecture with a two-stream condition module for text and scenes. To accurately localize target objects, they design auxiliary tasks like directly regressing object centers. In contrast to HUMANISE, we propose a two-stage pipeline where we first localize the target object with the help of ChatGPT [52] and then generate human motion using the object-centric representation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 502, + 167, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 502, + 167, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 502, + 167, + 514 + ], + "type": "text", + "content": "2.2. 3D visual grounding." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 522, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 289, + 713 + ], + "type": "text", + "content": "In recent years, visual grounding in 3D scenes has been explored [1, 11, 76] and also tackled in 3D question answering [13, 89]. Given the point cloud of 3D scenes, this task [1] requires models to locate the target object according to text instructions. Most works follow a two-stage scheme where multiple bounding boxes [9, 66, 102] or segmentation results [31, 90] are first predicted and then selecting the object according to language descriptions. [32] introduces multiview inputs, and [88] employs 2D semantics. [34] combines bottom-up [47] and top-down [7] detection methods. [48] designs a single-stage pipeline by progressively selecting key points. [20] extends [32] with the help of GPT [6] to generate multi-view text inputs. More recently, [30] proposes a neuro-symbolic framework with large language-tocode models [10]. Most works only localize a single object and [97] could localize a flexible number of objects. Differ" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 245, + 545, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 245, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 304, + 245, + 545, + 304 + ], + "type": "text", + "content": "ent from previous works which directly handle point clouds or multi-view images, we convert the scene into textual descriptions and leverage large language models to infer the target object. Like [30], in this work, we also leverage large language models for object localization." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 321, + 489, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 321, + 489, + 335 + ], + "spans": [ + { + "bbox": [ + 305, + 321, + 489, + 335 + ], + "type": "text", + "content": "3. Problem setup and preliminaries" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 343, + 545, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 343, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 304, + 343, + 545, + 380 + ], + "type": "text", + "content": "In this section, we discuss the definition of the task and preliminaries. We aim to generate human motion that is consistent with both the text description and the given scene." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 385, + 546, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 385, + 546, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 385, + 546, + 517 + ], + "type": "text", + "content": "Text descriptions. The text description follows the template in Sr3D [1] (e.g., “ [ ]”). There are four actions (walk, sit, stand up, and lie). The target represents the object that the agents need to interact with, and the anchor objects help to determine the target. A certain type of target furniture category usually has many instances in one scene, while the anchor furniture should be unique. To specify the exact target object, there are five types of spatial relations [1] between the target and the anchor: horizontal proximity, vertical proximity, between, allocentric, and support." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 522, + 545, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 545, + 559 + ], + "type": "text", + "content": "Scene representations. The scene is denoted as a point cloud of " + }, + { + "bbox": [ + 304, + 522, + 545, + 559 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 522, + 545, + 559 + ], + "type": "text", + "content": " points: " + }, + { + "bbox": [ + 304, + 522, + 545, + 559 + ], + "type": "inline_equation", + "content": "S \\in \\mathbf{R}^{N \\times 6}" + }, + { + "bbox": [ + 304, + 522, + 545, + 559 + ], + "type": "text", + "content": ", containing the position and normal direction information of each point." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 563, + 545, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 563, + 545, + 672 + ], + "spans": [ + { + "bbox": [ + 304, + 563, + 545, + 672 + ], + "type": "text", + "content": "Motion representations. The output motion sequence is represented as a sequence of SMPL-X [54] body meshes " + }, + { + "bbox": [ + 304, + 563, + 545, + 672 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 563, + 545, + 672 + ], + "type": "text", + "content": ". SMPL-X is a parametric human body model. In this work, body parameters include body shape parameters " + }, + { + "bbox": [ + 304, + 563, + 545, + 672 + ], + "type": "inline_equation", + "content": "\\beta \\in \\mathbf{R}^{10}" + }, + { + "bbox": [ + 304, + 563, + 545, + 672 + ], + "type": "text", + "content": ", global translation " + }, + { + "bbox": [ + 304, + 563, + 545, + 672 + ], + "type": "inline_equation", + "content": "r \\in \\mathbf{R}^3" + }, + { + "bbox": [ + 304, + 563, + 545, + 672 + ], + "type": "text", + "content": ", 6D global orientation " + }, + { + "bbox": [ + 304, + 563, + 545, + 672 + ], + "type": "inline_equation", + "content": "\\gamma \\in \\mathbf{R}^6" + }, + { + "bbox": [ + 304, + 563, + 545, + 672 + ], + "type": "text", + "content": ", and 6D pose parameters of " + }, + { + "bbox": [ + 304, + 563, + 545, + 672 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 304, + 563, + 545, + 672 + ], + "type": "text", + "content": " joints " + }, + { + "bbox": [ + 304, + 563, + 545, + 672 + ], + "type": "inline_equation", + "content": "\\theta \\in \\mathbf{R}^{J \\times 6}" + }, + { + "bbox": [ + 304, + 563, + 545, + 672 + ], + "type": "text", + "content": " [104]. Following [83], " + }, + { + "bbox": [ + 304, + 563, + 545, + 672 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 563, + 545, + 672 + ], + "type": "text", + "content": " is treated as a condition to model the effect of body shape, and we omit it for ease of notation. Note that start position and pose are also generated by the model." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "text", + "content": "Diffusion models. The diffusion [27] is defined as a Markov noising process " + }, + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_t\\}_{t=0}^T" + }, + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "text", + "content": " that follows " + }, + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x}_t \\mid \\mathbf{x}_0)" + }, + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0 \\sim q(\\mathbf{x}_0)" + }, + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "text", + "content": " is the data and " + }, + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "text", + "content": " is the noised data at" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1857" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 86, + 71, + 509, + 371 + ], + "blocks": [ + { + "bbox": [ + 86, + 71, + 509, + 371 + ], + "lines": [ + { + "bbox": [ + 86, + 71, + 509, + 371 + ], + "spans": [ + { + "bbox": [ + 86, + 71, + 509, + 371 + ], + "type": "image", + "image_path": "5383c1e92cfa5b5fcccbdc841b5a757e4db03d6e018bee428aa39de50b0ffe44.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 379, + 548, + 436 + ], + "lines": [ + { + "bbox": [ + 46, + 379, + 548, + 436 + ], + "spans": [ + { + "bbox": [ + 46, + 379, + 548, + 436 + ], + "type": "text", + "content": "Figure 3. Pipeline of localizing the target object. In stage 1, given the input text description and detected object bounding boxes (bbx), we construct the first prompt asking ChatGPT the categories of target objects and anchor objects. Based on the response, the scene graph can be simplified. In stage 2, we construct the second prompt with inputs and results from stage 1, including object relations derived from the simplified scene graph. The second prompt is designed for asking ChatGPT to infer the target object. Finally, we can get the target object bounding box from the response of ChatGPT." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 456, + 208, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 456, + 208, + 468 + ], + "spans": [ + { + "bbox": [ + 47, + 456, + 208, + 468 + ], + "type": "text", + "content": "noising step " + }, + { + "bbox": [ + 47, + 456, + 208, + 468 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 456, + 208, + 468 + ], + "type": "text", + "content": ". The formal definition is:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 80, + 478, + 287, + 493 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 478, + 287, + 493 + ], + "spans": [ + { + "bbox": [ + 80, + 478, + 287, + 493 + ], + "type": "interline_equation", + "content": "q \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {0}\\right) = \\mathcal {N} \\left(\\mathbf {x} _ {t}; \\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {x} _ {0}, (1 - \\bar {\\alpha} _ {t}) \\mathbf {I}\\right), \\tag {1}", + "image_path": "998b337b0136f532a797895d7845f54e445d622923ff17c7fa446105c6cc5dfc.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 502, + 287, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 502, + 287, + 551 + ], + "spans": [ + { + "bbox": [ + 47, + 502, + 287, + 551 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 502, + 287, + 551 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t" + }, + { + "bbox": [ + 47, + 502, + 287, + 551 + ], + "type": "text", + "content": " are constants with monotonically decreasing schedule. When " + }, + { + "bbox": [ + 47, + 502, + 287, + 551 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t" + }, + { + "bbox": [ + 47, + 502, + 287, + 551 + ], + "type": "text", + "content": " is small enough, we can approximate " + }, + { + "bbox": [ + 47, + 502, + 287, + 551 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 47, + 502, + 287, + 551 + ], + "type": "text", + "content": ". In our context, we use conditional diffusion models like [64, 75]. The training loss is defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 76, + 561, + 287, + 575 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 561, + 287, + 575 + ], + "spans": [ + { + "bbox": [ + 76, + 561, + 287, + 575 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathrm {E} _ {t \\in [ 1, T ], \\mathbf {x} _ {0} \\sim q (\\mathbf {x} _ {0})} \\left[ \\| \\mathbf {x} _ {0} - G (\\mathbf {x} _ {t}, t, \\mathbf {C}) \\| \\right], \\tag {2}", + "image_path": "ce858c25327c6671b020ad18c20fd9dd4967182d549f9eddd99aeb6e83a9823f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 584, + 277, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 584, + 277, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 584, + 277, + 597 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 584, + 277, + 597 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 47, + 584, + 277, + 597 + ], + "type": "text", + "content": " is the generative model, and " + }, + { + "bbox": [ + 47, + 584, + 277, + 597 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 47, + 584, + 277, + 597 + ], + "type": "text", + "content": " is the condition." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 609, + 103, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 103, + 620 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 103, + 620 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 629, + 288, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 712 + ], + "type": "text", + "content": "The overview of our method is shown in Fig. 2. In Sec. 4.1, we leverage ChatGPT to localize the target object given a 3D scene and a textual description. Based on the accurate localization, we can focus on the target object and employ an object-centric generation pipeline to separately synthesize trajectories (Sec. 4.2) and motions (Sec. 4.3). Implementation details are discussed in Sec. 4.4." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 455, + 534, + 469 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 455, + 534, + 469 + ], + "spans": [ + { + "bbox": [ + 306, + 455, + 534, + 469 + ], + "type": "text", + "content": "4.1. Language grounding of objects in 3D scenes" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 478, + 547, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 478, + 547, + 634 + ], + "spans": [ + { + "bbox": [ + 304, + 478, + 547, + 634 + ], + "type": "text", + "content": "Scene-aware motion generation from textual descriptions requires the scene-understanding ability and build the relationship between scenes and texts. Since the instructions describe how the character moves and interacts with a single target object, the majority of the scene might bear negligible relevance to the final motions [51]. Motivated by this, we propose to locate the target object to identify the most pertinent information. Inspired by the recent progress of LLM [42, 52, 53], ChatGPT [52] is utilized to find the specific objects in the given text. We first obtain textual descriptions of scenes by building scene graphs. Then we feed them with text instructions to ChatGPT with specially designed prompts and parse the response to get target objects." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": "Spatial scene graph extraction. To utilize LLM, the initial step involves converting a 3D scene into text. This is achieved by building a spatial scene graph. We utilize a pre-trained 3D object detection model in [47] to provide 3D box proposals. Subsequently, we follow the approach of [1] to obtain the relationships between objects. Specifically, for" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1858" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 78, + 74, + 156, + 152 + ], + "blocks": [ + { + "bbox": [ + 78, + 74, + 156, + 152 + ], + "lines": [ + { + "bbox": [ + 78, + 74, + 156, + 152 + ], + "spans": [ + { + "bbox": [ + 78, + 74, + 156, + 152 + ], + "type": "image", + "image_path": "c41b916797d7a3209a895dd23392fcf0cc621003b8abd6c000d825dfa04cb7aa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 185, + 75, + 261, + 152 + ], + "blocks": [ + { + "bbox": [ + 185, + 75, + 261, + 152 + ], + "lines": [ + { + "bbox": [ + 185, + 75, + 261, + 152 + ], + "spans": [ + { + "bbox": [ + 185, + 75, + 261, + 152 + ], + "type": "image", + "image_path": "d0b86e10e8804b71caabf3e792b51b9afa8c09f6f9d3abeb8812bc199a089d9b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 186, + 156, + 260, + 166 + ], + "lines": [ + { + "bbox": [ + 186, + 156, + 260, + 166 + ], + "spans": [ + { + "bbox": [ + 186, + 156, + 260, + 166 + ], + "type": "text", + "content": "(b) Target sensor" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 58, + 167, + 154, + 258 + ], + "blocks": [ + { + "bbox": [ + 69, + 155, + 160, + 166 + ], + "lines": [ + { + "bbox": [ + 69, + 155, + 160, + 166 + ], + "spans": [ + { + "bbox": [ + 69, + 155, + 160, + 166 + ], + "type": "text", + "content": "(a) Target object bbx" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 167, + 154, + 258 + ], + "lines": [ + { + "bbox": [ + 58, + 167, + 154, + 258 + ], + "spans": [ + { + "bbox": [ + 58, + 167, + 154, + 258 + ], + "type": "image", + "image_path": "0e4333041ce6956879ecff3d96021dceeebb273bfd1c9d4af66bf4f8c9aba3c5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 259, + 167, + 269 + ], + "lines": [ + { + "bbox": [ + 67, + 259, + 167, + 269 + ], + "spans": [ + { + "bbox": [ + 67, + 259, + 167, + 269 + ], + "type": "text", + "content": "(c) Environment sensor" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 278, + 287, + 333 + ], + "lines": [ + { + "bbox": [ + 46, + 278, + 287, + 333 + ], + "spans": [ + { + "bbox": [ + 46, + 278, + 287, + 333 + ], + "type": "text", + "content": "Figure 4. The visualization of the environment sensor, target sensor, and trajectory sensor. The target sensor (b) gives detailed geometry of the target object. The environment sensor (c) gives coarse spatial information around the target object. The trajectory sensor (d) is located around the human." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 186, + 175, + 261, + 252 + ], + "blocks": [ + { + "bbox": [ + 186, + 175, + 261, + 252 + ], + "lines": [ + { + "bbox": [ + 186, + 175, + 261, + 252 + ], + "spans": [ + { + "bbox": [ + 186, + 175, + 261, + 252 + ], + "type": "image", + "image_path": "213367d8f1ff555c0fc4f75d1e3512092ba8160ed76b9d3897bbe6c2066926c4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 193, + 258, + 257, + 269 + ], + "lines": [ + { + "bbox": [ + 193, + 258, + 257, + 269 + ], + "spans": [ + { + "bbox": [ + 193, + 258, + 257, + 269 + ], + "type": "text", + "content": "(d) Traj sensor" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 358, + 287, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 358, + 287, + 501 + ], + "spans": [ + { + "bbox": [ + 46, + 358, + 287, + 501 + ], + "type": "text", + "content": "every set of two objects, we can infer their relationships in three types from their bounding boxes: horizontal proximity, vertical proximity, and support as mentioned in Sec. 3; for every set of three objects, we can infer if one of them is in between of the other two objects. As detection results do not contain object poses, we do not construct allocentric relations (e.g., \"the shelf that is behind the sofa\"). Then, we build a scene graph, where each object is assigned as a node in the graph, and edges between nodes represent the relationships between objects. By converting 3D scenes into text, we can apply ChatGPT to extract meaningful insights from the data." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 508, + 286, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 508, + 286, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 508, + 286, + 675 + ], + "type": "text", + "content": "Leveraging ChatGPT to localize the target object. A simple approach is to directly input the entire scene graph into ChatGPT and ask it to select the target objects. However, scenes may contain many objects, resulting in extremely long textural descriptions. We observe that ChatGPT is often confused in such settings and fails to respond with the right answer. To narrow the search space, we first employ ChatGPT to recognize objects that correlate to the provided text. We then exclude the unrelated objects from the scene graph, focusing solely on those with pertinent information, enabling us to pinpoint the target object effectively. This approach has the advantage of reducing the number of objects that need to be considered, making it easier for ChatGPT to identify the target object." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": "As shown in Fig. 3, we construct two prompts in sequence. Take \"sit on the chair that is in the middle of the board and the end table\" as an example, we first need to nar" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 545, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 288 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 288 + ], + "type": "text", + "content": "row down the object category search space. In order to find which type of objects we care about, we construct the first prompt that asks ChatGPT to find out target objects and anchor objects. Target object is defined as the final object that we want the agent to interact with, which in this case, is the \"table\". Anchor objects are the objects that help with determining the target object, for there might be many chairs in one scene. Based on the target object \"chair\" and anchor objects \"board\" and \"end table\", we can filter out all the unrelated objects in the scene, only keeping chairs, the board, and the end table in our scene graph. Next, according to the simplified scene graph, we can describe object relations in text: every edge in the scene graph could be converted to an edge sentence like \"chair 4 is far from the end table 0\". Converting all edges to such edge sentences gives a full description of the current scene. Finally, we construct a second prompt by asking ChatGPT to infer the target object from the accumulated edge sentences." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 293, + 504, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 293, + 504, + 306 + ], + "spans": [ + { + "bbox": [ + 306, + 293, + 504, + 306 + ], + "type": "text", + "content": "4.2. Diffusion-based trajectory generation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 312, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 312, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 312, + 545, + 407 + ], + "type": "text", + "content": "Given the localized object from ChatGPT, we first generate the trajectory based on the instructions and then synthesize local human poses. Trajectory is defined as a sequence of characters' translations and orientations. As suggested by [51], not every point of the scene is relevant to the final human motions. Inspired by NSM [70] and ManipNet [92], we employ volumetric sensors (as shown in Fig. 4) around the target object to represent the scene." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 411, + 545, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 545, + 483 + ], + "type": "text", + "content": "Object-centric scene representation. Denote the target object center location as " + }, + { + "bbox": [ + 304, + 411, + 545, + 483 + ], + "type": "inline_equation", + "content": "c_{o} = (c_{x},c_{y},c_{z})" + }, + { + "bbox": [ + 304, + 411, + 545, + 483 + ], + "type": "text", + "content": ". We transform the point cloud of the scene " + }, + { + "bbox": [ + 304, + 411, + 545, + 483 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 411, + 545, + 483 + ], + "type": "text", + "content": " to an object-centered coordinate axis centered at " + }, + { + "bbox": [ + 304, + 411, + 545, + 483 + ], + "type": "inline_equation", + "content": "c_{o}" + }, + { + "bbox": [ + 304, + 411, + 545, + 483 + ], + "type": "text", + "content": ". To recognize the surrounding geometry of the target object, we create volumetric sensors called Environment Sensor and Target Sensor." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 487, + 545, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 487, + 545, + 571 + ], + "spans": [ + { + "bbox": [ + 304, + 487, + 545, + 571 + ], + "type": "text", + "content": "Environment sensor. The environment sensor is centered at " + }, + { + "bbox": [ + 304, + 487, + 545, + 571 + ], + "type": "inline_equation", + "content": "c_{o}" + }, + { + "bbox": [ + 304, + 487, + 545, + 571 + ], + "type": "text", + "content": " in a cubic shape with a volume of " + }, + { + "bbox": [ + 304, + 487, + 545, + 571 + ], + "type": "inline_equation", + "content": "4 \\times 4 \\times 4 \\, m^{3}" + }, + { + "bbox": [ + 304, + 487, + 545, + 571 + ], + "type": "text", + "content": ", containing " + }, + { + "bbox": [ + 304, + 487, + 545, + 571 + ], + "type": "inline_equation", + "content": "8 \\times 8 \\times 8" + }, + { + "bbox": [ + 304, + 487, + 545, + 571 + ], + "type": "text", + "content": " cubic voxels as shown in Fig. 4 (c). It is constructed by collecting all occupancies " + }, + { + "bbox": [ + 304, + 487, + 545, + 571 + ], + "type": "inline_equation", + "content": "o_{s}" + }, + { + "bbox": [ + 304, + 487, + 545, + 571 + ], + "type": "text", + "content": ", center positions " + }, + { + "bbox": [ + 304, + 487, + 545, + 571 + ], + "type": "inline_equation", + "content": "c_{v}" + }, + { + "bbox": [ + 304, + 487, + 545, + 571 + ], + "type": "text", + "content": ", and normal directions " + }, + { + "bbox": [ + 304, + 487, + 545, + 571 + ], + "type": "inline_equation", + "content": "n_{v}" + }, + { + "bbox": [ + 304, + 487, + 545, + 571 + ], + "type": "text", + "content": " of each voxel to form a feature vector " + }, + { + "bbox": [ + 304, + 487, + 545, + 571 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 487, + 545, + 571 + ], + "type": "text", + "content": ". Like [70], the scene occupancy " + }, + { + "bbox": [ + 304, + 487, + 545, + 571 + ], + "type": "inline_equation", + "content": "o_{s} \\in R^{1}" + }, + { + "bbox": [ + 304, + 487, + 545, + 571 + ], + "type": "text", + "content": " in each voxel is defined based on the given scene mesh:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 367, + 577, + 545, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 577, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 367, + 577, + 545, + 622 + ], + "type": "interline_equation", + "content": "o _ {s} = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} d _ {s} < 0, \\\\ 0 & \\text {i f} d _ {s} > a _ {s}, \\\\ 1 - \\frac {d _ {s}}{a _ {s}} & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {3}", + "image_path": "0d350024fd575905423059dd63bee63956ac6018b06bf9a4e233edaebbdcff68.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 625, + 545, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 625, + 545, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 625, + 545, + 686 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 625, + 545, + 686 + ], + "type": "inline_equation", + "content": "d_{s}" + }, + { + "bbox": [ + 304, + 625, + 545, + 686 + ], + "type": "text", + "content": " is the signed distance between the scene mesh and the voxel center, and " + }, + { + "bbox": [ + 304, + 625, + 545, + 686 + ], + "type": "inline_equation", + "content": "a_{s}" + }, + { + "bbox": [ + 304, + 625, + 545, + 686 + ], + "type": "text", + "content": " is the voxel edge length. " + }, + { + "bbox": [ + 304, + 625, + 545, + 686 + ], + "type": "inline_equation", + "content": "n_v" + }, + { + "bbox": [ + 304, + 625, + 545, + 686 + ], + "type": "text", + "content": " is the normal direction of the closest scene point to the voxel center. The environment sensor provides coarse scene geometry around the target object." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Target sensor. To capture the detailed geometry of the target object, we further build a target sensor. As we already" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1859" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "text", + "content": "obtain the 3D bounding box of the target object in Sec. 4.1, we crop the point clouds according to the bounding box and construct an " + }, + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "inline_equation", + "content": "8 \\times 8 \\times 8" + }, + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "text", + "content": " cubic volumes that cover the bounding box, as shown in Fig. 4 (b). Target sensor " + }, + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "text", + "content": " is in the same form as the environment sensor " + }, + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "text", + "content": ", except the target sensor has a different voxel size, as visualized in Fig. 4 (b) and (c)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 144, + 286, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 144, + 286, + 227 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 286, + 227 + ], + "type": "text", + "content": "Given the constructed object-centric representations, we follow [57] to employ a transformer decoder architecture [79], which enables arbitrary length motions. As for the text input, we use CLIP [62] text encoder to encode input text to the text feature " + }, + { + "bbox": [ + 46, + 144, + 286, + 227 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 144, + 286, + 227 + ], + "type": "text", + "content": ". The time-step " + }, + { + "bbox": [ + 46, + 144, + 286, + 227 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 144, + 286, + 227 + ], + "type": "text", + "content": " is injected into the decoder in sinusoidal position embeddings form [79]. In summary, the condition for this generation model is" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 130, + 236, + 286, + 249 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 286, + 249 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 286, + 249 + ], + "type": "interline_equation", + "content": "\\mathbf {C} _ {\\mathbf {t}} = \\{L, E, T \\}, \\tag {4}", + "image_path": "ee1f8e313d50632dcbc8d6dc1842e30db80a414641d91f983b999f9c8301b0cb.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 258, + 286, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 286, + 329 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 286, + 329 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 258, + 286, + 329 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 258, + 286, + 329 + ], + "type": "text", + "content": " is the text feature, " + }, + { + "bbox": [ + 46, + 258, + 286, + 329 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 46, + 258, + 286, + 329 + ], + "type": "text", + "content": " is the environment sensor, and " + }, + { + "bbox": [ + 46, + 258, + 286, + 329 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 258, + 286, + 329 + ], + "type": "text", + "content": " is the target sensor. All the conditions are projected to the same dimension " + }, + { + "bbox": [ + 46, + 258, + 286, + 329 + ], + "type": "inline_equation", + "content": "D = 512" + }, + { + "bbox": [ + 46, + 258, + 286, + 329 + ], + "type": "text", + "content": " by MLPs and summed with positional embeddings to form tokens. We use the simple objective described in Eq. 2 to train the trajectory generation model " + }, + { + "bbox": [ + 46, + 258, + 286, + 329 + ], + "type": "inline_equation", + "content": "G_{r}" + }, + { + "bbox": [ + 46, + 258, + 286, + 329 + ], + "type": "text", + "content": " to generate the trajectory " + }, + { + "bbox": [ + 46, + 258, + 286, + 329 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{1:N}" + }, + { + "bbox": [ + 46, + 258, + 286, + 329 + ], + "type": "text", + "content": " with length " + }, + { + "bbox": [ + 46, + 258, + 286, + 329 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 258, + 286, + 329 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 336, + 233, + 349 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 336, + 233, + 349 + ], + "spans": [ + { + "bbox": [ + 47, + 336, + 233, + 349 + ], + "type": "text", + "content": "4.3. Diffusion-based motion completion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 355, + 286, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 355, + 286, + 401 + ], + "spans": [ + { + "bbox": [ + 46, + 355, + 286, + 401 + ], + "type": "text", + "content": "Given the trajectory from Sec. 4.2, the next step is to complete the whole motion. Based on the generated trajectories, we construct a Trajectory Sensor " + }, + { + "bbox": [ + 46, + 355, + 286, + 401 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 46, + 355, + 286, + 401 + ], + "type": "text", + "content": " to explicitly reason about the interaction between the character and scenes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 407, + 286, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 407, + 286, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 286, + 514 + ], + "type": "text", + "content": "Trajectory sensor. The trajectory sensor is also a volumetric sensor which is similar to the form of the environment sensor but is designed for ego-centric perception [95] as shown in Fig. 4 (d). Specifically, trajectory sensors are positioned around characters in each frame. This sensor " + }, + { + "bbox": [ + 46, + 407, + 286, + 514 + ], + "type": "inline_equation", + "content": "O_{i}" + }, + { + "bbox": [ + 46, + 407, + 286, + 514 + ], + "type": "text", + "content": " is centered at the predicted root position of the " + }, + { + "bbox": [ + 46, + 407, + 286, + 514 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 407, + 286, + 514 + ], + "type": "text", + "content": "-th frame and faces to the " + }, + { + "bbox": [ + 46, + 407, + 286, + 514 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 407, + 286, + 514 + ], + "type": "text", + "content": "-th frame's predicted root orientation, containing " + }, + { + "bbox": [ + 46, + 407, + 286, + 514 + ], + "type": "inline_equation", + "content": "8 \\times 8 \\times 8" + }, + { + "bbox": [ + 46, + 407, + 286, + 514 + ], + "type": "text", + "content": " cubic voxels that store scene occupancy. The occupancy calculation is the same as Eq. 3." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 514, + 286, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 514, + 286, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 514, + 286, + 548 + ], + "type": "text", + "content": "Another transformer-based conditional diffusion model is used to synthesize local poses along trajectories. The condition is defined as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 104, + 559, + 286, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 559, + 286, + 571 + ], + "spans": [ + { + "bbox": [ + 104, + 559, + 286, + 571 + ], + "type": "interline_equation", + "content": "\\mathbf {C} _ {\\mathbf {m}} = \\{L, E, T, O _ {1}, \\dots , O _ {N} \\}, \\tag {5}", + "image_path": "b62e9e75ab1f607f0cd69799367ba7c6519c6b6c6f08d044ee30d9af4b335461.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 579, + 286, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 579, + 286, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 579, + 286, + 628 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 579, + 286, + 628 + ], + "type": "inline_equation", + "content": "L, E," + }, + { + "bbox": [ + 46, + 579, + 286, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 579, + 286, + 628 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 579, + 286, + 628 + ], + "type": "text", + "content": " have the same meanings in Eq. 4. Simple objective described in Eq. 2 is used to train the motion generation model " + }, + { + "bbox": [ + 46, + 579, + 286, + 628 + ], + "type": "inline_equation", + "content": "G_{m}" + }, + { + "bbox": [ + 46, + 579, + 286, + 628 + ], + "type": "text", + "content": " to generate global orientation " + }, + { + "bbox": [ + 46, + 579, + 286, + 628 + ], + "type": "inline_equation", + "content": "\\gamma_{1:N}" + }, + { + "bbox": [ + 46, + 579, + 286, + 628 + ], + "type": "text", + "content": " and local poses " + }, + { + "bbox": [ + 46, + 579, + 286, + 628 + ], + "type": "inline_equation", + "content": "\\theta_{1:N}" + }, + { + "bbox": [ + 46, + 579, + 286, + 628 + ], + "type": "text", + "content": " with length " + }, + { + "bbox": [ + 46, + 579, + 286, + 628 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 579, + 286, + 628 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 635, + 178, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 178, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 178, + 647 + ], + "type": "text", + "content": "4.4. Implementation details" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "text", + "content": "Following [38, 57, 80, 81], we use separate diffusion models for the trajectory generation and the motion generation. Because the HUMANISE dataset [83] contains a relatively small number (51 minutes) of pure motion samples from the AMASS dataset (3772 minutes) [49], we first pre-train the" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": "models on the whole AMASS dataset for 200 epochs and then fine-tune on the HUMANISE dataset for 200 epochs to improve the motion quality. During pretraining, the text feature " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": " is set to all zeros. Both models are trained with the AdamW optimizer [40], using a learning rate of 0.0001 on a single Nvidia RTX 3090 GPU. The batch size is set to 128. The version of ChatGPT is gpt-3.5-turbo. More details can be found in the supplementary material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 180, + 386, + 193 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 180, + 386, + 193 + ], + "spans": [ + { + "bbox": [ + 306, + 180, + 386, + 193 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 200, + 416, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 200, + 416, + 211 + ], + "spans": [ + { + "bbox": [ + 306, + 200, + 416, + 211 + ], + "type": "text", + "content": "5.1. Evaluation metrics" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 219, + 544, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 219, + 544, + 243 + ], + "spans": [ + { + "bbox": [ + 304, + 219, + 544, + 243 + ], + "type": "text", + "content": "We evaluate the generated motions in three aspects: scene-conditional, action-conditional, and pure motion quality." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 247, + 545, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 247, + 545, + 366 + ], + "spans": [ + { + "bbox": [ + 304, + 247, + 545, + 366 + ], + "type": "text", + "content": "Scene-conditional motion quality. To evaluate how the generated motion is aligned with the scene, we calculate the body-to-goal distance (goal dist.) [83] to measure how accurately the character interacts with target objects. However, goal dist. does not consider the consistency of the entire motion and scene (e.g., sitting on the sofa with incorrect orientation). To compensate for goal dist., a human perceptual study (indicated by scene score) is performed by randomly sampling 20 scenarios for each model, where ten workers are required to score each sample." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 371, + 545, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 371, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 304, + 371, + 545, + 441 + ], + "type": "text", + "content": "Action-conditional motion quality. To measure how the generated motion is aligned with the text, we follow [17] to evaluate action recognition accuracy (accuracy), diversity, and multimodality of the results. Calculating these metrics relies on a pre-trained action recognition model [87] and we train the recognition model on the HUMANISE dataset." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 447, + 545, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 447, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 447, + 545, + 518 + ], + "type": "text", + "content": "Pure motion quality. We evaluate the realism of generated motions using the metric suggested by [17], namely Frechet Inception Distance (FID). A lower FID indicates that generated motions are closer to the groundtruth motions. We also perform a human perceptual study (indicated by quality score) to measure pure motion quality." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 527, + 545, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 527, + 545, + 539 + ], + "spans": [ + { + "bbox": [ + 306, + 527, + 545, + 539 + ], + "type": "text", + "content": "5.2. Generating human motion from text and scene" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": "Our experiments are conducted on the HUMANISE dataset [83]. Following the previous setting [83], there are " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "16.5k" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": " motion sequences in 543 scenes for training and " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "3.1k" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": " motion sequences in 100 scenes for testing. We compare our method with four baselines: (1) " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{MDM}^*" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": " [75]: a diffusion-based motion generator. (2) " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{GMD}^*" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": " [38]: GMD proposes various techniques for enhancing the control and quality of MDM. We employ the single-stage setting of " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{GMD}^*" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": ". Similar to HUMANISE, we use a point-transformer to provide scene features for " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{MDM}^*" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{GMD}^*" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": ". (3) " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{GMD}^{\\mathbf{HC}}" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": ": we provide object centers predicted by HUMANISE (denoted by HC) to guide the motion generation process in GMD by adding a proximity loss to encourage the motion to be close to the predicted object center. The proximity loss is defined" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1860" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 87, + 70, + 509, + 175 + ], + "blocks": [ + { + "bbox": [ + 87, + 70, + 509, + 175 + ], + "lines": [ + { + "bbox": [ + 87, + 70, + 509, + 175 + ], + "spans": [ + { + "bbox": [ + 87, + 70, + 509, + 175 + ], + "type": "table", + "html": "
MethodsScene-conditionalAction-conditionalPure motion quality
goal dist.↓scene score↑accuracy↑diversity→multimodality→FID↓quality score↑
Real0.014-99.1%4.822.280.00-
MDM*2.0481.9996.0%4.832.570.162.45
GMD*1.2292.5096.6%4.912.280.242.64
GMDHC1.1302.4996.4%4.932.430.232.76
HUMANISE0.9952.3389.7%4.252.661.112.81
Ours0.3843.5497.1%4.782.310.123.50
", + "image_path": "73248fd43b5552d911fa93068a4ac3022915e6ec820bca60160882b34b132b98.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 182, + 547, + 228 + ], + "lines": [ + { + "bbox": [ + 46, + 182, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 547, + 228 + ], + "type": "text", + "content": "Table 1. Quantitative results on the HUMANISE dataset. We compare our method with four baselines (please refer to Sec. 5.2) and the real data. " + }, + { + "bbox": [ + 46, + 182, + 547, + 228 + ], + "type": "inline_equation", + "content": "\\mathrm{GMD}^{\\mathrm{HC}}" + }, + { + "bbox": [ + 46, + 182, + 547, + 228 + ], + "type": "text", + "content": " means using HUMANISE's predicted center to guide motion generation in " + }, + { + "bbox": [ + 46, + 182, + 547, + 228 + ], + "type": "inline_equation", + "content": "\\mathrm{GMD^{*}}" + }, + { + "bbox": [ + 46, + 182, + 547, + 228 + ], + "type": "text", + "content": ". Among the metrics, scene score and quality score are perceptual studies. " + }, + { + "bbox": [ + 46, + 182, + 547, + 228 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 46, + 182, + 547, + 228 + ], + "type": "text", + "content": " means higher is better and " + }, + { + "bbox": [ + 46, + 182, + 547, + 228 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 46, + 182, + 547, + 228 + ], + "type": "text", + "content": " means lower is better. " + }, + { + "bbox": [ + 46, + 182, + 547, + 228 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 182, + 547, + 228 + ], + "type": "text", + "content": " means closer to the real data is better. Bold indicates the best results.underline indicates the second best." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 50, + 237, + 545, + 389 + ], + "blocks": [ + { + "bbox": [ + 50, + 237, + 545, + 389 + ], + "lines": [ + { + "bbox": [ + 50, + 237, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 50, + 237, + 545, + 389 + ], + "type": "image", + "image_path": "24ed49104be155df0172295846b1e03d3e6fe6329a55c141d2c34fae0ea45a7e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 397, + 546, + 420 + ], + "lines": [ + { + "bbox": [ + 46, + 397, + 546, + 420 + ], + "spans": [ + { + "bbox": [ + 46, + 397, + 546, + 420 + ], + "type": "text", + "content": "Figure 5. Qualitative results. We compare our method with groundtruth and four baselines (please refer to Sec. 5.2) given the same text descriptions. Our method synthesizes motions that interact with the object precisely as the groundtruth data while the baselines fail." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 440, + 288, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 440, + 288, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 440, + 288, + 475 + ], + "type": "text", + "content": "as the distance from HC to the predicted human pelvis at the interacting frame. (4) HUMANISE: we directly use their released models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 476, + 287, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 476, + 287, + 573 + ], + "spans": [ + { + "bbox": [ + 46, + 476, + 287, + 573 + ], + "type": "text", + "content": "The quantitative results are shown in Tab. 1. Our method outperforms the baseline in terms of goal dist., scene score, accuracy, FID, and quality score and achieves competitive results in diversity and multimodality. The quantitative results show that our approach can generate better-quality motions and are more consistent with all the conditions. The qualitative results are demonstrated in Fig. 5. Please refer to the supplementary material for more visualizations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 582, + 139, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 582, + 139, + 594 + ], + "spans": [ + { + "bbox": [ + 47, + 582, + 139, + 594 + ], + "type": "text", + "content": "5.3. Ablation study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "Ablation of main components. Four variants are constructed to explore the effect of our design choice. (1) \"w/o localization\": the localization module is removed and trajectories and motions are generated directly using scene point clouds. Scene features are given as the same form of MDM* and GMD*. (2) \"w/o object-centric\": we remove our object-centric representation and predict motions in the scene coordinate. The scene features of our sensors are still employed. (3) \"w/o two-stage\": trajectories and" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 440, + 546, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 440, + 546, + 560 + ], + "spans": [ + { + "bbox": [ + 304, + 440, + 546, + 560 + ], + "type": "text", + "content": "motions are generated together. (4) \"w/o diffusion\": we employ cVAE instead of the diffusion model as the architecture. (5) \"w/o pretrain\": the models are not pre-trained on the AMASS dataset. As shown in Tab. 2, the localization enhances the precise interaction ability and the object-centric representation improves motion quality. Although the (\"w/o two-stage\") is more efficient and has competitive performance metrics comparing to our method, it is prone to collapse directly into the target position, lacking gradual transition and realism." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": "The design choice of object localization. For object localization, we compare our method with two baselines and five variants. (1) HUMANISE: predicts the object center in the auxiliary task. (2) BUTD-DETR [34]: is a 3D visual grounding method. (3) “ours w/o two-stage”: we employ a one-stage question-answering process. (4) “ours w/o few-shot”: the few-shot examples are not provided. (5) “ours using text matching”: we compute CLIP similarities instead of using ChatGPT. (6) “ours using LLaMA”: ChatGPT is replaced by LLaMA 2 7B model. (7) “ours using Mistral”: ChatGPT is replaced by Mistral 7B [35], an open-source LLM. The prompts for LLaMA 2 and Mistral" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "text", + "content": "1861" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 286, + 167 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 286, + 167 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 286, + 167 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 286, + 167 + ], + "type": "table", + "html": "
VariantsSceneActionPure
goal dist.↓accuracy↑diversity→mm→FID↓
Real0.01499.1%4.822.280.00
w/o localization0.59274.5%3.712.753.14
w/o object-centric0.41390.6%4.132.631.01
w/o two-stage0.38597.2%4.932.330.14
w/o diffusion0.39040.0%3.693.863.61
w/o pretrain0.39282.6%3.882.372.50
Ours0.38497.1%4.782.310.12
", + "image_path": "f12a4e427bdfb63e92e30a435b669bb51ab543226d4a2883839550df5783b546.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 241, + 287, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 241, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 241, + 287, + 384 + ], + "type": "text", + "content": "7B are slightly adjusted. We evaluate these variants under two scenarios: predicted detection [47] and groundtruth detection. The metrics include the accuracy (acc.) and the distance from predicted centers to the object center (center dist.). \"acc.\" is the percentage of times with an IoU (Intersection over Union) higher than the threshold (0.25) following [34]. The results are shown in Tab. 3. Since the results of Mistral are only slightly behind of ChatGPT, ChatGPT can be replaced by Mistral for better reproducibility, but cannot be replaced by the text matching method. Moreover, if groundtruth detection is provided, our method can achieve even better results." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 389, + 287, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 389, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 47, + 389, + 287, + 414 + ], + "type": "text", + "content": "The design choice of sensor density. Please refer to the supplementary material." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 422, + 139, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 422, + 139, + 434 + ], + "spans": [ + { + "bbox": [ + 47, + 422, + 139, + 434 + ], + "type": "text", + "content": "5.4. Generalization" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 441, + 287, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 441, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 46, + 441, + 287, + 525 + ], + "type": "text", + "content": "To validate the generalization ability of our method, we run our method directly on the unseen PROX dataset [22] without fine-tuning. We provide illustrations in Fig. 6. With our localization method based on ChatGPT and motion generation method based on volumetric sensors, our pipeline can easily generalize to other datasets. Please refer to the supplementary material for more results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 537, + 117, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 537, + 117, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 537, + 117, + 550 + ], + "type": "text", + "content": "6. Discussion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "content": "We have demonstrated that our approach could synthesize motions with better quality and more precise interactions. However, restricted by the dataset, the duration of motions is relatively short (60% are around 1-3s), and most texts follow the template [1] without detailed descriptions. We only tackle static scenes and extending our method to interact with moving objects is a future direction. We acknowledge that leveraging LLMs has several problems. ChatGPT may fail when the detector misses the objects and the behavior of ChatGPT is regulated by prompts [37]. Despite we have shown in Tab. 3 that ChatGPT can be replaced by an open source LLM Mistral 7B, using LLMs instead of classical parsing approaches is less efficient in the inference stage." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 309, + 70, + 545, + 183 + ], + "blocks": [ + { + "bbox": [ + 47, + 175, + 288, + 218 + ], + "lines": [ + { + "bbox": [ + 47, + 175, + 288, + 218 + ], + "spans": [ + { + "bbox": [ + 47, + 175, + 288, + 218 + ], + "type": "text", + "content": "Table 2. Ablation of main components. We compare our method with five variants (please refer to Sec. 5.3). Among them, " + }, + { + "bbox": [ + 47, + 175, + 288, + 218 + ], + "type": "inline_equation", + "content": "mm" + }, + { + "bbox": [ + 47, + 175, + 288, + 218 + ], + "type": "text", + "content": " indicates multimodality. Bold indicates the best results. Underline indicates the second best." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 70, + 545, + 183 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 545, + 183 + ], + "type": "table", + "html": "
VariantsPredicted detectionGT detection
acc. ↑center dist.↓acc. ↑center dist.↓
HUMANISE-1.48--
BUTD-DETR62.51.3363.11.23
Ours w/o two-stage49.01.3458.51.10
Ours w/o few-shot69.30.8086.30.45
Ours using text matching67.00.8979.70.52
Ours using LLaMA 240.31.5542.51.52
Ours using Mistral 7B72.40.7487.20.33
Ours using ChatGPT-3.575.60.6190.50.37
", + "image_path": "b609ce663ef4d30c04f4a988daf3c7807791a05064d63499bd468c7a400dd92d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 191, + 547, + 257 + ], + "lines": [ + { + "bbox": [ + 304, + 191, + 547, + 257 + ], + "spans": [ + { + "bbox": [ + 304, + 191, + 547, + 257 + ], + "type": "text", + "content": "Table 3. Design choices of object localization. We compare our method with two baselines and five variants (please refer to Sec. 5.3). Since HUMANISE directly regresses the coordinate of centers without utilizing groundtruth detection, we do not calculate acc. and only include their results under predicted detection. Bold indicates the best results." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 319, + 277, + 425, + 362 + ], + "blocks": [ + { + "bbox": [ + 319, + 277, + 425, + 362 + ], + "lines": [ + { + "bbox": [ + 319, + 277, + 425, + 362 + ], + "spans": [ + { + "bbox": [ + 319, + 277, + 425, + 362 + ], + "type": "image", + "image_path": "9cb3d200c0b4f617df9ff313c8771453e220e0b467a3816bf5906c45bfb74fb4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 426, + 277, + 533, + 361 + ], + "blocks": [ + { + "bbox": [ + 426, + 277, + 533, + 361 + ], + "lines": [ + { + "bbox": [ + 426, + 277, + 533, + 361 + ], + "spans": [ + { + "bbox": [ + 426, + 277, + 533, + 361 + ], + "type": "image", + "image_path": "4e37dd5c6339ff292e5995d045529d6ddf1c2f70df48c51cffe23d69863af90e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 318, + 363, + 425, + 450 + ], + "blocks": [ + { + "bbox": [ + 318, + 363, + 425, + 450 + ], + "lines": [ + { + "bbox": [ + 318, + 363, + 425, + 450 + ], + "spans": [ + { + "bbox": [ + 318, + 363, + 425, + 450 + ], + "type": "image", + "image_path": "3ad9ea70bd24639ce0169e49bf9d1b94fda76fffa27572b243d415862b60859b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 458, + 545, + 503 + ], + "lines": [ + { + "bbox": [ + 304, + 458, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 304, + 458, + 545, + 503 + ], + "type": "text", + "content": "Figure 6. Qualitative results of our method on the PROX dataset. We run our method on the scenes from the PROX dataset without fine-tuning. Results show that our method is capable to generalize to unseen scenes and objects." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 426, + 363, + 533, + 449 + ], + "blocks": [ + { + "bbox": [ + 426, + 363, + 533, + 449 + ], + "lines": [ + { + "bbox": [ + 426, + 363, + 533, + 449 + ], + "spans": [ + { + "bbox": [ + 426, + 363, + 533, + 449 + ], + "type": "image", + "image_path": "0bb381f1c3adbf42f2abb0023b608d75552d245318cc5c6225472174caf46fea.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 530, + 378, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 530, + 378, + 542 + ], + "spans": [ + { + "bbox": [ + 306, + 530, + 378, + 542 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 554, + 545, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 554, + 545, + 650 + ], + "spans": [ + { + "bbox": [ + 304, + 554, + 545, + 650 + ], + "type": "text", + "content": "In this work, we introduce a novel method for generating motion from text in a scene. To tackle this problem, we propose a two-step approach. The first step involves 3D visual grounding, where we identify the target object. In the second step, concentrating on the target object, we build a diffusion-based motion generation method. Our approach offers several advantages, including improved 3D visual grounding accuracy and motion quality." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 658, + 404, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 404, + 672 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 404, + 672 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": "This work was partially supported by the NSFC (No. 62172364), Ant Group and Information Technology Center and State Key Lab of CAD&CG, Zhejiang University." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1862" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 91, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 58, + 91, + 286, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 91, + 286, + 133 + ], + "spans": [ + { + "bbox": [ + 58, + 91, + 286, + 133 + ], + "type": "text", + "content": "[1] Panos Achlioptas, Ahmed Abdelreheem, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In ECCV, 2020. 2, 3, 4, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 137, + 286, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 137, + 286, + 168 + ], + "spans": [ + { + "bbox": [ + 58, + 137, + 286, + 168 + ], + "type": "text", + "content": "[2] Chaitanya Ahuja and Louis-Philippe Morency. Language2pose: Natural language grounded pose forecasting. In 3DV, 2019. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 171, + 286, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 171, + 286, + 203 + ], + "spans": [ + { + "bbox": [ + 58, + 171, + 286, + 203 + ], + "type": "text", + "content": "[3] Nikos Athanasiou, Mathis Petrovich, Michael J. Black, and Gül Varol. TEACH: Temporal Action Compositions for 3D Humans. In 3DV, 2022. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 205, + 286, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 205, + 286, + 237 + ], + "spans": [ + { + "bbox": [ + 58, + 205, + 286, + 237 + ], + "type": "text", + "content": "[4] Nikos Athanasiou, Mathis Petrovich, Michael J. Black, and Gül Varol. SINC: Spatial composition of 3D human motions for simultaneous action generation. ICCV, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 240, + 286, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 240, + 286, + 272 + ], + "spans": [ + { + "bbox": [ + 58, + 240, + 286, + 272 + ], + "type": "text", + "content": "[5] German Barquero, Sergio Escalera, and Cristina Palmero. Belfusion: Latent diffusion for behavior-driven human motion prediction. In ICCV, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 274, + 286, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 274, + 286, + 382 + ], + "spans": [ + { + "bbox": [ + 58, + 274, + 286, + 382 + ], + "type": "text", + "content": "[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In NeurIPS, 2020. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 385, + 286, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 385, + 286, + 427 + ], + "spans": [ + { + "bbox": [ + 58, + 385, + 286, + 427 + ], + "type": "text", + "content": "[7] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 58, + 430, + 286, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 430, + 286, + 462 + ], + "spans": [ + { + "bbox": [ + 58, + 430, + 286, + 462 + ], + "type": "text", + "content": "[8] Yu-Wei Chao, Jimei Yang, Weifeng Chen, and Jia Deng. Learning to sit: Synthesizing human-chair interactions via hierarchical control. In AAAI, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 58, + 464, + 286, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 464, + 286, + 497 + ], + "spans": [ + { + "bbox": [ + 58, + 464, + 286, + 497 + ], + "type": "text", + "content": "[9] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In ECCV, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 499, + 286, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 499, + 286, + 553 + ], + "spans": [ + { + "bbox": [ + 53, + 499, + 286, + 553 + ], + "type": "text", + "content": "[10] Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, et al. Evaluating large language models trained on code. arXiv preprint, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 555, + 286, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 555, + 286, + 598 + ], + "spans": [ + { + "bbox": [ + 53, + 555, + 286, + 598 + ], + "type": "text", + "content": "[11] Yixin Chen, Qing Li, Deqian Kong, Yik Lun Kei, Song-Chun Zhu, Tao Gao, Yixin Zhu, and Siyuan Huang. Yourfit: Embodied reference understanding with language and gesture. In ICCV, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 601, + 286, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 601, + 286, + 643 + ], + "spans": [ + { + "bbox": [ + 53, + 601, + 286, + 643 + ], + "type": "text", + "content": "[12] Rishabh Dabral, Muhammad Hamza Mughal, Vladislav Golyanik, and Christian Theobalt. Mofusion: A framework for denoising-diffusion-based motion synthesis. In CVPR, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 647, + 286, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 647, + 286, + 678 + ], + "spans": [ + { + "bbox": [ + 53, + 647, + 286, + 678 + ], + "type": "text", + "content": "[13] Abhishek Das, Georgia Gkioxari, Stefan Lee, Devi Parikh, and Dhruv Batra. Neural Modular Control for Embodied Question Answering. In CoRL, 2018. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 681, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 681, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 681, + 286, + 712 + ], + "type": "text", + "content": "[14] Katerina Fragkiadaki, Sergey Levine, Panna Felsen, and Jitendra Malik. Recurrent network models for human dynamics. In ICCV, 2015. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 312, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 545, + 106 + ], + "type": "text", + "content": "[15] Anindita Ghosh, Noshaba Cheema, Cennet Oguz, Christian Theobalt, and Philipp Slusallek. Synthesis of compositional animations from textual descriptions. In ICCV, 2021. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 107, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 107, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 312, + 107, + 545, + 149 + ], + "type": "text", + "content": "[16] Anindita Ghosh, Rishabh Dabral, Vladislav Golyanik, Christian Theobalt, and Philipp Slusallek. Imos: Intent-driven full-body motion synthesis for human-object interactions. In Computer Graphics Forum, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 151, + 545, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 151, + 545, + 193 + ], + "spans": [ + { + "bbox": [ + 312, + 151, + 545, + 193 + ], + "type": "text", + "content": "[17] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3d human motions. In ACM MM, 2020. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 195, + 545, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 195, + 545, + 226 + ], + "spans": [ + { + "bbox": [ + 312, + 195, + 545, + 226 + ], + "type": "text", + "content": "[18] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In CVPR, 2022. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 228, + 545, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 228, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 312, + 228, + 545, + 270 + ], + "type": "text", + "content": "[19] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In ECCV, 2022. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 272, + 545, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 272, + 545, + 314 + ], + "spans": [ + { + "bbox": [ + 312, + 272, + 545, + 314 + ], + "type": "text", + "content": "[20] Zoey Guo, Yiwen Tang, Ray Zhang, Dong Wang, Zhigang Wang, Bin Zhao, and Xuelong Li. Viewrefer: Grasp the multi-view knowledge for 3d visual grounding. In ICCV, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 316, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 316, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 312, + 316, + 545, + 348 + ], + "type": "text", + "content": "[21] Félix G Harvey, Mike Yurick, Derek Nowrouzezahrai, and Christopher Pal. Robust motion in-betweening. ACM Trans. Graph., 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 350, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 350, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 312, + 350, + 545, + 380 + ], + "type": "text", + "content": "[22] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J. Black. Resolving 3d human pose ambiguities with 3d scene constraints. In ICCV, 2019. 2, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 383, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 383, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 312, + 383, + 545, + 415 + ], + "type": "text", + "content": "[23] Mohamed Hassan, Duygu Ceylan, Ruben Villegas, Jun Saito, Jimei Yang, Yi Zhou, and Michael J. Black. Stochastic scene-aware motion prediction. In ICCV, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 416, + 545, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 416, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 312, + 416, + 545, + 447 + ], + "type": "text", + "content": "[24] Mohamed Hassan, Yunrong Guo, Tingwu Wang, Michael Black, Sanja Fidler, and Xue Bin Peng. Synthesizing physical character-scene interactions. In SIGGRAPH, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 449, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 449, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 312, + 449, + 545, + 480 + ], + "type": "text", + "content": "[25] Chengan He, Jun Saito, James Zachary, Holly Rushmeier, and Yi Zhou. Nemf: Neural motion fields for kinematic animation. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 482, + 545, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 482, + 545, + 523 + ], + "spans": [ + { + "bbox": [ + 312, + 482, + 545, + 523 + ], + "type": "text", + "content": "[26] Gustav Eje Henter, Simon Alexanderson, and Jonas Beskow. Moglow: Probabilistic and controllable motion synthesis using normalising flows. ACM Trans. Graph., 2020. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 526, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 526, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 312, + 526, + 545, + 547 + ], + "type": "text", + "content": "[27] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 548, + 545, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 548, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 312, + 548, + 545, + 580 + ], + "type": "text", + "content": "[28] Daniel Holden, Jun Saito, and Taku Komura. A deep learning framework for character motion synthesis and editing. ACM Trans. Graph., 2016. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 582, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 582, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 312, + 582, + 545, + 613 + ], + "type": "text", + "content": "[29] Daniel Holden, Taku Komura, and Jun Saito. Phase-functional neural networks for character control. ACM Trans. Graph., 2017. 1, 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 312, + 615, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 615, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 312, + 615, + 545, + 645 + ], + "type": "text", + "content": "[30] Joy Hsu, Jiayuan Mao, and Jiajun Wu. Ns3d: Neurosymbolic grounding of 3d objects and relations. In CVPR, 2023. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 312, + 647, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 647, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 312, + 647, + 545, + 679 + ], + "type": "text", + "content": "[31] Pin-Hao Huang, Han-Hung Lee, Hwann-Tzong Chen, and Tyng-Luh Liu. Text-guided graph neural networks for referring 3d instance segmentation. AAAI, 2021. 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 312, + 681, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 681, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 312, + 681, + 545, + 712 + ], + "type": "text", + "content": "[32] Shijia Huang, Yilun Chen, Jiaya Jia, and Liwei Wang. Multi-view transformer for 3d visual grounding. In CVPR, 2022. 3" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1863" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 73, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 73, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 73, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 53, + 73, + 287, + 116 + ], + "type": "text", + "content": "[33] Siyuan Huang, Zan Wang, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, and Song-Chun Zhu. Diffusion-based generation, optimization, and planning in 3d scenes. In CVPR, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 118, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 118, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 53, + 118, + 287, + 160 + ], + "type": "text", + "content": "[34] Ayush Jain, Nikolaos Gkanatsios, Ishita Mediratta, and Katerina Fragkiadaki. Bottom up top down detection transformers for language grounding in images and point clouds. In ECCV, 2022. 3, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 162, + 288, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 162, + 288, + 215 + ], + "spans": [ + { + "bbox": [ + 53, + 162, + 288, + 215 + ], + "type": "text", + "content": "[35] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 217, + 288, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 217, + 288, + 250 + ], + "spans": [ + { + "bbox": [ + 53, + 217, + 288, + 250 + ], + "type": "text", + "content": "[36] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. arXiv preprint arXiv:2306.14795, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 251, + 288, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 251, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 251, + 288, + 293 + ], + "type": "text", + "content": "[37] Sai Shashank Kalakonda, Shubh Maheshwari, and Ravi Kiran Sarvadevabhatla. Action-gpt: Leveraging large-scale language models for improved and generalized action generation. In ICME, 2023. 2, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 294, + 288, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 294, + 288, + 336 + ], + "spans": [ + { + "bbox": [ + 53, + 294, + 288, + 336 + ], + "type": "text", + "content": "[38] Korrawe Karunratanakul, Konpat Preechakul, Supasorn Suwajanakorn, and Siyu Tang. Guided motion diffusion for controllable human motion synthesis. In ICCV, 2023. 2, 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 338, + 288, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 338, + 288, + 370 + ], + "spans": [ + { + "bbox": [ + 53, + 338, + 288, + 370 + ], + "type": "text", + "content": "[39] Jihoon Kim, Jiseob Kim, and Sungjoon Choi. Flame: Freeform language-based motion synthesis & editing. In AAAI, 2023. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 371, + 288, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 371, + 288, + 393 + ], + "spans": [ + { + "bbox": [ + 53, + 371, + 288, + 393 + ], + "type": "text", + "content": "[40] Diederik P. Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization. arXiv e-prints, 2014. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 394, + 288, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 394, + 288, + 415 + ], + "spans": [ + { + "bbox": [ + 53, + 394, + 288, + 415 + ], + "type": "text", + "content": "[41] Diederik P Kingma and Max Welling. Auto-Encoding Variational Bayes. arXiv e-prints, 2013. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 416, + 288, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 416, + 288, + 448 + ], + "spans": [ + { + "bbox": [ + 53, + 416, + 288, + 448 + ], + "type": "text", + "content": "[42] Takeshi Kojima, Shixiang (Shane) Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners. In NeurIPS, 2022. 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 449, + 288, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 449, + 288, + 481 + ], + "spans": [ + { + "bbox": [ + 53, + 449, + 288, + 481 + ], + "type": "text", + "content": "[43] Hanyang Kong, Kehong Gong, Dongze Lian, Michael Bi Mi, and Xinchao Wang. Priority-centric human motion generation in discrete latent space. In ICCV, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 482, + 288, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 482, + 288, + 525 + ], + "spans": [ + { + "bbox": [ + 53, + 482, + 288, + 525 + ], + "type": "text", + "content": "[44] Peizhuo Li, Kfir Aberman, Zihan Zhang, Rana Hanocka, and Olga Sorkine-Hornung. Ganimator: Neural motion synthesis from a single sequence. ACM Trans. Graph., 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 526, + 288, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 526, + 288, + 568 + ], + "spans": [ + { + "bbox": [ + 53, + 526, + 288, + 568 + ], + "type": "text", + "content": "[45] Quanzhou Li, Jingbo Wang, Chen Change Loy, and Bo Dai. Task-oriented human-object interactions generation with implicit neural representations. arXiv preprint, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 570, + 288, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 570, + 288, + 602 + ], + "spans": [ + { + "bbox": [ + 53, + 570, + 288, + 602 + ], + "type": "text", + "content": "[46] Hung Yu Ling, Fabio Zinno, George Cheng, and Michiel Van De Panne. Character controllers using motion vaes. ACM Trans. Graph., 2020. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 604, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 604, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 53, + 604, + 288, + 635 + ], + "type": "text", + "content": "[47] Ze Liu, Zheng Zhang, Yue Cao, Han Hu, and Xin Tong. Group-free 3d object detection via transformers. In ICCV, 2021. 3, 4, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 53, + 637, + 288, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 637, + 288, + 679 + ], + "spans": [ + { + "bbox": [ + 53, + 637, + 288, + 679 + ], + "type": "text", + "content": "[48] Junyu Luo, Jiahui Fu, Xianghao Kong, Chen Gao, Haibing Ren, Hao Shen, Huaxia Xia, and Si Liu. 3d-sps: Single-stage 3d visual grounding via referred point progressive selection. In CVPR, 2022. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 53, + 681, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 681, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 681, + 288, + 713 + ], + "type": "text", + "content": "[49] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In ICCV, 2019. 2, 6" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 312, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 545, + 106 + ], + "type": "text", + "content": "[50] Julieta Martinez, Michael J. Black, and Javier Romero. On human motion prediction using recurrent neural networks. In CVPR, 2017. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 107, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 107, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 312, + 107, + 545, + 140 + ], + "type": "text", + "content": "[51] Aymen Mir, Xavier Puig, Angjoo Kanazawa, and Gerard Pons-Moll. Generating continual human motion in diverse 3d scenes. arXiv preprint, 2023. 2, 3, 4, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 141, + 545, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 141, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 312, + 141, + 545, + 163 + ], + "type": "text", + "content": "[52] OpenAI. Openai: Introducing chatgpt. https://openai.com/blog/chatgpt, 2022. 2, 3, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 164, + 489, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 164, + 489, + 175 + ], + "spans": [ + { + "bbox": [ + 312, + 164, + 489, + 175 + ], + "type": "text", + "content": "[53] OpenAI. Gpt-4 technical report, 2023. 2, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 176, + 545, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 176, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 312, + 176, + 545, + 219 + ], + "type": "text", + "content": "[54] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3d hands, face, and body from a single image. In CVPR, 2019. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 220, + 545, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 220, + 545, + 253 + ], + "spans": [ + { + "bbox": [ + 312, + 220, + 545, + 253 + ], + "type": "text", + "content": "[55] Mathis Petrovich, Michael J. Black, and Gül Varol. Action-conditioned 3D human motion synthesis with transformer VAE. In ICCV, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 255, + 545, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 255, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 312, + 255, + 545, + 287 + ], + "type": "text", + "content": "[56] Mathis Petrovich, Michael J. Black, and Gül Varol. TEMOS: Generating diverse human motions from textual descriptions. In ECCV, 2022. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 289, + 545, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 289, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 312, + 289, + 545, + 331 + ], + "type": "text", + "content": "[57] Huaijin Pi, Sida Peng, Minghui Yang, Xiaowei Zhou, and Hujun Bao. Hierarchical generation of human-object interactions with diffusion probabilistic models. In ICCV, 2023. 2, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 333, + 545, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 333, + 545, + 355 + ], + "spans": [ + { + "bbox": [ + 312, + 333, + 545, + 355 + ], + "type": "text", + "content": "[58] Matthias Plappert, Christian Mandery, and Tamim Asfour. The kit motion-language dataset. *Big data*, 2016. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 357, + 545, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 357, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 312, + 357, + 545, + 399 + ], + "type": "text", + "content": "[59] Abhinanda R. Punnakkal, Arjun Chandrasekaran, Nikos Athanasiou, Alejandra Quiros-Ramirez, and Michael J. Black. Babel: Bodies, action and behavior with english labels. In CVPR, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 400, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 400, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 312, + 400, + 545, + 443 + ], + "type": "text", + "content": "[60] Yijun Qian, Jack Urbanek, Alexander G Hauptmann, and Jungdam Won. Breaking the limits of text-conditioned 3d motion synthesis with elaborative descriptions. In ICCV, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 445, + 545, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 445, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 312, + 445, + 545, + 488 + ], + "type": "text", + "content": "[61] Sigal Raab, Inbal Leibovitch, Peizhuo Li, Kfir Aberman, Olga Sorkine-Hornung, and Daniel Cohen-Or. Modi: Unconditional motion synthesis from diverse data. In CVPR, 2023. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 491, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 491, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 312, + 491, + 545, + 544 + ], + "type": "text", + "content": "[62] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 312, + 546, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 546, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 312, + 546, + 545, + 601 + ], + "type": "text", + "content": "[63] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485-5551, 2020. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 312, + 602, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 602, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 312, + 602, + 545, + 644 + ], + "type": "text", + "content": "[64] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical Text-Conditional Image Generation with CLIP Latents. arXiv e-prints, 2022. 4" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 312, + 647, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 647, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 312, + 647, + 545, + 679 + ], + "type": "text", + "content": "[65] Ali Razavi, Aaron van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. In NeurIPS, 2019. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 312, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 681, + 545, + 713 + ], + "type": "text", + "content": "[66] Junha Roh, Karthik Desingh, Ali Farhadi, and Dieter Fox. Languagerefer: Spatial-language model for 3d visual grounding. In CoRL, 2022. 3" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1864" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 73, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 73, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 73, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 53, + 73, + 287, + 106 + ], + "type": "text", + "content": "[67] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 107, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 107, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 53, + 107, + 287, + 140 + ], + "type": "text", + "content": "[68] Yonatan Shafir, Guy Tevet, Roy Kapon, and Amit H Bermano. Human motion diffusion as a generative prior. arXiv preprint, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 141, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 141, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 53, + 141, + 287, + 174 + ], + "type": "text", + "content": "[69] Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. In NeurIPS, 2015. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 175, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 175, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 53, + 175, + 287, + 207 + ], + "type": "text", + "content": "[70] Sebastian Starke, He Zhang, Taku Komura, and Jun Saito. Neural state machine for character-scene interactions. ACM Trans. Graph., 2019. 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 209, + 287, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 209, + 287, + 241 + ], + "spans": [ + { + "bbox": [ + 53, + 209, + 287, + 241 + ], + "type": "text", + "content": "[71] Sebastian Starke, Yiwei Zhao, Taku Komura, and Kazi Zaman. Local motion phases for learning multi-contact character movements. ACM Trans. Graph., 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 243, + 287, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 243, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 53, + 243, + 287, + 275 + ], + "type": "text", + "content": "[72] Sebastian Starke, Ian Mason, and Taku Komura. Deepphase: Periodic autoencoders for learning motion phase manifolds. ACM Trans. Graph., 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 277, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 277, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 53, + 277, + 287, + 308 + ], + "type": "text", + "content": "[73] Omid Taheri, Vasileios Choutas, Michael J. Black, and Dimitrios Tzionas. Goal: Generating 4d whole-body motion for hand-object grasping. In CVPR, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 310, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 310, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 53, + 310, + 287, + 342 + ], + "type": "text", + "content": "[74] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. In ECCV, 2022. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 344, + 287, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 344, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 53, + 344, + 287, + 376 + ], + "type": "text", + "content": "[75] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In ICLR, 2023. 1, 2, 4, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 378, + 287, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 378, + 287, + 410 + ], + "spans": [ + { + "bbox": [ + 53, + 378, + 287, + 410 + ], + "type": "text", + "content": "[76] Jesse Thomason, Mohit Shridhar, Yonatan Bisk, Chris Paxton, and Luke Zettlemoyer. Language grounding with 3d objects. In CoRL, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 412, + 287, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 412, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 53, + 412, + 287, + 477 + ], + "type": "text", + "content": "[77] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models. arXiv preprint, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 479, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 479, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 53, + 479, + 287, + 510 + ], + "type": "text", + "content": "[78] Aaron van den Oord, Oriol Vinyals, and koray kavukcuoglu. Neural discrete representation learning. In NeurIPS, 2017. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 512, + 287, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 512, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 53, + 512, + 287, + 554 + ], + "type": "text", + "content": "[79] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 557, + 287, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 557, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 53, + 557, + 287, + 589 + ], + "type": "text", + "content": "[80] Jiashun Wang, Huazhe Xu, Jingwei Xu, Sifei Liu, and Xiaolong Wang. Synthesizing long-term 3d human motion and interaction in 3d scenes. In CVPR, 2021. 2, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 591, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 591, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 53, + 591, + 287, + 623 + ], + "type": "text", + "content": "[81] Jingbo Wang, Yu Rong, Jingyuan Liu, Sijie Yan, Dahua Lin, and Bo Dai. Towards diverse and natural scene-aware 3d human motion synthesis. In CVPR, 2022. 2, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 53, + 624, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 624, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 53, + 624, + 287, + 668 + ], + "type": "text", + "content": "[82] Xi Wang, Gen Li, Yen-Ling Kuo, Muhammed Kocabas, Emre Aksan, and Otmar Hilliges. Reconstructing action-conditioned human-object interactions using commonsense knowledge priors. In 3DV, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 53, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 670, + 287, + 712 + ], + "type": "text", + "content": "[83] Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, Wei Liang, and Siyuan Huang. Humanise: Language-conditioned human motion generation in 3d scenes. In NeurIPS, 2022. 2, 3, 6" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 312, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 545, + 106 + ], + "type": "text", + "content": "[84] Yan Wu, Jiahao Wang, Yan Zhang, Siwei Zhang, Otmar Hilliges, Fisher Yu, and Siyu Tang. Saga: Stochastic whole-body grasping with contact. In ECCV, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 107, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 107, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 312, + 107, + 545, + 139 + ], + "type": "text", + "content": "[85] Yiming Xie, Varun Jampani, Lei Zhong, Deqing Sun, and Huaizu Jiang. Omnicontrol: Control any joint at any time for human motion generation. arXiv preprint, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 140, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 140, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 312, + 140, + 545, + 172 + ], + "type": "text", + "content": "[86] Sirui Xu, Zhengyuan Li, Yu-Xiong Wang, and Liang-Yan Gui. Interdiff: Generating 3d human-object interactions with physics-informed diffusion. In ICCV, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 174, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 174, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 312, + 174, + 545, + 205 + ], + "type": "text", + "content": "[87] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In AAAI, 2018. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 206, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 206, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 312, + 206, + 545, + 239 + ], + "type": "text", + "content": "[88] Zhengyuan Yang, Songyang Zhang, Liwei Wang, and Jiebo Luo. Sat: 2d semantics assisted training for 3d visual grounding. In ICCV, pages 1856-1866, 2021. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 240, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 240, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 312, + 240, + 545, + 271 + ], + "type": "text", + "content": "[89] Shuquan Ye, Dongdong Chen, Songfang Han, and Jing Liao. 3d question answering. IEEE Transactions on Visualization and Computer Graphics, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 273, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 273, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 312, + 273, + 545, + 327 + ], + "type": "text", + "content": "[90] Zhihao Yuan, Xu Yan, Yinghong Liao, Ruimao Zhang, Sheng Wang, Zhen Li, and Shuguang Cui. Instancerefer: Cooperative holistic understanding for visual grounding on point clouds through instance multi-level contextual referring. In ICCV, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 328, + 545, + 359 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 328, + 545, + 359 + ], + "spans": [ + { + "bbox": [ + 312, + 328, + 545, + 359 + ], + "type": "text", + "content": "[91] He Zhang, Sebastian Starke, Taku Komura, and Jun Saito. Mode-adaptive neural networks for quadruped motion control. ACM Trans. Graph., 2018. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 361, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 361, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 312, + 361, + 545, + 403 + ], + "type": "text", + "content": "[92] He Zhang, Yuting Ye, Takaaki Shiratori, and Taku Komura. Manipnet: Neural manipulation synthesis with a hand-object spatial representation. ACM Trans. Graph., 2021. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 404, + 545, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 404, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 312, + 404, + 545, + 448 + ], + "type": "text", + "content": "[93] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Yong Zhang, Hongwei Zhao, Hongtao Lu, Xi Shen, and Ying Shan. Generating human motion from textual descriptions with discrete representations. In CVPR, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 449, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 449, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 312, + 449, + 545, + 491 + ], + "type": "text", + "content": "[94] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 493, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 493, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 312, + 493, + 545, + 525 + ], + "type": "text", + "content": "[95] Xiaohan Zhang, Bharat Lal Bhatnagar, Sebastian Starke, Vladimir Guzov, and Gerard Pons-Moll. Couch: Towards controllable human-chair interactions. In ECCV, 2022. 2, 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 526, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 526, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 312, + 526, + 545, + 547 + ], + "type": "text", + "content": "[96] Yan Zhang and Siyu Tang. The wanderings of odysseus in 3d scenes. In CVPR, 2022. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 312, + 548, + 545, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 548, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 312, + 548, + 545, + 580 + ], + "type": "text", + "content": "[97] Yiming Zhang, ZeMing Gong, and Angel X Chang. Multi3drefer: Grounding text description to multiple 3d objects. In ICCV, 2023. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 312, + 582, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 582, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 312, + 582, + 545, + 624 + ], + "type": "text", + "content": "[98] Yaqi Zhang, Di Huang, Bin Liu, Shixiang Tang, Yan Lu, Lu Chen, Lei Bai, Qi Chu, Nenghai Yu, and Wanli Ouyang. Motiongpt: Finetuned llms are general-purpose motion generators. arXiv preprint, 2023. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 312, + 625, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 625, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 312, + 625, + 545, + 646 + ], + "type": "text", + "content": "[99] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In ICCV, 2021. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 647, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 679 + ], + "type": "text", + "content": "[100] Kaifeng Zhao, Shaofei Wang, Yan Zhang, Thabo Beeler, and Siyu Tang. Compositional human-scene interaction synthesis with semantic control. In ECCV, 2022. 3" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "text", + "content": "[101] Kaifeng Zhao, Yan Zhang, Shaofei Wang, Thabo Beeler, and Siyu Tang. Synthesizing diverse human motions in 3d indoor scenes. In ICCV, 2023. 3" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1865" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 185 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[102] Lichen Zhao, Daigang Cai, Lu Sheng, and Dong Xu. 3dvg-transformer: Relation modeling for visual grounding on point clouds. In ICCV, 2021. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "text", + "content": "[103] Yang Zheng, Yanchao Yang, Kaichun Mo, Jiaman Li, Tao Yu, Yebin Liu, C Karen Liu, and Leonidas J Guibas. Gimo: Gaze-informed human motion prediction in context. In ECCV, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 185 + ], + "type": "text", + "content": "[104] Yi Zhou, Connelly Barnes, Lu Jingwan, Yang Jimei, and Li Hao. On the continuity of rotation representations in neural networks. In CVPR, 2019. 3" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1866" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Generating Illustrated Instructions/dd71c86a-390f-4c46-b6ae-0a13f2787fa0_content_list.json b/2024/Generating Illustrated Instructions/dd71c86a-390f-4c46-b6ae-0a13f2787fa0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..37e5b3cce056a3ca2c9c72fd0b3dddbe00ad2de2 --- /dev/null +++ b/2024/Generating Illustrated Instructions/dd71c86a-390f-4c46-b6ae-0a13f2787fa0_content_list.json @@ -0,0 +1,2162 @@ +[ + { + "type": "text", + "text": "Generating Illustrated Instructions", + "text_level": 1, + "bbox": [ + 305, + 130, + 663, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sachit Menon $^{1,2*}$", + "bbox": [ + 264, + 179, + 403, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ishan Misra", + "bbox": [ + 437, + 180, + 540, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rohit Girdhar1", + "bbox": [ + 583, + 180, + 700, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1GenAI, Meta", + "bbox": [ + 320, + 198, + 434, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2Columbia University", + "bbox": [ + 473, + 199, + 648, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://facebookresearch.github.io/IllustratedInstructions", + "bbox": [ + 251, + 219, + 710, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce a new task of generating \"Illustrated Instructions\", i.e. visual instructions customized to a user's needs. We identify desiderata unique to this task, and formalize it through a suite of automatic and human evaluation metrics, designed to measure the validity, consistency, and efficacy of the generations. We combine the power of large language models (LLMs) together with strong text-to-image generation diffusion models to propose a simple approach called StackedDiffusion, which generates such illustrated instructions given text as input. The resulting model strongly outperforms baseline approaches and state-of-the-art multimodal LLMs; and in $30\\%$ of cases, users even prefer it to human-generated articles. Most notably, it enables various new and exciting applications far beyond what static articles on the web can provide, such as personalized instructions complete with intermediate steps and pictures in response to a user's individual situation.", + "bbox": [ + 75, + 301, + 473, + 559 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 589, + 209, + 604 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The internet is a vast resource to find answers to all types of questions. It is often easy to find a webpage or a video that walks through the exact steps to achieve a user's goal. With the rise of Large Language Models (LLMs) trained on internet-scale data, users can just ask the LLM for instructions to achieve a goal. This allows the users to get answers for specific personalized queries for which there may not be an existing webpage on the internet, e.g., modifying cooking recipes with user specific dietary restrictions. Moreover, if the users make a mistake when following the instructions, simple follow-up questions to the LLM can generate alternate instructions, a major advantage over static web search.", + "bbox": [ + 75, + 616, + 468, + 797 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In spite of such advantages, LLMs still have one major limitation - they cannot generate visuals, which are critical for users to learn from and follow instructions for a wide range of tasks [6]. Consider, for instance, instructions that require visual inspection, e.g., a recipe that requires", + "bbox": [ + 75, + 797, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4fe67dc0ae5b1b060518dded246c2b77f59842388aaf6c09d967ab5aad81875d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 271, + 534, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "How can I bake a cake to celebrate an event in Seattle?", + "text_level": 1, + "bbox": [ + 542, + 277, + 834, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "One great way could be to make a coffee-infused cake to honor Seattle's coffee culture, with teal coloring for the Seattle flag.", + "bbox": [ + 511, + 297, + 844, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Step 1: Brew strong coffee; let cool.", + "bbox": [ + 511, + 327, + 612, + 347 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Step 2: Preheat the oven to $350\\mathrm{F}$", + "bbox": [ + 638, + 327, + 740, + 345 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Step 3: Prepare cake batter as usual.", + "bbox": [ + 767, + 325, + 875, + 345 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/852682748d3765eb82a702ea3eb732d06dc543bcbfb0a3a2c45dc54dc4fdd451.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 508, + 349, + 629, + 441 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Step 4: Mix in the coffee; fold in coffee grounds for texture.", + "bbox": [ + 511, + 444, + 614, + 474 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e19a4df7e34c8a6fff064a96695f89adb558a1d0115a4a06909e77c1e98a91c2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 637, + 349, + 756, + 441 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Step 5: Bake for 25- 30 minutes.", + "bbox": [ + 638, + 444, + 741, + 464 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/369dc6ae7426ff6b589954507b0a21efeed223528b8ae9e9a5880f6e5ba09956.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 349, + 885, + 441 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Step 6: Cover the cake in teal fondant for the Seattle flag.", + "bbox": [ + 766, + 444, + 867, + 474 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/03ec5dc17b70cd300b497e23cf15e210c38f3eacc6c5e2ca0a6809e7789ec75c.jpg", + "image_caption": [ + "Figure 1. StackedDiffusion generating Illustrated Instructions. Given a goal (or any textual user input), StackedDiffusion produces a customized instructional article complete with illustrations that not only tells the user how to achieve the goal in words, but also shows the user by providing illustrations." + ], + "image_footnote": [], + "bbox": [ + 506, + 474, + 629, + 569 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/987a18f0208641696c5ad4ea810c0b7bf3299c2c539484dd7750f5e35378d311.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 637, + 476, + 756, + 569 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5821d0db1c60bb551af3ca97b0d29079b4ef19842d877374e0d623776f483002.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 476, + 885, + 569 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the user to stir fry onions until golden brown, or searching for bubbles when a flat tire is submerged under water. An image accompanying such instructions can make following them significantly easier. Can we develop methods with the strengths of LLMs that can also generate such visuals?", + "bbox": [ + 496, + 671, + 890, + 746 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we tackle this challenge, developing models that can not only tell a user how to accomplish their task, but also show them how. We define the novel task of Illustrated Instructions: creating a set of steps with visualizations that showcase an approach to solve the user's task. We carefully consider the different dimensions of the problem, laying out three desiderata unique to this setting. To measure these desiderata, we develop automated metrics based on prior work in instructional article assessment [67] and image generation [50].", + "bbox": [ + 496, + 750, + 892, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work done during an internship at Meta.", + "bbox": [ + 94, + 886, + 316, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6274", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose a new model to solve the task of Illustrated Instructions that combines LLMs with a text-to-image diffusion model. We train our model on stacks of instructional images from websites such as WikiHow. The resulting StackedDiffusion model creates full instructional articles, complete with customized steps and a sequence of images uniquely generated to describe those steps. It leverages large-scale pretrained LLMs and finetuned text-to-image diffusion models, and employs techniques to accomplish the task without the need to introduce any new learnable parameters. This includes spatial tiling for simultaneous multi-image generation, text embedding concatenation to reduce information loss in long instructions, and a new \"step-positional encoding\" to better demarcate the different steps in the instructions. Thus, StackedDiffusion can generate useful illustrations even when trained even with limited instructional data (cf. Figure 7).", + "bbox": [ + 76, + 90, + 472, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We compare StackedDiffusion with various baseline approaches based on off-the-shelf tools, and find that they all fall short, even when used in conjunction with one another. Existing T2I models are incapable of generating visuals directly from a user query. Even when given more detailed instructional text, we show that existing T2I models fail to produce images that are simultaneously faithful to the goal, the step, and consistent among each other. Given the recent introduction of multimodal LLMs [1, 28], we also compare with a recent open-source model, GILL [27], and show such models also fall short of consistently generating useful visuals along with text. We posit that our approach of leveraging the spatial priors learned by pretrained diffusion models to generate multiple images together, in conjunction with pretrained LLMs, is significantly more compute and data efficient than an approach purely based on next-token prediction, without these priors. Our thorough ablations and human evaluations show that StackedDiffusion convincingly surpasses state-of-the-art models. Our final model even outperforms human-generated articles in $30\\%$ of cases, showing the strong potential of our approach.", + "bbox": [ + 76, + 352, + 472, + 671 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contributions: 1) We introduce the novel task of Illustrated Instructions, which requires generating a sequence of images and text that together describe how to achieve a goal ( $\\S 3$ and $\\S\\S 5.1$ and 5.2), along with desiderata and metrics for this task; 2) We propose a new approach StackedDiffusion for Illustrated Instructions, with novel modifications to the modeling procedure, enabling generation of visuals suitable for instructional articles for the first time without any additional parameters ( $\\S 4$ ); 3) We show that our proposed method achieves strong performance on all metrics, and confirm that human evaluators prefer it over existing methods by wide margins—even surpassing ground truth images in some cases ( $\\S\\S 5.3$ to 5.5); 4) Finally, we showcase new abilities that StackedDiffusion unlocks, including personalization, goal suggestion, and error correction, that go", + "bbox": [ + 76, + 674, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "far beyond what is possible with fixed articles (§ 5.6).", + "bbox": [ + 498, + 90, + 854, + 107 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 121, + 640, + 137 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Instructional data, tasks, and methods. In the text domain, learning language models on WikiHow [30, 72], has led to advances in tasks such as summarization [71], commonsense procedural knowledge [72, 75], question answering [11], and hierarchical reasoning [74]. In particular, Zhang et al. [72] introduce the goal inference task, in which a model is presented a goal text as input and asked which of 4 candidate steps is one that actually helps achieve that goal, as well as the analogous step inference task. In the multimodal setting, Yang et al. [67] introduce the Visual Goal-Step Inference (VGSI) dataset and task, which consider articles of interleaved text and images. In this task, a model is again presented a goal text as input but is asked which of 4 candidate images is one that actually helps achieve that goal. They show representations learned on this data aid in tasks related to instructional videos [39, 58].", + "bbox": [ + 496, + 147, + 893, + 388 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Learning representations from multimodal data. Multimodal data (including multimodal instructional data, such as from video [39, 58]) has proven to be a powerful source of signal for tasks such as zero-shot recognition [25, 47], text-image and text-video retrieval [3, 9, 16, 17, 34, 36, 37, 40, 60], temporal segmentation [36, 46, 56, 65, 76], activity localization [9, 33, 36, 63, 65, 77], anticipation [12, 15, 19, 52, 61], question-answering [31, 53, 64], summarization [41, 42], and even recipe personalization [14].", + "bbox": [ + 496, + 390, + 893, + 526 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing work on instructional data has centered around understanding, rather than generation. We instead focus on the novel setting of generating full multimodal articles complete with text and illustrations.", + "bbox": [ + 496, + 526, + 893, + 585 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generative models. Recent work has examined text-conditioned visual generation through autoregressive [8, 69] or diffusion models [4, 18, 21, 43, 48, 49, 51]. These advances have been leveraged to create text and images together with purely autoregressive [1, 70] or combined approaches [27]. Concurrent work [2, 57] enables the capability of generating multiple images together in the autoregressive framework, but requires substantial additional parameters and focus on creating images that adhere solely to the nearby text rather than enabling consistency. (Similar issues arise in the text-to-video setting; see Appendix 7.) StackedDiffusion, on the other hand, leverages the priors built into the T2I model to obtain consistency without any additional parameters.", + "bbox": [ + 496, + 588, + 895, + 800 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Illustrated Instructions", + "text_level": 1, + "bbox": [ + 500, + 814, + 718, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We now formalize our task and the corresponding desiderata. The input to the system is a goal text $\\mathbf{g}$ . As output, we would like to produce step text $\\mathbf{s}_i$ as well as step illustration $\\mathbf{I}_i$ for each step $i$ . The step text should be a natural language", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6275", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8bef10bc0808b264215a1b39b4debe4e84955a541d43247c8694557b6a9202d2.jpg", + "image_caption": [ + "Figure 2. Failure modes of a naive approach. A frozen T2I model is not able to capture both the goal and the step, showing only one or the other depending on how it is prompted. Further, it can not produce consistent images, leading to odd changes such as the color of the ice varying between images." + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 467, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "description of the step, while the step illustrations should be an image corresponding to the step.", + "bbox": [ + 75, + 414, + 467, + 444 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "On first glance, one might think that this task is a straightforward amalgamation of textual instruction generation and text to image generation. While these tasks are closely related and necessary substeps towards illustrated instructions, they are not sufficient. The naive approach—simply creating step text $\\mathbf{s}_i$ from goal text $\\mathbf{g}$ , and each image $\\mathbf{I}_i$ individually from each $\\mathbf{s}_i$ —fails to recognize the fresh new challenges that do not exist in either of these two other tasks. We identify three key requirements for useful illustrated illustrations: goal faithfulness, step faithfulness, and cross-image consistency.", + "bbox": [ + 75, + 445, + 467, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The first requirement is clear: if the images do not relate to the goal, they cannot be good illustrations. This motivates the desiderata of goal faithfulness, which requires that each image faithfully reflects the goal text.", + "bbox": [ + 75, + 612, + 467, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, goal faithfulness alone is not enough. Consider the first row of images in Figure 2. The images all reflect the ultimate goal—creating colored ice—but fail in their role as step illustrations. This in turn motivates step faithfulness, which requires that each image be faithful to the step text. We see in this example that baseline T2I models fail dramatically along this metric; every image reflects the goal rather than the specifics of the step requested.", + "bbox": [ + 75, + 672, + 467, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Finally, the generated images should be consistent with each other. While the second row of images in Figure 2 are all faithful to the step text, the color of the ice (and even the style of the images, cartoon or real) changes between images. This is jarring and confusing to the reader. This motivates the final criterion of cross-image consistency, which requires that each image be consistent with the other images", + "bbox": [ + 75, + 795, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/70e6cc6bb7533343b713e5d6e18f8c0cff12d414b7c8c2d9fb6f0164468e4763.jpg", + "image_caption": [ + "Figure 3. Overview of StackedDiffusion. At training time, we use the given goal and step text, and stack the encoded ground truth step-images. At inference time, we obtain the goal and step text from an LLM, and unstack denoised latents to produce the output images. See §4 for details and notation." + ], + "image_footnote": [], + "bbox": [ + 504, + 94, + 887, + 250 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "produced for a particular generation.", + "bbox": [ + 500, + 345, + 741, + 361 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. StackedDiffusion", + "text_level": 1, + "bbox": [ + 500, + 375, + 666, + 388 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose a new architecture, StackedDiffusion, to overcome the limitations of existing text-to-image (T2I) approaches for the task of generating interleaved text and images for instructional tasks. StackedDiffusion builds upon T2I models based on latent diffusion models (LDMs) [49], which are diffusion models [48] that operate on a low-resolution 'latent' encoding of the original images. Our primary desiderata is that the images generated must be faithful to both the goal and step text. However, in initial experiments we found that simply encoding the goal and step texts joined together (as strings) leads to an uninformative encoding. For instance, generations conditioned on this combined text tended to ignore certain steps or the goal. Furthermore, the length of this combined text will likely exceed the context length limitations of the text encoders [47] commonly used in T2I models, leading to undesired truncation and information loss.", + "bbox": [ + 496, + 398, + 890, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Hence, we elect to use a more general approach, applicable to text encoders with any context length. Rather than encoding the combined goal and step text to obtain the condition, we first separately encode the goal and step texts and then concatenate the encodings. This allows the model to learn to use the goal and step text independently, as well as in combination. We add a 'step-positional encoding' $\\Phi_{i}$ to this concatenation. It is broadcast across the dimensions pertaining to a particular step, to indicate to the model where a step begins and ends. $\\Phi_0$ denotes the positional encoding reserved to indicate the goal embedding. Hence, given a text encoder $\\psi$ , and $N$ steps in a given goal, we compute the overall text conditioning $\\mathcal{C}$ as", + "bbox": [ + 496, + 656, + 890, + 852 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} = \\left| \\right| \\left(\\psi (\\mathbf {g}) + \\Phi_ {0}, \\left. \\right| _ {i = 1} ^ {N} \\psi (\\mathbf {s} _ {i}) + \\Phi_ {i}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 862, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6276", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $||$ is the concatenation operation. This is shown in Figure 3 (top). This design decision is critical to obtain good goal and step faithfulness.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition, independently generating each image does not achieve the requirement of cross-image consistency as the model cannot exchange information across images. Thus, we generate all the $\\mathbf{I}_i$ images at once, which allows the model to jointly generate the sequence and achieve cross-image consistency. The remaining question, of course, is how to accomplish this simultaneous generation such that it gives rise to cross-image consistency. Our key observation is that T2I models already have a strong prior for consistency within a single image. Could we make use of this previously-learned knowledge?", + "bbox": [ + 75, + 136, + 467, + 301 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose a simple method to accomplish this: spatial tiling. As illustrated in Figure 3, the denoising U-Net is given latents $\\mathbf{z}_i$ corresponding to each of the output images simultaneously, tiled spatially as if a single image. At training time, training images are encoded into the latent space as usual, then reshaped into the tiled format. In detail,", + "bbox": [ + 75, + 301, + 467, + 393 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {i} = \\mathcal {E} \\left(\\mathbf {I} _ {i}\\right), \\quad \\mathbf {z} = \\left. \\right| _ {i = 1} ^ {N} \\mathbf {z} _ {i} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 401, + 467, + 441 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {L D M} := \\mathbb {E} _ {\\mathbf {z}, \\epsilon \\sim \\mathcal {N} (0, 1), t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {z} ^ {(t)}, t, \\mathcal {C}) \\| _ {2} ^ {2} \\right] \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 445, + 467, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{E}$ is the encoder to map the image to the latent space (for instance, using a VAE [26]), $L_{LDM}$ is the training objective for the LDM, $t$ denotes a timestep of the diffusion process, $\\epsilon$ denotes the noise added at a given timestep, and $\\theta$ denotes the parameters of the learned denoising U-Net. At inference, we use classifier-free guidance [43] to generate a latent $\\tilde{\\mathbf{z}}$ given conditioning goal and $N$ step texts. We split the latent into $\\tilde{\\mathbf{z}}_1,\\dots,\\tilde{\\mathbf{z}}_N$ , and decode into generated images using a decoder $\\mathcal{D}$ corresponding to the encoder $\\mathcal{E}$ , i.e. $\\tilde{\\mathbf{I}}_i = \\mathcal{D}(\\tilde{\\mathbf{z}}_i)$ . (We opt for tiling along a single spatial dimension for simplicity, and find that alternative tiling strategies do not significantly affect performance; see Appendix § 8.) We refer the reader to the Appendix § 9 and [49] for further details on LDMs. Given the stacked conditioning and generation operations, we refer to our final model as Stacked-Diffusion.", + "bbox": [ + 75, + 478, + 467, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During training, goal and step text are obtained from ground truth, while at inference, a pretrained LLM is used to transform arbitrary user input text to an inferred goal and generated step texts. We describe the LLM inference procedure and prompt engineering in the Appendix § 10. We initialize the U-Net using a pretrained T2I model, and fine-tune all layers when training with the stacked input. Our stacked conditioning only increases the spatial resolution of the input to the U-Net for which can be modeled entirely using the existing parameters (spatial convolutions, attention) of the U-Net. We use the T2I model's text encoder $(\\psi)$ , kept frozen, to encode the goal and step texts.", + "bbox": [ + 75, + 719, + 467, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0c73f9e5b449715bef0bcf29490b2d4f9cd7aff6a1231c003decf96a1d2c66f7.jpg", + "image_caption": [ + "Figure 4. Illustrated Instructions data. The histogram shows the distribution of step counts in the data. We find that more than $80\\%$ of articles consist of 6 or fewer steps." + ], + "image_footnote": [], + "bbox": [ + 511, + 87, + 880, + 184 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since the spatially stacked latent has a spatial resolution comparable to the ones used for high resolution image generation, we encounter issues observed by prior work [32] for high resolution training. Specifically, [32] finds that high resolution latents retain substantial information about the input being noised with typical noise addition schedules, even at the final diffusion timestep $t = T$ . This results in an unintentional distribution shift between train time (when all observed inputs contain signal) and test time (when the first inputs are pure noise with no signal). We address this by adjusting the training diffusion noise schedule such that the signal-to-noise ratio (SNR) at the final diffusion timestep $T$ is zero [32]. This ensures enough noise is added during training so as to mitigate this difference between train and test time usage.", + "bbox": [ + 496, + 262, + 890, + 488 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Implementation Details. We build upon a text-to-image latent diffusion model [49] trained on a large in-house image-text dataset. It leverages a VAE [26] to map the images to a 4D latent space, with a $8 \\times$ reduction in spatial resolution. The U-Net largely follows the implementation from [49], with minor architectural modifications. As conditioning, it uses the Flan-T5-XXL text encoder [10]. We train Stacked-Diffusion for 14000 steps using the AdamW optimizer [35] with a learning rate of $10^{-4}$ , weight decay of 0.01, and gradient clipping at $\\ell_2$ norm of 1.0. For classifier-free guidance, we use a conditioning dropout rate of 0.05. We choose to generate at most $N = 6$ images simultaneously, because as we will see in § 5.1, the vast majority of the data available has 6 or fewer steps. For shorter training sequences, we pad with empty frames and dummy step texts, and for longer, we drop the extra steps and images. We ablate this choice of $N$ in § 5.4. Hence in practice, for batch of 8 sets of 6 steps at 256px resolution, we get an encoded latent of $8 \\times 6 \\times 4 \\times 32 \\times 32$ . We concatenate the latents spatially into $8 \\times 4 \\times (6 * 32) \\times 32$ . The denoising U-Net is finetuned to denoise these stacked images, leveraging what it has previously learned about spatial consistency, and adapts to this new setting. At inference time, steps are generated using a pretrained LLM [44], prompted to generate at most $N$ steps ( $\\S 10$ ). To generate illustrations, noise is sampled in the shape of the tiled latent with $N$ steps, and the steps generated by the LLM are padded with dummy steps if less", + "bbox": [ + 496, + 493, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6277", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/abdd9c4211ed1cebf5401101ac227efbb6d2e36a7e02f360f2b02270f203e1cb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 157, + 209, + 258 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b85bd28800215fdabb0f01e6dfd03b75541a101b14794e1866016d50c4120439.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 209, + 157, + 336, + 258 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/69e3b1946c36b74b5383b6f23d0211d25bb483f9edc8239e70cb340ac12e54d1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 336, + 159, + 465, + 258 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ce3a95db7a595c6bade73f555e74213878dd23564813b3b498a0becd576fa658.jpg", + "image_caption": [ + "Figure 5. Metrics. We introduce three metrics, one for each desideratum presented in § 3. Goal faithfulness: the second image does not show muffins. Step faithfulness: the second image does not show the step. Cross-image consistency: the second image shows a different number of meatballs with different visuals." + ], + "image_footnote": [], + "bbox": [ + 80, + 258, + 207, + 359 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0fedcf6895c2292a06f1ceb20ffba2a1d5784ef75c7c2a170c7501cdd6f60a36.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 209, + 258, + 336, + 359 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c1e4cf8496c64bbf680b983e316127bd5eeba422f3f6da5bf10871fd0a4ad473.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 336, + 258, + 465, + 358 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "than $N$ . The resulting noise is denoised using the U-Net as usual, and finally reshaped to obtain the $N$ output images upon decoding. See Appendix § 15 for further details.", + "bbox": [ + 75, + 455, + 468, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 518, + 209, + 535 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We train and evaluate StackedDiffusion on web-based instructional data. We compare the generations to multiple established baselines and ground truth, using automatic metrics and human evaluations. We now describe the data, metrics, baselines, and the key results. Finally, we demonstrate some new applications that StackedDiffusion enables, showing how it goes beyond standard instructional articles.", + "bbox": [ + 75, + 545, + 468, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Illustrated Instructions Dataset", + "text_level": 1, + "bbox": [ + 76, + 662, + 354, + 679 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We introduce a new dataset to train and evaluate models for the Illustrated Instructions task, by repurposing the Visual Goal-Step Inference (VGSI) dataset [67]. VGSI consists of WikiHow articles, each of which have a high-level goal, 6-10 natural language steps (see Figure 4), and associated image illustrations. We observe that the same data can be used for generating instructional articles. It can provide signal for how the goal and step texts should map to output images, and how images should match with each other.", + "bbox": [ + 75, + 686, + 468, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For evaluation, we construct a held-out set from this same data. Used directly, however, the data is not well suited as many of the tasks are too high-level to be useful for consistent illustration. For instance, \"How to Start a Business in North Carolina\" may have steps \"Brainstorm", + "bbox": [ + 75, + 825, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/5669acf8d2a3b2761a8c7acd1772b0752af26497ee99e4bfba20eab8f209707e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
T2I (Frozen)22.092.943.251.369.3
T2I (Finetuned)33.378.852.451.553.5
StackedDiffusion(ref)74.361.550.739.5
Ground Truth82.581.773.750.6(N/A)
", + "bbox": [ + 501, + 104, + 895, + 172 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Comparison to baselines. Human evaluation is reported as win rate vs our full StackedDiffusion model. GF corresponds to goal faithfulness accuracy, SF to step faithfulness accuracy, CIC to cross-image consistency, and FID to Fréchet Inception Distance.", + "bbox": [ + 498, + 184, + 893, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ideas” and “Go to the courthouse” that have no shared visual content. As such, we filter the data for evaluation to the “Recipes” category, which always has a visually clear end state and where illustrations must be consistent with each other. Please see Appendix § 11 for more details.", + "bbox": [ + 498, + 268, + 892, + 345 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Metrics", + "text_level": 1, + "bbox": [ + 500, + 358, + 594, + 372 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Having established three desiderata in § 3, we now turn to the question of how to evaluate them. In addition to evaluating the quality of images using FID [22], we propose three metrics, one for each desideratum, that can be used to evaluate the faithfulness and consistency of the generated article. Finally, given the limitations of automatic metrics for evaluating generative modeling tasks [55], we use human evaluations as our primary metric for overall quality. We illustrate the metrics in Figure 5 and briefly describe them next. Please see the Appendix § 12 for more details.", + "bbox": [ + 496, + 382, + 892, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Goal Faithfulness (GF) measures how well the generated image is associated to the goal text. We evaluate this by constructing multiple-choice questions (MCQ) as in VGSI [67]. For each generated image, we compare its CLIP similarity [47] with the correct goal text vs the similarity with the texts of three other randomly selected goals. We compute the accuracy of the model in choosing the correct goal text. Step Faithfulness (SF) measures how faithfully the generated image illustrates the step it aims to depict. An image should match the text for the step it was made for, more than other steps. We measure this using CLIP similarity and a MCQ task similar to goal faithfulness, where the image should have higher CLIP similarity with the corresponding step text than the other step texts within the same goal.", + "bbox": [ + 496, + 535, + 893, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Cross-Image Consistency (CIC) evaluates how consistent the generated images for a goal are with each other and penalizes jarring inconsistencies across the images, such as objects changing color or number. For instance, if a particular set of ingredients are shown for \"gather the ingredients,\" then the same ingredients should be shown for \"mix the dry ingredients.\" We measure this by computing the average $\\ell_2$ distance between DINO [7] embeddings of the images for each step, as this considers the similarity of visual (rather than purely semantic) features. Like prior work [50], we", + "bbox": [ + 496, + 750, + 893, + 901 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Goal Faithfulness", + "bbox": [ + 96, + 92, + 192, + 104 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Step Faithfulness", + "bbox": [ + 223, + 93, + 320, + 104 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Cross-Image Consistency", + "bbox": [ + 330, + 93, + 467, + 104 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Goal: Baking Muffins", + "bbox": [ + 86, + 114, + 202, + 127 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Step: Put the muffins in the oven", + "bbox": [ + 210, + 116, + 334, + 138 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Previous Image:", + "bbox": [ + 344, + 116, + 390, + 138 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "#", + "bbox": [ + 401, + 104, + 467, + 159 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Human $(\\uparrow)$ GF $(\\uparrow)$ SF $(\\uparrow)$ CIC $(\\downarrow)$ FID $(\\downarrow)$", + "bbox": [ + 617, + 88, + 892, + 103 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6278", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8c7902d9d52dc15231ab9b688dcca9b7b887bd76069f368d1e1ce7da4c184769.jpg", + "image_caption": [ + "Goal: How to Bake Turkey Wings", + "Figure 6. Baseline comparison. StackedDiffusion images are preferred overwhelmingly over baselines. The frozen baseline tends to only produce images showing the goal, while the finetuned baseline produces images that are more faithful to the step text, but have no visual features in common." + ], + "image_footnote": [], + "bbox": [ + 83, + 99, + 467, + 289 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "also found that CLIP features do not work well for image similarity as they are invariant to critical aspects such as color, number of objects, style etc. DINO's self-supervised pretraining objective leads to features that are sensitive to the visual aspects particular to an image, rather than being invariant to aspects not captured by category or language.", + "bbox": [ + 75, + 383, + 468, + 474 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Human Evaluation is finally used to confirm that the automated evaluation corresponds to actual human preferences. We provide fully rendered articles from each model to human evaluators on Amazon Mechanical Turk (AMT), and ask them to choose which article they prefer, or if they are tied. We then consider the win rate as the proportion for which a majority of evaluators selected the given method compared to the articles produced by StackedDiffusion. A win rate of 50 denotes the methods are tied, whereas $< 50$ denotes the method is worse than StackedDiffusion.", + "bbox": [ + 75, + 474, + 468, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Baselines", + "text_level": 1, + "bbox": [ + 76, + 635, + 184, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We now describe some baseline approaches based on existing state-of-the-art image generation models, and compare them to StackedDiffusion in Table 1.", + "bbox": [ + 75, + 659, + 468, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "T2I (frozen): The most obvious choice for a baseline is to simply use a pretrained and frozen text-to-image (T2I) model. We use our in-house T2I model that is also used to initialize StackedDiffusion. To ensure this baseline gets same information as StackedDiffusion, we prompt the model with a concatenation of the goal with each step, $\\mathbf{g}||\\mathbf{s}_i$ , for each $i$ , and produce a single image $\\mathbf{I}_i$ . The sequence then is composed of $N$ independent generations.", + "bbox": [ + 75, + 703, + 468, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We find that these images are faithful to the goal text, but fail to be faithful to the step text and lack consistency. The goal faithfulness is substantially higher than even the ground truth (92.9% vs 81.7%). This is because the model is not capable of reasoning about what a particular step to-", + "bbox": [ + 75, + 825, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d01c9dd6964ee9155412d9b38e95567509bba414f3c9854c8639419d5446d72f.jpg", + "image_caption": [ + "Figure 7. Effect of training data. We find that having more training data improves StackedDiffusion's faithfulness and consistency. However even with $20\\%$ data it performs well, thanks to its effective use of pretrained T2I model weights." + ], + "image_footnote": [], + "bbox": [ + 516, + 87, + 879, + 215 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "wards a goal should look like. Instead, it simply generates an image that matches words in the input text, as illustrated in the first row of Figure 2. This results in all the images for e.g. \"Make Colored Ice\" showing the actual finished ice rather than any of the intermediate steps. In other words, the step text is eclipsed by the goal text.", + "bbox": [ + 496, + 301, + 890, + 392 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "T2I (finetuned): We finetune a T2I model with the goal and step text embeddings concatenated together as the condition, again producing $N$ independent generations. We find that this model is more faithful to the step text than the frozen T2I model, but still substantially less than our final model (52.4% vs 61.5%). This suggests the goal still overshadows the step text, but to a lesser extent. The goal faithfulness is also lower than for the frozen T2I model, as this model is trained to create images for steps rather than goals. The cross-image consistency is low, similar to the frozen model, as the images are still generated independently.", + "bbox": [ + 496, + 393, + 892, + 560 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "StackedDiffusion achieves substantially improved step faithfulness and cross-image consistency, while maintaining high goal faithfulness. We show an illustrative example in Figure 6. Unlike the baselines, the cross-image consistency is very close to that of the ground truth data; as generating them together allows for influence between them during the generation process. We also find that the FID of our generated images is substantially lower than that of other models, indicating that our generated images are more similar to the ground truth data. Finally on human evaluations, StackedDiffusion clearly outperforms them both by a preference of $78\\%$ and $66.6\\%$ respectively. Perhaps most notably, when compared to the ground truth images, human evaluators still picked StackedDiffusion $18.5\\%$ of the times, suggesting that our model generates some illustrations even better than the ground truth data. Further baselines can be found in Appendix § 13, with a comparison of training and inference costs in Appendix § 14.", + "bbox": [ + 496, + 560, + 892, + 833 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.4. Ablations", + "text_level": 1, + "bbox": [ + 500, + 845, + 609, + 859 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Step Count. As discussed in § 4, we elect to use a maximum step count of 6 as it allows sufficient data coverage", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6279", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f040870fd95e35069d76d20ae071203208224b8ec315b36438d5b3f58474fcbb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Win RateWin Rate
T2I (Frozen)33.8T2I (Finetuned)34.0
GILL [27]++3.9LLM + CLIP15.6
Goal Retrieval33.1Ground Truth70.0
", + "bbox": [ + 78, + 88, + 468, + 159 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and most articles fall under this value. We find that human evaluators prefer this model to the shorter step count model of length 4. It is slightly preferred over the model trained with a longer step count of 8, likely as there is insufficient data of long lengths, limiting any advantage that might be gained in longer length generations.", + "bbox": [ + 75, + 234, + 470, + 325 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/bb43af274d1595779da6b55e8c75dab6c1c33926e5ba89acc96b43d19bae226b.jpg", + "table_caption": [ + "Table 2. System-level comparison to prior work. Human evaluation results using fully generated outputs. Reported as human evaluation win rate of the associated method over StackedDiffusion (hence, higher is better for the reported method)." + ], + "table_footnote": [], + "table_body": "
Baseline4 Step (↑)8 Step (↑)
6 Step32.146.9
", + "bbox": [ + 147, + 335, + 400, + 375 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Training data. In Figure 7, we randomly sub-sample varying proportions of the total data, and train StackedDiffusion on each of those subsets. As the amount of training data increases, it generates more faithful and consistent images. However even with less data it performs well, owing to its effective use of the pretrained T2I model's initialization.", + "bbox": [ + 75, + 383, + 468, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Importance of 0SNR. We experiment with the 0SNR technique [32] due to the high spatial dimensionality of the latents. We find that not using 0SNR results in substantially reduced step faithfulness, dropping from $61.5\\%$ to $52.8\\%$ .", + "bbox": [ + 75, + 474, + 468, + 535 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Importance of Step-Positional Encoding. We also evaluate the importance of the step-positional encoding, $\\Phi_{i}$ . This gives the model a critical cue for which parts of the condition correspond to which steps. We find that not using the step-positional encoding results in substantially reduced step faithfulness as well, dropping from $61.5\\%$ to $49.8\\%$ .", + "bbox": [ + 75, + 535, + 468, + 626 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.5. Comparison to prior work", + "text_level": 1, + "bbox": [ + 76, + 635, + 316, + 651 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The previous metrics were computed with respect to fixed ground truth goal and step texts to provide direct comparisons. However, they do not encompass the full scope of our model's capabilities: the flexibility of the generated text is one of StackedDiffusion's core strengths. To evaluate in a closer setting to this real-world usage, we examine the quality of the articles generated by the full system, where the steps are generated with a LLM [44]. We perform system-level comparisons using human evaluators, including the quality of the generated text and the quality of the images created from the generated text.", + "bbox": [ + 75, + 659, + 468, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We show our results in Table 2. We first compare to the baselines introduced in § 5.3, and see results similar to Table 1. We see that even with LLM-generated text, Stacked-Diffusion outperforms baseline T2I based approaches, either frozen or finetuned, by more than $66\\%$ win-rate.", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/47d8b68fa6dad61088500ed14d3c77f59154410ea0b9c57e3bf7ef12f1180920.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 90, + 890, + 128 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/30e376064a07d20776c2aed48e05562dd3559b4e43dbd3f254616a1f596185ad.jpg", + "image_caption": [ + "Figure 8. Personalization. StackedDiffusion enables instructions personalized far beyond what is possible with preexisting articles." + ], + "image_footnote": [ + "Step 1: Drain and press tofu to remove excess water. Step 2: Crumble tofu; mix with bread crumbs and a binding agent. Step 3: Shape tofu mixture into burger patties. Step 4: Cook over medium heat until golden brown." + ], + "bbox": [ + 506, + 159, + 599, + 229 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2348732110f808aca1b589d07bbf4efcca036cfb0c16289a20171014c1d884d4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 604, + 159, + 694, + 229 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ec4155dfeef0eac5a21e03881668f562e1414ec8874f7671ca5b099fa54a8bd3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 159, + 792, + 229 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d177b9569dac71da510a5c24b37d1bad26592d5be34789f6db9a26cc052bfcbd.jpg", + "image_caption": [], + "image_footnote": [ + "$\\left( {x - {2x}}\\right) t - x{y}^{2} = \\left( {x - {2x}}\\right) {f}^{\\prime }t$" + ], + "bbox": [ + 794, + 159, + 885, + 229 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Additionally, we compare StackedDiffusion to recent state-of-the-art approaches in multimodal generation, as in principle those can also produce instructional text combined with illustrations. Specifically, we compare to GILL [29]-- a model trained to generate sequences of interleaved text and images. While other similar approaches have been proposed [1, 28], GILL is the best open-source model we could access. Using GILL directly, however, results in a human evaluation win rate of $0\\%$ . We posit this was for three primary reasons. First, the GILL model was unable to generate long enough generations to suffice as an article in comparison to our generations, despite our best efforts in modifying the generation parameters. Secondly, the model often did not produce any illustrations with the text, although we tried various prompting tactics. Thirdly, the text or images produced were often not of high enough quality to actually reflect the goal or steps. We thus introduce an alternative, which we call GILL++. This uses the (superior) text produced by the same language model as in our method, but passes each step text as input to GILL to generate the corresponding image. This results in a human evaluation win rate of $3.9\\%$ against StackedDiffusion. See Appendix § 16 for details on prompting GILL, and our GILL++ baseline.", + "bbox": [ + 496, + 292, + 892, + 640 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Next, we compare StackedDiffusion to a retrieval based approach, denoted by LLM+CLIP. Here we retrieve the closest image in the training set according to CLIP similarity to each of the steps created by the LLM (with each concatenated to the goal text, similar to how the frozen T2I model is used). This retrieval-based metric is not suitable for the automated metrics previously introduced as it is based on the same underlying similarity scores as the automated metrics, but we introduce it here for human evaluation. We find that StackedDiffusion is preferred overwhelmingly over this retrieval-based approach, despite the retrieval-based approach using real images from the training set. Similarly, we compare to another retrieval-based approach which retrieves the full article with the most similar goal text to the input goal, denoted Goal Retrieval. We find StackedDiffusion strongly outperforms this as well. These results together suggest that StackedDiffusion is able", + "bbox": [ + 496, + 643, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6280", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d4b653e492533a300057bdbe939d13e17d48e00c7e0b6735708fe389635afc0b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 90, + 109, + 114 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "I have a bunch of apples left from apple picking, what can I do with them?", + "bbox": [ + 117, + 93, + 447, + 114 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "If you have a lot of apples, you can turn them into a delicious apple crisp. Here's a simple recipe to get you started:", + "bbox": [ + 88, + 118, + 410, + 140 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e00d1a3038a1f738d21b711761e3aa64b83c497911abf45e70133c1ba0ae463f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 436, + 116, + 464, + 141 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Step 1: Preheat oven to $350^{\\circ}\\mathrm{F}$ $(175^{\\circ}\\mathrm{C})$", + "bbox": [ + 88, + 146, + 197, + 169 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Step 2: Peel, core, and slice apples; place in a baking dish.", + "bbox": [ + 215, + 146, + 326, + 179 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Step 3: Mix sugar, cinnamon, and a pinch of salt; sprinkle over apples.", + "bbox": [ + 344, + 146, + 459, + 188 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5f76b982470e95c1a04b54ce7ae0f0afdaba7682efdab2ea819e9352c3bca48a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 84, + 188, + 207, + 282 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b1c1e3407c5aa95a720bcb4c1c6c5943e545d72d586d640120c8b8f2eafa8863.jpg", + "image_caption": [ + "Step 5: Bake for 45 minutes or until golden brown." + ], + "image_footnote": [], + "bbox": [ + 212, + 188, + 334, + 282 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b05c9784cb602e72512f7c6d59ede22c192a020dd1bd392ecadebc28f2d9f715.jpg", + "image_caption": [ + "Step 6: Serve warm." + ], + "image_footnote": [], + "bbox": [ + 341, + 188, + 462, + 282 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/157a7442a98421ef62dc2388a1be819b14732a906eea65091a219119a50cdc22.jpg", + "image_caption": [ + "Step 4: Combine oats, flour, brown sugar, and butter; crumble over apples.", + "Figure 9. Goal suggestion and inference. StackedDiffusion can suggest what goal might be most relevant given other information." + ], + "image_footnote": [], + "bbox": [ + 84, + 325, + 207, + 420 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1321800589acd81068bc2c5de5b02f49a33f548f8744f3bf84dfb6aca7cd8204.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 212, + 325, + 334, + 420 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9c76d104f77fa03d55ab031c6807146db39a5db9c5d18bc3da885a9bdaa21beb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 341, + 325, + 462, + 420 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "to generate novel instructional articles that one can not simply retrieve from the training corpus.", + "bbox": [ + 75, + 474, + 468, + 505 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, we compare StackedDiffusion to the ground truth. Surprisingly again, we find that the human annotators pick our generations $30\\%$ of times even compared to ground truth. Note that these articles are handwritten and manually illustrated for the purpose of illustrating these goals. This shows the strong promise of our proposed approach. We believe with the rapid improvements in training data and generative modeling techniques, future iterations of our model could be indistinguishable or even better than manually created illustrated instructional articles.", + "bbox": [ + 75, + 505, + 468, + 655 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.6. Applications", + "text_level": 1, + "bbox": [ + 76, + 665, + 210, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Perhaps the biggest strength of a generative approach is its ability to handle unique user queries that may not fit into the boundaries posed by static articles on the web. Hence, we demonstrate the capabilities of the full system, combining the strengths of text generation capabilities from the language model with StackedDiffusion.", + "bbox": [ + 75, + 688, + 468, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Personalized instruction. A user can provide any situational information specific to their circumstances and obtain an article customized to that situation. This is not possible with fixed, existing articles. For example, in Figure 8, the user can specify a diet and obtain an article for the food they want that adheres to that diet.", + "bbox": [ + 75, + 780, + 468, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Goal suggestion. As StackedDiffusion accepts flexible input text, a user can provide higher-level information about", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8cad8b81aae740759c89e121102a47a3149f97aaa02e952b7f0084b992639edb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 90, + 535, + 113 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Oh no! I'm making bread and forgot to add salt when I made the dough! I'm at the kneading step.", + "bbox": [ + 540, + 90, + 879, + 113 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "If you're at the kneading step, don't worry—you can still incorporate it into the dough with more kneading. Here's how:", + "bbox": [ + 509, + 118, + 839, + 140 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/71ab28f90bc8e280ee1341f7124499b16a1260acdeae65b53746c69b2614a258.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 859, + 116, + 887, + 141 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Step 1: Determine the amount of salt needed.", + "bbox": [ + 509, + 146, + 622, + 179 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Step 2: Dissolve the salt in a small amount of warm water.", + "bbox": [ + 638, + 146, + 750, + 179 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Step 3: Flatten the dough into a rectangle.", + "bbox": [ + 766, + 146, + 880, + 169 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/10698da331ef96acbc1722d2850a1bffea0f43d33a1f952ff8ed107c2de59022.jpg", + "image_caption": [ + "Step 4: Spread the salt over the dough." + ], + "image_footnote": [], + "bbox": [ + 506, + 186, + 629, + 282 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5a9ec6801506eaa2482dff9b1df28cae824c3c53e6bb31f7365e12d82e9b060e.jpg", + "image_caption": [ + "Step 5: Fold and knead the dough to distribute." + ], + "image_footnote": [], + "bbox": [ + 635, + 186, + 756, + 282 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7aaf6a10495f2721bc93a9ad1d73b033ad5ff37563972cae78a779ae62c29f69.jpg", + "image_caption": [ + "Step 6: Proceed with the original bread recipe." + ], + "image_footnote": [], + "bbox": [ + 763, + 188, + 885, + 282 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4d3dd587c80162601dfb094dcfae5e45d3e615ea78c912df4e7fb6fa3d390f55.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 325, + 629, + 419 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e8b979dbd1ff2ba95c5013084b3030110152c106f174a67e1949ac6baff1f842.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 325, + 756, + 419 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9a4282b040e672208ad09baa2280751c0806c52f20519d106ff1dcc6b7b1822a.jpg", + "image_caption": [ + "Figure 10. Error correction. StackedDiffusion provides updated instructions in response to unexpected situations, like a user error." + ], + "image_footnote": [], + "bbox": [ + 763, + 325, + 885, + 420 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "what they want to do and obtain an article with a suggested specific goal that matches what the user would like to do. In Figure 9, the user describes their situation (having many apples) and StackedDiffusion suggests a goal that matches their situation (making apple crisps).", + "bbox": [ + 498, + 482, + 890, + 558 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Error correction. Oftentimes, a user in the course of performing a task may make a mistake. Fixed articles provide no avenue for recourse. However, StackedDiffusion can adapt to user error and create alternative instructions that best correct for and accommodate the user's situation. See Figure 10 for an example of this. A fixed article would not provide any alternative paths that adapt as the user performs the action. This highlights the new avenues opened up by our generative approach to this task.", + "bbox": [ + 498, + 561, + 890, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Limitations and Conclusions", + "text_level": 1, + "bbox": [ + 500, + 720, + 764, + 736 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While StackedDiffusion achieves strong performance in generating Illustrated Instructions, we note that there still remains a gap with ground truth data. Future work might examine how to obtain instructional data at greater scale and how this influences StackedDiffusion. Leveraging the latest T2I architectures to improve the faithfulness and quality of our generations would be another promising avenue for future work. Finally, leveraging improvements in text-to-video techniques to generate video clips describing each step would further improve the usability of such a system.", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6281", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Armen Aghajanyan, Bernie Huang, Candace Ross, Vladimir Karpukhin, Hu Xu, Naman Goyal, Dmytro Okhonko, Mandar Joshi, Gargi Ghosh, Mike Lewis, and Luke Zettlemoyer. CM3: A causal masked multimodal model of the internet. arXiv preprint arXiv:2201.07520, 2022.", + "[2] Emanuele Aiello, Lili Yu, Yixin Nie, Armen Aghajanyan, and Barlas Oguz. Jointly training large autoregressive multimodal models. In ICLR, 2024.", + "[3] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, 2021.", + "[4] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022.", + "[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your Latents: High-Resolution Video Synthesis with Latent Diffusion Models. In CVPR, 2023.", + "[6] Russell N. Carney and Joel R. Levin. Pictorial illustrations still improve students' learning from text. Educational Psychology Review, 14(1):5-26, 2002.", + "[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging Properties in Self-Supervised Vision Transformers. In ICCV, 2021.", + "[8] Huiwen Chang, Han Zhang, Jarred Barber, Aaron Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Patrick Murphy, William T. Freeman, Michael Rubinstein, Yanzhen Li, and Dilip Krishnan. Muse: Text-ToImage Generation via Masked Generative Transformers. In ICML, 2023.", + "[9] Brian Chen, Andrew Rouditchenko, Kevin Duarte, Hilde Kuehne, Samuel Thomas, Angie Boggust, Rameswar Panda, Brian Kingsbury, Rogerio Feris, David Harwath, et al. Multimodal clustering networks for self-supervised learning from unlabeled videos. In ICCV, 2021.", + "[10] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Alex Castro-Ros, Marie Pellat, Kevin Robinson, Dasha Valter, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416, 2022.", + "[11] Yang Deng, Wai Lam, Yuexiang Xie, Daoyuan Chen, Yaliang Li, Min Yang, and Ying Shen. Joint Learning of Answer Selection and Answer Summary Generation in Community Question Answering. In AAAI, 2020.", + "[12] Eadom Dessalene, Chinmaya Devaraj, Michael Maynord, Cornelia Fermuller, and Yiannis Aloimonos. Forecasting ac" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tion through contact representations from first person video. PAMI, 2021.", + "[13] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In ICCV, 2023.", + "[14] Bahare Fatemi, Quentin Duval, Rohit Girdhar, Michal Drozdzal, and Adriana Romero-Soriano. Learning to substitute ingredients in recipes. arXiv preprint arXiv:2302.07960, 2023.", + "[15] Antonino Furnari and Giovanni Maria Farinella. Rolling-unrolling lstms for action anticipation from first-person video. PAMI, 2020.", + "[16] Valentin Gabeur, Chen Sun, Karteek Alahari, and Cordelia Schmid. Multi-modal transformer for video retrieval. In ECCV, 2020.", + "[17] Valentin Gabeur, Arsha Nagrani, Chen Sun, Karteek Alahari, and Cordelia Schmid. Masking modalities for cross-modal video retrieval. In WACV, 2022.", + "[18] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. In ECCV, 2022.", + "[19] Rohit Girdhar and Kristen Grauman. Anticipative video transformer. In ICCV, 2021.", + "[20] Rohit Girdhar, Mannat Singh, Andrew Brown, Quentin Duval, Samaneh Azadi, Sai Saketh Rambhatla, Akbar Shah, Xi Yin, Devi Parikh, and Ishan Misra. Emu video: Factorizing text-to-video generation by explicit image conditioning. arXiv preprint arXiv:2311.10709, 2023.", + "[21] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In CVPR, 2022.", + "[22] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017.", + "[23] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen Video: High Definition Video Generation with Diffusion Models. arXiv preprint arXiv:2210.02303, 2022.", + "[24] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022.", + "[25] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021.", + "[26] Diederik P. Kingma and Max Welling. Auto-Encoding Variational Bayes. In ICLR, 2014.", + "[27] Jing Yu Koh, Daniel Fried, and Ruslan Salakhutdinov. Generating images with multimodal language models. In NeurIPS, 2023." + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6282", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Jing Yu Koh, Ruslan Salakhutdinov, and Daniel Fried. Grounding language models to images for multimodal inputs and outputs. In ICML, 2023.", + "[29] Jing Yu Koh, Ruslan Salakhutdinov, and Daniel Fried. Grounding Language Models to Images for Multimodal Generation. In NeurIPS, 2023.", + "[30] Mahnaz Koupae and William Yang Wang. WikiHow: A Large Scale Text Summarization Dataset. arXiv preprint arXiv:1810.09305, 2018.", + "[31] Linjie Li, Yen-Chun Chen, Yu Cheng, Zhe Gan, Licheng Yu, and Jingjing Liu. Hero: Hierarchical encoder for video+ language omni-representation pre-training. arXiv preprint arXiv:2005.00200, 2020.", + "[32] Shanchuan Lin, Bingchen Liu, Jiashi Li, and Xiao Yang. Common diffusion noise schedules and sample steps are flawed. arXiv preprint arXiv:2305.08891, 2023.", + "[33] Xudong Lin, Fabio Petroni, Gedas Bertasius, Marcus Rohrbach, Shih-Fu Chang, and Lorenzo Torresani. Learning To Recognize Procedural Activities with Distant Supervision. In CVPR, 2022.", + "[34] Song Liu, Haoqi Fan, Shengsheng Qian, Yiru Chen, Wenkui Ding, and Zhongyuan Wang. Hit: Hierarchical transformer with momentum contrast for video-text retrieval. In ICCV, 2021.", + "[35] Ilya Loshchilov and Frank Hutter. Decoupled Weight Decay Regularization. In ICLR, 2019.", + "[36] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon Bharti, and Ming Zhou. Univl: A unified video and language pre-training model for multimodal understanding and generation. arXiv preprint arXiv:2002.06353, 2020.", + "[37] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval. arXiv preprint arXiv:2104.08860, 2021.", + "[38] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Ti-niu Tan. Videofusion: Decomposed diffusion models for high-quality video generation. In CVPR, 2023.", + "[39] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. HowTo100M: Learning a Text-Video Embedding by Watching Hundred Million Narrated Video Clips. In ICCV, 2019.", + "[40] Antoine Miech, Jean-Baptiste Alayrac, Ivan Laptev, Josef Sivic, and Andrew Zisserman. Thinking fast and slow: Efficient text-to-visual retrieval with transformers. In CVPR, 2021.", + "[41] Medhini Narasimhan, Anna Rohrbach, and Trevor Darrell. Clip-it! language-guided video summarization. NIPS, 2021.", + "[42] Medhini Narasimhan, Arsha Nagrani, Chen Sun, Michael Rubinstein, Trevor Darrell, Anna Rohrbach, and Cordelia Schmid. Tl; dw? summarizing instructional videos with task relevance and cross-modal saliency. In ECCV, 2022.", + "[43] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2022.", + "[44] OpenAI. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023.", + "[45] Gaurav Parmar, Richard Zhang, and Jun-Yan Zhu. On aliased resizing and surprising subtleties in gan evaluation. In CVPR, 2022.", + "[46] AJ Piergiovanni, Anelia Angelova, Michael S Ryoo, and Irfan Essa. Unsupervised action segmentation for instructional videos. arXiv preprint arXiv:2106.03738, 2021.", + "[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021.", + "[48] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022.", + "[49] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bjorn Ommer. High-Resolution Image Synthesis with Latent Diffusion Models. In CVPR, 2022.", + "[50] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. DreamBooth: Fine Tuning Text-to-Image Diffusion Models for Subject-Driven Generation. In CVPR, 2023.", + "[51] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding. In NIPS, 2022.", + "[52] Fadime Sener and Angela Yao. Zero-shot anticipation for instructional activities. In ICCV, 2019.", + "[53] Paul Hongsuck Seo, Arsha Nagrani, and Cordelia Schmid. Look before you speak: Visually contextualized utterances. In CVPR, 2021.", + "[54] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-A-Video: Text-to-Video Generation without Text-Video Data. arXiv preprint arXiv:2209.14792, 2022.", + "[55] George Stein, Jesse C Cresswell, Rasa Hosseinzadeh, Yi Sui, Brendan Leigh Ross, Valentin Villecloze, Zhaoyan Liu, Anthony L Caterini, J Eric T Taylor, and Gabriel Loiza-Ganem. Exposing flaws of generative model evaluation metrics and their unfair treatment of diffusion models. arXiv preprint arXiv:2306.04675, 2023.", + "[56] Chen Sun, Fabien Baradel, Kevin Murphy, and Cordelia Schmid. Learning video representations using contrastive bidirectional transformer. arXiv preprint arXiv:1906.05743, 2019.", + "[57] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Emu: Generative pretraining in multimodality. In ICLR, 2024." + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6283", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[58] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. COIN: A Large-scale Dataset for Comprehensive Instructional Video Analysis. In CVPR, 2019.", + "[59] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable Length Video Generation from Open Domain Textual Descriptions. In ICLR, 2022.", + "[60] Wenzhe Wang, Mengdan Zhang, Runnan Chen, Guanyu Cai, Penghao Zhou, Pai Peng, Xiaowei Guo, Jian Wu, and Xing Sun. Dig into multi-modal cues for video retrieval with hierarchical alignment. In *IJCAI*, 2021.", + "[61] Chao-Yuan Wu, Yanghao Li, Karttikeya Mangalam, Haoqi Fan, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Memvit: Memory-augmented multiscale vision transformer for efficient long-term video recognition. arXiv preprint arXiv:2201.08383, 2022.", + "[62] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In ICCV, 2023.", + "[63] Mengmeng Xu, Juan-Manuel Pérez-Rúa, Victor Escorcia, Brais Martinez, Xiatian Zhu, Li Zhang, Bernard Ghanem, and Tao Xiang. Boundary-sensitive pre-training for temporal localization in videos. In ICCV, 2021.", + "[64] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Just ask: Learning to answer questions from millions of narrated videos. In ICCV, 2021.", + "[65] Jianwei Yang, Yonatan Bisk, and Jianfeng Gao. Taco: Token-aware cascade contrastive learning for video-text alignment. In ICCV, 2021.", + "[66] Mengjiao Yang, Yilun Du, Bo Dai, Dale Schuurmans, Joshua B Tenenbaum, and Pieter Abbeel. Probabilistic adaptation of text-to-video models. arXiv preprint arXiv:2306.01872, 2023.", + "[67] Yue Yang, Artemis Panagopoulou, Qing Lyu, Li Zhang, Mark Yatskar, and Chris Callison-Burch. Visual goal-step inference using wikihow. In EMNLP, 2021.", + "[68] Shengming Yin, Chenfei Wu, Huan Yang, Jianfeng Wang, Xiaodong Wang, Minheng Ni, Zhengyuan Yang, Linjie Li, Shuguang Liu, Fan Yang, Jianlong Fu, Gong Ming, Lijuan Wang, Zicheng Liu, Houqiang Li, and Nan Duan. NUWAXL: Diffusion over Diffusion for eXtremely Long Video Generation. arXiv preprint arXiv:2303.12346, 2023.", + "[69] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022.", + "[70] Lili Yu, Bowen Shi, Ramakanth Pasunuru, Benjamin Muller, Olga Golovneva, Tianlu Wang, Arun Babu, Binh Tang, Brian Karrer, Shelly Sheynin, et al. Scaling autoregressive multimodal models: Pretraining and instruction tuning. arXiv preprint arXiv:2309.02591, 2023." + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[71] Jingqing Zhang, Yao Zhao, Mohammad Saleh, and Peter J. Liu. PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization. In ICML, 2019.", + "[72] Li Zhang, Qing Lyu, and Chris Callison-Burch. Reasoning about Goals, Steps, and Temporal Ordering with WikiHow. In EMNLP, 2020.", + "[73] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022.", + "[74] Shuyan Zhou, Li Zhang, Yue Yang, Qing Lyu, Pengcheng Yin, Chris Callison-Burch, and Graham Neubig. Show Me More Details: Discovering Hierarchies of Procedures from Semi-structured Web Data. In ACL, 2022.", + "[75] Yilun Zhou, Julie Shah, and Steven Schockaert. Learning Household Task Knowledge from WikiHow Descriptions. In Proceedings of the 5th Workshop on Semantic Deep Learning (SemDeep-5), 2019.", + "[76] Linchao Zhu and Yi Yang. Actbert: Learning global-local video-text representations. In CVPR, 2020.", + "[77] Dimitri Zhukov, Jean-Baptiste Alayrac, Ivan Laptev, and Josef Sivic. Learning actionness via long-range temporal order verification. In ECCV, 2020." + ], + "bbox": [ + 501, + 92, + 890, + 417 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "6284", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Generating Illustrated Instructions/dd71c86a-390f-4c46-b6ae-0a13f2787fa0_model.json b/2024/Generating Illustrated Instructions/dd71c86a-390f-4c46-b6ae-0a13f2787fa0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1bedce2281b1eacfb983bc4deedf379393cd2c49 --- /dev/null +++ b/2024/Generating Illustrated Instructions/dd71c86a-390f-4c46-b6ae-0a13f2787fa0_model.json @@ -0,0 +1,3038 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.307, + 0.131, + 0.664, + 0.154 + ], + "angle": 0, + "content": "Generating Illustrated Instructions" + }, + { + "type": "text", + "bbox": [ + 0.266, + 0.18, + 0.404, + 0.198 + ], + "angle": 0, + "content": "Sachit Menon\\(^{1,2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.439, + 0.181, + 0.541, + 0.198 + ], + "angle": 0, + "content": "Ishan Misra" + }, + { + "type": "text", + "bbox": [ + 0.584, + 0.181, + 0.701, + 0.198 + ], + "angle": 0, + "content": "Rohit Girdhar1" + }, + { + "type": "text", + "bbox": [ + 0.321, + 0.199, + 0.435, + 0.216 + ], + "angle": 0, + "content": "1GenAI, Meta" + }, + { + "type": "text", + "bbox": [ + 0.475, + 0.2, + 0.65, + 0.218 + ], + "angle": 0, + "content": "2Columbia University" + }, + { + "type": "text", + "bbox": [ + 0.253, + 0.22, + 0.711, + 0.233 + ], + "angle": 0, + "content": "https://facebookresearch.github.io/IllustratedInstructions" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.302, + 0.474, + 0.56 + ], + "angle": 0, + "content": "We introduce a new task of generating \"Illustrated Instructions\", i.e. visual instructions customized to a user's needs. We identify desiderata unique to this task, and formalize it through a suite of automatic and human evaluation metrics, designed to measure the validity, consistency, and efficacy of the generations. We combine the power of large language models (LLMs) together with strong text-to-image generation diffusion models to propose a simple approach called StackedDiffusion, which generates such illustrated instructions given text as input. The resulting model strongly outperforms baseline approaches and state-of-the-art multimodal LLMs; and in \\(30\\%\\) of cases, users even prefer it to human-generated articles. Most notably, it enables various new and exciting applications far beyond what static articles on the web can provide, such as personalized instructions complete with intermediate steps and pictures in response to a user's individual situation." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.59, + 0.21, + 0.606 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.617, + 0.47, + 0.798 + ], + "angle": 0, + "content": "The internet is a vast resource to find answers to all types of questions. It is often easy to find a webpage or a video that walks through the exact steps to achieve a user's goal. With the rise of Large Language Models (LLMs) trained on internet-scale data, users can just ask the LLM for instructions to achieve a goal. This allows the users to get answers for specific personalized queries for which there may not be an existing webpage on the internet, e.g., modifying cooking recipes with user specific dietary restrictions. Moreover, if the users make a mistake when following the instructions, simple follow-up questions to the LLM can generate alternate instructions, a major advantage over static web search." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.799, + 0.47, + 0.876 + ], + "angle": 0, + "content": "In spite of such advantages, LLMs still have one major limitation - they cannot generate visuals, which are critical for users to learn from and follow instructions for a wide range of tasks [6]. Consider, for instance, instructions that require visual inspection, e.g., a recipe that requires" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.272, + 0.535, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.543, + 0.278, + 0.835, + 0.289 + ], + "angle": 0, + "content": "How can I bake a cake to celebrate an event in Seattle?" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.298, + 0.846, + 0.32 + ], + "angle": 0, + "content": "One great way could be to make a coffee-infused cake to honor Seattle's coffee culture, with teal coloring for the Seattle flag." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.328, + 0.614, + 0.348 + ], + "angle": 0, + "content": "Step 1: Brew strong coffee; let cool." + }, + { + "type": "text", + "bbox": [ + 0.64, + 0.328, + 0.741, + 0.347 + ], + "angle": 0, + "content": "Step 2: Preheat the oven to \\(350\\mathrm{F}\\)" + }, + { + "type": "text", + "bbox": [ + 0.768, + 0.327, + 0.876, + 0.347 + ], + "angle": 0, + "content": "Step 3: Prepare cake batter as usual." + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.35, + 0.63, + 0.443 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.445, + 0.616, + 0.475 + ], + "angle": 0, + "content": "Step 4: Mix in the coffee; fold in coffee grounds for texture." + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.35, + 0.758, + 0.443 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.64, + 0.445, + 0.743, + 0.465 + ], + "angle": 0, + "content": "Step 5: Bake for 25- 30 minutes." + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.35, + 0.886, + 0.443 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.767, + 0.445, + 0.868, + 0.476 + ], + "angle": 0, + "content": "Step 6: Cover the cake in teal fondant for the Seattle flag." + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.476, + 0.63, + 0.57 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.477, + 0.758, + 0.57 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.477, + 0.886, + 0.57 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.582, + 0.892, + 0.652 + ], + "angle": 0, + "content": "Figure 1. StackedDiffusion generating Illustrated Instructions. Given a goal (or any textual user input), StackedDiffusion produces a customized instructional article complete with illustrations that not only tells the user how to achieve the goal in words, but also shows the user by providing illustrations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.672, + 0.892, + 0.747 + ], + "angle": 0, + "content": "the user to stir fry onions until golden brown, or searching for bubbles when a flat tire is submerged under water. An image accompanying such instructions can make following them significantly easier. Can we develop methods with the strengths of LLMs that can also generate such visuals?" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.902 + ], + "angle": 0, + "content": "In this work, we tackle this challenge, developing models that can not only tell a user how to accomplish their task, but also show them how. We define the novel task of Illustrated Instructions: creating a set of steps with visualizations that showcase an approach to solve the user's task. We carefully consider the different dimensions of the problem, laying out three desiderata unique to this setting. To measure these desiderata, we develop automated metrics based on prior work in instructional article assessment [67] and image generation [50]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.318, + 0.9 + ], + "angle": 0, + "content": "*Work done during an internship at Meta." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6274" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.35 + ], + "angle": 0, + "content": "We propose a new model to solve the task of Illustrated Instructions that combines LLMs with a text-to-image diffusion model. We train our model on stacks of instructional images from websites such as WikiHow. The resulting StackedDiffusion model creates full instructional articles, complete with customized steps and a sequence of images uniquely generated to describe those steps. It leverages large-scale pretrained LLMs and finetuned text-to-image diffusion models, and employs techniques to accomplish the task without the need to introduce any new learnable parameters. This includes spatial tiling for simultaneous multi-image generation, text embedding concatenation to reduce information loss in long instructions, and a new \"step-positional encoding\" to better demarcate the different steps in the instructions. Thus, StackedDiffusion can generate useful illustrations even when trained even with limited instructional data (cf. Figure 7)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.353, + 0.473, + 0.672 + ], + "angle": 0, + "content": "We compare StackedDiffusion with various baseline approaches based on off-the-shelf tools, and find that they all fall short, even when used in conjunction with one another. Existing T2I models are incapable of generating visuals directly from a user query. Even when given more detailed instructional text, we show that existing T2I models fail to produce images that are simultaneously faithful to the goal, the step, and consistent among each other. Given the recent introduction of multimodal LLMs [1, 28], we also compare with a recent open-source model, GILL [27], and show such models also fall short of consistently generating useful visuals along with text. We posit that our approach of leveraging the spatial priors learned by pretrained diffusion models to generate multiple images together, in conjunction with pretrained LLMs, is significantly more compute and data efficient than an approach purely based on next-token prediction, without these priors. Our thorough ablations and human evaluations show that StackedDiffusion convincingly surpasses state-of-the-art models. Our final model even outperforms human-generated articles in \\(30\\%\\) of cases, showing the strong potential of our approach." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.675, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Contributions: 1) We introduce the novel task of Illustrated Instructions, which requires generating a sequence of images and text that together describe how to achieve a goal (\\(\\S 3\\) and \\(\\S\\S 5.1\\) and 5.2), along with desiderata and metrics for this task; 2) We propose a new approach StackedDiffusion for Illustrated Instructions, with novel modifications to the modeling procedure, enabling generation of visuals suitable for instructional articles for the first time without any additional parameters (\\(\\S 4\\)); 3) We show that our proposed method achieves strong performance on all metrics, and confirm that human evaluators prefer it over existing methods by wide margins—even surpassing ground truth images in some cases (\\(\\S\\S 5.3\\) to 5.5); 4) Finally, we showcase new abilities that StackedDiffusion unlocks, including personalization, goal suggestion, and error correction, that go" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.856, + 0.108 + ], + "angle": 0, + "content": "far beyond what is possible with fixed articles (§ 5.6)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.122, + 0.642, + 0.138 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.148, + 0.895, + 0.39 + ], + "angle": 0, + "content": "Instructional data, tasks, and methods. In the text domain, learning language models on WikiHow [30, 72], has led to advances in tasks such as summarization [71], commonsense procedural knowledge [72, 75], question answering [11], and hierarchical reasoning [74]. In particular, Zhang et al. [72] introduce the goal inference task, in which a model is presented a goal text as input and asked which of 4 candidate steps is one that actually helps achieve that goal, as well as the analogous step inference task. In the multimodal setting, Yang et al. [67] introduce the Visual Goal-Step Inference (VGSI) dataset and task, which consider articles of interleaved text and images. In this task, a model is again presented a goal text as input but is asked which of 4 candidate images is one that actually helps achieve that goal. They show representations learned on this data aid in tasks related to instructional videos [39, 58]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.391, + 0.895, + 0.527 + ], + "angle": 0, + "content": "Learning representations from multimodal data. Multimodal data (including multimodal instructional data, such as from video [39, 58]) has proven to be a powerful source of signal for tasks such as zero-shot recognition [25, 47], text-image and text-video retrieval [3, 9, 16, 17, 34, 36, 37, 40, 60], temporal segmentation [36, 46, 56, 65, 76], activity localization [9, 33, 36, 63, 65, 77], anticipation [12, 15, 19, 52, 61], question-answering [31, 53, 64], summarization [41, 42], and even recipe personalization [14]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.527, + 0.895, + 0.587 + ], + "angle": 0, + "content": "Existing work on instructional data has centered around understanding, rather than generation. We instead focus on the novel setting of generating full multimodal articles complete with text and illustrations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.589, + 0.896, + 0.801 + ], + "angle": 0, + "content": "Generative models. Recent work has examined text-conditioned visual generation through autoregressive [8, 69] or diffusion models [4, 18, 21, 43, 48, 49, 51]. These advances have been leveraged to create text and images together with purely autoregressive [1, 70] or combined approaches [27]. Concurrent work [2, 57] enables the capability of generating multiple images together in the autoregressive framework, but requires substantial additional parameters and focus on creating images that adhere solely to the nearby text rather than enabling consistency. (Similar issues arise in the text-to-video setting; see Appendix 7.) StackedDiffusion, on the other hand, leverages the priors built into the T2I model to obtain consistency without any additional parameters." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.815, + 0.72, + 0.831 + ], + "angle": 0, + "content": "3. Illustrated Instructions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We now formalize our task and the corresponding desiderata. The input to the system is a goal text \\(\\mathbf{g}\\). As output, we would like to produce step text \\(\\mathbf{s}_i\\) as well as step illustration \\(\\mathbf{I}_i\\) for each step \\(i\\). The step text should be a natural language" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "6275" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.468, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.329, + 0.47, + 0.399 + ], + "angle": 0, + "content": "Figure 2. Failure modes of a naive approach. A frozen T2I model is not able to capture both the goal and the step, showing only one or the other depending on how it is prompted. Further, it can not produce consistent images, leading to odd changes such as the color of the ice varying between images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.415, + 0.468, + 0.445 + ], + "angle": 0, + "content": "description of the step, while the step illustrations should be an image corresponding to the step." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.446, + 0.468, + 0.61 + ], + "angle": 0, + "content": "On first glance, one might think that this task is a straightforward amalgamation of textual instruction generation and text to image generation. While these tasks are closely related and necessary substeps towards illustrated instructions, they are not sufficient. The naive approach—simply creating step text \\(\\mathbf{s}_i\\) from goal text \\(\\mathbf{g}\\), and each image \\(\\mathbf{I}_i\\) individually from each \\(\\mathbf{s}_i\\)—fails to recognize the fresh new challenges that do not exist in either of these two other tasks. We identify three key requirements for useful illustrated illustrations: goal faithfulness, step faithfulness, and cross-image consistency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.613, + 0.468, + 0.673 + ], + "angle": 0, + "content": "The first requirement is clear: if the images do not relate to the goal, they cannot be good illustrations. This motivates the desiderata of goal faithfulness, which requires that each image faithfully reflects the goal text." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.468, + 0.794 + ], + "angle": 0, + "content": "However, goal faithfulness alone is not enough. Consider the first row of images in Figure 2. The images all reflect the ultimate goal—creating colored ice—but fail in their role as step illustrations. This in turn motivates step faithfulness, which requires that each image be faithful to the step text. We see in this example that baseline T2I models fail dramatically along this metric; every image reflects the goal rather than the specifics of the step requested." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Finally, the generated images should be consistent with each other. While the second row of images in Figure 2 are all faithful to the step text, the color of the ice (and even the style of the images, cartoon or real) changes between images. This is jarring and confusing to the reader. This motivates the final criterion of cross-image consistency, which requires that each image be consistent with the other images" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.095, + 0.888, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.263, + 0.892, + 0.333 + ], + "angle": 0, + "content": "Figure 3. Overview of StackedDiffusion. At training time, we use the given goal and step text, and stack the encoded ground truth step-images. At inference time, we obtain the goal and step text from an LLM, and unstack denoised latents to produce the output images. See §4 for details and notation." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.347, + 0.742, + 0.362 + ], + "angle": 0, + "content": "produced for a particular generation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.375, + 0.667, + 0.39 + ], + "angle": 0, + "content": "4. StackedDiffusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.4, + 0.892, + 0.655 + ], + "angle": 0, + "content": "We propose a new architecture, StackedDiffusion, to overcome the limitations of existing text-to-image (T2I) approaches for the task of generating interleaved text and images for instructional tasks. StackedDiffusion builds upon T2I models based on latent diffusion models (LDMs) [49], which are diffusion models [48] that operate on a low-resolution 'latent' encoding of the original images. Our primary desiderata is that the images generated must be faithful to both the goal and step text. However, in initial experiments we found that simply encoding the goal and step texts joined together (as strings) leads to an uninformative encoding. For instance, generations conditioned on this combined text tended to ignore certain steps or the goal. Furthermore, the length of this combined text will likely exceed the context length limitations of the text encoders [47] commonly used in T2I models, leading to undesired truncation and information loss." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.657, + 0.892, + 0.853 + ], + "angle": 0, + "content": "Hence, we elect to use a more general approach, applicable to text encoders with any context length. Rather than encoding the combined goal and step text to obtain the condition, we first separately encode the goal and step texts and then concatenate the encodings. This allows the model to learn to use the goal and step text independently, as well as in combination. We add a 'step-positional encoding' \\(\\Phi_{i}\\) to this concatenation. It is broadcast across the dimensions pertaining to a particular step, to indicate to the model where a step begins and ends. \\(\\Phi_0\\) denotes the positional encoding reserved to indicate the goal embedding. Hence, given a text encoder \\(\\psi\\), and \\(N\\) steps in a given goal, we compute the overall text conditioning \\(\\mathcal{C}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.863, + 0.892, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} = \\left| \\right| \\left(\\psi (\\mathbf {g}) + \\Phi_ {0}, \\left. \\right| _ {i = 1} ^ {N} \\psi (\\mathbf {s} _ {i}) + \\Phi_ {i}\\right) \\tag {1}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6276" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "where \\( || \\) is the concatenation operation. This is shown in Figure 3 (top). This design decision is critical to obtain good goal and step faithfulness." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.137, + 0.468, + 0.302 + ], + "angle": 0, + "content": "In addition, independently generating each image does not achieve the requirement of cross-image consistency as the model cannot exchange information across images. Thus, we generate all the \\(\\mathbf{I}_i\\) images at once, which allows the model to jointly generate the sequence and achieve cross-image consistency. The remaining question, of course, is how to accomplish this simultaneous generation such that it gives rise to cross-image consistency. Our key observation is that T2I models already have a strong prior for consistency within a single image. Could we make use of this previously-learned knowledge?" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.303, + 0.468, + 0.395 + ], + "angle": 0, + "content": "We propose a simple method to accomplish this: spatial tiling. As illustrated in Figure 3, the denoising U-Net is given latents \\(\\mathbf{z}_i\\) corresponding to each of the output images simultaneously, tiled spatially as if a single image. At training time, training images are encoded into the latent space as usual, then reshaped into the tiled format. In detail," + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.402, + 0.468, + 0.442 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {i} = \\mathcal {E} \\left(\\mathbf {I} _ {i}\\right), \\quad \\mathbf {z} = \\left. \\right| _ {i = 1} ^ {N} \\mathbf {z} _ {i} \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.446, + 0.468, + 0.472 + ], + "angle": 0, + "content": "\\[\nL _ {L D M} := \\mathbb {E} _ {\\mathbf {z}, \\epsilon \\sim \\mathcal {N} (0, 1), t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {z} ^ {(t)}, t, \\mathcal {C}) \\| _ {2} ^ {2} \\right] \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.479, + 0.468, + 0.718 + ], + "angle": 0, + "content": "where \\(\\mathcal{E}\\) is the encoder to map the image to the latent space (for instance, using a VAE [26]), \\(L_{LDM}\\) is the training objective for the LDM, \\(t\\) denotes a timestep of the diffusion process, \\(\\epsilon\\) denotes the noise added at a given timestep, and \\(\\theta\\) denotes the parameters of the learned denoising U-Net. At inference, we use classifier-free guidance [43] to generate a latent \\(\\tilde{\\mathbf{z}}\\) given conditioning goal and \\(N\\) step texts. We split the latent into \\(\\tilde{\\mathbf{z}}_1,\\dots,\\tilde{\\mathbf{z}}_N\\), and decode into generated images using a decoder \\(\\mathcal{D}\\) corresponding to the encoder \\(\\mathcal{E}\\), i.e. \\(\\tilde{\\mathbf{I}}_i = \\mathcal{D}(\\tilde{\\mathbf{z}}_i)\\). (We opt for tiling along a single spatial dimension for simplicity, and find that alternative tiling strategies do not significantly affect performance; see Appendix § 8.) We refer the reader to the Appendix § 9 and [49] for further details on LDMs. Given the stacked conditioning and generation operations, we refer to our final model as Stacked-Diffusion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.468, + 0.901 + ], + "angle": 0, + "content": "During training, goal and step text are obtained from ground truth, while at inference, a pretrained LLM is used to transform arbitrary user input text to an inferred goal and generated step texts. We describe the LLM inference procedure and prompt engineering in the Appendix § 10. We initialize the U-Net using a pretrained T2I model, and fine-tune all layers when training with the stacked input. Our stacked conditioning only increases the spatial resolution of the input to the U-Net for which can be modeled entirely using the existing parameters (spatial convolutions, attention) of the U-Net. We use the T2I model's text encoder \\((\\psi)\\), kept frozen, to encode the goal and step texts." + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.088, + 0.882, + 0.185 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.197, + 0.892, + 0.24 + ], + "angle": 0, + "content": "Figure 4. Illustrated Instructions data. The histogram shows the distribution of step counts in the data. We find that more than \\(80\\%\\) of articles consist of 6 or fewer steps." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.263, + 0.892, + 0.489 + ], + "angle": 0, + "content": "Since the spatially stacked latent has a spatial resolution comparable to the ones used for high resolution image generation, we encounter issues observed by prior work [32] for high resolution training. Specifically, [32] finds that high resolution latents retain substantial information about the input being noised with typical noise addition schedules, even at the final diffusion timestep \\( t = T \\). This results in an unintentional distribution shift between train time (when all observed inputs contain signal) and test time (when the first inputs are pure noise with no signal). We address this by adjusting the training diffusion noise schedule such that the signal-to-noise ratio (SNR) at the final diffusion timestep \\( T \\) is zero [32]. This ensures enough noise is added during training so as to mitigate this difference between train and test time usage." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.494, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. We build upon a text-to-image latent diffusion model [49] trained on a large in-house image-text dataset. It leverages a VAE [26] to map the images to a 4D latent space, with a \\(8 \\times\\) reduction in spatial resolution. The U-Net largely follows the implementation from [49], with minor architectural modifications. As conditioning, it uses the Flan-T5-XXL text encoder [10]. We train Stacked-Diffusion for 14000 steps using the AdamW optimizer [35] with a learning rate of \\(10^{-4}\\), weight decay of 0.01, and gradient clipping at \\(\\ell_2\\) norm of 1.0. For classifier-free guidance, we use a conditioning dropout rate of 0.05. We choose to generate at most \\(N = 6\\) images simultaneously, because as we will see in § 5.1, the vast majority of the data available has 6 or fewer steps. For shorter training sequences, we pad with empty frames and dummy step texts, and for longer, we drop the extra steps and images. We ablate this choice of \\(N\\) in § 5.4. Hence in practice, for batch of 8 sets of 6 steps at 256px resolution, we get an encoded latent of \\(8 \\times 6 \\times 4 \\times 32 \\times 32\\). We concatenate the latents spatially into \\(8 \\times 4 \\times (6 * 32) \\times 32\\). The denoising U-Net is finetuned to denoise these stacked images, leveraging what it has previously learned about spatial consistency, and adapts to this new setting. At inference time, steps are generated using a pretrained LLM [44], prompted to generate at most \\(N\\) steps (\\(\\S 10\\)). To generate illustrations, noise is sampled in the shape of the tiled latent with \\(N\\) steps, and the steps generated by the LLM are padded with dummy steps if less" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6277" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.097, + 0.093, + 0.193, + 0.105 + ], + "angle": 0, + "content": "Goal Faithfulness" + }, + { + "type": "header", + "bbox": [ + 0.225, + 0.094, + 0.321, + 0.106 + ], + "angle": 0, + "content": "Step Faithfulness" + }, + { + "type": "header", + "bbox": [ + 0.331, + 0.094, + 0.468, + 0.105 + ], + "angle": 0, + "content": "Cross-Image Consistency" + }, + { + "type": "header", + "bbox": [ + 0.088, + 0.116, + 0.203, + 0.128 + ], + "angle": 0, + "content": "Goal: Baking Muffins" + }, + { + "type": "header", + "bbox": [ + 0.212, + 0.117, + 0.335, + 0.14 + ], + "angle": 0, + "content": "Step: Put the muffins in the oven" + }, + { + "type": "header", + "bbox": [ + 0.346, + 0.117, + 0.392, + 0.14 + ], + "angle": 0, + "content": "Previous Image:" + }, + { + "type": "header", + "bbox": [ + 0.402, + 0.106, + 0.468, + 0.16 + ], + "angle": 0, + "content": "#" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.159, + 0.21, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.21, + 0.159, + 0.338, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.16, + 0.467, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.26, + 0.208, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.21, + 0.26, + 0.338, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.26, + 0.467, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.372, + 0.47, + 0.442 + ], + "angle": 0, + "content": "Figure 5. Metrics. We introduce three metrics, one for each desideratum presented in § 3. Goal faithfulness: the second image does not show muffins. Step faithfulness: the second image does not show the step. Cross-image consistency: the second image shows a different number of meatballs with different visuals." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.457, + 0.47, + 0.504 + ], + "angle": 0, + "content": "than \\(N\\). The resulting noise is denoised using the U-Net as usual, and finally reshaped to obtain the \\(N\\) output images upon decoding. See Appendix § 15 for further details." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.519, + 0.21, + 0.536 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.546, + 0.469, + 0.652 + ], + "angle": 0, + "content": "We train and evaluate StackedDiffusion on web-based instructional data. We compare the generations to multiple established baselines and ground truth, using automatic metrics and human evaluations. We now describe the data, metrics, baselines, and the key results. Finally, we demonstrate some new applications that StackedDiffusion enables, showing how it goes beyond standard instructional articles." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.664, + 0.355, + 0.68 + ], + "angle": 0, + "content": "5.1. Illustrated Instructions Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.688, + 0.47, + 0.824 + ], + "angle": 0, + "content": "We introduce a new dataset to train and evaluate models for the Illustrated Instructions task, by repurposing the Visual Goal-Step Inference (VGSI) dataset [67]. VGSI consists of WikiHow articles, each of which have a high-level goal, 6-10 natural language steps (see Figure 4), and associated image illustrations. We observe that the same data can be used for generating instructional articles. It can provide signal for how the goal and step texts should map to output images, and how images should match with each other." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.47, + 0.901 + ], + "angle": 0, + "content": "For evaluation, we construct a held-out set from this same data. Used directly, however, the data is not well suited as many of the tasks are too high-level to be useful for consistent illustration. For instance, \"How to Start a Business in North Carolina\" may have steps \"Brainstorm" + }, + { + "type": "header", + "bbox": [ + 0.618, + 0.089, + 0.893, + 0.104 + ], + "angle": 0, + "content": "Human \\((\\uparrow)\\) GF \\((\\uparrow)\\) SF \\((\\uparrow)\\) CIC \\((\\downarrow)\\) FID \\((\\downarrow)\\)" + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.106, + 0.896, + 0.173 + ], + "angle": 0, + "content": "
T2I (Frozen)22.092.943.251.369.3
T2I (Finetuned)33.378.852.451.553.5
StackedDiffusion(ref)74.361.550.739.5
Ground Truth82.581.773.750.6(N/A)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.185, + 0.894, + 0.241 + ], + "angle": 0, + "content": "Table 1. Comparison to baselines. Human evaluation is reported as win rate vs our full StackedDiffusion model. GF corresponds to goal faithfulness accuracy, SF to step faithfulness accuracy, CIC to cross-image consistency, and FID to Fréchet Inception Distance." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.269, + 0.893, + 0.346 + ], + "angle": 0, + "content": "ideas” and “Go to the courthouse” that have no shared visual content. As such, we filter the data for evaluation to the “Recipes” category, which always has a visually clear end state and where illustrations must be consistent with each other. Please see Appendix § 11 for more details." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.359, + 0.596, + 0.373 + ], + "angle": 0, + "content": "5.2. Metrics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.383, + 0.893, + 0.534 + ], + "angle": 0, + "content": "Having established three desiderata in § 3, we now turn to the question of how to evaluate them. In addition to evaluating the quality of images using FID [22], we propose three metrics, one for each desideratum, that can be used to evaluate the faithfulness and consistency of the generated article. Finally, given the limitations of automatic metrics for evaluating generative modeling tasks [55], we use human evaluations as our primary metric for overall quality. We illustrate the metrics in Figure 5 and briefly describe them next. Please see the Appendix § 12 for more details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.536, + 0.894, + 0.749 + ], + "angle": 0, + "content": "Goal Faithfulness (GF) measures how well the generated image is associated to the goal text. We evaluate this by constructing multiple-choice questions (MCQ) as in VGSI [67]. For each generated image, we compare its CLIP similarity [47] with the correct goal text vs the similarity with the texts of three other randomly selected goals. We compute the accuracy of the model in choosing the correct goal text. Step Faithfulness (SF) measures how faithfully the generated image illustrates the step it aims to depict. An image should match the text for the step it was made for, more than other steps. We measure this using CLIP similarity and a MCQ task similar to goal faithfulness, where the image should have higher CLIP similarity with the corresponding step text than the other step texts within the same goal." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Cross-Image Consistency (CIC) evaluates how consistent the generated images for a goal are with each other and penalizes jarring inconsistencies across the images, such as objects changing color or number. For instance, if a particular set of ingredients are shown for \"gather the ingredients,\" then the same ingredients should be shown for \"mix the dry ingredients.\" We measure this by computing the average \\(\\ell_2\\) distance between DINO [7] embeddings of the images for each step, as this considers the similarity of visual (rather than purely semantic) features. Like prior work [50], we" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6278" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.182, + 0.09, + 0.367, + 0.101 + ], + "angle": 0, + "content": "Goal: How to Bake Turkey Wings" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.101, + 0.468, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.301, + 0.47, + 0.371 + ], + "angle": 0, + "content": "Figure 6. Baseline comparison. StackedDiffusion images are preferred overwhelmingly over baselines. The frozen baseline tends to only produce images showing the goal, while the finetuned baseline produces images that are more faithful to the step text, but have no visual features in common." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.385, + 0.469, + 0.476 + ], + "angle": 0, + "content": "also found that CLIP features do not work well for image similarity as they are invariant to critical aspects such as color, number of objects, style etc. DINO's self-supervised pretraining objective leads to features that are sensitive to the visual aspects particular to an image, rather than being invariant to aspects not captured by category or language." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.476, + 0.47, + 0.626 + ], + "angle": 0, + "content": "Human Evaluation is finally used to confirm that the automated evaluation corresponds to actual human preferences. We provide fully rendered articles from each model to human evaluators on Amazon Mechanical Turk (AMT), and ask them to choose which article they prefer, or if they are tied. We then consider the win rate as the proportion for which a majority of evaluators selected the given method compared to the articles produced by StackedDiffusion. A win rate of 50 denotes the methods are tied, whereas \\(< 50\\) denotes the method is worse than StackedDiffusion." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.636, + 0.185, + 0.65 + ], + "angle": 0, + "content": "5.3. Baselines" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.469, + 0.703 + ], + "angle": 0, + "content": "We now describe some baseline approaches based on existing state-of-the-art image generation models, and compare them to StackedDiffusion in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.469, + 0.825 + ], + "angle": 0, + "content": "T2I (frozen): The most obvious choice for a baseline is to simply use a pretrained and frozen text-to-image (T2I) model. We use our in-house T2I model that is also used to initialize StackedDiffusion. To ensure this baseline gets same information as StackedDiffusion, we prompt the model with a concatenation of the goal with each step, \\(\\mathbf{g}||\\mathbf{s}_i\\), for each \\(i\\), and produce a single image \\(\\mathbf{I}_i\\). The sequence then is composed of \\(N\\) independent generations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We find that these images are faithful to the goal text, but fail to be faithful to the step text and lack consistency. The goal faithfulness is substantially higher than even the ground truth (92.9% vs 81.7%). This is because the model is not capable of reasoning about what a particular step to-" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.088, + 0.88, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.23, + 0.892, + 0.286 + ], + "angle": 0, + "content": "Figure 7. Effect of training data. We find that having more training data improves StackedDiffusion's faithfulness and consistency. However even with \\(20\\%\\) data it performs well, thanks to its effective use of pretrained T2I model weights." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.303, + 0.892, + 0.393 + ], + "angle": 0, + "content": "wards a goal should look like. Instead, it simply generates an image that matches words in the input text, as illustrated in the first row of Figure 2. This results in all the images for e.g. \"Make Colored Ice\" showing the actual finished ice rather than any of the intermediate steps. In other words, the step text is eclipsed by the goal text." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.394, + 0.893, + 0.561 + ], + "angle": 0, + "content": "T2I (finetuned): We finetune a T2I model with the goal and step text embeddings concatenated together as the condition, again producing \\( N \\) independent generations. We find that this model is more faithful to the step text than the frozen T2I model, but still substantially less than our final model (52.4% vs 61.5%). This suggests the goal still overshadows the step text, but to a lesser extent. The goal faithfulness is also lower than for the frozen T2I model, as this model is trained to create images for steps rather than goals. The cross-image consistency is low, similar to the frozen model, as the images are still generated independently." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.561, + 0.893, + 0.834 + ], + "angle": 0, + "content": "StackedDiffusion achieves substantially improved step faithfulness and cross-image consistency, while maintaining high goal faithfulness. We show an illustrative example in Figure 6. Unlike the baselines, the cross-image consistency is very close to that of the ground truth data; as generating them together allows for influence between them during the generation process. We also find that the FID of our generated images is substantially lower than that of other models, indicating that our generated images are more similar to the ground truth data. Finally on human evaluations, StackedDiffusion clearly outperforms them both by a preference of \\(78\\%\\) and \\(66.6\\%\\) respectively. Perhaps most notably, when compared to the ground truth images, human evaluators still picked StackedDiffusion \\(18.5\\%\\) of the times, suggesting that our model generates some illustrations even better than the ground truth data. Further baselines can be found in Appendix § 13, with a comparison of training and inference costs in Appendix § 14." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.846, + 0.611, + 0.861 + ], + "angle": 0, + "content": "5.4. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Step Count. As discussed in § 4, we elect to use a maximum step count of 6 as it allows sufficient data coverage" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6279" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.079, + 0.089, + 0.47, + 0.16 + ], + "angle": 0, + "content": "
Win RateWin Rate
T2I (Frozen)33.8T2I (Finetuned)34.0
GILL [27]++3.9LLM + CLIP15.6
Goal Retrieval33.1Ground Truth70.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.17, + 0.47, + 0.227 + ], + "angle": 0, + "content": "Table 2. System-level comparison to prior work. Human evaluation results using fully generated outputs. Reported as human evaluation win rate of the associated method over StackedDiffusion (hence, higher is better for the reported method)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.236, + 0.471, + 0.327 + ], + "angle": 0, + "content": "and most articles fall under this value. We find that human evaluators prefer this model to the shorter step count model of length 4. It is slightly preferred over the model trained with a longer step count of 8, likely as there is insufficient data of long lengths, limiting any advantage that might be gained in longer length generations." + }, + { + "type": "table", + "bbox": [ + 0.148, + 0.336, + 0.401, + 0.376 + ], + "angle": 0, + "content": "
Baseline4 Step (↑)8 Step (↑)
6 Step32.146.9
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.384, + 0.47, + 0.475 + ], + "angle": 0, + "content": "Training data. In Figure 7, we randomly sub-sample varying proportions of the total data, and train StackedDiffusion on each of those subsets. As the amount of training data increases, it generates more faithful and consistent images. However even with less data it performs well, owing to its effective use of the pretrained T2I model's initialization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.476, + 0.47, + 0.536 + ], + "angle": 0, + "content": "Importance of 0SNR. We experiment with the 0SNR technique [32] due to the high spatial dimensionality of the latents. We find that not using 0SNR results in substantially reduced step faithfulness, dropping from \\(61.5\\%\\) to \\(52.8\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.536, + 0.47, + 0.627 + ], + "angle": 0, + "content": "Importance of Step-Positional Encoding. We also evaluate the importance of the step-positional encoding, \\(\\Phi_{i}\\). This gives the model a critical cue for which parts of the condition correspond to which steps. We find that not using the step-positional encoding results in substantially reduced step faithfulness as well, dropping from \\(61.5\\%\\) to \\(49.8\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.636, + 0.317, + 0.652 + ], + "angle": 0, + "content": "5.5. Comparison to prior work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.47, + 0.825 + ], + "angle": 0, + "content": "The previous metrics were computed with respect to fixed ground truth goal and step texts to provide direct comparisons. However, they do not encompass the full scope of our model's capabilities: the flexibility of the generated text is one of StackedDiffusion's core strengths. To evaluate in a closer setting to this real-world usage, we examine the quality of the articles generated by the full system, where the steps are generated with a LLM [44]. We perform system-level comparisons using human evaluators, including the quality of the generated text and the quality of the images created from the generated text." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.901 + ], + "angle": 0, + "content": "We show our results in Table 2. We first compare to the baselines introduced in § 5.3, and see results similar to Table 1. We see that even with LLM-generated text, Stacked-Diffusion outperforms baseline T2I based approaches, either frozen or finetuned, by more than \\(66\\%\\) win-rate." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.091, + 0.892, + 0.129 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.507, + 0.133, + 0.864, + 0.158 + ], + "angle": 0, + "content": "Step 1: Drain and press tofu to remove excess water. Step 2: Crumble tofu; mix with bread crumbs and a binding agent. Step 3: Shape tofu mixture into burger patties. Step 4: Cook over medium heat until golden brown." + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.16, + 0.6, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.605, + 0.16, + 0.696, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.16, + 0.793, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.795, + 0.16, + 0.886, + 0.184 + ], + "angle": 0, + "content": "\\( \\left( {x - {2x}}\\right) t - x{y}^{2} = \\left( {x - {2x}}\\right) {f}^{\\prime }t \\)" + }, + { + "type": "image", + "bbox": [ + 0.795, + 0.16, + 0.886, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.244, + 0.892, + 0.272 + ], + "angle": 0, + "content": "Figure 8. Personalization. StackedDiffusion enables instructions personalized far beyond what is possible with preexisting articles." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.293, + 0.893, + 0.641 + ], + "angle": 0, + "content": "Additionally, we compare StackedDiffusion to recent state-of-the-art approaches in multimodal generation, as in principle those can also produce instructional text combined with illustrations. Specifically, we compare to GILL [29]-- a model trained to generate sequences of interleaved text and images. While other similar approaches have been proposed [1, 28], GILL is the best open-source model we could access. Using GILL directly, however, results in a human evaluation win rate of \\(0\\%\\). We posit this was for three primary reasons. First, the GILL model was unable to generate long enough generations to suffice as an article in comparison to our generations, despite our best efforts in modifying the generation parameters. Secondly, the model often did not produce any illustrations with the text, although we tried various prompting tactics. Thirdly, the text or images produced were often not of high enough quality to actually reflect the goal or steps. We thus introduce an alternative, which we call GILL++. This uses the (superior) text produced by the same language model as in our method, but passes each step text as input to GILL to generate the corresponding image. This results in a human evaluation win rate of \\(3.9\\%\\) against StackedDiffusion. See Appendix § 16 for details on prompting GILL, and our GILL++ baseline." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Next, we compare StackedDiffusion to a retrieval based approach, denoted by LLM+CLIP. Here we retrieve the closest image in the training set according to CLIP similarity to each of the steps created by the LLM (with each concatenated to the goal text, similar to how the frozen T2I model is used). This retrieval-based metric is not suitable for the automated metrics previously introduced as it is based on the same underlying similarity scores as the automated metrics, but we introduce it here for human evaluation. We find that StackedDiffusion is preferred overwhelmingly over this retrieval-based approach, despite the retrieval-based approach using real images from the training set. Similarly, we compare to another retrieval-based approach which retrieves the full article with the most similar goal text to the input goal, denoted Goal Retrieval. We find StackedDiffusion strongly outperforms this as well. These results together suggest that StackedDiffusion is able" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6280" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.092, + 0.111, + 0.115 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.119, + 0.094, + 0.448, + 0.115 + ], + "angle": 0, + "content": "I have a bunch of apples left from apple picking, what can I do with them?" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.119, + 0.411, + 0.141 + ], + "angle": 0, + "content": "If you have a lot of apples, you can turn them into a delicious apple crisp. Here's a simple recipe to get you started:" + }, + { + "type": "image", + "bbox": [ + 0.437, + 0.117, + 0.465, + 0.142 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.147, + 0.199, + 0.17 + ], + "angle": 0, + "content": "Step 1: Preheat oven to \\(350^{\\circ}\\mathrm{F}\\) \\((175^{\\circ}\\mathrm{C})\\)" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.328, + 0.18 + ], + "angle": 0, + "content": "Step 2: Peel, core, and slice apples; place in a baking dish." + }, + { + "type": "text", + "bbox": [ + 0.345, + 0.147, + 0.46, + 0.189 + ], + "angle": 0, + "content": "Step 3: Mix sugar, cinnamon, and a pinch of salt; sprinkle over apples." + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.189, + 0.208, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.285, + 0.205, + 0.326 + ], + "angle": 0, + "content": "Step 4: Combine oats, flour, brown sugar, and butter; crumble over apples." + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.189, + 0.335, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.285, + 0.331, + 0.315 + ], + "angle": 0, + "content": "Step 5: Bake for 45 minutes or until golden brown." + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.189, + 0.464, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.346, + 0.285, + 0.452, + 0.296 + ], + "angle": 0, + "content": "Step 6: Serve warm." + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.327, + 0.208, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.327, + 0.336, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.327, + 0.464, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.433, + 0.47, + 0.463 + ], + "angle": 0, + "content": "Figure 9. Goal suggestion and inference. StackedDiffusion can suggest what goal might be most relevant given other information." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.476, + 0.469, + 0.506 + ], + "angle": 0, + "content": "to generate novel instructional articles that one can not simply retrieve from the training corpus." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.506, + 0.469, + 0.656 + ], + "angle": 0, + "content": "Finally, we compare StackedDiffusion to the ground truth. Surprisingly again, we find that the human annotators pick our generations \\(30\\%\\) of times even compared to ground truth. Note that these articles are handwritten and manually illustrated for the purpose of illustrating these goals. This shows the strong promise of our proposed approach. We believe with the rapid improvements in training data and generative modeling techniques, future iterations of our model could be indistinguishable or even better than manually created illustrated instructional articles." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.666, + 0.212, + 0.683 + ], + "angle": 0, + "content": "5.6. Applications" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.469, + 0.78 + ], + "angle": 0, + "content": "Perhaps the biggest strength of a generative approach is its ability to handle unique user queries that may not fit into the boundaries posed by static articles on the web. Hence, we demonstrate the capabilities of the full system, combining the strengths of text generation capabilities from the language model with StackedDiffusion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.469, + 0.87 + ], + "angle": 0, + "content": "Personalized instruction. A user can provide any situational information specific to their circumstances and obtain an article customized to that situation. This is not possible with fixed, existing articles. For example, in Figure 8, the user can specify a diet and obtain an article for the food they want that adheres to that diet." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Goal suggestion. As StackedDiffusion accepts flexible input text, a user can provide higher-level information about" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.092, + 0.536, + 0.114 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.542, + 0.092, + 0.88, + 0.114 + ], + "angle": 0, + "content": "Oh no! I'm making bread and forgot to add salt when I made the dough! I'm at the kneading step." + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.119, + 0.84, + 0.141 + ], + "angle": 0, + "content": "If you're at the kneading step, don't worry—you can still incorporate it into the dough with more kneading. Here's how:" + }, + { + "type": "image", + "bbox": [ + 0.86, + 0.117, + 0.888, + 0.142 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.147, + 0.624, + 0.18 + ], + "angle": 0, + "content": "Step 1: Determine the amount of salt needed." + }, + { + "type": "text", + "bbox": [ + 0.639, + 0.147, + 0.75, + 0.18 + ], + "angle": 0, + "content": "Step 2: Dissolve the salt in a small amount of warm water." + }, + { + "type": "text", + "bbox": [ + 0.767, + 0.147, + 0.882, + 0.17 + ], + "angle": 0, + "content": "Step 3: Flatten the dough into a rectangle." + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.188, + 0.63, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.511, + 0.285, + 0.611, + 0.307 + ], + "angle": 0, + "content": "Step 4: Spread the salt over the dough." + }, + { + "type": "image", + "bbox": [ + 0.637, + 0.188, + 0.758, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.639, + 0.285, + 0.737, + 0.316 + ], + "angle": 0, + "content": "Step 5: Fold and knead the dough to distribute." + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.189, + 0.886, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.767, + 0.285, + 0.874, + 0.316 + ], + "angle": 0, + "content": "Step 6: Proceed with the original bread recipe." + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.326, + 0.63, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.326, + 0.758, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.326, + 0.886, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.434, + 0.892, + 0.462 + ], + "angle": 0, + "content": "Figure 10. Error correction. StackedDiffusion provides updated instructions in response to unexpected situations, like a user error." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.483, + 0.892, + 0.559 + ], + "angle": 0, + "content": "what they want to do and obtain an article with a suggested specific goal that matches what the user would like to do. In Figure 9, the user describes their situation (having many apples) and StackedDiffusion suggests a goal that matches their situation (making apple crisps)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.562, + 0.892, + 0.698 + ], + "angle": 0, + "content": "Error correction. Oftentimes, a user in the course of performing a task may make a mistake. Fixed articles provide no avenue for recourse. However, StackedDiffusion can adapt to user error and create alternative instructions that best correct for and accommodate the user's situation. See Figure 10 for an example of this. A fixed article would not provide any alternative paths that adapt as the user performs the action. This highlights the new avenues opened up by our generative approach to this task." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.721, + 0.766, + 0.737 + ], + "angle": 0, + "content": "6. Limitations and Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "While StackedDiffusion achieves strong performance in generating Illustrated Instructions, we note that there still remains a gap with ground truth data. Future work might examine how to obtain instructional data at greater scale and how this influences StackedDiffusion. Leveraging the latest T2I architectures to improve the faithfulness and quality of our generations would be another promising avenue for future work. Finally, leveraging improvements in text-to-video techniques to generate video clips describing each step would further improve the usability of such a system." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "6281" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.185 + ], + "angle": 0, + "content": "[1] Armen Aghajanyan, Bernie Huang, Candace Ross, Vladimir Karpukhin, Hu Xu, Naman Goyal, Dmytro Okhonko, Mandar Joshi, Gargi Ghosh, Mike Lewis, and Luke Zettlemoyer. CM3: A causal masked multimodal model of the internet. arXiv preprint arXiv:2201.07520, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.47, + 0.227 + ], + "angle": 0, + "content": "[2] Emanuele Aiello, Lili Yu, Yixin Nie, Armen Aghajanyan, and Barlas Oguz. Jointly training large autoregressive multimodal models. In ICLR, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.231, + 0.47, + 0.272 + ], + "angle": 0, + "content": "[3] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.274, + 0.47, + 0.343 + ], + "angle": 0, + "content": "[4] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.346, + 0.469, + 0.401 + ], + "angle": 0, + "content": "[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your Latents: High-Resolution Video Synthesis with Latent Diffusion Models. In CVPR, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.403, + 0.469, + 0.445 + ], + "angle": 0, + "content": "[6] Russell N. Carney and Joel R. Levin. Pictorial illustrations still improve students' learning from text. Educational Psychology Review, 14(1):5-26, 2002." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.447, + 0.469, + 0.501 + ], + "angle": 0, + "content": "[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging Properties in Self-Supervised Vision Transformers. In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.505, + 0.469, + 0.587 + ], + "angle": 0, + "content": "[8] Huiwen Chang, Han Zhang, Jarred Barber, Aaron Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Patrick Murphy, William T. Freeman, Michael Rubinstein, Yanzhen Li, and Dilip Krishnan. Muse: Text-ToImage Generation via Masked Generative Transformers. In ICML, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.59, + 0.469, + 0.658 + ], + "angle": 0, + "content": "[9] Brian Chen, Andrew Rouditchenko, Kevin Duarte, Hilde Kuehne, Samuel Thomas, Angie Boggust, Rameswar Panda, Brian Kingsbury, Rogerio Feris, David Harwath, et al. Multimodal clustering networks for self-supervised learning from unlabeled videos. In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.469, + 0.813 + ], + "angle": 0, + "content": "[10] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Alex Castro-Ros, Marie Pellat, Kevin Robinson, Dasha Valter, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[11] Yang Deng, Wai Lam, Yuexiang Xie, Daoyuan Chen, Yaliang Li, Min Yang, and Ying Shen. Joint Learning of Answer Selection and Answer Summary Generation in Community Question Answering. In AAAI, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[12] Eadom Dessalene, Chinmaya Devaraj, Michael Maynord, Cornelia Fermuller, and Yiannis Aloimonos. Forecasting ac" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.89, + 0.12 + ], + "angle": 0, + "content": "tion through contact representations from first person video. PAMI, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[13] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In ICCV, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[14] Bahare Fatemi, Quentin Duval, Rohit Girdhar, Michal Drozdzal, and Adriana Romero-Soriano. Learning to substitute ingredients in recipes. arXiv preprint arXiv:2302.07960, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.235, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[15] Antonino Furnari and Giovanni Maria Farinella. Rolling-unrolling lstms for action anticipation from first-person video. PAMI, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.278, + 0.892, + 0.319 + ], + "angle": 0, + "content": "[16] Valentin Gabeur, Chen Sun, Karteek Alahari, and Cordelia Schmid. Multi-modal transformer for video retrieval. In ECCV, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.321, + 0.892, + 0.361 + ], + "angle": 0, + "content": "[17] Valentin Gabeur, Arsha Nagrani, Chen Sun, Karteek Alahari, and Cordelia Schmid. Masking modalities for cross-modal video retrieval. In WACV, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.364, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[18] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. In ECCV, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.421, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[19] Rohit Girdhar and Kristen Grauman. Anticipative video transformer. In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.449, + 0.892, + 0.518 + ], + "angle": 0, + "content": "[20] Rohit Girdhar, Mannat Singh, Andrew Brown, Quentin Duval, Samaneh Azadi, Sai Saketh Rambhatla, Akbar Shah, Xi Yin, Devi Parikh, and Ishan Misra. Emu video: Factorizing text-to-video generation by explicit image conditioning. arXiv preprint arXiv:2311.10709, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.52, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[21] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.577, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[22] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.633, + 0.892, + 0.714 + ], + "angle": 0, + "content": "[23] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen Video: High Definition Video Generation with Diffusion Models. arXiv preprint arXiv:2210.02303, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.717, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[24] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.774, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[25] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.831, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[26] Diederik P. Kingma and Max Welling. Auto-Encoding Variational Bayes. In ICLR, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[27] Jing Yu Koh, Daniel Fried, and Ruslan Salakhutdinov. Generating images with multimodal language models. In NeurIPS, 2023." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6282" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "[28] Jing Yu Koh, Ruslan Salakhutdinov, and Daniel Fried. Grounding language models to images for multimodal inputs and outputs. In ICML, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.469, + 0.178 + ], + "angle": 0, + "content": "[29] Jing Yu Koh, Ruslan Salakhutdinov, and Daniel Fried. Grounding Language Models to Images for Multimodal Generation. In NeurIPS, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.181, + 0.469, + 0.221 + ], + "angle": 0, + "content": "[30] Mahnaz Koupae and William Yang Wang. WikiHow: A Large Scale Text Summarization Dataset. arXiv preprint arXiv:1810.09305, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.224, + 0.469, + 0.279 + ], + "angle": 0, + "content": "[31] Linjie Li, Yen-Chun Chen, Yu Cheng, Zhe Gan, Licheng Yu, and Jingjing Liu. Hero: Hierarchical encoder for video+ language omni-representation pre-training. arXiv preprint arXiv:2005.00200, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.282, + 0.469, + 0.323 + ], + "angle": 0, + "content": "[32] Shanchuan Lin, Bingchen Liu, Jiashi Li, and Xiao Yang. Common diffusion noise schedules and sample steps are flawed. arXiv preprint arXiv:2305.08891, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.326, + 0.469, + 0.379 + ], + "angle": 0, + "content": "[33] Xudong Lin, Fabio Petroni, Gedas Bertasius, Marcus Rohrbach, Shih-Fu Chang, and Lorenzo Torresani. Learning To Recognize Procedural Activities with Distant Supervision. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.383, + 0.469, + 0.436 + ], + "angle": 0, + "content": "[34] Song Liu, Haoqi Fan, Shengsheng Qian, Yiru Chen, Wenkui Ding, and Zhongyuan Wang. Hit: Hierarchical transformer with momentum contrast for video-text retrieval. In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.44, + 0.469, + 0.468 + ], + "angle": 0, + "content": "[35] Ilya Loshchilov and Frank Hutter. Decoupled Weight Decay Regularization. In ICLR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.471, + 0.469, + 0.538 + ], + "angle": 0, + "content": "[36] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon Bharti, and Ming Zhou. Univl: A unified video and language pre-training model for multimodal understanding and generation. arXiv preprint arXiv:2002.06353, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.542, + 0.469, + 0.596 + ], + "angle": 0, + "content": "[37] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval. arXiv preprint arXiv:2104.08860, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.6, + 0.469, + 0.654 + ], + "angle": 0, + "content": "[38] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Ti-niu Tan. Videofusion: Decomposed diffusion models for high-quality video generation. In CVPR, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.657, + 0.469, + 0.724 + ], + "angle": 0, + "content": "[39] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. HowTo100M: Learning a Text-Video Embedding by Watching Hundred Million Narrated Video Clips. In ICCV, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.729, + 0.469, + 0.782 + ], + "angle": 0, + "content": "[40] Antoine Miech, Jean-Baptiste Alayrac, Ivan Laptev, Josef Sivic, and Andrew Zisserman. Thinking fast and slow: Efficient text-to-visual retrieval with transformers. In CVPR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.786, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[41] Medhini Narasimhan, Anna Rohrbach, and Trevor Darrell. Clip-it! language-guided video summarization. NIPS, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.816, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[42] Medhini Narasimhan, Arsha Nagrani, Chen Sun, Michael Rubinstein, Trevor Darrell, Anna Rohrbach, and Cordelia Schmid. Tl; dw? summarizing instructional videos with task relevance and cross-modal saliency. In ECCV, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.874, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[43] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.16 + ], + "angle": 0, + "content": "[44] OpenAI. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.164, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[45] Gaurav Parmar, Richard Zhang, and Jun-Yan Zhu. On aliased resizing and surprising subtleties in gan evaluation. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.206, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[46] AJ Piergiovanni, Anelia Angelova, Michael S Ryoo, and Irfan Essa. Unsupervised action segmentation for instructional videos. arXiv preprint arXiv:2106.03738, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.248, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.317, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[48] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.373, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[49] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bjorn Ommer. High-Resolution Image Synthesis with Latent Diffusion Models. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.415, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[50] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. DreamBooth: Fine Tuning Text-to-Image Diffusion Models for Subject-Driven Generation. In CVPR, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.471, + 0.892, + 0.566 + ], + "angle": 0, + "content": "[51] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding. In NIPS, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.568, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[52] Fadime Sener and Angela Yao. Zero-shot anticipation for instructional activities. In ICCV, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.596, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[53] Paul Hongsuck Seo, Arsha Nagrani, and Cordelia Schmid. Look before you speak: Visually contextualized utterances. In CVPR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.638, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[54] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-A-Video: Text-to-Video Generation without Text-Video Data. arXiv preprint arXiv:2209.14792, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[55] George Stein, Jesse C Cresswell, Rasa Hosseinzadeh, Yi Sui, Brendan Leigh Ross, Valentin Villecloze, Zhaoyan Liu, Anthony L Caterini, J Eric T Taylor, and Gabriel Loiza-Ganem. Exposing flaws of generative model evaluation metrics and their unfair treatment of diffusion models. arXiv preprint arXiv:2306.04675, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.791, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[56] Chen Sun, Fabien Baradel, Kevin Murphy, and Cordelia Schmid. Learning video representations using contrastive bidirectional transformer. arXiv preprint arXiv:1906.05743, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[57] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Emu: Generative pretraining in multimodality. In ICLR, 2024." + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "6283" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[58] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. COIN: A Large-scale Dataset for Comprehensive Instructional Video Analysis. In CVPR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.151, + 0.47, + 0.22 + ], + "angle": 0, + "content": "[59] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable Length Video Generation from Open Domain Textual Descriptions. In ICLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.223, + 0.469, + 0.278 + ], + "angle": 0, + "content": "[60] Wenzhe Wang, Mengdan Zhang, Runnan Chen, Guanyu Cai, Penghao Zhou, Pai Peng, Xiaowei Guo, Jian Wu, and Xing Sun. Dig into multi-modal cues for video retrieval with hierarchical alignment. In *IJCAI*, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.281, + 0.469, + 0.349 + ], + "angle": 0, + "content": "[61] Chao-Yuan Wu, Yanghao Li, Karttikeya Mangalam, Haoqi Fan, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Memvit: Memory-augmented multiscale vision transformer for efficient long-term video recognition. arXiv preprint arXiv:2201.08383, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.353, + 0.469, + 0.421 + ], + "angle": 0, + "content": "[62] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In ICCV, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.424, + 0.469, + 0.48 + ], + "angle": 0, + "content": "[63] Mengmeng Xu, Juan-Manuel Pérez-Rúa, Victor Escorcia, Brais Martinez, Xiatian Zhu, Li Zhang, Bernard Ghanem, and Tao Xiang. Boundary-sensitive pre-training for temporal localization in videos. In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.483, + 0.469, + 0.524 + ], + "angle": 0, + "content": "[64] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Just ask: Learning to answer questions from millions of narrated videos. In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.527, + 0.469, + 0.568 + ], + "angle": 0, + "content": "[65] Jianwei Yang, Yonatan Bisk, and Jianfeng Gao. Taco: Token-aware cascade contrastive learning for video-text alignment. In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.572, + 0.469, + 0.626 + ], + "angle": 0, + "content": "[66] Mengjiao Yang, Yilun Du, Bo Dai, Dale Schuurmans, Joshua B Tenenbaum, and Pieter Abbeel. Probabilistic adaptation of text-to-video models. arXiv preprint arXiv:2306.01872, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.63, + 0.469, + 0.671 + ], + "angle": 0, + "content": "[67] Yue Yang, Artemis Panagopoulou, Qing Lyu, Li Zhang, Mark Yatskar, and Chris Callison-Burch. Visual goal-step inference using wikihow. In EMNLP, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.674, + 0.469, + 0.757 + ], + "angle": 0, + "content": "[68] Shengming Yin, Chenfei Wu, Huan Yang, Jianfeng Wang, Xiaodong Wang, Minheng Ni, Zhengyuan Yang, Linjie Li, Shuguang Liu, Fan Yang, Jianlong Fu, Gong Ming, Lijuan Wang, Zicheng Liu, Houqiang Li, and Nan Duan. NUWAXL: Diffusion over Diffusion for eXtremely Long Video Generation. arXiv preprint arXiv:2303.12346, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.469, + 0.829 + ], + "angle": 0, + "content": "[69] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[70] Lili Yu, Bowen Shi, Ramakanth Pasunuru, Benjamin Muller, Olga Golovneva, Tianlu Wang, Arun Babu, Binh Tang, Brian Karrer, Shelly Sheynin, et al. Scaling autoregressive multimodal models: Pretraining and instruction tuning. arXiv preprint arXiv:2309.02591, 2023." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[71] Jingqing Zhang, Yao Zhao, Mohammad Saleh, and Peter J. Liu. PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization. In ICML, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[72] Li Zhang, Qing Lyu, and Chris Callison-Burch. Reasoning about Goals, Steps, and Temporal Ordering with WikiHow. In EMNLP, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[73] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[74] Shuyan Zhou, Li Zhang, Yue Yang, Qing Lyu, Pengcheng Yin, Chris Callison-Burch, and Graham Neubig. Show Me More Details: Discovering Hierarchies of Procedures from Semi-structured Web Data. In ACL, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.292, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[75] Yilun Zhou, Julie Shah, and Steven Schockaert. Learning Household Task Knowledge from WikiHow Descriptions. In Proceedings of the 5th Workshop on Semantic Deep Learning (SemDeep-5), 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[76] Linchao Zhu and Yi Yang. Actbert: Learning global-local video-text representations. In CVPR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.377, + 0.892, + 0.418 + ], + "angle": 0, + "content": "[77] Dimitri Zhukov, Jean-Baptiste Alayrac, Ivan Laptev, and Josef Sivic. Learning actionness via long-range temporal order verification. In ECCV, 2020." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6284" + } + ] +] \ No newline at end of file diff --git a/2024/Generating Illustrated Instructions/dd71c86a-390f-4c46-b6ae-0a13f2787fa0_origin.pdf b/2024/Generating Illustrated Instructions/dd71c86a-390f-4c46-b6ae-0a13f2787fa0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6e26e855eb2c02a86e0a78a5181178355930b062 --- /dev/null +++ b/2024/Generating Illustrated Instructions/dd71c86a-390f-4c46-b6ae-0a13f2787fa0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:94a2dc8ac985adbc5f7f2039458ab888d713984f48c813885fba0da8b5710c52 +size 3330098 diff --git a/2024/Generating Illustrated Instructions/full.md b/2024/Generating Illustrated Instructions/full.md new file mode 100644 index 0000000000000000000000000000000000000000..deb233ca349866501a2d163029b26d11bd66c226 --- /dev/null +++ b/2024/Generating Illustrated Instructions/full.md @@ -0,0 +1,405 @@ +# Generating Illustrated Instructions + +Sachit Menon $^{1,2*}$ + +Ishan Misra + +Rohit Girdhar1 + +1GenAI, Meta + +2Columbia University + +https://facebookresearch.github.io/IllustratedInstructions + +# Abstract + +We introduce a new task of generating "Illustrated Instructions", i.e. visual instructions customized to a user's needs. We identify desiderata unique to this task, and formalize it through a suite of automatic and human evaluation metrics, designed to measure the validity, consistency, and efficacy of the generations. We combine the power of large language models (LLMs) together with strong text-to-image generation diffusion models to propose a simple approach called StackedDiffusion, which generates such illustrated instructions given text as input. The resulting model strongly outperforms baseline approaches and state-of-the-art multimodal LLMs; and in $30\%$ of cases, users even prefer it to human-generated articles. Most notably, it enables various new and exciting applications far beyond what static articles on the web can provide, such as personalized instructions complete with intermediate steps and pictures in response to a user's individual situation. + +# 1. Introduction + +The internet is a vast resource to find answers to all types of questions. It is often easy to find a webpage or a video that walks through the exact steps to achieve a user's goal. With the rise of Large Language Models (LLMs) trained on internet-scale data, users can just ask the LLM for instructions to achieve a goal. This allows the users to get answers for specific personalized queries for which there may not be an existing webpage on the internet, e.g., modifying cooking recipes with user specific dietary restrictions. Moreover, if the users make a mistake when following the instructions, simple follow-up questions to the LLM can generate alternate instructions, a major advantage over static web search. + +In spite of such advantages, LLMs still have one major limitation - they cannot generate visuals, which are critical for users to learn from and follow instructions for a wide range of tasks [6]. Consider, for instance, instructions that require visual inspection, e.g., a recipe that requires + +![](images/4fe67dc0ae5b1b060518dded246c2b77f59842388aaf6c09d967ab5aad81875d.jpg) + +# How can I bake a cake to celebrate an event in Seattle? + +One great way could be to make a coffee-infused cake to honor Seattle's coffee culture, with teal coloring for the Seattle flag. + +Step 1: Brew strong coffee; let cool. + +Step 2: Preheat the oven to $350\mathrm{F}$ + +Step 3: Prepare cake batter as usual. + +![](images/852682748d3765eb82a702ea3eb732d06dc543bcbfb0a3a2c45dc54dc4fdd451.jpg) + +Step 4: Mix in the coffee; fold in coffee grounds for texture. + +![](images/e19a4df7e34c8a6fff064a96695f89adb558a1d0115a4a06909e77c1e98a91c2.jpg) + +Step 5: Bake for 25- 30 minutes. + +![](images/369dc6ae7426ff6b589954507b0a21efeed223528b8ae9e9a5880f6e5ba09956.jpg) + +Step 6: Cover the cake in teal fondant for the Seattle flag. + +![](images/03ec5dc17b70cd300b497e23cf15e210c38f3eacc6c5e2ca0a6809e7789ec75c.jpg) +Figure 1. StackedDiffusion generating Illustrated Instructions. Given a goal (or any textual user input), StackedDiffusion produces a customized instructional article complete with illustrations that not only tells the user how to achieve the goal in words, but also shows the user by providing illustrations. + +![](images/987a18f0208641696c5ad4ea810c0b7bf3299c2c539484dd7750f5e35378d311.jpg) + +![](images/5821d0db1c60bb551af3ca97b0d29079b4ef19842d877374e0d623776f483002.jpg) + +the user to stir fry onions until golden brown, or searching for bubbles when a flat tire is submerged under water. An image accompanying such instructions can make following them significantly easier. Can we develop methods with the strengths of LLMs that can also generate such visuals? + +In this work, we tackle this challenge, developing models that can not only tell a user how to accomplish their task, but also show them how. We define the novel task of Illustrated Instructions: creating a set of steps with visualizations that showcase an approach to solve the user's task. We carefully consider the different dimensions of the problem, laying out three desiderata unique to this setting. To measure these desiderata, we develop automated metrics based on prior work in instructional article assessment [67] and image generation [50]. + +We propose a new model to solve the task of Illustrated Instructions that combines LLMs with a text-to-image diffusion model. We train our model on stacks of instructional images from websites such as WikiHow. The resulting StackedDiffusion model creates full instructional articles, complete with customized steps and a sequence of images uniquely generated to describe those steps. It leverages large-scale pretrained LLMs and finetuned text-to-image diffusion models, and employs techniques to accomplish the task without the need to introduce any new learnable parameters. This includes spatial tiling for simultaneous multi-image generation, text embedding concatenation to reduce information loss in long instructions, and a new "step-positional encoding" to better demarcate the different steps in the instructions. Thus, StackedDiffusion can generate useful illustrations even when trained even with limited instructional data (cf. Figure 7). + +We compare StackedDiffusion with various baseline approaches based on off-the-shelf tools, and find that they all fall short, even when used in conjunction with one another. Existing T2I models are incapable of generating visuals directly from a user query. Even when given more detailed instructional text, we show that existing T2I models fail to produce images that are simultaneously faithful to the goal, the step, and consistent among each other. Given the recent introduction of multimodal LLMs [1, 28], we also compare with a recent open-source model, GILL [27], and show such models also fall short of consistently generating useful visuals along with text. We posit that our approach of leveraging the spatial priors learned by pretrained diffusion models to generate multiple images together, in conjunction with pretrained LLMs, is significantly more compute and data efficient than an approach purely based on next-token prediction, without these priors. Our thorough ablations and human evaluations show that StackedDiffusion convincingly surpasses state-of-the-art models. Our final model even outperforms human-generated articles in $30\%$ of cases, showing the strong potential of our approach. + +Contributions: 1) We introduce the novel task of Illustrated Instructions, which requires generating a sequence of images and text that together describe how to achieve a goal ( $\S 3$ and $\S\S 5.1$ and 5.2), along with desiderata and metrics for this task; 2) We propose a new approach StackedDiffusion for Illustrated Instructions, with novel modifications to the modeling procedure, enabling generation of visuals suitable for instructional articles for the first time without any additional parameters ( $\S 4$ ); 3) We show that our proposed method achieves strong performance on all metrics, and confirm that human evaluators prefer it over existing methods by wide margins—even surpassing ground truth images in some cases ( $\S\S 5.3$ to 5.5); 4) Finally, we showcase new abilities that StackedDiffusion unlocks, including personalization, goal suggestion, and error correction, that go + +far beyond what is possible with fixed articles (§ 5.6). + +# 2. Related Work + +Instructional data, tasks, and methods. In the text domain, learning language models on WikiHow [30, 72], has led to advances in tasks such as summarization [71], commonsense procedural knowledge [72, 75], question answering [11], and hierarchical reasoning [74]. In particular, Zhang et al. [72] introduce the goal inference task, in which a model is presented a goal text as input and asked which of 4 candidate steps is one that actually helps achieve that goal, as well as the analogous step inference task. In the multimodal setting, Yang et al. [67] introduce the Visual Goal-Step Inference (VGSI) dataset and task, which consider articles of interleaved text and images. In this task, a model is again presented a goal text as input but is asked which of 4 candidate images is one that actually helps achieve that goal. They show representations learned on this data aid in tasks related to instructional videos [39, 58]. + +Learning representations from multimodal data. Multimodal data (including multimodal instructional data, such as from video [39, 58]) has proven to be a powerful source of signal for tasks such as zero-shot recognition [25, 47], text-image and text-video retrieval [3, 9, 16, 17, 34, 36, 37, 40, 60], temporal segmentation [36, 46, 56, 65, 76], activity localization [9, 33, 36, 63, 65, 77], anticipation [12, 15, 19, 52, 61], question-answering [31, 53, 64], summarization [41, 42], and even recipe personalization [14]. + +Existing work on instructional data has centered around understanding, rather than generation. We instead focus on the novel setting of generating full multimodal articles complete with text and illustrations. + +Generative models. Recent work has examined text-conditioned visual generation through autoregressive [8, 69] or diffusion models [4, 18, 21, 43, 48, 49, 51]. These advances have been leveraged to create text and images together with purely autoregressive [1, 70] or combined approaches [27]. Concurrent work [2, 57] enables the capability of generating multiple images together in the autoregressive framework, but requires substantial additional parameters and focus on creating images that adhere solely to the nearby text rather than enabling consistency. (Similar issues arise in the text-to-video setting; see Appendix 7.) StackedDiffusion, on the other hand, leverages the priors built into the T2I model to obtain consistency without any additional parameters. + +# 3. Illustrated Instructions + +We now formalize our task and the corresponding desiderata. The input to the system is a goal text $\mathbf{g}$ . As output, we would like to produce step text $\mathbf{s}_i$ as well as step illustration $\mathbf{I}_i$ for each step $i$ . The step text should be a natural language + +![](images/8bef10bc0808b264215a1b39b4debe4e84955a541d43247c8694557b6a9202d2.jpg) +Figure 2. Failure modes of a naive approach. A frozen T2I model is not able to capture both the goal and the step, showing only one or the other depending on how it is prompted. Further, it can not produce consistent images, leading to odd changes such as the color of the ice varying between images. + +description of the step, while the step illustrations should be an image corresponding to the step. + +On first glance, one might think that this task is a straightforward amalgamation of textual instruction generation and text to image generation. While these tasks are closely related and necessary substeps towards illustrated instructions, they are not sufficient. The naive approach—simply creating step text $\mathbf{s}_i$ from goal text $\mathbf{g}$ , and each image $\mathbf{I}_i$ individually from each $\mathbf{s}_i$ —fails to recognize the fresh new challenges that do not exist in either of these two other tasks. We identify three key requirements for useful illustrated illustrations: goal faithfulness, step faithfulness, and cross-image consistency. + +The first requirement is clear: if the images do not relate to the goal, they cannot be good illustrations. This motivates the desiderata of goal faithfulness, which requires that each image faithfully reflects the goal text. + +However, goal faithfulness alone is not enough. Consider the first row of images in Figure 2. The images all reflect the ultimate goal—creating colored ice—but fail in their role as step illustrations. This in turn motivates step faithfulness, which requires that each image be faithful to the step text. We see in this example that baseline T2I models fail dramatically along this metric; every image reflects the goal rather than the specifics of the step requested. + +Finally, the generated images should be consistent with each other. While the second row of images in Figure 2 are all faithful to the step text, the color of the ice (and even the style of the images, cartoon or real) changes between images. This is jarring and confusing to the reader. This motivates the final criterion of cross-image consistency, which requires that each image be consistent with the other images + +![](images/70e6cc6bb7533343b713e5d6e18f8c0cff12d414b7c8c2d9fb6f0164468e4763.jpg) +Figure 3. Overview of StackedDiffusion. At training time, we use the given goal and step text, and stack the encoded ground truth step-images. At inference time, we obtain the goal and step text from an LLM, and unstack denoised latents to produce the output images. See §4 for details and notation. + +produced for a particular generation. + +# 4. StackedDiffusion + +We propose a new architecture, StackedDiffusion, to overcome the limitations of existing text-to-image (T2I) approaches for the task of generating interleaved text and images for instructional tasks. StackedDiffusion builds upon T2I models based on latent diffusion models (LDMs) [49], which are diffusion models [48] that operate on a low-resolution 'latent' encoding of the original images. Our primary desiderata is that the images generated must be faithful to both the goal and step text. However, in initial experiments we found that simply encoding the goal and step texts joined together (as strings) leads to an uninformative encoding. For instance, generations conditioned on this combined text tended to ignore certain steps or the goal. Furthermore, the length of this combined text will likely exceed the context length limitations of the text encoders [47] commonly used in T2I models, leading to undesired truncation and information loss. + +Hence, we elect to use a more general approach, applicable to text encoders with any context length. Rather than encoding the combined goal and step text to obtain the condition, we first separately encode the goal and step texts and then concatenate the encodings. This allows the model to learn to use the goal and step text independently, as well as in combination. We add a 'step-positional encoding' $\Phi_{i}$ to this concatenation. It is broadcast across the dimensions pertaining to a particular step, to indicate to the model where a step begins and ends. $\Phi_0$ denotes the positional encoding reserved to indicate the goal embedding. Hence, given a text encoder $\psi$ , and $N$ steps in a given goal, we compute the overall text conditioning $\mathcal{C}$ as + +$$ +\mathcal {C} = \left| \right| \left(\psi (\mathbf {g}) + \Phi_ {0}, \left. \right| _ {i = 1} ^ {N} \psi (\mathbf {s} _ {i}) + \Phi_ {i}\right) \tag {1} +$$ + +where $||$ is the concatenation operation. This is shown in Figure 3 (top). This design decision is critical to obtain good goal and step faithfulness. + +In addition, independently generating each image does not achieve the requirement of cross-image consistency as the model cannot exchange information across images. Thus, we generate all the $\mathbf{I}_i$ images at once, which allows the model to jointly generate the sequence and achieve cross-image consistency. The remaining question, of course, is how to accomplish this simultaneous generation such that it gives rise to cross-image consistency. Our key observation is that T2I models already have a strong prior for consistency within a single image. Could we make use of this previously-learned knowledge? + +We propose a simple method to accomplish this: spatial tiling. As illustrated in Figure 3, the denoising U-Net is given latents $\mathbf{z}_i$ corresponding to each of the output images simultaneously, tiled spatially as if a single image. At training time, training images are encoded into the latent space as usual, then reshaped into the tiled format. In detail, + +$$ +\mathbf {z} _ {i} = \mathcal {E} \left(\mathbf {I} _ {i}\right), \quad \mathbf {z} = \left. \right| _ {i = 1} ^ {N} \mathbf {z} _ {i} \tag {2} +$$ + +$$ +L _ {L D M} := \mathbb {E} _ {\mathbf {z}, \epsilon \sim \mathcal {N} (0, 1), t} \left[ \| \epsilon - \epsilon_ {\theta} (\mathbf {z} ^ {(t)}, t, \mathcal {C}) \| _ {2} ^ {2} \right] \tag {3} +$$ + +where $\mathcal{E}$ is the encoder to map the image to the latent space (for instance, using a VAE [26]), $L_{LDM}$ is the training objective for the LDM, $t$ denotes a timestep of the diffusion process, $\epsilon$ denotes the noise added at a given timestep, and $\theta$ denotes the parameters of the learned denoising U-Net. At inference, we use classifier-free guidance [43] to generate a latent $\tilde{\mathbf{z}}$ given conditioning goal and $N$ step texts. We split the latent into $\tilde{\mathbf{z}}_1,\dots,\tilde{\mathbf{z}}_N$ , and decode into generated images using a decoder $\mathcal{D}$ corresponding to the encoder $\mathcal{E}$ , i.e. $\tilde{\mathbf{I}}_i = \mathcal{D}(\tilde{\mathbf{z}}_i)$ . (We opt for tiling along a single spatial dimension for simplicity, and find that alternative tiling strategies do not significantly affect performance; see Appendix § 8.) We refer the reader to the Appendix § 9 and [49] for further details on LDMs. Given the stacked conditioning and generation operations, we refer to our final model as Stacked-Diffusion. + +During training, goal and step text are obtained from ground truth, while at inference, a pretrained LLM is used to transform arbitrary user input text to an inferred goal and generated step texts. We describe the LLM inference procedure and prompt engineering in the Appendix § 10. We initialize the U-Net using a pretrained T2I model, and fine-tune all layers when training with the stacked input. Our stacked conditioning only increases the spatial resolution of the input to the U-Net for which can be modeled entirely using the existing parameters (spatial convolutions, attention) of the U-Net. We use the T2I model's text encoder $(\psi)$ , kept frozen, to encode the goal and step texts. + +![](images/0c73f9e5b449715bef0bcf29490b2d4f9cd7aff6a1231c003decf96a1d2c66f7.jpg) +Figure 4. Illustrated Instructions data. The histogram shows the distribution of step counts in the data. We find that more than $80\%$ of articles consist of 6 or fewer steps. + +Since the spatially stacked latent has a spatial resolution comparable to the ones used for high resolution image generation, we encounter issues observed by prior work [32] for high resolution training. Specifically, [32] finds that high resolution latents retain substantial information about the input being noised with typical noise addition schedules, even at the final diffusion timestep $t = T$ . This results in an unintentional distribution shift between train time (when all observed inputs contain signal) and test time (when the first inputs are pure noise with no signal). We address this by adjusting the training diffusion noise schedule such that the signal-to-noise ratio (SNR) at the final diffusion timestep $T$ is zero [32]. This ensures enough noise is added during training so as to mitigate this difference between train and test time usage. + +Implementation Details. We build upon a text-to-image latent diffusion model [49] trained on a large in-house image-text dataset. It leverages a VAE [26] to map the images to a 4D latent space, with a $8 \times$ reduction in spatial resolution. The U-Net largely follows the implementation from [49], with minor architectural modifications. As conditioning, it uses the Flan-T5-XXL text encoder [10]. We train Stacked-Diffusion for 14000 steps using the AdamW optimizer [35] with a learning rate of $10^{-4}$ , weight decay of 0.01, and gradient clipping at $\ell_2$ norm of 1.0. For classifier-free guidance, we use a conditioning dropout rate of 0.05. We choose to generate at most $N = 6$ images simultaneously, because as we will see in § 5.1, the vast majority of the data available has 6 or fewer steps. For shorter training sequences, we pad with empty frames and dummy step texts, and for longer, we drop the extra steps and images. We ablate this choice of $N$ in § 5.4. Hence in practice, for batch of 8 sets of 6 steps at 256px resolution, we get an encoded latent of $8 \times 6 \times 4 \times 32 \times 32$ . We concatenate the latents spatially into $8 \times 4 \times (6 * 32) \times 32$ . The denoising U-Net is finetuned to denoise these stacked images, leveraging what it has previously learned about spatial consistency, and adapts to this new setting. At inference time, steps are generated using a pretrained LLM [44], prompted to generate at most $N$ steps ( $\S 10$ ). To generate illustrations, noise is sampled in the shape of the tiled latent with $N$ steps, and the steps generated by the LLM are padded with dummy steps if less + +![](images/abdd9c4211ed1cebf5401101ac227efbb6d2e36a7e02f360f2b02270f203e1cb.jpg) + +![](images/b85bd28800215fdabb0f01e6dfd03b75541a101b14794e1866016d50c4120439.jpg) + +![](images/69e3b1946c36b74b5383b6f23d0211d25bb483f9edc8239e70cb340ac12e54d1.jpg) + +![](images/ce3a95db7a595c6bade73f555e74213878dd23564813b3b498a0becd576fa658.jpg) +Figure 5. Metrics. We introduce three metrics, one for each desideratum presented in § 3. Goal faithfulness: the second image does not show muffins. Step faithfulness: the second image does not show the step. Cross-image consistency: the second image shows a different number of meatballs with different visuals. + +![](images/0fedcf6895c2292a06f1ceb20ffba2a1d5784ef75c7c2a170c7501cdd6f60a36.jpg) + +![](images/c1e4cf8496c64bbf680b983e316127bd5eeba422f3f6da5bf10871fd0a4ad473.jpg) + +than $N$ . The resulting noise is denoised using the U-Net as usual, and finally reshaped to obtain the $N$ output images upon decoding. See Appendix § 15 for further details. + +# 5. Experiments + +We train and evaluate StackedDiffusion on web-based instructional data. We compare the generations to multiple established baselines and ground truth, using automatic metrics and human evaluations. We now describe the data, metrics, baselines, and the key results. Finally, we demonstrate some new applications that StackedDiffusion enables, showing how it goes beyond standard instructional articles. + +# 5.1. Illustrated Instructions Dataset + +We introduce a new dataset to train and evaluate models for the Illustrated Instructions task, by repurposing the Visual Goal-Step Inference (VGSI) dataset [67]. VGSI consists of WikiHow articles, each of which have a high-level goal, 6-10 natural language steps (see Figure 4), and associated image illustrations. We observe that the same data can be used for generating instructional articles. It can provide signal for how the goal and step texts should map to output images, and how images should match with each other. + +For evaluation, we construct a held-out set from this same data. Used directly, however, the data is not well suited as many of the tasks are too high-level to be useful for consistent illustration. For instance, "How to Start a Business in North Carolina" may have steps "Brainstorm + +
T2I (Frozen)22.092.943.251.369.3
T2I (Finetuned)33.378.852.451.553.5
StackedDiffusion(ref)74.361.550.739.5
Ground Truth82.581.773.750.6(N/A)
+ +Table 1. Comparison to baselines. Human evaluation is reported as win rate vs our full StackedDiffusion model. GF corresponds to goal faithfulness accuracy, SF to step faithfulness accuracy, CIC to cross-image consistency, and FID to Fréchet Inception Distance. + +ideas” and “Go to the courthouse” that have no shared visual content. As such, we filter the data for evaluation to the “Recipes” category, which always has a visually clear end state and where illustrations must be consistent with each other. Please see Appendix § 11 for more details. + +# 5.2. Metrics + +Having established three desiderata in § 3, we now turn to the question of how to evaluate them. In addition to evaluating the quality of images using FID [22], we propose three metrics, one for each desideratum, that can be used to evaluate the faithfulness and consistency of the generated article. Finally, given the limitations of automatic metrics for evaluating generative modeling tasks [55], we use human evaluations as our primary metric for overall quality. We illustrate the metrics in Figure 5 and briefly describe them next. Please see the Appendix § 12 for more details. + +Goal Faithfulness (GF) measures how well the generated image is associated to the goal text. We evaluate this by constructing multiple-choice questions (MCQ) as in VGSI [67]. For each generated image, we compare its CLIP similarity [47] with the correct goal text vs the similarity with the texts of three other randomly selected goals. We compute the accuracy of the model in choosing the correct goal text. Step Faithfulness (SF) measures how faithfully the generated image illustrates the step it aims to depict. An image should match the text for the step it was made for, more than other steps. We measure this using CLIP similarity and a MCQ task similar to goal faithfulness, where the image should have higher CLIP similarity with the corresponding step text than the other step texts within the same goal. + +Cross-Image Consistency (CIC) evaluates how consistent the generated images for a goal are with each other and penalizes jarring inconsistencies across the images, such as objects changing color or number. For instance, if a particular set of ingredients are shown for "gather the ingredients," then the same ingredients should be shown for "mix the dry ingredients." We measure this by computing the average $\ell_2$ distance between DINO [7] embeddings of the images for each step, as this considers the similarity of visual (rather than purely semantic) features. Like prior work [50], we + +![](images/8c7902d9d52dc15231ab9b688dcca9b7b887bd76069f368d1e1ce7da4c184769.jpg) +Goal: How to Bake Turkey Wings +Figure 6. Baseline comparison. StackedDiffusion images are preferred overwhelmingly over baselines. The frozen baseline tends to only produce images showing the goal, while the finetuned baseline produces images that are more faithful to the step text, but have no visual features in common. + +also found that CLIP features do not work well for image similarity as they are invariant to critical aspects such as color, number of objects, style etc. DINO's self-supervised pretraining objective leads to features that are sensitive to the visual aspects particular to an image, rather than being invariant to aspects not captured by category or language. + +Human Evaluation is finally used to confirm that the automated evaluation corresponds to actual human preferences. We provide fully rendered articles from each model to human evaluators on Amazon Mechanical Turk (AMT), and ask them to choose which article they prefer, or if they are tied. We then consider the win rate as the proportion for which a majority of evaluators selected the given method compared to the articles produced by StackedDiffusion. A win rate of 50 denotes the methods are tied, whereas $< 50$ denotes the method is worse than StackedDiffusion. + +# 5.3. Baselines + +We now describe some baseline approaches based on existing state-of-the-art image generation models, and compare them to StackedDiffusion in Table 1. + +T2I (frozen): The most obvious choice for a baseline is to simply use a pretrained and frozen text-to-image (T2I) model. We use our in-house T2I model that is also used to initialize StackedDiffusion. To ensure this baseline gets same information as StackedDiffusion, we prompt the model with a concatenation of the goal with each step, $\mathbf{g}||\mathbf{s}_i$ , for each $i$ , and produce a single image $\mathbf{I}_i$ . The sequence then is composed of $N$ independent generations. + +We find that these images are faithful to the goal text, but fail to be faithful to the step text and lack consistency. The goal faithfulness is substantially higher than even the ground truth (92.9% vs 81.7%). This is because the model is not capable of reasoning about what a particular step to- + +![](images/d01c9dd6964ee9155412d9b38e95567509bba414f3c9854c8639419d5446d72f.jpg) +Figure 7. Effect of training data. We find that having more training data improves StackedDiffusion's faithfulness and consistency. However even with $20\%$ data it performs well, thanks to its effective use of pretrained T2I model weights. + +wards a goal should look like. Instead, it simply generates an image that matches words in the input text, as illustrated in the first row of Figure 2. This results in all the images for e.g. "Make Colored Ice" showing the actual finished ice rather than any of the intermediate steps. In other words, the step text is eclipsed by the goal text. + +T2I (finetuned): We finetune a T2I model with the goal and step text embeddings concatenated together as the condition, again producing $N$ independent generations. We find that this model is more faithful to the step text than the frozen T2I model, but still substantially less than our final model (52.4% vs 61.5%). This suggests the goal still overshadows the step text, but to a lesser extent. The goal faithfulness is also lower than for the frozen T2I model, as this model is trained to create images for steps rather than goals. The cross-image consistency is low, similar to the frozen model, as the images are still generated independently. + +StackedDiffusion achieves substantially improved step faithfulness and cross-image consistency, while maintaining high goal faithfulness. We show an illustrative example in Figure 6. Unlike the baselines, the cross-image consistency is very close to that of the ground truth data; as generating them together allows for influence between them during the generation process. We also find that the FID of our generated images is substantially lower than that of other models, indicating that our generated images are more similar to the ground truth data. Finally on human evaluations, StackedDiffusion clearly outperforms them both by a preference of $78\%$ and $66.6\%$ respectively. Perhaps most notably, when compared to the ground truth images, human evaluators still picked StackedDiffusion $18.5\%$ of the times, suggesting that our model generates some illustrations even better than the ground truth data. Further baselines can be found in Appendix § 13, with a comparison of training and inference costs in Appendix § 14. + +# 5.4. Ablations + +Step Count. As discussed in § 4, we elect to use a maximum step count of 6 as it allows sufficient data coverage + +
Win RateWin Rate
T2I (Frozen)33.8T2I (Finetuned)34.0
GILL [27]++3.9LLM + CLIP15.6
Goal Retrieval33.1Ground Truth70.0
+ +and most articles fall under this value. We find that human evaluators prefer this model to the shorter step count model of length 4. It is slightly preferred over the model trained with a longer step count of 8, likely as there is insufficient data of long lengths, limiting any advantage that might be gained in longer length generations. + +Table 2. System-level comparison to prior work. Human evaluation results using fully generated outputs. Reported as human evaluation win rate of the associated method over StackedDiffusion (hence, higher is better for the reported method). + +
Baseline4 Step (↑)8 Step (↑)
6 Step32.146.9
+ +Training data. In Figure 7, we randomly sub-sample varying proportions of the total data, and train StackedDiffusion on each of those subsets. As the amount of training data increases, it generates more faithful and consistent images. However even with less data it performs well, owing to its effective use of the pretrained T2I model's initialization. + +Importance of 0SNR. We experiment with the 0SNR technique [32] due to the high spatial dimensionality of the latents. We find that not using 0SNR results in substantially reduced step faithfulness, dropping from $61.5\%$ to $52.8\%$ . + +Importance of Step-Positional Encoding. We also evaluate the importance of the step-positional encoding, $\Phi_{i}$ . This gives the model a critical cue for which parts of the condition correspond to which steps. We find that not using the step-positional encoding results in substantially reduced step faithfulness as well, dropping from $61.5\%$ to $49.8\%$ . + +# 5.5. Comparison to prior work + +The previous metrics were computed with respect to fixed ground truth goal and step texts to provide direct comparisons. However, they do not encompass the full scope of our model's capabilities: the flexibility of the generated text is one of StackedDiffusion's core strengths. To evaluate in a closer setting to this real-world usage, we examine the quality of the articles generated by the full system, where the steps are generated with a LLM [44]. We perform system-level comparisons using human evaluators, including the quality of the generated text and the quality of the images created from the generated text. + +We show our results in Table 2. We first compare to the baselines introduced in § 5.3, and see results similar to Table 1. We see that even with LLM-generated text, Stacked-Diffusion outperforms baseline T2I based approaches, either frozen or finetuned, by more than $66\%$ win-rate. + +![](images/47d8b68fa6dad61088500ed14d3c77f59154410ea0b9c57e3bf7ef12f1180920.jpg) + +Figure 8. Personalization. StackedDiffusion enables instructions personalized far beyond what is possible with preexisting articles. +![](images/30e376064a07d20776c2aed48e05562dd3559b4e43dbd3f254616a1f596185ad.jpg) +Step 1: Drain and press tofu to remove excess water. Step 2: Crumble tofu; mix with bread crumbs and a binding agent. Step 3: Shape tofu mixture into burger patties. Step 4: Cook over medium heat until golden brown. + +![](images/2348732110f808aca1b589d07bbf4efcca036cfb0c16289a20171014c1d884d4.jpg) + +![](images/ec4155dfeef0eac5a21e03881668f562e1414ec8874f7671ca5b099fa54a8bd3.jpg) + +![](images/d177b9569dac71da510a5c24b37d1bad26592d5be34789f6db9a26cc052bfcbd.jpg) +$\left( {x - {2x}}\right) t - x{y}^{2} = \left( {x - {2x}}\right) {f}^{\prime }t$ + +Additionally, we compare StackedDiffusion to recent state-of-the-art approaches in multimodal generation, as in principle those can also produce instructional text combined with illustrations. Specifically, we compare to GILL [29]-- a model trained to generate sequences of interleaved text and images. While other similar approaches have been proposed [1, 28], GILL is the best open-source model we could access. Using GILL directly, however, results in a human evaluation win rate of $0\%$ . We posit this was for three primary reasons. First, the GILL model was unable to generate long enough generations to suffice as an article in comparison to our generations, despite our best efforts in modifying the generation parameters. Secondly, the model often did not produce any illustrations with the text, although we tried various prompting tactics. Thirdly, the text or images produced were often not of high enough quality to actually reflect the goal or steps. We thus introduce an alternative, which we call GILL++. This uses the (superior) text produced by the same language model as in our method, but passes each step text as input to GILL to generate the corresponding image. This results in a human evaluation win rate of $3.9\%$ against StackedDiffusion. See Appendix § 16 for details on prompting GILL, and our GILL++ baseline. + +Next, we compare StackedDiffusion to a retrieval based approach, denoted by LLM+CLIP. Here we retrieve the closest image in the training set according to CLIP similarity to each of the steps created by the LLM (with each concatenated to the goal text, similar to how the frozen T2I model is used). This retrieval-based metric is not suitable for the automated metrics previously introduced as it is based on the same underlying similarity scores as the automated metrics, but we introduce it here for human evaluation. We find that StackedDiffusion is preferred overwhelmingly over this retrieval-based approach, despite the retrieval-based approach using real images from the training set. Similarly, we compare to another retrieval-based approach which retrieves the full article with the most similar goal text to the input goal, denoted Goal Retrieval. We find StackedDiffusion strongly outperforms this as well. These results together suggest that StackedDiffusion is able + +![](images/d4b653e492533a300057bdbe939d13e17d48e00c7e0b6735708fe389635afc0b.jpg) + +I have a bunch of apples left from apple picking, what can I do with them? + +If you have a lot of apples, you can turn them into a delicious apple crisp. Here's a simple recipe to get you started: + +![](images/e00d1a3038a1f738d21b711761e3aa64b83c497911abf45e70133c1ba0ae463f.jpg) + +Step 1: Preheat oven to $350^{\circ}\mathrm{F}$ $(175^{\circ}\mathrm{C})$ + +Step 2: Peel, core, and slice apples; place in a baking dish. + +Step 3: Mix sugar, cinnamon, and a pinch of salt; sprinkle over apples. + +![](images/5f76b982470e95c1a04b54ce7ae0f0afdaba7682efdab2ea819e9352c3bca48a.jpg) + +![](images/b1c1e3407c5aa95a720bcb4c1c6c5943e545d72d586d640120c8b8f2eafa8863.jpg) +Step 5: Bake for 45 minutes or until golden brown. + +![](images/b05c9784cb602e72512f7c6d59ede22c192a020dd1bd392ecadebc28f2d9f715.jpg) +Step 6: Serve warm. + +![](images/157a7442a98421ef62dc2388a1be819b14732a906eea65091a219119a50cdc22.jpg) +Step 4: Combine oats, flour, brown sugar, and butter; crumble over apples. +Figure 9. Goal suggestion and inference. StackedDiffusion can suggest what goal might be most relevant given other information. + +![](images/1321800589acd81068bc2c5de5b02f49a33f548f8744f3bf84dfb6aca7cd8204.jpg) + +![](images/9c76d104f77fa03d55ab031c6807146db39a5db9c5d18bc3da885a9bdaa21beb.jpg) + +to generate novel instructional articles that one can not simply retrieve from the training corpus. + +Finally, we compare StackedDiffusion to the ground truth. Surprisingly again, we find that the human annotators pick our generations $30\%$ of times even compared to ground truth. Note that these articles are handwritten and manually illustrated for the purpose of illustrating these goals. This shows the strong promise of our proposed approach. We believe with the rapid improvements in training data and generative modeling techniques, future iterations of our model could be indistinguishable or even better than manually created illustrated instructional articles. + +# 5.6. Applications + +Perhaps the biggest strength of a generative approach is its ability to handle unique user queries that may not fit into the boundaries posed by static articles on the web. Hence, we demonstrate the capabilities of the full system, combining the strengths of text generation capabilities from the language model with StackedDiffusion. + +Personalized instruction. A user can provide any situational information specific to their circumstances and obtain an article customized to that situation. This is not possible with fixed, existing articles. For example, in Figure 8, the user can specify a diet and obtain an article for the food they want that adheres to that diet. + +Goal suggestion. As StackedDiffusion accepts flexible input text, a user can provide higher-level information about + +![](images/8cad8b81aae740759c89e121102a47a3149f97aaa02e952b7f0084b992639edb.jpg) + +Oh no! I'm making bread and forgot to add salt when I made the dough! I'm at the kneading step. + +If you're at the kneading step, don't worry—you can still incorporate it into the dough with more kneading. Here's how: + +![](images/71ab28f90bc8e280ee1341f7124499b16a1260acdeae65b53746c69b2614a258.jpg) + +Step 1: Determine the amount of salt needed. + +Step 2: Dissolve the salt in a small amount of warm water. + +Step 3: Flatten the dough into a rectangle. + +![](images/10698da331ef96acbc1722d2850a1bffea0f43d33a1f952ff8ed107c2de59022.jpg) +Step 4: Spread the salt over the dough. + +![](images/5a9ec6801506eaa2482dff9b1df28cae824c3c53e6bb31f7365e12d82e9b060e.jpg) +Step 5: Fold and knead the dough to distribute. + +![](images/7aaf6a10495f2721bc93a9ad1d73b033ad5ff37563972cae78a779ae62c29f69.jpg) +Step 6: Proceed with the original bread recipe. + +![](images/4d3dd587c80162601dfb094dcfae5e45d3e615ea78c912df4e7fb6fa3d390f55.jpg) + +![](images/e8b979dbd1ff2ba95c5013084b3030110152c106f174a67e1949ac6baff1f842.jpg) + +![](images/9a4282b040e672208ad09baa2280751c0806c52f20519d106ff1dcc6b7b1822a.jpg) +Figure 10. Error correction. StackedDiffusion provides updated instructions in response to unexpected situations, like a user error. + +what they want to do and obtain an article with a suggested specific goal that matches what the user would like to do. In Figure 9, the user describes their situation (having many apples) and StackedDiffusion suggests a goal that matches their situation (making apple crisps). + +Error correction. Oftentimes, a user in the course of performing a task may make a mistake. Fixed articles provide no avenue for recourse. However, StackedDiffusion can adapt to user error and create alternative instructions that best correct for and accommodate the user's situation. See Figure 10 for an example of this. A fixed article would not provide any alternative paths that adapt as the user performs the action. This highlights the new avenues opened up by our generative approach to this task. + +# 6. Limitations and Conclusions + +While StackedDiffusion achieves strong performance in generating Illustrated Instructions, we note that there still remains a gap with ground truth data. Future work might examine how to obtain instructional data at greater scale and how this influences StackedDiffusion. Leveraging the latest T2I architectures to improve the faithfulness and quality of our generations would be another promising avenue for future work. Finally, leveraging improvements in text-to-video techniques to generate video clips describing each step would further improve the usability of such a system. + +# References + +[1] Armen Aghajanyan, Bernie Huang, Candace Ross, Vladimir Karpukhin, Hu Xu, Naman Goyal, Dmytro Okhonko, Mandar Joshi, Gargi Ghosh, Mike Lewis, and Luke Zettlemoyer. CM3: A causal masked multimodal model of the internet. arXiv preprint arXiv:2201.07520, 2022. +[2] Emanuele Aiello, Lili Yu, Yixin Nie, Armen Aghajanyan, and Barlas Oguz. Jointly training large autoregressive multimodal models. In ICLR, 2024. +[3] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, 2021. +[4] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. +[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your Latents: High-Resolution Video Synthesis with Latent Diffusion Models. In CVPR, 2023. +[6] Russell N. Carney and Joel R. Levin. Pictorial illustrations still improve students' learning from text. Educational Psychology Review, 14(1):5-26, 2002. +[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging Properties in Self-Supervised Vision Transformers. In ICCV, 2021. +[8] Huiwen Chang, Han Zhang, Jarred Barber, Aaron Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Patrick Murphy, William T. Freeman, Michael Rubinstein, Yanzhen Li, and Dilip Krishnan. Muse: Text-ToImage Generation via Masked Generative Transformers. In ICML, 2023. +[9] Brian Chen, Andrew Rouditchenko, Kevin Duarte, Hilde Kuehne, Samuel Thomas, Angie Boggust, Rameswar Panda, Brian Kingsbury, Rogerio Feris, David Harwath, et al. Multimodal clustering networks for self-supervised learning from unlabeled videos. In ICCV, 2021. +[10] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Alex Castro-Ros, Marie Pellat, Kevin Robinson, Dasha Valter, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416, 2022. +[11] Yang Deng, Wai Lam, Yuexiang Xie, Daoyuan Chen, Yaliang Li, Min Yang, and Ying Shen. Joint Learning of Answer Selection and Answer Summary Generation in Community Question Answering. In AAAI, 2020. +[12] Eadom Dessalene, Chinmaya Devaraj, Michael Maynord, Cornelia Fermuller, and Yiannis Aloimonos. Forecasting ac + +tion through contact representations from first person video. PAMI, 2021. +[13] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In ICCV, 2023. +[14] Bahare Fatemi, Quentin Duval, Rohit Girdhar, Michal Drozdzal, and Adriana Romero-Soriano. Learning to substitute ingredients in recipes. arXiv preprint arXiv:2302.07960, 2023. +[15] Antonino Furnari and Giovanni Maria Farinella. Rolling-unrolling lstms for action anticipation from first-person video. PAMI, 2020. +[16] Valentin Gabeur, Chen Sun, Karteek Alahari, and Cordelia Schmid. Multi-modal transformer for video retrieval. In ECCV, 2020. +[17] Valentin Gabeur, Arsha Nagrani, Chen Sun, Karteek Alahari, and Cordelia Schmid. Masking modalities for cross-modal video retrieval. In WACV, 2022. +[18] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. In ECCV, 2022. +[19] Rohit Girdhar and Kristen Grauman. Anticipative video transformer. In ICCV, 2021. +[20] Rohit Girdhar, Mannat Singh, Andrew Brown, Quentin Duval, Samaneh Azadi, Sai Saketh Rambhatla, Akbar Shah, Xi Yin, Devi Parikh, and Ishan Misra. Emu video: Factorizing text-to-video generation by explicit image conditioning. arXiv preprint arXiv:2311.10709, 2023. +[21] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In CVPR, 2022. +[22] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. +[23] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen Video: High Definition Video Generation with Diffusion Models. arXiv preprint arXiv:2210.02303, 2022. +[24] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. +[25] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. +[26] Diederik P. Kingma and Max Welling. Auto-Encoding Variational Bayes. In ICLR, 2014. +[27] Jing Yu Koh, Daniel Fried, and Ruslan Salakhutdinov. Generating images with multimodal language models. In NeurIPS, 2023. + +[28] Jing Yu Koh, Ruslan Salakhutdinov, and Daniel Fried. Grounding language models to images for multimodal inputs and outputs. In ICML, 2023. +[29] Jing Yu Koh, Ruslan Salakhutdinov, and Daniel Fried. Grounding Language Models to Images for Multimodal Generation. In NeurIPS, 2023. +[30] Mahnaz Koupae and William Yang Wang. WikiHow: A Large Scale Text Summarization Dataset. arXiv preprint arXiv:1810.09305, 2018. +[31] Linjie Li, Yen-Chun Chen, Yu Cheng, Zhe Gan, Licheng Yu, and Jingjing Liu. Hero: Hierarchical encoder for video+ language omni-representation pre-training. arXiv preprint arXiv:2005.00200, 2020. +[32] Shanchuan Lin, Bingchen Liu, Jiashi Li, and Xiao Yang. Common diffusion noise schedules and sample steps are flawed. arXiv preprint arXiv:2305.08891, 2023. +[33] Xudong Lin, Fabio Petroni, Gedas Bertasius, Marcus Rohrbach, Shih-Fu Chang, and Lorenzo Torresani. Learning To Recognize Procedural Activities with Distant Supervision. In CVPR, 2022. +[34] Song Liu, Haoqi Fan, Shengsheng Qian, Yiru Chen, Wenkui Ding, and Zhongyuan Wang. Hit: Hierarchical transformer with momentum contrast for video-text retrieval. In ICCV, 2021. +[35] Ilya Loshchilov and Frank Hutter. Decoupled Weight Decay Regularization. In ICLR, 2019. +[36] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon Bharti, and Ming Zhou. Univl: A unified video and language pre-training model for multimodal understanding and generation. arXiv preprint arXiv:2002.06353, 2020. +[37] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval. arXiv preprint arXiv:2104.08860, 2021. +[38] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Ti-niu Tan. Videofusion: Decomposed diffusion models for high-quality video generation. In CVPR, 2023. +[39] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. HowTo100M: Learning a Text-Video Embedding by Watching Hundred Million Narrated Video Clips. In ICCV, 2019. +[40] Antoine Miech, Jean-Baptiste Alayrac, Ivan Laptev, Josef Sivic, and Andrew Zisserman. Thinking fast and slow: Efficient text-to-visual retrieval with transformers. In CVPR, 2021. +[41] Medhini Narasimhan, Anna Rohrbach, and Trevor Darrell. Clip-it! language-guided video summarization. NIPS, 2021. +[42] Medhini Narasimhan, Arsha Nagrani, Chen Sun, Michael Rubinstein, Trevor Darrell, Anna Rohrbach, and Cordelia Schmid. Tl; dw? summarizing instructional videos with task relevance and cross-modal saliency. In ECCV, 2022. +[43] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and + +Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2022. +[44] OpenAI. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. +[45] Gaurav Parmar, Richard Zhang, and Jun-Yan Zhu. On aliased resizing and surprising subtleties in gan evaluation. In CVPR, 2022. +[46] AJ Piergiovanni, Anelia Angelova, Michael S Ryoo, and Irfan Essa. Unsupervised action segmentation for instructional videos. arXiv preprint arXiv:2106.03738, 2021. +[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. +[48] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. +[49] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bjorn Ommer. High-Resolution Image Synthesis with Latent Diffusion Models. In CVPR, 2022. +[50] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. DreamBooth: Fine Tuning Text-to-Image Diffusion Models for Subject-Driven Generation. In CVPR, 2023. +[51] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding. In NIPS, 2022. +[52] Fadime Sener and Angela Yao. Zero-shot anticipation for instructional activities. In ICCV, 2019. +[53] Paul Hongsuck Seo, Arsha Nagrani, and Cordelia Schmid. Look before you speak: Visually contextualized utterances. In CVPR, 2021. +[54] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-A-Video: Text-to-Video Generation without Text-Video Data. arXiv preprint arXiv:2209.14792, 2022. +[55] George Stein, Jesse C Cresswell, Rasa Hosseinzadeh, Yi Sui, Brendan Leigh Ross, Valentin Villecloze, Zhaoyan Liu, Anthony L Caterini, J Eric T Taylor, and Gabriel Loiza-Ganem. Exposing flaws of generative model evaluation metrics and their unfair treatment of diffusion models. arXiv preprint arXiv:2306.04675, 2023. +[56] Chen Sun, Fabien Baradel, Kevin Murphy, and Cordelia Schmid. Learning video representations using contrastive bidirectional transformer. arXiv preprint arXiv:1906.05743, 2019. +[57] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Emu: Generative pretraining in multimodality. In ICLR, 2024. + +[58] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. COIN: A Large-scale Dataset for Comprehensive Instructional Video Analysis. In CVPR, 2019. +[59] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable Length Video Generation from Open Domain Textual Descriptions. In ICLR, 2022. +[60] Wenzhe Wang, Mengdan Zhang, Runnan Chen, Guanyu Cai, Penghao Zhou, Pai Peng, Xiaowei Guo, Jian Wu, and Xing Sun. Dig into multi-modal cues for video retrieval with hierarchical alignment. In *IJCAI*, 2021. +[61] Chao-Yuan Wu, Yanghao Li, Karttikeya Mangalam, Haoqi Fan, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Memvit: Memory-augmented multiscale vision transformer for efficient long-term video recognition. arXiv preprint arXiv:2201.08383, 2022. +[62] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In ICCV, 2023. +[63] Mengmeng Xu, Juan-Manuel Pérez-Rúa, Victor Escorcia, Brais Martinez, Xiatian Zhu, Li Zhang, Bernard Ghanem, and Tao Xiang. Boundary-sensitive pre-training for temporal localization in videos. In ICCV, 2021. +[64] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Just ask: Learning to answer questions from millions of narrated videos. In ICCV, 2021. +[65] Jianwei Yang, Yonatan Bisk, and Jianfeng Gao. Taco: Token-aware cascade contrastive learning for video-text alignment. In ICCV, 2021. +[66] Mengjiao Yang, Yilun Du, Bo Dai, Dale Schuurmans, Joshua B Tenenbaum, and Pieter Abbeel. Probabilistic adaptation of text-to-video models. arXiv preprint arXiv:2306.01872, 2023. +[67] Yue Yang, Artemis Panagopoulou, Qing Lyu, Li Zhang, Mark Yatskar, and Chris Callison-Burch. Visual goal-step inference using wikihow. In EMNLP, 2021. +[68] Shengming Yin, Chenfei Wu, Huan Yang, Jianfeng Wang, Xiaodong Wang, Minheng Ni, Zhengyuan Yang, Linjie Li, Shuguang Liu, Fan Yang, Jianlong Fu, Gong Ming, Lijuan Wang, Zicheng Liu, Houqiang Li, and Nan Duan. NUWAXL: Diffusion over Diffusion for eXtremely Long Video Generation. arXiv preprint arXiv:2303.12346, 2023. +[69] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022. +[70] Lili Yu, Bowen Shi, Ramakanth Pasunuru, Benjamin Muller, Olga Golovneva, Tianlu Wang, Arun Babu, Binh Tang, Brian Karrer, Shelly Sheynin, et al. Scaling autoregressive multimodal models: Pretraining and instruction tuning. arXiv preprint arXiv:2309.02591, 2023. + +[71] Jingqing Zhang, Yao Zhao, Mohammad Saleh, and Peter J. Liu. PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization. In ICML, 2019. +[72] Li Zhang, Qing Lyu, and Chris Callison-Burch. Reasoning about Goals, Steps, and Temporal Ordering with WikiHow. In EMNLP, 2020. +[73] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. +[74] Shuyan Zhou, Li Zhang, Yue Yang, Qing Lyu, Pengcheng Yin, Chris Callison-Burch, and Graham Neubig. Show Me More Details: Discovering Hierarchies of Procedures from Semi-structured Web Data. In ACL, 2022. +[75] Yilun Zhou, Julie Shah, and Steven Schockaert. Learning Household Task Knowledge from WikiHow Descriptions. In Proceedings of the 5th Workshop on Semantic Deep Learning (SemDeep-5), 2019. +[76] Linchao Zhu and Yi Yang. Actbert: Learning global-local video-text representations. In CVPR, 2020. +[77] Dimitri Zhukov, Jean-Baptiste Alayrac, Ivan Laptev, and Josef Sivic. Learning actionness via long-range temporal order verification. In ECCV, 2020. \ No newline at end of file diff --git a/2024/Generating Illustrated Instructions/images.zip b/2024/Generating Illustrated Instructions/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..48578d7bb18f68eadc816234b033150632ec9969 --- /dev/null +++ b/2024/Generating Illustrated Instructions/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6faa671c074017ae28e2cce0387a5598d38aac733baebcb8574a83bc18cac40b +size 458646 diff --git a/2024/Generating Illustrated Instructions/layout.json b/2024/Generating Illustrated Instructions/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8ac82fbc4ed7a1ddbe99343ddc911bcb627dac9e --- /dev/null +++ b/2024/Generating Illustrated Instructions/layout.json @@ -0,0 +1,11075 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 187, + 103, + 406, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 103, + 406, + 121 + ], + "spans": [ + { + "bbox": [ + 187, + 103, + 406, + 121 + ], + "type": "text", + "content": "Generating Illustrated Instructions" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 162, + 142, + 247, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 142, + 247, + 156 + ], + "spans": [ + { + "bbox": [ + 162, + 142, + 247, + 156 + ], + "type": "text", + "content": "Sachit Menon" + }, + { + "bbox": [ + 162, + 142, + 247, + 156 + ], + "type": "inline_equation", + "content": "^{1,2*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 268, + 143, + 331, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 143, + 331, + 156 + ], + "spans": [ + { + "bbox": [ + 268, + 143, + 331, + 156 + ], + "type": "text", + "content": "Ishan Misra" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 357, + 143, + 429, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 143, + 429, + 156 + ], + "spans": [ + { + "bbox": [ + 357, + 143, + 429, + 156 + ], + "type": "text", + "content": "Rohit Girdhar1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 196, + 157, + 266, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 157, + 266, + 171 + ], + "spans": [ + { + "bbox": [ + 196, + 157, + 266, + 171 + ], + "type": "text", + "content": "1GenAI, Meta" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 290, + 158, + 397, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 290, + 158, + 397, + 172 + ], + "spans": [ + { + "bbox": [ + 290, + 158, + 397, + 172 + ], + "type": "text", + "content": "2Columbia University" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 154, + 174, + 435, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 174, + 435, + 184 + ], + "spans": [ + { + "bbox": [ + 154, + 174, + 435, + 184 + ], + "type": "text", + "content": "https://facebookresearch.github.io/IllustratedInstructions" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 239, + 290, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 290, + 443 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 290, + 443 + ], + "type": "text", + "content": "We introduce a new task of generating \"Illustrated Instructions\", i.e. visual instructions customized to a user's needs. We identify desiderata unique to this task, and formalize it through a suite of automatic and human evaluation metrics, designed to measure the validity, consistency, and efficacy of the generations. We combine the power of large language models (LLMs) together with strong text-to-image generation diffusion models to propose a simple approach called StackedDiffusion, which generates such illustrated instructions given text as input. The resulting model strongly outperforms baseline approaches and state-of-the-art multimodal LLMs; and in " + }, + { + "bbox": [ + 46, + 239, + 290, + 443 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 46, + 239, + 290, + 443 + ], + "type": "text", + "content": " of cases, users even prefer it to human-generated articles. Most notably, it enables various new and exciting applications far beyond what static articles on the web can provide, such as personalized instructions complete with intermediate steps and pictures in response to a user's individual situation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 467, + 128, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 467, + 128, + 479 + ], + "spans": [ + { + "bbox": [ + 47, + 467, + 128, + 479 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 488, + 287, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 488, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 46, + 488, + 287, + 632 + ], + "type": "text", + "content": "The internet is a vast resource to find answers to all types of questions. It is often easy to find a webpage or a video that walks through the exact steps to achieve a user's goal. With the rise of Large Language Models (LLMs) trained on internet-scale data, users can just ask the LLM for instructions to achieve a goal. This allows the users to get answers for specific personalized queries for which there may not be an existing webpage on the internet, e.g., modifying cooking recipes with user specific dietary restrictions. Moreover, if the users make a mistake when following the instructions, simple follow-up questions to the LLM can generate alternate instructions, a major advantage over static web search." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 632, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 632, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 632, + 287, + 693 + ], + "type": "text", + "content": "In spite of such advantages, LLMs still have one major limitation - they cannot generate visuals, which are critical for users to learn from and follow instructions for a wide range of tasks [6]. Consider, for instance, instructions that require visual inspection, e.g., a recipe that requires" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 315, + 215, + 327, + 232 + ], + "blocks": [ + { + "bbox": [ + 315, + 215, + 327, + 232 + ], + "lines": [ + { + "bbox": [ + 315, + 215, + 327, + 232 + ], + "spans": [ + { + "bbox": [ + 315, + 215, + 327, + 232 + ], + "type": "image", + "image_path": "4fe67dc0ae5b1b060518dded246c2b77f59842388aaf6c09d967ab5aad81875d.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 332, + 220, + 511, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 220, + 511, + 228 + ], + "spans": [ + { + "bbox": [ + 332, + 220, + 511, + 228 + ], + "type": "text", + "content": "How can I bake a cake to celebrate an event in Seattle?" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 236, + 517, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 517, + 253 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 517, + 253 + ], + "type": "text", + "content": "One great way could be to make a coffee-infused cake to honor Seattle's coffee culture, with teal coloring for the Seattle flag." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 259, + 375, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 259, + 375, + 275 + ], + "spans": [ + { + "bbox": [ + 313, + 259, + 375, + 275 + ], + "type": "text", + "content": "Step 1: Brew strong coffee; let cool." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 391, + 259, + 453, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 259, + 453, + 274 + ], + "spans": [ + { + "bbox": [ + 391, + 259, + 453, + 274 + ], + "type": "text", + "content": "Step 2: Preheat the oven to " + }, + { + "bbox": [ + 391, + 259, + 453, + 274 + ], + "type": "inline_equation", + "content": "350\\mathrm{F}" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 470, + 258, + 536, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 258, + 536, + 274 + ], + "spans": [ + { + "bbox": [ + 470, + 258, + 536, + 274 + ], + "type": "text", + "content": "Step 3: Prepare cake batter as usual." + } + ] + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 311, + 277, + 385, + 350 + ], + "blocks": [ + { + "bbox": [ + 311, + 277, + 385, + 350 + ], + "lines": [ + { + "bbox": [ + 311, + 277, + 385, + 350 + ], + "spans": [ + { + "bbox": [ + 311, + 277, + 385, + 350 + ], + "type": "image", + "image_path": "852682748d3765eb82a702ea3eb732d06dc543bcbfb0a3a2c45dc54dc4fdd451.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 352, + 376, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 352, + 376, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 352, + 376, + 376 + ], + "type": "text", + "content": "Step 4: Mix in the coffee; fold in coffee grounds for texture." + } + ] + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 390, + 277, + 463, + 350 + ], + "blocks": [ + { + "bbox": [ + 390, + 277, + 463, + 350 + ], + "lines": [ + { + "bbox": [ + 390, + 277, + 463, + 350 + ], + "spans": [ + { + "bbox": [ + 390, + 277, + 463, + 350 + ], + "type": "image", + "image_path": "e19a4df7e34c8a6fff064a96695f89adb558a1d0115a4a06909e77c1e98a91c2.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "bbox": [ + 391, + 352, + 454, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 352, + 454, + 368 + ], + "spans": [ + { + "bbox": [ + 391, + 352, + 454, + 368 + ], + "type": "text", + "content": "Step 5: Bake for 25- 30 minutes." + } + ] + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 468, + 277, + 542, + 350 + ], + "blocks": [ + { + "bbox": [ + 468, + 277, + 542, + 350 + ], + "lines": [ + { + "bbox": [ + 468, + 277, + 542, + 350 + ], + "spans": [ + { + "bbox": [ + 468, + 277, + 542, + 350 + ], + "type": "image", + "image_path": "369dc6ae7426ff6b589954507b0a21efeed223528b8ae9e9a5880f6e5ba09956.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 469, + 352, + 531, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 469, + 352, + 531, + 376 + ], + "spans": [ + { + "bbox": [ + 469, + 352, + 531, + 376 + ], + "type": "text", + "content": "Step 6: Cover the cake in teal fondant for the Seattle flag." + } + ] + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 310, + 376, + 385, + 451 + ], + "blocks": [ + { + "bbox": [ + 310, + 376, + 385, + 451 + ], + "lines": [ + { + "bbox": [ + 310, + 376, + 385, + 451 + ], + "spans": [ + { + "bbox": [ + 310, + 376, + 385, + 451 + ], + "type": "image", + "image_path": "03ec5dc17b70cd300b497e23cf15e210c38f3eacc6c5e2ca0a6809e7789ec75c.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 460, + 545, + 516 + ], + "lines": [ + { + "bbox": [ + 305, + 460, + 545, + 516 + ], + "spans": [ + { + "bbox": [ + 305, + 460, + 545, + 516 + ], + "type": "text", + "content": "Figure 1. StackedDiffusion generating Illustrated Instructions. Given a goal (or any textual user input), StackedDiffusion produces a customized instructional article complete with illustrations that not only tells the user how to achieve the goal in words, but also shows the user by providing illustrations." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 390, + 377, + 463, + 451 + ], + "blocks": [ + { + "bbox": [ + 390, + 377, + 463, + 451 + ], + "lines": [ + { + "bbox": [ + 390, + 377, + 463, + 451 + ], + "spans": [ + { + "bbox": [ + 390, + 377, + 463, + 451 + ], + "type": "image", + "image_path": "987a18f0208641696c5ad4ea810c0b7bf3299c2c539484dd7750f5e35378d311.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 468, + 377, + 542, + 451 + ], + "blocks": [ + { + "bbox": [ + 468, + 377, + 542, + 451 + ], + "lines": [ + { + "bbox": [ + 468, + 377, + 542, + 451 + ], + "spans": [ + { + "bbox": [ + 468, + 377, + 542, + 451 + ], + "type": "image", + "image_path": "5821d0db1c60bb551af3ca97b0d29079b4ef19842d877374e0d623776f483002.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 532, + 545, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 532, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 304, + 532, + 545, + 591 + ], + "type": "text", + "content": "the user to stir fry onions until golden brown, or searching for bubbles when a flat tire is submerged under water. An image accompanying such instructions can make following them significantly easier. Can we develop methods with the strengths of LLMs that can also generate such visuals?" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "content": "In this work, we tackle this challenge, developing models that can not only tell a user how to accomplish their task, but also show them how. We define the novel task of Illustrated Instructions: creating a set of steps with visualizations that showcase an approach to solve the user's task. We carefully consider the different dimensions of the problem, laying out three desiderata unique to this setting. To measure these desiderata, we develop automated metrics based on prior work in instructional article assessment [67] and image generation [50]." + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 194, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 194, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 194, + 712 + ], + "type": "text", + "content": "*Work done during an internship at Meta." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6274" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 277 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 277 + ], + "type": "text", + "content": "We propose a new model to solve the task of Illustrated Instructions that combines LLMs with a text-to-image diffusion model. We train our model on stacks of instructional images from websites such as WikiHow. The resulting StackedDiffusion model creates full instructional articles, complete with customized steps and a sequence of images uniquely generated to describe those steps. It leverages large-scale pretrained LLMs and finetuned text-to-image diffusion models, and employs techniques to accomplish the task without the need to introduce any new learnable parameters. This includes spatial tiling for simultaneous multi-image generation, text embedding concatenation to reduce information loss in long instructions, and a new \"step-positional encoding\" to better demarcate the different steps in the instructions. Thus, StackedDiffusion can generate useful illustrations even when trained even with limited instructional data (cf. Figure 7)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 279, + 289, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 279, + 289, + 532 + ], + "spans": [ + { + "bbox": [ + 47, + 279, + 289, + 532 + ], + "type": "text", + "content": "We compare StackedDiffusion with various baseline approaches based on off-the-shelf tools, and find that they all fall short, even when used in conjunction with one another. Existing T2I models are incapable of generating visuals directly from a user query. Even when given more detailed instructional text, we show that existing T2I models fail to produce images that are simultaneously faithful to the goal, the step, and consistent among each other. Given the recent introduction of multimodal LLMs [1, 28], we also compare with a recent open-source model, GILL [27], and show such models also fall short of consistently generating useful visuals along with text. We posit that our approach of leveraging the spatial priors learned by pretrained diffusion models to generate multiple images together, in conjunction with pretrained LLMs, is significantly more compute and data efficient than an approach purely based on next-token prediction, without these priors. Our thorough ablations and human evaluations show that StackedDiffusion convincingly surpasses state-of-the-art models. Our final model even outperforms human-generated articles in " + }, + { + "bbox": [ + 47, + 279, + 289, + 532 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 47, + 279, + 289, + 532 + ], + "type": "text", + "content": " of cases, showing the strong potential of our approach." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 534, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 534, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 534, + 289, + 715 + ], + "type": "text", + "content": "Contributions: 1) We introduce the novel task of Illustrated Instructions, which requires generating a sequence of images and text that together describe how to achieve a goal (" + }, + { + "bbox": [ + 47, + 534, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\S 3" + }, + { + "bbox": [ + 47, + 534, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 534, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\S\\S 5.1" + }, + { + "bbox": [ + 47, + 534, + 289, + 715 + ], + "type": "text", + "content": " and 5.2), along with desiderata and metrics for this task; 2) We propose a new approach StackedDiffusion for Illustrated Instructions, with novel modifications to the modeling procedure, enabling generation of visuals suitable for instructional articles for the first time without any additional parameters (" + }, + { + "bbox": [ + 47, + 534, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\S 4" + }, + { + "bbox": [ + 47, + 534, + 289, + 715 + ], + "type": "text", + "content": "); 3) We show that our proposed method achieves strong performance on all metrics, and confirm that human evaluators prefer it over existing methods by wide margins—even surpassing ground truth images in some cases (" + }, + { + "bbox": [ + 47, + 534, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\S\\S 5.3" + }, + { + "bbox": [ + 47, + 534, + 289, + 715 + ], + "type": "text", + "content": " to 5.5); 4) Finally, we showcase new abilities that StackedDiffusion unlocks, including personalization, goal suggestion, and error correction, that go" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 305, + 72, + 523, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 523, + 85 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 523, + 85 + ], + "type": "text", + "content": "far beyond what is possible with fixed articles (§ 5.6)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 96, + 392, + 109 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 96, + 392, + 109 + ], + "spans": [ + { + "bbox": [ + 306, + 96, + 392, + 109 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 117, + 547, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 117, + 547, + 308 + ], + "spans": [ + { + "bbox": [ + 304, + 117, + 547, + 308 + ], + "type": "text", + "content": "Instructional data, tasks, and methods. In the text domain, learning language models on WikiHow [30, 72], has led to advances in tasks such as summarization [71], commonsense procedural knowledge [72, 75], question answering [11], and hierarchical reasoning [74]. In particular, Zhang et al. [72] introduce the goal inference task, in which a model is presented a goal text as input and asked which of 4 candidate steps is one that actually helps achieve that goal, as well as the analogous step inference task. In the multimodal setting, Yang et al. [67] introduce the Visual Goal-Step Inference (VGSI) dataset and task, which consider articles of interleaved text and images. In this task, a model is again presented a goal text as input but is asked which of 4 candidate images is one that actually helps achieve that goal. They show representations learned on this data aid in tasks related to instructional videos [39, 58]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 309, + 547, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 309, + 547, + 417 + ], + "spans": [ + { + "bbox": [ + 304, + 309, + 547, + 417 + ], + "type": "text", + "content": "Learning representations from multimodal data. Multimodal data (including multimodal instructional data, such as from video [39, 58]) has proven to be a powerful source of signal for tasks such as zero-shot recognition [25, 47], text-image and text-video retrieval [3, 9, 16, 17, 34, 36, 37, 40, 60], temporal segmentation [36, 46, 56, 65, 76], activity localization [9, 33, 36, 63, 65, 77], anticipation [12, 15, 19, 52, 61], question-answering [31, 53, 64], summarization [41, 42], and even recipe personalization [14]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 417, + 547, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 417, + 547, + 464 + ], + "spans": [ + { + "bbox": [ + 304, + 417, + 547, + 464 + ], + "type": "text", + "content": "Existing work on instructional data has centered around understanding, rather than generation. We instead focus on the novel setting of generating full multimodal articles complete with text and illustrations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 466, + 548, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 466, + 548, + 634 + ], + "spans": [ + { + "bbox": [ + 304, + 466, + 548, + 634 + ], + "type": "text", + "content": "Generative models. Recent work has examined text-conditioned visual generation through autoregressive [8, 69] or diffusion models [4, 18, 21, 43, 48, 49, 51]. These advances have been leveraged to create text and images together with purely autoregressive [1, 70] or combined approaches [27]. Concurrent work [2, 57] enables the capability of generating multiple images together in the autoregressive framework, but requires substantial additional parameters and focus on creating images that adhere solely to the nearby text rather than enabling consistency. (Similar issues arise in the text-to-video setting; see Appendix 7.) StackedDiffusion, on the other hand, leverages the priors built into the T2I model to obtain consistency without any additional parameters." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 645, + 440, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 645, + 440, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 645, + 440, + 658 + ], + "type": "text", + "content": "3. Illustrated Instructions" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "We now formalize our task and the corresponding desiderata. The input to the system is a goal text " + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{g}" + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": ". As output, we would like to produce step text " + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_i" + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": " as well as step illustration " + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i" + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": " for each step " + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": ". The step text should be a natural language" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "6275" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 286, + 251 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 286, + 251 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 286, + 251 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 286, + 251 + ], + "type": "image", + "image_path": "8bef10bc0808b264215a1b39b4debe4e84955a541d43247c8694557b6a9202d2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 260, + 287, + 316 + ], + "lines": [ + { + "bbox": [ + 46, + 260, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 260, + 287, + 316 + ], + "type": "text", + "content": "Figure 2. Failure modes of a naive approach. A frozen T2I model is not able to capture both the goal and the step, showing only one or the other depending on how it is prompted. Further, it can not produce consistent images, leading to odd changes such as the color of the ice varying between images." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 328, + 286, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 328, + 286, + 352 + ], + "spans": [ + { + "bbox": [ + 46, + 328, + 286, + 352 + ], + "type": "text", + "content": "description of the step, while the step illustrations should be an image corresponding to the step." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 353, + 286, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 353, + 286, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 353, + 286, + 483 + ], + "type": "text", + "content": "On first glance, one might think that this task is a straightforward amalgamation of textual instruction generation and text to image generation. While these tasks are closely related and necessary substeps towards illustrated instructions, they are not sufficient. The naive approach—simply creating step text " + }, + { + "bbox": [ + 46, + 353, + 286, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_i" + }, + { + "bbox": [ + 46, + 353, + 286, + 483 + ], + "type": "text", + "content": " from goal text " + }, + { + "bbox": [ + 46, + 353, + 286, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{g}" + }, + { + "bbox": [ + 46, + 353, + 286, + 483 + ], + "type": "text", + "content": ", and each image " + }, + { + "bbox": [ + 46, + 353, + 286, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i" + }, + { + "bbox": [ + 46, + 353, + 286, + 483 + ], + "type": "text", + "content": " individually from each " + }, + { + "bbox": [ + 46, + 353, + 286, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_i" + }, + { + "bbox": [ + 46, + 353, + 286, + 483 + ], + "type": "text", + "content": "—fails to recognize the fresh new challenges that do not exist in either of these two other tasks. We identify three key requirements for useful illustrated illustrations: goal faithfulness, step faithfulness, and cross-image consistency." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 485, + 286, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 485, + 286, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 485, + 286, + 533 + ], + "type": "text", + "content": "The first requirement is clear: if the images do not relate to the goal, they cannot be good illustrations. This motivates the desiderata of goal faithfulness, which requires that each image faithfully reflects the goal text." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 533, + 286, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 286, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 286, + 628 + ], + "type": "text", + "content": "However, goal faithfulness alone is not enough. Consider the first row of images in Figure 2. The images all reflect the ultimate goal—creating colored ice—but fail in their role as step illustrations. This in turn motivates step faithfulness, which requires that each image be faithful to the step text. We see in this example that baseline T2I models fail dramatically along this metric; every image reflects the goal rather than the specifics of the step requested." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 630, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 286, + 713 + ], + "type": "text", + "content": "Finally, the generated images should be consistent with each other. While the second row of images in Figure 2 are all faithful to the step text, the color of the ice (and even the style of the images, cartoon or real) changes between images. This is jarring and confusing to the reader. This motivates the final criterion of cross-image consistency, which requires that each image be consistent with the other images" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 309, + 75, + 543, + 198 + ], + "blocks": [ + { + "bbox": [ + 309, + 75, + 543, + 198 + ], + "lines": [ + { + "bbox": [ + 309, + 75, + 543, + 198 + ], + "spans": [ + { + "bbox": [ + 309, + 75, + 543, + 198 + ], + "type": "image", + "image_path": "70e6cc6bb7533343b713e5d6e18f8c0cff12d414b7c8c2d9fb6f0164468e4763.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 208, + 545, + 263 + ], + "lines": [ + { + "bbox": [ + 305, + 208, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 305, + 208, + 545, + 263 + ], + "type": "text", + "content": "Figure 3. Overview of StackedDiffusion. At training time, we use the given goal and step text, and stack the encoded ground truth step-images. At inference time, we obtain the goal and step text from an LLM, and unstack denoised latents to produce the output images. See §4 for details and notation." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 274, + 454, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 274, + 454, + 286 + ], + "spans": [ + { + "bbox": [ + 306, + 274, + 454, + 286 + ], + "type": "text", + "content": "produced for a particular generation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 297, + 408, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 297, + 408, + 308 + ], + "spans": [ + { + "bbox": [ + 306, + 297, + 408, + 308 + ], + "type": "text", + "content": "4. StackedDiffusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 316, + 545, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 545, + 518 + ], + "type": "text", + "content": "We propose a new architecture, StackedDiffusion, to overcome the limitations of existing text-to-image (T2I) approaches for the task of generating interleaved text and images for instructional tasks. StackedDiffusion builds upon T2I models based on latent diffusion models (LDMs) [49], which are diffusion models [48] that operate on a low-resolution 'latent' encoding of the original images. Our primary desiderata is that the images generated must be faithful to both the goal and step text. However, in initial experiments we found that simply encoding the goal and step texts joined together (as strings) leads to an uninformative encoding. For instance, generations conditioned on this combined text tended to ignore certain steps or the goal. Furthermore, the length of this combined text will likely exceed the context length limitations of the text encoders [47] commonly used in T2I models, leading to undesired truncation and information loss." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 520, + 545, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 520, + 545, + 675 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 545, + 675 + ], + "type": "text", + "content": "Hence, we elect to use a more general approach, applicable to text encoders with any context length. Rather than encoding the combined goal and step text to obtain the condition, we first separately encode the goal and step texts and then concatenate the encodings. This allows the model to learn to use the goal and step text independently, as well as in combination. We add a 'step-positional encoding' " + }, + { + "bbox": [ + 304, + 520, + 545, + 675 + ], + "type": "inline_equation", + "content": "\\Phi_{i}" + }, + { + "bbox": [ + 304, + 520, + 545, + 675 + ], + "type": "text", + "content": " to this concatenation. It is broadcast across the dimensions pertaining to a particular step, to indicate to the model where a step begins and ends. " + }, + { + "bbox": [ + 304, + 520, + 545, + 675 + ], + "type": "inline_equation", + "content": "\\Phi_0" + }, + { + "bbox": [ + 304, + 520, + 545, + 675 + ], + "type": "text", + "content": " denotes the positional encoding reserved to indicate the goal embedding. Hence, given a text encoder " + }, + { + "bbox": [ + 304, + 520, + 545, + 675 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 304, + 520, + 545, + 675 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 520, + 545, + 675 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 520, + 545, + 675 + ], + "type": "text", + "content": " steps in a given goal, we compute the overall text conditioning " + }, + { + "bbox": [ + 304, + 520, + 545, + 675 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 304, + 520, + 545, + 675 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 346, + 683, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 683, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 346, + 683, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {C} = \\left| \\right| \\left(\\psi (\\mathbf {g}) + \\Phi_ {0}, \\left. \\right| _ {i = 1} ^ {N} \\psi (\\mathbf {s} _ {i}) + \\Phi_ {i}\\right) \\tag {1}", + "image_path": "429908dd1168f7c904c973a5ff7afbc134ae20967627fa34726cbf80e992c906.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6276" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "||" + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": " is the concatenation operation. This is shown in Figure 3 (top). This design decision is critical to obtain good goal and step faithfulness." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 108, + 286, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 108, + 286, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 108, + 286, + 239 + ], + "type": "text", + "content": "In addition, independently generating each image does not achieve the requirement of cross-image consistency as the model cannot exchange information across images. Thus, we generate all the " + }, + { + "bbox": [ + 46, + 108, + 286, + 239 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i" + }, + { + "bbox": [ + 46, + 108, + 286, + 239 + ], + "type": "text", + "content": " images at once, which allows the model to jointly generate the sequence and achieve cross-image consistency. The remaining question, of course, is how to accomplish this simultaneous generation such that it gives rise to cross-image consistency. Our key observation is that T2I models already have a strong prior for consistency within a single image. Could we make use of this previously-learned knowledge?" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 239, + 286, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 286, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 286, + 312 + ], + "type": "text", + "content": "We propose a simple method to accomplish this: spatial tiling. As illustrated in Figure 3, the denoising U-Net is given latents " + }, + { + "bbox": [ + 46, + 239, + 286, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i" + }, + { + "bbox": [ + 46, + 239, + 286, + 312 + ], + "type": "text", + "content": " corresponding to each of the output images simultaneously, tiled spatially as if a single image. At training time, training images are encoded into the latent space as usual, then reshaped into the tiled format. In detail," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 318, + 286, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 318, + 286, + 350 + ], + "spans": [ + { + "bbox": [ + 89, + 318, + 286, + 350 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {i} = \\mathcal {E} \\left(\\mathbf {I} _ {i}\\right), \\quad \\mathbf {z} = \\left. \\right| _ {i = 1} ^ {N} \\mathbf {z} _ {i} \\tag {2}", + "image_path": "4cf0b4e7172754007333e6f456d0612738a79f6c7970e44aea9c5080c6288418.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 71, + 353, + 286, + 373 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 353, + 286, + 373 + ], + "spans": [ + { + "bbox": [ + 71, + 353, + 286, + 373 + ], + "type": "interline_equation", + "content": "L _ {L D M} := \\mathbb {E} _ {\\mathbf {z}, \\epsilon \\sim \\mathcal {N} (0, 1), t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {z} ^ {(t)}, t, \\mathcal {C}) \\| _ {2} ^ {2} \\right] \\tag {3}", + "image_path": "bd328850e372584070d52f3b4161300f5c6b28bf1f97b30552301beaef0157a2.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 379, + 286, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 379, + 286, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 379, + 286, + 568 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 379, + 286, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 46, + 379, + 286, + 568 + ], + "type": "text", + "content": " is the encoder to map the image to the latent space (for instance, using a VAE [26]), " + }, + { + "bbox": [ + 46, + 379, + 286, + 568 + ], + "type": "inline_equation", + "content": "L_{LDM}" + }, + { + "bbox": [ + 46, + 379, + 286, + 568 + ], + "type": "text", + "content": " is the training objective for the LDM, " + }, + { + "bbox": [ + 46, + 379, + 286, + 568 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 379, + 286, + 568 + ], + "type": "text", + "content": " denotes a timestep of the diffusion process, " + }, + { + "bbox": [ + 46, + 379, + 286, + 568 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 379, + 286, + 568 + ], + "type": "text", + "content": " denotes the noise added at a given timestep, and " + }, + { + "bbox": [ + 46, + 379, + 286, + 568 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 379, + 286, + 568 + ], + "type": "text", + "content": " denotes the parameters of the learned denoising U-Net. At inference, we use classifier-free guidance [43] to generate a latent " + }, + { + "bbox": [ + 46, + 379, + 286, + 568 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{z}}" + }, + { + "bbox": [ + 46, + 379, + 286, + 568 + ], + "type": "text", + "content": " given conditioning goal and " + }, + { + "bbox": [ + 46, + 379, + 286, + 568 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 379, + 286, + 568 + ], + "type": "text", + "content": " step texts. We split the latent into " + }, + { + "bbox": [ + 46, + 379, + 286, + 568 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{z}}_1,\\dots,\\tilde{\\mathbf{z}}_N" + }, + { + "bbox": [ + 46, + 379, + 286, + 568 + ], + "type": "text", + "content": ", and decode into generated images using a decoder " + }, + { + "bbox": [ + 46, + 379, + 286, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 379, + 286, + 568 + ], + "type": "text", + "content": " corresponding to the encoder " + }, + { + "bbox": [ + 46, + 379, + 286, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 46, + 379, + 286, + 568 + ], + "type": "text", + "content": ", i.e. " + }, + { + "bbox": [ + 46, + 379, + 286, + 568 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{I}}_i = \\mathcal{D}(\\tilde{\\mathbf{z}}_i)" + }, + { + "bbox": [ + 46, + 379, + 286, + 568 + ], + "type": "text", + "content": ". (We opt for tiling along a single spatial dimension for simplicity, and find that alternative tiling strategies do not significantly affect performance; see Appendix § 8.) We refer the reader to the Appendix § 9 and [49] for further details on LDMs. Given the stacked conditioning and generation operations, we refer to our final model as Stacked-Diffusion." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "text", + "content": "During training, goal and step text are obtained from ground truth, while at inference, a pretrained LLM is used to transform arbitrary user input text to an inferred goal and generated step texts. We describe the LLM inference procedure and prompt engineering in the Appendix § 10. We initialize the U-Net using a pretrained T2I model, and fine-tune all layers when training with the stacked input. Our stacked conditioning only increases the spatial resolution of the input to the U-Net for which can be modeled entirely using the existing parameters (spatial convolutions, attention) of the U-Net. We use the T2I model's text encoder " + }, + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "inline_equation", + "content": "(\\psi)" + }, + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "text", + "content": ", kept frozen, to encode the goal and step texts." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 313, + 69, + 539, + 146 + ], + "blocks": [ + { + "bbox": [ + 313, + 69, + 539, + 146 + ], + "lines": [ + { + "bbox": [ + 313, + 69, + 539, + 146 + ], + "spans": [ + { + "bbox": [ + 313, + 69, + 539, + 146 + ], + "type": "image", + "image_path": "0c73f9e5b449715bef0bcf29490b2d4f9cd7aff6a1231c003decf96a1d2c66f7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 156, + 545, + 190 + ], + "lines": [ + { + "bbox": [ + 305, + 156, + 545, + 190 + ], + "spans": [ + { + "bbox": [ + 305, + 156, + 545, + 190 + ], + "type": "text", + "content": "Figure 4. Illustrated Instructions data. The histogram shows the distribution of step counts in the data. We find that more than " + }, + { + "bbox": [ + 305, + 156, + 545, + 190 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 305, + 156, + 545, + 190 + ], + "type": "text", + "content": " of articles consist of 6 or fewer steps." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 208, + 545, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 208, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 304, + 208, + 545, + 387 + ], + "type": "text", + "content": "Since the spatially stacked latent has a spatial resolution comparable to the ones used for high resolution image generation, we encounter issues observed by prior work [32] for high resolution training. Specifically, [32] finds that high resolution latents retain substantial information about the input being noised with typical noise addition schedules, even at the final diffusion timestep " + }, + { + "bbox": [ + 304, + 208, + 545, + 387 + ], + "type": "inline_equation", + "content": "t = T" + }, + { + "bbox": [ + 304, + 208, + 545, + 387 + ], + "type": "text", + "content": ". This results in an unintentional distribution shift between train time (when all observed inputs contain signal) and test time (when the first inputs are pure noise with no signal). We address this by adjusting the training diffusion noise schedule such that the signal-to-noise ratio (SNR) at the final diffusion timestep " + }, + { + "bbox": [ + 304, + 208, + 545, + 387 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 208, + 545, + 387 + ], + "type": "text", + "content": " is zero [32]. This ensures enough noise is added during training so as to mitigate this difference between train and test time usage." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "text", + "content": "Implementation Details. We build upon a text-to-image latent diffusion model [49] trained on a large in-house image-text dataset. It leverages a VAE [26] to map the images to a 4D latent space, with a " + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "inline_equation", + "content": "8 \\times" + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "text", + "content": " reduction in spatial resolution. The U-Net largely follows the implementation from [49], with minor architectural modifications. As conditioning, it uses the Flan-T5-XXL text encoder [10]. We train Stacked-Diffusion for 14000 steps using the AdamW optimizer [35] with a learning rate of " + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "text", + "content": ", weight decay of 0.01, and gradient clipping at " + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "text", + "content": " norm of 1.0. For classifier-free guidance, we use a conditioning dropout rate of 0.05. We choose to generate at most " + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "inline_equation", + "content": "N = 6" + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "text", + "content": " images simultaneously, because as we will see in § 5.1, the vast majority of the data available has 6 or fewer steps. For shorter training sequences, we pad with empty frames and dummy step texts, and for longer, we drop the extra steps and images. We ablate this choice of " + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "text", + "content": " in § 5.4. Hence in practice, for batch of 8 sets of 6 steps at 256px resolution, we get an encoded latent of " + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "inline_equation", + "content": "8 \\times 6 \\times 4 \\times 32 \\times 32" + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "text", + "content": ". We concatenate the latents spatially into " + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "inline_equation", + "content": "8 \\times 4 \\times (6 * 32) \\times 32" + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "text", + "content": ". The denoising U-Net is finetuned to denoise these stacked images, leveraging what it has previously learned about spatial consistency, and adapts to this new setting. At inference time, steps are generated using a pretrained LLM [44], prompted to generate at most " + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "text", + "content": " steps (" + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\S 10" + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "text", + "content": "). To generate illustrations, noise is sampled in the shape of the tiled latent with " + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "text", + "content": " steps, and the steps generated by the LLM are padded with dummy steps if less" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6277" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 125, + 128, + 205 + ], + "blocks": [ + { + "bbox": [ + 49, + 125, + 128, + 205 + ], + "lines": [ + { + "bbox": [ + 49, + 125, + 128, + 205 + ], + "spans": [ + { + "bbox": [ + 49, + 125, + 128, + 205 + ], + "type": "image", + "image_path": "abdd9c4211ed1cebf5401101ac227efbb6d2e36a7e02f360f2b02270f203e1cb.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 128, + 125, + 206, + 205 + ], + "blocks": [ + { + "bbox": [ + 128, + 125, + 206, + 205 + ], + "lines": [ + { + "bbox": [ + 128, + 125, + 206, + 205 + ], + "spans": [ + { + "bbox": [ + 128, + 125, + 206, + 205 + ], + "type": "image", + "image_path": "b85bd28800215fdabb0f01e6dfd03b75541a101b14794e1866016d50c4120439.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 206, + 126, + 285, + 205 + ], + "blocks": [ + { + "bbox": [ + 206, + 126, + 285, + 205 + ], + "lines": [ + { + "bbox": [ + 206, + 126, + 285, + 205 + ], + "spans": [ + { + "bbox": [ + 206, + 126, + 285, + 205 + ], + "type": "image", + "image_path": "69e3b1946c36b74b5383b6f23d0211d25bb483f9edc8239e70cb340ac12e54d1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 49, + 205, + 127, + 285 + ], + "blocks": [ + { + "bbox": [ + 49, + 205, + 127, + 285 + ], + "lines": [ + { + "bbox": [ + 49, + 205, + 127, + 285 + ], + "spans": [ + { + "bbox": [ + 49, + 205, + 127, + 285 + ], + "type": "image", + "image_path": "ce3a95db7a595c6bade73f555e74213878dd23564813b3b498a0becd576fa658.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 294, + 287, + 350 + ], + "lines": [ + { + "bbox": [ + 46, + 294, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 287, + 350 + ], + "type": "text", + "content": "Figure 5. Metrics. We introduce three metrics, one for each desideratum presented in § 3. Goal faithfulness: the second image does not show muffins. Step faithfulness: the second image does not show the step. Cross-image consistency: the second image shows a different number of meatballs with different visuals." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 128, + 205, + 206, + 285 + ], + "blocks": [ + { + "bbox": [ + 128, + 205, + 206, + 285 + ], + "lines": [ + { + "bbox": [ + 128, + 205, + 206, + 285 + ], + "spans": [ + { + "bbox": [ + 128, + 205, + 206, + 285 + ], + "type": "image", + "image_path": "0fedcf6895c2292a06f1ceb20ffba2a1d5784ef75c7c2a170c7501cdd6f60a36.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 206, + 205, + 285, + 284 + ], + "blocks": [ + { + "bbox": [ + 206, + 205, + 285, + 284 + ], + "lines": [ + { + "bbox": [ + 206, + 205, + 285, + 284 + ], + "spans": [ + { + "bbox": [ + 206, + 205, + 285, + 284 + ], + "type": "image", + "image_path": "c1e4cf8496c64bbf680b983e316127bd5eeba422f3f6da5bf10871fd0a4ad473.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 361, + 287, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 287, + 399 + ], + "type": "text", + "content": "than " + }, + { + "bbox": [ + 46, + 361, + 287, + 399 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 361, + 287, + 399 + ], + "type": "text", + "content": ". The resulting noise is denoised using the U-Net as usual, and finally reshaped to obtain the " + }, + { + "bbox": [ + 46, + 361, + 287, + 399 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 361, + 287, + 399 + ], + "type": "text", + "content": " output images upon decoding. See Appendix § 15 for further details." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 411, + 128, + 424 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 411, + 128, + 424 + ], + "spans": [ + { + "bbox": [ + 47, + 411, + 128, + 424 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 432, + 287, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 432, + 287, + 516 + ], + "spans": [ + { + "bbox": [ + 46, + 432, + 287, + 516 + ], + "type": "text", + "content": "We train and evaluate StackedDiffusion on web-based instructional data. We compare the generations to multiple established baselines and ground truth, using automatic metrics and human evaluations. We now describe the data, metrics, baselines, and the key results. Finally, we demonstrate some new applications that StackedDiffusion enables, showing how it goes beyond standard instructional articles." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 47, + 525, + 217, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 525, + 217, + 538 + ], + "spans": [ + { + "bbox": [ + 47, + 525, + 217, + 538 + ], + "type": "text", + "content": "5.1. Illustrated Instructions Dataset" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 544, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 652 + ], + "type": "text", + "content": "We introduce a new dataset to train and evaluate models for the Illustrated Instructions task, by repurposing the Visual Goal-Step Inference (VGSI) dataset [67]. VGSI consists of WikiHow articles, each of which have a high-level goal, 6-10 natural language steps (see Figure 4), and associated image illustrations. We observe that the same data can be used for generating instructional articles. It can provide signal for how the goal and step texts should map to output images, and how images should match with each other." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "content": "For evaluation, we construct a held-out set from this same data. Used directly, however, the data is not well suited as many of the tasks are too high-level to be useful for consistent illustration. For instance, \"How to Start a Business in North Carolina\" may have steps \"Brainstorm" + } + ] + } + ], + "index": 19 + }, + { + "type": "table", + "bbox": [ + 307, + 83, + 548, + 137 + ], + "blocks": [ + { + "bbox": [ + 307, + 83, + 548, + 137 + ], + "lines": [ + { + "bbox": [ + 307, + 83, + 548, + 137 + ], + "spans": [ + { + "bbox": [ + 307, + 83, + 548, + 137 + ], + "type": "table", + "html": "
T2I (Frozen)22.092.943.251.369.3
T2I (Finetuned)33.378.852.451.553.5
StackedDiffusion(ref)74.361.550.739.5
Ground Truth82.581.773.750.6(N/A)
", + "image_path": "5669acf8d2a3b2761a8c7acd1772b0752af26497ee99e4bfba20eab8f209707e.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 146, + 547, + 190 + ], + "lines": [ + { + "bbox": [ + 305, + 146, + 547, + 190 + ], + "spans": [ + { + "bbox": [ + 305, + 146, + 547, + 190 + ], + "type": "text", + "content": "Table 1. Comparison to baselines. Human evaluation is reported as win rate vs our full StackedDiffusion model. GF corresponds to goal faithfulness accuracy, SF to step faithfulness accuracy, CIC to cross-image consistency, and FID to Fréchet Inception Distance." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 213, + 546, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 213, + 546, + 274 + ], + "spans": [ + { + "bbox": [ + 305, + 213, + 546, + 274 + ], + "type": "text", + "content": "ideas” and “Go to the courthouse” that have no shared visual content. As such, we filter the data for evaluation to the “Recipes” category, which always has a visually clear end state and where illustrations must be consistent with each other. Please see Appendix § 11 for more details." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 284, + 364, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 284, + 364, + 295 + ], + "spans": [ + { + "bbox": [ + 306, + 284, + 364, + 295 + ], + "type": "text", + "content": "5.2. Metrics" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 303, + 546, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 303, + 546, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 303, + 546, + 422 + ], + "type": "text", + "content": "Having established three desiderata in § 3, we now turn to the question of how to evaluate them. In addition to evaluating the quality of images using FID [22], we propose three metrics, one for each desideratum, that can be used to evaluate the faithfulness and consistency of the generated article. Finally, given the limitations of automatic metrics for evaluating generative modeling tasks [55], we use human evaluations as our primary metric for overall quality. We illustrate the metrics in Figure 5 and briefly describe them next. Please see the Appendix § 12 for more details." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 424, + 547, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 424, + 547, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 424, + 547, + 593 + ], + "type": "text", + "content": "Goal Faithfulness (GF) measures how well the generated image is associated to the goal text. We evaluate this by constructing multiple-choice questions (MCQ) as in VGSI [67]. For each generated image, we compare its CLIP similarity [47] with the correct goal text vs the similarity with the texts of three other randomly selected goals. We compute the accuracy of the model in choosing the correct goal text. Step Faithfulness (SF) measures how faithfully the generated image illustrates the step it aims to depict. An image should match the text for the step it was made for, more than other steps. We measure this using CLIP similarity and a MCQ task similar to goal faithfulness, where the image should have higher CLIP similarity with the corresponding step text than the other step texts within the same goal." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": "Cross-Image Consistency (CIC) evaluates how consistent the generated images for a goal are with each other and penalizes jarring inconsistencies across the images, such as objects changing color or number. For instance, if a particular set of ingredients are shown for \"gather the ingredients,\" then the same ingredients should be shown for \"mix the dry ingredients.\" We measure this by computing the average " + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": " distance between DINO [7] embeddings of the images for each step, as this considers the similarity of visual (rather than purely semantic) features. Like prior work [50], we" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 59, + 73, + 118, + 83 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 73, + 118, + 83 + ], + "spans": [ + { + "bbox": [ + 59, + 73, + 118, + 83 + ], + "type": "text", + "content": "Goal Faithfulness" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 137, + 74, + 196, + 83 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 74, + 196, + 83 + ], + "spans": [ + { + "bbox": [ + 137, + 74, + 196, + 83 + ], + "type": "text", + "content": "Step Faithfulness" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 202, + 74, + 286, + 83 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 74, + 286, + 83 + ], + "spans": [ + { + "bbox": [ + 202, + 74, + 286, + 83 + ], + "type": "text", + "content": "Cross-Image Consistency" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 91, + 124, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 124, + 101 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 124, + 101 + ], + "type": "text", + "content": "Goal: Baking Muffins" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 129, + 92, + 205, + 110 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 92, + 205, + 110 + ], + "spans": [ + { + "bbox": [ + 129, + 92, + 205, + 110 + ], + "type": "text", + "content": "Step: Put the muffins in the oven" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 211, + 92, + 239, + 110 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 92, + 239, + 110 + ], + "spans": [ + { + "bbox": [ + 211, + 92, + 239, + 110 + ], + "type": "text", + "content": "Previous Image:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 246, + 83, + 286, + 126 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 83, + 286, + 126 + ], + "spans": [ + { + "bbox": [ + 246, + 83, + 286, + 126 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 378, + 70, + 546, + 82 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 70, + 546, + 82 + ], + "spans": [ + { + "bbox": [ + 378, + 70, + 546, + 82 + ], + "type": "text", + "content": "Human " + }, + { + "bbox": [ + 378, + 70, + 546, + 82 + ], + "type": "inline_equation", + "content": "(\\uparrow)" + }, + { + "bbox": [ + 378, + 70, + 546, + 82 + ], + "type": "text", + "content": " GF " + }, + { + "bbox": [ + 378, + 70, + 546, + 82 + ], + "type": "inline_equation", + "content": "(\\uparrow)" + }, + { + "bbox": [ + 378, + 70, + 546, + 82 + ], + "type": "text", + "content": " SF " + }, + { + "bbox": [ + 378, + 70, + 546, + 82 + ], + "type": "inline_equation", + "content": "(\\uparrow)" + }, + { + "bbox": [ + 378, + 70, + 546, + 82 + ], + "type": "text", + "content": " CIC " + }, + { + "bbox": [ + 378, + 70, + 546, + 82 + ], + "type": "inline_equation", + "content": "(\\downarrow)" + }, + { + "bbox": [ + 378, + 70, + 546, + 82 + ], + "type": "text", + "content": " FID " + }, + { + "bbox": [ + 378, + 70, + 546, + 82 + ], + "type": "inline_equation", + "content": "(\\downarrow)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6278" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 79, + 286, + 229 + ], + "blocks": [ + { + "bbox": [ + 111, + 71, + 224, + 79 + ], + "lines": [ + { + "bbox": [ + 111, + 71, + 224, + 79 + ], + "spans": [ + { + "bbox": [ + 111, + 71, + 224, + 79 + ], + "type": "text", + "content": "Goal: How to Bake Turkey Wings" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 51, + 79, + 286, + 229 + ], + "lines": [ + { + "bbox": [ + 51, + 79, + 286, + 229 + ], + "spans": [ + { + "bbox": [ + 51, + 79, + 286, + 229 + ], + "type": "image", + "image_path": "8c7902d9d52dc15231ab9b688dcca9b7b887bd76069f368d1e1ce7da4c184769.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 238, + 287, + 293 + ], + "lines": [ + { + "bbox": [ + 46, + 238, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 287, + 293 + ], + "type": "text", + "content": "Figure 6. Baseline comparison. StackedDiffusion images are preferred overwhelmingly over baselines. The frozen baseline tends to only produce images showing the goal, while the finetuned baseline produces images that are more faithful to the step text, but have no visual features in common." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 304, + 287, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 287, + 376 + ], + "type": "text", + "content": "also found that CLIP features do not work well for image similarity as they are invariant to critical aspects such as color, number of objects, style etc. DINO's self-supervised pretraining objective leads to features that are sensitive to the visual aspects particular to an image, rather than being invariant to aspects not captured by category or language." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 376, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 287, + 495 + ], + "type": "text", + "content": "Human Evaluation is finally used to confirm that the automated evaluation corresponds to actual human preferences. We provide fully rendered articles from each model to human evaluators on Amazon Mechanical Turk (AMT), and ask them to choose which article they prefer, or if they are tied. We then consider the win rate as the proportion for which a majority of evaluators selected the given method compared to the articles produced by StackedDiffusion. A win rate of 50 denotes the methods are tied, whereas " + }, + { + "bbox": [ + 46, + 376, + 287, + 495 + ], + "type": "inline_equation", + "content": "< 50" + }, + { + "bbox": [ + 46, + 376, + 287, + 495 + ], + "type": "text", + "content": " denotes the method is worse than StackedDiffusion." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 503, + 113, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 503, + 113, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 503, + 113, + 514 + ], + "type": "text", + "content": "5.3. Baselines" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 522, + 287, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 556 + ], + "type": "text", + "content": "We now describe some baseline approaches based on existing state-of-the-art image generation models, and compare them to StackedDiffusion in Table 1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 557, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 287, + 653 + ], + "type": "text", + "content": "T2I (frozen): The most obvious choice for a baseline is to simply use a pretrained and frozen text-to-image (T2I) model. We use our in-house T2I model that is also used to initialize StackedDiffusion. To ensure this baseline gets same information as StackedDiffusion, we prompt the model with a concatenation of the goal with each step, " + }, + { + "bbox": [ + 46, + 557, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{g}||\\mathbf{s}_i" + }, + { + "bbox": [ + 46, + 557, + 287, + 653 + ], + "type": "text", + "content": ", for each " + }, + { + "bbox": [ + 46, + 557, + 287, + 653 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 557, + 287, + 653 + ], + "type": "text", + "content": ", and produce a single image " + }, + { + "bbox": [ + 46, + 557, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i" + }, + { + "bbox": [ + 46, + 557, + 287, + 653 + ], + "type": "text", + "content": ". The sequence then is composed of " + }, + { + "bbox": [ + 46, + 557, + 287, + 653 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 557, + 287, + 653 + ], + "type": "text", + "content": " independent generations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "content": "We find that these images are faithful to the goal text, but fail to be faithful to the step text and lack consistency. The goal faithfulness is substantially higher than even the ground truth (92.9% vs 81.7%). This is because the model is not capable of reasoning about what a particular step to-" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 69, + 538, + 171 + ], + "blocks": [ + { + "bbox": [ + 316, + 69, + 538, + 171 + ], + "lines": [ + { + "bbox": [ + 316, + 69, + 538, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 69, + 538, + 171 + ], + "type": "image", + "image_path": "d01c9dd6964ee9155412d9b38e95567509bba414f3c9854c8639419d5446d72f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 182, + 545, + 226 + ], + "lines": [ + { + "bbox": [ + 305, + 182, + 545, + 226 + ], + "spans": [ + { + "bbox": [ + 305, + 182, + 545, + 226 + ], + "type": "text", + "content": "Figure 7. Effect of training data. We find that having more training data improves StackedDiffusion's faithfulness and consistency. However even with " + }, + { + "bbox": [ + 305, + 182, + 545, + 226 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 305, + 182, + 545, + 226 + ], + "type": "text", + "content": " data it performs well, thanks to its effective use of pretrained T2I model weights." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 239, + 545, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 239, + 545, + 311 + ], + "spans": [ + { + "bbox": [ + 304, + 239, + 545, + 311 + ], + "type": "text", + "content": "wards a goal should look like. Instead, it simply generates an image that matches words in the input text, as illustrated in the first row of Figure 2. This results in all the images for e.g. \"Make Colored Ice\" showing the actual finished ice rather than any of the intermediate steps. In other words, the step text is eclipsed by the goal text." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 312, + 546, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 312, + 546, + 444 + ], + "spans": [ + { + "bbox": [ + 304, + 312, + 546, + 444 + ], + "type": "text", + "content": "T2I (finetuned): We finetune a T2I model with the goal and step text embeddings concatenated together as the condition, again producing " + }, + { + "bbox": [ + 304, + 312, + 546, + 444 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 312, + 546, + 444 + ], + "type": "text", + "content": " independent generations. We find that this model is more faithful to the step text than the frozen T2I model, but still substantially less than our final model (52.4% vs 61.5%). This suggests the goal still overshadows the step text, but to a lesser extent. The goal faithfulness is also lower than for the frozen T2I model, as this model is trained to create images for steps rather than goals. The cross-image consistency is low, similar to the frozen model, as the images are still generated independently." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 444, + 546, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 444, + 546, + 660 + ], + "spans": [ + { + "bbox": [ + 304, + 444, + 546, + 660 + ], + "type": "text", + "content": "StackedDiffusion achieves substantially improved step faithfulness and cross-image consistency, while maintaining high goal faithfulness. We show an illustrative example in Figure 6. Unlike the baselines, the cross-image consistency is very close to that of the ground truth data; as generating them together allows for influence between them during the generation process. We also find that the FID of our generated images is substantially lower than that of other models, indicating that our generated images are more similar to the ground truth data. Finally on human evaluations, StackedDiffusion clearly outperforms them both by a preference of " + }, + { + "bbox": [ + 304, + 444, + 546, + 660 + ], + "type": "inline_equation", + "content": "78\\%" + }, + { + "bbox": [ + 304, + 444, + 546, + 660 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 444, + 546, + 660 + ], + "type": "inline_equation", + "content": "66.6\\%" + }, + { + "bbox": [ + 304, + 444, + 546, + 660 + ], + "type": "text", + "content": " respectively. Perhaps most notably, when compared to the ground truth images, human evaluators still picked StackedDiffusion " + }, + { + "bbox": [ + 304, + 444, + 546, + 660 + ], + "type": "inline_equation", + "content": "18.5\\%" + }, + { + "bbox": [ + 304, + 444, + 546, + 660 + ], + "type": "text", + "content": " of the times, suggesting that our model generates some illustrations even better than the ground truth data. Further baselines can be found in Appendix § 13, with a comparison of training and inference costs in Appendix § 14." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 670, + 373, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 373, + 681 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 373, + 681 + ], + "type": "text", + "content": "5.4. Ablations" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Step Count. As discussed in § 4, we elect to use a maximum step count of 6 as it allows sufficient data coverage" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6279" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 287, + 126 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 287, + 126 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 287, + 126 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 287, + 126 + ], + "type": "table", + "html": "
Win RateWin Rate
T2I (Frozen)33.8T2I (Finetuned)34.0
GILL [27]++3.9LLM + CLIP15.6
Goal Retrieval33.1Ground Truth70.0
", + "image_path": "f040870fd95e35069d76d20ae071203208224b8ec315b36438d5b3f58474fcbb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 186, + 288, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 186, + 288, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 186, + 288, + 258 + ], + "type": "text", + "content": "and most articles fall under this value. We find that human evaluators prefer this model to the shorter step count model of length 4. It is slightly preferred over the model trained with a longer step count of 8, likely as there is insufficient data of long lengths, limiting any advantage that might be gained in longer length generations." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 90, + 266, + 245, + 297 + ], + "blocks": [ + { + "bbox": [ + 46, + 134, + 287, + 179 + ], + "lines": [ + { + "bbox": [ + 46, + 134, + 287, + 179 + ], + "spans": [ + { + "bbox": [ + 46, + 134, + 287, + 179 + ], + "type": "text", + "content": "Table 2. System-level comparison to prior work. Human evaluation results using fully generated outputs. Reported as human evaluation win rate of the associated method over StackedDiffusion (hence, higher is better for the reported method)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 90, + 266, + 245, + 297 + ], + "lines": [ + { + "bbox": [ + 90, + 266, + 245, + 297 + ], + "spans": [ + { + "bbox": [ + 90, + 266, + 245, + 297 + ], + "type": "table", + "html": "
Baseline4 Step (↑)8 Step (↑)
6 Step32.146.9
", + "image_path": "bb43af274d1595779da6b55e8c75dab6c1c33926e5ba89acc96b43d19bae226b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 304, + 287, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 287, + 376 + ], + "type": "text", + "content": "Training data. In Figure 7, we randomly sub-sample varying proportions of the total data, and train StackedDiffusion on each of those subsets. As the amount of training data increases, it generates more faithful and consistent images. However even with less data it performs well, owing to its effective use of the pretrained T2I model's initialization." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 376, + 287, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 287, + 424 + ], + "type": "text", + "content": "Importance of 0SNR. We experiment with the 0SNR technique [32] due to the high spatial dimensionality of the latents. We find that not using 0SNR results in substantially reduced step faithfulness, dropping from " + }, + { + "bbox": [ + 46, + 376, + 287, + 424 + ], + "type": "inline_equation", + "content": "61.5\\%" + }, + { + "bbox": [ + 46, + 376, + 287, + 424 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 376, + 287, + 424 + ], + "type": "inline_equation", + "content": "52.8\\%" + }, + { + "bbox": [ + 46, + 376, + 287, + 424 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 424, + 287, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 424, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 46, + 424, + 287, + 496 + ], + "type": "text", + "content": "Importance of Step-Positional Encoding. We also evaluate the importance of the step-positional encoding, " + }, + { + "bbox": [ + 46, + 424, + 287, + 496 + ], + "type": "inline_equation", + "content": "\\Phi_{i}" + }, + { + "bbox": [ + 46, + 424, + 287, + 496 + ], + "type": "text", + "content": ". This gives the model a critical cue for which parts of the condition correspond to which steps. We find that not using the step-positional encoding results in substantially reduced step faithfulness as well, dropping from " + }, + { + "bbox": [ + 46, + 424, + 287, + 496 + ], + "type": "inline_equation", + "content": "61.5\\%" + }, + { + "bbox": [ + 46, + 424, + 287, + 496 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 424, + 287, + 496 + ], + "type": "inline_equation", + "content": "49.8\\%" + }, + { + "bbox": [ + 46, + 424, + 287, + 496 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 503, + 194, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 503, + 194, + 516 + ], + "spans": [ + { + "bbox": [ + 47, + 503, + 194, + 516 + ], + "type": "text", + "content": "5.5. Comparison to prior work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 522, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 653 + ], + "type": "text", + "content": "The previous metrics were computed with respect to fixed ground truth goal and step texts to provide direct comparisons. However, they do not encompass the full scope of our model's capabilities: the flexibility of the generated text is one of StackedDiffusion's core strengths. To evaluate in a closer setting to this real-world usage, we examine the quality of the articles generated by the full system, where the steps are generated with a LLM [44]. We perform system-level comparisons using human evaluators, including the quality of the generated text and the quality of the images created from the generated text." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "We show our results in Table 2. We first compare to the baselines introduced in § 5.3, and see results similar to Table 1. We see that even with LLM-generated text, Stacked-Diffusion outperforms baseline T2I based approaches, either frozen or finetuned, by more than " + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "66\\%" + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": " win-rate." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 545, + 102 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 545, + 102 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 545, + 102 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 545, + 102 + ], + "type": "image", + "image_path": "47d8b68fa6dad61088500ed14d3c77f59154410ea0b9c57e3bf7ef12f1180920.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 310, + 126, + 367, + 182 + ], + "blocks": [ + { + "bbox": [ + 310, + 105, + 528, + 125 + ], + "lines": [ + { + "bbox": [ + 310, + 105, + 528, + 125 + ], + "spans": [ + { + "bbox": [ + 310, + 105, + 528, + 125 + ], + "type": "text", + "content": "Step 1: Drain and press tofu to remove excess water. Step 2: Crumble tofu; mix with bread crumbs and a binding agent. Step 3: Shape tofu mixture into burger patties. Step 4: Cook over medium heat until golden brown." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 310, + 126, + 367, + 182 + ], + "lines": [ + { + "bbox": [ + 310, + 126, + 367, + 182 + ], + "spans": [ + { + "bbox": [ + 310, + 126, + 367, + 182 + ], + "type": "image", + "image_path": "30e376064a07d20776c2aed48e05562dd3559b4e43dbd3f254616a1f596185ad.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 193, + 545, + 215 + ], + "lines": [ + { + "bbox": [ + 306, + 193, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 306, + 193, + 545, + 215 + ], + "type": "text", + "content": "Figure 8. Personalization. StackedDiffusion enables instructions personalized far beyond what is possible with preexisting articles." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 370, + 126, + 425, + 182 + ], + "blocks": [ + { + "bbox": [ + 370, + 126, + 425, + 182 + ], + "lines": [ + { + "bbox": [ + 370, + 126, + 425, + 182 + ], + "spans": [ + { + "bbox": [ + 370, + 126, + 425, + 182 + ], + "type": "image", + "image_path": "2348732110f808aca1b589d07bbf4efcca036cfb0c16289a20171014c1d884d4.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 428, + 126, + 485, + 182 + ], + "blocks": [ + { + "bbox": [ + 428, + 126, + 485, + 182 + ], + "lines": [ + { + "bbox": [ + 428, + 126, + 485, + 182 + ], + "spans": [ + { + "bbox": [ + 428, + 126, + 485, + 182 + ], + "type": "image", + "image_path": "ec4155dfeef0eac5a21e03881668f562e1414ec8874f7671ca5b099fa54a8bd3.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 486, + 126, + 542, + 182 + ], + "blocks": [ + { + "bbox": [ + 486, + 126, + 542, + 145 + ], + "lines": [ + { + "bbox": [ + 486, + 126, + 542, + 145 + ], + "spans": [ + { + "bbox": [ + 486, + 126, + 542, + 145 + ], + "type": "inline_equation", + "content": "\\left( {x - {2x}}\\right) t - x{y}^{2} = \\left( {x - {2x}}\\right) {f}^{\\prime }t" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 486, + 126, + 542, + 182 + ], + "lines": [ + { + "bbox": [ + 486, + 126, + 542, + 182 + ], + "spans": [ + { + "bbox": [ + 486, + 126, + 542, + 182 + ], + "type": "image", + "image_path": "d177b9569dac71da510a5c24b37d1bad26592d5be34789f6db9a26cc052bfcbd.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 232, + 546, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 232, + 546, + 507 + ], + "spans": [ + { + "bbox": [ + 304, + 232, + 546, + 507 + ], + "type": "text", + "content": "Additionally, we compare StackedDiffusion to recent state-of-the-art approaches in multimodal generation, as in principle those can also produce instructional text combined with illustrations. Specifically, we compare to GILL [29]-- a model trained to generate sequences of interleaved text and images. While other similar approaches have been proposed [1, 28], GILL is the best open-source model we could access. Using GILL directly, however, results in a human evaluation win rate of " + }, + { + "bbox": [ + 304, + 232, + 546, + 507 + ], + "type": "inline_equation", + "content": "0\\%" + }, + { + "bbox": [ + 304, + 232, + 546, + 507 + ], + "type": "text", + "content": ". We posit this was for three primary reasons. First, the GILL model was unable to generate long enough generations to suffice as an article in comparison to our generations, despite our best efforts in modifying the generation parameters. Secondly, the model often did not produce any illustrations with the text, although we tried various prompting tactics. Thirdly, the text or images produced were often not of high enough quality to actually reflect the goal or steps. We thus introduce an alternative, which we call GILL++. This uses the (superior) text produced by the same language model as in our method, but passes each step text as input to GILL to generate the corresponding image. This results in a human evaluation win rate of " + }, + { + "bbox": [ + 304, + 232, + 546, + 507 + ], + "type": "inline_equation", + "content": "3.9\\%" + }, + { + "bbox": [ + 304, + 232, + 546, + 507 + ], + "type": "text", + "content": " against StackedDiffusion. See Appendix § 16 for details on prompting GILL, and our GILL++ baseline." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "type": "text", + "content": "Next, we compare StackedDiffusion to a retrieval based approach, denoted by LLM+CLIP. Here we retrieve the closest image in the training set according to CLIP similarity to each of the steps created by the LLM (with each concatenated to the goal text, similar to how the frozen T2I model is used). This retrieval-based metric is not suitable for the automated metrics previously introduced as it is based on the same underlying similarity scores as the automated metrics, but we introduce it here for human evaluation. We find that StackedDiffusion is preferred overwhelmingly over this retrieval-based approach, despite the retrieval-based approach using real images from the training set. Similarly, we compare to another retrieval-based approach which retrieves the full article with the most similar goal text to the input goal, denoted Goal Retrieval. We find StackedDiffusion strongly outperforms this as well. These results together suggest that StackedDiffusion is able" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6280" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 72, + 67, + 91 + ], + "blocks": [ + { + "bbox": [ + 51, + 72, + 67, + 91 + ], + "lines": [ + { + "bbox": [ + 51, + 72, + 67, + 91 + ], + "spans": [ + { + "bbox": [ + 51, + 72, + 67, + 91 + ], + "type": "image", + "image_path": "d4b653e492533a300057bdbe939d13e17d48e00c7e0b6735708fe389635afc0b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 72, + 74, + 274, + 91 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 74, + 274, + 91 + ], + "spans": [ + { + "bbox": [ + 72, + 74, + 274, + 91 + ], + "type": "text", + "content": "I have a bunch of apples left from apple picking, what can I do with them?" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 94, + 251, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 94, + 251, + 111 + ], + "spans": [ + { + "bbox": [ + 54, + 94, + 251, + 111 + ], + "type": "text", + "content": "If you have a lot of apples, you can turn them into a delicious apple crisp. Here's a simple recipe to get you started:" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 267, + 92, + 284, + 112 + ], + "blocks": [ + { + "bbox": [ + 267, + 92, + 284, + 112 + ], + "lines": [ + { + "bbox": [ + 267, + 92, + 284, + 112 + ], + "spans": [ + { + "bbox": [ + 267, + 92, + 284, + 112 + ], + "type": "image", + "image_path": "e00d1a3038a1f738d21b711761e3aa64b83c497911abf45e70133c1ba0ae463f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 116, + 121, + 134 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 116, + 121, + 134 + ], + "spans": [ + { + "bbox": [ + 54, + 116, + 121, + 134 + ], + "type": "text", + "content": "Step 1: Preheat oven to " + }, + { + "bbox": [ + 54, + 116, + 121, + 134 + ], + "type": "inline_equation", + "content": "350^{\\circ}\\mathrm{F}" + }, + { + "bbox": [ + 54, + 116, + 121, + 134 + ], + "type": "inline_equation", + "content": "(175^{\\circ}\\mathrm{C})" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 116, + 200, + 142 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 200, + 142 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 200, + 142 + ], + "type": "text", + "content": "Step 2: Peel, core, and slice apples; place in a baking dish." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 211, + 116, + 281, + 149 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 116, + 281, + 149 + ], + "spans": [ + { + "bbox": [ + 211, + 116, + 281, + 149 + ], + "type": "text", + "content": "Step 3: Mix sugar, cinnamon, and a pinch of salt; sprinkle over apples." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 52, + 149, + 127, + 224 + ], + "blocks": [ + { + "bbox": [ + 52, + 149, + 127, + 224 + ], + "lines": [ + { + "bbox": [ + 52, + 149, + 127, + 224 + ], + "spans": [ + { + "bbox": [ + 52, + 149, + 127, + 224 + ], + "type": "image", + "image_path": "5f76b982470e95c1a04b54ce7ae0f0afdaba7682efdab2ea819e9352c3bca48a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 130, + 149, + 205, + 224 + ], + "blocks": [ + { + "bbox": [ + 130, + 149, + 205, + 224 + ], + "lines": [ + { + "bbox": [ + 130, + 149, + 205, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 205, + 224 + ], + "type": "image", + "image_path": "b1c1e3407c5aa95a720bcb4c1c6c5943e545d72d586d640120c8b8f2eafa8863.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 133, + 225, + 202, + 249 + ], + "lines": [ + { + "bbox": [ + 133, + 225, + 202, + 249 + ], + "spans": [ + { + "bbox": [ + 133, + 225, + 202, + 249 + ], + "type": "text", + "content": "Step 5: Bake for 45 minutes or until golden brown." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 209, + 149, + 283, + 224 + ], + "blocks": [ + { + "bbox": [ + 209, + 149, + 283, + 224 + ], + "lines": [ + { + "bbox": [ + 209, + 149, + 283, + 224 + ], + "spans": [ + { + "bbox": [ + 209, + 149, + 283, + 224 + ], + "type": "image", + "image_path": "b05c9784cb602e72512f7c6d59ede22c192a020dd1bd392ecadebc28f2d9f715.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 211, + 225, + 276, + 234 + ], + "lines": [ + { + "bbox": [ + 211, + 225, + 276, + 234 + ], + "spans": [ + { + "bbox": [ + 211, + 225, + 276, + 234 + ], + "type": "text", + "content": "Step 6: Serve warm." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 52, + 258, + 127, + 333 + ], + "blocks": [ + { + "bbox": [ + 54, + 225, + 125, + 258 + ], + "lines": [ + { + "bbox": [ + 54, + 225, + 125, + 258 + ], + "spans": [ + { + "bbox": [ + 54, + 225, + 125, + 258 + ], + "type": "text", + "content": "Step 4: Combine oats, flour, brown sugar, and butter; crumble over apples." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 52, + 258, + 127, + 333 + ], + "lines": [ + { + "bbox": [ + 52, + 258, + 127, + 333 + ], + "spans": [ + { + "bbox": [ + 52, + 258, + 127, + 333 + ], + "type": "image", + "image_path": "157a7442a98421ef62dc2388a1be819b14732a906eea65091a219119a50cdc22.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 342, + 287, + 366 + ], + "lines": [ + { + "bbox": [ + 47, + 342, + 287, + 366 + ], + "spans": [ + { + "bbox": [ + 47, + 342, + 287, + 366 + ], + "type": "text", + "content": "Figure 9. Goal suggestion and inference. StackedDiffusion can suggest what goal might be most relevant given other information." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 130, + 258, + 205, + 333 + ], + "blocks": [ + { + "bbox": [ + 130, + 258, + 205, + 333 + ], + "lines": [ + { + "bbox": [ + 130, + 258, + 205, + 333 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 205, + 333 + ], + "type": "image", + "image_path": "1321800589acd81068bc2c5de5b02f49a33f548f8744f3bf84dfb6aca7cd8204.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 209, + 258, + 283, + 333 + ], + "blocks": [ + { + "bbox": [ + 209, + 258, + 283, + 333 + ], + "lines": [ + { + "bbox": [ + 209, + 258, + 283, + 333 + ], + "spans": [ + { + "bbox": [ + 209, + 258, + 283, + 333 + ], + "type": "image", + "image_path": "9c76d104f77fa03d55ab031c6807146db39a5db9c5d18bc3da885a9bdaa21beb.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 376, + 287, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 287, + 400 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 287, + 400 + ], + "type": "text", + "content": "to generate novel instructional articles that one can not simply retrieve from the training corpus." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 400, + 287, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 400, + 287, + 519 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 287, + 519 + ], + "type": "text", + "content": "Finally, we compare StackedDiffusion to the ground truth. Surprisingly again, we find that the human annotators pick our generations " + }, + { + "bbox": [ + 46, + 400, + 287, + 519 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 46, + 400, + 287, + 519 + ], + "type": "text", + "content": " of times even compared to ground truth. Note that these articles are handwritten and manually illustrated for the purpose of illustrating these goals. This shows the strong promise of our proposed approach. We believe with the rapid improvements in training data and generative modeling techniques, future iterations of our model could be indistinguishable or even better than manually created illustrated instructional articles." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 47, + 527, + 129, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 527, + 129, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 527, + 129, + 540 + ], + "type": "text", + "content": "5.6. Applications" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 545, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 617 + ], + "type": "text", + "content": "Perhaps the biggest strength of a generative approach is its ability to handle unique user queries that may not fit into the boundaries posed by static articles on the web. Hence, we demonstrate the capabilities of the full system, combining the strengths of text generation capabilities from the language model with StackedDiffusion." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 618, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 287, + 689 + ], + "type": "text", + "content": "Personalized instruction. A user can provide any situational information specific to their circumstances and obtain an article customized to that situation. This is not possible with fixed, existing articles. For example, in Figure 8, the user can specify a diet and obtain an article for the food they want that adheres to that diet." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "Goal suggestion. As StackedDiffusion accepts flexible input text, a user can provide higher-level information about" + } + ] + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 310, + 72, + 328, + 90 + ], + "blocks": [ + { + "bbox": [ + 310, + 72, + 328, + 90 + ], + "lines": [ + { + "bbox": [ + 310, + 72, + 328, + 90 + ], + "spans": [ + { + "bbox": [ + 310, + 72, + 328, + 90 + ], + "type": "image", + "image_path": "8cad8b81aae740759c89e121102a47a3149f97aaa02e952b7f0084b992639edb.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "bbox": [ + 331, + 72, + 538, + 90 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 72, + 538, + 90 + ], + "spans": [ + { + "bbox": [ + 331, + 72, + 538, + 90 + ], + "type": "text", + "content": "Oh no! I'm making bread and forgot to add salt when I made the dough! I'm at the kneading step." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 94, + 514, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 94, + 514, + 111 + ], + "spans": [ + { + "bbox": [ + 312, + 94, + 514, + 111 + ], + "type": "text", + "content": "If you're at the kneading step, don't worry—you can still incorporate it into the dough with more kneading. Here's how:" + } + ] + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 526, + 92, + 543, + 112 + ], + "blocks": [ + { + "bbox": [ + 526, + 92, + 543, + 112 + ], + "lines": [ + { + "bbox": [ + 526, + 92, + 543, + 112 + ], + "spans": [ + { + "bbox": [ + 526, + 92, + 543, + 112 + ], + "type": "image", + "image_path": "71ab28f90bc8e280ee1341f7124499b16a1260acdeae65b53746c69b2614a258.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 116, + 381, + 142 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 116, + 381, + 142 + ], + "spans": [ + { + "bbox": [ + 312, + 116, + 381, + 142 + ], + "type": "text", + "content": "Step 1: Determine the amount of salt needed." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 391, + 116, + 459, + 142 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 116, + 459, + 142 + ], + "spans": [ + { + "bbox": [ + 391, + 116, + 459, + 142 + ], + "type": "text", + "content": "Step 2: Dissolve the salt in a small amount of warm water." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 469, + 116, + 539, + 134 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 469, + 116, + 539, + 134 + ], + "spans": [ + { + "bbox": [ + 469, + 116, + 539, + 134 + ], + "type": "text", + "content": "Step 3: Flatten the dough into a rectangle." + } + ] + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 310, + 148, + 385, + 224 + ], + "blocks": [ + { + "bbox": [ + 310, + 148, + 385, + 224 + ], + "lines": [ + { + "bbox": [ + 310, + 148, + 385, + 224 + ], + "spans": [ + { + "bbox": [ + 310, + 148, + 385, + 224 + ], + "type": "image", + "image_path": "10698da331ef96acbc1722d2850a1bffea0f43d33a1f952ff8ed107c2de59022.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 225, + 373, + 243 + ], + "lines": [ + { + "bbox": [ + 312, + 225, + 373, + 243 + ], + "spans": [ + { + "bbox": [ + 312, + 225, + 373, + 243 + ], + "type": "text", + "content": "Step 4: Spread the salt over the dough." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 389, + 148, + 463, + 224 + ], + "blocks": [ + { + "bbox": [ + 389, + 148, + 463, + 224 + ], + "lines": [ + { + "bbox": [ + 389, + 148, + 463, + 224 + ], + "spans": [ + { + "bbox": [ + 389, + 148, + 463, + 224 + ], + "type": "image", + "image_path": "5a9ec6801506eaa2482dff9b1df28cae824c3c53e6bb31f7365e12d82e9b060e.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 391, + 225, + 451, + 250 + ], + "lines": [ + { + "bbox": [ + 391, + 225, + 451, + 250 + ], + "spans": [ + { + "bbox": [ + 391, + 225, + 451, + 250 + ], + "type": "text", + "content": "Step 5: Fold and knead the dough to distribute." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 467, + 149, + 542, + 224 + ], + "blocks": [ + { + "bbox": [ + 467, + 149, + 542, + 224 + ], + "lines": [ + { + "bbox": [ + 467, + 149, + 542, + 224 + ], + "spans": [ + { + "bbox": [ + 467, + 149, + 542, + 224 + ], + "type": "image", + "image_path": "7aaf6a10495f2721bc93a9ad1d73b033ad5ff37563972cae78a779ae62c29f69.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 469, + 225, + 534, + 250 + ], + "lines": [ + { + "bbox": [ + 469, + 225, + 534, + 250 + ], + "spans": [ + { + "bbox": [ + 469, + 225, + 534, + 250 + ], + "type": "text", + "content": "Step 6: Proceed with the original bread recipe." + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 310, + 258, + 385, + 332 + ], + "blocks": [ + { + "bbox": [ + 310, + 258, + 385, + 332 + ], + "lines": [ + { + "bbox": [ + 310, + 258, + 385, + 332 + ], + "spans": [ + { + "bbox": [ + 310, + 258, + 385, + 332 + ], + "type": "image", + "image_path": "4d3dd587c80162601dfb094dcfae5e45d3e615ea78c912df4e7fb6fa3d390f55.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 389, + 258, + 463, + 332 + ], + "blocks": [ + { + "bbox": [ + 389, + 258, + 463, + 332 + ], + "lines": [ + { + "bbox": [ + 389, + 258, + 463, + 332 + ], + "spans": [ + { + "bbox": [ + 389, + 258, + 463, + 332 + ], + "type": "image", + "image_path": "e8b979dbd1ff2ba95c5013084b3030110152c106f174a67e1949ac6baff1f842.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 467, + 258, + 542, + 333 + ], + "blocks": [ + { + "bbox": [ + 467, + 258, + 542, + 333 + ], + "lines": [ + { + "bbox": [ + 467, + 258, + 542, + 333 + ], + "spans": [ + { + "bbox": [ + 467, + 258, + 542, + 333 + ], + "type": "image", + "image_path": "9a4282b040e672208ad09baa2280751c0806c52f20519d106ff1dcc6b7b1822a.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 343, + 545, + 365 + ], + "lines": [ + { + "bbox": [ + 306, + 343, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 306, + 343, + 545, + 365 + ], + "type": "text", + "content": "Figure 10. Error correction. StackedDiffusion provides updated instructions in response to unexpected situations, like a user error." + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "bbox": [ + 305, + 382, + 545, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 382, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 305, + 382, + 545, + 442 + ], + "type": "text", + "content": "what they want to do and obtain an article with a suggested specific goal that matches what the user would like to do. In Figure 9, the user describes their situation (having many apples) and StackedDiffusion suggests a goal that matches their situation (making apple crisps)." + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 305, + 445, + 545, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 445, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 305, + 445, + 545, + 552 + ], + "type": "text", + "content": "Error correction. Oftentimes, a user in the course of performing a task may make a mistake. Fixed articles provide no avenue for recourse. However, StackedDiffusion can adapt to user error and create alternative instructions that best correct for and accommodate the user's situation. See Figure 10 for an example of this. A fixed article would not provide any alternative paths that adapt as the user performs the action. This highlights the new avenues opened up by our generative approach to this task." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 306, + 571, + 468, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 571, + 468, + 583 + ], + "spans": [ + { + "bbox": [ + 306, + 571, + 468, + 583 + ], + "type": "text", + "content": "6. Limitations and Conclusions" + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "While StackedDiffusion achieves strong performance in generating Illustrated Instructions, we note that there still remains a gap with ground truth data. Future work might examine how to obtain instructional data at greater scale and how this influences StackedDiffusion. Leveraging the latest T2I architectures to improve the faithfulness and quality of our generations would be another promising avenue for future work. Finally, leveraging improvements in text-to-video techniques to generate video clips describing each step would further improve the usability of such a system." + } + ] + } + ], + "index": 43 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "6281" + } + ] + } + ], + "index": 44 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "text", + "content": "[1] Armen Aghajanyan, Bernie Huang, Candace Ross, Vladimir Karpukhin, Hu Xu, Naman Goyal, Dmytro Okhonko, Mandar Joshi, Gargi Ghosh, Mike Lewis, and Luke Zettlemoyer. CM3: A causal masked multimodal model of the internet. arXiv preprint arXiv:2201.07520, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 287, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 287, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 287, + 179 + ], + "type": "text", + "content": "[2] Emanuele Aiello, Lili Yu, Yixin Nie, Armen Aghajanyan, and Barlas Oguz. Jointly training large autoregressive multimodal models. In ICLR, 2024." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 287, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 287, + 215 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 287, + 215 + ], + "type": "text", + "content": "[3] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, 2021." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 217, + 287, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 217, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 53, + 217, + 287, + 271 + ], + "type": "text", + "content": "[4] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 274, + 287, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 274, + 287, + 317 + ], + "spans": [ + { + "bbox": [ + 53, + 274, + 287, + 317 + ], + "type": "text", + "content": "[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your Latents: High-Resolution Video Synthesis with Latent Diffusion Models. In CVPR, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 319, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 319, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 53, + 319, + 287, + 352 + ], + "type": "text", + "content": "[6] Russell N. Carney and Joel R. Levin. Pictorial illustrations still improve students' learning from text. Educational Psychology Review, 14(1):5-26, 2002." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 354, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 354, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 53, + 354, + 287, + 396 + ], + "type": "text", + "content": "[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging Properties in Self-Supervised Vision Transformers. In ICCV, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 399, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 399, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 53, + 399, + 287, + 464 + ], + "type": "text", + "content": "[8] Huiwen Chang, Han Zhang, Jarred Barber, Aaron Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Patrick Murphy, William T. Freeman, Michael Rubinstein, Yanzhen Li, and Dilip Krishnan. Muse: Text-ToImage Generation via Masked Generative Transformers. In ICML, 2023." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 467, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 467, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 53, + 467, + 287, + 521 + ], + "type": "text", + "content": "[9] Brian Chen, Andrew Rouditchenko, Kevin Duarte, Hilde Kuehne, Samuel Thomas, Angie Boggust, Rameswar Panda, Brian Kingsbury, Rogerio Feris, David Harwath, et al. Multimodal clustering networks for self-supervised learning from unlabeled videos. In ICCV, 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 523, + 287, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 643 + ], + "type": "text", + "content": "[10] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Alex Castro-Ros, Marie Pellat, Kevin Robinson, Dasha Valter, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416, 2022." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "type": "text", + "content": "[11] Yang Deng, Wai Lam, Yuexiang Xie, Daoyuan Chen, Yaliang Li, Min Yang, and Ying Shen. Joint Learning of Answer Selection and Answer Summary Generation in Community Question Answering. In AAAI, 2020." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[12] Eadom Dessalene, Chinmaya Devaraj, Michael Maynord, Cornelia Fermuller, and Yiannis Aloimonos. Forecasting ac" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 73, + 544, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 544, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 544, + 95 + ], + "type": "text", + "content": "tion through contact representations from first person video. PAMI, 2021." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "type": "text", + "content": "[13] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In ICCV, 2023." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 141, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 183 + ], + "type": "text", + "content": "[14] Bahare Fatemi, Quentin Duval, Rohit Girdhar, Michal Drozdzal, and Adriana Romero-Soriano. Learning to substitute ingredients in recipes. arXiv preprint arXiv:2302.07960, 2023." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 186, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 545, + 217 + ], + "type": "text", + "content": "[15] Antonino Furnari and Giovanni Maria Farinella. Rolling-unrolling lstms for action anticipation from first-person video. PAMI, 2020." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 220, + 545, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 545, + 252 + ], + "type": "text", + "content": "[16] Valentin Gabeur, Chen Sun, Karteek Alahari, and Cordelia Schmid. Multi-modal transformer for video retrieval. In ECCV, 2020." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 254, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 254, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 254, + 545, + 285 + ], + "type": "text", + "content": "[17] Valentin Gabeur, Arsha Nagrani, Chen Sun, Karteek Alahari, and Cordelia Schmid. Masking modalities for cross-modal video retrieval. In WACV, 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 288, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 288, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 288, + 545, + 330 + ], + "type": "text", + "content": "[18] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. In ECCV, 2022." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 333, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 333, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 308, + 333, + 545, + 353 + ], + "type": "text", + "content": "[19] Rohit Girdhar and Kristen Grauman. Anticipative video transformer. In ICCV, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 355, + 545, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 355, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 308, + 355, + 545, + 410 + ], + "type": "text", + "content": "[20] Rohit Girdhar, Mannat Singh, Andrew Brown, Quentin Duval, Samaneh Azadi, Sai Saketh Rambhatla, Akbar Shah, Xi Yin, Devi Parikh, and Ishan Misra. Emu video: Factorizing text-to-video generation by explicit image conditioning. arXiv preprint arXiv:2311.10709, 2023." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 411, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 454 + ], + "type": "text", + "content": "[21] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In CVPR, 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 456, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 456, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 308, + 456, + 545, + 498 + ], + "type": "text", + "content": "[22] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 501, + 545, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 501, + 545, + 565 + ], + "spans": [ + { + "bbox": [ + 308, + 501, + 545, + 565 + ], + "type": "text", + "content": "[23] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen Video: High Definition Video Generation with Diffusion Models. arXiv preprint arXiv:2210.02303, 2022." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 567, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 567, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 308, + 567, + 545, + 610 + ], + "type": "text", + "content": "[24] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 613, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 656 + ], + "type": "text", + "content": "[25] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 658, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 678 + ], + "type": "text", + "content": "[26] Diederik P. Kingma and Max Welling. Auto-Encoding Variational Bayes. In ICLR, 2014." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 680, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 712 + ], + "type": "text", + "content": "[27] Jing Yu Koh, Daniel Fried, and Ruslan Salakhutdinov. Generating images with multimodal language models. In NeurIPS, 2023." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6282" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[28] Jing Yu Koh, Ruslan Salakhutdinov, and Daniel Fried. Grounding language models to images for multimodal inputs and outputs. In ICML, 2023." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 287, + 140 + ], + "type": "text", + "content": "[29] Jing Yu Koh, Ruslan Salakhutdinov, and Daniel Fried. Grounding Language Models to Images for Multimodal Generation. In NeurIPS, 2023." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 143, + 287, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 143, + 287, + 175 + ], + "spans": [ + { + "bbox": [ + 49, + 143, + 287, + 175 + ], + "type": "text", + "content": "[30] Mahnaz Koupae and William Yang Wang. WikiHow: A Large Scale Text Summarization Dataset. arXiv preprint arXiv:1810.09305, 2018." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 177, + 287, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 177, + 287, + 220 + ], + "spans": [ + { + "bbox": [ + 49, + 177, + 287, + 220 + ], + "type": "text", + "content": "[31] Linjie Li, Yen-Chun Chen, Yu Cheng, Zhe Gan, Licheng Yu, and Jingjing Liu. Hero: Hierarchical encoder for video+ language omni-representation pre-training. arXiv preprint arXiv:2005.00200, 2020." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 223, + 287, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 223, + 287, + 255 + ], + "spans": [ + { + "bbox": [ + 49, + 223, + 287, + 255 + ], + "type": "text", + "content": "[32] Shanchuan Lin, Bingchen Liu, Jiashi Li, and Xiao Yang. Common diffusion noise schedules and sample steps are flawed. arXiv preprint arXiv:2305.08891, 2023." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 258, + 287, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 258, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 49, + 258, + 287, + 300 + ], + "type": "text", + "content": "[33] Xudong Lin, Fabio Petroni, Gedas Bertasius, Marcus Rohrbach, Shih-Fu Chang, and Lorenzo Torresani. Learning To Recognize Procedural Activities with Distant Supervision. In CVPR, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 303, + 287, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 303, + 287, + 345 + ], + "spans": [ + { + "bbox": [ + 49, + 303, + 287, + 345 + ], + "type": "text", + "content": "[34] Song Liu, Haoqi Fan, Shengsheng Qian, Yiru Chen, Wenkui Ding, and Zhongyuan Wang. Hit: Hierarchical transformer with momentum contrast for video-text retrieval. In ICCV, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 348, + 287, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 348, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 49, + 348, + 287, + 370 + ], + "type": "text", + "content": "[35] Ilya Loshchilov and Frank Hutter. Decoupled Weight Decay Regularization. In ICLR, 2019." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 373, + 287, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 373, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 49, + 373, + 287, + 426 + ], + "type": "text", + "content": "[36] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon Bharti, and Ming Zhou. Univl: A unified video and language pre-training model for multimodal understanding and generation. arXiv preprint arXiv:2002.06353, 2020." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 429, + 287, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 429, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 49, + 429, + 287, + 472 + ], + "type": "text", + "content": "[37] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval. arXiv preprint arXiv:2104.08860, 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 475, + 287, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 475, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 49, + 475, + 287, + 517 + ], + "type": "text", + "content": "[38] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Ti-niu Tan. Videofusion: Decomposed diffusion models for high-quality video generation. In CVPR, 2023." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 520, + 287, + 573 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 520, + 287, + 573 + ], + "spans": [ + { + "bbox": [ + 49, + 520, + 287, + 573 + ], + "type": "text", + "content": "[39] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. HowTo100M: Learning a Text-Video Embedding by Watching Hundred Million Narrated Video Clips. In ICCV, 2019." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 577, + 287, + 619 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 577, + 287, + 619 + ], + "spans": [ + { + "bbox": [ + 49, + 577, + 287, + 619 + ], + "type": "text", + "content": "[40] Antoine Miech, Jean-Baptiste Alayrac, Ivan Laptev, Josef Sivic, and Andrew Zisserman. Thinking fast and slow: Efficient text-to-visual retrieval with transformers. In CVPR, 2021." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 622, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 622, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 49, + 622, + 287, + 644 + ], + "type": "text", + "content": "[41] Medhini Narasimhan, Anna Rohrbach, and Trevor Darrell. Clip-it! language-guided video summarization. NIPS, 2021." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 646, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 646, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 49, + 646, + 287, + 689 + ], + "type": "text", + "content": "[42] Medhini Narasimhan, Arsha Nagrani, Chen Sun, Michael Rubinstein, Trevor Darrell, Anna Rohrbach, and Cordelia Schmid. Tl; dw? summarizing instructional videos with task relevance and cross-modal saliency. In ECCV, 2022." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 49, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 692, + 287, + 713 + ], + "type": "text", + "content": "[43] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 105 + ], + "type": "text", + "content": "Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 107, + 545, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 126 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 126 + ], + "type": "text", + "content": "[44] OpenAI. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 129, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 129, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 308, + 129, + 545, + 160 + ], + "type": "text", + "content": "[45] Gaurav Parmar, Richard Zhang, and Jun-Yan Zhu. On aliased resizing and surprising subtleties in gan evaluation. In CVPR, 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 163, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 194 + ], + "type": "text", + "content": "[46] AJ Piergiovanni, Anelia Angelova, Michael S Ryoo, and Irfan Essa. Unsupervised action segmentation for instructional videos. arXiv preprint arXiv:2106.03738, 2021." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 196, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 196, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 308, + 196, + 545, + 249 + ], + "type": "text", + "content": "[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 251, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 251, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 251, + 545, + 293 + ], + "type": "text", + "content": "[48] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 295, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 295, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 308, + 295, + 545, + 327 + ], + "type": "text", + "content": "[49] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bjorn Ommer. High-Resolution Image Synthesis with Latent Diffusion Models. In CVPR, 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 328, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 328, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 308, + 328, + 545, + 370 + ], + "type": "text", + "content": "[50] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. DreamBooth: Fine Tuning Text-to-Image Diffusion Models for Subject-Driven Generation. In CVPR, 2023." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 373, + 545, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 373, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 308, + 373, + 545, + 448 + ], + "type": "text", + "content": "[51] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding. In NIPS, 2022." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 449, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 469 + ], + "type": "text", + "content": "[52] Fadime Sener and Angela Yao. Zero-shot anticipation for instructional activities. In ICCV, 2019." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 472, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 472, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 308, + 472, + 545, + 502 + ], + "type": "text", + "content": "[53] Paul Hongsuck Seo, Arsha Nagrani, and Cordelia Schmid. Look before you speak: Visually contextualized utterances. In CVPR, 2021." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 505, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 505, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 308, + 505, + 545, + 558 + ], + "type": "text", + "content": "[54] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-A-Video: Text-to-Video Generation without Text-Video Data. arXiv preprint arXiv:2209.14792, 2022." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 559, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 624 + ], + "type": "text", + "content": "[55] George Stein, Jesse C Cresswell, Rasa Hosseinzadeh, Yi Sui, Brendan Leigh Ross, Valentin Villecloze, Zhaoyan Liu, Anthony L Caterini, J Eric T Taylor, and Gabriel Loiza-Ganem. Exposing flaws of generative model evaluation metrics and their unfair treatment of diffusion models. arXiv preprint arXiv:2306.04675, 2023." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 626, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 626, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 308, + 626, + 545, + 667 + ], + "type": "text", + "content": "[56] Chen Sun, Fabien Baradel, Kevin Murphy, and Cordelia Schmid. Learning video representations using contrastive bidirectional transformer. arXiv preprint arXiv:1906.05743, 2019." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "text", + "content": "[57] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Emu: Generative pretraining in multimodality. In ICLR, 2024." + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6283" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[58] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. COIN: A Large-scale Dataset for Comprehensive Instructional Video Analysis. In CVPR, 2019." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 119, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 119, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 119, + 287, + 174 + ], + "type": "text", + "content": "[59] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable Length Video Generation from Open Domain Textual Descriptions. In ICLR, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 176, + 287, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 176, + 287, + 220 + ], + "spans": [ + { + "bbox": [ + 48, + 176, + 287, + 220 + ], + "type": "text", + "content": "[60] Wenzhe Wang, Mengdan Zhang, Runnan Chen, Guanyu Cai, Penghao Zhou, Pai Peng, Xiaowei Guo, Jian Wu, and Xing Sun. Dig into multi-modal cues for video retrieval with hierarchical alignment. In *IJCAI*, 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 222, + 287, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 222, + 287, + 276 + ], + "spans": [ + { + "bbox": [ + 48, + 222, + 287, + 276 + ], + "type": "text", + "content": "[61] Chao-Yuan Wu, Yanghao Li, Karttikeya Mangalam, Haoqi Fan, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Memvit: Memory-augmented multiscale vision transformer for efficient long-term video recognition. arXiv preprint arXiv:2201.08383, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 279, + 287, + 333 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 279, + 287, + 333 + ], + "spans": [ + { + "bbox": [ + 48, + 279, + 287, + 333 + ], + "type": "text", + "content": "[62] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In ICCV, 2023." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 335, + 287, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 335, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 48, + 335, + 287, + 380 + ], + "type": "text", + "content": "[63] Mengmeng Xu, Juan-Manuel Pérez-Rúa, Victor Escorcia, Brais Martinez, Xiatian Zhu, Li Zhang, Bernard Ghanem, and Tao Xiang. Boundary-sensitive pre-training for temporal localization in videos. In ICCV, 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 382, + 287, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 382, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 382, + 287, + 415 + ], + "type": "text", + "content": "[64] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Just ask: Learning to answer questions from millions of narrated videos. In ICCV, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 417, + 287, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 417, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 48, + 417, + 287, + 449 + ], + "type": "text", + "content": "[65] Jianwei Yang, Yonatan Bisk, and Jianfeng Gao. Taco: Token-aware cascade contrastive learning for video-text alignment. In ICCV, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 453, + 287, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 453, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 48, + 453, + 287, + 495 + ], + "type": "text", + "content": "[66] Mengjiao Yang, Yilun Du, Bo Dai, Dale Schuurmans, Joshua B Tenenbaum, and Pieter Abbeel. Probabilistic adaptation of text-to-video models. arXiv preprint arXiv:2306.01872, 2023." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 498, + 287, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 498, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 48, + 498, + 287, + 531 + ], + "type": "text", + "content": "[67] Yue Yang, Artemis Panagopoulou, Qing Lyu, Li Zhang, Mark Yatskar, and Chris Callison-Burch. Visual goal-step inference using wikihow. In EMNLP, 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 533, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 287, + 599 + ], + "type": "text", + "content": "[68] Shengming Yin, Chenfei Wu, Huan Yang, Jianfeng Wang, Xiaodong Wang, Minheng Ni, Zhengyuan Yang, Linjie Li, Shuguang Liu, Fan Yang, Jianlong Fu, Gong Ming, Lijuan Wang, Zicheng Liu, Houqiang Li, and Nan Duan. NUWAXL: Diffusion over Diffusion for eXtremely Long Video Generation. arXiv preprint arXiv:2303.12346, 2023." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 601, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 656 + ], + "type": "text", + "content": "[69] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[70] Lili Yu, Bowen Shi, Ramakanth Pasunuru, Benjamin Muller, Olga Golovneva, Tianlu Wang, Arun Babu, Binh Tang, Brian Karrer, Shelly Sheynin, et al. Scaling autoregressive multimodal models: Pretraining and instruction tuning. arXiv preprint arXiv:2309.02591, 2023." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 331 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[71] Jingqing Zhang, Yao Zhao, Mohammad Saleh, and Peter J. Liu. PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization. In ICML, 2019." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "type": "text", + "content": "[72] Li Zhang, Qing Lyu, and Chris Callison-Burch. Reasoning about Goals, Steps, and Temporal Ordering with WikiHow. In EMNLP, 2020." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "type": "text", + "content": "[73] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 186, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 228 + ], + "type": "text", + "content": "[74] Shuyan Zhou, Li Zhang, Yue Yang, Qing Lyu, Pengcheng Yin, Chris Callison-Burch, and Graham Neubig. Show Me More Details: Discovering Hierarchies of Procedures from Semi-structured Web Data. In ACL, 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 231, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 231, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 231, + 545, + 274 + ], + "type": "text", + "content": "[75] Yilun Zhou, Julie Shah, and Steven Schockaert. Learning Household Task Knowledge from WikiHow Descriptions. In Proceedings of the 5th Workshop on Semantic Deep Learning (SemDeep-5), 2019." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 276, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 297 + ], + "type": "text", + "content": "[76] Linchao Zhu and Yi Yang. Actbert: Learning global-local video-text representations. In CVPR, 2020." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 298, + 545, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 298, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 545, + 331 + ], + "type": "text", + "content": "[77] Dimitri Zhukov, Jean-Baptiste Alayrac, Ivan Laptev, and Josef Sivic. Learning actionness via long-range temporal order verification. In ECCV, 2020." + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6284" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Generating Non-Stationary Textures using Self-Rectification/3ddc2997-ec2e-480d-b380-c7012481f817_content_list.json b/2024/Generating Non-Stationary Textures using Self-Rectification/3ddc2997-ec2e-480d-b380-c7012481f817_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3b428c7b38ae410a7fc9894d11f07ca79767d238 --- /dev/null +++ b/2024/Generating Non-Stationary Textures using Self-Rectification/3ddc2997-ec2e-480d-b380-c7012481f817_content_list.json @@ -0,0 +1,1929 @@ +[ + { + "type": "text", + "text": "Generating Non-Stationary Textures using Self-Rectification", + "text_level": 1, + "bbox": [ + 179, + 130, + 789, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yang Zhou1 Rongjun Xiao1 Dani Lischinski2 Daniel Cohen-Or1,3 Hui Huang1* \n1Shenzhen University 2The Hebrew University of Jerusalem 3Tel Aviv University", + "bbox": [ + 99, + 179, + 875, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 251, + 313, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper addresses the challenge of example-based non-stationary texture synthesis. We introduce a novel two-step approach wherein users first modify a reference texture using standard image editing tools, yielding an initial rough target for the synthesis. Subsequently, our proposed method, termed \"self-rectification\", automatically refines this target into a coherent, seamless texture, while faithfully preserving the distinct visual characteristics of the reference exemplar. Our method leverages a pre-trained diffusion network, and uses self-attention mechanisms, to gradually align the synthesized texture with the reference, ensuring the retention of the structures in the provided target. Through experimental validation, our approach exhibits exceptional proficiency in handling non-stationary textures, demonstrating significant advancements in texture synthesis when compared to existing state-of-the-art techniques. Code is available at https://github.com/xiaorongjun000/Self-Rectification", + "bbox": [ + 76, + 282, + 473, + 554 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 583, + 209, + 598 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Example-based texture synthesis aims to generate a texture that faithfully captures all the visual characteristics of a provided reference texture exemplar. The key challenge is to generate a texture that visually mimics the reference, while avoiding exact replication and without producing conspicuous, unnatural artifacts.", + "bbox": [ + 75, + 608, + 468, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Over the past decades, numerous methods have emerged for synthesizing textures from examples, and many have demonstrated impressive results, particularly for homogeneous textures that can be accurately captured by stationary models [45]. Nonetheless, a significant challenge remains when dealing with real-world textures that are inherently inhomogeneous, and non-stationary.", + "bbox": [ + 75, + 699, + 468, + 804 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Non-stationary textures exhibit distinctive attributes, such as sprawling irregular large-scale structures or variations in attributes such as color, local orientation, and local scale. Fig. 1 (left) shows two examples. Mimicking such", + "bbox": [ + 75, + 805, + 468, + 866 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f31eb4945484ffe2712462bafa4b9d2d3b33e2d6deca38cb83dc50b03ad4b29d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 250, + 630, + 348 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e6f9c02fdb75b4b7b60fbd29e67750440ae9d53990ed575468762fc44b0bf71d.jpg", + "image_caption": [ + "Figure 1. Our method takes as input a reference texture (left), and a crude target texture provided by the user (middle column), which may lack coherence and completeness. Self-rectification is used to transform the target into a visually coherent texture (right) that complies with the structure of the crude target, while exhibiting the visual characteristics of the reference texture." + ], + "image_footnote": [], + "bbox": [ + 501, + 349, + 630, + 448 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/29ad14d7f1db57e93642510d70d71b1f0201738ea9f94457826b886bcb999a1a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 632, + 250, + 759, + 348 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1212ca72001c098789299d484ca37d8a6a2b066b53686bc5f5a3a3f0ad0aa22d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 632, + 349, + 759, + 448 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0f54f213629f7fedd67bce73e72d53484850dc2dd9e32e7693a9d75021168e7e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 761, + 250, + 890, + 348 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7ec8b08f3d3ce6cb996ca2d27bc176cda12f4b1a5cdb2c25e7b10708a04b2c26.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 761, + 349, + 890, + 448 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "complex structures and variations through example-based synthesis is a long-standing challenging task [31].", + "bbox": [ + 498, + 582, + 890, + 613 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The emergence of neural networks has provided powerful means to deal with non-stationary textures. Zhou et al. [51] introduced a method that involves overfitting a GAN, where the encoder extracts structural guidance for the decoder. This approach provides a viable means of spatially extending non-stationary textures while preserving the visual characteristics of the reference. However, this technique requires an extremely long optimization process for a single texture exemplar, and moreover, it fails to provide any controllability or editability.", + "bbox": [ + 496, + 613, + 892, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we introduce a two-step lazy-editing technique where the user first edits the given reference texture using conventional image editing tools, to obtain an extremely rough initial result, which may be incomplete and incoherent, as demonstrated in Fig. 1 (middle). Next, our technique automatically rectifies the edited texture into a regularized, coherent and seamless texture that follows the rough target, while retaining the local characteristics of the reference; see, e.g., Fig. 1 (right). We term this regulariza", + "bbox": [ + 496, + 763, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 875, + 220, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7767", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tion process self-rectification, since the texture is rectified based on the input texture itself.", + "bbox": [ + 75, + 90, + 468, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To self-rectify the crude target, we use a pre-trained diffusion network, and synthesize the final rectified result while utilizing the intermediate byproducts obtained by inverting both the initial target and the reference exemplar. Inspired by editing approaches that leverage self-attention [1,6,20], we extract certain self-attention features of the diffusion model during the inversion of the rough target, as well as the reference. These features are then injected into various steps in the inversion and denoising of the generated texture. Such feature injection forms \"cross-attention\" between the reference and target features [1, 6], and thereby, in the course of the diffusion process, the target texture is progressively refined to be increasingly locally similar to the reference, while retaining the global structure prescribed by the provided target. This self-rectification occurs in two passes, first addressing the larger scale structure and subsequently focusing on finer local details.", + "bbox": [ + 73, + 121, + 468, + 377 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To enhance the synthesis quality, we augment the reference texture with several transformed copies of it, which increases the diversity of the reference patterns, thereby improving the compatibility between the reference and the synthesized result. The augmented source features are injected into the corresponding target layers as well. Furthermore, we show that our method can also be applied to lazy editing of natural images, using the same means of synthesizing highly non-stationary textures. Experiments show that our method can deal with a large scope of nonstationary textures, with unprecedented flexibility and quality, compared to the state-of-the-art.", + "bbox": [ + 73, + 378, + 468, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 571, + 218, + 588 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Non-stationary texture synthesis. Early methods in example-based texture synthesis mainly focused on synthesizing textures with stationary characteristics [8, 9, 19, 21, 22, 43, 44]. To cope with non-stationary textures, researchers typically involve additional guidance for the control of distinctive attributes, e.g., using label maps to guide the layout synthesis of composite textures [15, 24, 31], representing the spatial variations of weathering textures with age/progression maps [3, 42], and describing the local orientations of directional textures with vector fields [26, 50]. Although the aforementioned methods have shown success, the guidance required from the user is tedious to provide, yet still limited.", + "bbox": [ + 73, + 598, + 467, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the deep learning era, neural methods for texture generation have rapidly emerged, either through new texture losses for texture optimization [12, 13, 49] or via training generative networks [4, 5, 10, 18, 23, 29, 32, 35-37, 40, 51]. Among these methods, Sendik and Cohen-Or [35] attempted to preserve the non-local structures of non-stationary textures by regularizing the feature correla", + "bbox": [ + 75, + 795, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/59cb17c6ec74110402a6ead0e05ea69d01522e1303202883d35bef37cad0d555.jpg", + "image_caption": [ + "Figure 2. Framework overview. Given a reference texture $I^R$ , we allow the user to quickly build a target image $I^{tar}$ in a lazy-editing manner. A coarse-to-fine synthesis is performed by running self-rectification twice. The coarse stage synthesizes a coarse yet complete overall structure, and the fine stage refines its output $I_{\\text{coarse}}^*$ with finer and more accurate details, producing the final result $I^*$ ." + ], + "image_footnote": [], + "bbox": [ + 506, + 85, + 893, + 220 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tion between different locations. Zhou et al. [51] proposed overfitting a GAN to learn the expansion from a small texture block to a larger one containing it. Their approach involves the GAN's encoder extracting the global structure of an input texture, duplicated by bottleneck residual blocks before decoding. Although these trained models can effectively extend non-stationary textures, their overfitting nature severely restricts their generalization and controllability.", + "bbox": [ + 498, + 344, + 890, + 467 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion-based image synthesis. The emergence of large-scale generative diffusion models, such as Stable Diffusion (SD) [30] and DALLE-2 [28], has revolutionized image synthesis due to their unprecedented generation power. To make the pre-trained diffusion models synthesize on image conditions, many solutions are proposed, including optimizing prompt tokens [2, 11], fine-tuning the entire model [33], or training an additional adapter [17,34,46-48].", + "bbox": [ + 498, + 484, + 890, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, researchers have investigated the role of the intermediate attention maps and features in the diffuser [7, 14,27,39], finding them crucial for layout/structure synthesis. The method we present in this paper, builds upon the mechanism of injecting keys and values from attention layers of one diffusion process into another, as a means for transferring visual features between images [1,6,20]. We use such injection to \"copy\" the local patterns from a source texture to a target one. In addition to the injection we also incorporate coarse-to-fine and data augmentation schemes, forming the self-rectification framework for generating nonstationary textures.", + "bbox": [ + 498, + 604, + 890, + 786 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 800, + 589, + 814 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our method employs a two-step process for synthesizing a new texture $I^{*}$ resembling a provided non-stationary reference texture $I^{R}$ , while guided by a user-provided rough target $I^{tar}$ , as depicted in Fig. 2. Initially, the user rapidly creates $I^{tar}$ , by assembling patches from the source refer", + "bbox": [ + 498, + 825, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7768", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e869299231132e9069bda6cf75424c234d4e42aa5bed010dde050968793b7191.jpg", + "image_caption": [ + "Figure 3. Our self-rectification synthesizes an output texture $I^{*}$ via structure-preserving inversion from a rough target image $I^{lar}$ and fine texture sampling using the reference $I^{R}$ . Both processes require the injection of self-attention features ( $KV$ ) from the DDIM inversion of a corresponding reference. More specifically, for structure-preserving inversion, the reference is the target image itself, denoted as $I^{IR}$ . For fine texture sampling, the input exemplar $I^{R}$ is used to inject features that help to synthesize a plausible output with fine texture details." + ], + "image_footnote": [], + "bbox": [ + 83, + 85, + 890, + 258 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ence texture. Such a lazy-edit forms a basic layout for the output image, $I^{*}$ . However, this initial sketch may be incoherent or incomplete, hence necessitating rectification. Subsequently, the texture undergoes self-rectification to align with both the user's coarse layout and the reference image's detailed texture characteristics. This self-rectification process is depicted in Fig. 2 and is executed in two stages: coarse rectification, followed by a finer one.", + "bbox": [ + 75, + 352, + 468, + 472 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Considering that the rough target $I^{tar}$ comprises patches derived from the reference texture $I^R$ , we implement a process of self-rectification, detailed below. In a broad sense, we utilize a pre-trained latent diffusion model to invert both the target image and the reference image into an initial latent noise. This is followed by employing feature injection during the sampling process of $I^*$ . The features from $I^{tar}$ guide the structural synthesis and the features from $I^R$ guide the fine texture details.", + "bbox": [ + 75, + 472, + 468, + 607 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Below, we first briefly review diffusion models and their self-attention and the cross-KV-injection mechanisms, and then proceed to present our self-rectification technique.", + "bbox": [ + 75, + 608, + 468, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. DDIM Sampling and Inversion", + "text_level": 1, + "bbox": [ + 76, + 662, + 352, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Denoising diffusion models [16, 38] involve two processes: a noising process that gradually transforms an input image into Gaussian noise, and a denoising/sampling process that generates images from Gaussian samples. Using DDIM sampling [38], starting from a noise sample $z_{\\mathrm{T}}$ , one can generate a clean sample $z_{0}$ via T deterministic steps:", + "bbox": [ + 75, + 686, + 468, + 777 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t - 1} = \\sqrt {\\bar {\\alpha} _ {t - 1}} f _ {\\theta} (z _ {t}, t) + \\sqrt {1 - \\bar {\\alpha} _ {t - 1}} \\epsilon_ {\\theta} (z _ {t}, t), \\quad (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 786, + 468, + 805 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\epsilon_{\\theta}$ is a noise prediction network conditioned on the current noisy sample $z_{t}$ and timestamp $t$ . $\\bar{\\alpha}_{t}$ is the noise scaling factor defined in [38], and $f_{\\theta}(z_t,t)$ is", + "bbox": [ + 75, + 815, + 468, + 859 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\theta} \\left(z _ {t}, t\\right) = \\frac {z _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta} \\left(z _ {t} , t\\right)}{\\sqrt {\\bar {\\alpha} _ {t}}}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 869, + 468, + 905 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the opposite direction, $DDIM$ inversion [38] of a given image $z_0$ is the process of incrementally adding deterministic noise, until obtaining $z_{\\mathrm{T}}$ :", + "bbox": [ + 498, + 352, + 890, + 397 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t + 1} = \\sqrt {\\bar {\\alpha} _ {t + 1}} f _ {\\theta} \\left(z _ {t}, t\\right) + \\sqrt {1 - \\bar {\\alpha} _ {t + 1}} \\epsilon_ {\\theta} \\left(z _ {t}, t\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 407, + 890, + 426 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Such inversion leads to a nearly faithful reconstruction [38].", + "bbox": [ + 500, + 435, + 890, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Stable Diffusion and Self-Attention.", + "text_level": 1, + "bbox": [ + 500, + 459, + 808, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We base our texture synthesis framework on Stable Diffusion (SD) [30], which is a pretrained latent diffusion model consisting of an encoder $\\mathcal{E}$ that maps an input image into the latent space, and a decoder $\\mathcal{D}$ that reconstructs a latent code back into image space. Both DDIM sampling and DDIM inversion are performed in the SD latent space.", + "bbox": [ + 496, + 482, + 890, + 573 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The noise predictor $\\epsilon_{\\theta}$ of SD is a large-scale U-Net that contains multiple self-attention modules [41]. Each self-attention layer transforms its input intermediate feature map (also called spatial features) into an attended representation by the following equation:", + "bbox": [ + 496, + 573, + 890, + 648 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t} (Q, K, V) = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d}}\\right) V, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 657, + 890, + 694 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $Q, K$ , and $V$ are the queries, keys, and values, respectively, obtained by learned linear projections of the same input spatial features, having dimension $d$ . The self-attention mechanism uses the similarities between the queries and keys as attention scores to weigh the importance or relevance of the values. Relevant info is thus aggregated as the attended representation.", + "bbox": [ + 496, + 702, + 890, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. KV-Injection", + "text_level": 1, + "bbox": [ + 500, + 816, + 637, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The self-attention features contain rich information about both the large-scale structures and local fine details of an input image [1, 6, 39]. Tumanyan et al. [39] demonstrate that by injecting the spatial features and the queries", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7769", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/01fc19e8ac5e749930664f6c79335d9cf8055d022cdb8134c0d0ccaadff0b0da.jpg", + "image_caption": [ + "Figure 4. Visualization of the intermediate latent codes in the inversion. For the standard DDIM inversion (top), the U-Net predicts noise to diffuse the distinctive structures so as to transform the input into Gaussian noise. In contrast, our structure-preserving inversion (bottom) reserves the distinctive patterns from user edits along the inversion process. See texts in Sec. 3.4 for more details." + ], + "image_footnote": [], + "bbox": [ + 76, + 85, + 472, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and keys $(QK)$ from the self-attention layers of a source (guidance) image into the corresponding layer of the generated target image during the sampling process, one can preserve the layout of the source while modifying its appearance. Injecting only the $KV$ features of a source reference, instead, transfers its appearance, including textures, to the generated target [1,6].", + "bbox": [ + 75, + 376, + 467, + 482 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our case, the target image is a rough and incomplete guidance, typically containing only a few source patches rotated and placed freely by the user. A complete overall structure needs to be synthesized reasonably yet conforming to user's constraints. Locally, however, the resulting texture should resemble the reference. To achieve these goals, we adapt the $KV$ -injection into both the DDIM inversion and sampling, bringing a novel self-rectification operation.", + "bbox": [ + 75, + 482, + 467, + 603 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Self-rectification", + "text_level": 1, + "bbox": [ + 76, + 612, + 240, + 627 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 3, our self-rectification consists of two parts: structure-preserving inversion and fine texture sampling. Both are performed in the SD latent space. The last latent code of inversion is used as the starting code of the sampling. The core modification to both processes is the $KV$ -injection of self-attention features.", + "bbox": [ + 75, + 635, + 468, + 726 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Structure-preserving inversion. The standard DDIM inversion (Eq. (3)) progressively transforms an input image (SD latent code) into pure Gaussian noise. At each time step, the noise to be added is predicted by the U-Net. As visualized in Fig. 4, for a given target image, the noise predicted by the U-Net for early time steps ( $e.g., t \\leq 20$ ) is mainly distributed to \"diffuse\" the prominent structures so that the distinctive patterns from user edits get scattered and random. As the inversion progresses, the magnitude of the noise added in each step becomes smaller and similar for", + "bbox": [ + 75, + 744, + 468, + 895 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/03503cdfdd39e5c6a714b99e7c1af312a21a75b8fe9d36674c93e7cfb801ca98.jpg", + "image_caption": [ + "Figure 5. Visualization of the intermediate latent codes in the fine texture sampling. Here, for the first 20 steps (from $t = 50$ to 30), we perform the standard DDIM sampling to reconstruct the target layout. Next, we perform $KV$ -injection in the remaining sampling steps ( $t = 30$ to 0), to synthesize fine textures for the output image. The rightmost shows the result produced by simply performing standard DDIM sampling for all steps, i.e., $S = 50$ . No additional structure is synthesized to complete the user edits." + ], + "image_footnote": [], + "bbox": [ + 498, + 88, + 893, + 143 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "all latent pixels, since the \"diffusion\" is getting close to done. As no text prompt is involved, the noise prediction in each inversion step is dominantly determined by the self-attention mechanism in the U-Net. Our key observation is that if we inject the $KV$ features from a large time step $t_1$ ( $\\gg \\mathrm{T} / 2$ ) into an early time step $t_2$ ( $\\ll \\mathrm{T} / 2$ ), the noise predicted at $t_2$ will be smaller and more spatially uniform. The distinctive patterns of $I^{tar}$ , reflecting the user's edits, are thus better preserved. Hence, we refer to this process as structure-preserving inversion.", + "bbox": [ + 496, + 287, + 890, + 439 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, given a rough target image $I^{tar}$ , we invert it twice. The first inversion is a standard DDIM inversion. The produced self-attention features along the noisig steps are regarded as the inversion reference (IR) for the second inversion. The self-attention during the second inversion, at time step $t$ , is now given by:", + "bbox": [ + 496, + 439, + 890, + 530 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t} \\left(Q _ {t} ^ {\\text {t a r}}, K _ {\\mathrm {T} - t} ^ {\\mathrm {I R}}, V _ {\\mathrm {T} - t} ^ {\\mathrm {I R}}\\right) = \\operatorname {S o f t m a x} \\left(\\frac {Q _ {t} ^ {\\text {t a r}} \\left(K _ {\\mathrm {T} - t} ^ {\\mathrm {I R}}\\right) ^ {T}}{\\sqrt {d}}\\right) V _ {\\mathrm {T} - t} ^ {\\mathrm {I R}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 540, + 893, + 594 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{T}$ denotes the total number of steps (we use $\\mathrm{T} = 50$ ). Here $KV$ features are injected in a reverse order in the second round. We would like to stress that we have also experimented with using an offset, i.e., replacing $\\mathrm{T} - t$ with $t + offset$ in the above equation, but found no advantage (see supplementary for more analysis and comparisons).", + "bbox": [ + 496, + 595, + 890, + 685 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During the second inversion, we do not perform $KV$ injections for all time steps. We set a parameter $P$ that defines the time step beyond which we continue with standard inversion. Fig. 4 shows the effect of our structure-preserving inversion, where $P = 30$ . Since the final latent code after the second inversion still contains rich information about the distinctive structures from $I^{tar}$ , the subsequent sampling that starts from this latent code is guided to synthesize a global structure that complies with $I^{tar}$ .", + "bbox": [ + 496, + 686, + 890, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fine texture sampling. We start the DDIM sampling from the final structure-preserving latent code. The sampling uses the U-Net to predict the noise to be removed, as defined in Eq. (1). However, simply performing standard", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7770", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "DDIM sampling for all denoising steps results in a nearly reconstructed target image since DDIM sampling is deterministic. Therefore, we set the first S steps (i.e., from time step T to T - S) to reconstruct the target layout to a certain extent. For the remaining T - S denoising steps, we synthesize the output image $I^{*}$ by matching fine textures from the reference via $KV$ -injection from the reference $I^{R}$ .", + "bbox": [ + 75, + 90, + 467, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To this end, we first DDIM-invert the reference texture $I^{R}$ . At denoising time step $t$ ( $t > \\mathrm{T} - \\mathrm{S}$ ), the $KV$ features extracted during the inversion are injected into the corresponding self-attention layers of the synthesized texture:", + "bbox": [ + 76, + 196, + 467, + 257 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t} \\left(Q _ {t} ^ {*}, K _ {t} ^ {R}, V _ {t} ^ {R}\\right) = \\operatorname {S o f t m a x} \\left(\\frac {Q _ {t} ^ {*} \\left(K _ {t} ^ {R}\\right) ^ {T}}{\\sqrt {d}}\\right) V _ {t} ^ {R}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 267, + 467, + 309 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This forms a cross-image attention, where corresponding fine local patterns in the reference are transferred to the output image in a plausible manner. Fig. 5 visualizes the intermediate process of our texture sampling, where $S = 20$ .", + "bbox": [ + 76, + 319, + 467, + 380 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Implementation details and data augmentation", + "text_level": 1, + "bbox": [ + 76, + 388, + 467, + 404 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the user editing phase, we fill the background of the target canvas with pixels randomly drawn from the source texture, such that the encoding of the target image would not deviate too far from the source in the SD latent space. As the self-rectification is performed twice, the output of coarse stage $I_{\\text{coarse}}^*$ , will be used as the input target image of fine stage to produce the final output $I^*$ . In contrast, for the inversion reference $I^{IR}$ involved in structure-preserving inversion of the two self-rectifications, we use the same initial target image that contains the original user edits. See the supplementary for the full algorithm pseudo-code.", + "bbox": [ + 75, + 412, + 467, + 578 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Considering the parameter settings: since the coarse stage aims for structure synthesis, relatively large $\\mathbf{P}$ and S are required in self-rectification, and vice versa in the fine stage. More specifically, let $\\mathrm{P_1}$ , $\\mathrm{P_2}$ , $\\mathrm{S_1}$ , and $\\mathrm{S_2}$ , denote the parameters used in the two rounds of structure-preserving inversion and fine texture sampling. We typically set $\\mathrm{P_1} = 20$ , $\\mathrm{P_2} = 5$ , $\\mathrm{S_1} = 20$ , and $\\mathrm{S_2} = 5$ . Following [6], we choose the $KV$ features from the 10th to 15th self-attention layers of the U-Net decoder part.", + "bbox": [ + 75, + 579, + 467, + 714 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To further improve the synthesis quality, especially when dealing with textures containing a dominant directional structure, such as the leaf shown in Fig. 1, we can introduce a few transformed images (flips and rotations) to augment the reference texture. We concatenate the new attention features from the augmentation to the original reference feature. For example, when we have $n$ augmented source textures, the $K^R$ and $V^R$ in Eq. (6) is now given by", + "bbox": [ + 75, + 714, + 467, + 835 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l l} K ^ {R} & = \\operatorname {C o n c a t} \\left(K _ {(0)} ^ {R}, K _ {(1)} ^ {R}, \\dots , K _ {(n)} ^ {R}\\right), \\\\ V ^ {R} & = \\operatorname {C o n c a t} \\left(V _ {(0)} ^ {R}, V _ {(1)} ^ {R}, \\dots , V _ {(n)} ^ {R}\\right) \\end{array} \\right. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 845, + 467, + 895 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $K_{(i)}^{R}$ , and $V_{(i)}^{R}$ are the self-attention features acquired from the DDIM inversion of the augmented references.", + "bbox": [ + 500, + 90, + 890, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 132, + 630, + 148 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We apply our method with Stable Diffusion with publicly available checkpoints v1.4. All experiments were conducted on a single Quadro P6000 24G GPU. The inference time synthesizing an image of $512 \\times 512$ pixels takes about three minutes for our coarse-to-fine self-rectification.", + "bbox": [ + 498, + 157, + 890, + 233 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Evaluations and Comparisons", + "text_level": 1, + "bbox": [ + 500, + 241, + 766, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our method with non-stationary textures released by [51]. Each example is resized to $512 \\times 512$ pixels as the reference, and we quickly built several different target images for each in PhotoShop with just a few lazy edits. Fig. 6 shows a gallery of results generated by our method. As can be seen our method faithfully reproduces the delicate textures of the exemplar, their global structure and yet respecting the sparse edits of the target image. More results are included in the supplementary.", + "bbox": [ + 498, + 263, + 890, + 400 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To compare with state-of-the-art texture synthesis methods, we fed the target images we sketched to the models trained by adversarial expansion (TexExp) [51]. As can be seen in Fig. 6, TexExp failed to reproduce the fine textures of the reference, as the edited target images are unseen data to its training. We also tested the texture optimization based on a recently proposed textural loss (GCD Loss) [49], where the target images are down-scaled as the initialization in its multi-resolution synthesis. Although plausible local textures are synthesized, the output of this method does not always conform to the user-edited layout (see Fig. 6).", + "bbox": [ + 496, + 400, + 890, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We have also applied our method on nearly homogeneous textures of stationary statistics. Since such textures do not have a prominent global structure, we can simply reshuffle patches of the reference to serve as the target. This, however, might cut some local elements in the example. Alternatively, we can use the reference image as the target and shuffle its inversion code before the sampling, which better preserves the local texture elements. Fig. 7 shows a few examples, demonstrating promising quality.", + "bbox": [ + 496, + 566, + 890, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 709, + 663, + 723 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we analyze the effect of the key components in our method through ablation studies.", + "bbox": [ + 500, + 732, + 888, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "KV-Injection in sampling. As $KV$ -Injection is applied both in the inversion and sampling process, we applied two ablations to study its effect. First, we set $P_1 = P_2 = 0$ , and explore the full parameter space for $S_1$ and $S_2$ , to study the effect of $KV$ -Injection in texture sampling without being affected by the inversion.", + "bbox": [ + 496, + 779, + 890, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fig. 8 shows a matrix of results of this parameter study. A smaller value of $S(S_1 \\& S_2)$ means performing more time", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7771", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/50443225708289e321da932961981303861c63ecd1f051c6a43b3fb8eb60797c.jpg", + "image_caption": [ + "Figure 6. We sketch two targets for each reference, denoted as Edit 1 and Edit 2, respectively. Our self-rectification method generates textures with global structures that faithfully respect the user edits of the target images, while still producing high-quality texture details. In contrast, adversarial expansion (TexExp) [51] does not capture the fine details well. Optimization by GCD Loss [49] reproduces the local textures, but does not always conform to the target image. Note that Edit 2 is used as the input for TexExp, and as the initialization in texture optimization of GCD Loss. More results are included in the supplementary." + ], + "image_footnote": [], + "bbox": [ + 76, + 87, + 893, + 829 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7772", + "bbox": [ + 482, + 945, + 516, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/afa97938b590bf80a91bd6da85eb186cbf05557a2080781c66cc2e17c52e75d7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 88, + 207, + 185 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0a9ba524973edd6fb6309680a5747ea336d91b95af73f2f36411686b4554a6bf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 210, + 88, + 334, + 185 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6262c145e78d74b2d1193ab93ee846d9a1bcaad0eb99fa55b938eb066eae3825.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 336, + 88, + 464, + 185 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/02ec0269d9b4bc0ee85f020804a3266a5ac00a135ae7355947be5a8d772d729b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 186, + 207, + 282 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7125e9d1284caf07b46a5a3842a719d263cc59995ec3dcddc22c48a5a7811bc8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 210, + 186, + 334, + 282 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/aeca6e4d5c72461cc0b796a32f4ad2aeb6f264c51ccdf1b543c33b99c136b7b6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 336, + 186, + 464, + 282 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bb90f141a104f75f975b91e18f326a429a4612cd44b9ebed7f4734109429fe5f.jpg", + "image_caption": [ + "Figure 7. For nearly homogeneous textures (left), we can use patch shuffle of the reference to define a random target layout, where the shuffling could be performed in image space before applying self-rectification (middle), or in latent space after the inversion in the first self-rectification (right)." + ], + "image_footnote": [], + "bbox": [ + 81, + 284, + 207, + 380 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d353d345b76b288015cfcbec2a1caaa096881502f9f0d10d51185a8352152b19.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 210, + 284, + 334, + 380 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1ef9999145be44a9ed679f675ddd306e26a270aa413c91cc046f0fd99d592e0a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 336, + 284, + 464, + 380 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "steps of $KV$ -Injection in sampling, and thus, the target image is rectified to be closer to the reference texture. On the contrary, a larger S reserves more of the target image's layout, however, at the same time, it introduces more structural errors or conflicts, yielding artifacts. None of these results reaches a good trade-off between synthesizing a reasonable global layout (especially considering the user edits) and reproducing local textures of the reference. Nevertheless, we can find a proper parameter setting for S, which is $10\\sim 20$ for $S_1$ , and 5 for $S_2$ . Hence we set the default values of $S_1$ and $S_2$ to be 20 and 5, and use it for all experiments.", + "bbox": [ + 75, + 474, + 468, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "KV-Injection in inversion. Next, we investigate the effect of $KV$ -Injection in our structure-preserving inversion. By fixing $S_1$ and $S$ to default values, we search the parameter space defined by $P_1$ and $P_2$ . As shown in Fig. 9, introducing $KV$ -Injection significantly improves the synthesis result. Both the structural errors and local artifacts are drastically reduced at the same time. Another important point is that, we may have a relatively wide range for setting the value of $P$ , which is empirically suggested by the results: $10\\sim 30$ for $P_1$ , and $5\\sim 15$ for $P_2$ . We usually set $P_1$ as 20, and $P_2$ as 5 in production. See supplementary for the full exploration results and more examples of this study.", + "bbox": [ + 75, + 659, + 467, + 840 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Data augmentation. In many cases, data augmentation may not be necessary, as the pre-trained Stable Diffusion", + "bbox": [ + 75, + 859, + 467, + 888 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4d8a2ebcb997becd41a36c72c0605fc46b8ea38eb747f06aa66b50c8003da72c.jpg", + "image_caption": [ + "Figure 8. Ablation study of $KV$ -Injection in texture sampling. We explore the parameter space for $S_1$ and $S_2$ , with $P_1 = P_2 = 0$ , and adopt the wood texture and its edited target shown in Fig. 1 for this test. The coarse self-rectification results are shown in the 1st column. We can see that changing the starting time-step of $KV$ -Injection in the sampling greatly affects the final output. However, none of these settings yields a qualified balance when considering both the target layout and the texture details. Note setting $S = 50$ (i.e., no $KV$ -Injection in sampling) only results in an approximate layout reconstruction without any details synthesized (see Fig. 5)." + ], + "image_footnote": [], + "bbox": [ + 501, + 85, + 888, + 308 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/352a9547a500faa95228f3b92b821c363d3421f0e9b368af4968f5bbef670f61.jpg", + "image_caption": [ + "Figure 9. Ablation study of $KV$ -Injection in structure-preserving inversion by exploring the parameter space defined by $\\mathrm{P}_1$ and $\\mathrm{P}_2$ . Here, we fixed $S_1 = 20$ and $S_2 = 5$ according to the test of Fig. 8. While compared with Fig. 8, the outputs are significantly improved after introducing $KV$ -Injection to the inversion. Many of these results can be considered good synthesis regarding the target image edited and the source texture of the wood." + ], + "image_footnote": [], + "bbox": [ + 503, + 479, + 888, + 704 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "model already has a certain ability to synthesize rotated texture patterns, except when the reference has dominant directional structures. To allow more free editing for directional textures, augmenting the reference is essential to synthesize", + "bbox": [ + 496, + 824, + 890, + 886 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7773", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a8afb72b0f76985592e857289bc6bff2d390b8120495bfd1888560005d2ae3d0.jpg", + "image_caption": [ + "Reference" + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 176, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b056391f41783179be12d7aaa5e5997cdd8d84f9547dba93611e99972aa3fe2c.jpg", + "image_caption": [ + "Target" + ], + "image_footnote": [], + "bbox": [ + 178, + 88, + 271, + 236 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c3e1520d8318dfa1d4da1127df86f99f02c5a9cced4d0ce74a97da3cc89bf549.jpg", + "image_caption": [ + "w/o Data Augmentation" + ], + "image_footnote": [], + "bbox": [ + 272, + 88, + 366, + 236 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2ea5f75c9435532f64f7830caa19a3197ca0ebcdcabff56c74cad9f4bce03490.jpg", + "image_caption": [ + "w/ Data Augmentation" + ], + "image_footnote": [], + "bbox": [ + 367, + 88, + 462, + 236 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/534b06e54f438892c9a367c5689f2083283f9dc4d17b14457ede7b9d1004e4c2.jpg", + "image_caption": [ + "Figure 10. For textures that contain dominant directional structures, data augmentation is essential to allow the output to admit with user edits. Specifically, we augment each source shown above with three images rotated at angles of $\\pm 45^{\\circ}$ and $90^{\\circ}$ , respectively." + ], + "image_footnote": [], + "bbox": [ + 83, + 354, + 207, + 449 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ab498a1a38151e4a3f98d3e6cb218d4f991081604ca797bf058204f5942fbb60.jpg", + "image_caption": [ + "Reference", + "Figure 11. Guided texture synthesis. The user provides a color layout to guide the output structure. The synthesized outputs follow the outlines well and still exhibit the reference texture details." + ], + "image_footnote": [], + "bbox": [ + 83, + 450, + 207, + 547 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cef84d6cc8c21d4b4e7f5d3e2c4bccdd9c409b51b787346a752f7f01e1581848.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 210, + 354, + 334, + 450 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bbb1d67a42a9f0e99f31263fa7e7f86afe82fb3e4c5cb23991c303673ba44a92.jpg", + "image_caption": [ + "Target Outlines" + ], + "image_footnote": [], + "bbox": [ + 210, + 452, + 334, + 547 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4324871efdf50b5f78be6a9811c3b74f3d7adf1af75b4db27673b10bd48d3710.jpg", + "image_caption": [ + "Figure 12. Our method can also be applied to image editing. For creating the target image, the user can either loosely sketch a layout on the target canvas, or directly edit the input image." + ], + "image_footnote": [], + "bbox": [ + 339, + 354, + 462, + 450 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/06471c22acbe139dc0dc719dc5b4b4e45ede3620947529d5f335f7300c602d1d.jpg", + "image_caption": [ + "Output" + ], + "image_footnote": [], + "bbox": [ + 339, + 452, + 462, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "a more consistent output structure conforming to user edits; see, e.g., Fig. 10 for a comparison on data augmentation, where additional rotated transformations are involved.", + "bbox": [ + 76, + 630, + 468, + 675 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Additional Applications", + "text_level": 1, + "bbox": [ + 76, + 685, + 295, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our self-rectification framework can extend its utility to seemingly other applications. For instance, as illustrated in Fig. 11, users can input a straightforward target image outlining a layout structure with colors compatible with reference textures. To avoid excessive smoothing, random noise is introduced to enhance the target image. The self-rectification process ensures that the target image incorporates texture details from the reference, while adhering to the user-provided layout structure.", + "bbox": [ + 75, + 708, + 468, + 843 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Moreover, our self-rectification method can be employed for image editing beyond textures. Users can edit a given image by employing basic cut-and-paste patch operations,", + "bbox": [ + 75, + 844, + 468, + 890 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9a51b9643c751b762b060d764ba814c86273b026c1b295514fde0c844cb5563b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 88, + 580, + 146 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2afe6df694be01756c8240f9179ef1299e06fc9a54447116826c51131ba8ed4d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 583, + 88, + 656, + 146 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bb60ddbaabe7c74cc98ee1aef82a5e5c78915c69a57f8e45a5d03df27cfb67ca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 658, + 88, + 732, + 146 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e238aa7d944ed2b4b557745d2a3afc45980c961ba1e64c2ee0664362d2aa95da.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 735, + 88, + 808, + 146 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/358d3490f70a74e6db88a679d7ae26e861438c5d9a7f5b69114a375982530850.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 812, + 88, + 885, + 146 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c8db08770fe2c85e968098a3a470d6eb5d60cecfa91569b505fc977e009f47a8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 148, + 580, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1a2bf52c8fe4e57ed4c58a4e6f202e86b891917e6d28a46581ea78904ac0e228.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 581, + 148, + 656, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bce3be17612a92ddcb44b9eb0a44057bec2c173e9aa516d57bf3469a1a74eebb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 658, + 148, + 732, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/02fdb624e503a3f6239afcb3869c7113f5922049123406920f0646bb4a821e58.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 735, + 148, + 808, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f8e5454382b1077ea07caa3c2c89ed6a65473f5abd983a49c60dcd2efe309a94.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 812, + 148, + 885, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d9c25b0cb2bd4c3eaf2895928ea8b6226aaf0fa49bb391ae06d63b1d7dfa6880.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 207, + 580, + 263 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a8fbc5542a9f94fc36d1c83a53229abd1ae4ca0414701b57e45bb95fe341e740.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 581, + 207, + 656, + 263 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7d4c643c70964ad9fb443f797913791934fe65944859c739a27d3e5dc935cba4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 658, + 207, + 732, + 263 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/41a1d0e9e1e4219f43ae2316918a889385ff34160e32da01b88fba46be5b031a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 735, + 207, + 808, + 263 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f9b59d1177084526d5b05a5dc55d39b0c9a5130f40d2fa80d6f6428d4232e83a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 812, + 207, + 885, + 263 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e4ef03c3ec70425ee66ca0d667d057b4fad343a038f1abb8090a2fc4619325c7.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 506, + 265, + 580, + 321 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3f45de8b56a1e0741dc611067adf2cc1110f9df159230430980544df39218e37.jpg", + "image_caption": [ + "Edit 1" + ], + "image_footnote": [], + "bbox": [ + 581, + 265, + 656, + 321 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/de5366bb400aab8215eeb4df37e15d0c6ea45df78f6bde3f504aead4f5eefedd.jpg", + "image_caption": [ + "Output 1" + ], + "image_footnote": [], + "bbox": [ + 658, + 265, + 732, + 321 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bdf22ee1553340991d3b29f835b78f6471b1bf0fbff8e69618fb910e3df95a9e.jpg", + "image_caption": [ + "Edit 2" + ], + "image_footnote": [], + "bbox": [ + 735, + 265, + 808, + 321 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ffc370ac281e528725a1f8b668f16a5d44e25a067add851875346c00a5590670.jpg", + "image_caption": [ + "Output 2" + ], + "image_footnote": [], + "bbox": [ + 812, + 265, + 885, + 321 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and the resulting crude edits undergo self-rectification. Fig. 12 presents several instances of such image editing operations, presenting two edits for each input image.", + "bbox": [ + 500, + 404, + 890, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 463, + 625, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our work addresses the intricate challenge of synthesizing non-stationary textures, offering a method that empowers users to efficiently design new textures with unprecedented controllability. This stands as a notable improvement over existing methods, providing a user-friendly process, which consists of two steps: Users begin with an initial rough edit using conventional image editing tools, followed by an automated self-rectification process. This process leverages a pre-trained diffusion network and injections of self-attention features, showcasing flexibility in synthesizing a diverse range of challenging non-stationary textures.", + "bbox": [ + 496, + 488, + 890, + 654 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the future, we would like to explore the extension of our approach to synthesize large-scale textures. Additionally, there is a promising avenue to incorporate semantic understanding into the self-rectification process, enhancing alignment with user intent. The integration of texture semantics holds the potential to yield contextually relevant and visually appealing synthesized textures.", + "bbox": [ + 496, + 655, + 890, + 760 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements This work was supported in parts by NSFC (62161146005, U21B2023, U2001206), GD NSF (2022A1515010221), DEGP Innovation Team (2022 KCXTD025), Shenzhen Science and Technology Program (KQTD20210811090044003, RCJC20200714114435012), ISF (3441/21), Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ) and Scientific Development Funds of Shenzhen University.", + "bbox": [ + 496, + 779, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7774", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yuval Alaluf, Daniel Garibi, Or Patashnik, Hadar Averbuch-Elor, and Daniel Cohen-Or. Cross-image attention for zero-shot appearance transfer, 2023. 2, 3, 4", + "[2] Yuval Alaluf, Elad Richardson, Gal Metzer, and Daniel Cohen-Or. A neural space-time representation for text-to-image personalization, 2023. 2", + "[3] Rachele Bellini, Yanir Kleiman, and Daniel Cohen-Or. Time-varying weathering in texture space. ACM Trans. on Graphics (Proc. of SIGGRAPH), 35(4):1-11, 2016. 2", + "[4] Sagie Benaim, Ron Mokady, Amit Bermano, and Lior Wolf. Structural analogy from a single image pair. Computer Graphics Forum, 40(1):249-265, 2021. 2", + "[5] Urs Bergmann, Nikolay Jetchev, and Roland Vollgraf. Learning texture manifolds with the periodic spatial gan. In Proc. of IEEE Int. Conf. on Machine Learning, page 469-477, 2017. 2", + "[6] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. MasaCtrl: tuning-free mutual self-attention control for consistent image synthesis and editing. In Proc. of Int. Conf. on Computer Vision, pages 22560–22570, October 2023. 2, 3, 4, 5", + "[7] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models, 2023. 2", + "[8] Alexei A. Efros and William T. Freeman. Image quilting for texture synthesis and transfer. In Proc. of SIGGRAPH, pages 341-346, 2001. 2", + "[9] Alexei A. Efros and Thomas K. Leung. Texture synthesis by non-parametric sampling. In Proc. of Int. Conf. on Computer Vision, 1999. 2", + "[10] Anna Frühstück, Ibraeem Alhashim, and Peter Wonka. Titegan: synthesis of large-scale non-homogeneous textures. ACM Trans. on Graphics (Proc. of SIGGRAPH), 38(4):1-11, 2019. 2", + "[11] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion, 2022. 2", + "[12] Leon Gatys, Alexander S Ecker, and Matthias Bethge. Texture synthesis using convolutional neural networks. In Advances in Neural Information Processing Systems, 2015. 2", + "[13] Eric Heitz, Kenneth Vanhoey, Thomas Chambon, and Laurent Belcour. A sliced wasserstein loss for neural texture synthesis. In Proc. of IEEE Conf. on Computer Vision & Pattern Recognition, June 2021. 2", + "[14] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. In Proc. of International Conference on Learning Representations, 2023. 2", + "[15] Aaron Hertzmann, Charles E. Jacobs, Nuria Oliver, Brian Curless, and David H. Salesin. Image analogies. In Proc. of SIGGRAPH, page 327-340, 2001. 2", + "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 3" + ], + "bbox": [ + 78, + 114, + 467, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models, 2021. 2", + "[18] Nikolay Jetchev, Urs Bergmann, and Roland Vollgraf. Texture synthesis with spatial generative adversarial networks. arXiv preprint arXiv:1611.08207, 2016. 2", + "[19] Alexandre Kaspar, Boris Neubert, Dani Lischinski, Mark Pauly, and Johannes Kopf. Self tuning texture optimization. Computer Graphics Forum (Proc. of Eurographics), 34(2):349-359, may 2015. 2", + "[20] Anant Khandelwal. Infusion: Inject and attention fusion for multi concept zero shot text based video editing. arXiv preprint arXiv:2308.00135, 2023. 2", + "[21] Vivek Kwatra, Irfan Essa, Aaron Bobick, and Nipun Kwatra. Texture optimization for example-based synthesis. ACM Trans. on Graphics (Proc. of SIGGRAPH), August 2005. 2", + "[22] Vivek Kwatra, Arno Schödl, Irfan Essa, Greg Turk, and Aaron Bobick. Graphcut textures: Image and video synthesis using graph cuts. ACM Trans. on Graphics (Proc. of SIGGRAPH), 22(3):277-286, jul 2003. 2", + "[23] Chuan Li and Michael Wand. Precomputed real-time texture synthesis with markovian generative adversarial networks. In Proc. of Euro. Conf. on Computer Vision, pages 702-716, 2016. 2", + "[24] Yitzchak David Lockerman, Basile Sauvage, Rémi Allègre, Jean-Michel Dischler, Julie Dorsey, and Holly Rushmeier. Multi-scale label-map extraction for texture synthesis. ACM Trans. on Graphics (Proc. of SIGGRAPH), 35(4):140:1-140:12, July 2016. 2", + "[25] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models, 2022.", + "[26] M. Lukáč, J. Fíšer, P. Asente, J. Lu, E. Shechtman, and D. Sykora. Brushables: Example-based edge-aware directional texture painting. Computer Graphics Forum (Proc. of Pacific Conf. on Computer Graphics & Applications), 34(7):257-267, oct 2015. 2", + "[27] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2", + "[28] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2):3, 2022. 2", + "[29] Carlos Rodriguez-Pardo and Elena Garces. Seamlessgan: Self-supervised synthesis of tileable texture maps. IEEE Trans. Visualization & Computer Graphics, 2022. 2", + "[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proc. of IEEE Conf. on Computer Vision & Pattern Recognition, pages 10684-10695, 2022. 2, 3", + "[31] Amir Rosenberger, Daniel Cohen-Or, and Dani Lischinski. Layered shape synthesis: Automatic generation of control maps for non-stationary textures. ACM Trans. on Graphics (Proc. of SIGGRAPH), 28(5):107:1-9, Dec. 2009. 1, 2" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7775", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Tamar Rott Shaham, Tali Dekel, and Tomer Michaeli. Singan: Learning a generative model from a single natural image. In Proc. of Int. Conf. on Computer Vision, 2019. 2", + "[33] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proc. of IEEE Conf. on Computer Vision & Pattern Recognition, 2023. 2", + "[34] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdream booth: Hypernetworks for fast personalization of text-to-image models, 2023. 2", + "[35] Omry Sendik and Daniel Cohen-Or. Deep correlations for texture synthesis. ACM Trans. on Graphics, 36(5):161:1-161:15, July 2017. 2", + "[36] Wu Shi and Yu Qiao. Fast texture synthesis via pseudo optimizer. In Proc. of IEEE Conf. on Computer Vision & Pattern Recognition, pages 5498-5507, 2020. 2", + "[37] Assaf Shocher, Shai Bagon, Phillip Isola, and Michal Irani. Ingan: Capturing and retargeting the \"dna\" of a natural image. In Proc. of Int. Conf. on Computer Vision, 2019. 2", + "[38] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3", + "[39] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proc. of IEEE Conf. on Computer Vision & Pattern Recognition, pages 1921-1930, June 2023. 2, 3", + "[40] Dmitry Ulyanov, Vadim Lebedev, Andrea Vedaldi, and Victor Lempitsky. Texture networks: Feed-forward synthesis of textures and stylized images. In Proc. of IEEE Int. Conf. on Machine Learning, page 1349-1357, 2016. 2", + "[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 30, 2017. 3", + "[42] Jiaping Wang, Xin Tong, Stephen Lin, Minghao Pan, Chao Wang, Hujun Bao, Baining Guo, and Heung-Yeung Shum. Appearance manifolds for modeling time-variant appearance of materials. ACM Trans. on Graphics (Proc. of SIGGRAPH), 25(3):754–761, 2006. 2", + "[43] Li-Yi Wei and Marc Levoy. Fast texture synthesis using tree-structured vector quantization. In Proc. of SIGGRAPH, pages 479–488, 2000. 2", + "[44] Yonatan Wexler, Eli Shechtman, and Michal Irani. Spacetime completion of video. IEEE Trans. Pattern Analysis & Machine Intelligence, 29(3):463-476, 2007. 2", + "[45] Li-Yi Wie, Sylvain Lefebvre, Vivek Kwatra, and Greg Turk. State of the Art in Example-based Texture Synthesis. In M. Pauly and G. Greiner, editors, Eurographics 2009 - State of the Art Reports. The Eurographics Association, 2009. 1", + "[46] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. arXiv preprint arXiv:2211.13227, 2022. 2" + ], + "bbox": [ + 78, + 90, + 468, + 883 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. In arXiv preprint arxiv:2308.06721, 2023. 2", + "[48] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 2", + "[49] Yang Zhou, Kaijian Chen, Rongjun Xiao, and Hui Huang. Neural texture synthesis with guided correspondence. In Proc. of IEEE Conf. on Computer Vision & Pattern Recognition, pages 18095-18104, 2023. 2, 5, 6", + "[50] Yang Zhou, Huajie Shi, Dani Lischinski, Minglun Gong, Johannes Kopf, and Hui Huang. Analysis and controlled synthesis of inhomogeneous textures. Computer Graphics Forum (Proc. of Eurographics), 36(2):199-212, 2017. 2", + "[51] Yang Zhou, Zhen Zhu, Xiang Bai, Dani Lischinski, Daniel Cohen-Or, and Hui Huang. Non-stationary texture synthesis by adversarial expansion. ACM Trans. on Graphics (Proc. of SIGGRAPH), 37(4):49:1-49:13, 2018. 1, 2, 5, 6" + ], + "bbox": [ + 501, + 90, + 890, + 359 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7776", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Generating Non-Stationary Textures using Self-Rectification/3ddc2997-ec2e-480d-b380-c7012481f817_model.json b/2024/Generating Non-Stationary Textures using Self-Rectification/3ddc2997-ec2e-480d-b380-c7012481f817_model.json new file mode 100644 index 0000000000000000000000000000000000000000..53f47d8aecba7d1cdeaf28cef1dd089a63670cb3 --- /dev/null +++ b/2024/Generating Non-Stationary Textures using Self-Rectification/3ddc2997-ec2e-480d-b380-c7012481f817_model.json @@ -0,0 +1,2541 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.181, + 0.131, + 0.79, + 0.154 + ], + "angle": 0, + "content": "Generating Non-Stationary Textures using Self-Rectification" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.18, + 0.877, + 0.22 + ], + "angle": 0, + "content": "Yang Zhou1 Rongjun Xiao1 Dani Lischinski2 Daniel Cohen-Or1,3 Hui Huang1* \n1Shenzhen University 2The Hebrew University of Jerusalem 3Tel Aviv University" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.252, + 0.314, + 0.267 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.283, + 0.474, + 0.555 + ], + "angle": 0, + "content": "This paper addresses the challenge of example-based non-stationary texture synthesis. We introduce a novel two-step approach wherein users first modify a reference texture using standard image editing tools, yielding an initial rough target for the synthesis. Subsequently, our proposed method, termed \"self-rectification\", automatically refines this target into a coherent, seamless texture, while faithfully preserving the distinct visual characteristics of the reference exemplar. Our method leverages a pre-trained diffusion network, and uses self-attention mechanisms, to gradually align the synthesized texture with the reference, ensuring the retention of the structures in the provided target. Through experimental validation, our approach exhibits exceptional proficiency in handling non-stationary textures, demonstrating significant advancements in texture synthesis when compared to existing state-of-the-art techniques. Code is available at https://github.com/xiaorongjun000/Self-Rectification" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.584, + 0.21, + 0.599 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.609, + 0.47, + 0.699 + ], + "angle": 0, + "content": "Example-based texture synthesis aims to generate a texture that faithfully captures all the visual characteristics of a provided reference texture exemplar. The key challenge is to generate a texture that visually mimics the reference, while avoiding exact replication and without producing conspicuous, unnatural artifacts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.7, + 0.47, + 0.805 + ], + "angle": 0, + "content": "Over the past decades, numerous methods have emerged for synthesizing textures from examples, and many have demonstrated impressive results, particularly for homogeneous textures that can be accurately captured by stationary models [45]. Nonetheless, a significant challenge remains when dealing with real-world textures that are inherently inhomogeneous, and non-stationary." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.806, + 0.47, + 0.867 + ], + "angle": 0, + "content": "Non-stationary textures exhibit distinctive attributes, such as sprawling irregular large-scale structures or variations in attributes such as color, local orientation, and local scale. Fig. 1 (left) shows two examples. Mimicking such" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.251, + 0.631, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.351, + 0.631, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.251, + 0.761, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.351, + 0.761, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.763, + 0.251, + 0.892, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.763, + 0.351, + 0.892, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.46, + 0.895, + 0.543 + ], + "angle": 0, + "content": "Figure 1. Our method takes as input a reference texture (left), and a crude target texture provided by the user (middle column), which may lack coherence and completeness. Self-rectification is used to transform the target into a visually coherent texture (right) that complies with the structure of the crude target, while exhibiting the visual characteristics of the reference texture." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.583, + 0.892, + 0.614 + ], + "angle": 0, + "content": "complex structures and variations through example-based synthesis is a long-standing challenging task [31]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.893, + 0.764 + ], + "angle": 0, + "content": "The emergence of neural networks has provided powerful means to deal with non-stationary textures. Zhou et al. [51] introduced a method that involves overfitting a GAN, where the encoder extracts structural guidance for the decoder. This approach provides a viable means of spatially extending non-stationary textures while preserving the visual characteristics of the reference. However, this technique requires an extremely long optimization process for a single texture exemplar, and moreover, it fails to provide any controllability or editability." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.764, + 0.895, + 0.901 + ], + "angle": 0, + "content": "In this paper, we introduce a two-step lazy-editing technique where the user first edits the given reference texture using conventional image editing tools, to obtain an extremely rough initial result, which may be incomplete and incoherent, as demonstrated in Fig. 1 (middle). Next, our technique automatically rectifies the edited texture into a regularized, coherent and seamless texture that follows the rough target, while retaining the local characteristics of the reference; see, e.g., Fig. 1 (right). We term this regulariza" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.222, + 0.888 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7767" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "tion process self-rectification, since the texture is rectified based on the input texture itself." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.122, + 0.47, + 0.378 + ], + "angle": 0, + "content": "To self-rectify the crude target, we use a pre-trained diffusion network, and synthesize the final rectified result while utilizing the intermediate byproducts obtained by inverting both the initial target and the reference exemplar. Inspired by editing approaches that leverage self-attention [1,6,20], we extract certain self-attention features of the diffusion model during the inversion of the rough target, as well as the reference. These features are then injected into various steps in the inversion and denoising of the generated texture. Such feature injection forms \"cross-attention\" between the reference and target features [1, 6], and thereby, in the course of the diffusion process, the target texture is progressively refined to be increasingly locally similar to the reference, while retaining the global structure prescribed by the provided target. This self-rectification occurs in two passes, first addressing the larger scale structure and subsequently focusing on finer local details." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.379, + 0.47, + 0.56 + ], + "angle": 0, + "content": "To enhance the synthesis quality, we augment the reference texture with several transformed copies of it, which increases the diversity of the reference patterns, thereby improving the compatibility between the reference and the synthesized result. The augmented source features are injected into the corresponding target layers as well. Furthermore, we show that our method can also be applied to lazy editing of natural images, using the same means of synthesizing highly non-stationary textures. Experiments show that our method can deal with a large scope of nonstationary textures, with unprecedented flexibility and quality, compared to the state-of-the-art." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.573, + 0.219, + 0.589 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.599, + 0.468, + 0.794 + ], + "angle": 0, + "content": "Non-stationary texture synthesis. Early methods in example-based texture synthesis mainly focused on synthesizing textures with stationary characteristics [8, 9, 19, 21, 22, 43, 44]. To cope with non-stationary textures, researchers typically involve additional guidance for the control of distinctive attributes, e.g., using label maps to guide the layout synthesis of composite textures [15, 24, 31], representing the spatial variations of weathering textures with age/progression maps [3, 42], and describing the local orientations of directional textures with vector fields [26, 50]. Although the aforementioned methods have shown success, the guidance required from the user is tedious to provide, yet still limited." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.47, + 0.901 + ], + "angle": 0, + "content": "In the deep learning era, neural methods for texture generation have rapidly emerged, either through new texture losses for texture optimization [12, 13, 49] or via training generative networks [4, 5, 10, 18, 23, 29, 32, 35-37, 40, 51]. Among these methods, Sendik and Cohen-Or [35] attempted to preserve the non-local structures of non-stationary textures by regularizing the feature correla" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.087, + 0.895, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.235, + 0.894, + 0.32 + ], + "angle": 0, + "content": "Figure 2. Framework overview. Given a reference texture \\( I^R \\), we allow the user to quickly build a target image \\( I^{tar} \\) in a lazy-editing manner. A coarse-to-fine synthesis is performed by running self-rectification twice. The coarse stage synthesizes a coarse yet complete overall structure, and the fine stage refines its output \\( I_{\\text{coarse}}^* \\) with finer and more accurate details, producing the final result \\( I^* \\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.345, + 0.892, + 0.468 + ], + "angle": 0, + "content": "tion between different locations. Zhou et al. [51] proposed overfitting a GAN to learn the expansion from a small texture block to a larger one containing it. Their approach involves the GAN's encoder extracting the global structure of an input texture, duplicated by bottleneck residual blocks before decoding. Although these trained models can effectively extend non-stationary textures, their overfitting nature severely restricts their generalization and controllability." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.485, + 0.892, + 0.606 + ], + "angle": 0, + "content": "Diffusion-based image synthesis. The emergence of large-scale generative diffusion models, such as Stable Diffusion (SD) [30] and DALLE-2 [28], has revolutionized image synthesis due to their unprecedented generation power. To make the pre-trained diffusion models synthesize on image conditions, many solutions are proposed, including optimizing prompt tokens [2, 11], fine-tuning the entire model [33], or training an additional adapter [17,34,46-48]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.606, + 0.892, + 0.787 + ], + "angle": 0, + "content": "Recently, researchers have investigated the role of the intermediate attention maps and features in the diffuser [7, 14,27,39], finding them crucial for layout/structure synthesis. The method we present in this paper, builds upon the mechanism of injecting keys and values from attention layers of one diffusion process into another, as a means for transferring visual features between images [1,6,20]. We use such injection to \"copy\" the local patterns from a source texture to a target one. In addition to the injection we also incorporate coarse-to-fine and data augmentation schemes, forming the self-rectification framework for generating nonstationary textures." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.801, + 0.591, + 0.815 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Our method employs a two-step process for synthesizing a new texture \\( I^{*} \\) resembling a provided non-stationary reference texture \\( I^{R} \\), while guided by a user-provided rough target \\( I^{tar} \\), as depicted in Fig. 2. Initially, the user rapidly creates \\( I^{tar} \\), by assembling patches from the source refer" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7768" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.087, + 0.891, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.27, + 0.895, + 0.327 + ], + "angle": 0, + "content": "Figure 3. Our self-rectification synthesizes an output texture \\( I^{*} \\) via structure-preserving inversion from a rough target image \\( I^{lar} \\) and fine texture sampling using the reference \\( I^{R} \\). Both processes require the injection of self-attention features (\\( KV \\)) from the DDIM inversion of a corresponding reference. More specifically, for structure-preserving inversion, the reference is the target image itself, denoted as \\( I^{IR} \\). For fine texture sampling, the input exemplar \\( I^{R} \\) is used to inject features that help to synthesize a plausible output with fine texture details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.353, + 0.47, + 0.473 + ], + "angle": 0, + "content": "ence texture. Such a lazy-edit forms a basic layout for the output image, \\( I^{*} \\). However, this initial sketch may be incoherent or incomplete, hence necessitating rectification. Subsequently, the texture undergoes self-rectification to align with both the user's coarse layout and the reference image's detailed texture characteristics. This self-rectification process is depicted in Fig. 2 and is executed in two stages: coarse rectification, followed by a finer one." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.473, + 0.47, + 0.608 + ], + "angle": 0, + "content": "Considering that the rough target \\( I^{tar} \\) comprises patches derived from the reference texture \\( I^R \\), we implement a process of self-rectification, detailed below. In a broad sense, we utilize a pre-trained latent diffusion model to invert both the target image and the reference image into an initial latent noise. This is followed by employing feature injection during the sampling process of \\( I^* \\). The features from \\( I^{tar} \\) guide the structural synthesis and the features from \\( I^R \\) guide the fine texture details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.609, + 0.47, + 0.655 + ], + "angle": 0, + "content": "Below, we first briefly review diffusion models and their self-attention and the cross-KV-injection mechanisms, and then proceed to present our self-rectification technique." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.663, + 0.353, + 0.68 + ], + "angle": 0, + "content": "3.1. DDIM Sampling and Inversion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.687, + 0.47, + 0.779 + ], + "angle": 0, + "content": "Denoising diffusion models [16, 38] involve two processes: a noising process that gradually transforms an input image into Gaussian noise, and a denoising/sampling process that generates images from Gaussian samples. Using DDIM sampling [38], starting from a noise sample \\( z_{\\mathrm{T}} \\), one can generate a clean sample \\( z_{0} \\) via T deterministic steps:" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.787, + 0.47, + 0.806 + ], + "angle": 0, + "content": "\\[\nz _ {t - 1} = \\sqrt {\\bar {\\alpha} _ {t - 1}} f _ {\\theta} (z _ {t}, t) + \\sqrt {1 - \\bar {\\alpha} _ {t - 1}} \\epsilon_ {\\theta} (z _ {t}, t), \\quad (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.816, + 0.47, + 0.861 + ], + "angle": 0, + "content": "where \\(\\epsilon_{\\theta}\\) is a noise prediction network conditioned on the current noisy sample \\(z_{t}\\) and timestamp \\(t\\). \\(\\bar{\\alpha}_{t}\\) is the noise scaling factor defined in [38], and \\(f_{\\theta}(z_t,t)\\) is" + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.871, + 0.47, + 0.906 + ], + "angle": 0, + "content": "\\[\nf _ {\\theta} \\left(z _ {t}, t\\right) = \\frac {z _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta} \\left(z _ {t} , t\\right)}{\\sqrt {\\bar {\\alpha} _ {t}}}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.353, + 0.892, + 0.398 + ], + "angle": 0, + "content": "In the opposite direction, \\(DDIM\\) inversion [38] of a given image \\(z_0\\) is the process of incrementally adding deterministic noise, until obtaining \\(z_{\\mathrm{T}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.408, + 0.892, + 0.427 + ], + "angle": 0, + "content": "\\[\nz _ {t + 1} = \\sqrt {\\bar {\\alpha} _ {t + 1}} f _ {\\theta} \\left(z _ {t}, t\\right) + \\sqrt {1 - \\bar {\\alpha} _ {t + 1}} \\epsilon_ {\\theta} \\left(z _ {t}, t\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.436, + 0.892, + 0.452 + ], + "angle": 0, + "content": "Such inversion leads to a nearly faithful reconstruction [38]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.46, + 0.81, + 0.475 + ], + "angle": 0, + "content": "3.2. Stable Diffusion and Self-Attention." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.483, + 0.892, + 0.574 + ], + "angle": 0, + "content": "We base our texture synthesis framework on Stable Diffusion (SD) [30], which is a pretrained latent diffusion model consisting of an encoder \\(\\mathcal{E}\\) that maps an input image into the latent space, and a decoder \\(\\mathcal{D}\\) that reconstructs a latent code back into image space. Both DDIM sampling and DDIM inversion are performed in the SD latent space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.574, + 0.892, + 0.65 + ], + "angle": 0, + "content": "The noise predictor \\(\\epsilon_{\\theta}\\) of SD is a large-scale U-Net that contains multiple self-attention modules [41]. Each self-attention layer transforms its input intermediate feature map (also called spatial features) into an attended representation by the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.659, + 0.892, + 0.695 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t} (Q, K, V) = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d}}\\right) V, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.892, + 0.809 + ], + "angle": 0, + "content": "where \\( Q, K \\), and \\( V \\) are the queries, keys, and values, respectively, obtained by learned linear projections of the same input spatial features, having dimension \\( d \\). The self-attention mechanism uses the similarities between the queries and keys as attention scores to weigh the importance or relevance of the values. Relevant info is thus aggregated as the attended representation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.817, + 0.638, + 0.833 + ], + "angle": 0, + "content": "3.3. KV-Injection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The self-attention features contain rich information about both the large-scale structures and local fine details of an input image [1, 6, 39]. Tumanyan et al. [39] demonstrate that by injecting the spatial features and the queries" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7769" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.087, + 0.473, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.267, + 0.471, + 0.351 + ], + "angle": 0, + "content": "Figure 4. Visualization of the intermediate latent codes in the inversion. For the standard DDIM inversion (top), the U-Net predicts noise to diffuse the distinctive structures so as to transform the input into Gaussian noise. In contrast, our structure-preserving inversion (bottom) reserves the distinctive patterns from user edits along the inversion process. See texts in Sec. 3.4 for more details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.377, + 0.468, + 0.483 + ], + "angle": 0, + "content": "and keys \\((QK)\\) from the self-attention layers of a source (guidance) image into the corresponding layer of the generated target image during the sampling process, one can preserve the layout of the source while modifying its appearance. Injecting only the \\(KV\\) features of a source reference, instead, transfers its appearance, including textures, to the generated target [1,6]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.483, + 0.468, + 0.604 + ], + "angle": 0, + "content": "In our case, the target image is a rough and incomplete guidance, typically containing only a few source patches rotated and placed freely by the user. A complete overall structure needs to be synthesized reasonably yet conforming to user's constraints. Locally, however, the resulting texture should resemble the reference. To achieve these goals, we adapt the \\( KV \\)-injection into both the DDIM inversion and sampling, bringing a novel self-rectification operation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.613, + 0.241, + 0.628 + ], + "angle": 0, + "content": "3.4. Self-rectification" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.636, + 0.47, + 0.727 + ], + "angle": 0, + "content": "As shown in Fig. 3, our self-rectification consists of two parts: structure-preserving inversion and fine texture sampling. Both are performed in the SD latent space. The last latent code of inversion is used as the starting code of the sampling. The core modification to both processes is the \\( KV \\)-injection of self-attention features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.745, + 0.47, + 0.896 + ], + "angle": 0, + "content": "Structure-preserving inversion. The standard DDIM inversion (Eq. (3)) progressively transforms an input image (SD latent code) into pure Gaussian noise. At each time step, the noise to be added is predicted by the U-Net. As visualized in Fig. 4, for a given target image, the noise predicted by the U-Net for early time steps (\\(e.g., t \\leq 20\\)) is mainly distributed to \"diffuse\" the prominent structures so that the distinctive patterns from user edits get scattered and random. As the inversion progresses, the magnitude of the noise added in each step becomes smaller and similar for" + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.089, + 0.895, + 0.144 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.151, + 0.894, + 0.263 + ], + "angle": 0, + "content": "Figure 5. Visualization of the intermediate latent codes in the fine texture sampling. Here, for the first 20 steps (from \\( t = 50 \\) to 30), we perform the standard DDIM sampling to reconstruct the target layout. Next, we perform \\( KV \\)-injection in the remaining sampling steps (\\( t = 30 \\) to 0), to synthesize fine textures for the output image. The rightmost shows the result produced by simply performing standard DDIM sampling for all steps, i.e., \\( S = 50 \\). No additional structure is synthesized to complete the user edits." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.289, + 0.892, + 0.44 + ], + "angle": 0, + "content": "all latent pixels, since the \"diffusion\" is getting close to done. As no text prompt is involved, the noise prediction in each inversion step is dominantly determined by the self-attention mechanism in the U-Net. Our key observation is that if we inject the \\( KV \\) features from a large time step \\( t_1 \\) (\\( \\gg \\mathrm{T} / 2 \\)) into an early time step \\( t_2 \\) (\\( \\ll \\mathrm{T} / 2 \\)), the noise predicted at \\( t_2 \\) will be smaller and more spatially uniform. The distinctive patterns of \\( I^{tar} \\), reflecting the user's edits, are thus better preserved. Hence, we refer to this process as structure-preserving inversion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.44, + 0.892, + 0.531 + ], + "angle": 0, + "content": "Specifically, given a rough target image \\(I^{tar}\\), we invert it twice. The first inversion is a standard DDIM inversion. The produced self-attention features along the noisig steps are regarded as the inversion reference (IR) for the second inversion. The self-attention during the second inversion, at time step \\(t\\), is now given by:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.541, + 0.895, + 0.595 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t} \\left(Q _ {t} ^ {\\text {t a r}}, K _ {\\mathrm {T} - t} ^ {\\mathrm {I R}}, V _ {\\mathrm {T} - t} ^ {\\mathrm {I R}}\\right) = \\operatorname {S o f t m a x} \\left(\\frac {Q _ {t} ^ {\\text {t a r}} \\left(K _ {\\mathrm {T} - t} ^ {\\mathrm {I R}}\\right) ^ {T}}{\\sqrt {d}}\\right) V _ {\\mathrm {T} - t} ^ {\\mathrm {I R}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.596, + 0.892, + 0.686 + ], + "angle": 0, + "content": "where \\( \\mathrm{T} \\) denotes the total number of steps (we use \\( \\mathrm{T} = 50 \\)). Here \\( KV \\) features are injected in a reverse order in the second round. We would like to stress that we have also experimented with using an offset, i.e., replacing \\( \\mathrm{T} - t \\) with \\( t + offset \\) in the above equation, but found no advantage (see supplementary for more analysis and comparisons)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.687, + 0.892, + 0.822 + ], + "angle": 0, + "content": "During the second inversion, we do not perform \\( KV \\) injections for all time steps. We set a parameter \\( P \\) that defines the time step beyond which we continue with standard inversion. Fig. 4 shows the effect of our structure-preserving inversion, where \\( P = 30 \\). Since the final latent code after the second inversion still contains rich information about the distinctive structures from \\( I^{tar} \\), the subsequent sampling that starts from this latent code is guided to synthesize a global structure that complies with \\( I^{tar} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Fine texture sampling. We start the DDIM sampling from the final structure-preserving latent code. The sampling uses the U-Net to predict the noise to be removed, as defined in Eq. (1). However, simply performing standard" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7770" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.198 + ], + "angle": 0, + "content": "DDIM sampling for all denoising steps results in a nearly reconstructed target image since DDIM sampling is deterministic. Therefore, we set the first S steps (i.e., from time step T to T - S) to reconstruct the target layout to a certain extent. For the remaining T - S denoising steps, we synthesize the output image \\( I^{*} \\) by matching fine textures from the reference via \\( KV \\)-injection from the reference \\( I^{R} \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.198, + 0.468, + 0.258 + ], + "angle": 0, + "content": "To this end, we first DDIM-invert the reference texture \\( I^{R} \\). At denoising time step \\( t \\) (\\( t > \\mathrm{T} - \\mathrm{S} \\)), the \\( KV \\) features extracted during the inversion are injected into the corresponding self-attention layers of the synthesized texture:" + }, + { + "type": "equation", + "bbox": [ + 0.098, + 0.268, + 0.468, + 0.31 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t} \\left(Q _ {t} ^ {*}, K _ {t} ^ {R}, V _ {t} ^ {R}\\right) = \\operatorname {S o f t m a x} \\left(\\frac {Q _ {t} ^ {*} \\left(K _ {t} ^ {R}\\right) ^ {T}}{\\sqrt {d}}\\right) V _ {t} ^ {R}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.32, + 0.468, + 0.381 + ], + "angle": 0, + "content": "This forms a cross-image attention, where corresponding fine local patterns in the reference are transferred to the output image in a plausible manner. Fig. 5 visualizes the intermediate process of our texture sampling, where \\( S = 20 \\)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.39, + 0.468, + 0.405 + ], + "angle": 0, + "content": "3.5. Implementation details and data augmentation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.413, + 0.468, + 0.579 + ], + "angle": 0, + "content": "In the user editing phase, we fill the background of the target canvas with pixels randomly drawn from the source texture, such that the encoding of the target image would not deviate too far from the source in the SD latent space. As the self-rectification is performed twice, the output of coarse stage \\( I_{\\text{coarse}}^* \\), will be used as the input target image of fine stage to produce the final output \\( I^* \\). In contrast, for the inversion reference \\( I^{IR} \\) involved in structure-preserving inversion of the two self-rectifications, we use the same initial target image that contains the original user edits. See the supplementary for the full algorithm pseudo-code." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.58, + 0.468, + 0.715 + ], + "angle": 0, + "content": "Considering the parameter settings: since the coarse stage aims for structure synthesis, relatively large \\(\\mathbf{P}\\) and S are required in self-rectification, and vice versa in the fine stage. More specifically, let \\(\\mathrm{P_1}\\), \\(\\mathrm{P_2}\\), \\(\\mathrm{S_1}\\), and \\(\\mathrm{S_2}\\), denote the parameters used in the two rounds of structure-preserving inversion and fine texture sampling. We typically set \\(\\mathrm{P_1} = 20\\), \\(\\mathrm{P_2} = 5\\), \\(\\mathrm{S_1} = 20\\), and \\(\\mathrm{S_2} = 5\\). Following [6], we choose the \\(KV\\) features from the 10th to 15th self-attention layers of the U-Net decoder part." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.715, + 0.468, + 0.837 + ], + "angle": 0, + "content": "To further improve the synthesis quality, especially when dealing with textures containing a dominant directional structure, such as the leaf shown in Fig. 1, we can introduce a few transformed images (flips and rotations) to augment the reference texture. We concatenate the new attention features from the augmentation to the original reference feature. For example, when we have \\( n \\) augmented source textures, the \\( K^R \\) and \\( V^R \\) in Eq. (6) is now given by" + }, + { + "type": "equation", + "bbox": [ + 0.127, + 0.847, + 0.468, + 0.896 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l l} K ^ {R} & = \\operatorname {C o n c a t} \\left(K _ {(0)} ^ {R}, K _ {(1)} ^ {R}, \\dots , K _ {(n)} ^ {R}\\right), \\\\ V ^ {R} & = \\operatorname {C o n c a t} \\left(V _ {(0)} ^ {R}, V _ {(1)} ^ {R}, \\dots , V _ {(n)} ^ {R}\\right) \\end{array} \\right. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.091, + 0.892, + 0.122 + ], + "angle": 0, + "content": "where \\(K_{(i)}^{R}\\), and \\(V_{(i)}^{R}\\) are the self-attention features acquired from the DDIM inversion of the augmented references." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.133, + 0.632, + 0.15 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.159, + 0.892, + 0.234 + ], + "angle": 0, + "content": "We apply our method with Stable Diffusion with publicly available checkpoints v1.4. All experiments were conducted on a single Quadro P6000 24G GPU. The inference time synthesizing an image of \\(512 \\times 512\\) pixels takes about three minutes for our coarse-to-fine self-rectification." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.242, + 0.767, + 0.258 + ], + "angle": 0, + "content": "4.1. Evaluations and Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.265, + 0.892, + 0.401 + ], + "angle": 0, + "content": "We evaluate our method with non-stationary textures released by [51]. Each example is resized to \\(512 \\times 512\\) pixels as the reference, and we quickly built several different target images for each in PhotoShop with just a few lazy edits. Fig. 6 shows a gallery of results generated by our method. As can be seen our method faithfully reproduces the delicate textures of the exemplar, their global structure and yet respecting the sparse edits of the target image. More results are included in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.401, + 0.892, + 0.567 + ], + "angle": 0, + "content": "To compare with state-of-the-art texture synthesis methods, we fed the target images we sketched to the models trained by adversarial expansion (TexExp) [51]. As can be seen in Fig. 6, TexExp failed to reproduce the fine textures of the reference, as the edited target images are unseen data to its training. We also tested the texture optimization based on a recently proposed textural loss (GCD Loss) [49], where the target images are down-scaled as the initialization in its multi-resolution synthesis. Although plausible local textures are synthesized, the output of this method does not always conform to the user-edited layout (see Fig. 6)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.892, + 0.703 + ], + "angle": 0, + "content": "We have also applied our method on nearly homogeneous textures of stationary statistics. Since such textures do not have a prominent global structure, we can simply reshuffle patches of the reference to serve as the target. This, however, might cut some local elements in the example. Alternatively, we can use the reference image as the target and shuffle its inversion code before the sampling, which better preserves the local texture elements. Fig. 7 shows a few examples, demonstrating promising quality." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.71, + 0.664, + 0.724 + ], + "angle": 0, + "content": "4.2. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.733, + 0.89, + 0.763 + ], + "angle": 0, + "content": "In this section, we analyze the effect of the key components in our method through ablation studies." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.87 + ], + "angle": 0, + "content": "KV-Injection in sampling. As \\( KV \\)-Injection is applied both in the inversion and sampling process, we applied two ablations to study its effect. First, we set \\( P_1 = P_2 = 0 \\), and explore the full parameter space for \\( S_1 \\) and \\( S_2 \\), to study the effect of \\( KV \\)-Injection in texture sampling without being affected by the inversion." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Fig. 8 shows a matrix of results of this parameter study. A smaller value of \\( S(S_1 \\& S_2) \\) means performing more time" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7771" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.088, + 0.895, + 0.83 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.837, + 0.897, + 0.907 + ], + "angle": 0, + "content": "Figure 6. We sketch two targets for each reference, denoted as Edit 1 and Edit 2, respectively. Our self-rectification method generates textures with global structures that faithfully respect the user edits of the target images, while still producing high-quality texture details. In contrast, adversarial expansion (TexExp) [51] does not capture the fine details well. Optimization by GCD Loss [49] reproduces the local textures, but does not always conform to the target image. Note that Edit 2 is used as the input for TexExp, and as the initialization in texture optimization of GCD Loss. More results are included in the supplementary." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.517, + 0.957 + ], + "angle": 0, + "content": "7772" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.083, + 0.089, + 0.209, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.211, + 0.089, + 0.336, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.089, + 0.465, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.187, + 0.209, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.211, + 0.187, + 0.336, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.187, + 0.465, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.285, + 0.209, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.211, + 0.285, + 0.336, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.285, + 0.465, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.392, + 0.47, + 0.462 + ], + "angle": 0, + "content": "Figure 7. For nearly homogeneous textures (left), we can use patch shuffle of the reference to define a random target layout, where the shuffling could be performed in image space before applying self-rectification (middle), or in latent space after the inversion in the first self-rectification (right)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.475, + 0.47, + 0.643 + ], + "angle": 0, + "content": "steps of \\( KV \\)-Injection in sampling, and thus, the target image is rectified to be closer to the reference texture. On the contrary, a larger S reserves more of the target image's layout, however, at the same time, it introduces more structural errors or conflicts, yielding artifacts. None of these results reaches a good trade-off between synthesizing a reasonable global layout (especially considering the user edits) and reproducing local textures of the reference. Nevertheless, we can find a proper parameter setting for S, which is \\( 10\\sim 20 \\) for \\( S_1 \\), and 5 for \\( S_2 \\). Hence we set the default values of \\( S_1 \\) and \\( S_2 \\) to be 20 and 5, and use it for all experiments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.468, + 0.842 + ], + "angle": 0, + "content": "KV-Injection in inversion. Next, we investigate the effect of \\( KV \\)-Injection in our structure-preserving inversion. By fixing \\( S_1 \\) and \\( S \\) to default values, we search the parameter space defined by \\( P_1 \\) and \\( P_2 \\). As shown in Fig. 9, introducing \\( KV \\)-Injection significantly improves the synthesis result. Both the structural errors and local artifacts are drastically reduced at the same time. Another important point is that, we may have a relatively wide range for setting the value of \\( P \\), which is empirically suggested by the results: \\( 10\\sim 30 \\) for \\( P_1 \\), and \\( 5\\sim 15 \\) for \\( P_2 \\). We usually set \\( P_1 \\) as 20, and \\( P_2 \\) as 5 in production. See supplementary for the full exploration results and more examples of this study." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.86, + 0.468, + 0.89 + ], + "angle": 0, + "content": "Data augmentation. In many cases, data augmentation may not be necessary, as the pre-trained Stable Diffusion" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.087, + 0.889, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.318, + 0.892, + 0.458 + ], + "angle": 0, + "content": "Figure 8. Ablation study of \\( KV \\)-Injection in texture sampling. We explore the parameter space for \\( S_1 \\) and \\( S_2 \\), with \\( P_1 = P_2 = 0 \\), and adopt the wood texture and its edited target shown in Fig. 1 for this test. The coarse self-rectification results are shown in the 1st column. We can see that changing the starting time-step of \\( KV \\)-Injection in the sampling greatly affects the final output. However, none of these settings yields a qualified balance when considering both the target layout and the texture details. Note setting \\( S = 50 \\) (i.e., no \\( KV \\)-Injection in sampling) only results in an approximate layout reconstruction without any details synthesized (see Fig. 5)." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.481, + 0.889, + 0.705 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.714, + 0.892, + 0.812 + ], + "angle": 0, + "content": "Figure 9. Ablation study of \\( KV \\)-Injection in structure-preserving inversion by exploring the parameter space defined by \\( \\mathrm{P}_1 \\) and \\( \\mathrm{P}_2 \\). Here, we fixed \\( S_1 = 20 \\) and \\( S_2 = 5 \\) according to the test of Fig. 8. While compared with Fig. 8, the outputs are significantly improved after introducing \\( KV \\)-Injection to the inversion. Many of these results can be considered good synthesis regarding the target image edited and the source texture of the wood." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.887 + ], + "angle": 0, + "content": "model already has a certain ability to synthesize rotated texture patterns, except when the reference has dominant directional structures. To allow more free editing for directional textures, augmenting the reference is essential to synthesize" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7773" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.177, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.243, + 0.16, + 0.254 + ], + "angle": 0, + "content": "Reference" + }, + { + "type": "image", + "bbox": [ + 0.179, + 0.089, + 0.272, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.207, + 0.243, + 0.245, + 0.255 + ], + "angle": 0, + "content": "Target" + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.089, + 0.367, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.28, + 0.238, + 0.362, + 0.261 + ], + "angle": 0, + "content": "w/o Data Augmentation" + }, + { + "type": "image", + "bbox": [ + 0.369, + 0.089, + 0.463, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.375, + 0.238, + 0.457, + 0.261 + ], + "angle": 0, + "content": "w/ Data Augmentation" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.272, + 0.47, + 0.328 + ], + "angle": 0, + "content": "Figure 10. For textures that contain dominant directional structures, data augmentation is essential to allow the output to admit with user edits. Specifically, we augment each source shown above with three images rotated at angles of \\(\\pm 45^{\\circ}\\) and \\(90^{\\circ}\\), respectively." + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.355, + 0.208, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.452, + 0.208, + 0.549 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.552, + 0.179, + 0.563 + ], + "angle": 0, + "content": "Reference" + }, + { + "type": "image", + "bbox": [ + 0.212, + 0.355, + 0.335, + 0.451 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.212, + 0.453, + 0.335, + 0.549 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.226, + 0.552, + 0.322, + 0.565 + ], + "angle": 0, + "content": "Target Outlines" + }, + { + "type": "image", + "bbox": [ + 0.34, + 0.355, + 0.464, + 0.451 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.34, + 0.453, + 0.464, + 0.549 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.38, + 0.552, + 0.425, + 0.565 + ], + "angle": 0, + "content": "Output" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.575, + 0.469, + 0.616 + ], + "angle": 0, + "content": "Figure 11. Guided texture synthesis. The user provides a color layout to guide the output structure. The synthesized outputs follow the outlines well and still exhibit the reference texture details." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.631, + 0.469, + 0.676 + ], + "angle": 0, + "content": "a more consistent output structure conforming to user edits; see, e.g., Fig. 10 for a comparison on data augmentation, where additional rotated transformations are involved." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.686, + 0.296, + 0.702 + ], + "angle": 0, + "content": "4.3. Additional Applications" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.709, + 0.469, + 0.844 + ], + "angle": 0, + "content": "Our self-rectification framework can extend its utility to seemingly other applications. For instance, as illustrated in Fig. 11, users can input a straightforward target image outlining a layout structure with colors compatible with reference textures. To avoid excessive smoothing, random noise is introduced to enhance the target image. The self-rectification process ensures that the target image incorporates texture details from the reference, while adhering to the user-provided layout structure." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.845, + 0.469, + 0.891 + ], + "angle": 0, + "content": "Moreover, our self-rectification method can be employed for image editing beyond textures. Users can edit a given image by employing basic cut-and-paste patch operations," + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.089, + 0.581, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.584, + 0.089, + 0.657, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.659, + 0.089, + 0.733, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.736, + 0.089, + 0.81, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.813, + 0.089, + 0.887, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.149, + 0.581, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.583, + 0.149, + 0.657, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.659, + 0.149, + 0.733, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.736, + 0.149, + 0.81, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.813, + 0.149, + 0.887, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.208, + 0.581, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.583, + 0.208, + 0.657, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.659, + 0.208, + 0.733, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.736, + 0.208, + 0.81, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.813, + 0.208, + 0.887, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.266, + 0.581, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.583, + 0.266, + 0.657, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.659, + 0.266, + 0.733, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.736, + 0.266, + 0.81, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.813, + 0.266, + 0.887, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.326, + 0.562, + 0.338 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image_caption", + "bbox": [ + 0.603, + 0.326, + 0.638, + 0.338 + ], + "angle": 0, + "content": "Edit 1" + }, + { + "type": "image_caption", + "bbox": [ + 0.674, + 0.326, + 0.728, + 0.338 + ], + "angle": 0, + "content": "Output 1" + }, + { + "type": "image_caption", + "bbox": [ + 0.758, + 0.326, + 0.796, + 0.337 + ], + "angle": 0, + "content": "Edit 2" + }, + { + "type": "image_caption", + "bbox": [ + 0.825, + 0.326, + 0.88, + 0.337 + ], + "angle": 0, + "content": "Output 2" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.349, + 0.892, + 0.391 + ], + "angle": 0, + "content": "Figure 12. Our method can also be applied to image editing. For creating the target image, the user can either loosely sketch a layout on the target canvas, or directly edit the input image." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.405, + 0.891, + 0.451 + ], + "angle": 0, + "content": "and the resulting crude edits undergo self-rectification. Fig. 12 presents several instances of such image editing operations, presenting two edits for each input image." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.464, + 0.627, + 0.48 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.489, + 0.892, + 0.655 + ], + "angle": 0, + "content": "Our work addresses the intricate challenge of synthesizing non-stationary textures, offering a method that empowers users to efficiently design new textures with unprecedented controllability. This stands as a notable improvement over existing methods, providing a user-friendly process, which consists of two steps: Users begin with an initial rough edit using conventional image editing tools, followed by an automated self-rectification process. This process leverages a pre-trained diffusion network and injections of self-attention features, showcasing flexibility in synthesizing a diverse range of challenging non-stationary textures." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.656, + 0.892, + 0.761 + ], + "angle": 0, + "content": "In the future, we would like to explore the extension of our approach to synthesize large-scale textures. Additionally, there is a promising avenue to incorporate semantic understanding into the self-rectification process, enhancing alignment with user intent. The integration of texture semantics holds the potential to yield contextually relevant and visually appealing synthesized textures." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgements This work was supported in parts by NSFC (62161146005, U21B2023, U2001206), GD NSF (2022A1515010221), DEGP Innovation Team (2022 KCXTD025), Shenzhen Science and Technology Program (KQTD20210811090044003, RCJC20200714114435012), ISF (3441/21), Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ) and Scientific Development Funds of Shenzhen University." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7774" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.156 + ], + "angle": 0, + "content": "[1] Yuval Alaluf, Daniel Garibi, Or Patashnik, Hadar Averbuch-Elor, and Daniel Cohen-Or. Cross-image attention for zero-shot appearance transfer, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.468, + 0.199 + ], + "angle": 0, + "content": "[2] Yuval Alaluf, Elad Richardson, Gal Metzer, and Daniel Cohen-Or. A neural space-time representation for text-to-image personalization, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.2, + 0.468, + 0.242 + ], + "angle": 0, + "content": "[3] Rachele Bellini, Yanir Kleiman, and Daniel Cohen-Or. Time-varying weathering in texture space. ACM Trans. on Graphics (Proc. of SIGGRAPH), 35(4):1-11, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.242, + 0.468, + 0.283 + ], + "angle": 0, + "content": "[4] Sagie Benaim, Ron Mokady, Amit Bermano, and Lior Wolf. Structural analogy from a single image pair. Computer Graphics Forum, 40(1):249-265, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.285, + 0.468, + 0.338 + ], + "angle": 0, + "content": "[5] Urs Bergmann, Nikolay Jetchev, and Roland Vollgraf. Learning texture manifolds with the periodic spatial gan. In Proc. of IEEE Int. Conf. on Machine Learning, page 469-477, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.341, + 0.468, + 0.409 + ], + "angle": 0, + "content": "[6] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. MasaCtrl: tuning-free mutual self-attention control for consistent image synthesis and editing. In Proc. of Int. Conf. on Computer Vision, pages 22560–22570, October 2023. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.41, + 0.468, + 0.464 + ], + "angle": 0, + "content": "[7] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.466, + 0.468, + 0.507 + ], + "angle": 0, + "content": "[8] Alexei A. Efros and William T. Freeman. Image quilting for texture synthesis and transfer. In Proc. of SIGGRAPH, pages 341-346, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.509, + 0.468, + 0.549 + ], + "angle": 0, + "content": "[9] Alexei A. Efros and Thomas K. Leung. Texture synthesis by non-parametric sampling. In Proc. of Int. Conf. on Computer Vision, 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.551, + 0.468, + 0.605 + ], + "angle": 0, + "content": "[10] Anna Frühstück, Ibraeem Alhashim, and Peter Wonka. Titegan: synthesis of large-scale non-homogeneous textures. ACM Trans. on Graphics (Proc. of SIGGRAPH), 38(4):1-11, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.607, + 0.468, + 0.661 + ], + "angle": 0, + "content": "[11] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.663, + 0.468, + 0.704 + ], + "angle": 0, + "content": "[12] Leon Gatys, Alexander S Ecker, and Matthias Bethge. Texture synthesis using convolutional neural networks. In Advances in Neural Information Processing Systems, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.468, + 0.76 + ], + "angle": 0, + "content": "[13] Eric Heitz, Kenneth Vanhoey, Thomas Chambon, and Laurent Belcour. A sliced wasserstein loss for neural texture synthesis. In Proc. of IEEE Conf. on Computer Vision & Pattern Recognition, June 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.468, + 0.816 + ], + "angle": 0, + "content": "[14] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. In Proc. of International Conference on Learning Representations, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.468, + 0.859 + ], + "angle": 0, + "content": "[15] Aaron Hertzmann, Charles E. Jacobs, Nuria Oliver, Brian Curless, and David H. Salesin. Image analogies. In Proc. of SIGGRAPH, page 327-340, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[17] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.149, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[18] Nikolay Jetchev, Urs Bergmann, and Roland Vollgraf. Texture synthesis with spatial generative adversarial networks. arXiv preprint arXiv:1611.08207, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.191, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[19] Alexandre Kaspar, Boris Neubert, Dani Lischinski, Mark Pauly, and Johannes Kopf. Self tuning texture optimization. Computer Graphics Forum (Proc. of Eurographics), 34(2):349-359, may 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.247, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[20] Anant Khandelwal. Infusion: Inject and attention fusion for multi concept zero shot text based video editing. arXiv preprint arXiv:2308.00135, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.288, + 0.892, + 0.329 + ], + "angle": 0, + "content": "[21] Vivek Kwatra, Irfan Essa, Aaron Bobick, and Nipun Kwatra. Texture optimization for example-based synthesis. ACM Trans. on Graphics (Proc. of SIGGRAPH), August 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.385 + ], + "angle": 0, + "content": "[22] Vivek Kwatra, Arno Schödl, Irfan Essa, Greg Turk, and Aaron Bobick. Graphcut textures: Image and video synthesis using graph cuts. ACM Trans. on Graphics (Proc. of SIGGRAPH), 22(3):277-286, jul 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.386, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[23] Chuan Li and Michael Wand. Precomputed real-time texture synthesis with markovian generative adversarial networks. In Proc. of Euro. Conf. on Computer Vision, pages 702-716, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.511 + ], + "angle": 0, + "content": "[24] Yitzchak David Lockerman, Basile Sauvage, Rémi Allègre, Jean-Michel Dischler, Julie Dorsey, and Holly Rushmeier. Multi-scale label-map extraction for texture synthesis. ACM Trans. on Graphics (Proc. of SIGGRAPH), 35(4):140:1-140:12, July 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.553 + ], + "angle": 0, + "content": "[25] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.554, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[26] M. Lukáč, J. Fíšer, P. Asente, J. Lu, E. Shechtman, and D. Sykora. Brushables: Example-based edge-aware directional texture painting. Computer Graphics Forum (Proc. of Pacific Conf. on Computer Graphics & Applications), 34(7):257-267, oct 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.622, + 0.892, + 0.678 + ], + "angle": 0, + "content": "[27] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[28] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2):3, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.892, + 0.776 + ], + "angle": 0, + "content": "[29] Carlos Rodriguez-Pardo and Elena Garces. Seamlessgan: Self-supervised synthesis of tileable texture maps. IEEE Trans. Visualization & Computer Graphics, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.777, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proc. of IEEE Conf. on Computer Vision & Pattern Recognition, pages 10684-10695, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[31] Amir Rosenberger, Daniel Cohen-Or, and Dani Lischinski. Layered shape synthesis: Automatic generation of control maps for non-stationary textures. ACM Trans. on Graphics (Proc. of SIGGRAPH), 28(5):107:1-9, Dec. 2009. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.957 + ], + "angle": 0, + "content": "7775" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "[32] Tamar Rott Shaham, Tali Dekel, and Tomer Michaeli. Singan: Learning a generative model from a single natural image. In Proc. of Int. Conf. on Computer Vision, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.469, + 0.205 + ], + "angle": 0, + "content": "[33] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proc. of IEEE Conf. on Computer Vision & Pattern Recognition, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.469, + 0.261 + ], + "angle": 0, + "content": "[34] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdream booth: Hypernetworks for fast personalization of text-to-image models, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.469, + 0.303 + ], + "angle": 0, + "content": "[35] Omry Sendik and Daniel Cohen-Or. Deep correlations for texture synthesis. ACM Trans. on Graphics, 36(5):161:1-161:15, July 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.469, + 0.347 + ], + "angle": 0, + "content": "[36] Wu Shi and Yu Qiao. Fast texture synthesis via pseudo optimizer. In Proc. of IEEE Conf. on Computer Vision & Pattern Recognition, pages 5498-5507, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.469, + 0.389 + ], + "angle": 0, + "content": "[37] Assaf Shocher, Shai Bagon, Phillip Isola, and Michal Irani. Ingan: Capturing and retargeting the \"dna\" of a natural image. In Proc. of Int. Conf. on Computer Vision, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.391, + 0.469, + 0.432 + ], + "angle": 0, + "content": "[38] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.469, + 0.501 + ], + "angle": 0, + "content": "[39] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proc. of IEEE Conf. on Computer Vision & Pattern Recognition, pages 1921-1930, June 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.469, + 0.559 + ], + "angle": 0, + "content": "[40] Dmitry Ulyanov, Vadim Lebedev, Andrea Vedaldi, and Victor Lempitsky. Texture networks: Feed-forward synthesis of textures and stylized images. In Proc. of IEEE Int. Conf. on Machine Learning, page 1349-1357, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.562, + 0.469, + 0.615 + ], + "angle": 0, + "content": "[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.618, + 0.469, + 0.685 + ], + "angle": 0, + "content": "[42] Jiaping Wang, Xin Tong, Stephen Lin, Minghao Pan, Chao Wang, Hujun Bao, Baining Guo, and Heung-Yeung Shum. Appearance manifolds for modeling time-variant appearance of materials. ACM Trans. on Graphics (Proc. of SIGGRAPH), 25(3):754–761, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.469, + 0.729 + ], + "angle": 0, + "content": "[43] Li-Yi Wei and Marc Levoy. Fast texture synthesis using tree-structured vector quantization. In Proc. of SIGGRAPH, pages 479–488, 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.731, + 0.469, + 0.771 + ], + "angle": 0, + "content": "[44] Yonatan Wexler, Eli Shechtman, and Michal Irani. Spacetime completion of video. IEEE Trans. Pattern Analysis & Machine Intelligence, 29(3):463-476, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.47, + 0.828 + ], + "angle": 0, + "content": "[45] Li-Yi Wie, Sylvain Lefebvre, Vivek Kwatra, and Greg Turk. State of the Art in Example-based Texture Synthesis. In M. Pauly and G. Greiner, editors, Eurographics 2009 - State of the Art Reports. The Eurographics Association, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.83, + 0.469, + 0.885 + ], + "angle": 0, + "content": "[46] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. arXiv preprint arXiv:2211.13227, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.885 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.146 + ], + "angle": 0, + "content": "[47] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. In arXiv preprint arxiv:2308.06721, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[48] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[49] Yang Zhou, Kaijian Chen, Rongjun Xiao, and Hui Huang. Neural texture synthesis with guided correspondence. In Proc. of IEEE Conf. on Computer Vision & Pattern Recognition, pages 18095-18104, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[50] Yang Zhou, Huajie Shi, Dani Lischinski, Minglun Gong, Johannes Kopf, and Hui Huang. Analysis and controlled synthesis of inhomogeneous textures. Computer Graphics Forum (Proc. of Eurographics), 36(2):199-212, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[51] Yang Zhou, Zhen Zhu, Xiang Bai, Dani Lischinski, Daniel Cohen-Or, and Hui Huang. Non-stationary texture synthesis by adversarial expansion. ACM Trans. on Graphics (Proc. of SIGGRAPH), 37(4):49:1-49:13, 2018. 1, 2, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7776" + } + ] +] \ No newline at end of file diff --git a/2024/Generating Non-Stationary Textures using Self-Rectification/3ddc2997-ec2e-480d-b380-c7012481f817_origin.pdf b/2024/Generating Non-Stationary Textures using Self-Rectification/3ddc2997-ec2e-480d-b380-c7012481f817_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6cd7c451e35c21d32bf0513cf3f917a7432cbf03 --- /dev/null +++ b/2024/Generating Non-Stationary Textures using Self-Rectification/3ddc2997-ec2e-480d-b380-c7012481f817_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8ed7221694fe0835b6d22bf64e62bf7217cd773507e8a15a44a441e130651849 +size 8674026 diff --git a/2024/Generating Non-Stationary Textures using Self-Rectification/full.md b/2024/Generating Non-Stationary Textures using Self-Rectification/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d86d0ef9bcc9a38d25a812647b7e55f280e66ef3 --- /dev/null +++ b/2024/Generating Non-Stationary Textures using Self-Rectification/full.md @@ -0,0 +1,367 @@ +# Generating Non-Stationary Textures using Self-Rectification + +Yang Zhou1 Rongjun Xiao1 Dani Lischinski2 Daniel Cohen-Or1,3 Hui Huang1* +1Shenzhen University 2The Hebrew University of Jerusalem 3Tel Aviv University + +# Abstract + +This paper addresses the challenge of example-based non-stationary texture synthesis. We introduce a novel two-step approach wherein users first modify a reference texture using standard image editing tools, yielding an initial rough target for the synthesis. Subsequently, our proposed method, termed "self-rectification", automatically refines this target into a coherent, seamless texture, while faithfully preserving the distinct visual characteristics of the reference exemplar. Our method leverages a pre-trained diffusion network, and uses self-attention mechanisms, to gradually align the synthesized texture with the reference, ensuring the retention of the structures in the provided target. Through experimental validation, our approach exhibits exceptional proficiency in handling non-stationary textures, demonstrating significant advancements in texture synthesis when compared to existing state-of-the-art techniques. Code is available at https://github.com/xiaorongjun000/Self-Rectification + +# 1. Introduction + +Example-based texture synthesis aims to generate a texture that faithfully captures all the visual characteristics of a provided reference texture exemplar. The key challenge is to generate a texture that visually mimics the reference, while avoiding exact replication and without producing conspicuous, unnatural artifacts. + +Over the past decades, numerous methods have emerged for synthesizing textures from examples, and many have demonstrated impressive results, particularly for homogeneous textures that can be accurately captured by stationary models [45]. Nonetheless, a significant challenge remains when dealing with real-world textures that are inherently inhomogeneous, and non-stationary. + +Non-stationary textures exhibit distinctive attributes, such as sprawling irregular large-scale structures or variations in attributes such as color, local orientation, and local scale. Fig. 1 (left) shows two examples. Mimicking such + +![](images/f31eb4945484ffe2712462bafa4b9d2d3b33e2d6deca38cb83dc50b03ad4b29d.jpg) + +![](images/e6f9c02fdb75b4b7b60fbd29e67750440ae9d53990ed575468762fc44b0bf71d.jpg) +Figure 1. Our method takes as input a reference texture (left), and a crude target texture provided by the user (middle column), which may lack coherence and completeness. Self-rectification is used to transform the target into a visually coherent texture (right) that complies with the structure of the crude target, while exhibiting the visual characteristics of the reference texture. + +![](images/29ad14d7f1db57e93642510d70d71b1f0201738ea9f94457826b886bcb999a1a.jpg) + +![](images/1212ca72001c098789299d484ca37d8a6a2b066b53686bc5f5a3a3f0ad0aa22d.jpg) + +![](images/0f54f213629f7fedd67bce73e72d53484850dc2dd9e32e7693a9d75021168e7e.jpg) + +![](images/7ec8b08f3d3ce6cb996ca2d27bc176cda12f4b1a5cdb2c25e7b10708a04b2c26.jpg) + +complex structures and variations through example-based synthesis is a long-standing challenging task [31]. + +The emergence of neural networks has provided powerful means to deal with non-stationary textures. Zhou et al. [51] introduced a method that involves overfitting a GAN, where the encoder extracts structural guidance for the decoder. This approach provides a viable means of spatially extending non-stationary textures while preserving the visual characteristics of the reference. However, this technique requires an extremely long optimization process for a single texture exemplar, and moreover, it fails to provide any controllability or editability. + +In this paper, we introduce a two-step lazy-editing technique where the user first edits the given reference texture using conventional image editing tools, to obtain an extremely rough initial result, which may be incomplete and incoherent, as demonstrated in Fig. 1 (middle). Next, our technique automatically rectifies the edited texture into a regularized, coherent and seamless texture that follows the rough target, while retaining the local characteristics of the reference; see, e.g., Fig. 1 (right). We term this regulariza + +tion process self-rectification, since the texture is rectified based on the input texture itself. + +To self-rectify the crude target, we use a pre-trained diffusion network, and synthesize the final rectified result while utilizing the intermediate byproducts obtained by inverting both the initial target and the reference exemplar. Inspired by editing approaches that leverage self-attention [1,6,20], we extract certain self-attention features of the diffusion model during the inversion of the rough target, as well as the reference. These features are then injected into various steps in the inversion and denoising of the generated texture. Such feature injection forms "cross-attention" between the reference and target features [1, 6], and thereby, in the course of the diffusion process, the target texture is progressively refined to be increasingly locally similar to the reference, while retaining the global structure prescribed by the provided target. This self-rectification occurs in two passes, first addressing the larger scale structure and subsequently focusing on finer local details. + +To enhance the synthesis quality, we augment the reference texture with several transformed copies of it, which increases the diversity of the reference patterns, thereby improving the compatibility between the reference and the synthesized result. The augmented source features are injected into the corresponding target layers as well. Furthermore, we show that our method can also be applied to lazy editing of natural images, using the same means of synthesizing highly non-stationary textures. Experiments show that our method can deal with a large scope of nonstationary textures, with unprecedented flexibility and quality, compared to the state-of-the-art. + +# 2. Related Work + +Non-stationary texture synthesis. Early methods in example-based texture synthesis mainly focused on synthesizing textures with stationary characteristics [8, 9, 19, 21, 22, 43, 44]. To cope with non-stationary textures, researchers typically involve additional guidance for the control of distinctive attributes, e.g., using label maps to guide the layout synthesis of composite textures [15, 24, 31], representing the spatial variations of weathering textures with age/progression maps [3, 42], and describing the local orientations of directional textures with vector fields [26, 50]. Although the aforementioned methods have shown success, the guidance required from the user is tedious to provide, yet still limited. + +In the deep learning era, neural methods for texture generation have rapidly emerged, either through new texture losses for texture optimization [12, 13, 49] or via training generative networks [4, 5, 10, 18, 23, 29, 32, 35-37, 40, 51]. Among these methods, Sendik and Cohen-Or [35] attempted to preserve the non-local structures of non-stationary textures by regularizing the feature correla + +![](images/59cb17c6ec74110402a6ead0e05ea69d01522e1303202883d35bef37cad0d555.jpg) +Figure 2. Framework overview. Given a reference texture $I^R$ , we allow the user to quickly build a target image $I^{tar}$ in a lazy-editing manner. A coarse-to-fine synthesis is performed by running self-rectification twice. The coarse stage synthesizes a coarse yet complete overall structure, and the fine stage refines its output $I_{\text{coarse}}^*$ with finer and more accurate details, producing the final result $I^*$ . + +tion between different locations. Zhou et al. [51] proposed overfitting a GAN to learn the expansion from a small texture block to a larger one containing it. Their approach involves the GAN's encoder extracting the global structure of an input texture, duplicated by bottleneck residual blocks before decoding. Although these trained models can effectively extend non-stationary textures, their overfitting nature severely restricts their generalization and controllability. + +Diffusion-based image synthesis. The emergence of large-scale generative diffusion models, such as Stable Diffusion (SD) [30] and DALLE-2 [28], has revolutionized image synthesis due to their unprecedented generation power. To make the pre-trained diffusion models synthesize on image conditions, many solutions are proposed, including optimizing prompt tokens [2, 11], fine-tuning the entire model [33], or training an additional adapter [17,34,46-48]. + +Recently, researchers have investigated the role of the intermediate attention maps and features in the diffuser [7, 14,27,39], finding them crucial for layout/structure synthesis. The method we present in this paper, builds upon the mechanism of injecting keys and values from attention layers of one diffusion process into another, as a means for transferring visual features between images [1,6,20]. We use such injection to "copy" the local patterns from a source texture to a target one. In addition to the injection we also incorporate coarse-to-fine and data augmentation schemes, forming the self-rectification framework for generating nonstationary textures. + +# 3. Method + +Our method employs a two-step process for synthesizing a new texture $I^{*}$ resembling a provided non-stationary reference texture $I^{R}$ , while guided by a user-provided rough target $I^{tar}$ , as depicted in Fig. 2. Initially, the user rapidly creates $I^{tar}$ , by assembling patches from the source refer + +![](images/e869299231132e9069bda6cf75424c234d4e42aa5bed010dde050968793b7191.jpg) +Figure 3. Our self-rectification synthesizes an output texture $I^{*}$ via structure-preserving inversion from a rough target image $I^{lar}$ and fine texture sampling using the reference $I^{R}$ . Both processes require the injection of self-attention features ( $KV$ ) from the DDIM inversion of a corresponding reference. More specifically, for structure-preserving inversion, the reference is the target image itself, denoted as $I^{IR}$ . For fine texture sampling, the input exemplar $I^{R}$ is used to inject features that help to synthesize a plausible output with fine texture details. + +ence texture. Such a lazy-edit forms a basic layout for the output image, $I^{*}$ . However, this initial sketch may be incoherent or incomplete, hence necessitating rectification. Subsequently, the texture undergoes self-rectification to align with both the user's coarse layout and the reference image's detailed texture characteristics. This self-rectification process is depicted in Fig. 2 and is executed in two stages: coarse rectification, followed by a finer one. + +Considering that the rough target $I^{tar}$ comprises patches derived from the reference texture $I^R$ , we implement a process of self-rectification, detailed below. In a broad sense, we utilize a pre-trained latent diffusion model to invert both the target image and the reference image into an initial latent noise. This is followed by employing feature injection during the sampling process of $I^*$ . The features from $I^{tar}$ guide the structural synthesis and the features from $I^R$ guide the fine texture details. + +Below, we first briefly review diffusion models and their self-attention and the cross-KV-injection mechanisms, and then proceed to present our self-rectification technique. + +# 3.1. DDIM Sampling and Inversion + +Denoising diffusion models [16, 38] involve two processes: a noising process that gradually transforms an input image into Gaussian noise, and a denoising/sampling process that generates images from Gaussian samples. Using DDIM sampling [38], starting from a noise sample $z_{\mathrm{T}}$ , one can generate a clean sample $z_{0}$ via T deterministic steps: + +$$ +z _ {t - 1} = \sqrt {\bar {\alpha} _ {t - 1}} f _ {\theta} (z _ {t}, t) + \sqrt {1 - \bar {\alpha} _ {t - 1}} \epsilon_ {\theta} (z _ {t}, t), \quad (1) +$$ + +where $\epsilon_{\theta}$ is a noise prediction network conditioned on the current noisy sample $z_{t}$ and timestamp $t$ . $\bar{\alpha}_{t}$ is the noise scaling factor defined in [38], and $f_{\theta}(z_t,t)$ is + +$$ +f _ {\theta} \left(z _ {t}, t\right) = \frac {z _ {t} - \sqrt {1 - \bar {\alpha} _ {t}} \epsilon_ {\theta} \left(z _ {t} , t\right)}{\sqrt {\bar {\alpha} _ {t}}}. \tag {2} +$$ + +In the opposite direction, $DDIM$ inversion [38] of a given image $z_0$ is the process of incrementally adding deterministic noise, until obtaining $z_{\mathrm{T}}$ : + +$$ +z _ {t + 1} = \sqrt {\bar {\alpha} _ {t + 1}} f _ {\theta} \left(z _ {t}, t\right) + \sqrt {1 - \bar {\alpha} _ {t + 1}} \epsilon_ {\theta} \left(z _ {t}, t\right). \tag {3} +$$ + +Such inversion leads to a nearly faithful reconstruction [38]. + +# 3.2. Stable Diffusion and Self-Attention. + +We base our texture synthesis framework on Stable Diffusion (SD) [30], which is a pretrained latent diffusion model consisting of an encoder $\mathcal{E}$ that maps an input image into the latent space, and a decoder $\mathcal{D}$ that reconstructs a latent code back into image space. Both DDIM sampling and DDIM inversion are performed in the SD latent space. + +The noise predictor $\epsilon_{\theta}$ of SD is a large-scale U-Net that contains multiple self-attention modules [41]. Each self-attention layer transforms its input intermediate feature map (also called spatial features) into an attended representation by the following equation: + +$$ +\operatorname {A t t} (Q, K, V) = \operatorname {S o f t m a x} \left(\frac {Q K ^ {T}}{\sqrt {d}}\right) V, \tag {4} +$$ + +where $Q, K$ , and $V$ are the queries, keys, and values, respectively, obtained by learned linear projections of the same input spatial features, having dimension $d$ . The self-attention mechanism uses the similarities between the queries and keys as attention scores to weigh the importance or relevance of the values. Relevant info is thus aggregated as the attended representation. + +# 3.3. KV-Injection + +The self-attention features contain rich information about both the large-scale structures and local fine details of an input image [1, 6, 39]. Tumanyan et al. [39] demonstrate that by injecting the spatial features and the queries + +![](images/01fc19e8ac5e749930664f6c79335d9cf8055d022cdb8134c0d0ccaadff0b0da.jpg) +Figure 4. Visualization of the intermediate latent codes in the inversion. For the standard DDIM inversion (top), the U-Net predicts noise to diffuse the distinctive structures so as to transform the input into Gaussian noise. In contrast, our structure-preserving inversion (bottom) reserves the distinctive patterns from user edits along the inversion process. See texts in Sec. 3.4 for more details. + +and keys $(QK)$ from the self-attention layers of a source (guidance) image into the corresponding layer of the generated target image during the sampling process, one can preserve the layout of the source while modifying its appearance. Injecting only the $KV$ features of a source reference, instead, transfers its appearance, including textures, to the generated target [1,6]. + +In our case, the target image is a rough and incomplete guidance, typically containing only a few source patches rotated and placed freely by the user. A complete overall structure needs to be synthesized reasonably yet conforming to user's constraints. Locally, however, the resulting texture should resemble the reference. To achieve these goals, we adapt the $KV$ -injection into both the DDIM inversion and sampling, bringing a novel self-rectification operation. + +# 3.4. Self-rectification + +As shown in Fig. 3, our self-rectification consists of two parts: structure-preserving inversion and fine texture sampling. Both are performed in the SD latent space. The last latent code of inversion is used as the starting code of the sampling. The core modification to both processes is the $KV$ -injection of self-attention features. + +Structure-preserving inversion. The standard DDIM inversion (Eq. (3)) progressively transforms an input image (SD latent code) into pure Gaussian noise. At each time step, the noise to be added is predicted by the U-Net. As visualized in Fig. 4, for a given target image, the noise predicted by the U-Net for early time steps ( $e.g., t \leq 20$ ) is mainly distributed to "diffuse" the prominent structures so that the distinctive patterns from user edits get scattered and random. As the inversion progresses, the magnitude of the noise added in each step becomes smaller and similar for + +![](images/03503cdfdd39e5c6a714b99e7c1af312a21a75b8fe9d36674c93e7cfb801ca98.jpg) +Figure 5. Visualization of the intermediate latent codes in the fine texture sampling. Here, for the first 20 steps (from $t = 50$ to 30), we perform the standard DDIM sampling to reconstruct the target layout. Next, we perform $KV$ -injection in the remaining sampling steps ( $t = 30$ to 0), to synthesize fine textures for the output image. The rightmost shows the result produced by simply performing standard DDIM sampling for all steps, i.e., $S = 50$ . No additional structure is synthesized to complete the user edits. + +all latent pixels, since the "diffusion" is getting close to done. As no text prompt is involved, the noise prediction in each inversion step is dominantly determined by the self-attention mechanism in the U-Net. Our key observation is that if we inject the $KV$ features from a large time step $t_1$ ( $\gg \mathrm{T} / 2$ ) into an early time step $t_2$ ( $\ll \mathrm{T} / 2$ ), the noise predicted at $t_2$ will be smaller and more spatially uniform. The distinctive patterns of $I^{tar}$ , reflecting the user's edits, are thus better preserved. Hence, we refer to this process as structure-preserving inversion. + +Specifically, given a rough target image $I^{tar}$ , we invert it twice. The first inversion is a standard DDIM inversion. The produced self-attention features along the noisig steps are regarded as the inversion reference (IR) for the second inversion. The self-attention during the second inversion, at time step $t$ , is now given by: + +$$ +\operatorname {A t t} \left(Q _ {t} ^ {\text {t a r}}, K _ {\mathrm {T} - t} ^ {\mathrm {I R}}, V _ {\mathrm {T} - t} ^ {\mathrm {I R}}\right) = \operatorname {S o f t m a x} \left(\frac {Q _ {t} ^ {\text {t a r}} \left(K _ {\mathrm {T} - t} ^ {\mathrm {I R}}\right) ^ {T}}{\sqrt {d}}\right) V _ {\mathrm {T} - t} ^ {\mathrm {I R}}, \tag {5} +$$ + +where $\mathrm{T}$ denotes the total number of steps (we use $\mathrm{T} = 50$ ). Here $KV$ features are injected in a reverse order in the second round. We would like to stress that we have also experimented with using an offset, i.e., replacing $\mathrm{T} - t$ with $t + offset$ in the above equation, but found no advantage (see supplementary for more analysis and comparisons). + +During the second inversion, we do not perform $KV$ injections for all time steps. We set a parameter $P$ that defines the time step beyond which we continue with standard inversion. Fig. 4 shows the effect of our structure-preserving inversion, where $P = 30$ . Since the final latent code after the second inversion still contains rich information about the distinctive structures from $I^{tar}$ , the subsequent sampling that starts from this latent code is guided to synthesize a global structure that complies with $I^{tar}$ . + +Fine texture sampling. We start the DDIM sampling from the final structure-preserving latent code. The sampling uses the U-Net to predict the noise to be removed, as defined in Eq. (1). However, simply performing standard + +DDIM sampling for all denoising steps results in a nearly reconstructed target image since DDIM sampling is deterministic. Therefore, we set the first S steps (i.e., from time step T to T - S) to reconstruct the target layout to a certain extent. For the remaining T - S denoising steps, we synthesize the output image $I^{*}$ by matching fine textures from the reference via $KV$ -injection from the reference $I^{R}$ . + +To this end, we first DDIM-invert the reference texture $I^{R}$ . At denoising time step $t$ ( $t > \mathrm{T} - \mathrm{S}$ ), the $KV$ features extracted during the inversion are injected into the corresponding self-attention layers of the synthesized texture: + +$$ +\operatorname {A t t} \left(Q _ {t} ^ {*}, K _ {t} ^ {R}, V _ {t} ^ {R}\right) = \operatorname {S o f t m a x} \left(\frac {Q _ {t} ^ {*} \left(K _ {t} ^ {R}\right) ^ {T}}{\sqrt {d}}\right) V _ {t} ^ {R}. \tag {6} +$$ + +This forms a cross-image attention, where corresponding fine local patterns in the reference are transferred to the output image in a plausible manner. Fig. 5 visualizes the intermediate process of our texture sampling, where $S = 20$ . + +# 3.5. Implementation details and data augmentation + +In the user editing phase, we fill the background of the target canvas with pixels randomly drawn from the source texture, such that the encoding of the target image would not deviate too far from the source in the SD latent space. As the self-rectification is performed twice, the output of coarse stage $I_{\text{coarse}}^*$ , will be used as the input target image of fine stage to produce the final output $I^*$ . In contrast, for the inversion reference $I^{IR}$ involved in structure-preserving inversion of the two self-rectifications, we use the same initial target image that contains the original user edits. See the supplementary for the full algorithm pseudo-code. + +Considering the parameter settings: since the coarse stage aims for structure synthesis, relatively large $\mathbf{P}$ and S are required in self-rectification, and vice versa in the fine stage. More specifically, let $\mathrm{P_1}$ , $\mathrm{P_2}$ , $\mathrm{S_1}$ , and $\mathrm{S_2}$ , denote the parameters used in the two rounds of structure-preserving inversion and fine texture sampling. We typically set $\mathrm{P_1} = 20$ , $\mathrm{P_2} = 5$ , $\mathrm{S_1} = 20$ , and $\mathrm{S_2} = 5$ . Following [6], we choose the $KV$ features from the 10th to 15th self-attention layers of the U-Net decoder part. + +To further improve the synthesis quality, especially when dealing with textures containing a dominant directional structure, such as the leaf shown in Fig. 1, we can introduce a few transformed images (flips and rotations) to augment the reference texture. We concatenate the new attention features from the augmentation to the original reference feature. For example, when we have $n$ augmented source textures, the $K^R$ and $V^R$ in Eq. (6) is now given by + +$$ +\left\{ \begin{array}{l l} K ^ {R} & = \operatorname {C o n c a t} \left(K _ {(0)} ^ {R}, K _ {(1)} ^ {R}, \dots , K _ {(n)} ^ {R}\right), \\ V ^ {R} & = \operatorname {C o n c a t} \left(V _ {(0)} ^ {R}, V _ {(1)} ^ {R}, \dots , V _ {(n)} ^ {R}\right) \end{array} \right. \tag {7} +$$ + +where $K_{(i)}^{R}$ , and $V_{(i)}^{R}$ are the self-attention features acquired from the DDIM inversion of the augmented references. + +# 4. Experiments + +We apply our method with Stable Diffusion with publicly available checkpoints v1.4. All experiments were conducted on a single Quadro P6000 24G GPU. The inference time synthesizing an image of $512 \times 512$ pixels takes about three minutes for our coarse-to-fine self-rectification. + +# 4.1. Evaluations and Comparisons + +We evaluate our method with non-stationary textures released by [51]. Each example is resized to $512 \times 512$ pixels as the reference, and we quickly built several different target images for each in PhotoShop with just a few lazy edits. Fig. 6 shows a gallery of results generated by our method. As can be seen our method faithfully reproduces the delicate textures of the exemplar, their global structure and yet respecting the sparse edits of the target image. More results are included in the supplementary. + +To compare with state-of-the-art texture synthesis methods, we fed the target images we sketched to the models trained by adversarial expansion (TexExp) [51]. As can be seen in Fig. 6, TexExp failed to reproduce the fine textures of the reference, as the edited target images are unseen data to its training. We also tested the texture optimization based on a recently proposed textural loss (GCD Loss) [49], where the target images are down-scaled as the initialization in its multi-resolution synthesis. Although plausible local textures are synthesized, the output of this method does not always conform to the user-edited layout (see Fig. 6). + +We have also applied our method on nearly homogeneous textures of stationary statistics. Since such textures do not have a prominent global structure, we can simply reshuffle patches of the reference to serve as the target. This, however, might cut some local elements in the example. Alternatively, we can use the reference image as the target and shuffle its inversion code before the sampling, which better preserves the local texture elements. Fig. 7 shows a few examples, demonstrating promising quality. + +# 4.2. Ablation Studies + +In this section, we analyze the effect of the key components in our method through ablation studies. + +KV-Injection in sampling. As $KV$ -Injection is applied both in the inversion and sampling process, we applied two ablations to study its effect. First, we set $P_1 = P_2 = 0$ , and explore the full parameter space for $S_1$ and $S_2$ , to study the effect of $KV$ -Injection in texture sampling without being affected by the inversion. + +Fig. 8 shows a matrix of results of this parameter study. A smaller value of $S(S_1 \& S_2)$ means performing more time + +![](images/50443225708289e321da932961981303861c63ecd1f051c6a43b3fb8eb60797c.jpg) +Figure 6. We sketch two targets for each reference, denoted as Edit 1 and Edit 2, respectively. Our self-rectification method generates textures with global structures that faithfully respect the user edits of the target images, while still producing high-quality texture details. In contrast, adversarial expansion (TexExp) [51] does not capture the fine details well. Optimization by GCD Loss [49] reproduces the local textures, but does not always conform to the target image. Note that Edit 2 is used as the input for TexExp, and as the initialization in texture optimization of GCD Loss. More results are included in the supplementary. + +![](images/afa97938b590bf80a91bd6da85eb186cbf05557a2080781c66cc2e17c52e75d7.jpg) + +![](images/0a9ba524973edd6fb6309680a5747ea336d91b95af73f2f36411686b4554a6bf.jpg) + +![](images/6262c145e78d74b2d1193ab93ee846d9a1bcaad0eb99fa55b938eb066eae3825.jpg) + +![](images/02ec0269d9b4bc0ee85f020804a3266a5ac00a135ae7355947be5a8d772d729b.jpg) + +![](images/7125e9d1284caf07b46a5a3842a719d263cc59995ec3dcddc22c48a5a7811bc8.jpg) + +![](images/aeca6e4d5c72461cc0b796a32f4ad2aeb6f264c51ccdf1b543c33b99c136b7b6.jpg) + +![](images/bb90f141a104f75f975b91e18f326a429a4612cd44b9ebed7f4734109429fe5f.jpg) +Figure 7. For nearly homogeneous textures (left), we can use patch shuffle of the reference to define a random target layout, where the shuffling could be performed in image space before applying self-rectification (middle), or in latent space after the inversion in the first self-rectification (right). + +![](images/d353d345b76b288015cfcbec2a1caaa096881502f9f0d10d51185a8352152b19.jpg) + +![](images/1ef9999145be44a9ed679f675ddd306e26a270aa413c91cc046f0fd99d592e0a.jpg) + +steps of $KV$ -Injection in sampling, and thus, the target image is rectified to be closer to the reference texture. On the contrary, a larger S reserves more of the target image's layout, however, at the same time, it introduces more structural errors or conflicts, yielding artifacts. None of these results reaches a good trade-off between synthesizing a reasonable global layout (especially considering the user edits) and reproducing local textures of the reference. Nevertheless, we can find a proper parameter setting for S, which is $10\sim 20$ for $S_1$ , and 5 for $S_2$ . Hence we set the default values of $S_1$ and $S_2$ to be 20 and 5, and use it for all experiments. + +KV-Injection in inversion. Next, we investigate the effect of $KV$ -Injection in our structure-preserving inversion. By fixing $S_1$ and $S$ to default values, we search the parameter space defined by $P_1$ and $P_2$ . As shown in Fig. 9, introducing $KV$ -Injection significantly improves the synthesis result. Both the structural errors and local artifacts are drastically reduced at the same time. Another important point is that, we may have a relatively wide range for setting the value of $P$ , which is empirically suggested by the results: $10\sim 30$ for $P_1$ , and $5\sim 15$ for $P_2$ . We usually set $P_1$ as 20, and $P_2$ as 5 in production. See supplementary for the full exploration results and more examples of this study. + +Data augmentation. In many cases, data augmentation may not be necessary, as the pre-trained Stable Diffusion + +![](images/4d8a2ebcb997becd41a36c72c0605fc46b8ea38eb747f06aa66b50c8003da72c.jpg) +Figure 8. Ablation study of $KV$ -Injection in texture sampling. We explore the parameter space for $S_1$ and $S_2$ , with $P_1 = P_2 = 0$ , and adopt the wood texture and its edited target shown in Fig. 1 for this test. The coarse self-rectification results are shown in the 1st column. We can see that changing the starting time-step of $KV$ -Injection in the sampling greatly affects the final output. However, none of these settings yields a qualified balance when considering both the target layout and the texture details. Note setting $S = 50$ (i.e., no $KV$ -Injection in sampling) only results in an approximate layout reconstruction without any details synthesized (see Fig. 5). + +![](images/352a9547a500faa95228f3b92b821c363d3421f0e9b368af4968f5bbef670f61.jpg) +Figure 9. Ablation study of $KV$ -Injection in structure-preserving inversion by exploring the parameter space defined by $\mathrm{P}_1$ and $\mathrm{P}_2$ . Here, we fixed $S_1 = 20$ and $S_2 = 5$ according to the test of Fig. 8. While compared with Fig. 8, the outputs are significantly improved after introducing $KV$ -Injection to the inversion. Many of these results can be considered good synthesis regarding the target image edited and the source texture of the wood. + +model already has a certain ability to synthesize rotated texture patterns, except when the reference has dominant directional structures. To allow more free editing for directional textures, augmenting the reference is essential to synthesize + +![](images/a8afb72b0f76985592e857289bc6bff2d390b8120495bfd1888560005d2ae3d0.jpg) +Reference + +![](images/b056391f41783179be12d7aaa5e5997cdd8d84f9547dba93611e99972aa3fe2c.jpg) +Target + +![](images/c3e1520d8318dfa1d4da1127df86f99f02c5a9cced4d0ce74a97da3cc89bf549.jpg) +w/o Data Augmentation + +![](images/2ea5f75c9435532f64f7830caa19a3197ca0ebcdcabff56c74cad9f4bce03490.jpg) +w/ Data Augmentation + +![](images/534b06e54f438892c9a367c5689f2083283f9dc4d17b14457ede7b9d1004e4c2.jpg) +Figure 10. For textures that contain dominant directional structures, data augmentation is essential to allow the output to admit with user edits. Specifically, we augment each source shown above with three images rotated at angles of $\pm 45^{\circ}$ and $90^{\circ}$ , respectively. + +![](images/ab498a1a38151e4a3f98d3e6cb218d4f991081604ca797bf058204f5942fbb60.jpg) +Reference +Figure 11. Guided texture synthesis. The user provides a color layout to guide the output structure. The synthesized outputs follow the outlines well and still exhibit the reference texture details. + +![](images/cef84d6cc8c21d4b4e7f5d3e2c4bccdd9c409b51b787346a752f7f01e1581848.jpg) + +![](images/bbb1d67a42a9f0e99f31263fa7e7f86afe82fb3e4c5cb23991c303673ba44a92.jpg) +Target Outlines + +![](images/4324871efdf50b5f78be6a9811c3b74f3d7adf1af75b4db27673b10bd48d3710.jpg) +Figure 12. Our method can also be applied to image editing. For creating the target image, the user can either loosely sketch a layout on the target canvas, or directly edit the input image. + +![](images/06471c22acbe139dc0dc719dc5b4b4e45ede3620947529d5f335f7300c602d1d.jpg) +Output + +a more consistent output structure conforming to user edits; see, e.g., Fig. 10 for a comparison on data augmentation, where additional rotated transformations are involved. + +# 4.3. Additional Applications + +Our self-rectification framework can extend its utility to seemingly other applications. For instance, as illustrated in Fig. 11, users can input a straightforward target image outlining a layout structure with colors compatible with reference textures. To avoid excessive smoothing, random noise is introduced to enhance the target image. The self-rectification process ensures that the target image incorporates texture details from the reference, while adhering to the user-provided layout structure. + +Moreover, our self-rectification method can be employed for image editing beyond textures. Users can edit a given image by employing basic cut-and-paste patch operations, + +![](images/9a51b9643c751b762b060d764ba814c86273b026c1b295514fde0c844cb5563b.jpg) + +![](images/2afe6df694be01756c8240f9179ef1299e06fc9a54447116826c51131ba8ed4d.jpg) + +![](images/bb60ddbaabe7c74cc98ee1aef82a5e5c78915c69a57f8e45a5d03df27cfb67ca.jpg) + +![](images/e238aa7d944ed2b4b557745d2a3afc45980c961ba1e64c2ee0664362d2aa95da.jpg) + +![](images/358d3490f70a74e6db88a679d7ae26e861438c5d9a7f5b69114a375982530850.jpg) + +![](images/c8db08770fe2c85e968098a3a470d6eb5d60cecfa91569b505fc977e009f47a8.jpg) + +![](images/1a2bf52c8fe4e57ed4c58a4e6f202e86b891917e6d28a46581ea78904ac0e228.jpg) + +![](images/bce3be17612a92ddcb44b9eb0a44057bec2c173e9aa516d57bf3469a1a74eebb.jpg) + +![](images/02fdb624e503a3f6239afcb3869c7113f5922049123406920f0646bb4a821e58.jpg) + +![](images/f8e5454382b1077ea07caa3c2c89ed6a65473f5abd983a49c60dcd2efe309a94.jpg) + +![](images/d9c25b0cb2bd4c3eaf2895928ea8b6226aaf0fa49bb391ae06d63b1d7dfa6880.jpg) + +![](images/a8fbc5542a9f94fc36d1c83a53229abd1ae4ca0414701b57e45bb95fe341e740.jpg) + +![](images/7d4c643c70964ad9fb443f797913791934fe65944859c739a27d3e5dc935cba4.jpg) + +![](images/41a1d0e9e1e4219f43ae2316918a889385ff34160e32da01b88fba46be5b031a.jpg) + +![](images/f9b59d1177084526d5b05a5dc55d39b0c9a5130f40d2fa80d6f6428d4232e83a.jpg) + +![](images/e4ef03c3ec70425ee66ca0d667d057b4fad343a038f1abb8090a2fc4619325c7.jpg) +Input + +![](images/3f45de8b56a1e0741dc611067adf2cc1110f9df159230430980544df39218e37.jpg) +Edit 1 + +![](images/de5366bb400aab8215eeb4df37e15d0c6ea45df78f6bde3f504aead4f5eefedd.jpg) +Output 1 + +![](images/bdf22ee1553340991d3b29f835b78f6471b1bf0fbff8e69618fb910e3df95a9e.jpg) +Edit 2 + +![](images/ffc370ac281e528725a1f8b668f16a5d44e25a067add851875346c00a5590670.jpg) +Output 2 + +and the resulting crude edits undergo self-rectification. Fig. 12 presents several instances of such image editing operations, presenting two edits for each input image. + +# 5. Conclusions + +Our work addresses the intricate challenge of synthesizing non-stationary textures, offering a method that empowers users to efficiently design new textures with unprecedented controllability. This stands as a notable improvement over existing methods, providing a user-friendly process, which consists of two steps: Users begin with an initial rough edit using conventional image editing tools, followed by an automated self-rectification process. This process leverages a pre-trained diffusion network and injections of self-attention features, showcasing flexibility in synthesizing a diverse range of challenging non-stationary textures. + +In the future, we would like to explore the extension of our approach to synthesize large-scale textures. Additionally, there is a promising avenue to incorporate semantic understanding into the self-rectification process, enhancing alignment with user intent. The integration of texture semantics holds the potential to yield contextually relevant and visually appealing synthesized textures. + +Acknowledgements This work was supported in parts by NSFC (62161146005, U21B2023, U2001206), GD NSF (2022A1515010221), DEGP Innovation Team (2022 KCXTD025), Shenzhen Science and Technology Program (KQTD20210811090044003, RCJC20200714114435012), ISF (3441/21), Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ) and Scientific Development Funds of Shenzhen University. + +# References + +[1] Yuval Alaluf, Daniel Garibi, Or Patashnik, Hadar Averbuch-Elor, and Daniel Cohen-Or. Cross-image attention for zero-shot appearance transfer, 2023. 2, 3, 4 +[2] Yuval Alaluf, Elad Richardson, Gal Metzer, and Daniel Cohen-Or. A neural space-time representation for text-to-image personalization, 2023. 2 +[3] Rachele Bellini, Yanir Kleiman, and Daniel Cohen-Or. Time-varying weathering in texture space. ACM Trans. on Graphics (Proc. of SIGGRAPH), 35(4):1-11, 2016. 2 +[4] Sagie Benaim, Ron Mokady, Amit Bermano, and Lior Wolf. Structural analogy from a single image pair. Computer Graphics Forum, 40(1):249-265, 2021. 2 +[5] Urs Bergmann, Nikolay Jetchev, and Roland Vollgraf. Learning texture manifolds with the periodic spatial gan. In Proc. of IEEE Int. Conf. on Machine Learning, page 469-477, 2017. 2 +[6] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. MasaCtrl: tuning-free mutual self-attention control for consistent image synthesis and editing. In Proc. of Int. Conf. on Computer Vision, pages 22560–22570, October 2023. 2, 3, 4, 5 +[7] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models, 2023. 2 +[8] Alexei A. Efros and William T. Freeman. Image quilting for texture synthesis and transfer. In Proc. of SIGGRAPH, pages 341-346, 2001. 2 +[9] Alexei A. Efros and Thomas K. Leung. Texture synthesis by non-parametric sampling. In Proc. of Int. Conf. on Computer Vision, 1999. 2 +[10] Anna Frühstück, Ibraeem Alhashim, and Peter Wonka. Titegan: synthesis of large-scale non-homogeneous textures. ACM Trans. on Graphics (Proc. of SIGGRAPH), 38(4):1-11, 2019. 2 +[11] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion, 2022. 2 +[12] Leon Gatys, Alexander S Ecker, and Matthias Bethge. Texture synthesis using convolutional neural networks. In Advances in Neural Information Processing Systems, 2015. 2 +[13] Eric Heitz, Kenneth Vanhoey, Thomas Chambon, and Laurent Belcour. A sliced wasserstein loss for neural texture synthesis. In Proc. of IEEE Conf. on Computer Vision & Pattern Recognition, June 2021. 2 +[14] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. In Proc. of International Conference on Learning Representations, 2023. 2 +[15] Aaron Hertzmann, Charles E. Jacobs, Nuria Oliver, Brian Curless, and David H. Salesin. Image analogies. In Proc. of SIGGRAPH, page 327-340, 2001. 2 +[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 3 + +[17] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models, 2021. 2 +[18] Nikolay Jetchev, Urs Bergmann, and Roland Vollgraf. Texture synthesis with spatial generative adversarial networks. arXiv preprint arXiv:1611.08207, 2016. 2 +[19] Alexandre Kaspar, Boris Neubert, Dani Lischinski, Mark Pauly, and Johannes Kopf. Self tuning texture optimization. Computer Graphics Forum (Proc. of Eurographics), 34(2):349-359, may 2015. 2 +[20] Anant Khandelwal. Infusion: Inject and attention fusion for multi concept zero shot text based video editing. arXiv preprint arXiv:2308.00135, 2023. 2 +[21] Vivek Kwatra, Irfan Essa, Aaron Bobick, and Nipun Kwatra. Texture optimization for example-based synthesis. ACM Trans. on Graphics (Proc. of SIGGRAPH), August 2005. 2 +[22] Vivek Kwatra, Arno Schödl, Irfan Essa, Greg Turk, and Aaron Bobick. Graphcut textures: Image and video synthesis using graph cuts. ACM Trans. on Graphics (Proc. of SIGGRAPH), 22(3):277-286, jul 2003. 2 +[23] Chuan Li and Michael Wand. Precomputed real-time texture synthesis with markovian generative adversarial networks. In Proc. of Euro. Conf. on Computer Vision, pages 702-716, 2016. 2 +[24] Yitzchak David Lockerman, Basile Sauvage, Rémi Allègre, Jean-Michel Dischler, Julie Dorsey, and Holly Rushmeier. Multi-scale label-map extraction for texture synthesis. ACM Trans. on Graphics (Proc. of SIGGRAPH), 35(4):140:1-140:12, July 2016. 2 +[25] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models, 2022. +[26] M. Lukáč, J. Fíšer, P. Asente, J. Lu, E. Shechtman, and D. Sykora. Brushables: Example-based edge-aware directional texture painting. Computer Graphics Forum (Proc. of Pacific Conf. on Computer Graphics & Applications), 34(7):257-267, oct 2015. 2 +[27] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2 +[28] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2):3, 2022. 2 +[29] Carlos Rodriguez-Pardo and Elena Garces. Seamlessgan: Self-supervised synthesis of tileable texture maps. IEEE Trans. Visualization & Computer Graphics, 2022. 2 +[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proc. of IEEE Conf. on Computer Vision & Pattern Recognition, pages 10684-10695, 2022. 2, 3 +[31] Amir Rosenberger, Daniel Cohen-Or, and Dani Lischinski. Layered shape synthesis: Automatic generation of control maps for non-stationary textures. ACM Trans. on Graphics (Proc. of SIGGRAPH), 28(5):107:1-9, Dec. 2009. 1, 2 + +[32] Tamar Rott Shaham, Tali Dekel, and Tomer Michaeli. Singan: Learning a generative model from a single natural image. In Proc. of Int. Conf. on Computer Vision, 2019. 2 +[33] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proc. of IEEE Conf. on Computer Vision & Pattern Recognition, 2023. 2 +[34] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdream booth: Hypernetworks for fast personalization of text-to-image models, 2023. 2 +[35] Omry Sendik and Daniel Cohen-Or. Deep correlations for texture synthesis. ACM Trans. on Graphics, 36(5):161:1-161:15, July 2017. 2 +[36] Wu Shi and Yu Qiao. Fast texture synthesis via pseudo optimizer. In Proc. of IEEE Conf. on Computer Vision & Pattern Recognition, pages 5498-5507, 2020. 2 +[37] Assaf Shocher, Shai Bagon, Phillip Isola, and Michal Irani. Ingan: Capturing and retargeting the "dna" of a natural image. In Proc. of Int. Conf. on Computer Vision, 2019. 2 +[38] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3 +[39] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proc. of IEEE Conf. on Computer Vision & Pattern Recognition, pages 1921-1930, June 2023. 2, 3 +[40] Dmitry Ulyanov, Vadim Lebedev, Andrea Vedaldi, and Victor Lempitsky. Texture networks: Feed-forward synthesis of textures and stylized images. In Proc. of IEEE Int. Conf. on Machine Learning, page 1349-1357, 2016. 2 +[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 30, 2017. 3 +[42] Jiaping Wang, Xin Tong, Stephen Lin, Minghao Pan, Chao Wang, Hujun Bao, Baining Guo, and Heung-Yeung Shum. Appearance manifolds for modeling time-variant appearance of materials. ACM Trans. on Graphics (Proc. of SIGGRAPH), 25(3):754–761, 2006. 2 +[43] Li-Yi Wei and Marc Levoy. Fast texture synthesis using tree-structured vector quantization. In Proc. of SIGGRAPH, pages 479–488, 2000. 2 +[44] Yonatan Wexler, Eli Shechtman, and Michal Irani. Spacetime completion of video. IEEE Trans. Pattern Analysis & Machine Intelligence, 29(3):463-476, 2007. 2 +[45] Li-Yi Wie, Sylvain Lefebvre, Vivek Kwatra, and Greg Turk. State of the Art in Example-based Texture Synthesis. In M. Pauly and G. Greiner, editors, Eurographics 2009 - State of the Art Reports. The Eurographics Association, 2009. 1 +[46] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. arXiv preprint arXiv:2211.13227, 2022. 2 + +[47] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. In arXiv preprint arxiv:2308.06721, 2023. 2 +[48] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 2 +[49] Yang Zhou, Kaijian Chen, Rongjun Xiao, and Hui Huang. Neural texture synthesis with guided correspondence. In Proc. of IEEE Conf. on Computer Vision & Pattern Recognition, pages 18095-18104, 2023. 2, 5, 6 +[50] Yang Zhou, Huajie Shi, Dani Lischinski, Minglun Gong, Johannes Kopf, and Hui Huang. Analysis and controlled synthesis of inhomogeneous textures. Computer Graphics Forum (Proc. of Eurographics), 36(2):199-212, 2017. 2 +[51] Yang Zhou, Zhen Zhu, Xiang Bai, Dani Lischinski, Daniel Cohen-Or, and Hui Huang. Non-stationary texture synthesis by adversarial expansion. ACM Trans. on Graphics (Proc. of SIGGRAPH), 37(4):49:1-49:13, 2018. 1, 2, 5, 6 \ No newline at end of file diff --git a/2024/Generating Non-Stationary Textures using Self-Rectification/images.zip b/2024/Generating Non-Stationary Textures using Self-Rectification/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..44a5263cdb812606f75aad16d9a1e03f9b1d38b1 --- /dev/null +++ b/2024/Generating Non-Stationary Textures using Self-Rectification/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:973f5095256724085ad316dd226a35f13d0cf06fd093f1c9ab6fe5b0d8dacb5f +size 1490212 diff --git a/2024/Generating Non-Stationary Textures using Self-Rectification/layout.json b/2024/Generating Non-Stationary Textures using Self-Rectification/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..75d97512ea8430010c1e5aa581b7a345d04f7d9c --- /dev/null +++ b/2024/Generating Non-Stationary Textures using Self-Rectification/layout.json @@ -0,0 +1,10868 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 110, + 103, + 483, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 103, + 483, + 121 + ], + "spans": [ + { + "bbox": [ + 110, + 103, + 483, + 121 + ], + "type": "text", + "content": "Generating Non-Stationary Textures using Self-Rectification" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 142, + 536, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 142, + 536, + 174 + ], + "spans": [ + { + "bbox": [ + 61, + 142, + 536, + 174 + ], + "type": "text", + "content": "Yang Zhou1 Rongjun Xiao1 Dani Lischinski2 Daniel Cohen-Or1,3 Hui Huang1* \n1Shenzhen University 2The Hebrew University of Jerusalem 3Tel Aviv University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 199, + 192, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 199, + 192, + 211 + ], + "spans": [ + { + "bbox": [ + 143, + 199, + 192, + 211 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 224, + 290, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 224, + 290, + 439 + ], + "spans": [ + { + "bbox": [ + 47, + 224, + 290, + 439 + ], + "type": "text", + "content": "This paper addresses the challenge of example-based non-stationary texture synthesis. We introduce a novel two-step approach wherein users first modify a reference texture using standard image editing tools, yielding an initial rough target for the synthesis. Subsequently, our proposed method, termed \"self-rectification\", automatically refines this target into a coherent, seamless texture, while faithfully preserving the distinct visual characteristics of the reference exemplar. Our method leverages a pre-trained diffusion network, and uses self-attention mechanisms, to gradually align the synthesized texture with the reference, ensuring the retention of the structures in the provided target. Through experimental validation, our approach exhibits exceptional proficiency in handling non-stationary textures, demonstrating significant advancements in texture synthesis when compared to existing state-of-the-art techniques. Code is available at https://github.com/xiaorongjun000/Self-Rectification" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 462, + 128, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 462, + 128, + 474 + ], + "spans": [ + { + "bbox": [ + 47, + 462, + 128, + 474 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 482, + 287, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 482, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 46, + 482, + 287, + 553 + ], + "type": "text", + "content": "Example-based texture synthesis aims to generate a texture that faithfully captures all the visual characteristics of a provided reference texture exemplar. The key challenge is to generate a texture that visually mimics the reference, while avoiding exact replication and without producing conspicuous, unnatural artifacts." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 554, + 287, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 554, + 287, + 637 + ], + "spans": [ + { + "bbox": [ + 46, + 554, + 287, + 637 + ], + "type": "text", + "content": "Over the past decades, numerous methods have emerged for synthesizing textures from examples, and many have demonstrated impressive results, particularly for homogeneous textures that can be accurately captured by stationary models [45]. Nonetheless, a significant challenge remains when dealing with real-world textures that are inherently inhomogeneous, and non-stationary." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 638, + 287, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 638, + 287, + 686 + ], + "spans": [ + { + "bbox": [ + 46, + 638, + 287, + 686 + ], + "type": "text", + "content": "Non-stationary textures exhibit distinctive attributes, such as sprawling irregular large-scale structures or variations in attributes such as color, local orientation, and local scale. Fig. 1 (left) shows two examples. Mimicking such" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 198, + 386, + 276 + ], + "blocks": [ + { + "bbox": [ + 307, + 198, + 386, + 276 + ], + "lines": [ + { + "bbox": [ + 307, + 198, + 386, + 276 + ], + "spans": [ + { + "bbox": [ + 307, + 198, + 386, + 276 + ], + "type": "image", + "image_path": "f31eb4945484ffe2712462bafa4b9d2d3b33e2d6deca38cb83dc50b03ad4b29d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 277, + 386, + 355 + ], + "blocks": [ + { + "bbox": [ + 307, + 277, + 386, + 355 + ], + "lines": [ + { + "bbox": [ + 307, + 277, + 386, + 355 + ], + "spans": [ + { + "bbox": [ + 307, + 277, + 386, + 355 + ], + "type": "image", + "image_path": "e6f9c02fdb75b4b7b60fbd29e67750440ae9d53990ed575468762fc44b0bf71d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 364, + 547, + 430 + ], + "lines": [ + { + "bbox": [ + 305, + 364, + 547, + 430 + ], + "spans": [ + { + "bbox": [ + 305, + 364, + 547, + 430 + ], + "type": "text", + "content": "Figure 1. Our method takes as input a reference texture (left), and a crude target texture provided by the user (middle column), which may lack coherence and completeness. Self-rectification is used to transform the target into a visually coherent texture (right) that complies with the structure of the crude target, while exhibiting the visual characteristics of the reference texture." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 387, + 198, + 465, + 276 + ], + "blocks": [ + { + "bbox": [ + 387, + 198, + 465, + 276 + ], + "lines": [ + { + "bbox": [ + 387, + 198, + 465, + 276 + ], + "spans": [ + { + "bbox": [ + 387, + 198, + 465, + 276 + ], + "type": "image", + "image_path": "29ad14d7f1db57e93642510d70d71b1f0201738ea9f94457826b886bcb999a1a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 387, + 277, + 465, + 355 + ], + "blocks": [ + { + "bbox": [ + 387, + 277, + 465, + 355 + ], + "lines": [ + { + "bbox": [ + 387, + 277, + 465, + 355 + ], + "spans": [ + { + "bbox": [ + 387, + 277, + 465, + 355 + ], + "type": "image", + "image_path": "1212ca72001c098789299d484ca37d8a6a2b066b53686bc5f5a3a3f0ad0aa22d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 466, + 198, + 545, + 276 + ], + "blocks": [ + { + "bbox": [ + 466, + 198, + 545, + 276 + ], + "lines": [ + { + "bbox": [ + 466, + 198, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 466, + 198, + 545, + 276 + ], + "type": "image", + "image_path": "0f54f213629f7fedd67bce73e72d53484850dc2dd9e32e7693a9d75021168e7e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 466, + 277, + 545, + 355 + ], + "blocks": [ + { + "bbox": [ + 466, + 277, + 545, + 355 + ], + "lines": [ + { + "bbox": [ + 466, + 277, + 545, + 355 + ], + "spans": [ + { + "bbox": [ + 466, + 277, + 545, + 355 + ], + "type": "image", + "image_path": "7ec8b08f3d3ce6cb996ca2d27bc176cda12f4b1a5cdb2c25e7b10708a04b2c26.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 461, + 545, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 461, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 305, + 461, + 545, + 486 + ], + "type": "text", + "content": "complex structures and variations through example-based synthesis is a long-standing challenging task [31]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 486, + 546, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 546, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 546, + 605 + ], + "type": "text", + "content": "The emergence of neural networks has provided powerful means to deal with non-stationary textures. Zhou et al. [51] introduced a method that involves overfitting a GAN, where the encoder extracts structural guidance for the decoder. This approach provides a viable means of spatially extending non-stationary textures while preserving the visual characteristics of the reference. However, this technique requires an extremely long optimization process for a single texture exemplar, and moreover, it fails to provide any controllability or editability." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": "In this paper, we introduce a two-step lazy-editing technique where the user first edits the given reference texture using conventional image editing tools, to obtain an extremely rough initial result, which may be incomplete and incoherent, as demonstrated in Fig. 1 (middle). Next, our technique automatically rectifies the edited texture into a regularized, coherent and seamless texture that follows the rough target, while retaining the local characteristics of the reference; see, e.g., Fig. 1 (right). We term this regulariza" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 135, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 135, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 135, + 703 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7767" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": "tion process self-rectification, since the texture is rectified based on the input texture itself." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 96, + 287, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 96, + 287, + 299 + ], + "spans": [ + { + "bbox": [ + 45, + 96, + 287, + 299 + ], + "type": "text", + "content": "To self-rectify the crude target, we use a pre-trained diffusion network, and synthesize the final rectified result while utilizing the intermediate byproducts obtained by inverting both the initial target and the reference exemplar. Inspired by editing approaches that leverage self-attention [1,6,20], we extract certain self-attention features of the diffusion model during the inversion of the rough target, as well as the reference. These features are then injected into various steps in the inversion and denoising of the generated texture. Such feature injection forms \"cross-attention\" between the reference and target features [1, 6], and thereby, in the course of the diffusion process, the target texture is progressively refined to be increasingly locally similar to the reference, while retaining the global structure prescribed by the provided target. This self-rectification occurs in two passes, first addressing the larger scale structure and subsequently focusing on finer local details." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 45, + 300, + 287, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 300, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 45, + 300, + 287, + 443 + ], + "type": "text", + "content": "To enhance the synthesis quality, we augment the reference texture with several transformed copies of it, which increases the diversity of the reference patterns, thereby improving the compatibility between the reference and the synthesized result. The augmented source features are injected into the corresponding target layers as well. Furthermore, we show that our method can also be applied to lazy editing of natural images, using the same means of synthesizing highly non-stationary textures. Experiments show that our method can deal with a large scope of nonstationary textures, with unprecedented flexibility and quality, compared to the state-of-the-art." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 453, + 134, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 453, + 134, + 466 + ], + "spans": [ + { + "bbox": [ + 47, + 453, + 134, + 466 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 474, + 286, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 474, + 286, + 628 + ], + "spans": [ + { + "bbox": [ + 45, + 474, + 286, + 628 + ], + "type": "text", + "content": "Non-stationary texture synthesis. Early methods in example-based texture synthesis mainly focused on synthesizing textures with stationary characteristics [8, 9, 19, 21, 22, 43, 44]. To cope with non-stationary textures, researchers typically involve additional guidance for the control of distinctive attributes, e.g., using label maps to guide the layout synthesis of composite textures [15, 24, 31], representing the spatial variations of weathering textures with age/progression maps [3, 42], and describing the local orientations of directional textures with vector fields [26, 50]. Although the aforementioned methods have shown success, the guidance required from the user is tedious to provide, yet still limited." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": "In the deep learning era, neural methods for texture generation have rapidly emerged, either through new texture losses for texture optimization [12, 13, 49] or via training generative networks [4, 5, 10, 18, 23, 29, 32, 35-37, 40, 51]. Among these methods, Sendik and Cohen-Or [35] attempted to preserve the non-local structures of non-stationary textures by regularizing the feature correla" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 310, + 68, + 547, + 175 + ], + "blocks": [ + { + "bbox": [ + 310, + 68, + 547, + 175 + ], + "lines": [ + { + "bbox": [ + 310, + 68, + 547, + 175 + ], + "spans": [ + { + "bbox": [ + 310, + 68, + 547, + 175 + ], + "type": "image", + "image_path": "59cb17c6ec74110402a6ead0e05ea69d01522e1303202883d35bef37cad0d555.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 186, + 547, + 253 + ], + "lines": [ + { + "bbox": [ + 305, + 186, + 547, + 253 + ], + "spans": [ + { + "bbox": [ + 305, + 186, + 547, + 253 + ], + "type": "text", + "content": "Figure 2. Framework overview. Given a reference texture " + }, + { + "bbox": [ + 305, + 186, + 547, + 253 + ], + "type": "inline_equation", + "content": "I^R" + }, + { + "bbox": [ + 305, + 186, + 547, + 253 + ], + "type": "text", + "content": ", we allow the user to quickly build a target image " + }, + { + "bbox": [ + 305, + 186, + 547, + 253 + ], + "type": "inline_equation", + "content": "I^{tar}" + }, + { + "bbox": [ + 305, + 186, + 547, + 253 + ], + "type": "text", + "content": " in a lazy-editing manner. A coarse-to-fine synthesis is performed by running self-rectification twice. The coarse stage synthesizes a coarse yet complete overall structure, and the fine stage refines its output " + }, + { + "bbox": [ + 305, + 186, + 547, + 253 + ], + "type": "inline_equation", + "content": "I_{\\text{coarse}}^*" + }, + { + "bbox": [ + 305, + 186, + 547, + 253 + ], + "type": "text", + "content": " with finer and more accurate details, producing the final result " + }, + { + "bbox": [ + 305, + 186, + 547, + 253 + ], + "type": "inline_equation", + "content": "I^*" + }, + { + "bbox": [ + 305, + 186, + 547, + 253 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 273, + 545, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 273, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 305, + 273, + 545, + 370 + ], + "type": "text", + "content": "tion between different locations. Zhou et al. [51] proposed overfitting a GAN to learn the expansion from a small texture block to a larger one containing it. Their approach involves the GAN's encoder extracting the global structure of an input texture, duplicated by bottleneck residual blocks before decoding. Although these trained models can effectively extend non-stationary textures, their overfitting nature severely restricts their generalization and controllability." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 384, + 545, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 384, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 305, + 384, + 545, + 479 + ], + "type": "text", + "content": "Diffusion-based image synthesis. The emergence of large-scale generative diffusion models, such as Stable Diffusion (SD) [30] and DALLE-2 [28], has revolutionized image synthesis due to their unprecedented generation power. To make the pre-trained diffusion models synthesize on image conditions, many solutions are proposed, including optimizing prompt tokens [2, 11], fine-tuning the entire model [33], or training an additional adapter [17,34,46-48]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 479, + 545, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 479, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 305, + 479, + 545, + 623 + ], + "type": "text", + "content": "Recently, researchers have investigated the role of the intermediate attention maps and features in the diffuser [7, 14,27,39], finding them crucial for layout/structure synthesis. The method we present in this paper, builds upon the mechanism of injecting keys and values from attention layers of one diffusion process into another, as a means for transferring visual features between images [1,6,20]. We use such injection to \"copy\" the local patterns from a source texture to a target one. In addition to the injection we also incorporate coarse-to-fine and data augmentation schemes, forming the self-rectification framework for generating nonstationary textures." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 634, + 361, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 634, + 361, + 645 + ], + "spans": [ + { + "bbox": [ + 306, + 634, + 361, + 645 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "content": "Our method employs a two-step process for synthesizing a new texture " + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "I^{*}" + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "content": " resembling a provided non-stationary reference texture " + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "I^{R}" + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "content": ", while guided by a user-provided rough target " + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "I^{tar}" + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "content": ", as depicted in Fig. 2. Initially, the user rapidly creates " + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "I^{tar}" + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "content": ", by assembling patches from the source refer" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7768" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 68, + 545, + 205 + ], + "blocks": [ + { + "bbox": [ + 51, + 68, + 545, + 205 + ], + "lines": [ + { + "bbox": [ + 51, + 68, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 51, + 68, + 545, + 205 + ], + "type": "image", + "image_path": "e869299231132e9069bda6cf75424c234d4e42aa5bed010dde050968793b7191.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 213, + 547, + 258 + ], + "lines": [ + { + "bbox": [ + 46, + 213, + 547, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 547, + 258 + ], + "type": "text", + "content": "Figure 3. Our self-rectification synthesizes an output texture " + }, + { + "bbox": [ + 46, + 213, + 547, + 258 + ], + "type": "inline_equation", + "content": "I^{*}" + }, + { + "bbox": [ + 46, + 213, + 547, + 258 + ], + "type": "text", + "content": " via structure-preserving inversion from a rough target image " + }, + { + "bbox": [ + 46, + 213, + 547, + 258 + ], + "type": "inline_equation", + "content": "I^{lar}" + }, + { + "bbox": [ + 46, + 213, + 547, + 258 + ], + "type": "text", + "content": " and fine texture sampling using the reference " + }, + { + "bbox": [ + 46, + 213, + 547, + 258 + ], + "type": "inline_equation", + "content": "I^{R}" + }, + { + "bbox": [ + 46, + 213, + 547, + 258 + ], + "type": "text", + "content": ". Both processes require the injection of self-attention features (" + }, + { + "bbox": [ + 46, + 213, + 547, + 258 + ], + "type": "inline_equation", + "content": "KV" + }, + { + "bbox": [ + 46, + 213, + 547, + 258 + ], + "type": "text", + "content": ") from the DDIM inversion of a corresponding reference. More specifically, for structure-preserving inversion, the reference is the target image itself, denoted as " + }, + { + "bbox": [ + 46, + 213, + 547, + 258 + ], + "type": "inline_equation", + "content": "I^{IR}" + }, + { + "bbox": [ + 46, + 213, + 547, + 258 + ], + "type": "text", + "content": ". For fine texture sampling, the input exemplar " + }, + { + "bbox": [ + 46, + 213, + 547, + 258 + ], + "type": "inline_equation", + "content": "I^{R}" + }, + { + "bbox": [ + 46, + 213, + 547, + 258 + ], + "type": "text", + "content": " is used to inject features that help to synthesize a plausible output with fine texture details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 279, + 287, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 279, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 46, + 279, + 287, + 374 + ], + "type": "text", + "content": "ence texture. Such a lazy-edit forms a basic layout for the output image, " + }, + { + "bbox": [ + 46, + 279, + 287, + 374 + ], + "type": "inline_equation", + "content": "I^{*}" + }, + { + "bbox": [ + 46, + 279, + 287, + 374 + ], + "type": "text", + "content": ". However, this initial sketch may be incoherent or incomplete, hence necessitating rectification. Subsequently, the texture undergoes self-rectification to align with both the user's coarse layout and the reference image's detailed texture characteristics. This self-rectification process is depicted in Fig. 2 and is executed in two stages: coarse rectification, followed by a finer one." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 374, + 287, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 374, + 287, + 481 + ], + "spans": [ + { + "bbox": [ + 46, + 374, + 287, + 481 + ], + "type": "text", + "content": "Considering that the rough target " + }, + { + "bbox": [ + 46, + 374, + 287, + 481 + ], + "type": "inline_equation", + "content": "I^{tar}" + }, + { + "bbox": [ + 46, + 374, + 287, + 481 + ], + "type": "text", + "content": " comprises patches derived from the reference texture " + }, + { + "bbox": [ + 46, + 374, + 287, + 481 + ], + "type": "inline_equation", + "content": "I^R" + }, + { + "bbox": [ + 46, + 374, + 287, + 481 + ], + "type": "text", + "content": ", we implement a process of self-rectification, detailed below. In a broad sense, we utilize a pre-trained latent diffusion model to invert both the target image and the reference image into an initial latent noise. This is followed by employing feature injection during the sampling process of " + }, + { + "bbox": [ + 46, + 374, + 287, + 481 + ], + "type": "inline_equation", + "content": "I^*" + }, + { + "bbox": [ + 46, + 374, + 287, + 481 + ], + "type": "text", + "content": ". The features from " + }, + { + "bbox": [ + 46, + 374, + 287, + 481 + ], + "type": "inline_equation", + "content": "I^{tar}" + }, + { + "bbox": [ + 46, + 374, + 287, + 481 + ], + "type": "text", + "content": " guide the structural synthesis and the features from " + }, + { + "bbox": [ + 46, + 374, + 287, + 481 + ], + "type": "inline_equation", + "content": "I^R" + }, + { + "bbox": [ + 46, + 374, + 287, + 481 + ], + "type": "text", + "content": " guide the fine texture details." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 482, + 287, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 482, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 482, + 287, + 518 + ], + "type": "text", + "content": "Below, we first briefly review diffusion models and their self-attention and the cross-KV-injection mechanisms, and then proceed to present our self-rectification technique." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 525, + 216, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 525, + 216, + 538 + ], + "spans": [ + { + "bbox": [ + 47, + 525, + 216, + 538 + ], + "type": "text", + "content": "3.1. DDIM Sampling and Inversion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 544, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 616 + ], + "type": "text", + "content": "Denoising diffusion models [16, 38] involve two processes: a noising process that gradually transforms an input image into Gaussian noise, and a denoising/sampling process that generates images from Gaussian samples. Using DDIM sampling [38], starting from a noise sample " + }, + { + "bbox": [ + 46, + 544, + 287, + 616 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{T}}" + }, + { + "bbox": [ + 46, + 544, + 287, + 616 + ], + "type": "text", + "content": ", one can generate a clean sample " + }, + { + "bbox": [ + 46, + 544, + 287, + 616 + ], + "type": "inline_equation", + "content": "z_{0}" + }, + { + "bbox": [ + 46, + 544, + 287, + 616 + ], + "type": "text", + "content": " via T deterministic steps:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 63, + 623, + 287, + 638 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 623, + 287, + 638 + ], + "spans": [ + { + "bbox": [ + 63, + 623, + 287, + 638 + ], + "type": "interline_equation", + "content": "z _ {t - 1} = \\sqrt {\\bar {\\alpha} _ {t - 1}} f _ {\\theta} (z _ {t}, t) + \\sqrt {1 - \\bar {\\alpha} _ {t - 1}} \\epsilon_ {\\theta} (z _ {t}, t), \\quad (1)", + "image_path": "72ac9f51fa41885ebd99aa6965f745b00cd87ed07b16c04a0fa849466759a8c3.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 646, + 287, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 646, + 287, + 681 + ], + "spans": [ + { + "bbox": [ + 46, + 646, + 287, + 681 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 646, + 287, + 681 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 46, + 646, + 287, + 681 + ], + "type": "text", + "content": " is a noise prediction network conditioned on the current noisy sample " + }, + { + "bbox": [ + 46, + 646, + 287, + 681 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 46, + 646, + 287, + 681 + ], + "type": "text", + "content": " and timestamp " + }, + { + "bbox": [ + 46, + 646, + 287, + 681 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 646, + 287, + 681 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 646, + 287, + 681 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_{t}" + }, + { + "bbox": [ + 46, + 646, + 287, + 681 + ], + "type": "text", + "content": " is the noise scaling factor defined in [38], and " + }, + { + "bbox": [ + 46, + 646, + 287, + 681 + ], + "type": "inline_equation", + "content": "f_{\\theta}(z_t,t)" + }, + { + "bbox": [ + 46, + 646, + 287, + 681 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 93, + 689, + 287, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 689, + 287, + 717 + ], + "spans": [ + { + "bbox": [ + 93, + 689, + 287, + 717 + ], + "type": "interline_equation", + "content": "f _ {\\theta} \\left(z _ {t}, t\\right) = \\frac {z _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta} \\left(z _ {t} , t\\right)}{\\sqrt {\\bar {\\alpha} _ {t}}}. \\tag {2}", + "image_path": "f6a40c009c6ce918350ab520b1feee083e13152a8c9fa27ea68ac583faf30a70.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 279, + 545, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 279, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 305, + 279, + 545, + 315 + ], + "type": "text", + "content": "In the opposite direction, " + }, + { + "bbox": [ + 305, + 279, + 545, + 315 + ], + "type": "inline_equation", + "content": "DDIM" + }, + { + "bbox": [ + 305, + 279, + 545, + 315 + ], + "type": "text", + "content": " inversion [38] of a given image " + }, + { + "bbox": [ + 305, + 279, + 545, + 315 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 305, + 279, + 545, + 315 + ], + "type": "text", + "content": " is the process of incrementally adding deterministic noise, until obtaining " + }, + { + "bbox": [ + 305, + 279, + 545, + 315 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{T}}" + }, + { + "bbox": [ + 305, + 279, + 545, + 315 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 321, + 323, + 545, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 323, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 321, + 323, + 545, + 338 + ], + "type": "interline_equation", + "content": "z _ {t + 1} = \\sqrt {\\bar {\\alpha} _ {t + 1}} f _ {\\theta} \\left(z _ {t}, t\\right) + \\sqrt {1 - \\bar {\\alpha} _ {t + 1}} \\epsilon_ {\\theta} \\left(z _ {t}, t\\right). \\tag {3}", + "image_path": "d2cedb1ae4113b43e26d7a2b058eb93aaa021c51b11be8d14b5f58db057dd02d.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 345, + 545, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 345, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 306, + 345, + 545, + 357 + ], + "type": "text", + "content": "Such inversion leads to a nearly faithful reconstruction [38]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 364, + 495, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 364, + 495, + 376 + ], + "spans": [ + { + "bbox": [ + 306, + 364, + 495, + 376 + ], + "type": "text", + "content": "3.2. Stable Diffusion and Self-Attention." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 382, + 545, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 382, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 304, + 382, + 545, + 454 + ], + "type": "text", + "content": "We base our texture synthesis framework on Stable Diffusion (SD) [30], which is a pretrained latent diffusion model consisting of an encoder " + }, + { + "bbox": [ + 304, + 382, + 545, + 454 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 304, + 382, + 545, + 454 + ], + "type": "text", + "content": " that maps an input image into the latent space, and a decoder " + }, + { + "bbox": [ + 304, + 382, + 545, + 454 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 382, + 545, + 454 + ], + "type": "text", + "content": " that reconstructs a latent code back into image space. Both DDIM sampling and DDIM inversion are performed in the SD latent space." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 454, + 545, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 454, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 304, + 454, + 545, + 514 + ], + "type": "text", + "content": "The noise predictor " + }, + { + "bbox": [ + 304, + 454, + 545, + 514 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 304, + 454, + 545, + 514 + ], + "type": "text", + "content": " of SD is a large-scale U-Net that contains multiple self-attention modules [41]. Each self-attention layer transforms its input intermediate feature map (also called spatial features) into an attended representation by the following equation:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 345, + 521, + 545, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 521, + 545, + 550 + ], + "spans": [ + { + "bbox": [ + 345, + 521, + 545, + 550 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t} (Q, K, V) = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d}}\\right) V, \\tag {4}", + "image_path": "3b91598f52de0882b99d90a43e32cd4e3ed9e888c55e9fea82f9d92d00641372.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 556, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 545, + 640 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 556, + 545, + 640 + ], + "type": "inline_equation", + "content": "Q, K" + }, + { + "bbox": [ + 304, + 556, + 545, + 640 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 556, + 545, + 640 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 304, + 556, + 545, + 640 + ], + "type": "text", + "content": " are the queries, keys, and values, respectively, obtained by learned linear projections of the same input spatial features, having dimension " + }, + { + "bbox": [ + 304, + 556, + 545, + 640 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 556, + 545, + 640 + ], + "type": "text", + "content": ". The self-attention mechanism uses the similarities between the queries and keys as attention scores to weigh the importance or relevance of the values. Relevant info is thus aggregated as the attended representation." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 647, + 390, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 390, + 659 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 390, + 659 + ], + "type": "text", + "content": "3.3. KV-Injection" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "The self-attention features contain rich information about both the large-scale structures and local fine details of an input image [1, 6, 39]. Tumanyan et al. [39] demonstrate that by injecting the spatial features and the queries" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7769" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 289, + 205 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 289, + 205 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 289, + 205 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 289, + 205 + ], + "type": "image", + "image_path": "01fc19e8ac5e749930664f6c79335d9cf8055d022cdb8134c0d0ccaadff0b0da.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 211, + 288, + 277 + ], + "lines": [ + { + "bbox": [ + 46, + 211, + 288, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 211, + 288, + 277 + ], + "type": "text", + "content": "Figure 4. Visualization of the intermediate latent codes in the inversion. For the standard DDIM inversion (top), the U-Net predicts noise to diffuse the distinctive structures so as to transform the input into Gaussian noise. In contrast, our structure-preserving inversion (bottom) reserves the distinctive patterns from user edits along the inversion process. See texts in Sec. 3.4 for more details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 298, + 286, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 298, + 286, + 382 + ], + "spans": [ + { + "bbox": [ + 46, + 298, + 286, + 382 + ], + "type": "text", + "content": "and keys " + }, + { + "bbox": [ + 46, + 298, + 286, + 382 + ], + "type": "inline_equation", + "content": "(QK)" + }, + { + "bbox": [ + 46, + 298, + 286, + 382 + ], + "type": "text", + "content": " from the self-attention layers of a source (guidance) image into the corresponding layer of the generated target image during the sampling process, one can preserve the layout of the source while modifying its appearance. Injecting only the " + }, + { + "bbox": [ + 46, + 298, + 286, + 382 + ], + "type": "inline_equation", + "content": "KV" + }, + { + "bbox": [ + 46, + 298, + 286, + 382 + ], + "type": "text", + "content": " features of a source reference, instead, transfers its appearance, including textures, to the generated target [1,6]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 382, + 286, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 382, + 286, + 478 + ], + "spans": [ + { + "bbox": [ + 46, + 382, + 286, + 478 + ], + "type": "text", + "content": "In our case, the target image is a rough and incomplete guidance, typically containing only a few source patches rotated and placed freely by the user. A complete overall structure needs to be synthesized reasonably yet conforming to user's constraints. Locally, however, the resulting texture should resemble the reference. To achieve these goals, we adapt the " + }, + { + "bbox": [ + 46, + 382, + 286, + 478 + ], + "type": "inline_equation", + "content": "KV" + }, + { + "bbox": [ + 46, + 382, + 286, + 478 + ], + "type": "text", + "content": "-injection into both the DDIM inversion and sampling, bringing a novel self-rectification operation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 485, + 147, + 497 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 485, + 147, + 497 + ], + "spans": [ + { + "bbox": [ + 47, + 485, + 147, + 497 + ], + "type": "text", + "content": "3.4. Self-rectification" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 503, + 287, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 503, + 287, + 575 + ], + "spans": [ + { + "bbox": [ + 46, + 503, + 287, + 575 + ], + "type": "text", + "content": "As shown in Fig. 3, our self-rectification consists of two parts: structure-preserving inversion and fine texture sampling. Both are performed in the SD latent space. The last latent code of inversion is used as the starting code of the sampling. The core modification to both processes is the " + }, + { + "bbox": [ + 46, + 503, + 287, + 575 + ], + "type": "inline_equation", + "content": "KV" + }, + { + "bbox": [ + 46, + 503, + 287, + 575 + ], + "type": "text", + "content": "-injection of self-attention features." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 590, + 287, + 709 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 590, + 287, + 709 + ], + "spans": [ + { + "bbox": [ + 46, + 590, + 287, + 709 + ], + "type": "text", + "content": "Structure-preserving inversion. The standard DDIM inversion (Eq. (3)) progressively transforms an input image (SD latent code) into pure Gaussian noise. At each time step, the noise to be added is predicted by the U-Net. As visualized in Fig. 4, for a given target image, the noise predicted by the U-Net for early time steps (" + }, + { + "bbox": [ + 46, + 590, + 287, + 709 + ], + "type": "inline_equation", + "content": "e.g., t \\leq 20" + }, + { + "bbox": [ + 46, + 590, + 287, + 709 + ], + "type": "text", + "content": ") is mainly distributed to \"diffuse\" the prominent structures so that the distinctive patterns from user edits get scattered and random. As the inversion progresses, the magnitude of the noise added in each step becomes smaller and similar for" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 305, + 70, + 547, + 114 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 547, + 114 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 547, + 114 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 547, + 114 + ], + "type": "image", + "image_path": "03503cdfdd39e5c6a714b99e7c1af312a21a75b8fe9d36674c93e7cfb801ca98.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 119, + 547, + 208 + ], + "lines": [ + { + "bbox": [ + 304, + 119, + 547, + 208 + ], + "spans": [ + { + "bbox": [ + 304, + 119, + 547, + 208 + ], + "type": "text", + "content": "Figure 5. Visualization of the intermediate latent codes in the fine texture sampling. Here, for the first 20 steps (from " + }, + { + "bbox": [ + 304, + 119, + 547, + 208 + ], + "type": "inline_equation", + "content": "t = 50" + }, + { + "bbox": [ + 304, + 119, + 547, + 208 + ], + "type": "text", + "content": " to 30), we perform the standard DDIM sampling to reconstruct the target layout. Next, we perform " + }, + { + "bbox": [ + 304, + 119, + 547, + 208 + ], + "type": "inline_equation", + "content": "KV" + }, + { + "bbox": [ + 304, + 119, + 547, + 208 + ], + "type": "text", + "content": "-injection in the remaining sampling steps (" + }, + { + "bbox": [ + 304, + 119, + 547, + 208 + ], + "type": "inline_equation", + "content": "t = 30" + }, + { + "bbox": [ + 304, + 119, + 547, + 208 + ], + "type": "text", + "content": " to 0), to synthesize fine textures for the output image. The rightmost shows the result produced by simply performing standard DDIM sampling for all steps, i.e., " + }, + { + "bbox": [ + 304, + 119, + 547, + 208 + ], + "type": "inline_equation", + "content": "S = 50" + }, + { + "bbox": [ + 304, + 119, + 547, + 208 + ], + "type": "text", + "content": ". No additional structure is synthesized to complete the user edits." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 228, + 545, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 545, + 348 + ], + "type": "text", + "content": "all latent pixels, since the \"diffusion\" is getting close to done. As no text prompt is involved, the noise prediction in each inversion step is dominantly determined by the self-attention mechanism in the U-Net. Our key observation is that if we inject the " + }, + { + "bbox": [ + 304, + 228, + 545, + 348 + ], + "type": "inline_equation", + "content": "KV" + }, + { + "bbox": [ + 304, + 228, + 545, + 348 + ], + "type": "text", + "content": " features from a large time step " + }, + { + "bbox": [ + 304, + 228, + 545, + 348 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 304, + 228, + 545, + 348 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 228, + 545, + 348 + ], + "type": "inline_equation", + "content": "\\gg \\mathrm{T} / 2" + }, + { + "bbox": [ + 304, + 228, + 545, + 348 + ], + "type": "text", + "content": ") into an early time step " + }, + { + "bbox": [ + 304, + 228, + 545, + 348 + ], + "type": "inline_equation", + "content": "t_2" + }, + { + "bbox": [ + 304, + 228, + 545, + 348 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 228, + 545, + 348 + ], + "type": "inline_equation", + "content": "\\ll \\mathrm{T} / 2" + }, + { + "bbox": [ + 304, + 228, + 545, + 348 + ], + "type": "text", + "content": "), the noise predicted at " + }, + { + "bbox": [ + 304, + 228, + 545, + 348 + ], + "type": "inline_equation", + "content": "t_2" + }, + { + "bbox": [ + 304, + 228, + 545, + 348 + ], + "type": "text", + "content": " will be smaller and more spatially uniform. The distinctive patterns of " + }, + { + "bbox": [ + 304, + 228, + 545, + 348 + ], + "type": "inline_equation", + "content": "I^{tar}" + }, + { + "bbox": [ + 304, + 228, + 545, + 348 + ], + "type": "text", + "content": ", reflecting the user's edits, are thus better preserved. Hence, we refer to this process as structure-preserving inversion." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 348, + 545, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 348, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 304, + 348, + 545, + 420 + ], + "type": "text", + "content": "Specifically, given a rough target image " + }, + { + "bbox": [ + 304, + 348, + 545, + 420 + ], + "type": "inline_equation", + "content": "I^{tar}" + }, + { + "bbox": [ + 304, + 348, + 545, + 420 + ], + "type": "text", + "content": ", we invert it twice. The first inversion is a standard DDIM inversion. The produced self-attention features along the noisig steps are regarded as the inversion reference (IR) for the second inversion. The self-attention during the second inversion, at time step " + }, + { + "bbox": [ + 304, + 348, + 545, + 420 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 348, + 545, + 420 + ], + "type": "text", + "content": ", is now given by:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 428, + 547, + 471 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 428, + 547, + 471 + ], + "spans": [ + { + "bbox": [ + 306, + 428, + 547, + 471 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t} \\left(Q _ {t} ^ {\\text {t a r}}, K _ {\\mathrm {T} - t} ^ {\\mathrm {I R}}, V _ {\\mathrm {T} - t} ^ {\\mathrm {I R}}\\right) = \\operatorname {S o f t m a x} \\left(\\frac {Q _ {t} ^ {\\text {t a r}} \\left(K _ {\\mathrm {T} - t} ^ {\\mathrm {I R}}\\right) ^ {T}}{\\sqrt {d}}\\right) V _ {\\mathrm {T} - t} ^ {\\mathrm {I R}}, \\tag {5}", + "image_path": "201be78212d8321c7c117fd66eae06a3b751ee59df88efa92a2b2140a8a4102e.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 472, + 545, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 545, + 543 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 472, + 545, + 543 + ], + "type": "inline_equation", + "content": "\\mathrm{T}" + }, + { + "bbox": [ + 304, + 472, + 545, + 543 + ], + "type": "text", + "content": " denotes the total number of steps (we use " + }, + { + "bbox": [ + 304, + 472, + 545, + 543 + ], + "type": "inline_equation", + "content": "\\mathrm{T} = 50" + }, + { + "bbox": [ + 304, + 472, + 545, + 543 + ], + "type": "text", + "content": "). Here " + }, + { + "bbox": [ + 304, + 472, + 545, + 543 + ], + "type": "inline_equation", + "content": "KV" + }, + { + "bbox": [ + 304, + 472, + 545, + 543 + ], + "type": "text", + "content": " features are injected in a reverse order in the second round. We would like to stress that we have also experimented with using an offset, i.e., replacing " + }, + { + "bbox": [ + 304, + 472, + 545, + 543 + ], + "type": "inline_equation", + "content": "\\mathrm{T} - t" + }, + { + "bbox": [ + 304, + 472, + 545, + 543 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 472, + 545, + 543 + ], + "type": "inline_equation", + "content": "t + offset" + }, + { + "bbox": [ + 304, + 472, + 545, + 543 + ], + "type": "text", + "content": " in the above equation, but found no advantage (see supplementary for more analysis and comparisons)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 544, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 545, + 651 + ], + "type": "text", + "content": "During the second inversion, we do not perform " + }, + { + "bbox": [ + 304, + 544, + 545, + 651 + ], + "type": "inline_equation", + "content": "KV" + }, + { + "bbox": [ + 304, + 544, + 545, + 651 + ], + "type": "text", + "content": " injections for all time steps. We set a parameter " + }, + { + "bbox": [ + 304, + 544, + 545, + 651 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 544, + 545, + 651 + ], + "type": "text", + "content": " that defines the time step beyond which we continue with standard inversion. Fig. 4 shows the effect of our structure-preserving inversion, where " + }, + { + "bbox": [ + 304, + 544, + 545, + 651 + ], + "type": "inline_equation", + "content": "P = 30" + }, + { + "bbox": [ + 304, + 544, + 545, + 651 + ], + "type": "text", + "content": ". Since the final latent code after the second inversion still contains rich information about the distinctive structures from " + }, + { + "bbox": [ + 304, + 544, + 545, + 651 + ], + "type": "inline_equation", + "content": "I^{tar}" + }, + { + "bbox": [ + 304, + 544, + 545, + 651 + ], + "type": "text", + "content": ", the subsequent sampling that starts from this latent code is guided to synthesize a global structure that complies with " + }, + { + "bbox": [ + 304, + 544, + 545, + 651 + ], + "type": "inline_equation", + "content": "I^{tar}" + }, + { + "bbox": [ + 304, + 544, + 545, + 651 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Fine texture sampling. We start the DDIM sampling from the final structure-preserving latent code. The sampling uses the U-Net to predict the noise to be removed, as defined in Eq. (1). However, simply performing standard" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7770" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "text", + "content": "DDIM sampling for all denoising steps results in a nearly reconstructed target image since DDIM sampling is deterministic. Therefore, we set the first S steps (i.e., from time step T to T - S) to reconstruct the target layout to a certain extent. For the remaining T - S denoising steps, we synthesize the output image " + }, + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "inline_equation", + "content": "I^{*}" + }, + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "text", + "content": " by matching fine textures from the reference via " + }, + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "inline_equation", + "content": "KV" + }, + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "text", + "content": "-injection from the reference " + }, + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "inline_equation", + "content": "I^{R}" + }, + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 156, + 286, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 156, + 286, + 204 + ], + "spans": [ + { + "bbox": [ + 47, + 156, + 286, + 204 + ], + "type": "text", + "content": "To this end, we first DDIM-invert the reference texture " + }, + { + "bbox": [ + 47, + 156, + 286, + 204 + ], + "type": "inline_equation", + "content": "I^{R}" + }, + { + "bbox": [ + 47, + 156, + 286, + 204 + ], + "type": "text", + "content": ". At denoising time step " + }, + { + "bbox": [ + 47, + 156, + 286, + 204 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 156, + 286, + 204 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 47, + 156, + 286, + 204 + ], + "type": "inline_equation", + "content": "t > \\mathrm{T} - \\mathrm{S}" + }, + { + "bbox": [ + 47, + 156, + 286, + 204 + ], + "type": "text", + "content": "), the " + }, + { + "bbox": [ + 47, + 156, + 286, + 204 + ], + "type": "inline_equation", + "content": "KV" + }, + { + "bbox": [ + 47, + 156, + 286, + 204 + ], + "type": "text", + "content": " features extracted during the inversion are injected into the corresponding self-attention layers of the synthesized texture:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 212, + 286, + 245 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 212, + 286, + 245 + ], + "spans": [ + { + "bbox": [ + 59, + 212, + 286, + 245 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t} \\left(Q _ {t} ^ {*}, K _ {t} ^ {R}, V _ {t} ^ {R}\\right) = \\operatorname {S o f t m a x} \\left(\\frac {Q _ {t} ^ {*} \\left(K _ {t} ^ {R}\\right) ^ {T}}{\\sqrt {d}}\\right) V _ {t} ^ {R}. \\tag {6}", + "image_path": "edc14e8fecb2e2abb3b3e8b62ca27940e954fc94094610f40509ff0ea3ec697f.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 253, + 286, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 253, + 286, + 301 + ], + "spans": [ + { + "bbox": [ + 47, + 253, + 286, + 301 + ], + "type": "text", + "content": "This forms a cross-image attention, where corresponding fine local patterns in the reference are transferred to the output image in a plausible manner. Fig. 5 visualizes the intermediate process of our texture sampling, where " + }, + { + "bbox": [ + 47, + 253, + 286, + 301 + ], + "type": "inline_equation", + "content": "S = 20" + }, + { + "bbox": [ + 47, + 253, + 286, + 301 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 308, + 286, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 308, + 286, + 320 + ], + "spans": [ + { + "bbox": [ + 47, + 308, + 286, + 320 + ], + "type": "text", + "content": "3.5. Implementation details and data augmentation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 327, + 286, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 286, + 458 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 286, + 458 + ], + "type": "text", + "content": "In the user editing phase, we fill the background of the target canvas with pixels randomly drawn from the source texture, such that the encoding of the target image would not deviate too far from the source in the SD latent space. As the self-rectification is performed twice, the output of coarse stage " + }, + { + "bbox": [ + 46, + 327, + 286, + 458 + ], + "type": "inline_equation", + "content": "I_{\\text{coarse}}^*" + }, + { + "bbox": [ + 46, + 327, + 286, + 458 + ], + "type": "text", + "content": ", will be used as the input target image of fine stage to produce the final output " + }, + { + "bbox": [ + 46, + 327, + 286, + 458 + ], + "type": "inline_equation", + "content": "I^*" + }, + { + "bbox": [ + 46, + 327, + 286, + 458 + ], + "type": "text", + "content": ". In contrast, for the inversion reference " + }, + { + "bbox": [ + 46, + 327, + 286, + 458 + ], + "type": "inline_equation", + "content": "I^{IR}" + }, + { + "bbox": [ + 46, + 327, + 286, + 458 + ], + "type": "text", + "content": " involved in structure-preserving inversion of the two self-rectifications, we use the same initial target image that contains the original user edits. See the supplementary for the full algorithm pseudo-code." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 459, + 286, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 459, + 286, + 566 + ], + "spans": [ + { + "bbox": [ + 46, + 459, + 286, + 566 + ], + "type": "text", + "content": "Considering the parameter settings: since the coarse stage aims for structure synthesis, relatively large " + }, + { + "bbox": [ + 46, + 459, + 286, + 566 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 46, + 459, + 286, + 566 + ], + "type": "text", + "content": " and S are required in self-rectification, and vice versa in the fine stage. More specifically, let " + }, + { + "bbox": [ + 46, + 459, + 286, + 566 + ], + "type": "inline_equation", + "content": "\\mathrm{P_1}" + }, + { + "bbox": [ + 46, + 459, + 286, + 566 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 459, + 286, + 566 + ], + "type": "inline_equation", + "content": "\\mathrm{P_2}" + }, + { + "bbox": [ + 46, + 459, + 286, + 566 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 459, + 286, + 566 + ], + "type": "inline_equation", + "content": "\\mathrm{S_1}" + }, + { + "bbox": [ + 46, + 459, + 286, + 566 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 459, + 286, + 566 + ], + "type": "inline_equation", + "content": "\\mathrm{S_2}" + }, + { + "bbox": [ + 46, + 459, + 286, + 566 + ], + "type": "text", + "content": ", denote the parameters used in the two rounds of structure-preserving inversion and fine texture sampling. We typically set " + }, + { + "bbox": [ + 46, + 459, + 286, + 566 + ], + "type": "inline_equation", + "content": "\\mathrm{P_1} = 20" + }, + { + "bbox": [ + 46, + 459, + 286, + 566 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 459, + 286, + 566 + ], + "type": "inline_equation", + "content": "\\mathrm{P_2} = 5" + }, + { + "bbox": [ + 46, + 459, + 286, + 566 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 459, + 286, + 566 + ], + "type": "inline_equation", + "content": "\\mathrm{S_1} = 20" + }, + { + "bbox": [ + 46, + 459, + 286, + 566 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 459, + 286, + 566 + ], + "type": "inline_equation", + "content": "\\mathrm{S_2} = 5" + }, + { + "bbox": [ + 46, + 459, + 286, + 566 + ], + "type": "text", + "content": ". Following [6], we choose the " + }, + { + "bbox": [ + 46, + 459, + 286, + 566 + ], + "type": "inline_equation", + "content": "KV" + }, + { + "bbox": [ + 46, + 459, + 286, + 566 + ], + "type": "text", + "content": " features from the 10th to 15th self-attention layers of the U-Net decoder part." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 566, + 286, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 566, + 286, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 566, + 286, + 662 + ], + "type": "text", + "content": "To further improve the synthesis quality, especially when dealing with textures containing a dominant directional structure, such as the leaf shown in Fig. 1, we can introduce a few transformed images (flips and rotations) to augment the reference texture. We concatenate the new attention features from the augmentation to the original reference feature. For example, when we have " + }, + { + "bbox": [ + 46, + 566, + 286, + 662 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 566, + 286, + 662 + ], + "type": "text", + "content": " augmented source textures, the " + }, + { + "bbox": [ + 46, + 566, + 286, + 662 + ], + "type": "inline_equation", + "content": "K^R" + }, + { + "bbox": [ + 46, + 566, + 286, + 662 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 566, + 286, + 662 + ], + "type": "inline_equation", + "content": "V^R" + }, + { + "bbox": [ + 46, + 566, + 286, + 662 + ], + "type": "text", + "content": " in Eq. (6) is now given by" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 77, + 670, + 286, + 709 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 670, + 286, + 709 + ], + "spans": [ + { + "bbox": [ + 77, + 670, + 286, + 709 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l l} K ^ {R} & = \\operatorname {C o n c a t} \\left(K _ {(0)} ^ {R}, K _ {(1)} ^ {R}, \\dots , K _ {(n)} ^ {R}\\right), \\\\ V ^ {R} & = \\operatorname {C o n c a t} \\left(V _ {(0)} ^ {R}, V _ {(1)} ^ {R}, \\dots , V _ {(n)} ^ {R}\\right) \\end{array} \\right. \\tag {7}", + "image_path": "bdb619e087e30421947278d9d4974864fe8fcd6ed6a2140618ce187eb335664a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 96 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "K_{(i)}^{R}" + }, + { + "bbox": [ + 306, + 72, + 545, + 96 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 306, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "V_{(i)}^{R}" + }, + { + "bbox": [ + 306, + 72, + 545, + 96 + ], + "type": "text", + "content": " are the self-attention features acquired from the DDIM inversion of the augmented references." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 105, + 386, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 105, + 386, + 118 + ], + "spans": [ + { + "bbox": [ + 306, + 105, + 386, + 118 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 125, + 545, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 125, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 305, + 125, + 545, + 185 + ], + "type": "text", + "content": "We apply our method with Stable Diffusion with publicly available checkpoints v1.4. All experiments were conducted on a single Quadro P6000 24G GPU. The inference time synthesizing an image of " + }, + { + "bbox": [ + 305, + 125, + 545, + 185 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 305, + 125, + 545, + 185 + ], + "type": "text", + "content": " pixels takes about three minutes for our coarse-to-fine self-rectification." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 191, + 469, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 191, + 469, + 204 + ], + "spans": [ + { + "bbox": [ + 306, + 191, + 469, + 204 + ], + "type": "text", + "content": "4.1. Evaluations and Comparisons" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 209, + 545, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 209, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 305, + 209, + 545, + 317 + ], + "type": "text", + "content": "We evaluate our method with non-stationary textures released by [51]. Each example is resized to " + }, + { + "bbox": [ + 305, + 209, + 545, + 317 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 305, + 209, + 545, + 317 + ], + "type": "text", + "content": " pixels as the reference, and we quickly built several different target images for each in PhotoShop with just a few lazy edits. Fig. 6 shows a gallery of results generated by our method. As can be seen our method faithfully reproduces the delicate textures of the exemplar, their global structure and yet respecting the sparse edits of the target image. More results are included in the supplementary." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 317, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 317, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 317, + 545, + 449 + ], + "type": "text", + "content": "To compare with state-of-the-art texture synthesis methods, we fed the target images we sketched to the models trained by adversarial expansion (TexExp) [51]. As can be seen in Fig. 6, TexExp failed to reproduce the fine textures of the reference, as the edited target images are unseen data to its training. We also tested the texture optimization based on a recently proposed textural loss (GCD Loss) [49], where the target images are down-scaled as the initialization in its multi-resolution synthesis. Although plausible local textures are synthesized, the output of this method does not always conform to the user-edited layout (see Fig. 6)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 449, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 556 + ], + "type": "text", + "content": "We have also applied our method on nearly homogeneous textures of stationary statistics. Since such textures do not have a prominent global structure, we can simply reshuffle patches of the reference to serve as the target. This, however, might cut some local elements in the example. Alternatively, we can use the reference image as the target and shuffle its inversion code before the sampling, which better preserves the local texture elements. Fig. 7 shows a few examples, demonstrating promising quality." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 562, + 406, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 562, + 406, + 573 + ], + "spans": [ + { + "bbox": [ + 306, + 562, + 406, + 573 + ], + "type": "text", + "content": "4.2. Ablation Studies" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 580, + 544, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 580, + 544, + 604 + ], + "spans": [ + { + "bbox": [ + 306, + 580, + 544, + 604 + ], + "type": "text", + "content": "In this section, we analyze the effect of the key components in our method through ablation studies." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "text", + "content": "KV-Injection in sampling. As " + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "inline_equation", + "content": "KV" + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "text", + "content": "-Injection is applied both in the inversion and sampling process, we applied two ablations to study its effect. First, we set " + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "inline_equation", + "content": "P_1 = P_2 = 0" + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "text", + "content": ", and explore the full parameter space for " + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "inline_equation", + "content": "S_1" + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "inline_equation", + "content": "S_2" + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "text", + "content": ", to study the effect of " + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "inline_equation", + "content": "KV" + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "text", + "content": "-Injection in texture sampling without being affected by the inversion." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Fig. 8 shows a matrix of results of this parameter study. A smaller value of " + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "S(S_1 \\& S_2)" + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": " means performing more time" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "7771" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 69, + 547, + 657 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 547, + 657 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 547, + 657 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 547, + 657 + ], + "type": "image", + "image_path": "50443225708289e321da932961981303861c63ecd1f051c6a43b3fb8eb60797c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 662, + 548, + 718 + ], + "lines": [ + { + "bbox": [ + 46, + 662, + 548, + 718 + ], + "spans": [ + { + "bbox": [ + 46, + 662, + 548, + 718 + ], + "type": "text", + "content": "Figure 6. We sketch two targets for each reference, denoted as Edit 1 and Edit 2, respectively. Our self-rectification method generates textures with global structures that faithfully respect the user edits of the target images, while still producing high-quality texture details. In contrast, adversarial expansion (TexExp) [51] does not capture the fine details well. Optimization by GCD Loss [49] reproduces the local textures, but does not always conform to the target image. Note that Edit 2 is used as the input for TexExp, and as the initialization in texture optimization of GCD Loss. More results are included in the supplementary." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "text", + "content": "7772" + } + ] + } + ], + "index": 2 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 127, + 147 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 127, + 147 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 127, + 147 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 127, + 147 + ], + "type": "image", + "image_path": "afa97938b590bf80a91bd6da85eb186cbf05557a2080781c66cc2e17c52e75d7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 129, + 70, + 205, + 147 + ], + "blocks": [ + { + "bbox": [ + 129, + 70, + 205, + 147 + ], + "lines": [ + { + "bbox": [ + 129, + 70, + 205, + 147 + ], + "spans": [ + { + "bbox": [ + 129, + 70, + 205, + 147 + ], + "type": "image", + "image_path": "0a9ba524973edd6fb6309680a5747ea336d91b95af73f2f36411686b4554a6bf.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 206, + 70, + 284, + 147 + ], + "blocks": [ + { + "bbox": [ + 206, + 70, + 284, + 147 + ], + "lines": [ + { + "bbox": [ + 206, + 70, + 284, + 147 + ], + "spans": [ + { + "bbox": [ + 206, + 70, + 284, + 147 + ], + "type": "image", + "image_path": "6262c145e78d74b2d1193ab93ee846d9a1bcaad0eb99fa55b938eb066eae3825.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 50, + 148, + 127, + 224 + ], + "blocks": [ + { + "bbox": [ + 50, + 148, + 127, + 224 + ], + "lines": [ + { + "bbox": [ + 50, + 148, + 127, + 224 + ], + "spans": [ + { + "bbox": [ + 50, + 148, + 127, + 224 + ], + "type": "image", + "image_path": "02ec0269d9b4bc0ee85f020804a3266a5ac00a135ae7355947be5a8d772d729b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 129, + 148, + 205, + 224 + ], + "blocks": [ + { + "bbox": [ + 129, + 148, + 205, + 224 + ], + "lines": [ + { + "bbox": [ + 129, + 148, + 205, + 224 + ], + "spans": [ + { + "bbox": [ + 129, + 148, + 205, + 224 + ], + "type": "image", + "image_path": "7125e9d1284caf07b46a5a3842a719d263cc59995ec3dcddc22c48a5a7811bc8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 206, + 148, + 284, + 224 + ], + "blocks": [ + { + "bbox": [ + 206, + 148, + 284, + 224 + ], + "lines": [ + { + "bbox": [ + 206, + 148, + 284, + 224 + ], + "spans": [ + { + "bbox": [ + 206, + 148, + 284, + 224 + ], + "type": "image", + "image_path": "aeca6e4d5c72461cc0b796a32f4ad2aeb6f264c51ccdf1b543c33b99c136b7b6.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 50, + 225, + 127, + 301 + ], + "blocks": [ + { + "bbox": [ + 50, + 225, + 127, + 301 + ], + "lines": [ + { + "bbox": [ + 50, + 225, + 127, + 301 + ], + "spans": [ + { + "bbox": [ + 50, + 225, + 127, + 301 + ], + "type": "image", + "image_path": "bb90f141a104f75f975b91e18f326a429a4612cd44b9ebed7f4734109429fe5f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 310, + 287, + 365 + ], + "lines": [ + { + "bbox": [ + 46, + 310, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 310, + 287, + 365 + ], + "type": "text", + "content": "Figure 7. For nearly homogeneous textures (left), we can use patch shuffle of the reference to define a random target layout, where the shuffling could be performed in image space before applying self-rectification (middle), or in latent space after the inversion in the first self-rectification (right)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 129, + 225, + 205, + 301 + ], + "blocks": [ + { + "bbox": [ + 129, + 225, + 205, + 301 + ], + "lines": [ + { + "bbox": [ + 129, + 225, + 205, + 301 + ], + "spans": [ + { + "bbox": [ + 129, + 225, + 205, + 301 + ], + "type": "image", + "image_path": "d353d345b76b288015cfcbec2a1caaa096881502f9f0d10d51185a8352152b19.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 206, + 225, + 284, + 301 + ], + "blocks": [ + { + "bbox": [ + 206, + 225, + 284, + 301 + ], + "lines": [ + { + "bbox": [ + 206, + 225, + 284, + 301 + ], + "spans": [ + { + "bbox": [ + 206, + 225, + 284, + 301 + ], + "type": "image", + "image_path": "1ef9999145be44a9ed679f675ddd306e26a270aa413c91cc046f0fd99d592e0a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 376, + 287, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 287, + 509 + ], + "type": "text", + "content": "steps of " + }, + { + "bbox": [ + 46, + 376, + 287, + 509 + ], + "type": "inline_equation", + "content": "KV" + }, + { + "bbox": [ + 46, + 376, + 287, + 509 + ], + "type": "text", + "content": "-Injection in sampling, and thus, the target image is rectified to be closer to the reference texture. On the contrary, a larger S reserves more of the target image's layout, however, at the same time, it introduces more structural errors or conflicts, yielding artifacts. None of these results reaches a good trade-off between synthesizing a reasonable global layout (especially considering the user edits) and reproducing local textures of the reference. Nevertheless, we can find a proper parameter setting for S, which is " + }, + { + "bbox": [ + 46, + 376, + 287, + 509 + ], + "type": "inline_equation", + "content": "10\\sim 20" + }, + { + "bbox": [ + 46, + 376, + 287, + 509 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 376, + 287, + 509 + ], + "type": "inline_equation", + "content": "S_1" + }, + { + "bbox": [ + 46, + 376, + 287, + 509 + ], + "type": "text", + "content": ", and 5 for " + }, + { + "bbox": [ + 46, + 376, + 287, + 509 + ], + "type": "inline_equation", + "content": "S_2" + }, + { + "bbox": [ + 46, + 376, + 287, + 509 + ], + "type": "text", + "content": ". Hence we set the default values of " + }, + { + "bbox": [ + 46, + 376, + 287, + 509 + ], + "type": "inline_equation", + "content": "S_1" + }, + { + "bbox": [ + 46, + 376, + 287, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 376, + 287, + 509 + ], + "type": "inline_equation", + "content": "S_2" + }, + { + "bbox": [ + 46, + 376, + 287, + 509 + ], + "type": "text", + "content": " to be 20 and 5, and use it for all experiments." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 522, + 286, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 286, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 286, + 666 + ], + "type": "text", + "content": "KV-Injection in inversion. Next, we investigate the effect of " + }, + { + "bbox": [ + 46, + 522, + 286, + 666 + ], + "type": "inline_equation", + "content": "KV" + }, + { + "bbox": [ + 46, + 522, + 286, + 666 + ], + "type": "text", + "content": "-Injection in our structure-preserving inversion. By fixing " + }, + { + "bbox": [ + 46, + 522, + 286, + 666 + ], + "type": "inline_equation", + "content": "S_1" + }, + { + "bbox": [ + 46, + 522, + 286, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 522, + 286, + 666 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 522, + 286, + 666 + ], + "type": "text", + "content": " to default values, we search the parameter space defined by " + }, + { + "bbox": [ + 46, + 522, + 286, + 666 + ], + "type": "inline_equation", + "content": "P_1" + }, + { + "bbox": [ + 46, + 522, + 286, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 522, + 286, + 666 + ], + "type": "inline_equation", + "content": "P_2" + }, + { + "bbox": [ + 46, + 522, + 286, + 666 + ], + "type": "text", + "content": ". As shown in Fig. 9, introducing " + }, + { + "bbox": [ + 46, + 522, + 286, + 666 + ], + "type": "inline_equation", + "content": "KV" + }, + { + "bbox": [ + 46, + 522, + 286, + 666 + ], + "type": "text", + "content": "-Injection significantly improves the synthesis result. Both the structural errors and local artifacts are drastically reduced at the same time. Another important point is that, we may have a relatively wide range for setting the value of " + }, + { + "bbox": [ + 46, + 522, + 286, + 666 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 522, + 286, + 666 + ], + "type": "text", + "content": ", which is empirically suggested by the results: " + }, + { + "bbox": [ + 46, + 522, + 286, + 666 + ], + "type": "inline_equation", + "content": "10\\sim 30" + }, + { + "bbox": [ + 46, + 522, + 286, + 666 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 522, + 286, + 666 + ], + "type": "inline_equation", + "content": "P_1" + }, + { + "bbox": [ + 46, + 522, + 286, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 522, + 286, + 666 + ], + "type": "inline_equation", + "content": "5\\sim 15" + }, + { + "bbox": [ + 46, + 522, + 286, + 666 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 522, + 286, + 666 + ], + "type": "inline_equation", + "content": "P_2" + }, + { + "bbox": [ + 46, + 522, + 286, + 666 + ], + "type": "text", + "content": ". We usually set " + }, + { + "bbox": [ + 46, + 522, + 286, + 666 + ], + "type": "inline_equation", + "content": "P_1" + }, + { + "bbox": [ + 46, + 522, + 286, + 666 + ], + "type": "text", + "content": " as 20, and " + }, + { + "bbox": [ + 46, + 522, + 286, + 666 + ], + "type": "inline_equation", + "content": "P_2" + }, + { + "bbox": [ + 46, + 522, + 286, + 666 + ], + "type": "text", + "content": " as 5 in production. See supplementary for the full exploration results and more examples of this study." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 681, + 286, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 681, + 286, + 704 + ], + "spans": [ + { + "bbox": [ + 46, + 681, + 286, + 704 + ], + "type": "text", + "content": "Data augmentation. In many cases, data augmentation may not be necessary, as the pre-trained Stable Diffusion" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 307, + 68, + 544, + 244 + ], + "blocks": [ + { + "bbox": [ + 307, + 68, + 544, + 244 + ], + "lines": [ + { + "bbox": [ + 307, + 68, + 544, + 244 + ], + "spans": [ + { + "bbox": [ + 307, + 68, + 544, + 244 + ], + "type": "image", + "image_path": "4d8a2ebcb997becd41a36c72c0605fc46b8ea38eb747f06aa66b50c8003da72c.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 251, + 545, + 362 + ], + "lines": [ + { + "bbox": [ + 305, + 251, + 545, + 362 + ], + "spans": [ + { + "bbox": [ + 305, + 251, + 545, + 362 + ], + "type": "text", + "content": "Figure 8. Ablation study of " + }, + { + "bbox": [ + 305, + 251, + 545, + 362 + ], + "type": "inline_equation", + "content": "KV" + }, + { + "bbox": [ + 305, + 251, + 545, + 362 + ], + "type": "text", + "content": "-Injection in texture sampling. We explore the parameter space for " + }, + { + "bbox": [ + 305, + 251, + 545, + 362 + ], + "type": "inline_equation", + "content": "S_1" + }, + { + "bbox": [ + 305, + 251, + 545, + 362 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 251, + 545, + 362 + ], + "type": "inline_equation", + "content": "S_2" + }, + { + "bbox": [ + 305, + 251, + 545, + 362 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 305, + 251, + 545, + 362 + ], + "type": "inline_equation", + "content": "P_1 = P_2 = 0" + }, + { + "bbox": [ + 305, + 251, + 545, + 362 + ], + "type": "text", + "content": ", and adopt the wood texture and its edited target shown in Fig. 1 for this test. The coarse self-rectification results are shown in the 1st column. We can see that changing the starting time-step of " + }, + { + "bbox": [ + 305, + 251, + 545, + 362 + ], + "type": "inline_equation", + "content": "KV" + }, + { + "bbox": [ + 305, + 251, + 545, + 362 + ], + "type": "text", + "content": "-Injection in the sampling greatly affects the final output. However, none of these settings yields a qualified balance when considering both the target layout and the texture details. Note setting " + }, + { + "bbox": [ + 305, + 251, + 545, + 362 + ], + "type": "inline_equation", + "content": "S = 50" + }, + { + "bbox": [ + 305, + 251, + 545, + 362 + ], + "type": "text", + "content": " (i.e., no " + }, + { + "bbox": [ + 305, + 251, + 545, + 362 + ], + "type": "inline_equation", + "content": "KV" + }, + { + "bbox": [ + 305, + 251, + 545, + 362 + ], + "type": "text", + "content": "-Injection in sampling) only results in an approximate layout reconstruction without any details synthesized (see Fig. 5)." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 308, + 380, + 544, + 558 + ], + "blocks": [ + { + "bbox": [ + 308, + 380, + 544, + 558 + ], + "lines": [ + { + "bbox": [ + 308, + 380, + 544, + 558 + ], + "spans": [ + { + "bbox": [ + 308, + 380, + 544, + 558 + ], + "type": "image", + "image_path": "352a9547a500faa95228f3b92b821c363d3421f0e9b368af4968f5bbef670f61.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 565, + 545, + 643 + ], + "lines": [ + { + "bbox": [ + 304, + 565, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 304, + 565, + 545, + 643 + ], + "type": "text", + "content": "Figure 9. Ablation study of " + }, + { + "bbox": [ + 304, + 565, + 545, + 643 + ], + "type": "inline_equation", + "content": "KV" + }, + { + "bbox": [ + 304, + 565, + 545, + 643 + ], + "type": "text", + "content": "-Injection in structure-preserving inversion by exploring the parameter space defined by " + }, + { + "bbox": [ + 304, + 565, + 545, + 643 + ], + "type": "inline_equation", + "content": "\\mathrm{P}_1" + }, + { + "bbox": [ + 304, + 565, + 545, + 643 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 565, + 545, + 643 + ], + "type": "inline_equation", + "content": "\\mathrm{P}_2" + }, + { + "bbox": [ + 304, + 565, + 545, + 643 + ], + "type": "text", + "content": ". Here, we fixed " + }, + { + "bbox": [ + 304, + 565, + 545, + 643 + ], + "type": "inline_equation", + "content": "S_1 = 20" + }, + { + "bbox": [ + 304, + 565, + 545, + 643 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 565, + 545, + 643 + ], + "type": "inline_equation", + "content": "S_2 = 5" + }, + { + "bbox": [ + 304, + 565, + 545, + 643 + ], + "type": "text", + "content": " according to the test of Fig. 8. While compared with Fig. 8, the outputs are significantly improved after introducing " + }, + { + "bbox": [ + 304, + 565, + 545, + 643 + ], + "type": "inline_equation", + "content": "KV" + }, + { + "bbox": [ + 304, + 565, + 545, + 643 + ], + "type": "text", + "content": "-Injection to the inversion. Many of these results can be considered good synthesis regarding the target image edited and the source texture of the wood." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 653, + 545, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 702 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 702 + ], + "type": "text", + "content": "model already has a certain ability to synthesize rotated texture patterns, except when the reference has dominant directional structures. To allow more free editing for directional textures, augmenting the reference is essential to synthesize" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7773" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 108, + 190 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 108, + 190 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 108, + 190 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 108, + 190 + ], + "type": "image", + "image_path": "a8afb72b0f76985592e857289bc6bff2d390b8120495bfd1888560005d2ae3d0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 192, + 97, + 201 + ], + "lines": [ + { + "bbox": [ + 59, + 192, + 97, + 201 + ], + "spans": [ + { + "bbox": [ + 59, + 192, + 97, + 201 + ], + "type": "text", + "content": "Reference" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 109, + 70, + 166, + 187 + ], + "blocks": [ + { + "bbox": [ + 109, + 70, + 166, + 187 + ], + "lines": [ + { + "bbox": [ + 109, + 70, + 166, + 187 + ], + "spans": [ + { + "bbox": [ + 109, + 70, + 166, + 187 + ], + "type": "image", + "image_path": "b056391f41783179be12d7aaa5e5997cdd8d84f9547dba93611e99972aa3fe2c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 126, + 192, + 149, + 201 + ], + "lines": [ + { + "bbox": [ + 126, + 192, + 149, + 201 + ], + "spans": [ + { + "bbox": [ + 126, + 192, + 149, + 201 + ], + "type": "text", + "content": "Target" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 167, + 70, + 224, + 187 + ], + "blocks": [ + { + "bbox": [ + 167, + 70, + 224, + 187 + ], + "lines": [ + { + "bbox": [ + 167, + 70, + 224, + 187 + ], + "spans": [ + { + "bbox": [ + 167, + 70, + 224, + 187 + ], + "type": "image", + "image_path": "c3e1520d8318dfa1d4da1127df86f99f02c5a9cced4d0ce74a97da3cc89bf549.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 171, + 188, + 221, + 206 + ], + "lines": [ + { + "bbox": [ + 171, + 188, + 221, + 206 + ], + "spans": [ + { + "bbox": [ + 171, + 188, + 221, + 206 + ], + "type": "text", + "content": "w/o Data Augmentation" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 225, + 70, + 283, + 187 + ], + "blocks": [ + { + "bbox": [ + 225, + 70, + 283, + 187 + ], + "lines": [ + { + "bbox": [ + 225, + 70, + 283, + 187 + ], + "spans": [ + { + "bbox": [ + 225, + 70, + 283, + 187 + ], + "type": "image", + "image_path": "2ea5f75c9435532f64f7830caa19a3197ca0ebcdcabff56c74cad9f4bce03490.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 229, + 188, + 279, + 206 + ], + "lines": [ + { + "bbox": [ + 229, + 188, + 279, + 206 + ], + "spans": [ + { + "bbox": [ + 229, + 188, + 279, + 206 + ], + "type": "text", + "content": "w/ Data Augmentation" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 51, + 281, + 127, + 356 + ], + "blocks": [ + { + "bbox": [ + 47, + 215, + 287, + 259 + ], + "lines": [ + { + "bbox": [ + 47, + 215, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 47, + 215, + 287, + 259 + ], + "type": "text", + "content": "Figure 10. For textures that contain dominant directional structures, data augmentation is essential to allow the output to admit with user edits. Specifically, we augment each source shown above with three images rotated at angles of " + }, + { + "bbox": [ + 47, + 215, + 287, + 259 + ], + "type": "inline_equation", + "content": "\\pm 45^{\\circ}" + }, + { + "bbox": [ + 47, + 215, + 287, + 259 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 215, + 287, + 259 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 47, + 215, + 287, + 259 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 51, + 281, + 127, + 356 + ], + "lines": [ + { + "bbox": [ + 51, + 281, + 127, + 356 + ], + "spans": [ + { + "bbox": [ + 51, + 281, + 127, + 356 + ], + "type": "image", + "image_path": "534b06e54f438892c9a367c5689f2083283f9dc4d17b14457ede7b9d1004e4c2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 51, + 357, + 127, + 434 + ], + "blocks": [ + { + "bbox": [ + 51, + 357, + 127, + 434 + ], + "lines": [ + { + "bbox": [ + 51, + 357, + 127, + 434 + ], + "spans": [ + { + "bbox": [ + 51, + 357, + 127, + 434 + ], + "type": "image", + "image_path": "ab498a1a38151e4a3f98d3e6cb218d4f991081604ca797bf058204f5942fbb60.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 437, + 109, + 445 + ], + "lines": [ + { + "bbox": [ + 69, + 437, + 109, + 445 + ], + "spans": [ + { + "bbox": [ + 69, + 437, + 109, + 445 + ], + "type": "text", + "content": "Reference" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 455, + 287, + 487 + ], + "lines": [ + { + "bbox": [ + 47, + 455, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 47, + 455, + 287, + 487 + ], + "type": "text", + "content": "Figure 11. Guided texture synthesis. The user provides a color layout to guide the output structure. The synthesized outputs follow the outlines well and still exhibit the reference texture details." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 129, + 281, + 205, + 357 + ], + "blocks": [ + { + "bbox": [ + 129, + 281, + 205, + 357 + ], + "lines": [ + { + "bbox": [ + 129, + 281, + 205, + 357 + ], + "spans": [ + { + "bbox": [ + 129, + 281, + 205, + 357 + ], + "type": "image", + "image_path": "cef84d6cc8c21d4b4e7f5d3e2c4bccdd9c409b51b787346a752f7f01e1581848.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 129, + 358, + 205, + 434 + ], + "blocks": [ + { + "bbox": [ + 129, + 358, + 205, + 434 + ], + "lines": [ + { + "bbox": [ + 129, + 358, + 205, + 434 + ], + "spans": [ + { + "bbox": [ + 129, + 358, + 205, + 434 + ], + "type": "image", + "image_path": "bbb1d67a42a9f0e99f31263fa7e7f86afe82fb3e4c5cb23991c303673ba44a92.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 437, + 197, + 447 + ], + "lines": [ + { + "bbox": [ + 138, + 437, + 197, + 447 + ], + "spans": [ + { + "bbox": [ + 138, + 437, + 197, + 447 + ], + "type": "text", + "content": "Target Outlines" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 208, + 281, + 283, + 357 + ], + "blocks": [ + { + "bbox": [ + 208, + 281, + 283, + 357 + ], + "lines": [ + { + "bbox": [ + 208, + 281, + 283, + 357 + ], + "spans": [ + { + "bbox": [ + 208, + 281, + 283, + 357 + ], + "type": "image", + "image_path": "4324871efdf50b5f78be6a9811c3b74f3d7adf1af75b4db27673b10bd48d3710.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 276, + 545, + 309 + ], + "lines": [ + { + "bbox": [ + 306, + 276, + 545, + 309 + ], + "spans": [ + { + "bbox": [ + 306, + 276, + 545, + 309 + ], + "type": "text", + "content": "Figure 12. Our method can also be applied to image editing. For creating the target image, the user can either loosely sketch a layout on the target canvas, or directly edit the input image." + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 208, + 358, + 283, + 434 + ], + "blocks": [ + { + "bbox": [ + 208, + 358, + 283, + 434 + ], + "lines": [ + { + "bbox": [ + 208, + 358, + 283, + 434 + ], + "spans": [ + { + "bbox": [ + 208, + 358, + 283, + 434 + ], + "type": "image", + "image_path": "06471c22acbe139dc0dc719dc5b4b4e45ede3620947529d5f335f7300c602d1d.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 232, + 437, + 260, + 447 + ], + "lines": [ + { + "bbox": [ + 232, + 437, + 260, + 447 + ], + "spans": [ + { + "bbox": [ + 232, + 437, + 260, + 447 + ], + "type": "text", + "content": "Output" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 47, + 499, + 287, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 499, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 47, + 499, + 287, + 535 + ], + "type": "text", + "content": "a more consistent output structure conforming to user edits; see, e.g., Fig. 10 for a comparison on data augmentation, where additional rotated transformations are involved." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 47, + 543, + 181, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 543, + 181, + 555 + ], + "spans": [ + { + "bbox": [ + 47, + 543, + 181, + 555 + ], + "type": "text", + "content": "4.3. Additional Applications" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 561, + 287, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 561, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 46, + 561, + 287, + 668 + ], + "type": "text", + "content": "Our self-rectification framework can extend its utility to seemingly other applications. For instance, as illustrated in Fig. 11, users can input a straightforward target image outlining a layout structure with colors compatible with reference textures. To avoid excessive smoothing, random noise is introduced to enhance the target image. The self-rectification process ensures that the target image incorporates texture details from the reference, while adhering to the user-provided layout structure." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 669, + 287, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 669, + 287, + 705 + ], + "spans": [ + { + "bbox": [ + 46, + 669, + 287, + 705 + ], + "type": "text", + "content": "Moreover, our self-rectification method can be employed for image editing beyond textures. Users can edit a given image by employing basic cut-and-paste patch operations," + } + ] + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 310, + 70, + 355, + 116 + ], + "blocks": [ + { + "bbox": [ + 310, + 70, + 355, + 116 + ], + "lines": [ + { + "bbox": [ + 310, + 70, + 355, + 116 + ], + "spans": [ + { + "bbox": [ + 310, + 70, + 355, + 116 + ], + "type": "image", + "image_path": "9a51b9643c751b762b060d764ba814c86273b026c1b295514fde0c844cb5563b.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 357, + 70, + 402, + 116 + ], + "blocks": [ + { + "bbox": [ + 357, + 70, + 402, + 116 + ], + "lines": [ + { + "bbox": [ + 357, + 70, + 402, + 116 + ], + "spans": [ + { + "bbox": [ + 357, + 70, + 402, + 116 + ], + "type": "image", + "image_path": "2afe6df694be01756c8240f9179ef1299e06fc9a54447116826c51131ba8ed4d.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 403, + 70, + 448, + 116 + ], + "blocks": [ + { + "bbox": [ + 403, + 70, + 448, + 116 + ], + "lines": [ + { + "bbox": [ + 403, + 70, + 448, + 116 + ], + "spans": [ + { + "bbox": [ + 403, + 70, + 448, + 116 + ], + "type": "image", + "image_path": "bb60ddbaabe7c74cc98ee1aef82a5e5c78915c69a57f8e45a5d03df27cfb67ca.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 450, + 70, + 495, + 116 + ], + "blocks": [ + { + "bbox": [ + 450, + 70, + 495, + 116 + ], + "lines": [ + { + "bbox": [ + 450, + 70, + 495, + 116 + ], + "spans": [ + { + "bbox": [ + 450, + 70, + 495, + 116 + ], + "type": "image", + "image_path": "e238aa7d944ed2b4b557745d2a3afc45980c961ba1e64c2ee0664362d2aa95da.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 497, + 70, + 542, + 116 + ], + "blocks": [ + { + "bbox": [ + 497, + 70, + 542, + 116 + ], + "lines": [ + { + "bbox": [ + 497, + 70, + 542, + 116 + ], + "spans": [ + { + "bbox": [ + 497, + 70, + 542, + 116 + ], + "type": "image", + "image_path": "358d3490f70a74e6db88a679d7ae26e861438c5d9a7f5b69114a375982530850.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 310, + 118, + 355, + 163 + ], + "blocks": [ + { + "bbox": [ + 310, + 118, + 355, + 163 + ], + "lines": [ + { + "bbox": [ + 310, + 118, + 355, + 163 + ], + "spans": [ + { + "bbox": [ + 310, + 118, + 355, + 163 + ], + "type": "image", + "image_path": "c8db08770fe2c85e968098a3a470d6eb5d60cecfa91569b505fc977e009f47a8.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 356, + 118, + 402, + 163 + ], + "blocks": [ + { + "bbox": [ + 356, + 118, + 402, + 163 + ], + "lines": [ + { + "bbox": [ + 356, + 118, + 402, + 163 + ], + "spans": [ + { + "bbox": [ + 356, + 118, + 402, + 163 + ], + "type": "image", + "image_path": "1a2bf52c8fe4e57ed4c58a4e6f202e86b891917e6d28a46581ea78904ac0e228.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 403, + 118, + 448, + 163 + ], + "blocks": [ + { + "bbox": [ + 403, + 118, + 448, + 163 + ], + "lines": [ + { + "bbox": [ + 403, + 118, + 448, + 163 + ], + "spans": [ + { + "bbox": [ + 403, + 118, + 448, + 163 + ], + "type": "image", + "image_path": "bce3be17612a92ddcb44b9eb0a44057bec2c173e9aa516d57bf3469a1a74eebb.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 450, + 118, + 495, + 163 + ], + "blocks": [ + { + "bbox": [ + 450, + 118, + 495, + 163 + ], + "lines": [ + { + "bbox": [ + 450, + 118, + 495, + 163 + ], + "spans": [ + { + "bbox": [ + 450, + 118, + 495, + 163 + ], + "type": "image", + "image_path": "02fdb624e503a3f6239afcb3869c7113f5922049123406920f0646bb4a821e58.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 497, + 118, + 542, + 163 + ], + "blocks": [ + { + "bbox": [ + 497, + 118, + 542, + 163 + ], + "lines": [ + { + "bbox": [ + 497, + 118, + 542, + 163 + ], + "spans": [ + { + "bbox": [ + 497, + 118, + 542, + 163 + ], + "type": "image", + "image_path": "f8e5454382b1077ea07caa3c2c89ed6a65473f5abd983a49c60dcd2efe309a94.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 310, + 164, + 355, + 209 + ], + "blocks": [ + { + "bbox": [ + 310, + 164, + 355, + 209 + ], + "lines": [ + { + "bbox": [ + 310, + 164, + 355, + 209 + ], + "spans": [ + { + "bbox": [ + 310, + 164, + 355, + 209 + ], + "type": "image", + "image_path": "d9c25b0cb2bd4c3eaf2895928ea8b6226aaf0fa49bb391ae06d63b1d7dfa6880.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 356, + 164, + 402, + 209 + ], + "blocks": [ + { + "bbox": [ + 356, + 164, + 402, + 209 + ], + "lines": [ + { + "bbox": [ + 356, + 164, + 402, + 209 + ], + "spans": [ + { + "bbox": [ + 356, + 164, + 402, + 209 + ], + "type": "image", + "image_path": "a8fbc5542a9f94fc36d1c83a53229abd1ae4ca0414701b57e45bb95fe341e740.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 403, + 164, + 448, + 209 + ], + "blocks": [ + { + "bbox": [ + 403, + 164, + 448, + 209 + ], + "lines": [ + { + "bbox": [ + 403, + 164, + 448, + 209 + ], + "spans": [ + { + "bbox": [ + 403, + 164, + 448, + 209 + ], + "type": "image", + "image_path": "7d4c643c70964ad9fb443f797913791934fe65944859c739a27d3e5dc935cba4.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 450, + 164, + 495, + 209 + ], + "blocks": [ + { + "bbox": [ + 450, + 164, + 495, + 209 + ], + "lines": [ + { + "bbox": [ + 450, + 164, + 495, + 209 + ], + "spans": [ + { + "bbox": [ + 450, + 164, + 495, + 209 + ], + "type": "image", + "image_path": "41a1d0e9e1e4219f43ae2316918a889385ff34160e32da01b88fba46be5b031a.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 497, + 164, + 542, + 209 + ], + "blocks": [ + { + "bbox": [ + 497, + 164, + 542, + 209 + ], + "lines": [ + { + "bbox": [ + 497, + 164, + 542, + 209 + ], + "spans": [ + { + "bbox": [ + 497, + 164, + 542, + 209 + ], + "type": "image", + "image_path": "f9b59d1177084526d5b05a5dc55d39b0c9a5130f40d2fa80d6f6428d4232e83a.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 310, + 210, + 355, + 255 + ], + "blocks": [ + { + "bbox": [ + 310, + 210, + 355, + 255 + ], + "lines": [ + { + "bbox": [ + 310, + 210, + 355, + 255 + ], + "spans": [ + { + "bbox": [ + 310, + 210, + 355, + 255 + ], + "type": "image", + "image_path": "e4ef03c3ec70425ee66ca0d667d057b4fad343a038f1abb8090a2fc4619325c7.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 258, + 343, + 267 + ], + "lines": [ + { + "bbox": [ + 323, + 258, + 343, + 267 + ], + "spans": [ + { + "bbox": [ + 323, + 258, + 343, + 267 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 356, + 210, + 402, + 255 + ], + "blocks": [ + { + "bbox": [ + 356, + 210, + 402, + 255 + ], + "lines": [ + { + "bbox": [ + 356, + 210, + 402, + 255 + ], + "spans": [ + { + "bbox": [ + 356, + 210, + 402, + 255 + ], + "type": "image", + "image_path": "3f45de8b56a1e0741dc611067adf2cc1110f9df159230430980544df39218e37.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 369, + 258, + 390, + 267 + ], + "lines": [ + { + "bbox": [ + 369, + 258, + 390, + 267 + ], + "spans": [ + { + "bbox": [ + 369, + 258, + 390, + 267 + ], + "type": "text", + "content": "Edit 1" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 403, + 210, + 448, + 255 + ], + "blocks": [ + { + "bbox": [ + 403, + 210, + 448, + 255 + ], + "lines": [ + { + "bbox": [ + 403, + 210, + 448, + 255 + ], + "spans": [ + { + "bbox": [ + 403, + 210, + 448, + 255 + ], + "type": "image", + "image_path": "de5366bb400aab8215eeb4df37e15d0c6ea45df78f6bde3f504aead4f5eefedd.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 412, + 258, + 445, + 267 + ], + "lines": [ + { + "bbox": [ + 412, + 258, + 445, + 267 + ], + "spans": [ + { + "bbox": [ + 412, + 258, + 445, + 267 + ], + "type": "text", + "content": "Output 1" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 450, + 210, + 495, + 255 + ], + "blocks": [ + { + "bbox": [ + 450, + 210, + 495, + 255 + ], + "lines": [ + { + "bbox": [ + 450, + 210, + 495, + 255 + ], + "spans": [ + { + "bbox": [ + 450, + 210, + 495, + 255 + ], + "type": "image", + "image_path": "bdf22ee1553340991d3b29f835b78f6471b1bf0fbff8e69618fb910e3df95a9e.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 463, + 258, + 487, + 266 + ], + "lines": [ + { + "bbox": [ + 463, + 258, + 487, + 266 + ], + "spans": [ + { + "bbox": [ + 463, + 258, + 487, + 266 + ], + "type": "text", + "content": "Edit 2" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 497, + 210, + 542, + 255 + ], + "blocks": [ + { + "bbox": [ + 497, + 210, + 542, + 255 + ], + "lines": [ + { + "bbox": [ + 497, + 210, + 542, + 255 + ], + "spans": [ + { + "bbox": [ + 497, + 210, + 542, + 255 + ], + "type": "image", + "image_path": "ffc370ac281e528725a1f8b668f16a5d44e25a067add851875346c00a5590670.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 504, + 258, + 538, + 266 + ], + "lines": [ + { + "bbox": [ + 504, + 258, + 538, + 266 + ], + "spans": [ + { + "bbox": [ + 504, + 258, + 538, + 266 + ], + "type": "text", + "content": "Output 2" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + } + ], + "index": 42 + }, + { + "bbox": [ + 306, + 320, + 545, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 320, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 306, + 320, + 545, + 357 + ], + "type": "text", + "content": "and the resulting crude edits undergo self-rectification. Fig. 12 presents several instances of such image editing operations, presenting two edits for each input image." + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 306, + 367, + 383, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 367, + 383, + 380 + ], + "spans": [ + { + "bbox": [ + 306, + 367, + 383, + 380 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 304, + 387, + 545, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 387, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 387, + 545, + 518 + ], + "type": "text", + "content": "Our work addresses the intricate challenge of synthesizing non-stationary textures, offering a method that empowers users to efficiently design new textures with unprecedented controllability. This stands as a notable improvement over existing methods, providing a user-friendly process, which consists of two steps: Users begin with an initial rough edit using conventional image editing tools, followed by an automated self-rectification process. This process leverages a pre-trained diffusion network and injections of self-attention features, showcasing flexibility in synthesizing a diverse range of challenging non-stationary textures." + } + ] + } + ], + "index": 51 + }, + { + "bbox": [ + 304, + 519, + 545, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 519, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 519, + 545, + 602 + ], + "type": "text", + "content": "In the future, we would like to explore the extension of our approach to synthesize large-scale textures. Additionally, there is a promising avenue to incorporate semantic understanding into the self-rectification process, enhancing alignment with user intent. The integration of texture semantics holds the potential to yield contextually relevant and visually appealing synthesized textures." + } + ] + } + ], + "index": 52 + }, + { + "bbox": [ + 304, + 617, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgements This work was supported in parts by NSFC (62161146005, U21B2023, U2001206), GD NSF (2022A1515010221), DEGP Innovation Team (2022 KCXTD025), Shenzhen Science and Technology Program (KQTD20210811090044003, RCJC20200714114435012), ISF (3441/21), Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ) and Scientific Development Funds of Shenzhen University." + } + ] + } + ], + "index": 53 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7774" + } + ] + } + ], + "index": 54 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "text", + "content": "[1] Yuval Alaluf, Daniel Garibi, Or Patashnik, Hadar Averbuch-Elor, and Daniel Cohen-Or. Cross-image attention for zero-shot appearance transfer, 2023. 2, 3, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 286, + 157 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 286, + 157 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 286, + 157 + ], + "type": "text", + "content": "[2] Yuval Alaluf, Elad Richardson, Gal Metzer, and Daniel Cohen-Or. A neural space-time representation for text-to-image personalization, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 158, + 286, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 158, + 286, + 191 + ], + "spans": [ + { + "bbox": [ + 54, + 158, + 286, + 191 + ], + "type": "text", + "content": "[3] Rachele Bellini, Yanir Kleiman, and Daniel Cohen-Or. Time-varying weathering in texture space. ACM Trans. on Graphics (Proc. of SIGGRAPH), 35(4):1-11, 2016. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 191, + 286, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 191, + 286, + 224 + ], + "spans": [ + { + "bbox": [ + 54, + 191, + 286, + 224 + ], + "type": "text", + "content": "[4] Sagie Benaim, Ron Mokady, Amit Bermano, and Lior Wolf. Structural analogy from a single image pair. Computer Graphics Forum, 40(1):249-265, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 225, + 286, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 225, + 286, + 267 + ], + "spans": [ + { + "bbox": [ + 54, + 225, + 286, + 267 + ], + "type": "text", + "content": "[5] Urs Bergmann, Nikolay Jetchev, and Roland Vollgraf. Learning texture manifolds with the periodic spatial gan. In Proc. of IEEE Int. Conf. on Machine Learning, page 469-477, 2017. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 270, + 286, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 270, + 286, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 270, + 286, + 323 + ], + "type": "text", + "content": "[6] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. MasaCtrl: tuning-free mutual self-attention control for consistent image synthesis and editing. In Proc. of Int. Conf. on Computer Vision, pages 22560–22570, October 2023. 2, 3, 4, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 324, + 286, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 324, + 286, + 367 + ], + "spans": [ + { + "bbox": [ + 53, + 324, + 286, + 367 + ], + "type": "text", + "content": "[7] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 369, + 286, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 286, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 286, + 401 + ], + "type": "text", + "content": "[8] Alexei A. Efros and William T. Freeman. Image quilting for texture synthesis and transfer. In Proc. of SIGGRAPH, pages 341-346, 2001. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 403, + 286, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 286, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 286, + 434 + ], + "type": "text", + "content": "[9] Alexei A. Efros and Thomas K. Leung. Texture synthesis by non-parametric sampling. In Proc. of Int. Conf. on Computer Vision, 1999. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 436, + 286, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 436, + 286, + 479 + ], + "spans": [ + { + "bbox": [ + 48, + 436, + 286, + 479 + ], + "type": "text", + "content": "[10] Anna Frühstück, Ibraeem Alhashim, and Peter Wonka. Titegan: synthesis of large-scale non-homogeneous textures. ACM Trans. on Graphics (Proc. of SIGGRAPH), 38(4):1-11, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 480, + 286, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 480, + 286, + 523 + ], + "spans": [ + { + "bbox": [ + 48, + 480, + 286, + 523 + ], + "type": "text", + "content": "[11] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 525, + 286, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 286, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 286, + 557 + ], + "type": "text", + "content": "[12] Leon Gatys, Alexander S Ecker, and Matthias Bethge. Texture synthesis using convolutional neural networks. In Advances in Neural Information Processing Systems, 2015. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 558, + 286, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 286, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 286, + 601 + ], + "type": "text", + "content": "[13] Eric Heitz, Kenneth Vanhoey, Thomas Chambon, and Laurent Belcour. A sliced wasserstein loss for neural texture synthesis. In Proc. of IEEE Conf. on Computer Vision & Pattern Recognition, June 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 602, + 286, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 286, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 286, + 646 + ], + "type": "text", + "content": "[14] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. In Proc. of International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 647, + 286, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 286, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 286, + 680 + ], + "type": "text", + "content": "[15] Aaron Hertzmann, Charles E. Jacobs, Nuria Oliver, Brian Curless, and David H. Salesin. Image analogies. In Proc. of SIGGRAPH, page 327-340, 2001. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 681, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 286, + 713 + ], + "type": "text", + "content": "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[17] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 545, + 150 + ], + "type": "text", + "content": "[18] Nikolay Jetchev, Urs Bergmann, and Roland Vollgraf. Texture synthesis with spatial generative adversarial networks. arXiv preprint arXiv:1611.08207, 2016. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 151, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 151, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 151, + 545, + 194 + ], + "type": "text", + "content": "[19] Alexandre Kaspar, Boris Neubert, Dani Lischinski, Mark Pauly, and Johannes Kopf. Self tuning texture optimization. Computer Graphics Forum (Proc. of Eurographics), 34(2):349-359, may 2015. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 195, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 545, + 227 + ], + "type": "text", + "content": "[20] Anant Khandelwal. Infusion: Inject and attention fusion for multi concept zero shot text based video editing. arXiv preprint arXiv:2308.00135, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 228, + 545, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 260 + ], + "type": "text", + "content": "[21] Vivek Kwatra, Irfan Essa, Aaron Bobick, and Nipun Kwatra. Texture optimization for example-based synthesis. ACM Trans. on Graphics (Proc. of SIGGRAPH), August 2005. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 261, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 304 + ], + "type": "text", + "content": "[22] Vivek Kwatra, Arno Schödl, Irfan Essa, Greg Turk, and Aaron Bobick. Graphcut textures: Image and video synthesis using graph cuts. ACM Trans. on Graphics (Proc. of SIGGRAPH), 22(3):277-286, jul 2003. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 305, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 305, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 308, + 305, + 545, + 348 + ], + "type": "text", + "content": "[23] Chuan Li and Michael Wand. Precomputed real-time texture synthesis with markovian generative adversarial networks. In Proc. of Euro. Conf. on Computer Vision, pages 702-716, 2016. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 350, + 545, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 404 + ], + "type": "text", + "content": "[24] Yitzchak David Lockerman, Basile Sauvage, Rémi Allègre, Jean-Michel Dischler, Julie Dorsey, and Holly Rushmeier. Multi-scale label-map extraction for texture synthesis. ACM Trans. on Graphics (Proc. of SIGGRAPH), 35(4):140:1-140:12, July 2016. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 404, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 437 + ], + "type": "text", + "content": "[25] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models, 2022." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 438, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 438, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 308, + 438, + 545, + 491 + ], + "type": "text", + "content": "[26] M. Lukáč, J. Fíšer, P. Asente, J. Lu, E. Shechtman, and D. Sykora. Brushables: Example-based edge-aware directional texture painting. Computer Graphics Forum (Proc. of Pacific Conf. on Computer Graphics & Applications), 34(7):257-267, oct 2015. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 492, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 492, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 492, + 545, + 536 + ], + "type": "text", + "content": "[27] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 537, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 545, + 581 + ], + "type": "text", + "content": "[28] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2):3, 2022. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 582, + 545, + 614 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 545, + 614 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 545, + 614 + ], + "type": "text", + "content": "[29] Carlos Rodriguez-Pardo and Elena Garces. Seamlessgan: Self-supervised synthesis of tileable texture maps. IEEE Trans. Visualization & Computer Graphics, 2022. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 615, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 615, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 308, + 615, + 545, + 668 + ], + "type": "text", + "content": "[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proc. of IEEE Conf. on Computer Vision & Pattern Recognition, pages 10684-10695, 2022. 2, 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "text", + "content": "[31] Amir Rosenberger, Daniel Cohen-Or, and Dani Lischinski. Layered shape synthesis: Automatic generation of control maps for non-stationary textures. ACM Trans. on Graphics (Proc. of SIGGRAPH), 28(5):107:1-9, Dec. 2009. 1, 2" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7775" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 700 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[32] Tamar Rott Shaham, Tali Dekel, and Tomer Michaeli. Singan: Learning a generative model from a single natural image. In Proc. of Int. Conf. on Computer Vision, 2019. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "text", + "content": "[33] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proc. of IEEE Conf. on Computer Vision & Pattern Recognition, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 206 + ], + "type": "text", + "content": "[34] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdream booth: Hypernetworks for fast personalization of text-to-image models, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 287, + 239 + ], + "type": "text", + "content": "[35] Omry Sendik and Daniel Cohen-Or. Deep correlations for texture synthesis. ACM Trans. on Graphics, 36(5):161:1-161:15, July 2017. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 241, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 287, + 274 + ], + "type": "text", + "content": "[36] Wu Shi and Yu Qiao. Fast texture synthesis via pseudo optimizer. In Proc. of IEEE Conf. on Computer Vision & Pattern Recognition, pages 5498-5507, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 276, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 287, + 308 + ], + "type": "text", + "content": "[37] Assaf Shocher, Shai Bagon, Phillip Isola, and Michal Irani. Ingan: Capturing and retargeting the \"dna\" of a natural image. In Proc. of Int. Conf. on Computer Vision, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 309, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 309, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 287, + 342 + ], + "type": "text", + "content": "[38] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 343, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 287, + 396 + ], + "type": "text", + "content": "[39] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proc. of IEEE Conf. on Computer Vision & Pattern Recognition, pages 1921-1930, June 2023. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 399, + 287, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 442 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 442 + ], + "type": "text", + "content": "[40] Dmitry Ulyanov, Vadim Lebedev, Andrea Vedaldi, and Victor Lempitsky. Texture networks: Feed-forward synthesis of textures and stylized images. In Proc. of IEEE Int. Conf. on Machine Learning, page 1349-1357, 2016. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 445, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 287, + 487 + ], + "type": "text", + "content": "[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 30, 2017. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 489, + 287, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 489, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 48, + 489, + 287, + 542 + ], + "type": "text", + "content": "[42] Jiaping Wang, Xin Tong, Stephen Lin, Minghao Pan, Chao Wang, Hujun Bao, Baining Guo, and Heung-Yeung Shum. Appearance manifolds for modeling time-variant appearance of materials. ACM Trans. on Graphics (Proc. of SIGGRAPH), 25(3):754–761, 2006. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 544, + 287, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 577 + ], + "type": "text", + "content": "[43] Li-Yi Wei and Marc Levoy. Fast texture synthesis using tree-structured vector quantization. In Proc. of SIGGRAPH, pages 479–488, 2000. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 578, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 610 + ], + "type": "text", + "content": "[44] Yonatan Wexler, Eli Shechtman, and Michal Irani. Spacetime completion of video. IEEE Trans. Pattern Analysis & Machine Intelligence, 29(3):463-476, 2007. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 613, + 287, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 655 + ], + "type": "text", + "content": "[45] Li-Yi Wie, Sylvain Lefebvre, Vivek Kwatra, and Greg Turk. State of the Art in Example-based Texture Synthesis. In M. Pauly and G. Greiner, editors, Eurographics 2009 - State of the Art Reports. The Eurographics Association, 2009. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 657, + 287, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 657, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 48, + 657, + 287, + 700 + ], + "type": "text", + "content": "[46] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. arXiv preprint arXiv:2211.13227, 2022. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 285 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 115 + ], + "type": "text", + "content": "[47] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. In arXiv preprint arxiv:2308.06721, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 149 + ], + "type": "text", + "content": "[48] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "text", + "content": "[49] Yang Zhou, Kaijian Chen, Rongjun Xiao, and Hui Huang. Neural texture synthesis with guided correspondence. In Proc. of IEEE Conf. on Computer Vision & Pattern Recognition, pages 18095-18104, 2023. 2, 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "type": "text", + "content": "[50] Yang Zhou, Huajie Shi, Dani Lischinski, Minglun Gong, Johannes Kopf, and Hui Huang. Analysis and controlled synthesis of inhomogeneous textures. Computer Graphics Forum (Proc. of Eurographics), 36(2):199-212, 2017. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "type": "text", + "content": "[51] Yang Zhou, Zhen Zhu, Xiang Bai, Dani Lischinski, Daniel Cohen-Or, and Hui Huang. Non-stationary texture synthesis by adversarial expansion. ACM Trans. on Graphics (Proc. of SIGGRAPH), 37(4):49:1-49:13, 2018. 1, 2, 5, 6" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7776" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Generative 3D Part Assembly via Part-Whole-Hierarchy Message Passing/40f117f5-c539-42c1-b970-aab5fe53da07_content_list.json b/2024/Generative 3D Part Assembly via Part-Whole-Hierarchy Message Passing/40f117f5-c539-42c1-b970-aab5fe53da07_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5831c1188f96808525ee686b85d56cfaf3a88f34 --- /dev/null +++ b/2024/Generative 3D Part Assembly via Part-Whole-Hierarchy Message Passing/40f117f5-c539-42c1-b970-aab5fe53da07_content_list.json @@ -0,0 +1,2116 @@ +[ + { + "type": "text", + "text": "Generative 3D Part Assembly via Part-Whole-Hierarchy Message Passing", + "text_level": 1, + "bbox": [ + 112, + 130, + 854, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bi'an Du1, Xiang Gao1, Wei Hu1*, Renjie Liao2,3,4", + "bbox": [ + 285, + 179, + 683, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Wangxuan Institute of Computer Technology, Peking University", + "bbox": [ + 227, + 198, + 741, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ University of British Columbia, $^{3}$ Vector Institute for AI, $^{4}$ Canada CIFAR AI Chair", + "bbox": [ + 153, + 215, + 813, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "pkudba@stu.pku.edu.cn, gyshgx868@pku.edu.cn, forhuwei@pku.edu.cn, rjliao@ece.ubc.ca", + "bbox": [ + 112, + 237, + 849, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generative 3D part assembly involves understanding part relationships and predicting their 6-DoF poses for assembling a realistic 3D shape. Prior work often focuses on the geometry of individual parts, neglecting part-whole hierarchies of objects. Leveraging two key observations: 1) super-part poses provide strong hints about part poses, and 2) predicting super-part poses is easier due to fewer superparts, we propose a part-whole-hierarchy message passing network for efficient 3D part assembly. We first introduce super-parts by grouping geometrically similar parts without any semantic labels. Then we employ a part-whole hierarchical encoder, wherein a super-part encoder predicts latent super-part poses based on input parts. Subsequently, we transform the point cloud using the latent poses, feeding it to the part encoder for aggregating super-part information and reasoning about part relationships to predict all part poses. In training, only ground-truth part poses are required. During inference, the predicted latent poses of super-parts enhance interpretability. Experimental results on the PartNet dataset show that our method achieves state-of-the-art performance in part and connectivity accuracy and enables an interpretable hierarchical part assembly. Code is available at https://github.com/pkudba/3DHPA.", + "bbox": [ + 75, + 318, + 473, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 696, + 209, + 713 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generative 3D Part Assembly [7, 15, 16, 25, 32] is an emerging research area that aims to generate complex 3D shapes via assembling simple 3D parts without relying on prior semantic knowledge. Different from traditional 3D shape generation, it focuses on generating diverse, plausible configurations of given parts. It facilitates the generation of complex objects and scenes with compositionality, flexibility, and efficiency. With the rapid evolution of 3D printing technology and the increasing demand for diverse 3D shapes, generative 3D part assembly finds applications in various sce", + "bbox": [ + 75, + 723, + 470, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "narios, attracting attention from experts in computer vision, graphics, robotics, and machine learning.", + "bbox": [ + 500, + 287, + 892, + 318 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, achieving generative 3D part assembly presents a significant challenge due to the vast number of potential part arrangements and orientations, coupled with the intricate dependencies among the parts. Adding to the complexity, part geometry exhibits notable variation even within the same object category, making it exceedingly difficult to generalize learned assembly patterns across different objects.", + "bbox": [ + 496, + 320, + 893, + 426 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Previous approaches primarily focus on designing architectures capable of learning powerful representations for individual 3D parts. The hope is that these learned representations could facilitate accurate part assembly, either in a one-shot manner or sequentially [14, 21, 33, 34]. However, these methods often ignore the inherent part-whole hierarchies in 3D shapes in representation learning. For instance, as illustrated in Figure 1, a chair consists of super-parts such as seats, backs, and legs, with each super-part further divisible into parts like seat surfaces and seat frames. Understanding the pose of a super-part provides insights into the poses of constituent parts within the same super-part, as they often share similar orientations or exhibit symmetry (e.g., left-right symmetry in chair legs and arms). Moreover, predicting super-part poses is typically easier due to the fewer number of super-parts compared to parts. Incorporating these hierarchies into the modeling would make the learning process and potentially improve the overall performance.", + "bbox": [ + 496, + 428, + 895, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we introduce a part-whole-hierarchy message passing network for 3D shape assembly, predicting 6 degrees of freedom (6-DoF) poses for super-parts and parts in a hierarchical manner. We establish the correspondence between parts and super-parts (i.e., subsets of parts) by grouping parts based on their geometric similarities, following the approach from [33, 34]. Importantly, we treat super-part poses as latent variables to be learned, thereby eliminating the requirement of ground-truth super-part poses in our work. Our model comprises two sequential modules: the super-part encoder and the part encoder.", + "bbox": [ + 496, + 702, + 895, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The first super-part encoder takes the 3D point clouds of all parts as input and employs the attention-based message", + "bbox": [ + 500, + 869, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author: Wei Hu (forhuwei@pku.edu.cn).", + "bbox": [ + 94, + 886, + 403, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20850", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a113e8d42477061455253f57651ce08ff4f9753ac7e0eaead656a8e64c613940.jpg", + "image_caption": [ + "Figure 1. Left: an illustration of the part-whole-hierarchy for 3D shapes; Right: the part assembly process via the proposed part-whole-hierarchy message passing network." + ], + "image_footnote": [], + "bbox": [ + 96, + 92, + 880, + 294 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "passing to predict super-parts poses. The representation of each super-part is aggregated from representations of parts within it. These super-part poses provide initial estimation for the poses of individual parts.", + "bbox": [ + 75, + 342, + 470, + 402 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Subsequently, the whole point clouds are transformed based on the predicted super-part poses and fed to the part encoder. This module, using a cross-attention mechanism, extracts features from the transformed point cloud, transferring super-part level information to the part level. It then leverages the attention-based message passing again to capture relationships among individual parts.", + "bbox": [ + 75, + 402, + 470, + 508 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Following previous works [14, 21, 33, 34], we train and evaluate our model on the PartNet dataset [18]. Overall, our model achieves the state-of-the-art performances, outperforming the strongest competitor [34] by a significant margin, i.e., with an almost $2\\%$ improvement in mean part accuracy and a $3\\%$ improvement in mean connectivity accuracy. Moreover, through visual analysis of the assembly process for both super-parts and parts, we not only showcase accurate generation of part poses but also demonstrate interpretability via the predicted super-part poses. This interpretability feature further improves the utility of our model.", + "bbox": [ + 75, + 510, + 470, + 676 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 691, + 218, + 709 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Assembly3D modeling Numerous prior studies have approached the challenging 3D part assembly through the joint estimation of part poses. The influential work by [7] introduces an intelligent scissoring technique to tackle this problem for part components. Subsequent research by [2, 10, 12] employs graphical models to capture the semantic and geometric relationships among shape components, enabling exploration in assembly-based shape modeling. PAGENet [14] presents a network that is aware of individual parts and generates semantic parts along with their poses. [31] proposes to decompose shapes using a library of 3D parts provided by the user. However, these studies either assumed prior knowl-", + "bbox": [ + 75, + 719, + 472, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "edge of part semantics or relied on existing shape databases. In a more practical setting, the authors in [3, 21, 33, 34] focus on the pose estimation of individual parts without relying on shape databases or known semantic information. Specifically, DGL [33] employs a dynamic part graph to iteratively refine the poses of individual parts. A progressive strategy using the recurrent graph learning framework has been investigated in [21]. Additionally, the authors in [3, 34] utilizes the Transformer [28] to model the structural relationships and performs the simultaneous assembly of all parts. We follow this line of work and propose novel improvement to generate structurally-coherent part assemblies.", + "bbox": [ + 496, + 342, + 893, + 523 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Structural Shape Generation In recent years, deep generative models, such as generative adversarial networks (GANs) [9] and variational autoencoders (VAEs) [5], have garnered significant attention for shape generation tasks. Notably, the work by [8] introduces a two-level variational autoencoder that simultaneously learns the overall shape structure and detailed part geometries. Both GRASS [13] and StructureNet [17] employ techniques to compress the shape structure into a latent space while considering the relationships between different parts. Additionally, [19] uses a part-tree decomposition to conditionally generate 3D shapes, and [11] adopts a procedural programmatic representation to establish connections between part cuboids. Inspired by the Seq2Seq networks in machine translation, [29] introduces a sequential encoding and decoding approach for the regression of shape parameters. While many of these approaches focus on directly generating new part shapes given random latent codes, our main focus revolves around the rigid transformation of existing parts to facilitate their assembly.", + "bbox": [ + 496, + 542, + 895, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Part-Whole-Hierarchy Message Passing", + "text_level": 1, + "bbox": [ + 498, + 844, + 857, + 862 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This section provides a detailed explanation of our proposed part-whole-hierarchy message passing network. We begin by", + "bbox": [ + 498, + 869, + 892, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20851", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/78accfc024b593cb18524471659a2a24ecd2cfd53c66d1699618b441aaca29ba.jpg", + "image_caption": [ + "Figure 2. The overall architecture of our model consists of two modules: super-part encoder and part encoder. We first obtain super-parts via grouping parts based on their geometric similarities in an unsupervised fashion. The super-part encoder takes point cloud as input and predicts latent super-part poses (no ground truth is needed). The point cloud is then transformed based on super-part poses and fed to the part encoder. We incorporate both cross-level and within-level attention in the part encoder to predict part poses." + ], + "image_footnote": [], + "bbox": [ + 89, + 88, + 880, + 390 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "presenting how we construct super-parts in an unsupervised manner. Then we introduce our super-part encoder that predicts the latent poses of super-parts. Subsequently, we describe how our part encoder leverages the latent poses and the part-whole hierarchy to predict part poses. The overall model is illustrated in Figure 2. Finally, we introduce the loss functions and the training process for diverse generation.", + "bbox": [ + 75, + 465, + 472, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Super-Part Construction", + "text_level": 1, + "bbox": [ + 76, + 585, + 305, + 603 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We denote the input point clouds as $\\mathcal{P} = \\{\\mathbf{P}_i|i = 1\\ldots N\\}$ , where $\\mathbf{P}_i\\in \\mathbb{R}^{d\\times 3}$ corresponds to the point cloud of the $i$ -th given part of the 3D shape. The values of $N$ and $d$ represent the number of parts and the number of points per part respectively. Note that their values may differ from object to object. For notation convenience, we assume objects are padded to the maximum number of parts, and each part is padded to the maximum number of points per part. The goal of our task is to predict a set of 6-DoF part poses $\\{(t_i,r_i)\\}_{i = 1}^N$ , where $t_i\\in \\mathbb{R}^3$ and $r_i\\in \\mathbb{R}^4$ represent the translation and rigid rotation for each part, respectively. The complete part assembly for a 3D shape is $\\mathcal{S} = \\cup_{i = 1}^{N}\\mathbf{T}_i(\\mathbf{P}_i)$ , and $\\mathbf{T}_i$ represents a transformation in $SE(3)$ that consists of a 3D rotation in the rotation group $SO(3)$ and a 3D translation in the translation group, induced by $(t_i,r_i)$ .", + "bbox": [ + 75, + 611, + 470, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A super-part is a subset of parts that is ideally semantically meaningful. However, since we do not have groundtruth super-parts, we need to construct them in an unsupervised manner. In particular, we compute axis-aligned", + "bbox": [ + 75, + 839, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "bounding boxes for each part and evaluate the similarity between these 3D boxes. Parts are grouped into the same super-part if the difference between their respective enclosing boxes is below a specified threshold. Following previous works [33, 34], we set the threshold to 0.2. Although the super-parts obtained through this method may lack semantic meaning, they provide a coarse abstraction grounded in geometry similarities without any supervised labels. Then we group the input point cloud parts $\\mathcal{P} = \\{\\mathbf{P}_i\\}_{i=1}^N$ into a set of $M$ super-parts $\\mathcal{P}' = \\{\\mathbf{P}_i'\\}_{i=1}^M$ based on the part-whole hierarchy of a given object. Here $\\mathbf{P}_i'$ represents the $i$ -th super-part. To maintain consistent notation, we again pad the number of parts per super-part to the maximum $M$ . Note that padding is used here for clarity and understanding. In practice, our model can handle sets of parts with varying sizes without padding.", + "bbox": [ + 496, + 465, + 893, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Super-Part Relation Reasoning", + "text_level": 1, + "bbox": [ + 498, + 723, + 777, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To reason about the relationships among super-parts, we utilize the super-part encoder. The computational process is illustrated in the top half of Figure 2.", + "bbox": [ + 496, + 748, + 890, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a set of point clouds of a super-part $\\mathbf{P}_i^{\\prime}$ , we first compute the set-level permutation-invariant representation $\\mathbf{F}_i^{\\prime}$ via a shared PointNet [22]. We use the same architecture and hyperparameters as in previous works [33, 34] for a fair comparison. This feature representation captures the global characteristics of each super-part while being invariant to the order of the parts within the set. Based on the features", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20852", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\mathbf{F}_i^{\\prime}$ extracted for each super-part from the PointNet, we use a two-layer multi-head attention (MHA) module to learn the relationships between super-parts. Specifically, following [28], we calculate an attention matrix $\\mathbf{A}^{\\prime}$ , where $\\mathbf{A}_{i,j}^{\\prime}$ represents the attention weight of the $i$ -th super-part to the $j$ -th super-part. Then we multiply each super-part feature $\\mathbf{F}_j^{\\prime}$ by $\\mathbf{A}_{i,j}^{\\prime}$ and sum them up to obtain the attention-weighted feature $\\mathbf{G}_i^{\\prime}$ . In our model, we set the embedding dimension to 256 and the number of heads for the MHA module to 8.", + "bbox": [ + 75, + 90, + 472, + 224 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For super-part pose prediction, we feed the attention-weighted feature $\\mathbf{G}_i'$ into a prediction head containing 4 fully-connected layers to obtain the pose $\\{(t_i', r_i')\\}_{i=1}^M$ . We apply tanh operation to the translation vector, which restricts the range of the part center offset as $(-1,1)$ . Additionally, for simplicity, we predict the quaternion vector $r_i' = (r_{i0}', r_{i1}', r_{i2}', r_{i3}')$ instead of the rotation matrix. We then use the Rodrigues formula [23] to obtain the rotation matrix $\\mathbf{R}_i'$ corresponding to $r_i'$ one by one as follows:", + "bbox": [ + 75, + 226, + 473, + 354 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {R} _ {i} ^ {\\prime} = \\left[ \\begin{array}{l} 1 - 2 r _ {i 2} ^ {\\prime 2} - 2 r _ {i 3} ^ {\\prime 2}, 2 r _ {i 1} ^ {\\prime} r _ {i 2} ^ {\\prime} - 2 r _ {i 0} ^ {\\prime} r _ {i 3} ^ {\\prime}, 2 r _ {i 1} ^ {\\prime} r _ {i 3} ^ {\\prime} + 2 r _ {i 0} ^ {\\prime} r _ {i 2} ^ {\\prime} \\\\ 2 r _ {i 1} ^ {\\prime} r _ {i 2} ^ {\\prime} + 2 r _ {i 0} ^ {\\prime} r _ {i 3} ^ {\\prime}, 1 - 2 r _ {i 1} ^ {\\prime 2} - 2 r _ {i 3} ^ {\\prime 2}, 2 r _ {i 2} ^ {\\prime} r _ {i 3} ^ {\\prime} - 2 r _ {i 0} ^ {\\prime} r _ {i 1} ^ {\\prime} \\\\ 2 r _ {i 1} ^ {\\prime} r _ {i 3} ^ {\\prime} - 2 r _ {i 0} ^ {\\prime} r _ {i 2} ^ {\\prime}, 2 r _ {i 2} ^ {\\prime} r _ {i 3} ^ {\\prime} + 2 r _ {i 0} ^ {\\prime} r _ {i 1} ^ {\\prime}, 1 - 2 r _ {i 1} ^ {\\prime 2} - 2 r _ {i 2} ^ {\\prime 2}. \\end{array} \\right]\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 364, + 475, + 412 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The transformed point cloud of the super-part is expressed as $\\mathbf{T}_i(\\mathbf{P}_i^{\\prime}) = \\mathbf{R}_i^{\\prime}\\mathbf{P}_i^{\\prime} + t_i^{\\prime}$ . To ensure that it is a unit quaternion, we normalize the rigid rotation vector such that $\\| r_i^\\prime \\| = 1$ .", + "bbox": [ + 75, + 420, + 470, + 467 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since we do not have ground-truth super-part poses, the supervision for this module solely comes from the backpropagation signal from the subsequent module. As a result, the super-part poses are considered as latent variables.", + "bbox": [ + 75, + 467, + 470, + 526 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Part Relation Reasoning", + "text_level": 1, + "bbox": [ + 76, + 536, + 302, + 553 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As previously mentioned, the poses of super-parts offer valuable clues about the poses of their corresponding parts. To leverage this information, we design a part encoder to incorporate the latent super-part poses to predict the part poses.", + "bbox": [ + 75, + 560, + 472, + 621 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Transformation via Latent Super-Part Poses Firstly, we apply the latent super-part poses to transform all point cloud parts, resulting in transformed parts denoted as $\\hat{\\mathcal{P}} = \\{\\hat{\\mathbf{P}}_i\\}_{i=1}^N$ . These transformed parts are then fed into the part encoder for further processing. Similarly to the super-part encoder, we employ another PointNet to extract part-level feature denoted as $\\hat{\\mathbf{F}}_i$ . In order to better utilize the part-whole hierarchy of objects, we design two ways to integrate the super-part level information to the part level, i.e., cross-level and within-level attention modules.", + "bbox": [ + 75, + 638, + 472, + 789 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Cross-Level Attention Part representations $\\hat{\\mathbf{F}}_i$ from the PointNet encoder does not leverage the part-whole hierarchy. To better fuse the information from the super-parts, we propose a cross-level attention module. Specifically, we treat the part representation $\\hat{\\mathbf{F}}_i$ as query and super-part representations $\\mathbf{F}_i'$ as keys and values. We then employ a MHA module", + "bbox": [ + 75, + 809, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "to perform an attention-weighted aggregation of super-part representations $\\mathbf{F}_i^{\\prime}$ to obtain the updated part representations $\\mathbf{F}_i$ . By doing so, the information from part-whole hierarchy is explicitly integrated into part representations. Here we specify the embedding dimension as 256 and the MHA module is set to utilize 8 heads.", + "bbox": [ + 496, + 90, + 893, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Within-Level Attention To model the part relationships, we leverage another within-level attention module, which is depicted in the bottom half of Figure 2. We first concatenate the representation from the previous module $F_{i}$ with an additional instance encoding vector that is unique for each part following [34]. This allows us to distinguish geometrically-equivalent parts. Then we perform message passing among parts using another MHA module where the attention computation is aware of the part-whole hierarchy. Specifically, we concatenate the latent super-part poses to the part-level feature before computing the attention scores at each layer. Consequently, in each message passing layer, a part receives pose hints from its parent super-part and updates its representation. The resulting part feature is denoted as $\\mathbf{G}_i = \\sum_{j=1}^N \\mathbf{A}_{i,j} \\mathbf{F}_i$ , where $\\mathbf{A}_{i,j}$ represents the attention weight of the $i$ -th part to the $j$ -th part. This design enables the part-to-part message passing to effectively leverage the part-whole hierarchy information. We set the number of heads for the MHA module to 8, the dimension of the part-level feature to 256, and concatenate it with the 40-dimensional instance encoding vector.", + "bbox": [ + 496, + 200, + 895, + 517 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Prediction Head Finally, we feed the attention-weighted part-level feature $\\mathbf{G}_i$ to a part pose prediction head, which consists of four fully-connected layers. The channel sizes of the first three fully-connected layers are set to 256, 256 and 1024, respectively, followed by ReLU activation functions. We use a linear projection layer to predict the part pose $(t_i, r_i)$ , encompassing a 4-dimensional rigid rotation and 3-dimensional translation. This prediction head outputs the part poses $\\{(t_i, r_i)\\}_{i=1}^N$ , where $t_i$ represents the translation vector and $r_i$ represents the rigid rotation vector for each part. We adopt the same design as in the super-part encoder to restrict the translation vectors to the range of $(-1, 1)$ and normalize the rigid rotation vector $r_i$ to have a unit norm.", + "bbox": [ + 496, + 535, + 893, + 731 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Training Objective for Diverse Generation", + "text_level": 1, + "bbox": [ + 498, + 739, + 861, + 757 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Considering the same set of input point cloud parts, geometrically equivalent parts (such as legs of a chair) can be interchanged and decorated parts may have multiple placement options, resulting in multiple possible shapes. For example, a semi-cylindrical part can be placed on top of a chair backrest as a headrest, or placed under the backrest as a lumbar pillow. To account for such diversified structural variations and configurations, we introduce random noise into the system following [21, 33, 34] and employ the Min-of-N", + "bbox": [ + 496, + 763, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20853", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(MoN) loss [6] to simultaneously consider assembly accuracy and assembly diversity. By considering the minimum of a set of N possible assemblies, we can encourage the model to generate diverse and valid shape configurations while maintaining overall accuracy in the assembly process. This helps address the challenge of capturing multiple plausible solutions that arise from the interchangeable and decoratable nature of the parts.", + "bbox": [ + 75, + 90, + 472, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let $\\mathcal{F}(\\mathcal{P},\\delta_j)$ denote our network outputs and $\\mathcal{F}^* (\\mathcal{P})$ denote the ground-truth point clouds, then the MoN loss is:", + "bbox": [ + 76, + 210, + 470, + 242 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {M o N}} = \\min _ {\\delta_ {j} \\sim \\mathcal {N} (0, 1)} \\mathcal {L} \\left(\\mathcal {F} \\left(\\mathcal {P}, \\delta_ {j}\\right), \\mathcal {F} ^ {*} (\\mathcal {P})\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 248, + 468, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\delta_{j}$ is a random noise vector drawn from standard Normal distribution in an IID fashion. Following [34], we sample 5 random vectors $\\delta_{j}$ during training. The loss function $\\mathcal{L}$ consists of the following components for both local part and global shape losses.", + "bbox": [ + 75, + 279, + 470, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Firstly, we supervise the translation via an $l_{2}$ loss between our prediction $t_{i}$ and the ground-truth translation vector $t_{i}^{*}$ for each part:", + "bbox": [ + 75, + 354, + 468, + 398 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t} = \\sum_ {i = 1} ^ {N} \\| t _ {i} - t _ {i} ^ {*} \\| _ {2} ^ {2} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 398, + 468, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then we compute the Chamfer distance [6] between the predicted and the ground-truth rigid rotation for each part:", + "bbox": [ + 76, + 441, + 468, + 472 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r} = \\sum_ {i = 1} ^ {N} d _ {c} \\left(r _ {i} \\left(\\mathbf {P} _ {i}\\right), r _ {i} ^ {*} \\left(\\mathbf {P} _ {i}\\right)\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 478, + 468, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Chamfer distance $d_{c}(\\mathcal{X},\\mathcal{Y})$ is a metric commonly used in point cloud comparison, which measures the dissimilarity between two point sets $\\mathcal{X}$ and $\\mathcal{Y}$ by calculating the average distance between each point in one set and its nearest neighbor in the other set as follows:", + "bbox": [ + 75, + 525, + 468, + 599 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nd _ {c} (\\mathcal {X}, \\mathcal {Y}) = \\sum_ {x \\in \\mathcal {X}} \\min _ {y \\in \\mathcal {Y}} \\| x - y \\| _ {2} ^ {2} + \\sum_ {y \\in \\mathcal {Y}} \\min _ {x \\in \\mathcal {X}} \\| x - y \\| _ {2} ^ {2} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 606, + 468, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this case, we use the Chamfer distance to assess the discrepancy between the predicted rotation of each part and its corresponding ground-truth rotation.", + "bbox": [ + 75, + 645, + 468, + 690 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similarly, the full shape $S$ , i.e., the set of point clouds belonging to all parts, is supervised via the Chamfer distance from the ground-truth shape $S^{*}$ :", + "bbox": [ + 75, + 690, + 468, + 734 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s} = d _ {c} \\left(\\mathcal {S}, \\mathcal {S} ^ {*}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 742, + 468, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In summary, the overall loss function is defined as:", + "bbox": [ + 96, + 766, + 433, + 780 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {t} \\mathcal {L} _ {t} + \\lambda_ {r} \\mathcal {L} _ {r} + \\lambda_ {s} \\mathcal {L} _ {s} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 787, + 468, + 803 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{t},\\lambda_{r}$ and $\\lambda_{s}$ represent the weights assigned to the three loss terms. Based on the cross validation, we empirically set $\\lambda_t = 1$ $\\lambda_r = 10$ ,and $\\lambda_s = 1$ in our experiments. By minimizing this loss function, we aim to improve the alignment of the predicted poses with the ground-truth poses, thereby enhancing the overall accuracy of the assembly.", + "bbox": [ + 75, + 809, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 90, + 633, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we demonstrate the effectiveness of the proposed model by comparing our results with state-of-the-art methods. We also provide visual analysis of the assembly process for super-parts and parts, which helps to illustrate the hierarchical assembly process from parts to the whole.", + "bbox": [ + 496, + 114, + 893, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Dataset", + "text_level": 1, + "bbox": [ + 500, + 203, + 594, + 217 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following [4, 14, 21, 24, 27, 29, 33, 34], we use the PartNet [18] dataset for both training and evaluation. This dataset consists of 26,671 shapes across 24 different 3D object categories. To effectively validate and compare different methods, we select the three largest categories, i.e., chairs, tables and lamps, following [4, 14, 21, 33, 34]. In total, we have 6,323 chairs, 8,218 tables, 2,207 lamps in the finest-grained level, and we adopt the official train/validation/test splits (70%/10%/20%) to conduct the experiments. For each part point cloud, 1000 points are sampled from the original part meshes using Farthest Point Sampling (FPS) [20]. All parts are transformed into its canonical space using PCA.", + "bbox": [ + 496, + 226, + 893, + 409 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Evaluation Metrics", + "text_level": 1, + "bbox": [ + 500, + 420, + 683, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We generate a variety of shapes by adding different Gaussian noises to a given set of input parts, and find the closest shape to the ground-truth using minimum matching distance (MMD) [1]. To measure part assembly quality, we follow [21, 33, 34] to use shape Chamfer distance (SCD), part accuracy (PA) and connectivity accuracy (CA). Shape Chamfer distance is defined in Equation (4) and Equation (5). Meanwhile, to compare the diversity of assembled parts, we propose two quantitative evaluation metrics including quality-diversity score (QDS) and the weighted quality-diversity score (WQDS) following [4]. We also visualize the generation results to qualitatively evaluate the diversity. More details of these metrics are provided in the appendix.", + "bbox": [ + 496, + 444, + 893, + 641 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Mean Part/Connectivity Accuracy (mPA/mCA). PA and CA depend on the Chamfer distance threshold $\\tau_{p}$ and $\\tau_{c}$ for judging whether the assembly is accurate. In order to provide a more comprehensive evaluation of the performance of the model, we average the results under multiple thresholds to get mean part accuracy (mPA) and mean connectivity accuracy (mCA), formally,", + "bbox": [ + 496, + 662, + 893, + 770 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {m P A} = \\frac {1}{T _ {p}} \\sum_ {\\tau_ {p} \\in T _ {p}} \\mathrm {P A} (\\tau_ {p}), \\quad \\mathrm {m C A} = \\frac {1}{T _ {c}} \\sum_ {\\tau_ {c} \\in T _ {c}} \\mathrm {C A} (\\tau_ {c}),\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 781, + 880, + 819 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $T_{p}$ and $T_{c}$ are set to $\\{0.01, 0.02, 0.03, 0.04, 0.05\\}$ .", + "bbox": [ + 500, + 832, + 872, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "QDS. Diversity score (DS) [19, 26] evaluates the diversity of the results as $\\mathrm{DS} = \\frac{1}{N^2}\\sum_{i,j=1}^{N}(d_c(\\mathbf{P}_i^*,\\mathbf{P}_j^*))$ , where $\\mathbf{P}_i^*$", + "bbox": [ + 496, + 868, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "footer", + "text": "20854", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/a471d791dd498fb8c933c60bd5b9591d41c32ec0d452fdb518eab7c7c35d69e6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SCD(10-2) ↓PA(%) ↑CA(%) ↑QDS(10-5) ↑WQDS(10-5) ↑
MethodsChairTableLampChairTableLampChairTableLampChairTableLampChairTableLamp
B-Global [14, 24]1.461.120.7915.7015.3722.619.9033.8418.600.150.200.761.251.400.58
B-LSTM [29]1.311.250.7721.7728.6420.786.8022.5614.050.040.270.631.071.431.54
B-Complement [27]2.412.981.508.782.3212.679.1915.5726.560.090.062.811.281.752.08
DGL [33]0.910.500.9339.0049.5133.3323.8739.9641.701.693.051.841.352.971.73
Score [4]0.710.421.1144.5152.7834.3230.3240.5949.073.369.176.831.703.812.82
RGL [21]0.870.480.7249.0654.1637.5632.2642.1557.343.557.636.822.124.072.96
IET [34]0.540.351.0362.8061.6738.6848.4556.1862.624.159.096.982.744.563.29
Ours0.510.280.7064.1364.8338.8049.2858.4564.165.629.587.123.064.813.90
", + "bbox": [ + 81, + 88, + 893, + 229 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0eaa9a2a4847629c1c81385e63fdf2ae7ce6f903536ab25e8e371e24f19154f4.jpg", + "table_caption": [ + "Table 1. Comparison between our approach and other methods under the Chamfer distance threshold 0.01." + ], + "table_footnote": [], + "table_body": "
mPA ↑mCA ↑
MethodsChairTableLampChairTableLamp
DGL [33]45.8457.8648.3229.1743.8851.75
IET [34]74.9273.2052.8062.3768.4975.94
Ours76.7976.1953.3165.3270.2678.15
", + "bbox": [ + 81, + 263, + 468, + 353 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Comparison between our approach and other methods under multiple Chamfer distance thresholds from 0.01 to 0.05.", + "bbox": [ + 76, + 359, + 468, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and $\\mathbf{P}_j^*$ represent any two assembled shapes. Based on DS, we propose the quality-diversity score (QDS) as below,", + "bbox": [ + 76, + 406, + 468, + 436 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {Q D S} = \\frac {1}{N ^ {2}} \\sum_ {i, j = 1} ^ {N} d _ {c} \\left(\\mathbf {P} _ {i} ^ {*}, \\mathbf {P} _ {j} ^ {*}\\right) \\mathbb {1} \\left(\\mathrm {C A} \\left(\\mathbf {P} _ {i} ^ {*} > \\tau_ {q}\\right)\\right) \\mathbb {1} \\left(\\mathrm {C A} \\left(\\mathbf {P} _ {j} ^ {*} > \\tau_ {q}\\right)\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 446, + 473, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "QDS imposes constraints to remove pairs that are of low assembly quality, thus assessing not only diversity but also the quality of generated shapes. Following [4], we adopt SCD as the distance metric. The value of $\\tau_{q}$ is set to 0.5 in both QDS and WQDS.", + "bbox": [ + 76, + 494, + 468, + 571 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "WQDS. Based on QDS, we further propose the weighted quality-diversity score (WQDS) as below,", + "bbox": [ + 76, + 590, + 468, + 619 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathrm {W Q D S} = \\frac {1}{N ^ {2}} \\sum_ {i, j = 1} ^ {N} d _ {c} \\left(\\mathbf {P} _ {i} ^ {*}, \\mathbf {P} _ {j} ^ {*}\\right) \\mathrm {C A} \\left(\\mathbf {P} _ {i}\\right) \\mathrm {C A} \\left(\\mathbf {P} _ {j}\\right) \\\\ \\mathbb {1} \\left(\\mathrm {C A} \\left(\\mathbf {P} _ {i} ^ {*} > \\tau_ {q}\\right)\\right) \\mathbb {1} \\left(\\mathrm {C A} \\left(\\mathbf {P} _ {j} ^ {*} > \\tau_ {q}\\right)\\right). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 630, + 441, + 689 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "WQDS weights the QDS of a pair by their connectivity accuracy, thus favoring assembled shapes that demonstrate a high-quality connection between each pair of parts.", + "bbox": [ + 76, + 699, + 468, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparisons with State-of-the-Art", + "text_level": 1, + "bbox": [ + 76, + 755, + 382, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our method with the state-of-the-art methods on PartNet [18] dataset. Following previous works, we first present the quantitative results under a certain Chamfer distance threshold of 0.01 in Table 1. As we can see, our method consistently outperforms all competitors. We also show the visualization of assembly results in Figure 3. It is clear that our method is better than others in terms of both the coherence of the assembled structure and the connectivity between", + "bbox": [ + 75, + 779, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "parts, e.g., the positioning of crossbeams between chair legs, and the connection between table legs and the tabletop.", + "bbox": [ + 496, + 267, + 890, + 297 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To eliminate the sensitivity of performances with respect to threshold values and provide a more comprehensive evaluation, we compare the mPA and mCA of our method with two competitive baselines in Table 2. For the Chair category, our method demonstrates a significant improvement over IET [34], with a $1.87\\%$ increase in mPA and a $2.95\\%$ increase in mCA, while the improvement in the Lamp category is relatively minor. This may be attributed to the limited variation in the geometry of lamp parts, making it difficult to clearly differentiate between various levels of the part-whole hierarchy. In addition, Figure 5 illustrates the performance of various methods on the Chair and Table categories using five different Chamfer distance thresholds ranging from 0.01 to 0.05. As the Chamfer distance threshold increases, the advantage of our method becomes more pronounced, which shows the superiority of our approach. We also evaluate the diversity of generative part assembly. As shown in Table 1, our method outperforms all competing methods in both QDS and WQDS. In Figure 4, we use the model to generate multiple assembly shapes for the same set of parts (with different noises) to show the variation of generated shapes.", + "bbox": [ + 496, + 299, + 893, + 617 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Human Study", + "text_level": 1, + "bbox": [ + 500, + 628, + 643, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since many inaccurate results are still valid visually, we also conduct a human study to evaluate the quality of generated assemblies. Specifically, we perform an A/B test with a group of students, utilizing the identical evaluation system as in [30, 35]. Participants were presented with pairs of randomly chosen assemblies of the same object from two different methods. In total, 28 participants labeled 93 assembly pairs. We compare the proposed method with three main baselines and present the outcomes in Table 5. In over $95\\%$ of the cases, our method is favored over other three competing methods, unequivocally demonstrating the superior visual quality of our method's output.", + "bbox": [ + 496, + 652, + 893, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 845, + 653, + 862 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We now investigate the importance of different loss components and verify the effectiveness of the proposed super-part", + "bbox": [ + 496, + 869, + 893, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20855", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c0a31a3bd067e53f1fbc39286632ef9f02b7469d3a791bfb754c93c45a61b6f5.jpg", + "image_caption": [ + "DGL" + ], + "image_footnote": [], + "bbox": [ + 145, + 103, + 205, + 164 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/98a9b109683bc0a7124b90dc96a7b1953e4e7618d8eb21019c6b1041876b3835.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 103, + 292, + 165 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d89e3dc458b7d1518e4047bb7a1c992606a7b462d2501e4f27a8e85951bf81f1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 305, + 103, + 387, + 166 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ee61dbac09230377db4af431ea67f53e540a5dd642a18814ce7bc67efac6b606.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 408, + 116, + 483, + 164 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b4bb8d5fb29c3d0a22a4df3c7c04ebeb1d1187dc437cc5587571b24848bc062a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 119, + 581, + 162 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1e78afbfb0f88373709f4c9316735357b0af95148c7ef12d7627722d8b5438e3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 598, + 119, + 671, + 165 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/43c670993759bf50271e36f064bc0c6fb7a11b729f16109fec0cd5e030704dd8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 98, + 758, + 159 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/99c5d6f186935f4fefd421ada073cef96c40885454d9ce31e8b1eb5810569f88.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 782, + 98, + 833, + 161 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6ba294a8c5df4bd1e23c6932d9ac5288d6b509d25a8a5b2d3ba0729bc26dd7a7.jpg", + "image_caption": [ + "IET" + ], + "image_footnote": [], + "bbox": [ + 147, + 180, + 204, + 236 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0bae73de0041777f7ee1b969aa33437d47c35caa9422819e2e638cc74d8139f5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 180, + 294, + 236 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ac472874acc8ebac9d6f81b5eb5f7a66221b5092c321d590518b453da52efa1e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 307, + 180, + 383, + 237 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/12158f99e369ffde7104a65f343f78a41484cfd787087238aebb4df934afc0df.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 410, + 194, + 483, + 231 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/88ca63764e1b942c822bdcdbbc9c52d2f9384ebfdb7e7eb5fb1de07c44611f82.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 199, + 581, + 237 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a2bdc168b759da5aad9855d257709cfaa3260cf877fe5ed8d811ad753b956bf0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 599, + 200, + 671, + 238 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1faffe13274ae5db24a8706a0c1e3d813cf37d12fe3be82f85a62ffd95cf4a5e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 700, + 174, + 754, + 239 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ea6131c54406dfdead26728adca6c4e08c25611ec10d19773a03cc21ce9e6d0d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 784, + 176, + 834, + 244 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1739980a2834bb31046dd9a000c112c3e99fdec541922734b2983357db39d502.jpg", + "image_caption": [ + "sno" + ], + "image_footnote": [], + "bbox": [ + 143, + 258, + 202, + 316 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/722b6727f01c40a86097a5fc6b3d21fee01b34874ceffcde11cada5439b65fe9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 230, + 258, + 290, + 316 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8efa1ddbf309bc904bdcc8e40bc31c97df75483289059040f6a284e10bf8940c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 308, + 260, + 383, + 319 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c1d2e9c85bc22f8d3d2b6ced980ebdcd46b7491489af02a1a7f647ed958b2891.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 411, + 275, + 486, + 311 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/627ec2a8d37f67b5cb84f78e360ede24e849e8641e9099b9583bc41569aebc84.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 277, + 581, + 318 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ba9517e33e3a63637b556c00f36563c0a63ea4aa1799912a2b5751db0b7353d8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 602, + 279, + 673, + 316 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1a4c047017a2ea569677c75349811c3726394ece7483574c26e3958eed412b1a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 253, + 750, + 321 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0c346b59483888665311185c00d91430dedb480506502cdf60dc0932e80e532b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 785, + 258, + 834, + 323 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/19c2f8696e111d1b4c9f6d63fe378b3cdd808708acb8183f2366ba8285e91370.jpg", + "image_caption": [ + "Grouu" + ], + "image_footnote": [], + "bbox": [ + 143, + 339, + 202, + 395 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/afdec642351463f0d6074e60d16ef556a3620a8797ceea4ff57985a7769a2e63.jpg", + "image_caption": [ + "Chair" + ], + "image_footnote": [], + "bbox": [ + 228, + 340, + 287, + 395 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7ab5fdf4dabc40bb14e1367c127c44aefe8ca066a1c5ec7da268f6301fee23f0.jpg", + "image_caption": [ + "Figure 3. The qualitative comparisons between our method and two most competitive baselines on PartNet [18]. We highlight some areas where the assembly quality of ours is clearly better." + ], + "image_footnote": [], + "bbox": [ + 308, + 340, + 383, + 396 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d48512d7f8df10e313a604ea140bdf2f0eac9426420b6977c466e2034911d677.jpg", + "image_caption": [ + "Table" + ], + "image_footnote": [], + "bbox": [ + 411, + 349, + 486, + 388 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/84447528fee5d37a3c41f3132709e45db88f8485e432ca3c63167f289205896d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 353, + 581, + 391 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/414a1b04931c331d180899687f57ff6984f4145f9322fa2e2ac1654bd96a4813.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 601, + 354, + 674, + 391 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1b23f45b2c1a881b9885b0dd01a05e30cd7687239a975d97b72dfa2a8a2825fa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 712, + 339, + 750, + 402 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/06eadb62e6faf4c9ca9148cd892598d47e5468a612596a1c4b7916748670f56b.jpg", + "image_caption": [ + "Lamp" + ], + "image_footnote": [], + "bbox": [ + 787, + 340, + 834, + 402 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/aa9bbca0fc68aa66d86d85c7d1765799b428039c0ee79ddc4e4406f8e7dc8995.jpg", + "image_caption": [ + "3" + ], + "image_footnote": [], + "bbox": [ + 130, + 481, + 196, + 534 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f7cc16ffc59690cdd255d5ddf4b95504c0e2fe0b8e691c9da25a725ca0700a0b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 222, + 482, + 271, + 536 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/662907be14770d8167de5685bfb25a14406bf262e75f5e273ba62617cca74bd2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 303, + 481, + 361, + 539 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/aff3a1d3db61d179ab0a8fa642e394fa5eb32603a8fabc4478e1db3ad6d09a49.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 385, + 481, + 450, + 541 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f559b0ae3511b22f652dde3b59d39d6963db5984fa57819b822122a3ef46c459.jpg", + "image_caption": [ + "Aeeaae 2" + ], + "image_footnote": [], + "bbox": [ + 130, + 561, + 192, + 612 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ba21a9cbc7ce0d1e2fb76659e2a183d4dec3717d7510c7ee540763dc0f3472b9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 220, + 554, + 269, + 613 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1cf5e07ab5cad61a3ba2b3accd347e4fec3326059627a5a83c9b20fc34ca969d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 303, + 553, + 359, + 613 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eac492ac95e1b1e54327be94a12117a5c47fd84d97deb925d390c91effb54ff6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 559, + 449, + 616 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/117bf18a41ad36240ccb0f8f80d9bb932bc2100c98b193af97c6ca6f8e698234.jpg", + "image_caption": [ + "aennnnnne aannnnnne", + "Figure 4. Diverse results on the unseen PartNet[18] test dataset generated by our network to demonstrate the structural variation in part assembly, providing different artistic results while maintaining reasonable object structures." + ], + "image_footnote": [], + "bbox": [ + 130, + 627, + 189, + 685 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1f0565af5277f064a2306c3c8634fcd09b5c3cbba8c16d96720580968b641785.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 220, + 627, + 269, + 686 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d1bfd76f71ff301a28bc796cf3878864398cdbc8fed06c5188f8f23f0bfeb219.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 297, + 626, + 356, + 688 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e337b53450f184be699e239f6a107408016ec5d77a57edfb80bab5cc90d7e843.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 628, + 447, + 689 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "encoder on the Chair category.", + "bbox": [ + 76, + 767, + 281, + 782 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Super-Part Encoder. Since the super-part encoder is at the core of our part-whole hierarchy message passing network, we conduct an experiment where we remove this module (denoted as w/o Super-Part Enc.). To mitigate the impact of the number of model parameters, we also create another baseline by increasing the number of parameters of the w/o", + "bbox": [ + 75, + 810, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f321efad3c61d3896280d1e60b50b10d36e1dbd0300bd12ac83f8d1b297651e0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SettingsSCD↓PA↑CA↑
w/o Super-Part Enc. + Augmented Param.0.003860.7053.46
w/o Super-Part Enc.0.003661.0455.39
Full Setting0.002864.8358.45
", + "bbox": [ + 537, + 460, + 848, + 551 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a23e510b00562bf9837875b7c1f13fc2ca27fc21791b1be5b61d70b93c711846.jpg", + "table_caption": [ + "Table 3. Ablation studies on part-whole hierarchy on the Table category under a certain Chamfer distance threshold 0.01." + ], + "table_footnote": [], + "table_body": "
LtLrLsSCD↓PA↑CA↑
✓✓0.007535.3632.73
✓✓0.003960.0254.21
✓✓0.003362.7456.98
✓✓✓0.002864.8358.45
", + "bbox": [ + 537, + 604, + 846, + 714 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Ablation studies on loss components on the Table category under a certain Chamfer distance threshold 0.01.", + "bbox": [ + 500, + 720, + 890, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Super-Part Enc. setting (denoted as w/o Super-Part Enc. + Augmented Param.). This involves increasing the number of parameters in the part encoder and the multi-head attention module, ensuring that the overall number of parameters is on the same order of magnitude as the full model setting. As shown in Table 3, the super-part encoder contributes significantly to the final performance. We also observe that augmenting the model parameters without incorporating the", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20856", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d0a5e2ee66ca2cc3e2903d93013d27c2a33ab727a5944cd5484f427dd9a12cef.jpg", + "image_caption": [ + "Figure 5. Performance on the Chair, Table and Lamp categories under multiple Chamfer distance thresholds." + ], + "image_footnote": [], + "bbox": [ + 81, + 102, + 346, + 255 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7f20821c27e88604feb16407c38088204dfa8b62ccc5aa54bcd3a73bf682d34e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 346, + 102, + 612, + 257 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d619800e14b090589a6a33359b79023d0d6a7449a4c38def3d739be7fa04f2d2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 102, + 885, + 257 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/803e4315084474964de1df8d5f820dea9199d7e64821a25b7618e1b340742a96.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPercent Prefer Ours
Ours vs IET[34]95.2%
Ours vs RGL [21]97.6%
Ours vs DGL [33]99.5%
", + "bbox": [ + 114, + 295, + 424, + 386 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Human study results on PartNet[18].", + "bbox": [ + 135, + 392, + 408, + 406 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/78d2d7c499f304c2361218850408503913e4789c3fb60a001317527d4be79150.jpg", + "image_caption": [ + "Figure 6. Visualization of our predicted hierarchical assembly process from parts to the whole." + ], + "image_footnote": [], + "bbox": [ + 81, + 420, + 460, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "super-part encoders actually leads to a performance drop. This finding further highlights the effectiveness of the proposed part-whole hierarchy network for part assembly.", + "bbox": [ + 75, + 623, + 470, + 670 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Loss Functions. We now evaluate the importance of individual loss terms in Equation (6). As shown in Table 4, removing the part translation loss $\\mathcal{L}_t$ leads to a significant performance degradation. Similarly, the remaining loss terms, including the rigid rotation loss $\\mathcal{L}_r$ and the shape assembly loss $\\mathcal{L}_s$ , demonstrate their significance in facilitating precise part rotations and ensuring the overall coherence of the assembled parts, respectively. This shows that each loss term does play an indispensable role in achieving high-quality assemblies.", + "bbox": [ + 75, + 686, + 470, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Hierarchical Part Assembly Analysis", + "text_level": 1, + "bbox": [ + 76, + 847, + 397, + 864 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We further provide a visual analysis of the assembly process for super-parts and parts in Figure 6. It is obvious that during", + "bbox": [ + 75, + 869, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the super-part assembly phase, important components such as chair seats and chair backs are often assembled correctly first. This aligns with our intuition that super-part assembly is relative easier and provides strong hints for predicting poses of parts like chair legs and arms, thus showing further evidence for the effectiveness of the part-whole-hierarchy message passing network.", + "bbox": [ + 496, + 299, + 893, + 405 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 417, + 619, + 434 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose the part-whole-hierarchy message passing network to address the challenging generative 3D part assembly task. We first group the point cloud of individual parts to form super-parts in an unsupervised way. Taking the point cloud as input, our super-part encoder predicts latent poses of super-parts which are used to transform the point cloud. We then feed the transformed point cloud to the part encoder. Relying on the cross-level and within-level attention based message passing, the part encoder takes the transformed point cloud as input and leverages the information from super-parts and predicts poses of parts. Experiments on the PartNet dataset demonstrate that our method achieves state-of-the-art performances as well as provides interpretable assembly process. In the future, we are interested in combining 3D part generation with our 3D part assembly model to generate 3D shapes from scratch.", + "bbox": [ + 496, + 443, + 893, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 500, + 704, + 633, + 718 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was funded, in part, by NSERC DG Grants (No. RGPIN-2022-04636), the Vector Institute for AI, and Canada CIFAR AI Chair. Resources used in preparing this research were provided, in part, by the Province of Ontario, the Government of Canada through the Digital Research Alliance of Canada alliance.can.ca, and companies sponsoring the Vector Institute www.vectorinstitute.ai/#partners, and Advanced Research Computing at the University of British Columbia. Additional hardware support was provided by John R. Evans Leaders Fund CFI grant.", + "bbox": [ + 496, + 726, + 895, + 878 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20857", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In International conference on machine learning, pages 40-49. PMLR, 2018. 5", + "[2] Siddhartha Chaudhuri, Evangelos Kalogerakis, Leonidas Guibas, and Vladlen Koltun. Probabilistic reasoning for assembly-based 3d modeling. In ACM SIGGRAPH 2011 papers, pages 1-10. 2011. 2", + "[3] Yun-Chun Chen, Haoda Li, Dylan Turpin, Alec Jacobson, and Animesh Garg. Neural shape mating: Self-supervised object assembly with adversarial shape priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12724-12733, 2022. 2", + "[4] Junfeng Cheng, Mingdong Wu, Ruiyuan Zhang, Guanqi Zhan, Chao Wu, and Hao Dong. Score-pa: Score-based 3d part assembly. arXiv preprint arXiv:2309.04220, 2023. 5, 6", + "[5] P Kingma Diederik, Max Welling, et al. Auto-encoding variational bayes. In Proceedings of the International Conference on Learning Representations (ICLR), 2014. 2", + "[6] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 605-613, 2017. 5", + "[7] Thomas Funkhouser, Michael Kazhdan, Philip Shilane, Patrick Min, William Kiefer, Ayellet Tal, Szymon Rusinkiewicz, and David Dobkin. Modeling by example. ACM transactions on graphics (TOG), 23(3):652-663, 2004. 1, 2", + "[8] Lin Gao, Jie Yang, Tong Wu, Yu-Jie Yuan, Hongbo Fu, Yu-Kun Lai, and Hao Zhang. Sdm-net: Deep generative network for structured deformable mesh. ACM Transactions on Graphics (TOG), 38(6):1-15, 2019. 2", + "[9] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 2", + "[10] Prakhar Jaiswal, Jinmiao Huang, and Rahul Rai. Assembly-based conceptual 3d modeling with unlabeled components using probabilistic factor graph. Computer-Aided Design, 74: 45-54, 2016. 2", + "[11] R Kenny Jones, Theresa Barton, Xianghao Xu, Kai Wang, Ellen Jiang, Paul Guerrero, Niloy J Mitra, and Daniel Ritchie. Shapeassembly: Learning to generate programs for 3d shape structure synthesis. ACM Transactions on Graphics (TOG), 39(6):1-20, 2020. 2", + "[12] Evangelos Kalogerakis, Siddhartha Chaudhuri, Daphne Koller, and Vladlen Koltun. A probabilistic model for component-based shape synthesis. Acm Transactions on Graphics (TOG), 31(4):1-11, 2012. 2", + "[13] Jun Li, Kai Xu, Siddhartha Chaudhuri, Ersin Yumer, Hao Zhang, and Leonidas Guibas. Grass: Generative recursive autoencoders for shape structures. ACM Transactions on Graphics (TOG), 36(4):1-14, 2017. 2", + "[14] Yichen Li, Kaichun Mo, Lin Shao, Minhyuk Sung, and Leonidas Guibas. Learning 3d part assembly from a single image. In Computer Vision-ECCV 2020: 16th European" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI 16, pages 664-682. Springer, 2020. 1, 2, 5, 6", + "[15] Yuval Litvak, Armin Biess, and Aharon Bar-Hillel. Learning pose estimation for high-precision robotic assembly using simulated depth images. In 2019 International Conference on Robotics and Automation (ICRA), pages 3521–3527. IEEE, 2019. 1", + "[16] Jianlan Luo, Eugen Solowjow, Chengtao Wen, Juan Aparicio Ojea, Alice M Agogino, Aviv Tamar, and Pieter Abbeel. Reinforcement learning on variable impedance controller for high-precision robotic assembly. In 2019 International Conference on Robotics and Automation (ICRA), pages 3080–3087. IEEE, 2019. 1", + "[17] Kaichun Mo, Paul Guerrero, Li Yi, Hao Su, Peter Wonka, Niloy Mitra, and Leonidas J Guibas. Structurenet: Hierarchical graph networks for 3d shape generation. arXiv preprint arXiv:1908.00575, 2019. 2", + "[18] Kaichun Mo, Shilin Zhu, Angel X Chang, Li Yi, Subarna Tripathi, Leonidas J Guibas, and Hao Su. Partnet: A large-scale benchmark for fine-grained and hierarchical part-level 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 909–918, 2019. 2, 5, 6, 7, 8", + "[19] Kaichun Mo, He Wang, Xinchen Yan, and Leonidas Guibas. Pt2pc: Learning to generate 3d point cloud shapes from part tree conditions. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI 16, pages 683-701. Springer, 2020. 2, 5", + "[20] Carsten Moenning and Neil A Dodgson. Fast marching farthest point sampling. Technical report, University of Cambridge, Computer Laboratory, 2003. 5", + "[21] Abhinav Narayan, Rajendra Nagar, and Shanmuganathan Raman. Rgl-net: A recurrent graph learning framework for progressive part assembly. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 78-87, 2022. 1, 2, 4, 5, 6, 8", + "[22] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 3", + "[23] Olinde Rodrigues. Des lois géométriques qui regissant les déplacements d'un système solide dans l'espace, et de la variation des coordonnées provenant de ces déplacements considérés indépendamment des causes qui peuvent les produit. Journal de mathématiques pures et appliquées, 5:380-440, 1840. 4", + "[24] Nadav Schor, Oren Katzir, Hao Zhang, and Daniel Cohen-Or. Componet: Learning to generate the unseen by part synthesis and composition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8759–8768, 2019. 5, 6", + "[25] Lin Shao, Toki Migimatsu, and Jeannette Bohg. Learning to scaffold the development of robotic manipulation skills. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 5671-5677. IEEE, 2020. 1", + "[26] Dong Wook Shu, Sung Woo Park, and Junseok Kwon. 3d point cloud generative adversarial network based on tree struc" + ], + "bbox": [ + 503, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20858", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tured graph convolutions. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3859-3868, 2019. 5", + "[27] Minhyuk Sung, Hao Su, Vladimir G Kim, Siddhartha Chaudhuri, and Leonidas Guibas. Complementme: Weakly-supervised component suggestions for 3d modeling. ACM Transactions on Graphics (TOG), 36(6):1-12, 2017. 5, 6", + "[28] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2, 4", + "[29] Rundi Wu, Yixin Zhuang, Kai Xu, Hao Zhang, and Baoquan Chen. Pq-net: A generative part seq2seq network for 3d shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 829-838, 2020. 2, 5, 6", + "[30] Yuwen Xiong, Wei-Chiu Ma, Jingkang Wang, and Raquel Urtasun. Learning compact representations for lidar completion and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1074-1083, 2023. 6", + "[31] Xianghao Xu, Paul Guerrero, Matthew Fisher, Siddhartha Chaudhuri, and Daniel Ritchie. Unsupervised 3d shape reconstruction by part retrieval and assembly. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8559-8567, 2023. 2", + "[32] Kevin Zakka, Andy Zeng, Johnny Lee, and Shuran Song. Form2fit: Learning shape priors for generalizable assembly from disassembly. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 9404–9410. IEEE, 2020. 1", + "[33] Guanqi Zhan, Qingnan Fan, Kaichun Mo, Lin Shao, Baoquan Chen, Leonidas J Guibas, Hao Dong, et al. Generative 3d part assembly via dynamic graph learning. Advances in Neural Information Processing Systems, 33:6315-6326, 2020. 1, 2, 3, 4, 5, 6, 8", + "[34] Rufeng Zhang, Tao Kong, Weihao Wang, Xuan Han, and Mingyu You. 3d part assembly generation with instance encoded transformer. IEEE Robotics and Automation Letters, 7(4):9051-9058, 2022. 1, 2, 3, 4, 5, 6, 8", + "[35] Vlas Zyrianov, Xiyue Zhu, and Shenlong Wang. Learning to generate realistic lidar point clouds. In European Conference on Computer Vision, pages 17-35. Springer, 2022. 6" + ], + "bbox": [ + 78, + 90, + 470, + 699 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20859", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Generative 3D Part Assembly via Part-Whole-Hierarchy Message Passing/40f117f5-c539-42c1-b970-aab5fe53da07_model.json b/2024/Generative 3D Part Assembly via Part-Whole-Hierarchy Message Passing/40f117f5-c539-42c1-b970-aab5fe53da07_model.json new file mode 100644 index 0000000000000000000000000000000000000000..477f3637dc8a3c410a3b4c37b35a7db924eef1d8 --- /dev/null +++ b/2024/Generative 3D Part Assembly via Part-Whole-Hierarchy Message Passing/40f117f5-c539-42c1-b970-aab5fe53da07_model.json @@ -0,0 +1,2519 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.131, + 0.856, + 0.154 + ], + "angle": 0, + "content": "Generative 3D Part Assembly via Part-Whole-Hierarchy Message Passing" + }, + { + "type": "text", + "bbox": [ + 0.286, + 0.18, + 0.684, + 0.199 + ], + "angle": 0, + "content": "Bi'an Du1, Xiang Gao1, Wei Hu1*, Renjie Liao2,3,4" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.199, + 0.743, + 0.218 + ], + "angle": 0, + "content": "\\(^{1}\\)Wangxuan Institute of Computer Technology, Peking University" + }, + { + "type": "text", + "bbox": [ + 0.155, + 0.217, + 0.815, + 0.234 + ], + "angle": 0, + "content": "\\(^{2}\\)University of British Columbia, \\(^{3}\\)Vector Institute for AI, \\(^{4}\\)Canada CIFAR AI Chair" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.238, + 0.85, + 0.251 + ], + "angle": 0, + "content": "pkudba@stu.pku.edu.cn, gyshgx868@pku.edu.cn, forhuwei@pku.edu.cn, rjliao@ece.ubc.ca" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.474, + 0.668 + ], + "angle": 0, + "content": "Generative 3D part assembly involves understanding part relationships and predicting their 6-DoF poses for assembling a realistic 3D shape. Prior work often focuses on the geometry of individual parts, neglecting part-whole hierarchies of objects. Leveraging two key observations: 1) super-part poses provide strong hints about part poses, and 2) predicting super-part poses is easier due to fewer superparts, we propose a part-whole-hierarchy message passing network for efficient 3D part assembly. We first introduce super-parts by grouping geometrically similar parts without any semantic labels. Then we employ a part-whole hierarchical encoder, wherein a super-part encoder predicts latent super-part poses based on input parts. Subsequently, we transform the point cloud using the latent poses, feeding it to the part encoder for aggregating super-part information and reasoning about part relationships to predict all part poses. In training, only ground-truth part poses are required. During inference, the predicted latent poses of super-parts enhance interpretability. Experimental results on the PartNet dataset show that our method achieves state-of-the-art performance in part and connectivity accuracy and enables an interpretable hierarchical part assembly. Code is available at https://github.com/pkudba/3DHPA." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.698, + 0.21, + 0.714 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.724, + 0.472, + 0.876 + ], + "angle": 0, + "content": "Generative 3D Part Assembly [7, 15, 16, 25, 32] is an emerging research area that aims to generate complex 3D shapes via assembling simple 3D parts without relying on prior semantic knowledge. Different from traditional 3D shape generation, it focuses on generating diverse, plausible configurations of given parts. It facilitates the generation of complex objects and scenes with compositionality, flexibility, and efficiency. With the rapid evolution of 3D printing technology and the increasing demand for diverse 3D shapes, generative 3D part assembly finds applications in various sce" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.289, + 0.893, + 0.319 + ], + "angle": 0, + "content": "narios, attracting attention from experts in computer vision, graphics, robotics, and machine learning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.321, + 0.895, + 0.427 + ], + "angle": 0, + "content": "However, achieving generative 3D part assembly presents a significant challenge due to the vast number of potential part arrangements and orientations, coupled with the intricate dependencies among the parts. Adding to the complexity, part geometry exhibits notable variation even within the same object category, making it exceedingly difficult to generalize learned assembly patterns across different objects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.429, + 0.896, + 0.701 + ], + "angle": 0, + "content": "Previous approaches primarily focus on designing architectures capable of learning powerful representations for individual 3D parts. The hope is that these learned representations could facilitate accurate part assembly, either in a one-shot manner or sequentially [14, 21, 33, 34]. However, these methods often ignore the inherent part-whole hierarchies in 3D shapes in representation learning. For instance, as illustrated in Figure 1, a chair consists of super-parts such as seats, backs, and legs, with each super-part further divisible into parts like seat surfaces and seat frames. Understanding the pose of a super-part provides insights into the poses of constituent parts within the same super-part, as they often share similar orientations or exhibit symmetry (e.g., left-right symmetry in chair legs and arms). Moreover, predicting super-part poses is typically easier due to the fewer number of super-parts compared to parts. Incorporating these hierarchies into the modeling would make the learning process and potentially improve the overall performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.897, + 0.869 + ], + "angle": 0, + "content": "In this paper, we introduce a part-whole-hierarchy message passing network for 3D shape assembly, predicting 6 degrees of freedom (6-DoF) poses for super-parts and parts in a hierarchical manner. We establish the correspondence between parts and super-parts (i.e., subsets of parts) by grouping parts based on their geometric similarities, following the approach from [33, 34]. Importantly, we treat super-part poses as latent variables to be learned, thereby eliminating the requirement of ground-truth super-part poses in our work. Our model comprises two sequential modules: the super-part encoder and the part encoder." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.895, + 0.902 + ], + "angle": 0, + "content": "The first super-part encoder takes the 3D point clouds of all parts as input and employs the attention-based message" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.405, + 0.901 + ], + "angle": 0, + "content": "*Corresponding Author: Wei Hu (forhuwei@pku.edu.cn)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20850" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.093, + 0.882, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.299, + 0.895, + 0.328 + ], + "angle": 0, + "content": "Figure 1. Left: an illustration of the part-whole-hierarchy for 3D shapes; Right: the part assembly process via the proposed part-whole-hierarchy message passing network." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.343, + 0.471, + 0.403 + ], + "angle": 0, + "content": "passing to predict super-parts poses. The representation of each super-part is aggregated from representations of parts within it. These super-part poses provide initial estimation for the poses of individual parts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.404, + 0.472, + 0.509 + ], + "angle": 0, + "content": "Subsequently, the whole point clouds are transformed based on the predicted super-part poses and fed to the part encoder. This module, using a cross-attention mechanism, extracts features from the transformed point cloud, transferring super-part level information to the part level. It then leverages the attention-based message passing again to capture relationships among individual parts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.511, + 0.472, + 0.678 + ], + "angle": 0, + "content": "Following previous works [14, 21, 33, 34], we train and evaluate our model on the PartNet dataset [18]. Overall, our model achieves the state-of-the-art performances, outperforming the strongest competitor [34] by a significant margin, i.e., with an almost \\(2\\%\\) improvement in mean part accuracy and a \\(3\\%\\) improvement in mean connectivity accuracy. Moreover, through visual analysis of the assembly process for both super-parts and parts, we not only showcase accurate generation of part poses but also demonstrate interpretability via the predicted super-part poses. This interpretability feature further improves the utility of our model." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.693, + 0.22, + 0.71 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.473, + 0.901 + ], + "angle": 0, + "content": "Assembly3D modeling Numerous prior studies have approached the challenging 3D part assembly through the joint estimation of part poses. The influential work by [7] introduces an intelligent scissoring technique to tackle this problem for part components. Subsequent research by [2, 10, 12] employs graphical models to capture the semantic and geometric relationships among shape components, enabling exploration in assembly-based shape modeling. PAGENet [14] presents a network that is aware of individual parts and generates semantic parts along with their poses. [31] proposes to decompose shapes using a library of 3D parts provided by the user. However, these studies either assumed prior knowl-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.343, + 0.895, + 0.525 + ], + "angle": 0, + "content": "edge of part semantics or relied on existing shape databases. In a more practical setting, the authors in [3, 21, 33, 34] focus on the pose estimation of individual parts without relying on shape databases or known semantic information. Specifically, DGL [33] employs a dynamic part graph to iteratively refine the poses of individual parts. A progressive strategy using the recurrent graph learning framework has been investigated in [21]. Additionally, the authors in [3, 34] utilizes the Transformer [28] to model the structural relationships and performs the simultaneous assembly of all parts. We follow this line of work and propose novel improvement to generate structurally-coherent part assemblies." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.544, + 0.897, + 0.832 + ], + "angle": 0, + "content": "Structural Shape Generation In recent years, deep generative models, such as generative adversarial networks (GANs) [9] and variational autoencoders (VAEs) [5], have garnered significant attention for shape generation tasks. Notably, the work by [8] introduces a two-level variational autoencoder that simultaneously learns the overall shape structure and detailed part geometries. Both GRASS [13] and StructureNet [17] employ techniques to compress the shape structure into a latent space while considering the relationships between different parts. Additionally, [19] uses a part-tree decomposition to conditionally generate 3D shapes, and [11] adopts a procedural programmatic representation to establish connections between part cuboids. Inspired by the Seq2Seq networks in machine translation, [29] introduces a sequential encoding and decoding approach for the regression of shape parameters. While many of these approaches focus on directly generating new part shapes given random latent codes, our main focus revolves around the rigid transformation of existing parts to facilitate their assembly." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.845, + 0.859, + 0.863 + ], + "angle": 0, + "content": "3. Part-Whole-Hierarchy Message Passing" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.893, + 0.903 + ], + "angle": 0, + "content": "This section provides a detailed explanation of our proposed part-whole-hierarchy message passing network. We begin by" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20851" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.089, + 0.882, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.395, + 0.895, + 0.452 + ], + "angle": 0, + "content": "Figure 2. The overall architecture of our model consists of two modules: super-part encoder and part encoder. We first obtain super-parts via grouping parts based on their geometric similarities in an unsupervised fashion. The super-part encoder takes point cloud as input and predicts latent super-part poses (no ground truth is needed). The point cloud is then transformed based on super-part poses and fed to the part encoder. We incorporate both cross-level and within-level attention in the part encoder to predict part poses." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.466, + 0.473, + 0.573 + ], + "angle": 0, + "content": "presenting how we construct super-parts in an unsupervised manner. Then we introduce our super-part encoder that predicts the latent poses of super-parts. Subsequently, we describe how our part encoder leverages the latent poses and the part-whole hierarchy to predict part poses. The overall model is illustrated in Figure 2. Finally, we introduce the loss functions and the training process for diverse generation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.587, + 0.307, + 0.604 + ], + "angle": 0, + "content": "3.1. Super-Part Construction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.612, + 0.472, + 0.839 + ], + "angle": 0, + "content": "We denote the input point clouds as \\(\\mathcal{P} = \\{\\mathbf{P}_i|i = 1\\ldots N\\}\\), where \\(\\mathbf{P}_i\\in \\mathbb{R}^{d\\times 3}\\) corresponds to the point cloud of the \\(i\\)-th given part of the 3D shape. The values of \\(N\\) and \\(d\\) represent the number of parts and the number of points per part respectively. Note that their values may differ from object to object. For notation convenience, we assume objects are padded to the maximum number of parts, and each part is padded to the maximum number of points per part. The goal of our task is to predict a set of 6-DoF part poses \\(\\{(t_i,r_i)\\}_{i = 1}^N\\), where \\(t_i\\in \\mathbb{R}^3\\) and \\(r_i\\in \\mathbb{R}^4\\) represent the translation and rigid rotation for each part, respectively. The complete part assembly for a 3D shape is \\(\\mathcal{S} = \\cup_{i = 1}^{N}\\mathbf{T}_i(\\mathbf{P}_i)\\), and \\(\\mathbf{T}_i\\) represents a transformation in \\(SE(3)\\) that consists of a 3D rotation in the rotation group \\(SO(3)\\) and a 3D translation in the translation group, induced by \\((t_i,r_i)\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.473, + 0.903 + ], + "angle": 0, + "content": "A super-part is a subset of parts that is ideally semantically meaningful. However, since we do not have groundtruth super-parts, we need to construct them in an unsupervised manner. In particular, we compute axis-aligned" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.466, + 0.895, + 0.711 + ], + "angle": 0, + "content": "bounding boxes for each part and evaluate the similarity between these 3D boxes. Parts are grouped into the same super-part if the difference between their respective enclosing boxes is below a specified threshold. Following previous works [33, 34], we set the threshold to 0.2. Although the super-parts obtained through this method may lack semantic meaning, they provide a coarse abstraction grounded in geometry similarities without any supervised labels. Then we group the input point cloud parts \\(\\mathcal{P} = \\{\\mathbf{P}_i\\}_{i=1}^N\\) into a set of \\(M\\) super-parts \\(\\mathcal{P}' = \\{\\mathbf{P}_i'\\}_{i=1}^M\\) based on the part-whole hierarchy of a given object. Here \\(\\mathbf{P}_i'\\) represents the \\(i\\)-th super-part. To maintain consistent notation, we again pad the number of parts per super-part to the maximum \\(M\\). Note that padding is used here for clarity and understanding. In practice, our model can handle sets of parts with varying sizes without padding." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.724, + 0.779, + 0.741 + ], + "angle": 0, + "content": "3.2. Super-Part Relation Reasoning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.749, + 0.892, + 0.794 + ], + "angle": 0, + "content": "To reason about the relationships among super-parts, we utilize the super-part encoder. The computational process is illustrated in the top half of Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Given a set of point clouds of a super-part \\(\\mathbf{P}_i^{\\prime}\\), we first compute the set-level permutation-invariant representation \\(\\mathbf{F}_i^{\\prime}\\) via a shared PointNet [22]. We use the same architecture and hyperparameters as in previous works [33, 34] for a fair comparison. This feature representation captures the global characteristics of each super-part while being invariant to the order of the parts within the set. Based on the features" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20852" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.226 + ], + "angle": 0, + "content": "\\(\\mathbf{F}_i^{\\prime}\\) extracted for each super-part from the PointNet, we use a two-layer multi-head attention (MHA) module to learn the relationships between super-parts. Specifically, following [28], we calculate an attention matrix \\(\\mathbf{A}^{\\prime}\\), where \\(\\mathbf{A}_{i,j}^{\\prime}\\) represents the attention weight of the \\(i\\)-th super-part to the \\(j\\)-th super-part. Then we multiply each super-part feature \\(\\mathbf{F}_j^{\\prime}\\) by \\(\\mathbf{A}_{i,j}^{\\prime}\\) and sum them up to obtain the attention-weighted feature \\(\\mathbf{G}_i^{\\prime}\\). In our model, we set the embedding dimension to 256 and the number of heads for the MHA module to 8." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.227, + 0.474, + 0.356 + ], + "angle": 0, + "content": "For super-part pose prediction, we feed the attention-weighted feature \\(\\mathbf{G}_i'\\) into a prediction head containing 4 fully-connected layers to obtain the pose \\(\\{(t_i', r_i')\\}_{i=1}^M\\). We apply tanh operation to the translation vector, which restricts the range of the part center offset as \\((-1,1)\\). Additionally, for simplicity, we predict the quaternion vector \\(r_i' = (r_{i0}', r_{i1}', r_{i2}', r_{i3}')\\) instead of the rotation matrix. We then use the Rodrigues formula [23] to obtain the rotation matrix \\(\\mathbf{R}_i'\\) corresponding to \\(r_i'\\) one by one as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.366, + 0.476, + 0.413 + ], + "angle": 0, + "content": "\\[\n\\mathbf {R} _ {i} ^ {\\prime} = \\left[ \\begin{array}{l} 1 - 2 r _ {i 2} ^ {\\prime 2} - 2 r _ {i 3} ^ {\\prime 2}, 2 r _ {i 1} ^ {\\prime} r _ {i 2} ^ {\\prime} - 2 r _ {i 0} ^ {\\prime} r _ {i 3} ^ {\\prime}, 2 r _ {i 1} ^ {\\prime} r _ {i 3} ^ {\\prime} + 2 r _ {i 0} ^ {\\prime} r _ {i 2} ^ {\\prime} \\\\ 2 r _ {i 1} ^ {\\prime} r _ {i 2} ^ {\\prime} + 2 r _ {i 0} ^ {\\prime} r _ {i 3} ^ {\\prime}, 1 - 2 r _ {i 1} ^ {\\prime 2} - 2 r _ {i 3} ^ {\\prime 2}, 2 r _ {i 2} ^ {\\prime} r _ {i 3} ^ {\\prime} - 2 r _ {i 0} ^ {\\prime} r _ {i 1} ^ {\\prime} \\\\ 2 r _ {i 1} ^ {\\prime} r _ {i 3} ^ {\\prime} - 2 r _ {i 0} ^ {\\prime} r _ {i 2} ^ {\\prime}, 2 r _ {i 2} ^ {\\prime} r _ {i 3} ^ {\\prime} + 2 r _ {i 0} ^ {\\prime} r _ {i 1} ^ {\\prime}, 1 - 2 r _ {i 1} ^ {\\prime 2} - 2 r _ {i 2} ^ {\\prime 2}. \\end{array} \\right]\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.421, + 0.471, + 0.468 + ], + "angle": 0, + "content": "The transformed point cloud of the super-part is expressed as \\(\\mathbf{T}_i(\\mathbf{P}_i^{\\prime}) = \\mathbf{R}_i^{\\prime}\\mathbf{P}_i^{\\prime} + t_i^{\\prime}\\). To ensure that it is a unit quaternion, we normalize the rigid rotation vector such that \\(\\| r_i^\\prime \\| = 1\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.468, + 0.472, + 0.527 + ], + "angle": 0, + "content": "Since we do not have ground-truth super-part poses, the supervision for this module solely comes from the backpropagation signal from the subsequent module. As a result, the super-part poses are considered as latent variables." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.537, + 0.303, + 0.554 + ], + "angle": 0, + "content": "3.3. Part Relation Reasoning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.561, + 0.473, + 0.622 + ], + "angle": 0, + "content": "As previously mentioned, the poses of super-parts offer valuable clues about the poses of their corresponding parts. To leverage this information, we design a part encoder to incorporate the latent super-part poses to predict the part poses." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.64, + 0.473, + 0.79 + ], + "angle": 0, + "content": "Transformation via Latent Super-Part Poses Firstly, we apply the latent super-part poses to transform all point cloud parts, resulting in transformed parts denoted as \\(\\hat{\\mathcal{P}} = \\{\\hat{\\mathbf{P}}_i\\}_{i=1}^N\\). These transformed parts are then fed into the part encoder for further processing. Similarly to the super-part encoder, we employ another PointNet to extract part-level feature denoted as \\(\\hat{\\mathbf{F}}_i\\). In order to better utilize the part-whole hierarchy of objects, we design two ways to integrate the super-part level information to the part level, i.e., cross-level and within-level attention modules." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Cross-Level Attention Part representations \\(\\hat{\\mathbf{F}}_i\\) from the PointNet encoder does not leverage the part-whole hierarchy. To better fuse the information from the super-parts, we propose a cross-level attention module. Specifically, we treat the part representation \\(\\hat{\\mathbf{F}}_i\\) as query and super-part representations \\(\\mathbf{F}_i'\\) as keys and values. We then employ a MHA module" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.182 + ], + "angle": 0, + "content": "to perform an attention-weighted aggregation of super-part representations \\(\\mathbf{F}_i^{\\prime}\\) to obtain the updated part representations \\(\\mathbf{F}_i\\). By doing so, the information from part-whole hierarchy is explicitly integrated into part representations. Here we specify the embedding dimension as 256 and the MHA module is set to utilize 8 heads." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.201, + 0.897, + 0.518 + ], + "angle": 0, + "content": "Within-Level Attention To model the part relationships, we leverage another within-level attention module, which is depicted in the bottom half of Figure 2. We first concatenate the representation from the previous module \\( F_{i} \\) with an additional instance encoding vector that is unique for each part following [34]. This allows us to distinguish geometrically-equivalent parts. Then we perform message passing among parts using another MHA module where the attention computation is aware of the part-whole hierarchy. Specifically, we concatenate the latent super-part poses to the part-level feature before computing the attention scores at each layer. Consequently, in each message passing layer, a part receives pose hints from its parent super-part and updates its representation. The resulting part feature is denoted as \\( \\mathbf{G}_i = \\sum_{j=1}^N \\mathbf{A}_{i,j} \\mathbf{F}_i \\), where \\( \\mathbf{A}_{i,j} \\) represents the attention weight of the \\( i \\)-th part to the \\( j \\)-th part. This design enables the part-to-part message passing to effectively leverage the part-whole hierarchy information. We set the number of heads for the MHA module to 8, the dimension of the part-level feature to 256, and concatenate it with the 40-dimensional instance encoding vector." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.536, + 0.895, + 0.732 + ], + "angle": 0, + "content": "Prediction Head Finally, we feed the attention-weighted part-level feature \\(\\mathbf{G}_i\\) to a part pose prediction head, which consists of four fully-connected layers. The channel sizes of the first three fully-connected layers are set to 256, 256 and 1024, respectively, followed by ReLU activation functions. We use a linear projection layer to predict the part pose \\((t_i, r_i)\\), encompassing a 4-dimensional rigid rotation and 3-dimensional translation. This prediction head outputs the part poses \\(\\{(t_i, r_i)\\}_{i=1}^N\\), where \\(t_i\\) represents the translation vector and \\(r_i\\) represents the rigid rotation vector for each part. We adopt the same design as in the super-part encoder to restrict the translation vectors to the range of \\((-1, 1)\\) and normalize the rigid rotation vector \\(r_i\\) to have a unit norm." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.741, + 0.862, + 0.758 + ], + "angle": 0, + "content": "3.4. Training Objective for Diverse Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Considering the same set of input point cloud parts, geometrically equivalent parts (such as legs of a chair) can be interchanged and decorated parts may have multiple placement options, resulting in multiple possible shapes. For example, a semi-cylindrical part can be placed on top of a chair backrest as a headrest, or placed under the backrest as a lumbar pillow. To account for such diversified structural variations and configurations, we introduce random noise into the system following [21, 33, 34] and employ the Min-of-N" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20853" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.212 + ], + "angle": 0, + "content": "(MoN) loss [6] to simultaneously consider assembly accuracy and assembly diversity. By considering the minimum of a set of N possible assemblies, we can encourage the model to generate diverse and valid shape configurations while maintaining overall accuracy in the assembly process. This helps address the challenge of capturing multiple plausible solutions that arise from the interchangeable and decoratable nature of the parts." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.212, + 0.472, + 0.243 + ], + "angle": 0, + "content": "Let \\(\\mathcal{F}(\\mathcal{P},\\delta_j)\\) denote our network outputs and \\(\\mathcal{F}^* (\\mathcal{P})\\) denote the ground-truth point clouds, then the MoN loss is:" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.249, + 0.47, + 0.274 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {M o N}} = \\min _ {\\delta_ {j} \\sim \\mathcal {N} (0, 1)} \\mathcal {L} \\left(\\mathcal {F} \\left(\\mathcal {P}, \\delta_ {j}\\right), \\mathcal {F} ^ {*} (\\mathcal {P})\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.28, + 0.472, + 0.355 + ], + "angle": 0, + "content": "where \\(\\delta_{j}\\) is a random noise vector drawn from standard Normal distribution in an IID fashion. Following [34], we sample 5 random vectors \\(\\delta_{j}\\) during training. The loss function \\(\\mathcal{L}\\) consists of the following components for both local part and global shape losses." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.355, + 0.47, + 0.4 + ], + "angle": 0, + "content": "Firstly, we supervise the translation via an \\( l_{2} \\) loss between our prediction \\( t_{i} \\) and the ground-truth translation vector \\( t_{i}^{*} \\) for each part:" + }, + { + "type": "equation", + "bbox": [ + 0.203, + 0.4, + 0.47, + 0.44 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {t} = \\sum_ {i = 1} ^ {N} \\| t _ {i} - t _ {i} ^ {*} \\| _ {2} ^ {2} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.443, + 0.47, + 0.473 + ], + "angle": 0, + "content": "Then we compute the Chamfer distance [6] between the predicted and the ground-truth rigid rotation for each part:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.479, + 0.47, + 0.519 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r} = \\sum_ {i = 1} ^ {N} d _ {c} \\left(r _ {i} \\left(\\mathbf {P} _ {i}\\right), r _ {i} ^ {*} \\left(\\mathbf {P} _ {i}\\right)\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.526, + 0.47, + 0.601 + ], + "angle": 0, + "content": "The Chamfer distance \\( d_{c}(\\mathcal{X},\\mathcal{Y}) \\) is a metric commonly used in point cloud comparison, which measures the dissimilarity between two point sets \\( \\mathcal{X} \\) and \\( \\mathcal{Y} \\) by calculating the average distance between each point in one set and its nearest neighbor in the other set as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.607, + 0.47, + 0.64 + ], + "angle": 0, + "content": "\\[\nd _ {c} (\\mathcal {X}, \\mathcal {Y}) = \\sum_ {x \\in \\mathcal {X}} \\min _ {y \\in \\mathcal {Y}} \\| x - y \\| _ {2} ^ {2} + \\sum_ {y \\in \\mathcal {Y}} \\min _ {x \\in \\mathcal {X}} \\| x - y \\| _ {2} ^ {2} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.646, + 0.47, + 0.691 + ], + "angle": 0, + "content": "In this case, we use the Chamfer distance to assess the discrepancy between the predicted rotation of each part and its corresponding ground-truth rotation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.691, + 0.47, + 0.736 + ], + "angle": 0, + "content": "Similarly, the full shape \\(S\\), i.e., the set of point clouds belonging to all parts, is supervised via the Chamfer distance from the ground-truth shape \\(S^{*}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.217, + 0.743, + 0.47, + 0.76 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s} = d _ {c} \\left(\\mathcal {S}, \\mathcal {S} ^ {*}\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.767, + 0.434, + 0.781 + ], + "angle": 0, + "content": "In summary, the overall loss function is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.183, + 0.788, + 0.47, + 0.804 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {t} \\mathcal {L} _ {t} + \\lambda_ {r} \\mathcal {L} _ {r} + \\lambda_ {s} \\mathcal {L} _ {s} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.472, + 0.902 + ], + "angle": 0, + "content": "where \\(\\lambda_{t},\\lambda_{r}\\) and \\(\\lambda_{s}\\) represent the weights assigned to the three loss terms. Based on the cross validation, we empirically set \\(\\lambda_t = 1\\) \\(\\lambda_r = 10\\) ,and \\(\\lambda_s = 1\\) in our experiments. By minimizing this loss function, we aim to improve the alignment of the predicted poses with the ground-truth poses, thereby enhancing the overall accuracy of the assembly." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.634, + 0.108 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.116, + 0.894, + 0.193 + ], + "angle": 0, + "content": "In this section, we demonstrate the effectiveness of the proposed model by comparing our results with state-of-the-art methods. We also provide visual analysis of the assembly process for super-parts and parts, which helps to illustrate the hierarchical assembly process from parts to the whole." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.204, + 0.596, + 0.218 + ], + "angle": 0, + "content": "4.1. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.227, + 0.895, + 0.41 + ], + "angle": 0, + "content": "Following [4, 14, 21, 24, 27, 29, 33, 34], we use the PartNet [18] dataset for both training and evaluation. This dataset consists of 26,671 shapes across 24 different 3D object categories. To effectively validate and compare different methods, we select the three largest categories, i.e., chairs, tables and lamps, following [4, 14, 21, 33, 34]. In total, we have 6,323 chairs, 8,218 tables, 2,207 lamps in the finest-grained level, and we adopt the official train/validation/test splits (70%/10%/20%) to conduct the experiments. For each part point cloud, 1000 points are sampled from the original part meshes using Farthest Point Sampling (FPS) [20]. All parts are transformed into its canonical space using PCA." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.421, + 0.684, + 0.435 + ], + "angle": 0, + "content": "4.2. Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.445, + 0.895, + 0.642 + ], + "angle": 0, + "content": "We generate a variety of shapes by adding different Gaussian noises to a given set of input parts, and find the closest shape to the ground-truth using minimum matching distance (MMD) [1]. To measure part assembly quality, we follow [21, 33, 34] to use shape Chamfer distance (SCD), part accuracy (PA) and connectivity accuracy (CA). Shape Chamfer distance is defined in Equation (4) and Equation (5). Meanwhile, to compare the diversity of assembled parts, we propose two quantitative evaluation metrics including quality-diversity score (QDS) and the weighted quality-diversity score (WQDS) following [4]. We also visualize the generation results to qualitatively evaluate the diversity. More details of these metrics are provided in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.664, + 0.894, + 0.771 + ], + "angle": 0, + "content": "Mean Part/Connectivity Accuracy (mPA/mCA). PA and CA depend on the Chamfer distance threshold \\(\\tau_{p}\\) and \\(\\tau_{c}\\) for judging whether the assembly is accurate. In order to provide a more comprehensive evaluation of the performance of the model, we average the results under multiple thresholds to get mean part accuracy (mPA) and mean connectivity accuracy (mCA), formally," + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.782, + 0.882, + 0.82 + ], + "angle": 0, + "content": "\\[\n\\mathrm {m P A} = \\frac {1}{T _ {p}} \\sum_ {\\tau_ {p} \\in T _ {p}} \\mathrm {P A} (\\tau_ {p}), \\quad \\mathrm {m C A} = \\frac {1}{T _ {c}} \\sum_ {\\tau_ {c} \\in T _ {c}} \\mathrm {C A} (\\tau_ {c}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.833, + 0.874, + 0.849 + ], + "angle": 0, + "content": "where \\(T_{p}\\) and \\(T_{c}\\) are set to \\(\\{0.01, 0.02, 0.03, 0.04, 0.05\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.869, + 0.894, + 0.904 + ], + "angle": 0, + "content": "QDS. Diversity score (DS) [19, 26] evaluates the diversity of the results as \\(\\mathrm{DS} = \\frac{1}{N^2}\\sum_{i,j=1}^{N}(d_c(\\mathbf{P}_i^*,\\mathbf{P}_j^*))\\), where \\(\\mathbf{P}_i^*\\)" + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20854" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.894, + 0.23 + ], + "angle": 0, + "content": "
SCD(10-2) ↓PA(%) ↑CA(%) ↑QDS(10-5) ↑WQDS(10-5) ↑
MethodsChairTableLampChairTableLampChairTableLampChairTableLampChairTableLamp
B-Global [14, 24]1.461.120.7915.7015.3722.619.9033.8418.600.150.200.761.251.400.58
B-LSTM [29]1.311.250.7721.7728.6420.786.8022.5614.050.040.270.631.071.431.54
B-Complement [27]2.412.981.508.782.3212.679.1915.5726.560.090.062.811.281.752.08
DGL [33]0.910.500.9339.0049.5133.3323.8739.9641.701.693.051.841.352.971.73
Score [4]0.710.421.1144.5152.7834.3230.3240.5949.073.369.176.831.703.812.82
RGL [21]0.870.480.7249.0654.1637.5632.2642.1557.343.557.636.822.124.072.96
IET [34]0.540.351.0362.8061.6738.6848.4556.1862.624.159.096.982.744.563.29
Ours0.510.280.7064.1364.8338.8049.2858.4564.165.629.587.123.064.813.90
" + }, + { + "type": "table_caption", + "bbox": [ + 0.169, + 0.238, + 0.798, + 0.252 + ], + "angle": 0, + "content": "Table 1. Comparison between our approach and other methods under the Chamfer distance threshold 0.01." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.265, + 0.47, + 0.354 + ], + "angle": 0, + "content": "
mPA ↑mCA ↑
MethodsChairTableLampChairTableLamp
DGL [33]45.8457.8648.3229.1743.8851.75
IET [34]74.9273.2052.8062.3768.4975.94
Ours76.7976.1953.3165.3270.2678.15
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.361, + 0.47, + 0.39 + ], + "angle": 0, + "content": "Table 2. Comparison between our approach and other methods under multiple Chamfer distance thresholds from 0.01 to 0.05." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.407, + 0.47, + 0.437 + ], + "angle": 0, + "content": "and \\(\\mathbf{P}_j^*\\) represent any two assembled shapes. Based on DS, we propose the quality-diversity score (QDS) as below," + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.447, + 0.474, + 0.485 + ], + "angle": 0, + "content": "\\[\n\\mathrm {Q D S} = \\frac {1}{N ^ {2}} \\sum_ {i, j = 1} ^ {N} d _ {c} \\left(\\mathbf {P} _ {i} ^ {*}, \\mathbf {P} _ {j} ^ {*}\\right) \\mathbb {1} \\left(\\mathrm {C A} \\left(\\mathbf {P} _ {i} ^ {*} > \\tau_ {q}\\right)\\right) \\mathbb {1} \\left(\\mathrm {C A} \\left(\\mathbf {P} _ {j} ^ {*} > \\tau_ {q}\\right)\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.495, + 0.47, + 0.572 + ], + "angle": 0, + "content": "QDS imposes constraints to remove pairs that are of low assembly quality, thus assessing not only diversity but also the quality of generated shapes. Following [4], we adopt SCD as the distance metric. The value of \\(\\tau_{q}\\) is set to 0.5 in both QDS and WQDS." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.591, + 0.47, + 0.621 + ], + "angle": 0, + "content": "WQDS. Based on QDS, we further propose the weighted quality-diversity score (WQDS) as below," + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.631, + 0.442, + 0.69 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathrm {W Q D S} = \\frac {1}{N ^ {2}} \\sum_ {i, j = 1} ^ {N} d _ {c} \\left(\\mathbf {P} _ {i} ^ {*}, \\mathbf {P} _ {j} ^ {*}\\right) \\mathrm {C A} \\left(\\mathbf {P} _ {i}\\right) \\mathrm {C A} \\left(\\mathbf {P} _ {j}\\right) \\\\ \\mathbb {1} \\left(\\mathrm {C A} \\left(\\mathbf {P} _ {i} ^ {*} > \\tau_ {q}\\right)\\right) \\mathbb {1} \\left(\\mathrm {C A} \\left(\\mathbf {P} _ {j} ^ {*} > \\tau_ {q}\\right)\\right). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.7, + 0.47, + 0.746 + ], + "angle": 0, + "content": "WQDS weights the QDS of a pair by their connectivity accuracy, thus favoring assembled shapes that demonstrate a high-quality connection between each pair of parts." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.756, + 0.383, + 0.772 + ], + "angle": 0, + "content": "4.3. Comparisons with State-of-the-Art" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.901 + ], + "angle": 0, + "content": "We compare our method with the state-of-the-art methods on PartNet [18] dataset. Following previous works, we first present the quantitative results under a certain Chamfer distance threshold of 0.01 in Table 1. As we can see, our method consistently outperforms all competitors. We also show the visualization of assembly results in Figure 3. It is clear that our method is better than others in terms of both the coherence of the assembled structure and the connectivity between" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.268, + 0.892, + 0.299 + ], + "angle": 0, + "content": "parts, e.g., the positioning of crossbeams between chair legs, and the connection between table legs and the tabletop." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.3, + 0.895, + 0.618 + ], + "angle": 0, + "content": "To eliminate the sensitivity of performances with respect to threshold values and provide a more comprehensive evaluation, we compare the mPA and mCA of our method with two competitive baselines in Table 2. For the Chair category, our method demonstrates a significant improvement over IET [34], with a \\(1.87\\%\\) increase in mPA and a \\(2.95\\%\\) increase in mCA, while the improvement in the Lamp category is relatively minor. This may be attributed to the limited variation in the geometry of lamp parts, making it difficult to clearly differentiate between various levels of the part-whole hierarchy. In addition, Figure 5 illustrates the performance of various methods on the Chair and Table categories using five different Chamfer distance thresholds ranging from 0.01 to 0.05. As the Chamfer distance threshold increases, the advantage of our method becomes more pronounced, which shows the superiority of our approach. We also evaluate the diversity of generative part assembly. As shown in Table 1, our method outperforms all competing methods in both QDS and WQDS. In Figure 4, we use the model to generate multiple assembly shapes for the same set of parts (with different noises) to show the variation of generated shapes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.629, + 0.644, + 0.645 + ], + "angle": 0, + "content": "4.4. Human Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.653, + 0.895, + 0.835 + ], + "angle": 0, + "content": "Since many inaccurate results are still valid visually, we also conduct a human study to evaluate the quality of generated assemblies. Specifically, we perform an A/B test with a group of students, utilizing the identical evaluation system as in [30, 35]. Participants were presented with pairs of randomly chosen assemblies of the same object from two different methods. In total, 28 participants labeled 93 assembly pairs. We compare the proposed method with three main baselines and present the outcomes in Table 5. In over \\(95\\%\\) of the cases, our method is favored over other three competing methods, unequivocally demonstrating the superior visual quality of our method's output." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.654, + 0.863 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.895, + 0.902 + ], + "angle": 0, + "content": "We now investigate the importance of different loss components and verify the effectiveness of the proposed super-part" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20855" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.092, + 0.117, + 0.113, + 0.145 + ], + "angle": 0, + "content": "DGL" + }, + { + "type": "image", + "bbox": [ + 0.146, + 0.104, + 0.207, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.104, + 0.293, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.307, + 0.104, + 0.388, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.409, + 0.117, + 0.484, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.121, + 0.582, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.121, + 0.672, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.099, + 0.759, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.783, + 0.099, + 0.834, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.096, + 0.193, + 0.114, + 0.213 + ], + "angle": 0, + "content": "IET" + }, + { + "type": "image", + "bbox": [ + 0.148, + 0.181, + 0.205, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.181, + 0.295, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.308, + 0.181, + 0.385, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.195, + 0.485, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.2, + 0.582, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.6, + 0.201, + 0.672, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.175, + 0.755, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.785, + 0.177, + 0.836, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.096, + 0.267, + 0.112, + 0.291 + ], + "angle": 0, + "content": "sno" + }, + { + "type": "image", + "bbox": [ + 0.144, + 0.259, + 0.204, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.232, + 0.26, + 0.292, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.309, + 0.261, + 0.385, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.412, + 0.276, + 0.487, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.279, + 0.583, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.604, + 0.28, + 0.674, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.255, + 0.75, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.787, + 0.26, + 0.835, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.327, + 0.114, + 0.388 + ], + "angle": 0, + "content": "Grouu" + }, + { + "type": "image", + "bbox": [ + 0.144, + 0.34, + 0.204, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.341, + 0.288, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.309, + 0.342, + 0.385, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.412, + 0.351, + 0.487, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.354, + 0.582, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.355, + 0.675, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.34, + 0.751, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.788, + 0.341, + 0.835, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.235, + 0.408, + 0.271, + 0.42 + ], + "angle": 0, + "content": "Chair" + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.408, + 0.563, + 0.42 + ], + "angle": 0, + "content": "Table" + }, + { + "type": "image_caption", + "bbox": [ + 0.756, + 0.408, + 0.793, + 0.422 + ], + "angle": 0, + "content": "Lamp" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.428, + 0.892, + 0.456 + ], + "angle": 0, + "content": "Figure 3. The qualitative comparisons between our method and two most competitive baselines on PartNet [18]. We highlight some areas where the assembly quality of ours is clearly better." + }, + { + "type": "image_caption", + "bbox": [ + 0.095, + 0.489, + 0.112, + 0.534 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.131, + 0.482, + 0.197, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.483, + 0.272, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.482, + 0.362, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.482, + 0.452, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.096, + 0.568, + 0.112, + 0.611 + ], + "angle": 0, + "content": "Aeeaae 2" + }, + { + "type": "image", + "bbox": [ + 0.131, + 0.562, + 0.194, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.555, + 0.271, + 0.614 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.554, + 0.361, + 0.614 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.385, + 0.56, + 0.45, + 0.617 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.638, + 0.112, + 0.68 + ], + "angle": 0, + "content": "aennnnnne aannnnnne" + }, + { + "type": "image", + "bbox": [ + 0.131, + 0.628, + 0.19, + 0.686 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.628, + 0.27, + 0.687 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.298, + 0.627, + 0.357, + 0.689 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.385, + 0.629, + 0.448, + 0.69 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.7, + 0.47, + 0.756 + ], + "angle": 0, + "content": "Figure 4. Diverse results on the unseen PartNet[18] test dataset generated by our network to demonstrate the structural variation in part assembly, providing different artistic results while maintaining reasonable object structures." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.768, + 0.282, + 0.784 + ], + "angle": 0, + "content": "encoder on the Chair category." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Super-Part Encoder. Since the super-part encoder is at the core of our part-whole hierarchy message passing network, we conduct an experiment where we remove this module (denoted as w/o Super-Part Enc.). To mitigate the impact of the number of model parameters, we also create another baseline by increasing the number of parameters of the w/o" + }, + { + "type": "table", + "bbox": [ + 0.539, + 0.462, + 0.849, + 0.553 + ], + "angle": 0, + "content": "
SettingsSCD↓PA↑CA↑
w/o Super-Part Enc. + Augmented Param.0.003860.7053.46
w/o Super-Part Enc.0.003661.0455.39
Full Setting0.002864.8358.45
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.559, + 0.892, + 0.588 + ], + "angle": 0, + "content": "Table 3. Ablation studies on part-whole hierarchy on the Table category under a certain Chamfer distance threshold 0.01." + }, + { + "type": "table", + "bbox": [ + 0.539, + 0.605, + 0.848, + 0.715 + ], + "angle": 0, + "content": "
LtLrLsSCD↓PA↑CA↑
✓✓0.007535.3632.73
✓✓0.003960.0254.21
✓✓0.003362.7456.98
✓✓✓0.002864.8358.45
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.722, + 0.892, + 0.75 + ], + "angle": 0, + "content": "Table 4. Ablation studies on loss components on the Table category under a certain Chamfer distance threshold 0.01." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Super-Part Enc. setting (denoted as w/o Super-Part Enc. + Augmented Param.). This involves increasing the number of parameters in the part encoder and the multi-head attention module, ensuring that the overall number of parameters is on the same order of magnitude as the full model setting. As shown in Table 3, the super-part encoder contributes significantly to the final performance. We also observe that augmenting the model parameters without incorporating the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "20856" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.103, + 0.347, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.347, + 0.103, + 0.614, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.616, + 0.103, + 0.887, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.157, + 0.269, + 0.812, + 0.285 + ], + "angle": 0, + "content": "Figure 5. Performance on the Chair, Table and Lamp categories under multiple Chamfer distance thresholds." + }, + { + "type": "table", + "bbox": [ + 0.115, + 0.296, + 0.426, + 0.387 + ], + "angle": 0, + "content": "
MethodPercent Prefer Ours
Ours vs IET[34]95.2%
Ours vs RGL [21]97.6%
Ours vs DGL [33]99.5%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.137, + 0.393, + 0.409, + 0.407 + ], + "angle": 0, + "content": "Table 5. Human study results on PartNet[18]." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.421, + 0.462, + 0.576 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.587, + 0.471, + 0.616 + ], + "angle": 0, + "content": "Figure 6. Visualization of our predicted hierarchical assembly process from parts to the whole." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.624, + 0.472, + 0.671 + ], + "angle": 0, + "content": "super-part encoders actually leads to a performance drop. This finding further highlights the effectiveness of the proposed part-whole hierarchy network for part assembly." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.687, + 0.472, + 0.839 + ], + "angle": 0, + "content": "Loss Functions. We now evaluate the importance of individual loss terms in Equation (6). As shown in Table 4, removing the part translation loss \\(\\mathcal{L}_t\\) leads to a significant performance degradation. Similarly, the remaining loss terms, including the rigid rotation loss \\(\\mathcal{L}_r\\) and the shape assembly loss \\(\\mathcal{L}_s\\), demonstrate their significance in facilitating precise part rotations and ensuring the overall coherence of the assembled parts, respectively. This shows that each loss term does play an indispensable role in achieving high-quality assemblies." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.398, + 0.865 + ], + "angle": 0, + "content": "4.6. Hierarchical Part Assembly Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We further provide a visual analysis of the assembly process for super-parts and parts in Figure 6. It is obvious that during" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.3, + 0.895, + 0.406 + ], + "angle": 0, + "content": "the super-part assembly phase, important components such as chair seats and chair backs are often assembled correctly first. This aligns with our intuition that super-part assembly is relative easier and provides strong hints for predicting poses of parts like chair legs and arms, thus showing further evidence for the effectiveness of the part-whole-hierarchy message passing network." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.419, + 0.62, + 0.435 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.444, + 0.895, + 0.687 + ], + "angle": 0, + "content": "In this paper, we propose the part-whole-hierarchy message passing network to address the challenging generative 3D part assembly task. We first group the point cloud of individual parts to form super-parts in an unsupervised way. Taking the point cloud as input, our super-part encoder predicts latent poses of super-parts which are used to transform the point cloud. We then feed the transformed point cloud to the part encoder. Relying on the cross-level and within-level attention based message passing, the part encoder takes the transformed point cloud as input and leverages the information from super-parts and predicts poses of parts. Experiments on the PartNet dataset demonstrate that our method achieves state-of-the-art performances as well as provides interpretable assembly process. In the future, we are interested in combining 3D part generation with our 3D part assembly model to generate 3D shapes from scratch." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.705, + 0.634, + 0.719 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.727, + 0.897, + 0.88 + ], + "angle": 0, + "content": "This work was funded, in part, by NSERC DG Grants (No. RGPIN-2022-04636), the Vector Institute for AI, and Canada CIFAR AI Chair. Resources used in preparing this research were provided, in part, by the Province of Ontario, the Government of Canada through the Digital Research Alliance of Canada alliance.can.ca, and companies sponsoring the Vector Institute www.vectorinstitute.ai/#partners, and Advanced Research Computing at the University of British Columbia. Additional hardware support was provided by John R. Evans Leaders Fund CFI grant." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20857" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In International conference on machine learning, pages 40-49. PMLR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.472, + 0.227 + ], + "angle": 0, + "content": "[2] Siddhartha Chaudhuri, Evangelos Kalogerakis, Leonidas Guibas, and Vladlen Koltun. Probabilistic reasoning for assembly-based 3d modeling. In ACM SIGGRAPH 2011 papers, pages 1-10. 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.472, + 0.297 + ], + "angle": 0, + "content": "[3] Yun-Chun Chen, Haoda Li, Dylan Turpin, Alec Jacobson, and Animesh Garg. Neural shape mating: Self-supervised object assembly with adversarial shape priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12724-12733, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.471, + 0.339 + ], + "angle": 0, + "content": "[4] Junfeng Cheng, Mingdong Wu, Ruiyuan Zhang, Guanqi Zhan, Chao Wu, and Hao Dong. Score-pa: Score-based 3d part assembly. arXiv preprint arXiv:2309.04220, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.34, + 0.471, + 0.382 + ], + "angle": 0, + "content": "[5] P Kingma Diederik, Max Welling, et al. Auto-encoding variational bayes. In Proceedings of the International Conference on Learning Representations (ICLR), 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.383, + 0.471, + 0.438 + ], + "angle": 0, + "content": "[6] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 605-613, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.439, + 0.471, + 0.507 + ], + "angle": 0, + "content": "[7] Thomas Funkhouser, Michael Kazhdan, Philip Shilane, Patrick Min, William Kiefer, Ayellet Tal, Szymon Rusinkiewicz, and David Dobkin. Modeling by example. ACM transactions on graphics (TOG), 23(3):652-663, 2004. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.508, + 0.471, + 0.564 + ], + "angle": 0, + "content": "[8] Lin Gao, Jie Yang, Tong Wu, Yu-Jie Yuan, Hongbo Fu, Yu-Kun Lai, and Hao Zhang. Sdm-net: Deep generative network for structured deformable mesh. ACM Transactions on Graphics (TOG), 38(6):1-15, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.471, + 0.62 + ], + "angle": 0, + "content": "[9] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.621, + 0.471, + 0.675 + ], + "angle": 0, + "content": "[10] Prakhar Jaiswal, Jinmiao Huang, and Rahul Rai. Assembly-based conceptual 3d modeling with unlabeled components using probabilistic factor graph. Computer-Aided Design, 74: 45-54, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.471, + 0.746 + ], + "angle": 0, + "content": "[11] R Kenny Jones, Theresa Barton, Xianghao Xu, Kai Wang, Ellen Jiang, Paul Guerrero, Niloy J Mitra, and Daniel Ritchie. Shapeassembly: Learning to generate programs for 3d shape structure synthesis. ACM Transactions on Graphics (TOG), 39(6):1-20, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.471, + 0.803 + ], + "angle": 0, + "content": "[12] Evangelos Kalogerakis, Siddhartha Chaudhuri, Daphne Koller, and Vladlen Koltun. A probabilistic model for component-based shape synthesis. Acm Transactions on Graphics (TOG), 31(4):1-11, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.471, + 0.859 + ], + "angle": 0, + "content": "[13] Jun Li, Kai Xu, Siddhartha Chaudhuri, Ersin Yumer, Hao Zhang, and Leonidas Guibas. Grass: Generative recursive autoencoders for shape structures. ACM Transactions on Graphics (TOG), 36(4):1-14, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[14] Yichen Li, Kaichun Mo, Lin Shao, Minhyuk Sung, and Leonidas Guibas. Learning 3d part assembly from a single image. In Computer Vision-ECCV 2020: 16th European" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.893, + 0.121 + ], + "angle": 0, + "content": "Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI 16, pages 664-682. Springer, 2020. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.894, + 0.188 + ], + "angle": 0, + "content": "[15] Yuval Litvak, Armin Biess, and Aharon Bar-Hillel. Learning pose estimation for high-precision robotic assembly using simulated depth images. In 2019 International Conference on Robotics and Automation (ICRA), pages 3521–3527. IEEE, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.19, + 0.894, + 0.272 + ], + "angle": 0, + "content": "[16] Jianlan Luo, Eugen Solowjow, Chengtao Wen, Juan Aparicio Ojea, Alice M Agogino, Aviv Tamar, and Pieter Abbeel. Reinforcement learning on variable impedance controller for high-precision robotic assembly. In 2019 International Conference on Robotics and Automation (ICRA), pages 3080–3087. IEEE, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.274, + 0.894, + 0.328 + ], + "angle": 0, + "content": "[17] Kaichun Mo, Paul Guerrero, Li Yi, Hao Su, Peter Wonka, Niloy Mitra, and Leonidas J Guibas. Structurenet: Hierarchical graph networks for 3d shape generation. arXiv preprint arXiv:1908.00575, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.894, + 0.412 + ], + "angle": 0, + "content": "[18] Kaichun Mo, Shilin Zhu, Angel X Chang, Li Yi, Subarna Tripathi, Leonidas J Guibas, and Hao Su. Partnet: A large-scale benchmark for fine-grained and hierarchical part-level 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 909–918, 2019. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.413, + 0.894, + 0.483 + ], + "angle": 0, + "content": "[19] Kaichun Mo, He Wang, Xinchen Yan, and Leonidas Guibas. Pt2pc: Learning to generate 3d point cloud shapes from part tree conditions. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI 16, pages 683-701. Springer, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.894, + 0.525 + ], + "angle": 0, + "content": "[20] Carsten Moenning and Neil A Dodgson. Fast marching farthest point sampling. Technical report, University of Cambridge, Computer Laboratory, 2003. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.526, + 0.893, + 0.593 + ], + "angle": 0, + "content": "[21] Abhinav Narayan, Rajendra Nagar, and Shanmuganathan Raman. Rgl-net: A recurrent graph learning framework for progressive part assembly. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 78-87, 2022. 1, 2, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.893, + 0.662 + ], + "angle": 0, + "content": "[22] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.664, + 0.894, + 0.746 + ], + "angle": 0, + "content": "[23] Olinde Rodrigues. Des lois géométriques qui regissant les déplacements d'un système solide dans l'espace, et de la variation des coordonnées provenant de ces déplacements considérés indépendamment des causes qui peuvent les produit. Journal de mathématiques pures et appliquées, 5:380-440, 1840. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.748, + 0.894, + 0.816 + ], + "angle": 0, + "content": "[24] Nadav Schor, Oren Katzir, Hao Zhang, and Daniel Cohen-Or. Componet: Learning to generate the unseen by part synthesis and composition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8759–8768, 2019. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.894, + 0.872 + ], + "angle": 0, + "content": "[25] Lin Shao, Toki Migimatsu, and Jeannette Bohg. Learning to scaffold the development of robotic manipulation skills. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 5671-5677. IEEE, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.894, + 0.902 + ], + "angle": 0, + "content": "[26] Dong Wook Shu, Sung Woo Park, and Junseok Kwon. 3d point cloud generative adversarial network based on tree struc" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20858" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.472, + 0.133 + ], + "angle": 0, + "content": "tured graph convolutions. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3859-3868, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.472, + 0.19 + ], + "angle": 0, + "content": "[27] Minhyuk Sung, Hao Su, Vladimir G Kim, Siddhartha Chaudhuri, and Leonidas Guibas. Complementme: Weakly-supervised component suggestions for 3d modeling. ACM Transactions on Graphics (TOG), 36(6):1-12, 2017. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.472, + 0.247 + ], + "angle": 0, + "content": "[28] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.472, + 0.317 + ], + "angle": 0, + "content": "[29] Rundi Wu, Yixin Zhuang, Kai Xu, Hao Zhang, and Baoquan Chen. Pq-net: A generative part seq2seq network for 3d shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 829-838, 2020. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.32, + 0.472, + 0.387 + ], + "angle": 0, + "content": "[30] Yuwen Xiong, Wei-Chiu Ma, Jingkang Wang, and Raquel Urtasun. Learning compact representations for lidar completion and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1074-1083, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.39, + 0.472, + 0.459 + ], + "angle": 0, + "content": "[31] Xianghao Xu, Paul Guerrero, Matthew Fisher, Siddhartha Chaudhuri, and Daniel Ritchie. Unsupervised 3d shape reconstruction by part retrieval and assembly. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8559-8567, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.461, + 0.472, + 0.528 + ], + "angle": 0, + "content": "[32] Kevin Zakka, Andy Zeng, Johnny Lee, and Shuran Song. Form2fit: Learning shape priors for generalizable assembly from disassembly. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 9404–9410. IEEE, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.531, + 0.472, + 0.599 + ], + "angle": 0, + "content": "[33] Guanqi Zhan, Qingnan Fan, Kaichun Mo, Lin Shao, Baoquan Chen, Leonidas J Guibas, Hao Dong, et al. Generative 3d part assembly via dynamic graph learning. Advances in Neural Information Processing Systems, 33:6315-6326, 2020. 1, 2, 3, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.601, + 0.472, + 0.656 + ], + "angle": 0, + "content": "[34] Rufeng Zhang, Tao Kong, Weihao Wang, Xuan Han, and Mingyu You. 3d part assembly generation with instance encoded transformer. IEEE Robotics and Automation Letters, 7(4):9051-9058, 2022. 1, 2, 3, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.658, + 0.472, + 0.7 + ], + "angle": 0, + "content": "[35] Vlas Zyrianov, Xiyue Zhu, and Shenlong Wang. Learning to generate realistic lidar point clouds. In European Conference on Computer Vision, pages 17-35. Springer, 2022. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.7 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20859" + } + ] +] \ No newline at end of file diff --git a/2024/Generative 3D Part Assembly via Part-Whole-Hierarchy Message Passing/40f117f5-c539-42c1-b970-aab5fe53da07_origin.pdf b/2024/Generative 3D Part Assembly via Part-Whole-Hierarchy Message Passing/40f117f5-c539-42c1-b970-aab5fe53da07_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f9a0ff5a248107f32bc18e40bb081656118e8de3 --- /dev/null +++ b/2024/Generative 3D Part Assembly via Part-Whole-Hierarchy Message Passing/40f117f5-c539-42c1-b970-aab5fe53da07_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cac02dafd68e4a11330cea80be4c614984e7b791569baf909be58cdf8d547f8a +size 3793474 diff --git a/2024/Generative 3D Part Assembly via Part-Whole-Hierarchy Message Passing/full.md b/2024/Generative 3D Part Assembly via Part-Whole-Hierarchy Message Passing/full.md new file mode 100644 index 0000000000000000000000000000000000000000..96861980f74029f4982e00c216e36683ab8e2290 --- /dev/null +++ b/2024/Generative 3D Part Assembly via Part-Whole-Hierarchy Message Passing/full.md @@ -0,0 +1,391 @@ +# Generative 3D Part Assembly via Part-Whole-Hierarchy Message Passing + +Bi'an Du1, Xiang Gao1, Wei Hu1*, Renjie Liao2,3,4 + +$^{1}$ Wangxuan Institute of Computer Technology, Peking University + +$^{2}$ University of British Columbia, $^{3}$ Vector Institute for AI, $^{4}$ Canada CIFAR AI Chair + +pkudba@stu.pku.edu.cn, gyshgx868@pku.edu.cn, forhuwei@pku.edu.cn, rjliao@ece.ubc.ca + +# Abstract + +Generative 3D part assembly involves understanding part relationships and predicting their 6-DoF poses for assembling a realistic 3D shape. Prior work often focuses on the geometry of individual parts, neglecting part-whole hierarchies of objects. Leveraging two key observations: 1) super-part poses provide strong hints about part poses, and 2) predicting super-part poses is easier due to fewer superparts, we propose a part-whole-hierarchy message passing network for efficient 3D part assembly. We first introduce super-parts by grouping geometrically similar parts without any semantic labels. Then we employ a part-whole hierarchical encoder, wherein a super-part encoder predicts latent super-part poses based on input parts. Subsequently, we transform the point cloud using the latent poses, feeding it to the part encoder for aggregating super-part information and reasoning about part relationships to predict all part poses. In training, only ground-truth part poses are required. During inference, the predicted latent poses of super-parts enhance interpretability. Experimental results on the PartNet dataset show that our method achieves state-of-the-art performance in part and connectivity accuracy and enables an interpretable hierarchical part assembly. Code is available at https://github.com/pkudba/3DHPA. + +# 1. Introduction + +Generative 3D Part Assembly [7, 15, 16, 25, 32] is an emerging research area that aims to generate complex 3D shapes via assembling simple 3D parts without relying on prior semantic knowledge. Different from traditional 3D shape generation, it focuses on generating diverse, plausible configurations of given parts. It facilitates the generation of complex objects and scenes with compositionality, flexibility, and efficiency. With the rapid evolution of 3D printing technology and the increasing demand for diverse 3D shapes, generative 3D part assembly finds applications in various sce + +narios, attracting attention from experts in computer vision, graphics, robotics, and machine learning. + +However, achieving generative 3D part assembly presents a significant challenge due to the vast number of potential part arrangements and orientations, coupled with the intricate dependencies among the parts. Adding to the complexity, part geometry exhibits notable variation even within the same object category, making it exceedingly difficult to generalize learned assembly patterns across different objects. + +Previous approaches primarily focus on designing architectures capable of learning powerful representations for individual 3D parts. The hope is that these learned representations could facilitate accurate part assembly, either in a one-shot manner or sequentially [14, 21, 33, 34]. However, these methods often ignore the inherent part-whole hierarchies in 3D shapes in representation learning. For instance, as illustrated in Figure 1, a chair consists of super-parts such as seats, backs, and legs, with each super-part further divisible into parts like seat surfaces and seat frames. Understanding the pose of a super-part provides insights into the poses of constituent parts within the same super-part, as they often share similar orientations or exhibit symmetry (e.g., left-right symmetry in chair legs and arms). Moreover, predicting super-part poses is typically easier due to the fewer number of super-parts compared to parts. Incorporating these hierarchies into the modeling would make the learning process and potentially improve the overall performance. + +In this paper, we introduce a part-whole-hierarchy message passing network for 3D shape assembly, predicting 6 degrees of freedom (6-DoF) poses for super-parts and parts in a hierarchical manner. We establish the correspondence between parts and super-parts (i.e., subsets of parts) by grouping parts based on their geometric similarities, following the approach from [33, 34]. Importantly, we treat super-part poses as latent variables to be learned, thereby eliminating the requirement of ground-truth super-part poses in our work. Our model comprises two sequential modules: the super-part encoder and the part encoder. + +The first super-part encoder takes the 3D point clouds of all parts as input and employs the attention-based message + +![](images/a113e8d42477061455253f57651ce08ff4f9753ac7e0eaead656a8e64c613940.jpg) +Figure 1. Left: an illustration of the part-whole-hierarchy for 3D shapes; Right: the part assembly process via the proposed part-whole-hierarchy message passing network. + +passing to predict super-parts poses. The representation of each super-part is aggregated from representations of parts within it. These super-part poses provide initial estimation for the poses of individual parts. + +Subsequently, the whole point clouds are transformed based on the predicted super-part poses and fed to the part encoder. This module, using a cross-attention mechanism, extracts features from the transformed point cloud, transferring super-part level information to the part level. It then leverages the attention-based message passing again to capture relationships among individual parts. + +Following previous works [14, 21, 33, 34], we train and evaluate our model on the PartNet dataset [18]. Overall, our model achieves the state-of-the-art performances, outperforming the strongest competitor [34] by a significant margin, i.e., with an almost $2\%$ improvement in mean part accuracy and a $3\%$ improvement in mean connectivity accuracy. Moreover, through visual analysis of the assembly process for both super-parts and parts, we not only showcase accurate generation of part poses but also demonstrate interpretability via the predicted super-part poses. This interpretability feature further improves the utility of our model. + +# 2. Related Work + +Assembly3D modeling Numerous prior studies have approached the challenging 3D part assembly through the joint estimation of part poses. The influential work by [7] introduces an intelligent scissoring technique to tackle this problem for part components. Subsequent research by [2, 10, 12] employs graphical models to capture the semantic and geometric relationships among shape components, enabling exploration in assembly-based shape modeling. PAGENet [14] presents a network that is aware of individual parts and generates semantic parts along with their poses. [31] proposes to decompose shapes using a library of 3D parts provided by the user. However, these studies either assumed prior knowl- + +edge of part semantics or relied on existing shape databases. In a more practical setting, the authors in [3, 21, 33, 34] focus on the pose estimation of individual parts without relying on shape databases or known semantic information. Specifically, DGL [33] employs a dynamic part graph to iteratively refine the poses of individual parts. A progressive strategy using the recurrent graph learning framework has been investigated in [21]. Additionally, the authors in [3, 34] utilizes the Transformer [28] to model the structural relationships and performs the simultaneous assembly of all parts. We follow this line of work and propose novel improvement to generate structurally-coherent part assemblies. + +Structural Shape Generation In recent years, deep generative models, such as generative adversarial networks (GANs) [9] and variational autoencoders (VAEs) [5], have garnered significant attention for shape generation tasks. Notably, the work by [8] introduces a two-level variational autoencoder that simultaneously learns the overall shape structure and detailed part geometries. Both GRASS [13] and StructureNet [17] employ techniques to compress the shape structure into a latent space while considering the relationships between different parts. Additionally, [19] uses a part-tree decomposition to conditionally generate 3D shapes, and [11] adopts a procedural programmatic representation to establish connections between part cuboids. Inspired by the Seq2Seq networks in machine translation, [29] introduces a sequential encoding and decoding approach for the regression of shape parameters. While many of these approaches focus on directly generating new part shapes given random latent codes, our main focus revolves around the rigid transformation of existing parts to facilitate their assembly. + +# 3. Part-Whole-Hierarchy Message Passing + +This section provides a detailed explanation of our proposed part-whole-hierarchy message passing network. We begin by + +![](images/78accfc024b593cb18524471659a2a24ecd2cfd53c66d1699618b441aaca29ba.jpg) +Figure 2. The overall architecture of our model consists of two modules: super-part encoder and part encoder. We first obtain super-parts via grouping parts based on their geometric similarities in an unsupervised fashion. The super-part encoder takes point cloud as input and predicts latent super-part poses (no ground truth is needed). The point cloud is then transformed based on super-part poses and fed to the part encoder. We incorporate both cross-level and within-level attention in the part encoder to predict part poses. + +presenting how we construct super-parts in an unsupervised manner. Then we introduce our super-part encoder that predicts the latent poses of super-parts. Subsequently, we describe how our part encoder leverages the latent poses and the part-whole hierarchy to predict part poses. The overall model is illustrated in Figure 2. Finally, we introduce the loss functions and the training process for diverse generation. + +# 3.1. Super-Part Construction + +We denote the input point clouds as $\mathcal{P} = \{\mathbf{P}_i|i = 1\ldots N\}$ , where $\mathbf{P}_i\in \mathbb{R}^{d\times 3}$ corresponds to the point cloud of the $i$ -th given part of the 3D shape. The values of $N$ and $d$ represent the number of parts and the number of points per part respectively. Note that their values may differ from object to object. For notation convenience, we assume objects are padded to the maximum number of parts, and each part is padded to the maximum number of points per part. The goal of our task is to predict a set of 6-DoF part poses $\{(t_i,r_i)\}_{i = 1}^N$ , where $t_i\in \mathbb{R}^3$ and $r_i\in \mathbb{R}^4$ represent the translation and rigid rotation for each part, respectively. The complete part assembly for a 3D shape is $\mathcal{S} = \cup_{i = 1}^{N}\mathbf{T}_i(\mathbf{P}_i)$ , and $\mathbf{T}_i$ represents a transformation in $SE(3)$ that consists of a 3D rotation in the rotation group $SO(3)$ and a 3D translation in the translation group, induced by $(t_i,r_i)$ . + +A super-part is a subset of parts that is ideally semantically meaningful. However, since we do not have groundtruth super-parts, we need to construct them in an unsupervised manner. In particular, we compute axis-aligned + +bounding boxes for each part and evaluate the similarity between these 3D boxes. Parts are grouped into the same super-part if the difference between their respective enclosing boxes is below a specified threshold. Following previous works [33, 34], we set the threshold to 0.2. Although the super-parts obtained through this method may lack semantic meaning, they provide a coarse abstraction grounded in geometry similarities without any supervised labels. Then we group the input point cloud parts $\mathcal{P} = \{\mathbf{P}_i\}_{i=1}^N$ into a set of $M$ super-parts $\mathcal{P}' = \{\mathbf{P}_i'\}_{i=1}^M$ based on the part-whole hierarchy of a given object. Here $\mathbf{P}_i'$ represents the $i$ -th super-part. To maintain consistent notation, we again pad the number of parts per super-part to the maximum $M$ . Note that padding is used here for clarity and understanding. In practice, our model can handle sets of parts with varying sizes without padding. + +# 3.2. Super-Part Relation Reasoning + +To reason about the relationships among super-parts, we utilize the super-part encoder. The computational process is illustrated in the top half of Figure 2. + +Given a set of point clouds of a super-part $\mathbf{P}_i^{\prime}$ , we first compute the set-level permutation-invariant representation $\mathbf{F}_i^{\prime}$ via a shared PointNet [22]. We use the same architecture and hyperparameters as in previous works [33, 34] for a fair comparison. This feature representation captures the global characteristics of each super-part while being invariant to the order of the parts within the set. Based on the features + +$\mathbf{F}_i^{\prime}$ extracted for each super-part from the PointNet, we use a two-layer multi-head attention (MHA) module to learn the relationships between super-parts. Specifically, following [28], we calculate an attention matrix $\mathbf{A}^{\prime}$ , where $\mathbf{A}_{i,j}^{\prime}$ represents the attention weight of the $i$ -th super-part to the $j$ -th super-part. Then we multiply each super-part feature $\mathbf{F}_j^{\prime}$ by $\mathbf{A}_{i,j}^{\prime}$ and sum them up to obtain the attention-weighted feature $\mathbf{G}_i^{\prime}$ . In our model, we set the embedding dimension to 256 and the number of heads for the MHA module to 8. + +For super-part pose prediction, we feed the attention-weighted feature $\mathbf{G}_i'$ into a prediction head containing 4 fully-connected layers to obtain the pose $\{(t_i', r_i')\}_{i=1}^M$ . We apply tanh operation to the translation vector, which restricts the range of the part center offset as $(-1,1)$ . Additionally, for simplicity, we predict the quaternion vector $r_i' = (r_{i0}', r_{i1}', r_{i2}', r_{i3}')$ instead of the rotation matrix. We then use the Rodrigues formula [23] to obtain the rotation matrix $\mathbf{R}_i'$ corresponding to $r_i'$ one by one as follows: + +$$ +\mathbf {R} _ {i} ^ {\prime} = \left[ \begin{array}{l} 1 - 2 r _ {i 2} ^ {\prime 2} - 2 r _ {i 3} ^ {\prime 2}, 2 r _ {i 1} ^ {\prime} r _ {i 2} ^ {\prime} - 2 r _ {i 0} ^ {\prime} r _ {i 3} ^ {\prime}, 2 r _ {i 1} ^ {\prime} r _ {i 3} ^ {\prime} + 2 r _ {i 0} ^ {\prime} r _ {i 2} ^ {\prime} \\ 2 r _ {i 1} ^ {\prime} r _ {i 2} ^ {\prime} + 2 r _ {i 0} ^ {\prime} r _ {i 3} ^ {\prime}, 1 - 2 r _ {i 1} ^ {\prime 2} - 2 r _ {i 3} ^ {\prime 2}, 2 r _ {i 2} ^ {\prime} r _ {i 3} ^ {\prime} - 2 r _ {i 0} ^ {\prime} r _ {i 1} ^ {\prime} \\ 2 r _ {i 1} ^ {\prime} r _ {i 3} ^ {\prime} - 2 r _ {i 0} ^ {\prime} r _ {i 2} ^ {\prime}, 2 r _ {i 2} ^ {\prime} r _ {i 3} ^ {\prime} + 2 r _ {i 0} ^ {\prime} r _ {i 1} ^ {\prime}, 1 - 2 r _ {i 1} ^ {\prime 2} - 2 r _ {i 2} ^ {\prime 2}. \end{array} \right] +$$ + +The transformed point cloud of the super-part is expressed as $\mathbf{T}_i(\mathbf{P}_i^{\prime}) = \mathbf{R}_i^{\prime}\mathbf{P}_i^{\prime} + t_i^{\prime}$ . To ensure that it is a unit quaternion, we normalize the rigid rotation vector such that $\| r_i^\prime \| = 1$ . + +Since we do not have ground-truth super-part poses, the supervision for this module solely comes from the backpropagation signal from the subsequent module. As a result, the super-part poses are considered as latent variables. + +# 3.3. Part Relation Reasoning + +As previously mentioned, the poses of super-parts offer valuable clues about the poses of their corresponding parts. To leverage this information, we design a part encoder to incorporate the latent super-part poses to predict the part poses. + +Transformation via Latent Super-Part Poses Firstly, we apply the latent super-part poses to transform all point cloud parts, resulting in transformed parts denoted as $\hat{\mathcal{P}} = \{\hat{\mathbf{P}}_i\}_{i=1}^N$ . These transformed parts are then fed into the part encoder for further processing. Similarly to the super-part encoder, we employ another PointNet to extract part-level feature denoted as $\hat{\mathbf{F}}_i$ . In order to better utilize the part-whole hierarchy of objects, we design two ways to integrate the super-part level information to the part level, i.e., cross-level and within-level attention modules. + +Cross-Level Attention Part representations $\hat{\mathbf{F}}_i$ from the PointNet encoder does not leverage the part-whole hierarchy. To better fuse the information from the super-parts, we propose a cross-level attention module. Specifically, we treat the part representation $\hat{\mathbf{F}}_i$ as query and super-part representations $\mathbf{F}_i'$ as keys and values. We then employ a MHA module + +to perform an attention-weighted aggregation of super-part representations $\mathbf{F}_i^{\prime}$ to obtain the updated part representations $\mathbf{F}_i$ . By doing so, the information from part-whole hierarchy is explicitly integrated into part representations. Here we specify the embedding dimension as 256 and the MHA module is set to utilize 8 heads. + +Within-Level Attention To model the part relationships, we leverage another within-level attention module, which is depicted in the bottom half of Figure 2. We first concatenate the representation from the previous module $F_{i}$ with an additional instance encoding vector that is unique for each part following [34]. This allows us to distinguish geometrically-equivalent parts. Then we perform message passing among parts using another MHA module where the attention computation is aware of the part-whole hierarchy. Specifically, we concatenate the latent super-part poses to the part-level feature before computing the attention scores at each layer. Consequently, in each message passing layer, a part receives pose hints from its parent super-part and updates its representation. The resulting part feature is denoted as $\mathbf{G}_i = \sum_{j=1}^N \mathbf{A}_{i,j} \mathbf{F}_i$ , where $\mathbf{A}_{i,j}$ represents the attention weight of the $i$ -th part to the $j$ -th part. This design enables the part-to-part message passing to effectively leverage the part-whole hierarchy information. We set the number of heads for the MHA module to 8, the dimension of the part-level feature to 256, and concatenate it with the 40-dimensional instance encoding vector. + +Prediction Head Finally, we feed the attention-weighted part-level feature $\mathbf{G}_i$ to a part pose prediction head, which consists of four fully-connected layers. The channel sizes of the first three fully-connected layers are set to 256, 256 and 1024, respectively, followed by ReLU activation functions. We use a linear projection layer to predict the part pose $(t_i, r_i)$ , encompassing a 4-dimensional rigid rotation and 3-dimensional translation. This prediction head outputs the part poses $\{(t_i, r_i)\}_{i=1}^N$ , where $t_i$ represents the translation vector and $r_i$ represents the rigid rotation vector for each part. We adopt the same design as in the super-part encoder to restrict the translation vectors to the range of $(-1, 1)$ and normalize the rigid rotation vector $r_i$ to have a unit norm. + +# 3.4. Training Objective for Diverse Generation + +Considering the same set of input point cloud parts, geometrically equivalent parts (such as legs of a chair) can be interchanged and decorated parts may have multiple placement options, resulting in multiple possible shapes. For example, a semi-cylindrical part can be placed on top of a chair backrest as a headrest, or placed under the backrest as a lumbar pillow. To account for such diversified structural variations and configurations, we introduce random noise into the system following [21, 33, 34] and employ the Min-of-N + +(MoN) loss [6] to simultaneously consider assembly accuracy and assembly diversity. By considering the minimum of a set of N possible assemblies, we can encourage the model to generate diverse and valid shape configurations while maintaining overall accuracy in the assembly process. This helps address the challenge of capturing multiple plausible solutions that arise from the interchangeable and decoratable nature of the parts. + +Let $\mathcal{F}(\mathcal{P},\delta_j)$ denote our network outputs and $\mathcal{F}^* (\mathcal{P})$ denote the ground-truth point clouds, then the MoN loss is: + +$$ +\mathcal {L} _ {\mathrm {M o N}} = \min _ {\delta_ {j} \sim \mathcal {N} (0, 1)} \mathcal {L} \left(\mathcal {F} \left(\mathcal {P}, \delta_ {j}\right), \mathcal {F} ^ {*} (\mathcal {P})\right) \tag {1} +$$ + +where $\delta_{j}$ is a random noise vector drawn from standard Normal distribution in an IID fashion. Following [34], we sample 5 random vectors $\delta_{j}$ during training. The loss function $\mathcal{L}$ consists of the following components for both local part and global shape losses. + +Firstly, we supervise the translation via an $l_{2}$ loss between our prediction $t_{i}$ and the ground-truth translation vector $t_{i}^{*}$ for each part: + +$$ +\mathcal {L} _ {t} = \sum_ {i = 1} ^ {N} \| t _ {i} - t _ {i} ^ {*} \| _ {2} ^ {2} \tag {2} +$$ + +Then we compute the Chamfer distance [6] between the predicted and the ground-truth rigid rotation for each part: + +$$ +\mathcal {L} _ {r} = \sum_ {i = 1} ^ {N} d _ {c} \left(r _ {i} \left(\mathbf {P} _ {i}\right), r _ {i} ^ {*} \left(\mathbf {P} _ {i}\right)\right) \tag {3} +$$ + +The Chamfer distance $d_{c}(\mathcal{X},\mathcal{Y})$ is a metric commonly used in point cloud comparison, which measures the dissimilarity between two point sets $\mathcal{X}$ and $\mathcal{Y}$ by calculating the average distance between each point in one set and its nearest neighbor in the other set as follows: + +$$ +d _ {c} (\mathcal {X}, \mathcal {Y}) = \sum_ {x \in \mathcal {X}} \min _ {y \in \mathcal {Y}} \| x - y \| _ {2} ^ {2} + \sum_ {y \in \mathcal {Y}} \min _ {x \in \mathcal {X}} \| x - y \| _ {2} ^ {2} \tag {4} +$$ + +In this case, we use the Chamfer distance to assess the discrepancy between the predicted rotation of each part and its corresponding ground-truth rotation. + +Similarly, the full shape $S$ , i.e., the set of point clouds belonging to all parts, is supervised via the Chamfer distance from the ground-truth shape $S^{*}$ : + +$$ +\mathcal {L} _ {s} = d _ {c} \left(\mathcal {S}, \mathcal {S} ^ {*}\right) \tag {5} +$$ + +In summary, the overall loss function is defined as: + +$$ +\mathcal {L} = \lambda_ {t} \mathcal {L} _ {t} + \lambda_ {r} \mathcal {L} _ {r} + \lambda_ {s} \mathcal {L} _ {s} \tag {6} +$$ + +where $\lambda_{t},\lambda_{r}$ and $\lambda_{s}$ represent the weights assigned to the three loss terms. Based on the cross validation, we empirically set $\lambda_t = 1$ $\lambda_r = 10$ ,and $\lambda_s = 1$ in our experiments. By minimizing this loss function, we aim to improve the alignment of the predicted poses with the ground-truth poses, thereby enhancing the overall accuracy of the assembly. + +# 4. Experiments + +In this section, we demonstrate the effectiveness of the proposed model by comparing our results with state-of-the-art methods. We also provide visual analysis of the assembly process for super-parts and parts, which helps to illustrate the hierarchical assembly process from parts to the whole. + +# 4.1. Dataset + +Following [4, 14, 21, 24, 27, 29, 33, 34], we use the PartNet [18] dataset for both training and evaluation. This dataset consists of 26,671 shapes across 24 different 3D object categories. To effectively validate and compare different methods, we select the three largest categories, i.e., chairs, tables and lamps, following [4, 14, 21, 33, 34]. In total, we have 6,323 chairs, 8,218 tables, 2,207 lamps in the finest-grained level, and we adopt the official train/validation/test splits (70%/10%/20%) to conduct the experiments. For each part point cloud, 1000 points are sampled from the original part meshes using Farthest Point Sampling (FPS) [20]. All parts are transformed into its canonical space using PCA. + +# 4.2. Evaluation Metrics + +We generate a variety of shapes by adding different Gaussian noises to a given set of input parts, and find the closest shape to the ground-truth using minimum matching distance (MMD) [1]. To measure part assembly quality, we follow [21, 33, 34] to use shape Chamfer distance (SCD), part accuracy (PA) and connectivity accuracy (CA). Shape Chamfer distance is defined in Equation (4) and Equation (5). Meanwhile, to compare the diversity of assembled parts, we propose two quantitative evaluation metrics including quality-diversity score (QDS) and the weighted quality-diversity score (WQDS) following [4]. We also visualize the generation results to qualitatively evaluate the diversity. More details of these metrics are provided in the appendix. + +Mean Part/Connectivity Accuracy (mPA/mCA). PA and CA depend on the Chamfer distance threshold $\tau_{p}$ and $\tau_{c}$ for judging whether the assembly is accurate. In order to provide a more comprehensive evaluation of the performance of the model, we average the results under multiple thresholds to get mean part accuracy (mPA) and mean connectivity accuracy (mCA), formally, + +$$ +\mathrm {m P A} = \frac {1}{T _ {p}} \sum_ {\tau_ {p} \in T _ {p}} \mathrm {P A} (\tau_ {p}), \quad \mathrm {m C A} = \frac {1}{T _ {c}} \sum_ {\tau_ {c} \in T _ {c}} \mathrm {C A} (\tau_ {c}), +$$ + +where $T_{p}$ and $T_{c}$ are set to $\{0.01, 0.02, 0.03, 0.04, 0.05\}$ . + +QDS. Diversity score (DS) [19, 26] evaluates the diversity of the results as $\mathrm{DS} = \frac{1}{N^2}\sum_{i,j=1}^{N}(d_c(\mathbf{P}_i^*,\mathbf{P}_j^*))$ , where $\mathbf{P}_i^*$ + +
SCD(10-2) ↓PA(%) ↑CA(%) ↑QDS(10-5) ↑WQDS(10-5) ↑
MethodsChairTableLampChairTableLampChairTableLampChairTableLampChairTableLamp
B-Global [14, 24]1.461.120.7915.7015.3722.619.9033.8418.600.150.200.761.251.400.58
B-LSTM [29]1.311.250.7721.7728.6420.786.8022.5614.050.040.270.631.071.431.54
B-Complement [27]2.412.981.508.782.3212.679.1915.5726.560.090.062.811.281.752.08
DGL [33]0.910.500.9339.0049.5133.3323.8739.9641.701.693.051.841.352.971.73
Score [4]0.710.421.1144.5152.7834.3230.3240.5949.073.369.176.831.703.812.82
RGL [21]0.870.480.7249.0654.1637.5632.2642.1557.343.557.636.822.124.072.96
IET [34]0.540.351.0362.8061.6738.6848.4556.1862.624.159.096.982.744.563.29
Ours0.510.280.7064.1364.8338.8049.2858.4564.165.629.587.123.064.813.90
+ +Table 1. Comparison between our approach and other methods under the Chamfer distance threshold 0.01. + +
mPA ↑mCA ↑
MethodsChairTableLampChairTableLamp
DGL [33]45.8457.8648.3229.1743.8851.75
IET [34]74.9273.2052.8062.3768.4975.94
Ours76.7976.1953.3165.3270.2678.15
+ +Table 2. Comparison between our approach and other methods under multiple Chamfer distance thresholds from 0.01 to 0.05. + +and $\mathbf{P}_j^*$ represent any two assembled shapes. Based on DS, we propose the quality-diversity score (QDS) as below, + +$$ +\mathrm {Q D S} = \frac {1}{N ^ {2}} \sum_ {i, j = 1} ^ {N} d _ {c} \left(\mathbf {P} _ {i} ^ {*}, \mathbf {P} _ {j} ^ {*}\right) \mathbb {1} \left(\mathrm {C A} \left(\mathbf {P} _ {i} ^ {*} > \tau_ {q}\right)\right) \mathbb {1} \left(\mathrm {C A} \left(\mathbf {P} _ {j} ^ {*} > \tau_ {q}\right)\right). +$$ + +QDS imposes constraints to remove pairs that are of low assembly quality, thus assessing not only diversity but also the quality of generated shapes. Following [4], we adopt SCD as the distance metric. The value of $\tau_{q}$ is set to 0.5 in both QDS and WQDS. + +WQDS. Based on QDS, we further propose the weighted quality-diversity score (WQDS) as below, + +$$ +\begin{array}{l} \mathrm {W Q D S} = \frac {1}{N ^ {2}} \sum_ {i, j = 1} ^ {N} d _ {c} \left(\mathbf {P} _ {i} ^ {*}, \mathbf {P} _ {j} ^ {*}\right) \mathrm {C A} \left(\mathbf {P} _ {i}\right) \mathrm {C A} \left(\mathbf {P} _ {j}\right) \\ \mathbb {1} \left(\mathrm {C A} \left(\mathbf {P} _ {i} ^ {*} > \tau_ {q}\right)\right) \mathbb {1} \left(\mathrm {C A} \left(\mathbf {P} _ {j} ^ {*} > \tau_ {q}\right)\right). \\ \end{array} +$$ + +WQDS weights the QDS of a pair by their connectivity accuracy, thus favoring assembled shapes that demonstrate a high-quality connection between each pair of parts. + +# 4.3. Comparisons with State-of-the-Art + +We compare our method with the state-of-the-art methods on PartNet [18] dataset. Following previous works, we first present the quantitative results under a certain Chamfer distance threshold of 0.01 in Table 1. As we can see, our method consistently outperforms all competitors. We also show the visualization of assembly results in Figure 3. It is clear that our method is better than others in terms of both the coherence of the assembled structure and the connectivity between + +parts, e.g., the positioning of crossbeams between chair legs, and the connection between table legs and the tabletop. + +To eliminate the sensitivity of performances with respect to threshold values and provide a more comprehensive evaluation, we compare the mPA and mCA of our method with two competitive baselines in Table 2. For the Chair category, our method demonstrates a significant improvement over IET [34], with a $1.87\%$ increase in mPA and a $2.95\%$ increase in mCA, while the improvement in the Lamp category is relatively minor. This may be attributed to the limited variation in the geometry of lamp parts, making it difficult to clearly differentiate between various levels of the part-whole hierarchy. In addition, Figure 5 illustrates the performance of various methods on the Chair and Table categories using five different Chamfer distance thresholds ranging from 0.01 to 0.05. As the Chamfer distance threshold increases, the advantage of our method becomes more pronounced, which shows the superiority of our approach. We also evaluate the diversity of generative part assembly. As shown in Table 1, our method outperforms all competing methods in both QDS and WQDS. In Figure 4, we use the model to generate multiple assembly shapes for the same set of parts (with different noises) to show the variation of generated shapes. + +# 4.4. Human Study + +Since many inaccurate results are still valid visually, we also conduct a human study to evaluate the quality of generated assemblies. Specifically, we perform an A/B test with a group of students, utilizing the identical evaluation system as in [30, 35]. Participants were presented with pairs of randomly chosen assemblies of the same object from two different methods. In total, 28 participants labeled 93 assembly pairs. We compare the proposed method with three main baselines and present the outcomes in Table 5. In over $95\%$ of the cases, our method is favored over other three competing methods, unequivocally demonstrating the superior visual quality of our method's output. + +# 4.5. Ablation Study + +We now investigate the importance of different loss components and verify the effectiveness of the proposed super-part + +![](images/c0a31a3bd067e53f1fbc39286632ef9f02b7469d3a791bfb754c93c45a61b6f5.jpg) +DGL + +![](images/98a9b109683bc0a7124b90dc96a7b1953e4e7618d8eb21019c6b1041876b3835.jpg) + +![](images/d89e3dc458b7d1518e4047bb7a1c992606a7b462d2501e4f27a8e85951bf81f1.jpg) + +![](images/ee61dbac09230377db4af431ea67f53e540a5dd642a18814ce7bc67efac6b606.jpg) + +![](images/b4bb8d5fb29c3d0a22a4df3c7c04ebeb1d1187dc437cc5587571b24848bc062a.jpg) + +![](images/1e78afbfb0f88373709f4c9316735357b0af95148c7ef12d7627722d8b5438e3.jpg) + +![](images/43c670993759bf50271e36f064bc0c6fb7a11b729f16109fec0cd5e030704dd8.jpg) + +![](images/99c5d6f186935f4fefd421ada073cef96c40885454d9ce31e8b1eb5810569f88.jpg) + +![](images/6ba294a8c5df4bd1e23c6932d9ac5288d6b509d25a8a5b2d3ba0729bc26dd7a7.jpg) +IET + +![](images/0bae73de0041777f7ee1b969aa33437d47c35caa9422819e2e638cc74d8139f5.jpg) + +![](images/ac472874acc8ebac9d6f81b5eb5f7a66221b5092c321d590518b453da52efa1e.jpg) + +![](images/12158f99e369ffde7104a65f343f78a41484cfd787087238aebb4df934afc0df.jpg) + +![](images/88ca63764e1b942c822bdcdbbc9c52d2f9384ebfdb7e7eb5fb1de07c44611f82.jpg) + +![](images/a2bdc168b759da5aad9855d257709cfaa3260cf877fe5ed8d811ad753b956bf0.jpg) + +![](images/1faffe13274ae5db24a8706a0c1e3d813cf37d12fe3be82f85a62ffd95cf4a5e.jpg) + +![](images/ea6131c54406dfdead26728adca6c4e08c25611ec10d19773a03cc21ce9e6d0d.jpg) + +![](images/1739980a2834bb31046dd9a000c112c3e99fdec541922734b2983357db39d502.jpg) +sno + +![](images/722b6727f01c40a86097a5fc6b3d21fee01b34874ceffcde11cada5439b65fe9.jpg) + +![](images/8efa1ddbf309bc904bdcc8e40bc31c97df75483289059040f6a284e10bf8940c.jpg) + +![](images/c1d2e9c85bc22f8d3d2b6ced980ebdcd46b7491489af02a1a7f647ed958b2891.jpg) + +![](images/627ec2a8d37f67b5cb84f78e360ede24e849e8641e9099b9583bc41569aebc84.jpg) + +![](images/ba9517e33e3a63637b556c00f36563c0a63ea4aa1799912a2b5751db0b7353d8.jpg) + +![](images/1a4c047017a2ea569677c75349811c3726394ece7483574c26e3958eed412b1a.jpg) + +![](images/0c346b59483888665311185c00d91430dedb480506502cdf60dc0932e80e532b.jpg) + +![](images/19c2f8696e111d1b4c9f6d63fe378b3cdd808708acb8183f2366ba8285e91370.jpg) +Grouu + +![](images/afdec642351463f0d6074e60d16ef556a3620a8797ceea4ff57985a7769a2e63.jpg) +Chair + +![](images/7ab5fdf4dabc40bb14e1367c127c44aefe8ca066a1c5ec7da268f6301fee23f0.jpg) +Figure 3. The qualitative comparisons between our method and two most competitive baselines on PartNet [18]. We highlight some areas where the assembly quality of ours is clearly better. + +![](images/d48512d7f8df10e313a604ea140bdf2f0eac9426420b6977c466e2034911d677.jpg) +Table + +![](images/84447528fee5d37a3c41f3132709e45db88f8485e432ca3c63167f289205896d.jpg) + +![](images/414a1b04931c331d180899687f57ff6984f4145f9322fa2e2ac1654bd96a4813.jpg) + +![](images/1b23f45b2c1a881b9885b0dd01a05e30cd7687239a975d97b72dfa2a8a2825fa.jpg) + +![](images/06eadb62e6faf4c9ca9148cd892598d47e5468a612596a1c4b7916748670f56b.jpg) +Lamp + +![](images/aa9bbca0fc68aa66d86d85c7d1765799b428039c0ee79ddc4e4406f8e7dc8995.jpg) +3 + +![](images/f7cc16ffc59690cdd255d5ddf4b95504c0e2fe0b8e691c9da25a725ca0700a0b.jpg) + +![](images/662907be14770d8167de5685bfb25a14406bf262e75f5e273ba62617cca74bd2.jpg) + +![](images/aff3a1d3db61d179ab0a8fa642e394fa5eb32603a8fabc4478e1db3ad6d09a49.jpg) + +![](images/f559b0ae3511b22f652dde3b59d39d6963db5984fa57819b822122a3ef46c459.jpg) +Aeeaae 2 + +![](images/ba21a9cbc7ce0d1e2fb76659e2a183d4dec3717d7510c7ee540763dc0f3472b9.jpg) + +![](images/1cf5e07ab5cad61a3ba2b3accd347e4fec3326059627a5a83c9b20fc34ca969d.jpg) + +![](images/eac492ac95e1b1e54327be94a12117a5c47fd84d97deb925d390c91effb54ff6.jpg) + +![](images/117bf18a41ad36240ccb0f8f80d9bb932bc2100c98b193af97c6ca6f8e698234.jpg) +aennnnnne aannnnnne +Figure 4. Diverse results on the unseen PartNet[18] test dataset generated by our network to demonstrate the structural variation in part assembly, providing different artistic results while maintaining reasonable object structures. + +![](images/1f0565af5277f064a2306c3c8634fcd09b5c3cbba8c16d96720580968b641785.jpg) + +![](images/d1bfd76f71ff301a28bc796cf3878864398cdbc8fed06c5188f8f23f0bfeb219.jpg) + +![](images/e337b53450f184be699e239f6a107408016ec5d77a57edfb80bab5cc90d7e843.jpg) + +encoder on the Chair category. + +Super-Part Encoder. Since the super-part encoder is at the core of our part-whole hierarchy message passing network, we conduct an experiment where we remove this module (denoted as w/o Super-Part Enc.). To mitigate the impact of the number of model parameters, we also create another baseline by increasing the number of parameters of the w/o + +
SettingsSCD↓PA↑CA↑
w/o Super-Part Enc. + Augmented Param.0.003860.7053.46
w/o Super-Part Enc.0.003661.0455.39
Full Setting0.002864.8358.45
+ +Table 3. Ablation studies on part-whole hierarchy on the Table category under a certain Chamfer distance threshold 0.01. + +
LtLrLsSCD↓PA↑CA↑
✓✓0.007535.3632.73
✓✓0.003960.0254.21
✓✓0.003362.7456.98
✓✓✓0.002864.8358.45
+ +Table 4. Ablation studies on loss components on the Table category under a certain Chamfer distance threshold 0.01. + +Super-Part Enc. setting (denoted as w/o Super-Part Enc. + Augmented Param.). This involves increasing the number of parameters in the part encoder and the multi-head attention module, ensuring that the overall number of parameters is on the same order of magnitude as the full model setting. As shown in Table 3, the super-part encoder contributes significantly to the final performance. We also observe that augmenting the model parameters without incorporating the + +![](images/d0a5e2ee66ca2cc3e2903d93013d27c2a33ab727a5944cd5484f427dd9a12cef.jpg) +Figure 5. Performance on the Chair, Table and Lamp categories under multiple Chamfer distance thresholds. + +![](images/7f20821c27e88604feb16407c38088204dfa8b62ccc5aa54bcd3a73bf682d34e.jpg) + +![](images/d619800e14b090589a6a33359b79023d0d6a7449a4c38def3d739be7fa04f2d2.jpg) + +
MethodPercent Prefer Ours
Ours vs IET[34]95.2%
Ours vs RGL [21]97.6%
Ours vs DGL [33]99.5%
+ +Table 5. Human study results on PartNet[18]. + +![](images/78d2d7c499f304c2361218850408503913e4789c3fb60a001317527d4be79150.jpg) +Figure 6. Visualization of our predicted hierarchical assembly process from parts to the whole. + +super-part encoders actually leads to a performance drop. This finding further highlights the effectiveness of the proposed part-whole hierarchy network for part assembly. + +Loss Functions. We now evaluate the importance of individual loss terms in Equation (6). As shown in Table 4, removing the part translation loss $\mathcal{L}_t$ leads to a significant performance degradation. Similarly, the remaining loss terms, including the rigid rotation loss $\mathcal{L}_r$ and the shape assembly loss $\mathcal{L}_s$ , demonstrate their significance in facilitating precise part rotations and ensuring the overall coherence of the assembled parts, respectively. This shows that each loss term does play an indispensable role in achieving high-quality assemblies. + +# 4.6. Hierarchical Part Assembly Analysis + +We further provide a visual analysis of the assembly process for super-parts and parts in Figure 6. It is obvious that during + +the super-part assembly phase, important components such as chair seats and chair backs are often assembled correctly first. This aligns with our intuition that super-part assembly is relative easier and provides strong hints for predicting poses of parts like chair legs and arms, thus showing further evidence for the effectiveness of the part-whole-hierarchy message passing network. + +# 5. Conclusion + +In this paper, we propose the part-whole-hierarchy message passing network to address the challenging generative 3D part assembly task. We first group the point cloud of individual parts to form super-parts in an unsupervised way. Taking the point cloud as input, our super-part encoder predicts latent poses of super-parts which are used to transform the point cloud. We then feed the transformed point cloud to the part encoder. Relying on the cross-level and within-level attention based message passing, the part encoder takes the transformed point cloud as input and leverages the information from super-parts and predicts poses of parts. Experiments on the PartNet dataset demonstrate that our method achieves state-of-the-art performances as well as provides interpretable assembly process. In the future, we are interested in combining 3D part generation with our 3D part assembly model to generate 3D shapes from scratch. + +# Acknowledgments + +This work was funded, in part, by NSERC DG Grants (No. RGPIN-2022-04636), the Vector Institute for AI, and Canada CIFAR AI Chair. Resources used in preparing this research were provided, in part, by the Province of Ontario, the Government of Canada through the Digital Research Alliance of Canada alliance.can.ca, and companies sponsoring the Vector Institute www.vectorinstitute.ai/#partners, and Advanced Research Computing at the University of British Columbia. Additional hardware support was provided by John R. Evans Leaders Fund CFI grant. + +# References + +[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In International conference on machine learning, pages 40-49. PMLR, 2018. 5 +[2] Siddhartha Chaudhuri, Evangelos Kalogerakis, Leonidas Guibas, and Vladlen Koltun. Probabilistic reasoning for assembly-based 3d modeling. In ACM SIGGRAPH 2011 papers, pages 1-10. 2011. 2 +[3] Yun-Chun Chen, Haoda Li, Dylan Turpin, Alec Jacobson, and Animesh Garg. Neural shape mating: Self-supervised object assembly with adversarial shape priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12724-12733, 2022. 2 +[4] Junfeng Cheng, Mingdong Wu, Ruiyuan Zhang, Guanqi Zhan, Chao Wu, and Hao Dong. Score-pa: Score-based 3d part assembly. arXiv preprint arXiv:2309.04220, 2023. 5, 6 +[5] P Kingma Diederik, Max Welling, et al. Auto-encoding variational bayes. In Proceedings of the International Conference on Learning Representations (ICLR), 2014. 2 +[6] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 605-613, 2017. 5 +[7] Thomas Funkhouser, Michael Kazhdan, Philip Shilane, Patrick Min, William Kiefer, Ayellet Tal, Szymon Rusinkiewicz, and David Dobkin. Modeling by example. ACM transactions on graphics (TOG), 23(3):652-663, 2004. 1, 2 +[8] Lin Gao, Jie Yang, Tong Wu, Yu-Jie Yuan, Hongbo Fu, Yu-Kun Lai, and Hao Zhang. Sdm-net: Deep generative network for structured deformable mesh. ACM Transactions on Graphics (TOG), 38(6):1-15, 2019. 2 +[9] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 2 +[10] Prakhar Jaiswal, Jinmiao Huang, and Rahul Rai. Assembly-based conceptual 3d modeling with unlabeled components using probabilistic factor graph. Computer-Aided Design, 74: 45-54, 2016. 2 +[11] R Kenny Jones, Theresa Barton, Xianghao Xu, Kai Wang, Ellen Jiang, Paul Guerrero, Niloy J Mitra, and Daniel Ritchie. Shapeassembly: Learning to generate programs for 3d shape structure synthesis. ACM Transactions on Graphics (TOG), 39(6):1-20, 2020. 2 +[12] Evangelos Kalogerakis, Siddhartha Chaudhuri, Daphne Koller, and Vladlen Koltun. A probabilistic model for component-based shape synthesis. Acm Transactions on Graphics (TOG), 31(4):1-11, 2012. 2 +[13] Jun Li, Kai Xu, Siddhartha Chaudhuri, Ersin Yumer, Hao Zhang, and Leonidas Guibas. Grass: Generative recursive autoencoders for shape structures. ACM Transactions on Graphics (TOG), 36(4):1-14, 2017. 2 +[14] Yichen Li, Kaichun Mo, Lin Shao, Minhyuk Sung, and Leonidas Guibas. Learning 3d part assembly from a single image. In Computer Vision-ECCV 2020: 16th European + +Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI 16, pages 664-682. Springer, 2020. 1, 2, 5, 6 +[15] Yuval Litvak, Armin Biess, and Aharon Bar-Hillel. Learning pose estimation for high-precision robotic assembly using simulated depth images. In 2019 International Conference on Robotics and Automation (ICRA), pages 3521–3527. IEEE, 2019. 1 +[16] Jianlan Luo, Eugen Solowjow, Chengtao Wen, Juan Aparicio Ojea, Alice M Agogino, Aviv Tamar, and Pieter Abbeel. Reinforcement learning on variable impedance controller for high-precision robotic assembly. In 2019 International Conference on Robotics and Automation (ICRA), pages 3080–3087. IEEE, 2019. 1 +[17] Kaichun Mo, Paul Guerrero, Li Yi, Hao Su, Peter Wonka, Niloy Mitra, and Leonidas J Guibas. Structurenet: Hierarchical graph networks for 3d shape generation. arXiv preprint arXiv:1908.00575, 2019. 2 +[18] Kaichun Mo, Shilin Zhu, Angel X Chang, Li Yi, Subarna Tripathi, Leonidas J Guibas, and Hao Su. Partnet: A large-scale benchmark for fine-grained and hierarchical part-level 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 909–918, 2019. 2, 5, 6, 7, 8 +[19] Kaichun Mo, He Wang, Xinchen Yan, and Leonidas Guibas. Pt2pc: Learning to generate 3d point cloud shapes from part tree conditions. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI 16, pages 683-701. Springer, 2020. 2, 5 +[20] Carsten Moenning and Neil A Dodgson. Fast marching farthest point sampling. Technical report, University of Cambridge, Computer Laboratory, 2003. 5 +[21] Abhinav Narayan, Rajendra Nagar, and Shanmuganathan Raman. Rgl-net: A recurrent graph learning framework for progressive part assembly. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 78-87, 2022. 1, 2, 4, 5, 6, 8 +[22] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 3 +[23] Olinde Rodrigues. Des lois géométriques qui regissant les déplacements d'un système solide dans l'espace, et de la variation des coordonnées provenant de ces déplacements considérés indépendamment des causes qui peuvent les produit. Journal de mathématiques pures et appliquées, 5:380-440, 1840. 4 +[24] Nadav Schor, Oren Katzir, Hao Zhang, and Daniel Cohen-Or. Componet: Learning to generate the unseen by part synthesis and composition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8759–8768, 2019. 5, 6 +[25] Lin Shao, Toki Migimatsu, and Jeannette Bohg. Learning to scaffold the development of robotic manipulation skills. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 5671-5677. IEEE, 2020. 1 +[26] Dong Wook Shu, Sung Woo Park, and Junseok Kwon. 3d point cloud generative adversarial network based on tree struc + +tured graph convolutions. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3859-3868, 2019. 5 +[27] Minhyuk Sung, Hao Su, Vladimir G Kim, Siddhartha Chaudhuri, and Leonidas Guibas. Complementme: Weakly-supervised component suggestions for 3d modeling. ACM Transactions on Graphics (TOG), 36(6):1-12, 2017. 5, 6 +[28] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2, 4 +[29] Rundi Wu, Yixin Zhuang, Kai Xu, Hao Zhang, and Baoquan Chen. Pq-net: A generative part seq2seq network for 3d shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 829-838, 2020. 2, 5, 6 +[30] Yuwen Xiong, Wei-Chiu Ma, Jingkang Wang, and Raquel Urtasun. Learning compact representations for lidar completion and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1074-1083, 2023. 6 +[31] Xianghao Xu, Paul Guerrero, Matthew Fisher, Siddhartha Chaudhuri, and Daniel Ritchie. Unsupervised 3d shape reconstruction by part retrieval and assembly. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8559-8567, 2023. 2 +[32] Kevin Zakka, Andy Zeng, Johnny Lee, and Shuran Song. Form2fit: Learning shape priors for generalizable assembly from disassembly. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 9404–9410. IEEE, 2020. 1 +[33] Guanqi Zhan, Qingnan Fan, Kaichun Mo, Lin Shao, Baoquan Chen, Leonidas J Guibas, Hao Dong, et al. Generative 3d part assembly via dynamic graph learning. Advances in Neural Information Processing Systems, 33:6315-6326, 2020. 1, 2, 3, 4, 5, 6, 8 +[34] Rufeng Zhang, Tao Kong, Weihao Wang, Xuan Han, and Mingyu You. 3d part assembly generation with instance encoded transformer. IEEE Robotics and Automation Letters, 7(4):9051-9058, 2022. 1, 2, 3, 4, 5, 6, 8 +[35] Vlas Zyrianov, Xiyue Zhu, and Shenlong Wang. Learning to generate realistic lidar point clouds. In European Conference on Computer Vision, pages 17-35. Springer, 2022. 6 \ No newline at end of file diff --git a/2024/Generative 3D Part Assembly via Part-Whole-Hierarchy Message Passing/images.zip b/2024/Generative 3D Part Assembly via Part-Whole-Hierarchy Message Passing/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..581d98500eb446a5b673220f6aabf03e824bf603 --- /dev/null +++ b/2024/Generative 3D Part Assembly via Part-Whole-Hierarchy Message Passing/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:807275c0d403865cffaf16c96146c32e19f0188c8c6827ded413589f53a63788 +size 609983 diff --git a/2024/Generative 3D Part Assembly via Part-Whole-Hierarchy Message Passing/layout.json b/2024/Generative 3D Part Assembly via Part-Whole-Hierarchy Message Passing/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d7f79428137a311dda2e5dfacbb20da7c4bbae93 --- /dev/null +++ b/2024/Generative 3D Part Assembly via Part-Whole-Hierarchy Message Passing/layout.json @@ -0,0 +1,10075 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 69, + 103, + 523, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 103, + 523, + 121 + ], + "spans": [ + { + "bbox": [ + 69, + 103, + 523, + 121 + ], + "type": "text", + "content": "Generative 3D Part Assembly via Part-Whole-Hierarchy Message Passing" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 175, + 142, + 418, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 142, + 418, + 157 + ], + "spans": [ + { + "bbox": [ + 175, + 142, + 418, + 157 + ], + "type": "text", + "content": "Bi'an Du1, Xiang Gao1, Wei Hu1*, Renjie Liao2,3,4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 139, + 157, + 454, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 157, + 454, + 172 + ], + "spans": [ + { + "bbox": [ + 139, + 157, + 454, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 139, + 157, + 454, + 172 + ], + "type": "text", + "content": "Wangxuan Institute of Computer Technology, Peking University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 94, + 171, + 498, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 171, + 498, + 185 + ], + "spans": [ + { + "bbox": [ + 94, + 171, + 498, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 94, + 171, + 498, + 185 + ], + "type": "text", + "content": "University of British Columbia, " + }, + { + "bbox": [ + 94, + 171, + 498, + 185 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 94, + 171, + 498, + 185 + ], + "type": "text", + "content": "Vector Institute for AI, " + }, + { + "bbox": [ + 94, + 171, + 498, + 185 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 94, + 171, + 498, + 185 + ], + "type": "text", + "content": "Canada CIFAR AI Chair" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 188, + 520, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 188, + 520, + 198 + ], + "spans": [ + { + "bbox": [ + 69, + 188, + 520, + 198 + ], + "type": "text", + "content": "pkudba@stu.pku.edu.cn, gyshgx868@pku.edu.cn, forhuwei@pku.edu.cn, rjliao@ece.ubc.ca" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 252, + 290, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 290, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 290, + 529 + ], + "type": "text", + "content": "Generative 3D part assembly involves understanding part relationships and predicting their 6-DoF poses for assembling a realistic 3D shape. Prior work often focuses on the geometry of individual parts, neglecting part-whole hierarchies of objects. Leveraging two key observations: 1) super-part poses provide strong hints about part poses, and 2) predicting super-part poses is easier due to fewer superparts, we propose a part-whole-hierarchy message passing network for efficient 3D part assembly. We first introduce super-parts by grouping geometrically similar parts without any semantic labels. Then we employ a part-whole hierarchical encoder, wherein a super-part encoder predicts latent super-part poses based on input parts. Subsequently, we transform the point cloud using the latent poses, feeding it to the part encoder for aggregating super-part information and reasoning about part relationships to predict all part poses. In training, only ground-truth part poses are required. During inference, the predicted latent poses of super-parts enhance interpretability. Experimental results on the PartNet dataset show that our method achieves state-of-the-art performance in part and connectivity accuracy and enables an interpretable hierarchical part assembly. Code is available at https://github.com/pkudba/3DHPA." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 552, + 128, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 128, + 565 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 128, + 565 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 573, + 288, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 573, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 573, + 288, + 693 + ], + "type": "text", + "content": "Generative 3D Part Assembly [7, 15, 16, 25, 32] is an emerging research area that aims to generate complex 3D shapes via assembling simple 3D parts without relying on prior semantic knowledge. Different from traditional 3D shape generation, it focuses on generating diverse, plausible configurations of given parts. It facilitates the generation of complex objects and scenes with compositionality, flexibility, and efficiency. With the rapid evolution of 3D printing technology and the increasing demand for diverse 3D shapes, generative 3D part assembly finds applications in various sce" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 228, + 546, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 228, + 546, + 252 + ], + "spans": [ + { + "bbox": [ + 306, + 228, + 546, + 252 + ], + "type": "text", + "content": "narios, attracting attention from experts in computer vision, graphics, robotics, and machine learning." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 254, + 547, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 254, + 547, + 338 + ], + "spans": [ + { + "bbox": [ + 304, + 254, + 547, + 338 + ], + "type": "text", + "content": "However, achieving generative 3D part assembly presents a significant challenge due to the vast number of potential part arrangements and orientations, coupled with the intricate dependencies among the parts. Adding to the complexity, part geometry exhibits notable variation even within the same object category, making it exceedingly difficult to generalize learned assembly patterns across different objects." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 339, + 548, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 339, + 548, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 339, + 548, + 555 + ], + "type": "text", + "content": "Previous approaches primarily focus on designing architectures capable of learning powerful representations for individual 3D parts. The hope is that these learned representations could facilitate accurate part assembly, either in a one-shot manner or sequentially [14, 21, 33, 34]. However, these methods often ignore the inherent part-whole hierarchies in 3D shapes in representation learning. For instance, as illustrated in Figure 1, a chair consists of super-parts such as seats, backs, and legs, with each super-part further divisible into parts like seat surfaces and seat frames. Understanding the pose of a super-part provides insights into the poses of constituent parts within the same super-part, as they often share similar orientations or exhibit symmetry (e.g., left-right symmetry in chair legs and arms). Moreover, predicting super-part poses is typically easier due to the fewer number of super-parts compared to parts. Incorporating these hierarchies into the modeling would make the learning process and potentially improve the overall performance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 556, + 548, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 548, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 548, + 688 + ], + "type": "text", + "content": "In this paper, we introduce a part-whole-hierarchy message passing network for 3D shape assembly, predicting 6 degrees of freedom (6-DoF) poses for super-parts and parts in a hierarchical manner. We establish the correspondence between parts and super-parts (i.e., subsets of parts) by grouping parts based on their geometric similarities, following the approach from [33, 34]. Importantly, we treat super-part poses as latent variables to be learned, thereby eliminating the requirement of ground-truth super-part poses in our work. Our model comprises two sequential modules: the super-part encoder and the part encoder." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "content": "The first super-part encoder takes the 3D point clouds of all parts as input and employs the attention-based message" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 247, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 247, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 247, + 713 + ], + "type": "text", + "content": "*Corresponding Author: Wei Hu (forhuwei@pku.edu.cn)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20850" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 73, + 539, + 233 + ], + "blocks": [ + { + "bbox": [ + 59, + 73, + 539, + 233 + ], + "lines": [ + { + "bbox": [ + 59, + 73, + 539, + 233 + ], + "spans": [ + { + "bbox": [ + 59, + 73, + 539, + 233 + ], + "type": "image", + "image_path": "a113e8d42477061455253f57651ce08ff4f9753ac7e0eaead656a8e64c613940.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 236, + 547, + 259 + ], + "lines": [ + { + "bbox": [ + 46, + 236, + 547, + 259 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 547, + 259 + ], + "type": "text", + "content": "Figure 1. Left: an illustration of the part-whole-hierarchy for 3D shapes; Right: the part assembly process via the proposed part-whole-hierarchy message passing network." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 271, + 288, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 271, + 288, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 271, + 288, + 319 + ], + "type": "text", + "content": "passing to predict super-parts poses. The representation of each super-part is aggregated from representations of parts within it. These super-part poses provide initial estimation for the poses of individual parts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 319, + 288, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 319, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 288, + 403 + ], + "type": "text", + "content": "Subsequently, the whole point clouds are transformed based on the predicted super-part poses and fed to the part encoder. This module, using a cross-attention mechanism, extracts features from the transformed point cloud, transferring super-part level information to the part level. It then leverages the attention-based message passing again to capture relationships among individual parts." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 404, + 288, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 404, + 288, + 536 + ], + "spans": [ + { + "bbox": [ + 46, + 404, + 288, + 536 + ], + "type": "text", + "content": "Following previous works [14, 21, 33, 34], we train and evaluate our model on the PartNet dataset [18]. Overall, our model achieves the state-of-the-art performances, outperforming the strongest competitor [34] by a significant margin, i.e., with an almost " + }, + { + "bbox": [ + 46, + 404, + 288, + 536 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 46, + 404, + 288, + 536 + ], + "type": "text", + "content": " improvement in mean part accuracy and a " + }, + { + "bbox": [ + 46, + 404, + 288, + 536 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 46, + 404, + 288, + 536 + ], + "type": "text", + "content": " improvement in mean connectivity accuracy. Moreover, through visual analysis of the assembly process for both super-parts and parts, we not only showcase accurate generation of part poses but also demonstrate interpretability via the predicted super-part poses. This interpretability feature further improves the utility of our model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 548, + 134, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 548, + 134, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 548, + 134, + 562 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 570, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 289, + 713 + ], + "type": "text", + "content": "Assembly3D modeling Numerous prior studies have approached the challenging 3D part assembly through the joint estimation of part poses. The influential work by [7] introduces an intelligent scissoring technique to tackle this problem for part components. Subsequent research by [2, 10, 12] employs graphical models to capture the semantic and geometric relationships among shape components, enabling exploration in assembly-based shape modeling. PAGENet [14] presents a network that is aware of individual parts and generates semantic parts along with their poses. [31] proposes to decompose shapes using a library of 3D parts provided by the user. However, these studies either assumed prior knowl-" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 271, + 547, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 271, + 547, + 415 + ], + "spans": [ + { + "bbox": [ + 304, + 271, + 547, + 415 + ], + "type": "text", + "content": "edge of part semantics or relied on existing shape databases. In a more practical setting, the authors in [3, 21, 33, 34] focus on the pose estimation of individual parts without relying on shape databases or known semantic information. Specifically, DGL [33] employs a dynamic part graph to iteratively refine the poses of individual parts. A progressive strategy using the recurrent graph learning framework has been investigated in [21]. Additionally, the authors in [3, 34] utilizes the Transformer [28] to model the structural relationships and performs the simultaneous assembly of all parts. We follow this line of work and propose novel improvement to generate structurally-coherent part assemblies." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 430, + 548, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 430, + 548, + 658 + ], + "spans": [ + { + "bbox": [ + 304, + 430, + 548, + 658 + ], + "type": "text", + "content": "Structural Shape Generation In recent years, deep generative models, such as generative adversarial networks (GANs) [9] and variational autoencoders (VAEs) [5], have garnered significant attention for shape generation tasks. Notably, the work by [8] introduces a two-level variational autoencoder that simultaneously learns the overall shape structure and detailed part geometries. Both GRASS [13] and StructureNet [17] employ techniques to compress the shape structure into a latent space while considering the relationships between different parts. Additionally, [19] uses a part-tree decomposition to conditionally generate 3D shapes, and [11] adopts a procedural programmatic representation to establish connections between part cuboids. Inspired by the Seq2Seq networks in machine translation, [29] introduces a sequential encoding and decoding approach for the regression of shape parameters. While many of these approaches focus on directly generating new part shapes given random latent codes, our main focus revolves around the rigid transformation of existing parts to facilitate their assembly." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 669, + 525, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 669, + 525, + 683 + ], + "spans": [ + { + "bbox": [ + 305, + 669, + 525, + 683 + ], + "type": "text", + "content": "3. Part-Whole-Hierarchy Message Passing" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 689, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 546, + 715 + ], + "type": "text", + "content": "This section provides a detailed explanation of our proposed part-whole-hierarchy message passing network. We begin by" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20851" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 70, + 539, + 309 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 539, + 309 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 539, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 539, + 309 + ], + "type": "image", + "image_path": "78accfc024b593cb18524471659a2a24ecd2cfd53c66d1699618b441aaca29ba.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 312, + 547, + 357 + ], + "lines": [ + { + "bbox": [ + 46, + 312, + 547, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 547, + 357 + ], + "type": "text", + "content": "Figure 2. The overall architecture of our model consists of two modules: super-part encoder and part encoder. We first obtain super-parts via grouping parts based on their geometric similarities in an unsupervised fashion. The super-part encoder takes point cloud as input and predicts latent super-part poses (no ground truth is needed). The point cloud is then transformed based on super-part poses and fed to the part encoder. We incorporate both cross-level and within-level attention in the part encoder to predict part poses." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 369, + 289, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 289, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 289, + 453 + ], + "type": "text", + "content": "presenting how we construct super-parts in an unsupervised manner. Then we introduce our super-part encoder that predicts the latent poses of super-parts. Subsequently, we describe how our part encoder leverages the latent poses and the part-whole hierarchy to predict part poses. The overall model is illustrated in Figure 2. Finally, we introduce the loss functions and the training process for diverse generation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 464, + 187, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 464, + 187, + 478 + ], + "spans": [ + { + "bbox": [ + 47, + 464, + 187, + 478 + ], + "type": "text", + "content": "3.1. Super-Part Construction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 484, + 288, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 484, + 288, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 484, + 288, + 664 + ], + "type": "text", + "content": "We denote the input point clouds as " + }, + { + "bbox": [ + 46, + 484, + 288, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\{\\mathbf{P}_i|i = 1\\ldots N\\}" + }, + { + "bbox": [ + 46, + 484, + 288, + 664 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 484, + 288, + 664 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_i\\in \\mathbb{R}^{d\\times 3}" + }, + { + "bbox": [ + 46, + 484, + 288, + 664 + ], + "type": "text", + "content": " corresponds to the point cloud of the " + }, + { + "bbox": [ + 46, + 484, + 288, + 664 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 484, + 288, + 664 + ], + "type": "text", + "content": "-th given part of the 3D shape. The values of " + }, + { + "bbox": [ + 46, + 484, + 288, + 664 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 484, + 288, + 664 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 484, + 288, + 664 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 484, + 288, + 664 + ], + "type": "text", + "content": " represent the number of parts and the number of points per part respectively. Note that their values may differ from object to object. For notation convenience, we assume objects are padded to the maximum number of parts, and each part is padded to the maximum number of points per part. The goal of our task is to predict a set of 6-DoF part poses " + }, + { + "bbox": [ + 46, + 484, + 288, + 664 + ], + "type": "inline_equation", + "content": "\\{(t_i,r_i)\\}_{i = 1}^N" + }, + { + "bbox": [ + 46, + 484, + 288, + 664 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 484, + 288, + 664 + ], + "type": "inline_equation", + "content": "t_i\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 484, + 288, + 664 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 484, + 288, + 664 + ], + "type": "inline_equation", + "content": "r_i\\in \\mathbb{R}^4" + }, + { + "bbox": [ + 46, + 484, + 288, + 664 + ], + "type": "text", + "content": " represent the translation and rigid rotation for each part, respectively. The complete part assembly for a 3D shape is " + }, + { + "bbox": [ + 46, + 484, + 288, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{S} = \\cup_{i = 1}^{N}\\mathbf{T}_i(\\mathbf{P}_i)" + }, + { + "bbox": [ + 46, + 484, + 288, + 664 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 484, + 288, + 664 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_i" + }, + { + "bbox": [ + 46, + 484, + 288, + 664 + ], + "type": "text", + "content": " represents a transformation in " + }, + { + "bbox": [ + 46, + 484, + 288, + 664 + ], + "type": "inline_equation", + "content": "SE(3)" + }, + { + "bbox": [ + 46, + 484, + 288, + 664 + ], + "type": "text", + "content": " that consists of a 3D rotation in the rotation group " + }, + { + "bbox": [ + 46, + 484, + 288, + 664 + ], + "type": "inline_equation", + "content": "SO(3)" + }, + { + "bbox": [ + 46, + 484, + 288, + 664 + ], + "type": "text", + "content": " and a 3D translation in the translation group, induced by " + }, + { + "bbox": [ + 46, + 484, + 288, + 664 + ], + "type": "inline_equation", + "content": "(t_i,r_i)" + }, + { + "bbox": [ + 46, + 484, + 288, + 664 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "text", + "content": "A super-part is a subset of parts that is ideally semantically meaningful. However, since we do not have groundtruth super-parts, we need to construct them in an unsupervised manner. In particular, we compute axis-aligned" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 369, + 547, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 369, + 547, + 563 + ], + "spans": [ + { + "bbox": [ + 304, + 369, + 547, + 563 + ], + "type": "text", + "content": "bounding boxes for each part and evaluate the similarity between these 3D boxes. Parts are grouped into the same super-part if the difference between their respective enclosing boxes is below a specified threshold. Following previous works [33, 34], we set the threshold to 0.2. Although the super-parts obtained through this method may lack semantic meaning, they provide a coarse abstraction grounded in geometry similarities without any supervised labels. Then we group the input point cloud parts " + }, + { + "bbox": [ + 304, + 369, + 547, + 563 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\{\\mathbf{P}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 304, + 369, + 547, + 563 + ], + "type": "text", + "content": " into a set of " + }, + { + "bbox": [ + 304, + 369, + 547, + 563 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 369, + 547, + 563 + ], + "type": "text", + "content": " super-parts " + }, + { + "bbox": [ + 304, + 369, + 547, + 563 + ], + "type": "inline_equation", + "content": "\\mathcal{P}' = \\{\\mathbf{P}_i'\\}_{i=1}^M" + }, + { + "bbox": [ + 304, + 369, + 547, + 563 + ], + "type": "text", + "content": " based on the part-whole hierarchy of a given object. Here " + }, + { + "bbox": [ + 304, + 369, + 547, + 563 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_i'" + }, + { + "bbox": [ + 304, + 369, + 547, + 563 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 304, + 369, + 547, + 563 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 369, + 547, + 563 + ], + "type": "text", + "content": "-th super-part. To maintain consistent notation, we again pad the number of parts per super-part to the maximum " + }, + { + "bbox": [ + 304, + 369, + 547, + 563 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 369, + 547, + 563 + ], + "type": "text", + "content": ". Note that padding is used here for clarity and understanding. In practice, our model can handle sets of parts with varying sizes without padding." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 573, + 476, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 573, + 476, + 586 + ], + "spans": [ + { + "bbox": [ + 305, + 573, + 476, + 586 + ], + "type": "text", + "content": "3.2. Super-Part Relation Reasoning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 593, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 593, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 593, + 545, + 628 + ], + "type": "text", + "content": "To reason about the relationships among super-parts, we utilize the super-part encoder. The computational process is illustrated in the top half of Figure 2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "Given a set of point clouds of a super-part " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_i^{\\prime}" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": ", we first compute the set-level permutation-invariant representation " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i^{\\prime}" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": " via a shared PointNet [22]. We use the same architecture and hyperparameters as in previous works [33, 34] for a fair comparison. This feature representation captures the global characteristics of each super-part while being invariant to the order of the parts within the set. Based on the features" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20852" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i^{\\prime}" + }, + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "type": "text", + "content": " extracted for each super-part from the PointNet, we use a two-layer multi-head attention (MHA) module to learn the relationships between super-parts. Specifically, following [28], we calculate an attention matrix " + }, + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "type": "inline_equation", + "content": "\\mathbf{A}^{\\prime}" + }, + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{i,j}^{\\prime}" + }, + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "type": "text", + "content": " represents the attention weight of the " + }, + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "type": "text", + "content": "-th super-part to the " + }, + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "type": "text", + "content": "-th super-part. Then we multiply each super-part feature " + }, + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_j^{\\prime}" + }, + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{i,j}^{\\prime}" + }, + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "type": "text", + "content": " and sum them up to obtain the attention-weighted feature " + }, + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_i^{\\prime}" + }, + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "type": "text", + "content": ". In our model, we set the embedding dimension to 256 and the number of heads for the MHA module to 8." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 179, + 290, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 179, + 290, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 179, + 290, + 281 + ], + "type": "text", + "content": "For super-part pose prediction, we feed the attention-weighted feature " + }, + { + "bbox": [ + 46, + 179, + 290, + 281 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_i'" + }, + { + "bbox": [ + 46, + 179, + 290, + 281 + ], + "type": "text", + "content": " into a prediction head containing 4 fully-connected layers to obtain the pose " + }, + { + "bbox": [ + 46, + 179, + 290, + 281 + ], + "type": "inline_equation", + "content": "\\{(t_i', r_i')\\}_{i=1}^M" + }, + { + "bbox": [ + 46, + 179, + 290, + 281 + ], + "type": "text", + "content": ". We apply tanh operation to the translation vector, which restricts the range of the part center offset as " + }, + { + "bbox": [ + 46, + 179, + 290, + 281 + ], + "type": "inline_equation", + "content": "(-1,1)" + }, + { + "bbox": [ + 46, + 179, + 290, + 281 + ], + "type": "text", + "content": ". Additionally, for simplicity, we predict the quaternion vector " + }, + { + "bbox": [ + 46, + 179, + 290, + 281 + ], + "type": "inline_equation", + "content": "r_i' = (r_{i0}', r_{i1}', r_{i2}', r_{i3}')" + }, + { + "bbox": [ + 46, + 179, + 290, + 281 + ], + "type": "text", + "content": " instead of the rotation matrix. We then use the Rodrigues formula [23] to obtain the rotation matrix " + }, + { + "bbox": [ + 46, + 179, + 290, + 281 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_i'" + }, + { + "bbox": [ + 46, + 179, + 290, + 281 + ], + "type": "text", + "content": " corresponding to " + }, + { + "bbox": [ + 46, + 179, + 290, + 281 + ], + "type": "inline_equation", + "content": "r_i'" + }, + { + "bbox": [ + 46, + 179, + 290, + 281 + ], + "type": "text", + "content": " one by one as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 289, + 291, + 327 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 289, + 291, + 327 + ], + "spans": [ + { + "bbox": [ + 47, + 289, + 291, + 327 + ], + "type": "interline_equation", + "content": "\\mathbf {R} _ {i} ^ {\\prime} = \\left[ \\begin{array}{l} 1 - 2 r _ {i 2} ^ {\\prime 2} - 2 r _ {i 3} ^ {\\prime 2}, 2 r _ {i 1} ^ {\\prime} r _ {i 2} ^ {\\prime} - 2 r _ {i 0} ^ {\\prime} r _ {i 3} ^ {\\prime}, 2 r _ {i 1} ^ {\\prime} r _ {i 3} ^ {\\prime} + 2 r _ {i 0} ^ {\\prime} r _ {i 2} ^ {\\prime} \\\\ 2 r _ {i 1} ^ {\\prime} r _ {i 2} ^ {\\prime} + 2 r _ {i 0} ^ {\\prime} r _ {i 3} ^ {\\prime}, 1 - 2 r _ {i 1} ^ {\\prime 2} - 2 r _ {i 3} ^ {\\prime 2}, 2 r _ {i 2} ^ {\\prime} r _ {i 3} ^ {\\prime} - 2 r _ {i 0} ^ {\\prime} r _ {i 1} ^ {\\prime} \\\\ 2 r _ {i 1} ^ {\\prime} r _ {i 3} ^ {\\prime} - 2 r _ {i 0} ^ {\\prime} r _ {i 2} ^ {\\prime}, 2 r _ {i 2} ^ {\\prime} r _ {i 3} ^ {\\prime} + 2 r _ {i 0} ^ {\\prime} r _ {i 1} ^ {\\prime}, 1 - 2 r _ {i 1} ^ {\\prime 2} - 2 r _ {i 2} ^ {\\prime 2}. \\end{array} \\right]", + "image_path": "bb67998e240ae4a03f51c37d537ab78e715a6eaa335fbfd80b986d151156f1d6.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 333, + 288, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 333, + 288, + 370 + ], + "spans": [ + { + "bbox": [ + 46, + 333, + 288, + 370 + ], + "type": "text", + "content": "The transformed point cloud of the super-part is expressed as " + }, + { + "bbox": [ + 46, + 333, + 288, + 370 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_i(\\mathbf{P}_i^{\\prime}) = \\mathbf{R}_i^{\\prime}\\mathbf{P}_i^{\\prime} + t_i^{\\prime}" + }, + { + "bbox": [ + 46, + 333, + 288, + 370 + ], + "type": "text", + "content": ". To ensure that it is a unit quaternion, we normalize the rigid rotation vector such that " + }, + { + "bbox": [ + 46, + 333, + 288, + 370 + ], + "type": "inline_equation", + "content": "\\| r_i^\\prime \\| = 1" + }, + { + "bbox": [ + 46, + 333, + 288, + 370 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 370, + 288, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 370, + 288, + 417 + ], + "spans": [ + { + "bbox": [ + 46, + 370, + 288, + 417 + ], + "type": "text", + "content": "Since we do not have ground-truth super-part poses, the supervision for this module solely comes from the backpropagation signal from the subsequent module. As a result, the super-part poses are considered as latent variables." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 425, + 185, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 425, + 185, + 438 + ], + "spans": [ + { + "bbox": [ + 47, + 425, + 185, + 438 + ], + "type": "text", + "content": "3.3. Part Relation Reasoning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 444, + 289, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 444, + 289, + 492 + ], + "spans": [ + { + "bbox": [ + 46, + 444, + 289, + 492 + ], + "type": "text", + "content": "As previously mentioned, the poses of super-parts offer valuable clues about the poses of their corresponding parts. To leverage this information, we design a part encoder to incorporate the latent super-part poses to predict the part poses." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 506, + 289, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 289, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 289, + 625 + ], + "type": "text", + "content": "Transformation via Latent Super-Part Poses Firstly, we apply the latent super-part poses to transform all point cloud parts, resulting in transformed parts denoted as " + }, + { + "bbox": [ + 46, + 506, + 289, + 625 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{P}} = \\{\\hat{\\mathbf{P}}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 46, + 506, + 289, + 625 + ], + "type": "text", + "content": ". These transformed parts are then fed into the part encoder for further processing. Similarly to the super-part encoder, we employ another PointNet to extract part-level feature denoted as " + }, + { + "bbox": [ + 46, + 506, + 289, + 625 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_i" + }, + { + "bbox": [ + 46, + 506, + 289, + 625 + ], + "type": "text", + "content": ". In order to better utilize the part-whole hierarchy of objects, we design two ways to integrate the super-part level information to the part level, i.e., cross-level and within-level attention modules." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 641, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 289, + 715 + ], + "type": "text", + "content": "Cross-Level Attention Part representations " + }, + { + "bbox": [ + 46, + 641, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_i" + }, + { + "bbox": [ + 46, + 641, + 289, + 715 + ], + "type": "text", + "content": " from the PointNet encoder does not leverage the part-whole hierarchy. To better fuse the information from the super-parts, we propose a cross-level attention module. Specifically, we treat the part representation " + }, + { + "bbox": [ + 46, + 641, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_i" + }, + { + "bbox": [ + 46, + 641, + 289, + 715 + ], + "type": "text", + "content": " as query and super-part representations " + }, + { + "bbox": [ + 46, + 641, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i'" + }, + { + "bbox": [ + 46, + 641, + 289, + 715 + ], + "type": "text", + "content": " as keys and values. We then employ a MHA module" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "content": "to perform an attention-weighted aggregation of super-part representations " + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i^{\\prime}" + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "content": " to obtain the updated part representations " + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i" + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "content": ". By doing so, the information from part-whole hierarchy is explicitly integrated into part representations. Here we specify the embedding dimension as 256 and the MHA module is set to utilize 8 heads." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 159, + 548, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 159, + 548, + 410 + ], + "spans": [ + { + "bbox": [ + 304, + 159, + 548, + 410 + ], + "type": "text", + "content": "Within-Level Attention To model the part relationships, we leverage another within-level attention module, which is depicted in the bottom half of Figure 2. We first concatenate the representation from the previous module " + }, + { + "bbox": [ + 304, + 159, + 548, + 410 + ], + "type": "inline_equation", + "content": "F_{i}" + }, + { + "bbox": [ + 304, + 159, + 548, + 410 + ], + "type": "text", + "content": " with an additional instance encoding vector that is unique for each part following [34]. This allows us to distinguish geometrically-equivalent parts. Then we perform message passing among parts using another MHA module where the attention computation is aware of the part-whole hierarchy. Specifically, we concatenate the latent super-part poses to the part-level feature before computing the attention scores at each layer. Consequently, in each message passing layer, a part receives pose hints from its parent super-part and updates its representation. The resulting part feature is denoted as " + }, + { + "bbox": [ + 304, + 159, + 548, + 410 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_i = \\sum_{j=1}^N \\mathbf{A}_{i,j} \\mathbf{F}_i" + }, + { + "bbox": [ + 304, + 159, + 548, + 410 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 159, + 548, + 410 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{i,j}" + }, + { + "bbox": [ + 304, + 159, + 548, + 410 + ], + "type": "text", + "content": " represents the attention weight of the " + }, + { + "bbox": [ + 304, + 159, + 548, + 410 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 159, + 548, + 410 + ], + "type": "text", + "content": "-th part to the " + }, + { + "bbox": [ + 304, + 159, + 548, + 410 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 159, + 548, + 410 + ], + "type": "text", + "content": "-th part. This design enables the part-to-part message passing to effectively leverage the part-whole hierarchy information. We set the number of heads for the MHA module to 8, the dimension of the part-level feature to 256, and concatenate it with the 40-dimensional instance encoding vector." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 424, + 547, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 424, + 547, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 424, + 547, + 579 + ], + "type": "text", + "content": "Prediction Head Finally, we feed the attention-weighted part-level feature " + }, + { + "bbox": [ + 304, + 424, + 547, + 579 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_i" + }, + { + "bbox": [ + 304, + 424, + 547, + 579 + ], + "type": "text", + "content": " to a part pose prediction head, which consists of four fully-connected layers. The channel sizes of the first three fully-connected layers are set to 256, 256 and 1024, respectively, followed by ReLU activation functions. We use a linear projection layer to predict the part pose " + }, + { + "bbox": [ + 304, + 424, + 547, + 579 + ], + "type": "inline_equation", + "content": "(t_i, r_i)" + }, + { + "bbox": [ + 304, + 424, + 547, + 579 + ], + "type": "text", + "content": ", encompassing a 4-dimensional rigid rotation and 3-dimensional translation. This prediction head outputs the part poses " + }, + { + "bbox": [ + 304, + 424, + 547, + 579 + ], + "type": "inline_equation", + "content": "\\{(t_i, r_i)\\}_{i=1}^N" + }, + { + "bbox": [ + 304, + 424, + 547, + 579 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 424, + 547, + 579 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 304, + 424, + 547, + 579 + ], + "type": "text", + "content": " represents the translation vector and " + }, + { + "bbox": [ + 304, + 424, + 547, + 579 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 304, + 424, + 547, + 579 + ], + "type": "text", + "content": " represents the rigid rotation vector for each part. We adopt the same design as in the super-part encoder to restrict the translation vectors to the range of " + }, + { + "bbox": [ + 304, + 424, + 547, + 579 + ], + "type": "inline_equation", + "content": "(-1, 1)" + }, + { + "bbox": [ + 304, + 424, + 547, + 579 + ], + "type": "text", + "content": " and normalize the rigid rotation vector " + }, + { + "bbox": [ + 304, + 424, + 547, + 579 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 304, + 424, + 547, + 579 + ], + "type": "text", + "content": " to have a unit norm." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 586, + 527, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 586, + 527, + 600 + ], + "spans": [ + { + "bbox": [ + 305, + 586, + 527, + 600 + ], + "type": "text", + "content": "3.4. Training Objective for Diverse Generation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": "Considering the same set of input point cloud parts, geometrically equivalent parts (such as legs of a chair) can be interchanged and decorated parts may have multiple placement options, resulting in multiple possible shapes. For example, a semi-cylindrical part can be placed on top of a chair backrest as a headrest, or placed under the backrest as a lumbar pillow. To account for such diversified structural variations and configurations, we introduce random noise into the system following [21, 33, 34] and employ the Min-of-N" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20853" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": "(MoN) loss [6] to simultaneously consider assembly accuracy and assembly diversity. By considering the minimum of a set of N possible assemblies, we can encourage the model to generate diverse and valid shape configurations while maintaining overall accuracy in the assembly process. This helps address the challenge of capturing multiple plausible solutions that arise from the interchangeable and decoratable nature of the parts." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 167, + 288, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 167, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 47, + 167, + 288, + 192 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 47, + 167, + 288, + 192 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\mathcal{P},\\delta_j)" + }, + { + "bbox": [ + 47, + 167, + 288, + 192 + ], + "type": "text", + "content": " denote our network outputs and " + }, + { + "bbox": [ + 47, + 167, + 288, + 192 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^* (\\mathcal{P})" + }, + { + "bbox": [ + 47, + 167, + 288, + 192 + ], + "type": "text", + "content": " denote the ground-truth point clouds, then the MoN loss is:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 83, + 197, + 287, + 217 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 197, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 83, + 197, + 287, + 217 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {M o N}} = \\min _ {\\delta_ {j} \\sim \\mathcal {N} (0, 1)} \\mathcal {L} \\left(\\mathcal {F} \\left(\\mathcal {P}, \\delta_ {j}\\right), \\mathcal {F} ^ {*} (\\mathcal {P})\\right) \\tag {1}", + "image_path": "8c9534f51660d187b208d5f5e01f5f4ae2b15fedcf7c63aa59bbdc7e1d1e2c87.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 221, + 288, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 221, + 288, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 221, + 288, + 281 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 221, + 288, + 281 + ], + "type": "inline_equation", + "content": "\\delta_{j}" + }, + { + "bbox": [ + 46, + 221, + 288, + 281 + ], + "type": "text", + "content": " is a random noise vector drawn from standard Normal distribution in an IID fashion. Following [34], we sample 5 random vectors " + }, + { + "bbox": [ + 46, + 221, + 288, + 281 + ], + "type": "inline_equation", + "content": "\\delta_{j}" + }, + { + "bbox": [ + 46, + 221, + 288, + 281 + ], + "type": "text", + "content": " during training. The loss function " + }, + { + "bbox": [ + 46, + 221, + 288, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 46, + 221, + 288, + 281 + ], + "type": "text", + "content": " consists of the following components for both local part and global shape losses." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 281, + 287, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 287, + 316 + ], + "type": "text", + "content": "Firstly, we supervise the translation via an " + }, + { + "bbox": [ + 46, + 281, + 287, + 316 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 46, + 281, + 287, + 316 + ], + "type": "text", + "content": " loss between our prediction " + }, + { + "bbox": [ + 46, + 281, + 287, + 316 + ], + "type": "inline_equation", + "content": "t_{i}" + }, + { + "bbox": [ + 46, + 281, + 287, + 316 + ], + "type": "text", + "content": " and the ground-truth translation vector " + }, + { + "bbox": [ + 46, + 281, + 287, + 316 + ], + "type": "inline_equation", + "content": "t_{i}^{*}" + }, + { + "bbox": [ + 46, + 281, + 287, + 316 + ], + "type": "text", + "content": " for each part:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 124, + 316, + 287, + 348 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 316, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 124, + 316, + 287, + 348 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {t} = \\sum_ {i = 1} ^ {N} \\| t _ {i} - t _ {i} ^ {*} \\| _ {2} ^ {2} \\tag {2}", + "image_path": "de013c0ce545643f95bd664b6e2465abe36283b130e603994dc707333d1843c4.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 350, + 287, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 350, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 47, + 350, + 287, + 374 + ], + "type": "text", + "content": "Then we compute the Chamfer distance [6] between the predicted and the ground-truth rigid rotation for each part:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 379, + 287, + 411 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 379, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 104, + 379, + 287, + 411 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r} = \\sum_ {i = 1} ^ {N} d _ {c} \\left(r _ {i} \\left(\\mathbf {P} _ {i}\\right), r _ {i} ^ {*} \\left(\\mathbf {P} _ {i}\\right)\\right) \\tag {3}", + "image_path": "c4a0c8c6c6f21cdbc20364a52f581ffbbc8f2e18aa61edd6e977c79d6b11cf1a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 416, + 287, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 416, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 416, + 287, + 475 + ], + "type": "text", + "content": "The Chamfer distance " + }, + { + "bbox": [ + 46, + 416, + 287, + 475 + ], + "type": "inline_equation", + "content": "d_{c}(\\mathcal{X},\\mathcal{Y})" + }, + { + "bbox": [ + 46, + 416, + 287, + 475 + ], + "type": "text", + "content": " is a metric commonly used in point cloud comparison, which measures the dissimilarity between two point sets " + }, + { + "bbox": [ + 46, + 416, + 287, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 46, + 416, + 287, + 475 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 416, + 287, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}" + }, + { + "bbox": [ + 46, + 416, + 287, + 475 + ], + "type": "text", + "content": " by calculating the average distance between each point in one set and its nearest neighbor in the other set as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 480, + 287, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 480, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 53, + 480, + 287, + 506 + ], + "type": "interline_equation", + "content": "d _ {c} (\\mathcal {X}, \\mathcal {Y}) = \\sum_ {x \\in \\mathcal {X}} \\min _ {y \\in \\mathcal {Y}} \\| x - y \\| _ {2} ^ {2} + \\sum_ {y \\in \\mathcal {Y}} \\min _ {x \\in \\mathcal {X}} \\| x - y \\| _ {2} ^ {2} \\tag {4}", + "image_path": "7b66b26d3b53fdf6cee5cf3ded988178492312036dd1283ea604735ca64e79d3.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 511, + 287, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 511, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 46, + 511, + 287, + 547 + ], + "type": "text", + "content": "In this case, we use the Chamfer distance to assess the discrepancy between the predicted rotation of each part and its corresponding ground-truth rotation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 547, + 287, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 547, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 547, + 287, + 582 + ], + "type": "text", + "content": "Similarly, the full shape " + }, + { + "bbox": [ + 46, + 547, + 287, + 582 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 547, + 287, + 582 + ], + "type": "text", + "content": ", i.e., the set of point clouds belonging to all parts, is supervised via the Chamfer distance from the ground-truth shape " + }, + { + "bbox": [ + 46, + 547, + 287, + 582 + ], + "type": "inline_equation", + "content": "S^{*}" + }, + { + "bbox": [ + 46, + 547, + 287, + 582 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 588, + 287, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 287, + 601 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s} = d _ {c} \\left(\\mathcal {S}, \\mathcal {S} ^ {*}\\right) \\tag {5}", + "image_path": "a7f40cc504d0e95996f0012675461ec0dc962ae54470730af0f35ed73cf569f8.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 59, + 607, + 265, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 607, + 265, + 618 + ], + "spans": [ + { + "bbox": [ + 59, + 607, + 265, + 618 + ], + "type": "text", + "content": "In summary, the overall loss function is defined as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 111, + 624, + 287, + 636 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 624, + 287, + 636 + ], + "spans": [ + { + "bbox": [ + 111, + 624, + 287, + 636 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {t} \\mathcal {L} _ {t} + \\lambda_ {r} \\mathcal {L} _ {r} + \\lambda_ {s} \\mathcal {L} _ {s} \\tag {6}", + "image_path": "c9ff867650d1876a47bb4d3829557267411880b7d5a749658cec31dcd8e4c106.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{t},\\lambda_{r}" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{s}" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": " represent the weights assigned to the three loss terms. Based on the cross validation, we empirically set " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_t = 1" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_r = 10" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": " ,and " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_s = 1" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": " in our experiments. By minimizing this loss function, we aim to improve the alignment of the predicted poses with the ground-truth poses, thereby enhancing the overall accuracy of the assembly." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 72, + 388, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 388, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 388, + 85 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 91, + 547, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 547, + 152 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 547, + 152 + ], + "type": "text", + "content": "In this section, we demonstrate the effectiveness of the proposed model by comparing our results with state-of-the-art methods. We also provide visual analysis of the assembly process for super-parts and parts, which helps to illustrate the hierarchical assembly process from parts to the whole." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 161, + 364, + 172 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 161, + 364, + 172 + ], + "spans": [ + { + "bbox": [ + 306, + 161, + 364, + 172 + ], + "type": "text", + "content": "4.1. Dataset" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 179, + 547, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 179, + 547, + 324 + ], + "spans": [ + { + "bbox": [ + 304, + 179, + 547, + 324 + ], + "type": "text", + "content": "Following [4, 14, 21, 24, 27, 29, 33, 34], we use the PartNet [18] dataset for both training and evaluation. This dataset consists of 26,671 shapes across 24 different 3D object categories. To effectively validate and compare different methods, we select the three largest categories, i.e., chairs, tables and lamps, following [4, 14, 21, 33, 34]. In total, we have 6,323 chairs, 8,218 tables, 2,207 lamps in the finest-grained level, and we adopt the official train/validation/test splits (70%/10%/20%) to conduct the experiments. For each part point cloud, 1000 points are sampled from the original part meshes using Farthest Point Sampling (FPS) [20]. All parts are transformed into its canonical space using PCA." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 333, + 418, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 333, + 418, + 344 + ], + "spans": [ + { + "bbox": [ + 306, + 333, + 418, + 344 + ], + "type": "text", + "content": "4.2. Evaluation Metrics" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 352, + 547, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 352, + 547, + 508 + ], + "spans": [ + { + "bbox": [ + 304, + 352, + 547, + 508 + ], + "type": "text", + "content": "We generate a variety of shapes by adding different Gaussian noises to a given set of input parts, and find the closest shape to the ground-truth using minimum matching distance (MMD) [1]. To measure part assembly quality, we follow [21, 33, 34] to use shape Chamfer distance (SCD), part accuracy (PA) and connectivity accuracy (CA). Shape Chamfer distance is defined in Equation (4) and Equation (5). Meanwhile, to compare the diversity of assembled parts, we propose two quantitative evaluation metrics including quality-diversity score (QDS) and the weighted quality-diversity score (WQDS) following [4]. We also visualize the generation results to qualitatively evaluate the diversity. More details of these metrics are provided in the appendix." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 525, + 547, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 525, + 547, + 610 + ], + "spans": [ + { + "bbox": [ + 304, + 525, + 547, + 610 + ], + "type": "text", + "content": "Mean Part/Connectivity Accuracy (mPA/mCA). PA and CA depend on the Chamfer distance threshold " + }, + { + "bbox": [ + 304, + 525, + 547, + 610 + ], + "type": "inline_equation", + "content": "\\tau_{p}" + }, + { + "bbox": [ + 304, + 525, + 547, + 610 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 525, + 547, + 610 + ], + "type": "inline_equation", + "content": "\\tau_{c}" + }, + { + "bbox": [ + 304, + 525, + 547, + 610 + ], + "type": "text", + "content": " for judging whether the assembly is accurate. In order to provide a more comprehensive evaluation of the performance of the model, we average the results under multiple thresholds to get mean part accuracy (mPA) and mean connectivity accuracy (mCA), formally," + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 310, + 619, + 539, + 649 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 619, + 539, + 649 + ], + "spans": [ + { + "bbox": [ + 310, + 619, + 539, + 649 + ], + "type": "interline_equation", + "content": "\\mathrm {m P A} = \\frac {1}{T _ {p}} \\sum_ {\\tau_ {p} \\in T _ {p}} \\mathrm {P A} (\\tau_ {p}), \\quad \\mathrm {m C A} = \\frac {1}{T _ {c}} \\sum_ {\\tau_ {c} \\in T _ {c}} \\mathrm {C A} (\\tau_ {c}),", + "image_path": "b2f118b807193675c22ea698f0552c2db658c97a85833968683209e931ace5b8.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 659, + 534, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 659, + 534, + 672 + ], + "spans": [ + { + "bbox": [ + 306, + 659, + 534, + 672 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 659, + 534, + 672 + ], + "type": "inline_equation", + "content": "T_{p}" + }, + { + "bbox": [ + 306, + 659, + 534, + 672 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 659, + 534, + 672 + ], + "type": "inline_equation", + "content": "T_{c}" + }, + { + "bbox": [ + 306, + 659, + 534, + 672 + ], + "type": "text", + "content": " are set to " + }, + { + "bbox": [ + 306, + 659, + 534, + 672 + ], + "type": "inline_equation", + "content": "\\{0.01, 0.02, 0.03, 0.04, 0.05\\}" + }, + { + "bbox": [ + 306, + 659, + 534, + 672 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 688, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 688, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 688, + 547, + 715 + ], + "type": "text", + "content": "QDS. Diversity score (DS) [19, 26] evaluates the diversity of the results as " + }, + { + "bbox": [ + 304, + 688, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{DS} = \\frac{1}{N^2}\\sum_{i,j=1}^{N}(d_c(\\mathbf{P}_i^*,\\mathbf{P}_j^*))" + }, + { + "bbox": [ + 304, + 688, + 547, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 688, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_i^*" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20854" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 547, + 182 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 547, + 182 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 547, + 182 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 547, + 182 + ], + "type": "table", + "html": "
SCD(10-2) ↓PA(%) ↑CA(%) ↑QDS(10-5) ↑WQDS(10-5) ↑
MethodsChairTableLampChairTableLampChairTableLampChairTableLampChairTableLamp
B-Global [14, 24]1.461.120.7915.7015.3722.619.9033.8418.600.150.200.761.251.400.58
B-LSTM [29]1.311.250.7721.7728.6420.786.8022.5614.050.040.270.631.071.431.54
B-Complement [27]2.412.981.508.782.3212.679.1915.5726.560.090.062.811.281.752.08
DGL [33]0.910.500.9339.0049.5133.3323.8739.9641.701.693.051.841.352.971.73
Score [4]0.710.421.1144.5152.7834.3230.3240.5949.073.369.176.831.703.812.82
RGL [21]0.870.480.7249.0654.1637.5632.2642.1557.343.557.636.822.124.072.96
IET [34]0.540.351.0362.8061.6738.6848.4556.1862.624.159.096.982.744.563.29
Ours0.510.280.7064.1364.8338.8049.2858.4564.165.629.587.123.064.813.90
", + "image_path": "a471d791dd498fb8c933c60bd5b9591d41c32ec0d452fdb518eab7c7c35d69e6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 209, + 287, + 280 + ], + "blocks": [ + { + "bbox": [ + 103, + 188, + 488, + 199 + ], + "lines": [ + { + "bbox": [ + 103, + 188, + 488, + 199 + ], + "spans": [ + { + "bbox": [ + 103, + 188, + 488, + 199 + ], + "type": "text", + "content": "Table 1. Comparison between our approach and other methods under the Chamfer distance threshold 0.01." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 209, + 287, + 280 + ], + "lines": [ + { + "bbox": [ + 50, + 209, + 287, + 280 + ], + "spans": [ + { + "bbox": [ + 50, + 209, + 287, + 280 + ], + "type": "table", + "html": "
mPA ↑mCA ↑
MethodsChairTableLampChairTableLamp
DGL [33]45.8457.8648.3229.1743.8851.75
IET [34]74.9273.2052.8062.3768.4975.94
Ours76.7976.1953.3165.3270.2678.15
", + "image_path": "0eaa9a2a4847629c1c81385e63fdf2ae7ce6f903536ab25e8e371e24f19154f4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 285, + 287, + 308 + ], + "lines": [ + { + "bbox": [ + 47, + 285, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 47, + 285, + 287, + 308 + ], + "type": "text", + "content": "Table 2. Comparison between our approach and other methods under multiple Chamfer distance thresholds from 0.01 to 0.05." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 322, + 287, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 322, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 47, + 322, + 287, + 346 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 47, + 322, + 287, + 346 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_j^*" + }, + { + "bbox": [ + 47, + 322, + 287, + 346 + ], + "type": "text", + "content": " represent any two assembled shapes. Based on DS, we propose the quality-diversity score (QDS) as below," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 354, + 290, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 354, + 290, + 384 + ], + "spans": [ + { + "bbox": [ + 47, + 354, + 290, + 384 + ], + "type": "interline_equation", + "content": "\\mathrm {Q D S} = \\frac {1}{N ^ {2}} \\sum_ {i, j = 1} ^ {N} d _ {c} \\left(\\mathbf {P} _ {i} ^ {*}, \\mathbf {P} _ {j} ^ {*}\\right) \\mathbb {1} \\left(\\mathrm {C A} \\left(\\mathbf {P} _ {i} ^ {*} > \\tau_ {q}\\right)\\right) \\mathbb {1} \\left(\\mathrm {C A} \\left(\\mathbf {P} _ {j} ^ {*} > \\tau_ {q}\\right)\\right).", + "image_path": "e3a2f6bd7e8df3d4a5427cbe04ba593a6d428f9773c1f83eb58ca517204df5c0.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 392, + 287, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 392, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 47, + 392, + 287, + 453 + ], + "type": "text", + "content": "QDS imposes constraints to remove pairs that are of low assembly quality, thus assessing not only diversity but also the quality of generated shapes. Following [4], we adopt SCD as the distance metric. The value of " + }, + { + "bbox": [ + 47, + 392, + 287, + 453 + ], + "type": "inline_equation", + "content": "\\tau_{q}" + }, + { + "bbox": [ + 47, + 392, + 287, + 453 + ], + "type": "text", + "content": " is set to 0.5 in both QDS and WQDS." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 468, + 287, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 468, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 468, + 287, + 491 + ], + "type": "text", + "content": "WQDS. Based on QDS, we further propose the weighted quality-diversity score (WQDS) as below," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 499, + 270, + 546 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 499, + 270, + 546 + ], + "spans": [ + { + "bbox": [ + 62, + 499, + 270, + 546 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathrm {W Q D S} = \\frac {1}{N ^ {2}} \\sum_ {i, j = 1} ^ {N} d _ {c} \\left(\\mathbf {P} _ {i} ^ {*}, \\mathbf {P} _ {j} ^ {*}\\right) \\mathrm {C A} \\left(\\mathbf {P} _ {i}\\right) \\mathrm {C A} \\left(\\mathbf {P} _ {j}\\right) \\\\ \\mathbb {1} \\left(\\mathrm {C A} \\left(\\mathbf {P} _ {i} ^ {*} > \\tau_ {q}\\right)\\right) \\mathbb {1} \\left(\\mathrm {C A} \\left(\\mathbf {P} _ {j} ^ {*} > \\tau_ {q}\\right)\\right). \\\\ \\end{array}", + "image_path": "c0b1616d5d5191daa8b2393b0ee0f0ce440ef388a5ade6247326ee0a937a71e9.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 554, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 554, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 47, + 554, + 287, + 590 + ], + "type": "text", + "content": "WQDS weights the QDS of a pair by their connectivity accuracy, thus favoring assembled shapes that demonstrate a high-quality connection between each pair of parts." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 598, + 234, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 598, + 234, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 598, + 234, + 611 + ], + "type": "text", + "content": "4.3. Comparisons with State-of-the-Art" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "content": "We compare our method with the state-of-the-art methods on PartNet [18] dataset. Following previous works, we first present the quantitative results under a certain Chamfer distance threshold of 0.01 in Table 1. As we can see, our method consistently outperforms all competitors. We also show the visualization of assembly results in Figure 3. It is clear that our method is better than others in terms of both the coherence of the assembled structure and the connectivity between" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 212, + 545, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 212, + 545, + 236 + ], + "spans": [ + { + "bbox": [ + 304, + 212, + 545, + 236 + ], + "type": "text", + "content": "parts, e.g., the positioning of crossbeams between chair legs, and the connection between table legs and the tabletop." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 237, + 547, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 237, + 547, + 489 + ], + "spans": [ + { + "bbox": [ + 304, + 237, + 547, + 489 + ], + "type": "text", + "content": "To eliminate the sensitivity of performances with respect to threshold values and provide a more comprehensive evaluation, we compare the mPA and mCA of our method with two competitive baselines in Table 2. For the Chair category, our method demonstrates a significant improvement over IET [34], with a " + }, + { + "bbox": [ + 304, + 237, + 547, + 489 + ], + "type": "inline_equation", + "content": "1.87\\%" + }, + { + "bbox": [ + 304, + 237, + 547, + 489 + ], + "type": "text", + "content": " increase in mPA and a " + }, + { + "bbox": [ + 304, + 237, + 547, + 489 + ], + "type": "inline_equation", + "content": "2.95\\%" + }, + { + "bbox": [ + 304, + 237, + 547, + 489 + ], + "type": "text", + "content": " increase in mCA, while the improvement in the Lamp category is relatively minor. This may be attributed to the limited variation in the geometry of lamp parts, making it difficult to clearly differentiate between various levels of the part-whole hierarchy. In addition, Figure 5 illustrates the performance of various methods on the Chair and Table categories using five different Chamfer distance thresholds ranging from 0.01 to 0.05. As the Chamfer distance threshold increases, the advantage of our method becomes more pronounced, which shows the superiority of our approach. We also evaluate the diversity of generative part assembly. As shown in Table 1, our method outperforms all competing methods in both QDS and WQDS. In Figure 4, we use the model to generate multiple assembly shapes for the same set of parts (with different noises) to show the variation of generated shapes." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 498, + 394, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 498, + 394, + 510 + ], + "spans": [ + { + "bbox": [ + 306, + 498, + 394, + 510 + ], + "type": "text", + "content": "4.4. Human Study" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 517, + 547, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 517, + 547, + 661 + ], + "spans": [ + { + "bbox": [ + 304, + 517, + 547, + 661 + ], + "type": "text", + "content": "Since many inaccurate results are still valid visually, we also conduct a human study to evaluate the quality of generated assemblies. Specifically, we perform an A/B test with a group of students, utilizing the identical evaluation system as in [30, 35]. Participants were presented with pairs of randomly chosen assemblies of the same object from two different methods. In total, 28 participants labeled 93 assembly pairs. We compare the proposed method with three main baselines and present the outcomes in Table 5. In over " + }, + { + "bbox": [ + 304, + 517, + 547, + 661 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 304, + 517, + 547, + 661 + ], + "type": "text", + "content": " of the cases, our method is favored over other three competing methods, unequivocally demonstrating the superior visual quality of our method's output." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 670, + 400, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 400, + 683 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 400, + 683 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 547, + 714 + ], + "type": "text", + "content": "We now investigate the importance of different loss components and verify the effectiveness of the proposed super-part" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20855" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 89, + 82, + 126, + 130 + ], + "blocks": [ + { + "bbox": [ + 56, + 92, + 69, + 114 + ], + "lines": [ + { + "bbox": [ + 56, + 92, + 69, + 114 + ], + "spans": [ + { + "bbox": [ + 56, + 92, + 69, + 114 + ], + "type": "text", + "content": "DGL" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 89, + 82, + 126, + 130 + ], + "lines": [ + { + "bbox": [ + 89, + 82, + 126, + 130 + ], + "spans": [ + { + "bbox": [ + 89, + 82, + 126, + 130 + ], + "type": "image", + "image_path": "c0a31a3bd067e53f1fbc39286632ef9f02b7469d3a791bfb754c93c45a61b6f5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 143, + 82, + 179, + 131 + ], + "blocks": [ + { + "bbox": [ + 143, + 82, + 179, + 131 + ], + "lines": [ + { + "bbox": [ + 143, + 82, + 179, + 131 + ], + "spans": [ + { + "bbox": [ + 143, + 82, + 179, + 131 + ], + "type": "image", + "image_path": "98a9b109683bc0a7124b90dc96a7b1953e4e7618d8eb21019c6b1041876b3835.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 187, + 82, + 237, + 132 + ], + "blocks": [ + { + "bbox": [ + 187, + 82, + 237, + 132 + ], + "lines": [ + { + "bbox": [ + 187, + 82, + 237, + 132 + ], + "spans": [ + { + "bbox": [ + 187, + 82, + 237, + 132 + ], + "type": "image", + "image_path": "d89e3dc458b7d1518e4047bb7a1c992606a7b462d2501e4f27a8e85951bf81f1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 250, + 92, + 296, + 130 + ], + "blocks": [ + { + "bbox": [ + 250, + 92, + 296, + 130 + ], + "lines": [ + { + "bbox": [ + 250, + 92, + 296, + 130 + ], + "spans": [ + { + "bbox": [ + 250, + 92, + 296, + 130 + ], + "type": "image", + "image_path": "ee61dbac09230377db4af431ea67f53e540a5dd642a18814ce7bc67efac6b606.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 95, + 356, + 129 + ], + "blocks": [ + { + "bbox": [ + 307, + 95, + 356, + 129 + ], + "lines": [ + { + "bbox": [ + 307, + 95, + 356, + 129 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 356, + 129 + ], + "type": "image", + "image_path": "b4bb8d5fb29c3d0a22a4df3c7c04ebeb1d1187dc437cc5587571b24848bc062a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 366, + 95, + 411, + 131 + ], + "blocks": [ + { + "bbox": [ + 366, + 95, + 411, + 131 + ], + "lines": [ + { + "bbox": [ + 366, + 95, + 411, + 131 + ], + "spans": [ + { + "bbox": [ + 366, + 95, + 411, + 131 + ], + "type": "image", + "image_path": "1e78afbfb0f88373709f4c9316735357b0af95148c7ef12d7627722d8b5438e3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 425, + 78, + 464, + 126 + ], + "blocks": [ + { + "bbox": [ + 425, + 78, + 464, + 126 + ], + "lines": [ + { + "bbox": [ + 425, + 78, + 464, + 126 + ], + "spans": [ + { + "bbox": [ + 425, + 78, + 464, + 126 + ], + "type": "image", + "image_path": "43c670993759bf50271e36f064bc0c6fb7a11b729f16109fec0cd5e030704dd8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 479, + 78, + 510, + 128 + ], + "blocks": [ + { + "bbox": [ + 479, + 78, + 510, + 128 + ], + "lines": [ + { + "bbox": [ + 479, + 78, + 510, + 128 + ], + "spans": [ + { + "bbox": [ + 479, + 78, + 510, + 128 + ], + "type": "image", + "image_path": "99c5d6f186935f4fefd421ada073cef96c40885454d9ce31e8b1eb5810569f88.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 90, + 143, + 125, + 187 + ], + "blocks": [ + { + "bbox": [ + 58, + 152, + 69, + 168 + ], + "lines": [ + { + "bbox": [ + 58, + 152, + 69, + 168 + ], + "spans": [ + { + "bbox": [ + 58, + 152, + 69, + 168 + ], + "type": "text", + "content": "IET" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 90, + 143, + 125, + 187 + ], + "lines": [ + { + "bbox": [ + 90, + 143, + 125, + 187 + ], + "spans": [ + { + "bbox": [ + 90, + 143, + 125, + 187 + ], + "type": "image", + "image_path": "6ba294a8c5df4bd1e23c6932d9ac5288d6b509d25a8a5b2d3ba0729bc26dd7a7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 143, + 143, + 180, + 187 + ], + "blocks": [ + { + "bbox": [ + 143, + 143, + 180, + 187 + ], + "lines": [ + { + "bbox": [ + 143, + 143, + 180, + 187 + ], + "spans": [ + { + "bbox": [ + 143, + 143, + 180, + 187 + ], + "type": "image", + "image_path": "0bae73de0041777f7ee1b969aa33437d47c35caa9422819e2e638cc74d8139f5.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 188, + 143, + 235, + 188 + ], + "blocks": [ + { + "bbox": [ + 188, + 143, + 235, + 188 + ], + "lines": [ + { + "bbox": [ + 188, + 143, + 235, + 188 + ], + "spans": [ + { + "bbox": [ + 188, + 143, + 235, + 188 + ], + "type": "image", + "image_path": "ac472874acc8ebac9d6f81b5eb5f7a66221b5092c321d590518b453da52efa1e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 251, + 154, + 296, + 183 + ], + "blocks": [ + { + "bbox": [ + 251, + 154, + 296, + 183 + ], + "lines": [ + { + "bbox": [ + 251, + 154, + 296, + 183 + ], + "spans": [ + { + "bbox": [ + 251, + 154, + 296, + 183 + ], + "type": "image", + "image_path": "12158f99e369ffde7104a65f343f78a41484cfd787087238aebb4df934afc0df.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 309, + 158, + 356, + 188 + ], + "blocks": [ + { + "bbox": [ + 309, + 158, + 356, + 188 + ], + "lines": [ + { + "bbox": [ + 309, + 158, + 356, + 188 + ], + "spans": [ + { + "bbox": [ + 309, + 158, + 356, + 188 + ], + "type": "image", + "image_path": "88ca63764e1b942c822bdcdbbc9c52d2f9384ebfdb7e7eb5fb1de07c44611f82.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 367, + 159, + 411, + 189 + ], + "blocks": [ + { + "bbox": [ + 367, + 159, + 411, + 189 + ], + "lines": [ + { + "bbox": [ + 367, + 159, + 411, + 189 + ], + "spans": [ + { + "bbox": [ + 367, + 159, + 411, + 189 + ], + "type": "image", + "image_path": "a2bdc168b759da5aad9855d257709cfaa3260cf877fe5ed8d811ad753b956bf0.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 429, + 138, + 462, + 190 + ], + "blocks": [ + { + "bbox": [ + 429, + 138, + 462, + 190 + ], + "lines": [ + { + "bbox": [ + 429, + 138, + 462, + 190 + ], + "spans": [ + { + "bbox": [ + 429, + 138, + 462, + 190 + ], + "type": "image", + "image_path": "1faffe13274ae5db24a8706a0c1e3d813cf37d12fe3be82f85a62ffd95cf4a5e.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 480, + 140, + 511, + 194 + ], + "blocks": [ + { + "bbox": [ + 480, + 140, + 511, + 194 + ], + "lines": [ + { + "bbox": [ + 480, + 140, + 511, + 194 + ], + "spans": [ + { + "bbox": [ + 480, + 140, + 511, + 194 + ], + "type": "image", + "image_path": "ea6131c54406dfdead26728adca6c4e08c25611ec10d19773a03cc21ce9e6d0d.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 88, + 205, + 124, + 251 + ], + "blocks": [ + { + "bbox": [ + 58, + 211, + 68, + 230 + ], + "lines": [ + { + "bbox": [ + 58, + 211, + 68, + 230 + ], + "spans": [ + { + "bbox": [ + 58, + 211, + 68, + 230 + ], + "type": "text", + "content": "sno" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 88, + 205, + 124, + 251 + ], + "lines": [ + { + "bbox": [ + 88, + 205, + 124, + 251 + ], + "spans": [ + { + "bbox": [ + 88, + 205, + 124, + 251 + ], + "type": "image", + "image_path": "1739980a2834bb31046dd9a000c112c3e99fdec541922734b2983357db39d502.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 141, + 205, + 178, + 251 + ], + "blocks": [ + { + "bbox": [ + 141, + 205, + 178, + 251 + ], + "lines": [ + { + "bbox": [ + 141, + 205, + 178, + 251 + ], + "spans": [ + { + "bbox": [ + 141, + 205, + 178, + 251 + ], + "type": "image", + "image_path": "722b6727f01c40a86097a5fc6b3d21fee01b34874ceffcde11cada5439b65fe9.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 189, + 206, + 235, + 253 + ], + "blocks": [ + { + "bbox": [ + 189, + 206, + 235, + 253 + ], + "lines": [ + { + "bbox": [ + 189, + 206, + 235, + 253 + ], + "spans": [ + { + "bbox": [ + 189, + 206, + 235, + 253 + ], + "type": "image", + "image_path": "8efa1ddbf309bc904bdcc8e40bc31c97df75483289059040f6a284e10bf8940c.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 252, + 218, + 298, + 247 + ], + "blocks": [ + { + "bbox": [ + 252, + 218, + 298, + 247 + ], + "lines": [ + { + "bbox": [ + 252, + 218, + 298, + 247 + ], + "spans": [ + { + "bbox": [ + 252, + 218, + 298, + 247 + ], + "type": "image", + "image_path": "c1d2e9c85bc22f8d3d2b6ced980ebdcd46b7491489af02a1a7f647ed958b2891.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 310, + 220, + 356, + 252 + ], + "blocks": [ + { + "bbox": [ + 310, + 220, + 356, + 252 + ], + "lines": [ + { + "bbox": [ + 310, + 220, + 356, + 252 + ], + "spans": [ + { + "bbox": [ + 310, + 220, + 356, + 252 + ], + "type": "image", + "image_path": "627ec2a8d37f67b5cb84f78e360ede24e849e8641e9099b9583bc41569aebc84.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 369, + 221, + 412, + 251 + ], + "blocks": [ + { + "bbox": [ + 369, + 221, + 412, + 251 + ], + "lines": [ + { + "bbox": [ + 369, + 221, + 412, + 251 + ], + "spans": [ + { + "bbox": [ + 369, + 221, + 412, + 251 + ], + "type": "image", + "image_path": "ba9517e33e3a63637b556c00f36563c0a63ea4aa1799912a2b5751db0b7353d8.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 434, + 201, + 459, + 255 + ], + "blocks": [ + { + "bbox": [ + 434, + 201, + 459, + 255 + ], + "lines": [ + { + "bbox": [ + 434, + 201, + 459, + 255 + ], + "spans": [ + { + "bbox": [ + 434, + 201, + 459, + 255 + ], + "type": "image", + "image_path": "1a4c047017a2ea569677c75349811c3726394ece7483574c26e3958eed412b1a.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 481, + 205, + 511, + 256 + ], + "blocks": [ + { + "bbox": [ + 481, + 205, + 511, + 256 + ], + "lines": [ + { + "bbox": [ + 481, + 205, + 511, + 256 + ], + "spans": [ + { + "bbox": [ + 481, + 205, + 511, + 256 + ], + "type": "image", + "image_path": "0c346b59483888665311185c00d91430dedb480506502cdf60dc0932e80e532b.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 88, + 269, + 124, + 313 + ], + "blocks": [ + { + "bbox": [ + 59, + 258, + 69, + 307 + ], + "lines": [ + { + "bbox": [ + 59, + 258, + 69, + 307 + ], + "spans": [ + { + "bbox": [ + 59, + 258, + 69, + 307 + ], + "type": "text", + "content": "Grouu" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 88, + 269, + 124, + 313 + ], + "lines": [ + { + "bbox": [ + 88, + 269, + 124, + 313 + ], + "spans": [ + { + "bbox": [ + 88, + 269, + 124, + 313 + ], + "type": "image", + "image_path": "19c2f8696e111d1b4c9f6d63fe378b3cdd808708acb8183f2366ba8285e91370.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 140, + 270, + 176, + 313 + ], + "blocks": [ + { + "bbox": [ + 140, + 270, + 176, + 313 + ], + "lines": [ + { + "bbox": [ + 140, + 270, + 176, + 313 + ], + "spans": [ + { + "bbox": [ + 140, + 270, + 176, + 313 + ], + "type": "image", + "image_path": "afdec642351463f0d6074e60d16ef556a3620a8797ceea4ff57985a7769a2e63.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 143, + 323, + 165, + 332 + ], + "lines": [ + { + "bbox": [ + 143, + 323, + 165, + 332 + ], + "spans": [ + { + "bbox": [ + 143, + 323, + 165, + 332 + ], + "type": "text", + "content": "Chair" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 189, + 270, + 235, + 314 + ], + "blocks": [ + { + "bbox": [ + 189, + 270, + 235, + 314 + ], + "lines": [ + { + "bbox": [ + 189, + 270, + 235, + 314 + ], + "spans": [ + { + "bbox": [ + 189, + 270, + 235, + 314 + ], + "type": "image", + "image_path": "7ab5fdf4dabc40bb14e1367c127c44aefe8ca066a1c5ec7da268f6301fee23f0.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 338, + 545, + 361 + ], + "lines": [ + { + "bbox": [ + 46, + 338, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 545, + 361 + ], + "type": "text", + "content": "Figure 3. The qualitative comparisons between our method and two most competitive baselines on PartNet [18]. We highlight some areas where the assembly quality of ours is clearly better." + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 252, + 277, + 298, + 308 + ], + "blocks": [ + { + "bbox": [ + 252, + 277, + 298, + 308 + ], + "lines": [ + { + "bbox": [ + 252, + 277, + 298, + 308 + ], + "spans": [ + { + "bbox": [ + 252, + 277, + 298, + 308 + ], + "type": "image", + "image_path": "d48512d7f8df10e313a604ea140bdf2f0eac9426420b6977c466e2034911d677.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 323, + 344, + 332 + ], + "lines": [ + { + "bbox": [ + 323, + 323, + 344, + 332 + ], + "spans": [ + { + "bbox": [ + 323, + 323, + 344, + 332 + ], + "type": "text", + "content": "Table" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 309, + 280, + 356, + 310 + ], + "blocks": [ + { + "bbox": [ + 309, + 280, + 356, + 310 + ], + "lines": [ + { + "bbox": [ + 309, + 280, + 356, + 310 + ], + "spans": [ + { + "bbox": [ + 309, + 280, + 356, + 310 + ], + "type": "image", + "image_path": "84447528fee5d37a3c41f3132709e45db88f8485e432ca3c63167f289205896d.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 368, + 281, + 413, + 310 + ], + "blocks": [ + { + "bbox": [ + 368, + 281, + 413, + 310 + ], + "lines": [ + { + "bbox": [ + 368, + 281, + 413, + 310 + ], + "spans": [ + { + "bbox": [ + 368, + 281, + 413, + 310 + ], + "type": "image", + "image_path": "414a1b04931c331d180899687f57ff6984f4145f9322fa2e2ac1654bd96a4813.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 436, + 269, + 459, + 319 + ], + "blocks": [ + { + "bbox": [ + 436, + 269, + 459, + 319 + ], + "lines": [ + { + "bbox": [ + 436, + 269, + 459, + 319 + ], + "spans": [ + { + "bbox": [ + 436, + 269, + 459, + 319 + ], + "type": "image", + "image_path": "1b23f45b2c1a881b9885b0dd01a05e30cd7687239a975d97b72dfa2a8a2825fa.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 482, + 270, + 511, + 319 + ], + "blocks": [ + { + "bbox": [ + 482, + 270, + 511, + 319 + ], + "lines": [ + { + "bbox": [ + 482, + 270, + 511, + 319 + ], + "spans": [ + { + "bbox": [ + 482, + 270, + 511, + 319 + ], + "type": "image", + "image_path": "06eadb62e6faf4c9ca9148cd892598d47e5468a612596a1c4b7916748670f56b.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 462, + 323, + 485, + 334 + ], + "lines": [ + { + "bbox": [ + 462, + 323, + 485, + 334 + ], + "spans": [ + { + "bbox": [ + 462, + 323, + 485, + 334 + ], + "type": "text", + "content": "Lamp" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 80, + 381, + 120, + 423 + ], + "blocks": [ + { + "bbox": [ + 58, + 387, + 68, + 422 + ], + "lines": [ + { + "bbox": [ + 58, + 387, + 68, + 422 + ], + "spans": [ + { + "bbox": [ + 58, + 387, + 68, + 422 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 80, + 381, + 120, + 423 + ], + "lines": [ + { + "bbox": [ + 80, + 381, + 120, + 423 + ], + "spans": [ + { + "bbox": [ + 80, + 381, + 120, + 423 + ], + "type": "image", + "image_path": "aa9bbca0fc68aa66d86d85c7d1765799b428039c0ee79ddc4e4406f8e7dc8995.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 136, + 382, + 166, + 425 + ], + "blocks": [ + { + "bbox": [ + 136, + 382, + 166, + 425 + ], + "lines": [ + { + "bbox": [ + 136, + 382, + 166, + 425 + ], + "spans": [ + { + "bbox": [ + 136, + 382, + 166, + 425 + ], + "type": "image", + "image_path": "f7cc16ffc59690cdd255d5ddf4b95504c0e2fe0b8e691c9da25a725ca0700a0b.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 186, + 381, + 221, + 427 + ], + "blocks": [ + { + "bbox": [ + 186, + 381, + 221, + 427 + ], + "lines": [ + { + "bbox": [ + 186, + 381, + 221, + 427 + ], + "spans": [ + { + "bbox": [ + 186, + 381, + 221, + 427 + ], + "type": "image", + "image_path": "662907be14770d8167de5685bfb25a14406bf262e75f5e273ba62617cca74bd2.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 236, + 381, + 276, + 429 + ], + "blocks": [ + { + "bbox": [ + 236, + 381, + 276, + 429 + ], + "lines": [ + { + "bbox": [ + 236, + 381, + 276, + 429 + ], + "spans": [ + { + "bbox": [ + 236, + 381, + 276, + 429 + ], + "type": "image", + "image_path": "aff3a1d3db61d179ab0a8fa642e394fa5eb32603a8fabc4478e1db3ad6d09a49.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 80, + 445, + 118, + 485 + ], + "blocks": [ + { + "bbox": [ + 58, + 449, + 68, + 483 + ], + "lines": [ + { + "bbox": [ + 58, + 449, + 68, + 483 + ], + "spans": [ + { + "bbox": [ + 58, + 449, + 68, + 483 + ], + "type": "text", + "content": "Aeeaae 2" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 80, + 445, + 118, + 485 + ], + "lines": [ + { + "bbox": [ + 80, + 445, + 118, + 485 + ], + "spans": [ + { + "bbox": [ + 80, + 445, + 118, + 485 + ], + "type": "image", + "image_path": "f559b0ae3511b22f652dde3b59d39d6963db5984fa57819b822122a3ef46c459.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 135, + 439, + 165, + 486 + ], + "blocks": [ + { + "bbox": [ + 135, + 439, + 165, + 486 + ], + "lines": [ + { + "bbox": [ + 135, + 439, + 165, + 486 + ], + "spans": [ + { + "bbox": [ + 135, + 439, + 165, + 486 + ], + "type": "image", + "image_path": "ba21a9cbc7ce0d1e2fb76659e2a183d4dec3717d7510c7ee540763dc0f3472b9.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + } + ], + "index": 47 + }, + { + "type": "image", + "bbox": [ + 186, + 438, + 220, + 486 + ], + "blocks": [ + { + "bbox": [ + 186, + 438, + 220, + 486 + ], + "lines": [ + { + "bbox": [ + 186, + 438, + 220, + 486 + ], + "spans": [ + { + "bbox": [ + 186, + 438, + 220, + 486 + ], + "type": "image", + "image_path": "1cf5e07ab5cad61a3ba2b3accd347e4fec3326059627a5a83c9b20fc34ca969d.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 235, + 443, + 275, + 488 + ], + "blocks": [ + { + "bbox": [ + 235, + 443, + 275, + 488 + ], + "lines": [ + { + "bbox": [ + 235, + 443, + 275, + 488 + ], + "spans": [ + { + "bbox": [ + 235, + 443, + 275, + 488 + ], + "type": "image", + "image_path": "eac492ac95e1b1e54327be94a12117a5c47fd84d97deb925d390c91effb54ff6.jpg" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_body" + } + ], + "index": 49 + }, + { + "type": "image", + "bbox": [ + 80, + 497, + 116, + 543 + ], + "blocks": [ + { + "bbox": [ + 59, + 505, + 68, + 538 + ], + "lines": [ + { + "bbox": [ + 59, + 505, + 68, + 538 + ], + "spans": [ + { + "bbox": [ + 59, + 505, + 68, + 538 + ], + "type": "text", + "content": "aennnnnne aannnnnne" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 80, + 497, + 116, + 543 + ], + "lines": [ + { + "bbox": [ + 80, + 497, + 116, + 543 + ], + "spans": [ + { + "bbox": [ + 80, + 497, + 116, + 543 + ], + "type": "image", + "image_path": "117bf18a41ad36240ccb0f8f80d9bb932bc2100c98b193af97c6ca6f8e698234.jpg" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 554, + 287, + 598 + ], + "lines": [ + { + "bbox": [ + 46, + 554, + 287, + 598 + ], + "spans": [ + { + "bbox": [ + 46, + 554, + 287, + 598 + ], + "type": "text", + "content": "Figure 4. Diverse results on the unseen PartNet[18] test dataset generated by our network to demonstrate the structural variation in part assembly, providing different artistic results while maintaining reasonable object structures." + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_caption" + } + ], + "index": 51 + }, + { + "type": "image", + "bbox": [ + 135, + 497, + 165, + 544 + ], + "blocks": [ + { + "bbox": [ + 135, + 497, + 165, + 544 + ], + "lines": [ + { + "bbox": [ + 135, + 497, + 165, + 544 + ], + "spans": [ + { + "bbox": [ + 135, + 497, + 165, + 544 + ], + "type": "image", + "image_path": "1f0565af5277f064a2306c3c8634fcd09b5c3cbba8c16d96720580968b641785.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + } + ], + "index": 52 + }, + { + "type": "image", + "bbox": [ + 182, + 496, + 218, + 545 + ], + "blocks": [ + { + "bbox": [ + 182, + 496, + 218, + 545 + ], + "lines": [ + { + "bbox": [ + 182, + 496, + 218, + 545 + ], + "spans": [ + { + "bbox": [ + 182, + 496, + 218, + 545 + ], + "type": "image", + "image_path": "d1bfd76f71ff301a28bc796cf3878864398cdbc8fed06c5188f8f23f0bfeb219.jpg" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_body" + } + ], + "index": 53 + }, + { + "type": "image", + "bbox": [ + 235, + 498, + 274, + 546 + ], + "blocks": [ + { + "bbox": [ + 235, + 498, + 274, + 546 + ], + "lines": [ + { + "bbox": [ + 235, + 498, + 274, + 546 + ], + "spans": [ + { + "bbox": [ + 235, + 498, + 274, + 546 + ], + "type": "image", + "image_path": "e337b53450f184be699e239f6a107408016ec5d77a57edfb80bab5cc90d7e843.jpg" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_body" + } + ], + "index": 54 + }, + { + "bbox": [ + 47, + 608, + 172, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 608, + 172, + 620 + ], + "spans": [ + { + "bbox": [ + 47, + 608, + 172, + 620 + ], + "type": "text", + "content": "encoder on the Chair category." + } + ] + } + ], + "index": 56 + }, + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "content": "Super-Part Encoder. Since the super-part encoder is at the core of our part-whole hierarchy message passing network, we conduct an experiment where we remove this module (denoted as w/o Super-Part Enc.). To mitigate the impact of the number of model parameters, we also create another baseline by increasing the number of parameters of the w/o" + } + ] + } + ], + "index": 57 + }, + { + "type": "table", + "bbox": [ + 329, + 365, + 519, + 437 + ], + "blocks": [ + { + "bbox": [ + 329, + 365, + 519, + 437 + ], + "lines": [ + { + "bbox": [ + 329, + 365, + 519, + 437 + ], + "spans": [ + { + "bbox": [ + 329, + 365, + 519, + 437 + ], + "type": "table", + "html": "
SettingsSCD↓PA↑CA↑
w/o Super-Part Enc. + Augmented Param.0.003860.7053.46
w/o Super-Part Enc.0.003661.0455.39
Full Setting0.002864.8358.45
", + "image_path": "f321efad3c61d3896280d1e60b50b10d36e1dbd0300bd12ac83f8d1b297651e0.jpg" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "table_body" + } + ], + "index": 58 + }, + { + "type": "table", + "bbox": [ + 329, + 479, + 518, + 566 + ], + "blocks": [ + { + "bbox": [ + 306, + 442, + 545, + 465 + ], + "lines": [ + { + "bbox": [ + 306, + 442, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 306, + 442, + 545, + 465 + ], + "type": "text", + "content": "Table 3. Ablation studies on part-whole hierarchy on the Table category under a certain Chamfer distance threshold 0.01." + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 329, + 479, + 518, + 566 + ], + "lines": [ + { + "bbox": [ + 329, + 479, + 518, + 566 + ], + "spans": [ + { + "bbox": [ + 329, + 479, + 518, + 566 + ], + "type": "table", + "html": "
LtLrLsSCD↓PA↑CA↑
✓✓0.007535.3632.73
✓✓0.003960.0254.21
✓✓0.003362.7456.98
✓✓✓0.002864.8358.45
", + "image_path": "a23e510b00562bf9837875b7c1f13fc2ca27fc21791b1be5b61d70b93c711846.jpg" + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "table_body" + } + ], + "index": 60 + }, + { + "bbox": [ + 306, + 571, + 545, + 594 + ], + "lines": [ + { + "bbox": [ + 306, + 571, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 306, + 571, + 545, + 594 + ], + "type": "text", + "content": "Table 4. Ablation studies on loss components on the Table category under a certain Chamfer distance threshold 0.01." + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "Super-Part Enc. setting (denoted as w/o Super-Part Enc. + Augmented Param.). This involves increasing the number of parameters in the part encoder and the multi-head attention module, ensuring that the overall number of parameters is on the same order of magnitude as the full model setting. As shown in Table 3, the super-part encoder contributes significantly to the final performance. We also observe that augmenting the model parameters without incorporating the" + } + ] + } + ], + "index": 62 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20856" + } + ] + } + ], + "index": 63 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 81, + 212, + 202 + ], + "blocks": [ + { + "bbox": [ + 50, + 81, + 212, + 202 + ], + "lines": [ + { + "bbox": [ + 50, + 81, + 212, + 202 + ], + "spans": [ + { + "bbox": [ + 50, + 81, + 212, + 202 + ], + "type": "image", + "image_path": "d0a5e2ee66ca2cc3e2903d93013d27c2a33ab727a5944cd5484f427dd9a12cef.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 96, + 213, + 496, + 225 + ], + "lines": [ + { + "bbox": [ + 96, + 213, + 496, + 225 + ], + "spans": [ + { + "bbox": [ + 96, + 213, + 496, + 225 + ], + "type": "text", + "content": "Figure 5. Performance on the Chair, Table and Lamp categories under multiple Chamfer distance thresholds." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 212, + 81, + 375, + 204 + ], + "blocks": [ + { + "bbox": [ + 212, + 81, + 375, + 204 + ], + "lines": [ + { + "bbox": [ + 212, + 81, + 375, + 204 + ], + "spans": [ + { + "bbox": [ + 212, + 81, + 375, + 204 + ], + "type": "image", + "image_path": "7f20821c27e88604feb16407c38088204dfa8b62ccc5aa54bcd3a73bf682d34e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 376, + 81, + 542, + 204 + ], + "blocks": [ + { + "bbox": [ + 376, + 81, + 542, + 204 + ], + "lines": [ + { + "bbox": [ + 376, + 81, + 542, + 204 + ], + "spans": [ + { + "bbox": [ + 376, + 81, + 542, + 204 + ], + "type": "image", + "image_path": "d619800e14b090589a6a33359b79023d0d6a7449a4c38def3d739be7fa04f2d2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 70, + 234, + 260, + 306 + ], + "blocks": [ + { + "bbox": [ + 70, + 234, + 260, + 306 + ], + "lines": [ + { + "bbox": [ + 70, + 234, + 260, + 306 + ], + "spans": [ + { + "bbox": [ + 70, + 234, + 260, + 306 + ], + "type": "table", + "html": "
MethodPercent Prefer Ours
Ours vs IET[34]95.2%
Ours vs RGL [21]97.6%
Ours vs DGL [33]99.5%
", + "image_path": "803e4315084474964de1df8d5f820dea9199d7e64821a25b7618e1b340742a96.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 83, + 311, + 250, + 322 + ], + "lines": [ + { + "bbox": [ + 83, + 311, + 250, + 322 + ], + "spans": [ + { + "bbox": [ + 83, + 311, + 250, + 322 + ], + "type": "text", + "content": "Table 5. Human study results on PartNet[18]." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 50, + 333, + 282, + 456 + ], + "blocks": [ + { + "bbox": [ + 50, + 333, + 282, + 456 + ], + "lines": [ + { + "bbox": [ + 50, + 333, + 282, + 456 + ], + "spans": [ + { + "bbox": [ + 50, + 333, + 282, + 456 + ], + "type": "image", + "image_path": "78d2d7c499f304c2361218850408503913e4789c3fb60a001317527d4be79150.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 464, + 288, + 487 + ], + "lines": [ + { + "bbox": [ + 46, + 464, + 288, + 487 + ], + "spans": [ + { + "bbox": [ + 46, + 464, + 288, + 487 + ], + "type": "text", + "content": "Figure 6. Visualization of our predicted hierarchical assembly process from parts to the whole." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 494, + 288, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 494, + 288, + 531 + ], + "spans": [ + { + "bbox": [ + 46, + 494, + 288, + 531 + ], + "type": "text", + "content": "super-part encoders actually leads to a performance drop. This finding further highlights the effectiveness of the proposed part-whole hierarchy network for part assembly." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 544, + 288, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 288, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 288, + 664 + ], + "type": "text", + "content": "Loss Functions. We now evaluate the importance of individual loss terms in Equation (6). As shown in Table 4, removing the part translation loss " + }, + { + "bbox": [ + 46, + 544, + 288, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_t" + }, + { + "bbox": [ + 46, + 544, + 288, + 664 + ], + "type": "text", + "content": " leads to a significant performance degradation. Similarly, the remaining loss terms, including the rigid rotation loss " + }, + { + "bbox": [ + 46, + 544, + 288, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_r" + }, + { + "bbox": [ + 46, + 544, + 288, + 664 + ], + "type": "text", + "content": " and the shape assembly loss " + }, + { + "bbox": [ + 46, + 544, + 288, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 46, + 544, + 288, + 664 + ], + "type": "text", + "content": ", demonstrate their significance in facilitating precise part rotations and ensuring the overall coherence of the assembled parts, respectively. This shows that each loss term does play an indispensable role in achieving high-quality assemblies." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 671, + 243, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 243, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 243, + 685 + ], + "type": "text", + "content": "4.6. Hierarchical Part Assembly Analysis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "content": "We further provide a visual analysis of the assembly process for super-parts and parts in Figure 6. It is obvious that during" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 237, + 547, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 237, + 547, + 321 + ], + "spans": [ + { + "bbox": [ + 304, + 237, + 547, + 321 + ], + "type": "text", + "content": "the super-part assembly phase, important components such as chair seats and chair backs are often assembled correctly first. This aligns with our intuition that super-part assembly is relative easier and provides strong hints for predicting poses of parts like chair legs and arms, thus showing further evidence for the effectiveness of the part-whole-hierarchy message passing network." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 331, + 379, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 331, + 379, + 344 + ], + "spans": [ + { + "bbox": [ + 306, + 331, + 379, + 344 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 351, + 547, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 351, + 547, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 351, + 547, + 544 + ], + "type": "text", + "content": "In this paper, we propose the part-whole-hierarchy message passing network to address the challenging generative 3D part assembly task. We first group the point cloud of individual parts to form super-parts in an unsupervised way. Taking the point cloud as input, our super-part encoder predicts latent poses of super-parts which are used to transform the point cloud. We then feed the transformed point cloud to the part encoder. Relying on the cross-level and within-level attention based message passing, the part encoder takes the transformed point cloud as input and leverages the information from super-parts and predicts poses of parts. Experiments on the PartNet dataset demonstrate that our method achieves state-of-the-art performances as well as provides interpretable assembly process. In the future, we are interested in combining 3D part generation with our 3D part assembly model to generate 3D shapes from scratch." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 558, + 388, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 558, + 388, + 569 + ], + "spans": [ + { + "bbox": [ + 306, + 558, + 388, + 569 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 575, + 548, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 575, + 548, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 575, + 548, + 696 + ], + "type": "text", + "content": "This work was funded, in part, by NSERC DG Grants (No. RGPIN-2022-04636), the Vector Institute for AI, and Canada CIFAR AI Chair. Resources used in preparing this research were provided, in part, by the Province of Ontario, the Government of Canada through the Digital Research Alliance of Canada alliance.can.ca, and companies sponsoring the Vector Institute www.vectorinstitute.ai/#partners, and Advanced Research Computing at the University of British Columbia. Additional hardware support was provided by John R. Evans Leaders Fund CFI grant." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20857" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In International conference on machine learning, pages 40-49. PMLR, 2018. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 288, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 288, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 288, + 179 + ], + "type": "text", + "content": "[2] Siddhartha Chaudhuri, Evangelos Kalogerakis, Leonidas Guibas, and Vladlen Koltun. Probabilistic reasoning for assembly-based 3d modeling. In ACM SIGGRAPH 2011 papers, pages 1-10. 2011. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 288, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 288, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 288, + 235 + ], + "type": "text", + "content": "[3] Yun-Chun Chen, Haoda Li, Dylan Turpin, Alec Jacobson, and Animesh Garg. Neural shape mating: Self-supervised object assembly with adversarial shape priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12724-12733, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 288, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 288, + 268 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 288, + 268 + ], + "type": "text", + "content": "[4] Junfeng Cheng, Mingdong Wu, Ruiyuan Zhang, Guanqi Zhan, Chao Wu, and Hao Dong. Score-pa: Score-based 3d part assembly. arXiv preprint arXiv:2309.04220, 2023. 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 269, + 288, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 269, + 288, + 302 + ], + "spans": [ + { + "bbox": [ + 53, + 269, + 288, + 302 + ], + "type": "text", + "content": "[5] P Kingma Diederik, Max Welling, et al. Auto-encoding variational bayes. In Proceedings of the International Conference on Learning Representations (ICLR), 2014. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 303, + 288, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 303, + 288, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 303, + 288, + 346 + ], + "type": "text", + "content": "[6] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 605-613, 2017. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 347, + 288, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 347, + 288, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 347, + 288, + 401 + ], + "type": "text", + "content": "[7] Thomas Funkhouser, Michael Kazhdan, Philip Shilane, Patrick Min, William Kiefer, Ayellet Tal, Szymon Rusinkiewicz, and David Dobkin. Modeling by example. ACM transactions on graphics (TOG), 23(3):652-663, 2004. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 402, + 288, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 402, + 288, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 402, + 288, + 446 + ], + "type": "text", + "content": "[8] Lin Gao, Jie Yang, Tong Wu, Yu-Jie Yuan, Hongbo Fu, Yu-Kun Lai, and Hao Zhang. Sdm-net: Deep generative network for structured deformable mesh. ACM Transactions on Graphics (TOG), 38(6):1-15, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 447, + 288, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 288, + 491 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 288, + 491 + ], + "type": "text", + "content": "[9] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 491, + 288, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 288, + 534 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 288, + 534 + ], + "type": "text", + "content": "[10] Prakhar Jaiswal, Jinmiao Huang, and Rahul Rai. Assembly-based conceptual 3d modeling with unlabeled components using probabilistic factor graph. Computer-Aided Design, 74: 45-54, 2016. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 536, + 288, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 288, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 288, + 590 + ], + "type": "text", + "content": "[11] R Kenny Jones, Theresa Barton, Xianghao Xu, Kai Wang, Ellen Jiang, Paul Guerrero, Niloy J Mitra, and Daniel Ritchie. Shapeassembly: Learning to generate programs for 3d shape structure synthesis. ACM Transactions on Graphics (TOG), 39(6):1-20, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 591, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 288, + 635 + ], + "type": "text", + "content": "[12] Evangelos Kalogerakis, Siddhartha Chaudhuri, Daphne Koller, and Vladlen Koltun. A probabilistic model for component-based shape synthesis. Acm Transactions on Graphics (TOG), 31(4):1-11, 2012. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 636, + 288, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 288, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 288, + 680 + ], + "type": "text", + "content": "[13] Jun Li, Kai Xu, Siddhartha Chaudhuri, Ersin Yumer, Hao Zhang, and Leonidas Guibas. Grass: Generative recursive autoencoders for shape structures. ACM Transactions on Graphics (TOG), 36(4):1-14, 2017. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 681, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 288, + 714 + ], + "type": "text", + "content": "[14] Yichen Li, Kaichun Mo, Lin Shao, Minhyuk Sung, and Leonidas Guibas. Learning 3d part assembly from a single image. In Computer Vision-ECCV 2020: 16th European" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 546, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 546, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 546, + 95 + ], + "type": "text", + "content": "Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI 16, pages 664-682. Springer, 2020. 1, 2, 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 96, + 547, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 547, + 148 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 547, + 148 + ], + "type": "text", + "content": "[15] Yuval Litvak, Armin Biess, and Aharon Bar-Hillel. Learning pose estimation for high-precision robotic assembly using simulated depth images. In 2019 International Conference on Robotics and Automation (ICRA), pages 3521–3527. IEEE, 2019. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 150, + 547, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 150, + 547, + 215 + ], + "spans": [ + { + "bbox": [ + 308, + 150, + 547, + 215 + ], + "type": "text", + "content": "[16] Jianlan Luo, Eugen Solowjow, Chengtao Wen, Juan Aparicio Ojea, Alice M Agogino, Aviv Tamar, and Pieter Abbeel. Reinforcement learning on variable impedance controller for high-precision robotic assembly. In 2019 International Conference on Robotics and Automation (ICRA), pages 3080–3087. IEEE, 2019. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 217, + 547, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 547, + 259 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 547, + 259 + ], + "type": "text", + "content": "[17] Kaichun Mo, Paul Guerrero, Li Yi, Hao Su, Peter Wonka, Niloy Mitra, and Leonidas J Guibas. Structurenet: Hierarchical graph networks for 3d shape generation. arXiv preprint arXiv:1908.00575, 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 261, + 547, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 547, + 326 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 547, + 326 + ], + "type": "text", + "content": "[18] Kaichun Mo, Shilin Zhu, Angel X Chang, Li Yi, Subarna Tripathi, Leonidas J Guibas, and Hao Su. Partnet: A large-scale benchmark for fine-grained and hierarchical part-level 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 909–918, 2019. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 327, + 547, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 547, + 382 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 547, + 382 + ], + "type": "text", + "content": "[19] Kaichun Mo, He Wang, Xinchen Yan, and Leonidas Guibas. Pt2pc: Learning to generate 3d point cloud shapes from part tree conditions. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI 16, pages 683-701. Springer, 2020. 2, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 383, + 547, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 547, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 547, + 415 + ], + "type": "text", + "content": "[20] Carsten Moenning and Neil A Dodgson. Fast marching farthest point sampling. Technical report, University of Cambridge, Computer Laboratory, 2003. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 416, + 546, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 416, + 546, + 469 + ], + "spans": [ + { + "bbox": [ + 308, + 416, + 546, + 469 + ], + "type": "text", + "content": "[21] Abhinav Narayan, Rajendra Nagar, and Shanmuganathan Raman. Rgl-net: A recurrent graph learning framework for progressive part assembly. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 78-87, 2022. 1, 2, 4, 5, 6, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 471, + 546, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 546, + 524 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 546, + 524 + ], + "type": "text", + "content": "[22] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 525, + 547, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 525, + 547, + 590 + ], + "spans": [ + { + "bbox": [ + 308, + 525, + 547, + 590 + ], + "type": "text", + "content": "[23] Olinde Rodrigues. Des lois géométriques qui regissant les déplacements d'un système solide dans l'espace, et de la variation des coordonnées provenant de ces déplacements considérés indépendamment des causes qui peuvent les produit. Journal de mathématiques pures et appliquées, 5:380-440, 1840. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 592, + 547, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 592, + 547, + 646 + ], + "spans": [ + { + "bbox": [ + 308, + 592, + 547, + 646 + ], + "type": "text", + "content": "[24] Nadav Schor, Oren Katzir, Hao Zhang, and Daniel Cohen-Or. Componet: Learning to generate the unseen by part synthesis and composition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8759–8768, 2019. 5, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 647, + 547, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 547, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 547, + 690 + ], + "type": "text", + "content": "[25] Lin Shao, Toki Migimatsu, and Jeannette Bohg. Learning to scaffold the development of robotic manipulation skills. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 5671-5677. IEEE, 2020. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 691, + 547, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 547, + 714 + ], + "type": "text", + "content": "[26] Dong Wook Shu, Sung Woo Park, and Junseok Kwon. 3d point cloud generative adversarial network based on tree struc" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20858" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 554 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 66, + 72, + 288, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 288, + 105 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 288, + 105 + ], + "type": "text", + "content": "tured graph convolutions. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3859-3868, 2019. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 288, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 288, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 288, + 150 + ], + "type": "text", + "content": "[27] Minhyuk Sung, Hao Su, Vladimir G Kim, Siddhartha Chaudhuri, and Leonidas Guibas. Complementme: Weakly-supervised component suggestions for 3d modeling. ACM Transactions on Graphics (TOG), 36(6):1-12, 2017. 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 288, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 288, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 288, + 195 + ], + "type": "text", + "content": "[28] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 197, + 288, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 288, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 288, + 251 + ], + "type": "text", + "content": "[29] Rundi Wu, Yixin Zhuang, Kai Xu, Hao Zhang, and Baoquan Chen. Pq-net: A generative part seq2seq network for 3d shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 829-838, 2020. 2, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 253, + 288, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 253, + 288, + 306 + ], + "spans": [ + { + "bbox": [ + 48, + 253, + 288, + 306 + ], + "type": "text", + "content": "[30] Yuwen Xiong, Wei-Chiu Ma, Jingkang Wang, and Raquel Urtasun. Learning compact representations for lidar completion and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1074-1083, 2023. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 308, + 288, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 308, + 288, + 363 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 288, + 363 + ], + "type": "text", + "content": "[31] Xianghao Xu, Paul Guerrero, Matthew Fisher, Siddhartha Chaudhuri, and Daniel Ritchie. Unsupervised 3d shape reconstruction by part retrieval and assembly. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8559-8567, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 365, + 288, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 288, + 418 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 288, + 418 + ], + "type": "text", + "content": "[32] Kevin Zakka, Andy Zeng, Johnny Lee, and Shuran Song. Form2fit: Learning shape priors for generalizable assembly from disassembly. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 9404–9410. IEEE, 2020. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 420, + 288, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 420, + 288, + 474 + ], + "spans": [ + { + "bbox": [ + 48, + 420, + 288, + 474 + ], + "type": "text", + "content": "[33] Guanqi Zhan, Qingnan Fan, Kaichun Mo, Lin Shao, Baoquan Chen, Leonidas J Guibas, Hao Dong, et al. Generative 3d part assembly via dynamic graph learning. Advances in Neural Information Processing Systems, 33:6315-6326, 2020. 1, 2, 3, 4, 5, 6, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 475, + 288, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 475, + 288, + 519 + ], + "spans": [ + { + "bbox": [ + 48, + 475, + 288, + 519 + ], + "type": "text", + "content": "[34] Rufeng Zhang, Tao Kong, Weihao Wang, Xuan Han, and Mingyu You. 3d part assembly generation with instance encoded transformer. IEEE Robotics and Automation Letters, 7(4):9051-9058, 2022. 1, 2, 3, 4, 5, 6, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 521, + 288, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 521, + 288, + 554 + ], + "spans": [ + { + "bbox": [ + 48, + 521, + 288, + 554 + ], + "type": "text", + "content": "[35] Vlas Zyrianov, Xiyue Zhu, and Shenlong Wang. Learning to generate realistic lidar point clouds. In European Conference on Computer Vision, pages 17-35. Springer, 2022. 6" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20859" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Generative Image Dynamics/34d8fada-6dfb-4124-8738-6ec5f099992d_content_list.json b/2024/Generative Image Dynamics/34d8fada-6dfb-4124-8738-6ec5f099992d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4a8420dd6323e1bc9c624af225ad134aa991ddad --- /dev/null +++ b/2024/Generative Image Dynamics/34d8fada-6dfb-4124-8738-6ec5f099992d_content_list.json @@ -0,0 +1,1824 @@ +[ + { + "type": "text", + "text": "Generative Image Dynamics", + "text_level": 1, + "bbox": [ + 341, + 130, + 629, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhengqi Li", + "bbox": [ + 174, + 180, + 267, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Richard Tucker", + "bbox": [ + 305, + 180, + 431, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Noah Snavely", + "bbox": [ + 470, + 181, + 584, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Aleksander Holynski", + "bbox": [ + 625, + 181, + 792, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Google Research", + "bbox": [ + 398, + 205, + 535, + 223 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1970c681d962b8853f36ccc8ddaa928629b646be28135670f9eaa973dda86f4c.jpg", + "image_caption": [ + "Input Picture" + ], + "image_footnote": [], + "bbox": [ + 98, + 325, + 330, + 426 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/59ff8290e3cfb9113942f4c942cf80ea62a05c73470708eecd62a494792d00db.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 333, + 364, + 372, + 387 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/78a247a00539d0c217235990f6b0f8246232cb2fa9eb2aef343d17d5599a3585.jpg", + "image_caption": [ + "Spectral Volume (Ge-Space Modal Basis)" + ], + "image_footnote": [], + "bbox": [ + 375, + 305, + 482, + 354 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5a42bd78bf68b0384b71f855a8a0777cec05b33cfd30a08f546469f3abb83c7d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 485, + 305, + 593, + 354 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/34db33c57aca6d04fb45265e6306b77ab81015b48f000e0d3b5a555a59ad3b49.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 375, + 356, + 482, + 404 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/455907a99362c2f7ae7a8d16b5365affcae8744b009709fcfea82bbcd5348b35.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 485, + 354, + 593, + 404 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7a85c481136ceda92f81229273fb6140ebf9894317f442717658aa20a53d3f1b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 421, + 410, + 431, + 424 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d19703c80e58541080695e33803a4edf3861d9c7d7bba0f397bec0177c4387b5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 532, + 410, + 542, + 425 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ac2065fee6863ad784779beaccf133f097476adc78b8661117189f30e348e101.jpg", + "image_caption": [ + "X coefficients" + ], + "image_footnote": [], + "bbox": [ + 375, + 433, + 482, + 481 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6ca5b677fbf8d3e4bf3bec7c9223c52dee59e51042843b48f57d2890205a9704.jpg", + "image_caption": [ + "Y coefficients" + ], + "image_footnote": [], + "bbox": [ + 485, + 433, + 593, + 481 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/37c84aaa5014cb7cd271e56cfcb527993742f81f69768eafa7ac4fd800fdb55f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 277, + 867, + 380 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6f14a0e77e624585c13fd3df632bca6bfc17717dbd0a100efa2417ab8765a6cb.jpg", + "image_caption": [ + "Figure 1. We model a generative image-space prior on scene motion: from a single RGB image, our method generates a spectral volume [23], a motion representation that models dense, long-term pixel trajectories in the Fourier domain. Our learned motion priors can be used to turn a single picture into a seamlessly looping video, or into an interactive simulation of dynamics that responds to user inputs like dragging and releasing points. On the right, we visualize output videos as space-time $X - t$ slices (along the input scanline shown on the left)." + ], + "image_footnote": [], + "bbox": [ + 647, + 383, + 867, + 487 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 566, + 313, + 583 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present an approach to modeling an image-space prior on scene motion. Our prior is learned from a collection of motion trajectories extracted from real video sequences depicting natural, oscillatory dynamics of objects such as trees, flowers, candles, and clothes swaying in the wind. We model dense, long-term motion in the Fourier domain as spectral volumes, which we find are well-suited to prediction with diffusion models. Given a single image, our trained model uses a frequency-coordinated diffusion sampling process to predict a spectral volume, which can be converted into a motion texture that spans an entire video. Along with an image-based rendering module, the predicted motion representation can be used for a number of downstream applications, such as turning still images into seamlessly looping videos, or allowing users to interact with objects in real images, producing realistic simulated dynamics (by interpreting the spectral volumes as image-space modal bases). See our project page for more results: generative-dynamics.github.io.", + "bbox": [ + 73, + 612, + 473, + 885 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 566, + 630, + 583 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The natural world is always in motion, with even seemingly static scenes containing subtle oscillations as a result of wind, water currents, respiration, or other natural rhythms. Emulating this motion is crucial in visual content synthesis—human sensitivity to motion can cause imagery without motion (or with slightly unrealistic motion) to seem uncanny or unreal.", + "bbox": [ + 496, + 595, + 893, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While it is easy for humans to interpret or imagine motion in scenes, training a model to learn or produce realistic scene motion is far from trivial. The motion we observe in the world is the result of a scene's underlying physical dynamics, i.e., forces applied to objects that respond according to their unique physical properties—their mass, elasticity, etc—quantities that are hard to measure and capture at scale. Fortunately, measuring them is not necessary for certain applications: e.g., one can simulate plausible dynamics in a scene by simply analyzing some observed 2D motion [23].", + "bbox": [ + 496, + 688, + 895, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This same observed motion can also serve as a supervisory signal in learning dynamics across scenes—because although observed motion is multi-modal and grounded in complex physical effects, it is nevertheless often predictable:", + "bbox": [ + 496, + 840, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24142", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "candles will flicker in certain ways, trees will sway, and their leaves will rustle. As humans, this predictability is ingrained in our systems of perception: by viewing a still image, we can imagine plausible motions—or, since there might have been many possible such motions, a distribution of natural motions conditioned on that image. Given the facility with which humans are able to model these distributions, a natural research problem is to model them computationally.", + "bbox": [ + 75, + 90, + 470, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent advances in generative models, in particular conditional diffusion models [44, 84, 86], have enabled us to model rich distributions, including distributions of real images conditioned on text [72-74]. This capability has enabled several new applications, such as text-conditioned generation of diverse and realistic image content. Following the success of these image models, recent work has extended these models to other domains, such as videos [7, 43] and 3D geometry [76, 99, 100, 102].", + "bbox": [ + 75, + 214, + 470, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we model a generative prior for image-space scene motion, i.e., the motion of all pixels in a single image. This model is trained on motion trajectories automatically extracted from a large collection of real video sequences. In particular, from each training video we compute motion in the form of a spectral volume [22, 23], a frequency-domain representation of dense, long-range pixel trajectories. Spectral volumes are well-suited to scenes that exhibit oscillatory dynamics, e.g., trees and flowers moving in the wind. We find that this representation is also highly effective as an output of a diffusion model for modeling scene motion. We train a generative model that, conditioned on a single image, can sample spectral volumes from its learned distribution. A predicted spectral volume can then be directly transformed into a motion texture—a set of long-range, per-pixel motion trajectories—that can be used to animate the image. The spectral volume can also be interpreted as an image-space modal basis for use in simulating interactive dynamics [22].", + "bbox": [ + 75, + 352, + 470, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We predict spectral volumes from input images using a diffusion model that generates coefficients one frequency at a time, but coordinates these predictions across frequency bands through a shared attention module. The predicted motions can be used to synthesize future frames (via an image-based rendering model)—turning still images into realistic animations, as illustrated in Fig. 1.", + "bbox": [ + 75, + 626, + 468, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Compared with priors over raw RGB pixels, priors over motion capture more fundamental, lower-dimensional structure that efficiently explains long-range variations in pixel values. Hence, generating intermediate motion leads to more coherent long-term generation and more fine-grained control over animations. We demonstrate the use of our trained model in several downstream applications, such as creating seamless looping videos, editing the generated motions, and enabling interactive dynamic images via image-space modal bases, i.e., simulating the response of object dynamics to user-applied forces [22].", + "bbox": [ + 75, + 734, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 89, + 640, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generative synthesis. Recent advances in generative models have enabled photorealistic synthesis of images conditioned on text prompts [16, 17, 24, 72-74]. These text-to-image models can be augmented to synthesize video sequences by extending the generated image tensors along a time dimension [7, 9, 43, 62, 83, 105, 105, 110]. While these methods can produce video sequences that capture the spatiotemporal statistics of real footage, these videos often suffer from artifacts like incoherent motion, unrealistic temporal variation in textures, and violations of physical constraints like preservation of mass.", + "bbox": [ + 496, + 114, + 893, + 282 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Animating images. Instead of generating videos entirely from text, other techniques take as input a still picture and animate it. Many recent deep learning methods adopt a 3D-Unet architecture to produce video volumes directly [27, 36, 40, 47, 53, 92]. These models are effectively the same video generation models (but conditioned on image information instead of text), and exhibit similar artifacts to those mentioned above. One way to overcome these limitations is to not directly generate the video content itself, but instead animate an input source image through image-based rendering, i.e., moving the image content around according to motion derived from external sources such as a driving video [51, 79-81, 98], motion or 3D geometry priors [8, 29, 46, 63, 64, 66, 89, 96, 100, 101, 103, 108], or user annotations [6, 18, 20, 33, 38, 97, 104, 107]. Animating images according to motion fields yields greater temporal coherence and realism, but these prior methods either require additional guidance signals or user input, or utilize limited motion representations.", + "bbox": [ + 496, + 287, + 895, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motion models and motion priors. In computer graphics, natural, oscillatory 3D motion (e.g., water rippling or trees waving in the wind) can be modeled with noise that is shaped in the Fourier domain and then converted to time-domain motion fields [78, 87]. Some of these methods rely on a modal analysis of the underlying dynamics of the system being simulated [22, 25, 88]. These spectral techniques were adapted to animate plants, water, and clouds from single 2D pictures by Chuang et al. [20], given user annotations. Our work is especially inspired by Davis [23], who connected modal analysis of a scene with the motions observed in a video of that scene, and used this analysis to simulate interactive dynamics from a video. We adopt the frequency-space spectral volume motion representation from Davis et al., extract this representation from a large set of training videos, and show that spectral volumes are suitable for predicting motion from single images with diffusion models.", + "bbox": [ + 496, + 582, + 893, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Other methods have used various motion representations in prediction tasks, where an image or video is used to inform a deterministic future motion estimate [34, 70], or a more rich distribution of possible motions [93, 95, 103]. However,", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24143", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "many of these methods predict an optical flow motion estimate (i.e., the instantaneous motion of each pixel), not full motion trajectories. In addition, much of this prior work is focused on tasks like activity recognition, not on synthesis tasks. More recent work has demonstrated the advantages of modeling and predicting motion using generative models in a number of closed-domain settings such as humans and animals [2, 19, 28, 71, 90, 106].", + "bbox": [ + 75, + 90, + 472, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Videos as textures. Certain moving scenes can be thought of as a kind of texture—termed dynamic textures [26]—that model videos as space-time samples of a stochastic process. Dynamic textures can represent smooth, natural motions like waves, flames, or moving trees, and have been widely used for video classification, segmentation or encoding [12-15, 75]. A related kind of texture, called a video texture, represents a moving scene as a set of input video frames along with transition probabilities between any pair of frames [65, 77]. A number of methods estimate dynamic or video textures through analysis of scene motion and pixel statistics, with the aim of generating seamlessly looping or infinitely varying output videos [1, 21, 32, 58, 59, 77]. In contrast to much of this work, our method learns priors in advance that can then be applied to single images.", + "bbox": [ + 75, + 223, + 472, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Overview", + "text_level": 1, + "bbox": [ + 76, + 468, + 181, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a single picture $I_0$ , our goal is to generate a video $\\{\\hat{I}_1, \\hat{I}_2, \\dots, \\hat{I}_T\\}$ featuring oscillatory motions such as those of trees, flowers, or candle flames swaying in the breeze. Our system consists of two modules: a motion prediction module and an image-based rendering module. Our pipeline begins by using a latent diffusion model (LDM) to predict a spectral volume $S = (S_{f_0}, S_{f_1}, \\dots, S_{f_{K-1}})$ for the input $I_0$ . The predicted spectral volume is then transformed to a motion texture $\\mathcal{F} = (F_1, F_2, \\dots, F_T)$ through an inverse discrete Fourier transform. This motion determines the position of each input pixel at every future time step.", + "bbox": [ + 75, + 496, + 468, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a predicted motion texture, we then animate the input RGB image using a neural image-based rendering technique (Sec. 5). We explore applications of this method, including producing seamless looping animations and simulating interactive dynamics, in Sec. 6.", + "bbox": [ + 75, + 664, + 470, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Predicting motion", + "text_level": 1, + "bbox": [ + 76, + 758, + 251, + 776 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Motion representation", + "text_level": 1, + "bbox": [ + 76, + 784, + 284, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Formally, a motion texture is a sequence of time-varying 2D displacement maps $\\mathcal{F} = \\{F_t|t = 1,\\dots,T\\}$ , where the 2D displacement vector $F_{t}(\\mathbf{p})$ at each pixel coordinate $\\mathbf{p}$ from input image $I_0$ defines the position of that pixel at a future time $t$ [20]. To generate a future frame at time $t$ , one can sput pixels from $I_0$ using the corresponding displacement", + "bbox": [ + 75, + 809, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/54d1208cb540e2bb8f79089f292952a53d70139d7f5cacf10203728b31d89d51.jpg", + "image_caption": [ + "Figure 2. Left: We visualize the average power spectrum for the $X$ and $Y$ motion components extracted from real videos, shown as the blue and green curves. Natural oscillation motions are composed primarily of low-frequency components, and so we use the first $K = 16$ terms, marked with red dots. Right: we show a histogram of the amplitude of Fourier terms at $3.0\\mathrm{Hz}$ after (1) scaling amplitude by image width and height (blue), or (2) frequency adaptive normalization (red). Our adaptive normalization prevents the coefficients from concentrating at extreme values." + ], + "image_footnote": [], + "bbox": [ + 501, + 89, + 697, + 176 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/77ce6d12a9675cdf33837894d68f11fb0ca1d62780bdf8d19002f05cbe16bb42.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 89, + 890, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "map $D_{t}$ , resulting in a forward-warped image $I_t'$ :", + "bbox": [ + 498, + 338, + 823, + 354 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI _ {t} ^ {\\prime} (\\mathbf {p} + F _ {t} (\\mathbf {p})) = I _ {0} (\\mathbf {p}). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 357, + 890, + 375 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "If our goal is to produce a video via a motion texture, then one choice would be to predict a time-domain motion texture directly from an input image. However, the size of the motion texture would need to scale with the length of the video: generating $T$ output frames implies predicting $T$ displacement fields. To avoid predicting such a large output representation for long videos, many prior animation methods either generate video frames autoregressively [7, 29, 57, 60, 92], or predict each future output frame independently via an extra time embedding [4]. However, neither strategy ensures long-term temporal consistency of generated videos.", + "bbox": [ + 496, + 378, + 893, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fortunately, many natural motions can be described as a superposition of a small number of harmonic oscillators represented with different frequencies, amplitude and phases [20, 23, 25, 50, 68]. Because these underlying motions are quasi-periodic, it is natural to model them in the frequency domain. Hence, we adopt from Davis et al. [23] an efficient frequency space representation of motion in a video called a spectral volume, visualized in Fig. 3. A spectral volume is the temporal Fourier transform of per-pixel trajectories extracted from a video.", + "bbox": [ + 496, + 544, + 893, + 694 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given this motion representation, we formulate the motion prediction problem as a multi-modal image-to-image translation task: from an input image to an output motion spectral volume. We adopt latent diffusion models (LDMs) to generate spectral volumes comprised of a $4K$ -channel 2D motion spectrum map, where $K \\ll T$ is the number of frequencies modeled, and where at each frequency we need four scalars to represent the complex Fourier coefficients for the $x$ - and $y$ -dimensions. Note that the motion trajectory of a pixel at future time steps $\\mathcal{F}(\\mathbf{p}) = \\{F_t(\\mathbf{p})|t = 1,2,\\dots T\\}$ and its representation as a spectral volume $\\mathcal{S}(\\mathbf{p}) = \\{S_{f_k}(\\mathbf{p})|k = 0,1,\\dots \\frac{T}{2} -1\\}$ are related by the Fast Fourier transform (FFT):", + "bbox": [ + 496, + 695, + 893, + 877 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} (\\mathbf {p}) = \\operatorname {F F T} (\\mathcal {F} (\\mathbf {p})). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 880, + 890, + 897 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24144", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5744e0ab4ea703956d9398b30c1ee12d84e4d34f4e6d9a5b3f42a74356566278.jpg", + "image_caption": [ + "Figure 3. Motion prediction module. We predict a spectral volume $S$ through a frequency-coordinated denoising model. Each block of the diffusion network $\\epsilon_{\\theta}$ interleaves 2D spatial layers with attention layers (red box, right), and iteratively denoises latent features $z^n$ . The denoised features are fed to a decoder $\\mathcal{D}$ to produce $S$ . During training, we concatenate the downsampled input $I_0$ with noisy latent features encoded from a real motion texture via an encoder $\\mathcal{E}$ , and replace the noisy features with Gaussian noise $z^N$ during inference (left)." + ], + "image_footnote": [], + "bbox": [ + 106, + 89, + 872, + 281 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "How should we select the $K$ output frequencies? Prior work in real-time animation has observed that most natural oscillation motions are composed primarily of low-frequency components [25, 68]. To validate this observation, we computed the average power spectrum of the motion extracted from 1,000 randomly sampled 5-second real video clips. As shown in the left plot of Fig. 2, the power spectrum of the motion decreases exponentially with increasing frequency. This suggests that most natural oscillation motions can indeed be well represented by low-frequency terms. In practice, we found that the first $K = 16$ Fourier coefficients are sufficient to realistically reproduce the original natural motion in a range of real videos and scenes.", + "bbox": [ + 75, + 376, + 473, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Predicting motion with a diffusion model", + "text_level": 1, + "bbox": [ + 76, + 587, + 428, + 603 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We choose a latent diffusion model (LDM) [73] as the backbone for our motion prediction module, as LDMs are more computationally efficient than pixel-space diffusion models, while preserving synthesis quality. A standard LDM consists of two main modules: (1) a variational autoencoder (VAE) that compresses the input image to a latent space through an encoder $z = E(I)$ , then reconstructs the input from the latent features via a decoder $I = D(z)$ , and (2) a U-Net based diffusion model that learns to iteratively denoise features starting from Gaussian noise. Our training applies this process not to RGB images but to spectral volumes from real video sequences, which are encoded and then diffused for $n$ steps with a pre-defined variance schedule to produce noisy latents $z^n$ . The 2D U-Nets are trained to denoise the noisy latents by iteratively estimating the noise $\\epsilon_{\\theta}(z^{n};n,c)$ used to update the latent feature at each step $n\\in (1,2,\\dots,N)$ . The training loss for the LDM is written as", + "bbox": [ + 75, + 611, + 473, + 868 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {L D M}} = \\mathbb {E} _ {n \\in \\mathcal {U} [ 1, N ], \\epsilon^ {n} \\in \\mathcal {N} (0, 1)} \\left[ \\left| \\left| \\epsilon^ {n} - \\epsilon_ {\\theta} \\left(z ^ {n}; n, c\\right) \\right| \\right| ^ {2} \\right] \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 88, + 878, + 468, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $c$ is the embedding of any conditional signal, such as text, or, in our case, the first frame of the training video sequence, $I_0$ . The clean latent features $z^0$ are then passed through the decoder to recover the spectral volume.", + "bbox": [ + 498, + 376, + 893, + 438 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Frequency adaptive normalization. One issue we observed is that motion textures have particular distribution characteristics across frequencies. As visualized in the left plot of Fig. 2, the amplitude of the spectral volumes spans a range of 0 to 100 and decays approximately exponentially with increasing frequency. As diffusion models require that the absolute values of the output are between -1 and 1 for stable training and denoising [44], we must normalize the coefficients of $S$ extracted from real videos before using them for training. If we scale the magnitudes of these coefficients to [0,1] based on the image dimensions as in prior work [29, 76], nearly all the coefficients at higher frequencies will end up close to zero, as shown in the right plot of Fig. 2. Models trained on such data can produce inaccurate motions, since during inference, even small prediction errors can cause large relative errors after denormalization.", + "bbox": [ + 496, + 470, + 893, + 714 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address this issue, we employ a simple but effective frequency adaptive normalization method: First, we independently normalize Fourier coefficients at each frequency based on statistics computed from the training set. Namely, for each individual frequency $f_{j}$ , we compute the $95^{\\mathrm{th}}$ percentile of Fourier coefficient magnitudes over all input samples and use that value as a per-frequency scaling factor $s f_{j}$ . We then apply a power transformation to each scaled Fourier coefficient to pull it away from extreme values. In practice, we observe that a square root performs better than other nonlinear transformations such as log or reciprocal. In summary, the final coefficient values of spectral volume $S(\\mathbf{p})$ at", + "bbox": [ + 496, + 719, + 895, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24145", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "frequency $f_{j}$ (used for training our LDM) are computed as", + "bbox": [ + 76, + 90, + 465, + 107 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS _ {f _ {j}} ^ {\\prime} (\\mathbf {p}) = \\operatorname {s i g n} \\left(S _ {f _ {j}}\\right) \\sqrt {\\left| \\frac {S _ {f _ {j}} (\\mathbf {p})}{s _ {f _ {j}}} \\right|}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 114, + 468, + 154 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown on the right plot of Fig. 2, after applying frequency adaptive normalization, the spectral volume coefficients distribute more evenly.", + "bbox": [ + 75, + 161, + 468, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Frequency-coordinated denoising. The straightforward way to predict a spectral volume $S$ with $K$ frequency bands is to output a tensor of $4K$ channels from a single diffusion U-Net. However, as in prior work [7], we observe that training a model to produce a large number of channels can yield over-smoothed, inaccurate outputs. An alternative would be to independently predict each individual frequency slice by injecting an extra frequency embedding into the LDM [4], but this design choice would result in uncorrelated predictions in the frequency domain, leading to unrealistic motion.", + "bbox": [ + 75, + 210, + 470, + 361 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Therefore, inspired by recent video diffusion work [7], we propose a frequency-coordinated denoising strategy, illustrated in Fig. 3. In particular, given an input image $I_0$ , we first train an LDM $\\epsilon_{\\theta}$ to predict a single 4-channel frequency slice of spectral volume $S_{f_j}$ , where we inject an extra frequency embedding along with the time-step embedding into the LDM. We then freeze the parameters of this LDM $\\epsilon_{\\theta}$ , introduce attention layers interleaved with the 2D spatial layers of $\\epsilon_{\\theta}$ across the $K$ frequency bands, and fine-tune. Specifically, for a batch size $B$ , the 2D spatial layers of $\\epsilon_{\\theta}$ treat the corresponding $B \\cdot K$ noisy latent features of channel size $C$ as independent samples with shape $\\mathcal{R}^{(B \\cdot K) \\times C \\times H \\times W}$ . The attention layer then interprets these as consecutive features spanning the frequency axis, and we reshape the latent features from previous 2D spatial layers to $\\mathcal{R}^{B \\times K \\times C \\times H \\times W}$ before feeding them to the attention layers. In other words, the frequency attention layers are fine-tuned to coordinate all frequency slices so as to produce coherent spectral volumes. In our experiments, we see that the average VAE reconstruction error improves from 0.024 to 0.018 when we switch from a single 2D U-Net to a frequency-coordinated denoising module, suggesting an improved upper bound on LDM prediction accuracy; in Sec. 7.3, we also show that this design choice improves video generation quality.", + "bbox": [ + 75, + 362, + 472, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Image-based rendering", + "text_level": 1, + "bbox": [ + 76, + 739, + 295, + 757 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We now describe how we take a spectral volume $S$ predicted for a given input image $I_0$ and render a future frame $\\hat{I}_t$ at time $t$ . We first derive a motion texture in the time domain using the inverse temporal FFT applied at each pixel $\\mathcal{F}(\\mathbf{p}) = \\mathrm{FFT}^{-1}(S(\\mathbf{p}))$ . To produce a future frame $\\hat{I}_t$ , we adopt a deep image-based rendering technique and perform splatting with the predicted motion field $F_t$ to forward-warp the encoded $I_0$ , as shown in Fig. 4. Since forward warping can lead to holes, and multiple source pixels can map to the same output", + "bbox": [ + 75, + 763, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/983a2734f6cefff5069132c49a005bc46464b113a35388810637e96b3a9af55f.jpg", + "image_caption": [ + "Figure 4. Rendering module. We fill in missing content and refine the warped input image using a deep image-based rendering module, where multi-scale features are extracted from the input image $I_0$ . Softmax splatting is then applied over the features with a motion field $F_t$ from time 0 to $t$ (subject to the weights $W$ ). The warped features are fed to an image synthesis network to produce the rendered image $\\hat{I}_t$ ." + ], + "image_footnote": [], + "bbox": [ + 506, + 92, + 883, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2D location, we adopt the feature pyramid softmax splatting strategy proposed in prior work on frame interpolation [67].", + "bbox": [ + 498, + 393, + 893, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, we encode $I_0$ through a feature extractor network to produce a multi-scale feature map. For each individual feature map at scale $j$ , we resize and scale the predicted 2D motion field $F_{t}$ according to the resolution. As in Davis et al. [22], we use predicted flow magnitude as a proxy for depth to determine the contributing weight of each source pixel mapped to its destination location. In particular, we compute a per-pixel weight, $W(\\mathbf{p}) = \\frac{1}{T}\\sum_{t}||F_{t}(\\mathbf{p})||_{2}^{2}$ as the average magnitude of the predicted motion texture. In other words, we assume large motions correspond to moving foreground objects, and small or zero motions correspond to background. We use motion-derived weights instead of learnable ones as [46] because we observe that in the single-view case, learnable weights are not effective for addressing disocclusion ambiguities.", + "bbox": [ + 496, + 424, + 892, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the motion field $F_{t}$ and weights $W$ , we apply softmax splatting to warp the feature map at each scale to produce a warped feature. The warped features are then injected into the corresponding blocks of an image synthesis decoder to produce a final rendered image $\\hat{I}_{t}$ .", + "bbox": [ + 496, + 650, + 893, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We jointly train the feature extractor and synthesis networks with start and target frames $(I_0, I_t)$ randomly sampled from real videos, using the estimated flow field from $I_0$ to $I_t$ to warp encoded features from $I_0$ , and supervising predictions $\\hat{I}_t$ against $I_t$ with a VGG perceptual loss [49].", + "bbox": [ + 496, + 726, + 893, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6. Applications", + "text_level": 1, + "bbox": [ + 500, + 815, + 630, + 832 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Image-to-video. Our system enables the animation of a single still picture by first predicting a motion spectral volume from the input image and generating an animation by applying our image-based rendering module to the motion", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24146", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "texture transformed from the spectral volume. Since we explicitly model scene motion, this allows us to produce slow-motion videos by linearly interpolating the motion texture, or to magnify (or minify) animated motions by adjusting the amplitude of the predicted spectral volume coefficients.", + "bbox": [ + 75, + 90, + 470, + 167 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Seamless looping. Many applications require videos that loop seamlessly, where there is no discontinuity between the start and end of the video. Unfortunately, it is hard to find a large collection of seamlessly looping videos for training. Instead, we devise a method to use our motion diffusion model, trained on regular non-looping video clips, to produce seamless looping video. Inspired by recent work on guidance for image editing [3, 30], our method is a motion self-guidance technique that guides the motion denoising sampling processing using explicit looping constraints. In particular, at each iterative denoising step during inference, we incorporate an additional motion guidance signal alongside standard classifier-free guidance [45], where we enforce each pixel's position and velocity at the start and end frames to be as similar as possible:", + "bbox": [ + 75, + 167, + 470, + 393 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\epsilon} ^ {n} = (1 + w) \\epsilon_ {\\theta} \\left(z ^ {n}; n, c\\right) - w \\epsilon_ {\\theta} \\left(z ^ {n}; n, \\emptyset\\right) + u \\sigma^ {n} \\nabla_ {z ^ {n}} \\mathcal {L} _ {g} ^ {n}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 402, + 467, + 421 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {g} ^ {n} = \\left\\| F _ {T} ^ {n} - F _ {1} ^ {n} \\right\\| _ {1} + \\left\\| \\nabla F _ {T} ^ {n} - \\nabla F _ {1} ^ {n} \\right\\| _ {1} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 422, + 468, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $F_{t}^{n}$ is the predicted 2D displacement field at time $t$ and denoising step $n$ . $w$ is the classifier-free guidance weight, and $u$ is the motion self-guidance weight. In the supplemental video, we apply a baseline appearance-based looping algorithm [58] to generate a looping video from our non-looping output, and show that our motion self-guidance technique produces seamless looping videos with less distortion and fewer artifacts.", + "bbox": [ + 75, + 445, + 468, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Interactive dynamics from a single image. Davis et al. [22] show that the spectral volume, evaluated at certain resonant frequencies, can approximate an image-space modal basis that is a projection of the vibration modes of the underlying scene (or, more generally, captures spatial and temporal correlations in oscillatory dynamics), and can be used to simulate the object's response to a user-defined force. We adopt this modal analysis method [22, 69], allowing us to write the image-space 2D motion displacement field for the object's physical response as a weighted sum of motion spectrum coefficients $S_{f_j}$ modulated by the state of complex modal coordinates $\\mathbf{q}_{f_j}(t)$ at each simulated time step $t$ :", + "bbox": [ + 75, + 569, + 470, + 751 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF _ {t} (\\mathbf {p}) = \\sum_ {f _ {j}} S _ {f _ {j}} (\\mathbf {p}) \\mathbf {q} _ {f _ {j}} (t) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 756, + 468, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We simulate the state of the modal coordinates $\\mathbf{q}_{f_j}(t)$ via an explicit Euler method applied to the equations of motion for a decoupled mass-spring-damper system represented in modal space [22, 23, 69]. We refer readers to supplementary material and original work for a full derivation. Note that our method produces an interactive scene from a single picture, whereas these prior methods required a video as input.", + "bbox": [ + 75, + 795, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/00340d8fafd80fcd22e9f4f54eafc484a83f64cde4b17786fa02e4449adef3ca.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodImage SynthesisVideo Synthesis
FIDKIDFVDFVD32DTFVDDTFVD32
TATS [35]65.81.67265.6419.622.640.7
Stochastic I2V [27]68.33.12253.5320.916.741.7
MCVD [92]63.42.97208.6270.419.553.9
LFDM [66]47.61.70187.5254.313.045.6
DMVFN [48]37.91.09206.5316.311.254.5
Endo et al. [29]10.40.19166.0231.65.3565.1
Holynski et al. [46]11.20.20179.0253.77.2346.8
Ours4.030.0847.162.92.536.75
", + "bbox": [ + 501, + 90, + 898, + 224 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparisons on the test set. We report both image synthesis and video synthesis quality. Here, KID is scaled by 100. Lower is better for all error. See Sec. 7.1 for descriptions of baselines and error metrics.", + "bbox": [ + 496, + 234, + 893, + 292 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "7. Experiments", + "text_level": 1, + "bbox": [ + 500, + 329, + 632, + 345 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. We use an LDM [73] as the backbone for predicting spectral volumes, for which we use a VAE with a continuous latent space of dimension 4. We train the VAE with an $L_{1}$ reconstruction loss, a multi-scale gradient consistency loss [54-56], and a KL-divergence loss with respective weights of $1,0.2,10^{-6}$ . We train the same 2D U-Net used in the original LDM work to perform iterative denoising with a simple MSE loss [44], and adopt the attention layers from [41] for frequency-coordinated denoising. For quantitative evaluation, we train both VAE and LDM on images of size $256 \\times 160$ from scratch for fair comparisons, and it takes around 6 days to converge using 16 Nvidia A100 GPUs. For our main quantitative and qualitative results, we run the motion diffusion model with DDIM [85] for 250 steps. We also show generated videos of up to a resolution of $512 \\times 288$ , created by fine-tuning a pre-trained image inpainting LDM model [73] on our dataset.", + "bbox": [ + 496, + 354, + 893, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We adopt ResNet-34 [39] for the feature extractor in our IBR module. Our image synthesis network is based on an architecture for conditional image inpainting [57, 109]. Our rendering module runs in real-time at 25FPS on a Nvidia V100 GPU during inference. We adopt universal guidance [3] to produce seamless looping videos, where we set weights $w = 1.75$ , $u = 200$ , and use 500 DDIM steps with 2 self-recurrence iterations.", + "bbox": [ + 496, + 611, + 893, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Data. We collect and process a set of 3,015 videos of natural scenes exhibiting oscillatory motions from online sources our own captures. We withhold $10\\%$ of videos for testing and use the remainder for training. To extract ground truth motion trajectories, we apply a coarse-to-fine flow method [10, 61] between each selected starting image and every future frame of the video. As training data, we take every 10th video frame as input images and derive the corresponding ground truth spectral volumes using the computed motion trajectories across the following 149 frames. In total, our data consists of over 150K image-motion pairs.", + "bbox": [ + 496, + 734, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24147", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/efc2a8f8a62808bfbc24690cd316a86b0c962bc313f6cbd97826670372d514cf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 89, + 197, + 143 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/687b4757f4f7ff8f19ece485a333350c452fc6d5c64eb4fbcc8c046edab9c00c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 197, + 90, + 883, + 143 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7839a926526e3e4dd3a6af660f6a59596f3b2f4d27111727c750b3aacf0b78ef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 146, + 197, + 203 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5b28f4e228885a1d8e872a723429e8503719246c835cfa996b23a72fda71a388.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 197, + 146, + 883, + 203 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8145382f33d8ec80598737642c71212ac9f35bf8987751b375c4cfd7fb429695.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 205, + 197, + 262 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9ec4504c6cb8a6f632d254066af28657078a4fead1f3d0b52dcd4bf665bee686.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 197, + 205, + 883, + 261 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7dc991b385b536cee35167fb8d7974d35270558ad365c6855f1f8fefb7d8c4c0.jpg", + "image_caption": [ + "Input image" + ], + "image_footnote": [], + "bbox": [ + 83, + 263, + 197, + 321 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d99e174881d2b858ec871d89b560ce036d391c4194bf134bdbd875d6a53ba017.jpg", + "image_caption": [ + "Reference", + "Stochastic-I2V [27]", + "MCVD[92]", + "Endo et al. [29]", + "Holynski et al. [46].", + "Ours", + "Figure 5. X - t slices of videos generated by different approaches. From left to right: input image and corresponding X - t video slices from the ground truth video, from videos generated by three baselines [27, 29, 46, 92], and finally videos generated by our approach." + ], + "image_footnote": [], + "bbox": [ + 197, + 263, + 883, + 320 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/878d42b430e0dff41b50c1a4bd46edb1567a538fbaae11017e6b474aee4599d7.jpg", + "image_caption": [ + "Figure 6. Sliding window FID and DTFVD. We show sliding window FID with window size 30 frames, and DTFVD with size 16 frames, for videos generated by different methods." + ], + "image_footnote": [], + "bbox": [ + 78, + 383, + 274, + 484 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/74230c64b94baadab59b4863b66b451dff1512f6d155c3d062f9c5de1350609c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 383, + 467, + 484 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines. We compare our approach to recent single-image animation and video prediction methods. Endo et al. [29] and DMVFN [48] predict instantaneous 2D motion fields and render future frames auto-regressively. Holynski et al. [46] instead simulates motion through a single static Eulerian motion description. Other recent work such as Stochastic Image-to-Video (Stochastic-I2V) [27], TATS [35], and MCVD [92] adopt VAEs, transformers, or diffusion models to directly predict raw video frames; LFDM [66] generates future frames by predicting flow volumes and warping latents in a diffusion model. We train all the above methods on our data using their respective open-source implementations.", + "bbox": [ + 75, + 560, + 468, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate the quality of the videos generated by our approach and by prior baselines in two ways. First, we evaluate the quality of individual synthesized frames using metrics designed for image synthesis tasks. We adopt the Fréchet Inception Distance (FID) [42] and Kernel Inception Distance (KID) [5] to measure the average distance between the distributions of generated frames and ground truth frames.", + "bbox": [ + 75, + 742, + 470, + 847 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Second, to evaluate the quality and temporal coherence of synthesized videos, we adopt the Fréchet Video Dis", + "bbox": [ + 75, + 848, + 470, + 878 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/736a084db9418f093c03885dc2c16f7b5577ef5e5104d6acefaa0b40fd101295.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodImage SynthesisVideo Synthesis
FIDKIDFVDFVD32DTFVDDTFVD32
Repeat I0--237.5316.75.3045.6
K=43.920.0760.378.43.128.59
K=83.950.0752.168.72.717.37
K=244.090.0848.265.12.506.94
w/o adaptive norm.4.530.0962.780.13.168.19
Independent pred.4.000.0852.571.32.707.40
Volume pred.4.740.0953.771.12.837.79
Baseline splat [46]4.250.0949.566.82.837.27
Full (K=16)4.030.0847.162.92.536.75
", + "bbox": [ + 500, + 383, + 897, + 558 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Ablation study. Sec. 7.3 describes each configuration.", + "bbox": [ + 506, + 561, + 880, + 577 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tance [91] with window size 16 (FVD) and 32 $(\\mathrm{FVD}_{32})$ based on an I3D model [11] trained on the Human Kinetics datasets [52]. To more faithfully reflect synthesis quality for the natural oscillation motions we seek to generate, we also adopt the Dynamic Texture Frechet Video Distance [27], which measures distance from videos with window size 16 (DTFVD) and size 32 $(\\mathrm{DTFVD}_{32})$ , using a I3D model trained on the Dynamic Textures Database [37], a dataset consisting primarily of natural motion textures.", + "bbox": [ + 496, + 618, + 893, + 755 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further use a sliding window FID of window size 30 frames, and a sliding window DTFVD with window size 16 frames, as in [57, 60], to measure how generated video quality degrades over time. For all methods, we evaluate metrics at $256 \\times 128$ resolution by center-cropping.", + "bbox": [ + 496, + 756, + 893, + 833 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "7.1.Quantitative results", + "text_level": 1, + "bbox": [ + 500, + 845, + 687, + 861 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1 shows quantitative comparisons between our approach and baselines on our test set. Our approach signifi-", + "bbox": [ + 496, + 869, + 893, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "1We use the open-source implementation of [46] from Fan et al. [82].", + "bbox": [ + 94, + 886, + 460, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24148", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1e82f1326227507e720cca0dca4200e264301081fbda44457fe31c861c4cb02a.jpg", + "image_caption": [ + "Figure 7. We show generated future frames from three recent large video diffusion models [31, 36, 97]." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 470, + 148 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "cantly outperforms prior single-image animation baselines in terms of both image and video synthesis quality. Specifically, our much lower FVD and DT-FVD distances suggest that the videos generated by our approach are more realistic and more temporally coherent. Further, Fig. 6 shows the sliding window FID and sliding window DT-FVD distances of generated videos from different methods. Thanks to the global spectral volume representation, videos generated by our approach do not suffer from degradation over time.", + "bbox": [ + 75, + 209, + 470, + 345 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7.2. Qualitative results", + "text_level": 1, + "bbox": [ + 76, + 357, + 253, + 372 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We visualize qualitative comparisons between videos as spatio-temporal $X$ - $t$ slices of the generated videos, a standard way of visualizing small motions in a video [94]. As shown in Fig. 5, our generated video dynamics more strongly resemble the motion patterns observed in the corresponding real reference videos (second column), compared to other methods. Baselines such as Stochastic I2V [27] and MCVD [92] fail to model both appearance and motion realistically over time. Endo et al. [29] and Holynski et al. [46] produce video frames with fewer artifacts but exhibits oversmooth or non-oscillatory motion over time. We refer readers to supplementary material to assess the quality of generated video frames and estimated motion across different methods.", + "bbox": [ + 75, + 381, + 470, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7.3. Ablation study", + "text_level": 1, + "bbox": [ + 76, + 589, + 225, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct an ablation study to validate the major design choices in our motion prediction and rendering modules, comparing our full configuration with different variants. Specifically, we evaluate results using different numbers of frequency bands $K = 4, 8, 16, 24$ . We observe that increasing the number of frequency bands improves video prediction quality, but the improvement is marginal with more than 16 frequencies. Next, we remove adaptive frequency normalization from the ground truth spectral volumes, and instead just scale them based on input image width and height (w/o adaptive norm.). Additionally, we remove the frequency coordinated-denoising module (Independent pred.), or replace it with a simpler DM where a tensor volume of $4K$ channel spectral volumes are predicted jointly via a single 2D U-net diffusion model (Volume pred.). Finally, we compare results where we use a baseline rendering method that applies softmax splatting over single-scale features subject to learnable weights as used by [46] (Baseline splat). We also add a baseline where the generated video is a volume by", + "bbox": [ + 75, + 613, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fade8472536fcf69e69f99edd72d2d97d82ea8a2b72a746f509115affa77424b.jpg", + "image_caption": [ + "Figure 8. Limitations. We show examples of rendered future frames (even), and overlay of input and rendered images (odd). Our method can produce artifacts in regions of thin objects or large motions, and regions requiring filling large amount of new contents." + ], + "image_footnote": [], + "bbox": [ + 501, + 89, + 893, + 132 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "repeating input image $N$ times (Repeat $I_0$ ). From Table 2, we observe that all simpler or alternative configurations lead to worse performance compared with our full model.", + "bbox": [ + 498, + 220, + 892, + 265 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7.4. Comparing to large video models", + "text_level": 1, + "bbox": [ + 500, + 273, + 790, + 290 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We further perform a user study and compare our generated animations with ones from recent large video diffusion models:AnimateDiff [36], ModelScope [97] and Gen-2 [31], which predict video volumes directly. On a randomly selected 30 videos from the test set, we ask users \"which video is more realistic?\". Users report a $80.9\\%$ preference for our approach over others. Moreover, as shown in Fig. 7, we observe that the generated videos from these baselines are either unable to adhere to the input image content, or exhibit gradual color drift and distortion over time. We refer readers to the supplementary material for full comparisons.", + "bbox": [ + 498, + 297, + 893, + 464 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8. Discussion and conclusion", + "text_level": 1, + "bbox": [ + 500, + 478, + 741, + 493 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Since our approach only predicts lower frequencies of a spectral volume, it can fail to model non-oscillating motions or high-frequency vibrations—this may be resolved by using learned motion bases. Furthermore, the quality of generated videos relies on the quality of underlying motion trajectories, which may degrade in scenes with thin moving objects or objects with large displacements. Even if correct, motions that require generating large amounts of new unseen content may also cause degradations (Fig. 8).", + "bbox": [ + 498, + 503, + 893, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Conclusion. We present a new approach for modeling natural oscillation dynamics from a single still picture. Our image-space motion prior is represented with spectral volumes, a frequency representation of per-pixel motion trajectories, which we find to be efficient and effective for prediction with diffusion models, and which we learn from collections of real world videos. Spectral volumes are predicted using frequency-coordinated latent diffusion model and are used to animate future video frames through an image-based rendering module. We show that our approach produces photo-realistic animations from a single picture and significantly outperforms prior baselines, and that it can enable several downstream applications such as creating seamlessly looping or interactive image dynamics.", + "bbox": [ + 498, + 641, + 893, + 852 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. We thank Abe Davis, Rick Szeliski, Andrew Liu, Boyang Deng, Qianqian Wang, Xuan Luo, and Lucy Chai for fruitful discussions and helpful comments.", + "bbox": [ + 500, + 854, + 892, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24149", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Aseem Agarwala, Ke Colin Zheng, Chris Pal, Maneesh Agrawala, Michael Cohen, Brian Curless, David Salesin, and Richard Szeliski. Panoramic video textures. In ACM SIGGRAPH 2005 Papers, pages 821-827. 2005.", + "[2] Hyemin Ahn, Esteve Valls Mascaro, and Dongheui Lee. Can we use diffusion probabilistic models for 3d motion prediction? arXiv preprint arXiv:2302.14503, 2023.", + "[3] Arpit Bansal, Hong-Min Chu, Avi Schwarzschild, Soumyadip Sengupta, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Universal guidance for diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 843-852, 2023.", + "[4] Hugo Bertiche, Niloy J Mitra, Kuldeep Kulkarni, ChunHao P Huang, Tuanfeng Y Wang, Meysam Madadi, Sergio Escalera, and Duygu Ceylan. Blowing in the wind: Cyclenet for human cinemagrams from still images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 459-468, 2023.", + "[5] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying MMD GANs. arXiv preprint arXiv:1801.01401, 2018.", + "[6] Andreas Blattmann, Timo Milbich, Michael Dorkenwald, and Björn Ommer. ipoke: Poking a still image for controlled stochastic video synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14707-14717, 2021.", + "[7] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023.", + "[8] Richard Strong Bowen, Richard Tucker, Ramin Zabih, and Noah Snavely. Dimensions of motion: Monocular prediction through flow subspaces. In 2022 International Conference on 3D Vision (3DV), pages 454-464. IEEE, 2022.", + "[9] Tim Brooks, Janne Hellsten, Miika Aittala, Ting-Chun Wang, Timo Aila, Jaakko Lehtinen, Ming-Yu Liu, Alexei Efros, and Tero Karras. Generating long videos of dynamic scenes. Advances in Neural Information Processing Systems, 35:31769-31781, 2022.", + "[10] Thomas Brox, Andrés Bruhn, Nils Papenberg, and Joachim Weickert. High accuracy optical flow estimation based on a theory for warping. In Proc. European Conf. on Computer Vision (ECCV), pages 25-36. Springer, 2004.", + "[11] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017.", + "[12] Dan Casas, Marco Volino, John Collomosse, and Adrian Hilton. 4d video textures for interactive character appearance. In Computer Graphics Forum, volume 33, pages 371-380. Wiley Online Library, 2014.", + "[13] Antoni B Chan and Nuno Vasconcelos. Mixtures of dynamic textures. In Tenth IEEE International Conference on Computer Vision (ICCV'05) Volume 1, volume 1, pages 641-647. IEEE, 2005." + ], + "bbox": [ + 86, + 114, + 470, + 887 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Antoni B Chan and Nuno Vasconcelos. Classifying video with kernel dynamic textures. In 2007 IEEE Conference on Computer Vision and Pattern Recognition, pages 1-6. IEEE, 2007.", + "[15] Antoni B Chan and Nuno Vasconcelos. Modeling, clustering, and segmenting video with mixtures of dynamic textures. IEEE transactions on pattern analysis and machine intelligence, 30(5):909-926, 2008.", + "[16] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023.", + "[17] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11315-11325, 2022.", + "[18] Tsai-Shien Chen, Chieh Hubert Lin, Hung-Yu Tseng, Tsung-Yi Lin, and Ming-Hsuan Yang. Motion-conditioned diffusion model for controllable video synthesis. arXiv preprint arXiv:2304.14404, 2023.", + "[19] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18000-18010, 2023.", + "[20] Yung-Yu Chuang, Dan B Goldman, Ke Colin Zheng, Brian Curless, David H Salesin, and Richard Szeliski. Animating pictures with stochastic motion textures. In ACM SIGGRAPH 2005 Papers, pages 853-860. 2005.", + "[21] Vincent C Couture, Michael S Langer, and Sebastien Roy. Omnistereo video textures without ghosting. In 2013 International Conference on 3D Vision-3DV 2013, pages 64–70. IEEE, 2013.", + "[22] Abe Davis, Justin G Chen, and Frédo Durand. Image-space modal bases for plausible manipulation of objects in video. ACM Transactions on Graphics (TOG), 34(6):1-7, 2015.", + "[23] Myers Abraham Davis. Visual vibration analysis. PhD thesis, Massachusetts Institute of Technology, 2016.", + "[24] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021.", + "[25] Julien Diener, Mathieu Rodriguez, Lionel Baboud, and Lionel Reveret. Wind projection basis for real-time animation of trees. In Computer graphics forum, volume 28, pages 533-540. Wiley Online Library, 2009.", + "[26] Gianfranco Doretto, Alessandro Chiuso, Ying Nian Wu, and Stefano Soatto. Dynamic textures. International journal of computer vision, 51:91-109, 2003.", + "[27] Michael Dorkenwald, Timo Milbich, Andreas Blattmann, Robin Rombach, Konstantinos G. Derpanis, and Bjorn Ommer. Stochastic image-to-video synthesis using cinns. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3742-3753, June 2021.", + "[28] Yuming Du, Robin Kips, Albert Pumarola, Sebastian Starke, Ali Thabet, and Artsiom Sanakoyeu. Avatars grow legs: Generating smooth human motion from sparse tracking inputs with diffusion model. In Proceedings of the IEEE/CVF" + ], + "bbox": [ + 509, + 92, + 893, + 892 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24150", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Computer Vision and Pattern Recognition, pages 481-490, 2023.", + "[29] Yuki Endo, Yoshihiro Kanamori, and Shigeru Kuriyama. Animating landscape: Self-supervised learning of decoupled motion and appearance for single-image video synthesis. ACM Transactions on Graphics (Proceedings of ACM SIGGRAPH Asia 2019), 38(6):175:1-175:19, 2019.", + "[30] Dave Epstein, Allan Jabri, Ben Poole, Alexei A Efros, and Aleksander Holynski. Diffusion self-guidance for controllable image generation. arXiv preprint arXiv:2306.00986, 2023.", + "[31] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7346-7356, 2023.", + "[32] Matthew Flagg, Atsushi Nakazawa, Qiushuang Zhang, Sing Bing Kang, Young Kee Ryu, Irfan Essa, and James M Rehg. Human video textures. In Proceedings of the 2009 symposium on Interactive 3D graphics and games, pages 199-206, 2009.", + "[33] Jean-Yves Franceschi, Edouard Delasalles, Mickaël Chen, Sylvain Lamprier, and Patrick Gallinari. Stochastic latent residual video prediction. In International Conference on Machine Learning, pages 3233-3246. PMLR, 2020.", + "[34] Ruohan Gao, Bo Xiong, and Kristen Grauman. Im2Flow: Motion hallucination from static images for action recognition. In Proc. Computer Vision and Pattern Recognition (CVPR), 2018.", + "[35] Songwei Ge, Thomas Hayes, Harry Yang, Xi Yin, Guan Pang, David Jacobs, Jia-Bin Huang, and Devi Parikh. Long video generation with time-agnostic vqgan and timesensitive transformer. arXiv preprint arXiv:2204.03638, 2022.", + "[36] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023.", + "[37] Isma Hadji and Richard P Wildes. A new large scale dynamic texture dataset with application to convnet understanding. In Proceedings of the European Conference on Computer Vision (ECCV), pages 320-335, 2018.", + "[38] Zekun Hao, Xun Huang, and Serge Belongie. Controllable video generation with sparse trajectories. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7854-7863, 2018.", + "[39] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016.", + "[40] Yingqing He, Menghan Xia, Haoxin Chen, Xiaodong Cun, Yuan Gong, Jinbo Xing, Yong Zhang, Xintao Wang, Chao Weng, Ying Shan, et al. animate-a-story: Storytelling with retrieval-augmented video generation. arXiv preprint arXiv:2307.06940, 2023.", + "[41] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2022." + ], + "bbox": [ + 86, + 92, + 470, + 891 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017.", + "[43] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022.", + "[44] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020.", + "[45] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022.", + "[46] Aleksander Holynski, Brian L Curless, Steven M Seitz, and Richard Szeliski. Animating pictures with Eulerian motion fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5810-5819, 2021.", + "[47] Tobias Hoppe, Arash Mehrjou, Stefan Bauer, Didrik Nielsen, and Andrea Dittadi. Diffusion models for video prediction and infilling. Trans. Mach. Learn. Res., 2022, 2022.", + "[48] Xiaotao Hu, Zhewei Huang, Ailin Huang, Jun Xu, and Shuchang Zhou. A dynamic multi-scale voxel flow network for video prediction. ArXiv, abs/2303.09875, 2023.", + "[49] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016.", + "[50] Hitoshi Kanda and Jun Ohya. Efficient, realistic method for animating dynamic behaviors of 3d botanical trees. In 2003 International Conference on Multimedia and Expo. ICME'03. Proceedings (Cat. No. 03TH8698), volume 2, pages II-89. IEEE, 2003.", + "[51] Johanna Karras, Aleksander Holynski, Ting-Chun Wang, and Ira Kemelmacher-Shlizerman. Dreampose: Fashion image-to-video synthesis via stable diffusion. arXiv preprint arXiv:2304.06025, 2023.", + "[52] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017.", + "[53] Alex X Lee, Richard Zhang, Frederik Ebert, Pieter Abbeel, Chelsea Finn, and Sergey Levine. Stochastic adversarial video prediction. arXiv preprint arXiv:1804.01523, 2018.", + "[54] Zhengqi Li, Tali Dekel, Forrester Cole, Richard Tucker, Noah Snavely, Ce Liu, and William T Freeman. Learning the depths of moving people by watching frozen people. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4521-4530, 2019.", + "[55] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2041-2050, 2018.", + "[56] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snively. Dynibar: Neural dynamic image-based" + ], + "bbox": [ + 509, + 92, + 893, + 891 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24151", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4273-4284, 2023.", + "[57] Zhengqi Li, Qianqian Wang, Noah Snavely, and Angjoo Kanazawa. Infinitenature-zero: Learning perpetual view generation of natural scenes from single images. In European Conference on Computer Vision, pages 515-534. Springer, 2022.", + "[58] Jing Liao, Mark Finch, and Hugues Hoppe. Fast computation of seamless video loops. ACM Transactions on Graphics (TOG), 34(6):1-10, 2015.", + "[59] Zicheng Liao, Neel Joshi, and Hugues Hoppe. Automated video looping with progressive dynamism. ACM Transactions on Graphics (TOG), 32(4):1-10, 2013.", + "[60] Andrew Liu, Richard Tucker, Varun Jampani, Ameesh Makadia, Noah Snavely, and Angjoo Kanazawa. Infinite nature: Perpetual view generation of natural scenes from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14458-14467, 2021.", + "[61] Ce Liu. Beyond pixels: exploring new representations and applications for motion analysis. PhD thesis, Massachusetts Institute of Technology, 2009.", + "[62] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Tieniu Tan. Videofusion: Decomposed diffusion models for high-quality video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10209-10218, 2023.", + "[63] Aniruddha Mahapatra and Kuldeep Kulkarni. Controllable animation of fluid elements in still images. In Proc. Computer Vision and Pattern Recognition (CVPR), 2022.", + "[64] Arun Mallya, Ting-Chun Wang, and Ming-Yu Liu. Implicit warping for animation with image sets. Advances in Neural Information Processing Systems, 35:22438-22450, 2022.", + "[65] Medhini Narasimhan, Shiry Ginosar, Andrew Owens, Alexei A Efros, and Trevor Darrell. Strumming to the beat: Audio-conditioned contrastive video textures. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3761-3770, 2022.", + "[66] Haomiao Ni, Changhao Shi, Kai Li, Sharon X Huang, and Martin Renqiang Min. Conditional image-to-video generation with latent flow diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18444-18455, 2023.", + "[67] Simon Niklaus and Feng Liu. Softmax splatting for video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5437-5446, 2020.", + "[68] Shin Ota, Machiko Tamura, Kunihiko Fujita, T Fujimoto, K Muraoka, and Norishige Chiba. 1/f/sup/spl beta//noise-based real-time animation of trees swaying in wind fields. In Proceedings Computer Graphics International 2003, pages 52-59. IEEE, 2003.", + "[69] Automne Petitjean, Yohan Poirier-Ginter, Ayush Tewari, Guillaume Cordonnier, and George Drettakis. Modalnerf: Neural modal analysis and synthesis for free-viewpoint navigation in dynamically vibrating scenes. In Computer Graphics Forum, volume 42, 2023.", + "[70] Silvia L. Pintea, Jan C. van Gemert, and Arnold W. M." + ], + "bbox": [ + 86, + 92, + 470, + 890 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Smeulders. Déjà vu: Motion prediction in static images. In Proc. European Conf. on Computer Vision (ECCV), 2014.", + "[71] Sigal Raab, Inbal Leibovitch, Guy Tevet, Moab Arar, Amit H Bermano, and Daniel Cohen-Or. Single motion diffusion. arXiv preprint arXiv:2302.05905, 2023.", + "[72] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2):3, 2022.", + "[73] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022.", + "[74] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022.", + "[75] Payam Saisan, Gianfranco Doretto, Ying Nian Wu, and Stefano Soatto. Dynamic texture recognition. In Proceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition. CVPR 2001, volume 2, pages II-II. IEEE, 2001.", + "[76] Saurabh Saxena, Charles Herrmann, Junhwa Hur, Abhishek Kar, Mohammad Norouzi, Deqing Sun, and David J. Fleet. The surprising effectiveness of diffusion models for optical flow and monocular depth estimation, 2023.", + "[77] Arno Schödl, Richard Szeliski, David H Salesin, and Irfan Essa. Video textures. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 489-498, 2000.", + "[78] Mikio Shinya and Alain Fournier. Stochastic motion—motion under the influence of wind. Computer Graphics Forum, 11(3), 1992.", + "[79] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. Animating arbitrary objects via deep motion transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2377-2386, 2019.", + "[80] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. Advances in neural information processing systems, 32, 2019.", + "[81] Aliaksandr Siarohin, Oliver J Woodford, Jian Ren, Mengei Chai, and Sergey Tulyakov. Motion representations for articulated animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13653-13662, 2021.", + "[82] Chen Qian Kwan-Yee Lin Hongsheng Li Siming Fan, Jingtan Piao. Simulating fluids in real-world still images. arXiv preprint, arXiv:2204.11335, 2022.", + "[83] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3626-3636, 2022.", + "[84] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan," + ], + "bbox": [ + 509, + 92, + 893, + 890 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "24152", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015.", + "[85] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv:2010.02502, October 2020.", + "[86] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020.", + "[87] Jos Stam. Multi-scale stochastic modelling of complex natural phenomena. PhD thesis, 1995.", + "[88] Jos Stam. Stochastic dynamics: Simulating the effects of turbulence on flexible structures. Computer Graphics Forum, 16(3), 1997.", + "[89] Ryusuke Sugimoto, Mingming He, Jing Liao, and Pedro V Sander. Water simulation and rendering from a still photograph. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022.", + "[90] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022.", + "[91] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018.", + "[92] Vikram Voleti, Alexia Jolicoeur-Martineau, and Christopher Pal. Mcvd: Masked conditional video diffusion for prediction, generation, and interpolation. In (NeurIPS) Advances in Neural Information Processing Systems, 2022.", + "[93] Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba. Generating videos with scene dynamics. In Neural Information Processing Systems, 2016.", + "[94] Neal Wadhwa, Michael Rubinstein, Frédo Durand, and William T Freeman. Phase-based video motion processing. ACM Transactions on Graphics (ToG), 32(4):1-10, 2013.", + "[95] Jacob Walker, Carl Doersch, Abhinav Gupta, and Martial Hebert. An uncertain future: Forecasting from static images using variational autoencoders. In Proc. European Conf. on Computer Vision (ECCV), 2016.", + "[96] Jacob Walker, Abhinav Gupta, and Martial Hebert. Dense optical flow prediction from a static image. In Proceedings of the IEEE International Conference on Computer Vision, pages 2443-2451, 2015.", + "[97] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. arXiv preprint arXiv:2306.02018, 2023.", + "[98] Yaohui Wang, Di Yang, Francois Bremond, and Antitza Dantcheva. Latent image animator: Learning to animate images via latent space navigation. arXiv preprint arXiv:2203.09043, 2022.", + "[99] Frederik Warburg, Ethan Weber, Matthew Tancik, Aleksander Holynski, and Angjoo Kanazawa. Nerfbusters: Removing ghostly artifacts from casually captured nerfs. arXiv preprint arXiv:2304.10532, 2023.", + "[100] Daniel Watson, William Chan, Ricardo Martin-Brualla," + ], + "bbox": [ + 78, + 92, + 470, + 888 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. arXiv preprint arXiv:2210.04628, 2022.", + "[101] Chung-Yi Weng, Brian Curless, and Ira Kemelmacher-Shlizerman. Photo wake-up: 3d character animation from a single photo. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5908-5917, 2019.", + "[102] Jamie Wynn and Daniyar Turmukhambetov. DiffusioNeRF: Regularizing Neural Radiance Fields with Denoising Diffusion Models. In CVPR, 2023.", + "[103] Tianfan Xue, Jiajun Wu, Katherine L Bouman, and William T Freeman. Visual dynamics: Stochastic future generation via layered cross convolutional networks. Trans. Pattern Analysis and Machine Intelligence, 41(9):2236-2250, 2019.", + "[104] Shengming Yin, Chenfei Wu, Jian Liang, Jie Shi, Houqiang Li, Gong Ming, and Nan Duan. Dragnuwa: Fine-grained control in video generation by integrating text, image, and trajectory. arXiv preprint arXiv:2308.08089, 2023.", + "[105] Sihyun Yu, Kihyuk Sohn, Subin Kim, and Jinwoo Shin. Video probabilistic diffusion models in projected latent space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18456-18466, 2023.", + "[106] Mingyuan Zhang, Xinying Guo, Liang Pan, Zhongang Cai, Fangzhou Hong, Huirong Li, Lei Yang, and Ziwei Liu. Remodiffuse: Retrieval-augmented motion diffusion model. arXiv preprint arXiv:2304.01116, 2023.", + "[107] Yabo Zhang, Yuxiang Wei, Dongsheng Jiang, Xiaopeng Zhang, Wangmeng Zuo, and Qi Tian. Controlvideo: Training-free controllable text-to-video generation. arXiv preprint arXiv:2305.13077, 2023.", + "[108] Jian Zhao and Hui Zhang. Thin-plate spline motion model for image animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3657–3666, 2022.", + "[109] Shengyu Zhao, Jonathan Cui, Yilun Sheng, Yue Dong, Xiao Liang, Eric I Chang, and Yan Xu. Large scale image completion via co-modulated generative adversarial networks. In International Conference on Learning Representations (ICLR), 2021.", + "[110] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022." + ], + "bbox": [ + 503, + 92, + 893, + 715 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "24153", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2024/Generative Image Dynamics/34d8fada-6dfb-4124-8738-6ec5f099992d_model.json b/2024/Generative Image Dynamics/34d8fada-6dfb-4124-8738-6ec5f099992d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..43fdd675a4c1741e091cc9f90e15dca7f0ed6378 --- /dev/null +++ b/2024/Generative Image Dynamics/34d8fada-6dfb-4124-8738-6ec5f099992d_model.json @@ -0,0 +1,3073 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.342, + 0.131, + 0.63, + 0.154 + ], + "angle": 0, + "content": "Generative Image Dynamics" + }, + { + "type": "text", + "bbox": [ + 0.175, + 0.181, + 0.269, + 0.201 + ], + "angle": 0, + "content": "Zhengqi Li" + }, + { + "type": "text", + "bbox": [ + 0.307, + 0.181, + 0.433, + 0.199 + ], + "angle": 0, + "content": "Richard Tucker" + }, + { + "type": "text", + "bbox": [ + 0.472, + 0.182, + 0.585, + 0.199 + ], + "angle": 0, + "content": "Noah Snavely" + }, + { + "type": "text", + "bbox": [ + 0.626, + 0.182, + 0.794, + 0.199 + ], + "angle": 0, + "content": "Aleksander Holynski" + }, + { + "type": "text", + "bbox": [ + 0.399, + 0.207, + 0.537, + 0.224 + ], + "angle": 0, + "content": "Google Research" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.326, + 0.331, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.173, + 0.429, + 0.255, + 0.44 + ], + "angle": 0, + "content": "Input Picture" + }, + { + "type": "image", + "bbox": [ + 0.334, + 0.366, + 0.374, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.435, + 0.279, + 0.568, + 0.305 + ], + "angle": 0, + "content": "Spectral Volume (Ge-Space Modal Basis)" + }, + { + "type": "image", + "bbox": [ + 0.376, + 0.306, + 0.483, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.486, + 0.306, + 0.594, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.357, + 0.483, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.486, + 0.356, + 0.594, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.423, + 0.411, + 0.432, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.411, + 0.543, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.376, + 0.434, + 0.483, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.396, + 0.484, + 0.463, + 0.493 + ], + "angle": 0, + "content": "X coefficients" + }, + { + "type": "image", + "bbox": [ + 0.486, + 0.434, + 0.594, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.484, + 0.573, + 0.493 + ], + "angle": 0, + "content": "Y coefficients" + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.278, + 0.868, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.385, + 0.868, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.493, + 0.894, + 0.551 + ], + "angle": 0, + "content": "Figure 1. We model a generative image-space prior on scene motion: from a single RGB image, our method generates a spectral volume [23], a motion representation that models dense, long-term pixel trajectories in the Fourier domain. Our learned motion priors can be used to turn a single picture into a seamlessly looping video, or into an interactive simulation of dynamics that responds to user inputs like dragging and releasing points. On the right, we visualize output videos as space-time \\( X - t \\) slices (along the input scanline shown on the left)." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.568, + 0.314, + 0.584 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.613, + 0.474, + 0.886 + ], + "angle": 0, + "content": "We present an approach to modeling an image-space prior on scene motion. Our prior is learned from a collection of motion trajectories extracted from real video sequences depicting natural, oscillatory dynamics of objects such as trees, flowers, candles, and clothes swaying in the wind. We model dense, long-term motion in the Fourier domain as spectral volumes, which we find are well-suited to prediction with diffusion models. Given a single image, our trained model uses a frequency-coordinated diffusion sampling process to predict a spectral volume, which can be converted into a motion texture that spans an entire video. Along with an image-based rendering module, the predicted motion representation can be used for a number of downstream applications, such as turning still images into seamlessly looping videos, or allowing users to interact with objects in real images, producing realistic simulated dynamics (by interpreting the spectral volumes as image-space modal bases). See our project page for more results: generative-dynamics.github.io." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.568, + 0.631, + 0.584 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.596, + 0.895, + 0.687 + ], + "angle": 0, + "content": "The natural world is always in motion, with even seemingly static scenes containing subtle oscillations as a result of wind, water currents, respiration, or other natural rhythms. Emulating this motion is crucial in visual content synthesis—human sensitivity to motion can cause imagery without motion (or with slightly unrealistic motion) to seem uncanny or unreal." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.896, + 0.84 + ], + "angle": 0, + "content": "While it is easy for humans to interpret or imagine motion in scenes, training a model to learn or produce realistic scene motion is far from trivial. The motion we observe in the world is the result of a scene's underlying physical dynamics, i.e., forces applied to objects that respond according to their unique physical properties—their mass, elasticity, etc—quantities that are hard to measure and capture at scale. Fortunately, measuring them is not necessary for certain applications: e.g., one can simulate plausible dynamics in a scene by simply analyzing some observed 2D motion [23]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.842, + 0.895, + 0.903 + ], + "angle": 0, + "content": "This same observed motion can also serve as a supervisory signal in learning dynamics across scenes—because although observed motion is multi-modal and grounded in complex physical effects, it is nevertheless often predictable:" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "24142" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.212 + ], + "angle": 0, + "content": "candles will flicker in certain ways, trees will sway, and their leaves will rustle. As humans, this predictability is ingrained in our systems of perception: by viewing a still image, we can imagine plausible motions—or, since there might have been many possible such motions, a distribution of natural motions conditioned on that image. Given the facility with which humans are able to model these distributions, a natural research problem is to model them computationally." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.215, + 0.471, + 0.35 + ], + "angle": 0, + "content": "Recent advances in generative models, in particular conditional diffusion models [44, 84, 86], have enabled us to model rich distributions, including distributions of real images conditioned on text [72-74]. This capability has enabled several new applications, such as text-conditioned generation of diverse and realistic image content. Following the success of these image models, recent work has extended these models to other domains, such as videos [7, 43] and 3D geometry [76, 99, 100, 102]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.353, + 0.471, + 0.625 + ], + "angle": 0, + "content": "In this paper, we model a generative prior for image-space scene motion, i.e., the motion of all pixels in a single image. This model is trained on motion trajectories automatically extracted from a large collection of real video sequences. In particular, from each training video we compute motion in the form of a spectral volume [22, 23], a frequency-domain representation of dense, long-range pixel trajectories. Spectral volumes are well-suited to scenes that exhibit oscillatory dynamics, e.g., trees and flowers moving in the wind. We find that this representation is also highly effective as an output of a diffusion model for modeling scene motion. We train a generative model that, conditioned on a single image, can sample spectral volumes from its learned distribution. A predicted spectral volume can then be directly transformed into a motion texture—a set of long-range, per-pixel motion trajectories—that can be used to animate the image. The spectral volume can also be interpreted as an image-space modal basis for use in simulating interactive dynamics [22]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.627, + 0.47, + 0.734 + ], + "angle": 0, + "content": "We predict spectral volumes from input images using a diffusion model that generates coefficients one frequency at a time, but coordinates these predictions across frequency bands through a shared attention module. The predicted motions can be used to synthesize future frames (via an image-based rendering model)—turning still images into realistic animations, as illustrated in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Compared with priors over raw RGB pixels, priors over motion capture more fundamental, lower-dimensional structure that efficiently explains long-range variations in pixel values. Hence, generating intermediate motion leads to more coherent long-term generation and more fine-grained control over animations. We demonstrate the use of our trained model in several downstream applications, such as creating seamless looping videos, editing the generated motions, and enabling interactive dynamic images via image-space modal bases, i.e., simulating the response of object dynamics to user-applied forces [22]." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.09, + 0.642, + 0.107 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.116, + 0.895, + 0.283 + ], + "angle": 0, + "content": "Generative synthesis. Recent advances in generative models have enabled photorealistic synthesis of images conditioned on text prompts [16, 17, 24, 72-74]. These text-to-image models can be augmented to synthesize video sequences by extending the generated image tensors along a time dimension [7, 9, 43, 62, 83, 105, 105, 110]. While these methods can produce video sequences that capture the spatiotemporal statistics of real footage, these videos often suffer from artifacts like incoherent motion, unrealistic temporal variation in textures, and violations of physical constraints like preservation of mass." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.289, + 0.897, + 0.577 + ], + "angle": 0, + "content": "Animating images. Instead of generating videos entirely from text, other techniques take as input a still picture and animate it. Many recent deep learning methods adopt a 3D-Unet architecture to produce video volumes directly [27, 36, 40, 47, 53, 92]. These models are effectively the same video generation models (but conditioned on image information instead of text), and exhibit similar artifacts to those mentioned above. One way to overcome these limitations is to not directly generate the video content itself, but instead animate an input source image through image-based rendering, i.e., moving the image content around according to motion derived from external sources such as a driving video [51, 79-81, 98], motion or 3D geometry priors [8, 29, 46, 63, 64, 66, 89, 96, 100, 101, 103, 108], or user annotations [6, 18, 20, 33, 38, 97, 104, 107]. Animating images according to motion fields yields greater temporal coherence and realism, but these prior methods either require additional guidance signals or user input, or utilize limited motion representations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.895, + 0.84 + ], + "angle": 0, + "content": "Motion models and motion priors. In computer graphics, natural, oscillatory 3D motion (e.g., water rippling or trees waving in the wind) can be modeled with noise that is shaped in the Fourier domain and then converted to time-domain motion fields [78, 87]. Some of these methods rely on a modal analysis of the underlying dynamics of the system being simulated [22, 25, 88]. These spectral techniques were adapted to animate plants, water, and clouds from single 2D pictures by Chuang et al. [20], given user annotations. Our work is especially inspired by Davis [23], who connected modal analysis of a scene with the motions observed in a video of that scene, and used this analysis to simulate interactive dynamics from a video. We adopt the frequency-space spectral volume motion representation from Davis et al., extract this representation from a large set of training videos, and show that spectral volumes are suitable for predicting motion from single images with diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Other methods have used various motion representations in prediction tasks, where an image or video is used to inform a deterministic future motion estimate [34, 70], or a more rich distribution of possible motions [93, 95, 103]. However," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24143" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.213 + ], + "angle": 0, + "content": "many of these methods predict an optical flow motion estimate (i.e., the instantaneous motion of each pixel), not full motion trajectories. In addition, much of this prior work is focused on tasks like activity recognition, not on synthesis tasks. More recent work has demonstrated the advantages of modeling and predicting motion using generative models in a number of closed-domain settings such as humans and animals [2, 19, 28, 71, 90, 106]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.224, + 0.473, + 0.452 + ], + "angle": 0, + "content": "Videos as textures. Certain moving scenes can be thought of as a kind of texture—termed dynamic textures [26]—that model videos as space-time samples of a stochastic process. Dynamic textures can represent smooth, natural motions like waves, flames, or moving trees, and have been widely used for video classification, segmentation or encoding [12-15, 75]. A related kind of texture, called a video texture, represents a moving scene as a set of input video frames along with transition probabilities between any pair of frames [65, 77]. A number of methods estimate dynamic or video textures through analysis of scene motion and pixel statistics, with the aim of generating seamlessly looping or infinitely varying output videos [1, 21, 32, 58, 59, 77]. In contrast to much of this work, our method learns priors in advance that can then be applied to single images." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.469, + 0.183, + 0.485 + ], + "angle": 0, + "content": "3. Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.497, + 0.47, + 0.663 + ], + "angle": 0, + "content": "Given a single picture \\( I_0 \\), our goal is to generate a video \\( \\{\\hat{I}_1, \\hat{I}_2, \\dots, \\hat{I}_T\\} \\) featuring oscillatory motions such as those of trees, flowers, or candle flames swaying in the breeze. Our system consists of two modules: a motion prediction module and an image-based rendering module. Our pipeline begins by using a latent diffusion model (LDM) to predict a spectral volume \\( S = (S_{f_0}, S_{f_1}, \\dots, S_{f_{K-1}}) \\) for the input \\( I_0 \\). The predicted spectral volume is then transformed to a motion texture \\( \\mathcal{F} = (F_1, F_2, \\dots, F_T) \\) through an inverse discrete Fourier transform. This motion determines the position of each input pixel at every future time step." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.665, + 0.471, + 0.741 + ], + "angle": 0, + "content": "Given a predicted motion texture, we then animate the input RGB image using a neural image-based rendering technique (Sec. 5). We explore applications of this method, including producing seamless looping animations and simulating interactive dynamics, in Sec. 6." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.759, + 0.253, + 0.777 + ], + "angle": 0, + "content": "4. Predicting motion" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.785, + 0.285, + 0.802 + ], + "angle": 0, + "content": "4.1. Motion representation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Formally, a motion texture is a sequence of time-varying 2D displacement maps \\(\\mathcal{F} = \\{F_t|t = 1,\\dots,T\\}\\), where the 2D displacement vector \\(F_{t}(\\mathbf{p})\\) at each pixel coordinate \\(\\mathbf{p}\\) from input image \\(I_0\\) defines the position of that pixel at a future time \\(t\\) [20]. To generate a future frame at time \\(t\\), one can sput pixels from \\(I_0\\) using the corresponding displacement" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.09, + 0.698, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.09, + 0.891, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.191, + 0.895, + 0.317 + ], + "angle": 0, + "content": "Figure 2. Left: We visualize the average power spectrum for the \\(X\\) and \\(Y\\) motion components extracted from real videos, shown as the blue and green curves. Natural oscillation motions are composed primarily of low-frequency components, and so we use the first \\(K = 16\\) terms, marked with red dots. Right: we show a histogram of the amplitude of Fourier terms at \\(3.0\\mathrm{Hz}\\) after (1) scaling amplitude by image width and height (blue), or (2) frequency adaptive normalization (red). Our adaptive normalization prevents the coefficients from concentrating at extreme values." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.339, + 0.825, + 0.355 + ], + "angle": 0, + "content": "map \\(D_{t}\\), resulting in a forward-warped image \\(I_t'\\):" + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.358, + 0.892, + 0.375 + ], + "angle": 0, + "content": "\\[\nI _ {t} ^ {\\prime} (\\mathbf {p} + F _ {t} (\\mathbf {p})) = I _ {0} (\\mathbf {p}). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.379, + 0.895, + 0.544 + ], + "angle": 0, + "content": "If our goal is to produce a video via a motion texture, then one choice would be to predict a time-domain motion texture directly from an input image. However, the size of the motion texture would need to scale with the length of the video: generating \\( T \\) output frames implies predicting \\( T \\) displacement fields. To avoid predicting such a large output representation for long videos, many prior animation methods either generate video frames autoregressively [7, 29, 57, 60, 92], or predict each future output frame independently via an extra time embedding [4]. However, neither strategy ensures long-term temporal consistency of generated videos." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.545, + 0.895, + 0.695 + ], + "angle": 0, + "content": "Fortunately, many natural motions can be described as a superposition of a small number of harmonic oscillators represented with different frequencies, amplitude and phases [20, 23, 25, 50, 68]. Because these underlying motions are quasi-periodic, it is natural to model them in the frequency domain. Hence, we adopt from Davis et al. [23] an efficient frequency space representation of motion in a video called a spectral volume, visualized in Fig. 3. A spectral volume is the temporal Fourier transform of per-pixel trajectories extracted from a video." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.696, + 0.895, + 0.878 + ], + "angle": 0, + "content": "Given this motion representation, we formulate the motion prediction problem as a multi-modal image-to-image translation task: from an input image to an output motion spectral volume. We adopt latent diffusion models (LDMs) to generate spectral volumes comprised of a \\(4K\\)-channel 2D motion spectrum map, where \\(K \\ll T\\) is the number of frequencies modeled, and where at each frequency we need four scalars to represent the complex Fourier coefficients for the \\(x\\)- and \\(y\\)-dimensions. Note that the motion trajectory of a pixel at future time steps \\(\\mathcal{F}(\\mathbf{p}) = \\{F_t(\\mathbf{p})|t = 1,2,\\dots T\\}\\) and its representation as a spectral volume \\(\\mathcal{S}(\\mathbf{p}) = \\{S_{f_k}(\\mathbf{p})|k = 0,1,\\dots \\frac{T}{2} -1\\}\\) are related by the Fast Fourier transform (FFT):" + }, + { + "type": "equation", + "bbox": [ + 0.624, + 0.881, + 0.891, + 0.898 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} (\\mathbf {p}) = \\operatorname {F F T} (\\mathcal {F} (\\mathbf {p})). \\tag {2}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24144" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.09, + 0.873, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.295, + 0.895, + 0.355 + ], + "angle": 0, + "content": "Figure 3. Motion prediction module. We predict a spectral volume \\(S\\) through a frequency-coordinated denoising model. Each block of the diffusion network \\(\\epsilon_{\\theta}\\) interleaves 2D spatial layers with attention layers (red box, right), and iteratively denoises latent features \\(z^n\\). The denoised features are fed to a decoder \\(\\mathcal{D}\\) to produce \\(S\\). During training, we concatenate the downsampled input \\(I_0\\) with noisy latent features encoded from a real motion texture via an encoder \\(\\mathcal{E}\\), and replace the noisy features with Gaussian noise \\(z^N\\) during inference (left)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.377, + 0.475, + 0.573 + ], + "angle": 0, + "content": "How should we select the \\( K \\) output frequencies? Prior work in real-time animation has observed that most natural oscillation motions are composed primarily of low-frequency components [25, 68]. To validate this observation, we computed the average power spectrum of the motion extracted from 1,000 randomly sampled 5-second real video clips. As shown in the left plot of Fig. 2, the power spectrum of the motion decreases exponentially with increasing frequency. This suggests that most natural oscillation motions can indeed be well represented by low-frequency terms. In practice, we found that the first \\( K = 16 \\) Fourier coefficients are sufficient to realistically reproduce the original natural motion in a range of real videos and scenes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.588, + 0.429, + 0.604 + ], + "angle": 0, + "content": "4.2. Predicting motion with a diffusion model" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.612, + 0.474, + 0.869 + ], + "angle": 0, + "content": "We choose a latent diffusion model (LDM) [73] as the backbone for our motion prediction module, as LDMs are more computationally efficient than pixel-space diffusion models, while preserving synthesis quality. A standard LDM consists of two main modules: (1) a variational autoencoder (VAE) that compresses the input image to a latent space through an encoder \\( z = E(I) \\), then reconstructs the input from the latent features via a decoder \\( I = D(z) \\), and (2) a U-Net based diffusion model that learns to iteratively denoise features starting from Gaussian noise. Our training applies this process not to RGB images but to spectral volumes from real video sequences, which are encoded and then diffused for \\( n \\) steps with a pre-defined variance schedule to produce noisy latents \\( z^n \\). The 2D U-Nets are trained to denoise the noisy latents by iteratively estimating the noise \\( \\epsilon_{\\theta}(z^{n};n,c) \\) used to update the latent feature at each step \\( n\\in (1,2,\\dots,N) \\). The training loss for the LDM is written as" + }, + { + "type": "equation", + "bbox": [ + 0.089, + 0.88, + 0.47, + 0.9 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {L D M}} = \\mathbb {E} _ {n \\in \\mathcal {U} [ 1, N ], \\epsilon^ {n} \\in \\mathcal {N} (0, 1)} \\left[ \\left| \\left| \\epsilon^ {n} - \\epsilon_ {\\theta} \\left(z ^ {n}; n, c\\right) \\right| \\right| ^ {2} \\right] \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.377, + 0.895, + 0.439 + ], + "angle": 0, + "content": "where \\( c \\) is the embedding of any conditional signal, such as text, or, in our case, the first frame of the training video sequence, \\( I_0 \\). The clean latent features \\( z^0 \\) are then passed through the decoder to recover the spectral volume." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.472, + 0.895, + 0.715 + ], + "angle": 0, + "content": "Frequency adaptive normalization. One issue we observed is that motion textures have particular distribution characteristics across frequencies. As visualized in the left plot of Fig. 2, the amplitude of the spectral volumes spans a range of 0 to 100 and decays approximately exponentially with increasing frequency. As diffusion models require that the absolute values of the output are between -1 and 1 for stable training and denoising [44], we must normalize the coefficients of \\( S \\) extracted from real videos before using them for training. If we scale the magnitudes of these coefficients to [0,1] based on the image dimensions as in prior work [29, 76], nearly all the coefficients at higher frequencies will end up close to zero, as shown in the right plot of Fig. 2. Models trained on such data can produce inaccurate motions, since during inference, even small prediction errors can cause large relative errors after denormalization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.897, + 0.903 + ], + "angle": 0, + "content": "To address this issue, we employ a simple but effective frequency adaptive normalization method: First, we independently normalize Fourier coefficients at each frequency based on statistics computed from the training set. Namely, for each individual frequency \\( f_{j} \\), we compute the \\( 95^{\\mathrm{th}} \\) percentile of Fourier coefficient magnitudes over all input samples and use that value as a per-frequency scaling factor \\( s f_{j} \\). We then apply a power transformation to each scaled Fourier coefficient to pull it away from extreme values. In practice, we observe that a square root performs better than other nonlinear transformations such as log or reciprocal. In summary, the final coefficient values of spectral volume \\( S(\\mathbf{p}) \\) at" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24145" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.467, + 0.108 + ], + "angle": 0, + "content": "frequency \\( f_{j} \\) (used for training our LDM) are computed as" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.115, + 0.47, + 0.155 + ], + "angle": 0, + "content": "\\[\nS _ {f _ {j}} ^ {\\prime} (\\mathbf {p}) = \\operatorname {s i g n} \\left(S _ {f _ {j}}\\right) \\sqrt {\\left| \\frac {S _ {f _ {j}} (\\mathbf {p})}{s _ {f _ {j}}} \\right|}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.162, + 0.47, + 0.207 + ], + "angle": 0, + "content": "As shown on the right plot of Fig. 2, after applying frequency adaptive normalization, the spectral volume coefficients distribute more evenly." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.211, + 0.471, + 0.362 + ], + "angle": 0, + "content": "Frequency-coordinated denoising. The straightforward way to predict a spectral volume \\(S\\) with \\(K\\) frequency bands is to output a tensor of \\(4K\\) channels from a single diffusion U-Net. However, as in prior work [7], we observe that training a model to produce a large number of channels can yield over-smoothed, inaccurate outputs. An alternative would be to independently predict each individual frequency slice by injecting an extra frequency embedding into the LDM [4], but this design choice would result in uncorrelated predictions in the frequency domain, leading to unrealistic motion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.363, + 0.473, + 0.726 + ], + "angle": 0, + "content": "Therefore, inspired by recent video diffusion work [7], we propose a frequency-coordinated denoising strategy, illustrated in Fig. 3. In particular, given an input image \\(I_0\\), we first train an LDM \\(\\epsilon_{\\theta}\\) to predict a single 4-channel frequency slice of spectral volume \\(S_{f_j}\\), where we inject an extra frequency embedding along with the time-step embedding into the LDM. We then freeze the parameters of this LDM \\(\\epsilon_{\\theta}\\), introduce attention layers interleaved with the 2D spatial layers of \\(\\epsilon_{\\theta}\\) across the \\(K\\) frequency bands, and fine-tune. Specifically, for a batch size \\(B\\), the 2D spatial layers of \\(\\epsilon_{\\theta}\\) treat the corresponding \\(B \\cdot K\\) noisy latent features of channel size \\(C\\) as independent samples with shape \\(\\mathcal{R}^{(B \\cdot K) \\times C \\times H \\times W}\\). The attention layer then interprets these as consecutive features spanning the frequency axis, and we reshape the latent features from previous 2D spatial layers to \\(\\mathcal{R}^{B \\times K \\times C \\times H \\times W}\\) before feeding them to the attention layers. In other words, the frequency attention layers are fine-tuned to coordinate all frequency slices so as to produce coherent spectral volumes. In our experiments, we see that the average VAE reconstruction error improves from 0.024 to 0.018 when we switch from a single 2D U-Net to a frequency-coordinated denoising module, suggesting an improved upper bound on LDM prediction accuracy; in Sec. 7.3, we also show that this design choice improves video generation quality." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.74, + 0.296, + 0.758 + ], + "angle": 0, + "content": "5. Image-based rendering" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.902 + ], + "angle": 0, + "content": "We now describe how we take a spectral volume \\(S\\) predicted for a given input image \\(I_0\\) and render a future frame \\(\\hat{I}_t\\) at time \\(t\\). We first derive a motion texture in the time domain using the inverse temporal FFT applied at each pixel \\(\\mathcal{F}(\\mathbf{p}) = \\mathrm{FFT}^{-1}(S(\\mathbf{p}))\\). To produce a future frame \\(\\hat{I}_t\\), we adopt a deep image-based rendering technique and perform splatting with the predicted motion field \\(F_t\\) to forward-warp the encoded \\(I_0\\), as shown in Fig. 4. Since forward warping can lead to holes, and multiple source pixels can map to the same output" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.093, + 0.885, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.271, + 0.894, + 0.37 + ], + "angle": 0, + "content": "Figure 4. Rendering module. We fill in missing content and refine the warped input image using a deep image-based rendering module, where multi-scale features are extracted from the input image \\( I_0 \\). Softmax splatting is then applied over the features with a motion field \\( F_t \\) from time 0 to \\( t \\) (subject to the weights \\( W \\)). The warped features are fed to an image synthesis network to produce the rendered image \\( \\hat{I}_t \\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.394, + 0.894, + 0.424 + ], + "angle": 0, + "content": "2D location, we adopt the feature pyramid softmax splatting strategy proposed in prior work on frame interpolation [67]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.425, + 0.893, + 0.65 + ], + "angle": 0, + "content": "Specifically, we encode \\( I_0 \\) through a feature extractor network to produce a multi-scale feature map. For each individual feature map at scale \\( j \\), we resize and scale the predicted 2D motion field \\( F_{t} \\) according to the resolution. As in Davis et al. [22], we use predicted flow magnitude as a proxy for depth to determine the contributing weight of each source pixel mapped to its destination location. In particular, we compute a per-pixel weight, \\( W(\\mathbf{p}) = \\frac{1}{T}\\sum_{t}||F_{t}(\\mathbf{p})||_{2}^{2} \\) as the average magnitude of the predicted motion texture. In other words, we assume large motions correspond to moving foreground objects, and small or zero motions correspond to background. We use motion-derived weights instead of learnable ones as [46] because we observe that in the single-view case, learnable weights are not effective for addressing disocclusion ambiguities." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.651, + 0.894, + 0.726 + ], + "angle": 0, + "content": "With the motion field \\( F_{t} \\) and weights \\( W \\), we apply softmax splatting to warp the feature map at each scale to produce a warped feature. The warped features are then injected into the corresponding blocks of an image synthesis decoder to produce a final rendered image \\( \\hat{I}_{t} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.727, + 0.894, + 0.803 + ], + "angle": 0, + "content": "We jointly train the feature extractor and synthesis networks with start and target frames \\((I_0, I_t)\\) randomly sampled from real videos, using the estimated flow field from \\(I_0\\) to \\(I_t\\) to warp encoded features from \\(I_0\\), and supervising predictions \\(\\hat{I}_t\\) against \\(I_t\\) with a VGG perceptual loss [49]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.816, + 0.631, + 0.833 + ], + "angle": 0, + "content": "6. Applications" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Image-to-video. Our system enables the animation of a single still picture by first predicting a motion spectral volume from the input image and generating an animation by applying our image-based rendering module to the motion" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24146" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.168 + ], + "angle": 0, + "content": "texture transformed from the spectral volume. Since we explicitly model scene motion, this allows us to produce slow-motion videos by linearly interpolating the motion texture, or to magnify (or minify) animated motions by adjusting the amplitude of the predicted spectral volume coefficients." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.169, + 0.472, + 0.395 + ], + "angle": 0, + "content": "Seamless looping. Many applications require videos that loop seamlessly, where there is no discontinuity between the start and end of the video. Unfortunately, it is hard to find a large collection of seamlessly looping videos for training. Instead, we devise a method to use our motion diffusion model, trained on regular non-looping video clips, to produce seamless looping video. Inspired by recent work on guidance for image editing [3, 30], our method is a motion self-guidance technique that guides the motion denoising sampling processing using explicit looping constraints. In particular, at each iterative denoising step during inference, we incorporate an additional motion guidance signal alongside standard classifier-free guidance [45], where we enforce each pixel's position and velocity at the start and end frames to be as similar as possible:" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.404, + 0.468, + 0.422 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\epsilon} ^ {n} = (1 + w) \\epsilon_ {\\theta} \\left(z ^ {n}; n, c\\right) - w \\epsilon_ {\\theta} \\left(z ^ {n}; n, \\emptyset\\right) + u \\sigma^ {n} \\nabla_ {z ^ {n}} \\mathcal {L} _ {g} ^ {n}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.424, + 0.469, + 0.443 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {g} ^ {n} = \\left\\| F _ {T} ^ {n} - F _ {1} ^ {n} \\right\\| _ {1} + \\left\\| \\nabla F _ {T} ^ {n} - \\nabla F _ {1} ^ {n} \\right\\| _ {1} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.446, + 0.47, + 0.567 + ], + "angle": 0, + "content": "where \\( F_{t}^{n} \\) is the predicted 2D displacement field at time \\( t \\) and denoising step \\( n \\). \\( w \\) is the classifier-free guidance weight, and \\( u \\) is the motion self-guidance weight. In the supplemental video, we apply a baseline appearance-based looping algorithm [58] to generate a looping video from our non-looping output, and show that our motion self-guidance technique produces seamless looping videos with less distortion and fewer artifacts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.57, + 0.471, + 0.752 + ], + "angle": 0, + "content": "Interactive dynamics from a single image. Davis et al. [22] show that the spectral volume, evaluated at certain resonant frequencies, can approximate an image-space modal basis that is a projection of the vibration modes of the underlying scene (or, more generally, captures spatial and temporal correlations in oscillatory dynamics), and can be used to simulate the object's response to a user-defined force. We adopt this modal analysis method [22, 69], allowing us to write the image-space 2D motion displacement field for the object's physical response as a weighted sum of motion spectrum coefficients \\( S_{f_j} \\) modulated by the state of complex modal coordinates \\( \\mathbf{q}_{f_j}(t) \\) at each simulated time step \\( t \\):" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.757, + 0.469, + 0.79 + ], + "angle": 0, + "content": "\\[\nF _ {t} (\\mathbf {p}) = \\sum_ {f _ {j}} S _ {f _ {j}} (\\mathbf {p}) \\mathbf {q} _ {f _ {j}} (t) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.471, + 0.901 + ], + "angle": 0, + "content": "We simulate the state of the modal coordinates \\(\\mathbf{q}_{f_j}(t)\\) via an explicit Euler method applied to the equations of motion for a decoupled mass-spring-damper system represented in modal space [22, 23, 69]. We refer readers to supplementary material and original work for a full derivation. Note that our method produces an interactive scene from a single picture, whereas these prior methods required a video as input." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.091, + 0.899, + 0.226 + ], + "angle": 0, + "content": "
MethodImage SynthesisVideo Synthesis
FIDKIDFVDFVD32DTFVDDTFVD32
TATS [35]65.81.67265.6419.622.640.7
Stochastic I2V [27]68.33.12253.5320.916.741.7
MCVD [92]63.42.97208.6270.419.553.9
LFDM [66]47.61.70187.5254.313.045.6
DMVFN [48]37.91.09206.5316.311.254.5
Endo et al. [29]10.40.19166.0231.65.3565.1
Holynski et al. [46]11.20.20179.0253.77.2346.8
Ours4.030.0847.162.92.536.75
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.236, + 0.894, + 0.293 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons on the test set. We report both image synthesis and video synthesis quality. Here, KID is scaled by 100. Lower is better for all error. See Sec. 7.1 for descriptions of baselines and error metrics." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.33, + 0.633, + 0.347 + ], + "angle": 0, + "content": "7. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.355, + 0.895, + 0.61 + ], + "angle": 0, + "content": "Implementation details. We use an LDM [73] as the backbone for predicting spectral volumes, for which we use a VAE with a continuous latent space of dimension 4. We train the VAE with an \\(L_{1}\\) reconstruction loss, a multi-scale gradient consistency loss [54-56], and a KL-divergence loss with respective weights of \\(1,0.2,10^{-6}\\). We train the same 2D U-Net used in the original LDM work to perform iterative denoising with a simple MSE loss [44], and adopt the attention layers from [41] for frequency-coordinated denoising. For quantitative evaluation, we train both VAE and LDM on images of size \\(256 \\times 160\\) from scratch for fair comparisons, and it takes around 6 days to converge using 16 Nvidia A100 GPUs. For our main quantitative and qualitative results, we run the motion diffusion model with DDIM [85] for 250 steps. We also show generated videos of up to a resolution of \\(512 \\times 288\\), created by fine-tuning a pre-trained image inpainting LDM model [73] on our dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.894, + 0.732 + ], + "angle": 0, + "content": "We adopt ResNet-34 [39] for the feature extractor in our IBR module. Our image synthesis network is based on an architecture for conditional image inpainting [57, 109]. Our rendering module runs in real-time at 25FPS on a Nvidia V100 GPU during inference. We adopt universal guidance [3] to produce seamless looping videos, where we set weights \\( w = 1.75 \\), \\( u = 200 \\), and use 500 DDIM steps with 2 self-recurrence iterations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Data. We collect and process a set of 3,015 videos of natural scenes exhibiting oscillatory motions from online sources our own captures. We withhold \\(10\\%\\) of videos for testing and use the remainder for training. To extract ground truth motion trajectories, we apply a coarse-to-fine flow method [10, 61] between each selected starting image and every future frame of the video. As training data, we take every 10th video frame as input images and derive the corresponding ground truth spectral volumes using the computed motion trajectories across the following 149 frames. In total, our data consists of over 150K image-motion pairs." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24147" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.09, + 0.199, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.199, + 0.091, + 0.885, + 0.144 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.147, + 0.198, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.198, + 0.147, + 0.885, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.206, + 0.199, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.199, + 0.207, + 0.885, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.265, + 0.199, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.199, + 0.265, + 0.885, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.326, + 0.178, + 0.338 + ], + "angle": 0, + "content": "Input image" + }, + { + "type": "image_caption", + "bbox": [ + 0.229, + 0.326, + 0.286, + 0.337 + ], + "angle": 0, + "content": "Reference" + }, + { + "type": "image_caption", + "bbox": [ + 0.319, + 0.326, + 0.425, + 0.338 + ], + "angle": 0, + "content": "Stochastic-I2V [27]" + }, + { + "type": "image_caption", + "bbox": [ + 0.452, + 0.326, + 0.519, + 0.338 + ], + "angle": 0, + "content": "MCVD[92]" + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.326, + 0.641, + 0.338 + ], + "angle": 0, + "content": "Endo et al. [29]" + }, + { + "type": "image_caption", + "bbox": [ + 0.661, + 0.326, + 0.767, + 0.338 + ], + "angle": 0, + "content": "Holynski et al. [46]." + }, + { + "type": "image_caption", + "bbox": [ + 0.813, + 0.326, + 0.841, + 0.337 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.338, + 0.892, + 0.362 + ], + "angle": 0, + "content": "Figure 5. X - t slices of videos generated by different approaches. From left to right: input image and corresponding X - t video slices from the ground truth video, from videos generated by three baselines [27, 29, 46, 92], and finally videos generated by our approach." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.385, + 0.275, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.275, + 0.385, + 0.468, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.494, + 0.471, + 0.536 + ], + "angle": 0, + "content": "Figure 6. Sliding window FID and DTFVD. We show sliding window FID with window size 30 frames, and DTFVD with size 16 frames, for videos generated by different methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.561, + 0.47, + 0.743 + ], + "angle": 0, + "content": "Baselines. We compare our approach to recent single-image animation and video prediction methods. Endo et al. [29] and DMVFN [48] predict instantaneous 2D motion fields and render future frames auto-regressively. Holynski et al. [46] instead simulates motion through a single static Eulerian motion description. Other recent work such as Stochastic Image-to-Video (Stochastic-I2V) [27], TATS [35], and MCVD [92] adopt VAEs, transformers, or diffusion models to directly predict raw video frames; LFDM [66] generates future frames by predicting flow volumes and warping latents in a diffusion model. We train all the above methods on our data using their respective open-source implementations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.743, + 0.471, + 0.848 + ], + "angle": 0, + "content": "We evaluate the quality of the videos generated by our approach and by prior baselines in two ways. First, we evaluate the quality of individual synthesized frames using metrics designed for image synthesis tasks. We adopt the Fréchet Inception Distance (FID) [42] and Kernel Inception Distance (KID) [5] to measure the average distance between the distributions of generated frames and ground truth frames." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.849, + 0.471, + 0.879 + ], + "angle": 0, + "content": "Second, to evaluate the quality and temporal coherence of synthesized videos, we adopt the Fréchet Video Dis" + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.385, + 0.898, + 0.559 + ], + "angle": 0, + "content": "
MethodImage SynthesisVideo Synthesis
FIDKIDFVDFVD32DTFVDDTFVD32
Repeat I0--237.5316.75.3045.6
K=43.920.0760.378.43.128.59
K=83.950.0752.168.72.717.37
K=244.090.0848.265.12.506.94
w/o adaptive norm.4.530.0962.780.13.168.19
Independent pred.4.000.0852.571.32.707.40
Volume pred.4.740.0953.771.12.837.79
Baseline splat [46]4.250.0949.566.82.837.27
Full (K=16)4.030.0847.162.92.536.75
" + }, + { + "type": "table_caption", + "bbox": [ + 0.507, + 0.563, + 0.882, + 0.578 + ], + "angle": 0, + "content": "Table 2. Ablation study. Sec. 7.3 describes each configuration." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.619, + 0.894, + 0.756 + ], + "angle": 0, + "content": "tance [91] with window size 16 (FVD) and 32 \\((\\mathrm{FVD}_{32})\\) based on an I3D model [11] trained on the Human Kinetics datasets [52]. To more faithfully reflect synthesis quality for the natural oscillation motions we seek to generate, we also adopt the Dynamic Texture Frechet Video Distance [27], which measures distance from videos with window size 16 (DTFVD) and size 32 \\((\\mathrm{DTFVD}_{32})\\), using a I3D model trained on the Dynamic Textures Database [37], a dataset consisting primarily of natural motion textures." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.757, + 0.894, + 0.834 + ], + "angle": 0, + "content": "We further use a sliding window FID of window size 30 frames, and a sliding window DTFVD with window size 16 frames, as in [57, 60], to measure how generated video quality degrades over time. For all methods, we evaluate metrics at \\(256 \\times 128\\) resolution by center-cropping." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.846, + 0.689, + 0.862 + ], + "angle": 0, + "content": "7.1.Quantitative results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Table 1 shows quantitative comparisons between our approach and baselines on our test set. Our approach signifi-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.462, + 0.901 + ], + "angle": 0, + "content": "1We use the open-source implementation of [46] from Fan et al. [82]." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24148" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.472, + 0.149 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.155, + 0.47, + 0.182 + ], + "angle": 0, + "content": "Figure 7. We show generated future frames from three recent large video diffusion models [31, 36, 97]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.21, + 0.471, + 0.346 + ], + "angle": 0, + "content": "cantly outperforms prior single-image animation baselines in terms of both image and video synthesis quality. Specifically, our much lower FVD and DT-FVD distances suggest that the videos generated by our approach are more realistic and more temporally coherent. Further, Fig. 6 shows the sliding window FID and sliding window DT-FVD distances of generated videos from different methods. Thanks to the global spectral volume representation, videos generated by our approach do not suffer from degradation over time." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.358, + 0.254, + 0.373 + ], + "angle": 0, + "content": "7.2. Qualitative results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.382, + 0.472, + 0.577 + ], + "angle": 0, + "content": "We visualize qualitative comparisons between videos as spatio-temporal \\(X\\)-\\(t\\) slices of the generated videos, a standard way of visualizing small motions in a video [94]. As shown in Fig. 5, our generated video dynamics more strongly resemble the motion patterns observed in the corresponding real reference videos (second column), compared to other methods. Baselines such as Stochastic I2V [27] and MCVD [92] fail to model both appearance and motion realistically over time. Endo et al. [29] and Holynski et al. [46] produce video frames with fewer artifacts but exhibits oversmooth or non-oscillatory motion over time. We refer readers to supplementary material to assess the quality of generated video frames and estimated motion across different methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.59, + 0.227, + 0.606 + ], + "angle": 0, + "content": "7.3. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.471, + 0.901 + ], + "angle": 0, + "content": "We conduct an ablation study to validate the major design choices in our motion prediction and rendering modules, comparing our full configuration with different variants. Specifically, we evaluate results using different numbers of frequency bands \\( K = 4, 8, 16, 24 \\). We observe that increasing the number of frequency bands improves video prediction quality, but the improvement is marginal with more than 16 frequencies. Next, we remove adaptive frequency normalization from the ground truth spectral volumes, and instead just scale them based on input image width and height (w/o adaptive norm.). Additionally, we remove the frequency coordinated-denoising module (Independent pred.), or replace it with a simpler DM where a tensor volume of \\( 4K \\) channel spectral volumes are predicted jointly via a single 2D U-net diffusion model (Volume pred.). Finally, we compare results where we use a baseline rendering method that applies softmax splatting over single-scale features subject to learnable weights as used by [46] (Baseline splat). We also add a baseline where the generated video is a volume by" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.09, + 0.895, + 0.133 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.139, + 0.894, + 0.195 + ], + "angle": 0, + "content": "Figure 8. Limitations. We show examples of rendered future frames (even), and overlay of input and rendered images (odd). Our method can produce artifacts in regions of thin objects or large motions, and regions requiring filling large amount of new contents." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.221, + 0.893, + 0.266 + ], + "angle": 0, + "content": "repeating input image \\(N\\) times (Repeat \\(I_0\\)). From Table 2, we observe that all simpler or alternative configurations lead to worse performance compared with our full model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.275, + 0.791, + 0.291 + ], + "angle": 0, + "content": "7.4. Comparing to large video models" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.299, + 0.895, + 0.465 + ], + "angle": 0, + "content": "We further perform a user study and compare our generated animations with ones from recent large video diffusion models:AnimateDiff [36], ModelScope [97] and Gen-2 [31], which predict video volumes directly. On a randomly selected 30 videos from the test set, we ask users \"which video is more realistic?\". Users report a \\(80.9\\%\\) preference for our approach over others. Moreover, as shown in Fig. 7, we observe that the generated videos from these baselines are either unable to adhere to the input image content, or exhibit gradual color drift and distortion over time. We refer readers to the supplementary material for full comparisons." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.479, + 0.742, + 0.494 + ], + "angle": 0, + "content": "8. Discussion and conclusion" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.504, + 0.894, + 0.64 + ], + "angle": 0, + "content": "Limitations. Since our approach only predicts lower frequencies of a spectral volume, it can fail to model non-oscillating motions or high-frequency vibrations—this may be resolved by using learned motion bases. Furthermore, the quality of generated videos relies on the quality of underlying motion trajectories, which may degrade in scenes with thin moving objects or objects with large displacements. Even if correct, motions that require generating large amounts of new unseen content may also cause degradations (Fig. 8)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.642, + 0.895, + 0.853 + ], + "angle": 0, + "content": "Conclusion. We present a new approach for modeling natural oscillation dynamics from a single still picture. Our image-space motion prior is represented with spectral volumes, a frequency representation of per-pixel motion trajectories, which we find to be efficient and effective for prediction with diffusion models, and which we learn from collections of real world videos. Spectral volumes are predicted using frequency-coordinated latent diffusion model and are used to animate future video frames through an image-based rendering module. We show that our approach produces photo-realistic animations from a single picture and significantly outperforms prior baselines, and that it can enable several downstream applications such as creating seamlessly looping or interactive image dynamics." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.856, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. We thank Abe Davis, Rick Szeliski, Andrew Liu, Boyang Deng, Qianqian Wang, Xuan Luo, and Lucy Chai for fruitful discussions and helpful comments." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24149" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.116, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Aseem Agarwala, Ke Colin Zheng, Chris Pal, Maneesh Agrawala, Michael Cohen, Brian Curless, David Salesin, and Richard Szeliski. Panoramic video textures. In ACM SIGGRAPH 2005 Papers, pages 821-827. 2005." + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.17, + 0.471, + 0.21 + ], + "angle": 0, + "content": "[2] Hyemin Ahn, Esteve Valls Mascaro, and Dongheui Lee. Can we use diffusion probabilistic models for 3d motion prediction? arXiv preprint arXiv:2302.14503, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.21, + 0.471, + 0.278 + ], + "angle": 0, + "content": "[3] Arpit Bansal, Hong-Min Chu, Avi Schwarzschild, Soumyadip Sengupta, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Universal guidance for diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 843-852, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.278, + 0.471, + 0.359 + ], + "angle": 0, + "content": "[4] Hugo Bertiche, Niloy J Mitra, Kuldeep Kulkarni, ChunHao P Huang, Tuanfeng Y Wang, Meysam Madadi, Sergio Escalera, and Duygu Ceylan. Blowing in the wind: Cyclenet for human cinemagrams from still images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 459-468, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.36, + 0.471, + 0.4 + ], + "angle": 0, + "content": "[5] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying MMD GANs. arXiv preprint arXiv:1801.01401, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.4, + 0.471, + 0.467 + ], + "angle": 0, + "content": "[6] Andreas Blattmann, Timo Milbich, Michael Dorkenwald, and Björn Ommer. ipoke: Poking a still image for controlled stochastic video synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14707-14717, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.468, + 0.471, + 0.55 + ], + "angle": 0, + "content": "[7] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.55, + 0.471, + 0.605 + ], + "angle": 0, + "content": "[8] Richard Strong Bowen, Richard Tucker, Ramin Zabih, and Noah Snavely. Dimensions of motion: Monocular prediction through flow subspaces. In 2022 International Conference on 3D Vision (3DV), pages 454-464. IEEE, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.605, + 0.471, + 0.671 + ], + "angle": 0, + "content": "[9] Tim Brooks, Janne Hellsten, Miika Aittala, Ting-Chun Wang, Timo Aila, Jaakko Lehtinen, Ming-Yu Liu, Alexei Efros, and Tero Karras. Generating long videos of dynamic scenes. Advances in Neural Information Processing Systems, 35:31769-31781, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.673, + 0.471, + 0.727 + ], + "angle": 0, + "content": "[10] Thomas Brox, Andrés Bruhn, Nils Papenberg, and Joachim Weickert. High accuracy optical flow estimation based on a theory for warping. In Proc. European Conf. on Computer Vision (ECCV), pages 25-36. Springer, 2004." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.727, + 0.471, + 0.781 + ], + "angle": 0, + "content": "[11] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.781, + 0.471, + 0.835 + ], + "angle": 0, + "content": "[12] Dan Casas, Marco Volino, John Collomosse, and Adrian Hilton. 4d video textures for interactive character appearance. In Computer Graphics Forum, volume 33, pages 371-380. Wiley Online Library, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.835, + 0.471, + 0.888 + ], + "angle": 0, + "content": "[13] Antoni B Chan and Nuno Vasconcelos. Mixtures of dynamic textures. In Tenth IEEE International Conference on Computer Vision (ICCV'05) Volume 1, volume 1, pages 641-647. IEEE, 2005." + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.116, + 0.471, + 0.888 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.146 + ], + "angle": 0, + "content": "[14] Antoni B Chan and Nuno Vasconcelos. Classifying video with kernel dynamic textures. In 2007 IEEE Conference on Computer Vision and Pattern Recognition, pages 1-6. IEEE, 2007." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.148, + 0.894, + 0.202 + ], + "angle": 0, + "content": "[15] Antoni B Chan and Nuno Vasconcelos. Modeling, clustering, and segmenting video with mixtures of dynamic textures. IEEE transactions on pattern analysis and machine intelligence, 30(5):909-926, 2008." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.202, + 0.894, + 0.269 + ], + "angle": 0, + "content": "[16] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.269, + 0.894, + 0.324 + ], + "angle": 0, + "content": "[17] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11315-11325, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.324, + 0.894, + 0.377 + ], + "angle": 0, + "content": "[18] Tsai-Shien Chen, Chieh Hubert Lin, Hung-Yu Tseng, Tsung-Yi Lin, and Ming-Hsuan Yang. Motion-conditioned diffusion model for controllable video synthesis. arXiv preprint arXiv:2304.14404, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.378, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[19] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18000-18010, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.446, + 0.894, + 0.5 + ], + "angle": 0, + "content": "[20] Yung-Yu Chuang, Dan B Goldman, Ke Colin Zheng, Brian Curless, David H Salesin, and Richard Szeliski. Animating pictures with stochastic motion textures. In ACM SIGGRAPH 2005 Papers, pages 853-860. 2005." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.5, + 0.894, + 0.553 + ], + "angle": 0, + "content": "[21] Vincent C Couture, Michael S Langer, and Sebastien Roy. Omnistereo video textures without ghosting. In 2013 International Conference on 3D Vision-3DV 2013, pages 64–70. IEEE, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.554, + 0.892, + 0.594 + ], + "angle": 0, + "content": "[22] Abe Davis, Justin G Chen, and Frédo Durand. Image-space modal bases for plausible manipulation of objects in video. ACM Transactions on Graphics (TOG), 34(6):1-7, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.594, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[23] Myers Abraham Davis. Visual vibration analysis. PhD thesis, Massachusetts Institute of Technology, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.621, + 0.892, + 0.661 + ], + "angle": 0, + "content": "[24] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.661, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[25] Julien Diener, Mathieu Rodriguez, Lionel Baboud, and Lionel Reveret. Wind projection basis for real-time animation of trees. In Computer graphics forum, volume 28, pages 533-540. Wiley Online Library, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.715, + 0.892, + 0.755 + ], + "angle": 0, + "content": "[26] Gianfranco Doretto, Alessandro Chiuso, Ying Nian Wu, and Stefano Soatto. Dynamic textures. International journal of computer vision, 51:91-109, 2003." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.755, + 0.894, + 0.835 + ], + "angle": 0, + "content": "[27] Michael Dorkenwald, Timo Milbich, Andreas Blattmann, Robin Rombach, Konstantinos G. Derpanis, and Bjorn Ommer. Stochastic image-to-video synthesis using cinns. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3742-3753, June 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.835, + 0.894, + 0.893 + ], + "angle": 0, + "content": "[28] Yuming Du, Robin Kips, Albert Pumarola, Sebastian Starke, Ali Thabet, and Artsiom Sanakoyeu. Avatars grow legs: Generating smooth human motion from sparse tracking inputs with diffusion model. In Proceedings of the IEEE/CVF" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.093, + 0.894, + 0.893 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24150" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.093, + 0.47, + 0.119 + ], + "angle": 0, + "content": "Conference on Computer Vision and Pattern Recognition, pages 481-490, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.121, + 0.472, + 0.187 + ], + "angle": 0, + "content": "[29] Yuki Endo, Yoshihiro Kanamori, and Shigeru Kuriyama. Animating landscape: Self-supervised learning of decoupled motion and appearance for single-image video synthesis. ACM Transactions on Graphics (Proceedings of ACM SIGGRAPH Asia 2019), 38(6):175:1-175:19, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.188, + 0.472, + 0.24 + ], + "angle": 0, + "content": "[30] Dave Epstein, Allan Jabri, Ben Poole, Alexei A Efros, and Aleksander Holynski. Diffusion self-guidance for controllable image generation. arXiv preprint arXiv:2306.00986, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.242, + 0.471, + 0.31 + ], + "angle": 0, + "content": "[31] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7346-7356, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.31, + 0.471, + 0.377 + ], + "angle": 0, + "content": "[32] Matthew Flagg, Atsushi Nakazawa, Qiushuang Zhang, Sing Bing Kang, Young Kee Ryu, Irfan Essa, and James M Rehg. Human video textures. In Proceedings of the 2009 symposium on Interactive 3D graphics and games, pages 199-206, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.378, + 0.471, + 0.432 + ], + "angle": 0, + "content": "[33] Jean-Yves Franceschi, Edouard Delasalles, Mickaël Chen, Sylvain Lamprier, and Patrick Gallinari. Stochastic latent residual video prediction. In International Conference on Machine Learning, pages 3233-3246. PMLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.432, + 0.471, + 0.486 + ], + "angle": 0, + "content": "[34] Ruohan Gao, Bo Xiong, and Kristen Grauman. Im2Flow: Motion hallucination from static images for action recognition. In Proc. Computer Vision and Pattern Recognition (CVPR), 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.486, + 0.471, + 0.553 + ], + "angle": 0, + "content": "[35] Songwei Ge, Thomas Hayes, Harry Yang, Xi Yin, Guan Pang, David Jacobs, Jia-Bin Huang, and Devi Parikh. Long video generation with time-agnostic vqgan and timesensitive transformer. arXiv preprint arXiv:2204.03638, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.554, + 0.471, + 0.608 + ], + "angle": 0, + "content": "[36] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.608, + 0.471, + 0.662 + ], + "angle": 0, + "content": "[37] Isma Hadji and Richard P Wildes. A new large scale dynamic texture dataset with application to convnet understanding. In Proceedings of the European Conference on Computer Vision (ECCV), pages 320-335, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.662, + 0.471, + 0.717 + ], + "angle": 0, + "content": "[38] Zekun Hao, Xun Huang, and Serge Belongie. Controllable video generation with sparse trajectories. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7854-7863, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.717, + 0.471, + 0.77 + ], + "angle": 0, + "content": "[39] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.771, + 0.471, + 0.837 + ], + "angle": 0, + "content": "[40] Yingqing He, Menghan Xia, Haoxin Chen, Xiaodong Cun, Yuan Gong, Jinbo Xing, Yong Zhang, Xintao Wang, Chao Weng, Ying Shan, et al. animate-a-story: Storytelling with retrieval-augmented video generation. arXiv preprint arXiv:2307.06940, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.838, + 0.471, + 0.892 + ], + "angle": 0, + "content": "[41] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2022." + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.093, + 0.472, + 0.892 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.093, + 0.894, + 0.16 + ], + "angle": 0, + "content": "[42] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.161, + 0.894, + 0.229 + ], + "angle": 0, + "content": "[43] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.229, + 0.894, + 0.27 + ], + "angle": 0, + "content": "[44] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.27, + 0.894, + 0.296 + ], + "angle": 0, + "content": "[45] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.296, + 0.894, + 0.363 + ], + "angle": 0, + "content": "[46] Aleksander Holynski, Brian L Curless, Steven M Seitz, and Richard Szeliski. Animating pictures with Eulerian motion fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5810-5819, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.364, + 0.894, + 0.404 + ], + "angle": 0, + "content": "[47] Tobias Hoppe, Arash Mehrjou, Stefan Bauer, Didrik Nielsen, and Andrea Dittadi. Diffusion models for video prediction and infilling. Trans. Mach. Learn. Res., 2022, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.405, + 0.894, + 0.445 + ], + "angle": 0, + "content": "[48] Xiaotao Hu, Zhewei Huang, Ailin Huang, Jun Xu, and Shuchang Zhou. A dynamic multi-scale voxel flow network for video prediction. ArXiv, abs/2303.09875, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.445, + 0.894, + 0.513 + ], + "angle": 0, + "content": "[49] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.513, + 0.894, + 0.58 + ], + "angle": 0, + "content": "[50] Hitoshi Kanda and Jun Ohya. Efficient, realistic method for animating dynamic behaviors of 3d botanical trees. In 2003 International Conference on Multimedia and Expo. ICME'03. Proceedings (Cat. No. 03TH8698), volume 2, pages II-89. IEEE, 2003." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.58, + 0.894, + 0.634 + ], + "angle": 0, + "content": "[51] Johanna Karras, Aleksander Holynski, Ting-Chun Wang, and Ira Kemelmacher-Shlizerman. Dreampose: Fashion image-to-video synthesis via stable diffusion. arXiv preprint arXiv:2304.06025, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.634, + 0.894, + 0.701 + ], + "angle": 0, + "content": "[52] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.701, + 0.894, + 0.742 + ], + "angle": 0, + "content": "[53] Alex X Lee, Richard Zhang, Frederik Ebert, Pieter Abbeel, Chelsea Finn, and Sergey Levine. Stochastic adversarial video prediction. arXiv preprint arXiv:1804.01523, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.742, + 0.894, + 0.811 + ], + "angle": 0, + "content": "[54] Zhengqi Li, Tali Dekel, Forrester Cole, Richard Tucker, Noah Snavely, Ce Liu, and William T Freeman. Learning the depths of moving people by watching frozen people. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4521-4530, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.811, + 0.894, + 0.865 + ], + "angle": 0, + "content": "[55] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2041-2050, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.865, + 0.894, + 0.892 + ], + "angle": 0, + "content": "[56] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snively. Dynibar: Neural dynamic image-based" + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.093, + 0.894, + 0.892 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24151" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.093, + 0.47, + 0.133 + ], + "angle": 0, + "content": "rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4273-4284, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.134, + 0.472, + 0.2 + ], + "angle": 0, + "content": "[57] Zhengqi Li, Qianqian Wang, Noah Snavely, and Angjoo Kanazawa. Infinitenature-zero: Learning perpetual view generation of natural scenes from single images. In European Conference on Computer Vision, pages 515-534. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.202, + 0.471, + 0.242 + ], + "angle": 0, + "content": "[58] Jing Liao, Mark Finch, and Hugues Hoppe. Fast computation of seamless video loops. ACM Transactions on Graphics (TOG), 34(6):1-10, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.243, + 0.471, + 0.282 + ], + "angle": 0, + "content": "[59] Zicheng Liao, Neel Joshi, and Hugues Hoppe. Automated video looping with progressive dynamism. ACM Transactions on Graphics (TOG), 32(4):1-10, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.283, + 0.471, + 0.35 + ], + "angle": 0, + "content": "[60] Andrew Liu, Richard Tucker, Varun Jampani, Ameesh Makadia, Noah Snavely, and Angjoo Kanazawa. Infinite nature: Perpetual view generation of natural scenes from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14458-14467, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.351, + 0.471, + 0.39 + ], + "angle": 0, + "content": "[61] Ce Liu. Beyond pixels: exploring new representations and applications for motion analysis. PhD thesis, Massachusetts Institute of Technology, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.391, + 0.471, + 0.471 + ], + "angle": 0, + "content": "[62] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Tieniu Tan. Videofusion: Decomposed diffusion models for high-quality video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10209-10218, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.473, + 0.471, + 0.512 + ], + "angle": 0, + "content": "[63] Aniruddha Mahapatra and Kuldeep Kulkarni. Controllable animation of fluid elements in still images. In Proc. Computer Vision and Pattern Recognition (CVPR), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.513, + 0.471, + 0.552 + ], + "angle": 0, + "content": "[64] Arun Mallya, Ting-Chun Wang, and Ming-Yu Liu. Implicit warping for animation with image sets. Advances in Neural Information Processing Systems, 35:22438-22450, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.553, + 0.471, + 0.62 + ], + "angle": 0, + "content": "[65] Medhini Narasimhan, Shiry Ginosar, Andrew Owens, Alexei A Efros, and Trevor Darrell. Strumming to the beat: Audio-conditioned contrastive video textures. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3761-3770, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.621, + 0.471, + 0.688 + ], + "angle": 0, + "content": "[66] Haomiao Ni, Changhao Shi, Kai Li, Sharon X Huang, and Martin Renqiang Min. Conditional image-to-video generation with latent flow diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18444-18455, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.689, + 0.471, + 0.741 + ], + "angle": 0, + "content": "[67] Simon Niklaus and Feng Liu. Softmax splatting for video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5437-5446, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.743, + 0.471, + 0.81 + ], + "angle": 0, + "content": "[68] Shin Ota, Machiko Tamura, Kunihiko Fujita, T Fujimoto, K Muraoka, and Norishige Chiba. 1/f/sup/spl beta//noise-based real-time animation of trees swaying in wind fields. In Proceedings Computer Graphics International 2003, pages 52-59. IEEE, 2003." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.811, + 0.471, + 0.878 + ], + "angle": 0, + "content": "[69] Automne Petitjean, Yohan Poirier-Ginter, Ayush Tewari, Guillaume Cordonnier, and George Drettakis. Modalnerf: Neural modal analysis and synthesis for free-viewpoint navigation in dynamically vibrating scenes. In Computer Graphics Forum, volume 42, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.879, + 0.471, + 0.891 + ], + "angle": 0, + "content": "[70] Silvia L. Pintea, Jan C. van Gemert, and Arnold W. M." + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.093, + 0.472, + 0.891 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Smeulders. Déjà vu: Motion prediction in static images. In Proc. European Conf. on Computer Vision (ECCV), 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.121, + 0.894, + 0.159 + ], + "angle": 0, + "content": "[71] Sigal Raab, Inbal Leibovitch, Guy Tevet, Moab Arar, Amit H Bermano, and Daniel Cohen-Or. Single motion diffusion. arXiv preprint arXiv:2302.05905, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.16, + 0.894, + 0.213 + ], + "angle": 0, + "content": "[72] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2):3, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.215, + 0.894, + 0.282 + ], + "angle": 0, + "content": "[73] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.283, + 0.894, + 0.363 + ], + "angle": 0, + "content": "[74] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.365, + 0.894, + 0.432 + ], + "angle": 0, + "content": "[75] Payam Saisan, Gianfranco Doretto, Ying Nian Wu, and Stefano Soatto. Dynamic texture recognition. In Proceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition. CVPR 2001, volume 2, pages II-II. IEEE, 2001." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.433, + 0.894, + 0.486 + ], + "angle": 0, + "content": "[76] Saurabh Saxena, Charles Herrmann, Junhwa Hur, Abhishek Kar, Mohammad Norouzi, Deqing Sun, and David J. Fleet. The surprising effectiveness of diffusion models for optical flow and monocular depth estimation, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.486, + 0.894, + 0.54 + ], + "angle": 0, + "content": "[77] Arno Schödl, Richard Szeliski, David H Salesin, and Irfan Essa. Video textures. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 489-498, 2000." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.541, + 0.894, + 0.58 + ], + "angle": 0, + "content": "[78] Mikio Shinya and Alain Fournier. Stochastic motion—motion under the influence of wind. Computer Graphics Forum, 11(3), 1992." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.581, + 0.894, + 0.648 + ], + "angle": 0, + "content": "[79] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. Animating arbitrary objects via deep motion transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2377-2386, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.649, + 0.894, + 0.702 + ], + "angle": 0, + "content": "[80] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. Advances in neural information processing systems, 32, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.703, + 0.894, + 0.77 + ], + "angle": 0, + "content": "[81] Aliaksandr Siarohin, Oliver J Woodford, Jian Ren, Mengei Chai, and Sergey Tulyakov. Motion representations for articulated animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13653-13662, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.771, + 0.894, + 0.811 + ], + "angle": 0, + "content": "[82] Chen Qian Kwan-Yee Lin Hongsheng Li Siming Fan, Jingtan Piao. Simulating fluids in real-world still images. arXiv preprint, arXiv:2204.11335, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.811, + 0.894, + 0.878 + ], + "angle": 0, + "content": "[83] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3626-3636, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.879, + 0.894, + 0.891 + ], + "angle": 0, + "content": "[84] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan," + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.093, + 0.894, + 0.891 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24152" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.093, + 0.471, + 0.134 + ], + "angle": 0, + "content": "and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.134, + 0.472, + 0.173 + ], + "angle": 0, + "content": "[85] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv:2010.02502, October 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.173, + 0.472, + 0.227 + ], + "angle": 0, + "content": "[86] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.228, + 0.472, + 0.254 + ], + "angle": 0, + "content": "[87] Jos Stam. Multi-scale stochastic modelling of complex natural phenomena. PhD thesis, 1995." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.254, + 0.472, + 0.294 + ], + "angle": 0, + "content": "[88] Jos Stam. Stochastic dynamics: Simulating the effects of turbulence on flexible structures. Computer Graphics Forum, 16(3), 1997." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.295, + 0.471, + 0.348 + ], + "angle": 0, + "content": "[89] Ryusuke Sugimoto, Mingming He, Jing Liao, and Pedro V Sander. Water simulation and rendering from a still photograph. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.348, + 0.471, + 0.389 + ], + "angle": 0, + "content": "[90] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.389, + 0.471, + 0.443 + ], + "angle": 0, + "content": "[91] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.443, + 0.471, + 0.497 + ], + "angle": 0, + "content": "[92] Vikram Voleti, Alexia Jolicoeur-Martineau, and Christopher Pal. Mcvd: Masked conditional video diffusion for prediction, generation, and interpolation. In (NeurIPS) Advances in Neural Information Processing Systems, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.497, + 0.471, + 0.537 + ], + "angle": 0, + "content": "[93] Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba. Generating videos with scene dynamics. In Neural Information Processing Systems, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.537, + 0.471, + 0.59 + ], + "angle": 0, + "content": "[94] Neal Wadhwa, Michael Rubinstein, Frédo Durand, and William T Freeman. Phase-based video motion processing. ACM Transactions on Graphics (ToG), 32(4):1-10, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.591, + 0.471, + 0.646 + ], + "angle": 0, + "content": "[95] Jacob Walker, Carl Doersch, Abhinav Gupta, and Martial Hebert. An uncertain future: Forecasting from static images using variational autoencoders. In Proc. European Conf. on Computer Vision (ECCV), 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.646, + 0.471, + 0.699 + ], + "angle": 0, + "content": "[96] Jacob Walker, Abhinav Gupta, and Martial Hebert. Dense optical flow prediction from a static image. In Proceedings of the IEEE International Conference on Computer Vision, pages 2443-2451, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.7, + 0.471, + 0.767 + ], + "angle": 0, + "content": "[97] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. arXiv preprint arXiv:2306.02018, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.767, + 0.471, + 0.821 + ], + "angle": 0, + "content": "[98] Yaohui Wang, Di Yang, Francois Bremond, and Antitza Dantcheva. Latent image animator: Learning to animate images via latent space navigation. arXiv preprint arXiv:2203.09043, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.821, + 0.471, + 0.876 + ], + "angle": 0, + "content": "[99] Frederik Warburg, Ethan Weber, Matthew Tancik, Aleksander Holynski, and Angjoo Kanazawa. Nerfbusters: Removing ghostly artifacts from casually captured nerfs. arXiv preprint arXiv:2304.10532, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.876, + 0.471, + 0.89 + ], + "angle": 0, + "content": "[100] Daniel Watson, William Chan, Ricardo Martin-Brualla," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.472, + 0.89 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.093, + 0.894, + 0.134 + ], + "angle": 0, + "content": "Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. arXiv preprint arXiv:2210.04628, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.134, + 0.894, + 0.201 + ], + "angle": 0, + "content": "[101] Chung-Yi Weng, Brian Curless, and Ira Kemelmacher-Shlizerman. Photo wake-up: 3d character animation from a single photo. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5908-5917, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.201, + 0.894, + 0.242 + ], + "angle": 0, + "content": "[102] Jamie Wynn and Daniyar Turmukhambetov. DiffusioNeRF: Regularizing Neural Radiance Fields with Denoising Diffusion Models. In CVPR, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.242, + 0.894, + 0.309 + ], + "angle": 0, + "content": "[103] Tianfan Xue, Jiajun Wu, Katherine L Bouman, and William T Freeman. Visual dynamics: Stochastic future generation via layered cross convolutional networks. Trans. Pattern Analysis and Machine Intelligence, 41(9):2236-2250, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.31, + 0.893, + 0.364 + ], + "angle": 0, + "content": "[104] Shengming Yin, Chenfei Wu, Jian Liang, Jie Shi, Houqiang Li, Gong Ming, and Nan Duan. Dragnuwa: Fine-grained control in video generation by integrating text, image, and trajectory. arXiv preprint arXiv:2308.08089, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.364, + 0.894, + 0.431 + ], + "angle": 0, + "content": "[105] Sihyun Yu, Kihyuk Sohn, Subin Kim, and Jinwoo Shin. Video probabilistic diffusion models in projected latent space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18456-18466, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.431, + 0.894, + 0.486 + ], + "angle": 0, + "content": "[106] Mingyuan Zhang, Xinying Guo, Liang Pan, Zhongang Cai, Fangzhou Hong, Huirong Li, Lei Yang, and Ziwei Liu. Remodiffuse: Retrieval-augmented motion diffusion model. arXiv preprint arXiv:2304.01116, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.486, + 0.893, + 0.54 + ], + "angle": 0, + "content": "[107] Yabo Zhang, Yuxiang Wei, Dongsheng Jiang, Xiaopeng Zhang, Wangmeng Zuo, and Qi Tian. Controlvideo: Training-free controllable text-to-video generation. arXiv preprint arXiv:2305.13077, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.54, + 0.893, + 0.594 + ], + "angle": 0, + "content": "[108] Jian Zhao and Hui Zhang. Thin-plate spline motion model for image animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3657–3666, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.594, + 0.893, + 0.662 + ], + "angle": 0, + "content": "[109] Shengyu Zhao, Jonathan Cui, Yilun Sheng, Yue Dong, Xiao Liang, Eric I Chang, and Yan Xu. Large scale image completion via co-modulated generative adversarial networks. In International Conference on Learning Representations (ICLR), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.662, + 0.893, + 0.716 + ], + "angle": 0, + "content": "[110] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022." + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.716 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24153" + } + ] +] \ No newline at end of file diff --git a/2024/Generative Image Dynamics/34d8fada-6dfb-4124-8738-6ec5f099992d_origin.pdf b/2024/Generative Image Dynamics/34d8fada-6dfb-4124-8738-6ec5f099992d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..afab970a46b6470fdf5cbd6cb59df7f10aacaf6d --- /dev/null +++ b/2024/Generative Image Dynamics/34d8fada-6dfb-4124-8738-6ec5f099992d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:38812f9e3295c3ebb683c33f412a832e5108d25bdac9e16e7da8adcb692891a7 +size 4003794 diff --git a/2024/Generative Image Dynamics/full.md b/2024/Generative Image Dynamics/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9128c1f972c0972401996f2c38cacc9f59d5c52e --- /dev/null +++ b/2024/Generative Image Dynamics/full.md @@ -0,0 +1,395 @@ +# Generative Image Dynamics + +Zhengqi Li + +Richard Tucker + +Noah Snavely + +Aleksander Holynski + +Google Research + +![](images/1970c681d962b8853f36ccc8ddaa928629b646be28135670f9eaa973dda86f4c.jpg) +Input Picture + +![](images/59ff8290e3cfb9113942f4c942cf80ea62a05c73470708eecd62a494792d00db.jpg) + +![](images/78a247a00539d0c217235990f6b0f8246232cb2fa9eb2aef343d17d5599a3585.jpg) +Spectral Volume (Ge-Space Modal Basis) + +![](images/5a42bd78bf68b0384b71f855a8a0777cec05b33cfd30a08f546469f3abb83c7d.jpg) + +![](images/34db33c57aca6d04fb45265e6306b77ab81015b48f000e0d3b5a555a59ad3b49.jpg) + +![](images/455907a99362c2f7ae7a8d16b5365affcae8744b009709fcfea82bbcd5348b35.jpg) + +![](images/7a85c481136ceda92f81229273fb6140ebf9894317f442717658aa20a53d3f1b.jpg) + +![](images/d19703c80e58541080695e33803a4edf3861d9c7d7bba0f397bec0177c4387b5.jpg) + +![](images/ac2065fee6863ad784779beaccf133f097476adc78b8661117189f30e348e101.jpg) +X coefficients + +![](images/6ca5b677fbf8d3e4bf3bec7c9223c52dee59e51042843b48f57d2890205a9704.jpg) +Y coefficients + +![](images/37c84aaa5014cb7cd271e56cfcb527993742f81f69768eafa7ac4fd800fdb55f.jpg) + +![](images/6f14a0e77e624585c13fd3df632bca6bfc17717dbd0a100efa2417ab8765a6cb.jpg) +Figure 1. We model a generative image-space prior on scene motion: from a single RGB image, our method generates a spectral volume [23], a motion representation that models dense, long-term pixel trajectories in the Fourier domain. Our learned motion priors can be used to turn a single picture into a seamlessly looping video, or into an interactive simulation of dynamics that responds to user inputs like dragging and releasing points. On the right, we visualize output videos as space-time $X - t$ slices (along the input scanline shown on the left). + +# Abstract + +We present an approach to modeling an image-space prior on scene motion. Our prior is learned from a collection of motion trajectories extracted from real video sequences depicting natural, oscillatory dynamics of objects such as trees, flowers, candles, and clothes swaying in the wind. We model dense, long-term motion in the Fourier domain as spectral volumes, which we find are well-suited to prediction with diffusion models. Given a single image, our trained model uses a frequency-coordinated diffusion sampling process to predict a spectral volume, which can be converted into a motion texture that spans an entire video. Along with an image-based rendering module, the predicted motion representation can be used for a number of downstream applications, such as turning still images into seamlessly looping videos, or allowing users to interact with objects in real images, producing realistic simulated dynamics (by interpreting the spectral volumes as image-space modal bases). See our project page for more results: generative-dynamics.github.io. + +# 1. Introduction + +The natural world is always in motion, with even seemingly static scenes containing subtle oscillations as a result of wind, water currents, respiration, or other natural rhythms. Emulating this motion is crucial in visual content synthesis—human sensitivity to motion can cause imagery without motion (or with slightly unrealistic motion) to seem uncanny or unreal. + +While it is easy for humans to interpret or imagine motion in scenes, training a model to learn or produce realistic scene motion is far from trivial. The motion we observe in the world is the result of a scene's underlying physical dynamics, i.e., forces applied to objects that respond according to their unique physical properties—their mass, elasticity, etc—quantities that are hard to measure and capture at scale. Fortunately, measuring them is not necessary for certain applications: e.g., one can simulate plausible dynamics in a scene by simply analyzing some observed 2D motion [23]. + +This same observed motion can also serve as a supervisory signal in learning dynamics across scenes—because although observed motion is multi-modal and grounded in complex physical effects, it is nevertheless often predictable: + +candles will flicker in certain ways, trees will sway, and their leaves will rustle. As humans, this predictability is ingrained in our systems of perception: by viewing a still image, we can imagine plausible motions—or, since there might have been many possible such motions, a distribution of natural motions conditioned on that image. Given the facility with which humans are able to model these distributions, a natural research problem is to model them computationally. + +Recent advances in generative models, in particular conditional diffusion models [44, 84, 86], have enabled us to model rich distributions, including distributions of real images conditioned on text [72-74]. This capability has enabled several new applications, such as text-conditioned generation of diverse and realistic image content. Following the success of these image models, recent work has extended these models to other domains, such as videos [7, 43] and 3D geometry [76, 99, 100, 102]. + +In this paper, we model a generative prior for image-space scene motion, i.e., the motion of all pixels in a single image. This model is trained on motion trajectories automatically extracted from a large collection of real video sequences. In particular, from each training video we compute motion in the form of a spectral volume [22, 23], a frequency-domain representation of dense, long-range pixel trajectories. Spectral volumes are well-suited to scenes that exhibit oscillatory dynamics, e.g., trees and flowers moving in the wind. We find that this representation is also highly effective as an output of a diffusion model for modeling scene motion. We train a generative model that, conditioned on a single image, can sample spectral volumes from its learned distribution. A predicted spectral volume can then be directly transformed into a motion texture—a set of long-range, per-pixel motion trajectories—that can be used to animate the image. The spectral volume can also be interpreted as an image-space modal basis for use in simulating interactive dynamics [22]. + +We predict spectral volumes from input images using a diffusion model that generates coefficients one frequency at a time, but coordinates these predictions across frequency bands through a shared attention module. The predicted motions can be used to synthesize future frames (via an image-based rendering model)—turning still images into realistic animations, as illustrated in Fig. 1. + +Compared with priors over raw RGB pixels, priors over motion capture more fundamental, lower-dimensional structure that efficiently explains long-range variations in pixel values. Hence, generating intermediate motion leads to more coherent long-term generation and more fine-grained control over animations. We demonstrate the use of our trained model in several downstream applications, such as creating seamless looping videos, editing the generated motions, and enabling interactive dynamic images via image-space modal bases, i.e., simulating the response of object dynamics to user-applied forces [22]. + +# 2. Related Work + +Generative synthesis. Recent advances in generative models have enabled photorealistic synthesis of images conditioned on text prompts [16, 17, 24, 72-74]. These text-to-image models can be augmented to synthesize video sequences by extending the generated image tensors along a time dimension [7, 9, 43, 62, 83, 105, 105, 110]. While these methods can produce video sequences that capture the spatiotemporal statistics of real footage, these videos often suffer from artifacts like incoherent motion, unrealistic temporal variation in textures, and violations of physical constraints like preservation of mass. + +Animating images. Instead of generating videos entirely from text, other techniques take as input a still picture and animate it. Many recent deep learning methods adopt a 3D-Unet architecture to produce video volumes directly [27, 36, 40, 47, 53, 92]. These models are effectively the same video generation models (but conditioned on image information instead of text), and exhibit similar artifacts to those mentioned above. One way to overcome these limitations is to not directly generate the video content itself, but instead animate an input source image through image-based rendering, i.e., moving the image content around according to motion derived from external sources such as a driving video [51, 79-81, 98], motion or 3D geometry priors [8, 29, 46, 63, 64, 66, 89, 96, 100, 101, 103, 108], or user annotations [6, 18, 20, 33, 38, 97, 104, 107]. Animating images according to motion fields yields greater temporal coherence and realism, but these prior methods either require additional guidance signals or user input, or utilize limited motion representations. + +Motion models and motion priors. In computer graphics, natural, oscillatory 3D motion (e.g., water rippling or trees waving in the wind) can be modeled with noise that is shaped in the Fourier domain and then converted to time-domain motion fields [78, 87]. Some of these methods rely on a modal analysis of the underlying dynamics of the system being simulated [22, 25, 88]. These spectral techniques were adapted to animate plants, water, and clouds from single 2D pictures by Chuang et al. [20], given user annotations. Our work is especially inspired by Davis [23], who connected modal analysis of a scene with the motions observed in a video of that scene, and used this analysis to simulate interactive dynamics from a video. We adopt the frequency-space spectral volume motion representation from Davis et al., extract this representation from a large set of training videos, and show that spectral volumes are suitable for predicting motion from single images with diffusion models. + +Other methods have used various motion representations in prediction tasks, where an image or video is used to inform a deterministic future motion estimate [34, 70], or a more rich distribution of possible motions [93, 95, 103]. However, + +many of these methods predict an optical flow motion estimate (i.e., the instantaneous motion of each pixel), not full motion trajectories. In addition, much of this prior work is focused on tasks like activity recognition, not on synthesis tasks. More recent work has demonstrated the advantages of modeling and predicting motion using generative models in a number of closed-domain settings such as humans and animals [2, 19, 28, 71, 90, 106]. + +Videos as textures. Certain moving scenes can be thought of as a kind of texture—termed dynamic textures [26]—that model videos as space-time samples of a stochastic process. Dynamic textures can represent smooth, natural motions like waves, flames, or moving trees, and have been widely used for video classification, segmentation or encoding [12-15, 75]. A related kind of texture, called a video texture, represents a moving scene as a set of input video frames along with transition probabilities between any pair of frames [65, 77]. A number of methods estimate dynamic or video textures through analysis of scene motion and pixel statistics, with the aim of generating seamlessly looping or infinitely varying output videos [1, 21, 32, 58, 59, 77]. In contrast to much of this work, our method learns priors in advance that can then be applied to single images. + +# 3. Overview + +Given a single picture $I_0$ , our goal is to generate a video $\{\hat{I}_1, \hat{I}_2, \dots, \hat{I}_T\}$ featuring oscillatory motions such as those of trees, flowers, or candle flames swaying in the breeze. Our system consists of two modules: a motion prediction module and an image-based rendering module. Our pipeline begins by using a latent diffusion model (LDM) to predict a spectral volume $S = (S_{f_0}, S_{f_1}, \dots, S_{f_{K-1}})$ for the input $I_0$ . The predicted spectral volume is then transformed to a motion texture $\mathcal{F} = (F_1, F_2, \dots, F_T)$ through an inverse discrete Fourier transform. This motion determines the position of each input pixel at every future time step. + +Given a predicted motion texture, we then animate the input RGB image using a neural image-based rendering technique (Sec. 5). We explore applications of this method, including producing seamless looping animations and simulating interactive dynamics, in Sec. 6. + +# 4. Predicting motion + +# 4.1. Motion representation + +Formally, a motion texture is a sequence of time-varying 2D displacement maps $\mathcal{F} = \{F_t|t = 1,\dots,T\}$ , where the 2D displacement vector $F_{t}(\mathbf{p})$ at each pixel coordinate $\mathbf{p}$ from input image $I_0$ defines the position of that pixel at a future time $t$ [20]. To generate a future frame at time $t$ , one can sput pixels from $I_0$ using the corresponding displacement + +![](images/54d1208cb540e2bb8f79089f292952a53d70139d7f5cacf10203728b31d89d51.jpg) +Figure 2. Left: We visualize the average power spectrum for the $X$ and $Y$ motion components extracted from real videos, shown as the blue and green curves. Natural oscillation motions are composed primarily of low-frequency components, and so we use the first $K = 16$ terms, marked with red dots. Right: we show a histogram of the amplitude of Fourier terms at $3.0\mathrm{Hz}$ after (1) scaling amplitude by image width and height (blue), or (2) frequency adaptive normalization (red). Our adaptive normalization prevents the coefficients from concentrating at extreme values. + +![](images/77ce6d12a9675cdf33837894d68f11fb0ca1d62780bdf8d19002f05cbe16bb42.jpg) + +map $D_{t}$ , resulting in a forward-warped image $I_t'$ : + +$$ +I _ {t} ^ {\prime} (\mathbf {p} + F _ {t} (\mathbf {p})) = I _ {0} (\mathbf {p}). \tag {1} +$$ + +If our goal is to produce a video via a motion texture, then one choice would be to predict a time-domain motion texture directly from an input image. However, the size of the motion texture would need to scale with the length of the video: generating $T$ output frames implies predicting $T$ displacement fields. To avoid predicting such a large output representation for long videos, many prior animation methods either generate video frames autoregressively [7, 29, 57, 60, 92], or predict each future output frame independently via an extra time embedding [4]. However, neither strategy ensures long-term temporal consistency of generated videos. + +Fortunately, many natural motions can be described as a superposition of a small number of harmonic oscillators represented with different frequencies, amplitude and phases [20, 23, 25, 50, 68]. Because these underlying motions are quasi-periodic, it is natural to model them in the frequency domain. Hence, we adopt from Davis et al. [23] an efficient frequency space representation of motion in a video called a spectral volume, visualized in Fig. 3. A spectral volume is the temporal Fourier transform of per-pixel trajectories extracted from a video. + +Given this motion representation, we formulate the motion prediction problem as a multi-modal image-to-image translation task: from an input image to an output motion spectral volume. We adopt latent diffusion models (LDMs) to generate spectral volumes comprised of a $4K$ -channel 2D motion spectrum map, where $K \ll T$ is the number of frequencies modeled, and where at each frequency we need four scalars to represent the complex Fourier coefficients for the $x$ - and $y$ -dimensions. Note that the motion trajectory of a pixel at future time steps $\mathcal{F}(\mathbf{p}) = \{F_t(\mathbf{p})|t = 1,2,\dots T\}$ and its representation as a spectral volume $\mathcal{S}(\mathbf{p}) = \{S_{f_k}(\mathbf{p})|k = 0,1,\dots \frac{T}{2} -1\}$ are related by the Fast Fourier transform (FFT): + +$$ +\mathcal {S} (\mathbf {p}) = \operatorname {F F T} (\mathcal {F} (\mathbf {p})). \tag {2} +$$ + +![](images/5744e0ab4ea703956d9398b30c1ee12d84e4d34f4e6d9a5b3f42a74356566278.jpg) +Figure 3. Motion prediction module. We predict a spectral volume $S$ through a frequency-coordinated denoising model. Each block of the diffusion network $\epsilon_{\theta}$ interleaves 2D spatial layers with attention layers (red box, right), and iteratively denoises latent features $z^n$ . The denoised features are fed to a decoder $\mathcal{D}$ to produce $S$ . During training, we concatenate the downsampled input $I_0$ with noisy latent features encoded from a real motion texture via an encoder $\mathcal{E}$ , and replace the noisy features with Gaussian noise $z^N$ during inference (left). + +How should we select the $K$ output frequencies? Prior work in real-time animation has observed that most natural oscillation motions are composed primarily of low-frequency components [25, 68]. To validate this observation, we computed the average power spectrum of the motion extracted from 1,000 randomly sampled 5-second real video clips. As shown in the left plot of Fig. 2, the power spectrum of the motion decreases exponentially with increasing frequency. This suggests that most natural oscillation motions can indeed be well represented by low-frequency terms. In practice, we found that the first $K = 16$ Fourier coefficients are sufficient to realistically reproduce the original natural motion in a range of real videos and scenes. + +# 4.2. Predicting motion with a diffusion model + +We choose a latent diffusion model (LDM) [73] as the backbone for our motion prediction module, as LDMs are more computationally efficient than pixel-space diffusion models, while preserving synthesis quality. A standard LDM consists of two main modules: (1) a variational autoencoder (VAE) that compresses the input image to a latent space through an encoder $z = E(I)$ , then reconstructs the input from the latent features via a decoder $I = D(z)$ , and (2) a U-Net based diffusion model that learns to iteratively denoise features starting from Gaussian noise. Our training applies this process not to RGB images but to spectral volumes from real video sequences, which are encoded and then diffused for $n$ steps with a pre-defined variance schedule to produce noisy latents $z^n$ . The 2D U-Nets are trained to denoise the noisy latents by iteratively estimating the noise $\epsilon_{\theta}(z^{n};n,c)$ used to update the latent feature at each step $n\in (1,2,\dots,N)$ . The training loss for the LDM is written as + +$$ +\mathcal {L} _ {\mathrm {L D M}} = \mathbb {E} _ {n \in \mathcal {U} [ 1, N ], \epsilon^ {n} \in \mathcal {N} (0, 1)} \left[ \left| \left| \epsilon^ {n} - \epsilon_ {\theta} \left(z ^ {n}; n, c\right) \right| \right| ^ {2} \right] \tag {3} +$$ + +where $c$ is the embedding of any conditional signal, such as text, or, in our case, the first frame of the training video sequence, $I_0$ . The clean latent features $z^0$ are then passed through the decoder to recover the spectral volume. + +Frequency adaptive normalization. One issue we observed is that motion textures have particular distribution characteristics across frequencies. As visualized in the left plot of Fig. 2, the amplitude of the spectral volumes spans a range of 0 to 100 and decays approximately exponentially with increasing frequency. As diffusion models require that the absolute values of the output are between -1 and 1 for stable training and denoising [44], we must normalize the coefficients of $S$ extracted from real videos before using them for training. If we scale the magnitudes of these coefficients to [0,1] based on the image dimensions as in prior work [29, 76], nearly all the coefficients at higher frequencies will end up close to zero, as shown in the right plot of Fig. 2. Models trained on such data can produce inaccurate motions, since during inference, even small prediction errors can cause large relative errors after denormalization. + +To address this issue, we employ a simple but effective frequency adaptive normalization method: First, we independently normalize Fourier coefficients at each frequency based on statistics computed from the training set. Namely, for each individual frequency $f_{j}$ , we compute the $95^{\mathrm{th}}$ percentile of Fourier coefficient magnitudes over all input samples and use that value as a per-frequency scaling factor $s f_{j}$ . We then apply a power transformation to each scaled Fourier coefficient to pull it away from extreme values. In practice, we observe that a square root performs better than other nonlinear transformations such as log or reciprocal. In summary, the final coefficient values of spectral volume $S(\mathbf{p})$ at + +frequency $f_{j}$ (used for training our LDM) are computed as + +$$ +S _ {f _ {j}} ^ {\prime} (\mathbf {p}) = \operatorname {s i g n} \left(S _ {f _ {j}}\right) \sqrt {\left| \frac {S _ {f _ {j}} (\mathbf {p})}{s _ {f _ {j}}} \right|}. \tag {4} +$$ + +As shown on the right plot of Fig. 2, after applying frequency adaptive normalization, the spectral volume coefficients distribute more evenly. + +Frequency-coordinated denoising. The straightforward way to predict a spectral volume $S$ with $K$ frequency bands is to output a tensor of $4K$ channels from a single diffusion U-Net. However, as in prior work [7], we observe that training a model to produce a large number of channels can yield over-smoothed, inaccurate outputs. An alternative would be to independently predict each individual frequency slice by injecting an extra frequency embedding into the LDM [4], but this design choice would result in uncorrelated predictions in the frequency domain, leading to unrealistic motion. + +Therefore, inspired by recent video diffusion work [7], we propose a frequency-coordinated denoising strategy, illustrated in Fig. 3. In particular, given an input image $I_0$ , we first train an LDM $\epsilon_{\theta}$ to predict a single 4-channel frequency slice of spectral volume $S_{f_j}$ , where we inject an extra frequency embedding along with the time-step embedding into the LDM. We then freeze the parameters of this LDM $\epsilon_{\theta}$ , introduce attention layers interleaved with the 2D spatial layers of $\epsilon_{\theta}$ across the $K$ frequency bands, and fine-tune. Specifically, for a batch size $B$ , the 2D spatial layers of $\epsilon_{\theta}$ treat the corresponding $B \cdot K$ noisy latent features of channel size $C$ as independent samples with shape $\mathcal{R}^{(B \cdot K) \times C \times H \times W}$ . The attention layer then interprets these as consecutive features spanning the frequency axis, and we reshape the latent features from previous 2D spatial layers to $\mathcal{R}^{B \times K \times C \times H \times W}$ before feeding them to the attention layers. In other words, the frequency attention layers are fine-tuned to coordinate all frequency slices so as to produce coherent spectral volumes. In our experiments, we see that the average VAE reconstruction error improves from 0.024 to 0.018 when we switch from a single 2D U-Net to a frequency-coordinated denoising module, suggesting an improved upper bound on LDM prediction accuracy; in Sec. 7.3, we also show that this design choice improves video generation quality. + +# 5. Image-based rendering + +We now describe how we take a spectral volume $S$ predicted for a given input image $I_0$ and render a future frame $\hat{I}_t$ at time $t$ . We first derive a motion texture in the time domain using the inverse temporal FFT applied at each pixel $\mathcal{F}(\mathbf{p}) = \mathrm{FFT}^{-1}(S(\mathbf{p}))$ . To produce a future frame $\hat{I}_t$ , we adopt a deep image-based rendering technique and perform splatting with the predicted motion field $F_t$ to forward-warp the encoded $I_0$ , as shown in Fig. 4. Since forward warping can lead to holes, and multiple source pixels can map to the same output + +![](images/983a2734f6cefff5069132c49a005bc46464b113a35388810637e96b3a9af55f.jpg) +Figure 4. Rendering module. We fill in missing content and refine the warped input image using a deep image-based rendering module, where multi-scale features are extracted from the input image $I_0$ . Softmax splatting is then applied over the features with a motion field $F_t$ from time 0 to $t$ (subject to the weights $W$ ). The warped features are fed to an image synthesis network to produce the rendered image $\hat{I}_t$ . + +2D location, we adopt the feature pyramid softmax splatting strategy proposed in prior work on frame interpolation [67]. + +Specifically, we encode $I_0$ through a feature extractor network to produce a multi-scale feature map. For each individual feature map at scale $j$ , we resize and scale the predicted 2D motion field $F_{t}$ according to the resolution. As in Davis et al. [22], we use predicted flow magnitude as a proxy for depth to determine the contributing weight of each source pixel mapped to its destination location. In particular, we compute a per-pixel weight, $W(\mathbf{p}) = \frac{1}{T}\sum_{t}||F_{t}(\mathbf{p})||_{2}^{2}$ as the average magnitude of the predicted motion texture. In other words, we assume large motions correspond to moving foreground objects, and small or zero motions correspond to background. We use motion-derived weights instead of learnable ones as [46] because we observe that in the single-view case, learnable weights are not effective for addressing disocclusion ambiguities. + +With the motion field $F_{t}$ and weights $W$ , we apply softmax splatting to warp the feature map at each scale to produce a warped feature. The warped features are then injected into the corresponding blocks of an image synthesis decoder to produce a final rendered image $\hat{I}_{t}$ . + +We jointly train the feature extractor and synthesis networks with start and target frames $(I_0, I_t)$ randomly sampled from real videos, using the estimated flow field from $I_0$ to $I_t$ to warp encoded features from $I_0$ , and supervising predictions $\hat{I}_t$ against $I_t$ with a VGG perceptual loss [49]. + +# 6. Applications + +Image-to-video. Our system enables the animation of a single still picture by first predicting a motion spectral volume from the input image and generating an animation by applying our image-based rendering module to the motion + +texture transformed from the spectral volume. Since we explicitly model scene motion, this allows us to produce slow-motion videos by linearly interpolating the motion texture, or to magnify (or minify) animated motions by adjusting the amplitude of the predicted spectral volume coefficients. + +Seamless looping. Many applications require videos that loop seamlessly, where there is no discontinuity between the start and end of the video. Unfortunately, it is hard to find a large collection of seamlessly looping videos for training. Instead, we devise a method to use our motion diffusion model, trained on regular non-looping video clips, to produce seamless looping video. Inspired by recent work on guidance for image editing [3, 30], our method is a motion self-guidance technique that guides the motion denoising sampling processing using explicit looping constraints. In particular, at each iterative denoising step during inference, we incorporate an additional motion guidance signal alongside standard classifier-free guidance [45], where we enforce each pixel's position and velocity at the start and end frames to be as similar as possible: + +$$ +\tilde {\epsilon} ^ {n} = (1 + w) \epsilon_ {\theta} \left(z ^ {n}; n, c\right) - w \epsilon_ {\theta} \left(z ^ {n}; n, \emptyset\right) + u \sigma^ {n} \nabla_ {z ^ {n}} \mathcal {L} _ {g} ^ {n} +$$ + +$$ +\mathcal {L} _ {g} ^ {n} = \left\| F _ {T} ^ {n} - F _ {1} ^ {n} \right\| _ {1} + \left\| \nabla F _ {T} ^ {n} - \nabla F _ {1} ^ {n} \right\| _ {1} \tag {5} +$$ + +where $F_{t}^{n}$ is the predicted 2D displacement field at time $t$ and denoising step $n$ . $w$ is the classifier-free guidance weight, and $u$ is the motion self-guidance weight. In the supplemental video, we apply a baseline appearance-based looping algorithm [58] to generate a looping video from our non-looping output, and show that our motion self-guidance technique produces seamless looping videos with less distortion and fewer artifacts. + +Interactive dynamics from a single image. Davis et al. [22] show that the spectral volume, evaluated at certain resonant frequencies, can approximate an image-space modal basis that is a projection of the vibration modes of the underlying scene (or, more generally, captures spatial and temporal correlations in oscillatory dynamics), and can be used to simulate the object's response to a user-defined force. We adopt this modal analysis method [22, 69], allowing us to write the image-space 2D motion displacement field for the object's physical response as a weighted sum of motion spectrum coefficients $S_{f_j}$ modulated by the state of complex modal coordinates $\mathbf{q}_{f_j}(t)$ at each simulated time step $t$ : + +$$ +F _ {t} (\mathbf {p}) = \sum_ {f _ {j}} S _ {f _ {j}} (\mathbf {p}) \mathbf {q} _ {f _ {j}} (t) \tag {6} +$$ + +We simulate the state of the modal coordinates $\mathbf{q}_{f_j}(t)$ via an explicit Euler method applied to the equations of motion for a decoupled mass-spring-damper system represented in modal space [22, 23, 69]. We refer readers to supplementary material and original work for a full derivation. Note that our method produces an interactive scene from a single picture, whereas these prior methods required a video as input. + +
MethodImage SynthesisVideo Synthesis
FIDKIDFVDFVD32DTFVDDTFVD32
TATS [35]65.81.67265.6419.622.640.7
Stochastic I2V [27]68.33.12253.5320.916.741.7
MCVD [92]63.42.97208.6270.419.553.9
LFDM [66]47.61.70187.5254.313.045.6
DMVFN [48]37.91.09206.5316.311.254.5
Endo et al. [29]10.40.19166.0231.65.3565.1
Holynski et al. [46]11.20.20179.0253.77.2346.8
Ours4.030.0847.162.92.536.75
+ +Table 1. Quantitative comparisons on the test set. We report both image synthesis and video synthesis quality. Here, KID is scaled by 100. Lower is better for all error. See Sec. 7.1 for descriptions of baselines and error metrics. + +# 7. Experiments + +Implementation details. We use an LDM [73] as the backbone for predicting spectral volumes, for which we use a VAE with a continuous latent space of dimension 4. We train the VAE with an $L_{1}$ reconstruction loss, a multi-scale gradient consistency loss [54-56], and a KL-divergence loss with respective weights of $1,0.2,10^{-6}$ . We train the same 2D U-Net used in the original LDM work to perform iterative denoising with a simple MSE loss [44], and adopt the attention layers from [41] for frequency-coordinated denoising. For quantitative evaluation, we train both VAE and LDM on images of size $256 \times 160$ from scratch for fair comparisons, and it takes around 6 days to converge using 16 Nvidia A100 GPUs. For our main quantitative and qualitative results, we run the motion diffusion model with DDIM [85] for 250 steps. We also show generated videos of up to a resolution of $512 \times 288$ , created by fine-tuning a pre-trained image inpainting LDM model [73] on our dataset. + +We adopt ResNet-34 [39] for the feature extractor in our IBR module. Our image synthesis network is based on an architecture for conditional image inpainting [57, 109]. Our rendering module runs in real-time at 25FPS on a Nvidia V100 GPU during inference. We adopt universal guidance [3] to produce seamless looping videos, where we set weights $w = 1.75$ , $u = 200$ , and use 500 DDIM steps with 2 self-recurrence iterations. + +Data. We collect and process a set of 3,015 videos of natural scenes exhibiting oscillatory motions from online sources our own captures. We withhold $10\%$ of videos for testing and use the remainder for training. To extract ground truth motion trajectories, we apply a coarse-to-fine flow method [10, 61] between each selected starting image and every future frame of the video. As training data, we take every 10th video frame as input images and derive the corresponding ground truth spectral volumes using the computed motion trajectories across the following 149 frames. In total, our data consists of over 150K image-motion pairs. + +![](images/efc2a8f8a62808bfbc24690cd316a86b0c962bc313f6cbd97826670372d514cf.jpg) + +![](images/687b4757f4f7ff8f19ece485a333350c452fc6d5c64eb4fbcc8c046edab9c00c.jpg) + +![](images/7839a926526e3e4dd3a6af660f6a59596f3b2f4d27111727c750b3aacf0b78ef.jpg) + +![](images/5b28f4e228885a1d8e872a723429e8503719246c835cfa996b23a72fda71a388.jpg) + +![](images/8145382f33d8ec80598737642c71212ac9f35bf8987751b375c4cfd7fb429695.jpg) + +![](images/9ec4504c6cb8a6f632d254066af28657078a4fead1f3d0b52dcd4bf665bee686.jpg) + +![](images/7dc991b385b536cee35167fb8d7974d35270558ad365c6855f1f8fefb7d8c4c0.jpg) +Input image + +![](images/d99e174881d2b858ec871d89b560ce036d391c4194bf134bdbd875d6a53ba017.jpg) +Reference +Stochastic-I2V [27] +MCVD[92] +Endo et al. [29] +Holynski et al. [46]. +Ours +Figure 5. X - t slices of videos generated by different approaches. From left to right: input image and corresponding X - t video slices from the ground truth video, from videos generated by three baselines [27, 29, 46, 92], and finally videos generated by our approach. + +![](images/878d42b430e0dff41b50c1a4bd46edb1567a538fbaae11017e6b474aee4599d7.jpg) +Figure 6. Sliding window FID and DTFVD. We show sliding window FID with window size 30 frames, and DTFVD with size 16 frames, for videos generated by different methods. + +![](images/74230c64b94baadab59b4863b66b451dff1512f6d155c3d062f9c5de1350609c.jpg) + +Baselines. We compare our approach to recent single-image animation and video prediction methods. Endo et al. [29] and DMVFN [48] predict instantaneous 2D motion fields and render future frames auto-regressively. Holynski et al. [46] instead simulates motion through a single static Eulerian motion description. Other recent work such as Stochastic Image-to-Video (Stochastic-I2V) [27], TATS [35], and MCVD [92] adopt VAEs, transformers, or diffusion models to directly predict raw video frames; LFDM [66] generates future frames by predicting flow volumes and warping latents in a diffusion model. We train all the above methods on our data using their respective open-source implementations. + +We evaluate the quality of the videos generated by our approach and by prior baselines in two ways. First, we evaluate the quality of individual synthesized frames using metrics designed for image synthesis tasks. We adopt the Fréchet Inception Distance (FID) [42] and Kernel Inception Distance (KID) [5] to measure the average distance between the distributions of generated frames and ground truth frames. + +Second, to evaluate the quality and temporal coherence of synthesized videos, we adopt the Fréchet Video Dis + +
MethodImage SynthesisVideo Synthesis
FIDKIDFVDFVD32DTFVDDTFVD32
Repeat I0--237.5316.75.3045.6
K=43.920.0760.378.43.128.59
K=83.950.0752.168.72.717.37
K=244.090.0848.265.12.506.94
w/o adaptive norm.4.530.0962.780.13.168.19
Independent pred.4.000.0852.571.32.707.40
Volume pred.4.740.0953.771.12.837.79
Baseline splat [46]4.250.0949.566.82.837.27
Full (K=16)4.030.0847.162.92.536.75
+ +Table 2. Ablation study. Sec. 7.3 describes each configuration. + +tance [91] with window size 16 (FVD) and 32 $(\mathrm{FVD}_{32})$ based on an I3D model [11] trained on the Human Kinetics datasets [52]. To more faithfully reflect synthesis quality for the natural oscillation motions we seek to generate, we also adopt the Dynamic Texture Frechet Video Distance [27], which measures distance from videos with window size 16 (DTFVD) and size 32 $(\mathrm{DTFVD}_{32})$ , using a I3D model trained on the Dynamic Textures Database [37], a dataset consisting primarily of natural motion textures. + +We further use a sliding window FID of window size 30 frames, and a sliding window DTFVD with window size 16 frames, as in [57, 60], to measure how generated video quality degrades over time. For all methods, we evaluate metrics at $256 \times 128$ resolution by center-cropping. + +# 7.1.Quantitative results + +Table 1 shows quantitative comparisons between our approach and baselines on our test set. Our approach signifi- + +![](images/1e82f1326227507e720cca0dca4200e264301081fbda44457fe31c861c4cb02a.jpg) +Figure 7. We show generated future frames from three recent large video diffusion models [31, 36, 97]. + +cantly outperforms prior single-image animation baselines in terms of both image and video synthesis quality. Specifically, our much lower FVD and DT-FVD distances suggest that the videos generated by our approach are more realistic and more temporally coherent. Further, Fig. 6 shows the sliding window FID and sliding window DT-FVD distances of generated videos from different methods. Thanks to the global spectral volume representation, videos generated by our approach do not suffer from degradation over time. + +# 7.2. Qualitative results + +We visualize qualitative comparisons between videos as spatio-temporal $X$ - $t$ slices of the generated videos, a standard way of visualizing small motions in a video [94]. As shown in Fig. 5, our generated video dynamics more strongly resemble the motion patterns observed in the corresponding real reference videos (second column), compared to other methods. Baselines such as Stochastic I2V [27] and MCVD [92] fail to model both appearance and motion realistically over time. Endo et al. [29] and Holynski et al. [46] produce video frames with fewer artifacts but exhibits oversmooth or non-oscillatory motion over time. We refer readers to supplementary material to assess the quality of generated video frames and estimated motion across different methods. + +# 7.3. Ablation study + +We conduct an ablation study to validate the major design choices in our motion prediction and rendering modules, comparing our full configuration with different variants. Specifically, we evaluate results using different numbers of frequency bands $K = 4, 8, 16, 24$ . We observe that increasing the number of frequency bands improves video prediction quality, but the improvement is marginal with more than 16 frequencies. Next, we remove adaptive frequency normalization from the ground truth spectral volumes, and instead just scale them based on input image width and height (w/o adaptive norm.). Additionally, we remove the frequency coordinated-denoising module (Independent pred.), or replace it with a simpler DM where a tensor volume of $4K$ channel spectral volumes are predicted jointly via a single 2D U-net diffusion model (Volume pred.). Finally, we compare results where we use a baseline rendering method that applies softmax splatting over single-scale features subject to learnable weights as used by [46] (Baseline splat). We also add a baseline where the generated video is a volume by + +![](images/fade8472536fcf69e69f99edd72d2d97d82ea8a2b72a746f509115affa77424b.jpg) +Figure 8. Limitations. We show examples of rendered future frames (even), and overlay of input and rendered images (odd). Our method can produce artifacts in regions of thin objects or large motions, and regions requiring filling large amount of new contents. + +repeating input image $N$ times (Repeat $I_0$ ). From Table 2, we observe that all simpler or alternative configurations lead to worse performance compared with our full model. + +# 7.4. Comparing to large video models + +We further perform a user study and compare our generated animations with ones from recent large video diffusion models:AnimateDiff [36], ModelScope [97] and Gen-2 [31], which predict video volumes directly. On a randomly selected 30 videos from the test set, we ask users "which video is more realistic?". Users report a $80.9\%$ preference for our approach over others. Moreover, as shown in Fig. 7, we observe that the generated videos from these baselines are either unable to adhere to the input image content, or exhibit gradual color drift and distortion over time. We refer readers to the supplementary material for full comparisons. + +# 8. Discussion and conclusion + +Limitations. Since our approach only predicts lower frequencies of a spectral volume, it can fail to model non-oscillating motions or high-frequency vibrations—this may be resolved by using learned motion bases. Furthermore, the quality of generated videos relies on the quality of underlying motion trajectories, which may degrade in scenes with thin moving objects or objects with large displacements. Even if correct, motions that require generating large amounts of new unseen content may also cause degradations (Fig. 8). + +Conclusion. We present a new approach for modeling natural oscillation dynamics from a single still picture. Our image-space motion prior is represented with spectral volumes, a frequency representation of per-pixel motion trajectories, which we find to be efficient and effective for prediction with diffusion models, and which we learn from collections of real world videos. Spectral volumes are predicted using frequency-coordinated latent diffusion model and are used to animate future video frames through an image-based rendering module. We show that our approach produces photo-realistic animations from a single picture and significantly outperforms prior baselines, and that it can enable several downstream applications such as creating seamlessly looping or interactive image dynamics. + +Acknowledgements. We thank Abe Davis, Rick Szeliski, Andrew Liu, Boyang Deng, Qianqian Wang, Xuan Luo, and Lucy Chai for fruitful discussions and helpful comments. + +# References + +[1] Aseem Agarwala, Ke Colin Zheng, Chris Pal, Maneesh Agrawala, Michael Cohen, Brian Curless, David Salesin, and Richard Szeliski. Panoramic video textures. In ACM SIGGRAPH 2005 Papers, pages 821-827. 2005. +[2] Hyemin Ahn, Esteve Valls Mascaro, and Dongheui Lee. Can we use diffusion probabilistic models for 3d motion prediction? arXiv preprint arXiv:2302.14503, 2023. +[3] Arpit Bansal, Hong-Min Chu, Avi Schwarzschild, Soumyadip Sengupta, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Universal guidance for diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 843-852, 2023. +[4] Hugo Bertiche, Niloy J Mitra, Kuldeep Kulkarni, ChunHao P Huang, Tuanfeng Y Wang, Meysam Madadi, Sergio Escalera, and Duygu Ceylan. Blowing in the wind: Cyclenet for human cinemagrams from still images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 459-468, 2023. +[5] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying MMD GANs. arXiv preprint arXiv:1801.01401, 2018. +[6] Andreas Blattmann, Timo Milbich, Michael Dorkenwald, and Björn Ommer. ipoke: Poking a still image for controlled stochastic video synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14707-14717, 2021. +[7] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. +[8] Richard Strong Bowen, Richard Tucker, Ramin Zabih, and Noah Snavely. Dimensions of motion: Monocular prediction through flow subspaces. In 2022 International Conference on 3D Vision (3DV), pages 454-464. IEEE, 2022. +[9] Tim Brooks, Janne Hellsten, Miika Aittala, Ting-Chun Wang, Timo Aila, Jaakko Lehtinen, Ming-Yu Liu, Alexei Efros, and Tero Karras. Generating long videos of dynamic scenes. Advances in Neural Information Processing Systems, 35:31769-31781, 2022. +[10] Thomas Brox, Andrés Bruhn, Nils Papenberg, and Joachim Weickert. High accuracy optical flow estimation based on a theory for warping. In Proc. European Conf. on Computer Vision (ECCV), pages 25-36. Springer, 2004. +[11] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. +[12] Dan Casas, Marco Volino, John Collomosse, and Adrian Hilton. 4d video textures for interactive character appearance. In Computer Graphics Forum, volume 33, pages 371-380. Wiley Online Library, 2014. +[13] Antoni B Chan and Nuno Vasconcelos. Mixtures of dynamic textures. In Tenth IEEE International Conference on Computer Vision (ICCV'05) Volume 1, volume 1, pages 641-647. IEEE, 2005. + +[14] Antoni B Chan and Nuno Vasconcelos. Classifying video with kernel dynamic textures. In 2007 IEEE Conference on Computer Vision and Pattern Recognition, pages 1-6. IEEE, 2007. +[15] Antoni B Chan and Nuno Vasconcelos. Modeling, clustering, and segmenting video with mixtures of dynamic textures. IEEE transactions on pattern analysis and machine intelligence, 30(5):909-926, 2008. +[16] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. +[17] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11315-11325, 2022. +[18] Tsai-Shien Chen, Chieh Hubert Lin, Hung-Yu Tseng, Tsung-Yi Lin, and Ming-Hsuan Yang. Motion-conditioned diffusion model for controllable video synthesis. arXiv preprint arXiv:2304.14404, 2023. +[19] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18000-18010, 2023. +[20] Yung-Yu Chuang, Dan B Goldman, Ke Colin Zheng, Brian Curless, David H Salesin, and Richard Szeliski. Animating pictures with stochastic motion textures. In ACM SIGGRAPH 2005 Papers, pages 853-860. 2005. +[21] Vincent C Couture, Michael S Langer, and Sebastien Roy. Omnistereo video textures without ghosting. In 2013 International Conference on 3D Vision-3DV 2013, pages 64–70. IEEE, 2013. +[22] Abe Davis, Justin G Chen, and Frédo Durand. Image-space modal bases for plausible manipulation of objects in video. ACM Transactions on Graphics (TOG), 34(6):1-7, 2015. +[23] Myers Abraham Davis. Visual vibration analysis. PhD thesis, Massachusetts Institute of Technology, 2016. +[24] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. +[25] Julien Diener, Mathieu Rodriguez, Lionel Baboud, and Lionel Reveret. Wind projection basis for real-time animation of trees. In Computer graphics forum, volume 28, pages 533-540. Wiley Online Library, 2009. +[26] Gianfranco Doretto, Alessandro Chiuso, Ying Nian Wu, and Stefano Soatto. Dynamic textures. International journal of computer vision, 51:91-109, 2003. +[27] Michael Dorkenwald, Timo Milbich, Andreas Blattmann, Robin Rombach, Konstantinos G. Derpanis, and Bjorn Ommer. Stochastic image-to-video synthesis using cinns. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3742-3753, June 2021. +[28] Yuming Du, Robin Kips, Albert Pumarola, Sebastian Starke, Ali Thabet, and Artsiom Sanakoyeu. Avatars grow legs: Generating smooth human motion from sparse tracking inputs with diffusion model. In Proceedings of the IEEE/CVF + +Conference on Computer Vision and Pattern Recognition, pages 481-490, 2023. +[29] Yuki Endo, Yoshihiro Kanamori, and Shigeru Kuriyama. Animating landscape: Self-supervised learning of decoupled motion and appearance for single-image video synthesis. ACM Transactions on Graphics (Proceedings of ACM SIGGRAPH Asia 2019), 38(6):175:1-175:19, 2019. +[30] Dave Epstein, Allan Jabri, Ben Poole, Alexei A Efros, and Aleksander Holynski. Diffusion self-guidance for controllable image generation. arXiv preprint arXiv:2306.00986, 2023. +[31] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7346-7356, 2023. +[32] Matthew Flagg, Atsushi Nakazawa, Qiushuang Zhang, Sing Bing Kang, Young Kee Ryu, Irfan Essa, and James M Rehg. Human video textures. In Proceedings of the 2009 symposium on Interactive 3D graphics and games, pages 199-206, 2009. +[33] Jean-Yves Franceschi, Edouard Delasalles, Mickaël Chen, Sylvain Lamprier, and Patrick Gallinari. Stochastic latent residual video prediction. In International Conference on Machine Learning, pages 3233-3246. PMLR, 2020. +[34] Ruohan Gao, Bo Xiong, and Kristen Grauman. Im2Flow: Motion hallucination from static images for action recognition. In Proc. Computer Vision and Pattern Recognition (CVPR), 2018. +[35] Songwei Ge, Thomas Hayes, Harry Yang, Xi Yin, Guan Pang, David Jacobs, Jia-Bin Huang, and Devi Parikh. Long video generation with time-agnostic vqgan and timesensitive transformer. arXiv preprint arXiv:2204.03638, 2022. +[36] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. +[37] Isma Hadji and Richard P Wildes. A new large scale dynamic texture dataset with application to convnet understanding. In Proceedings of the European Conference on Computer Vision (ECCV), pages 320-335, 2018. +[38] Zekun Hao, Xun Huang, and Serge Belongie. Controllable video generation with sparse trajectories. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7854-7863, 2018. +[39] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. +[40] Yingqing He, Menghan Xia, Haoxin Chen, Xiaodong Cun, Yuan Gong, Jinbo Xing, Yong Zhang, Xintao Wang, Chao Weng, Ying Shan, et al. animate-a-story: Storytelling with retrieval-augmented video generation. arXiv preprint arXiv:2307.06940, 2023. +[41] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2022. + +[42] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. +[43] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. +[44] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. +[45] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. +[46] Aleksander Holynski, Brian L Curless, Steven M Seitz, and Richard Szeliski. Animating pictures with Eulerian motion fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5810-5819, 2021. +[47] Tobias Hoppe, Arash Mehrjou, Stefan Bauer, Didrik Nielsen, and Andrea Dittadi. Diffusion models for video prediction and infilling. Trans. Mach. Learn. Res., 2022, 2022. +[48] Xiaotao Hu, Zhewei Huang, Ailin Huang, Jun Xu, and Shuchang Zhou. A dynamic multi-scale voxel flow network for video prediction. ArXiv, abs/2303.09875, 2023. +[49] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016. +[50] Hitoshi Kanda and Jun Ohya. Efficient, realistic method for animating dynamic behaviors of 3d botanical trees. In 2003 International Conference on Multimedia and Expo. ICME'03. Proceedings (Cat. No. 03TH8698), volume 2, pages II-89. IEEE, 2003. +[51] Johanna Karras, Aleksander Holynski, Ting-Chun Wang, and Ira Kemelmacher-Shlizerman. Dreampose: Fashion image-to-video synthesis via stable diffusion. arXiv preprint arXiv:2304.06025, 2023. +[52] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. +[53] Alex X Lee, Richard Zhang, Frederik Ebert, Pieter Abbeel, Chelsea Finn, and Sergey Levine. Stochastic adversarial video prediction. arXiv preprint arXiv:1804.01523, 2018. +[54] Zhengqi Li, Tali Dekel, Forrester Cole, Richard Tucker, Noah Snavely, Ce Liu, and William T Freeman. Learning the depths of moving people by watching frozen people. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4521-4530, 2019. +[55] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2041-2050, 2018. +[56] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snively. Dynibar: Neural dynamic image-based + +rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4273-4284, 2023. +[57] Zhengqi Li, Qianqian Wang, Noah Snavely, and Angjoo Kanazawa. Infinitenature-zero: Learning perpetual view generation of natural scenes from single images. In European Conference on Computer Vision, pages 515-534. Springer, 2022. +[58] Jing Liao, Mark Finch, and Hugues Hoppe. Fast computation of seamless video loops. ACM Transactions on Graphics (TOG), 34(6):1-10, 2015. +[59] Zicheng Liao, Neel Joshi, and Hugues Hoppe. Automated video looping with progressive dynamism. ACM Transactions on Graphics (TOG), 32(4):1-10, 2013. +[60] Andrew Liu, Richard Tucker, Varun Jampani, Ameesh Makadia, Noah Snavely, and Angjoo Kanazawa. Infinite nature: Perpetual view generation of natural scenes from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14458-14467, 2021. +[61] Ce Liu. Beyond pixels: exploring new representations and applications for motion analysis. PhD thesis, Massachusetts Institute of Technology, 2009. +[62] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Tieniu Tan. Videofusion: Decomposed diffusion models for high-quality video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10209-10218, 2023. +[63] Aniruddha Mahapatra and Kuldeep Kulkarni. Controllable animation of fluid elements in still images. In Proc. Computer Vision and Pattern Recognition (CVPR), 2022. +[64] Arun Mallya, Ting-Chun Wang, and Ming-Yu Liu. Implicit warping for animation with image sets. Advances in Neural Information Processing Systems, 35:22438-22450, 2022. +[65] Medhini Narasimhan, Shiry Ginosar, Andrew Owens, Alexei A Efros, and Trevor Darrell. Strumming to the beat: Audio-conditioned contrastive video textures. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3761-3770, 2022. +[66] Haomiao Ni, Changhao Shi, Kai Li, Sharon X Huang, and Martin Renqiang Min. Conditional image-to-video generation with latent flow diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18444-18455, 2023. +[67] Simon Niklaus and Feng Liu. Softmax splatting for video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5437-5446, 2020. +[68] Shin Ota, Machiko Tamura, Kunihiko Fujita, T Fujimoto, K Muraoka, and Norishige Chiba. 1/f/sup/spl beta//noise-based real-time animation of trees swaying in wind fields. In Proceedings Computer Graphics International 2003, pages 52-59. IEEE, 2003. +[69] Automne Petitjean, Yohan Poirier-Ginter, Ayush Tewari, Guillaume Cordonnier, and George Drettakis. Modalnerf: Neural modal analysis and synthesis for free-viewpoint navigation in dynamically vibrating scenes. In Computer Graphics Forum, volume 42, 2023. +[70] Silvia L. Pintea, Jan C. van Gemert, and Arnold W. M. + +Smeulders. Déjà vu: Motion prediction in static images. In Proc. European Conf. on Computer Vision (ECCV), 2014. +[71] Sigal Raab, Inbal Leibovitch, Guy Tevet, Moab Arar, Amit H Bermano, and Daniel Cohen-Or. Single motion diffusion. arXiv preprint arXiv:2302.05905, 2023. +[72] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2):3, 2022. +[73] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. +[74] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. +[75] Payam Saisan, Gianfranco Doretto, Ying Nian Wu, and Stefano Soatto. Dynamic texture recognition. In Proceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition. CVPR 2001, volume 2, pages II-II. IEEE, 2001. +[76] Saurabh Saxena, Charles Herrmann, Junhwa Hur, Abhishek Kar, Mohammad Norouzi, Deqing Sun, and David J. Fleet. The surprising effectiveness of diffusion models for optical flow and monocular depth estimation, 2023. +[77] Arno Schödl, Richard Szeliski, David H Salesin, and Irfan Essa. Video textures. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 489-498, 2000. +[78] Mikio Shinya and Alain Fournier. Stochastic motion—motion under the influence of wind. Computer Graphics Forum, 11(3), 1992. +[79] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. Animating arbitrary objects via deep motion transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2377-2386, 2019. +[80] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. Advances in neural information processing systems, 32, 2019. +[81] Aliaksandr Siarohin, Oliver J Woodford, Jian Ren, Mengei Chai, and Sergey Tulyakov. Motion representations for articulated animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13653-13662, 2021. +[82] Chen Qian Kwan-Yee Lin Hongsheng Li Siming Fan, Jingtan Piao. Simulating fluids in real-world still images. arXiv preprint, arXiv:2204.11335, 2022. +[83] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3626-3636, 2022. +[84] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, + +and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. +[85] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv:2010.02502, October 2020. +[86] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. +[87] Jos Stam. Multi-scale stochastic modelling of complex natural phenomena. PhD thesis, 1995. +[88] Jos Stam. Stochastic dynamics: Simulating the effects of turbulence on flexible structures. Computer Graphics Forum, 16(3), 1997. +[89] Ryusuke Sugimoto, Mingming He, Jing Liao, and Pedro V Sander. Water simulation and rendering from a still photograph. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. +[90] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. +[91] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. +[92] Vikram Voleti, Alexia Jolicoeur-Martineau, and Christopher Pal. Mcvd: Masked conditional video diffusion for prediction, generation, and interpolation. In (NeurIPS) Advances in Neural Information Processing Systems, 2022. +[93] Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba. Generating videos with scene dynamics. In Neural Information Processing Systems, 2016. +[94] Neal Wadhwa, Michael Rubinstein, Frédo Durand, and William T Freeman. Phase-based video motion processing. ACM Transactions on Graphics (ToG), 32(4):1-10, 2013. +[95] Jacob Walker, Carl Doersch, Abhinav Gupta, and Martial Hebert. An uncertain future: Forecasting from static images using variational autoencoders. In Proc. European Conf. on Computer Vision (ECCV), 2016. +[96] Jacob Walker, Abhinav Gupta, and Martial Hebert. Dense optical flow prediction from a static image. In Proceedings of the IEEE International Conference on Computer Vision, pages 2443-2451, 2015. +[97] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. arXiv preprint arXiv:2306.02018, 2023. +[98] Yaohui Wang, Di Yang, Francois Bremond, and Antitza Dantcheva. Latent image animator: Learning to animate images via latent space navigation. arXiv preprint arXiv:2203.09043, 2022. +[99] Frederik Warburg, Ethan Weber, Matthew Tancik, Aleksander Holynski, and Angjoo Kanazawa. Nerfbusters: Removing ghostly artifacts from casually captured nerfs. arXiv preprint arXiv:2304.10532, 2023. +[100] Daniel Watson, William Chan, Ricardo Martin-Brualla, + +Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. arXiv preprint arXiv:2210.04628, 2022. +[101] Chung-Yi Weng, Brian Curless, and Ira Kemelmacher-Shlizerman. Photo wake-up: 3d character animation from a single photo. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5908-5917, 2019. +[102] Jamie Wynn and Daniyar Turmukhambetov. DiffusioNeRF: Regularizing Neural Radiance Fields with Denoising Diffusion Models. In CVPR, 2023. +[103] Tianfan Xue, Jiajun Wu, Katherine L Bouman, and William T Freeman. Visual dynamics: Stochastic future generation via layered cross convolutional networks. Trans. Pattern Analysis and Machine Intelligence, 41(9):2236-2250, 2019. +[104] Shengming Yin, Chenfei Wu, Jian Liang, Jie Shi, Houqiang Li, Gong Ming, and Nan Duan. Dragnuwa: Fine-grained control in video generation by integrating text, image, and trajectory. arXiv preprint arXiv:2308.08089, 2023. +[105] Sihyun Yu, Kihyuk Sohn, Subin Kim, and Jinwoo Shin. Video probabilistic diffusion models in projected latent space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18456-18466, 2023. +[106] Mingyuan Zhang, Xinying Guo, Liang Pan, Zhongang Cai, Fangzhou Hong, Huirong Li, Lei Yang, and Ziwei Liu. Remodiffuse: Retrieval-augmented motion diffusion model. arXiv preprint arXiv:2304.01116, 2023. +[107] Yabo Zhang, Yuxiang Wei, Dongsheng Jiang, Xiaopeng Zhang, Wangmeng Zuo, and Qi Tian. Controlvideo: Training-free controllable text-to-video generation. arXiv preprint arXiv:2305.13077, 2023. +[108] Jian Zhao and Hui Zhang. Thin-plate spline motion model for image animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3657–3666, 2022. +[109] Shengyu Zhao, Jonathan Cui, Yilun Sheng, Yue Dong, Xiao Liang, Eric I Chang, and Yan Xu. Large scale image completion via co-modulated generative adversarial networks. In International Conference on Learning Representations (ICLR), 2021. +[110] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. \ No newline at end of file diff --git a/2024/Generative Image Dynamics/images.zip b/2024/Generative Image Dynamics/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0c334e44dfa9ecc7cf8b9cc738e74a95259f2b93 --- /dev/null +++ b/2024/Generative Image Dynamics/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4acfad54a5e58705967889bc07d9d678f76400582c875d5bd773db56dde4770b +size 481388 diff --git a/2024/Generative Image Dynamics/layout.json b/2024/Generative Image Dynamics/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f8c34075d56bd97af625b7c092f0086d1f850b04 --- /dev/null +++ b/2024/Generative Image Dynamics/layout.json @@ -0,0 +1,11876 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 209, + 103, + 385, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 103, + 385, + 121 + ], + "spans": [ + { + "bbox": [ + 209, + 103, + 385, + 121 + ], + "type": "text", + "content": "Generative Image Dynamics" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 143, + 164, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 143, + 164, + 159 + ], + "spans": [ + { + "bbox": [ + 107, + 143, + 164, + 159 + ], + "type": "text", + "content": "Zhengqi Li" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 187, + 143, + 264, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 143, + 264, + 157 + ], + "spans": [ + { + "bbox": [ + 187, + 143, + 264, + 157 + ], + "type": "text", + "content": "Richard Tucker" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 288, + 144, + 358, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 288, + 144, + 358, + 157 + ], + "spans": [ + { + "bbox": [ + 288, + 144, + 358, + 157 + ], + "type": "text", + "content": "Noah Snavely" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 383, + 144, + 485, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 144, + 485, + 157 + ], + "spans": [ + { + "bbox": [ + 383, + 144, + 485, + 157 + ], + "type": "text", + "content": "Aleksander Holynski" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 244, + 163, + 328, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 163, + 328, + 177 + ], + "spans": [ + { + "bbox": [ + 244, + 163, + 328, + 177 + ], + "type": "text", + "content": "Google Research" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 60, + 258, + 202, + 338 + ], + "blocks": [ + { + "bbox": [ + 60, + 258, + 202, + 338 + ], + "lines": [ + { + "bbox": [ + 60, + 258, + 202, + 338 + ], + "spans": [ + { + "bbox": [ + 60, + 258, + 202, + 338 + ], + "type": "image", + "image_path": "1970c681d962b8853f36ccc8ddaa928629b646be28135670f9eaa973dda86f4c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 105, + 339, + 156, + 348 + ], + "lines": [ + { + "bbox": [ + 105, + 339, + 156, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 156, + 348 + ], + "type": "text", + "content": "Input Picture" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 204, + 289, + 228, + 307 + ], + "blocks": [ + { + "bbox": [ + 204, + 289, + 228, + 307 + ], + "lines": [ + { + "bbox": [ + 204, + 289, + 228, + 307 + ], + "spans": [ + { + "bbox": [ + 204, + 289, + 228, + 307 + ], + "type": "image", + "image_path": "59ff8290e3cfb9113942f4c942cf80ea62a05c73470708eecd62a494792d00db.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 230, + 242, + 295, + 281 + ], + "blocks": [ + { + "bbox": [ + 266, + 220, + 347, + 241 + ], + "lines": [ + { + "bbox": [ + 266, + 220, + 347, + 241 + ], + "spans": [ + { + "bbox": [ + 266, + 220, + 347, + 241 + ], + "type": "text", + "content": "Spectral Volume (Ge-Space Modal Basis)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 230, + 242, + 295, + 281 + ], + "lines": [ + { + "bbox": [ + 230, + 242, + 295, + 281 + ], + "spans": [ + { + "bbox": [ + 230, + 242, + 295, + 281 + ], + "type": "image", + "image_path": "78a247a00539d0c217235990f6b0f8246232cb2fa9eb2aef343d17d5599a3585.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 297, + 242, + 363, + 281 + ], + "blocks": [ + { + "bbox": [ + 297, + 242, + 363, + 281 + ], + "lines": [ + { + "bbox": [ + 297, + 242, + 363, + 281 + ], + "spans": [ + { + "bbox": [ + 297, + 242, + 363, + 281 + ], + "type": "image", + "image_path": "5a42bd78bf68b0384b71f855a8a0777cec05b33cfd30a08f546469f3abb83c7d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 230, + 282, + 295, + 320 + ], + "blocks": [ + { + "bbox": [ + 230, + 282, + 295, + 320 + ], + "lines": [ + { + "bbox": [ + 230, + 282, + 295, + 320 + ], + "spans": [ + { + "bbox": [ + 230, + 282, + 295, + 320 + ], + "type": "image", + "image_path": "34db33c57aca6d04fb45265e6306b77ab81015b48f000e0d3b5a555a59ad3b49.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 297, + 281, + 363, + 320 + ], + "blocks": [ + { + "bbox": [ + 297, + 281, + 363, + 320 + ], + "lines": [ + { + "bbox": [ + 297, + 281, + 363, + 320 + ], + "spans": [ + { + "bbox": [ + 297, + 281, + 363, + 320 + ], + "type": "image", + "image_path": "455907a99362c2f7ae7a8d16b5365affcae8744b009709fcfea82bbcd5348b35.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 258, + 325, + 264, + 336 + ], + "blocks": [ + { + "bbox": [ + 258, + 325, + 264, + 336 + ], + "lines": [ + { + "bbox": [ + 258, + 325, + 264, + 336 + ], + "spans": [ + { + "bbox": [ + 258, + 325, + 264, + 336 + ], + "type": "image", + "image_path": "7a85c481136ceda92f81229273fb6140ebf9894317f442717658aa20a53d3f1b.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 326, + 325, + 332, + 337 + ], + "blocks": [ + { + "bbox": [ + 326, + 325, + 332, + 337 + ], + "lines": [ + { + "bbox": [ + 326, + 325, + 332, + 337 + ], + "spans": [ + { + "bbox": [ + 326, + 325, + 332, + 337 + ], + "type": "image", + "image_path": "d19703c80e58541080695e33803a4edf3861d9c7d7bba0f397bec0177c4387b5.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 230, + 343, + 295, + 381 + ], + "blocks": [ + { + "bbox": [ + 230, + 343, + 295, + 381 + ], + "lines": [ + { + "bbox": [ + 230, + 343, + 295, + 381 + ], + "spans": [ + { + "bbox": [ + 230, + 343, + 295, + 381 + ], + "type": "image", + "image_path": "ac2065fee6863ad784779beaccf133f097476adc78b8661117189f30e348e101.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 242, + 383, + 283, + 390 + ], + "lines": [ + { + "bbox": [ + 242, + 383, + 283, + 390 + ], + "spans": [ + { + "bbox": [ + 242, + 383, + 283, + 390 + ], + "type": "text", + "content": "X coefficients" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 297, + 343, + 363, + 381 + ], + "blocks": [ + { + "bbox": [ + 297, + 343, + 363, + 381 + ], + "lines": [ + { + "bbox": [ + 297, + 343, + 363, + 381 + ], + "spans": [ + { + "bbox": [ + 297, + 343, + 363, + 381 + ], + "type": "image", + "image_path": "6ca5b677fbf8d3e4bf3bec7c9223c52dee59e51042843b48f57d2890205a9704.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 383, + 350, + 390 + ], + "lines": [ + { + "bbox": [ + 310, + 383, + 350, + 390 + ], + "spans": [ + { + "bbox": [ + 310, + 383, + 350, + 390 + ], + "type": "text", + "content": "Y coefficients" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 395, + 220, + 531, + 301 + ], + "blocks": [ + { + "bbox": [ + 395, + 220, + 531, + 301 + ], + "lines": [ + { + "bbox": [ + 395, + 220, + 531, + 301 + ], + "spans": [ + { + "bbox": [ + 395, + 220, + 531, + 301 + ], + "type": "image", + "image_path": "37c84aaa5014cb7cd271e56cfcb527993742f81f69768eafa7ac4fd800fdb55f.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 396, + 304, + 531, + 386 + ], + "blocks": [ + { + "bbox": [ + 396, + 304, + 531, + 386 + ], + "lines": [ + { + "bbox": [ + 396, + 304, + 531, + 386 + ], + "spans": [ + { + "bbox": [ + 396, + 304, + 531, + 386 + ], + "type": "image", + "image_path": "6f14a0e77e624585c13fd3df632bca6bfc17717dbd0a100efa2417ab8765a6cb.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 390, + 547, + 436 + ], + "lines": [ + { + "bbox": [ + 45, + 390, + 547, + 436 + ], + "spans": [ + { + "bbox": [ + 45, + 390, + 547, + 436 + ], + "type": "text", + "content": "Figure 1. We model a generative image-space prior on scene motion: from a single RGB image, our method generates a spectral volume [23], a motion representation that models dense, long-term pixel trajectories in the Fourier domain. Our learned motion priors can be used to turn a single picture into a seamlessly looping video, or into an interactive simulation of dynamics that responds to user inputs like dragging and releasing points. On the right, we visualize output videos as space-time " + }, + { + "bbox": [ + 45, + 390, + 547, + 436 + ], + "type": "inline_equation", + "content": "X - t" + }, + { + "bbox": [ + 45, + 390, + 547, + 436 + ], + "type": "text", + "content": " slices (along the input scanline shown on the left)." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "bbox": [ + 143, + 449, + 192, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 449, + 192, + 462 + ], + "spans": [ + { + "bbox": [ + 143, + 449, + 192, + 462 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 45, + 485, + 290, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 485, + 290, + 701 + ], + "spans": [ + { + "bbox": [ + 45, + 485, + 290, + 701 + ], + "type": "text", + "content": "We present an approach to modeling an image-space prior on scene motion. Our prior is learned from a collection of motion trajectories extracted from real video sequences depicting natural, oscillatory dynamics of objects such as trees, flowers, candles, and clothes swaying in the wind. We model dense, long-term motion in the Fourier domain as spectral volumes, which we find are well-suited to prediction with diffusion models. Given a single image, our trained model uses a frequency-coordinated diffusion sampling process to predict a spectral volume, which can be converted into a motion texture that spans an entire video. Along with an image-based rendering module, the predicted motion representation can be used for a number of downstream applications, such as turning still images into seamlessly looping videos, or allowing users to interact with objects in real images, producing realistic simulated dynamics (by interpreting the spectral volumes as image-space modal bases). See our project page for more results: generative-dynamics.github.io." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 449, + 386, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 386, + 462 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 386, + 462 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 472, + 547, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 547, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 547, + 544 + ], + "type": "text", + "content": "The natural world is always in motion, with even seemingly static scenes containing subtle oscillations as a result of wind, water currents, respiration, or other natural rhythms. Emulating this motion is crucial in visual content synthesis—human sensitivity to motion can cause imagery without motion (or with slightly unrealistic motion) to seem uncanny or unreal." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 545, + 548, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 548, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 548, + 665 + ], + "type": "text", + "content": "While it is easy for humans to interpret or imagine motion in scenes, training a model to learn or produce realistic scene motion is far from trivial. The motion we observe in the world is the result of a scene's underlying physical dynamics, i.e., forces applied to objects that respond according to their unique physical properties—their mass, elasticity, etc—quantities that are hard to measure and capture at scale. Fortunately, measuring them is not necessary for certain applications: e.g., one can simulate plausible dynamics in a scene by simply analyzing some observed 2D motion [23]." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 666, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 547, + 715 + ], + "type": "text", + "content": "This same observed motion can also serve as a supervisory signal in learning dynamics across scenes—because although observed motion is multi-modal and grounded in complex physical effects, it is nevertheless often predictable:" + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24142" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 167 + ], + "type": "text", + "content": "candles will flicker in certain ways, trees will sway, and their leaves will rustle. As humans, this predictability is ingrained in our systems of perception: by viewing a still image, we can imagine plausible motions—or, since there might have been many possible such motions, a distribution of natural motions conditioned on that image. Given the facility with which humans are able to model these distributions, a natural research problem is to model them computationally." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 170, + 288, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 170, + 288, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 170, + 288, + 277 + ], + "type": "text", + "content": "Recent advances in generative models, in particular conditional diffusion models [44, 84, 86], have enabled us to model rich distributions, including distributions of real images conditioned on text [72-74]. This capability has enabled several new applications, such as text-conditioned generation of diverse and realistic image content. Following the success of these image models, recent work has extended these models to other domains, such as videos [7, 43] and 3D geometry [76, 99, 100, 102]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 279, + 288, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 279, + 288, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 279, + 288, + 495 + ], + "type": "text", + "content": "In this paper, we model a generative prior for image-space scene motion, i.e., the motion of all pixels in a single image. This model is trained on motion trajectories automatically extracted from a large collection of real video sequences. In particular, from each training video we compute motion in the form of a spectral volume [22, 23], a frequency-domain representation of dense, long-range pixel trajectories. Spectral volumes are well-suited to scenes that exhibit oscillatory dynamics, e.g., trees and flowers moving in the wind. We find that this representation is also highly effective as an output of a diffusion model for modeling scene motion. We train a generative model that, conditioned on a single image, can sample spectral volumes from its learned distribution. A predicted spectral volume can then be directly transformed into a motion texture—a set of long-range, per-pixel motion trajectories—that can be used to animate the image. The spectral volume can also be interpreted as an image-space modal basis for use in simulating interactive dynamics [22]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 496, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 496, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 496, + 287, + 581 + ], + "type": "text", + "content": "We predict spectral volumes from input images using a diffusion model that generates coefficients one frequency at a time, but coordinates these predictions across frequency bands through a shared attention module. The predicted motions can be used to synthesize future frames (via an image-based rendering model)—turning still images into realistic animations, as illustrated in Fig. 1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": "Compared with priors over raw RGB pixels, priors over motion capture more fundamental, lower-dimensional structure that efficiently explains long-range variations in pixel values. Hence, generating intermediate motion leads to more coherent long-term generation and more fine-grained control over animations. We demonstrate the use of our trained model in several downstream applications, such as creating seamless looping videos, editing the generated motions, and enabling interactive dynamic images via image-space modal bases, i.e., simulating the response of object dynamics to user-applied forces [22]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 91, + 547, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 547, + 224 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 547, + 224 + ], + "type": "text", + "content": "Generative synthesis. Recent advances in generative models have enabled photorealistic synthesis of images conditioned on text prompts [16, 17, 24, 72-74]. These text-to-image models can be augmented to synthesize video sequences by extending the generated image tensors along a time dimension [7, 9, 43, 62, 83, 105, 105, 110]. While these methods can produce video sequences that capture the spatiotemporal statistics of real footage, these videos often suffer from artifacts like incoherent motion, unrealistic temporal variation in textures, and violations of physical constraints like preservation of mass." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 228, + 548, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 548, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 548, + 456 + ], + "type": "text", + "content": "Animating images. Instead of generating videos entirely from text, other techniques take as input a still picture and animate it. Many recent deep learning methods adopt a 3D-Unet architecture to produce video volumes directly [27, 36, 40, 47, 53, 92]. These models are effectively the same video generation models (but conditioned on image information instead of text), and exhibit similar artifacts to those mentioned above. One way to overcome these limitations is to not directly generate the video content itself, but instead animate an input source image through image-based rendering, i.e., moving the image content around according to motion derived from external sources such as a driving video [51, 79-81, 98], motion or 3D geometry priors [8, 29, 46, 63, 64, 66, 89, 96, 100, 101, 103, 108], or user annotations [6, 18, 20, 33, 38, 97, 104, 107]. Animating images according to motion fields yields greater temporal coherence and realism, but these prior methods either require additional guidance signals or user input, or utilize limited motion representations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 461, + 547, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 547, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 547, + 665 + ], + "type": "text", + "content": "Motion models and motion priors. In computer graphics, natural, oscillatory 3D motion (e.g., water rippling or trees waving in the wind) can be modeled with noise that is shaped in the Fourier domain and then converted to time-domain motion fields [78, 87]. Some of these methods rely on a modal analysis of the underlying dynamics of the system being simulated [22, 25, 88]. These spectral techniques were adapted to animate plants, water, and clouds from single 2D pictures by Chuang et al. [20], given user annotations. Our work is especially inspired by Davis [23], who connected modal analysis of a scene with the motions observed in a video of that scene, and used this analysis to simulate interactive dynamics from a video. We adopt the frequency-space spectral volume motion representation from Davis et al., extract this representation from a large set of training videos, and show that spectral volumes are suitable for predicting motion from single images with diffusion models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "Other methods have used various motion representations in prediction tasks, where an image or video is used to inform a deterministic future motion estimate [34, 70], or a more rich distribution of possible motions [93, 95, 103]. However," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24143" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": "many of these methods predict an optical flow motion estimate (i.e., the instantaneous motion of each pixel), not full motion trajectories. In addition, much of this prior work is focused on tasks like activity recognition, not on synthesis tasks. More recent work has demonstrated the advantages of modeling and predicting motion using generative models in a number of closed-domain settings such as humans and animals [2, 19, 28, 71, 90, 106]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 177, + 289, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 177, + 289, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 177, + 289, + 357 + ], + "type": "text", + "content": "Videos as textures. Certain moving scenes can be thought of as a kind of texture—termed dynamic textures [26]—that model videos as space-time samples of a stochastic process. Dynamic textures can represent smooth, natural motions like waves, flames, or moving trees, and have been widely used for video classification, segmentation or encoding [12-15, 75]. A related kind of texture, called a video texture, represents a moving scene as a set of input video frames along with transition probabilities between any pair of frames [65, 77]. A number of methods estimate dynamic or video textures through analysis of scene motion and pixel statistics, with the aim of generating seamlessly looping or infinitely varying output videos [1, 21, 32, 58, 59, 77]. In contrast to much of this work, our method learns priors in advance that can then be applied to single images." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 371, + 111, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 371, + 111, + 384 + ], + "spans": [ + { + "bbox": [ + 47, + 371, + 111, + 384 + ], + "type": "text", + "content": "3. Overview" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 393, + 287, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 393, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 46, + 393, + 287, + 525 + ], + "type": "text", + "content": "Given a single picture " + }, + { + "bbox": [ + 46, + 393, + 287, + 525 + ], + "type": "inline_equation", + "content": "I_0" + }, + { + "bbox": [ + 46, + 393, + 287, + 525 + ], + "type": "text", + "content": ", our goal is to generate a video " + }, + { + "bbox": [ + 46, + 393, + 287, + 525 + ], + "type": "inline_equation", + "content": "\\{\\hat{I}_1, \\hat{I}_2, \\dots, \\hat{I}_T\\}" + }, + { + "bbox": [ + 46, + 393, + 287, + 525 + ], + "type": "text", + "content": " featuring oscillatory motions such as those of trees, flowers, or candle flames swaying in the breeze. Our system consists of two modules: a motion prediction module and an image-based rendering module. Our pipeline begins by using a latent diffusion model (LDM) to predict a spectral volume " + }, + { + "bbox": [ + 46, + 393, + 287, + 525 + ], + "type": "inline_equation", + "content": "S = (S_{f_0}, S_{f_1}, \\dots, S_{f_{K-1}})" + }, + { + "bbox": [ + 46, + 393, + 287, + 525 + ], + "type": "text", + "content": " for the input " + }, + { + "bbox": [ + 46, + 393, + 287, + 525 + ], + "type": "inline_equation", + "content": "I_0" + }, + { + "bbox": [ + 46, + 393, + 287, + 525 + ], + "type": "text", + "content": ". The predicted spectral volume is then transformed to a motion texture " + }, + { + "bbox": [ + 46, + 393, + 287, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{F} = (F_1, F_2, \\dots, F_T)" + }, + { + "bbox": [ + 46, + 393, + 287, + 525 + ], + "type": "text", + "content": " through an inverse discrete Fourier transform. This motion determines the position of each input pixel at every future time step." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 526, + 288, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 526, + 288, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 526, + 288, + 586 + ], + "type": "text", + "content": "Given a predicted motion texture, we then animate the input RGB image using a neural image-based rendering technique (Sec. 5). We explore applications of this method, including producing seamless looping animations and simulating interactive dynamics, in Sec. 6." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 601, + 154, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 601, + 154, + 615 + ], + "spans": [ + { + "bbox": [ + 47, + 601, + 154, + 615 + ], + "type": "text", + "content": "4. Predicting motion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 621, + 174, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 174, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 174, + 635 + ], + "type": "text", + "content": "4.1. Motion representation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": "Formally, a motion texture is a sequence of time-varying 2D displacement maps " + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{F} = \\{F_t|t = 1,\\dots,T\\}" + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": ", where the 2D displacement vector " + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "F_{t}(\\mathbf{p})" + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": " at each pixel coordinate " + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": " from input image " + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "I_0" + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": " defines the position of that pixel at a future time " + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": " [20]. To generate a future frame at time " + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": ", one can sput pixels from " + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "I_0" + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": " using the corresponding displacement" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 71, + 427, + 140 + ], + "blocks": [ + { + "bbox": [ + 307, + 71, + 427, + 140 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 427, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 427, + 140 + ], + "type": "image", + "image_path": "54d1208cb540e2bb8f79089f292952a53d70139d7f5cacf10203728b31d89d51.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 151, + 547, + 251 + ], + "lines": [ + { + "bbox": [ + 304, + 151, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 151, + 547, + 251 + ], + "type": "text", + "content": "Figure 2. Left: We visualize the average power spectrum for the " + }, + { + "bbox": [ + 304, + 151, + 547, + 251 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 151, + 547, + 251 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 151, + 547, + 251 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 304, + 151, + 547, + 251 + ], + "type": "text", + "content": " motion components extracted from real videos, shown as the blue and green curves. Natural oscillation motions are composed primarily of low-frequency components, and so we use the first " + }, + { + "bbox": [ + 304, + 151, + 547, + 251 + ], + "type": "inline_equation", + "content": "K = 16" + }, + { + "bbox": [ + 304, + 151, + 547, + 251 + ], + "type": "text", + "content": " terms, marked with red dots. Right: we show a histogram of the amplitude of Fourier terms at " + }, + { + "bbox": [ + 304, + 151, + 547, + 251 + ], + "type": "inline_equation", + "content": "3.0\\mathrm{Hz}" + }, + { + "bbox": [ + 304, + 151, + 547, + 251 + ], + "type": "text", + "content": " after (1) scaling amplitude by image width and height (blue), or (2) frequency adaptive normalization (red). Our adaptive normalization prevents the coefficients from concentrating at extreme values." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 427, + 71, + 545, + 140 + ], + "blocks": [ + { + "bbox": [ + 427, + 71, + 545, + 140 + ], + "lines": [ + { + "bbox": [ + 427, + 71, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 427, + 71, + 545, + 140 + ], + "type": "image", + "image_path": "77ce6d12a9675cdf33837894d68f11fb0ca1d62780bdf8d19002f05cbe16bb42.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 268, + 504, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 268, + 504, + 281 + ], + "spans": [ + { + "bbox": [ + 305, + 268, + 504, + 281 + ], + "type": "text", + "content": "map " + }, + { + "bbox": [ + 305, + 268, + 504, + 281 + ], + "type": "inline_equation", + "content": "D_{t}" + }, + { + "bbox": [ + 305, + 268, + 504, + 281 + ], + "type": "text", + "content": ", resulting in a forward-warped image " + }, + { + "bbox": [ + 305, + 268, + 504, + 281 + ], + "type": "inline_equation", + "content": "I_t'" + }, + { + "bbox": [ + 305, + 268, + 504, + 281 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 375, + 283, + 545, + 297 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 283, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 375, + 283, + 545, + 297 + ], + "type": "interline_equation", + "content": "I _ {t} ^ {\\prime} (\\mathbf {p} + F _ {t} (\\mathbf {p})) = I _ {0} (\\mathbf {p}). \\tag {1}", + "image_path": "ccdd04f5a8108047d9e4faebb14321c9cab4c89679d8e3a1f5130fec46bfd5db.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 300, + 547, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 547, + 430 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 547, + 430 + ], + "type": "text", + "content": "If our goal is to produce a video via a motion texture, then one choice would be to predict a time-domain motion texture directly from an input image. However, the size of the motion texture would need to scale with the length of the video: generating " + }, + { + "bbox": [ + 304, + 300, + 547, + 430 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 300, + 547, + 430 + ], + "type": "text", + "content": " output frames implies predicting " + }, + { + "bbox": [ + 304, + 300, + 547, + 430 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 300, + 547, + 430 + ], + "type": "text", + "content": " displacement fields. To avoid predicting such a large output representation for long videos, many prior animation methods either generate video frames autoregressively [7, 29, 57, 60, 92], or predict each future output frame independently via an extra time embedding [4]. However, neither strategy ensures long-term temporal consistency of generated videos." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 431, + 547, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 431, + 547, + 550 + ], + "spans": [ + { + "bbox": [ + 304, + 431, + 547, + 550 + ], + "type": "text", + "content": "Fortunately, many natural motions can be described as a superposition of a small number of harmonic oscillators represented with different frequencies, amplitude and phases [20, 23, 25, 50, 68]. Because these underlying motions are quasi-periodic, it is natural to model them in the frequency domain. Hence, we adopt from Davis et al. [23] an efficient frequency space representation of motion in a video called a spectral volume, visualized in Fig. 3. A spectral volume is the temporal Fourier transform of per-pixel trajectories extracted from a video." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 551, + 547, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 551, + 547, + 695 + ], + "spans": [ + { + "bbox": [ + 304, + 551, + 547, + 695 + ], + "type": "text", + "content": "Given this motion representation, we formulate the motion prediction problem as a multi-modal image-to-image translation task: from an input image to an output motion spectral volume. We adopt latent diffusion models (LDMs) to generate spectral volumes comprised of a " + }, + { + "bbox": [ + 304, + 551, + 547, + 695 + ], + "type": "inline_equation", + "content": "4K" + }, + { + "bbox": [ + 304, + 551, + 547, + 695 + ], + "type": "text", + "content": "-channel 2D motion spectrum map, where " + }, + { + "bbox": [ + 304, + 551, + 547, + 695 + ], + "type": "inline_equation", + "content": "K \\ll T" + }, + { + "bbox": [ + 304, + 551, + 547, + 695 + ], + "type": "text", + "content": " is the number of frequencies modeled, and where at each frequency we need four scalars to represent the complex Fourier coefficients for the " + }, + { + "bbox": [ + 304, + 551, + 547, + 695 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 551, + 547, + 695 + ], + "type": "text", + "content": "- and " + }, + { + "bbox": [ + 304, + 551, + 547, + 695 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 551, + 547, + 695 + ], + "type": "text", + "content": "-dimensions. Note that the motion trajectory of a pixel at future time steps " + }, + { + "bbox": [ + 304, + 551, + 547, + 695 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\mathbf{p}) = \\{F_t(\\mathbf{p})|t = 1,2,\\dots T\\}" + }, + { + "bbox": [ + 304, + 551, + 547, + 695 + ], + "type": "text", + "content": " and its representation as a spectral volume " + }, + { + "bbox": [ + 304, + 551, + 547, + 695 + ], + "type": "inline_equation", + "content": "\\mathcal{S}(\\mathbf{p}) = \\{S_{f_k}(\\mathbf{p})|k = 0,1,\\dots \\frac{T}{2} -1\\}" + }, + { + "bbox": [ + 304, + 551, + 547, + 695 + ], + "type": "text", + "content": " are related by the Fast Fourier transform (FFT):" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 381, + 697, + 545, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 697, + 545, + 711 + ], + "spans": [ + { + "bbox": [ + 381, + 697, + 545, + 711 + ], + "type": "interline_equation", + "content": "\\mathcal {S} (\\mathbf {p}) = \\operatorname {F F T} (\\mathcal {F} (\\mathbf {p})). \\tag {2}", + "image_path": "a675f5d3303f57d683543f6cd6f4c2b1bd9779ec14a4cb82c53933ee20a0b0f8.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24144" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 71, + 534, + 223 + ], + "blocks": [ + { + "bbox": [ + 65, + 71, + 534, + 223 + ], + "lines": [ + { + "bbox": [ + 65, + 71, + 534, + 223 + ], + "spans": [ + { + "bbox": [ + 65, + 71, + 534, + 223 + ], + "type": "image", + "image_path": "5744e0ab4ea703956d9398b30c1ee12d84e4d34f4e6d9a5b3f42a74356566278.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 233, + 547, + 281 + ], + "lines": [ + { + "bbox": [ + 46, + 233, + 547, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 233, + 547, + 281 + ], + "type": "text", + "content": "Figure 3. Motion prediction module. We predict a spectral volume " + }, + { + "bbox": [ + 46, + 233, + 547, + 281 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 233, + 547, + 281 + ], + "type": "text", + "content": " through a frequency-coordinated denoising model. Each block of the diffusion network " + }, + { + "bbox": [ + 46, + 233, + 547, + 281 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 46, + 233, + 547, + 281 + ], + "type": "text", + "content": " interleaves 2D spatial layers with attention layers (red box, right), and iteratively denoises latent features " + }, + { + "bbox": [ + 46, + 233, + 547, + 281 + ], + "type": "inline_equation", + "content": "z^n" + }, + { + "bbox": [ + 46, + 233, + 547, + 281 + ], + "type": "text", + "content": ". The denoised features are fed to a decoder " + }, + { + "bbox": [ + 46, + 233, + 547, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 233, + 547, + 281 + ], + "type": "text", + "content": " to produce " + }, + { + "bbox": [ + 46, + 233, + 547, + 281 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 233, + 547, + 281 + ], + "type": "text", + "content": ". During training, we concatenate the downsampled input " + }, + { + "bbox": [ + 46, + 233, + 547, + 281 + ], + "type": "inline_equation", + "content": "I_0" + }, + { + "bbox": [ + 46, + 233, + 547, + 281 + ], + "type": "text", + "content": " with noisy latent features encoded from a real motion texture via an encoder " + }, + { + "bbox": [ + 46, + 233, + 547, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 46, + 233, + 547, + 281 + ], + "type": "text", + "content": ", and replace the noisy features with Gaussian noise " + }, + { + "bbox": [ + 46, + 233, + 547, + 281 + ], + "type": "inline_equation", + "content": "z^N" + }, + { + "bbox": [ + 46, + 233, + 547, + 281 + ], + "type": "text", + "content": " during inference (left)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 298, + 290, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 298, + 290, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 298, + 290, + 453 + ], + "type": "text", + "content": "How should we select the " + }, + { + "bbox": [ + 46, + 298, + 290, + 453 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 298, + 290, + 453 + ], + "type": "text", + "content": " output frequencies? Prior work in real-time animation has observed that most natural oscillation motions are composed primarily of low-frequency components [25, 68]. To validate this observation, we computed the average power spectrum of the motion extracted from 1,000 randomly sampled 5-second real video clips. As shown in the left plot of Fig. 2, the power spectrum of the motion decreases exponentially with increasing frequency. This suggests that most natural oscillation motions can indeed be well represented by low-frequency terms. In practice, we found that the first " + }, + { + "bbox": [ + 46, + 298, + 290, + 453 + ], + "type": "inline_equation", + "content": "K = 16" + }, + { + "bbox": [ + 46, + 298, + 290, + 453 + ], + "type": "text", + "content": " Fourier coefficients are sufficient to realistically reproduce the original natural motion in a range of real videos and scenes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 465, + 262, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 465, + 262, + 478 + ], + "spans": [ + { + "bbox": [ + 47, + 465, + 262, + 478 + ], + "type": "text", + "content": "4.2. Predicting motion with a diffusion model" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 484, + 290, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 484, + 290, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 484, + 290, + 688 + ], + "type": "text", + "content": "We choose a latent diffusion model (LDM) [73] as the backbone for our motion prediction module, as LDMs are more computationally efficient than pixel-space diffusion models, while preserving synthesis quality. A standard LDM consists of two main modules: (1) a variational autoencoder (VAE) that compresses the input image to a latent space through an encoder " + }, + { + "bbox": [ + 46, + 484, + 290, + 688 + ], + "type": "inline_equation", + "content": "z = E(I)" + }, + { + "bbox": [ + 46, + 484, + 290, + 688 + ], + "type": "text", + "content": ", then reconstructs the input from the latent features via a decoder " + }, + { + "bbox": [ + 46, + 484, + 290, + 688 + ], + "type": "inline_equation", + "content": "I = D(z)" + }, + { + "bbox": [ + 46, + 484, + 290, + 688 + ], + "type": "text", + "content": ", and (2) a U-Net based diffusion model that learns to iteratively denoise features starting from Gaussian noise. Our training applies this process not to RGB images but to spectral volumes from real video sequences, which are encoded and then diffused for " + }, + { + "bbox": [ + 46, + 484, + 290, + 688 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 484, + 290, + 688 + ], + "type": "text", + "content": " steps with a pre-defined variance schedule to produce noisy latents " + }, + { + "bbox": [ + 46, + 484, + 290, + 688 + ], + "type": "inline_equation", + "content": "z^n" + }, + { + "bbox": [ + 46, + 484, + 290, + 688 + ], + "type": "text", + "content": ". The 2D U-Nets are trained to denoise the noisy latents by iteratively estimating the noise " + }, + { + "bbox": [ + 46, + 484, + 290, + 688 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(z^{n};n,c)" + }, + { + "bbox": [ + 46, + 484, + 290, + 688 + ], + "type": "text", + "content": " used to update the latent feature at each step " + }, + { + "bbox": [ + 46, + 484, + 290, + 688 + ], + "type": "inline_equation", + "content": "n\\in (1,2,\\dots,N)" + }, + { + "bbox": [ + 46, + 484, + 290, + 688 + ], + "type": "text", + "content": ". The training loss for the LDM is written as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 696, + 287, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 696, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 54, + 696, + 287, + 712 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {L D M}} = \\mathbb {E} _ {n \\in \\mathcal {U} [ 1, N ], \\epsilon^ {n} \\in \\mathcal {N} (0, 1)} \\left[ \\left| \\left| \\epsilon^ {n} - \\epsilon_ {\\theta} \\left(z ^ {n}; n, c\\right) \\right| \\right| ^ {2} \\right] \\tag {3}", + "image_path": "04db2fd5e79faa7eb95190c6c2302bdbb19d9dd22f47c2af26c6e3daa9cee473.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 298, + 547, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 298, + 547, + 347 + ], + "spans": [ + { + "bbox": [ + 305, + 298, + 547, + 347 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 298, + 547, + 347 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 305, + 298, + 547, + 347 + ], + "type": "text", + "content": " is the embedding of any conditional signal, such as text, or, in our case, the first frame of the training video sequence, " + }, + { + "bbox": [ + 305, + 298, + 547, + 347 + ], + "type": "inline_equation", + "content": "I_0" + }, + { + "bbox": [ + 305, + 298, + 547, + 347 + ], + "type": "text", + "content": ". The clean latent features " + }, + { + "bbox": [ + 305, + 298, + 547, + 347 + ], + "type": "inline_equation", + "content": "z^0" + }, + { + "bbox": [ + 305, + 298, + 547, + 347 + ], + "type": "text", + "content": " are then passed through the decoder to recover the spectral volume." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 373, + 547, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 547, + 566 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 547, + 566 + ], + "type": "text", + "content": "Frequency adaptive normalization. One issue we observed is that motion textures have particular distribution characteristics across frequencies. As visualized in the left plot of Fig. 2, the amplitude of the spectral volumes spans a range of 0 to 100 and decays approximately exponentially with increasing frequency. As diffusion models require that the absolute values of the output are between -1 and 1 for stable training and denoising [44], we must normalize the coefficients of " + }, + { + "bbox": [ + 304, + 373, + 547, + 566 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 373, + 547, + 566 + ], + "type": "text", + "content": " extracted from real videos before using them for training. If we scale the magnitudes of these coefficients to [0,1] based on the image dimensions as in prior work [29, 76], nearly all the coefficients at higher frequencies will end up close to zero, as shown in the right plot of Fig. 2. Models trained on such data can produce inaccurate motions, since during inference, even small prediction errors can cause large relative errors after denormalization." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 570, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 548, + 715 + ], + "type": "text", + "content": "To address this issue, we employ a simple but effective frequency adaptive normalization method: First, we independently normalize Fourier coefficients at each frequency based on statistics computed from the training set. Namely, for each individual frequency " + }, + { + "bbox": [ + 304, + 570, + 548, + 715 + ], + "type": "inline_equation", + "content": "f_{j}" + }, + { + "bbox": [ + 304, + 570, + 548, + 715 + ], + "type": "text", + "content": ", we compute the " + }, + { + "bbox": [ + 304, + 570, + 548, + 715 + ], + "type": "inline_equation", + "content": "95^{\\mathrm{th}}" + }, + { + "bbox": [ + 304, + 570, + 548, + 715 + ], + "type": "text", + "content": " percentile of Fourier coefficient magnitudes over all input samples and use that value as a per-frequency scaling factor " + }, + { + "bbox": [ + 304, + 570, + 548, + 715 + ], + "type": "inline_equation", + "content": "s f_{j}" + }, + { + "bbox": [ + 304, + 570, + 548, + 715 + ], + "type": "text", + "content": ". We then apply a power transformation to each scaled Fourier coefficient to pull it away from extreme values. In practice, we observe that a square root performs better than other nonlinear transformations such as log or reciprocal. In summary, the final coefficient values of spectral volume " + }, + { + "bbox": [ + 304, + 570, + 548, + 715 + ], + "type": "inline_equation", + "content": "S(\\mathbf{p})" + }, + { + "bbox": [ + 304, + 570, + 548, + 715 + ], + "type": "text", + "content": " at" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24145" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 285, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 285, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 285, + 85 + ], + "type": "text", + "content": "frequency " + }, + { + "bbox": [ + 47, + 72, + 285, + 85 + ], + "type": "inline_equation", + "content": "f_{j}" + }, + { + "bbox": [ + 47, + 72, + 285, + 85 + ], + "type": "text", + "content": " (used for training our LDM) are computed as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 99, + 91, + 287, + 122 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 91, + 287, + 122 + ], + "spans": [ + { + "bbox": [ + 99, + 91, + 287, + 122 + ], + "type": "interline_equation", + "content": "S _ {f _ {j}} ^ {\\prime} (\\mathbf {p}) = \\operatorname {s i g n} \\left(S _ {f _ {j}}\\right) \\sqrt {\\left| \\frac {S _ {f _ {j}} (\\mathbf {p})}{s _ {f _ {j}}} \\right|}. \\tag {4}", + "image_path": "c047e88c26b0d3654ce49fd0c5731026ca106c9dad4de6766189c2466c3c9921.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 128, + 287, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 128, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 46, + 128, + 287, + 163 + ], + "type": "text", + "content": "As shown on the right plot of Fig. 2, after applying frequency adaptive normalization, the spectral volume coefficients distribute more evenly." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 167, + 288, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 167, + 288, + 286 + ], + "spans": [ + { + "bbox": [ + 46, + 167, + 288, + 286 + ], + "type": "text", + "content": "Frequency-coordinated denoising. The straightforward way to predict a spectral volume " + }, + { + "bbox": [ + 46, + 167, + 288, + 286 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 167, + 288, + 286 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 167, + 288, + 286 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 167, + 288, + 286 + ], + "type": "text", + "content": " frequency bands is to output a tensor of " + }, + { + "bbox": [ + 46, + 167, + 288, + 286 + ], + "type": "inline_equation", + "content": "4K" + }, + { + "bbox": [ + 46, + 167, + 288, + 286 + ], + "type": "text", + "content": " channels from a single diffusion U-Net. However, as in prior work [7], we observe that training a model to produce a large number of channels can yield over-smoothed, inaccurate outputs. An alternative would be to independently predict each individual frequency slice by injecting an extra frequency embedding into the LDM [4], but this design choice would result in uncorrelated predictions in the frequency domain, leading to unrealistic motion." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 287, + 289, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 287, + 289, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 287, + 289, + 574 + ], + "type": "text", + "content": "Therefore, inspired by recent video diffusion work [7], we propose a frequency-coordinated denoising strategy, illustrated in Fig. 3. In particular, given an input image " + }, + { + "bbox": [ + 46, + 287, + 289, + 574 + ], + "type": "inline_equation", + "content": "I_0" + }, + { + "bbox": [ + 46, + 287, + 289, + 574 + ], + "type": "text", + "content": ", we first train an LDM " + }, + { + "bbox": [ + 46, + 287, + 289, + 574 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 46, + 287, + 289, + 574 + ], + "type": "text", + "content": " to predict a single 4-channel frequency slice of spectral volume " + }, + { + "bbox": [ + 46, + 287, + 289, + 574 + ], + "type": "inline_equation", + "content": "S_{f_j}" + }, + { + "bbox": [ + 46, + 287, + 289, + 574 + ], + "type": "text", + "content": ", where we inject an extra frequency embedding along with the time-step embedding into the LDM. We then freeze the parameters of this LDM " + }, + { + "bbox": [ + 46, + 287, + 289, + 574 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 46, + 287, + 289, + 574 + ], + "type": "text", + "content": ", introduce attention layers interleaved with the 2D spatial layers of " + }, + { + "bbox": [ + 46, + 287, + 289, + 574 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 46, + 287, + 289, + 574 + ], + "type": "text", + "content": " across the " + }, + { + "bbox": [ + 46, + 287, + 289, + 574 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 287, + 289, + 574 + ], + "type": "text", + "content": " frequency bands, and fine-tune. Specifically, for a batch size " + }, + { + "bbox": [ + 46, + 287, + 289, + 574 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 287, + 289, + 574 + ], + "type": "text", + "content": ", the 2D spatial layers of " + }, + { + "bbox": [ + 46, + 287, + 289, + 574 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 46, + 287, + 289, + 574 + ], + "type": "text", + "content": " treat the corresponding " + }, + { + "bbox": [ + 46, + 287, + 289, + 574 + ], + "type": "inline_equation", + "content": "B \\cdot K" + }, + { + "bbox": [ + 46, + 287, + 289, + 574 + ], + "type": "text", + "content": " noisy latent features of channel size " + }, + { + "bbox": [ + 46, + 287, + 289, + 574 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 287, + 289, + 574 + ], + "type": "text", + "content": " as independent samples with shape " + }, + { + "bbox": [ + 46, + 287, + 289, + 574 + ], + "type": "inline_equation", + "content": "\\mathcal{R}^{(B \\cdot K) \\times C \\times H \\times W}" + }, + { + "bbox": [ + 46, + 287, + 289, + 574 + ], + "type": "text", + "content": ". The attention layer then interprets these as consecutive features spanning the frequency axis, and we reshape the latent features from previous 2D spatial layers to " + }, + { + "bbox": [ + 46, + 287, + 289, + 574 + ], + "type": "inline_equation", + "content": "\\mathcal{R}^{B \\times K \\times C \\times H \\times W}" + }, + { + "bbox": [ + 46, + 287, + 289, + 574 + ], + "type": "text", + "content": " before feeding them to the attention layers. In other words, the frequency attention layers are fine-tuned to coordinate all frequency slices so as to produce coherent spectral volumes. In our experiments, we see that the average VAE reconstruction error improves from 0.024 to 0.018 when we switch from a single 2D U-Net to a frequency-coordinated denoising module, suggesting an improved upper bound on LDM prediction accuracy; in Sec. 7.3, we also show that this design choice improves video generation quality." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 586, + 181, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 181, + 600 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 181, + 600 + ], + "type": "text", + "content": "5. Image-based rendering" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": "We now describe how we take a spectral volume " + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": " predicted for a given input image " + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "I_0" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": " and render a future frame " + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\hat{I}_t" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": ". We first derive a motion texture in the time domain using the inverse temporal FFT applied at each pixel " + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\mathbf{p}) = \\mathrm{FFT}^{-1}(S(\\mathbf{p}))" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": ". To produce a future frame " + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\hat{I}_t" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": ", we adopt a deep image-based rendering technique and perform splatting with the predicted motion field " + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "F_t" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": " to forward-warp the encoded " + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "I_0" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": ", as shown in Fig. 4. Since forward warping can lead to holes, and multiple source pixels can map to the same output" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 310, + 73, + 541, + 205 + ], + "blocks": [ + { + "bbox": [ + 310, + 73, + 541, + 205 + ], + "lines": [ + { + "bbox": [ + 310, + 73, + 541, + 205 + ], + "spans": [ + { + "bbox": [ + 310, + 73, + 541, + 205 + ], + "type": "image", + "image_path": "983a2734f6cefff5069132c49a005bc46464b113a35388810637e96b3a9af55f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 214, + 547, + 293 + ], + "lines": [ + { + "bbox": [ + 304, + 214, + 547, + 293 + ], + "spans": [ + { + "bbox": [ + 304, + 214, + 547, + 293 + ], + "type": "text", + "content": "Figure 4. Rendering module. We fill in missing content and refine the warped input image using a deep image-based rendering module, where multi-scale features are extracted from the input image " + }, + { + "bbox": [ + 304, + 214, + 547, + 293 + ], + "type": "inline_equation", + "content": "I_0" + }, + { + "bbox": [ + 304, + 214, + 547, + 293 + ], + "type": "text", + "content": ". Softmax splatting is then applied over the features with a motion field " + }, + { + "bbox": [ + 304, + 214, + 547, + 293 + ], + "type": "inline_equation", + "content": "F_t" + }, + { + "bbox": [ + 304, + 214, + 547, + 293 + ], + "type": "text", + "content": " from time 0 to " + }, + { + "bbox": [ + 304, + 214, + 547, + 293 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 214, + 547, + 293 + ], + "type": "text", + "content": " (subject to the weights " + }, + { + "bbox": [ + 304, + 214, + 547, + 293 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 214, + 547, + 293 + ], + "type": "text", + "content": "). The warped features are fed to an image synthesis network to produce the rendered image " + }, + { + "bbox": [ + 304, + 214, + 547, + 293 + ], + "type": "inline_equation", + "content": "\\hat{I}_t" + }, + { + "bbox": [ + 304, + 214, + 547, + 293 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 312, + 547, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 312, + 547, + 335 + ], + "spans": [ + { + "bbox": [ + 305, + 312, + 547, + 335 + ], + "type": "text", + "content": "2D location, we adopt the feature pyramid softmax splatting strategy proposed in prior work on frame interpolation [67]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 336, + 546, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 336, + 546, + 514 + ], + "spans": [ + { + "bbox": [ + 304, + 336, + 546, + 514 + ], + "type": "text", + "content": "Specifically, we encode " + }, + { + "bbox": [ + 304, + 336, + 546, + 514 + ], + "type": "inline_equation", + "content": "I_0" + }, + { + "bbox": [ + 304, + 336, + 546, + 514 + ], + "type": "text", + "content": " through a feature extractor network to produce a multi-scale feature map. For each individual feature map at scale " + }, + { + "bbox": [ + 304, + 336, + 546, + 514 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 336, + 546, + 514 + ], + "type": "text", + "content": ", we resize and scale the predicted 2D motion field " + }, + { + "bbox": [ + 304, + 336, + 546, + 514 + ], + "type": "inline_equation", + "content": "F_{t}" + }, + { + "bbox": [ + 304, + 336, + 546, + 514 + ], + "type": "text", + "content": " according to the resolution. As in Davis et al. [22], we use predicted flow magnitude as a proxy for depth to determine the contributing weight of each source pixel mapped to its destination location. In particular, we compute a per-pixel weight, " + }, + { + "bbox": [ + 304, + 336, + 546, + 514 + ], + "type": "inline_equation", + "content": "W(\\mathbf{p}) = \\frac{1}{T}\\sum_{t}||F_{t}(\\mathbf{p})||_{2}^{2}" + }, + { + "bbox": [ + 304, + 336, + 546, + 514 + ], + "type": "text", + "content": " as the average magnitude of the predicted motion texture. In other words, we assume large motions correspond to moving foreground objects, and small or zero motions correspond to background. We use motion-derived weights instead of learnable ones as [46] because we observe that in the single-view case, learnable weights are not effective for addressing disocclusion ambiguities." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 515, + 547, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 515, + 547, + 574 + ], + "spans": [ + { + "bbox": [ + 304, + 515, + 547, + 574 + ], + "type": "text", + "content": "With the motion field " + }, + { + "bbox": [ + 304, + 515, + 547, + 574 + ], + "type": "inline_equation", + "content": "F_{t}" + }, + { + "bbox": [ + 304, + 515, + 547, + 574 + ], + "type": "text", + "content": " and weights " + }, + { + "bbox": [ + 304, + 515, + 547, + 574 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 515, + 547, + 574 + ], + "type": "text", + "content": ", we apply softmax splatting to warp the feature map at each scale to produce a warped feature. The warped features are then injected into the corresponding blocks of an image synthesis decoder to produce a final rendered image " + }, + { + "bbox": [ + 304, + 515, + 547, + 574 + ], + "type": "inline_equation", + "content": "\\hat{I}_{t}" + }, + { + "bbox": [ + 304, + 515, + 547, + 574 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 575, + 547, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 575, + 547, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 575, + 547, + 635 + ], + "type": "text", + "content": "We jointly train the feature extractor and synthesis networks with start and target frames " + }, + { + "bbox": [ + 304, + 575, + 547, + 635 + ], + "type": "inline_equation", + "content": "(I_0, I_t)" + }, + { + "bbox": [ + 304, + 575, + 547, + 635 + ], + "type": "text", + "content": " randomly sampled from real videos, using the estimated flow field from " + }, + { + "bbox": [ + 304, + 575, + 547, + 635 + ], + "type": "inline_equation", + "content": "I_0" + }, + { + "bbox": [ + 304, + 575, + 547, + 635 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 575, + 547, + 635 + ], + "type": "inline_equation", + "content": "I_t" + }, + { + "bbox": [ + 304, + 575, + 547, + 635 + ], + "type": "text", + "content": " to warp encoded features from " + }, + { + "bbox": [ + 304, + 575, + 547, + 635 + ], + "type": "inline_equation", + "content": "I_0" + }, + { + "bbox": [ + 304, + 575, + 547, + 635 + ], + "type": "text", + "content": ", and supervising predictions " + }, + { + "bbox": [ + 304, + 575, + 547, + 635 + ], + "type": "inline_equation", + "content": "\\hat{I}_t" + }, + { + "bbox": [ + 304, + 575, + 547, + 635 + ], + "type": "text", + "content": " against " + }, + { + "bbox": [ + 304, + 575, + 547, + 635 + ], + "type": "inline_equation", + "content": "I_t" + }, + { + "bbox": [ + 304, + 575, + 547, + 635 + ], + "type": "text", + "content": " with a VGG perceptual loss [49]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 646, + 386, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 646, + 386, + 659 + ], + "spans": [ + { + "bbox": [ + 306, + 646, + 386, + 659 + ], + "type": "text", + "content": "6. Applications" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "Image-to-video. Our system enables the animation of a single still picture by first predicting a motion spectral volume from the input image and generating an animation by applying our image-based rendering module to the motion" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24146" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": "texture transformed from the spectral volume. Since we explicitly model scene motion, this allows us to produce slow-motion videos by linearly interpolating the motion texture, or to magnify (or minify) animated motions by adjusting the amplitude of the predicted spectral volume coefficients." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 288, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 288, + 312 + ], + "type": "text", + "content": "Seamless looping. Many applications require videos that loop seamlessly, where there is no discontinuity between the start and end of the video. Unfortunately, it is hard to find a large collection of seamlessly looping videos for training. Instead, we devise a method to use our motion diffusion model, trained on regular non-looping video clips, to produce seamless looping video. Inspired by recent work on guidance for image editing [3, 30], our method is a motion self-guidance technique that guides the motion denoising sampling processing using explicit looping constraints. In particular, at each iterative denoising step during inference, we incorporate an additional motion guidance signal alongside standard classifier-free guidance [45], where we enforce each pixel's position and velocity at the start and end frames to be as similar as possible:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 319, + 286, + 334 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 286, + 334 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 286, + 334 + ], + "type": "interline_equation", + "content": "\\tilde {\\epsilon} ^ {n} = (1 + w) \\epsilon_ {\\theta} \\left(z ^ {n}; n, c\\right) - w \\epsilon_ {\\theta} \\left(z ^ {n}; n, \\emptyset\\right) + u \\sigma^ {n} \\nabla_ {z ^ {n}} \\mathcal {L} _ {g} ^ {n}", + "image_path": "d2fc04b6df32c783245335540e37ea480e0512e842abd4b2d8cda9f45affae3f.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 100, + 335, + 287, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 335, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 100, + 335, + 287, + 350 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {g} ^ {n} = \\left\\| F _ {T} ^ {n} - F _ {1} ^ {n} \\right\\| _ {1} + \\left\\| \\nabla F _ {T} ^ {n} - \\nabla F _ {1} ^ {n} \\right\\| _ {1} \\tag {5}", + "image_path": "ecef6e8d74c07912d1e72f485921516b9137f9d25e73a9414f557197f1400355.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 353, + 287, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 353, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 353, + 287, + 449 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 353, + 287, + 449 + ], + "type": "inline_equation", + "content": "F_{t}^{n}" + }, + { + "bbox": [ + 46, + 353, + 287, + 449 + ], + "type": "text", + "content": " is the predicted 2D displacement field at time " + }, + { + "bbox": [ + 46, + 353, + 287, + 449 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 353, + 287, + 449 + ], + "type": "text", + "content": " and denoising step " + }, + { + "bbox": [ + 46, + 353, + 287, + 449 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 353, + 287, + 449 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 353, + 287, + 449 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 353, + 287, + 449 + ], + "type": "text", + "content": " is the classifier-free guidance weight, and " + }, + { + "bbox": [ + 46, + 353, + 287, + 449 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 46, + 353, + 287, + 449 + ], + "type": "text", + "content": " is the motion self-guidance weight. In the supplemental video, we apply a baseline appearance-based looping algorithm [58] to generate a looping video from our non-looping output, and show that our motion self-guidance technique produces seamless looping videos with less distortion and fewer artifacts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 451, + 288, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 451, + 288, + 595 + ], + "spans": [ + { + "bbox": [ + 46, + 451, + 288, + 595 + ], + "type": "text", + "content": "Interactive dynamics from a single image. Davis et al. [22] show that the spectral volume, evaluated at certain resonant frequencies, can approximate an image-space modal basis that is a projection of the vibration modes of the underlying scene (or, more generally, captures spatial and temporal correlations in oscillatory dynamics), and can be used to simulate the object's response to a user-defined force. We adopt this modal analysis method [22, 69], allowing us to write the image-space 2D motion displacement field for the object's physical response as a weighted sum of motion spectrum coefficients " + }, + { + "bbox": [ + 46, + 451, + 288, + 595 + ], + "type": "inline_equation", + "content": "S_{f_j}" + }, + { + "bbox": [ + 46, + 451, + 288, + 595 + ], + "type": "text", + "content": " modulated by the state of complex modal coordinates " + }, + { + "bbox": [ + 46, + 451, + 288, + 595 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_{f_j}(t)" + }, + { + "bbox": [ + 46, + 451, + 288, + 595 + ], + "type": "text", + "content": " at each simulated time step " + }, + { + "bbox": [ + 46, + 451, + 288, + 595 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 451, + 288, + 595 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 111, + 599, + 287, + 625 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 599, + 287, + 625 + ], + "spans": [ + { + "bbox": [ + 111, + 599, + 287, + 625 + ], + "type": "interline_equation", + "content": "F _ {t} (\\mathbf {p}) = \\sum_ {f _ {j}} S _ {f _ {j}} (\\mathbf {p}) \\mathbf {q} _ {f _ {j}} (t) \\tag {6}", + "image_path": "033220da353f3cd7fa6677aa0fcb9457a1e5433a073468762be73d4935483201.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "type": "text", + "content": "We simulate the state of the modal coordinates " + }, + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_{f_j}(t)" + }, + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "type": "text", + "content": " via an explicit Euler method applied to the equations of motion for a decoupled mass-spring-damper system represented in modal space [22, 23, 69]. We refer readers to supplementary material and original work for a full derivation. Note that our method produces an interactive scene from a single picture, whereas these prior methods required a video as input." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 307, + 72, + 550, + 178 + ], + "blocks": [ + { + "bbox": [ + 307, + 72, + 550, + 178 + ], + "lines": [ + { + "bbox": [ + 307, + 72, + 550, + 178 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 550, + 178 + ], + "type": "table", + "html": "
MethodImage SynthesisVideo Synthesis
FIDKIDFVDFVD32DTFVDDTFVD32
TATS [35]65.81.67265.6419.622.640.7
Stochastic I2V [27]68.33.12253.5320.916.741.7
MCVD [92]63.42.97208.6270.419.553.9
LFDM [66]47.61.70187.5254.313.045.6
DMVFN [48]37.91.09206.5316.311.254.5
Endo et al. [29]10.40.19166.0231.65.3565.1
Holynski et al. [46]11.20.20179.0253.77.2346.8
Ours4.030.0847.162.92.536.75
", + "image_path": "00340d8fafd80fcd22e9f4f54eafc484a83f64cde4b17786fa02e4449adef3ca.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 186, + 547, + 232 + ], + "lines": [ + { + "bbox": [ + 304, + 186, + 547, + 232 + ], + "spans": [ + { + "bbox": [ + 304, + 186, + 547, + 232 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons on the test set. We report both image synthesis and video synthesis quality. Here, KID is scaled by 100. Lower is better for all error. See Sec. 7.1 for descriptions of baselines and error metrics." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 261, + 387, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 261, + 387, + 274 + ], + "spans": [ + { + "bbox": [ + 306, + 261, + 387, + 274 + ], + "type": "text", + "content": "7. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 281, + 547, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 547, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 547, + 483 + ], + "type": "text", + "content": "Implementation details. We use an LDM [73] as the backbone for predicting spectral volumes, for which we use a VAE with a continuous latent space of dimension 4. We train the VAE with an " + }, + { + "bbox": [ + 304, + 281, + 547, + 483 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 304, + 281, + 547, + 483 + ], + "type": "text", + "content": " reconstruction loss, a multi-scale gradient consistency loss [54-56], and a KL-divergence loss with respective weights of " + }, + { + "bbox": [ + 304, + 281, + 547, + 483 + ], + "type": "inline_equation", + "content": "1,0.2,10^{-6}" + }, + { + "bbox": [ + 304, + 281, + 547, + 483 + ], + "type": "text", + "content": ". We train the same 2D U-Net used in the original LDM work to perform iterative denoising with a simple MSE loss [44], and adopt the attention layers from [41] for frequency-coordinated denoising. For quantitative evaluation, we train both VAE and LDM on images of size " + }, + { + "bbox": [ + 304, + 281, + 547, + 483 + ], + "type": "inline_equation", + "content": "256 \\times 160" + }, + { + "bbox": [ + 304, + 281, + 547, + 483 + ], + "type": "text", + "content": " from scratch for fair comparisons, and it takes around 6 days to converge using 16 Nvidia A100 GPUs. For our main quantitative and qualitative results, we run the motion diffusion model with DDIM [85] for 250 steps. We also show generated videos of up to a resolution of " + }, + { + "bbox": [ + 304, + 281, + 547, + 483 + ], + "type": "inline_equation", + "content": "512 \\times 288" + }, + { + "bbox": [ + 304, + 281, + 547, + 483 + ], + "type": "text", + "content": ", created by fine-tuning a pre-trained image inpainting LDM model [73] on our dataset." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 484, + 547, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 547, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 547, + 579 + ], + "type": "text", + "content": "We adopt ResNet-34 [39] for the feature extractor in our IBR module. Our image synthesis network is based on an architecture for conditional image inpainting [57, 109]. Our rendering module runs in real-time at 25FPS on a Nvidia V100 GPU during inference. We adopt universal guidance [3] to produce seamless looping videos, where we set weights " + }, + { + "bbox": [ + 304, + 484, + 547, + 579 + ], + "type": "inline_equation", + "content": "w = 1.75" + }, + { + "bbox": [ + 304, + 484, + 547, + 579 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 484, + 547, + 579 + ], + "type": "inline_equation", + "content": "u = 200" + }, + { + "bbox": [ + 304, + 484, + 547, + 579 + ], + "type": "text", + "content": ", and use 500 DDIM steps with 2 self-recurrence iterations." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "content": "Data. We collect and process a set of 3,015 videos of natural scenes exhibiting oscillatory motions from online sources our own captures. We withhold " + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "content": " of videos for testing and use the remainder for training. To extract ground truth motion trajectories, we apply a coarse-to-fine flow method [10, 61] between each selected starting image and every future frame of the video. As training data, we take every 10th video frame as input images and derive the corresponding ground truth spectral volumes using the computed motion trajectories across the following 149 frames. In total, our data consists of over 150K image-motion pairs." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24147" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 71, + 121, + 114 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 121, + 114 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 121, + 114 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 121, + 114 + ], + "type": "image", + "image_path": "efc2a8f8a62808bfbc24690cd316a86b0c962bc313f6cbd97826670372d514cf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 121, + 72, + 541, + 114 + ], + "blocks": [ + { + "bbox": [ + 121, + 72, + 541, + 114 + ], + "lines": [ + { + "bbox": [ + 121, + 72, + 541, + 114 + ], + "spans": [ + { + "bbox": [ + 121, + 72, + 541, + 114 + ], + "type": "image", + "image_path": "687b4757f4f7ff8f19ece485a333350c452fc6d5c64eb4fbcc8c046edab9c00c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 51, + 116, + 121, + 161 + ], + "blocks": [ + { + "bbox": [ + 51, + 116, + 121, + 161 + ], + "lines": [ + { + "bbox": [ + 51, + 116, + 121, + 161 + ], + "spans": [ + { + "bbox": [ + 51, + 116, + 121, + 161 + ], + "type": "image", + "image_path": "7839a926526e3e4dd3a6af660f6a59596f3b2f4d27111727c750b3aacf0b78ef.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 121, + 116, + 541, + 161 + ], + "blocks": [ + { + "bbox": [ + 121, + 116, + 541, + 161 + ], + "lines": [ + { + "bbox": [ + 121, + 116, + 541, + 161 + ], + "spans": [ + { + "bbox": [ + 121, + 116, + 541, + 161 + ], + "type": "image", + "image_path": "5b28f4e228885a1d8e872a723429e8503719246c835cfa996b23a72fda71a388.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 51, + 163, + 121, + 208 + ], + "blocks": [ + { + "bbox": [ + 51, + 163, + 121, + 208 + ], + "lines": [ + { + "bbox": [ + 51, + 163, + 121, + 208 + ], + "spans": [ + { + "bbox": [ + 51, + 163, + 121, + 208 + ], + "type": "image", + "image_path": "8145382f33d8ec80598737642c71212ac9f35bf8987751b375c4cfd7fb429695.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 121, + 163, + 541, + 207 + ], + "blocks": [ + { + "bbox": [ + 121, + 163, + 541, + 207 + ], + "lines": [ + { + "bbox": [ + 121, + 163, + 541, + 207 + ], + "spans": [ + { + "bbox": [ + 121, + 163, + 541, + 207 + ], + "type": "image", + "image_path": "9ec4504c6cb8a6f632d254066af28657078a4fead1f3d0b52dcd4bf665bee686.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 51, + 209, + 121, + 255 + ], + "blocks": [ + { + "bbox": [ + 51, + 209, + 121, + 255 + ], + "lines": [ + { + "bbox": [ + 51, + 209, + 121, + 255 + ], + "spans": [ + { + "bbox": [ + 51, + 209, + 121, + 255 + ], + "type": "image", + "image_path": "7dc991b385b536cee35167fb8d7974d35270558ad365c6855f1f8fefb7d8c4c0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 258, + 108, + 267 + ], + "lines": [ + { + "bbox": [ + 67, + 258, + 108, + 267 + ], + "spans": [ + { + "bbox": [ + 67, + 258, + 108, + 267 + ], + "type": "text", + "content": "Input image" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 121, + 209, + 541, + 254 + ], + "blocks": [ + { + "bbox": [ + 121, + 209, + 541, + 254 + ], + "lines": [ + { + "bbox": [ + 121, + 209, + 541, + 254 + ], + "spans": [ + { + "bbox": [ + 121, + 209, + 541, + 254 + ], + "type": "image", + "image_path": "d99e174881d2b858ec871d89b560ce036d391c4194bf134bdbd875d6a53ba017.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 140, + 258, + 175, + 266 + ], + "lines": [ + { + "bbox": [ + 140, + 258, + 175, + 266 + ], + "spans": [ + { + "bbox": [ + 140, + 258, + 175, + 266 + ], + "type": "text", + "content": "Reference" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 195, + 258, + 260, + 267 + ], + "lines": [ + { + "bbox": [ + 195, + 258, + 260, + 267 + ], + "spans": [ + { + "bbox": [ + 195, + 258, + 260, + 267 + ], + "type": "text", + "content": "Stochastic-I2V [27]" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 276, + 258, + 317, + 267 + ], + "lines": [ + { + "bbox": [ + 276, + 258, + 317, + 267 + ], + "spans": [ + { + "bbox": [ + 276, + 258, + 317, + 267 + ], + "type": "text", + "content": "MCVD[92]" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 340, + 258, + 392, + 267 + ], + "lines": [ + { + "bbox": [ + 340, + 258, + 392, + 267 + ], + "spans": [ + { + "bbox": [ + 340, + 258, + 392, + 267 + ], + "type": "text", + "content": "Endo et al. [29]" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 404, + 258, + 469, + 267 + ], + "lines": [ + { + "bbox": [ + 404, + 258, + 469, + 267 + ], + "spans": [ + { + "bbox": [ + 404, + 258, + 469, + 267 + ], + "type": "text", + "content": "Holynski et al. [46]." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 497, + 258, + 514, + 266 + ], + "lines": [ + { + "bbox": [ + 497, + 258, + 514, + 266 + ], + "spans": [ + { + "bbox": [ + 497, + 258, + 514, + 266 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 267, + 545, + 286 + ], + "lines": [ + { + "bbox": [ + 47, + 267, + 545, + 286 + ], + "spans": [ + { + "bbox": [ + 47, + 267, + 545, + 286 + ], + "type": "text", + "content": "Figure 5. X - t slices of videos generated by different approaches. From left to right: input image and corresponding X - t video slices from the ground truth video, from videos generated by three baselines [27, 29, 46, 92], and finally videos generated by our approach." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 48, + 304, + 168, + 384 + ], + "blocks": [ + { + "bbox": [ + 48, + 304, + 168, + 384 + ], + "lines": [ + { + "bbox": [ + 48, + 304, + 168, + 384 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 168, + 384 + ], + "type": "image", + "image_path": "878d42b430e0dff41b50c1a4bd46edb1567a538fbaae11017e6b474aee4599d7.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 391, + 288, + 424 + ], + "lines": [ + { + "bbox": [ + 46, + 391, + 288, + 424 + ], + "spans": [ + { + "bbox": [ + 46, + 391, + 288, + 424 + ], + "type": "text", + "content": "Figure 6. Sliding window FID and DTFVD. We show sliding window FID with window size 30 frames, and DTFVD with size 16 frames, for videos generated by different methods." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 168, + 304, + 286, + 384 + ], + "blocks": [ + { + "bbox": [ + 168, + 304, + 286, + 384 + ], + "lines": [ + { + "bbox": [ + 168, + 304, + 286, + 384 + ], + "spans": [ + { + "bbox": [ + 168, + 304, + 286, + 384 + ], + "type": "image", + "image_path": "74230c64b94baadab59b4863b66b451dff1512f6d155c3d062f9c5de1350609c.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 444, + 287, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 444, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 46, + 444, + 287, + 588 + ], + "type": "text", + "content": "Baselines. We compare our approach to recent single-image animation and video prediction methods. Endo et al. [29] and DMVFN [48] predict instantaneous 2D motion fields and render future frames auto-regressively. Holynski et al. [46] instead simulates motion through a single static Eulerian motion description. Other recent work such as Stochastic Image-to-Video (Stochastic-I2V) [27], TATS [35], and MCVD [92] adopt VAEs, transformers, or diffusion models to directly predict raw video frames; LFDM [66] generates future frames by predicting flow volumes and warping latents in a diffusion model. We train all the above methods on our data using their respective open-source implementations." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 588, + 288, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 588, + 288, + 671 + ], + "spans": [ + { + "bbox": [ + 46, + 588, + 288, + 671 + ], + "type": "text", + "content": "We evaluate the quality of the videos generated by our approach and by prior baselines in two ways. First, we evaluate the quality of individual synthesized frames using metrics designed for image synthesis tasks. We adopt the Fréchet Inception Distance (FID) [42] and Kernel Inception Distance (KID) [5] to measure the average distance between the distributions of generated frames and ground truth frames." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 672, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 672, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 672, + 288, + 696 + ], + "type": "text", + "content": "Second, to evaluate the quality and temporal coherence of synthesized videos, we adopt the Fréchet Video Dis" + } + ] + } + ], + "index": 21 + }, + { + "type": "table", + "bbox": [ + 306, + 304, + 549, + 442 + ], + "blocks": [ + { + "bbox": [ + 306, + 304, + 549, + 442 + ], + "lines": [ + { + "bbox": [ + 306, + 304, + 549, + 442 + ], + "spans": [ + { + "bbox": [ + 306, + 304, + 549, + 442 + ], + "type": "table", + "html": "
MethodImage SynthesisVideo Synthesis
FIDKIDFVDFVD32DTFVDDTFVD32
Repeat I0--237.5316.75.3045.6
K=43.920.0760.378.43.128.59
K=83.950.0752.168.72.717.37
K=244.090.0848.265.12.506.94
w/o adaptive norm.4.530.0962.780.13.168.19
Independent pred.4.000.0852.571.32.707.40
Volume pred.4.740.0953.771.12.837.79
Baseline splat [46]4.250.0949.566.82.837.27
Full (K=16)4.030.0847.162.92.536.75
", + "image_path": "736a084db9418f093c03885dc2c16f7b5577ef5e5104d6acefaa0b40fd101295.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "table_body" + } + ], + "index": 22 + }, + { + "bbox": [ + 310, + 445, + 539, + 457 + ], + "lines": [ + { + "bbox": [ + 310, + 445, + 539, + 457 + ], + "spans": [ + { + "bbox": [ + 310, + 445, + 539, + 457 + ], + "type": "text", + "content": "Table 2. Ablation study. Sec. 7.3 describes each configuration." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 490, + 547, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 490, + 547, + 598 + ], + "spans": [ + { + "bbox": [ + 304, + 490, + 547, + 598 + ], + "type": "text", + "content": "tance [91] with window size 16 (FVD) and 32 " + }, + { + "bbox": [ + 304, + 490, + 547, + 598 + ], + "type": "inline_equation", + "content": "(\\mathrm{FVD}_{32})" + }, + { + "bbox": [ + 304, + 490, + 547, + 598 + ], + "type": "text", + "content": " based on an I3D model [11] trained on the Human Kinetics datasets [52]. To more faithfully reflect synthesis quality for the natural oscillation motions we seek to generate, we also adopt the Dynamic Texture Frechet Video Distance [27], which measures distance from videos with window size 16 (DTFVD) and size 32 " + }, + { + "bbox": [ + 304, + 490, + 547, + 598 + ], + "type": "inline_equation", + "content": "(\\mathrm{DTFVD}_{32})" + }, + { + "bbox": [ + 304, + 490, + 547, + 598 + ], + "type": "text", + "content": ", using a I3D model trained on the Dynamic Textures Database [37], a dataset consisting primarily of natural motion textures." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 599, + 547, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 599, + 547, + 660 + ], + "spans": [ + { + "bbox": [ + 304, + 599, + 547, + 660 + ], + "type": "text", + "content": "We further use a sliding window FID of window size 30 frames, and a sliding window DTFVD with window size 16 frames, as in [57, 60], to measure how generated video quality degrades over time. For all methods, we evaluate metrics at " + }, + { + "bbox": [ + 304, + 599, + 547, + 660 + ], + "type": "inline_equation", + "content": "256 \\times 128" + }, + { + "bbox": [ + 304, + 599, + 547, + 660 + ], + "type": "text", + "content": " resolution by center-cropping." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 670, + 421, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 421, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 421, + 682 + ], + "type": "text", + "content": "7.1.Quantitative results" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 547, + 714 + ], + "type": "text", + "content": "Table 1 shows quantitative comparisons between our approach and baselines on our test set. Our approach signifi-" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 282, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 282, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 282, + 713 + ], + "type": "text", + "content": "1We use the open-source implementation of [46] from Fan et al. [82]." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24148" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 288, + 118 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 288, + 118 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 288, + 118 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 288, + 118 + ], + "type": "image", + "image_path": "1e82f1326227507e720cca0dca4200e264301081fbda44457fe31c861c4cb02a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 122, + 287, + 144 + ], + "lines": [ + { + "bbox": [ + 46, + 122, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 122, + 287, + 144 + ], + "type": "text", + "content": "Figure 7. We show generated future frames from three recent large video diffusion models [31, 36, 97]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 166, + 288, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 166, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 46, + 166, + 288, + 274 + ], + "type": "text", + "content": "cantly outperforms prior single-image animation baselines in terms of both image and video synthesis quality. Specifically, our much lower FVD and DT-FVD distances suggest that the videos generated by our approach are more realistic and more temporally coherent. Further, Fig. 6 shows the sliding window FID and sliding window DT-FVD distances of generated videos from different methods. Thanks to the global spectral volume representation, videos generated by our approach do not suffer from degradation over time." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 283, + 155, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 283, + 155, + 295 + ], + "spans": [ + { + "bbox": [ + 47, + 283, + 155, + 295 + ], + "type": "text", + "content": "7.2. Qualitative results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 302, + 288, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 302, + 288, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 302, + 288, + 456 + ], + "type": "text", + "content": "We visualize qualitative comparisons between videos as spatio-temporal " + }, + { + "bbox": [ + 46, + 302, + 288, + 456 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 302, + 288, + 456 + ], + "type": "text", + "content": "-" + }, + { + "bbox": [ + 46, + 302, + 288, + 456 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 302, + 288, + 456 + ], + "type": "text", + "content": " slices of the generated videos, a standard way of visualizing small motions in a video [94]. As shown in Fig. 5, our generated video dynamics more strongly resemble the motion patterns observed in the corresponding real reference videos (second column), compared to other methods. Baselines such as Stochastic I2V [27] and MCVD [92] fail to model both appearance and motion realistically over time. Endo et al. [29] and Holynski et al. [46] produce video frames with fewer artifacts but exhibits oversmooth or non-oscillatory motion over time. We refer readers to supplementary material to assess the quality of generated video frames and estimated motion across different methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 467, + 138, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 467, + 138, + 479 + ], + "spans": [ + { + "bbox": [ + 47, + 467, + 138, + 479 + ], + "type": "text", + "content": "7.3. Ablation study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 486, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 288, + 713 + ], + "type": "text", + "content": "We conduct an ablation study to validate the major design choices in our motion prediction and rendering modules, comparing our full configuration with different variants. Specifically, we evaluate results using different numbers of frequency bands " + }, + { + "bbox": [ + 46, + 486, + 288, + 713 + ], + "type": "inline_equation", + "content": "K = 4, 8, 16, 24" + }, + { + "bbox": [ + 46, + 486, + 288, + 713 + ], + "type": "text", + "content": ". We observe that increasing the number of frequency bands improves video prediction quality, but the improvement is marginal with more than 16 frequencies. Next, we remove adaptive frequency normalization from the ground truth spectral volumes, and instead just scale them based on input image width and height (w/o adaptive norm.). Additionally, we remove the frequency coordinated-denoising module (Independent pred.), or replace it with a simpler DM where a tensor volume of " + }, + { + "bbox": [ + 46, + 486, + 288, + 713 + ], + "type": "inline_equation", + "content": "4K" + }, + { + "bbox": [ + 46, + 486, + 288, + 713 + ], + "type": "text", + "content": " channel spectral volumes are predicted jointly via a single 2D U-net diffusion model (Volume pred.). Finally, we compare results where we use a baseline rendering method that applies softmax splatting over single-scale features subject to learnable weights as used by [46] (Baseline splat). We also add a baseline where the generated video is a volume by" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 71, + 547, + 105 + ], + "blocks": [ + { + "bbox": [ + 307, + 71, + 547, + 105 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 547, + 105 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 547, + 105 + ], + "type": "image", + "image_path": "fade8472536fcf69e69f99edd72d2d97d82ea8a2b72a746f509115affa77424b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 110, + 547, + 154 + ], + "lines": [ + { + "bbox": [ + 305, + 110, + 547, + 154 + ], + "spans": [ + { + "bbox": [ + 305, + 110, + 547, + 154 + ], + "type": "text", + "content": "Figure 8. Limitations. We show examples of rendered future frames (even), and overlay of input and rendered images (odd). Our method can produce artifacts in regions of thin objects or large motions, and regions requiring filling large amount of new contents." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 175, + 546, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 175, + 546, + 210 + ], + "spans": [ + { + "bbox": [ + 305, + 175, + 546, + 210 + ], + "type": "text", + "content": "repeating input image " + }, + { + "bbox": [ + 305, + 175, + 546, + 210 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 175, + 546, + 210 + ], + "type": "text", + "content": " times (Repeat " + }, + { + "bbox": [ + 305, + 175, + 546, + 210 + ], + "type": "inline_equation", + "content": "I_0" + }, + { + "bbox": [ + 305, + 175, + 546, + 210 + ], + "type": "text", + "content": "). From Table 2, we observe that all simpler or alternative configurations lead to worse performance compared with our full model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 217, + 484, + 230 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 217, + 484, + 230 + ], + "spans": [ + { + "bbox": [ + 306, + 217, + 484, + 230 + ], + "type": "text", + "content": "7.4. Comparing to large video models" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 236, + 547, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 236, + 547, + 368 + ], + "spans": [ + { + "bbox": [ + 305, + 236, + 547, + 368 + ], + "type": "text", + "content": "We further perform a user study and compare our generated animations with ones from recent large video diffusion models:AnimateDiff [36], ModelScope [97] and Gen-2 [31], which predict video volumes directly. On a randomly selected 30 videos from the test set, we ask users \"which video is more realistic?\". Users report a " + }, + { + "bbox": [ + 305, + 236, + 547, + 368 + ], + "type": "inline_equation", + "content": "80.9\\%" + }, + { + "bbox": [ + 305, + 236, + 547, + 368 + ], + "type": "text", + "content": " preference for our approach over others. Moreover, as shown in Fig. 7, we observe that the generated videos from these baselines are either unable to adhere to the input image content, or exhibit gradual color drift and distortion over time. We refer readers to the supplementary material for full comparisons." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 379, + 454, + 391 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 379, + 454, + 391 + ], + "spans": [ + { + "bbox": [ + 306, + 379, + 454, + 391 + ], + "type": "text", + "content": "8. Discussion and conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 399, + 547, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 399, + 547, + 506 + ], + "spans": [ + { + "bbox": [ + 305, + 399, + 547, + 506 + ], + "type": "text", + "content": "Limitations. Since our approach only predicts lower frequencies of a spectral volume, it can fail to model non-oscillating motions or high-frequency vibrations—this may be resolved by using learned motion bases. Furthermore, the quality of generated videos relies on the quality of underlying motion trajectories, which may degrade in scenes with thin moving objects or objects with large displacements. Even if correct, motions that require generating large amounts of new unseen content may also cause degradations (Fig. 8)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 508, + 547, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 508, + 547, + 675 + ], + "spans": [ + { + "bbox": [ + 305, + 508, + 547, + 675 + ], + "type": "text", + "content": "Conclusion. We present a new approach for modeling natural oscillation dynamics from a single still picture. Our image-space motion prior is represented with spectral volumes, a frequency representation of per-pixel motion trajectories, which we find to be efficient and effective for prediction with diffusion models, and which we learn from collections of real world videos. Spectral volumes are predicted using frequency-coordinated latent diffusion model and are used to animate future video frames through an image-based rendering module. We show that our approach produces photo-realistic animations from a single picture and significantly outperforms prior baselines, and that it can enable several downstream applications such as creating seamlessly looping or interactive image dynamics." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 677, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 677, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 677, + 546, + 713 + ], + "type": "text", + "content": "Acknowledgements. We thank Abe Davis, Rick Szeliski, Andrew Liu, Boyang Deng, Qianqian Wang, Xuan Luo, and Lucy Chai for fruitful discussions and helpful comments." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24149" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 91, + 288, + 703 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 58, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 58, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Aseem Agarwala, Ke Colin Zheng, Chris Pal, Maneesh Agrawala, Michael Cohen, Brian Curless, David Salesin, and Richard Szeliski. Panoramic video textures. In ACM SIGGRAPH 2005 Papers, pages 821-827. 2005." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 134, + 288, + 166 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 134, + 288, + 166 + ], + "spans": [ + { + "bbox": [ + 58, + 134, + 288, + 166 + ], + "type": "text", + "content": "[2] Hyemin Ahn, Esteve Valls Mascaro, and Dongheui Lee. Can we use diffusion probabilistic models for 3d motion prediction? arXiv preprint arXiv:2302.14503, 2023." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 166, + 288, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 166, + 288, + 220 + ], + "spans": [ + { + "bbox": [ + 58, + 166, + 288, + 220 + ], + "type": "text", + "content": "[3] Arpit Bansal, Hong-Min Chu, Avi Schwarzschild, Soumyadip Sengupta, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Universal guidance for diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 843-852, 2023." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 220, + 288, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 220, + 288, + 284 + ], + "spans": [ + { + "bbox": [ + 58, + 220, + 288, + 284 + ], + "type": "text", + "content": "[4] Hugo Bertiche, Niloy J Mitra, Kuldeep Kulkarni, ChunHao P Huang, Tuanfeng Y Wang, Meysam Madadi, Sergio Escalera, and Duygu Ceylan. Blowing in the wind: Cyclenet for human cinemagrams from still images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 459-468, 2023." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 285, + 288, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 285, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 58, + 285, + 288, + 316 + ], + "type": "text", + "content": "[5] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying MMD GANs. arXiv preprint arXiv:1801.01401, 2018." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 316, + 288, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 316, + 288, + 369 + ], + "spans": [ + { + "bbox": [ + 58, + 316, + 288, + 369 + ], + "type": "text", + "content": "[6] Andreas Blattmann, Timo Milbich, Michael Dorkenwald, and Björn Ommer. ipoke: Poking a still image for controlled stochastic video synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14707-14717, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 370, + 288, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 370, + 288, + 435 + ], + "spans": [ + { + "bbox": [ + 58, + 370, + 288, + 435 + ], + "type": "text", + "content": "[7] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 58, + 435, + 288, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 435, + 288, + 479 + ], + "spans": [ + { + "bbox": [ + 58, + 435, + 288, + 479 + ], + "type": "text", + "content": "[8] Richard Strong Bowen, Richard Tucker, Ramin Zabih, and Noah Snavely. Dimensions of motion: Monocular prediction through flow subspaces. In 2022 International Conference on 3D Vision (3DV), pages 454-464. IEEE, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 58, + 479, + 288, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 479, + 288, + 531 + ], + "spans": [ + { + "bbox": [ + 58, + 479, + 288, + 531 + ], + "type": "text", + "content": "[9] Tim Brooks, Janne Hellsten, Miika Aittala, Ting-Chun Wang, Timo Aila, Jaakko Lehtinen, Ming-Yu Liu, Alexei Efros, and Tero Karras. Generating long videos of dynamic scenes. Advances in Neural Information Processing Systems, 35:31769-31781, 2022." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 533, + 288, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 533, + 288, + 575 + ], + "spans": [ + { + "bbox": [ + 53, + 533, + 288, + 575 + ], + "type": "text", + "content": "[10] Thomas Brox, Andrés Bruhn, Nils Papenberg, and Joachim Weickert. High accuracy optical flow estimation based on a theory for warping. In Proc. European Conf. on Computer Vision (ECCV), pages 25-36. Springer, 2004." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 575, + 288, + 618 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 575, + 288, + 618 + ], + "spans": [ + { + "bbox": [ + 53, + 575, + 288, + 618 + ], + "type": "text", + "content": "[11] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 618, + 288, + 661 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 618, + 288, + 661 + ], + "spans": [ + { + "bbox": [ + 53, + 618, + 288, + 661 + ], + "type": "text", + "content": "[12] Dan Casas, Marco Volino, John Collomosse, and Adrian Hilton. 4d video textures for interactive character appearance. In Computer Graphics Forum, volume 33, pages 371-380. Wiley Online Library, 2014." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 661, + 288, + 703 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 661, + 288, + 703 + ], + "spans": [ + { + "bbox": [ + 53, + 661, + 288, + 703 + ], + "type": "text", + "content": "[13] Antoni B Chan and Nuno Vasconcelos. Mixtures of dynamic textures. In Tenth IEEE International Conference on Computer Vision (ICCV'05) Volume 1, volume 1, pages 641-647. IEEE, 2005." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 547, + 707 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 312, + 73, + 545, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 545, + 115 + ], + "type": "text", + "content": "[14] Antoni B Chan and Nuno Vasconcelos. Classifying video with kernel dynamic textures. In 2007 IEEE Conference on Computer Vision and Pattern Recognition, pages 1-6. IEEE, 2007." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 117, + 547, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 117, + 547, + 159 + ], + "spans": [ + { + "bbox": [ + 312, + 117, + 547, + 159 + ], + "type": "text", + "content": "[15] Antoni B Chan and Nuno Vasconcelos. Modeling, clustering, and segmenting video with mixtures of dynamic textures. IEEE transactions on pattern analysis and machine intelligence, 30(5):909-926, 2008." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 159, + 547, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 159, + 547, + 213 + ], + "spans": [ + { + "bbox": [ + 312, + 159, + 547, + 213 + ], + "type": "text", + "content": "[16] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 213, + 547, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 213, + 547, + 256 + ], + "spans": [ + { + "bbox": [ + 312, + 213, + 547, + 256 + ], + "type": "text", + "content": "[17] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11315-11325, 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 256, + 547, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 256, + 547, + 298 + ], + "spans": [ + { + "bbox": [ + 312, + 256, + 547, + 298 + ], + "type": "text", + "content": "[18] Tsai-Shien Chen, Chieh Hubert Lin, Hung-Yu Tseng, Tsung-Yi Lin, and Ming-Hsuan Yang. Motion-conditioned diffusion model for controllable video synthesis. arXiv preprint arXiv:2304.14404, 2023." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 299, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 299, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 312, + 299, + 545, + 353 + ], + "type": "text", + "content": "[19] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18000-18010, 2023." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 353, + 547, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 353, + 547, + 396 + ], + "spans": [ + { + "bbox": [ + 312, + 353, + 547, + 396 + ], + "type": "text", + "content": "[20] Yung-Yu Chuang, Dan B Goldman, Ke Colin Zheng, Brian Curless, David H Salesin, and Richard Szeliski. Animating pictures with stochastic motion textures. In ACM SIGGRAPH 2005 Papers, pages 853-860. 2005." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 396, + 547, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 396, + 547, + 437 + ], + "spans": [ + { + "bbox": [ + 312, + 396, + 547, + 437 + ], + "type": "text", + "content": "[21] Vincent C Couture, Michael S Langer, and Sebastien Roy. Omnistereo video textures without ghosting. In 2013 International Conference on 3D Vision-3DV 2013, pages 64–70. IEEE, 2013." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 438, + 545, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 438, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 312, + 438, + 545, + 470 + ], + "type": "text", + "content": "[22] Abe Davis, Justin G Chen, and Frédo Durand. Image-space modal bases for plausible manipulation of objects in video. ACM Transactions on Graphics (TOG), 34(6):1-7, 2015." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 470, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 470, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 312, + 470, + 545, + 491 + ], + "type": "text", + "content": "[23] Myers Abraham Davis. Visual vibration analysis. PhD thesis, Massachusetts Institute of Technology, 2016." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 491, + 545, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 491, + 545, + 523 + ], + "spans": [ + { + "bbox": [ + 312, + 491, + 545, + 523 + ], + "type": "text", + "content": "[24] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 523, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 523, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 312, + 523, + 545, + 566 + ], + "type": "text", + "content": "[25] Julien Diener, Mathieu Rodriguez, Lionel Baboud, and Lionel Reveret. Wind projection basis for real-time animation of trees. In Computer graphics forum, volume 28, pages 533-540. Wiley Online Library, 2009." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 566, + 545, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 566, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 312, + 566, + 545, + 597 + ], + "type": "text", + "content": "[26] Gianfranco Doretto, Alessandro Chiuso, Ying Nian Wu, and Stefano Soatto. Dynamic textures. International journal of computer vision, 51:91-109, 2003." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 597, + 547, + 661 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 597, + 547, + 661 + ], + "spans": [ + { + "bbox": [ + 312, + 597, + 547, + 661 + ], + "type": "text", + "content": "[27] Michael Dorkenwald, Timo Milbich, Andreas Blattmann, Robin Rombach, Konstantinos G. Derpanis, and Bjorn Ommer. Stochastic image-to-video synthesis using cinns. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3742-3753, June 2021." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 661, + 547, + 707 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 661, + 547, + 707 + ], + "spans": [ + { + "bbox": [ + 312, + 661, + 547, + 707 + ], + "type": "text", + "content": "[28] Yuming Du, Robin Kips, Albert Pumarola, Sebastian Starke, Ali Thabet, and Artsiom Sanakoyeu. Avatars grow legs: Generating smooth human motion from sparse tracking inputs with diffusion model. In Proceedings of the IEEE/CVF" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24150" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 73, + 288, + 706 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 72, + 73, + 287, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 73, + 287, + 94 + ], + "spans": [ + { + "bbox": [ + 72, + 73, + 287, + 94 + ], + "type": "text", + "content": "Conference on Computer Vision and Pattern Recognition, pages 481-490, 2023." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 95, + 288, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 95, + 288, + 148 + ], + "spans": [ + { + "bbox": [ + 53, + 95, + 288, + 148 + ], + "type": "text", + "content": "[29] Yuki Endo, Yoshihiro Kanamori, and Shigeru Kuriyama. Animating landscape: Self-supervised learning of decoupled motion and appearance for single-image video synthesis. ACM Transactions on Graphics (Proceedings of ACM SIGGRAPH Asia 2019), 38(6):175:1-175:19, 2019." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 190 + ], + "type": "text", + "content": "[30] Dave Epstein, Allan Jabri, Ben Poole, Alexei A Efros, and Aleksander Holynski. Diffusion self-guidance for controllable image generation. arXiv preprint arXiv:2306.00986, 2023." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 288, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 288, + 245 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 288, + 245 + ], + "type": "text", + "content": "[31] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7346-7356, 2023." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 245, + 288, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 245, + 288, + 298 + ], + "spans": [ + { + "bbox": [ + 53, + 245, + 288, + 298 + ], + "type": "text", + "content": "[32] Matthew Flagg, Atsushi Nakazawa, Qiushuang Zhang, Sing Bing Kang, Young Kee Ryu, Irfan Essa, and James M Rehg. Human video textures. In Proceedings of the 2009 symposium on Interactive 3D graphics and games, pages 199-206, 2009." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 299, + 288, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 299, + 288, + 342 + ], + "spans": [ + { + "bbox": [ + 53, + 299, + 288, + 342 + ], + "type": "text", + "content": "[33] Jean-Yves Franceschi, Edouard Delasalles, Mickaël Chen, Sylvain Lamprier, and Patrick Gallinari. Stochastic latent residual video prediction. In International Conference on Machine Learning, pages 3233-3246. PMLR, 2020." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 342, + 288, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 342, + 288, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 342, + 288, + 384 + ], + "type": "text", + "content": "[34] Ruohan Gao, Bo Xiong, and Kristen Grauman. Im2Flow: Motion hallucination from static images for action recognition. In Proc. Computer Vision and Pattern Recognition (CVPR), 2018." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 384, + 288, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 384, + 288, + 437 + ], + "spans": [ + { + "bbox": [ + 53, + 384, + 288, + 437 + ], + "type": "text", + "content": "[35] Songwei Ge, Thomas Hayes, Harry Yang, Xi Yin, Guan Pang, David Jacobs, Jia-Bin Huang, and Devi Parikh. Long video generation with time-agnostic vqgan and timesensitive transformer. arXiv preprint arXiv:2204.03638, 2022." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 438, + 288, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 438, + 288, + 481 + ], + "spans": [ + { + "bbox": [ + 53, + 438, + 288, + 481 + ], + "type": "text", + "content": "[36] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 481, + 288, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 481, + 288, + 524 + ], + "spans": [ + { + "bbox": [ + 53, + 481, + 288, + 524 + ], + "type": "text", + "content": "[37] Isma Hadji and Richard P Wildes. A new large scale dynamic texture dataset with application to convnet understanding. In Proceedings of the European Conference on Computer Vision (ECCV), pages 320-335, 2018." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 524, + 288, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 524, + 288, + 567 + ], + "spans": [ + { + "bbox": [ + 53, + 524, + 288, + 567 + ], + "type": "text", + "content": "[38] Zekun Hao, Xun Huang, and Serge Belongie. Controllable video generation with sparse trajectories. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7854-7863, 2018." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 567, + 288, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 567, + 288, + 609 + ], + "spans": [ + { + "bbox": [ + 53, + 567, + 288, + 609 + ], + "type": "text", + "content": "[39] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 610, + 288, + 662 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 610, + 288, + 662 + ], + "spans": [ + { + "bbox": [ + 53, + 610, + 288, + 662 + ], + "type": "text", + "content": "[40] Yingqing He, Menghan Xia, Haoxin Chen, Xiaodong Cun, Yuan Gong, Jinbo Xing, Yong Zhang, Xintao Wang, Chao Weng, Ying Shan, et al. animate-a-story: Storytelling with retrieval-augmented video generation. arXiv preprint arXiv:2307.06940, 2023." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 663, + 288, + 706 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 663, + 288, + 706 + ], + "spans": [ + { + "bbox": [ + 53, + 663, + 288, + 706 + ], + "type": "text", + "content": "[41] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2022." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 547, + 706 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 312, + 73, + 547, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 547, + 126 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 547, + 126 + ], + "type": "text", + "content": "[42] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 127, + 547, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 127, + 547, + 181 + ], + "spans": [ + { + "bbox": [ + 312, + 127, + 547, + 181 + ], + "type": "text", + "content": "[43] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 181, + 547, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 181, + 547, + 213 + ], + "spans": [ + { + "bbox": [ + 312, + 181, + 547, + 213 + ], + "type": "text", + "content": "[44] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 213, + 547, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 213, + 547, + 234 + ], + "spans": [ + { + "bbox": [ + 312, + 213, + 547, + 234 + ], + "type": "text", + "content": "[45] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 234, + 547, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 234, + 547, + 287 + ], + "spans": [ + { + "bbox": [ + 312, + 234, + 547, + 287 + ], + "type": "text", + "content": "[46] Aleksander Holynski, Brian L Curless, Steven M Seitz, and Richard Szeliski. Animating pictures with Eulerian motion fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5810-5819, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 288, + 547, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 288, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 312, + 288, + 547, + 319 + ], + "type": "text", + "content": "[47] Tobias Hoppe, Arash Mehrjou, Stefan Bauer, Didrik Nielsen, and Andrea Dittadi. Diffusion models for video prediction and infilling. Trans. Mach. Learn. Res., 2022, 2022." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 320, + 547, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 320, + 547, + 352 + ], + "spans": [ + { + "bbox": [ + 312, + 320, + 547, + 352 + ], + "type": "text", + "content": "[48] Xiaotao Hu, Zhewei Huang, Ailin Huang, Jun Xu, and Shuchang Zhou. A dynamic multi-scale voxel flow network for video prediction. ArXiv, abs/2303.09875, 2023." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 352, + 547, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 352, + 547, + 406 + ], + "spans": [ + { + "bbox": [ + 312, + 352, + 547, + 406 + ], + "type": "text", + "content": "[49] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 406, + 547, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 406, + 547, + 459 + ], + "spans": [ + { + "bbox": [ + 312, + 406, + 547, + 459 + ], + "type": "text", + "content": "[50] Hitoshi Kanda and Jun Ohya. Efficient, realistic method for animating dynamic behaviors of 3d botanical trees. In 2003 International Conference on Multimedia and Expo. ICME'03. Proceedings (Cat. No. 03TH8698), volume 2, pages II-89. IEEE, 2003." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 459, + 547, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 459, + 547, + 502 + ], + "spans": [ + { + "bbox": [ + 312, + 459, + 547, + 502 + ], + "type": "text", + "content": "[51] Johanna Karras, Aleksander Holynski, Ting-Chun Wang, and Ira Kemelmacher-Shlizerman. Dreampose: Fashion image-to-video synthesis via stable diffusion. arXiv preprint arXiv:2304.06025, 2023." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 502, + 547, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 502, + 547, + 555 + ], + "spans": [ + { + "bbox": [ + 312, + 502, + 547, + 555 + ], + "type": "text", + "content": "[52] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 555, + 547, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 555, + 547, + 587 + ], + "spans": [ + { + "bbox": [ + 312, + 555, + 547, + 587 + ], + "type": "text", + "content": "[53] Alex X Lee, Richard Zhang, Frederik Ebert, Pieter Abbeel, Chelsea Finn, and Sergey Levine. Stochastic adversarial video prediction. arXiv preprint arXiv:1804.01523, 2018." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 587, + 547, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 587, + 547, + 642 + ], + "spans": [ + { + "bbox": [ + 312, + 587, + 547, + 642 + ], + "type": "text", + "content": "[54] Zhengqi Li, Tali Dekel, Forrester Cole, Richard Tucker, Noah Snavely, Ce Liu, and William T Freeman. Learning the depths of moving people by watching frozen people. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4521-4530, 2019." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 642, + 547, + 685 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 642, + 547, + 685 + ], + "spans": [ + { + "bbox": [ + 312, + 642, + 547, + 685 + ], + "type": "text", + "content": "[55] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2041-2050, 2018." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 685, + 547, + 706 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 685, + 547, + 706 + ], + "spans": [ + { + "bbox": [ + 312, + 685, + 547, + 706 + ], + "type": "text", + "content": "[56] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snively. Dynibar: Neural dynamic image-based" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24151" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 73, + 288, + 705 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 72, + 73, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 73, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 72, + 73, + 287, + 105 + ], + "type": "text", + "content": "rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4273-4284, 2023." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 106, + 288, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 106, + 288, + 158 + ], + "spans": [ + { + "bbox": [ + 53, + 106, + 288, + 158 + ], + "type": "text", + "content": "[57] Zhengqi Li, Qianqian Wang, Noah Snavely, and Angjoo Kanazawa. Infinitenature-zero: Learning perpetual view generation of natural scenes from single images. In European Conference on Computer Vision, pages 515-534. Springer, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 159, + 288, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 159, + 288, + 191 + ], + "spans": [ + { + "bbox": [ + 54, + 159, + 288, + 191 + ], + "type": "text", + "content": "[58] Jing Liao, Mark Finch, and Hugues Hoppe. Fast computation of seamless video loops. ACM Transactions on Graphics (TOG), 34(6):1-10, 2015." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 192, + 288, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 192, + 288, + 223 + ], + "spans": [ + { + "bbox": [ + 53, + 192, + 288, + 223 + ], + "type": "text", + "content": "[59] Zicheng Liao, Neel Joshi, and Hugues Hoppe. Automated video looping with progressive dynamism. ACM Transactions on Graphics (TOG), 32(4):1-10, 2013." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 224, + 288, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 224, + 288, + 277 + ], + "spans": [ + { + "bbox": [ + 54, + 224, + 288, + 277 + ], + "type": "text", + "content": "[60] Andrew Liu, Richard Tucker, Varun Jampani, Ameesh Makadia, Noah Snavely, and Angjoo Kanazawa. Infinite nature: Perpetual view generation of natural scenes from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14458-14467, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 277, + 288, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 277, + 288, + 308 + ], + "spans": [ + { + "bbox": [ + 53, + 277, + 288, + 308 + ], + "type": "text", + "content": "[61] Ce Liu. Beyond pixels: exploring new representations and applications for motion analysis. PhD thesis, Massachusetts Institute of Technology, 2009." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 309, + 288, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 309, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 54, + 309, + 288, + 373 + ], + "type": "text", + "content": "[62] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Tieniu Tan. Videofusion: Decomposed diffusion models for high-quality video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10209-10218, 2023." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 374, + 288, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 374, + 288, + 405 + ], + "spans": [ + { + "bbox": [ + 54, + 374, + 288, + 405 + ], + "type": "text", + "content": "[63] Aniruddha Mahapatra and Kuldeep Kulkarni. Controllable animation of fluid elements in still images. In Proc. Computer Vision and Pattern Recognition (CVPR), 2022." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 406, + 288, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 406, + 288, + 437 + ], + "spans": [ + { + "bbox": [ + 53, + 406, + 288, + 437 + ], + "type": "text", + "content": "[64] Arun Mallya, Ting-Chun Wang, and Ming-Yu Liu. Implicit warping for animation with image sets. Advances in Neural Information Processing Systems, 35:22438-22450, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 437, + 288, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 437, + 288, + 491 + ], + "spans": [ + { + "bbox": [ + 53, + 437, + 288, + 491 + ], + "type": "text", + "content": "[65] Medhini Narasimhan, Shiry Ginosar, Andrew Owens, Alexei A Efros, and Trevor Darrell. Strumming to the beat: Audio-conditioned contrastive video textures. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3761-3770, 2022." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 491, + 288, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 491, + 288, + 544 + ], + "spans": [ + { + "bbox": [ + 53, + 491, + 288, + 544 + ], + "type": "text", + "content": "[66] Haomiao Ni, Changhao Shi, Kai Li, Sharon X Huang, and Martin Renqiang Min. Conditional image-to-video generation with latent flow diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18444-18455, 2023." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 545, + 288, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 545, + 288, + 586 + ], + "spans": [ + { + "bbox": [ + 53, + 545, + 288, + 586 + ], + "type": "text", + "content": "[67] Simon Niklaus and Feng Liu. Softmax splatting for video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5437-5446, 2020." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 588, + 288, + 641 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 588, + 288, + 641 + ], + "spans": [ + { + "bbox": [ + 53, + 588, + 288, + 641 + ], + "type": "text", + "content": "[68] Shin Ota, Machiko Tamura, Kunihiko Fujita, T Fujimoto, K Muraoka, and Norishige Chiba. 1/f/sup/spl beta//noise-based real-time animation of trees swaying in wind fields. In Proceedings Computer Graphics International 2003, pages 52-59. IEEE, 2003." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 642, + 288, + 695 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 642, + 288, + 695 + ], + "spans": [ + { + "bbox": [ + 53, + 642, + 288, + 695 + ], + "type": "text", + "content": "[69] Automne Petitjean, Yohan Poirier-Ginter, Ayush Tewari, Guillaume Cordonnier, and George Drettakis. Modalnerf: Neural modal analysis and synthesis for free-viewpoint navigation in dynamically vibrating scenes. In Computer Graphics Forum, volume 42, 2023." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 696, + 288, + 705 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 696, + 288, + 705 + ], + "spans": [ + { + "bbox": [ + 53, + 696, + 288, + 705 + ], + "type": "text", + "content": "[70] Silvia L. Pintea, Jan C. van Gemert, and Arnold W. M." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 547, + 705 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "type": "text", + "content": "Smeulders. Déjà vu: Motion prediction in static images. In Proc. European Conf. on Computer Vision (ECCV), 2014." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 95, + 547, + 125 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 95, + 547, + 125 + ], + "spans": [ + { + "bbox": [ + 312, + 95, + 547, + 125 + ], + "type": "text", + "content": "[71] Sigal Raab, Inbal Leibovitch, Guy Tevet, Moab Arar, Amit H Bermano, and Daniel Cohen-Or. Single motion diffusion. arXiv preprint arXiv:2302.05905, 2023." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 126, + 547, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 126, + 547, + 168 + ], + "spans": [ + { + "bbox": [ + 312, + 126, + 547, + 168 + ], + "type": "text", + "content": "[72] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2):3, 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 170, + 547, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 170, + 547, + 223 + ], + "spans": [ + { + "bbox": [ + 312, + 170, + 547, + 223 + ], + "type": "text", + "content": "[73] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 224, + 547, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 224, + 547, + 287 + ], + "spans": [ + { + "bbox": [ + 312, + 224, + 547, + 287 + ], + "type": "text", + "content": "[74] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 289, + 547, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 289, + 547, + 342 + ], + "spans": [ + { + "bbox": [ + 312, + 289, + 547, + 342 + ], + "type": "text", + "content": "[75] Payam Saisan, Gianfranco Doretto, Ying Nian Wu, and Stefano Soatto. Dynamic texture recognition. In Proceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition. CVPR 2001, volume 2, pages II-II. IEEE, 2001." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 342, + 547, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 342, + 547, + 384 + ], + "spans": [ + { + "bbox": [ + 312, + 342, + 547, + 384 + ], + "type": "text", + "content": "[76] Saurabh Saxena, Charles Herrmann, Junhwa Hur, Abhishek Kar, Mohammad Norouzi, Deqing Sun, and David J. Fleet. The surprising effectiveness of diffusion models for optical flow and monocular depth estimation, 2023." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 384, + 547, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 384, + 547, + 427 + ], + "spans": [ + { + "bbox": [ + 312, + 384, + 547, + 427 + ], + "type": "text", + "content": "[77] Arno Schödl, Richard Szeliski, David H Salesin, and Irfan Essa. Video textures. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 489-498, 2000." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 428, + 547, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 428, + 547, + 459 + ], + "spans": [ + { + "bbox": [ + 312, + 428, + 547, + 459 + ], + "type": "text", + "content": "[78] Mikio Shinya and Alain Fournier. Stochastic motion—motion under the influence of wind. Computer Graphics Forum, 11(3), 1992." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 460, + 547, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 460, + 547, + 513 + ], + "spans": [ + { + "bbox": [ + 312, + 460, + 547, + 513 + ], + "type": "text", + "content": "[79] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. Animating arbitrary objects via deep motion transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2377-2386, 2019." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 514, + 547, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 514, + 547, + 555 + ], + "spans": [ + { + "bbox": [ + 312, + 514, + 547, + 555 + ], + "type": "text", + "content": "[80] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. Advances in neural information processing systems, 32, 2019." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 556, + 547, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 556, + 547, + 609 + ], + "spans": [ + { + "bbox": [ + 312, + 556, + 547, + 609 + ], + "type": "text", + "content": "[81] Aliaksandr Siarohin, Oliver J Woodford, Jian Ren, Mengei Chai, and Sergey Tulyakov. Motion representations for articulated animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13653-13662, 2021." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 610, + 547, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 610, + 547, + 642 + ], + "spans": [ + { + "bbox": [ + 312, + 610, + 547, + 642 + ], + "type": "text", + "content": "[82] Chen Qian Kwan-Yee Lin Hongsheng Li Siming Fan, Jingtan Piao. Simulating fluids in real-world still images. arXiv preprint, arXiv:2204.11335, 2022." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 642, + 547, + 695 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 642, + 547, + 695 + ], + "spans": [ + { + "bbox": [ + 312, + 642, + 547, + 695 + ], + "type": "text", + "content": "[83] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3626-3636, 2022." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 696, + 547, + 705 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 696, + 547, + 705 + ], + "spans": [ + { + "bbox": [ + 312, + 696, + 547, + 705 + ], + "type": "text", + "content": "[84] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan," + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "24152" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 288, + 704 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 72, + 73, + 288, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 73, + 288, + 106 + ], + "spans": [ + { + "bbox": [ + 72, + 73, + 288, + 106 + ], + "type": "text", + "content": "and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 106, + 288, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 106, + 288, + 137 + ], + "spans": [ + { + "bbox": [ + 53, + 106, + 288, + 137 + ], + "type": "text", + "content": "[85] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv:2010.02502, October 2020." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 137, + 288, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 137, + 288, + 179 + ], + "spans": [ + { + "bbox": [ + 54, + 137, + 288, + 179 + ], + "type": "text", + "content": "[86] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 180, + 288, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 180, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 54, + 180, + 288, + 201 + ], + "type": "text", + "content": "[87] Jos Stam. Multi-scale stochastic modelling of complex natural phenomena. PhD thesis, 1995." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 201, + 288, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 201, + 288, + 232 + ], + "spans": [ + { + "bbox": [ + 54, + 201, + 288, + 232 + ], + "type": "text", + "content": "[88] Jos Stam. Stochastic dynamics: Simulating the effects of turbulence on flexible structures. Computer Graphics Forum, 16(3), 1997." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 233, + 288, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 233, + 288, + 275 + ], + "spans": [ + { + "bbox": [ + 54, + 233, + 288, + 275 + ], + "type": "text", + "content": "[89] Ryusuke Sugimoto, Mingming He, Jing Liao, and Pedro V Sander. Water simulation and rendering from a still photograph. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 275, + 288, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 275, + 288, + 308 + ], + "spans": [ + { + "bbox": [ + 54, + 275, + 288, + 308 + ], + "type": "text", + "content": "[90] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 308, + 288, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 308, + 288, + 350 + ], + "spans": [ + { + "bbox": [ + 54, + 308, + 288, + 350 + ], + "type": "text", + "content": "[91] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 350, + 288, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 350, + 288, + 393 + ], + "spans": [ + { + "bbox": [ + 54, + 350, + 288, + 393 + ], + "type": "text", + "content": "[92] Vikram Voleti, Alexia Jolicoeur-Martineau, and Christopher Pal. Mcvd: Masked conditional video diffusion for prediction, generation, and interpolation. In (NeurIPS) Advances in Neural Information Processing Systems, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 393, + 288, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 393, + 288, + 425 + ], + "spans": [ + { + "bbox": [ + 54, + 393, + 288, + 425 + ], + "type": "text", + "content": "[93] Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba. Generating videos with scene dynamics. In Neural Information Processing Systems, 2016." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 425, + 288, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 425, + 288, + 467 + ], + "spans": [ + { + "bbox": [ + 54, + 425, + 288, + 467 + ], + "type": "text", + "content": "[94] Neal Wadhwa, Michael Rubinstein, Frédo Durand, and William T Freeman. Phase-based video motion processing. ACM Transactions on Graphics (ToG), 32(4):1-10, 2013." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 54, + 468, + 288, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 468, + 288, + 511 + ], + "spans": [ + { + "bbox": [ + 54, + 468, + 288, + 511 + ], + "type": "text", + "content": "[95] Jacob Walker, Carl Doersch, Abhinav Gupta, and Martial Hebert. An uncertain future: Forecasting from static images using variational autoencoders. In Proc. European Conf. on Computer Vision (ECCV), 2016." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 54, + 511, + 288, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 511, + 288, + 553 + ], + "spans": [ + { + "bbox": [ + 54, + 511, + 288, + 553 + ], + "type": "text", + "content": "[96] Jacob Walker, Abhinav Gupta, and Martial Hebert. Dense optical flow prediction from a static image. In Proceedings of the IEEE International Conference on Computer Vision, pages 2443-2451, 2015." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 54, + 554, + 288, + 607 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 554, + 288, + 607 + ], + "spans": [ + { + "bbox": [ + 54, + 554, + 288, + 607 + ], + "type": "text", + "content": "[97] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. arXiv preprint arXiv:2306.02018, 2023." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 54, + 607, + 288, + 650 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 607, + 288, + 650 + ], + "spans": [ + { + "bbox": [ + 54, + 607, + 288, + 650 + ], + "type": "text", + "content": "[98] Yaohui Wang, Di Yang, Francois Bremond, and Antitza Dantcheva. Latent image animator: Learning to animate images via latent space navigation. arXiv preprint arXiv:2203.09043, 2022." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 54, + 650, + 288, + 693 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 650, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 54, + 650, + 288, + 693 + ], + "type": "text", + "content": "[99] Frederik Warburg, Ethan Weber, Matthew Tancik, Aleksander Holynski, and Angjoo Kanazawa. Nerfbusters: Removing ghostly artifacts from casually captured nerfs. arXiv preprint arXiv:2304.10532, 2023." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 693, + 288, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 693, + 288, + 704 + ], + "spans": [ + { + "bbox": [ + 48, + 693, + 288, + 704 + ], + "type": "text", + "content": "[100] Daniel Watson, William Chan, Ricardo Martin-Brualla," + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 567 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 331, + 73, + 547, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 73, + 547, + 106 + ], + "spans": [ + { + "bbox": [ + 331, + 73, + 547, + 106 + ], + "type": "text", + "content": "Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. arXiv preprint arXiv:2210.04628, 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 106, + 547, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 106, + 547, + 159 + ], + "spans": [ + { + "bbox": [ + 308, + 106, + 547, + 159 + ], + "type": "text", + "content": "[101] Chung-Yi Weng, Brian Curless, and Ira Kemelmacher-Shlizerman. Photo wake-up: 3d character animation from a single photo. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5908-5917, 2019." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 159, + 547, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 159, + 547, + 191 + ], + "spans": [ + { + "bbox": [ + 308, + 159, + 547, + 191 + ], + "type": "text", + "content": "[102] Jamie Wynn and Daniyar Turmukhambetov. DiffusioNeRF: Regularizing Neural Radiance Fields with Denoising Diffusion Models. In CVPR, 2023." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 191, + 547, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 191, + 547, + 244 + ], + "spans": [ + { + "bbox": [ + 308, + 191, + 547, + 244 + ], + "type": "text", + "content": "[103] Tianfan Xue, Jiajun Wu, Katherine L Bouman, and William T Freeman. Visual dynamics: Stochastic future generation via layered cross convolutional networks. Trans. Pattern Analysis and Machine Intelligence, 41(9):2236-2250, 2019." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 245, + 546, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 245, + 546, + 288 + ], + "spans": [ + { + "bbox": [ + 308, + 245, + 546, + 288 + ], + "type": "text", + "content": "[104] Shengming Yin, Chenfei Wu, Jian Liang, Jie Shi, Houqiang Li, Gong Ming, and Nan Duan. Dragnuwa: Fine-grained control in video generation by integrating text, image, and trajectory. arXiv preprint arXiv:2308.08089, 2023." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 288, + 547, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 288, + 547, + 341 + ], + "spans": [ + { + "bbox": [ + 308, + 288, + 547, + 341 + ], + "type": "text", + "content": "[105] Sihyun Yu, Kihyuk Sohn, Subin Kim, and Jinwoo Shin. Video probabilistic diffusion models in projected latent space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18456-18466, 2023." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 341, + 547, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 341, + 547, + 384 + ], + "spans": [ + { + "bbox": [ + 308, + 341, + 547, + 384 + ], + "type": "text", + "content": "[106] Mingyuan Zhang, Xinying Guo, Liang Pan, Zhongang Cai, Fangzhou Hong, Huirong Li, Lei Yang, and Ziwei Liu. Remodiffuse: Retrieval-augmented motion diffusion model. arXiv preprint arXiv:2304.01116, 2023." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 384, + 546, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 384, + 546, + 427 + ], + "spans": [ + { + "bbox": [ + 308, + 384, + 546, + 427 + ], + "type": "text", + "content": "[107] Yabo Zhang, Yuxiang Wei, Dongsheng Jiang, Xiaopeng Zhang, Wangmeng Zuo, and Qi Tian. Controlvideo: Training-free controllable text-to-video generation. arXiv preprint arXiv:2305.13077, 2023." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 427, + 546, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 427, + 546, + 470 + ], + "spans": [ + { + "bbox": [ + 308, + 427, + 546, + 470 + ], + "type": "text", + "content": "[108] Jian Zhao and Hui Zhang. Thin-plate spline motion model for image animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3657–3666, 2022." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 470, + 546, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 470, + 546, + 524 + ], + "spans": [ + { + "bbox": [ + 308, + 470, + 546, + 524 + ], + "type": "text", + "content": "[109] Shengyu Zhao, Jonathan Cui, Yilun Sheng, Yue Dong, Xiao Liang, Eric I Chang, and Yan Xu. Large scale image completion via co-modulated generative adversarial networks. In International Conference on Learning Representations (ICLR), 2021." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 524, + 546, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 524, + 546, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 524, + 546, + 567 + ], + "type": "text", + "content": "[110] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022." + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24153" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Generative Latent Coding for Ultra-Low Bitrate Image Compression/a71ce275-c18c-4862-b210-a735f3bb6944_content_list.json b/2024/Generative Latent Coding for Ultra-Low Bitrate Image Compression/a71ce275-c18c-4862-b210-a735f3bb6944_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..30603d61c1206c3a572fe2c07fc5153cf4a1ff6d --- /dev/null +++ b/2024/Generative Latent Coding for Ultra-Low Bitrate Image Compression/a71ce275-c18c-4862-b210-a735f3bb6944_content_list.json @@ -0,0 +1,2059 @@ +[ + { + "type": "text", + "text": "Generative Latent Coding for Ultra-Low Bitrate Image Compression", + "text_level": 1, + "bbox": [ + 137, + 130, + 831, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhaoyang Jia $^{1*}$ Jiahao Li $^{2}$ Bin Li $^{2}$ Houqiang Li $^{1}$ Yan Lu $^{2}$", + "bbox": [ + 181, + 179, + 782, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 University of Science and Technology of China 2 Microsoft Research Asia", + "bbox": [ + 181, + 199, + 789, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "jzy_ustc@mail.ustc.edu.cn, lihq@ustc.edu.cn, {li.jiahao, libin, yanlu}@microsoft.com", + "bbox": [ + 114, + 219, + 849, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Most existing image compression approaches perform transform coding in the pixel space to reduce its spatial redundancy. However, they encounter difficulties in achieving both high-realism and high-fidelity at low bitrate, as the pixel-space distortion may not align with human perception. To address this issue, we introduce a Generative Latent Coding (GLC) architecture, which performs transform coding in the latent space of a generative vector-quantized variational auto-encoder (VQ-VAE), instead of in the pixel space. The generative latent space is characterized by greater sparsity, richer semantic and better alignment with human perception, rendering it advantageous for achieving high-realism and high-fidelity compression. Additionally, we introduce a categorical hyper module to reduce the bit cost of hyper-information, and a code-prediction-based supervision to enhance the semantic consistency. Experiments demonstrate that our GLC maintains high visual quality with less than 0.04 bpp on natural images and less than 0.01 bpp on facial images. On the CLIC2020 test set, we achieve the same FID as MS-ILLM with $45\\%$ fewer bits. Furthermore, the powerful generative latent space enables various applications built on our GLC pipeline, such as image restoration and style transfer.", + "bbox": [ + 75, + 301, + 473, + 650 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 669, + 209, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Amid the ongoing surge of digital visual data, the importance of achieving high-efficiency image compression becomes increasingly paramount. From the traditional compression standards [54, 55] to the emerging learned image compression models [4, 5, 11, 20, 41, 45], most compression algorithms follow the pixel-space transform coding [4, 19] paradigm. Specifically, they convert pixels into compact representations through a transform module, which eliminates the redundancy to reduce the bit cost in the subsequent entropy coding process.", + "bbox": [ + 75, + 695, + 468, + 847 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, we observe a common inherent limitation in", + "bbox": [ + 96, + 848, + 468, + 863 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a2ba8fbb87b5b0653a8d05ef617dd3e5b676dab018b51f1cf11a13168b793beb.jpg", + "image_caption": [ + "Figure 1. Generative latent space of VQ-VAE exhibits better alignment with human perception than pixel space for ultra-low bitrate compression. Under comparable distortion levels (measured by signal-to-noise ratio, SNR), latent-space compression produces reconstructions with superior perceptual quality (measured by DISTS [15]) than pixel-space generative codec MS-ILLM [46], as the compressed latents remain in the same latent code space." + ], + "image_footnote": [], + "bbox": [ + 508, + 268, + 885, + 494 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "these methods: the pixel-space distortion is not always consistent with the human perception, especially at low bitrate. In practice, human observers prioritize the semantic consistency and the texture realism of an image, but this information is difficult to be adequately exploited solely by a pixel-space transform module. As shown in the left of Figure 1, pixel-space generative image codec MS-ILLM [46] struggles to guarantee visual quality at low bitrate, even after it incorporates perceptual supervision [29] and adversarial supervision [18] within the pixel space.", + "bbox": [ + 496, + 612, + 893, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Based on such observation, a natural problem arises: how can we compress images in a way that aligns with human perception? To address this challenge, we introduce a Generative Latent Coding (GLC) paradigm. In GLC, we first encode images into a generative latent space that aligns with human perception, and subsequently perform transform coding in this latent space, instead of in pixel space. To concretize this concept, we adopt a generative vector-quantized variational auto-encoder (VQ-VAE) [17, 53] to", + "bbox": [ + 496, + 763, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*This work was done when Zhaoyang Jia was an intern at Microsoft Research Asia.", + "bbox": [ + 76, + 875, + 468, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "26088", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fa2f7ad2599bdd232dd9af00fa873e08404ca7c186b9f152ea0b146474fe767e.jpg", + "image_caption": [ + "Figure 2. A qualitative comparison between HiFiC [44], MS-ILLM [46], Text+Sketch [35] and the proposed GLC. GLC produces images with high visual quality, even in regions with complex texture. In contrast, HiFiC and MS-ILLM exhibit noticeable artifacts, and Text+Sketch generates results that deviate significantly from the input. Best viewed when zoomed in." + ], + "image_footnote": [], + "bbox": [ + 83, + 89, + 885, + 385 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "produce the latent space, which offers three significant advantages: 1) The discrete codes of VQ-VAE encapsulate semantic visual components [53], allowing GLC to focus on compressing the semantic content to guarantee fidelity. 2) Generative VQ-VAE exhibits remarkable generative capabilities [17] for high-realism texture reconstruction. 3) The discrete variational bottleneck naturally brings a low-entropy and distortion-robust latent space for compression. Thanks to such characteristics, GLC is more aligned with human perception to achieve enhanced visual quality, as demonstrated in the right of Figure 1.", + "bbox": [ + 75, + 459, + 472, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "When implementing GLC, two crucial questions persist: How to effectively compress the generative latents? And how to supervise the generative latent coding? A straightforward approach to compress VQ-VAE latents is indices-map coding [27, 28, 43], but it is limited by the ineffective redundancy reduction between indices and the lack of rate-variable coding support. In this paper, we propose a novel generative-latent-space transform coding approach, where an effective rate-variable structure is adopted to reduce latent redundancy for higher compression ratio. In addition, a categorical hyper module is introduced to model the distribution of $z$ with a discrete codebook, which significantly reduces the bitrate of $z$ when compared to the factorized hyper module [5]. As for the supervision of GLC, inspired by recent code prediction transformers [27, 28, 60], a code-prediction-based supervision is proposed. It serves as an auxiliary supervision employed solely in the training process to greatly enhance the semantic consistency.", + "bbox": [ + 75, + 628, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Benefited from these advanced designs, our GLC achieves excellent performance on both natural and facial images. In the CLIC 2020 test set [51], GLC attains a bitrate less than 0.04 bpp while delivering high visual quality. It obtains $45\\%$ bit savings compared to MS-ILLM at the same FID. In the CelebAHQ [30] dataset, GLC achieves an even lower bitrate of less than 0.01 bpp. As shown in Figure 2, compared with recent advanced generative image compression approaches MS-ILLM [46] and Text+Sketch [35], GLC produces more visually appealing compression results with a lower bit cost.", + "bbox": [ + 496, + 459, + 890, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, leveraging the representative generative latent space, GLC supports various vision applications such as image restoration and style transfer. By replacing the compression encoder with a restoration encoder, the proposed restoration application surpasses the performance of cascading a restoration model and a neural codec. We hope such versatility of generative latent space will help connect image compression with other vision tasks in the future.", + "bbox": [ + 496, + 625, + 890, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our main contributions are :", + "bbox": [ + 517, + 747, + 790, + 760 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present a generative latent coding (GLC) scheme, which performs transform coding in the generative latent space of a VQ-VAE to achieve high-fidelity and high-realism compression at ultra-low bitrate.", + "- We introduce a categorical hyper module to significantly reduce the bit cost of hyper information. Additionally, a code-prediction-based supervision is adopted to enhance the perceptual quality.", + "- GLC obtains $45\\%$ bit reduction on CLIC2020 with the" + ], + "bbox": [ + 500, + 763, + 890, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "26089", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/752991a6a6488088bb4007ad888aa69defed14166a7fc044e13ee45a4aeb716a.jpg", + "image_caption": [ + "Figure 3. Illustration of the proposed Generative Latent Coding (GLC) framework. (Left) GLC firstly encodes the image into a generative latent representation (Section 3.2), then compresses the latent with transform coding (Section 3.3), and finally reconstructs image from the compressed latent. (Right) We progressively train GLC in three stages (Section 4): In stage I, we train a generative VQ-VAE to obtain a human-perception-aligned latent space. In stage II, the transform coding module learns to compress the latent with a code-prediction-based latent supervision (Figure 6). Finally, in stage III, the entire network is fine-tuned jointly with a code-prediction-based pixel supervision to further enhance the compression performance." + ], + "image_footnote": [], + "bbox": [ + 84, + 97, + 385, + 291 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2f1e3acd8c9ce4c5ce81e9bd16fd4673e8ab73f694a8683bdf1aaad66607829a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 395, + 97, + 519, + 287 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e9a7896199c88e7736807b01977446915c7124d5a9e6e6b418b8185296c2d2f4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 522, + 98, + 666, + 286 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0a0612e6323cbe6ddcedcc5580590af275e60cfcf2e2398da48b3985f50e2232.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 98, + 883, + 286 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "same FID as the previous SOTA. Furthermore, GLC enables various application within its latent space.", + "bbox": [ + 89, + 401, + 470, + 433 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 76, + 444, + 227, + 460 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Learned Image Compression", + "text_level": 1, + "bbox": [ + 76, + 469, + 334, + 486 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Lossy image compression is grounded on Shannon's rate-distortion theory [12]. Balle et al. [4] first proposed to utilize neural networks for pixel-space transform coding [19], employing analysis and synthesis transform modules to convert images into compact representations for entropy coding. Subsequently, some researches make strides in improving the probability model [4, 5, 11, 34, 38, 45] for more accurate estimation, while others explore the network structure [11, 41], optimization algorithm [58, 59] and rate-variable coding [13, 20] for improved compression performance and practicality.", + "bbox": [ + 75, + 492, + 468, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A recently raised critical challenge in image compression is how to improve the perceptual quality of the reconstruction. Agustsson et al. [2] first introduced the concept of generative compression, which compresses essential image features and generate distorted details using GAN. Some subsequent works [9, 25] extract image sketches and latent codes to ensure geometry-consistency. Text+Sketch [35] utilizes a conditional-diffusion model to generate image based on image captions and sketches, achieving superior perceptual quality. While these schemes produce visual appealing results, they often deviate significantly from the input and cannot guarantee the semantic consistency.", + "bbox": [ + 75, + 659, + 470, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To achieve generative compression with high-fidelity, Mentzer et al. [44] further studied the network structure and generative adversarial loss to enable high-fidelity compression. Subsequent researches further enhance the trans", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "form coding [16, 22], generative post-processing [24] or focus on controlling the trade-off between fidelity and realism [3, 26]. Recently, MS-ILLM [46] introduces a noisy discriminator which is conditioned on quantized local image representations to greatly enhance the statistical fidelity.", + "bbox": [ + 496, + 401, + 893, + 492 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Latent Space Modeling", + "text_level": 1, + "bbox": [ + 500, + 508, + 715, + 526 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Latent space image modeling means modeling the distribution of image within the latent space of a neural network. This technique has been primarily developed for image generation. Chen et al. [10] and Oord et al. [53] introduced PixelCNN [52] in the latent space of VAE and VQ-VAE for image generation. Esser et al. [17] took it a step further by incorporating transformers into VQ-VAE latent space for high-quality generation. More recently, Rombach et al. [48] employed diffusion models to model the latent space of VQ-VAE, achieving remarkable results in high-resolution image generation. These studies underscore the potential of image processing within the generative latent space, particularly in the latent space of VQ-VAE.", + "bbox": [ + 496, + 535, + 893, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, the concept of latent space modeling has been extended to other tasks. CodeFormer [60] introduces a code prediction transformer that takes distorted latents as input and predicts the high-quality VQ-VAE index for facial restoration. Building upon this, Jiang et al. [27, 28] proposed transmitting the predicted indices to achieve restoration-based facial conferencing. In this paper, we explore the characteristics of latent space modeling in the realm of generative image compression. We specially design a transform coding paradigm in the latent space, which demonstrates superior effectiveness.", + "bbox": [ + 496, + 734, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "26090", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8b3b196db7993380f4bd7fac041dc1d0e812942d29f6640bae52acfac4d3d834.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 395, + 213 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3ff63c56751249968e11c1c1c51505c547412dcc4985e1ea5f7d574356d56700.jpg", + "image_caption": [ + "(b)", + "Figure 4. Illustration of the transform coding in latent space. (a) The model structure of the transform coding module. We further compare it with other coding schemes in operational diagrams : (b) indices-map coding [27, 28, 43], (c) transform coding with factorized hyper module [5, 21, 38] and (d) proposed transform coding with categorical hyper module. Here AE and AD denote arithmetic encoding and decoding, VQ-E and VQ-D stand for VQ-indices-map encoding and decoding, Q refers to scalar quantization, U signifies the addition of uniform noise as a differential simulation of Q, and S denotes the spatial context entropy module." + ], + "image_footnote": [], + "bbox": [ + 418, + 130, + 547, + 210 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/69fe87705e8f0aef6b5878c6b940b55d3f2c1a71d562bf3eed581a7b650006af.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 565, + 90, + 715, + 210 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8c14a71bca9fa0a737d269db45feda0d7cd8091ed7b118e487613a365934fff0.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 723, + 90, + 872, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 334, + 166, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 76, + 359, + 187, + 373 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we introduce the details of the proposed Generative Latent Coding (GLC) architecture. To achieve high-perceptual-quality compression, GLC encodes image into a perception-aligned latent space through a generative latent auto-encoder, and perform transform coding on the latent representations for lower bitrate. As depicted in the left of Figure 3, the input image $x$ is firstly encoded into the latent $l$ using the latent encoder $E$ . Then $l$ undergoes an analysis transform $g_{a}$ to produce the code $y$ , which is further scalar-quantized to $\\hat{y}$ for entropy coding. Following that, a synthesis transform $g_{s}$ is employed to transform $\\hat{y}$ back to $l$ , and finally, the reconstruction $\\hat{x}$ is generated by the latent decoder $D$ . This entire process is formulated as:", + "bbox": [ + 75, + 383, + 470, + 579 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} l = E (x), y = g _ {a} (l) \\\\ \\hat {y} = Q (y) \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 585, + 468, + 622 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {l} = g _ {s} (\\hat {y}), \\hat {x} = D (\\hat {l})\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 625, + 343, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Generative Latent Auto-Encoder", + "text_level": 1, + "bbox": [ + 76, + 650, + 366, + 664 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To achieve high-quality generative latent coding, how to obtain a human-perception-aligned latent space is a crucial challenge. In GLC, we address it by employing the generative VQ-VAE [17] as the latent auto-encoder $(E$ and $D)$ . By mapping images into visual semantic elements within a codebook $C$ and incorporating a generative image decoding process, both semantic consistency and texture realism can be well guaranteed. Additionally, it contributes to the compression process through a sparse yet robust latent space, which is achieved by training with the discrete codebook $C$ as a variational bottleneck.", + "bbox": [ + 75, + 672, + 468, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Transform Coding in Latent Space", + "text_level": 1, + "bbox": [ + 76, + 847, + 380, + 863 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To compress the latent representations $l$ , a direct approach is VQ-indices-map coding [27, 28, 43] (Figure 4b). How-", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/afbc0991a5095cefef3137c9837b15d425912a21a0a1d60cfa2d6d50f4bf4747.jpg", + "image_caption": [ + "Original Image" + ], + "image_footnote": [], + "bbox": [ + 506, + 392, + 612, + 473 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/104b96d7fa6777e12ed848e401740afd8bdfefb4a8c50a8a72355f4634a2cf58.jpg", + "image_caption": [ + "$y = 0$ bits $z = 1000$ bits" + ], + "image_footnote": [], + "bbox": [ + 627, + 342, + 736, + 422 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/460943cf38de08e5b4ae78112309d23b141f4b44b0ab5c198f09e76c39f6453c.jpg", + "image_caption": [ + "$y = 0$ bits $z = 160$ bits" + ], + "image_footnote": [], + "bbox": [ + 761, + 342, + 867, + 422 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a9e64e9ac1fcf78ce2221a773ae859d7eff1aedcd0fd33877f0e2af639dca0e9.jpg", + "image_caption": [ + "$y = 1800$ bits $z = 1000$ bits", + "Factorized z", + "Figure 5. Example of comparison between factorized and categorical hyper modules. The proposed categorical $z$ encodes essential semantic and structural information with much less bits." + ], + "image_footnote": [], + "bbox": [ + 630, + 438, + 736, + 518 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7fc7a3ee1aa2ba6b560866f36b38ac79bd69924bab8af1fb7088576a00c6b53d.jpg", + "image_caption": [ + "$y = 1640$ bits $z = 160$ bits", + "Categorical z" + ], + "image_footnote": [], + "bbox": [ + 764, + 438, + 867, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ever, these methods often lack a careful consideration of the correlation among the latents, resulting in an insufficient redundancy reduction and consequently a high bit cost. In GLC, we introduce a transform coding module to compress the latent, replacing the vector-quantization step for more effective reduction of latent redundancy. As shown in Figure 4a, the latents are transformed into code $y$ using transformations $g_{a}$ , $g_{s}$ and then quantized to $\\hat{y}$ . Entropy coding is applied to $\\hat{y}$ based on a probability $p(\\hat{y})$ , which is estimated by a categorical hyper module (Section 3.3.1) and a quadtree-partition-based spatial context module [38].", + "bbox": [ + 496, + 628, + 890, + 796 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1 Categorical Hyper Module", + "text_level": 1, + "bbox": [ + 500, + 815, + 741, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Factorized hyper module [5] (Figure 4c) is commonly employed in recent image compression schemes. However, at ultra-low bitrate, we notice that the factorized $z$ tends to encode low-level information such as color and texture, incur", + "bbox": [ + 496, + 839, + 890, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "26091", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ring a high bit cost, as illustrated in Figure 5. To address it, we propose a categorical hyper module (Figure 4d), which utilizes a hyper codebook to store the basic semantic elements rather than the low-level information. This module comprises a hyper analysis transform $h_a$ , a hyper synthesis transform $h_s$ and a hyper codebook $C_h$ . The transformations are formulated as:", + "bbox": [ + 75, + 90, + 472, + 196 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nz = h _ {a} (y), \\hat {z} = V Q (z, C _ {h}), p r i o r _ {z} = h _ {s} (\\hat {z}) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 204, + 468, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $z$ and $\\hat{z}$ denote hyper-codes. $VQ(\\cdot, C_h)$ represents vector-quantization by nearest lookup in $C_h$ . As shown in the right of Figure 5, the categorical $z$ is more inclined to capture high-level semantic information, which can be encoded with significantly fewer bits.", + "bbox": [ + 75, + 227, + 468, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.2 Rate-Variable Transformation", + "text_level": 1, + "bbox": [ + 76, + 321, + 344, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A notable advantage of transform coding over VQ-indices-map coding is its capability for rate-variable compression, which is a core functionality for a practical image codec. Indices-map coding is limited since the codebook can only model one specified distribution, but different rates naturally need different distributions. In contrast, transform coding converts latent into a unified Gaussian distribution, and variable-rate can be achieved by variable parameters (e.g., means and scales) of Gaussian. In GLC, we follow DCVC series [36-39, 47, 49] to incorporate a scalar $q$ in the transform coding to achieve rate-variable compression.", + "bbox": [ + 75, + 344, + 472, + 512 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Progressive Training", + "text_level": 1, + "bbox": [ + 76, + 523, + 276, + 540 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As depicted in the right of Figure 3, we adopt a three-stage progressive training manner to fully leverage the potential of GLC. We initially learn a human-perception-aligned latent space to guarantee the perceptual quality, subsequently learn to perform transform coding on this latent space to achieve low bitrate, and finally fine-tune the entire network for superior compression performance. At each stage, distinct loss functions are adopted to guide different modules.", + "bbox": [ + 75, + 547, + 468, + 670 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Stage I: Auto-Encoder Learning", + "text_level": 1, + "bbox": [ + 76, + 676, + 366, + 693 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To obtain a human-perception-aligned latent space for compression, we begin with training a generative VQ-VAE as the initialization of $E$ and $D$ . To ensure the sparsity of the latent space, an auxiliary codebook $C$ is employed to perform nearest vector-quantization, transforming $l$ to $\\widetilde{l}$ . The supervision comprises reconstruction loss, perceptual loss [29], adversarial loss [18] and codebook loss [53]:", + "bbox": [ + 75, + 700, + 468, + 808 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {S t a g e I}} = \\left| \\left| x - \\hat {x} \\right| \\right| + \\mathcal {L} _ {p e r} (x, \\hat {x}) \\tag {3} \\\\ + \\lambda_ {a d v} \\cdot \\mathcal {L} _ {a d v} (x, \\hat {x}) + \\mathcal {L} _ {\\text {c o d e b o o k}} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 813, + 468, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $\\mathcal{L}_{per}$ corresponds to the LPIPS loss calculated using VGG [50] extracted features. $\\mathcal{L}_{adv}$ is the adaptive PatchGAN adversarial loss [17] with a weight of $\\lambda_{adv} = 0.8$ .", + "bbox": [ + 75, + 854, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/686772d25e674bd1209af91b09224489b63f2e1a1818e78f9f6dcf8bb2ca6c84.jpg", + "image_caption": [ + "Figure 6. Illustration of code-prediction-based supervision." + ], + "image_footnote": [], + "bbox": [ + 506, + 92, + 697, + 282 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/42fa88018f2af32acdaf021d1d840493356db5e66eec6dbae839dee3a623d4b3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 92, + 883, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The codebook loss is formulated as", + "bbox": [ + 500, + 338, + 736, + 352 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o d e b o o k}} = \\left| \\left| \\operatorname {s g} (l) - \\tilde {l} \\right| \\right| + \\beta \\cdot \\left| \\left| \\operatorname {s g} (\\tilde {l}) - l \\right| \\right| \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 362, + 890, + 381 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{sg}(\\cdot)$ denotes the stop-gradient operator and $\\beta = 0.25$ controls the update rates of the $E$ and $C$ .", + "bbox": [ + 498, + 392, + 892, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Stage II: Transform Coding Learning", + "text_level": 1, + "bbox": [ + 498, + 431, + 830, + 449 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the trained latent space, we further learn to perform transform coding to achieve low-bitrate latent compression, while fixing the auto-encoder $E$ and $D$ . We introduce an auxiliary code predictor $CP$ to enhance the semantic consistency by necessitating the latent to possess the capability to predict the correct VQ-indices. As shown in Figure 6, we encode $l$ into VQ-indices by $M_l = VQ(l,C)$ and subsequently predict these indices by $\\hat{M}_{\\hat{l}} = CP(\\hat{l})$ . So the code-prediction-based loss can be formulated by", + "bbox": [ + 496, + 455, + 890, + 592 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} _ {c o d e} (l, \\hat {l}) = \\alpha \\cdot C E (M _ {l}, \\hat {M _ {l}}) + | | l - \\hat {l} | | _ {2} ^ {2} \\qquad (5)\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 602, + 890, + 622 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $CE$ denotes the cross entropy loss and we set $\\alpha = 0.5$ by default. Then the transform coding module can be supervised by the rate-distortion trade-off", + "bbox": [ + 498, + 632, + 892, + 676 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {S t a g e I I}} = \\mathbf {E} _ {x \\sim p _ {X}} [ \\mathcal {R} (\\hat {y}) + \\lambda \\cdot \\mathcal {D} _ {\\text {c o d e}} (l, \\hat {l}) ] \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 688, + 890, + 707 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{R}$ is the estimated rate and $\\lambda$ is used to control the trade-off. Note that a codebook loss (as formulated in Equation 4) is required to train the hyper codebook $C_h$ in the categorical hyper module. We omit it from the loss functions of both stage II and stage III for the sake of conciseness.", + "bbox": [ + 496, + 715, + 890, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Stage III: Joint Training", + "text_level": 1, + "bbox": [ + 500, + 801, + 730, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we fine-tune the entire network with the pixel space supervision to achieve better compression performance. As shown in Figure 6, we extend the code-prediction-based latent supervision into the pixel space. Specifically, we utilize the encoder $E_{VQ}$ trained from stage I to encode $x$ and", + "bbox": [ + 496, + 825, + 893, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "26092", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\hat{x}$ into latent space by $\\hat{l}^p = E_{VQ}(\\hat{x})$ and $l^p = E_{VQ}(x)$ , so the code-prediction-based pixel loss can be calculated by $\\mathcal{D}_{code}(l^p,\\hat{l}^p)$ in the same formulation with Equation 5. Here we use $E_{VQ}$ since it can map the input data to a compatible latent space with the codebook $C$ for code prediction. The overall pixel supervision is defined as:", + "bbox": [ + 76, + 89, + 472, + 183 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {D} _ {\\text {S t a g e I I I}} = | | x - \\hat {x} | | + \\mathcal {L} _ {p e r} (x, \\hat {x}) \\\\ + \\lambda_ {a d v} \\cdot \\mathcal {L} _ {a d v} (x, \\hat {x}) + \\lambda_ {c o d e} \\cdot \\mathcal {D} _ {c o d e} \\left(l ^ {p}, \\hat {l} ^ {p}\\right) \\tag {7} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 190, + 467, + 229 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where we set $\\lambda_{code} = 0.05$ by default. The rate-distortion trade-off supervision is:", + "bbox": [ + 76, + 238, + 468, + 268 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {S t a g e I I I}} = \\mathbf {E} _ {x \\sim p _ {X}} \\left[ \\mathcal {R} (\\hat {y}) + \\lambda \\cdot \\mathcal {D} _ {\\text {S t a g e I I I}} \\right] \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 277, + 468, + 296 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Discussion of Code-Prediction-Based Loss", + "text_level": 1, + "bbox": [ + 76, + 303, + 434, + 318 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Code prediction transformers [27, 28, 60] have demonstrated their effectiveness in high-quality image reconstruction. They typically input the predicted latent directly into the decoder for reconstruction. Different from these methods, in GLC, we suggest to consider code prediction solely as an auxiliary supervision during training, but not used in the inference process of the compression pipeline.", + "bbox": [ + 75, + 325, + 468, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This design is based on an observation: if a code prediction module is introduced before the decoder, the fine-tuning process in stage III cannot enhance compression performance further. It appears that the codebook becomes a performance bottleneck, restricting the decoder to receiving only the vector-quantized latent as input, which has already been well-trained in stage I. In GLC, by utilizing code prediction solely as auxiliary supervision during training, we eliminate this bottleneck, allowing the decoder to receive more flexible input for additional fine-tuning. We find that this code-prediction-based supervision effectively enhances the semantic consistency of the reconstructions, as shown in Figure 7. By necessitating the latent to possess the capability to predict the code index, the latent contains more semantic information, such as gestures and attributes.", + "bbox": [ + 75, + 431, + 470, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 671, + 209, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 696, + 294, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training details. We train GLC for both natural image compression and facial image compression. For natural images, we conduct stage I on ImageNet training set [14], stage II and III on OpenImage test set [33], using randomly cropped $256 \\times 256$ patches. For facial images, GLC is trained on FFHQ dataset [31] for all stages with a resolution of $512 \\times 512$ . Both models are optimized by AdamW [42] with a batch size of 8. For each batch, we train the model with different $\\lambda$ to achieve rate-variable compression.", + "bbox": [ + 75, + 719, + 468, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation dataset. We evaluate GLC on CLIC 2020 test set [51] with original resolution for natural image compression, and evaluate on CelebAHQ [30] with a resolution of", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/00d8dc75728ecd5a4dd2019b3c3782075ae5c1ad855f8e025c6ee047839a158b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 508, + 98, + 583, + 156 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f326b6496b3a638e84c51a36711dc49e29f4175d70904798e09cf73328ffe87d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 586, + 98, + 679, + 156 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b15f4e3adcc6340552b7de0bbadba090e319570c0343c0c514b089b4f12acdc9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 684, + 98, + 779, + 156 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e0ee7f99d6570edc677c69dadb82f476e243090f014f0a737286c329e1f5818c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 785, + 98, + 880, + 156 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/40998eeda359794e5baf91901d8a42342f6fc60c723d22a53ecccb9640f2e2c4.jpg", + "image_caption": [ + "Original Image" + ], + "image_footnote": [], + "bbox": [ + 508, + 161, + 581, + 219 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0f182d6dbb3af5bdeff9923461cd2d699eb95fb21062e1e7cfc1bed5b261c68d.jpg", + "image_caption": [ + "w/o code-pred-based loss" + ], + "image_footnote": [], + "bbox": [ + 584, + 161, + 679, + 219 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7229844ecc1d6331c262e46852d43a9bbe6dbb292fd7dd3ba4e3466f0ceb9de9.jpg", + "image_caption": [ + "w/ code-pred-based loss" + ], + "image_footnote": [], + "bbox": [ + 684, + 161, + 779, + 219 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5e12b06237e39622632a63022ecc04fa662a1ee2990caddf564d374b35cc4a26.jpg", + "image_caption": [ + "Figure 7. An example of using code-prediction-based supervision after stage II. Code prediction loss enhances the semantic consistency of the compressed latent, such as the color of eye shadow and the opening of the mouth." + ], + "image_footnote": [], + "bbox": [ + 785, + 161, + 880, + 219 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$512 \\times 512$ for facial image. We also show the results on Kodak [32], DIV2K [1] and MS-COCO 30K [40] in the supplementary material.", + "bbox": [ + 496, + 337, + 890, + 381 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics. We measure bit-stream size by bits per pixel (bpp), and measure visual quality by reference perceptual metrics LPIPS [29] and DISTS [15] and no-reference perceptual metrics FID [23] and KID [6]. We also provide PSNR and MS-SSIM [56] results in the supplementary material for completeness. Nevertheless, it is worth note that these pixel-level distortion metrics PSNR, MS-SSIM and LPIPS have strong limitations when evaluating image compression at ultra-low bitrate, which is also mentioned in other works [15, 35]. We provide a clearer demonstration of it in the supplementary material.", + "bbox": [ + 496, + 382, + 892, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baseline methods. We compare with traditional codec VVC [54], neural codec TCM [41], EVC [20], and generative codec FCC [26], Text+Sketch [35], HiFiC [44], MSILLM [46]. As some methods do not release models for ultra-low bitrate, we either retrain or fine-tune their models to suit such low bitrate. Text+Sketch is not evaluated on CLIC since it does not support compression in high resolution. In addition, we also compare with recent works HFD [24] and PerCo [8] in the supplementary material. For facial compression, we fine-tune EVC, TCM, HiFiC and MSILLM using FFHQ dataset for a fair comparison.", + "bbox": [ + 496, + 547, + 890, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Main Results", + "text_level": 1, + "bbox": [ + 500, + 724, + 637, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 8 shows the performance of the proposal and compared methods at ultra-low bitrate. On CLIC 2020, GLC demonstrates superiority in terms of DISTS, FID and KID than other methods. Specifically, GLC saves about $45\\%$ bits compared to previous SOTA method MS-ILLM while maintaining an equivalent FID. When comparing the pixel-level metric LPIPS, GLC also achieves comparable performance with high-fidelity generative codec such as HiFiC and MS-ILLM. On CelebAHQ, GLC outperforms all other methods across all metrics by a large margin.", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "26093", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6dc3f1d89dc3d53e894365ea39211e3f98302e0e20bab5ad21895a9677a918dd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 197, + 93, + 781, + 119 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f7078527f2c23414802e3e37ab8349a8747802b61e3123a6f806040ce85bbced.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 84, + 128, + 279, + 277 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/403dea6682aea885d23c00031f4f1bb80a92592acace05378a3cd688554aca01.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 128, + 480, + 279 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/80672e66c6094bdfa9f4a44624598d1dc510a0050c4bc129ae9849a8d1c6afa8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 128, + 684, + 279 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c8a5fafb393863e956b79750a9026a773aa0a4131b936da7a97fa256b99c4958.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 686, + 128, + 888, + 279 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8c80faa331ad8711fbaed3614179b17c2c042137be2acfc8351cd5091af22d6b.jpg", + "image_caption": [ + "Figure 8. Comparison of methods on CLIC 2020 test set and CelebAHQ." + ], + "image_footnote": [], + "bbox": [ + 84, + 281, + 272, + 430 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ae20a0313821ff26b38593ad0d02e89f4c5a0d7b7deed04d5daf2d4ad3a6c31a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 281, + 475, + 430 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e80b271631dc5deb6d58ce9f702be0191ddf8330fe45ca7bf64b58494f86682e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 281, + 679, + 430 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/577b515a6b710eda02570e14559ef65a960210e3cab3d0fb6384a201fbd8b4d7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 281, + 880, + 430 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b37e9dfeebaa98abf9ab3dc3c1c60019c490ede60ff70276c8716ee6cbd26de1.jpg", + "image_caption": [ + "Original" + ], + "image_footnote": [], + "bbox": [ + 89, + 463, + 220, + 561 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/da3c2f2a90f317114dd0f66db38db904d79ecf5bf865f2ab23ea0c83e1415466.jpg", + "image_caption": [ + "[Kodak]", + "Original", + "[CelebAHQ]", + "Figure 9. Qualitative examples of different methods on Kodak and CelebAHQ. More comparisons are in supplementary materials." + ], + "image_footnote": [], + "bbox": [ + 89, + 584, + 220, + 681 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3630dd87ce80b6f61463bac3f632d77d788ea541db3329aaefccb9de0dfef2d3.jpg", + "image_caption": [ + "Text+Sketch" + ], + "image_footnote": [], + "bbox": [ + 222, + 464, + 352, + 561 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0399d03b2df8f80592f27b944e28fab3433bdc46373cb409c4414e6febeb597c.jpg", + "image_caption": [ + "0.0320 bpp", + "Wc", + "0.0115bpp" + ], + "image_footnote": [], + "bbox": [ + 222, + 584, + 352, + 681 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ef1f42cbf9449f47bda56a2527d4b43d21518871d5d715761d0fd4dca403364c.jpg", + "image_caption": [ + "TCM" + ], + "image_footnote": [], + "bbox": [ + 352, + 464, + 482, + 561 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/aa91f8e1cd8218b10f2d83d40672a7e7cd4d543c6fdfd73c3ef7aec3190c5c69.jpg", + "image_caption": [ + "0.0787 bpp", + "TCM", + "0.0094 bpp" + ], + "image_footnote": [], + "bbox": [ + 352, + 583, + 482, + 681 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e5c92f3959b1da52eac29b5ad6773ab19e61782fd90e7d1a5f1a69d78738c18f.jpg", + "image_caption": [ + "HiFiC" + ], + "image_footnote": [], + "bbox": [ + 485, + 464, + 614, + 561 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3b518627e47e28da9b24fa31891946af601b5acf1d0e198d78b0182385b51843.jpg", + "image_caption": [ + "0.0426 bpp", + "HiFiC", + "0.0085 bpp" + ], + "image_footnote": [], + "bbox": [ + 486, + 583, + 614, + 681 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ac1626058217735b60a17c04045e006124a996d73fb066643cb95dfa6a8125af.jpg", + "image_caption": [ + "MS-ILLM" + ], + "image_footnote": [], + "bbox": [ + 617, + 464, + 746, + 561 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7fea7b78e033cb5524ff7be0cf7209f15d4deb05b698bd0359b6afcd0233cf3a.jpg", + "image_caption": [ + "0.0375 bpp", + "MS-ILLM", + "0.0088 bpp" + ], + "image_footnote": [], + "bbox": [ + 617, + 583, + 746, + 681 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9d1ad4d9804d1eb22ce311ef547f926a08263eed0ee6a5962a8a4ec8c59f7702.jpg", + "image_caption": [ + "GLC" + ], + "image_footnote": [], + "bbox": [ + 750, + 464, + 880, + 561 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/611305432135b369f03cbe1954580b906c7afb1bd89a7af3521880ba727acf1c.jpg", + "image_caption": [ + "0.0300 bpp", + "GLC", + "0.0080 bpp" + ], + "image_footnote": [], + "bbox": [ + 750, + 584, + 880, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 9 shows the qualitative comparison results. For natural image compression, Text+Sketch generates low-fidelity reconstructions, while TCM, HiFiC, and MS-ILLM produce blurry results at ultra low bitrate. In contrast, GLC achieves high-fidelity and high-realism results. In the case of facial image compression, we find existing methods cannot produce satisfactory reconstruction due to the severely distorted information at even more extreme bitrate limitation (e.g., 0.01 bpp or lower). TCM and HiFiC fall short in generating highly realistic results. Even though MS-ILLM", + "bbox": [ + 75, + 748, + 472, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "provides clearer details, it struggles to preserve correct facial attribute. Compared to them, GLC excels in both realism and fidelity at even lower bitrate.", + "bbox": [ + 496, + 748, + 890, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 801, + 653, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we conduct ablation studies to demonstrate the effectiveness of each proposed component. To provide a clearer comparison, we evaluate the BD-Rate [7] on the FID-BPP curve on the CLIC 2020 test set.", + "bbox": [ + 496, + 825, + 890, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Transform coding. A straightforward approach to com", + "bbox": [ + 500, + 886, + 890, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "26094", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9d9f943e15242e2150b5a0df00560180e099bcd671766c2061e0e5dfe1637580.jpg", + "table_caption": [ + "Table 1. Ablation study on latent-space compression." + ], + "table_footnote": [], + "table_body": "
Latent coding schemeProbability model of zBD-Rate ↓
Indices-map coding-66.2%
Transform codingFactorized prior17.7%
Categorical prior0%
", + "bbox": [ + 89, + 116, + 450, + 186 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/354d576cf5c2a39926f4b8e331c25b05776b714df2382868b213489d263ab2c6.jpg", + "table_caption": [ + "Table 2. Ablation study on the code prediction module." + ], + "table_footnote": [], + "table_body": "
code prediction usageBD-Rate ↓
w/o code pred.13.1%
code pred. in network60.7%
code pred. as supervision0%
", + "bbox": [ + 147, + 229, + 392, + 295 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "press the VQ-VAE latents is indices-map coding [27, 28, 43]. However, it causes $66.2\\%$ performance loss compared with transform coding, as shown in Table 1. It shows the effectiveness of transform coding on reducing redundancy.", + "bbox": [ + 75, + 321, + 467, + 382 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Categorical hyper module. In Section 3.3, we illustrate the superiority of employing a categorical prior for $z$ compared to the commonly used factorized prior. Table 1 further provides a quantitative comparison, demonstrating a significant improvement of $17.7\\%$ with such design.", + "bbox": [ + 75, + 382, + 467, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Code-prediction-based supervision. In Section 4, we suggest employing the code prediction module as an auxiliary loss during training, instead of during the inference process of the model pipeline as in [27, 28, 60]. As shown in Table 2, incorporating the code prediction module directly into the network leads to a $60.7\\%$ performance drop. We further remove the code-prediction-based supervision for comparison, and results show that adopting the code-prediction-based supervision brings a $13.1\\%$ improvement.", + "bbox": [ + 75, + 457, + 467, + 592 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Applications", + "text_level": 1, + "bbox": [ + 76, + 614, + 207, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Leveraging the potent latent space, our GLC pipeline opens avenues for exploring various vision applications. In this paper, we implement image restoration and style transfer as examples to show its potential. As depicted in Figure 10, for image restoration, we train a restoration encoder to map distorted images into clean latents, allowing users to directly compress a noisy image and decompress a clean one without additional cost. We compare our restoration application with a straightforward scheme of cascading an additional restoration network [57] with the GLC codec. The results in Table 3 indicate superior performance for our restoration application without the need for extra model parameters. Similarly, users can directly decode the latent into another style through a stylization decoder to achieve style transfer. We hope such versatility of the generative latent space will foster connections between image compression and other vision tasks in future research.", + "bbox": [ + 75, + 643, + 468, + 898 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0ce959ac2b89800cefdb94c256a4db5c43ac631d62c91fcdb1a4633fcee24b2e.jpg", + "image_caption": [ + "Figure 10. Generative latent applications built on GLC. In the practical image compression system, users can choose different encoders and decoders, to compress an image (medium), compress a distorted image and decompress a clear one (top), or decode the image into another style (bottom)." + ], + "image_footnote": [], + "bbox": [ + 504, + 90, + 890, + 369 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0244d4bd419e38c8b0144199c3fc32e35b03527b9a9961a08c8ccb5a70cd4047.jpg", + "table_caption": [ + "Table 3. Comparison for different joint restoration and compression schemes on CLIC 2020 test set." + ], + "table_footnote": [], + "table_body": "
SchemeBPP ↓FID ↓DISTS ↓Parameters
Restormer [57] + GLC Codec0.031410.790.117425M + 109M
GLC Restoration Application0.02998.620.1081109M
", + "bbox": [ + 503, + 515, + 888, + 563 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion and Limitation", + "text_level": 1, + "bbox": [ + 500, + 594, + 748, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce a generative latent coding (GLC) scheme to achieve high-fidelity and high-realism generative compression at ultra-low bitrate. Unlike most existing pixel-domain codecs, GLC performs transform coding on the latent domain of a generative VQ-VAE. By incorporating a categorical hyper module and a code-prediction-based supervision, GLC demonstrates state-of-the-art performance on several benchmarks. We further develop several vision applications on the GLC pipeline to demonstrate its practical potential.", + "bbox": [ + 496, + 622, + 890, + 772 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "However, as a generative image codec trained on specified datasets, the generalization capability of GLC is not always satisfactory. For instance, GLC cannot guarantee a clear and accurate reconstruction of screen contents, as illustrated in the supplementary material. Future work will focus on addressing this limitation, to enhance the generalization ability of GLC by improving the model structure and the training strategy.", + "bbox": [ + 496, + 773, + 890, + 887 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "26095", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 126-135, 2017. 6", + "[2] Eirikur Agustsson, Michael Tschannen, Fabian Mentzer, Radu Timofte, and Luc Van Gool. Generative adversarial networks for extreme learned image compression. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 221-231, 2019. 3", + "[3] Eirikur Agustsson, David Minnen, George Toderici, and Fabian Mentzer. Multi-realism image compression with a conditional generator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22324-22333, 2023. 3", + "[4] Johannes Balle, Valero Laparra, and Eero P Simoncelli. End-to-end optimized image compression. In 5th International Conference on Learning Representations, ICLR 2017, 2017. 1, 3", + "[5] Johannes Balle, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. In International Conference on Learning Representations, 2018. 1, 2, 3, 4", + "[6] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. arXiv preprint arXiv:1801.01401, 2018. 6", + "[7] Gisle Bjontegaard. Calculation of average psnr differences between rd-curves. ITU SG16 Doc. VCEG-M33, 2001. 7", + "[8] Marlene Careil, Matthew J Muckley, Jakob Verbeek, and Stéphane Lathuilière. Towards image compression with perfect realism at ultra-low bitrates. In The Twelfth International Conference on Learning Representations, 2023. 6", + "[9] Jianhui Chang, Zhenghui Zhao, Chuanmin Jia, Shiqi Wang, Lingbo Yang, Qi Mao, Jian Zhang, and Siwei Ma. Conceptual compression via deep structure and texture synthesis. IEEE Transactions on Image Processing, 31:2809-2823, 2022. 3", + "[10] Xi Chen, Diederik P Kingma, Tim Salimans, Yan Duan, Prafulla Dhariwal, John Schulman, Ilya Sutskever, and Pieter Abbeel. Variational lossy autoencoder. arXiv preprint arXiv:1611.02731, 2016. 3", + "[11] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned image compression with discretized gaussian mixture likelihoods and attention modules. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7939-7948, 2020. 1, 3", + "[12] Thomas M Cover. Elements of information theory. John Wiley & Sons, 1999. 3", + "[13] Ze Cui, Jing Wang, Shangyin Gao, Tiansheng Guo, Yihui Feng, and Bo Bai. Asymmetric gained deep image compression with continuous rate adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10532-10541, 2021. 3", + "[14] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. IEEE transactions on pattern analysis and machine intelligence, 44(5):2567-2581, 2020. 1, 6", + "[16] Alaaeldin El-Nouby, Matthew J Muckley, Karen Ullrich, Ivan Laptev, Jakob Verbeek, and Herve Jegou. Image compression with product quantized masked image modeling. Transactions on Machine Learning Research, 2022. 3", + "[17] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 1, 2, 3, 4, 5", + "[18] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1, 5", + "[19] Vivek K Goyal. Theoretical foundations of transform coding. IEEE Signal Processing Magazine, 18(5):9-21, 2001. 1, 3", + "[20] Wang Guo-Hua, Jiahao Li, Bin Li, and Yan Lu. Evc: Towards real-time neural image compression with mask decay. In The Eleventh International Conference on Learning Representations, 2022. 1, 3, 6", + "[21] Dailan He, Yaoyan Zheng, Baocheng Sun, Yan Wang, and Hongwei Qin. Checkerboard context model for efficient learned image compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14771-14780, 2021. 4", + "[22] Dailan He, Ziming Yang, Hongjiu Yu, Tongda Xu, Jixiang Luo, Yuan Chen, Chenjian Gao, Xinjie Shi, Hongwei Qin, and Yan Wang. Po-elic: Perception-oriented efficient learned image coding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1764–1769, 2022. 3", + "[23] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6", + "[24] Emiel Hoogeboom, Eirikur Agustsson, Fabian Mentzer, Luca Versari, George Toderici, and Lucas Theis. High-fidelity image compression with score-based generative models. arXiv preprint arXiv:2305.18231, 2023. 3, 6", + "[25] Yueyu Hu, Shuai Yang, Wenhan Yang, Ling-Yu Duan, and Jiaying Liu. Towards coding for human and machine vision: A scalable image coding approach. In 2020 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2020. 3", + "[26] Shoma Iwai, Tomo Miyazaki, Yoshihiro Sugaya, and Shinichiro Omachi. Fidelity-controllable extreme image compression with generative adversarial networks. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 8235-8242. IEEE, 2021. 3, 6", + "[27] Wei Jiang, Hyomin Choi, and Fabien Racapé. Adaptive human-centric video compression for humans and machines. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1121-1129, 2023. 2, 3, 4, 6, 8" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "26096", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Wei Jiang, Hyomin Choi, Fabien Racapé, Simon Feltman, and Fatih Kamisli. Face restoration-based scalable quality coding for video conferencing. In 2023 IEEE International Conference on Multimedia and Expo Workshops (ICMEW), pages 206-211. IEEE, 2023. 2, 3, 4, 6, 8", + "[29] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016. 1, 5, 6", + "[30] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In International Conference on Learning Representations, 2018. 2, 6", + "[31] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 6", + "[32] Kodak Lossless True Color Image Suite. http://r0k.us/graphics/kodak/.6", + "[33] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision, 128(7):1956-1981, 2020. 6", + "[34] Jooyoung Lee, Seunghyun Cho, and Seung-Kwon Beck. Context-adaptive entropy model for end-to-end optimized image compression. In International Conference on Learning Representations, 2018. 3", + "[35] Eric Lei, Yigit Berkay Uslu, Hamed Hassani, and Shirin Saeedi Bidokhti. Text+ sketch: Image compression at ultra low rates. In ICML 2023 Workshop Neural Compression: From Information Theory to Applications, 2023. 2, 3, 6", + "[36] Jiahao Li, Bin Li, and Yan Lu. Deep contextual video compression. Advances in Neural Information Processing Systems, 34:18114-18125, 2021. 5", + "[37] Jiahao Li, Bin Li, and Yan Lu. Hybrid spatial-temporal entropy modelling for neural video compression. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1503-1511, 2022.", + "[38] Jiahao Li, Bin Li, and Yan Lu. Neural video compression with diverse contexts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22616-22626, 2023. 3, 4", + "[39] Jiahao Li, Bin Li, and Yan Lu. Neural video compression with feature modulation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 17-21, 2024, 2024. 5", + "[40] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Jinming Liu, Heming Sun, and Jiro Katto. Learned image compression with mixed transformer-cnn architectures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14388-14397, 2023. 1, 3, 6", + "[42] Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam. 2018. 6", + "[43] Qi Mao, Tinghan Yang, Yinuo Zhang, Shuyin Pan, Meng Wang, Shiqi Wang, and Siwei Ma. Extreme image compression using fine-tuned vqgan models. arXiv preprint arXiv:2307.08265, 2023. 2, 4, 8", + "[44] Fabian Mentzer, George D Toderici, Michael Tschannen, and Eirikur Agustsson. High-fidelity generative image compression. Advances in Neural Information Processing Systems, 33:11913-11924, 2020. 2, 3, 6", + "[45] David Minnen, Johannes Balle, and George D Toderici. Joint autoregressive and hierarchical priors for learned image compression. Advances in neural information processing systems, 31, 2018. 1, 3", + "[46] Matthew J. Muckley, Alaaeldin El-Nouby, Karen Ullrich, Hervé Jégou, and Jakob Verbeek. Improving statistical fidelity for neural image compression with implicit local likelihood models. In International Conference on Machine Learning, 2023. 1, 2, 3, 6", + "[47] Linfeng Qi, Jiahao Li, Bin Li, Houqiang Li, and Yan Lu. Motion information propagation for neural video compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6111-6120, 2023. 5", + "[48] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3", + "[49] Xihua Sheng, Jiahao Li, Bin Li, Li Li, Dong Liu, and Yan Lu. Temporal context mining for learned video compression. IEEE Transactions on Multimedia, 2022. 5", + "[50] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 5", + "[51] George Toderici, Lucas Theis, Nick Johnston, Eirikur Agustsson, Fabian Mentzer, Johannes Balle, Wenzhe Shi, and Radu Timofte. Clic 2020: Challenge on learned image compression, 2020, 2020. 2, 6", + "[52] Aäron Van Den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks. In International conference on machine learning, pages 1747-1756. PMLR, 2016. 3", + "[53] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 1, 2, 3, 5", + "[54] VVC-21.2. https://vcgit.hhi.fraunhofer.de/jvet/VVCSoftware_VTM/~/tree/VTM-21.2. Accessed: 10/23/2023, 2023. 1, 6", + "[55] Gregory K Wallace. TheJPEG still picture compression standard.Communications of the ACM, 34(4):30-44, 1991.1", + "[56] Zhou Wang, Eero P Simoncelli, and Alan C Bovik. Multiscale structural similarity for image quality assessment. In" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "26097", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "The Thrity-Seventh Asilomar Conference on Signals, Systems & Computers, 2003, pages 1398-1402. IEEE, 2003. 6", + "[57] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728-5739, 2022. 8", + "[58] Jing Zhao, Bin Li, Jiahao Li, Ruiqin Xiong, and Yan Lu. A universal encoder rate distortion optimization framework for learned compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1880-1884, 2021. 3", + "[59] Jing Zhao, Bin Li, Jiahao Li, Ruiqin Xiong, and Yan Lu. A universal optimization framework for learning-based image codec. ACM Transactions on Multimedia Computing, Communications and Applications, 20(1):1-19, 2023. 3", + "[60] Shangchen Zhou, Kelvin Chan, Chongyi Li, and Chen Change Loy. Towards robust blind face restoration with codebook lookup transformer. Advances in Neural Information Processing Systems, 35:30599-30611, 2022. 2, 3, 6, 8" + ], + "bbox": [ + 78, + 90, + 468, + 400 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "26098", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Generative Latent Coding for Ultra-Low Bitrate Image Compression/a71ce275-c18c-4862-b210-a735f3bb6944_model.json b/2024/Generative Latent Coding for Ultra-Low Bitrate Image Compression/a71ce275-c18c-4862-b210-a735f3bb6944_model.json new file mode 100644 index 0000000000000000000000000000000000000000..968d9cc0ad171866500d9a76f6328d431a19256e --- /dev/null +++ b/2024/Generative Latent Coding for Ultra-Low Bitrate Image Compression/a71ce275-c18c-4862-b210-a735f3bb6944_model.json @@ -0,0 +1,3049 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.138, + 0.131, + 0.833, + 0.154 + ], + "angle": 0, + "content": "Generative Latent Coding for Ultra-Low Bitrate Image Compression" + }, + { + "type": "text", + "bbox": [ + 0.183, + 0.18, + 0.783, + 0.2 + ], + "angle": 0, + "content": "Zhaoyang Jia\\(^{1*}\\) Jiahao Li\\(^{2}\\) Bin Li\\(^{2}\\) Houqiang Li\\(^{1}\\) Yan Lu\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.182, + 0.2, + 0.79, + 0.217 + ], + "angle": 0, + "content": "1 University of Science and Technology of China 2 Microsoft Research Asia" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.22, + 0.851, + 0.234 + ], + "angle": 0, + "content": "jzy_ustc@mail.ustc.edu.cn, lihq@ustc.edu.cn, {li.jiahao, libin, yanlu}@microsoft.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.302, + 0.474, + 0.651 + ], + "angle": 0, + "content": "Most existing image compression approaches perform transform coding in the pixel space to reduce its spatial redundancy. However, they encounter difficulties in achieving both high-realism and high-fidelity at low bitrate, as the pixel-space distortion may not align with human perception. To address this issue, we introduce a Generative Latent Coding (GLC) architecture, which performs transform coding in the latent space of a generative vector-quantized variational auto-encoder (VQ-VAE), instead of in the pixel space. The generative latent space is characterized by greater sparsity, richer semantic and better alignment with human perception, rendering it advantageous for achieving high-realism and high-fidelity compression. Additionally, we introduce a categorical hyper module to reduce the bit cost of hyper-information, and a code-prediction-based supervision to enhance the semantic consistency. Experiments demonstrate that our GLC maintains high visual quality with less than 0.04 bpp on natural images and less than 0.01 bpp on facial images. On the CLIC2020 test set, we achieve the same FID as MS-ILLM with \\(45\\%\\) fewer bits. Furthermore, the powerful generative latent space enables various applications built on our GLC pipeline, such as image restoration and style transfer." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.67, + 0.21, + 0.685 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.696, + 0.47, + 0.848 + ], + "angle": 0, + "content": "Amid the ongoing surge of digital visual data, the importance of achieving high-efficiency image compression becomes increasingly paramount. From the traditional compression standards [54, 55] to the emerging learned image compression models [4, 5, 11, 20, 41, 45], most compression algorithms follow the pixel-space transform coding [4, 19] paradigm. Specifically, they convert pixels into compact representations through a transform module, which eliminates the redundancy to reduce the bit cost in the subsequent entropy coding process." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.849, + 0.47, + 0.864 + ], + "angle": 0, + "content": "However, we observe a common inherent limitation in" + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.269, + 0.887, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.499, + 0.895, + 0.598 + ], + "angle": 0, + "content": "Figure 1. Generative latent space of VQ-VAE exhibits better alignment with human perception than pixel space for ultra-low bitrate compression. Under comparable distortion levels (measured by signal-to-noise ratio, SNR), latent-space compression produces reconstructions with superior perceptual quality (measured by DISTS [15]) than pixel-space generative codec MS-ILLM [46], as the compressed latents remain in the same latent code space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.895, + 0.765 + ], + "angle": 0, + "content": "these methods: the pixel-space distortion is not always consistent with the human perception, especially at low bitrate. In practice, human observers prioritize the semantic consistency and the texture realism of an image, but this information is difficult to be adequately exploited solely by a pixel-space transform module. As shown in the left of Figure 1, pixel-space generative image codec MS-ILLM [46] struggles to guarantee visual quality at low bitrate, even after it incorporates perceptual supervision [29] and adversarial supervision [18] within the pixel space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.896, + 0.903 + ], + "angle": 0, + "content": "Based on such observation, a natural problem arises: how can we compress images in a way that aligns with human perception? To address this challenge, we introduce a Generative Latent Coding (GLC) paradigm. In GLC, we first encode images into a generative latent space that aligns with human perception, and subsequently perform transform coding in this latent space, instead of in pixel space. To concretize this concept, we adopt a generative vector-quantized variational auto-encoder (VQ-VAE) [17, 53] to" + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.875, + 0.47, + 0.9 + ], + "angle": 0, + "content": "*This work was done when Zhaoyang Jia was an intern at Microsoft Research Asia." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26088" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.09, + 0.887, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.398, + 0.895, + 0.442 + ], + "angle": 0, + "content": "Figure 2. A qualitative comparison between HiFiC [44], MS-ILLM [46], Text+Sketch [35] and the proposed GLC. GLC produces images with high visual quality, even in regions with complex texture. In contrast, HiFiC and MS-ILLM exhibit noticeable artifacts, and Text+Sketch generates results that deviate significantly from the input. Best viewed when zoomed in." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.46, + 0.473, + 0.626 + ], + "angle": 0, + "content": "produce the latent space, which offers three significant advantages: 1) The discrete codes of VQ-VAE encapsulate semantic visual components [53], allowing GLC to focus on compressing the semantic content to guarantee fidelity. 2) Generative VQ-VAE exhibits remarkable generative capabilities [17] for high-realism texture reconstruction. 3) The discrete variational bottleneck naturally brings a low-entropy and distortion-robust latent space for compression. Thanks to such characteristics, GLC is more aligned with human perception to achieve enhanced visual quality, as demonstrated in the right of Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.473, + 0.903 + ], + "angle": 0, + "content": "When implementing GLC, two crucial questions persist: How to effectively compress the generative latents? And how to supervise the generative latent coding? A straightforward approach to compress VQ-VAE latents is indices-map coding [27, 28, 43], but it is limited by the ineffective redundancy reduction between indices and the lack of rate-variable coding support. In this paper, we propose a novel generative-latent-space transform coding approach, where an effective rate-variable structure is adopted to reduce latent redundancy for higher compression ratio. In addition, a categorical hyper module is introduced to model the distribution of \\( z \\) with a discrete codebook, which significantly reduces the bitrate of \\( z \\) when compared to the factorized hyper module [5]. As for the supervision of GLC, inspired by recent code prediction transformers [27, 28, 60], a code-prediction-based supervision is proposed. It serves as an auxiliary supervision employed solely in the training process to greatly enhance the semantic consistency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.46, + 0.892, + 0.625 + ], + "angle": 0, + "content": "Benefited from these advanced designs, our GLC achieves excellent performance on both natural and facial images. In the CLIC 2020 test set [51], GLC attains a bitrate less than 0.04 bpp while delivering high visual quality. It obtains \\(45\\%\\) bit savings compared to MS-ILLM at the same FID. In the CelebAHQ [30] dataset, GLC achieves an even lower bitrate of less than 0.01 bpp. As shown in Figure 2, compared with recent advanced generative image compression approaches MS-ILLM [46] and Text+Sketch [35], GLC produces more visually appealing compression results with a lower bit cost." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.626, + 0.892, + 0.746 + ], + "angle": 0, + "content": "Furthermore, leveraging the representative generative latent space, GLC supports various vision applications such as image restoration and style transfer. By replacing the compression encoder with a restoration encoder, the proposed restoration application surpasses the performance of cascading a restoration model and a neural codec. We hope such versatility of generative latent space will help connect image compression with other vision tasks in the future." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.748, + 0.792, + 0.761 + ], + "angle": 0, + "content": "In summary, our main contributions are :" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.765, + 0.892, + 0.825 + ], + "angle": 0, + "content": "- We present a generative latent coding (GLC) scheme, which performs transform coding in the generative latent space of a VQ-VAE to achieve high-fidelity and high-realism compression at ultra-low bitrate." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.826, + 0.892, + 0.886 + ], + "angle": 0, + "content": "- We introduce a categorical hyper module to significantly reduce the bit cost of hyper information. Additionally, a code-prediction-based supervision is adopted to enhance the perceptual quality." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.9 + ], + "angle": 0, + "content": "- GLC obtains \\(45\\%\\) bit reduction on CLIC2020 with the" + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.765, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26089" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.098, + 0.387, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.098, + 0.52, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.524, + 0.099, + 0.668, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.099, + 0.885, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.302, + 0.895, + 0.386 + ], + "angle": 0, + "content": "Figure 3. Illustration of the proposed Generative Latent Coding (GLC) framework. (Left) GLC firstly encodes the image into a generative latent representation (Section 3.2), then compresses the latent with transform coding (Section 3.3), and finally reconstructs image from the compressed latent. (Right) We progressively train GLC in three stages (Section 4): In stage I, we train a generative VQ-VAE to obtain a human-perception-aligned latent space. In stage II, the transform coding module learns to compress the latent with a code-prediction-based latent supervision (Figure 6). Finally, in stage III, the entire network is fine-tuned jointly with a code-prediction-based pixel supervision to further enhance the compression performance." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.402, + 0.471, + 0.434 + ], + "angle": 0, + "content": "same FID as the previous SOTA. Furthermore, GLC enables various application within its latent space." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.445, + 0.228, + 0.461 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.47, + 0.336, + 0.487 + ], + "angle": 0, + "content": "2.1. Learned Image Compression" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.493, + 0.47, + 0.659 + ], + "angle": 0, + "content": "Lossy image compression is grounded on Shannon's rate-distortion theory [12]. Balle et al. [4] first proposed to utilize neural networks for pixel-space transform coding [19], employing analysis and synthesis transform modules to convert images into compact representations for entropy coding. Subsequently, some researches make strides in improving the probability model [4, 5, 11, 34, 38, 45] for more accurate estimation, while others explore the network structure [11, 41], optimization algorithm [58, 59] and rate-variable coding [13, 20] for improved compression performance and practicality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.471, + 0.84 + ], + "angle": 0, + "content": "A recently raised critical challenge in image compression is how to improve the perceptual quality of the reconstruction. Agustsson et al. [2] first introduced the concept of generative compression, which compresses essential image features and generate distorted details using GAN. Some subsequent works [9, 25] extract image sketches and latent codes to ensure geometry-consistency. Text+Sketch [35] utilizes a conditional-diffusion model to generate image based on image captions and sketches, achieving superior perceptual quality. While these schemes produce visual appealing results, they often deviate significantly from the input and cannot guarantee the semantic consistency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "To achieve generative compression with high-fidelity, Mentzer et al. [44] further studied the network structure and generative adversarial loss to enable high-fidelity compression. Subsequent researches further enhance the trans" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.402, + 0.895, + 0.493 + ], + "angle": 0, + "content": "form coding [16, 22], generative post-processing [24] or focus on controlling the trade-off between fidelity and realism [3, 26]. Recently, MS-ILLM [46] introduces a noisy discriminator which is conditioned on quantized local image representations to greatly enhance the statistical fidelity." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.51, + 0.716, + 0.527 + ], + "angle": 0, + "content": "2.2. Latent Space Modeling" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.536, + 0.894, + 0.734 + ], + "angle": 0, + "content": "Latent space image modeling means modeling the distribution of image within the latent space of a neural network. This technique has been primarily developed for image generation. Chen et al. [10] and Oord et al. [53] introduced PixelCNN [52] in the latent space of VAE and VQ-VAE for image generation. Esser et al. [17] took it a step further by incorporating transformers into VQ-VAE latent space for high-quality generation. More recently, Rombach et al. [48] employed diffusion models to model the latent space of VQ-VAE, achieving remarkable results in high-resolution image generation. These studies underscore the potential of image processing within the generative latent space, particularly in the latent space of VQ-VAE." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Recently, the concept of latent space modeling has been extended to other tasks. CodeFormer [60] introduces a code prediction transformer that takes distorted latents as input and predicts the high-quality VQ-VAE index for facial restoration. Building upon this, Jiang et al. [27, 28] proposed transmitting the predicted indices to achieve restoration-based facial conferencing. In this paper, we explore the characteristics of latent space modeling in the realm of generative image compression. We specially design a transform coding paradigm in the latent space, which demonstrates superior effectiveness." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "26090" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.086, + 0.396, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.238, + 0.217, + 0.255, + 0.228 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.131, + 0.548, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.476, + 0.217, + 0.493, + 0.228 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.566, + 0.092, + 0.716, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.625, + 0.217, + 0.642, + 0.228 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.725, + 0.092, + 0.874, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.782, + 0.217, + 0.8, + 0.228 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.241, + 0.894, + 0.311 + ], + "angle": 0, + "content": "Figure 4. Illustration of the transform coding in latent space. (a) The model structure of the transform coding module. We further compare it with other coding schemes in operational diagrams : (b) indices-map coding [27, 28, 43], (c) transform coding with factorized hyper module [5, 21, 38] and (d) proposed transform coding with categorical hyper module. Here AE and AD denote arithmetic encoding and decoding, VQ-E and VQ-D stand for VQ-indices-map encoding and decoding, Q refers to scalar quantization, U signifies the addition of uniform noise as a differential simulation of Q, and S denotes the spatial context entropy module." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.335, + 0.168, + 0.35 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.36, + 0.188, + 0.374 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.384, + 0.471, + 0.58 + ], + "angle": 0, + "content": "In this section, we introduce the details of the proposed Generative Latent Coding (GLC) architecture. To achieve high-perceptual-quality compression, GLC encodes image into a perception-aligned latent space through a generative latent auto-encoder, and perform transform coding on the latent representations for lower bitrate. As depicted in the left of Figure 3, the input image \\( x \\) is firstly encoded into the latent \\( l \\) using the latent encoder \\( E \\). Then \\( l \\) undergoes an analysis transform \\( g_{a} \\) to produce the code \\( y \\), which is further scalar-quantized to \\( \\hat{y} \\) for entropy coding. Following that, a synthesis transform \\( g_{s} \\) is employed to transform \\( \\hat{y} \\) back to \\( l \\), and finally, the reconstruction \\( \\hat{x} \\) is generated by the latent decoder \\( D \\). This entire process is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.201, + 0.587, + 0.469, + 0.623 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} l = E (x), y = g _ {a} (l) \\\\ \\hat {y} = Q (y) \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.201, + 0.626, + 0.344, + 0.644 + ], + "angle": 0, + "content": "\\[\n\\hat {l} = g _ {s} (\\hat {y}), \\hat {x} = D (\\hat {l})\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.651, + 0.367, + 0.665 + ], + "angle": 0, + "content": "3.2. Generative Latent Auto-Encoder" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.47, + 0.839 + ], + "angle": 0, + "content": "To achieve high-quality generative latent coding, how to obtain a human-perception-aligned latent space is a crucial challenge. In GLC, we address it by employing the generative VQ-VAE [17] as the latent auto-encoder \\((E\\) and \\(D)\\). By mapping images into visual semantic elements within a codebook \\(C\\) and incorporating a generative image decoding process, both semantic consistency and texture realism can be well guaranteed. Additionally, it contributes to the compression process through a sparse yet robust latent space, which is achieved by training with the discrete codebook \\(C\\) as a variational bottleneck." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.382, + 0.864 + ], + "angle": 0, + "content": "3.3. Transform Coding in Latent Space" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "To compress the latent representations \\(l\\), a direct approach is VQ-indices-map coding [27, 28, 43] (Figure 4b). How-" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.393, + 0.614, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.53, + 0.477, + 0.598, + 0.488 + ], + "angle": 0, + "content": "Original Image" + }, + { + "type": "image", + "bbox": [ + 0.629, + 0.343, + 0.738, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.627, + 0.425, + 0.74, + 0.434 + ], + "angle": 0, + "content": "\\(y = 0\\) bits \\(z = 1000\\) bits" + }, + { + "type": "image", + "bbox": [ + 0.762, + 0.343, + 0.869, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.766, + 0.425, + 0.874, + 0.434 + ], + "angle": 0, + "content": "\\(y = 0\\) bits \\(z = 160\\) bits" + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.439, + 0.738, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.626, + 0.521, + 0.75, + 0.53 + ], + "angle": 0, + "content": "\\(y = 1800\\) bits \\(z = 1000\\) bits" + }, + { + "type": "image_caption", + "bbox": [ + 0.659, + 0.535, + 0.716, + 0.545 + ], + "angle": 0, + "content": "Factorized z" + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.439, + 0.869, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.762, + 0.521, + 0.879, + 0.53 + ], + "angle": 0, + "content": "\\(y = 1640\\) bits \\(z = 160\\) bits" + }, + { + "type": "image_caption", + "bbox": [ + 0.787, + 0.535, + 0.847, + 0.545 + ], + "angle": 0, + "content": "Categorical z" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.561, + 0.892, + 0.603 + ], + "angle": 0, + "content": "Figure 5. Example of comparison between factorized and categorical hyper modules. The proposed categorical \\( z \\) encodes essential semantic and structural information with much less bits." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.892, + 0.797 + ], + "angle": 0, + "content": "ever, these methods often lack a careful consideration of the correlation among the latents, resulting in an insufficient redundancy reduction and consequently a high bit cost. In GLC, we introduce a transform coding module to compress the latent, replacing the vector-quantization step for more effective reduction of latent redundancy. As shown in Figure 4a, the latents are transformed into code \\( y \\) using transformations \\( g_{a} \\), \\( g_{s} \\) and then quantized to \\( \\hat{y} \\). Entropy coding is applied to \\( \\hat{y} \\) based on a probability \\( p(\\hat{y}) \\), which is estimated by a categorical hyper module (Section 3.3.1) and a quadtree-partition-based spatial context module [38]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.816, + 0.742, + 0.832 + ], + "angle": 0, + "content": "3.3.1 Categorical Hyper Module" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Factorized hyper module [5] (Figure 4c) is commonly employed in recent image compression schemes. However, at ultra-low bitrate, we notice that the factorized \\( z \\) tends to encode low-level information such as color and texture, incur" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26091" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.198 + ], + "angle": 0, + "content": "ring a high bit cost, as illustrated in Figure 5. To address it, we propose a categorical hyper module (Figure 4d), which utilizes a hyper codebook to store the basic semantic elements rather than the low-level information. This module comprises a hyper analysis transform \\( h_a \\), a hyper synthesis transform \\( h_s \\) and a hyper codebook \\( C_h \\). The transformations are formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.205, + 0.47, + 0.222 + ], + "angle": 0, + "content": "\\[\nz = h _ {a} (y), \\hat {z} = V Q (z, C _ {h}), p r i o r _ {z} = h _ {s} (\\hat {z}) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.228, + 0.47, + 0.304 + ], + "angle": 0, + "content": "where \\( z \\) and \\( \\hat{z} \\) denote hyper-codes. \\( VQ(\\cdot, C_h) \\) represents vector-quantization by nearest lookup in \\( C_h \\). As shown in the right of Figure 5, the categorical \\( z \\) is more inclined to capture high-level semantic information, which can be encoded with significantly fewer bits." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.322, + 0.345, + 0.336 + ], + "angle": 0, + "content": "3.3.2 Rate-Variable Transformation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.345, + 0.473, + 0.513 + ], + "angle": 0, + "content": "A notable advantage of transform coding over VQ-indices-map coding is its capability for rate-variable compression, which is a core functionality for a practical image codec. Indices-map coding is limited since the codebook can only model one specified distribution, but different rates naturally need different distributions. In contrast, transform coding converts latent into a unified Gaussian distribution, and variable-rate can be achieved by variable parameters (e.g., means and scales) of Gaussian. In GLC, we follow DCVC series [36-39, 47, 49] to incorporate a scalar \\( q \\) in the transform coding to achieve rate-variable compression." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.524, + 0.277, + 0.541 + ], + "angle": 0, + "content": "4. Progressive Training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.549, + 0.47, + 0.671 + ], + "angle": 0, + "content": "As depicted in the right of Figure 3, we adopt a three-stage progressive training manner to fully leverage the potential of GLC. We initially learn a human-perception-aligned latent space to guarantee the perceptual quality, subsequently learn to perform transform coding on this latent space to achieve low bitrate, and finally fine-tune the entire network for superior compression performance. At each stage, distinct loss functions are adopted to guide different modules." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.677, + 0.367, + 0.694 + ], + "angle": 0, + "content": "4.1. Stage I: Auto-Encoder Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.701, + 0.47, + 0.809 + ], + "angle": 0, + "content": "To obtain a human-perception-aligned latent space for compression, we begin with training a generative VQ-VAE as the initialization of \\( E \\) and \\( D \\). To ensure the sparsity of the latent space, an auxiliary codebook \\( C \\) is employed to perform nearest vector-quantization, transforming \\( l \\) to \\( \\widetilde{l} \\). The supervision comprises reconstruction loss, perceptual loss [29], adversarial loss [18] and codebook loss [53]:" + }, + { + "type": "equation", + "bbox": [ + 0.139, + 0.814, + 0.47, + 0.849 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {S t a g e I}} = \\left| \\left| x - \\hat {x} \\right| \\right| + \\mathcal {L} _ {p e r} (x, \\hat {x}) \\tag {3} \\\\ + \\lambda_ {a d v} \\cdot \\mathcal {L} _ {a d v} (x, \\hat {x}) + \\mathcal {L} _ {\\text {c o d e b o o k}} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Here, \\(\\mathcal{L}_{per}\\) corresponds to the LPIPS loss calculated using VGG [50] extracted features. \\(\\mathcal{L}_{adv}\\) is the adaptive PatchGAN adversarial loss [17] with a weight of \\(\\lambda_{adv} = 0.8\\)." + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.093, + 0.699, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.093, + 0.885, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.52, + 0.298, + 0.871, + 0.313 + ], + "angle": 0, + "content": "Figure 6. Illustration of code-prediction-based supervision." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.339, + 0.738, + 0.353 + ], + "angle": 0, + "content": "The codebook loss is formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.363, + 0.892, + 0.382 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o d e b o o k}} = \\left| \\left| \\operatorname {s g} (l) - \\tilde {l} \\right| \\right| + \\beta \\cdot \\left| \\left| \\operatorname {s g} (\\tilde {l}) - l \\right| \\right| \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.393, + 0.893, + 0.424 + ], + "angle": 0, + "content": "where \\(\\mathrm{sg}(\\cdot)\\) denotes the stop-gradient operator and \\(\\beta = 0.25\\) controls the update rates of the \\(E\\) and \\(C\\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.432, + 0.831, + 0.45 + ], + "angle": 0, + "content": "4.2. Stage II: Transform Coding Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.456, + 0.892, + 0.593 + ], + "angle": 0, + "content": "Given the trained latent space, we further learn to perform transform coding to achieve low-bitrate latent compression, while fixing the auto-encoder \\(E\\) and \\(D\\). We introduce an auxiliary code predictor \\(CP\\) to enhance the semantic consistency by necessitating the latent to possess the capability to predict the correct VQ-indices. As shown in Figure 6, we encode \\(l\\) into VQ-indices by \\(M_l = VQ(l,C)\\) and subsequently predict these indices by \\(\\hat{M}_{\\hat{l}} = CP(\\hat{l})\\). So the code-prediction-based loss can be formulated by" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.603, + 0.892, + 0.623 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} _ {c o d e} (l, \\hat {l}) = \\alpha \\cdot C E (M _ {l}, \\hat {M _ {l}}) + | | l - \\hat {l} | | _ {2} ^ {2} \\qquad (5)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.633, + 0.893, + 0.678 + ], + "angle": 0, + "content": "where \\( CE \\) denotes the cross entropy loss and we set \\( \\alpha = 0.5 \\) by default. Then the transform coding module can be supervised by the rate-distortion trade-off" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.689, + 0.892, + 0.708 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {S t a g e I I}} = \\mathbf {E} _ {x \\sim p _ {X}} [ \\mathcal {R} (\\hat {y}) + \\lambda \\cdot \\mathcal {D} _ {\\text {c o d e}} (l, \\hat {l}) ] \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.717, + 0.892, + 0.793 + ], + "angle": 0, + "content": "where \\(\\mathcal{R}\\) is the estimated rate and \\(\\lambda\\) is used to control the trade-off. Note that a codebook loss (as formulated in Equation 4) is required to train the hyper codebook \\(C_h\\) in the categorical hyper module. We omit it from the loss functions of both stage II and stage III for the sake of conciseness." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.731, + 0.818 + ], + "angle": 0, + "content": "4.3. Stage III: Joint Training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Finally, we fine-tune the entire network with the pixel space supervision to achieve better compression performance. As shown in Figure 6, we extend the code-prediction-based latent supervision into the pixel space. Specifically, we utilize the encoder \\( E_{VQ} \\) trained from stage I to encode \\( x \\) and" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26092" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.09, + 0.473, + 0.184 + ], + "angle": 0, + "content": "\\(\\hat{x}\\) into latent space by \\(\\hat{l}^p = E_{VQ}(\\hat{x})\\) and \\(l^p = E_{VQ}(x)\\), so the code-prediction-based pixel loss can be calculated by \\(\\mathcal{D}_{code}(l^p,\\hat{l}^p)\\) in the same formulation with Equation 5. Here we use \\(E_{VQ}\\) since it can map the input data to a compatible latent space with the codebook \\(C\\) for code prediction. The overall pixel supervision is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.191, + 0.468, + 0.23 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {D} _ {\\text {S t a g e I I I}} = | | x - \\hat {x} | | + \\mathcal {L} _ {p e r} (x, \\hat {x}) \\\\ + \\lambda_ {a d v} \\cdot \\mathcal {L} _ {a d v} (x, \\hat {x}) + \\lambda_ {c o d e} \\cdot \\mathcal {D} _ {c o d e} \\left(l ^ {p}, \\hat {l} ^ {p}\\right) \\tag {7} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.239, + 0.47, + 0.269 + ], + "angle": 0, + "content": "where we set \\(\\lambda_{code} = 0.05\\) by default. The rate-distortion trade-off supervision is:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.279, + 0.47, + 0.297 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {S t a g e I I I}} = \\mathbf {E} _ {x \\sim p _ {X}} \\left[ \\mathcal {R} (\\hat {y}) + \\lambda \\cdot \\mathcal {D} _ {\\text {S t a g e I I I}} \\right] \\tag {8}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.304, + 0.436, + 0.319 + ], + "angle": 0, + "content": "4.4. Discussion of Code-Prediction-Based Loss" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.327, + 0.47, + 0.433 + ], + "angle": 0, + "content": "Code prediction transformers [27, 28, 60] have demonstrated their effectiveness in high-quality image reconstruction. They typically input the predicted latent directly into the decoder for reconstruction. Different from these methods, in GLC, we suggest to consider code prediction solely as an auxiliary supervision during training, but not used in the inference process of the compression pipeline." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.433, + 0.471, + 0.659 + ], + "angle": 0, + "content": "This design is based on an observation: if a code prediction module is introduced before the decoder, the fine-tuning process in stage III cannot enhance compression performance further. It appears that the codebook becomes a performance bottleneck, restricting the decoder to receiving only the vector-quantized latent as input, which has already been well-trained in stage I. In GLC, by utilizing code prediction solely as auxiliary supervision during training, we eliminate this bottleneck, allowing the decoder to receive more flexible input for additional fine-tuning. We find that this code-prediction-based supervision effectively enhances the semantic consistency of the reconstructions, as shown in Figure 7. By necessitating the latent to possess the capability to predict the code index, the latent contains more semantic information, such as gestures and attributes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.672, + 0.21, + 0.688 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.697, + 0.295, + 0.712 + ], + "angle": 0, + "content": "5.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.469, + 0.855 + ], + "angle": 0, + "content": "Training details. We train GLC for both natural image compression and facial image compression. For natural images, we conduct stage I on ImageNet training set [14], stage II and III on OpenImage test set [33], using randomly cropped \\(256 \\times 256\\) patches. For facial images, GLC is trained on FFHQ dataset [31] for all stages with a resolution of \\(512 \\times 512\\). Both models are optimized by AdamW [42] with a batch size of 8. For each batch, we train the model with different \\(\\lambda\\) to achieve rate-variable compression." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Evaluation dataset. We evaluate GLC on CLIC 2020 test set [51] with original resolution for natural image compression, and evaluate on CelebAHQ [30] with a resolution of" + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.099, + 0.584, + 0.157 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.099, + 0.681, + 0.157 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.685, + 0.099, + 0.78, + 0.157 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.786, + 0.099, + 0.882, + 0.157 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.162, + 0.583, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.586, + 0.162, + 0.68, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.685, + 0.162, + 0.78, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.786, + 0.162, + 0.882, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.595, + 0.226, + 0.671, + 0.24 + ], + "angle": 0, + "content": "Original Image" + }, + { + "type": "image_caption", + "bbox": [ + 0.692, + 0.225, + 0.773, + 0.247 + ], + "angle": 0, + "content": "w/o code-pred-based loss" + }, + { + "type": "image_caption", + "bbox": [ + 0.795, + 0.225, + 0.871, + 0.247 + ], + "angle": 0, + "content": "w/ code-pred-based loss" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.261, + 0.892, + 0.317 + ], + "angle": 0, + "content": "Figure 7. An example of using code-prediction-based supervision after stage II. Code prediction loss enhances the semantic consistency of the compressed latent, such as the color of eye shadow and the opening of the mouth." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.338, + 0.892, + 0.382 + ], + "angle": 0, + "content": "\\(512 \\times 512\\) for facial image. We also show the results on Kodak [32], DIV2K [1] and MS-COCO 30K [40] in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.383, + 0.893, + 0.549 + ], + "angle": 0, + "content": "Evaluation metrics. We measure bit-stream size by bits per pixel (bpp), and measure visual quality by reference perceptual metrics LPIPS [29] and DISTS [15] and no-reference perceptual metrics FID [23] and KID [6]. We also provide PSNR and MS-SSIM [56] results in the supplementary material for completeness. Nevertheless, it is worth note that these pixel-level distortion metrics PSNR, MS-SSIM and LPIPS have strong limitations when evaluating image compression at ultra-low bitrate, which is also mentioned in other works [15, 35]. We provide a clearer demonstration of it in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.549, + 0.892, + 0.716 + ], + "angle": 0, + "content": "Baseline methods. We compare with traditional codec VVC [54], neural codec TCM [41], EVC [20], and generative codec FCC [26], Text+Sketch [35], HiFiC [44], MSILLM [46]. As some methods do not release models for ultra-low bitrate, we either retrain or fine-tune their models to suit such low bitrate. Text+Sketch is not evaluated on CLIC since it does not support compression in high resolution. In addition, we also compare with recent works HFD [24] and PerCo [8] in the supplementary material. For facial compression, we fine-tune EVC, TCM, HiFiC and MSILLM using FFHQ dataset for a fair comparison." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.726, + 0.638, + 0.74 + ], + "angle": 0, + "content": "5.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Figure 8 shows the performance of the proposal and compared methods at ultra-low bitrate. On CLIC 2020, GLC demonstrates superiority in terms of DISTS, FID and KID than other methods. Specifically, GLC saves about \\(45\\%\\) bits compared to previous SOTA method MS-ILLM while maintaining an equivalent FID. When comparing the pixel-level metric LPIPS, GLC also achieves comparable performance with high-fidelity generative codec such as HiFiC and MS-ILLM. On CelebAHQ, GLC outperforms all other methods across all metrics by a large margin." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26093" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.199, + 0.094, + 0.782, + 0.121 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.129, + 0.28, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.129, + 0.482, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.129, + 0.685, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.129, + 0.889, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.282, + 0.273, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.282, + 0.477, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.282, + 0.68, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.282, + 0.882, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.267, + 0.437, + 0.702, + 0.452 + ], + "angle": 0, + "content": "Figure 8. Comparison of methods on CLIC 2020 test set and CelebAHQ." + }, + { + "type": "image", + "bbox": [ + 0.09, + 0.464, + 0.222, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.565, + 0.175, + 0.574 + ], + "angle": 0, + "content": "Original" + }, + { + "type": "image_caption", + "bbox": [ + 0.137, + 0.574, + 0.174, + 0.585 + ], + "angle": 0, + "content": "[Kodak]" + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.585, + 0.221, + 0.683 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.685, + 0.175, + 0.693 + ], + "angle": 0, + "content": "Original" + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.694, + 0.183, + 0.705 + ], + "angle": 0, + "content": "[CelebAHQ]" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.465, + 0.353, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.259, + 0.565, + 0.315, + 0.574 + ], + "angle": 0, + "content": "Text+Sketch" + }, + { + "type": "image_caption", + "bbox": [ + 0.26, + 0.574, + 0.314, + 0.585 + ], + "angle": 0, + "content": "0.0320 bpp" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.585, + 0.353, + 0.683 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.277, + 0.685, + 0.302, + 0.694 + ], + "angle": 0, + "content": "Wc" + }, + { + "type": "image_caption", + "bbox": [ + 0.26, + 0.694, + 0.315, + 0.705 + ], + "angle": 0, + "content": "0.0115bpp" + }, + { + "type": "image", + "bbox": [ + 0.354, + 0.465, + 0.483, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.408, + 0.565, + 0.432, + 0.574 + ], + "angle": 0, + "content": "TCM" + }, + { + "type": "image_caption", + "bbox": [ + 0.392, + 0.574, + 0.446, + 0.584 + ], + "angle": 0, + "content": "0.0787 bpp" + }, + { + "type": "image", + "bbox": [ + 0.354, + 0.584, + 0.483, + 0.683 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.408, + 0.685, + 0.431, + 0.694 + ], + "angle": 0, + "content": "TCM" + }, + { + "type": "image_caption", + "bbox": [ + 0.392, + 0.694, + 0.446, + 0.704 + ], + "angle": 0, + "content": "0.0094 bpp" + }, + { + "type": "image", + "bbox": [ + 0.486, + 0.465, + 0.616, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.565, + 0.564, + 0.574 + ], + "angle": 0, + "content": "HiFiC" + }, + { + "type": "image_caption", + "bbox": [ + 0.524, + 0.574, + 0.578, + 0.584 + ], + "angle": 0, + "content": "0.0426 bpp" + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.584, + 0.616, + 0.683 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.685, + 0.564, + 0.693 + ], + "angle": 0, + "content": "HiFiC" + }, + { + "type": "image_caption", + "bbox": [ + 0.524, + 0.694, + 0.578, + 0.704 + ], + "angle": 0, + "content": "0.0085 bpp" + }, + { + "type": "image", + "bbox": [ + 0.618, + 0.465, + 0.748, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.661, + 0.565, + 0.705, + 0.574 + ], + "angle": 0, + "content": "MS-ILLM" + }, + { + "type": "image_caption", + "bbox": [ + 0.658, + 0.574, + 0.71, + 0.584 + ], + "angle": 0, + "content": "0.0375 bpp" + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.584, + 0.748, + 0.683 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.662, + 0.685, + 0.706, + 0.693 + ], + "angle": 0, + "content": "MS-ILLM" + }, + { + "type": "image_caption", + "bbox": [ + 0.656, + 0.694, + 0.711, + 0.704 + ], + "angle": 0, + "content": "0.0088 bpp" + }, + { + "type": "image", + "bbox": [ + 0.75, + 0.465, + 0.881, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.806, + 0.565, + 0.826, + 0.574 + ], + "angle": 0, + "content": "GLC" + }, + { + "type": "image_caption", + "bbox": [ + 0.788, + 0.575, + 0.842, + 0.585 + ], + "angle": 0, + "content": "0.0300 bpp" + }, + { + "type": "image", + "bbox": [ + 0.751, + 0.585, + 0.881, + 0.683 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.805, + 0.685, + 0.826, + 0.693 + ], + "angle": 0, + "content": "GLC" + }, + { + "type": "image_caption", + "bbox": [ + 0.788, + 0.694, + 0.842, + 0.704 + ], + "angle": 0, + "content": "0.0080 bpp" + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.715, + 0.868, + 0.73 + ], + "angle": 0, + "content": "Figure 9. Qualitative examples of different methods on Kodak and CelebAHQ. More comparisons are in supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.749, + 0.473, + 0.901 + ], + "angle": 0, + "content": "Figure 9 shows the qualitative comparison results. For natural image compression, Text+Sketch generates low-fidelity reconstructions, while TCM, HiFiC, and MS-ILLM produce blurry results at ultra low bitrate. In contrast, GLC achieves high-fidelity and high-realism results. In the case of facial image compression, we find existing methods cannot produce satisfactory reconstruction due to the severely distorted information at even more extreme bitrate limitation (e.g., 0.01 bpp or lower). TCM and HiFiC fall short in generating highly realistic results. Even though MS-ILLM" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.749, + 0.892, + 0.794 + ], + "angle": 0, + "content": "provides clearer details, it struggles to preserve correct facial attribute. Compared to them, GLC excels in both realism and fidelity at even lower bitrate." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.654, + 0.818 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.885 + ], + "angle": 0, + "content": "In this section, we conduct ablation studies to demonstrate the effectiveness of each proposed component. To provide a clearer comparison, we evaluate the BD-Rate [7] on the FID-BPP curve on the CLIC 2020 test set." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.887, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Transform coding. A straightforward approach to com" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "26094" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.115, + 0.09, + 0.432, + 0.104 + ], + "angle": 0, + "content": "Table 1. Ablation study on latent-space compression." + }, + { + "type": "table", + "bbox": [ + 0.09, + 0.117, + 0.451, + 0.188 + ], + "angle": 0, + "content": "
Latent coding schemeProbability model of zBD-Rate ↓
Indices-map coding-66.2%
Transform codingFactorized prior17.7%
Categorical prior0%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.109, + 0.202, + 0.437, + 0.217 + ], + "angle": 0, + "content": "Table 2. Ablation study on the code prediction module." + }, + { + "type": "table", + "bbox": [ + 0.148, + 0.23, + 0.393, + 0.296 + ], + "angle": 0, + "content": "
code prediction usageBD-Rate ↓
w/o code pred.13.1%
code pred. in network60.7%
code pred. as supervision0%
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.322, + 0.468, + 0.383 + ], + "angle": 0, + "content": "press the VQ-VAE latents is indices-map coding [27, 28, 43]. However, it causes \\(66.2\\%\\) performance loss compared with transform coding, as shown in Table 1. It shows the effectiveness of transform coding on reducing redundancy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.383, + 0.468, + 0.457 + ], + "angle": 0, + "content": "Categorical hyper module. In Section 3.3, we illustrate the superiority of employing a categorical prior for \\( z \\) compared to the commonly used factorized prior. Table 1 further provides a quantitative comparison, demonstrating a significant improvement of \\( 17.7\\% \\) with such design." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.458, + 0.468, + 0.593 + ], + "angle": 0, + "content": "Code-prediction-based supervision. In Section 4, we suggest employing the code prediction module as an auxiliary loss during training, instead of during the inference process of the model pipeline as in [27, 28, 60]. As shown in Table 2, incorporating the code prediction module directly into the network leads to a \\(60.7\\%\\) performance drop. We further remove the code-prediction-based supervision for comparison, and results show that adopting the code-prediction-based supervision brings a \\(13.1\\%\\) improvement." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.616, + 0.208, + 0.634 + ], + "angle": 0, + "content": "6. Applications" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.47, + 0.9 + ], + "angle": 0, + "content": "Leveraging the potent latent space, our GLC pipeline opens avenues for exploring various vision applications. In this paper, we implement image restoration and style transfer as examples to show its potential. As depicted in Figure 10, for image restoration, we train a restoration encoder to map distorted images into clean latents, allowing users to directly compress a noisy image and decompress a clean one without additional cost. We compare our restoration application with a straightforward scheme of cascading an additional restoration network [57] with the GLC codec. The results in Table 3 indicate superior performance for our restoration application without the need for extra model parameters. Similarly, users can directly decode the latent into another style through a stylization decoder to achieve style transfer. We hope such versatility of the generative latent space will foster connections between image compression and other vision tasks in future research." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.092, + 0.891, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.386, + 0.892, + 0.456 + ], + "angle": 0, + "content": "Figure 10. Generative latent applications built on GLC. In the practical image compression system, users can choose different encoders and decoders, to compress an image (medium), compress a distorted image and decompress a clear one (top), or decode the image into another style (bottom)." + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.476, + 0.891, + 0.503 + ], + "angle": 0, + "content": "Table 3. Comparison for different joint restoration and compression schemes on CLIC 2020 test set." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.516, + 0.89, + 0.564 + ], + "angle": 0, + "content": "
SchemeBPP ↓FID ↓DISTS ↓Parameters
Restormer [57] + GLC Codec0.031410.790.117425M + 109M
GLC Restoration Application0.02998.620.1081109M
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.595, + 0.749, + 0.611 + ], + "angle": 0, + "content": "7. Conclusion and Limitation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.623, + 0.892, + 0.773 + ], + "angle": 0, + "content": "In this paper, we introduce a generative latent coding (GLC) scheme to achieve high-fidelity and high-realism generative compression at ultra-low bitrate. Unlike most existing pixel-domain codecs, GLC performs transform coding on the latent domain of a generative VQ-VAE. By incorporating a categorical hyper module and a code-prediction-based supervision, GLC demonstrates state-of-the-art performance on several benchmarks. We further develop several vision applications on the GLC pipeline to demonstrate its practical potential." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.775, + 0.892, + 0.888 + ], + "angle": 0, + "content": "However, as a generative image codec trained on specified datasets, the generalization capability of GLC is not always satisfactory. For instance, GLC cannot guarantee a clear and accurate reconstruction of screen contents, as illustrated in the supplementary material. Future work will focus on addressing this limitation, to enhance the generalization ability of GLC by improving the model structure and the training strategy." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26095" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 126-135, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.472, + 0.242 + ], + "angle": 0, + "content": "[2] Eirikur Agustsson, Michael Tschannen, Fabian Mentzer, Radu Timofte, and Luc Van Gool. Generative adversarial networks for extreme learned image compression. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 221-231, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.471, + 0.311 + ], + "angle": 0, + "content": "[3] Eirikur Agustsson, David Minnen, George Toderici, and Fabian Mentzer. Multi-realism image compression with a conditional generator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22324-22333, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.312, + 0.471, + 0.367 + ], + "angle": 0, + "content": "[4] Johannes Balle, Valero Laparra, and Eero P Simoncelli. End-to-end optimized image compression. In 5th International Conference on Learning Representations, ICLR 2017, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.471, + 0.424 + ], + "angle": 0, + "content": "[5] Johannes Balle, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. In International Conference on Learning Representations, 2018. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.471, + 0.465 + ], + "angle": 0, + "content": "[6] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. arXiv preprint arXiv:1801.01401, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.466, + 0.471, + 0.494 + ], + "angle": 0, + "content": "[7] Gisle Bjontegaard. Calculation of average psnr differences between rd-curves. ITU SG16 Doc. VCEG-M33, 2001. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.495, + 0.471, + 0.55 + ], + "angle": 0, + "content": "[8] Marlene Careil, Matthew J Muckley, Jakob Verbeek, and Stéphane Lathuilière. Towards image compression with perfect realism at ultra-low bitrates. In The Twelfth International Conference on Learning Representations, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.551, + 0.471, + 0.619 + ], + "angle": 0, + "content": "[9] Jianhui Chang, Zhenghui Zhao, Chuanmin Jia, Shiqi Wang, Lingbo Yang, Qi Mao, Jian Zhang, and Siwei Ma. Conceptual compression via deep structure and texture synthesis. IEEE Transactions on Image Processing, 31:2809-2823, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.621, + 0.47, + 0.676 + ], + "angle": 0, + "content": "[10] Xi Chen, Diederik P Kingma, Tim Salimans, Yan Duan, Prafulla Dhariwal, John Schulman, Ilya Sutskever, and Pieter Abbeel. Variational lossy autoencoder. arXiv preprint arXiv:1611.02731, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.47, + 0.747 + ], + "angle": 0, + "content": "[11] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned image compression with discretized gaussian mixture likelihoods and attention modules. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7939-7948, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.47, + 0.774 + ], + "angle": 0, + "content": "[12] Thomas M Cover. Elements of information theory. John Wiley & Sons, 1999. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.471, + 0.845 + ], + "angle": 0, + "content": "[13] Ze Cui, Jing Wang, Shangyin Gao, Tiansheng Guo, Yihui Feng, and Bo Bai. Asymmetric gained deep image compression with continuous rate adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10532-10541, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[14] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[15] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. IEEE transactions on pattern analysis and machine intelligence, 44(5):2567-2581, 2020. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[16] Alaaeldin El-Nouby, Matthew J Muckley, Karen Ullrich, Ivan Laptev, Jakob Verbeek, and Herve Jegou. Image compression with product quantized masked image modeling. Transactions on Machine Learning Research, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.272 + ], + "angle": 0, + "content": "[17] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.274, + 0.892, + 0.33 + ], + "angle": 0, + "content": "[18] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.358 + ], + "angle": 0, + "content": "[19] Vivek K Goyal. Theoretical foundations of transform coding. IEEE Signal Processing Magazine, 18(5):9-21, 2001. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.359, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[20] Wang Guo-Hua, Jiahao Li, Bin Li, and Yan Lu. Evc: Towards real-time neural image compression with mask decay. In The Eleventh International Conference on Learning Representations, 2022. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.892, + 0.483 + ], + "angle": 0, + "content": "[21] Dailan He, Yaoyan Zheng, Baocheng Sun, Yan Wang, and Hongwei Qin. Checkerboard context model for efficient learned image compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14771-14780, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.567 + ], + "angle": 0, + "content": "[22] Dailan He, Ziming Yang, Hongjiu Yu, Tongda Xu, Jixiang Luo, Yuan Chen, Chenjian Gao, Xinjie Shi, Hongwei Qin, and Yan Wang. Po-elic: Perception-oriented efficient learned image coding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1764–1769, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.568, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[23] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.636, + 0.892, + 0.692 + ], + "angle": 0, + "content": "[24] Emiel Hoogeboom, Eirikur Agustsson, Fabian Mentzer, Luca Versari, George Toderici, and Lucas Theis. High-fidelity image compression with score-based generative models. arXiv preprint arXiv:2305.18231, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.892, + 0.762 + ], + "angle": 0, + "content": "[25] Yueyu Hu, Shuai Yang, Wenhan Yang, Ling-Yu Duan, and Jiaying Liu. Towards coding for human and machine vision: A scalable image coding approach. In 2020 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.762, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[26] Shoma Iwai, Tomo Miyazaki, Yoshihiro Sugaya, and Shinichiro Omachi. Fidelity-controllable extreme image compression with generative adversarial networks. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 8235-8242. IEEE, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[27] Wei Jiang, Hyomin Choi, and Fabien Racapé. Adaptive human-centric video compression for humans and machines. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1121-1129, 2023. 2, 3, 4, 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26096" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[28] Wei Jiang, Hyomin Choi, Fabien Racapé, Simon Feltman, and Fatih Kamisli. Face restoration-based scalable quality coding for video conferencing. In 2023 IEEE International Conference on Multimedia and Expo Workshops (ICMEW), pages 206-211. IEEE, 2023. 2, 3, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.471, + 0.244 + ], + "angle": 0, + "content": "[29] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.246, + 0.471, + 0.3 + ], + "angle": 0, + "content": "[30] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In International Conference on Learning Representations, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.301, + 0.471, + 0.357 + ], + "angle": 0, + "content": "[31] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.358, + 0.468, + 0.384 + ], + "angle": 0, + "content": "[32] Kodak Lossless True Color Image Suite. http://r0k.us/graphics/kodak/.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.385, + 0.471, + 0.481 + ], + "angle": 0, + "content": "[33] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision, 128(7):1956-1981, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.483, + 0.471, + 0.538 + ], + "angle": 0, + "content": "[34] Jooyoung Lee, Seunghyun Cho, and Seung-Kwon Beck. Context-adaptive entropy model for end-to-end optimized image compression. In International Conference on Learning Representations, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.471, + 0.607 + ], + "angle": 0, + "content": "[35] Eric Lei, Yigit Berkay Uslu, Hamed Hassani, and Shirin Saeedi Bidokhti. Text+ sketch: Image compression at ultra low rates. In ICML 2023 Workshop Neural Compression: From Information Theory to Applications, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.471, + 0.649 + ], + "angle": 0, + "content": "[36] Jiahao Li, Bin Li, and Yan Lu. Deep contextual video compression. Advances in Neural Information Processing Systems, 34:18114-18125, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.471, + 0.705 + ], + "angle": 0, + "content": "[37] Jiahao Li, Bin Li, and Yan Lu. Hybrid spatial-temporal entropy modelling for neural video compression. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1503-1511, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.471, + 0.761 + ], + "angle": 0, + "content": "[38] Jiahao Li, Bin Li, and Yan Lu. Neural video compression with diverse contexts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22616-22626, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.471, + 0.816 + ], + "angle": 0, + "content": "[39] Jiahao Li, Bin Li, and Yan Lu. Neural video compression with feature modulation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 17-21, 2024, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[40] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[41] Jinming Liu, Heming Sun, and Jiro Katto. Learned image compression with mixed transformer-cnn architectures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14388-14397, 2023. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[42] Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam. 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[43] Qi Mao, Tinghan Yang, Yinuo Zhang, Shuyin Pan, Meng Wang, Shiqi Wang, and Siwei Ma. Extreme image compression using fine-tuned vqgan models. arXiv preprint arXiv:2307.08265, 2023. 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[44] Fabian Mentzer, George D Toderici, Michael Tschannen, and Eirikur Agustsson. High-fidelity generative image compression. Advances in Neural Information Processing Systems, 33:11913-11924, 2020. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[45] David Minnen, Johannes Balle, and George D Toderici. Joint autoregressive and hierarchical priors for learned image compression. Advances in neural information processing systems, 31, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[46] Matthew J. Muckley, Alaaeldin El-Nouby, Karen Ullrich, Hervé Jégou, and Jakob Verbeek. Improving statistical fidelity for neural image compression with implicit local likelihood models. In International Conference on Machine Learning, 2023. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.433, + 0.892, + 0.488 + ], + "angle": 0, + "content": "[47] Linfeng Qi, Jiahao Li, Bin Li, Houqiang Li, and Yan Lu. Motion information propagation for neural video compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6111-6120, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.49, + 0.892, + 0.559 + ], + "angle": 0, + "content": "[48] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.56, + 0.892, + 0.601 + ], + "angle": 0, + "content": "[49] Xihua Sheng, Jiahao Li, Bin Li, Li Li, Dong Liu, and Yan Lu. Temporal context mining for learned video compression. IEEE Transactions on Multimedia, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.602, + 0.892, + 0.644 + ], + "angle": 0, + "content": "[50] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.645, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[51] George Toderici, Lucas Theis, Nick Johnston, Eirikur Agustsson, Fabian Mentzer, Johannes Balle, Wenzhe Shi, and Radu Timofte. Clic 2020: Challenge on learned image compression, 2020, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.702, + 0.892, + 0.757 + ], + "angle": 0, + "content": "[52] Aäron Van Den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks. In International conference on machine learning, pages 1747-1756. PMLR, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.758, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[53] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.801, + 0.892, + 0.842 + ], + "angle": 0, + "content": "[54] VVC-21.2. https://vcgit.hhi.fraunhofer.de/jvet/VVCSoftware_VTM/~/tree/VTM-21.2. Accessed: 10/23/2023, 2023. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[55] Gregory K Wallace. TheJPEG still picture compression standard.Communications of the ACM, 34(4):30-44, 1991.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[56] Zhou Wang, Eero P Simoncelli, and Alan C Bovik. Multiscale structural similarity for image quality assessment. In" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26097" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "The Thrity-Seventh Asilomar Conference on Signals, Systems & Computers, 2003, pages 1398-1402. IEEE, 2003. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.203 + ], + "angle": 0, + "content": "[57] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728-5739, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.206, + 0.469, + 0.274 + ], + "angle": 0, + "content": "[58] Jing Zhao, Bin Li, Jiahao Li, Ruiqin Xiong, and Yan Lu. A universal encoder rate distortion optimization framework for learned compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1880-1884, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.277, + 0.469, + 0.331 + ], + "angle": 0, + "content": "[59] Jing Zhao, Bin Li, Jiahao Li, Ruiqin Xiong, and Yan Lu. A universal optimization framework for learning-based image codec. ACM Transactions on Multimedia Computing, Communications and Applications, 20(1):1-19, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.333, + 0.469, + 0.401 + ], + "angle": 0, + "content": "[60] Shangchen Zhou, Kelvin Chan, Chongyi Li, and Chen Change Loy. Towards robust blind face restoration with codebook lookup transformer. Advances in Neural Information Processing Systems, 35:30599-30611, 2022. 2, 3, 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "26098" + } + ] +] \ No newline at end of file diff --git a/2024/Generative Latent Coding for Ultra-Low Bitrate Image Compression/a71ce275-c18c-4862-b210-a735f3bb6944_origin.pdf b/2024/Generative Latent Coding for Ultra-Low Bitrate Image Compression/a71ce275-c18c-4862-b210-a735f3bb6944_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..842bba42975275e3acfdce4ef2bf14c7982bfa2c --- /dev/null +++ b/2024/Generative Latent Coding for Ultra-Low Bitrate Image Compression/a71ce275-c18c-4862-b210-a735f3bb6944_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b367e33b94917d899ef8886c4560ae94d6ecccea7f4c1b344f6ce549362b3be6 +size 3114958 diff --git a/2024/Generative Latent Coding for Ultra-Low Bitrate Image Compression/full.md b/2024/Generative Latent Coding for Ultra-Low Bitrate Image Compression/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5a1b8978d15e5e27bba3d90cb762683249b27f65 --- /dev/null +++ b/2024/Generative Latent Coding for Ultra-Low Bitrate Image Compression/full.md @@ -0,0 +1,421 @@ +# Generative Latent Coding for Ultra-Low Bitrate Image Compression + +Zhaoyang Jia $^{1*}$ Jiahao Li $^{2}$ Bin Li $^{2}$ Houqiang Li $^{1}$ Yan Lu $^{2}$ + +1 University of Science and Technology of China 2 Microsoft Research Asia + +jzy_ustc@mail.ustc.edu.cn, lihq@ustc.edu.cn, {li.jiahao, libin, yanlu}@microsoft.com + +# Abstract + +Most existing image compression approaches perform transform coding in the pixel space to reduce its spatial redundancy. However, they encounter difficulties in achieving both high-realism and high-fidelity at low bitrate, as the pixel-space distortion may not align with human perception. To address this issue, we introduce a Generative Latent Coding (GLC) architecture, which performs transform coding in the latent space of a generative vector-quantized variational auto-encoder (VQ-VAE), instead of in the pixel space. The generative latent space is characterized by greater sparsity, richer semantic and better alignment with human perception, rendering it advantageous for achieving high-realism and high-fidelity compression. Additionally, we introduce a categorical hyper module to reduce the bit cost of hyper-information, and a code-prediction-based supervision to enhance the semantic consistency. Experiments demonstrate that our GLC maintains high visual quality with less than 0.04 bpp on natural images and less than 0.01 bpp on facial images. On the CLIC2020 test set, we achieve the same FID as MS-ILLM with $45\%$ fewer bits. Furthermore, the powerful generative latent space enables various applications built on our GLC pipeline, such as image restoration and style transfer. + +# 1. Introduction + +Amid the ongoing surge of digital visual data, the importance of achieving high-efficiency image compression becomes increasingly paramount. From the traditional compression standards [54, 55] to the emerging learned image compression models [4, 5, 11, 20, 41, 45], most compression algorithms follow the pixel-space transform coding [4, 19] paradigm. Specifically, they convert pixels into compact representations through a transform module, which eliminates the redundancy to reduce the bit cost in the subsequent entropy coding process. + +However, we observe a common inherent limitation in + +![](images/a2ba8fbb87b5b0653a8d05ef617dd3e5b676dab018b51f1cf11a13168b793beb.jpg) +Figure 1. Generative latent space of VQ-VAE exhibits better alignment with human perception than pixel space for ultra-low bitrate compression. Under comparable distortion levels (measured by signal-to-noise ratio, SNR), latent-space compression produces reconstructions with superior perceptual quality (measured by DISTS [15]) than pixel-space generative codec MS-ILLM [46], as the compressed latents remain in the same latent code space. + +these methods: the pixel-space distortion is not always consistent with the human perception, especially at low bitrate. In practice, human observers prioritize the semantic consistency and the texture realism of an image, but this information is difficult to be adequately exploited solely by a pixel-space transform module. As shown in the left of Figure 1, pixel-space generative image codec MS-ILLM [46] struggles to guarantee visual quality at low bitrate, even after it incorporates perceptual supervision [29] and adversarial supervision [18] within the pixel space. + +Based on such observation, a natural problem arises: how can we compress images in a way that aligns with human perception? To address this challenge, we introduce a Generative Latent Coding (GLC) paradigm. In GLC, we first encode images into a generative latent space that aligns with human perception, and subsequently perform transform coding in this latent space, instead of in pixel space. To concretize this concept, we adopt a generative vector-quantized variational auto-encoder (VQ-VAE) [17, 53] to + +![](images/fa2f7ad2599bdd232dd9af00fa873e08404ca7c186b9f152ea0b146474fe767e.jpg) +Figure 2. A qualitative comparison between HiFiC [44], MS-ILLM [46], Text+Sketch [35] and the proposed GLC. GLC produces images with high visual quality, even in regions with complex texture. In contrast, HiFiC and MS-ILLM exhibit noticeable artifacts, and Text+Sketch generates results that deviate significantly from the input. Best viewed when zoomed in. + +produce the latent space, which offers three significant advantages: 1) The discrete codes of VQ-VAE encapsulate semantic visual components [53], allowing GLC to focus on compressing the semantic content to guarantee fidelity. 2) Generative VQ-VAE exhibits remarkable generative capabilities [17] for high-realism texture reconstruction. 3) The discrete variational bottleneck naturally brings a low-entropy and distortion-robust latent space for compression. Thanks to such characteristics, GLC is more aligned with human perception to achieve enhanced visual quality, as demonstrated in the right of Figure 1. + +When implementing GLC, two crucial questions persist: How to effectively compress the generative latents? And how to supervise the generative latent coding? A straightforward approach to compress VQ-VAE latents is indices-map coding [27, 28, 43], but it is limited by the ineffective redundancy reduction between indices and the lack of rate-variable coding support. In this paper, we propose a novel generative-latent-space transform coding approach, where an effective rate-variable structure is adopted to reduce latent redundancy for higher compression ratio. In addition, a categorical hyper module is introduced to model the distribution of $z$ with a discrete codebook, which significantly reduces the bitrate of $z$ when compared to the factorized hyper module [5]. As for the supervision of GLC, inspired by recent code prediction transformers [27, 28, 60], a code-prediction-based supervision is proposed. It serves as an auxiliary supervision employed solely in the training process to greatly enhance the semantic consistency. + +Benefited from these advanced designs, our GLC achieves excellent performance on both natural and facial images. In the CLIC 2020 test set [51], GLC attains a bitrate less than 0.04 bpp while delivering high visual quality. It obtains $45\%$ bit savings compared to MS-ILLM at the same FID. In the CelebAHQ [30] dataset, GLC achieves an even lower bitrate of less than 0.01 bpp. As shown in Figure 2, compared with recent advanced generative image compression approaches MS-ILLM [46] and Text+Sketch [35], GLC produces more visually appealing compression results with a lower bit cost. + +Furthermore, leveraging the representative generative latent space, GLC supports various vision applications such as image restoration and style transfer. By replacing the compression encoder with a restoration encoder, the proposed restoration application surpasses the performance of cascading a restoration model and a neural codec. We hope such versatility of generative latent space will help connect image compression with other vision tasks in the future. + +In summary, our main contributions are : + +- We present a generative latent coding (GLC) scheme, which performs transform coding in the generative latent space of a VQ-VAE to achieve high-fidelity and high-realism compression at ultra-low bitrate. +- We introduce a categorical hyper module to significantly reduce the bit cost of hyper information. Additionally, a code-prediction-based supervision is adopted to enhance the perceptual quality. +- GLC obtains $45\%$ bit reduction on CLIC2020 with the + +![](images/752991a6a6488088bb4007ad888aa69defed14166a7fc044e13ee45a4aeb716a.jpg) +Figure 3. Illustration of the proposed Generative Latent Coding (GLC) framework. (Left) GLC firstly encodes the image into a generative latent representation (Section 3.2), then compresses the latent with transform coding (Section 3.3), and finally reconstructs image from the compressed latent. (Right) We progressively train GLC in three stages (Section 4): In stage I, we train a generative VQ-VAE to obtain a human-perception-aligned latent space. In stage II, the transform coding module learns to compress the latent with a code-prediction-based latent supervision (Figure 6). Finally, in stage III, the entire network is fine-tuned jointly with a code-prediction-based pixel supervision to further enhance the compression performance. + +![](images/2f1e3acd8c9ce4c5ce81e9bd16fd4673e8ab73f694a8683bdf1aaad66607829a.jpg) + +![](images/e9a7896199c88e7736807b01977446915c7124d5a9e6e6b418b8185296c2d2f4.jpg) + +![](images/0a0612e6323cbe6ddcedcc5580590af275e60cfcf2e2398da48b3985f50e2232.jpg) + +same FID as the previous SOTA. Furthermore, GLC enables various application within its latent space. + +# 2. Related Works + +# 2.1. Learned Image Compression + +Lossy image compression is grounded on Shannon's rate-distortion theory [12]. Balle et al. [4] first proposed to utilize neural networks for pixel-space transform coding [19], employing analysis and synthesis transform modules to convert images into compact representations for entropy coding. Subsequently, some researches make strides in improving the probability model [4, 5, 11, 34, 38, 45] for more accurate estimation, while others explore the network structure [11, 41], optimization algorithm [58, 59] and rate-variable coding [13, 20] for improved compression performance and practicality. + +A recently raised critical challenge in image compression is how to improve the perceptual quality of the reconstruction. Agustsson et al. [2] first introduced the concept of generative compression, which compresses essential image features and generate distorted details using GAN. Some subsequent works [9, 25] extract image sketches and latent codes to ensure geometry-consistency. Text+Sketch [35] utilizes a conditional-diffusion model to generate image based on image captions and sketches, achieving superior perceptual quality. While these schemes produce visual appealing results, they often deviate significantly from the input and cannot guarantee the semantic consistency. + +To achieve generative compression with high-fidelity, Mentzer et al. [44] further studied the network structure and generative adversarial loss to enable high-fidelity compression. Subsequent researches further enhance the trans + +form coding [16, 22], generative post-processing [24] or focus on controlling the trade-off between fidelity and realism [3, 26]. Recently, MS-ILLM [46] introduces a noisy discriminator which is conditioned on quantized local image representations to greatly enhance the statistical fidelity. + +# 2.2. Latent Space Modeling + +Latent space image modeling means modeling the distribution of image within the latent space of a neural network. This technique has been primarily developed for image generation. Chen et al. [10] and Oord et al. [53] introduced PixelCNN [52] in the latent space of VAE and VQ-VAE for image generation. Esser et al. [17] took it a step further by incorporating transformers into VQ-VAE latent space for high-quality generation. More recently, Rombach et al. [48] employed diffusion models to model the latent space of VQ-VAE, achieving remarkable results in high-resolution image generation. These studies underscore the potential of image processing within the generative latent space, particularly in the latent space of VQ-VAE. + +Recently, the concept of latent space modeling has been extended to other tasks. CodeFormer [60] introduces a code prediction transformer that takes distorted latents as input and predicts the high-quality VQ-VAE index for facial restoration. Building upon this, Jiang et al. [27, 28] proposed transmitting the predicted indices to achieve restoration-based facial conferencing. In this paper, we explore the characteristics of latent space modeling in the realm of generative image compression. We specially design a transform coding paradigm in the latent space, which demonstrates superior effectiveness. + +![](images/8b3b196db7993380f4bd7fac041dc1d0e812942d29f6640bae52acfac4d3d834.jpg) +(a) + +![](images/3ff63c56751249968e11c1c1c51505c547412dcc4985e1ea5f7d574356d56700.jpg) +(b) +Figure 4. Illustration of the transform coding in latent space. (a) The model structure of the transform coding module. We further compare it with other coding schemes in operational diagrams : (b) indices-map coding [27, 28, 43], (c) transform coding with factorized hyper module [5, 21, 38] and (d) proposed transform coding with categorical hyper module. Here AE and AD denote arithmetic encoding and decoding, VQ-E and VQ-D stand for VQ-indices-map encoding and decoding, Q refers to scalar quantization, U signifies the addition of uniform noise as a differential simulation of Q, and S denotes the spatial context entropy module. + +![](images/69fe87705e8f0aef6b5878c6b940b55d3f2c1a71d562bf3eed581a7b650006af.jpg) +(c) + +![](images/8c14a71bca9fa0a737d269db45feda0d7cd8091ed7b118e487613a365934fff0.jpg) +(d) + +# 3. Method + +# 3.1. Overview + +In this section, we introduce the details of the proposed Generative Latent Coding (GLC) architecture. To achieve high-perceptual-quality compression, GLC encodes image into a perception-aligned latent space through a generative latent auto-encoder, and perform transform coding on the latent representations for lower bitrate. As depicted in the left of Figure 3, the input image $x$ is firstly encoded into the latent $l$ using the latent encoder $E$ . Then $l$ undergoes an analysis transform $g_{a}$ to produce the code $y$ , which is further scalar-quantized to $\hat{y}$ for entropy coding. Following that, a synthesis transform $g_{s}$ is employed to transform $\hat{y}$ back to $l$ , and finally, the reconstruction $\hat{x}$ is generated by the latent decoder $D$ . This entire process is formulated as: + +$$ +\begin{array}{l} l = E (x), y = g _ {a} (l) \\ \hat {y} = Q (y) \tag {1} \\ \end{array} +$$ + +$$ +\hat {l} = g _ {s} (\hat {y}), \hat {x} = D (\hat {l}) +$$ + +# 3.2. Generative Latent Auto-Encoder + +To achieve high-quality generative latent coding, how to obtain a human-perception-aligned latent space is a crucial challenge. In GLC, we address it by employing the generative VQ-VAE [17] as the latent auto-encoder $(E$ and $D)$ . By mapping images into visual semantic elements within a codebook $C$ and incorporating a generative image decoding process, both semantic consistency and texture realism can be well guaranteed. Additionally, it contributes to the compression process through a sparse yet robust latent space, which is achieved by training with the discrete codebook $C$ as a variational bottleneck. + +# 3.3. Transform Coding in Latent Space + +To compress the latent representations $l$ , a direct approach is VQ-indices-map coding [27, 28, 43] (Figure 4b). How- + +![](images/afbc0991a5095cefef3137c9837b15d425912a21a0a1d60cfa2d6d50f4bf4747.jpg) +Original Image + +![](images/104b96d7fa6777e12ed848e401740afd8bdfefb4a8c50a8a72355f4634a2cf58.jpg) +$y = 0$ bits $z = 1000$ bits + +![](images/460943cf38de08e5b4ae78112309d23b141f4b44b0ab5c198f09e76c39f6453c.jpg) +$y = 0$ bits $z = 160$ bits + +![](images/a9e64e9ac1fcf78ce2221a773ae859d7eff1aedcd0fd33877f0e2af639dca0e9.jpg) +$y = 1800$ bits $z = 1000$ bits +Factorized z +Figure 5. Example of comparison between factorized and categorical hyper modules. The proposed categorical $z$ encodes essential semantic and structural information with much less bits. + +![](images/7fc7a3ee1aa2ba6b560866f36b38ac79bd69924bab8af1fb7088576a00c6b53d.jpg) +$y = 1640$ bits $z = 160$ bits +Categorical z + +ever, these methods often lack a careful consideration of the correlation among the latents, resulting in an insufficient redundancy reduction and consequently a high bit cost. In GLC, we introduce a transform coding module to compress the latent, replacing the vector-quantization step for more effective reduction of latent redundancy. As shown in Figure 4a, the latents are transformed into code $y$ using transformations $g_{a}$ , $g_{s}$ and then quantized to $\hat{y}$ . Entropy coding is applied to $\hat{y}$ based on a probability $p(\hat{y})$ , which is estimated by a categorical hyper module (Section 3.3.1) and a quadtree-partition-based spatial context module [38]. + +# 3.3.1 Categorical Hyper Module + +Factorized hyper module [5] (Figure 4c) is commonly employed in recent image compression schemes. However, at ultra-low bitrate, we notice that the factorized $z$ tends to encode low-level information such as color and texture, incur + +ring a high bit cost, as illustrated in Figure 5. To address it, we propose a categorical hyper module (Figure 4d), which utilizes a hyper codebook to store the basic semantic elements rather than the low-level information. This module comprises a hyper analysis transform $h_a$ , a hyper synthesis transform $h_s$ and a hyper codebook $C_h$ . The transformations are formulated as: + +$$ +z = h _ {a} (y), \hat {z} = V Q (z, C _ {h}), p r i o r _ {z} = h _ {s} (\hat {z}) \tag {2} +$$ + +where $z$ and $\hat{z}$ denote hyper-codes. $VQ(\cdot, C_h)$ represents vector-quantization by nearest lookup in $C_h$ . As shown in the right of Figure 5, the categorical $z$ is more inclined to capture high-level semantic information, which can be encoded with significantly fewer bits. + +# 3.3.2 Rate-Variable Transformation + +A notable advantage of transform coding over VQ-indices-map coding is its capability for rate-variable compression, which is a core functionality for a practical image codec. Indices-map coding is limited since the codebook can only model one specified distribution, but different rates naturally need different distributions. In contrast, transform coding converts latent into a unified Gaussian distribution, and variable-rate can be achieved by variable parameters (e.g., means and scales) of Gaussian. In GLC, we follow DCVC series [36-39, 47, 49] to incorporate a scalar $q$ in the transform coding to achieve rate-variable compression. + +# 4. Progressive Training + +As depicted in the right of Figure 3, we adopt a three-stage progressive training manner to fully leverage the potential of GLC. We initially learn a human-perception-aligned latent space to guarantee the perceptual quality, subsequently learn to perform transform coding on this latent space to achieve low bitrate, and finally fine-tune the entire network for superior compression performance. At each stage, distinct loss functions are adopted to guide different modules. + +# 4.1. Stage I: Auto-Encoder Learning + +To obtain a human-perception-aligned latent space for compression, we begin with training a generative VQ-VAE as the initialization of $E$ and $D$ . To ensure the sparsity of the latent space, an auxiliary codebook $C$ is employed to perform nearest vector-quantization, transforming $l$ to $\widetilde{l}$ . The supervision comprises reconstruction loss, perceptual loss [29], adversarial loss [18] and codebook loss [53]: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {S t a g e I}} = \left| \left| x - \hat {x} \right| \right| + \mathcal {L} _ {p e r} (x, \hat {x}) \tag {3} \\ + \lambda_ {a d v} \cdot \mathcal {L} _ {a d v} (x, \hat {x}) + \mathcal {L} _ {\text {c o d e b o o k}} \\ \end{array} +$$ + +Here, $\mathcal{L}_{per}$ corresponds to the LPIPS loss calculated using VGG [50] extracted features. $\mathcal{L}_{adv}$ is the adaptive PatchGAN adversarial loss [17] with a weight of $\lambda_{adv} = 0.8$ . + +![](images/686772d25e674bd1209af91b09224489b63f2e1a1818e78f9f6dcf8bb2ca6c84.jpg) +Figure 6. Illustration of code-prediction-based supervision. + +![](images/42fa88018f2af32acdaf021d1d840493356db5e66eec6dbae839dee3a623d4b3.jpg) + +The codebook loss is formulated as + +$$ +\mathcal {L} _ {\text {c o d e b o o k}} = \left| \left| \operatorname {s g} (l) - \tilde {l} \right| \right| + \beta \cdot \left| \left| \operatorname {s g} (\tilde {l}) - l \right| \right| \tag {4} +$$ + +where $\mathrm{sg}(\cdot)$ denotes the stop-gradient operator and $\beta = 0.25$ controls the update rates of the $E$ and $C$ . + +# 4.2. Stage II: Transform Coding Learning + +Given the trained latent space, we further learn to perform transform coding to achieve low-bitrate latent compression, while fixing the auto-encoder $E$ and $D$ . We introduce an auxiliary code predictor $CP$ to enhance the semantic consistency by necessitating the latent to possess the capability to predict the correct VQ-indices. As shown in Figure 6, we encode $l$ into VQ-indices by $M_l = VQ(l,C)$ and subsequently predict these indices by $\hat{M}_{\hat{l}} = CP(\hat{l})$ . So the code-prediction-based loss can be formulated by + +$$ +\mathcal {D} _ {c o d e} (l, \hat {l}) = \alpha \cdot C E (M _ {l}, \hat {M _ {l}}) + | | l - \hat {l} | | _ {2} ^ {2} \qquad (5) +$$ + +where $CE$ denotes the cross entropy loss and we set $\alpha = 0.5$ by default. Then the transform coding module can be supervised by the rate-distortion trade-off + +$$ +\mathcal {L} _ {\text {S t a g e I I}} = \mathbf {E} _ {x \sim p _ {X}} [ \mathcal {R} (\hat {y}) + \lambda \cdot \mathcal {D} _ {\text {c o d e}} (l, \hat {l}) ] \tag {6} +$$ + +where $\mathcal{R}$ is the estimated rate and $\lambda$ is used to control the trade-off. Note that a codebook loss (as formulated in Equation 4) is required to train the hyper codebook $C_h$ in the categorical hyper module. We omit it from the loss functions of both stage II and stage III for the sake of conciseness. + +# 4.3. Stage III: Joint Training + +Finally, we fine-tune the entire network with the pixel space supervision to achieve better compression performance. As shown in Figure 6, we extend the code-prediction-based latent supervision into the pixel space. Specifically, we utilize the encoder $E_{VQ}$ trained from stage I to encode $x$ and + +$\hat{x}$ into latent space by $\hat{l}^p = E_{VQ}(\hat{x})$ and $l^p = E_{VQ}(x)$ , so the code-prediction-based pixel loss can be calculated by $\mathcal{D}_{code}(l^p,\hat{l}^p)$ in the same formulation with Equation 5. Here we use $E_{VQ}$ since it can map the input data to a compatible latent space with the codebook $C$ for code prediction. The overall pixel supervision is defined as: + +$$ +\begin{array}{l} \mathcal {D} _ {\text {S t a g e I I I}} = | | x - \hat {x} | | + \mathcal {L} _ {p e r} (x, \hat {x}) \\ + \lambda_ {a d v} \cdot \mathcal {L} _ {a d v} (x, \hat {x}) + \lambda_ {c o d e} \cdot \mathcal {D} _ {c o d e} \left(l ^ {p}, \hat {l} ^ {p}\right) \tag {7} \\ \end{array} +$$ + +where we set $\lambda_{code} = 0.05$ by default. The rate-distortion trade-off supervision is: + +$$ +\mathcal {L} _ {\text {S t a g e I I I}} = \mathbf {E} _ {x \sim p _ {X}} \left[ \mathcal {R} (\hat {y}) + \lambda \cdot \mathcal {D} _ {\text {S t a g e I I I}} \right] \tag {8} +$$ + +# 4.4. Discussion of Code-Prediction-Based Loss + +Code prediction transformers [27, 28, 60] have demonstrated their effectiveness in high-quality image reconstruction. They typically input the predicted latent directly into the decoder for reconstruction. Different from these methods, in GLC, we suggest to consider code prediction solely as an auxiliary supervision during training, but not used in the inference process of the compression pipeline. + +This design is based on an observation: if a code prediction module is introduced before the decoder, the fine-tuning process in stage III cannot enhance compression performance further. It appears that the codebook becomes a performance bottleneck, restricting the decoder to receiving only the vector-quantized latent as input, which has already been well-trained in stage I. In GLC, by utilizing code prediction solely as auxiliary supervision during training, we eliminate this bottleneck, allowing the decoder to receive more flexible input for additional fine-tuning. We find that this code-prediction-based supervision effectively enhances the semantic consistency of the reconstructions, as shown in Figure 7. By necessitating the latent to possess the capability to predict the code index, the latent contains more semantic information, such as gestures and attributes. + +# 5. Experiments + +# 5.1. Implementation Details + +Training details. We train GLC for both natural image compression and facial image compression. For natural images, we conduct stage I on ImageNet training set [14], stage II and III on OpenImage test set [33], using randomly cropped $256 \times 256$ patches. For facial images, GLC is trained on FFHQ dataset [31] for all stages with a resolution of $512 \times 512$ . Both models are optimized by AdamW [42] with a batch size of 8. For each batch, we train the model with different $\lambda$ to achieve rate-variable compression. + +Evaluation dataset. We evaluate GLC on CLIC 2020 test set [51] with original resolution for natural image compression, and evaluate on CelebAHQ [30] with a resolution of + +![](images/00d8dc75728ecd5a4dd2019b3c3782075ae5c1ad855f8e025c6ee047839a158b.jpg) + +![](images/f326b6496b3a638e84c51a36711dc49e29f4175d70904798e09cf73328ffe87d.jpg) + +![](images/b15f4e3adcc6340552b7de0bbadba090e319570c0343c0c514b089b4f12acdc9.jpg) + +![](images/e0ee7f99d6570edc677c69dadb82f476e243090f014f0a737286c329e1f5818c.jpg) + +![](images/40998eeda359794e5baf91901d8a42342f6fc60c723d22a53ecccb9640f2e2c4.jpg) +Original Image + +![](images/0f182d6dbb3af5bdeff9923461cd2d699eb95fb21062e1e7cfc1bed5b261c68d.jpg) +w/o code-pred-based loss + +![](images/7229844ecc1d6331c262e46852d43a9bbe6dbb292fd7dd3ba4e3466f0ceb9de9.jpg) +w/ code-pred-based loss + +![](images/5e12b06237e39622632a63022ecc04fa662a1ee2990caddf564d374b35cc4a26.jpg) +Figure 7. An example of using code-prediction-based supervision after stage II. Code prediction loss enhances the semantic consistency of the compressed latent, such as the color of eye shadow and the opening of the mouth. + +$512 \times 512$ for facial image. We also show the results on Kodak [32], DIV2K [1] and MS-COCO 30K [40] in the supplementary material. + +Evaluation metrics. We measure bit-stream size by bits per pixel (bpp), and measure visual quality by reference perceptual metrics LPIPS [29] and DISTS [15] and no-reference perceptual metrics FID [23] and KID [6]. We also provide PSNR and MS-SSIM [56] results in the supplementary material for completeness. Nevertheless, it is worth note that these pixel-level distortion metrics PSNR, MS-SSIM and LPIPS have strong limitations when evaluating image compression at ultra-low bitrate, which is also mentioned in other works [15, 35]. We provide a clearer demonstration of it in the supplementary material. + +Baseline methods. We compare with traditional codec VVC [54], neural codec TCM [41], EVC [20], and generative codec FCC [26], Text+Sketch [35], HiFiC [44], MSILLM [46]. As some methods do not release models for ultra-low bitrate, we either retrain or fine-tune their models to suit such low bitrate. Text+Sketch is not evaluated on CLIC since it does not support compression in high resolution. In addition, we also compare with recent works HFD [24] and PerCo [8] in the supplementary material. For facial compression, we fine-tune EVC, TCM, HiFiC and MSILLM using FFHQ dataset for a fair comparison. + +# 5.2. Main Results + +Figure 8 shows the performance of the proposal and compared methods at ultra-low bitrate. On CLIC 2020, GLC demonstrates superiority in terms of DISTS, FID and KID than other methods. Specifically, GLC saves about $45\%$ bits compared to previous SOTA method MS-ILLM while maintaining an equivalent FID. When comparing the pixel-level metric LPIPS, GLC also achieves comparable performance with high-fidelity generative codec such as HiFiC and MS-ILLM. On CelebAHQ, GLC outperforms all other methods across all metrics by a large margin. + +![](images/6dc3f1d89dc3d53e894365ea39211e3f98302e0e20bab5ad21895a9677a918dd.jpg) + +![](images/f7078527f2c23414802e3e37ab8349a8747802b61e3123a6f806040ce85bbced.jpg) + +![](images/403dea6682aea885d23c00031f4f1bb80a92592acace05378a3cd688554aca01.jpg) + +![](images/80672e66c6094bdfa9f4a44624598d1dc510a0050c4bc129ae9849a8d1c6afa8.jpg) + +![](images/c8a5fafb393863e956b79750a9026a773aa0a4131b936da7a97fa256b99c4958.jpg) + +![](images/8c80faa331ad8711fbaed3614179b17c2c042137be2acfc8351cd5091af22d6b.jpg) +Figure 8. Comparison of methods on CLIC 2020 test set and CelebAHQ. + +![](images/ae20a0313821ff26b38593ad0d02e89f4c5a0d7b7deed04d5daf2d4ad3a6c31a.jpg) + +![](images/e80b271631dc5deb6d58ce9f702be0191ddf8330fe45ca7bf64b58494f86682e.jpg) + +![](images/577b515a6b710eda02570e14559ef65a960210e3cab3d0fb6384a201fbd8b4d7.jpg) + +![](images/b37e9dfeebaa98abf9ab3dc3c1c60019c490ede60ff70276c8716ee6cbd26de1.jpg) +Original + +![](images/da3c2f2a90f317114dd0f66db38db904d79ecf5bf865f2ab23ea0c83e1415466.jpg) +[Kodak] +Original +[CelebAHQ] +Figure 9. Qualitative examples of different methods on Kodak and CelebAHQ. More comparisons are in supplementary materials. + +![](images/3630dd87ce80b6f61463bac3f632d77d788ea541db3329aaefccb9de0dfef2d3.jpg) +Text+Sketch + +![](images/0399d03b2df8f80592f27b944e28fab3433bdc46373cb409c4414e6febeb597c.jpg) +0.0320 bpp +Wc +0.0115bpp + +![](images/ef1f42cbf9449f47bda56a2527d4b43d21518871d5d715761d0fd4dca403364c.jpg) +TCM + +![](images/aa91f8e1cd8218b10f2d83d40672a7e7cd4d543c6fdfd73c3ef7aec3190c5c69.jpg) +0.0787 bpp +TCM +0.0094 bpp + +![](images/e5c92f3959b1da52eac29b5ad6773ab19e61782fd90e7d1a5f1a69d78738c18f.jpg) +HiFiC + +![](images/3b518627e47e28da9b24fa31891946af601b5acf1d0e198d78b0182385b51843.jpg) +0.0426 bpp +HiFiC +0.0085 bpp + +![](images/ac1626058217735b60a17c04045e006124a996d73fb066643cb95dfa6a8125af.jpg) +MS-ILLM + +![](images/7fea7b78e033cb5524ff7be0cf7209f15d4deb05b698bd0359b6afcd0233cf3a.jpg) +0.0375 bpp +MS-ILLM +0.0088 bpp + +![](images/9d1ad4d9804d1eb22ce311ef547f926a08263eed0ee6a5962a8a4ec8c59f7702.jpg) +GLC + +![](images/611305432135b369f03cbe1954580b906c7afb1bd89a7af3521880ba727acf1c.jpg) +0.0300 bpp +GLC +0.0080 bpp + +Figure 9 shows the qualitative comparison results. For natural image compression, Text+Sketch generates low-fidelity reconstructions, while TCM, HiFiC, and MS-ILLM produce blurry results at ultra low bitrate. In contrast, GLC achieves high-fidelity and high-realism results. In the case of facial image compression, we find existing methods cannot produce satisfactory reconstruction due to the severely distorted information at even more extreme bitrate limitation (e.g., 0.01 bpp or lower). TCM and HiFiC fall short in generating highly realistic results. Even though MS-ILLM + +provides clearer details, it struggles to preserve correct facial attribute. Compared to them, GLC excels in both realism and fidelity at even lower bitrate. + +# 5.3. Ablation Study + +In this section, we conduct ablation studies to demonstrate the effectiveness of each proposed component. To provide a clearer comparison, we evaluate the BD-Rate [7] on the FID-BPP curve on the CLIC 2020 test set. + +Transform coding. A straightforward approach to com + +Table 1. Ablation study on latent-space compression. + +
Latent coding schemeProbability model of zBD-Rate ↓
Indices-map coding-66.2%
Transform codingFactorized prior17.7%
Categorical prior0%
+ +Table 2. Ablation study on the code prediction module. + +
code prediction usageBD-Rate ↓
w/o code pred.13.1%
code pred. in network60.7%
code pred. as supervision0%
+ +press the VQ-VAE latents is indices-map coding [27, 28, 43]. However, it causes $66.2\%$ performance loss compared with transform coding, as shown in Table 1. It shows the effectiveness of transform coding on reducing redundancy. + +Categorical hyper module. In Section 3.3, we illustrate the superiority of employing a categorical prior for $z$ compared to the commonly used factorized prior. Table 1 further provides a quantitative comparison, demonstrating a significant improvement of $17.7\%$ with such design. + +Code-prediction-based supervision. In Section 4, we suggest employing the code prediction module as an auxiliary loss during training, instead of during the inference process of the model pipeline as in [27, 28, 60]. As shown in Table 2, incorporating the code prediction module directly into the network leads to a $60.7\%$ performance drop. We further remove the code-prediction-based supervision for comparison, and results show that adopting the code-prediction-based supervision brings a $13.1\%$ improvement. + +# 6. Applications + +Leveraging the potent latent space, our GLC pipeline opens avenues for exploring various vision applications. In this paper, we implement image restoration and style transfer as examples to show its potential. As depicted in Figure 10, for image restoration, we train a restoration encoder to map distorted images into clean latents, allowing users to directly compress a noisy image and decompress a clean one without additional cost. We compare our restoration application with a straightforward scheme of cascading an additional restoration network [57] with the GLC codec. The results in Table 3 indicate superior performance for our restoration application without the need for extra model parameters. Similarly, users can directly decode the latent into another style through a stylization decoder to achieve style transfer. We hope such versatility of the generative latent space will foster connections between image compression and other vision tasks in future research. + +![](images/0ce959ac2b89800cefdb94c256a4db5c43ac631d62c91fcdb1a4633fcee24b2e.jpg) +Figure 10. Generative latent applications built on GLC. In the practical image compression system, users can choose different encoders and decoders, to compress an image (medium), compress a distorted image and decompress a clear one (top), or decode the image into another style (bottom). + +Table 3. Comparison for different joint restoration and compression schemes on CLIC 2020 test set. + +
SchemeBPP ↓FID ↓DISTS ↓Parameters
Restormer [57] + GLC Codec0.031410.790.117425M + 109M
GLC Restoration Application0.02998.620.1081109M
+ +# 7. Conclusion and Limitation + +In this paper, we introduce a generative latent coding (GLC) scheme to achieve high-fidelity and high-realism generative compression at ultra-low bitrate. Unlike most existing pixel-domain codecs, GLC performs transform coding on the latent domain of a generative VQ-VAE. By incorporating a categorical hyper module and a code-prediction-based supervision, GLC demonstrates state-of-the-art performance on several benchmarks. We further develop several vision applications on the GLC pipeline to demonstrate its practical potential. + +However, as a generative image codec trained on specified datasets, the generalization capability of GLC is not always satisfactory. For instance, GLC cannot guarantee a clear and accurate reconstruction of screen contents, as illustrated in the supplementary material. Future work will focus on addressing this limitation, to enhance the generalization ability of GLC by improving the model structure and the training strategy. + +# References + +[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 126-135, 2017. 6 +[2] Eirikur Agustsson, Michael Tschannen, Fabian Mentzer, Radu Timofte, and Luc Van Gool. Generative adversarial networks for extreme learned image compression. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 221-231, 2019. 3 +[3] Eirikur Agustsson, David Minnen, George Toderici, and Fabian Mentzer. Multi-realism image compression with a conditional generator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22324-22333, 2023. 3 +[4] Johannes Balle, Valero Laparra, and Eero P Simoncelli. End-to-end optimized image compression. In 5th International Conference on Learning Representations, ICLR 2017, 2017. 1, 3 +[5] Johannes Balle, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. In International Conference on Learning Representations, 2018. 1, 2, 3, 4 +[6] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. arXiv preprint arXiv:1801.01401, 2018. 6 +[7] Gisle Bjontegaard. Calculation of average psnr differences between rd-curves. ITU SG16 Doc. VCEG-M33, 2001. 7 +[8] Marlene Careil, Matthew J Muckley, Jakob Verbeek, and Stéphane Lathuilière. Towards image compression with perfect realism at ultra-low bitrates. In The Twelfth International Conference on Learning Representations, 2023. 6 +[9] Jianhui Chang, Zhenghui Zhao, Chuanmin Jia, Shiqi Wang, Lingbo Yang, Qi Mao, Jian Zhang, and Siwei Ma. Conceptual compression via deep structure and texture synthesis. IEEE Transactions on Image Processing, 31:2809-2823, 2022. 3 +[10] Xi Chen, Diederik P Kingma, Tim Salimans, Yan Duan, Prafulla Dhariwal, John Schulman, Ilya Sutskever, and Pieter Abbeel. Variational lossy autoencoder. arXiv preprint arXiv:1611.02731, 2016. 3 +[11] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned image compression with discretized gaussian mixture likelihoods and attention modules. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7939-7948, 2020. 1, 3 +[12] Thomas M Cover. Elements of information theory. John Wiley & Sons, 1999. 3 +[13] Ze Cui, Jing Wang, Shangyin Gao, Tiansheng Guo, Yihui Feng, and Bo Bai. Asymmetric gained deep image compression with continuous rate adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10532-10541, 2021. 3 +[14] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6 + +[15] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. IEEE transactions on pattern analysis and machine intelligence, 44(5):2567-2581, 2020. 1, 6 +[16] Alaaeldin El-Nouby, Matthew J Muckley, Karen Ullrich, Ivan Laptev, Jakob Verbeek, and Herve Jegou. Image compression with product quantized masked image modeling. Transactions on Machine Learning Research, 2022. 3 +[17] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 1, 2, 3, 4, 5 +[18] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1, 5 +[19] Vivek K Goyal. Theoretical foundations of transform coding. IEEE Signal Processing Magazine, 18(5):9-21, 2001. 1, 3 +[20] Wang Guo-Hua, Jiahao Li, Bin Li, and Yan Lu. Evc: Towards real-time neural image compression with mask decay. In The Eleventh International Conference on Learning Representations, 2022. 1, 3, 6 +[21] Dailan He, Yaoyan Zheng, Baocheng Sun, Yan Wang, and Hongwei Qin. Checkerboard context model for efficient learned image compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14771-14780, 2021. 4 +[22] Dailan He, Ziming Yang, Hongjiu Yu, Tongda Xu, Jixiang Luo, Yuan Chen, Chenjian Gao, Xinjie Shi, Hongwei Qin, and Yan Wang. Po-elic: Perception-oriented efficient learned image coding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1764–1769, 2022. 3 +[23] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6 +[24] Emiel Hoogeboom, Eirikur Agustsson, Fabian Mentzer, Luca Versari, George Toderici, and Lucas Theis. High-fidelity image compression with score-based generative models. arXiv preprint arXiv:2305.18231, 2023. 3, 6 +[25] Yueyu Hu, Shuai Yang, Wenhan Yang, Ling-Yu Duan, and Jiaying Liu. Towards coding for human and machine vision: A scalable image coding approach. In 2020 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2020. 3 +[26] Shoma Iwai, Tomo Miyazaki, Yoshihiro Sugaya, and Shinichiro Omachi. Fidelity-controllable extreme image compression with generative adversarial networks. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 8235-8242. IEEE, 2021. 3, 6 +[27] Wei Jiang, Hyomin Choi, and Fabien Racapé. Adaptive human-centric video compression for humans and machines. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1121-1129, 2023. 2, 3, 4, 6, 8 + +[28] Wei Jiang, Hyomin Choi, Fabien Racapé, Simon Feltman, and Fatih Kamisli. Face restoration-based scalable quality coding for video conferencing. In 2023 IEEE International Conference on Multimedia and Expo Workshops (ICMEW), pages 206-211. IEEE, 2023. 2, 3, 4, 6, 8 +[29] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016. 1, 5, 6 +[30] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In International Conference on Learning Representations, 2018. 2, 6 +[31] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 6 +[32] Kodak Lossless True Color Image Suite. http://r0k.us/graphics/kodak/.6 +[33] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision, 128(7):1956-1981, 2020. 6 +[34] Jooyoung Lee, Seunghyun Cho, and Seung-Kwon Beck. Context-adaptive entropy model for end-to-end optimized image compression. In International Conference on Learning Representations, 2018. 3 +[35] Eric Lei, Yigit Berkay Uslu, Hamed Hassani, and Shirin Saeedi Bidokhti. Text+ sketch: Image compression at ultra low rates. In ICML 2023 Workshop Neural Compression: From Information Theory to Applications, 2023. 2, 3, 6 +[36] Jiahao Li, Bin Li, and Yan Lu. Deep contextual video compression. Advances in Neural Information Processing Systems, 34:18114-18125, 2021. 5 +[37] Jiahao Li, Bin Li, and Yan Lu. Hybrid spatial-temporal entropy modelling for neural video compression. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1503-1511, 2022. +[38] Jiahao Li, Bin Li, and Yan Lu. Neural video compression with diverse contexts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22616-22626, 2023. 3, 4 +[39] Jiahao Li, Bin Li, and Yan Lu. Neural video compression with feature modulation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 17-21, 2024, 2024. 5 +[40] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6 + +[41] Jinming Liu, Heming Sun, and Jiro Katto. Learned image compression with mixed transformer-cnn architectures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14388-14397, 2023. 1, 3, 6 +[42] Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam. 2018. 6 +[43] Qi Mao, Tinghan Yang, Yinuo Zhang, Shuyin Pan, Meng Wang, Shiqi Wang, and Siwei Ma. Extreme image compression using fine-tuned vqgan models. arXiv preprint arXiv:2307.08265, 2023. 2, 4, 8 +[44] Fabian Mentzer, George D Toderici, Michael Tschannen, and Eirikur Agustsson. High-fidelity generative image compression. Advances in Neural Information Processing Systems, 33:11913-11924, 2020. 2, 3, 6 +[45] David Minnen, Johannes Balle, and George D Toderici. Joint autoregressive and hierarchical priors for learned image compression. Advances in neural information processing systems, 31, 2018. 1, 3 +[46] Matthew J. Muckley, Alaaeldin El-Nouby, Karen Ullrich, Hervé Jégou, and Jakob Verbeek. Improving statistical fidelity for neural image compression with implicit local likelihood models. In International Conference on Machine Learning, 2023. 1, 2, 3, 6 +[47] Linfeng Qi, Jiahao Li, Bin Li, Houqiang Li, and Yan Lu. Motion information propagation for neural video compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6111-6120, 2023. 5 +[48] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3 +[49] Xihua Sheng, Jiahao Li, Bin Li, Li Li, Dong Liu, and Yan Lu. Temporal context mining for learned video compression. IEEE Transactions on Multimedia, 2022. 5 +[50] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 5 +[51] George Toderici, Lucas Theis, Nick Johnston, Eirikur Agustsson, Fabian Mentzer, Johannes Balle, Wenzhe Shi, and Radu Timofte. Clic 2020: Challenge on learned image compression, 2020, 2020. 2, 6 +[52] Aäron Van Den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks. In International conference on machine learning, pages 1747-1756. PMLR, 2016. 3 +[53] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 1, 2, 3, 5 +[54] VVC-21.2. https://vcgit.hhi.fraunhofer.de/jvet/VVCSoftware_VTM/~/tree/VTM-21.2. Accessed: 10/23/2023, 2023. 1, 6 +[55] Gregory K Wallace. TheJPEG still picture compression standard.Communications of the ACM, 34(4):30-44, 1991.1 +[56] Zhou Wang, Eero P Simoncelli, and Alan C Bovik. Multiscale structural similarity for image quality assessment. In + +The Thrity-Seventh Asilomar Conference on Signals, Systems & Computers, 2003, pages 1398-1402. IEEE, 2003. 6 +[57] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728-5739, 2022. 8 +[58] Jing Zhao, Bin Li, Jiahao Li, Ruiqin Xiong, and Yan Lu. A universal encoder rate distortion optimization framework for learned compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1880-1884, 2021. 3 +[59] Jing Zhao, Bin Li, Jiahao Li, Ruiqin Xiong, and Yan Lu. A universal optimization framework for learning-based image codec. ACM Transactions on Multimedia Computing, Communications and Applications, 20(1):1-19, 2023. 3 +[60] Shangchen Zhou, Kelvin Chan, Chongyi Li, and Chen Change Loy. Towards robust blind face restoration with codebook lookup transformer. Advances in Neural Information Processing Systems, 35:30599-30611, 2022. 2, 3, 6, 8 \ No newline at end of file diff --git a/2024/Generative Latent Coding for Ultra-Low Bitrate Image Compression/images.zip b/2024/Generative Latent Coding for Ultra-Low Bitrate Image Compression/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ffaaa26da5323f26f6c19ca56cf041f30b3168d8 --- /dev/null +++ b/2024/Generative Latent Coding for Ultra-Low Bitrate Image Compression/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:970cdfaf00ee356c675bde9a5b6a83e862f651fb14b1114bfc485913a6609444 +size 899063 diff --git a/2024/Generative Latent Coding for Ultra-Low Bitrate Image Compression/layout.json b/2024/Generative Latent Coding for Ultra-Low Bitrate Image Compression/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bf00c878f210d5565d4ca8ee0f465483c5e1d968 --- /dev/null +++ b/2024/Generative Latent Coding for Ultra-Low Bitrate Image Compression/layout.json @@ -0,0 +1,11510 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 84, + 103, + 509, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 103, + 509, + 121 + ], + "spans": [ + { + "bbox": [ + 84, + 103, + 509, + 121 + ], + "type": "text", + "content": "Generative Latent Coding for Ultra-Low Bitrate Image Compression" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 142, + 479, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 142, + 479, + 158 + ], + "spans": [ + { + "bbox": [ + 111, + 142, + 479, + 158 + ], + "type": "text", + "content": "Zhaoyang Jia" + }, + { + "bbox": [ + 111, + 142, + 479, + 158 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 111, + 142, + 479, + 158 + ], + "type": "text", + "content": " Jiahao Li" + }, + { + "bbox": [ + 111, + 142, + 479, + 158 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 111, + 142, + 479, + 158 + ], + "type": "text", + "content": " Bin Li" + }, + { + "bbox": [ + 111, + 142, + 479, + 158 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 111, + 142, + 479, + 158 + ], + "type": "text", + "content": " Houqiang Li" + }, + { + "bbox": [ + 111, + 142, + 479, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 111, + 142, + 479, + 158 + ], + "type": "text", + "content": " Yan Lu" + }, + { + "bbox": [ + 111, + 142, + 479, + 158 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 111, + 158, + 483, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 158, + 483, + 171 + ], + "spans": [ + { + "bbox": [ + 111, + 158, + 483, + 171 + ], + "type": "text", + "content": "1 University of Science and Technology of China 2 Microsoft Research Asia" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 70, + 174, + 520, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 174, + 520, + 185 + ], + "spans": [ + { + "bbox": [ + 70, + 174, + 520, + 185 + ], + "type": "text", + "content": "jzy_ustc@mail.ustc.edu.cn, lihq@ustc.edu.cn, {li.jiahao, libin, yanlu}@microsoft.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 239, + 290, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 290, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 290, + 515 + ], + "type": "text", + "content": "Most existing image compression approaches perform transform coding in the pixel space to reduce its spatial redundancy. However, they encounter difficulties in achieving both high-realism and high-fidelity at low bitrate, as the pixel-space distortion may not align with human perception. To address this issue, we introduce a Generative Latent Coding (GLC) architecture, which performs transform coding in the latent space of a generative vector-quantized variational auto-encoder (VQ-VAE), instead of in the pixel space. The generative latent space is characterized by greater sparsity, richer semantic and better alignment with human perception, rendering it advantageous for achieving high-realism and high-fidelity compression. Additionally, we introduce a categorical hyper module to reduce the bit cost of hyper-information, and a code-prediction-based supervision to enhance the semantic consistency. Experiments demonstrate that our GLC maintains high visual quality with less than 0.04 bpp on natural images and less than 0.01 bpp on facial images. On the CLIC2020 test set, we achieve the same FID as MS-ILLM with " + }, + { + "bbox": [ + 46, + 239, + 290, + 515 + ], + "type": "inline_equation", + "content": "45\\%" + }, + { + "bbox": [ + 46, + 239, + 290, + 515 + ], + "type": "text", + "content": " fewer bits. Furthermore, the powerful generative latent space enables various applications built on our GLC pipeline, such as image restoration and style transfer." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 530, + 128, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 530, + 128, + 542 + ], + "spans": [ + { + "bbox": [ + 47, + 530, + 128, + 542 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 551, + 287, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 551, + 287, + 671 + ], + "spans": [ + { + "bbox": [ + 46, + 551, + 287, + 671 + ], + "type": "text", + "content": "Amid the ongoing surge of digital visual data, the importance of achieving high-efficiency image compression becomes increasingly paramount. From the traditional compression standards [54, 55] to the emerging learned image compression models [4, 5, 11, 20, 41, 45], most compression algorithms follow the pixel-space transform coding [4, 19] paradigm. Specifically, they convert pixels into compact representations through a transform module, which eliminates the redundancy to reduce the bit cost in the subsequent entropy coding process." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 59, + 672, + 287, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 672, + 287, + 684 + ], + "spans": [ + { + "bbox": [ + 59, + 672, + 287, + 684 + ], + "type": "text", + "content": "However, we observe a common inherent limitation in" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 311, + 213, + 542, + 392 + ], + "blocks": [ + { + "bbox": [ + 311, + 213, + 542, + 392 + ], + "lines": [ + { + "bbox": [ + 311, + 213, + 542, + 392 + ], + "spans": [ + { + "bbox": [ + 311, + 213, + 542, + 392 + ], + "type": "image", + "image_path": "a2ba8fbb87b5b0653a8d05ef617dd3e5b676dab018b51f1cf11a13168b793beb.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 395, + 547, + 473 + ], + "lines": [ + { + "bbox": [ + 304, + 395, + 547, + 473 + ], + "spans": [ + { + "bbox": [ + 304, + 395, + 547, + 473 + ], + "type": "text", + "content": "Figure 1. Generative latent space of VQ-VAE exhibits better alignment with human perception than pixel space for ultra-low bitrate compression. Under comparable distortion levels (measured by signal-to-noise ratio, SNR), latent-space compression produces reconstructions with superior perceptual quality (measured by DISTS [15]) than pixel-space generative codec MS-ILLM [46], as the compressed latents remain in the same latent code space." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 485, + 547, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 547, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 547, + 605 + ], + "type": "text", + "content": "these methods: the pixel-space distortion is not always consistent with the human perception, especially at low bitrate. In practice, human observers prioritize the semantic consistency and the texture realism of an image, but this information is difficult to be adequately exploited solely by a pixel-space transform module. As shown in the left of Figure 1, pixel-space generative image codec MS-ILLM [46] struggles to guarantee visual quality at low bitrate, even after it incorporates perceptual supervision [29] and adversarial supervision [18] within the pixel space." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "content": "Based on such observation, a natural problem arises: how can we compress images in a way that aligns with human perception? To address this challenge, we introduce a Generative Latent Coding (GLC) paradigm. In GLC, we first encode images into a generative latent space that aligns with human perception, and subsequently perform transform coding in this latent space, instead of in pixel space. To concretize this concept, we adopt a generative vector-quantized variational auto-encoder (VQ-VAE) [17, 53] to" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "type": "text", + "content": "*This work was done when Zhaoyang Jia was an intern at Microsoft Research Asia." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26088" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 71, + 542, + 305 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 542, + 305 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 542, + 305 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 542, + 305 + ], + "type": "image", + "image_path": "fa2f7ad2599bdd232dd9af00fa873e08404ca7c186b9f152ea0b146474fe767e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 315, + 547, + 350 + ], + "lines": [ + { + "bbox": [ + 46, + 315, + 547, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 315, + 547, + 350 + ], + "type": "text", + "content": "Figure 2. A qualitative comparison between HiFiC [44], MS-ILLM [46], Text+Sketch [35] and the proposed GLC. GLC produces images with high visual quality, even in regions with complex texture. In contrast, HiFiC and MS-ILLM exhibit noticeable artifacts, and Text+Sketch generates results that deviate significantly from the input. Best viewed when zoomed in." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 364, + 289, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 364, + 289, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 364, + 289, + 495 + ], + "type": "text", + "content": "produce the latent space, which offers three significant advantages: 1) The discrete codes of VQ-VAE encapsulate semantic visual components [53], allowing GLC to focus on compressing the semantic content to guarantee fidelity. 2) Generative VQ-VAE exhibits remarkable generative capabilities [17] for high-realism texture reconstruction. 3) The discrete variational bottleneck naturally brings a low-entropy and distortion-robust latent space for compression. Thanks to such characteristics, GLC is more aligned with human perception to achieve enhanced visual quality, as demonstrated in the right of Figure 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "text", + "content": "When implementing GLC, two crucial questions persist: How to effectively compress the generative latents? And how to supervise the generative latent coding? A straightforward approach to compress VQ-VAE latents is indices-map coding [27, 28, 43], but it is limited by the ineffective redundancy reduction between indices and the lack of rate-variable coding support. In this paper, we propose a novel generative-latent-space transform coding approach, where an effective rate-variable structure is adopted to reduce latent redundancy for higher compression ratio. In addition, a categorical hyper module is introduced to model the distribution of " + }, + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "text", + "content": " with a discrete codebook, which significantly reduces the bitrate of " + }, + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "text", + "content": " when compared to the factorized hyper module [5]. As for the supervision of GLC, inspired by recent code prediction transformers [27, 28, 60], a code-prediction-based supervision is proposed. It serves as an auxiliary supervision employed solely in the training process to greatly enhance the semantic consistency." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 364, + 545, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 364, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 364, + 545, + 495 + ], + "type": "text", + "content": "Benefited from these advanced designs, our GLC achieves excellent performance on both natural and facial images. In the CLIC 2020 test set [51], GLC attains a bitrate less than 0.04 bpp while delivering high visual quality. It obtains " + }, + { + "bbox": [ + 304, + 364, + 545, + 495 + ], + "type": "inline_equation", + "content": "45\\%" + }, + { + "bbox": [ + 304, + 364, + 545, + 495 + ], + "type": "text", + "content": " bit savings compared to MS-ILLM at the same FID. In the CelebAHQ [30] dataset, GLC achieves an even lower bitrate of less than 0.01 bpp. As shown in Figure 2, compared with recent advanced generative image compression approaches MS-ILLM [46] and Text+Sketch [35], GLC produces more visually appealing compression results with a lower bit cost." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 495, + 545, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 545, + 590 + ], + "type": "text", + "content": "Furthermore, leveraging the representative generative latent space, GLC supports various vision applications such as image restoration and style transfer. By replacing the compression encoder with a restoration encoder, the proposed restoration application surpasses the performance of cascading a restoration model and a neural codec. We hope such versatility of generative latent space will help connect image compression with other vision tasks in the future." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 592, + 484, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 592, + 484, + 602 + ], + "spans": [ + { + "bbox": [ + 317, + 592, + 484, + 602 + ], + "type": "text", + "content": "In summary, our main contributions are :" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 605, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 306, + 605, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 605, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 306, + 605, + 545, + 653 + ], + "type": "text", + "content": "- We present a generative latent coding (GLC) scheme, which performs transform coding in the generative latent space of a VQ-VAE to achieve high-fidelity and high-realism compression at ultra-low bitrate." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 654, + 545, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 654, + 545, + 701 + ], + "spans": [ + { + "bbox": [ + 306, + 654, + 545, + 701 + ], + "type": "text", + "content": "- We introduce a categorical hyper module to significantly reduce the bit cost of hyper information. Additionally, a code-prediction-based supervision is adopted to enhance the perceptual quality." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 701, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 712 + ], + "type": "text", + "content": "- GLC obtains " + }, + { + "bbox": [ + 306, + 701, + 545, + 712 + ], + "type": "inline_equation", + "content": "45\\%" + }, + { + "bbox": [ + 306, + 701, + 545, + 712 + ], + "type": "text", + "content": " bit reduction on CLIC2020 with the" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26089" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 77, + 236, + 231 + ], + "blocks": [ + { + "bbox": [ + 52, + 77, + 236, + 231 + ], + "lines": [ + { + "bbox": [ + 52, + 77, + 236, + 231 + ], + "spans": [ + { + "bbox": [ + 52, + 77, + 236, + 231 + ], + "type": "image", + "image_path": "752991a6a6488088bb4007ad888aa69defed14166a7fc044e13ee45a4aeb716a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 239, + 547, + 305 + ], + "lines": [ + { + "bbox": [ + 46, + 239, + 547, + 305 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 547, + 305 + ], + "type": "text", + "content": "Figure 3. Illustration of the proposed Generative Latent Coding (GLC) framework. (Left) GLC firstly encodes the image into a generative latent representation (Section 3.2), then compresses the latent with transform coding (Section 3.3), and finally reconstructs image from the compressed latent. (Right) We progressively train GLC in three stages (Section 4): In stage I, we train a generative VQ-VAE to obtain a human-perception-aligned latent space. In stage II, the transform coding module learns to compress the latent with a code-prediction-based latent supervision (Figure 6). Finally, in stage III, the entire network is fine-tuned jointly with a code-prediction-based pixel supervision to further enhance the compression performance." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 242, + 77, + 318, + 228 + ], + "blocks": [ + { + "bbox": [ + 242, + 77, + 318, + 228 + ], + "lines": [ + { + "bbox": [ + 242, + 77, + 318, + 228 + ], + "spans": [ + { + "bbox": [ + 242, + 77, + 318, + 228 + ], + "type": "image", + "image_path": "2f1e3acd8c9ce4c5ce81e9bd16fd4673e8ab73f694a8683bdf1aaad66607829a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 320, + 78, + 408, + 227 + ], + "blocks": [ + { + "bbox": [ + 320, + 78, + 408, + 227 + ], + "lines": [ + { + "bbox": [ + 320, + 78, + 408, + 227 + ], + "spans": [ + { + "bbox": [ + 320, + 78, + 408, + 227 + ], + "type": "image", + "image_path": "e9a7896199c88e7736807b01977446915c7124d5a9e6e6b418b8185296c2d2f4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 410, + 78, + 541, + 227 + ], + "blocks": [ + { + "bbox": [ + 410, + 78, + 541, + 227 + ], + "lines": [ + { + "bbox": [ + 410, + 78, + 541, + 227 + ], + "spans": [ + { + "bbox": [ + 410, + 78, + 541, + 227 + ], + "type": "image", + "image_path": "0a0612e6323cbe6ddcedcc5580590af275e60cfcf2e2398da48b3985f50e2232.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 318, + 288, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 318, + 288, + 343 + ], + "spans": [ + { + "bbox": [ + 55, + 318, + 288, + 343 + ], + "type": "text", + "content": "same FID as the previous SOTA. Furthermore, GLC enables various application within its latent space." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 352, + 139, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 352, + 139, + 365 + ], + "spans": [ + { + "bbox": [ + 47, + 352, + 139, + 365 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 372, + 205, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 372, + 205, + 385 + ], + "spans": [ + { + "bbox": [ + 47, + 372, + 205, + 385 + ], + "type": "text", + "content": "2.1. Learned Image Compression" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 390, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 390, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 390, + 287, + 521 + ], + "type": "text", + "content": "Lossy image compression is grounded on Shannon's rate-distortion theory [12]. Balle et al. [4] first proposed to utilize neural networks for pixel-space transform coding [19], employing analysis and synthesis transform modules to convert images into compact representations for entropy coding. Subsequently, some researches make strides in improving the probability model [4, 5, 11, 34, 38, 45] for more accurate estimation, while others explore the network structure [11, 41], optimization algorithm [58, 59] and rate-variable coding [13, 20] for improved compression performance and practicality." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 522, + 288, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 288, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 288, + 665 + ], + "type": "text", + "content": "A recently raised critical challenge in image compression is how to improve the perceptual quality of the reconstruction. Agustsson et al. [2] first introduced the concept of generative compression, which compresses essential image features and generate distorted details using GAN. Some subsequent works [9, 25] extract image sketches and latent codes to ensure geometry-consistency. Text+Sketch [35] utilizes a conditional-diffusion model to generate image based on image captions and sketches, achieving superior perceptual quality. While these schemes produce visual appealing results, they often deviate significantly from the input and cannot guarantee the semantic consistency." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "To achieve generative compression with high-fidelity, Mentzer et al. [44] further studied the network structure and generative adversarial loss to enable high-fidelity compression. Subsequent researches further enhance the trans" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 318, + 547, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 318, + 547, + 390 + ], + "spans": [ + { + "bbox": [ + 304, + 318, + 547, + 390 + ], + "type": "text", + "content": "form coding [16, 22], generative post-processing [24] or focus on controlling the trade-off between fidelity and realism [3, 26]. Recently, MS-ILLM [46] introduces a noisy discriminator which is conditioned on quantized local image representations to greatly enhance the statistical fidelity." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 403, + 438, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 403, + 438, + 417 + ], + "spans": [ + { + "bbox": [ + 306, + 403, + 438, + 417 + ], + "type": "text", + "content": "2.2. Latent Space Modeling" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 424, + 547, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 424, + 547, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 424, + 547, + 581 + ], + "type": "text", + "content": "Latent space image modeling means modeling the distribution of image within the latent space of a neural network. This technique has been primarily developed for image generation. Chen et al. [10] and Oord et al. [53] introduced PixelCNN [52] in the latent space of VAE and VQ-VAE for image generation. Esser et al. [17] took it a step further by incorporating transformers into VQ-VAE latent space for high-quality generation. More recently, Rombach et al. [48] employed diffusion models to model the latent space of VQ-VAE, achieving remarkable results in high-resolution image generation. These studies underscore the potential of image processing within the generative latent space, particularly in the latent space of VQ-VAE." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "content": "Recently, the concept of latent space modeling has been extended to other tasks. CodeFormer [60] introduces a code prediction transformer that takes distorted latents as input and predicts the high-quality VQ-VAE index for facial restoration. Building upon this, Jiang et al. [27, 28] proposed transmitting the predicted indices to achieve restoration-based facial conferencing. In this paper, we explore the characteristics of latent space modeling in the realm of generative image compression. We specially design a transform coding paradigm in the latent space, which demonstrates superior effectiveness." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26090" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 242, + 169 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 242, + 169 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 242, + 169 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 242, + 169 + ], + "type": "image", + "image_path": "8b3b196db7993380f4bd7fac041dc1d0e812942d29f6640bae52acfac4d3d834.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 145, + 171, + 156, + 180 + ], + "lines": [ + { + "bbox": [ + 145, + 171, + 156, + 180 + ], + "spans": [ + { + "bbox": [ + 145, + 171, + 156, + 180 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 256, + 103, + 335, + 167 + ], + "blocks": [ + { + "bbox": [ + 256, + 103, + 335, + 167 + ], + "lines": [ + { + "bbox": [ + 256, + 103, + 335, + 167 + ], + "spans": [ + { + "bbox": [ + 256, + 103, + 335, + 167 + ], + "type": "image", + "image_path": "3ff63c56751249968e11c1c1c51505c547412dcc4985e1ea5f7d574356d56700.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 291, + 171, + 301, + 180 + ], + "lines": [ + { + "bbox": [ + 291, + 171, + 301, + 180 + ], + "spans": [ + { + "bbox": [ + 291, + 171, + 301, + 180 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 190, + 547, + 246 + ], + "lines": [ + { + "bbox": [ + 46, + 190, + 547, + 246 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 547, + 246 + ], + "type": "text", + "content": "Figure 4. Illustration of the transform coding in latent space. (a) The model structure of the transform coding module. We further compare it with other coding schemes in operational diagrams : (b) indices-map coding [27, 28, 43], (c) transform coding with factorized hyper module [5, 21, 38] and (d) proposed transform coding with categorical hyper module. Here AE and AD denote arithmetic encoding and decoding, VQ-E and VQ-D stand for VQ-indices-map encoding and decoding, Q refers to scalar quantization, U signifies the addition of uniform noise as a differential simulation of Q, and S denotes the spatial context entropy module." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 346, + 72, + 438, + 167 + ], + "blocks": [ + { + "bbox": [ + 346, + 72, + 438, + 167 + ], + "lines": [ + { + "bbox": [ + 346, + 72, + 438, + 167 + ], + "spans": [ + { + "bbox": [ + 346, + 72, + 438, + 167 + ], + "type": "image", + "image_path": "69fe87705e8f0aef6b5878c6b940b55d3f2c1a71d562bf3eed581a7b650006af.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 382, + 171, + 392, + 180 + ], + "lines": [ + { + "bbox": [ + 382, + 171, + 392, + 180 + ], + "spans": [ + { + "bbox": [ + 382, + 171, + 392, + 180 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 443, + 72, + 534, + 167 + ], + "blocks": [ + { + "bbox": [ + 443, + 72, + 534, + 167 + ], + "lines": [ + { + "bbox": [ + 443, + 72, + 534, + 167 + ], + "spans": [ + { + "bbox": [ + 443, + 72, + 534, + 167 + ], + "type": "image", + "image_path": "8c14a71bca9fa0a737d269db45feda0d7cd8091ed7b118e487613a365934fff0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 478, + 171, + 489, + 180 + ], + "lines": [ + { + "bbox": [ + 478, + 171, + 489, + 180 + ], + "spans": [ + { + "bbox": [ + 478, + 171, + 489, + 180 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 265, + 102, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 265, + 102, + 277 + ], + "spans": [ + { + "bbox": [ + 47, + 265, + 102, + 277 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 285, + 115, + 296 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 285, + 115, + 296 + ], + "spans": [ + { + "bbox": [ + 47, + 285, + 115, + 296 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 304, + 288, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 288, + 459 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 288, + 459 + ], + "type": "text", + "content": "In this section, we introduce the details of the proposed Generative Latent Coding (GLC) architecture. To achieve high-perceptual-quality compression, GLC encodes image into a perception-aligned latent space through a generative latent auto-encoder, and perform transform coding on the latent representations for lower bitrate. As depicted in the left of Figure 3, the input image " + }, + { + "bbox": [ + 46, + 304, + 288, + 459 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 304, + 288, + 459 + ], + "type": "text", + "content": " is firstly encoded into the latent " + }, + { + "bbox": [ + 46, + 304, + 288, + 459 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 304, + 288, + 459 + ], + "type": "text", + "content": " using the latent encoder " + }, + { + "bbox": [ + 46, + 304, + 288, + 459 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 46, + 304, + 288, + 459 + ], + "type": "text", + "content": ". Then " + }, + { + "bbox": [ + 46, + 304, + 288, + 459 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 304, + 288, + 459 + ], + "type": "text", + "content": " undergoes an analysis transform " + }, + { + "bbox": [ + 46, + 304, + 288, + 459 + ], + "type": "inline_equation", + "content": "g_{a}" + }, + { + "bbox": [ + 46, + 304, + 288, + 459 + ], + "type": "text", + "content": " to produce the code " + }, + { + "bbox": [ + 46, + 304, + 288, + 459 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 304, + 288, + 459 + ], + "type": "text", + "content": ", which is further scalar-quantized to " + }, + { + "bbox": [ + 46, + 304, + 288, + 459 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 46, + 304, + 288, + 459 + ], + "type": "text", + "content": " for entropy coding. Following that, a synthesis transform " + }, + { + "bbox": [ + 46, + 304, + 288, + 459 + ], + "type": "inline_equation", + "content": "g_{s}" + }, + { + "bbox": [ + 46, + 304, + 288, + 459 + ], + "type": "text", + "content": " is employed to transform " + }, + { + "bbox": [ + 46, + 304, + 288, + 459 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 46, + 304, + 288, + 459 + ], + "type": "text", + "content": " back to " + }, + { + "bbox": [ + 46, + 304, + 288, + 459 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 304, + 288, + 459 + ], + "type": "text", + "content": ", and finally, the reconstruction " + }, + { + "bbox": [ + 46, + 304, + 288, + 459 + ], + "type": "inline_equation", + "content": "\\hat{x}" + }, + { + "bbox": [ + 46, + 304, + 288, + 459 + ], + "type": "text", + "content": " is generated by the latent decoder " + }, + { + "bbox": [ + 46, + 304, + 288, + 459 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 304, + 288, + 459 + ], + "type": "text", + "content": ". This entire process is formulated as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 123, + 464, + 287, + 493 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 464, + 287, + 493 + ], + "spans": [ + { + "bbox": [ + 123, + 464, + 287, + 493 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} l = E (x), y = g _ {a} (l) \\\\ \\hat {y} = Q (y) \\tag {1} \\\\ \\end{array}", + "image_path": "7a8aaa2b45c57a3f415df7a0f65d798ea37ee83b66d254e5cfb74b36a7510803.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 123, + 495, + 210, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 495, + 210, + 510 + ], + "spans": [ + { + "bbox": [ + 123, + 495, + 210, + 510 + ], + "type": "interline_equation", + "content": "\\hat {l} = g _ {s} (\\hat {y}), \\hat {x} = D (\\hat {l})", + "image_path": "fa3ecf91dd73b8e477a2c25b2c625a5883ab818d1b2a86f236d611000ce30e25.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 515, + 224, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 515, + 224, + 526 + ], + "spans": [ + { + "bbox": [ + 47, + 515, + 224, + 526 + ], + "type": "text", + "content": "3.2. Generative Latent Auto-Encoder" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 533, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 664 + ], + "type": "text", + "content": "To achieve high-quality generative latent coding, how to obtain a human-perception-aligned latent space is a crucial challenge. In GLC, we address it by employing the generative VQ-VAE [17] as the latent auto-encoder " + }, + { + "bbox": [ + 46, + 533, + 287, + 664 + ], + "type": "inline_equation", + "content": "(E" + }, + { + "bbox": [ + 46, + 533, + 287, + 664 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 533, + 287, + 664 + ], + "type": "inline_equation", + "content": "D)" + }, + { + "bbox": [ + 46, + 533, + 287, + 664 + ], + "type": "text", + "content": ". By mapping images into visual semantic elements within a codebook " + }, + { + "bbox": [ + 46, + 533, + 287, + 664 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 533, + 287, + 664 + ], + "type": "text", + "content": " and incorporating a generative image decoding process, both semantic consistency and texture realism can be well guaranteed. Additionally, it contributes to the compression process through a sparse yet robust latent space, which is achieved by training with the discrete codebook " + }, + { + "bbox": [ + 46, + 533, + 287, + 664 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 533, + 287, + 664 + ], + "type": "text", + "content": " as a variational bottleneck." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 671, + 233, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 233, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 233, + 684 + ], + "type": "text", + "content": "3.3. Transform Coding in Latent Space" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "To compress the latent representations " + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": ", a direct approach is VQ-indices-map coding [27, 28, 43] (Figure 4b). How-" + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 310, + 311, + 375, + 375 + ], + "blocks": [ + { + "bbox": [ + 310, + 311, + 375, + 375 + ], + "lines": [ + { + "bbox": [ + 310, + 311, + 375, + 375 + ], + "spans": [ + { + "bbox": [ + 310, + 311, + 375, + 375 + ], + "type": "image", + "image_path": "afbc0991a5095cefef3137c9837b15d425912a21a0a1d60cfa2d6d50f4bf4747.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 377, + 365, + 386 + ], + "lines": [ + { + "bbox": [ + 324, + 377, + 365, + 386 + ], + "spans": [ + { + "bbox": [ + 324, + 377, + 365, + 386 + ], + "type": "text", + "content": "Original Image" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 384, + 271, + 451, + 335 + ], + "blocks": [ + { + "bbox": [ + 384, + 271, + 451, + 335 + ], + "lines": [ + { + "bbox": [ + 384, + 271, + 451, + 335 + ], + "spans": [ + { + "bbox": [ + 384, + 271, + 451, + 335 + ], + "type": "image", + "image_path": "104b96d7fa6777e12ed848e401740afd8bdfefb4a8c50a8a72355f4634a2cf58.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 383, + 336, + 452, + 343 + ], + "lines": [ + { + "bbox": [ + 383, + 336, + 452, + 343 + ], + "spans": [ + { + "bbox": [ + 383, + 336, + 452, + 343 + ], + "type": "inline_equation", + "content": "y = 0" + }, + { + "bbox": [ + 383, + 336, + 452, + 343 + ], + "type": "text", + "content": " bits " + }, + { + "bbox": [ + 383, + 336, + 452, + 343 + ], + "type": "inline_equation", + "content": "z = 1000" + }, + { + "bbox": [ + 383, + 336, + 452, + 343 + ], + "type": "text", + "content": " bits" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 466, + 271, + 531, + 335 + ], + "blocks": [ + { + "bbox": [ + 466, + 271, + 531, + 335 + ], + "lines": [ + { + "bbox": [ + 466, + 271, + 531, + 335 + ], + "spans": [ + { + "bbox": [ + 466, + 271, + 531, + 335 + ], + "type": "image", + "image_path": "460943cf38de08e5b4ae78112309d23b141f4b44b0ab5c198f09e76c39f6453c.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 468, + 336, + 534, + 343 + ], + "lines": [ + { + "bbox": [ + 468, + 336, + 534, + 343 + ], + "spans": [ + { + "bbox": [ + 468, + 336, + 534, + 343 + ], + "type": "inline_equation", + "content": "y = 0" + }, + { + "bbox": [ + 468, + 336, + 534, + 343 + ], + "type": "text", + "content": " bits " + }, + { + "bbox": [ + 468, + 336, + 534, + 343 + ], + "type": "inline_equation", + "content": "z = 160" + }, + { + "bbox": [ + 468, + 336, + 534, + 343 + ], + "type": "text", + "content": " bits" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 386, + 347, + 451, + 411 + ], + "blocks": [ + { + "bbox": [ + 386, + 347, + 451, + 411 + ], + "lines": [ + { + "bbox": [ + 386, + 347, + 451, + 411 + ], + "spans": [ + { + "bbox": [ + 386, + 347, + 451, + 411 + ], + "type": "image", + "image_path": "a9e64e9ac1fcf78ce2221a773ae859d7eff1aedcd0fd33877f0e2af639dca0e9.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 383, + 412, + 459, + 419 + ], + "lines": [ + { + "bbox": [ + 383, + 412, + 459, + 419 + ], + "spans": [ + { + "bbox": [ + 383, + 412, + 459, + 419 + ], + "type": "inline_equation", + "content": "y = 1800" + }, + { + "bbox": [ + 383, + 412, + 459, + 419 + ], + "type": "text", + "content": " bits " + }, + { + "bbox": [ + 383, + 412, + 459, + 419 + ], + "type": "inline_equation", + "content": "z = 1000" + }, + { + "bbox": [ + 383, + 412, + 459, + 419 + ], + "type": "text", + "content": " bits" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 403, + 423, + 438, + 431 + ], + "lines": [ + { + "bbox": [ + 403, + 423, + 438, + 431 + ], + "spans": [ + { + "bbox": [ + 403, + 423, + 438, + 431 + ], + "type": "text", + "content": "Factorized z" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 444, + 545, + 477 + ], + "lines": [ + { + "bbox": [ + 306, + 444, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 306, + 444, + 545, + 477 + ], + "type": "text", + "content": "Figure 5. Example of comparison between factorized and categorical hyper modules. The proposed categorical " + }, + { + "bbox": [ + 306, + 444, + 545, + 477 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 306, + 444, + 545, + 477 + ], + "type": "text", + "content": " encodes essential semantic and structural information with much less bits." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 468, + 347, + 531, + 411 + ], + "blocks": [ + { + "bbox": [ + 468, + 347, + 531, + 411 + ], + "lines": [ + { + "bbox": [ + 468, + 347, + 531, + 411 + ], + "spans": [ + { + "bbox": [ + 468, + 347, + 531, + 411 + ], + "type": "image", + "image_path": "7fc7a3ee1aa2ba6b560866f36b38ac79bd69924bab8af1fb7088576a00c6b53d.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 466, + 412, + 537, + 419 + ], + "lines": [ + { + "bbox": [ + 466, + 412, + 537, + 419 + ], + "spans": [ + { + "bbox": [ + 466, + 412, + 537, + 419 + ], + "type": "inline_equation", + "content": "y = 1640" + }, + { + "bbox": [ + 466, + 412, + 537, + 419 + ], + "type": "text", + "content": " bits " + }, + { + "bbox": [ + 466, + 412, + 537, + 419 + ], + "type": "inline_equation", + "content": "z = 160" + }, + { + "bbox": [ + 466, + 412, + 537, + 419 + ], + "type": "text", + "content": " bits" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 481, + 423, + 518, + 431 + ], + "lines": [ + { + "bbox": [ + 481, + 423, + 518, + 431 + ], + "spans": [ + { + "bbox": [ + 481, + 423, + 518, + 431 + ], + "type": "text", + "content": "Categorical z" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 498, + 545, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 631 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 631 + ], + "type": "text", + "content": "ever, these methods often lack a careful consideration of the correlation among the latents, resulting in an insufficient redundancy reduction and consequently a high bit cost. In GLC, we introduce a transform coding module to compress the latent, replacing the vector-quantization step for more effective reduction of latent redundancy. As shown in Figure 4a, the latents are transformed into code " + }, + { + "bbox": [ + 304, + 498, + 545, + 631 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 498, + 545, + 631 + ], + "type": "text", + "content": " using transformations " + }, + { + "bbox": [ + 304, + 498, + 545, + 631 + ], + "type": "inline_equation", + "content": "g_{a}" + }, + { + "bbox": [ + 304, + 498, + 545, + 631 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 498, + 545, + 631 + ], + "type": "inline_equation", + "content": "g_{s}" + }, + { + "bbox": [ + 304, + 498, + 545, + 631 + ], + "type": "text", + "content": " and then quantized to " + }, + { + "bbox": [ + 304, + 498, + 545, + 631 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 304, + 498, + 545, + 631 + ], + "type": "text", + "content": ". Entropy coding is applied to " + }, + { + "bbox": [ + 304, + 498, + 545, + 631 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 304, + 498, + 545, + 631 + ], + "type": "text", + "content": " based on a probability " + }, + { + "bbox": [ + 304, + 498, + 545, + 631 + ], + "type": "inline_equation", + "content": "p(\\hat{y})" + }, + { + "bbox": [ + 304, + 498, + 545, + 631 + ], + "type": "text", + "content": ", which is estimated by a categorical hyper module (Section 3.3.1) and a quadtree-partition-based spatial context module [38]." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 306, + 646, + 454, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 646, + 454, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 646, + 454, + 658 + ], + "type": "text", + "content": "3.3.1 Categorical Hyper Module" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "content": "Factorized hyper module [5] (Figure 4c) is commonly employed in recent image compression schemes. However, at ultra-low bitrate, we notice that the factorized " + }, + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "content": " tends to encode low-level information such as color and texture, incur" + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26091" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": "ring a high bit cost, as illustrated in Figure 5. To address it, we propose a categorical hyper module (Figure 4d), which utilizes a hyper codebook to store the basic semantic elements rather than the low-level information. This module comprises a hyper analysis transform " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "h_a" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": ", a hyper synthesis transform " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "h_s" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " and a hyper codebook " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "C_h" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": ". The transformations are formulated as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 71, + 162, + 287, + 175 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 162, + 287, + 175 + ], + "spans": [ + { + "bbox": [ + 71, + 162, + 287, + 175 + ], + "type": "interline_equation", + "content": "z = h _ {a} (y), \\hat {z} = V Q (z, C _ {h}), p r i o r _ {z} = h _ {s} (\\hat {z}) \\tag {2}", + "image_path": "ac65bc7f4c8eaeffc28e0e17abe2fe3ada805be9d4bd667b25282b3651607b9b.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 180, + 287, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 180, + 287, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 287, + 240 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 180, + 287, + 240 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 180, + 287, + 240 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 180, + 287, + 240 + ], + "type": "inline_equation", + "content": "\\hat{z}" + }, + { + "bbox": [ + 46, + 180, + 287, + 240 + ], + "type": "text", + "content": " denote hyper-codes. " + }, + { + "bbox": [ + 46, + 180, + 287, + 240 + ], + "type": "inline_equation", + "content": "VQ(\\cdot, C_h)" + }, + { + "bbox": [ + 46, + 180, + 287, + 240 + ], + "type": "text", + "content": " represents vector-quantization by nearest lookup in " + }, + { + "bbox": [ + 46, + 180, + 287, + 240 + ], + "type": "inline_equation", + "content": "C_h" + }, + { + "bbox": [ + 46, + 180, + 287, + 240 + ], + "type": "text", + "content": ". As shown in the right of Figure 5, the categorical " + }, + { + "bbox": [ + 46, + 180, + 287, + 240 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 180, + 287, + 240 + ], + "type": "text", + "content": " is more inclined to capture high-level semantic information, which can be encoded with significantly fewer bits." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 255, + 211, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 255, + 211, + 266 + ], + "spans": [ + { + "bbox": [ + 47, + 255, + 211, + 266 + ], + "type": "text", + "content": "3.3.2 Rate-Variable Transformation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 273, + 289, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 273, + 289, + 406 + ], + "spans": [ + { + "bbox": [ + 46, + 273, + 289, + 406 + ], + "type": "text", + "content": "A notable advantage of transform coding over VQ-indices-map coding is its capability for rate-variable compression, which is a core functionality for a practical image codec. Indices-map coding is limited since the codebook can only model one specified distribution, but different rates naturally need different distributions. In contrast, transform coding converts latent into a unified Gaussian distribution, and variable-rate can be achieved by variable parameters (e.g., means and scales) of Gaussian. In GLC, we follow DCVC series [36-39, 47, 49] to incorporate a scalar " + }, + { + "bbox": [ + 46, + 273, + 289, + 406 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 46, + 273, + 289, + 406 + ], + "type": "text", + "content": " in the transform coding to achieve rate-variable compression." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 415, + 169, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 415, + 169, + 428 + ], + "spans": [ + { + "bbox": [ + 47, + 415, + 169, + 428 + ], + "type": "text", + "content": "4. Progressive Training" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 434, + 287, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 287, + 531 + ], + "type": "text", + "content": "As depicted in the right of Figure 3, we adopt a three-stage progressive training manner to fully leverage the potential of GLC. We initially learn a human-perception-aligned latent space to guarantee the perceptual quality, subsequently learn to perform transform coding on this latent space to achieve low bitrate, and finally fine-tune the entire network for superior compression performance. At each stage, distinct loss functions are adopted to guide different modules." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 536, + 224, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 224, + 549 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 224, + 549 + ], + "type": "text", + "content": "4.1. Stage I: Auto-Encoder Learning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 555, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 555, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 555, + 287, + 640 + ], + "type": "text", + "content": "To obtain a human-perception-aligned latent space for compression, we begin with training a generative VQ-VAE as the initialization of " + }, + { + "bbox": [ + 46, + 555, + 287, + 640 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 46, + 555, + 287, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 555, + 287, + 640 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 555, + 287, + 640 + ], + "type": "text", + "content": ". To ensure the sparsity of the latent space, an auxiliary codebook " + }, + { + "bbox": [ + 46, + 555, + 287, + 640 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 555, + 287, + 640 + ], + "type": "text", + "content": " is employed to perform nearest vector-quantization, transforming " + }, + { + "bbox": [ + 46, + 555, + 287, + 640 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 555, + 287, + 640 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 555, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\widetilde{l}" + }, + { + "bbox": [ + 46, + 555, + 287, + 640 + ], + "type": "text", + "content": ". The supervision comprises reconstruction loss, perceptual loss [29], adversarial loss [18] and codebook loss [53]:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 85, + 644, + 287, + 672 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 644, + 287, + 672 + ], + "spans": [ + { + "bbox": [ + 85, + 644, + 287, + 672 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {S t a g e I}} = \\left| \\left| x - \\hat {x} \\right| \\right| + \\mathcal {L} _ {p e r} (x, \\hat {x}) \\tag {3} \\\\ + \\lambda_ {a d v} \\cdot \\mathcal {L} _ {a d v} (x, \\hat {x}) + \\mathcal {L} _ {\\text {c o d e b o o k}} \\\\ \\end{array}", + "image_path": "a779d9125adcddc82f9b8f0696f0f23cee594ac1db6e09ccb07a36824c79e8ad.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{per}" + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": " corresponds to the LPIPS loss calculated using VGG [50] extracted features. " + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{adv}" + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": " is the adaptive PatchGAN adversarial loss [17] with a weight of " + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{adv} = 0.8" + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 310, + 73, + 427, + 224 + ], + "blocks": [ + { + "bbox": [ + 310, + 73, + 427, + 224 + ], + "lines": [ + { + "bbox": [ + 310, + 73, + 427, + 224 + ], + "spans": [ + { + "bbox": [ + 310, + 73, + 427, + 224 + ], + "type": "image", + "image_path": "686772d25e674bd1209af91b09224489b63f2e1a1818e78f9f6dcf8bb2ca6c84.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 236, + 533, + 247 + ], + "lines": [ + { + "bbox": [ + 318, + 236, + 533, + 247 + ], + "spans": [ + { + "bbox": [ + 318, + 236, + 533, + 247 + ], + "type": "text", + "content": "Figure 6. Illustration of code-prediction-based supervision." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 430, + 73, + 541, + 224 + ], + "blocks": [ + { + "bbox": [ + 430, + 73, + 541, + 224 + ], + "lines": [ + { + "bbox": [ + 430, + 73, + 541, + 224 + ], + "spans": [ + { + "bbox": [ + 430, + 73, + 541, + 224 + ], + "type": "image", + "image_path": "42fa88018f2af32acdaf021d1d840493356db5e66eec6dbae839dee3a623d4b3.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 268, + 451, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 268, + 451, + 279 + ], + "spans": [ + { + "bbox": [ + 306, + 268, + 451, + 279 + ], + "type": "text", + "content": "The codebook loss is formulated as" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 339, + 287, + 545, + 302 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 287, + 545, + 302 + ], + "spans": [ + { + "bbox": [ + 339, + 287, + 545, + 302 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o d e b o o k}} = \\left| \\left| \\operatorname {s g} (l) - \\tilde {l} \\right| \\right| + \\beta \\cdot \\left| \\left| \\operatorname {s g} (\\tilde {l}) - l \\right| \\right| \\tag {4}", + "image_path": "7bccb15f86eb5604225dbe887b6bb88efd354ef7c09fb1786cab49f06ebaf9ae.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 311, + 546, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 311, + 546, + 335 + ], + "spans": [ + { + "bbox": [ + 305, + 311, + 546, + 335 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 311, + 546, + 335 + ], + "type": "inline_equation", + "content": "\\mathrm{sg}(\\cdot)" + }, + { + "bbox": [ + 305, + 311, + 546, + 335 + ], + "type": "text", + "content": " denotes the stop-gradient operator and " + }, + { + "bbox": [ + 305, + 311, + 546, + 335 + ], + "type": "inline_equation", + "content": "\\beta = 0.25" + }, + { + "bbox": [ + 305, + 311, + 546, + 335 + ], + "type": "text", + "content": " controls the update rates of the " + }, + { + "bbox": [ + 305, + 311, + 546, + 335 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 305, + 311, + 546, + 335 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 311, + 546, + 335 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 305, + 311, + 546, + 335 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 342, + 508, + 356 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 342, + 508, + 356 + ], + "spans": [ + { + "bbox": [ + 305, + 342, + 508, + 356 + ], + "type": "text", + "content": "4.2. Stage II: Transform Coding Learning" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 361, + 545, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 545, + 469 + ], + "type": "text", + "content": "Given the trained latent space, we further learn to perform transform coding to achieve low-bitrate latent compression, while fixing the auto-encoder " + }, + { + "bbox": [ + 304, + 361, + 545, + 469 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 361, + 545, + 469 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 361, + 545, + 469 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 361, + 545, + 469 + ], + "type": "text", + "content": ". We introduce an auxiliary code predictor " + }, + { + "bbox": [ + 304, + 361, + 545, + 469 + ], + "type": "inline_equation", + "content": "CP" + }, + { + "bbox": [ + 304, + 361, + 545, + 469 + ], + "type": "text", + "content": " to enhance the semantic consistency by necessitating the latent to possess the capability to predict the correct VQ-indices. As shown in Figure 6, we encode " + }, + { + "bbox": [ + 304, + 361, + 545, + 469 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 361, + 545, + 469 + ], + "type": "text", + "content": " into VQ-indices by " + }, + { + "bbox": [ + 304, + 361, + 545, + 469 + ], + "type": "inline_equation", + "content": "M_l = VQ(l,C)" + }, + { + "bbox": [ + 304, + 361, + 545, + 469 + ], + "type": "text", + "content": " and subsequently predict these indices by " + }, + { + "bbox": [ + 304, + 361, + 545, + 469 + ], + "type": "inline_equation", + "content": "\\hat{M}_{\\hat{l}} = CP(\\hat{l})" + }, + { + "bbox": [ + 304, + 361, + 545, + 469 + ], + "type": "text", + "content": ". So the code-prediction-based loss can be formulated by" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 339, + 477, + 545, + 493 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 477, + 545, + 493 + ], + "spans": [ + { + "bbox": [ + 339, + 477, + 545, + 493 + ], + "type": "interline_equation", + "content": "\\mathcal {D} _ {c o d e} (l, \\hat {l}) = \\alpha \\cdot C E (M _ {l}, \\hat {M _ {l}}) + | | l - \\hat {l} | | _ {2} ^ {2} \\qquad (5)", + "image_path": "a8a19b56cb2b0608098ff7231c08886f8b55889050d759dc8981c85aa173607d.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 501, + 546, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 501, + 546, + 536 + ], + "spans": [ + { + "bbox": [ + 305, + 501, + 546, + 536 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 501, + 546, + 536 + ], + "type": "inline_equation", + "content": "CE" + }, + { + "bbox": [ + 305, + 501, + 546, + 536 + ], + "type": "text", + "content": " denotes the cross entropy loss and we set " + }, + { + "bbox": [ + 305, + 501, + 546, + 536 + ], + "type": "inline_equation", + "content": "\\alpha = 0.5" + }, + { + "bbox": [ + 305, + 501, + 546, + 536 + ], + "type": "text", + "content": " by default. Then the transform coding module can be supervised by the rate-distortion trade-off" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 341, + 545, + 545, + 560 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 545, + 545, + 560 + ], + "spans": [ + { + "bbox": [ + 341, + 545, + 545, + 560 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {S t a g e I I}} = \\mathbf {E} _ {x \\sim p _ {X}} [ \\mathcal {R} (\\hat {y}) + \\lambda \\cdot \\mathcal {D} _ {\\text {c o d e}} (l, \\hat {l}) ] \\tag {6}", + "image_path": "50781dad7a3c741b8cda41496269a28d6655d6991b37d4a929a2c86828d336ce.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 567, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 567, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 567, + 545, + 628 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 567, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 304, + 567, + 545, + 628 + ], + "type": "text", + "content": " is the estimated rate and " + }, + { + "bbox": [ + 304, + 567, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 567, + 545, + 628 + ], + "type": "text", + "content": " is used to control the trade-off. Note that a codebook loss (as formulated in Equation 4) is required to train the hyper codebook " + }, + { + "bbox": [ + 304, + 567, + 545, + 628 + ], + "type": "inline_equation", + "content": "C_h" + }, + { + "bbox": [ + 304, + 567, + 545, + 628 + ], + "type": "text", + "content": " in the categorical hyper module. We omit it from the loss functions of both stage II and stage III for the sake of conciseness." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 635, + 447, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 447, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 447, + 647 + ], + "type": "text", + "content": "4.3. Stage III: Joint Training" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 654, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 714 + ], + "type": "text", + "content": "Finally, we fine-tune the entire network with the pixel space supervision to achieve better compression performance. As shown in Figure 6, we extend the code-prediction-based latent supervision into the pixel space. Specifically, we utilize the encoder " + }, + { + "bbox": [ + 304, + 654, + 547, + 714 + ], + "type": "inline_equation", + "content": "E_{VQ}" + }, + { + "bbox": [ + 304, + 654, + 547, + 714 + ], + "type": "text", + "content": " trained from stage I to encode " + }, + { + "bbox": [ + 304, + 654, + 547, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 654, + 547, + 714 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26092" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 289, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 289, + 145 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 289, + 145 + ], + "type": "inline_equation", + "content": "\\hat{x}" + }, + { + "bbox": [ + 47, + 71, + 289, + 145 + ], + "type": "text", + "content": " into latent space by " + }, + { + "bbox": [ + 47, + 71, + 289, + 145 + ], + "type": "inline_equation", + "content": "\\hat{l}^p = E_{VQ}(\\hat{x})" + }, + { + "bbox": [ + 47, + 71, + 289, + 145 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 71, + 289, + 145 + ], + "type": "inline_equation", + "content": "l^p = E_{VQ}(x)" + }, + { + "bbox": [ + 47, + 71, + 289, + 145 + ], + "type": "text", + "content": ", so the code-prediction-based pixel loss can be calculated by " + }, + { + "bbox": [ + 47, + 71, + 289, + 145 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{code}(l^p,\\hat{l}^p)" + }, + { + "bbox": [ + 47, + 71, + 289, + 145 + ], + "type": "text", + "content": " in the same formulation with Equation 5. Here we use " + }, + { + "bbox": [ + 47, + 71, + 289, + 145 + ], + "type": "inline_equation", + "content": "E_{VQ}" + }, + { + "bbox": [ + 47, + 71, + 289, + 145 + ], + "type": "text", + "content": " since it can map the input data to a compatible latent space with the codebook " + }, + { + "bbox": [ + 47, + 71, + 289, + 145 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 71, + 289, + 145 + ], + "type": "text", + "content": " for code prediction. The overall pixel supervision is defined as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 151, + 286, + 182 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 151, + 286, + 182 + ], + "spans": [ + { + "bbox": [ + 55, + 151, + 286, + 182 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {D} _ {\\text {S t a g e I I I}} = | | x - \\hat {x} | | + \\mathcal {L} _ {p e r} (x, \\hat {x}) \\\\ + \\lambda_ {a d v} \\cdot \\mathcal {L} _ {a d v} (x, \\hat {x}) + \\lambda_ {c o d e} \\cdot \\mathcal {D} _ {c o d e} \\left(l ^ {p}, \\hat {l} ^ {p}\\right) \\tag {7} \\\\ \\end{array}", + "image_path": "b0d5c40829afdd09919844bc9e4305fa6ebd3124a7ea343ee877b429fc185888.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 189, + 287, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 189, + 287, + 213 + ], + "spans": [ + { + "bbox": [ + 47, + 189, + 287, + 213 + ], + "type": "text", + "content": "where we set " + }, + { + "bbox": [ + 47, + 189, + 287, + 213 + ], + "type": "inline_equation", + "content": "\\lambda_{code} = 0.05" + }, + { + "bbox": [ + 47, + 189, + 287, + 213 + ], + "type": "text", + "content": " by default. The rate-distortion trade-off supervision is:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 85, + 220, + 287, + 235 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 220, + 287, + 235 + ], + "spans": [ + { + "bbox": [ + 85, + 220, + 287, + 235 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {S t a g e I I I}} = \\mathbf {E} _ {x \\sim p _ {X}} \\left[ \\mathcal {R} (\\hat {y}) + \\lambda \\cdot \\mathcal {D} _ {\\text {S t a g e I I I}} \\right] \\tag {8}", + "image_path": "cee059710a6aef56c96c0645753e9e234e4ebf2c608df786d55a88586345581e.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 240, + 266, + 252 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 240, + 266, + 252 + ], + "spans": [ + { + "bbox": [ + 47, + 240, + 266, + 252 + ], + "type": "text", + "content": "4.4. Discussion of Code-Prediction-Based Loss" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 258, + 287, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 287, + 342 + ], + "type": "text", + "content": "Code prediction transformers [27, 28, 60] have demonstrated their effectiveness in high-quality image reconstruction. They typically input the predicted latent directly into the decoder for reconstruction. Different from these methods, in GLC, we suggest to consider code prediction solely as an auxiliary supervision during training, but not used in the inference process of the compression pipeline." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 342, + 288, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 288, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 288, + 521 + ], + "type": "text", + "content": "This design is based on an observation: if a code prediction module is introduced before the decoder, the fine-tuning process in stage III cannot enhance compression performance further. It appears that the codebook becomes a performance bottleneck, restricting the decoder to receiving only the vector-quantized latent as input, which has already been well-trained in stage I. In GLC, by utilizing code prediction solely as auxiliary supervision during training, we eliminate this bottleneck, allowing the decoder to receive more flexible input for additional fine-tuning. We find that this code-prediction-based supervision effectively enhances the semantic consistency of the reconstructions, as shown in Figure 7. By necessitating the latent to possess the capability to predict the code index, the latent contains more semantic information, such as gestures and attributes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 532, + 128, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 532, + 128, + 544 + ], + "spans": [ + { + "bbox": [ + 47, + 532, + 128, + 544 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 552, + 180, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 180, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 180, + 563 + ], + "type": "text", + "content": "5.1. Implementation Details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 570, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 677 + ], + "type": "text", + "content": "Training details. We train GLC for both natural image compression and facial image compression. For natural images, we conduct stage I on ImageNet training set [14], stage II and III on OpenImage test set [33], using randomly cropped " + }, + { + "bbox": [ + 46, + 570, + 287, + 677 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 570, + 287, + 677 + ], + "type": "text", + "content": " patches. For facial images, GLC is trained on FFHQ dataset [31] for all stages with a resolution of " + }, + { + "bbox": [ + 46, + 570, + 287, + 677 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 46, + 570, + 287, + 677 + ], + "type": "text", + "content": ". Both models are optimized by AdamW [42] with a batch size of 8. For each batch, we train the model with different " + }, + { + "bbox": [ + 46, + 570, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 570, + 287, + 677 + ], + "type": "text", + "content": " to achieve rate-variable compression." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Evaluation dataset. We evaluate GLC on CLIC 2020 test set [51] with original resolution for natural image compression, and evaluate on CelebAHQ [30] with a resolution of" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 311, + 78, + 357, + 124 + ], + "blocks": [ + { + "bbox": [ + 311, + 78, + 357, + 124 + ], + "lines": [ + { + "bbox": [ + 311, + 78, + 357, + 124 + ], + "spans": [ + { + "bbox": [ + 311, + 78, + 357, + 124 + ], + "type": "image", + "image_path": "00d8dc75728ecd5a4dd2019b3c3782075ae5c1ad855f8e025c6ee047839a158b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 359, + 78, + 416, + 124 + ], + "blocks": [ + { + "bbox": [ + 359, + 78, + 416, + 124 + ], + "lines": [ + { + "bbox": [ + 359, + 78, + 416, + 124 + ], + "spans": [ + { + "bbox": [ + 359, + 78, + 416, + 124 + ], + "type": "image", + "image_path": "f326b6496b3a638e84c51a36711dc49e29f4175d70904798e09cf73328ffe87d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 419, + 78, + 477, + 124 + ], + "blocks": [ + { + "bbox": [ + 419, + 78, + 477, + 124 + ], + "lines": [ + { + "bbox": [ + 419, + 78, + 477, + 124 + ], + "spans": [ + { + "bbox": [ + 419, + 78, + 477, + 124 + ], + "type": "image", + "image_path": "b15f4e3adcc6340552b7de0bbadba090e319570c0343c0c514b089b4f12acdc9.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 481, + 78, + 539, + 124 + ], + "blocks": [ + { + "bbox": [ + 481, + 78, + 539, + 124 + ], + "lines": [ + { + "bbox": [ + 481, + 78, + 539, + 124 + ], + "spans": [ + { + "bbox": [ + 481, + 78, + 539, + 124 + ], + "type": "image", + "image_path": "e0ee7f99d6570edc677c69dadb82f476e243090f014f0a737286c329e1f5818c.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 311, + 128, + 356, + 174 + ], + "blocks": [ + { + "bbox": [ + 311, + 128, + 356, + 174 + ], + "lines": [ + { + "bbox": [ + 311, + 128, + 356, + 174 + ], + "spans": [ + { + "bbox": [ + 311, + 128, + 356, + 174 + ], + "type": "image", + "image_path": "40998eeda359794e5baf91901d8a42342f6fc60c723d22a53ecccb9640f2e2c4.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 364, + 178, + 410, + 190 + ], + "lines": [ + { + "bbox": [ + 364, + 178, + 410, + 190 + ], + "spans": [ + { + "bbox": [ + 364, + 178, + 410, + 190 + ], + "type": "text", + "content": "Original Image" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 358, + 128, + 416, + 174 + ], + "blocks": [ + { + "bbox": [ + 358, + 128, + 416, + 174 + ], + "lines": [ + { + "bbox": [ + 358, + 128, + 416, + 174 + ], + "spans": [ + { + "bbox": [ + 358, + 128, + 416, + 174 + ], + "type": "image", + "image_path": "0f182d6dbb3af5bdeff9923461cd2d699eb95fb21062e1e7cfc1bed5b261c68d.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 423, + 178, + 473, + 195 + ], + "lines": [ + { + "bbox": [ + 423, + 178, + 473, + 195 + ], + "spans": [ + { + "bbox": [ + 423, + 178, + 473, + 195 + ], + "type": "text", + "content": "w/o code-pred-based loss" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 419, + 128, + 477, + 174 + ], + "blocks": [ + { + "bbox": [ + 419, + 128, + 477, + 174 + ], + "lines": [ + { + "bbox": [ + 419, + 128, + 477, + 174 + ], + "spans": [ + { + "bbox": [ + 419, + 128, + 477, + 174 + ], + "type": "image", + "image_path": "7229844ecc1d6331c262e46852d43a9bbe6dbb292fd7dd3ba4e3466f0ceb9de9.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 486, + 178, + 533, + 195 + ], + "lines": [ + { + "bbox": [ + 486, + 178, + 533, + 195 + ], + "spans": [ + { + "bbox": [ + 486, + 178, + 533, + 195 + ], + "type": "text", + "content": "w/ code-pred-based loss" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 481, + 128, + 539, + 174 + ], + "blocks": [ + { + "bbox": [ + 481, + 128, + 539, + 174 + ], + "lines": [ + { + "bbox": [ + 481, + 128, + 539, + 174 + ], + "spans": [ + { + "bbox": [ + 481, + 128, + 539, + 174 + ], + "type": "image", + "image_path": "5e12b06237e39622632a63022ecc04fa662a1ee2990caddf564d374b35cc4a26.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 206, + 545, + 251 + ], + "lines": [ + { + "bbox": [ + 304, + 206, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 206, + 545, + 251 + ], + "type": "text", + "content": "Figure 7. An example of using code-prediction-based supervision after stage II. Code prediction loss enhances the semantic consistency of the compressed latent, such as the color of eye shadow and the opening of the mouth." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 267, + 545, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 267, + 545, + 302 + ], + "spans": [ + { + "bbox": [ + 304, + 267, + 545, + 302 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 267, + 545, + 302 + ], + "type": "text", + "content": " for facial image. We also show the results on Kodak [32], DIV2K [1] and MS-COCO 30K [40] in the supplementary material." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 303, + 546, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 303, + 546, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 303, + 546, + 434 + ], + "type": "text", + "content": "Evaluation metrics. We measure bit-stream size by bits per pixel (bpp), and measure visual quality by reference perceptual metrics LPIPS [29] and DISTS [15] and no-reference perceptual metrics FID [23] and KID [6]. We also provide PSNR and MS-SSIM [56] results in the supplementary material for completeness. Nevertheless, it is worth note that these pixel-level distortion metrics PSNR, MS-SSIM and LPIPS have strong limitations when evaluating image compression at ultra-low bitrate, which is also mentioned in other works [15, 35]. We provide a clearer demonstration of it in the supplementary material." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 434, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 545, + 567 + ], + "type": "text", + "content": "Baseline methods. We compare with traditional codec VVC [54], neural codec TCM [41], EVC [20], and generative codec FCC [26], Text+Sketch [35], HiFiC [44], MSILLM [46]. As some methods do not release models for ultra-low bitrate, we either retrain or fine-tune their models to suit such low bitrate. Text+Sketch is not evaluated on CLIC since it does not support compression in high resolution. In addition, we also compare with recent works HFD [24] and PerCo [8] in the supplementary material. For facial compression, we fine-tune EVC, TCM, HiFiC and MSILLM using FFHQ dataset for a fair comparison." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 574, + 390, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 390, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 390, + 586 + ], + "type": "text", + "content": "5.2. Main Results" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "Figure 8 shows the performance of the proposal and compared methods at ultra-low bitrate. On CLIC 2020, GLC demonstrates superiority in terms of DISTS, FID and KID than other methods. Specifically, GLC saves about " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "45\\%" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " bits compared to previous SOTA method MS-ILLM while maintaining an equivalent FID. When comparing the pixel-level metric LPIPS, GLC also achieves comparable performance with high-fidelity generative codec such as HiFiC and MS-ILLM. On CelebAHQ, GLC outperforms all other methods across all metrics by a large margin." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26093" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 121, + 74, + 478, + 95 + ], + "blocks": [ + { + "bbox": [ + 121, + 74, + 478, + 95 + ], + "lines": [ + { + "bbox": [ + 121, + 74, + 478, + 95 + ], + "spans": [ + { + "bbox": [ + 121, + 74, + 478, + 95 + ], + "type": "image", + "image_path": "6dc3f1d89dc3d53e894365ea39211e3f98302e0e20bab5ad21895a9677a918dd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 52, + 102, + 171, + 220 + ], + "blocks": [ + { + "bbox": [ + 52, + 102, + 171, + 220 + ], + "lines": [ + { + "bbox": [ + 52, + 102, + 171, + 220 + ], + "spans": [ + { + "bbox": [ + 52, + 102, + 171, + 220 + ], + "type": "image", + "image_path": "f7078527f2c23414802e3e37ab8349a8747802b61e3123a6f806040ce85bbced.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 173, + 102, + 294, + 221 + ], + "blocks": [ + { + "bbox": [ + 173, + 102, + 294, + 221 + ], + "lines": [ + { + "bbox": [ + 173, + 102, + 294, + 221 + ], + "spans": [ + { + "bbox": [ + 173, + 102, + 294, + 221 + ], + "type": "image", + "image_path": "403dea6682aea885d23c00031f4f1bb80a92592acace05378a3cd688554aca01.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 296, + 102, + 419, + 221 + ], + "blocks": [ + { + "bbox": [ + 296, + 102, + 419, + 221 + ], + "lines": [ + { + "bbox": [ + 296, + 102, + 419, + 221 + ], + "spans": [ + { + "bbox": [ + 296, + 102, + 419, + 221 + ], + "type": "image", + "image_path": "80672e66c6094bdfa9f4a44624598d1dc510a0050c4bc129ae9849a8d1c6afa8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 420, + 102, + 544, + 221 + ], + "blocks": [ + { + "bbox": [ + 420, + 102, + 544, + 221 + ], + "lines": [ + { + "bbox": [ + 420, + 102, + 544, + 221 + ], + "spans": [ + { + "bbox": [ + 420, + 102, + 544, + 221 + ], + "type": "image", + "image_path": "c8a5fafb393863e956b79750a9026a773aa0a4131b936da7a97fa256b99c4958.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 52, + 223, + 167, + 341 + ], + "blocks": [ + { + "bbox": [ + 52, + 223, + 167, + 341 + ], + "lines": [ + { + "bbox": [ + 52, + 223, + 167, + 341 + ], + "spans": [ + { + "bbox": [ + 52, + 223, + 167, + 341 + ], + "type": "image", + "image_path": "8c80faa331ad8711fbaed3614179b17c2c042137be2acfc8351cd5091af22d6b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 163, + 346, + 429, + 357 + ], + "lines": [ + { + "bbox": [ + 163, + 346, + 429, + 357 + ], + "spans": [ + { + "bbox": [ + 163, + 346, + 429, + 357 + ], + "type": "text", + "content": "Figure 8. Comparison of methods on CLIC 2020 test set and CelebAHQ." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 173, + 223, + 291, + 341 + ], + "blocks": [ + { + "bbox": [ + 173, + 223, + 291, + 341 + ], + "lines": [ + { + "bbox": [ + 173, + 223, + 291, + 341 + ], + "spans": [ + { + "bbox": [ + 173, + 223, + 291, + 341 + ], + "type": "image", + "image_path": "ae20a0313821ff26b38593ad0d02e89f4c5a0d7b7deed04d5daf2d4ad3a6c31a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 296, + 223, + 416, + 341 + ], + "blocks": [ + { + "bbox": [ + 296, + 223, + 416, + 341 + ], + "lines": [ + { + "bbox": [ + 296, + 223, + 416, + 341 + ], + "spans": [ + { + "bbox": [ + 296, + 223, + 416, + 341 + ], + "type": "image", + "image_path": "e80b271631dc5deb6d58ce9f702be0191ddf8330fe45ca7bf64b58494f86682e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 422, + 223, + 539, + 341 + ], + "blocks": [ + { + "bbox": [ + 422, + 223, + 539, + 341 + ], + "lines": [ + { + "bbox": [ + 422, + 223, + 539, + 341 + ], + "spans": [ + { + "bbox": [ + 422, + 223, + 539, + 341 + ], + "type": "image", + "image_path": "577b515a6b710eda02570e14559ef65a960210e3cab3d0fb6384a201fbd8b4d7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 55, + 367, + 135, + 445 + ], + "blocks": [ + { + "bbox": [ + 55, + 367, + 135, + 445 + ], + "lines": [ + { + "bbox": [ + 55, + 367, + 135, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 367, + 135, + 445 + ], + "type": "image", + "image_path": "b37e9dfeebaa98abf9ab3dc3c1c60019c490ede60ff70276c8716ee6cbd26de1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 447, + 107, + 454 + ], + "lines": [ + { + "bbox": [ + 83, + 447, + 107, + 454 + ], + "spans": [ + { + "bbox": [ + 83, + 447, + 107, + 454 + ], + "type": "text", + "content": "Original" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 55, + 463, + 135, + 540 + ], + "blocks": [ + { + "bbox": [ + 83, + 454, + 106, + 463 + ], + "lines": [ + { + "bbox": [ + 83, + 454, + 106, + 463 + ], + "spans": [ + { + "bbox": [ + 83, + 454, + 106, + 463 + ], + "type": "text", + "content": "[Kodak]" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 463, + 135, + 540 + ], + "lines": [ + { + "bbox": [ + 55, + 463, + 135, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 135, + 540 + ], + "type": "image", + "image_path": "da3c2f2a90f317114dd0f66db38db904d79ecf5bf865f2ab23ea0c83e1415466.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 542, + 107, + 548 + ], + "lines": [ + { + "bbox": [ + 83, + 542, + 107, + 548 + ], + "spans": [ + { + "bbox": [ + 83, + 542, + 107, + 548 + ], + "type": "text", + "content": "Original" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 78, + 549, + 111, + 558 + ], + "lines": [ + { + "bbox": [ + 78, + 549, + 111, + 558 + ], + "spans": [ + { + "bbox": [ + 78, + 549, + 111, + 558 + ], + "type": "text", + "content": "[CelebAHQ]" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 61, + 566, + 531, + 578 + ], + "lines": [ + { + "bbox": [ + 61, + 566, + 531, + 578 + ], + "spans": [ + { + "bbox": [ + 61, + 566, + 531, + 578 + ], + "type": "text", + "content": "Figure 9. Qualitative examples of different methods on Kodak and CelebAHQ. More comparisons are in supplementary materials." + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 136, + 368, + 216, + 445 + ], + "blocks": [ + { + "bbox": [ + 136, + 368, + 216, + 445 + ], + "lines": [ + { + "bbox": [ + 136, + 368, + 216, + 445 + ], + "spans": [ + { + "bbox": [ + 136, + 368, + 216, + 445 + ], + "type": "image", + "image_path": "3630dd87ce80b6f61463bac3f632d77d788ea541db3329aaefccb9de0dfef2d3.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 158, + 447, + 192, + 454 + ], + "lines": [ + { + "bbox": [ + 158, + 447, + 192, + 454 + ], + "spans": [ + { + "bbox": [ + 158, + 447, + 192, + 454 + ], + "type": "text", + "content": "Text+Sketch" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 136, + 463, + 216, + 540 + ], + "blocks": [ + { + "bbox": [ + 159, + 454, + 192, + 463 + ], + "lines": [ + { + "bbox": [ + 159, + 454, + 192, + 463 + ], + "spans": [ + { + "bbox": [ + 159, + 454, + 192, + 463 + ], + "type": "text", + "content": "0.0320 bpp" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 136, + 463, + 216, + 540 + ], + "lines": [ + { + "bbox": [ + 136, + 463, + 216, + 540 + ], + "spans": [ + { + "bbox": [ + 136, + 463, + 216, + 540 + ], + "type": "image", + "image_path": "0399d03b2df8f80592f27b944e28fab3433bdc46373cb409c4414e6febeb597c.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 169, + 542, + 184, + 549 + ], + "lines": [ + { + "bbox": [ + 169, + 542, + 184, + 549 + ], + "spans": [ + { + "bbox": [ + 169, + 542, + 184, + 549 + ], + "type": "text", + "content": "Wc" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 159, + 549, + 192, + 558 + ], + "lines": [ + { + "bbox": [ + 159, + 549, + 192, + 558 + ], + "spans": [ + { + "bbox": [ + 159, + 549, + 192, + 558 + ], + "type": "text", + "content": "0.0115bpp" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 216, + 368, + 295, + 445 + ], + "blocks": [ + { + "bbox": [ + 216, + 368, + 295, + 445 + ], + "lines": [ + { + "bbox": [ + 216, + 368, + 295, + 445 + ], + "spans": [ + { + "bbox": [ + 216, + 368, + 295, + 445 + ], + "type": "image", + "image_path": "ef1f42cbf9449f47bda56a2527d4b43d21518871d5d715761d0fd4dca403364c.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 249, + 447, + 264, + 454 + ], + "lines": [ + { + "bbox": [ + 249, + 447, + 264, + 454 + ], + "spans": [ + { + "bbox": [ + 249, + 447, + 264, + 454 + ], + "type": "text", + "content": "TCM" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 216, + 462, + 295, + 540 + ], + "blocks": [ + { + "bbox": [ + 239, + 454, + 272, + 462 + ], + "lines": [ + { + "bbox": [ + 239, + 454, + 272, + 462 + ], + "spans": [ + { + "bbox": [ + 239, + 454, + 272, + 462 + ], + "type": "text", + "content": "0.0787 bpp" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 216, + 462, + 295, + 540 + ], + "lines": [ + { + "bbox": [ + 216, + 462, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 216, + 462, + 295, + 540 + ], + "type": "image", + "image_path": "aa91f8e1cd8218b10f2d83d40672a7e7cd4d543c6fdfd73c3ef7aec3190c5c69.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 249, + 542, + 263, + 549 + ], + "lines": [ + { + "bbox": [ + 249, + 542, + 263, + 549 + ], + "spans": [ + { + "bbox": [ + 249, + 542, + 263, + 549 + ], + "type": "text", + "content": "TCM" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 239, + 549, + 272, + 557 + ], + "lines": [ + { + "bbox": [ + 239, + 549, + 272, + 557 + ], + "spans": [ + { + "bbox": [ + 239, + 549, + 272, + 557 + ], + "type": "text", + "content": "0.0094 bpp" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 297, + 368, + 376, + 445 + ], + "blocks": [ + { + "bbox": [ + 297, + 368, + 376, + 445 + ], + "lines": [ + { + "bbox": [ + 297, + 368, + 376, + 445 + ], + "spans": [ + { + "bbox": [ + 297, + 368, + 376, + 445 + ], + "type": "image", + "image_path": "e5c92f3959b1da52eac29b5ad6773ab19e61782fd90e7d1a5f1a69d78738c18f.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 447, + 345, + 454 + ], + "lines": [ + { + "bbox": [ + 329, + 447, + 345, + 454 + ], + "spans": [ + { + "bbox": [ + 329, + 447, + 345, + 454 + ], + "type": "text", + "content": "HiFiC" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 298, + 462, + 376, + 540 + ], + "blocks": [ + { + "bbox": [ + 320, + 454, + 353, + 462 + ], + "lines": [ + { + "bbox": [ + 320, + 454, + 353, + 462 + ], + "spans": [ + { + "bbox": [ + 320, + 454, + 353, + 462 + ], + "type": "text", + "content": "0.0426 bpp" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 298, + 462, + 376, + 540 + ], + "lines": [ + { + "bbox": [ + 298, + 462, + 376, + 540 + ], + "spans": [ + { + "bbox": [ + 298, + 462, + 376, + 540 + ], + "type": "image", + "image_path": "3b518627e47e28da9b24fa31891946af601b5acf1d0e198d78b0182385b51843.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 542, + 345, + 548 + ], + "lines": [ + { + "bbox": [ + 329, + 542, + 345, + 548 + ], + "spans": [ + { + "bbox": [ + 329, + 542, + 345, + 548 + ], + "type": "text", + "content": "HiFiC" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 320, + 549, + 353, + 557 + ], + "lines": [ + { + "bbox": [ + 320, + 549, + 353, + 557 + ], + "spans": [ + { + "bbox": [ + 320, + 549, + 353, + 557 + ], + "type": "text", + "content": "0.0085 bpp" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 378, + 368, + 457, + 445 + ], + "blocks": [ + { + "bbox": [ + 378, + 368, + 457, + 445 + ], + "lines": [ + { + "bbox": [ + 378, + 368, + 457, + 445 + ], + "spans": [ + { + "bbox": [ + 378, + 368, + 457, + 445 + ], + "type": "image", + "image_path": "ac1626058217735b60a17c04045e006124a996d73fb066643cb95dfa6a8125af.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 404, + 447, + 431, + 454 + ], + "lines": [ + { + "bbox": [ + 404, + 447, + 431, + 454 + ], + "spans": [ + { + "bbox": [ + 404, + 447, + 431, + 454 + ], + "type": "text", + "content": "MS-ILLM" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 378, + 462, + 457, + 540 + ], + "blocks": [ + { + "bbox": [ + 402, + 454, + 434, + 462 + ], + "lines": [ + { + "bbox": [ + 402, + 454, + 434, + 462 + ], + "spans": [ + { + "bbox": [ + 402, + 454, + 434, + 462 + ], + "type": "text", + "content": "0.0375 bpp" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 378, + 462, + 457, + 540 + ], + "lines": [ + { + "bbox": [ + 378, + 462, + 457, + 540 + ], + "spans": [ + { + "bbox": [ + 378, + 462, + 457, + 540 + ], + "type": "image", + "image_path": "7fea7b78e033cb5524ff7be0cf7209f15d4deb05b698bd0359b6afcd0233cf3a.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 405, + 542, + 432, + 548 + ], + "lines": [ + { + "bbox": [ + 405, + 542, + 432, + 548 + ], + "spans": [ + { + "bbox": [ + 405, + 542, + 432, + 548 + ], + "type": "text", + "content": "MS-ILLM" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 401, + 549, + 435, + 557 + ], + "lines": [ + { + "bbox": [ + 401, + 549, + 435, + 557 + ], + "spans": [ + { + "bbox": [ + 401, + 549, + 435, + 557 + ], + "type": "text", + "content": "0.0088 bpp" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 459, + 368, + 539, + 445 + ], + "blocks": [ + { + "bbox": [ + 459, + 368, + 539, + 445 + ], + "lines": [ + { + "bbox": [ + 459, + 368, + 539, + 445 + ], + "spans": [ + { + "bbox": [ + 459, + 368, + 539, + 445 + ], + "type": "image", + "image_path": "9d1ad4d9804d1eb22ce311ef547f926a08263eed0ee6a5962a8a4ec8c59f7702.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 493, + 447, + 505, + 454 + ], + "lines": [ + { + "bbox": [ + 493, + 447, + 505, + 454 + ], + "spans": [ + { + "bbox": [ + 493, + 447, + 505, + 454 + ], + "type": "text", + "content": "GLC" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 459, + 463, + 539, + 540 + ], + "blocks": [ + { + "bbox": [ + 482, + 455, + 515, + 463 + ], + "lines": [ + { + "bbox": [ + 482, + 455, + 515, + 463 + ], + "spans": [ + { + "bbox": [ + 482, + 455, + 515, + 463 + ], + "type": "text", + "content": "0.0300 bpp" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 459, + 463, + 539, + 540 + ], + "lines": [ + { + "bbox": [ + 459, + 463, + 539, + 540 + ], + "spans": [ + { + "bbox": [ + 459, + 463, + 539, + 540 + ], + "type": "image", + "image_path": "611305432135b369f03cbe1954580b906c7afb1bd89a7af3521880ba727acf1c.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 492, + 542, + 505, + 548 + ], + "lines": [ + { + "bbox": [ + 492, + 542, + 505, + 548 + ], + "spans": [ + { + "bbox": [ + 492, + 542, + 505, + 548 + ], + "type": "text", + "content": "GLC" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 482, + 549, + 515, + 557 + ], + "lines": [ + { + "bbox": [ + 482, + 549, + 515, + 557 + ], + "spans": [ + { + "bbox": [ + 482, + 549, + 515, + 557 + ], + "type": "text", + "content": "0.0080 bpp" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + } + ], + "index": 43 + }, + { + "bbox": [ + 46, + 593, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 593, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 593, + 289, + 713 + ], + "type": "text", + "content": "Figure 9 shows the qualitative comparison results. For natural image compression, Text+Sketch generates low-fidelity reconstructions, while TCM, HiFiC, and MS-ILLM produce blurry results at ultra low bitrate. In contrast, GLC achieves high-fidelity and high-realism results. In the case of facial image compression, we find existing methods cannot produce satisfactory reconstruction due to the severely distorted information at even more extreme bitrate limitation (e.g., 0.01 bpp or lower). TCM and HiFiC fall short in generating highly realistic results. Even though MS-ILLM" + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 304, + 593, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 593, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 593, + 545, + 628 + ], + "type": "text", + "content": "provides clearer details, it struggles to preserve correct facial attribute. Compared to them, GLC excels in both realism and fidelity at even lower bitrate." + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 306, + 635, + 400, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 400, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 400, + 647 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 304, + 654, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 700 + ], + "type": "text", + "content": "In this section, we conduct ablation studies to demonstrate the effectiveness of each proposed component. To provide a clearer comparison, we evaluate the BD-Rate [7] on the FID-BPP curve on the CLIC 2020 test set." + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 306, + 702, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 702, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 702, + 545, + 714 + ], + "type": "text", + "content": "Transform coding. A straightforward approach to com" + } + ] + } + ], + "index": 51 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26094" + } + ] + } + ], + "index": 52 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 92, + 276, + 148 + ], + "blocks": [ + { + "bbox": [ + 70, + 71, + 264, + 82 + ], + "lines": [ + { + "bbox": [ + 70, + 71, + 264, + 82 + ], + "spans": [ + { + "bbox": [ + 70, + 71, + 264, + 82 + ], + "type": "text", + "content": "Table 1. Ablation study on latent-space compression." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 55, + 92, + 276, + 148 + ], + "lines": [ + { + "bbox": [ + 55, + 92, + 276, + 148 + ], + "spans": [ + { + "bbox": [ + 55, + 92, + 276, + 148 + ], + "type": "table", + "html": "
Latent coding schemeProbability model of zBD-Rate ↓
Indices-map coding-66.2%
Transform codingFactorized prior17.7%
Categorical prior0%
", + "image_path": "9d9f943e15242e2150b5a0df00560180e099bcd671766c2061e0e5dfe1637580.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 90, + 182, + 240, + 234 + ], + "blocks": [ + { + "bbox": [ + 66, + 159, + 267, + 171 + ], + "lines": [ + { + "bbox": [ + 66, + 159, + 267, + 171 + ], + "spans": [ + { + "bbox": [ + 66, + 159, + 267, + 171 + ], + "type": "text", + "content": "Table 2. Ablation study on the code prediction module." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 90, + 182, + 240, + 234 + ], + "lines": [ + { + "bbox": [ + 90, + 182, + 240, + 234 + ], + "spans": [ + { + "bbox": [ + 90, + 182, + 240, + 234 + ], + "type": "table", + "html": "
code prediction usageBD-Rate ↓
w/o code pred.13.1%
code pred. in network60.7%
code pred. as supervision0%
", + "image_path": "354d576cf5c2a39926f4b8e331c25b05776b714df2382868b213489d263ab2c6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 255, + 286, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 286, + 303 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 286, + 303 + ], + "type": "text", + "content": "press the VQ-VAE latents is indices-map coding [27, 28, 43]. However, it causes " + }, + { + "bbox": [ + 46, + 255, + 286, + 303 + ], + "type": "inline_equation", + "content": "66.2\\%" + }, + { + "bbox": [ + 46, + 255, + 286, + 303 + ], + "type": "text", + "content": " performance loss compared with transform coding, as shown in Table 1. It shows the effectiveness of transform coding on reducing redundancy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 303, + 286, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 303, + 286, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 303, + 286, + 361 + ], + "type": "text", + "content": "Categorical hyper module. In Section 3.3, we illustrate the superiority of employing a categorical prior for " + }, + { + "bbox": [ + 46, + 303, + 286, + 361 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 303, + 286, + 361 + ], + "type": "text", + "content": " compared to the commonly used factorized prior. Table 1 further provides a quantitative comparison, demonstrating a significant improvement of " + }, + { + "bbox": [ + 46, + 303, + 286, + 361 + ], + "type": "inline_equation", + "content": "17.7\\%" + }, + { + "bbox": [ + 46, + 303, + 286, + 361 + ], + "type": "text", + "content": " with such design." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 362, + 286, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 362, + 286, + 469 + ], + "spans": [ + { + "bbox": [ + 46, + 362, + 286, + 469 + ], + "type": "text", + "content": "Code-prediction-based supervision. In Section 4, we suggest employing the code prediction module as an auxiliary loss during training, instead of during the inference process of the model pipeline as in [27, 28, 60]. As shown in Table 2, incorporating the code prediction module directly into the network leads to a " + }, + { + "bbox": [ + 46, + 362, + 286, + 469 + ], + "type": "inline_equation", + "content": "60.7\\%" + }, + { + "bbox": [ + 46, + 362, + 286, + 469 + ], + "type": "text", + "content": " performance drop. We further remove the code-prediction-based supervision for comparison, and results show that adopting the code-prediction-based supervision brings a " + }, + { + "bbox": [ + 46, + 362, + 286, + 469 + ], + "type": "inline_equation", + "content": "13.1\\%" + }, + { + "bbox": [ + 46, + 362, + 286, + 469 + ], + "type": "text", + "content": " improvement." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 487, + 127, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 487, + 127, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 487, + 127, + 502 + ], + "type": "text", + "content": "6. Applications" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 510, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 712 + ], + "type": "text", + "content": "Leveraging the potent latent space, our GLC pipeline opens avenues for exploring various vision applications. In this paper, we implement image restoration and style transfer as examples to show its potential. As depicted in Figure 10, for image restoration, we train a restoration encoder to map distorted images into clean latents, allowing users to directly compress a noisy image and decompress a clean one without additional cost. We compare our restoration application with a straightforward scheme of cascading an additional restoration network [57] with the GLC codec. The results in Table 3 indicate superior performance for our restoration application without the need for extra model parameters. Similarly, users can directly decode the latent into another style through a stylization decoder to achieve style transfer. We hope such versatility of the generative latent space will foster connections between image compression and other vision tasks in future research." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 545, + 293 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 545, + 293 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 545, + 293 + ], + "type": "image", + "image_path": "0ce959ac2b89800cefdb94c256a4db5c43ac631d62c91fcdb1a4633fcee24b2e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 305, + 545, + 361 + ], + "lines": [ + { + "bbox": [ + 305, + 305, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 305, + 305, + 545, + 361 + ], + "type": "text", + "content": "Figure 10. Generative latent applications built on GLC. In the practical image compression system, users can choose different encoders and decoders, to compress an image (medium), compress a distorted image and decompress a clear one (top), or decode the image into another style (bottom)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 308, + 408, + 544, + 446 + ], + "blocks": [ + { + "bbox": [ + 306, + 376, + 545, + 398 + ], + "lines": [ + { + "bbox": [ + 306, + 376, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 306, + 376, + 545, + 398 + ], + "type": "text", + "content": "Table 3. Comparison for different joint restoration and compression schemes on CLIC 2020 test set." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 408, + 544, + 446 + ], + "lines": [ + { + "bbox": [ + 308, + 408, + 544, + 446 + ], + "spans": [ + { + "bbox": [ + 308, + 408, + 544, + 446 + ], + "type": "table", + "html": "
SchemeBPP ↓FID ↓DISTS ↓Parameters
Restormer [57] + GLC Codec0.031410.790.117425M + 109M
GLC Restoration Application0.02998.620.1081109M
", + "image_path": "0244d4bd419e38c8b0144199c3fc32e35b03527b9a9961a08c8ccb5a70cd4047.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 471, + 458, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 471, + 458, + 483 + ], + "spans": [ + { + "bbox": [ + 306, + 471, + 458, + 483 + ], + "type": "text", + "content": "7. Conclusion and Limitation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 493, + 545, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 493, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 304, + 493, + 545, + 612 + ], + "type": "text", + "content": "In this paper, we introduce a generative latent coding (GLC) scheme to achieve high-fidelity and high-realism generative compression at ultra-low bitrate. Unlike most existing pixel-domain codecs, GLC performs transform coding on the latent domain of a generative VQ-VAE. By incorporating a categorical hyper module and a code-prediction-based supervision, GLC demonstrates state-of-the-art performance on several benchmarks. We further develop several vision applications on the GLC pipeline to demonstrate its practical potential." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 613, + 545, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 613, + 545, + 703 + ], + "spans": [ + { + "bbox": [ + 304, + 613, + 545, + 703 + ], + "type": "text", + "content": "However, as a generative image codec trained on specified datasets, the generalization capability of GLC is not always satisfactory. For instance, GLC cannot guarantee a clear and accurate reconstruction of screen contents, as illustrated in the supplementary material. Future work will focus on addressing this limitation, to enhance the generalization ability of GLC by improving the model structure and the training strategy." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26095" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 126-135, 2017. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 288, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 288, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 288, + 191 + ], + "type": "text", + "content": "[2] Eirikur Agustsson, Michael Tschannen, Fabian Mentzer, Radu Timofte, and Luc Van Gool. Generative adversarial networks for extreme learned image compression. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 221-231, 2019. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 288, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 288, + 246 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 288, + 246 + ], + "type": "text", + "content": "[3] Eirikur Agustsson, David Minnen, George Toderici, and Fabian Mentzer. Multi-realism image compression with a conditional generator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22324-22333, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 247, + 288, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 288, + 290 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 288, + 290 + ], + "type": "text", + "content": "[4] Johannes Balle, Valero Laparra, and Eero P Simoncelli. End-to-end optimized image compression. In 5th International Conference on Learning Representations, ICLR 2017, 2017. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 291, + 288, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 288, + 335 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 288, + 335 + ], + "type": "text", + "content": "[5] Johannes Balle, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. In International Conference on Learning Representations, 2018. 1, 2, 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 335, + 288, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 288, + 368 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 288, + 368 + ], + "type": "text", + "content": "[6] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. arXiv preprint arXiv:1801.01401, 2018. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 369, + 288, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 288, + 391 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 288, + 391 + ], + "type": "text", + "content": "[7] Gisle Bjontegaard. Calculation of average psnr differences between rd-curves. ITU SG16 Doc. VCEG-M33, 2001. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 392, + 288, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 392, + 288, + 435 + ], + "spans": [ + { + "bbox": [ + 53, + 392, + 288, + 435 + ], + "type": "text", + "content": "[8] Marlene Careil, Matthew J Muckley, Jakob Verbeek, and Stéphane Lathuilière. Towards image compression with perfect realism at ultra-low bitrates. In The Twelfth International Conference on Learning Representations, 2023. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 436, + 288, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 436, + 288, + 490 + ], + "spans": [ + { + "bbox": [ + 53, + 436, + 288, + 490 + ], + "type": "text", + "content": "[9] Jianhui Chang, Zhenghui Zhao, Chuanmin Jia, Shiqi Wang, Lingbo Yang, Qi Mao, Jian Zhang, and Siwei Ma. Conceptual compression via deep structure and texture synthesis. IEEE Transactions on Image Processing, 31:2809-2823, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 491, + 287, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 287, + 535 + ], + "type": "text", + "content": "[10] Xi Chen, Diederik P Kingma, Tim Salimans, Yan Duan, Prafulla Dhariwal, John Schulman, Ilya Sutskever, and Pieter Abbeel. Variational lossy autoencoder. arXiv preprint arXiv:1611.02731, 2016. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 536, + 287, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 591 + ], + "type": "text", + "content": "[11] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned image compression with discretized gaussian mixture likelihoods and attention modules. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7939-7948, 2020. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 591, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 287, + 613 + ], + "type": "text", + "content": "[12] Thomas M Cover. Elements of information theory. John Wiley & Sons, 1999. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 614, + 288, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 288, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 288, + 669 + ], + "type": "text", + "content": "[13] Ze Cui, Jing Wang, Shangyin Gao, Tiansheng Guo, Yihui Feng, and Bo Bai. Asymmetric gained deep image compression with continuous rate adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10532-10541, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 714 + ], + "type": "text", + "content": "[14] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[15] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. IEEE transactions on pattern analysis and machine intelligence, 44(5):2567-2581, 2020. 1, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "text", + "content": "[16] Alaaeldin El-Nouby, Matthew J Muckley, Karen Ullrich, Ivan Laptev, Jakob Verbeek, and Herve Jegou. Image compression with product quantized masked image modeling. Transactions on Machine Learning Research, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 162, + 545, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 215 + ], + "type": "text", + "content": "[17] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 217, + 545, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 261 + ], + "type": "text", + "content": "[18] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 261, + 545, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 283 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 283 + ], + "type": "text", + "content": "[19] Vivek K Goyal. Theoretical foundations of transform coding. IEEE Signal Processing Magazine, 18(5):9-21, 2001. 1, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 284, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 284, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 308, + 284, + 545, + 327 + ], + "type": "text", + "content": "[20] Wang Guo-Hua, Jiahao Li, Bin Li, and Yan Lu. Evc: Towards real-time neural image compression with mask decay. In The Eleventh International Conference on Learning Representations, 2022. 1, 3, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 327, + 545, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 382 + ], + "type": "text", + "content": "[21] Dailan He, Yaoyan Zheng, Baocheng Sun, Yan Wang, and Hongwei Qin. Checkerboard context model for efficient learned image compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14771-14780, 2021. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 383, + 545, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 449 + ], + "type": "text", + "content": "[22] Dailan He, Ziming Yang, Hongjiu Yu, Tongda Xu, Jixiang Luo, Yuan Chen, Chenjian Gao, Xinjie Shi, Hongwei Qin, and Yan Wang. Po-elic: Perception-oriented efficient learned image coding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1764–1769, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 449, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 502 + ], + "type": "text", + "content": "[23] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 503, + 545, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 503, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 308, + 503, + 545, + 548 + ], + "type": "text", + "content": "[24] Emiel Hoogeboom, Eirikur Agustsson, Fabian Mentzer, Luca Versari, George Toderici, and Lucas Theis. High-fidelity image compression with score-based generative models. arXiv preprint arXiv:2305.18231, 2023. 3, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 548, + 545, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 603 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 603 + ], + "type": "text", + "content": "[25] Yueyu Hu, Shuai Yang, Wenhan Yang, Ling-Yu Duan, and Jiaying Liu. Towards coding for human and machine vision: A scalable image coding approach. In 2020 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2020. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 603, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 603, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 308, + 603, + 545, + 658 + ], + "type": "text", + "content": "[26] Shoma Iwai, Tomo Miyazaki, Yoshihiro Sugaya, and Shinichiro Omachi. Fidelity-controllable extreme image compression with generative adversarial networks. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 8235-8242. IEEE, 2021. 3, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "text", + "content": "[27] Wei Jiang, Hyomin Choi, and Fabien Racapé. Adaptive human-centric video compression for humans and machines. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1121-1129, 2023. 2, 3, 4, 6, 8" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "26096" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[28] Wei Jiang, Hyomin Choi, Fabien Racapé, Simon Feltman, and Fatih Kamisli. Face restoration-based scalable quality coding for video conferencing. In 2023 IEEE International Conference on Multimedia and Expo Workshops (ICMEW), pages 206-211. IEEE, 2023. 2, 3, 4, 6, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 288, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 288, + 193 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 288, + 193 + ], + "type": "text", + "content": "[29] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016. 1, 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 194, + 288, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 194, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 48, + 194, + 288, + 237 + ], + "type": "text", + "content": "[30] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In International Conference on Learning Representations, 2018. 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 238, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 238, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 238, + 288, + 282 + ], + "type": "text", + "content": "[31] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 283, + 286, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 283, + 286, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 283, + 286, + 304 + ], + "type": "text", + "content": "[32] Kodak Lossless True Color Image Suite. http://r0k.us/graphics/kodak/.6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 304, + 288, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 304, + 288, + 380 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 288, + 380 + ], + "type": "text", + "content": "[33] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision, 128(7):1956-1981, 2020. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 382, + 288, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 382, + 288, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 382, + 288, + 426 + ], + "type": "text", + "content": "[34] Jooyoung Lee, Seunghyun Cho, and Seung-Kwon Beck. Context-adaptive entropy model for end-to-end optimized image compression. In International Conference on Learning Representations, 2018. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 426, + 288, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 288, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 288, + 480 + ], + "type": "text", + "content": "[35] Eric Lei, Yigit Berkay Uslu, Hamed Hassani, and Shirin Saeedi Bidokhti. Text+ sketch: Image compression at ultra low rates. In ICML 2023 Workshop Neural Compression: From Information Theory to Applications, 2023. 2, 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 481, + 288, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 288, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 288, + 514 + ], + "type": "text", + "content": "[36] Jiahao Li, Bin Li, and Yan Lu. Deep contextual video compression. Advances in Neural Information Processing Systems, 34:18114-18125, 2021. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 514, + 288, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 288, + 558 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 288, + 558 + ], + "type": "text", + "content": "[37] Jiahao Li, Bin Li, and Yan Lu. Hybrid spatial-temporal entropy modelling for neural video compression. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1503-1511, 2022." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 559, + 288, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 288, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 288, + 602 + ], + "type": "text", + "content": "[38] Jiahao Li, Bin Li, and Yan Lu. Neural video compression with diverse contexts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22616-22626, 2023. 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 603, + 288, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 288, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 288, + 646 + ], + "type": "text", + "content": "[39] Jiahao Li, Bin Li, and Yan Lu. Neural video compression with feature modulation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 17-21, 2024, 2024. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 713 + ], + "type": "text", + "content": "[40] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[41] Jinming Liu, Heming Sun, and Jiro Katto. Learned image compression with mixed transformer-cnn architectures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14388-14397, 2023. 1, 3, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 150 + ], + "type": "text", + "content": "[42] Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam. 2018. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "text", + "content": "[43] Qi Mao, Tinghan Yang, Yinuo Zhang, Shuyin Pan, Meng Wang, Shiqi Wang, and Siwei Ma. Extreme image compression using fine-tuned vqgan models. arXiv preprint arXiv:2307.08265, 2023. 2, 4, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "type": "text", + "content": "[44] Fabian Mentzer, George D Toderici, Michael Tschannen, and Eirikur Agustsson. High-fidelity generative image compression. Advances in Neural Information Processing Systems, 33:11913-11924, 2020. 2, 3, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "type": "text", + "content": "[45] David Minnen, Johannes Balle, and George D Toderici. Joint autoregressive and hierarchical priors for learned image compression. Advances in neural information processing systems, 31, 2018. 1, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 286, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 545, + 341 + ], + "type": "text", + "content": "[46] Matthew J. Muckley, Alaaeldin El-Nouby, Karen Ullrich, Hervé Jégou, and Jakob Verbeek. Improving statistical fidelity for neural image compression with implicit local likelihood models. In International Conference on Machine Learning, 2023. 1, 2, 3, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 342, + 545, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 545, + 386 + ], + "type": "text", + "content": "[47] Linfeng Qi, Jiahao Li, Bin Li, Houqiang Li, and Yan Lu. Motion information propagation for neural video compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6111-6120, 2023. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 388, + 545, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 442 + ], + "type": "text", + "content": "[48] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 443, + 545, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 443, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 307, + 443, + 545, + 475 + ], + "type": "text", + "content": "[49] Xihua Sheng, Jiahao Li, Bin Li, Li Li, Dong Liu, and Yan Lu. Temporal context mining for learned video compression. IEEE Transactions on Multimedia, 2022. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 476, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 476, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 476, + 545, + 510 + ], + "type": "text", + "content": "[50] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 510, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 510, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 510, + 545, + 555 + ], + "type": "text", + "content": "[51] George Toderici, Lucas Theis, Nick Johnston, Eirikur Agustsson, Fabian Mentzer, Johannes Balle, Wenzhe Shi, and Radu Timofte. Clic 2020: Challenge on learned image compression, 2020, 2020. 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 555, + 545, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 555, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 307, + 555, + 545, + 599 + ], + "type": "text", + "content": "[52] Aäron Van Den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks. In International conference on machine learning, pages 1747-1756. PMLR, 2016. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 600, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 600, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 307, + 600, + 545, + 633 + ], + "type": "text", + "content": "[53] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 1, 2, 3, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 634, + 545, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 634, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 307, + 634, + 545, + 666 + ], + "type": "text", + "content": "[54] VVC-21.2. https://vcgit.hhi.fraunhofer.de/jvet/VVCSoftware_VTM/~/tree/VTM-21.2. Accessed: 10/23/2023, 2023. 1, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 669, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 689 + ], + "type": "text", + "content": "[55] Gregory K Wallace. TheJPEG still picture compression standard.Communications of the ACM, 34(4):30-44, 1991.1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[56] Zhou Wang, Eero P Simoncelli, and Alan C Bovik. Multiscale structural similarity for image quality assessment. In" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26097" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 317 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "text", + "content": "The Thrity-Seventh Asilomar Conference on Signals, Systems & Computers, 2003, pages 1398-1402. IEEE, 2003. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 160 + ], + "type": "text", + "content": "[57] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728-5739, 2022. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 217 + ], + "type": "text", + "content": "[58] Jing Zhao, Bin Li, Jiahao Li, Ruiqin Xiong, and Yan Lu. A universal encoder rate distortion optimization framework for learned compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1880-1884, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 219, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 219, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 287, + 262 + ], + "type": "text", + "content": "[59] Jing Zhao, Bin Li, Jiahao Li, Ruiqin Xiong, and Yan Lu. A universal optimization framework for learning-based image codec. ACM Transactions on Multimedia Computing, Communications and Applications, 20(1):1-19, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 263, + 287, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 263, + 287, + 317 + ], + "spans": [ + { + "bbox": [ + 48, + 263, + 287, + 317 + ], + "type": "text", + "content": "[60] Shangchen Zhou, Kelvin Chan, Chongyi Li, and Chen Change Loy. Towards robust blind face restoration with codebook lookup transformer. Advances in Neural Information Processing Systems, 35:30599-30611, 2022. 2, 3, 6, 8" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "26098" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Generative Multi-modal Models are Good Class Incremental Learners/a19d9ccd-1dff-4a86-83d4-a672741b4e0a_content_list.json b/2024/Generative Multi-modal Models are Good Class Incremental Learners/a19d9ccd-1dff-4a86-83d4-a672741b4e0a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6df96dcb576ffa69b3be4c695f7819c46596e93f --- /dev/null +++ b/2024/Generative Multi-modal Models are Good Class Incremental Learners/a19d9ccd-1dff-4a86-83d4-a672741b4e0a_content_list.json @@ -0,0 +1,1559 @@ +[ + { + "type": "text", + "text": "Generative Multi-modal Models are Good Class-Incremental Learners", + "text_level": 1, + "bbox": [ + 127, + 130, + 841, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xusheng Cao $^{1}$ , Haori Lu $^{1}$ , Linlan Huang $^{1}$ , Xialei Liu $^{2,1*}$ , Ming-Ming Cheng $^{2,1}$ $^{1}$ VCIP, CS, Nankai University $^{2}$ NKIARI, Shenzhen Futian", + "bbox": [ + 174, + 179, + 792, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{caoxusheng, luhaori, huanglinlan}@mail.nankai.edu.cn, {xialei, cmm}@nankai.edu.cn", + "bbox": [ + 120, + 218, + 841, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In class-incremental learning (CIL) scenarios, the phenomenon of catastrophic forgetting caused by the classifier's bias towards the current task has long posed a significant challenge. It is mainly caused by the characteristic of discriminative models. With the growing popularity of the generative multi-modal models, we would explore replacing discriminative models with generative ones for CIL. However, transitioning from discriminative to generative models requires addressing two key challenges. The primary challenge lies in transferring the generated textual information into the classification of distinct categories. Additionally, it requires formulating the task of CIL within a generative framework. To this end, we propose a novel generative multi-modal model (GMM) framework for class-incremental learning. Our approach directly generates labels for images using an adapted generative model. After obtaining the detailed text, we use a text encoder to extract text features and employ feature matching to determine the most similar label as the classification prediction. In the conventional CIL settings, we achieve significantly better results in long-sequence task scenarios. Under the Few-shot CIL setting, we have improved by at least $14\\%$ accuracy over all the current state-of-the-art methods with significantly less forgetting. Our code is available at https://github.com/DoubleClass/GMM.", + "bbox": [ + 75, + 301, + 473, + 680 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 710, + 209, + 726 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep neural networks [19, 33, 56] have made remarkable strides in numerous applications, primarily owing to the vast amounts of data and computational resources at their disposal. Nonetheless, these accomplishments are predominantly contingent on having access to all the required data simultaneously for training on various tasks. In cases where data is acquired incrementally, these networks often encounter the challenge of catastrophic forgetting [43]. Hence, the capacity to seamlessly incorporate new knowl", + "bbox": [ + 75, + 737, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2073830f54ab6f9c767edec7e8011939c055c676c9226f2286e014db165462d5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 267, + 885, + 338 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/faa7a05aa9b57d9a251de60b9c1b9de4a9bef9f4ffae2aec368dcb51f6459e26.jpg", + "image_caption": [ + "Figure 1. Illustration of conventional discriminative models for class-incremental learning (CIL) and our generative multi-modal models (GMM) for CIL. Discriminative models pose potential risk of classifier bias toward current task with network expansion. Our GMM framework consists generation and classification phases. And It is adapted to CIL based on the similarity of generated text and the true category names." + ], + "image_footnote": [], + "bbox": [ + 504, + 339, + 888, + 469 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "edge while retaining previously acquired knowledge is a highly desirable attribute for future artificial intelligence systems. Continual learning [45, 67, 77, 84] is a subject of study aimed at advancing the evolution of neural networks toward this goal.", + "bbox": [ + 496, + 608, + 893, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Numerous studies have delved into continual learning, categorizing their approaches into three main groups [14]: rehearsal-based, architecture-based, and regularization-based methods. Additionally, hybrid methods are gaining popularity as they combine insights from different perspectives. Within this research landscape, three primary scenarios [66] have received extensive attention, with class-incremental learning (CIL) [41] being one of the most demanding settings. In our work, we concentrate on CIL, wherein each task comprises a distinct set of classes, and the primary challenge is to enable the network to recognize new classes without forgetting knowledge of previously encountered ones.", + "bbox": [ + 496, + 686, + 893, + 881 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The majority of existing research on class-incremental", + "bbox": [ + 519, + 885, + 890, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 887, + 215, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "28706", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "learning (CIL) focuses on training models from scratch, relying solely on data from the current tasks [7, 15, 16, 24, 26, 29, 48, 74, 76, 89]. In contrast, humans accumulate knowledge over extended periods, drawing upon a wealth of prior world knowledge. Consequently, there has been a growing interest in pre-trained models for CIL [55, 71, 72, 81, 86], harnessing the knowledge acquired from extensive, pre-existing datasets to address the immediate tasks at hand. For instance, prompt-based methods [55, 71, 72] utilize prompt-tuning to summarize task-specific knowledge from prior pretrained knowledge. SLCA [81] and ADAM [85], on the other hand, solely fine-tuning the pre-trained model to adapt the pre-existing knowledge into the objectives of immediate tasks.", + "bbox": [ + 75, + 90, + 472, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle image classification downstream tasks, pretrained models are traditionally derived from discriminative tasks, like supervised learning on datasets such as ImageNet-21K [49], or they may stem from self-supervised learning efforts [6, 9, 10, 22]. However, in our research, we venture into the paradigm of generative multi-modal models to address image classification tasks. Generative models like GPT4 [44] and LLaVa [32] have garnered significant attention in recent years due to their capacity to produce highly informative descriptions of input images. On the one hand, it can harness the wealth of semantic correspondences between texts and images, while on the other hand, there's no requirement to expand the classifier with each new task, unlike in the case of discriminative models for CIL.", + "bbox": [ + 75, + 305, + 472, + 517 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Nonetheless, harnessing the knowledge from pre-trained generative models for downstream class-incremental learning (CIL) tasks presents a nontrivial endeavor. The primary challenge lies in transferring the generated textual information into the classification of distinct categories. Additionally, there's the task of formulating CIL within a generative framework, which poses a second significant challenge. Shao et al. [52] presents the VAG system, which formulates CIL as a continual label generation problem, preserving the language model's ability to learn new classes. However, it only works in the field of Natural Language Processing (NLP), which is inherently suited to Large Language Models (LLM). To the best of our knowledge, we are the first to apply this generative approach to incremental learning in the field of image classification.", + "bbox": [ + 75, + 520, + 472, + 747 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose Generative Multi-modal Models (GMM) for class-incremental learning. As illustrated in Fig. 1 (a), conventional discriminative methods extract image features with a network backbone, then forward them to a classifier to obtain the probability of the image belonging to each label, with the label having the highest probability being the output of the discriminative models. While in Fig. 1 (b), we adopt a generative approach, which directly produces a descriptive sentence for the given image, which is then compared with the actual label texts with a", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "text encoder. The most similar label becomes the predicted result of our generative model. This approach allows us to leverage the rich pre-training knowledge in generative multi-modal models while avoiding the use of the expanded classification head, which mitigates the risk of the model bias towards the current task and reduces catastrophic forgetting.", + "bbox": [ + 496, + 90, + 893, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions of this paper are:", + "bbox": [ + 517, + 196, + 792, + 212 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel generative approach (GMM) to address class-incremental learning by leveraging multimodal models.", + "- We reformulate GMM for image classification and adapt it for the downstream benchmarks. Without an expanded classification head like in discriminative models, our model significantly mitigates the issue of bias towards current tasks, resulting in significantly reduced forgetting in CIL.", + "- Our model achieves state-of-the-art performance across multiple datasets in both conventional and few-shot CIL settings." + ], + "bbox": [ + 500, + 215, + 893, + 396 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 410, + 640, + 426 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Class-Incremental Learning", + "text_level": 1, + "bbox": [ + 500, + 435, + 754, + 452 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In Class-Incremental Learning, tasks arrive sequentially, and each class is exclusive to a specific task without any overlap. The goal is to acquire knowledge from new classes while preserving information from previously encountered classes. There are three primary branches in CIL [14], including rehearsal-based, architecture-based, and regularization-based methods. Rehearsal-based methods [1, 7, 48, 75] store a small set of data derived from old classes to represent knowledge from previous tasks. These exemplar data can be either original data [48], generative data [18, 54] or hidden features [20]. Architecture-based methods focus on modifying network architecture to alleviate forgetting. Approaches include learning redundant network architecture [17, 47], learning different expert networks [3, 50] or parameters [38, 40, 51] for each task, dynamically expanding network parameters to accumulate incremental knowledge [76]. Regularization-based methods introduce an additional regularization term to restrict network updates when adapting to new tasks. In such cases, EWC [26], SDC [78] and Rotated-EWC [35] expect that parameters essential for the old tasks should not be updated excessively. Moreover, from the perspective of network output consistency, numerous studies [25, 30, 36, 61, 80] incorporate distillation to prevent forgetting.", + "bbox": [ + 496, + 458, + 893, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Few-shot CIL Few-shot Class-Incremental Learning (FSCIL) [42, 62] explores few-shot learning in an incremental context, with all data samples available for base session and very limited data in each incremental session. Some FSCIL methods [11, 62, 83] train the model in both base and incre", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "28707", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "mental sessions, aiming to mitigate overfitting challenges caused by the limited data in incremental learning. Other strategies [53, 79, 90] primarily train the model in the base session and make minimal adjustments in the incremental sessions, thereby reducing forgetting but may come at the cost of decreased precision in the incremental sessions.", + "bbox": [ + 75, + 90, + 470, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Pre-trained models for CIL", + "text_level": 1, + "bbox": [ + 76, + 191, + 323, + 208 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "There are many methods [63, 72, 73] having shown that pretrained models are effective for continual learning. One main branch trains a set of prompts to retain previous knowledge [55, 68, 71, 72]. A selected subset of prompts are fed into the model during forward to prompt model the past knowledge. Additionally, methods like SLCA [81] and ADAM [85] fine-tune pre-trained models, achieving impressive results with less forgetting. Continual-CLIP [63] demonstrates that the CLIP [46] model is capable of performing continual learning without any extra training. This highlights the significant potential of multi-modal pretrained models in the realm of continual learning. Inspired by this, many methods [37, 86] employ CLIP as the backbone to utilize the multi-modal information. However, if not using a classifier directly, these approaches need to utilize expanded text features to calculate distances with image features for classification. This will exacerbate the model's bias towards current data, consequently leading to the forgetting of previously acquired knowledge. To avoid this bias, we use a generative model to directly generate prediction text. The fixed text decoder will function as the classifier, significantly alleviating the bias.", + "bbox": [ + 75, + 215, + 473, + 549 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Vision Language Models", + "text_level": 1, + "bbox": [ + 76, + 559, + 303, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In recent years, vision-language multi-modal models have made significant progress and achieved impressive results in various downstream tasks [5, 28, 34, 70]. Traditional vision-language models employ different types of encoders to extract information from vision and language models, including single-stream [57], dual-stream [39] and fusion [60] encoders. A key aspect of vision-language models is the alignment of multi-modal features. CLIP [46], for example, extracts image and text features separately using respective encoders and enforces alignment through a contrastive loss, ensuring alignment between positive image-text pairs in the feature space. VisualGPT [8] and Frozen [65] leverage pre-trained models as encoders for visual-language tasks. From then on, the utilization of pre-trained models in vision-language tasks became more and more popular. For instance, Flamingo [2] and BLIP-2 [27] align the pre-trained image and text encoders employing gated cross-attention and Q-Former, respectively. Furthermore, LLaVA [32] and MiniGPT-4 [88] leverage more robust Large Language Models (LLM) [13, 64] as text encoders, while only training a projection layer for alignment. With", + "bbox": [ + 75, + 583, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the increasing popularity of LLM, an increasing number of studies [4, 69, 87] explore the potential of multi-modal LLMs for vision-language tasks.", + "bbox": [ + 496, + 90, + 890, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 498, + 151, + 591, + 166 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce the preliminaries of class-incremental learning and generative multi-modal models. Then, we present our approach to leverage generative models for CIL and the corresponding learning process.", + "bbox": [ + 496, + 176, + 890, + 238 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 498, + 248, + 640, + 263 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Class-Incremental Learning. Given N tasks $T = \\{T_{1}, T_{2}, \\dots, T_{N}\\}$ , the goal of class-incremental learning is to learn each task $T_{t}$ with its associated data $\\{\\mathbf{X}_{t}, \\mathbf{Y}_{t}\\}$ in a sequential order. For each task, it contains samples $\\{\\mathbf{x}_i, \\mathbf{y}_i\\}$ , $i = 1, \\dots, n_t$ , where $\\mathbf{x}_i$ is the images and $\\mathbf{y}_i$ is the corresponding one-hot labels. Typically, $\\mathbf{X}_i \\cap \\mathbf{X}_j = \\emptyset, \\forall i \\neq j$ . At inference, the model is tested on all seen tasks without task IDs. In some scenarios, fixed memory storage is set to keep a few samples of previous tasks to prevent forgetting.", + "bbox": [ + 496, + 275, + 890, + 411 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Normally, a CIL model consists of a feature extractor and a classifier head $F = \\{f_{\\theta}, \\mathcal{H}_{\\phi}\\}$ which are parameterized by $\\{\\theta, \\phi\\}$ . In traditional class-incremental learning, $\\theta$ is usually a modified ResNet [21] with all parameters tunable. In pre-trained or prompt-based methods, $\\theta$ represents fewer trainable parameters like a linear adaptor or a couple of prompts. $\\phi$ is a linear classifier head projecting image features to probability predictions, which has to be expanded for each new task in order to make predictions for the new classes. The conventional Cross Entropy loss is often used for updating $\\theta$ and $\\phi$ , which for task $t$ is:", + "bbox": [ + 496, + 411, + 890, + 578 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C E}} \\left(\\mathbf {X} _ {t}, \\mathbf {Y} _ {t}; \\theta , \\phi\\right) = - \\frac {1}{n _ {t}} \\sum_ {i = 1} ^ {n _ {t}} \\mathbf {y} _ {i} \\cdot \\log \\mathcal {H} (f (\\mathbf {x} _ {i}; \\theta); \\phi). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 603, + 890, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the continual learning process, the parameter $\\phi$ can easily deviate to the data of the current task due to the absence or scarcity of old samples in the previous tasks, resulting in forgetting the previously acquired knowledge and deteriorating the overall performance.", + "bbox": [ + 496, + 655, + 890, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generative Multi-Modal Models (GMM). Multi-modal models have demonstrated exceptional performance in generating detailed image descriptions by incorporating both visual and textual information. Notably, GPT-4 [44] stands out as an advanced model proficient in generating comprehensive image descriptions and providing explanations for the depicted content. Furthermore, MiniGPT-4 [88] proposes a two-stage fine-tuning process that aligns image features and large language models, enabling LLaMa [64] to recognize images and conduct further dialogue based on the image content.", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "28708", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0c76f35c4278ab83e6618a5a6e79752f02729fc226ffe59b2694249f5ebdc057.jpg", + "image_caption": [ + "Figure 2. The overview structure of our proposed method. The conceptual illustration of the generative multi-modal model (GMM) is shown on the left. In order to adapt this model for CIL, We have to turn the GMM model for classification and further adapt to our objective benchmark for learning (see Sec. 3.2). On the right side, we demonstrate how the final evaluation is taken for CIL with all seen classes. Text encoder is used for obtaining the embeddings for similarity prediction." + ], + "image_footnote": [], + "bbox": [ + 89, + 93, + 529, + 354 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/114042b0e1f8d66755f9eb4f73250f247fb96697ae64dece69a0f5a6f6208f5d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 560, + 93, + 870, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 2, these models consist of an encoder $f_{enc}$ to generate content embeddings including image embedding and text embedding, which is further used as the input of an auto-aggressive decoder $f_{dec}$ to generate image descriptions. Input image $\\mathbf{x}_i$ is encoded as an image embedding $e_i$ , and question embeddings $\\mathbf{q}$ with $q_1, \\dots, q_l$ can be concatenated with the image embedding together to generate answer embeddings $\\mathbf{s}$ with $s_1, \\dots, s_m$ . The output tokens are generated one by one with the condition of previously generated tokens. For instance, $s_m$ is generated with all previous $m - 1$ tokens (see the decoder in Fig. 2).", + "bbox": [ + 75, + 449, + 472, + 617 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Generative Multi-Modal Models for CIL", + "text_level": 1, + "bbox": [ + 76, + 628, + 424, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We adhere to the foundational settings of MiniGPT-4, incorporating a frozen image encoder $f_{enc}$ followed by a trainable projection layer for adaptation to downstream tasks, as shown in Fig. 2. Our primary innovation involves the direct utilization of generative models to produce text, which can then serve as a basis for discriminative classification. However, two major challenges need addressing. First, there is the issue of adapting generative multi-modal models for classification, given that the generated text may differ significantly from class names. Second, we must devise a mechanism for our classification benchmarks to learn in a manner consistent with generative multi-modal models. We introduce these two aspects as follows.", + "bbox": [ + 75, + 652, + 468, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Turning GMM for classification. We employ a distance metric to bridge the gap between generative and discriminative models. During training, we use the real label's text to", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "encourage the model to predict the label of an image with a concise and accurate sentence in the format of \"This is a photo of [CLS].\" avoiding detailed descriptions of all contents in the image. During testing, the model follows the format to output the category text for a given image. We extract the content in \"[CLS]\", then obtain its text features using the CLIP [46] text encoder $f_{\\text{text}}$ , and compute the distance with text features of all categories seen by now. The closest class was then considered the final prediction of the generative model.", + "bbox": [ + 496, + 449, + 890, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Converting CIL Benchmarks for adaptation. CIL is usually evaluated on ImageNet, CIFAR-100, and ImageNet-R datasets. These datasets usually consist of images and corresponding one-hot labels $\\{\\mathbf{X}_t,\\mathbf{Y}_t\\}$ . Using the CIFAR100 dataset as an example, we pair each image with a sentence to form an image-text pair format $\\{\\mathbf{X}_t,\\mathbf{S}_t\\}$ with the template: \"This is a photo of [CLS]\", where \"[CLS]\" is the label name of that category, such as apple, dog, etc. Next, we partitioned the 100 classes into various tasks based on different settings and fed them into the model sequentially. After completing the training on task $T$ , the model should be capable of classifying all the classes encompassed from task 0 to task $T$ . Note that only the linear projection layer is updated for further adaptation.", + "bbox": [ + 496, + 604, + 892, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Optimization and Inference", + "text_level": 1, + "bbox": [ + 500, + 827, + 750, + 843 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Optimization. For each task $t$ , we obtain the current task's image-text pair $\\{\\mathbf{X}_t, \\mathbf{S}_t\\}$ , where $\\mathbf{S}_t$ contains the corresponding sentence of each image. During training, we", + "bbox": [ + 496, + 854, + 892, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "28709", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/795324d2883931c7dfc59b1422c50d7b51d83c092b2f94fc9e08d72aec9c2eaa.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TypeMethodExemplarTiny-ImageNetImageNet-R
5 tasks10 tasks20 tasks10 tasks
AvgLastAvgLastAvgLastLast
ConventionalEWC [26]X19.016.0015.823.7912.354.7335.00
LwF [29]X22.317.3417.344.7312.484.2638.50
iCaRL [48]✓45.9534.6043.2233.2237.8527.54-
EEIL [7]✓47.1735.1245.0334.6440.4129.72-
UCIR [24]✓50.3039.4248.5837.2942.8430.85-
PASS [89]X49.5441.6447.1939.2742.0132.93-
DyTox [16]✓55.5847.2352.2642.7946.1836.21-
Discriminative PT modelsContinual-CLIP[63]X70.4966.4370.5566.4370.5166.4372.00
L2P [72]X83.5378.3276.3765.7868.0452.4072.92
L2P [72]✓80.2472.8980.0872.6179.4470.4159.78
DualPrompt [71]X85.1581.0181.3873.7373.4560.1668.82
DualPrompt [71]✓79.9272.8379.1573.2180.1771.7457.02
CODA-Prompt [55]X85.9181.3682.8075.2877.4366.3273.88
Linear ProbeX74.3865.4069.7358.3160.1449.7245.17
Linear Probe✓70.1061.1169.3564.1971.6470.5055.72
Generative PT modelsZero-shotX58.1653.7258.1053.7258.1353.7267.38
GMM (Ours)X83.4276.9882.4976.5181.7076.0380.72
GMM (Ours)✓84.1678.4683.9578.6484.2379.1789.41
", + "bbox": [ + 98, + 88, + 874, + 445 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Comparison results of our method with other conventional baselines and methods learned with discriminative pre-trained (PT) models on Tiny-ImageNet and ImageNet-R under the conventional CIL setting. \"Avg\" represents the averaged performance after training each task, and \"Last\" represents the performance on all test samples after training the last task.", + "bbox": [ + 75, + 455, + 893, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "first utilize a tokenizer to tokenize and acquire the embedding of the questions and answers. We leverage pre-trained encoder $f_{enc}$ and the projection layer to obtain the corresponding features for the input images:", + "bbox": [ + 75, + 518, + 470, + 580 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {e} _ {i} = f _ {\\text {e n c}} \\left(\\mathbf {x} _ {i}; \\theta_ {\\text {e n c}}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 601, + 468, + 618 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, the question embedding and the ground-truth embedding of this question, e.g., \"This is a photo of [CLS]\", is concatenated with image embedding. The final input of LLM Decoder $f_{dec}$ is:", + "bbox": [ + 75, + 631, + 468, + 694 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {e}} _ {i} = \\operatorname {C O N C A T E} (b o s, \\mathbf {e} _ {i}, \\mathbf {q}, \\mathbf {s}, e o s). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 714, + 468, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "bos is the symbol of the sentence beginning, and $eos$ is the symbol for the end of the sentence. This encourages tokens at positions $m - 1$ to predict token $m$ :", + "bbox": [ + 75, + 742, + 468, + 787 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP \\left(\\hat {s} _ {1}, \\hat {s} _ {2}, \\dots , \\hat {s} _ {m} \\mid \\mathbf {x} _ {i}, \\mathbf {q}, \\mathbf {s}\\right) = \\prod_ {j = 1} ^ {m - 1} P \\left(s _ {j} \\mid \\mathbf {e} _ {i}, \\mathbf {q}, s _ {1}, s _ {2}, \\dots , s _ {j - 1}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 81, + 803, + 468, + 852 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $s_j$ indicates the ground-truth answer token and $\\hat{s}_m$ is the generated prediction. Then, we can compute the Cross Entropy loss as follows:", + "bbox": [ + 75, + 854, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C E}} = - \\frac {1}{m} \\sum_ {j = 1} ^ {m} s _ {j} \\cdot \\log \\hat {s} _ {j}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 529, + 890, + 570 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inference. During inference, we use the updated projection layer in conjunction with the pre-trained encoder to obtain image features. These image features, combined with the question embeddings are then passed to the LLM Decoder to obtain the text output.", + "bbox": [ + 496, + 575, + 890, + 652 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {p r e d} = \\operatorname {a r g m a x} < f _ {\\text {t e x t}} (\\mathbf {s}), f _ {\\text {t e x t}} (\\hat {\\mathbf {s}}) >, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 666, + 890, + 683 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $f_{\\text{text}}$ is the text encoder, $<, >$ is the cosine similarity used to calculate the final predictions $pred$ .", + "bbox": [ + 496, + 686, + 890, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 728, + 633, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental setups", + "text_level": 1, + "bbox": [ + 500, + 753, + 694, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets and Baselines. We conduct experiments in both conventional CIL and Few-shot CIL scenarios. In conventional CIL, we evaluate on three datasets. CIFAR100, TinyImageNet and ImageNet-R. CIFAR100 contains 60,000 images of $32 \\times 32$ pixels in 100 categories. Each category has 600 images, of which 500 are for the training set, and 100 are for the test set. We experiment with two settings, B0-n and B50-n. The former splits 100 classes into $n$ tasks,", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "28710", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5fc1dac2541b878b33a4ae6f17eed45101f963f3375d3362b89593ea76028575.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 104, + 117, + 356, + 296 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/68c76ab96796a0d53861c66ed96a3e434a3c01621d1bf11651bbd215f8d9b009.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 356, + 117, + 609, + 297 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/09d896fa62c4964f2a353a78e90082618608d797913106feb397f67719e05963.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 609, + 117, + 861, + 297 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/df95b220d179064318f406d516e71fd714fddeab26fb2cfbd4b2bcdcb40999e6.jpg", + "image_caption": [ + "Figure 3. Comparison of our method with other SOTA baselines on CIFAR100 and Tiny-ImageNet under the conventional CIL setting." + ], + "image_footnote": [], + "bbox": [ + 102, + 296, + 354, + 460 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/54928a2f405f63f3ca0218c5723973effad0d62dcd2e329ad50350010bf0ac2a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 357, + 296, + 607, + 460 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bf7495e757d995760444f4bfcac38caef78453cbdbd5fc4c3c8c8d308c775fab.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 611, + 296, + 861, + 460 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "while the latter trains on 50 classes first and then distributes the other 50 classes across 5/10 tasks.", + "bbox": [ + 75, + 506, + 468, + 536 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tiny-ImageNet contains 200 classes out of the original 1000 classes in ImageNet, with 550 images per class, of which 500 are in the training set and 50 are in the test set. The images are down-sampled to $64 \\times 64$ pixels, which makes them easier to process and analyze. We train the first task in half 100 classes and split the other 100 classes into 5/10/20 tasks following [89].", + "bbox": [ + 75, + 537, + 468, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ImageNet-R [23] contains 200 classes of images, which are included in the original ImageNet 1000 classes. However, many images are newly added and have various styles, such as sketch, painting, misc, etc. The dataset poses a great challenge for continual learning, as it has a wide diversity of image categories and styles and an uneven distribution of samples ranging from 45 to 500 per category. We follow [71] to split the dataset into 10 tasks, each containing 20 classes.", + "bbox": [ + 75, + 643, + 468, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Few-shot CIL, we use CIFAR100 and miniImageNet [49] following the split proposed by [62]. For both datasets, we partition the data into two parts: base session and incremental sessions. The base session comprises 60 classes with all data available, while the incremental session follows a 5-way 5-shot setting, meaning that each session consists of only 5 classes with 5 samples each.", + "bbox": [ + 75, + 779, + 468, + 885 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In both conventional and few-shot scenarios, we com", + "bbox": [ + 96, + 886, + 468, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "pare our method with some of the current state-of-the-art approaches, including conventional methods [7, 15, 16, 24, 29, 48, 53, 74, 76, 82, 89], pre-trained and prompt-based methods [55, 71, 72], and some specifically designed methods [12, 53, 62, 79] for few-shot scenarios. Additionally, we compare with a linear probe baseline, wherein features obtained from the image encoder are connected to a classifier for classification. We also consider the Zero-shot approach, where the generated text is directly used for classification without further fine-tuning.", + "bbox": [ + 496, + 506, + 890, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We follow BLIP2 [27] to use the EVA-CLIP [59] pre-trained ViT-g/14 and BLIP2 pretrained Qfomer. We also used the MiniGPT-4 pre-trained projection layer checkpoint as our initial parameters. Under the many-shot \"B0\" setting, we employ a learning rate of 3e-7 and use a scheduler with cosine decay. The total training process consists of 2 epochs only. In B50 or B100 settings, we first train the linear layer with a learning rate of 3e-6 on the base classes, and then on the subsequent tasks, we adopt a lower learning rate of 3e-7, both employing a cosine decay scheduler. For the few-shot setting, we use a learning rate of 3e-6 for both base task and incremental tasks. We train one epoch for the base task and two epochs for incremental tasks.", + "bbox": [ + 496, + 688, + 892, + 898 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Continual CLIP L2P DualPrompt CODA-Prompt GMM", + "bbox": [ + 243, + 89, + 712, + 104 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "28711", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5f239806a72dca730dddc92934ec918d1ed8fd2713199e779dc667914d97d49a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
012345678PD↓
iCaRL [48]61.3146.3242.9437.6330.4924.0020.8918.8017.2144.10
EEIL [7]61.3146.5844.0037.2933.1427.1224.1021.5719.5841.73
LUCIR [24]61.3147.8039.3131.9125.6821.3518.6717.2414.1747.14
TOPIC [62]61.3150.0945.1741.1637.4835.5232.1929.4624.4236.89
CEC [79]72.0066.8362.9759.4356.7053.7351.1949.2447.6324.37
F2M [53]72.0567.4763.1659.7056.7153.7751.1149.2147.8424.21
MetaFSCIL [12]72.0467.9463.7760.2957.5855.1652.9050.7949.1922.85
Entropy-reg [31]71.8467.1263.2159.7757.0153.9551.5549.5248.2123.63
L2P* [72]94.1287.2080.9975.6770.9466.7663.1159.8156.8337.29
DualPrompt* [71]93.9786.8580.6775.3170.6166.4462.7759.5856.8037.17
CODA-Prompt* [55]95.3788.8682.6977.8774.4770.1666.4663.7361.1434.23
Zero-shot58.0858.9557.7657.8958.1957.4256.2654.8254.953.13
GMM (Ours)89.3588.4086.1185.0783.6181.3578.9777.3475.1814.17
", + "bbox": [ + 119, + 87, + 854, + 380 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Comparison results of our method with other SOTA baselines on mini-ImageNet under the few-shot CIL setting. Phase 0 is the base task with all samples available, and the following 1-8 phases are the incremental 5-way 5-shot tasks. The reported accuracy is the test result on all seen classes after each session of training. \"PD\" represents Performance Prop between session 0 and session 8, with lower values indicating lower forgetting. * indicates our re-implementation based on PILOT [58].", + "bbox": [ + 75, + 388, + 893, + 446 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Experiments on Conventional CIL", + "text_level": 1, + "bbox": [ + 76, + 470, + 377, + 488 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Table 1, we can see that our method outperforms all conventional methods by a large margin, including ResNet-based method DER and ViT-based DyTox. Note that without exemplar, our performance is a bit lower than Dual-Prompt and CODA-Prompt at B100-5 setting. We argue that their performance is mainly due to the backbone pretrained on ImageNet-21K, which largely overlaps with CIFAR100 and Tiny-ImageNet. Another interesting observation is that our method has better performance than all baselines under longer sequence settings (B100-10, B100-20). We believe this is because generative models do not rely on classification heads, making them less prone to bias toward the current task, resulting in less forgetting of past tasks. The Linear probe setting performs less than our method, indicating that our main contribution is not from the Large pre-trained ViT but the generation pipeline. In addition, the Zero-shot performance is superior to many traditional baselines, meaning that the Generative Multi-modal Models are indeed efficient Class-Incremental learners, but its output is less concise without fine-tuning (see Fig. 4).", + "bbox": [ + 75, + 501, + 472, + 805 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Fig. 3, we compared our approach with some pretrained models on CIFAR100 and Tiny-ImageNet in terms of last task accuracy (all baselines are based on PILOT [58] and use 2000 exemplars). It can be observed that our method does not outperform other approaches in the initial tasks (0-2) and short sequence settings (B0-5, B100-5).", + "bbox": [ + 75, + 809, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This is because we do not rely on a supervised ImageNet-21K pre-trained backbone. Besides, we trained each task for only 1-2 epochs to ensure efficiency without sacrificing generalization. However, our method exhibits significant advantages in long sequences and later tasks. For instance, under the CIFAR100 B0-20 setting, we outperform CODA-Prompt by 10 points and DualPrompt by 7 points.", + "bbox": [ + 496, + 472, + 893, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Experiments on Few-shot CIL", + "text_level": 1, + "bbox": [ + 498, + 589, + 769, + 606 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Table 2, we compare our method with several baselines in the few-shot setting on mini-ImageNet. The evaluation metric is the model's accuracy across all the classes it has encountered so far. Our method outperforms conventional methods by a substantial margin in the final task, achieving an increase of more than $26\\%$ . Furthermore, we surpass the best discriminative pre-trained approach CODA-Prompt by more than $14\\%$ points. It's important to note that our accuracy in the first task (89.35) might not be as high as CODA-Prompt (95.37). However, in subsequent sessions, we consistently perform better than CODA-Prompt due to our ability to learn new tasks and retain knowledge of old tasks simultaneously.", + "bbox": [ + 496, + 612, + 893, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Table 3, our method outperforms all other baselines on the CIFAR100 dataset of the few-shot setting, achieving a remarkably lower Performance Drop (PD) of 10.06. Furthermore, the Zero-shot baseline can achieve a very low PD due to the absence of forgetting. However, its overall performance is not very satisfying, as the length and con", + "bbox": [ + 496, + 810, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "28712", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9eb814575495c5472761290ad93bc8cfc615a60b8c99f2b6b6049b8f77342136.jpg", + "image_caption": [], + "image_footnote": [ + "This is a photo of a statue of a bird with a long beak and a long neck. It is made of stone and has a blue sky background.", + "This is a pelican." + ], + "bbox": [ + 101, + 89, + 305, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0326910f8b5be09208084d044d3a89818adb9175b3fb5d387183603fb7634fe0.jpg", + "image_caption": [ + "Figure 4. Visual comparison examples of our method against frozen MiniGPT-4 [88] (Zero-shot). Text with a gray background is generated by MiniGPT-4 based on the image, while text with a orange background represents our method's output. The ground-truth labels are displayed on the left side of each row of images. All images displayed here are random sampled from ImageNet-R." + ], + "image_footnote": [ + "This is a painting of a fire truck. The fire truck is red and has a ladder on the side. The ladder is painted in white. The fire truck has a red and white stripe on the side.", + "This is a fire engine." + ], + "bbox": [ + 101, + 199, + 305, + 306 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/14fd87dd31df3cc82c57882b4e2ce0b2f5b8a5823f7f3cf9c39fcba719bc9230.jpg", + "image_caption": [], + "image_footnote": [ + "This is a sculpture of a bird with its wings spread wide, as if it is flying. The bird is made of bronze and has a green patina. It is located in a park.", + "This is a pelican." + ], + "bbox": [ + 500, + 90, + 681, + 195 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/37cebaae52d016cdf65733a87f2a55a88532afcd0422f8209dfcf44daef9238c.jpg", + "image_caption": [], + "image_footnote": [ + "This is a toy car with a duck in the driver's seat. The car is red and has a fire truck on the back. The duck is wearing a firefighter's helmet and has a fire hose in its hand.", + "This is a fire truck." + ], + "bbox": [ + 501, + 199, + 683, + 305 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/523d670f10c74d96c889250453c0c41b1d9cbef9027b9c3729c8e23cf422ef57.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method048PD↓
iCaRL [48]64.1027.9313.7350.37
EEIL [7]64.1028.9615.8548.25
LUCIR [24]64.1031.6113.5450.56
TOPIC [62]64.1040.1129.3734.73
CEC [79]73.0758.0949.1423.93
F2M [53]71.4557.7649.3522.06
MetaFSCIL [12]74.5059.4849.9724.53
Entropy-reg [31]74.4059.7150.1424.26
L2P* [72]91.2268.6654.8936.33
DualPrompt* [71]91.0868.4554.6736.41
CODA-Prompt* [55]93.5571.9159.3234.23
Zero-shot74.1372.5967.936.20
GMM (Ours)91.5385.6581.4710.06
", + "bbox": [ + 84, + 373, + 460, + 654 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Comparison results of our method with other SOTA baselines on CIFAR100 under the few-shot CIL setting.", + "bbox": [ + 76, + 664, + 468, + 693 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tent of its output are inconsistent and unpredictable without fine-tuning.", + "bbox": [ + 75, + 700, + 468, + 732 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Visualizations", + "text_level": 1, + "bbox": [ + 76, + 739, + 220, + 755 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Fig. 4, we present some comparison examples of our methods against GMM without fine-tuning [88]. We can see that GMM without fine-tuning provides an intuitive description of the overall image content with varying lengths of output text. However, it tends to recognize only broad categories (e.g., biar, car) and struggles with fine-grained categorization (e.g., pelican, truck). The descriptions are sometimes somewhat repetitive (e.g., first fire engine). In contrast, our fine-tuned method accurately identifies the im", + "bbox": [ + 75, + 763, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "age's real category, even if there are occasional discrepancies with the true labels (e.g., \"fire engine\" vs. \"fire truck\"). Besides, with the assistance of the text encoder during the testing phase, our model can achieve correct classification results even when predicting similar but not identical text.", + "bbox": [ + 496, + 376, + 890, + 452 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 474, + 617, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose GMM to use generative models for class-incremental learning. By fine-tuning the Generative Multi-modal Model (GMM), we directly generate the label text of the images to be classified. Then we select the label most similar to the generated text by its features. Our experiments demonstrate that this method, which does not require a classification head, is highly effective in addressing classification biases in continual learning.", + "bbox": [ + 496, + 502, + 890, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Since we are the first to introduce generative models to class-incremental learning, the overall design of our method is embarrassingly simple. We believe that with more focused efforts in this direction, there will be significant advancements in the field of continual learning.", + "bbox": [ + 496, + 630, + 890, + 705 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Broader impact. We believe that introducing GMM into continual learning (CL) is both necessary and urgent. With the rapid development of GMM, we can leverage their capabilities to improve the performance of continual learning. Besides, integrating CL methods into the training process of GMM could significantly reduce training costs.", + "bbox": [ + 496, + 712, + 890, + 803 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This work is funded by NSFC (NO. 62206135, 62225604), Young Elite Scientists Sponsorship Program by CAST (2023QNRC001), and the Fundamental Research Funds for the Central Universities (Nankai University, 070-63233085). Computation is supported by the Supercomputing Center of Nankai University.", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "28713", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hongjoon Ahn, Jihwan Kwak, Subin Lim, Hyeonsu Bang, Hyojun Kim, and Taesup Moon. Ss-il: Separated softmax for incremental learning. In ICCV, 2021. 2", + "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 3", + "[3] Rahaf Aljundi, Punarjay Chakravarty, and Tinne Tuytelaars. Expert gate: Lifelong learning with a network of experts. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3366-3375, 2017. 2", + "[4] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2(3):4, 2023. 3", + "[5] Hedi Ben-Younes, Rémi Cadene, Matthieu Cord, and Nicolas Thome. Mutan: Multimodal tucker fusion for visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 2612-2620, 2017. 3", + "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2", + "[7] Francisco M Castro, Manuel J Marín-Jiménez, Nicolas Guil, Cordelia Schmid, and Karteek Alahari. End-to-end incremental learning. In Proceedings of the European conference on computer vision (ECCV), pages 233-248, 2018. 2, 5, 6, 7, 8", + "[8] Jun Chen, Han Guo, Kai Yi, Boyang Li, and Mohamed Elhoseiny. Visualgpt: Data-efficient adaptation of pretrained language models for image captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18030-18040, 2022. 3", + "[9] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 2", + "[10] Xinlei Chen*, Saining Xie*, and Kaiming He. An empirical study of training self-supervised vision transformers. arXiv preprint arXiv:2104.02057, 2021. 2", + "[11] Ali Cheraghian, Shafin Rahman, Pengfei Fang, Soumava Kumar Roy, Lars Petersson, and Mehrtash Harandi. Semantic-aware knowledge distillation for few-shot class-incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2534-2543, 2021. 2", + "[12] Zhixiang Chi, Li Gu, Huan Liu, Yang Wang, Yuanhao Yu, and Jin Tang. Meta-scil: A meta-learning approach for few-shot class incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14166-14175, 2022. 6, 7, 8" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2023. 3", + "[14] Matthias Delange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Ales Leonardis, Greg Slabaugh, and Tinne Tuytelaars. A continual learning survey: Defying forgetting in classification tasks. TPAMI, 2021. 1, 2", + "[15] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In ECCV, 2020. 2, 6", + "[16] Arthur Douillard, Alexandre Ramé, Guillaume Couairon, and Matthieu Cord. Dytox: Transformers for continual learning with dynamic token expansion. In CVPR, 2022. 2, 5, 6", + "[17] Chrisantha Fernando, Dylan Banarse, Charles Blundell, Yori Zwols, David Ha, Andrei A Rusu, Alexander Pritzel, and Daan Wierstra. Pathnet: Evolution channels gradient descent in super neural networks. arXiv preprint arXiv:1701.08734, 2017. 2", + "[18] Rui Gao and Weiwei Liu. Ddgr: Continual learning with deep diffusion-based generative replay. In ICML, 2023. 2", + "[19] Meng-Hao Guo, Cheng-Ze Lu, Zheng-Ning Liu, Ming-Ming Cheng, and Shi-Min Hu. Visual attention network. Computational Visual Media, 9(4):733-752, 2023. 1", + "[20] Tyler L Hayes, Kushal Kafle, Robik Shrestha, Manoj Acharya, and Christopher Kanan. Remind your neural network to prevent catastrophic forgetting. In ECCV, 2020. 2", + "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3", + "[22] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2", + "[23] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, Dawn Song, Jacob Steinhardt, and Justin Gilmer. The many faces of robustness: A critical analysis of out-of-distribution generalization. ICCV, 2021. 6", + "[24] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In CVPR, 2019. 2, 5, 6, 7, 8", + "[25] Xinting Hu, Kaihua Tang, Chunyan Miao, Xian-Sheng Hua, and Hanwang Zhang. Distilling causal effect of data in class incremental learning. In CVPR, 2021. 2", + "[26] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. 2, 5", + "[27] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with" + ], + "bbox": [ + 501, + 92, + 890, + 902 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "28714", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 3, 6", + "[28] Wenbo Li, Pengchuan Zhang, Lei Zhang, Qiuyuan Huang, Xiaodong He, Siwei Lyu, and Jianfeng Gao. Object-driven text-to-image synthesis via adversarial training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12174-12182, 2019. 3", + "[29] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 2, 5, 6", + "[30] Zhizhong Li and Derek Hoiem. Learning without forgetting. TPAMI, 2018. 2", + "[31] Huan Liu, Li Gu, Zhixiang Chi, Yang Wang, Yuanhao Yu, Jun Chen, and Jin Tang. Few-shot class-incremental learning via entropy-regularized data-free replay. In European Conference on Computer Vision, pages 146–162. Springer, 2022. 7, 8", + "[32] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 2, 3", + "[33] Ming Liu, Yuxiang Wei, Xiaohe Wu, Wangmeng Zuo, and Lei Zhang. Survey on leveraging pre-trained generative adversarial networks for image editing and restoration. Science China Information Sciences, 66(5):151101, 2023. 1", + "[34] Wei Liu, Sihan Chen, Longteng Guo, Xinxin Zhu, and Jing Liu. Cptr: Full transformer network for image captioning. arXiv preprint arXiv:2101.10804, 2021. 3", + "[35] Xialei Liu, Marc Masana, Luis Herranz, Joost Van de Weijer, Antonio M Lopez, and Andrew D Bagdanov. Rotate your networks: Better weight consolidation and less catastrophic forgetting. In ICPR, 2018. 2", + "[36] Xialei Liu, Yu-Song Hu, Xu-Sheng Cao, Andrew D Bagdanov, Ke Li, and Ming-Ming Cheng. Long-tailed class incremental learning. In European Conference on Computer Vision, pages 495-512. Springer, 2022. 2", + "[37] Xialei Liu, Xusheng Cao, Haori Lu, Jia-wen Xiao, Andrew D Bagdanov, and Ming-Ming Cheng. Class incremental learning with pre-trained vision-language models. arXiv preprint arXiv:2310.20348, 2023. 3", + "[38] Yaoyao Liu, Bernt Schiele, and Qianru Sun. Adaptive aggregation networks for class-incremental learning. In CVPR, 2021. 2", + "[39] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019. 3", + "[40] Arun Mallya, Dillon Davis, and Svetlana Lazebnik. Piggyback: Adapting a single network to multiple tasks by learning to mask weights. In ECCV, 2018. 2", + "[41] Marc Masana, Xialei Liu, Bartlomiej Twardowski, Mikel Menta, Andrew D Bagdanov, and Joost Van De Weijer. Class-incremental learning: survey and performance evaluation on image classification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5):5513-5533, 2022. 1", + "[42] Pratik Mazumder, Pravendra Singh, and Piyush Rai. Few-shot lifelong learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2337-2345, 2021. 2" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Michael McCloskey and Neal J Cohen. Catastrophic interference in connectionist networks: The sequential learning problem. In *Psychology of learning and motivation*, pages 109–165. Elsevier, 1989. 1", + "[44] OpenAI. Gpt-4 technical report, 2023. 2, 3", + "[45] Haoxuan Qu, Hossein Rahmani, Li Xu, Bryan Williams, and Jun Liu. Recent advances of continual learning in computer vision: An overview. arXiv preprint arXiv:2109.11369, 2021. 1", + "[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3, 4", + "[47] Jathushan Rajasegaran, Munawar Hayat, Salman Khan, Fahad Shahbaz Khan, and Ling Shao. Random path selection for incremental learning. In NIPS, 2019. 2", + "[48] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In CVPR, 2017. 2, 5, 6, 7, 8", + "[49] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 2, 6", + "[50] Jonathan Schwarz, Wojciech Czarnecki, Jelena Luketina, Agnieszka Grabska-Barwinska, Yee Whye Teh, Razvan Pascanu, and Raia Hadsell. Progress & compress: A scalable framework for continual learning. In ICML, 2018. 2", + "[51] Joan Serra, Didac Suris, Marius Miron, and Alexandros Karatzoglou. Overcoming catastrophic forgetting with hard attention to the task. In ICML, 2018. 2", + "[52] Yijia Shao, Yiduo Guo, Dongyan Zhao, and Bing Liu. Class incremental learning based on label generation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 1263-1276, Toronto, Canada, 2023. Association for Computational Linguistics. 2", + "[53] Guangyuan Shi, Jiaxin Chen, Wenlong Zhang, Li-Ming Zhan, and Xiao-Ming Wu. Overcoming catastrophic forgetting in incremental few-shot learning by finding flat minima. Advances in neural information processing systems, 34: 6747-6761, 2021. 3, 6, 7, 8", + "[54] Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual learning with deep generative replay. Advances in neural information processing systems, 30, 2017. 2", + "[55] James Seale Smith, Leonid Karlinsky, Vyshnavi Gutta, Paola Cascante-Bonilla, Donghyun Kim, Assaf Arbelle, Rameswar Panda, Rogerio Feris, and Zsolt Kira. Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11909-11919, 2023. 2, 3, 5, 6, 7, 8", + "[56] Boyuan Sun, Yuqi Yang, Le Zhang, Ming-Ming Cheng, and Qibin Hou. Corrmatch: Label propagation via correlation matching for semi-supervised semantic segmentation. IEEE Computer Vision and Pattern Recognition (CVPR), 2024. 1" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "28715", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[57] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7464-7473, 2019. 3", + "[58] Hai-Long Sun, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Pilot: A pre-trained model-based continual learning toolbox. arXiv preprint arXiv:2309.07117, 2023. 7", + "[59] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389, 2023. 6", + "[60] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019. 3", + "[61] Xiaoyu Tao, Xinyuan Chang, Xiaopeng Hong, Xing Wei, and Yihong Gong. Topology-preserving class-incremental learning. In ECCV, 2020. 2", + "[62] Xiaoyu Tao, Xiaopeng Hong, Xinyuan Chang, Songlin Dong, Xing Wei, and Yihong Gong. Few-shot class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12183-12192, 2020. 2, 6, 7, 8", + "[63] Vishal Thengane, Salman Khan, Munawar Hayat, and Fahad Khan. Clip model is an efficient continual learner. arXiv preprint arXiv:2210.03114, 2022. 3, 5", + "[64] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3", + "[65] Maria Tsimpoukelli, Jacob L Menick, Serkan Cabi, SM Eslami, Oriol Vinyls, and Felix Hill. Multimodal few-shot learning with frozen language models. Advances in Neural Information Processing Systems, 34:200-212, 2021. 3", + "[66] Gido M Van de Ven and Andreas S Tolias. Three scenarios for continual learning. NIPS Workshops, 2019. 1", + "[67] Liuyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. A comprehensive survey of continual learning: Theory, method and application. arXiv preprint arXiv:2302.00487, 2023. 1", + "[68] Liyuan Wang, Jingyi Xie, Xingxing Zhang, Mingyi Huang, Hang Su, and Jun Zhu. Hierarchical decomposition of prompt-based continual learning: Rethinking obscured suboptimality. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[69] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, et al. Vision: Large language model is also an open-ended decoder for vision-centric tasks. arXiv preprint arXiv:2305.11175, 2023. 3", + "[70] Xiao Wang, Guangyao Chen, Guangwu Qian, Pengcheng Gao, Xiao-Yong Wei, Yaowei Wang, Yonghong Tian, and Wen Gao. Large-scale multi-modal pre-trained models: A comprehensive survey. Machine Intelligence Research, 20 (4):447-482, 2023. 3", + "[71] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "prompting for rehearsal-free continual learning. In ECCV, 2022. 2, 3, 5, 6, 7, 8", + "[72] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer G. Dy, and Tomas Pfister. Learning to prompt for continual learning. CVPR, 2022. 2, 3, 5, 6, 7, 8", + "[73] Tz-Ying Wu, Gurumurthy Swaminathan, Zhizhong Li, Avinash Ravichandran, Nuno Vasconcelos, Rahul Bhotika, and Stefano Soatto. Class-incremental learning with strong pre-trained models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9601-9610, 2022. 3", + "[74] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In CVPR, 2019. 2, 6", + "[75] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In ICCV, 2019. 2", + "[76] Shipeng Yan, Jiangwei Xie, and Xuming He. Der: Dynamically expandable representation for class incremental learning. In CVPR, 2021. 2, 6", + "[77] Yang Yang, Zhiying Cui, Junjie Xu, Changhong Zhong, Wei-Shi Zheng, and Ruixuan Wang. Continual learning with bayesian model based on a fixed pre-trained feature extractor. Visual Intelligence, 1(1):5, 2023. 1", + "[78] Lu Yu, Bartlomiej Twardowski, Xialei Liu, Luis Herranz, Kai Wang, Yongmei Cheng, Shangling Jui, and Joost van de Weijer. Semantic drift compensation for class-incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6982-6991, 2020. 2", + "[79] Chi Zhang, Nan Song, Guosheng Lin, Yun Zheng, Pan Pan, and Yinghui Xu. Few-shot incremental learning with continually evolved classifiers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12455-12464, 2021. 3, 6, 7, 8", + "[80] Chang-Bin Zhang, Jia-Wen Xiao, Xialei Liu, Ying-Cong Chen, and Ming-Ming Cheng. Representation compensation networks for continual semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7053-7064, 2022. 2", + "[81] Gengwei Zhang, Liyuan Wang, Guoliang Kang, Ling Chen, and Yunchao Wei. Slca: Slow learner with classifier alignment for continual learning on a pre-trained model. arXiv preprint arXiv:2303.05118, 2023. 2, 3", + "[82] Bowen Zhao, Xi Xiao, Guojun Gan, Bin Zhang, and Shu-Tao Xia. Maintaining discrimination and fairness in class incremental learning. In CVPR, 2020. 6", + "[83] Hanbin Zhao, Yongjian Fu, Mintong Kang, Qi Tian, Fei Wu, and Xi Li. Mgmtf: Multi-grained slow vs. fast framework for few-shot class-incremental learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 2", + "[84] Da-Wei Zhou, Qi-Wei Wang, Zhi-Hong Qi, Han-Jia Ye, De-Chuan Zhan, and Ziwei Liu. Deep class-incremental learning: A survey. arXiv preprint arXiv:2302.03648, 2023. 1", + "[85] Da-Wei Zhou, Han-Jia Ye, De-Chuan Zhan, and Ziwei Liu. Revisiting class-incremental learning with pre-trained mod" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "28716", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "els: Generalizability and adaptivity are all you need, 2023. 2, 3", + "[86] Da-Wei Zhou, Yuanhan Zhang, Jingyi Ning, Han-Jia Ye, De-Chuan Zhan, and Ziwei Liu. Learning without forgetting for vision-language models. arXiv preprint arXiv:2305.19270, 2023. 2, 3", + "[87] Qiang Zhou, Chaohui Yu, Shaofeng Zhang, Sitong Wu, Zhibing Wang, and Fan Wang. Regionblip: A unified multimodal pre-training framework for holistic and regional comprehension. arXiv preprint arXiv:2308.02299, 2023. 3", + "[88] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 3, 8", + "[89] Fei Zhu, Xu-Yao Zhang, Chuang Wang, Fei Yin, and ChengLin Liu. Prototype augmentation and self-supervision for incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5871-5880, 2021. 2, 5, 6", + "[90] Kai Zhu, Yang Cao, Wei Zhai, Jie Cheng, and Zheng-Jun Zha. Self-promoted prototype refinement for few-shot class incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6801-6810, 2021. 3" + ], + "bbox": [ + 78, + 90, + 468, + 429 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "28717", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2024/Generative Multi-modal Models are Good Class Incremental Learners/a19d9ccd-1dff-4a86-83d4-a672741b4e0a_model.json b/2024/Generative Multi-modal Models are Good Class Incremental Learners/a19d9ccd-1dff-4a86-83d4-a672741b4e0a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8cd5507f64a6d8f6457e78a8224c264fea3f1ca0 --- /dev/null +++ b/2024/Generative Multi-modal Models are Good Class Incremental Learners/a19d9ccd-1dff-4a86-83d4-a672741b4e0a_model.json @@ -0,0 +1,2578 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.129, + 0.131, + 0.842, + 0.152 + ], + "angle": 0, + "content": "Generative Multi-modal Models are Good Class-Incremental Learners" + }, + { + "type": "text", + "bbox": [ + 0.175, + 0.18, + 0.794, + 0.217 + ], + "angle": 0, + "content": "Xusheng Cao\\(^{1}\\), Haori Lu\\(^{1}\\), Linlan Huang\\(^{1}\\), Xialei Liu\\(^{2,1*}\\), Ming-Ming Cheng\\(^{2,1}\\) \n\\(^{1}\\)VCIP, CS, Nankai University \\(^{2}\\)NKIARI, Shenzhen Futian" + }, + { + "type": "text", + "bbox": [ + 0.121, + 0.219, + 0.842, + 0.235 + ], + "angle": 0, + "content": "{caoxusheng, luhaori, huanglinlan}@mail.nankai.edu.cn, {xialei, cmm}@nankai.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.302, + 0.474, + 0.681 + ], + "angle": 0, + "content": "In class-incremental learning (CIL) scenarios, the phenomenon of catastrophic forgetting caused by the classifier's bias towards the current task has long posed a significant challenge. It is mainly caused by the characteristic of discriminative models. With the growing popularity of the generative multi-modal models, we would explore replacing discriminative models with generative ones for CIL. However, transitioning from discriminative to generative models requires addressing two key challenges. The primary challenge lies in transferring the generated textual information into the classification of distinct categories. Additionally, it requires formulating the task of CIL within a generative framework. To this end, we propose a novel generative multi-modal model (GMM) framework for class-incremental learning. Our approach directly generates labels for images using an adapted generative model. After obtaining the detailed text, we use a text encoder to extract text features and employ feature matching to determine the most similar label as the classification prediction. In the conventional CIL settings, we achieve significantly better results in long-sequence task scenarios. Under the Few-shot CIL setting, we have improved by at least \\(14\\%\\) accuracy over all the current state-of-the-art methods with significantly less forgetting. Our code is available at https://github.com/DoubleClass/GMM." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.712, + 0.21, + 0.727 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.738, + 0.47, + 0.875 + ], + "angle": 0, + "content": "Deep neural networks [19, 33, 56] have made remarkable strides in numerous applications, primarily owing to the vast amounts of data and computational resources at their disposal. Nonetheless, these accomplishments are predominantly contingent on having access to all the required data simultaneously for training on various tasks. In cases where data is acquired incrementally, these networks often encounter the challenge of catastrophic forgetting [43]. Hence, the capacity to seamlessly incorporate new knowl" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.268, + 0.887, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.34, + 0.89, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.48, + 0.895, + 0.579 + ], + "angle": 0, + "content": "Figure 1. Illustration of conventional discriminative models for class-incremental learning (CIL) and our generative multi-modal models (GMM) for CIL. Discriminative models pose potential risk of classifier bias toward current task with network expansion. Our GMM framework consists generation and classification phases. And It is adapted to CIL based on the similarity of generated text and the true category names." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.609, + 0.894, + 0.685 + ], + "angle": 0, + "content": "edge while retaining previously acquired knowledge is a highly desirable attribute for future artificial intelligence systems. Continual learning [45, 67, 77, 84] is a subject of study aimed at advancing the evolution of neural networks toward this goal." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.687, + 0.895, + 0.882 + ], + "angle": 0, + "content": "Numerous studies have delved into continual learning, categorizing their approaches into three main groups [14]: rehearsal-based, architecture-based, and regularization-based methods. Additionally, hybrid methods are gaining popularity as they combine insights from different perspectives. Within this research landscape, three primary scenarios [66] have received extensive attention, with class-incremental learning (CIL) [41] being one of the most demanding settings. In our work, we concentrate on CIL, wherein each task comprises a distinct set of classes, and the primary challenge is to enable the network to recognize new classes without forgetting knowledge of previously encountered ones." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "The majority of existing research on class-incremental" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.216, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28706" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.302 + ], + "angle": 0, + "content": "learning (CIL) focuses on training models from scratch, relying solely on data from the current tasks [7, 15, 16, 24, 26, 29, 48, 74, 76, 89]. In contrast, humans accumulate knowledge over extended periods, drawing upon a wealth of prior world knowledge. Consequently, there has been a growing interest in pre-trained models for CIL [55, 71, 72, 81, 86], harnessing the knowledge acquired from extensive, pre-existing datasets to address the immediate tasks at hand. For instance, prompt-based methods [55, 71, 72] utilize prompt-tuning to summarize task-specific knowledge from prior pretrained knowledge. SLCA [81] and ADAM [85], on the other hand, solely fine-tuning the pre-trained model to adapt the pre-existing knowledge into the objectives of immediate tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.306, + 0.473, + 0.518 + ], + "angle": 0, + "content": "To tackle image classification downstream tasks, pretrained models are traditionally derived from discriminative tasks, like supervised learning on datasets such as ImageNet-21K [49], or they may stem from self-supervised learning efforts [6, 9, 10, 22]. However, in our research, we venture into the paradigm of generative multi-modal models to address image classification tasks. Generative models like GPT4 [44] and LLaVa [32] have garnered significant attention in recent years due to their capacity to produce highly informative descriptions of input images. On the one hand, it can harness the wealth of semantic correspondences between texts and images, while on the other hand, there's no requirement to expand the classifier with each new task, unlike in the case of discriminative models for CIL." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.521, + 0.473, + 0.748 + ], + "angle": 0, + "content": "Nonetheless, harnessing the knowledge from pre-trained generative models for downstream class-incremental learning (CIL) tasks presents a nontrivial endeavor. The primary challenge lies in transferring the generated textual information into the classification of distinct categories. Additionally, there's the task of formulating CIL within a generative framework, which poses a second significant challenge. Shao et al. [52] presents the VAG system, which formulates CIL as a continual label generation problem, preserving the language model's ability to learn new classes. However, it only works in the field of Natural Language Processing (NLP), which is inherently suited to Large Language Models (LLM). To the best of our knowledge, we are the first to apply this generative approach to incremental learning in the field of image classification." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.903 + ], + "angle": 0, + "content": "In this work, we propose Generative Multi-modal Models (GMM) for class-incremental learning. As illustrated in Fig. 1 (a), conventional discriminative methods extract image features with a network backbone, then forward them to a classifier to obtain the probability of the image belonging to each label, with the label having the highest probability being the output of the discriminative models. While in Fig. 1 (b), we adopt a generative approach, which directly produces a descriptive sentence for the given image, which is then compared with the actual label texts with a" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.198 + ], + "angle": 0, + "content": "text encoder. The most similar label becomes the predicted result of our generative model. This approach allows us to leverage the rich pre-training knowledge in generative multi-modal models while avoiding the use of the expanded classification head, which mitigates the risk of the model bias towards the current task and reduces catastrophic forgetting." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.198, + 0.793, + 0.213 + ], + "angle": 0, + "content": "The main contributions of this paper are:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.216, + 0.892, + 0.26 + ], + "angle": 0, + "content": "- We propose a novel generative approach (GMM) to address class-incremental learning by leveraging multimodal models." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.261, + 0.895, + 0.35 + ], + "angle": 0, + "content": "- We reformulate GMM for image classification and adapt it for the downstream benchmarks. Without an expanded classification head like in discriminative models, our model significantly mitigates the issue of bias towards current tasks, resulting in significantly reduced forgetting in CIL." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.352, + 0.895, + 0.397 + ], + "angle": 0, + "content": "- Our model achieves state-of-the-art performance across multiple datasets in both conventional and few-shot CIL settings." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.216, + 0.895, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.411, + 0.642, + 0.427 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.436, + 0.755, + 0.453 + ], + "angle": 0, + "content": "2.1. Class-Incremental Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.459, + 0.895, + 0.822 + ], + "angle": 0, + "content": "In Class-Incremental Learning, tasks arrive sequentially, and each class is exclusive to a specific task without any overlap. The goal is to acquire knowledge from new classes while preserving information from previously encountered classes. There are three primary branches in CIL [14], including rehearsal-based, architecture-based, and regularization-based methods. Rehearsal-based methods [1, 7, 48, 75] store a small set of data derived from old classes to represent knowledge from previous tasks. These exemplar data can be either original data [48], generative data [18, 54] or hidden features [20]. Architecture-based methods focus on modifying network architecture to alleviate forgetting. Approaches include learning redundant network architecture [17, 47], learning different expert networks [3, 50] or parameters [38, 40, 51] for each task, dynamically expanding network parameters to accumulate incremental knowledge [76]. Regularization-based methods introduce an additional regularization term to restrict network updates when adapting to new tasks. In such cases, EWC [26], SDC [78] and Rotated-EWC [35] expect that parameters essential for the old tasks should not be updated excessively. Moreover, from the perspective of network output consistency, numerous studies [25, 30, 36, 61, 80] incorporate distillation to prevent forgetting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Few-shot CIL Few-shot Class-Incremental Learning (FSCIL) [42, 62] explores few-shot learning in an incremental context, with all data samples available for base session and very limited data in each incremental session. Some FSCIL methods [11, 62, 83] train the model in both base and incre" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28707" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.182 + ], + "angle": 0, + "content": "mental sessions, aiming to mitigate overfitting challenges caused by the limited data in incremental learning. Other strategies [53, 79, 90] primarily train the model in the base session and make minimal adjustments in the incremental sessions, thereby reducing forgetting but may come at the cost of decreased precision in the incremental sessions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.193, + 0.325, + 0.209 + ], + "angle": 0, + "content": "2.2. Pre-trained models for CIL" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.217, + 0.474, + 0.55 + ], + "angle": 0, + "content": "There are many methods [63, 72, 73] having shown that pretrained models are effective for continual learning. One main branch trains a set of prompts to retain previous knowledge [55, 68, 71, 72]. A selected subset of prompts are fed into the model during forward to prompt model the past knowledge. Additionally, methods like SLCA [81] and ADAM [85] fine-tune pre-trained models, achieving impressive results with less forgetting. Continual-CLIP [63] demonstrates that the CLIP [46] model is capable of performing continual learning without any extra training. This highlights the significant potential of multi-modal pretrained models in the realm of continual learning. Inspired by this, many methods [37, 86] employ CLIP as the backbone to utilize the multi-modal information. However, if not using a classifier directly, these approaches need to utilize expanded text features to calculate distances with image features for classification. This will exacerbate the model's bias towards current data, consequently leading to the forgetting of previously acquired knowledge. To avoid this bias, we use a generative model to directly generate prediction text. The fixed text decoder will function as the classifier, significantly alleviating the bias." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.56, + 0.304, + 0.577 + ], + "angle": 0, + "content": "2.3. Vision Language Models" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In recent years, vision-language multi-modal models have made significant progress and achieved impressive results in various downstream tasks [5, 28, 34, 70]. Traditional vision-language models employ different types of encoders to extract information from vision and language models, including single-stream [57], dual-stream [39] and fusion [60] encoders. A key aspect of vision-language models is the alignment of multi-modal features. CLIP [46], for example, extracts image and text features separately using respective encoders and enforces alignment through a contrastive loss, ensuring alignment between positive image-text pairs in the feature space. VisualGPT [8] and Frozen [65] leverage pre-trained models as encoders for visual-language tasks. From then on, the utilization of pre-trained models in vision-language tasks became more and more popular. For instance, Flamingo [2] and BLIP-2 [27] align the pre-trained image and text encoders employing gated cross-attention and Q-Former, respectively. Furthermore, LLaVA [32] and MiniGPT-4 [88] leverage more robust Large Language Models (LLM) [13, 64] as text encoders, while only training a projection layer for alignment. With" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.138 + ], + "angle": 0, + "content": "the increasing popularity of LLM, an increasing number of studies [4, 69, 87] explore the potential of multi-modal LLMs for vision-language tasks." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.152, + 0.593, + 0.167 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.178, + 0.892, + 0.239 + ], + "angle": 0, + "content": "In this section, we introduce the preliminaries of class-incremental learning and generative multi-modal models. Then, we present our approach to leverage generative models for CIL and the corresponding learning process." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.249, + 0.642, + 0.264 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.276, + 0.892, + 0.412 + ], + "angle": 0, + "content": "Class-Incremental Learning. Given N tasks \\( T = \\{T_{1}, T_{2}, \\dots, T_{N}\\} \\), the goal of class-incremental learning is to learn each task \\( T_{t} \\) with its associated data \\( \\{\\mathbf{X}_{t}, \\mathbf{Y}_{t}\\} \\) in a sequential order. For each task, it contains samples \\( \\{\\mathbf{x}_i, \\mathbf{y}_i\\} \\), \\( i = 1, \\dots, n_t \\), where \\( \\mathbf{x}_i \\) is the images and \\( \\mathbf{y}_i \\) is the corresponding one-hot labels. Typically, \\( \\mathbf{X}_i \\cap \\mathbf{X}_j = \\emptyset, \\forall i \\neq j \\). At inference, the model is tested on all seen tasks without task IDs. In some scenarios, fixed memory storage is set to keep a few samples of previous tasks to prevent forgetting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.412, + 0.892, + 0.579 + ], + "angle": 0, + "content": "Normally, a CIL model consists of a feature extractor and a classifier head \\( F = \\{f_{\\theta}, \\mathcal{H}_{\\phi}\\} \\) which are parameterized by \\( \\{\\theta, \\phi\\} \\). In traditional class-incremental learning, \\( \\theta \\) is usually a modified ResNet [21] with all parameters tunable. In pre-trained or prompt-based methods, \\( \\theta \\) represents fewer trainable parameters like a linear adaptor or a couple of prompts. \\( \\phi \\) is a linear classifier head projecting image features to probability predictions, which has to be expanded for each new task in order to make predictions for the new classes. The conventional Cross Entropy loss is often used for updating \\( \\theta \\) and \\( \\phi \\), which for task \\( t \\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.51, + 0.604, + 0.892, + 0.655 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {C E}} \\left(\\mathbf {X} _ {t}, \\mathbf {Y} _ {t}; \\theta , \\phi\\right) = - \\frac {1}{n _ {t}} \\sum_ {i = 1} ^ {n _ {t}} \\mathbf {y} _ {i} \\cdot \\log \\mathcal {H} (f (\\mathbf {x} _ {i}; \\theta); \\phi). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.656, + 0.892, + 0.732 + ], + "angle": 0, + "content": "In the continual learning process, the parameter \\(\\phi\\) can easily deviate to the data of the current task due to the absence or scarcity of old samples in the previous tasks, resulting in forgetting the previously acquired knowledge and deteriorating the overall performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Generative Multi-Modal Models (GMM). Multi-modal models have demonstrated exceptional performance in generating detailed image descriptions by incorporating both visual and textual information. Notably, GPT-4 [44] stands out as an advanced model proficient in generating comprehensive image descriptions and providing explanations for the depicted content. Furthermore, MiniGPT-4 [88] proposes a two-stage fine-tuning process that aligns image features and large language models, enabling LLaMa [64] to recognize images and conduct further dialogue based on the image content." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28708" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.09, + 0.094, + 0.531, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.561, + 0.094, + 0.871, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.367, + 0.895, + 0.424 + ], + "angle": 0, + "content": "Figure 2. The overview structure of our proposed method. The conceptual illustration of the generative multi-modal model (GMM) is shown on the left. In order to adapt this model for CIL, We have to turn the GMM model for classification and further adapt to our objective benchmark for learning (see Sec. 3.2). On the right side, we demonstrate how the final evaluation is taken for CIL with all seen classes. Text encoder is used for obtaining the embeddings for similarity prediction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.45, + 0.473, + 0.618 + ], + "angle": 0, + "content": "As shown in Fig. 2, these models consist of an encoder \\( f_{enc} \\) to generate content embeddings including image embedding and text embedding, which is further used as the input of an auto-aggressive decoder \\( f_{dec} \\) to generate image descriptions. Input image \\( \\mathbf{x}_i \\) is encoded as an image embedding \\( e_i \\), and question embeddings \\( \\mathbf{q} \\) with \\( q_1, \\dots, q_l \\) can be concatenated with the image embedding together to generate answer embeddings \\( \\mathbf{s} \\) with \\( s_1, \\dots, s_m \\). The output tokens are generated one by one with the condition of previously generated tokens. For instance, \\( s_m \\) is generated with all previous \\( m - 1 \\) tokens (see the decoder in Fig. 2)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.629, + 0.426, + 0.645 + ], + "angle": 0, + "content": "3.2. Generative Multi-Modal Models for CIL" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.654, + 0.47, + 0.851 + ], + "angle": 0, + "content": "We adhere to the foundational settings of MiniGPT-4, incorporating a frozen image encoder \\( f_{enc} \\) followed by a trainable projection layer for adaptation to downstream tasks, as shown in Fig. 2. Our primary innovation involves the direct utilization of generative models to produce text, which can then serve as a basis for discriminative classification. However, two major challenges need addressing. First, there is the issue of adapting generative multi-modal models for classification, given that the generated text may differ significantly from class names. Second, we must devise a mechanism for our classification benchmarks to learn in a manner consistent with generative multi-modal models. We introduce these two aspects as follows." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Turning GMM for classification. We employ a distance metric to bridge the gap between generative and discriminative models. During training, we use the real label's text to" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.45, + 0.892, + 0.601 + ], + "angle": 0, + "content": "encourage the model to predict the label of an image with a concise and accurate sentence in the format of \"This is a photo of [CLS].\" avoiding detailed descriptions of all contents in the image. During testing, the model follows the format to output the category text for a given image. We extract the content in \"[CLS]\", then obtain its text features using the CLIP [46] text encoder \\( f_{\\text{text}} \\), and compute the distance with text features of all categories seen by now. The closest class was then considered the final prediction of the generative model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.605, + 0.893, + 0.817 + ], + "angle": 0, + "content": "Converting CIL Benchmarks for adaptation. CIL is usually evaluated on ImageNet, CIFAR-100, and ImageNet-R datasets. These datasets usually consist of images and corresponding one-hot labels \\(\\{\\mathbf{X}_t,\\mathbf{Y}_t\\}\\). Using the CIFAR100 dataset as an example, we pair each image with a sentence to form an image-text pair format \\(\\{\\mathbf{X}_t,\\mathbf{S}_t\\}\\) with the template: \"This is a photo of [CLS]\", where \"[CLS]\" is the label name of that category, such as apple, dog, etc. Next, we partitioned the 100 classes into various tasks based on different settings and fed them into the model sequentially. After completing the training on task \\(T\\), the model should be capable of classifying all the classes encompassed from task 0 to task \\(T\\). Note that only the linear projection layer is updated for further adaptation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.828, + 0.75, + 0.844 + ], + "angle": 0, + "content": "3.3. Optimization and Inference" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Optimization. For each task \\( t \\), we obtain the current task's image-text pair \\( \\{\\mathbf{X}_t, \\mathbf{S}_t\\} \\), where \\( \\mathbf{S}_t \\) contains the corresponding sentence of each image. During training, we" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28709" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.099, + 0.089, + 0.875, + 0.446 + ], + "angle": 0, + "content": "
TypeMethodExemplarTiny-ImageNetImageNet-R
5 tasks10 tasks20 tasks10 tasks
AvgLastAvgLastAvgLastLast
ConventionalEWC [26]X19.016.0015.823.7912.354.7335.00
LwF [29]X22.317.3417.344.7312.484.2638.50
iCaRL [48]✓45.9534.6043.2233.2237.8527.54-
EEIL [7]✓47.1735.1245.0334.6440.4129.72-
UCIR [24]✓50.3039.4248.5837.2942.8430.85-
PASS [89]X49.5441.6447.1939.2742.0132.93-
DyTox [16]✓55.5847.2352.2642.7946.1836.21-
Discriminative PT modelsContinual-CLIP[63]X70.4966.4370.5566.4370.5166.4372.00
L2P [72]X83.5378.3276.3765.7868.0452.4072.92
L2P [72]✓80.2472.8980.0872.6179.4470.4159.78
DualPrompt [71]X85.1581.0181.3873.7373.4560.1668.82
DualPrompt [71]✓79.9272.8379.1573.2180.1771.7457.02
CODA-Prompt [55]X85.9181.3682.8075.2877.4366.3273.88
Linear ProbeX74.3865.4069.7358.3160.1449.7245.17
Linear Probe✓70.1061.1169.3564.1971.6470.5055.72
Generative PT modelsZero-shotX58.1653.7258.1053.7258.1353.7267.38
GMM (Ours)X83.4276.9882.4976.5181.7076.0380.72
GMM (Ours)✓84.1678.4683.9578.6484.2379.1789.41
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.456, + 0.895, + 0.501 + ], + "angle": 0, + "content": "Table 1. Comparison results of our method with other conventional baselines and methods learned with discriminative pre-trained (PT) models on Tiny-ImageNet and ImageNet-R under the conventional CIL setting. \"Avg\" represents the averaged performance after training each task, and \"Last\" represents the performance on all test samples after training the last task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.519, + 0.472, + 0.581 + ], + "angle": 0, + "content": "first utilize a tokenizer to tokenize and acquire the embedding of the questions and answers. We leverage pre-trained encoder \\( f_{enc} \\) and the projection layer to obtain the corresponding features for the input images:" + }, + { + "type": "equation", + "bbox": [ + 0.203, + 0.602, + 0.47, + 0.619 + ], + "angle": 0, + "content": "\\[\n\\mathbf {e} _ {i} = f _ {\\text {e n c}} \\left(\\mathbf {x} _ {i}; \\theta_ {\\text {e n c}}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.632, + 0.47, + 0.695 + ], + "angle": 0, + "content": "Then, the question embedding and the ground-truth embedding of this question, e.g., \"This is a photo of [CLS]\", is concatenated with image embedding. The final input of LLM Decoder \\( f_{dec} \\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.715, + 0.47, + 0.731 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {e}} _ {i} = \\operatorname {C O N C A T E} (b o s, \\mathbf {e} _ {i}, \\mathbf {q}, \\mathbf {s}, e o s). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.743, + 0.47, + 0.789 + ], + "angle": 0, + "content": "bos is the symbol of the sentence beginning, and \\(eos\\) is the symbol for the end of the sentence. This encourages tokens at positions \\(m - 1\\) to predict token \\(m\\):" + }, + { + "type": "equation", + "bbox": [ + 0.082, + 0.804, + 0.469, + 0.853 + ], + "angle": 0, + "content": "\\[\nP \\left(\\hat {s} _ {1}, \\hat {s} _ {2}, \\dots , \\hat {s} _ {m} \\mid \\mathbf {x} _ {i}, \\mathbf {q}, \\mathbf {s}\\right) = \\prod_ {j = 1} ^ {m - 1} P \\left(s _ {j} \\mid \\mathbf {e} _ {i}, \\mathbf {q}, s _ {1}, s _ {2}, \\dots , s _ {j - 1}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.903 + ], + "angle": 0, + "content": "where \\( s_j \\) indicates the ground-truth answer token and \\( \\hat{s}_m \\) is the generated prediction. Then, we can compute the Cross Entropy loss as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.53, + 0.892, + 0.571 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {C E}} = - \\frac {1}{m} \\sum_ {j = 1} ^ {m} s _ {j} \\cdot \\log \\hat {s} _ {j}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.576, + 0.892, + 0.653 + ], + "angle": 0, + "content": "Inference. During inference, we use the updated projection layer in conjunction with the pre-trained encoder to obtain image features. These image features, combined with the question embeddings are then passed to the LLM Decoder to obtain the text output." + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.667, + 0.891, + 0.684 + ], + "angle": 0, + "content": "\\[\n\\operatorname {p r e d} = \\operatorname {a r g m a x} < f _ {\\text {t e x t}} (\\mathbf {s}), f _ {\\text {t e x t}} (\\hat {\\mathbf {s}}) >, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.687, + 0.892, + 0.718 + ], + "angle": 0, + "content": "where \\( f_{\\text{text}} \\) is the text encoder, \\( <, > \\) is the cosine similarity used to calculate the final predictions \\( pred \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.729, + 0.634, + 0.745 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.754, + 0.695, + 0.77 + ], + "angle": 0, + "content": "4.1. Experimental setups" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Datasets and Baselines. We conduct experiments in both conventional CIL and Few-shot CIL scenarios. In conventional CIL, we evaluate on three datasets. CIFAR100, TinyImageNet and ImageNet-R. CIFAR100 contains 60,000 images of \\(32 \\times 32\\) pixels in 100 categories. Each category has 600 images, of which 500 are for the training set, and 100 are for the test set. We experiment with two settings, B0-n and B50-n. The former splits 100 classes into \\(n\\) tasks," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28710" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.245, + 0.09, + 0.713, + 0.105 + ], + "angle": 0, + "content": "Continual CLIP L2P DualPrompt CODA-Prompt GMM" + }, + { + "type": "image", + "bbox": [ + 0.105, + 0.118, + 0.357, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.357, + 0.118, + 0.61, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.118, + 0.862, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.297, + 0.356, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.297, + 0.609, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.297, + 0.862, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.085, + 0.466, + 0.882, + 0.481 + ], + "angle": 0, + "content": "Figure 3. Comparison of our method with other SOTA baselines on CIFAR100 and Tiny-ImageNet under the conventional CIL setting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.469, + 0.537 + ], + "angle": 0, + "content": "while the latter trains on 50 classes first and then distributes the other 50 classes across 5/10 tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.47, + 0.644 + ], + "angle": 0, + "content": "Tiny-ImageNet contains 200 classes out of the original 1000 classes in ImageNet, with 550 images per class, of which 500 are in the training set and 50 are in the test set. The images are down-sampled to \\(64 \\times 64\\) pixels, which makes them easier to process and analyze. We train the first task in half 100 classes and split the other 100 classes into 5/10/20 tasks following [89]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.469, + 0.779 + ], + "angle": 0, + "content": "ImageNet-R [23] contains 200 classes of images, which are included in the original ImageNet 1000 classes. However, many images are newly added and have various styles, such as sketch, painting, misc, etc. The dataset poses a great challenge for continual learning, as it has a wide diversity of image categories and styles and an uneven distribution of samples ranging from 45 to 500 per category. We follow [71] to split the dataset into 10 tasks, each containing 20 classes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.469, + 0.886 + ], + "angle": 0, + "content": "In Few-shot CIL, we use CIFAR100 and miniImageNet [49] following the split proposed by [62]. For both datasets, we partition the data into two parts: base session and incremental sessions. The base session comprises 60 classes with all data available, while the incremental session follows a 5-way 5-shot setting, meaning that each session consists of only 5 classes with 5 samples each." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.887, + 0.469, + 0.9 + ], + "angle": 0, + "content": "In both conventional and few-shot scenarios, we com" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.507, + 0.892, + 0.659 + ], + "angle": 0, + "content": "pare our method with some of the current state-of-the-art approaches, including conventional methods [7, 15, 16, 24, 29, 48, 53, 74, 76, 82, 89], pre-trained and prompt-based methods [55, 71, 72], and some specifically designed methods [12, 53, 62, 79] for few-shot scenarios. Additionally, we compare with a linear probe baseline, wherein features obtained from the image encoder are connected to a classifier for classification. We also consider the Zero-shot approach, where the generated text is directly used for classification without further fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.893, + 0.9 + ], + "angle": 0, + "content": "Implementation Details. We follow BLIP2 [27] to use the EVA-CLIP [59] pre-trained ViT-g/14 and BLIP2 pretrained Qfomer. We also used the MiniGPT-4 pre-trained projection layer checkpoint as our initial parameters. Under the many-shot \"B0\" setting, we employ a learning rate of 3e-7 and use a scheduler with cosine decay. The total training process consists of 2 epochs only. In B50 or B100 settings, we first train the linear layer with a learning rate of 3e-6 on the base classes, and then on the subsequent tasks, we adopt a lower learning rate of 3e-7, both employing a cosine decay scheduler. For the few-shot setting, we use a learning rate of 3e-6 for both base task and incremental tasks. We train one epoch for the base task and two epochs for incremental tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28711" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.12, + 0.088, + 0.856, + 0.381 + ], + "angle": 0, + "content": "
012345678PD↓
iCaRL [48]61.3146.3242.9437.6330.4924.0020.8918.8017.2144.10
EEIL [7]61.3146.5844.0037.2933.1427.1224.1021.5719.5841.73
LUCIR [24]61.3147.8039.3131.9125.6821.3518.6717.2414.1747.14
TOPIC [62]61.3150.0945.1741.1637.4835.5232.1929.4624.4236.89
CEC [79]72.0066.8362.9759.4356.7053.7351.1949.2447.6324.37
F2M [53]72.0567.4763.1659.7056.7153.7751.1149.2147.8424.21
MetaFSCIL [12]72.0467.9463.7760.2957.5855.1652.9050.7949.1922.85
Entropy-reg [31]71.8467.1263.2159.7757.0153.9551.5549.5248.2123.63
L2P* [72]94.1287.2080.9975.6770.9466.7663.1159.8156.8337.29
DualPrompt* [71]93.9786.8580.6775.3170.6166.4462.7759.5856.8037.17
CODA-Prompt* [55]95.3788.8682.6977.8774.4770.1666.4663.7361.1434.23
Zero-shot58.0858.9557.7657.8958.1957.4256.2654.8254.953.13
GMM (Ours)89.3588.4086.1185.0783.6181.3578.9777.3475.1814.17
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.39, + 0.895, + 0.448 + ], + "angle": 0, + "content": "Table 2. Comparison results of our method with other SOTA baselines on mini-ImageNet under the few-shot CIL setting. Phase 0 is the base task with all samples available, and the following 1-8 phases are the incremental 5-way 5-shot tasks. The reported accuracy is the test result on all seen classes after each session of training. \"PD\" represents Performance Prop between session 0 and session 8, with lower values indicating lower forgetting. * indicates our re-implementation based on PILOT [58]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.472, + 0.379, + 0.489 + ], + "angle": 0, + "content": "4.2. Experiments on Conventional CIL" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.502, + 0.473, + 0.806 + ], + "angle": 0, + "content": "In Table 1, we can see that our method outperforms all conventional methods by a large margin, including ResNet-based method DER and ViT-based DyTox. Note that without exemplar, our performance is a bit lower than Dual-Prompt and CODA-Prompt at B100-5 setting. We argue that their performance is mainly due to the backbone pretrained on ImageNet-21K, which largely overlaps with CIFAR100 and Tiny-ImageNet. Another interesting observation is that our method has better performance than all baselines under longer sequence settings (B100-10, B100-20). We believe this is because generative models do not rely on classification heads, making them less prone to bias toward the current task, resulting in less forgetting of past tasks. The Linear probe setting performs less than our method, indicating that our main contribution is not from the Large pre-trained ViT but the generation pipeline. In addition, the Zero-shot performance is superior to many traditional baselines, meaning that the Generative Multi-modal Models are indeed efficient Class-Incremental learners, but its output is less concise without fine-tuning (see Fig. 4)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In Fig. 3, we compared our approach with some pretrained models on CIFAR100 and Tiny-ImageNet in terms of last task accuracy (all baselines are based on PILOT [58] and use 2000 exemplars). It can be observed that our method does not outperform other approaches in the initial tasks (0-2) and short sequence settings (B0-5, B100-5)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.473, + 0.895, + 0.581 + ], + "angle": 0, + "content": "This is because we do not rely on a supervised ImageNet-21K pre-trained backbone. Besides, we trained each task for only 1-2 epochs to ensure efficiency without sacrificing generalization. However, our method exhibits significant advantages in long sequences and later tasks. For instance, under the CIFAR100 B0-20 setting, we outperform CODA-Prompt by 10 points and DualPrompt by 7 points." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.59, + 0.771, + 0.607 + ], + "angle": 0, + "content": "4.3. Experiments on Few-shot CIL" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.895, + 0.81 + ], + "angle": 0, + "content": "In Table 2, we compare our method with several baselines in the few-shot setting on mini-ImageNet. The evaluation metric is the model's accuracy across all the classes it has encountered so far. Our method outperforms conventional methods by a substantial margin in the final task, achieving an increase of more than \\(26\\%\\). Furthermore, we surpass the best discriminative pre-trained approach CODA-Prompt by more than \\(14\\%\\) points. It's important to note that our accuracy in the first task (89.35) might not be as high as CODA-Prompt (95.37). However, in subsequent sessions, we consistently perform better than CODA-Prompt due to our ability to learn new tasks and retain knowledge of old tasks simultaneously." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In Table 3, our method outperforms all other baselines on the CIFAR100 dataset of the few-shot setting, achieving a remarkably lower Performance Drop (PD) of 10.06. Furthermore, the Zero-shot baseline can achieve a very low PD due to the absence of forgetting. However, its overall performance is not very satisfying, as the length and con" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28712" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.09, + 0.306, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.312, + 0.095, + 0.484, + 0.145 + ], + "angle": 0, + "content": "This is a photo of a statue of a bird with a long beak and a long neck. It is made of stone and has a blue sky background." + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.2, + 0.306, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.313, + 0.178, + 0.433, + 0.194 + ], + "angle": 0, + "content": "This is a pelican." + }, + { + "type": "image_footnote", + "bbox": [ + 0.312, + 0.205, + 0.48, + 0.28 + ], + "angle": 0, + "content": "This is a painting of a fire truck. The fire truck is red and has a ladder on the side. The ladder is painted in white. The fire truck has a red and white stripe on the side." + }, + { + "type": "image_footnote", + "bbox": [ + 0.315, + 0.288, + 0.457, + 0.304 + ], + "angle": 0, + "content": "This is a fire engine." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.091, + 0.683, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.689, + 0.095, + 0.856, + 0.158 + ], + "angle": 0, + "content": "This is a sculpture of a bird with its wings spread wide, as if it is flying. The bird is made of bronze and has a green patina. It is located in a park." + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.2, + 0.684, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.69, + 0.178, + 0.808, + 0.194 + ], + "angle": 0, + "content": "This is a pelican." + }, + { + "type": "image_footnote", + "bbox": [ + 0.689, + 0.205, + 0.863, + 0.278 + ], + "angle": 0, + "content": "This is a toy car with a duck in the driver's seat. The car is red and has a fire truck on the back. The duck is wearing a firefighter's helmet and has a fire hose in its hand." + }, + { + "type": "image_footnote", + "bbox": [ + 0.692, + 0.288, + 0.824, + 0.302 + ], + "angle": 0, + "content": "This is a fire truck." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.32, + 0.894, + 0.362 + ], + "angle": 0, + "content": "Figure 4. Visual comparison examples of our method against frozen MiniGPT-4 [88] (Zero-shot). Text with a gray background is generated by MiniGPT-4 based on the image, while text with a orange background represents our method's output. The ground-truth labels are displayed on the left side of each row of images. All images displayed here are random sampled from ImageNet-R." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.374, + 0.462, + 0.655 + ], + "angle": 0, + "content": "
Method048PD↓
iCaRL [48]64.1027.9313.7350.37
EEIL [7]64.1028.9615.8548.25
LUCIR [24]64.1031.6113.5450.56
TOPIC [62]64.1040.1129.3734.73
CEC [79]73.0758.0949.1423.93
F2M [53]71.4557.7649.3522.06
MetaFSCIL [12]74.5059.4849.9724.53
Entropy-reg [31]74.4059.7150.1424.26
L2P* [72]91.2268.6654.8936.33
DualPrompt* [71]91.0868.4554.6736.41
CODA-Prompt* [55]93.5571.9159.3234.23
Zero-shot74.1372.5967.936.20
GMM (Ours)91.5385.6581.4710.06
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.665, + 0.47, + 0.694 + ], + "angle": 0, + "content": "Table 3. Comparison results of our method with other SOTA baselines on CIFAR100 under the few-shot CIL setting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.702, + 0.47, + 0.733 + ], + "angle": 0, + "content": "tent of its output are inconsistent and unpredictable without fine-tuning." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.741, + 0.221, + 0.756 + ], + "angle": 0, + "content": "4.4. Visualizations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.901 + ], + "angle": 0, + "content": "In Fig. 4, we present some comparison examples of our methods against GMM without fine-tuning [88]. We can see that GMM without fine-tuning provides an intuitive description of the overall image content with varying lengths of output text. However, it tends to recognize only broad categories (e.g., biar, car) and struggles with fine-grained categorization (e.g., pelican, truck). The descriptions are sometimes somewhat repetitive (e.g., first fire engine). In contrast, our fine-tuned method accurately identifies the im" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.377, + 0.892, + 0.453 + ], + "angle": 0, + "content": "age's real category, even if there are occasional discrepancies with the true labels (e.g., \"fire engine\" vs. \"fire truck\"). Besides, with the assistance of the text encoder during the testing phase, our model can achieve correct classification results even when predicting similar but not identical text." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.475, + 0.619, + 0.491 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.503, + 0.892, + 0.625 + ], + "angle": 0, + "content": "In this paper, we propose GMM to use generative models for class-incremental learning. By fine-tuning the Generative Multi-modal Model (GMM), we directly generate the label text of the images to be classified. Then we select the label most similar to the generated text by its features. Our experiments demonstrate that this method, which does not require a classification head, is highly effective in addressing classification biases in continual learning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.631, + 0.892, + 0.707 + ], + "angle": 0, + "content": "Limitations. Since we are the first to introduce generative models to class-incremental learning, the overall design of our method is embarrassingly simple. We believe that with more focused efforts in this direction, there will be significant advancements in the field of continual learning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.713, + 0.892, + 0.804 + ], + "angle": 0, + "content": "Broader impact. We believe that introducing GMM into continual learning (CL) is both necessary and urgent. With the rapid development of GMM, we can leverage their capabilities to improve the performance of continual learning. Besides, integrating CL methods into the training process of GMM could significantly reduce training costs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgments. This work is funded by NSFC (NO. 62206135, 62225604), Young Elite Scientists Sponsorship Program by CAST (2023QNRC001), and the Fundamental Research Funds for the Central Universities (Nankai University, 070-63233085). Computation is supported by the Supercomputing Center of Nankai University." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28713" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.157 + ], + "angle": 0, + "content": "[1] Hongjoon Ahn, Jihwan Kwak, Subin Lim, Hyeonsu Bang, Hyojun Kim, and Taesup Moon. Ss-il: Separated softmax for incremental learning. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.472, + 0.241 + ], + "angle": 0, + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.471, + 0.3 + ], + "angle": 0, + "content": "[3] Rahaf Aljundi, Punarjay Chakravarty, and Tinne Tuytelaars. Expert gate: Lifelong learning with a network of experts. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3366-3375, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.302, + 0.471, + 0.372 + ], + "angle": 0, + "content": "[4] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2(3):4, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.373, + 0.471, + 0.442 + ], + "angle": 0, + "content": "[5] Hedi Ben-Younes, Rémi Cadene, Matthieu Cord, and Nicolas Thome. Mutan: Multimodal tucker fusion for visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 2612-2620, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.445, + 0.47, + 0.514 + ], + "angle": 0, + "content": "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.516, + 0.47, + 0.584 + ], + "angle": 0, + "content": "[7] Francisco M Castro, Manuel J Marín-Jiménez, Nicolas Guil, Cordelia Schmid, and Karteek Alahari. End-to-end incremental learning. In Proceedings of the European conference on computer vision (ECCV), pages 233-248, 2018. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.587, + 0.471, + 0.657 + ], + "angle": 0, + "content": "[8] Jun Chen, Han Guo, Kai Yi, Boyang Li, and Mohamed Elhoseiny. Visualgpt: Data-efficient adaptation of pretrained language models for image captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18030-18040, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.659, + 0.469, + 0.7 + ], + "angle": 0, + "content": "[9] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.702, + 0.469, + 0.744 + ], + "angle": 0, + "content": "[10] Xinlei Chen*, Saining Xie*, and Kaiming He. An empirical study of training self-supervised vision transformers. arXiv preprint arXiv:2104.02057, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.469, + 0.83 + ], + "angle": 0, + "content": "[11] Ali Cheraghian, Shafin Rahman, Pengfei Fang, Soumava Kumar Roy, Lars Petersson, and Mehrtash Harandi. Semantic-aware knowledge distillation for few-shot class-incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2534-2543, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[12] Zhixiang Chi, Li Gu, Huan Liu, Yang Wang, Yuanhao Yu, and Jin Tang. Meta-scil: A meta-learning approach for few-shot class incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14166-14175, 2022. 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.174 + ], + "angle": 0, + "content": "[13] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.176, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[14] Matthias Delange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Ales Leonardis, Greg Slabaugh, and Tinne Tuytelaars. A continual learning survey: Defying forgetting in classification tasks. TPAMI, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.892, + 0.285 + ], + "angle": 0, + "content": "[15] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In ECCV, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.288, + 0.892, + 0.329 + ], + "angle": 0, + "content": "[16] Arthur Douillard, Alexandre Ramé, Guillaume Couairon, and Matthieu Cord. Dytox: Transformers for continual learning with dynamic token expansion. In CVPR, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.397 + ], + "angle": 0, + "content": "[17] Chrisantha Fernando, Dylan Banarse, Charles Blundell, Yori Zwols, David Ha, Andrei A Rusu, Alexander Pritzel, and Daan Wierstra. Pathnet: Evolution channels gradient descent in super neural networks. arXiv preprint arXiv:1701.08734, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.892, + 0.427 + ], + "angle": 0, + "content": "[18] Rui Gao and Weiwei Liu. Ddgr: Continual learning with deep diffusion-based generative replay. In ICML, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.428, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[19] Meng-Hao Guo, Cheng-Ze Lu, Zheng-Ning Liu, Ming-Ming Cheng, and Shi-Min Hu. Visual attention network. Computational Visual Media, 9(4):733-752, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.47, + 0.892, + 0.511 + ], + "angle": 0, + "content": "[20] Tyler L Hayes, Kushal Kafle, Robik Shrestha, Manoj Acharya, and Christopher Kanan. Remind your neural network to prevent catastrophic forgetting. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.512, + 0.892, + 0.567 + ], + "angle": 0, + "content": "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.568, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[22] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[23] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, Dawn Song, Jacob Steinhardt, and Justin Gilmer. The many faces of robustness: A critical analysis of out-of-distribution generalization. ICCV, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.892, + 0.748 + ], + "angle": 0, + "content": "[24] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In CVPR, 2019. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.749, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[25] Xinting Hu, Kaihua Tang, Chunyan Miao, Xian-Sheng Hua, and Hanwang Zhang. Distilling causal effect of data in class incremental learning. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[26] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.903 + ], + "angle": 0, + "content": "[27] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28714" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.472, + 0.191 + ], + "angle": 0, + "content": "[28] Wenbo Li, Pengchuan Zhang, Lei Zhang, Qiuyuan Huang, Xiaodong He, Siwei Lyu, and Jianfeng Gao. Object-driven text-to-image synthesis via adversarial training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12174-12182, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.47, + 0.233 + ], + "angle": 0, + "content": "[29] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.469, + 0.261 + ], + "angle": 0, + "content": "[30] Zhizhong Li and Derek Hoiem. Learning without forgetting. TPAMI, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.469, + 0.331 + ], + "angle": 0, + "content": "[31] Huan Liu, Li Gu, Zhixiang Chi, Yang Wang, Yuanhao Yu, Jun Chen, and Jin Tang. Few-shot class-incremental learning via entropy-regularized data-free replay. In European Conference on Computer Vision, pages 146–162. Springer, 2022. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.469, + 0.362 + ], + "angle": 0, + "content": "[32] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.364, + 0.469, + 0.418 + ], + "angle": 0, + "content": "[33] Ming Liu, Yuxiang Wei, Xiaohe Wu, Wangmeng Zuo, and Lei Zhang. Survey on leveraging pre-trained generative adversarial networks for image editing and restoration. Science China Information Sciences, 66(5):151101, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.421, + 0.469, + 0.462 + ], + "angle": 0, + "content": "[34] Wei Liu, Sihan Chen, Longteng Guo, Xinxin Zhu, and Jing Liu. Cptr: Full transformer network for image captioning. arXiv preprint arXiv:2101.10804, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.463, + 0.469, + 0.517 + ], + "angle": 0, + "content": "[35] Xialei Liu, Marc Masana, Luis Herranz, Joost Van de Weijer, Antonio M Lopez, and Andrew D Bagdanov. Rotate your networks: Better weight consolidation and less catastrophic forgetting. In ICPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.519, + 0.469, + 0.575 + ], + "angle": 0, + "content": "[36] Xialei Liu, Yu-Song Hu, Xu-Sheng Cao, Andrew D Bagdanov, Ke Li, and Ming-Ming Cheng. Long-tailed class incremental learning. In European Conference on Computer Vision, pages 495-512. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.577, + 0.469, + 0.63 + ], + "angle": 0, + "content": "[37] Xialei Liu, Xusheng Cao, Haori Lu, Jia-wen Xiao, Andrew D Bagdanov, and Ming-Ming Cheng. Class incremental learning with pre-trained vision-language models. arXiv preprint arXiv:2310.20348, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.469, + 0.673 + ], + "angle": 0, + "content": "[38] Yaoyao Liu, Bernt Schiele, and Qianru Sun. Adaptive aggregation networks for class-incremental learning. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.676, + 0.469, + 0.731 + ], + "angle": 0, + "content": "[39] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.469, + 0.773 + ], + "angle": 0, + "content": "[40] Arun Mallya, Dillon Davis, and Svetlana Lazebnik. Piggyback: Adapting a single network to multiple tasks by learning to mask weights. In ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.469, + 0.856 + ], + "angle": 0, + "content": "[41] Marc Masana, Xialei Liu, Bartlomiej Twardowski, Mikel Menta, Andrew D Bagdanov, and Joost Van De Weijer. Class-incremental learning: survey and performance evaluation on image classification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5):5513-5533, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[42] Pratik Mazumder, Pravendra Singh, and Piyush Rai. Few-shot lifelong learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2337-2345, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[43] Michael McCloskey and Neal J Cohen. Catastrophic interference in connectionist networks: The sequential learning problem. In *Psychology of learning and motivation*, pages 109–165. Elsevier, 1989. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.791, + 0.162 + ], + "angle": 0, + "content": "[44] OpenAI. Gpt-4 technical report, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[45] Haoxuan Qu, Hossein Rahmani, Li Xu, Bryan Williams, and Jun Liu. Recent advances of continual learning in computer vision: An overview. arXiv preprint arXiv:2109.11369, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.289, + 0.892, + 0.329 + ], + "angle": 0, + "content": "[47] Jathushan Rajasegaran, Munawar Hayat, Salman Khan, Fahad Shahbaz Khan, and Ling Shao. Random path selection for incremental learning. In NIPS, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.331, + 0.892, + 0.383 + ], + "angle": 0, + "content": "[48] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In CVPR, 2017. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.386, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[49] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.457, + 0.892, + 0.511 + ], + "angle": 0, + "content": "[50] Jonathan Schwarz, Wojciech Czarnecki, Jelena Luketina, Agnieszka Grabska-Barwinska, Yee Whye Teh, Razvan Pascanu, and Raia Hadsell. Progress & compress: A scalable framework for continual learning. In ICML, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.512, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[51] Joan Serra, Didac Suris, Marius Miron, and Alexandros Karatzoglou. Overcoming catastrophic forgetting with hard attention to the task. In ICML, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.554, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[52] Yijia Shao, Yiduo Guo, Dongyan Zhao, and Bing Liu. Class incremental learning based on label generation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 1263-1276, Toronto, Canada, 2023. Association for Computational Linguistics. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.637, + 0.892, + 0.704 + ], + "angle": 0, + "content": "[53] Guangyuan Shi, Jiaxin Chen, Wenlong Zhang, Li-Ming Zhan, and Xiao-Ming Wu. Overcoming catastrophic forgetting in incremental few-shot learning by finding flat minima. Advances in neural information processing systems, 34: 6747-6761, 2021. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.707, + 0.892, + 0.748 + ], + "angle": 0, + "content": "[54] Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual learning with deep generative replay. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.749, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[55] James Seale Smith, Leonid Karlinsky, Vyshnavi Gutta, Paola Cascante-Bonilla, Donghyun Kim, Assaf Arbelle, Rameswar Panda, Rogerio Feris, and Zsolt Kira. Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11909-11919, 2023. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[56] Boyuan Sun, Yuqi Yang, Le Zhang, Ming-Ming Cheng, and Qibin Hou. Corrmatch: Label propagation via correlation matching for semi-supervised semantic segmentation. IEEE Computer Vision and Pattern Recognition (CVPR), 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28715" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.161 + ], + "angle": 0, + "content": "[57] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7464-7473, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.163, + 0.47, + 0.204 + ], + "angle": 0, + "content": "[58] Hai-Long Sun, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Pilot: A pre-trained model-based continual learning toolbox. arXiv preprint arXiv:2309.07117, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.207, + 0.469, + 0.247 + ], + "angle": 0, + "content": "[59] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.25, + 0.469, + 0.291 + ], + "angle": 0, + "content": "[60] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.293, + 0.469, + 0.332 + ], + "angle": 0, + "content": "[61] Xiaoyu Tao, Xinyuan Chang, Xiaopeng Hong, Xing Wei, and Yihong Gong. Topology-preserving class-incremental learning. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.335, + 0.469, + 0.403 + ], + "angle": 0, + "content": "[62] Xiaoyu Tao, Xiaopeng Hong, Xinyuan Chang, Songlin Dong, Xing Wei, and Yihong Gong. Few-shot class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12183-12192, 2020. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.406, + 0.469, + 0.447 + ], + "angle": 0, + "content": "[63] Vishal Thengane, Salman Khan, Munawar Hayat, and Fahad Khan. Clip model is an efficient continual learner. arXiv preprint arXiv:2210.03114, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.449, + 0.469, + 0.517 + ], + "angle": 0, + "content": "[64] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.519, + 0.469, + 0.574 + ], + "angle": 0, + "content": "[65] Maria Tsimpoukelli, Jacob L Menick, Serkan Cabi, SM Eslami, Oriol Vinyls, and Felix Hill. Multimodal few-shot learning with frozen language models. Advances in Neural Information Processing Systems, 34:200-212, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.576, + 0.469, + 0.604 + ], + "angle": 0, + "content": "[66] Gido M Van de Ven and Andreas S Tolias. Three scenarios for continual learning. NIPS Workshops, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.606, + 0.469, + 0.646 + ], + "angle": 0, + "content": "[67] Liuyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. A comprehensive survey of continual learning: Theory, method and application. arXiv preprint arXiv:2302.00487, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.648, + 0.469, + 0.716 + ], + "angle": 0, + "content": "[68] Liyuan Wang, Jingyi Xie, Xingxing Zhang, Mingyi Huang, Hang Su, and Jun Zhu. Hierarchical decomposition of prompt-based continual learning: Rethinking obscured suboptimality. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.719, + 0.469, + 0.786 + ], + "angle": 0, + "content": "[69] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, et al. Vision: Large language model is also an open-ended decoder for vision-centric tasks. arXiv preprint arXiv:2305.11175, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.789, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[70] Xiao Wang, Guangyao Chen, Guangwu Qian, Pengcheng Gao, Xiao-Yong Wei, Yaowei Wang, Yonghong Tian, and Wen Gao. Large-scale multi-modal pre-trained models: A comprehensive survey. Machine Intelligence Research, 20 (4):447-482, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[71] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "prompting for rehearsal-free continual learning. In ECCV, 2022. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[72] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer G. Dy, and Tomas Pfister. Learning to prompt for continual learning. CVPR, 2022. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.892, + 0.26 + ], + "angle": 0, + "content": "[73] Tz-Ying Wu, Gurumurthy Swaminathan, Zhizhong Li, Avinash Ravichandran, Nuno Vasconcelos, Rahul Bhotika, and Stefano Soatto. Class-incremental learning with strong pre-trained models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9601-9610, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.263, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[74] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In CVPR, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.305, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[75] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.892, + 0.39 + ], + "angle": 0, + "content": "[76] Shipeng Yan, Jiangwei Xie, and Xuming He. Der: Dynamically expandable representation for class incremental learning. In CVPR, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.392, + 0.892, + 0.447 + ], + "angle": 0, + "content": "[77] Yang Yang, Zhiying Cui, Junjie Xu, Changhong Zhong, Wei-Shi Zheng, and Ruixuan Wang. Continual learning with bayesian model based on a fixed pre-trained feature extractor. Visual Intelligence, 1(1):5, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.449, + 0.892, + 0.529 + ], + "angle": 0, + "content": "[78] Lu Yu, Bartlomiej Twardowski, Xialei Liu, Luis Herranz, Kai Wang, Yongmei Cheng, Shangling Jui, and Joost van de Weijer. Semantic drift compensation for class-incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6982-6991, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.533, + 0.892, + 0.602 + ], + "angle": 0, + "content": "[79] Chi Zhang, Nan Song, Guosheng Lin, Yun Zheng, Pan Pan, and Yinghui Xu. Few-shot incremental learning with continually evolved classifiers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12455-12464, 2021. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.604, + 0.892, + 0.672 + ], + "angle": 0, + "content": "[80] Chang-Bin Zhang, Jia-Wen Xiao, Xialei Liu, Ying-Cong Chen, and Ming-Ming Cheng. Representation compensation networks for continual semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7053-7064, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.674, + 0.892, + 0.729 + ], + "angle": 0, + "content": "[81] Gengwei Zhang, Liyuan Wang, Guoliang Kang, Ling Chen, and Yunchao Wei. Slca: Slow learner with classifier alignment for continual learning on a pre-trained model. arXiv preprint arXiv:2303.05118, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.731, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[82] Bowen Zhao, Xi Xiao, Guojun Gan, Bin Zhang, and Shu-Tao Xia. Maintaining discrimination and fairness in class incremental learning. In CVPR, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.774, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[83] Hanbin Zhao, Yongjian Fu, Mintong Kang, Qi Tian, Fei Wu, and Xi Li. Mgmtf: Multi-grained slow vs. fast framework for few-shot class-incremental learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.831, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[84] Da-Wei Zhou, Qi-Wei Wang, Zhi-Hong Qi, Han-Jia Ye, De-Chuan Zhan, and Ziwei Liu. Deep class-incremental learning: A survey. arXiv preprint arXiv:2302.03648, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[85] Da-Wei Zhou, Han-Jia Ye, De-Chuan Zhan, and Ziwei Liu. Revisiting class-incremental learning with pre-trained mod" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28716" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.468, + 0.119 + ], + "angle": 0, + "content": "els: Generalizability and adaptivity are all you need, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.176 + ], + "angle": 0, + "content": "[86] Da-Wei Zhou, Yuanhan Zhang, Jingyi Ning, Han-Jia Ye, De-Chuan Zhan, and Ziwei Liu. Learning without forgetting for vision-language models. arXiv preprint arXiv:2305.19270, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.233 + ], + "angle": 0, + "content": "[87] Qiang Zhou, Chaohui Yu, Shaofeng Zhang, Sitong Wu, Zhibing Wang, and Fan Wang. Regionblip: A unified multimodal pre-training framework for holistic and regional comprehension. arXiv preprint arXiv:2308.02299, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.469, + 0.29 + ], + "angle": 0, + "content": "[88] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.292, + 0.469, + 0.36 + ], + "angle": 0, + "content": "[89] Fei Zhu, Xu-Yao Zhang, Chuang Wang, Fei Yin, and ChengLin Liu. Prototype augmentation and self-supervision for incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5871-5880, 2021. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.469, + 0.43 + ], + "angle": 0, + "content": "[90] Kai Zhu, Yang Cao, Wei Zhai, Jie Cheng, and Zheng-Jun Zha. Self-promoted prototype refinement for few-shot class incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6801-6810, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28717" + } + ] +] \ No newline at end of file diff --git a/2024/Generative Multi-modal Models are Good Class Incremental Learners/a19d9ccd-1dff-4a86-83d4-a672741b4e0a_origin.pdf b/2024/Generative Multi-modal Models are Good Class Incremental Learners/a19d9ccd-1dff-4a86-83d4-a672741b4e0a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4d9c28a28e155c8cd6a5852c27759f50a85e11af --- /dev/null +++ b/2024/Generative Multi-modal Models are Good Class Incremental Learners/a19d9ccd-1dff-4a86-83d4-a672741b4e0a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:97ba7b6daa62a55f415f33de4271f2ebc90828838ee4397baa4beada6fe691b8 +size 3829207 diff --git a/2024/Generative Multi-modal Models are Good Class Incremental Learners/full.md b/2024/Generative Multi-modal Models are Good Class Incremental Learners/full.md new file mode 100644 index 0000000000000000000000000000000000000000..686d247a9c151769caaa82c01625cedf07d2d637 --- /dev/null +++ b/2024/Generative Multi-modal Models are Good Class Incremental Learners/full.md @@ -0,0 +1,327 @@ +# Generative Multi-modal Models are Good Class-Incremental Learners + +Xusheng Cao $^{1}$ , Haori Lu $^{1}$ , Linlan Huang $^{1}$ , Xialei Liu $^{2,1*}$ , Ming-Ming Cheng $^{2,1}$ $^{1}$ VCIP, CS, Nankai University $^{2}$ NKIARI, Shenzhen Futian + +{caoxusheng, luhaori, huanglinlan}@mail.nankai.edu.cn, {xialei, cmm}@nankai.edu.cn + +# Abstract + +In class-incremental learning (CIL) scenarios, the phenomenon of catastrophic forgetting caused by the classifier's bias towards the current task has long posed a significant challenge. It is mainly caused by the characteristic of discriminative models. With the growing popularity of the generative multi-modal models, we would explore replacing discriminative models with generative ones for CIL. However, transitioning from discriminative to generative models requires addressing two key challenges. The primary challenge lies in transferring the generated textual information into the classification of distinct categories. Additionally, it requires formulating the task of CIL within a generative framework. To this end, we propose a novel generative multi-modal model (GMM) framework for class-incremental learning. Our approach directly generates labels for images using an adapted generative model. After obtaining the detailed text, we use a text encoder to extract text features and employ feature matching to determine the most similar label as the classification prediction. In the conventional CIL settings, we achieve significantly better results in long-sequence task scenarios. Under the Few-shot CIL setting, we have improved by at least $14\%$ accuracy over all the current state-of-the-art methods with significantly less forgetting. Our code is available at https://github.com/DoubleClass/GMM. + +# 1. Introduction + +Deep neural networks [19, 33, 56] have made remarkable strides in numerous applications, primarily owing to the vast amounts of data and computational resources at their disposal. Nonetheless, these accomplishments are predominantly contingent on having access to all the required data simultaneously for training on various tasks. In cases where data is acquired incrementally, these networks often encounter the challenge of catastrophic forgetting [43]. Hence, the capacity to seamlessly incorporate new knowl + +![](images/2073830f54ab6f9c767edec7e8011939c055c676c9226f2286e014db165462d5.jpg) + +![](images/faa7a05aa9b57d9a251de60b9c1b9de4a9bef9f4ffae2aec368dcb51f6459e26.jpg) +Figure 1. Illustration of conventional discriminative models for class-incremental learning (CIL) and our generative multi-modal models (GMM) for CIL. Discriminative models pose potential risk of classifier bias toward current task with network expansion. Our GMM framework consists generation and classification phases. And It is adapted to CIL based on the similarity of generated text and the true category names. + +edge while retaining previously acquired knowledge is a highly desirable attribute for future artificial intelligence systems. Continual learning [45, 67, 77, 84] is a subject of study aimed at advancing the evolution of neural networks toward this goal. + +Numerous studies have delved into continual learning, categorizing their approaches into three main groups [14]: rehearsal-based, architecture-based, and regularization-based methods. Additionally, hybrid methods are gaining popularity as they combine insights from different perspectives. Within this research landscape, three primary scenarios [66] have received extensive attention, with class-incremental learning (CIL) [41] being one of the most demanding settings. In our work, we concentrate on CIL, wherein each task comprises a distinct set of classes, and the primary challenge is to enable the network to recognize new classes without forgetting knowledge of previously encountered ones. + +The majority of existing research on class-incremental + +learning (CIL) focuses on training models from scratch, relying solely on data from the current tasks [7, 15, 16, 24, 26, 29, 48, 74, 76, 89]. In contrast, humans accumulate knowledge over extended periods, drawing upon a wealth of prior world knowledge. Consequently, there has been a growing interest in pre-trained models for CIL [55, 71, 72, 81, 86], harnessing the knowledge acquired from extensive, pre-existing datasets to address the immediate tasks at hand. For instance, prompt-based methods [55, 71, 72] utilize prompt-tuning to summarize task-specific knowledge from prior pretrained knowledge. SLCA [81] and ADAM [85], on the other hand, solely fine-tuning the pre-trained model to adapt the pre-existing knowledge into the objectives of immediate tasks. + +To tackle image classification downstream tasks, pretrained models are traditionally derived from discriminative tasks, like supervised learning on datasets such as ImageNet-21K [49], or they may stem from self-supervised learning efforts [6, 9, 10, 22]. However, in our research, we venture into the paradigm of generative multi-modal models to address image classification tasks. Generative models like GPT4 [44] and LLaVa [32] have garnered significant attention in recent years due to their capacity to produce highly informative descriptions of input images. On the one hand, it can harness the wealth of semantic correspondences between texts and images, while on the other hand, there's no requirement to expand the classifier with each new task, unlike in the case of discriminative models for CIL. + +Nonetheless, harnessing the knowledge from pre-trained generative models for downstream class-incremental learning (CIL) tasks presents a nontrivial endeavor. The primary challenge lies in transferring the generated textual information into the classification of distinct categories. Additionally, there's the task of formulating CIL within a generative framework, which poses a second significant challenge. Shao et al. [52] presents the VAG system, which formulates CIL as a continual label generation problem, preserving the language model's ability to learn new classes. However, it only works in the field of Natural Language Processing (NLP), which is inherently suited to Large Language Models (LLM). To the best of our knowledge, we are the first to apply this generative approach to incremental learning in the field of image classification. + +In this work, we propose Generative Multi-modal Models (GMM) for class-incremental learning. As illustrated in Fig. 1 (a), conventional discriminative methods extract image features with a network backbone, then forward them to a classifier to obtain the probability of the image belonging to each label, with the label having the highest probability being the output of the discriminative models. While in Fig. 1 (b), we adopt a generative approach, which directly produces a descriptive sentence for the given image, which is then compared with the actual label texts with a + +text encoder. The most similar label becomes the predicted result of our generative model. This approach allows us to leverage the rich pre-training knowledge in generative multi-modal models while avoiding the use of the expanded classification head, which mitigates the risk of the model bias towards the current task and reduces catastrophic forgetting. + +The main contributions of this paper are: + +- We propose a novel generative approach (GMM) to address class-incremental learning by leveraging multimodal models. +- We reformulate GMM for image classification and adapt it for the downstream benchmarks. Without an expanded classification head like in discriminative models, our model significantly mitigates the issue of bias towards current tasks, resulting in significantly reduced forgetting in CIL. +- Our model achieves state-of-the-art performance across multiple datasets in both conventional and few-shot CIL settings. + +# 2. Related Work + +# 2.1. Class-Incremental Learning + +In Class-Incremental Learning, tasks arrive sequentially, and each class is exclusive to a specific task without any overlap. The goal is to acquire knowledge from new classes while preserving information from previously encountered classes. There are three primary branches in CIL [14], including rehearsal-based, architecture-based, and regularization-based methods. Rehearsal-based methods [1, 7, 48, 75] store a small set of data derived from old classes to represent knowledge from previous tasks. These exemplar data can be either original data [48], generative data [18, 54] or hidden features [20]. Architecture-based methods focus on modifying network architecture to alleviate forgetting. Approaches include learning redundant network architecture [17, 47], learning different expert networks [3, 50] or parameters [38, 40, 51] for each task, dynamically expanding network parameters to accumulate incremental knowledge [76]. Regularization-based methods introduce an additional regularization term to restrict network updates when adapting to new tasks. In such cases, EWC [26], SDC [78] and Rotated-EWC [35] expect that parameters essential for the old tasks should not be updated excessively. Moreover, from the perspective of network output consistency, numerous studies [25, 30, 36, 61, 80] incorporate distillation to prevent forgetting. + +Few-shot CIL Few-shot Class-Incremental Learning (FSCIL) [42, 62] explores few-shot learning in an incremental context, with all data samples available for base session and very limited data in each incremental session. Some FSCIL methods [11, 62, 83] train the model in both base and incre + +mental sessions, aiming to mitigate overfitting challenges caused by the limited data in incremental learning. Other strategies [53, 79, 90] primarily train the model in the base session and make minimal adjustments in the incremental sessions, thereby reducing forgetting but may come at the cost of decreased precision in the incremental sessions. + +# 2.2. Pre-trained models for CIL + +There are many methods [63, 72, 73] having shown that pretrained models are effective for continual learning. One main branch trains a set of prompts to retain previous knowledge [55, 68, 71, 72]. A selected subset of prompts are fed into the model during forward to prompt model the past knowledge. Additionally, methods like SLCA [81] and ADAM [85] fine-tune pre-trained models, achieving impressive results with less forgetting. Continual-CLIP [63] demonstrates that the CLIP [46] model is capable of performing continual learning without any extra training. This highlights the significant potential of multi-modal pretrained models in the realm of continual learning. Inspired by this, many methods [37, 86] employ CLIP as the backbone to utilize the multi-modal information. However, if not using a classifier directly, these approaches need to utilize expanded text features to calculate distances with image features for classification. This will exacerbate the model's bias towards current data, consequently leading to the forgetting of previously acquired knowledge. To avoid this bias, we use a generative model to directly generate prediction text. The fixed text decoder will function as the classifier, significantly alleviating the bias. + +# 2.3. Vision Language Models + +In recent years, vision-language multi-modal models have made significant progress and achieved impressive results in various downstream tasks [5, 28, 34, 70]. Traditional vision-language models employ different types of encoders to extract information from vision and language models, including single-stream [57], dual-stream [39] and fusion [60] encoders. A key aspect of vision-language models is the alignment of multi-modal features. CLIP [46], for example, extracts image and text features separately using respective encoders and enforces alignment through a contrastive loss, ensuring alignment between positive image-text pairs in the feature space. VisualGPT [8] and Frozen [65] leverage pre-trained models as encoders for visual-language tasks. From then on, the utilization of pre-trained models in vision-language tasks became more and more popular. For instance, Flamingo [2] and BLIP-2 [27] align the pre-trained image and text encoders employing gated cross-attention and Q-Former, respectively. Furthermore, LLaVA [32] and MiniGPT-4 [88] leverage more robust Large Language Models (LLM) [13, 64] as text encoders, while only training a projection layer for alignment. With + +the increasing popularity of LLM, an increasing number of studies [4, 69, 87] explore the potential of multi-modal LLMs for vision-language tasks. + +# 3. Method + +In this section, we introduce the preliminaries of class-incremental learning and generative multi-modal models. Then, we present our approach to leverage generative models for CIL and the corresponding learning process. + +# 3.1. Preliminaries + +Class-Incremental Learning. Given N tasks $T = \{T_{1}, T_{2}, \dots, T_{N}\}$ , the goal of class-incremental learning is to learn each task $T_{t}$ with its associated data $\{\mathbf{X}_{t}, \mathbf{Y}_{t}\}$ in a sequential order. For each task, it contains samples $\{\mathbf{x}_i, \mathbf{y}_i\}$ , $i = 1, \dots, n_t$ , where $\mathbf{x}_i$ is the images and $\mathbf{y}_i$ is the corresponding one-hot labels. Typically, $\mathbf{X}_i \cap \mathbf{X}_j = \emptyset, \forall i \neq j$ . At inference, the model is tested on all seen tasks without task IDs. In some scenarios, fixed memory storage is set to keep a few samples of previous tasks to prevent forgetting. + +Normally, a CIL model consists of a feature extractor and a classifier head $F = \{f_{\theta}, \mathcal{H}_{\phi}\}$ which are parameterized by $\{\theta, \phi\}$ . In traditional class-incremental learning, $\theta$ is usually a modified ResNet [21] with all parameters tunable. In pre-trained or prompt-based methods, $\theta$ represents fewer trainable parameters like a linear adaptor or a couple of prompts. $\phi$ is a linear classifier head projecting image features to probability predictions, which has to be expanded for each new task in order to make predictions for the new classes. The conventional Cross Entropy loss is often used for updating $\theta$ and $\phi$ , which for task $t$ is: + +$$ +\mathcal {L} _ {\mathrm {C E}} \left(\mathbf {X} _ {t}, \mathbf {Y} _ {t}; \theta , \phi\right) = - \frac {1}{n _ {t}} \sum_ {i = 1} ^ {n _ {t}} \mathbf {y} _ {i} \cdot \log \mathcal {H} (f (\mathbf {x} _ {i}; \theta); \phi). \tag {1} +$$ + +In the continual learning process, the parameter $\phi$ can easily deviate to the data of the current task due to the absence or scarcity of old samples in the previous tasks, resulting in forgetting the previously acquired knowledge and deteriorating the overall performance. + +Generative Multi-Modal Models (GMM). Multi-modal models have demonstrated exceptional performance in generating detailed image descriptions by incorporating both visual and textual information. Notably, GPT-4 [44] stands out as an advanced model proficient in generating comprehensive image descriptions and providing explanations for the depicted content. Furthermore, MiniGPT-4 [88] proposes a two-stage fine-tuning process that aligns image features and large language models, enabling LLaMa [64] to recognize images and conduct further dialogue based on the image content. + +![](images/0c76f35c4278ab83e6618a5a6e79752f02729fc226ffe59b2694249f5ebdc057.jpg) +Figure 2. The overview structure of our proposed method. The conceptual illustration of the generative multi-modal model (GMM) is shown on the left. In order to adapt this model for CIL, We have to turn the GMM model for classification and further adapt to our objective benchmark for learning (see Sec. 3.2). On the right side, we demonstrate how the final evaluation is taken for CIL with all seen classes. Text encoder is used for obtaining the embeddings for similarity prediction. + +![](images/114042b0e1f8d66755f9eb4f73250f247fb96697ae64dece69a0f5a6f6208f5d.jpg) + +As shown in Fig. 2, these models consist of an encoder $f_{enc}$ to generate content embeddings including image embedding and text embedding, which is further used as the input of an auto-aggressive decoder $f_{dec}$ to generate image descriptions. Input image $\mathbf{x}_i$ is encoded as an image embedding $e_i$ , and question embeddings $\mathbf{q}$ with $q_1, \dots, q_l$ can be concatenated with the image embedding together to generate answer embeddings $\mathbf{s}$ with $s_1, \dots, s_m$ . The output tokens are generated one by one with the condition of previously generated tokens. For instance, $s_m$ is generated with all previous $m - 1$ tokens (see the decoder in Fig. 2). + +# 3.2. Generative Multi-Modal Models for CIL + +We adhere to the foundational settings of MiniGPT-4, incorporating a frozen image encoder $f_{enc}$ followed by a trainable projection layer for adaptation to downstream tasks, as shown in Fig. 2. Our primary innovation involves the direct utilization of generative models to produce text, which can then serve as a basis for discriminative classification. However, two major challenges need addressing. First, there is the issue of adapting generative multi-modal models for classification, given that the generated text may differ significantly from class names. Second, we must devise a mechanism for our classification benchmarks to learn in a manner consistent with generative multi-modal models. We introduce these two aspects as follows. + +Turning GMM for classification. We employ a distance metric to bridge the gap between generative and discriminative models. During training, we use the real label's text to + +encourage the model to predict the label of an image with a concise and accurate sentence in the format of "This is a photo of [CLS]." avoiding detailed descriptions of all contents in the image. During testing, the model follows the format to output the category text for a given image. We extract the content in "[CLS]", then obtain its text features using the CLIP [46] text encoder $f_{\text{text}}$ , and compute the distance with text features of all categories seen by now. The closest class was then considered the final prediction of the generative model. + +Converting CIL Benchmarks for adaptation. CIL is usually evaluated on ImageNet, CIFAR-100, and ImageNet-R datasets. These datasets usually consist of images and corresponding one-hot labels $\{\mathbf{X}_t,\mathbf{Y}_t\}$ . Using the CIFAR100 dataset as an example, we pair each image with a sentence to form an image-text pair format $\{\mathbf{X}_t,\mathbf{S}_t\}$ with the template: "This is a photo of [CLS]", where "[CLS]" is the label name of that category, such as apple, dog, etc. Next, we partitioned the 100 classes into various tasks based on different settings and fed them into the model sequentially. After completing the training on task $T$ , the model should be capable of classifying all the classes encompassed from task 0 to task $T$ . Note that only the linear projection layer is updated for further adaptation. + +# 3.3. Optimization and Inference + +Optimization. For each task $t$ , we obtain the current task's image-text pair $\{\mathbf{X}_t, \mathbf{S}_t\}$ , where $\mathbf{S}_t$ contains the corresponding sentence of each image. During training, we + +
TypeMethodExemplarTiny-ImageNetImageNet-R
5 tasks10 tasks20 tasks10 tasks
AvgLastAvgLastAvgLastLast
ConventionalEWC [26]X19.016.0015.823.7912.354.7335.00
LwF [29]X22.317.3417.344.7312.484.2638.50
iCaRL [48]✓45.9534.6043.2233.2237.8527.54-
EEIL [7]✓47.1735.1245.0334.6440.4129.72-
UCIR [24]✓50.3039.4248.5837.2942.8430.85-
PASS [89]X49.5441.6447.1939.2742.0132.93-
DyTox [16]✓55.5847.2352.2642.7946.1836.21-
Discriminative PT modelsContinual-CLIP[63]X70.4966.4370.5566.4370.5166.4372.00
L2P [72]X83.5378.3276.3765.7868.0452.4072.92
L2P [72]✓80.2472.8980.0872.6179.4470.4159.78
DualPrompt [71]X85.1581.0181.3873.7373.4560.1668.82
DualPrompt [71]✓79.9272.8379.1573.2180.1771.7457.02
CODA-Prompt [55]X85.9181.3682.8075.2877.4366.3273.88
Linear ProbeX74.3865.4069.7358.3160.1449.7245.17
Linear Probe✓70.1061.1169.3564.1971.6470.5055.72
Generative PT modelsZero-shotX58.1653.7258.1053.7258.1353.7267.38
GMM (Ours)X83.4276.9882.4976.5181.7076.0380.72
GMM (Ours)✓84.1678.4683.9578.6484.2379.1789.41
+ +Table 1. Comparison results of our method with other conventional baselines and methods learned with discriminative pre-trained (PT) models on Tiny-ImageNet and ImageNet-R under the conventional CIL setting. "Avg" represents the averaged performance after training each task, and "Last" represents the performance on all test samples after training the last task. + +first utilize a tokenizer to tokenize and acquire the embedding of the questions and answers. We leverage pre-trained encoder $f_{enc}$ and the projection layer to obtain the corresponding features for the input images: + +$$ +\mathbf {e} _ {i} = f _ {\text {e n c}} \left(\mathbf {x} _ {i}; \theta_ {\text {e n c}}\right). \tag {2} +$$ + +Then, the question embedding and the ground-truth embedding of this question, e.g., "This is a photo of [CLS]", is concatenated with image embedding. The final input of LLM Decoder $f_{dec}$ is: + +$$ +\hat {\mathbf {e}} _ {i} = \operatorname {C O N C A T E} (b o s, \mathbf {e} _ {i}, \mathbf {q}, \mathbf {s}, e o s). \tag {3} +$$ + +bos is the symbol of the sentence beginning, and $eos$ is the symbol for the end of the sentence. This encourages tokens at positions $m - 1$ to predict token $m$ : + +$$ +P \left(\hat {s} _ {1}, \hat {s} _ {2}, \dots , \hat {s} _ {m} \mid \mathbf {x} _ {i}, \mathbf {q}, \mathbf {s}\right) = \prod_ {j = 1} ^ {m - 1} P \left(s _ {j} \mid \mathbf {e} _ {i}, \mathbf {q}, s _ {1}, s _ {2}, \dots , s _ {j - 1}\right), \tag {4} +$$ + +where $s_j$ indicates the ground-truth answer token and $\hat{s}_m$ is the generated prediction. Then, we can compute the Cross Entropy loss as follows: + +$$ +\mathcal {L} _ {\mathrm {C E}} = - \frac {1}{m} \sum_ {j = 1} ^ {m} s _ {j} \cdot \log \hat {s} _ {j}. \tag {5} +$$ + +Inference. During inference, we use the updated projection layer in conjunction with the pre-trained encoder to obtain image features. These image features, combined with the question embeddings are then passed to the LLM Decoder to obtain the text output. + +$$ +\operatorname {p r e d} = \operatorname {a r g m a x} < f _ {\text {t e x t}} (\mathbf {s}), f _ {\text {t e x t}} (\hat {\mathbf {s}}) >, \tag {6} +$$ + +where $f_{\text{text}}$ is the text encoder, $<, >$ is the cosine similarity used to calculate the final predictions $pred$ . + +# 4. Experiments + +# 4.1. Experimental setups + +Datasets and Baselines. We conduct experiments in both conventional CIL and Few-shot CIL scenarios. In conventional CIL, we evaluate on three datasets. CIFAR100, TinyImageNet and ImageNet-R. CIFAR100 contains 60,000 images of $32 \times 32$ pixels in 100 categories. Each category has 600 images, of which 500 are for the training set, and 100 are for the test set. We experiment with two settings, B0-n and B50-n. The former splits 100 classes into $n$ tasks, + +![](images/5fc1dac2541b878b33a4ae6f17eed45101f963f3375d3362b89593ea76028575.jpg) + +![](images/68c76ab96796a0d53861c66ed96a3e434a3c01621d1bf11651bbd215f8d9b009.jpg) + +![](images/09d896fa62c4964f2a353a78e90082618608d797913106feb397f67719e05963.jpg) + +![](images/df95b220d179064318f406d516e71fd714fddeab26fb2cfbd4b2bcdcb40999e6.jpg) +Figure 3. Comparison of our method with other SOTA baselines on CIFAR100 and Tiny-ImageNet under the conventional CIL setting. + +![](images/54928a2f405f63f3ca0218c5723973effad0d62dcd2e329ad50350010bf0ac2a.jpg) + +![](images/bf7495e757d995760444f4bfcac38caef78453cbdbd5fc4c3c8c8d308c775fab.jpg) + +while the latter trains on 50 classes first and then distributes the other 50 classes across 5/10 tasks. + +Tiny-ImageNet contains 200 classes out of the original 1000 classes in ImageNet, with 550 images per class, of which 500 are in the training set and 50 are in the test set. The images are down-sampled to $64 \times 64$ pixels, which makes them easier to process and analyze. We train the first task in half 100 classes and split the other 100 classes into 5/10/20 tasks following [89]. + +ImageNet-R [23] contains 200 classes of images, which are included in the original ImageNet 1000 classes. However, many images are newly added and have various styles, such as sketch, painting, misc, etc. The dataset poses a great challenge for continual learning, as it has a wide diversity of image categories and styles and an uneven distribution of samples ranging from 45 to 500 per category. We follow [71] to split the dataset into 10 tasks, each containing 20 classes. + +In Few-shot CIL, we use CIFAR100 and miniImageNet [49] following the split proposed by [62]. For both datasets, we partition the data into two parts: base session and incremental sessions. The base session comprises 60 classes with all data available, while the incremental session follows a 5-way 5-shot setting, meaning that each session consists of only 5 classes with 5 samples each. + +In both conventional and few-shot scenarios, we com + +pare our method with some of the current state-of-the-art approaches, including conventional methods [7, 15, 16, 24, 29, 48, 53, 74, 76, 82, 89], pre-trained and prompt-based methods [55, 71, 72], and some specifically designed methods [12, 53, 62, 79] for few-shot scenarios. Additionally, we compare with a linear probe baseline, wherein features obtained from the image encoder are connected to a classifier for classification. We also consider the Zero-shot approach, where the generated text is directly used for classification without further fine-tuning. + +Implementation Details. We follow BLIP2 [27] to use the EVA-CLIP [59] pre-trained ViT-g/14 and BLIP2 pretrained Qfomer. We also used the MiniGPT-4 pre-trained projection layer checkpoint as our initial parameters. Under the many-shot "B0" setting, we employ a learning rate of 3e-7 and use a scheduler with cosine decay. The total training process consists of 2 epochs only. In B50 or B100 settings, we first train the linear layer with a learning rate of 3e-6 on the base classes, and then on the subsequent tasks, we adopt a lower learning rate of 3e-7, both employing a cosine decay scheduler. For the few-shot setting, we use a learning rate of 3e-6 for both base task and incremental tasks. We train one epoch for the base task and two epochs for incremental tasks. + +
012345678PD↓
iCaRL [48]61.3146.3242.9437.6330.4924.0020.8918.8017.2144.10
EEIL [7]61.3146.5844.0037.2933.1427.1224.1021.5719.5841.73
LUCIR [24]61.3147.8039.3131.9125.6821.3518.6717.2414.1747.14
TOPIC [62]61.3150.0945.1741.1637.4835.5232.1929.4624.4236.89
CEC [79]72.0066.8362.9759.4356.7053.7351.1949.2447.6324.37
F2M [53]72.0567.4763.1659.7056.7153.7751.1149.2147.8424.21
MetaFSCIL [12]72.0467.9463.7760.2957.5855.1652.9050.7949.1922.85
Entropy-reg [31]71.8467.1263.2159.7757.0153.9551.5549.5248.2123.63
L2P* [72]94.1287.2080.9975.6770.9466.7663.1159.8156.8337.29
DualPrompt* [71]93.9786.8580.6775.3170.6166.4462.7759.5856.8037.17
CODA-Prompt* [55]95.3788.8682.6977.8774.4770.1666.4663.7361.1434.23
Zero-shot58.0858.9557.7657.8958.1957.4256.2654.8254.953.13
GMM (Ours)89.3588.4086.1185.0783.6181.3578.9777.3475.1814.17
+ +Table 2. Comparison results of our method with other SOTA baselines on mini-ImageNet under the few-shot CIL setting. Phase 0 is the base task with all samples available, and the following 1-8 phases are the incremental 5-way 5-shot tasks. The reported accuracy is the test result on all seen classes after each session of training. "PD" represents Performance Prop between session 0 and session 8, with lower values indicating lower forgetting. * indicates our re-implementation based on PILOT [58]. + +# 4.2. Experiments on Conventional CIL + +In Table 1, we can see that our method outperforms all conventional methods by a large margin, including ResNet-based method DER and ViT-based DyTox. Note that without exemplar, our performance is a bit lower than Dual-Prompt and CODA-Prompt at B100-5 setting. We argue that their performance is mainly due to the backbone pretrained on ImageNet-21K, which largely overlaps with CIFAR100 and Tiny-ImageNet. Another interesting observation is that our method has better performance than all baselines under longer sequence settings (B100-10, B100-20). We believe this is because generative models do not rely on classification heads, making them less prone to bias toward the current task, resulting in less forgetting of past tasks. The Linear probe setting performs less than our method, indicating that our main contribution is not from the Large pre-trained ViT but the generation pipeline. In addition, the Zero-shot performance is superior to many traditional baselines, meaning that the Generative Multi-modal Models are indeed efficient Class-Incremental learners, but its output is less concise without fine-tuning (see Fig. 4). + +In Fig. 3, we compared our approach with some pretrained models on CIFAR100 and Tiny-ImageNet in terms of last task accuracy (all baselines are based on PILOT [58] and use 2000 exemplars). It can be observed that our method does not outperform other approaches in the initial tasks (0-2) and short sequence settings (B0-5, B100-5). + +This is because we do not rely on a supervised ImageNet-21K pre-trained backbone. Besides, we trained each task for only 1-2 epochs to ensure efficiency without sacrificing generalization. However, our method exhibits significant advantages in long sequences and later tasks. For instance, under the CIFAR100 B0-20 setting, we outperform CODA-Prompt by 10 points and DualPrompt by 7 points. + +# 4.3. Experiments on Few-shot CIL + +In Table 2, we compare our method with several baselines in the few-shot setting on mini-ImageNet. The evaluation metric is the model's accuracy across all the classes it has encountered so far. Our method outperforms conventional methods by a substantial margin in the final task, achieving an increase of more than $26\%$ . Furthermore, we surpass the best discriminative pre-trained approach CODA-Prompt by more than $14\%$ points. It's important to note that our accuracy in the first task (89.35) might not be as high as CODA-Prompt (95.37). However, in subsequent sessions, we consistently perform better than CODA-Prompt due to our ability to learn new tasks and retain knowledge of old tasks simultaneously. + +In Table 3, our method outperforms all other baselines on the CIFAR100 dataset of the few-shot setting, achieving a remarkably lower Performance Drop (PD) of 10.06. Furthermore, the Zero-shot baseline can achieve a very low PD due to the absence of forgetting. However, its overall performance is not very satisfying, as the length and con + +![](images/9eb814575495c5472761290ad93bc8cfc615a60b8c99f2b6b6049b8f77342136.jpg) +This is a photo of a statue of a bird with a long beak and a long neck. It is made of stone and has a blue sky background. +This is a pelican. + +Figure 4. Visual comparison examples of our method against frozen MiniGPT-4 [88] (Zero-shot). Text with a gray background is generated by MiniGPT-4 based on the image, while text with a orange background represents our method's output. The ground-truth labels are displayed on the left side of each row of images. All images displayed here are random sampled from ImageNet-R. +![](images/0326910f8b5be09208084d044d3a89818adb9175b3fb5d387183603fb7634fe0.jpg) +This is a painting of a fire truck. The fire truck is red and has a ladder on the side. The ladder is painted in white. The fire truck has a red and white stripe on the side. +This is a fire engine. + +![](images/14fd87dd31df3cc82c57882b4e2ce0b2f5b8a5823f7f3cf9c39fcba719bc9230.jpg) +This is a sculpture of a bird with its wings spread wide, as if it is flying. The bird is made of bronze and has a green patina. It is located in a park. +This is a pelican. + +![](images/37cebaae52d016cdf65733a87f2a55a88532afcd0422f8209dfcf44daef9238c.jpg) +This is a toy car with a duck in the driver's seat. The car is red and has a fire truck on the back. The duck is wearing a firefighter's helmet and has a fire hose in its hand. +This is a fire truck. + +
Method048PD↓
iCaRL [48]64.1027.9313.7350.37
EEIL [7]64.1028.9615.8548.25
LUCIR [24]64.1031.6113.5450.56
TOPIC [62]64.1040.1129.3734.73
CEC [79]73.0758.0949.1423.93
F2M [53]71.4557.7649.3522.06
MetaFSCIL [12]74.5059.4849.9724.53
Entropy-reg [31]74.4059.7150.1424.26
L2P* [72]91.2268.6654.8936.33
DualPrompt* [71]91.0868.4554.6736.41
CODA-Prompt* [55]93.5571.9159.3234.23
Zero-shot74.1372.5967.936.20
GMM (Ours)91.5385.6581.4710.06
+ +Table 3. Comparison results of our method with other SOTA baselines on CIFAR100 under the few-shot CIL setting. + +tent of its output are inconsistent and unpredictable without fine-tuning. + +# 4.4. Visualizations + +In Fig. 4, we present some comparison examples of our methods against GMM without fine-tuning [88]. We can see that GMM without fine-tuning provides an intuitive description of the overall image content with varying lengths of output text. However, it tends to recognize only broad categories (e.g., biar, car) and struggles with fine-grained categorization (e.g., pelican, truck). The descriptions are sometimes somewhat repetitive (e.g., first fire engine). In contrast, our fine-tuned method accurately identifies the im + +age's real category, even if there are occasional discrepancies with the true labels (e.g., "fire engine" vs. "fire truck"). Besides, with the assistance of the text encoder during the testing phase, our model can achieve correct classification results even when predicting similar but not identical text. + +# 5. Conclusion + +In this paper, we propose GMM to use generative models for class-incremental learning. By fine-tuning the Generative Multi-modal Model (GMM), we directly generate the label text of the images to be classified. Then we select the label most similar to the generated text by its features. Our experiments demonstrate that this method, which does not require a classification head, is highly effective in addressing classification biases in continual learning. + +Limitations. Since we are the first to introduce generative models to class-incremental learning, the overall design of our method is embarrassingly simple. We believe that with more focused efforts in this direction, there will be significant advancements in the field of continual learning. + +Broader impact. We believe that introducing GMM into continual learning (CL) is both necessary and urgent. With the rapid development of GMM, we can leverage their capabilities to improve the performance of continual learning. Besides, integrating CL methods into the training process of GMM could significantly reduce training costs. + +Acknowledgments. This work is funded by NSFC (NO. 62206135, 62225604), Young Elite Scientists Sponsorship Program by CAST (2023QNRC001), and the Fundamental Research Funds for the Central Universities (Nankai University, 070-63233085). Computation is supported by the Supercomputing Center of Nankai University. + +# References + +[1] Hongjoon Ahn, Jihwan Kwak, Subin Lim, Hyeonsu Bang, Hyojun Kim, and Taesup Moon. Ss-il: Separated softmax for incremental learning. In ICCV, 2021. 2 +[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 3 +[3] Rahaf Aljundi, Punarjay Chakravarty, and Tinne Tuytelaars. Expert gate: Lifelong learning with a network of experts. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3366-3375, 2017. 2 +[4] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2(3):4, 2023. 3 +[5] Hedi Ben-Younes, Rémi Cadene, Matthieu Cord, and Nicolas Thome. Mutan: Multimodal tucker fusion for visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 2612-2620, 2017. 3 +[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2 +[7] Francisco M Castro, Manuel J Marín-Jiménez, Nicolas Guil, Cordelia Schmid, and Karteek Alahari. End-to-end incremental learning. In Proceedings of the European conference on computer vision (ECCV), pages 233-248, 2018. 2, 5, 6, 7, 8 +[8] Jun Chen, Han Guo, Kai Yi, Boyang Li, and Mohamed Elhoseiny. Visualgpt: Data-efficient adaptation of pretrained language models for image captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18030-18040, 2022. 3 +[9] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 2 +[10] Xinlei Chen*, Saining Xie*, and Kaiming He. An empirical study of training self-supervised vision transformers. arXiv preprint arXiv:2104.02057, 2021. 2 +[11] Ali Cheraghian, Shafin Rahman, Pengfei Fang, Soumava Kumar Roy, Lars Petersson, and Mehrtash Harandi. Semantic-aware knowledge distillation for few-shot class-incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2534-2543, 2021. 2 +[12] Zhixiang Chi, Li Gu, Huan Liu, Yang Wang, Yuanhao Yu, and Jin Tang. Meta-scil: A meta-learning approach for few-shot class incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14166-14175, 2022. 6, 7, 8 + +[13] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2023. 3 +[14] Matthias Delange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Ales Leonardis, Greg Slabaugh, and Tinne Tuytelaars. A continual learning survey: Defying forgetting in classification tasks. TPAMI, 2021. 1, 2 +[15] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In ECCV, 2020. 2, 6 +[16] Arthur Douillard, Alexandre Ramé, Guillaume Couairon, and Matthieu Cord. Dytox: Transformers for continual learning with dynamic token expansion. In CVPR, 2022. 2, 5, 6 +[17] Chrisantha Fernando, Dylan Banarse, Charles Blundell, Yori Zwols, David Ha, Andrei A Rusu, Alexander Pritzel, and Daan Wierstra. Pathnet: Evolution channels gradient descent in super neural networks. arXiv preprint arXiv:1701.08734, 2017. 2 +[18] Rui Gao and Weiwei Liu. Ddgr: Continual learning with deep diffusion-based generative replay. In ICML, 2023. 2 +[19] Meng-Hao Guo, Cheng-Ze Lu, Zheng-Ning Liu, Ming-Ming Cheng, and Shi-Min Hu. Visual attention network. Computational Visual Media, 9(4):733-752, 2023. 1 +[20] Tyler L Hayes, Kushal Kafle, Robik Shrestha, Manoj Acharya, and Christopher Kanan. Remind your neural network to prevent catastrophic forgetting. In ECCV, 2020. 2 +[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3 +[22] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2 +[23] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, Dawn Song, Jacob Steinhardt, and Justin Gilmer. The many faces of robustness: A critical analysis of out-of-distribution generalization. ICCV, 2021. 6 +[24] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In CVPR, 2019. 2, 5, 6, 7, 8 +[25] Xinting Hu, Kaihua Tang, Chunyan Miao, Xian-Sheng Hua, and Hanwang Zhang. Distilling causal effect of data in class incremental learning. In CVPR, 2021. 2 +[26] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. 2, 5 +[27] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with + +frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 3, 6 +[28] Wenbo Li, Pengchuan Zhang, Lei Zhang, Qiuyuan Huang, Xiaodong He, Siwei Lyu, and Jianfeng Gao. Object-driven text-to-image synthesis via adversarial training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12174-12182, 2019. 3 +[29] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 2, 5, 6 +[30] Zhizhong Li and Derek Hoiem. Learning without forgetting. TPAMI, 2018. 2 +[31] Huan Liu, Li Gu, Zhixiang Chi, Yang Wang, Yuanhao Yu, Jun Chen, and Jin Tang. Few-shot class-incremental learning via entropy-regularized data-free replay. In European Conference on Computer Vision, pages 146–162. Springer, 2022. 7, 8 +[32] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 2, 3 +[33] Ming Liu, Yuxiang Wei, Xiaohe Wu, Wangmeng Zuo, and Lei Zhang. Survey on leveraging pre-trained generative adversarial networks for image editing and restoration. Science China Information Sciences, 66(5):151101, 2023. 1 +[34] Wei Liu, Sihan Chen, Longteng Guo, Xinxin Zhu, and Jing Liu. Cptr: Full transformer network for image captioning. arXiv preprint arXiv:2101.10804, 2021. 3 +[35] Xialei Liu, Marc Masana, Luis Herranz, Joost Van de Weijer, Antonio M Lopez, and Andrew D Bagdanov. Rotate your networks: Better weight consolidation and less catastrophic forgetting. In ICPR, 2018. 2 +[36] Xialei Liu, Yu-Song Hu, Xu-Sheng Cao, Andrew D Bagdanov, Ke Li, and Ming-Ming Cheng. Long-tailed class incremental learning. In European Conference on Computer Vision, pages 495-512. Springer, 2022. 2 +[37] Xialei Liu, Xusheng Cao, Haori Lu, Jia-wen Xiao, Andrew D Bagdanov, and Ming-Ming Cheng. Class incremental learning with pre-trained vision-language models. arXiv preprint arXiv:2310.20348, 2023. 3 +[38] Yaoyao Liu, Bernt Schiele, and Qianru Sun. Adaptive aggregation networks for class-incremental learning. In CVPR, 2021. 2 +[39] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019. 3 +[40] Arun Mallya, Dillon Davis, and Svetlana Lazebnik. Piggyback: Adapting a single network to multiple tasks by learning to mask weights. In ECCV, 2018. 2 +[41] Marc Masana, Xialei Liu, Bartlomiej Twardowski, Mikel Menta, Andrew D Bagdanov, and Joost Van De Weijer. Class-incremental learning: survey and performance evaluation on image classification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5):5513-5533, 2022. 1 +[42] Pratik Mazumder, Pravendra Singh, and Piyush Rai. Few-shot lifelong learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2337-2345, 2021. 2 + +[43] Michael McCloskey and Neal J Cohen. Catastrophic interference in connectionist networks: The sequential learning problem. In *Psychology of learning and motivation*, pages 109–165. Elsevier, 1989. 1 +[44] OpenAI. Gpt-4 technical report, 2023. 2, 3 +[45] Haoxuan Qu, Hossein Rahmani, Li Xu, Bryan Williams, and Jun Liu. Recent advances of continual learning in computer vision: An overview. arXiv preprint arXiv:2109.11369, 2021. 1 +[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3, 4 +[47] Jathushan Rajasegaran, Munawar Hayat, Salman Khan, Fahad Shahbaz Khan, and Ling Shao. Random path selection for incremental learning. In NIPS, 2019. 2 +[48] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In CVPR, 2017. 2, 5, 6, 7, 8 +[49] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 2, 6 +[50] Jonathan Schwarz, Wojciech Czarnecki, Jelena Luketina, Agnieszka Grabska-Barwinska, Yee Whye Teh, Razvan Pascanu, and Raia Hadsell. Progress & compress: A scalable framework for continual learning. In ICML, 2018. 2 +[51] Joan Serra, Didac Suris, Marius Miron, and Alexandros Karatzoglou. Overcoming catastrophic forgetting with hard attention to the task. In ICML, 2018. 2 +[52] Yijia Shao, Yiduo Guo, Dongyan Zhao, and Bing Liu. Class incremental learning based on label generation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 1263-1276, Toronto, Canada, 2023. Association for Computational Linguistics. 2 +[53] Guangyuan Shi, Jiaxin Chen, Wenlong Zhang, Li-Ming Zhan, and Xiao-Ming Wu. Overcoming catastrophic forgetting in incremental few-shot learning by finding flat minima. Advances in neural information processing systems, 34: 6747-6761, 2021. 3, 6, 7, 8 +[54] Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual learning with deep generative replay. Advances in neural information processing systems, 30, 2017. 2 +[55] James Seale Smith, Leonid Karlinsky, Vyshnavi Gutta, Paola Cascante-Bonilla, Donghyun Kim, Assaf Arbelle, Rameswar Panda, Rogerio Feris, and Zsolt Kira. Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11909-11919, 2023. 2, 3, 5, 6, 7, 8 +[56] Boyuan Sun, Yuqi Yang, Le Zhang, Ming-Ming Cheng, and Qibin Hou. Corrmatch: Label propagation via correlation matching for semi-supervised semantic segmentation. IEEE Computer Vision and Pattern Recognition (CVPR), 2024. 1 + +[57] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7464-7473, 2019. 3 +[58] Hai-Long Sun, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Pilot: A pre-trained model-based continual learning toolbox. arXiv preprint arXiv:2309.07117, 2023. 7 +[59] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389, 2023. 6 +[60] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019. 3 +[61] Xiaoyu Tao, Xinyuan Chang, Xiaopeng Hong, Xing Wei, and Yihong Gong. Topology-preserving class-incremental learning. In ECCV, 2020. 2 +[62] Xiaoyu Tao, Xiaopeng Hong, Xinyuan Chang, Songlin Dong, Xing Wei, and Yihong Gong. Few-shot class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12183-12192, 2020. 2, 6, 7, 8 +[63] Vishal Thengane, Salman Khan, Munawar Hayat, and Fahad Khan. Clip model is an efficient continual learner. arXiv preprint arXiv:2210.03114, 2022. 3, 5 +[64] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3 +[65] Maria Tsimpoukelli, Jacob L Menick, Serkan Cabi, SM Eslami, Oriol Vinyls, and Felix Hill. Multimodal few-shot learning with frozen language models. Advances in Neural Information Processing Systems, 34:200-212, 2021. 3 +[66] Gido M Van de Ven and Andreas S Tolias. Three scenarios for continual learning. NIPS Workshops, 2019. 1 +[67] Liuyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. A comprehensive survey of continual learning: Theory, method and application. arXiv preprint arXiv:2302.00487, 2023. 1 +[68] Liyuan Wang, Jingyi Xie, Xingxing Zhang, Mingyi Huang, Hang Su, and Jun Zhu. Hierarchical decomposition of prompt-based continual learning: Rethinking obscured suboptimality. Advances in Neural Information Processing Systems, 36, 2024. 3 +[69] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, et al. Vision: Large language model is also an open-ended decoder for vision-centric tasks. arXiv preprint arXiv:2305.11175, 2023. 3 +[70] Xiao Wang, Guangyao Chen, Guangwu Qian, Pengcheng Gao, Xiao-Yong Wei, Yaowei Wang, Yonghong Tian, and Wen Gao. Large-scale multi-modal pre-trained models: A comprehensive survey. Machine Intelligence Research, 20 (4):447-482, 2023. 3 +[71] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary + +prompting for rehearsal-free continual learning. In ECCV, 2022. 2, 3, 5, 6, 7, 8 +[72] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer G. Dy, and Tomas Pfister. Learning to prompt for continual learning. CVPR, 2022. 2, 3, 5, 6, 7, 8 +[73] Tz-Ying Wu, Gurumurthy Swaminathan, Zhizhong Li, Avinash Ravichandran, Nuno Vasconcelos, Rahul Bhotika, and Stefano Soatto. Class-incremental learning with strong pre-trained models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9601-9610, 2022. 3 +[74] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In CVPR, 2019. 2, 6 +[75] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In ICCV, 2019. 2 +[76] Shipeng Yan, Jiangwei Xie, and Xuming He. Der: Dynamically expandable representation for class incremental learning. In CVPR, 2021. 2, 6 +[77] Yang Yang, Zhiying Cui, Junjie Xu, Changhong Zhong, Wei-Shi Zheng, and Ruixuan Wang. Continual learning with bayesian model based on a fixed pre-trained feature extractor. Visual Intelligence, 1(1):5, 2023. 1 +[78] Lu Yu, Bartlomiej Twardowski, Xialei Liu, Luis Herranz, Kai Wang, Yongmei Cheng, Shangling Jui, and Joost van de Weijer. Semantic drift compensation for class-incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6982-6991, 2020. 2 +[79] Chi Zhang, Nan Song, Guosheng Lin, Yun Zheng, Pan Pan, and Yinghui Xu. Few-shot incremental learning with continually evolved classifiers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12455-12464, 2021. 3, 6, 7, 8 +[80] Chang-Bin Zhang, Jia-Wen Xiao, Xialei Liu, Ying-Cong Chen, and Ming-Ming Cheng. Representation compensation networks for continual semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7053-7064, 2022. 2 +[81] Gengwei Zhang, Liyuan Wang, Guoliang Kang, Ling Chen, and Yunchao Wei. Slca: Slow learner with classifier alignment for continual learning on a pre-trained model. arXiv preprint arXiv:2303.05118, 2023. 2, 3 +[82] Bowen Zhao, Xi Xiao, Guojun Gan, Bin Zhang, and Shu-Tao Xia. Maintaining discrimination and fairness in class incremental learning. In CVPR, 2020. 6 +[83] Hanbin Zhao, Yongjian Fu, Mintong Kang, Qi Tian, Fei Wu, and Xi Li. Mgmtf: Multi-grained slow vs. fast framework for few-shot class-incremental learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 2 +[84] Da-Wei Zhou, Qi-Wei Wang, Zhi-Hong Qi, Han-Jia Ye, De-Chuan Zhan, and Ziwei Liu. Deep class-incremental learning: A survey. arXiv preprint arXiv:2302.03648, 2023. 1 +[85] Da-Wei Zhou, Han-Jia Ye, De-Chuan Zhan, and Ziwei Liu. Revisiting class-incremental learning with pre-trained mod + +els: Generalizability and adaptivity are all you need, 2023. 2, 3 +[86] Da-Wei Zhou, Yuanhan Zhang, Jingyi Ning, Han-Jia Ye, De-Chuan Zhan, and Ziwei Liu. Learning without forgetting for vision-language models. arXiv preprint arXiv:2305.19270, 2023. 2, 3 +[87] Qiang Zhou, Chaohui Yu, Shaofeng Zhang, Sitong Wu, Zhibing Wang, and Fan Wang. Regionblip: A unified multimodal pre-training framework for holistic and regional comprehension. arXiv preprint arXiv:2308.02299, 2023. 3 +[88] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 3, 8 +[89] Fei Zhu, Xu-Yao Zhang, Chuang Wang, Fei Yin, and ChengLin Liu. Prototype augmentation and self-supervision for incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5871-5880, 2021. 2, 5, 6 +[90] Kai Zhu, Yang Cao, Wei Zhai, Jie Cheng, and Zheng-Jun Zha. Self-promoted prototype refinement for few-shot class incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6801-6810, 2021. 3 \ No newline at end of file diff --git a/2024/Generative Multi-modal Models are Good Class Incremental Learners/images.zip b/2024/Generative Multi-modal Models are Good Class Incremental Learners/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e7af8a00024a50d15f624ccd0514d12158270510 --- /dev/null +++ b/2024/Generative Multi-modal Models are Good Class Incremental Learners/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1b5eae37ed814808ae52638c49c25c2f02561a55f7906c9847a72372e83b5135 +size 694338 diff --git a/2024/Generative Multi-modal Models are Good Class Incremental Learners/layout.json b/2024/Generative Multi-modal Models are Good Class Incremental Learners/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..08a2b74493bf5af9243bff3018fe52e5ef88638f --- /dev/null +++ b/2024/Generative Multi-modal Models are Good Class Incremental Learners/layout.json @@ -0,0 +1,9045 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 78, + 103, + 515, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 103, + 515, + 120 + ], + "spans": [ + { + "bbox": [ + 78, + 103, + 515, + 120 + ], + "type": "text", + "content": "Generative Multi-modal Models are Good Class-Incremental Learners" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 142, + 485, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 142, + 485, + 171 + ], + "spans": [ + { + "bbox": [ + 107, + 142, + 485, + 171 + ], + "type": "text", + "content": "Xusheng Cao" + }, + { + "bbox": [ + 107, + 142, + 485, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 107, + 142, + 485, + 171 + ], + "type": "text", + "content": ", Haori Lu" + }, + { + "bbox": [ + 107, + 142, + 485, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 107, + 142, + 485, + 171 + ], + "type": "text", + "content": ", Linlan Huang" + }, + { + "bbox": [ + 107, + 142, + 485, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 107, + 142, + 485, + 171 + ], + "type": "text", + "content": ", Xialei Liu" + }, + { + "bbox": [ + 107, + 142, + 485, + 171 + ], + "type": "inline_equation", + "content": "^{2,1*}" + }, + { + "bbox": [ + 107, + 142, + 485, + 171 + ], + "type": "text", + "content": ", Ming-Ming Cheng" + }, + { + "bbox": [ + 107, + 142, + 485, + 171 + ], + "type": "inline_equation", + "content": "^{2,1}" + }, + { + "bbox": [ + 107, + 142, + 485, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 107, + 142, + 485, + 171 + ], + "type": "text", + "content": "VCIP, CS, Nankai University " + }, + { + "bbox": [ + 107, + 142, + 485, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 107, + 142, + 485, + 171 + ], + "type": "text", + "content": "NKIARI, Shenzhen Futian" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 74, + 173, + 515, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 173, + 515, + 186 + ], + "spans": [ + { + "bbox": [ + 74, + 173, + 515, + 186 + ], + "type": "text", + "content": "{caoxusheng, luhaori, huanglinlan}@mail.nankai.edu.cn, {xialei, cmm}@nankai.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 239, + 290, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 290, + 539 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 290, + 539 + ], + "type": "text", + "content": "In class-incremental learning (CIL) scenarios, the phenomenon of catastrophic forgetting caused by the classifier's bias towards the current task has long posed a significant challenge. It is mainly caused by the characteristic of discriminative models. With the growing popularity of the generative multi-modal models, we would explore replacing discriminative models with generative ones for CIL. However, transitioning from discriminative to generative models requires addressing two key challenges. The primary challenge lies in transferring the generated textual information into the classification of distinct categories. Additionally, it requires formulating the task of CIL within a generative framework. To this end, we propose a novel generative multi-modal model (GMM) framework for class-incremental learning. Our approach directly generates labels for images using an adapted generative model. After obtaining the detailed text, we use a text encoder to extract text features and employ feature matching to determine the most similar label as the classification prediction. In the conventional CIL settings, we achieve significantly better results in long-sequence task scenarios. Under the Few-shot CIL setting, we have improved by at least " + }, + { + "bbox": [ + 46, + 239, + 290, + 539 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 46, + 239, + 290, + 539 + ], + "type": "text", + "content": " accuracy over all the current state-of-the-art methods with significantly less forgetting. Our code is available at https://github.com/DoubleClass/GMM." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 563, + 128, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 128, + 575 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 128, + 575 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 584, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 584, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 584, + 287, + 693 + ], + "type": "text", + "content": "Deep neural networks [19, 33, 56] have made remarkable strides in numerous applications, primarily owing to the vast amounts of data and computational resources at their disposal. Nonetheless, these accomplishments are predominantly contingent on having access to all the required data simultaneously for training on various tasks. In cases where data is acquired incrementally, these networks often encounter the challenge of catastrophic forgetting [43]. Hence, the capacity to seamlessly incorporate new knowl" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 212, + 542, + 268 + ], + "blocks": [ + { + "bbox": [ + 309, + 212, + 542, + 268 + ], + "lines": [ + { + "bbox": [ + 309, + 212, + 542, + 268 + ], + "spans": [ + { + "bbox": [ + 309, + 212, + 542, + 268 + ], + "type": "image", + "image_path": "2073830f54ab6f9c767edec7e8011939c055c676c9226f2286e014db165462d5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 309, + 269, + 544, + 372 + ], + "blocks": [ + { + "bbox": [ + 309, + 269, + 544, + 372 + ], + "lines": [ + { + "bbox": [ + 309, + 269, + 544, + 372 + ], + "spans": [ + { + "bbox": [ + 309, + 269, + 544, + 372 + ], + "type": "image", + "image_path": "faa7a05aa9b57d9a251de60b9c1b9de4a9bef9f4ffae2aec368dcb51f6459e26.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 380, + 547, + 458 + ], + "lines": [ + { + "bbox": [ + 305, + 380, + 547, + 458 + ], + "spans": [ + { + "bbox": [ + 305, + 380, + 547, + 458 + ], + "type": "text", + "content": "Figure 1. Illustration of conventional discriminative models for class-incremental learning (CIL) and our generative multi-modal models (GMM) for CIL. Discriminative models pose potential risk of classifier bias toward current task with network expansion. Our GMM framework consists generation and classification phases. And It is adapted to CIL based on the similarity of generated text and the true category names." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 482, + 547, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 482, + 547, + 542 + ], + "spans": [ + { + "bbox": [ + 304, + 482, + 547, + 542 + ], + "type": "text", + "content": "edge while retaining previously acquired knowledge is a highly desirable attribute for future artificial intelligence systems. Continual learning [45, 67, 77, 84] is a subject of study aimed at advancing the evolution of neural networks toward this goal." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 544, + 547, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 547, + 698 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 547, + 698 + ], + "type": "text", + "content": "Numerous studies have delved into continual learning, categorizing their approaches into three main groups [14]: rehearsal-based, architecture-based, and regularization-based methods. Additionally, hybrid methods are gaining popularity as they combine insights from different perspectives. Within this research landscape, three primary scenarios [66] have received extensive attention, with class-incremental learning (CIL) [41] being one of the most demanding settings. In our work, we concentrate on CIL, wherein each task comprises a distinct set of classes, and the primary challenge is to enable the network to recognize new classes without forgetting knowledge of previously encountered ones." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "text", + "content": "The majority of existing research on class-incremental" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 703, + 132, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 132, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 132, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28706" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "text", + "content": "learning (CIL) focuses on training models from scratch, relying solely on data from the current tasks [7, 15, 16, 24, 26, 29, 48, 74, 76, 89]. In contrast, humans accumulate knowledge over extended periods, drawing upon a wealth of prior world knowledge. Consequently, there has been a growing interest in pre-trained models for CIL [55, 71, 72, 81, 86], harnessing the knowledge acquired from extensive, pre-existing datasets to address the immediate tasks at hand. For instance, prompt-based methods [55, 71, 72] utilize prompt-tuning to summarize task-specific knowledge from prior pretrained knowledge. SLCA [81] and ADAM [85], on the other hand, solely fine-tuning the pre-trained model to adapt the pre-existing knowledge into the objectives of immediate tasks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 242, + 289, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 242, + 289, + 410 + ], + "spans": [ + { + "bbox": [ + 46, + 242, + 289, + 410 + ], + "type": "text", + "content": "To tackle image classification downstream tasks, pretrained models are traditionally derived from discriminative tasks, like supervised learning on datasets such as ImageNet-21K [49], or they may stem from self-supervised learning efforts [6, 9, 10, 22]. However, in our research, we venture into the paradigm of generative multi-modal models to address image classification tasks. Generative models like GPT4 [44] and LLaVa [32] have garnered significant attention in recent years due to their capacity to produce highly informative descriptions of input images. On the one hand, it can harness the wealth of semantic correspondences between texts and images, while on the other hand, there's no requirement to expand the classifier with each new task, unlike in the case of discriminative models for CIL." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 412, + 289, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 412, + 289, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 412, + 289, + 592 + ], + "type": "text", + "content": "Nonetheless, harnessing the knowledge from pre-trained generative models for downstream class-incremental learning (CIL) tasks presents a nontrivial endeavor. The primary challenge lies in transferring the generated textual information into the classification of distinct categories. Additionally, there's the task of formulating CIL within a generative framework, which poses a second significant challenge. Shao et al. [52] presents the VAG system, which formulates CIL as a continual label generation problem, preserving the language model's ability to learn new classes. However, it only works in the field of Natural Language Processing (NLP), which is inherently suited to Large Language Models (LLM). To the best of our knowledge, we are the first to apply this generative approach to incremental learning in the field of image classification." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "In this work, we propose Generative Multi-modal Models (GMM) for class-incremental learning. As illustrated in Fig. 1 (a), conventional discriminative methods extract image features with a network backbone, then forward them to a classifier to obtain the probability of the image belonging to each label, with the label having the highest probability being the output of the discriminative models. While in Fig. 1 (b), we adopt a generative approach, which directly produces a descriptive sentence for the given image, which is then compared with the actual label texts with a" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "content": "text encoder. The most similar label becomes the predicted result of our generative model. This approach allows us to leverage the rich pre-training knowledge in generative multi-modal models while avoiding the use of the expanded classification head, which mitigates the risk of the model bias towards the current task and reduces catastrophic forgetting." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 156, + 485, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 156, + 485, + 168 + ], + "spans": [ + { + "bbox": [ + 317, + 156, + 485, + 168 + ], + "type": "text", + "content": "The main contributions of this paper are:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 171, + 547, + 314 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 306, + 171, + 545, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 171, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 306, + 171, + 545, + 205 + ], + "type": "text", + "content": "- We propose a novel generative approach (GMM) to address class-incremental learning by leveraging multimodal models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 206, + 547, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 206, + 547, + 277 + ], + "spans": [ + { + "bbox": [ + 306, + 206, + 547, + 277 + ], + "type": "text", + "content": "- We reformulate GMM for image classification and adapt it for the downstream benchmarks. Without an expanded classification head like in discriminative models, our model significantly mitigates the issue of bias towards current tasks, resulting in significantly reduced forgetting in CIL." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 278, + 547, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 278, + 547, + 314 + ], + "spans": [ + { + "bbox": [ + 306, + 278, + 547, + 314 + ], + "type": "text", + "content": "- Our model achieves state-of-the-art performance across multiple datasets in both conventional and few-shot CIL settings." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 325, + 392, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 325, + 392, + 338 + ], + "spans": [ + { + "bbox": [ + 306, + 325, + 392, + 338 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 345, + 462, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 345, + 462, + 358 + ], + "spans": [ + { + "bbox": [ + 306, + 345, + 462, + 358 + ], + "type": "text", + "content": "2.1. Class-Incremental Learning" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 363, + 547, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 363, + 547, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 547, + 651 + ], + "type": "text", + "content": "In Class-Incremental Learning, tasks arrive sequentially, and each class is exclusive to a specific task without any overlap. The goal is to acquire knowledge from new classes while preserving information from previously encountered classes. There are three primary branches in CIL [14], including rehearsal-based, architecture-based, and regularization-based methods. Rehearsal-based methods [1, 7, 48, 75] store a small set of data derived from old classes to represent knowledge from previous tasks. These exemplar data can be either original data [48], generative data [18, 54] or hidden features [20]. Architecture-based methods focus on modifying network architecture to alleviate forgetting. Approaches include learning redundant network architecture [17, 47], learning different expert networks [3, 50] or parameters [38, 40, 51] for each task, dynamically expanding network parameters to accumulate incremental knowledge [76]. Regularization-based methods introduce an additional regularization term to restrict network updates when adapting to new tasks. In such cases, EWC [26], SDC [78] and Rotated-EWC [35] expect that parameters essential for the old tasks should not be updated excessively. Moreover, from the perspective of network output consistency, numerous studies [25, 30, 36, 61, 80] incorporate distillation to prevent forgetting." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "Few-shot CIL Few-shot Class-Incremental Learning (FSCIL) [42, 62] explores few-shot learning in an incremental context, with all data samples available for base session and very limited data in each incremental session. Some FSCIL methods [11, 62, 83] train the model in both base and incre" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28707" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": "mental sessions, aiming to mitigate overfitting challenges caused by the limited data in incremental learning. Other strategies [53, 79, 90] primarily train the model in the base session and make minimal adjustments in the incremental sessions, thereby reducing forgetting but may come at the cost of decreased precision in the incremental sessions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 152, + 198, + 165 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 152, + 198, + 165 + ], + "spans": [ + { + "bbox": [ + 47, + 152, + 198, + 165 + ], + "type": "text", + "content": "2.2. Pre-trained models for CIL" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 171, + 290, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 171, + 290, + 435 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 290, + 435 + ], + "type": "text", + "content": "There are many methods [63, 72, 73] having shown that pretrained models are effective for continual learning. One main branch trains a set of prompts to retain previous knowledge [55, 68, 71, 72]. A selected subset of prompts are fed into the model during forward to prompt model the past knowledge. Additionally, methods like SLCA [81] and ADAM [85] fine-tune pre-trained models, achieving impressive results with less forgetting. Continual-CLIP [63] demonstrates that the CLIP [46] model is capable of performing continual learning without any extra training. This highlights the significant potential of multi-modal pretrained models in the realm of continual learning. Inspired by this, many methods [37, 86] employ CLIP as the backbone to utilize the multi-modal information. However, if not using a classifier directly, these approaches need to utilize expanded text features to calculate distances with image features for classification. This will exacerbate the model's bias towards current data, consequently leading to the forgetting of previously acquired knowledge. To avoid this bias, we use a generative model to directly generate prediction text. The fixed text decoder will function as the classifier, significantly alleviating the bias." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 443, + 186, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 443, + 186, + 456 + ], + "spans": [ + { + "bbox": [ + 47, + 443, + 186, + 456 + ], + "type": "text", + "content": "2.3. Vision Language Models" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "text", + "content": "In recent years, vision-language multi-modal models have made significant progress and achieved impressive results in various downstream tasks [5, 28, 34, 70]. Traditional vision-language models employ different types of encoders to extract information from vision and language models, including single-stream [57], dual-stream [39] and fusion [60] encoders. A key aspect of vision-language models is the alignment of multi-modal features. CLIP [46], for example, extracts image and text features separately using respective encoders and enforces alignment through a contrastive loss, ensuring alignment between positive image-text pairs in the feature space. VisualGPT [8] and Frozen [65] leverage pre-trained models as encoders for visual-language tasks. From then on, the utilization of pre-trained models in vision-language tasks became more and more popular. For instance, Flamingo [2] and BLIP-2 [27] align the pre-trained image and text encoders employing gated cross-attention and Q-Former, respectively. Furthermore, LLaVA [32] and MiniGPT-4 [88] leverage more robust Large Language Models (LLM) [13, 64] as text encoders, while only training a projection layer for alignment. With" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "content": "the increasing popularity of LLM, an increasing number of studies [4, 69, 87] explore the potential of multi-modal LLMs for vision-language tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 120, + 362, + 132 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 120, + 362, + 132 + ], + "spans": [ + { + "bbox": [ + 305, + 120, + 362, + 132 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 140, + 545, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 140, + 545, + 189 + ], + "spans": [ + { + "bbox": [ + 304, + 140, + 545, + 189 + ], + "type": "text", + "content": "In this section, we introduce the preliminaries of class-incremental learning and generative multi-modal models. Then, we present our approach to leverage generative models for CIL and the corresponding learning process." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 197, + 392, + 209 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 197, + 392, + 209 + ], + "spans": [ + { + "bbox": [ + 305, + 197, + 392, + 209 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 218, + 545, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 218, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 304, + 218, + 545, + 326 + ], + "type": "text", + "content": "Class-Incremental Learning. Given N tasks " + }, + { + "bbox": [ + 304, + 218, + 545, + 326 + ], + "type": "inline_equation", + "content": "T = \\{T_{1}, T_{2}, \\dots, T_{N}\\}" + }, + { + "bbox": [ + 304, + 218, + 545, + 326 + ], + "type": "text", + "content": ", the goal of class-incremental learning is to learn each task " + }, + { + "bbox": [ + 304, + 218, + 545, + 326 + ], + "type": "inline_equation", + "content": "T_{t}" + }, + { + "bbox": [ + 304, + 218, + 545, + 326 + ], + "type": "text", + "content": " with its associated data " + }, + { + "bbox": [ + 304, + 218, + 545, + 326 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{X}_{t}, \\mathbf{Y}_{t}\\}" + }, + { + "bbox": [ + 304, + 218, + 545, + 326 + ], + "type": "text", + "content": " in a sequential order. For each task, it contains samples " + }, + { + "bbox": [ + 304, + 218, + 545, + 326 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_i, \\mathbf{y}_i\\}" + }, + { + "bbox": [ + 304, + 218, + 545, + 326 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 218, + 545, + 326 + ], + "type": "inline_equation", + "content": "i = 1, \\dots, n_t" + }, + { + "bbox": [ + 304, + 218, + 545, + 326 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 218, + 545, + 326 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 304, + 218, + 545, + 326 + ], + "type": "text", + "content": " is the images and " + }, + { + "bbox": [ + 304, + 218, + 545, + 326 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_i" + }, + { + "bbox": [ + 304, + 218, + 545, + 326 + ], + "type": "text", + "content": " is the corresponding one-hot labels. Typically, " + }, + { + "bbox": [ + 304, + 218, + 545, + 326 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_i \\cap \\mathbf{X}_j = \\emptyset, \\forall i \\neq j" + }, + { + "bbox": [ + 304, + 218, + 545, + 326 + ], + "type": "text", + "content": ". At inference, the model is tested on all seen tasks without task IDs. In some scenarios, fixed memory storage is set to keep a few samples of previous tasks to prevent forgetting." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 326, + 545, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 326, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 304, + 326, + 545, + 458 + ], + "type": "text", + "content": "Normally, a CIL model consists of a feature extractor and a classifier head " + }, + { + "bbox": [ + 304, + 326, + 545, + 458 + ], + "type": "inline_equation", + "content": "F = \\{f_{\\theta}, \\mathcal{H}_{\\phi}\\}" + }, + { + "bbox": [ + 304, + 326, + 545, + 458 + ], + "type": "text", + "content": " which are parameterized by " + }, + { + "bbox": [ + 304, + 326, + 545, + 458 + ], + "type": "inline_equation", + "content": "\\{\\theta, \\phi\\}" + }, + { + "bbox": [ + 304, + 326, + 545, + 458 + ], + "type": "text", + "content": ". In traditional class-incremental learning, " + }, + { + "bbox": [ + 304, + 326, + 545, + 458 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 326, + 545, + 458 + ], + "type": "text", + "content": " is usually a modified ResNet [21] with all parameters tunable. In pre-trained or prompt-based methods, " + }, + { + "bbox": [ + 304, + 326, + 545, + 458 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 326, + 545, + 458 + ], + "type": "text", + "content": " represents fewer trainable parameters like a linear adaptor or a couple of prompts. " + }, + { + "bbox": [ + 304, + 326, + 545, + 458 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 326, + 545, + 458 + ], + "type": "text", + "content": " is a linear classifier head projecting image features to probability predictions, which has to be expanded for each new task in order to make predictions for the new classes. The conventional Cross Entropy loss is often used for updating " + }, + { + "bbox": [ + 304, + 326, + 545, + 458 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 326, + 545, + 458 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 326, + 545, + 458 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 326, + 545, + 458 + ], + "type": "text", + "content": ", which for task " + }, + { + "bbox": [ + 304, + 326, + 545, + 458 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 326, + 545, + 458 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 478, + 545, + 518 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 478, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 312, + 478, + 545, + 518 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {C E}} \\left(\\mathbf {X} _ {t}, \\mathbf {Y} _ {t}; \\theta , \\phi\\right) = - \\frac {1}{n _ {t}} \\sum_ {i = 1} ^ {n _ {t}} \\mathbf {y} _ {i} \\cdot \\log \\mathcal {H} (f (\\mathbf {x} _ {i}; \\theta); \\phi). \\tag {1}", + "image_path": "baab134f539c82db80b0989e28460397140b443022ca9a6b21a75aa2de7ec84f.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 519, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 519, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 519, + 545, + 579 + ], + "type": "text", + "content": "In the continual learning process, the parameter " + }, + { + "bbox": [ + 304, + 519, + 545, + 579 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 519, + 545, + 579 + ], + "type": "text", + "content": " can easily deviate to the data of the current task due to the absence or scarcity of old samples in the previous tasks, resulting in forgetting the previously acquired knowledge and deteriorating the overall performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "Generative Multi-Modal Models (GMM). Multi-modal models have demonstrated exceptional performance in generating detailed image descriptions by incorporating both visual and textual information. Notably, GPT-4 [44] stands out as an advanced model proficient in generating comprehensive image descriptions and providing explanations for the depicted content. Furthermore, MiniGPT-4 [88] proposes a two-stage fine-tuning process that aligns image features and large language models, enabling LLaMa [64] to recognize images and conduct further dialogue based on the image content." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28708" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 74, + 324, + 281 + ], + "blocks": [ + { + "bbox": [ + 55, + 74, + 324, + 281 + ], + "lines": [ + { + "bbox": [ + 55, + 74, + 324, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 74, + 324, + 281 + ], + "type": "image", + "image_path": "0c76f35c4278ab83e6618a5a6e79752f02729fc226ffe59b2694249f5ebdc057.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 290, + 547, + 335 + ], + "lines": [ + { + "bbox": [ + 46, + 290, + 547, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 290, + 547, + 335 + ], + "type": "text", + "content": "Figure 2. The overview structure of our proposed method. The conceptual illustration of the generative multi-modal model (GMM) is shown on the left. In order to adapt this model for CIL, We have to turn the GMM model for classification and further adapt to our objective benchmark for learning (see Sec. 3.2). On the right side, we demonstrate how the final evaluation is taken for CIL with all seen classes. Text encoder is used for obtaining the embeddings for similarity prediction." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 343, + 74, + 533, + 281 + ], + "blocks": [ + { + "bbox": [ + 343, + 74, + 533, + 281 + ], + "lines": [ + { + "bbox": [ + 343, + 74, + 533, + 281 + ], + "spans": [ + { + "bbox": [ + 343, + 74, + 533, + 281 + ], + "type": "image", + "image_path": "114042b0e1f8d66755f9eb4f73250f247fb96697ae64dece69a0f5a6f6208f5d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 356, + 289, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 356, + 289, + 489 + ], + "spans": [ + { + "bbox": [ + 46, + 356, + 289, + 489 + ], + "type": "text", + "content": "As shown in Fig. 2, these models consist of an encoder " + }, + { + "bbox": [ + 46, + 356, + 289, + 489 + ], + "type": "inline_equation", + "content": "f_{enc}" + }, + { + "bbox": [ + 46, + 356, + 289, + 489 + ], + "type": "text", + "content": " to generate content embeddings including image embedding and text embedding, which is further used as the input of an auto-aggressive decoder " + }, + { + "bbox": [ + 46, + 356, + 289, + 489 + ], + "type": "inline_equation", + "content": "f_{dec}" + }, + { + "bbox": [ + 46, + 356, + 289, + 489 + ], + "type": "text", + "content": " to generate image descriptions. Input image " + }, + { + "bbox": [ + 46, + 356, + 289, + 489 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 46, + 356, + 289, + 489 + ], + "type": "text", + "content": " is encoded as an image embedding " + }, + { + "bbox": [ + 46, + 356, + 289, + 489 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 46, + 356, + 289, + 489 + ], + "type": "text", + "content": ", and question embeddings " + }, + { + "bbox": [ + 46, + 356, + 289, + 489 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 46, + 356, + 289, + 489 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 356, + 289, + 489 + ], + "type": "inline_equation", + "content": "q_1, \\dots, q_l" + }, + { + "bbox": [ + 46, + 356, + 289, + 489 + ], + "type": "text", + "content": " can be concatenated with the image embedding together to generate answer embeddings " + }, + { + "bbox": [ + 46, + 356, + 289, + 489 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 46, + 356, + 289, + 489 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 356, + 289, + 489 + ], + "type": "inline_equation", + "content": "s_1, \\dots, s_m" + }, + { + "bbox": [ + 46, + 356, + 289, + 489 + ], + "type": "text", + "content": ". The output tokens are generated one by one with the condition of previously generated tokens. For instance, " + }, + { + "bbox": [ + 46, + 356, + 289, + 489 + ], + "type": "inline_equation", + "content": "s_m" + }, + { + "bbox": [ + 46, + 356, + 289, + 489 + ], + "type": "text", + "content": " is generated with all previous " + }, + { + "bbox": [ + 46, + 356, + 289, + 489 + ], + "type": "inline_equation", + "content": "m - 1" + }, + { + "bbox": [ + 46, + 356, + 289, + 489 + ], + "type": "text", + "content": " tokens (see the decoder in Fig. 2)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 498, + 260, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 498, + 260, + 510 + ], + "spans": [ + { + "bbox": [ + 47, + 498, + 260, + 510 + ], + "type": "text", + "content": "3.2. Generative Multi-Modal Models for CIL" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 517, + 287, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 517, + 287, + 673 + ], + "spans": [ + { + "bbox": [ + 46, + 517, + 287, + 673 + ], + "type": "text", + "content": "We adhere to the foundational settings of MiniGPT-4, incorporating a frozen image encoder " + }, + { + "bbox": [ + 46, + 517, + 287, + 673 + ], + "type": "inline_equation", + "content": "f_{enc}" + }, + { + "bbox": [ + 46, + 517, + 287, + 673 + ], + "type": "text", + "content": " followed by a trainable projection layer for adaptation to downstream tasks, as shown in Fig. 2. Our primary innovation involves the direct utilization of generative models to produce text, which can then serve as a basis for discriminative classification. However, two major challenges need addressing. First, there is the issue of adapting generative multi-modal models for classification, given that the generated text may differ significantly from class names. Second, we must devise a mechanism for our classification benchmarks to learn in a manner consistent with generative multi-modal models. We introduce these two aspects as follows." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "Turning GMM for classification. We employ a distance metric to bridge the gap between generative and discriminative models. During training, we use the real label's text to" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 356, + 545, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 356, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 304, + 356, + 545, + 475 + ], + "type": "text", + "content": "encourage the model to predict the label of an image with a concise and accurate sentence in the format of \"This is a photo of [CLS].\" avoiding detailed descriptions of all contents in the image. During testing, the model follows the format to output the category text for a given image. We extract the content in \"[CLS]\", then obtain its text features using the CLIP [46] text encoder " + }, + { + "bbox": [ + 304, + 356, + 545, + 475 + ], + "type": "inline_equation", + "content": "f_{\\text{text}}" + }, + { + "bbox": [ + 304, + 356, + 545, + 475 + ], + "type": "text", + "content": ", and compute the distance with text features of all categories seen by now. The closest class was then considered the final prediction of the generative model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 479, + 546, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 546, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 546, + 647 + ], + "type": "text", + "content": "Converting CIL Benchmarks for adaptation. CIL is usually evaluated on ImageNet, CIFAR-100, and ImageNet-R datasets. These datasets usually consist of images and corresponding one-hot labels " + }, + { + "bbox": [ + 304, + 479, + 546, + 647 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{X}_t,\\mathbf{Y}_t\\}" + }, + { + "bbox": [ + 304, + 479, + 546, + 647 + ], + "type": "text", + "content": ". Using the CIFAR100 dataset as an example, we pair each image with a sentence to form an image-text pair format " + }, + { + "bbox": [ + 304, + 479, + 546, + 647 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{X}_t,\\mathbf{S}_t\\}" + }, + { + "bbox": [ + 304, + 479, + 546, + 647 + ], + "type": "text", + "content": " with the template: \"This is a photo of [CLS]\", where \"[CLS]\" is the label name of that category, such as apple, dog, etc. Next, we partitioned the 100 classes into various tasks based on different settings and fed them into the model sequentially. After completing the training on task " + }, + { + "bbox": [ + 304, + 479, + 546, + 647 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 479, + 546, + 647 + ], + "type": "text", + "content": ", the model should be capable of classifying all the classes encompassed from task 0 to task " + }, + { + "bbox": [ + 304, + 479, + 546, + 647 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 479, + 546, + 647 + ], + "type": "text", + "content": ". Note that only the linear projection layer is updated for further adaptation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 655, + 459, + 668 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 655, + 459, + 668 + ], + "spans": [ + { + "bbox": [ + 306, + 655, + 459, + 668 + ], + "type": "text", + "content": "3.3. Optimization and Inference" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 677, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 546, + 715 + ], + "type": "text", + "content": "Optimization. For each task " + }, + { + "bbox": [ + 304, + 677, + 546, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 677, + 546, + 715 + ], + "type": "text", + "content": ", we obtain the current task's image-text pair " + }, + { + "bbox": [ + 304, + 677, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{X}_t, \\mathbf{S}_t\\}" + }, + { + "bbox": [ + 304, + 677, + 546, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 677, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_t" + }, + { + "bbox": [ + 304, + 677, + 546, + 715 + ], + "type": "text", + "content": " contains the corresponding sentence of each image. During training, we" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28709" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 60, + 70, + 535, + 353 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 535, + 353 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 535, + 353 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 535, + 353 + ], + "type": "table", + "html": "
TypeMethodExemplarTiny-ImageNetImageNet-R
5 tasks10 tasks20 tasks10 tasks
AvgLastAvgLastAvgLastLast
ConventionalEWC [26]X19.016.0015.823.7912.354.7335.00
LwF [29]X22.317.3417.344.7312.484.2638.50
iCaRL [48]✓45.9534.6043.2233.2237.8527.54-
EEIL [7]✓47.1735.1245.0334.6440.4129.72-
UCIR [24]✓50.3039.4248.5837.2942.8430.85-
PASS [89]X49.5441.6447.1939.2742.0132.93-
DyTox [16]✓55.5847.2352.2642.7946.1836.21-
Discriminative PT modelsContinual-CLIP[63]X70.4966.4370.5566.4370.5166.4372.00
L2P [72]X83.5378.3276.3765.7868.0452.4072.92
L2P [72]✓80.2472.8980.0872.6179.4470.4159.78
DualPrompt [71]X85.1581.0181.3873.7373.4560.1668.82
DualPrompt [71]✓79.9272.8379.1573.2180.1771.7457.02
CODA-Prompt [55]X85.9181.3682.8075.2877.4366.3273.88
Linear ProbeX74.3865.4069.7358.3160.1449.7245.17
Linear Probe✓70.1061.1169.3564.1971.6470.5055.72
Generative PT modelsZero-shotX58.1653.7258.1053.7258.1353.7267.38
GMM (Ours)X83.4276.9882.4976.5181.7076.0380.72
GMM (Ours)✓84.1678.4683.9578.6484.2379.1789.41
", + "image_path": "795324d2883931c7dfc59b1422c50d7b51d83c092b2f94fc9e08d72aec9c2eaa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 361, + 547, + 396 + ], + "lines": [ + { + "bbox": [ + 46, + 361, + 547, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 547, + 396 + ], + "type": "text", + "content": "Table 1. Comparison results of our method with other conventional baselines and methods learned with discriminative pre-trained (PT) models on Tiny-ImageNet and ImageNet-R under the conventional CIL setting. \"Avg\" represents the averaged performance after training each task, and \"Last\" represents the performance on all test samples after training the last task." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 411, + 288, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 288, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 288, + 460 + ], + "type": "text", + "content": "first utilize a tokenizer to tokenize and acquire the embedding of the questions and answers. We leverage pre-trained encoder " + }, + { + "bbox": [ + 46, + 411, + 288, + 460 + ], + "type": "inline_equation", + "content": "f_{enc}" + }, + { + "bbox": [ + 46, + 411, + 288, + 460 + ], + "type": "text", + "content": " and the projection layer to obtain the corresponding features for the input images:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 124, + 476, + 287, + 490 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 476, + 287, + 490 + ], + "spans": [ + { + "bbox": [ + 124, + 476, + 287, + 490 + ], + "type": "interline_equation", + "content": "\\mathbf {e} _ {i} = f _ {\\text {e n c}} \\left(\\mathbf {x} _ {i}; \\theta_ {\\text {e n c}}\\right). \\tag {2}", + "image_path": "4d6790444e8a2997d7c5edcf7a116d06d175d9f02a56303c426360d56baa550d.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 500, + 287, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 500, + 287, + 550 + ], + "spans": [ + { + "bbox": [ + 46, + 500, + 287, + 550 + ], + "type": "text", + "content": "Then, the question embedding and the ground-truth embedding of this question, e.g., \"This is a photo of [CLS]\", is concatenated with image embedding. The final input of LLM Decoder " + }, + { + "bbox": [ + 46, + 500, + 287, + 550 + ], + "type": "inline_equation", + "content": "f_{dec}" + }, + { + "bbox": [ + 46, + 500, + 287, + 550 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 93, + 566, + 287, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 566, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 93, + 566, + 287, + 578 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {e}} _ {i} = \\operatorname {C O N C A T E} (b o s, \\mathbf {e} _ {i}, \\mathbf {q}, \\mathbf {s}, e o s). \\tag {3}", + "image_path": "52888a09a79883183e53a0f13696cf46f11c17ec148db7f7c7479fe66c280ca9.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 588, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 588, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 588, + 287, + 624 + ], + "type": "text", + "content": "bos is the symbol of the sentence beginning, and " + }, + { + "bbox": [ + 46, + 588, + 287, + 624 + ], + "type": "inline_equation", + "content": "eos" + }, + { + "bbox": [ + 46, + 588, + 287, + 624 + ], + "type": "text", + "content": " is the symbol for the end of the sentence. This encourages tokens at positions " + }, + { + "bbox": [ + 46, + 588, + 287, + 624 + ], + "type": "inline_equation", + "content": "m - 1" + }, + { + "bbox": [ + 46, + 588, + 287, + 624 + ], + "type": "text", + "content": " to predict token " + }, + { + "bbox": [ + 46, + 588, + 287, + 624 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 588, + 287, + 624 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 50, + 636, + 287, + 675 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 636, + 287, + 675 + ], + "spans": [ + { + "bbox": [ + 50, + 636, + 287, + 675 + ], + "type": "interline_equation", + "content": "P \\left(\\hat {s} _ {1}, \\hat {s} _ {2}, \\dots , \\hat {s} _ {m} \\mid \\mathbf {x} _ {i}, \\mathbf {q}, \\mathbf {s}\\right) = \\prod_ {j = 1} ^ {m - 1} P \\left(s _ {j} \\mid \\mathbf {e} _ {i}, \\mathbf {q}, s _ {1}, s _ {2}, \\dots , s _ {j - 1}\\right), \\tag {4}", + "image_path": "59a81f761d2974d084d118627cec4837645e163b47df34b5b6df9dec89726b37.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "inline_equation", + "content": "s_j" + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": " indicates the ground-truth answer token and " + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\hat{s}_m" + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": " is the generated prediction. Then, we can compute the Cross Entropy loss as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 369, + 419, + 545, + 452 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 419, + 545, + 452 + ], + "spans": [ + { + "bbox": [ + 369, + 419, + 545, + 452 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {C E}} = - \\frac {1}{m} \\sum_ {j = 1} ^ {m} s _ {j} \\cdot \\log \\hat {s} _ {j}. \\tag {5}", + "image_path": "1680e56b6c142f8364d9721f87f8d1322dc27c9e1e567faea3de4638d8548bdc.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 456, + 545, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 456, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 456, + 545, + 517 + ], + "type": "text", + "content": "Inference. During inference, we use the updated projection layer in conjunction with the pre-trained encoder to obtain image features. These image features, combined with the question embeddings are then passed to the LLM Decoder to obtain the text output." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 348, + 528, + 545, + 541 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 528, + 545, + 541 + ], + "spans": [ + { + "bbox": [ + 348, + 528, + 545, + 541 + ], + "type": "interline_equation", + "content": "\\operatorname {p r e d} = \\operatorname {a r g m a x} < f _ {\\text {t e x t}} (\\mathbf {s}), f _ {\\text {t e x t}} (\\hat {\\mathbf {s}}) >, \\tag {6}", + "image_path": "ed98d08fb666dd4610c6538012d84644c3cfa998973c18ac815ca1d08642dfdc.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 544, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 545, + 568 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 544, + 545, + 568 + ], + "type": "inline_equation", + "content": "f_{\\text{text}}" + }, + { + "bbox": [ + 304, + 544, + 545, + 568 + ], + "type": "text", + "content": " is the text encoder, " + }, + { + "bbox": [ + 304, + 544, + 545, + 568 + ], + "type": "inline_equation", + "content": "<, >" + }, + { + "bbox": [ + 304, + 544, + 545, + 568 + ], + "type": "text", + "content": " is the cosine similarity used to calculate the final predictions " + }, + { + "bbox": [ + 304, + 544, + 545, + 568 + ], + "type": "inline_equation", + "content": "pred" + }, + { + "bbox": [ + 304, + 544, + 545, + 568 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 577, + 388, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 577, + 388, + 590 + ], + "spans": [ + { + "bbox": [ + 306, + 577, + 388, + 590 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 597, + 425, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 597, + 425, + 609 + ], + "spans": [ + { + "bbox": [ + 306, + 597, + 425, + 609 + ], + "type": "text", + "content": "4.1. Experimental setups" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Datasets and Baselines. We conduct experiments in both conventional CIL and Few-shot CIL scenarios. In conventional CIL, we evaluate on three datasets. CIFAR100, TinyImageNet and ImageNet-R. CIFAR100 contains 60,000 images of " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " pixels in 100 categories. Each category has 600 images, of which 500 are for the training set, and 100 are for the test set. We experiment with two settings, B0-n and B50-n. The former splits 100 classes into " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " tasks," + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28710" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 64, + 93, + 218, + 235 + ], + "blocks": [ + { + "bbox": [ + 64, + 93, + 218, + 235 + ], + "lines": [ + { + "bbox": [ + 64, + 93, + 218, + 235 + ], + "spans": [ + { + "bbox": [ + 64, + 93, + 218, + 235 + ], + "type": "image", + "image_path": "5fc1dac2541b878b33a4ae6f17eed45101f963f3375d3362b89593ea76028575.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 218, + 93, + 373, + 236 + ], + "blocks": [ + { + "bbox": [ + 218, + 93, + 373, + 236 + ], + "lines": [ + { + "bbox": [ + 218, + 93, + 373, + 236 + ], + "spans": [ + { + "bbox": [ + 218, + 93, + 373, + 236 + ], + "type": "image", + "image_path": "68c76ab96796a0d53861c66ed96a3e434a3c01621d1bf11651bbd215f8d9b009.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 373, + 93, + 527, + 236 + ], + "blocks": [ + { + "bbox": [ + 373, + 93, + 527, + 236 + ], + "lines": [ + { + "bbox": [ + 373, + 93, + 527, + 236 + ], + "spans": [ + { + "bbox": [ + 373, + 93, + 527, + 236 + ], + "type": "image", + "image_path": "09d896fa62c4964f2a353a78e90082618608d797913106feb397f67719e05963.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 63, + 235, + 217, + 365 + ], + "blocks": [ + { + "bbox": [ + 63, + 235, + 217, + 365 + ], + "lines": [ + { + "bbox": [ + 63, + 235, + 217, + 365 + ], + "spans": [ + { + "bbox": [ + 63, + 235, + 217, + 365 + ], + "type": "image", + "image_path": "df95b220d179064318f406d516e71fd714fddeab26fb2cfbd4b2bcdcb40999e6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 52, + 369, + 539, + 380 + ], + "lines": [ + { + "bbox": [ + 52, + 369, + 539, + 380 + ], + "spans": [ + { + "bbox": [ + 52, + 369, + 539, + 380 + ], + "type": "text", + "content": "Figure 3. Comparison of our method with other SOTA baselines on CIFAR100 and Tiny-ImageNet under the conventional CIL setting." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 219, + 235, + 372, + 365 + ], + "blocks": [ + { + "bbox": [ + 219, + 235, + 372, + 365 + ], + "lines": [ + { + "bbox": [ + 219, + 235, + 372, + 365 + ], + "spans": [ + { + "bbox": [ + 219, + 235, + 372, + 365 + ], + "type": "image", + "image_path": "54928a2f405f63f3ca0218c5723973effad0d62dcd2e329ad50350010bf0ac2a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 374, + 235, + 527, + 365 + ], + "blocks": [ + { + "bbox": [ + 374, + 235, + 527, + 365 + ], + "lines": [ + { + "bbox": [ + 374, + 235, + 527, + 365 + ], + "spans": [ + { + "bbox": [ + 374, + 235, + 527, + 365 + ], + "type": "image", + "image_path": "bf7495e757d995760444f4bfcac38caef78453cbdbd5fc4c3c8c8d308c775fab.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 401, + 287, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 287, + 425 + ], + "type": "text", + "content": "while the latter trains on 50 classes first and then distributes the other 50 classes across 5/10 tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 426, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 510 + ], + "type": "text", + "content": "Tiny-ImageNet contains 200 classes out of the original 1000 classes in ImageNet, with 550 images per class, of which 500 are in the training set and 50 are in the test set. The images are down-sampled to " + }, + { + "bbox": [ + 46, + 426, + 287, + 510 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 46, + 426, + 287, + 510 + ], + "type": "text", + "content": " pixels, which makes them easier to process and analyze. We train the first task in half 100 classes and split the other 100 classes into 5/10/20 tasks following [89]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 510, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 616 + ], + "type": "text", + "content": "ImageNet-R [23] contains 200 classes of images, which are included in the original ImageNet 1000 classes. However, many images are newly added and have various styles, such as sketch, painting, misc, etc. The dataset poses a great challenge for continual learning, as it has a wide diversity of image categories and styles and an uneven distribution of samples ranging from 45 to 500 per category. We follow [71] to split the dataset into 10 tasks, each containing 20 classes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 617, + 287, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 701 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 701 + ], + "type": "text", + "content": "In Few-shot CIL, we use CIFAR100 and miniImageNet [49] following the split proposed by [62]. For both datasets, we partition the data into two parts: base session and incremental sessions. The base session comprises 60 classes with all data available, while the incremental session follows a 5-way 5-shot setting, meaning that each session consists of only 5 classes with 5 samples each." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 59, + 702, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 702, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 702, + 287, + 712 + ], + "type": "text", + "content": "In both conventional and few-shot scenarios, we com" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 401, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 401, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 545, + 521 + ], + "type": "text", + "content": "pare our method with some of the current state-of-the-art approaches, including conventional methods [7, 15, 16, 24, 29, 48, 53, 74, 76, 82, 89], pre-trained and prompt-based methods [55, 71, 72], and some specifically designed methods [12, 53, 62, 79] for few-shot scenarios. Additionally, we compare with a linear probe baseline, wherein features obtained from the image encoder are connected to a classifier for classification. We also consider the Zero-shot approach, where the generated text is directly used for classification without further fine-tuning." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 545, + 546, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 546, + 712 + ], + "type": "text", + "content": "Implementation Details. We follow BLIP2 [27] to use the EVA-CLIP [59] pre-trained ViT-g/14 and BLIP2 pretrained Qfomer. We also used the MiniGPT-4 pre-trained projection layer checkpoint as our initial parameters. Under the many-shot \"B0\" setting, we employ a learning rate of 3e-7 and use a scheduler with cosine decay. The total training process consists of 2 epochs only. In B50 or B100 settings, we first train the linear layer with a learning rate of 3e-6 on the base classes, and then on the subsequent tasks, we adopt a lower learning rate of 3e-7, both employing a cosine decay scheduler. For the few-shot setting, we use a learning rate of 3e-6 for both base task and incremental tasks. We train one epoch for the base task and two epochs for incremental tasks." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 149, + 71, + 436, + 83 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 71, + 436, + 83 + ], + "spans": [ + { + "bbox": [ + 149, + 71, + 436, + 83 + ], + "type": "text", + "content": "Continual CLIP L2P DualPrompt CODA-Prompt GMM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28711" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 73, + 69, + 523, + 301 + ], + "blocks": [ + { + "bbox": [ + 73, + 69, + 523, + 301 + ], + "lines": [ + { + "bbox": [ + 73, + 69, + 523, + 301 + ], + "spans": [ + { + "bbox": [ + 73, + 69, + 523, + 301 + ], + "type": "table", + "html": "
012345678PD↓
iCaRL [48]61.3146.3242.9437.6330.4924.0020.8918.8017.2144.10
EEIL [7]61.3146.5844.0037.2933.1427.1224.1021.5719.5841.73
LUCIR [24]61.3147.8039.3131.9125.6821.3518.6717.2414.1747.14
TOPIC [62]61.3150.0945.1741.1637.4835.5232.1929.4624.4236.89
CEC [79]72.0066.8362.9759.4356.7053.7351.1949.2447.6324.37
F2M [53]72.0567.4763.1659.7056.7153.7751.1149.2147.8424.21
MetaFSCIL [12]72.0467.9463.7760.2957.5855.1652.9050.7949.1922.85
Entropy-reg [31]71.8467.1263.2159.7757.0153.9551.5549.5248.2123.63
L2P* [72]94.1287.2080.9975.6770.9466.7663.1159.8156.8337.29
DualPrompt* [71]93.9786.8580.6775.3170.6166.4462.7759.5856.8037.17
CODA-Prompt* [55]95.3788.8682.6977.8774.4770.1666.4663.7361.1434.23
Zero-shot58.0858.9557.7657.8958.1957.4256.2654.8254.953.13
GMM (Ours)89.3588.4086.1185.0783.6181.3578.9777.3475.1814.17
", + "image_path": "5f239806a72dca730dddc92934ec918d1ed8fd2713199e779dc667914d97d49a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 308, + 547, + 354 + ], + "lines": [ + { + "bbox": [ + 46, + 308, + 547, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 547, + 354 + ], + "type": "text", + "content": "Table 2. Comparison results of our method with other SOTA baselines on mini-ImageNet under the few-shot CIL setting. Phase 0 is the base task with all samples available, and the following 1-8 phases are the incremental 5-way 5-shot tasks. The reported accuracy is the test result on all seen classes after each session of training. \"PD\" represents Performance Prop between session 0 and session 8, with lower values indicating lower forgetting. * indicates our re-implementation based on PILOT [58]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 373, + 231, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 373, + 231, + 387 + ], + "spans": [ + { + "bbox": [ + 47, + 373, + 231, + 387 + ], + "type": "text", + "content": "4.2. Experiments on Conventional CIL" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 397, + 289, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 397, + 289, + 638 + ], + "spans": [ + { + "bbox": [ + 46, + 397, + 289, + 638 + ], + "type": "text", + "content": "In Table 1, we can see that our method outperforms all conventional methods by a large margin, including ResNet-based method DER and ViT-based DyTox. Note that without exemplar, our performance is a bit lower than Dual-Prompt and CODA-Prompt at B100-5 setting. We argue that their performance is mainly due to the backbone pretrained on ImageNet-21K, which largely overlaps with CIFAR100 and Tiny-ImageNet. Another interesting observation is that our method has better performance than all baselines under longer sequence settings (B100-10, B100-20). We believe this is because generative models do not rely on classification heads, making them less prone to bias toward the current task, resulting in less forgetting of past tasks. The Linear probe setting performs less than our method, indicating that our main contribution is not from the Large pre-trained ViT but the generation pipeline. In addition, the Zero-shot performance is superior to many traditional baselines, meaning that the Generative Multi-modal Models are indeed efficient Class-Incremental learners, but its output is less concise without fine-tuning (see Fig. 4)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": "In Fig. 3, we compared our approach with some pretrained models on CIFAR100 and Tiny-ImageNet in terms of last task accuracy (all baselines are based on PILOT [58] and use 2000 exemplars). It can be observed that our method does not outperform other approaches in the initial tasks (0-2) and short sequence settings (B0-5, B100-5)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 374, + 547, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 374, + 547, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 374, + 547, + 460 + ], + "type": "text", + "content": "This is because we do not rely on a supervised ImageNet-21K pre-trained backbone. Besides, we trained each task for only 1-2 epochs to ensure efficiency without sacrificing generalization. However, our method exhibits significant advantages in long sequences and later tasks. For instance, under the CIFAR100 B0-20 setting, we outperform CODA-Prompt by 10 points and DualPrompt by 7 points." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 467, + 471, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 467, + 471, + 480 + ], + "spans": [ + { + "bbox": [ + 305, + 467, + 471, + 480 + ], + "type": "text", + "content": "4.3. Experiments on Few-shot CIL" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 485, + 547, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 547, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 547, + 641 + ], + "type": "text", + "content": "In Table 2, we compare our method with several baselines in the few-shot setting on mini-ImageNet. The evaluation metric is the model's accuracy across all the classes it has encountered so far. Our method outperforms conventional methods by a substantial margin in the final task, achieving an increase of more than " + }, + { + "bbox": [ + 304, + 485, + 547, + 641 + ], + "type": "inline_equation", + "content": "26\\%" + }, + { + "bbox": [ + 304, + 485, + 547, + 641 + ], + "type": "text", + "content": ". Furthermore, we surpass the best discriminative pre-trained approach CODA-Prompt by more than " + }, + { + "bbox": [ + 304, + 485, + 547, + 641 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 304, + 485, + 547, + 641 + ], + "type": "text", + "content": " points. It's important to note that our accuracy in the first task (89.35) might not be as high as CODA-Prompt (95.37). However, in subsequent sessions, we consistently perform better than CODA-Prompt due to our ability to learn new tasks and retain knowledge of old tasks simultaneously." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "content": "In Table 3, our method outperforms all other baselines on the CIFAR100 dataset of the few-shot setting, achieving a remarkably lower Performance Drop (PD) of 10.06. Furthermore, the Zero-shot baseline can achieve a very low PD due to the absence of forgetting. However, its overall performance is not very satisfying, as the length and con" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28712" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 71, + 187, + 156 + ], + "blocks": [ + { + "bbox": [ + 62, + 71, + 187, + 156 + ], + "lines": [ + { + "bbox": [ + 62, + 71, + 187, + 156 + ], + "spans": [ + { + "bbox": [ + 62, + 71, + 187, + 156 + ], + "type": "image", + "image_path": "9eb814575495c5472761290ad93bc8cfc615a60b8c99f2b6b6049b8f77342136.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 190, + 75, + 296, + 114 + ], + "lines": [ + { + "bbox": [ + 190, + 75, + 296, + 114 + ], + "spans": [ + { + "bbox": [ + 190, + 75, + 296, + 114 + ], + "type": "text", + "content": "This is a photo of a statue of a bird with a long beak and a long neck. It is made of stone and has a blue sky background." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 191, + 140, + 264, + 153 + ], + "lines": [ + { + "bbox": [ + 191, + 140, + 264, + 153 + ], + "spans": [ + { + "bbox": [ + 191, + 140, + 264, + 153 + ], + "type": "text", + "content": "This is a pelican." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 62, + 158, + 187, + 243 + ], + "blocks": [ + { + "bbox": [ + 62, + 158, + 187, + 243 + ], + "lines": [ + { + "bbox": [ + 62, + 158, + 187, + 243 + ], + "spans": [ + { + "bbox": [ + 62, + 158, + 187, + 243 + ], + "type": "image", + "image_path": "0326910f8b5be09208084d044d3a89818adb9175b3fb5d387183603fb7634fe0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 190, + 162, + 293, + 221 + ], + "lines": [ + { + "bbox": [ + 190, + 162, + 293, + 221 + ], + "spans": [ + { + "bbox": [ + 190, + 162, + 293, + 221 + ], + "type": "text", + "content": "This is a painting of a fire truck. The fire truck is red and has a ladder on the side. The ladder is painted in white. The fire truck has a red and white stripe on the side." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 192, + 228, + 279, + 240 + ], + "lines": [ + { + "bbox": [ + 192, + 228, + 279, + 240 + ], + "spans": [ + { + "bbox": [ + 192, + 228, + 279, + 240 + ], + "type": "text", + "content": "This is a fire engine." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 46, + 253, + 547, + 286 + ], + "lines": [ + { + "bbox": [ + 46, + 253, + 547, + 286 + ], + "spans": [ + { + "bbox": [ + 46, + 253, + 547, + 286 + ], + "type": "text", + "content": "Figure 4. Visual comparison examples of our method against frozen MiniGPT-4 [88] (Zero-shot). Text with a gray background is generated by MiniGPT-4 based on the image, while text with a orange background represents our method's output. The ground-truth labels are displayed on the left side of each row of images. All images displayed here are random sampled from ImageNet-R." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 306, + 72, + 417, + 155 + ], + "blocks": [ + { + "bbox": [ + 306, + 72, + 417, + 155 + ], + "lines": [ + { + "bbox": [ + 306, + 72, + 417, + 155 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 417, + 155 + ], + "type": "image", + "image_path": "14fd87dd31df3cc82c57882b4e2ce0b2f5b8a5823f7f3cf9c39fcba719bc9230.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 421, + 75, + 523, + 125 + ], + "lines": [ + { + "bbox": [ + 421, + 75, + 523, + 125 + ], + "spans": [ + { + "bbox": [ + 421, + 75, + 523, + 125 + ], + "type": "text", + "content": "This is a sculpture of a bird with its wings spread wide, as if it is flying. The bird is made of bronze and has a green patina. It is located in a park." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 422, + 140, + 494, + 153 + ], + "lines": [ + { + "bbox": [ + 422, + 140, + 494, + 153 + ], + "spans": [ + { + "bbox": [ + 422, + 140, + 494, + 153 + ], + "type": "text", + "content": "This is a pelican." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 158, + 418, + 242 + ], + "blocks": [ + { + "bbox": [ + 307, + 158, + 418, + 242 + ], + "lines": [ + { + "bbox": [ + 307, + 158, + 418, + 242 + ], + "spans": [ + { + "bbox": [ + 307, + 158, + 418, + 242 + ], + "type": "image", + "image_path": "37cebaae52d016cdf65733a87f2a55a88532afcd0422f8209dfcf44daef9238c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 421, + 162, + 528, + 220 + ], + "lines": [ + { + "bbox": [ + 421, + 162, + 528, + 220 + ], + "spans": [ + { + "bbox": [ + 421, + 162, + 528, + 220 + ], + "type": "text", + "content": "This is a toy car with a duck in the driver's seat. The car is red and has a fire truck on the back. The duck is wearing a firefighter's helmet and has a fire hose in its hand." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 423, + 228, + 504, + 239 + ], + "lines": [ + { + "bbox": [ + 423, + 228, + 504, + 239 + ], + "spans": [ + { + "bbox": [ + 423, + 228, + 504, + 239 + ], + "type": "text", + "content": "This is a fire truck." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 52, + 296, + 282, + 518 + ], + "blocks": [ + { + "bbox": [ + 52, + 296, + 282, + 518 + ], + "lines": [ + { + "bbox": [ + 52, + 296, + 282, + 518 + ], + "spans": [ + { + "bbox": [ + 52, + 296, + 282, + 518 + ], + "type": "table", + "html": "
Method048PD↓
iCaRL [48]64.1027.9313.7350.37
EEIL [7]64.1028.9615.8548.25
LUCIR [24]64.1031.6113.5450.56
TOPIC [62]64.1040.1129.3734.73
CEC [79]73.0758.0949.1423.93
F2M [53]71.4557.7649.3522.06
MetaFSCIL [12]74.5059.4849.9724.53
Entropy-reg [31]74.4059.7150.1424.26
L2P* [72]91.2268.6654.8936.33
DualPrompt* [71]91.0868.4554.6736.41
CODA-Prompt* [55]93.5571.9159.3234.23
Zero-shot74.1372.5967.936.20
GMM (Ours)91.5385.6581.4710.06
", + "image_path": "523d670f10c74d96c889250453c0c41b1d9cbef9027b9c3729c8e23cf422ef57.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 526, + 287, + 549 + ], + "lines": [ + { + "bbox": [ + 47, + 526, + 287, + 549 + ], + "spans": [ + { + "bbox": [ + 47, + 526, + 287, + 549 + ], + "type": "text", + "content": "Table 3. Comparison results of our method with other SOTA baselines on CIFAR100 under the few-shot CIL setting." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 555, + 287, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 555, + 287, + 580 + ], + "spans": [ + { + "bbox": [ + 46, + 555, + 287, + 580 + ], + "type": "text", + "content": "tent of its output are inconsistent and unpredictable without fine-tuning." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 586, + 135, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 135, + 598 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 135, + 598 + ], + "type": "text", + "content": "4.4. Visualizations" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "text", + "content": "In Fig. 4, we present some comparison examples of our methods against GMM without fine-tuning [88]. We can see that GMM without fine-tuning provides an intuitive description of the overall image content with varying lengths of output text. However, it tends to recognize only broad categories (e.g., biar, car) and struggles with fine-grained categorization (e.g., pelican, truck). The descriptions are sometimes somewhat repetitive (e.g., first fire engine). In contrast, our fine-tuned method accurately identifies the im" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 298, + 545, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 298, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 304, + 298, + 545, + 358 + ], + "type": "text", + "content": "age's real category, even if there are occasional discrepancies with the true labels (e.g., \"fire engine\" vs. \"fire truck\"). Besides, with the assistance of the text encoder during the testing phase, our model can achieve correct classification results even when predicting similar but not identical text." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 376, + 378, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 376, + 378, + 388 + ], + "spans": [ + { + "bbox": [ + 306, + 376, + 378, + 388 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 398, + 545, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 398, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 398, + 545, + 495 + ], + "type": "text", + "content": "In this paper, we propose GMM to use generative models for class-incremental learning. By fine-tuning the Generative Multi-modal Model (GMM), we directly generate the label text of the images to be classified. Then we select the label most similar to the generated text by its features. Our experiments demonstrate that this method, which does not require a classification head, is highly effective in addressing classification biases in continual learning." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 499, + 545, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 499, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 304, + 499, + 545, + 559 + ], + "type": "text", + "content": "Limitations. Since we are the first to introduce generative models to class-incremental learning, the overall design of our method is embarrassingly simple. We believe that with more focused efforts in this direction, there will be significant advancements in the field of continual learning." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 564, + 545, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 564, + 545, + 636 + ], + "spans": [ + { + "bbox": [ + 304, + 564, + 545, + 636 + ], + "type": "text", + "content": "Broader impact. We believe that introducing GMM into continual learning (CL) is both necessary and urgent. With the rapid development of GMM, we can leverage their capabilities to improve the performance of continual learning. Besides, integrating CL methods into the training process of GMM could significantly reduce training costs." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgments. This work is funded by NSFC (NO. 62206135, 62225604), Young Elite Scientists Sponsorship Program by CAST (2023QNRC001), and the Fundamental Research Funds for the Central Universities (Nankai University, 070-63233085). Computation is supported by the Supercomputing Center of Nankai University." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28713" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "text", + "content": "[1] Hongjoon Ahn, Jihwan Kwak, Subin Lim, Hyeonsu Bang, Hyojun Kim, and Taesup Moon. Ss-il: Separated softmax for incremental learning. In ICCV, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 288, + 190 + ], + "type": "text", + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 288, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 288, + 237 + ], + "type": "text", + "content": "[3] Rahaf Aljundi, Punarjay Chakravarty, and Tinne Tuytelaars. Expert gate: Lifelong learning with a network of experts. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3366-3375, 2017. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 239, + 288, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 288, + 294 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 288, + 294 + ], + "type": "text", + "content": "[4] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2(3):4, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 295, + 288, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 295, + 288, + 350 + ], + "spans": [ + { + "bbox": [ + 53, + 295, + 288, + 350 + ], + "type": "text", + "content": "[5] Hedi Ben-Younes, Rémi Cadene, Matthieu Cord, and Nicolas Thome. Mutan: Multimodal tucker fusion for visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 2612-2620, 2017. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 352, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 352, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 352, + 287, + 407 + ], + "type": "text", + "content": "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 408, + 287, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 408, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 53, + 408, + 287, + 462 + ], + "type": "text", + "content": "[7] Francisco M Castro, Manuel J Marín-Jiménez, Nicolas Guil, Cordelia Schmid, and Karteek Alahari. End-to-end incremental learning. In Proceedings of the European conference on computer vision (ECCV), pages 233-248, 2018. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 464, + 288, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 464, + 288, + 520 + ], + "spans": [ + { + "bbox": [ + 53, + 464, + 288, + 520 + ], + "type": "text", + "content": "[8] Jun Chen, Han Guo, Kai Yi, Boyang Li, and Mohamed Elhoseiny. Visualgpt: Data-efficient adaptation of pretrained language models for image captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18030-18040, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 521, + 287, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 521, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 53, + 521, + 287, + 554 + ], + "type": "text", + "content": "[9] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 555, + 287, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 589 + ], + "type": "text", + "content": "[10] Xinlei Chen*, Saining Xie*, and Kaiming He. An empirical study of training self-supervised vision transformers. arXiv preprint arXiv:2104.02057, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 590, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 657 + ], + "type": "text", + "content": "[11] Ali Cheraghian, Shafin Rahman, Pengfei Fang, Soumava Kumar Roy, Lars Petersson, and Mehrtash Harandi. Semantic-aware knowledge distillation for few-shot class-incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2534-2543, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "text", + "content": "[12] Zhixiang Chi, Li Gu, Huan Liu, Yang Wang, Yuanhao Yu, and Jin Tang. Meta-scil: A meta-learning approach for few-shot class incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14166-14175, 2022. 6, 7, 8" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 715 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 137 + ], + "type": "text", + "content": "[13] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 307, + 73, + 545, + 137 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 307, + 73, + 545, + 137 + ], + "type": "text", + "content": " chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 139, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 139, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 139, + 545, + 183 + ], + "type": "text", + "content": "[14] Matthias Delange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Ales Leonardis, Greg Slabaugh, and Tinne Tuytelaars. A continual learning survey: Defying forgetting in classification tasks. TPAMI, 2021. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 184, + 545, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 545, + 225 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 545, + 225 + ], + "type": "text", + "content": "[15] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In ECCV, 2020. 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 228, + 545, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 260 + ], + "type": "text", + "content": "[16] Arthur Douillard, Alexandre Ramé, Guillaume Couairon, and Matthieu Cord. Dytox: Transformers for continual learning with dynamic token expansion. In CVPR, 2022. 2, 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 261, + 545, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 314 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 314 + ], + "type": "text", + "content": "[17] Chrisantha Fernando, Dylan Banarse, Charles Blundell, Yori Zwols, David Ha, Andrei A Rusu, Alexander Pritzel, and Daan Wierstra. Pathnet: Evolution channels gradient descent in super neural networks. arXiv preprint arXiv:1701.08734, 2017. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 316, + 545, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 338 + ], + "type": "text", + "content": "[18] Rui Gao and Weiwei Liu. Ddgr: Continual learning with deep diffusion-based generative replay. In ICML, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 338, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 545, + 370 + ], + "type": "text", + "content": "[19] Meng-Hao Guo, Cheng-Ze Lu, Zheng-Ning Liu, Ming-Ming Cheng, and Shi-Min Hu. Visual attention network. Computational Visual Media, 9(4):733-752, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 372, + 545, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 372, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 308, + 372, + 545, + 404 + ], + "type": "text", + "content": "[20] Tyler L Hayes, Kushal Kafle, Robik Shrestha, Manoj Acharya, and Christopher Kanan. Remind your neural network to prevent catastrophic forgetting. In ECCV, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 405, + 545, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 405, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 308, + 405, + 545, + 449 + ], + "type": "text", + "content": "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 449, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 502 + ], + "type": "text", + "content": "[22] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 504, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 545, + 558 + ], + "type": "text", + "content": "[23] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, Dawn Song, Jacob Steinhardt, and Justin Gilmer. The many faces of robustness: A critical analysis of out-of-distribution generalization. ICCV, 2021. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 559, + 545, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 592 + ], + "type": "text", + "content": "[24] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In CVPR, 2019. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 593, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 593, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 593, + 545, + 624 + ], + "type": "text", + "content": "[25] Xinting Hu, Kaihua Tang, Chunyan Miao, Xian-Sheng Hua, and Hanwang Zhang. Distilling causal effect of data in class incremental learning. In CVPR, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 625, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 690 + ], + "type": "text", + "content": "[26] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. 2, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 691, + 545, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 715 + ], + "type": "text", + "content": "[27] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28714" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 3, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 288, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 288, + 151 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 288, + 151 + ], + "type": "text", + "content": "[28] Wenbo Li, Pengchuan Zhang, Lei Zhang, Qiuyuan Huang, Xiaodong He, Siwei Lyu, and Jianfeng Gao. Object-driven text-to-image synthesis via adversarial training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12174-12182, 2019. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 184 + ], + "type": "text", + "content": "[29] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 2, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 287, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 287, + 206 + ], + "type": "text", + "content": "[30] Zhizhong Li and Derek Hoiem. Learning without forgetting. TPAMI, 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 209, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 262 + ], + "type": "text", + "content": "[31] Huan Liu, Li Gu, Zhixiang Chi, Yang Wang, Yuanhao Yu, Jun Chen, and Jin Tang. Few-shot class-incremental learning via entropy-regularized data-free replay. In European Conference on Computer Vision, pages 146–162. Springer, 2022. 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 264, + 287, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 286 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 286 + ], + "type": "text", + "content": "[32] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 288, + 287, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 288, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 48, + 288, + 287, + 331 + ], + "type": "text", + "content": "[33] Ming Liu, Yuxiang Wei, Xiaohe Wu, Wangmeng Zuo, and Lei Zhang. Survey on leveraging pre-trained generative adversarial networks for image editing and restoration. Science China Information Sciences, 66(5):151101, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 333, + 287, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 333, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 48, + 333, + 287, + 365 + ], + "type": "text", + "content": "[34] Wei Liu, Sihan Chen, Longteng Guo, Xinxin Zhu, and Jing Liu. Cptr: Full transformer network for image captioning. arXiv preprint arXiv:2101.10804, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 366, + 287, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 366, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 48, + 366, + 287, + 409 + ], + "type": "text", + "content": "[35] Xialei Liu, Marc Masana, Luis Herranz, Joost Van de Weijer, Antonio M Lopez, and Andrew D Bagdanov. Rotate your networks: Better weight consolidation and less catastrophic forgetting. In ICPR, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 411, + 287, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 287, + 455 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 287, + 455 + ], + "type": "text", + "content": "[36] Xialei Liu, Yu-Song Hu, Xu-Sheng Cao, Andrew D Bagdanov, Ke Li, and Ming-Ming Cheng. Long-tailed class incremental learning. In European Conference on Computer Vision, pages 495-512. Springer, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 456, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 498 + ], + "type": "text", + "content": "[37] Xialei Liu, Xusheng Cao, Haori Lu, Jia-wen Xiao, Andrew D Bagdanov, and Ming-Ming Cheng. Class incremental learning with pre-trained vision-language models. arXiv preprint arXiv:2310.20348, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 501, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 287, + 533 + ], + "type": "text", + "content": "[38] Yaoyao Liu, Bernt Schiele, and Qianru Sun. Adaptive aggregation networks for class-incremental learning. In CVPR, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 535, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 535, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 535, + 287, + 578 + ], + "type": "text", + "content": "[39] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 580, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 287, + 612 + ], + "type": "text", + "content": "[40] Arun Mallya, Dillon Davis, and Svetlana Lazebnik. Piggyback: Adapting a single network to multiple tasks by learning to mask weights. In ECCV, 2018. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 614, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 677 + ], + "type": "text", + "content": "[41] Marc Masana, Xialei Liu, Bartlomiej Twardowski, Mikel Menta, Andrew D Bagdanov, and Joost Van De Weijer. Class-incremental learning: survey and performance evaluation on image classification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5):5513-5533, 2022. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[42] Pratik Mazumder, Pravendra Singh, and Piyush Rai. Few-shot lifelong learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2337-2345, 2021. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[43] Michael McCloskey and Neal J Cohen. Catastrophic interference in connectionist networks: The sequential learning problem. In *Psychology of learning and motivation*, pages 109–165. Elsevier, 1989. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 118, + 484, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 484, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 484, + 128 + ], + "type": "text", + "content": "[44] OpenAI. Gpt-4 technical report, 2023. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 129, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 171 + ], + "type": "text", + "content": "[45] Haoxuan Qu, Hossein Rahmani, Li Xu, Bryan Williams, and Jun Liu. Recent advances of continual learning in computer vision: An overview. arXiv preprint arXiv:2109.11369, 2021. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 173, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 545, + 227 + ], + "type": "text", + "content": "[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 228, + 545, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 545, + 260 + ], + "type": "text", + "content": "[47] Jathushan Rajasegaran, Munawar Hayat, Salman Khan, Fahad Shahbaz Khan, and Ling Shao. Random path selection for incremental learning. In NIPS, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 262, + 545, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 262, + 545, + 303 + ], + "spans": [ + { + "bbox": [ + 307, + 262, + 545, + 303 + ], + "type": "text", + "content": "[48] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In CVPR, 2017. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 305, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 305, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 307, + 305, + 545, + 360 + ], + "type": "text", + "content": "[49] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 361, + 545, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 361, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 307, + 361, + 545, + 404 + ], + "type": "text", + "content": "[50] Jonathan Schwarz, Wojciech Czarnecki, Jelena Luketina, Agnieszka Grabska-Barwinska, Yee Whye Teh, Razvan Pascanu, and Raia Hadsell. Progress & compress: A scalable framework for continual learning. In ICML, 2018. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 405, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 405, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 405, + 545, + 437 + ], + "type": "text", + "content": "[51] Joan Serra, Didac Suris, Marius Miron, and Alexandros Karatzoglou. Overcoming catastrophic forgetting with hard attention to the task. In ICML, 2018. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 438, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 438, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 307, + 438, + 545, + 503 + ], + "type": "text", + "content": "[52] Yijia Shao, Yiduo Guo, Dongyan Zhao, and Bing Liu. Class incremental learning based on label generation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 1263-1276, Toronto, Canada, 2023. Association for Computational Linguistics. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 504, + 545, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 504, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 307, + 504, + 545, + 557 + ], + "type": "text", + "content": "[53] Guangyuan Shi, Jiaxin Chen, Wenlong Zhang, Li-Ming Zhan, and Xiao-Ming Wu. Overcoming catastrophic forgetting in incremental few-shot learning by finding flat minima. Advances in neural information processing systems, 34: 6747-6761, 2021. 3, 6, 7, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 559, + 545, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 559, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 307, + 559, + 545, + 592 + ], + "type": "text", + "content": "[54] Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual learning with deep generative replay. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 593, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 593, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 593, + 545, + 668 + ], + "type": "text", + "content": "[55] James Seale Smith, Leonid Karlinsky, Vyshnavi Gutta, Paola Cascante-Bonilla, Donghyun Kim, Assaf Arbelle, Rameswar Panda, Rogerio Feris, and Zsolt Kira. Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11909-11919, 2023. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "type": "text", + "content": "[56] Boyuan Sun, Yuqi Yang, Le Zhang, Ming-Ming Cheng, and Qibin Hou. Corrmatch: Label propagation via correlation matching for semi-supervised semantic segmentation. IEEE Computer Vision and Pattern Recognition (CVPR), 2024. 1" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28715" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "text", + "content": "[57] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7464-7473, 2019. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 129, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 129, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 49, + 129, + 287, + 161 + ], + "type": "text", + "content": "[58] Hai-Long Sun, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Pilot: A pre-trained model-based continual learning toolbox. arXiv preprint arXiv:2309.07117, 2023. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 287, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 287, + 195 + ], + "type": "text", + "content": "[59] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389, 2023. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 198, + 287, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 198, + 287, + 230 + ], + "spans": [ + { + "bbox": [ + 49, + 198, + 287, + 230 + ], + "type": "text", + "content": "[60] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 232, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 232, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 49, + 232, + 287, + 262 + ], + "type": "text", + "content": "[61] Xiaoyu Tao, Xinyuan Chang, Xiaopeng Hong, Xing Wei, and Yihong Gong. Topology-preserving class-incremental learning. In ECCV, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 265, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 265, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 49, + 265, + 287, + 319 + ], + "type": "text", + "content": "[62] Xiaoyu Tao, Xiaopeng Hong, Xinyuan Chang, Songlin Dong, Xing Wei, and Yihong Gong. Few-shot class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12183-12192, 2020. 2, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 321, + 287, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 321, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 49, + 321, + 287, + 354 + ], + "type": "text", + "content": "[63] Vishal Thengane, Salman Khan, Munawar Hayat, and Fahad Khan. Clip model is an efficient continual learner. arXiv preprint arXiv:2210.03114, 2022. 3, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 355, + 287, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 355, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 49, + 355, + 287, + 409 + ], + "type": "text", + "content": "[64] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 411, + 287, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 411, + 287, + 454 + ], + "spans": [ + { + "bbox": [ + 49, + 411, + 287, + 454 + ], + "type": "text", + "content": "[65] Maria Tsimpoukelli, Jacob L Menick, Serkan Cabi, SM Eslami, Oriol Vinyls, and Felix Hill. Multimodal few-shot learning with frozen language models. Advances in Neural Information Processing Systems, 34:200-212, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 456, + 287, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 456, + 287, + 478 + ], + "spans": [ + { + "bbox": [ + 49, + 456, + 287, + 478 + ], + "type": "text", + "content": "[66] Gido M Van de Ven and Andreas S Tolias. Three scenarios for continual learning. NIPS Workshops, 2019. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 479, + 287, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 479, + 287, + 511 + ], + "spans": [ + { + "bbox": [ + 49, + 479, + 287, + 511 + ], + "type": "text", + "content": "[67] Liuyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. A comprehensive survey of continual learning: Theory, method and application. arXiv preprint arXiv:2302.00487, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 513, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 513, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 49, + 513, + 287, + 567 + ], + "type": "text", + "content": "[68] Liyuan Wang, Jingyi Xie, Xingxing Zhang, Mingyi Huang, Hang Su, and Jun Zhu. Hierarchical decomposition of prompt-based continual learning: Rethinking obscured suboptimality. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 569, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 569, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 49, + 569, + 287, + 622 + ], + "type": "text", + "content": "[69] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, et al. Vision: Large language model is also an open-ended decoder for vision-centric tasks. arXiv preprint arXiv:2305.11175, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 624, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 624, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 49, + 624, + 287, + 678 + ], + "type": "text", + "content": "[70] Xiao Wang, Guangyao Chen, Guangwu Qian, Pengcheng Gao, Xiao-Yong Wei, Yaowei Wang, Yonghong Tian, and Wen Gao. Large-scale multi-modal pre-trained models: A comprehensive survey. Machine Intelligence Research, 20 (4):447-482, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 681, + 287, + 713 + ], + "type": "text", + "content": "[71] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "text", + "content": "prompting for rehearsal-free continual learning. In ECCV, 2022. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "type": "text", + "content": "[72] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer G. Dy, and Tomas Pfister. Learning to prompt for continual learning. CVPR, 2022. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 141, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 545, + 205 + ], + "type": "text", + "content": "[73] Tz-Ying Wu, Gurumurthy Swaminathan, Zhizhong Li, Avinash Ravichandran, Nuno Vasconcelos, Rahul Bhotika, and Stefano Soatto. Class-incremental learning with strong pre-trained models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9601-9610, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 208, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 208, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 308, + 208, + 545, + 240 + ], + "type": "text", + "content": "[74] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In CVPR, 2019. 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 241, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 241, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 241, + 545, + 274 + ], + "type": "text", + "content": "[75] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In ICCV, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 276, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 545, + 308 + ], + "type": "text", + "content": "[76] Shipeng Yan, Jiangwei Xie, and Xuming He. Der: Dynamically expandable representation for class incremental learning. In CVPR, 2021. 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 310, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 310, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 308, + 310, + 545, + 354 + ], + "type": "text", + "content": "[77] Yang Yang, Zhiying Cui, Junjie Xu, Changhong Zhong, Wei-Shi Zheng, and Ruixuan Wang. Continual learning with bayesian model based on a fixed pre-trained feature extractor. Visual Intelligence, 1(1):5, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 355, + 545, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 355, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 308, + 355, + 545, + 418 + ], + "type": "text", + "content": "[78] Lu Yu, Bartlomiej Twardowski, Xialei Liu, Luis Herranz, Kai Wang, Yongmei Cheng, Shangling Jui, and Joost van de Weijer. Semantic drift compensation for class-incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6982-6991, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 422, + 545, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 545, + 476 + ], + "type": "text", + "content": "[79] Chi Zhang, Nan Song, Guosheng Lin, Yun Zheng, Pan Pan, and Yinghui Xu. Few-shot incremental learning with continually evolved classifiers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12455-12464, 2021. 3, 6, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 478, + 545, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 478, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 308, + 478, + 545, + 532 + ], + "type": "text", + "content": "[80] Chang-Bin Zhang, Jia-Wen Xiao, Xialei Liu, Ying-Cong Chen, and Ming-Ming Cheng. Representation compensation networks for continual semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7053-7064, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 533, + 545, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 533, + 545, + 577 + ], + "spans": [ + { + "bbox": [ + 308, + 533, + 545, + 577 + ], + "type": "text", + "content": "[81] Gengwei Zhang, Liyuan Wang, Guoliang Kang, Ling Chen, and Yunchao Wei. Slca: Slow learner with classifier alignment for continual learning on a pre-trained model. arXiv preprint arXiv:2303.05118, 2023. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 578, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 578, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 308, + 578, + 545, + 611 + ], + "type": "text", + "content": "[82] Bowen Zhao, Xi Xiao, Guojun Gan, Bin Zhang, and Shu-Tao Xia. Maintaining discrimination and fairness in class incremental learning. In CVPR, 2020. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 613, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 656 + ], + "type": "text", + "content": "[83] Hanbin Zhao, Yongjian Fu, Mintong Kang, Qi Tian, Fei Wu, and Xi Li. Mgmtf: Multi-grained slow vs. fast framework for few-shot class-incremental learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "type": "text", + "content": "[84] Da-Wei Zhou, Qi-Wei Wang, Zhi-Hong Qi, Han-Jia Ye, De-Chuan Zhan, and Ziwei Liu. Deep class-incremental learning: A survey. arXiv preprint arXiv:2302.03648, 2023. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[85] Da-Wei Zhou, Han-Jia Ye, De-Chuan Zhan, and Ziwei Liu. Revisiting class-incremental learning with pre-trained mod" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "28716" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 340 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "type": "text", + "content": "els: Generalizability and adaptivity are all you need, 2023. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "type": "text", + "content": "[86] Da-Wei Zhou, Yuanhan Zhang, Jingyi Ning, Han-Jia Ye, De-Chuan Zhan, and Ziwei Liu. Learning without forgetting for vision-language models. arXiv preprint arXiv:2305.19270, 2023. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "type": "text", + "content": "[87] Qiang Zhou, Chaohui Yu, Shaofeng Zhang, Sitong Wu, Zhibing Wang, and Fan Wang. Regionblip: A unified multimodal pre-training framework for holistic and regional comprehension. arXiv preprint arXiv:2308.02299, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 287, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 287, + 229 + ], + "type": "text", + "content": "[88] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 3, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 231, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 231, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 231, + 287, + 285 + ], + "type": "text", + "content": "[89] Fei Zhu, Xu-Yao Zhang, Chuang Wang, Fei Yin, and ChengLin Liu. Prototype augmentation and self-supervision for incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5871-5880, 2021. 2, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 286, + 287, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 287, + 340 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 287, + 340 + ], + "type": "text", + "content": "[90] Kai Zhu, Yang Cao, Wei Zhai, Jie Cheng, and Zheng-Jun Zha. Self-promoted prototype refinement for few-shot class incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6801-6810, 2021. 3" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28717" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Generative Multimodal Models are In-Context Learners/1a0461d9-1d47-4e83-8d07-3f1f380c508b_content_list.json b/2024/Generative Multimodal Models are In-Context Learners/1a0461d9-1d47-4e83-8d07-3f1f380c508b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a42b98fe29029a223947c168c1719d4398f2ea56 --- /dev/null +++ b/2024/Generative Multimodal Models are In-Context Learners/1a0461d9-1d47-4e83-8d07-3f1f380c508b_content_list.json @@ -0,0 +1,2289 @@ +[ + { + "type": "text", + "text": "Generative Multimodal Models are In-Context Learners", + "text_level": 1, + "bbox": [ + 199, + 130, + 772, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Quan Sun $^{1*}$ Yufeng Cui $^{1*}$ Xiaosong Zhang $^{1*}$ Fan Zhang $^{1*}$ Qiying Yu $^{2,1*}$ Yueze Wang $^{1}$ Yongming Rao $^{1}$ Jingjing Liu $^{2}$ Tiejun Huang $^{1,3}$ Xinlong Wang $^{1\\dagger}$", + "bbox": [ + 84, + 178, + 883, + 218 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Beijing Academy of Artificial Intelligence $^{2}$ Tsinghua University $^{3}$ Peking University *equal contribution †project lead", + "bbox": [ + 161, + 223, + 816, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "code & models: https://github.com/baaivision/Emu", + "bbox": [ + 277, + 272, + 689, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 321, + 313, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The human ability to easily solve multimodal tasks in context (i.e., with only a few demonstrations or simple instructions), is what current multimodal systems have largely struggled to imitate. In this work, we demonstrate that the task-agnostic in-context learning capabilities of large multimodal models can be significantly enhanced by effective scaling-up. We introduce Emu2, a generative multimodal model with 37 billion parameters, trained on large-scale multimodal sequences with a unified autoregressive objective. Emu2 exhibits strong multimodal in-context learning abilities, even emerging to solve tasks that require on-the-fly reasoning, such as visual prompting and object-grounded generation. The model sets a new record on multiple multimodal understanding tasks in few-shot settings. When instruction-tuned to follow specific instructions, Emu2 further achieves new state-of-the-art on challenging tasks such as question answering benchmarks for large multimodal models and open-ended subject-driven generation. These achievements demonstrate that Emu2 can serve as a base model and general-purpose interface for a wide range of multimodal tasks. Code and models are publicly available to facilitate future research.", + "bbox": [ + 75, + 353, + 473, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 715, + 209, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multimodal tasks [25, 41] encompass anything involving understanding and generation in single or multiple modalities [5, 19, 58], which can be highly diverse and long-tail. Previous multimodal systems largely rely on designing task-specific architecture and collecting a sizable supervised training set, both of which are difficult to scale, particularly when this process needs to be repeated for each new task encountered. By contrast, humans can solve a new task in context, i.e., with only a few demonstrations or simple", + "bbox": [ + 75, + 739, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "instructions - a capability that current multimodal models have yet to learn.", + "bbox": [ + 500, + 323, + 890, + 353 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, generative pretrained language models have demonstrated strong in-context learning abilities [11, 21, 73]. By training a 37-billion-parameter model Emu2 and thoroughly evaluating it on diverse multimodal tasks, we demonstrate that a scaled-up multimodal generative pretrained model can harness similar in-context learning abilities and effectively generalize to unseen multimodal tasks. Emu2 is trained with a unified autoregressive objective: predict-the next-multimodal-element (either visual embeddings or textual tokens). In this unified generative pretraining process, large-scale multimodal sequences (e.g., text, image-text pairs, and interleaved image-text-video) are used for model training.", + "bbox": [ + 496, + 354, + 892, + 550 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We measure Emu2's capabilities of learning from a few examples or instructions on standard multimodal datasets, as well as new tasks unseen in the training set. Specifically, Emu2 is evaluated under two scenarios: (a) few-shot setting, where we allow as many examples as possible to fit the context window of the model; and (b) instruction tuning, where the model is tuned to follow specific instructions.", + "bbox": [ + 496, + 551, + 893, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Emu2 achieves promising results in the few-shot setting on a wide range of vision-language tasks. For example, it demonstrates state-of-the-art few-shot performance on multiple visual question-answering datasets. We observe a performance improvement when the number of examples in context increases. Figure 1 illustrates Emu2's strong multimodal reasoning capabilities for tasks in the wild, e.g., recognition and counting in a specific format. Emu2 also learns to follow visual prompting in context (e.g., the circles laid on the images in Figure 1), even although it struggles at a smaller scale or at zero shot.", + "bbox": [ + 496, + 657, + 893, + 823 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As Emu2 is inherently equipped to handle interleaved text-image-video at both input and output, it serves as a powerful and versatile base model for diverse multimodal tasks, by following specific task instructions. For example, after instruct tuning with conversational data,", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Correspondence to wangxinlong@baai.ac.cn", + "bbox": [ + 94, + 886, + 339, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14398", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4d1282fe13868ac44b6e10da23a599442d29205b9894a5f26b5835062e24bf21.jpg", + "image_caption": [ + "[dog: 1, frisbee: 1]." + ], + "image_footnote": [], + "bbox": [ + 125, + 114, + 207, + 160 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4c2ef205e12e6ea72513547bb6aa5691fd2aad11fc7cfce9ec1db0d23f8af9e6.jpg", + "image_caption": [ + "[burger: 1, glass: 1, bottle: 1]." + ], + "image_footnote": [], + "bbox": [ + 300, + 114, + 339, + 157 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e18a18357157aafcc0876e4d567730bb90498ce80ddb55e2f659af8654967ba8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 119, + 547, + 156 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bc74e1252b4635fccb7ae97a3523505ff7f2a16ca917c7d301d0cce46a8dfb95.jpg", + "image_caption": [ + "[cat: 3].", + "The text in the red circle: Do Not?", + "Completion", + "[beer: 3, banana: 2].", + "The text in the red circle: 'Lynn'." + ], + "image_footnote": [], + "bbox": [ + 642, + 114, + 720, + 204 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8f2304618c0af1da4f98ed5456c6033dde1bf907a936c8b9e8bbaadc877042d2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 205, + 720, + 247 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2ff32f6c5fc5c8c0e7d9c838edef336146ad3d5d2d452df3cd8a96eb5f848b2e.jpg", + "image_caption": [ + "The text in the red circle: 'Rights'." + ], + "image_footnote": [], + "bbox": [ + 297, + 160, + 377, + 247 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b6f719711669f26965c0c0aab109f6ff4de38cc022efa75785e1bfff82d10509.jpg", + "image_caption": [ + "The text in the red circle: 'Ave'." + ], + "image_footnote": [], + "bbox": [ + 478, + 159, + 553, + 247 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a984b813cd72fc5a86e70deb01b4d3b49ab8e913846529a250c48ce4643a60a9.jpg", + "image_caption": [ + "a photo of a yellow backpack:" + ], + "image_footnote": [], + "bbox": [ + 199, + 250, + 254, + 292 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8d8b3d9a362c1cc3ca1888b4780271d0c8095498d8b52997a2db79f753101225.jpg", + "image_caption": [ + "The subject A with a city in the background:" + ], + "image_footnote": [], + "bbox": [ + 243, + 292, + 300, + 337 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ae4d936285f56781589dc0c08f8b1f9de2abbb8c279041c27b5e4fbafaf51303.jpg", + "image_caption": [ + "a photo of a blue backpack:" + ], + "image_footnote": [], + "bbox": [ + 326, + 250, + 383, + 292 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f3f3282f01c52853b8596959d8781b36847ee25556ebf3e9c4c85bb732f55bc6.jpg", + "image_caption": [ + "The subject A in a purple wizard outfit:" + ], + "image_footnote": [], + "bbox": [ + 421, + 250, + 478, + 292 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b30403a6a631b901f5292a76ea31235533551cc8be5c0aa63f1586206c078bae.jpg", + "image_caption": [ + "a photo of a brown backpack:" + ], + "image_footnote": [], + "bbox": [ + 480, + 250, + 511, + 292 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/81e4a0bd18ca12fc80a72a2a4143e50915625f04ca02987532551a38d97e9ddc.jpg", + "image_caption": [ + "of", + "a photo of a blue and red backpack:" + ], + "image_footnote": [], + "bbox": [ + 583, + 247, + 640, + 292 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d742bdad6bcf6a4c6004af35598c0cfb1d51bc43e843b792562535df65a5b081.jpg", + "image_caption": [ + "The subject A wearing a santa hat:" + ], + "image_footnote": [], + "bbox": [ + 419, + 294, + 478, + 338 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ab5cbb65ae2ddcd90db53f9385e58ada3cdf0ff187222a07549312d913a3fd8e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 583, + 301, + 640, + 335 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a779de923f8cba57b2cfdde5bacdbe9d56b1752bd0eadf0c831a5494eb133946.jpg", + "image_caption": [ + "motorcycle's headlight." + ], + "image_footnote": [], + "bbox": [ + 764, + 248, + 825, + 292 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/14e89c157256675bb4a8e3278a36e973f488147d308e5040932e0c0808a6d2cf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 297, + 823, + 338 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4a708d896cb4fa258a03419dd1c1d491df7b0b979056ea67b760ea64eb4bbf99.jpg", + "image_caption": [ + "Look at this sequence of three shapes. What shape should come as the fourth shape? Explain your reasoning with detailed descriptions of the first shapes." + ], + "image_footnote": [], + "bbox": [ + 156, + 363, + 238, + 411 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The first shape is a triangle, the second shape is a square, and the third shape is a pentagon. To find the next shape in the sequence, we need to identify a pattern. In this case, the pattern is that each shape has one more side than the previous shape. Therefore, the next shape in the sequence should have six sides, which is a hexagon.", + "bbox": [ + 272, + 368, + 400, + 469 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/345c2ff9a7706acaced5a87e494f53efc5d3d183c9d9df39d4263a04e26b43d5.jpg", + "image_caption": [ + "What is funny about this image?" + ], + "image_footnote": [], + "bbox": [ + 439, + 366, + 521, + 414 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The funny aspect of this image is that a squirrel is holding a camera and appears to be taking a picture of a bird. This is amusing because animals, such as squirrels and birds, do not typically use cameras or engage in human-like activities like photography. The scene is set up in a way that makes it look like the squirrel is intentionally capturing the bird's image, which adds a humorous twist to the scene. Additionally, the bird is perched on a mushroom, further enhancing the whimsical and playful nature of the image.", + "bbox": [ + 557, + 362, + 687, + 479 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5e49f2038c304b38fb988de176d8fa07331dd47731cc1c8820e381bd495b3f42.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 353, + 867, + 483 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b8771eba724fe87515faa21c975df27701a76a07e0a20141161ad134e9c4e539.jpg", + "image_caption": [ + "impressionist painting of an astronaut in a jungle" + ], + "image_footnote": [], + "bbox": [ + 124, + 498, + 267, + 609 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/575e7154d1415b51921064162802d7ad53b7703bd2f0245a3910adde8cee94b5.jpg", + "image_caption": [ + "An image of wearing a big sunglasses on the beach" + ], + "image_footnote": [], + "bbox": [ + 436, + 498, + 581, + 607 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/dc5f36c519e6e38dbc6b601e0d5233a9d933c7ca79abcefde199813ee8ed810e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 308, + 507, + 387, + 601 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/56e48415d335188b4e7a492cdb006a2680e263fdcaeca29c72fd9dfe218ea6a5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 581, + 500, + 720, + 609 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8631505cc0b1e5ae9da3fbadd84a9da737df6128a09e7df04523cad4b6683db6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 722, + 498, + 867, + 609 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e6daaa233f23264b059d0d6fd6b483e222aef55470fd14ca0e668e5a4710fbbb.jpg", + "image_caption": [ + "An anime draw of " + ], + "image_footnote": [], + "bbox": [ + 279, + 676, + 348, + 729 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/471cad8f46d65acc63f695dd251e311729e43cdc21a99e756df02e1b315f1afb.jpg", + "image_caption": [ + "and $\\text{a dog}$" + ], + "image_footnote": [], + "bbox": [ + 352, + 678, + 387, + 729 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4d04223adb671d60a30154108673737b9de3adcdcd2fe98da48d3797d22e5833.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 450, + 676, + 516, + 729 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ac7d0b522d218ac19c15f874ba476b6c20751be158e2ac5048376a2498259e1e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 676, + 588, + 729 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b719d4b5086e773d384f307c4c5d3406accab4e3d9f78f03dd0989ff5c0a63d3.jpg", + "image_caption": [ + ", in the style of" + ], + "image_footnote": [], + "bbox": [ + 658, + 686, + 723, + 720 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/03e2ddada0a6b3036cd26a09b3401d0d546ea6fa19e4db0c96652fa9318b98af.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 730, + 647, + 867, + 752 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c81c5676d24211777e288e9f0df17aa44ec37ecb3bffc4db67f074b0bbe0313f.jpg", + "image_caption": [ + " and [USER]: [ASSISTANT]: ”. Here represents system message and varies between the two major task categories (academic-task-oriented and multimodal chat). The section comprises multimodal tokens, including images, videos, and text. Only tokens in the section will be supervised by cross-entropy loss during training.", + "bbox": [ + 75, + 137, + 468, + 318 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Setup. We use a global batch size of 768 and train for 8k steps. The learning rate linearly warms up to $1 \\times 10^{-5}$ in the first 100 steps, then decays to zero with a cosine schedule. The model is trained using the AdamW optimizer with $\\beta_{1} = 0.9$ , $\\beta_{2} = 0.98$ , $\\epsilon = 1 \\times 10^{-6}$ , and a gradient clipping of 5.0. The sequence length during training is limited to 2048, and any excess beyond that is truncated directly. We consistently employed an input image/video resolution of $448 \\times 448$ . For video data, we uniformly sample frames in time as input to the model. The number of sampled frames for each video is randomly chosen from 8, 12, and 16. To capture more intricate spatial details, following the visual encoder stage, we apply mean-pooling to each static image, dividing it into $16 \\times 16$ tokens during instruction fine-tuning. This differs from the pre-training phase, where $8 \\times 8$ tokens were utilized.", + "bbox": [ + 75, + 319, + 473, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2.3.2 Controllable Visual Generation", + "text_level": 1, + "bbox": [ + 76, + 585, + 351, + 602 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Data. We leverage a mix of high-quality datasets to unleash the potential of controllable generation in context. We use a grounded image-text pair dataset CapsFusion-grounded-100M and GRIT [57] for grounded text-to-image generation. To mitigate the impact of image backgrounds on the effectiveness of multi-entity subject-driven generation, we employ SAM [36] to preprocess the grounding data, yielding a subset of approximately 5 million samples with segmentation results. Additionally, we leverage InstructPix2Pix constructed by [10] for image editing tasks. For the text-to-image task, we use a filtered subset of CapsFusion [87], LAION-Aesthetics [1], SA-1B [36], and LAION-High-Resolution [3].", + "bbox": [ + 75, + 612, + 468, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We also collect data from premium sources (e.g., Unsplash [20]) and outputs from advanced text-to-image systems (e.g., Midjourney-V5 [54] and DALL-E-3 [9]) for quality fine-tuning. This diverse dataset includes around 500k high-quality image-text pairs. For all the data above, during the training, only samples with image resolutions", + "bbox": [ + 75, + 810, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "higher than $448 \\times 448$ were retained to ensure generation quality. More details can be found in the supplementary.", + "bbox": [ + 498, + 90, + 890, + 122 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/5ddefc160cf822c8872b36f202ac4897bf5aa162984612ce85ecae06a175c411.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelShotVQAv2OKVQAVizWizTextVQAHateful Memes
Kosmos-1 (1.6B)051.0-29.2--
451.8-35.3--
851.4-39.0--
Flamingo (9B)0*51.844.728.831.857.0
456.349.334.933.662.7
858.050.039.433.663.9
1659.450.843.033.564.5
Flamingo (80B)0*56.350.631.635.046.4
463.157.439.636.568.6
865.657.544.837.370.0
1666.857.848.437.670.0
IDEFICS (80B)0*60.045.236.030.960.6
463.652.440.434.457.8
864.855.146.135.758.2
1665.456.848.336.357.8
Emu (14B)0*52.942.834.4--
458.4-41.3--
859.0-43.9--
16-----
Emu2 (37B)033.526.740.426.452.2
467.053.254.648.262.4
867.854.154.749.365.8
1668.857.157.050.366.0
", + "bbox": [ + 506, + 135, + 890, + 467 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Zero-shot and few-shot evaluations of Emu2. $0^{*}$ denotes text two-shot and image zero-shot results following Flamingo [5]. The best results are in bold and the second best are underlined.", + "bbox": [ + 498, + 476, + 890, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Objective. We use the same unified generative pretraining objective to adapt to diverse generation tasks in context. Specifically, a training sample for generation is formulated as: “A photo of

a man

image embedding of object localization image[IMG] image embedding of man[/IMG] sitting next to

a dog

image embedding of object localization image[IMG] image embedding of dog[/IMG][IMG] image embedding of the whole image[/IMG”. We represent the coordinates of each object directly in image form by drawing the bounding box of each object at its specified location on a black image. Our Emu2-Gen conducts unified multimodal modeling of the text, object image, and corresponding object localization image. The regression loss only applies to the visual embeddings of the last image. We freeze the Visual Encoder during fine-tuning. We randomly drop tokens of entities and object localization image to enhance model adaptability and robustness. Additionally, we apply data augmentation to each object image, incorporating random background variations and random crop, aiming to reduce the reliance on image backgrounds. Training Setup. We use a global batch size of 4,096 and", + "bbox": [ + 496, + 537, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14402", + "bbox": [ + 480, + 926, + 519, + 936 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/365871c6dd5272dc4273fc556f6098340601702f5e796cb4ef00d51286185620.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelVisual Question AnswerLMM Benchmarks
VQAv2 [28]OKVQA [53]GQA [32]VizWiz [29]TextVQA [68]MSVD [82]MSRVTT [82]SEED [39]MM-Vet [88]TS [7]MMMU [89]
Flamingo-9B [5]51.844.7-28.8-30.213.7----
Flamingo-80B [5]56.350.6-31.6-35.617.4----
Kosmos-1 [31]51.0--29.2-------
Kosmos-2 [57]51.1------50.0--26.6
BLIP-2-13B [42]--41.019.642.520.310.346.422.4--
InstructBLIP-13B [22]--49.533.450.741.224.8-25.6552.4-
IDEFICS-9B [38]50.938.4-35.525.9------
IDEFICS-80B [38]60.045.2-36.030.9------
Shikra-13B [15]77.4*47.2---------
Qwen-VL-13B-Chat [6]78.2*56.6*57.5*38.961.5*--58.2-645.2-
LLaVA-1.5-13B [48]80.0*-63.3*53.661.3--61.635.4-33.6
CogVLM [77]83.4*58.9*--68.1*----662.630.1
Emu-I [71]62.049.246.038.3-37.021.2-36.3--
Emu2-Chat84.9*64.8*65.1*54.966.6*49.031.462.848.5703.834.1
", + "bbox": [ + 81, + 88, + 890, + 334 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Results on visual question answering and LMM benchmarks. * indicates that samples from this task's training set have been trained. SEED and TS respectively represent SEED-Bench [39] and TouchStone [7]. For MM-Vet, we present the average result of five scoring runs.", + "bbox": [ + 75, + 345, + 892, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "train for 3k steps. The learning rate linearly warms up to $5 \\times 10^{-5}$ in the first 100 steps, then decays to zero with a cosine schedule. We further fine-tune for 900 steps using the 500k high-quality pairs with a batch size of 2048.", + "bbox": [ + 75, + 414, + 468, + 474 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3. Evaluation", + "text_level": 1, + "bbox": [ + 76, + 488, + 192, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1. Pretrained Base Model", + "text_level": 1, + "bbox": [ + 76, + 512, + 290, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate zero-shot and few-shot abilities of Emu2 on OKVQA [53], VQAv2 [28], VizWiz [29], TextVQA [68], and HatefulMemes [35] tasks. Details of the datasets and prompts can be found in supplementary materials. The results are presented in Table 1. Emu2 demonstrates remarkable in-context ability, showcasing improved performance with more in-context samples seen. Specifically, on VQAv2, VizWiz and TextVQA datasets, Emu2 outperforms Flamingo-80B and IDEFICS-80B under all few-shot settings with a much smaller model scale (37B).", + "bbox": [ + 75, + 535, + 468, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 1 demonstrates Emu2's few-shot capabilities in the wild. For example, the model learns to classify and count simultaneously in a specific format via a few examples (row 1). Additionally, Emu2 is capable of following visual prompts in context, e.g., the red circles laid on the images (row 2 and 3).", + "bbox": [ + 75, + 686, + 468, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2. Instruction-Following Chat", + "text_level": 1, + "bbox": [ + 76, + 786, + 321, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our Emu2-Chat is evaluated on academic-task-oriented benchmarks including image question-answering datasets (VQAv2 [28], OKVQA [53], GQA [32], VizWiz [29], TextVQA [68]) and video question-answering datasets (MSVD [82] and MSRvTT [82]). The evaluation also encompassed recent benchmarks for large multimodal models,", + "bbox": [ + 75, + 809, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "including SEED-Bench [39], MM-Vet [88], TouchStone [7] and MMMU [89]. When evaluated on SEED-Bench, we followed the setup of LLaVa-1.5 [48] by presenting options to the model for completing multiple-choice tasks.", + "bbox": [ + 496, + 414, + 890, + 474 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Table 2, Emu2-Chat consistently outperforms other models in image question-answering tasks, encompassing well-established benchmarks like VQAv2 and GQA. Notably, it shows a noticeable improvement in the OKVQA task, which requires the utilization of external knowledge, showcasing the advantage of our model for mastering real-world knowledge. For video question-answering, Emu2-Chat demonstrated advantages even though it did not use video question-answering data for training. It achieved an accuracy of 49.0 and 31.4 on the MSVD-QA and MSRVTT-QA tasks, respectively, surpassing InstructBLIP and the larger Flamingo-80B. More importantly, our model has also achieved better results on LMM benchmarks. LMM benchmarks such as MM-Vet provide a more comprehensive evaluation of model abilities, including solving complicated tasks. Emu2-Chat achieves a score of 48.5 in MM-Vet and 703.8 in Touch-Stone, confirming its superior capability in understanding and solving multimodal problems.", + "bbox": [ + 496, + 476, + 892, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition, we demonstrated the visual grounding capability of our model using the refer expression comprehension benchmarks. In Table 3, Emu2-Chat achieved the best results among generalist models on RefCOCO [34], RefCOCO+ [52] and RefCOCOg [52]. Its most notable advantage was observed in RefCOCO+, which focused solely on purely appearance-based descriptions without allowing the use of position references. This highlights our model's powerful perceptual abilities in capturing intricate details.", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14403", + "bbox": [ + 478, + 925, + 517, + 936 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8cc37d98935c4f44559647e055e5f0294afab6627bad7cb4697c07728efeac58.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 89, + 166, + 140 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f49789c75d3cf1d5b93bf13793e3a023eb895543a8af4f2a5b2930fc81ddbdfd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 141, + 166, + 191 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c243d98ec61ed32f69243e80a599b8a9c5fb00f121a214bf3369082167fad9d3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 191, + 158, + 241 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e35517afe733d538ccf4abdc92ef411ca9a0e99a964baaafae17d14b8d268c6c.jpg", + "image_caption": [ + "A photo of the first dog>, the second dog>, the third dog on the grass" + ], + "image_footnote": [], + "bbox": [ + 179, + 90, + 318, + 196 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cf05a40c61a0659c785072d14443bdd2adf475adde31682340a08b7a756a1698.jpg", + "image_caption": [ + "An oil painting of the first dog>, the second dog>, the third dog>" + ], + "image_footnote": [], + "bbox": [ + 321, + 90, + 460, + 198 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3d8dbce32e38d79ef07ae1922ac658afe4324c4f9d5c27275e43f0118f0e2846.jpg", + "image_caption": [ + "A photo of the first dog>, the second dog>, the third dog> on the beach" + ], + "image_footnote": [], + "bbox": [ + 464, + 90, + 602, + 198 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/01eeb3d932b34db42a4a83ad3bf541fe2cc358e560872b8eb9e0f9b01c743ac2.jpg", + "image_caption": [ + "as an oil painting by Monet" + ], + "image_footnote": [], + "bbox": [ + 604, + 90, + 743, + 198 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/005d7167506e27504e53f460afd17bcf90099b1c73eebf5344af922093879ae2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 611, + 207, + 645, + 234 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0cd54a6d7deaf963723f17b3fb76825c5236b86ffa07e4d515fa1518840ce7dc.jpg", + "image_caption": [ + "in a dark forest in 3D" + ], + "image_footnote": [], + "bbox": [ + 746, + 90, + 885, + 198 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9e567b0a83330b66bae6b4cc1437fded089295b75a239fef9e4f5c54cbea2b24.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 750, + 208, + 779, + 234 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4a23fa92e42fe9417d7aabb40a922353a9b7b3f04942f0384346061cb73e86b1.jpg", + "image_caption": [ + "a bear and a sunflower" + ], + "image_footnote": [], + "bbox": [ + 83, + 243, + 254, + 376 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f3224340123304d8763ef9a5dd124aa5ebf91c211ecbfc6dff103f8c9847ba64.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 256, + 243, + 429, + 376 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e6bacea7baf4fbd6906a22b63e90333636fda2da2df53b1470d32aaa70c78242.jpg", + "image_caption": [ + "Figure 4. Visualization of Emu2-Gen's controllable generation capability. The model is capable of accepting a mix of text, locations and images as input, and generating images in context. The presented examples include text- and subject-grounded generation, stylization, multi-entity composition, subject-driven editing, and text-to-image generation." + ], + "image_footnote": [], + "bbox": [ + 259, + 378, + 428, + 420 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9a15867f22c2fc32224497c3d97c80b1999a30acde3f37a3d56c88e747a209e7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 429, + 243, + 602, + 376 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/46c7465f93b0dc2005ac0d6f2fc237cad65f7b6e4b397ea06add7a4379628794.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 431, + 378, + 598, + 420 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5c79ae0cdef25fe4e2b8769659255fdca36387d774d3a21bb355df5e16fab49e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 604, + 243, + 743, + 349 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3dc820fd695e8c875e65609b25c378a1b990c43c26ed2812ac2427b1855f56a3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 604, + 352, + 743, + 421 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/31ee407d139f4e9740f54ce47a40de1cd41b635024c73154551387b368021693.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 746, + 243, + 885, + 349 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d77f6c2bc5f9b120d3614cba07753f7c2f472f93935ba6a3ee954a5beba1054e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 746, + 352, + 885, + 421 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7f27792e67ded5ea3b9c6dfe6e7e0ff949fbefc79e7c32e5f91ddc2b90a55a6b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelRefCOCORefCOCO+RefCOCOg
valtestAtestBvaltestAtestBvaltest
OFA-L [75]79.9683.6776.3968.2976.0061.7567.5767.58
Shikra-7B [15]87.0190.6180.2481.6087.3672.1282.2782.19
Shikra-13B [15]87.8391.1181.8182.8987.7974.4182.6483.16
Qwen-VL-7B [6]89.3692.2685.3483.1288.2577.2185.5885.48
Emu2-Chat90.4093.8885.9787.0591.4380.4787.6488.11
CogVLM [77]92.5193.9588.7387.5291.8181.4389.4690.09
", + "bbox": [ + 81, + 502, + 465, + 614 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Results on referring expression comprehension. We grayed out CogVLM because its generalist grounding-enhanced model was specialist trained on high-quality grounding data.", + "bbox": [ + 75, + 625, + 468, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3. Controllable Visual Generation", + "text_level": 1, + "bbox": [ + 76, + 688, + 354, + 704 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Results. Figure 3 presents a visualization of Emu2's autoencoding results. With Emu2's Visual Encoder and Visual Decoder, we can tokenize an image into visual embeddings and detokenize them back. Compared with SEED [27] and Emu [71], Emu2 shows significantly superior results. We also evaluate our image autoencoding results on MS-COCO [47] and achieve a strong 0.907 CLIP-I [59] score. More results are in the supplementary.", + "bbox": [ + 75, + 715, + 468, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As depicted in Figure 4, Emu2-Gen is capable of accepting a mixture of text, locations and images as input, and generating images in context. The model skillfully engages in various controllable visual generation tasks in", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "a zero-shot setting, capitalizing on the in-context learning capabilities in multimodality. Examples in Figure 4 show generated images of three dogs conditioned on different subjects, locations and scenarios. The presented visual samples demonstrate the model's proficiency in tasks such as re-contextualization, stylization, modification, region-controllable generation, and multi-entity composition.", + "bbox": [ + 496, + 503, + 890, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Zero-shot Text-to-image Generation. We evaluate the zero-shot text-to-image generation capability on 30k randomly sampled data from the MS-COCO [47] validation set. We employ CLIP-ViT-B [60], following the approach in DALL-E 3[9], to calculate the CLIP-T score to assess prompt-following ability. Additionally, we utilize CLIP-ViT-L, as in GILL[37], to compute the CLIP-I score for measuring image similarity. A higher score means the generated image is more similar to the prompt or the real image. Table 4 shows that Emu2-Gen achieves the state-of-the-art performance in terms of both CLIP-I and CLIP-T scores compared to various unimodal generation models and multimodal models. More text-to-image generation cases can be found in supplementary.", + "bbox": [ + 496, + 611, + 892, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Zero-shot Subject-driven Generation. Following Kosmos-G [56], we also evaluate our model's subject-driven image editing ability on DreamBench [63]. We generate four images for each prompt, resulting in a total of 3,000 images for a comprehensive evaluation. We employ", + "bbox": [ + 496, + 824, + 892, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14404", + "bbox": [ + 480, + 926, + 517, + 936 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/91e1531e7a5b2390ff9e39991f7c1a04f99612797b7a2a0634fe857b3343c96a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelsCLIP-I ↑CLIP-T ↑
unimodal generation models
MUSE [13]-0.320
Imagen [64]-0.270
DALL-E 2 † [61]-0.314
DALL-E 3 † [9]-0.320
SDv1.5 [62]0.6670.302
SDXL [58]0.6740.310
multimodal generation models
GILL [37]0.684-
SEED [27]0.682-
Emu [71]0.6560.286
Emu2-Gen0.6860.297
", + "bbox": [ + 147, + 88, + 401, + 286 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6d08b44ae618af4ac578ab497c003150a35e2e283d9dddf5f58694f4109f72b3.jpg", + "table_caption": [ + "Table 4. Quantitative comparison of zero-shot text-to-image generation on MS-COCO [47] validation set. 30k samples are randomly sampled. $\\dagger$ CLIP-T score is calculated on 4,096 samples. We also evaluate our image autoencoding results on MS-COCO which achieves a strong 0.907 CLIP-I score." + ], + "table_footnote": [], + "table_body": "
MethodsDINO ↑CLIP-I ↑CLIP-T ↑
Real Images (Oracle)0.7740.885-
Fine-Tuning
Textual Inversion [24]0.5690.7800.255
DreamBooth [63]0.6680.8030.305
BLIP-Diffusion [42]0.6700.8050.302
Test Time Tuning Free
Re-Imagen* [16]0.6000.7400.270
SuTI [17]0.7410.8190.304
BLIP-Diffusion* [42]0.5940.7790.300
Kosmos-G* (single image input)0.6940.8470.287
Emu2-Gen * (single image input)0.7660.8500.287
", + "bbox": [ + 84, + 382, + 465, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Quantitative comparison of zero-shot single-entity subject-driven generation on DreamBench. * denotes zero-shot methods.", + "bbox": [ + 75, + 590, + 468, + 632 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "DINO [12] and CLIP-I [59] to evaluate subject fidelity, and CLIP-T [59] to evaluate text fidelity, aligning with the methodology established by DreamBooth. Notably, Emu2-Gen excels in subject fidelity, as evidenced by its superior performance on DINO and CLIP-I metrics compared to methods like BLIP-Diffusion and Kosmos-G. Emu2-Gen impressively reconstructs subjects with just one image input in zero-shot setting, demonstrating superior subject fidelity through powerful visual decoding. Further illustrative cases are provided in the supplementary, showcasing Emu2-Gen's proficiency in multi-entity generation.", + "bbox": [ + 75, + 660, + 468, + 828 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4. Related Work", + "text_level": 1, + "bbox": [ + 76, + 844, + 218, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Large Multimodal Models. Recent years have witnessed the rapid growth of large multimodal models [5, 19, 31,", + "bbox": [ + 75, + 869, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "71]. CLIP [59] pioneered the learning of LMMs with a contrastive learning objective on massive image-text pair data. Flamingo [5] and Kosmos [31, 57] exhibit promising zero-shot and few-shot multi-modal understanding performance by training on large-scale image-text interleaved data. With the remarkable progress in open-sourced LLMs, [42, 43, 49, 94] show promising results by connecting vision encoders and LLMs with a small intermediate model. A school of successive efforts [69, 76, 84, 90, 91] further improves visual instruction tuning with better overall training pipelines [6, 40], grounding annotations [14, 15, 85, 92], and extra tasks [6]. There are early studies on training more unified large multimodal models [23, 27, 71, 86] that are capable of performing visual understanding and generation simultaneously. In this paper, we further explore the distinct solution proposed in Emu [71]: learning large multimodal models with generative objectives on both texts and images.", + "bbox": [ + 496, + 90, + 890, + 349 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In-Context Learning. Recent advancements in large language models [11, 21] underscore their capacity for in-context learning [11]. This phenomenon, particularly evident as LLMs scale up in size and data, has been exploited for complex challenges such as mathematical reasoning [81], signaling new emergent ability in model behavior [80]. Flamingo [5] integrates visual inputs to LLMs, enabling the in-context learning of visual-linguistic tasks such as image captioning and OCR through language-based interfacing. Painter [78] and SegGPT [79] conduct an early study of visual in-context learning. Inspired by the emerging abilities of large language models, in this work we study the problem of multimodal in-context learning by scaling up generative multimodal models and demonstrating strong results in broad understanding and generation tasks.", + "bbox": [ + 496, + 358, + 890, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 613, + 617, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present a 37 billion-parameter generative multimodal model Emu2 that shows strong performance and versatility on many multimodal tasks in the in-context settings. Emu2 serves as a base model and a general-purpose interface for a variety of multimodal tasks. We demonstrate state-of-the-art results on a broad range of benchmarks of multimodal understanding and generation. Specifically, our model largely surpasses prior work on the lately proposed LMM benchmarks that require more advanced capability compared to classic academic benchmarks. Emu2 also shows remarkable capability of controllable visual generation in multimodal context, e.g., subject-/text-grounded generation. Additionally, we review the limitations and broader social impact of Emu2. Despite discussed weaknesses, these results suggest that generative multimodal model at scale may be an important step towards the development of adaptable, general multimodal systems.", + "bbox": [ + 496, + 643, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14405", + "bbox": [ + 480, + 926, + 517, + 936 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Laion-aesthetics. https://laion.ai/blog/laion-aesthetics/, .4,5,2,3", + "[2] Laion coco: 600m synthetic captions from laion2b-en. https://laion.ai/blog/laion-coco/,4,2", + "[3] Laion-high-resolution. https://huggingface.co/datasets/laion/laion-high-resolution, 5, 3", + "[4] Sharegpt. https://sharegpt.com/.5,2", + "[5] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 1, 5, 6, 8, 3", + "[6] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 6, 7, 8", + "[7] Shuai Bai, Shusheng Yang, Jinze Bai, Peng Wang, Xingxuan Zhang, Junyang Lin, Xinggang Wang, Chang Zhou, and Jingren Zhou. Touchstone: Evaluating vision-language models by language models. arXiv preprint arXiv:2308.16890, 2023. 6", + "[8] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1728-1738, 2021. 3, 1", + "[9] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, Wesam Manassra, Prafulla Dhariwal, Casey Chu, Yunxin Jiao, and Aditya Ramesh. Improving image generation with better captions. 2023. 5, 7, 8, 3", + "[10] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 5, 3", + "[11] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1, 8", + "[12] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 8", + "[13] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. 8", + "[14] Chi Chen, Ruoyu Qin, Fuwen Luo, Xiaoyue Mi, Peng Li, Maosong Sun, and Yang Liu. Position-enhanced visual" + ], + "bbox": [ + 76, + 114, + 467, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "instruction tuning for multimodal large language models. arXiv preprint arXiv:2308.13437, 2023. 8", + "[15] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 6, 7, 8", + "[16] Wenhu Chen, Hexiang Hu, Chitwan Sahara, and William W Cohen. Re-imagen: Retrieval-augmented text-to-image generator. arXiv preprint arXiv:2209.14491, 2022. 8", + "[17] Wenhu Chen, Hexiang Hu, Yandong Li, Nataniel Rui, Xuhui Jia, Ming-Wei Chang, and William W Cohen. Subject-driven text-to-image generation via apprenticeship learning. arXiv preprint arXiv:2304.00186, 2023. 8", + "[18] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 4, 2", + "[19] Xi Chen, Xiao Wang, Lucas Beyer, Alexander Kolesnikov, Jialin Wu, Paul Voigtlaender, Basil Mustafa, Sebastian Goodman, Ibrahim Alabdulmohsin, Piotr Padlewski, et al. Pali-3 vision language models: Smaller, faster, stronger. arXiv preprint arXiv:2310.09199, 2023. 1, 8", + "[20] Luke Chesser and Timothy Carbone. Unsplash. https://github.com/unsplash/datasets.5,3", + "[21] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 1, 8", + "[22] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Albert Li, Pascale Fung, and Steven C. H. Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500, 2023. 6", + "[23] Runpei Dong, Chunrui Han, Yuang Peng, Zekun Qi, Zheng Ge, Jinrong Yang, Liang Zhao, Jianjian Sun, Hongyu Zhou, Haoran Wei, Xiangwen Kong, Xiangyu Zhang, Kaisheng Ma, and Li Yi. Dreamllm: Synergistic multimodal comprehension and creation. arXiv preprint arXiv:2309.11499, 2023.8", + "[24] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 8", + "[25] Zhe Gan, Linjie Li, Chunyuan Li, Lijuan Wang, Zicheng Liu, Jianfeng Gao, et al. Vision-language pre-training: Basics, recent advances, and future trends. Foundations and Trends® in Computer Graphics and Vision, 14(3-4):163-352, 2022. 1", + "[26] Leo Gao, Stella Biderman, Sid Black, Laurence Golding, Travis Hoppe, Charles Foster, Jason Phang, Horace He, Anish Thite, Noa Nabeshima, Shawn Presser, and Connor Leahy. The Pile: An 800gb dataset of diverse text for language modeling. arXiv preprint arXiv:2101.00027, 2020. 3, 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14406", + "bbox": [ + 480, + 926, + 519, + 936 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Yuying Ge, Yixiao Ge, Ziyun Zeng, Xintao Wang, and Ying Shan. Planting a seed of vision in large language model. arXiv preprint arXiv:2307.08041, 2023. 4, 7, 8", + "[28] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 4, 6, 2", + "[29] Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. Vizwiz grand challenge: Answering visual questions from blind people. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3608-3617, 2018. 6", + "[30] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 4", + "[31] Shaohan Huang, Li Dong, Wenhui Wang, Yaru Hao, Saksham Singhal, Shuming Ma, Tengchao Lv, Lei Cui, Owais Khan Mohammed, Qiang Liu, et al. Language is not all you need: Aligning perception with language models. arXiv preprint arXiv:2302.14045, 2023. 6, 8", + "[32] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 4, 6, 2", + "[33] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in Neural Information Processing Systems, 35:26565-26577, 2022. 3", + "[34] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014. 5, 6, 2", + "[35] Douwe Kiela, Hamed Firooz, Aravind Mohan, Vedanuj Goswami, Amanpreet Singh, Pratik Ringshia, and Davide Testuggine. The hateful memes challenge: Detecting hate speech in multimodal memes. Advances in neural information processing systems, 33:2611-2624, 2020. 6", + "[36] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 5, 3", + "[37] Jing Yu Koh, Daniel Fried, and Ruslan Salakhutdinov. Generating images with multimodal language models. arXiv preprint arXiv:2305.17216, 2023. 7, 8", + "[38] Hugo Laurençon, Lucile Saulnier, Léo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, Thomas Wang, Siddharth Karamcheti, Alexander M Rush, Douwe Kiela, et al. Obelics: An open web-scale filtered dataset of interleaved image-text documents. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 6", + "[39] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking mul" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "timodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 6", + "[40] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726, 2023.8", + "[41] Chunyuan Li, Zhe Gan, Zhengyuan Yang, Jianwei Yang, Linjie Li, Lijuan Wang, and Jianfeng Gao. Multimodal foundation models: From specialists to general-purpose assistants. arXiv preprint arXiv:2309.10020, 1(2):2, 2023. 1", + "[42] Dongxu Li, Junnan Li, and Steven CH Hoi. Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. arXiv preprint arXiv:2305.14720, 2023. 6, 8", + "[43] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 8", + "[44] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 5, 2", + "[45] Lei Li, Yuwei Yin, Shicheng Li, Liang Chen, Peiyi Wang, Shuhuai Ren, Mukai Li, Yazheng Yang, Jingjing Xu, Xu Sun, et al. M3it: A large-scale dataset towards multimodal multilingual instruction tuning. arXiv preprint arXiv:2306.04387, 2023. 4, 2", + "[46] Xin Li, Wenqing Chu, Ye Wu, Weihang Yuan, Fanglong Liu, Qi Zhang, Fu Li, Haocheng Feng, Errui Ding, and Jingdong Wang. Videogen: A reference-guided latent diffusion approach for high definition text-to-video generation. arXiv preprint arXiv:2309.00398, 2023. 4", + "[47] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 7, 8", + "[48] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 4, 6, 3", + "[49] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 5, 8, 2", + "[50] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 3, 4", + "[51] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems, 35:2507–2521, 2022. 4", + "[52] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 5, 6, 2" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14407", + "bbox": [ + 480, + 926, + 517, + 936 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 3195-3204, 2019. 4, 6, 2", + "[54] Midjourney. Midjourney. https://www.midjourney.com. 5, 3", + "[55] Leonardo Nicoletti and Dina Bass. Humans are biased: Generative ai is even worse. *Bloomberg Technology+ Equality*. Accessed June, 23:2023, 2023. 1", + "[56] Xichen Pan, Li Dong, Shaohan Huang, Zhiliang Peng, Wenhu Chen, and Furu Wei. Kosmos-g: Generating images in context with multimodal large language models. arXiv preprint arXiv:2310.02992, 2023. 7, 4", + "[57] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 3, 5, 6, 8, 1, 2", + "[58] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023. 1, 3, 4, 8, 2", + "[59] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7, 8", + "[60] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7", + "[61] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022.8", + "[62] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 4, 8", + "[63] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 7, 8, 4", + "[64] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 8", + "[65] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. arXiv preprint arXiv:2210.08402, 2022. 3, 1", + "[66] Oleksii Sidorov, Ronghang Hu, Marcus Rohrbach, and Amanpreet Singh. Textcaps: a dataset for image captioning with reading comprehension. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 742-758. Springer, 2020. 4, 2", + "[67] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 4", + "[68] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 4, 6, 2", + "[69] Yixuan Su, Tian Lan, Huayang Li, Jialu Xu, Yan Wang, and Deng Cai. Pandagpt: One model to instruction-follow them all. arXiv preprint arXiv:2305.16355, 2023. 8", + "[70] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389, 2023. 3", + "[71] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative pretraining in multimodality. arXiv preprint arXiv:2307.05222, 2023. 3, 4, 6, 7, 8, 1", + "[72] Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B Hashimoto. Stanford alpaca: An instruction-following llama model, 2023. 5, 2", + "[73] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 3", + "[74] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 4", + "[75] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In International Conference on Machine Learning, pages 23318-23340. PMLR, 2022. 7", + "[76] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, et al. Vision: Large language model is also an open-ended decoder for vision-centric tasks. arXiv preprint arXiv:2305.11175, 2023. 8", + "[77] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "14408", + "bbox": [ + 480, + 925, + 517, + 935 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Song, et al. Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079, 2023. 6, 7", + "[78] Xinlong Wang, Wen Wang, Yue Cao, Chunhua Shen, and Tiejun Huang. Images speak in images: A generalist painter for in-context visual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6830-6839, 2023. 8", + "[79] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Segmenting everything in context. arXiv preprint arXiv:2304.03284, 2023. 8", + "[80] Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, et al. Emergent abilities of large language models. arXiv preprint arXiv:2206.07682, 2022. 8", + "[81] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 35:24824-24837, 2022. 8", + "[82] Dejing Xu, Zhou Zhao, Jun Xiao, Fei Wu, Hanwang Zhang, Xiangnan He, and Yueting Zhuang. Video question answering via gradually refined attention over appearance and motion. In Proceedings of the 25th ACM international conference on Multimedia, pages 1645-1653, 2017. 6", + "[83] Charig Yang, Weidi Xie, and Andrew Zisserman. It's about time: Analog clock reading in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2508-2517, 2022. 5, 2", + "[84] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. MPLug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 8", + "[85] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704, 2023. 8", + "[86] Lili Yu, Bowen Shi, Ramakanth Pasunuru, Benjamin Muller, Olga Golovneva, Tianlu Wang, Arun Babu, Binh Tang, Brian Karrer, Shelly Sheynin, et al. Scaling autoregressive multimodal models: Pretraining and instruction tuning. arXiv preprint arXiv:2309.02591, 2023. 8", + "[87] Qiying Yu, Quan Sun, Xiaosong Zhang, Yufeng Cui, Fan Zhang, Xinlong Wang, and Jingjing Liu. Capsfusion: Rethinking image-text data at scale. arXiv preprint arXiv:2310.20550, 2023. 3, 5, 1", + "[88] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490, 2023. 6", + "[89] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. arXiv preprint arXiv:2311.16502, 2023. 6" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[90] Ao Zhang, Hao Fei, Yuan Yao, Wei Ji, Li Li, Zhiyuan Liu, and Tat-Seng Chua. Transfer visual prompt generator across llms. arXiv preprint arXiv:2305.01278, 2023. 8", + "[91] Pan Zhang, Xiaoyi Dong Bin Wang, Yuhang Cao, Chao Xu, Linke Ouyang, Zhiyuan Zhao, Shuangrui Ding, Songyang Zhang, Haodong Duan, Hang Yan, et al. Internlmxcomposer: A vision-language large model for advanced text-image comprehension and composition. arXiv preprint arXiv:2309.15112, 2023. 8", + "[92] Shilong Zhang, Peize Sun, Shoufa Chen, Min Xiao, Wenqi Shao, Wenwei Zhang, Kai Chen, and Ping Luo. Gpt4roi: Instruction tuning large language model on region-of-interest. arXiv preprint arXiv:2307.03601, 2023. 8", + "[93] Yanzhe Zhang, Ruiyi Zhang, Jiuxiang Gu, Yufan Zhou, Nedim Lipka, Diyi Yang, and Tong Sun. Llavar: Enhanced visual instruction tuning for text-rich image understanding. arXiv preprint arXiv:2306.17107, 2023. 5, 2", + "[94] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 8", + "[95] Wanrong Zhu, Jack Hessel, Anas Awadalla, Samir Yitzhak Gadre, Jesse Dodge, Alex Fang, Youngjae Yu, Ludwig Schmidt, William Yang Wang, and Yejin Choi. Multimodal c4: An open, billion-scale corpus of images interleaved with text. arXiv preprint arXiv:2304.06939, 2023. 3, 1" + ], + "bbox": [ + 501, + 92, + 890, + 458 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "14409", + "bbox": [ + 480, + 924, + 519, + 936 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2024/Generative Multimodal Models are In-Context Learners/1a0461d9-1d47-4e83-8d07-3f1f380c508b_model.json b/2024/Generative Multimodal Models are In-Context Learners/1a0461d9-1d47-4e83-8d07-3f1f380c508b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..705ebf631f613baa12ada302216b6478829a28f4 --- /dev/null +++ b/2024/Generative Multimodal Models are In-Context Learners/1a0461d9-1d47-4e83-8d07-3f1f380c508b_model.json @@ -0,0 +1,3502 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.2, + 0.131, + 0.773, + 0.153 + ], + "angle": 0, + "content": "Generative Multimodal Models are In-Context Learners" + }, + { + "type": "text", + "bbox": [ + 0.086, + 0.179, + 0.884, + 0.219 + ], + "angle": 0, + "content": "Quan Sun\\(^{1*}\\) Yufeng Cui\\(^{1*}\\) Xiaosong Zhang\\(^{1*}\\) Fan Zhang\\(^{1*}\\) Qiying Yu\\(^{2,1*}\\) Yueze Wang\\(^{1}\\) Yongming Rao\\(^{1}\\) Jingjing Liu\\(^{2}\\) Tiejun Huang\\(^{1,3}\\) Xinlong Wang\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.162, + 0.224, + 0.817, + 0.264 + ], + "angle": 0, + "content": "\\(^{1}\\) Beijing Academy of Artificial Intelligence \\(^{2}\\) Tsinghua University \\(^{3}\\) Peking University *equal contribution †project lead" + }, + { + "type": "text", + "bbox": [ + 0.278, + 0.273, + 0.691, + 0.288 + ], + "angle": 0, + "content": "code & models: https://github.com/baaivision/Emu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.323, + 0.314, + 0.339 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.354, + 0.474, + 0.688 + ], + "angle": 0, + "content": "The human ability to easily solve multimodal tasks in context (i.e., with only a few demonstrations or simple instructions), is what current multimodal systems have largely struggled to imitate. In this work, we demonstrate that the task-agnostic in-context learning capabilities of large multimodal models can be significantly enhanced by effective scaling-up. We introduce Emu2, a generative multimodal model with 37 billion parameters, trained on large-scale multimodal sequences with a unified autoregressive objective. Emu2 exhibits strong multimodal in-context learning abilities, even emerging to solve tasks that require on-the-fly reasoning, such as visual prompting and object-grounded generation. The model sets a new record on multiple multimodal understanding tasks in few-shot settings. When instruction-tuned to follow specific instructions, Emu2 further achieves new state-of-the-art on challenging tasks such as question answering benchmarks for large multimodal models and open-ended subject-driven generation. These achievements demonstrate that Emu2 can serve as a base model and general-purpose interface for a wide range of multimodal tasks. Code and models are publicly available to facilitate future research." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.716, + 0.21, + 0.731 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.741, + 0.47, + 0.879 + ], + "angle": 0, + "content": "Multimodal tasks [25, 41] encompass anything involving understanding and generation in single or multiple modalities [5, 19, 58], which can be highly diverse and long-tail. Previous multimodal systems largely rely on designing task-specific architecture and collecting a sizable supervised training set, both of which are difficult to scale, particularly when this process needs to be repeated for each new task encountered. By contrast, humans can solve a new task in context, i.e., with only a few demonstrations or simple" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.324, + 0.892, + 0.354 + ], + "angle": 0, + "content": "instructions - a capability that current multimodal models have yet to learn." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.355, + 0.893, + 0.551 + ], + "angle": 0, + "content": "Recently, generative pretrained language models have demonstrated strong in-context learning abilities [11, 21, 73]. By training a 37-billion-parameter model Emu2 and thoroughly evaluating it on diverse multimodal tasks, we demonstrate that a scaled-up multimodal generative pretrained model can harness similar in-context learning abilities and effectively generalize to unseen multimodal tasks. Emu2 is trained with a unified autoregressive objective: predict-the next-multimodal-element (either visual embeddings or textual tokens). In this unified generative pretraining process, large-scale multimodal sequences (e.g., text, image-text pairs, and interleaved image-text-video) are used for model training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.552, + 0.894, + 0.658 + ], + "angle": 0, + "content": "We measure Emu2's capabilities of learning from a few examples or instructions on standard multimodal datasets, as well as new tasks unseen in the training set. Specifically, Emu2 is evaluated under two scenarios: (a) few-shot setting, where we allow as many examples as possible to fit the context window of the model; and (b) instruction tuning, where the model is tuned to follow specific instructions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.895, + 0.824 + ], + "angle": 0, + "content": "Emu2 achieves promising results in the few-shot setting on a wide range of vision-language tasks. For example, it demonstrates state-of-the-art few-shot performance on multiple visual question-answering datasets. We observe a performance improvement when the number of examples in context increases. Figure 1 illustrates Emu2's strong multimodal reasoning capabilities for tasks in the wild, e.g., recognition and counting in a specific format. Emu2 also learns to follow visual prompting in context (e.g., the circles laid on the images in Figure 1), even although it struggles at a smaller scale or at zero shot." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.903 + ], + "angle": 0, + "content": "As Emu2 is inherently equipped to handle interleaved text-image-video at both input and output, it serves as a powerful and versatile base model for diverse multimodal tasks, by following specific task instructions. For example, after instruct tuning with conversational data," + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.887, + 0.341, + 0.901 + ], + "angle": 0, + "content": "†Correspondence to wangxinlong@baai.ac.cn" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.52, + 0.937 + ], + "angle": 0, + "content": "14398" + } + ], + [ + { + "type": "aside_text", + "bbox": [ + 0.101, + 0.19, + 0.117, + 0.261 + ], + "angle": 270, + "content": "In-context Completion" + }, + { + "type": "image", + "bbox": [ + 0.126, + 0.116, + 0.208, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.132, + 0.293, + 0.143 + ], + "angle": 0, + "content": "[dog: 1, frisbee: 1]." + }, + { + "type": "image", + "bbox": [ + 0.301, + 0.115, + 0.34, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.35, + 0.132, + 0.468, + 0.143 + ], + "angle": 0, + "content": "[burger: 1, glass: 1, bottle: 1]." + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.12, + 0.548, + 0.157 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.583, + 0.134, + 0.615, + 0.143 + ], + "angle": 0, + "content": "[cat: 3]." + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.116, + 0.721, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.173, + 0.635, + 0.192 + ], + "angle": 0, + "content": "The text in the red circle: Do Not?" + }, + { + "type": "image", + "bbox": [ + 0.644, + 0.206, + 0.721, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.172, + 0.292, + 0.192 + ], + "angle": 0, + "content": "The text in the red circle: 'Rights'." + }, + { + "type": "image", + "bbox": [ + 0.299, + 0.161, + 0.379, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.392, + 0.173, + 0.466, + 0.192 + ], + "angle": 0, + "content": "The text in the red circle: 'Ave'." + }, + { + "type": "image", + "bbox": [ + 0.479, + 0.16, + 0.555, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.56, + 0.251, + 0.635, + 0.272 + ], + "angle": 0, + "content": "of" + }, + { + "type": "image_caption", + "bbox": [ + 0.653, + 0.259, + 0.712, + 0.286 + ], + "angle": 0, + "content": "a photo of a blue and red backpack:" + }, + { + "type": "image_caption", + "bbox": [ + 0.771, + 0.099, + 0.821, + 0.11 + ], + "angle": 0, + "content": "Completion" + }, + { + "type": "image_caption", + "bbox": [ + 0.756, + 0.133, + 0.836, + 0.143 + ], + "angle": 0, + "content": "[beer: 3, banana: 2]." + }, + { + "type": "image_caption", + "bbox": [ + 0.729, + 0.178, + 0.861, + 0.188 + ], + "angle": 0, + "content": "The text in the red circle: 'Lynn'." + }, + { + "type": "image_caption", + "bbox": [ + 0.749, + 0.223, + 0.842, + 0.233 + ], + "angle": 0, + "content": "motorcycle's headlight." + }, + { + "type": "image_caption", + "bbox": [ + 0.142, + 0.257, + 0.186, + 0.286 + ], + "angle": 0, + "content": "a photo of a yellow backpack:" + }, + { + "type": "image_caption", + "bbox": [ + 0.151, + 0.303, + 0.22, + 0.331 + ], + "angle": 0, + "content": "The subject A with a city in the background:" + }, + { + "type": "image", + "bbox": [ + 0.2, + 0.25, + 0.255, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.27, + 0.259, + 0.313, + 0.286 + ], + "angle": 0, + "content": "a photo of a blue backpack:" + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.293, + 0.301, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.328, + 0.25, + 0.385, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.312, + 0.308, + 0.41, + 0.326 + ], + "angle": 0, + "content": "The subject A wearing a santa hat:" + }, + { + "type": "image", + "bbox": [ + 0.422, + 0.25, + 0.48, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.481, + 0.25, + 0.513, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.258, + 0.571, + 0.287 + ], + "angle": 0, + "content": "a photo of a brown backpack:" + }, + { + "type": "image", + "bbox": [ + 0.584, + 0.248, + 0.642, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.421, + 0.295, + 0.48, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.49, + 0.308, + 0.573, + 0.327 + ], + "angle": 0, + "content": "The subject A in a purple wizard outfit:" + }, + { + "type": "image", + "bbox": [ + 0.584, + 0.302, + 0.642, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.766, + 0.249, + 0.826, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.766, + 0.298, + 0.825, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "aside_text", + "bbox": [ + 0.106, + 0.376, + 0.119, + 0.46 + ], + "angle": 270, + "content": "Multimodal understanding" + }, + { + "type": "image", + "bbox": [ + 0.158, + 0.364, + 0.239, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.13, + 0.425, + 0.264, + 0.471 + ], + "angle": 0, + "content": "Look at this sequence of three shapes. What shape should come as the fourth shape? Explain your reasoning with detailed descriptions of the first shapes." + }, + { + "type": "text", + "bbox": [ + 0.273, + 0.369, + 0.401, + 0.47 + ], + "angle": 0, + "content": "The first shape is a triangle, the second shape is a square, and the third shape is a pentagon. To find the next shape in the sequence, we need to identify a pattern. In this case, the pattern is that each shape has one more side than the previous shape. Therefore, the next shape in the sequence should have six sides, which is a hexagon." + }, + { + "type": "image", + "bbox": [ + 0.44, + 0.367, + 0.522, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.419, + 0.443, + 0.549, + 0.453 + ], + "angle": 0, + "content": "What is funny about this image?" + }, + { + "type": "text", + "bbox": [ + 0.558, + 0.363, + 0.688, + 0.48 + ], + "angle": 0, + "content": "The funny aspect of this image is that a squirrel is holding a camera and appears to be taking a picture of a bird. This is amusing because animals, such as squirrels and birds, do not typically use cameras or engage in human-like activities like photography. The scene is set up in a way that makes it look like the squirrel is intentionally capturing the bird's image, which adds a humorous twist to the scene. Additionally, the bird is perched on a mushroom, further enhancing the whimsical and playful nature of the image." + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.354, + 0.868, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "aside_text", + "bbox": [ + 0.104, + 0.614, + 0.117, + 0.728 + ], + "angle": 270, + "content": "Generate from any prompt sequence" + }, + { + "type": "image", + "bbox": [ + 0.125, + 0.499, + 0.268, + 0.61 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.142, + 0.62, + 0.253, + 0.639 + ], + "angle": 0, + "content": "impressionist painting of an astronaut in a jungle" + }, + { + "type": "image", + "bbox": [ + 0.437, + 0.499, + 0.582, + 0.608 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.31, + 0.508, + 0.388, + 0.602 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.31, + 0.624, + 0.543, + 0.634 + ], + "angle": 0, + "content": "An image of wearing a big sunglasses on the beach" + }, + { + "type": "image", + "bbox": [ + 0.583, + 0.5, + 0.721, + 0.61 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.723, + 0.499, + 0.868, + 0.61 + ], + "angle": 0, + "content": null + }, + { + "type": "aside_text", + "bbox": [ + 0.104, + 0.614, + 0.117, + 0.728 + ], + "angle": 270, + "content": "Generate from any prompt sequence" + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.697, + 0.277, + 0.707 + ], + "angle": 0, + "content": "An anime draw of " + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.677, + 0.349, + 0.731 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.679, + 0.388, + 0.73 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.392, + 0.698, + 0.447, + 0.708 + ], + "angle": 0, + "content": "and \\(\\text{a dog}\\)" + }, + { + "type": "image", + "bbox": [ + 0.452, + 0.677, + 0.517, + 0.731 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.677, + 0.589, + 0.731 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.593, + 0.699, + 0.658, + 0.708 + ], + "angle": 0, + "content": ", in the style of" + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.687, + 0.725, + 0.722 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.731, + 0.648, + 0.868, + 0.753 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.797, + 0.279, + 0.816 + ], + "angle": 0, + "content": " and [USER]: [ASSISTANT]: ”. Here represents system message and varies between the two major task categories (academic-task-oriented and multimodal chat). The section comprises multimodal tokens, including images, videos, and text. Only tokens in the section will be supervised by cross-entropy loss during training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.32, + 0.474, + 0.562 + ], + "angle": 0, + "content": "Training Setup. We use a global batch size of 768 and train for 8k steps. The learning rate linearly warms up to \\(1 \\times 10^{-5}\\) in the first 100 steps, then decays to zero with a cosine schedule. The model is trained using the AdamW optimizer with \\(\\beta_{1} = 0.9\\), \\(\\beta_{2} = 0.98\\), \\(\\epsilon = 1 \\times 10^{-6}\\), and a gradient clipping of 5.0. The sequence length during training is limited to 2048, and any excess beyond that is truncated directly. We consistently employed an input image/video resolution of \\(448 \\times 448\\). For video data, we uniformly sample frames in time as input to the model. The number of sampled frames for each video is randomly chosen from 8, 12, and 16. To capture more intricate spatial details, following the visual encoder stage, we apply mean-pooling to each static image, dividing it into \\(16 \\times 16\\) tokens during instruction fine-tuning. This differs from the pre-training phase, where \\(8 \\times 8\\) tokens were utilized." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.587, + 0.352, + 0.603 + ], + "angle": 0, + "content": "2.3.2 Controllable Visual Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.613, + 0.47, + 0.809 + ], + "angle": 0, + "content": "Training Data. We leverage a mix of high-quality datasets to unleash the potential of controllable generation in context. We use a grounded image-text pair dataset CapsFusion-grounded-100M and GRIT [57] for grounded text-to-image generation. To mitigate the impact of image backgrounds on the effectiveness of multi-entity subject-driven generation, we employ SAM [36] to preprocess the grounding data, yielding a subset of approximately 5 million samples with segmentation results. Additionally, we leverage InstructPix2Pix constructed by [10] for image editing tasks. For the text-to-image task, we use a filtered subset of CapsFusion [87], LAION-Aesthetics [1], SA-1B [36], and LAION-High-Resolution [3]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.902 + ], + "angle": 0, + "content": "We also collect data from premium sources (e.g., Unsplash [20]) and outputs from advanced text-to-image systems (e.g., Midjourney-V5 [54] and DALL-E-3 [9]) for quality fine-tuning. This diverse dataset includes around 500k high-quality image-text pairs. For all the data above, during the training, only samples with image resolutions" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.123 + ], + "angle": 0, + "content": "higher than \\(448 \\times 448\\) were retained to ensure generation quality. More details can be found in the supplementary." + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.136, + 0.891, + 0.468 + ], + "angle": 0, + "content": "
ModelShotVQAv2OKVQAVizWizTextVQAHateful Memes
Kosmos-1 (1.6B)051.0-29.2--
451.8-35.3--
851.4-39.0--
Flamingo (9B)0*51.844.728.831.857.0
456.349.334.933.662.7
858.050.039.433.663.9
1659.450.843.033.564.5
Flamingo (80B)0*56.350.631.635.046.4
463.157.439.636.568.6
865.657.544.837.370.0
1666.857.848.437.670.0
IDEFICS (80B)0*60.045.236.030.960.6
463.652.440.434.457.8
864.855.146.135.758.2
1665.456.848.336.357.8
Emu (14B)0*52.942.834.4--
458.4-41.3--
859.0-43.9--
16-----
Emu2 (37B)033.526.740.426.452.2
467.053.254.648.262.4
867.854.154.749.365.8
1668.857.157.050.366.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.477, + 0.892, + 0.52 + ], + "angle": 0, + "content": "Table 1. Zero-shot and few-shot evaluations of Emu2. \\(0^{*}\\) denotes text two-shot and image zero-shot results following Flamingo [5]. The best results are in bold and the second best are underlined." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Training Objective. We use the same unified generative pretraining objective to adapt to diverse generation tasks in context. Specifically, a training sample for generation is formulated as: “A photo of

a man

image embedding of object localization image[IMG] image embedding of man[/IMG] sitting next to

a dog

image embedding of object localization image[IMG] image embedding of dog[/IMG][IMG] image embedding of the whole image[/IMG”. We represent the coordinates of each object directly in image form by drawing the bounding box of each object at its specified location on a black image. Our Emu2-Gen conducts unified multimodal modeling of the text, object image, and corresponding object localization image. The regression loss only applies to the visual embeddings of the last image. We freeze the Visual Encoder during fine-tuning. We randomly drop tokens of entities and object localization image to enhance model adaptability and robustness. Additionally, we apply data augmentation to each object image, incorporating random background variations and random crop, aiming to reduce the reliance on image backgrounds. Training Setup. We use a global batch size of 4,096 and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.927, + 0.52, + 0.938 + ], + "angle": 0, + "content": "14402" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.891, + 0.335 + ], + "angle": 0, + "content": "
ModelVisual Question AnswerLMM Benchmarks
VQAv2 [28]OKVQA [53]GQA [32]VizWiz [29]TextVQA [68]MSVD [82]MSRVTT [82]SEED [39]MM-Vet [88]TS [7]MMMU [89]
Flamingo-9B [5]51.844.7-28.8-30.213.7----
Flamingo-80B [5]56.350.6-31.6-35.617.4----
Kosmos-1 [31]51.0--29.2-------
Kosmos-2 [57]51.1------50.0--26.6
BLIP-2-13B [42]--41.019.642.520.310.346.422.4--
InstructBLIP-13B [22]--49.533.450.741.224.8-25.6552.4-
IDEFICS-9B [38]50.938.4-35.525.9------
IDEFICS-80B [38]60.045.2-36.030.9------
Shikra-13B [15]77.4*47.2---------
Qwen-VL-13B-Chat [6]78.2*56.6*57.5*38.961.5*--58.2-645.2-
LLaVA-1.5-13B [48]80.0*-63.3*53.661.3--61.635.4-33.6
CogVLM [77]83.4*58.9*--68.1*----662.630.1
Emu-I [71]62.049.246.038.3-37.021.2-36.3--
Emu2-Chat84.9*64.8*65.1*54.966.6*49.031.462.848.5703.834.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.346, + 0.893, + 0.389 + ], + "angle": 0, + "content": "Table 2. Results on visual question answering and LMM benchmarks. * indicates that samples from this task's training set have been trained. SEED and TS respectively represent SEED-Bench [39] and TouchStone [7]. For MM-Vet, we present the average result of five scoring runs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.415, + 0.47, + 0.476 + ], + "angle": 0, + "content": "train for 3k steps. The learning rate linearly warms up to \\(5 \\times 10^{-5}\\) in the first 100 steps, then decays to zero with a cosine schedule. We further fine-tune for 900 steps using the 500k high-quality pairs with a batch size of 2048." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.489, + 0.194, + 0.504 + ], + "angle": 0, + "content": "3. Evaluation" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.513, + 0.291, + 0.529 + ], + "angle": 0, + "content": "3.1. Pretrained Base Model" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.536, + 0.47, + 0.687 + ], + "angle": 0, + "content": "We evaluate zero-shot and few-shot abilities of Emu2 on OKVQA [53], VQAv2 [28], VizWiz [29], TextVQA [68], and HatefulMemes [35] tasks. Details of the datasets and prompts can be found in supplementary materials. The results are presented in Table 1. Emu2 demonstrates remarkable in-context ability, showcasing improved performance with more in-context samples seen. Specifically, on VQAv2, VizWiz and TextVQA datasets, Emu2 outperforms Flamingo-80B and IDEFICS-80B under all few-shot settings with a much smaller model scale (37B)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.687, + 0.47, + 0.779 + ], + "angle": 0, + "content": "Figure 1 demonstrates Emu2's few-shot capabilities in the wild. For example, the model learns to classify and count simultaneously in a specific format via a few examples (row 1). Additionally, Emu2 is capable of following visual prompts in context, e.g., the red circles laid on the images (row 2 and 3)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.323, + 0.803 + ], + "angle": 0, + "content": "3.2. Instruction-Following Chat" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Our Emu2-Chat is evaluated on academic-task-oriented benchmarks including image question-answering datasets (VQAv2 [28], OKVQA [53], GQA [32], VizWiz [29], TextVQA [68]) and video question-answering datasets (MSVD [82] and MSRvTT [82]). The evaluation also encompassed recent benchmarks for large multimodal models," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.415, + 0.892, + 0.476 + ], + "angle": 0, + "content": "including SEED-Bench [39], MM-Vet [88], TouchStone [7] and MMMU [89]. When evaluated on SEED-Bench, we followed the setup of LLaVa-1.5 [48] by presenting options to the model for completing multiple-choice tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.477, + 0.893, + 0.764 + ], + "angle": 0, + "content": "As shown in Table 2, Emu2-Chat consistently outperforms other models in image question-answering tasks, encompassing well-established benchmarks like VQAv2 and GQA. Notably, it shows a noticeable improvement in the OKVQA task, which requires the utilization of external knowledge, showcasing the advantage of our model for mastering real-world knowledge. For video question-answering, Emu2-Chat demonstrated advantages even though it did not use video question-answering data for training. It achieved an accuracy of 49.0 and 31.4 on the MSVD-QA and MSRVTT-QA tasks, respectively, surpassing InstructBLIP and the larger Flamingo-80B. More importantly, our model has also achieved better results on LMM benchmarks. LMM benchmarks such as MM-Vet provide a more comprehensive evaluation of model abilities, including solving complicated tasks. Emu2-Chat achieves a score of 48.5 in MM-Vet and 703.8 in Touch-Stone, confirming its superior capability in understanding and solving multimodal problems." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.894, + 0.903 + ], + "angle": 0, + "content": "In addition, we demonstrated the visual grounding capability of our model using the refer expression comprehension benchmarks. In Table 3, Emu2-Chat achieved the best results among generalist models on RefCOCO [34], RefCOCO+ [52] and RefCOCOg [52]. Its most notable advantage was observed in RefCOCO+, which focused solely on purely appearance-based descriptions without allowing the use of position references. This highlights our model's powerful perceptual abilities in capturing intricate details." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.926, + 0.519, + 0.938 + ], + "angle": 0, + "content": "14403" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.09, + 0.168, + 0.141 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.142, + 0.167, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.192, + 0.159, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.181, + 0.091, + 0.319, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.192, + 0.208, + 0.309, + 0.237 + ], + "angle": 0, + "content": "A photo of the first dog>, the second dog>, the third dog on the grass" + }, + { + "type": "image", + "bbox": [ + 0.322, + 0.091, + 0.462, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.331, + 0.208, + 0.453, + 0.237 + ], + "angle": 0, + "content": "An oil painting of the first dog>, the second dog>, the third dog>" + }, + { + "type": "image", + "bbox": [ + 0.465, + 0.091, + 0.604, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.475, + 0.208, + 0.593, + 0.237 + ], + "angle": 0, + "content": "A photo of the first dog>, the second dog>, the third dog> on the beach" + }, + { + "type": "image", + "bbox": [ + 0.606, + 0.091, + 0.745, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.208, + 0.646, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.654, + 0.214, + 0.724, + 0.232 + ], + "angle": 0, + "content": "as an oil painting by Monet" + }, + { + "type": "image", + "bbox": [ + 0.748, + 0.091, + 0.886, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.751, + 0.209, + 0.78, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.79, + 0.217, + 0.874, + 0.226 + ], + "angle": 0, + "content": "in a dark forest in 3D" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.244, + 0.256, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.395, + 0.221, + 0.405 + ], + "angle": 0, + "content": "a bear and a sunflower" + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.244, + 0.43, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.261, + 0.38, + 0.429, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.431, + 0.244, + 0.604, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.432, + 0.38, + 0.599, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.606, + 0.244, + 0.744, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.606, + 0.353, + 0.744, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.748, + 0.244, + 0.886, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.748, + 0.353, + 0.886, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.436, + 0.893, + 0.479 + ], + "angle": 0, + "content": "Figure 4. Visualization of Emu2-Gen's controllable generation capability. The model is capable of accepting a mix of text, locations and images as input, and generating images in context. The presented examples include text- and subject-grounded generation, stylization, multi-entity composition, subject-driven editing, and text-to-image generation." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.503, + 0.466, + 0.616 + ], + "angle": 0, + "content": "
ModelRefCOCORefCOCO+RefCOCOg
valtestAtestBvaltestAtestBvaltest
OFA-L [75]79.9683.6776.3968.2976.0061.7567.5767.58
Shikra-7B [15]87.0190.6180.2481.6087.3672.1282.2782.19
Shikra-13B [15]87.8391.1181.8182.8987.7974.4182.6483.16
Qwen-VL-7B [6]89.3692.2685.3483.1288.2577.2185.5885.48
Emu2-Chat90.4093.8885.9787.0591.4380.4787.6488.11
CogVLM [77]92.5193.9588.7387.5291.8181.4389.4690.09
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.626, + 0.47, + 0.669 + ], + "angle": 0, + "content": "Table 3. Results on referring expression comprehension. We grayed out CogVLM because its generalist grounding-enhanced model was specialist trained on high-quality grounding data." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.689, + 0.355, + 0.705 + ], + "angle": 0, + "content": "3.3. Controllable Visual Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.716, + 0.47, + 0.837 + ], + "angle": 0, + "content": "Qualitative Results. Figure 3 presents a visualization of Emu2's autoencoding results. With Emu2's Visual Encoder and Visual Decoder, we can tokenize an image into visual embeddings and detokenize them back. Compared with SEED [27] and Emu [71], Emu2 shows significantly superior results. We also evaluate our image autoencoding results on MS-COCO [47] and achieve a strong 0.907 CLIP-I [59] score. More results are in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "As depicted in Figure 4, Emu2-Gen is capable of accepting a mixture of text, locations and images as input, and generating images in context. The model skillfully engages in various controllable visual generation tasks in" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.505, + 0.892, + 0.611 + ], + "angle": 0, + "content": "a zero-shot setting, capitalizing on the in-context learning capabilities in multimodality. Examples in Figure 4 show generated images of three dogs conditioned on different subjects, locations and scenarios. The presented visual samples demonstrate the model's proficiency in tasks such as re-contextualization, stylization, modification, region-controllable generation, and multi-entity composition." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.893, + 0.824 + ], + "angle": 0, + "content": "Zero-shot Text-to-image Generation. We evaluate the zero-shot text-to-image generation capability on 30k randomly sampled data from the MS-COCO [47] validation set. We employ CLIP-ViT-B [60], following the approach in DALL-E 3[9], to calculate the CLIP-T score to assess prompt-following ability. Additionally, we utilize CLIP-ViT-L, as in GILL[37], to compute the CLIP-I score for measuring image similarity. A higher score means the generated image is more similar to the prompt or the real image. Table 4 shows that Emu2-Gen achieves the state-of-the-art performance in terms of both CLIP-I and CLIP-T scores compared to various unimodal generation models and multimodal models. More text-to-image generation cases can be found in supplementary." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Zero-shot Subject-driven Generation. Following Kosmos-G [56], we also evaluate our model's subject-driven image editing ability on DreamBench [63]. We generate four images for each prompt, resulting in a total of 3,000 images for a comprehensive evaluation. We employ" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.927, + 0.519, + 0.938 + ], + "angle": 0, + "content": "14404" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.148, + 0.089, + 0.402, + 0.287 + ], + "angle": 0, + "content": "
ModelsCLIP-I ↑CLIP-T ↑
unimodal generation models
MUSE [13]-0.320
Imagen [64]-0.270
DALL-E 2 † [61]-0.314
DALL-E 3 † [9]-0.320
SDv1.5 [62]0.6670.302
SDXL [58]0.6740.310
multimodal generation models
GILL [37]0.684-
SEED [27]0.682-
Emu [71]0.6560.286
Emu2-Gen0.6860.297
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.298, + 0.472, + 0.368 + ], + "angle": 0, + "content": "Table 4. Quantitative comparison of zero-shot text-to-image generation on MS-COCO [47] validation set. 30k samples are randomly sampled. \\(\\dagger\\)CLIP-T score is calculated on 4,096 samples. We also evaluate our image autoencoding results on MS-COCO which achieves a strong 0.907 CLIP-I score." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.383, + 0.467, + 0.581 + ], + "angle": 0, + "content": "
MethodsDINO ↑CLIP-I ↑CLIP-T ↑
Real Images (Oracle)0.7740.885-
Fine-Tuning
Textual Inversion [24]0.5690.7800.255
DreamBooth [63]0.6680.8030.305
BLIP-Diffusion [42]0.6700.8050.302
Test Time Tuning Free
Re-Imagen* [16]0.6000.7400.270
SuTI [17]0.7410.8190.304
BLIP-Diffusion* [42]0.5940.7790.300
Kosmos-G* (single image input)0.6940.8470.287
Emu2-Gen * (single image input)0.7660.8500.287
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.591, + 0.47, + 0.633 + ], + "angle": 0, + "content": "Table 5. Quantitative comparison of zero-shot single-entity subject-driven generation on DreamBench. * denotes zero-shot methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.661, + 0.47, + 0.829 + ], + "angle": 0, + "content": "DINO [12] and CLIP-I [59] to evaluate subject fidelity, and CLIP-T [59] to evaluate text fidelity, aligning with the methodology established by DreamBooth. Notably, Emu2-Gen excels in subject fidelity, as evidenced by its superior performance on DINO and CLIP-I metrics compared to methods like BLIP-Diffusion and Kosmos-G. Emu2-Gen impressively reconstructs subjects with just one image input in zero-shot setting, demonstrating superior subject fidelity through powerful visual decoding. Further illustrative cases are provided in the supplementary, showcasing Emu2-Gen's proficiency in multi-entity generation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.845, + 0.22, + 0.86 + ], + "angle": 0, + "content": "4. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Large Multimodal Models. Recent years have witnessed the rapid growth of large multimodal models [5, 19, 31," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.091, + 0.892, + 0.35 + ], + "angle": 0, + "content": "71]. CLIP [59] pioneered the learning of LMMs with a contrastive learning objective on massive image-text pair data. Flamingo [5] and Kosmos [31, 57] exhibit promising zero-shot and few-shot multi-modal understanding performance by training on large-scale image-text interleaved data. With the remarkable progress in open-sourced LLMs, [42, 43, 49, 94] show promising results by connecting vision encoders and LLMs with a small intermediate model. A school of successive efforts [69, 76, 84, 90, 91] further improves visual instruction tuning with better overall training pipelines [6, 40], grounding annotations [14, 15, 85, 92], and extra tasks [6]. There are early studies on training more unified large multimodal models [23, 27, 71, 86] that are capable of performing visual understanding and generation simultaneously. In this paper, we further explore the distinct solution proposed in Emu [71]: learning large multimodal models with generative objectives on both texts and images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.359, + 0.892, + 0.586 + ], + "angle": 0, + "content": "In-Context Learning. Recent advancements in large language models [11, 21] underscore their capacity for in-context learning [11]. This phenomenon, particularly evident as LLMs scale up in size and data, has been exploited for complex challenges such as mathematical reasoning [81], signaling new emergent ability in model behavior [80]. Flamingo [5] integrates visual inputs to LLMs, enabling the in-context learning of visual-linguistic tasks such as image captioning and OCR through language-based interfacing. Painter [78] and SegGPT [79] conduct an early study of visual in-context learning. Inspired by the emerging abilities of large language models, in this work we study the problem of multimodal in-context learning by scaling up generative multimodal models and demonstrating strong results in broad understanding and generation tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.614, + 0.619, + 0.63 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We present a 37 billion-parameter generative multimodal model Emu2 that shows strong performance and versatility on many multimodal tasks in the in-context settings. Emu2 serves as a base model and a general-purpose interface for a variety of multimodal tasks. We demonstrate state-of-the-art results on a broad range of benchmarks of multimodal understanding and generation. Specifically, our model largely surpasses prior work on the lately proposed LMM benchmarks that require more advanced capability compared to classic academic benchmarks. Emu2 also shows remarkable capability of controllable visual generation in multimodal context, e.g., subject-/text-grounded generation. Additionally, we review the limitations and broader social impact of Emu2. Despite discussed weaknesses, these results suggest that generative multimodal model at scale may be an important step towards the development of adaptable, general multimodal systems." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.927, + 0.519, + 0.938 + ], + "angle": 0, + "content": "14405" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.142 + ], + "angle": 0, + "content": "[1] Laion-aesthetics. https://laion.ai/blog/laion-aesthetics/, .4,5,2,3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.144, + 0.468, + 0.171 + ], + "angle": 0, + "content": "[2] Laion coco: 600m synthetic captions from laion2b-en. https://laion.ai/blog/laion-coco/,4,2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.468, + 0.212 + ], + "angle": 0, + "content": "[3] Laion-high-resolution. https://huggingface.co/datasets/laion/laion-high-resolution, 5, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.215, + 0.393, + 0.228 + ], + "angle": 0, + "content": "[4] Sharegpt. https://sharegpt.com/.5,2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.23, + 0.468, + 0.311 + ], + "angle": 0, + "content": "[5] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 1, 5, 6, 8, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.314, + 0.468, + 0.38 + ], + "angle": 0, + "content": "[6] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.383, + 0.468, + 0.45 + ], + "angle": 0, + "content": "[7] Shuai Bai, Shusheng Yang, Jinze Bai, Peng Wang, Xingxuan Zhang, Junyang Lin, Xinggang Wang, Chang Zhou, and Jingren Zhou. Touchstone: Evaluating vision-language models by language models. arXiv preprint arXiv:2308.16890, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.453, + 0.468, + 0.52 + ], + "angle": 0, + "content": "[8] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1728-1738, 2021. 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.524, + 0.468, + 0.591 + ], + "angle": 0, + "content": "[9] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, Wesam Manassra, Prafulla Dhariwal, Casey Chu, Yunxin Jiao, and Aditya Ramesh. Improving image generation with better captions. 2023. 5, 7, 8, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.594, + 0.468, + 0.661 + ], + "angle": 0, + "content": "[10] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 5, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.664, + 0.468, + 0.732 + ], + "angle": 0, + "content": "[11] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.734, + 0.468, + 0.802 + ], + "angle": 0, + "content": "[12] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.804, + 0.468, + 0.871 + ], + "angle": 0, + "content": "[13] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.873, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[14] Chi Chen, Ruoyu Qin, Fuwen Luo, Xiaoyue Mi, Peng Li, Maosong Sun, and Yang Liu. Position-enhanced visual" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.115, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.89, + 0.12 + ], + "angle": 0, + "content": "instruction tuning for multimodal large language models. arXiv preprint arXiv:2308.13437, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.124, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[15] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.181, + 0.892, + 0.222 + ], + "angle": 0, + "content": "[16] Wenhu Chen, Hexiang Hu, Chitwan Sahara, and William W Cohen. Re-imagen: Retrieval-augmented text-to-image generator. arXiv preprint arXiv:2209.14491, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.226, + 0.892, + 0.28 + ], + "angle": 0, + "content": "[17] Wenhu Chen, Hexiang Hu, Yandong Li, Nataniel Rui, Xuhui Jia, Ming-Wei Chang, and William W Cohen. Subject-driven text-to-image generation via apprenticeship learning. arXiv preprint arXiv:2304.00186, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.284, + 0.892, + 0.338 + ], + "angle": 0, + "content": "[18] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 4, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.342, + 0.892, + 0.41 + ], + "angle": 0, + "content": "[19] Xi Chen, Xiao Wang, Lucas Beyer, Alexander Kolesnikov, Jialin Wu, Paul Voigtlaender, Basil Mustafa, Sebastian Goodman, Ibrahim Alabdulmohsin, Piotr Padlewski, et al. Pali-3 vision language models: Smaller, faster, stronger. arXiv preprint arXiv:2310.09199, 2023. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.414, + 0.888, + 0.441 + ], + "angle": 0, + "content": "[20] Luke Chesser and Timothy Carbone. Unsplash. https://github.com/unsplash/datasets.5,3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.445, + 0.892, + 0.513 + ], + "angle": 0, + "content": "[21] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.516, + 0.892, + 0.585 + ], + "angle": 0, + "content": "[22] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Albert Li, Pascale Fung, and Steven C. H. Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.588, + 0.892, + 0.669 + ], + "angle": 0, + "content": "[23] Runpei Dong, Chunrui Han, Yuang Peng, Zekun Qi, Zheng Ge, Jinrong Yang, Liang Zhao, Jianjian Sun, Hongyu Zhou, Haoran Wei, Xiangwen Kong, Xiangyu Zhang, Kaisheng Ma, and Li Yi. Dreamllm: Synergistic multimodal comprehension and creation. arXiv preprint arXiv:2309.11499, 2023.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.674, + 0.892, + 0.741 + ], + "angle": 0, + "content": "[24] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.746, + 0.892, + 0.813 + ], + "angle": 0, + "content": "[25] Zhe Gan, Linjie Li, Chunyuan Li, Lijuan Wang, Zicheng Liu, Jianfeng Gao, et al. Vision-language pre-training: Basics, recent advances, and future trends. Foundations and Trends® in Computer Graphics and Vision, 14(3-4):163-352, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[26] Leo Gao, Stella Biderman, Sid Black, Laurence Golding, Travis Hoppe, Charles Foster, Jason Phang, Horace He, Anish Thite, Noa Nabeshima, Shawn Presser, and Connor Leahy. The Pile: An 800gb dataset of diverse text for language modeling. arXiv preprint arXiv:2101.00027, 2020. 3, 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.927, + 0.52, + 0.938 + ], + "angle": 0, + "content": "14406" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "[27] Yuying Ge, Yixiao Ge, Ziyun Zeng, Xintao Wang, and Ying Shan. Planting a seed of vision in large language model. arXiv preprint arXiv:2307.08041, 2023. 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.469, + 0.219 + ], + "angle": 0, + "content": "[28] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 4, 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.469, + 0.302 + ], + "angle": 0, + "content": "[29] Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. Vizwiz grand challenge: Answering visual questions from blind people. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3608-3617, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.469, + 0.333 + ], + "angle": 0, + "content": "[30] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.469, + 0.404 + ], + "angle": 0, + "content": "[31] Shaohan Huang, Li Dong, Wenhui Wang, Yaru Hao, Saksham Singhal, Shuming Ma, Tengchao Lv, Lei Cui, Owais Khan Mohammed, Qiang Liu, et al. Language is not all you need: Aligning perception with language models. arXiv preprint arXiv:2302.14045, 2023. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.469, + 0.473 + ], + "angle": 0, + "content": "[32] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 4, 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.476, + 0.469, + 0.531 + ], + "angle": 0, + "content": "[33] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in Neural Information Processing Systems, 35:26565-26577, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.533, + 0.469, + 0.602 + ], + "angle": 0, + "content": "[34] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014. 5, 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.604, + 0.469, + 0.673 + ], + "angle": 0, + "content": "[35] Douwe Kiela, Hamed Firooz, Aravind Mohan, Vedanuj Goswami, Amanpreet Singh, Pratik Ringshia, and Davide Testuggine. The hateful memes challenge: Detecting hate speech in multimodal memes. Advances in neural information processing systems, 33:2611-2624, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.469, + 0.731 + ], + "angle": 0, + "content": "[36] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 5, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.469, + 0.773 + ], + "angle": 0, + "content": "[37] Jing Yu Koh, Daniel Fried, and Ruslan Salakhutdinov. Generating images with multimodal language models. arXiv preprint arXiv:2305.17216, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[38] Hugo Laurençon, Lucile Saulnier, Léo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, Thomas Wang, Siddharth Karamcheti, Alexander M Rush, Douwe Kiela, et al. Obelics: An open web-scale filtered dataset of interleaved image-text documents. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[39] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking mul" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "timodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[40] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726, 2023.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.177, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[41] Chunyuan Li, Zhe Gan, Zhengyuan Yang, Jianwei Yang, Linjie Li, Lijuan Wang, and Jianfeng Gao. Multimodal foundation models: From specialists to general-purpose assistants. arXiv preprint arXiv:2309.10020, 1(2):2, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.233, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[42] Dongxu Li, Junnan Li, and Steven CH Hoi. Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. arXiv preprint arXiv:2305.14720, 2023. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.288, + 0.892, + 0.356 + ], + "angle": 0, + "content": "[43] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.358, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[44] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 5, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.413, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[45] Lei Li, Yuwei Yin, Shicheng Li, Liang Chen, Peiyi Wang, Shuhuai Ren, Mukai Li, Yazheng Yang, Jingjing Xu, Xu Sun, et al. M3it: A large-scale dataset towards multimodal multilingual instruction tuning. arXiv preprint arXiv:2306.04387, 2023. 4, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.484, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[46] Xin Li, Wenqing Chu, Ye Wu, Weihang Yuan, Fanglong Liu, Qi Zhang, Fu Li, Haocheng Feng, Errui Ding, and Jingdong Wang. Videogen: A reference-guided latent diffusion approach for high definition text-to-video generation. arXiv preprint arXiv:2309.00398, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.553, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[47] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.637, + 0.892, + 0.678 + ], + "angle": 0, + "content": "[48] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 4, 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.679, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[49] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 5, 8, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.72, + 0.892, + 0.748 + ], + "angle": 0, + "content": "[50] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.749, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[51] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems, 35:2507–2521, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[52] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 5, 6, 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.927, + 0.519, + 0.938 + ], + "angle": 0, + "content": "14407" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.162 + ], + "angle": 0, + "content": "[53] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 3195-3204, 2019. 4, 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.164, + 0.466, + 0.19 + ], + "angle": 0, + "content": "[54] Midjourney. Midjourney. https://www.midjourney.com. 5, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.193, + 0.468, + 0.234 + ], + "angle": 0, + "content": "[55] Leonardo Nicoletti and Dina Bass. Humans are biased: Generative ai is even worse. *Bloomberg Technology+ Equality*. Accessed June, 23:2023, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.236, + 0.469, + 0.291 + ], + "angle": 0, + "content": "[56] Xichen Pan, Li Dong, Shaohan Huang, Zhiliang Peng, Wenhu Chen, and Furu Wei. Kosmos-g: Generating images in context with multimodal large language models. arXiv preprint arXiv:2310.02992, 2023. 7, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.293, + 0.469, + 0.348 + ], + "angle": 0, + "content": "[57] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 3, 5, 6, 8, 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.35, + 0.469, + 0.405 + ], + "angle": 0, + "content": "[58] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023. 1, 3, 4, 8, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.407, + 0.469, + 0.489 + ], + "angle": 0, + "content": "[59] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.491, + 0.469, + 0.573 + ], + "angle": 0, + "content": "[60] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.576, + 0.469, + 0.63 + ], + "angle": 0, + "content": "[61] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.633, + 0.47, + 0.702 + ], + "angle": 0, + "content": "[62] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.704, + 0.469, + 0.786 + ], + "angle": 0, + "content": "[63] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 7, 8, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.788, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[64] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[65] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. arXiv preprint arXiv:2210.08402, 2022. 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.149, + 0.892, + 0.231 + ], + "angle": 0, + "content": "[66] Oleksii Sidorov, Ronghang Hu, Marcus Rohrbach, and Amanpreet Singh. Textcaps: a dataset for image captioning with reading comprehension. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 742-758. Springer, 2020. 4, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[67] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.302, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[68] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 4, 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[69] Yixuan Su, Tian Lan, Huayang Li, Jialu Xu, Yan Wang, and Deng Cai. Pandagpt: One model to instruction-follow them all. arXiv preprint arXiv:2305.16355, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[70] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.456, + 0.892, + 0.523 + ], + "angle": 0, + "content": "[71] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative pretraining in multimodality. arXiv preprint arXiv:2307.05222, 2023. 3, 4, 6, 7, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.525, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[72] Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B Hashimoto. Stanford alpaca: An instruction-following llama model, 2023. 5, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.676 + ], + "angle": 0, + "content": "[73] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.892, + 0.719 + ], + "angle": 0, + "content": "[74] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.721, + 0.892, + 0.803 + ], + "angle": 0, + "content": "[75] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In International Conference on Machine Learning, pages 23318-23340. PMLR, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[76] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, et al. Vision: Large language model is also an open-ended decoder for vision-centric tasks. arXiv preprint arXiv:2305.11175, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[77] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.926, + 0.519, + 0.936 + ], + "angle": 0, + "content": "14408" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "Song, et al. Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.472, + 0.189 + ], + "angle": 0, + "content": "[78] Xinlong Wang, Wen Wang, Yue Cao, Chunhua Shen, and Tiejun Huang. Images speak in images: A generalist painter for in-context visual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6830-6839, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.191, + 0.469, + 0.245 + ], + "angle": 0, + "content": "[79] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Segmenting everything in context. arXiv preprint arXiv:2304.03284, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.247, + 0.469, + 0.314 + ], + "angle": 0, + "content": "[80] Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, et al. Emergent abilities of large language models. arXiv preprint arXiv:2206.07682, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.317, + 0.469, + 0.384 + ], + "angle": 0, + "content": "[81] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 35:24824-24837, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.386, + 0.469, + 0.455 + ], + "angle": 0, + "content": "[82] Dejing Xu, Zhou Zhao, Jun Xiao, Fei Wu, Hanwang Zhang, Xiangnan He, and Yueting Zhuang. Video question answering via gradually refined attention over appearance and motion. In Proceedings of the 25th ACM international conference on Multimedia, pages 1645-1653, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.457, + 0.471, + 0.511 + ], + "angle": 0, + "content": "[83] Charig Yang, Weidi Xie, and Andrew Zisserman. It's about time: Analog clock reading in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2508-2517, 2022. 5, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.512, + 0.469, + 0.58 + ], + "angle": 0, + "content": "[84] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. MPLug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.469, + 0.649 + ], + "angle": 0, + "content": "[85] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.651, + 0.469, + 0.719 + ], + "angle": 0, + "content": "[86] Lili Yu, Bowen Shi, Ramakanth Pasunuru, Benjamin Muller, Olga Golovneva, Tianlu Wang, Arun Babu, Binh Tang, Brian Karrer, Shelly Sheynin, et al. Scaling autoregressive multimodal models: Pretraining and instruction tuning. arXiv preprint arXiv:2309.02591, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.721, + 0.469, + 0.775 + ], + "angle": 0, + "content": "[87] Qiying Yu, Quan Sun, Xiaosong Zhang, Yufeng Cui, Fan Zhang, Xinlong Wang, and Jingjing Liu. Capsfusion: Rethinking image-text data at scale. arXiv preprint arXiv:2310.20550, 2023. 3, 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.777, + 0.469, + 0.831 + ], + "angle": 0, + "content": "[88] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[89] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. arXiv preprint arXiv:2311.16502, 2023. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[90] Ao Zhang, Hao Fei, Yuan Yao, Wei Ji, Li Li, Zhiyuan Liu, and Tat-Seng Chua. Transfer visual prompt generator across llms. arXiv preprint arXiv:2305.01278, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[91] Pan Zhang, Xiaoyi Dong Bin Wang, Yuhang Cao, Chao Xu, Linke Ouyang, Zhiyuan Zhao, Shuangrui Ding, Songyang Zhang, Haodong Duan, Hang Yan, et al. Internlmxcomposer: A vision-language large model for advanced text-image comprehension and composition. arXiv preprint arXiv:2309.15112, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.22, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[92] Shilong Zhang, Peize Sun, Shoufa Chen, Min Xiao, Wenqi Shao, Wenwei Zhang, Kai Chen, and Ping Luo. Gpt4roi: Instruction tuning large language model on region-of-interest. arXiv preprint arXiv:2307.03601, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[93] Yanzhe Zhang, Ruiyi Zhang, Jiuxiang Gu, Yufan Zhou, Nedim Lipka, Diyi Yang, and Tong Sun. Llavar: Enhanced visual instruction tuning for text-rich image understanding. arXiv preprint arXiv:2306.17107, 2023. 5, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.333, + 0.892, + 0.388 + ], + "angle": 0, + "content": "[94] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.39, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[95] Wanrong Zhu, Jack Hessel, Anas Awadalla, Samir Yitzhak Gadre, Jesse Dodge, Alex Fang, Youngjae Yu, Ludwig Schmidt, William Yang Wang, and Yejin Choi. Multimodal c4: An open, billion-scale corpus of images interleaved with text. arXiv preprint arXiv:2304.06939, 2023. 3, 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.925, + 0.52, + 0.937 + ], + "angle": 0, + "content": "14409" + } + ] +] \ No newline at end of file diff --git a/2024/Generative Multimodal Models are In-Context Learners/1a0461d9-1d47-4e83-8d07-3f1f380c508b_origin.pdf b/2024/Generative Multimodal Models are In-Context Learners/1a0461d9-1d47-4e83-8d07-3f1f380c508b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..34a6025efe3d4b5e13fbb5fad301881467e8d923 --- /dev/null +++ b/2024/Generative Multimodal Models are In-Context Learners/1a0461d9-1d47-4e83-8d07-3f1f380c508b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cd1cdaa1df8c1a5d35cc0870951d5bd7981fbd094b34c72d7ac1b6e7563ca69e +size 7828386 diff --git a/2024/Generative Multimodal Models are In-Context Learners/full.md b/2024/Generative Multimodal Models are In-Context Learners/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c9a372668c58ebdfa1758dbf720dfdf75ef536cc --- /dev/null +++ b/2024/Generative Multimodal Models are In-Context Learners/full.md @@ -0,0 +1,445 @@ +# Generative Multimodal Models are In-Context Learners + +Quan Sun $^{1*}$ Yufeng Cui $^{1*}$ Xiaosong Zhang $^{1*}$ Fan Zhang $^{1*}$ Qiying Yu $^{2,1*}$ Yueze Wang $^{1}$ Yongming Rao $^{1}$ Jingjing Liu $^{2}$ Tiejun Huang $^{1,3}$ Xinlong Wang $^{1\dagger}$ + +$^{1}$ Beijing Academy of Artificial Intelligence $^{2}$ Tsinghua University $^{3}$ Peking University *equal contribution †project lead + +code & models: https://github.com/baaivision/Emu + +# Abstract + +The human ability to easily solve multimodal tasks in context (i.e., with only a few demonstrations or simple instructions), is what current multimodal systems have largely struggled to imitate. In this work, we demonstrate that the task-agnostic in-context learning capabilities of large multimodal models can be significantly enhanced by effective scaling-up. We introduce Emu2, a generative multimodal model with 37 billion parameters, trained on large-scale multimodal sequences with a unified autoregressive objective. Emu2 exhibits strong multimodal in-context learning abilities, even emerging to solve tasks that require on-the-fly reasoning, such as visual prompting and object-grounded generation. The model sets a new record on multiple multimodal understanding tasks in few-shot settings. When instruction-tuned to follow specific instructions, Emu2 further achieves new state-of-the-art on challenging tasks such as question answering benchmarks for large multimodal models and open-ended subject-driven generation. These achievements demonstrate that Emu2 can serve as a base model and general-purpose interface for a wide range of multimodal tasks. Code and models are publicly available to facilitate future research. + +# 1. Introduction + +Multimodal tasks [25, 41] encompass anything involving understanding and generation in single or multiple modalities [5, 19, 58], which can be highly diverse and long-tail. Previous multimodal systems largely rely on designing task-specific architecture and collecting a sizable supervised training set, both of which are difficult to scale, particularly when this process needs to be repeated for each new task encountered. By contrast, humans can solve a new task in context, i.e., with only a few demonstrations or simple + +instructions - a capability that current multimodal models have yet to learn. + +Recently, generative pretrained language models have demonstrated strong in-context learning abilities [11, 21, 73]. By training a 37-billion-parameter model Emu2 and thoroughly evaluating it on diverse multimodal tasks, we demonstrate that a scaled-up multimodal generative pretrained model can harness similar in-context learning abilities and effectively generalize to unseen multimodal tasks. Emu2 is trained with a unified autoregressive objective: predict-the next-multimodal-element (either visual embeddings or textual tokens). In this unified generative pretraining process, large-scale multimodal sequences (e.g., text, image-text pairs, and interleaved image-text-video) are used for model training. + +We measure Emu2's capabilities of learning from a few examples or instructions on standard multimodal datasets, as well as new tasks unseen in the training set. Specifically, Emu2 is evaluated under two scenarios: (a) few-shot setting, where we allow as many examples as possible to fit the context window of the model; and (b) instruction tuning, where the model is tuned to follow specific instructions. + +Emu2 achieves promising results in the few-shot setting on a wide range of vision-language tasks. For example, it demonstrates state-of-the-art few-shot performance on multiple visual question-answering datasets. We observe a performance improvement when the number of examples in context increases. Figure 1 illustrates Emu2's strong multimodal reasoning capabilities for tasks in the wild, e.g., recognition and counting in a specific format. Emu2 also learns to follow visual prompting in context (e.g., the circles laid on the images in Figure 1), even although it struggles at a smaller scale or at zero shot. + +As Emu2 is inherently equipped to handle interleaved text-image-video at both input and output, it serves as a powerful and versatile base model for diverse multimodal tasks, by following specific task instructions. For example, after instruct tuning with conversational data, + +![](images/4d1282fe13868ac44b6e10da23a599442d29205b9894a5f26b5835062e24bf21.jpg) +[dog: 1, frisbee: 1]. + +![](images/4c2ef205e12e6ea72513547bb6aa5691fd2aad11fc7cfce9ec1db0d23f8af9e6.jpg) +[burger: 1, glass: 1, bottle: 1]. + +![](images/e18a18357157aafcc0876e4d567730bb90498ce80ddb55e2f659af8654967ba8.jpg) + +![](images/bc74e1252b4635fccb7ae97a3523505ff7f2a16ca917c7d301d0cce46a8dfb95.jpg) +[cat: 3]. +The text in the red circle: Do Not? +Completion +[beer: 3, banana: 2]. +The text in the red circle: 'Lynn'. + +![](images/8f2304618c0af1da4f98ed5456c6033dde1bf907a936c8b9e8bbaadc877042d2.jpg) + +![](images/2ff32f6c5fc5c8c0e7d9c838edef336146ad3d5d2d452df3cd8a96eb5f848b2e.jpg) +The text in the red circle: 'Rights'. + +![](images/b6f719711669f26965c0c0aab109f6ff4de38cc022efa75785e1bfff82d10509.jpg) +The text in the red circle: 'Ave'. + +![](images/a984b813cd72fc5a86e70deb01b4d3b49ab8e913846529a250c48ce4643a60a9.jpg) +a photo of a yellow backpack: + +![](images/8d8b3d9a362c1cc3ca1888b4780271d0c8095498d8b52997a2db79f753101225.jpg) +The subject A with a city in the background: + +![](images/ae4d936285f56781589dc0c08f8b1f9de2abbb8c279041c27b5e4fbafaf51303.jpg) +a photo of a blue backpack: + +![](images/f3f3282f01c52853b8596959d8781b36847ee25556ebf3e9c4c85bb732f55bc6.jpg) +The subject A in a purple wizard outfit: + +![](images/b30403a6a631b901f5292a76ea31235533551cc8be5c0aa63f1586206c078bae.jpg) +a photo of a brown backpack: + +![](images/81e4a0bd18ca12fc80a72a2a4143e50915625f04ca02987532551a38d97e9ddc.jpg) +of +a photo of a blue and red backpack: + +![](images/d742bdad6bcf6a4c6004af35598c0cfb1d51bc43e843b792562535df65a5b081.jpg) +The subject A wearing a santa hat: + +![](images/ab5cbb65ae2ddcd90db53f9385e58ada3cdf0ff187222a07549312d913a3fd8e.jpg) + +![](images/a779de923f8cba57b2cfdde5bacdbe9d56b1752bd0eadf0c831a5494eb133946.jpg) +motorcycle's headlight. + +![](images/14e89c157256675bb4a8e3278a36e973f488147d308e5040932e0c0808a6d2cf.jpg) + +![](images/4a708d896cb4fa258a03419dd1c1d491df7b0b979056ea67b760ea64eb4bbf99.jpg) +Look at this sequence of three shapes. What shape should come as the fourth shape? Explain your reasoning with detailed descriptions of the first shapes. + +The first shape is a triangle, the second shape is a square, and the third shape is a pentagon. To find the next shape in the sequence, we need to identify a pattern. In this case, the pattern is that each shape has one more side than the previous shape. Therefore, the next shape in the sequence should have six sides, which is a hexagon. + +![](images/345c2ff9a7706acaced5a87e494f53efc5d3d183c9d9df39d4263a04e26b43d5.jpg) +What is funny about this image? + +The funny aspect of this image is that a squirrel is holding a camera and appears to be taking a picture of a bird. This is amusing because animals, such as squirrels and birds, do not typically use cameras or engage in human-like activities like photography. The scene is set up in a way that makes it look like the squirrel is intentionally capturing the bird's image, which adds a humorous twist to the scene. Additionally, the bird is perched on a mushroom, further enhancing the whimsical and playful nature of the image. + +![](images/5e49f2038c304b38fb988de176d8fa07331dd47731cc1c8820e381bd495b3f42.jpg) + +![](images/b8771eba724fe87515faa21c975df27701a76a07e0a20141161ad134e9c4e539.jpg) +impressionist painting of an astronaut in a jungle + +![](images/575e7154d1415b51921064162802d7ad53b7703bd2f0245a3910adde8cee94b5.jpg) +An image of wearing a big sunglasses on the beach + +![](images/dc5f36c519e6e38dbc6b601e0d5233a9d933c7ca79abcefde199813ee8ed810e.jpg) + +![](images/56e48415d335188b4e7a492cdb006a2680e263fdcaeca29c72fd9dfe218ea6a5.jpg) + +![](images/8631505cc0b1e5ae9da3fbadd84a9da737df6128a09e7df04523cad4b6683db6.jpg) + +![](images/e6daaa233f23264b059d0d6fd6b483e222aef55470fd14ca0e668e5a4710fbbb.jpg) +An anime draw of + +![](images/471cad8f46d65acc63f695dd251e311729e43cdc21a99e756df02e1b315f1afb.jpg) +and $\text{a dog}$ + +![](images/4d04223adb671d60a30154108673737b9de3adcdcd2fe98da48d3797d22e5833.jpg) + +![](images/ac7d0b522d218ac19c15f874ba476b6c20751be158e2ac5048376a2498259e1e.jpg) + +![](images/b719d4b5086e773d384f307c4c5d3406accab4e3d9f78f03dd0989ff5c0a63d3.jpg) +, in the style of + +![](images/03e2ddada0a6b3036cd26a09b3401d0d546ea6fa19e4db0c96652fa9318b98af.jpg) + +![](images/c81c5676d24211777e288e9f0df17aa44ec37ecb3bffc4db67f074b0bbe0313f.jpg) + and [USER]: [ASSISTANT]: ”. Here represents system message and varies between the two major task categories (academic-task-oriented and multimodal chat). The section comprises multimodal tokens, including images, videos, and text. Only tokens in the section will be supervised by cross-entropy loss during training. + +Training Setup. We use a global batch size of 768 and train for 8k steps. The learning rate linearly warms up to $1 \times 10^{-5}$ in the first 100 steps, then decays to zero with a cosine schedule. The model is trained using the AdamW optimizer with $\beta_{1} = 0.9$ , $\beta_{2} = 0.98$ , $\epsilon = 1 \times 10^{-6}$ , and a gradient clipping of 5.0. The sequence length during training is limited to 2048, and any excess beyond that is truncated directly. We consistently employed an input image/video resolution of $448 \times 448$ . For video data, we uniformly sample frames in time as input to the model. The number of sampled frames for each video is randomly chosen from 8, 12, and 16. To capture more intricate spatial details, following the visual encoder stage, we apply mean-pooling to each static image, dividing it into $16 \times 16$ tokens during instruction fine-tuning. This differs from the pre-training phase, where $8 \times 8$ tokens were utilized. + +# 2.3.2 Controllable Visual Generation + +Training Data. We leverage a mix of high-quality datasets to unleash the potential of controllable generation in context. We use a grounded image-text pair dataset CapsFusion-grounded-100M and GRIT [57] for grounded text-to-image generation. To mitigate the impact of image backgrounds on the effectiveness of multi-entity subject-driven generation, we employ SAM [36] to preprocess the grounding data, yielding a subset of approximately 5 million samples with segmentation results. Additionally, we leverage InstructPix2Pix constructed by [10] for image editing tasks. For the text-to-image task, we use a filtered subset of CapsFusion [87], LAION-Aesthetics [1], SA-1B [36], and LAION-High-Resolution [3]. + +We also collect data from premium sources (e.g., Unsplash [20]) and outputs from advanced text-to-image systems (e.g., Midjourney-V5 [54] and DALL-E-3 [9]) for quality fine-tuning. This diverse dataset includes around 500k high-quality image-text pairs. For all the data above, during the training, only samples with image resolutions + +higher than $448 \times 448$ were retained to ensure generation quality. More details can be found in the supplementary. + +
ModelShotVQAv2OKVQAVizWizTextVQAHateful Memes
Kosmos-1 (1.6B)051.0-29.2--
451.8-35.3--
851.4-39.0--
Flamingo (9B)0*51.844.728.831.857.0
456.349.334.933.662.7
858.050.039.433.663.9
1659.450.843.033.564.5
Flamingo (80B)0*56.350.631.635.046.4
463.157.439.636.568.6
865.657.544.837.370.0
1666.857.848.437.670.0
IDEFICS (80B)0*60.045.236.030.960.6
463.652.440.434.457.8
864.855.146.135.758.2
1665.456.848.336.357.8
Emu (14B)0*52.942.834.4--
458.4-41.3--
859.0-43.9--
16-----
Emu2 (37B)033.526.740.426.452.2
467.053.254.648.262.4
867.854.154.749.365.8
1668.857.157.050.366.0
+ +Table 1. Zero-shot and few-shot evaluations of Emu2. $0^{*}$ denotes text two-shot and image zero-shot results following Flamingo [5]. The best results are in bold and the second best are underlined. + +Training Objective. We use the same unified generative pretraining objective to adapt to diverse generation tasks in context. Specifically, a training sample for generation is formulated as: “A photo of

a man

image embedding of object localization image[IMG] image embedding of man[/IMG] sitting next to

a dog

image embedding of object localization image[IMG] image embedding of dog[/IMG][IMG] image embedding of the whole image[/IMG”. We represent the coordinates of each object directly in image form by drawing the bounding box of each object at its specified location on a black image. Our Emu2-Gen conducts unified multimodal modeling of the text, object image, and corresponding object localization image. The regression loss only applies to the visual embeddings of the last image. We freeze the Visual Encoder during fine-tuning. We randomly drop tokens of entities and object localization image to enhance model adaptability and robustness. Additionally, we apply data augmentation to each object image, incorporating random background variations and random crop, aiming to reduce the reliance on image backgrounds. Training Setup. We use a global batch size of 4,096 and + +
ModelVisual Question AnswerLMM Benchmarks
VQAv2 [28]OKVQA [53]GQA [32]VizWiz [29]TextVQA [68]MSVD [82]MSRVTT [82]SEED [39]MM-Vet [88]TS [7]MMMU [89]
Flamingo-9B [5]51.844.7-28.8-30.213.7----
Flamingo-80B [5]56.350.6-31.6-35.617.4----
Kosmos-1 [31]51.0--29.2-------
Kosmos-2 [57]51.1------50.0--26.6
BLIP-2-13B [42]--41.019.642.520.310.346.422.4--
InstructBLIP-13B [22]--49.533.450.741.224.8-25.6552.4-
IDEFICS-9B [38]50.938.4-35.525.9------
IDEFICS-80B [38]60.045.2-36.030.9------
Shikra-13B [15]77.4*47.2---------
Qwen-VL-13B-Chat [6]78.2*56.6*57.5*38.961.5*--58.2-645.2-
LLaVA-1.5-13B [48]80.0*-63.3*53.661.3--61.635.4-33.6
CogVLM [77]83.4*58.9*--68.1*----662.630.1
Emu-I [71]62.049.246.038.3-37.021.2-36.3--
Emu2-Chat84.9*64.8*65.1*54.966.6*49.031.462.848.5703.834.1
+ +Table 2. Results on visual question answering and LMM benchmarks. * indicates that samples from this task's training set have been trained. SEED and TS respectively represent SEED-Bench [39] and TouchStone [7]. For MM-Vet, we present the average result of five scoring runs. + +train for 3k steps. The learning rate linearly warms up to $5 \times 10^{-5}$ in the first 100 steps, then decays to zero with a cosine schedule. We further fine-tune for 900 steps using the 500k high-quality pairs with a batch size of 2048. + +# 3. Evaluation + +# 3.1. Pretrained Base Model + +We evaluate zero-shot and few-shot abilities of Emu2 on OKVQA [53], VQAv2 [28], VizWiz [29], TextVQA [68], and HatefulMemes [35] tasks. Details of the datasets and prompts can be found in supplementary materials. The results are presented in Table 1. Emu2 demonstrates remarkable in-context ability, showcasing improved performance with more in-context samples seen. Specifically, on VQAv2, VizWiz and TextVQA datasets, Emu2 outperforms Flamingo-80B and IDEFICS-80B under all few-shot settings with a much smaller model scale (37B). + +Figure 1 demonstrates Emu2's few-shot capabilities in the wild. For example, the model learns to classify and count simultaneously in a specific format via a few examples (row 1). Additionally, Emu2 is capable of following visual prompts in context, e.g., the red circles laid on the images (row 2 and 3). + +# 3.2. Instruction-Following Chat + +Our Emu2-Chat is evaluated on academic-task-oriented benchmarks including image question-answering datasets (VQAv2 [28], OKVQA [53], GQA [32], VizWiz [29], TextVQA [68]) and video question-answering datasets (MSVD [82] and MSRvTT [82]). The evaluation also encompassed recent benchmarks for large multimodal models, + +including SEED-Bench [39], MM-Vet [88], TouchStone [7] and MMMU [89]. When evaluated on SEED-Bench, we followed the setup of LLaVa-1.5 [48] by presenting options to the model for completing multiple-choice tasks. + +As shown in Table 2, Emu2-Chat consistently outperforms other models in image question-answering tasks, encompassing well-established benchmarks like VQAv2 and GQA. Notably, it shows a noticeable improvement in the OKVQA task, which requires the utilization of external knowledge, showcasing the advantage of our model for mastering real-world knowledge. For video question-answering, Emu2-Chat demonstrated advantages even though it did not use video question-answering data for training. It achieved an accuracy of 49.0 and 31.4 on the MSVD-QA and MSRVTT-QA tasks, respectively, surpassing InstructBLIP and the larger Flamingo-80B. More importantly, our model has also achieved better results on LMM benchmarks. LMM benchmarks such as MM-Vet provide a more comprehensive evaluation of model abilities, including solving complicated tasks. Emu2-Chat achieves a score of 48.5 in MM-Vet and 703.8 in Touch-Stone, confirming its superior capability in understanding and solving multimodal problems. + +In addition, we demonstrated the visual grounding capability of our model using the refer expression comprehension benchmarks. In Table 3, Emu2-Chat achieved the best results among generalist models on RefCOCO [34], RefCOCO+ [52] and RefCOCOg [52]. Its most notable advantage was observed in RefCOCO+, which focused solely on purely appearance-based descriptions without allowing the use of position references. This highlights our model's powerful perceptual abilities in capturing intricate details. + +![](images/8cc37d98935c4f44559647e055e5f0294afab6627bad7cb4697c07728efeac58.jpg) + +![](images/f49789c75d3cf1d5b93bf13793e3a023eb895543a8af4f2a5b2930fc81ddbdfd.jpg) + +![](images/c243d98ec61ed32f69243e80a599b8a9c5fb00f121a214bf3369082167fad9d3.jpg) + +![](images/e35517afe733d538ccf4abdc92ef411ca9a0e99a964baaafae17d14b8d268c6c.jpg) +A photo of the first dog>, the second dog>, the third dog on the grass + +![](images/cf05a40c61a0659c785072d14443bdd2adf475adde31682340a08b7a756a1698.jpg) +An oil painting of the first dog>, the second dog>, the third dog> + +![](images/3d8dbce32e38d79ef07ae1922ac658afe4324c4f9d5c27275e43f0118f0e2846.jpg) +A photo of the first dog>, the second dog>, the third dog> on the beach + +![](images/01eeb3d932b34db42a4a83ad3bf541fe2cc358e560872b8eb9e0f9b01c743ac2.jpg) +as an oil painting by Monet + +![](images/005d7167506e27504e53f460afd17bcf90099b1c73eebf5344af922093879ae2.jpg) + +![](images/0cd54a6d7deaf963723f17b3fb76825c5236b86ffa07e4d515fa1518840ce7dc.jpg) +in a dark forest in 3D + +![](images/9e567b0a83330b66bae6b4cc1437fded089295b75a239fef9e4f5c54cbea2b24.jpg) + +![](images/4a23fa92e42fe9417d7aabb40a922353a9b7b3f04942f0384346061cb73e86b1.jpg) +a bear and a sunflower + +![](images/f3224340123304d8763ef9a5dd124aa5ebf91c211ecbfc6dff103f8c9847ba64.jpg) + +![](images/e6bacea7baf4fbd6906a22b63e90333636fda2da2df53b1470d32aaa70c78242.jpg) +Figure 4. Visualization of Emu2-Gen's controllable generation capability. The model is capable of accepting a mix of text, locations and images as input, and generating images in context. The presented examples include text- and subject-grounded generation, stylization, multi-entity composition, subject-driven editing, and text-to-image generation. + +![](images/9a15867f22c2fc32224497c3d97c80b1999a30acde3f37a3d56c88e747a209e7.jpg) + +![](images/46c7465f93b0dc2005ac0d6f2fc237cad65f7b6e4b397ea06add7a4379628794.jpg) + +![](images/5c79ae0cdef25fe4e2b8769659255fdca36387d774d3a21bb355df5e16fab49e.jpg) + +![](images/3dc820fd695e8c875e65609b25c378a1b990c43c26ed2812ac2427b1855f56a3.jpg) + +![](images/31ee407d139f4e9740f54ce47a40de1cd41b635024c73154551387b368021693.jpg) + +![](images/d77f6c2bc5f9b120d3614cba07753f7c2f472f93935ba6a3ee954a5beba1054e.jpg) + +
ModelRefCOCORefCOCO+RefCOCOg
valtestAtestBvaltestAtestBvaltest
OFA-L [75]79.9683.6776.3968.2976.0061.7567.5767.58
Shikra-7B [15]87.0190.6180.2481.6087.3672.1282.2782.19
Shikra-13B [15]87.8391.1181.8182.8987.7974.4182.6483.16
Qwen-VL-7B [6]89.3692.2685.3483.1288.2577.2185.5885.48
Emu2-Chat90.4093.8885.9787.0591.4380.4787.6488.11
CogVLM [77]92.5193.9588.7387.5291.8181.4389.4690.09
+ +Table 3. Results on referring expression comprehension. We grayed out CogVLM because its generalist grounding-enhanced model was specialist trained on high-quality grounding data. + +# 3.3. Controllable Visual Generation + +Qualitative Results. Figure 3 presents a visualization of Emu2's autoencoding results. With Emu2's Visual Encoder and Visual Decoder, we can tokenize an image into visual embeddings and detokenize them back. Compared with SEED [27] and Emu [71], Emu2 shows significantly superior results. We also evaluate our image autoencoding results on MS-COCO [47] and achieve a strong 0.907 CLIP-I [59] score. More results are in the supplementary. + +As depicted in Figure 4, Emu2-Gen is capable of accepting a mixture of text, locations and images as input, and generating images in context. The model skillfully engages in various controllable visual generation tasks in + +a zero-shot setting, capitalizing on the in-context learning capabilities in multimodality. Examples in Figure 4 show generated images of three dogs conditioned on different subjects, locations and scenarios. The presented visual samples demonstrate the model's proficiency in tasks such as re-contextualization, stylization, modification, region-controllable generation, and multi-entity composition. + +Zero-shot Text-to-image Generation. We evaluate the zero-shot text-to-image generation capability on 30k randomly sampled data from the MS-COCO [47] validation set. We employ CLIP-ViT-B [60], following the approach in DALL-E 3[9], to calculate the CLIP-T score to assess prompt-following ability. Additionally, we utilize CLIP-ViT-L, as in GILL[37], to compute the CLIP-I score for measuring image similarity. A higher score means the generated image is more similar to the prompt or the real image. Table 4 shows that Emu2-Gen achieves the state-of-the-art performance in terms of both CLIP-I and CLIP-T scores compared to various unimodal generation models and multimodal models. More text-to-image generation cases can be found in supplementary. + +Zero-shot Subject-driven Generation. Following Kosmos-G [56], we also evaluate our model's subject-driven image editing ability on DreamBench [63]. We generate four images for each prompt, resulting in a total of 3,000 images for a comprehensive evaluation. We employ + +
ModelsCLIP-I ↑CLIP-T ↑
unimodal generation models
MUSE [13]-0.320
Imagen [64]-0.270
DALL-E 2 † [61]-0.314
DALL-E 3 † [9]-0.320
SDv1.5 [62]0.6670.302
SDXL [58]0.6740.310
multimodal generation models
GILL [37]0.684-
SEED [27]0.682-
Emu [71]0.6560.286
Emu2-Gen0.6860.297
+ +Table 4. Quantitative comparison of zero-shot text-to-image generation on MS-COCO [47] validation set. 30k samples are randomly sampled. $\dagger$ CLIP-T score is calculated on 4,096 samples. We also evaluate our image autoencoding results on MS-COCO which achieves a strong 0.907 CLIP-I score. + +
MethodsDINO ↑CLIP-I ↑CLIP-T ↑
Real Images (Oracle)0.7740.885-
Fine-Tuning
Textual Inversion [24]0.5690.7800.255
DreamBooth [63]0.6680.8030.305
BLIP-Diffusion [42]0.6700.8050.302
Test Time Tuning Free
Re-Imagen* [16]0.6000.7400.270
SuTI [17]0.7410.8190.304
BLIP-Diffusion* [42]0.5940.7790.300
Kosmos-G* (single image input)0.6940.8470.287
Emu2-Gen * (single image input)0.7660.8500.287
+ +Table 5. Quantitative comparison of zero-shot single-entity subject-driven generation on DreamBench. * denotes zero-shot methods. + +DINO [12] and CLIP-I [59] to evaluate subject fidelity, and CLIP-T [59] to evaluate text fidelity, aligning with the methodology established by DreamBooth. Notably, Emu2-Gen excels in subject fidelity, as evidenced by its superior performance on DINO and CLIP-I metrics compared to methods like BLIP-Diffusion and Kosmos-G. Emu2-Gen impressively reconstructs subjects with just one image input in zero-shot setting, demonstrating superior subject fidelity through powerful visual decoding. Further illustrative cases are provided in the supplementary, showcasing Emu2-Gen's proficiency in multi-entity generation. + +# 4. Related Work + +Large Multimodal Models. Recent years have witnessed the rapid growth of large multimodal models [5, 19, 31, + +71]. CLIP [59] pioneered the learning of LMMs with a contrastive learning objective on massive image-text pair data. Flamingo [5] and Kosmos [31, 57] exhibit promising zero-shot and few-shot multi-modal understanding performance by training on large-scale image-text interleaved data. With the remarkable progress in open-sourced LLMs, [42, 43, 49, 94] show promising results by connecting vision encoders and LLMs with a small intermediate model. A school of successive efforts [69, 76, 84, 90, 91] further improves visual instruction tuning with better overall training pipelines [6, 40], grounding annotations [14, 15, 85, 92], and extra tasks [6]. There are early studies on training more unified large multimodal models [23, 27, 71, 86] that are capable of performing visual understanding and generation simultaneously. In this paper, we further explore the distinct solution proposed in Emu [71]: learning large multimodal models with generative objectives on both texts and images. + +In-Context Learning. Recent advancements in large language models [11, 21] underscore their capacity for in-context learning [11]. This phenomenon, particularly evident as LLMs scale up in size and data, has been exploited for complex challenges such as mathematical reasoning [81], signaling new emergent ability in model behavior [80]. Flamingo [5] integrates visual inputs to LLMs, enabling the in-context learning of visual-linguistic tasks such as image captioning and OCR through language-based interfacing. Painter [78] and SegGPT [79] conduct an early study of visual in-context learning. Inspired by the emerging abilities of large language models, in this work we study the problem of multimodal in-context learning by scaling up generative multimodal models and demonstrating strong results in broad understanding and generation tasks. + +# 5. Conclusion + +We present a 37 billion-parameter generative multimodal model Emu2 that shows strong performance and versatility on many multimodal tasks in the in-context settings. Emu2 serves as a base model and a general-purpose interface for a variety of multimodal tasks. We demonstrate state-of-the-art results on a broad range of benchmarks of multimodal understanding and generation. Specifically, our model largely surpasses prior work on the lately proposed LMM benchmarks that require more advanced capability compared to classic academic benchmarks. Emu2 also shows remarkable capability of controllable visual generation in multimodal context, e.g., subject-/text-grounded generation. Additionally, we review the limitations and broader social impact of Emu2. Despite discussed weaknesses, these results suggest that generative multimodal model at scale may be an important step towards the development of adaptable, general multimodal systems. + +# References + +[1] Laion-aesthetics. https://laion.ai/blog/laion-aesthetics/, .4,5,2,3 +[2] Laion coco: 600m synthetic captions from laion2b-en. https://laion.ai/blog/laion-coco/,4,2 +[3] Laion-high-resolution. https://huggingface.co/datasets/laion/laion-high-resolution, 5, 3 +[4] Sharegpt. https://sharegpt.com/.5,2 +[5] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 1, 5, 6, 8, 3 +[6] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 6, 7, 8 +[7] Shuai Bai, Shusheng Yang, Jinze Bai, Peng Wang, Xingxuan Zhang, Junyang Lin, Xinggang Wang, Chang Zhou, and Jingren Zhou. Touchstone: Evaluating vision-language models by language models. arXiv preprint arXiv:2308.16890, 2023. 6 +[8] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1728-1738, 2021. 3, 1 +[9] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, Wesam Manassra, Prafulla Dhariwal, Casey Chu, Yunxin Jiao, and Aditya Ramesh. Improving image generation with better captions. 2023. 5, 7, 8, 3 +[10] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 5, 3 +[11] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1, 8 +[12] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 8 +[13] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. 8 +[14] Chi Chen, Ruoyu Qin, Fuwen Luo, Xiaoyue Mi, Peng Li, Maosong Sun, and Yang Liu. Position-enhanced visual + +instruction tuning for multimodal large language models. arXiv preprint arXiv:2308.13437, 2023. 8 +[15] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 6, 7, 8 +[16] Wenhu Chen, Hexiang Hu, Chitwan Sahara, and William W Cohen. Re-imagen: Retrieval-augmented text-to-image generator. arXiv preprint arXiv:2209.14491, 2022. 8 +[17] Wenhu Chen, Hexiang Hu, Yandong Li, Nataniel Rui, Xuhui Jia, Ming-Wei Chang, and William W Cohen. Subject-driven text-to-image generation via apprenticeship learning. arXiv preprint arXiv:2304.00186, 2023. 8 +[18] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 4, 2 +[19] Xi Chen, Xiao Wang, Lucas Beyer, Alexander Kolesnikov, Jialin Wu, Paul Voigtlaender, Basil Mustafa, Sebastian Goodman, Ibrahim Alabdulmohsin, Piotr Padlewski, et al. Pali-3 vision language models: Smaller, faster, stronger. arXiv preprint arXiv:2310.09199, 2023. 1, 8 +[20] Luke Chesser and Timothy Carbone. Unsplash. https://github.com/unsplash/datasets.5,3 +[21] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 1, 8 +[22] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Albert Li, Pascale Fung, and Steven C. H. Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500, 2023. 6 +[23] Runpei Dong, Chunrui Han, Yuang Peng, Zekun Qi, Zheng Ge, Jinrong Yang, Liang Zhao, Jianjian Sun, Hongyu Zhou, Haoran Wei, Xiangwen Kong, Xiangyu Zhang, Kaisheng Ma, and Li Yi. Dreamllm: Synergistic multimodal comprehension and creation. arXiv preprint arXiv:2309.11499, 2023.8 +[24] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 8 +[25] Zhe Gan, Linjie Li, Chunyuan Li, Lijuan Wang, Zicheng Liu, Jianfeng Gao, et al. Vision-language pre-training: Basics, recent advances, and future trends. Foundations and Trends® in Computer Graphics and Vision, 14(3-4):163-352, 2022. 1 +[26] Leo Gao, Stella Biderman, Sid Black, Laurence Golding, Travis Hoppe, Charles Foster, Jason Phang, Horace He, Anish Thite, Noa Nabeshima, Shawn Presser, and Connor Leahy. The Pile: An 800gb dataset of diverse text for language modeling. arXiv preprint arXiv:2101.00027, 2020. 3, 2 + +[27] Yuying Ge, Yixiao Ge, Ziyun Zeng, Xintao Wang, and Ying Shan. Planting a seed of vision in large language model. arXiv preprint arXiv:2307.08041, 2023. 4, 7, 8 +[28] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 4, 6, 2 +[29] Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. Vizwiz grand challenge: Answering visual questions from blind people. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3608-3617, 2018. 6 +[30] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 4 +[31] Shaohan Huang, Li Dong, Wenhui Wang, Yaru Hao, Saksham Singhal, Shuming Ma, Tengchao Lv, Lei Cui, Owais Khan Mohammed, Qiang Liu, et al. Language is not all you need: Aligning perception with language models. arXiv preprint arXiv:2302.14045, 2023. 6, 8 +[32] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 4, 6, 2 +[33] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in Neural Information Processing Systems, 35:26565-26577, 2022. 3 +[34] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014. 5, 6, 2 +[35] Douwe Kiela, Hamed Firooz, Aravind Mohan, Vedanuj Goswami, Amanpreet Singh, Pratik Ringshia, and Davide Testuggine. The hateful memes challenge: Detecting hate speech in multimodal memes. Advances in neural information processing systems, 33:2611-2624, 2020. 6 +[36] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 5, 3 +[37] Jing Yu Koh, Daniel Fried, and Ruslan Salakhutdinov. Generating images with multimodal language models. arXiv preprint arXiv:2305.17216, 2023. 7, 8 +[38] Hugo Laurençon, Lucile Saulnier, Léo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, Thomas Wang, Siddharth Karamcheti, Alexander M Rush, Douwe Kiela, et al. Obelics: An open web-scale filtered dataset of interleaved image-text documents. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 6 +[39] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking mul + +timodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 6 +[40] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726, 2023.8 +[41] Chunyuan Li, Zhe Gan, Zhengyuan Yang, Jianwei Yang, Linjie Li, Lijuan Wang, and Jianfeng Gao. Multimodal foundation models: From specialists to general-purpose assistants. arXiv preprint arXiv:2309.10020, 1(2):2, 2023. 1 +[42] Dongxu Li, Junnan Li, and Steven CH Hoi. Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. arXiv preprint arXiv:2305.14720, 2023. 6, 8 +[43] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 8 +[44] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 5, 2 +[45] Lei Li, Yuwei Yin, Shicheng Li, Liang Chen, Peiyi Wang, Shuhuai Ren, Mukai Li, Yazheng Yang, Jingjing Xu, Xu Sun, et al. M3it: A large-scale dataset towards multimodal multilingual instruction tuning. arXiv preprint arXiv:2306.04387, 2023. 4, 2 +[46] Xin Li, Wenqing Chu, Ye Wu, Weihang Yuan, Fanglong Liu, Qi Zhang, Fu Li, Haocheng Feng, Errui Ding, and Jingdong Wang. Videogen: A reference-guided latent diffusion approach for high definition text-to-video generation. arXiv preprint arXiv:2309.00398, 2023. 4 +[47] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 7, 8 +[48] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 4, 6, 3 +[49] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 5, 8, 2 +[50] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 3, 4 +[51] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems, 35:2507–2521, 2022. 4 +[52] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 5, 6, 2 + +[53] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 3195-3204, 2019. 4, 6, 2 +[54] Midjourney. Midjourney. https://www.midjourney.com. 5, 3 +[55] Leonardo Nicoletti and Dina Bass. Humans are biased: Generative ai is even worse. *Bloomberg Technology+ Equality*. Accessed June, 23:2023, 2023. 1 +[56] Xichen Pan, Li Dong, Shaohan Huang, Zhiliang Peng, Wenhu Chen, and Furu Wei. Kosmos-g: Generating images in context with multimodal large language models. arXiv preprint arXiv:2310.02992, 2023. 7, 4 +[57] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 3, 5, 6, 8, 1, 2 +[58] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023. 1, 3, 4, 8, 2 +[59] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7, 8 +[60] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7 +[61] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022.8 +[62] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 4, 8 +[63] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 7, 8, 4 +[64] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 8 +[65] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo + +Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. arXiv preprint arXiv:2210.08402, 2022. 3, 1 +[66] Oleksii Sidorov, Ronghang Hu, Marcus Rohrbach, and Amanpreet Singh. Textcaps: a dataset for image captioning with reading comprehension. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 742-758. Springer, 2020. 4, 2 +[67] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 4 +[68] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 4, 6, 2 +[69] Yixuan Su, Tian Lan, Huayang Li, Jialu Xu, Yan Wang, and Deng Cai. Pandagpt: One model to instruction-follow them all. arXiv preprint arXiv:2305.16355, 2023. 8 +[70] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389, 2023. 3 +[71] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative pretraining in multimodality. arXiv preprint arXiv:2307.05222, 2023. 3, 4, 6, 7, 8, 1 +[72] Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B Hashimoto. Stanford alpaca: An instruction-following llama model, 2023. 5, 2 +[73] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 3 +[74] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 4 +[75] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In International Conference on Machine Learning, pages 23318-23340. PMLR, 2022. 7 +[76] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, et al. Vision: Large language model is also an open-ended decoder for vision-centric tasks. arXiv preprint arXiv:2305.11175, 2023. 8 +[77] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan + +Song, et al. Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079, 2023. 6, 7 +[78] Xinlong Wang, Wen Wang, Yue Cao, Chunhua Shen, and Tiejun Huang. Images speak in images: A generalist painter for in-context visual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6830-6839, 2023. 8 +[79] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Segmenting everything in context. arXiv preprint arXiv:2304.03284, 2023. 8 +[80] Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, et al. Emergent abilities of large language models. arXiv preprint arXiv:2206.07682, 2022. 8 +[81] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 35:24824-24837, 2022. 8 +[82] Dejing Xu, Zhou Zhao, Jun Xiao, Fei Wu, Hanwang Zhang, Xiangnan He, and Yueting Zhuang. Video question answering via gradually refined attention over appearance and motion. In Proceedings of the 25th ACM international conference on Multimedia, pages 1645-1653, 2017. 6 +[83] Charig Yang, Weidi Xie, and Andrew Zisserman. It's about time: Analog clock reading in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2508-2517, 2022. 5, 2 +[84] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. MPLug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 8 +[85] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704, 2023. 8 +[86] Lili Yu, Bowen Shi, Ramakanth Pasunuru, Benjamin Muller, Olga Golovneva, Tianlu Wang, Arun Babu, Binh Tang, Brian Karrer, Shelly Sheynin, et al. Scaling autoregressive multimodal models: Pretraining and instruction tuning. arXiv preprint arXiv:2309.02591, 2023. 8 +[87] Qiying Yu, Quan Sun, Xiaosong Zhang, Yufeng Cui, Fan Zhang, Xinlong Wang, and Jingjing Liu. Capsfusion: Rethinking image-text data at scale. arXiv preprint arXiv:2310.20550, 2023. 3, 5, 1 +[88] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490, 2023. 6 +[89] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. arXiv preprint arXiv:2311.16502, 2023. 6 + +[90] Ao Zhang, Hao Fei, Yuan Yao, Wei Ji, Li Li, Zhiyuan Liu, and Tat-Seng Chua. Transfer visual prompt generator across llms. arXiv preprint arXiv:2305.01278, 2023. 8 +[91] Pan Zhang, Xiaoyi Dong Bin Wang, Yuhang Cao, Chao Xu, Linke Ouyang, Zhiyuan Zhao, Shuangrui Ding, Songyang Zhang, Haodong Duan, Hang Yan, et al. Internlmxcomposer: A vision-language large model for advanced text-image comprehension and composition. arXiv preprint arXiv:2309.15112, 2023. 8 +[92] Shilong Zhang, Peize Sun, Shoufa Chen, Min Xiao, Wenqi Shao, Wenwei Zhang, Kai Chen, and Ping Luo. Gpt4roi: Instruction tuning large language model on region-of-interest. arXiv preprint arXiv:2307.03601, 2023. 8 +[93] Yanzhe Zhang, Ruiyi Zhang, Jiuxiang Gu, Yufan Zhou, Nedim Lipka, Diyi Yang, and Tong Sun. Llavar: Enhanced visual instruction tuning for text-rich image understanding. arXiv preprint arXiv:2306.17107, 2023. 5, 2 +[94] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 8 +[95] Wanrong Zhu, Jack Hessel, Anas Awadalla, Samir Yitzhak Gadre, Jesse Dodge, Alex Fang, Youngjae Yu, Ludwig Schmidt, William Yang Wang, and Yejin Choi. Multimodal c4: An open, billion-scale corpus of images interleaved with text. arXiv preprint arXiv:2304.06939, 2023. 3, 1 \ No newline at end of file diff --git a/2024/Generative Multimodal Models are In-Context Learners/images.zip b/2024/Generative Multimodal Models are In-Context Learners/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..46e2b4b78cd16f2e66cb3bb76a19e93f9bd73d21 --- /dev/null +++ b/2024/Generative Multimodal Models are In-Context Learners/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6826f1db3ba7e64cff5df7ae0ac4894ab906a50087c26e1f5ca2a51384e197f4 +size 771225 diff --git a/2024/Generative Multimodal Models are In-Context Learners/layout.json b/2024/Generative Multimodal Models are In-Context Learners/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4733409c23d1010df59001d78f2b1b700ba1c7be --- /dev/null +++ b/2024/Generative Multimodal Models are In-Context Learners/layout.json @@ -0,0 +1,12185 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 122, + 103, + 473, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 103, + 473, + 121 + ], + "spans": [ + { + "bbox": [ + 122, + 103, + 473, + 121 + ], + "type": "text", + "content": "Generative Multimodal Models are In-Context Learners" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 52, + 141, + 541, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 141, + 541, + 173 + ], + "spans": [ + { + "bbox": [ + 52, + 141, + 541, + 173 + ], + "type": "text", + "content": "Quan Sun" + }, + { + "bbox": [ + 52, + 141, + 541, + 173 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 52, + 141, + 541, + 173 + ], + "type": "text", + "content": " Yufeng Cui" + }, + { + "bbox": [ + 52, + 141, + 541, + 173 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 52, + 141, + 541, + 173 + ], + "type": "text", + "content": " Xiaosong Zhang" + }, + { + "bbox": [ + 52, + 141, + 541, + 173 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 52, + 141, + 541, + 173 + ], + "type": "text", + "content": " Fan Zhang" + }, + { + "bbox": [ + 52, + 141, + 541, + 173 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 52, + 141, + 541, + 173 + ], + "type": "text", + "content": " Qiying Yu" + }, + { + "bbox": [ + 52, + 141, + 541, + 173 + ], + "type": "inline_equation", + "content": "^{2,1*}" + }, + { + "bbox": [ + 52, + 141, + 541, + 173 + ], + "type": "text", + "content": " Yueze Wang" + }, + { + "bbox": [ + 52, + 141, + 541, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 52, + 141, + 541, + 173 + ], + "type": "text", + "content": " Yongming Rao" + }, + { + "bbox": [ + 52, + 141, + 541, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 52, + 141, + 541, + 173 + ], + "type": "text", + "content": " Jingjing Liu" + }, + { + "bbox": [ + 52, + 141, + 541, + 173 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 52, + 141, + 541, + 173 + ], + "type": "text", + "content": " Tiejun Huang" + }, + { + "bbox": [ + 52, + 141, + 541, + 173 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 52, + 141, + 541, + 173 + ], + "type": "text", + "content": " Xinlong Wang" + }, + { + "bbox": [ + 52, + 141, + 541, + 173 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 99, + 177, + 500, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 177, + 500, + 209 + ], + "spans": [ + { + "bbox": [ + 99, + 177, + 500, + 209 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 99, + 177, + 500, + 209 + ], + "type": "text", + "content": " Beijing Academy of Artificial Intelligence " + }, + { + "bbox": [ + 99, + 177, + 500, + 209 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 99, + 177, + 500, + 209 + ], + "type": "text", + "content": " Tsinghua University " + }, + { + "bbox": [ + 99, + 177, + 500, + 209 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 99, + 177, + 500, + 209 + ], + "type": "text", + "content": " Peking University *equal contribution †project lead" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 170, + 216, + 422, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 216, + 422, + 228 + ], + "spans": [ + { + "bbox": [ + 170, + 216, + 422, + 228 + ], + "type": "text", + "content": "code & models: https://github.com/baaivision/Emu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "spans": [ + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 280, + 290, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 280, + 290, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 280, + 290, + 544 + ], + "type": "text", + "content": "The human ability to easily solve multimodal tasks in context (i.e., with only a few demonstrations or simple instructions), is what current multimodal systems have largely struggled to imitate. In this work, we demonstrate that the task-agnostic in-context learning capabilities of large multimodal models can be significantly enhanced by effective scaling-up. We introduce Emu2, a generative multimodal model with 37 billion parameters, trained on large-scale multimodal sequences with a unified autoregressive objective. Emu2 exhibits strong multimodal in-context learning abilities, even emerging to solve tasks that require on-the-fly reasoning, such as visual prompting and object-grounded generation. The model sets a new record on multiple multimodal understanding tasks in few-shot settings. When instruction-tuned to follow specific instructions, Emu2 further achieves new state-of-the-art on challenging tasks such as question answering benchmarks for large multimodal models and open-ended subject-driven generation. These achievements demonstrate that Emu2 can serve as a base model and general-purpose interface for a wide range of multimodal tasks. Code and models are publicly available to facilitate future research." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 567, + 128, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 567, + 128, + 578 + ], + "spans": [ + { + "bbox": [ + 47, + 567, + 128, + 578 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 586, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 586, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 586, + 287, + 696 + ], + "type": "text", + "content": "Multimodal tasks [25, 41] encompass anything involving understanding and generation in single or multiple modalities [5, 19, 58], which can be highly diverse and long-tail. Previous multimodal systems largely rely on designing task-specific architecture and collecting a sizable supervised training set, both of which are difficult to scale, particularly when this process needs to be repeated for each new task encountered. By contrast, humans can solve a new task in context, i.e., with only a few demonstrations or simple" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 256, + 545, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 256, + 545, + 280 + ], + "spans": [ + { + "bbox": [ + 306, + 256, + 545, + 280 + ], + "type": "text", + "content": "instructions - a capability that current multimodal models have yet to learn." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 281, + 546, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 546, + 436 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 546, + 436 + ], + "type": "text", + "content": "Recently, generative pretrained language models have demonstrated strong in-context learning abilities [11, 21, 73]. By training a 37-billion-parameter model Emu2 and thoroughly evaluating it on diverse multimodal tasks, we demonstrate that a scaled-up multimodal generative pretrained model can harness similar in-context learning abilities and effectively generalize to unseen multimodal tasks. Emu2 is trained with a unified autoregressive objective: predict-the next-multimodal-element (either visual embeddings or textual tokens). In this unified generative pretraining process, large-scale multimodal sequences (e.g., text, image-text pairs, and interleaved image-text-video) are used for model training." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 437, + 547, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 547, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 547, + 521 + ], + "type": "text", + "content": "We measure Emu2's capabilities of learning from a few examples or instructions on standard multimodal datasets, as well as new tasks unseen in the training set. Specifically, Emu2 is evaluated under two scenarios: (a) few-shot setting, where we allow as many examples as possible to fit the context window of the model; and (b) instruction tuning, where the model is tuned to follow specific instructions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 521, + 547, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 547, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 547, + 652 + ], + "type": "text", + "content": "Emu2 achieves promising results in the few-shot setting on a wide range of vision-language tasks. For example, it demonstrates state-of-the-art few-shot performance on multiple visual question-answering datasets. We observe a performance improvement when the number of examples in context increases. Figure 1 illustrates Emu2's strong multimodal reasoning capabilities for tasks in the wild, e.g., recognition and counting in a specific format. Emu2 also learns to follow visual prompting in context (e.g., the circles laid on the images in Figure 1), even although it struggles at a smaller scale or at zero shot." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "As Emu2 is inherently equipped to handle interleaved text-image-video at both input and output, it serves as a powerful and versatile base model for diverse multimodal tasks, by following specific task instructions. For example, after instruct tuning with conversational data," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 208, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 208, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 208, + 713 + ], + "type": "text", + "content": "†Correspondence to wangxinlong@baai.ac.cn" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "14398" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 91, + 127, + 127 + ], + "blocks": [ + { + "bbox": [ + 77, + 91, + 127, + 127 + ], + "lines": [ + { + "bbox": [ + 77, + 91, + 127, + 127 + ], + "spans": [ + { + "bbox": [ + 77, + 91, + 127, + 127 + ], + "type": "image", + "image_path": "4d1282fe13868ac44b6e10da23a599442d29205b9894a5f26b5835062e24bf21.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 104, + 179, + 113 + ], + "lines": [ + { + "bbox": [ + 132, + 104, + 179, + 113 + ], + "spans": [ + { + "bbox": [ + 132, + 104, + 179, + 113 + ], + "type": "text", + "content": "[dog: 1, frisbee: 1]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 184, + 91, + 208, + 125 + ], + "blocks": [ + { + "bbox": [ + 184, + 91, + 208, + 125 + ], + "lines": [ + { + "bbox": [ + 184, + 91, + 208, + 125 + ], + "spans": [ + { + "bbox": [ + 184, + 91, + 208, + 125 + ], + "type": "image", + "image_path": "4c2ef205e12e6ea72513547bb6aa5691fd2aad11fc7cfce9ec1db0d23f8af9e6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 214, + 104, + 286, + 113 + ], + "lines": [ + { + "bbox": [ + 214, + 104, + 286, + 113 + ], + "spans": [ + { + "bbox": [ + 214, + 104, + 286, + 113 + ], + "type": "text", + "content": "[burger: 1, glass: 1, bottle: 1]." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 296, + 95, + 335, + 124 + ], + "blocks": [ + { + "bbox": [ + 296, + 95, + 335, + 124 + ], + "lines": [ + { + "bbox": [ + 296, + 95, + 335, + 124 + ], + "spans": [ + { + "bbox": [ + 296, + 95, + 335, + 124 + ], + "type": "image", + "image_path": "e18a18357157aafcc0876e4d567730bb90498ce80ddb55e2f659af8654967ba8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 393, + 91, + 441, + 162 + ], + "blocks": [ + { + "bbox": [ + 356, + 106, + 376, + 113 + ], + "lines": [ + { + "bbox": [ + 356, + 106, + 376, + 113 + ], + "spans": [ + { + "bbox": [ + 356, + 106, + 376, + 113 + ], + "type": "text", + "content": "[cat: 3]." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 393, + 91, + 441, + 162 + ], + "lines": [ + { + "bbox": [ + 393, + 91, + 441, + 162 + ], + "spans": [ + { + "bbox": [ + 393, + 91, + 441, + 162 + ], + "type": "image", + "image_path": "bc74e1252b4635fccb7ae97a3523505ff7f2a16ca917c7d301d0cce46a8dfb95.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 137, + 388, + 152 + ], + "lines": [ + { + "bbox": [ + 343, + 137, + 388, + 152 + ], + "spans": [ + { + "bbox": [ + 343, + 137, + 388, + 152 + ], + "type": "text", + "content": "The text in the red circle: Do Not?" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 471, + 78, + 502, + 87 + ], + "lines": [ + { + "bbox": [ + 471, + 78, + 502, + 87 + ], + "spans": [ + { + "bbox": [ + 471, + 78, + 502, + 87 + ], + "type": "text", + "content": "Completion" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 462, + 105, + 511, + 113 + ], + "lines": [ + { + "bbox": [ + 462, + 105, + 511, + 113 + ], + "spans": [ + { + "bbox": [ + 462, + 105, + 511, + 113 + ], + "type": "text", + "content": "[beer: 3, banana: 2]." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 446, + 140, + 526, + 148 + ], + "lines": [ + { + "bbox": [ + 446, + 140, + 526, + 148 + ], + "spans": [ + { + "bbox": [ + 446, + 140, + 526, + 148 + ], + "type": "text", + "content": "The text in the red circle: 'Lynn'." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 394, + 163, + 441, + 196 + ], + "blocks": [ + { + "bbox": [ + 394, + 163, + 441, + 196 + ], + "lines": [ + { + "bbox": [ + 394, + 163, + 441, + 196 + ], + "spans": [ + { + "bbox": [ + 394, + 163, + 441, + 196 + ], + "type": "image", + "image_path": "8f2304618c0af1da4f98ed5456c6033dde1bf907a936c8b9e8bbaadc877042d2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 182, + 127, + 231, + 196 + ], + "blocks": [ + { + "bbox": [ + 133, + 136, + 178, + 152 + ], + "lines": [ + { + "bbox": [ + 133, + 136, + 178, + 152 + ], + "spans": [ + { + "bbox": [ + 133, + 136, + 178, + 152 + ], + "type": "text", + "content": "The text in the red circle: 'Rights'." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 182, + 127, + 231, + 196 + ], + "lines": [ + { + "bbox": [ + 182, + 127, + 231, + 196 + ], + "spans": [ + { + "bbox": [ + 182, + 127, + 231, + 196 + ], + "type": "image", + "image_path": "2ff32f6c5fc5c8c0e7d9c838edef336146ad3d5d2d452df3cd8a96eb5f848b2e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 293, + 126, + 339, + 196 + ], + "blocks": [ + { + "bbox": [ + 239, + 137, + 285, + 152 + ], + "lines": [ + { + "bbox": [ + 239, + 137, + 285, + 152 + ], + "spans": [ + { + "bbox": [ + 239, + 137, + 285, + 152 + ], + "type": "text", + "content": "The text in the red circle: 'Ave'." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 293, + 126, + 339, + 196 + ], + "lines": [ + { + "bbox": [ + 293, + 126, + 339, + 196 + ], + "spans": [ + { + "bbox": [ + 293, + 126, + 339, + 196 + ], + "type": "image", + "image_path": "b6f719711669f26965c0c0aab109f6ff4de38cc022efa75785e1bfff82d10509.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 122, + 198, + 156, + 232 + ], + "blocks": [ + { + "bbox": [ + 86, + 203, + 113, + 226 + ], + "lines": [ + { + "bbox": [ + 86, + 203, + 113, + 226 + ], + "spans": [ + { + "bbox": [ + 86, + 203, + 113, + 226 + ], + "type": "text", + "content": "a photo of a yellow backpack:" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 122, + 198, + 156, + 232 + ], + "lines": [ + { + "bbox": [ + 122, + 198, + 156, + 232 + ], + "spans": [ + { + "bbox": [ + 122, + 198, + 156, + 232 + ], + "type": "image", + "image_path": "a984b813cd72fc5a86e70deb01b4d3b49ab8e913846529a250c48ce4643a60a9.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 149, + 232, + 184, + 267 + ], + "blocks": [ + { + "bbox": [ + 92, + 239, + 134, + 262 + ], + "lines": [ + { + "bbox": [ + 92, + 239, + 134, + 262 + ], + "spans": [ + { + "bbox": [ + 92, + 239, + 134, + 262 + ], + "type": "text", + "content": "The subject A with a city in the background:" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 149, + 232, + 184, + 267 + ], + "lines": [ + { + "bbox": [ + 149, + 232, + 184, + 267 + ], + "spans": [ + { + "bbox": [ + 149, + 232, + 184, + 267 + ], + "type": "image", + "image_path": "8d8b3d9a362c1cc3ca1888b4780271d0c8095498d8b52997a2db79f753101225.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 200, + 198, + 235, + 232 + ], + "blocks": [ + { + "bbox": [ + 165, + 205, + 191, + 226 + ], + "lines": [ + { + "bbox": [ + 165, + 205, + 191, + 226 + ], + "spans": [ + { + "bbox": [ + 165, + 205, + 191, + 226 + ], + "type": "text", + "content": "a photo of a blue backpack:" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 200, + 198, + 235, + 232 + ], + "lines": [ + { + "bbox": [ + 200, + 198, + 235, + 232 + ], + "spans": [ + { + "bbox": [ + 200, + 198, + 235, + 232 + ], + "type": "image", + "image_path": "ae4d936285f56781589dc0c08f8b1f9de2abbb8c279041c27b5e4fbafaf51303.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 258, + 198, + 293, + 232 + ], + "blocks": [ + { + "bbox": [ + 258, + 198, + 293, + 232 + ], + "lines": [ + { + "bbox": [ + 258, + 198, + 293, + 232 + ], + "spans": [ + { + "bbox": [ + 258, + 198, + 293, + 232 + ], + "type": "image", + "image_path": "f3f3282f01c52853b8596959d8781b36847ee25556ebf3e9c4c85bb732f55bc6.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 299, + 243, + 350, + 258 + ], + "lines": [ + { + "bbox": [ + 299, + 243, + 350, + 258 + ], + "spans": [ + { + "bbox": [ + 299, + 243, + 350, + 258 + ], + "type": "text", + "content": "The subject A in a purple wizard outfit:" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 294, + 198, + 313, + 232 + ], + "blocks": [ + { + "bbox": [ + 294, + 198, + 313, + 232 + ], + "lines": [ + { + "bbox": [ + 294, + 198, + 313, + 232 + ], + "spans": [ + { + "bbox": [ + 294, + 198, + 313, + 232 + ], + "type": "image", + "image_path": "b30403a6a631b901f5292a76ea31235533551cc8be5c0aa63f1586206c078bae.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 322, + 204, + 349, + 227 + ], + "lines": [ + { + "bbox": [ + 322, + 204, + 349, + 227 + ], + "spans": [ + { + "bbox": [ + 322, + 204, + 349, + 227 + ], + "type": "text", + "content": "a photo of a brown backpack:" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 357, + 196, + 392, + 232 + ], + "blocks": [ + { + "bbox": [ + 342, + 198, + 388, + 215 + ], + "lines": [ + { + "bbox": [ + 342, + 198, + 388, + 215 + ], + "spans": [ + { + "bbox": [ + 342, + 198, + 388, + 215 + ], + "type": "text", + "content": "of" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 399, + 205, + 435, + 226 + ], + "lines": [ + { + "bbox": [ + 399, + 205, + 435, + 226 + ], + "spans": [ + { + "bbox": [ + 399, + 205, + 435, + 226 + ], + "type": "text", + "content": "a photo of a blue and red backpack:" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 357, + 196, + 392, + 232 + ], + "lines": [ + { + "bbox": [ + 357, + 196, + 392, + 232 + ], + "spans": [ + { + "bbox": [ + 357, + 196, + 392, + 232 + ], + "type": "image", + "image_path": "81e4a0bd18ca12fc80a72a2a4143e50915625f04ca02987532551a38d97e9ddc.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 257, + 233, + 293, + 268 + ], + "blocks": [ + { + "bbox": [ + 190, + 243, + 250, + 258 + ], + "lines": [ + { + "bbox": [ + 190, + 243, + 250, + 258 + ], + "spans": [ + { + "bbox": [ + 190, + 243, + 250, + 258 + ], + "type": "text", + "content": "The subject A wearing a santa hat:" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 257, + 233, + 293, + 268 + ], + "lines": [ + { + "bbox": [ + 257, + 233, + 293, + 268 + ], + "spans": [ + { + "bbox": [ + 257, + 233, + 293, + 268 + ], + "type": "image", + "image_path": "d742bdad6bcf6a4c6004af35598c0cfb1d51bc43e843b792562535df65a5b081.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 357, + 239, + 392, + 266 + ], + "blocks": [ + { + "bbox": [ + 357, + 239, + 392, + 266 + ], + "lines": [ + { + "bbox": [ + 357, + 239, + 392, + 266 + ], + "spans": [ + { + "bbox": [ + 357, + 239, + 392, + 266 + ], + "type": "image", + "image_path": "ab5cbb65ae2ddcd90db53f9385e58ada3cdf0ff187222a07549312d913a3fd8e.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 468, + 197, + 505, + 232 + ], + "blocks": [ + { + "bbox": [ + 458, + 176, + 515, + 184 + ], + "lines": [ + { + "bbox": [ + 458, + 176, + 515, + 184 + ], + "spans": [ + { + "bbox": [ + 458, + 176, + 515, + 184 + ], + "type": "text", + "content": "motorcycle's headlight." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 468, + 197, + 505, + 232 + ], + "lines": [ + { + "bbox": [ + 468, + 197, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 468, + 197, + 505, + 232 + ], + "type": "image", + "image_path": "a779de923f8cba57b2cfdde5bacdbe9d56b1752bd0eadf0c831a5494eb133946.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 468, + 236, + 504, + 268 + ], + "blocks": [ + { + "bbox": [ + 468, + 236, + 504, + 268 + ], + "lines": [ + { + "bbox": [ + 468, + 236, + 504, + 268 + ], + "spans": [ + { + "bbox": [ + 468, + 236, + 504, + 268 + ], + "type": "image", + "image_path": "14e89c157256675bb4a8e3278a36e973f488147d308e5040932e0c0808a6d2cf.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 96, + 288, + 146, + 326 + ], + "blocks": [ + { + "bbox": [ + 96, + 288, + 146, + 326 + ], + "lines": [ + { + "bbox": [ + 96, + 288, + 146, + 326 + ], + "spans": [ + { + "bbox": [ + 96, + 288, + 146, + 326 + ], + "type": "image", + "image_path": "4a708d896cb4fa258a03419dd1c1d491df7b0b979056ea67b760ea64eb4bbf99.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 79, + 336, + 161, + 373 + ], + "lines": [ + { + "bbox": [ + 79, + 336, + 161, + 373 + ], + "spans": [ + { + "bbox": [ + 79, + 336, + 161, + 373 + ], + "type": "text", + "content": "Look at this sequence of three shapes. What shape should come as the fourth shape? Explain your reasoning with detailed descriptions of the first shapes." + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "bbox": [ + 167, + 292, + 245, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 292, + 245, + 372 + ], + "spans": [ + { + "bbox": [ + 167, + 292, + 245, + 372 + ], + "type": "text", + "content": "The first shape is a triangle, the second shape is a square, and the third shape is a pentagon. To find the next shape in the sequence, we need to identify a pattern. In this case, the pattern is that each shape has one more side than the previous shape. Therefore, the next shape in the sequence should have six sides, which is a hexagon." + } + ] + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 269, + 290, + 319, + 328 + ], + "blocks": [ + { + "bbox": [ + 269, + 290, + 319, + 328 + ], + "lines": [ + { + "bbox": [ + 269, + 290, + 319, + 328 + ], + "spans": [ + { + "bbox": [ + 269, + 290, + 319, + 328 + ], + "type": "image", + "image_path": "345c2ff9a7706acaced5a87e494f53efc5d3d183c9d9df39d4263a04e26b43d5.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 256, + 350, + 335, + 358 + ], + "lines": [ + { + "bbox": [ + 256, + 350, + 335, + 358 + ], + "spans": [ + { + "bbox": [ + 256, + 350, + 335, + 358 + ], + "type": "text", + "content": "What is funny about this image?" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + } + ], + "index": 40 + }, + { + "bbox": [ + 341, + 287, + 421, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 287, + 421, + 380 + ], + "spans": [ + { + "bbox": [ + 341, + 287, + 421, + 380 + ], + "type": "text", + "content": "The funny aspect of this image is that a squirrel is holding a camera and appears to be taking a picture of a bird. This is amusing because animals, such as squirrels and birds, do not typically use cameras or engage in human-like activities like photography. The scene is set up in a way that makes it look like the squirrel is intentionally capturing the bird's image, which adds a humorous twist to the scene. Additionally, the bird is perched on a mushroom, further enhancing the whimsical and playful nature of the image." + } + ] + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 430, + 280, + 531, + 383 + ], + "blocks": [ + { + "bbox": [ + 430, + 280, + 531, + 383 + ], + "lines": [ + { + "bbox": [ + 430, + 280, + 531, + 383 + ], + "spans": [ + { + "bbox": [ + 430, + 280, + 531, + 383 + ], + "type": "image", + "image_path": "5e49f2038c304b38fb988de176d8fa07331dd47731cc1c8820e381bd495b3f42.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 76, + 395, + 164, + 483 + ], + "blocks": [ + { + "bbox": [ + 76, + 395, + 164, + 483 + ], + "lines": [ + { + "bbox": [ + 76, + 395, + 164, + 483 + ], + "spans": [ + { + "bbox": [ + 76, + 395, + 164, + 483 + ], + "type": "image", + "image_path": "b8771eba724fe87515faa21c975df27701a76a07e0a20141161ad134e9c4e539.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 86, + 491, + 154, + 506 + ], + "lines": [ + { + "bbox": [ + 86, + 491, + 154, + 506 + ], + "spans": [ + { + "bbox": [ + 86, + 491, + 154, + 506 + ], + "type": "text", + "content": "impressionist painting of an astronaut in a jungle" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 267, + 395, + 356, + 481 + ], + "blocks": [ + { + "bbox": [ + 267, + 395, + 356, + 481 + ], + "lines": [ + { + "bbox": [ + 267, + 395, + 356, + 481 + ], + "spans": [ + { + "bbox": [ + 267, + 395, + 356, + 481 + ], + "type": "image", + "image_path": "575e7154d1415b51921064162802d7ad53b7703bd2f0245a3910adde8cee94b5.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 189, + 494, + 332, + 502 + ], + "lines": [ + { + "bbox": [ + 189, + 494, + 332, + 502 + ], + "spans": [ + { + "bbox": [ + 189, + 494, + 332, + 502 + ], + "type": "text", + "content": "An image of wearing a big sunglasses on the beach" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + } + ], + "index": 47 + }, + { + "type": "image", + "bbox": [ + 189, + 402, + 237, + 476 + ], + "blocks": [ + { + "bbox": [ + 189, + 402, + 237, + 476 + ], + "lines": [ + { + "bbox": [ + 189, + 402, + 237, + 476 + ], + "spans": [ + { + "bbox": [ + 189, + 402, + 237, + 476 + ], + "type": "image", + "image_path": "dc5f36c519e6e38dbc6b601e0d5233a9d933c7ca79abcefde199813ee8ed810e.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 356, + 396, + 441, + 483 + ], + "blocks": [ + { + "bbox": [ + 356, + 396, + 441, + 483 + ], + "lines": [ + { + "bbox": [ + 356, + 396, + 441, + 483 + ], + "spans": [ + { + "bbox": [ + 356, + 396, + 441, + 483 + ], + "type": "image", + "image_path": "56e48415d335188b4e7a492cdb006a2680e263fdcaeca29c72fd9dfe218ea6a5.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 442, + 395, + 531, + 483 + ], + "blocks": [ + { + "bbox": [ + 442, + 395, + 531, + 483 + ], + "lines": [ + { + "bbox": [ + 442, + 395, + 531, + 483 + ], + "spans": [ + { + "bbox": [ + 442, + 395, + 531, + 483 + ], + "type": "image", + "image_path": "8631505cc0b1e5ae9da3fbadd84a9da737df6128a09e7df04523cad4b6683db6.jpg" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_body" + } + ], + "index": 51 + }, + { + "type": "image", + "bbox": [ + 171, + 536, + 213, + 578 + ], + "blocks": [ + { + "bbox": [ + 78, + 552, + 169, + 559 + ], + "lines": [ + { + "bbox": [ + 78, + 552, + 169, + 559 + ], + "spans": [ + { + "bbox": [ + 78, + 552, + 169, + 559 + ], + "type": "text", + "content": "An anime draw of " + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 171, + 536, + 213, + 578 + ], + "lines": [ + { + "bbox": [ + 171, + 536, + 213, + 578 + ], + "spans": [ + { + "bbox": [ + 171, + 536, + 213, + 578 + ], + "type": "image", + "image_path": "e6daaa233f23264b059d0d6fd6b483e222aef55470fd14ca0e668e5a4710fbbb.jpg" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_body" + } + ], + "index": 54 + }, + { + "type": "image", + "bbox": [ + 216, + 537, + 237, + 578 + ], + "blocks": [ + { + "bbox": [ + 216, + 537, + 237, + 578 + ], + "lines": [ + { + "bbox": [ + 216, + 537, + 237, + 578 + ], + "spans": [ + { + "bbox": [ + 216, + 537, + 237, + 578 + ], + "type": "image", + "image_path": "471cad8f46d65acc63f695dd251e311729e43cdc21a99e756df02e1b315f1afb.jpg" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 239, + 552, + 273, + 560 + ], + "lines": [ + { + "bbox": [ + 239, + 552, + 273, + 560 + ], + "spans": [ + { + "bbox": [ + 239, + 552, + 273, + 560 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 239, + 552, + 273, + 560 + ], + "type": "inline_equation", + "content": "\\text{a dog}" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_caption" + } + ], + "index": 55 + }, + { + "type": "image", + "bbox": [ + 276, + 536, + 316, + 578 + ], + "blocks": [ + { + "bbox": [ + 276, + 536, + 316, + 578 + ], + "lines": [ + { + "bbox": [ + 276, + 536, + 316, + 578 + ], + "spans": [ + { + "bbox": [ + 276, + 536, + 316, + 578 + ], + "type": "image", + "image_path": "4d04223adb671d60a30154108673737b9de3adcdcd2fe98da48d3797d22e5833.jpg" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_body" + } + ], + "index": 57 + }, + { + "type": "image", + "bbox": [ + 317, + 536, + 360, + 578 + ], + "blocks": [ + { + "bbox": [ + 317, + 536, + 360, + 578 + ], + "lines": [ + { + "bbox": [ + 317, + 536, + 360, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 536, + 360, + 578 + ], + "type": "image", + "image_path": "ac7d0b522d218ac19c15f874ba476b6c20751be158e2ac5048376a2498259e1e.jpg" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_body" + } + ], + "index": 58 + }, + { + "type": "image", + "bbox": [ + 403, + 544, + 443, + 571 + ], + "blocks": [ + { + "bbox": [ + 362, + 553, + 402, + 560 + ], + "lines": [ + { + "bbox": [ + 362, + 553, + 402, + 560 + ], + "spans": [ + { + "bbox": [ + 362, + 553, + 402, + 560 + ], + "type": "text", + "content": ", in the style of" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 403, + 544, + 443, + 571 + ], + "lines": [ + { + "bbox": [ + 403, + 544, + 443, + 571 + ], + "spans": [ + { + "bbox": [ + 403, + 544, + 443, + 571 + ], + "type": "image", + "image_path": "b719d4b5086e773d384f307c4c5d3406accab4e3d9f78f03dd0989ff5c0a63d3.jpg" + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_body" + } + ], + "index": 60 + }, + { + "type": "image", + "bbox": [ + 447, + 513, + 531, + 596 + ], + "blocks": [ + { + "bbox": [ + 447, + 513, + 531, + 596 + ], + "lines": [ + { + "bbox": [ + 447, + 513, + 531, + 596 + ], + "spans": [ + { + "bbox": [ + 447, + 513, + 531, + 596 + ], + "type": "image", + "image_path": "03e2ddada0a6b3036cd26a09b3401d0d546ea6fa19e4db0c96652fa9318b98af.jpg" + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "image_body" + } + ], + "index": 61 + }, + { + "type": "image", + "bbox": [ + 176, + 623, + 220, + 647 + ], + "blocks": [ + { + "bbox": [ + 78, + 631, + 170, + 646 + ], + "lines": [ + { + "bbox": [ + 78, + 631, + 170, + 646 + ], + "spans": [ + { + "bbox": [ + 78, + 631, + 170, + 646 + ], + "type": "text", + "content": " and [USER]: [ASSISTANT]: ”. Here represents system message and varies between the two major task categories (academic-task-oriented and multimodal chat). The section comprises multimodal tokens, including images, videos, and text. Only tokens in the section will be supervised by cross-entropy loss during training." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 253, + 290, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 253, + 290, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 253, + 290, + 445 + ], + "type": "text", + "content": "Training Setup. We use a global batch size of 768 and train for 8k steps. The learning rate linearly warms up to " + }, + { + "bbox": [ + 46, + 253, + 290, + 445 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-5}" + }, + { + "bbox": [ + 46, + 253, + 290, + 445 + ], + "type": "text", + "content": " in the first 100 steps, then decays to zero with a cosine schedule. The model is trained using the AdamW optimizer with " + }, + { + "bbox": [ + 46, + 253, + 290, + 445 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 46, + 253, + 290, + 445 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 253, + 290, + 445 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.98" + }, + { + "bbox": [ + 46, + 253, + 290, + 445 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 253, + 290, + 445 + ], + "type": "inline_equation", + "content": "\\epsilon = 1 \\times 10^{-6}" + }, + { + "bbox": [ + 46, + 253, + 290, + 445 + ], + "type": "text", + "content": ", and a gradient clipping of 5.0. The sequence length during training is limited to 2048, and any excess beyond that is truncated directly. We consistently employed an input image/video resolution of " + }, + { + "bbox": [ + 46, + 253, + 290, + 445 + ], + "type": "inline_equation", + "content": "448 \\times 448" + }, + { + "bbox": [ + 46, + 253, + 290, + 445 + ], + "type": "text", + "content": ". For video data, we uniformly sample frames in time as input to the model. The number of sampled frames for each video is randomly chosen from 8, 12, and 16. To capture more intricate spatial details, following the visual encoder stage, we apply mean-pooling to each static image, dividing it into " + }, + { + "bbox": [ + 46, + 253, + 290, + 445 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 46, + 253, + 290, + 445 + ], + "type": "text", + "content": " tokens during instruction fine-tuning. This differs from the pre-training phase, where " + }, + { + "bbox": [ + 46, + 253, + 290, + 445 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 46, + 253, + 290, + 445 + ], + "type": "text", + "content": " tokens were utilized." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 464, + 215, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 464, + 215, + 477 + ], + "spans": [ + { + "bbox": [ + 47, + 464, + 215, + 477 + ], + "type": "text", + "content": "2.3.2 Controllable Visual Generation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 485, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 485, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 485, + 287, + 640 + ], + "type": "text", + "content": "Training Data. We leverage a mix of high-quality datasets to unleash the potential of controllable generation in context. We use a grounded image-text pair dataset CapsFusion-grounded-100M and GRIT [57] for grounded text-to-image generation. To mitigate the impact of image backgrounds on the effectiveness of multi-entity subject-driven generation, we employ SAM [36] to preprocess the grounding data, yielding a subset of approximately 5 million samples with segmentation results. Additionally, we leverage InstructPix2Pix constructed by [10] for image editing tasks. For the text-to-image task, we use a filtered subset of CapsFusion [87], LAION-Aesthetics [1], SA-1B [36], and LAION-High-Resolution [3]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": "We also collect data from premium sources (e.g., Unsplash [20]) and outputs from advanced text-to-image systems (e.g., Midjourney-V5 [54] and DALL-E-3 [9]) for quality fine-tuning. This diverse dataset includes around 500k high-quality image-text pairs. For all the data above, during the training, only samples with image resolutions" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "content": "higher than " + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "inline_equation", + "content": "448 \\times 448" + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "content": " were retained to ensure generation quality. More details can be found in the supplementary." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 310, + 107, + 545, + 370 + ], + "blocks": [ + { + "bbox": [ + 310, + 107, + 545, + 370 + ], + "lines": [ + { + "bbox": [ + 310, + 107, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 310, + 107, + 545, + 370 + ], + "type": "table", + "html": "
ModelShotVQAv2OKVQAVizWizTextVQAHateful Memes
Kosmos-1 (1.6B)051.0-29.2--
451.8-35.3--
851.4-39.0--
Flamingo (9B)0*51.844.728.831.857.0
456.349.334.933.662.7
858.050.039.433.663.9
1659.450.843.033.564.5
Flamingo (80B)0*56.350.631.635.046.4
463.157.439.636.568.6
865.657.544.837.370.0
1666.857.848.437.670.0
IDEFICS (80B)0*60.045.236.030.960.6
463.652.440.434.457.8
864.855.146.135.758.2
1665.456.848.336.357.8
Emu (14B)0*52.942.834.4--
458.4-41.3--
859.0-43.9--
16-----
Emu2 (37B)033.526.740.426.452.2
467.053.254.648.262.4
867.854.154.749.365.8
1668.857.157.050.366.0
", + "image_path": "5ddefc160cf822c8872b36f202ac4897bf5aa162984612ce85ecae06a175c411.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 377, + 545, + 411 + ], + "lines": [ + { + "bbox": [ + 305, + 377, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 305, + 377, + 545, + 411 + ], + "type": "text", + "content": "Table 1. Zero-shot and few-shot evaluations of Emu2. " + }, + { + "bbox": [ + 305, + 377, + 545, + 411 + ], + "type": "inline_equation", + "content": "0^{*}" + }, + { + "bbox": [ + 305, + 377, + 545, + 411 + ], + "type": "text", + "content": " denotes text two-shot and image zero-shot results following Flamingo [5]. The best results are in bold and the second best are underlined." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 426, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 713 + ], + "type": "text", + "content": "Training Objective. We use the same unified generative pretraining objective to adapt to diverse generation tasks in context. Specifically, a training sample for generation is formulated as: “A photo of

a man

image embedding of object localization image[IMG] image embedding of man[/IMG] sitting next to

a dog

image embedding of object localization image[IMG] image embedding of dog[/IMG][IMG] image embedding of the whole image[/IMG”. We represent the coordinates of each object directly in image form by drawing the bounding box of each object at its specified location on a black image. Our Emu2-Gen conducts unified multimodal modeling of the text, object image, and corresponding object localization image. The regression loss only applies to the visual embeddings of the last image. We freeze the Visual Encoder during fine-tuning. We randomly drop tokens of entities and object localization image to enhance model adaptability and robustness. Additionally, we apply data augmentation to each object image, incorporating random background variations and random crop, aiming to reduce the reliance on image backgrounds. Training Setup. We use a global batch size of 4,096 and" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 734, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 734, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 734, + 318, + 742 + ], + "type": "text", + "content": "14402" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 545, + 265 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 545, + 265 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 545, + 265 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 545, + 265 + ], + "type": "table", + "html": "
ModelVisual Question AnswerLMM Benchmarks
VQAv2 [28]OKVQA [53]GQA [32]VizWiz [29]TextVQA [68]MSVD [82]MSRVTT [82]SEED [39]MM-Vet [88]TS [7]MMMU [89]
Flamingo-9B [5]51.844.7-28.8-30.213.7----
Flamingo-80B [5]56.350.6-31.6-35.617.4----
Kosmos-1 [31]51.0--29.2-------
Kosmos-2 [57]51.1------50.0--26.6
BLIP-2-13B [42]--41.019.642.520.310.346.422.4--
InstructBLIP-13B [22]--49.533.450.741.224.8-25.6552.4-
IDEFICS-9B [38]50.938.4-35.525.9------
IDEFICS-80B [38]60.045.2-36.030.9------
Shikra-13B [15]77.4*47.2---------
Qwen-VL-13B-Chat [6]78.2*56.6*57.5*38.961.5*--58.2-645.2-
LLaVA-1.5-13B [48]80.0*-63.3*53.661.3--61.635.4-33.6
CogVLM [77]83.4*58.9*--68.1*----662.630.1
Emu-I [71]62.049.246.038.3-37.021.2-36.3--
Emu2-Chat84.9*64.8*65.1*54.966.6*49.031.462.848.5703.834.1
", + "image_path": "365871c6dd5272dc4273fc556f6098340601702f5e796cb4ef00d51286185620.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 274, + 546, + 308 + ], + "lines": [ + { + "bbox": [ + 46, + 274, + 546, + 308 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 546, + 308 + ], + "type": "text", + "content": "Table 2. Results on visual question answering and LMM benchmarks. * indicates that samples from this task's training set have been trained. SEED and TS respectively represent SEED-Bench [39] and TouchStone [7]. For MM-Vet, we present the average result of five scoring runs." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 328, + 287, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 328, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 328, + 287, + 376 + ], + "type": "text", + "content": "train for 3k steps. The learning rate linearly warms up to " + }, + { + "bbox": [ + 46, + 328, + 287, + 376 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-5}" + }, + { + "bbox": [ + 46, + 328, + 287, + 376 + ], + "type": "text", + "content": " in the first 100 steps, then decays to zero with a cosine schedule. We further fine-tune for 900 steps using the 500k high-quality pairs with a batch size of 2048." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 387, + 118, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 387, + 118, + 399 + ], + "spans": [ + { + "bbox": [ + 47, + 387, + 118, + 399 + ], + "type": "text", + "content": "3. Evaluation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 406, + 178, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 406, + 178, + 418 + ], + "spans": [ + { + "bbox": [ + 47, + 406, + 178, + 418 + ], + "type": "text", + "content": "3.1. Pretrained Base Model" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 424, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 424, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 424, + 287, + 544 + ], + "type": "text", + "content": "We evaluate zero-shot and few-shot abilities of Emu2 on OKVQA [53], VQAv2 [28], VizWiz [29], TextVQA [68], and HatefulMemes [35] tasks. Details of the datasets and prompts can be found in supplementary materials. The results are presented in Table 1. Emu2 demonstrates remarkable in-context ability, showcasing improved performance with more in-context samples seen. Specifically, on VQAv2, VizWiz and TextVQA datasets, Emu2 outperforms Flamingo-80B and IDEFICS-80B under all few-shot settings with a much smaller model scale (37B)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 544, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 616 + ], + "type": "text", + "content": "Figure 1 demonstrates Emu2's few-shot capabilities in the wild. For example, the model learns to classify and count simultaneously in a specific format via a few examples (row 1). Additionally, Emu2 is capable of following visual prompts in context, e.g., the red circles laid on the images (row 2 and 3)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 623, + 197, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 197, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 197, + 635 + ], + "type": "text", + "content": "3.2. Instruction-Following Chat" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": "Our Emu2-Chat is evaluated on academic-task-oriented benchmarks including image question-answering datasets (VQAv2 [28], OKVQA [53], GQA [32], VizWiz [29], TextVQA [68]) and video question-answering datasets (MSVD [82] and MSRvTT [82]). The evaluation also encompassed recent benchmarks for large multimodal models," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 328, + 545, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 328, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 304, + 328, + 545, + 376 + ], + "type": "text", + "content": "including SEED-Bench [39], MM-Vet [88], TouchStone [7] and MMMU [89]. When evaluated on SEED-Bench, we followed the setup of LLaVa-1.5 [48] by presenting options to the model for completing multiple-choice tasks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 377, + 546, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 377, + 546, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 377, + 546, + 605 + ], + "type": "text", + "content": "As shown in Table 2, Emu2-Chat consistently outperforms other models in image question-answering tasks, encompassing well-established benchmarks like VQAv2 and GQA. Notably, it shows a noticeable improvement in the OKVQA task, which requires the utilization of external knowledge, showcasing the advantage of our model for mastering real-world knowledge. For video question-answering, Emu2-Chat demonstrated advantages even though it did not use video question-answering data for training. It achieved an accuracy of 49.0 and 31.4 on the MSVD-QA and MSRVTT-QA tasks, respectively, surpassing InstructBLIP and the larger Flamingo-80B. More importantly, our model has also achieved better results on LMM benchmarks. LMM benchmarks such as MM-Vet provide a more comprehensive evaluation of model abilities, including solving complicated tasks. Emu2-Chat achieves a score of 48.5 in MM-Vet and 703.8 in Touch-Stone, confirming its superior capability in understanding and solving multimodal problems." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "In addition, we demonstrated the visual grounding capability of our model using the refer expression comprehension benchmarks. In Table 3, Emu2-Chat achieved the best results among generalist models on RefCOCO [34], RefCOCO+ [52] and RefCOCOg [52]. Its most notable advantage was observed in RefCOCO+, which focused solely on purely appearance-based descriptions without allowing the use of position references. This highlights our model's powerful perceptual abilities in capturing intricate details." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "type": "text", + "content": "14403" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 71, + 102, + 111 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 102, + 111 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 102, + 111 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 102, + 111 + ], + "type": "image", + "image_path": "8cc37d98935c4f44559647e055e5f0294afab6627bad7cb4697c07728efeac58.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 112, + 102, + 152 + ], + "blocks": [ + { + "bbox": [ + 58, + 112, + 102, + 152 + ], + "lines": [ + { + "bbox": [ + 58, + 112, + 102, + 152 + ], + "spans": [ + { + "bbox": [ + 58, + 112, + 102, + 152 + ], + "type": "image", + "image_path": "f49789c75d3cf1d5b93bf13793e3a023eb895543a8af4f2a5b2930fc81ddbdfd.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 62, + 152, + 97, + 191 + ], + "blocks": [ + { + "bbox": [ + 62, + 152, + 97, + 191 + ], + "lines": [ + { + "bbox": [ + 62, + 152, + 97, + 191 + ], + "spans": [ + { + "bbox": [ + 62, + 152, + 97, + 191 + ], + "type": "image", + "image_path": "c243d98ec61ed32f69243e80a599b8a9c5fb00f121a214bf3369082167fad9d3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 110, + 72, + 195, + 156 + ], + "blocks": [ + { + "bbox": [ + 110, + 72, + 195, + 156 + ], + "lines": [ + { + "bbox": [ + 110, + 72, + 195, + 156 + ], + "spans": [ + { + "bbox": [ + 110, + 72, + 195, + 156 + ], + "type": "image", + "image_path": "e35517afe733d538ccf4abdc92ef411ca9a0e99a964baaafae17d14b8d268c6c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 117, + 164, + 189, + 187 + ], + "lines": [ + { + "bbox": [ + 117, + 164, + 189, + 187 + ], + "spans": [ + { + "bbox": [ + 117, + 164, + 189, + 187 + ], + "type": "text", + "content": "A photo of the first dog>, the second dog>, the third dog on the grass" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 197, + 72, + 282, + 157 + ], + "blocks": [ + { + "bbox": [ + 197, + 72, + 282, + 157 + ], + "lines": [ + { + "bbox": [ + 197, + 72, + 282, + 157 + ], + "spans": [ + { + "bbox": [ + 197, + 72, + 282, + 157 + ], + "type": "image", + "image_path": "cf05a40c61a0659c785072d14443bdd2adf475adde31682340a08b7a756a1698.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 202, + 164, + 277, + 187 + ], + "lines": [ + { + "bbox": [ + 202, + 164, + 277, + 187 + ], + "spans": [ + { + "bbox": [ + 202, + 164, + 277, + 187 + ], + "type": "text", + "content": "An oil painting of the first dog>, the second dog>, the third dog>" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 284, + 72, + 369, + 157 + ], + "blocks": [ + { + "bbox": [ + 284, + 72, + 369, + 157 + ], + "lines": [ + { + "bbox": [ + 284, + 72, + 369, + 157 + ], + "spans": [ + { + "bbox": [ + 284, + 72, + 369, + 157 + ], + "type": "image", + "image_path": "3d8dbce32e38d79ef07ae1922ac658afe4324c4f9d5c27275e43f0118f0e2846.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 290, + 164, + 362, + 187 + ], + "lines": [ + { + "bbox": [ + 290, + 164, + 362, + 187 + ], + "spans": [ + { + "bbox": [ + 290, + 164, + 362, + 187 + ], + "type": "text", + "content": "A photo of the first dog>, the second dog>, the third dog> on the beach" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 370, + 72, + 455, + 157 + ], + "blocks": [ + { + "bbox": [ + 370, + 72, + 455, + 157 + ], + "lines": [ + { + "bbox": [ + 370, + 72, + 455, + 157 + ], + "spans": [ + { + "bbox": [ + 370, + 72, + 455, + 157 + ], + "type": "image", + "image_path": "01eeb3d932b34db42a4a83ad3bf541fe2cc358e560872b8eb9e0f9b01c743ac2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 400, + 169, + 443, + 183 + ], + "lines": [ + { + "bbox": [ + 400, + 169, + 443, + 183 + ], + "spans": [ + { + "bbox": [ + 400, + 169, + 443, + 183 + ], + "type": "text", + "content": "as an oil painting by Monet" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 374, + 164, + 395, + 186 + ], + "blocks": [ + { + "bbox": [ + 374, + 164, + 395, + 186 + ], + "lines": [ + { + "bbox": [ + 374, + 164, + 395, + 186 + ], + "spans": [ + { + "bbox": [ + 374, + 164, + 395, + 186 + ], + "type": "image", + "image_path": "005d7167506e27504e53f460afd17bcf90099b1c73eebf5344af922093879ae2.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 457, + 72, + 542, + 157 + ], + "blocks": [ + { + "bbox": [ + 457, + 72, + 542, + 157 + ], + "lines": [ + { + "bbox": [ + 457, + 72, + 542, + 157 + ], + "spans": [ + { + "bbox": [ + 457, + 72, + 542, + 157 + ], + "type": "image", + "image_path": "0cd54a6d7deaf963723f17b3fb76825c5236b86ffa07e4d515fa1518840ce7dc.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 483, + 171, + 534, + 178 + ], + "lines": [ + { + "bbox": [ + 483, + 171, + 534, + 178 + ], + "spans": [ + { + "bbox": [ + 483, + 171, + 534, + 178 + ], + "type": "text", + "content": "in a dark forest in 3D" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 459, + 165, + 477, + 186 + ], + "blocks": [ + { + "bbox": [ + 459, + 165, + 477, + 186 + ], + "lines": [ + { + "bbox": [ + 459, + 165, + 477, + 186 + ], + "spans": [ + { + "bbox": [ + 459, + 165, + 477, + 186 + ], + "type": "image", + "image_path": "9e567b0a83330b66bae6b4cc1437fded089295b75a239fef9e4f5c54cbea2b24.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 51, + 193, + 156, + 298 + ], + "blocks": [ + { + "bbox": [ + 51, + 193, + 156, + 298 + ], + "lines": [ + { + "bbox": [ + 51, + 193, + 156, + 298 + ], + "spans": [ + { + "bbox": [ + 51, + 193, + 156, + 298 + ], + "type": "image", + "image_path": "4a23fa92e42fe9417d7aabb40a922353a9b7b3f04942f0384346061cb73e86b1.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 71, + 312, + 135, + 320 + ], + "lines": [ + { + "bbox": [ + 71, + 312, + 135, + 320 + ], + "spans": [ + { + "bbox": [ + 71, + 312, + 135, + 320 + ], + "type": "text", + "content": "a bear and a sunflower" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 157, + 193, + 263, + 298 + ], + "blocks": [ + { + "bbox": [ + 157, + 193, + 263, + 298 + ], + "lines": [ + { + "bbox": [ + 157, + 193, + 263, + 298 + ], + "spans": [ + { + "bbox": [ + 157, + 193, + 263, + 298 + ], + "type": "image", + "image_path": "f3224340123304d8763ef9a5dd124aa5ebf91c211ecbfc6dff103f8c9847ba64.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 159, + 300, + 262, + 333 + ], + "blocks": [ + { + "bbox": [ + 159, + 300, + 262, + 333 + ], + "lines": [ + { + "bbox": [ + 159, + 300, + 262, + 333 + ], + "spans": [ + { + "bbox": [ + 159, + 300, + 262, + 333 + ], + "type": "image", + "image_path": "e6bacea7baf4fbd6906a22b63e90333636fda2da2df53b1470d32aaa70c78242.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 345, + 546, + 379 + ], + "lines": [ + { + "bbox": [ + 46, + 345, + 546, + 379 + ], + "spans": [ + { + "bbox": [ + 46, + 345, + 546, + 379 + ], + "type": "text", + "content": "Figure 4. Visualization of Emu2-Gen's controllable generation capability. The model is capable of accepting a mix of text, locations and images as input, and generating images in context. The presented examples include text- and subject-grounded generation, stylization, multi-entity composition, subject-driven editing, and text-to-image generation." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 263, + 193, + 369, + 298 + ], + "blocks": [ + { + "bbox": [ + 263, + 193, + 369, + 298 + ], + "lines": [ + { + "bbox": [ + 263, + 193, + 369, + 298 + ], + "spans": [ + { + "bbox": [ + 263, + 193, + 369, + 298 + ], + "type": "image", + "image_path": "9a15867f22c2fc32224497c3d97c80b1999a30acde3f37a3d56c88e747a209e7.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 264, + 300, + 366, + 333 + ], + "blocks": [ + { + "bbox": [ + 264, + 300, + 366, + 333 + ], + "lines": [ + { + "bbox": [ + 264, + 300, + 366, + 333 + ], + "spans": [ + { + "bbox": [ + 264, + 300, + 366, + 333 + ], + "type": "image", + "image_path": "46c7465f93b0dc2005ac0d6f2fc237cad65f7b6e4b397ea06add7a4379628794.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 370, + 193, + 455, + 277 + ], + "blocks": [ + { + "bbox": [ + 370, + 193, + 455, + 277 + ], + "lines": [ + { + "bbox": [ + 370, + 193, + 455, + 277 + ], + "spans": [ + { + "bbox": [ + 370, + 193, + 455, + 277 + ], + "type": "image", + "image_path": "5c79ae0cdef25fe4e2b8769659255fdca36387d774d3a21bb355df5e16fab49e.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 370, + 279, + 455, + 334 + ], + "blocks": [ + { + "bbox": [ + 370, + 279, + 455, + 334 + ], + "lines": [ + { + "bbox": [ + 370, + 279, + 455, + 334 + ], + "spans": [ + { + "bbox": [ + 370, + 279, + 455, + 334 + ], + "type": "image", + "image_path": "3dc820fd695e8c875e65609b25c378a1b990c43c26ed2812ac2427b1855f56a3.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 457, + 193, + 542, + 277 + ], + "blocks": [ + { + "bbox": [ + 457, + 193, + 542, + 277 + ], + "lines": [ + { + "bbox": [ + 457, + 193, + 542, + 277 + ], + "spans": [ + { + "bbox": [ + 457, + 193, + 542, + 277 + ], + "type": "image", + "image_path": "31ee407d139f4e9740f54ce47a40de1cd41b635024c73154551387b368021693.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 457, + 279, + 542, + 334 + ], + "blocks": [ + { + "bbox": [ + 457, + 279, + 542, + 334 + ], + "lines": [ + { + "bbox": [ + 457, + 279, + 542, + 334 + ], + "spans": [ + { + "bbox": [ + 457, + 279, + 542, + 334 + ], + "type": "image", + "image_path": "d77f6c2bc5f9b120d3614cba07753f7c2f472f93935ba6a3ee954a5beba1054e.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "table", + "bbox": [ + 50, + 398, + 285, + 487 + ], + "blocks": [ + { + "bbox": [ + 50, + 398, + 285, + 487 + ], + "lines": [ + { + "bbox": [ + 50, + 398, + 285, + 487 + ], + "spans": [ + { + "bbox": [ + 50, + 398, + 285, + 487 + ], + "type": "table", + "html": "
ModelRefCOCORefCOCO+RefCOCOg
valtestAtestBvaltestAtestBvaltest
OFA-L [75]79.9683.6776.3968.2976.0061.7567.5767.58
Shikra-7B [15]87.0190.6180.2481.6087.3672.1282.2782.19
Shikra-13B [15]87.8391.1181.8182.8987.7974.4182.6483.16
Qwen-VL-7B [6]89.3692.2685.3483.1288.2577.2185.5885.48
Emu2-Chat90.4093.8885.9787.0591.4380.4787.6488.11
CogVLM [77]92.5193.9588.7387.5291.8181.4389.4690.09
", + "image_path": "7f27792e67ded5ea3b9c6dfe6e7e0ff949fbefc79e7c32e5f91ddc2b90a55a6b.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "table_body" + } + ], + "index": 26 + }, + { + "bbox": [ + 46, + 495, + 287, + 529 + ], + "lines": [ + { + "bbox": [ + 46, + 495, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 287, + 529 + ], + "type": "text", + "content": "Table 3. Results on referring expression comprehension. We grayed out CogVLM because its generalist grounding-enhanced model was specialist trained on high-quality grounding data." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 545, + 217, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 545, + 217, + 558 + ], + "spans": [ + { + "bbox": [ + 47, + 545, + 217, + 558 + ], + "type": "text", + "content": "3.3. Controllable Visual Generation" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 46, + 567, + 287, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 567, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 567, + 287, + 662 + ], + "type": "text", + "content": "Qualitative Results. Figure 3 presents a visualization of Emu2's autoencoding results. With Emu2's Visual Encoder and Visual Decoder, we can tokenize an image into visual embeddings and detokenize them back. Compared with SEED [27] and Emu [71], Emu2 shows significantly superior results. We also evaluate our image autoencoding results on MS-COCO [47] and achieve a strong 0.907 CLIP-I [59] score. More results are in the supplementary." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "As depicted in Figure 4, Emu2-Gen is capable of accepting a mixture of text, locations and images as input, and generating images in context. The model skillfully engages in various controllable visual generation tasks in" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 399, + 545, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 545, + 483 + ], + "type": "text", + "content": "a zero-shot setting, capitalizing on the in-context learning capabilities in multimodality. Examples in Figure 4 show generated images of three dogs conditioned on different subjects, locations and scenarios. The presented visual samples demonstrate the model's proficiency in tasks such as re-contextualization, stylization, modification, region-controllable generation, and multi-entity composition." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 304, + 484, + 546, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 546, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 546, + 652 + ], + "type": "text", + "content": "Zero-shot Text-to-image Generation. We evaluate the zero-shot text-to-image generation capability on 30k randomly sampled data from the MS-COCO [47] validation set. We employ CLIP-ViT-B [60], following the approach in DALL-E 3[9], to calculate the CLIP-T score to assess prompt-following ability. Additionally, we utilize CLIP-ViT-L, as in GILL[37], to compute the CLIP-I score for measuring image similarity. A higher score means the generated image is more similar to the prompt or the real image. Table 4 shows that Emu2-Gen achieves the state-of-the-art performance in terms of both CLIP-I and CLIP-T scores compared to various unimodal generation models and multimodal models. More text-to-image generation cases can be found in supplementary." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 653, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 546, + 714 + ], + "type": "text", + "content": "Zero-shot Subject-driven Generation. Following Kosmos-G [56], we also evaluate our model's subject-driven image editing ability on DreamBench [63]. We generate four images for each prompt, resulting in a total of 3,000 images for a comprehensive evaluation. We employ" + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "type": "text", + "content": "14404" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 90, + 70, + 246, + 227 + ], + "blocks": [ + { + "bbox": [ + 90, + 70, + 246, + 227 + ], + "lines": [ + { + "bbox": [ + 90, + 70, + 246, + 227 + ], + "spans": [ + { + "bbox": [ + 90, + 70, + 246, + 227 + ], + "type": "table", + "html": "
ModelsCLIP-I ↑CLIP-T ↑
unimodal generation models
MUSE [13]-0.320
Imagen [64]-0.270
DALL-E 2 † [61]-0.314
DALL-E 3 † [9]-0.320
SDv1.5 [62]0.6670.302
SDXL [58]0.6740.310
multimodal generation models
GILL [37]0.684-
SEED [27]0.682-
Emu [71]0.6560.286
Emu2-Gen0.6860.297
", + "image_path": "91e1531e7a5b2390ff9e39991f7c1a04f99612797b7a2a0634fe857b3343c96a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 52, + 303, + 285, + 460 + ], + "blocks": [ + { + "bbox": [ + 46, + 236, + 288, + 291 + ], + "lines": [ + { + "bbox": [ + 46, + 236, + 288, + 291 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 288, + 291 + ], + "type": "text", + "content": "Table 4. Quantitative comparison of zero-shot text-to-image generation on MS-COCO [47] validation set. 30k samples are randomly sampled. " + }, + { + "bbox": [ + 46, + 236, + 288, + 291 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 46, + 236, + 288, + 291 + ], + "type": "text", + "content": "CLIP-T score is calculated on 4,096 samples. We also evaluate our image autoencoding results on MS-COCO which achieves a strong 0.907 CLIP-I score." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 303, + 285, + 460 + ], + "lines": [ + { + "bbox": [ + 52, + 303, + 285, + 460 + ], + "spans": [ + { + "bbox": [ + 52, + 303, + 285, + 460 + ], + "type": "table", + "html": "
MethodsDINO ↑CLIP-I ↑CLIP-T ↑
Real Images (Oracle)0.7740.885-
Fine-Tuning
Textual Inversion [24]0.5690.7800.255
DreamBooth [63]0.6680.8030.305
BLIP-Diffusion [42]0.6700.8050.302
Test Time Tuning Free
Re-Imagen* [16]0.6000.7400.270
SuTI [17]0.7410.8190.304
BLIP-Diffusion* [42]0.5940.7790.300
Kosmos-G* (single image input)0.6940.8470.287
Emu2-Gen * (single image input)0.7660.8500.287
", + "image_path": "6d08b44ae618af4ac578ab497c003150a35e2e283d9dddf5f58694f4109f72b3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 468, + 287, + 501 + ], + "lines": [ + { + "bbox": [ + 46, + 468, + 287, + 501 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 287, + 501 + ], + "type": "text", + "content": "Table 5. Quantitative comparison of zero-shot single-entity subject-driven generation on DreamBench. * denotes zero-shot methods." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 523, + 287, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 523, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 46, + 523, + 287, + 656 + ], + "type": "text", + "content": "DINO [12] and CLIP-I [59] to evaluate subject fidelity, and CLIP-T [59] to evaluate text fidelity, aligning with the methodology established by DreamBooth. Notably, Emu2-Gen excels in subject fidelity, as evidenced by its superior performance on DINO and CLIP-I metrics compared to methods like BLIP-Diffusion and Kosmos-G. Emu2-Gen impressively reconstructs subjects with just one image input in zero-shot setting, demonstrating superior subject fidelity through powerful visual decoding. Further illustrative cases are provided in the supplementary, showcasing Emu2-Gen's proficiency in multi-entity generation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 669, + 134, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 134, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 134, + 681 + ], + "type": "text", + "content": "4. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": "Large Multimodal Models. Recent years have witnessed the rapid growth of large multimodal models [5, 19, 31," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 277 + ], + "type": "text", + "content": "71]. CLIP [59] pioneered the learning of LMMs with a contrastive learning objective on massive image-text pair data. Flamingo [5] and Kosmos [31, 57] exhibit promising zero-shot and few-shot multi-modal understanding performance by training on large-scale image-text interleaved data. With the remarkable progress in open-sourced LLMs, [42, 43, 49, 94] show promising results by connecting vision encoders and LLMs with a small intermediate model. A school of successive efforts [69, 76, 84, 90, 91] further improves visual instruction tuning with better overall training pipelines [6, 40], grounding annotations [14, 15, 85, 92], and extra tasks [6]. There are early studies on training more unified large multimodal models [23, 27, 71, 86] that are capable of performing visual understanding and generation simultaneously. In this paper, we further explore the distinct solution proposed in Emu [71]: learning large multimodal models with generative objectives on both texts and images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 284, + 545, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 284, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 304, + 284, + 545, + 464 + ], + "type": "text", + "content": "In-Context Learning. Recent advancements in large language models [11, 21] underscore their capacity for in-context learning [11]. This phenomenon, particularly evident as LLMs scale up in size and data, has been exploited for complex challenges such as mathematical reasoning [81], signaling new emergent ability in model behavior [80]. Flamingo [5] integrates visual inputs to LLMs, enabling the in-context learning of visual-linguistic tasks such as image captioning and OCR through language-based interfacing. Painter [78] and SegGPT [79] conduct an early study of visual in-context learning. Inspired by the emerging abilities of large language models, in this work we study the problem of multimodal in-context learning by scaling up generative multimodal models and demonstrating strong results in broad understanding and generation tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 486, + 378, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 486, + 378, + 498 + ], + "spans": [ + { + "bbox": [ + 306, + 486, + 378, + 498 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "We present a 37 billion-parameter generative multimodal model Emu2 that shows strong performance and versatility on many multimodal tasks in the in-context settings. Emu2 serves as a base model and a general-purpose interface for a variety of multimodal tasks. We demonstrate state-of-the-art results on a broad range of benchmarks of multimodal understanding and generation. Specifically, our model largely surpasses prior work on the lately proposed LMM benchmarks that require more advanced capability compared to classic academic benchmarks. Emu2 also shows remarkable capability of controllable visual generation in multimodal context, e.g., subject-/text-grounded generation. Additionally, we review the limitations and broader social impact of Emu2. Despite discussed weaknesses, these results suggest that generative multimodal model at scale may be an important step towards the development of adaptable, general multimodal systems." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "type": "text", + "content": "14405" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 112 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 112 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 112 + ], + "type": "text", + "content": "[1] Laion-aesthetics. https://laion.ai/blog/laion-aesthetics/, .4,5,2,3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 114, + 286, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 286, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 286, + 135 + ], + "type": "text", + "content": "[2] Laion coco: 600m synthetic captions from laion2b-en. https://laion.ai/blog/laion-coco/,4,2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 137, + 286, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 286, + 167 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 286, + 167 + ], + "type": "text", + "content": "[3] Laion-high-resolution. https://huggingface.co/datasets/laion/laion-high-resolution, 5, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 170, + 240, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 170, + 240, + 180 + ], + "spans": [ + { + "bbox": [ + 53, + 170, + 240, + 180 + ], + "type": "text", + "content": "[4] Sharegpt. https://sharegpt.com/.5,2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 182, + 286, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 286, + 246 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 286, + 246 + ], + "type": "text", + "content": "[5] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 1, 5, 6, 8, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 248, + 286, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 248, + 286, + 300 + ], + "spans": [ + { + "bbox": [ + 53, + 248, + 286, + 300 + ], + "type": "text", + "content": "[6] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 303, + 286, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 303, + 286, + 356 + ], + "spans": [ + { + "bbox": [ + 53, + 303, + 286, + 356 + ], + "type": "text", + "content": "[7] Shuai Bai, Shusheng Yang, Jinze Bai, Peng Wang, Xingxuan Zhang, Junyang Lin, Xinggang Wang, Chang Zhou, and Jingren Zhou. Touchstone: Evaluating vision-language models by language models. arXiv preprint arXiv:2308.16890, 2023. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 358, + 286, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 358, + 286, + 411 + ], + "spans": [ + { + "bbox": [ + 53, + 358, + 286, + 411 + ], + "type": "text", + "content": "[8] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1728-1738, 2021. 3, 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 415, + 286, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 415, + 286, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 415, + 286, + 468 + ], + "type": "text", + "content": "[9] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, Wesam Manassra, Prafulla Dhariwal, Casey Chu, Yunxin Jiao, and Aditya Ramesh. Improving image generation with better captions. 2023. 5, 7, 8, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 470, + 286, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 470, + 286, + 523 + ], + "spans": [ + { + "bbox": [ + 47, + 470, + 286, + 523 + ], + "type": "text", + "content": "[10] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 5, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 525, + 286, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 525, + 286, + 579 + ], + "spans": [ + { + "bbox": [ + 47, + 525, + 286, + 579 + ], + "type": "text", + "content": "[11] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 581, + 286, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 581, + 286, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 581, + 286, + 635 + ], + "type": "text", + "content": "[12] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 636, + 286, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 636, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 47, + 636, + 286, + 689 + ], + "type": "text", + "content": "[13] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 691, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 691, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 691, + 286, + 712 + ], + "type": "text", + "content": "[14] Chi Chen, Ruoyu Qin, Fuwen Luo, Xiaoyue Mi, Peng Li, Maosong Sun, and Yang Liu. Position-enhanced visual" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 544, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 544, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 544, + 95 + ], + "type": "text", + "content": "instruction tuning for multimodal large language models. arXiv preprint arXiv:2308.13437, 2023. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 98, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 98, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 98, + 545, + 140 + ], + "type": "text", + "content": "[15] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 6, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 143, + 545, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 143, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 307, + 143, + 545, + 175 + ], + "type": "text", + "content": "[16] Wenhu Chen, Hexiang Hu, Chitwan Sahara, and William W Cohen. Re-imagen: Retrieval-augmented text-to-image generator. arXiv preprint arXiv:2209.14491, 2022. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 178, + 545, + 221 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 178, + 545, + 221 + ], + "spans": [ + { + "bbox": [ + 307, + 178, + 545, + 221 + ], + "type": "text", + "content": "[17] Wenhu Chen, Hexiang Hu, Yandong Li, Nataniel Rui, Xuhui Jia, Ming-Wei Chang, and William W Cohen. Subject-driven text-to-image generation via apprenticeship learning. arXiv preprint arXiv:2304.00186, 2023. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 224, + 545, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 224, + 545, + 267 + ], + "spans": [ + { + "bbox": [ + 307, + 224, + 545, + 267 + ], + "type": "text", + "content": "[18] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 4, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 270, + 545, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 270, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 307, + 270, + 545, + 324 + ], + "type": "text", + "content": "[19] Xi Chen, Xiao Wang, Lucas Beyer, Alexander Kolesnikov, Jialin Wu, Paul Voigtlaender, Basil Mustafa, Sebastian Goodman, Ibrahim Alabdulmohsin, Piotr Padlewski, et al. Pali-3 vision language models: Smaller, faster, stronger. arXiv preprint arXiv:2310.09199, 2023. 1, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 327, + 543, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 543, + 349 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 543, + 349 + ], + "type": "text", + "content": "[20] Luke Chesser and Timothy Carbone. Unsplash. https://github.com/unsplash/datasets.5,3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 352, + 545, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 352, + 545, + 406 + ], + "spans": [ + { + "bbox": [ + 307, + 352, + 545, + 406 + ], + "type": "text", + "content": "[21] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 1, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 408, + 545, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 408, + 545, + 463 + ], + "spans": [ + { + "bbox": [ + 307, + 408, + 545, + 463 + ], + "type": "text", + "content": "[22] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Albert Li, Pascale Fung, and Steven C. H. Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500, 2023. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 465, + 545, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 465, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 307, + 465, + 545, + 529 + ], + "type": "text", + "content": "[23] Runpei Dong, Chunrui Han, Yuang Peng, Zekun Qi, Zheng Ge, Jinrong Yang, Liang Zhao, Jianjian Sun, Hongyu Zhou, Haoran Wei, Xiangwen Kong, Xiangyu Zhang, Kaisheng Ma, and Li Yi. Dreamllm: Synergistic multimodal comprehension and creation. arXiv preprint arXiv:2309.11499, 2023.8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 533, + 545, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 533, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 307, + 533, + 545, + 586 + ], + "type": "text", + "content": "[24] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 590, + 545, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 643 + ], + "type": "text", + "content": "[25] Zhe Gan, Linjie Li, Chunyuan Li, Lijuan Wang, Zicheng Liu, Jianfeng Gao, et al. Vision-language pre-training: Basics, recent advances, and future trends. Foundations and Trends® in Computer Graphics and Vision, 14(3-4):163-352, 2022. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "type": "text", + "content": "[26] Leo Gao, Stella Biderman, Sid Black, Laurence Golding, Travis Hoppe, Charles Foster, Jason Phang, Horace He, Anish Thite, Noa Nabeshima, Shawn Presser, and Connor Leahy. The Pile: An 800gb dataset of diverse text for language modeling. arXiv preprint arXiv:2101.00027, 2020. 3, 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 734, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 734, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 734, + 318, + 742 + ], + "type": "text", + "content": "14406" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[27] Yuying Ge, Yixiao Ge, Ziyun Zeng, Xintao Wang, and Ying Shan. Planting a seed of vision in large language model. arXiv preprint arXiv:2307.08041, 2023. 4, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 173 + ], + "type": "text", + "content": "[28] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 4, 6, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 239 + ], + "type": "text", + "content": "[29] Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. Vizwiz grand challenge: Answering visual questions from blind people. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3608-3617, 2018. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 241, + 287, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 287, + 263 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 287, + 263 + ], + "type": "text", + "content": "[30] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 319 + ], + "type": "text", + "content": "[31] Shaohan Huang, Li Dong, Wenhui Wang, Yaru Hao, Saksham Singhal, Shuming Ma, Tengchao Lv, Lei Cui, Owais Khan Mohammed, Qiang Liu, et al. Language is not all you need: Aligning perception with language models. arXiv preprint arXiv:2302.14045, 2023. 6, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 320, + 287, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 287, + 374 + ], + "type": "text", + "content": "[32] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 4, 6, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 376, + 287, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 287, + 420 + ], + "type": "text", + "content": "[33] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in Neural Information Processing Systems, 35:26565-26577, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 422, + 287, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 476 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 476 + ], + "type": "text", + "content": "[34] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014. 5, 6, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 478, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 478, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 478, + 287, + 533 + ], + "type": "text", + "content": "[35] Douwe Kiela, Hamed Firooz, Aravind Mohan, Vedanuj Goswami, Amanpreet Singh, Pratik Ringshia, and Davide Testuggine. The hateful memes challenge: Detecting hate speech in multimodal memes. Advances in neural information processing systems, 33:2611-2624, 2020. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 534, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 287, + 578 + ], + "type": "text", + "content": "[36] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 5, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 579, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 287, + 612 + ], + "type": "text", + "content": "[37] Jing Yu Koh, Daniel Fried, and Ruslan Salakhutdinov. Generating images with multimodal language models. arXiv preprint arXiv:2305.17216, 2023. 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 689 + ], + "type": "text", + "content": "[38] Hugo Laurençon, Lucile Saulnier, Léo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, Thomas Wang, Siddharth Karamcheti, Alexander M Rush, Douwe Kiela, et al. Obelics: An open web-scale filtered dataset of interleaved image-text documents. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[39] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking mul" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "timodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 95, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 545, + 139 + ], + "type": "text", + "content": "[40] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726, 2023.8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 140, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 183 + ], + "type": "text", + "content": "[41] Chunyuan Li, Zhe Gan, Zhengyuan Yang, Jianwei Yang, Linjie Li, Lijuan Wang, and Jianfeng Gao. Multimodal foundation models: From specialists to general-purpose assistants. arXiv preprint arXiv:2309.10020, 1(2):2, 2023. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 184, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 184, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 307, + 184, + 545, + 227 + ], + "type": "text", + "content": "[42] Dongxu Li, Junnan Li, and Steven CH Hoi. Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. arXiv preprint arXiv:2305.14720, 2023. 6, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 228, + 545, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 545, + 281 + ], + "type": "text", + "content": "[43] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 283, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 283, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 307, + 283, + 545, + 326 + ], + "type": "text", + "content": "[44] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 5, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 327, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 545, + 381 + ], + "type": "text", + "content": "[45] Lei Li, Yuwei Yin, Shicheng Li, Liang Chen, Peiyi Wang, Shuhuai Ren, Mukai Li, Yazheng Yang, Jingjing Xu, Xu Sun, et al. M3it: A large-scale dataset towards multimodal multilingual instruction tuning. arXiv preprint arXiv:2306.04387, 2023. 4, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 383, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 383, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 383, + 545, + 437 + ], + "type": "text", + "content": "[46] Xin Li, Wenqing Chu, Ye Wu, Weihang Yuan, Fanglong Liu, Qi Zhang, Fu Li, Haocheng Feng, Errui Ding, and Jingdong Wang. Videogen: A reference-guided latent diffusion approach for high definition text-to-video generation. arXiv preprint arXiv:2309.00398, 2023. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 437, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 437, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 307, + 437, + 545, + 503 + ], + "type": "text", + "content": "[47] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 504, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 504, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 307, + 504, + 545, + 536 + ], + "type": "text", + "content": "[48] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 4, 6, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 537, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 537, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 307, + 537, + 545, + 568 + ], + "type": "text", + "content": "[49] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 5, 8, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 570, + 545, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 570, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 307, + 570, + 545, + 592 + ], + "type": "text", + "content": "[50] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 3, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 593, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 593, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 307, + 593, + 545, + 657 + ], + "type": "text", + "content": "[51] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems, 35:2507–2521, 2022. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "type": "text", + "content": "[52] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 5, 6, 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "type": "text", + "content": "14407" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 128 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 128 + ], + "type": "text", + "content": "[53] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 3195-3204, 2019. 4, 6, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 129, + 285, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 129, + 285, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 129, + 285, + 150 + ], + "type": "text", + "content": "[54] Midjourney. Midjourney. https://www.midjourney.com. 5, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 152, + 286, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 152, + 286, + 185 + ], + "spans": [ + { + "bbox": [ + 49, + 152, + 286, + 185 + ], + "type": "text", + "content": "[55] Leonardo Nicoletti and Dina Bass. Humans are biased: Generative ai is even worse. *Bloomberg Technology+ Equality*. Accessed June, 23:2023, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 186, + 287, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 186, + 287, + 230 + ], + "spans": [ + { + "bbox": [ + 49, + 186, + 287, + 230 + ], + "type": "text", + "content": "[56] Xichen Pan, Li Dong, Shaohan Huang, Zhiliang Peng, Wenhu Chen, and Furu Wei. Kosmos-g: Generating images in context with multimodal large language models. arXiv preprint arXiv:2310.02992, 2023. 7, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 232, + 287, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 232, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 49, + 232, + 287, + 275 + ], + "type": "text", + "content": "[57] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 3, 5, 6, 8, 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 277, + 287, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 277, + 287, + 320 + ], + "spans": [ + { + "bbox": [ + 49, + 277, + 287, + 320 + ], + "type": "text", + "content": "[58] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023. 1, 3, 4, 8, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 322, + 287, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 322, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 49, + 322, + 287, + 387 + ], + "type": "text", + "content": "[59] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 388, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 388, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 49, + 388, + 287, + 453 + ], + "type": "text", + "content": "[60] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 456, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 456, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 49, + 456, + 287, + 498 + ], + "type": "text", + "content": "[61] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022.8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 501, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 501, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 49, + 501, + 287, + 555 + ], + "type": "text", + "content": "[62] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 4, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 557, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 557, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 49, + 557, + 287, + 622 + ], + "type": "text", + "content": "[63] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 7, 8, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 624, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 624, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 49, + 624, + 287, + 689 + ], + "type": "text", + "content": "[64] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 691, + 287, + 713 + ], + "type": "text", + "content": "[65] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 116 + ], + "type": "text", + "content": "Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. arXiv preprint arXiv:2210.08402, 2022. 3, 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 118, + 545, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 545, + 182 + ], + "type": "text", + "content": "[66] Oleksii Sidorov, Ronghang Hu, Marcus Rohrbach, and Amanpreet Singh. Textcaps: a dataset for image captioning with reading comprehension. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 742-758. Springer, 2020. 4, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 184, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 545, + 237 + ], + "type": "text", + "content": "[67] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 239, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 293 + ], + "type": "text", + "content": "[68] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 4, 6, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 294, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 545, + 326 + ], + "type": "text", + "content": "[69] Yixuan Su, Tian Lan, Huayang Li, Jialu Xu, Yan Wang, and Deng Cai. Pandagpt: One model to instruction-follow them all. arXiv preprint arXiv:2305.16355, 2023. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 327, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 360 + ], + "type": "text", + "content": "[70] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 361, + 545, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 361, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 308, + 361, + 545, + 414 + ], + "type": "text", + "content": "[71] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative pretraining in multimodality. arXiv preprint arXiv:2307.05222, 2023. 3, 4, 6, 7, 8, 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 415, + 545, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 415, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 308, + 415, + 545, + 458 + ], + "type": "text", + "content": "[72] Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B Hashimoto. Stanford alpaca: An instruction-following llama model, 2023. 5, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 460, + 545, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 535 + ], + "type": "text", + "content": "[73] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 537, + 545, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 545, + 569 + ], + "type": "text", + "content": "[74] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 571, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 571, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 571, + 545, + 635 + ], + "type": "text", + "content": "[75] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In International Conference on Machine Learning, pages 23318-23340. PMLR, 2022. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 636, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 689 + ], + "type": "text", + "content": "[76] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, et al. Vision: Large language model is also an open-ended decoder for vision-centric tasks. arXiv preprint arXiv:2305.11175, 2023. 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[77] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 733, + 317, + 741 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 733, + 317, + 741 + ], + "spans": [ + { + "bbox": [ + 294, + 733, + 317, + 741 + ], + "type": "text", + "content": "14408" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "Song, et al. Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079, 2023. 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 288, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 288, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 288, + 149 + ], + "type": "text", + "content": "[78] Xinlong Wang, Wen Wang, Yue Cao, Chunhua Shen, and Tiejun Huang. Images speak in images: A generalist painter for in-context visual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6830-6839, 2023. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 151, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 151, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 151, + 287, + 194 + ], + "type": "text", + "content": "[79] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Segmenting everything in context. arXiv preprint arXiv:2304.03284, 2023. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 195, + 287, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 195, + 287, + 248 + ], + "spans": [ + { + "bbox": [ + 48, + 195, + 287, + 248 + ], + "type": "text", + "content": "[80] Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, et al. Emergent abilities of large language models. arXiv preprint arXiv:2206.07682, 2022. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 251, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 251, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 251, + 287, + 304 + ], + "type": "text", + "content": "[81] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 35:24824-24837, 2022. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 305, + 287, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 305, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 305, + 287, + 360 + ], + "type": "text", + "content": "[82] Dejing Xu, Zhou Zhao, Jun Xiao, Fei Wu, Hanwang Zhang, Xiangnan He, and Yueting Zhuang. Video question answering via gradually refined attention over appearance and motion. In Proceedings of the 25th ACM international conference on Multimedia, pages 1645-1653, 2017. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 361, + 288, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 361, + 288, + 404 + ], + "spans": [ + { + "bbox": [ + 48, + 361, + 288, + 404 + ], + "type": "text", + "content": "[83] Charig Yang, Weidi Xie, and Andrew Zisserman. It's about time: Analog clock reading in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2508-2517, 2022. 5, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 405, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 405, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 405, + 287, + 459 + ], + "type": "text", + "content": "[84] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. MPLug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 460, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 287, + 514 + ], + "type": "text", + "content": "[85] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704, 2023. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 515, + 287, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 515, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 48, + 515, + 287, + 569 + ], + "type": "text", + "content": "[86] Lili Yu, Bowen Shi, Ramakanth Pasunuru, Benjamin Muller, Olga Golovneva, Tianlu Wang, Arun Babu, Binh Tang, Brian Karrer, Shelly Sheynin, et al. Scaling autoregressive multimodal models: Pretraining and instruction tuning. arXiv preprint arXiv:2309.02591, 2023. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 571, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 571, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 571, + 287, + 613 + ], + "type": "text", + "content": "[87] Qiying Yu, Quan Sun, Xiaosong Zhang, Yufeng Cui, Fan Zhang, Xinlong Wang, and Jingjing Liu. Capsfusion: Rethinking image-text data at scale. arXiv preprint arXiv:2310.20550, 2023. 3, 5, 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 615, + 287, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 615, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 615, + 287, + 658 + ], + "type": "text", + "content": "[88] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490, 2023. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[89] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. arXiv preprint arXiv:2311.16502, 2023. 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 363 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[90] Ao Zhang, Hao Fei, Yuan Yao, Wei Ji, Li Li, Zhiyuan Liu, and Tat-Seng Chua. Transfer visual prompt generator across llms. arXiv preprint arXiv:2305.01278, 2023. 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 107, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 172 + ], + "type": "text", + "content": "[91] Pan Zhang, Xiaoyi Dong Bin Wang, Yuhang Cao, Chao Xu, Linke Ouyang, Zhiyuan Zhao, Shuangrui Ding, Songyang Zhang, Haodong Duan, Hang Yan, et al. Internlmxcomposer: A vision-language large model for advanced text-image comprehension and composition. arXiv preprint arXiv:2309.15112, 2023. 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 174, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 174, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 174, + 545, + 217 + ], + "type": "text", + "content": "[92] Shilong Zhang, Peize Sun, Shoufa Chen, Min Xiao, Wenqi Shao, Wenwei Zhang, Kai Chen, and Ping Luo. Gpt4roi: Instruction tuning large language model on region-of-interest. arXiv preprint arXiv:2307.03601, 2023. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 219, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 262 + ], + "type": "text", + "content": "[93] Yanzhe Zhang, Ruiyi Zhang, Jiuxiang Gu, Yufan Zhou, Nedim Lipka, Diyi Yang, and Tong Sun. Llavar: Enhanced visual instruction tuning for text-rich image understanding. arXiv preprint arXiv:2306.17107, 2023. 5, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 263, + 545, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 263, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 307, + 263, + 545, + 307 + ], + "type": "text", + "content": "[94] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 308, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 545, + 363 + ], + "type": "text", + "content": "[95] Wanrong Zhu, Jack Hessel, Anas Awadalla, Samir Yitzhak Gadre, Jesse Dodge, Alex Fang, Youngjae Yu, Ludwig Schmidt, William Yang Wang, and Yejin Choi. Multimodal c4: An open, billion-scale corpus of images interleaved with text. arXiv preprint arXiv:2304.06939, 2023. 3, 1" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 318, + 742 + ], + "type": "text", + "content": "14409" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Generative Powers of Ten/2ccba12d-1f5a-40cf-8686-79a45cb775c2_content_list.json b/2024/Generative Powers of Ten/2ccba12d-1f5a-40cf-8686-79a45cb775c2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a76613add07278cd60d9efdc1c8603d7bc31b46f --- /dev/null +++ b/2024/Generative Powers of Ten/2ccba12d-1f5a-40cf-8686-79a45cb775c2_content_list.json @@ -0,0 +1,1717 @@ +[ + { + "type": "text", + "text": "Generative Powers of Ten", + "text_level": 1, + "bbox": [ + 354, + 130, + 614, + 150 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiaojuan Wang $^{1}$ Janne Kontkanen $^{2}$ Brian Curless $^{1,2}$ Steven M. Seitz $^{1,2}$ Ira Kemelmacher-Shlizerman $^{1,2}$ Ben Mildenhall $^{2}$ Pratul Srinivasan $^{2}$ Dor Verbin $^{2}$ Aleksander Holynski $^{2,3}$", + "bbox": [ + 84, + 178, + 929, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Washington 2Google Research 3UC Berkeley powers-of-ten.github.io", + "bbox": [ + 238, + 218, + 779, + 258 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/50180d19baaad31e78afc0d2392daff08fa72b1cc31485e9c8c658d57328b5ff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 295, + 277, + 450 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c93968793eedac9115882213db7031d5e9f450d07e344f826fa6727edf97c705.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 295, + 480, + 446 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e926d347caa8dc1f46df55655fea69c5650f89e87e00ed42d2d14b33cc1e3832.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 296, + 684, + 446 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6dacc97b07d88436b55f860d7c262390d535f63ff68078a1ec7a77b362ba2f38.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 296, + 887, + 446 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ce99e3d73cecd198574053c0e8a5b6e7de7aa1f01eb02d0ebc2a7cce5f763700.jpg", + "image_caption": [ + "Figure 1. Given a series of prompts describing a scene at varying zoom levels, e.g., from a distant galaxy to the surface of an alien planet, our method uses a pre-trained text-to-image diffusion model to generate a continuously zooming video sequence." + ], + "image_footnote": [], + "bbox": [ + 78, + 462, + 885, + 503 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 558, + 313, + 574 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present a method that uses a text-to-image model to generate consistent content across multiple image scales, enabling extreme semantic zooms into a scene, e.g. ranging from a wide-angle landscape view of a forest to a macro shot of an insect sitting on one of the tree branches. We achieve this through a joint multi-scale diffusion sampling approach that encourages consistency across different scales while preserving the integrity of each individual sampling process. Since each generated scale is guided by a different text prompt, our method enables deeper levels of zoom than traditional super-resolution methods that may struggle to create new contextual structure at vastly different scales. We compare our method qualitatively with alternative techniques in image super-resolution and outpainting, and show that our method is most effective at generating consistent multi-scale content.", + "bbox": [ + 75, + 589, + 472, + 830 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 844, + 209, + 859 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in text-to-image models [3, 6, 7, 15, 18, 19, 29] have been transformative in enabling applications", + "bbox": [ + 76, + 869, + 468, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "like image generation from a single text prompt. But while digital images exist at a fixed resolution, the real world can be experienced at many different levels of scale. Few things exemplify this better than the classic 1977 short film \"Powers of Ten\", shown in Figure 2, which showcases the sheer magnitudes of scale that exist in the universe by visualizing a continuous zoom from the outermost depths of the galaxy to the cells inside our bodies1. Unfortunately, producing animations or interactive experiences like these has traditionally required trained artists and many hours of tedious labor—and although we might want to replace this process with a generative model, existing methods have not yet demonstrated the ability to generate consistent content across multiple zoom levels.", + "bbox": [ + 496, + 559, + 892, + 768 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unlike traditional super-resolution methods, which generate higher-resolution content conditioned on the pixels of the original image, extreme zooms expose entirely new structures, e.g., magnifying a hand to reveal its underlying skin cells. Generating such a zoom requires semantic knowledge of human anatomy. In this paper, we focus on solving this semantic zoom problem, i.e., enabling text-", + "bbox": [ + 496, + 770, + 893, + 878 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1https://www.youtube.com/watch?v=0fKBhvdjuy0", + "bbox": [ + 517, + 887, + 864, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7173", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b41583abfe9cc70afc96c0d8222ec83e1b3ea766bca1cde5ce5b124d9743105a.jpg", + "image_caption": [ + "Figure 2. Powers of Ten (1977) This documentary film illustrates the relative scale of the universe as a single shot that gradually zooms out from a human to the universe, and then back again to the microscopic molecular level." + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 153, + 209 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c1b100c55418feae215da7ec2658f518d55eab12e6264a450a6ff23b0e3825e3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 156, + 89, + 232, + 209 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6817a7b802fd6130cd15068c3eb14534728cff9193d25731e9831e46a11ad452.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 235, + 89, + 310, + 209 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4bdda37128e8e51968fb42bd45bca02cb22621797d4dcf221c77d66f0ceb5912.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 313, + 89, + 390, + 209 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ac2ea4825cd227dcd1d72290dc3d4886875eab89d76b95a8d7aa1218d3d7d1ca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 393, + 89, + 467, + 209 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "conditioned multi-scale image generation, to create Powers of Ten-like zoom videos. As input, our method expects a series of text prompts that describe different scales of the scene, and produces as output a multi-scale image representation that can be explored interactively or rendered to a seamless zooming video. These text prompts can be user-defined (allowing for creative control over the content at different zoom levels) or crafted with the help of a large language model (e.g., by querying the model with an image caption and a prompt like \"describe what might you see if you zoomed in by $2x$ )", + "bbox": [ + 75, + 280, + 470, + 445 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "At its core, our method relies on a joint sampling algorithm that uses a set of parallel diffusion sampling processes distributed across zoom levels. These sampling processes are coordinated to be consistent through an iterative frequency-band consolidation process, in which intermediate image predictions are consistently combined across scales. Unlike existing approaches that accomplish similar goals by repeatedly increasing the effective image resolution (e.g., through super-resolution or image outpainting), our sampling process jointly optimizes for the content of all scales at once, allowing for both (1) plausible images at each scale and (2) consistent content across scales. Furthermore, existing methods are limited in their ability to explore wide ranges of scale, since they rely primarily on the input image content to determine the added details at subsequent zoom levels. In many cases, image patches contain insufficient contextual information to inform detail at deeper (e.g., $10\\mathrm{x}$ or $100\\mathrm{x}$ ) zoom levels. On the other hand, our method grounds each scale in a text prompt, allowing for new structures and content to be conceived across extreme zoom levels. In our experiments, we compare our work qualitatively to these existing methods, and demonstrate that the zoom videos that our method produces are notably more consistent. Finally, we showcase a number of ways in which our algorithm can be used, e.g., by conditioning purely on text or grounding the generation in a known (real) image.", + "bbox": [ + 75, + 448, + 468, + 843 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Prior Work", + "text_level": 1, + "bbox": [ + 76, + 858, + 199, + 875 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Super-resolution and inpainting. Existing text-to-image", + "bbox": [ + 76, + 885, + 467, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "based super resolution models [1, 22] and outpainting models [1, 16, 20, 27] can be adapted to the zoom task as autoregressive processes, i.e., by progressively outpainting a zoomed-in image, or progressively super-resolving a zoomed-out image. One significant drawback of these approaches is that later-generated images have no influence on the previously generated ones, which can often lead to suboptimal results, as certain structures may be entirely incompatible with subsequent levels of detail, causing error accumulation across recurrent network applications.", + "bbox": [ + 496, + 90, + 890, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Perpetual view generation. Starting from a single view RGB image, perpetual view generation methods like Infinite Nature [11] and InfiniteNature-Zero [12] learn to generate unbounded flythrough videos of natural scenes. These methods differ from our generative zoom in two key ways: (1) they translate the camera in 3D, causing a \"fly-through\" effect with perspective effects, rather than the \"zoom in\" our method produces, and (2) they synthesize the fly-through starting from a single image by progressively inpainting unknown parts of novel views, whereas we generate the entire zoom sequence simultaneously and coherently across scales, with text-guided semantic control.", + "bbox": [ + 496, + 248, + 890, + 429 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion joint sampling for consistent generation. Recent research [2, 10, 28, 30] leverages pretrained diffusion models to generate arbitrary-sized images or panoramas from smaller pieces using joint diffusion processes. These processes involve concurrently generating these multiple images by merging their intermediate results within the sampling process. In particular, DiffCollage [30] introduces a factor graph formulation to express spatial constraints among these images, representing each image as a node, and overlapping areas with additional nodes. Each sampling step involves aggregating individual predictions based on the factor graph. For this to be possible, a given diffusion model needs to be finetuned for different factor nodes. Other works such as MultiDiffusion [2] reconcile different denoising steps by solving for a least squares optimal solution: i.e., averaging the diffusion model predictions at overlapping areas. However, none of these approaches can be applied to our problem, where our jointly sampled images have spatial correspondence at vastly different spatial scales.", + "bbox": [ + 496, + 436, + 890, + 737 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 500, + 762, + 637, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion models [5, 8, 23-26] generate images from random noise through a sequential sampling process. This sampling process reverses a destructive process that gradually adds Gaussian noise on a clean image $\\mathbf{x}$ . The intermediate noisy image at time step $t$ is expressed as:", + "bbox": [ + 496, + 790, + 890, + 867 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {t} = \\alpha_ {t} \\mathbf {x} + \\sigma_ {t} \\boldsymbol {\\epsilon} _ {t},\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 887, + 754, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7174", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\epsilon_{t} \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ is a standard Gaussian noise, and $\\alpha_{t}$ and $\\sigma_{t}$ define a fixed noise schedule, with larger $t$ corresponding to more noise. A diffusion model is a neural network $\\epsilon_{\\theta}$ that predicts the approximate clean image $\\hat{\\mathbf{x}}$ directly, or equivalently the added noise $\\epsilon_{t}$ in $\\mathbf{z}_{t}$ . The network is trained with the loss:", + "bbox": [ + 75, + 90, + 472, + 181 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\boldsymbol {\\theta}) = \\mathbb {E} _ {t \\sim U [ 1, T ], \\boldsymbol {\\epsilon} _ {t} \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I})} [ w (t) \\| \\boldsymbol {\\epsilon} _ {\\boldsymbol {\\theta}} (\\mathbf {z} _ {t}; t, y) - \\boldsymbol {\\epsilon} _ {t} \\| _ {2} ^ {2} ],\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 196, + 455, + 215 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $y$ is an additional conditioning signal like text [16, 17, 21], and $w(t)$ is a weighting function typically set to 1 [8]. A standard choice for $\\epsilon_{\\theta}$ is a U-Net with self-attention and cross-attention operations attending to the conditioning $y$ .", + "bbox": [ + 75, + 229, + 468, + 304 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Once the diffusion model is trained, various sampling methods [8, 13, 24] are designed to sample efficiently from the model, starting from pure noise $\\mathbf{z}_T \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})$ and iteratively denoising it to a clean image. These sampling methods often rely on classifier-free guidance [8], a process which uses a linear combination of the text-conditional and unconditional predictions to achieve better adherence to the conditioning signal:", + "bbox": [ + 75, + 306, + 470, + 428 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {\\epsilon}} _ {t} = (1 + \\omega) \\boldsymbol {\\epsilon} _ {\\theta} (\\mathbf {z} _ {t}; t, y) - \\omega \\boldsymbol {\\epsilon} _ {\\theta} (\\mathbf {z} _ {t}; t).\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 443, + 398, + 459 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This revised $\\hat{\\epsilon}_t$ is used as the noise prediction to update the noisy image $\\mathbf{z}_t$ . Given a noisy image and a noise prediction, the estimated clean image $\\hat{\\mathbf{x}}_t$ is computed as $\\hat{\\mathbf{x}}_t = (\\mathbf{z}_t - \\sigma_t\\hat{\\epsilon}_t) / \\alpha_t$ . The iterative update function in the sampling process depends on the sampler used; in this paper we use DDPM [8].", + "bbox": [ + 75, + 474, + 470, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 76, + 584, + 168, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $y_0, \\ldots, y_{N-1}$ be a series of prompts describing a single scene at varying, corresponding zoom levels $p_0, \\ldots, p_{N-1}$ forming a geometric progression, i.e., $p_i = p^i$ (we typically set $p$ to 2 or 4). Our objective is to generate a sequence of corresponding $H \\times W \\times C$ images $\\mathbf{x}_0, \\ldots, \\mathbf{x}_{N-1}$ from an existing, pre-trained, text-to-image diffusion model. We aim to generate the entire set of images jointly in a zoom-consistent way. This means that the image $\\mathbf{x}_i$ at any specific zoom level $p_i$ , should be consistent with the center $H / p \\times W / p$ crop of the zoomed-out image $\\mathbf{x}_{i-1}$ .", + "bbox": [ + 75, + 611, + 468, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose a multi-scale joint sampling approach and a corresponding zoom stack representation that gets updated in the diffusion-based sampling process. In Sec. 4.1, we introduce our zoom stack representation and the process that allows us to render it into an image at any given zoom level. In Sec. 4.2, we present an approach for consolidating multiple diffusion estimates into this representation in a consistent way. Finally, in Sec. 4.3, we show how these components are used in the complete sampling process.", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/09ddcebdb8bed0e3cd4b3164df5c681211e845a15a1131301bb9faf1e7d3cbf0.jpg", + "image_caption": [ + "Figure 3. Zoom stack. Our representation consists of $N$ layer images $L_{i}$ of constant resolution (left). These layers are arranged in a pyramid-like structure, with layers representing finer details corresponding to a smaller spatial extent (middle). These layers are composited to form an image at any zoom level (right)." + ], + "image_footnote": [], + "bbox": [ + 498, + 87, + 893, + 224 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Zoom Stack Representation", + "text_level": 1, + "bbox": [ + 498, + 321, + 750, + 338 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our zoom stack representation, which we denote by $\\mathcal{L} = (L_0,\\dots,L_{N - 1})$ , is designed to allow rendering images at any zoom level $p_0,\\ldots ,p_{N - 1}$ . The representation, illustrated in Fig. 3, contains $N$ images of shape $H\\times W$ , one for each zoom level, where the $i$ th image $L_{i}$ stores the pixels corresponding to the $i$ th zoom level $p_i$ .", + "bbox": [ + 496, + 345, + 893, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image rendering. The rendering operator, which we denote by $\\Pi_{\\mathrm{image}}(\\mathcal{L};i)$ , takes a zoom stack $\\mathcal{L}$ and returns the image at the $i$ th zoom level $p_i = p^i$ . We denote by $\\mathcal{D}_i(\\mathbf{x})$ the operator which downscales the image $\\mathbf{x}$ by factor $p_i$ , and zero-pads the image back to size $H\\times W$ ; and we denote by $M_{i}$ the corresponding $H\\times W$ binary image which has value 1 at the center $H / p_{i}\\times W / p_{i}$ patch and value 0 at padded pixels. The operator $\\mathcal{D}_i$ operates by prefiltering the image with a truncated Gaussian kernel of size $p_i\\times p_i$ and resampling with a stride of $p_i$ . As described in Alg. 1, an image $\\mathbf{x}_i$ at the $i$ th zoom level is rendered by starting with $L_{i}$ , and iteratively replacing its central $H / p_j\\times W / p_j$ crop with $\\mathcal{D}_{j - i}(L_j)$ , for $j = i + 1,\\dots,N - 1$ . (In Alg. 1 we denote by $\\odot$ the elementwise multiplication of a binary mask $M$ with an image.) This process guarantees that rendering at different zoom levels will be consistent at overlapping central regions.", + "bbox": [ + 496, + 441, + 892, + 699 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Noise rendering. At every denoising iteration of DDPM [8], each pixel is corrupted by globally-scaled i.i.d. Gaussian noise $\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ . Since we would like images rendered at different zoom levels to be consistent, it is essential to make sure the added noise is also consistent, with overlapping region across different zoom levels sharing the same noise structure. Therefore, we use a rendering operator similar to $\\Pi_{\\mathrm{image}}$ which converts a set of independent noise images, $\\mathcal{E} = (E_0,\\dots,E_{N - 1})$ into a single zoom-consistent noise $\\epsilon_{i} = \\Pi_{\\mathrm{noise}}(\\mathcal{E};i)$ . However, because downsampling involves prefiltering, which modifies the statistics of the resulting noise, we upscale the $j$ th down-scaled noise component by $p_j / p_i$ to preserve the variance,", + "bbox": [ + 496, + 704, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7175", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d9d3fe23998ad3633f582da5ce01a14cad339d4d2d4f5c59f77e16dbe45f1494.jpg", + "image_caption": [ + "Figure 4. Overview of a single sampling step. (1) Noisy images $\\mathbf{z}_{i,t}$ from each zoom level, along with the respective prompts $y_{i}$ are simultaneously fed into the same pretrained diffusion model, returning estimates of the corresponding clean images $\\hat{\\mathbf{x}}_{i,t}$ . These images may have inconsistent estimates for the overlapping regions that they all observe. We employ multi-resolution blending to fuse these regions into a consistent zoom stack $\\mathcal{L}_t$ and re-render the different zoom levels from the consistent representation. These re-rendered images $\\Pi_{\\mathrm{image}}(\\mathcal{L}_t;i)$ are then used as the clean image estimates in the DDPM sampling step." + ], + "image_footnote": [], + "bbox": [ + 84, + 87, + 893, + 262 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ensuring that the noise satisfies the standard Gaussian distribution assumption, i.e., that $\\epsilon_{i} = \\Pi_{\\mathrm{noise}}(\\mathcal{E};i)\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ for all levels $i$ .", + "bbox": [ + 75, + 357, + 470, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 Image and noise rendering at scale $i$ .", + "text_level": 1, + "bbox": [ + 78, + 412, + 410, + 429 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Set $\\mathbf{x} \\gets L_i, \\epsilon \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})$", + "2: for $j = i + 1, \\ldots, N - 1$ do", + "3: $\\mathbf{x} \\gets M_{j - i} \\odot \\mathcal{D}_{j - i}(L_j) + (1 - M_{j - i}) \\odot \\mathbf{x}$", + "4: $\\epsilon \\gets (p_j / p_i)M_{j - i}\\odot \\mathcal{D}_{j - i}(E_j) + (1 - M_{j - i})\\odot \\epsilon$", + "5: end for", + "6: return x, $\\epsilon$" + ], + "bbox": [ + 86, + 438, + 465, + 530 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Multi-resolution blending", + "text_level": 1, + "bbox": [ + 76, + 560, + 313, + 577 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Equipped with a method for rendering a zoom stack and sampling noise at any given zoom level, we now describe a mechanism for integrating multiple observations of the same scene $\\mathbf{x}_0,\\dots ,\\mathbf{x}_{N - 1}$ at varying zoom levels $p_0,\\dots ,p_{N - 1}$ into a consistent zoom stack $\\mathcal{L}$ . This process is a necessary component of the consistent sampling process, as the diffusion model applied at various zoom levels will produce inconsistent content in the overlapping regions. Specifically, the $j$ th zoom stack level $L_{j}$ is used in rendering multiple images at all zoom levels $i\\leq j$ and therefore its value should be consistent with multiple image observations (or diffusion model samples), namely $\\{\\mathbf{x}_i:i\\leq j\\}$ . The simplest possible solution to this is to naively average the overlapping regions across all observations. This approach, however, results in blurry zoom stack images, since coarser-scale observations of overlapping regions contain fewer pixels, and therefore only lower-frequency information.", + "bbox": [ + 75, + 583, + 468, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To solve this, we propose an approach we call multi-resolution blending, which uses Laplacian pyramids to selectively fuse the appropriate frequency bands of each ob", + "bbox": [ + 75, + 854, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e4e25f8f3dfa282aeb1252e2d1e9282814a45e534cffbfb0a1279fc907684e12.jpg", + "image_caption": [ + "Figure 5. Multi-resolution blending. We produce a consistent estimate for Layer $L_{i}$ in the zoom stack by merging the $H / p_{j} \\times W / p_{j}$ central region of the corresponding zoomed out images $\\mathbf{x}_j$ for $j \\leq i$ . This merging process involves (1) creating a Laplacian pyramid from each observation, and blending together the corresponding frequency bands to create a blended pyramid. This blended pyramid is recomposed into an image, which is used to update the layer $L_{i}$ ." + ], + "image_footnote": [], + "bbox": [ + 504, + 354, + 890, + 661 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "servation level, which prevents aliasing as well as overblurring. We show an outline of this process in Fig. 5. More concretely, to update the $i$ th layer in the zoom stack, we begin by cropping all samples $j \\geq i$ to match with the content of the $i$ th level, and rescaling them back to $H \\times W$ . We then analyze each of these $N - i - 1$ images into a Laplacian pyramid [4], and average across corresponding frequency bands", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7176", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(see Figure 5), resulting in an average Laplacian pyramid, which can be recomposed into an image and assigned to the $i$ th level of the zoom stack. This process is applied for each layer of the zoom stack $L_{i}$ , collecting from all further zoomed-out levels $j \\geq i$ .", + "bbox": [ + 75, + 90, + 472, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Multi-scale consistent sampling", + "text_level": 1, + "bbox": [ + 76, + 174, + 356, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our complete multi-scale joint sampling process is shown in Alg. 2. Fig. 4 illustrates a single sampling step $t$ : Noisy images $\\mathbf{z}_{i,t}$ in each zoom level along with the respective prompt $y_{i}$ are fed into the pretrained diffusion model in parallel to predict the noise $\\hat{\\epsilon}_{i,t-1}$ , and thus to compute the estimated clean images $\\hat{\\mathbf{x}}_{i,t}$ . Equipped with our multi-resolution blending technique, the clean images are consolidated into a zoom stack, which is then rendered at all zoom levels, yielding consistent images $\\Pi_{\\mathrm{image}}(\\mathcal{L}_t; i)$ . These images are then used in a DDPM update step along with the input $\\mathbf{z}_t$ to compute the next $\\mathbf{z}_{t-1}$ .", + "bbox": [ + 75, + 196, + 473, + 364 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 2 Multi-scale joint sampling." + ], + "code_body": "1: Set $\\mathcal{L}_T\\gets 0,\\mathbf{z}_{i,T}\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I}),\\forall i = 0,\\dots ,N - 1$ \n2: for $t = T,\\ldots ,1$ do \n3: $\\mathcal{E}\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ \n4: parfor $i = 0,\\dots ,N - 1$ do \n5: $\\mathbf{x}_{i,t} = \\Pi_{\\mathrm{image}}(\\mathcal{L}_t;i)$ \n6: $\\epsilon_{i} = \\Pi_{\\mathrm{noise}}(\\mathcal{E};i)$ \n7: $\\mathbf{z}_{i,t - 1} = \\mathrm{DDPM\\_update}(\\mathbf{z}_{i,t},\\mathbf{x}_{i,t},\\boldsymbol {\\epsilon}_i)$ \n8: $\\hat{\\epsilon}_{i,t - 1} = (1 + \\omega)\\epsilon_{\\theta}(\\mathbf{z}_{i,t - 1};t - 1,y_i)$ \n9: $-\\omega \\epsilon_{\\theta}(\\mathbf{z}_{i,t - 1};t - 1)$ \n10: $\\hat{\\mathbf{x}}_{i,t - 1} = (\\mathbf{z}_{i,t - 1} - \\sigma_{t - 1}\\hat{\\epsilon}_{i,t - 1}) / \\alpha_{t - 1}$ \n11: end parfor \n12: $\\mathcal{L}_{t - 1}\\gets \\mathrm{Blending}(\\{\\hat{\\mathbf{x}}_{i,t - 1}\\}_{i = 0}^{N - 1})$ \n13: end for \n14: return $\\mathcal{L}_0$", + "bbox": [ + 84, + 401, + 426, + 616 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Photograph-based Zoom", + "text_level": 1, + "bbox": [ + 76, + 640, + 303, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition to using text prompts to generate the entire zoom stack from scratch, our approach can also generate a sequence zooming into an existing photograph. Given the most zoomed-out input image $\\xi$ , we still use Alg. 2, but we additionally update the denoised images to minimize the following loss function before every blending operation:", + "bbox": [ + 75, + 662, + 468, + 755 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\ell \\left(\\hat {\\mathbf {x}} _ {0, t}, \\dots , \\hat {\\mathbf {x}} _ {N - 1, t}\\right) = \\sum_ {i = 0} ^ {N - 1} \\left\\| \\mathcal {D} _ {i} \\left(\\hat {\\mathbf {x}} _ {i, t}\\right) - M _ {i} \\odot \\boldsymbol {\\xi}\\right) \\| _ {2} ^ {2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 761, + 468, + 804 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where, as we defined in Sec. 4.1, $\\mathcal{D}_i(\\mathbf{x})$ downscales the image $\\mathbf{x}$ by a factor $p_i$ and pads the result back to $H\\times W$ , and $M_{i}$ is a binary mask with 1 at the center $H / p_{i}\\times W / p_{i}$ square and 0 otherwise. Before every blending operation we apply 5 Adam [9] steps at a learning rate of 0.1. This simple optimization-based strategy encourages the estimated", + "bbox": [ + 75, + 809, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/eaa9a31aaa72a963efdd16d687e2839fe8ce4a46fdd5adb424edacad7d4b1bd7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 88, + 686, + 229 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/bfc0e925fba6e95d3b312b5fbdf6702b97a3e60de83a9470a800062b87fe7a1a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 229, + 681, + 376 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8175f2014ca7d774b099d36197995e27f52cdee3d5d7aec44d724608402981ee.jpg", + "image_caption": [ + "A girl is holding a maple leaf in front of her face." + ], + "image_footnote": [], + "bbox": [ + 687, + 88, + 870, + 229 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b18d46674b65b1a8312bc7f3a2622f3ac21de9cbbce1e2fd3dd331e690689ca7.jpg", + "image_caption": [ + "A closeup of the surface of skin of the back hand" + ], + "image_footnote": [], + "bbox": [ + 504, + 383, + 686, + 523 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9d0501a2ef72bfb900a737f51b36cbd64be308aefed0294f5981970f46ff2838.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 687, + 236, + 867, + 376 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/071d14903dd88940f2d07a591ca292f2797a5ed0eb784bd15d6837dcc05bec36.jpg", + "image_caption": [ + "Epidermal layer of multiple rows of tiny skin cells", + "A single round skin cell with its nucleus" + ], + "image_footnote": [], + "bbox": [ + 504, + 532, + 686, + 672 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0baf321a12cec849ef0f3b8528dd0a6f3bc1d93a751962df9949dc30a8c58b16.jpg", + "image_caption": [ + "A brightly colored autumn maple leaf" + ], + "image_footnote": [], + "bbox": [ + 687, + 383, + 867, + 523 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cee112abab610b2cdbf1820a059a5972c428f689928bf082b9631d8aeb6d84c1.jpg", + "image_caption": [ + "A nucleus within a skin cell", + "Figure 6. Selected images of our generated zoom sequences beginning with a provided real image. Left: Zoom from a man on a picnic blanket into the skin cells on his hand. Right: Zoom from a girl holding a leaf into the intricate vein patterns on the leaf. Face is blurred for anonymity." + ], + "image_footnote": [], + "bbox": [ + 504, + 680, + 686, + 819 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f43c5e6302dfec8775403418c6ab804b6e965741d30ad8c5902080f83fcc3f74.jpg", + "image_caption": [ + "Orange maple leaf texture with lots of veins", + "Macrophoto of the veins pattern on the maple leaf" + ], + "image_footnote": [], + "bbox": [ + 687, + 532, + 867, + 672 + ], + "page_idx": 4 + }, + { + "type": "aside_text", + "text": "Zoomed out", + "bbox": [ + 877, + 89, + 890, + 135 + ], + "page_idx": 4 + }, + { + "type": "aside_text", + "text": "Zoomed in", + "bbox": [ + 877, + 787, + 890, + 830 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7177", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/65391a8619a1c7df0f938363f722701c01afd365872e4cdad8631aa291f5f3e4.jpg", + "image_caption": [ + "Galactic core" + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 287, + 234 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3102531ab35c33709222e73438bac246320c93ff98847ce366eedad5af1a1bdf.jpg", + "image_caption": [ + "Satellite image of the Earth's surface" + ], + "image_footnote": [], + "bbox": [ + 287, + 88, + 473, + 233 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9f6fb647d84af58d6621c80423c91cc1bc85cbb054c37ba3e6ba094e0a725e19.jpg", + "image_caption": [ + "An aerial photo capturing Hawaii's islands" + ], + "image_footnote": [], + "bbox": [ + 475, + 88, + 665, + 234 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/42ee3e044d7608bd775b99d25786bded72e9226d5ebdab4e5e3d999608992a95.jpg", + "image_caption": [ + "A straight road alpine forests on the sides", + "no paaooZ" + ], + "image_footnote": [], + "bbox": [ + 665, + 88, + 854, + 234 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/488449011145f4e274b8f429dbd3d516e8df83b20f1f96cbb5d72e25506b007f.jpg", + "image_caption": [ + "Dark starry sky" + ], + "image_footnote": [], + "bbox": [ + 96, + 243, + 285, + 388 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5800ab47dd2dfc938c2370532c62ee36d81affbf53b938a521c8560645934ccb.jpg", + "image_caption": [ + "Satellite image of a landmass of the Earth's surface" + ], + "image_footnote": [], + "bbox": [ + 285, + 243, + 473, + 388 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6b63d39a8012abaff02989e5b2aa299057b63f963370a29007e69c7802948ad0.jpg", + "image_caption": [ + "An aerial photo of Hawaii's mountains and rain forest" + ], + "image_footnote": [], + "bbox": [ + 475, + 243, + 663, + 388 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c8a937fd92677936c9d3137743bc76b54da5a3fb1b4557f5309683387580ef2e.jpg", + "image_caption": [ + "Alpine forest road with Mount Rainier in the end" + ], + "image_footnote": [], + "bbox": [ + 665, + 243, + 854, + 388 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4533042ab7be98b9e46bf4f3dde873bf98af9685961026832f47faaaeae05fc6.jpg", + "image_caption": [ + "Far view of alien solar system" + ], + "image_footnote": [], + "bbox": [ + 96, + 397, + 285, + 542 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ab4d258ca1fb869134cce3d1ddb9b91ef456010f9620934ce95c8e941b2ea49e.jpg", + "image_caption": [ + "Satellite image of a quaint American countryside" + ], + "image_footnote": [], + "bbox": [ + 285, + 397, + 473, + 542 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5fdc2d268c78cd41e0cca1fdc2c8bb6be16bb50832756d6e2a9b59911c349991.jpg", + "image_caption": [ + "An aerial close-up of the volcano's caldera" + ], + "image_footnote": [], + "bbox": [ + 475, + 397, + 665, + 542 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d9221a0a2d01b0c14f47feff75bb53634ee83a43649fd323ddbf0d9edd99bc36.jpg", + "image_caption": [ + "Alpine meadows against the massive Mount Rainier" + ], + "image_footnote": [], + "bbox": [ + 665, + 397, + 854, + 542 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2acb4e28e8d0fad8066dec15f7fd5cca3dc92014b69c0e3fe899844472524d7c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 551, + 285, + 696 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/76dece77728ba7efd05a8d601324842fefc1ac0f8372034bb66ca746289ba7ea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 285, + 551, + 473, + 696 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/42dfe2c0f608eb2897eef3e6b0b939c5f3e80b5cad0c40dfb63291a9efdef785.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 475, + 551, + 665, + 696 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e2a53f25efa467a52f9dcbee6bc5259e2819ab01a6d960865221ed313d3cb72c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 665, + 551, + 854, + 696 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d2ce1aae24960dc53fe311203086bf2570378e92b2d16331600ec8eebd6fe156.jpg", + "image_caption": [ + "An exoplanet of a foreign solar system", + "Top-down aerial image of deserted continents", + "Figure 7. Selected stills from our generated zoom videos (columns). Please refer to the supplementary materials for complete text prompts." + ], + "image_footnote": [], + "bbox": [ + 96, + 705, + 285, + 849 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cf14ddaceddf48b468726f94fb1d97c439f1e2392d47abb94e6ef09a41dd142a.jpg", + "image_caption": [ + "Satellite image of a foggy forest", + "Top down view of a lake with a person kayaking" + ], + "image_footnote": [], + "bbox": [ + 285, + 705, + 473, + 849 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/87e25a44e21ba0a2c7d664eba517574443f38c785d23e20befc9413da7adb21b.jpg", + "image_caption": [ + "An aerial close-up of the rim of a volcano's caldera", + "A man standing on the edge of a volcano's caldera" + ], + "image_footnote": [], + "bbox": [ + 475, + 705, + 665, + 849 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ff7b55765cb2d66e776cf69e259b8dc51ef7f0892ee2f4802ef5d9ee95a4023a.jpg", + "image_caption": [ + "Steep cliffs and rocky outcrops of a snow mountain", + "A team of climbers climbing on the rugged cliffs", + "uipuaooZ" + ], + "image_footnote": [], + "bbox": [ + 665, + 705, + 854, + 849 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7178", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "clean images $\\{\\hat{\\mathbf{x}}_{i,t - 1}\\}_{i = 0}^{N - 1}$ to match with the content provided in $\\pmb{\\xi}$ in a zoom-consistent way. We show our generated photograph-based zoom sequences in Fig. 6.", + "bbox": [ + 76, + 89, + 468, + 137 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 143, + 294, + 159 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For the underlying text-to-image diffusion model, we use a version of Imagen [21] trained on internal data sources, which is a cascaded diffusion model consisting of (1) a base model conditioned on a text prompt embedding and (2) a super resolution model additionally conditioned the low resolution output from the base model. We use its default DDPM sampling procedure with 256 sampling steps, and we employ our multi-scale joint sampling to the base model only. We use the super resolution model to upsample each generated image independently.", + "bbox": [ + 75, + 167, + 470, + 318 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 330, + 209, + 347 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Figs. 6, 7, 8, 9, and 10, we demonstrate that our approach successfully generates consistent high quality zoom sequences for arbitrary relative zoom factors and a diverse set of scenes. Please see our supplementary materials for a full collection of videos. Sec. 5.1 describes how we generate text prompts, Sec. 5.2 demonstrates how our method outperforms diffusion-based outpainting and super-resolution models, and Sec. 5.3 justifies our design decisions with an ablation study.", + "bbox": [ + 75, + 354, + 468, + 491 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Text Prompt Generation", + "text_level": 1, + "bbox": [ + 76, + 500, + 302, + 515 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We generate a collection of text prompts that describe scenes at varying levels of scales using a combination of ChatGPT [14] and manual editing. We start with prompting ChatGPT with a description of a scene, and asking it to formulate the sequence of prompts we might need for different zoom levels. While the results from this query are often plausible, they often (1) do not accurately match the corresponding requested scales, or (2) do not match the distribution of text prompts that the text-to-image model is able to most effectively generate. As such, we manually refine the prompts. A comprehensive collection of the prompts used to generate results in the paper are provided in the supplemental materials, along with the initial versions automatically produced by ChatGPT. In the future, we expect LLMs (and in particular, multimodal models) to automatically produce a sequence of prompts well suited for this application. In total, we collect a total of 10 examples, with the prompts sequence length varying form 6 to 16.", + "bbox": [ + 75, + 522, + 470, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Baseline Comparisons", + "text_level": 1, + "bbox": [ + 76, + 801, + 284, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 8 compares zoom sequences generated with our method and without (i.e., independently sampling each scale). When compared to our results, the independently-generated images similarly follow the text prompt, but clearly do not correspond to a single consistent underlying scene.", + "bbox": [ + 75, + 824, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Zoomed out", + "bbox": [ + 514, + 89, + 573, + 99 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Zoomed in", + "bbox": [ + 823, + 89, + 879, + 98 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/23db433e18c53431031553f13b04936af6831f6286e0a5f89ca88b180af0c341.jpg", + "image_caption": [ + "Figure 8. Generated zoom sequences with independent sampling (top) and our multi-scale sampling (bottom). Our method encourages different levels to depict a consistent underlying scene, while not compromising the image quality." + ], + "image_footnote": [], + "bbox": [ + 504, + 104, + 887, + 253 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Next, we compare our method to two autogressive generation approaches for generating zoom sequences: (1) Stable Diffusion's [1] outpainting model and (2) Stable Diffusion's \"upscale\" super-resolution model. We show representative qualitative results in Fig. 9.", + "bbox": [ + 498, + 338, + 890, + 412 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison to progressive outpainting. The outpainting baseline starts with generating the most zoomed-in image and progressively generates coarser scales by downsampling the previous generated image and outpainting the surrounding area. As in our method, the inpainting of each level is conditioned on the corresponding text prompt. In Fig. 9, we show that because of the causality of the autoregressive process, the outpainting approach suffers from gradually accumulating errors, i.e., when a mistake is made at a given step, later outpainting iterations may struggle to produce a consistent image.", + "bbox": [ + 496, + 417, + 892, + 583 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison to progressive super-resolution. The superresolution baseline starts with the most zoomed-out image and generates subsequent scales by super-resolving the upscaled central image region, conditioned on the corresponding text prompt. The low resolution input provides strong structural information which constrains the layout of the next zoomed-in image. As we can see in Fig. 9, this super-resolution baseline is not able to synthesize new objects that would only appear in the finer, zoomed-in scales.", + "bbox": [ + 496, + 588, + 890, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablations", + "text_level": 1, + "bbox": [ + 500, + 736, + 609, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Fig. 10, we show comparisons to simpler versions of our method to examine the effect of our design decisions.", + "bbox": [ + 498, + 760, + 890, + 790 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Joint vs. Iterative update. Instead of performing multiscale blending approach, we can instead iteratively cycle through the images in the zoom stack, and perform one sampling step at each level independently. Unlike fully independent sampling, this process does allow for sharing of information between scales, since the steps are still applied to renders from the zoom stack. We find that although this pro", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7179", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a15c9ebdc8fd47af793464f0374eeea9f3f39735fa2304f7eb4732a5d2503ce9.jpg", + "image_caption": [ + "Figure 9. Comparisons with Stable Diffusion Outpainting and super-resolution (SR) models." + ], + "image_footnote": [], + "bbox": [ + 96, + 87, + 883, + 414 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7c9b052a7db879198c21b45b06c744b667218175e889b665f3af0ddd9ddddff0.jpg", + "image_caption": [ + "Figure 10. Ablations. We evaluate other options for multi-scale consistency: (1) iteratively updating each level separately, (2) naive multi-scale blending, (3) removing the shared noise." + ], + "image_footnote": [], + "bbox": [ + 86, + 462, + 465, + 746 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "duces more consistent results than independent sampling, there remain inconsistencies at stack layer boundaries.", + "bbox": [ + 75, + 819, + 467, + 849 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Shared vs. random noise Instead of using a shared noise $\\Pi_{\\mathrm{noise}}$ , noise can be sampled independently for each zoom level. We find that this leads to blur in the output samples.", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison with naïve blending. Instead of our multiscale blending, we can instead naively blend the observations together, e.g., as in MultiDiffusion [2]. We find that this leads to blurry outputs at deeper zoom levels.", + "bbox": [ + 498, + 465, + 890, + 526 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Discussion & Limitations", + "text_level": 1, + "bbox": [ + 500, + 547, + 736, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A significant challenge in our work is discovering the appropriate set of text prompts that (1) agree with each other across a set of fixed scales, and (2) can be effectively generated consistently by a given text-to-image model. One possible avenue of improvement could be to, along with sampling, optimize for suitable geometric transformations between successive zoom levels. These transformations could include translation, rotation, and even scale, to find better alignment between the zoom levels and the prompts.", + "bbox": [ + 496, + 575, + 890, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Alternatively, one can optimize the text embeddings, to find better descriptions that correspond to subsequent zoom levels. Or, instead, use the LLM for in-the-loop generation, i.e., by giving LLM the generated image content, and asking it to refine its prompts to produce images which are closer in correspondence given the set of pre-defined scales.", + "bbox": [ + 496, + 713, + 890, + 803 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. We thank Ben Poole, Jon Barron, Luyang Zhu, Ruiqi Gao, Tong He, Grace Luo, Angjoo Kanazawa, Vickie Ye, Songwei Ge, Keunhong Park, and David Salesin for helpful discussions and feedback. This work was supported in part by UW Reality Lab, Meta, Google, OPPO, and Amazon.", + "bbox": [ + 496, + 810, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7180", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Stability AI. Stable-diffusion-2-inpainting. https://huggingface.co/stabilityai/stable-diffusion-2-inpainting.2,7", + "[2] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. arXiv preprint arXiv:2302.08113, 2023. 2, 8", + "[3] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 1", + "[4] Peter J Burt and Edward H Adelson. The laplacian pyramid as a compact image code. In Readings in computer vision, pages 671-679. Elsevier, 1987. 4", + "[5] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2", + "[6] Dave Epstein, Allan Jabri, Ben Poole, Alexei A Efros, and Aleksander Holynski. Diffusion self-guidance for controllable image generation. arXiv preprint arXiv:2306.00986, 2023. 1", + "[7] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 1", + "[8] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 3", + "[9] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5", + "[10] Yuseung Lee, Kunho Kim, Hyunjin Kim, and Minhyuk Sung. Syncdiffusion: Coherent montage via synchronized joint diffusions. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2", + "[11] Zhengqi Li, Qianqian Wang, Noah Snavely, and Angjoo Kanazawa. Infinitenature-zero: Learning perpetual view generation of natural scenes from single images. In European Conference on Computer Vision, pages 515-534. Springer, 2022. 2", + "[12] Andrew Liu, Richard Tucker, Varun Jampani, Ameesh Makadia, Noah Snavely, and Angjoo Kanazawa. Infinite nature: Perpetual view generation of natural scenes from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14458-14467, 2021. 2", + "[13] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. arXiv preprint arXiv:2202.09778, 2022. 3", + "[14] OpenAI. Chatgpt [large language model]. https://chat.openai.com/chat.7", + "[15] Ryan Po, Wang Yifan, Vladislav Golyanik, Kfir Aberman, Jonathan T Barron, Amit H Bermano, Eric Ryan Chan, Tali Dekel, Aleksander Holynski, Angjoo Kanazawa, et al. State of the art on diffusion models for visual computing. arXiv preprint arXiv:2310.07204, 2023. 1" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2, 3", + "[17] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3", + "[18] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 1", + "[19] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models. arXiv preprint arXiv:2307.06949, 2023. 1", + "[20] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 2", + "[21] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 3, 7", + "[22] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image super-resolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4713-4726, 2022. 2", + "[23] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2", + "[24] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.3", + "[25] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019.", + "[26] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2", + "[27] Luming Tang, Nataniel Ruiz, Chu Qinghao, Yuanzhen Li, Aleksander Holynski, David E Jacobs, Bharath Hariharan, Yael Pritch, Neal Wadhwa, Kfir Aberman, and Michael Rubinstein. Realfill: Reference-driven generation for authentic image completion. arXiv preprint arXiv:2309.16668, 2023. 2" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7181", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Shitao Tang, Fuyang Zhang, Jiacheng Chen, Peng Wang, and Yasutaka Furukawa. Mvdiffusion: Enabling holistic multiview image generation with correspondence-aware diffusion. arXiv preprint arXiv:2307.01097, 2023. 2", + "[29] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1", + "[30] Qinsheng Zhang, Jiaming Song, Xun Huang, Yongxin Chen, and Ming-Yu Liu. Diffcollage: Parallel generation of large content with diffusion models. arXiv preprint arXiv:2303.17076, 2023. 2" + ], + "bbox": [ + 78, + 90, + 468, + 260 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7182", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Generative Powers of Ten/2ccba12d-1f5a-40cf-8686-79a45cb775c2_model.json b/2024/Generative Powers of Ten/2ccba12d-1f5a-40cf-8686-79a45cb775c2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a2c337b57298a198fee4ec820c3679a4c5517fc1 --- /dev/null +++ b/2024/Generative Powers of Ten/2ccba12d-1f5a-40cf-8686-79a45cb775c2_model.json @@ -0,0 +1,2354 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.355, + 0.131, + 0.616, + 0.151 + ], + "angle": 0, + "content": "Generative Powers of Ten" + }, + { + "type": "text", + "bbox": [ + 0.086, + 0.179, + 0.93, + 0.217 + ], + "angle": 0, + "content": "Xiaojuan Wang\\(^{1}\\) Janne Kontkanen\\(^{2}\\) Brian Curless\\(^{1,2}\\) Steven M. Seitz\\(^{1,2}\\) Ira Kemelmacher-Shlizerman\\(^{1,2}\\) Ben Mildenhall\\(^{2}\\) Pratul Srinivasan\\(^{2}\\) Dor Verbin\\(^{2}\\) Aleksander Holynski\\(^{2,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.219, + 0.78, + 0.26 + ], + "angle": 0, + "content": "1University of Washington 2Google Research 3UC Berkeley powers-of-ten.github.io" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.296, + 0.278, + 0.451 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.296, + 0.482, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.489, + 0.297, + 0.686, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.297, + 0.888, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.463, + 0.887, + 0.504 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.517, + 0.893, + 0.547 + ], + "angle": 0, + "content": "Figure 1. Given a series of prompts describing a scene at varying zoom levels, e.g., from a distant galaxy to the surface of an alien planet, our method uses a pre-trained text-to-image diffusion model to generate a continuously zooming video sequence." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.559, + 0.314, + 0.575 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.59, + 0.473, + 0.832 + ], + "angle": 0, + "content": "We present a method that uses a text-to-image model to generate consistent content across multiple image scales, enabling extreme semantic zooms into a scene, e.g. ranging from a wide-angle landscape view of a forest to a macro shot of an insect sitting on one of the tree branches. We achieve this through a joint multi-scale diffusion sampling approach that encourages consistency across different scales while preserving the integrity of each individual sampling process. Since each generated scale is guided by a different text prompt, our method enables deeper levels of zoom than traditional super-resolution methods that may struggle to create new contextual structure at vastly different scales. We compare our method qualitatively with alternative techniques in image super-resolution and outpainting, and show that our method is most effective at generating consistent multi-scale content." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.845, + 0.21, + 0.861 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Recent advances in text-to-image models [3, 6, 7, 15, 18, 19, 29] have been transformative in enabling applications" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.56, + 0.893, + 0.77 + ], + "angle": 0, + "content": "like image generation from a single text prompt. But while digital images exist at a fixed resolution, the real world can be experienced at many different levels of scale. Few things exemplify this better than the classic 1977 short film \"Powers of Ten\", shown in Figure 2, which showcases the sheer magnitudes of scale that exist in the universe by visualizing a continuous zoom from the outermost depths of the galaxy to the cells inside our bodies1. Unfortunately, producing animations or interactive experiences like these has traditionally required trained artists and many hours of tedious labor—and although we might want to replace this process with a generative model, existing methods have not yet demonstrated the ability to generate consistent content across multiple zoom levels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.771, + 0.895, + 0.879 + ], + "angle": 0, + "content": "Unlike traditional super-resolution methods, which generate higher-resolution content conditioned on the pixels of the original image, extreme zooms expose entirely new structures, e.g., magnifying a hand to reveal its underlying skin cells. Generating such a zoom requires semantic knowledge of human anatomy. In this paper, we focus on solving this semantic zoom problem, i.e., enabling text-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.888, + 0.866, + 0.901 + ], + "angle": 0, + "content": "1https://www.youtube.com/watch?v=0fKBhvdjuy0" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7173" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.155, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.157, + 0.09, + 0.233, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.236, + 0.09, + 0.312, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.09, + 0.391, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.395, + 0.09, + 0.468, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.221, + 0.47, + 0.276 + ], + "angle": 0, + "content": "Figure 2. Powers of Ten (1977) This documentary film illustrates the relative scale of the universe as a single shot that gradually zooms out from a human to the universe, and then back again to the microscopic molecular level." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.281, + 0.471, + 0.446 + ], + "angle": 0, + "content": "conditioned multi-scale image generation, to create Powers of Ten-like zoom videos. As input, our method expects a series of text prompts that describe different scales of the scene, and produces as output a multi-scale image representation that can be explored interactively or rendered to a seamless zooming video. These text prompts can be user-defined (allowing for creative control over the content at different zoom levels) or crafted with the help of a large language model (e.g., by querying the model with an image caption and a prompt like \"describe what might you see if you zoomed in by \\(2x\\))" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.449, + 0.47, + 0.844 + ], + "angle": 0, + "content": "At its core, our method relies on a joint sampling algorithm that uses a set of parallel diffusion sampling processes distributed across zoom levels. These sampling processes are coordinated to be consistent through an iterative frequency-band consolidation process, in which intermediate image predictions are consistently combined across scales. Unlike existing approaches that accomplish similar goals by repeatedly increasing the effective image resolution (e.g., through super-resolution or image outpainting), our sampling process jointly optimizes for the content of all scales at once, allowing for both (1) plausible images at each scale and (2) consistent content across scales. Furthermore, existing methods are limited in their ability to explore wide ranges of scale, since they rely primarily on the input image content to determine the added details at subsequent zoom levels. In many cases, image patches contain insufficient contextual information to inform detail at deeper (e.g., \\(10\\mathrm{x}\\) or \\(100\\mathrm{x}\\)) zoom levels. On the other hand, our method grounds each scale in a text prompt, allowing for new structures and content to be conceived across extreme zoom levels. In our experiments, we compare our work qualitatively to these existing methods, and demonstrate that the zoom videos that our method produces are notably more consistent. Finally, we showcase a number of ways in which our algorithm can be used, e.g., by conditioning purely on text or grounding the generation in a known (real) image." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.859, + 0.2, + 0.875 + ], + "angle": 0, + "content": "2. Prior Work" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.468, + 0.902 + ], + "angle": 0, + "content": "Super-resolution and inpainting. Existing text-to-image" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.243 + ], + "angle": 0, + "content": "based super resolution models [1, 22] and outpainting models [1, 16, 20, 27] can be adapted to the zoom task as autoregressive processes, i.e., by progressively outpainting a zoomed-in image, or progressively super-resolving a zoomed-out image. One significant drawback of these approaches is that later-generated images have no influence on the previously generated ones, which can often lead to suboptimal results, as certain structures may be entirely incompatible with subsequent levels of detail, causing error accumulation across recurrent network applications." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.249, + 0.892, + 0.43 + ], + "angle": 0, + "content": "Perpetual view generation. Starting from a single view RGB image, perpetual view generation methods like Infinite Nature [11] and InfiniteNature-Zero [12] learn to generate unbounded flythrough videos of natural scenes. These methods differ from our generative zoom in two key ways: (1) they translate the camera in 3D, causing a \"fly-through\" effect with perspective effects, rather than the \"zoom in\" our method produces, and (2) they synthesize the fly-through starting from a single image by progressively inpainting unknown parts of novel views, whereas we generate the entire zoom sequence simultaneously and coherently across scales, with text-guided semantic control." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.437, + 0.892, + 0.738 + ], + "angle": 0, + "content": "Diffusion joint sampling for consistent generation. Recent research [2, 10, 28, 30] leverages pretrained diffusion models to generate arbitrary-sized images or panoramas from smaller pieces using joint diffusion processes. These processes involve concurrently generating these multiple images by merging their intermediate results within the sampling process. In particular, DiffCollage [30] introduces a factor graph formulation to express spatial constraints among these images, representing each image as a node, and overlapping areas with additional nodes. Each sampling step involves aggregating individual predictions based on the factor graph. For this to be possible, a given diffusion model needs to be finetuned for different factor nodes. Other works such as MultiDiffusion [2] reconcile different denoising steps by solving for a least squares optimal solution: i.e., averaging the diffusion model predictions at overlapping areas. However, none of these approaches can be applied to our problem, where our jointly sampled images have spatial correspondence at vastly different spatial scales." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.763, + 0.638, + 0.779 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.791, + 0.892, + 0.868 + ], + "angle": 0, + "content": "Diffusion models [5, 8, 23-26] generate images from random noise through a sequential sampling process. This sampling process reverses a destructive process that gradually adds Gaussian noise on a clean image \\(\\mathbf{x}\\). The intermediate noisy image at time step \\(t\\) is expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.636, + 0.888, + 0.756, + 0.901 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {t} = \\alpha_ {t} \\mathbf {x} + \\sigma_ {t} \\boldsymbol {\\epsilon} _ {t},\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7174" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.182 + ], + "angle": 0, + "content": "where \\(\\epsilon_{t} \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\) is a standard Gaussian noise, and \\(\\alpha_{t}\\) and \\(\\sigma_{t}\\) define a fixed noise schedule, with larger \\(t\\) corresponding to more noise. A diffusion model is a neural network \\(\\epsilon_{\\theta}\\) that predicts the approximate clean image \\(\\hat{\\mathbf{x}}\\) directly, or equivalently the added noise \\(\\epsilon_{t}\\) in \\(\\mathbf{z}_{t}\\). The network is trained with the loss:" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.197, + 0.456, + 0.217 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\boldsymbol {\\theta}) = \\mathbb {E} _ {t \\sim U [ 1, T ], \\boldsymbol {\\epsilon} _ {t} \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I})} [ w (t) \\| \\boldsymbol {\\epsilon} _ {\\boldsymbol {\\theta}} (\\mathbf {z} _ {t}; t, y) - \\boldsymbol {\\epsilon} _ {t} \\| _ {2} ^ {2} ],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.23, + 0.47, + 0.305 + ], + "angle": 0, + "content": "where \\(y\\) is an additional conditioning signal like text [16, 17, 21], and \\(w(t)\\) is a weighting function typically set to 1 [8]. A standard choice for \\(\\epsilon_{\\theta}\\) is a U-Net with self-attention and cross-attention operations attending to the conditioning \\(y\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.307, + 0.471, + 0.429 + ], + "angle": 0, + "content": "Once the diffusion model is trained, various sampling methods [8, 13, 24] are designed to sample efficiently from the model, starting from pure noise \\(\\mathbf{z}_T \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})\\) and iteratively denoising it to a clean image. These sampling methods often rely on classifier-free guidance [8], a process which uses a linear combination of the text-conditional and unconditional predictions to achieve better adherence to the conditioning signal:" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.444, + 0.4, + 0.46 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {\\epsilon}} _ {t} = (1 + \\omega) \\boldsymbol {\\epsilon} _ {\\theta} (\\mathbf {z} _ {t}; t, y) - \\omega \\boldsymbol {\\epsilon} _ {\\theta} (\\mathbf {z} _ {t}; t).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.475, + 0.471, + 0.567 + ], + "angle": 0, + "content": "This revised \\(\\hat{\\epsilon}_t\\) is used as the noise prediction to update the noisy image \\(\\mathbf{z}_t\\). Given a noisy image and a noise prediction, the estimated clean image \\(\\hat{\\mathbf{x}}_t\\) is computed as \\(\\hat{\\mathbf{x}}_t = (\\mathbf{z}_t - \\sigma_t\\hat{\\epsilon}_t) / \\alpha_t\\). The iterative update function in the sampling process depends on the sampler used; in this paper we use DDPM [8]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.585, + 0.169, + 0.601 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.612, + 0.47, + 0.764 + ], + "angle": 0, + "content": "Let \\( y_0, \\ldots, y_{N-1} \\) be a series of prompts describing a single scene at varying, corresponding zoom levels \\( p_0, \\ldots, p_{N-1} \\) forming a geometric progression, i.e., \\( p_i = p^i \\) (we typically set \\( p \\) to 2 or 4). Our objective is to generate a sequence of corresponding \\( H \\times W \\times C \\) images \\( \\mathbf{x}_0, \\ldots, \\mathbf{x}_{N-1} \\) from an existing, pre-trained, text-to-image diffusion model. We aim to generate the entire set of images jointly in a zoom-consistent way. This means that the image \\( \\mathbf{x}_i \\) at any specific zoom level \\( p_i \\), should be consistent with the center \\( H / p \\times W / p \\) crop of the zoomed-out image \\( \\mathbf{x}_{i-1} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We propose a multi-scale joint sampling approach and a corresponding zoom stack representation that gets updated in the diffusion-based sampling process. In Sec. 4.1, we introduce our zoom stack representation and the process that allows us to render it into an image at any given zoom level. In Sec. 4.2, we present an approach for consolidating multiple diffusion estimates into this representation in a consistent way. Finally, in Sec. 4.3, we show how these components are used in the complete sampling process." + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.088, + 0.895, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.234, + 0.895, + 0.306 + ], + "angle": 0, + "content": "Figure 3. Zoom stack. Our representation consists of \\( N \\) layer images \\( L_{i} \\) of constant resolution (left). These layers are arranged in a pyramid-like structure, with layers representing finer details corresponding to a smaller spatial extent (middle). These layers are composited to form an image at any zoom level (right)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.322, + 0.75, + 0.339 + ], + "angle": 0, + "content": "4.1. Zoom Stack Representation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.347, + 0.894, + 0.438 + ], + "angle": 0, + "content": "Our zoom stack representation, which we denote by \\(\\mathcal{L} = (L_0,\\dots,L_{N - 1})\\), is designed to allow rendering images at any zoom level \\(p_0,\\ldots ,p_{N - 1}\\). The representation, illustrated in Fig. 3, contains \\(N\\) images of shape \\(H\\times W\\), one for each zoom level, where the \\(i\\)th image \\(L_{i}\\) stores the pixels corresponding to the \\(i\\)th zoom level \\(p_i\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.442, + 0.893, + 0.7 + ], + "angle": 0, + "content": "Image rendering. The rendering operator, which we denote by \\(\\Pi_{\\mathrm{image}}(\\mathcal{L};i)\\), takes a zoom stack \\(\\mathcal{L}\\) and returns the image at the \\(i\\)th zoom level \\(p_i = p^i\\). We denote by \\(\\mathcal{D}_i(\\mathbf{x})\\) the operator which downscales the image \\(\\mathbf{x}\\) by factor \\(p_i\\), and zero-pads the image back to size \\(H\\times W\\); and we denote by \\(M_{i}\\) the corresponding \\(H\\times W\\) binary image which has value 1 at the center \\(H / p_{i}\\times W / p_{i}\\) patch and value 0 at padded pixels. The operator \\(\\mathcal{D}_i\\) operates by prefiltering the image with a truncated Gaussian kernel of size \\(p_i\\times p_i\\) and resampling with a stride of \\(p_i\\). As described in Alg. 1, an image \\(\\mathbf{x}_i\\) at the \\(i\\)th zoom level is rendered by starting with \\(L_{i}\\), and iteratively replacing its central \\(H / p_j\\times W / p_j\\) crop with \\(\\mathcal{D}_{j - i}(L_j)\\), for \\(j = i + 1,\\dots,N - 1\\). (In Alg. 1 we denote by \\(\\odot\\) the elementwise multiplication of a binary mask \\(M\\) with an image.) This process guarantees that rendering at different zoom levels will be consistent at overlapping central regions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Noise rendering. At every denoising iteration of DDPM [8], each pixel is corrupted by globally-scaled i.i.d. Gaussian noise \\(\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\). Since we would like images rendered at different zoom levels to be consistent, it is essential to make sure the added noise is also consistent, with overlapping region across different zoom levels sharing the same noise structure. Therefore, we use a rendering operator similar to \\(\\Pi_{\\mathrm{image}}\\) which converts a set of independent noise images, \\(\\mathcal{E} = (E_0,\\dots,E_{N - 1})\\) into a single zoom-consistent noise \\(\\epsilon_{i} = \\Pi_{\\mathrm{noise}}(\\mathcal{E};i)\\). However, because downsampling involves prefiltering, which modifies the statistics of the resulting noise, we upscale the \\(j\\)th down-scaled noise component by \\(p_j / p_i\\) to preserve the variance," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7175" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.088, + 0.895, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.274, + 0.895, + 0.346 + ], + "angle": 0, + "content": "Figure 4. Overview of a single sampling step. (1) Noisy images \\(\\mathbf{z}_{i,t}\\) from each zoom level, along with the respective prompts \\(y_{i}\\) are simultaneously fed into the same pretrained diffusion model, returning estimates of the corresponding clean images \\(\\hat{\\mathbf{x}}_{i,t}\\). These images may have inconsistent estimates for the overlapping regions that they all observe. We employ multi-resolution blending to fuse these regions into a consistent zoom stack \\(\\mathcal{L}_t\\) and re-render the different zoom levels from the consistent representation. These re-rendered images \\(\\Pi_{\\mathrm{image}}(\\mathcal{L}_t;i)\\) are then used as the clean image estimates in the DDPM sampling step." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.358, + 0.471, + 0.403 + ], + "angle": 0, + "content": "ensuring that the noise satisfies the standard Gaussian distribution assumption, i.e., that \\(\\epsilon_{i} = \\Pi_{\\mathrm{noise}}(\\mathcal{E};i)\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\) for all levels \\(i\\)." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.414, + 0.411, + 0.43 + ], + "angle": 0, + "content": "Algorithm 1 Image and noise rendering at scale \\(i\\)." + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.439, + 0.276, + 0.455 + ], + "angle": 0, + "content": "1: Set \\(\\mathbf{x} \\gets L_i, \\epsilon \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})\\)" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.456, + 0.299, + 0.47 + ], + "angle": 0, + "content": "2: for \\( j = i + 1, \\ldots, N - 1 \\) do" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.47, + 0.42, + 0.486 + ], + "angle": 0, + "content": "3: \\(\\mathbf{x} \\gets M_{j - i} \\odot \\mathcal{D}_{j - i}(L_j) + (1 - M_{j - i}) \\odot \\mathbf{x}\\)" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.486, + 0.467, + 0.502 + ], + "angle": 0, + "content": "4: \\(\\epsilon \\gets (p_j / p_i)M_{j - i}\\odot \\mathcal{D}_{j - i}(E_j) + (1 - M_{j - i})\\odot \\epsilon\\)" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.502, + 0.161, + 0.515 + ], + "angle": 0, + "content": "5: end for" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.517, + 0.185, + 0.531 + ], + "angle": 0, + "content": "6: return x, \\(\\epsilon\\)" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.439, + 0.467, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.561, + 0.314, + 0.578 + ], + "angle": 0, + "content": "4.2. Multi-resolution blending" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.47, + 0.854 + ], + "angle": 0, + "content": "Equipped with a method for rendering a zoom stack and sampling noise at any given zoom level, we now describe a mechanism for integrating multiple observations of the same scene \\(\\mathbf{x}_0,\\dots ,\\mathbf{x}_{N - 1}\\) at varying zoom levels \\(p_0,\\dots ,p_{N - 1}\\) into a consistent zoom stack \\(\\mathcal{L}\\). This process is a necessary component of the consistent sampling process, as the diffusion model applied at various zoom levels will produce inconsistent content in the overlapping regions. Specifically, the \\(j\\)th zoom stack level \\(L_{j}\\) is used in rendering multiple images at all zoom levels \\(i\\leq j\\) and therefore its value should be consistent with multiple image observations (or diffusion model samples), namely \\(\\{\\mathbf{x}_i:i\\leq j\\}\\). The simplest possible solution to this is to naively average the overlapping regions across all observations. This approach, however, results in blurry zoom stack images, since coarser-scale observations of overlapping regions contain fewer pixels, and therefore only lower-frequency information." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.902 + ], + "angle": 0, + "content": "To solve this, we propose an approach we call multi-resolution blending, which uses Laplacian pyramids to selectively fuse the appropriate frequency bands of each ob" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.356, + 0.891, + 0.662 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.673, + 0.895, + 0.786 + ], + "angle": 0, + "content": "Figure 5. Multi-resolution blending. We produce a consistent estimate for Layer \\( L_{i} \\) in the zoom stack by merging the \\( H / p_{j} \\times W / p_{j} \\) central region of the corresponding zoomed out images \\( \\mathbf{x}_j \\) for \\( j \\leq i \\). This merging process involves (1) creating a Laplacian pyramid from each observation, and blending together the corresponding frequency bands to create a blended pyramid. This blended pyramid is recomposed into an image, which is used to update the layer \\( L_{i} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.903 + ], + "angle": 0, + "content": "servation level, which prevents aliasing as well as overblurring. We show an outline of this process in Fig. 5. More concretely, to update the \\(i\\)th layer in the zoom stack, we begin by cropping all samples \\(j \\geq i\\) to match with the content of the \\(i\\)th level, and rescaling them back to \\(H \\times W\\). We then analyze each of these \\(N - i - 1\\) images into a Laplacian pyramid [4], and average across corresponding frequency bands" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "7176" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.168 + ], + "angle": 0, + "content": "(see Figure 5), resulting in an average Laplacian pyramid, which can be recomposed into an image and assigned to the \\(i\\)th level of the zoom stack. This process is applied for each layer of the zoom stack \\(L_{i}\\), collecting from all further zoomed-out levels \\(j \\geq i\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.175, + 0.357, + 0.192 + ], + "angle": 0, + "content": "4.3. Multi-scale consistent sampling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.198, + 0.474, + 0.366 + ], + "angle": 0, + "content": "Our complete multi-scale joint sampling process is shown in Alg. 2. Fig. 4 illustrates a single sampling step \\( t \\): Noisy images \\( \\mathbf{z}_{i,t} \\) in each zoom level along with the respective prompt \\( y_{i} \\) are fed into the pretrained diffusion model in parallel to predict the noise \\( \\hat{\\epsilon}_{i,t-1} \\), and thus to compute the estimated clean images \\( \\hat{\\mathbf{x}}_{i,t} \\). Equipped with our multi-resolution blending technique, the clean images are consolidated into a zoom stack, which is then rendered at all zoom levels, yielding consistent images \\( \\Pi_{\\mathrm{image}}(\\mathcal{L}_t; i) \\). These images are then used in a DDPM update step along with the input \\( \\mathbf{z}_t \\) to compute the next \\( \\mathbf{z}_{t-1} \\)." + }, + { + "type": "code_caption", + "bbox": [ + 0.08, + 0.378, + 0.347, + 0.394 + ], + "angle": 0, + "content": "Algorithm 2 Multi-scale joint sampling." + }, + { + "type": "algorithm", + "bbox": [ + 0.085, + 0.402, + 0.428, + 0.617 + ], + "angle": 0, + "content": "1: Set \\(\\mathcal{L}_T\\gets 0,\\mathbf{z}_{i,T}\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I}),\\forall i = 0,\\dots ,N - 1\\) \n2: for \\(t = T,\\ldots ,1\\) do \n3: \\(\\mathcal{E}\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\) \n4: parfor \\(i = 0,\\dots ,N - 1\\) do \n5: \\(\\mathbf{x}_{i,t} = \\Pi_{\\mathrm{image}}(\\mathcal{L}_t;i)\\) \n6: \\(\\epsilon_{i} = \\Pi_{\\mathrm{noise}}(\\mathcal{E};i)\\) \n7: \\(\\mathbf{z}_{i,t - 1} = \\mathrm{DDPM\\_update}(\\mathbf{z}_{i,t},\\mathbf{x}_{i,t},\\boldsymbol {\\epsilon}_i)\\) \n8: \\(\\hat{\\epsilon}_{i,t - 1} = (1 + \\omega)\\epsilon_{\\theta}(\\mathbf{z}_{i,t - 1};t - 1,y_i)\\) \n9: \\(-\\omega \\epsilon_{\\theta}(\\mathbf{z}_{i,t - 1};t - 1)\\) \n10: \\(\\hat{\\mathbf{x}}_{i,t - 1} = (\\mathbf{z}_{i,t - 1} - \\sigma_{t - 1}\\hat{\\epsilon}_{i,t - 1}) / \\alpha_{t - 1}\\) \n11: end parfor \n12: \\(\\mathcal{L}_{t - 1}\\gets \\mathrm{Blending}(\\{\\hat{\\mathbf{x}}_{i,t - 1}\\}_{i = 0}^{N - 1})\\) \n13: end for \n14: return \\(\\mathcal{L}_0\\)" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.641, + 0.305, + 0.658 + ], + "angle": 0, + "content": "4.4. Photograph-based Zoom" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.664, + 0.47, + 0.756 + ], + "angle": 0, + "content": "In addition to using text prompts to generate the entire zoom stack from scratch, our approach can also generate a sequence zooming into an existing photograph. Given the most zoomed-out input image \\(\\xi\\), we still use Alg. 2, but we additionally update the denoised images to minimize the following loss function before every blending operation:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.762, + 0.47, + 0.805 + ], + "angle": 0, + "content": "\\[\n\\ell \\left(\\hat {\\mathbf {x}} _ {0, t}, \\dots , \\hat {\\mathbf {x}} _ {N - 1, t}\\right) = \\sum_ {i = 0} ^ {N - 1} \\left\\| \\mathcal {D} _ {i} \\left(\\hat {\\mathbf {x}} _ {i, t}\\right) - M _ {i} \\odot \\boldsymbol {\\xi}\\right) \\| _ {2} ^ {2}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.903 + ], + "angle": 0, + "content": "where, as we defined in Sec. 4.1, \\(\\mathcal{D}_i(\\mathbf{x})\\) downscales the image \\(\\mathbf{x}\\) by a factor \\(p_i\\) and pads the result back to \\(H\\times W\\), and \\(M_{i}\\) is a binary mask with 1 at the center \\(H / p_{i}\\times W / p_{i}\\) square and 0 otherwise. Before every blending operation we apply 5 Adam [9] steps at a learning rate of 0.1. This simple optimization-based strategy encourages the estimated" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.089, + 0.687, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.23, + 0.683, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.378, + 0.68, + 0.385 + ], + "angle": 0, + "content": "A closeup of the surface of skin of the back hand" + }, + { + "type": "image", + "bbox": [ + 0.688, + 0.089, + 0.871, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.698, + 0.23, + 0.858, + 0.237 + ], + "angle": 0, + "content": "A girl is holding a maple leaf in front of her face." + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.385, + 0.687, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.526, + 0.68, + 0.533 + ], + "angle": 0, + "content": "Epidermal layer of multiple rows of tiny skin cells" + }, + { + "type": "image", + "bbox": [ + 0.688, + 0.237, + 0.869, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.715, + 0.378, + 0.842, + 0.385 + ], + "angle": 0, + "content": "A brightly colored autumn maple leaf" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.533, + 0.687, + 0.673 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.673, + 0.664, + 0.68 + ], + "angle": 0, + "content": "A single round skin cell with its nucleus" + }, + { + "type": "image", + "bbox": [ + 0.688, + 0.385, + 0.869, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.706, + 0.526, + 0.85, + 0.533 + ], + "angle": 0, + "content": "Orange maple leaf texture with lots of veins" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.681, + 0.687, + 0.82 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.549, + 0.821, + 0.643, + 0.829 + ], + "angle": 0, + "content": "A nucleus within a skin cell" + }, + { + "type": "image", + "bbox": [ + 0.688, + 0.533, + 0.869, + 0.673 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.696, + 0.673, + 0.861, + 0.681 + ], + "angle": 0, + "content": "Macrophoto of the veins pattern on the maple leaf" + }, + { + "type": "image_caption", + "bbox": [ + 0.506, + 0.843, + 0.893, + 0.913 + ], + "angle": 0, + "content": "Figure 6. Selected images of our generated zoom sequences beginning with a provided real image. Left: Zoom from a man on a picnic blanket into the skin cells on his hand. Right: Zoom from a girl holding a leaf into the intricate vein patterns on the leaf. Face is blurred for anonymity." + }, + { + "type": "aside_text", + "bbox": [ + 0.878, + 0.09, + 0.892, + 0.136 + ], + "angle": 90, + "content": "Zoomed out" + }, + { + "type": "aside_text", + "bbox": [ + 0.878, + 0.789, + 0.892, + 0.831 + ], + "angle": 90, + "content": "Zoomed in" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7177" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.288, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.169, + 0.236, + 0.215, + 0.243 + ], + "angle": 0, + "content": "Galactic core" + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.089, + 0.475, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.321, + 0.236, + 0.442, + 0.243 + ], + "angle": 0, + "content": "Satellite image of the Earth's surface" + }, + { + "type": "image", + "bbox": [ + 0.476, + 0.089, + 0.666, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.236, + 0.641, + 0.244 + ], + "angle": 0, + "content": "An aerial photo capturing Hawaii's islands" + }, + { + "type": "image", + "bbox": [ + 0.666, + 0.089, + 0.855, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.691, + 0.236, + 0.828, + 0.243 + ], + "angle": 0, + "content": "A straight road alpine forests on the sides" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.244, + 0.286, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.166, + 0.39, + 0.218, + 0.396 + ], + "angle": 0, + "content": "Dark starry sky" + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.244, + 0.475, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.297, + 0.39, + 0.466, + 0.397 + ], + "angle": 0, + "content": "Satellite image of a landmass of the Earth's surface" + }, + { + "type": "image", + "bbox": [ + 0.476, + 0.244, + 0.665, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.483, + 0.39, + 0.659, + 0.397 + ], + "angle": 0, + "content": "An aerial photo of Hawaii's mountains and rain forest" + }, + { + "type": "image", + "bbox": [ + 0.666, + 0.244, + 0.855, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.679, + 0.39, + 0.841, + 0.397 + ], + "angle": 0, + "content": "Alpine forest road with Mount Rainier in the end" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.398, + 0.286, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.142, + 0.544, + 0.242, + 0.551 + ], + "angle": 0, + "content": "Far view of alien solar system" + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.398, + 0.475, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.301, + 0.544, + 0.462, + 0.551 + ], + "angle": 0, + "content": "Satellite image of a quaint American countryside" + }, + { + "type": "image", + "bbox": [ + 0.476, + 0.398, + 0.666, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.501, + 0.544, + 0.641, + 0.551 + ], + "angle": 0, + "content": "An aerial close-up of the volcano's caldera" + }, + { + "type": "image", + "bbox": [ + 0.666, + 0.398, + 0.855, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.675, + 0.544, + 0.846, + 0.55 + ], + "angle": 0, + "content": "Alpine meadows against the massive Mount Rainier" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.552, + 0.286, + 0.698 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.13, + 0.699, + 0.256, + 0.706 + ], + "angle": 0, + "content": "An exoplanet of a foreign solar system" + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.552, + 0.475, + 0.698 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.328, + 0.699, + 0.434, + 0.706 + ], + "angle": 0, + "content": "Satellite image of a foggy forest" + }, + { + "type": "image", + "bbox": [ + 0.476, + 0.552, + 0.666, + 0.698 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.486, + 0.699, + 0.655, + 0.706 + ], + "angle": 0, + "content": "An aerial close-up of the rim of a volcano's caldera" + }, + { + "type": "image", + "bbox": [ + 0.666, + 0.552, + 0.855, + 0.698 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.675, + 0.699, + 0.844, + 0.706 + ], + "angle": 0, + "content": "Steep cliffs and rocky outcrops of a snow mountain" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.707, + 0.286, + 0.851 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.852, + 0.267, + 0.86 + ], + "angle": 0, + "content": "Top-down aerial image of deserted continents" + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.707, + 0.475, + 0.851 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.301, + 0.852, + 0.461, + 0.86 + ], + "angle": 0, + "content": "Top down view of a lake with a person kayaking" + }, + { + "type": "image", + "bbox": [ + 0.476, + 0.707, + 0.666, + 0.851 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.487, + 0.852, + 0.654, + 0.86 + ], + "angle": 0, + "content": "A man standing on the edge of a volcano's caldera" + }, + { + "type": "image", + "bbox": [ + 0.666, + 0.707, + 0.855, + 0.851 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.679, + 0.852, + 0.841, + 0.86 + ], + "angle": 0, + "content": "A team of climbers climbing on the rugged cliffs" + }, + { + "type": "image_caption", + "bbox": [ + 0.865, + 0.092, + 0.878, + 0.137 + ], + "angle": 270, + "content": "no paaooZ" + }, + { + "type": "image_caption", + "bbox": [ + 0.865, + 0.819, + 0.877, + 0.859 + ], + "angle": 270, + "content": "uipuaooZ" + }, + { + "type": "image_caption", + "bbox": [ + 0.079, + 0.874, + 0.892, + 0.889 + ], + "angle": 0, + "content": "Figure 7. Selected stills from our generated zoom videos (columns). Please refer to the supplementary materials for complete text prompts." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7178" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.09, + 0.47, + 0.138 + ], + "angle": 0, + "content": "clean images \\(\\{\\hat{\\mathbf{x}}_{i,t - 1}\\}_{i = 0}^{N - 1}\\) to match with the content provided in \\(\\pmb{\\xi}\\) in a zoom-consistent way. We show our generated photograph-based zoom sequences in Fig. 6." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.145, + 0.295, + 0.16 + ], + "angle": 0, + "content": "4.5. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.471, + 0.319 + ], + "angle": 0, + "content": "For the underlying text-to-image diffusion model, we use a version of Imagen [21] trained on internal data sources, which is a cascaded diffusion model consisting of (1) a base model conditioned on a text prompt embedding and (2) a super resolution model additionally conditioned the low resolution output from the base model. We use its default DDPM sampling procedure with 256 sampling steps, and we employ our multi-scale joint sampling to the base model only. We use the super resolution model to upsample each generated image independently." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.331, + 0.21, + 0.348 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.356, + 0.47, + 0.492 + ], + "angle": 0, + "content": "In Figs. 6, 7, 8, 9, and 10, we demonstrate that our approach successfully generates consistent high quality zoom sequences for arbitrary relative zoom factors and a diverse set of scenes. Please see our supplementary materials for a full collection of videos. Sec. 5.1 describes how we generate text prompts, Sec. 5.2 demonstrates how our method outperforms diffusion-based outpainting and super-resolution models, and Sec. 5.3 justifies our design decisions with an ablation study." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.5, + 0.303, + 0.516 + ], + "angle": 0, + "content": "5.1. Text Prompt Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.523, + 0.471, + 0.795 + ], + "angle": 0, + "content": "We generate a collection of text prompts that describe scenes at varying levels of scales using a combination of ChatGPT [14] and manual editing. We start with prompting ChatGPT with a description of a scene, and asking it to formulate the sequence of prompts we might need for different zoom levels. While the results from this query are often plausible, they often (1) do not accurately match the corresponding requested scales, or (2) do not match the distribution of text prompts that the text-to-image model is able to most effectively generate. As such, we manually refine the prompts. A comprehensive collection of the prompts used to generate results in the paper are provided in the supplemental materials, along with the initial versions automatically produced by ChatGPT. In the future, we expect LLMs (and in particular, multimodal models) to automatically produce a sequence of prompts well suited for this application. In total, we collect a total of 10 examples, with the prompts sequence length varying form 6 to 16." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.285, + 0.818 + ], + "angle": 0, + "content": "5.2. Baseline Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Fig. 8 compares zoom sequences generated with our method and without (i.e., independently sampling each scale). When compared to our results, the independently-generated images similarly follow the text prompt, but clearly do not correspond to a single consistent underlying scene." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.09, + 0.574, + 0.1 + ], + "angle": 0, + "content": "Zoomed out" + }, + { + "type": "text", + "bbox": [ + 0.825, + 0.09, + 0.88, + 0.099 + ], + "angle": 0, + "content": "Zoomed in" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.105, + 0.888, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.267, + 0.892, + 0.323 + ], + "angle": 0, + "content": "Figure 8. Generated zoom sequences with independent sampling (top) and our multi-scale sampling (bottom). Our method encourages different levels to depict a consistent underlying scene, while not compromising the image quality." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.339, + 0.892, + 0.414 + ], + "angle": 0, + "content": "Next, we compare our method to two autogressive generation approaches for generating zoom sequences: (1) Stable Diffusion's [1] outpainting model and (2) Stable Diffusion's \"upscale\" super-resolution model. We show representative qualitative results in Fig. 9." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.419, + 0.893, + 0.584 + ], + "angle": 0, + "content": "Comparison to progressive outpainting. The outpainting baseline starts with generating the most zoomed-in image and progressively generates coarser scales by downsampling the previous generated image and outpainting the surrounding area. As in our method, the inpainting of each level is conditioned on the corresponding text prompt. In Fig. 9, we show that because of the causality of the autoregressive process, the outpainting approach suffers from gradually accumulating errors, i.e., when a mistake is made at a given step, later outpainting iterations may struggle to produce a consistent image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.589, + 0.892, + 0.725 + ], + "angle": 0, + "content": "Comparison to progressive super-resolution. The superresolution baseline starts with the most zoomed-out image and generates subsequent scales by super-resolving the upscaled central image region, conditioned on the corresponding text prompt. The low resolution input provides strong structural information which constrains the layout of the next zoomed-in image. As we can see in Fig. 9, this super-resolution baseline is not able to synthesize new objects that would only appear in the finer, zoomed-in scales." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.737, + 0.611, + 0.751 + ], + "angle": 0, + "content": "5.3. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.761, + 0.892, + 0.791 + ], + "angle": 0, + "content": "In Fig. 10, we show comparisons to simpler versions of our method to examine the effect of our design decisions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Joint vs. Iterative update. Instead of performing multiscale blending approach, we can instead iteratively cycle through the images in the zoom stack, and perform one sampling step at each level independently. Unlike fully independent sampling, this process does allow for sharing of information between scales, since the steps are still applied to renders from the zoom stack. We find that although this pro" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7179" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.088, + 0.885, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.21, + 0.426, + 0.759, + 0.44 + ], + "angle": 0, + "content": "Figure 9. Comparisons with Stable Diffusion Outpainting and super-resolution (SR) models." + }, + { + "type": "image", + "bbox": [ + 0.087, + 0.463, + 0.466, + 0.747 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.759, + 0.47, + 0.801 + ], + "angle": 0, + "content": "Figure 10. Ablations. We evaluate other options for multi-scale consistency: (1) iteratively updating each level separately, (2) naive multi-scale blending, (3) removing the shared noise." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.82, + 0.468, + 0.85 + ], + "angle": 0, + "content": "duces more consistent results than independent sampling, there remain inconsistencies at stack layer boundaries." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Shared vs. random noise Instead of using a shared noise \\(\\Pi_{\\mathrm{noise}}\\), noise can be sampled independently for each zoom level. We find that this leads to blur in the output samples." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.466, + 0.892, + 0.527 + ], + "angle": 0, + "content": "Comparison with naïve blending. Instead of our multiscale blending, we can instead naively blend the observations together, e.g., as in MultiDiffusion [2]. We find that this leads to blurry outputs at deeper zoom levels." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.548, + 0.738, + 0.563 + ], + "angle": 0, + "content": "6. Discussion & Limitations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.576, + 0.892, + 0.712 + ], + "angle": 0, + "content": "A significant challenge in our work is discovering the appropriate set of text prompts that (1) agree with each other across a set of fixed scales, and (2) can be effectively generated consistently by a given text-to-image model. One possible avenue of improvement could be to, along with sampling, optimize for suitable geometric transformations between successive zoom levels. These transformations could include translation, rotation, and even scale, to find better alignment between the zoom levels and the prompts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.714, + 0.892, + 0.804 + ], + "angle": 0, + "content": "Alternatively, one can optimize the text embeddings, to find better descriptions that correspond to subsequent zoom levels. Or, instead, use the LLM for in-the-loop generation, i.e., by giving LLM the generated image content, and asking it to refine its prompts to produce images which are closer in correspondence given the set of pre-defined scales." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgements. We thank Ben Poole, Jon Barron, Luyang Zhu, Ruiqi Gao, Tong He, Grace Luo, Angjoo Kanazawa, Vickie Ye, Songwei Ge, Keunhong Park, and David Salesin for helpful discussions and feedback. This work was supported in part by UW Reality Lab, Meta, Google, OPPO, and Amazon." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7180" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Stability AI. Stable-diffusion-2-inpainting. https://huggingface.co/stabilityai/stable-diffusion-2-inpainting.2,7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.47, + 0.199 + ], + "angle": 0, + "content": "[2] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. arXiv preprint arXiv:2302.08113, 2023. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.471, + 0.268 + ], + "angle": 0, + "content": "[3] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.269, + 0.471, + 0.311 + ], + "angle": 0, + "content": "[4] Peter J Burt and Edward H Adelson. The laplacian pyramid as a compact image code. In Readings in computer vision, pages 671-679. Elsevier, 1987. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.312, + 0.471, + 0.354 + ], + "angle": 0, + "content": "[5] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.355, + 0.47, + 0.408 + ], + "angle": 0, + "content": "[6] Dave Epstein, Allan Jabri, Ben Poole, Alexei A Efros, and Aleksander Holynski. Diffusion self-guidance for controllable image generation. arXiv preprint arXiv:2306.00986, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.41, + 0.471, + 0.465 + ], + "angle": 0, + "content": "[7] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.466, + 0.471, + 0.508 + ], + "angle": 0, + "content": "[8] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.509, + 0.471, + 0.55 + ], + "angle": 0, + "content": "[9] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.551, + 0.47, + 0.607 + ], + "angle": 0, + "content": "[10] Yuseung Lee, Kunho Kim, Hyunjin Kim, and Minhyuk Sung. Syncdiffusion: Coherent montage via synchronized joint diffusions. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.47, + 0.676 + ], + "angle": 0, + "content": "[11] Zhengqi Li, Qianqian Wang, Noah Snavely, and Angjoo Kanazawa. Infinitenature-zero: Learning perpetual view generation of natural scenes from single images. In European Conference on Computer Vision, pages 515-534. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.47, + 0.759 + ], + "angle": 0, + "content": "[12] Andrew Liu, Richard Tucker, Varun Jampani, Ameesh Makadia, Noah Snavely, and Angjoo Kanazawa. Infinite nature: Perpetual view generation of natural scenes from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14458-14467, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.47, + 0.802 + ], + "angle": 0, + "content": "[13] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. arXiv preprint arXiv:2202.09778, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.47, + 0.831 + ], + "angle": 0, + "content": "[14] OpenAI. Chatgpt [large language model]. https://chat.openai.com/chat.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[15] Ryan Po, Wang Yifan, Vladislav Golyanik, Kfir Aberman, Jonathan T Barron, Amit H Bermano, Eric Ryan Chan, Tali Dekel, Aleksander Holynski, Angjoo Kanazawa, et al. State of the art on diffusion models for visual computing. arXiv preprint arXiv:2310.07204, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[16] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.894, + 0.219 + ], + "angle": 0, + "content": "[17] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[18] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[19] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models. arXiv preprint arXiv:2307.06949, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.375, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[20] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.447, + 0.892, + 0.53 + ], + "angle": 0, + "content": "[21] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.532, + 0.892, + 0.6 + ], + "angle": 0, + "content": "[22] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image super-resolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4713-4726, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.602, + 0.892, + 0.671 + ], + "angle": 0, + "content": "[23] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.673, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[24] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.717, + 0.892, + 0.759 + ], + "angle": 0, + "content": "[25] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.76, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[26] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[27] Luming Tang, Nataniel Ruiz, Chu Qinghao, Yuanzhen Li, Aleksander Holynski, David E Jacobs, Bharath Hariharan, Yael Pritch, Neal Wadhwa, Kfir Aberman, and Michael Rubinstein. Realfill: Reference-driven generation for authentic image completion. arXiv preprint arXiv:2309.16668, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7181" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[28] Shitao Tang, Fuyang Zhang, Jiacheng Chen, Peng Wang, and Yasutaka Furukawa. Mvdiffusion: Enabling holistic multiview image generation with correspondence-aware diffusion. arXiv preprint arXiv:2307.01097, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[29] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.206, + 0.47, + 0.261 + ], + "angle": 0, + "content": "[30] Qinsheng Zhang, Jiaming Song, Xun Huang, Yongxin Chen, and Ming-Yu Liu. Diffcollage: Parallel generation of large content with diffusion models. arXiv preprint arXiv:2303.17076, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "7182" + } + ] +] \ No newline at end of file diff --git a/2024/Generative Powers of Ten/2ccba12d-1f5a-40cf-8686-79a45cb775c2_origin.pdf b/2024/Generative Powers of Ten/2ccba12d-1f5a-40cf-8686-79a45cb775c2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4bf85a56c202b0d4e95c5baf5d6b8afc0037aa67 --- /dev/null +++ b/2024/Generative Powers of Ten/2ccba12d-1f5a-40cf-8686-79a45cb775c2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b28d23ed065eb5da28172f5ffef76ac4f1f74e8cd08a5fdbe46f1686cbfb42fc +size 9369553 diff --git a/2024/Generative Powers of Ten/full.md b/2024/Generative Powers of Ten/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0dc7f1072889a475d2e6127b58a9499232752dea --- /dev/null +++ b/2024/Generative Powers of Ten/full.md @@ -0,0 +1,333 @@ +# Generative Powers of Ten + +Xiaojuan Wang $^{1}$ Janne Kontkanen $^{2}$ Brian Curless $^{1,2}$ Steven M. Seitz $^{1,2}$ Ira Kemelmacher-Shlizerman $^{1,2}$ Ben Mildenhall $^{2}$ Pratul Srinivasan $^{2}$ Dor Verbin $^{2}$ Aleksander Holynski $^{2,3}$ + +1University of Washington 2Google Research 3UC Berkeley powers-of-ten.github.io + +![](images/50180d19baaad31e78afc0d2392daff08fa72b1cc31485e9c8c658d57328b5ff.jpg) + +![](images/c93968793eedac9115882213db7031d5e9f450d07e344f826fa6727edf97c705.jpg) + +![](images/e926d347caa8dc1f46df55655fea69c5650f89e87e00ed42d2d14b33cc1e3832.jpg) + +![](images/6dacc97b07d88436b55f860d7c262390d535f63ff68078a1ec7a77b362ba2f38.jpg) + +![](images/ce99e3d73cecd198574053c0e8a5b6e7de7aa1f01eb02d0ebc2a7cce5f763700.jpg) +Figure 1. Given a series of prompts describing a scene at varying zoom levels, e.g., from a distant galaxy to the surface of an alien planet, our method uses a pre-trained text-to-image diffusion model to generate a continuously zooming video sequence. + +# Abstract + +We present a method that uses a text-to-image model to generate consistent content across multiple image scales, enabling extreme semantic zooms into a scene, e.g. ranging from a wide-angle landscape view of a forest to a macro shot of an insect sitting on one of the tree branches. We achieve this through a joint multi-scale diffusion sampling approach that encourages consistency across different scales while preserving the integrity of each individual sampling process. Since each generated scale is guided by a different text prompt, our method enables deeper levels of zoom than traditional super-resolution methods that may struggle to create new contextual structure at vastly different scales. We compare our method qualitatively with alternative techniques in image super-resolution and outpainting, and show that our method is most effective at generating consistent multi-scale content. + +# 1. Introduction + +Recent advances in text-to-image models [3, 6, 7, 15, 18, 19, 29] have been transformative in enabling applications + +like image generation from a single text prompt. But while digital images exist at a fixed resolution, the real world can be experienced at many different levels of scale. Few things exemplify this better than the classic 1977 short film "Powers of Ten", shown in Figure 2, which showcases the sheer magnitudes of scale that exist in the universe by visualizing a continuous zoom from the outermost depths of the galaxy to the cells inside our bodies1. Unfortunately, producing animations or interactive experiences like these has traditionally required trained artists and many hours of tedious labor—and although we might want to replace this process with a generative model, existing methods have not yet demonstrated the ability to generate consistent content across multiple zoom levels. + +Unlike traditional super-resolution methods, which generate higher-resolution content conditioned on the pixels of the original image, extreme zooms expose entirely new structures, e.g., magnifying a hand to reveal its underlying skin cells. Generating such a zoom requires semantic knowledge of human anatomy. In this paper, we focus on solving this semantic zoom problem, i.e., enabling text- + +![](images/b41583abfe9cc70afc96c0d8222ec83e1b3ea766bca1cde5ce5b124d9743105a.jpg) +Figure 2. Powers of Ten (1977) This documentary film illustrates the relative scale of the universe as a single shot that gradually zooms out from a human to the universe, and then back again to the microscopic molecular level. + +![](images/c1b100c55418feae215da7ec2658f518d55eab12e6264a450a6ff23b0e3825e3.jpg) + +![](images/6817a7b802fd6130cd15068c3eb14534728cff9193d25731e9831e46a11ad452.jpg) + +![](images/4bdda37128e8e51968fb42bd45bca02cb22621797d4dcf221c77d66f0ceb5912.jpg) + +![](images/ac2ea4825cd227dcd1d72290dc3d4886875eab89d76b95a8d7aa1218d3d7d1ca.jpg) + +conditioned multi-scale image generation, to create Powers of Ten-like zoom videos. As input, our method expects a series of text prompts that describe different scales of the scene, and produces as output a multi-scale image representation that can be explored interactively or rendered to a seamless zooming video. These text prompts can be user-defined (allowing for creative control over the content at different zoom levels) or crafted with the help of a large language model (e.g., by querying the model with an image caption and a prompt like "describe what might you see if you zoomed in by $2x$ ) + +At its core, our method relies on a joint sampling algorithm that uses a set of parallel diffusion sampling processes distributed across zoom levels. These sampling processes are coordinated to be consistent through an iterative frequency-band consolidation process, in which intermediate image predictions are consistently combined across scales. Unlike existing approaches that accomplish similar goals by repeatedly increasing the effective image resolution (e.g., through super-resolution or image outpainting), our sampling process jointly optimizes for the content of all scales at once, allowing for both (1) plausible images at each scale and (2) consistent content across scales. Furthermore, existing methods are limited in their ability to explore wide ranges of scale, since they rely primarily on the input image content to determine the added details at subsequent zoom levels. In many cases, image patches contain insufficient contextual information to inform detail at deeper (e.g., $10\mathrm{x}$ or $100\mathrm{x}$ ) zoom levels. On the other hand, our method grounds each scale in a text prompt, allowing for new structures and content to be conceived across extreme zoom levels. In our experiments, we compare our work qualitatively to these existing methods, and demonstrate that the zoom videos that our method produces are notably more consistent. Finally, we showcase a number of ways in which our algorithm can be used, e.g., by conditioning purely on text or grounding the generation in a known (real) image. + +# 2. Prior Work + +Super-resolution and inpainting. Existing text-to-image + +based super resolution models [1, 22] and outpainting models [1, 16, 20, 27] can be adapted to the zoom task as autoregressive processes, i.e., by progressively outpainting a zoomed-in image, or progressively super-resolving a zoomed-out image. One significant drawback of these approaches is that later-generated images have no influence on the previously generated ones, which can often lead to suboptimal results, as certain structures may be entirely incompatible with subsequent levels of detail, causing error accumulation across recurrent network applications. + +Perpetual view generation. Starting from a single view RGB image, perpetual view generation methods like Infinite Nature [11] and InfiniteNature-Zero [12] learn to generate unbounded flythrough videos of natural scenes. These methods differ from our generative zoom in two key ways: (1) they translate the camera in 3D, causing a "fly-through" effect with perspective effects, rather than the "zoom in" our method produces, and (2) they synthesize the fly-through starting from a single image by progressively inpainting unknown parts of novel views, whereas we generate the entire zoom sequence simultaneously and coherently across scales, with text-guided semantic control. + +Diffusion joint sampling for consistent generation. Recent research [2, 10, 28, 30] leverages pretrained diffusion models to generate arbitrary-sized images or panoramas from smaller pieces using joint diffusion processes. These processes involve concurrently generating these multiple images by merging their intermediate results within the sampling process. In particular, DiffCollage [30] introduces a factor graph formulation to express spatial constraints among these images, representing each image as a node, and overlapping areas with additional nodes. Each sampling step involves aggregating individual predictions based on the factor graph. For this to be possible, a given diffusion model needs to be finetuned for different factor nodes. Other works such as MultiDiffusion [2] reconcile different denoising steps by solving for a least squares optimal solution: i.e., averaging the diffusion model predictions at overlapping areas. However, none of these approaches can be applied to our problem, where our jointly sampled images have spatial correspondence at vastly different spatial scales. + +# 3. Preliminaries + +Diffusion models [5, 8, 23-26] generate images from random noise through a sequential sampling process. This sampling process reverses a destructive process that gradually adds Gaussian noise on a clean image $\mathbf{x}$ . The intermediate noisy image at time step $t$ is expressed as: + +$$ +\mathbf {z} _ {t} = \alpha_ {t} \mathbf {x} + \sigma_ {t} \boldsymbol {\epsilon} _ {t}, +$$ + +where $\epsilon_{t} \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ is a standard Gaussian noise, and $\alpha_{t}$ and $\sigma_{t}$ define a fixed noise schedule, with larger $t$ corresponding to more noise. A diffusion model is a neural network $\epsilon_{\theta}$ that predicts the approximate clean image $\hat{\mathbf{x}}$ directly, or equivalently the added noise $\epsilon_{t}$ in $\mathbf{z}_{t}$ . The network is trained with the loss: + +$$ +\mathcal {L} (\boldsymbol {\theta}) = \mathbb {E} _ {t \sim U [ 1, T ], \boldsymbol {\epsilon} _ {t} \sim \mathcal {N} (\mathbf {0}, \mathbf {I})} [ w (t) \| \boldsymbol {\epsilon} _ {\boldsymbol {\theta}} (\mathbf {z} _ {t}; t, y) - \boldsymbol {\epsilon} _ {t} \| _ {2} ^ {2} ], +$$ + +where $y$ is an additional conditioning signal like text [16, 17, 21], and $w(t)$ is a weighting function typically set to 1 [8]. A standard choice for $\epsilon_{\theta}$ is a U-Net with self-attention and cross-attention operations attending to the conditioning $y$ . + +Once the diffusion model is trained, various sampling methods [8, 13, 24] are designed to sample efficiently from the model, starting from pure noise $\mathbf{z}_T \sim \mathcal{N}(\mathbf{0}, \mathbf{I})$ and iteratively denoising it to a clean image. These sampling methods often rely on classifier-free guidance [8], a process which uses a linear combination of the text-conditional and unconditional predictions to achieve better adherence to the conditioning signal: + +$$ +\hat {\boldsymbol {\epsilon}} _ {t} = (1 + \omega) \boldsymbol {\epsilon} _ {\theta} (\mathbf {z} _ {t}; t, y) - \omega \boldsymbol {\epsilon} _ {\theta} (\mathbf {z} _ {t}; t). +$$ + +This revised $\hat{\epsilon}_t$ is used as the noise prediction to update the noisy image $\mathbf{z}_t$ . Given a noisy image and a noise prediction, the estimated clean image $\hat{\mathbf{x}}_t$ is computed as $\hat{\mathbf{x}}_t = (\mathbf{z}_t - \sigma_t\hat{\epsilon}_t) / \alpha_t$ . The iterative update function in the sampling process depends on the sampler used; in this paper we use DDPM [8]. + +# 4. Method + +Let $y_0, \ldots, y_{N-1}$ be a series of prompts describing a single scene at varying, corresponding zoom levels $p_0, \ldots, p_{N-1}$ forming a geometric progression, i.e., $p_i = p^i$ (we typically set $p$ to 2 or 4). Our objective is to generate a sequence of corresponding $H \times W \times C$ images $\mathbf{x}_0, \ldots, \mathbf{x}_{N-1}$ from an existing, pre-trained, text-to-image diffusion model. We aim to generate the entire set of images jointly in a zoom-consistent way. This means that the image $\mathbf{x}_i$ at any specific zoom level $p_i$ , should be consistent with the center $H / p \times W / p$ crop of the zoomed-out image $\mathbf{x}_{i-1}$ . + +We propose a multi-scale joint sampling approach and a corresponding zoom stack representation that gets updated in the diffusion-based sampling process. In Sec. 4.1, we introduce our zoom stack representation and the process that allows us to render it into an image at any given zoom level. In Sec. 4.2, we present an approach for consolidating multiple diffusion estimates into this representation in a consistent way. Finally, in Sec. 4.3, we show how these components are used in the complete sampling process. + +![](images/09ddcebdb8bed0e3cd4b3164df5c681211e845a15a1131301bb9faf1e7d3cbf0.jpg) +Figure 3. Zoom stack. Our representation consists of $N$ layer images $L_{i}$ of constant resolution (left). These layers are arranged in a pyramid-like structure, with layers representing finer details corresponding to a smaller spatial extent (middle). These layers are composited to form an image at any zoom level (right). + +# 4.1. Zoom Stack Representation + +Our zoom stack representation, which we denote by $\mathcal{L} = (L_0,\dots,L_{N - 1})$ , is designed to allow rendering images at any zoom level $p_0,\ldots ,p_{N - 1}$ . The representation, illustrated in Fig. 3, contains $N$ images of shape $H\times W$ , one for each zoom level, where the $i$ th image $L_{i}$ stores the pixels corresponding to the $i$ th zoom level $p_i$ . + +Image rendering. The rendering operator, which we denote by $\Pi_{\mathrm{image}}(\mathcal{L};i)$ , takes a zoom stack $\mathcal{L}$ and returns the image at the $i$ th zoom level $p_i = p^i$ . We denote by $\mathcal{D}_i(\mathbf{x})$ the operator which downscales the image $\mathbf{x}$ by factor $p_i$ , and zero-pads the image back to size $H\times W$ ; and we denote by $M_{i}$ the corresponding $H\times W$ binary image which has value 1 at the center $H / p_{i}\times W / p_{i}$ patch and value 0 at padded pixels. The operator $\mathcal{D}_i$ operates by prefiltering the image with a truncated Gaussian kernel of size $p_i\times p_i$ and resampling with a stride of $p_i$ . As described in Alg. 1, an image $\mathbf{x}_i$ at the $i$ th zoom level is rendered by starting with $L_{i}$ , and iteratively replacing its central $H / p_j\times W / p_j$ crop with $\mathcal{D}_{j - i}(L_j)$ , for $j = i + 1,\dots,N - 1$ . (In Alg. 1 we denote by $\odot$ the elementwise multiplication of a binary mask $M$ with an image.) This process guarantees that rendering at different zoom levels will be consistent at overlapping central regions. + +Noise rendering. At every denoising iteration of DDPM [8], each pixel is corrupted by globally-scaled i.i.d. Gaussian noise $\epsilon \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ . Since we would like images rendered at different zoom levels to be consistent, it is essential to make sure the added noise is also consistent, with overlapping region across different zoom levels sharing the same noise structure. Therefore, we use a rendering operator similar to $\Pi_{\mathrm{image}}$ which converts a set of independent noise images, $\mathcal{E} = (E_0,\dots,E_{N - 1})$ into a single zoom-consistent noise $\epsilon_{i} = \Pi_{\mathrm{noise}}(\mathcal{E};i)$ . However, because downsampling involves prefiltering, which modifies the statistics of the resulting noise, we upscale the $j$ th down-scaled noise component by $p_j / p_i$ to preserve the variance, + +![](images/d9d3fe23998ad3633f582da5ce01a14cad339d4d2d4f5c59f77e16dbe45f1494.jpg) +Figure 4. Overview of a single sampling step. (1) Noisy images $\mathbf{z}_{i,t}$ from each zoom level, along with the respective prompts $y_{i}$ are simultaneously fed into the same pretrained diffusion model, returning estimates of the corresponding clean images $\hat{\mathbf{x}}_{i,t}$ . These images may have inconsistent estimates for the overlapping regions that they all observe. We employ multi-resolution blending to fuse these regions into a consistent zoom stack $\mathcal{L}_t$ and re-render the different zoom levels from the consistent representation. These re-rendered images $\Pi_{\mathrm{image}}(\mathcal{L}_t;i)$ are then used as the clean image estimates in the DDPM sampling step. + +ensuring that the noise satisfies the standard Gaussian distribution assumption, i.e., that $\epsilon_{i} = \Pi_{\mathrm{noise}}(\mathcal{E};i)\sim \mathcal{N}(\mathbf{0},\mathbf{I})$ for all levels $i$ . + +# Algorithm 1 Image and noise rendering at scale $i$ . + +1: Set $\mathbf{x} \gets L_i, \epsilon \sim \mathcal{N}(\mathbf{0}, \mathbf{I})$ +2: for $j = i + 1, \ldots, N - 1$ do +3: $\mathbf{x} \gets M_{j - i} \odot \mathcal{D}_{j - i}(L_j) + (1 - M_{j - i}) \odot \mathbf{x}$ +4: $\epsilon \gets (p_j / p_i)M_{j - i}\odot \mathcal{D}_{j - i}(E_j) + (1 - M_{j - i})\odot \epsilon$ +5: end for +6: return x, $\epsilon$ + +# 4.2. Multi-resolution blending + +Equipped with a method for rendering a zoom stack and sampling noise at any given zoom level, we now describe a mechanism for integrating multiple observations of the same scene $\mathbf{x}_0,\dots ,\mathbf{x}_{N - 1}$ at varying zoom levels $p_0,\dots ,p_{N - 1}$ into a consistent zoom stack $\mathcal{L}$ . This process is a necessary component of the consistent sampling process, as the diffusion model applied at various zoom levels will produce inconsistent content in the overlapping regions. Specifically, the $j$ th zoom stack level $L_{j}$ is used in rendering multiple images at all zoom levels $i\leq j$ and therefore its value should be consistent with multiple image observations (or diffusion model samples), namely $\{\mathbf{x}_i:i\leq j\}$ . The simplest possible solution to this is to naively average the overlapping regions across all observations. This approach, however, results in blurry zoom stack images, since coarser-scale observations of overlapping regions contain fewer pixels, and therefore only lower-frequency information. + +To solve this, we propose an approach we call multi-resolution blending, which uses Laplacian pyramids to selectively fuse the appropriate frequency bands of each ob + +![](images/e4e25f8f3dfa282aeb1252e2d1e9282814a45e534cffbfb0a1279fc907684e12.jpg) +Figure 5. Multi-resolution blending. We produce a consistent estimate for Layer $L_{i}$ in the zoom stack by merging the $H / p_{j} \times W / p_{j}$ central region of the corresponding zoomed out images $\mathbf{x}_j$ for $j \leq i$ . This merging process involves (1) creating a Laplacian pyramid from each observation, and blending together the corresponding frequency bands to create a blended pyramid. This blended pyramid is recomposed into an image, which is used to update the layer $L_{i}$ . + +servation level, which prevents aliasing as well as overblurring. We show an outline of this process in Fig. 5. More concretely, to update the $i$ th layer in the zoom stack, we begin by cropping all samples $j \geq i$ to match with the content of the $i$ th level, and rescaling them back to $H \times W$ . We then analyze each of these $N - i - 1$ images into a Laplacian pyramid [4], and average across corresponding frequency bands + +(see Figure 5), resulting in an average Laplacian pyramid, which can be recomposed into an image and assigned to the $i$ th level of the zoom stack. This process is applied for each layer of the zoom stack $L_{i}$ , collecting from all further zoomed-out levels $j \geq i$ . + +# 4.3. Multi-scale consistent sampling + +Our complete multi-scale joint sampling process is shown in Alg. 2. Fig. 4 illustrates a single sampling step $t$ : Noisy images $\mathbf{z}_{i,t}$ in each zoom level along with the respective prompt $y_{i}$ are fed into the pretrained diffusion model in parallel to predict the noise $\hat{\epsilon}_{i,t-1}$ , and thus to compute the estimated clean images $\hat{\mathbf{x}}_{i,t}$ . Equipped with our multi-resolution blending technique, the clean images are consolidated into a zoom stack, which is then rendered at all zoom levels, yielding consistent images $\Pi_{\mathrm{image}}(\mathcal{L}_t; i)$ . These images are then used in a DDPM update step along with the input $\mathbf{z}_t$ to compute the next $\mathbf{z}_{t-1}$ . + +Algorithm 2 Multi-scale joint sampling. +1: Set $\mathcal{L}_T\gets 0,\mathbf{z}_{i,T}\sim \mathcal{N}(\mathbf{0},\mathbf{I}),\forall i = 0,\dots ,N - 1$ +2: for $t = T,\ldots ,1$ do +3: $\mathcal{E}\sim \mathcal{N}(\mathbf{0},\mathbf{I})$ +4: parfor $i = 0,\dots ,N - 1$ do +5: $\mathbf{x}_{i,t} = \Pi_{\mathrm{image}}(\mathcal{L}_t;i)$ +6: $\epsilon_{i} = \Pi_{\mathrm{noise}}(\mathcal{E};i)$ +7: $\mathbf{z}_{i,t - 1} = \mathrm{DDPM\_update}(\mathbf{z}_{i,t},\mathbf{x}_{i,t},\boldsymbol {\epsilon}_i)$ +8: $\hat{\epsilon}_{i,t - 1} = (1 + \omega)\epsilon_{\theta}(\mathbf{z}_{i,t - 1};t - 1,y_i)$ +9: $-\omega \epsilon_{\theta}(\mathbf{z}_{i,t - 1};t - 1)$ +10: $\hat{\mathbf{x}}_{i,t - 1} = (\mathbf{z}_{i,t - 1} - \sigma_{t - 1}\hat{\epsilon}_{i,t - 1}) / \alpha_{t - 1}$ +11: end parfor +12: $\mathcal{L}_{t - 1}\gets \mathrm{Blending}(\{\hat{\mathbf{x}}_{i,t - 1}\}_{i = 0}^{N - 1})$ +13: end for +14: return $\mathcal{L}_0$ + +# 4.4. Photograph-based Zoom + +In addition to using text prompts to generate the entire zoom stack from scratch, our approach can also generate a sequence zooming into an existing photograph. Given the most zoomed-out input image $\xi$ , we still use Alg. 2, but we additionally update the denoised images to minimize the following loss function before every blending operation: + +$$ +\ell \left(\hat {\mathbf {x}} _ {0, t}, \dots , \hat {\mathbf {x}} _ {N - 1, t}\right) = \sum_ {i = 0} ^ {N - 1} \left\| \mathcal {D} _ {i} \left(\hat {\mathbf {x}} _ {i, t}\right) - M _ {i} \odot \boldsymbol {\xi}\right) \| _ {2} ^ {2}, \tag {1} +$$ + +where, as we defined in Sec. 4.1, $\mathcal{D}_i(\mathbf{x})$ downscales the image $\mathbf{x}$ by a factor $p_i$ and pads the result back to $H\times W$ , and $M_{i}$ is a binary mask with 1 at the center $H / p_{i}\times W / p_{i}$ square and 0 otherwise. Before every blending operation we apply 5 Adam [9] steps at a learning rate of 0.1. This simple optimization-based strategy encourages the estimated + +![](images/eaa9a31aaa72a963efdd16d687e2839fe8ce4a46fdd5adb424edacad7d4b1bd7.jpg) + +![](images/bfc0e925fba6e95d3b312b5fbdf6702b97a3e60de83a9470a800062b87fe7a1a.jpg) + +![](images/8175f2014ca7d774b099d36197995e27f52cdee3d5d7aec44d724608402981ee.jpg) +A girl is holding a maple leaf in front of her face. + +![](images/b18d46674b65b1a8312bc7f3a2622f3ac21de9cbbce1e2fd3dd331e690689ca7.jpg) +A closeup of the surface of skin of the back hand + +![](images/9d0501a2ef72bfb900a737f51b36cbd64be308aefed0294f5981970f46ff2838.jpg) + +![](images/071d14903dd88940f2d07a591ca292f2797a5ed0eb784bd15d6837dcc05bec36.jpg) +Epidermal layer of multiple rows of tiny skin cells +A single round skin cell with its nucleus + +![](images/0baf321a12cec849ef0f3b8528dd0a6f3bc1d93a751962df9949dc30a8c58b16.jpg) +A brightly colored autumn maple leaf + +![](images/cee112abab610b2cdbf1820a059a5972c428f689928bf082b9631d8aeb6d84c1.jpg) +A nucleus within a skin cell +Figure 6. Selected images of our generated zoom sequences beginning with a provided real image. Left: Zoom from a man on a picnic blanket into the skin cells on his hand. Right: Zoom from a girl holding a leaf into the intricate vein patterns on the leaf. Face is blurred for anonymity. + +![](images/f43c5e6302dfec8775403418c6ab804b6e965741d30ad8c5902080f83fcc3f74.jpg) +Orange maple leaf texture with lots of veins +Macrophoto of the veins pattern on the maple leaf + +![](images/65391a8619a1c7df0f938363f722701c01afd365872e4cdad8631aa291f5f3e4.jpg) +Galactic core + +![](images/3102531ab35c33709222e73438bac246320c93ff98847ce366eedad5af1a1bdf.jpg) +Satellite image of the Earth's surface + +![](images/9f6fb647d84af58d6621c80423c91cc1bc85cbb054c37ba3e6ba094e0a725e19.jpg) +An aerial photo capturing Hawaii's islands + +![](images/42ee3e044d7608bd775b99d25786bded72e9226d5ebdab4e5e3d999608992a95.jpg) +A straight road alpine forests on the sides +no paaooZ + +![](images/488449011145f4e274b8f429dbd3d516e8df83b20f1f96cbb5d72e25506b007f.jpg) +Dark starry sky + +![](images/5800ab47dd2dfc938c2370532c62ee36d81affbf53b938a521c8560645934ccb.jpg) +Satellite image of a landmass of the Earth's surface + +![](images/6b63d39a8012abaff02989e5b2aa299057b63f963370a29007e69c7802948ad0.jpg) +An aerial photo of Hawaii's mountains and rain forest + +![](images/c8a937fd92677936c9d3137743bc76b54da5a3fb1b4557f5309683387580ef2e.jpg) +Alpine forest road with Mount Rainier in the end + +![](images/4533042ab7be98b9e46bf4f3dde873bf98af9685961026832f47faaaeae05fc6.jpg) +Far view of alien solar system + +![](images/ab4d258ca1fb869134cce3d1ddb9b91ef456010f9620934ce95c8e941b2ea49e.jpg) +Satellite image of a quaint American countryside + +![](images/5fdc2d268c78cd41e0cca1fdc2c8bb6be16bb50832756d6e2a9b59911c349991.jpg) +An aerial close-up of the volcano's caldera + +![](images/d9221a0a2d01b0c14f47feff75bb53634ee83a43649fd323ddbf0d9edd99bc36.jpg) +Alpine meadows against the massive Mount Rainier + +![](images/2acb4e28e8d0fad8066dec15f7fd5cca3dc92014b69c0e3fe899844472524d7c.jpg) + +![](images/76dece77728ba7efd05a8d601324842fefc1ac0f8372034bb66ca746289ba7ea.jpg) + +![](images/42dfe2c0f608eb2897eef3e6b0b939c5f3e80b5cad0c40dfb63291a9efdef785.jpg) + +![](images/e2a53f25efa467a52f9dcbee6bc5259e2819ab01a6d960865221ed313d3cb72c.jpg) + +![](images/d2ce1aae24960dc53fe311203086bf2570378e92b2d16331600ec8eebd6fe156.jpg) +An exoplanet of a foreign solar system +Top-down aerial image of deserted continents +Figure 7. Selected stills from our generated zoom videos (columns). Please refer to the supplementary materials for complete text prompts. + +![](images/cf14ddaceddf48b468726f94fb1d97c439f1e2392d47abb94e6ef09a41dd142a.jpg) +Satellite image of a foggy forest +Top down view of a lake with a person kayaking + +![](images/87e25a44e21ba0a2c7d664eba517574443f38c785d23e20befc9413da7adb21b.jpg) +An aerial close-up of the rim of a volcano's caldera +A man standing on the edge of a volcano's caldera + +![](images/ff7b55765cb2d66e776cf69e259b8dc51ef7f0892ee2f4802ef5d9ee95a4023a.jpg) +Steep cliffs and rocky outcrops of a snow mountain +A team of climbers climbing on the rugged cliffs +uipuaooZ + +clean images $\{\hat{\mathbf{x}}_{i,t - 1}\}_{i = 0}^{N - 1}$ to match with the content provided in $\pmb{\xi}$ in a zoom-consistent way. We show our generated photograph-based zoom sequences in Fig. 6. + +# 4.5. Implementation Details + +For the underlying text-to-image diffusion model, we use a version of Imagen [21] trained on internal data sources, which is a cascaded diffusion model consisting of (1) a base model conditioned on a text prompt embedding and (2) a super resolution model additionally conditioned the low resolution output from the base model. We use its default DDPM sampling procedure with 256 sampling steps, and we employ our multi-scale joint sampling to the base model only. We use the super resolution model to upsample each generated image independently. + +# 5. Experiments + +In Figs. 6, 7, 8, 9, and 10, we demonstrate that our approach successfully generates consistent high quality zoom sequences for arbitrary relative zoom factors and a diverse set of scenes. Please see our supplementary materials for a full collection of videos. Sec. 5.1 describes how we generate text prompts, Sec. 5.2 demonstrates how our method outperforms diffusion-based outpainting and super-resolution models, and Sec. 5.3 justifies our design decisions with an ablation study. + +# 5.1. Text Prompt Generation + +We generate a collection of text prompts that describe scenes at varying levels of scales using a combination of ChatGPT [14] and manual editing. We start with prompting ChatGPT with a description of a scene, and asking it to formulate the sequence of prompts we might need for different zoom levels. While the results from this query are often plausible, they often (1) do not accurately match the corresponding requested scales, or (2) do not match the distribution of text prompts that the text-to-image model is able to most effectively generate. As such, we manually refine the prompts. A comprehensive collection of the prompts used to generate results in the paper are provided in the supplemental materials, along with the initial versions automatically produced by ChatGPT. In the future, we expect LLMs (and in particular, multimodal models) to automatically produce a sequence of prompts well suited for this application. In total, we collect a total of 10 examples, with the prompts sequence length varying form 6 to 16. + +# 5.2. Baseline Comparisons + +Fig. 8 compares zoom sequences generated with our method and without (i.e., independently sampling each scale). When compared to our results, the independently-generated images similarly follow the text prompt, but clearly do not correspond to a single consistent underlying scene. + +Zoomed out + +Zoomed in + +![](images/23db433e18c53431031553f13b04936af6831f6286e0a5f89ca88b180af0c341.jpg) +Figure 8. Generated zoom sequences with independent sampling (top) and our multi-scale sampling (bottom). Our method encourages different levels to depict a consistent underlying scene, while not compromising the image quality. + +Next, we compare our method to two autogressive generation approaches for generating zoom sequences: (1) Stable Diffusion's [1] outpainting model and (2) Stable Diffusion's "upscale" super-resolution model. We show representative qualitative results in Fig. 9. + +Comparison to progressive outpainting. The outpainting baseline starts with generating the most zoomed-in image and progressively generates coarser scales by downsampling the previous generated image and outpainting the surrounding area. As in our method, the inpainting of each level is conditioned on the corresponding text prompt. In Fig. 9, we show that because of the causality of the autoregressive process, the outpainting approach suffers from gradually accumulating errors, i.e., when a mistake is made at a given step, later outpainting iterations may struggle to produce a consistent image. + +Comparison to progressive super-resolution. The superresolution baseline starts with the most zoomed-out image and generates subsequent scales by super-resolving the upscaled central image region, conditioned on the corresponding text prompt. The low resolution input provides strong structural information which constrains the layout of the next zoomed-in image. As we can see in Fig. 9, this super-resolution baseline is not able to synthesize new objects that would only appear in the finer, zoomed-in scales. + +# 5.3. Ablations + +In Fig. 10, we show comparisons to simpler versions of our method to examine the effect of our design decisions. + +Joint vs. Iterative update. Instead of performing multiscale blending approach, we can instead iteratively cycle through the images in the zoom stack, and perform one sampling step at each level independently. Unlike fully independent sampling, this process does allow for sharing of information between scales, since the steps are still applied to renders from the zoom stack. We find that although this pro + +![](images/a15c9ebdc8fd47af793464f0374eeea9f3f39735fa2304f7eb4732a5d2503ce9.jpg) +Figure 9. Comparisons with Stable Diffusion Outpainting and super-resolution (SR) models. + +![](images/7c9b052a7db879198c21b45b06c744b667218175e889b665f3af0ddd9ddddff0.jpg) +Figure 10. Ablations. We evaluate other options for multi-scale consistency: (1) iteratively updating each level separately, (2) naive multi-scale blending, (3) removing the shared noise. + +duces more consistent results than independent sampling, there remain inconsistencies at stack layer boundaries. + +Shared vs. random noise Instead of using a shared noise $\Pi_{\mathrm{noise}}$ , noise can be sampled independently for each zoom level. We find that this leads to blur in the output samples. + +Comparison with naïve blending. Instead of our multiscale blending, we can instead naively blend the observations together, e.g., as in MultiDiffusion [2]. We find that this leads to blurry outputs at deeper zoom levels. + +# 6. Discussion & Limitations + +A significant challenge in our work is discovering the appropriate set of text prompts that (1) agree with each other across a set of fixed scales, and (2) can be effectively generated consistently by a given text-to-image model. One possible avenue of improvement could be to, along with sampling, optimize for suitable geometric transformations between successive zoom levels. These transformations could include translation, rotation, and even scale, to find better alignment between the zoom levels and the prompts. + +Alternatively, one can optimize the text embeddings, to find better descriptions that correspond to subsequent zoom levels. Or, instead, use the LLM for in-the-loop generation, i.e., by giving LLM the generated image content, and asking it to refine its prompts to produce images which are closer in correspondence given the set of pre-defined scales. + +Acknowledgements. We thank Ben Poole, Jon Barron, Luyang Zhu, Ruiqi Gao, Tong He, Grace Luo, Angjoo Kanazawa, Vickie Ye, Songwei Ge, Keunhong Park, and David Salesin for helpful discussions and feedback. This work was supported in part by UW Reality Lab, Meta, Google, OPPO, and Amazon. + +# References + +[1] Stability AI. Stable-diffusion-2-inpainting. https://huggingface.co/stabilityai/stable-diffusion-2-inpainting.2,7 +[2] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. arXiv preprint arXiv:2302.08113, 2023. 2, 8 +[3] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 1 +[4] Peter J Burt and Edward H Adelson. The laplacian pyramid as a compact image code. In Readings in computer vision, pages 671-679. Elsevier, 1987. 4 +[5] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2 +[6] Dave Epstein, Allan Jabri, Ben Poole, Alexei A Efros, and Aleksander Holynski. Diffusion self-guidance for controllable image generation. arXiv preprint arXiv:2306.00986, 2023. 1 +[7] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 1 +[8] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 3 +[9] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5 +[10] Yuseung Lee, Kunho Kim, Hyunjin Kim, and Minhyuk Sung. Syncdiffusion: Coherent montage via synchronized joint diffusions. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2 +[11] Zhengqi Li, Qianqian Wang, Noah Snavely, and Angjoo Kanazawa. Infinitenature-zero: Learning perpetual view generation of natural scenes from single images. In European Conference on Computer Vision, pages 515-534. Springer, 2022. 2 +[12] Andrew Liu, Richard Tucker, Varun Jampani, Ameesh Makadia, Noah Snavely, and Angjoo Kanazawa. Infinite nature: Perpetual view generation of natural scenes from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14458-14467, 2021. 2 +[13] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. arXiv preprint arXiv:2202.09778, 2022. 3 +[14] OpenAI. Chatgpt [large language model]. https://chat.openai.com/chat.7 +[15] Ryan Po, Wang Yifan, Vladislav Golyanik, Kfir Aberman, Jonathan T Barron, Amit H Bermano, Eric Ryan Chan, Tali Dekel, Aleksander Holynski, Angjoo Kanazawa, et al. State of the art on diffusion models for visual computing. arXiv preprint arXiv:2310.07204, 2023. 1 + +[16] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2, 3 +[17] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3 +[18] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 1 +[19] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models. arXiv preprint arXiv:2307.06949, 2023. 1 +[20] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 2 +[21] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 3, 7 +[22] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image super-resolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4713-4726, 2022. 2 +[23] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2 +[24] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.3 +[25] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. +[26] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2 +[27] Luming Tang, Nataniel Ruiz, Chu Qinghao, Yuanzhen Li, Aleksander Holynski, David E Jacobs, Bharath Hariharan, Yael Pritch, Neal Wadhwa, Kfir Aberman, and Michael Rubinstein. Realfill: Reference-driven generation for authentic image completion. arXiv preprint arXiv:2309.16668, 2023. 2 + +[28] Shitao Tang, Fuyang Zhang, Jiacheng Chen, Peng Wang, and Yasutaka Furukawa. Mvdiffusion: Enabling holistic multiview image generation with correspondence-aware diffusion. arXiv preprint arXiv:2307.01097, 2023. 2 +[29] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1 +[30] Qinsheng Zhang, Jiaming Song, Xun Huang, Yongxin Chen, and Ming-Yu Liu. Diffcollage: Parallel generation of large content with diffusion models. arXiv preprint arXiv:2303.17076, 2023. 2 \ No newline at end of file diff --git a/2024/Generative Powers of Ten/images.zip b/2024/Generative Powers of Ten/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5bfacb75f49d2cbab595a369204650f79c24e67a --- /dev/null +++ b/2024/Generative Powers of Ten/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3b5200a2debeb4f6c53b4376c7c2576e41a5ecd8ac2abea6a1a87f1d1d04ca18 +size 1681048 diff --git a/2024/Generative Powers of Ten/layout.json b/2024/Generative Powers of Ten/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..176ba8cd53a1a64f094eec6c9fc202542aafbfc4 --- /dev/null +++ b/2024/Generative Powers of Ten/layout.json @@ -0,0 +1,10373 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 217, + 103, + 376, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 103, + 376, + 119 + ], + "spans": [ + { + "bbox": [ + 217, + 103, + 376, + 119 + ], + "type": "text", + "content": "Generative Powers of Ten" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 52, + 141, + 569, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 141, + 569, + 171 + ], + "spans": [ + { + "bbox": [ + 52, + 141, + 569, + 171 + ], + "type": "text", + "content": "Xiaojuan Wang" + }, + { + "bbox": [ + 52, + 141, + 569, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 52, + 141, + 569, + 171 + ], + "type": "text", + "content": " Janne Kontkanen" + }, + { + "bbox": [ + 52, + 141, + 569, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 52, + 141, + 569, + 171 + ], + "type": "text", + "content": " Brian Curless" + }, + { + "bbox": [ + 52, + 141, + 569, + 171 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 52, + 141, + 569, + 171 + ], + "type": "text", + "content": " Steven M. Seitz" + }, + { + "bbox": [ + 52, + 141, + 569, + 171 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 52, + 141, + 569, + 171 + ], + "type": "text", + "content": " Ira Kemelmacher-Shlizerman" + }, + { + "bbox": [ + 52, + 141, + 569, + 171 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 52, + 141, + 569, + 171 + ], + "type": "text", + "content": " Ben Mildenhall" + }, + { + "bbox": [ + 52, + 141, + 569, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 52, + 141, + 569, + 171 + ], + "type": "text", + "content": " Pratul Srinivasan" + }, + { + "bbox": [ + 52, + 141, + 569, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 52, + 141, + 569, + 171 + ], + "type": "text", + "content": " Dor Verbin" + }, + { + "bbox": [ + 52, + 141, + 569, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 52, + 141, + 569, + 171 + ], + "type": "text", + "content": " Aleksander Holynski" + }, + { + "bbox": [ + 52, + 141, + 569, + 171 + ], + "type": "inline_equation", + "content": "^{2,3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 173, + 477, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 173, + 477, + 205 + ], + "spans": [ + { + "bbox": [ + 146, + 173, + 477, + 205 + ], + "type": "text", + "content": "1University of Washington 2Google Research 3UC Berkeley powers-of-ten.github.io" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 50, + 234, + 170, + 357 + ], + "blocks": [ + { + "bbox": [ + 50, + 234, + 170, + 357 + ], + "lines": [ + { + "bbox": [ + 50, + 234, + 170, + 357 + ], + "spans": [ + { + "bbox": [ + 50, + 234, + 170, + 357 + ], + "type": "image", + "image_path": "50180d19baaad31e78afc0d2392daff08fa72b1cc31485e9c8c658d57328b5ff.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 173, + 234, + 294, + 354 + ], + "blocks": [ + { + "bbox": [ + 173, + 234, + 294, + 354 + ], + "lines": [ + { + "bbox": [ + 173, + 234, + 294, + 354 + ], + "spans": [ + { + "bbox": [ + 173, + 234, + 294, + 354 + ], + "type": "image", + "image_path": "c93968793eedac9115882213db7031d5e9f450d07e344f826fa6727edf97c705.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 299, + 235, + 419, + 354 + ], + "blocks": [ + { + "bbox": [ + 299, + 235, + 419, + 354 + ], + "lines": [ + { + "bbox": [ + 299, + 235, + 419, + 354 + ], + "spans": [ + { + "bbox": [ + 299, + 235, + 419, + 354 + ], + "type": "image", + "image_path": "e926d347caa8dc1f46df55655fea69c5650f89e87e00ed42d2d14b33cc1e3832.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 425, + 235, + 543, + 354 + ], + "blocks": [ + { + "bbox": [ + 425, + 235, + 543, + 354 + ], + "lines": [ + { + "bbox": [ + 425, + 235, + 543, + 354 + ], + "spans": [ + { + "bbox": [ + 425, + 235, + 543, + 354 + ], + "type": "image", + "image_path": "6dacc97b07d88436b55f860d7c262390d535f63ff68078a1ec7a77b362ba2f38.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 48, + 366, + 542, + 399 + ], + "blocks": [ + { + "bbox": [ + 48, + 366, + 542, + 399 + ], + "lines": [ + { + "bbox": [ + 48, + 366, + 542, + 399 + ], + "spans": [ + { + "bbox": [ + 48, + 366, + 542, + 399 + ], + "type": "image", + "image_path": "ce99e3d73cecd198574053c0e8a5b6e7de7aa1f01eb02d0ebc2a7cce5f763700.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 409, + 546, + 433 + ], + "lines": [ + { + "bbox": [ + 46, + 409, + 546, + 433 + ], + "spans": [ + { + "bbox": [ + 46, + 409, + 546, + 433 + ], + "type": "text", + "content": "Figure 1. Given a series of prompts describing a scene at varying zoom levels, e.g., from a distant galaxy to the surface of an alien planet, our method uses a pre-trained text-to-image diffusion model to generate a continuously zooming video sequence." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 143, + 442, + 192, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 442, + 192, + 455 + ], + "spans": [ + { + "bbox": [ + 143, + 442, + 192, + 455 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 467, + 289, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 467, + 289, + 658 + ], + "spans": [ + { + "bbox": [ + 46, + 467, + 289, + 658 + ], + "type": "text", + "content": "We present a method that uses a text-to-image model to generate consistent content across multiple image scales, enabling extreme semantic zooms into a scene, e.g. ranging from a wide-angle landscape view of a forest to a macro shot of an insect sitting on one of the tree branches. We achieve this through a joint multi-scale diffusion sampling approach that encourages consistency across different scales while preserving the integrity of each individual sampling process. Since each generated scale is guided by a different text prompt, our method enables deeper levels of zoom than traditional super-resolution methods that may struggle to create new contextual structure at vastly different scales. We compare our method qualitatively with alternative techniques in image super-resolution and outpainting, and show that our method is most effective at generating consistent multi-scale content." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 669, + 128, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 128, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 128, + 681 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": "Recent advances in text-to-image models [3, 6, 7, 15, 18, 19, 29] have been transformative in enabling applications" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 443, + 546, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 443, + 546, + 609 + ], + "spans": [ + { + "bbox": [ + 304, + 443, + 546, + 609 + ], + "type": "text", + "content": "like image generation from a single text prompt. But while digital images exist at a fixed resolution, the real world can be experienced at many different levels of scale. Few things exemplify this better than the classic 1977 short film \"Powers of Ten\", shown in Figure 2, which showcases the sheer magnitudes of scale that exist in the universe by visualizing a continuous zoom from the outermost depths of the galaxy to the cells inside our bodies1. Unfortunately, producing animations or interactive experiences like these has traditionally required trained artists and many hours of tedious labor—and although we might want to replace this process with a generative model, existing methods have not yet demonstrated the ability to generate consistent content across multiple zoom levels." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 610, + 547, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 610, + 547, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 610, + 547, + 696 + ], + "type": "text", + "content": "Unlike traditional super-resolution methods, which generate higher-resolution content conditioned on the pixels of the original image, extreme zooms expose entirely new structures, e.g., magnifying a hand to reveal its underlying skin cells. Generating such a zoom requires semantic knowledge of human anatomy. In this paper, we focus on solving this semantic zoom problem, i.e., enabling text-" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 317, + 703, + 529, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 703, + 529, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 703, + 529, + 713 + ], + "type": "text", + "content": "1https://www.youtube.com/watch?v=0fKBhvdjuy0" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7173" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 94, + 166 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 94, + 166 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 94, + 166 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 94, + 166 + ], + "type": "image", + "image_path": "b41583abfe9cc70afc96c0d8222ec83e1b3ea766bca1cde5ce5b124d9743105a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 175, + 287, + 218 + ], + "lines": [ + { + "bbox": [ + 47, + 175, + 287, + 218 + ], + "spans": [ + { + "bbox": [ + 47, + 175, + 287, + 218 + ], + "type": "text", + "content": "Figure 2. Powers of Ten (1977) This documentary film illustrates the relative scale of the universe as a single shot that gradually zooms out from a human to the universe, and then back again to the microscopic molecular level." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 96, + 71, + 142, + 166 + ], + "blocks": [ + { + "bbox": [ + 96, + 71, + 142, + 166 + ], + "lines": [ + { + "bbox": [ + 96, + 71, + 142, + 166 + ], + "spans": [ + { + "bbox": [ + 96, + 71, + 142, + 166 + ], + "type": "image", + "image_path": "c1b100c55418feae215da7ec2658f518d55eab12e6264a450a6ff23b0e3825e3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 144, + 71, + 190, + 166 + ], + "blocks": [ + { + "bbox": [ + 144, + 71, + 190, + 166 + ], + "lines": [ + { + "bbox": [ + 144, + 71, + 190, + 166 + ], + "spans": [ + { + "bbox": [ + 144, + 71, + 190, + 166 + ], + "type": "image", + "image_path": "6817a7b802fd6130cd15068c3eb14534728cff9193d25731e9831e46a11ad452.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 192, + 71, + 239, + 166 + ], + "blocks": [ + { + "bbox": [ + 192, + 71, + 239, + 166 + ], + "lines": [ + { + "bbox": [ + 192, + 71, + 239, + 166 + ], + "spans": [ + { + "bbox": [ + 192, + 71, + 239, + 166 + ], + "type": "image", + "image_path": "4bdda37128e8e51968fb42bd45bca02cb22621797d4dcf221c77d66f0ceb5912.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 241, + 71, + 286, + 166 + ], + "blocks": [ + { + "bbox": [ + 241, + 71, + 286, + 166 + ], + "lines": [ + { + "bbox": [ + 241, + 71, + 286, + 166 + ], + "spans": [ + { + "bbox": [ + 241, + 71, + 286, + 166 + ], + "type": "image", + "image_path": "ac2ea4825cd227dcd1d72290dc3d4886875eab89d76b95a8d7aa1218d3d7d1ca.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 222, + 288, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 222, + 288, + 353 + ], + "spans": [ + { + "bbox": [ + 46, + 222, + 288, + 353 + ], + "type": "text", + "content": "conditioned multi-scale image generation, to create Powers of Ten-like zoom videos. As input, our method expects a series of text prompts that describe different scales of the scene, and produces as output a multi-scale image representation that can be explored interactively or rendered to a seamless zooming video. These text prompts can be user-defined (allowing for creative control over the content at different zoom levels) or crafted with the help of a large language model (e.g., by querying the model with an image caption and a prompt like \"describe what might you see if you zoomed in by " + }, + { + "bbox": [ + 46, + 222, + 288, + 353 + ], + "type": "inline_equation", + "content": "2x" + }, + { + "bbox": [ + 46, + 222, + 288, + 353 + ], + "type": "text", + "content": ")" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 355, + 287, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 355, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 46, + 355, + 287, + 668 + ], + "type": "text", + "content": "At its core, our method relies on a joint sampling algorithm that uses a set of parallel diffusion sampling processes distributed across zoom levels. These sampling processes are coordinated to be consistent through an iterative frequency-band consolidation process, in which intermediate image predictions are consistently combined across scales. Unlike existing approaches that accomplish similar goals by repeatedly increasing the effective image resolution (e.g., through super-resolution or image outpainting), our sampling process jointly optimizes for the content of all scales at once, allowing for both (1) plausible images at each scale and (2) consistent content across scales. Furthermore, existing methods are limited in their ability to explore wide ranges of scale, since they rely primarily on the input image content to determine the added details at subsequent zoom levels. In many cases, image patches contain insufficient contextual information to inform detail at deeper (e.g., " + }, + { + "bbox": [ + 46, + 355, + 287, + 668 + ], + "type": "inline_equation", + "content": "10\\mathrm{x}" + }, + { + "bbox": [ + 46, + 355, + 287, + 668 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 355, + 287, + 668 + ], + "type": "inline_equation", + "content": "100\\mathrm{x}" + }, + { + "bbox": [ + 46, + 355, + 287, + 668 + ], + "type": "text", + "content": ") zoom levels. On the other hand, our method grounds each scale in a text prompt, allowing for new structures and content to be conceived across extreme zoom levels. In our experiments, we compare our work qualitatively to these existing methods, and demonstrate that the zoom videos that our method produces are notably more consistent. Finally, we showcase a number of ways in which our algorithm can be used, e.g., by conditioning purely on text or grounding the generation in a known (real) image." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 680, + 122, + 693 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 680, + 122, + 693 + ], + "spans": [ + { + "bbox": [ + 47, + 680, + 122, + 693 + ], + "type": "text", + "content": "2. Prior Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 701, + 286, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 286, + 714 + ], + "type": "text", + "content": "Super-resolution and inpainting. Existing text-to-image" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "content": "based super resolution models [1, 22] and outpainting models [1, 16, 20, 27] can be adapted to the zoom task as autoregressive processes, i.e., by progressively outpainting a zoomed-in image, or progressively super-resolving a zoomed-out image. One significant drawback of these approaches is that later-generated images have no influence on the previously generated ones, which can often lead to suboptimal results, as certain structures may be entirely incompatible with subsequent levels of detail, causing error accumulation across recurrent network applications." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 197, + 545, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 197, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 304, + 197, + 545, + 340 + ], + "type": "text", + "content": "Perpetual view generation. Starting from a single view RGB image, perpetual view generation methods like Infinite Nature [11] and InfiniteNature-Zero [12] learn to generate unbounded flythrough videos of natural scenes. These methods differ from our generative zoom in two key ways: (1) they translate the camera in 3D, causing a \"fly-through\" effect with perspective effects, rather than the \"zoom in\" our method produces, and (2) they synthesize the fly-through starting from a single image by progressively inpainting unknown parts of novel views, whereas we generate the entire zoom sequence simultaneously and coherently across scales, with text-guided semantic control." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 346, + 545, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 346, + 545, + 584 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 545, + 584 + ], + "type": "text", + "content": "Diffusion joint sampling for consistent generation. Recent research [2, 10, 28, 30] leverages pretrained diffusion models to generate arbitrary-sized images or panoramas from smaller pieces using joint diffusion processes. These processes involve concurrently generating these multiple images by merging their intermediate results within the sampling process. In particular, DiffCollage [30] introduces a factor graph formulation to express spatial constraints among these images, representing each image as a node, and overlapping areas with additional nodes. Each sampling step involves aggregating individual predictions based on the factor graph. For this to be possible, a given diffusion model needs to be finetuned for different factor nodes. Other works such as MultiDiffusion [2] reconcile different denoising steps by solving for a least squares optimal solution: i.e., averaging the diffusion model predictions at overlapping areas. However, none of these approaches can be applied to our problem, where our jointly sampled images have spatial correspondence at vastly different spatial scales." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 604, + 390, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 604, + 390, + 616 + ], + "spans": [ + { + "bbox": [ + 306, + 604, + 390, + 616 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 626, + 545, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 626, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 626, + 545, + 687 + ], + "type": "text", + "content": "Diffusion models [5, 8, 23-26] generate images from random noise through a sequential sampling process. This sampling process reverses a destructive process that gradually adds Gaussian noise on a clean image " + }, + { + "bbox": [ + 304, + 626, + 545, + 687 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 626, + 545, + 687 + ], + "type": "text", + "content": ". The intermediate noisy image at time step " + }, + { + "bbox": [ + 304, + 626, + 545, + 687 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 626, + 545, + 687 + ], + "type": "text", + "content": " is expressed as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 389, + 703, + 462, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 703, + 462, + 713 + ], + "spans": [ + { + "bbox": [ + 389, + 703, + 462, + 713 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {t} = \\alpha_ {t} \\mathbf {x} + \\sigma_ {t} \\boldsymbol {\\epsilon} _ {t},", + "image_path": "a64d5577698837e3513544f5b63bfa0980137d57e6ec156f0775df9a854144c6.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7174" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "type": "inline_equation", + "content": "\\epsilon_{t} \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "type": "text", + "content": " is a standard Gaussian noise, and " + }, + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "type": "inline_equation", + "content": "\\sigma_{t}" + }, + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "type": "text", + "content": " define a fixed noise schedule, with larger " + }, + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "type": "text", + "content": " corresponding to more noise. A diffusion model is a neural network " + }, + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "type": "text", + "content": " that predicts the approximate clean image " + }, + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}" + }, + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "type": "text", + "content": " directly, or equivalently the added noise " + }, + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "type": "inline_equation", + "content": "\\epsilon_{t}" + }, + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{t}" + }, + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "type": "text", + "content": ". The network is trained with the loss:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 156, + 279, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 156, + 279, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 156, + 279, + 171 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\boldsymbol {\\theta}) = \\mathbb {E} _ {t \\sim U [ 1, T ], \\boldsymbol {\\epsilon} _ {t} \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I})} [ w (t) \\| \\boldsymbol {\\epsilon} _ {\\boldsymbol {\\theta}} (\\mathbf {z} _ {t}; t, y) - \\boldsymbol {\\epsilon} _ {t} \\| _ {2} ^ {2} ],", + "image_path": "9faa55aa62b6608ee0df132e6fc7bab8f85406d09508cb0a3b78c8bc3f832985.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 182, + 287, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 182, + 287, + 241 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 287, + 241 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 182, + 287, + 241 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 182, + 287, + 241 + ], + "type": "text", + "content": " is an additional conditioning signal like text [16, 17, 21], and " + }, + { + "bbox": [ + 46, + 182, + 287, + 241 + ], + "type": "inline_equation", + "content": "w(t)" + }, + { + "bbox": [ + 46, + 182, + 287, + 241 + ], + "type": "text", + "content": " is a weighting function typically set to 1 [8]. A standard choice for " + }, + { + "bbox": [ + 46, + 182, + 287, + 241 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 46, + 182, + 287, + 241 + ], + "type": "text", + "content": " is a U-Net with self-attention and cross-attention operations attending to the conditioning " + }, + { + "bbox": [ + 46, + 182, + 287, + 241 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 182, + 287, + 241 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 243, + 288, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 243, + 288, + 339 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 288, + 339 + ], + "type": "text", + "content": "Once the diffusion model is trained, various sampling methods [8, 13, 24] are designed to sample efficiently from the model, starting from pure noise " + }, + { + "bbox": [ + 46, + 243, + 288, + 339 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_T \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})" + }, + { + "bbox": [ + 46, + 243, + 288, + 339 + ], + "type": "text", + "content": " and iteratively denoising it to a clean image. These sampling methods often rely on classifier-free guidance [8], a process which uses a linear combination of the text-conditional and unconditional predictions to achieve better adherence to the conditioning signal:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 89, + 351, + 244, + 364 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 351, + 244, + 364 + ], + "spans": [ + { + "bbox": [ + 89, + 351, + 244, + 364 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {\\epsilon}} _ {t} = (1 + \\omega) \\boldsymbol {\\epsilon} _ {\\theta} (\\mathbf {z} _ {t}; t, y) - \\omega \\boldsymbol {\\epsilon} _ {\\theta} (\\mathbf {z} _ {t}; t).", + "image_path": "099ebab67c69a6004222908397038d6431ef21630906deb8b061faceea0fd246.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 376, + 288, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 288, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 288, + 449 + ], + "type": "text", + "content": "This revised " + }, + { + "bbox": [ + 46, + 376, + 288, + 449 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_t" + }, + { + "bbox": [ + 46, + 376, + 288, + 449 + ], + "type": "text", + "content": " is used as the noise prediction to update the noisy image " + }, + { + "bbox": [ + 46, + 376, + 288, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 46, + 376, + 288, + 449 + ], + "type": "text", + "content": ". Given a noisy image and a noise prediction, the estimated clean image " + }, + { + "bbox": [ + 46, + 376, + 288, + 449 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_t" + }, + { + "bbox": [ + 46, + 376, + 288, + 449 + ], + "type": "text", + "content": " is computed as " + }, + { + "bbox": [ + 46, + 376, + 288, + 449 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_t = (\\mathbf{z}_t - \\sigma_t\\hat{\\epsilon}_t) / \\alpha_t" + }, + { + "bbox": [ + 46, + 376, + 288, + 449 + ], + "type": "text", + "content": ". The iterative update function in the sampling process depends on the sampler used; in this paper we use DDPM [8]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 463, + 103, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 463, + 103, + 475 + ], + "spans": [ + { + "bbox": [ + 47, + 463, + 103, + 475 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 484, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 484, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 484, + 287, + 605 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 46, + 484, + 287, + 605 + ], + "type": "inline_equation", + "content": "y_0, \\ldots, y_{N-1}" + }, + { + "bbox": [ + 46, + 484, + 287, + 605 + ], + "type": "text", + "content": " be a series of prompts describing a single scene at varying, corresponding zoom levels " + }, + { + "bbox": [ + 46, + 484, + 287, + 605 + ], + "type": "inline_equation", + "content": "p_0, \\ldots, p_{N-1}" + }, + { + "bbox": [ + 46, + 484, + 287, + 605 + ], + "type": "text", + "content": " forming a geometric progression, i.e., " + }, + { + "bbox": [ + 46, + 484, + 287, + 605 + ], + "type": "inline_equation", + "content": "p_i = p^i" + }, + { + "bbox": [ + 46, + 484, + 287, + 605 + ], + "type": "text", + "content": " (we typically set " + }, + { + "bbox": [ + 46, + 484, + 287, + 605 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 484, + 287, + 605 + ], + "type": "text", + "content": " to 2 or 4). Our objective is to generate a sequence of corresponding " + }, + { + "bbox": [ + 46, + 484, + 287, + 605 + ], + "type": "inline_equation", + "content": "H \\times W \\times C" + }, + { + "bbox": [ + 46, + 484, + 287, + 605 + ], + "type": "text", + "content": " images " + }, + { + "bbox": [ + 46, + 484, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0, \\ldots, \\mathbf{x}_{N-1}" + }, + { + "bbox": [ + 46, + 484, + 287, + 605 + ], + "type": "text", + "content": " from an existing, pre-trained, text-to-image diffusion model. We aim to generate the entire set of images jointly in a zoom-consistent way. This means that the image " + }, + { + "bbox": [ + 46, + 484, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 46, + 484, + 287, + 605 + ], + "type": "text", + "content": " at any specific zoom level " + }, + { + "bbox": [ + 46, + 484, + 287, + 605 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 46, + 484, + 287, + 605 + ], + "type": "text", + "content": ", should be consistent with the center " + }, + { + "bbox": [ + 46, + 484, + 287, + 605 + ], + "type": "inline_equation", + "content": "H / p \\times W / p" + }, + { + "bbox": [ + 46, + 484, + 287, + 605 + ], + "type": "text", + "content": " crop of the zoomed-out image " + }, + { + "bbox": [ + 46, + 484, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i-1}" + }, + { + "bbox": [ + 46, + 484, + 287, + 605 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "We propose a multi-scale joint sampling approach and a corresponding zoom stack representation that gets updated in the diffusion-based sampling process. In Sec. 4.1, we introduce our zoom stack representation and the process that allows us to render it into an image at any given zoom level. In Sec. 4.2, we present an approach for consolidating multiple diffusion estimates into this representation in a consistent way. Finally, in Sec. 4.3, we show how these components are used in the complete sampling process." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 305, + 69, + 547, + 178 + ], + "blocks": [ + { + "bbox": [ + 305, + 69, + 547, + 178 + ], + "lines": [ + { + "bbox": [ + 305, + 69, + 547, + 178 + ], + "spans": [ + { + "bbox": [ + 305, + 69, + 547, + 178 + ], + "type": "image", + "image_path": "09ddcebdb8bed0e3cd4b3164df5c681211e845a15a1131301bb9faf1e7d3cbf0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 185, + 547, + 242 + ], + "lines": [ + { + "bbox": [ + 304, + 185, + 547, + 242 + ], + "spans": [ + { + "bbox": [ + 304, + 185, + 547, + 242 + ], + "type": "text", + "content": "Figure 3. Zoom stack. Our representation consists of " + }, + { + "bbox": [ + 304, + 185, + 547, + 242 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 185, + 547, + 242 + ], + "type": "text", + "content": " layer images " + }, + { + "bbox": [ + 304, + 185, + 547, + 242 + ], + "type": "inline_equation", + "content": "L_{i}" + }, + { + "bbox": [ + 304, + 185, + 547, + 242 + ], + "type": "text", + "content": " of constant resolution (left). These layers are arranged in a pyramid-like structure, with layers representing finer details corresponding to a smaller spatial extent (middle). These layers are composited to form an image at any zoom level (right)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 255, + 459, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 255, + 459, + 268 + ], + "spans": [ + { + "bbox": [ + 305, + 255, + 459, + 268 + ], + "type": "text", + "content": "4.1. Zoom Stack Representation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 274, + 547, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 274, + 547, + 346 + ], + "spans": [ + { + "bbox": [ + 304, + 274, + 547, + 346 + ], + "type": "text", + "content": "Our zoom stack representation, which we denote by " + }, + { + "bbox": [ + 304, + 274, + 547, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = (L_0,\\dots,L_{N - 1})" + }, + { + "bbox": [ + 304, + 274, + 547, + 346 + ], + "type": "text", + "content": ", is designed to allow rendering images at any zoom level " + }, + { + "bbox": [ + 304, + 274, + 547, + 346 + ], + "type": "inline_equation", + "content": "p_0,\\ldots ,p_{N - 1}" + }, + { + "bbox": [ + 304, + 274, + 547, + 346 + ], + "type": "text", + "content": ". The representation, illustrated in Fig. 3, contains " + }, + { + "bbox": [ + 304, + 274, + 547, + 346 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 274, + 547, + 346 + ], + "type": "text", + "content": " images of shape " + }, + { + "bbox": [ + 304, + 274, + 547, + 346 + ], + "type": "inline_equation", + "content": "H\\times W" + }, + { + "bbox": [ + 304, + 274, + 547, + 346 + ], + "type": "text", + "content": ", one for each zoom level, where the " + }, + { + "bbox": [ + 304, + 274, + 547, + 346 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 274, + 547, + 346 + ], + "type": "text", + "content": "th image " + }, + { + "bbox": [ + 304, + 274, + 547, + 346 + ], + "type": "inline_equation", + "content": "L_{i}" + }, + { + "bbox": [ + 304, + 274, + 547, + 346 + ], + "type": "text", + "content": " stores the pixels corresponding to the " + }, + { + "bbox": [ + 304, + 274, + 547, + 346 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 274, + 547, + 346 + ], + "type": "text", + "content": "th zoom level " + }, + { + "bbox": [ + 304, + 274, + 547, + 346 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 304, + 274, + 547, + 346 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 350, + 546, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 350, + 546, + 554 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 546, + 554 + ], + "type": "text", + "content": "Image rendering. The rendering operator, which we denote by " + }, + { + "bbox": [ + 304, + 350, + 546, + 554 + ], + "type": "inline_equation", + "content": "\\Pi_{\\mathrm{image}}(\\mathcal{L};i)" + }, + { + "bbox": [ + 304, + 350, + 546, + 554 + ], + "type": "text", + "content": ", takes a zoom stack " + }, + { + "bbox": [ + 304, + 350, + 546, + 554 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 304, + 350, + 546, + 554 + ], + "type": "text", + "content": " and returns the image at the " + }, + { + "bbox": [ + 304, + 350, + 546, + 554 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 350, + 546, + 554 + ], + "type": "text", + "content": "th zoom level " + }, + { + "bbox": [ + 304, + 350, + 546, + 554 + ], + "type": "inline_equation", + "content": "p_i = p^i" + }, + { + "bbox": [ + 304, + 350, + 546, + 554 + ], + "type": "text", + "content": ". We denote by " + }, + { + "bbox": [ + 304, + 350, + 546, + 554 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_i(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 350, + 546, + 554 + ], + "type": "text", + "content": " the operator which downscales the image " + }, + { + "bbox": [ + 304, + 350, + 546, + 554 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 350, + 546, + 554 + ], + "type": "text", + "content": " by factor " + }, + { + "bbox": [ + 304, + 350, + 546, + 554 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 304, + 350, + 546, + 554 + ], + "type": "text", + "content": ", and zero-pads the image back to size " + }, + { + "bbox": [ + 304, + 350, + 546, + 554 + ], + "type": "inline_equation", + "content": "H\\times W" + }, + { + "bbox": [ + 304, + 350, + 546, + 554 + ], + "type": "text", + "content": "; and we denote by " + }, + { + "bbox": [ + 304, + 350, + 546, + 554 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 304, + 350, + 546, + 554 + ], + "type": "text", + "content": " the corresponding " + }, + { + "bbox": [ + 304, + 350, + 546, + 554 + ], + "type": "inline_equation", + "content": "H\\times W" + }, + { + "bbox": [ + 304, + 350, + 546, + 554 + ], + "type": "text", + "content": " binary image which has value 1 at the center " + }, + { + "bbox": [ + 304, + 350, + 546, + 554 + ], + "type": "inline_equation", + "content": "H / p_{i}\\times W / p_{i}" + }, + { + "bbox": [ + 304, + 350, + 546, + 554 + ], + "type": "text", + "content": " patch and value 0 at padded pixels. The operator " + }, + { + "bbox": [ + 304, + 350, + 546, + 554 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_i" + }, + { + "bbox": [ + 304, + 350, + 546, + 554 + ], + "type": "text", + "content": " operates by prefiltering the image with a truncated Gaussian kernel of size " + }, + { + "bbox": [ + 304, + 350, + 546, + 554 + ], + "type": "inline_equation", + "content": "p_i\\times p_i" + }, + { + "bbox": [ + 304, + 350, + 546, + 554 + ], + "type": "text", + "content": " and resampling with a stride of " + }, + { + "bbox": [ + 304, + 350, + 546, + 554 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 304, + 350, + 546, + 554 + ], + "type": "text", + "content": ". As described in Alg. 1, an image " + }, + { + "bbox": [ + 304, + 350, + 546, + 554 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 304, + 350, + 546, + 554 + ], + "type": "text", + "content": " at the " + }, + { + "bbox": [ + 304, + 350, + 546, + 554 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 350, + 546, + 554 + ], + "type": "text", + "content": "th zoom level is rendered by starting with " + }, + { + "bbox": [ + 304, + 350, + 546, + 554 + ], + "type": "inline_equation", + "content": "L_{i}" + }, + { + "bbox": [ + 304, + 350, + 546, + 554 + ], + "type": "text", + "content": ", and iteratively replacing its central " + }, + { + "bbox": [ + 304, + 350, + 546, + 554 + ], + "type": "inline_equation", + "content": "H / p_j\\times W / p_j" + }, + { + "bbox": [ + 304, + 350, + 546, + 554 + ], + "type": "text", + "content": " crop with " + }, + { + "bbox": [ + 304, + 350, + 546, + 554 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{j - i}(L_j)" + }, + { + "bbox": [ + 304, + 350, + 546, + 554 + ], + "type": "text", + "content": ", for " + }, + { + "bbox": [ + 304, + 350, + 546, + 554 + ], + "type": "inline_equation", + "content": "j = i + 1,\\dots,N - 1" + }, + { + "bbox": [ + 304, + 350, + 546, + 554 + ], + "type": "text", + "content": ". (In Alg. 1 we denote by " + }, + { + "bbox": [ + 304, + 350, + 546, + 554 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 304, + 350, + 546, + 554 + ], + "type": "text", + "content": " the elementwise multiplication of a binary mask " + }, + { + "bbox": [ + 304, + 350, + 546, + 554 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 350, + 546, + 554 + ], + "type": "text", + "content": " with an image.) This process guarantees that rendering at different zoom levels will be consistent at overlapping central regions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 558, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 547, + 714 + ], + "type": "text", + "content": "Noise rendering. At every denoising iteration of DDPM [8], each pixel is corrupted by globally-scaled i.i.d. Gaussian noise " + }, + { + "bbox": [ + 304, + 558, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 304, + 558, + 547, + 714 + ], + "type": "text", + "content": ". Since we would like images rendered at different zoom levels to be consistent, it is essential to make sure the added noise is also consistent, with overlapping region across different zoom levels sharing the same noise structure. Therefore, we use a rendering operator similar to " + }, + { + "bbox": [ + 304, + 558, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\Pi_{\\mathrm{image}}" + }, + { + "bbox": [ + 304, + 558, + 547, + 714 + ], + "type": "text", + "content": " which converts a set of independent noise images, " + }, + { + "bbox": [ + 304, + 558, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{E} = (E_0,\\dots,E_{N - 1})" + }, + { + "bbox": [ + 304, + 558, + 547, + 714 + ], + "type": "text", + "content": " into a single zoom-consistent noise " + }, + { + "bbox": [ + 304, + 558, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon_{i} = \\Pi_{\\mathrm{noise}}(\\mathcal{E};i)" + }, + { + "bbox": [ + 304, + 558, + 547, + 714 + ], + "type": "text", + "content": ". However, because downsampling involves prefiltering, which modifies the statistics of the resulting noise, we upscale the " + }, + { + "bbox": [ + 304, + 558, + 547, + 714 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 558, + 547, + 714 + ], + "type": "text", + "content": "th down-scaled noise component by " + }, + { + "bbox": [ + 304, + 558, + 547, + 714 + ], + "type": "inline_equation", + "content": "p_j / p_i" + }, + { + "bbox": [ + 304, + 558, + 547, + 714 + ], + "type": "text", + "content": " to preserve the variance," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7175" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 69, + 547, + 208 + ], + "blocks": [ + { + "bbox": [ + 52, + 69, + 547, + 208 + ], + "lines": [ + { + "bbox": [ + 52, + 69, + 547, + 208 + ], + "spans": [ + { + "bbox": [ + 52, + 69, + 547, + 208 + ], + "type": "image", + "image_path": "d9d3fe23998ad3633f582da5ce01a14cad339d4d2d4f5c59f77e16dbe45f1494.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 217, + 547, + 274 + ], + "lines": [ + { + "bbox": [ + 46, + 217, + 547, + 274 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 547, + 274 + ], + "type": "text", + "content": "Figure 4. Overview of a single sampling step. (1) Noisy images " + }, + { + "bbox": [ + 46, + 217, + 547, + 274 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{i,t}" + }, + { + "bbox": [ + 46, + 217, + 547, + 274 + ], + "type": "text", + "content": " from each zoom level, along with the respective prompts " + }, + { + "bbox": [ + 46, + 217, + 547, + 274 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 46, + 217, + 547, + 274 + ], + "type": "text", + "content": " are simultaneously fed into the same pretrained diffusion model, returning estimates of the corresponding clean images " + }, + { + "bbox": [ + 46, + 217, + 547, + 274 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_{i,t}" + }, + { + "bbox": [ + 46, + 217, + 547, + 274 + ], + "type": "text", + "content": ". These images may have inconsistent estimates for the overlapping regions that they all observe. We employ multi-resolution blending to fuse these regions into a consistent zoom stack " + }, + { + "bbox": [ + 46, + 217, + 547, + 274 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_t" + }, + { + "bbox": [ + 46, + 217, + 547, + 274 + ], + "type": "text", + "content": " and re-render the different zoom levels from the consistent representation. These re-rendered images " + }, + { + "bbox": [ + 46, + 217, + 547, + 274 + ], + "type": "inline_equation", + "content": "\\Pi_{\\mathrm{image}}(\\mathcal{L}_t;i)" + }, + { + "bbox": [ + 46, + 217, + 547, + 274 + ], + "type": "text", + "content": " are then used as the clean image estimates in the DDPM sampling step." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 283, + 288, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 283, + 288, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 283, + 288, + 319 + ], + "type": "text", + "content": "ensuring that the noise satisfies the standard Gaussian distribution assumption, i.e., that " + }, + { + "bbox": [ + 46, + 283, + 288, + 319 + ], + "type": "inline_equation", + "content": "\\epsilon_{i} = \\Pi_{\\mathrm{noise}}(\\mathcal{E};i)\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 46, + 283, + 288, + 319 + ], + "type": "text", + "content": " for all levels " + }, + { + "bbox": [ + 46, + 283, + 288, + 319 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 283, + 288, + 319 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 327, + 251, + 340 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 251, + 340 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 251, + 340 + ], + "type": "text", + "content": "Algorithm 1 Image and noise rendering at scale " + }, + { + "bbox": [ + 48, + 327, + 251, + 340 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 48, + 327, + 251, + 340 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 347, + 285, + 420 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 53, + 347, + 168, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 347, + 168, + 360 + ], + "spans": [ + { + "bbox": [ + 53, + 347, + 168, + 360 + ], + "type": "text", + "content": "1: Set " + }, + { + "bbox": [ + 53, + 347, + 168, + 360 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\gets L_i, \\epsilon \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 361, + 182, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 361, + 182, + 372 + ], + "spans": [ + { + "bbox": [ + 53, + 361, + 182, + 372 + ], + "type": "text", + "content": "2: for " + }, + { + "bbox": [ + 53, + 361, + 182, + 372 + ], + "type": "inline_equation", + "content": "j = i + 1, \\ldots, N - 1" + }, + { + "bbox": [ + 53, + 361, + 182, + 372 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 372, + 257, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 372, + 257, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 372, + 257, + 384 + ], + "type": "text", + "content": "3: " + }, + { + "bbox": [ + 53, + 372, + 257, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\gets M_{j - i} \\odot \\mathcal{D}_{j - i}(L_j) + (1 - M_{j - i}) \\odot \\mathbf{x}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 384, + 285, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 384, + 285, + 397 + ], + "spans": [ + { + "bbox": [ + 53, + 384, + 285, + 397 + ], + "type": "text", + "content": "4: " + }, + { + "bbox": [ + 53, + 384, + 285, + 397 + ], + "type": "inline_equation", + "content": "\\epsilon \\gets (p_j / p_i)M_{j - i}\\odot \\mathcal{D}_{j - i}(E_j) + (1 - M_{j - i})\\odot \\epsilon" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 397, + 98, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 397, + 98, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 397, + 98, + 407 + ], + "type": "text", + "content": "5: end for" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 409, + 113, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 409, + 113, + 420 + ], + "spans": [ + { + "bbox": [ + 53, + 409, + 113, + 420 + ], + "type": "text", + "content": "6: return x, " + }, + { + "bbox": [ + 53, + 409, + 113, + 420 + ], + "type": "inline_equation", + "content": "\\epsilon" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 444, + 192, + 457 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 444, + 192, + 457 + ], + "spans": [ + { + "bbox": [ + 47, + 444, + 192, + 457 + ], + "type": "text", + "content": "4.2. Multi-resolution blending" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 462, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 287, + 676 + ], + "type": "text", + "content": "Equipped with a method for rendering a zoom stack and sampling noise at any given zoom level, we now describe a mechanism for integrating multiple observations of the same scene " + }, + { + "bbox": [ + 46, + 462, + 287, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0,\\dots ,\\mathbf{x}_{N - 1}" + }, + { + "bbox": [ + 46, + 462, + 287, + 676 + ], + "type": "text", + "content": " at varying zoom levels " + }, + { + "bbox": [ + 46, + 462, + 287, + 676 + ], + "type": "inline_equation", + "content": "p_0,\\dots ,p_{N - 1}" + }, + { + "bbox": [ + 46, + 462, + 287, + 676 + ], + "type": "text", + "content": " into a consistent zoom stack " + }, + { + "bbox": [ + 46, + 462, + 287, + 676 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 46, + 462, + 287, + 676 + ], + "type": "text", + "content": ". This process is a necessary component of the consistent sampling process, as the diffusion model applied at various zoom levels will produce inconsistent content in the overlapping regions. Specifically, the " + }, + { + "bbox": [ + 46, + 462, + 287, + 676 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 462, + 287, + 676 + ], + "type": "text", + "content": "th zoom stack level " + }, + { + "bbox": [ + 46, + 462, + 287, + 676 + ], + "type": "inline_equation", + "content": "L_{j}" + }, + { + "bbox": [ + 46, + 462, + 287, + 676 + ], + "type": "text", + "content": " is used in rendering multiple images at all zoom levels " + }, + { + "bbox": [ + 46, + 462, + 287, + 676 + ], + "type": "inline_equation", + "content": "i\\leq j" + }, + { + "bbox": [ + 46, + 462, + 287, + 676 + ], + "type": "text", + "content": " and therefore its value should be consistent with multiple image observations (or diffusion model samples), namely " + }, + { + "bbox": [ + 46, + 462, + 287, + 676 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_i:i\\leq j\\}" + }, + { + "bbox": [ + 46, + 462, + 287, + 676 + ], + "type": "text", + "content": ". The simplest possible solution to this is to naively average the overlapping regions across all observations. This approach, however, results in blurry zoom stack images, since coarser-scale observations of overlapping regions contain fewer pixels, and therefore only lower-frequency information." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "content": "To solve this, we propose an approach we call multi-resolution blending, which uses Laplacian pyramids to selectively fuse the appropriate frequency bands of each ob" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 309, + 281, + 545, + 524 + ], + "blocks": [ + { + "bbox": [ + 309, + 281, + 545, + 524 + ], + "lines": [ + { + "bbox": [ + 309, + 281, + 545, + 524 + ], + "spans": [ + { + "bbox": [ + 309, + 281, + 545, + 524 + ], + "type": "image", + "image_path": "e4e25f8f3dfa282aeb1252e2d1e9282814a45e534cffbfb0a1279fc907684e12.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 533, + 547, + 622 + ], + "lines": [ + { + "bbox": [ + 304, + 533, + 547, + 622 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 547, + 622 + ], + "type": "text", + "content": "Figure 5. Multi-resolution blending. We produce a consistent estimate for Layer " + }, + { + "bbox": [ + 304, + 533, + 547, + 622 + ], + "type": "inline_equation", + "content": "L_{i}" + }, + { + "bbox": [ + 304, + 533, + 547, + 622 + ], + "type": "text", + "content": " in the zoom stack by merging the " + }, + { + "bbox": [ + 304, + 533, + 547, + 622 + ], + "type": "inline_equation", + "content": "H / p_{j} \\times W / p_{j}" + }, + { + "bbox": [ + 304, + 533, + 547, + 622 + ], + "type": "text", + "content": " central region of the corresponding zoomed out images " + }, + { + "bbox": [ + 304, + 533, + 547, + 622 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_j" + }, + { + "bbox": [ + 304, + 533, + 547, + 622 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 304, + 533, + 547, + 622 + ], + "type": "inline_equation", + "content": "j \\leq i" + }, + { + "bbox": [ + 304, + 533, + 547, + 622 + ], + "type": "text", + "content": ". This merging process involves (1) creating a Laplacian pyramid from each observation, and blending together the corresponding frequency bands to create a blended pyramid. This blended pyramid is recomposed into an image, which is used to update the layer " + }, + { + "bbox": [ + 304, + 533, + 547, + 622 + ], + "type": "inline_equation", + "content": "L_{i}" + }, + { + "bbox": [ + 304, + 533, + 547, + 622 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "servation level, which prevents aliasing as well as overblurring. We show an outline of this process in Fig. 5. More concretely, to update the " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "th layer in the zoom stack, we begin by cropping all samples " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "j \\geq i" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": " to match with the content of the " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "th level, and rescaling them back to " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": ". We then analyze each of these " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "N - i - 1" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": " images into a Laplacian pyramid [4], and average across corresponding frequency bands" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7176" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 133 + ], + "type": "text", + "content": "(see Figure 5), resulting in an average Laplacian pyramid, which can be recomposed into an image and assigned to the " + }, + { + "bbox": [ + 46, + 72, + 289, + 133 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 72, + 289, + 133 + ], + "type": "text", + "content": "th level of the zoom stack. This process is applied for each layer of the zoom stack " + }, + { + "bbox": [ + 46, + 72, + 289, + 133 + ], + "type": "inline_equation", + "content": "L_{i}" + }, + { + "bbox": [ + 46, + 72, + 289, + 133 + ], + "type": "text", + "content": ", collecting from all further zoomed-out levels " + }, + { + "bbox": [ + 46, + 72, + 289, + 133 + ], + "type": "inline_equation", + "content": "j \\geq i" + }, + { + "bbox": [ + 46, + 72, + 289, + 133 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 138, + 218, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 138, + 218, + 152 + ], + "spans": [ + { + "bbox": [ + 47, + 138, + 218, + 152 + ], + "type": "text", + "content": "4.3. Multi-scale consistent sampling" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 156, + 290, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 156, + 290, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 156, + 290, + 289 + ], + "type": "text", + "content": "Our complete multi-scale joint sampling process is shown in Alg. 2. Fig. 4 illustrates a single sampling step " + }, + { + "bbox": [ + 46, + 156, + 290, + 289 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 156, + 290, + 289 + ], + "type": "text", + "content": ": Noisy images " + }, + { + "bbox": [ + 46, + 156, + 290, + 289 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{i,t}" + }, + { + "bbox": [ + 46, + 156, + 290, + 289 + ], + "type": "text", + "content": " in each zoom level along with the respective prompt " + }, + { + "bbox": [ + 46, + 156, + 290, + 289 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 46, + 156, + 290, + 289 + ], + "type": "text", + "content": " are fed into the pretrained diffusion model in parallel to predict the noise " + }, + { + "bbox": [ + 46, + 156, + 290, + 289 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_{i,t-1}" + }, + { + "bbox": [ + 46, + 156, + 290, + 289 + ], + "type": "text", + "content": ", and thus to compute the estimated clean images " + }, + { + "bbox": [ + 46, + 156, + 290, + 289 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_{i,t}" + }, + { + "bbox": [ + 46, + 156, + 290, + 289 + ], + "type": "text", + "content": ". Equipped with our multi-resolution blending technique, the clean images are consolidated into a zoom stack, which is then rendered at all zoom levels, yielding consistent images " + }, + { + "bbox": [ + 46, + 156, + 290, + 289 + ], + "type": "inline_equation", + "content": "\\Pi_{\\mathrm{image}}(\\mathcal{L}_t; i)" + }, + { + "bbox": [ + 46, + 156, + 290, + 289 + ], + "type": "text", + "content": ". These images are then used in a DDPM update step along with the input " + }, + { + "bbox": [ + 46, + 156, + 290, + 289 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 46, + 156, + 290, + 289 + ], + "type": "text", + "content": " to compute the next " + }, + { + "bbox": [ + 46, + 156, + 290, + 289 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{t-1}" + }, + { + "bbox": [ + 46, + 156, + 290, + 289 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "type": "code", + "bbox": [ + 52, + 318, + 261, + 488 + ], + "blocks": [ + { + "bbox": [ + 48, + 299, + 212, + 312 + ], + "lines": [ + { + "bbox": [ + 48, + 299, + 212, + 312 + ], + "spans": [ + { + "bbox": [ + 48, + 299, + 212, + 312 + ], + "type": "text", + "content": "Algorithm 2 Multi-scale joint sampling." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 52, + 318, + 261, + 488 + ], + "lines": [ + { + "bbox": [ + 52, + 318, + 261, + 488 + ], + "spans": [ + { + "bbox": [ + 52, + 318, + 261, + 488 + ], + "type": "text", + "content": "1: Set " + }, + { + "bbox": [ + 52, + 318, + 261, + 488 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_T\\gets 0,\\mathbf{z}_{i,T}\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I}),\\forall i = 0,\\dots ,N - 1" + }, + { + "bbox": [ + 52, + 318, + 261, + 488 + ], + "type": "text", + "content": " \n2: for " + }, + { + "bbox": [ + 52, + 318, + 261, + 488 + ], + "type": "inline_equation", + "content": "t = T,\\ldots ,1" + }, + { + "bbox": [ + 52, + 318, + 261, + 488 + ], + "type": "text", + "content": " do \n3: " + }, + { + "bbox": [ + 52, + 318, + 261, + 488 + ], + "type": "inline_equation", + "content": "\\mathcal{E}\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 52, + 318, + 261, + 488 + ], + "type": "text", + "content": " \n4: parfor " + }, + { + "bbox": [ + 52, + 318, + 261, + 488 + ], + "type": "inline_equation", + "content": "i = 0,\\dots ,N - 1" + }, + { + "bbox": [ + 52, + 318, + 261, + 488 + ], + "type": "text", + "content": " do \n5: " + }, + { + "bbox": [ + 52, + 318, + 261, + 488 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i,t} = \\Pi_{\\mathrm{image}}(\\mathcal{L}_t;i)" + }, + { + "bbox": [ + 52, + 318, + 261, + 488 + ], + "type": "text", + "content": " \n6: " + }, + { + "bbox": [ + 52, + 318, + 261, + 488 + ], + "type": "inline_equation", + "content": "\\epsilon_{i} = \\Pi_{\\mathrm{noise}}(\\mathcal{E};i)" + }, + { + "bbox": [ + 52, + 318, + 261, + 488 + ], + "type": "text", + "content": " \n7: " + }, + { + "bbox": [ + 52, + 318, + 261, + 488 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{i,t - 1} = \\mathrm{DDPM\\_update}(\\mathbf{z}_{i,t},\\mathbf{x}_{i,t},\\boldsymbol {\\epsilon}_i)" + }, + { + "bbox": [ + 52, + 318, + 261, + 488 + ], + "type": "text", + "content": " \n8: " + }, + { + "bbox": [ + 52, + 318, + 261, + 488 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_{i,t - 1} = (1 + \\omega)\\epsilon_{\\theta}(\\mathbf{z}_{i,t - 1};t - 1,y_i)" + }, + { + "bbox": [ + 52, + 318, + 261, + 488 + ], + "type": "text", + "content": " \n9: " + }, + { + "bbox": [ + 52, + 318, + 261, + 488 + ], + "type": "inline_equation", + "content": "-\\omega \\epsilon_{\\theta}(\\mathbf{z}_{i,t - 1};t - 1)" + }, + { + "bbox": [ + 52, + 318, + 261, + 488 + ], + "type": "text", + "content": " \n10: " + }, + { + "bbox": [ + 52, + 318, + 261, + 488 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_{i,t - 1} = (\\mathbf{z}_{i,t - 1} - \\sigma_{t - 1}\\hat{\\epsilon}_{i,t - 1}) / \\alpha_{t - 1}" + }, + { + "bbox": [ + 52, + 318, + 261, + 488 + ], + "type": "text", + "content": " \n11: end parfor \n12: " + }, + { + "bbox": [ + 52, + 318, + 261, + 488 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{t - 1}\\gets \\mathrm{Blending}(\\{\\hat{\\mathbf{x}}_{i,t - 1}\\}_{i = 0}^{N - 1})" + }, + { + "bbox": [ + 52, + 318, + 261, + 488 + ], + "type": "text", + "content": " \n13: end for \n14: return " + }, + { + "bbox": [ + 52, + 318, + 261, + 488 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_0" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "code_body" + } + ], + "index": 4, + "sub_type": "algorithm" + }, + { + "bbox": [ + 47, + 507, + 186, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 507, + 186, + 521 + ], + "spans": [ + { + "bbox": [ + 47, + 507, + 186, + 521 + ], + "type": "text", + "content": "4.4. Photograph-based Zoom" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 525, + 287, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 525, + 287, + 598 + ], + "spans": [ + { + "bbox": [ + 46, + 525, + 287, + 598 + ], + "type": "text", + "content": "In addition to using text prompts to generate the entire zoom stack from scratch, our approach can also generate a sequence zooming into an existing photograph. Given the most zoomed-out input image " + }, + { + "bbox": [ + 46, + 525, + 287, + 598 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 46, + 525, + 287, + 598 + ], + "type": "text", + "content": ", we still use Alg. 2, but we additionally update the denoised images to minimize the following loss function before every blending operation:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 603, + 287, + 637 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 603, + 287, + 637 + ], + "spans": [ + { + "bbox": [ + 58, + 603, + 287, + 637 + ], + "type": "interline_equation", + "content": "\\ell \\left(\\hat {\\mathbf {x}} _ {0, t}, \\dots , \\hat {\\mathbf {x}} _ {N - 1, t}\\right) = \\sum_ {i = 0} ^ {N - 1} \\left\\| \\mathcal {D} _ {i} \\left(\\hat {\\mathbf {x}} _ {i, t}\\right) - M _ {i} \\odot \\boldsymbol {\\xi}\\right) \\| _ {2} ^ {2}, \\tag {1}", + "image_path": "cf8a0930a0dba3efbc4a4add95807841141aa079ac8b4a3b0882c1ea26b65144.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": "where, as we defined in Sec. 4.1, " + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_i(\\mathbf{x})" + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": " downscales the image " + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": " by a factor " + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": " and pads the result back to " + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "H\\times W" + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": " is a binary mask with 1 at the center " + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "H / p_{i}\\times W / p_{i}" + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": " square and 0 otherwise. Before every blending operation we apply 5 Adam [9] steps at a learning rate of 0.1. This simple optimization-based strategy encourages the estimated" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 70, + 420, + 182 + ], + "blocks": [ + { + "bbox": [ + 310, + 70, + 420, + 182 + ], + "lines": [ + { + "bbox": [ + 310, + 70, + 420, + 182 + ], + "spans": [ + { + "bbox": [ + 310, + 70, + 420, + 182 + ], + "type": "image", + "image_path": "eaa9a31aaa72a963efdd16d687e2839fe8ce4a46fdd5adb424edacad7d4b1bd7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 312, + 182, + 417, + 298 + ], + "blocks": [ + { + "bbox": [ + 312, + 182, + 417, + 298 + ], + "lines": [ + { + "bbox": [ + 312, + 182, + 417, + 298 + ], + "spans": [ + { + "bbox": [ + 312, + 182, + 417, + 298 + ], + "type": "image", + "image_path": "bfc0e925fba6e95d3b312b5fbdf6702b97a3e60de83a9470a800062b87fe7a1a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 421, + 70, + 533, + 182 + ], + "blocks": [ + { + "bbox": [ + 421, + 70, + 533, + 182 + ], + "lines": [ + { + "bbox": [ + 421, + 70, + 533, + 182 + ], + "spans": [ + { + "bbox": [ + 421, + 70, + 533, + 182 + ], + "type": "image", + "image_path": "8175f2014ca7d774b099d36197995e27f52cdee3d5d7aec44d724608402981ee.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 427, + 182, + 525, + 187 + ], + "lines": [ + { + "bbox": [ + 427, + 182, + 525, + 187 + ], + "spans": [ + { + "bbox": [ + 427, + 182, + 525, + 187 + ], + "type": "text", + "content": "A girl is holding a maple leaf in front of her face." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 309, + 304, + 420, + 415 + ], + "blocks": [ + { + "bbox": [ + 315, + 299, + 416, + 304 + ], + "lines": [ + { + "bbox": [ + 315, + 299, + 416, + 304 + ], + "spans": [ + { + "bbox": [ + 315, + 299, + 416, + 304 + ], + "type": "text", + "content": "A closeup of the surface of skin of the back hand" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 309, + 304, + 420, + 415 + ], + "lines": [ + { + "bbox": [ + 309, + 304, + 420, + 415 + ], + "spans": [ + { + "bbox": [ + 309, + 304, + 420, + 415 + ], + "type": "image", + "image_path": "b18d46674b65b1a8312bc7f3a2622f3ac21de9cbbce1e2fd3dd331e690689ca7.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 421, + 187, + 531, + 298 + ], + "blocks": [ + { + "bbox": [ + 421, + 187, + 531, + 298 + ], + "lines": [ + { + "bbox": [ + 421, + 187, + 531, + 298 + ], + "spans": [ + { + "bbox": [ + 421, + 187, + 531, + 298 + ], + "type": "image", + "image_path": "9d0501a2ef72bfb900a737f51b36cbd64be308aefed0294f5981970f46ff2838.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 309, + 422, + 420, + 533 + ], + "blocks": [ + { + "bbox": [ + 313, + 416, + 416, + 422 + ], + "lines": [ + { + "bbox": [ + 313, + 416, + 416, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 416, + 416, + 422 + ], + "type": "text", + "content": "Epidermal layer of multiple rows of tiny skin cells" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 309, + 422, + 420, + 533 + ], + "lines": [ + { + "bbox": [ + 309, + 422, + 420, + 533 + ], + "spans": [ + { + "bbox": [ + 309, + 422, + 420, + 533 + ], + "type": "image", + "image_path": "071d14903dd88940f2d07a591ca292f2797a5ed0eb784bd15d6837dcc05bec36.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 533, + 406, + 538 + ], + "lines": [ + { + "bbox": [ + 323, + 533, + 406, + 538 + ], + "spans": [ + { + "bbox": [ + 323, + 533, + 406, + 538 + ], + "type": "text", + "content": "A single round skin cell with its nucleus" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 421, + 304, + 531, + 415 + ], + "blocks": [ + { + "bbox": [ + 437, + 299, + 515, + 304 + ], + "lines": [ + { + "bbox": [ + 437, + 299, + 515, + 304 + ], + "spans": [ + { + "bbox": [ + 437, + 299, + 515, + 304 + ], + "type": "text", + "content": "A brightly colored autumn maple leaf" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 421, + 304, + 531, + 415 + ], + "lines": [ + { + "bbox": [ + 421, + 304, + 531, + 415 + ], + "spans": [ + { + "bbox": [ + 421, + 304, + 531, + 415 + ], + "type": "image", + "image_path": "0baf321a12cec849ef0f3b8528dd0a6f3bc1d93a751962df9949dc30a8c58b16.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 309, + 539, + 420, + 649 + ], + "blocks": [ + { + "bbox": [ + 309, + 539, + 420, + 649 + ], + "lines": [ + { + "bbox": [ + 309, + 539, + 420, + 649 + ], + "spans": [ + { + "bbox": [ + 309, + 539, + 420, + 649 + ], + "type": "image", + "image_path": "cee112abab610b2cdbf1820a059a5972c428f689928bf082b9631d8aeb6d84c1.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 650, + 393, + 656 + ], + "lines": [ + { + "bbox": [ + 335, + 650, + 393, + 656 + ], + "spans": [ + { + "bbox": [ + 335, + 650, + 393, + 656 + ], + "type": "text", + "content": "A nucleus within a skin cell" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 309, + 667, + 546, + 723 + ], + "lines": [ + { + "bbox": [ + 309, + 667, + 546, + 723 + ], + "spans": [ + { + "bbox": [ + 309, + 667, + 546, + 723 + ], + "type": "text", + "content": "Figure 6. Selected images of our generated zoom sequences beginning with a provided real image. Left: Zoom from a man on a picnic blanket into the skin cells on his hand. Right: Zoom from a girl holding a leaf into the intricate vein patterns on the leaf. Face is blurred for anonymity." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 421, + 422, + 531, + 533 + ], + "blocks": [ + { + "bbox": [ + 432, + 416, + 520, + 422 + ], + "lines": [ + { + "bbox": [ + 432, + 416, + 520, + 422 + ], + "spans": [ + { + "bbox": [ + 432, + 416, + 520, + 422 + ], + "type": "text", + "content": "Orange maple leaf texture with lots of veins" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 421, + 422, + 531, + 533 + ], + "lines": [ + { + "bbox": [ + 421, + 422, + 531, + 533 + ], + "spans": [ + { + "bbox": [ + 421, + 422, + 531, + 533 + ], + "type": "image", + "image_path": "f43c5e6302dfec8775403418c6ab804b6e965741d30ad8c5902080f83fcc3f74.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 425, + 533, + 526, + 539 + ], + "lines": [ + { + "bbox": [ + 425, + 533, + 526, + 539 + ], + "spans": [ + { + "bbox": [ + 425, + 533, + 526, + 539 + ], + "type": "text", + "content": "Macrophoto of the veins pattern on the maple leaf" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 537, + 71, + 545, + 107 + ], + "type": "aside_text", + "angle": 90, + "lines": [ + { + "bbox": [ + 537, + 71, + 545, + 107 + ], + "spans": [ + { + "bbox": [ + 537, + 71, + 545, + 107 + ], + "type": "text", + "content": "Zoomed out" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 537, + 624, + 545, + 658 + ], + "type": "aside_text", + "angle": 90, + "lines": [ + { + "bbox": [ + 537, + 624, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 537, + 624, + 545, + 658 + ], + "type": "text", + "content": "Zoomed in" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7177" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 176, + 186 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 176, + 186 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 176, + 186 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 176, + 186 + ], + "type": "image", + "image_path": "65391a8619a1c7df0f938363f722701c01afd365872e4cdad8631aa291f5f3e4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 103, + 186, + 131, + 192 + ], + "lines": [ + { + "bbox": [ + 103, + 186, + 131, + 192 + ], + "spans": [ + { + "bbox": [ + 103, + 186, + 131, + 192 + ], + "type": "text", + "content": "Galactic core" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 176, + 70, + 290, + 185 + ], + "blocks": [ + { + "bbox": [ + 176, + 70, + 290, + 185 + ], + "lines": [ + { + "bbox": [ + 176, + 70, + 290, + 185 + ], + "spans": [ + { + "bbox": [ + 176, + 70, + 290, + 185 + ], + "type": "image", + "image_path": "3102531ab35c33709222e73438bac246320c93ff98847ce366eedad5af1a1bdf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 196, + 186, + 270, + 192 + ], + "lines": [ + { + "bbox": [ + 196, + 186, + 270, + 192 + ], + "spans": [ + { + "bbox": [ + 196, + 186, + 270, + 192 + ], + "type": "text", + "content": "Satellite image of the Earth's surface" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 291, + 70, + 407, + 186 + ], + "blocks": [ + { + "bbox": [ + 291, + 70, + 407, + 186 + ], + "lines": [ + { + "bbox": [ + 291, + 70, + 407, + 186 + ], + "spans": [ + { + "bbox": [ + 291, + 70, + 407, + 186 + ], + "type": "image", + "image_path": "9f6fb647d84af58d6621c80423c91cc1bc85cbb054c37ba3e6ba094e0a725e19.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 186, + 392, + 193 + ], + "lines": [ + { + "bbox": [ + 306, + 186, + 392, + 193 + ], + "spans": [ + { + "bbox": [ + 306, + 186, + 392, + 193 + ], + "type": "text", + "content": "An aerial photo capturing Hawaii's islands" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 407, + 70, + 523, + 186 + ], + "blocks": [ + { + "bbox": [ + 407, + 70, + 523, + 186 + ], + "lines": [ + { + "bbox": [ + 407, + 70, + 523, + 186 + ], + "spans": [ + { + "bbox": [ + 407, + 70, + 523, + 186 + ], + "type": "image", + "image_path": "42ee3e044d7608bd775b99d25786bded72e9226d5ebdab4e5e3d999608992a95.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 422, + 186, + 506, + 192 + ], + "lines": [ + { + "bbox": [ + 422, + 186, + 506, + 192 + ], + "spans": [ + { + "bbox": [ + 422, + 186, + 506, + 192 + ], + "type": "text", + "content": "A straight road alpine forests on the sides" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 529, + 72, + 537, + 108 + ], + "lines": [ + { + "bbox": [ + 529, + 72, + 537, + 108 + ], + "spans": [ + { + "bbox": [ + 529, + 72, + 537, + 108 + ], + "type": "text", + "content": "no paaooZ" + } + ] + } + ], + "index": 40, + "angle": 270, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 59, + 193, + 175, + 308 + ], + "blocks": [ + { + "bbox": [ + 59, + 193, + 175, + 308 + ], + "lines": [ + { + "bbox": [ + 59, + 193, + 175, + 308 + ], + "spans": [ + { + "bbox": [ + 59, + 193, + 175, + 308 + ], + "type": "image", + "image_path": "488449011145f4e274b8f429dbd3d516e8df83b20f1f96cbb5d72e25506b007f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 101, + 308, + 133, + 313 + ], + "lines": [ + { + "bbox": [ + 101, + 308, + 133, + 313 + ], + "spans": [ + { + "bbox": [ + 101, + 308, + 133, + 313 + ], + "type": "text", + "content": "Dark starry sky" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 175, + 193, + 290, + 308 + ], + "blocks": [ + { + "bbox": [ + 175, + 193, + 290, + 308 + ], + "lines": [ + { + "bbox": [ + 175, + 193, + 290, + 308 + ], + "spans": [ + { + "bbox": [ + 175, + 193, + 290, + 308 + ], + "type": "image", + "image_path": "5800ab47dd2dfc938c2370532c62ee36d81affbf53b938a521c8560645934ccb.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 308, + 285, + 314 + ], + "lines": [ + { + "bbox": [ + 181, + 308, + 285, + 314 + ], + "spans": [ + { + "bbox": [ + 181, + 308, + 285, + 314 + ], + "type": "text", + "content": "Satellite image of a landmass of the Earth's surface" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 291, + 193, + 406, + 308 + ], + "blocks": [ + { + "bbox": [ + 291, + 193, + 406, + 308 + ], + "lines": [ + { + "bbox": [ + 291, + 193, + 406, + 308 + ], + "spans": [ + { + "bbox": [ + 291, + 193, + 406, + 308 + ], + "type": "image", + "image_path": "6b63d39a8012abaff02989e5b2aa299057b63f963370a29007e69c7802948ad0.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 295, + 308, + 403, + 314 + ], + "lines": [ + { + "bbox": [ + 295, + 308, + 403, + 314 + ], + "spans": [ + { + "bbox": [ + 295, + 308, + 403, + 314 + ], + "type": "text", + "content": "An aerial photo of Hawaii's mountains and rain forest" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 407, + 193, + 523, + 308 + ], + "blocks": [ + { + "bbox": [ + 407, + 193, + 523, + 308 + ], + "lines": [ + { + "bbox": [ + 407, + 193, + 523, + 308 + ], + "spans": [ + { + "bbox": [ + 407, + 193, + 523, + 308 + ], + "type": "image", + "image_path": "c8a937fd92677936c9d3137743bc76b54da5a3fb1b4557f5309683387580ef2e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 415, + 308, + 514, + 314 + ], + "lines": [ + { + "bbox": [ + 415, + 308, + 514, + 314 + ], + "spans": [ + { + "bbox": [ + 415, + 308, + 514, + 314 + ], + "type": "text", + "content": "Alpine forest road with Mount Rainier in the end" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 59, + 315, + 175, + 430 + ], + "blocks": [ + { + "bbox": [ + 59, + 315, + 175, + 430 + ], + "lines": [ + { + "bbox": [ + 59, + 315, + 175, + 430 + ], + "spans": [ + { + "bbox": [ + 59, + 315, + 175, + 430 + ], + "type": "image", + "image_path": "4533042ab7be98b9e46bf4f3dde873bf98af9685961026832f47faaaeae05fc6.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 86, + 430, + 148, + 436 + ], + "lines": [ + { + "bbox": [ + 86, + 430, + 148, + 436 + ], + "spans": [ + { + "bbox": [ + 86, + 430, + 148, + 436 + ], + "type": "text", + "content": "Far view of alien solar system" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 175, + 315, + 290, + 430 + ], + "blocks": [ + { + "bbox": [ + 175, + 315, + 290, + 430 + ], + "lines": [ + { + "bbox": [ + 175, + 315, + 290, + 430 + ], + "spans": [ + { + "bbox": [ + 175, + 315, + 290, + 430 + ], + "type": "image", + "image_path": "ab4d258ca1fb869134cce3d1ddb9b91ef456010f9620934ce95c8e941b2ea49e.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 430, + 282, + 436 + ], + "lines": [ + { + "bbox": [ + 184, + 430, + 282, + 436 + ], + "spans": [ + { + "bbox": [ + 184, + 430, + 282, + 436 + ], + "type": "text", + "content": "Satellite image of a quaint American countryside" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 291, + 315, + 407, + 430 + ], + "blocks": [ + { + "bbox": [ + 291, + 315, + 407, + 430 + ], + "lines": [ + { + "bbox": [ + 291, + 315, + 407, + 430 + ], + "spans": [ + { + "bbox": [ + 291, + 315, + 407, + 430 + ], + "type": "image", + "image_path": "5fdc2d268c78cd41e0cca1fdc2c8bb6be16bb50832756d6e2a9b59911c349991.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 430, + 392, + 436 + ], + "lines": [ + { + "bbox": [ + 306, + 430, + 392, + 436 + ], + "spans": [ + { + "bbox": [ + 306, + 430, + 392, + 436 + ], + "type": "text", + "content": "An aerial close-up of the volcano's caldera" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 407, + 315, + 523, + 430 + ], + "blocks": [ + { + "bbox": [ + 407, + 315, + 523, + 430 + ], + "lines": [ + { + "bbox": [ + 407, + 315, + 523, + 430 + ], + "spans": [ + { + "bbox": [ + 407, + 315, + 523, + 430 + ], + "type": "image", + "image_path": "d9221a0a2d01b0c14f47feff75bb53634ee83a43649fd323ddbf0d9edd99bc36.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 413, + 430, + 517, + 435 + ], + "lines": [ + { + "bbox": [ + 413, + 430, + 517, + 435 + ], + "spans": [ + { + "bbox": [ + 413, + 430, + 517, + 435 + ], + "type": "text", + "content": "Alpine meadows against the massive Mount Rainier" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 59, + 437, + 175, + 552 + ], + "blocks": [ + { + "bbox": [ + 59, + 437, + 175, + 552 + ], + "lines": [ + { + "bbox": [ + 59, + 437, + 175, + 552 + ], + "spans": [ + { + "bbox": [ + 59, + 437, + 175, + 552 + ], + "type": "image", + "image_path": "2acb4e28e8d0fad8066dec15f7fd5cca3dc92014b69c0e3fe899844472524d7c.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 175, + 437, + 290, + 552 + ], + "blocks": [ + { + "bbox": [ + 175, + 437, + 290, + 552 + ], + "lines": [ + { + "bbox": [ + 175, + 437, + 290, + 552 + ], + "spans": [ + { + "bbox": [ + 175, + 437, + 290, + 552 + ], + "type": "image", + "image_path": "76dece77728ba7efd05a8d601324842fefc1ac0f8372034bb66ca746289ba7ea.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 291, + 437, + 407, + 552 + ], + "blocks": [ + { + "bbox": [ + 291, + 437, + 407, + 552 + ], + "lines": [ + { + "bbox": [ + 291, + 437, + 407, + 552 + ], + "spans": [ + { + "bbox": [ + 291, + 437, + 407, + 552 + ], + "type": "image", + "image_path": "42dfe2c0f608eb2897eef3e6b0b939c5f3e80b5cad0c40dfb63291a9efdef785.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 407, + 437, + 523, + 552 + ], + "blocks": [ + { + "bbox": [ + 407, + 437, + 523, + 552 + ], + "lines": [ + { + "bbox": [ + 407, + 437, + 523, + 552 + ], + "spans": [ + { + "bbox": [ + 407, + 437, + 523, + 552 + ], + "type": "image", + "image_path": "e2a53f25efa467a52f9dcbee6bc5259e2819ab01a6d960865221ed313d3cb72c.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 59, + 559, + 175, + 673 + ], + "blocks": [ + { + "bbox": [ + 79, + 553, + 156, + 559 + ], + "lines": [ + { + "bbox": [ + 79, + 553, + 156, + 559 + ], + "spans": [ + { + "bbox": [ + 79, + 553, + 156, + 559 + ], + "type": "text", + "content": "An exoplanet of a foreign solar system" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 559, + 175, + 673 + ], + "lines": [ + { + "bbox": [ + 59, + 559, + 175, + 673 + ], + "spans": [ + { + "bbox": [ + 59, + 559, + 175, + 673 + ], + "type": "image", + "image_path": "d2ce1aae24960dc53fe311203086bf2570378e92b2d16331600ec8eebd6fe156.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 71, + 674, + 163, + 681 + ], + "lines": [ + { + "bbox": [ + 71, + 674, + 163, + 681 + ], + "spans": [ + { + "bbox": [ + 71, + 674, + 163, + 681 + ], + "type": "text", + "content": "Top-down aerial image of deserted continents" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 48, + 692, + 545, + 704 + ], + "lines": [ + { + "bbox": [ + 48, + 692, + 545, + 704 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 545, + 704 + ], + "type": "text", + "content": "Figure 7. Selected stills from our generated zoom videos (columns). Please refer to the supplementary materials for complete text prompts." + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 175, + 559, + 290, + 673 + ], + "blocks": [ + { + "bbox": [ + 200, + 553, + 265, + 559 + ], + "lines": [ + { + "bbox": [ + 200, + 553, + 265, + 559 + ], + "spans": [ + { + "bbox": [ + 200, + 553, + 265, + 559 + ], + "type": "text", + "content": "Satellite image of a foggy forest" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 175, + 559, + 290, + 673 + ], + "lines": [ + { + "bbox": [ + 175, + 559, + 290, + 673 + ], + "spans": [ + { + "bbox": [ + 175, + 559, + 290, + 673 + ], + "type": "image", + "image_path": "cf14ddaceddf48b468726f94fb1d97c439f1e2392d47abb94e6ef09a41dd142a.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 674, + 282, + 681 + ], + "lines": [ + { + "bbox": [ + 184, + 674, + 282, + 681 + ], + "spans": [ + { + "bbox": [ + 184, + 674, + 282, + 681 + ], + "type": "text", + "content": "Top down view of a lake with a person kayaking" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 291, + 559, + 407, + 673 + ], + "blocks": [ + { + "bbox": [ + 297, + 553, + 400, + 559 + ], + "lines": [ + { + "bbox": [ + 297, + 553, + 400, + 559 + ], + "spans": [ + { + "bbox": [ + 297, + 553, + 400, + 559 + ], + "type": "text", + "content": "An aerial close-up of the rim of a volcano's caldera" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 291, + 559, + 407, + 673 + ], + "lines": [ + { + "bbox": [ + 291, + 559, + 407, + 673 + ], + "spans": [ + { + "bbox": [ + 291, + 559, + 407, + 673 + ], + "type": "image", + "image_path": "87e25a44e21ba0a2c7d664eba517574443f38c785d23e20befc9413da7adb21b.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 298, + 674, + 400, + 681 + ], + "lines": [ + { + "bbox": [ + 298, + 674, + 400, + 681 + ], + "spans": [ + { + "bbox": [ + 298, + 674, + 400, + 681 + ], + "type": "text", + "content": "A man standing on the edge of a volcano's caldera" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 407, + 559, + 523, + 673 + ], + "blocks": [ + { + "bbox": [ + 413, + 553, + 516, + 559 + ], + "lines": [ + { + "bbox": [ + 413, + 553, + 516, + 559 + ], + "spans": [ + { + "bbox": [ + 413, + 553, + 516, + 559 + ], + "type": "text", + "content": "Steep cliffs and rocky outcrops of a snow mountain" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 407, + 559, + 523, + 673 + ], + "lines": [ + { + "bbox": [ + 407, + 559, + 523, + 673 + ], + "spans": [ + { + "bbox": [ + 407, + 559, + 523, + 673 + ], + "type": "image", + "image_path": "ff7b55765cb2d66e776cf69e259b8dc51ef7f0892ee2f4802ef5d9ee95a4023a.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 415, + 674, + 514, + 681 + ], + "lines": [ + { + "bbox": [ + 415, + 674, + 514, + 681 + ], + "spans": [ + { + "bbox": [ + 415, + 674, + 514, + 681 + ], + "type": "text", + "content": "A team of climbers climbing on the rugged cliffs" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 529, + 648, + 536, + 680 + ], + "lines": [ + { + "bbox": [ + 529, + 648, + 536, + 680 + ], + "spans": [ + { + "bbox": [ + 529, + 648, + 536, + 680 + ], + "type": "text", + "content": "uipuaooZ" + } + ] + } + ], + "index": 41, + "angle": 270, + "type": "image_caption" + } + ], + "index": 38 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7178" + } + ] + } + ], + "index": 43 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 287, + 109 + ], + "type": "text", + "content": "clean images " + }, + { + "bbox": [ + 47, + 71, + 287, + 109 + ], + "type": "inline_equation", + "content": "\\{\\hat{\\mathbf{x}}_{i,t - 1}\\}_{i = 0}^{N - 1}" + }, + { + "bbox": [ + 47, + 71, + 287, + 109 + ], + "type": "text", + "content": " to match with the content provided in " + }, + { + "bbox": [ + 47, + 71, + 287, + 109 + ], + "type": "inline_equation", + "content": "\\pmb{\\xi}" + }, + { + "bbox": [ + 47, + 71, + 287, + 109 + ], + "type": "text", + "content": " in a zoom-consistent way. We show our generated photograph-based zoom sequences in Fig. 6." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 114, + 180, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 114, + 180, + 126 + ], + "spans": [ + { + "bbox": [ + 47, + 114, + 180, + 126 + ], + "type": "text", + "content": "4.5. Implementation Details" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 133, + 288, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 288, + 252 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 288, + 252 + ], + "type": "text", + "content": "For the underlying text-to-image diffusion model, we use a version of Imagen [21] trained on internal data sources, which is a cascaded diffusion model consisting of (1) a base model conditioned on a text prompt embedding and (2) a super resolution model additionally conditioned the low resolution output from the base model. We use its default DDPM sampling procedure with 256 sampling steps, and we employ our multi-scale joint sampling to the base model only. We use the super resolution model to upsample each generated image independently." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 262, + 128, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 262, + 128, + 275 + ], + "spans": [ + { + "bbox": [ + 47, + 262, + 128, + 275 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 281, + 287, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 287, + 389 + ], + "type": "text", + "content": "In Figs. 6, 7, 8, 9, and 10, we demonstrate that our approach successfully generates consistent high quality zoom sequences for arbitrary relative zoom factors and a diverse set of scenes. Please see our supplementary materials for a full collection of videos. Sec. 5.1 describes how we generate text prompts, Sec. 5.2 demonstrates how our method outperforms diffusion-based outpainting and super-resolution models, and Sec. 5.3 justifies our design decisions with an ablation study." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 396, + 185, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 396, + 185, + 408 + ], + "spans": [ + { + "bbox": [ + 47, + 396, + 185, + 408 + ], + "type": "text", + "content": "5.1. Text Prompt Generation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 414, + 288, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 414, + 288, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 414, + 288, + 629 + ], + "type": "text", + "content": "We generate a collection of text prompts that describe scenes at varying levels of scales using a combination of ChatGPT [14] and manual editing. We start with prompting ChatGPT with a description of a scene, and asking it to formulate the sequence of prompts we might need for different zoom levels. While the results from this query are often plausible, they often (1) do not accurately match the corresponding requested scales, or (2) do not match the distribution of text prompts that the text-to-image model is able to most effectively generate. As such, we manually refine the prompts. A comprehensive collection of the prompts used to generate results in the paper are provided in the supplemental materials, along with the initial versions automatically produced by ChatGPT. In the future, we expect LLMs (and in particular, multimodal models) to automatically produce a sequence of prompts well suited for this application. In total, we collect a total of 10 examples, with the prompts sequence length varying form 6 to 16." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 635, + 174, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 174, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 174, + 647 + ], + "type": "text", + "content": "5.2. Baseline Comparisons" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": "Fig. 8 compares zoom sequences generated with our method and without (i.e., independently sampling each scale). When compared to our results, the independently-generated images similarly follow the text prompt, but clearly do not correspond to a single consistent underlying scene." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 71, + 351, + 79 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 71, + 351, + 79 + ], + "spans": [ + { + "bbox": [ + 315, + 71, + 351, + 79 + ], + "type": "text", + "content": "Zoomed out" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 504, + 71, + 538, + 78 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 504, + 71, + 538, + 78 + ], + "spans": [ + { + "bbox": [ + 504, + 71, + 538, + 78 + ], + "type": "text", + "content": "Zoomed in" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 309, + 83, + 543, + 201 + ], + "blocks": [ + { + "bbox": [ + 309, + 83, + 543, + 201 + ], + "lines": [ + { + "bbox": [ + 309, + 83, + 543, + 201 + ], + "spans": [ + { + "bbox": [ + 309, + 83, + 543, + 201 + ], + "type": "image", + "image_path": "23db433e18c53431031553f13b04936af6831f6286e0a5f89ca88b180af0c341.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 211, + 545, + 255 + ], + "lines": [ + { + "bbox": [ + 305, + 211, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 305, + 211, + 545, + 255 + ], + "type": "text", + "content": "Figure 8. Generated zoom sequences with independent sampling (top) and our multi-scale sampling (bottom). Our method encourages different levels to depict a consistent underlying scene, while not compromising the image quality." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 268, + 545, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 268, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 305, + 268, + 545, + 327 + ], + "type": "text", + "content": "Next, we compare our method to two autogressive generation approaches for generating zoom sequences: (1) Stable Diffusion's [1] outpainting model and (2) Stable Diffusion's \"upscale\" super-resolution model. We show representative qualitative results in Fig. 9." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 331, + 546, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 331, + 546, + 462 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 546, + 462 + ], + "type": "text", + "content": "Comparison to progressive outpainting. The outpainting baseline starts with generating the most zoomed-in image and progressively generates coarser scales by downsampling the previous generated image and outpainting the surrounding area. As in our method, the inpainting of each level is conditioned on the corresponding text prompt. In Fig. 9, we show that because of the causality of the autoregressive process, the outpainting approach suffers from gradually accumulating errors, i.e., when a mistake is made at a given step, later outpainting iterations may struggle to produce a consistent image." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 466, + 545, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 466, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 304, + 466, + 545, + 574 + ], + "type": "text", + "content": "Comparison to progressive super-resolution. The superresolution baseline starts with the most zoomed-out image and generates subsequent scales by super-resolving the upscaled central image region, conditioned on the corresponding text prompt. The low resolution input provides strong structural information which constrains the layout of the next zoomed-in image. As we can see in Fig. 9, this super-resolution baseline is not able to synthesize new objects that would only appear in the finer, zoomed-in scales." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 583, + 373, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 583, + 373, + 594 + ], + "spans": [ + { + "bbox": [ + 306, + 583, + 373, + 594 + ], + "type": "text", + "content": "5.3. Ablations" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 602, + 545, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 602, + 545, + 626 + ], + "spans": [ + { + "bbox": [ + 305, + 602, + 545, + 626 + ], + "type": "text", + "content": "In Fig. 10, we show comparisons to simpler versions of our method to examine the effect of our design decisions." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Joint vs. Iterative update. Instead of performing multiscale blending approach, we can instead iteratively cycle through the images in the zoom stack, and perform one sampling step at each level independently. Unlike fully independent sampling, this process does allow for sharing of information between scales, since the steps are still applied to renders from the zoom stack. We find that although this pro" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7179" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 69, + 541, + 328 + ], + "blocks": [ + { + "bbox": [ + 59, + 69, + 541, + 328 + ], + "lines": [ + { + "bbox": [ + 59, + 69, + 541, + 328 + ], + "spans": [ + { + "bbox": [ + 59, + 69, + 541, + 328 + ], + "type": "image", + "image_path": "a15c9ebdc8fd47af793464f0374eeea9f3f39735fa2304f7eb4732a5d2503ce9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 128, + 337, + 464, + 348 + ], + "lines": [ + { + "bbox": [ + 128, + 337, + 464, + 348 + ], + "spans": [ + { + "bbox": [ + 128, + 337, + 464, + 348 + ], + "type": "text", + "content": "Figure 9. Comparisons with Stable Diffusion Outpainting and super-resolution (SR) models." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 53, + 366, + 285, + 591 + ], + "blocks": [ + { + "bbox": [ + 53, + 366, + 285, + 591 + ], + "lines": [ + { + "bbox": [ + 53, + 366, + 285, + 591 + ], + "spans": [ + { + "bbox": [ + 53, + 366, + 285, + 591 + ], + "type": "image", + "image_path": "7c9b052a7db879198c21b45b06c744b667218175e889b665f3af0ddd9ddddff0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 601, + 287, + 634 + ], + "lines": [ + { + "bbox": [ + 46, + 601, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 46, + 601, + 287, + 634 + ], + "type": "text", + "content": "Figure 10. Ablations. We evaluate other options for multi-scale consistency: (1) iteratively updating each level separately, (2) naive multi-scale blending, (3) removing the shared noise." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 649, + 286, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 649, + 286, + 673 + ], + "spans": [ + { + "bbox": [ + 46, + 649, + 286, + 673 + ], + "type": "text", + "content": "duces more consistent results than independent sampling, there remain inconsistencies at stack layer boundaries." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Shared vs. random noise Instead of using a shared noise " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\Pi_{\\mathrm{noise}}" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": ", noise can be sampled independently for each zoom level. We find that this leads to blur in the output samples." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 369, + 545, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 369, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 305, + 369, + 545, + 417 + ], + "type": "text", + "content": "Comparison with naïve blending. Instead of our multiscale blending, we can instead naively blend the observations together, e.g., as in MultiDiffusion [2]. We find that this leads to blurry outputs at deeper zoom levels." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 434, + 451, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 434, + 451, + 445 + ], + "spans": [ + { + "bbox": [ + 306, + 434, + 451, + 445 + ], + "type": "text", + "content": "6. Discussion & Limitations" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 456, + 545, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 456, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 304, + 456, + 545, + 563 + ], + "type": "text", + "content": "A significant challenge in our work is discovering the appropriate set of text prompts that (1) agree with each other across a set of fixed scales, and (2) can be effectively generated consistently by a given text-to-image model. One possible avenue of improvement could be to, along with sampling, optimize for suitable geometric transformations between successive zoom levels. These transformations could include translation, rotation, and even scale, to find better alignment between the zoom levels and the prompts." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 565, + 545, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 565, + 545, + 636 + ], + "spans": [ + { + "bbox": [ + 304, + 565, + 545, + 636 + ], + "type": "text", + "content": "Alternatively, one can optimize the text embeddings, to find better descriptions that correspond to subsequent zoom levels. Or, instead, use the LLM for in-the-loop generation, i.e., by giving LLM the generated image content, and asking it to refine its prompts to produce images which are closer in correspondence given the set of pre-defined scales." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgements. We thank Ben Poole, Jon Barron, Luyang Zhu, Ruiqi Gao, Tong He, Grace Luo, Angjoo Kanazawa, Vickie Ye, Songwei Ge, Keunhong Park, and David Salesin for helpful discussions and feedback. This work was supported in part by UW Reality Lab, Meta, Google, OPPO, and Amazon." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7180" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Stability AI. Stable-diffusion-2-inpainting. https://huggingface.co/stabilityai/stable-diffusion-2-inpainting.2,7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 287, + 157 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 287, + 157 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 287, + 157 + ], + "type": "text", + "content": "[2] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. arXiv preprint arXiv:2302.08113, 2023. 2, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 158, + 288, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 288, + 212 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 288, + 212 + ], + "type": "text", + "content": "[3] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 213, + 288, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 213, + 288, + 246 + ], + "spans": [ + { + "bbox": [ + 53, + 213, + 288, + 246 + ], + "type": "text", + "content": "[4] Peter J Burt and Edward H Adelson. The laplacian pyramid as a compact image code. In Readings in computer vision, pages 671-679. Elsevier, 1987. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 247, + 288, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 288, + 280 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 288, + 280 + ], + "type": "text", + "content": "[5] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 281, + 287, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 281, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 281, + 287, + 323 + ], + "type": "text", + "content": "[6] Dave Epstein, Allan Jabri, Ben Poole, Alexei A Efros, and Aleksander Holynski. Diffusion self-guidance for controllable image generation. arXiv preprint arXiv:2306.00986, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 324, + 288, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 324, + 288, + 368 + ], + "spans": [ + { + "bbox": [ + 53, + 324, + 288, + 368 + ], + "type": "text", + "content": "[7] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 369, + 288, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 288, + 402 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 288, + 402 + ], + "type": "text", + "content": "[8] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 403, + 288, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 288, + 435 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 288, + 435 + ], + "type": "text", + "content": "[9] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 436, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 436, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 436, + 287, + 480 + ], + "type": "text", + "content": "[10] Yuseung Lee, Kunho Kim, Hyunjin Kim, and Minhyuk Sung. Syncdiffusion: Coherent montage via synchronized joint diffusions. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 481, + 287, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 287, + 535 + ], + "type": "text", + "content": "[11] Zhengqi Li, Qianqian Wang, Noah Snavely, and Angjoo Kanazawa. Infinitenature-zero: Learning perpetual view generation of natural scenes from single images. In European Conference on Computer Vision, pages 515-534. Springer, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 536, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 601 + ], + "type": "text", + "content": "[12] Andrew Liu, Richard Tucker, Varun Jampani, Ameesh Makadia, Noah Snavely, and Angjoo Kanazawa. Infinite nature: Perpetual view generation of natural scenes from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14458-14467, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 602, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 635 + ], + "type": "text", + "content": "[13] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. arXiv preprint arXiv:2202.09778, 2022. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 635, + 287, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 658 + ], + "type": "text", + "content": "[14] OpenAI. Chatgpt [large language model]. https://chat.openai.com/chat.7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[15] Ryan Po, Wang Yifan, Vladislav Golyanik, Kfir Aberman, Jonathan T Barron, Amit H Bermano, Eric Ryan Chan, Tali Dekel, Aleksander Holynski, Angjoo Kanazawa, et al. State of the art on diffusion models for visual computing. arXiv preprint arXiv:2310.07204, 2023. 1" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[16] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 118, + 547, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 173 + ], + "type": "text", + "content": "[17] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 175, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 239 + ], + "type": "text", + "content": "[18] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 241, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 296 + ], + "type": "text", + "content": "[19] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models. arXiv preprint arXiv:2307.06949, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 297, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 545, + 352 + ], + "type": "text", + "content": "[20] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 354, + 545, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 419 + ], + "type": "text", + "content": "[21] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 3, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 421, + 545, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 421, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 307, + 421, + 545, + 475 + ], + "type": "text", + "content": "[22] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image super-resolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4713-4726, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 476, + 545, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 476, + 545, + 531 + ], + "spans": [ + { + "bbox": [ + 307, + 476, + 545, + 531 + ], + "type": "text", + "content": "[23] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 533, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 533, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 307, + 533, + 545, + 566 + ], + "type": "text", + "content": "[24] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 567, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 545, + 601 + ], + "type": "text", + "content": "[25] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 601, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 646 + ], + "type": "text", + "content": "[26] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "type": "text", + "content": "[27] Luming Tang, Nataniel Ruiz, Chu Qinghao, Yuanzhen Li, Aleksander Holynski, David E Jacobs, Bharath Hariharan, Yael Pritch, Neal Wadhwa, Kfir Aberman, and Michael Rubinstein. Realfill: Reference-driven generation for authentic image completion. arXiv preprint arXiv:2309.16668, 2023. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7181" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 206 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[28] Shitao Tang, Fuyang Zhang, Jiacheng Chen, Peng Wang, and Yasutaka Furukawa. Mvdiffusion: Enabling holistic multiview image generation with correspondence-aware diffusion. arXiv preprint arXiv:2307.01097, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "type": "text", + "content": "[29] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 206 + ], + "type": "text", + "content": "[30] Qinsheng Zhang, Jiaming Song, Xun Huang, Yongxin Chen, and Ming-Yu Liu. Diffcollage: Parallel generation of large content with diffusion models. arXiv preprint arXiv:2303.17076, 2023. 2" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7182" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Generative Proxemics_ A Prior for 3D Social Interaction from Images/b086fde4-b439-4f25-92c2-a2dc5214a0a3_content_list.json b/2024/Generative Proxemics_ A Prior for 3D Social Interaction from Images/b086fde4-b439-4f25-92c2-a2dc5214a0a3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d19b45ee6077ac3474df27194b8d87dffaeb01b8 --- /dev/null +++ b/2024/Generative Proxemics_ A Prior for 3D Social Interaction from Images/b086fde4-b439-4f25-92c2-a2dc5214a0a3_content_list.json @@ -0,0 +1,1449 @@ +[ + { + "type": "text", + "text": "Generative Proxemics: A Prior for 3D Social Interaction from Images", + "text_level": 1, + "bbox": [ + 133, + 130, + 834, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lea Müller1,4 Vickie Ye1 Georgios Pavlakos2,5 Michael Black3 Angjoo Kanazawa1 \n1UC Berkeley UT Austin MPI for Intelligent Systems, Tübingen", + "bbox": [ + 119, + 179, + 846, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{mueller,vye,kanazawa}@berkeley.edu,pavlakos@cs.utexas.edu,black@tuebingen.mpg.de", + "bbox": [ + 119, + 219, + 849, + 233 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c0d00b3a2ccfd2dd657e0577d5fd876ab6aded41daf4aa370a7375e745a53a7a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 84, + 268, + 300, + 381 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7551801e4d9bec1616e7c25563a5b0c87c38f8d7b7fcd96128c82486f122ef2d.jpg", + "image_caption": [ + "Figure 1. Generative Proxemics. We propose a diffusion model that learns a 3D generative model of two people in close social interaction. We show how the model can be used to generate samples or as a social prior in the downstream task of reconstructing two people in close proximity from images without any user annotation at test time. Shown here are input test images (left) and our predicted 3D bodies (right)." + ], + "image_footnote": [], + "bbox": [ + 84, + 383, + 297, + 488 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2b2f6d447af0d544d5155398d0267e5d85cbf06841bda94c9e767aca4ac89f82.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 305, + 268, + 517, + 381 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6acf03a0036e78e42b534f2144268dd15cda7fdc198e68edcef69dbc5886257f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 305, + 383, + 517, + 488 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ecf016e0e6c59b999bc208db9869de9c6e96af850bb0e44addfc1bf809af0425.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 552, + 268, + 864, + 489 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 554, + 313, + 569 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Social interaction is a fundamental aspect of human behavior and communication. The way individuals position themselves in relation to others, also known as proxemics, conveys social cues and affects the dynamics of social interaction. Reconstructing such interaction from images presents challenges because of mutual occlusion and the limited availability of large training datasets. To address this, we present a novel approach that learns a prior over the 3D proxemics two people in close social interaction and demonstrate its use for single-view 3D reconstruction. We start by creating 3D training data of interacting people using image datasets with contact annotations. We then model the proxemics using a novel denoising diffusion model called BUDDI that learns the joint distribution over the poses of two people in close social interaction. Sampling from our generative proxemics model produces realistic 3D human interactions, which we validate through a perceptual study. We use BUDDI in reconstructing two people in close proximity from an image without any contact annotation via an optimization approach that uses the diffusion model as a prior. Our approach recovers accurate 3D so", + "bbox": [ + 75, + 583, + 472, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cial interactions from noisy initial estimates, outperforming state-of-the-art methods. Our code, data, and model are available at: muelea.github.io/buddy.", + "bbox": [ + 500, + 556, + 890, + 599 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 628, + 630, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Humans are social creatures, and physical interaction plays a crucial role in our daily lives, shaping our relationships. For example, research in behavioral science has shown that a slight touch between two people can cause a more friendly behaviour towards the touch-giver and lead to increased tips in restaurants [5]. Capturing and modeling scenarios of physical social interaction through computer vision will enable advancements in augmented and virtual reality and impact other fields like robotics and behavioral science. The recent progress in single-person mesh estimation is unfortunately not sufficient to model close interaction, as these methods are not trained to reason about the relative depth of people and the intricate interplay between two people's body poses, shapes, and proximity.", + "bbox": [ + 496, + 654, + 890, + 866 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "4Portions of this work were done while LM was at MPI-IS, Tübingen.", + "bbox": [ + 514, + 875, + 887, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "5This work was done while GP was at UC Berkeley.", + "bbox": [ + 517, + 887, + 795, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9687", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we present the first approach that learns a generative model for 3D social proxemics and demonstrate its use as data-driven prior during an optimization routine. The diffusion model is trained using 3D human poses and shapes reconstructed from a large-scale image collection [8] using contact annotation, as well as using motion-capture (MoCap) data [8, 54]. The resulting model is able to generate the 3D pose and shape parameters of pairs of interacting people. When trained on bodies recovered from images, the model learns interactions depicted in photographs, such as people standing close together, playing sports, hugging, etc., see Figure 1. We further demonstrate the effectiveness of the learned prior by applying it to the challenging task of 3D human pose and shape reconstruction from a single image containing people engaged in social interaction.", + "bbox": [ + 80, + 90, + 467, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, we propose BUDDI: a \"BUDdies Diffusion Model\". Diffusion models are established methods for image generation and are often used to model 3D human motion. In this work we use them to model 3D social proxemics. The majority of state-of-the-art diffusion-based methods for 3D human mesh generation operate on 3D joint locations [45]. This representation lacks information about the human body surface, which, intuitively, is important for reasoning about interpersonal contact. Our approach, in contrast, operates on the parameters of two parametric human body models, which represent the surfaces of two people closely interacting. After training, our model is able to generate samples of plausible pairs of 3D bodies in social interaction from pure noise. The model can also be conditioned in the output of a human pose and shape regressor. In this conditional case, the model takes the noisy output and generates similar poses but with realistic social interaction.", + "bbox": [ + 80, + 320, + 467, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We then demonstrate how exploit BUDDI's knowledge of human proxemics to guide 3D mesh reconstruction of people in a close social interaction from a single image. To this end, we introduce a novel optimization-based approach, which uses BUDDI as a data-driven prior. We initialize our optimization routine with samples from BUDDI, conditioned to the output of a state-of-the-art multi-person human mesh regressor [42]. We then optimize over SMPL-X pose, shape, and translation parameters to match detected 2D joint locations. We incorporate guidance from the diffusion model using a loss inspired by the Score-Distillation loss from the 3D object creation literature [35]: In each optimization step, BUDDI refines the current estimate towards a more plausible social interaction conditioned on the initial predictions. The refined pose, shape and translation serve as prior in the overall objective function.", + "bbox": [ + 80, + 580, + 467, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions include (1) presenting the first generative model of a pair of 3D people in close social interaction and (2) a novel approach for reconstructing 3D human meshes from images without relying on ground-truth contact annotations. We perform extensive experiments with", + "bbox": [ + 80, + 825, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "BUDDI to evaluate its performance on the FlickrCI3D Signatures dataset [8] as well as CHI3D, and the recent Hi4D dataset [54] and find that it outperforms the state of the art as well as strong baselines. We also evaluate the unconditional samples from the diffusion model in a perceptual study, where people find our samples more realistic $44.4\\%$ when compared over real samples, where $50\\%$ is the upperbound where they do not see any difference. Importantly, we find that our optimization approach significantly improves the results of [42] both quantitatively and qualitatively. This work opens up a new avenue of research on digital human synthesis, laying the foundation for a deeper understanding of human social behavior derived from image data. Our data, code, and model will be available for research.", + "bbox": [ + 503, + 90, + 888, + 315 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 503, + 332, + 638, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generating 3D humans. There has been recent interest in generating 3D humans, in different contexts. Several methods automatically populate static 3D scenes with 3D humans [14, 60, 61], while more recent methods generate both body and hand poses to interact with 3D objects [43, 44, 49]. Other work generates human motions conditioned on different inputs such as audio [26, 47] or text [33, 34, 45]. Concurrent work proposes text-to-3D diffusion-based approaches to generate motion of two interacting humans [27, 40]. Neither method predicts the full body surface, but rather they synthesize either 3D joint locations or SMPL pose parameters for the average body. These methods are not used as priors for reconstructing interacting people from images.", + "bbox": [ + 503, + 356, + 888, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To model 3D human proxemics probabilistically, we employ diffusion models, which achieve impressive performance on image generation tasks [7, 15, 38, 39]. They have recently been adopted in 3D human motion generation scenarios: MDM [45] generates plausible motions conditioned on text input; PhysDiff [55] incorporates physical constraints in the diffusion process to generate physically plausible motions; and EDGE [47] uses a transformer-based diffusion model for dance generation. Related work [4, 6, 28] has investigated different modalities for the conditioning, e.g., audio, text, or action classes. EgoEgo [25] generates plausible full-body motions conditioned on the head motion. SceneDiffuser [16] focuses on the scene-conditioned setting. We also rely on techniques from the diffusion literature, but consider the unique setting where two people are in close interaction and leverage this for single-image 3D reconstruction.", + "bbox": [ + 503, + 568, + 888, + 823 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-person 3D human mesh estimation. An extensive line of work focuses on reconstructing the 3D human pose and shape of a single person from images using optimization [2, 11, 24, 32, 37, 46, 50] or regression approaches [1, 12, 19, 20, 22, 29, 31, 51, 58, 59]. Capitalizing", + "bbox": [ + 503, + 825, + 888, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9688", + "bbox": [ + 483, + 946, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/309b64c2bc74f21d55748b73879f9648e1a0223602603cf9c9bac43a9ce72380.jpg", + "image_caption": [ + "Figure 2. BUDDI: BUddies Diffusion model. On the left, we illustrate the architecture of BUDDI, our diffusion model for modeling 3D social proxemics between two people in close interaction. The diffusion process is applied directly on SMPL-X body parameters. To condition BUDDI on estimated body model parameters, $c_{\\mathrm{H}}$ , we concatenate the parameters along the token dimension. On the right, we illustrate the optimization method with BUDDI as prior. Our optimization takes detected keypoints [3, 52] and an initial regressor estimate [42] as input. Given the regressor estimate, we sample from BUDDI to obtain $\\tilde{x}$ which we use to initialize the optimization routine. In each optimization iteration, we take a single diffuse-denoise step on the current estimate using the learned denoiser model $D$ . Our losses encourage the current estimate to be close to the refined meshes $(L_{\\mathrm{diffusion}})$ and to the initial estimate and detected keypoints $(L_{\\mathrm{fitting}})$ ." + ], + "image_footnote": [], + "bbox": [ + 78, + 87, + 424, + 261 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f3e68861a4b5e3b31c4203da0d49cd48ce47bffb1125ffb6f0dce333107e371e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 428, + 87, + 890, + 262 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "on these techniques, recent approaches focus explicitly on reconstructing multiple people jointly from a single image. Zanfir et al. [56] propose an optimization solution, while Jiang et al. [18] and Sun et al. [41] rely on deep networks to regress the pose and shape for all people in the image. BEV [42] extends ROMP [41] to reason about the depth of people while taking age/height into account.", + "bbox": [ + 75, + 398, + 468, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The above methods do not address contact between people. To do so, Fieraru et al. [8] introduce the first datasets with ground-truth labels for the body regions in contact between humans. Labels are collected using MoCap (CHI3D) or human annotators (FlickrCI3D Signatures). They propose an optimization approach that requires the ground-truth contact map to reconstruct people in close proximity at test time. More recently, REMIPS [9] introduces a transformer-based method that regresses the 3D pose of multiple people. REMIPS is trained using the above datasets while taking into account contact and interpenetration. In this work, we take a very different approach by learning and exploiting a 3D generative proxemics prior. We use the ground-truth contact maps to generate pseudoground truth 3D human fits from which we learn the diffusion model; once this is learned, we show that it can be used as a prior to recover plausible bodies in close proximity from images without explicit knowledge of contact maps.", + "bbox": [ + 75, + 506, + 468, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Data-driven priors in optimization. Optimization-based methods for 3D human pose and shape estimation, like SMPLify [2], are versatile and allow different data-driven prior terms to be incorporated in the objective function. Different methods have been used to learn pose priors including GMMs [2], VAEs [32], neural distance fields [46], and normalizing flows [57]. ProHMR [23] learns a pose prior conditioned on image pixels. HuMoR [37] incor", + "bbox": [ + 75, + 780, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "porates a data-driven motion prior in the iterative optimization. POSA [14] learns a prior for human-scene interaction from PROX data [13] and uses it in their optimization. In contrast to these methods, we use a diffusion model to capture the joint distribution over SMPL-X parameters for two people interacting and show that we can both sample from the model and use during optimization to improve the pose estimates of interacting people.", + "bbox": [ + 496, + 398, + 890, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 547, + 589, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We introduce BUDDI, a generative model of two people in close social interaction. Because of the complexity and multimodality of the data, we turn to denoising diffusion probabilistic models [15] to address this task. In Sec. 3.1, we describe the basics of diffusion, and the parameterization we employ to model people in contact. In addition to sampling new body meshes from our model, our generative model can serve as a prior for reconstructing 3D humans from images. In Sec. 3.2, we describe an optimization procedure that incorporates BUDDI as a prior to recover two SMPL-X meshes from observed 2D keypoints.", + "bbox": [ + 496, + 578, + 890, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For all of the following, we use the SMPL-X [32] body model to represent the human bodies. SMPL-X is a differentiable function that maps pose, $\\pmb{\\theta} \\in \\mathbb{R}^{21 \\times 3}$ , shape, $\\beta \\in \\mathbb{R}^{10}$ , and expression, $\\psi \\in \\mathbb{R}^{10}$ parameters to a mesh consisting of $N_v = 10,475$ vertices $V \\in \\mathbb{R}^{N_v \\times 3}$ . We place the generated meshes in the world by rotating and translating them by $\\phi \\in \\mathbb{R}^3$ and $\\gamma \\in \\mathbb{R}^3$ . We denote person $a$ 's parameters as $X^a = [\\phi^a, \\pmb{\\theta}^a, \\pmb{\\beta}^a, \\gamma^a]$ and $X^b = [\\phi^b, \\pmb{\\theta}^b, \\pmb{\\beta}^b, \\gamma^b]$ . For simplicity, we refer to both people when no index is specified, e.g., $X$ refers $X^a$ and $X^b$", + "bbox": [ + 496, + 750, + 892, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9689", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Diffusion Model for 3D Proxemics", + "text_level": 1, + "bbox": [ + 76, + 90, + 375, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Denoising diffusion models are latent variable generative models that learn to transform random noise into the desired data distribution $p_{\\mathrm{data}}$ through a forward and reverse process. The forward inference process is a Markov chain over $T$ steps given by transitions $q(\\mathbf{x}_{t + 1}|\\mathbf{x}_t)$ , which gradually adds Gaussian noise to clean samples $\\mathbf{x}_0$ from the data distribution according to a fixed variance schedule $\\sigma_t$ .", + "bbox": [ + 75, + 112, + 468, + 219 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The reverse process $q(\\mathbf{x}_{t - 1}|\\mathbf{x}_t)$ then gradually denoises noisy samples back into the data distribution. The reverse process transitions follow a Gaussian distribution when conditioned on $x_0$ , but must be inferred during the generative process. Following Ramesh et al. [36], we train a neural network $D$ that predicts a sample $\\hat{\\mathbf{x}}_0 = D(\\mathbf{x}_t; t)$ from a noisy sample $\\mathbf{x}_t$ given the noise level $t$ . For the task of reconstructing humans from images, when we have initial estimates of the SMPL-X parameters of two humans, we condition the denoising network $D$ on $\\mathbf{c}_H$ , the predicted SMPL-X parameters of the two humans by a regressor.", + "bbox": [ + 75, + 219, + 468, + 385 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We refer to the process of adding noise as diffusion and the process of removing the noise via $D$ as denoising. Specifically, we diffuse a ground-truth sample $\\mathbf{x}_0$ by uniformly sampling a noise level $t$ with $\\epsilon_t \\sim \\mathcal{N}(0, \\mathbf{I})$ to obtain the noisy sample $\\mathbf{x}_t = \\sqrt{\\sigma_t'}\\mathbf{x}_0 + \\sqrt{1 - \\sigma_t'}\\epsilon_t$ with $\\sigma_t' = \\prod_{i=1}^{t} (1 - \\sigma_t)$ .", + "bbox": [ + 76, + 385, + 468, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We then train $D$ to minimize", + "bbox": [ + 96, + 479, + 290, + 492 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {\\mathbf {x} _ {0} \\sim p _ {\\text {d a t a}}} \\mathbb {E} _ {t \\sim \\mathcal {U} \\{0, T \\}, \\mathbf {x} _ {t} \\sim q (\\cdot | \\mathbf {x} _ {0})} | | D (\\mathbf {x} _ {t}; t, \\mathbf {c} _ {H}) - \\mathbf {x} _ {0} | |, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 503, + 468, + 520 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where we set $\\mathbf{c}_H = \\varnothing$ for $20\\%$ of conditional model training, and all of unconditional model training.", + "bbox": [ + 76, + 529, + 468, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Architecture. Because we aim to model close contact between people, we choose a model state space that can express the full surface of the human body. Specifically, in contrast to prior work in human motion diffusion that operate only on joint angles and locations [45, 55], we directly operate on the full SMPL-X parameters of the two people. A sample $\\mathbf{x}$ thus corresponds to the concatenation of two bodies:", + "bbox": [ + 75, + 577, + 468, + 696 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} = \\left[ X ^ {a}, X ^ {b} \\right] = \\left[ \\phi^ {a}, \\boldsymbol {\\theta} ^ {a}, \\boldsymbol {\\beta} ^ {a}, \\gamma^ {a}, \\phi^ {b}, \\boldsymbol {\\theta} ^ {b}, \\boldsymbol {\\beta} ^ {b}, \\gamma^ {b} \\right].\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 705, + 434, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We denoise a sample $\\mathbf{x}_t$ with a transformer encoder block on tokenized parameters. Specifically, each parameter of each person is tokenized into 152-dimensional latent vectors with per-parameter and per-person embedding layers. We tokenize the noise level $t$ similarly with a noise embedding. When conditioning is available, i.e. SMPL-X estimates for person $a$ and $b$ , we similarly tokenize the parameters to be used as additional tokens. We pass the available tokens into the transformer encoder, and similarly decode the output tokens with per-token embeddings. We illustrate the denoiser architecture in Figure 2.", + "bbox": [ + 75, + 734, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Losses. We employ standard human pose and shape regularization losses. We write our training objective as", + "bbox": [ + 500, + 90, + 890, + 121 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {D} = L _ {\\theta} + L _ {\\beta} + L _ {\\gamma} + L _ {v 2 v}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 132, + 890, + 148 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $L_{\\theta}, L_{\\beta}, L_{\\gamma}$ denote squared L2-losses on respective body model parameters, and $L_{v2v}$ denotes a squared L2 loss on model vertices. We use 6D rotation representations [62] for global orientation and pose, and model the relative translation between $a$ and $b$ . We show generated samples from our unconditional model in Fig. 3.", + "bbox": [ + 496, + 159, + 890, + 250 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Optimization with the Proxemics Prior", + "text_level": 1, + "bbox": [ + 498, + 258, + 833, + 273 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reconstructing 3D human meshes from a single image is an extremely under-constrained problem, and priors over human pose and shape are crucial in an optimization based framework for recovering plausible meshes [2, 32, 56]. Our problem involves people in close contact, which requires correctly placing the meshes in context with each other, which has only been done when given ground truth contact annotations at test time [8]. We remove the need for ground-truth contact maps by using our generative model as a prior during reconstruction with a score distillation approach [35, 48].", + "bbox": [ + 496, + 282, + 890, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During inference, we observe detected 2D keypoints $\\tilde{J}_{2D}$ and initial body model parameter estimates $\\mathbf{c}_H$ from a regressor [42]. We then optimize the body parameters of two people to minimize", + "bbox": [ + 496, + 446, + 890, + 508 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {O p t i z i m a t i o n w . B U D D I}} = L _ {\\text {f i t t i n g}} + L _ {\\text {d i f f u s i o n}}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 520, + 890, + 536 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$L_{\\mathrm{fitting}}$ ensures that the solution stays close to the image evidence, while $L_{\\mathrm{diffusion}}$ is a data-driven prior using our conditional diffusion model. We treat this prior as similar to those used for 3D pose in previous works such as GMM [2] and V-Poser [32], but for 3D proxemics. We illustrate the optimization procedure in Fig. 2 right.", + "bbox": [ + 496, + 546, + 890, + 637 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We initialize our optimization by generating a sample $\\tilde{\\mathbf{x}}$ from the conditional model. We sample with DDIM sampling with 100 evenly spaced steps. We then use the data fitting loss:", + "bbox": [ + 496, + 637, + 890, + 696 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {f i t t i n g}} = \\lambda_ {J} L _ {J} + \\lambda_ {\\tilde {\\theta}} L _ {\\tilde {\\theta}} + \\lambda_ {P} L _ {P}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 710, + 890, + 727 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $L_{J}$ denotes 2D re-projection error between the reprojected 3D joints of the current estimate and the detected 2D keypoints, and $L_{\\tilde{\\theta}}$ is a prior for the solution to be close to the denoised initialization.", + "bbox": [ + 496, + 739, + 890, + 797 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$L_{P}$ denotes an interpenetration loss between two people that pushes inside vertices to the surface, which we use winding numbers [17] to find intersecting vertices between low-resolution SMPL-X meshes of the current estimates:", + "bbox": [ + 496, + 799, + 890, + 859 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {P} = \\sum_ {v \\in V _ {I} ^ {a}} \\min _ {u \\in V ^ {b}} \\| v - u \\| ^ {2} + \\sum_ {v \\in V _ {I} ^ {b}} \\min _ {u \\in V ^ {a}} \\| v - u \\| ^ {2}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 867, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9690", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/67b8969f2bc4b92fc703e2a3e2e368be28f9a12b7b2b08cc057d7f18397191d6.jpg", + "image_caption": [ + "Figure 3. Generative Proxemics: Samples from BUDDI. All samples are unconditionally generated from pure noise using the trained diffusion model. We select several representative examples and show two views per sample. These samples reveal that BUDDI has learned the distribution of people in close contact including embracing each other, playing sports, sitting side by side, and taking photographs." + ], + "image_footnote": [], + "bbox": [ + 125, + 109, + 867, + 327 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $V_{I}^{a}$ denotes vertices of $M^a$ intersecting the lowresolution mesh of $M^b$ ; and vice versa for $V_{I}^{b}$ .", + "bbox": [ + 75, + 406, + 468, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To use the prior on human interaction into account, we use the learned denoising model $D$ from BUDDI and perform a single diffuse-denoise step, with a noise level at $t = 10$ , on the current estimate. The denoised estimate, $\\hat{\\mathbf{x}}_0 = D(\\mathbf{x}_t; t, \\mathbf{c}_H)$ , regularizes the current estimate via", + "bbox": [ + 75, + 436, + 470, + 513 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {d i f f u s i o n}} = \\left\\| D \\left(\\mathbf {x} _ {t}; t, \\mathbf {c} _ {H}\\right) - \\mathbf {x} \\right\\|, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 525, + 468, + 544 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{x}_t = \\sqrt{\\sigma_t^{\\prime}}\\mathbf{x}_{\\mathrm{no - grad}} + \\sqrt{1 - \\sigma_t^{\\prime}}\\epsilon_t$ denotes the diffused body model parameters of the current estimate, and $\\mathbf{x}_{\\mathrm{no - grad}}$ denotes the current estimate with detached gradients. $\\hat{\\mathbf{x}}_0$ , and encourages $\\mathbf{x}$ to be close to $\\hat{\\mathbf{x}}_0$ . In practice, we penalize the decoded parameters of $\\mathbf{x}$ and $\\hat{\\mathbf{x}}_0$ directly as", + "bbox": [ + 75, + 553, + 472, + 632 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {\\text {d i f f u s i o n}} = \\lambda_ {\\hat {\\phi}} | | \\hat {\\phi} _ {0} - \\phi | | + \\lambda_ {\\hat {\\theta}} | | \\hat {\\theta} _ {0} - \\theta | | \\tag {7} \\\\ + \\lambda_ {\\hat {\\boldsymbol {\\beta}}} | | \\hat {\\boldsymbol {\\beta}} _ {0} - \\boldsymbol {\\beta} | | + \\lambda_ {\\hat {\\boldsymbol {\\gamma}}} | | \\hat {\\boldsymbol {\\gamma}} _ {0} - \\boldsymbol {\\gamma} | |. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 642, + 468, + 685 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Intuitively, this loss uses the learned denoiser $D$ to take a step from the current estimate towards the data distribution of two people in close proximity, conditioned on the regressor prediction.", + "bbox": [ + 75, + 695, + 470, + 756 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 768, + 299, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Data. There are few datasets containing 3D ground truth of humans in close social interaction [8, 54]. Such datasets are usually captured in lab environments, consequently they are small and do not contain the variety of interactions between humans \"in the wild,\" e.g. when playing sports or taking social pictures. To address this lack of data, we create Flickr Fits, i.e. SMPL-X fits", + "bbox": [ + 75, + 794, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "for Flickr images portraying humans in contact scenarios. For this, we use FlickrCI3D Signatures [8], a dataset of images showing interacting humans collected from Flickr with discrete human-human contact annotations. Specifically, the SMPL-X body surface is divided into $R = 75$ regions such that each region, $r$ , roughly covers a similar area. For a given photo, the human annotators assign a binary label indicating contact between a region on one person and a region on the other. For two meshes, $M^a$ and $M^b$ , the annotation can be represented as a binary contact map $\\mathcal{C}^D \\in \\{0,1\\}^{R \\times R}$ , where", + "bbox": [ + 496, + 406, + 893, + 574 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} _ {i j} ^ {D} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} r _ {i} \\text {o f} M ^ {a} \\text {i s i n c o n t a c t w i t h} r _ {j} \\text {o f} M ^ {b} \\\\ 0, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 584, + 890, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use these ground-truth contact maps in an optimization routine for fitting two people to detected keypoints, similar to Sec. 3.2 but replace the diffusion model prior with standard image fitting priors. The dataset contains 10,631/1,139 train/test images, with one image containing multiple contact annotations. Note that we use pairs of people in contact as labeled in the FlickrCI3D Signatures dataset [8] and only use images containing matching regressor estimates, 2D keypoints, and contact labels. See Fig. 4 for qualitative examples and the Sup. Mat. for more details about the selection process and an evaluation of the optimization method with ground-truth contact labels.", + "bbox": [ + 496, + 628, + 893, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We also augment our training data with available MoCap data, which is considerably smaller than those obtained from image fits: CHI3D [8] contains 3/2 pairs of training/test subjects performing 127 sequences of two-person interactions like hugs or kicks with ground-truth SMPL-X bodies. One frame per sequence has contact map annota", + "bbox": [ + 496, + 810, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9691", + "bbox": [ + 480, + 944, + 513, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f72b40b5e6a3ff82197ae5ea6209b1143354e3b1afb0147d938623591cec6c62.jpg", + "image_caption": [ + "Figure 4. Flickr Fits. We visualize the output of the optimization process that reconstructs two people in close proximity using ground-truth contact maps, shown from four different views. We use these 3D fits as training data for BUDDI." + ], + "image_footnote": [], + "bbox": [ + 104, + 88, + 467, + 262 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tions. We use only the contact frame of the sequences from two subject pairs, resulting in 247 mesh pairs for training, and the third pair for evaluation. Hi4D [54] contains sequences of 20 pairs of people interacting with each other. The interactions include actions like hugging, dancing, and fighting. We randomly split the data into $14/3/3$ pairs for train/val/test and use every fifth frame of the subsequence involving contact as labeled in Hi4D, resulting in about 1K mesh pairs for training. The body representation format in Hi4D is SMPL, which we transfer to SMPL-X using the SMPL-X code repository [32]. Please see the Sup. Mat. for more details of the datasets. Note that while we use SMPL-X model, BUDDI is not trained on hands because none of these datasets contain hand poses.", + "bbox": [ + 75, + 340, + 468, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "BUDDI Training. BUDDI is trained with meshes from FlickrCI3D Signatures Fits, CHI3D, and Hi4D. We use $60\\%$ Flickr, $20\\%$ CHI3D, and $20\\%$ Hi4D data distribution per batch with batch size 512. The transformer backbone has six layers and eight heads; we use $10\\%$ dropout and randomly shuffle the order of people during training. To train BUDDI, we randomly sample noise levels $t$ up to 1000 using a cosine noise schedule [30]. We use the Adam optimizer [21] with learning rate $10^{-4}$ . We train two separate networks: an unconditional model for generation and the conditional version for reconstruction. For the conditional model, we use all camera views of the MoCap datasets, i.e. 4/8 cameras for CHI3D/Hi4D. The unconditional model is trained on 3D MoCap fits in the world coordinate system. To sample new poses, we use DDIM sampling starting at noise levels $t = 1000$ in steps of 10.", + "bbox": [ + 75, + 551, + 468, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Optimization Details. During optimization, we experiment with different noise levels, between 10 and 100, and find that $t = 10$ does not disturb the inputs too much, but enough for $D$ to generate new configurations. We use detected 2D keypoints from OpenPose [3] and ViTpose [52] and BEV [42] estimates as conditioning. Unlike single-person mesh regressors, BEV is designed to predict multiple", + "bbox": [ + 75, + 795, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "people including their relative depth. Please see Sup. Mat. for more details.", + "bbox": [ + 498, + 90, + 890, + 119 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 133, + 632, + 150 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare our reconstruction method with BEV [42], which is also used as an input to our conditional model. Since there is no other available work that reasons about people in close social interaction, we experiment with simple but effective baselines. We train the transformer model of BUDDI to directly predict SMPL-X parameters of people in contact from BEV input, essentially a deterministic, single-step ablation of BUDDI. We also evaluate the direct conditional denoised output of BEV by BUDDI without any optimization. As another baseline, we propose an optimization routine that replaces $L_{\\text{diffusion}}$ with a simple heuristic that takes the minimal distances between two meshes predicted by BEV and minimizes their distance during optimization along with the other energy terms. Finally, to compare the generation ability we train a VAE which we also use during the optimization routine in a similar manner to VPoser [32] but for two people by optimizing the VAE latent space instead of SMPL-X parameters. We refer to these models as Transformer, BUDDI (gen.), Contact Heuristic, and VAE, respectively. All baselines are trained on the same datasets as BUDDI with the same sampling strategies. Details about our baselines are provided in the Sup. Mat..", + "bbox": [ + 496, + 159, + 890, + 491 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. We use standard evaluation metrics from the human pose and shape estimation literature. We also report the joint PA-MPJPE computed by performing Procrustes alignment of both people together. In addition to per-person metrics, this captures the relative orientation and translation of the two people. Since our method directly estimates 3D humans we propose a new metric similar to PCK [53] from the 2D pose literature called PCC, the percentage of correct contact points with respect to a radius $r$ . Specifically, given two meshes, $M^a / M^b$ and a contact map $\\mathcal{C}^D$ we compute the pairwise vertex-to-vertex Euclidean distances $d_{\\mathrm{eucl}}(\\mathcal{C}^D)$ between annotated contact regions and consider the pair to be correct when $\\min(d_{\\mathrm{eucl}}(\\mathcal{C}^D)) < r$ .", + "bbox": [ + 496, + 491, + 892, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Unconditional Generation", + "text_level": 1, + "bbox": [ + 500, + 696, + 736, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We qualitatively evaluate BUDDI by showing samples generated from the model in Fig. 1 and 3. Our approach is able to generate people in close proximity including embraces, handshakes, having a conversation, sitting side by side, and, in general, plausibly interacting with each other. Since it is trained on Internet image collections, it also learns to generate people posing for photographs or playing sports.", + "bbox": [ + 496, + 719, + 890, + 825 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We further run a perceptual study to evaluate the realism of the generated social interactions against other methods. In a forced choice study, we compare our generated samples with samples from the real data distribution according to the 60/20/20 per-batch ratio for Flickr/CHI3D/Hi4D", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9692", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b6a5bf1fac02cd6bc4f974a89c2775a5b3fdac8ac3486b88bd98b1cc84f37845.jpg", + "image_caption": [ + "Figure 5. Automatic reconstruction of people in close social interaction. We show qualitative results from a) BEV, b) contact heuristics, which takes the BEV output and encourages the closest parts to be in contact, and c) our method, which optimizes the BEV estimates against the image evidence with the BUDDI prior. Our approach recovers a plausible reconstruction with subtle details." + ], + "image_footnote": [], + "bbox": [ + 76, + 88, + 898, + 510 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3ed800cd20658cd4c76b956fb93f83acaabe971b7e5fea8f422689c070d3856b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
JOINT ↓\nPA-MPJPEPCC at radius ↑
510152025
BEV106-----
Transformer861440607382
BUDDI (gen.)921539587180
Heuristic681434496170
VAE1011128425565
BUDDI661944627381
", + "bbox": [ + 138, + 585, + 410, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. 3D Pose Evaluation on FlickrCI3D Signatures. We evaluate methods against the Flickr fits using their joint (two-person) PA-MPJPE expressed in mm. We also evaluate the percentage of correct contact points (PCC) for radius $r$ mm.", + "bbox": [ + 75, + 707, + 470, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "used during training. We also compare BUDDI against generations from the VAE and a non-parametric random baseline that samples meshes from the pseudo-ground truth after centering the two people. We do a forced choice comparison between BUDDI and these three other methods, asking workers on Amazon Mechanical Turk to choose the sample that shows a more realistic close social interaction. We use 256 samples per method. We collect ratings for 768", + "bbox": [ + 75, + 780, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "pairwise comparisons. In this study, BUDDI was chosen over random in $71.23\\%$ of the comparisons, over the VAE in $60.17\\%$ , and over the training data in $44.4\\%$ . Note that $50\\%$ is the upper bound for such forced choice comparisons, in which participants cannot tell the difference between real and generated samples.", + "bbox": [ + 496, + 589, + 893, + 680 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For a quantitative evaluation, we compute the FID score between samples from BUDDI and samples from the $VAE$ on concatenated SMPL-X parameters. We sample 8K examples per method and from our training data following the dataset ratio per batch. BUDDI has a lower FID score (1.6) compared to the $VAE$ (3.3).", + "bbox": [ + 496, + 681, + 893, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Fitting with BUDDI", + "text_level": 1, + "bbox": [ + 500, + 785, + 689, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We show qualitative results in Fig. 5 comparing BUDDI against BEV and the Contact Heuristic. Our approach is able to generate various types of human interactions with plausible contact and depth placement. It is also able to capture close interaction between a child and a parent. Although the Contact Heuristic (center) is able to move two", + "bbox": [ + 496, + 809, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9693", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8530125c9b4849884ec11b92ce8d77a98514eddc418d37d23c0edf4d0c4173ec.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PER PERSON ↓ PA-MPJPEJOINT ↓ PA-MPJPEJOINT PA-MPJPE
backhugbasketballcheersdancefighthighfivehugkissposesidehugtalk
BEV78 / 84136200126109135121106163139142131118
Heuristic67 / 7112116883941319468159159118113109
BUDDI (F, C)70 / 77115200949212810810013311410410791
Transformer79 / 85120161141103138123128117106120105100
BUDDI (gen.)82 / 901171521391201371309610197115102101
VAE80 / 8213817513311414111987176162135140113
BUDDI70 / 769812795921131097210585889681
", + "bbox": [ + 78, + 88, + 903, + 202 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e6af8957d89256d4d240ee5897730bb78f2e76518a5abeec2417623653445f0d.jpg", + "table_caption": [ + "Table 2. Evaluation of BUDDI on Hi4D. We compare the output of BUDDI to the proposed baseline methods on the Hi4D challenge. The first block shows methods that do not use Hi4D data during training or are optimization based without access to priors trained on Hi4D. BUDDI (F,C) in particular, is our model BUDDI trained on Flickr and CHI3D data only. All errors are reported in mm for 3D Joints." + ], + "table_footnote": [], + "table_body": "
PER PERSON ↓\nPA-MPJPEJOINT ↓\nPA-MPJPE
BEV50 5296
Transformer54 56105
BUDDI (gen.)53 5380
Heuristic49 46105
VAE54 54103
BUDDI48 4768
", + "bbox": [ + 138, + 277, + 411, + 392 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Quantitative Evaluation on CHI3D. We compare the output of our model to the baselines on CHI3D (pair s03). All errors reported in mm for 3D Joints.", + "bbox": [ + 75, + 402, + 470, + 445 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "people closer together, which helps with image alignment, upon close observation it is not able to capture the subtle interaction between people that happens during intimate interaction. BUDDI's estimates are more realistic and better capture the subtle details of interaction. We provide additional qualitative examples in the Sup. Mat.", + "bbox": [ + 75, + 457, + 468, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We further report the percentage of correct contact (PCC) with respect to the ground truth contact map on the FlickrCI3D Signatures test set in Table 1. The table also shows the pose reconstruction accuracy against our Flickr Fits. All metrics show improvement over BEV, in particular the joint PA-MPJPE. Non-optimization methods, i.e. Transformer and BUDDI (gen.), are able to predict plausible contacts, with similar PCC accuracy to BUDDI, but struggle to reconstruct the data with a worse joint PA-MPJPE. The Heuristic, in contrast, achieves a lower reconstruction error, but worse PCC. Our approach which leverages the learned prior during optimization can recover both the relative positions and contacts between the two people. To provide insights into the performance of single-person mesh regressors when evaluated on the two-person reconstruction task, we run 4D Humans [10] on Flickr Fits. The joint PA-MPJPE is $344\\mathrm{mm}$ which is high, as expected, since these methods are not trained to reason about proximity.", + "bbox": [ + 73, + 550, + 468, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We further evaluate our model against ground truth MoCap data in Tables 2 and 3. Optimization with BUDDI consistently improves the two-person reconstruction error over BEV and other baselines. When evaluated per action, the strongest improvements over BEV come from complex", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "close social interactions like hugging or kissing, at $58\\mathrm{mm}$ and $54\\mathrm{mm}$ absolute improvement over BUDDI respectively. The Heuristic baseline achieves a low PA-MPJPE reconstruction error on all three datasets in particular for poses with a few physical contact points, such as a handshake, whereas more complex contact, such as a hug, requires data-driven priors like BUDDI; we provide an analysis in the Sup. Mat. to quantify this hypothesis. Transformer and BUDDI (gen.) have lower joint PA-MPJPE errors than BEV and the Heuristic, but worse per-person reconstruction errors. The VAE results suggest that directly operating in the latent space of a generative model is challenging and not sufficient to accurately recover close social interactions. BUDDI, in contrast, is able to model a wide variety of poses, as supported by the numerical results.", + "bbox": [ + 496, + 282, + 893, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 525, + 619, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose BUDDI, a diffusion model for close human-human interaction. We train BUDDI from 3D fits obtained from a large-scale dataset of images with ground truth contact annotations as well as a small set of available mocap data. BUDDI enables unconditional sampling of people in close social interaction. More importantly, we also demonstrate how BUDDI can be used as effective prior for single-view 3D reconstruction of two people in close proximity.", + "bbox": [ + 496, + 551, + 893, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our core contribution of a generative proxemics prior provides the foundation for future work on modeling and capturing human interaction. For example, future work could iteratively apply our method to reconstruct more than two people in close proximity, explore other conditioning modalities like pixel features, text, action labels, or, by taking the outputs of recent single-person mesh regressors into account, address reconstructing fine-grained interaction including finger pose and facial expressions.", + "bbox": [ + 496, + 674, + 893, + 810 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements: We thank A. Holynski, E. Weber, F. Warburg, B. Peebles, J. Rajasegaran, K. Mangalam and N. Athanasiou for insightful discussions, and A. Cseke, T. McConnell and T. Alexiadis for running the user study.", + "bbox": [ + 496, + 811, + 890, + 872 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Disclosure: https://files.is.tue.mpg.de/black/CoI_CVPR_2024.txt", + "bbox": [ + 496, + 872, + 890, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9694", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Anurag Arnab, Carl Doersch, and Andrew Zisserman. Exploiting temporal context for 3D human pose estimation in the wild. In Computer Vision and Pattern Recognition (CVPR), 2019.", + "[2] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J. Black. Keep it SMPL: Automatic estimation of 3D human pose and shape from a single image. In European Conference on Computer Vision (ECCV), pages 561-578. Springer International Publishing, 2016.", + "[3] Zhe Cao, Gines Hidalgo Martinez, Tomas Simon, Shih-En Wei, and Yaser Sheikh. OpenPose: Realtime multi-person 2D pose estimation using part affinity fields. Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 43 (1):172-186, 2019.", + "[4] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18000-18010, 2023.", + "[5] April H Crusco and Christopher G Wetzel. The midas touch: The effects of interpersonal touch on restaurant tipping. *Personality and Social Psychology Bulletin*, 10(4): 512-517, 1984.", + "[6] Rishabh Dabral, Muhammad Hamza Mughal, Vladislav Golyanik, and Christian Theobalt. Mofusion: A framework for denoising-diffusion-based motion synthesis. In Computer Vision and Pattern Recognition (CVPR), 2023.", + "[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat GANs on image synthesis. Conference on Neural Information Processing Systems (NeurIPS), 34:8780-8794, 2021.", + "[8] Mihai Fieraru, Mihai Zanfir, Elisabeta Oneata, Alin-Ionut Popa, Vlad Olaru, and Cristian Sminchisescu. Three-dimensional reconstruction of human interactions. In Computer Vision and Pattern Recognition (CVPR), pages 7214-7223, 2020.", + "[9] Mihai Fieraru, Mihai Zanfir, Teodor Szente, Eduard Bazavan, Vlad Olaru, and Cristian Sminchisescu. REMIPS: Physically consistent 3D reconstruction of multiple interacting people under weak supervision. In Conference on Neural Information Processing Systems (NeurIPS), pages 19385-19397, 2021.", + "[10] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa*, and Jitendra Malik*. Humans in 4D: Reconstructing and tracking humans with transformers. In International Conference on Computer Vision (ICCV), 2023.", + "[11] Peng Guan, Alexander Weiss, Alexandru Balan, and Michael J. Black. Estimating human shape and pose from a single image. In International Conference on Computer Vision (ICCV), pages 1381-1388, 2009.", + "[12] Riza Alp Guler and Iasonas Kokkinos. HoloPose: Holistic 3D human reconstruction in-the-wild. In Computer Vision and Pattern Recognition (CVPR), 2019.", + "[13] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J Black. Resolving 3D human pose ambigui" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ties with 3D scene constraints. In International Conference on Computer Vision (ICCV), pages 2282-2292, 2019.", + "[14] Mohamed Hassan, Partha Ghosh, Joachim Tesch, Dimitrios Tzionas, and Michael J Black. Populating 3D scenes by learning human-scene interaction. In Computer Vision and Pattern Recognition (CVPR), pages 14708-14718, 2021.", + "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Conference on Neural Information Processing Systems (NeurIPS), pages 6840-6851, 2020.", + "[16] Siyuan Huang, Zan Wang, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, and Song-Chun Zhu. Diffusion-based generation, optimization, and planning in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023.", + "[17] Alec Jacobson, Ladislav Kavan, and Olga Sorkine-Hornung. Robust inside-outside segmentation using generalized winding numbers. ACM Transactions on Graphics (TOG), 32(4): 1-12, 2013.", + "[18] Wen Jiang, Nikos Kolotouros, Georgios Pavlakos, Xiaowei Zhou, and Kostas Daniilidis. Coherent reconstruction of multiple humans from a single image. In Computer Vision and Pattern Recognition (CVPR), pages 5578-5587, 2020.", + "[19] Hanbyul Joo, Natalia Neverova, and Andrea Vedaldi. Exemplar fine-tuning for 3D human model fitting towards in-the-wild 3D human pose estimation. In International Conference on 3D Vision (3DV), 2021.", + "[20] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In Computer Vision and Pattern Recognition (CVPR), pages 7122-7131, 2018.", + "[21] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.", + "[22] Nikos Kolotouros, Georgios Pavlakos, Michael J. Black, and Kostas Daniilidis. Learning to reconstruct 3D human pose and shape via model-fitting in the loop. In International Conference on Computer Vision (ICCV), pages 2252-2261, 2019.", + "[23] Nikos Kolotouros, Georgios Pavlakos, Dinesh Jayaraman, and Kostas Daniilidis. Probabilistic modeling for human mesh recovery. In International Conference on Computer Vision (ICCV), 2021.", + "[24] Christoph Lassner, Javier Romero, Martin Kiefel, Federica Bogo, Michael J. Black, and Peter V. Gehler. Unite the people: Closing the loop between 3D and 2D human representations. In Computer Vision and Pattern Recognition (CVPR), pages 6050-6059, 2017.", + "[25] Jiaman Li, Karen Liu, and Jiajun Wu. Ego-body pose estimation via ego-head pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17142-17151, 2023.", + "[26] Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. AI choreographer: Music conditioned 3D dance generation with AIST++. In International Conference on Computer Vision (ICCV), pages 13401-13412, 2021.", + "[27] Han Liang, Wenqian Zhang, Wenxuan Li, Jingyi Yu, and Lan Xu. Intergen: Diffusion-based multi-human motion" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9695", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "generation under complex interactions. arXiv preprint arXiv:2304.05684, 2023.", + "[28] Jianxin Ma, Shuai Bai, and Chang Zhou. Pretrained diffusion models for unified human motion synthesis. arXiv preprint arXiv:2212.02837, 2022.", + "[29] Lea Müller, Ahmed A. A. Osman, Siyu Tang, Chun-Hao P. Huang, and Michael J. Black. On self-contact and human pose. In Computer Vision and Pattern Recognition (CVPR), pages 9990-9999, 2021.", + "[30] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning (ICML), pages 8162-8171, 2021.", + "[31] Mohamed Omran, Christoph Lassner, Gerard Pons-Moll, Peter V. Gehler, and Bernt Schiele. Neural body fitting: Unifying deep learning and model based human pose and shape estimation. In International Conference on 3D Vision (3DV), pages 484-494, 2018.", + "[32] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019.", + "[33] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3D human motion synthesis with transformer VAE. In Computer Vision and Pattern Recognition (CVPR), pages 10985-10995, 2021.", + "[34] Mathis Petrovich, Michael J Black, and Gül Varol. TEMOS: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision (ECCV), pages 480-497, 2022.", + "[35] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv, 2022.", + "[36] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents, 2022.", + "[37] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J Guibas. HuMoR: 3D human motion model for robust pose estimation. In International Conference on Computer Vision (ICCV), 2021.", + "[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022.", + "[39] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022.", + "[40] Yonatan Shafir, Guy Tevet, Roy Kapon, and Amit H Bermano. Human motion diffusion as a generative prior. arXiv preprint arXiv:2303.01418, 2023.", + "[41] Yu Sun, Qian Bao, Wu Liu, Yili Fu, Michael J Black, and Tao Mei. Monocular, one-stage, regression of multiple 3D people. In Computer Vision and Pattern Recognition (CVPR), pages 11179-11188, 2021." + ], + "bbox": [ + 78, + 92, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Yu Sun, Wu Liu, Qian Bao, Yili Fu, Tao Mei, and Michael J Black. Putting people in their place: Monocular regression of 3d people in depth. In Computer Vision and Pattern Recognition (CVPR), pages 13243-13252, 2022.", + "[43] Omid Taheri, Vasileios Choutas, Michael J. Black, and Dimitrios Tzionas. GOAL: Generating 4D whole-body motion for hand-object grasping. In Computer Vision and Pattern Recognition (CVPR), pages 13263-13273, 2022.", + "[44] Purva Tendulkar, Dídac Surís, and Carl Vondrick. FLEX: Full-body grasping without full-body grasps. In Computer Vision and Pattern Recognition (CVPR), 2023.", + "[45] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022.", + "[46] Garvita Tiwari, Dimitrije Antic, Jan Eric Lenssen, Nikolaos Sarafianos, Tony Tung, and Gerard Pons-Moll. Pose-ndf: Modeling human pose manifolds with neural distance fields. In European Conference on Computer Vision (ECCV), 2022.", + "[47] Jonathan Tseng, Rodrigo Castellon, and C Karen Liu. EDGE: Editable dance generation from music. Computer Vision and Pattern Recognition (CVPR), 2023.", + "[48] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12619-12629, 2023.", + "[49] Yan Wu, Jiahao Wang, Yan Zhang, Siwei Zhang, Otmar Hilliges, Fisher Yu, and Siyu Tang. Saga: Stochastic whole-body grasping with contact. In European Conference on Computer Vision (ECCV), pages 257-274, 2022.", + "[50] Hongyi Xu, Eduard Gabriel Bazavan, Andrei Zanfir, William T. Freeman, Rahul Sukthankar, and Cristian Sminchisescu. GHUM & GHUML: Generative 3D human shape and articulated pose models. In Computer Vision and Pattern Recognition (CVPR), pages 6183-6192, 2020.", + "[51] Yuanlu Xu, Song-Chun Zhu, and Tony Tung. DenseRaC: Joint 3D pose and shape estimation by dense render-and compare. In International Conference on Computer Vision (ICCV), 2019.", + "[52] Yufei Xu, Jing Zhang, Qiming Zhang, and Dacheng Tao. Vitpose: Simple vision transformer baselines for human pose estimation. Advances in Neural Information Processing Systems, 35:38571-38584, 2022.", + "[53] Yi Yang and Deva Ramanan. Articulated human detection with flexible mixtures of parts. IEEE transactions on pattern analysis and machine intelligence, 35(12):2878-2890, 2012.", + "[54] Yifei Yin, Chen Guo, Manuel Kaufmann, Juan Zarate, Jie Song, and Otmar Hilliges. Hi4d: 4d instance segmentation of close human interaction. In Computer Vision and Pattern Recognition (CVPR), 2023.", + "[55] Ye Yuan, Jiaming Song, Umar Iqbal, Arash Vahdat, and Jan Kautz. Physdiff: Physics-guided human motion diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16010-16021, 2023.", + "[56] Andrei Zanfir, Elisabeta Maroiniu, and Cristian Sminchisescu. Monocular 3d pose and shape estimation of multiple people in natural scenes-the importance of multiple scene" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9696", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "constraints. In Computer Vision and Pattern Recognition (CVPR), pages 2148-2157, 2018.", + "[57] Andrei Zanfir, Eduard Gabriel Bazavan, Hongyi Xu, William T Freeman, Rahul Sukthankar, and Cristian Sminchisescu. Weakly supervised 3D human pose and shape reconstruction with normalizing flows. In European Conference on Computer Vision (ECCV), pages 465-481, 2020.", + "[58] Andrei Zanfir, Eduard Gabriel Bazavan, Mihai Zanfir, William T Freeman, Rahul Sukthankar, and Cristian Sminchisescu. Neural descent for visual 3D human pose and shape. In Computer Vision and Pattern Recognition (CVPR), pages 14484-14493, 2021.", + "[59] Hongwen Zhang, Yating Tian, Xinchi Zhou, Wanli Ouyang, Yebin Liu, Limin Wang, and Zhenan Sun. PyMAF: 3D human pose and shape regression with pyramidal mesh alignment feedback loop. In International Conference on Computer Vision (ICCV), pages 11446-11456, 2021.", + "[60] Siwei Zhang, Yan Zhang, Qianli Ma, Michael J Black, and Siyu Tang. PLACE: Proximity learning of articulation and contact in 3D environments. In International Conference on 3D Vision (3DV), pages 642-651, 2020.", + "[61] Yan Zhang, Mohamed Hassan, Heiko Neumann, Michael J Black, and Siyu Tang. Generating 3D people in scenes without people. In Computer Vision and Pattern Recognition (CVPR), pages 6194-6204, 2020.", + "[62] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Computer Vision and Pattern Recognition (CVPR), pages 5745-5753, 2019." + ], + "bbox": [ + 78, + 92, + 468, + 501 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "9697", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Generative Proxemics_ A Prior for 3D Social Interaction from Images/b086fde4-b439-4f25-92c2-a2dc5214a0a3_model.json b/2024/Generative Proxemics_ A Prior for 3D Social Interaction from Images/b086fde4-b439-4f25-92c2-a2dc5214a0a3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..52ad88423bb96e798c3b888d9b582679834fc6ff --- /dev/null +++ b/2024/Generative Proxemics_ A Prior for 3D Social Interaction from Images/b086fde4-b439-4f25-92c2-a2dc5214a0a3_model.json @@ -0,0 +1,2125 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.134, + 0.131, + 0.836, + 0.152 + ], + "angle": 0, + "content": "Generative Proxemics: A Prior for 3D Social Interaction from Images" + }, + { + "type": "text", + "bbox": [ + 0.12, + 0.18, + 0.847, + 0.217 + ], + "angle": 0, + "content": "Lea Müller1,4 Vickie Ye1 Georgios Pavlakos2,5 Michael Black3 Angjoo Kanazawa1 \n1UC Berkeley UT Austin MPI for Intelligent Systems, Tübingen" + }, + { + "type": "text", + "bbox": [ + 0.12, + 0.22, + 0.85, + 0.234 + ], + "angle": 0, + "content": "{mueller,vye,kanazawa}@berkeley.edu,pavlakos@cs.utexas.edu,black@tuebingen.mpg.de" + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.27, + 0.301, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.384, + 0.299, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.27, + 0.519, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.384, + 0.519, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.553, + 0.269, + 0.866, + 0.491 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.5, + 0.892, + 0.543 + ], + "angle": 0, + "content": "Figure 1. Generative Proxemics. We propose a diffusion model that learns a 3D generative model of two people in close social interaction. We show how the model can be used to generate samples or as a social prior in the downstream task of reconstructing two people in close proximity from images without any user annotation at test time. Shown here are input test images (left) and our predicted 3D bodies (right)." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.555, + 0.314, + 0.57 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Social interaction is a fundamental aspect of human behavior and communication. The way individuals position themselves in relation to others, also known as proxemics, conveys social cues and affects the dynamics of social interaction. Reconstructing such interaction from images presents challenges because of mutual occlusion and the limited availability of large training datasets. To address this, we present a novel approach that learns a prior over the 3D proxemics two people in close social interaction and demonstrate its use for single-view 3D reconstruction. We start by creating 3D training data of interacting people using image datasets with contact annotations. We then model the proxemics using a novel denoising diffusion model called BUDDI that learns the joint distribution over the poses of two people in close social interaction. Sampling from our generative proxemics model produces realistic 3D human interactions, which we validate through a perceptual study. We use BUDDI in reconstructing two people in close proximity from an image without any contact annotation via an optimization approach that uses the diffusion model as a prior. Our approach recovers accurate 3D so" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.557, + 0.892, + 0.601 + ], + "angle": 0, + "content": "cial interactions from noisy initial estimates, outperforming state-of-the-art methods. Our code, data, and model are available at: muelea.github.io/buddy." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.63, + 0.631, + 0.644 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.655, + 0.892, + 0.867 + ], + "angle": 0, + "content": "Humans are social creatures, and physical interaction plays a crucial role in our daily lives, shaping our relationships. For example, research in behavioral science has shown that a slight touch between two people can cause a more friendly behaviour towards the touch-giver and lead to increased tips in restaurants [5]. Capturing and modeling scenarios of physical social interaction through computer vision will enable advancements in augmented and virtual reality and impact other fields like robotics and behavioral science. The recent progress in single-person mesh estimation is unfortunately not sufficient to model close interaction, as these methods are not trained to reason about the relative depth of people and the intricate interplay between two people's body poses, shapes, and proximity." + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.875, + 0.888, + 0.888 + ], + "angle": 0, + "content": "4Portions of this work were done while LM was at MPI-IS, Tübingen." + }, + { + "type": "page_footnote", + "bbox": [ + 0.519, + 0.888, + 0.796, + 0.9 + ], + "angle": 0, + "content": "5This work was done while GP was at UC Berkeley." + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.875, + 0.888, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9687" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.081, + 0.092, + 0.468, + 0.317 + ], + "angle": 0, + "content": "In this work, we present the first approach that learns a generative model for 3D social proxemics and demonstrate its use as data-driven prior during an optimization routine. The diffusion model is trained using 3D human poses and shapes reconstructed from a large-scale image collection [8] using contact annotation, as well as using motion-capture (MoCap) data [8, 54]. The resulting model is able to generate the 3D pose and shape parameters of pairs of interacting people. When trained on bodies recovered from images, the model learns interactions depicted in photographs, such as people standing close together, playing sports, hugging, etc., see Figure 1. We further demonstrate the effectiveness of the learned prior by applying it to the challenging task of 3D human pose and shape reconstruction from a single image containing people engaged in social interaction." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.321, + 0.468, + 0.577 + ], + "angle": 0, + "content": "Specifically, we propose BUDDI: a \"BUDdies Diffusion Model\". Diffusion models are established methods for image generation and are often used to model 3D human motion. In this work we use them to model 3D social proxemics. The majority of state-of-the-art diffusion-based methods for 3D human mesh generation operate on 3D joint locations [45]. This representation lacks information about the human body surface, which, intuitively, is important for reasoning about interpersonal contact. Our approach, in contrast, operates on the parameters of two parametric human body models, which represent the surfaces of two people closely interacting. After training, our model is able to generate samples of plausible pairs of 3D bodies in social interaction from pure noise. The model can also be conditioned in the output of a human pose and shape regressor. In this conditional case, the model takes the noisy output and generates similar poses but with realistic social interaction." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.581, + 0.468, + 0.822 + ], + "angle": 0, + "content": "We then demonstrate how exploit BUDDI's knowledge of human proxemics to guide 3D mesh reconstruction of people in a close social interaction from a single image. To this end, we introduce a novel optimization-based approach, which uses BUDDI as a data-driven prior. We initialize our optimization routine with samples from BUDDI, conditioned to the output of a state-of-the-art multi-person human mesh regressor [42]. We then optimize over SMPL-X pose, shape, and translation parameters to match detected 2D joint locations. We incorporate guidance from the diffusion model using a loss inspired by the Score-Distillation loss from the 3D object creation literature [35]: In each optimization step, BUDDI refines the current estimate towards a more plausible social interaction conditioned on the initial predictions. The refined pose, shape and translation serve as prior in the overall objective function." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.826, + 0.468, + 0.899 + ], + "angle": 0, + "content": "Our contributions include (1) presenting the first generative model of a pair of 3D people in close social interaction and (2) a novel approach for reconstructing 3D human meshes from images without relying on ground-truth contact annotations. We perform extensive experiments with" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.092, + 0.89, + 0.316 + ], + "angle": 0, + "content": "BUDDI to evaluate its performance on the FlickrCI3D Signatures dataset [8] as well as CHI3D, and the recent Hi4D dataset [54] and find that it outperforms the state of the art as well as strong baselines. We also evaluate the unconditional samples from the diffusion model in a perceptual study, where people find our samples more realistic \\(44.4\\%\\) when compared over real samples, where \\(50\\%\\) is the upperbound where they do not see any difference. Importantly, we find that our optimization approach significantly improves the results of [42] both quantitatively and qualitatively. This work opens up a new avenue of research on digital human synthesis, laying the foundation for a deeper understanding of human social behavior derived from image data. Our data, code, and model will be available for research." + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.333, + 0.64, + 0.348 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.357, + 0.89, + 0.568 + ], + "angle": 0, + "content": "Generating 3D humans. There has been recent interest in generating 3D humans, in different contexts. Several methods automatically populate static 3D scenes with 3D humans [14, 60, 61], while more recent methods generate both body and hand poses to interact with 3D objects [43, 44, 49]. Other work generates human motions conditioned on different inputs such as audio [26, 47] or text [33, 34, 45]. Concurrent work proposes text-to-3D diffusion-based approaches to generate motion of two interacting humans [27, 40]. Neither method predicts the full body surface, but rather they synthesize either 3D joint locations or SMPL pose parameters for the average body. These methods are not used as priors for reconstructing interacting people from images." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.569, + 0.89, + 0.824 + ], + "angle": 0, + "content": "To model 3D human proxemics probabilistically, we employ diffusion models, which achieve impressive performance on image generation tasks [7, 15, 38, 39]. They have recently been adopted in 3D human motion generation scenarios: MDM [45] generates plausible motions conditioned on text input; PhysDiff [55] incorporates physical constraints in the diffusion process to generate physically plausible motions; and EDGE [47] uses a transformer-based diffusion model for dance generation. Related work [4, 6, 28] has investigated different modalities for the conditioning, e.g., audio, text, or action classes. EgoEgo [25] generates plausible full-body motions conditioned on the head motion. SceneDiffuser [16] focuses on the scene-conditioned setting. We also rely on techniques from the diffusion literature, but consider the unique setting where two people are in close interaction and leverage this for single-image 3D reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.826, + 0.89, + 0.9 + ], + "angle": 0, + "content": "Multi-person 3D human mesh estimation. An extensive line of work focuses on reconstructing the 3D human pose and shape of a single person from images using optimization [2, 11, 24, 32, 37, 46, 50] or regression approaches [1, 12, 19, 20, 22, 29, 31, 51, 58, 59]. Capitalizing" + }, + { + "type": "page_number", + "bbox": [ + 0.485, + 0.947, + 0.514, + 0.956 + ], + "angle": 0, + "content": "9688" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.088, + 0.425, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.429, + 0.088, + 0.891, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.275, + 0.893, + 0.374 + ], + "angle": 0, + "content": "Figure 2. BUDDI: BUddies Diffusion model. On the left, we illustrate the architecture of BUDDI, our diffusion model for modeling 3D social proxemics between two people in close interaction. The diffusion process is applied directly on SMPL-X body parameters. To condition BUDDI on estimated body model parameters, \\( c_{\\mathrm{H}} \\), we concatenate the parameters along the token dimension. On the right, we illustrate the optimization method with BUDDI as prior. Our optimization takes detected keypoints [3, 52] and an initial regressor estimate [42] as input. Given the regressor estimate, we sample from BUDDI to obtain \\( \\tilde{x} \\) which we use to initialize the optimization routine. In each optimization iteration, we take a single diffuse-denoise step on the current estimate using the learned denoiser model \\( D \\). Our losses encourage the current estimate to be close to the refined meshes \\( (L_{\\mathrm{diffusion}}) \\) and to the initial estimate and detected keypoints \\( (L_{\\mathrm{fitting}}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.4, + 0.47, + 0.505 + ], + "angle": 0, + "content": "on these techniques, recent approaches focus explicitly on reconstructing multiple people jointly from a single image. Zanfir et al. [56] propose an optimization solution, while Jiang et al. [18] and Sun et al. [41] rely on deep networks to regress the pose and shape for all people in the image. BEV [42] extends ROMP [41] to reason about the depth of people while taking age/height into account." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.47, + 0.779 + ], + "angle": 0, + "content": "The above methods do not address contact between people. To do so, Fieraru et al. [8] introduce the first datasets with ground-truth labels for the body regions in contact between humans. Labels are collected using MoCap (CHI3D) or human annotators (FlickrCI3D Signatures). They propose an optimization approach that requires the ground-truth contact map to reconstruct people in close proximity at test time. More recently, REMIPS [9] introduces a transformer-based method that regresses the 3D pose of multiple people. REMIPS is trained using the above datasets while taking into account contact and interpenetration. In this work, we take a very different approach by learning and exploiting a 3D generative proxemics prior. We use the ground-truth contact maps to generate pseudoground truth 3D human fits from which we learn the diffusion model; once this is learned, we show that it can be used as a prior to recover plausible bodies in close proximity from images without explicit knowledge of contact maps." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Data-driven priors in optimization. Optimization-based methods for 3D human pose and shape estimation, like SMPLify [2], are versatile and allow different data-driven prior terms to be incorporated in the objective function. Different methods have been used to learn pose priors including GMMs [2], VAEs [32], neural distance fields [46], and normalizing flows [57]. ProHMR [23] learns a pose prior conditioned on image pixels. HuMoR [37] incor" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.4, + 0.892, + 0.522 + ], + "angle": 0, + "content": "porates a data-driven motion prior in the iterative optimization. POSA [14] learns a prior for human-scene interaction from PROX data [13] and uses it in their optimization. In contrast to these methods, we use a diffusion model to capture the joint distribution over SMPL-X parameters for two people interacting and show that we can both sample from the model and use during optimization to improve the pose estimates of interacting people." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.548, + 0.591, + 0.563 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.579, + 0.892, + 0.746 + ], + "angle": 0, + "content": "We introduce BUDDI, a generative model of two people in close social interaction. Because of the complexity and multimodality of the data, we turn to denoising diffusion probabilistic models [15] to address this task. In Sec. 3.1, we describe the basics of diffusion, and the parameterization we employ to model people in contact. In addition to sampling new body meshes from our model, our generative model can serve as a prior for reconstructing 3D humans from images. In Sec. 3.2, we describe an optimization procedure that incorporates BUDDI as a prior to recover two SMPL-X meshes from observed 2D keypoints." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.901 + ], + "angle": 0, + "content": "For all of the following, we use the SMPL-X [32] body model to represent the human bodies. SMPL-X is a differentiable function that maps pose, \\(\\pmb{\\theta} \\in \\mathbb{R}^{21 \\times 3}\\), shape, \\(\\beta \\in \\mathbb{R}^{10}\\), and expression, \\(\\psi \\in \\mathbb{R}^{10}\\) parameters to a mesh consisting of \\(N_v = 10,475\\) vertices \\(V \\in \\mathbb{R}^{N_v \\times 3}\\). We place the generated meshes in the world by rotating and translating them by \\(\\phi \\in \\mathbb{R}^3\\) and \\(\\gamma \\in \\mathbb{R}^3\\). We denote person \\(a\\)'s parameters as \\(X^a = [\\phi^a, \\pmb{\\theta}^a, \\pmb{\\beta}^a, \\gamma^a]\\) and \\(X^b = [\\phi^b, \\pmb{\\theta}^b, \\pmb{\\beta}^b, \\gamma^b]\\). For simplicity, we refer to both people when no index is specified, e.g., \\(X\\) refers \\(X^a\\) and \\(X^b\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9689" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.091, + 0.377, + 0.106 + ], + "angle": 0, + "content": "3.1. Diffusion Model for 3D Proxemics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.113, + 0.469, + 0.22 + ], + "angle": 0, + "content": "Denoising diffusion models are latent variable generative models that learn to transform random noise into the desired data distribution \\( p_{\\mathrm{data}} \\) through a forward and reverse process. The forward inference process is a Markov chain over \\( T \\) steps given by transitions \\( q(\\mathbf{x}_{t + 1}|\\mathbf{x}_t) \\), which gradually adds Gaussian noise to clean samples \\( \\mathbf{x}_0 \\) from the data distribution according to a fixed variance schedule \\( \\sigma_t \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.22, + 0.469, + 0.386 + ], + "angle": 0, + "content": "The reverse process \\( q(\\mathbf{x}_{t - 1}|\\mathbf{x}_t) \\) then gradually denoises noisy samples back into the data distribution. The reverse process transitions follow a Gaussian distribution when conditioned on \\( x_0 \\), but must be inferred during the generative process. Following Ramesh et al. [36], we train a neural network \\( D \\) that predicts a sample \\( \\hat{\\mathbf{x}}_0 = D(\\mathbf{x}_t; t) \\) from a noisy sample \\( \\mathbf{x}_t \\) given the noise level \\( t \\). For the task of reconstructing humans from images, when we have initial estimates of the SMPL-X parameters of two humans, we condition the denoising network \\( D \\) on \\( \\mathbf{c}_H \\), the predicted SMPL-X parameters of the two humans by a regressor." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.386, + 0.469, + 0.48 + ], + "angle": 0, + "content": "We refer to the process of adding noise as diffusion and the process of removing the noise via \\(D\\) as denoising. Specifically, we diffuse a ground-truth sample \\(\\mathbf{x}_0\\) by uniformly sampling a noise level \\(t\\) with \\(\\epsilon_t \\sim \\mathcal{N}(0, \\mathbf{I})\\) to obtain the noisy sample \\(\\mathbf{x}_t = \\sqrt{\\sigma_t'}\\mathbf{x}_0 + \\sqrt{1 - \\sigma_t'}\\epsilon_t\\) with \\(\\sigma_t' = \\prod_{i=1}^{t} (1 - \\sigma_t)\\)." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.48, + 0.291, + 0.493 + ], + "angle": 0, + "content": "We then train \\(D\\) to minimize" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.504, + 0.469, + 0.521 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {\\mathbf {x} _ {0} \\sim p _ {\\text {d a t a}}} \\mathbb {E} _ {t \\sim \\mathcal {U} \\{0, T \\}, \\mathbf {x} _ {t} \\sim q (\\cdot | \\mathbf {x} _ {0})} | | D (\\mathbf {x} _ {t}; t, \\mathbf {c} _ {H}) - \\mathbf {x} _ {0} | |, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.53, + 0.469, + 0.56 + ], + "angle": 0, + "content": "where we set \\(\\mathbf{c}_H = \\varnothing\\) for \\(20\\%\\) of conditional model training, and all of unconditional model training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.578, + 0.469, + 0.697 + ], + "angle": 0, + "content": "Architecture. Because we aim to model close contact between people, we choose a model state space that can express the full surface of the human body. Specifically, in contrast to prior work in human motion diffusion that operate only on joint angles and locations [45, 55], we directly operate on the full SMPL-X parameters of the two people. A sample \\( \\mathbf{x} \\) thus corresponds to the concatenation of two bodies:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.707, + 0.436, + 0.726 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} = \\left[ X ^ {a}, X ^ {b} \\right] = \\left[ \\phi^ {a}, \\boldsymbol {\\theta} ^ {a}, \\boldsymbol {\\beta} ^ {a}, \\gamma^ {a}, \\phi^ {b}, \\boldsymbol {\\theta} ^ {b}, \\boldsymbol {\\beta} ^ {b}, \\gamma^ {b} \\right].\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We denoise a sample \\(\\mathbf{x}_t\\) with a transformer encoder block on tokenized parameters. Specifically, each parameter of each person is tokenized into 152-dimensional latent vectors with per-parameter and per-person embedding layers. We tokenize the noise level \\(t\\) similarly with a noise embedding. When conditioning is available, i.e. SMPL-X estimates for person \\(a\\) and \\(b\\), we similarly tokenize the parameters to be used as additional tokens. We pass the available tokens into the transformer encoder, and similarly decode the output tokens with per-token embeddings. We illustrate the denoiser architecture in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "Losses. We employ standard human pose and shape regularization losses. We write our training objective as" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.133, + 0.892, + 0.15 + ], + "angle": 0, + "content": "\\[\nL _ {D} = L _ {\\theta} + L _ {\\beta} + L _ {\\gamma} + L _ {v 2 v}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.16, + 0.892, + 0.251 + ], + "angle": 0, + "content": "where \\(L_{\\theta}, L_{\\beta}, L_{\\gamma}\\) denote squared L2-losses on respective body model parameters, and \\(L_{v2v}\\) denotes a squared L2 loss on model vertices. We use 6D rotation representations [62] for global orientation and pose, and model the relative translation between \\(a\\) and \\(b\\). We show generated samples from our unconditional model in Fig. 3." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.26, + 0.834, + 0.275 + ], + "angle": 0, + "content": "3.2. Optimization with the Proxemics Prior" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.283, + 0.892, + 0.448 + ], + "angle": 0, + "content": "Reconstructing 3D human meshes from a single image is an extremely under-constrained problem, and priors over human pose and shape are crucial in an optimization based framework for recovering plausible meshes [2, 32, 56]. Our problem involves people in close contact, which requires correctly placing the meshes in context with each other, which has only been done when given ground truth contact annotations at test time [8]. We remove the need for ground-truth contact maps by using our generative model as a prior during reconstruction with a score distillation approach [35, 48]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.448, + 0.892, + 0.509 + ], + "angle": 0, + "content": "During inference, we observe detected 2D keypoints \\(\\tilde{J}_{2D}\\) and initial body model parameter estimates \\(\\mathbf{c}_H\\) from a regressor [42]. We then optimize the body parameters of two people to minimize" + }, + { + "type": "equation", + "bbox": [ + 0.562, + 0.521, + 0.892, + 0.537 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {O p t i z i m a t i o n w . B U D D I}} = L _ {\\text {f i t t i n g}} + L _ {\\text {d i f f u s i o n}}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.547, + 0.892, + 0.638 + ], + "angle": 0, + "content": "\\(L_{\\mathrm{fitting}}\\) ensures that the solution stays close to the image evidence, while \\(L_{\\mathrm{diffusion}}\\) is a data-driven prior using our conditional diffusion model. We treat this prior as similar to those used for 3D pose in previous works such as GMM [2] and V-Poser [32], but for 3D proxemics. We illustrate the optimization procedure in Fig. 2 right." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.638, + 0.892, + 0.698 + ], + "angle": 0, + "content": "We initialize our optimization by generating a sample \\(\\tilde{\\mathbf{x}}\\) from the conditional model. We sample with DDIM sampling with 100 evenly spaced steps. We then use the data fitting loss:" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.712, + 0.892, + 0.728 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {f i t t i n g}} = \\lambda_ {J} L _ {J} + \\lambda_ {\\tilde {\\theta}} L _ {\\tilde {\\theta}} + \\lambda_ {P} L _ {P}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.74, + 0.892, + 0.799 + ], + "angle": 0, + "content": "where \\( L_{J} \\) denotes 2D re-projection error between the reprojected 3D joints of the current estimate and the detected 2D keypoints, and \\( L_{\\tilde{\\theta}} \\) is a prior for the solution to be close to the denoised initialization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.8, + 0.892, + 0.86 + ], + "angle": 0, + "content": "\\(L_{P}\\) denotes an interpenetration loss between two people that pushes inside vertices to the surface, which we use winding numbers [17] to find intersecting vertices between low-resolution SMPL-X meshes of the current estimates:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.868, + 0.892, + 0.904 + ], + "angle": 0, + "content": "\\[\nL _ {P} = \\sum_ {v \\in V _ {I} ^ {a}} \\min _ {u \\in V ^ {b}} \\| v - u \\| ^ {2} + \\sum_ {v \\in V _ {I} ^ {b}} \\min _ {u \\in V ^ {a}} \\| v - u \\| ^ {2}, \\tag {5}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9690" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.127, + 0.111, + 0.868, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.338, + 0.895, + 0.385 + ], + "angle": 0, + "content": "Figure 3. Generative Proxemics: Samples from BUDDI. All samples are unconditionally generated from pure noise using the trained diffusion model. We select several representative examples and show two views per sample. These samples reveal that BUDDI has learned the distribution of people in close contact including embracing each other, playing sports, sitting side by side, and taking photographs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.407, + 0.469, + 0.438 + ], + "angle": 0, + "content": "where \\(V_{I}^{a}\\) denotes vertices of \\(M^a\\) intersecting the lowresolution mesh of \\(M^b\\); and vice versa for \\(V_{I}^{b}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.438, + 0.471, + 0.514 + ], + "angle": 0, + "content": "To use the prior on human interaction into account, we use the learned denoising model \\(D\\) from BUDDI and perform a single diffuse-denoise step, with a noise level at \\(t = 10\\), on the current estimate. The denoised estimate, \\(\\hat{\\mathbf{x}}_0 = D(\\mathbf{x}_t; t, \\mathbf{c}_H)\\), regularizes the current estimate via" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.526, + 0.47, + 0.545 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {d i f f u s i o n}} = \\left\\| D \\left(\\mathbf {x} _ {t}; t, \\mathbf {c} _ {H}\\right) - \\mathbf {x} \\right\\|, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.554, + 0.473, + 0.633 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}_t = \\sqrt{\\sigma_t^{\\prime}}\\mathbf{x}_{\\mathrm{no - grad}} + \\sqrt{1 - \\sigma_t^{\\prime}}\\epsilon_t\\) denotes the diffused body model parameters of the current estimate, and \\(\\mathbf{x}_{\\mathrm{no - grad}}\\) denotes the current estimate with detached gradients. \\(\\hat{\\mathbf{x}}_0\\) , and encourages \\(\\mathbf{x}\\) to be close to \\(\\hat{\\mathbf{x}}_0\\) . In practice, we penalize the decoded parameters of \\(\\mathbf{x}\\) and \\(\\hat{\\mathbf{x}}_0\\) directly as" + }, + { + "type": "equation", + "bbox": [ + 0.128, + 0.643, + 0.47, + 0.686 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {\\text {d i f f u s i o n}} = \\lambda_ {\\hat {\\phi}} | | \\hat {\\phi} _ {0} - \\phi | | + \\lambda_ {\\hat {\\theta}} | | \\hat {\\theta} _ {0} - \\theta | | \\tag {7} \\\\ + \\lambda_ {\\hat {\\boldsymbol {\\beta}}} | | \\hat {\\boldsymbol {\\beta}} _ {0} - \\boldsymbol {\\beta} | | + \\lambda_ {\\hat {\\boldsymbol {\\gamma}}} | | \\hat {\\boldsymbol {\\gamma}} _ {0} - \\boldsymbol {\\gamma} | |. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.696, + 0.471, + 0.757 + ], + "angle": 0, + "content": "Intuitively, this loss uses the learned denoiser \\( D \\) to take a step from the current estimate towards the data distribution of two people in close proximity, conditioned on the regressor prediction." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.77, + 0.3, + 0.788 + ], + "angle": 0, + "content": "4. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Training Data. There are few datasets containing 3D ground truth of humans in close social interaction [8, 54]. Such datasets are usually captured in lab environments, consequently they are small and do not contain the variety of interactions between humans \"in the wild,\" e.g. when playing sports or taking social pictures. To address this lack of data, we create Flickr Fits, i.e. SMPL-X fits" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.407, + 0.895, + 0.575 + ], + "angle": 0, + "content": "for Flickr images portraying humans in contact scenarios. For this, we use FlickrCI3D Signatures [8], a dataset of images showing interacting humans collected from Flickr with discrete human-human contact annotations. Specifically, the SMPL-X body surface is divided into \\( R = 75 \\) regions such that each region, \\( r \\), roughly covers a similar area. For a given photo, the human annotators assign a binary label indicating contact between a region on one person and a region on the other. For two meshes, \\( M^a \\) and \\( M^b \\), the annotation can be represented as a binary contact map \\( \\mathcal{C}^D \\in \\{0,1\\}^{R \\times R} \\), where" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.585, + 0.892, + 0.62 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} _ {i j} ^ {D} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} r _ {i} \\text {o f} M ^ {a} \\text {i s i n c o n t a c t w i t h} r _ {j} \\text {o f} M ^ {b} \\\\ 0, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.895, + 0.81 + ], + "angle": 0, + "content": "We use these ground-truth contact maps in an optimization routine for fitting two people to detected keypoints, similar to Sec. 3.2 but replace the diffusion model prior with standard image fitting priors. The dataset contains 10,631/1,139 train/test images, with one image containing multiple contact annotations. Note that we use pairs of people in contact as labeled in the FlickrCI3D Signatures dataset [8] and only use images containing matching regressor estimates, 2D keypoints, and contact labels. See Fig. 4 for qualitative examples and the Sup. Mat. for more details about the selection process and an evaluation of the optimization method with ground-truth contact labels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We also augment our training data with available MoCap data, which is considerably smaller than those obtained from image fits: CHI3D [8] contains 3/2 pairs of training/test subjects performing 127 sequences of two-person interactions like hugs or kicks with ground-truth SMPL-X bodies. One frame per sequence has contact map annota" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.514, + 0.958 + ], + "angle": 0, + "content": "9691" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.105, + 0.089, + 0.468, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.274, + 0.472, + 0.331 + ], + "angle": 0, + "content": "Figure 4. Flickr Fits. We visualize the output of the optimization process that reconstructs two people in close proximity using ground-truth contact maps, shown from four different views. We use these 3D fits as training data for BUDDI." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.341, + 0.47, + 0.552 + ], + "angle": 0, + "content": "tions. We use only the contact frame of the sequences from two subject pairs, resulting in 247 mesh pairs for training, and the third pair for evaluation. Hi4D [54] contains sequences of 20 pairs of people interacting with each other. The interactions include actions like hugging, dancing, and fighting. We randomly split the data into \\(14/3/3\\) pairs for train/val/test and use every fifth frame of the subsequence involving contact as labeled in Hi4D, resulting in about 1K mesh pairs for training. The body representation format in Hi4D is SMPL, which we transfer to SMPL-X using the SMPL-X code repository [32]. Please see the Sup. Mat. for more details of the datasets. Note that while we use SMPL-X model, BUDDI is not trained on hands because none of these datasets contain hand poses." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.47, + 0.794 + ], + "angle": 0, + "content": "BUDDI Training. BUDDI is trained with meshes from FlickrCI3D Signatures Fits, CHI3D, and Hi4D. We use \\(60\\%\\) Flickr, \\(20\\%\\) CHI3D, and \\(20\\%\\) Hi4D data distribution per batch with batch size 512. The transformer backbone has six layers and eight heads; we use \\(10\\%\\) dropout and randomly shuffle the order of people during training. To train BUDDI, we randomly sample noise levels \\(t\\) up to 1000 using a cosine noise schedule [30]. We use the Adam optimizer [21] with learning rate \\(10^{-4}\\). We train two separate networks: an unconditional model for generation and the conditional version for reconstruction. For the conditional model, we use all camera views of the MoCap datasets, i.e. 4/8 cameras for CHI3D/Hi4D. The unconditional model is trained on 3D MoCap fits in the world coordinate system. To sample new poses, we use DDIM sampling starting at noise levels \\(t = 1000\\) in steps of 10." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Optimization Details. During optimization, we experiment with different noise levels, between 10 and 100, and find that \\( t = 10 \\) does not disturb the inputs too much, but enough for \\( D \\) to generate new configurations. We use detected 2D keypoints from OpenPose [3] and ViTpose [52] and BEV [42] estimates as conditioning. Unlike single-person mesh regressors, BEV is designed to predict multiple" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "people including their relative depth. Please see Sup. Mat. for more details." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.134, + 0.633, + 0.151 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.16, + 0.892, + 0.492 + ], + "angle": 0, + "content": "Baselines. We compare our reconstruction method with BEV [42], which is also used as an input to our conditional model. Since there is no other available work that reasons about people in close social interaction, we experiment with simple but effective baselines. We train the transformer model of BUDDI to directly predict SMPL-X parameters of people in contact from BEV input, essentially a deterministic, single-step ablation of BUDDI. We also evaluate the direct conditional denoised output of BEV by BUDDI without any optimization. As another baseline, we propose an optimization routine that replaces \\( L_{\\text{diffusion}} \\) with a simple heuristic that takes the minimal distances between two meshes predicted by BEV and minimizes their distance during optimization along with the other energy terms. Finally, to compare the generation ability we train a VAE which we also use during the optimization routine in a similar manner to VPoser [32] but for two people by optimizing the VAE latent space instead of SMPL-X parameters. We refer to these models as Transformer, BUDDI (gen.), Contact Heuristic, and VAE, respectively. All baselines are trained on the same datasets as BUDDI with the same sampling strategies. Details about our baselines are provided in the Sup. Mat.." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.492, + 0.893, + 0.689 + ], + "angle": 0, + "content": "Metrics. We use standard evaluation metrics from the human pose and shape estimation literature. We also report the joint PA-MPJPE computed by performing Procrustes alignment of both people together. In addition to per-person metrics, this captures the relative orientation and translation of the two people. Since our method directly estimates 3D humans we propose a new metric similar to PCK [53] from the 2D pose literature called PCC, the percentage of correct contact points with respect to a radius \\( r \\). Specifically, given two meshes, \\( M^a / M^b \\) and a contact map \\( \\mathcal{C}^D \\) we compute the pairwise vertex-to-vertex Euclidean distances \\( d_{\\mathrm{eucl}}(\\mathcal{C}^D) \\) between annotated contact regions and consider the pair to be correct when \\( \\min(d_{\\mathrm{eucl}}(\\mathcal{C}^D)) < r \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.697, + 0.738, + 0.712 + ], + "angle": 0, + "content": "5.1. Unconditional Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.826 + ], + "angle": 0, + "content": "We qualitatively evaluate BUDDI by showing samples generated from the model in Fig. 1 and 3. Our approach is able to generate people in close proximity including embraces, handshakes, having a conversation, sitting side by side, and, in general, plausibly interacting with each other. Since it is trained on Internet image collections, it also learns to generate people posing for photographs or playing sports." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We further run a perceptual study to evaluate the realism of the generated social interactions against other methods. In a forced choice study, we compare our generated samples with samples from the real data distribution according to the 60/20/20 per-batch ratio for Flickr/CHI3D/Hi4D" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9692" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.089, + 0.899, + 0.511 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.521, + 0.895, + 0.568 + ], + "angle": 0, + "content": "Figure 5. Automatic reconstruction of people in close social interaction. We show qualitative results from a) BEV, b) contact heuristics, which takes the BEV output and encourages the closest parts to be in contact, and c) our method, which optimizes the BEV estimates against the image evidence with the BUDDI prior. Our approach recovers a plausible reconstruction with subtle details." + }, + { + "type": "table", + "bbox": [ + 0.139, + 0.587, + 0.411, + 0.693 + ], + "angle": 0, + "content": "
JOINT ↓\nPA-MPJPEPCC at radius ↑
510152025
BEV106-----
Transformer861440607382
BUDDI (gen.)921539587180
Heuristic681434496170
VAE1011128425565
BUDDI661944627381
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.708, + 0.472, + 0.765 + ], + "angle": 0, + "content": "Table 1. 3D Pose Evaluation on FlickrCI3D Signatures. We evaluate methods against the Flickr fits using their joint (two-person) PA-MPJPE expressed in mm. We also evaluate the percentage of correct contact points (PCC) for radius \\( r \\) mm." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.473, + 0.903 + ], + "angle": 0, + "content": "used during training. We also compare BUDDI against generations from the VAE and a non-parametric random baseline that samples meshes from the pseudo-ground truth after centering the two people. We do a forced choice comparison between BUDDI and these three other methods, asking workers on Amazon Mechanical Turk to choose the sample that shows a more realistic close social interaction. We use 256 samples per method. We collect ratings for 768" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.59, + 0.895, + 0.681 + ], + "angle": 0, + "content": "pairwise comparisons. In this study, BUDDI was chosen over random in \\(71.23\\%\\) of the comparisons, over the VAE in \\(60.17\\%\\), and over the training data in \\(44.4\\%\\). Note that \\(50\\%\\) is the upper bound for such forced choice comparisons, in which participants cannot tell the difference between real and generated samples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.682, + 0.895, + 0.773 + ], + "angle": 0, + "content": "For a quantitative evaluation, we compute the FID score between samples from BUDDI and samples from the \\( VAE \\) on concatenated SMPL-X parameters. We sample 8K examples per method and from our training data following the dataset ratio per batch. BUDDI has a lower FID score (1.6) compared to the \\( VAE \\) (3.3)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.786, + 0.691, + 0.803 + ], + "angle": 0, + "content": "5.2. Fitting with BUDDI" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We show qualitative results in Fig. 5 comparing BUDDI against BEV and the Contact Heuristic. Our approach is able to generate various types of human interactions with plausible contact and depth placement. It is also able to capture close interaction between a child and a parent. Although the Contact Heuristic (center) is able to move two" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "9693" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.079, + 0.089, + 0.904, + 0.203 + ], + "angle": 0, + "content": "
PER PERSON ↓ PA-MPJPEJOINT ↓ PA-MPJPEJOINT PA-MPJPE
backhugbasketballcheersdancefighthighfivehugkissposesidehugtalk
BEV78 / 84136200126109135121106163139142131118
Heuristic67 / 7112116883941319468159159118113109
BUDDI (F, C)70 / 77115200949212810810013311410410791
Transformer79 / 85120161141103138123128117106120105100
BUDDI (gen.)82 / 901171521391201371309610197115102101
VAE80 / 8213817513311414111987176162135140113
BUDDI70 / 769812795921131097210585889681
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.214, + 0.895, + 0.258 + ], + "angle": 0, + "content": "Table 2. Evaluation of BUDDI on Hi4D. We compare the output of BUDDI to the proposed baseline methods on the Hi4D challenge. The first block shows methods that do not use Hi4D data during training or are optimization based without access to priors trained on Hi4D. BUDDI (F,C) in particular, is our model BUDDI trained on Flickr and CHI3D data only. All errors are reported in mm for 3D Joints." + }, + { + "type": "table", + "bbox": [ + 0.14, + 0.279, + 0.413, + 0.393 + ], + "angle": 0, + "content": "
PER PERSON ↓\nPA-MPJPEJOINT ↓\nPA-MPJPE
BEV50 5296
Transformer54 56105
BUDDI (gen.)53 5380
Heuristic49 46105
VAE54 54103
BUDDI48 4768
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.403, + 0.472, + 0.446 + ], + "angle": 0, + "content": "Table 3. Quantitative Evaluation on CHI3D. We compare the output of our model to the baselines on CHI3D (pair s03). All errors reported in mm for 3D Joints." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.458, + 0.47, + 0.549 + ], + "angle": 0, + "content": "people closer together, which helps with image alignment, upon close observation it is not able to capture the subtle interaction between people that happens during intimate interaction. BUDDI's estimates are more realistic and better capture the subtle details of interaction. We provide additional qualitative examples in the Sup. Mat." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.551, + 0.47, + 0.824 + ], + "angle": 0, + "content": "We further report the percentage of correct contact (PCC) with respect to the ground truth contact map on the FlickrCI3D Signatures test set in Table 1. The table also shows the pose reconstruction accuracy against our Flickr Fits. All metrics show improvement over BEV, in particular the joint PA-MPJPE. Non-optimization methods, i.e. Transformer and BUDDI (gen.), are able to predict plausible contacts, with similar PCC accuracy to BUDDI, but struggle to reconstruct the data with a worse joint PA-MPJPE. The Heuristic, in contrast, achieves a lower reconstruction error, but worse PCC. Our approach which leverages the learned prior during optimization can recover both the relative positions and contacts between the two people. To provide insights into the performance of single-person mesh regressors when evaluated on the two-person reconstruction task, we run 4D Humans [10] on Flickr Fits. The joint PA-MPJPE is \\(344\\mathrm{mm}\\) which is high, as expected, since these methods are not trained to reason about proximity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We further evaluate our model against ground truth MoCap data in Tables 2 and 3. Optimization with BUDDI consistently improves the two-person reconstruction error over BEV and other baselines. When evaluated per action, the strongest improvements over BEV come from complex" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.283, + 0.895, + 0.51 + ], + "angle": 0, + "content": "close social interactions like hugging or kissing, at \\(58\\mathrm{mm}\\) and \\(54\\mathrm{mm}\\) absolute improvement over BUDDI respectively. The Heuristic baseline achieves a low PA-MPJPE reconstruction error on all three datasets in particular for poses with a few physical contact points, such as a handshake, whereas more complex contact, such as a hug, requires data-driven priors like BUDDI; we provide an analysis in the Sup. Mat. to quantify this hypothesis. Transformer and BUDDI (gen.) have lower joint PA-MPJPE errors than BEV and the Heuristic, but worse per-person reconstruction errors. The VAE results suggest that directly operating in the latent space of a generative model is challenging and not sufficient to accurately recover close social interactions. BUDDI, in contrast, is able to model a wide variety of poses, as supported by the numerical results." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.526, + 0.62, + 0.543 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.894, + 0.674 + ], + "angle": 0, + "content": "We propose BUDDI, a diffusion model for close human-human interaction. We train BUDDI from 3D fits obtained from a large-scale dataset of images with ground truth contact annotations as well as a small set of available mocap data. BUDDI enables unconditional sampling of people in close social interaction. More importantly, we also demonstrate how BUDDI can be used as effective prior for single-view 3D reconstruction of two people in close proximity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.894, + 0.811 + ], + "angle": 0, + "content": "Our core contribution of a generative proxemics prior provides the foundation for future work on modeling and capturing human interaction. For example, future work could iteratively apply our method to reconstruct more than two people in close proximity, explore other conditioning modalities like pixel features, text, action labels, or, by taking the outputs of recent single-person mesh regressors into account, address reconstructing fine-grained interaction including finger pose and facial expressions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.812, + 0.892, + 0.873 + ], + "angle": 0, + "content": "Acknowledgements: We thank A. Holynski, E. Weber, F. Warburg, B. Peebles, J. Rajasegaran, K. Mangalam and N. Athanasiou for insightful discussions, and A. Cseke, T. McConnell and T. Alexiadis for running the user study." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.873, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Disclosure: https://files.is.tue.mpg.de/black/CoI_CVPR_2024.txt" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "9694" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Anurag Arnab, Carl Doersch, and Andrew Zisserman. Exploiting temporal context for 3D human pose estimation in the wild. In Computer Vision and Pattern Recognition (CVPR), 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.255 + ], + "angle": 0, + "content": "[2] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J. Black. Keep it SMPL: Automatic estimation of 3D human pose and shape from a single image. In European Conference on Computer Vision (ECCV), pages 561-578. Springer International Publishing, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.258, + 0.47, + 0.327 + ], + "angle": 0, + "content": "[3] Zhe Cao, Gines Hidalgo Martinez, Tomas Simon, Shih-En Wei, and Yaser Sheikh. OpenPose: Realtime multi-person 2D pose estimation using part affinity fields. Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 43 (1):172-186, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.329, + 0.47, + 0.398 + ], + "angle": 0, + "content": "[4] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18000-18010, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.401, + 0.468, + 0.455 + ], + "angle": 0, + "content": "[5] April H Crusco and Christopher G Wetzel. The midas touch: The effects of interpersonal touch on restaurant tipping. *Personality and Social Psychology Bulletin*, 10(4): 512-517, 1984." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.458, + 0.468, + 0.514 + ], + "angle": 0, + "content": "[6] Rishabh Dabral, Muhammad Hamza Mughal, Vladislav Golyanik, and Christian Theobalt. Mofusion: A framework for denoising-diffusion-based motion synthesis. In Computer Vision and Pattern Recognition (CVPR), 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.516, + 0.468, + 0.557 + ], + "angle": 0, + "content": "[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat GANs on image synthesis. Conference on Neural Information Processing Systems (NeurIPS), 34:8780-8794, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.559, + 0.468, + 0.627 + ], + "angle": 0, + "content": "[8] Mihai Fieraru, Mihai Zanfir, Elisabeta Oneata, Alin-Ionut Popa, Vlad Olaru, and Cristian Sminchisescu. Three-dimensional reconstruction of human interactions. In Computer Vision and Pattern Recognition (CVPR), pages 7214-7223, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.63, + 0.468, + 0.712 + ], + "angle": 0, + "content": "[9] Mihai Fieraru, Mihai Zanfir, Teodor Szente, Eduard Bazavan, Vlad Olaru, and Cristian Sminchisescu. REMIPS: Physically consistent 3D reconstruction of multiple interacting people under weak supervision. In Conference on Neural Information Processing Systems (NeurIPS), pages 19385-19397, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.715, + 0.468, + 0.77 + ], + "angle": 0, + "content": "[10] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa*, and Jitendra Malik*. Humans in 4D: Reconstructing and tracking humans with transformers. In International Conference on Computer Vision (ICCV), 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.773, + 0.468, + 0.828 + ], + "angle": 0, + "content": "[11] Peng Guan, Alexander Weiss, Alexandru Balan, and Michael J. Black. Estimating human shape and pose from a single image. In International Conference on Computer Vision (ICCV), pages 1381-1388, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.83, + 0.468, + 0.871 + ], + "angle": 0, + "content": "[12] Riza Alp Guler and Iasonas Kokkinos. HoloPose: Holistic 3D human reconstruction in-the-wild. In Computer Vision and Pattern Recognition (CVPR), 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.468, + 0.902 + ], + "angle": 0, + "content": "[13] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J Black. Resolving 3D human pose ambigui" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "ties with 3D scene constraints. In International Conference on Computer Vision (ICCV), pages 2282-2292, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.123, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[14] Mohamed Hassan, Partha Ghosh, Joachim Tesch, Dimitrios Tzionas, and Michael J Black. Populating 3D scenes by learning human-scene interaction. In Computer Vision and Pattern Recognition (CVPR), pages 14708-14718, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.892, + 0.22 + ], + "angle": 0, + "content": "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Conference on Neural Information Processing Systems (NeurIPS), pages 6840-6851, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.222, + 0.892, + 0.291 + ], + "angle": 0, + "content": "[16] Siyuan Huang, Zan Wang, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, and Song-Chun Zhu. Diffusion-based generation, optimization, and planning in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.293, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[17] Alec Jacobson, Ladislav Kavan, and Olga Sorkine-Hornung. Robust inside-outside segmentation using generalized winding numbers. ACM Transactions on Graphics (TOG), 32(4): 1-12, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[18] Wen Jiang, Nikos Kolotouros, Georgios Pavlakos, Xiaowei Zhou, and Kostas Daniilidis. Coherent reconstruction of multiple humans from a single image. In Computer Vision and Pattern Recognition (CVPR), pages 5578-5587, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.406, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[19] Hanbyul Joo, Natalia Neverova, and Andrea Vedaldi. Exemplar fine-tuning for 3D human model fitting towards in-the-wild 3D human pose estimation. In International Conference on 3D Vision (3DV), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.462, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[20] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In Computer Vision and Pattern Recognition (CVPR), pages 7122-7131, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.519, + 0.892, + 0.559 + ], + "angle": 0, + "content": "[21] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.562, + 0.892, + 0.629 + ], + "angle": 0, + "content": "[22] Nikos Kolotouros, Georgios Pavlakos, Michael J. Black, and Kostas Daniilidis. Learning to reconstruct 3D human pose and shape via model-fitting in the loop. In International Conference on Computer Vision (ICCV), pages 2252-2261, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.633, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[23] Nikos Kolotouros, Georgios Pavlakos, Dinesh Jayaraman, and Kostas Daniilidis. Probabilistic modeling for human mesh recovery. In International Conference on Computer Vision (ICCV), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.689, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[24] Christoph Lassner, Javier Romero, Martin Kiefel, Federica Bogo, Michael J. Black, and Peter V. Gehler. Unite the people: Closing the loop between 3D and 2D human representations. In Computer Vision and Pattern Recognition (CVPR), pages 6050-6059, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.76, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[25] Jiaman Li, Karen Liu, and Jiajun Wu. Ego-body pose estimation via ego-head pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17142-17151, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.817, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[26] Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. AI choreographer: Music conditioned 3D dance generation with AIST++. In International Conference on Computer Vision (ICCV), pages 13401-13412, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[27] Han Liang, Wenqian Zhang, Wenxuan Li, Jingyi Yu, and Lan Xu. Intergen: Diffusion-based multi-human motion" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.957 + ], + "angle": 0, + "content": "9695" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.47, + 0.12 + ], + "angle": 0, + "content": "generation under complex interactions. arXiv preprint arXiv:2304.05684, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.47, + 0.162 + ], + "angle": 0, + "content": "[28] Jianxin Ma, Shuai Bai, and Chang Zhou. Pretrained diffusion models for unified human motion synthesis. arXiv preprint arXiv:2212.02837, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.469, + 0.218 + ], + "angle": 0, + "content": "[29] Lea Müller, Ahmed A. A. Osman, Siyu Tang, Chun-Hao P. Huang, and Michael J. Black. On self-contact and human pose. In Computer Vision and Pattern Recognition (CVPR), pages 9990-9999, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.219, + 0.469, + 0.273 + ], + "angle": 0, + "content": "[30] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning (ICML), pages 8162-8171, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.469, + 0.344 + ], + "angle": 0, + "content": "[31] Mohamed Omran, Christoph Lassner, Gerard Pons-Moll, Peter V. Gehler, and Bernt Schiele. Neural body fitting: Unifying deep learning and model based human pose and shape estimation. In International Conference on 3D Vision (3DV), pages 484-494, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.469, + 0.413 + ], + "angle": 0, + "content": "[32] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.414, + 0.469, + 0.468 + ], + "angle": 0, + "content": "[33] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3D human motion synthesis with transformer VAE. In Computer Vision and Pattern Recognition (CVPR), pages 10985-10995, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.469, + 0.523 + ], + "angle": 0, + "content": "[34] Mathis Petrovich, Michael J Black, and Gül Varol. TEMOS: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision (ECCV), pages 480-497, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.469, + 0.564 + ], + "angle": 0, + "content": "[35] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.469, + 0.607 + ], + "angle": 0, + "content": "[36] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.609, + 0.469, + 0.663 + ], + "angle": 0, + "content": "[37] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J Guibas. HuMoR: 3D human motion model for robust pose estimation. In International Conference on Computer Vision (ICCV), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.665, + 0.469, + 0.72 + ], + "angle": 0, + "content": "[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.721, + 0.469, + 0.802 + ], + "angle": 0, + "content": "[39] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.469, + 0.845 + ], + "angle": 0, + "content": "[40] Yonatan Shafir, Guy Tevet, Roy Kapon, and Amit H Bermano. Human motion diffusion as a generative prior. arXiv preprint arXiv:2303.01418, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[41] Yu Sun, Qian Bao, Wu Liu, Yili Fu, Michael J Black, and Tao Mei. Monocular, one-stage, regression of multiple 3D people. In Computer Vision and Pattern Recognition (CVPR), pages 11179-11188, 2021." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[42] Yu Sun, Wu Liu, Qian Bao, Yili Fu, Tao Mei, and Michael J Black. Putting people in their place: Monocular regression of 3d people in depth. In Computer Vision and Pattern Recognition (CVPR), pages 13243-13252, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[43] Omid Taheri, Vasileios Choutas, Michael J. Black, and Dimitrios Tzionas. GOAL: Generating 4D whole-body motion for hand-object grasping. In Computer Vision and Pattern Recognition (CVPR), pages 13263-13273, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[44] Purva Tendulkar, Dídac Surís, and Carl Vondrick. FLEX: Full-body grasping without full-body grasps. In Computer Vision and Pattern Recognition (CVPR), 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.291 + ], + "angle": 0, + "content": "[45] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.292, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[46] Garvita Tiwari, Dimitrije Antic, Jan Eric Lenssen, Nikolaos Sarafianos, Tony Tung, and Gerard Pons-Moll. Pose-ndf: Modeling human pose manifolds with neural distance fields. In European Conference on Computer Vision (ECCV), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.39 + ], + "angle": 0, + "content": "[47] Jonathan Tseng, Rodrigo Castellon, and C Karen Liu. EDGE: Editable dance generation from music. Computer Vision and Pattern Recognition (CVPR), 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.392, + 0.892, + 0.461 + ], + "angle": 0, + "content": "[48] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12619-12629, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.462, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[49] Yan Wu, Jiahao Wang, Yan Zhang, Siwei Zhang, Otmar Hilliges, Fisher Yu, and Siyu Tang. Saga: Stochastic whole-body grasping with contact. In European Conference on Computer Vision (ECCV), pages 257-274, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.519, + 0.892, + 0.588 + ], + "angle": 0, + "content": "[50] Hongyi Xu, Eduard Gabriel Bazavan, Andrei Zanfir, William T. Freeman, Rahul Sukthankar, and Cristian Sminchisescu. GHUM & GHUML: Generative 3D human shape and articulated pose models. In Computer Vision and Pattern Recognition (CVPR), pages 6183-6192, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.59, + 0.892, + 0.644 + ], + "angle": 0, + "content": "[51] Yuanlu Xu, Song-Chun Zhu, and Tony Tung. DenseRaC: Joint 3D pose and shape estimation by dense render-and compare. In International Conference on Computer Vision (ICCV), 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.646, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[52] Yufei Xu, Jing Zhang, Qiming Zhang, and Dacheng Tao. Vitpose: Simple vision transformer baselines for human pose estimation. Advances in Neural Information Processing Systems, 35:38571-38584, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.703, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[53] Yi Yang and Deva Ramanan. Articulated human detection with flexible mixtures of parts. IEEE transactions on pattern analysis and machine intelligence, 35(12):2878-2890, 2012." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.746, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[54] Yifei Yin, Chen Guo, Manuel Kaufmann, Juan Zarate, Jie Song, and Otmar Hilliges. Hi4d: 4d instance segmentation of close human interaction. In Computer Vision and Pattern Recognition (CVPR), 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[55] Ye Yuan, Jiaming Song, Umar Iqbal, Arash Vahdat, and Jan Kautz. Physdiff: Physics-guided human motion diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16010-16021, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[56] Andrei Zanfir, Elisabeta Maroiniu, and Cristian Sminchisescu. Monocular 3d pose and shape estimation of multiple people in natural scenes-the importance of multiple scene" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9696" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.468, + 0.12 + ], + "angle": 0, + "content": "constraints. In Computer Vision and Pattern Recognition (CVPR), pages 2148-2157, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.191 + ], + "angle": 0, + "content": "[57] Andrei Zanfir, Eduard Gabriel Bazavan, Hongyi Xu, William T Freeman, Rahul Sukthankar, and Cristian Sminchisescu. Weakly supervised 3D human pose and shape reconstruction with normalizing flows. In European Conference on Computer Vision (ECCV), pages 465-481, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.261 + ], + "angle": 0, + "content": "[58] Andrei Zanfir, Eduard Gabriel Bazavan, Mihai Zanfir, William T Freeman, Rahul Sukthankar, and Cristian Sminchisescu. Neural descent for visual 3D human pose and shape. In Computer Vision and Pattern Recognition (CVPR), pages 14484-14493, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.469, + 0.331 + ], + "angle": 0, + "content": "[59] Hongwen Zhang, Yating Tian, Xinchi Zhou, Wanli Ouyang, Yebin Liu, Limin Wang, and Zhenan Sun. PyMAF: 3D human pose and shape regression with pyramidal mesh alignment feedback loop. In International Conference on Computer Vision (ICCV), pages 11446-11456, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.333, + 0.469, + 0.389 + ], + "angle": 0, + "content": "[60] Siwei Zhang, Yan Zhang, Qianli Ma, Michael J Black, and Siyu Tang. PLACE: Proximity learning of articulation and contact in 3D environments. In International Conference on 3D Vision (3DV), pages 642-651, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.39, + 0.469, + 0.445 + ], + "angle": 0, + "content": "[61] Yan Zhang, Mohamed Hassan, Heiko Neumann, Michael J Black, and Siyu Tang. Generating 3D people in scenes without people. In Computer Vision and Pattern Recognition (CVPR), pages 6194-6204, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.447, + 0.469, + 0.502 + ], + "angle": 0, + "content": "[62] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Computer Vision and Pattern Recognition (CVPR), pages 5745-5753, 2019." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.469, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9697" + } + ] +] \ No newline at end of file diff --git a/2024/Generative Proxemics_ A Prior for 3D Social Interaction from Images/b086fde4-b439-4f25-92c2-a2dc5214a0a3_origin.pdf b/2024/Generative Proxemics_ A Prior for 3D Social Interaction from Images/b086fde4-b439-4f25-92c2-a2dc5214a0a3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b4eba5b4615e989a8faa1eac30cfc1466048deba --- /dev/null +++ b/2024/Generative Proxemics_ A Prior for 3D Social Interaction from Images/b086fde4-b439-4f25-92c2-a2dc5214a0a3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d387a4584c50f34d95f3155dfaff8ce35487c5fe0c52cab57b4f99d15395c700 +size 8293775 diff --git a/2024/Generative Proxemics_ A Prior for 3D Social Interaction from Images/full.md b/2024/Generative Proxemics_ A Prior for 3D Social Interaction from Images/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2bb87c8c3c05f2ac149af57e931c9a04fe69ebb7 --- /dev/null +++ b/2024/Generative Proxemics_ A Prior for 3D Social Interaction from Images/full.md @@ -0,0 +1,296 @@ +# Generative Proxemics: A Prior for 3D Social Interaction from Images + +Lea Müller1,4 Vickie Ye1 Georgios Pavlakos2,5 Michael Black3 Angjoo Kanazawa1 +1UC Berkeley UT Austin MPI for Intelligent Systems, Tübingen + +{mueller,vye,kanazawa}@berkeley.edu,pavlakos@cs.utexas.edu,black@tuebingen.mpg.de + +![](images/c0d00b3a2ccfd2dd657e0577d5fd876ab6aded41daf4aa370a7375e745a53a7a.jpg) + +![](images/7551801e4d9bec1616e7c25563a5b0c87c38f8d7b7fcd96128c82486f122ef2d.jpg) +Figure 1. Generative Proxemics. We propose a diffusion model that learns a 3D generative model of two people in close social interaction. We show how the model can be used to generate samples or as a social prior in the downstream task of reconstructing two people in close proximity from images without any user annotation at test time. Shown here are input test images (left) and our predicted 3D bodies (right). + +![](images/2b2f6d447af0d544d5155398d0267e5d85cbf06841bda94c9e767aca4ac89f82.jpg) + +![](images/6acf03a0036e78e42b534f2144268dd15cda7fdc198e68edcef69dbc5886257f.jpg) + +![](images/ecf016e0e6c59b999bc208db9869de9c6e96af850bb0e44addfc1bf809af0425.jpg) + +# Abstract + +Social interaction is a fundamental aspect of human behavior and communication. The way individuals position themselves in relation to others, also known as proxemics, conveys social cues and affects the dynamics of social interaction. Reconstructing such interaction from images presents challenges because of mutual occlusion and the limited availability of large training datasets. To address this, we present a novel approach that learns a prior over the 3D proxemics two people in close social interaction and demonstrate its use for single-view 3D reconstruction. We start by creating 3D training data of interacting people using image datasets with contact annotations. We then model the proxemics using a novel denoising diffusion model called BUDDI that learns the joint distribution over the poses of two people in close social interaction. Sampling from our generative proxemics model produces realistic 3D human interactions, which we validate through a perceptual study. We use BUDDI in reconstructing two people in close proximity from an image without any contact annotation via an optimization approach that uses the diffusion model as a prior. Our approach recovers accurate 3D so + +cial interactions from noisy initial estimates, outperforming state-of-the-art methods. Our code, data, and model are available at: muelea.github.io/buddy. + +# 1. Introduction + +Humans are social creatures, and physical interaction plays a crucial role in our daily lives, shaping our relationships. For example, research in behavioral science has shown that a slight touch between two people can cause a more friendly behaviour towards the touch-giver and lead to increased tips in restaurants [5]. Capturing and modeling scenarios of physical social interaction through computer vision will enable advancements in augmented and virtual reality and impact other fields like robotics and behavioral science. The recent progress in single-person mesh estimation is unfortunately not sufficient to model close interaction, as these methods are not trained to reason about the relative depth of people and the intricate interplay between two people's body poses, shapes, and proximity. + +In this work, we present the first approach that learns a generative model for 3D social proxemics and demonstrate its use as data-driven prior during an optimization routine. The diffusion model is trained using 3D human poses and shapes reconstructed from a large-scale image collection [8] using contact annotation, as well as using motion-capture (MoCap) data [8, 54]. The resulting model is able to generate the 3D pose and shape parameters of pairs of interacting people. When trained on bodies recovered from images, the model learns interactions depicted in photographs, such as people standing close together, playing sports, hugging, etc., see Figure 1. We further demonstrate the effectiveness of the learned prior by applying it to the challenging task of 3D human pose and shape reconstruction from a single image containing people engaged in social interaction. + +Specifically, we propose BUDDI: a "BUDdies Diffusion Model". Diffusion models are established methods for image generation and are often used to model 3D human motion. In this work we use them to model 3D social proxemics. The majority of state-of-the-art diffusion-based methods for 3D human mesh generation operate on 3D joint locations [45]. This representation lacks information about the human body surface, which, intuitively, is important for reasoning about interpersonal contact. Our approach, in contrast, operates on the parameters of two parametric human body models, which represent the surfaces of two people closely interacting. After training, our model is able to generate samples of plausible pairs of 3D bodies in social interaction from pure noise. The model can also be conditioned in the output of a human pose and shape regressor. In this conditional case, the model takes the noisy output and generates similar poses but with realistic social interaction. + +We then demonstrate how exploit BUDDI's knowledge of human proxemics to guide 3D mesh reconstruction of people in a close social interaction from a single image. To this end, we introduce a novel optimization-based approach, which uses BUDDI as a data-driven prior. We initialize our optimization routine with samples from BUDDI, conditioned to the output of a state-of-the-art multi-person human mesh regressor [42]. We then optimize over SMPL-X pose, shape, and translation parameters to match detected 2D joint locations. We incorporate guidance from the diffusion model using a loss inspired by the Score-Distillation loss from the 3D object creation literature [35]: In each optimization step, BUDDI refines the current estimate towards a more plausible social interaction conditioned on the initial predictions. The refined pose, shape and translation serve as prior in the overall objective function. + +Our contributions include (1) presenting the first generative model of a pair of 3D people in close social interaction and (2) a novel approach for reconstructing 3D human meshes from images without relying on ground-truth contact annotations. We perform extensive experiments with + +BUDDI to evaluate its performance on the FlickrCI3D Signatures dataset [8] as well as CHI3D, and the recent Hi4D dataset [54] and find that it outperforms the state of the art as well as strong baselines. We also evaluate the unconditional samples from the diffusion model in a perceptual study, where people find our samples more realistic $44.4\%$ when compared over real samples, where $50\%$ is the upperbound where they do not see any difference. Importantly, we find that our optimization approach significantly improves the results of [42] both quantitatively and qualitatively. This work opens up a new avenue of research on digital human synthesis, laying the foundation for a deeper understanding of human social behavior derived from image data. Our data, code, and model will be available for research. + +# 2. Related Work + +Generating 3D humans. There has been recent interest in generating 3D humans, in different contexts. Several methods automatically populate static 3D scenes with 3D humans [14, 60, 61], while more recent methods generate both body and hand poses to interact with 3D objects [43, 44, 49]. Other work generates human motions conditioned on different inputs such as audio [26, 47] or text [33, 34, 45]. Concurrent work proposes text-to-3D diffusion-based approaches to generate motion of two interacting humans [27, 40]. Neither method predicts the full body surface, but rather they synthesize either 3D joint locations or SMPL pose parameters for the average body. These methods are not used as priors for reconstructing interacting people from images. + +To model 3D human proxemics probabilistically, we employ diffusion models, which achieve impressive performance on image generation tasks [7, 15, 38, 39]. They have recently been adopted in 3D human motion generation scenarios: MDM [45] generates plausible motions conditioned on text input; PhysDiff [55] incorporates physical constraints in the diffusion process to generate physically plausible motions; and EDGE [47] uses a transformer-based diffusion model for dance generation. Related work [4, 6, 28] has investigated different modalities for the conditioning, e.g., audio, text, or action classes. EgoEgo [25] generates plausible full-body motions conditioned on the head motion. SceneDiffuser [16] focuses on the scene-conditioned setting. We also rely on techniques from the diffusion literature, but consider the unique setting where two people are in close interaction and leverage this for single-image 3D reconstruction. + +Multi-person 3D human mesh estimation. An extensive line of work focuses on reconstructing the 3D human pose and shape of a single person from images using optimization [2, 11, 24, 32, 37, 46, 50] or regression approaches [1, 12, 19, 20, 22, 29, 31, 51, 58, 59]. Capitalizing + +![](images/309b64c2bc74f21d55748b73879f9648e1a0223602603cf9c9bac43a9ce72380.jpg) +Figure 2. BUDDI: BUddies Diffusion model. On the left, we illustrate the architecture of BUDDI, our diffusion model for modeling 3D social proxemics between two people in close interaction. The diffusion process is applied directly on SMPL-X body parameters. To condition BUDDI on estimated body model parameters, $c_{\mathrm{H}}$ , we concatenate the parameters along the token dimension. On the right, we illustrate the optimization method with BUDDI as prior. Our optimization takes detected keypoints [3, 52] and an initial regressor estimate [42] as input. Given the regressor estimate, we sample from BUDDI to obtain $\tilde{x}$ which we use to initialize the optimization routine. In each optimization iteration, we take a single diffuse-denoise step on the current estimate using the learned denoiser model $D$ . Our losses encourage the current estimate to be close to the refined meshes $(L_{\mathrm{diffusion}})$ and to the initial estimate and detected keypoints $(L_{\mathrm{fitting}})$ . + +![](images/f3e68861a4b5e3b31c4203da0d49cd48ce47bffb1125ffb6f0dce333107e371e.jpg) + +on these techniques, recent approaches focus explicitly on reconstructing multiple people jointly from a single image. Zanfir et al. [56] propose an optimization solution, while Jiang et al. [18] and Sun et al. [41] rely on deep networks to regress the pose and shape for all people in the image. BEV [42] extends ROMP [41] to reason about the depth of people while taking age/height into account. + +The above methods do not address contact between people. To do so, Fieraru et al. [8] introduce the first datasets with ground-truth labels for the body regions in contact between humans. Labels are collected using MoCap (CHI3D) or human annotators (FlickrCI3D Signatures). They propose an optimization approach that requires the ground-truth contact map to reconstruct people in close proximity at test time. More recently, REMIPS [9] introduces a transformer-based method that regresses the 3D pose of multiple people. REMIPS is trained using the above datasets while taking into account contact and interpenetration. In this work, we take a very different approach by learning and exploiting a 3D generative proxemics prior. We use the ground-truth contact maps to generate pseudoground truth 3D human fits from which we learn the diffusion model; once this is learned, we show that it can be used as a prior to recover plausible bodies in close proximity from images without explicit knowledge of contact maps. + +Data-driven priors in optimization. Optimization-based methods for 3D human pose and shape estimation, like SMPLify [2], are versatile and allow different data-driven prior terms to be incorporated in the objective function. Different methods have been used to learn pose priors including GMMs [2], VAEs [32], neural distance fields [46], and normalizing flows [57]. ProHMR [23] learns a pose prior conditioned on image pixels. HuMoR [37] incor + +porates a data-driven motion prior in the iterative optimization. POSA [14] learns a prior for human-scene interaction from PROX data [13] and uses it in their optimization. In contrast to these methods, we use a diffusion model to capture the joint distribution over SMPL-X parameters for two people interacting and show that we can both sample from the model and use during optimization to improve the pose estimates of interacting people. + +# 3. Method + +We introduce BUDDI, a generative model of two people in close social interaction. Because of the complexity and multimodality of the data, we turn to denoising diffusion probabilistic models [15] to address this task. In Sec. 3.1, we describe the basics of diffusion, and the parameterization we employ to model people in contact. In addition to sampling new body meshes from our model, our generative model can serve as a prior for reconstructing 3D humans from images. In Sec. 3.2, we describe an optimization procedure that incorporates BUDDI as a prior to recover two SMPL-X meshes from observed 2D keypoints. + +For all of the following, we use the SMPL-X [32] body model to represent the human bodies. SMPL-X is a differentiable function that maps pose, $\pmb{\theta} \in \mathbb{R}^{21 \times 3}$ , shape, $\beta \in \mathbb{R}^{10}$ , and expression, $\psi \in \mathbb{R}^{10}$ parameters to a mesh consisting of $N_v = 10,475$ vertices $V \in \mathbb{R}^{N_v \times 3}$ . We place the generated meshes in the world by rotating and translating them by $\phi \in \mathbb{R}^3$ and $\gamma \in \mathbb{R}^3$ . We denote person $a$ 's parameters as $X^a = [\phi^a, \pmb{\theta}^a, \pmb{\beta}^a, \gamma^a]$ and $X^b = [\phi^b, \pmb{\theta}^b, \pmb{\beta}^b, \gamma^b]$ . For simplicity, we refer to both people when no index is specified, e.g., $X$ refers $X^a$ and $X^b$ + +# 3.1. Diffusion Model for 3D Proxemics + +Denoising diffusion models are latent variable generative models that learn to transform random noise into the desired data distribution $p_{\mathrm{data}}$ through a forward and reverse process. The forward inference process is a Markov chain over $T$ steps given by transitions $q(\mathbf{x}_{t + 1}|\mathbf{x}_t)$ , which gradually adds Gaussian noise to clean samples $\mathbf{x}_0$ from the data distribution according to a fixed variance schedule $\sigma_t$ . + +The reverse process $q(\mathbf{x}_{t - 1}|\mathbf{x}_t)$ then gradually denoises noisy samples back into the data distribution. The reverse process transitions follow a Gaussian distribution when conditioned on $x_0$ , but must be inferred during the generative process. Following Ramesh et al. [36], we train a neural network $D$ that predicts a sample $\hat{\mathbf{x}}_0 = D(\mathbf{x}_t; t)$ from a noisy sample $\mathbf{x}_t$ given the noise level $t$ . For the task of reconstructing humans from images, when we have initial estimates of the SMPL-X parameters of two humans, we condition the denoising network $D$ on $\mathbf{c}_H$ , the predicted SMPL-X parameters of the two humans by a regressor. + +We refer to the process of adding noise as diffusion and the process of removing the noise via $D$ as denoising. Specifically, we diffuse a ground-truth sample $\mathbf{x}_0$ by uniformly sampling a noise level $t$ with $\epsilon_t \sim \mathcal{N}(0, \mathbf{I})$ to obtain the noisy sample $\mathbf{x}_t = \sqrt{\sigma_t'}\mathbf{x}_0 + \sqrt{1 - \sigma_t'}\epsilon_t$ with $\sigma_t' = \prod_{i=1}^{t} (1 - \sigma_t)$ . + +We then train $D$ to minimize + +$$ +\mathbb {E} _ {\mathbf {x} _ {0} \sim p _ {\text {d a t a}}} \mathbb {E} _ {t \sim \mathcal {U} \{0, T \}, \mathbf {x} _ {t} \sim q (\cdot | \mathbf {x} _ {0})} | | D (\mathbf {x} _ {t}; t, \mathbf {c} _ {H}) - \mathbf {x} _ {0} | |, \tag {1} +$$ + +where we set $\mathbf{c}_H = \varnothing$ for $20\%$ of conditional model training, and all of unconditional model training. + +Architecture. Because we aim to model close contact between people, we choose a model state space that can express the full surface of the human body. Specifically, in contrast to prior work in human motion diffusion that operate only on joint angles and locations [45, 55], we directly operate on the full SMPL-X parameters of the two people. A sample $\mathbf{x}$ thus corresponds to the concatenation of two bodies: + +$$ +\mathbf {x} = \left[ X ^ {a}, X ^ {b} \right] = \left[ \phi^ {a}, \boldsymbol {\theta} ^ {a}, \boldsymbol {\beta} ^ {a}, \gamma^ {a}, \phi^ {b}, \boldsymbol {\theta} ^ {b}, \boldsymbol {\beta} ^ {b}, \gamma^ {b} \right]. +$$ + +We denoise a sample $\mathbf{x}_t$ with a transformer encoder block on tokenized parameters. Specifically, each parameter of each person is tokenized into 152-dimensional latent vectors with per-parameter and per-person embedding layers. We tokenize the noise level $t$ similarly with a noise embedding. When conditioning is available, i.e. SMPL-X estimates for person $a$ and $b$ , we similarly tokenize the parameters to be used as additional tokens. We pass the available tokens into the transformer encoder, and similarly decode the output tokens with per-token embeddings. We illustrate the denoiser architecture in Figure 2. + +Losses. We employ standard human pose and shape regularization losses. We write our training objective as + +$$ +L _ {D} = L _ {\theta} + L _ {\beta} + L _ {\gamma} + L _ {v 2 v}, \tag {2} +$$ + +where $L_{\theta}, L_{\beta}, L_{\gamma}$ denote squared L2-losses on respective body model parameters, and $L_{v2v}$ denotes a squared L2 loss on model vertices. We use 6D rotation representations [62] for global orientation and pose, and model the relative translation between $a$ and $b$ . We show generated samples from our unconditional model in Fig. 3. + +# 3.2. Optimization with the Proxemics Prior + +Reconstructing 3D human meshes from a single image is an extremely under-constrained problem, and priors over human pose and shape are crucial in an optimization based framework for recovering plausible meshes [2, 32, 56]. Our problem involves people in close contact, which requires correctly placing the meshes in context with each other, which has only been done when given ground truth contact annotations at test time [8]. We remove the need for ground-truth contact maps by using our generative model as a prior during reconstruction with a score distillation approach [35, 48]. + +During inference, we observe detected 2D keypoints $\tilde{J}_{2D}$ and initial body model parameter estimates $\mathbf{c}_H$ from a regressor [42]. We then optimize the body parameters of two people to minimize + +$$ +L _ {\text {O p t i z i m a t i o n w . B U D D I}} = L _ {\text {f i t t i n g}} + L _ {\text {d i f f u s i o n}}. \tag {3} +$$ + +$L_{\mathrm{fitting}}$ ensures that the solution stays close to the image evidence, while $L_{\mathrm{diffusion}}$ is a data-driven prior using our conditional diffusion model. We treat this prior as similar to those used for 3D pose in previous works such as GMM [2] and V-Poser [32], but for 3D proxemics. We illustrate the optimization procedure in Fig. 2 right. + +We initialize our optimization by generating a sample $\tilde{\mathbf{x}}$ from the conditional model. We sample with DDIM sampling with 100 evenly spaced steps. We then use the data fitting loss: + +$$ +L _ {\text {f i t t i n g}} = \lambda_ {J} L _ {J} + \lambda_ {\tilde {\theta}} L _ {\tilde {\theta}} + \lambda_ {P} L _ {P}, \tag {4} +$$ + +where $L_{J}$ denotes 2D re-projection error between the reprojected 3D joints of the current estimate and the detected 2D keypoints, and $L_{\tilde{\theta}}$ is a prior for the solution to be close to the denoised initialization. + +$L_{P}$ denotes an interpenetration loss between two people that pushes inside vertices to the surface, which we use winding numbers [17] to find intersecting vertices between low-resolution SMPL-X meshes of the current estimates: + +$$ +L _ {P} = \sum_ {v \in V _ {I} ^ {a}} \min _ {u \in V ^ {b}} \| v - u \| ^ {2} + \sum_ {v \in V _ {I} ^ {b}} \min _ {u \in V ^ {a}} \| v - u \| ^ {2}, \tag {5} +$$ + +![](images/67b8969f2bc4b92fc703e2a3e2e368be28f9a12b7b2b08cc057d7f18397191d6.jpg) +Figure 3. Generative Proxemics: Samples from BUDDI. All samples are unconditionally generated from pure noise using the trained diffusion model. We select several representative examples and show two views per sample. These samples reveal that BUDDI has learned the distribution of people in close contact including embracing each other, playing sports, sitting side by side, and taking photographs. + +where $V_{I}^{a}$ denotes vertices of $M^a$ intersecting the lowresolution mesh of $M^b$ ; and vice versa for $V_{I}^{b}$ . + +To use the prior on human interaction into account, we use the learned denoising model $D$ from BUDDI and perform a single diffuse-denoise step, with a noise level at $t = 10$ , on the current estimate. The denoised estimate, $\hat{\mathbf{x}}_0 = D(\mathbf{x}_t; t, \mathbf{c}_H)$ , regularizes the current estimate via + +$$ +L _ {\text {d i f f u s i o n}} = \left\| D \left(\mathbf {x} _ {t}; t, \mathbf {c} _ {H}\right) - \mathbf {x} \right\|, \tag {6} +$$ + +where $\mathbf{x}_t = \sqrt{\sigma_t^{\prime}}\mathbf{x}_{\mathrm{no - grad}} + \sqrt{1 - \sigma_t^{\prime}}\epsilon_t$ denotes the diffused body model parameters of the current estimate, and $\mathbf{x}_{\mathrm{no - grad}}$ denotes the current estimate with detached gradients. $\hat{\mathbf{x}}_0$ , and encourages $\mathbf{x}$ to be close to $\hat{\mathbf{x}}_0$ . In practice, we penalize the decoded parameters of $\mathbf{x}$ and $\hat{\mathbf{x}}_0$ directly as + +$$ +\begin{array}{l} L _ {\text {d i f f u s i o n}} = \lambda_ {\hat {\phi}} | | \hat {\phi} _ {0} - \phi | | + \lambda_ {\hat {\theta}} | | \hat {\theta} _ {0} - \theta | | \tag {7} \\ + \lambda_ {\hat {\boldsymbol {\beta}}} | | \hat {\boldsymbol {\beta}} _ {0} - \boldsymbol {\beta} | | + \lambda_ {\hat {\boldsymbol {\gamma}}} | | \hat {\boldsymbol {\gamma}} _ {0} - \boldsymbol {\gamma} | |. \\ \end{array} +$$ + +Intuitively, this loss uses the learned denoiser $D$ to take a step from the current estimate towards the data distribution of two people in close proximity, conditioned on the regressor prediction. + +# 4. Implementation Details + +Training Data. There are few datasets containing 3D ground truth of humans in close social interaction [8, 54]. Such datasets are usually captured in lab environments, consequently they are small and do not contain the variety of interactions between humans "in the wild," e.g. when playing sports or taking social pictures. To address this lack of data, we create Flickr Fits, i.e. SMPL-X fits + +for Flickr images portraying humans in contact scenarios. For this, we use FlickrCI3D Signatures [8], a dataset of images showing interacting humans collected from Flickr with discrete human-human contact annotations. Specifically, the SMPL-X body surface is divided into $R = 75$ regions such that each region, $r$ , roughly covers a similar area. For a given photo, the human annotators assign a binary label indicating contact between a region on one person and a region on the other. For two meshes, $M^a$ and $M^b$ , the annotation can be represented as a binary contact map $\mathcal{C}^D \in \{0,1\}^{R \times R}$ , where + +$$ +\mathcal {C} _ {i j} ^ {D} = \left\{ \begin{array}{l l} 1, & \text {i f} r _ {i} \text {o f} M ^ {a} \text {i s i n c o n t a c t w i t h} r _ {j} \text {o f} M ^ {b} \\ 0, & \text {o t h e r w i s e .} \end{array} \right. \tag {8} +$$ + +We use these ground-truth contact maps in an optimization routine for fitting two people to detected keypoints, similar to Sec. 3.2 but replace the diffusion model prior with standard image fitting priors. The dataset contains 10,631/1,139 train/test images, with one image containing multiple contact annotations. Note that we use pairs of people in contact as labeled in the FlickrCI3D Signatures dataset [8] and only use images containing matching regressor estimates, 2D keypoints, and contact labels. See Fig. 4 for qualitative examples and the Sup. Mat. for more details about the selection process and an evaluation of the optimization method with ground-truth contact labels. + +We also augment our training data with available MoCap data, which is considerably smaller than those obtained from image fits: CHI3D [8] contains 3/2 pairs of training/test subjects performing 127 sequences of two-person interactions like hugs or kicks with ground-truth SMPL-X bodies. One frame per sequence has contact map annota + +![](images/f72b40b5e6a3ff82197ae5ea6209b1143354e3b1afb0147d938623591cec6c62.jpg) +Figure 4. Flickr Fits. We visualize the output of the optimization process that reconstructs two people in close proximity using ground-truth contact maps, shown from four different views. We use these 3D fits as training data for BUDDI. + +tions. We use only the contact frame of the sequences from two subject pairs, resulting in 247 mesh pairs for training, and the third pair for evaluation. Hi4D [54] contains sequences of 20 pairs of people interacting with each other. The interactions include actions like hugging, dancing, and fighting. We randomly split the data into $14/3/3$ pairs for train/val/test and use every fifth frame of the subsequence involving contact as labeled in Hi4D, resulting in about 1K mesh pairs for training. The body representation format in Hi4D is SMPL, which we transfer to SMPL-X using the SMPL-X code repository [32]. Please see the Sup. Mat. for more details of the datasets. Note that while we use SMPL-X model, BUDDI is not trained on hands because none of these datasets contain hand poses. + +BUDDI Training. BUDDI is trained with meshes from FlickrCI3D Signatures Fits, CHI3D, and Hi4D. We use $60\%$ Flickr, $20\%$ CHI3D, and $20\%$ Hi4D data distribution per batch with batch size 512. The transformer backbone has six layers and eight heads; we use $10\%$ dropout and randomly shuffle the order of people during training. To train BUDDI, we randomly sample noise levels $t$ up to 1000 using a cosine noise schedule [30]. We use the Adam optimizer [21] with learning rate $10^{-4}$ . We train two separate networks: an unconditional model for generation and the conditional version for reconstruction. For the conditional model, we use all camera views of the MoCap datasets, i.e. 4/8 cameras for CHI3D/Hi4D. The unconditional model is trained on 3D MoCap fits in the world coordinate system. To sample new poses, we use DDIM sampling starting at noise levels $t = 1000$ in steps of 10. + +Optimization Details. During optimization, we experiment with different noise levels, between 10 and 100, and find that $t = 10$ does not disturb the inputs too much, but enough for $D$ to generate new configurations. We use detected 2D keypoints from OpenPose [3] and ViTpose [52] and BEV [42] estimates as conditioning. Unlike single-person mesh regressors, BEV is designed to predict multiple + +people including their relative depth. Please see Sup. Mat. for more details. + +# 5. Experiments + +Baselines. We compare our reconstruction method with BEV [42], which is also used as an input to our conditional model. Since there is no other available work that reasons about people in close social interaction, we experiment with simple but effective baselines. We train the transformer model of BUDDI to directly predict SMPL-X parameters of people in contact from BEV input, essentially a deterministic, single-step ablation of BUDDI. We also evaluate the direct conditional denoised output of BEV by BUDDI without any optimization. As another baseline, we propose an optimization routine that replaces $L_{\text{diffusion}}$ with a simple heuristic that takes the minimal distances between two meshes predicted by BEV and minimizes their distance during optimization along with the other energy terms. Finally, to compare the generation ability we train a VAE which we also use during the optimization routine in a similar manner to VPoser [32] but for two people by optimizing the VAE latent space instead of SMPL-X parameters. We refer to these models as Transformer, BUDDI (gen.), Contact Heuristic, and VAE, respectively. All baselines are trained on the same datasets as BUDDI with the same sampling strategies. Details about our baselines are provided in the Sup. Mat.. + +Metrics. We use standard evaluation metrics from the human pose and shape estimation literature. We also report the joint PA-MPJPE computed by performing Procrustes alignment of both people together. In addition to per-person metrics, this captures the relative orientation and translation of the two people. Since our method directly estimates 3D humans we propose a new metric similar to PCK [53] from the 2D pose literature called PCC, the percentage of correct contact points with respect to a radius $r$ . Specifically, given two meshes, $M^a / M^b$ and a contact map $\mathcal{C}^D$ we compute the pairwise vertex-to-vertex Euclidean distances $d_{\mathrm{eucl}}(\mathcal{C}^D)$ between annotated contact regions and consider the pair to be correct when $\min(d_{\mathrm{eucl}}(\mathcal{C}^D)) < r$ . + +# 5.1. Unconditional Generation + +We qualitatively evaluate BUDDI by showing samples generated from the model in Fig. 1 and 3. Our approach is able to generate people in close proximity including embraces, handshakes, having a conversation, sitting side by side, and, in general, plausibly interacting with each other. Since it is trained on Internet image collections, it also learns to generate people posing for photographs or playing sports. + +We further run a perceptual study to evaluate the realism of the generated social interactions against other methods. In a forced choice study, we compare our generated samples with samples from the real data distribution according to the 60/20/20 per-batch ratio for Flickr/CHI3D/Hi4D + +![](images/b6a5bf1fac02cd6bc4f974a89c2775a5b3fdac8ac3486b88bd98b1cc84f37845.jpg) +Figure 5. Automatic reconstruction of people in close social interaction. We show qualitative results from a) BEV, b) contact heuristics, which takes the BEV output and encourages the closest parts to be in contact, and c) our method, which optimizes the BEV estimates against the image evidence with the BUDDI prior. Our approach recovers a plausible reconstruction with subtle details. + +
JOINT ↓ +PA-MPJPEPCC at radius ↑
510152025
BEV106-----
Transformer861440607382
BUDDI (gen.)921539587180
Heuristic681434496170
VAE1011128425565
BUDDI661944627381
+ +Table 1. 3D Pose Evaluation on FlickrCI3D Signatures. We evaluate methods against the Flickr fits using their joint (two-person) PA-MPJPE expressed in mm. We also evaluate the percentage of correct contact points (PCC) for radius $r$ mm. + +used during training. We also compare BUDDI against generations from the VAE and a non-parametric random baseline that samples meshes from the pseudo-ground truth after centering the two people. We do a forced choice comparison between BUDDI and these three other methods, asking workers on Amazon Mechanical Turk to choose the sample that shows a more realistic close social interaction. We use 256 samples per method. We collect ratings for 768 + +pairwise comparisons. In this study, BUDDI was chosen over random in $71.23\%$ of the comparisons, over the VAE in $60.17\%$ , and over the training data in $44.4\%$ . Note that $50\%$ is the upper bound for such forced choice comparisons, in which participants cannot tell the difference between real and generated samples. + +For a quantitative evaluation, we compute the FID score between samples from BUDDI and samples from the $VAE$ on concatenated SMPL-X parameters. We sample 8K examples per method and from our training data following the dataset ratio per batch. BUDDI has a lower FID score (1.6) compared to the $VAE$ (3.3). + +# 5.2. Fitting with BUDDI + +We show qualitative results in Fig. 5 comparing BUDDI against BEV and the Contact Heuristic. Our approach is able to generate various types of human interactions with plausible contact and depth placement. It is also able to capture close interaction between a child and a parent. Although the Contact Heuristic (center) is able to move two + +
PER PERSON ↓ PA-MPJPEJOINT ↓ PA-MPJPEJOINT PA-MPJPE
backhugbasketballcheersdancefighthighfivehugkissposesidehugtalk
BEV78 / 84136200126109135121106163139142131118
Heuristic67 / 7112116883941319468159159118113109
BUDDI (F, C)70 / 77115200949212810810013311410410791
Transformer79 / 85120161141103138123128117106120105100
BUDDI (gen.)82 / 901171521391201371309610197115102101
VAE80 / 8213817513311414111987176162135140113
BUDDI70 / 769812795921131097210585889681
+ +Table 2. Evaluation of BUDDI on Hi4D. We compare the output of BUDDI to the proposed baseline methods on the Hi4D challenge. The first block shows methods that do not use Hi4D data during training or are optimization based without access to priors trained on Hi4D. BUDDI (F,C) in particular, is our model BUDDI trained on Flickr and CHI3D data only. All errors are reported in mm for 3D Joints. + +
PER PERSON ↓ +PA-MPJPEJOINT ↓ +PA-MPJPE
BEV50 5296
Transformer54 56105
BUDDI (gen.)53 5380
Heuristic49 46105
VAE54 54103
BUDDI48 4768
+ +Table 3. Quantitative Evaluation on CHI3D. We compare the output of our model to the baselines on CHI3D (pair s03). All errors reported in mm for 3D Joints. + +people closer together, which helps with image alignment, upon close observation it is not able to capture the subtle interaction between people that happens during intimate interaction. BUDDI's estimates are more realistic and better capture the subtle details of interaction. We provide additional qualitative examples in the Sup. Mat. + +We further report the percentage of correct contact (PCC) with respect to the ground truth contact map on the FlickrCI3D Signatures test set in Table 1. The table also shows the pose reconstruction accuracy against our Flickr Fits. All metrics show improvement over BEV, in particular the joint PA-MPJPE. Non-optimization methods, i.e. Transformer and BUDDI (gen.), are able to predict plausible contacts, with similar PCC accuracy to BUDDI, but struggle to reconstruct the data with a worse joint PA-MPJPE. The Heuristic, in contrast, achieves a lower reconstruction error, but worse PCC. Our approach which leverages the learned prior during optimization can recover both the relative positions and contacts between the two people. To provide insights into the performance of single-person mesh regressors when evaluated on the two-person reconstruction task, we run 4D Humans [10] on Flickr Fits. The joint PA-MPJPE is $344\mathrm{mm}$ which is high, as expected, since these methods are not trained to reason about proximity. + +We further evaluate our model against ground truth MoCap data in Tables 2 and 3. Optimization with BUDDI consistently improves the two-person reconstruction error over BEV and other baselines. When evaluated per action, the strongest improvements over BEV come from complex + +close social interactions like hugging or kissing, at $58\mathrm{mm}$ and $54\mathrm{mm}$ absolute improvement over BUDDI respectively. The Heuristic baseline achieves a low PA-MPJPE reconstruction error on all three datasets in particular for poses with a few physical contact points, such as a handshake, whereas more complex contact, such as a hug, requires data-driven priors like BUDDI; we provide an analysis in the Sup. Mat. to quantify this hypothesis. Transformer and BUDDI (gen.) have lower joint PA-MPJPE errors than BEV and the Heuristic, but worse per-person reconstruction errors. The VAE results suggest that directly operating in the latent space of a generative model is challenging and not sufficient to accurately recover close social interactions. BUDDI, in contrast, is able to model a wide variety of poses, as supported by the numerical results. + +# 6. Conclusion + +We propose BUDDI, a diffusion model for close human-human interaction. We train BUDDI from 3D fits obtained from a large-scale dataset of images with ground truth contact annotations as well as a small set of available mocap data. BUDDI enables unconditional sampling of people in close social interaction. More importantly, we also demonstrate how BUDDI can be used as effective prior for single-view 3D reconstruction of two people in close proximity. + +Our core contribution of a generative proxemics prior provides the foundation for future work on modeling and capturing human interaction. For example, future work could iteratively apply our method to reconstruct more than two people in close proximity, explore other conditioning modalities like pixel features, text, action labels, or, by taking the outputs of recent single-person mesh regressors into account, address reconstructing fine-grained interaction including finger pose and facial expressions. + +Acknowledgements: We thank A. Holynski, E. Weber, F. Warburg, B. Peebles, J. Rajasegaran, K. Mangalam and N. Athanasiou for insightful discussions, and A. Cseke, T. McConnell and T. Alexiadis for running the user study. + +Disclosure: https://files.is.tue.mpg.de/black/CoI_CVPR_2024.txt + +# References + +[1] Anurag Arnab, Carl Doersch, and Andrew Zisserman. Exploiting temporal context for 3D human pose estimation in the wild. In Computer Vision and Pattern Recognition (CVPR), 2019. +[2] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J. Black. Keep it SMPL: Automatic estimation of 3D human pose and shape from a single image. In European Conference on Computer Vision (ECCV), pages 561-578. Springer International Publishing, 2016. +[3] Zhe Cao, Gines Hidalgo Martinez, Tomas Simon, Shih-En Wei, and Yaser Sheikh. OpenPose: Realtime multi-person 2D pose estimation using part affinity fields. Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 43 (1):172-186, 2019. +[4] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18000-18010, 2023. +[5] April H Crusco and Christopher G Wetzel. The midas touch: The effects of interpersonal touch on restaurant tipping. *Personality and Social Psychology Bulletin*, 10(4): 512-517, 1984. +[6] Rishabh Dabral, Muhammad Hamza Mughal, Vladislav Golyanik, and Christian Theobalt. Mofusion: A framework for denoising-diffusion-based motion synthesis. In Computer Vision and Pattern Recognition (CVPR), 2023. +[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat GANs on image synthesis. Conference on Neural Information Processing Systems (NeurIPS), 34:8780-8794, 2021. +[8] Mihai Fieraru, Mihai Zanfir, Elisabeta Oneata, Alin-Ionut Popa, Vlad Olaru, and Cristian Sminchisescu. Three-dimensional reconstruction of human interactions. In Computer Vision and Pattern Recognition (CVPR), pages 7214-7223, 2020. +[9] Mihai Fieraru, Mihai Zanfir, Teodor Szente, Eduard Bazavan, Vlad Olaru, and Cristian Sminchisescu. REMIPS: Physically consistent 3D reconstruction of multiple interacting people under weak supervision. In Conference on Neural Information Processing Systems (NeurIPS), pages 19385-19397, 2021. +[10] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa*, and Jitendra Malik*. Humans in 4D: Reconstructing and tracking humans with transformers. In International Conference on Computer Vision (ICCV), 2023. +[11] Peng Guan, Alexander Weiss, Alexandru Balan, and Michael J. Black. Estimating human shape and pose from a single image. In International Conference on Computer Vision (ICCV), pages 1381-1388, 2009. +[12] Riza Alp Guler and Iasonas Kokkinos. HoloPose: Holistic 3D human reconstruction in-the-wild. In Computer Vision and Pattern Recognition (CVPR), 2019. +[13] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J Black. Resolving 3D human pose ambigui + +ties with 3D scene constraints. In International Conference on Computer Vision (ICCV), pages 2282-2292, 2019. +[14] Mohamed Hassan, Partha Ghosh, Joachim Tesch, Dimitrios Tzionas, and Michael J Black. Populating 3D scenes by learning human-scene interaction. In Computer Vision and Pattern Recognition (CVPR), pages 14708-14718, 2021. +[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Conference on Neural Information Processing Systems (NeurIPS), pages 6840-6851, 2020. +[16] Siyuan Huang, Zan Wang, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, and Song-Chun Zhu. Diffusion-based generation, optimization, and planning in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. +[17] Alec Jacobson, Ladislav Kavan, and Olga Sorkine-Hornung. Robust inside-outside segmentation using generalized winding numbers. ACM Transactions on Graphics (TOG), 32(4): 1-12, 2013. +[18] Wen Jiang, Nikos Kolotouros, Georgios Pavlakos, Xiaowei Zhou, and Kostas Daniilidis. Coherent reconstruction of multiple humans from a single image. In Computer Vision and Pattern Recognition (CVPR), pages 5578-5587, 2020. +[19] Hanbyul Joo, Natalia Neverova, and Andrea Vedaldi. Exemplar fine-tuning for 3D human model fitting towards in-the-wild 3D human pose estimation. In International Conference on 3D Vision (3DV), 2021. +[20] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In Computer Vision and Pattern Recognition (CVPR), pages 7122-7131, 2018. +[21] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. +[22] Nikos Kolotouros, Georgios Pavlakos, Michael J. Black, and Kostas Daniilidis. Learning to reconstruct 3D human pose and shape via model-fitting in the loop. In International Conference on Computer Vision (ICCV), pages 2252-2261, 2019. +[23] Nikos Kolotouros, Georgios Pavlakos, Dinesh Jayaraman, and Kostas Daniilidis. Probabilistic modeling for human mesh recovery. In International Conference on Computer Vision (ICCV), 2021. +[24] Christoph Lassner, Javier Romero, Martin Kiefel, Federica Bogo, Michael J. Black, and Peter V. Gehler. Unite the people: Closing the loop between 3D and 2D human representations. In Computer Vision and Pattern Recognition (CVPR), pages 6050-6059, 2017. +[25] Jiaman Li, Karen Liu, and Jiajun Wu. Ego-body pose estimation via ego-head pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17142-17151, 2023. +[26] Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. AI choreographer: Music conditioned 3D dance generation with AIST++. In International Conference on Computer Vision (ICCV), pages 13401-13412, 2021. +[27] Han Liang, Wenqian Zhang, Wenxuan Li, Jingyi Yu, and Lan Xu. Intergen: Diffusion-based multi-human motion + +generation under complex interactions. arXiv preprint arXiv:2304.05684, 2023. +[28] Jianxin Ma, Shuai Bai, and Chang Zhou. Pretrained diffusion models for unified human motion synthesis. arXiv preprint arXiv:2212.02837, 2022. +[29] Lea Müller, Ahmed A. A. Osman, Siyu Tang, Chun-Hao P. Huang, and Michael J. Black. On self-contact and human pose. In Computer Vision and Pattern Recognition (CVPR), pages 9990-9999, 2021. +[30] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning (ICML), pages 8162-8171, 2021. +[31] Mohamed Omran, Christoph Lassner, Gerard Pons-Moll, Peter V. Gehler, and Bernt Schiele. Neural body fitting: Unifying deep learning and model based human pose and shape estimation. In International Conference on 3D Vision (3DV), pages 484-494, 2018. +[32] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. +[33] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3D human motion synthesis with transformer VAE. In Computer Vision and Pattern Recognition (CVPR), pages 10985-10995, 2021. +[34] Mathis Petrovich, Michael J Black, and Gül Varol. TEMOS: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision (ECCV), pages 480-497, 2022. +[35] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv, 2022. +[36] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents, 2022. +[37] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J Guibas. HuMoR: 3D human motion model for robust pose estimation. In International Conference on Computer Vision (ICCV), 2021. +[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. +[39] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. +[40] Yonatan Shafir, Guy Tevet, Roy Kapon, and Amit H Bermano. Human motion diffusion as a generative prior. arXiv preprint arXiv:2303.01418, 2023. +[41] Yu Sun, Qian Bao, Wu Liu, Yili Fu, Michael J Black, and Tao Mei. Monocular, one-stage, regression of multiple 3D people. In Computer Vision and Pattern Recognition (CVPR), pages 11179-11188, 2021. + +[42] Yu Sun, Wu Liu, Qian Bao, Yili Fu, Tao Mei, and Michael J Black. Putting people in their place: Monocular regression of 3d people in depth. In Computer Vision and Pattern Recognition (CVPR), pages 13243-13252, 2022. +[43] Omid Taheri, Vasileios Choutas, Michael J. Black, and Dimitrios Tzionas. GOAL: Generating 4D whole-body motion for hand-object grasping. In Computer Vision and Pattern Recognition (CVPR), pages 13263-13273, 2022. +[44] Purva Tendulkar, Dídac Surís, and Carl Vondrick. FLEX: Full-body grasping without full-body grasps. In Computer Vision and Pattern Recognition (CVPR), 2023. +[45] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. +[46] Garvita Tiwari, Dimitrije Antic, Jan Eric Lenssen, Nikolaos Sarafianos, Tony Tung, and Gerard Pons-Moll. Pose-ndf: Modeling human pose manifolds with neural distance fields. In European Conference on Computer Vision (ECCV), 2022. +[47] Jonathan Tseng, Rodrigo Castellon, and C Karen Liu. EDGE: Editable dance generation from music. Computer Vision and Pattern Recognition (CVPR), 2023. +[48] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12619-12629, 2023. +[49] Yan Wu, Jiahao Wang, Yan Zhang, Siwei Zhang, Otmar Hilliges, Fisher Yu, and Siyu Tang. Saga: Stochastic whole-body grasping with contact. In European Conference on Computer Vision (ECCV), pages 257-274, 2022. +[50] Hongyi Xu, Eduard Gabriel Bazavan, Andrei Zanfir, William T. Freeman, Rahul Sukthankar, and Cristian Sminchisescu. GHUM & GHUML: Generative 3D human shape and articulated pose models. In Computer Vision and Pattern Recognition (CVPR), pages 6183-6192, 2020. +[51] Yuanlu Xu, Song-Chun Zhu, and Tony Tung. DenseRaC: Joint 3D pose and shape estimation by dense render-and compare. In International Conference on Computer Vision (ICCV), 2019. +[52] Yufei Xu, Jing Zhang, Qiming Zhang, and Dacheng Tao. Vitpose: Simple vision transformer baselines for human pose estimation. Advances in Neural Information Processing Systems, 35:38571-38584, 2022. +[53] Yi Yang and Deva Ramanan. Articulated human detection with flexible mixtures of parts. IEEE transactions on pattern analysis and machine intelligence, 35(12):2878-2890, 2012. +[54] Yifei Yin, Chen Guo, Manuel Kaufmann, Juan Zarate, Jie Song, and Otmar Hilliges. Hi4d: 4d instance segmentation of close human interaction. In Computer Vision and Pattern Recognition (CVPR), 2023. +[55] Ye Yuan, Jiaming Song, Umar Iqbal, Arash Vahdat, and Jan Kautz. Physdiff: Physics-guided human motion diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16010-16021, 2023. +[56] Andrei Zanfir, Elisabeta Maroiniu, and Cristian Sminchisescu. Monocular 3d pose and shape estimation of multiple people in natural scenes-the importance of multiple scene + +constraints. In Computer Vision and Pattern Recognition (CVPR), pages 2148-2157, 2018. +[57] Andrei Zanfir, Eduard Gabriel Bazavan, Hongyi Xu, William T Freeman, Rahul Sukthankar, and Cristian Sminchisescu. Weakly supervised 3D human pose and shape reconstruction with normalizing flows. In European Conference on Computer Vision (ECCV), pages 465-481, 2020. +[58] Andrei Zanfir, Eduard Gabriel Bazavan, Mihai Zanfir, William T Freeman, Rahul Sukthankar, and Cristian Sminchisescu. Neural descent for visual 3D human pose and shape. In Computer Vision and Pattern Recognition (CVPR), pages 14484-14493, 2021. +[59] Hongwen Zhang, Yating Tian, Xinchi Zhou, Wanli Ouyang, Yebin Liu, Limin Wang, and Zhenan Sun. PyMAF: 3D human pose and shape regression with pyramidal mesh alignment feedback loop. In International Conference on Computer Vision (ICCV), pages 11446-11456, 2021. +[60] Siwei Zhang, Yan Zhang, Qianli Ma, Michael J Black, and Siyu Tang. PLACE: Proximity learning of articulation and contact in 3D environments. In International Conference on 3D Vision (3DV), pages 642-651, 2020. +[61] Yan Zhang, Mohamed Hassan, Heiko Neumann, Michael J Black, and Siyu Tang. Generating 3D people in scenes without people. In Computer Vision and Pattern Recognition (CVPR), pages 6194-6204, 2020. +[62] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Computer Vision and Pattern Recognition (CVPR), pages 5745-5753, 2019. \ No newline at end of file diff --git a/2024/Generative Proxemics_ A Prior for 3D Social Interaction from Images/images.zip b/2024/Generative Proxemics_ A Prior for 3D Social Interaction from Images/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d4a87cc50fecf054f2e7a5e19ef46fde9c55af3c --- /dev/null +++ b/2024/Generative Proxemics_ A Prior for 3D Social Interaction from Images/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8e75755c4fb1210b41f30d00925eb26f5bf718d0c0f44e1f54cee0eb9645fd57 +size 604121 diff --git a/2024/Generative Proxemics_ A Prior for 3D Social Interaction from Images/layout.json b/2024/Generative Proxemics_ A Prior for 3D Social Interaction from Images/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..21697948cb2e7fdb9be8e54c4cf4594afb983371 --- /dev/null +++ b/2024/Generative Proxemics_ A Prior for 3D Social Interaction from Images/layout.json @@ -0,0 +1,8540 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 82, + 103, + 511, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 103, + 511, + 120 + ], + "spans": [ + { + "bbox": [ + 82, + 103, + 511, + 120 + ], + "type": "text", + "content": "Generative Proxemics: A Prior for 3D Social Interaction from Images" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 73, + 142, + 518, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 142, + 518, + 171 + ], + "spans": [ + { + "bbox": [ + 73, + 142, + 518, + 171 + ], + "type": "text", + "content": "Lea Müller1,4 Vickie Ye1 Georgios Pavlakos2,5 Michael Black3 Angjoo Kanazawa1 \n1UC Berkeley UT Austin MPI for Intelligent Systems, Tübingen" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 73, + 174, + 520, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 174, + 520, + 185 + ], + "spans": [ + { + "bbox": [ + 73, + 174, + 520, + 185 + ], + "type": "text", + "content": "{mueller,vye,kanazawa}@berkeley.edu,pavlakos@cs.utexas.edu,black@tuebingen.mpg.de" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 52, + 213, + 184, + 302 + ], + "blocks": [ + { + "bbox": [ + 52, + 213, + 184, + 302 + ], + "lines": [ + { + "bbox": [ + 52, + 213, + 184, + 302 + ], + "spans": [ + { + "bbox": [ + 52, + 213, + 184, + 302 + ], + "type": "image", + "image_path": "c0d00b3a2ccfd2dd657e0577d5fd876ab6aded41daf4aa370a7375e745a53a7a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 52, + 304, + 182, + 387 + ], + "blocks": [ + { + "bbox": [ + 52, + 304, + 182, + 387 + ], + "lines": [ + { + "bbox": [ + 52, + 304, + 182, + 387 + ], + "spans": [ + { + "bbox": [ + 52, + 304, + 182, + 387 + ], + "type": "image", + "image_path": "7551801e4d9bec1616e7c25563a5b0c87c38f8d7b7fcd96128c82486f122ef2d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 396, + 545, + 430 + ], + "lines": [ + { + "bbox": [ + 46, + 396, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 545, + 430 + ], + "type": "text", + "content": "Figure 1. Generative Proxemics. We propose a diffusion model that learns a 3D generative model of two people in close social interaction. We show how the model can be used to generate samples or as a social prior in the downstream task of reconstructing two people in close proximity from images without any user annotation at test time. Shown here are input test images (left) and our predicted 3D bodies (right)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 187, + 213, + 317, + 302 + ], + "blocks": [ + { + "bbox": [ + 187, + 213, + 317, + 302 + ], + "lines": [ + { + "bbox": [ + 187, + 213, + 317, + 302 + ], + "spans": [ + { + "bbox": [ + 187, + 213, + 317, + 302 + ], + "type": "image", + "image_path": "2b2f6d447af0d544d5155398d0267e5d85cbf06841bda94c9e767aca4ac89f82.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 187, + 304, + 317, + 387 + ], + "blocks": [ + { + "bbox": [ + 187, + 304, + 317, + 387 + ], + "lines": [ + { + "bbox": [ + 187, + 304, + 317, + 387 + ], + "spans": [ + { + "bbox": [ + 187, + 304, + 317, + 387 + ], + "type": "image", + "image_path": "6acf03a0036e78e42b534f2144268dd15cda7fdc198e68edcef69dbc5886257f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 338, + 213, + 529, + 388 + ], + "blocks": [ + { + "bbox": [ + 338, + 213, + 529, + 388 + ], + "lines": [ + { + "bbox": [ + 338, + 213, + 529, + 388 + ], + "spans": [ + { + "bbox": [ + 338, + 213, + 529, + 388 + ], + "type": "image", + "image_path": "ecf016e0e6c59b999bc208db9869de9c6e96af850bb0e44addfc1bf809af0425.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 143, + 439, + 192, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 439, + 192, + 451 + ], + "spans": [ + { + "bbox": [ + 143, + 439, + 192, + 451 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 462, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 289, + 714 + ], + "type": "text", + "content": "Social interaction is a fundamental aspect of human behavior and communication. The way individuals position themselves in relation to others, also known as proxemics, conveys social cues and affects the dynamics of social interaction. Reconstructing such interaction from images presents challenges because of mutual occlusion and the limited availability of large training datasets. To address this, we present a novel approach that learns a prior over the 3D proxemics two people in close social interaction and demonstrate its use for single-view 3D reconstruction. We start by creating 3D training data of interacting people using image datasets with contact annotations. We then model the proxemics using a novel denoising diffusion model called BUDDI that learns the joint distribution over the poses of two people in close social interaction. Sampling from our generative proxemics model produces realistic 3D human interactions, which we validate through a perceptual study. We use BUDDI in reconstructing two people in close proximity from an image without any contact annotation via an optimization approach that uses the diffusion model as a prior. Our approach recovers accurate 3D so" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 441, + 545, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 441, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 306, + 441, + 545, + 475 + ], + "type": "text", + "content": "cial interactions from noisy initial estimates, outperforming state-of-the-art methods. Our code, data, and model are available at: muelea.github.io/buddy." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 498, + 386, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 498, + 386, + 510 + ], + "spans": [ + { + "bbox": [ + 306, + 498, + 386, + 510 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 518, + 545, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 518, + 545, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 518, + 545, + 686 + ], + "type": "text", + "content": "Humans are social creatures, and physical interaction plays a crucial role in our daily lives, shaping our relationships. For example, research in behavioral science has shown that a slight touch between two people can cause a more friendly behaviour towards the touch-giver and lead to increased tips in restaurants [5]. Capturing and modeling scenarios of physical social interaction through computer vision will enable advancements in augmented and virtual reality and impact other fields like robotics and behavioral science. The recent progress in single-person mesh estimation is unfortunately not sufficient to model close interaction, as these methods are not trained to reason about the relative depth of people and the intricate interplay between two people's body poses, shapes, and proximity." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 315, + 693, + 543, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 693, + 543, + 703 + ], + "spans": [ + { + "bbox": [ + 315, + 693, + 543, + 703 + ], + "type": "text", + "content": "4Portions of this work were done while LM was at MPI-IS, Tübingen." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 703, + 487, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 703, + 487, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 703, + 487, + 712 + ], + "type": "text", + "content": "5This work was done while GP was at UC Berkeley." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9687" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 49, + 72, + 286, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 72, + 286, + 251 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 286, + 251 + ], + "type": "text", + "content": "In this work, we present the first approach that learns a generative model for 3D social proxemics and demonstrate its use as data-driven prior during an optimization routine. The diffusion model is trained using 3D human poses and shapes reconstructed from a large-scale image collection [8] using contact annotation, as well as using motion-capture (MoCap) data [8, 54]. The resulting model is able to generate the 3D pose and shape parameters of pairs of interacting people. When trained on bodies recovered from images, the model learns interactions depicted in photographs, such as people standing close together, playing sports, hugging, etc., see Figure 1. We further demonstrate the effectiveness of the learned prior by applying it to the challenging task of 3D human pose and shape reconstruction from a single image containing people engaged in social interaction." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 254, + 286, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 254, + 286, + 456 + ], + "spans": [ + { + "bbox": [ + 49, + 254, + 286, + 456 + ], + "type": "text", + "content": "Specifically, we propose BUDDI: a \"BUDdies Diffusion Model\". Diffusion models are established methods for image generation and are often used to model 3D human motion. In this work we use them to model 3D social proxemics. The majority of state-of-the-art diffusion-based methods for 3D human mesh generation operate on 3D joint locations [45]. This representation lacks information about the human body surface, which, intuitively, is important for reasoning about interpersonal contact. Our approach, in contrast, operates on the parameters of two parametric human body models, which represent the surfaces of two people closely interacting. After training, our model is able to generate samples of plausible pairs of 3D bodies in social interaction from pure noise. The model can also be conditioned in the output of a human pose and shape regressor. In this conditional case, the model takes the noisy output and generates similar poses but with realistic social interaction." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 460, + 286, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 460, + 286, + 651 + ], + "spans": [ + { + "bbox": [ + 49, + 460, + 286, + 651 + ], + "type": "text", + "content": "We then demonstrate how exploit BUDDI's knowledge of human proxemics to guide 3D mesh reconstruction of people in a close social interaction from a single image. To this end, we introduce a novel optimization-based approach, which uses BUDDI as a data-driven prior. We initialize our optimization routine with samples from BUDDI, conditioned to the output of a state-of-the-art multi-person human mesh regressor [42]. We then optimize over SMPL-X pose, shape, and translation parameters to match detected 2D joint locations. We incorporate guidance from the diffusion model using a loss inspired by the Score-Distillation loss from the 3D object creation literature [35]: In each optimization step, BUDDI refines the current estimate towards a more plausible social interaction conditioned on the initial predictions. The refined pose, shape and translation serve as prior in the overall objective function." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 654, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 654, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 654, + 286, + 712 + ], + "type": "text", + "content": "Our contributions include (1) presenting the first generative model of a pair of 3D people in close social interaction and (2) a novel approach for reconstructing 3D human meshes from images without relying on ground-truth contact annotations. We perform extensive experiments with" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 308, + 72, + 544, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 72, + 544, + 250 + ], + "spans": [ + { + "bbox": [ + 308, + 72, + 544, + 250 + ], + "type": "text", + "content": "BUDDI to evaluate its performance on the FlickrCI3D Signatures dataset [8] as well as CHI3D, and the recent Hi4D dataset [54] and find that it outperforms the state of the art as well as strong baselines. We also evaluate the unconditional samples from the diffusion model in a perceptual study, where people find our samples more realistic " + }, + { + "bbox": [ + 308, + 72, + 544, + 250 + ], + "type": "inline_equation", + "content": "44.4\\%" + }, + { + "bbox": [ + 308, + 72, + 544, + 250 + ], + "type": "text", + "content": " when compared over real samples, where " + }, + { + "bbox": [ + 308, + 72, + 544, + 250 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 308, + 72, + 544, + 250 + ], + "type": "text", + "content": " is the upperbound where they do not see any difference. Importantly, we find that our optimization approach significantly improves the results of [42] both quantitatively and qualitatively. This work opens up a new avenue of research on digital human synthesis, laying the foundation for a deeper understanding of human social behavior derived from image data. Our data, code, and model will be available for research." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 308, + 263, + 391, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 263, + 391, + 275 + ], + "spans": [ + { + "bbox": [ + 308, + 263, + 391, + 275 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 308, + 282, + 544, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 282, + 544, + 449 + ], + "spans": [ + { + "bbox": [ + 308, + 282, + 544, + 449 + ], + "type": "text", + "content": "Generating 3D humans. There has been recent interest in generating 3D humans, in different contexts. Several methods automatically populate static 3D scenes with 3D humans [14, 60, 61], while more recent methods generate both body and hand poses to interact with 3D objects [43, 44, 49]. Other work generates human motions conditioned on different inputs such as audio [26, 47] or text [33, 34, 45]. Concurrent work proposes text-to-3D diffusion-based approaches to generate motion of two interacting humans [27, 40]. Neither method predicts the full body surface, but rather they synthesize either 3D joint locations or SMPL pose parameters for the average body. These methods are not used as priors for reconstructing interacting people from images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 308, + 450, + 544, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 450, + 544, + 652 + ], + "spans": [ + { + "bbox": [ + 308, + 450, + 544, + 652 + ], + "type": "text", + "content": "To model 3D human proxemics probabilistically, we employ diffusion models, which achieve impressive performance on image generation tasks [7, 15, 38, 39]. They have recently been adopted in 3D human motion generation scenarios: MDM [45] generates plausible motions conditioned on text input; PhysDiff [55] incorporates physical constraints in the diffusion process to generate physically plausible motions; and EDGE [47] uses a transformer-based diffusion model for dance generation. Related work [4, 6, 28] has investigated different modalities for the conditioning, e.g., audio, text, or action classes. EgoEgo [25] generates plausible full-body motions conditioned on the head motion. SceneDiffuser [16] focuses on the scene-conditioned setting. We also rely on techniques from the diffusion literature, but consider the unique setting where two people are in close interaction and leverage this for single-image 3D reconstruction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 308, + 654, + 544, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 654, + 544, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 654, + 544, + 712 + ], + "type": "text", + "content": "Multi-person 3D human mesh estimation. An extensive line of work focuses on reconstructing the 3D human pose and shape of a single person from images using optimization [2, 11, 24, 32, 37, 46, 50] or regression approaches [1, 12, 19, 20, 22, 29, 31, 51, 58, 59]. Capitalizing" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "type": "text", + "content": "9688" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 69, + 260, + 207 + ], + "blocks": [ + { + "bbox": [ + 48, + 69, + 260, + 207 + ], + "lines": [ + { + "bbox": [ + 48, + 69, + 260, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 69, + 260, + 207 + ], + "type": "image", + "image_path": "309b64c2bc74f21d55748b73879f9648e1a0223602603cf9c9bac43a9ce72380.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 217, + 546, + 296 + ], + "lines": [ + { + "bbox": [ + 46, + 217, + 546, + 296 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 546, + 296 + ], + "type": "text", + "content": "Figure 2. BUDDI: BUddies Diffusion model. On the left, we illustrate the architecture of BUDDI, our diffusion model for modeling 3D social proxemics between two people in close interaction. The diffusion process is applied directly on SMPL-X body parameters. To condition BUDDI on estimated body model parameters, " + }, + { + "bbox": [ + 46, + 217, + 546, + 296 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{H}}" + }, + { + "bbox": [ + 46, + 217, + 546, + 296 + ], + "type": "text", + "content": ", we concatenate the parameters along the token dimension. On the right, we illustrate the optimization method with BUDDI as prior. Our optimization takes detected keypoints [3, 52] and an initial regressor estimate [42] as input. Given the regressor estimate, we sample from BUDDI to obtain " + }, + { + "bbox": [ + 46, + 217, + 546, + 296 + ], + "type": "inline_equation", + "content": "\\tilde{x}" + }, + { + "bbox": [ + 46, + 217, + 546, + 296 + ], + "type": "text", + "content": " which we use to initialize the optimization routine. In each optimization iteration, we take a single diffuse-denoise step on the current estimate using the learned denoiser model " + }, + { + "bbox": [ + 46, + 217, + 546, + 296 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 217, + 546, + 296 + ], + "type": "text", + "content": ". Our losses encourage the current estimate to be close to the refined meshes " + }, + { + "bbox": [ + 46, + 217, + 546, + 296 + ], + "type": "inline_equation", + "content": "(L_{\\mathrm{diffusion}})" + }, + { + "bbox": [ + 46, + 217, + 546, + 296 + ], + "type": "text", + "content": " and to the initial estimate and detected keypoints " + }, + { + "bbox": [ + 46, + 217, + 546, + 296 + ], + "type": "inline_equation", + "content": "(L_{\\mathrm{fitting}})" + }, + { + "bbox": [ + 46, + 217, + 546, + 296 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 262, + 69, + 545, + 208 + ], + "blocks": [ + { + "bbox": [ + 262, + 69, + 545, + 208 + ], + "lines": [ + { + "bbox": [ + 262, + 69, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 262, + 69, + 545, + 208 + ], + "type": "image", + "image_path": "f3e68861a4b5e3b31c4203da0d49cd48ce47bffb1125ffb6f0dce333107e371e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 316, + 287, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 287, + 399 + ], + "type": "text", + "content": "on these techniques, recent approaches focus explicitly on reconstructing multiple people jointly from a single image. Zanfir et al. [56] propose an optimization solution, while Jiang et al. [18] and Sun et al. [41] rely on deep networks to regress the pose and shape for all people in the image. BEV [42] extends ROMP [41] to reason about the depth of people while taking age/height into account." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 401, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 287, + 616 + ], + "type": "text", + "content": "The above methods do not address contact between people. To do so, Fieraru et al. [8] introduce the first datasets with ground-truth labels for the body regions in contact between humans. Labels are collected using MoCap (CHI3D) or human annotators (FlickrCI3D Signatures). They propose an optimization approach that requires the ground-truth contact map to reconstruct people in close proximity at test time. More recently, REMIPS [9] introduces a transformer-based method that regresses the 3D pose of multiple people. REMIPS is trained using the above datasets while taking into account contact and interpenetration. In this work, we take a very different approach by learning and exploiting a 3D generative proxemics prior. We use the ground-truth contact maps to generate pseudoground truth 3D human fits from which we learn the diffusion model; once this is learned, we show that it can be used as a prior to recover plausible bodies in close proximity from images without explicit knowledge of contact maps." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "content": "Data-driven priors in optimization. Optimization-based methods for 3D human pose and shape estimation, like SMPLify [2], are versatile and allow different data-driven prior terms to be incorporated in the objective function. Different methods have been used to learn pose priors including GMMs [2], VAEs [32], neural distance fields [46], and normalizing flows [57]. ProHMR [23] learns a pose prior conditioned on image pixels. HuMoR [37] incor" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 316, + 545, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 545, + 413 + ], + "type": "text", + "content": "porates a data-driven motion prior in the iterative optimization. POSA [14] learns a prior for human-scene interaction from PROX data [13] and uses it in their optimization. In contrast to these methods, we use a diffusion model to capture the joint distribution over SMPL-X parameters for two people interacting and show that we can both sample from the model and use during optimization to improve the pose estimates of interacting people." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 434, + 361, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 434, + 361, + 445 + ], + "spans": [ + { + "bbox": [ + 306, + 434, + 361, + 445 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 458, + 545, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 458, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 458, + 545, + 590 + ], + "type": "text", + "content": "We introduce BUDDI, a generative model of two people in close social interaction. Because of the complexity and multimodality of the data, we turn to denoising diffusion probabilistic models [15] to address this task. In Sec. 3.1, we describe the basics of diffusion, and the parameterization we employ to model people in contact. In addition to sampling new body meshes from our model, our generative model can serve as a prior for reconstructing 3D humans from images. In Sec. 3.2, we describe an optimization procedure that incorporates BUDDI as a prior to recover two SMPL-X meshes from observed 2D keypoints." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": "For all of the following, we use the SMPL-X [32] body model to represent the human bodies. SMPL-X is a differentiable function that maps pose, " + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta} \\in \\mathbb{R}^{21 \\times 3}" + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": ", shape, " + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\beta \\in \\mathbb{R}^{10}" + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": ", and expression, " + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\psi \\in \\mathbb{R}^{10}" + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": " parameters to a mesh consisting of " + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "inline_equation", + "content": "N_v = 10,475" + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": " vertices " + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "inline_equation", + "content": "V \\in \\mathbb{R}^{N_v \\times 3}" + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": ". We place the generated meshes in the world by rotating and translating them by " + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\phi \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\gamma \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": ". We denote person " + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": "'s parameters as " + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "inline_equation", + "content": "X^a = [\\phi^a, \\pmb{\\theta}^a, \\pmb{\\beta}^a, \\gamma^a]" + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "inline_equation", + "content": "X^b = [\\phi^b, \\pmb{\\theta}^b, \\pmb{\\beta}^b, \\gamma^b]" + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": ". For simplicity, we refer to both people when no index is specified, e.g., " + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": " refers " + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "inline_equation", + "content": "X^a" + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "inline_equation", + "content": "X^b" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9689" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 230, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 230, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 230, + 83 + ], + "type": "text", + "content": "3.1. Diffusion Model for 3D Proxemics" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 89, + 287, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 89, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 46, + 89, + 287, + 174 + ], + "type": "text", + "content": "Denoising diffusion models are latent variable generative models that learn to transform random noise into the desired data distribution " + }, + { + "bbox": [ + 46, + 89, + 287, + 174 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{data}}" + }, + { + "bbox": [ + 46, + 89, + 287, + 174 + ], + "type": "text", + "content": " through a forward and reverse process. The forward inference process is a Markov chain over " + }, + { + "bbox": [ + 46, + 89, + 287, + 174 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 89, + 287, + 174 + ], + "type": "text", + "content": " steps given by transitions " + }, + { + "bbox": [ + 46, + 89, + 287, + 174 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x}_{t + 1}|\\mathbf{x}_t)" + }, + { + "bbox": [ + 46, + 89, + 287, + 174 + ], + "type": "text", + "content": ", which gradually adds Gaussian noise to clean samples " + }, + { + "bbox": [ + 46, + 89, + 287, + 174 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 46, + 89, + 287, + 174 + ], + "type": "text", + "content": " from the data distribution according to a fixed variance schedule " + }, + { + "bbox": [ + 46, + 89, + 287, + 174 + ], + "type": "inline_equation", + "content": "\\sigma_t" + }, + { + "bbox": [ + 46, + 89, + 287, + 174 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 174, + 287, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 174, + 287, + 305 + ], + "spans": [ + { + "bbox": [ + 46, + 174, + 287, + 305 + ], + "type": "text", + "content": "The reverse process " + }, + { + "bbox": [ + 46, + 174, + 287, + 305 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x}_{t - 1}|\\mathbf{x}_t)" + }, + { + "bbox": [ + 46, + 174, + 287, + 305 + ], + "type": "text", + "content": " then gradually denoises noisy samples back into the data distribution. The reverse process transitions follow a Gaussian distribution when conditioned on " + }, + { + "bbox": [ + 46, + 174, + 287, + 305 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 46, + 174, + 287, + 305 + ], + "type": "text", + "content": ", but must be inferred during the generative process. Following Ramesh et al. [36], we train a neural network " + }, + { + "bbox": [ + 46, + 174, + 287, + 305 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 174, + 287, + 305 + ], + "type": "text", + "content": " that predicts a sample " + }, + { + "bbox": [ + 46, + 174, + 287, + 305 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0 = D(\\mathbf{x}_t; t)" + }, + { + "bbox": [ + 46, + 174, + 287, + 305 + ], + "type": "text", + "content": " from a noisy sample " + }, + { + "bbox": [ + 46, + 174, + 287, + 305 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 46, + 174, + 287, + 305 + ], + "type": "text", + "content": " given the noise level " + }, + { + "bbox": [ + 46, + 174, + 287, + 305 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 174, + 287, + 305 + ], + "type": "text", + "content": ". For the task of reconstructing humans from images, when we have initial estimates of the SMPL-X parameters of two humans, we condition the denoising network " + }, + { + "bbox": [ + 46, + 174, + 287, + 305 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 174, + 287, + 305 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 46, + 174, + 287, + 305 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_H" + }, + { + "bbox": [ + 46, + 174, + 287, + 305 + ], + "type": "text", + "content": ", the predicted SMPL-X parameters of the two humans by a regressor." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 305, + 287, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 305, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 47, + 305, + 287, + 380 + ], + "type": "text", + "content": "We refer to the process of adding noise as diffusion and the process of removing the noise via " + }, + { + "bbox": [ + 47, + 305, + 287, + 380 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 47, + 305, + 287, + 380 + ], + "type": "text", + "content": " as denoising. Specifically, we diffuse a ground-truth sample " + }, + { + "bbox": [ + 47, + 305, + 287, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 47, + 305, + 287, + 380 + ], + "type": "text", + "content": " by uniformly sampling a noise level " + }, + { + "bbox": [ + 47, + 305, + 287, + 380 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 305, + 287, + 380 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 47, + 305, + 287, + 380 + ], + "type": "inline_equation", + "content": "\\epsilon_t \\sim \\mathcal{N}(0, \\mathbf{I})" + }, + { + "bbox": [ + 47, + 305, + 287, + 380 + ], + "type": "text", + "content": " to obtain the noisy sample " + }, + { + "bbox": [ + 47, + 305, + 287, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t = \\sqrt{\\sigma_t'}\\mathbf{x}_0 + \\sqrt{1 - \\sigma_t'}\\epsilon_t" + }, + { + "bbox": [ + 47, + 305, + 287, + 380 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 47, + 305, + 287, + 380 + ], + "type": "inline_equation", + "content": "\\sigma_t' = \\prod_{i=1}^{t} (1 - \\sigma_t)" + }, + { + "bbox": [ + 47, + 305, + 287, + 380 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 380, + 178, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 380, + 178, + 390 + ], + "spans": [ + { + "bbox": [ + 59, + 380, + 178, + 390 + ], + "type": "text", + "content": "We then train " + }, + { + "bbox": [ + 59, + 380, + 178, + 390 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 59, + 380, + 178, + 390 + ], + "type": "text", + "content": " to minimize" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 399, + 287, + 412 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 287, + 412 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 287, + 412 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {\\mathbf {x} _ {0} \\sim p _ {\\text {d a t a}}} \\mathbb {E} _ {t \\sim \\mathcal {U} \\{0, T \\}, \\mathbf {x} _ {t} \\sim q (\\cdot | \\mathbf {x} _ {0})} | | D (\\mathbf {x} _ {t}; t, \\mathbf {c} _ {H}) - \\mathbf {x} _ {0} | |, \\tag {1}", + "image_path": "84a67e59eaced9e58ad8016e0175ff6104db43778bb1ff5d7728d3a34b24dfe4.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 419, + 287, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 419, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 47, + 419, + 287, + 443 + ], + "type": "text", + "content": "where we set " + }, + { + "bbox": [ + 47, + 419, + 287, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_H = \\varnothing" + }, + { + "bbox": [ + 47, + 419, + 287, + 443 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 47, + 419, + 287, + 443 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 47, + 419, + 287, + 443 + ], + "type": "text", + "content": " of conditional model training, and all of unconditional model training." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 457, + 287, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 457, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 457, + 287, + 552 + ], + "type": "text", + "content": "Architecture. Because we aim to model close contact between people, we choose a model state space that can express the full surface of the human body. Specifically, in contrast to prior work in human motion diffusion that operate only on joint angles and locations [45, 55], we directly operate on the full SMPL-X parameters of the two people. A sample " + }, + { + "bbox": [ + 46, + 457, + 287, + 552 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 457, + 287, + 552 + ], + "type": "text", + "content": " thus corresponds to the concatenation of two bodies:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 559, + 266, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 559, + 266, + 574 + ], + "spans": [ + { + "bbox": [ + 67, + 559, + 266, + 574 + ], + "type": "interline_equation", + "content": "\\mathbf {x} = \\left[ X ^ {a}, X ^ {b} \\right] = \\left[ \\phi^ {a}, \\boldsymbol {\\theta} ^ {a}, \\boldsymbol {\\beta} ^ {a}, \\gamma^ {a}, \\phi^ {b}, \\boldsymbol {\\theta} ^ {b}, \\boldsymbol {\\beta} ^ {b}, \\gamma^ {b} \\right].", + "image_path": "787883abb8162be6c8e8fc3202a40d3613aef0a17457ccc0ac526e7a487dd0a6.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "content": "We denoise a sample " + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "content": " with a transformer encoder block on tokenized parameters. Specifically, each parameter of each person is tokenized into 152-dimensional latent vectors with per-parameter and per-person embedding layers. We tokenize the noise level " + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "content": " similarly with a noise embedding. When conditioning is available, i.e. SMPL-X estimates for person " + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "content": ", we similarly tokenize the parameters to be used as additional tokens. We pass the available tokens into the transformer encoder, and similarly decode the output tokens with per-token embeddings. We illustrate the denoiser architecture in Figure 2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 96 + ], + "type": "text", + "content": "Losses. We employ standard human pose and shape regularization losses. We write our training objective as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 361, + 105, + 545, + 118 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 105, + 545, + 118 + ], + "spans": [ + { + "bbox": [ + 361, + 105, + 545, + 118 + ], + "type": "interline_equation", + "content": "L _ {D} = L _ {\\theta} + L _ {\\beta} + L _ {\\gamma} + L _ {v 2 v}, \\tag {2}", + "image_path": "dfa7b72fbda7140277820ce1f859d3e2357af34299291b2c8b307599cf91fecc.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 126, + 545, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 126, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 304, + 126, + 545, + 198 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 126, + 545, + 198 + ], + "type": "inline_equation", + "content": "L_{\\theta}, L_{\\beta}, L_{\\gamma}" + }, + { + "bbox": [ + 304, + 126, + 545, + 198 + ], + "type": "text", + "content": " denote squared L2-losses on respective body model parameters, and " + }, + { + "bbox": [ + 304, + 126, + 545, + 198 + ], + "type": "inline_equation", + "content": "L_{v2v}" + }, + { + "bbox": [ + 304, + 126, + 545, + 198 + ], + "type": "text", + "content": " denotes a squared L2 loss on model vertices. We use 6D rotation representations [62] for global orientation and pose, and model the relative translation between " + }, + { + "bbox": [ + 304, + 126, + 545, + 198 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 304, + 126, + 545, + 198 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 126, + 545, + 198 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 304, + 126, + 545, + 198 + ], + "type": "text", + "content": ". We show generated samples from our unconditional model in Fig. 3." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 205, + 510, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 205, + 510, + 217 + ], + "spans": [ + { + "bbox": [ + 305, + 205, + 510, + 217 + ], + "type": "text", + "content": "3.2. Optimization with the Proxemics Prior" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 224, + 545, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 224, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 304, + 224, + 545, + 354 + ], + "type": "text", + "content": "Reconstructing 3D human meshes from a single image is an extremely under-constrained problem, and priors over human pose and shape are crucial in an optimization based framework for recovering plausible meshes [2, 32, 56]. Our problem involves people in close contact, which requires correctly placing the meshes in context with each other, which has only been done when given ground truth contact annotations at test time [8]. We remove the need for ground-truth contact maps by using our generative model as a prior during reconstruction with a score distillation approach [35, 48]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 354, + 545, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 545, + 403 + ], + "type": "text", + "content": "During inference, we observe detected 2D keypoints " + }, + { + "bbox": [ + 304, + 354, + 545, + 403 + ], + "type": "inline_equation", + "content": "\\tilde{J}_{2D}" + }, + { + "bbox": [ + 304, + 354, + 545, + 403 + ], + "type": "text", + "content": " and initial body model parameter estimates " + }, + { + "bbox": [ + 304, + 354, + 545, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_H" + }, + { + "bbox": [ + 304, + 354, + 545, + 403 + ], + "type": "text", + "content": " from a regressor [42]. We then optimize the body parameters of two people to minimize" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 343, + 412, + 545, + 425 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 412, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 343, + 412, + 545, + 425 + ], + "type": "interline_equation", + "content": "L _ {\\text {O p t i z i m a t i o n w . B U D D I}} = L _ {\\text {f i t t i n g}} + L _ {\\text {d i f f u s i o n}}. \\tag {3}", + "image_path": "0a69e781e336e83cbc865444fc19094fdab703c345e43283ac2fb23ac60986c6.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 433, + 545, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 433, + 545, + 505 + ], + "spans": [ + { + "bbox": [ + 304, + 433, + 545, + 505 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{fitting}}" + }, + { + "bbox": [ + 304, + 433, + 545, + 505 + ], + "type": "text", + "content": " ensures that the solution stays close to the image evidence, while " + }, + { + "bbox": [ + 304, + 433, + 545, + 505 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{diffusion}}" + }, + { + "bbox": [ + 304, + 433, + 545, + 505 + ], + "type": "text", + "content": " is a data-driven prior using our conditional diffusion model. We treat this prior as similar to those used for 3D pose in previous works such as GMM [2] and V-Poser [32], but for 3D proxemics. We illustrate the optimization procedure in Fig. 2 right." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 505, + 545, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 505, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 505, + 545, + 552 + ], + "type": "text", + "content": "We initialize our optimization by generating a sample " + }, + { + "bbox": [ + 304, + 505, + 545, + 552 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}" + }, + { + "bbox": [ + 304, + 505, + 545, + 552 + ], + "type": "text", + "content": " from the conditional model. We sample with DDIM sampling with 100 evenly spaced steps. We then use the data fitting loss:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 358, + 563, + 545, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 563, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 358, + 563, + 545, + 576 + ], + "type": "interline_equation", + "content": "L _ {\\text {f i t t i n g}} = \\lambda_ {J} L _ {J} + \\lambda_ {\\tilde {\\theta}} L _ {\\tilde {\\theta}} + \\lambda_ {P} L _ {P}, \\tag {4}", + "image_path": "bb805dba6c7687d6d60a66534d847e84492770d28754871cbb0d95fdacff055d.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 586, + 545, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 586, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 304, + 586, + 545, + 632 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 586, + 545, + 632 + ], + "type": "inline_equation", + "content": "L_{J}" + }, + { + "bbox": [ + 304, + 586, + 545, + 632 + ], + "type": "text", + "content": " denotes 2D re-projection error between the reprojected 3D joints of the current estimate and the detected 2D keypoints, and " + }, + { + "bbox": [ + 304, + 586, + 545, + 632 + ], + "type": "inline_equation", + "content": "L_{\\tilde{\\theta}}" + }, + { + "bbox": [ + 304, + 586, + 545, + 632 + ], + "type": "text", + "content": " is a prior for the solution to be close to the denoised initialization." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 633, + 545, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 633, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 304, + 633, + 545, + 681 + ], + "type": "inline_equation", + "content": "L_{P}" + }, + { + "bbox": [ + 304, + 633, + 545, + 681 + ], + "type": "text", + "content": " denotes an interpenetration loss between two people that pushes inside vertices to the surface, which we use winding numbers [17] to find intersecting vertices between low-resolution SMPL-X meshes of the current estimates:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 687, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 687, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 315, + 687, + 545, + 715 + ], + "type": "interline_equation", + "content": "L _ {P} = \\sum_ {v \\in V _ {I} ^ {a}} \\min _ {u \\in V ^ {b}} \\| v - u \\| ^ {2} + \\sum_ {v \\in V _ {I} ^ {b}} \\min _ {u \\in V ^ {a}} \\| v - u \\| ^ {2}, \\tag {5}", + "image_path": "b54747c88890306c4cb780d9aac959b92269bed5f5381c69c0f266daaa1d029f.jpg" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9690" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 87, + 531, + 259 + ], + "blocks": [ + { + "bbox": [ + 77, + 87, + 531, + 259 + ], + "lines": [ + { + "bbox": [ + 77, + 87, + 531, + 259 + ], + "spans": [ + { + "bbox": [ + 77, + 87, + 531, + 259 + ], + "type": "image", + "image_path": "67b8969f2bc4b92fc703e2a3e2e368be28f9a12b7b2b08cc057d7f18397191d6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 267, + 547, + 304 + ], + "lines": [ + { + "bbox": [ + 46, + 267, + 547, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 267, + 547, + 304 + ], + "type": "text", + "content": "Figure 3. Generative Proxemics: Samples from BUDDI. All samples are unconditionally generated from pure noise using the trained diffusion model. We select several representative examples and show two views per sample. These samples reveal that BUDDI has learned the distribution of people in close contact including embracing each other, playing sports, sitting side by side, and taking photographs." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 322, + 287, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 322, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 322, + 287, + 346 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 322, + 287, + 346 + ], + "type": "inline_equation", + "content": "V_{I}^{a}" + }, + { + "bbox": [ + 46, + 322, + 287, + 346 + ], + "type": "text", + "content": " denotes vertices of " + }, + { + "bbox": [ + 46, + 322, + 287, + 346 + ], + "type": "inline_equation", + "content": "M^a" + }, + { + "bbox": [ + 46, + 322, + 287, + 346 + ], + "type": "text", + "content": " intersecting the lowresolution mesh of " + }, + { + "bbox": [ + 46, + 322, + 287, + 346 + ], + "type": "inline_equation", + "content": "M^b" + }, + { + "bbox": [ + 46, + 322, + 287, + 346 + ], + "type": "text", + "content": "; and vice versa for " + }, + { + "bbox": [ + 46, + 322, + 287, + 346 + ], + "type": "inline_equation", + "content": "V_{I}^{b}" + }, + { + "bbox": [ + 46, + 322, + 287, + 346 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 346, + 288, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 346, + 288, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 288, + 407 + ], + "type": "text", + "content": "To use the prior on human interaction into account, we use the learned denoising model " + }, + { + "bbox": [ + 46, + 346, + 288, + 407 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 346, + 288, + 407 + ], + "type": "text", + "content": " from BUDDI and perform a single diffuse-denoise step, with a noise level at " + }, + { + "bbox": [ + 46, + 346, + 288, + 407 + ], + "type": "inline_equation", + "content": "t = 10" + }, + { + "bbox": [ + 46, + 346, + 288, + 407 + ], + "type": "text", + "content": ", on the current estimate. The denoised estimate, " + }, + { + "bbox": [ + 46, + 346, + 288, + 407 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0 = D(\\mathbf{x}_t; t, \\mathbf{c}_H)" + }, + { + "bbox": [ + 46, + 346, + 288, + 407 + ], + "type": "text", + "content": ", regularizes the current estimate via" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 101, + 416, + 287, + 431 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 416, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 101, + 416, + 287, + 431 + ], + "type": "interline_equation", + "content": "L _ {\\text {d i f f u s i o n}} = \\left\\| D \\left(\\mathbf {x} _ {t}; t, \\mathbf {c} _ {H}\\right) - \\mathbf {x} \\right\\|, \\tag {6}", + "image_path": "b6a541632519c2d78479c432767a963811efc70e3c3dd83408a34671af055972.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 438, + 289, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 438, + 289, + 501 + ], + "spans": [ + { + "bbox": [ + 46, + 438, + 289, + 501 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 438, + 289, + 501 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t = \\sqrt{\\sigma_t^{\\prime}}\\mathbf{x}_{\\mathrm{no - grad}} + \\sqrt{1 - \\sigma_t^{\\prime}}\\epsilon_t" + }, + { + "bbox": [ + 46, + 438, + 289, + 501 + ], + "type": "text", + "content": " denotes the diffused body model parameters of the current estimate, and " + }, + { + "bbox": [ + 46, + 438, + 289, + 501 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{no - grad}}" + }, + { + "bbox": [ + 46, + 438, + 289, + 501 + ], + "type": "text", + "content": " denotes the current estimate with detached gradients. " + }, + { + "bbox": [ + 46, + 438, + 289, + 501 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0" + }, + { + "bbox": [ + 46, + 438, + 289, + 501 + ], + "type": "text", + "content": " , and encourages " + }, + { + "bbox": [ + 46, + 438, + 289, + 501 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 438, + 289, + 501 + ], + "type": "text", + "content": " to be close to " + }, + { + "bbox": [ + 46, + 438, + 289, + 501 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0" + }, + { + "bbox": [ + 46, + 438, + 289, + 501 + ], + "type": "text", + "content": " . In practice, we penalize the decoded parameters of " + }, + { + "bbox": [ + 46, + 438, + 289, + 501 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 438, + 289, + 501 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 438, + 289, + 501 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0" + }, + { + "bbox": [ + 46, + 438, + 289, + 501 + ], + "type": "text", + "content": " directly as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 78, + 509, + 287, + 543 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 509, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 78, + 509, + 287, + 543 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {\\text {d i f f u s i o n}} = \\lambda_ {\\hat {\\phi}} | | \\hat {\\phi} _ {0} - \\phi | | + \\lambda_ {\\hat {\\theta}} | | \\hat {\\theta} _ {0} - \\theta | | \\tag {7} \\\\ + \\lambda_ {\\hat {\\boldsymbol {\\beta}}} | | \\hat {\\boldsymbol {\\beta}} _ {0} - \\boldsymbol {\\beta} | | + \\lambda_ {\\hat {\\boldsymbol {\\gamma}}} | | \\hat {\\boldsymbol {\\gamma}} _ {0} - \\boldsymbol {\\gamma} | |. \\\\ \\end{array}", + "image_path": "6ddb5ff9beda5fdb4297b24b06556f4608e5dfd07d5be2382a0d239ae8aaf124.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 551, + 288, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 551, + 288, + 599 + ], + "spans": [ + { + "bbox": [ + 46, + 551, + 288, + 599 + ], + "type": "text", + "content": "Intuitively, this loss uses the learned denoiser " + }, + { + "bbox": [ + 46, + 551, + 288, + 599 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 551, + 288, + 599 + ], + "type": "text", + "content": " to take a step from the current estimate towards the data distribution of two people in close proximity, conditioned on the regressor prediction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 609, + 183, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 183, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 183, + 624 + ], + "type": "text", + "content": "4. Implementation Details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "content": "Training Data. There are few datasets containing 3D ground truth of humans in close social interaction [8, 54]. Such datasets are usually captured in lab environments, consequently they are small and do not contain the variety of interactions between humans \"in the wild,\" e.g. when playing sports or taking social pictures. To address this lack of data, we create Flickr Fits, i.e. SMPL-X fits" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 322, + 547, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 322, + 547, + 455 + ], + "spans": [ + { + "bbox": [ + 304, + 322, + 547, + 455 + ], + "type": "text", + "content": "for Flickr images portraying humans in contact scenarios. For this, we use FlickrCI3D Signatures [8], a dataset of images showing interacting humans collected from Flickr with discrete human-human contact annotations. Specifically, the SMPL-X body surface is divided into " + }, + { + "bbox": [ + 304, + 322, + 547, + 455 + ], + "type": "inline_equation", + "content": "R = 75" + }, + { + "bbox": [ + 304, + 322, + 547, + 455 + ], + "type": "text", + "content": " regions such that each region, " + }, + { + "bbox": [ + 304, + 322, + 547, + 455 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 322, + 547, + 455 + ], + "type": "text", + "content": ", roughly covers a similar area. For a given photo, the human annotators assign a binary label indicating contact between a region on one person and a region on the other. For two meshes, " + }, + { + "bbox": [ + 304, + 322, + 547, + 455 + ], + "type": "inline_equation", + "content": "M^a" + }, + { + "bbox": [ + 304, + 322, + 547, + 455 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 322, + 547, + 455 + ], + "type": "inline_equation", + "content": "M^b" + }, + { + "bbox": [ + 304, + 322, + 547, + 455 + ], + "type": "text", + "content": ", the annotation can be represented as a binary contact map " + }, + { + "bbox": [ + 304, + 322, + 547, + 455 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^D \\in \\{0,1\\}^{R \\times R}" + }, + { + "bbox": [ + 304, + 322, + 547, + 455 + ], + "type": "text", + "content": ", where" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 463, + 545, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 545, + 491 + ], + "type": "interline_equation", + "content": "\\mathcal {C} _ {i j} ^ {D} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} r _ {i} \\text {o f} M ^ {a} \\text {i s i n c o n t a c t w i t h} r _ {j} \\text {o f} M ^ {b} \\\\ 0, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {8}", + "image_path": "40ecff8c0cae1f7ee09ca927f91dc11d0f6160c830e9cac53a58eba73e4222d1.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 498, + 547, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 641 + ], + "type": "text", + "content": "We use these ground-truth contact maps in an optimization routine for fitting two people to detected keypoints, similar to Sec. 3.2 but replace the diffusion model prior with standard image fitting priors. The dataset contains 10,631/1,139 train/test images, with one image containing multiple contact annotations. Note that we use pairs of people in contact as labeled in the FlickrCI3D Signatures dataset [8] and only use images containing matching regressor estimates, 2D keypoints, and contact labels. See Fig. 4 for qualitative examples and the Sup. Mat. for more details about the selection process and an evaluation of the optimization method with ground-truth contact labels." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "content": "We also augment our training data with available MoCap data, which is considerably smaller than those obtained from image fits: CHI3D [8] contains 3/2 pairs of training/test subjects performing 127 sequences of two-person interactions like hugs or kicks with ground-truth SMPL-X bodies. One frame per sequence has contact map annota" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 314, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 314, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 314, + 758 + ], + "type": "text", + "content": "9691" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 64, + 70, + 286, + 208 + ], + "blocks": [ + { + "bbox": [ + 64, + 70, + 286, + 208 + ], + "lines": [ + { + "bbox": [ + 64, + 70, + 286, + 208 + ], + "spans": [ + { + "bbox": [ + 64, + 70, + 286, + 208 + ], + "type": "image", + "image_path": "f72b40b5e6a3ff82197ae5ea6209b1143354e3b1afb0147d938623591cec6c62.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 217, + 288, + 262 + ], + "lines": [ + { + "bbox": [ + 46, + 217, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 288, + 262 + ], + "type": "text", + "content": "Figure 4. Flickr Fits. We visualize the output of the optimization process that reconstructs two people in close proximity using ground-truth contact maps, shown from four different views. We use these 3D fits as training data for BUDDI." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 270, + 287, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 287, + 437 + ], + "type": "text", + "content": "tions. We use only the contact frame of the sequences from two subject pairs, resulting in 247 mesh pairs for training, and the third pair for evaluation. Hi4D [54] contains sequences of 20 pairs of people interacting with each other. The interactions include actions like hugging, dancing, and fighting. We randomly split the data into " + }, + { + "bbox": [ + 46, + 270, + 287, + 437 + ], + "type": "inline_equation", + "content": "14/3/3" + }, + { + "bbox": [ + 46, + 270, + 287, + 437 + ], + "type": "text", + "content": " pairs for train/val/test and use every fifth frame of the subsequence involving contact as labeled in Hi4D, resulting in about 1K mesh pairs for training. The body representation format in Hi4D is SMPL, which we transfer to SMPL-X using the SMPL-X code repository [32]. Please see the Sup. Mat. for more details of the datasets. Note that while we use SMPL-X model, BUDDI is not trained on hands because none of these datasets contain hand poses." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 437, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 628 + ], + "type": "text", + "content": "BUDDI Training. BUDDI is trained with meshes from FlickrCI3D Signatures Fits, CHI3D, and Hi4D. We use " + }, + { + "bbox": [ + 46, + 437, + 287, + 628 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 46, + 437, + 287, + 628 + ], + "type": "text", + "content": " Flickr, " + }, + { + "bbox": [ + 46, + 437, + 287, + 628 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 46, + 437, + 287, + 628 + ], + "type": "text", + "content": " CHI3D, and " + }, + { + "bbox": [ + 46, + 437, + 287, + 628 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 46, + 437, + 287, + 628 + ], + "type": "text", + "content": " Hi4D data distribution per batch with batch size 512. The transformer backbone has six layers and eight heads; we use " + }, + { + "bbox": [ + 46, + 437, + 287, + 628 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 437, + 287, + 628 + ], + "type": "text", + "content": " dropout and randomly shuffle the order of people during training. To train BUDDI, we randomly sample noise levels " + }, + { + "bbox": [ + 46, + 437, + 287, + 628 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 437, + 287, + 628 + ], + "type": "text", + "content": " up to 1000 using a cosine noise schedule [30]. We use the Adam optimizer [21] with learning rate " + }, + { + "bbox": [ + 46, + 437, + 287, + 628 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 46, + 437, + 287, + 628 + ], + "type": "text", + "content": ". We train two separate networks: an unconditional model for generation and the conditional version for reconstruction. For the conditional model, we use all camera views of the MoCap datasets, i.e. 4/8 cameras for CHI3D/Hi4D. The unconditional model is trained on 3D MoCap fits in the world coordinate system. To sample new poses, we use DDIM sampling starting at noise levels " + }, + { + "bbox": [ + 46, + 437, + 287, + 628 + ], + "type": "inline_equation", + "content": "t = 1000" + }, + { + "bbox": [ + 46, + 437, + 287, + 628 + ], + "type": "text", + "content": " in steps of 10." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "text", + "content": "Optimization Details. During optimization, we experiment with different noise levels, between 10 and 100, and find that " + }, + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "inline_equation", + "content": "t = 10" + }, + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "text", + "content": " does not disturb the inputs too much, but enough for " + }, + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "text", + "content": " to generate new configurations. We use detected 2D keypoints from OpenPose [3] and ViTpose [52] and BEV [42] estimates as conditioning. Unlike single-person mesh regressors, BEV is designed to predict multiple" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": "people including their relative depth. Please see Sup. Mat. for more details." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 106, + 387, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 106, + 387, + 119 + ], + "spans": [ + { + "bbox": [ + 306, + 106, + 387, + 119 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 126, + 545, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 126, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 304, + 126, + 545, + 389 + ], + "type": "text", + "content": "Baselines. We compare our reconstruction method with BEV [42], which is also used as an input to our conditional model. Since there is no other available work that reasons about people in close social interaction, we experiment with simple but effective baselines. We train the transformer model of BUDDI to directly predict SMPL-X parameters of people in contact from BEV input, essentially a deterministic, single-step ablation of BUDDI. We also evaluate the direct conditional denoised output of BEV by BUDDI without any optimization. As another baseline, we propose an optimization routine that replaces " + }, + { + "bbox": [ + 304, + 126, + 545, + 389 + ], + "type": "inline_equation", + "content": "L_{\\text{diffusion}}" + }, + { + "bbox": [ + 304, + 126, + 545, + 389 + ], + "type": "text", + "content": " with a simple heuristic that takes the minimal distances between two meshes predicted by BEV and minimizes their distance during optimization along with the other energy terms. Finally, to compare the generation ability we train a VAE which we also use during the optimization routine in a similar manner to VPoser [32] but for two people by optimizing the VAE latent space instead of SMPL-X parameters. We refer to these models as Transformer, BUDDI (gen.), Contact Heuristic, and VAE, respectively. All baselines are trained on the same datasets as BUDDI with the same sampling strategies. Details about our baselines are provided in the Sup. Mat.." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 389, + 546, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 389, + 546, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 389, + 546, + 545 + ], + "type": "text", + "content": "Metrics. We use standard evaluation metrics from the human pose and shape estimation literature. We also report the joint PA-MPJPE computed by performing Procrustes alignment of both people together. In addition to per-person metrics, this captures the relative orientation and translation of the two people. Since our method directly estimates 3D humans we propose a new metric similar to PCK [53] from the 2D pose literature called PCC, the percentage of correct contact points with respect to a radius " + }, + { + "bbox": [ + 304, + 389, + 546, + 545 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 389, + 546, + 545 + ], + "type": "text", + "content": ". Specifically, given two meshes, " + }, + { + "bbox": [ + 304, + 389, + 546, + 545 + ], + "type": "inline_equation", + "content": "M^a / M^b" + }, + { + "bbox": [ + 304, + 389, + 546, + 545 + ], + "type": "text", + "content": " and a contact map " + }, + { + "bbox": [ + 304, + 389, + 546, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^D" + }, + { + "bbox": [ + 304, + 389, + 546, + 545 + ], + "type": "text", + "content": " we compute the pairwise vertex-to-vertex Euclidean distances " + }, + { + "bbox": [ + 304, + 389, + 546, + 545 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{eucl}}(\\mathcal{C}^D)" + }, + { + "bbox": [ + 304, + 389, + 546, + 545 + ], + "type": "text", + "content": " between annotated contact regions and consider the pair to be correct when " + }, + { + "bbox": [ + 304, + 389, + 546, + 545 + ], + "type": "inline_equation", + "content": "\\min(d_{\\mathrm{eucl}}(\\mathcal{C}^D)) < r" + }, + { + "bbox": [ + 304, + 389, + 546, + 545 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 552, + 451, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 552, + 451, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 552, + 451, + 563 + ], + "type": "text", + "content": "5.1. Unconditional Generation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 570, + 545, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 654 + ], + "type": "text", + "content": "We qualitatively evaluate BUDDI by showing samples generated from the model in Fig. 1 and 3. Our approach is able to generate people in close proximity including embraces, handshakes, having a conversation, sitting side by side, and, in general, plausibly interacting with each other. Since it is trained on Internet image collections, it also learns to generate people posing for photographs or playing sports." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "We further run a perceptual study to evaluate the realism of the generated social interactions against other methods. In a forced choice study, we compare our generated samples with samples from the real data distribution according to the 60/20/20 per-batch ratio for Flickr/CHI3D/Hi4D" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9692" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 70, + 550, + 404 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 550, + 404 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 550, + 404 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 550, + 404 + ], + "type": "image", + "image_path": "b6a5bf1fac02cd6bc4f974a89c2775a5b3fdac8ac3486b88bd98b1cc84f37845.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 412, + 547, + 449 + ], + "lines": [ + { + "bbox": [ + 46, + 412, + 547, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 412, + 547, + 449 + ], + "type": "text", + "content": "Figure 5. Automatic reconstruction of people in close social interaction. We show qualitative results from a) BEV, b) contact heuristics, which takes the BEV output and encourages the closest parts to be in contact, and c) our method, which optimizes the BEV estimates against the image evidence with the BUDDI prior. Our approach recovers a plausible reconstruction with subtle details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 85, + 464, + 251, + 548 + ], + "blocks": [ + { + "bbox": [ + 85, + 464, + 251, + 548 + ], + "lines": [ + { + "bbox": [ + 85, + 464, + 251, + 548 + ], + "spans": [ + { + "bbox": [ + 85, + 464, + 251, + 548 + ], + "type": "table", + "html": "
JOINT ↓\nPA-MPJPEPCC at radius ↑
510152025
BEV106-----
Transformer861440607382
BUDDI (gen.)921539587180
Heuristic681434496170
VAE1011128425565
BUDDI661944627381
", + "image_path": "3ed800cd20658cd4c76b956fb93f83acaabe971b7e5fea8f422689c070d3856b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 560, + 288, + 605 + ], + "lines": [ + { + "bbox": [ + 46, + 560, + 288, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 560, + 288, + 605 + ], + "type": "text", + "content": "Table 1. 3D Pose Evaluation on FlickrCI3D Signatures. We evaluate methods against the Flickr fits using their joint (two-person) PA-MPJPE expressed in mm. We also evaluate the percentage of correct contact points (PCC) for radius " + }, + { + "bbox": [ + 46, + 560, + 288, + 605 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 560, + 288, + 605 + ], + "type": "text", + "content": " mm." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 618, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 289, + 715 + ], + "type": "text", + "content": "used during training. We also compare BUDDI against generations from the VAE and a non-parametric random baseline that samples meshes from the pseudo-ground truth after centering the two people. We do a forced choice comparison between BUDDI and these three other methods, asking workers on Amazon Mechanical Turk to choose the sample that shows a more realistic close social interaction. We use 256 samples per method. We collect ratings for 768" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 467, + 547, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 467, + 547, + 539 + ], + "spans": [ + { + "bbox": [ + 304, + 467, + 547, + 539 + ], + "type": "text", + "content": "pairwise comparisons. In this study, BUDDI was chosen over random in " + }, + { + "bbox": [ + 304, + 467, + 547, + 539 + ], + "type": "inline_equation", + "content": "71.23\\%" + }, + { + "bbox": [ + 304, + 467, + 547, + 539 + ], + "type": "text", + "content": " of the comparisons, over the VAE in " + }, + { + "bbox": [ + 304, + 467, + 547, + 539 + ], + "type": "inline_equation", + "content": "60.17\\%" + }, + { + "bbox": [ + 304, + 467, + 547, + 539 + ], + "type": "text", + "content": ", and over the training data in " + }, + { + "bbox": [ + 304, + 467, + 547, + 539 + ], + "type": "inline_equation", + "content": "44.4\\%" + }, + { + "bbox": [ + 304, + 467, + 547, + 539 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 304, + 467, + 547, + 539 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 467, + 547, + 539 + ], + "type": "text", + "content": " is the upper bound for such forced choice comparisons, in which participants cannot tell the difference between real and generated samples." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 540, + 547, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 540, + 547, + 612 + ], + "spans": [ + { + "bbox": [ + 304, + 540, + 547, + 612 + ], + "type": "text", + "content": "For a quantitative evaluation, we compute the FID score between samples from BUDDI and samples from the " + }, + { + "bbox": [ + 304, + 540, + 547, + 612 + ], + "type": "inline_equation", + "content": "VAE" + }, + { + "bbox": [ + 304, + 540, + 547, + 612 + ], + "type": "text", + "content": " on concatenated SMPL-X parameters. We sample 8K examples per method and from our training data following the dataset ratio per batch. BUDDI has a lower FID score (1.6) compared to the " + }, + { + "bbox": [ + 304, + 540, + 547, + 612 + ], + "type": "inline_equation", + "content": "VAE" + }, + { + "bbox": [ + 304, + 540, + 547, + 612 + ], + "type": "text", + "content": " (3.3)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 622, + 422, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 622, + 422, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 622, + 422, + 635 + ], + "type": "text", + "content": "5.2. Fitting with BUDDI" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": "We show qualitative results in Fig. 5 comparing BUDDI against BEV and the Contact Heuristic. Our approach is able to generate various types of human interactions with plausible contact and depth placement. It is also able to capture close interaction between a child and a parent. Although the Contact Heuristic (center) is able to move two" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "9693" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 553, + 160 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 553, + 160 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 553, + 160 + ], + "type": "table", + "html": "
PER PERSON ↓ PA-MPJPEJOINT ↓ PA-MPJPEJOINT PA-MPJPE
backhugbasketballcheersdancefighthighfivehugkissposesidehugtalk
BEV78 / 84136200126109135121106163139142131118
Heuristic67 / 7112116883941319468159159118113109
BUDDI (F, C)70 / 77115200949212810810013311410410791
Transformer79 / 85120161141103138123128117106120105100
BUDDI (gen.)82 / 901171521391201371309610197115102101
VAE80 / 8213817513311414111987176162135140113
BUDDI70 / 769812795921131097210585889681
", + "image_path": "8530125c9b4849884ec11b92ce8d77a98514eddc418d37d23c0edf4d0c4173ec.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 85, + 220, + 252, + 311 + ], + "blocks": [ + { + "bbox": [ + 46, + 169, + 547, + 204 + ], + "lines": [ + { + "bbox": [ + 46, + 169, + 547, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 169, + 547, + 204 + ], + "type": "text", + "content": "Table 2. Evaluation of BUDDI on Hi4D. We compare the output of BUDDI to the proposed baseline methods on the Hi4D challenge. The first block shows methods that do not use Hi4D data during training or are optimization based without access to priors trained on Hi4D. BUDDI (F,C) in particular, is our model BUDDI trained on Flickr and CHI3D data only. All errors are reported in mm for 3D Joints." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 85, + 220, + 252, + 311 + ], + "lines": [ + { + "bbox": [ + 85, + 220, + 252, + 311 + ], + "spans": [ + { + "bbox": [ + 85, + 220, + 252, + 311 + ], + "type": "table", + "html": "
PER PERSON ↓\nPA-MPJPEJOINT ↓\nPA-MPJPE
BEV50 5296
Transformer54 56105
BUDDI (gen.)53 5380
Heuristic49 46105
VAE54 54103
BUDDI48 4768
", + "image_path": "e6af8957d89256d4d240ee5897730bb78f2e76518a5abeec2417623653445f0d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 319, + 288, + 353 + ], + "lines": [ + { + "bbox": [ + 46, + 319, + 288, + 353 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 288, + 353 + ], + "type": "text", + "content": "Table 3. Quantitative Evaluation on CHI3D. We compare the output of our model to the baselines on CHI3D (pair s03). All errors reported in mm for 3D Joints." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 362, + 287, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 362, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 362, + 287, + 434 + ], + "type": "text", + "content": "people closer together, which helps with image alignment, upon close observation it is not able to capture the subtle interaction between people that happens during intimate interaction. BUDDI's estimates are more realistic and better capture the subtle details of interaction. We provide additional qualitative examples in the Sup. Mat." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 436, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 436, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 45, + 436, + 287, + 652 + ], + "type": "text", + "content": "We further report the percentage of correct contact (PCC) with respect to the ground truth contact map on the FlickrCI3D Signatures test set in Table 1. The table also shows the pose reconstruction accuracy against our Flickr Fits. All metrics show improvement over BEV, in particular the joint PA-MPJPE. Non-optimization methods, i.e. Transformer and BUDDI (gen.), are able to predict plausible contacts, with similar PCC accuracy to BUDDI, but struggle to reconstruct the data with a worse joint PA-MPJPE. The Heuristic, in contrast, achieves a lower reconstruction error, but worse PCC. Our approach which leverages the learned prior during optimization can recover both the relative positions and contacts between the two people. To provide insights into the performance of single-person mesh regressors when evaluated on the two-person reconstruction task, we run 4D Humans [10] on Flickr Fits. The joint PA-MPJPE is " + }, + { + "bbox": [ + 45, + 436, + 287, + 652 + ], + "type": "inline_equation", + "content": "344\\mathrm{mm}" + }, + { + "bbox": [ + 45, + 436, + 287, + 652 + ], + "type": "text", + "content": " which is high, as expected, since these methods are not trained to reason about proximity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "We further evaluate our model against ground truth MoCap data in Tables 2 and 3. Optimization with BUDDI consistently improves the two-person reconstruction error over BEV and other baselines. When evaluated per action, the strongest improvements over BEV come from complex" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 224, + 547, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 224, + 547, + 403 + ], + "spans": [ + { + "bbox": [ + 304, + 224, + 547, + 403 + ], + "type": "text", + "content": "close social interactions like hugging or kissing, at " + }, + { + "bbox": [ + 304, + 224, + 547, + 403 + ], + "type": "inline_equation", + "content": "58\\mathrm{mm}" + }, + { + "bbox": [ + 304, + 224, + 547, + 403 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 224, + 547, + 403 + ], + "type": "inline_equation", + "content": "54\\mathrm{mm}" + }, + { + "bbox": [ + 304, + 224, + 547, + 403 + ], + "type": "text", + "content": " absolute improvement over BUDDI respectively. The Heuristic baseline achieves a low PA-MPJPE reconstruction error on all three datasets in particular for poses with a few physical contact points, such as a handshake, whereas more complex contact, such as a hug, requires data-driven priors like BUDDI; we provide an analysis in the Sup. Mat. to quantify this hypothesis. Transformer and BUDDI (gen.) have lower joint PA-MPJPE errors than BEV and the Heuristic, but worse per-person reconstruction errors. The VAE results suggest that directly operating in the latent space of a generative model is challenging and not sufficient to accurately recover close social interactions. BUDDI, in contrast, is able to model a wide variety of poses, as supported by the numerical results." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 416, + 379, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 416, + 379, + 430 + ], + "spans": [ + { + "bbox": [ + 306, + 416, + 379, + 430 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 437, + 547, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 547, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 547, + 533 + ], + "type": "text", + "content": "We propose BUDDI, a diffusion model for close human-human interaction. We train BUDDI from 3D fits obtained from a large-scale dataset of images with ground truth contact annotations as well as a small set of available mocap data. BUDDI enables unconditional sampling of people in close social interaction. More importantly, we also demonstrate how BUDDI can be used as effective prior for single-view 3D reconstruction of two people in close proximity." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 534, + 547, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 547, + 642 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 547, + 642 + ], + "type": "text", + "content": "Our core contribution of a generative proxemics prior provides the foundation for future work on modeling and capturing human interaction. For example, future work could iteratively apply our method to reconstruct more than two people in close proximity, explore other conditioning modalities like pixel features, text, action labels, or, by taking the outputs of recent single-person mesh regressors into account, address reconstructing fine-grained interaction including finger pose and facial expressions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 643, + 545, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 643, + 545, + 691 + ], + "spans": [ + { + "bbox": [ + 304, + 643, + 545, + 691 + ], + "type": "text", + "content": "Acknowledgements: We thank A. Holynski, E. Weber, F. Warburg, B. Peebles, J. Rajasegaran, K. Mangalam and N. Athanasiou for insightful discussions, and A. Cseke, T. McConnell and T. Alexiadis for running the user study." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 691, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 691, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 691, + 545, + 715 + ], + "type": "text", + "content": "Disclosure: https://files.is.tue.mpg.de/black/CoI_CVPR_2024.txt" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "9694" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Anurag Arnab, Carl Doersch, and Andrew Zisserman. Exploiting temporal context for 3D human pose estimation in the wild. In Computer Vision and Pattern Recognition (CVPR), 2019." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 201 + ], + "type": "text", + "content": "[2] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J. Black. Keep it SMPL: Automatic estimation of 3D human pose and shape from a single image. In European Conference on Computer Vision (ECCV), pages 561-578. Springer International Publishing, 2016." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 204, + 287, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 204, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 53, + 204, + 287, + 258 + ], + "type": "text", + "content": "[3] Zhe Cao, Gines Hidalgo Martinez, Tomas Simon, Shih-En Wei, and Yaser Sheikh. OpenPose: Realtime multi-person 2D pose estimation using part affinity fields. Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 43 (1):172-186, 2019." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 260, + 287, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 260, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 53, + 260, + 287, + 315 + ], + "type": "text", + "content": "[4] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18000-18010, 2023." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 317, + 286, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 317, + 286, + 360 + ], + "spans": [ + { + "bbox": [ + 53, + 317, + 286, + 360 + ], + "type": "text", + "content": "[5] April H Crusco and Christopher G Wetzel. The midas touch: The effects of interpersonal touch on restaurant tipping. *Personality and Social Psychology Bulletin*, 10(4): 512-517, 1984." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 362, + 286, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 362, + 286, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 362, + 286, + 407 + ], + "type": "text", + "content": "[6] Rishabh Dabral, Muhammad Hamza Mughal, Vladislav Golyanik, and Christian Theobalt. Mofusion: A framework for denoising-diffusion-based motion synthesis. In Computer Vision and Pattern Recognition (CVPR), 2023." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 408, + 286, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 408, + 286, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 408, + 286, + 441 + ], + "type": "text", + "content": "[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat GANs on image synthesis. Conference on Neural Information Processing Systems (NeurIPS), 34:8780-8794, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 442, + 286, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 442, + 286, + 496 + ], + "spans": [ + { + "bbox": [ + 53, + 442, + 286, + 496 + ], + "type": "text", + "content": "[8] Mihai Fieraru, Mihai Zanfir, Elisabeta Oneata, Alin-Ionut Popa, Vlad Olaru, and Cristian Sminchisescu. Three-dimensional reconstruction of human interactions. In Computer Vision and Pattern Recognition (CVPR), pages 7214-7223, 2020." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 498, + 286, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 498, + 286, + 563 + ], + "spans": [ + { + "bbox": [ + 53, + 498, + 286, + 563 + ], + "type": "text", + "content": "[9] Mihai Fieraru, Mihai Zanfir, Teodor Szente, Eduard Bazavan, Vlad Olaru, and Cristian Sminchisescu. REMIPS: Physically consistent 3D reconstruction of multiple interacting people under weak supervision. In Conference on Neural Information Processing Systems (NeurIPS), pages 19385-19397, 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 566, + 286, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 566, + 286, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 566, + 286, + 609 + ], + "type": "text", + "content": "[10] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa*, and Jitendra Malik*. Humans in 4D: Reconstructing and tracking humans with transformers. In International Conference on Computer Vision (ICCV), 2023." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 612, + 286, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 612, + 286, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 612, + 286, + 655 + ], + "type": "text", + "content": "[11] Peng Guan, Alexander Weiss, Alexandru Balan, and Michael J. Black. Estimating human shape and pose from a single image. In International Conference on Computer Vision (ICCV), pages 1381-1388, 2009." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 657, + 286, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 657, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 657, + 286, + 689 + ], + "type": "text", + "content": "[12] Riza Alp Guler and Iasonas Kokkinos. HoloPose: Holistic 3D human reconstruction in-the-wild. In Computer Vision and Pattern Recognition (CVPR), 2019." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 286, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 286, + 714 + ], + "type": "text", + "content": "[13] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J Black. Resolving 3D human pose ambigui" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "ties with 3D scene constraints. In International Conference on Computer Vision (ICCV), pages 2282-2292, 2019." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 97, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 97, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 97, + 545, + 140 + ], + "type": "text", + "content": "[14] Mohamed Hassan, Partha Ghosh, Joachim Tesch, Dimitrios Tzionas, and Michael J Black. Populating 3D scenes by learning human-scene interaction. In Computer Vision and Pattern Recognition (CVPR), pages 14708-14718, 2021." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 141, + 545, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 545, + 174 + ], + "type": "text", + "content": "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Conference on Neural Information Processing Systems (NeurIPS), pages 6840-6851, 2020." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 175, + 545, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 175, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 308, + 175, + 545, + 230 + ], + "type": "text", + "content": "[16] Siyuan Huang, Zan Wang, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, and Song-Chun Zhu. Diffusion-based generation, optimization, and planning in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 232, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 232, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 232, + 545, + 274 + ], + "type": "text", + "content": "[17] Alec Jacobson, Ladislav Kavan, and Olga Sorkine-Hornung. Robust inside-outside segmentation using generalized winding numbers. ACM Transactions on Graphics (TOG), 32(4): 1-12, 2013." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 276, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 545, + 319 + ], + "type": "text", + "content": "[18] Wen Jiang, Nikos Kolotouros, Georgios Pavlakos, Xiaowei Zhou, and Kostas Daniilidis. Coherent reconstruction of multiple humans from a single image. In Computer Vision and Pattern Recognition (CVPR), pages 5578-5587, 2020." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 321, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 321, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 308, + 321, + 545, + 364 + ], + "type": "text", + "content": "[19] Hanbyul Joo, Natalia Neverova, and Andrea Vedaldi. Exemplar fine-tuning for 3D human model fitting towards in-the-wild 3D human pose estimation. In International Conference on 3D Vision (3DV), 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 365, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 545, + 409 + ], + "type": "text", + "content": "[20] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In Computer Vision and Pattern Recognition (CVPR), pages 7122-7131, 2018." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 411, + 545, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 442 + ], + "type": "text", + "content": "[21] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 445, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 445, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 545, + 498 + ], + "type": "text", + "content": "[22] Nikos Kolotouros, Georgios Pavlakos, Michael J. Black, and Kostas Daniilidis. Learning to reconstruct 3D human pose and shape via model-fitting in the loop. In International Conference on Computer Vision (ICCV), pages 2252-2261, 2019." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 501, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 501, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 501, + 545, + 544 + ], + "type": "text", + "content": "[23] Nikos Kolotouros, Georgios Pavlakos, Dinesh Jayaraman, and Kostas Daniilidis. Probabilistic modeling for human mesh recovery. In International Conference on Computer Vision (ICCV), 2021." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 545, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 545, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 308, + 545, + 545, + 600 + ], + "type": "text", + "content": "[24] Christoph Lassner, Javier Romero, Martin Kiefel, Federica Bogo, Michael J. Black, and Peter V. Gehler. Unite the people: Closing the loop between 3D and 2D human representations. In Computer Vision and Pattern Recognition (CVPR), pages 6050-6059, 2017." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 601, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 601, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 308, + 601, + 545, + 645 + ], + "type": "text", + "content": "[25] Jiaman Li, Karen Liu, and Jiajun Wu. Ego-body pose estimation via ego-head pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17142-17151, 2023." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "text", + "content": "[26] Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. AI choreographer: Music conditioned 3D dance generation with AIST++. In International Conference on Computer Vision (ICCV), pages 13401-13412, 2021." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[27] Han Liang, Wenqian Zhang, Wenxuan Li, Jingyi Yu, and Lan Xu. Intergen: Diffusion-based multi-human motion" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9695" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "text", + "content": "generation under complex interactions. arXiv preprint arXiv:2304.05684, 2023." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 128 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 128 + ], + "type": "text", + "content": "[28] Jianxin Ma, Shuai Bai, and Chang Zhou. Pretrained diffusion models for unified human motion synthesis. arXiv preprint arXiv:2212.02837, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 129, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 172 + ], + "type": "text", + "content": "[29] Lea Müller, Ahmed A. A. Osman, Siyu Tang, Chun-Hao P. Huang, and Michael J. Black. On self-contact and human pose. In Computer Vision and Pattern Recognition (CVPR), pages 9990-9999, 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 173, + 287, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 173, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 173, + 287, + 216 + ], + "type": "text", + "content": "[30] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning (ICML), pages 8162-8171, 2021." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 217, + 287, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 272 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 272 + ], + "type": "text", + "content": "[31] Mohamed Omran, Christoph Lassner, Gerard Pons-Moll, Peter V. Gehler, and Bernt Schiele. Neural body fitting: Unifying deep learning and model based human pose and shape estimation. In International Conference on 3D Vision (3DV), pages 484-494, 2018." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 272, + 287, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 287, + 327 + ], + "type": "text", + "content": "[32] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 327, + 287, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 287, + 370 + ], + "type": "text", + "content": "[33] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3D human motion synthesis with transformer VAE. In Computer Vision and Pattern Recognition (CVPR), pages 10985-10995, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 371, + 287, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 414 + ], + "type": "text", + "content": "[34] Mathis Petrovich, Michael J Black, and Gül Varol. TEMOS: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision (ECCV), pages 480-497, 2022." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 415, + 287, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 446 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 446 + ], + "type": "text", + "content": "[35] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 449, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 480 + ], + "type": "text", + "content": "[36] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents, 2022." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 482, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 482, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 482, + 287, + 525 + ], + "type": "text", + "content": "[37] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J Guibas. HuMoR: 3D human motion model for robust pose estimation. In International Conference on Computer Vision (ICCV), 2021." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 526, + 287, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 526, + 287, + 570 + ], + "spans": [ + { + "bbox": [ + 48, + 526, + 287, + 570 + ], + "type": "text", + "content": "[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 571, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 571, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 571, + 287, + 635 + ], + "type": "text", + "content": "[39] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 636, + 287, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 669 + ], + "type": "text", + "content": "[40] Yonatan Shafir, Guy Tevet, Roy Kapon, and Amit H Bermano. Human motion diffusion as a generative prior. arXiv preprint arXiv:2303.01418, 2023." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[41] Yu Sun, Qian Bao, Wu Liu, Yili Fu, Michael J Black, and Tao Mei. Monocular, one-stage, regression of multiple 3D people. In Computer Vision and Pattern Recognition (CVPR), pages 11179-11188, 2021." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[42] Yu Sun, Wu Liu, Qian Bao, Yili Fu, Tao Mei, and Michael J Black. Putting people in their place: Monocular regression of 3d people in depth. In Computer Vision and Pattern Recognition (CVPR), pages 13243-13252, 2022." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "text", + "content": "[43] Omid Taheri, Vasileios Choutas, Michael J. Black, and Dimitrios Tzionas. GOAL: Generating 4D whole-body motion for hand-object grasping. In Computer Vision and Pattern Recognition (CVPR), pages 13263-13273, 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 163, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 196 + ], + "type": "text", + "content": "[44] Purva Tendulkar, Dídac Surís, and Carl Vondrick. FLEX: Full-body grasping without full-body grasps. In Computer Vision and Pattern Recognition (CVPR), 2023." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 197, + 545, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 230 + ], + "type": "text", + "content": "[45] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 231, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 231, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 231, + 545, + 274 + ], + "type": "text", + "content": "[46] Garvita Tiwari, Dimitrije Antic, Jan Eric Lenssen, Nikolaos Sarafianos, Tony Tung, and Gerard Pons-Moll. Pose-ndf: Modeling human pose manifolds with neural distance fields. In European Conference on Computer Vision (ECCV), 2022." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 276, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 308 + ], + "type": "text", + "content": "[47] Jonathan Tseng, Rodrigo Castellon, and C Karen Liu. EDGE: Editable dance generation from music. Computer Vision and Pattern Recognition (CVPR), 2023." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 310, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 310, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 307, + 310, + 545, + 365 + ], + "type": "text", + "content": "[48] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12619-12629, 2023." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 365, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 409 + ], + "type": "text", + "content": "[49] Yan Wu, Jiahao Wang, Yan Zhang, Siwei Zhang, Otmar Hilliges, Fisher Yu, and Siyu Tang. Saga: Stochastic whole-body grasping with contact. In European Conference on Computer Vision (ECCV), pages 257-274, 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 411, + 545, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 465 + ], + "type": "text", + "content": "[50] Hongyi Xu, Eduard Gabriel Bazavan, Andrei Zanfir, William T. Freeman, Rahul Sukthankar, and Cristian Sminchisescu. GHUM & GHUML: Generative 3D human shape and articulated pose models. In Computer Vision and Pattern Recognition (CVPR), pages 6183-6192, 2020." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 467, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 467, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 467, + 545, + 510 + ], + "type": "text", + "content": "[51] Yuanlu Xu, Song-Chun Zhu, and Tony Tung. DenseRaC: Joint 3D pose and shape estimation by dense render-and compare. In International Conference on Computer Vision (ICCV), 2019." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 511, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 511, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 511, + 545, + 555 + ], + "type": "text", + "content": "[52] Yufei Xu, Jing Zhang, Qiming Zhang, and Dacheng Tao. Vitpose: Simple vision transformer baselines for human pose estimation. Advances in Neural Information Processing Systems, 35:38571-38584, 2022." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 556, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 556, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 307, + 556, + 545, + 589 + ], + "type": "text", + "content": "[53] Yi Yang and Deva Ramanan. Articulated human detection with flexible mixtures of parts. IEEE transactions on pattern analysis and machine intelligence, 35(12):2878-2890, 2012." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 590, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 634 + ], + "type": "text", + "content": "[54] Yifei Yin, Chen Guo, Manuel Kaufmann, Juan Zarate, Jie Song, and Otmar Hilliges. Hi4d: 4d instance segmentation of close human interaction. In Computer Vision and Pattern Recognition (CVPR), 2023." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 635, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 679 + ], + "type": "text", + "content": "[55] Ye Yuan, Jiaming Song, Umar Iqbal, Arash Vahdat, and Jan Kautz. Physdiff: Physics-guided human motion diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16010-16021, 2023." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "text", + "content": "[56] Andrei Zanfir, Elisabeta Maroiniu, and Cristian Sminchisescu. Monocular 3d pose and shape estimation of multiple people in natural scenes-the importance of multiple scene" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9696" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 397 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 67, + 73, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 286, + 95 + ], + "type": "text", + "content": "constraints. In Computer Vision and Pattern Recognition (CVPR), pages 2148-2157, 2018." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "type": "text", + "content": "[57] Andrei Zanfir, Eduard Gabriel Bazavan, Hongyi Xu, William T Freeman, Rahul Sukthankar, and Cristian Sminchisescu. Weakly supervised 3D human pose and shape reconstruction with normalizing flows. In European Conference on Computer Vision (ECCV), pages 465-481, 2020." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 206 + ], + "type": "text", + "content": "[58] Andrei Zanfir, Eduard Gabriel Bazavan, Mihai Zanfir, William T Freeman, Rahul Sukthankar, and Cristian Sminchisescu. Neural descent for visual 3D human pose and shape. In Computer Vision and Pattern Recognition (CVPR), pages 14484-14493, 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 287, + 262 + ], + "type": "text", + "content": "[59] Hongwen Zhang, Yating Tian, Xinchi Zhou, Wanli Ouyang, Yebin Liu, Limin Wang, and Zhenan Sun. PyMAF: 3D human pose and shape regression with pyramidal mesh alignment feedback loop. In International Conference on Computer Vision (ICCV), pages 11446-11456, 2021." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 263, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 263, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 263, + 287, + 308 + ], + "type": "text", + "content": "[60] Siwei Zhang, Yan Zhang, Qianli Ma, Michael J Black, and Siyu Tang. PLACE: Proximity learning of articulation and contact in 3D environments. In International Conference on 3D Vision (3DV), pages 642-651, 2020." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 308, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 308, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 287, + 352 + ], + "type": "text", + "content": "[61] Yan Zhang, Mohamed Hassan, Heiko Neumann, Michael J Black, and Siyu Tang. Generating 3D people in scenes without people. In Computer Vision and Pattern Recognition (CVPR), pages 6194-6204, 2020." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 354, + 287, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 397 + ], + "type": "text", + "content": "[62] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Computer Vision and Pattern Recognition (CVPR), pages 5745-5753, 2019." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9697" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Generative Quanta Color Imaging/9125a07a-d19f-455b-bcb9-5f2fd8cc25e9_content_list.json b/2024/Generative Quanta Color Imaging/9125a07a-d19f-455b-bcb9-5f2fd8cc25e9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..39ceb40b472625bb120ea724e209ad590b73807e --- /dev/null +++ b/2024/Generative Quanta Color Imaging/9125a07a-d19f-455b-bcb9-5f2fd8cc25e9_content_list.json @@ -0,0 +1,1367 @@ +[ + { + "type": "text", + "text": "Generative Quanta Color Imaging", + "text_level": 1, + "bbox": [ + 310, + 130, + 658, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vishal Purohit, Junjie Luo, Yiheng Chi, Qi Guo, Stanley H. Chan, Qiang Qiu Elmore Family School of Electrical and Computer Engineering Purdue University, West Lafayette, USA.", + "bbox": [ + 181, + 180, + 789, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{purohitv,luo330,chi14,guo675,stanchan,qqiu}@purdue.edu", + "bbox": [ + 238, + 236, + 723, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The astonishing development of single-photon cameras has created an unprecedented opportunity for scientific and industrial imaging. However, the high data throughput generated by these 1-bit sensors creates a significant bottleneck for low-power applications. In this paper, we explore the possibility of generating a color image from a single binary frame of a single-photon camera. We evidently find this problem being particularly difficult to standard colorization approaches due to the substantial degree of exposure variation. The core innovation of our paper is an exposure synthesis model framed under a neural ordinary differential equation (Neural ODE) that allows us to generate a continuum of exposures from a single observation. This innovation ensures consistent exposure in binary images that colorizers take on, resulting in notably enhanced colorization. We demonstrate applications of the method in single-image and burst colorization and show superior generative performance over baselines. Project website can be found at https://vishal-s-p.github.io/projects/2023/generative_quanta_color.html", + "bbox": [ + 73, + 318, + 473, + 621 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 631, + 209, + 646 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Single-photon image sensors are gaining a strong momentum over the past decade due to their impeccable photon-counting capability. This not only opens the door for advanced scientific imaging in low-light environment, but also creates the opportunity for new imaging in high-speed, high dynamic range, and low-bit conditions. Single-photon imaging today are realized through single-photon avalanche diodes (SPAD) and quanta image sensors (QIS). The latest development of QIS has demonstrated a mega-pixel resolution [48] which is compatible with mainstream cell phone cameras, whereas SPAD has achieved hundreds of thousands of (binary) frames per second which has enabled a plethora of industrial and 3D imaging applications [16].", + "bbox": [ + 75, + 657, + 468, + 854 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Problem Statement. In the context of 1-bit single-photon imaging, the data volume generated is huge. Taking a $1000 \\times 1000$ SPAD array as an example, a video of", + "bbox": [ + 75, + 854, + 470, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Binary Input Pix2PixDNI Our Method Groundtruth", + "bbox": [ + 501, + 285, + 777, + 296 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/900a82ebe2a17662b8c37752d6039080a656a4dc93c8d4757103bd60fa57fafd.jpg", + "image_caption": [ + "Figure 1. We introduce generative quanta color imaging. Given a binary frame captured by a single-photon camera (quanta image sensor (QIS) in this example), the proposed method generates a continuum of exposures using a neural ordinary differential equation framework. Color are then generated based on these exposures. Left: A qualitative comparison between our approach and existing methods is shown. Right: Comparison between colorization results and image captured using a RGB CMOS camera under similar conditions." + ], + "image_footnote": [], + "bbox": [ + 501, + 297, + 779, + 460 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e5467101037743f5643ff514b2579ea22fa07905613a732a24a27f0aa1ec7d75.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 792, + 296, + 890, + 460 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "500 binary frames requires a $500\\mathrm{Mb / s}$ data transfer rate; if the frame rate becomes $150\\mathrm{k}$ fps, the data transfer rate increases to $150\\mathrm{Gb / s}$ . Since data throughput scales proportionally with the power consumption, it quickly becomes infeasible to use these sensors continuously without considering power supply, memory, and transmission issues. While solid progress are being made across the sensor and signal processing community [1, 11, 17, 23, 24, 45-47, 49, 50], the data throughput remains an unsolved problem for single-photon sensors to be considered applicable in many applications, including low-power devices, wireless devices, underwater vehicles, spacecrafts, etc., where bandwidth is the bottleneck.", + "bbox": [ + 496, + 617, + 892, + 813 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, instead of compressing data or redesigning the sensor, we ask the question", + "bbox": [ + 500, + 816, + 890, + 847 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Is it possible to recover a color image from a single 1-bit image?", + "bbox": [ + 524, + 862, + 867, + 893 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "25138", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Significance. There are several reasons why this problem is significant. Firstly, while there is an ocean of literature on image generation (through methods such as Generative Adversarial Networks [26] and diffusion [31]), our problem is unique in the sense none of the existing work addresses it. Besides being binary (and hence colorless), in practice the input data usually also suffers from under- or over-exposures. Therefore, before generating color, it is even more critical to synthesize the unseen exposure. Secondly, any solution offered to our problem will provide a path to generative imaging in the future. From the point of view of compressed acquisition, our method demonstrates the extreme case of compression where the incoming data is a single bit. For future AR/VR applications where power is severely limited, our approach shows the possibility of extreme data compression.", + "bbox": [ + 75, + 90, + 472, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contributions. The core algorithmic contribution of this paper is two-fold", + "bbox": [ + 76, + 332, + 468, + 362 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a neural ordinary differential equation (neural ODE) based framework to synthesize a continuum of exposures that are not available in the measurement. By controlling the integration interval of neural ODE, we obtain convolutional filters that are responsible for converting input representation to image representation of desired exposure.", + "- We theoretically and empirically show that a controlled variation of filter atoms leads to controlled exposure changes in the output image." + ], + "bbox": [ + 76, + 363, + 468, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We verify our proposed method using both synthetic and real single-photon image sensor data. A snapshot of the results is shown in Fig. 1.", + "bbox": [ + 75, + 513, + 468, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 573, + 218, + 588 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Single-Photon Imaging. The context of this paper is single-photon image sensors, of which we are mostly interested in one-bit sensing because of its potential in high-speed and high dynamic range imaging [2]. Today, 1-bit sensing has been realized using quanta image sensors (QIS) [19-21, 50, 67] and single-photon avalanche diodes (SPAD) [1, 16, 17]. From an image reconstruction point of view, since 1-bit data contains so little information, all existing image reconstruction methods need to use multiple frames. For QIS, these methods range from classical optimization [3, 5], transform-denoise [4], neural network [12], color demosaicking [18], frequency demodulation [18], and student-teacher learning [11, 22]. For SPAD, quanta burst photography [51] and its variants [30, 32, 35, 57] first align frames and then apply image fusions [27-29, 59]. The question we ask in this paper is to push the input requirement to the limit: can we recover/generate a color image from a binary input?", + "bbox": [ + 75, + 598, + 472, + 868 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image-to-Image Translation. This paper focuses on using generative models. Generative Adversarial Networks", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(GANs) [26] have been a significant breakthrough in the field of image translation. Pix2Pix [33] and CycleGAN [68] demonstrated the effectiveness of GANs in image-to-image translation tasks, with paired and unpaired datasets, respectively. Several subsequent works have focused on improving image fidelity, disentangling style and content, and generating images with varying styles, such as Pix2PixHD [61], StarGAN [13], StyleGAN [37], and StyleGANv2 [14]. Recent advancements in continuous cross-domain translation techniques have further improved the output quality by leveraging images from intermediate domains. These techniques include interpolation of networks [40, 41], interpolation of two latent vectors [14, 37], and exploiting the interpolation path and translation of manifold [9, 42, 55].", + "bbox": [ + 496, + 90, + 890, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Convolutional Filter Decomposition. In the proposed method, one of the key components is the convolutional filter decomposition. Several techniques have been proposed to decompose filters of Convolutional Neural Networks (CNNs) in order to reduce the computation complexity by exploiting the redundancy in convolutional filters [10, 39, 53, 54, 63, 69]. Filter decomposition methods achieve similar performance as compared to the original CNNs while offering speedups in computation. As it is expensive to model the high-dimension space of convolutional filters directly, Qiu et al. [56] observed that the convolutional filters can often be approximated using a linear combination small set of basis elements. By decomposing filters into a small set of basis, Wang et al. [64, 65] were able to sample the basis and demonstrated theoretical and empirical results on stochastic image generation tasks. Futher, Chen et al. [7] showed that filter decomposition can be used to calculate network similarity and large convolutional models can be fine tuned using filter decomposition [8].", + "bbox": [ + 496, + 303, + 892, + 590 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Proposed Approach", + "text_level": 1, + "bbox": [ + 500, + 608, + 691, + 626 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. One-Bit Image Formation Model", + "text_level": 1, + "bbox": [ + 500, + 633, + 790, + 650 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We consider a simplified image formation model for 1-bit image sensors. Letting $\\theta$ be the underlying quanta exposure [20], the observed signal is defined by", + "bbox": [ + 496, + 657, + 890, + 705 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nY = \\operatorname {A D C} \\left\\{\\text {P o i s s o n} (\\theta) + \\text {G a u s s i a n} \\left(0, \\sigma_ {r} ^ {2}\\right) \\right\\},\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 717, + 843, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\sigma_r$ denotes the standard deviation of read noise, and ADC is the analog-to-digital conversion. For 1-bit sensing, the ADC can be modeled as a threshold such that $\\mathrm{ADC}(x) = 1$ if $x \\geq q$ and $\\mathrm{ADC}(x) = 0$ if $x < q$ for some threshold $q$ . Since $Y$ is a binary random variable, it is fully characterized by the probability mass function, which is [2]", + "bbox": [ + 496, + 756, + 890, + 861 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\np _ {Y} (1) = \\sum_ {k = 0} ^ {\\infty} \\frac {e ^ {- \\theta} \\theta^ {k}}{k !} \\Phi \\left(\\frac {k - q}{\\sigma_ {r}}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 864, + 807, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "25139", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0a412ef20fc6665ff386e6b92ac579b7530768973b6fe897ccb70c41383edb7d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 89, + 588, + 205 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d19eea29fd9623bbe343a5cfd25e45c74b76b4ed0d542eaf8effb270fc75c590.jpg", + "image_caption": [ + "Figure 2. Illustration of exposure correction and colorization of binary images using neural networks. (a) depicts a range of images from overexposed to underexposed, illustrating the degradation of image details due to exposure variation. (b) contrasts the standard colorization workflow and our proposed approach. (i) In standard colorization approaches, a neural network learns to map a binary image $\\mathbf{Y}$ to the corresponding color image $\\mathbf{X}_c$ via a neural network, $\\mathcal{F}^{\\mathrm{aug}}$ , where superscript 'aug' indicates the colorizer is trained using dataset with augmented exposure images. (ii) In contrast, our approach does not require training colorizer with augmented exposure images. (c) compares the colorization results: the first row is the output of a colorizer trained without augmentation, the second row is the output of colorizer trained with augmented data, the third row corresponds to the results of our method and the last provides the ground truth images for reference." + ], + "image_footnote": [], + "bbox": [ + 80, + 205, + 588, + 383 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/49d6687fb09d1b4eec2811a3145e43b77894b9acfbf2b080262ed524ede45462.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 99, + 888, + 385 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\Phi(z) = \\int_{-\\infty}^{z} \\frac{1}{\\sqrt{2\\pi}} \\exp\\{-t^2/2\\} dt$ is the cumulative distribution of a standard Gaussian. It then follows that the mean (or the bit density) is $\\mu = \\mathbb{E}[Y] = p_Y(1)$ .", + "bbox": [ + 75, + 531, + 470, + 582 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/0108bdf016ace58c99108d00d4ec4dd407feba3e8002819849009497c5a661c1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodGenerator SizeDataset SizeTraining TimeFID
Standard colorizationResNet-93x1x62.34
+ Increased training timeResNet-93x1.5x54.25
+ Increased model sizeResNet-103x1x74.58
Proposed methodResNet-91x1x49.88
", + "bbox": [ + 76, + 594, + 470, + 659 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Performance comparison of the standard colorization approach under various training conditions versus the proposed method.", + "bbox": [ + 75, + 667, + 470, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Why Exposure Synthesis is Required for Colorization?", + "text_level": 1, + "bbox": [ + 76, + 739, + 468, + 771 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a binary image $\\mathbf{Y} \\in \\mathbb{R}^N$ , our goal is to recover a color image $\\mathbf{X}_c \\in \\mathbb{R}^{N \\times 3}$ . However, since the measurement does not contain any color information, our problem fits the classical work of image colorization [38, 58, 60, 66]. However, it is non-trivial to deploy such an extension in real world scenarios, because very likely there is a discrepancy in the exposure levels of the input images compared to those used during training.", + "bbox": [ + 75, + 779, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An example of binary images of different exposures are shown in Figure 2(a) illustrating the effect of exposure change on the details in the captured binary binary image. Consider a typical colorization network for example, an Image-to-Image translation model like Pix2Pix [34] trained on well exposed image but tested on overexposed images, the colorization results are notably poorer (shown in Figure 2(c) row 1). This is a realistic concern because the underlying exposure $\\theta$ can fluctuate dramatically in real life.", + "bbox": [ + 496, + 532, + 893, + 670 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Possible solutions include (i) Exposure Augmentation : training dataset is enriched with images of varying exposures, at a cost of additional training time. We find, however, this approach partially addresses the issue, as demonstrated in our results shown in Figure 2(b) row 2. In particular, this approach falls short at over- or under-exposure. One can alleviate this issue by applying multiple colorizers, each handling a narrow range of exposures, but this is not ideal as it demands substantially more computational resources. (ii) Increased Model size : Another solution is to increase the size of generator used in Pix2Pix [34] and train for longer period of time. In Table 1, we present the colorization performance represented by the FID metric under various training scenarios. Under increased generator size or increased training time the standard colorization ap", + "bbox": [ + 496, + 674, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "25140", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "proach performs poorly compared to proposed method detailed in subsequent sections.", + "bbox": [ + 75, + 90, + 468, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proposed Solution. We address the challenge by a dedicated exposure correction based on neural ordinary differential equations (neural ODEs) to standardize the exposure of binary images. This specialization allows the colorizer to focus on consistent-exposure images, resulting in notably enhanced colorization, as evidenced Figure 2(c) row 3. Additionally, our exposure synthesis can generate multiple exposures from a single binary image, enabling more effective multi-exposure colorization. Some of the added advantages of our approach include: (a) Modularity. The colorizer in our design is interchangeable, allowing for upgrades to more advanced architectures without requiring retraining on a large, augmented dataset. (b) Parameter Efficiency. Our method achieves its objectives without resorting to multiple networks with large parameter sets, ensuring a more streamlined and resource-efficient process.", + "bbox": [ + 75, + 121, + 472, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Continuous Exposure Synthesis using Filter Decomposed Neural ODEs", + "text_level": 1, + "bbox": [ + 76, + 369, + 468, + 401 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The image formation model defined in Section 3.1 indicates that we can formulate the continuum burst of binary images as $\\mathbf{Y}(\\widetilde{\\theta})$ , where the binary image $\\mathbf{Y}$ is a function of a continuous exposure value $\\widetilde{\\theta} \\in (0,1]$ . A small $\\widetilde{\\theta}$ value indicates a higher exposed image $\\mathbf{Y}$ . One of the core contributions of this work is a deep neural network architecture that synthesizes a binary image $\\mathbf{Y}_t$ at a target exposure $\\widetilde{\\theta}_{\\mathrm{target}}$ given an input binary image $\\mathbf{Y}_i$ measured at the exposure value $\\widetilde{\\theta}_{\\mathrm{input}}$ .", + "bbox": [ + 75, + 407, + 468, + 549 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As the exposure values are continuous, we may need an exhaustive training set that densely samples every possible pair of input exposure value $\\widetilde{\\theta}_{\\mathrm{input}}$ and target exposure value $\\widetilde{\\theta}_{\\mathrm{target}}$ for every scene to train our network. To tackle this problem, we designed a convolutional architecture that enforces the parameter of the network to smoothly change according to the exposures.", + "bbox": [ + 75, + 549, + 468, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use the convolution filter decomposition technique [56, 61, 64] to represent the convolutional filters $\\mathbf{F}$ with a low number of parameters, and we leverage neural ODEs [6] to associate parameters with image exposures. With these techniques, we enable the adaptive control of filter parameters $\\mathbf{F}$ and thereby the learning of the exposure synthesis operation $\\mathcal{G}$ efficiently.", + "bbox": [ + 75, + 659, + 468, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Efficient Modelling of Filter Subspace. Specifically, we formulate the filter parameters $\\mathbf{F}$ using the atom-coefficient [56] decomposition, which permits the smallest computational bottleneck in modeling filter subspaces among the candidates. Given a convolutional filter $\\mathbf{F} \\in \\mathbb{R}^{c' \\times c \\times k \\times k}$ where $c, c'$ and $k$ are the output, input channels, and spatial size, respectively. The atom-coefficient decomposition over a set of $m$ filter atoms decomposes $\\mathbf{F}$ into coefficients, denoted by $\\phi \\in \\mathbb{R}^{c' \\times c \\times m}$ , and atoms, denoted", + "bbox": [ + 75, + 763, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\text {b y} \\Lambda \\in \\mathbb {R} ^ {m \\times k \\times k} \\\\ \\mathbf {F} = \\phi \\boldsymbol {\\Lambda}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 498, + 89, + 728, + 133 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When $m \\ll c$ , the dimension of the parameters is significantly reduced. Extending the decomposition to a L-layer convolutional network gives a set of $L$ filter atoms $\\Lambda = \\{\\Lambda^l\\}_{l=1}^L$ and $L$ coefficients $\\phi = \\{\\phi^l\\}_{l=1}^L$ .", + "bbox": [ + 498, + 148, + 890, + 209 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Enforcing Filter Atom Continuity. We enforce the filter atoms $\\Lambda$ to smoothly change according to the input and target exposure levels $\\widetilde{\\theta}_{\\mathrm{input}}$ and $\\widetilde{\\theta}_{\\mathrm{target}}$ . To do so, we incorporate the continuous-time networks like neural ODEs to describe the relationship of $\\Lambda (\\widetilde{\\theta}_{\\mathrm{input}},\\widetilde{\\theta}_{\\mathrm{target}})$ , as the neural ODEs provide advantages in the number of parameters required, memory consumption, and computational adaptivity [6]. The derivative of the filter atoms $d\\Lambda /d\\widetilde{\\theta}$ w.r.t. the exposure value $\\widetilde{\\theta}$ as a neural network $g(\\cdot)$", + "bbox": [ + 496, + 209, + 890, + 349 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nd \\boldsymbol {\\Lambda} / d \\widetilde {\\theta} = g (\\widetilde {\\theta} _ {\\text {i n p u t}}, \\widetilde {\\theta} _ {\\text {t a r g e t}}, \\boldsymbol {\\Lambda} _ {\\text {i n i t}}; \\mathcal {W}),\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 362, + 812, + 381 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\Lambda_{\\mathrm{init}}$ is the initial state of filter atoms, and $\\mathcal{W}$ is the learnable parameters of $g(\\cdot)$ . The initial condition $\\Lambda_{\\mathrm{init}}$ is a learnable parameter and is optimized jointly with $\\mathcal{W}$ . Given the input and target exposure values $\\widetilde{\\theta}_{\\mathrm{input}}$ and $\\widetilde{\\theta}_{\\mathrm{target}}$ , the corresponding filter atoms can be calculated as", + "bbox": [ + 496, + 392, + 890, + 470 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Lambda} (\\widetilde {\\theta} _ {\\text {t a r g e t}}, \\widetilde {\\theta} _ {\\text {i n p u t}}) = \\boldsymbol {\\Lambda} _ {\\text {i n i t}} + \\int_ {\\widetilde {\\theta} _ {\\text {i n p u t}}} ^ {\\widetilde {\\theta} _ {\\text {t a r g e t}}} g (\\boldsymbol {\\Lambda} _ {\\text {i n i t}}, \\mathcal {W}) d \\widetilde {\\theta}.\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 481, + 854, + 521 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "According to Picard-Lindelöf existence theorem [15], for a given initial condition, the solution to the IVP is unique. Theoretically, we prove continuity in filter atoms induces controlled changes in image exposure.", + "bbox": [ + 496, + 531, + 890, + 592 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 3.1 If $\\Lambda_{1}$ and $\\Lambda_{2}$ are filter atoms generated with neural ODE using integration intervals $(\\widetilde{\\theta}_{0},\\widetilde{\\theta}_{l})$ and $(\\widetilde{\\theta}_0,\\widetilde{\\theta}_2)$ , respectively such that $\\| \\Lambda_1 - \\Lambda_2\\| \\leq \\epsilon |\\widetilde{\\theta}_1 - \\widetilde{\\theta}_2|$ , for $\\epsilon >0$ . If $\\theta_{1}$ and $\\theta_{2}$ are the exposures at $\\widetilde{\\theta}_{1}$ and $\\widetilde{\\theta}_{2}$ , and activation function $\\sigma$ is non-expansive, then change in exposure $\\Delta \\theta$ is continuous in filter atoms.", + "bbox": [ + 496, + 606, + 890, + 699 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Details of Theorem 3.1 and its proof can be found in the Appendix section A1. An illustration of our parameter efficient approach is shown in Figure 3.", + "bbox": [ + 496, + 712, + 890, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Exposure Adaptive Colorization", + "text_level": 1, + "bbox": [ + 500, + 768, + 782, + 784 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our proposed exposure adaptive colorization consists of an exposure correction realized by the stated synthesis and exposure level specific colorizer.", + "bbox": [ + 496, + 791, + 890, + 835 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Exposure correction. Given a colorization method specific to exposure $\\widetilde{\\theta}_{\\mathrm{target}}$ and a input sample $(\\mathbf{Y}_i,\\widetilde{\\theta}_{\\mathrm{input}})$ , the proposed exposure correction generates image $\\hat{\\mathbf{Y}}_t$ at the target exposure in two steps. With the operation $\\mathcal{G}$ carried out", + "bbox": [ + 496, + 837, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "25141", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e285dcaa0552d4ee4cbacedca07a215de1f7c4629c2282772c00c7ae5a8cf969.jpg", + "image_caption": [ + "Figure 3. An illustration of our proposed method for exposure adaptive colorization: A binary image, $\\mathbf{Y}$ , which can be overexposed or underexposed, is input into the proposed exposure synthesis module. The colorization of binary image can be achieved using Single Image Colorization (SIC) or Burst Image Colorization (BIC). (i) SIC: Based on the input and target exposure levels, $\\widetilde{\\theta}_{\\mathrm{input}}$ and $\\widetilde{\\theta}_{\\mathrm{target}}$ , this module adjusts the weights of a exposure synthesis network $\\mathcal{G}$ , which then generates an exposure-corrected image. Note that corrected image is not necessarily a binary image. Since the colorization module $\\mathcal{F}$ is trained to colorize only image of specific exposure, the exposure synthesis module ensures corrected binary image has similar exposure to the one on which $\\mathcal{F}$ is trained. (ii) BIC: For BIC we generate images with varying exposures as input to the burst image colorization network, $\\mathcal{F}_{\\mathrm{burst}}$ . The trained network is able to exploit the complementary information across multiple exposures with the help of Cross Non-Local Fusion blocks [44] to synthesize colors in regions of the image that is otherwise not possible by SIC approach." + ], + "image_footnote": [], + "bbox": [ + 101, + 87, + 890, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "by some generative network, first, the filter atoms of the network's convolutional filters are computed using the neural ODEs based on the $(\\widetilde{\\theta}_{\\mathrm{input}},\\widetilde{\\theta}_{\\mathrm{target}})$ pair; next, the convolutional filter weights of the network are updated using the filter atoms and their corresponding coefficients, and the network generate the exposure corrected image $\\hat{\\mathbf{Y}}_t$ based on $\\mathbf{Y}_i$ . In the case of burst colorization, multiple images are generated corresponding to a set of target exposures.", + "bbox": [ + 75, + 484, + 468, + 607 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Colorization. Both single-image-based colorization and burst-based colorization are considered in our work. The first approach involves a colorization network $\\mathcal{F}:\\hat{\\mathbf{Y}}_t\\to$ $\\mathbf{X}_c$ which takes a exposure corrected binary image as input and outputs corresponding RGB image $\\mathbf{X}_c$ . The network is trained exclusively at a fixed exposure level, and variations in the exposure of the input image is pre-adjusted by the exposure correction step. The second approach uses a colorization network $\\mathcal{F}_{\\mathrm{burst}}:\\{\\hat{\\mathbf{Y}}_t\\}_{t = 1}^K\\to \\mathbf{X}_c$ that takes in a burst of $K$ binary images $\\{\\hat{\\mathbf{Y}}_t\\}_{t = 1}^K$ corresponding to predefined varying exposure values $\\{\\hat{\\theta}_t\\}_{t = 1}^K$ and outputs one RGB image $\\mathbf{X}_c$ .", + "bbox": [ + 75, + 607, + 470, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 806, + 209, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 76, + 830, + 266, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate the performance of both the exposure synthesis module standalone and the entire proposed approach. We conduct exposure burst recovery experiment to evalu", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ate the exposure synthesis and Single Image Colorization (SIC) and Burst Image Colorization (BIC) experiments to evaluate the final colorization results.", + "bbox": [ + 498, + 484, + 890, + 529 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation of Exposure Adaptive Colorization. We implement the neural ODEs using a 6-layer multilayer perceptron. The generative networks we use for exposure synthesis are variants of Pix2Pix [34] and CycleGAN [68], with which the models are named AtomODE-Pix2Pix and AtomODE-CycleGAN. For colorization, we train a Pix2Pix colorizer dedicated to SIC task and another to BIC. For BIC, the colorizer is additionally equipped with a Cross Non-Local Fusion (CNLF) [44] to constructively fuse the representations obtained from the burst. Note that the colorizers are trained directly on pairs of binary and color image data at fixed $\\widetilde{\\theta}$ , independent of exposure synthesis. In our colorization experiment, the same trained colorizer is applied on all exposure correction methods for comparison. Details can be found in the Appendix section A2.", + "bbox": [ + 496, + 532, + 892, + 760 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets and Synthetic Data Generation. We train and evaluate the proposed method as well as the baselines on synthetic data generated using images from two datasets: AFHQ [14] and CelebA-HQ [36]. AFHQ dataset contains images of three classes, cats, dogs, and wild animals, and CelebA-HQ contains images of male and female human faces.", + "bbox": [ + 496, + 762, + 890, + 867 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To synthesize binary images, we use the image formation model described in Section 3.1 and apply exposure bracket", + "bbox": [ + 500, + 869, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "25142", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/39e4ee4f0f7c272e37bb46f576a893cb877a5cb36e3b831dcaa1866cdd58fb0c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Exposure Synthesis MethodAFHQ (512 x 512)CelebA-HQ (256 x 256)
CatDogWildMaleFemale
MSE (↓)/RL (↑)/FID (↓)MSE (↓)/RL (↑)/FID (↓)MSE (↓)/RL (↑)/FID (↓)MSE (↓)/RL (↑)/FID (↓)MSE (↓)/RL (↑)/FID (↓)
Pix2Pix-DNI3.75 / 0.9816 / 61.895.91 / 0.9844 / 95.438.13 / 0.9780 / 39.293.74 / 0.9563 / 56.134.78 / 0.9681 / 39.81
CycleGAN-DNI9.61 / 0.9816 / 207.8216.79 / 0.9824 / 220.909.62 / 0.9820 / 336.389.31 / 0.9824 / 94.7814.26 / 0.9810 / 250.61
SAVI2I41.15 / 46.27 / 178.2946.27 / 0.9804 / 217.9838.87 / 0.9019 / 139.4767.26 / 0.9563 / 111.0569.21 / 0.9698 / 261.45
DLOW2.44 / 0.9654 / 92.302.32 / 0.9489 / 173.652.75 / 0.9632 / 179.231.87 / 0.9655 / 62.571.37 / 0.9798 / 63.31
AtomODE-Pix2Pix (ours)2.29 / 0.9997 / 57.883.17 / 0.9994 / 140.232.15 / 0.9998 / 79.762.15 / 0.9995 / 53.443.20 / 0.9995 / 40.08
AtomODE-CycleGAN (ours)1.15 / 0.9948 / 60.541.18 / 0.9971 / 87.431.23 / 0.9979 / 35.751.58 / 0.9986 / 77.821.87 / 0.9994 / 62.59
", + "bbox": [ + 78, + 89, + 893, + 188 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Colorization results on AFHQ and CelebA-HQ datasets. The input image for all these experiments is an overexposed image at index $\\tau = 0$ of exposure burst. All the colorizers are trained using an image corresponding to $\\tau = 10$ in the exposure burst. At test time we correct the overexposed input to correspond to the one used during training. In the first column of the table, we note the combination of the exposure correction method used and colorizer is same across all the methods.", + "bbox": [ + 75, + 194, + 893, + 250 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/435800515c977b9ae215fe545e14af136908b2fb5937ebbdd23b1f132a724d40.jpg", + "image_caption": [ + "Figure 4. Qualitative results for exposure burst recovery experiments for AFHQ dataset. We compare our methods, AtomODE-Pix2Pix and AtomODE-CycleGAN, with DLOW, SAVI2I and DNI." + ], + "image_footnote": [], + "bbox": [ + 81, + 260, + 468, + 661 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ing to generate various exposure. Details of data synthesis can be found in the Appendix section A2. Our compiled test datasets consist of 15-frame bursts of exposure varying from overexposed to underexposed. We index the exposure of these images in a burst by $\\tau$ , a discretized label of $\\widetilde{\\theta}$ .", + "bbox": [ + 75, + 731, + 468, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. In the exposure burst recovery experiment, we compare our approach against DLOW [25] and SAVI2I [52]. We further consider DNI [62] for its capability of interpolating network parameters to generate continuous variations of outputs. We prepare baselines Pix2Pix-DNI and CycleGAN-DNI by coupling DNI with Pix2Pix and Cycle", + "bbox": [ + 75, + 810, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "GAN exposure generators. In colorization experiments, we apply the independently trained Pix2Pix colorizers to the exposure correction results generated by all methods. For more details on baseline training and hyperparameters, we refer readers to the Appendix section A2.", + "bbox": [ + 496, + 263, + 890, + 338 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. In the burst recovery experiment, we choose to use Mean Squared Error (MSE) and Relative Linearity (RL) [43] to distinguish a smoothly varying exposure burst with consistent appearance changes. A high RL score indicates a better transition in exposure among consecutive samples. In the colorization experiments, we assess the quality of exposure correction by computing the Fréchet Inception Distance (FID) between the colorized image and ground truth.", + "bbox": [ + 496, + 339, + 892, + 459 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Exposure Burst Recovery Results", + "text_level": 1, + "bbox": [ + 500, + 483, + 792, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 4 presents the qualitative comparison of all methods on AFHQ dataset. We find our methods, AtomODE Pix2Pix and AtomODE CycleGAN are able to generate consistent set of smoothly varying exposure images than baseline methods and match ground truth images generated from exposure bracketing. We also observe some baselines struggle with generation of overexposed frames. All methods for comparison are further assessed on both datasets using MSE and RL, with the quantitative results shown in Table 2. Results suggest that the proposed AtomODE provides consistently better performance in all testing conditions across multiple datasets. Additional qualitative results can be found in Appendix section A4.", + "bbox": [ + 496, + 506, + 890, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Single and Burst Image Colorization Results", + "text_level": 1, + "bbox": [ + 498, + 710, + 877, + 727 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The SIC colorization results of all methods on AFHQ dataset are shown in Figure 5, where the exposure of the test samples is different from the that of the colorizer's training samples. Corrupted colored images are generated by most of the baselines, while the proposed method produces visually appealing outputs. Figure 7 shows the SIC colorization results on CelebA-HQ dataset. Additionally, we also present the qualitative results for BIC in Figure 6 where $K$ denotes the number of binary frames input to the colorizer. The quantitative results are shown in Table 2, assessed with the FID metric.", + "bbox": [ + 496, + 734, + 890, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "25143", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/548330aef06c66cc200aa89566c82076cec156c5839c7327bbb55a0d36983a12.jpg", + "image_caption": [ + "Figure 5. Colorized results for AFHQ dataset using various combinations of exposure correction methods and colorizer. The first column shows the overexposed binary input image for exposure correction. The results of our methods AtomODE Pix2Pix and AtomODE CycleGAN are shown in columns six and seven." + ], + "image_footnote": [], + "bbox": [ + 81, + 87, + 890, + 664 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Real Data Experiments", + "text_level": 1, + "bbox": [ + 76, + 724, + 290, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As our approach is trained on synthetic data, it is essential to evaluate its performance on real-world data. To this end, we conduct experiments using both a CMOS camera and a prototype QIS camera. We refer the readers to Appendix for further details about the experiment setup. We print 100 images from the training set and 25 from the test set, and we capture pictures of them using both cameras, collecting 125 pairs of grayscale and RGB images. We fine-tune the colorization model with the 100 real data samples, and we evaluate our proposed approach on the set of 25 test data, in", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "cluding both binarized CMOS data and QIS data. In Figure 8, we present qualitative results from the evaluation on the real data obtained by the CMOS camera. Despite that the proposed method was not trained with vast real sensor data, it is able to produce visually pleasing colorized images. We observe that fine-tuning the model using a small training dataset significantly improves the colorization quality. We further present results from the evaluation on images captured by the QIS camera in Figure 9. These results demonstrate the potential of the proposed method for real-world single-photon imaging applications.", + "bbox": [ + 496, + 724, + 893, + 893 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "25144", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4c800f0b820fd94bab360555cc73142aedd6a6e07fa8f77a825db0c3fb3ad13a.jpg", + "image_caption": [ + "Groundtruth $K = 5$ $K = 1$" + ], + "image_footnote": [], + "bbox": [ + 104, + 94, + 467, + 263 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/678f953f4e3d069d606f7f18b9f3f692d3fe60ed38c26f9723a3671b7762c12b.jpg", + "image_caption": [ + "Figure 6. An illustration of burst image colorization for AFHQ dataset using multiple binary images of varying exposures as input. The colorizer is trained using images of exposure $\\tau = \\{0,4,8,10,14\\}$ . An overexposed input image is used to synthesize multiple images of varying exposures and passed to the colorizer.", + "Figure 7. Colorized results for CelebA-HQ dataset using various combinations of exposure correction methods and colorizers. The results of our method are shown in columns four and five." + ], + "image_footnote": [], + "bbox": [ + 81, + 380, + 462, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Benefits for Practical Applications", + "text_level": 1, + "bbox": [ + 76, + 709, + 375, + 726 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our methodology is specifically designed with modularity and parameter efficiency in mind, features crucial for real-world industry applications. The Pix2PixDNI method, the second-best performing approach, necessitates training multiple networks and either interpolating model weights on-the-fly during inference or precomputing and storing them to generate images with varied exposures. Conversely, our method circumvents the additional storage requirements and the need to train multiple model parameters by incorporating the exposure correction into the neural ODE framework.", + "bbox": [ + 75, + 734, + 468, + 898 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e76a678abea8de61fab3bce9e4dcc81f118d15b6886ffcd9854fe69d984c9e0f.jpg", + "image_caption": [ + "Figure 8. Results on real-world images captured from a CMOS camera (images shown along the row with a red line are colorized images, and along the green line are the ground-truth images captured using an RGB CMOS camera)." + ], + "image_footnote": [], + "bbox": [ + 501, + 90, + 883, + 321 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7f5b1a31e523c5b163bc17b15a131fb6ada6f0d10867c42e87a120f5e9a4a8c3.jpg", + "image_caption": [ + "Figure 9. Colorized output using input from prototype QIS camera." + ], + "image_footnote": [], + "bbox": [ + 501, + 407, + 887, + 549 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 606, + 617, + 622 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we proposed a generative exposure correction approach that can synthesize a spectrum of exposure given a single 1-bit image. We modeled the binary signal variation with respect to exposure change as a continuous function by adapting a small set of parameters of a network, known as filter atoms, to desired exposure using neural ODE. Theoretically and empirically, we demonstrated successful recovery of bursts of 1-bit images from a single input image, and we verified its application on single-image-based and burst-image-based 1-bit image colorization. Experiments on real CMOS and QIS data further indicated the potentials of the proposed approach on real-world single-photon imaging applications.", + "bbox": [ + 496, + 632, + 892, + 829 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "25145", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Samuel Burri, Yuki Maruyama, Xavier Michalet, Francesco Regazzoni, Claudio Bruschini, and Edoardo Charbon. Architecture and applications of a high resolution gated SPAD image sensor. Optics Express, 22(14):17573-17589, 2014. 1, 2", + "[2] Stanley H. Chan. What does a one-bit Quanta Image Sensor offer? IEEE Transactions on Computational Imaging, 8: 770-783, 2022. 2", + "[3] Stanley H. Chan and Yue M. Lu. Efficient image reconstruction for gigapixel quantum image sensors. In IEEE Global Conference on Signal and Information Processing, pages 312-316, 2014. 2", + "[4] Stanley H. Chan, Omar A. Elgendy, and Xiran Wang. Images from bits: Non-iterative image reconstruction for Quanta Image Sensors. MDPI Sensors, 16(11), 2016. 2", + "[5] Stanley H. Chan, Xiran Wang, and Omar A. Elgendy. Plug-and-Play ADMM for image restoration: Fixed-point convergence and applications. IEEE Transactions on Computational Imaging, 3:84–98, 2016. 2", + "[6] Ricky T. Q. Chen, Yulia Rubanova, Jesse Bettencourt, and David K. Duvenaud. Neural Ordinary Differential Equations. In NeurIPS, 2018. 4", + "[7] Wei Chen, Zichen Miao, and Qiang Qiu. Inner product-based neural network similarity. In NeurIPS, 2023. 2", + "[8] Wei Chen, Zichen Miao, and Qiang Qiu. Large convolutional model tuning via filter subspace. Available online: https://arxiv.org/abs/2403.00269, 2024. Accessed 03/06/2024. 2", + "[9] Ying-Cong Chen, Xiaogang Xu, Zhuotao Tian, and Jiaya Jia. Homomorphic latent space interpolation for unpaired Image-To-Image translation. In CVPR, pages 2403-2411, 2019. 2", + "[10] Xiuyuan Cheng, Qiang Qiu, Robert Calderbank, and Guillermo Sapiro. RotDCF: Decomposition of convolutional filters for rotation-equivariant deep networks. In ICLR, 2019. 2", + "[11] Yiheng Chi, Abhiram Gnanasambandam, Vladlen Koltun, and Stanley H. Chan. Dynamic low-light imaging with Quanta Image Sensors. In ECCV, pages 122-138, 2020. 1, 2", + "[12] Joon Hee Choi, Omar A. Elgendy, and Stanley H. Chan. Image reconstruction for Quanta Image Sensors using deep neural networks. In IEEE ICASSP, pages 6543-6547, 2018. 2", + "[13] Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. StarGAN: Unified Generative Adversarial Networks for multi-domain Image-to-Image translation. In CVPR, pages 8789–8797, 2018. 2", + "[14] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. StarGAN v2: Diverse image synthesis for multiple domains. In CVPR, pages 8185-8194, 2020. 2, 5", + "[15] Earl A. Coddington and Norman Levinson. Theory of Ordinary Differential Equations. McGraw-Hill, 1955. 4", + "[16] Neale A. W. Dutton, Istvan Gyongy, Luca Parmesan, Salvatore Gnecchi, Neil Calder, Bruce R. Rae, Sara Pellegrini, Lindsay A. Grant, and Robert K. Henderson. A SPAD-based QVGA image sensor for single-photon counting and quanta" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "imaging. IEEE Transactions on Electron Devices, 63(1): 189-196, 2016. 1, 2", + "[17] Neale A. W. Dutton, Istvan Gyongy, Luca Parmesan, and Robert K. Henderson. Single photon counting performance and noise analysis of CMOS SPAD-based image sensors. MDPI Sensors, 16(7), 2016. 1, 2", + "[18] Omar A. Elgendy, Abhiram Gnanasambandam, Stanley H. Chan, and Jiaju Ma. Low-light demosaicking and denoising for small pixels using learned frequency selection. IEEE Transactions on Computational Imaging, 7:137-150, 2021. 2", + "[19] Eric R. Fossum. Some thoughts on future digital still cameras. In Image Sensors and Signal Processing for Digital Still Cameras, pages 305-314. Taylor and Francis Group, 2005. 2", + "[20] Eric R. Fossum. Modeling the performance of single-bit and multi-bit Quanta Image Sensors. IEEE Journal of the Electron Devices Society, 1(9):166-174, 2013. 2", + "[21] Eric R. Fossum. Analog read noise and quantizer threshold estimation from Quanta Image Sensor bit density. IEEE Journal of the Electron Devices Society, 10:269-274, 2022. 2", + "[22] Abhiram Gnanasambandam and Stanley H. Chan. Image classification in the dark using Quanta Image Sensors. In ECCV, pages 484-501, 2020. 2", + "[23] Abhiram Gnanasambandam and Stanley H. Chan. HDR imaging with Quanta Image Sensors: Theoretical limits and optimal reconstruction. IEEE Transactions on Computational Imaging, 6:1571-1585, 2020. 1", + "[24] Abhiram Gnanasambandam and Stanley H. Chan. Exposure-referred signal-to-noise ratio for digital image sensors. IEEE Transactions on Computational Imaging, 8:561-575, 2022. 1", + "[25] Rui Gong, Wen Li, Yuhua Chen, and Luc Van Gool. DLOW: Domain flow for adaptation and generalization. In CVPR, pages 2472-2481, 2019. 6", + "[26] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative Adversarial Nets. In NeurIPS, 2014. 2", + "[27] Anant Gupta, Atul Ingle, and Mohit Gupta. Asynchronous single-photon 3D imaging. In ICCV, pages 7908-7917, 2019. 2", + "[28] Anant Gupta, Atul Ingle, Andreas Velten, and Mohit Gupta. Photon-flooded single-photon 3D cameras. In CVPR, pages 6763-6772, 2019.", + "[29] Felipe Gutierrez-Barragan, Atul Ingle, Trevor Seets, Mohit Gupta, and Andreas Velten. Compressive single-photon 3D cameras. In CVPR, pages 17833-17843, 2022. 2", + "[30] Istvan Gyongy, Neale AW Dutton, and Robert K Henderson. Single-photon tracking for high-speed vision. MDPI Sensors, 18:323, 2018. 2", + "[31] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 2", + "[32] Atul Ingle, Andreas Velten, and Mohit Gupta. High flux passive imaging with single-photon sensors. In CVPR, pages 6753-6762, 2021. 2" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "25146", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-Image translation with conditional adversarial networks. In CVPR, pages 5967–5976, 2017. 2", + "[34] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A. Efros. Image-to-Image translation with conditional adversarial networks. In CVPR, pages 5967–5976, 2017. 3, 5", + "[35] Kiyotaka Iwabuchi, Yusuke Kameda, and Takayuki Hamamoto. Image quality improvements based on motion-based deblurring for single-photon imaging. IEEE Access, 9:30080-30094, 2021. 2", + "[36] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In ICLR, 2018. 5", + "[37] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for Generative Adversarial Networks. IEEE TPAMI, 43(12):4217-4228, 2021. 2", + "[38] Manoj Kumar, Dirk Weissenborn, and Nal Kalchbrenner. Colorization transformer. In ICLR, 2021. 3", + "[39] Vadim Lebedev, Yaroslav Ganin, Maksim Rakhuba, Ivan Oseledets, and Victor Lempitsky. Speeding-up convolutional neural networks using fine-tuned CP-decomposition. In ICLR, 2015. 2", + "[40] Yu-Jing Lin, Po-Wei Wu, Che-Han Chang, Edward Chang, and Shih-Wei Liao. RelGAN: Multi-domain Image-to-Image translation via relative attributes. In ICCV, pages 5913–5921, 2019. 2", + "[41] Wallace Lira, Johannes Merz, Daniel Ritchie, Daniel Cohen-Or, and Hao Zhang. GANHopper: Multi-hop GAN for unsupervised Image-to-Image translation. In ECCV, pages 363–379, 2020. 2", + "[42] Yahui Liu, Enver Sangineto, Yajing Chen, Linchao Bao, Haoxian Zhang, Nicu Sebe, Bruno Lepri, Wei Wang, and Marco De Nadai. Smoothing the disentangled latent style space for unsupervised Image-to-Image translation. In CVPR, pages 10780–10789, 2021. 2", + "[43] Kun Lu, Rongpeng Li, and Honggang Zhang. Contrastive monotonic pixel-level modulation. In ECCV, pages 493-510, 2022. 6", + "[44] Ziwei Luo, Lei Yu, Xuan Mo, Youwei Li, Lanpeng Jia, Haoqiang Fan, Jian Sun, and Shuaicheng Liu. EBSR: Feature Enhanced Burst Super-Resolution with deformable alignment. In CVPR, pages 471–478, 2021. 5", + "[45] Jiaju Ma and Eric Fossum. Quanta Image Sensor jot with sub 0.3e- r.m.s. read noise and photon counting capability. IEEE Electron Device Letters, 36(9):926-928, 2015. 1", + "[46] Jiaju Ma and Eric R. Fossum. A Pump-Gate Jot device with high conversion gain for a Quanta Image Sensor. IEEE Journal of the Electron Devices Society, 3:73-77, 2015.", + "[47] Jiaju Ma, Dakota Starkey, Arun Rao, Kofi Odame, and Eric R. Fossum. Characterization of Quanta Image Sensor pump-gate jots with deep sub-electron read noise. IEEE Journal of the Electron Devices Society, 3(6):472-480, 2015. 1", + "[48] Jiaju Ma, Saleh Masoodian, Dakota A Starkey, and Eric R. Fossum. Photon-number-resolving megapixel image sensor at room temperature without avalanche gain. Optica, 4(12): 1474-1481, 2017. 1" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Jiaju Ma, Dexue Zhang, Omar Elgendy, and Saleh Masoodian. A 0.19e-rms read noise 16.7mpixel stacked Quanta Image Sensor with $1.1\\mu \\mathrm{m}$ -pitch backside illuminated pixels. IEEE Electron Device Letters, 42(6):891-894, 2021. 1", + "[50] Jiaju Ma, Stanley H. Chan, and Eric R. Fossum. Review of Quanta Image Sensors for ultralow-light imaging. IEEE Transactions on Electron Devices, 69(6):2824-2839, 2022. 1, 2", + "[51] Sizhuo Ma, Paul Mos, Edoardo Charbon, and Mohit Gupta. Burst vision using single-photon cameras. In WACV, pages 5364-5374, 2021. 2", + "[52] Qi Mao, Hsin-Ying Lee, Hung-Yu Tseng, Jia-Bin Huang, Siwei Ma, and Ming-Hsuan Yang. Continuous and diverse Image-to-Image translation via signed attribute vectors. IJCV, 130(2):517-549, 2020. 6", + "[53] Zichen Miao, Ze Wang, Xiuyuan Cheng, and Qiang Qiu. Spatiotemporal joint filter decomposition in 3D convolutional neural networks. In NeurIPS, 2021. 2", + "[54] Zichen Miao, Ze Wang, Wei Chen, and Qiang Qiu. Continual learning with filter atom swapping. In ICLR, 2022. 2", + "[55] Fabio Pizzati, Pietro Cerri, and Raoul de Charette. CoMoGAN: Continuous model-guided Image-to-Image translation. In CVPR, pages 14283–14293, 2021. 2", + "[56] Qiang Qiu, Xiuyuan Cheng, Robert Calderbank, and Guillermo Sapiro. DCFNet: Deep neural network with decomposed convolutional filters. In ICML, 2018. 2, 4", + "[57] Trevor Seets, Atul Ingle, Martin Laurenzis, and Andreas Velten. Motion adaptive deblurring with single-photon cameras. In WACV, pages 1944–1953, 2021. 2", + "[58] Jheng-Wei Su, Hung-Kuo Chu, and Jia-Bin Huang. Instance-aware image colorization. In CVPR, pages 7965–7974, 2020. 3", + "[59] Varun Sundar, Sizhuo Ma, Aswin Sankaranarayanan, and Mohit Gupta. Single-photon structured light. In CVPR, pages 17844-17854, 2022. 2", + "[60] Patricia Vitoria, Lara Raad, and Coloma Ballester. ChromaGAN: Adversarial picture colorization with semantic class distribution. In WACV, pages 2434-2443, 2020. 3", + "[61] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional GANs. In CVPR, pages 8798-8807, 2018. 2, 4", + "[62] X. Wang, K. Yu, C. Dong, X. Tang, and C. Loy. Deep network interpolation for continuous imagery effect transition. In CVPR, pages 1692-1701, 2019. 6", + "[63] Ze Wang. Efficient Adaptation of Deep Vision Models. PhD thesis, Purdue University, 2023. Available online: https://hammer.purdue.edu/articles/thesis/Efficient_Adaptation_of_Deep_Vision_CModels/22696129.2", + "[64] Ze Wang, Xiuyuan Cheng, Guillermo Sapiro, and Qiang Qiu. Stochastic conditional generative networks with basis decomposition. In ICLR, 2020. 2, 4", + "[65] Ze Wang, Seunghyun Hwang, Zichen Miao, and Qiang Qiu. Image generation using continuous filter atoms. In NeurIPS, 2021. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "25147", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[66] Yanze Wu, Xintao Wang, Yu Li, Honglun Zhang, Xun Zhao, and Ying Shan. Towards vivid and diverse image colorization with generative color prior. In ICCV, pages 14357-14366, 2021. 3", + "[67] Zhaoyang Yin, Jiaju Ma, Saleh Masoodian, and Eric R Fossum. Threshold uniformity improvement in 1b Quanta Image Sensor readout circuit. MDPI Sensors, 22(7):2578, 2022. 2", + "[68] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In ICCV, pages 2242-2251, 2017. 2, 5", + "[69] Wei Zhu, Qiang Qiu, Robert Calderbank, Guillermo Sapiro, and Xiuyuan Cheng. Scaling-translation-equivariant networks with decomposed convolutional filters. Journal of Machine Learning Research, 23(68):1-45, 2022. 2" + ], + "bbox": [ + 78, + 90, + 468, + 301 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "25148", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Generative Quanta Color Imaging/9125a07a-d19f-455b-bcb9-5f2fd8cc25e9_model.json b/2024/Generative Quanta Color Imaging/9125a07a-d19f-455b-bcb9-5f2fd8cc25e9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4be0e8c27291c1fa6bb27a1af8e6942f4b7b274f --- /dev/null +++ b/2024/Generative Quanta Color Imaging/9125a07a-d19f-455b-bcb9-5f2fd8cc25e9_model.json @@ -0,0 +1,2136 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.311, + 0.131, + 0.66, + 0.154 + ], + "angle": 0, + "content": "Generative Quanta Color Imaging" + }, + { + "type": "text", + "bbox": [ + 0.182, + 0.181, + 0.79, + 0.234 + ], + "angle": 0, + "content": "Vishal Purohit, Junjie Luo, Yiheng Chi, Qi Guo, Stanley H. Chan, Qiang Qiu Elmore Family School of Electrical and Computer Engineering Purdue University, West Lafayette, USA." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.237, + 0.725, + 0.252 + ], + "angle": 0, + "content": "{purohitv,luo330,chi14,guo675,stanchan,qqiu}@purdue.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.319, + 0.474, + 0.622 + ], + "angle": 0, + "content": "The astonishing development of single-photon cameras has created an unprecedented opportunity for scientific and industrial imaging. However, the high data throughput generated by these 1-bit sensors creates a significant bottleneck for low-power applications. In this paper, we explore the possibility of generating a color image from a single binary frame of a single-photon camera. We evidently find this problem being particularly difficult to standard colorization approaches due to the substantial degree of exposure variation. The core innovation of our paper is an exposure synthesis model framed under a neural ordinary differential equation (Neural ODE) that allows us to generate a continuum of exposures from a single observation. This innovation ensures consistent exposure in binary images that colorizers take on, resulting in notably enhanced colorization. We demonstrate applications of the method in single-image and burst colorization and show superior generative performance over baselines. Project website can be found at https://vishal-s-p.github.io/projects/2023/generative_quanta_color.html" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.632, + 0.21, + 0.647 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.658, + 0.47, + 0.856 + ], + "angle": 0, + "content": "Single-photon image sensors are gaining a strong momentum over the past decade due to their impeccable photon-counting capability. This not only opens the door for advanced scientific imaging in low-light environment, but also creates the opportunity for new imaging in high-speed, high dynamic range, and low-bit conditions. Single-photon imaging today are realized through single-photon avalanche diodes (SPAD) and quanta image sensors (QIS). The latest development of QIS has demonstrated a mega-pixel resolution [48] which is compatible with mainstream cell phone cameras, whereas SPAD has achieved hundreds of thousands of (binary) frames per second which has enabled a plethora of industrial and 3D imaging applications [16]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Problem Statement. In the context of 1-bit single-photon imaging, the data volume generated is huge. Taking a \\(1000 \\times 1000\\) SPAD array as an example, a video of" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.286, + 0.779, + 0.297 + ], + "angle": 0, + "content": "Binary Input Pix2PixDNI Our Method Groundtruth" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.299, + 0.781, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.794, + 0.297, + 0.892, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.473, + 0.893, + 0.598 + ], + "angle": 0, + "content": "Figure 1. We introduce generative quanta color imaging. Given a binary frame captured by a single-photon camera (quanta image sensor (QIS) in this example), the proposed method generates a continuum of exposures using a neural ordinary differential equation framework. Color are then generated based on these exposures. Left: A qualitative comparison between our approach and existing methods is shown. Right: Comparison between colorization results and image captured using a RGB CMOS camera under similar conditions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.618, + 0.893, + 0.814 + ], + "angle": 0, + "content": "500 binary frames requires a \\(500\\mathrm{Mb / s}\\) data transfer rate; if the frame rate becomes \\(150\\mathrm{k}\\) fps, the data transfer rate increases to \\(150\\mathrm{Gb / s}\\). Since data throughput scales proportionally with the power consumption, it quickly becomes infeasible to use these sensors continuously without considering power supply, memory, and transmission issues. While solid progress are being made across the sensor and signal processing community [1, 11, 17, 23, 24, 45-47, 49, 50], the data throughput remains an unsolved problem for single-photon sensors to be considered applicable in many applications, including low-power devices, wireless devices, underwater vehicles, spacecrafts, etc., where bandwidth is the bottleneck." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.817, + 0.892, + 0.848 + ], + "angle": 0, + "content": "In this paper, instead of compressing data or redesigning the sensor, we ask the question" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.863, + 0.868, + 0.894 + ], + "angle": 0, + "content": "Is it possible to recover a color image from a single 1-bit image?" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25138" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.332 + ], + "angle": 0, + "content": "Significance. There are several reasons why this problem is significant. Firstly, while there is an ocean of literature on image generation (through methods such as Generative Adversarial Networks [26] and diffusion [31]), our problem is unique in the sense none of the existing work addresses it. Besides being binary (and hence colorless), in practice the input data usually also suffers from under- or over-exposures. Therefore, before generating color, it is even more critical to synthesize the unseen exposure. Secondly, any solution offered to our problem will provide a path to generative imaging in the future. From the point of view of compressed acquisition, our method demonstrates the extreme case of compression where the incoming data is a single bit. For future AR/VR applications where power is severely limited, our approach shows the possibility of extreme data compression." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.333, + 0.47, + 0.363 + ], + "angle": 0, + "content": "Contributions. The core algorithmic contribution of this paper is two-fold" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.364, + 0.47, + 0.468 + ], + "angle": 0, + "content": "- We propose a neural ordinary differential equation (neural ODE) based framework to synthesize a continuum of exposures that are not available in the measurement. By controlling the integration interval of neural ODE, we obtain convolutional filters that are responsible for converting input representation to image representation of desired exposure." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.469, + 0.469, + 0.514 + ], + "angle": 0, + "content": "- We theoretically and empirically show that a controlled variation of filter atoms leads to controlled exposure changes in the output image." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.364, + 0.47, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.515, + 0.47, + 0.56 + ], + "angle": 0, + "content": "We verify our proposed method using both synthetic and real single-photon image sensor data. A snapshot of the results is shown in Fig. 1." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.574, + 0.219, + 0.589 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.473, + 0.869 + ], + "angle": 0, + "content": "Single-Photon Imaging. The context of this paper is single-photon image sensors, of which we are mostly interested in one-bit sensing because of its potential in high-speed and high dynamic range imaging [2]. Today, 1-bit sensing has been realized using quanta image sensors (QIS) [19-21, 50, 67] and single-photon avalanche diodes (SPAD) [1, 16, 17]. From an image reconstruction point of view, since 1-bit data contains so little information, all existing image reconstruction methods need to use multiple frames. For QIS, these methods range from classical optimization [3, 5], transform-denoise [4], neural network [12], color demosaicking [18], frequency demodulation [18], and student-teacher learning [11, 22]. For SPAD, quanta burst photography [51] and its variants [30, 32, 35, 57] first align frames and then apply image fusions [27-29, 59]. The question we ask in this paper is to push the input requirement to the limit: can we recover/generate a color image from a binary input?" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Image-to-Image Translation. This paper focuses on using generative models. Generative Adversarial Networks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.303 + ], + "angle": 0, + "content": "(GANs) [26] have been a significant breakthrough in the field of image translation. Pix2Pix [33] and CycleGAN [68] demonstrated the effectiveness of GANs in image-to-image translation tasks, with paired and unpaired datasets, respectively. Several subsequent works have focused on improving image fidelity, disentangling style and content, and generating images with varying styles, such as Pix2PixHD [61], StarGAN [13], StyleGAN [37], and StyleGANv2 [14]. Recent advancements in continuous cross-domain translation techniques have further improved the output quality by leveraging images from intermediate domains. These techniques include interpolation of networks [40, 41], interpolation of two latent vectors [14, 37], and exploiting the interpolation path and translation of manifold [9, 42, 55]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.304, + 0.893, + 0.591 + ], + "angle": 0, + "content": "Convolutional Filter Decomposition. In the proposed method, one of the key components is the convolutional filter decomposition. Several techniques have been proposed to decompose filters of Convolutional Neural Networks (CNNs) in order to reduce the computation complexity by exploiting the redundancy in convolutional filters [10, 39, 53, 54, 63, 69]. Filter decomposition methods achieve similar performance as compared to the original CNNs while offering speedups in computation. As it is expensive to model the high-dimension space of convolutional filters directly, Qiu et al. [56] observed that the convolutional filters can often be approximated using a linear combination small set of basis elements. By decomposing filters into a small set of basis, Wang et al. [64, 65] were able to sample the basis and demonstrated theoretical and empirical results on stochastic image generation tasks. Futher, Chen et al. [7] showed that filter decomposition can be used to calculate network similarity and large convolutional models can be fine tuned using filter decomposition [8]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.609, + 0.692, + 0.627 + ], + "angle": 0, + "content": "3. Proposed Approach" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.635, + 0.791, + 0.651 + ], + "angle": 0, + "content": "3.1. One-Bit Image Formation Model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.706 + ], + "angle": 0, + "content": "We consider a simplified image formation model for 1-bit image sensors. Letting \\(\\theta\\) be the underlying quanta exposure [20], the observed signal is defined by" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.718, + 0.844, + 0.745 + ], + "angle": 0, + "content": "\\[\nY = \\operatorname {A D C} \\left\\{\\text {P o i s s o n} (\\theta) + \\text {G a u s s i a n} \\left(0, \\sigma_ {r} ^ {2}\\right) \\right\\},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.757, + 0.892, + 0.862 + ], + "angle": 0, + "content": "where \\(\\sigma_r\\) denotes the standard deviation of read noise, and ADC is the analog-to-digital conversion. For 1-bit sensing, the ADC can be modeled as a threshold such that \\(\\mathrm{ADC}(x) = 1\\) if \\(x \\geq q\\) and \\(\\mathrm{ADC}(x) = 0\\) if \\(x < q\\) for some threshold \\(q\\). Since \\(Y\\) is a binary random variable, it is fully characterized by the probability mass function, which is [2]" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.865, + 0.808, + 0.904 + ], + "angle": 0, + "content": "\\[\np _ {Y} (1) = \\sum_ {k = 0} ^ {\\infty} \\frac {e ^ {- \\theta} \\theta^ {k}}{k !} \\Phi \\left(\\frac {k - q}{\\sigma_ {r}}\\right),\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25139" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.589, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.207, + 0.589, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.101, + 0.889, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.396, + 0.896, + 0.507 + ], + "angle": 0, + "content": "Figure 2. Illustration of exposure correction and colorization of binary images using neural networks. (a) depicts a range of images from overexposed to underexposed, illustrating the degradation of image details due to exposure variation. (b) contrasts the standard colorization workflow and our proposed approach. (i) In standard colorization approaches, a neural network learns to map a binary image \\(\\mathbf{Y}\\) to the corresponding color image \\(\\mathbf{X}_c\\) via a neural network, \\(\\mathcal{F}^{\\mathrm{aug}}\\), where superscript 'aug' indicates the colorizer is trained using dataset with augmented exposure images. (ii) In contrast, our approach does not require training colorizer with augmented exposure images. (c) compares the colorization results: the first row is the output of a colorizer trained without augmentation, the second row is the output of colorizer trained with augmented data, the third row corresponds to the results of our method and the last provides the ground truth images for reference." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.532, + 0.471, + 0.583 + ], + "angle": 0, + "content": "where \\(\\Phi(z) = \\int_{-\\infty}^{z} \\frac{1}{\\sqrt{2\\pi}} \\exp\\{-t^2/2\\} dt\\) is the cumulative distribution of a standard Gaussian. It then follows that the mean (or the bit density) is \\(\\mu = \\mathbb{E}[Y] = p_Y(1)\\)." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.595, + 0.472, + 0.66 + ], + "angle": 0, + "content": "
MethodGenerator SizeDataset SizeTraining TimeFID
Standard colorizationResNet-93x1x62.34
+ Increased training timeResNet-93x1.5x54.25
+ Increased model sizeResNet-103x1x74.58
Proposed methodResNet-91x1x49.88
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.669, + 0.471, + 0.711 + ], + "angle": 0, + "content": "Table 1. Performance comparison of the standard colorization approach under various training conditions versus the proposed method." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.74, + 0.47, + 0.772 + ], + "angle": 0, + "content": "3.2. Why Exposure Synthesis is Required for Colorization?" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Given a binary image \\(\\mathbf{Y} \\in \\mathbb{R}^N\\), our goal is to recover a color image \\(\\mathbf{X}_c \\in \\mathbb{R}^{N \\times 3}\\). However, since the measurement does not contain any color information, our problem fits the classical work of image colorization [38, 58, 60, 66]. However, it is non-trivial to deploy such an extension in real world scenarios, because very likely there is a discrepancy in the exposure levels of the input images compared to those used during training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.533, + 0.895, + 0.671 + ], + "angle": 0, + "content": "An example of binary images of different exposures are shown in Figure 2(a) illustrating the effect of exposure change on the details in the captured binary binary image. Consider a typical colorization network for example, an Image-to-Image translation model like Pix2Pix [34] trained on well exposed image but tested on overexposed images, the colorization results are notably poorer (shown in Figure 2(c) row 1). This is a realistic concern because the underlying exposure \\(\\theta\\) can fluctuate dramatically in real life." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Possible solutions include (i) Exposure Augmentation : training dataset is enriched with images of varying exposures, at a cost of additional training time. We find, however, this approach partially addresses the issue, as demonstrated in our results shown in Figure 2(b) row 2. In particular, this approach falls short at over- or under-exposure. One can alleviate this issue by applying multiple colorizers, each handling a narrow range of exposures, but this is not ideal as it demands substantially more computational resources. (ii) Increased Model size : Another solution is to increase the size of generator used in Pix2Pix [34] and train for longer period of time. In Table 1, we present the colorization performance represented by the FID metric under various training scenarios. Under increased generator size or increased training time the standard colorization ap" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25140" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "proach performs poorly compared to proposed method detailed in subsequent sections." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.122, + 0.473, + 0.365 + ], + "angle": 0, + "content": "Proposed Solution. We address the challenge by a dedicated exposure correction based on neural ordinary differential equations (neural ODEs) to standardize the exposure of binary images. This specialization allows the colorizer to focus on consistent-exposure images, resulting in notably enhanced colorization, as evidenced Figure 2(c) row 3. Additionally, our exposure synthesis can generate multiple exposures from a single binary image, enabling more effective multi-exposure colorization. Some of the added advantages of our approach include: (a) Modularity. The colorizer in our design is interchangeable, allowing for upgrades to more advanced architectures without requiring retraining on a large, augmented dataset. (b) Parameter Efficiency. Our method achieves its objectives without resorting to multiple networks with large parameter sets, ensuring a more streamlined and resource-efficient process." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.37, + 0.47, + 0.402 + ], + "angle": 0, + "content": "3.3. Continuous Exposure Synthesis using Filter Decomposed Neural ODEs" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.408, + 0.47, + 0.55 + ], + "angle": 0, + "content": "The image formation model defined in Section 3.1 indicates that we can formulate the continuum burst of binary images as \\(\\mathbf{Y}(\\widetilde{\\theta})\\), where the binary image \\(\\mathbf{Y}\\) is a function of a continuous exposure value \\(\\widetilde{\\theta} \\in (0,1]\\). A small \\(\\widetilde{\\theta}\\) value indicates a higher exposed image \\(\\mathbf{Y}\\). One of the core contributions of this work is a deep neural network architecture that synthesizes a binary image \\(\\mathbf{Y}_t\\) at a target exposure \\(\\widetilde{\\theta}_{\\mathrm{target}}\\) given an input binary image \\(\\mathbf{Y}_i\\) measured at the exposure value \\(\\widetilde{\\theta}_{\\mathrm{input}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.55, + 0.47, + 0.659 + ], + "angle": 0, + "content": "As the exposure values are continuous, we may need an exhaustive training set that densely samples every possible pair of input exposure value \\(\\widetilde{\\theta}_{\\mathrm{input}}\\) and target exposure value \\(\\widetilde{\\theta}_{\\mathrm{target}}\\) for every scene to train our network. To tackle this problem, we designed a convolutional architecture that enforces the parameter of the network to smoothly change according to the exposures." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.47, + 0.765 + ], + "angle": 0, + "content": "We use the convolution filter decomposition technique [56, 61, 64] to represent the convolutional filters \\(\\mathbf{F}\\) with a low number of parameters, and we leverage neural ODEs [6] to associate parameters with image exposures. With these techniques, we enable the adaptive control of filter parameters \\(\\mathbf{F}\\) and thereby the learning of the exposure synthesis operation \\(\\mathcal{G}\\) efficiently." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Efficient Modelling of Filter Subspace. Specifically, we formulate the filter parameters \\(\\mathbf{F}\\) using the atom-coefficient [56] decomposition, which permits the smallest computational bottleneck in modeling filter subspaces among the candidates. Given a convolutional filter \\(\\mathbf{F} \\in \\mathbb{R}^{c' \\times c \\times k \\times k}\\) where \\(c, c'\\) and \\(k\\) are the output, input channels, and spatial size, respectively. The atom-coefficient decomposition over a set of \\(m\\) filter atoms decomposes \\(\\mathbf{F}\\) into coefficients, denoted by \\(\\phi \\in \\mathbb{R}^{c' \\times c \\times m}\\), and atoms, denoted" + }, + { + "type": "equation", + "bbox": [ + 0.499, + 0.09, + 0.729, + 0.134 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\text {b y} \\Lambda \\in \\mathbb {R} ^ {m \\times k \\times k} \\\\ \\mathbf {F} = \\phi \\boldsymbol {\\Lambda}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.149, + 0.892, + 0.21 + ], + "angle": 0, + "content": "When \\(m \\ll c\\), the dimension of the parameters is significantly reduced. Extending the decomposition to a L-layer convolutional network gives a set of \\(L\\) filter atoms \\(\\Lambda = \\{\\Lambda^l\\}_{l=1}^L\\) and \\(L\\) coefficients \\(\\phi = \\{\\phi^l\\}_{l=1}^L\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.21, + 0.892, + 0.351 + ], + "angle": 0, + "content": "Enforcing Filter Atom Continuity. We enforce the filter atoms \\(\\Lambda\\) to smoothly change according to the input and target exposure levels \\(\\widetilde{\\theta}_{\\mathrm{input}}\\) and \\(\\widetilde{\\theta}_{\\mathrm{target}}\\). To do so, we incorporate the continuous-time networks like neural ODEs to describe the relationship of \\(\\Lambda (\\widetilde{\\theta}_{\\mathrm{input}},\\widetilde{\\theta}_{\\mathrm{target}})\\), as the neural ODEs provide advantages in the number of parameters required, memory consumption, and computational adaptivity [6]. The derivative of the filter atoms \\(d\\Lambda /d\\widetilde{\\theta}\\) w.r.t. the exposure value \\(\\widetilde{\\theta}\\) as a neural network \\(g(\\cdot)\\)" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.363, + 0.813, + 0.382 + ], + "angle": 0, + "content": "\\[\nd \\boldsymbol {\\Lambda} / d \\widetilde {\\theta} = g (\\widetilde {\\theta} _ {\\text {i n p u t}}, \\widetilde {\\theta} _ {\\text {t a r g e t}}, \\boldsymbol {\\Lambda} _ {\\text {i n i t}}; \\mathcal {W}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.393, + 0.892, + 0.471 + ], + "angle": 0, + "content": "where \\(\\Lambda_{\\mathrm{init}}\\) is the initial state of filter atoms, and \\(\\mathcal{W}\\) is the learnable parameters of \\(g(\\cdot)\\). The initial condition \\(\\Lambda_{\\mathrm{init}}\\) is a learnable parameter and is optimized jointly with \\(\\mathcal{W}\\). Given the input and target exposure values \\(\\widetilde{\\theta}_{\\mathrm{input}}\\) and \\(\\widetilde{\\theta}_{\\mathrm{target}}\\), the corresponding filter atoms can be calculated as" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.482, + 0.856, + 0.522 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Lambda} (\\widetilde {\\theta} _ {\\text {t a r g e t}}, \\widetilde {\\theta} _ {\\text {i n p u t}}) = \\boldsymbol {\\Lambda} _ {\\text {i n i t}} + \\int_ {\\widetilde {\\theta} _ {\\text {i n p u t}}} ^ {\\widetilde {\\theta} _ {\\text {t a r g e t}}} g (\\boldsymbol {\\Lambda} _ {\\text {i n i t}}, \\mathcal {W}) d \\widetilde {\\theta}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.532, + 0.892, + 0.593 + ], + "angle": 0, + "content": "According to Picard-Lindelöf existence theorem [15], for a given initial condition, the solution to the IVP is unique. Theoretically, we prove continuity in filter atoms induces controlled changes in image exposure." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.607, + 0.892, + 0.7 + ], + "angle": 0, + "content": "Theorem 3.1 If \\(\\Lambda_{1}\\) and \\(\\Lambda_{2}\\) are filter atoms generated with neural ODE using integration intervals \\((\\widetilde{\\theta}_{0},\\widetilde{\\theta}_{l})\\) and \\((\\widetilde{\\theta}_0,\\widetilde{\\theta}_2)\\), respectively such that \\(\\| \\Lambda_1 - \\Lambda_2\\| \\leq \\epsilon |\\widetilde{\\theta}_1 - \\widetilde{\\theta}_2|\\), for \\(\\epsilon >0\\). If \\(\\theta_{1}\\) and \\(\\theta_{2}\\) are the exposures at \\(\\widetilde{\\theta}_{1}\\) and \\(\\widetilde{\\theta}_{2}\\), and activation function \\(\\sigma\\) is non-expansive, then change in exposure \\(\\Delta \\theta\\) is continuous in filter atoms." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.713, + 0.892, + 0.759 + ], + "angle": 0, + "content": "Details of Theorem 3.1 and its proof can be found in the Appendix section A1. An illustration of our parameter efficient approach is shown in Figure 3." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.769, + 0.783, + 0.785 + ], + "angle": 0, + "content": "3.4. Exposure Adaptive Colorization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.792, + 0.892, + 0.837 + ], + "angle": 0, + "content": "Our proposed exposure adaptive colorization consists of an exposure correction realized by the stated synthesis and exposure level specific colorizer." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.838, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Exposure correction. Given a colorization method specific to exposure \\(\\widetilde{\\theta}_{\\mathrm{target}}\\) and a input sample \\((\\mathbf{Y}_i,\\widetilde{\\theta}_{\\mathrm{input}})\\), the proposed exposure correction generates image \\(\\hat{\\mathbf{Y}}_t\\) at the target exposure in two steps. With the operation \\(\\mathcal{G}\\) carried out" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25141" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.088, + 0.891, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.333, + 0.895, + 0.461 + ], + "angle": 0, + "content": "Figure 3. An illustration of our proposed method for exposure adaptive colorization: A binary image, \\(\\mathbf{Y}\\), which can be overexposed or underexposed, is input into the proposed exposure synthesis module. The colorization of binary image can be achieved using Single Image Colorization (SIC) or Burst Image Colorization (BIC). (i) SIC: Based on the input and target exposure levels, \\(\\widetilde{\\theta}_{\\mathrm{input}}\\) and \\(\\widetilde{\\theta}_{\\mathrm{target}}\\), this module adjusts the weights of a exposure synthesis network \\(\\mathcal{G}\\), which then generates an exposure-corrected image. Note that corrected image is not necessarily a binary image. Since the colorization module \\(\\mathcal{F}\\) is trained to colorize only image of specific exposure, the exposure synthesis module ensures corrected binary image has similar exposure to the one on which \\(\\mathcal{F}\\) is trained. (ii) BIC: For BIC we generate images with varying exposures as input to the burst image colorization network, \\(\\mathcal{F}_{\\mathrm{burst}}\\). The trained network is able to exploit the complementary information across multiple exposures with the help of Cross Non-Local Fusion blocks [44] to synthesize colors in regions of the image that is otherwise not possible by SIC approach." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.485, + 0.47, + 0.608 + ], + "angle": 0, + "content": "by some generative network, first, the filter atoms of the network's convolutional filters are computed using the neural ODEs based on the \\((\\widetilde{\\theta}_{\\mathrm{input}},\\widetilde{\\theta}_{\\mathrm{target}})\\) pair; next, the convolutional filter weights of the network are updated using the filter atoms and their corresponding coefficients, and the network generate the exposure corrected image \\(\\hat{\\mathbf{Y}}_t\\) based on \\(\\mathbf{Y}_i\\). In the case of burst colorization, multiple images are generated corresponding to a set of target exposures." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.608, + 0.471, + 0.793 + ], + "angle": 0, + "content": "Colorization. Both single-image-based colorization and burst-based colorization are considered in our work. The first approach involves a colorization network \\(\\mathcal{F}:\\hat{\\mathbf{Y}}_t\\to\\) \\(\\mathbf{X}_c\\) which takes a exposure corrected binary image as input and outputs corresponding RGB image \\(\\mathbf{X}_c\\). The network is trained exclusively at a fixed exposure level, and variations in the exposure of the input image is pre-adjusted by the exposure correction step. The second approach uses a colorization network \\(\\mathcal{F}_{\\mathrm{burst}}:\\{\\hat{\\mathbf{Y}}_t\\}_{t = 1}^K\\to \\mathbf{X}_c\\) that takes in a burst of \\(K\\) binary images \\(\\{\\hat{\\mathbf{Y}}_t\\}_{t = 1}^K\\) corresponding to predefined varying exposure values \\(\\{\\hat{\\theta}_t\\}_{t = 1}^K\\) and outputs one RGB image \\(\\mathbf{X}_c\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.807, + 0.21, + 0.824 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.267, + 0.849 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "We evaluate the performance of both the exposure synthesis module standalone and the entire proposed approach. We conduct exposure burst recovery experiment to evalu" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.485, + 0.892, + 0.53 + ], + "angle": 0, + "content": "ate the exposure synthesis and Single Image Colorization (SIC) and Burst Image Colorization (BIC) experiments to evaluate the final colorization results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.533, + 0.893, + 0.761 + ], + "angle": 0, + "content": "Implementation of Exposure Adaptive Colorization. We implement the neural ODEs using a 6-layer multilayer perceptron. The generative networks we use for exposure synthesis are variants of Pix2Pix [34] and CycleGAN [68], with which the models are named AtomODE-Pix2Pix and AtomODE-CycleGAN. For colorization, we train a Pix2Pix colorizer dedicated to SIC task and another to BIC. For BIC, the colorizer is additionally equipped with a Cross Non-Local Fusion (CNLF) [44] to constructively fuse the representations obtained from the burst. Note that the colorizers are trained directly on pairs of binary and color image data at fixed \\(\\widetilde{\\theta}\\), independent of exposure synthesis. In our colorization experiment, the same trained colorizer is applied on all exposure correction methods for comparison. Details can be found in the Appendix section A2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.763, + 0.892, + 0.868 + ], + "angle": 0, + "content": "Datasets and Synthetic Data Generation. We train and evaluate the proposed method as well as the baselines on synthetic data generated using images from two datasets: AFHQ [14] and CelebA-HQ [36]. AFHQ dataset contains images of three classes, cats, dogs, and wild animals, and CelebA-HQ contains images of male and female human faces." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "To synthesize binary images, we use the image formation model described in Section 3.1 and apply exposure bracket" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25142" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.09, + 0.894, + 0.189 + ], + "angle": 0, + "content": "
Exposure Synthesis MethodAFHQ (512 x 512)CelebA-HQ (256 x 256)
CatDogWildMaleFemale
MSE (↓)/RL (↑)/FID (↓)MSE (↓)/RL (↑)/FID (↓)MSE (↓)/RL (↑)/FID (↓)MSE (↓)/RL (↑)/FID (↓)MSE (↓)/RL (↑)/FID (↓)
Pix2Pix-DNI3.75 / 0.9816 / 61.895.91 / 0.9844 / 95.438.13 / 0.9780 / 39.293.74 / 0.9563 / 56.134.78 / 0.9681 / 39.81
CycleGAN-DNI9.61 / 0.9816 / 207.8216.79 / 0.9824 / 220.909.62 / 0.9820 / 336.389.31 / 0.9824 / 94.7814.26 / 0.9810 / 250.61
SAVI2I41.15 / 46.27 / 178.2946.27 / 0.9804 / 217.9838.87 / 0.9019 / 139.4767.26 / 0.9563 / 111.0569.21 / 0.9698 / 261.45
DLOW2.44 / 0.9654 / 92.302.32 / 0.9489 / 173.652.75 / 0.9632 / 179.231.87 / 0.9655 / 62.571.37 / 0.9798 / 63.31
AtomODE-Pix2Pix (ours)2.29 / 0.9997 / 57.883.17 / 0.9994 / 140.232.15 / 0.9998 / 79.762.15 / 0.9995 / 53.443.20 / 0.9995 / 40.08
AtomODE-CycleGAN (ours)1.15 / 0.9948 / 60.541.18 / 0.9971 / 87.431.23 / 0.9979 / 35.751.58 / 0.9986 / 77.821.87 / 0.9994 / 62.59
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.195, + 0.895, + 0.251 + ], + "angle": 0, + "content": "Table 2. Colorization results on AFHQ and CelebA-HQ datasets. The input image for all these experiments is an overexposed image at index \\(\\tau = 0\\) of exposure burst. All the colorizers are trained using an image corresponding to \\(\\tau = 10\\) in the exposure burst. At test time we correct the overexposed input to correspond to the one used during training. In the first column of the table, we note the combination of the exposure correction method used and colorizer is same across all the methods." + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.261, + 0.47, + 0.662 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.675, + 0.47, + 0.729 + ], + "angle": 0, + "content": "Figure 4. Qualitative results for exposure burst recovery experiments for AFHQ dataset. We compare our methods, AtomODE-Pix2Pix and AtomODE-CycleGAN, with DLOW, SAVI2I and DNI." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.732, + 0.47, + 0.809 + ], + "angle": 0, + "content": "ing to generate various exposure. Details of data synthesis can be found in the Appendix section A2. Our compiled test datasets consist of 15-frame bursts of exposure varying from overexposed to underexposed. We index the exposure of these images in a burst by \\(\\tau\\), a discretized label of \\(\\widetilde{\\theta}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Baselines. In the exposure burst recovery experiment, we compare our approach against DLOW [25] and SAVI2I [52]. We further consider DNI [62] for its capability of interpolating network parameters to generate continuous variations of outputs. We prepare baselines Pix2Pix-DNI and CycleGAN-DNI by coupling DNI with Pix2Pix and Cycle" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.264, + 0.892, + 0.339 + ], + "angle": 0, + "content": "GAN exposure generators. In colorization experiments, we apply the independently trained Pix2Pix colorizers to the exposure correction results generated by all methods. For more details on baseline training and hyperparameters, we refer readers to the Appendix section A2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.34, + 0.893, + 0.46 + ], + "angle": 0, + "content": "Metrics. In the burst recovery experiment, we choose to use Mean Squared Error (MSE) and Relative Linearity (RL) [43] to distinguish a smoothly varying exposure burst with consistent appearance changes. A high RL score indicates a better transition in exposure among consecutive samples. In the colorization experiments, we assess the quality of exposure correction by computing the Fréchet Inception Distance (FID) between the colorized image and ground truth." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.484, + 0.794, + 0.5 + ], + "angle": 0, + "content": "4.2. Exposure Burst Recovery Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.507, + 0.892, + 0.703 + ], + "angle": 0, + "content": "Figure 4 presents the qualitative comparison of all methods on AFHQ dataset. We find our methods, AtomODE Pix2Pix and AtomODE CycleGAN are able to generate consistent set of smoothly varying exposure images than baseline methods and match ground truth images generated from exposure bracketing. We also observe some baselines struggle with generation of overexposed frames. All methods for comparison are further assessed on both datasets using MSE and RL, with the quantitative results shown in Table 2. Results suggest that the proposed AtomODE provides consistently better performance in all testing conditions across multiple datasets. Additional qualitative results can be found in Appendix section A4." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.712, + 0.878, + 0.728 + ], + "angle": 0, + "content": "4.3. Single and Burst Image Colorization Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.9 + ], + "angle": 0, + "content": "The SIC colorization results of all methods on AFHQ dataset are shown in Figure 5, where the exposure of the test samples is different from the that of the colorizer's training samples. Corrupted colored images are generated by most of the baselines, while the proposed method produces visually appealing outputs. Figure 7 shows the SIC colorization results on CelebA-HQ dataset. Additionally, we also present the qualitative results for BIC in Figure 6 where \\( K \\) denotes the number of binary frames input to the colorizer. The quantitative results are shown in Table 2, assessed with the FID metric." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25143" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.088, + 0.891, + 0.665 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.675, + 0.893, + 0.719 + ], + "angle": 0, + "content": "Figure 5. Colorized results for AFHQ dataset using various combinations of exposure correction methods and colorizer. The first column shows the overexposed binary input image for exposure correction. The results of our methods AtomODE Pix2Pix and AtomODE CycleGAN are shown in columns six and seven." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.726, + 0.292, + 0.743 + ], + "angle": 0, + "content": "4.4. Real Data Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.903 + ], + "angle": 0, + "content": "As our approach is trained on synthetic data, it is essential to evaluate its performance on real-world data. To this end, we conduct experiments using both a CMOS camera and a prototype QIS camera. We refer the readers to Appendix for further details about the experiment setup. We print 100 images from the training set and 25 from the test set, and we capture pictures of them using both cameras, collecting 125 pairs of grayscale and RGB images. We fine-tune the colorization model with the 100 real data samples, and we evaluate our proposed approach on the set of 25 test data, in" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.726, + 0.895, + 0.895 + ], + "angle": 0, + "content": "cluding both binarized CMOS data and QIS data. In Figure 8, we present qualitative results from the evaluation on the real data obtained by the CMOS camera. Despite that the proposed method was not trained with vast real sensor data, it is able to produce visually pleasing colorized images. We observe that fine-tuning the model using a small training dataset significantly improves the colorization quality. We further present results from the evaluation on images captured by the QIS camera in Figure 9. These results demonstrate the potential of the proposed method for real-world single-photon imaging applications." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25144" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.086, + 0.111, + 0.1, + 0.258 + ], + "angle": 270, + "content": "Groundtruth \\(K = 5\\) \\(K = 1\\)" + }, + { + "type": "image", + "bbox": [ + 0.105, + 0.095, + 0.468, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.279, + 0.47, + 0.361 + ], + "angle": 0, + "content": "Figure 6. An illustration of burst image colorization for AFHQ dataset using multiple binary images of varying exposures as input. The colorizer is trained using images of exposure \\(\\tau = \\{0,4,8,10,14\\}\\). An overexposed input image is used to synthesize multiple images of varying exposures and passed to the colorizer." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.381, + 0.464, + 0.64 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.652, + 0.47, + 0.694 + ], + "angle": 0, + "content": "Figure 7. Colorized results for CelebA-HQ dataset using various combinations of exposure correction methods and colorizers. The results of our method are shown in columns four and five." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.71, + 0.377, + 0.727 + ], + "angle": 0, + "content": "4.5. Benefits for Practical Applications" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.9 + ], + "angle": 0, + "content": "Our methodology is specifically designed with modularity and parameter efficiency in mind, features crucial for real-world industry applications. The Pix2PixDNI method, the second-best performing approach, necessitates training multiple networks and either interpolating model weights on-the-fly during inference or precomputing and storing them to generate images with varied exposures. Conversely, our method circumvents the additional storage requirements and the need to train multiple model parameters by incorporating the exposure correction into the neural ODE framework." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.091, + 0.885, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.334, + 0.893, + 0.391 + ], + "angle": 0, + "content": "Figure 8. Results on real-world images captured from a CMOS camera (images shown along the row with a red line are colorized images, and along the green line are the ground-truth images captured using an RGB CMOS camera)." + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.409, + 0.888, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.564, + 0.892, + 0.592 + ], + "angle": 0, + "content": "Figure 9. Colorized output using input from prototype QIS camera." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.607, + 0.619, + 0.623 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.633, + 0.893, + 0.83 + ], + "angle": 0, + "content": "In this work, we proposed a generative exposure correction approach that can synthesize a spectrum of exposure given a single 1-bit image. We modeled the binary signal variation with respect to exposure change as a continuous function by adapting a small set of parameters of a network, known as filter atoms, to desired exposure using neural ODE. Theoretically and empirically, we demonstrated successful recovery of bursts of 1-bit images from a single input image, and we verified its application on single-image-based and burst-image-based 1-bit image colorization. Experiments on real CMOS and QIS data further indicated the potentials of the proposed approach on real-world single-photon imaging applications." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "25145" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Samuel Burri, Yuki Maruyama, Xavier Michalet, Francesco Regazzoni, Claudio Bruschini, and Edoardo Charbon. Architecture and applications of a high resolution gated SPAD image sensor. Optics Express, 22(14):17573-17589, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.472, + 0.227 + ], + "angle": 0, + "content": "[2] Stanley H. Chan. What does a one-bit Quanta Image Sensor offer? IEEE Transactions on Computational Imaging, 8: 770-783, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.229, + 0.471, + 0.285 + ], + "angle": 0, + "content": "[3] Stanley H. Chan and Yue M. Lu. Efficient image reconstruction for gigapixel quantum image sensors. In IEEE Global Conference on Signal and Information Processing, pages 312-316, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.287, + 0.471, + 0.328 + ], + "angle": 0, + "content": "[4] Stanley H. Chan, Omar A. Elgendy, and Xiran Wang. Images from bits: Non-iterative image reconstruction for Quanta Image Sensors. MDPI Sensors, 16(11), 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.329, + 0.471, + 0.385 + ], + "angle": 0, + "content": "[5] Stanley H. Chan, Xiran Wang, and Omar A. Elgendy. Plug-and-Play ADMM for image restoration: Fixed-point convergence and applications. IEEE Transactions on Computational Imaging, 3:84–98, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.387, + 0.471, + 0.428 + ], + "angle": 0, + "content": "[6] Ricky T. Q. Chen, Yulia Rubanova, Jesse Bettencourt, and David K. Duvenaud. Neural Ordinary Differential Equations. In NeurIPS, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.43, + 0.471, + 0.457 + ], + "angle": 0, + "content": "[7] Wei Chen, Zichen Miao, and Qiang Qiu. Inner product-based neural network similarity. In NeurIPS, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.459, + 0.471, + 0.514 + ], + "angle": 0, + "content": "[8] Wei Chen, Zichen Miao, and Qiang Qiu. Large convolutional model tuning via filter subspace. Available online: https://arxiv.org/abs/2403.00269, 2024. Accessed 03/06/2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.516, + 0.471, + 0.558 + ], + "angle": 0, + "content": "[9] Ying-Cong Chen, Xiaogang Xu, Zhuotao Tian, and Jiaya Jia. Homomorphic latent space interpolation for unpaired Image-To-Image translation. In CVPR, pages 2403-2411, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.559, + 0.47, + 0.613 + ], + "angle": 0, + "content": "[10] Xiuyuan Cheng, Qiang Qiu, Robert Calderbank, and Guillermo Sapiro. RotDCF: Decomposition of convolutional filters for rotation-equivariant deep networks. In ICLR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.616, + 0.47, + 0.67 + ], + "angle": 0, + "content": "[11] Yiheng Chi, Abhiram Gnanasambandam, Vladlen Koltun, and Stanley H. Chan. Dynamic low-light imaging with Quanta Image Sensors. In ECCV, pages 122-138, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.673, + 0.47, + 0.715 + ], + "angle": 0, + "content": "[12] Joon Hee Choi, Omar A. Elgendy, and Stanley H. Chan. Image reconstruction for Quanta Image Sensors using deep neural networks. In IEEE ICASSP, pages 6543-6547, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.716, + 0.47, + 0.771 + ], + "angle": 0, + "content": "[13] Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. StarGAN: Unified Generative Adversarial Networks for multi-domain Image-to-Image translation. In CVPR, pages 8789–8797, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.773, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[14] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. StarGAN v2: Diverse image synthesis for multiple domains. In CVPR, pages 8185-8194, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.47, + 0.844 + ], + "angle": 0, + "content": "[15] Earl A. Coddington and Norman Levinson. Theory of Ordinary Differential Equations. McGraw-Hill, 1955. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[16] Neale A. W. Dutton, Istvan Gyongy, Luca Parmesan, Salvatore Gnecchi, Neil Calder, Bruce R. Rae, Sara Pellegrini, Lindsay A. Grant, and Robert K. Henderson. A SPAD-based QVGA image sensor for single-photon counting and quanta" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "imaging. IEEE Transactions on Electron Devices, 63(1): 189-196, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[17] Neale A. W. Dutton, Istvan Gyongy, Luca Parmesan, and Robert K. Henderson. Single photon counting performance and noise analysis of CMOS SPAD-based image sensors. MDPI Sensors, 16(7), 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.246 + ], + "angle": 0, + "content": "[18] Omar A. Elgendy, Abhiram Gnanasambandam, Stanley H. Chan, and Jiaju Ma. Low-light demosaicking and denoising for small pixels using learned frequency selection. IEEE Transactions on Computational Imaging, 7:137-150, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.249, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[19] Eric R. Fossum. Some thoughts on future digital still cameras. In Image Sensors and Signal Processing for Digital Still Cameras, pages 305-314. Taylor and Francis Group, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.305, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[20] Eric R. Fossum. Modeling the performance of single-bit and multi-bit Quanta Image Sensors. IEEE Journal of the Electron Devices Society, 1(9):166-174, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.892, + 0.401 + ], + "angle": 0, + "content": "[21] Eric R. Fossum. Analog read noise and quantizer threshold estimation from Quanta Image Sensor bit density. IEEE Journal of the Electron Devices Society, 10:269-274, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.404, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[22] Abhiram Gnanasambandam and Stanley H. Chan. Image classification in the dark using Quanta Image Sensors. In ECCV, pages 484-501, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.448, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[23] Abhiram Gnanasambandam and Stanley H. Chan. HDR imaging with Quanta Image Sensors: Theoretical limits and optimal reconstruction. IEEE Transactions on Computational Imaging, 6:1571-1585, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.504, + 0.892, + 0.558 + ], + "angle": 0, + "content": "[24] Abhiram Gnanasambandam and Stanley H. Chan. Exposure-referred signal-to-noise ratio for digital image sensors. IEEE Transactions on Computational Imaging, 8:561-575, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.56, + 0.892, + 0.602 + ], + "angle": 0, + "content": "[25] Rui Gong, Wen Li, Yuhua Chen, and Luc Van Gool. DLOW: Domain flow for adaptation and generalization. In CVPR, pages 2472-2481, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.604, + 0.892, + 0.657 + ], + "angle": 0, + "content": "[26] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative Adversarial Nets. In NeurIPS, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.66, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[27] Anant Gupta, Atul Ingle, and Mohit Gupta. Asynchronous single-photon 3D imaging. In ICCV, pages 7908-7917, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.703, + 0.892, + 0.743 + ], + "angle": 0, + "content": "[28] Anant Gupta, Atul Ingle, Andreas Velten, and Mohit Gupta. Photon-flooded single-photon 3D cameras. In CVPR, pages 6763-6772, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.746, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[29] Felipe Gutierrez-Barragan, Atul Ingle, Trevor Seets, Mohit Gupta, and Andreas Velten. Compressive single-photon 3D cameras. In CVPR, pages 17833-17843, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[30] Istvan Gyongy, Neale AW Dutton, and Robert K Henderson. Single-photon tracking for high-speed vision. MDPI Sensors, 18:323, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.831, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[31] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[32] Atul Ingle, Andreas Velten, and Mohit Gupta. High flux passive imaging with single-photon sensors. In CVPR, pages 6753-6762, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "25146" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[33] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-Image translation with conditional adversarial networks. In CVPR, pages 5967–5976, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.47, + 0.177 + ], + "angle": 0, + "content": "[34] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A. Efros. Image-to-Image translation with conditional adversarial networks. In CVPR, pages 5967–5976, 2017. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.179, + 0.469, + 0.233 + ], + "angle": 0, + "content": "[35] Kiyotaka Iwabuchi, Yusuke Kameda, and Takayuki Hamamoto. Image quality improvements based on motion-based deblurring for single-photon imaging. IEEE Access, 9:30080-30094, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.235, + 0.469, + 0.275 + ], + "angle": 0, + "content": "[36] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In ICLR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.278, + 0.469, + 0.319 + ], + "angle": 0, + "content": "[37] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for Generative Adversarial Networks. IEEE TPAMI, 43(12):4217-4228, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.321, + 0.468, + 0.347 + ], + "angle": 0, + "content": "[38] Manoj Kumar, Dirk Weissenborn, and Nal Kalchbrenner. Colorization transformer. In ICLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.35, + 0.468, + 0.404 + ], + "angle": 0, + "content": "[39] Vadim Lebedev, Yaroslav Ganin, Maksim Rakhuba, Ivan Oseledets, and Victor Lempitsky. Speeding-up convolutional neural networks using fine-tuned CP-decomposition. In ICLR, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.407, + 0.468, + 0.46 + ], + "angle": 0, + "content": "[40] Yu-Jing Lin, Po-Wei Wu, Che-Han Chang, Edward Chang, and Shih-Wei Liao. RelGAN: Multi-domain Image-to-Image translation via relative attributes. In ICCV, pages 5913–5921, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.463, + 0.469, + 0.517 + ], + "angle": 0, + "content": "[41] Wallace Lira, Johannes Merz, Daniel Ritchie, Daniel Cohen-Or, and Hao Zhang. GANHopper: Multi-hop GAN for unsupervised Image-to-Image translation. In ECCV, pages 363–379, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.52, + 0.469, + 0.588 + ], + "angle": 0, + "content": "[42] Yahui Liu, Enver Sangineto, Yajing Chen, Linchao Bao, Haoxian Zhang, Nicu Sebe, Bruno Lepri, Wei Wang, and Marco De Nadai. Smoothing the disentangled latent style space for unsupervised Image-to-Image translation. In CVPR, pages 10780–10789, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.591, + 0.469, + 0.63 + ], + "angle": 0, + "content": "[43] Kun Lu, Rongpeng Li, and Honggang Zhang. Contrastive monotonic pixel-level modulation. In ECCV, pages 493-510, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.633, + 0.469, + 0.688 + ], + "angle": 0, + "content": "[44] Ziwei Luo, Lei Yu, Xuan Mo, Youwei Li, Lanpeng Jia, Haoqiang Fan, Jian Sun, and Shuaicheng Liu. EBSR: Feature Enhanced Burst Super-Resolution with deformable alignment. In CVPR, pages 471–478, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.69, + 0.469, + 0.731 + ], + "angle": 0, + "content": "[45] Jiaju Ma and Eric Fossum. Quanta Image Sensor jot with sub 0.3e- r.m.s. read noise and photon counting capability. IEEE Electron Device Letters, 36(9):926-928, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.733, + 0.469, + 0.773 + ], + "angle": 0, + "content": "[46] Jiaju Ma and Eric R. Fossum. A Pump-Gate Jot device with high conversion gain for a Quanta Image Sensor. IEEE Journal of the Electron Devices Society, 3:73-77, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.776, + 0.469, + 0.842 + ], + "angle": 0, + "content": "[47] Jiaju Ma, Dakota Starkey, Arun Rao, Kofi Odame, and Eric R. Fossum. Characterization of Quanta Image Sensor pump-gate jots with deep sub-electron read noise. IEEE Journal of the Electron Devices Society, 3(6):472-480, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[48] Jiaju Ma, Saleh Masoodian, Dakota A Starkey, and Eric R. Fossum. Photon-number-resolving megapixel image sensor at room temperature without avalanche gain. Optica, 4(12): 1474-1481, 2017. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[49] Jiaju Ma, Dexue Zhang, Omar Elgendy, and Saleh Masoodian. A 0.19e-rms read noise 16.7mpixel stacked Quanta Image Sensor with \\(1.1\\mu \\mathrm{m}\\)-pitch backside illuminated pixels. IEEE Electron Device Letters, 42(6):891-894, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.151, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[50] Jiaju Ma, Stanley H. Chan, and Eric R. Fossum. Review of Quanta Image Sensors for ultralow-light imaging. IEEE Transactions on Electron Devices, 69(6):2824-2839, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.208, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[51] Sizhuo Ma, Paul Mos, Edoardo Charbon, and Mohit Gupta. Burst vision using single-photon cameras. In WACV, pages 5364-5374, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.251, + 0.892, + 0.306 + ], + "angle": 0, + "content": "[52] Qi Mao, Hsin-Ying Lee, Hung-Yu Tseng, Jia-Bin Huang, Siwei Ma, and Ming-Hsuan Yang. Continuous and diverse Image-to-Image translation via signed attribute vectors. IJCV, 130(2):517-549, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.309, + 0.892, + 0.349 + ], + "angle": 0, + "content": "[53] Zichen Miao, Ze Wang, Xiuyuan Cheng, and Qiang Qiu. Spatiotemporal joint filter decomposition in 3D convolutional neural networks. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.352, + 0.892, + 0.38 + ], + "angle": 0, + "content": "[54] Zichen Miao, Ze Wang, Wei Chen, and Qiang Qiu. Continual learning with filter atom swapping. In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.383, + 0.892, + 0.422 + ], + "angle": 0, + "content": "[55] Fabio Pizzati, Pietro Cerri, and Raoul de Charette. CoMoGAN: Continuous model-guided Image-to-Image translation. In CVPR, pages 14283–14293, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.426, + 0.892, + 0.466 + ], + "angle": 0, + "content": "[56] Qiang Qiu, Xiuyuan Cheng, Robert Calderbank, and Guillermo Sapiro. DCFNet: Deep neural network with decomposed convolutional filters. In ICML, 2018. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.47, + 0.892, + 0.51 + ], + "angle": 0, + "content": "[57] Trevor Seets, Atul Ingle, Martin Laurenzis, and Andreas Velten. Motion adaptive deblurring with single-photon cameras. In WACV, pages 1944–1953, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.513, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[58] Jheng-Wei Su, Hung-Kuo Chu, and Jia-Bin Huang. Instance-aware image colorization. In CVPR, pages 7965–7974, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.556, + 0.892, + 0.597 + ], + "angle": 0, + "content": "[59] Varun Sundar, Sizhuo Ma, Aswin Sankaranarayanan, and Mohit Gupta. Single-photon structured light. In CVPR, pages 17844-17854, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.601, + 0.892, + 0.641 + ], + "angle": 0, + "content": "[60] Patricia Vitoria, Lara Raad, and Coloma Ballester. ChromaGAN: Adversarial picture colorization with semantic class distribution. In WACV, pages 2434-2443, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.644, + 0.892, + 0.699 + ], + "angle": 0, + "content": "[61] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional GANs. In CVPR, pages 8798-8807, 2018. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.702, + 0.892, + 0.742 + ], + "angle": 0, + "content": "[62] X. Wang, K. Yu, C. Dong, X. Tang, and C. Loy. Deep network interpolation for continuous imagery effect transition. In CVPR, pages 1692-1701, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.745, + 0.892, + 0.813 + ], + "angle": 0, + "content": "[63] Ze Wang. Efficient Adaptation of Deep Vision Models. PhD thesis, Purdue University, 2023. Available online: https://hammer.purdue.edu/articles/thesis/Efficient_Adaptation_of_Deep_Vision_CModels/22696129.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.816, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[64] Ze Wang, Xiuyuan Cheng, Guillermo Sapiro, and Qiang Qiu. Stochastic conditional generative networks with basis decomposition. In ICLR, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[65] Ze Wang, Seunghyun Hwang, Zichen Miao, and Qiang Qiu. Image generation using continuous filter atoms. In NeurIPS, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25147" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.146 + ], + "angle": 0, + "content": "[66] Yanze Wu, Xintao Wang, Yu Li, Honglun Zhang, Xun Zhao, and Ying Shan. Towards vivid and diverse image colorization with generative color prior. In ICCV, pages 14357-14366, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[67] Zhaoyang Yin, Jiaju Ma, Saleh Masoodian, and Eric R Fossum. Threshold uniformity improvement in 1b Quanta Image Sensor readout circuit. MDPI Sensors, 22(7):2578, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.246 + ], + "angle": 0, + "content": "[68] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In ICCV, pages 2242-2251, 2017. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.469, + 0.303 + ], + "angle": 0, + "content": "[69] Wei Zhu, Qiang Qiu, Robert Calderbank, Guillermo Sapiro, and Xiuyuan Cheng. Scaling-translation-equivariant networks with decomposed convolutional filters. Journal of Machine Learning Research, 23(68):1-45, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "25148" + } + ] +] \ No newline at end of file diff --git a/2024/Generative Quanta Color Imaging/9125a07a-d19f-455b-bcb9-5f2fd8cc25e9_origin.pdf b/2024/Generative Quanta Color Imaging/9125a07a-d19f-455b-bcb9-5f2fd8cc25e9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b0b4e3ddc82cf5c47edfe16d614e623bcf9e77f0 --- /dev/null +++ b/2024/Generative Quanta Color Imaging/9125a07a-d19f-455b-bcb9-5f2fd8cc25e9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ae4e4c20cc2624db28bab717300a430027e82b4a69b64037963faaa263c0ddd3 +size 4240548 diff --git a/2024/Generative Quanta Color Imaging/full.md b/2024/Generative Quanta Color Imaging/full.md new file mode 100644 index 0000000000000000000000000000000000000000..365aef3e69c4f38cd38f83c341b8b62bbb8182ea --- /dev/null +++ b/2024/Generative Quanta Color Imaging/full.md @@ -0,0 +1,279 @@ +# Generative Quanta Color Imaging + +Vishal Purohit, Junjie Luo, Yiheng Chi, Qi Guo, Stanley H. Chan, Qiang Qiu Elmore Family School of Electrical and Computer Engineering Purdue University, West Lafayette, USA. + +{purohitv,luo330,chi14,guo675,stanchan,qqiu}@purdue.edu + +# Abstract + +The astonishing development of single-photon cameras has created an unprecedented opportunity for scientific and industrial imaging. However, the high data throughput generated by these 1-bit sensors creates a significant bottleneck for low-power applications. In this paper, we explore the possibility of generating a color image from a single binary frame of a single-photon camera. We evidently find this problem being particularly difficult to standard colorization approaches due to the substantial degree of exposure variation. The core innovation of our paper is an exposure synthesis model framed under a neural ordinary differential equation (Neural ODE) that allows us to generate a continuum of exposures from a single observation. This innovation ensures consistent exposure in binary images that colorizers take on, resulting in notably enhanced colorization. We demonstrate applications of the method in single-image and burst colorization and show superior generative performance over baselines. Project website can be found at https://vishal-s-p.github.io/projects/2023/generative_quanta_color.html + +# 1. Introduction + +Single-photon image sensors are gaining a strong momentum over the past decade due to their impeccable photon-counting capability. This not only opens the door for advanced scientific imaging in low-light environment, but also creates the opportunity for new imaging in high-speed, high dynamic range, and low-bit conditions. Single-photon imaging today are realized through single-photon avalanche diodes (SPAD) and quanta image sensors (QIS). The latest development of QIS has demonstrated a mega-pixel resolution [48] which is compatible with mainstream cell phone cameras, whereas SPAD has achieved hundreds of thousands of (binary) frames per second which has enabled a plethora of industrial and 3D imaging applications [16]. + +Problem Statement. In the context of 1-bit single-photon imaging, the data volume generated is huge. Taking a $1000 \times 1000$ SPAD array as an example, a video of + +Binary Input Pix2PixDNI Our Method Groundtruth + +![](images/900a82ebe2a17662b8c37752d6039080a656a4dc93c8d4757103bd60fa57fafd.jpg) +Figure 1. We introduce generative quanta color imaging. Given a binary frame captured by a single-photon camera (quanta image sensor (QIS) in this example), the proposed method generates a continuum of exposures using a neural ordinary differential equation framework. Color are then generated based on these exposures. Left: A qualitative comparison between our approach and existing methods is shown. Right: Comparison between colorization results and image captured using a RGB CMOS camera under similar conditions. + +![](images/e5467101037743f5643ff514b2579ea22fa07905613a732a24a27f0aa1ec7d75.jpg) + +500 binary frames requires a $500\mathrm{Mb / s}$ data transfer rate; if the frame rate becomes $150\mathrm{k}$ fps, the data transfer rate increases to $150\mathrm{Gb / s}$ . Since data throughput scales proportionally with the power consumption, it quickly becomes infeasible to use these sensors continuously without considering power supply, memory, and transmission issues. While solid progress are being made across the sensor and signal processing community [1, 11, 17, 23, 24, 45-47, 49, 50], the data throughput remains an unsolved problem for single-photon sensors to be considered applicable in many applications, including low-power devices, wireless devices, underwater vehicles, spacecrafts, etc., where bandwidth is the bottleneck. + +In this paper, instead of compressing data or redesigning the sensor, we ask the question + +Is it possible to recover a color image from a single 1-bit image? + +Significance. There are several reasons why this problem is significant. Firstly, while there is an ocean of literature on image generation (through methods such as Generative Adversarial Networks [26] and diffusion [31]), our problem is unique in the sense none of the existing work addresses it. Besides being binary (and hence colorless), in practice the input data usually also suffers from under- or over-exposures. Therefore, before generating color, it is even more critical to synthesize the unseen exposure. Secondly, any solution offered to our problem will provide a path to generative imaging in the future. From the point of view of compressed acquisition, our method demonstrates the extreme case of compression where the incoming data is a single bit. For future AR/VR applications where power is severely limited, our approach shows the possibility of extreme data compression. + +Contributions. The core algorithmic contribution of this paper is two-fold + +- We propose a neural ordinary differential equation (neural ODE) based framework to synthesize a continuum of exposures that are not available in the measurement. By controlling the integration interval of neural ODE, we obtain convolutional filters that are responsible for converting input representation to image representation of desired exposure. +- We theoretically and empirically show that a controlled variation of filter atoms leads to controlled exposure changes in the output image. + +We verify our proposed method using both synthetic and real single-photon image sensor data. A snapshot of the results is shown in Fig. 1. + +# 2. Related Work + +Single-Photon Imaging. The context of this paper is single-photon image sensors, of which we are mostly interested in one-bit sensing because of its potential in high-speed and high dynamic range imaging [2]. Today, 1-bit sensing has been realized using quanta image sensors (QIS) [19-21, 50, 67] and single-photon avalanche diodes (SPAD) [1, 16, 17]. From an image reconstruction point of view, since 1-bit data contains so little information, all existing image reconstruction methods need to use multiple frames. For QIS, these methods range from classical optimization [3, 5], transform-denoise [4], neural network [12], color demosaicking [18], frequency demodulation [18], and student-teacher learning [11, 22]. For SPAD, quanta burst photography [51] and its variants [30, 32, 35, 57] first align frames and then apply image fusions [27-29, 59]. The question we ask in this paper is to push the input requirement to the limit: can we recover/generate a color image from a binary input? + +Image-to-Image Translation. This paper focuses on using generative models. Generative Adversarial Networks + +(GANs) [26] have been a significant breakthrough in the field of image translation. Pix2Pix [33] and CycleGAN [68] demonstrated the effectiveness of GANs in image-to-image translation tasks, with paired and unpaired datasets, respectively. Several subsequent works have focused on improving image fidelity, disentangling style and content, and generating images with varying styles, such as Pix2PixHD [61], StarGAN [13], StyleGAN [37], and StyleGANv2 [14]. Recent advancements in continuous cross-domain translation techniques have further improved the output quality by leveraging images from intermediate domains. These techniques include interpolation of networks [40, 41], interpolation of two latent vectors [14, 37], and exploiting the interpolation path and translation of manifold [9, 42, 55]. + +Convolutional Filter Decomposition. In the proposed method, one of the key components is the convolutional filter decomposition. Several techniques have been proposed to decompose filters of Convolutional Neural Networks (CNNs) in order to reduce the computation complexity by exploiting the redundancy in convolutional filters [10, 39, 53, 54, 63, 69]. Filter decomposition methods achieve similar performance as compared to the original CNNs while offering speedups in computation. As it is expensive to model the high-dimension space of convolutional filters directly, Qiu et al. [56] observed that the convolutional filters can often be approximated using a linear combination small set of basis elements. By decomposing filters into a small set of basis, Wang et al. [64, 65] were able to sample the basis and demonstrated theoretical and empirical results on stochastic image generation tasks. Futher, Chen et al. [7] showed that filter decomposition can be used to calculate network similarity and large convolutional models can be fine tuned using filter decomposition [8]. + +# 3. Proposed Approach + +# 3.1. One-Bit Image Formation Model + +We consider a simplified image formation model for 1-bit image sensors. Letting $\theta$ be the underlying quanta exposure [20], the observed signal is defined by + +$$ +Y = \operatorname {A D C} \left\{\text {P o i s s o n} (\theta) + \text {G a u s s i a n} \left(0, \sigma_ {r} ^ {2}\right) \right\}, +$$ + +where $\sigma_r$ denotes the standard deviation of read noise, and ADC is the analog-to-digital conversion. For 1-bit sensing, the ADC can be modeled as a threshold such that $\mathrm{ADC}(x) = 1$ if $x \geq q$ and $\mathrm{ADC}(x) = 0$ if $x < q$ for some threshold $q$ . Since $Y$ is a binary random variable, it is fully characterized by the probability mass function, which is [2] + +$$ +p _ {Y} (1) = \sum_ {k = 0} ^ {\infty} \frac {e ^ {- \theta} \theta^ {k}}{k !} \Phi \left(\frac {k - q}{\sigma_ {r}}\right), +$$ + +![](images/0a412ef20fc6665ff386e6b92ac579b7530768973b6fe897ccb70c41383edb7d.jpg) + +![](images/d19eea29fd9623bbe343a5cfd25e45c74b76b4ed0d542eaf8effb270fc75c590.jpg) +Figure 2. Illustration of exposure correction and colorization of binary images using neural networks. (a) depicts a range of images from overexposed to underexposed, illustrating the degradation of image details due to exposure variation. (b) contrasts the standard colorization workflow and our proposed approach. (i) In standard colorization approaches, a neural network learns to map a binary image $\mathbf{Y}$ to the corresponding color image $\mathbf{X}_c$ via a neural network, $\mathcal{F}^{\mathrm{aug}}$ , where superscript 'aug' indicates the colorizer is trained using dataset with augmented exposure images. (ii) In contrast, our approach does not require training colorizer with augmented exposure images. (c) compares the colorization results: the first row is the output of a colorizer trained without augmentation, the second row is the output of colorizer trained with augmented data, the third row corresponds to the results of our method and the last provides the ground truth images for reference. + +![](images/49d6687fb09d1b4eec2811a3145e43b77894b9acfbf2b080262ed524ede45462.jpg) + +where $\Phi(z) = \int_{-\infty}^{z} \frac{1}{\sqrt{2\pi}} \exp\{-t^2/2\} dt$ is the cumulative distribution of a standard Gaussian. It then follows that the mean (or the bit density) is $\mu = \mathbb{E}[Y] = p_Y(1)$ . + +
MethodGenerator SizeDataset SizeTraining TimeFID
Standard colorizationResNet-93x1x62.34
+ Increased training timeResNet-93x1.5x54.25
+ Increased model sizeResNet-103x1x74.58
Proposed methodResNet-91x1x49.88
+ +Table 1. Performance comparison of the standard colorization approach under various training conditions versus the proposed method. + +# 3.2. Why Exposure Synthesis is Required for Colorization? + +Given a binary image $\mathbf{Y} \in \mathbb{R}^N$ , our goal is to recover a color image $\mathbf{X}_c \in \mathbb{R}^{N \times 3}$ . However, since the measurement does not contain any color information, our problem fits the classical work of image colorization [38, 58, 60, 66]. However, it is non-trivial to deploy such an extension in real world scenarios, because very likely there is a discrepancy in the exposure levels of the input images compared to those used during training. + +An example of binary images of different exposures are shown in Figure 2(a) illustrating the effect of exposure change on the details in the captured binary binary image. Consider a typical colorization network for example, an Image-to-Image translation model like Pix2Pix [34] trained on well exposed image but tested on overexposed images, the colorization results are notably poorer (shown in Figure 2(c) row 1). This is a realistic concern because the underlying exposure $\theta$ can fluctuate dramatically in real life. + +Possible solutions include (i) Exposure Augmentation : training dataset is enriched with images of varying exposures, at a cost of additional training time. We find, however, this approach partially addresses the issue, as demonstrated in our results shown in Figure 2(b) row 2. In particular, this approach falls short at over- or under-exposure. One can alleviate this issue by applying multiple colorizers, each handling a narrow range of exposures, but this is not ideal as it demands substantially more computational resources. (ii) Increased Model size : Another solution is to increase the size of generator used in Pix2Pix [34] and train for longer period of time. In Table 1, we present the colorization performance represented by the FID metric under various training scenarios. Under increased generator size or increased training time the standard colorization ap + +proach performs poorly compared to proposed method detailed in subsequent sections. + +Proposed Solution. We address the challenge by a dedicated exposure correction based on neural ordinary differential equations (neural ODEs) to standardize the exposure of binary images. This specialization allows the colorizer to focus on consistent-exposure images, resulting in notably enhanced colorization, as evidenced Figure 2(c) row 3. Additionally, our exposure synthesis can generate multiple exposures from a single binary image, enabling more effective multi-exposure colorization. Some of the added advantages of our approach include: (a) Modularity. The colorizer in our design is interchangeable, allowing for upgrades to more advanced architectures without requiring retraining on a large, augmented dataset. (b) Parameter Efficiency. Our method achieves its objectives without resorting to multiple networks with large parameter sets, ensuring a more streamlined and resource-efficient process. + +# 3.3. Continuous Exposure Synthesis using Filter Decomposed Neural ODEs + +The image formation model defined in Section 3.1 indicates that we can formulate the continuum burst of binary images as $\mathbf{Y}(\widetilde{\theta})$ , where the binary image $\mathbf{Y}$ is a function of a continuous exposure value $\widetilde{\theta} \in (0,1]$ . A small $\widetilde{\theta}$ value indicates a higher exposed image $\mathbf{Y}$ . One of the core contributions of this work is a deep neural network architecture that synthesizes a binary image $\mathbf{Y}_t$ at a target exposure $\widetilde{\theta}_{\mathrm{target}}$ given an input binary image $\mathbf{Y}_i$ measured at the exposure value $\widetilde{\theta}_{\mathrm{input}}$ . + +As the exposure values are continuous, we may need an exhaustive training set that densely samples every possible pair of input exposure value $\widetilde{\theta}_{\mathrm{input}}$ and target exposure value $\widetilde{\theta}_{\mathrm{target}}$ for every scene to train our network. To tackle this problem, we designed a convolutional architecture that enforces the parameter of the network to smoothly change according to the exposures. + +We use the convolution filter decomposition technique [56, 61, 64] to represent the convolutional filters $\mathbf{F}$ with a low number of parameters, and we leverage neural ODEs [6] to associate parameters with image exposures. With these techniques, we enable the adaptive control of filter parameters $\mathbf{F}$ and thereby the learning of the exposure synthesis operation $\mathcal{G}$ efficiently. + +Efficient Modelling of Filter Subspace. Specifically, we formulate the filter parameters $\mathbf{F}$ using the atom-coefficient [56] decomposition, which permits the smallest computational bottleneck in modeling filter subspaces among the candidates. Given a convolutional filter $\mathbf{F} \in \mathbb{R}^{c' \times c \times k \times k}$ where $c, c'$ and $k$ are the output, input channels, and spatial size, respectively. The atom-coefficient decomposition over a set of $m$ filter atoms decomposes $\mathbf{F}$ into coefficients, denoted by $\phi \in \mathbb{R}^{c' \times c \times m}$ , and atoms, denoted + +$$ +\begin{array}{l} \text {b y} \Lambda \in \mathbb {R} ^ {m \times k \times k} \\ \mathbf {F} = \phi \boldsymbol {\Lambda}. \\ \end{array} +$$ + +When $m \ll c$ , the dimension of the parameters is significantly reduced. Extending the decomposition to a L-layer convolutional network gives a set of $L$ filter atoms $\Lambda = \{\Lambda^l\}_{l=1}^L$ and $L$ coefficients $\phi = \{\phi^l\}_{l=1}^L$ . + +Enforcing Filter Atom Continuity. We enforce the filter atoms $\Lambda$ to smoothly change according to the input and target exposure levels $\widetilde{\theta}_{\mathrm{input}}$ and $\widetilde{\theta}_{\mathrm{target}}$ . To do so, we incorporate the continuous-time networks like neural ODEs to describe the relationship of $\Lambda (\widetilde{\theta}_{\mathrm{input}},\widetilde{\theta}_{\mathrm{target}})$ , as the neural ODEs provide advantages in the number of parameters required, memory consumption, and computational adaptivity [6]. The derivative of the filter atoms $d\Lambda /d\widetilde{\theta}$ w.r.t. the exposure value $\widetilde{\theta}$ as a neural network $g(\cdot)$ + +$$ +d \boldsymbol {\Lambda} / d \widetilde {\theta} = g (\widetilde {\theta} _ {\text {i n p u t}}, \widetilde {\theta} _ {\text {t a r g e t}}, \boldsymbol {\Lambda} _ {\text {i n i t}}; \mathcal {W}), +$$ + +where $\Lambda_{\mathrm{init}}$ is the initial state of filter atoms, and $\mathcal{W}$ is the learnable parameters of $g(\cdot)$ . The initial condition $\Lambda_{\mathrm{init}}$ is a learnable parameter and is optimized jointly with $\mathcal{W}$ . Given the input and target exposure values $\widetilde{\theta}_{\mathrm{input}}$ and $\widetilde{\theta}_{\mathrm{target}}$ , the corresponding filter atoms can be calculated as + +$$ +\boldsymbol {\Lambda} (\widetilde {\theta} _ {\text {t a r g e t}}, \widetilde {\theta} _ {\text {i n p u t}}) = \boldsymbol {\Lambda} _ {\text {i n i t}} + \int_ {\widetilde {\theta} _ {\text {i n p u t}}} ^ {\widetilde {\theta} _ {\text {t a r g e t}}} g (\boldsymbol {\Lambda} _ {\text {i n i t}}, \mathcal {W}) d \widetilde {\theta}. +$$ + +According to Picard-Lindelöf existence theorem [15], for a given initial condition, the solution to the IVP is unique. Theoretically, we prove continuity in filter atoms induces controlled changes in image exposure. + +Theorem 3.1 If $\Lambda_{1}$ and $\Lambda_{2}$ are filter atoms generated with neural ODE using integration intervals $(\widetilde{\theta}_{0},\widetilde{\theta}_{l})$ and $(\widetilde{\theta}_0,\widetilde{\theta}_2)$ , respectively such that $\| \Lambda_1 - \Lambda_2\| \leq \epsilon |\widetilde{\theta}_1 - \widetilde{\theta}_2|$ , for $\epsilon >0$ . If $\theta_{1}$ and $\theta_{2}$ are the exposures at $\widetilde{\theta}_{1}$ and $\widetilde{\theta}_{2}$ , and activation function $\sigma$ is non-expansive, then change in exposure $\Delta \theta$ is continuous in filter atoms. + +Details of Theorem 3.1 and its proof can be found in the Appendix section A1. An illustration of our parameter efficient approach is shown in Figure 3. + +# 3.4. Exposure Adaptive Colorization + +Our proposed exposure adaptive colorization consists of an exposure correction realized by the stated synthesis and exposure level specific colorizer. + +Exposure correction. Given a colorization method specific to exposure $\widetilde{\theta}_{\mathrm{target}}$ and a input sample $(\mathbf{Y}_i,\widetilde{\theta}_{\mathrm{input}})$ , the proposed exposure correction generates image $\hat{\mathbf{Y}}_t$ at the target exposure in two steps. With the operation $\mathcal{G}$ carried out + +![](images/e285dcaa0552d4ee4cbacedca07a215de1f7c4629c2282772c00c7ae5a8cf969.jpg) +Figure 3. An illustration of our proposed method for exposure adaptive colorization: A binary image, $\mathbf{Y}$ , which can be overexposed or underexposed, is input into the proposed exposure synthesis module. The colorization of binary image can be achieved using Single Image Colorization (SIC) or Burst Image Colorization (BIC). (i) SIC: Based on the input and target exposure levels, $\widetilde{\theta}_{\mathrm{input}}$ and $\widetilde{\theta}_{\mathrm{target}}$ , this module adjusts the weights of a exposure synthesis network $\mathcal{G}$ , which then generates an exposure-corrected image. Note that corrected image is not necessarily a binary image. Since the colorization module $\mathcal{F}$ is trained to colorize only image of specific exposure, the exposure synthesis module ensures corrected binary image has similar exposure to the one on which $\mathcal{F}$ is trained. (ii) BIC: For BIC we generate images with varying exposures as input to the burst image colorization network, $\mathcal{F}_{\mathrm{burst}}$ . The trained network is able to exploit the complementary information across multiple exposures with the help of Cross Non-Local Fusion blocks [44] to synthesize colors in regions of the image that is otherwise not possible by SIC approach. + +by some generative network, first, the filter atoms of the network's convolutional filters are computed using the neural ODEs based on the $(\widetilde{\theta}_{\mathrm{input}},\widetilde{\theta}_{\mathrm{target}})$ pair; next, the convolutional filter weights of the network are updated using the filter atoms and their corresponding coefficients, and the network generate the exposure corrected image $\hat{\mathbf{Y}}_t$ based on $\mathbf{Y}_i$ . In the case of burst colorization, multiple images are generated corresponding to a set of target exposures. + +Colorization. Both single-image-based colorization and burst-based colorization are considered in our work. The first approach involves a colorization network $\mathcal{F}:\hat{\mathbf{Y}}_t\to$ $\mathbf{X}_c$ which takes a exposure corrected binary image as input and outputs corresponding RGB image $\mathbf{X}_c$ . The network is trained exclusively at a fixed exposure level, and variations in the exposure of the input image is pre-adjusted by the exposure correction step. The second approach uses a colorization network $\mathcal{F}_{\mathrm{burst}}:\{\hat{\mathbf{Y}}_t\}_{t = 1}^K\to \mathbf{X}_c$ that takes in a burst of $K$ binary images $\{\hat{\mathbf{Y}}_t\}_{t = 1}^K$ corresponding to predefined varying exposure values $\{\hat{\theta}_t\}_{t = 1}^K$ and outputs one RGB image $\mathbf{X}_c$ . + +# 4. Experiments + +# 4.1. Experimental Setup + +We evaluate the performance of both the exposure synthesis module standalone and the entire proposed approach. We conduct exposure burst recovery experiment to evalu + +ate the exposure synthesis and Single Image Colorization (SIC) and Burst Image Colorization (BIC) experiments to evaluate the final colorization results. + +Implementation of Exposure Adaptive Colorization. We implement the neural ODEs using a 6-layer multilayer perceptron. The generative networks we use for exposure synthesis are variants of Pix2Pix [34] and CycleGAN [68], with which the models are named AtomODE-Pix2Pix and AtomODE-CycleGAN. For colorization, we train a Pix2Pix colorizer dedicated to SIC task and another to BIC. For BIC, the colorizer is additionally equipped with a Cross Non-Local Fusion (CNLF) [44] to constructively fuse the representations obtained from the burst. Note that the colorizers are trained directly on pairs of binary and color image data at fixed $\widetilde{\theta}$ , independent of exposure synthesis. In our colorization experiment, the same trained colorizer is applied on all exposure correction methods for comparison. Details can be found in the Appendix section A2. + +Datasets and Synthetic Data Generation. We train and evaluate the proposed method as well as the baselines on synthetic data generated using images from two datasets: AFHQ [14] and CelebA-HQ [36]. AFHQ dataset contains images of three classes, cats, dogs, and wild animals, and CelebA-HQ contains images of male and female human faces. + +To synthesize binary images, we use the image formation model described in Section 3.1 and apply exposure bracket + +
Exposure Synthesis MethodAFHQ (512 x 512)CelebA-HQ (256 x 256)
CatDogWildMaleFemale
MSE (↓)/RL (↑)/FID (↓)MSE (↓)/RL (↑)/FID (↓)MSE (↓)/RL (↑)/FID (↓)MSE (↓)/RL (↑)/FID (↓)MSE (↓)/RL (↑)/FID (↓)
Pix2Pix-DNI3.75 / 0.9816 / 61.895.91 / 0.9844 / 95.438.13 / 0.9780 / 39.293.74 / 0.9563 / 56.134.78 / 0.9681 / 39.81
CycleGAN-DNI9.61 / 0.9816 / 207.8216.79 / 0.9824 / 220.909.62 / 0.9820 / 336.389.31 / 0.9824 / 94.7814.26 / 0.9810 / 250.61
SAVI2I41.15 / 46.27 / 178.2946.27 / 0.9804 / 217.9838.87 / 0.9019 / 139.4767.26 / 0.9563 / 111.0569.21 / 0.9698 / 261.45
DLOW2.44 / 0.9654 / 92.302.32 / 0.9489 / 173.652.75 / 0.9632 / 179.231.87 / 0.9655 / 62.571.37 / 0.9798 / 63.31
AtomODE-Pix2Pix (ours)2.29 / 0.9997 / 57.883.17 / 0.9994 / 140.232.15 / 0.9998 / 79.762.15 / 0.9995 / 53.443.20 / 0.9995 / 40.08
AtomODE-CycleGAN (ours)1.15 / 0.9948 / 60.541.18 / 0.9971 / 87.431.23 / 0.9979 / 35.751.58 / 0.9986 / 77.821.87 / 0.9994 / 62.59
+ +Table 2. Colorization results on AFHQ and CelebA-HQ datasets. The input image for all these experiments is an overexposed image at index $\tau = 0$ of exposure burst. All the colorizers are trained using an image corresponding to $\tau = 10$ in the exposure burst. At test time we correct the overexposed input to correspond to the one used during training. In the first column of the table, we note the combination of the exposure correction method used and colorizer is same across all the methods. + +![](images/435800515c977b9ae215fe545e14af136908b2fb5937ebbdd23b1f132a724d40.jpg) +Figure 4. Qualitative results for exposure burst recovery experiments for AFHQ dataset. We compare our methods, AtomODE-Pix2Pix and AtomODE-CycleGAN, with DLOW, SAVI2I and DNI. + +ing to generate various exposure. Details of data synthesis can be found in the Appendix section A2. Our compiled test datasets consist of 15-frame bursts of exposure varying from overexposed to underexposed. We index the exposure of these images in a burst by $\tau$ , a discretized label of $\widetilde{\theta}$ . + +Baselines. In the exposure burst recovery experiment, we compare our approach against DLOW [25] and SAVI2I [52]. We further consider DNI [62] for its capability of interpolating network parameters to generate continuous variations of outputs. We prepare baselines Pix2Pix-DNI and CycleGAN-DNI by coupling DNI with Pix2Pix and Cycle + +GAN exposure generators. In colorization experiments, we apply the independently trained Pix2Pix colorizers to the exposure correction results generated by all methods. For more details on baseline training and hyperparameters, we refer readers to the Appendix section A2. + +Metrics. In the burst recovery experiment, we choose to use Mean Squared Error (MSE) and Relative Linearity (RL) [43] to distinguish a smoothly varying exposure burst with consistent appearance changes. A high RL score indicates a better transition in exposure among consecutive samples. In the colorization experiments, we assess the quality of exposure correction by computing the Fréchet Inception Distance (FID) between the colorized image and ground truth. + +# 4.2. Exposure Burst Recovery Results + +Figure 4 presents the qualitative comparison of all methods on AFHQ dataset. We find our methods, AtomODE Pix2Pix and AtomODE CycleGAN are able to generate consistent set of smoothly varying exposure images than baseline methods and match ground truth images generated from exposure bracketing. We also observe some baselines struggle with generation of overexposed frames. All methods for comparison are further assessed on both datasets using MSE and RL, with the quantitative results shown in Table 2. Results suggest that the proposed AtomODE provides consistently better performance in all testing conditions across multiple datasets. Additional qualitative results can be found in Appendix section A4. + +# 4.3. Single and Burst Image Colorization Results + +The SIC colorization results of all methods on AFHQ dataset are shown in Figure 5, where the exposure of the test samples is different from the that of the colorizer's training samples. Corrupted colored images are generated by most of the baselines, while the proposed method produces visually appealing outputs. Figure 7 shows the SIC colorization results on CelebA-HQ dataset. Additionally, we also present the qualitative results for BIC in Figure 6 where $K$ denotes the number of binary frames input to the colorizer. The quantitative results are shown in Table 2, assessed with the FID metric. + +![](images/548330aef06c66cc200aa89566c82076cec156c5839c7327bbb55a0d36983a12.jpg) +Figure 5. Colorized results for AFHQ dataset using various combinations of exposure correction methods and colorizer. The first column shows the overexposed binary input image for exposure correction. The results of our methods AtomODE Pix2Pix and AtomODE CycleGAN are shown in columns six and seven. + +# 4.4. Real Data Experiments + +As our approach is trained on synthetic data, it is essential to evaluate its performance on real-world data. To this end, we conduct experiments using both a CMOS camera and a prototype QIS camera. We refer the readers to Appendix for further details about the experiment setup. We print 100 images from the training set and 25 from the test set, and we capture pictures of them using both cameras, collecting 125 pairs of grayscale and RGB images. We fine-tune the colorization model with the 100 real data samples, and we evaluate our proposed approach on the set of 25 test data, in + +cluding both binarized CMOS data and QIS data. In Figure 8, we present qualitative results from the evaluation on the real data obtained by the CMOS camera. Despite that the proposed method was not trained with vast real sensor data, it is able to produce visually pleasing colorized images. We observe that fine-tuning the model using a small training dataset significantly improves the colorization quality. We further present results from the evaluation on images captured by the QIS camera in Figure 9. These results demonstrate the potential of the proposed method for real-world single-photon imaging applications. + +![](images/4c800f0b820fd94bab360555cc73142aedd6a6e07fa8f77a825db0c3fb3ad13a.jpg) +Groundtruth $K = 5$ $K = 1$ + +![](images/678f953f4e3d069d606f7f18b9f3f692d3fe60ed38c26f9723a3671b7762c12b.jpg) +Figure 6. An illustration of burst image colorization for AFHQ dataset using multiple binary images of varying exposures as input. The colorizer is trained using images of exposure $\tau = \{0,4,8,10,14\}$ . An overexposed input image is used to synthesize multiple images of varying exposures and passed to the colorizer. +Figure 7. Colorized results for CelebA-HQ dataset using various combinations of exposure correction methods and colorizers. The results of our method are shown in columns four and five. + +# 4.5. Benefits for Practical Applications + +Our methodology is specifically designed with modularity and parameter efficiency in mind, features crucial for real-world industry applications. The Pix2PixDNI method, the second-best performing approach, necessitates training multiple networks and either interpolating model weights on-the-fly during inference or precomputing and storing them to generate images with varied exposures. Conversely, our method circumvents the additional storage requirements and the need to train multiple model parameters by incorporating the exposure correction into the neural ODE framework. + +![](images/e76a678abea8de61fab3bce9e4dcc81f118d15b6886ffcd9854fe69d984c9e0f.jpg) +Figure 8. Results on real-world images captured from a CMOS camera (images shown along the row with a red line are colorized images, and along the green line are the ground-truth images captured using an RGB CMOS camera). + +![](images/7f5b1a31e523c5b163bc17b15a131fb6ada6f0d10867c42e87a120f5e9a4a8c3.jpg) +Figure 9. Colorized output using input from prototype QIS camera. + +# 5. Conclusion + +In this work, we proposed a generative exposure correction approach that can synthesize a spectrum of exposure given a single 1-bit image. We modeled the binary signal variation with respect to exposure change as a continuous function by adapting a small set of parameters of a network, known as filter atoms, to desired exposure using neural ODE. Theoretically and empirically, we demonstrated successful recovery of bursts of 1-bit images from a single input image, and we verified its application on single-image-based and burst-image-based 1-bit image colorization. Experiments on real CMOS and QIS data further indicated the potentials of the proposed approach on real-world single-photon imaging applications. + +# References + +[1] Samuel Burri, Yuki Maruyama, Xavier Michalet, Francesco Regazzoni, Claudio Bruschini, and Edoardo Charbon. Architecture and applications of a high resolution gated SPAD image sensor. Optics Express, 22(14):17573-17589, 2014. 1, 2 +[2] Stanley H. Chan. What does a one-bit Quanta Image Sensor offer? IEEE Transactions on Computational Imaging, 8: 770-783, 2022. 2 +[3] Stanley H. Chan and Yue M. Lu. Efficient image reconstruction for gigapixel quantum image sensors. In IEEE Global Conference on Signal and Information Processing, pages 312-316, 2014. 2 +[4] Stanley H. Chan, Omar A. Elgendy, and Xiran Wang. Images from bits: Non-iterative image reconstruction for Quanta Image Sensors. MDPI Sensors, 16(11), 2016. 2 +[5] Stanley H. Chan, Xiran Wang, and Omar A. Elgendy. Plug-and-Play ADMM for image restoration: Fixed-point convergence and applications. IEEE Transactions on Computational Imaging, 3:84–98, 2016. 2 +[6] Ricky T. Q. Chen, Yulia Rubanova, Jesse Bettencourt, and David K. Duvenaud. Neural Ordinary Differential Equations. In NeurIPS, 2018. 4 +[7] Wei Chen, Zichen Miao, and Qiang Qiu. Inner product-based neural network similarity. In NeurIPS, 2023. 2 +[8] Wei Chen, Zichen Miao, and Qiang Qiu. Large convolutional model tuning via filter subspace. Available online: https://arxiv.org/abs/2403.00269, 2024. Accessed 03/06/2024. 2 +[9] Ying-Cong Chen, Xiaogang Xu, Zhuotao Tian, and Jiaya Jia. Homomorphic latent space interpolation for unpaired Image-To-Image translation. In CVPR, pages 2403-2411, 2019. 2 +[10] Xiuyuan Cheng, Qiang Qiu, Robert Calderbank, and Guillermo Sapiro. RotDCF: Decomposition of convolutional filters for rotation-equivariant deep networks. In ICLR, 2019. 2 +[11] Yiheng Chi, Abhiram Gnanasambandam, Vladlen Koltun, and Stanley H. Chan. Dynamic low-light imaging with Quanta Image Sensors. In ECCV, pages 122-138, 2020. 1, 2 +[12] Joon Hee Choi, Omar A. Elgendy, and Stanley H. Chan. Image reconstruction for Quanta Image Sensors using deep neural networks. In IEEE ICASSP, pages 6543-6547, 2018. 2 +[13] Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. StarGAN: Unified Generative Adversarial Networks for multi-domain Image-to-Image translation. In CVPR, pages 8789–8797, 2018. 2 +[14] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. StarGAN v2: Diverse image synthesis for multiple domains. In CVPR, pages 8185-8194, 2020. 2, 5 +[15] Earl A. Coddington and Norman Levinson. Theory of Ordinary Differential Equations. McGraw-Hill, 1955. 4 +[16] Neale A. W. Dutton, Istvan Gyongy, Luca Parmesan, Salvatore Gnecchi, Neil Calder, Bruce R. Rae, Sara Pellegrini, Lindsay A. Grant, and Robert K. Henderson. A SPAD-based QVGA image sensor for single-photon counting and quanta + +imaging. IEEE Transactions on Electron Devices, 63(1): 189-196, 2016. 1, 2 +[17] Neale A. W. Dutton, Istvan Gyongy, Luca Parmesan, and Robert K. Henderson. Single photon counting performance and noise analysis of CMOS SPAD-based image sensors. MDPI Sensors, 16(7), 2016. 1, 2 +[18] Omar A. Elgendy, Abhiram Gnanasambandam, Stanley H. Chan, and Jiaju Ma. Low-light demosaicking and denoising for small pixels using learned frequency selection. IEEE Transactions on Computational Imaging, 7:137-150, 2021. 2 +[19] Eric R. Fossum. Some thoughts on future digital still cameras. In Image Sensors and Signal Processing for Digital Still Cameras, pages 305-314. Taylor and Francis Group, 2005. 2 +[20] Eric R. Fossum. Modeling the performance of single-bit and multi-bit Quanta Image Sensors. IEEE Journal of the Electron Devices Society, 1(9):166-174, 2013. 2 +[21] Eric R. Fossum. Analog read noise and quantizer threshold estimation from Quanta Image Sensor bit density. IEEE Journal of the Electron Devices Society, 10:269-274, 2022. 2 +[22] Abhiram Gnanasambandam and Stanley H. Chan. Image classification in the dark using Quanta Image Sensors. In ECCV, pages 484-501, 2020. 2 +[23] Abhiram Gnanasambandam and Stanley H. Chan. HDR imaging with Quanta Image Sensors: Theoretical limits and optimal reconstruction. IEEE Transactions on Computational Imaging, 6:1571-1585, 2020. 1 +[24] Abhiram Gnanasambandam and Stanley H. Chan. Exposure-referred signal-to-noise ratio for digital image sensors. IEEE Transactions on Computational Imaging, 8:561-575, 2022. 1 +[25] Rui Gong, Wen Li, Yuhua Chen, and Luc Van Gool. DLOW: Domain flow for adaptation and generalization. In CVPR, pages 2472-2481, 2019. 6 +[26] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative Adversarial Nets. In NeurIPS, 2014. 2 +[27] Anant Gupta, Atul Ingle, and Mohit Gupta. Asynchronous single-photon 3D imaging. In ICCV, pages 7908-7917, 2019. 2 +[28] Anant Gupta, Atul Ingle, Andreas Velten, and Mohit Gupta. Photon-flooded single-photon 3D cameras. In CVPR, pages 6763-6772, 2019. +[29] Felipe Gutierrez-Barragan, Atul Ingle, Trevor Seets, Mohit Gupta, and Andreas Velten. Compressive single-photon 3D cameras. In CVPR, pages 17833-17843, 2022. 2 +[30] Istvan Gyongy, Neale AW Dutton, and Robert K Henderson. Single-photon tracking for high-speed vision. MDPI Sensors, 18:323, 2018. 2 +[31] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 2 +[32] Atul Ingle, Andreas Velten, and Mohit Gupta. High flux passive imaging with single-photon sensors. In CVPR, pages 6753-6762, 2021. 2 + +[33] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-Image translation with conditional adversarial networks. In CVPR, pages 5967–5976, 2017. 2 +[34] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A. Efros. Image-to-Image translation with conditional adversarial networks. In CVPR, pages 5967–5976, 2017. 3, 5 +[35] Kiyotaka Iwabuchi, Yusuke Kameda, and Takayuki Hamamoto. Image quality improvements based on motion-based deblurring for single-photon imaging. IEEE Access, 9:30080-30094, 2021. 2 +[36] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In ICLR, 2018. 5 +[37] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for Generative Adversarial Networks. IEEE TPAMI, 43(12):4217-4228, 2021. 2 +[38] Manoj Kumar, Dirk Weissenborn, and Nal Kalchbrenner. Colorization transformer. In ICLR, 2021. 3 +[39] Vadim Lebedev, Yaroslav Ganin, Maksim Rakhuba, Ivan Oseledets, and Victor Lempitsky. Speeding-up convolutional neural networks using fine-tuned CP-decomposition. In ICLR, 2015. 2 +[40] Yu-Jing Lin, Po-Wei Wu, Che-Han Chang, Edward Chang, and Shih-Wei Liao. RelGAN: Multi-domain Image-to-Image translation via relative attributes. In ICCV, pages 5913–5921, 2019. 2 +[41] Wallace Lira, Johannes Merz, Daniel Ritchie, Daniel Cohen-Or, and Hao Zhang. GANHopper: Multi-hop GAN for unsupervised Image-to-Image translation. In ECCV, pages 363–379, 2020. 2 +[42] Yahui Liu, Enver Sangineto, Yajing Chen, Linchao Bao, Haoxian Zhang, Nicu Sebe, Bruno Lepri, Wei Wang, and Marco De Nadai. Smoothing the disentangled latent style space for unsupervised Image-to-Image translation. In CVPR, pages 10780–10789, 2021. 2 +[43] Kun Lu, Rongpeng Li, and Honggang Zhang. Contrastive monotonic pixel-level modulation. In ECCV, pages 493-510, 2022. 6 +[44] Ziwei Luo, Lei Yu, Xuan Mo, Youwei Li, Lanpeng Jia, Haoqiang Fan, Jian Sun, and Shuaicheng Liu. EBSR: Feature Enhanced Burst Super-Resolution with deformable alignment. In CVPR, pages 471–478, 2021. 5 +[45] Jiaju Ma and Eric Fossum. Quanta Image Sensor jot with sub 0.3e- r.m.s. read noise and photon counting capability. IEEE Electron Device Letters, 36(9):926-928, 2015. 1 +[46] Jiaju Ma and Eric R. Fossum. A Pump-Gate Jot device with high conversion gain for a Quanta Image Sensor. IEEE Journal of the Electron Devices Society, 3:73-77, 2015. +[47] Jiaju Ma, Dakota Starkey, Arun Rao, Kofi Odame, and Eric R. Fossum. Characterization of Quanta Image Sensor pump-gate jots with deep sub-electron read noise. IEEE Journal of the Electron Devices Society, 3(6):472-480, 2015. 1 +[48] Jiaju Ma, Saleh Masoodian, Dakota A Starkey, and Eric R. Fossum. Photon-number-resolving megapixel image sensor at room temperature without avalanche gain. Optica, 4(12): 1474-1481, 2017. 1 + +[49] Jiaju Ma, Dexue Zhang, Omar Elgendy, and Saleh Masoodian. A 0.19e-rms read noise 16.7mpixel stacked Quanta Image Sensor with $1.1\mu \mathrm{m}$ -pitch backside illuminated pixels. IEEE Electron Device Letters, 42(6):891-894, 2021. 1 +[50] Jiaju Ma, Stanley H. Chan, and Eric R. Fossum. Review of Quanta Image Sensors for ultralow-light imaging. IEEE Transactions on Electron Devices, 69(6):2824-2839, 2022. 1, 2 +[51] Sizhuo Ma, Paul Mos, Edoardo Charbon, and Mohit Gupta. Burst vision using single-photon cameras. In WACV, pages 5364-5374, 2021. 2 +[52] Qi Mao, Hsin-Ying Lee, Hung-Yu Tseng, Jia-Bin Huang, Siwei Ma, and Ming-Hsuan Yang. Continuous and diverse Image-to-Image translation via signed attribute vectors. IJCV, 130(2):517-549, 2020. 6 +[53] Zichen Miao, Ze Wang, Xiuyuan Cheng, and Qiang Qiu. Spatiotemporal joint filter decomposition in 3D convolutional neural networks. In NeurIPS, 2021. 2 +[54] Zichen Miao, Ze Wang, Wei Chen, and Qiang Qiu. Continual learning with filter atom swapping. In ICLR, 2022. 2 +[55] Fabio Pizzati, Pietro Cerri, and Raoul de Charette. CoMoGAN: Continuous model-guided Image-to-Image translation. In CVPR, pages 14283–14293, 2021. 2 +[56] Qiang Qiu, Xiuyuan Cheng, Robert Calderbank, and Guillermo Sapiro. DCFNet: Deep neural network with decomposed convolutional filters. In ICML, 2018. 2, 4 +[57] Trevor Seets, Atul Ingle, Martin Laurenzis, and Andreas Velten. Motion adaptive deblurring with single-photon cameras. In WACV, pages 1944–1953, 2021. 2 +[58] Jheng-Wei Su, Hung-Kuo Chu, and Jia-Bin Huang. Instance-aware image colorization. In CVPR, pages 7965–7974, 2020. 3 +[59] Varun Sundar, Sizhuo Ma, Aswin Sankaranarayanan, and Mohit Gupta. Single-photon structured light. In CVPR, pages 17844-17854, 2022. 2 +[60] Patricia Vitoria, Lara Raad, and Coloma Ballester. ChromaGAN: Adversarial picture colorization with semantic class distribution. In WACV, pages 2434-2443, 2020. 3 +[61] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional GANs. In CVPR, pages 8798-8807, 2018. 2, 4 +[62] X. Wang, K. Yu, C. Dong, X. Tang, and C. Loy. Deep network interpolation for continuous imagery effect transition. In CVPR, pages 1692-1701, 2019. 6 +[63] Ze Wang. Efficient Adaptation of Deep Vision Models. PhD thesis, Purdue University, 2023. Available online: https://hammer.purdue.edu/articles/thesis/Efficient_Adaptation_of_Deep_Vision_CModels/22696129.2 +[64] Ze Wang, Xiuyuan Cheng, Guillermo Sapiro, and Qiang Qiu. Stochastic conditional generative networks with basis decomposition. In ICLR, 2020. 2, 4 +[65] Ze Wang, Seunghyun Hwang, Zichen Miao, and Qiang Qiu. Image generation using continuous filter atoms. In NeurIPS, 2021. 2 + +[66] Yanze Wu, Xintao Wang, Yu Li, Honglun Zhang, Xun Zhao, and Ying Shan. Towards vivid and diverse image colorization with generative color prior. In ICCV, pages 14357-14366, 2021. 3 +[67] Zhaoyang Yin, Jiaju Ma, Saleh Masoodian, and Eric R Fossum. Threshold uniformity improvement in 1b Quanta Image Sensor readout circuit. MDPI Sensors, 22(7):2578, 2022. 2 +[68] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In ICCV, pages 2242-2251, 2017. 2, 5 +[69] Wei Zhu, Qiang Qiu, Robert Calderbank, Guillermo Sapiro, and Xiuyuan Cheng. Scaling-translation-equivariant networks with decomposed convolutional filters. Journal of Machine Learning Research, 23(68):1-45, 2022. 2 \ No newline at end of file diff --git a/2024/Generative Quanta Color Imaging/images.zip b/2024/Generative Quanta Color Imaging/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1128f80aaa63df9856df0b118abf1a0ed8570121 --- /dev/null +++ b/2024/Generative Quanta Color Imaging/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c24bbfdbec35887bc2fe4043f5bf74d081cfafae5a649508a96070fd2ea6843a +size 1018159 diff --git a/2024/Generative Quanta Color Imaging/layout.json b/2024/Generative Quanta Color Imaging/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a0965d3ea56e954a1e93b655e5b11e8fc52ec2a4 --- /dev/null +++ b/2024/Generative Quanta Color Imaging/layout.json @@ -0,0 +1,8564 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 190, + 103, + 403, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 103, + 403, + 121 + ], + "spans": [ + { + "bbox": [ + 190, + 103, + 403, + 121 + ], + "type": "text", + "content": "Generative Quanta Color Imaging" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 143, + 483, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 143, + 483, + 185 + ], + "spans": [ + { + "bbox": [ + 111, + 143, + 483, + 185 + ], + "type": "text", + "content": "Vishal Purohit, Junjie Luo, Yiheng Chi, Qi Guo, Stanley H. Chan, Qiang Qiu Elmore Family School of Electrical and Computer Engineering Purdue University, West Lafayette, USA." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 187, + 443, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 187, + 443, + 199 + ], + "spans": [ + { + "bbox": [ + 146, + 187, + 443, + 199 + ], + "type": "text", + "content": "{purohitv,luo330,chi14,guo675,stanchan,qqiu}@purdue.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 252, + 290, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 252, + 290, + 492 + ], + "spans": [ + { + "bbox": [ + 45, + 252, + 290, + 492 + ], + "type": "text", + "content": "The astonishing development of single-photon cameras has created an unprecedented opportunity for scientific and industrial imaging. However, the high data throughput generated by these 1-bit sensors creates a significant bottleneck for low-power applications. In this paper, we explore the possibility of generating a color image from a single binary frame of a single-photon camera. We evidently find this problem being particularly difficult to standard colorization approaches due to the substantial degree of exposure variation. The core innovation of our paper is an exposure synthesis model framed under a neural ordinary differential equation (Neural ODE) that allows us to generate a continuum of exposures from a single observation. This innovation ensures consistent exposure in binary images that colorizers take on, resulting in notably enhanced colorization. We demonstrate applications of the method in single-image and burst colorization and show superior generative performance over baselines. Project website can be found at https://vishal-s-p.github.io/projects/2023/generative_quanta_color.html" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 500, + 128, + 512 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 500, + 128, + 512 + ], + "spans": [ + { + "bbox": [ + 47, + 500, + 128, + 512 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 521, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 677 + ], + "type": "text", + "content": "Single-photon image sensors are gaining a strong momentum over the past decade due to their impeccable photon-counting capability. This not only opens the door for advanced scientific imaging in low-light environment, but also creates the opportunity for new imaging in high-speed, high dynamic range, and low-bit conditions. Single-photon imaging today are realized through single-photon avalanche diodes (SPAD) and quanta image sensors (QIS). The latest development of QIS has demonstrated a mega-pixel resolution [48] which is compatible with mainstream cell phone cameras, whereas SPAD has achieved hundreds of thousands of (binary) frames per second which has enabled a plethora of industrial and 3D imaging applications [16]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "content": "Problem Statement. In the context of 1-bit single-photon imaging, the data volume generated is huge. Taking a " + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "inline_equation", + "content": "1000 \\times 1000" + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "content": " SPAD array as an example, a video of" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 307, + 226, + 476, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 226, + 476, + 235 + ], + "spans": [ + { + "bbox": [ + 307, + 226, + 476, + 235 + ], + "type": "text", + "content": "Binary Input Pix2PixDNI Our Method Groundtruth" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 236, + 477, + 365 + ], + "blocks": [ + { + "bbox": [ + 307, + 236, + 477, + 365 + ], + "lines": [ + { + "bbox": [ + 307, + 236, + 477, + 365 + ], + "spans": [ + { + "bbox": [ + 307, + 236, + 477, + 365 + ], + "type": "image", + "image_path": "900a82ebe2a17662b8c37752d6039080a656a4dc93c8d4757103bd60fa57fafd.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 374, + 546, + 473 + ], + "lines": [ + { + "bbox": [ + 305, + 374, + 546, + 473 + ], + "spans": [ + { + "bbox": [ + 305, + 374, + 546, + 473 + ], + "type": "text", + "content": "Figure 1. We introduce generative quanta color imaging. Given a binary frame captured by a single-photon camera (quanta image sensor (QIS) in this example), the proposed method generates a continuum of exposures using a neural ordinary differential equation framework. Color are then generated based on these exposures. Left: A qualitative comparison between our approach and existing methods is shown. Right: Comparison between colorization results and image captured using a RGB CMOS camera under similar conditions." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 485, + 235, + 545, + 365 + ], + "blocks": [ + { + "bbox": [ + 485, + 235, + 545, + 365 + ], + "lines": [ + { + "bbox": [ + 485, + 235, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 485, + 235, + 545, + 365 + ], + "type": "image", + "image_path": "e5467101037743f5643ff514b2579ea22fa07905613a732a24a27f0aa1ec7d75.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 489, + 546, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 489, + 546, + 644 + ], + "spans": [ + { + "bbox": [ + 304, + 489, + 546, + 644 + ], + "type": "text", + "content": "500 binary frames requires a " + }, + { + "bbox": [ + 304, + 489, + 546, + 644 + ], + "type": "inline_equation", + "content": "500\\mathrm{Mb / s}" + }, + { + "bbox": [ + 304, + 489, + 546, + 644 + ], + "type": "text", + "content": " data transfer rate; if the frame rate becomes " + }, + { + "bbox": [ + 304, + 489, + 546, + 644 + ], + "type": "inline_equation", + "content": "150\\mathrm{k}" + }, + { + "bbox": [ + 304, + 489, + 546, + 644 + ], + "type": "text", + "content": " fps, the data transfer rate increases to " + }, + { + "bbox": [ + 304, + 489, + 546, + 644 + ], + "type": "inline_equation", + "content": "150\\mathrm{Gb / s}" + }, + { + "bbox": [ + 304, + 489, + 546, + 644 + ], + "type": "text", + "content": ". Since data throughput scales proportionally with the power consumption, it quickly becomes infeasible to use these sensors continuously without considering power supply, memory, and transmission issues. While solid progress are being made across the sensor and signal processing community [1, 11, 17, 23, 24, 45-47, 49, 50], the data throughput remains an unsolved problem for single-photon sensors to be considered applicable in many applications, including low-power devices, wireless devices, underwater vehicles, spacecrafts, etc., where bandwidth is the bottleneck." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 647, + 545, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 545, + 671 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 545, + 671 + ], + "type": "text", + "content": "In this paper, instead of compressing data or redesigning the sensor, we ask the question" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 683, + 531, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 683, + 531, + 708 + ], + "spans": [ + { + "bbox": [ + 321, + 683, + 531, + 708 + ], + "type": "text", + "content": "Is it possible to recover a color image from a single 1-bit image?" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25138" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 262 + ], + "type": "text", + "content": "Significance. There are several reasons why this problem is significant. Firstly, while there is an ocean of literature on image generation (through methods such as Generative Adversarial Networks [26] and diffusion [31]), our problem is unique in the sense none of the existing work addresses it. Besides being binary (and hence colorless), in practice the input data usually also suffers from under- or over-exposures. Therefore, before generating color, it is even more critical to synthesize the unseen exposure. Secondly, any solution offered to our problem will provide a path to generative imaging in the future. From the point of view of compressed acquisition, our method demonstrates the extreme case of compression where the incoming data is a single bit. For future AR/VR applications where power is severely limited, our approach shows the possibility of extreme data compression." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 263, + 287, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 263, + 287, + 287 + ], + "spans": [ + { + "bbox": [ + 47, + 263, + 287, + 287 + ], + "type": "text", + "content": "Contributions. The core algorithmic contribution of this paper is two-fold" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 288, + 287, + 407 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 47, + 288, + 287, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 288, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 47, + 288, + 287, + 370 + ], + "type": "text", + "content": "- We propose a neural ordinary differential equation (neural ODE) based framework to synthesize a continuum of exposures that are not available in the measurement. By controlling the integration interval of neural ODE, we obtain convolutional filters that are responsible for converting input representation to image representation of desired exposure." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 371, + 287, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 371, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 47, + 371, + 287, + 407 + ], + "type": "text", + "content": "- We theoretically and empirically show that a controlled variation of filter atoms leads to controlled exposure changes in the output image." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 407, + 287, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 407, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 287, + 443 + ], + "type": "text", + "content": "We verify our proposed method using both synthetic and real single-photon image sensor data. A snapshot of the results is shown in Fig. 1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 454, + 134, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 454, + 134, + 466 + ], + "spans": [ + { + "bbox": [ + 47, + 454, + 134, + 466 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 474, + 289, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 289, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 289, + 688 + ], + "type": "text", + "content": "Single-Photon Imaging. The context of this paper is single-photon image sensors, of which we are mostly interested in one-bit sensing because of its potential in high-speed and high dynamic range imaging [2]. Today, 1-bit sensing has been realized using quanta image sensors (QIS) [19-21, 50, 67] and single-photon avalanche diodes (SPAD) [1, 16, 17]. From an image reconstruction point of view, since 1-bit data contains so little information, all existing image reconstruction methods need to use multiple frames. For QIS, these methods range from classical optimization [3, 5], transform-denoise [4], neural network [12], color demosaicking [18], frequency demodulation [18], and student-teacher learning [11, 22]. For SPAD, quanta burst photography [51] and its variants [30, 32, 35, 57] first align frames and then apply image fusions [27-29, 59]. The question we ask in this paper is to push the input requirement to the limit: can we recover/generate a color image from a binary input?" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "Image-to-Image Translation. This paper focuses on using generative models. Generative Adversarial Networks" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 239 + ], + "type": "text", + "content": "(GANs) [26] have been a significant breakthrough in the field of image translation. Pix2Pix [33] and CycleGAN [68] demonstrated the effectiveness of GANs in image-to-image translation tasks, with paired and unpaired datasets, respectively. Several subsequent works have focused on improving image fidelity, disentangling style and content, and generating images with varying styles, such as Pix2PixHD [61], StarGAN [13], StyleGAN [37], and StyleGANv2 [14]. Recent advancements in continuous cross-domain translation techniques have further improved the output quality by leveraging images from intermediate domains. These techniques include interpolation of networks [40, 41], interpolation of two latent vectors [14, 37], and exploiting the interpolation path and translation of manifold [9, 42, 55]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 240, + 546, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 240, + 546, + 468 + ], + "spans": [ + { + "bbox": [ + 304, + 240, + 546, + 468 + ], + "type": "text", + "content": "Convolutional Filter Decomposition. In the proposed method, one of the key components is the convolutional filter decomposition. Several techniques have been proposed to decompose filters of Convolutional Neural Networks (CNNs) in order to reduce the computation complexity by exploiting the redundancy in convolutional filters [10, 39, 53, 54, 63, 69]. Filter decomposition methods achieve similar performance as compared to the original CNNs while offering speedups in computation. As it is expensive to model the high-dimension space of convolutional filters directly, Qiu et al. [56] observed that the convolutional filters can often be approximated using a linear combination small set of basis elements. By decomposing filters into a small set of basis, Wang et al. [64, 65] were able to sample the basis and demonstrated theoretical and empirical results on stochastic image generation tasks. Futher, Chen et al. [7] showed that filter decomposition can be used to calculate network similarity and large convolutional models can be fine tuned using filter decomposition [8]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 482, + 423, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 482, + 423, + 496 + ], + "spans": [ + { + "bbox": [ + 306, + 482, + 423, + 496 + ], + "type": "text", + "content": "3. Proposed Approach" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 502, + 484, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 502, + 484, + 515 + ], + "spans": [ + { + "bbox": [ + 306, + 502, + 484, + 515 + ], + "type": "text", + "content": "3.1. One-Bit Image Formation Model" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 521, + 545, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 559 + ], + "type": "text", + "content": "We consider a simplified image formation model for 1-bit image sensors. Letting " + }, + { + "bbox": [ + 304, + 521, + 545, + 559 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 521, + 545, + 559 + ], + "type": "text", + "content": " be the underlying quanta exposure [20], the observed signal is defined by" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 334, + 568, + 516, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 568, + 516, + 590 + ], + "spans": [ + { + "bbox": [ + 334, + 568, + 516, + 590 + ], + "type": "interline_equation", + "content": "Y = \\operatorname {A D C} \\left\\{\\text {P o i s s o n} (\\theta) + \\text {G a u s s i a n} \\left(0, \\sigma_ {r} ^ {2}\\right) \\right\\},", + "image_path": "aed275e930ddd2c09d339cfb6240a840dcbae18a2965c64a5f6c1eab80dde02f.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 599, + 545, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 599, + 545, + 682 + ], + "spans": [ + { + "bbox": [ + 304, + 599, + 545, + 682 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 599, + 545, + 682 + ], + "type": "inline_equation", + "content": "\\sigma_r" + }, + { + "bbox": [ + 304, + 599, + 545, + 682 + ], + "type": "text", + "content": " denotes the standard deviation of read noise, and ADC is the analog-to-digital conversion. For 1-bit sensing, the ADC can be modeled as a threshold such that " + }, + { + "bbox": [ + 304, + 599, + 545, + 682 + ], + "type": "inline_equation", + "content": "\\mathrm{ADC}(x) = 1" + }, + { + "bbox": [ + 304, + 599, + 545, + 682 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 304, + 599, + 545, + 682 + ], + "type": "inline_equation", + "content": "x \\geq q" + }, + { + "bbox": [ + 304, + 599, + 545, + 682 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 599, + 545, + 682 + ], + "type": "inline_equation", + "content": "\\mathrm{ADC}(x) = 0" + }, + { + "bbox": [ + 304, + 599, + 545, + 682 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 304, + 599, + 545, + 682 + ], + "type": "inline_equation", + "content": "x < q" + }, + { + "bbox": [ + 304, + 599, + 545, + 682 + ], + "type": "text", + "content": " for some threshold " + }, + { + "bbox": [ + 304, + 599, + 545, + 682 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 304, + 599, + 545, + 682 + ], + "type": "text", + "content": ". Since " + }, + { + "bbox": [ + 304, + 599, + 545, + 682 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 304, + 599, + 545, + 682 + ], + "type": "text", + "content": " is a binary random variable, it is fully characterized by the probability mass function, which is [2]" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 356, + 685, + 494, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 685, + 494, + 715 + ], + "spans": [ + { + "bbox": [ + 356, + 685, + 494, + 715 + ], + "type": "interline_equation", + "content": "p _ {Y} (1) = \\sum_ {k = 0} ^ {\\infty} \\frac {e ^ {- \\theta} \\theta^ {k}}{k !} \\Phi \\left(\\frac {k - q}{\\sigma_ {r}}\\right),", + "image_path": "25075a8d95ae362fa36d1d780cd8e8ac359ba7cc866ab62a64ba721ccf212e71.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25139" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 360, + 163 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 360, + 163 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 360, + 163 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 360, + 163 + ], + "type": "image", + "image_path": "0a412ef20fc6665ff386e6b92ac579b7530768973b6fe897ccb70c41383edb7d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 49, + 163, + 360, + 304 + ], + "blocks": [ + { + "bbox": [ + 49, + 163, + 360, + 304 + ], + "lines": [ + { + "bbox": [ + 49, + 163, + 360, + 304 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 360, + 304 + ], + "type": "image", + "image_path": "d19eea29fd9623bbe343a5cfd25e45c74b76b4ed0d542eaf8effb270fc75c590.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 313, + 548, + 401 + ], + "lines": [ + { + "bbox": [ + 46, + 313, + 548, + 401 + ], + "spans": [ + { + "bbox": [ + 46, + 313, + 548, + 401 + ], + "type": "text", + "content": "Figure 2. Illustration of exposure correction and colorization of binary images using neural networks. (a) depicts a range of images from overexposed to underexposed, illustrating the degradation of image details due to exposure variation. (b) contrasts the standard colorization workflow and our proposed approach. (i) In standard colorization approaches, a neural network learns to map a binary image " + }, + { + "bbox": [ + 46, + 313, + 548, + 401 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 46, + 313, + 548, + 401 + ], + "type": "text", + "content": " to the corresponding color image " + }, + { + "bbox": [ + 46, + 313, + 548, + 401 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_c" + }, + { + "bbox": [ + 46, + 313, + 548, + 401 + ], + "type": "text", + "content": " via a neural network, " + }, + { + "bbox": [ + 46, + 313, + 548, + 401 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{\\mathrm{aug}}" + }, + { + "bbox": [ + 46, + 313, + 548, + 401 + ], + "type": "text", + "content": ", where superscript 'aug' indicates the colorizer is trained using dataset with augmented exposure images. (ii) In contrast, our approach does not require training colorizer with augmented exposure images. (c) compares the colorization results: the first row is the output of a colorizer trained without augmentation, the second row is the output of colorizer trained with augmented data, the third row corresponds to the results of our method and the last provides the ground truth images for reference." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 365, + 79, + 544, + 305 + ], + "blocks": [ + { + "bbox": [ + 365, + 79, + 544, + 305 + ], + "lines": [ + { + "bbox": [ + 365, + 79, + 544, + 305 + ], + "spans": [ + { + "bbox": [ + 365, + 79, + 544, + 305 + ], + "type": "image", + "image_path": "49d6687fb09d1b4eec2811a3145e43b77894b9acfbf2b080262ed524ede45462.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 421, + 288, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 421, + 288, + 461 + ], + "spans": [ + { + "bbox": [ + 46, + 421, + 288, + 461 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 421, + 288, + 461 + ], + "type": "inline_equation", + "content": "\\Phi(z) = \\int_{-\\infty}^{z} \\frac{1}{\\sqrt{2\\pi}} \\exp\\{-t^2/2\\} dt" + }, + { + "bbox": [ + 46, + 421, + 288, + 461 + ], + "type": "text", + "content": " is the cumulative distribution of a standard Gaussian. It then follows that the mean (or the bit density) is " + }, + { + "bbox": [ + 46, + 421, + 288, + 461 + ], + "type": "inline_equation", + "content": "\\mu = \\mathbb{E}[Y] = p_Y(1)" + }, + { + "bbox": [ + 46, + 421, + 288, + 461 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 47, + 471, + 288, + 522 + ], + "blocks": [ + { + "bbox": [ + 47, + 471, + 288, + 522 + ], + "lines": [ + { + "bbox": [ + 47, + 471, + 288, + 522 + ], + "spans": [ + { + "bbox": [ + 47, + 471, + 288, + 522 + ], + "type": "table", + "html": "
MethodGenerator SizeDataset SizeTraining TimeFID
Standard colorizationResNet-93x1x62.34
+ Increased training timeResNet-93x1.5x54.25
+ Increased model sizeResNet-103x1x74.58
Proposed methodResNet-91x1x49.88
", + "image_path": "0108bdf016ace58c99108d00d4ec4dd407feba3e8002819849009497c5a661c1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 529, + 288, + 563 + ], + "lines": [ + { + "bbox": [ + 46, + 529, + 288, + 563 + ], + "spans": [ + { + "bbox": [ + 46, + 529, + 288, + 563 + ], + "type": "text", + "content": "Table 1. Performance comparison of the standard colorization approach under various training conditions versus the proposed method." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 586, + 287, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 287, + 611 + ], + "type": "text", + "content": "3.2. Why Exposure Synthesis is Required for Colorization?" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": "Given a binary image " + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{Y} \\in \\mathbb{R}^N" + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": ", our goal is to recover a color image " + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_c \\in \\mathbb{R}^{N \\times 3}" + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": ". However, since the measurement does not contain any color information, our problem fits the classical work of image colorization [38, 58, 60, 66]. However, it is non-trivial to deploy such an extension in real world scenarios, because very likely there is a discrepancy in the exposure levels of the input images compared to those used during training." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 422, + 547, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 547, + 531 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 547, + 531 + ], + "type": "text", + "content": "An example of binary images of different exposures are shown in Figure 2(a) illustrating the effect of exposure change on the details in the captured binary binary image. Consider a typical colorization network for example, an Image-to-Image translation model like Pix2Pix [34] trained on well exposed image but tested on overexposed images, the colorization results are notably poorer (shown in Figure 2(c) row 1). This is a realistic concern because the underlying exposure " + }, + { + "bbox": [ + 304, + 422, + 547, + 531 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 422, + 547, + 531 + ], + "type": "text", + "content": " can fluctuate dramatically in real life." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": "Possible solutions include (i) Exposure Augmentation : training dataset is enriched with images of varying exposures, at a cost of additional training time. We find, however, this approach partially addresses the issue, as demonstrated in our results shown in Figure 2(b) row 2. In particular, this approach falls short at over- or under-exposure. One can alleviate this issue by applying multiple colorizers, each handling a narrow range of exposures, but this is not ideal as it demands substantially more computational resources. (ii) Increased Model size : Another solution is to increase the size of generator used in Pix2Pix [34] and train for longer period of time. In Table 1, we present the colorization performance represented by the FID metric under various training scenarios. Under increased generator size or increased training time the standard colorization ap" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25140" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": "proach performs poorly compared to proposed method detailed in subsequent sections." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 96, + 289, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 96, + 289, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 96, + 289, + 289 + ], + "type": "text", + "content": "Proposed Solution. We address the challenge by a dedicated exposure correction based on neural ordinary differential equations (neural ODEs) to standardize the exposure of binary images. This specialization allows the colorizer to focus on consistent-exposure images, resulting in notably enhanced colorization, as evidenced Figure 2(c) row 3. Additionally, our exposure synthesis can generate multiple exposures from a single binary image, enabling more effective multi-exposure colorization. Some of the added advantages of our approach include: (a) Modularity. The colorizer in our design is interchangeable, allowing for upgrades to more advanced architectures without requiring retraining on a large, augmented dataset. (b) Parameter Efficiency. Our method achieves its objectives without resorting to multiple networks with large parameter sets, ensuring a more streamlined and resource-efficient process." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 293, + 287, + 318 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 293, + 287, + 318 + ], + "spans": [ + { + "bbox": [ + 47, + 293, + 287, + 318 + ], + "type": "text", + "content": "3.3. Continuous Exposure Synthesis using Filter Decomposed Neural ODEs" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 323, + 287, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 323, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 287, + 435 + ], + "type": "text", + "content": "The image formation model defined in Section 3.1 indicates that we can formulate the continuum burst of binary images as " + }, + { + "bbox": [ + 46, + 323, + 287, + 435 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}(\\widetilde{\\theta})" + }, + { + "bbox": [ + 46, + 323, + 287, + 435 + ], + "type": "text", + "content": ", where the binary image " + }, + { + "bbox": [ + 46, + 323, + 287, + 435 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 46, + 323, + 287, + 435 + ], + "type": "text", + "content": " is a function of a continuous exposure value " + }, + { + "bbox": [ + 46, + 323, + 287, + 435 + ], + "type": "inline_equation", + "content": "\\widetilde{\\theta} \\in (0,1]" + }, + { + "bbox": [ + 46, + 323, + 287, + 435 + ], + "type": "text", + "content": ". A small " + }, + { + "bbox": [ + 46, + 323, + 287, + 435 + ], + "type": "inline_equation", + "content": "\\widetilde{\\theta}" + }, + { + "bbox": [ + 46, + 323, + 287, + 435 + ], + "type": "text", + "content": " value indicates a higher exposed image " + }, + { + "bbox": [ + 46, + 323, + 287, + 435 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 46, + 323, + 287, + 435 + ], + "type": "text", + "content": ". One of the core contributions of this work is a deep neural network architecture that synthesizes a binary image " + }, + { + "bbox": [ + 46, + 323, + 287, + 435 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_t" + }, + { + "bbox": [ + 46, + 323, + 287, + 435 + ], + "type": "text", + "content": " at a target exposure " + }, + { + "bbox": [ + 46, + 323, + 287, + 435 + ], + "type": "inline_equation", + "content": "\\widetilde{\\theta}_{\\mathrm{target}}" + }, + { + "bbox": [ + 46, + 323, + 287, + 435 + ], + "type": "text", + "content": " given an input binary image " + }, + { + "bbox": [ + 46, + 323, + 287, + 435 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_i" + }, + { + "bbox": [ + 46, + 323, + 287, + 435 + ], + "type": "text", + "content": " measured at the exposure value " + }, + { + "bbox": [ + 46, + 323, + 287, + 435 + ], + "type": "inline_equation", + "content": "\\widetilde{\\theta}_{\\mathrm{input}}" + }, + { + "bbox": [ + 46, + 323, + 287, + 435 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 435, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 435, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 435, + 287, + 521 + ], + "type": "text", + "content": "As the exposure values are continuous, we may need an exhaustive training set that densely samples every possible pair of input exposure value " + }, + { + "bbox": [ + 46, + 435, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\widetilde{\\theta}_{\\mathrm{input}}" + }, + { + "bbox": [ + 46, + 435, + 287, + 521 + ], + "type": "text", + "content": " and target exposure value " + }, + { + "bbox": [ + 46, + 435, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\widetilde{\\theta}_{\\mathrm{target}}" + }, + { + "bbox": [ + 46, + 435, + 287, + 521 + ], + "type": "text", + "content": " for every scene to train our network. To tackle this problem, we designed a convolutional architecture that enforces the parameter of the network to smoothly change according to the exposures." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 522, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 605 + ], + "type": "text", + "content": "We use the convolution filter decomposition technique [56, 61, 64] to represent the convolutional filters " + }, + { + "bbox": [ + 46, + 522, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 46, + 522, + 287, + 605 + ], + "type": "text", + "content": " with a low number of parameters, and we leverage neural ODEs [6] to associate parameters with image exposures. With these techniques, we enable the adaptive control of filter parameters " + }, + { + "bbox": [ + 46, + 522, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 46, + 522, + 287, + 605 + ], + "type": "text", + "content": " and thereby the learning of the exposure synthesis operation " + }, + { + "bbox": [ + 46, + 522, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 46, + 522, + 287, + 605 + ], + "type": "text", + "content": " efficiently." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": "Efficient Modelling of Filter Subspace. Specifically, we formulate the filter parameters " + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": " using the atom-coefficient [56] decomposition, which permits the smallest computational bottleneck in modeling filter subspaces among the candidates. Given a convolutional filter " + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{F} \\in \\mathbb{R}^{c' \\times c \\times k \\times k}" + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "c, c'" + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": " are the output, input channels, and spatial size, respectively. The atom-coefficient decomposition over a set of " + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": " filter atoms decomposes " + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": " into coefficients, denoted by " + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\phi \\in \\mathbb{R}^{c' \\times c \\times m}" + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": ", and atoms, denoted" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 71, + 446, + 106 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 71, + 446, + 106 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 446, + 106 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\text {b y} \\Lambda \\in \\mathbb {R} ^ {m \\times k \\times k} \\\\ \\mathbf {F} = \\phi \\boldsymbol {\\Lambda}. \\\\ \\end{array}", + "image_path": "d7442b71d1920b59f539c846db7e645052a9557c9a82f4261f763aa5c29e68b4.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 118, + 545, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 118, + 545, + 166 + ], + "spans": [ + { + "bbox": [ + 305, + 118, + 545, + 166 + ], + "type": "text", + "content": "When " + }, + { + "bbox": [ + 305, + 118, + 545, + 166 + ], + "type": "inline_equation", + "content": "m \\ll c" + }, + { + "bbox": [ + 305, + 118, + 545, + 166 + ], + "type": "text", + "content": ", the dimension of the parameters is significantly reduced. Extending the decomposition to a L-layer convolutional network gives a set of " + }, + { + "bbox": [ + 305, + 118, + 545, + 166 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 305, + 118, + 545, + 166 + ], + "type": "text", + "content": " filter atoms " + }, + { + "bbox": [ + 305, + 118, + 545, + 166 + ], + "type": "inline_equation", + "content": "\\Lambda = \\{\\Lambda^l\\}_{l=1}^L" + }, + { + "bbox": [ + 305, + 118, + 545, + 166 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 118, + 545, + 166 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 305, + 118, + 545, + 166 + ], + "type": "text", + "content": " coefficients " + }, + { + "bbox": [ + 305, + 118, + 545, + 166 + ], + "type": "inline_equation", + "content": "\\phi = \\{\\phi^l\\}_{l=1}^L" + }, + { + "bbox": [ + 305, + 118, + 545, + 166 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 166, + 545, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 166, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 304, + 166, + 545, + 277 + ], + "type": "text", + "content": "Enforcing Filter Atom Continuity. We enforce the filter atoms " + }, + { + "bbox": [ + 304, + 166, + 545, + 277 + ], + "type": "inline_equation", + "content": "\\Lambda" + }, + { + "bbox": [ + 304, + 166, + 545, + 277 + ], + "type": "text", + "content": " to smoothly change according to the input and target exposure levels " + }, + { + "bbox": [ + 304, + 166, + 545, + 277 + ], + "type": "inline_equation", + "content": "\\widetilde{\\theta}_{\\mathrm{input}}" + }, + { + "bbox": [ + 304, + 166, + 545, + 277 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 166, + 545, + 277 + ], + "type": "inline_equation", + "content": "\\widetilde{\\theta}_{\\mathrm{target}}" + }, + { + "bbox": [ + 304, + 166, + 545, + 277 + ], + "type": "text", + "content": ". To do so, we incorporate the continuous-time networks like neural ODEs to describe the relationship of " + }, + { + "bbox": [ + 304, + 166, + 545, + 277 + ], + "type": "inline_equation", + "content": "\\Lambda (\\widetilde{\\theta}_{\\mathrm{input}},\\widetilde{\\theta}_{\\mathrm{target}})" + }, + { + "bbox": [ + 304, + 166, + 545, + 277 + ], + "type": "text", + "content": ", as the neural ODEs provide advantages in the number of parameters required, memory consumption, and computational adaptivity [6]. The derivative of the filter atoms " + }, + { + "bbox": [ + 304, + 166, + 545, + 277 + ], + "type": "inline_equation", + "content": "d\\Lambda /d\\widetilde{\\theta}" + }, + { + "bbox": [ + 304, + 166, + 545, + 277 + ], + "type": "text", + "content": " w.r.t. the exposure value " + }, + { + "bbox": [ + 304, + 166, + 545, + 277 + ], + "type": "inline_equation", + "content": "\\widetilde{\\theta}" + }, + { + "bbox": [ + 304, + 166, + 545, + 277 + ], + "type": "text", + "content": " as a neural network " + }, + { + "bbox": [ + 304, + 166, + 545, + 277 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 354, + 287, + 497, + 302 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 287, + 497, + 302 + ], + "spans": [ + { + "bbox": [ + 354, + 287, + 497, + 302 + ], + "type": "interline_equation", + "content": "d \\boldsymbol {\\Lambda} / d \\widetilde {\\theta} = g (\\widetilde {\\theta} _ {\\text {i n p u t}}, \\widetilde {\\theta} _ {\\text {t a r g e t}}, \\boldsymbol {\\Lambda} _ {\\text {i n i t}}; \\mathcal {W}),", + "image_path": "203fe6c8d4eb5c61c60f418370f2843788496eddd5d69d9c8a2b1d5898f7dbfd.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 311, + 545, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 311, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 304, + 311, + 545, + 373 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 311, + 545, + 373 + ], + "type": "inline_equation", + "content": "\\Lambda_{\\mathrm{init}}" + }, + { + "bbox": [ + 304, + 311, + 545, + 373 + ], + "type": "text", + "content": " is the initial state of filter atoms, and " + }, + { + "bbox": [ + 304, + 311, + 545, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 304, + 311, + 545, + 373 + ], + "type": "text", + "content": " is the learnable parameters of " + }, + { + "bbox": [ + 304, + 311, + 545, + 373 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 304, + 311, + 545, + 373 + ], + "type": "text", + "content": ". The initial condition " + }, + { + "bbox": [ + 304, + 311, + 545, + 373 + ], + "type": "inline_equation", + "content": "\\Lambda_{\\mathrm{init}}" + }, + { + "bbox": [ + 304, + 311, + 545, + 373 + ], + "type": "text", + "content": " is a learnable parameter and is optimized jointly with " + }, + { + "bbox": [ + 304, + 311, + 545, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 304, + 311, + 545, + 373 + ], + "type": "text", + "content": ". Given the input and target exposure values " + }, + { + "bbox": [ + 304, + 311, + 545, + 373 + ], + "type": "inline_equation", + "content": "\\widetilde{\\theta}_{\\mathrm{input}}" + }, + { + "bbox": [ + 304, + 311, + 545, + 373 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 311, + 545, + 373 + ], + "type": "inline_equation", + "content": "\\widetilde{\\theta}_{\\mathrm{target}}" + }, + { + "bbox": [ + 304, + 311, + 545, + 373 + ], + "type": "text", + "content": ", the corresponding filter atoms can be calculated as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 328, + 381, + 523, + 413 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 381, + 523, + 413 + ], + "spans": [ + { + "bbox": [ + 328, + 381, + 523, + 413 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Lambda} (\\widetilde {\\theta} _ {\\text {t a r g e t}}, \\widetilde {\\theta} _ {\\text {i n p u t}}) = \\boldsymbol {\\Lambda} _ {\\text {i n i t}} + \\int_ {\\widetilde {\\theta} _ {\\text {i n p u t}}} ^ {\\widetilde {\\theta} _ {\\text {t a r g e t}}} g (\\boldsymbol {\\Lambda} _ {\\text {i n i t}}, \\mathcal {W}) d \\widetilde {\\theta}.", + "image_path": "3c1183cdf9d486d1972ad3a3b64bb08e65de74bd553e35f0f84684bf630a2426.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 421, + 545, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 421, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 304, + 421, + 545, + 469 + ], + "type": "text", + "content": "According to Picard-Lindelöf existence theorem [15], for a given initial condition, the solution to the IVP is unique. Theoretically, we prove continuity in filter atoms induces controlled changes in image exposure." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 480, + 545, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 480, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 304, + 480, + 545, + 554 + ], + "type": "text", + "content": "Theorem 3.1 If " + }, + { + "bbox": [ + 304, + 480, + 545, + 554 + ], + "type": "inline_equation", + "content": "\\Lambda_{1}" + }, + { + "bbox": [ + 304, + 480, + 545, + 554 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 480, + 545, + 554 + ], + "type": "inline_equation", + "content": "\\Lambda_{2}" + }, + { + "bbox": [ + 304, + 480, + 545, + 554 + ], + "type": "text", + "content": " are filter atoms generated with neural ODE using integration intervals " + }, + { + "bbox": [ + 304, + 480, + 545, + 554 + ], + "type": "inline_equation", + "content": "(\\widetilde{\\theta}_{0},\\widetilde{\\theta}_{l})" + }, + { + "bbox": [ + 304, + 480, + 545, + 554 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 480, + 545, + 554 + ], + "type": "inline_equation", + "content": "(\\widetilde{\\theta}_0,\\widetilde{\\theta}_2)" + }, + { + "bbox": [ + 304, + 480, + 545, + 554 + ], + "type": "text", + "content": ", respectively such that " + }, + { + "bbox": [ + 304, + 480, + 545, + 554 + ], + "type": "inline_equation", + "content": "\\| \\Lambda_1 - \\Lambda_2\\| \\leq \\epsilon |\\widetilde{\\theta}_1 - \\widetilde{\\theta}_2|" + }, + { + "bbox": [ + 304, + 480, + 545, + 554 + ], + "type": "text", + "content": ", for " + }, + { + "bbox": [ + 304, + 480, + 545, + 554 + ], + "type": "inline_equation", + "content": "\\epsilon >0" + }, + { + "bbox": [ + 304, + 480, + 545, + 554 + ], + "type": "text", + "content": ". If " + }, + { + "bbox": [ + 304, + 480, + 545, + 554 + ], + "type": "inline_equation", + "content": "\\theta_{1}" + }, + { + "bbox": [ + 304, + 480, + 545, + 554 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 480, + 545, + 554 + ], + "type": "inline_equation", + "content": "\\theta_{2}" + }, + { + "bbox": [ + 304, + 480, + 545, + 554 + ], + "type": "text", + "content": " are the exposures at " + }, + { + "bbox": [ + 304, + 480, + 545, + 554 + ], + "type": "inline_equation", + "content": "\\widetilde{\\theta}_{1}" + }, + { + "bbox": [ + 304, + 480, + 545, + 554 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 480, + 545, + 554 + ], + "type": "inline_equation", + "content": "\\widetilde{\\theta}_{2}" + }, + { + "bbox": [ + 304, + 480, + 545, + 554 + ], + "type": "text", + "content": ", and activation function " + }, + { + "bbox": [ + 304, + 480, + 545, + 554 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 480, + 545, + 554 + ], + "type": "text", + "content": " is non-expansive, then change in exposure " + }, + { + "bbox": [ + 304, + 480, + 545, + 554 + ], + "type": "inline_equation", + "content": "\\Delta \\theta" + }, + { + "bbox": [ + 304, + 480, + 545, + 554 + ], + "type": "text", + "content": " is continuous in filter atoms." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 564, + 545, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 564, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 564, + 545, + 601 + ], + "type": "text", + "content": "Details of Theorem 3.1 and its proof can be found in the Appendix section A1. An illustration of our parameter efficient approach is shown in Figure 3." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 609, + 479, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 609, + 479, + 621 + ], + "spans": [ + { + "bbox": [ + 306, + 609, + 479, + 621 + ], + "type": "text", + "content": "3.4. Exposure Adaptive Colorization" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 627, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 627, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 627, + 545, + 662 + ], + "type": "text", + "content": "Our proposed exposure adaptive colorization consists of an exposure correction realized by the stated synthesis and exposure level specific colorizer." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 663, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 663, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 663, + 545, + 714 + ], + "type": "text", + "content": "Exposure correction. Given a colorization method specific to exposure " + }, + { + "bbox": [ + 304, + 663, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\widetilde{\\theta}_{\\mathrm{target}}" + }, + { + "bbox": [ + 304, + 663, + 545, + 714 + ], + "type": "text", + "content": " and a input sample " + }, + { + "bbox": [ + 304, + 663, + 545, + 714 + ], + "type": "inline_equation", + "content": "(\\mathbf{Y}_i,\\widetilde{\\theta}_{\\mathrm{input}})" + }, + { + "bbox": [ + 304, + 663, + 545, + 714 + ], + "type": "text", + "content": ", the proposed exposure correction generates image " + }, + { + "bbox": [ + 304, + 663, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Y}}_t" + }, + { + "bbox": [ + 304, + 663, + 545, + 714 + ], + "type": "text", + "content": " at the target exposure in two steps. With the operation " + }, + { + "bbox": [ + 304, + 663, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 304, + 663, + 545, + 714 + ], + "type": "text", + "content": " carried out" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25141" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 69, + 545, + 255 + ], + "blocks": [ + { + "bbox": [ + 62, + 69, + 545, + 255 + ], + "lines": [ + { + "bbox": [ + 62, + 69, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 62, + 69, + 545, + 255 + ], + "type": "image", + "image_path": "e285dcaa0552d4ee4cbacedca07a215de1f7c4629c2282772c00c7ae5a8cf969.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 263, + 547, + 365 + ], + "lines": [ + { + "bbox": [ + 46, + 263, + 547, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 263, + 547, + 365 + ], + "type": "text", + "content": "Figure 3. An illustration of our proposed method for exposure adaptive colorization: A binary image, " + }, + { + "bbox": [ + 46, + 263, + 547, + 365 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 46, + 263, + 547, + 365 + ], + "type": "text", + "content": ", which can be overexposed or underexposed, is input into the proposed exposure synthesis module. The colorization of binary image can be achieved using Single Image Colorization (SIC) or Burst Image Colorization (BIC). (i) SIC: Based on the input and target exposure levels, " + }, + { + "bbox": [ + 46, + 263, + 547, + 365 + ], + "type": "inline_equation", + "content": "\\widetilde{\\theta}_{\\mathrm{input}}" + }, + { + "bbox": [ + 46, + 263, + 547, + 365 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 263, + 547, + 365 + ], + "type": "inline_equation", + "content": "\\widetilde{\\theta}_{\\mathrm{target}}" + }, + { + "bbox": [ + 46, + 263, + 547, + 365 + ], + "type": "text", + "content": ", this module adjusts the weights of a exposure synthesis network " + }, + { + "bbox": [ + 46, + 263, + 547, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 46, + 263, + 547, + 365 + ], + "type": "text", + "content": ", which then generates an exposure-corrected image. Note that corrected image is not necessarily a binary image. Since the colorization module " + }, + { + "bbox": [ + 46, + 263, + 547, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 46, + 263, + 547, + 365 + ], + "type": "text", + "content": " is trained to colorize only image of specific exposure, the exposure synthesis module ensures corrected binary image has similar exposure to the one on which " + }, + { + "bbox": [ + 46, + 263, + 547, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 46, + 263, + 547, + 365 + ], + "type": "text", + "content": " is trained. (ii) BIC: For BIC we generate images with varying exposures as input to the burst image colorization network, " + }, + { + "bbox": [ + 46, + 263, + 547, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{burst}}" + }, + { + "bbox": [ + 46, + 263, + 547, + 365 + ], + "type": "text", + "content": ". The trained network is able to exploit the complementary information across multiple exposures with the help of Cross Non-Local Fusion blocks [44] to synthesize colors in regions of the image that is otherwise not possible by SIC approach." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 384, + 287, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 287, + 481 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 287, + 481 + ], + "type": "text", + "content": "by some generative network, first, the filter atoms of the network's convolutional filters are computed using the neural ODEs based on the " + }, + { + "bbox": [ + 46, + 384, + 287, + 481 + ], + "type": "inline_equation", + "content": "(\\widetilde{\\theta}_{\\mathrm{input}},\\widetilde{\\theta}_{\\mathrm{target}})" + }, + { + "bbox": [ + 46, + 384, + 287, + 481 + ], + "type": "text", + "content": " pair; next, the convolutional filter weights of the network are updated using the filter atoms and their corresponding coefficients, and the network generate the exposure corrected image " + }, + { + "bbox": [ + 46, + 384, + 287, + 481 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Y}}_t" + }, + { + "bbox": [ + 46, + 384, + 287, + 481 + ], + "type": "text", + "content": " based on " + }, + { + "bbox": [ + 46, + 384, + 287, + 481 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_i" + }, + { + "bbox": [ + 46, + 384, + 287, + 481 + ], + "type": "text", + "content": ". In the case of burst colorization, multiple images are generated corresponding to a set of target exposures." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 481, + 288, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 481, + 288, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 481, + 288, + 628 + ], + "type": "text", + "content": "Colorization. Both single-image-based colorization and burst-based colorization are considered in our work. The first approach involves a colorization network " + }, + { + "bbox": [ + 46, + 481, + 288, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{F}:\\hat{\\mathbf{Y}}_t\\to" + }, + { + "bbox": [ + 46, + 481, + 288, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_c" + }, + { + "bbox": [ + 46, + 481, + 288, + 628 + ], + "type": "text", + "content": " which takes a exposure corrected binary image as input and outputs corresponding RGB image " + }, + { + "bbox": [ + 46, + 481, + 288, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_c" + }, + { + "bbox": [ + 46, + 481, + 288, + 628 + ], + "type": "text", + "content": ". The network is trained exclusively at a fixed exposure level, and variations in the exposure of the input image is pre-adjusted by the exposure correction step. The second approach uses a colorization network " + }, + { + "bbox": [ + 46, + 481, + 288, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{burst}}:\\{\\hat{\\mathbf{Y}}_t\\}_{t = 1}^K\\to \\mathbf{X}_c" + }, + { + "bbox": [ + 46, + 481, + 288, + 628 + ], + "type": "text", + "content": " that takes in a burst of " + }, + { + "bbox": [ + 46, + 481, + 288, + 628 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 481, + 288, + 628 + ], + "type": "text", + "content": " binary images " + }, + { + "bbox": [ + 46, + 481, + 288, + 628 + ], + "type": "inline_equation", + "content": "\\{\\hat{\\mathbf{Y}}_t\\}_{t = 1}^K" + }, + { + "bbox": [ + 46, + 481, + 288, + 628 + ], + "type": "text", + "content": " corresponding to predefined varying exposure values " + }, + { + "bbox": [ + 46, + 481, + 288, + 628 + ], + "type": "inline_equation", + "content": "\\{\\hat{\\theta}_t\\}_{t = 1}^K" + }, + { + "bbox": [ + 46, + 481, + 288, + 628 + ], + "type": "text", + "content": " and outputs one RGB image " + }, + { + "bbox": [ + 46, + 481, + 288, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_c" + }, + { + "bbox": [ + 46, + 481, + 288, + 628 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 639, + 128, + 652 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 639, + 128, + 652 + ], + "spans": [ + { + "bbox": [ + 47, + 639, + 128, + 652 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 658, + 163, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 163, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 163, + 672 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "We evaluate the performance of both the exposure synthesis module standalone and the entire proposed approach. We conduct exposure burst recovery experiment to evalu" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 384, + 545, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 384, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 305, + 384, + 545, + 419 + ], + "type": "text", + "content": "ate the exposure synthesis and Single Image Colorization (SIC) and Burst Image Colorization (BIC) experiments to evaluate the final colorization results." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 422, + 546, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 546, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 546, + 602 + ], + "type": "text", + "content": "Implementation of Exposure Adaptive Colorization. We implement the neural ODEs using a 6-layer multilayer perceptron. The generative networks we use for exposure synthesis are variants of Pix2Pix [34] and CycleGAN [68], with which the models are named AtomODE-Pix2Pix and AtomODE-CycleGAN. For colorization, we train a Pix2Pix colorizer dedicated to SIC task and another to BIC. For BIC, the colorizer is additionally equipped with a Cross Non-Local Fusion (CNLF) [44] to constructively fuse the representations obtained from the burst. Note that the colorizers are trained directly on pairs of binary and color image data at fixed " + }, + { + "bbox": [ + 304, + 422, + 546, + 602 + ], + "type": "inline_equation", + "content": "\\widetilde{\\theta}" + }, + { + "bbox": [ + 304, + 422, + 546, + 602 + ], + "type": "text", + "content": ", independent of exposure synthesis. In our colorization experiment, the same trained colorizer is applied on all exposure correction methods for comparison. Details can be found in the Appendix section A2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 604, + 545, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 604, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 604, + 545, + 687 + ], + "type": "text", + "content": "Datasets and Synthetic Data Generation. We train and evaluate the proposed method as well as the baselines on synthetic data generated using images from two datasets: AFHQ [14] and CelebA-HQ [36]. AFHQ dataset contains images of three classes, cats, dogs, and wild animals, and CelebA-HQ contains images of male and female human faces." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": "To synthesize binary images, we use the image formation model described in Section 3.1 and apply exposure bracket" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25142" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 71, + 547, + 149 + ], + "blocks": [ + { + "bbox": [ + 48, + 71, + 547, + 149 + ], + "lines": [ + { + "bbox": [ + 48, + 71, + 547, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 547, + 149 + ], + "type": "table", + "html": "
Exposure Synthesis MethodAFHQ (512 x 512)CelebA-HQ (256 x 256)
CatDogWildMaleFemale
MSE (↓)/RL (↑)/FID (↓)MSE (↓)/RL (↑)/FID (↓)MSE (↓)/RL (↑)/FID (↓)MSE (↓)/RL (↑)/FID (↓)MSE (↓)/RL (↑)/FID (↓)
Pix2Pix-DNI3.75 / 0.9816 / 61.895.91 / 0.9844 / 95.438.13 / 0.9780 / 39.293.74 / 0.9563 / 56.134.78 / 0.9681 / 39.81
CycleGAN-DNI9.61 / 0.9816 / 207.8216.79 / 0.9824 / 220.909.62 / 0.9820 / 336.389.31 / 0.9824 / 94.7814.26 / 0.9810 / 250.61
SAVI2I41.15 / 46.27 / 178.2946.27 / 0.9804 / 217.9838.87 / 0.9019 / 139.4767.26 / 0.9563 / 111.0569.21 / 0.9698 / 261.45
DLOW2.44 / 0.9654 / 92.302.32 / 0.9489 / 173.652.75 / 0.9632 / 179.231.87 / 0.9655 / 62.571.37 / 0.9798 / 63.31
AtomODE-Pix2Pix (ours)2.29 / 0.9997 / 57.883.17 / 0.9994 / 140.232.15 / 0.9998 / 79.762.15 / 0.9995 / 53.443.20 / 0.9995 / 40.08
AtomODE-CycleGAN (ours)1.15 / 0.9948 / 60.541.18 / 0.9971 / 87.431.23 / 0.9979 / 35.751.58 / 0.9986 / 77.821.87 / 0.9994 / 62.59
", + "image_path": "39e4ee4f0f7c272e37bb46f576a893cb877a5cb36e3b831dcaa1866cdd58fb0c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 154, + 547, + 198 + ], + "lines": [ + { + "bbox": [ + 46, + 154, + 547, + 198 + ], + "spans": [ + { + "bbox": [ + 46, + 154, + 547, + 198 + ], + "type": "text", + "content": "Table 2. Colorization results on AFHQ and CelebA-HQ datasets. The input image for all these experiments is an overexposed image at index " + }, + { + "bbox": [ + 46, + 154, + 547, + 198 + ], + "type": "inline_equation", + "content": "\\tau = 0" + }, + { + "bbox": [ + 46, + 154, + 547, + 198 + ], + "type": "text", + "content": " of exposure burst. All the colorizers are trained using an image corresponding to " + }, + { + "bbox": [ + 46, + 154, + 547, + 198 + ], + "type": "inline_equation", + "content": "\\tau = 10" + }, + { + "bbox": [ + 46, + 154, + 547, + 198 + ], + "type": "text", + "content": " in the exposure burst. At test time we correct the overexposed input to correspond to the one used during training. In the first column of the table, we note the combination of the exposure correction method used and colorizer is same across all the methods." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 50, + 206, + 287, + 524 + ], + "blocks": [ + { + "bbox": [ + 50, + 206, + 287, + 524 + ], + "lines": [ + { + "bbox": [ + 50, + 206, + 287, + 524 + ], + "spans": [ + { + "bbox": [ + 50, + 206, + 287, + 524 + ], + "type": "image", + "image_path": "435800515c977b9ae215fe545e14af136908b2fb5937ebbdd23b1f132a724d40.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 534, + 287, + 577 + ], + "lines": [ + { + "bbox": [ + 46, + 534, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 287, + 577 + ], + "type": "text", + "content": "Figure 4. Qualitative results for exposure burst recovery experiments for AFHQ dataset. We compare our methods, AtomODE-Pix2Pix and AtomODE-CycleGAN, with DLOW, SAVI2I and DNI." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 579, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 579, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 579, + 287, + 640 + ], + "type": "text", + "content": "ing to generate various exposure. Details of data synthesis can be found in the Appendix section A2. Our compiled test datasets consist of 15-frame bursts of exposure varying from overexposed to underexposed. We index the exposure of these images in a burst by " + }, + { + "bbox": [ + 46, + 579, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 579, + 287, + 640 + ], + "type": "text", + "content": ", a discretized label of " + }, + { + "bbox": [ + 46, + 579, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\widetilde{\\theta}" + }, + { + "bbox": [ + 46, + 579, + 287, + 640 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": "Baselines. In the exposure burst recovery experiment, we compare our approach against DLOW [25] and SAVI2I [52]. We further consider DNI [62] for its capability of interpolating network parameters to generate continuous variations of outputs. We prepare baselines Pix2Pix-DNI and CycleGAN-DNI by coupling DNI with Pix2Pix and Cycle" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 209, + 545, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 209, + 545, + 268 + ], + "spans": [ + { + "bbox": [ + 304, + 209, + 545, + 268 + ], + "type": "text", + "content": "GAN exposure generators. In colorization experiments, we apply the independently trained Pix2Pix colorizers to the exposure correction results generated by all methods. For more details on baseline training and hyperparameters, we refer readers to the Appendix section A2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 269, + 546, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 269, + 546, + 364 + ], + "spans": [ + { + "bbox": [ + 304, + 269, + 546, + 364 + ], + "type": "text", + "content": "Metrics. In the burst recovery experiment, we choose to use Mean Squared Error (MSE) and Relative Linearity (RL) [43] to distinguish a smoothly varying exposure burst with consistent appearance changes. A high RL score indicates a better transition in exposure among consecutive samples. In the colorization experiments, we assess the quality of exposure correction by computing the Fréchet Inception Distance (FID) between the colorized image and ground truth." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 383, + 485, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 383, + 485, + 396 + ], + "spans": [ + { + "bbox": [ + 306, + 383, + 485, + 396 + ], + "type": "text", + "content": "4.2. Exposure Burst Recovery Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 401, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 401, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 545, + 556 + ], + "type": "text", + "content": "Figure 4 presents the qualitative comparison of all methods on AFHQ dataset. We find our methods, AtomODE Pix2Pix and AtomODE CycleGAN are able to generate consistent set of smoothly varying exposure images than baseline methods and match ground truth images generated from exposure bracketing. We also observe some baselines struggle with generation of overexposed frames. All methods for comparison are further assessed on both datasets using MSE and RL, with the quantitative results shown in Table 2. Results suggest that the proposed AtomODE provides consistently better performance in all testing conditions across multiple datasets. Additional qualitative results can be found in Appendix section A4." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 563, + 537, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 563, + 537, + 576 + ], + "spans": [ + { + "bbox": [ + 305, + 563, + 537, + 576 + ], + "type": "text", + "content": "4.3. Single and Burst Image Colorization Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 582, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 712 + ], + "type": "text", + "content": "The SIC colorization results of all methods on AFHQ dataset are shown in Figure 5, where the exposure of the test samples is different from the that of the colorizer's training samples. Corrupted colored images are generated by most of the baselines, while the proposed method produces visually appealing outputs. Figure 7 shows the SIC colorization results on CelebA-HQ dataset. Additionally, we also present the qualitative results for BIC in Figure 6 where " + }, + { + "bbox": [ + 304, + 582, + 545, + 712 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 582, + 545, + 712 + ], + "type": "text", + "content": " denotes the number of binary frames input to the colorizer. The quantitative results are shown in Table 2, assessed with the FID metric." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25143" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 69, + 545, + 526 + ], + "blocks": [ + { + "bbox": [ + 50, + 69, + 545, + 526 + ], + "lines": [ + { + "bbox": [ + 50, + 69, + 545, + 526 + ], + "spans": [ + { + "bbox": [ + 50, + 69, + 545, + 526 + ], + "type": "image", + "image_path": "548330aef06c66cc200aa89566c82076cec156c5839c7327bbb55a0d36983a12.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 534, + 546, + 569 + ], + "lines": [ + { + "bbox": [ + 46, + 534, + 546, + 569 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 546, + 569 + ], + "type": "text", + "content": "Figure 5. Colorized results for AFHQ dataset using various combinations of exposure correction methods and colorizer. The first column shows the overexposed binary input image for exposure correction. The results of our methods AtomODE Pix2Pix and AtomODE CycleGAN are shown in columns six and seven." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 574, + 178, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 178, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 178, + 588 + ], + "type": "text", + "content": "4.4. Real Data Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "As our approach is trained on synthetic data, it is essential to evaluate its performance on real-world data. To this end, we conduct experiments using both a CMOS camera and a prototype QIS camera. We refer the readers to Appendix for further details about the experiment setup. We print 100 images from the training set and 25 from the test set, and we capture pictures of them using both cameras, collecting 125 pairs of grayscale and RGB images. We fine-tune the colorization model with the 100 real data samples, and we evaluate our proposed approach on the set of 25 test data, in" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 574, + 547, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 574, + 547, + 708 + ], + "spans": [ + { + "bbox": [ + 304, + 574, + 547, + 708 + ], + "type": "text", + "content": "cluding both binarized CMOS data and QIS data. In Figure 8, we present qualitative results from the evaluation on the real data obtained by the CMOS camera. Despite that the proposed method was not trained with vast real sensor data, it is able to produce visually pleasing colorized images. We observe that fine-tuning the model using a small training dataset significantly improves the colorization quality. We further present results from the evaluation on images captured by the QIS camera in Figure 9. These results demonstrate the potential of the proposed method for real-world single-photon imaging applications." + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25144" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 64, + 75, + 286, + 209 + ], + "blocks": [ + { + "bbox": [ + 52, + 87, + 61, + 204 + ], + "lines": [ + { + "bbox": [ + 52, + 87, + 61, + 204 + ], + "spans": [ + { + "bbox": [ + 52, + 87, + 61, + 204 + ], + "type": "text", + "content": "Groundtruth " + }, + { + "bbox": [ + 52, + 87, + 61, + 204 + ], + "type": "inline_equation", + "content": "K = 5" + }, + { + "bbox": [ + 52, + 87, + 61, + 204 + ], + "type": "inline_equation", + "content": "K = 1" + } + ] + } + ], + "index": 0, + "angle": 270, + "type": "image_caption" + }, + { + "bbox": [ + 64, + 75, + 286, + 209 + ], + "lines": [ + { + "bbox": [ + 64, + 75, + 286, + 209 + ], + "spans": [ + { + "bbox": [ + 64, + 75, + 286, + 209 + ], + "type": "image", + "image_path": "4c800f0b820fd94bab360555cc73142aedd6a6e07fa8f77a825db0c3fb3ad13a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 50, + 301, + 283, + 506 + ], + "blocks": [ + { + "bbox": [ + 47, + 220, + 287, + 285 + ], + "lines": [ + { + "bbox": [ + 47, + 220, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 47, + 220, + 287, + 285 + ], + "type": "text", + "content": "Figure 6. An illustration of burst image colorization for AFHQ dataset using multiple binary images of varying exposures as input. The colorizer is trained using images of exposure " + }, + { + "bbox": [ + 47, + 220, + 287, + 285 + ], + "type": "inline_equation", + "content": "\\tau = \\{0,4,8,10,14\\}" + }, + { + "bbox": [ + 47, + 220, + 287, + 285 + ], + "type": "text", + "content": ". An overexposed input image is used to synthesize multiple images of varying exposures and passed to the colorizer." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 301, + 283, + 506 + ], + "lines": [ + { + "bbox": [ + 50, + 301, + 283, + 506 + ], + "spans": [ + { + "bbox": [ + 50, + 301, + 283, + 506 + ], + "type": "image", + "image_path": "678f953f4e3d069d606f7f18b9f3f692d3fe60ed38c26f9723a3671b7762c12b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 516, + 287, + 549 + ], + "lines": [ + { + "bbox": [ + 46, + 516, + 287, + 549 + ], + "spans": [ + { + "bbox": [ + 46, + 516, + 287, + 549 + ], + "type": "text", + "content": "Figure 7. Colorized results for CelebA-HQ dataset using various combinations of exposure correction methods and colorizers. The results of our method are shown in columns four and five." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 562, + 230, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 562, + 230, + 575 + ], + "spans": [ + { + "bbox": [ + 47, + 562, + 230, + 575 + ], + "type": "text", + "content": "4.5. Benefits for Practical Applications" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 582, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 712 + ], + "type": "text", + "content": "Our methodology is specifically designed with modularity and parameter efficiency in mind, features crucial for real-world industry applications. The Pix2PixDNI method, the second-best performing approach, necessitates training multiple networks and either interpolating model weights on-the-fly during inference or precomputing and storing them to generate images with varied exposures. Conversely, our method circumvents the additional storage requirements and the need to train multiple model parameters by incorporating the exposure correction into the neural ODE framework." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 72, + 541, + 255 + ], + "blocks": [ + { + "bbox": [ + 307, + 72, + 541, + 255 + ], + "lines": [ + { + "bbox": [ + 307, + 72, + 541, + 255 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 541, + 255 + ], + "type": "image", + "image_path": "e76a678abea8de61fab3bce9e4dcc81f118d15b6886ffcd9854fe69d984c9e0f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 264, + 546, + 309 + ], + "lines": [ + { + "bbox": [ + 305, + 264, + 546, + 309 + ], + "spans": [ + { + "bbox": [ + 305, + 264, + 546, + 309 + ], + "type": "text", + "content": "Figure 8. Results on real-world images captured from a CMOS camera (images shown along the row with a red line are colorized images, and along the green line are the ground-truth images captured using an RGB CMOS camera)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 323, + 543, + 435 + ], + "blocks": [ + { + "bbox": [ + 307, + 323, + 543, + 435 + ], + "lines": [ + { + "bbox": [ + 307, + 323, + 543, + 435 + ], + "spans": [ + { + "bbox": [ + 307, + 323, + 543, + 435 + ], + "type": "image", + "image_path": "7f5b1a31e523c5b163bc17b15a131fb6ada6f0d10867c42e87a120f5e9a4a8c3.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 446, + 545, + 468 + ], + "lines": [ + { + "bbox": [ + 306, + 446, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 306, + 446, + 545, + 468 + ], + "type": "text", + "content": "Figure 9. Colorized output using input from prototype QIS camera." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 480, + 378, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 480, + 378, + 493 + ], + "spans": [ + { + "bbox": [ + 306, + 480, + 378, + 493 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 501, + 546, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 501, + 546, + 657 + ], + "spans": [ + { + "bbox": [ + 304, + 501, + 546, + 657 + ], + "type": "text", + "content": "In this work, we proposed a generative exposure correction approach that can synthesize a spectrum of exposure given a single 1-bit image. We modeled the binary signal variation with respect to exposure change as a continuous function by adapting a small set of parameters of a network, known as filter atoms, to desired exposure using neural ODE. Theoretically and empirically, we demonstrated successful recovery of bursts of 1-bit images from a single input image, and we verified its application on single-image-based and burst-image-based 1-bit image colorization. Experiments on real CMOS and QIS data further indicated the potentials of the proposed approach on real-world single-photon imaging applications." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "25145" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Samuel Burri, Yuki Maruyama, Xavier Michalet, Francesco Regazzoni, Claudio Bruschini, and Edoardo Charbon. Architecture and applications of a high resolution gated SPAD image sensor. Optics Express, 22(14):17573-17589, 2014. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 288, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 288, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 288, + 179 + ], + "type": "text", + "content": "[2] Stanley H. Chan. What does a one-bit Quanta Image Sensor offer? IEEE Transactions on Computational Imaging, 8: 770-783, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 181, + 288, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 181, + 288, + 225 + ], + "spans": [ + { + "bbox": [ + 53, + 181, + 288, + 225 + ], + "type": "text", + "content": "[3] Stanley H. Chan and Yue M. Lu. Efficient image reconstruction for gigapixel quantum image sensors. In IEEE Global Conference on Signal and Information Processing, pages 312-316, 2014. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 227, + 288, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 227, + 288, + 259 + ], + "spans": [ + { + "bbox": [ + 53, + 227, + 288, + 259 + ], + "type": "text", + "content": "[4] Stanley H. Chan, Omar A. Elgendy, and Xiran Wang. Images from bits: Non-iterative image reconstruction for Quanta Image Sensors. MDPI Sensors, 16(11), 2016. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 260, + 288, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 260, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 53, + 260, + 288, + 304 + ], + "type": "text", + "content": "[5] Stanley H. Chan, Xiran Wang, and Omar A. Elgendy. Plug-and-Play ADMM for image restoration: Fixed-point convergence and applications. IEEE Transactions on Computational Imaging, 3:84–98, 2016. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 306, + 288, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 306, + 288, + 338 + ], + "spans": [ + { + "bbox": [ + 53, + 306, + 288, + 338 + ], + "type": "text", + "content": "[6] Ricky T. Q. Chen, Yulia Rubanova, Jesse Bettencourt, and David K. Duvenaud. Neural Ordinary Differential Equations. In NeurIPS, 2018. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 340, + 288, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 340, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 53, + 340, + 288, + 361 + ], + "type": "text", + "content": "[7] Wei Chen, Zichen Miao, and Qiang Qiu. Inner product-based neural network similarity. In NeurIPS, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 363, + 288, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 363, + 288, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 363, + 288, + 407 + ], + "type": "text", + "content": "[8] Wei Chen, Zichen Miao, and Qiang Qiu. Large convolutional model tuning via filter subspace. Available online: https://arxiv.org/abs/2403.00269, 2024. Accessed 03/06/2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 408, + 288, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 408, + 288, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 408, + 288, + 441 + ], + "type": "text", + "content": "[9] Ying-Cong Chen, Xiaogang Xu, Zhuotao Tian, and Jiaya Jia. Homomorphic latent space interpolation for unpaired Image-To-Image translation. In CVPR, pages 2403-2411, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 442, + 287, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 442, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 48, + 442, + 287, + 485 + ], + "type": "text", + "content": "[10] Xiuyuan Cheng, Qiang Qiu, Robert Calderbank, and Guillermo Sapiro. RotDCF: Decomposition of convolutional filters for rotation-equivariant deep networks. In ICLR, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 487, + 287, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 487, + 287, + 530 + ], + "spans": [ + { + "bbox": [ + 48, + 487, + 287, + 530 + ], + "type": "text", + "content": "[11] Yiheng Chi, Abhiram Gnanasambandam, Vladlen Koltun, and Stanley H. Chan. Dynamic low-light imaging with Quanta Image Sensors. In ECCV, pages 122-138, 2020. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 533, + 287, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 287, + 566 + ], + "type": "text", + "content": "[12] Joon Hee Choi, Omar A. Elgendy, and Stanley H. Chan. Image reconstruction for Quanta Image Sensors using deep neural networks. In IEEE ICASSP, pages 6543-6547, 2018. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 567, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 610 + ], + "type": "text", + "content": "[13] Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. StarGAN: Unified Generative Adversarial Networks for multi-domain Image-to-Image translation. In CVPR, pages 8789–8797, 2018. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 612, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 612, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 612, + 287, + 645 + ], + "type": "text", + "content": "[14] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. StarGAN v2: Diverse image synthesis for multiple domains. In CVPR, pages 8185-8194, 2020. 2, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 647, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 668 + ], + "type": "text", + "content": "[15] Earl A. Coddington and Norman Levinson. Theory of Ordinary Differential Equations. McGraw-Hill, 1955. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 670, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 714 + ], + "type": "text", + "content": "[16] Neale A. W. Dutton, Istvan Gyongy, Luca Parmesan, Salvatore Gnecchi, Neil Calder, Bruce R. Rae, Sara Pellegrini, Lindsay A. Grant, and Robert K. Henderson. A SPAD-based QVGA image sensor for single-photon counting and quanta" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "imaging. IEEE Transactions on Electron Devices, 63(1): 189-196, 2016. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "type": "text", + "content": "[17] Neale A. W. Dutton, Istvan Gyongy, Luca Parmesan, and Robert K. Henderson. Single photon counting performance and noise analysis of CMOS SPAD-based image sensors. MDPI Sensors, 16(7), 2016. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 141, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 194 + ], + "type": "text", + "content": "[18] Omar A. Elgendy, Abhiram Gnanasambandam, Stanley H. Chan, and Jiaju Ma. Low-light demosaicking and denoising for small pixels using learned frequency selection. IEEE Transactions on Computational Imaging, 7:137-150, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 197, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 197, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 197, + 545, + 239 + ], + "type": "text", + "content": "[19] Eric R. Fossum. Some thoughts on future digital still cameras. In Image Sensors and Signal Processing for Digital Still Cameras, pages 305-314. Taylor and Francis Group, 2005. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 241, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 241, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 241, + 545, + 274 + ], + "type": "text", + "content": "[20] Eric R. Fossum. Modeling the performance of single-bit and multi-bit Quanta Image Sensors. IEEE Journal of the Electron Devices Society, 1(9):166-174, 2013. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 276, + 545, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 545, + 317 + ], + "type": "text", + "content": "[21] Eric R. Fossum. Analog read noise and quantizer threshold estimation from Quanta Image Sensor bit density. IEEE Journal of the Electron Devices Society, 10:269-274, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 319, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 319, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 308, + 319, + 545, + 353 + ], + "type": "text", + "content": "[22] Abhiram Gnanasambandam and Stanley H. Chan. Image classification in the dark using Quanta Image Sensors. In ECCV, pages 484-501, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 354, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 354, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 308, + 354, + 545, + 398 + ], + "type": "text", + "content": "[23] Abhiram Gnanasambandam and Stanley H. Chan. HDR imaging with Quanta Image Sensors: Theoretical limits and optimal reconstruction. IEEE Transactions on Computational Imaging, 6:1571-1585, 2020. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 399, + 545, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 441 + ], + "type": "text", + "content": "[24] Abhiram Gnanasambandam and Stanley H. Chan. Exposure-referred signal-to-noise ratio for digital image sensors. IEEE Transactions on Computational Imaging, 8:561-575, 2022. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 443, + 545, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 443, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 308, + 443, + 545, + 476 + ], + "type": "text", + "content": "[25] Rui Gong, Wen Li, Yuhua Chen, and Luc Van Gool. DLOW: Domain flow for adaptation and generalization. In CVPR, pages 2472-2481, 2019. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 478, + 545, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 478, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 308, + 478, + 545, + 520 + ], + "type": "text", + "content": "[26] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative Adversarial Nets. In NeurIPS, 2014. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 522, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 522, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 522, + 545, + 555 + ], + "type": "text", + "content": "[27] Anant Gupta, Atul Ingle, and Mohit Gupta. Asynchronous single-photon 3D imaging. In ICCV, pages 7908-7917, 2019. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 556, + 545, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 556, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 308, + 556, + 545, + 588 + ], + "type": "text", + "content": "[28] Anant Gupta, Atul Ingle, Andreas Velten, and Mohit Gupta. Photon-flooded single-photon 3D cameras. In CVPR, pages 6763-6772, 2019." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 590, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 545, + 623 + ], + "type": "text", + "content": "[29] Felipe Gutierrez-Barragan, Atul Ingle, Trevor Seets, Mohit Gupta, and Andreas Velten. Compressive single-photon 3D cameras. In CVPR, pages 17833-17843, 2022. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 624, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 656 + ], + "type": "text", + "content": "[30] Istvan Gyongy, Neale AW Dutton, and Robert K Henderson. Single-photon tracking for high-speed vision. MDPI Sensors, 18:323, 2018. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 658, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 679 + ], + "type": "text", + "content": "[31] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "text", + "content": "[32] Atul Ingle, Andreas Velten, and Mohit Gupta. High flux passive imaging with single-photon sensors. In CVPR, pages 6753-6762, 2021. 2" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "25146" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[33] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-Image translation with conditional adversarial networks. In CVPR, pages 5967–5976, 2017. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 287, + 140 + ], + "type": "text", + "content": "[34] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A. Efros. Image-to-Image translation with conditional adversarial networks. In CVPR, pages 5967–5976, 2017. 3, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 141, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 141, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 49, + 141, + 287, + 184 + ], + "type": "text", + "content": "[35] Kiyotaka Iwabuchi, Yusuke Kameda, and Takayuki Hamamoto. Image quality improvements based on motion-based deblurring for single-photon imaging. IEEE Access, 9:30080-30094, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 186, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 186, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 49, + 186, + 287, + 217 + ], + "type": "text", + "content": "[36] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In ICLR, 2018. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 220, + 287, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 220, + 287, + 252 + ], + "spans": [ + { + "bbox": [ + 49, + 220, + 287, + 252 + ], + "type": "text", + "content": "[37] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for Generative Adversarial Networks. IEEE TPAMI, 43(12):4217-4228, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 254, + 286, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 254, + 286, + 274 + ], + "spans": [ + { + "bbox": [ + 49, + 254, + 286, + 274 + ], + "type": "text", + "content": "[38] Manoj Kumar, Dirk Weissenborn, and Nal Kalchbrenner. Colorization transformer. In ICLR, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 277, + 286, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 277, + 286, + 319 + ], + "spans": [ + { + "bbox": [ + 49, + 277, + 286, + 319 + ], + "type": "text", + "content": "[39] Vadim Lebedev, Yaroslav Ganin, Maksim Rakhuba, Ivan Oseledets, and Victor Lempitsky. Speeding-up convolutional neural networks using fine-tuned CP-decomposition. In ICLR, 2015. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 322, + 286, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 322, + 286, + 364 + ], + "spans": [ + { + "bbox": [ + 49, + 322, + 286, + 364 + ], + "type": "text", + "content": "[40] Yu-Jing Lin, Po-Wei Wu, Che-Han Chang, Edward Chang, and Shih-Wei Liao. RelGAN: Multi-domain Image-to-Image translation via relative attributes. In ICCV, pages 5913–5921, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 366, + 287, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 366, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 49, + 366, + 287, + 409 + ], + "type": "text", + "content": "[41] Wallace Lira, Johannes Merz, Daniel Ritchie, Daniel Cohen-Or, and Hao Zhang. GANHopper: Multi-hop GAN for unsupervised Image-to-Image translation. In ECCV, pages 363–379, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 411, + 287, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 411, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 49, + 411, + 287, + 465 + ], + "type": "text", + "content": "[42] Yahui Liu, Enver Sangineto, Yajing Chen, Linchao Bao, Haoxian Zhang, Nicu Sebe, Bruno Lepri, Wei Wang, and Marco De Nadai. Smoothing the disentangled latent style space for unsupervised Image-to-Image translation. In CVPR, pages 10780–10789, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 468, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 468, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 49, + 468, + 287, + 498 + ], + "type": "text", + "content": "[43] Kun Lu, Rongpeng Li, and Honggang Zhang. Contrastive monotonic pixel-level modulation. In ECCV, pages 493-510, 2022. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 501, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 501, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 49, + 501, + 287, + 544 + ], + "type": "text", + "content": "[44] Ziwei Luo, Lei Yu, Xuan Mo, Youwei Li, Lanpeng Jia, Haoqiang Fan, Jian Sun, and Shuaicheng Liu. EBSR: Feature Enhanced Burst Super-Resolution with deformable alignment. In CVPR, pages 471–478, 2021. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 546, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 546, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 49, + 546, + 287, + 578 + ], + "type": "text", + "content": "[45] Jiaju Ma and Eric Fossum. Quanta Image Sensor jot with sub 0.3e- r.m.s. read noise and photon counting capability. IEEE Electron Device Letters, 36(9):926-928, 2015. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 580, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 580, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 49, + 580, + 287, + 612 + ], + "type": "text", + "content": "[46] Jiaju Ma and Eric R. Fossum. A Pump-Gate Jot device with high conversion gain for a Quanta Image Sensor. IEEE Journal of the Electron Devices Society, 3:73-77, 2015." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 614, + 287, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 614, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 49, + 614, + 287, + 666 + ], + "type": "text", + "content": "[47] Jiaju Ma, Dakota Starkey, Arun Rao, Kofi Odame, and Eric R. Fossum. Characterization of Quanta Image Sensor pump-gate jots with deep sub-electron read noise. IEEE Journal of the Electron Devices Society, 3(6):472-480, 2015. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "type": "text", + "content": "[48] Jiaju Ma, Saleh Masoodian, Dakota A Starkey, and Eric R. Fossum. Photon-number-resolving megapixel image sensor at room temperature without avalanche gain. Optica, 4(12): 1474-1481, 2017. 1" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[49] Jiaju Ma, Dexue Zhang, Omar Elgendy, and Saleh Masoodian. A 0.19e-rms read noise 16.7mpixel stacked Quanta Image Sensor with " + }, + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "inline_equation", + "content": "1.1\\mu \\mathrm{m}" + }, + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "-pitch backside illuminated pixels. IEEE Electron Device Letters, 42(6):891-894, 2021. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 119, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 119, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 119, + 545, + 162 + ], + "type": "text", + "content": "[50] Jiaju Ma, Stanley H. Chan, and Eric R. Fossum. Review of Quanta Image Sensors for ultralow-light imaging. IEEE Transactions on Electron Devices, 69(6):2824-2839, 2022. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 164, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 164, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 308, + 164, + 545, + 195 + ], + "type": "text", + "content": "[51] Sizhuo Ma, Paul Mos, Edoardo Charbon, and Mohit Gupta. Burst vision using single-photon cameras. In WACV, pages 5364-5374, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 198, + 545, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 198, + 545, + 242 + ], + "spans": [ + { + "bbox": [ + 308, + 198, + 545, + 242 + ], + "type": "text", + "content": "[52] Qi Mao, Hsin-Ying Lee, Hung-Yu Tseng, Jia-Bin Huang, Siwei Ma, and Ming-Hsuan Yang. Continuous and diverse Image-to-Image translation via signed attribute vectors. IJCV, 130(2):517-549, 2020. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 244, + 545, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 244, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 308, + 244, + 545, + 276 + ], + "type": "text", + "content": "[53] Zichen Miao, Ze Wang, Xiuyuan Cheng, and Qiang Qiu. Spatiotemporal joint filter decomposition in 3D convolutional neural networks. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 278, + 545, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 278, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 308, + 278, + 545, + 300 + ], + "type": "text", + "content": "[54] Zichen Miao, Ze Wang, Wei Chen, and Qiang Qiu. Continual learning with filter atom swapping. In ICLR, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 303, + 545, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 303, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 308, + 303, + 545, + 334 + ], + "type": "text", + "content": "[55] Fabio Pizzati, Pietro Cerri, and Raoul de Charette. CoMoGAN: Continuous model-guided Image-to-Image translation. In CVPR, pages 14283–14293, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 337, + 545, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 337, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 308, + 337, + 545, + 369 + ], + "type": "text", + "content": "[56] Qiang Qiu, Xiuyuan Cheng, Robert Calderbank, and Guillermo Sapiro. DCFNet: Deep neural network with decomposed convolutional filters. In ICML, 2018. 2, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 372, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 372, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 372, + 545, + 403 + ], + "type": "text", + "content": "[57] Trevor Seets, Atul Ingle, Martin Laurenzis, and Andreas Velten. Motion adaptive deblurring with single-photon cameras. In WACV, pages 1944–1953, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 406, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 406, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 406, + 545, + 437 + ], + "type": "text", + "content": "[58] Jheng-Wei Su, Hung-Kuo Chu, and Jia-Bin Huang. Instance-aware image colorization. In CVPR, pages 7965–7974, 2020. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 440, + 545, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 440, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 308, + 440, + 545, + 472 + ], + "type": "text", + "content": "[59] Varun Sundar, Sizhuo Ma, Aswin Sankaranarayanan, and Mohit Gupta. Single-photon structured light. In CVPR, pages 17844-17854, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 475, + 545, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 475, + 545, + 507 + ], + "spans": [ + { + "bbox": [ + 308, + 475, + 545, + 507 + ], + "type": "text", + "content": "[60] Patricia Vitoria, Lara Raad, and Coloma Ballester. ChromaGAN: Adversarial picture colorization with semantic class distribution. In WACV, pages 2434-2443, 2020. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 510, + 545, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 510, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 308, + 510, + 545, + 553 + ], + "type": "text", + "content": "[61] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional GANs. In CVPR, pages 8798-8807, 2018. 2, 4" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 555, + 545, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 555, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 308, + 555, + 545, + 587 + ], + "type": "text", + "content": "[62] X. Wang, K. Yu, C. Dong, X. Tang, and C. Loy. Deep network interpolation for continuous imagery effect transition. In CVPR, pages 1692-1701, 2019. 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 590, + 545, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 545, + 643 + ], + "type": "text", + "content": "[63] Ze Wang. Efficient Adaptation of Deep Vision Models. PhD thesis, Purdue University, 2023. Available online: https://hammer.purdue.edu/articles/thesis/Efficient_Adaptation_of_Deep_Vision_CModels/22696129.2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 646, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 646, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 646, + 545, + 678 + ], + "type": "text", + "content": "[64] Ze Wang, Xiuyuan Cheng, Guillermo Sapiro, and Qiang Qiu. Stochastic conditional generative networks with basis decomposition. In ICLR, 2020. 2, 4" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "text", + "content": "[65] Ze Wang, Seunghyun Hwang, Zichen Miao, and Qiang Qiu. Image generation using continuous filter atoms. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25147" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 239 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "text", + "content": "[66] Yanze Wu, Xintao Wang, Yu Li, Honglun Zhang, Xun Zhao, and Ying Shan. Towards vivid and diverse image colorization with generative color prior. In ICCV, pages 14357-14366, 2021. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "type": "text", + "content": "[67] Zhaoyang Yin, Jiaju Ma, Saleh Masoodian, and Eric R Fossum. Threshold uniformity improvement in 1b Quanta Image Sensor readout circuit. MDPI Sensors, 22(7):2578, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 194 + ], + "type": "text", + "content": "[68] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In ICCV, pages 2242-2251, 2017. 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 197, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 287, + 239 + ], + "type": "text", + "content": "[69] Wei Zhu, Qiang Qiu, Robert Calderbank, Guillermo Sapiro, and Xiuyuan Cheng. Scaling-translation-equivariant networks with decomposed convolutional filters. Journal of Machine Learning Research, 23(68):1-45, 2022. 2" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25148" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Generative Region-Language Pretraining for Open-Ended Object Detection/bb1e00ce-acc5-43f7-b817-49b140b640c0_content_list.json b/2024/Generative Region-Language Pretraining for Open-Ended Object Detection/bb1e00ce-acc5-43f7-b817-49b140b640c0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..aa656bb226b001d190f236cede3b805035765fa6 --- /dev/null +++ b/2024/Generative Region-Language Pretraining for Open-Ended Object Detection/bb1e00ce-acc5-43f7-b817-49b140b640c0_content_list.json @@ -0,0 +1,1494 @@ +[ + { + "type": "text", + "text": "Generative Region-Language Pretraining for Open-Ended Object Detection", + "text_level": 1, + "bbox": [ + 102, + 130, + 867, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chuang Lin1", + "bbox": [ + 163, + 178, + 274, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yi Jiang*", + "bbox": [ + 310, + 179, + 398, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lizhen $\\mathbf{Q}\\mathbf{u}^{1}$", + "bbox": [ + 428, + 180, + 524, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zehuan Yuan", + "bbox": [ + 563, + 180, + 684, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jianfei Cai*", + "bbox": [ + 722, + 180, + 828, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Monash University", + "bbox": [ + 320, + 198, + 485, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 ByteDance Inc.", + "bbox": [ + 524, + 198, + 661, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 251, + 312, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent research, significant attention has been devoted to the open-vocabulary object detection task, aiming to generalize beyond the limited number of classes labeled during training and detect objects described by arbitrary category names at inference. Compared with conventional object detection, open vocabulary object detection largely extends the object detection categories. However, it relies on calculating the similarity between image regions and a set of arbitrary category names with a pretrained vision-and-language model. This implies that, despite its open-set nature, the task still needs the predefined object categories during the inference stage. This raises the question: What if we do not have exact knowledge of object categories during inference? In this paper, we call such a new setting as generative open-ended object detection, which is a more general and practical problem. To address it, we formulate object detection as a generative problem and propose a simple framework named GenerateU, which can detect dense objects and generate their names in a free-form way. Particularly, we employ Deformable DETR as a region proposal generator with a language model translating visual regions to object names. To assess the free-form object detection task, we introduce an evaluation method designed to quantitatively measure the performance of generative outcomes. Extensive experiments demonstrate strong zero-shot detection performance of our GenerateU. For example, on the LVIS dataset, our GenerateU achieves comparable results to the open-vocabulary object detection method GLIP, even though the category names are not seen by GenerateU during inference. Code is available at: https://github.com/FoundationVision/GenerateU.", + "bbox": [ + 76, + 282, + 473, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 777, + 207, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Object detection, a crucial task in computer vision, has made significant progress in recent years, particularly with the advent of deep learning [6, 8, 18, 41, 46]. The objective of this task is to locate objects in images with tight bounding boxes and recognize their respective categories. However,", + "bbox": [ + 75, + 803, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/25e562aad69e873247e74514ab066c5b0cdcd46222ccbfd44926dd867c5681f8.jpg", + "image_caption": [ + "Predefined category", + "(a) Open-Vocabulary Object Detection" + ], + "image_footnote": [], + "bbox": [ + 509, + 265, + 836, + 344 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9ddb8aacd764116ab0bde92874b7dc211320af2d0e83f2d2064ee6d99e95c1b8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 508, + 369, + 921, + 459 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b6a34afda8e313e9d2d66ec8e143e972b2e349ac9d2c7743256850291ac94d45.jpg", + "image_caption": [ + "(b) Phrase Grounding", + "(c) Generative Open-Ended Object Detection", + "Figure 1. Comparing generative open-ended object detection with other open-set object detection tasks. Open-vocabulary object detection and phrase grounding typically require predefined categories or phrases in text prompts to align with image regions. In contrast, our introduced generative open-ended object detection is a more general and practical setting where categorical information is not explicitly defined. Such a setting is especially meaningful for scenarios where users lack precise knowledge of object categories during inference." + ], + "image_footnote": [], + "bbox": [ + 509, + 477, + 919, + 537 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "most existing object detection algorithms are restricted to a predetermined set of object categories defined in detection datasets [9, 14, 17, 26, 42]. For example, an object detector trained on COCO [9] can well identify 80 pre-defined classes but encounters difficulty in recognizing new classes that have not been seen or defined during training.", + "bbox": [ + 496, + 744, + 893, + 837 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Open-vocabulary learning has been introduced to remedy this problem and considerable efforts have been made [7, 22, 28, 29, 35, 55, 65]. Open-vocabulary object detection addresses the closed-set limitation typically via", + "bbox": [ + 496, + 839, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Authors", + "bbox": [ + 94, + 886, + 228, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13958", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "weak supervision from language, i.e., image-text pairs, or large pretrained Vision-Language Models (VLMs), such as CLIP [39]. The text embeddings from the text encoder of CLIP can well align with the visual regions for the novel categories or the noun phrases in image captions, as shown in Fig. 1. However, despite open-set nature, open-vocabulary object detection still requires predefined object categories during the inference stage, while in many practical scenarios, we might not have exact knowledge of test images. Even if we have prior knowledge of test images, manually defining object categories might introduce language ambiguities (e.g., similar object names like \"person\", \"a man\", \"young boy\" in Fig. 1) or be not comprehensive enough (e.g. missing \"chain link fence\" in Fig. 1), ultimately restricting flexibility. If we do not have prior knowledge of test images, to cover all possible objects that may appear in the images, the common solution to design a large and comprehensive label set that would be complex and time-consuming. The above discussions raise the research question: Can we do open-world dense object detection that does not require predefined object categories during inference? We call this problem open-ended object detection.", + "bbox": [ + 76, + 90, + 472, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Therefore, in this paper, we formulate open-ended object detection as a generative problem and propose a simple architecture termed GenerateU to address it. Our GenerateU contains two major components: a visual object detector to localize image regions and a language model to translate visual regions to object names. This model is trained end-to-end, optimizing both components simultaneously. To broaden the knowledge of GenerateU, we train it using a small set of human-annotated object-language paired data and scale up the vocabulary size with massive image-text pairs of data. Furthermore, due to the limitation in image captions whose nouns often fall short of accurately representing all objects in the images, we employ a pseudolabeling method to supplement missing objects in images.", + "bbox": [ + 76, + 431, + 472, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our major contributions are summarized as follows.", + "bbox": [ + 96, + 651, + 439, + 666 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a new and more practical object detection problem: open-ended object detection, and formulate it as a generative problem. The reformulation avoids manual effort in predefining object categories during both training and inference, aiming at a more general and flexible architecture.", + "- We develop a novel end-to-end learning framework, termed GenerateU, which directly generates object names in a free-form manner. GenerateU is trained with a small set of human-annotated object-language paired data and massive image-text paired data.", + "- Compared with open-vocabulary object detection models, GenerateU achieves comparable results on zero-shot LVIS, even though it does not see object categories during inference." + ], + "bbox": [ + 76, + 674, + 468, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 89, + 640, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Open-Vocabulary Object Detection", + "text_level": 1, + "bbox": [ + 500, + 114, + 805, + 131 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Learning visual models from language supervision becomes popular in image recognition tasks [19, 20, 39, 40, 58, 59, 62, 64]. Recently, many works [23, 23-25, 34, 43-45, 48, 49, 52-54, 56, 57] have extended the success in open-vocabulary image recognition to object detection, i.e., open-vocabulary object detection (OVOD). As the pioneering work of OVOD, OVRCNN [61] successfully transfers the vision and language model trained on image-text pairs to the detection framework. To leverage the capability of CLIP [39], ViLD [16] and RegionCLIP [63] turn to learn the visual region feature from classification-oriented models by knowledge distillation. OV-DETR [60] proposes a novel open-vocabulary detector built upon DETR [8] by formulating the classification as a binary matching between the input queries and the referring objects. Different from distilling frozen vision encoder, OWL [35] proposes a two-step framework for contrastive image-text pre-training and end-to-end detection fine-tuning, using a standard Vision Transformer [13]. Detic [65] proposes to extend the detector vocabulary with large-scale image classification data by applying supervision to the proposal with the largest spatial size. To reduce the annotation cost for open-vocabulary object detection, VLDet [29] proposes a novel framework to directly learn from image-text paired data by formulating object-language alignment as a set matching problem.", + "bbox": [ + 496, + 138, + 893, + 517 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Open vocabulary object detection aims at generalization to novel categories that were not seen during the training phase. Existing OVOD works typically rely on calculating the similarity between image regions and a set of arbitrary category names with a pretrained vision-language model. It remains necessary to predefine the categories during inference. In contrast, this work delves deeper into a new challenge - how to effectively handle scenarios where the exact categories remain unknown during the inference process.", + "bbox": [ + 496, + 517, + 893, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Multimodal Large Language Model", + "text_level": 1, + "bbox": [ + 500, + 665, + 812, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by the remarkable capability of LLM [1, 4, 47, 50], multimodal foundation models [2, 15, 27, 30, 37, 38] leverage LLM as the core intelligence and are trained with extensive image-text pairs, resulting in enhanced performance across various vision-and-language tasks such as VQA and image captioning. BLIP-2 [27] proposes a generic and efficient pre-training strategy that bootstraps vision-language pre-training from off-the-shelf frozen pre-trained image encoders and frozen large language models. By inserting adapters into LLaMA's transformer, LLaMA-Adapter [15] introduces only 1.2M learnable parameters, and turns a LLaMA into an instruction-following model. LLaVA [30] represents a novel end-to-end trained large multimodal model that combines a vision encoder and Vicuna [36]", + "bbox": [ + 496, + 688, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13959", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "for general-purpose visual and language understanding. It achieves impressive chat capabilities, emulating the essence of the multimodal GPT-4.", + "bbox": [ + 75, + 90, + 468, + 135 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Based on this amazing progress, subsequent studies further explore enhancing the object localization capabilities of multimodal models. Kosmos-2 [37] unlocks the grounding capability, enabling the model to provide visual answers such as bounding boxes, thereby supporting a wider array of vision-language tasks, including referring expression comprehension. UNINEXT[55] reformulates diverse instance perception tasks into a unified object discovery and retrieval paradigm and can flexibly perceive different types of objects by simply changing the input prompts. DetGPT [38] locates the objects of interest based on the language instructions and the LLM enables users to interact with the system, allowing for a higher level of interactivity like visual reasoning. In contrast to these works that focus on understanding natural language expressions referring to specific objects in an image, our work aims to identify and localize all objects of interest in an image, associating each with a class name.", + "bbox": [ + 75, + 136, + 470, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Dense Captioning", + "text_level": 1, + "bbox": [ + 76, + 401, + 250, + 417 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dense captioning aims to predict a set of descriptions across various image regions. DenseCap [21] develops a fully convolutional localization network to extract image regions, which are then described using an RNN language model. In dense captioning, it treats each region as an individual image and tries to generate captions encompassing multiple objects and their attribute, such as \"cat riding a skateboard\" and \"bear is wearing a red hat\". To evaluate the generated captions, dense captioning employs the METEOR [21] metric, originally designed for machine translation, which assesses the generation results from the sentence perspective. In comparison, our proposed generative open-ended object detection focuses on describing individual objects as category names in a zero-shot setting.", + "bbox": [ + 75, + 425, + 468, + 635 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Note that recently, CapDet [32] and GRiT [51] attempted to integrate open-vocabulary object detection and dense captioning by introducing an additional language model as a captioning head. However, for open-world object detection, they either still require a predefined category space during inference, or are limited to the close-set object detection problem.", + "bbox": [ + 75, + 636, + 468, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 753, + 166, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to build an open-world object detector capable of localizing all objects in an image and providing their corresponding category names in a free-form way. Figure 2 gives an overview of our proposed open-ended object detection model, GenerateU, which comprises two major components: an open-world object detector and a language model. We describe the open-world object detector in Sec. 3.1, which consists of an image encoder and a detection head,", + "bbox": [ + 75, + 780, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and the language model, which can be directly borrowed from an existing Multimodal Large Language Model (including a V-to-L adaptor and a language model, Sec. 3.2) or retrained (Sec. 3.3). An additional pseudo-labeling method is described in Sec. 3.4 for enriching the label diversity.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Open-World Object Detection", + "text_level": 1, + "bbox": [ + 498, + 176, + 766, + 193 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The first crucial step is to extract accurate object regions from images, for which we develop an open-world object detector. Although we are not constrained to any specific object detection model, we choose Deformable DETR [66] for the following reasons. First, it alleviates heuristics when compared to other popular one or two-stage frameworks, and provides a versatile query-to-instance pipeline. Besides, Deformable DETR [66] demonstrates faster convergence speed and enhanced accuracy compared to the original DETR [8]. Deformable DETR or the original one employs Hungarian matching to learn a mapping from predicted queries to ground truth objects, followed by training the matched object query to regress to its corresponding ground truth with a combination of classification loss and bounding box regression loss. In the context of open-set problems, we do not rely on the object class information. Instead, we adopt an open-world detection approach, i.e. class-agnostic object detector, wherein matched queries are only classified into foreground or background. Consequently, the detection process involves three losses: binary cross-entropy, generalized IoU, and L1 regression loss.", + "bbox": [ + 496, + 200, + 890, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Transferring from a Frozen Multimodal LLM", + "text_level": 1, + "bbox": [ + 498, + 529, + 888, + 545 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Once object candidates are localized, a natural idea is to leverage a pretrained multimodal large language model (MLLM) to reduce the training cost and offer strong zero-shot language abilities for generating object names. Recently developed MLLMs [2, 15, 27, 30, 37, 38] typically consist of three components: an image encoder to extract visual features; an adaptor network to bridge the modality gap and reduce the number of image embeddings, converting visual feature into token representations; a language model to generate predictions for vision-centric tasks. As shown in Figure 2 Top, we integrate a class-agnostic Deformable DETR, with a frozen image encoder to extract object queries, into an MLLM. We keep the MLLM frozen, and the only trainable module is the detection head for learning object representations. Besides, images are preprocessed to a resolution of $224 \\times 224$ pixels during training in KOSMOS-2 [37]. In object detection, it is common to employ a larger input size for multi-scale feature extraction. To handle changes in image size, we utilize linear interpolation to enlarge the original position embedding.", + "bbox": [ + 496, + 551, + 890, + 853 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Despite its simplicity, our empirical study shows that such a direct open-end object detection based on pretrained MLLM results in poor performance large due to the do", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13960", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e12d69e0c9c9cc249edb36316903e6ffc8e8b7993cb7b370fc7e2246ed3b7574.jpg", + "image_caption": [ + "(a) Training with frozen MLLM" + ], + "image_footnote": [], + "bbox": [ + 86, + 87, + 890, + 167 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3f9b49cba6219db8d05f7a320af1c6dbbb631c9358b78aad3cfc286fc2316452.jpg", + "image_caption": [ + "(b) Training End-to-End", + "Figure 2. Overview of our proposed open-ended object detection model, GenerateU, which comprises two major components: an object detector and a language model. We compare two training strategies: (Top) We incorporate the class-agnostic DETR (with frozen image encoder) into a pre-trained and frozen Multimodal Large Language Model (including Adaptor and Language Model), to facilitate a smooth transfer of knowledge from the language domain to object detection; (Bottom) We activate the image encoder and language model as trainable components, taking an end-to-end approach to seamlessly integrate region-level understanding into the language model." + ], + "image_footnote": [], + "bbox": [ + 86, + 205, + 750, + 286 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "main shift. MLLMs were trained on large-scale image-text paired data, where they generate text descriptions given each whole image as a condition, without a fine-grained understanding of individual objects.", + "bbox": [ + 75, + 398, + 470, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Generative Region-Language Pretraining", + "text_level": 1, + "bbox": [ + 76, + 469, + 431, + 486 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Thus, instead of employing a frozen MLLM, we propose GenerateU to directly link the open-world object detector with a language model, and activate both the image encoder and the language model as trainable components, as shown in Figure 2. Particularly, we employ an encoder-decoder-based language model, where visual representation serves as input for the encoder, and the associated text serves as the generation target for the language decoder. Here we only use the object queries that have been successfully matched with ground truth through Hungarian matching, ignoring the unmatched queries in the generating process. Specifically, each matched object query is first treated as an individual region, fed to a projection function to align its dimensions with the language model. The encoder comprises blocks with self-attention layers and feed-forward networks. The decoder, similar in structure, adds a cross-attention mechanism for interfacing with the encoder. The decoder employs an autoregressive self-attention mechanism, limiting the model's attention to past outputs, and its final output is processed through a dense layer with a softmax function. We apply the language modeling loss to train the model.", + "bbox": [ + 75, + 492, + 468, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Exploring on the region-word alignment loss. Benefiting from the pretrained vision-language model, we incorporate a region-word alignment loss during training to aid in learning to distinguish region features. In detail, for an object query, we treat its corresponding word as a positive sample", + "bbox": [ + 75, + 825, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and consider other words in the same minibatch as negative samples. By calculating the similarity between the region features from class-agnostic Deformable DETR and text features from the fixed CLIP text encoder [39], the model is optimized by a binary cross-entropy loss. This region-word alignment loss is jointly trained with the object detection losses and the language modeling loss.", + "bbox": [ + 496, + 398, + 893, + 506 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Enrich Label Diversity", + "text_level": 1, + "bbox": [ + 498, + 537, + 712, + 555 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Considerable efforts have been invested in collecting detection data that is both semantically rich and voluminous. However, human annotation proves to be both costly and resource-restrictive. Previous works, such as Kosmos2 [37] and GLIP [28], leverage image-text pairs to expand the vocabulary scale. They use pretrained grounding models to generate bounding boxes aligned with noun phrases from captions. These boxes, in turn, serve as pseudo detection labels for training the model. However, given that captions may not comprehensively describe every object in an image, the count of nouns in the caption is typically much lower than the actual number of objects present in the corresponding image. Consequently, the quantity of pseudolabels generated is constrained. To address this, we use GenerateU pretrained on the available labels to generate pseudo labels as a supplement for missing objects in images. Furthermore, we observed that beam search in language models naturally generates synonyms, thereby providing diverse object labels. Figure 3 provides pseudolabel examples, illustrating the generation of boxes covering nearly all objects in the image, along with a diverse set of rich vocabulary labels.", + "bbox": [ + 496, + 568, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13961", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a5ab296f5fe245b9c1426f2d047d4f778caff974d4dcad1083fd693282fe2735.jpg", + "image_caption": [ + "\"Belcher family you and me we got this poster A1.\"" + ], + "image_footnote": [], + "bbox": [ + 169, + 88, + 344, + 215 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7a06a49ef61887658a4ba9fb018be6b94279cd338191f5ba2cec2fa862a8f080.jpg", + "image_caption": [ + "\"A gorgeous wedding centrepiece with feathers, moss, greenery, white blooms and greenery.\"" + ], + "image_footnote": [], + "bbox": [ + 377, + 88, + 506, + 217 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d99c7767488e108e6140b15d842d9c08213d4e88ce57711141a2f008505021d1.jpg", + "image_caption": [ + "\"Just before Christmas, Thornton's issued a profit warning after suffering low order levels from supermarket customers and problems at its new warehouse.\"" + ], + "image_footnote": [], + "bbox": [ + 558, + 88, + 790, + 217 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5d65a0938866d9eb1b9c62b29449f25f26559c7f8a2f165ace301636a1ab7a1b.jpg", + "image_caption": [ + "\"Boy in red shirt catching a ball.\"" + ], + "image_footnote": [], + "bbox": [ + 168, + 244, + 364, + 359 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b3e42baa73626b78b379e1c29c344166bb06ffb522d97476cbd013f28ff88b35.jpg", + "image_caption": [ + "\"Adams, seen here celebrating after scoring against Chesterfield.\"" + ], + "image_footnote": [], + "bbox": [ + 390, + 244, + 511, + 359 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d87f46a45fa9811eba6cd90f94d1cffe6329d5a6b1c0fcdefbf85822521200c8.jpg", + "image_caption": [ + "\"Blonde woman with a colourful umbrella being blown away by the wind with her umbrella snowy landscape.\"" + ], + "image_footnote": [], + "bbox": [ + 570, + 244, + 790, + 359 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c577a96246ee033c561625452eb3f432f40a7870178791c960ac6a17b32268d9.jpg", + "image_caption": [ + "\"Meatballs on top of spaghetti squash on a white plate with rest behind it in skillet.\"" + ], + "image_footnote": [], + "bbox": [ + 168, + 377, + 307, + 512 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/58dfc839dd86f12068800f1167777d5eb6855852ea47a466edddecd173fe1296.jpg", + "image_caption": [ + "\"How to make fresh brewed iced tea sweetened or without sugar. Homemade recipe using tea bags with a squeeze of lemon or orange juice and a sprig of mint.\"" + ], + "image_footnote": [], + "bbox": [ + 364, + 377, + 493, + 510 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/dc103776b5d647320b6693445d711adfee7765776020cd8e68ae0f3488221cb6.jpg", + "image_caption": [ + "\"Close up picture of a plate of tuna noodle casserole with a bowl of salad in the background.\"" + ], + "image_footnote": [], + "bbox": [ + 555, + 377, + 790, + 512 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7173af299b8b3acc84c36f764fbaba2a031585e6c6afe5549ed808c915a8479d.jpg", + "image_caption": [ + "\"One of a kind: meet the only BMW E3 Estate in the world.\"", + "Figure 3. Selected pseudo-label examples highlight the generation of bounding boxes covering nearly all objects in the images. A varied set of descriptive labels showcases the model's capability in producing diverse and linguistically rich vocabulary. A white underline in a figure indicates that the pseudo label comes from the noun phrase (black underline) in the caption." + ], + "image_footnote": [], + "bbox": [ + 168, + 540, + 390, + 652 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c4351951b984b01fef65660c6550c75ac0f86e2df99623997fe920ed5b900988.jpg", + "image_caption": [ + "\"The last time with his S15, Matt was always in the need of more power.\"" + ], + "image_footnote": [], + "bbox": [ + 410, + 540, + 596, + 652 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/968e8f549491e2fca42421ffd39c110a7a842c196083f6e8e0be81d69742f149.jpg", + "image_caption": [ + "\"Construction Junction box truck parked in front of the loading dock.\"" + ], + "image_footnote": [], + "bbox": [ + 622, + 541, + 789, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 756, + 207, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 76, + 784, + 178, + 797 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our model is trained with hybrid supervision from two types of data: detection data and image-text data. For detection data, we use Visual Genome (VG) [26] which contains 77,398 images for training. Different from previous works [32, 51], we do not use the region captioning annotations of VG, since they may contain descriptions that", + "bbox": [ + 75, + 809, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "encompass multiple objects or even describe actions. Our focus is directed towards the object detection task, and thus we leverage the object annotations of VG which annotate the same bounding box regions with concise and specific object names, covering a wide range of objects. For pretraining, we use about 5 million image-text pairs GRIT [37] from COYO700M [5] with pseudo grounding labels generated by KOSMOS-2 [37] using GLIP [28].", + "bbox": [ + 496, + 757, + 890, + 878 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13962", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f4bc7fb1dc95aeb4ff379d415c99db5e599ae0713d8d012da44e135dfc7084a8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBackbonePre-Train DataOpen-endedOpen-setLVIS
APrAPcAPfAP
MDETR [22]RN101LVIS--20.924.924.324.2
MaskRCNN [18]RN101LVIS--26.334.033.933.3
Deformable DETR [66]Swin-TLVIS--24.236.038.236.0
GLIP (C) [28]Swin-TO365, GoldG-✓17.719.531.024.9
GLIP [28]Swin-TO365, GoldG, CAP4M-✓20.821.431.026.0
GenerateUSwin-TVG✓✓17.422.429.625.4
GenerateUSwin-TVG, GRIT5M✓✓20.024.929.826.8
GenerateUSwin-LVG, GRIT5M✓✓22.325.231.427.9
", + "bbox": [ + 76, + 88, + 893, + 268 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Zero-shot domain transfer to LVIS. Our method achieves comparable performance to prior models in the close-ended setting, without requiring access to category names during inference.", + "bbox": [ + 75, + 277, + 892, + 309 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Evaluation protocol and metrics", + "text_level": 1, + "bbox": [ + 76, + 333, + 361, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Addressing an open-ended problem presents challenges in evaluating performance. For instance, when annotating a laptop, the human-annotated ground truth category may be \"laptop\", \"computer\", \"PC\", \"Microcomputer\", etc. To assess the effectiveness of the open-end generative results, we propose two settings: (a) We calculate the similarity score between the generated and annotated category names with a fixed pretrained text encoder. Notice that this text encoder is not utilized by the generative detection model; rather, it is only used for quantitative performance measurement. Any text encoder capable of computing similarity between two sentences can be adopted for this purpose. In this study, we select the popular CLIP [39] text encoder and BERTlarge [12] model. In this way, we can map the generative results to predefined categories in a dataset for evaluation. (b) We use METEOR [3] to assess the generated text quality, which is the same as the evaluation in dense captioning [21]. METEOR is widely used in NLP for evaluating machine translations. Following [21], we measure the mean Average Precision (AP) for both localization, using Intersection over Union (IoU) thresholds of .3, .4, .5, .6, and .7, and language accuracy, applying METEOR score thresholds at 0, .05, .1, .15, .2, and .25.", + "bbox": [ + 73, + 356, + 472, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following GLIP [28], we evaluate our GenerateU primarily on LVIS [17] which contains 1203 categories and report APs for rare categories $(AP_r)$ , common categories $(AP_c)$ , frequency categories $(AP_f)$ and all categories $(mAP)$ , respectively. Specifically, following GLIP [28] and MDETR [22], we evaluate on the $5\\mathrm{k}$ minival subset and report the zero-shot fixed AP [11] for a fair comparison.", + "bbox": [ + 75, + 704, + 468, + 811 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 816, + 294, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We attempt two different backbones, Swin-Tiny and Swin-Large [31] as the visual encoder. The Deformable DETR architecture follows [66] with 6 encoder layers and 6 decoder layers. The number of object queries $N$ is set to 300.", + "bbox": [ + 75, + 839, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We adopt FlanT5-base [10] as the language model and initialize the weights from it. The optimizer is AdamW [33] with a weight decay of 0.05. The learning rate is set to $2 \\times 10^{-4}$ for the parameters of the visual backbone and the detection head, and $3 \\times 10^{-4}$ for the language model. For the frozen MLLM approach in Sec 3.2, we explore the state-of-the-art model KOSMOS-2 [37] which was trained with the multimodal corpora. We fix all parameters and set the learning rate to $2 \\times 10^{-4}$ for the detection head. Our model is trained on 16 A100 GPUs.", + "bbox": [ + 496, + 333, + 893, + 486 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Generative Open-Ended Detection Results", + "text_level": 1, + "bbox": [ + 498, + 497, + 861, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We assess the model's capacity to identify diverse objects in a zero-shot setting on LVIS, presenting the results in Table 1, where rows 1-3 include methods utilizing LVIS as pre-training data, representing fully supervised models. Remarkably, we observe that even with only VG as training data, our model demonstrates commendable performance on zero-shot LVIS. This result suggests that predefined class names might not be necessary for open-world object detection during inference, especially when models are trained with extensive visual concepts. Furthermore, introducing additional training data from cost-effective image-text pairs consistently improves performance. For instance, when GenerateU trained with additional image-text pair data (GRIT [37]), we observe a substantial performance gain of 3.6 on $AP_{r}$ and 1.4 on $AP$ . In summary, the semantic richness derived from image-text data significantly enhances the model's ability to recognize rare objects. Notably, our model achieves comparable performance to GLIP under the same training data size but without requiring access to category names during inference. Moreover, we enhanced our model by integrating a larger backbone, i.e. Swin-Large, confirming that our method can scale effectively with increased model complexity.", + "bbox": [ + 496, + 521, + 893, + 868 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Transfer to other datasets. Our generative open-ended object detector can transfer to any dataset without any mod-", + "bbox": [ + 498, + 869, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13963", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3a762135de65307a5b3b15cb680707e0c32c27bbcecd32a8f251c2c01f3f8e01.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsImage EncoderDetection HeadLanguage ModelLVIS
APrAPcAPfAP
Init from MLLMFreezeTrainFreeze*5.310.219.714.3
TrainTrainFreeze*10.917.228.121.9
Init from LLMTrainTrainFreeze11.220.127.422.9
TrainTrainTrain20.024.929.826.8
", + "bbox": [ + 78, + 88, + 890, + 202 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ef069a0dc76299e1d5b347d997cd481c49a189a43d6de33cff188909422e0636.jpg", + "table_caption": [ + "Table 2. Comparison of our GenerateU with two different training strategies on a pretrained LLM on zero-shot LVIS. The results highlight generative object detection benefits from end-to-end training. Here, $\\star$ refers to freezing both V-to-L adaptor and language model." + ], + "table_footnote": [], + "table_body": "
MethodsPre-Train DataCOCOObjects365
Supervised-46.525.6
GenerateUVG33.010.1
GenerateUVG, GRIT33.610.5
", + "bbox": [ + 76, + 263, + 468, + 340 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a2b83b2455ab360e1f2e80ef7b5db93b1f7bed44f589a5c0894a46b7c9301a53.jpg", + "table_caption": [ + "Table 3. We evaluate our GenerateU on COCO and Objects365 validation set in a zero-shot setting. GenerateU can transfer to various downstream datasets without requiring any modifications." + ], + "table_footnote": [], + "table_body": "
MetricsAPrAPcAPfAP
CLIP [39]20.024.929.826.8
BERT [12]12.519.925.121.7
METEOR [3]---9.2
", + "bbox": [ + 76, + 407, + 464, + 484 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Effect of different evaluation metrics on zero-shot LVIS.", + "bbox": [ + 76, + 494, + 465, + 508 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ifications. In contrast to open vocabulary object detection methods, we do not need to switch the classifier to the specific category text embeddings. We evaluate the transferability of the proposed GenerateU on COCO [9] and Objects365 [42] validation sets, as shown in Table 3. The results demonstrate the effectiveness of our method and underscore the importance of scaling up the training data.", + "bbox": [ + 73, + 535, + 468, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 75, + 650, + 230, + 666 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with transferring from a Frozen MLLM. In Table 2, we evaluate the approach discussed in Sec 3.2, i.e. transferring a frozen MLLM to the generative object detection task. Following the state-of-the-art MLLM model KOSMOS-2 [37], we maintain its parameters fixed and only train the detection head. However, we observe that such direct transfer of the MLLM to object detection leads to poor performance in zero-shot LVIS. We attribute this result to the pretrained image encoder, which is trained on images of size $224 \\times 224$ , which is insufficient for the object detection task. Besides, the MLLM originally supervised by image-level annotations such as image captions, lacks understanding in both object localization and fine-grained object categories. We also experiment with making the image encoder trainable while keeping the language model frozen. This", + "bbox": [ + 73, + 672, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c40eadc437b848cf075267e1b41a2fcbd93ad9730c7e694909684c4d3a8d55c2.jpg", + "image_caption": [ + "Figure 4. Effect of beam size. Beam search plays a crucial role in generating rare object names, effectively addressing the long-tail problem." + ], + "image_footnote": [], + "bbox": [ + 504, + 268, + 883, + 420 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "adjustment significantly improves the results. We further fine-tune the entire model, initializing the Large Language Model (LLM) from FlanT5-base [10], which achieves the best performance. The above analysis underscores the importance of end-to-end training in the open-ended object detection task.", + "bbox": [ + 496, + 503, + 890, + 594 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Beam Search. Towards a general and practical object detection system, we believe that beam search plays a crucial role in generating outputs that are closely aligned with human intuition. This is because it explores a broader array of possibilities during the decoding phase. Beam search naturally contains the multi-level object names and their synonyms, enhancing the system's ability to recognize objects with varied linguistic expressions. This contributes to a more adaptable and nuanced understanding of the objects being detected. As shown in Figure 4, beam search largely improves the rare category performance, proving its importance for the long tail problem, and emphasizing the system's effectiveness in handling less frequent object categories. In our experiments, we set the default beam size to 3.", + "bbox": [ + 496, + 595, + 892, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation Metrics. To evaluate the results of generative object detection, we use different metrics: METEOR, evaluating the quality of text, and similarity scores, calculated by a fixed text encoder. Note that the off-the-shelf text encoder is not used for the proposed method, but just for the", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13964", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/22645868d459770bef30c5708c58d3a11e89663802b14a6e9377ee97b5ac3905.jpg", + "image_caption": [ + "Figure 5. Qualitative prediction results from GenerateU and ground truth on LVIS [17]. GenerateU produces complete and precise predictions, showcasing its ability to go beyond fixed vocabulary constraints." + ], + "image_footnote": [], + "bbox": [ + 130, + 88, + 849, + 435 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d60254961e0b8c1b11c632257a3800e91d999add5d465cd464837da6a0d82236.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Pre-Train DataRegion-word AlignmentAPrAPcAPfAP
VG-16.019.126.722.5
✓17.4↑1.422.4↑3.929.6↑2.925.4↑2.9
VG, GRIT-19.223.528.925.7
✓20.0↑0.824.9↑1.429.8↑0.926.8↑0.9
", + "bbox": [ + 76, + 497, + 486, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Effect of region-word alignment loss.", + "bbox": [ + 135, + 621, + 410, + 636 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "evaluation. Here we compare two offline models: CLIP text encoder [39] and BERT [12], as presented in Table 4. We find that the CLIP text encoder is particularly adept at handling noun phrases, while the BERT encoder, predominantly trained on sentence corpora, may exhibit different strengths in processing textual information.", + "bbox": [ + 75, + 670, + 468, + 761 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Region-word Alignment. We conduct ablation experiments to evaluate the effectiveness of the region-word alignment loss in our method, as presented in Table 5. We observe that training with only the detection loss and the language modeling loss yields promising results. Adding the region-word alignment loss provides additional supervision and further improves the performance. This further demonstrates the effectiveness of the proposed generative object detection method.", + "bbox": [ + 75, + 763, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Qualitative Results", + "text_level": 1, + "bbox": [ + 500, + 500, + 681, + 515 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Here we provide qualitative visualizations in showcasing the superior object detection capabilities of our model. As shown in Figure 5, our model can identify a more diverse range of objects compared to ground truth annotations. These visualizations not only affirm the reliability of our detection framework but also demonstrate its ability to detect a wider variety of objects, providing valuable insights into the model's strengths.", + "bbox": [ + 496, + 532, + 890, + 655 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 698, + 617, + 714 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper introduces GenerateU, a novel approach for generative open-ended object detection. By reformulating object detection as a generative task, the model eliminates the need for predefined categories during inference, offering a more practical solution. Our method achieves comparable results with open-vocabulary models on zero-shot LVIS, even though categories are not seen in our GenerateU. Overall, this work contributes to the fast-evolving area of object detection towards open-world scenarios. Future works include investigating the effect of training data scales and exploring methods beyond naive pseudo-labeling.", + "bbox": [ + 496, + 732, + 890, + 888 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13965", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Rohan Anil, Andrew M Dai, Orhan First, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, et al. Palm 2 technical report. arXiv preprint arXiv:2305.10403, 2023. 2", + "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 2, 3", + "[3] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 6, 7", + "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 2", + "[5] Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, and Saehoon Kim. Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset, 2022.5", + "[6] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6154-6162, 2018. 1", + "[7] Zhaowei Cai, Gukyeong Kwon, Avinash Ravichandran, Erhan Bas, Zhuowen Tu, Rahul Bhotika, and Stefano Soatto. X-detr: A versatile architecture for instance-wise vision-language tasks. arXiv preprint arXiv:2204.05626, 2022. 1", + "[8] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 1, 2, 3", + "[9] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 1, 7", + "[10] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416, 2022. 6, 7", + "[11] Achal Dave, Piotr Dollar, Deva Ramanan, Alexander Kirilov, and Ross Girshick. Evaluating large-vocabulary object detectors: The devil is in the details. arXiv preprint arXiv:2102.01066, 2021. 6", + "[12] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 6, 7, 8", + "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner," + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2", + "[14] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303-338, 2010. 1", + "[15] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xi-angyu Yue, et al. Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010, 2023. 2, 3", + "[16] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021. 2", + "[17] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 1, 6, 8", + "[18] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE International Conference on Computer Vision, pages 2961-2969, 2017. 1, 6", + "[19] Dat Huynh, Jason Kuen, Zhe Lin, Jiumiang Gu, and Ehsan Elhamifar. Open-vocabulary instance segmentation via robust cross-modal pseudo-labeling. arXiv preprint arXiv:2111.12698, 2021. 2", + "[20] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 2", + "[21] Justin Johnson, Andrej Karpathy, and Li Fei-Fei. Densecap: Fully convolutional localization networks for dense captioning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4565–4574, 2016. 3, 6", + "[22] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. MDetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1780-1790, 2021. 1, 6", + "[23] Prannay Kaul, Weidi Xie, and Andrew Zisserman. Multimodal classifiers for open-vocabulary object detection. arXiv preprint arXiv:2306.05493, 2023. 2", + "[24] Dahun Kim, Anelia Angelova, and Weicheng Kuo. Contrastive feature masking open-vocabulary vision transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15602-15612, 2023.", + "[25] Dahun Kim, Anelia Angelova, and Weicheng Kuo. Region-aware pretraining for open-vocabulary object detection with vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11144-11154, 2023. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13966", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123(1):32-73, 2017. 1, 5", + "[27] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 2, 3", + "[28] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. arXiv preprint arXiv:2112.03857, 2021. 1, 4, 5, 6", + "[29] Chuang Lin, Peize Sun, Yi Jiang, Ping Luo, Lizhen Qu, Gholamreza Haffari, Zehuan Yuan, and Jianfei Cai. Learning object-language alignments for open-vocabulary object detection. arXiv preprint arXiv:2211.14843, 2022. 1, 2", + "[30] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 2, 3", + "[31] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 6", + "[32] Yanxin Long, Youpeng Wen, Jianhua Han, Hang Xu, Pengzhen Ren, Wei Zhang, Shen Zhao, and Xiaodan Liang. Capdet: Unifying dense captioning and open-world detection pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15233-15243, 2023. 3, 5", + "[33] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6", + "[34] Chuofan Ma, Yi Jiang, Xin Wen, Zehuan Yuan, and Xiaojuan Qi. Codet: Co-occurrence guided region-word alignment for open-vocabulary object detection. arXiv preprint arXiv:2310.16667, 2023. 2", + "[35] Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, et al. Simple open-vocabulary object detection with vision transformers. arXiv preprint arXiv:2205.06230, 2022. 1, 2", + "[36] Baolin Peng, Chunyuan Li, Pengcheng He, Michel Galley, and Jianfeng Gao. Instruction tuning with gpt-4. arXiv preprint arXiv:2304.03277, 2023. 2", + "[37] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 2, 3, 4, 5, 6, 7", + "[38] Renjie Pi, Jiahui Gao, Shizhe Diao, Rui Pan, Hanze Dong, Jipeng Zhang, Lewei Yao, Jianhua Han, Hang Xu, and Lingpeng Kong Tong Zhang. Detgpt: Detect what you need via reasoning. arXiv preprint arXiv:2305.14167, 2023. 2, 3", + "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry," + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 4, 6, 7, 8", + "[40] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. arXiv preprint arXiv:2112.01518, 2021. 2", + "[41] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In Advances in Neural Information Processing Systems, pages 91-99, 2015. 1", + "[42] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019. 1, 7", + "[43] Cheng Shi and Sibei Yang. Edadet: Open-vocabulary object detection using early dense alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15724-15734, 2023. 2", + "[44] Hengcan Shi, Munawar Hayat, and Jianfei Cai. Open-vocabulary object detection via scene graph discovery. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4012-4021, 2023.", + "[45] Hengcan Shi, Munawar Hayat, and Jianfei Cai. Unified open-vocabulary dense visual prediction. arXiv preprint arXiv:2307.08238, 2023. 2", + "[46] Peize Sun, Rufeng Zhang, Yi Jiang, Tao Kong, Chenfeng Xu, Wei Zhan, Masayoshi Tomizuka, Lei Li, Zehuan Yuan, Changhu Wang, et al. Sparse r-cnn: End-to-end object detection with learnable proposals. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14454-14463, 2021. 1", + "[47] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2", + "[48] Luting Wang, Yi Liu, Penghui Du, Zihan Ding, Yue Liao, Qiaosong Qi, Biaolong Chen, and Si Liu. Object-aware distillation pyramid for open-vocabulary object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11186-11196, 2023. 2", + "[49] Tao Wang. Learning to detect and segment for open vocabulary object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7051-7060, 2023. 2", + "[50] BigScience Workshop, Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ilic, Daniel Hesslow, Roman Castagné, Alexandra Sasha Luccioni, François Yvon, et al. Bloom: A 176b-parameter open-access multilingual language model. arXiv preprint arXiv:2211.05100, 2022. 2", + "[51] Jialian Wu, Jianfeng Wang, Zhengyuan Yang, Zhe Gan, Zicheng Liu, Junsong Yuan, and Lijuan Wang. Grit: A generative region-to-text transformer for object understanding. arXiv preprint arXiv:2212.00280, 2022. 3, 5" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13967", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Jianzong Wu, Xiangtai Li, Henghui Ding, Xia Li, Guangliang Cheng, Yunhai Tong, and Chen Change Loy. Betrayed by captions: Joint caption grounding and generation for open vocabulary instance segmentation. arXiv preprint arXiv:2301.00805, 2023. 2", + "[53] Size Wu, Wenwei Zhang, Sheng Jin, Wentao Liu, and Chen Change Loy. Aligning bag of regions for open-vocabulary object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15254-15264, 2023.", + "[54] Xiaoshi Wu, Feng Zhu, Rui Zhao, and Hongsheng Li. Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7031-7040, 2023. 2", + "[55] Bin Yan, Yi Jiang, Jiannan Wu, Dong Wang, Ping Luo, Zehuan Yuan, and Huchuan Lu. Universal instance perception as object discovery and retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15325-15336, 2023. 1, 3", + "[56] Lewei Yao, Jianhua Han, Youpeng Wen, Xiaodan Liang, Dan Xu, Wei Zhang, Zhenguo Li, Chunjing Xu, and Hang Xu. Detclip: Dictionary-enriched visual-concept paralleled pretraining for open-world detection. Advances in Neural Information Processing Systems, 35:9125-9138, 2022. 2", + "[57] Lewei Yao, Jianhua Han, Xiaodan Liang, Dan Xu, Wei Zhang, Zhenguo Li, and Hang Xu. Detclipv2: Scalable open-vocabulary object detection pre-training via word-region alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23497-23506, 2023. 2", + "[58] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 2", + "[59] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 2", + "[60] Yuhang Zang, Wei Li, Kaiyang Zhou, Chen Huang, and Chen Change Loy. Open-vocabulary detr with conditional matching. In European Conference on Computer Vision, pages 106-122. Springer, 2022. 2", + "[61] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-vocabulary object detection using captions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14393-14402, 2021. 2", + "[62] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. arXiv preprint arXiv:2111.07991, 2021. 2", + "[63] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In Proceedings of the" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16793-16803, 2022. 2", + "[64] Chong Zhou, Chen Change Loy, and Bo Dai. Denseclip: Extract free dense labels from clip. arXiv preprint arXiv:2112.01071, 2021. 2", + "[65] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Phillip Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. arXiv preprint arXiv:2201.02605, 2022. 1, 2", + "[66] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020.3, 6" + ], + "bbox": [ + 501, + 92, + 893, + 273 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "13968", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Generative Region-Language Pretraining for Open-Ended Object Detection/bb1e00ce-acc5-43f7-b817-49b140b640c0_model.json b/2024/Generative Region-Language Pretraining for Open-Ended Object Detection/bb1e00ce-acc5-43f7-b817-49b140b640c0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0c7bb67df25afb769526608c88e1c7be011fe393 --- /dev/null +++ b/2024/Generative Region-Language Pretraining for Open-Ended Object Detection/bb1e00ce-acc5-43f7-b817-49b140b640c0_model.json @@ -0,0 +1,2378 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.104, + 0.131, + 0.868, + 0.154 + ], + "angle": 0, + "content": "Generative Region-Language Pretraining for Open-Ended Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.165, + 0.179, + 0.276, + 0.199 + ], + "angle": 0, + "content": "Chuang Lin1" + }, + { + "type": "text", + "bbox": [ + 0.312, + 0.18, + 0.4, + 0.199 + ], + "angle": 0, + "content": "Yi Jiang*" + }, + { + "type": "text", + "bbox": [ + 0.429, + 0.181, + 0.526, + 0.199 + ], + "angle": 0, + "content": "Lizhen \\(\\mathbf{Q}\\mathbf{u}^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.564, + 0.181, + 0.686, + 0.198 + ], + "angle": 0, + "content": "Zehuan Yuan" + }, + { + "type": "text", + "bbox": [ + 0.723, + 0.181, + 0.83, + 0.198 + ], + "angle": 0, + "content": "Jianfei Cai*" + }, + { + "type": "text", + "bbox": [ + 0.321, + 0.199, + 0.486, + 0.217 + ], + "angle": 0, + "content": "1 Monash University" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.199, + 0.662, + 0.217 + ], + "angle": 0, + "content": "2 ByteDance Inc." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.252, + 0.313, + 0.267 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.283, + 0.474, + 0.751 + ], + "angle": 0, + "content": "In recent research, significant attention has been devoted to the open-vocabulary object detection task, aiming to generalize beyond the limited number of classes labeled during training and detect objects described by arbitrary category names at inference. Compared with conventional object detection, open vocabulary object detection largely extends the object detection categories. However, it relies on calculating the similarity between image regions and a set of arbitrary category names with a pretrained vision-and-language model. This implies that, despite its open-set nature, the task still needs the predefined object categories during the inference stage. This raises the question: What if we do not have exact knowledge of object categories during inference? In this paper, we call such a new setting as generative open-ended object detection, which is a more general and practical problem. To address it, we formulate object detection as a generative problem and propose a simple framework named GenerateU, which can detect dense objects and generate their names in a free-form way. Particularly, we employ Deformable DETR as a region proposal generator with a language model translating visual regions to object names. To assess the free-form object detection task, we introduce an evaluation method designed to quantitatively measure the performance of generative outcomes. Extensive experiments demonstrate strong zero-shot detection performance of our GenerateU. For example, on the LVIS dataset, our GenerateU achieves comparable results to the open-vocabulary object detection method GLIP, even though the category names are not seen by GenerateU during inference. Code is available at: https://github.com/FoundationVision/GenerateU." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.779, + 0.208, + 0.794 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.804, + 0.47, + 0.88 + ], + "angle": 0, + "content": "Object detection, a crucial task in computer vision, has made significant progress in recent years, particularly with the advent of deep learning [6, 8, 18, 41, 46]. The objective of this task is to locate objects in images with tight bounding boxes and recognize their respective categories. However," + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.254, + 0.604, + 0.266 + ], + "angle": 0, + "content": "Predefined category" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.266, + 0.837, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.51, + 0.352, + 0.716, + 0.364 + ], + "angle": 0, + "content": "(a) Open-Vocabulary Object Detection" + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.37, + 0.922, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.464, + 0.632, + 0.476 + ], + "angle": 0, + "content": "(b) Phrase Grounding" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.478, + 0.921, + 0.539 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.51, + 0.541, + 0.752, + 0.553 + ], + "angle": 0, + "content": "(c) Generative Open-Ended Object Detection" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.569, + 0.895, + 0.695 + ], + "angle": 0, + "content": "Figure 1. Comparing generative open-ended object detection with other open-set object detection tasks. Open-vocabulary object detection and phrase grounding typically require predefined categories or phrases in text prompts to align with image regions. In contrast, our introduced generative open-ended object detection is a more general and practical setting where categorical information is not explicitly defined. Such a setting is especially meaningful for scenarios where users lack precise knowledge of object categories during inference." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.746, + 0.894, + 0.838 + ], + "angle": 0, + "content": "most existing object detection algorithms are restricted to a predetermined set of object categories defined in detection datasets [9, 14, 17, 26, 42]. For example, an object detector trained on COCO [9] can well identify 80 pre-defined classes but encounters difficulty in recognizing new classes that have not been seen or defined during training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Open-vocabulary learning has been introduced to remedy this problem and considerable efforts have been made [7, 22, 28, 29, 35, 55, 65]. Open-vocabulary object detection addresses the closed-set limitation typically via" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.229, + 0.9 + ], + "angle": 0, + "content": "*Corresponding Authors" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13958" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.426 + ], + "angle": 0, + "content": "weak supervision from language, i.e., image-text pairs, or large pretrained Vision-Language Models (VLMs), such as CLIP [39]. The text embeddings from the text encoder of CLIP can well align with the visual regions for the novel categories or the noun phrases in image captions, as shown in Fig. 1. However, despite open-set nature, open-vocabulary object detection still requires predefined object categories during the inference stage, while in many practical scenarios, we might not have exact knowledge of test images. Even if we have prior knowledge of test images, manually defining object categories might introduce language ambiguities (e.g., similar object names like \"person\", \"a man\", \"young boy\" in Fig. 1) or be not comprehensive enough (e.g. missing \"chain link fence\" in Fig. 1), ultimately restricting flexibility. If we do not have prior knowledge of test images, to cover all possible objects that may appear in the images, the common solution to design a large and comprehensive label set that would be complex and time-consuming. The above discussions raise the research question: Can we do open-world dense object detection that does not require predefined object categories during inference? We call this problem open-ended object detection." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.432, + 0.473, + 0.645 + ], + "angle": 0, + "content": "Therefore, in this paper, we formulate open-ended object detection as a generative problem and propose a simple architecture termed GenerateU to address it. Our GenerateU contains two major components: a visual object detector to localize image regions and a language model to translate visual regions to object names. This model is trained end-to-end, optimizing both components simultaneously. To broaden the knowledge of GenerateU, we train it using a small set of human-annotated object-language paired data and scale up the vocabulary size with massive image-text pairs of data. Furthermore, due to the limitation in image captions whose nouns often fall short of accurately representing all objects in the images, we employ a pseudolabeling method to supplement missing objects in images." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.652, + 0.441, + 0.667 + ], + "angle": 0, + "content": "Our major contributions are summarized as follows." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.675, + 0.47, + 0.763 + ], + "angle": 0, + "content": "- We introduce a new and more practical object detection problem: open-ended object detection, and formulate it as a generative problem. The reformulation avoids manual effort in predefining object categories during both training and inference, aiming at a more general and flexible architecture." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.766, + 0.47, + 0.84 + ], + "angle": 0, + "content": "- We develop a novel end-to-end learning framework, termed GenerateU, which directly generates object names in a free-form manner. GenerateU is trained with a small set of human-annotated object-language paired data and massive image-text paired data." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.841, + 0.47, + 0.9 + ], + "angle": 0, + "content": "- Compared with open-vocabulary object detection models, GenerateU achieves comparable results on zero-shot LVIS, even though it does not see object categories during inference." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.675, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.642, + 0.107 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.115, + 0.807, + 0.132 + ], + "angle": 0, + "content": "2.1. Open-Vocabulary Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.14, + 0.895, + 0.518 + ], + "angle": 0, + "content": "Learning visual models from language supervision becomes popular in image recognition tasks [19, 20, 39, 40, 58, 59, 62, 64]. Recently, many works [23, 23-25, 34, 43-45, 48, 49, 52-54, 56, 57] have extended the success in open-vocabulary image recognition to object detection, i.e., open-vocabulary object detection (OVOD). As the pioneering work of OVOD, OVRCNN [61] successfully transfers the vision and language model trained on image-text pairs to the detection framework. To leverage the capability of CLIP [39], ViLD [16] and RegionCLIP [63] turn to learn the visual region feature from classification-oriented models by knowledge distillation. OV-DETR [60] proposes a novel open-vocabulary detector built upon DETR [8] by formulating the classification as a binary matching between the input queries and the referring objects. Different from distilling frozen vision encoder, OWL [35] proposes a two-step framework for contrastive image-text pre-training and end-to-end detection fine-tuning, using a standard Vision Transformer [13]. Detic [65] proposes to extend the detector vocabulary with large-scale image classification data by applying supervision to the proposal with the largest spatial size. To reduce the annotation cost for open-vocabulary object detection, VLDet [29] proposes a novel framework to directly learn from image-text paired data by formulating object-language alignment as a set matching problem." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.518, + 0.895, + 0.655 + ], + "angle": 0, + "content": "Open vocabulary object detection aims at generalization to novel categories that were not seen during the training phase. Existing OVOD works typically rely on calculating the similarity between image regions and a set of arbitrary category names with a pretrained vision-language model. It remains necessary to predefine the categories during inference. In contrast, this work delves deeper into a new challenge - how to effectively handle scenarios where the exact categories remain unknown during the inference process." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.666, + 0.813, + 0.683 + ], + "angle": 0, + "content": "2.2. Multimodal Large Language Model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Inspired by the remarkable capability of LLM [1, 4, 47, 50], multimodal foundation models [2, 15, 27, 30, 37, 38] leverage LLM as the core intelligence and are trained with extensive image-text pairs, resulting in enhanced performance across various vision-and-language tasks such as VQA and image captioning. BLIP-2 [27] proposes a generic and efficient pre-training strategy that bootstraps vision-language pre-training from off-the-shelf frozen pre-trained image encoders and frozen large language models. By inserting adapters into LLaMA's transformer, LLaMA-Adapter [15] introduces only 1.2M learnable parameters, and turns a LLaMA into an instruction-following model. LLaVA [30] represents a novel end-to-end trained large multimodal model that combines a vision encoder and Vicuna [36]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13959" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.136 + ], + "angle": 0, + "content": "for general-purpose visual and language understanding. It achieves impressive chat capabilities, emulating the essence of the multimodal GPT-4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.137, + 0.471, + 0.394 + ], + "angle": 0, + "content": "Based on this amazing progress, subsequent studies further explore enhancing the object localization capabilities of multimodal models. Kosmos-2 [37] unlocks the grounding capability, enabling the model to provide visual answers such as bounding boxes, thereby supporting a wider array of vision-language tasks, including referring expression comprehension. UNINEXT[55] reformulates diverse instance perception tasks into a unified object discovery and retrieval paradigm and can flexibly perceive different types of objects by simply changing the input prompts. DetGPT [38] locates the objects of interest based on the language instructions and the LLM enables users to interact with the system, allowing for a higher level of interactivity like visual reasoning. In contrast to these works that focus on understanding natural language expressions referring to specific objects in an image, our work aims to identify and localize all objects of interest in an image, associating each with a class name." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.402, + 0.251, + 0.419 + ], + "angle": 0, + "content": "2.3. Dense Captioning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.426, + 0.47, + 0.636 + ], + "angle": 0, + "content": "Dense captioning aims to predict a set of descriptions across various image regions. DenseCap [21] develops a fully convolutional localization network to extract image regions, which are then described using an RNN language model. In dense captioning, it treats each region as an individual image and tries to generate captions encompassing multiple objects and their attribute, such as \"cat riding a skateboard\" and \"bear is wearing a red hat\". To evaluate the generated captions, dense captioning employs the METEOR [21] metric, originally designed for machine translation, which assesses the generation results from the sentence perspective. In comparison, our proposed generative open-ended object detection focuses on describing individual objects as category names in a zero-shot setting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.637, + 0.47, + 0.743 + ], + "angle": 0, + "content": "Note that recently, CapDet [32] and GRiT [51] attempted to integrate open-vocabulary object detection and dense captioning by introducing an additional language model as a captioning head. However, for open-world object detection, they either still require a predefined category space during inference, or are limited to the close-set object detection problem." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.755, + 0.168, + 0.77 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Our goal is to build an open-world object detector capable of localizing all objects in an image and providing their corresponding category names in a free-form way. Figure 2 gives an overview of our proposed open-ended object detection model, GenerateU, which comprises two major components: an open-world object detector and a language model. We describe the open-world object detector in Sec. 3.1, which consists of an image encoder and a detection head," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "and the language model, which can be directly borrowed from an existing Multimodal Large Language Model (including a V-to-L adaptor and a language model, Sec. 3.2) or retrained (Sec. 3.3). An additional pseudo-labeling method is described in Sec. 3.4 for enriching the label diversity." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.178, + 0.767, + 0.194 + ], + "angle": 0, + "content": "3.1. Open-World Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.201, + 0.892, + 0.519 + ], + "angle": 0, + "content": "The first crucial step is to extract accurate object regions from images, for which we develop an open-world object detector. Although we are not constrained to any specific object detection model, we choose Deformable DETR [66] for the following reasons. First, it alleviates heuristics when compared to other popular one or two-stage frameworks, and provides a versatile query-to-instance pipeline. Besides, Deformable DETR [66] demonstrates faster convergence speed and enhanced accuracy compared to the original DETR [8]. Deformable DETR or the original one employs Hungarian matching to learn a mapping from predicted queries to ground truth objects, followed by training the matched object query to regress to its corresponding ground truth with a combination of classification loss and bounding box regression loss. In the context of open-set problems, we do not rely on the object class information. Instead, we adopt an open-world detection approach, i.e. class-agnostic object detector, wherein matched queries are only classified into foreground or background. Consequently, the detection process involves three losses: binary cross-entropy, generalized IoU, and L1 regression loss." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.53, + 0.889, + 0.546 + ], + "angle": 0, + "content": "3.2. Transferring from a Frozen Multimodal LLM" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.854 + ], + "angle": 0, + "content": "Once object candidates are localized, a natural idea is to leverage a pretrained multimodal large language model (MLLM) to reduce the training cost and offer strong zero-shot language abilities for generating object names. Recently developed MLLMs [2, 15, 27, 30, 37, 38] typically consist of three components: an image encoder to extract visual features; an adaptor network to bridge the modality gap and reduce the number of image embeddings, converting visual feature into token representations; a language model to generate predictions for vision-centric tasks. As shown in Figure 2 Top, we integrate a class-agnostic Deformable DETR, with a frozen image encoder to extract object queries, into an MLLM. We keep the MLLM frozen, and the only trainable module is the detection head for learning object representations. Besides, images are preprocessed to a resolution of \\(224 \\times 224\\) pixels during training in KOSMOS-2 [37]. In object detection, it is common to employ a larger input size for multi-scale feature extraction. To handle changes in image size, we utilize linear interpolation to enlarge the original position embedding." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Despite its simplicity, our empirical study shows that such a direct open-end object detection based on pretrained MLLM results in poor performance large due to the do" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13960" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.088, + 0.892, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.17, + 0.273, + 0.183 + ], + "angle": 0, + "content": "(a) Training with frozen MLLM" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.19, + 0.235, + 0.203 + ], + "angle": 0, + "content": "(b) Training End-to-End" + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.206, + 0.75, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.303, + 0.895, + 0.376 + ], + "angle": 0, + "content": "Figure 2. Overview of our proposed open-ended object detection model, GenerateU, which comprises two major components: an object detector and a language model. We compare two training strategies: (Top) We incorporate the class-agnostic DETR (with frozen image encoder) into a pre-trained and frozen Multimodal Large Language Model (including Adaptor and Language Model), to facilitate a smooth transfer of knowledge from the language domain to object detection; (Bottom) We activate the image encoder and language model as trainable components, taking an end-to-end approach to seamlessly integrate region-level understanding into the language model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.4, + 0.471, + 0.462 + ], + "angle": 0, + "content": "main shift. MLLMs were trained on large-scale image-text paired data, where they generate text descriptions given each whole image as a condition, without a fine-grained understanding of individual objects." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.47, + 0.433, + 0.487 + ], + "angle": 0, + "content": "3.3. Generative Region-Language Pretraining" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.493, + 0.47, + 0.825 + ], + "angle": 0, + "content": "Thus, instead of employing a frozen MLLM, we propose GenerateU to directly link the open-world object detector with a language model, and activate both the image encoder and the language model as trainable components, as shown in Figure 2. Particularly, we employ an encoder-decoder-based language model, where visual representation serves as input for the encoder, and the associated text serves as the generation target for the language decoder. Here we only use the object queries that have been successfully matched with ground truth through Hungarian matching, ignoring the unmatched queries in the generating process. Specifically, each matched object query is first treated as an individual region, fed to a projection function to align its dimensions with the language model. The encoder comprises blocks with self-attention layers and feed-forward networks. The decoder, similar in structure, adds a cross-attention mechanism for interfacing with the encoder. The decoder employs an autoregressive self-attention mechanism, limiting the model's attention to past outputs, and its final output is processed through a dense layer with a softmax function. We apply the language modeling loss to train the model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Exploring on the region-word alignment loss. Benefiting from the pretrained vision-language model, we incorporate a region-word alignment loss during training to aid in learning to distinguish region features. In detail, for an object query, we treat its corresponding word as a positive sample" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.4, + 0.895, + 0.507 + ], + "angle": 0, + "content": "and consider other words in the same minibatch as negative samples. By calculating the similarity between the region features from class-agnostic Deformable DETR and text features from the fixed CLIP text encoder [39], the model is optimized by a binary cross-entropy loss. This region-word alignment loss is jointly trained with the object detection losses and the language modeling loss." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.538, + 0.713, + 0.556 + ], + "angle": 0, + "content": "3.4. Enrich Label Diversity" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Considerable efforts have been invested in collecting detection data that is both semantically rich and voluminous. However, human annotation proves to be both costly and resource-restrictive. Previous works, such as Kosmos2 [37] and GLIP [28], leverage image-text pairs to expand the vocabulary scale. They use pretrained grounding models to generate bounding boxes aligned with noun phrases from captions. These boxes, in turn, serve as pseudo detection labels for training the model. However, given that captions may not comprehensively describe every object in an image, the count of nouns in the caption is typically much lower than the actual number of objects present in the corresponding image. Consequently, the quantity of pseudolabels generated is constrained. To address this, we use GenerateU pretrained on the available labels to generate pseudo labels as a supplement for missing objects in images. Furthermore, we observed that beam search in language models naturally generates synonyms, thereby providing diverse object labels. Figure 3 provides pseudolabel examples, illustrating the generation of boxes covering nearly all objects in the image, along with a diverse set of rich vocabulary labels." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13961" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.171, + 0.089, + 0.346, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.168, + 0.218, + 0.328, + 0.236 + ], + "angle": 0, + "content": "\"Belcher family you and me we got this poster A1.\"" + }, + { + "type": "image", + "bbox": [ + 0.379, + 0.089, + 0.507, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.358, + 0.218, + 0.547, + 0.236 + ], + "angle": 0, + "content": "\"A gorgeous wedding centrepiece with feathers, moss, greenery, white blooms and greenery.\"" + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.089, + 0.791, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.559, + 0.218, + 0.796, + 0.245 + ], + "angle": 0, + "content": "\"Just before Christmas, Thornton's issued a profit warning after suffering low order levels from supermarket customers and problems at its new warehouse.\"" + }, + { + "type": "image", + "bbox": [ + 0.169, + 0.246, + 0.365, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.175, + 0.36, + 0.306, + 0.369 + ], + "angle": 0, + "content": "\"Boy in red shirt catching a ball.\"" + }, + { + "type": "image", + "bbox": [ + 0.391, + 0.246, + 0.512, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.369, + 0.36, + 0.541, + 0.377 + ], + "angle": 0, + "content": "\"Adams, seen here celebrating after scoring against Chesterfield.\"" + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.246, + 0.791, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.36, + 0.796, + 0.377 + ], + "angle": 0, + "content": "\"Blonde woman with a colourful umbrella being blown away by the wind with her umbrella snowy landscape.\"" + }, + { + "type": "image", + "bbox": [ + 0.169, + 0.378, + 0.308, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.172, + 0.513, + 0.302, + 0.539 + ], + "angle": 0, + "content": "\"Meatballs on top of spaghetti squash on a white plate with rest behind it in skillet.\"" + }, + { + "type": "image", + "bbox": [ + 0.366, + 0.378, + 0.495, + 0.511 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.311, + 0.512, + 0.542, + 0.54 + ], + "angle": 0, + "content": "\"How to make fresh brewed iced tea sweetened or without sugar. Homemade recipe using tea bags with a squeeze of lemon or orange juice and a sprig of mint.\"" + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.378, + 0.791, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.56, + 0.513, + 0.785, + 0.531 + ], + "angle": 0, + "content": "\"Close up picture of a plate of tuna noodle casserole with a bowl of salad in the background.\"" + }, + { + "type": "image", + "bbox": [ + 0.169, + 0.541, + 0.392, + 0.654 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.172, + 0.655, + 0.381, + 0.672 + ], + "angle": 0, + "content": "\"One of a kind: meet the only BMW E3 Estate in the world.\"" + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.541, + 0.598, + 0.654 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.413, + 0.655, + 0.603, + 0.673 + ], + "angle": 0, + "content": "\"The last time with his S15, Matt was always in the need of more power.\"" + }, + { + "type": "image", + "bbox": [ + 0.623, + 0.542, + 0.79, + 0.654 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.624, + 0.655, + 0.784, + 0.673 + ], + "angle": 0, + "content": "\"Construction Junction box truck parked in front of the loading dock.\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.69, + 0.894, + 0.732 + ], + "angle": 0, + "content": "Figure 3. Selected pseudo-label examples highlight the generation of bounding boxes covering nearly all objects in the images. A varied set of descriptive labels showcases the model's capability in producing diverse and linguistically rich vocabulary. A white underline in a figure indicates that the pseudo label comes from the noun phrase (black underline) in the caption." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.757, + 0.208, + 0.774 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.785, + 0.179, + 0.799 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Our model is trained with hybrid supervision from two types of data: detection data and image-text data. For detection data, we use Visual Genome (VG) [26] which contains 77,398 images for training. Different from previous works [32, 51], we do not use the region captioning annotations of VG, since they may contain descriptions that" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.758, + 0.892, + 0.88 + ], + "angle": 0, + "content": "encompass multiple objects or even describe actions. Our focus is directed towards the object detection task, and thus we leverage the object annotations of VG which annotate the same bounding box regions with concise and specific object names, covering a wide range of objects. For pretraining, we use about 5 million image-text pairs GRIT [37] from COYO700M [5] with pseudo grounding labels generated by KOSMOS-2 [37] using GLIP [28]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13962" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.078, + 0.089, + 0.895, + 0.27 + ], + "angle": 0, + "content": "
MethodBackbonePre-Train DataOpen-endedOpen-setLVIS
APrAPcAPfAP
MDETR [22]RN101LVIS--20.924.924.324.2
MaskRCNN [18]RN101LVIS--26.334.033.933.3
Deformable DETR [66]Swin-TLVIS--24.236.038.236.0
GLIP (C) [28]Swin-TO365, GoldG-✓17.719.531.024.9
GLIP [28]Swin-TO365, GoldG, CAP4M-✓20.821.431.026.0
GenerateUSwin-TVG✓✓17.422.429.625.4
GenerateUSwin-TVG, GRIT5M✓✓20.024.929.826.8
GenerateUSwin-LVG, GRIT5M✓✓22.325.231.427.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.279, + 0.893, + 0.31 + ], + "angle": 0, + "content": "Table 1. Zero-shot domain transfer to LVIS. Our method achieves comparable performance to prior models in the close-ended setting, without requiring access to category names during inference." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.334, + 0.362, + 0.35 + ], + "angle": 0, + "content": "4.2. Evaluation protocol and metrics" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.357, + 0.473, + 0.703 + ], + "angle": 0, + "content": "Addressing an open-ended problem presents challenges in evaluating performance. For instance, when annotating a laptop, the human-annotated ground truth category may be \"laptop\", \"computer\", \"PC\", \"Microcomputer\", etc. To assess the effectiveness of the open-end generative results, we propose two settings: (a) We calculate the similarity score between the generated and annotated category names with a fixed pretrained text encoder. Notice that this text encoder is not utilized by the generative detection model; rather, it is only used for quantitative performance measurement. Any text encoder capable of computing similarity between two sentences can be adopted for this purpose. In this study, we select the popular CLIP [39] text encoder and BERTlarge [12] model. In this way, we can map the generative results to predefined categories in a dataset for evaluation. (b) We use METEOR [3] to assess the generated text quality, which is the same as the evaluation in dense captioning [21]. METEOR is widely used in NLP for evaluating machine translations. Following [21], we measure the mean Average Precision (AP) for both localization, using Intersection over Union (IoU) thresholds of .3, .4, .5, .6, and .7, and language accuracy, applying METEOR score thresholds at 0, .05, .1, .15, .2, and .25." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.47, + 0.812 + ], + "angle": 0, + "content": "Following GLIP [28], we evaluate our GenerateU primarily on LVIS [17] which contains 1203 categories and report APs for rare categories \\((AP_r)\\), common categories \\((AP_c)\\), frequency categories \\((AP_f)\\) and all categories \\((mAP)\\), respectively. Specifically, following GLIP [28] and MDETR [22], we evaluate on the \\(5\\mathrm{k}\\) minival subset and report the zero-shot fixed AP [11] for a fair comparison." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.818, + 0.295, + 0.833 + ], + "angle": 0, + "content": "4.3. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.903 + ], + "angle": 0, + "content": "We attempt two different backbones, Swin-Tiny and Swin-Large [31] as the visual encoder. The Deformable DETR architecture follows [66] with 6 encoder layers and 6 decoder layers. The number of object queries \\(N\\) is set to 300." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.334, + 0.895, + 0.487 + ], + "angle": 0, + "content": "We adopt FlanT5-base [10] as the language model and initialize the weights from it. The optimizer is AdamW [33] with a weight decay of 0.05. The learning rate is set to \\(2 \\times 10^{-4}\\) for the parameters of the visual backbone and the detection head, and \\(3 \\times 10^{-4}\\) for the language model. For the frozen MLLM approach in Sec 3.2, we explore the state-of-the-art model KOSMOS-2 [37] which was trained with the multimodal corpora. We fix all parameters and set the learning rate to \\(2 \\times 10^{-4}\\) for the detection head. Our model is trained on 16 A100 GPUs." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.498, + 0.862, + 0.515 + ], + "angle": 0, + "content": "4.4. Generative Open-Ended Detection Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.895, + 0.869 + ], + "angle": 0, + "content": "We assess the model's capacity to identify diverse objects in a zero-shot setting on LVIS, presenting the results in Table 1, where rows 1-3 include methods utilizing LVIS as pre-training data, representing fully supervised models. Remarkably, we observe that even with only VG as training data, our model demonstrates commendable performance on zero-shot LVIS. This result suggests that predefined class names might not be necessary for open-world object detection during inference, especially when models are trained with extensive visual concepts. Furthermore, introducing additional training data from cost-effective image-text pairs consistently improves performance. For instance, when GenerateU trained with additional image-text pair data (GRIT [37]), we observe a substantial performance gain of 3.6 on \\(AP_{r}\\) and 1.4 on \\(AP\\). In summary, the semantic richness derived from image-text data significantly enhances the model's ability to recognize rare objects. Notably, our model achieves comparable performance to GLIP under the same training data size but without requiring access to category names during inference. Moreover, we enhanced our model by integrating a larger backbone, i.e. Swin-Large, confirming that our method can scale effectively with increased model complexity." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Transfer to other datasets. Our generative open-ended object detector can transfer to any dataset without any mod-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13963" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.089, + 0.891, + 0.203 + ], + "angle": 0, + "content": "
MethodsImage EncoderDetection HeadLanguage ModelLVIS
APrAPcAPfAP
Init from MLLMFreezeTrainFreeze*5.310.219.714.3
TrainTrainFreeze*10.917.228.121.9
Init from LLMTrainTrainFreeze11.220.127.422.9
TrainTrainTrain20.024.929.826.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.213, + 0.893, + 0.242 + ], + "angle": 0, + "content": "Table 2. Comparison of our GenerateU with two different training strategies on a pretrained LLM on zero-shot LVIS. The results highlight generative object detection benefits from end-to-end training. Here, \\(\\star\\) refers to freezing both V-to-L adaptor and language model." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.265, + 0.47, + 0.342 + ], + "angle": 0, + "content": "
MethodsPre-Train DataCOCOObjects365
Supervised-46.525.6
GenerateUVG33.010.1
GenerateUVG, GRIT33.610.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.353, + 0.47, + 0.396 + ], + "angle": 0, + "content": "Table 3. We evaluate our GenerateU on COCO and Objects365 validation set in a zero-shot setting. GenerateU can transfer to various downstream datasets without requiring any modifications." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.409, + 0.465, + 0.485 + ], + "angle": 0, + "content": "
MetricsAPrAPcAPfAP
CLIP [39]20.024.929.826.8
BERT [12]12.519.925.121.7
METEOR [3]---9.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.496, + 0.467, + 0.51 + ], + "angle": 0, + "content": "Table 4. Effect of different evaluation metrics on zero-shot LVIS." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.536, + 0.469, + 0.643 + ], + "angle": 0, + "content": "ifications. In contrast to open vocabulary object detection methods, we do not need to switch the classifier to the specific category text embeddings. We evaluate the transferability of the proposed GenerateU on COCO [9] and Objects365 [42] validation sets, as shown in Table 3. The results demonstrate the effectiveness of our method and underscore the importance of scaling up the training data." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.651, + 0.231, + 0.667 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.674, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Comparison with transferring from a Frozen MLLM. In Table 2, we evaluate the approach discussed in Sec 3.2, i.e. transferring a frozen MLLM to the generative object detection task. Following the state-of-the-art MLLM model KOSMOS-2 [37], we maintain its parameters fixed and only train the detection head. However, we observe that such direct transfer of the MLLM to object detection leads to poor performance in zero-shot LVIS. We attribute this result to the pretrained image encoder, which is trained on images of size \\(224 \\times 224\\), which is insufficient for the object detection task. Besides, the MLLM originally supervised by image-level annotations such as image captions, lacks understanding in both object localization and fine-grained object categories. We also experiment with making the image encoder trainable while keeping the language model frozen. This" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.269, + 0.885, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.433, + 0.893, + 0.476 + ], + "angle": 0, + "content": "Figure 4. Effect of beam size. Beam search plays a crucial role in generating rare object names, effectively addressing the long-tail problem." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.505, + 0.892, + 0.595 + ], + "angle": 0, + "content": "adjustment significantly improves the results. We further fine-tune the entire model, initializing the Large Language Model (LLM) from FlanT5-base [10], which achieves the best performance. The above analysis underscores the importance of end-to-end training in the open-ended object detection task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.893, + 0.822 + ], + "angle": 0, + "content": "Beam Search. Towards a general and practical object detection system, we believe that beam search plays a crucial role in generating outputs that are closely aligned with human intuition. This is because it explores a broader array of possibilities during the decoding phase. Beam search naturally contains the multi-level object names and their synonyms, enhancing the system's ability to recognize objects with varied linguistic expressions. This contributes to a more adaptable and nuanced understanding of the objects being detected. As shown in Figure 4, beam search largely improves the rare category performance, proving its importance for the long tail problem, and emphasizing the system's effectiveness in handling less frequent object categories. In our experiments, we set the default beam size to 3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Evaluation Metrics. To evaluate the results of generative object detection, we use different metrics: METEOR, evaluating the quality of text, and similarity scores, calculated by a fixed text encoder. Note that the off-the-shelf text encoder is not used for the proposed method, but just for the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13964" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.131, + 0.089, + 0.85, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.446, + 0.893, + 0.476 + ], + "angle": 0, + "content": "Figure 5. Qualitative prediction results from GenerateU and ground truth on LVIS [17]. GenerateU produces complete and precise predictions, showcasing its ability to go beyond fixed vocabulary constraints." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.498, + 0.487, + 0.613 + ], + "angle": 0, + "content": "
Pre-Train DataRegion-word AlignmentAPrAPcAPfAP
VG-16.019.126.722.5
✓17.4↑1.422.4↑3.929.6↑2.925.4↑2.9
VG, GRIT-19.223.528.925.7
✓20.0↑0.824.9↑1.429.8↑0.926.8↑0.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.136, + 0.622, + 0.411, + 0.637 + ], + "angle": 0, + "content": "Table 5. Effect of region-word alignment loss." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.671, + 0.47, + 0.762 + ], + "angle": 0, + "content": "evaluation. Here we compare two offline models: CLIP text encoder [39] and BERT [12], as presented in Table 4. We find that the CLIP text encoder is particularly adept at handling noun phrases, while the BERT encoder, predominantly trained on sentence corpora, may exhibit different strengths in processing textual information." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Region-word Alignment. We conduct ablation experiments to evaluate the effectiveness of the region-word alignment loss in our method, as presented in Table 5. We observe that training with only the detection loss and the language modeling loss yields promising results. Adding the region-word alignment loss provides additional supervision and further improves the performance. This further demonstrates the effectiveness of the proposed generative object detection method." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.5, + 0.683, + 0.516 + ], + "angle": 0, + "content": "4.6. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.534, + 0.892, + 0.656 + ], + "angle": 0, + "content": "Here we provide qualitative visualizations in showcasing the superior object detection capabilities of our model. As shown in Figure 5, our model can identify a more diverse range of objects compared to ground truth annotations. These visualizations not only affirm the reliability of our detection framework but also demonstrate its ability to detect a wider variety of objects, providing valuable insights into the model's strengths." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.699, + 0.619, + 0.715 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.733, + 0.892, + 0.889 + ], + "angle": 0, + "content": "This paper introduces GenerateU, a novel approach for generative open-ended object detection. By reformulating object detection as a generative task, the model eliminates the need for predefined categories during inference, offering a more practical solution. Our method achieves comparable results with open-vocabulary models on zero-shot LVIS, even though categories are not seen in our GenerateU. Overall, this work contributes to the fast-evolving area of object detection towards open-world scenarios. Future works include investigating the effect of training data scales and exploring methods beyond naive pseudo-labeling." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13965" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Rohan Anil, Andrew M Dai, Orhan First, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, et al. Palm 2 technical report. arXiv preprint arXiv:2305.10403, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.24 + ], + "angle": 0, + "content": "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.244, + 0.471, + 0.313 + ], + "angle": 0, + "content": "[3] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.315, + 0.471, + 0.384 + ], + "angle": 0, + "content": "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.387, + 0.469, + 0.442 + ], + "angle": 0, + "content": "[5] Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, and Saehoon Kim. Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset, 2022.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.444, + 0.469, + 0.499 + ], + "angle": 0, + "content": "[6] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6154-6162, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.501, + 0.469, + 0.557 + ], + "angle": 0, + "content": "[7] Zhaowei Cai, Gukyeong Kwon, Avinash Ravichandran, Erhan Bas, Zhuowen Tu, Rahul Bhotika, and Stefano Soatto. X-detr: A versatile architecture for instance-wise vision-language tasks. arXiv preprint arXiv:2204.05626, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.559, + 0.469, + 0.626 + ], + "angle": 0, + "content": "[8] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.63, + 0.469, + 0.685 + ], + "angle": 0, + "content": "[9] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.687, + 0.469, + 0.755 + ], + "angle": 0, + "content": "[10] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.758, + 0.469, + 0.813 + ], + "angle": 0, + "content": "[11] Achal Dave, Piotr Dollar, Deva Ramanan, Alexander Kirilov, and Ross Girshick. Evaluating large-vocabulary object detectors: The devil is in the details. arXiv preprint arXiv:2102.01066, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[12] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.151, + 0.892, + 0.206 + ], + "angle": 0, + "content": "[14] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303-338, 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.209, + 0.892, + 0.277 + ], + "angle": 0, + "content": "[15] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xi-angyu Yue, et al. Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.281, + 0.892, + 0.335 + ], + "angle": 0, + "content": "[16] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.339, + 0.892, + 0.395 + ], + "angle": 0, + "content": "[17] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 1, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.398, + 0.892, + 0.451 + ], + "angle": 0, + "content": "[18] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE International Conference on Computer Vision, pages 2961-2969, 2017. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.455, + 0.892, + 0.51 + ], + "angle": 0, + "content": "[19] Dat Huynh, Jason Kuen, Zhe Lin, Jiumiang Gu, and Ehsan Elhamifar. Open-vocabulary instance segmentation via robust cross-modal pseudo-labeling. arXiv preprint arXiv:2111.12698, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.513, + 0.892, + 0.595 + ], + "angle": 0, + "content": "[20] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.6, + 0.892, + 0.655 + ], + "angle": 0, + "content": "[21] Justin Johnson, Andrej Karpathy, and Li Fei-Fei. Densecap: Fully convolutional localization networks for dense captioning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4565–4574, 2016. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.657, + 0.892, + 0.727 + ], + "angle": 0, + "content": "[22] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. MDetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1780-1790, 2021. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.73, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[23] Prannay Kaul, Weidi Xie, and Andrew Zisserman. Multimodal classifiers for open-vocabulary object detection. arXiv preprint arXiv:2306.05493, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[24] Dahun Kim, Anelia Angelova, and Weicheng Kuo. Contrastive feature masking open-vocabulary vision transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15602-15612, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[25] Dahun Kim, Anelia Angelova, and Weicheng Kuo. Region-aware pretraining for open-vocabulary object detection with vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11144-11154, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13966" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.175 + ], + "angle": 0, + "content": "[26] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123(1):32-73, 2017. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.176, + 0.47, + 0.232 + ], + "angle": 0, + "content": "[27] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.471, + 0.3 + ], + "angle": 0, + "content": "[28] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. arXiv preprint arXiv:2112.03857, 2021. 1, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.301, + 0.471, + 0.357 + ], + "angle": 0, + "content": "[29] Chuang Lin, Peize Sun, Yi Jiang, Ping Luo, Lizhen Qu, Gholamreza Haffari, Zehuan Yuan, and Jianfei Cai. Learning object-language alignments for open-vocabulary object detection. arXiv preprint arXiv:2211.14843, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.358, + 0.47, + 0.398 + ], + "angle": 0, + "content": "[30] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.471, + 0.468 + ], + "angle": 0, + "content": "[31] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.47, + 0.55 + ], + "angle": 0, + "content": "[32] Yanxin Long, Youpeng Wen, Jianhua Han, Hang Xu, Pengzhen Ren, Wei Zhang, Shen Zhao, and Xiaodan Liang. Capdet: Unifying dense captioning and open-world detection pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15233-15243, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.552, + 0.469, + 0.58 + ], + "angle": 0, + "content": "[33] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.47, + 0.635 + ], + "angle": 0, + "content": "[34] Chuofan Ma, Yi Jiang, Xin Wen, Zehuan Yuan, and Xiaojuan Qi. Codet: Co-occurrence guided region-word alignment for open-vocabulary object detection. arXiv preprint arXiv:2310.16667, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.636, + 0.47, + 0.718 + ], + "angle": 0, + "content": "[35] Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, et al. Simple open-vocabulary object detection with vision transformers. arXiv preprint arXiv:2205.06230, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.47, + 0.761 + ], + "angle": 0, + "content": "[36] Baolin Peng, Chunyuan Li, Pengcheng He, Michel Galley, and Jianfeng Gao. Instruction tuning with gpt-4. arXiv preprint arXiv:2304.03277, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.47, + 0.817 + ], + "angle": 0, + "content": "[37] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.47, + 0.872 + ], + "angle": 0, + "content": "[38] Renjie Pi, Jiahui Gao, Shizhe Diao, Rui Pan, Hanze Dong, Jipeng Zhang, Lewei Yao, Jianhua Han, Hang Xu, and Lingpeng Kong Tong Zhang. Detgpt: Detect what you need via reasoning. arXiv preprint arXiv:2305.14167, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.216 + ], + "angle": 0, + "content": "[40] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. arXiv preprint arXiv:2112.01518, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.218, + 0.892, + 0.273 + ], + "angle": 0, + "content": "[41] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In Advances in Neural Information Processing Systems, pages 91-99, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.274, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[42] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.344, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[43] Cheng Shi and Sibei Yang. Edadet: Open-vocabulary object detection using early dense alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15724-15734, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.4, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[44] Hengcan Shi, Munawar Hayat, and Jianfei Cai. Open-vocabulary object detection via scene graph discovery. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4012-4021, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.456, + 0.892, + 0.495 + ], + "angle": 0, + "content": "[45] Hengcan Shi, Munawar Hayat, and Jianfei Cai. Unified open-vocabulary dense visual prediction. arXiv preprint arXiv:2307.08238, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.497, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[46] Peize Sun, Rufeng Zhang, Yi Jiang, Tao Kong, Chenfeng Xu, Wei Zhan, Masayoshi Tomizuka, Lei Li, Zehuan Yuan, Changhu Wang, et al. Sparse r-cnn: End-to-end object detection with learnable proposals. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14454-14463, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.58, + 0.892, + 0.65 + ], + "angle": 0, + "content": "[47] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.65, + 0.892, + 0.719 + ], + "angle": 0, + "content": "[48] Luting Wang, Yi Liu, Penghui Du, Zihan Ding, Yue Liao, Qiaosong Qi, Biaolong Chen, and Si Liu. Object-aware distillation pyramid for open-vocabulary object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11186-11196, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.72, + 0.892, + 0.774 + ], + "angle": 0, + "content": "[49] Tao Wang. Learning to detect and segment for open vocabulary object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7051-7060, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.775, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[50] BigScience Workshop, Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ilic, Daniel Hesslow, Roman Castagné, Alexandra Sasha Luccioni, François Yvon, et al. Bloom: A 176b-parameter open-access multilingual language model. arXiv preprint arXiv:2211.05100, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[51] Jialian Wu, Jianfeng Wang, Zhengyuan Yang, Zhe Gan, Zicheng Liu, Junsong Yuan, and Lijuan Wang. Grit: A generative region-to-text transformer for object understanding. arXiv preprint arXiv:2212.00280, 2022. 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13967" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[52] Jianzong Wu, Xiangtai Li, Henghui Ding, Xia Li, Guangliang Cheng, Yunhai Tong, and Chen Change Loy. Betrayed by captions: Joint caption grounding and generation for open vocabulary instance segmentation. arXiv preprint arXiv:2301.00805, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.164, + 0.472, + 0.234 + ], + "angle": 0, + "content": "[53] Size Wu, Wenwei Zhang, Sheng Jin, Wentao Liu, and Chen Change Loy. Aligning bag of regions for open-vocabulary object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15254-15264, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.471, + 0.303 + ], + "angle": 0, + "content": "[54] Xiaoshi Wu, Feng Zhu, Rui Zhao, and Hongsheng Li. Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7031-7040, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.472, + 0.375 + ], + "angle": 0, + "content": "[55] Bin Yan, Yi Jiang, Jiannan Wu, Dong Wang, Ping Luo, Zehuan Yuan, and Huchuan Lu. Universal instance perception as object discovery and retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15325-15336, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.47, + 0.446 + ], + "angle": 0, + "content": "[56] Lewei Yao, Jianhua Han, Youpeng Wen, Xiaodan Liang, Dan Xu, Wei Zhang, Zhenguo Li, Chunjing Xu, and Hang Xu. Detclip: Dictionary-enriched visual-concept paralleled pretraining for open-world detection. Advances in Neural Information Processing Systems, 35:9125-9138, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.447, + 0.47, + 0.53 + ], + "angle": 0, + "content": "[57] Lewei Yao, Jianhua Han, Xiaodan Liang, Dan Xu, Wei Zhang, Zhenguo Li, and Hang Xu. Detclipv2: Scalable open-vocabulary object detection pre-training via word-region alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23497-23506, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.532, + 0.47, + 0.587 + ], + "angle": 0, + "content": "[58] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.589, + 0.47, + 0.658 + ], + "angle": 0, + "content": "[59] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.47, + 0.716 + ], + "angle": 0, + "content": "[60] Yuhang Zang, Wei Li, Kaiyang Zhou, Chen Huang, and Chen Change Loy. Open-vocabulary detr with conditional matching. In European Conference on Computer Vision, pages 106-122. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.47, + 0.785 + ], + "angle": 0, + "content": "[61] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-vocabulary object detection using captions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14393-14402, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[62] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. arXiv preprint arXiv:2111.07991, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[63] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16793-16803, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.162 + ], + "angle": 0, + "content": "[64] Chong Zhou, Chen Change Loy, and Bo Dai. Denseclip: Extract free dense labels from clip. arXiv preprint arXiv:2112.01071, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.894, + 0.219 + ], + "angle": 0, + "content": "[65] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Phillip Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. arXiv preprint arXiv:2201.02605, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.894, + 0.275 + ], + "angle": 0, + "content": "[66] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020.3, 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13968" + } + ] +] \ No newline at end of file diff --git a/2024/Generative Region-Language Pretraining for Open-Ended Object Detection/bb1e00ce-acc5-43f7-b817-49b140b640c0_origin.pdf b/2024/Generative Region-Language Pretraining for Open-Ended Object Detection/bb1e00ce-acc5-43f7-b817-49b140b640c0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f97d499eb3eec1884922795aae3ea20b63e22f2f --- /dev/null +++ b/2024/Generative Region-Language Pretraining for Open-Ended Object Detection/bb1e00ce-acc5-43f7-b817-49b140b640c0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7096b1b04453537977bfc6005cc0b6b3383aa142c758bb37dccf18f98537b906 +size 4619937 diff --git a/2024/Generative Region-Language Pretraining for Open-Ended Object Detection/full.md b/2024/Generative Region-Language Pretraining for Open-Ended Object Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d3a15cfe8ef26dd328efc2f6394392110938b66c --- /dev/null +++ b/2024/Generative Region-Language Pretraining for Open-Ended Object Detection/full.md @@ -0,0 +1,297 @@ +# Generative Region-Language Pretraining for Open-Ended Object Detection + +Chuang Lin1 + +Yi Jiang* + +Lizhen $\mathbf{Q}\mathbf{u}^{1}$ + +Zehuan Yuan + +Jianfei Cai* + +1 Monash University + +2 ByteDance Inc. + +# Abstract + +In recent research, significant attention has been devoted to the open-vocabulary object detection task, aiming to generalize beyond the limited number of classes labeled during training and detect objects described by arbitrary category names at inference. Compared with conventional object detection, open vocabulary object detection largely extends the object detection categories. However, it relies on calculating the similarity between image regions and a set of arbitrary category names with a pretrained vision-and-language model. This implies that, despite its open-set nature, the task still needs the predefined object categories during the inference stage. This raises the question: What if we do not have exact knowledge of object categories during inference? In this paper, we call such a new setting as generative open-ended object detection, which is a more general and practical problem. To address it, we formulate object detection as a generative problem and propose a simple framework named GenerateU, which can detect dense objects and generate their names in a free-form way. Particularly, we employ Deformable DETR as a region proposal generator with a language model translating visual regions to object names. To assess the free-form object detection task, we introduce an evaluation method designed to quantitatively measure the performance of generative outcomes. Extensive experiments demonstrate strong zero-shot detection performance of our GenerateU. For example, on the LVIS dataset, our GenerateU achieves comparable results to the open-vocabulary object detection method GLIP, even though the category names are not seen by GenerateU during inference. Code is available at: https://github.com/FoundationVision/GenerateU. + +# 1. Introduction + +Object detection, a crucial task in computer vision, has made significant progress in recent years, particularly with the advent of deep learning [6, 8, 18, 41, 46]. The objective of this task is to locate objects in images with tight bounding boxes and recognize their respective categories. However, + +![](images/25e562aad69e873247e74514ab066c5b0cdcd46222ccbfd44926dd867c5681f8.jpg) +Predefined category +(a) Open-Vocabulary Object Detection + +![](images/9ddb8aacd764116ab0bde92874b7dc211320af2d0e83f2d2064ee6d99e95c1b8.jpg) + +![](images/b6a34afda8e313e9d2d66ec8e143e972b2e349ac9d2c7743256850291ac94d45.jpg) +(b) Phrase Grounding +(c) Generative Open-Ended Object Detection +Figure 1. Comparing generative open-ended object detection with other open-set object detection tasks. Open-vocabulary object detection and phrase grounding typically require predefined categories or phrases in text prompts to align with image regions. In contrast, our introduced generative open-ended object detection is a more general and practical setting where categorical information is not explicitly defined. Such a setting is especially meaningful for scenarios where users lack precise knowledge of object categories during inference. + +most existing object detection algorithms are restricted to a predetermined set of object categories defined in detection datasets [9, 14, 17, 26, 42]. For example, an object detector trained on COCO [9] can well identify 80 pre-defined classes but encounters difficulty in recognizing new classes that have not been seen or defined during training. + +Open-vocabulary learning has been introduced to remedy this problem and considerable efforts have been made [7, 22, 28, 29, 35, 55, 65]. Open-vocabulary object detection addresses the closed-set limitation typically via + +weak supervision from language, i.e., image-text pairs, or large pretrained Vision-Language Models (VLMs), such as CLIP [39]. The text embeddings from the text encoder of CLIP can well align with the visual regions for the novel categories or the noun phrases in image captions, as shown in Fig. 1. However, despite open-set nature, open-vocabulary object detection still requires predefined object categories during the inference stage, while in many practical scenarios, we might not have exact knowledge of test images. Even if we have prior knowledge of test images, manually defining object categories might introduce language ambiguities (e.g., similar object names like "person", "a man", "young boy" in Fig. 1) or be not comprehensive enough (e.g. missing "chain link fence" in Fig. 1), ultimately restricting flexibility. If we do not have prior knowledge of test images, to cover all possible objects that may appear in the images, the common solution to design a large and comprehensive label set that would be complex and time-consuming. The above discussions raise the research question: Can we do open-world dense object detection that does not require predefined object categories during inference? We call this problem open-ended object detection. + +Therefore, in this paper, we formulate open-ended object detection as a generative problem and propose a simple architecture termed GenerateU to address it. Our GenerateU contains two major components: a visual object detector to localize image regions and a language model to translate visual regions to object names. This model is trained end-to-end, optimizing both components simultaneously. To broaden the knowledge of GenerateU, we train it using a small set of human-annotated object-language paired data and scale up the vocabulary size with massive image-text pairs of data. Furthermore, due to the limitation in image captions whose nouns often fall short of accurately representing all objects in the images, we employ a pseudolabeling method to supplement missing objects in images. + +Our major contributions are summarized as follows. + +- We introduce a new and more practical object detection problem: open-ended object detection, and formulate it as a generative problem. The reformulation avoids manual effort in predefining object categories during both training and inference, aiming at a more general and flexible architecture. +- We develop a novel end-to-end learning framework, termed GenerateU, which directly generates object names in a free-form manner. GenerateU is trained with a small set of human-annotated object-language paired data and massive image-text paired data. +- Compared with open-vocabulary object detection models, GenerateU achieves comparable results on zero-shot LVIS, even though it does not see object categories during inference. + +# 2. Related Work + +# 2.1. Open-Vocabulary Object Detection + +Learning visual models from language supervision becomes popular in image recognition tasks [19, 20, 39, 40, 58, 59, 62, 64]. Recently, many works [23, 23-25, 34, 43-45, 48, 49, 52-54, 56, 57] have extended the success in open-vocabulary image recognition to object detection, i.e., open-vocabulary object detection (OVOD). As the pioneering work of OVOD, OVRCNN [61] successfully transfers the vision and language model trained on image-text pairs to the detection framework. To leverage the capability of CLIP [39], ViLD [16] and RegionCLIP [63] turn to learn the visual region feature from classification-oriented models by knowledge distillation. OV-DETR [60] proposes a novel open-vocabulary detector built upon DETR [8] by formulating the classification as a binary matching between the input queries and the referring objects. Different from distilling frozen vision encoder, OWL [35] proposes a two-step framework for contrastive image-text pre-training and end-to-end detection fine-tuning, using a standard Vision Transformer [13]. Detic [65] proposes to extend the detector vocabulary with large-scale image classification data by applying supervision to the proposal with the largest spatial size. To reduce the annotation cost for open-vocabulary object detection, VLDet [29] proposes a novel framework to directly learn from image-text paired data by formulating object-language alignment as a set matching problem. + +Open vocabulary object detection aims at generalization to novel categories that were not seen during the training phase. Existing OVOD works typically rely on calculating the similarity between image regions and a set of arbitrary category names with a pretrained vision-language model. It remains necessary to predefine the categories during inference. In contrast, this work delves deeper into a new challenge - how to effectively handle scenarios where the exact categories remain unknown during the inference process. + +# 2.2. Multimodal Large Language Model + +Inspired by the remarkable capability of LLM [1, 4, 47, 50], multimodal foundation models [2, 15, 27, 30, 37, 38] leverage LLM as the core intelligence and are trained with extensive image-text pairs, resulting in enhanced performance across various vision-and-language tasks such as VQA and image captioning. BLIP-2 [27] proposes a generic and efficient pre-training strategy that bootstraps vision-language pre-training from off-the-shelf frozen pre-trained image encoders and frozen large language models. By inserting adapters into LLaMA's transformer, LLaMA-Adapter [15] introduces only 1.2M learnable parameters, and turns a LLaMA into an instruction-following model. LLaVA [30] represents a novel end-to-end trained large multimodal model that combines a vision encoder and Vicuna [36] + +for general-purpose visual and language understanding. It achieves impressive chat capabilities, emulating the essence of the multimodal GPT-4. + +Based on this amazing progress, subsequent studies further explore enhancing the object localization capabilities of multimodal models. Kosmos-2 [37] unlocks the grounding capability, enabling the model to provide visual answers such as bounding boxes, thereby supporting a wider array of vision-language tasks, including referring expression comprehension. UNINEXT[55] reformulates diverse instance perception tasks into a unified object discovery and retrieval paradigm and can flexibly perceive different types of objects by simply changing the input prompts. DetGPT [38] locates the objects of interest based on the language instructions and the LLM enables users to interact with the system, allowing for a higher level of interactivity like visual reasoning. In contrast to these works that focus on understanding natural language expressions referring to specific objects in an image, our work aims to identify and localize all objects of interest in an image, associating each with a class name. + +# 2.3. Dense Captioning + +Dense captioning aims to predict a set of descriptions across various image regions. DenseCap [21] develops a fully convolutional localization network to extract image regions, which are then described using an RNN language model. In dense captioning, it treats each region as an individual image and tries to generate captions encompassing multiple objects and their attribute, such as "cat riding a skateboard" and "bear is wearing a red hat". To evaluate the generated captions, dense captioning employs the METEOR [21] metric, originally designed for machine translation, which assesses the generation results from the sentence perspective. In comparison, our proposed generative open-ended object detection focuses on describing individual objects as category names in a zero-shot setting. + +Note that recently, CapDet [32] and GRiT [51] attempted to integrate open-vocabulary object detection and dense captioning by introducing an additional language model as a captioning head. However, for open-world object detection, they either still require a predefined category space during inference, or are limited to the close-set object detection problem. + +# 3. Method + +Our goal is to build an open-world object detector capable of localizing all objects in an image and providing their corresponding category names in a free-form way. Figure 2 gives an overview of our proposed open-ended object detection model, GenerateU, which comprises two major components: an open-world object detector and a language model. We describe the open-world object detector in Sec. 3.1, which consists of an image encoder and a detection head, + +and the language model, which can be directly borrowed from an existing Multimodal Large Language Model (including a V-to-L adaptor and a language model, Sec. 3.2) or retrained (Sec. 3.3). An additional pseudo-labeling method is described in Sec. 3.4 for enriching the label diversity. + +# 3.1. Open-World Object Detection + +The first crucial step is to extract accurate object regions from images, for which we develop an open-world object detector. Although we are not constrained to any specific object detection model, we choose Deformable DETR [66] for the following reasons. First, it alleviates heuristics when compared to other popular one or two-stage frameworks, and provides a versatile query-to-instance pipeline. Besides, Deformable DETR [66] demonstrates faster convergence speed and enhanced accuracy compared to the original DETR [8]. Deformable DETR or the original one employs Hungarian matching to learn a mapping from predicted queries to ground truth objects, followed by training the matched object query to regress to its corresponding ground truth with a combination of classification loss and bounding box regression loss. In the context of open-set problems, we do not rely on the object class information. Instead, we adopt an open-world detection approach, i.e. class-agnostic object detector, wherein matched queries are only classified into foreground or background. Consequently, the detection process involves three losses: binary cross-entropy, generalized IoU, and L1 regression loss. + +# 3.2. Transferring from a Frozen Multimodal LLM + +Once object candidates are localized, a natural idea is to leverage a pretrained multimodal large language model (MLLM) to reduce the training cost and offer strong zero-shot language abilities for generating object names. Recently developed MLLMs [2, 15, 27, 30, 37, 38] typically consist of three components: an image encoder to extract visual features; an adaptor network to bridge the modality gap and reduce the number of image embeddings, converting visual feature into token representations; a language model to generate predictions for vision-centric tasks. As shown in Figure 2 Top, we integrate a class-agnostic Deformable DETR, with a frozen image encoder to extract object queries, into an MLLM. We keep the MLLM frozen, and the only trainable module is the detection head for learning object representations. Besides, images are preprocessed to a resolution of $224 \times 224$ pixels during training in KOSMOS-2 [37]. In object detection, it is common to employ a larger input size for multi-scale feature extraction. To handle changes in image size, we utilize linear interpolation to enlarge the original position embedding. + +Despite its simplicity, our empirical study shows that such a direct open-end object detection based on pretrained MLLM results in poor performance large due to the do + +![](images/e12d69e0c9c9cc249edb36316903e6ffc8e8b7993cb7b370fc7e2246ed3b7574.jpg) +(a) Training with frozen MLLM + +![](images/3f9b49cba6219db8d05f7a320af1c6dbbb631c9358b78aad3cfc286fc2316452.jpg) +(b) Training End-to-End +Figure 2. Overview of our proposed open-ended object detection model, GenerateU, which comprises two major components: an object detector and a language model. We compare two training strategies: (Top) We incorporate the class-agnostic DETR (with frozen image encoder) into a pre-trained and frozen Multimodal Large Language Model (including Adaptor and Language Model), to facilitate a smooth transfer of knowledge from the language domain to object detection; (Bottom) We activate the image encoder and language model as trainable components, taking an end-to-end approach to seamlessly integrate region-level understanding into the language model. + +main shift. MLLMs were trained on large-scale image-text paired data, where they generate text descriptions given each whole image as a condition, without a fine-grained understanding of individual objects. + +# 3.3. Generative Region-Language Pretraining + +Thus, instead of employing a frozen MLLM, we propose GenerateU to directly link the open-world object detector with a language model, and activate both the image encoder and the language model as trainable components, as shown in Figure 2. Particularly, we employ an encoder-decoder-based language model, where visual representation serves as input for the encoder, and the associated text serves as the generation target for the language decoder. Here we only use the object queries that have been successfully matched with ground truth through Hungarian matching, ignoring the unmatched queries in the generating process. Specifically, each matched object query is first treated as an individual region, fed to a projection function to align its dimensions with the language model. The encoder comprises blocks with self-attention layers and feed-forward networks. The decoder, similar in structure, adds a cross-attention mechanism for interfacing with the encoder. The decoder employs an autoregressive self-attention mechanism, limiting the model's attention to past outputs, and its final output is processed through a dense layer with a softmax function. We apply the language modeling loss to train the model. + +Exploring on the region-word alignment loss. Benefiting from the pretrained vision-language model, we incorporate a region-word alignment loss during training to aid in learning to distinguish region features. In detail, for an object query, we treat its corresponding word as a positive sample + +and consider other words in the same minibatch as negative samples. By calculating the similarity between the region features from class-agnostic Deformable DETR and text features from the fixed CLIP text encoder [39], the model is optimized by a binary cross-entropy loss. This region-word alignment loss is jointly trained with the object detection losses and the language modeling loss. + +# 3.4. Enrich Label Diversity + +Considerable efforts have been invested in collecting detection data that is both semantically rich and voluminous. However, human annotation proves to be both costly and resource-restrictive. Previous works, such as Kosmos2 [37] and GLIP [28], leverage image-text pairs to expand the vocabulary scale. They use pretrained grounding models to generate bounding boxes aligned with noun phrases from captions. These boxes, in turn, serve as pseudo detection labels for training the model. However, given that captions may not comprehensively describe every object in an image, the count of nouns in the caption is typically much lower than the actual number of objects present in the corresponding image. Consequently, the quantity of pseudolabels generated is constrained. To address this, we use GenerateU pretrained on the available labels to generate pseudo labels as a supplement for missing objects in images. Furthermore, we observed that beam search in language models naturally generates synonyms, thereby providing diverse object labels. Figure 3 provides pseudolabel examples, illustrating the generation of boxes covering nearly all objects in the image, along with a diverse set of rich vocabulary labels. + +![](images/a5ab296f5fe245b9c1426f2d047d4f778caff974d4dcad1083fd693282fe2735.jpg) +"Belcher family you and me we got this poster A1." + +![](images/7a06a49ef61887658a4ba9fb018be6b94279cd338191f5ba2cec2fa862a8f080.jpg) +"A gorgeous wedding centrepiece with feathers, moss, greenery, white blooms and greenery." + +![](images/d99c7767488e108e6140b15d842d9c08213d4e88ce57711141a2f008505021d1.jpg) +"Just before Christmas, Thornton's issued a profit warning after suffering low order levels from supermarket customers and problems at its new warehouse." + +![](images/5d65a0938866d9eb1b9c62b29449f25f26559c7f8a2f165ace301636a1ab7a1b.jpg) +"Boy in red shirt catching a ball." + +![](images/b3e42baa73626b78b379e1c29c344166bb06ffb522d97476cbd013f28ff88b35.jpg) +"Adams, seen here celebrating after scoring against Chesterfield." + +![](images/d87f46a45fa9811eba6cd90f94d1cffe6329d5a6b1c0fcdefbf85822521200c8.jpg) +"Blonde woman with a colourful umbrella being blown away by the wind with her umbrella snowy landscape." + +![](images/c577a96246ee033c561625452eb3f432f40a7870178791c960ac6a17b32268d9.jpg) +"Meatballs on top of spaghetti squash on a white plate with rest behind it in skillet." + +![](images/58dfc839dd86f12068800f1167777d5eb6855852ea47a466edddecd173fe1296.jpg) +"How to make fresh brewed iced tea sweetened or without sugar. Homemade recipe using tea bags with a squeeze of lemon or orange juice and a sprig of mint." + +![](images/dc103776b5d647320b6693445d711adfee7765776020cd8e68ae0f3488221cb6.jpg) +"Close up picture of a plate of tuna noodle casserole with a bowl of salad in the background." + +![](images/7173af299b8b3acc84c36f764fbaba2a031585e6c6afe5549ed808c915a8479d.jpg) +"One of a kind: meet the only BMW E3 Estate in the world." +Figure 3. Selected pseudo-label examples highlight the generation of bounding boxes covering nearly all objects in the images. A varied set of descriptive labels showcases the model's capability in producing diverse and linguistically rich vocabulary. A white underline in a figure indicates that the pseudo label comes from the noun phrase (black underline) in the caption. + +![](images/c4351951b984b01fef65660c6550c75ac0f86e2df99623997fe920ed5b900988.jpg) +"The last time with his S15, Matt was always in the need of more power." + +![](images/968e8f549491e2fca42421ffd39c110a7a842c196083f6e8e0be81d69742f149.jpg) +"Construction Junction box truck parked in front of the loading dock." + +# 4. Experiments + +# 4.1. Datasets + +Our model is trained with hybrid supervision from two types of data: detection data and image-text data. For detection data, we use Visual Genome (VG) [26] which contains 77,398 images for training. Different from previous works [32, 51], we do not use the region captioning annotations of VG, since they may contain descriptions that + +encompass multiple objects or even describe actions. Our focus is directed towards the object detection task, and thus we leverage the object annotations of VG which annotate the same bounding box regions with concise and specific object names, covering a wide range of objects. For pretraining, we use about 5 million image-text pairs GRIT [37] from COYO700M [5] with pseudo grounding labels generated by KOSMOS-2 [37] using GLIP [28]. + +
MethodBackbonePre-Train DataOpen-endedOpen-setLVIS
APrAPcAPfAP
MDETR [22]RN101LVIS--20.924.924.324.2
MaskRCNN [18]RN101LVIS--26.334.033.933.3
Deformable DETR [66]Swin-TLVIS--24.236.038.236.0
GLIP (C) [28]Swin-TO365, GoldG-✓17.719.531.024.9
GLIP [28]Swin-TO365, GoldG, CAP4M-✓20.821.431.026.0
GenerateUSwin-TVG✓✓17.422.429.625.4
GenerateUSwin-TVG, GRIT5M✓✓20.024.929.826.8
GenerateUSwin-LVG, GRIT5M✓✓22.325.231.427.9
+ +Table 1. Zero-shot domain transfer to LVIS. Our method achieves comparable performance to prior models in the close-ended setting, without requiring access to category names during inference. + +# 4.2. Evaluation protocol and metrics + +Addressing an open-ended problem presents challenges in evaluating performance. For instance, when annotating a laptop, the human-annotated ground truth category may be "laptop", "computer", "PC", "Microcomputer", etc. To assess the effectiveness of the open-end generative results, we propose two settings: (a) We calculate the similarity score between the generated and annotated category names with a fixed pretrained text encoder. Notice that this text encoder is not utilized by the generative detection model; rather, it is only used for quantitative performance measurement. Any text encoder capable of computing similarity between two sentences can be adopted for this purpose. In this study, we select the popular CLIP [39] text encoder and BERTlarge [12] model. In this way, we can map the generative results to predefined categories in a dataset for evaluation. (b) We use METEOR [3] to assess the generated text quality, which is the same as the evaluation in dense captioning [21]. METEOR is widely used in NLP for evaluating machine translations. Following [21], we measure the mean Average Precision (AP) for both localization, using Intersection over Union (IoU) thresholds of .3, .4, .5, .6, and .7, and language accuracy, applying METEOR score thresholds at 0, .05, .1, .15, .2, and .25. + +Following GLIP [28], we evaluate our GenerateU primarily on LVIS [17] which contains 1203 categories and report APs for rare categories $(AP_r)$ , common categories $(AP_c)$ , frequency categories $(AP_f)$ and all categories $(mAP)$ , respectively. Specifically, following GLIP [28] and MDETR [22], we evaluate on the $5\mathrm{k}$ minival subset and report the zero-shot fixed AP [11] for a fair comparison. + +# 4.3. Implementation Details + +We attempt two different backbones, Swin-Tiny and Swin-Large [31] as the visual encoder. The Deformable DETR architecture follows [66] with 6 encoder layers and 6 decoder layers. The number of object queries $N$ is set to 300. + +We adopt FlanT5-base [10] as the language model and initialize the weights from it. The optimizer is AdamW [33] with a weight decay of 0.05. The learning rate is set to $2 \times 10^{-4}$ for the parameters of the visual backbone and the detection head, and $3 \times 10^{-4}$ for the language model. For the frozen MLLM approach in Sec 3.2, we explore the state-of-the-art model KOSMOS-2 [37] which was trained with the multimodal corpora. We fix all parameters and set the learning rate to $2 \times 10^{-4}$ for the detection head. Our model is trained on 16 A100 GPUs. + +# 4.4. Generative Open-Ended Detection Results + +We assess the model's capacity to identify diverse objects in a zero-shot setting on LVIS, presenting the results in Table 1, where rows 1-3 include methods utilizing LVIS as pre-training data, representing fully supervised models. Remarkably, we observe that even with only VG as training data, our model demonstrates commendable performance on zero-shot LVIS. This result suggests that predefined class names might not be necessary for open-world object detection during inference, especially when models are trained with extensive visual concepts. Furthermore, introducing additional training data from cost-effective image-text pairs consistently improves performance. For instance, when GenerateU trained with additional image-text pair data (GRIT [37]), we observe a substantial performance gain of 3.6 on $AP_{r}$ and 1.4 on $AP$ . In summary, the semantic richness derived from image-text data significantly enhances the model's ability to recognize rare objects. Notably, our model achieves comparable performance to GLIP under the same training data size but without requiring access to category names during inference. Moreover, we enhanced our model by integrating a larger backbone, i.e. Swin-Large, confirming that our method can scale effectively with increased model complexity. + +Transfer to other datasets. Our generative open-ended object detector can transfer to any dataset without any mod- + +
MethodsImage EncoderDetection HeadLanguage ModelLVIS
APrAPcAPfAP
Init from MLLMFreezeTrainFreeze*5.310.219.714.3
TrainTrainFreeze*10.917.228.121.9
Init from LLMTrainTrainFreeze11.220.127.422.9
TrainTrainTrain20.024.929.826.8
+ +Table 2. Comparison of our GenerateU with two different training strategies on a pretrained LLM on zero-shot LVIS. The results highlight generative object detection benefits from end-to-end training. Here, $\star$ refers to freezing both V-to-L adaptor and language model. + +
MethodsPre-Train DataCOCOObjects365
Supervised-46.525.6
GenerateUVG33.010.1
GenerateUVG, GRIT33.610.5
+ +Table 3. We evaluate our GenerateU on COCO and Objects365 validation set in a zero-shot setting. GenerateU can transfer to various downstream datasets without requiring any modifications. + +
MetricsAPrAPcAPfAP
CLIP [39]20.024.929.826.8
BERT [12]12.519.925.121.7
METEOR [3]---9.2
+ +Table 4. Effect of different evaluation metrics on zero-shot LVIS. + +ifications. In contrast to open vocabulary object detection methods, we do not need to switch the classifier to the specific category text embeddings. We evaluate the transferability of the proposed GenerateU on COCO [9] and Objects365 [42] validation sets, as shown in Table 3. The results demonstrate the effectiveness of our method and underscore the importance of scaling up the training data. + +# 4.5. Ablation Study + +Comparison with transferring from a Frozen MLLM. In Table 2, we evaluate the approach discussed in Sec 3.2, i.e. transferring a frozen MLLM to the generative object detection task. Following the state-of-the-art MLLM model KOSMOS-2 [37], we maintain its parameters fixed and only train the detection head. However, we observe that such direct transfer of the MLLM to object detection leads to poor performance in zero-shot LVIS. We attribute this result to the pretrained image encoder, which is trained on images of size $224 \times 224$ , which is insufficient for the object detection task. Besides, the MLLM originally supervised by image-level annotations such as image captions, lacks understanding in both object localization and fine-grained object categories. We also experiment with making the image encoder trainable while keeping the language model frozen. This + +![](images/c40eadc437b848cf075267e1b41a2fcbd93ad9730c7e694909684c4d3a8d55c2.jpg) +Figure 4. Effect of beam size. Beam search plays a crucial role in generating rare object names, effectively addressing the long-tail problem. + +adjustment significantly improves the results. We further fine-tune the entire model, initializing the Large Language Model (LLM) from FlanT5-base [10], which achieves the best performance. The above analysis underscores the importance of end-to-end training in the open-ended object detection task. + +Beam Search. Towards a general and practical object detection system, we believe that beam search plays a crucial role in generating outputs that are closely aligned with human intuition. This is because it explores a broader array of possibilities during the decoding phase. Beam search naturally contains the multi-level object names and their synonyms, enhancing the system's ability to recognize objects with varied linguistic expressions. This contributes to a more adaptable and nuanced understanding of the objects being detected. As shown in Figure 4, beam search largely improves the rare category performance, proving its importance for the long tail problem, and emphasizing the system's effectiveness in handling less frequent object categories. In our experiments, we set the default beam size to 3. + +Evaluation Metrics. To evaluate the results of generative object detection, we use different metrics: METEOR, evaluating the quality of text, and similarity scores, calculated by a fixed text encoder. Note that the off-the-shelf text encoder is not used for the proposed method, but just for the + +![](images/22645868d459770bef30c5708c58d3a11e89663802b14a6e9377ee97b5ac3905.jpg) +Figure 5. Qualitative prediction results from GenerateU and ground truth on LVIS [17]. GenerateU produces complete and precise predictions, showcasing its ability to go beyond fixed vocabulary constraints. + +
Pre-Train DataRegion-word AlignmentAPrAPcAPfAP
VG-16.019.126.722.5
✓17.4↑1.422.4↑3.929.6↑2.925.4↑2.9
VG, GRIT-19.223.528.925.7
✓20.0↑0.824.9↑1.429.8↑0.926.8↑0.9
+ +Table 5. Effect of region-word alignment loss. + +evaluation. Here we compare two offline models: CLIP text encoder [39] and BERT [12], as presented in Table 4. We find that the CLIP text encoder is particularly adept at handling noun phrases, while the BERT encoder, predominantly trained on sentence corpora, may exhibit different strengths in processing textual information. + +Region-word Alignment. We conduct ablation experiments to evaluate the effectiveness of the region-word alignment loss in our method, as presented in Table 5. We observe that training with only the detection loss and the language modeling loss yields promising results. Adding the region-word alignment loss provides additional supervision and further improves the performance. This further demonstrates the effectiveness of the proposed generative object detection method. + +# 4.6. Qualitative Results + +Here we provide qualitative visualizations in showcasing the superior object detection capabilities of our model. As shown in Figure 5, our model can identify a more diverse range of objects compared to ground truth annotations. These visualizations not only affirm the reliability of our detection framework but also demonstrate its ability to detect a wider variety of objects, providing valuable insights into the model's strengths. + +# 5. Conclusion + +This paper introduces GenerateU, a novel approach for generative open-ended object detection. By reformulating object detection as a generative task, the model eliminates the need for predefined categories during inference, offering a more practical solution. Our method achieves comparable results with open-vocabulary models on zero-shot LVIS, even though categories are not seen in our GenerateU. Overall, this work contributes to the fast-evolving area of object detection towards open-world scenarios. Future works include investigating the effect of training data scales and exploring methods beyond naive pseudo-labeling. + +# References + +[1] Rohan Anil, Andrew M Dai, Orhan First, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, et al. Palm 2 technical report. arXiv preprint arXiv:2305.10403, 2023. 2 +[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 2, 3 +[3] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 6, 7 +[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 2 +[5] Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, and Saehoon Kim. Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset, 2022.5 +[6] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6154-6162, 2018. 1 +[7] Zhaowei Cai, Gukyeong Kwon, Avinash Ravichandran, Erhan Bas, Zhuowen Tu, Rahul Bhotika, and Stefano Soatto. X-detr: A versatile architecture for instance-wise vision-language tasks. arXiv preprint arXiv:2204.05626, 2022. 1 +[8] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 1, 2, 3 +[9] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 1, 7 +[10] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416, 2022. 6, 7 +[11] Achal Dave, Piotr Dollar, Deva Ramanan, Alexander Kirilov, and Ross Girshick. Evaluating large-vocabulary object detectors: The devil is in the details. arXiv preprint arXiv:2102.01066, 2021. 6 +[12] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 6, 7, 8 +[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, + +Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2 +[14] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303-338, 2010. 1 +[15] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xi-angyu Yue, et al. Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010, 2023. 2, 3 +[16] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021. 2 +[17] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 1, 6, 8 +[18] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE International Conference on Computer Vision, pages 2961-2969, 2017. 1, 6 +[19] Dat Huynh, Jason Kuen, Zhe Lin, Jiumiang Gu, and Ehsan Elhamifar. Open-vocabulary instance segmentation via robust cross-modal pseudo-labeling. arXiv preprint arXiv:2111.12698, 2021. 2 +[20] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 2 +[21] Justin Johnson, Andrej Karpathy, and Li Fei-Fei. Densecap: Fully convolutional localization networks for dense captioning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4565–4574, 2016. 3, 6 +[22] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. MDetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1780-1790, 2021. 1, 6 +[23] Prannay Kaul, Weidi Xie, and Andrew Zisserman. Multimodal classifiers for open-vocabulary object detection. arXiv preprint arXiv:2306.05493, 2023. 2 +[24] Dahun Kim, Anelia Angelova, and Weicheng Kuo. Contrastive feature masking open-vocabulary vision transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15602-15612, 2023. +[25] Dahun Kim, Anelia Angelova, and Weicheng Kuo. Region-aware pretraining for open-vocabulary object detection with vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11144-11154, 2023. 2 + +[26] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123(1):32-73, 2017. 1, 5 +[27] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 2, 3 +[28] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. arXiv preprint arXiv:2112.03857, 2021. 1, 4, 5, 6 +[29] Chuang Lin, Peize Sun, Yi Jiang, Ping Luo, Lizhen Qu, Gholamreza Haffari, Zehuan Yuan, and Jianfei Cai. Learning object-language alignments for open-vocabulary object detection. arXiv preprint arXiv:2211.14843, 2022. 1, 2 +[30] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 2, 3 +[31] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 6 +[32] Yanxin Long, Youpeng Wen, Jianhua Han, Hang Xu, Pengzhen Ren, Wei Zhang, Shen Zhao, and Xiaodan Liang. Capdet: Unifying dense captioning and open-world detection pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15233-15243, 2023. 3, 5 +[33] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6 +[34] Chuofan Ma, Yi Jiang, Xin Wen, Zehuan Yuan, and Xiaojuan Qi. Codet: Co-occurrence guided region-word alignment for open-vocabulary object detection. arXiv preprint arXiv:2310.16667, 2023. 2 +[35] Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, et al. Simple open-vocabulary object detection with vision transformers. arXiv preprint arXiv:2205.06230, 2022. 1, 2 +[36] Baolin Peng, Chunyuan Li, Pengcheng He, Michel Galley, and Jianfeng Gao. Instruction tuning with gpt-4. arXiv preprint arXiv:2304.03277, 2023. 2 +[37] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 2, 3, 4, 5, 6, 7 +[38] Renjie Pi, Jiahui Gao, Shizhe Diao, Rui Pan, Hanze Dong, Jipeng Zhang, Lewei Yao, Jianhua Han, Hang Xu, and Lingpeng Kong Tong Zhang. Detgpt: Detect what you need via reasoning. arXiv preprint arXiv:2305.14167, 2023. 2, 3 +[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, + +Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 4, 6, 7, 8 +[40] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. arXiv preprint arXiv:2112.01518, 2021. 2 +[41] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In Advances in Neural Information Processing Systems, pages 91-99, 2015. 1 +[42] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019. 1, 7 +[43] Cheng Shi and Sibei Yang. Edadet: Open-vocabulary object detection using early dense alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15724-15734, 2023. 2 +[44] Hengcan Shi, Munawar Hayat, and Jianfei Cai. Open-vocabulary object detection via scene graph discovery. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4012-4021, 2023. +[45] Hengcan Shi, Munawar Hayat, and Jianfei Cai. Unified open-vocabulary dense visual prediction. arXiv preprint arXiv:2307.08238, 2023. 2 +[46] Peize Sun, Rufeng Zhang, Yi Jiang, Tao Kong, Chenfeng Xu, Wei Zhan, Masayoshi Tomizuka, Lei Li, Zehuan Yuan, Changhu Wang, et al. Sparse r-cnn: End-to-end object detection with learnable proposals. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14454-14463, 2021. 1 +[47] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2 +[48] Luting Wang, Yi Liu, Penghui Du, Zihan Ding, Yue Liao, Qiaosong Qi, Biaolong Chen, and Si Liu. Object-aware distillation pyramid for open-vocabulary object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11186-11196, 2023. 2 +[49] Tao Wang. Learning to detect and segment for open vocabulary object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7051-7060, 2023. 2 +[50] BigScience Workshop, Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ilic, Daniel Hesslow, Roman Castagné, Alexandra Sasha Luccioni, François Yvon, et al. Bloom: A 176b-parameter open-access multilingual language model. arXiv preprint arXiv:2211.05100, 2022. 2 +[51] Jialian Wu, Jianfeng Wang, Zhengyuan Yang, Zhe Gan, Zicheng Liu, Junsong Yuan, and Lijuan Wang. Grit: A generative region-to-text transformer for object understanding. arXiv preprint arXiv:2212.00280, 2022. 3, 5 + +[52] Jianzong Wu, Xiangtai Li, Henghui Ding, Xia Li, Guangliang Cheng, Yunhai Tong, and Chen Change Loy. Betrayed by captions: Joint caption grounding and generation for open vocabulary instance segmentation. arXiv preprint arXiv:2301.00805, 2023. 2 +[53] Size Wu, Wenwei Zhang, Sheng Jin, Wentao Liu, and Chen Change Loy. Aligning bag of regions for open-vocabulary object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15254-15264, 2023. +[54] Xiaoshi Wu, Feng Zhu, Rui Zhao, and Hongsheng Li. Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7031-7040, 2023. 2 +[55] Bin Yan, Yi Jiang, Jiannan Wu, Dong Wang, Ping Luo, Zehuan Yuan, and Huchuan Lu. Universal instance perception as object discovery and retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15325-15336, 2023. 1, 3 +[56] Lewei Yao, Jianhua Han, Youpeng Wen, Xiaodan Liang, Dan Xu, Wei Zhang, Zhenguo Li, Chunjing Xu, and Hang Xu. Detclip: Dictionary-enriched visual-concept paralleled pretraining for open-world detection. Advances in Neural Information Processing Systems, 35:9125-9138, 2022. 2 +[57] Lewei Yao, Jianhua Han, Xiaodan Liang, Dan Xu, Wei Zhang, Zhenguo Li, and Hang Xu. Detclipv2: Scalable open-vocabulary object detection pre-training via word-region alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23497-23506, 2023. 2 +[58] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 2 +[59] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 2 +[60] Yuhang Zang, Wei Li, Kaiyang Zhou, Chen Huang, and Chen Change Loy. Open-vocabulary detr with conditional matching. In European Conference on Computer Vision, pages 106-122. Springer, 2022. 2 +[61] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-vocabulary object detection using captions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14393-14402, 2021. 2 +[62] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. arXiv preprint arXiv:2111.07991, 2021. 2 +[63] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In Proceedings of the + +IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16793-16803, 2022. 2 +[64] Chong Zhou, Chen Change Loy, and Bo Dai. Denseclip: Extract free dense labels from clip. arXiv preprint arXiv:2112.01071, 2021. 2 +[65] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Phillip Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. arXiv preprint arXiv:2201.02605, 2022. 1, 2 +[66] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020.3, 6 \ No newline at end of file diff --git a/2024/Generative Region-Language Pretraining for Open-Ended Object Detection/images.zip b/2024/Generative Region-Language Pretraining for Open-Ended Object Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1b2f9027a3bc02aca148822bd927f64761c83a1a --- /dev/null +++ b/2024/Generative Region-Language Pretraining for Open-Ended Object Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b76304dc02a0ad7fd8eb22a15afe43b94401a446ea7e12e0a79395afc63638b7 +size 707066 diff --git a/2024/Generative Region-Language Pretraining for Open-Ended Object Detection/layout.json b/2024/Generative Region-Language Pretraining for Open-Ended Object Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0c46b9d7214cf27ddfcded1e00efbd41763e4b8a --- /dev/null +++ b/2024/Generative Region-Language Pretraining for Open-Ended Object Detection/layout.json @@ -0,0 +1,7639 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 63, + 103, + 531, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 103, + 531, + 121 + ], + "spans": [ + { + "bbox": [ + 63, + 103, + 531, + 121 + ], + "type": "text", + "content": "Generative Region-Language Pretraining for Open-Ended Object Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 100, + 141, + 168, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 141, + 168, + 157 + ], + "spans": [ + { + "bbox": [ + 100, + 141, + 168, + 157 + ], + "type": "text", + "content": "Chuang Lin1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 190, + 142, + 244, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 142, + 244, + 157 + ], + "spans": [ + { + "bbox": [ + 190, + 142, + 244, + 157 + ], + "type": "text", + "content": "Yi Jiang*" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 262, + 143, + 321, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 262, + 143, + 321, + 157 + ], + "spans": [ + { + "bbox": [ + 262, + 143, + 321, + 157 + ], + "type": "text", + "content": "Lizhen " + }, + { + "bbox": [ + 262, + 143, + 321, + 157 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}\\mathbf{u}^{1}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 345, + 143, + 419, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 143, + 419, + 156 + ], + "spans": [ + { + "bbox": [ + 345, + 143, + 419, + 156 + ], + "type": "text", + "content": "Zehuan Yuan" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 442, + 143, + 507, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 442, + 143, + 507, + 156 + ], + "spans": [ + { + "bbox": [ + 442, + 143, + 507, + 156 + ], + "type": "text", + "content": "Jianfei Cai*" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 196, + 157, + 297, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 157, + 297, + 171 + ], + "spans": [ + { + "bbox": [ + 196, + 157, + 297, + 171 + ], + "type": "text", + "content": "1 Monash University" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 321, + 157, + 405, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 157, + 405, + 171 + ], + "spans": [ + { + "bbox": [ + 321, + 157, + 405, + 171 + ], + "type": "text", + "content": "2 ByteDance Inc." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 143, + 199, + 191, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 199, + 191, + 211 + ], + "spans": [ + { + "bbox": [ + 143, + 199, + 191, + 211 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 224, + 290, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 224, + 290, + 594 + ], + "spans": [ + { + "bbox": [ + 47, + 224, + 290, + 594 + ], + "type": "text", + "content": "In recent research, significant attention has been devoted to the open-vocabulary object detection task, aiming to generalize beyond the limited number of classes labeled during training and detect objects described by arbitrary category names at inference. Compared with conventional object detection, open vocabulary object detection largely extends the object detection categories. However, it relies on calculating the similarity between image regions and a set of arbitrary category names with a pretrained vision-and-language model. This implies that, despite its open-set nature, the task still needs the predefined object categories during the inference stage. This raises the question: What if we do not have exact knowledge of object categories during inference? In this paper, we call such a new setting as generative open-ended object detection, which is a more general and practical problem. To address it, we formulate object detection as a generative problem and propose a simple framework named GenerateU, which can detect dense objects and generate their names in a free-form way. Particularly, we employ Deformable DETR as a region proposal generator with a language model translating visual regions to object names. To assess the free-form object detection task, we introduce an evaluation method designed to quantitatively measure the performance of generative outcomes. Extensive experiments demonstrate strong zero-shot detection performance of our GenerateU. For example, on the LVIS dataset, our GenerateU achieves comparable results to the open-vocabulary object detection method GLIP, even though the category names are not seen by GenerateU during inference. Code is available at: https://github.com/FoundationVision/GenerateU." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 616, + 127, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 616, + 127, + 628 + ], + "spans": [ + { + "bbox": [ + 47, + 616, + 127, + 628 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 636, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 636, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 636, + 287, + 696 + ], + "type": "text", + "content": "Object detection, a crucial task in computer vision, has made significant progress in recent years, particularly with the advent of deep learning [6, 8, 18, 41, 46]. The objective of this task is to locate objects in images with tight bounding boxes and recognize their respective categories. However," + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 312, + 210, + 512, + 273 + ], + "blocks": [ + { + "bbox": [ + 311, + 201, + 369, + 210 + ], + "lines": [ + { + "bbox": [ + 311, + 201, + 369, + 210 + ], + "spans": [ + { + "bbox": [ + 311, + 201, + 369, + 210 + ], + "type": "text", + "content": "Predefined category" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 312, + 210, + 512, + 273 + ], + "lines": [ + { + "bbox": [ + 312, + 210, + 512, + 273 + ], + "spans": [ + { + "bbox": [ + 312, + 210, + 512, + 273 + ], + "type": "image", + "image_path": "25e562aad69e873247e74514ab066c5b0cdcd46222ccbfd44926dd867c5681f8.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 278, + 438, + 288 + ], + "lines": [ + { + "bbox": [ + 312, + 278, + 438, + 288 + ], + "spans": [ + { + "bbox": [ + 312, + 278, + 438, + 288 + ], + "type": "text", + "content": "(a) Open-Vocabulary Object Detection" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 311, + 293, + 564, + 364 + ], + "blocks": [ + { + "bbox": [ + 311, + 293, + 564, + 364 + ], + "lines": [ + { + "bbox": [ + 311, + 293, + 564, + 364 + ], + "spans": [ + { + "bbox": [ + 311, + 293, + 564, + 364 + ], + "type": "image", + "image_path": "9ddb8aacd764116ab0bde92874b7dc211320af2d0e83f2d2064ee6d99e95c1b8.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 312, + 378, + 563, + 426 + ], + "blocks": [ + { + "bbox": [ + 313, + 367, + 386, + 376 + ], + "lines": [ + { + "bbox": [ + 313, + 367, + 386, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 386, + 376 + ], + "type": "text", + "content": "(b) Phrase Grounding" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 312, + 378, + 563, + 426 + ], + "lines": [ + { + "bbox": [ + 312, + 378, + 563, + 426 + ], + "spans": [ + { + "bbox": [ + 312, + 378, + 563, + 426 + ], + "type": "image", + "image_path": "b6a34afda8e313e9d2d66ec8e143e972b2e349ac9d2c7743256850291ac94d45.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 428, + 460, + 437 + ], + "lines": [ + { + "bbox": [ + 312, + 428, + 460, + 437 + ], + "spans": [ + { + "bbox": [ + 312, + 428, + 460, + 437 + ], + "type": "text", + "content": "(c) Generative Open-Ended Object Detection" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 450, + 547, + 550 + ], + "lines": [ + { + "bbox": [ + 304, + 450, + 547, + 550 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 547, + 550 + ], + "type": "text", + "content": "Figure 1. Comparing generative open-ended object detection with other open-set object detection tasks. Open-vocabulary object detection and phrase grounding typically require predefined categories or phrases in text prompts to align with image regions. In contrast, our introduced generative open-ended object detection is a more general and practical setting where categorical information is not explicitly defined. Such a setting is especially meaningful for scenarios where users lack precise knowledge of object categories during inference." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 590, + 547, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 590, + 547, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 590, + 547, + 663 + ], + "type": "text", + "content": "most existing object detection algorithms are restricted to a predetermined set of object categories defined in detection datasets [9, 14, 17, 26, 42]. For example, an object detector trained on COCO [9] can well identify 80 pre-defined classes but encounters difficulty in recognizing new classes that have not been seen or defined during training." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "content": "Open-vocabulary learning has been introduced to remedy this problem and considerable efforts have been made [7, 22, 28, 29, 35, 55, 65]. Open-vocabulary object detection addresses the closed-set limitation typically via" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 140, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 140, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 140, + 712 + ], + "type": "text", + "content": "*Corresponding Authors" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13958" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 337 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 337 + ], + "type": "text", + "content": "weak supervision from language, i.e., image-text pairs, or large pretrained Vision-Language Models (VLMs), such as CLIP [39]. The text embeddings from the text encoder of CLIP can well align with the visual regions for the novel categories or the noun phrases in image captions, as shown in Fig. 1. However, despite open-set nature, open-vocabulary object detection still requires predefined object categories during the inference stage, while in many practical scenarios, we might not have exact knowledge of test images. Even if we have prior knowledge of test images, manually defining object categories might introduce language ambiguities (e.g., similar object names like \"person\", \"a man\", \"young boy\" in Fig. 1) or be not comprehensive enough (e.g. missing \"chain link fence\" in Fig. 1), ultimately restricting flexibility. If we do not have prior knowledge of test images, to cover all possible objects that may appear in the images, the common solution to design a large and comprehensive label set that would be complex and time-consuming. The above discussions raise the research question: Can we do open-world dense object detection that does not require predefined object categories during inference? We call this problem open-ended object detection." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 342, + 289, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 342, + 289, + 510 + ], + "spans": [ + { + "bbox": [ + 47, + 342, + 289, + 510 + ], + "type": "text", + "content": "Therefore, in this paper, we formulate open-ended object detection as a generative problem and propose a simple architecture termed GenerateU to address it. Our GenerateU contains two major components: a visual object detector to localize image regions and a language model to translate visual regions to object names. This model is trained end-to-end, optimizing both components simultaneously. To broaden the knowledge of GenerateU, we train it using a small set of human-annotated object-language paired data and scale up the vocabulary size with massive image-text pairs of data. Furthermore, due to the limitation in image captions whose nouns often fall short of accurately representing all objects in the images, we employ a pseudolabeling method to supplement missing objects in images." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 516, + 269, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 516, + 269, + 528 + ], + "spans": [ + { + "bbox": [ + 59, + 516, + 269, + 528 + ], + "type": "text", + "content": "Our major contributions are summarized as follows." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 534, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 534, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 534, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 47, + 534, + 287, + 604 + ], + "type": "text", + "content": "- We introduce a new and more practical object detection problem: open-ended object detection, and formulate it as a generative problem. The reformulation avoids manual effort in predefining object categories during both training and inference, aiming at a more general and flexible architecture." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 606, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 606, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 47, + 606, + 287, + 665 + ], + "type": "text", + "content": "- We develop a novel end-to-end learning framework, termed GenerateU, which directly generates object names in a free-form manner. GenerateU is trained with a small set of human-annotated object-language paired data and massive image-text paired data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 666, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 666, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 666, + 287, + 712 + ], + "type": "text", + "content": "- Compared with open-vocabulary object detection models, GenerateU achieves comparable results on zero-shot LVIS, even though it does not see object categories during inference." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 91, + 493, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 91, + 493, + 104 + ], + "spans": [ + { + "bbox": [ + 306, + 91, + 493, + 104 + ], + "type": "text", + "content": "2.1. Open-Vocabulary Object Detection" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 110, + 547, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 110, + 547, + 410 + ], + "spans": [ + { + "bbox": [ + 304, + 110, + 547, + 410 + ], + "type": "text", + "content": "Learning visual models from language supervision becomes popular in image recognition tasks [19, 20, 39, 40, 58, 59, 62, 64]. Recently, many works [23, 23-25, 34, 43-45, 48, 49, 52-54, 56, 57] have extended the success in open-vocabulary image recognition to object detection, i.e., open-vocabulary object detection (OVOD). As the pioneering work of OVOD, OVRCNN [61] successfully transfers the vision and language model trained on image-text pairs to the detection framework. To leverage the capability of CLIP [39], ViLD [16] and RegionCLIP [63] turn to learn the visual region feature from classification-oriented models by knowledge distillation. OV-DETR [60] proposes a novel open-vocabulary detector built upon DETR [8] by formulating the classification as a binary matching between the input queries and the referring objects. Different from distilling frozen vision encoder, OWL [35] proposes a two-step framework for contrastive image-text pre-training and end-to-end detection fine-tuning, using a standard Vision Transformer [13]. Detic [65] proposes to extend the detector vocabulary with large-scale image classification data by applying supervision to the proposal with the largest spatial size. To reduce the annotation cost for open-vocabulary object detection, VLDet [29] proposes a novel framework to directly learn from image-text paired data by formulating object-language alignment as a set matching problem." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 410, + 547, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 410, + 547, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 410, + 547, + 518 + ], + "type": "text", + "content": "Open vocabulary object detection aims at generalization to novel categories that were not seen during the training phase. Existing OVOD works typically rely on calculating the similarity between image regions and a set of arbitrary category names with a pretrained vision-language model. It remains necessary to predefine the categories during inference. In contrast, this work delves deeper into a new challenge - how to effectively handle scenarios where the exact categories remain unknown during the inference process." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 527, + 497, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 527, + 497, + 540 + ], + "spans": [ + { + "bbox": [ + 306, + 527, + 497, + 540 + ], + "type": "text", + "content": "2.2. Multimodal Large Language Model" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": "Inspired by the remarkable capability of LLM [1, 4, 47, 50], multimodal foundation models [2, 15, 27, 30, 37, 38] leverage LLM as the core intelligence and are trained with extensive image-text pairs, resulting in enhanced performance across various vision-and-language tasks such as VQA and image captioning. BLIP-2 [27] proposes a generic and efficient pre-training strategy that bootstraps vision-language pre-training from off-the-shelf frozen pre-trained image encoders and frozen large language models. By inserting adapters into LLaMA's transformer, LLaMA-Adapter [15] introduces only 1.2M learnable parameters, and turns a LLaMA into an instruction-following model. LLaVA [30] represents a novel end-to-end trained large multimodal model that combines a vision encoder and Vicuna [36]" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13959" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "content": "for general-purpose visual and language understanding. It achieves impressive chat capabilities, emulating the essence of the multimodal GPT-4." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 108, + 288, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 108, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 108, + 288, + 312 + ], + "type": "text", + "content": "Based on this amazing progress, subsequent studies further explore enhancing the object localization capabilities of multimodal models. Kosmos-2 [37] unlocks the grounding capability, enabling the model to provide visual answers such as bounding boxes, thereby supporting a wider array of vision-language tasks, including referring expression comprehension. UNINEXT[55] reformulates diverse instance perception tasks into a unified object discovery and retrieval paradigm and can flexibly perceive different types of objects by simply changing the input prompts. DetGPT [38] locates the objects of interest based on the language instructions and the LLM enables users to interact with the system, allowing for a higher level of interactivity like visual reasoning. In contrast to these works that focus on understanding natural language expressions referring to specific objects in an image, our work aims to identify and localize all objects of interest in an image, associating each with a class name." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 318, + 153, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 318, + 153, + 331 + ], + "spans": [ + { + "bbox": [ + 47, + 318, + 153, + 331 + ], + "type": "text", + "content": "2.3. Dense Captioning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 337, + 287, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 337, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 46, + 337, + 287, + 503 + ], + "type": "text", + "content": "Dense captioning aims to predict a set of descriptions across various image regions. DenseCap [21] develops a fully convolutional localization network to extract image regions, which are then described using an RNN language model. In dense captioning, it treats each region as an individual image and tries to generate captions encompassing multiple objects and their attribute, such as \"cat riding a skateboard\" and \"bear is wearing a red hat\". To evaluate the generated captions, dense captioning employs the METEOR [21] metric, originally designed for machine translation, which assesses the generation results from the sentence perspective. In comparison, our proposed generative open-ended object detection focuses on describing individual objects as category names in a zero-shot setting." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 504, + 287, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 504, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 46, + 504, + 287, + 588 + ], + "type": "text", + "content": "Note that recently, CapDet [32] and GRiT [51] attempted to integrate open-vocabulary object detection and dense captioning by introducing an additional language model as a captioning head. However, for open-world object detection, they either still require a predefined category space during inference, or are limited to the close-set object detection problem." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 597, + 102, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 102, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 102, + 609 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "content": "Our goal is to build an open-world object detector capable of localizing all objects in an image and providing their corresponding category names in a free-form way. Figure 2 gives an overview of our proposed open-ended object detection model, GenerateU, which comprises two major components: an open-world object detector and a language model. We describe the open-world object detector in Sec. 3.1, which consists of an image encoder and a detection head," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "and the language model, which can be directly borrowed from an existing Multimodal Large Language Model (including a V-to-L adaptor and a language model, Sec. 3.2) or retrained (Sec. 3.3). An additional pseudo-labeling method is described in Sec. 3.4 for enriching the label diversity." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 140, + 469, + 153 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 140, + 469, + 153 + ], + "spans": [ + { + "bbox": [ + 305, + 140, + 469, + 153 + ], + "type": "text", + "content": "3.1. Open-World Object Detection" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 159, + 545, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 159, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 159, + 545, + 411 + ], + "type": "text", + "content": "The first crucial step is to extract accurate object regions from images, for which we develop an open-world object detector. Although we are not constrained to any specific object detection model, we choose Deformable DETR [66] for the following reasons. First, it alleviates heuristics when compared to other popular one or two-stage frameworks, and provides a versatile query-to-instance pipeline. Besides, Deformable DETR [66] demonstrates faster convergence speed and enhanced accuracy compared to the original DETR [8]. Deformable DETR or the original one employs Hungarian matching to learn a mapping from predicted queries to ground truth objects, followed by training the matched object query to regress to its corresponding ground truth with a combination of classification loss and bounding box regression loss. In the context of open-set problems, we do not rely on the object class information. Instead, we adopt an open-world detection approach, i.e. class-agnostic object detector, wherein matched queries are only classified into foreground or background. Consequently, the detection process involves three losses: binary cross-entropy, generalized IoU, and L1 regression loss." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 419, + 544, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 419, + 544, + 432 + ], + "spans": [ + { + "bbox": [ + 305, + 419, + 544, + 432 + ], + "type": "text", + "content": "3.2. Transferring from a Frozen Multimodal LLM" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 437, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 676 + ], + "type": "text", + "content": "Once object candidates are localized, a natural idea is to leverage a pretrained multimodal large language model (MLLM) to reduce the training cost and offer strong zero-shot language abilities for generating object names. Recently developed MLLMs [2, 15, 27, 30, 37, 38] typically consist of three components: an image encoder to extract visual features; an adaptor network to bridge the modality gap and reduce the number of image embeddings, converting visual feature into token representations; a language model to generate predictions for vision-centric tasks. As shown in Figure 2 Top, we integrate a class-agnostic Deformable DETR, with a frozen image encoder to extract object queries, into an MLLM. We keep the MLLM frozen, and the only trainable module is the detection head for learning object representations. Besides, images are preprocessed to a resolution of " + }, + { + "bbox": [ + 304, + 437, + 545, + 676 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 304, + 437, + 545, + 676 + ], + "type": "text", + "content": " pixels during training in KOSMOS-2 [37]. In object detection, it is common to employ a larger input size for multi-scale feature extraction. To handle changes in image size, we utilize linear interpolation to enlarge the original position embedding." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "Despite its simplicity, our empirical study shows that such a direct open-end object detection based on pretrained MLLM results in poor performance large due to the do" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13960" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 69, + 545, + 133 + ], + "blocks": [ + { + "bbox": [ + 53, + 69, + 545, + 133 + ], + "lines": [ + { + "bbox": [ + 53, + 69, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 53, + 69, + 545, + 133 + ], + "type": "image", + "image_path": "e12d69e0c9c9cc249edb36316903e6ffc8e8b7993cb7b370fc7e2246ed3b7574.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 134, + 167, + 144 + ], + "lines": [ + { + "bbox": [ + 55, + 134, + 167, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 134, + 167, + 144 + ], + "type": "text", + "content": "(a) Training with frozen MLLM" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 53, + 163, + 459, + 227 + ], + "blocks": [ + { + "bbox": [ + 54, + 150, + 143, + 160 + ], + "lines": [ + { + "bbox": [ + 54, + 150, + 143, + 160 + ], + "spans": [ + { + "bbox": [ + 54, + 150, + 143, + 160 + ], + "type": "text", + "content": "(b) Training End-to-End" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 53, + 163, + 459, + 227 + ], + "lines": [ + { + "bbox": [ + 53, + 163, + 459, + 227 + ], + "spans": [ + { + "bbox": [ + 53, + 163, + 459, + 227 + ], + "type": "image", + "image_path": "3f9b49cba6219db8d05f7a320af1c6dbbb631c9358b78aad3cfc286fc2316452.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 239, + 547, + 297 + ], + "lines": [ + { + "bbox": [ + 46, + 239, + 547, + 297 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 547, + 297 + ], + "type": "text", + "content": "Figure 2. Overview of our proposed open-ended object detection model, GenerateU, which comprises two major components: an object detector and a language model. We compare two training strategies: (Top) We incorporate the class-agnostic DETR (with frozen image encoder) into a pre-trained and frozen Multimodal Large Language Model (including Adaptor and Language Model), to facilitate a smooth transfer of knowledge from the language domain to object detection; (Bottom) We activate the image encoder and language model as trainable components, taking an end-to-end approach to seamlessly integrate region-level understanding into the language model." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 316, + 288, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 288, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 288, + 365 + ], + "type": "text", + "content": "main shift. MLLMs were trained on large-scale image-text paired data, where they generate text descriptions given each whole image as a condition, without a fine-grained understanding of individual objects." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 372, + 264, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 372, + 264, + 385 + ], + "spans": [ + { + "bbox": [ + 47, + 372, + 264, + 385 + ], + "type": "text", + "content": "3.3. Generative Region-Language Pretraining" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 390, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 390, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 390, + 287, + 653 + ], + "type": "text", + "content": "Thus, instead of employing a frozen MLLM, we propose GenerateU to directly link the open-world object detector with a language model, and activate both the image encoder and the language model as trainable components, as shown in Figure 2. Particularly, we employ an encoder-decoder-based language model, where visual representation serves as input for the encoder, and the associated text serves as the generation target for the language decoder. Here we only use the object queries that have been successfully matched with ground truth through Hungarian matching, ignoring the unmatched queries in the generating process. Specifically, each matched object query is first treated as an individual region, fed to a projection function to align its dimensions with the language model. The encoder comprises blocks with self-attention layers and feed-forward networks. The decoder, similar in structure, adds a cross-attention mechanism for interfacing with the encoder. The decoder employs an autoregressive self-attention mechanism, limiting the model's attention to past outputs, and its final output is processed through a dense layer with a softmax function. We apply the language modeling loss to train the model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "type": "text", + "content": "Exploring on the region-word alignment loss. Benefiting from the pretrained vision-language model, we incorporate a region-word alignment loss during training to aid in learning to distinguish region features. In detail, for an object query, we treat its corresponding word as a positive sample" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 316, + 547, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 547, + 401 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 547, + 401 + ], + "type": "text", + "content": "and consider other words in the same minibatch as negative samples. By calculating the similarity between the region features from class-agnostic Deformable DETR and text features from the fixed CLIP text encoder [39], the model is optimized by a binary cross-entropy loss. This region-word alignment loss is jointly trained with the object detection losses and the language modeling loss." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 426, + 436, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 426, + 436, + 440 + ], + "spans": [ + { + "bbox": [ + 305, + 426, + 436, + 440 + ], + "type": "text", + "content": "3.4. Enrich Label Diversity" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 450, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 545, + 715 + ], + "type": "text", + "content": "Considerable efforts have been invested in collecting detection data that is both semantically rich and voluminous. However, human annotation proves to be both costly and resource-restrictive. Previous works, such as Kosmos2 [37] and GLIP [28], leverage image-text pairs to expand the vocabulary scale. They use pretrained grounding models to generate bounding boxes aligned with noun phrases from captions. These boxes, in turn, serve as pseudo detection labels for training the model. However, given that captions may not comprehensively describe every object in an image, the count of nouns in the caption is typically much lower than the actual number of objects present in the corresponding image. Consequently, the quantity of pseudolabels generated is constrained. To address this, we use GenerateU pretrained on the available labels to generate pseudo labels as a supplement for missing objects in images. Furthermore, we observed that beam search in language models naturally generates synonyms, thereby providing diverse object labels. Figure 3 provides pseudolabel examples, illustrating the generation of boxes covering nearly all objects in the image, along with a diverse set of rich vocabulary labels." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13961" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 104, + 70, + 211, + 171 + ], + "blocks": [ + { + "bbox": [ + 104, + 70, + 211, + 171 + ], + "lines": [ + { + "bbox": [ + 104, + 70, + 211, + 171 + ], + "spans": [ + { + "bbox": [ + 104, + 70, + 211, + 171 + ], + "type": "image", + "image_path": "a5ab296f5fe245b9c1426f2d047d4f778caff974d4dcad1083fd693282fe2735.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 102, + 172, + 200, + 186 + ], + "lines": [ + { + "bbox": [ + 102, + 172, + 200, + 186 + ], + "spans": [ + { + "bbox": [ + 102, + 172, + 200, + 186 + ], + "type": "text", + "content": "\"Belcher family you and me we got this poster A1.\"" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 231, + 70, + 310, + 172 + ], + "blocks": [ + { + "bbox": [ + 231, + 70, + 310, + 172 + ], + "lines": [ + { + "bbox": [ + 231, + 70, + 310, + 172 + ], + "spans": [ + { + "bbox": [ + 231, + 70, + 310, + 172 + ], + "type": "image", + "image_path": "7a06a49ef61887658a4ba9fb018be6b94279cd338191f5ba2cec2fa862a8f080.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 219, + 172, + 334, + 186 + ], + "lines": [ + { + "bbox": [ + 219, + 172, + 334, + 186 + ], + "spans": [ + { + "bbox": [ + 219, + 172, + 334, + 186 + ], + "type": "text", + "content": "\"A gorgeous wedding centrepiece with feathers, moss, greenery, white blooms and greenery.\"" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 342, + 70, + 484, + 172 + ], + "blocks": [ + { + "bbox": [ + 342, + 70, + 484, + 172 + ], + "lines": [ + { + "bbox": [ + 342, + 70, + 484, + 172 + ], + "spans": [ + { + "bbox": [ + 342, + 70, + 484, + 172 + ], + "type": "image", + "image_path": "d99c7767488e108e6140b15d842d9c08213d4e88ce57711141a2f008505021d1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 342, + 172, + 487, + 194 + ], + "lines": [ + { + "bbox": [ + 342, + 172, + 487, + 194 + ], + "spans": [ + { + "bbox": [ + 342, + 172, + 487, + 194 + ], + "type": "text", + "content": "\"Just before Christmas, Thornton's issued a profit warning after suffering low order levels from supermarket customers and problems at its new warehouse.\"" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 103, + 194, + 223, + 285 + ], + "blocks": [ + { + "bbox": [ + 103, + 194, + 223, + 285 + ], + "lines": [ + { + "bbox": [ + 103, + 194, + 223, + 285 + ], + "spans": [ + { + "bbox": [ + 103, + 194, + 223, + 285 + ], + "type": "image", + "image_path": "5d65a0938866d9eb1b9c62b29449f25f26559c7f8a2f165ace301636a1ab7a1b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 107, + 285, + 187, + 292 + ], + "lines": [ + { + "bbox": [ + 107, + 285, + 187, + 292 + ], + "spans": [ + { + "bbox": [ + 107, + 285, + 187, + 292 + ], + "type": "text", + "content": "\"Boy in red shirt catching a ball.\"" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 239, + 194, + 313, + 285 + ], + "blocks": [ + { + "bbox": [ + 239, + 194, + 313, + 285 + ], + "lines": [ + { + "bbox": [ + 239, + 194, + 313, + 285 + ], + "spans": [ + { + "bbox": [ + 239, + 194, + 313, + 285 + ], + "type": "image", + "image_path": "b3e42baa73626b78b379e1c29c344166bb06ffb522d97476cbd013f28ff88b35.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 225, + 285, + 331, + 298 + ], + "lines": [ + { + "bbox": [ + 225, + 285, + 331, + 298 + ], + "spans": [ + { + "bbox": [ + 225, + 285, + 331, + 298 + ], + "type": "text", + "content": "\"Adams, seen here celebrating after scoring against Chesterfield.\"" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 349, + 194, + 484, + 285 + ], + "blocks": [ + { + "bbox": [ + 349, + 194, + 484, + 285 + ], + "lines": [ + { + "bbox": [ + 349, + 194, + 484, + 285 + ], + "spans": [ + { + "bbox": [ + 349, + 194, + 484, + 285 + ], + "type": "image", + "image_path": "d87f46a45fa9811eba6cd90f94d1cffe6329d5a6b1c0fcdefbf85822521200c8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 285, + 487, + 298 + ], + "lines": [ + { + "bbox": [ + 340, + 285, + 487, + 298 + ], + "spans": [ + { + "bbox": [ + 340, + 285, + 487, + 298 + ], + "type": "text", + "content": "\"Blonde woman with a colourful umbrella being blown away by the wind with her umbrella snowy landscape.\"" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 103, + 299, + 188, + 406 + ], + "blocks": [ + { + "bbox": [ + 103, + 299, + 188, + 406 + ], + "lines": [ + { + "bbox": [ + 103, + 299, + 188, + 406 + ], + "spans": [ + { + "bbox": [ + 103, + 299, + 188, + 406 + ], + "type": "image", + "image_path": "c577a96246ee033c561625452eb3f432f40a7870178791c960ac6a17b32268d9.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 105, + 406, + 184, + 426 + ], + "lines": [ + { + "bbox": [ + 105, + 406, + 184, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 184, + 426 + ], + "type": "text", + "content": "\"Meatballs on top of spaghetti squash on a white plate with rest behind it in skillet.\"" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 223, + 299, + 302, + 404 + ], + "blocks": [ + { + "bbox": [ + 223, + 299, + 302, + 404 + ], + "lines": [ + { + "bbox": [ + 223, + 299, + 302, + 404 + ], + "spans": [ + { + "bbox": [ + 223, + 299, + 302, + 404 + ], + "type": "image", + "image_path": "58dfc839dd86f12068800f1167777d5eb6855852ea47a466edddecd173fe1296.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 190, + 405, + 331, + 427 + ], + "lines": [ + { + "bbox": [ + 190, + 405, + 331, + 427 + ], + "spans": [ + { + "bbox": [ + 190, + 405, + 331, + 427 + ], + "type": "text", + "content": "\"How to make fresh brewed iced tea sweetened or without sugar. Homemade recipe using tea bags with a squeeze of lemon or orange juice and a sprig of mint.\"" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 340, + 299, + 484, + 406 + ], + "blocks": [ + { + "bbox": [ + 340, + 299, + 484, + 406 + ], + "lines": [ + { + "bbox": [ + 340, + 299, + 484, + 406 + ], + "spans": [ + { + "bbox": [ + 340, + 299, + 484, + 406 + ], + "type": "image", + "image_path": "dc103776b5d647320b6693445d711adfee7765776020cd8e68ae0f3488221cb6.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 342, + 406, + 480, + 420 + ], + "lines": [ + { + "bbox": [ + 342, + 406, + 480, + 420 + ], + "spans": [ + { + "bbox": [ + 342, + 406, + 480, + 420 + ], + "type": "text", + "content": "\"Close up picture of a plate of tuna noodle casserole with a bowl of salad in the background.\"" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 103, + 428, + 239, + 517 + ], + "blocks": [ + { + "bbox": [ + 103, + 428, + 239, + 517 + ], + "lines": [ + { + "bbox": [ + 103, + 428, + 239, + 517 + ], + "spans": [ + { + "bbox": [ + 103, + 428, + 239, + 517 + ], + "type": "image", + "image_path": "7173af299b8b3acc84c36f764fbaba2a031585e6c6afe5549ed808c915a8479d.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 105, + 518, + 233, + 532 + ], + "lines": [ + { + "bbox": [ + 105, + 518, + 233, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 233, + 532 + ], + "type": "text", + "content": "\"One of a kind: meet the only BMW E3 Estate in the world.\"" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 546, + 547, + 579 + ], + "lines": [ + { + "bbox": [ + 46, + 546, + 547, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 546, + 547, + 579 + ], + "type": "text", + "content": "Figure 3. Selected pseudo-label examples highlight the generation of bounding boxes covering nearly all objects in the images. A varied set of descriptive labels showcases the model's capability in producing diverse and linguistically rich vocabulary. A white underline in a figure indicates that the pseudo label comes from the noun phrase (black underline) in the caption." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 251, + 428, + 365, + 517 + ], + "blocks": [ + { + "bbox": [ + 251, + 428, + 365, + 517 + ], + "lines": [ + { + "bbox": [ + 251, + 428, + 365, + 517 + ], + "spans": [ + { + "bbox": [ + 251, + 428, + 365, + 517 + ], + "type": "image", + "image_path": "c4351951b984b01fef65660c6550c75ac0f86e2df99623997fe920ed5b900988.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 252, + 518, + 369, + 533 + ], + "lines": [ + { + "bbox": [ + 252, + 518, + 369, + 533 + ], + "spans": [ + { + "bbox": [ + 252, + 518, + 369, + 533 + ], + "type": "text", + "content": "\"The last time with his S15, Matt was always in the need of more power.\"" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 381, + 429, + 483, + 517 + ], + "blocks": [ + { + "bbox": [ + 381, + 429, + 483, + 517 + ], + "lines": [ + { + "bbox": [ + 381, + 429, + 483, + 517 + ], + "spans": [ + { + "bbox": [ + 381, + 429, + 483, + 517 + ], + "type": "image", + "image_path": "968e8f549491e2fca42421ffd39c110a7a842c196083f6e8e0be81d69742f149.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 381, + 518, + 479, + 533 + ], + "lines": [ + { + "bbox": [ + 381, + 518, + 479, + 533 + ], + "spans": [ + { + "bbox": [ + 381, + 518, + 479, + 533 + ], + "type": "text", + "content": "\"Construction Junction box truck parked in front of the loading dock.\"" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 47, + 599, + 127, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 127, + 613 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 127, + 613 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 47, + 621, + 109, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 109, + 632 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 109, + 632 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "content": "Our model is trained with hybrid supervision from two types of data: detection data and image-text data. For detection data, we use Visual Genome (VG) [26] which contains 77,398 images for training. Different from previous works [32, 51], we do not use the region captioning annotations of VG, since they may contain descriptions that" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 600, + 545, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 600, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 600, + 545, + 696 + ], + "type": "text", + "content": "encompass multiple objects or even describe actions. Our focus is directed towards the object detection task, and thus we leverage the object annotations of VG which annotate the same bounding box regions with concise and specific object names, covering a wide range of objects. For pretraining, we use about 5 million image-text pairs GRIT [37] from COYO700M [5] with pseudo grounding labels generated by KOSMOS-2 [37] using GLIP [28]." + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13962" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 70, + 547, + 213 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 547, + 213 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 547, + 213 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 547, + 213 + ], + "type": "table", + "html": "
MethodBackbonePre-Train DataOpen-endedOpen-setLVIS
APrAPcAPfAP
MDETR [22]RN101LVIS--20.924.924.324.2
MaskRCNN [18]RN101LVIS--26.334.033.933.3
Deformable DETR [66]Swin-TLVIS--24.236.038.236.0
GLIP (C) [28]Swin-TO365, GoldG-✓17.719.531.024.9
GLIP [28]Swin-TO365, GoldG, CAP4M-✓20.821.431.026.0
GenerateUSwin-TVG✓✓17.422.429.625.4
GenerateUSwin-TVG, GRIT5M✓✓20.024.929.826.8
GenerateUSwin-LVG, GRIT5M✓✓22.325.231.427.9
", + "image_path": "f4bc7fb1dc95aeb4ff379d415c99db5e599ae0713d8d012da44e135dfc7084a8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 220, + 546, + 245 + ], + "lines": [ + { + "bbox": [ + 46, + 220, + 546, + 245 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 546, + 245 + ], + "type": "text", + "content": "Table 1. Zero-shot domain transfer to LVIS. Our method achieves comparable performance to prior models in the close-ended setting, without requiring access to category names during inference." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 264, + 221, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 264, + 221, + 277 + ], + "spans": [ + { + "bbox": [ + 47, + 264, + 221, + 277 + ], + "type": "text", + "content": "4.2. Evaluation protocol and metrics" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 282, + 289, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 282, + 289, + 556 + ], + "spans": [ + { + "bbox": [ + 45, + 282, + 289, + 556 + ], + "type": "text", + "content": "Addressing an open-ended problem presents challenges in evaluating performance. For instance, when annotating a laptop, the human-annotated ground truth category may be \"laptop\", \"computer\", \"PC\", \"Microcomputer\", etc. To assess the effectiveness of the open-end generative results, we propose two settings: (a) We calculate the similarity score between the generated and annotated category names with a fixed pretrained text encoder. Notice that this text encoder is not utilized by the generative detection model; rather, it is only used for quantitative performance measurement. Any text encoder capable of computing similarity between two sentences can be adopted for this purpose. In this study, we select the popular CLIP [39] text encoder and BERTlarge [12] model. In this way, we can map the generative results to predefined categories in a dataset for evaluation. (b) We use METEOR [3] to assess the generated text quality, which is the same as the evaluation in dense captioning [21]. METEOR is widely used in NLP for evaluating machine translations. Following [21], we measure the mean Average Precision (AP) for both localization, using Intersection over Union (IoU) thresholds of .3, .4, .5, .6, and .7, and language accuracy, applying METEOR score thresholds at 0, .05, .1, .15, .2, and .25." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 558, + 287, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 287, + 643 + ], + "type": "text", + "content": "Following GLIP [28], we evaluate our GenerateU primarily on LVIS [17] which contains 1203 categories and report APs for rare categories " + }, + { + "bbox": [ + 46, + 558, + 287, + 643 + ], + "type": "inline_equation", + "content": "(AP_r)" + }, + { + "bbox": [ + 46, + 558, + 287, + 643 + ], + "type": "text", + "content": ", common categories " + }, + { + "bbox": [ + 46, + 558, + 287, + 643 + ], + "type": "inline_equation", + "content": "(AP_c)" + }, + { + "bbox": [ + 46, + 558, + 287, + 643 + ], + "type": "text", + "content": ", frequency categories " + }, + { + "bbox": [ + 46, + 558, + 287, + 643 + ], + "type": "inline_equation", + "content": "(AP_f)" + }, + { + "bbox": [ + 46, + 558, + 287, + 643 + ], + "type": "text", + "content": " and all categories " + }, + { + "bbox": [ + 46, + 558, + 287, + 643 + ], + "type": "inline_equation", + "content": "(mAP)" + }, + { + "bbox": [ + 46, + 558, + 287, + 643 + ], + "type": "text", + "content": ", respectively. Specifically, following GLIP [28] and MDETR [22], we evaluate on the " + }, + { + "bbox": [ + 46, + 558, + 287, + 643 + ], + "type": "inline_equation", + "content": "5\\mathrm{k}" + }, + { + "bbox": [ + 46, + 558, + 287, + 643 + ], + "type": "text", + "content": " minival subset and report the zero-shot fixed AP [11] for a fair comparison." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 647, + 180, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 180, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 180, + 659 + ], + "type": "text", + "content": "4.3. Implementation Details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "content": "We attempt two different backbones, Swin-Tiny and Swin-Large [31] as the visual encoder. The Deformable DETR architecture follows [66] with 6 encoder layers and 6 decoder layers. The number of object queries " + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "content": " is set to 300." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 264, + 547, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 264, + 547, + 385 + ], + "spans": [ + { + "bbox": [ + 304, + 264, + 547, + 385 + ], + "type": "text", + "content": "We adopt FlanT5-base [10] as the language model and initialize the weights from it. The optimizer is AdamW [33] with a weight decay of 0.05. The learning rate is set to " + }, + { + "bbox": [ + 304, + 264, + 547, + 385 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-4}" + }, + { + "bbox": [ + 304, + 264, + 547, + 385 + ], + "type": "text", + "content": " for the parameters of the visual backbone and the detection head, and " + }, + { + "bbox": [ + 304, + 264, + 547, + 385 + ], + "type": "inline_equation", + "content": "3 \\times 10^{-4}" + }, + { + "bbox": [ + 304, + 264, + 547, + 385 + ], + "type": "text", + "content": " for the language model. For the frozen MLLM approach in Sec 3.2, we explore the state-of-the-art model KOSMOS-2 [37] which was trained with the multimodal corpora. We fix all parameters and set the learning rate to " + }, + { + "bbox": [ + 304, + 264, + 547, + 385 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-4}" + }, + { + "bbox": [ + 304, + 264, + 547, + 385 + ], + "type": "text", + "content": " for the detection head. Our model is trained on 16 A100 GPUs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 394, + 527, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 394, + 527, + 407 + ], + "spans": [ + { + "bbox": [ + 305, + 394, + 527, + 407 + ], + "type": "text", + "content": "4.4. Generative Open-Ended Detection Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 413, + 547, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 547, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 547, + 688 + ], + "type": "text", + "content": "We assess the model's capacity to identify diverse objects in a zero-shot setting on LVIS, presenting the results in Table 1, where rows 1-3 include methods utilizing LVIS as pre-training data, representing fully supervised models. Remarkably, we observe that even with only VG as training data, our model demonstrates commendable performance on zero-shot LVIS. This result suggests that predefined class names might not be necessary for open-world object detection during inference, especially when models are trained with extensive visual concepts. Furthermore, introducing additional training data from cost-effective image-text pairs consistently improves performance. For instance, when GenerateU trained with additional image-text pair data (GRIT [37]), we observe a substantial performance gain of 3.6 on " + }, + { + "bbox": [ + 304, + 413, + 547, + 688 + ], + "type": "inline_equation", + "content": "AP_{r}" + }, + { + "bbox": [ + 304, + 413, + 547, + 688 + ], + "type": "text", + "content": " and 1.4 on " + }, + { + "bbox": [ + 304, + 413, + 547, + 688 + ], + "type": "inline_equation", + "content": "AP" + }, + { + "bbox": [ + 304, + 413, + 547, + 688 + ], + "type": "text", + "content": ". In summary, the semantic richness derived from image-text data significantly enhances the model's ability to recognize rare objects. Notably, our model achieves comparable performance to GLIP under the same training data size but without requiring access to category names during inference. Moreover, we enhanced our model by integrating a larger backbone, i.e. Swin-Large, confirming that our method can scale effectively with increased model complexity." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "type": "text", + "content": "Transfer to other datasets. Our generative open-ended object detector can transfer to any dataset without any mod-" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13963" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 545, + 160 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 545, + 160 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 545, + 160 + ], + "type": "table", + "html": "
MethodsImage EncoderDetection HeadLanguage ModelLVIS
APrAPcAPfAP
Init from MLLMFreezeTrainFreeze*5.310.219.714.3
TrainTrainFreeze*10.917.228.121.9
Init from LLMTrainTrainFreeze11.220.127.422.9
TrainTrainTrain20.024.929.826.8
", + "image_path": "3a762135de65307a5b3b15cb680707e0c32c27bbcecd32a8f251c2c01f3f8e01.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 47, + 209, + 287, + 270 + ], + "blocks": [ + { + "bbox": [ + 46, + 168, + 546, + 191 + ], + "lines": [ + { + "bbox": [ + 46, + 168, + 546, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 546, + 191 + ], + "type": "text", + "content": "Table 2. Comparison of our GenerateU with two different training strategies on a pretrained LLM on zero-shot LVIS. The results highlight generative object detection benefits from end-to-end training. Here, " + }, + { + "bbox": [ + 46, + 168, + 546, + 191 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 46, + 168, + 546, + 191 + ], + "type": "text", + "content": " refers to freezing both V-to-L adaptor and language model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 209, + 287, + 270 + ], + "lines": [ + { + "bbox": [ + 47, + 209, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 47, + 209, + 287, + 270 + ], + "type": "table", + "html": "
MethodsPre-Train DataCOCOObjects365
Supervised-46.525.6
GenerateUVG33.010.1
GenerateUVG, GRIT33.610.5
", + "image_path": "ef069a0dc76299e1d5b347d997cd481c49a189a43d6de33cff188909422e0636.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 47, + 323, + 284, + 384 + ], + "blocks": [ + { + "bbox": [ + 46, + 279, + 287, + 313 + ], + "lines": [ + { + "bbox": [ + 46, + 279, + 287, + 313 + ], + "spans": [ + { + "bbox": [ + 46, + 279, + 287, + 313 + ], + "type": "text", + "content": "Table 3. We evaluate our GenerateU on COCO and Objects365 validation set in a zero-shot setting. GenerateU can transfer to various downstream datasets without requiring any modifications." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 323, + 284, + 384 + ], + "lines": [ + { + "bbox": [ + 47, + 323, + 284, + 384 + ], + "spans": [ + { + "bbox": [ + 47, + 323, + 284, + 384 + ], + "type": "table", + "html": "
MetricsAPrAPcAPfAP
CLIP [39]20.024.929.826.8
BERT [12]12.519.925.121.7
METEOR [3]---9.2
", + "image_path": "a2b83b2455ab360e1f2e80ef7b5db93b1f7bed44f589a5c0894a46b7c9301a53.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 392, + 285, + 403 + ], + "lines": [ + { + "bbox": [ + 47, + 392, + 285, + 403 + ], + "spans": [ + { + "bbox": [ + 47, + 392, + 285, + 403 + ], + "type": "text", + "content": "Table 4. Effect of different evaluation metrics on zero-shot LVIS." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 45, + 424, + 287, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 424, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 45, + 424, + 287, + 509 + ], + "type": "text", + "content": "ifications. In contrast to open vocabulary object detection methods, we do not need to switch the classifier to the specific category text embeddings. We evaluate the transferability of the proposed GenerateU on COCO [9] and Objects365 [42] validation sets, as shown in Table 3. The results demonstrate the effectiveness of our method and underscore the importance of scaling up the training data." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 515, + 141, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 515, + 141, + 528 + ], + "spans": [ + { + "bbox": [ + 46, + 515, + 141, + 528 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 533, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 533, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 45, + 533, + 287, + 714 + ], + "type": "text", + "content": "Comparison with transferring from a Frozen MLLM. In Table 2, we evaluate the approach discussed in Sec 3.2, i.e. transferring a frozen MLLM to the generative object detection task. Following the state-of-the-art MLLM model KOSMOS-2 [37], we maintain its parameters fixed and only train the detection head. However, we observe that such direct transfer of the MLLM to object detection leads to poor performance in zero-shot LVIS. We attribute this result to the pretrained image encoder, which is trained on images of size " + }, + { + "bbox": [ + 45, + 533, + 287, + 714 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 45, + 533, + 287, + 714 + ], + "type": "text", + "content": ", which is insufficient for the object detection task. Besides, the MLLM originally supervised by image-level annotations such as image captions, lacks understanding in both object localization and fine-grained object categories. We also experiment with making the image encoder trainable while keeping the language model frozen. This" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 213, + 541, + 333 + ], + "blocks": [ + { + "bbox": [ + 309, + 213, + 541, + 333 + ], + "lines": [ + { + "bbox": [ + 309, + 213, + 541, + 333 + ], + "spans": [ + { + "bbox": [ + 309, + 213, + 541, + 333 + ], + "type": "image", + "image_path": "c40eadc437b848cf075267e1b41a2fcbd93ad9730c7e694909684c4d3a8d55c2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 342, + 546, + 376 + ], + "lines": [ + { + "bbox": [ + 305, + 342, + 546, + 376 + ], + "spans": [ + { + "bbox": [ + 305, + 342, + 546, + 376 + ], + "type": "text", + "content": "Figure 4. Effect of beam size. Beam search plays a crucial role in generating rare object names, effectively addressing the long-tail problem." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 399, + 545, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 545, + 471 + ], + "type": "text", + "content": "adjustment significantly improves the results. We further fine-tune the entire model, initializing the Large Language Model (LLM) from FlanT5-base [10], which achieves the best performance. The above analysis underscores the importance of end-to-end training in the open-ended object detection task." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 472, + 546, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 546, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 546, + 651 + ], + "type": "text", + "content": "Beam Search. Towards a general and practical object detection system, we believe that beam search plays a crucial role in generating outputs that are closely aligned with human intuition. This is because it explores a broader array of possibilities during the decoding phase. Beam search naturally contains the multi-level object names and their synonyms, enhancing the system's ability to recognize objects with varied linguistic expressions. This contributes to a more adaptable and nuanced understanding of the objects being detected. As shown in Figure 4, beam search largely improves the rare category performance, proving its importance for the long tail problem, and emphasizing the system's effectiveness in handling less frequent object categories. In our experiments, we set the default beam size to 3." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "Evaluation Metrics. To evaluate the results of generative object detection, we use different metrics: METEOR, evaluating the quality of text, and similarity scores, calculated by a fixed text encoder. Note that the off-the-shelf text encoder is not used for the proposed method, but just for the" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13964" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 70, + 520, + 345 + ], + "blocks": [ + { + "bbox": [ + 80, + 70, + 520, + 345 + ], + "lines": [ + { + "bbox": [ + 80, + 70, + 520, + 345 + ], + "spans": [ + { + "bbox": [ + 80, + 70, + 520, + 345 + ], + "type": "image", + "image_path": "22645868d459770bef30c5708c58d3a11e89663802b14a6e9377ee97b5ac3905.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 353, + 546, + 376 + ], + "lines": [ + { + "bbox": [ + 46, + 353, + 546, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 353, + 546, + 376 + ], + "type": "text", + "content": "Figure 5. Qualitative prediction results from GenerateU and ground truth on LVIS [17]. GenerateU produces complete and precise predictions, showcasing its ability to go beyond fixed vocabulary constraints." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 47, + 394, + 298, + 485 + ], + "blocks": [ + { + "bbox": [ + 47, + 394, + 298, + 485 + ], + "lines": [ + { + "bbox": [ + 47, + 394, + 298, + 485 + ], + "spans": [ + { + "bbox": [ + 47, + 394, + 298, + 485 + ], + "type": "table", + "html": "
Pre-Train DataRegion-word AlignmentAPrAPcAPfAP
VG-16.019.126.722.5
✓17.4↑1.422.4↑3.929.6↑2.925.4↑2.9
VG, GRIT-19.223.528.925.7
✓20.0↑0.824.9↑1.429.8↑0.926.8↑0.9
", + "image_path": "d60254961e0b8c1b11c632257a3800e91d999add5d465cd464837da6a0d82236.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 83, + 492, + 251, + 504 + ], + "lines": [ + { + "bbox": [ + 83, + 492, + 251, + 504 + ], + "spans": [ + { + "bbox": [ + 83, + 492, + 251, + 504 + ], + "type": "text", + "content": "Table 5. Effect of region-word alignment loss." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 531, + 287, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 531, + 287, + 603 + ], + "spans": [ + { + "bbox": [ + 46, + 531, + 287, + 603 + ], + "type": "text", + "content": "evaluation. Here we compare two offline models: CLIP text encoder [39] and BERT [12], as presented in Table 4. We find that the CLIP text encoder is particularly adept at handling noun phrases, while the BERT encoder, predominantly trained on sentence corpora, may exhibit different strengths in processing textual information." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "text", + "content": "Region-word Alignment. We conduct ablation experiments to evaluate the effectiveness of the region-word alignment loss in our method, as presented in Table 5. We observe that training with only the detection loss and the language modeling loss yields promising results. Adding the region-word alignment loss provides additional supervision and further improves the performance. This further demonstrates the effectiveness of the proposed generative object detection method." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 396, + 417, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 396, + 417, + 408 + ], + "spans": [ + { + "bbox": [ + 306, + 396, + 417, + 408 + ], + "type": "text", + "content": "4.6. Qualitative Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 422, + 545, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 545, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 545, + 519 + ], + "type": "text", + "content": "Here we provide qualitative visualizations in showcasing the superior object detection capabilities of our model. As shown in Figure 5, our model can identify a more diverse range of objects compared to ground truth annotations. These visualizations not only affirm the reliability of our detection framework but also demonstrate its ability to detect a wider variety of objects, providing valuable insights into the model's strengths." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 553, + 378, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 553, + 378, + 566 + ], + "spans": [ + { + "bbox": [ + 306, + 553, + 378, + 566 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 580, + 545, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 580, + 545, + 704 + ], + "spans": [ + { + "bbox": [ + 304, + 580, + 545, + 704 + ], + "type": "text", + "content": "This paper introduces GenerateU, a novel approach for generative open-ended object detection. By reformulating object detection as a generative task, the model eliminates the need for predefined categories during inference, offering a more practical solution. Our method achieves comparable results with open-vocabulary models on zero-shot LVIS, even though categories are not seen in our GenerateU. Overall, this work contributes to the fast-evolving area of object detection towards open-world scenarios. Future works include investigating the effect of training data scales and exploring methods beyond naive pseudo-labeling." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13965" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Rohan Anil, Andrew M Dai, Orhan First, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, et al. Palm 2 technical report. arXiv preprint arXiv:2305.10403, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 190 + ], + "type": "text", + "content": "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 193, + 288, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 193, + 288, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 193, + 288, + 247 + ], + "type": "text", + "content": "[3] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 249, + 288, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 249, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 53, + 249, + 288, + 304 + ], + "type": "text", + "content": "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 306, + 287, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 306, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 53, + 306, + 287, + 350 + ], + "type": "text", + "content": "[5] Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, and Saehoon Kim. Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset, 2022.5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 351, + 287, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 351, + 287, + 395 + ], + "spans": [ + { + "bbox": [ + 53, + 351, + 287, + 395 + ], + "type": "text", + "content": "[6] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6154-6162, 2018. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 396, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 396, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 396, + 287, + 441 + ], + "type": "text", + "content": "[7] Zhaowei Cai, Gukyeong Kwon, Avinash Ravichandran, Erhan Bas, Zhuowen Tu, Rahul Bhotika, and Stefano Soatto. X-detr: A versatile architecture for instance-wise vision-language tasks. arXiv preprint arXiv:2204.05626, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 442, + 287, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 442, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 53, + 442, + 287, + 495 + ], + "type": "text", + "content": "[8] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 1, 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 498, + 287, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 498, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 53, + 498, + 287, + 542 + ], + "type": "text", + "content": "[9] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 1, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 544, + 287, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 597 + ], + "type": "text", + "content": "[10] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416, 2022. 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 600, + 287, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 600, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 600, + 287, + 643 + ], + "type": "text", + "content": "[11] Achal Dave, Piotr Dollar, Deva Ramanan, Alexander Kirilov, and Ross Girshick. Evaluating large-vocabulary object detectors: The devil is in the details. arXiv preprint arXiv:2102.01066, 2021. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "type": "text", + "content": "[12] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 6, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 712 + ], + "type": "text", + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner," + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 116 + ], + "type": "text", + "content": "Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 119, + 545, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 119, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 119, + 545, + 163 + ], + "type": "text", + "content": "[14] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303-338, 2010. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 165, + 545, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 165, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 307, + 165, + 545, + 219 + ], + "type": "text", + "content": "[15] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xi-angyu Yue, et al. Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010, 2023. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 222, + 545, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 222, + 545, + 265 + ], + "spans": [ + { + "bbox": [ + 307, + 222, + 545, + 265 + ], + "type": "text", + "content": "[16] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 268, + 545, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 268, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 307, + 268, + 545, + 312 + ], + "type": "text", + "content": "[17] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 1, 6, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 315, + 545, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 315, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 307, + 315, + 545, + 357 + ], + "type": "text", + "content": "[18] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE International Conference on Computer Vision, pages 2961-2969, 2017. 1, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 360, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 360, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 360, + 545, + 403 + ], + "type": "text", + "content": "[19] Dat Huynh, Jason Kuen, Zhe Lin, Jiumiang Gu, and Ehsan Elhamifar. Open-vocabulary instance segmentation via robust cross-modal pseudo-labeling. arXiv preprint arXiv:2111.12698, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 406, + 545, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 406, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 307, + 406, + 545, + 471 + ], + "type": "text", + "content": "[20] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 475, + 545, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 475, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 307, + 475, + 545, + 518 + ], + "type": "text", + "content": "[21] Justin Johnson, Andrej Karpathy, and Li Fei-Fei. Densecap: Fully convolutional localization networks for dense captioning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4565–4574, 2016. 3, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 520, + 545, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 520, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 307, + 520, + 545, + 575 + ], + "type": "text", + "content": "[22] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. MDetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1780-1790, 2021. 1, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 578, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 578, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 307, + 578, + 545, + 610 + ], + "type": "text", + "content": "[23] Prannay Kaul, Weidi Xie, and Andrew Zisserman. Multimodal classifiers for open-vocabulary object detection. arXiv preprint arXiv:2306.05493, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 613, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 656 + ], + "type": "text", + "content": "[24] Dahun Kim, Anelia Angelova, and Weicheng Kuo. Contrastive feature masking open-vocabulary vision transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15602-15612, 2023." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "text", + "content": "[25] Dahun Kim, Anelia Angelova, and Weicheng Kuo. Region-aware pretraining for open-vocabulary object detection with vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11144-11154, 2023. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "13966" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 138 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 138 + ], + "type": "text", + "content": "[26] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123(1):32-73, 2017. 1, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 139, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 139, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 139, + 287, + 183 + ], + "type": "text", + "content": "[27] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 184, + 288, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 288, + 237 + ], + "type": "text", + "content": "[28] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. arXiv preprint arXiv:2112.03857, 2021. 1, 4, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 238, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 238, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 238, + 288, + 282 + ], + "type": "text", + "content": "[29] Chuang Lin, Peize Sun, Yi Jiang, Ping Luo, Lizhen Qu, Gholamreza Haffari, Zehuan Yuan, and Jianfei Cai. Learning object-language alignments for open-vocabulary object detection. arXiv preprint arXiv:2211.14843, 2022. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 283, + 287, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 283, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 48, + 283, + 287, + 315 + ], + "type": "text", + "content": "[30] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 316, + 288, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 288, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 288, + 370 + ], + "type": "text", + "content": "[31] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 371, + 287, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 435 + ], + "type": "text", + "content": "[32] Yanxin Long, Youpeng Wen, Jianhua Han, Hang Xu, Pengzhen Ren, Wei Zhang, Shen Zhao, and Xiaodan Liang. Capdet: Unifying dense captioning and open-world detection pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15233-15243, 2023. 3, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 437, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 287, + 459 + ], + "type": "text", + "content": "[33] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 460, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 287, + 502 + ], + "type": "text", + "content": "[34] Chuofan Ma, Yi Jiang, Xin Wen, Zehuan Yuan, and Xiaojuan Qi. Codet: Co-occurrence guided region-word alignment for open-vocabulary object detection. arXiv preprint arXiv:2310.16667, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 503, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 503, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 503, + 287, + 568 + ], + "type": "text", + "content": "[35] Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, et al. Simple open-vocabulary object detection with vision transformers. arXiv preprint arXiv:2205.06230, 2022. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 570, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 602 + ], + "type": "text", + "content": "[36] Baolin Peng, Chunyuan Li, Pengcheng He, Michel Galley, and Jianfeng Gao. Instruction tuning with gpt-4. arXiv preprint arXiv:2304.03277, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 603, + 287, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 287, + 647 + ], + "type": "text", + "content": "[37] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "type": "text", + "content": "[38] Renjie Pi, Jiahui Gao, Shizhe Diao, Rui Pan, Hanze Dong, Jipeng Zhang, Lewei Yao, Jianhua Han, Hang Xu, and Lingpeng Kong Tong Zhang. Detgpt: Detect what you need via reasoning. arXiv preprint arXiv:2305.14167, 2023. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "text", + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry," + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 117 + ], + "type": "text", + "content": "Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 4, 6, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "type": "text", + "content": "[40] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. arXiv preprint arXiv:2112.01518, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 172, + 545, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 172, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 307, + 172, + 545, + 216 + ], + "type": "text", + "content": "[41] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In Advances in Neural Information Processing Systems, pages 91-99, 2015. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 217, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 217, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 307, + 217, + 545, + 271 + ], + "type": "text", + "content": "[42] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019. 1, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 272, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 272, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 307, + 272, + 545, + 316 + ], + "type": "text", + "content": "[43] Cheng Shi and Sibei Yang. Edadet: Open-vocabulary object detection using early dense alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15724-15734, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 316, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 316, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 307, + 316, + 545, + 360 + ], + "type": "text", + "content": "[44] Hengcan Shi, Munawar Hayat, and Jianfei Cai. Open-vocabulary object detection via scene graph discovery. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4012-4021, 2023." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 361, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 361, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 361, + 545, + 392 + ], + "type": "text", + "content": "[45] Hengcan Shi, Munawar Hayat, and Jianfei Cai. Unified open-vocabulary dense visual prediction. arXiv preprint arXiv:2307.08238, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 393, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 393, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 307, + 393, + 545, + 459 + ], + "type": "text", + "content": "[46] Peize Sun, Rufeng Zhang, Yi Jiang, Tao Kong, Chenfeng Xu, Wei Zhan, Masayoshi Tomizuka, Lei Li, Zehuan Yuan, Changhu Wang, et al. Sparse r-cnn: End-to-end object detection with learnable proposals. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14454-14463, 2021. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 459, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 459, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 307, + 459, + 545, + 514 + ], + "type": "text", + "content": "[47] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 514, + 545, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 514, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 307, + 514, + 545, + 569 + ], + "type": "text", + "content": "[48] Luting Wang, Yi Liu, Penghui Du, Zihan Ding, Yue Liao, Qiaosong Qi, Biaolong Chen, and Si Liu. Object-aware distillation pyramid for open-vocabulary object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11186-11196, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 570, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 570, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 307, + 570, + 545, + 613 + ], + "type": "text", + "content": "[49] Tao Wang. Learning to detect and segment for open vocabulary object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7051-7060, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 613, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 669 + ], + "type": "text", + "content": "[50] BigScience Workshop, Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ilic, Daniel Hesslow, Roman Castagné, Alexandra Sasha Luccioni, François Yvon, et al. Bloom: A 176b-parameter open-access multilingual language model. arXiv preprint arXiv:2211.05100, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "type": "text", + "content": "[51] Jialian Wu, Jianfeng Wang, Zhengyuan Yang, Zhe Gan, Zicheng Liu, Junsong Yuan, and Lijuan Wang. Grit: A generative region-to-text transformer for object understanding. arXiv preprint arXiv:2212.00280, 2022. 3, 5" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13967" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[52] Jianzong Wu, Xiangtai Li, Henghui Ding, Xia Li, Guangliang Cheng, Yunhai Tong, and Chen Change Loy. Betrayed by captions: Joint caption grounding and generation for open vocabulary instance segmentation. arXiv preprint arXiv:2301.00805, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 288, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 185 + ], + "type": "text", + "content": "[53] Size Wu, Wenwei Zhang, Sheng Jin, Wentao Liu, and Chen Change Loy. Aligning bag of regions for open-vocabulary object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15254-15264, 2023." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 186, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 288, + 239 + ], + "type": "text", + "content": "[54] Xiaoshi Wu, Feng Zhu, Rui Zhao, and Hongsheng Li. Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7031-7040, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 241, + 288, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 288, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 288, + 297 + ], + "type": "text", + "content": "[55] Bin Yan, Yi Jiang, Jiannan Wu, Dong Wang, Ping Luo, Zehuan Yuan, and Huchuan Lu. Universal instance perception as object discovery and retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15325-15336, 2023. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 298, + 287, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 287, + 353 + ], + "type": "text", + "content": "[56] Lewei Yao, Jianhua Han, Youpeng Wen, Xiaodan Liang, Dan Xu, Wei Zhang, Zhenguo Li, Chunjing Xu, and Hang Xu. Detclip: Dictionary-enriched visual-concept paralleled pretraining for open-world detection. Advances in Neural Information Processing Systems, 35:9125-9138, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 354, + 287, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 419 + ], + "type": "text", + "content": "[57] Lewei Yao, Jianhua Han, Xiaodan Liang, Dan Xu, Wei Zhang, Zhenguo Li, and Hang Xu. Detclipv2: Scalable open-vocabulary object detection pre-training via word-region alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23497-23506, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 421, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 421, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 421, + 287, + 464 + ], + "type": "text", + "content": "[58] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 466, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 466, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 466, + 287, + 521 + ], + "type": "text", + "content": "[59] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 522, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 287, + 567 + ], + "type": "text", + "content": "[60] Yuhang Zang, Wei Li, Kaiyang Zhou, Chen Huang, and Chen Change Loy. Open-vocabulary detr with conditional matching. In European Conference on Computer Vision, pages 106-122. Springer, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 568, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 621 + ], + "type": "text", + "content": "[61] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-vocabulary object detection using captions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14393-14402, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "text", + "content": "[62] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. arXiv preprint arXiv:2111.07991, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "text", + "content": "[63] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In Proceedings of the" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 217 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16793-16803, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 96, + 547, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 128 + ], + "type": "text", + "content": "[64] Chong Zhou, Chen Change Loy, and Bo Dai. Denseclip: Extract free dense labels from clip. arXiv preprint arXiv:2112.01071, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 547, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 173 + ], + "type": "text", + "content": "[65] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Phillip Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. arXiv preprint arXiv:2201.02605, 2022. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 175, + 547, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 547, + 217 + ], + "type": "text", + "content": "[66] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020.3, 6" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "13968" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Generative Rendering_ Controllable 4D-Guided Video Generation with 2D Diffusion Models/038dcdfb-1eda-4a35-919e-6cd09ccad282_content_list.json b/2024/Generative Rendering_ Controllable 4D-Guided Video Generation with 2D Diffusion Models/038dcdfb-1eda-4a35-919e-6cd09ccad282_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d350c9dd0aa98a5b3d2a9d64d644cc18f0344670 --- /dev/null +++ b/2024/Generative Rendering_ Controllable 4D-Guided Video Generation with 2D Diffusion Models/038dcdfb-1eda-4a35-919e-6cd09ccad282_content_list.json @@ -0,0 +1,1467 @@ +[ + { + "type": "text", + "text": "Generative Rendering: Controllable 4D-Guided Video Generation with 2D Diffusion Models", + "text_level": 1, + "bbox": [ + 109, + 128, + 859, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shengqu Cai $^{1,2,\\triangle}$ Duygu Ceylan $^{2*}$ Matheus Gadelha $^{2*}$ Chun-Hao Paul Huang $^{2}$ Tuanfeng Yang Wang $^{2}$ Gordon Wetzstein $^{1}$ Stanford University $^{2}$ Adobe Research", + "bbox": [ + 220, + 202, + 767, + 257 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/99b9d01934059d1cd0df4a1c6db8b3593326c3ff3b269d1cdb334d9ba6487fe6.jpg", + "image_caption": [ + "Figure 1. Generative Rendering is a novel diffusion-based approach to render 3D animated but textured scenes (left) directly to a stylized animation (right), with styles specified by text prompts. Our approach offers new levels of user control to image generation models." + ], + "image_footnote": [], + "bbox": [ + 78, + 290, + 895, + 435 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 492, + 313, + 507 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Traditional 3D content creation tools empower users to bring their imagination to life by giving them direct control over a scene's geometry, appearance, motion, and camera path. Creating computer-generated videos, however, is a tedious manual process, which can be automated by emerging text-to-video diffusion models. Despite great promise, video diffusion models are difficult to control, hindering a user to apply their own creativity rather than amplifying it. To address this challenge, we present a novel approach that combines the controllability of dynamic 3D meshes with the expressivity and editability of emerging diffusion models. For this purpose, our approach takes an animated, low-fidelity rendered mesh as input and injects the ground truth correspondence information obtained from the dynamic mesh into various stages of a pre-trained text-to-image generation model to output high-quality and temporally consistent frames. We demonstrate our approach on various examples where motion can be obtained by animating rigged assets or changing the camera path. Project page: primecai.github.io/generative_rendering.", + "bbox": [ + 73, + 527, + 473, + 832 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 492, + 630, + 507 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Artists, designers, architects, and other creators rely on traditional 3D content creation tools to render computer-generated videos. Unfortunately, existing 3D workflows are laborious, time consuming, and require expertise. Emerging generative artificial intelligence tools, such as text-to-image (T2I) and text-to-video (T2V) models, solve these issues by automating many of the manual steps of traditional workflows. Video generation, however, is difficult to control in that it is not easily possible to specify scene layout and motion in a temporally consistent manner.", + "bbox": [ + 496, + 518, + 895, + 671 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent approaches have attempted to control diffusion models. For example, ControlNet [39] uses a pre-trained T2I diffusion model and finetunes an adapter network that is conditioned on depth, pose, or edge images to control the layout. This strategy is successful for generating individual frames, but results in flicker for video generation. Other approaches aim at learning the complex types of motions encountered in natural videos directly [3, 9, 11-13, 21, 33, 35, 38, 40]. While successful in generating smooth motions, these approaches are not easily controllable. Finally, video-to-video diffusion models [6, 10, 20, 37] enable video editing and stylization, but they require a high-fidelity video as input, which is not always available.", + "bbox": [ + 496, + 671, + 895, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we aim to combine the power of 3D workflows with T2I models for generating 4D-guided stylized", + "bbox": [ + 500, + 869, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\triangle$ : Part of this work was done during an internship at Adobe Research.", + "bbox": [ + 99, + 875, + 467, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*: Equal contribution.", + "bbox": [ + 101, + 888, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7611", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "animations. Specifically, we leverage 3D tools to rapidly prototype proxy geometry and motion of a scene (e.g., camera paths, physically based simulation, or character animation) while utilizing the T2I generation model as a renderer to output the final stylized animations (see Fig. 1). At the core of our method is the ability to leverage the ground truth 4D spatio-temporal correspondences that can be obtained from an animated 3D scene to guide the image generation. First, we leverage the correspondence information to effectively perform noise initialization which is critical for temporal consistency. Many recent video-to-video stylization methods resort to inverting the source video to obtain consistent noise initialization [6, 10]. However, this is not possible in our setup since we do not have an initial source video and inversion does not work well on untextured renderings of the 3D scene. Instead, we propose to utilize the canonical representation (i.e., UV space) of the 3D scene to initialize random noise which is then projected to each frame of the animation. Next, we enrich the self-attention layers of the image diffusion model which are known to contain spatial appearance information vital for obtaining consistent renderings [6, 37]. Specifically, we propagate both the input and output of self-attention layers across the frames via known correspondences to enforce more consistent results. Finally, similar to previous work [39], we use the depth cues rendered from the 3D scene in conjunction with control models to provide structure guidance. In summary, our framework treats the pre-trained T2I diffusion model as a multi-frame renderer that is able to capture the correspondences while maintaining high-fidelity and consistent generations.", + "bbox": [ + 76, + 90, + 472, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate our method on a variety of scenes and demonstrate its advantages with comparisons to state-of-the-art methods. In summary, our contributions include:", + "bbox": [ + 76, + 545, + 470, + 589 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present a novel framework for 4D-guided animation synthesis utilizing the pre-trained T2I generation models as a multi-frame renderer.", + "- We enhance the self-attention layers of the image generation model by performing correspondence-aware blending of both input and output features to enforce consistent appearance synthesis.", + "- We introduce a UV-space noise initialization mechanism. Combined with the correspondence-aware attention, this enables better consistency across frame generations." + ], + "bbox": [ + 76, + 601, + 468, + 761 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 786, + 217, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we discuss the methods most closely related to ours. A comprehensive review of diffusion models for visual computing can be found in [26].", + "bbox": [ + 76, + 811, + 470, + 858 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video Generation. T2V diffusion models extend T2I models by generating dynamic scenes. The majority of works", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "in this area focus on extending pre-trained T2I diffusion models and finetuning on video data [3, 9, 13, 15]. A typical method of lifting image-based models to videos is to add a time-aware module. This can include 3D convolutions [3], or pseudo-3D convolutions factorized over spatial and temporal dimensions to reduce computational costs [7, 7, 9, 11, 33]. Recently, [11] has shown very compelling video generation results by training a motion module with such temporal layers on top a frozen StableDiffusion model [32]. All of these approaches, however, lack fine-grained structural control of the video generation process in a temporally stable manner. Moreover, training a high-quality T2V model is computationally demanding and challenging, in part due to the limited sizes of available captioned high-quality video datasets. This is part of the reason why existing T2V models are limited to outputting only a few seconds of animation. For these reasons, we build on T2I diffusion models and directly leverage ground truth correspondences of the input meshes to generate highly expressive and temporally consistent animations of arbitrary length.", + "bbox": [ + 496, + 90, + 893, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video Editing and Stylization. Another line of work focuses on video editing and stylization. Given an input video, video editing/stylization aims to change the input video according to a user-defined prompt, while keeping other components harmonious. Earlier work, such as [2], decomposes a video to atlases and performs editing directly on these atlases. This approach obtains perfect consistency, but it is only applicable to certain types of scenes and the decomposition is very slow to compute. Exemplar works based on diffusion models [6, 7, 20, 28, 37] along this path generally rely heavily on extracting temporal information from the input video, e.g., via DDIM inversion [17], then inflating the self-attention modules to attend to multiple frames, and propagating the edited contents across different frames. These techniques typically overfit to video inputs and are not directly applicable to our setting. For this reason, we adopt these baselines ensuring a fair comparison to their core methods, and demonstrate that our approach outperforms them. In addition, as the temporal prior is either extracted from the input video or learned from video training data, these works generally lack 3D awareness and may not generate 3D-aware animations. A recent attempt designed a 4D representation that employs a canonical field and deformation field to each frame, such that one can easily edit the video by changing contents on the canonical field which propagates naturally to the other frames [24]. However, this representation still has a relatively weak generalization ability and tends to only work well on very short video clips.", + "bbox": [ + 496, + 411, + 895, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, while related, none of these video editing approaches operate in the same setting as our approach, which takes a low-fidelity, untextured rendering of a scene as input to generate a video.", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7612", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/88d918ba1f719f64344148808899064c4a0f56b5162ddc819196335eb933f60a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 88, + 359, + 186 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/642b780861f1cafcdc2001c30634eb32eb11b8368da59b4ff0cb131d9ad1aa78.jpg", + "image_caption": [ + "Figure 2. Overview of our pipeline. Our system takes as input a set of UV and depths maps rendered from an animated 3D scene. We use a depth-conditioned ControlNet [39] to generate corresponding frames while using the UV correspondences to preserve the consistency. We initialize the noise in the UV space of each object which we then render into each image (Sec. 3.3). For each diffusion step, we first use extended attention for a set of keyframes and extract their pre- and post-attention features (Sec. 3.1). The post-attention features are projected to the UV space and unified. Finally, all frames are generated using a weighted combination of the outputs of the extended attention with the pre-attention features of the keyframe, and the UV-composed post-attention features from the keyframes (Sec. 3.2)." + ], + "image_footnote": [], + "bbox": [ + 83, + 188, + 359, + 324 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a658fd35eea5eecc17ccf46024ba7a49054856f51be34097d7d5d64cb3cc50cf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 362, + 88, + 883, + 324 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Generation. This class of methods is able to achieve near-perfect 3D consistency by using underlying 3D representations. One line of work [4,8,16] focuses on scene generation using diffusion models. These works typically predict or estimate the scene geometry on the fly using monocular predictors such as MiDaS [30] and use that information to constitute individual generations into a scene. This approach is applicable only in specific domains such as indoor scenes or nature scenes, with limited motion control ability due to depth-scale ambiguity. The closest form of 3D generation to our task is mesh texturing. Given an explicit representation such as a mesh, prior works [5,31,36] demonstrate convincing asset mesh texturization using the powerful inpainting function of Stable Diffusion [32]. Other works [25,27] use Score Distillation [27] to directly optimize a neural radiance field [23], but they suffer from quality degradation potentially due to heavy classifier-free guidance [14]. Follow-up works [22] attempt to combine the strength of score distillation and mesh texturization. Since these methods generate a static texture for a given mesh, they are not able to capture appearance interactions (e.g., reflections or shadows) in case of scene-level content and motions.", + "bbox": [ + 75, + 440, + 472, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 784, + 168, + 799 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The goal of generative rendering is to transfer creator-defined scene-level proxy meshes (without appearance) and motions directly to a convincing animation. Given a 3D scene composed of animated meshes, we render guiding channels that consist of depth maps $\\mathcal{D} = [D^1,\\dots,D^N]$ and UV coordinate maps $\\mathcal{U}\\mathcal{V} = [UV^{1},\\dots,UV^{N}]$ where $N$ is", + "bbox": [ + 75, + 810, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the length of the sequence and each UV map contains the 2D texture coordinates as well as an object ID in its RGB channels. Using a depth-conditioned 2D image generation model such as StableDiffusion [32] with depth ControlNet [39], we aim to transfer the sequence $(D^i,UV^i)$ into a set of generated images that depict a stylized animation. To achieve this goal, we perform feature blending over the input and output features of the self-attention modules of the diffusion model (Sec. 3.2) and also perform UV-guided noise initialization (Sec. 3.3) to improve the temporal consistency. We provide an overview of our framework in Fig. 2 and our overall algorithm in Alg. 1. Next, we provide some background on the self-attention features of a diffusion model before discussing the different parts of our method in detail.", + "bbox": [ + 496, + 440, + 895, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Self-attention feature injection", + "text_level": 1, + "bbox": [ + 500, + 674, + 769, + 690 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent work has demonstrated that self-attention features encode spatial self-similarity [34] and they are effective in terms of maintaining temporal consistency when they are used to perform cross-frame attention [6, 37]. A typical approach is to extend the diffusion network to process multiple frames jointly where they can attend to each other's features. Furthermore, it is possible to manipulate both the input and output features of a self-attention module to further enforce consistency. We denote these operations as pre-attention and post-attention feature injections respectively and provide a description in the following.", + "bbox": [ + 496, + 702, + 893, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pre-attention Feature Injection. An attention module is", + "bbox": [ + 500, + 885, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7613", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "defined as:", + "bbox": [ + 76, + 90, + 151, + 104 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {F} = \\operatorname {A t t n} (\\boldsymbol {Q}; \\boldsymbol {K}; \\boldsymbol {V}) = \\operatorname {S o f t m a x} \\left(\\frac {\\boldsymbol {Q} \\boldsymbol {K} ^ {\\top}}{\\sqrt {d}}\\right) \\cdot \\boldsymbol {V}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 113, + 470, + 148 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $Q, K$ , and $V$ denote the query, key, and value features respectively. In case of a self-attention module, $Q, K$ and $V$ are computed by projecting the same features $f^{(i,l)}$ , via the corresponding attention matrices, i.e.:", + "bbox": [ + 76, + 157, + 468, + 218 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {Q} ^ {(i, l)} = \\mathbf {W} ^ {Q} \\boldsymbol {f} ^ {(i, l)}, \\boldsymbol {K} ^ {(i, l)} = \\mathbf {W} ^ {K} \\boldsymbol {f} ^ {(i, l)}, \\boldsymbol {V} ^ {(i, l)} = \\mathbf {W} ^ {V} \\boldsymbol {f} ^ {(i, l)} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 226, + 470, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $i$ denotes the frame index and $l$ denotes the diffusion step and $\\mathbf{W}^{Q / K / V}$ are weight matrices for query, key, and values, respectively. In other words, for each query feature an attention weight is computed based on the similarity between the query and each key feature. Then, the value features are combined based on this attention weight.", + "bbox": [ + 75, + 260, + 468, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "One can manipulate the input features provided to the self-attention module, specifically the key and value features, which we denote as pre-attention feature injection. As discussed by [10], in case of generating $N$ frames, one naive option is to perform extended attention where the features $\\pmb{f}^{(i,l)}$ of all the frames are concatenated and utilized as key and value pairs:", + "bbox": [ + 75, + 351, + 468, + 455 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\boldsymbol {K} ^ {(i, l)} = \\mathbf {W} ^ {K} \\left[ \\boldsymbol {f} ^ {(1, l)}, \\dots , \\boldsymbol {f} ^ {(N, l)} \\right], \\tag {3} \\\\ \\boldsymbol {V} ^ {(i, l)} = \\mathbf {W} ^ {V} [ \\boldsymbol {f} ^ {(1, l)}, \\dots , \\boldsymbol {f} ^ {(N, l)} ]. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 465, + 468, + 507 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Extended attention is a synchronized operation, where all frames attend to features of each other resulting in a more consistent generation. However, generating multiple frames together incurs a high computational cost, which in practice limits the number of frames that can be processed in a sequence. For this reason, we utilize extended attention only for a subset of keyframes and propose mechanisms to propagate the keyframe features to the rest of the sequence, as discussed in the next section.", + "bbox": [ + 75, + 516, + 468, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Post-attention Feature Injection. While it is possible to manipulate the input of the self-attention layer, another option is to directly manipulate the output, which we denote as post-attention feature injection. In particular, given known correspondences between two frames $i$ and $j$ , one can reproject $F^{(i,l)}$ , the output of the self-attention module of frame $i$ at diffusion step $l$ , to obtain $F^{(j,l)}$ :", + "bbox": [ + 75, + 665, + 470, + 771 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {F} ^ {(j, l)} = \\pi_ {i, j} (\\boldsymbol {F} ^ {(i, l)}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 780, + 468, + 800 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pi_{i,j}$ denotes the re-projection operation, which is enabled by the ground truth pixel-level correspondences between frames given by our UV maps.", + "bbox": [ + 75, + 810, + 468, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While post-attention feature injection is effective in achieving consistency, it tends to create blurry results and also avoids generating of certain interactions between objects", + "bbox": [ + 75, + 854, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(e.g., shadows, reflections) since it overwrites the output the self attention. Hence, as described next, our method combines pre- and post-attention feature injection in an effective way to generate consistent but also high-quality frames.", + "bbox": [ + 496, + 90, + 893, + 152 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. UV-space feature injection", + "text_level": 1, + "bbox": [ + 498, + 159, + 740, + 175 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given that pre- and post-attention features of the self-attention module can be manipulated to enhance consistency, our main goal is to perform this manipulation guided by the ground truth correspondences we have from the 3D scene. Specifically, we utilize the canonical UV space to bring the features from the different frames into correspondence which we denote as UV-space feature blending. Given a set of feature maps for each frame $i$ , we project them to a canonical UV space. In particular, for each texel in the UV space, we find its closest point correspondence in each frame. Having such correspondences from multiple frames, a straightforward approach is to take the (weighted) average to blend the features [10]. However, we find this consistently leads to sub-optimal blurry results. Therefore, to combat over-smoothing of the features, we additionally perform blending sequentially and fill a certain texel with the features of its corresponding pixel in a frame only if it has not been filled before. The final unified texture is then the mean of the inpainted texture and the average texture. After features from all frames are blended, we obtain a UV-space feature map which we denote as $T^{(i,l)}$ . Once such a unified feature map is obtained, we can project it back to each frame to obtain $\\bar{f}^{(i,l)}$ .", + "bbox": [ + 496, + 181, + 893, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the UV-space feature blending mechanism, we next describe how we utilize it during the image generation process. For each diffusion step, we first sample a random batch of $m$ keyframes and perform extended attention on these keyframes, extracting both their pre-attention features $\\pmb{f}_{KF}^{(i,l)}$ and post-attention features $\\pmb{F}_{KF}^{(i,l)}$ . We further composite a UV space feature map, $\\bar{T}_{KF}^{l}$ , by blending $\\pmb{F}_{KF}^{(i,l)}$ as described previously. We then perform the diffusion step on all the frames (in case of keyframes, the diffusion step is simply repeated) in a sequential manner. For each frame, we compose the pre-attention features by concatenating the pre-attention features of the keyframes, $\\mathbf{f}_{KF}^{l} = [f_{KF}^{(1,l)},\\dots,f_{KF}^{(m,l)}]$ , with the pre-attention features of the current frame $\\pmb{f}^{(i,l)}$ resulting in:", + "bbox": [ + 496, + 530, + 893, + 750 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {K} ^ {(i, l)} = \\mathbf {W} ^ {K} \\left[ \\boldsymbol {f} ^ {(i, l)}, \\mathbf {f} _ {K F} ^ {l} \\right], \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 753, + 890, + 781 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {V} ^ {(i, l)} = \\mathbf {W} ^ {V} [ \\mathbf {f} ^ {(i, l)}, \\mathbf {f} _ {K F} ^ {l} ].\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 777, + 785, + 795 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After executing the self-attention module, we blend the output, i.e., post-attention features of the current frame $\\hat{\\pmb{F}}^{(i,l)}$ , with the projection of the blended post-attention features to the current frame $\\bar{\\pmb{F}}^{(i,l)}$ . Hence, the updated output of the module becomes:", + "bbox": [ + 496, + 801, + 893, + 875 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {F} _ {\\text {o u t}} ^ {(i, l)} = \\alpha \\cdot \\bar {\\boldsymbol {F}} ^ {(i, l)} + (1 - \\alpha) \\cdot \\hat {\\boldsymbol {F}} ^ {(i, l)} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 882, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7614", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{\\pmb{F}}^{(i,l)}$ is the output of the inflated self-attention block computed using keys and values in Eq. 5.", + "bbox": [ + 76, + 89, + 468, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. UV noise initialization", + "text_level": 1, + "bbox": [ + 76, + 132, + 284, + 146 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Prior video editing works [6, 10] generally rely on inversion methods, such as DDIM inversion [17], to invert each source frame to the noise space which are then used as initialization for editing. Since source frames in the input video are coherent, this inversion strategy yields relatively coherent noise patterns that in turn helps with consistent editing. However, in our case, since we have an untextured scene, inverting rendered frames that lack appearance does not lead to coherent noisy latents. Another option is to keep the noise fixed across all frames. However, this results in severe texture-sticking issues as we demonstrate in the supplementary material. We seek an alternative to capture temporal correlations in initial noise patterns. Inspired by previous work [9, 19], we leverage the fact that we have a canonical UV space by initializing the noise in this space by creating a Gaussian noise texture. We then project this noise to each frame by utilizing the frame-UV correspondences. As we illustrate in our evaluations, this significantly boosts the coherency of the generated outputs.", + "bbox": [ + 75, + 155, + 472, + 444 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Latent normalization", + "text_level": 1, + "bbox": [ + 76, + 454, + 276, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We notice a color flickering issue, where the overall color distribution and contrast change between frames, even if the motion is limited. Therefore, we apply an AdaIN operation on the predicted latents to make every frame's overall distribution similar to the first frame. The simple AdaIN operation solves this issue for some samples, but we find that calculating the distribution over all pixels is very sensitive to the motions. Therefore, inspired by [1], we calculate the statistics only on the background pixels, which are easy to separate in our setup.", + "bbox": [ + 75, + 478, + 468, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 643, + 209, + 660 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We evaluate our method on scenes with various complexity demonstrating three main types of animation sequences: 1) Camera rotations: given a single static object we create a camera path with 360-rotation around the object; 2) Physical simulation: we prepare a motion sequence by utilizing physically based simulation, e.g., bouncing balls where we use rigid body dynamics for collision handling; 3) Character animation: we create scenes with rigged and animated characters obtained from Mixamo and render them with various dance motions.", + "bbox": [ + 75, + 681, + 468, + 833 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. To our knowledge, no published work directly focuses on the same task our method does. Therefore, we", + "bbox": [ + 76, + 845, + 468, + 876 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1 Generative Rendering", + "text_level": 1, + "bbox": [ + 500, + 90, + 735, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Input: UV maps $\\mathcal{U}\\mathcal{V} = [UV^1,\\dots,UV^N]$ ; depth maps $\\pmb{\\mathcal{D}} = [D^{1},\\dots,D^{N}]$ ; text prompt $c$ ; diffusion model $\\Phi$ ; projection from UV to image $\\pi_{\\mathcal{U}\\mathcal{V}}$ ; $T$ diffusion steps; $k$ keyframes", + "bbox": [ + 500, + 112, + 890, + 159 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nz _ {\\mathbf {U V}} \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I}) \\quad / / \\text {S a m p l e n o i s e i n U V s p a c e}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 166, + 890, + 181 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} = \\mathbf {z} _ {T} = \\pi_ {\\mathbf {u v}} (z _ {\\mathbf {u v}})\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 181, + 648, + 198 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For $t = T,\\dots ,1$ do", + "bbox": [ + 500, + 198, + 638, + 210 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} K F = \\text {s h u f f l e} (\\{i _ {1}, \\dots , i _ {k} \\}) \\\\ \\boldsymbol {f} _ {K F}, \\boldsymbol {F} _ {K F} \\leftarrow \\Phi (\\mathbf {z} _ {K F}, \\boldsymbol {\\mathcal {D}} _ {K F}, c) \\quad / / \\text {S a m p l e k e y f r a m e s} \\\\ \\bar {\\boldsymbol {T}} _ {K F} = \\text {b l e n d} (\\pi_ {\\boldsymbol {u v}} ^ {- 1} (\\boldsymbol {F} _ {K F})) \\quad / / \\text {B l e n d s f e a t u r e s i n U V s p a c e} \\\\ \\mathbf {z} _ {t - 1} = [ ] \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 214, + 890, + 275 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For $z^i$ in $\\mathbf{z}$ do", + "bbox": [ + 517, + 276, + 614, + 287 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\boldsymbol {F}} ^ {i} = \\pi_ {U V ^ {i}} (\\bar {\\boldsymbol {T}} _ {K F})\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 289, + 655, + 305 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t - 1} ^ {i} = \\Phi (z ^ {i}, \\boldsymbol {\\mathcal {D}} _ {K F}, c, \\boldsymbol {f} _ {K F}, \\bar {\\boldsymbol {F}} ^ {i})\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 305, + 746, + 321 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {t - 1}. \\text {a p p e n d} \\left(z _ {t - 1} ^ {i}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 321, + 656, + 335 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} = \\mathbf {z} _ {t - 1}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 337, + 591, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Output: z", + "bbox": [ + 500, + 353, + 573, + 368 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "adapt two recent methods to our setting as baselines. 1) we use the pre-attention feature propagation mechanism of Pix2Video [6] (i.e., each frame attends to an anchor and previous frame) together with depth conditioning, and 2) we apply the extended attention mechanism and token propagation from TokenFlow [10] using nearest neighbor matching. We note that the original TokenFlow method computes correspondences by matching features obtained via inverting original frames. Since this is not meaningful in our case, we instead provide the ground truth correspondences we obtain from 3D. Both Pix2Video and TokenFlow rely on inversion for noise initialization. Since this is not doable in our setting, we instead use identical noise to initialize.", + "bbox": [ + 496, + 398, + 893, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation metrics. Following prior video editing evaluation protocols [3,6,7,10,37], we report two metrics based on CLIP [29]. To measure frame consistency, we compute CLIP image embeddings for each frame of the output and compute the average cosine similarity between all pairs of consecutive video frames. To measure prompt fidelity, we compute the mean CLIP embedding similarity score between each frame of the output and the corresponding input prompt.", + "bbox": [ + 496, + 609, + 893, + 731 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. We use the StableDiffusion [32] v1.5 model and a depth-conditioned ControlNet [39] as our image generator. We generate all results at $512 \\times 512$ resolution using Karras scheduler [18] with 50 denoising steps.", + "bbox": [ + 496, + 744, + 893, + 805 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Results", + "text_level": 1, + "bbox": [ + 500, + 818, + 591, + 832 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Qualitative results. Qualitative comparisons to the baselines are shown in Fig. 3. While Pix2Video [6] produces high-quality frames it shows limited spatio-temporal con", + "bbox": [ + 496, + 854, + 893, + 901 + ], + "page_idx": 4 + }, + { + "type": "footer", + "text": "https://www MIXamo.com/", + "bbox": [ + 99, + 887, + 245, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7615", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9eab1f9c742bbf615015b9eea21104269038310540baaaeb5424c35e3922a66a.jpg", + "image_caption": [ + "a photo realistic basketball bouncing under dreamlike sunlight" + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 483, + 459 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c7b145a57890474794a0209f7c80e5fe9b57f1631d6b5d9848a0bcdb1bcb903e.jpg", + "image_caption": [ + "a hyper realistic fluffy Swarovski blue crystal fox running" + ], + "image_footnote": [], + "bbox": [ + 486, + 88, + 887, + 459 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/34da284c4522893b8ea3d00630986bb7d5a1e1b74c0e43ce970f78d13c12ae56.jpg", + "image_caption": [ + "Figure 3. Qualitative comparisons. We adapt the feature blending mechanisms proposed by Pix2Video [6] and TokenFlow [10] to our setting. While Pix2Video produces reasonable results, it lacks a mechanism to maintain frame-level consistency. TokenFlow uses a token propagation & interpolation mechanism between nearby keyframes, but this results in blurred outputs in our setting.", + "Figure 4. Camera and object rotation. Our algorithm supports camera and object rotations for static scenarios." + ], + "image_footnote": [], + "bbox": [ + 83, + 530, + 887, + 645 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "sistency, because it purely relies on the cross-frame attention mechanism to attend to corresponding features. Token-Flow [10] can produce smoother results, but we find that it consistently produces blurry textures. This is in part due to the lack of DDIM inversion [17], which provides a decent level of feature similarities across frames. In our setting, this DDIM inversion is not meaningful and skipping the self-attention module to directly blend the features of keyframes means averaging less similar features, resulting in blurry appearance. Without accurate DDIM inversion and correct nearest neighbor matching, the blending oversmoothes the features and lead to blurry outputs of low quality. In addition, unlike nearest neighbor correspondences computed", + "bbox": [ + 75, + 691, + 472, + 888 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "using DDIM inversion, UV correspondences are not guaranteed to be an exact match in the low-dimensional feature space, which is only $64 \\times 64$ pixels in our model, especially around object edges. Consequently, a straight blending of the post-attention features also creates blurry results.", + "bbox": [ + 498, + 691, + 893, + 767 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our method can handle camera rotation with reasonable quality in static object-level scenarios, as shown in Fig. 4. In Fig. 5, we provide additional results on the same input meshes but with different prompts. Our method shows robustness across a diverse set of prompts and styles. Our inflated attention modules can capture motion-specific traits, such as shadows (Row 3, Row 7), lighting changes (Row 1), and reasonable deformations that are not modeled by the", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7616", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/62a31c81aead55562e78dc8ee695237fda05241d8a94c0a991a141f4d7a9d304.jpg", + "image_caption": [ + "a shining diamond ball bouncing in a chamber" + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 885, + 167 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5ae8460dad3f11cd0a0712af65f36559a9aedde509b48522e15902f472ea0425.jpg", + "image_caption": [ + "a fox running in a lively green forest" + ], + "image_footnote": [], + "bbox": [ + 83, + 176, + 885, + 255 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a56ffbfad99db634f52080aac741a62ef4f99417af0791059c27e57a8e7e5ae2.jpg", + "image_caption": [ + "the Joker dancing" + ], + "image_footnote": [], + "bbox": [ + 83, + 265, + 885, + 343 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/83cc8ddd47be6de75b625f28ed4bdba481870dcc5a0b26f12d465c5298d17837.jpg", + "image_caption": [ + "a little girl walking around, Van Gogh style" + ], + "image_footnote": [], + "bbox": [ + 83, + 352, + 885, + 431 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7a846c22002c0546bf3281d5e63c540a22fbf2e07e90d2071ec98dcc710032f7.jpg", + "image_caption": [ + "a little girl walking around, Da Vinci style" + ], + "image_footnote": [], + "bbox": [ + 83, + 440, + 885, + 517 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f9f6183bcf9dd417ebb16915e1c8d32ffd097e3a6b2aca8fc16078f4da1d6825.jpg", + "image_caption": [ + "a stormtrooper swimming" + ], + "image_footnote": [], + "bbox": [ + 83, + 527, + 885, + 604 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/907f33d096ce5e1401d8cec552912f50bc0d6561d80ce2351e0c5bd1fdebc2e2.jpg", + "image_caption": [ + "three slendermen dancing", + "Figure 5. Qualitative results. Our method generalizes to a diverse set of input meshes, styles, and prompts." + ], + "image_footnote": [], + "bbox": [ + 83, + 614, + 885, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "input low-fidelity meshes (Row 1 ground regions).", + "bbox": [ + 76, + 750, + 410, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative results. We show quantitative evaluations for frame consistency and prompt fidelity in Tab. 1. We observe that our frame consistency is better than all baselines using T2I models. As expected, Pix2Video [6] suffers more from temporal changes due to the lack of a correspondence injection module. Even though TensorFlow's token propagation encourages smoothness over fidelity, our method still shows better consistency. Only the T2V model Gen1 has a better", + "bbox": [ + 75, + 780, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "frame consistency, which is expected because it is trained on video data. All current T2V models, including Gen1, can only generate a few frames and are limited in other ways. Our approach is an attempt to overcome this deficiency using 2D T2I models exclusively. Fidelity-wise, we are slightly better than Pix2Video [6], because it reuses the set of pre-attention features from previous frames, while we only reuse a portion of the pre-attention features to generate each frame. Since TokenFlow [10]'s token propagation mechanism in", + "bbox": [ + 496, + 750, + 893, + 886 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7617", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e6e68b1da0bc89dc23021d313a39e5e63e302acba770b95f8dce37886bd13aee.jpg", + "image_caption": [ + "Figure 6. Ablation study. We disabled several key building blocks and observe clear artifacts or inconsistencies in all ablations." + ], + "image_footnote": [], + "bbox": [ + 84, + 88, + 483, + 238 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f85826a741fc862bbae0d6869a5fd1a0de4ce57fe9fa26ef49b3ab5c5571bb01.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 88, + 887, + 238 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6f560122ea8d7d1be833d42190bc8929aa3cd044f6864ede6b0abfc0cf9741eb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodFrame Consistency ↑Prompt Fidelity ↑
Per-Frame0.95470.3233
Pix2Video*0.96300.2983
TokenFlow*0.98220.2737
Gen10.99070.3029
Ours0.98450.3227
", + "bbox": [ + 78, + 280, + 475, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 1. Quantitative evaluation. Our approach offers the best frame consistency among approaches using T2I models. As expected, only the proprietary T2V model Gen1 observes better temporal consistency. Our method also shows a higher prompt fidelity than relevant baselines that offer a reasonable amount of temporal consistency. *baseline adapted to our setting.", + "bbox": [ + 75, + 398, + 470, + 483 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "produces over-smoothing effects, its results are blurry for many of the samples, hence achieving lower fidelity scores. Our generative rendering utilizes the benefits of pre-attention inflation to gain overall global structural and semantic consistency, while also combining with post-attention feature warping as local guidance. Combined with our UV noise initialization, we outperform or are on par with all relevant baselines in the proposed setting.", + "bbox": [ + 75, + 508, + 470, + 632 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation studies. We perform a thorough ablation study to verify our design choices, shown in Fig. 6. We study three aspects of our pipeline: 1) comparing with pre-attention injection only, 2) comparing with post-attention injection only, and 3) the effectiveness of UV noise initialization. We observe that directly unifying and projecting the post-attention features without passing through the attention module produces chaotic results (severe artifacts in \"no pre-attn\", the joker's face and clothes regions). This observation is in line with our hypothesis for the poor performance of the adapted TokenFlow mechanism. When the motion is relatively large, geometrically matched features may not be good matches in the feature space. Therefore, directly warping a feature from one frame to another based on the geometrical correspondence has a high chance of not producing reasonable outputs. However, when we combine post-attention feature warping with pre-attention feature injection, the post-attention fea", + "bbox": [ + 75, + 643, + 472, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tures can act as guidance for the next network layer. Without post-attention feature warping, purely pre-attention feature injection does not guarantee pixel-wise consistency since the attention operation merely computes a correspondence based on feature similarities (in “no post-attn”, the joker’s clothes change over frames). Our UV noise initialization is well-aligned with our setup. If we use fixed noise instead, the blending of the features may cause unwanted artifacts (in “fixed noise”, the joker’s face has blending artifacts and is not recognizable in the third image). Our UV noise initialization serves generally and can work beyond our setup. Additional details on ablations are included in the supplementary.", + "bbox": [ + 496, + 282, + 893, + 467 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussion", + "text_level": 1, + "bbox": [ + 500, + 498, + 612, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce Generative Rendering, a novel zero-shot pipeline based on 2D diffusion models for 4D-conditioned animation generation. Our method can animate creator-defined low-fidelity meshes and motion sequences, thereby bypassing steps requiring significant manual labor such as detailing, texturing, physical simulation, etc. The key idea of generative rendering is to utilize the prior within a depth-conditioned 2D diffusion model to provide the basic structure and physical fidelity to make convincing animations. This is accomplished by injecting the correspondences into the diffusion model using a combination of pre- and post-attention feature injection while unifying these features in the UV space. Our model demonstrates better frame consistency and prompt fidelity than relevant baselines.", + "bbox": [ + 496, + 530, + 893, + 742 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Conclusion. Diffusion models have emerged as state-of-the-art generative methods across multiple domains, such as 2D, 3D, and video generation and editing. generative rendering adds unprecedented levels of control to video generation with T2I diffusion models by bridging the gap between the traditional rendering pipeline and diffusion models.", + "bbox": [ + 496, + 760, + 893, + 852 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This project was in part supported by Samsung, Google and Stanford HAI.", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7618", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yuval Alaluf, Daniel Garibi, Or Patashnik, Hadar Averbuch-Elor, and Daniel Cohen-Or. Cross-image attention for zero-shot appearance transfer. In arXiv, 2023. 5", + "[2] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In ECCV, 2022. 2", + "[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 1, 2, 5", + "[4] Shengqu Cai, Eric Ryan Chan, Songyou Peng, Mohamad Shahbazi, Anton Obukhov, Luc Van Gool, and Gordon Wetzstein. Diffdreamer: Towards consistent unsupervised single-view scene extrapolation with conditional diffusion models. In ICCV, 2023. 3", + "[5] Tianshi Cao, Karsten Kreis, Sanja Fidler, Nicholas Sharp, and Kangxue Yin. Texfusion: Synthesizing 3d textures with text-guided image diffusion models. In ICCV, 2023. 3", + "[6] Duygu Ceylan, Chun-Hao Huang, and Niloy J. Mitra. Pix2video: Video editing using image diffusion. In ICCV, 2023. 1, 2, 3, 5, 6, 7", + "[7] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In ICCV, 2023. 2, 5", + "[8] Rafail Fridman, Amit Abecasis, Yoni Kasten, and Tali Dekel. Scenescape: Text-driven consistent scene generation. In arXiv, 2023. 3", + "[9] Songwei Ge, Seungjun Nah, Guilin Liu, Tyler Poon, Andrew Tao, Bryan Catanzaro, David Jacobs, Jia-Bin Huang, Ming-Yu Liu, and Yogesh Balaji. Preserve your own correlation: A noise prior for video diffusion models. In ICCV, 2023. 1, 2, 5", + "[10] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. In arXiv, 2023. 1, 2, 4, 5, 6, 7", + "[11] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. In arXiv, 2023. 1, 2", + "[12] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2022. 1", + "[13] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Image video: High definition video generation with diffusion models. In arXiv, 2022. 1, 2", + "[14] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In arXiv, 2022. 3", + "[15] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. In arXiv, 2022. 2", + "[16] Lukas Hollein, Ang Cao, Andrew Owens, Justin Johnson, and Matthias Nießner. Text2room: Extracting textured 3d meshes from 2d text-to-image models. In ICCV, 2023. 3" + ], + "bbox": [ + 78, + 114, + 470, + 896 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly ddpm noise space: Inversion and manipulations. In arXiv, 2023. 2, 5, 6", + "[18] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In NeurIPS, 2022. 5", + "[19] Michael Kass and Davide Pesare. Coherent noise for nonphotorealistic rendering. In TOG, 2011. 5", + "[20] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. In arXiv, 2023. 1, 2", + "[21] Xin Li, Wenqing Chu, Ye Wu, Weihang Yuan, Fanglong Liu, Qi Zhang, Fu Li, Haocheng Feng, Errui Ding, and Jingdong Wang. Videogen: A reference-guided latent diffusion approach for high definition text-to-video generation, 2023. 1", + "[22] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, 2023. 3", + "[23] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 3", + "[24] Hao Ouyang, Qiuyu Wang, Yuxi Xiao, Qingyan Bai, Juntao Zhang, Kecheng Zheng, Xiaowei Zhou, Qifeng Chen, and Yujun Shen. Codef: Content deformation fields for temporally consistent video processing. In arXiv, 2023. 2", + "[25] Ryan Po and Gordon Wetzstein. Compositional 3d scene generation using locally conditioned diffusion. In arXiv, 2023. 3", + "[26] Ryan Po, Wang Yifan, Vladislav Golyanik, Kfir Aberman, Jonathan T Barron, Amit H Bermano, Eric Ryan Chan, Tali Dekel, Aleksander Holynski, Angjoo Kanazawa, et al. State of the art on diffusion models for visual computing. arXiv preprint arXiv:2310.07204, 2023. 2", + "[27] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 3", + "[28] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. In ICCV, 2023. 2", + "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In CoRR, 2021. 5", + "[30] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. In TPAMI, 2022. 3", + "[31] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. In SIGGRAPH, 2023. 3" + ], + "bbox": [ + 501, + 92, + 890, + 883 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7619", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3, 5", + "[33] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-video: Text-to-video generation without text-video data. In arXiv, 2022. 1, 2", + "[34] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1921-1930, June 2023. 3", + "[35] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 1", + "[36] Tianfu Wang, Menelaos Kanakis, Konrad Schindler, Luc Van" + ], + "bbox": [ + 78, + 90, + 470, + 345 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Gool, and Anton Obukhov. Breathing new life into 3d assets with generative repainting. In BMVC, 2023. 3", + "[37] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In arXiv, 2022. 1, 2, 3, 5", + "[38] Sihyun Yu, Kihyuk Sohn, Subin Kim, and Jinwoo Shin. Video probabilistic diffusion models in projected latent space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18456-18466, 2023. 1", + "[39] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 1, 2, 3, 5", + "[40] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 1" + ], + "bbox": [ + 501, + 90, + 893, + 345 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7620", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Generative Rendering_ Controllable 4D-Guided Video Generation with 2D Diffusion Models/038dcdfb-1eda-4a35-919e-6cd09ccad282_model.json b/2024/Generative Rendering_ Controllable 4D-Guided Video Generation with 2D Diffusion Models/038dcdfb-1eda-4a35-919e-6cd09ccad282_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3c418e732ae3800a6cf15e26ad1e112512c377b9 --- /dev/null +++ b/2024/Generative Rendering_ Controllable 4D-Guided Video Generation with 2D Diffusion Models/038dcdfb-1eda-4a35-919e-6cd09ccad282_model.json @@ -0,0 +1,2002 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.11, + 0.13, + 0.861, + 0.175 + ], + "angle": 0, + "content": "Generative Rendering: Controllable 4D-Guided Video Generation with 2D Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.203, + 0.768, + 0.258 + ], + "angle": 0, + "content": "Shengqu Cai\\(^{1,2,\\triangle}\\) Duygu Ceylan\\(^{2*}\\) Matheus Gadelha\\(^{2*}\\) Chun-Hao Paul Huang\\(^{2}\\) Tuanfeng Yang Wang\\(^{2}\\) Gordon Wetzstein\\(^{1}\\) Stanford University\\(^{2}\\) Adobe Research" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.291, + 0.896, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.443, + 0.893, + 0.472 + ], + "angle": 0, + "content": "Figure 1. Generative Rendering is a novel diffusion-based approach to render 3D animated but textured scenes (left) directly to a stylized animation (right), with styles specified by text prompts. Our approach offers new levels of user control to image generation models." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.493, + 0.314, + 0.508 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.529, + 0.474, + 0.833 + ], + "angle": 0, + "content": "Traditional 3D content creation tools empower users to bring their imagination to life by giving them direct control over a scene's geometry, appearance, motion, and camera path. Creating computer-generated videos, however, is a tedious manual process, which can be automated by emerging text-to-video diffusion models. Despite great promise, video diffusion models are difficult to control, hindering a user to apply their own creativity rather than amplifying it. To address this challenge, we present a novel approach that combines the controllability of dynamic 3D meshes with the expressivity and editability of emerging diffusion models. For this purpose, our approach takes an animated, low-fidelity rendered mesh as input and injects the ground truth correspondence information obtained from the dynamic mesh into various stages of a pre-trained text-to-image generation model to output high-quality and temporally consistent frames. We demonstrate our approach on various examples where motion can be obtained by animating rigged assets or changing the camera path. Project page: primecai.github.io/generative_rendering." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.493, + 0.631, + 0.508 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.52, + 0.896, + 0.672 + ], + "angle": 0, + "content": "Artists, designers, architects, and other creators rely on traditional 3D content creation tools to render computer-generated videos. Unfortunately, existing 3D workflows are laborious, time consuming, and require expertise. Emerging generative artificial intelligence tools, such as text-to-image (T2I) and text-to-video (T2V) models, solve these issues by automating many of the manual steps of traditional workflows. Video generation, however, is difficult to control in that it is not easily possible to specify scene layout and motion in a temporally consistent manner." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.672, + 0.896, + 0.868 + ], + "angle": 0, + "content": "Recent approaches have attempted to control diffusion models. For example, ControlNet [39] uses a pre-trained T2I diffusion model and finetunes an adapter network that is conditioned on depth, pose, or edge images to control the layout. This strategy is successful for generating individual frames, but results in flicker for video generation. Other approaches aim at learning the complex types of motions encountered in natural videos directly [3, 9, 11-13, 21, 33, 35, 38, 40]. While successful in generating smooth motions, these approaches are not easily controllable. Finally, video-to-video diffusion models [6, 10, 20, 37] enable video editing and stylization, but they require a high-fidelity video as input, which is not always available." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.895, + 0.902 + ], + "angle": 0, + "content": "In this paper, we aim to combine the power of 3D workflows with T2I models for generating 4D-guided stylized" + }, + { + "type": "page_footnote", + "bbox": [ + 0.1, + 0.875, + 0.468, + 0.888 + ], + "angle": 0, + "content": "\\(\\triangle\\) : Part of this work was done during an internship at Adobe Research." + }, + { + "type": "page_footnote", + "bbox": [ + 0.102, + 0.889, + 0.222, + 0.9 + ], + "angle": 0, + "content": "*: Equal contribution." + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.871, + 0.468, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7611" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.544 + ], + "angle": 0, + "content": "animations. Specifically, we leverage 3D tools to rapidly prototype proxy geometry and motion of a scene (e.g., camera paths, physically based simulation, or character animation) while utilizing the T2I generation model as a renderer to output the final stylized animations (see Fig. 1). At the core of our method is the ability to leverage the ground truth 4D spatio-temporal correspondences that can be obtained from an animated 3D scene to guide the image generation. First, we leverage the correspondence information to effectively perform noise initialization which is critical for temporal consistency. Many recent video-to-video stylization methods resort to inverting the source video to obtain consistent noise initialization [6, 10]. However, this is not possible in our setup since we do not have an initial source video and inversion does not work well on untextured renderings of the 3D scene. Instead, we propose to utilize the canonical representation (i.e., UV space) of the 3D scene to initialize random noise which is then projected to each frame of the animation. Next, we enrich the self-attention layers of the image diffusion model which are known to contain spatial appearance information vital for obtaining consistent renderings [6, 37]. Specifically, we propagate both the input and output of self-attention layers across the frames via known correspondences to enforce more consistent results. Finally, similar to previous work [39], we use the depth cues rendered from the 3D scene in conjunction with control models to provide structure guidance. In summary, our framework treats the pre-trained T2I diffusion model as a multi-frame renderer that is able to capture the correspondences while maintaining high-fidelity and consistent generations." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.546, + 0.472, + 0.59 + ], + "angle": 0, + "content": "We evaluate our method on a variety of scenes and demonstrate its advantages with comparisons to state-of-the-art methods. In summary, our contributions include:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.602, + 0.469, + 0.647 + ], + "angle": 0, + "content": "- We present a novel framework for 4D-guided animation synthesis utilizing the pre-trained T2I generation models as a multi-frame renderer." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.652, + 0.47, + 0.711 + ], + "angle": 0, + "content": "- We enhance the self-attention layers of the image generation model by performing correspondence-aware blending of both input and output features to enforce consistent appearance synthesis." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.716, + 0.47, + 0.762 + ], + "angle": 0, + "content": "- We introduce a UV-space noise initialization mechanism. Combined with the correspondence-aware attention, this enables better consistency across frame generations." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.602, + 0.47, + 0.762 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.787, + 0.218, + 0.803 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.813, + 0.472, + 0.859 + ], + "angle": 0, + "content": "In this section, we discuss the methods most closely related to ours. A comprehensive review of diffusion models for visual computing can be found in [26]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Video Generation. T2V diffusion models extend T2I models by generating dynamic scenes. The majority of works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.395 + ], + "angle": 0, + "content": "in this area focus on extending pre-trained T2I diffusion models and finetuning on video data [3, 9, 13, 15]. A typical method of lifting image-based models to videos is to add a time-aware module. This can include 3D convolutions [3], or pseudo-3D convolutions factorized over spatial and temporal dimensions to reduce computational costs [7, 7, 9, 11, 33]. Recently, [11] has shown very compelling video generation results by training a motion module with such temporal layers on top a frozen StableDiffusion model [32]. All of these approaches, however, lack fine-grained structural control of the video generation process in a temporally stable manner. Moreover, training a high-quality T2V model is computationally demanding and challenging, in part due to the limited sizes of available captioned high-quality video datasets. This is part of the reason why existing T2V models are limited to outputting only a few seconds of animation. For these reasons, we build on T2I diffusion models and directly leverage ground truth correspondences of the input meshes to generate highly expressive and temporally consistent animations of arbitrary length." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.412, + 0.897, + 0.836 + ], + "angle": 0, + "content": "Video Editing and Stylization. Another line of work focuses on video editing and stylization. Given an input video, video editing/stylization aims to change the input video according to a user-defined prompt, while keeping other components harmonious. Earlier work, such as [2], decomposes a video to atlases and performs editing directly on these atlases. This approach obtains perfect consistency, but it is only applicable to certain types of scenes and the decomposition is very slow to compute. Exemplar works based on diffusion models [6, 7, 20, 28, 37] along this path generally rely heavily on extracting temporal information from the input video, e.g., via DDIM inversion [17], then inflating the self-attention modules to attend to multiple frames, and propagating the edited contents across different frames. These techniques typically overfit to video inputs and are not directly applicable to our setting. For this reason, we adopt these baselines ensuring a fair comparison to their core methods, and demonstrate that our approach outperforms them. In addition, as the temporal prior is either extracted from the input video or learned from video training data, these works generally lack 3D awareness and may not generate 3D-aware animations. A recent attempt designed a 4D representation that employs a canonical field and deformation field to each frame, such that one can easily edit the video by changing contents on the canonical field which propagates naturally to the other frames [24]. However, this representation still has a relatively weak generalization ability and tends to only work well on very short video clips." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In summary, while related, none of these video editing approaches operate in the same setting as our approach, which takes a low-fidelity, untextured rendering of a scene as input to generate a video." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7612" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.36, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.189, + 0.36, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.364, + 0.089, + 0.885, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.331, + 0.895, + 0.415 + ], + "angle": 0, + "content": "Figure 2. Overview of our pipeline. Our system takes as input a set of UV and depths maps rendered from an animated 3D scene. We use a depth-conditioned ControlNet [39] to generate corresponding frames while using the UV correspondences to preserve the consistency. We initialize the noise in the UV space of each object which we then render into each image (Sec. 3.3). For each diffusion step, we first use extended attention for a set of keyframes and extract their pre- and post-attention features (Sec. 3.1). The post-attention features are projected to the UV space and unified. Finally, all frames are generated using a weighted combination of the outputs of the extended attention with the pre-attention features of the keyframe, and the UV-composed post-attention features from the keyframes (Sec. 3.2)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.441, + 0.473, + 0.773 + ], + "angle": 0, + "content": "3D Generation. This class of methods is able to achieve near-perfect 3D consistency by using underlying 3D representations. One line of work [4,8,16] focuses on scene generation using diffusion models. These works typically predict or estimate the scene geometry on the fly using monocular predictors such as MiDaS [30] and use that information to constitute individual generations into a scene. This approach is applicable only in specific domains such as indoor scenes or nature scenes, with limited motion control ability due to depth-scale ambiguity. The closest form of 3D generation to our task is mesh texturing. Given an explicit representation such as a mesh, prior works [5,31,36] demonstrate convincing asset mesh texturization using the powerful inpainting function of Stable Diffusion [32]. Other works [25,27] use Score Distillation [27] to directly optimize a neural radiance field [23], but they suffer from quality degradation potentially due to heavy classifier-free guidance [14]. Follow-up works [22] attempt to combine the strength of score distillation and mesh texturization. Since these methods generate a static texture for a given mesh, they are not able to capture appearance interactions (e.g., reflections or shadows) in case of scene-level content and motions." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.785, + 0.169, + 0.8 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.472, + 0.903 + ], + "angle": 0, + "content": "The goal of generative rendering is to transfer creator-defined scene-level proxy meshes (without appearance) and motions directly to a convincing animation. Given a 3D scene composed of animated meshes, we render guiding channels that consist of depth maps \\(\\mathcal{D} = [D^1,\\dots,D^N]\\) and UV coordinate maps \\(\\mathcal{U}\\mathcal{V} = [UV^{1},\\dots,UV^{N}]\\) where \\(N\\) is" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.441, + 0.896, + 0.653 + ], + "angle": 0, + "content": "the length of the sequence and each UV map contains the 2D texture coordinates as well as an object ID in its RGB channels. Using a depth-conditioned 2D image generation model such as StableDiffusion [32] with depth ControlNet [39], we aim to transfer the sequence \\((D^i,UV^i)\\) into a set of generated images that depict a stylized animation. To achieve this goal, we perform feature blending over the input and output features of the self-attention modules of the diffusion model (Sec. 3.2) and also perform UV-guided noise initialization (Sec. 3.3) to improve the temporal consistency. We provide an overview of our framework in Fig. 2 and our overall algorithm in Alg. 1. Next, we provide some background on the self-attention features of a diffusion model before discussing the different parts of our method in detail." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.675, + 0.771, + 0.691 + ], + "angle": 0, + "content": "3.1. Self-attention feature injection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.895, + 0.87 + ], + "angle": 0, + "content": "Recent work has demonstrated that self-attention features encode spatial self-similarity [34] and they are effective in terms of maintaining temporal consistency when they are used to perform cross-frame attention [6, 37]. A typical approach is to extend the diffusion network to process multiple frames jointly where they can attend to each other's features. Furthermore, it is possible to manipulate both the input and output features of a self-attention module to further enforce consistency. We denote these operations as pre-attention and post-attention feature injections respectively and provide a description in the following." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Pre-attention Feature Injection. An attention module is" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7613" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.152, + 0.105 + ], + "angle": 0, + "content": "defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.114, + 0.471, + 0.15 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {F} = \\operatorname {A t t n} (\\boldsymbol {Q}; \\boldsymbol {K}; \\boldsymbol {V}) = \\operatorname {S o f t m a x} \\left(\\frac {\\boldsymbol {Q} \\boldsymbol {K} ^ {\\top}}{\\sqrt {d}}\\right) \\cdot \\boldsymbol {V}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.158, + 0.47, + 0.219 + ], + "angle": 0, + "content": "where \\(Q, K\\), and \\(V\\) denote the query, key, and value features respectively. In case of a self-attention module, \\(Q, K\\) and \\(V\\) are computed by projecting the same features \\(f^{(i,l)}\\), via the corresponding attention matrices, i.e.:" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.227, + 0.472, + 0.26 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {Q} ^ {(i, l)} = \\mathbf {W} ^ {Q} \\boldsymbol {f} ^ {(i, l)}, \\boldsymbol {K} ^ {(i, l)} = \\mathbf {W} ^ {K} \\boldsymbol {f} ^ {(i, l)}, \\boldsymbol {V} ^ {(i, l)} = \\mathbf {W} ^ {V} \\boldsymbol {f} ^ {(i, l)} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.261, + 0.469, + 0.35 + ], + "angle": 0, + "content": "where \\(i\\) denotes the frame index and \\(l\\) denotes the diffusion step and \\(\\mathbf{W}^{Q / K / V}\\) are weight matrices for query, key, and values, respectively. In other words, for each query feature an attention weight is computed based on the similarity between the query and each key feature. Then, the value features are combined based on this attention weight." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.352, + 0.47, + 0.457 + ], + "angle": 0, + "content": "One can manipulate the input features provided to the self-attention module, specifically the key and value features, which we denote as pre-attention feature injection. As discussed by [10], in case of generating \\(N\\) frames, one naive option is to perform extended attention where the features \\(\\pmb{f}^{(i,l)}\\) of all the frames are concatenated and utilized as key and value pairs:" + }, + { + "type": "equation", + "bbox": [ + 0.164, + 0.467, + 0.47, + 0.508 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\boldsymbol {K} ^ {(i, l)} = \\mathbf {W} ^ {K} \\left[ \\boldsymbol {f} ^ {(1, l)}, \\dots , \\boldsymbol {f} ^ {(N, l)} \\right], \\tag {3} \\\\ \\boldsymbol {V} ^ {(i, l)} = \\mathbf {W} ^ {V} [ \\boldsymbol {f} ^ {(1, l)}, \\dots , \\boldsymbol {f} ^ {(N, l)} ]. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.517, + 0.47, + 0.653 + ], + "angle": 0, + "content": "Extended attention is a synchronized operation, where all frames attend to features of each other resulting in a more consistent generation. However, generating multiple frames together incurs a high computational cost, which in practice limits the number of frames that can be processed in a sequence. For this reason, we utilize extended attention only for a subset of keyframes and propose mechanisms to propagate the keyframe features to the rest of the sequence, as discussed in the next section." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.666, + 0.472, + 0.772 + ], + "angle": 0, + "content": "Post-attention Feature Injection. While it is possible to manipulate the input of the self-attention layer, another option is to directly manipulate the output, which we denote as post-attention feature injection. In particular, given known correspondences between two frames \\( i \\) and \\( j \\), one can reproject \\( F^{(i,l)} \\), the output of the self-attention module of frame \\( i \\) at diffusion step \\( l \\), to obtain \\( F^{(j,l)} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.201, + 0.781, + 0.47, + 0.801 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {F} ^ {(j, l)} = \\pi_ {i, j} (\\boldsymbol {F} ^ {(i, l)}), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.469, + 0.856 + ], + "angle": 0, + "content": "where \\(\\pi_{i,j}\\) denotes the re-projection operation, which is enabled by the ground truth pixel-level correspondences between frames given by our UV maps." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "While post-attention feature injection is effective in achieving consistency, it tends to create blurry results and also avoids generating of certain interactions between objects" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.894, + 0.153 + ], + "angle": 0, + "content": "(e.g., shadows, reflections) since it overwrites the output the self attention. Hence, as described next, our method combines pre- and post-attention feature injection in an effective way to generate consistent but also high-quality frames." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.16, + 0.741, + 0.176 + ], + "angle": 0, + "content": "3.2. UV-space feature injection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.183, + 0.895, + 0.529 + ], + "angle": 0, + "content": "Given that pre- and post-attention features of the self-attention module can be manipulated to enhance consistency, our main goal is to perform this manipulation guided by the ground truth correspondences we have from the 3D scene. Specifically, we utilize the canonical UV space to bring the features from the different frames into correspondence which we denote as UV-space feature blending. Given a set of feature maps for each frame \\( i \\), we project them to a canonical UV space. In particular, for each texel in the UV space, we find its closest point correspondence in each frame. Having such correspondences from multiple frames, a straightforward approach is to take the (weighted) average to blend the features [10]. However, we find this consistently leads to sub-optimal blurry results. Therefore, to combat over-smoothing of the features, we additionally perform blending sequentially and fill a certain texel with the features of its corresponding pixel in a frame only if it has not been filled before. The final unified texture is then the mean of the inpainted texture and the average texture. After features from all frames are blended, we obtain a UV-space feature map which we denote as \\( T^{(i,l)} \\). Once such a unified feature map is obtained, we can project it back to each frame to obtain \\( \\bar{f}^{(i,l)} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.531, + 0.895, + 0.75 + ], + "angle": 0, + "content": "Given the UV-space feature blending mechanism, we next describe how we utilize it during the image generation process. For each diffusion step, we first sample a random batch of \\(m\\) keyframes and perform extended attention on these keyframes, extracting both their pre-attention features \\(\\pmb{f}_{KF}^{(i,l)}\\) and post-attention features \\(\\pmb{F}_{KF}^{(i,l)}\\). We further composite a UV space feature map, \\(\\bar{T}_{KF}^{l}\\), by blending \\(\\pmb{F}_{KF}^{(i,l)}\\) as described previously. We then perform the diffusion step on all the frames (in case of keyframes, the diffusion step is simply repeated) in a sequential manner. For each frame, we compose the pre-attention features by concatenating the pre-attention features of the keyframes, \\(\\mathbf{f}_{KF}^{l} = [f_{KF}^{(1,l)},\\dots,f_{KF}^{(m,l)}]\\), with the pre-attention features of the current frame \\(\\pmb{f}^{(i,l)}\\) resulting in:" + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.755, + 0.892, + 0.782 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {K} ^ {(i, l)} = \\mathbf {W} ^ {K} \\left[ \\boldsymbol {f} ^ {(i, l)}, \\mathbf {f} _ {K F} ^ {l} \\right], \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.609, + 0.778, + 0.786, + 0.796 + ], + "angle": 0, + "content": "\\[\n\\mathbf {V} ^ {(i, l)} = \\mathbf {W} ^ {V} [ \\mathbf {f} ^ {(i, l)}, \\mathbf {f} _ {K F} ^ {l} ].\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.802, + 0.894, + 0.876 + ], + "angle": 0, + "content": "After executing the self-attention module, we blend the output, i.e., post-attention features of the current frame \\(\\hat{\\pmb{F}}^{(i,l)}\\), with the projection of the blended post-attention features to the current frame \\(\\bar{\\pmb{F}}^{(i,l)}\\). Hence, the updated output of the module becomes:" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.883, + 0.892, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {F} _ {\\text {o u t}} ^ {(i, l)} = \\alpha \\cdot \\bar {\\boldsymbol {F}} ^ {(i, l)} + (1 - \\alpha) \\cdot \\hat {\\boldsymbol {F}} ^ {(i, l)} \\tag {6}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7614" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.09, + 0.47, + 0.122 + ], + "angle": 0, + "content": "where \\(\\hat{\\pmb{F}}^{(i,l)}\\) is the output of the inflated self-attention block computed using keys and values in Eq. 5." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.133, + 0.285, + 0.147 + ], + "angle": 0, + "content": "3.3. UV noise initialization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.156, + 0.473, + 0.445 + ], + "angle": 0, + "content": "Prior video editing works [6, 10] generally rely on inversion methods, such as DDIM inversion [17], to invert each source frame to the noise space which are then used as initialization for editing. Since source frames in the input video are coherent, this inversion strategy yields relatively coherent noise patterns that in turn helps with consistent editing. However, in our case, since we have an untextured scene, inverting rendered frames that lack appearance does not lead to coherent noisy latents. Another option is to keep the noise fixed across all frames. However, this results in severe texture-sticking issues as we demonstrate in the supplementary material. We seek an alternative to capture temporal correlations in initial noise patterns. Inspired by previous work [9, 19], we leverage the fact that we have a canonical UV space by initializing the noise in this space by creating a Gaussian noise texture. We then project this noise to each frame by utilizing the frame-UV correspondences. As we illustrate in our evaluations, this significantly boosts the coherency of the generated outputs." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.455, + 0.277, + 0.469 + ], + "angle": 0, + "content": "3.4. Latent normalization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.479, + 0.47, + 0.63 + ], + "angle": 0, + "content": "We notice a color flickering issue, where the overall color distribution and contrast change between frames, even if the motion is limited. Therefore, we apply an AdaIN operation on the predicted latents to make every frame's overall distribution similar to the first frame. The simple AdaIN operation solves this issue for some samples, but we find that calculating the distribution over all pixels is very sensitive to the motions. Therefore, inspired by [1], we calculate the statistics only on the background pixels, which are easy to separate in our setup." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.645, + 0.21, + 0.661 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.683, + 0.47, + 0.834 + ], + "angle": 0, + "content": "Datasets. We evaluate our method on scenes with various complexity demonstrating three main types of animation sequences: 1) Camera rotations: given a single static object we create a camera path with 360-rotation around the object; 2) Physical simulation: we prepare a motion sequence by utilizing physically based simulation, e.g., bouncing balls where we use rigid body dynamics for collision handling; 3) Character animation: we create scenes with rigged and animated characters obtained from Mixamo and render them with various dance motions." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.847, + 0.469, + 0.877 + ], + "angle": 0, + "content": "Baselines. To our knowledge, no published work directly focuses on the same task our method does. Therefore, we" + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.092, + 0.736, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Generative Rendering" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.113, + 0.892, + 0.16 + ], + "angle": 0, + "content": "Input: UV maps \\(\\mathcal{U}\\mathcal{V} = [UV^1,\\dots,UV^N]\\); depth maps \\(\\pmb{\\mathcal{D}} = [D^{1},\\dots,D^{N}]\\); text prompt \\(c\\); diffusion model \\(\\Phi\\); projection from UV to image \\(\\pi_{\\mathcal{U}\\mathcal{V}}\\); \\(T\\) diffusion steps; \\(k\\) keyframes" + }, + { + "type": "equation", + "bbox": [ + 0.501, + 0.167, + 0.892, + 0.183 + ], + "angle": 0, + "content": "\\[\nz _ {\\mathbf {U V}} \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I}) \\quad / / \\text {S a m p l e n o i s e i n U V s p a c e}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.501, + 0.183, + 0.649, + 0.199 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} = \\mathbf {z} _ {T} = \\pi_ {\\mathbf {u v}} (z _ {\\mathbf {u v}})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.199, + 0.64, + 0.212 + ], + "angle": 0, + "content": "For \\(t = T,\\dots ,1\\) do" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.215, + 0.892, + 0.276 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} K F = \\text {s h u f f l e} (\\{i _ {1}, \\dots , i _ {k} \\}) \\\\ \\boldsymbol {f} _ {K F}, \\boldsymbol {F} _ {K F} \\leftarrow \\Phi (\\mathbf {z} _ {K F}, \\boldsymbol {\\mathcal {D}} _ {K F}, c) \\quad / / \\text {S a m p l e k e y f r a m e s} \\\\ \\bar {\\boldsymbol {T}} _ {K F} = \\text {b l e n d} (\\pi_ {\\boldsymbol {u v}} ^ {- 1} (\\boldsymbol {F} _ {K F})) \\quad / / \\text {B l e n d s f e a t u r e s i n U V s p a c e} \\\\ \\mathbf {z} _ {t - 1} = [ ] \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.277, + 0.616, + 0.289 + ], + "angle": 0, + "content": "For \\(z^i\\) in \\(\\mathbf{z}\\) do" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.29, + 0.656, + 0.306 + ], + "angle": 0, + "content": "\\[\n\\bar {\\boldsymbol {F}} ^ {i} = \\pi_ {U V ^ {i}} (\\bar {\\boldsymbol {T}} _ {K F})\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.306, + 0.748, + 0.322 + ], + "angle": 0, + "content": "\\[\nz _ {t - 1} ^ {i} = \\Phi (z ^ {i}, \\boldsymbol {\\mathcal {D}} _ {K F}, c, \\boldsymbol {f} _ {K F}, \\bar {\\boldsymbol {F}} ^ {i})\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.322, + 0.658, + 0.337 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {t - 1}. \\text {a p p e n d} \\left(z _ {t - 1} ^ {i}\\right)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.338, + 0.593, + 0.35 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} = \\mathbf {z} _ {t - 1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.354, + 0.575, + 0.369 + ], + "angle": 0, + "content": "Output: z" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.399, + 0.895, + 0.596 + ], + "angle": 0, + "content": "adapt two recent methods to our setting as baselines. 1) we use the pre-attention feature propagation mechanism of Pix2Video [6] (i.e., each frame attends to an anchor and previous frame) together with depth conditioning, and 2) we apply the extended attention mechanism and token propagation from TokenFlow [10] using nearest neighbor matching. We note that the original TokenFlow method computes correspondences by matching features obtained via inverting original frames. Since this is not meaningful in our case, we instead provide the ground truth correspondences we obtain from 3D. Both Pix2Video and TokenFlow rely on inversion for noise initialization. Since this is not doable in our setting, we instead use identical noise to initialize." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.61, + 0.894, + 0.732 + ], + "angle": 0, + "content": "Evaluation metrics. Following prior video editing evaluation protocols [3,6,7,10,37], we report two metrics based on CLIP [29]. To measure frame consistency, we compute CLIP image embeddings for each frame of the output and compute the average cosine similarity between all pairs of consecutive video frames. To measure prompt fidelity, we compute the mean CLIP embedding similarity score between each frame of the output and the corresponding input prompt." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.745, + 0.894, + 0.806 + ], + "angle": 0, + "content": "Implementation details. We use the StableDiffusion [32] v1.5 model and a depth-conditioned ControlNet [39] as our image generator. We generate all results at \\( 512 \\times 512 \\) resolution using Karras scheduler [18] with 50 denoising steps." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.819, + 0.593, + 0.833 + ], + "angle": 0, + "content": "4.1. Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Qualitative results. Qualitative comparisons to the baselines are shown in Fig. 3. While Pix2Video [6] produces high-quality frames it shows limited spatio-temporal con" + }, + { + "type": "footer", + "bbox": [ + 0.101, + 0.888, + 0.246, + 0.9 + ], + "angle": 0, + "content": "https://www MIXamo.com/" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7615" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.485, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.158, + 0.461, + 0.424, + 0.47 + ], + "angle": 0, + "content": "a photo realistic basketball bouncing under dreamlike sunlight" + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.089, + 0.888, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.573, + 0.461, + 0.816, + 0.47 + ], + "angle": 0, + "content": "a hyper realistic fluffy Swarovski blue crystal fox running" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.475, + 0.893, + 0.517 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparisons. We adapt the feature blending mechanisms proposed by Pix2Video [6] and TokenFlow [10] to our setting. While Pix2Video produces reasonable results, it lacks a mechanism to maintain frame-level consistency. TokenFlow uses a token propagation & interpolation mechanism between nearby keyframes, but this results in blurred outputs in our setting." + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.531, + 0.888, + 0.646 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.152, + 0.651, + 0.816, + 0.665 + ], + "angle": 0, + "content": "Figure 4. Camera and object rotation. Our algorithm supports camera and object rotations for static scenarios." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.692, + 0.473, + 0.889 + ], + "angle": 0, + "content": "sistency, because it purely relies on the cross-frame attention mechanism to attend to corresponding features. Token-Flow [10] can produce smoother results, but we find that it consistently produces blurry textures. This is in part due to the lack of DDIM inversion [17], which provides a decent level of feature similarities across frames. In our setting, this DDIM inversion is not meaningful and skipping the self-attention module to directly blend the features of keyframes means averaging less similar features, resulting in blurry appearance. Without accurate DDIM inversion and correct nearest neighbor matching, the blending oversmoothes the features and lead to blurry outputs of low quality. In addition, unlike nearest neighbor correspondences computed" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.692, + 0.895, + 0.768 + ], + "angle": 0, + "content": "using DDIM inversion, UV correspondences are not guaranteed to be an exact match in the low-dimensional feature space, which is only \\(64 \\times 64\\) pixels in our model, especially around object edges. Consequently, a straight blending of the post-attention features also creates blurry results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Our method can handle camera rotation with reasonable quality in static object-level scenarios, as shown in Fig. 4. In Fig. 5, we provide additional results on the same input meshes but with different prompts. Our method shows robustness across a diverse set of prompts and styles. Our inflated attention modules can capture motion-specific traits, such as shadows (Row 3, Row 7), lighting changes (Row 1), and reasonable deformations that are not modeled by the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7616" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.887, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.349, + 0.168, + 0.621, + 0.178 + ], + "angle": 0, + "content": "a shining diamond ball bouncing in a chamber" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.178, + 0.887, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.379, + 0.255, + 0.592, + 0.266 + ], + "angle": 0, + "content": "a fox running in a lively green forest" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.266, + 0.887, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.432, + 0.343, + 0.539, + 0.353 + ], + "angle": 0, + "content": "the Joker dancing" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.353, + 0.887, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.361, + 0.431, + 0.61, + 0.441 + ], + "angle": 0, + "content": "a little girl walking around, Van Gogh style" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.441, + 0.887, + 0.518 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.365, + 0.518, + 0.606, + 0.529 + ], + "angle": 0, + "content": "a little girl walking around, Da Vinci style" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.529, + 0.887, + 0.606 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.409, + 0.606, + 0.562, + 0.616 + ], + "angle": 0, + "content": "a stormtrooper swimming" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.616, + 0.887, + 0.693 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.407, + 0.693, + 0.564, + 0.704 + ], + "angle": 0, + "content": "three slendermen dancing" + }, + { + "type": "image_caption", + "bbox": [ + 0.165, + 0.71, + 0.802, + 0.724 + ], + "angle": 0, + "content": "Figure 5. Qualitative results. Our method generalizes to a diverse set of input meshes, styles, and prompts." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.75, + 0.411, + 0.765 + ], + "angle": 0, + "content": "input low-fidelity meshes (Row 1 ground regions)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Quantitative results. We show quantitative evaluations for frame consistency and prompt fidelity in Tab. 1. We observe that our frame consistency is better than all baselines using T2I models. As expected, Pix2Video [6] suffers more from temporal changes due to the lack of a correspondence injection module. Even though TensorFlow's token propagation encourages smoothness over fidelity, our method still shows better consistency. Only the T2V model Gen1 has a better" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.887 + ], + "angle": 0, + "content": "frame consistency, which is expected because it is trained on video data. All current T2V models, including Gen1, can only generate a few frames and are limited in other ways. Our approach is an attempt to overcome this deficiency using 2D T2I models exclusively. Fidelity-wise, we are slightly better than Pix2Video [6], because it reuses the set of pre-attention features from previous frames, while we only reuse a portion of the pre-attention features to generate each frame. Since TokenFlow [10]'s token propagation mechanism in" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7617" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.089, + 0.484, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.089, + 0.888, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.105, + 0.243, + 0.862, + 0.258 + ], + "angle": 0, + "content": "Figure 6. Ablation study. We disabled several key building blocks and observe clear artifacts or inconsistencies in all ablations." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.281, + 0.477, + 0.39 + ], + "angle": 0, + "content": "
MethodFrame Consistency ↑Prompt Fidelity ↑
Per-Frame0.95470.3233
Pix2Video*0.96300.2983
TokenFlow*0.98220.2737
Gen10.99070.3029
Ours0.98450.3227
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.399, + 0.472, + 0.484 + ], + "angle": 0, + "content": "Table 1. Quantitative evaluation. Our approach offers the best frame consistency among approaches using T2I models. As expected, only the proprietary T2V model Gen1 observes better temporal consistency. Our method also shows a higher prompt fidelity than relevant baselines that offer a reasonable amount of temporal consistency. *baseline adapted to our setting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.51, + 0.471, + 0.633 + ], + "angle": 0, + "content": "produces over-smoothing effects, its results are blurry for many of the samples, hence achieving lower fidelity scores. Our generative rendering utilizes the benefits of pre-attention inflation to gain overall global structural and semantic consistency, while also combining with post-attention feature warping as local guidance. Combined with our UV noise initialization, we outperform or are on par with all relevant baselines in the proposed setting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Ablation studies. We perform a thorough ablation study to verify our design choices, shown in Fig. 6. We study three aspects of our pipeline: 1) comparing with pre-attention injection only, 2) comparing with post-attention injection only, and 3) the effectiveness of UV noise initialization. We observe that directly unifying and projecting the post-attention features without passing through the attention module produces chaotic results (severe artifacts in \"no pre-attn\", the joker's face and clothes regions). This observation is in line with our hypothesis for the poor performance of the adapted TokenFlow mechanism. When the motion is relatively large, geometrically matched features may not be good matches in the feature space. Therefore, directly warping a feature from one frame to another based on the geometrical correspondence has a high chance of not producing reasonable outputs. However, when we combine post-attention feature warping with pre-attention feature injection, the post-attention fea" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.284, + 0.895, + 0.468 + ], + "angle": 0, + "content": "tures can act as guidance for the next network layer. Without post-attention feature warping, purely pre-attention feature injection does not guarantee pixel-wise consistency since the attention operation merely computes a correspondence based on feature similarities (in “no post-attn”, the joker’s clothes change over frames). Our UV noise initialization is well-aligned with our setup. If we use fixed noise instead, the blending of the features may cause unwanted artifacts (in “fixed noise”, the joker’s face has blending artifacts and is not recognizable in the third image). Our UV noise initialization serves generally and can work beyond our setup. Additional details on ablations are included in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.499, + 0.614, + 0.515 + ], + "angle": 0, + "content": "5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.531, + 0.895, + 0.743 + ], + "angle": 0, + "content": "In this paper, we introduce Generative Rendering, a novel zero-shot pipeline based on 2D diffusion models for 4D-conditioned animation generation. Our method can animate creator-defined low-fidelity meshes and motion sequences, thereby bypassing steps requiring significant manual labor such as detailing, texturing, physical simulation, etc. The key idea of generative rendering is to utilize the prior within a depth-conditioned 2D diffusion model to provide the basic structure and physical fidelity to make convincing animations. This is accomplished by injecting the correspondences into the diffusion model using a combination of pre- and post-attention feature injection while unifying these features in the UV space. Our model demonstrates better frame consistency and prompt fidelity than relevant baselines." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.761, + 0.895, + 0.853 + ], + "angle": 0, + "content": "Conclusion. Diffusion models have emerged as state-of-the-art generative methods across multiple domains, such as 2D, 3D, and video generation and editing. generative rendering adds unprecedented levels of control to video generation with T2I diffusion models by bridging the gap between the traditional rendering pipeline and diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Acknowledgement. This project was in part supported by Samsung, Google and Stanford HAI." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7618" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.472, + 0.156 + ], + "angle": 0, + "content": "[1] Yuval Alaluf, Daniel Garibi, Or Patashnik, Hadar Averbuch-Elor, and Daniel Cohen-Or. Cross-image attention for zero-shot appearance transfer. In arXiv, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.472, + 0.198 + ], + "angle": 0, + "content": "[2] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.472, + 0.253 + ], + "angle": 0, + "content": "[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.255, + 0.472, + 0.323 + ], + "angle": 0, + "content": "[4] Shengqu Cai, Eric Ryan Chan, Songyou Peng, Mohamad Shahbazi, Anton Obukhov, Luc Van Gool, and Gordon Wetzstein. Diffdreamer: Towards consistent unsupervised single-view scene extrapolation with conditional diffusion models. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.325, + 0.472, + 0.366 + ], + "angle": 0, + "content": "[5] Tianshi Cao, Karsten Kreis, Sanja Fidler, Nicholas Sharp, and Kangxue Yin. Texfusion: Synthesizing 3d textures with text-guided image diffusion models. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.367, + 0.472, + 0.407 + ], + "angle": 0, + "content": "[6] Duygu Ceylan, Chun-Hao Huang, and Niloy J. Mitra. Pix2video: Video editing using image diffusion. In ICCV, 2023. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.409, + 0.472, + 0.463 + ], + "angle": 0, + "content": "[7] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In ICCV, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.465, + 0.472, + 0.505 + ], + "angle": 0, + "content": "[8] Rafail Fridman, Amit Abecasis, Yoni Kasten, and Tali Dekel. Scenescape: Text-driven consistent scene generation. In arXiv, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.507, + 0.472, + 0.561 + ], + "angle": 0, + "content": "[9] Songwei Ge, Seungjun Nah, Guilin Liu, Tyler Poon, Andrew Tao, Bryan Catanzaro, David Jacobs, Jia-Bin Huang, Ming-Yu Liu, and Yogesh Balaji. Preserve your own correlation: A noise prior for video diffusion models. In ICCV, 2023. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.563, + 0.472, + 0.603 + ], + "angle": 0, + "content": "[10] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. In arXiv, 2023. 1, 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.604, + 0.472, + 0.658 + ], + "angle": 0, + "content": "[11] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. In arXiv, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.472, + 0.714 + ], + "angle": 0, + "content": "[12] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.716, + 0.472, + 0.784 + ], + "angle": 0, + "content": "[13] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Image video: High definition video generation with diffusion models. In arXiv, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.785, + 0.472, + 0.813 + ], + "angle": 0, + "content": "[14] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In arXiv, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.814, + 0.472, + 0.853 + ], + "angle": 0, + "content": "[15] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. In arXiv, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.855, + 0.472, + 0.897 + ], + "angle": 0, + "content": "[16] Lukas Hollein, Ang Cao, Andrew Owens, Justin Johnson, and Matthias Nießner. Text2room: Extracting textured 3d meshes from 2d text-to-image models. In ICCV, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.897 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[17] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly ddpm noise space: Inversion and manipulations. In arXiv, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[18] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In NeurIPS, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.206 + ], + "angle": 0, + "content": "[19] Michael Kass and Davide Pesare. Coherent noise for nonphotorealistic rendering. In TOG, 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.208, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[20] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. In arXiv, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[21] Xin Li, Wenqing Chu, Ye Wu, Weihang Yuan, Fanglong Liu, Qi Zhang, Fu Li, Haocheng Feng, Errui Ding, and Jingdong Wang. Videogen: A reference-guided latent diffusion approach for high definition text-to-video generation, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.388 + ], + "angle": 0, + "content": "[22] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[23] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.448, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[24] Hao Ouyang, Qiuyu Wang, Yuxi Xiao, Qingyan Bai, Juntao Zhang, Kecheng Zheng, Xiaowei Zhou, Qifeng Chen, and Yujun Shen. Codef: Content deformation fields for temporally consistent video processing. In arXiv, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[25] Ryan Po and Gordon Wetzstein. Compositional 3d scene generation using locally conditioned diffusion. In arXiv, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.547, + 0.892, + 0.615 + ], + "angle": 0, + "content": "[26] Ryan Po, Wang Yifan, Vladislav Golyanik, Kfir Aberman, Jonathan T Barron, Amit H Bermano, Eric Ryan Chan, Tali Dekel, Aleksander Holynski, Angjoo Kanazawa, et al. State of the art on diffusion models for visual computing. arXiv preprint arXiv:2310.07204, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.617, + 0.892, + 0.657 + ], + "angle": 0, + "content": "[27] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.66, + 0.892, + 0.714 + ], + "angle": 0, + "content": "[28] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.716, + 0.892, + 0.785 + ], + "angle": 0, + "content": "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In CoRR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.787, + 0.892, + 0.842 + ], + "angle": 0, + "content": "[30] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. In TPAMI, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.844, + 0.892, + 0.885 + ], + "angle": 0, + "content": "[31] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. In SIGGRAPH, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.885 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7619" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.146 + ], + "angle": 0, + "content": "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.472, + 0.219 + ], + "angle": 0, + "content": "[33] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-video: Text-to-video generation without text-video data. In arXiv, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.22, + 0.472, + 0.289 + ], + "angle": 0, + "content": "[34] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1921-1930, June 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.29, + 0.472, + 0.332 + ], + "angle": 0, + "content": "[35] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.333, + 0.472, + 0.347 + ], + "angle": 0, + "content": "[36] Tianfu Wang, Menelaos Kanakis, Konrad Schindler, Luc Van" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "Gool, and Anton Obukhov. Breathing new life into 3d assets with generative repainting. In BMVC, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.189 + ], + "angle": 0, + "content": "[37] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In arXiv, 2022. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.894, + 0.248 + ], + "angle": 0, + "content": "[38] Sihyun Yu, Kihyuk Sohn, Subin Kim, and Jinwoo Shin. Video probabilistic diffusion models in projected latent space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18456-18466, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[39] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.292, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[40] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.894, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7620" + } + ] +] \ No newline at end of file diff --git a/2024/Generative Rendering_ Controllable 4D-Guided Video Generation with 2D Diffusion Models/038dcdfb-1eda-4a35-919e-6cd09ccad282_origin.pdf b/2024/Generative Rendering_ Controllable 4D-Guided Video Generation with 2D Diffusion Models/038dcdfb-1eda-4a35-919e-6cd09ccad282_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b7c342afa785dc998dd5400f4c467689d2212cca --- /dev/null +++ b/2024/Generative Rendering_ Controllable 4D-Guided Video Generation with 2D Diffusion Models/038dcdfb-1eda-4a35-919e-6cd09ccad282_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab9cb2da77914d50d27846354838772f7fac79d8168f03e8e0b9ffde4adf425d +size 5096435 diff --git a/2024/Generative Rendering_ Controllable 4D-Guided Video Generation with 2D Diffusion Models/full.md b/2024/Generative Rendering_ Controllable 4D-Guided Video Generation with 2D Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e0c1157793f97f135defef85bb37790550a4d67f --- /dev/null +++ b/2024/Generative Rendering_ Controllable 4D-Guided Video Generation with 2D Diffusion Models/full.md @@ -0,0 +1,291 @@ +# Generative Rendering: Controllable 4D-Guided Video Generation with 2D Diffusion Models + +Shengqu Cai $^{1,2,\triangle}$ Duygu Ceylan $^{2*}$ Matheus Gadelha $^{2*}$ Chun-Hao Paul Huang $^{2}$ Tuanfeng Yang Wang $^{2}$ Gordon Wetzstein $^{1}$ Stanford University $^{2}$ Adobe Research + +![](images/99b9d01934059d1cd0df4a1c6db8b3593326c3ff3b269d1cdb334d9ba6487fe6.jpg) +Figure 1. Generative Rendering is a novel diffusion-based approach to render 3D animated but textured scenes (left) directly to a stylized animation (right), with styles specified by text prompts. Our approach offers new levels of user control to image generation models. + +# Abstract + +Traditional 3D content creation tools empower users to bring their imagination to life by giving them direct control over a scene's geometry, appearance, motion, and camera path. Creating computer-generated videos, however, is a tedious manual process, which can be automated by emerging text-to-video diffusion models. Despite great promise, video diffusion models are difficult to control, hindering a user to apply their own creativity rather than amplifying it. To address this challenge, we present a novel approach that combines the controllability of dynamic 3D meshes with the expressivity and editability of emerging diffusion models. For this purpose, our approach takes an animated, low-fidelity rendered mesh as input and injects the ground truth correspondence information obtained from the dynamic mesh into various stages of a pre-trained text-to-image generation model to output high-quality and temporally consistent frames. We demonstrate our approach on various examples where motion can be obtained by animating rigged assets or changing the camera path. Project page: primecai.github.io/generative_rendering. + +# 1. Introduction + +Artists, designers, architects, and other creators rely on traditional 3D content creation tools to render computer-generated videos. Unfortunately, existing 3D workflows are laborious, time consuming, and require expertise. Emerging generative artificial intelligence tools, such as text-to-image (T2I) and text-to-video (T2V) models, solve these issues by automating many of the manual steps of traditional workflows. Video generation, however, is difficult to control in that it is not easily possible to specify scene layout and motion in a temporally consistent manner. + +Recent approaches have attempted to control diffusion models. For example, ControlNet [39] uses a pre-trained T2I diffusion model and finetunes an adapter network that is conditioned on depth, pose, or edge images to control the layout. This strategy is successful for generating individual frames, but results in flicker for video generation. Other approaches aim at learning the complex types of motions encountered in natural videos directly [3, 9, 11-13, 21, 33, 35, 38, 40]. While successful in generating smooth motions, these approaches are not easily controllable. Finally, video-to-video diffusion models [6, 10, 20, 37] enable video editing and stylization, but they require a high-fidelity video as input, which is not always available. + +In this paper, we aim to combine the power of 3D workflows with T2I models for generating 4D-guided stylized + +animations. Specifically, we leverage 3D tools to rapidly prototype proxy geometry and motion of a scene (e.g., camera paths, physically based simulation, or character animation) while utilizing the T2I generation model as a renderer to output the final stylized animations (see Fig. 1). At the core of our method is the ability to leverage the ground truth 4D spatio-temporal correspondences that can be obtained from an animated 3D scene to guide the image generation. First, we leverage the correspondence information to effectively perform noise initialization which is critical for temporal consistency. Many recent video-to-video stylization methods resort to inverting the source video to obtain consistent noise initialization [6, 10]. However, this is not possible in our setup since we do not have an initial source video and inversion does not work well on untextured renderings of the 3D scene. Instead, we propose to utilize the canonical representation (i.e., UV space) of the 3D scene to initialize random noise which is then projected to each frame of the animation. Next, we enrich the self-attention layers of the image diffusion model which are known to contain spatial appearance information vital for obtaining consistent renderings [6, 37]. Specifically, we propagate both the input and output of self-attention layers across the frames via known correspondences to enforce more consistent results. Finally, similar to previous work [39], we use the depth cues rendered from the 3D scene in conjunction with control models to provide structure guidance. In summary, our framework treats the pre-trained T2I diffusion model as a multi-frame renderer that is able to capture the correspondences while maintaining high-fidelity and consistent generations. + +We evaluate our method on a variety of scenes and demonstrate its advantages with comparisons to state-of-the-art methods. In summary, our contributions include: + +- We present a novel framework for 4D-guided animation synthesis utilizing the pre-trained T2I generation models as a multi-frame renderer. +- We enhance the self-attention layers of the image generation model by performing correspondence-aware blending of both input and output features to enforce consistent appearance synthesis. +- We introduce a UV-space noise initialization mechanism. Combined with the correspondence-aware attention, this enables better consistency across frame generations. + +# 2. Related Work + +In this section, we discuss the methods most closely related to ours. A comprehensive review of diffusion models for visual computing can be found in [26]. + +Video Generation. T2V diffusion models extend T2I models by generating dynamic scenes. The majority of works + +in this area focus on extending pre-trained T2I diffusion models and finetuning on video data [3, 9, 13, 15]. A typical method of lifting image-based models to videos is to add a time-aware module. This can include 3D convolutions [3], or pseudo-3D convolutions factorized over spatial and temporal dimensions to reduce computational costs [7, 7, 9, 11, 33]. Recently, [11] has shown very compelling video generation results by training a motion module with such temporal layers on top a frozen StableDiffusion model [32]. All of these approaches, however, lack fine-grained structural control of the video generation process in a temporally stable manner. Moreover, training a high-quality T2V model is computationally demanding and challenging, in part due to the limited sizes of available captioned high-quality video datasets. This is part of the reason why existing T2V models are limited to outputting only a few seconds of animation. For these reasons, we build on T2I diffusion models and directly leverage ground truth correspondences of the input meshes to generate highly expressive and temporally consistent animations of arbitrary length. + +Video Editing and Stylization. Another line of work focuses on video editing and stylization. Given an input video, video editing/stylization aims to change the input video according to a user-defined prompt, while keeping other components harmonious. Earlier work, such as [2], decomposes a video to atlases and performs editing directly on these atlases. This approach obtains perfect consistency, but it is only applicable to certain types of scenes and the decomposition is very slow to compute. Exemplar works based on diffusion models [6, 7, 20, 28, 37] along this path generally rely heavily on extracting temporal information from the input video, e.g., via DDIM inversion [17], then inflating the self-attention modules to attend to multiple frames, and propagating the edited contents across different frames. These techniques typically overfit to video inputs and are not directly applicable to our setting. For this reason, we adopt these baselines ensuring a fair comparison to their core methods, and demonstrate that our approach outperforms them. In addition, as the temporal prior is either extracted from the input video or learned from video training data, these works generally lack 3D awareness and may not generate 3D-aware animations. A recent attempt designed a 4D representation that employs a canonical field and deformation field to each frame, such that one can easily edit the video by changing contents on the canonical field which propagates naturally to the other frames [24]. However, this representation still has a relatively weak generalization ability and tends to only work well on very short video clips. + +In summary, while related, none of these video editing approaches operate in the same setting as our approach, which takes a low-fidelity, untextured rendering of a scene as input to generate a video. + +![](images/88d918ba1f719f64344148808899064c4a0f56b5162ddc819196335eb933f60a.jpg) + +![](images/642b780861f1cafcdc2001c30634eb32eb11b8368da59b4ff0cb131d9ad1aa78.jpg) +Figure 2. Overview of our pipeline. Our system takes as input a set of UV and depths maps rendered from an animated 3D scene. We use a depth-conditioned ControlNet [39] to generate corresponding frames while using the UV correspondences to preserve the consistency. We initialize the noise in the UV space of each object which we then render into each image (Sec. 3.3). For each diffusion step, we first use extended attention for a set of keyframes and extract their pre- and post-attention features (Sec. 3.1). The post-attention features are projected to the UV space and unified. Finally, all frames are generated using a weighted combination of the outputs of the extended attention with the pre-attention features of the keyframe, and the UV-composed post-attention features from the keyframes (Sec. 3.2). + +![](images/a658fd35eea5eecc17ccf46024ba7a49054856f51be34097d7d5d64cb3cc50cf.jpg) + +3D Generation. This class of methods is able to achieve near-perfect 3D consistency by using underlying 3D representations. One line of work [4,8,16] focuses on scene generation using diffusion models. These works typically predict or estimate the scene geometry on the fly using monocular predictors such as MiDaS [30] and use that information to constitute individual generations into a scene. This approach is applicable only in specific domains such as indoor scenes or nature scenes, with limited motion control ability due to depth-scale ambiguity. The closest form of 3D generation to our task is mesh texturing. Given an explicit representation such as a mesh, prior works [5,31,36] demonstrate convincing asset mesh texturization using the powerful inpainting function of Stable Diffusion [32]. Other works [25,27] use Score Distillation [27] to directly optimize a neural radiance field [23], but they suffer from quality degradation potentially due to heavy classifier-free guidance [14]. Follow-up works [22] attempt to combine the strength of score distillation and mesh texturization. Since these methods generate a static texture for a given mesh, they are not able to capture appearance interactions (e.g., reflections or shadows) in case of scene-level content and motions. + +# 3. Method + +The goal of generative rendering is to transfer creator-defined scene-level proxy meshes (without appearance) and motions directly to a convincing animation. Given a 3D scene composed of animated meshes, we render guiding channels that consist of depth maps $\mathcal{D} = [D^1,\dots,D^N]$ and UV coordinate maps $\mathcal{U}\mathcal{V} = [UV^{1},\dots,UV^{N}]$ where $N$ is + +the length of the sequence and each UV map contains the 2D texture coordinates as well as an object ID in its RGB channels. Using a depth-conditioned 2D image generation model such as StableDiffusion [32] with depth ControlNet [39], we aim to transfer the sequence $(D^i,UV^i)$ into a set of generated images that depict a stylized animation. To achieve this goal, we perform feature blending over the input and output features of the self-attention modules of the diffusion model (Sec. 3.2) and also perform UV-guided noise initialization (Sec. 3.3) to improve the temporal consistency. We provide an overview of our framework in Fig. 2 and our overall algorithm in Alg. 1. Next, we provide some background on the self-attention features of a diffusion model before discussing the different parts of our method in detail. + +# 3.1. Self-attention feature injection + +Recent work has demonstrated that self-attention features encode spatial self-similarity [34] and they are effective in terms of maintaining temporal consistency when they are used to perform cross-frame attention [6, 37]. A typical approach is to extend the diffusion network to process multiple frames jointly where they can attend to each other's features. Furthermore, it is possible to manipulate both the input and output features of a self-attention module to further enforce consistency. We denote these operations as pre-attention and post-attention feature injections respectively and provide a description in the following. + +Pre-attention Feature Injection. An attention module is + +defined as: + +$$ +\boldsymbol {F} = \operatorname {A t t n} (\boldsymbol {Q}; \boldsymbol {K}; \boldsymbol {V}) = \operatorname {S o f t m a x} \left(\frac {\boldsymbol {Q} \boldsymbol {K} ^ {\top}}{\sqrt {d}}\right) \cdot \boldsymbol {V}, \tag {1} +$$ + +where $Q, K$ , and $V$ denote the query, key, and value features respectively. In case of a self-attention module, $Q, K$ and $V$ are computed by projecting the same features $f^{(i,l)}$ , via the corresponding attention matrices, i.e.: + +$$ +\boldsymbol {Q} ^ {(i, l)} = \mathbf {W} ^ {Q} \boldsymbol {f} ^ {(i, l)}, \boldsymbol {K} ^ {(i, l)} = \mathbf {W} ^ {K} \boldsymbol {f} ^ {(i, l)}, \boldsymbol {V} ^ {(i, l)} = \mathbf {W} ^ {V} \boldsymbol {f} ^ {(i, l)} \tag {2} +$$ + +where $i$ denotes the frame index and $l$ denotes the diffusion step and $\mathbf{W}^{Q / K / V}$ are weight matrices for query, key, and values, respectively. In other words, for each query feature an attention weight is computed based on the similarity between the query and each key feature. Then, the value features are combined based on this attention weight. + +One can manipulate the input features provided to the self-attention module, specifically the key and value features, which we denote as pre-attention feature injection. As discussed by [10], in case of generating $N$ frames, one naive option is to perform extended attention where the features $\pmb{f}^{(i,l)}$ of all the frames are concatenated and utilized as key and value pairs: + +$$ +\begin{array}{l} \boldsymbol {K} ^ {(i, l)} = \mathbf {W} ^ {K} \left[ \boldsymbol {f} ^ {(1, l)}, \dots , \boldsymbol {f} ^ {(N, l)} \right], \tag {3} \\ \boldsymbol {V} ^ {(i, l)} = \mathbf {W} ^ {V} [ \boldsymbol {f} ^ {(1, l)}, \dots , \boldsymbol {f} ^ {(N, l)} ]. \\ \end{array} +$$ + +Extended attention is a synchronized operation, where all frames attend to features of each other resulting in a more consistent generation. However, generating multiple frames together incurs a high computational cost, which in practice limits the number of frames that can be processed in a sequence. For this reason, we utilize extended attention only for a subset of keyframes and propose mechanisms to propagate the keyframe features to the rest of the sequence, as discussed in the next section. + +Post-attention Feature Injection. While it is possible to manipulate the input of the self-attention layer, another option is to directly manipulate the output, which we denote as post-attention feature injection. In particular, given known correspondences between two frames $i$ and $j$ , one can reproject $F^{(i,l)}$ , the output of the self-attention module of frame $i$ at diffusion step $l$ , to obtain $F^{(j,l)}$ : + +$$ +\boldsymbol {F} ^ {(j, l)} = \pi_ {i, j} (\boldsymbol {F} ^ {(i, l)}), \tag {4} +$$ + +where $\pi_{i,j}$ denotes the re-projection operation, which is enabled by the ground truth pixel-level correspondences between frames given by our UV maps. + +While post-attention feature injection is effective in achieving consistency, it tends to create blurry results and also avoids generating of certain interactions between objects + +(e.g., shadows, reflections) since it overwrites the output the self attention. Hence, as described next, our method combines pre- and post-attention feature injection in an effective way to generate consistent but also high-quality frames. + +# 3.2. UV-space feature injection + +Given that pre- and post-attention features of the self-attention module can be manipulated to enhance consistency, our main goal is to perform this manipulation guided by the ground truth correspondences we have from the 3D scene. Specifically, we utilize the canonical UV space to bring the features from the different frames into correspondence which we denote as UV-space feature blending. Given a set of feature maps for each frame $i$ , we project them to a canonical UV space. In particular, for each texel in the UV space, we find its closest point correspondence in each frame. Having such correspondences from multiple frames, a straightforward approach is to take the (weighted) average to blend the features [10]. However, we find this consistently leads to sub-optimal blurry results. Therefore, to combat over-smoothing of the features, we additionally perform blending sequentially and fill a certain texel with the features of its corresponding pixel in a frame only if it has not been filled before. The final unified texture is then the mean of the inpainted texture and the average texture. After features from all frames are blended, we obtain a UV-space feature map which we denote as $T^{(i,l)}$ . Once such a unified feature map is obtained, we can project it back to each frame to obtain $\bar{f}^{(i,l)}$ . + +Given the UV-space feature blending mechanism, we next describe how we utilize it during the image generation process. For each diffusion step, we first sample a random batch of $m$ keyframes and perform extended attention on these keyframes, extracting both their pre-attention features $\pmb{f}_{KF}^{(i,l)}$ and post-attention features $\pmb{F}_{KF}^{(i,l)}$ . We further composite a UV space feature map, $\bar{T}_{KF}^{l}$ , by blending $\pmb{F}_{KF}^{(i,l)}$ as described previously. We then perform the diffusion step on all the frames (in case of keyframes, the diffusion step is simply repeated) in a sequential manner. For each frame, we compose the pre-attention features by concatenating the pre-attention features of the keyframes, $\mathbf{f}_{KF}^{l} = [f_{KF}^{(1,l)},\dots,f_{KF}^{(m,l)}]$ , with the pre-attention features of the current frame $\pmb{f}^{(i,l)}$ resulting in: + +$$ +\boldsymbol {K} ^ {(i, l)} = \mathbf {W} ^ {K} \left[ \boldsymbol {f} ^ {(i, l)}, \mathbf {f} _ {K F} ^ {l} \right], \tag {5} +$$ + +$$ +\mathbf {V} ^ {(i, l)} = \mathbf {W} ^ {V} [ \mathbf {f} ^ {(i, l)}, \mathbf {f} _ {K F} ^ {l} ]. +$$ + +After executing the self-attention module, we blend the output, i.e., post-attention features of the current frame $\hat{\pmb{F}}^{(i,l)}$ , with the projection of the blended post-attention features to the current frame $\bar{\pmb{F}}^{(i,l)}$ . Hence, the updated output of the module becomes: + +$$ +\boldsymbol {F} _ {\text {o u t}} ^ {(i, l)} = \alpha \cdot \bar {\boldsymbol {F}} ^ {(i, l)} + (1 - \alpha) \cdot \hat {\boldsymbol {F}} ^ {(i, l)} \tag {6} +$$ + +where $\hat{\pmb{F}}^{(i,l)}$ is the output of the inflated self-attention block computed using keys and values in Eq. 5. + +# 3.3. UV noise initialization + +Prior video editing works [6, 10] generally rely on inversion methods, such as DDIM inversion [17], to invert each source frame to the noise space which are then used as initialization for editing. Since source frames in the input video are coherent, this inversion strategy yields relatively coherent noise patterns that in turn helps with consistent editing. However, in our case, since we have an untextured scene, inverting rendered frames that lack appearance does not lead to coherent noisy latents. Another option is to keep the noise fixed across all frames. However, this results in severe texture-sticking issues as we demonstrate in the supplementary material. We seek an alternative to capture temporal correlations in initial noise patterns. Inspired by previous work [9, 19], we leverage the fact that we have a canonical UV space by initializing the noise in this space by creating a Gaussian noise texture. We then project this noise to each frame by utilizing the frame-UV correspondences. As we illustrate in our evaluations, this significantly boosts the coherency of the generated outputs. + +# 3.4. Latent normalization + +We notice a color flickering issue, where the overall color distribution and contrast change between frames, even if the motion is limited. Therefore, we apply an AdaIN operation on the predicted latents to make every frame's overall distribution similar to the first frame. The simple AdaIN operation solves this issue for some samples, but we find that calculating the distribution over all pixels is very sensitive to the motions. Therefore, inspired by [1], we calculate the statistics only on the background pixels, which are easy to separate in our setup. + +# 4. Experiments + +Datasets. We evaluate our method on scenes with various complexity demonstrating three main types of animation sequences: 1) Camera rotations: given a single static object we create a camera path with 360-rotation around the object; 2) Physical simulation: we prepare a motion sequence by utilizing physically based simulation, e.g., bouncing balls where we use rigid body dynamics for collision handling; 3) Character animation: we create scenes with rigged and animated characters obtained from Mixamo and render them with various dance motions. + +Baselines. To our knowledge, no published work directly focuses on the same task our method does. Therefore, we + +# Algorithm 1 Generative Rendering + +Input: UV maps $\mathcal{U}\mathcal{V} = [UV^1,\dots,UV^N]$ ; depth maps $\pmb{\mathcal{D}} = [D^{1},\dots,D^{N}]$ ; text prompt $c$ ; diffusion model $\Phi$ ; projection from UV to image $\pi_{\mathcal{U}\mathcal{V}}$ ; $T$ diffusion steps; $k$ keyframes + +$$ +z _ {\mathbf {U V}} \sim \mathcal {N} (\mathbf {0}, \mathbf {I}) \quad / / \text {S a m p l e n o i s e i n U V s p a c e} +$$ + +$$ +\mathbf {z} = \mathbf {z} _ {T} = \pi_ {\mathbf {u v}} (z _ {\mathbf {u v}}) +$$ + +For $t = T,\dots ,1$ do + +$$ +\begin{array}{l} K F = \text {s h u f f l e} (\{i _ {1}, \dots , i _ {k} \}) \\ \boldsymbol {f} _ {K F}, \boldsymbol {F} _ {K F} \leftarrow \Phi (\mathbf {z} _ {K F}, \boldsymbol {\mathcal {D}} _ {K F}, c) \quad / / \text {S a m p l e k e y f r a m e s} \\ \bar {\boldsymbol {T}} _ {K F} = \text {b l e n d} (\pi_ {\boldsymbol {u v}} ^ {- 1} (\boldsymbol {F} _ {K F})) \quad / / \text {B l e n d s f e a t u r e s i n U V s p a c e} \\ \mathbf {z} _ {t - 1} = [ ] \end{array} +$$ + +For $z^i$ in $\mathbf{z}$ do + +$$ +\bar {\boldsymbol {F}} ^ {i} = \pi_ {U V ^ {i}} (\bar {\boldsymbol {T}} _ {K F}) +$$ + +$$ +z _ {t - 1} ^ {i} = \Phi (z ^ {i}, \boldsymbol {\mathcal {D}} _ {K F}, c, \boldsymbol {f} _ {K F}, \bar {\boldsymbol {F}} ^ {i}) +$$ + +$$ +\mathbf {z} _ {t - 1}. \text {a p p e n d} \left(z _ {t - 1} ^ {i}\right) +$$ + +$$ +\mathbf {z} = \mathbf {z} _ {t - 1} +$$ + +Output: z + +adapt two recent methods to our setting as baselines. 1) we use the pre-attention feature propagation mechanism of Pix2Video [6] (i.e., each frame attends to an anchor and previous frame) together with depth conditioning, and 2) we apply the extended attention mechanism and token propagation from TokenFlow [10] using nearest neighbor matching. We note that the original TokenFlow method computes correspondences by matching features obtained via inverting original frames. Since this is not meaningful in our case, we instead provide the ground truth correspondences we obtain from 3D. Both Pix2Video and TokenFlow rely on inversion for noise initialization. Since this is not doable in our setting, we instead use identical noise to initialize. + +Evaluation metrics. Following prior video editing evaluation protocols [3,6,7,10,37], we report two metrics based on CLIP [29]. To measure frame consistency, we compute CLIP image embeddings for each frame of the output and compute the average cosine similarity between all pairs of consecutive video frames. To measure prompt fidelity, we compute the mean CLIP embedding similarity score between each frame of the output and the corresponding input prompt. + +Implementation details. We use the StableDiffusion [32] v1.5 model and a depth-conditioned ControlNet [39] as our image generator. We generate all results at $512 \times 512$ resolution using Karras scheduler [18] with 50 denoising steps. + +# 4.1. Results + +Qualitative results. Qualitative comparisons to the baselines are shown in Fig. 3. While Pix2Video [6] produces high-quality frames it shows limited spatio-temporal con + +![](images/9eab1f9c742bbf615015b9eea21104269038310540baaaeb5424c35e3922a66a.jpg) +a photo realistic basketball bouncing under dreamlike sunlight + +![](images/c7b145a57890474794a0209f7c80e5fe9b57f1631d6b5d9848a0bcdb1bcb903e.jpg) +a hyper realistic fluffy Swarovski blue crystal fox running + +![](images/34da284c4522893b8ea3d00630986bb7d5a1e1b74c0e43ce970f78d13c12ae56.jpg) +Figure 3. Qualitative comparisons. We adapt the feature blending mechanisms proposed by Pix2Video [6] and TokenFlow [10] to our setting. While Pix2Video produces reasonable results, it lacks a mechanism to maintain frame-level consistency. TokenFlow uses a token propagation & interpolation mechanism between nearby keyframes, but this results in blurred outputs in our setting. +Figure 4. Camera and object rotation. Our algorithm supports camera and object rotations for static scenarios. + +sistency, because it purely relies on the cross-frame attention mechanism to attend to corresponding features. Token-Flow [10] can produce smoother results, but we find that it consistently produces blurry textures. This is in part due to the lack of DDIM inversion [17], which provides a decent level of feature similarities across frames. In our setting, this DDIM inversion is not meaningful and skipping the self-attention module to directly blend the features of keyframes means averaging less similar features, resulting in blurry appearance. Without accurate DDIM inversion and correct nearest neighbor matching, the blending oversmoothes the features and lead to blurry outputs of low quality. In addition, unlike nearest neighbor correspondences computed + +using DDIM inversion, UV correspondences are not guaranteed to be an exact match in the low-dimensional feature space, which is only $64 \times 64$ pixels in our model, especially around object edges. Consequently, a straight blending of the post-attention features also creates blurry results. + +Our method can handle camera rotation with reasonable quality in static object-level scenarios, as shown in Fig. 4. In Fig. 5, we provide additional results on the same input meshes but with different prompts. Our method shows robustness across a diverse set of prompts and styles. Our inflated attention modules can capture motion-specific traits, such as shadows (Row 3, Row 7), lighting changes (Row 1), and reasonable deformations that are not modeled by the + +![](images/62a31c81aead55562e78dc8ee695237fda05241d8a94c0a991a141f4d7a9d304.jpg) +a shining diamond ball bouncing in a chamber + +![](images/5ae8460dad3f11cd0a0712af65f36559a9aedde509b48522e15902f472ea0425.jpg) +a fox running in a lively green forest + +![](images/a56ffbfad99db634f52080aac741a62ef4f99417af0791059c27e57a8e7e5ae2.jpg) +the Joker dancing + +![](images/83cc8ddd47be6de75b625f28ed4bdba481870dcc5a0b26f12d465c5298d17837.jpg) +a little girl walking around, Van Gogh style + +![](images/7a846c22002c0546bf3281d5e63c540a22fbf2e07e90d2071ec98dcc710032f7.jpg) +a little girl walking around, Da Vinci style + +![](images/f9f6183bcf9dd417ebb16915e1c8d32ffd097e3a6b2aca8fc16078f4da1d6825.jpg) +a stormtrooper swimming + +![](images/907f33d096ce5e1401d8cec552912f50bc0d6561d80ce2351e0c5bd1fdebc2e2.jpg) +three slendermen dancing +Figure 5. Qualitative results. Our method generalizes to a diverse set of input meshes, styles, and prompts. + +input low-fidelity meshes (Row 1 ground regions). + +Quantitative results. We show quantitative evaluations for frame consistency and prompt fidelity in Tab. 1. We observe that our frame consistency is better than all baselines using T2I models. As expected, Pix2Video [6] suffers more from temporal changes due to the lack of a correspondence injection module. Even though TensorFlow's token propagation encourages smoothness over fidelity, our method still shows better consistency. Only the T2V model Gen1 has a better + +frame consistency, which is expected because it is trained on video data. All current T2V models, including Gen1, can only generate a few frames and are limited in other ways. Our approach is an attempt to overcome this deficiency using 2D T2I models exclusively. Fidelity-wise, we are slightly better than Pix2Video [6], because it reuses the set of pre-attention features from previous frames, while we only reuse a portion of the pre-attention features to generate each frame. Since TokenFlow [10]'s token propagation mechanism in + +![](images/e6e68b1da0bc89dc23021d313a39e5e63e302acba770b95f8dce37886bd13aee.jpg) +Figure 6. Ablation study. We disabled several key building blocks and observe clear artifacts or inconsistencies in all ablations. + +![](images/f85826a741fc862bbae0d6869a5fd1a0de4ce57fe9fa26ef49b3ab5c5571bb01.jpg) + +
MethodFrame Consistency ↑Prompt Fidelity ↑
Per-Frame0.95470.3233
Pix2Video*0.96300.2983
TokenFlow*0.98220.2737
Gen10.99070.3029
Ours0.98450.3227
+ +Table 1. Quantitative evaluation. Our approach offers the best frame consistency among approaches using T2I models. As expected, only the proprietary T2V model Gen1 observes better temporal consistency. Our method also shows a higher prompt fidelity than relevant baselines that offer a reasonable amount of temporal consistency. *baseline adapted to our setting. + +produces over-smoothing effects, its results are blurry for many of the samples, hence achieving lower fidelity scores. Our generative rendering utilizes the benefits of pre-attention inflation to gain overall global structural and semantic consistency, while also combining with post-attention feature warping as local guidance. Combined with our UV noise initialization, we outperform or are on par with all relevant baselines in the proposed setting. + +Ablation studies. We perform a thorough ablation study to verify our design choices, shown in Fig. 6. We study three aspects of our pipeline: 1) comparing with pre-attention injection only, 2) comparing with post-attention injection only, and 3) the effectiveness of UV noise initialization. We observe that directly unifying and projecting the post-attention features without passing through the attention module produces chaotic results (severe artifacts in "no pre-attn", the joker's face and clothes regions). This observation is in line with our hypothesis for the poor performance of the adapted TokenFlow mechanism. When the motion is relatively large, geometrically matched features may not be good matches in the feature space. Therefore, directly warping a feature from one frame to another based on the geometrical correspondence has a high chance of not producing reasonable outputs. However, when we combine post-attention feature warping with pre-attention feature injection, the post-attention fea + +tures can act as guidance for the next network layer. Without post-attention feature warping, purely pre-attention feature injection does not guarantee pixel-wise consistency since the attention operation merely computes a correspondence based on feature similarities (in “no post-attn”, the joker’s clothes change over frames). Our UV noise initialization is well-aligned with our setup. If we use fixed noise instead, the blending of the features may cause unwanted artifacts (in “fixed noise”, the joker’s face has blending artifacts and is not recognizable in the third image). Our UV noise initialization serves generally and can work beyond our setup. Additional details on ablations are included in the supplementary. + +# 5. Discussion + +In this paper, we introduce Generative Rendering, a novel zero-shot pipeline based on 2D diffusion models for 4D-conditioned animation generation. Our method can animate creator-defined low-fidelity meshes and motion sequences, thereby bypassing steps requiring significant manual labor such as detailing, texturing, physical simulation, etc. The key idea of generative rendering is to utilize the prior within a depth-conditioned 2D diffusion model to provide the basic structure and physical fidelity to make convincing animations. This is accomplished by injecting the correspondences into the diffusion model using a combination of pre- and post-attention feature injection while unifying these features in the UV space. Our model demonstrates better frame consistency and prompt fidelity than relevant baselines. + +Conclusion. Diffusion models have emerged as state-of-the-art generative methods across multiple domains, such as 2D, 3D, and video generation and editing. generative rendering adds unprecedented levels of control to video generation with T2I diffusion models by bridging the gap between the traditional rendering pipeline and diffusion models. + +Acknowledgement. This project was in part supported by Samsung, Google and Stanford HAI. + +# References + +[1] Yuval Alaluf, Daniel Garibi, Or Patashnik, Hadar Averbuch-Elor, and Daniel Cohen-Or. Cross-image attention for zero-shot appearance transfer. In arXiv, 2023. 5 +[2] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In ECCV, 2022. 2 +[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 1, 2, 5 +[4] Shengqu Cai, Eric Ryan Chan, Songyou Peng, Mohamad Shahbazi, Anton Obukhov, Luc Van Gool, and Gordon Wetzstein. Diffdreamer: Towards consistent unsupervised single-view scene extrapolation with conditional diffusion models. In ICCV, 2023. 3 +[5] Tianshi Cao, Karsten Kreis, Sanja Fidler, Nicholas Sharp, and Kangxue Yin. Texfusion: Synthesizing 3d textures with text-guided image diffusion models. In ICCV, 2023. 3 +[6] Duygu Ceylan, Chun-Hao Huang, and Niloy J. Mitra. Pix2video: Video editing using image diffusion. In ICCV, 2023. 1, 2, 3, 5, 6, 7 +[7] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In ICCV, 2023. 2, 5 +[8] Rafail Fridman, Amit Abecasis, Yoni Kasten, and Tali Dekel. Scenescape: Text-driven consistent scene generation. In arXiv, 2023. 3 +[9] Songwei Ge, Seungjun Nah, Guilin Liu, Tyler Poon, Andrew Tao, Bryan Catanzaro, David Jacobs, Jia-Bin Huang, Ming-Yu Liu, and Yogesh Balaji. Preserve your own correlation: A noise prior for video diffusion models. In ICCV, 2023. 1, 2, 5 +[10] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. In arXiv, 2023. 1, 2, 4, 5, 6, 7 +[11] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. In arXiv, 2023. 1, 2 +[12] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2022. 1 +[13] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Image video: High definition video generation with diffusion models. In arXiv, 2022. 1, 2 +[14] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In arXiv, 2022. 3 +[15] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. In arXiv, 2022. 2 +[16] Lukas Hollein, Ang Cao, Andrew Owens, Justin Johnson, and Matthias Nießner. Text2room: Extracting textured 3d meshes from 2d text-to-image models. In ICCV, 2023. 3 + +[17] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly ddpm noise space: Inversion and manipulations. In arXiv, 2023. 2, 5, 6 +[18] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In NeurIPS, 2022. 5 +[19] Michael Kass and Davide Pesare. Coherent noise for nonphotorealistic rendering. In TOG, 2011. 5 +[20] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. In arXiv, 2023. 1, 2 +[21] Xin Li, Wenqing Chu, Ye Wu, Weihang Yuan, Fanglong Liu, Qi Zhang, Fu Li, Haocheng Feng, Errui Ding, and Jingdong Wang. Videogen: A reference-guided latent diffusion approach for high definition text-to-video generation, 2023. 1 +[22] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, 2023. 3 +[23] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 3 +[24] Hao Ouyang, Qiuyu Wang, Yuxi Xiao, Qingyan Bai, Juntao Zhang, Kecheng Zheng, Xiaowei Zhou, Qifeng Chen, and Yujun Shen. Codef: Content deformation fields for temporally consistent video processing. In arXiv, 2023. 2 +[25] Ryan Po and Gordon Wetzstein. Compositional 3d scene generation using locally conditioned diffusion. In arXiv, 2023. 3 +[26] Ryan Po, Wang Yifan, Vladislav Golyanik, Kfir Aberman, Jonathan T Barron, Amit H Bermano, Eric Ryan Chan, Tali Dekel, Aleksander Holynski, Angjoo Kanazawa, et al. State of the art on diffusion models for visual computing. arXiv preprint arXiv:2310.07204, 2023. 2 +[27] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 3 +[28] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. In ICCV, 2023. 2 +[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In CoRR, 2021. 5 +[30] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. In TPAMI, 2022. 3 +[31] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. In SIGGRAPH, 2023. 3 + +[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3, 5 +[33] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-video: Text-to-video generation without text-video data. In arXiv, 2022. 1, 2 +[34] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1921-1930, June 2023. 3 +[35] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 1 +[36] Tianfu Wang, Menelaos Kanakis, Konrad Schindler, Luc Van + +Gool, and Anton Obukhov. Breathing new life into 3d assets with generative repainting. In BMVC, 2023. 3 +[37] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In arXiv, 2022. 1, 2, 3, 5 +[38] Sihyun Yu, Kihyuk Sohn, Subin Kim, and Jinwoo Shin. Video probabilistic diffusion models in projected latent space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18456-18466, 2023. 1 +[39] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 1, 2, 3, 5 +[40] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 1 \ No newline at end of file diff --git a/2024/Generative Rendering_ Controllable 4D-Guided Video Generation with 2D Diffusion Models/images.zip b/2024/Generative Rendering_ Controllable 4D-Guided Video Generation with 2D Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..daf4ed610dc612c40c74fd5660b8276587bbee44 --- /dev/null +++ b/2024/Generative Rendering_ Controllable 4D-Guided Video Generation with 2D Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:22cbe3f8b48af550bf156df7e1b24a35557de5f85204dd479167325f035cd59c +size 996178 diff --git a/2024/Generative Rendering_ Controllable 4D-Guided Video Generation with 2D Diffusion Models/layout.json b/2024/Generative Rendering_ Controllable 4D-Guided Video Generation with 2D Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..357884b575945dd6d7502f06e3999a9ecd77e6fb --- /dev/null +++ b/2024/Generative Rendering_ Controllable 4D-Guided Video Generation with 2D Diffusion Models/layout.json @@ -0,0 +1,7217 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 102, + 526, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 102, + 526, + 138 + ], + "spans": [ + { + "bbox": [ + 67, + 102, + 526, + 138 + ], + "type": "text", + "content": "Generative Rendering: Controllable 4D-Guided Video Generation with 2D Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 135, + 160, + 470, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 160, + 470, + 204 + ], + "spans": [ + { + "bbox": [ + 135, + 160, + 470, + 204 + ], + "type": "text", + "content": "Shengqu Cai" + }, + { + "bbox": [ + 135, + 160, + 470, + 204 + ], + "type": "inline_equation", + "content": "^{1,2,\\triangle}" + }, + { + "bbox": [ + 135, + 160, + 470, + 204 + ], + "type": "text", + "content": " Duygu Ceylan" + }, + { + "bbox": [ + 135, + 160, + 470, + 204 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 135, + 160, + 470, + 204 + ], + "type": "text", + "content": " Matheus Gadelha" + }, + { + "bbox": [ + 135, + 160, + 470, + 204 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 135, + 160, + 470, + 204 + ], + "type": "text", + "content": " Chun-Hao Paul Huang" + }, + { + "bbox": [ + 135, + 160, + 470, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 135, + 160, + 470, + 204 + ], + "type": "text", + "content": " Tuanfeng Yang Wang" + }, + { + "bbox": [ + 135, + 160, + 470, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 135, + 160, + 470, + 204 + ], + "type": "text", + "content": " Gordon Wetzstein" + }, + { + "bbox": [ + 135, + 160, + 470, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 135, + 160, + 470, + 204 + ], + "type": "text", + "content": " Stanford University" + }, + { + "bbox": [ + 135, + 160, + 470, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 135, + 160, + 470, + 204 + ], + "type": "text", + "content": " Adobe Research" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 48, + 230, + 548, + 345 + ], + "blocks": [ + { + "bbox": [ + 48, + 230, + 548, + 345 + ], + "lines": [ + { + "bbox": [ + 48, + 230, + 548, + 345 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 548, + 345 + ], + "type": "image", + "image_path": "99b9d01934059d1cd0df4a1c6db8b3593326c3ff3b269d1cdb334d9ba6487fe6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 350, + 546, + 373 + ], + "lines": [ + { + "bbox": [ + 46, + 350, + 546, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 546, + 373 + ], + "type": "text", + "content": "Figure 1. Generative Rendering is a novel diffusion-based approach to render 3D animated but textured scenes (left) directly to a stylized animation (right), with styles specified by text prompts. Our approach offers new levels of user control to image generation models." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 390, + 192, + 402 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 390, + 192, + 402 + ], + "spans": [ + { + "bbox": [ + 143, + 390, + 192, + 402 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 418, + 290, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 418, + 290, + 659 + ], + "spans": [ + { + "bbox": [ + 45, + 418, + 290, + 659 + ], + "type": "text", + "content": "Traditional 3D content creation tools empower users to bring their imagination to life by giving them direct control over a scene's geometry, appearance, motion, and camera path. Creating computer-generated videos, however, is a tedious manual process, which can be automated by emerging text-to-video diffusion models. Despite great promise, video diffusion models are difficult to control, hindering a user to apply their own creativity rather than amplifying it. To address this challenge, we present a novel approach that combines the controllability of dynamic 3D meshes with the expressivity and editability of emerging diffusion models. For this purpose, our approach takes an animated, low-fidelity rendered mesh as input and injects the ground truth correspondence information obtained from the dynamic mesh into various stages of a pre-trained text-to-image generation model to output high-quality and temporally consistent frames. We demonstrate our approach on various examples where motion can be obtained by animating rigged assets or changing the camera path. Project page: primecai.github.io/generative_rendering." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 307, + 390, + 386, + 402 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 390, + 386, + 402 + ], + "spans": [ + { + "bbox": [ + 307, + 390, + 386, + 402 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 411, + 548, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 548, + 532 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 548, + 532 + ], + "type": "text", + "content": "Artists, designers, architects, and other creators rely on traditional 3D content creation tools to render computer-generated videos. Unfortunately, existing 3D workflows are laborious, time consuming, and require expertise. Emerging generative artificial intelligence tools, such as text-to-image (T2I) and text-to-video (T2V) models, solve these issues by automating many of the manual steps of traditional workflows. Video generation, however, is difficult to control in that it is not easily possible to specify scene layout and motion in a temporally consistent manner." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 532, + 548, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 532, + 548, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 532, + 548, + 687 + ], + "type": "text", + "content": "Recent approaches have attempted to control diffusion models. For example, ControlNet [39] uses a pre-trained T2I diffusion model and finetunes an adapter network that is conditioned on depth, pose, or edge images to control the layout. This strategy is successful for generating individual frames, but results in flicker for video generation. Other approaches aim at learning the complex types of motions encountered in natural videos directly [3, 9, 11-13, 21, 33, 35, 38, 40]. While successful in generating smooth motions, these approaches are not easily controllable. Finally, video-to-video diffusion models [6, 10, 20, 37] enable video editing and stylization, but they require a high-fidelity video as input, which is not always available." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "content": "In this paper, we aim to combine the power of 3D workflows with T2I models for generating 4D-guided stylized" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 693, + 286, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 693, + 286, + 703 + ], + "spans": [ + { + "bbox": [ + 61, + 693, + 286, + 703 + ], + "type": "inline_equation", + "content": "\\triangle" + }, + { + "bbox": [ + 61, + 693, + 286, + 703 + ], + "type": "text", + "content": " : Part of this work was done during an internship at Adobe Research." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 704, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 704, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 704, + 135, + 712 + ], + "type": "text", + "content": "*: Equal contribution." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7611" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 430 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 430 + ], + "type": "text", + "content": "animations. Specifically, we leverage 3D tools to rapidly prototype proxy geometry and motion of a scene (e.g., camera paths, physically based simulation, or character animation) while utilizing the T2I generation model as a renderer to output the final stylized animations (see Fig. 1). At the core of our method is the ability to leverage the ground truth 4D spatio-temporal correspondences that can be obtained from an animated 3D scene to guide the image generation. First, we leverage the correspondence information to effectively perform noise initialization which is critical for temporal consistency. Many recent video-to-video stylization methods resort to inverting the source video to obtain consistent noise initialization [6, 10]. However, this is not possible in our setup since we do not have an initial source video and inversion does not work well on untextured renderings of the 3D scene. Instead, we propose to utilize the canonical representation (i.e., UV space) of the 3D scene to initialize random noise which is then projected to each frame of the animation. Next, we enrich the self-attention layers of the image diffusion model which are known to contain spatial appearance information vital for obtaining consistent renderings [6, 37]. Specifically, we propagate both the input and output of self-attention layers across the frames via known correspondences to enforce more consistent results. Finally, similar to previous work [39], we use the depth cues rendered from the 3D scene in conjunction with control models to provide structure guidance. In summary, our framework treats the pre-trained T2I diffusion model as a multi-frame renderer that is able to capture the correspondences while maintaining high-fidelity and consistent generations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 432, + 288, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 432, + 288, + 467 + ], + "spans": [ + { + "bbox": [ + 47, + 432, + 288, + 467 + ], + "type": "text", + "content": "We evaluate our method on a variety of scenes and demonstrate its advantages with comparisons to state-of-the-art methods. In summary, our contributions include:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 476, + 287, + 603 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 47, + 476, + 287, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 476, + 287, + 512 + ], + "spans": [ + { + "bbox": [ + 47, + 476, + 287, + 512 + ], + "type": "text", + "content": "- We present a novel framework for 4D-guided animation synthesis utilizing the pre-trained T2I generation models as a multi-frame renderer." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 516, + 287, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 516, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 516, + 287, + 563 + ], + "type": "text", + "content": "- We enhance the self-attention layers of the image generation model by performing correspondence-aware blending of both input and output features to enforce consistent appearance synthesis." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 567, + 287, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 567, + 287, + 603 + ], + "spans": [ + { + "bbox": [ + 47, + 567, + 287, + 603 + ], + "type": "text", + "content": "- We introduce a UV-space noise initialization mechanism. Combined with the correspondence-aware attention, this enables better consistency across frame generations." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 623, + 133, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 133, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 133, + 635 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 643, + 288, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 643, + 288, + 680 + ], + "spans": [ + { + "bbox": [ + 47, + 643, + 288, + 680 + ], + "type": "text", + "content": "In this section, we discuss the methods most closely related to ours. A comprehensive review of diffusion models for visual computing can be found in [26]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "Video Generation. T2V diffusion models extend T2I models by generating dynamic scenes. The majority of works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 547, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 312 + ], + "type": "text", + "content": "in this area focus on extending pre-trained T2I diffusion models and finetuning on video data [3, 9, 13, 15]. A typical method of lifting image-based models to videos is to add a time-aware module. This can include 3D convolutions [3], or pseudo-3D convolutions factorized over spatial and temporal dimensions to reduce computational costs [7, 7, 9, 11, 33]. Recently, [11] has shown very compelling video generation results by training a motion module with such temporal layers on top a frozen StableDiffusion model [32]. All of these approaches, however, lack fine-grained structural control of the video generation process in a temporally stable manner. Moreover, training a high-quality T2V model is computationally demanding and challenging, in part due to the limited sizes of available captioned high-quality video datasets. This is part of the reason why existing T2V models are limited to outputting only a few seconds of animation. For these reasons, we build on T2I diffusion models and directly leverage ground truth correspondences of the input meshes to generate highly expressive and temporally consistent animations of arbitrary length." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 326, + 548, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 326, + 548, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 326, + 548, + 662 + ], + "type": "text", + "content": "Video Editing and Stylization. Another line of work focuses on video editing and stylization. Given an input video, video editing/stylization aims to change the input video according to a user-defined prompt, while keeping other components harmonious. Earlier work, such as [2], decomposes a video to atlases and performs editing directly on these atlases. This approach obtains perfect consistency, but it is only applicable to certain types of scenes and the decomposition is very slow to compute. Exemplar works based on diffusion models [6, 7, 20, 28, 37] along this path generally rely heavily on extracting temporal information from the input video, e.g., via DDIM inversion [17], then inflating the self-attention modules to attend to multiple frames, and propagating the edited contents across different frames. These techniques typically overfit to video inputs and are not directly applicable to our setting. For this reason, we adopt these baselines ensuring a fair comparison to their core methods, and demonstrate that our approach outperforms them. In addition, as the temporal prior is either extracted from the input video or learned from video training data, these works generally lack 3D awareness and may not generate 3D-aware animations. A recent attempt designed a 4D representation that employs a canonical field and deformation field to each frame, such that one can easily edit the video by changing contents on the canonical field which propagates naturally to the other frames [24]. However, this representation still has a relatively weak generalization ability and tends to only work well on very short video clips." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "In summary, while related, none of these video editing approaches operate in the same setting as our approach, which takes a low-fidelity, untextured rendering of a scene as input to generate a video." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7612" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 220, + 148 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 220, + 148 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 220, + 148 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 220, + 148 + ], + "type": "image", + "image_path": "88d918ba1f719f64344148808899064c4a0f56b5162ddc819196335eb933f60a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 51, + 149, + 220, + 257 + ], + "blocks": [ + { + "bbox": [ + 51, + 149, + 220, + 257 + ], + "lines": [ + { + "bbox": [ + 51, + 149, + 220, + 257 + ], + "spans": [ + { + "bbox": [ + 51, + 149, + 220, + 257 + ], + "type": "image", + "image_path": "642b780861f1cafcdc2001c30634eb32eb11b8368da59b4ff0cb131d9ad1aa78.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 262, + 547, + 328 + ], + "lines": [ + { + "bbox": [ + 46, + 262, + 547, + 328 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 547, + 328 + ], + "type": "text", + "content": "Figure 2. Overview of our pipeline. Our system takes as input a set of UV and depths maps rendered from an animated 3D scene. We use a depth-conditioned ControlNet [39] to generate corresponding frames while using the UV correspondences to preserve the consistency. We initialize the noise in the UV space of each object which we then render into each image (Sec. 3.3). For each diffusion step, we first use extended attention for a set of keyframes and extract their pre- and post-attention features (Sec. 3.1). The post-attention features are projected to the UV space and unified. Finally, all frames are generated using a weighted combination of the outputs of the extended attention with the pre-attention features of the keyframe, and the UV-composed post-attention features from the keyframes (Sec. 3.2)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 222, + 70, + 541, + 257 + ], + "blocks": [ + { + "bbox": [ + 222, + 70, + 541, + 257 + ], + "lines": [ + { + "bbox": [ + 222, + 70, + 541, + 257 + ], + "spans": [ + { + "bbox": [ + 222, + 70, + 541, + 257 + ], + "type": "image", + "image_path": "a658fd35eea5eecc17ccf46024ba7a49054856f51be34097d7d5d64cb3cc50cf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 349, + 289, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 349, + 289, + 612 + ], + "spans": [ + { + "bbox": [ + 46, + 349, + 289, + 612 + ], + "type": "text", + "content": "3D Generation. This class of methods is able to achieve near-perfect 3D consistency by using underlying 3D representations. One line of work [4,8,16] focuses on scene generation using diffusion models. These works typically predict or estimate the scene geometry on the fly using monocular predictors such as MiDaS [30] and use that information to constitute individual generations into a scene. This approach is applicable only in specific domains such as indoor scenes or nature scenes, with limited motion control ability due to depth-scale ambiguity. The closest form of 3D generation to our task is mesh texturing. Given an explicit representation such as a mesh, prior works [5,31,36] demonstrate convincing asset mesh texturization using the powerful inpainting function of Stable Diffusion [32]. Other works [25,27] use Score Distillation [27] to directly optimize a neural radiance field [23], but they suffer from quality degradation potentially due to heavy classifier-free guidance [14]. Follow-up works [22] attempt to combine the strength of score distillation and mesh texturization. Since these methods generate a static texture for a given mesh, they are not able to capture appearance interactions (e.g., reflections or shadows) in case of scene-level content and motions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 621, + 103, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 103, + 633 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 103, + 633 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "content": "The goal of generative rendering is to transfer creator-defined scene-level proxy meshes (without appearance) and motions directly to a convincing animation. Given a 3D scene composed of animated meshes, we render guiding channels that consist of depth maps " + }, + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = [D^1,\\dots,D^N]" + }, + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "content": " and UV coordinate maps " + }, + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{U}\\mathcal{V} = [UV^{1},\\dots,UV^{N}]" + }, + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 349, + 548, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 349, + 548, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 349, + 548, + 517 + ], + "type": "text", + "content": "the length of the sequence and each UV map contains the 2D texture coordinates as well as an object ID in its RGB channels. Using a depth-conditioned 2D image generation model such as StableDiffusion [32] with depth ControlNet [39], we aim to transfer the sequence " + }, + { + "bbox": [ + 304, + 349, + 548, + 517 + ], + "type": "inline_equation", + "content": "(D^i,UV^i)" + }, + { + "bbox": [ + 304, + 349, + 548, + 517 + ], + "type": "text", + "content": " into a set of generated images that depict a stylized animation. To achieve this goal, we perform feature blending over the input and output features of the self-attention modules of the diffusion model (Sec. 3.2) and also perform UV-guided noise initialization (Sec. 3.3) to improve the temporal consistency. We provide an overview of our framework in Fig. 2 and our overall algorithm in Alg. 1. Next, we provide some background on the self-attention features of a diffusion model before discussing the different parts of our method in detail." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 534, + 471, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 534, + 471, + 547 + ], + "spans": [ + { + "bbox": [ + 306, + 534, + 471, + 547 + ], + "type": "text", + "content": "3.1. Self-attention feature injection" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 556, + 547, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 547, + 689 + ], + "type": "text", + "content": "Recent work has demonstrated that self-attention features encode spatial self-similarity [34] and they are effective in terms of maintaining temporal consistency when they are used to perform cross-frame attention [6, 37]. A typical approach is to extend the diffusion network to process multiple frames jointly where they can attend to each other's features. Furthermore, it is possible to manipulate both the input and output features of a self-attention module to further enforce consistency. We denote these operations as pre-attention and post-attention feature injections respectively and provide a description in the following." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "content": "Pre-attention Feature Injection. An attention module is" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7613" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 93, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 93, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 93, + 83 + ], + "type": "text", + "content": "defined as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 90, + 288, + 118 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 90, + 288, + 118 + ], + "spans": [ + { + "bbox": [ + 56, + 90, + 288, + 118 + ], + "type": "interline_equation", + "content": "\\boldsymbol {F} = \\operatorname {A t t n} (\\boldsymbol {Q}; \\boldsymbol {K}; \\boldsymbol {V}) = \\operatorname {S o f t m a x} \\left(\\frac {\\boldsymbol {Q} \\boldsymbol {K} ^ {\\top}}{\\sqrt {d}}\\right) \\cdot \\boldsymbol {V}, \\tag {1}", + "image_path": "ce9e3168d77d6c0c70ede21e399ef850f0866b42c35a8c6ea9f21444aab96d5c.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 125, + 287, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 125, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 47, + 125, + 287, + 173 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 125, + 287, + 173 + ], + "type": "inline_equation", + "content": "Q, K" + }, + { + "bbox": [ + 47, + 125, + 287, + 173 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 125, + 287, + 173 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 47, + 125, + 287, + 173 + ], + "type": "text", + "content": " denote the query, key, and value features respectively. In case of a self-attention module, " + }, + { + "bbox": [ + 47, + 125, + 287, + 173 + ], + "type": "inline_equation", + "content": "Q, K" + }, + { + "bbox": [ + 47, + 125, + 287, + 173 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 125, + 287, + 173 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 47, + 125, + 287, + 173 + ], + "type": "text", + "content": " are computed by projecting the same features " + }, + { + "bbox": [ + 47, + 125, + 287, + 173 + ], + "type": "inline_equation", + "content": "f^{(i,l)}" + }, + { + "bbox": [ + 47, + 125, + 287, + 173 + ], + "type": "text", + "content": ", via the corresponding attention matrices, i.e.:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 179, + 288, + 205 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 179, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 47, + 179, + 288, + 205 + ], + "type": "interline_equation", + "content": "\\boldsymbol {Q} ^ {(i, l)} = \\mathbf {W} ^ {Q} \\boldsymbol {f} ^ {(i, l)}, \\boldsymbol {K} ^ {(i, l)} = \\mathbf {W} ^ {K} \\boldsymbol {f} ^ {(i, l)}, \\boldsymbol {V} ^ {(i, l)} = \\mathbf {W} ^ {V} \\boldsymbol {f} ^ {(i, l)} \\tag {2}", + "image_path": "50c19850fd8b68e0ffbda8fb32c679b495cd891afdf0c71b8c44d8595d79e1db.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 206, + 287, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 206, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 206, + 287, + 277 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 206, + 287, + 277 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 206, + 287, + 277 + ], + "type": "text", + "content": " denotes the frame index and " + }, + { + "bbox": [ + 46, + 206, + 287, + 277 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 206, + 287, + 277 + ], + "type": "text", + "content": " denotes the diffusion step and " + }, + { + "bbox": [ + 46, + 206, + 287, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^{Q / K / V}" + }, + { + "bbox": [ + 46, + 206, + 287, + 277 + ], + "type": "text", + "content": " are weight matrices for query, key, and values, respectively. In other words, for each query feature an attention weight is computed based on the similarity between the query and each key feature. Then, the value features are combined based on this attention weight." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 278, + 287, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 278, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 278, + 287, + 361 + ], + "type": "text", + "content": "One can manipulate the input features provided to the self-attention module, specifically the key and value features, which we denote as pre-attention feature injection. As discussed by [10], in case of generating " + }, + { + "bbox": [ + 46, + 278, + 287, + 361 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 278, + 287, + 361 + ], + "type": "text", + "content": " frames, one naive option is to perform extended attention where the features " + }, + { + "bbox": [ + 46, + 278, + 287, + 361 + ], + "type": "inline_equation", + "content": "\\pmb{f}^{(i,l)}" + }, + { + "bbox": [ + 46, + 278, + 287, + 361 + ], + "type": "text", + "content": " of all the frames are concatenated and utilized as key and value pairs:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 100, + 369, + 287, + 402 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 369, + 287, + 402 + ], + "spans": [ + { + "bbox": [ + 100, + 369, + 287, + 402 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\boldsymbol {K} ^ {(i, l)} = \\mathbf {W} ^ {K} \\left[ \\boldsymbol {f} ^ {(1, l)}, \\dots , \\boldsymbol {f} ^ {(N, l)} \\right], \\tag {3} \\\\ \\boldsymbol {V} ^ {(i, l)} = \\mathbf {W} ^ {V} [ \\boldsymbol {f} ^ {(1, l)}, \\dots , \\boldsymbol {f} ^ {(N, l)} ]. \\\\ \\end{array}", + "image_path": "b3f0f8bcd5f28c34ef42c851ea153ad77c29a5cfdd57d23945e975081d5842c1.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 409, + 287, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 409, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 409, + 287, + 517 + ], + "type": "text", + "content": "Extended attention is a synchronized operation, where all frames attend to features of each other resulting in a more consistent generation. However, generating multiple frames together incurs a high computational cost, which in practice limits the number of frames that can be processed in a sequence. For this reason, we utilize extended attention only for a subset of keyframes and propose mechanisms to propagate the keyframe features to the rest of the sequence, as discussed in the next section." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 527, + 288, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 527, + 288, + 611 + ], + "spans": [ + { + "bbox": [ + 46, + 527, + 288, + 611 + ], + "type": "text", + "content": "Post-attention Feature Injection. While it is possible to manipulate the input of the self-attention layer, another option is to directly manipulate the output, which we denote as post-attention feature injection. In particular, given known correspondences between two frames " + }, + { + "bbox": [ + 46, + 527, + 288, + 611 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 527, + 288, + 611 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 527, + 288, + 611 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 527, + 288, + 611 + ], + "type": "text", + "content": ", one can reproject " + }, + { + "bbox": [ + 46, + 527, + 288, + 611 + ], + "type": "inline_equation", + "content": "F^{(i,l)}" + }, + { + "bbox": [ + 46, + 527, + 288, + 611 + ], + "type": "text", + "content": ", the output of the self-attention module of frame " + }, + { + "bbox": [ + 46, + 527, + 288, + 611 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 527, + 288, + 611 + ], + "type": "text", + "content": " at diffusion step " + }, + { + "bbox": [ + 46, + 527, + 288, + 611 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 527, + 288, + 611 + ], + "type": "text", + "content": ", to obtain " + }, + { + "bbox": [ + 46, + 527, + 288, + 611 + ], + "type": "inline_equation", + "content": "F^{(j,l)}" + }, + { + "bbox": [ + 46, + 527, + 288, + 611 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 123, + 618, + 287, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 618, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 123, + 618, + 287, + 634 + ], + "type": "interline_equation", + "content": "\\boldsymbol {F} ^ {(j, l)} = \\pi_ {i, j} (\\boldsymbol {F} ^ {(i, l)}), \\tag {4}", + "image_path": "d2c1b955a4abda901b5ba9e3b354358970b70c62708cfa20b3000a2620facbac.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 642, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 642, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\pi_{i,j}" + }, + { + "bbox": [ + 46, + 642, + 287, + 677 + ], + "type": "text", + "content": " denotes the re-projection operation, which is enabled by the ground truth pixel-level correspondences between frames given by our UV maps." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": "While post-attention feature injection is effective in achieving consistency, it tends to create blurry results and also avoids generating of certain interactions between objects" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 547, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 121 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 121 + ], + "type": "text", + "content": "(e.g., shadows, reflections) since it overwrites the output the self attention. Hence, as described next, our method combines pre- and post-attention feature injection in an effective way to generate consistent but also high-quality frames." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 126, + 453, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 126, + 453, + 139 + ], + "spans": [ + { + "bbox": [ + 305, + 126, + 453, + 139 + ], + "type": "text", + "content": "3.2. UV-space feature injection" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 144, + 547, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 144, + 547, + 418 + ], + "spans": [ + { + "bbox": [ + 304, + 144, + 547, + 418 + ], + "type": "text", + "content": "Given that pre- and post-attention features of the self-attention module can be manipulated to enhance consistency, our main goal is to perform this manipulation guided by the ground truth correspondences we have from the 3D scene. Specifically, we utilize the canonical UV space to bring the features from the different frames into correspondence which we denote as UV-space feature blending. Given a set of feature maps for each frame " + }, + { + "bbox": [ + 304, + 144, + 547, + 418 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 144, + 547, + 418 + ], + "type": "text", + "content": ", we project them to a canonical UV space. In particular, for each texel in the UV space, we find its closest point correspondence in each frame. Having such correspondences from multiple frames, a straightforward approach is to take the (weighted) average to blend the features [10]. However, we find this consistently leads to sub-optimal blurry results. Therefore, to combat over-smoothing of the features, we additionally perform blending sequentially and fill a certain texel with the features of its corresponding pixel in a frame only if it has not been filled before. The final unified texture is then the mean of the inpainted texture and the average texture. After features from all frames are blended, we obtain a UV-space feature map which we denote as " + }, + { + "bbox": [ + 304, + 144, + 547, + 418 + ], + "type": "inline_equation", + "content": "T^{(i,l)}" + }, + { + "bbox": [ + 304, + 144, + 547, + 418 + ], + "type": "text", + "content": ". Once such a unified feature map is obtained, we can project it back to each frame to obtain " + }, + { + "bbox": [ + 304, + 144, + 547, + 418 + ], + "type": "inline_equation", + "content": "\\bar{f}^{(i,l)}" + }, + { + "bbox": [ + 304, + 144, + 547, + 418 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 420, + 547, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 420, + 547, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 420, + 547, + 594 + ], + "type": "text", + "content": "Given the UV-space feature blending mechanism, we next describe how we utilize it during the image generation process. For each diffusion step, we first sample a random batch of " + }, + { + "bbox": [ + 304, + 420, + 547, + 594 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 420, + 547, + 594 + ], + "type": "text", + "content": " keyframes and perform extended attention on these keyframes, extracting both their pre-attention features " + }, + { + "bbox": [ + 304, + 420, + 547, + 594 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{KF}^{(i,l)}" + }, + { + "bbox": [ + 304, + 420, + 547, + 594 + ], + "type": "text", + "content": " and post-attention features " + }, + { + "bbox": [ + 304, + 420, + 547, + 594 + ], + "type": "inline_equation", + "content": "\\pmb{F}_{KF}^{(i,l)}" + }, + { + "bbox": [ + 304, + 420, + 547, + 594 + ], + "type": "text", + "content": ". We further composite a UV space feature map, " + }, + { + "bbox": [ + 304, + 420, + 547, + 594 + ], + "type": "inline_equation", + "content": "\\bar{T}_{KF}^{l}" + }, + { + "bbox": [ + 304, + 420, + 547, + 594 + ], + "type": "text", + "content": ", by blending " + }, + { + "bbox": [ + 304, + 420, + 547, + 594 + ], + "type": "inline_equation", + "content": "\\pmb{F}_{KF}^{(i,l)}" + }, + { + "bbox": [ + 304, + 420, + 547, + 594 + ], + "type": "text", + "content": " as described previously. We then perform the diffusion step on all the frames (in case of keyframes, the diffusion step is simply repeated) in a sequential manner. For each frame, we compose the pre-attention features by concatenating the pre-attention features of the keyframes, " + }, + { + "bbox": [ + 304, + 420, + 547, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{KF}^{l} = [f_{KF}^{(1,l)},\\dots,f_{KF}^{(m,l)}]" + }, + { + "bbox": [ + 304, + 420, + 547, + 594 + ], + "type": "text", + "content": ", with the pre-attention features of the current frame " + }, + { + "bbox": [ + 304, + 420, + 547, + 594 + ], + "type": "inline_equation", + "content": "\\pmb{f}^{(i,l)}" + }, + { + "bbox": [ + 304, + 420, + 547, + 594 + ], + "type": "text", + "content": " resulting in:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 370, + 597, + 545, + 619 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 597, + 545, + 619 + ], + "spans": [ + { + "bbox": [ + 370, + 597, + 545, + 619 + ], + "type": "interline_equation", + "content": "\\boldsymbol {K} ^ {(i, l)} = \\mathbf {W} ^ {K} \\left[ \\boldsymbol {f} ^ {(i, l)}, \\mathbf {f} _ {K F} ^ {l} \\right], \\tag {5}", + "image_path": "20965c2d972033e0301f8e68e575e8d8ad533c9e4a1aa2f2ddfed82548bdf276.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 372, + 616, + 481, + 630 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 616, + 481, + 630 + ], + "spans": [ + { + "bbox": [ + 372, + 616, + 481, + 630 + ], + "type": "interline_equation", + "content": "\\mathbf {V} ^ {(i, l)} = \\mathbf {W} ^ {V} [ \\mathbf {f} ^ {(i, l)}, \\mathbf {f} _ {K F} ^ {l} ].", + "image_path": "a2a3e40d8437234e9126fcc5837da30ec9db678f0309ea16dfff561132cae12c.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 635, + 547, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 635, + 547, + 693 + ], + "spans": [ + { + "bbox": [ + 304, + 635, + 547, + 693 + ], + "type": "text", + "content": "After executing the self-attention module, we blend the output, i.e., post-attention features of the current frame " + }, + { + "bbox": [ + 304, + 635, + 547, + 693 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{F}}^{(i,l)}" + }, + { + "bbox": [ + 304, + 635, + 547, + 693 + ], + "type": "text", + "content": ", with the projection of the blended post-attention features to the current frame " + }, + { + "bbox": [ + 304, + 635, + 547, + 693 + ], + "type": "inline_equation", + "content": "\\bar{\\pmb{F}}^{(i,l)}" + }, + { + "bbox": [ + 304, + 635, + 547, + 693 + ], + "type": "text", + "content": ". Hence, the updated output of the module becomes:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 351, + 699, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 699, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 351, + 699, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\boldsymbol {F} _ {\\text {o u t}} ^ {(i, l)} = \\alpha \\cdot \\bar {\\boldsymbol {F}} ^ {(i, l)} + (1 - \\alpha) \\cdot \\hat {\\boldsymbol {F}} ^ {(i, l)} \\tag {6}", + "image_path": "d4cd9f684bf3bf9086d31ab2915d1868667d9ebaf6f08934e7bd83c84787bd2a.jpg" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7614" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 287, + 96 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 71, + 287, + 96 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{F}}^{(i,l)}" + }, + { + "bbox": [ + 47, + 71, + 287, + 96 + ], + "type": "text", + "content": " is the output of the inflated self-attention block computed using keys and values in Eq. 5." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 105, + 174, + 116 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 105, + 174, + 116 + ], + "spans": [ + { + "bbox": [ + 47, + 105, + 174, + 116 + ], + "type": "text", + "content": "3.3. UV noise initialization" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 123, + 289, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 123, + 289, + 352 + ], + "spans": [ + { + "bbox": [ + 46, + 123, + 289, + 352 + ], + "type": "text", + "content": "Prior video editing works [6, 10] generally rely on inversion methods, such as DDIM inversion [17], to invert each source frame to the noise space which are then used as initialization for editing. Since source frames in the input video are coherent, this inversion strategy yields relatively coherent noise patterns that in turn helps with consistent editing. However, in our case, since we have an untextured scene, inverting rendered frames that lack appearance does not lead to coherent noisy latents. Another option is to keep the noise fixed across all frames. However, this results in severe texture-sticking issues as we demonstrate in the supplementary material. We seek an alternative to capture temporal correlations in initial noise patterns. Inspired by previous work [9, 19], we leverage the fact that we have a canonical UV space by initializing the noise in this space by creating a Gaussian noise texture. We then project this noise to each frame by utilizing the frame-UV correspondences. As we illustrate in our evaluations, this significantly boosts the coherency of the generated outputs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 360, + 169, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 360, + 169, + 371 + ], + "spans": [ + { + "bbox": [ + 47, + 360, + 169, + 371 + ], + "type": "text", + "content": "3.4. Latent normalization" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 379, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 379, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 379, + 287, + 498 + ], + "type": "text", + "content": "We notice a color flickering issue, where the overall color distribution and contrast change between frames, even if the motion is limited. Therefore, we apply an AdaIN operation on the predicted latents to make every frame's overall distribution similar to the first frame. The simple AdaIN operation solves this issue for some samples, but we find that calculating the distribution over all pixels is very sensitive to the motions. Therefore, inspired by [1], we calculate the statistics only on the background pixels, which are easy to separate in our setup." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 510, + 128, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 510, + 128, + 523 + ], + "spans": [ + { + "bbox": [ + 47, + 510, + 128, + 523 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 540, + 287, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 287, + 660 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 287, + 660 + ], + "type": "text", + "content": "Datasets. We evaluate our method on scenes with various complexity demonstrating three main types of animation sequences: 1) Camera rotations: given a single static object we create a camera path with 360-rotation around the object; 2) Physical simulation: we prepare a motion sequence by utilizing physically based simulation, e.g., bouncing balls where we use rigid body dynamics for collision handling; 3) Character animation: we create scenes with rigged and animated characters obtained from Mixamo and render them with various dance motions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 670, + 287, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 287, + 694 + ], + "type": "text", + "content": "Baselines. To our knowledge, no published work directly focuses on the same task our method does. Therefore, we" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 72, + 450, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 450, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 450, + 84 + ], + "type": "text", + "content": "Algorithm 1 Generative Rendering" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 89, + 545, + 126 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 89, + 545, + 126 + ], + "spans": [ + { + "bbox": [ + 306, + 89, + 545, + 126 + ], + "type": "text", + "content": "Input: UV maps " + }, + { + "bbox": [ + 306, + 89, + 545, + 126 + ], + "type": "inline_equation", + "content": "\\mathcal{U}\\mathcal{V} = [UV^1,\\dots,UV^N]" + }, + { + "bbox": [ + 306, + 89, + 545, + 126 + ], + "type": "text", + "content": "; depth maps " + }, + { + "bbox": [ + 306, + 89, + 545, + 126 + ], + "type": "inline_equation", + "content": "\\pmb{\\mathcal{D}} = [D^{1},\\dots,D^{N}]" + }, + { + "bbox": [ + 306, + 89, + 545, + 126 + ], + "type": "text", + "content": "; text prompt " + }, + { + "bbox": [ + 306, + 89, + 545, + 126 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 306, + 89, + 545, + 126 + ], + "type": "text", + "content": "; diffusion model " + }, + { + "bbox": [ + 306, + 89, + 545, + 126 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 306, + 89, + 545, + 126 + ], + "type": "text", + "content": "; projection from UV to image " + }, + { + "bbox": [ + 306, + 89, + 545, + 126 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathcal{U}\\mathcal{V}}" + }, + { + "bbox": [ + 306, + 89, + 545, + 126 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 306, + 89, + 545, + 126 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 306, + 89, + 545, + 126 + ], + "type": "text", + "content": " diffusion steps; " + }, + { + "bbox": [ + 306, + 89, + 545, + 126 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 306, + 89, + 545, + 126 + ], + "type": "text", + "content": " keyframes" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 132, + 545, + 144 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 132, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 306, + 132, + 545, + 144 + ], + "type": "interline_equation", + "content": "z _ {\\mathbf {U V}} \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I}) \\quad / / \\text {S a m p l e n o i s e i n U V s p a c e}", + "image_path": "3090728af32e1fdde28cf7acb839f87240aa0d387772b90cbc1386050f4c56f3.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 144, + 397, + 157 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 144, + 397, + 157 + ], + "spans": [ + { + "bbox": [ + 306, + 144, + 397, + 157 + ], + "type": "interline_equation", + "content": "\\mathbf {z} = \\mathbf {z} _ {T} = \\pi_ {\\mathbf {u v}} (z _ {\\mathbf {u v}})", + "image_path": "4992fd0729c9bc0b864049c85ab26a3bae507014a18499eb67e0089001450b56.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 157, + 391, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 157, + 391, + 167 + ], + "spans": [ + { + "bbox": [ + 306, + 157, + 391, + 167 + ], + "type": "text", + "content": "For " + }, + { + "bbox": [ + 306, + 157, + 391, + 167 + ], + "type": "inline_equation", + "content": "t = T,\\dots ,1" + }, + { + "bbox": [ + 306, + 157, + 391, + 167 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 170, + 545, + 218 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 170, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 317, + 170, + 545, + 218 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} K F = \\text {s h u f f l e} (\\{i _ {1}, \\dots , i _ {k} \\}) \\\\ \\boldsymbol {f} _ {K F}, \\boldsymbol {F} _ {K F} \\leftarrow \\Phi (\\mathbf {z} _ {K F}, \\boldsymbol {\\mathcal {D}} _ {K F}, c) \\quad / / \\text {S a m p l e k e y f r a m e s} \\\\ \\bar {\\boldsymbol {T}} _ {K F} = \\text {b l e n d} (\\pi_ {\\boldsymbol {u v}} ^ {- 1} (\\boldsymbol {F} _ {K F})) \\quad / / \\text {B l e n d s f e a t u r e s i n U V s p a c e} \\\\ \\mathbf {z} _ {t - 1} = [ ] \\end{array}", + "image_path": "13b9c4f402100099c99434a51d9872b665973ea75264e629fd5243627bde415b.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 219, + 376, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 219, + 376, + 228 + ], + "spans": [ + { + "bbox": [ + 317, + 219, + 376, + 228 + ], + "type": "text", + "content": "For " + }, + { + "bbox": [ + 317, + 219, + 376, + 228 + ], + "type": "inline_equation", + "content": "z^i" + }, + { + "bbox": [ + 317, + 219, + 376, + 228 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 317, + 219, + 376, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 317, + 219, + 376, + 228 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 325, + 229, + 401, + 242 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 229, + 401, + 242 + ], + "spans": [ + { + "bbox": [ + 325, + 229, + 401, + 242 + ], + "type": "interline_equation", + "content": "\\bar {\\boldsymbol {F}} ^ {i} = \\pi_ {U V ^ {i}} (\\bar {\\boldsymbol {T}} _ {K F})", + "image_path": "d75021b03bd7704a13df2a75f37047ab6c584cd5570fb966fc0a7efef8188abb.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 325, + 242, + 457, + 255 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 242, + 457, + 255 + ], + "spans": [ + { + "bbox": [ + 325, + 242, + 457, + 255 + ], + "type": "interline_equation", + "content": "z _ {t - 1} ^ {i} = \\Phi (z ^ {i}, \\boldsymbol {\\mathcal {D}} _ {K F}, c, \\boldsymbol {f} _ {K F}, \\bar {\\boldsymbol {F}} ^ {i})", + "image_path": "7e243f86dd2c90de8e210b5ca345fecee9eb036afa0563a0be9685650247dbf0.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 325, + 255, + 402, + 266 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 255, + 402, + 266 + ], + "spans": [ + { + "bbox": [ + 325, + 255, + 402, + 266 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {t - 1}. \\text {a p p e n d} \\left(z _ {t - 1} ^ {i}\\right)", + "image_path": "ac0c8a4bc3357b6a249b503c11fd71ab746d0ff43afaea77746e208b13227b65.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 325, + 267, + 362, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 267, + 362, + 277 + ], + "spans": [ + { + "bbox": [ + 325, + 267, + 362, + 277 + ], + "type": "interline_equation", + "content": "\\mathbf {z} = \\mathbf {z} _ {t - 1}", + "image_path": "cea15c7dcee22a594863f78af2e3da87aa13d36be5cf6f2566de1ed602af59ec.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 280, + 351, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 280, + 351, + 292 + ], + "spans": [ + { + "bbox": [ + 306, + 280, + 351, + 292 + ], + "type": "text", + "content": "Output: z" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 316, + 547, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 547, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 547, + 472 + ], + "type": "text", + "content": "adapt two recent methods to our setting as baselines. 1) we use the pre-attention feature propagation mechanism of Pix2Video [6] (i.e., each frame attends to an anchor and previous frame) together with depth conditioning, and 2) we apply the extended attention mechanism and token propagation from TokenFlow [10] using nearest neighbor matching. We note that the original TokenFlow method computes correspondences by matching features obtained via inverting original frames. Since this is not meaningful in our case, we instead provide the ground truth correspondences we obtain from 3D. Both Pix2Video and TokenFlow rely on inversion for noise initialization. Since this is not doable in our setting, we instead use identical noise to initialize." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 483, + 547, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 547, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 547, + 579 + ], + "type": "text", + "content": "Evaluation metrics. Following prior video editing evaluation protocols [3,6,7,10,37], we report two metrics based on CLIP [29]. To measure frame consistency, we compute CLIP image embeddings for each frame of the output and compute the average cosine similarity between all pairs of consecutive video frames. To measure prompt fidelity, we compute the mean CLIP embedding similarity score between each frame of the output and the corresponding input prompt." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 590, + 547, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 590, + 547, + 638 + ], + "spans": [ + { + "bbox": [ + 304, + 590, + 547, + 638 + ], + "type": "text", + "content": "Implementation details. We use the StableDiffusion [32] v1.5 model and a depth-conditioned ControlNet [39] as our image generator. We generate all results at " + }, + { + "bbox": [ + 304, + 590, + 547, + 638 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 590, + 547, + 638 + ], + "type": "text", + "content": " resolution using Karras scheduler [18] with 50 denoising steps." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 648, + 362, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 648, + 362, + 659 + ], + "spans": [ + { + "bbox": [ + 306, + 648, + 362, + 659 + ], + "type": "text", + "content": "4.1. Results" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 677, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 714 + ], + "type": "text", + "content": "Qualitative results. Qualitative comparisons to the baselines are shown in Fig. 3. While Pix2Video [6] produces high-quality frames it shows limited spatio-temporal con" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 61, + 703, + 150, + 712 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 703, + 150, + 712 + ], + "spans": [ + { + "bbox": [ + 61, + 703, + 150, + 712 + ], + "type": "text", + "content": "https://www MIXamo.com/" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7615" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 296, + 364 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 296, + 364 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 296, + 364 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 296, + 364 + ], + "type": "image", + "image_path": "9eab1f9c742bbf615015b9eea21104269038310540baaaeb5424c35e3922a66a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 96, + 365, + 259, + 372 + ], + "lines": [ + { + "bbox": [ + 96, + 365, + 259, + 372 + ], + "spans": [ + { + "bbox": [ + 96, + 365, + 259, + 372 + ], + "type": "text", + "content": "a photo realistic basketball bouncing under dreamlike sunlight" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 298, + 70, + 543, + 364 + ], + "blocks": [ + { + "bbox": [ + 298, + 70, + 543, + 364 + ], + "lines": [ + { + "bbox": [ + 298, + 70, + 543, + 364 + ], + "spans": [ + { + "bbox": [ + 298, + 70, + 543, + 364 + ], + "type": "image", + "image_path": "c7b145a57890474794a0209f7c80e5fe9b57f1631d6b5d9848a0bcdb1bcb903e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 350, + 365, + 499, + 372 + ], + "lines": [ + { + "bbox": [ + 350, + 365, + 499, + 372 + ], + "spans": [ + { + "bbox": [ + 350, + 365, + 499, + 372 + ], + "type": "text", + "content": "a hyper realistic fluffy Swarovski blue crystal fox running" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 51, + 420, + 543, + 511 + ], + "blocks": [ + { + "bbox": [ + 46, + 376, + 546, + 409 + ], + "lines": [ + { + "bbox": [ + 46, + 376, + 546, + 409 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 546, + 409 + ], + "type": "text", + "content": "Figure 3. Qualitative comparisons. We adapt the feature blending mechanisms proposed by Pix2Video [6] and TokenFlow [10] to our setting. While Pix2Video produces reasonable results, it lacks a mechanism to maintain frame-level consistency. TokenFlow uses a token propagation & interpolation mechanism between nearby keyframes, but this results in blurred outputs in our setting." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 51, + 420, + 543, + 511 + ], + "lines": [ + { + "bbox": [ + 51, + 420, + 543, + 511 + ], + "spans": [ + { + "bbox": [ + 51, + 420, + 543, + 511 + ], + "type": "image", + "image_path": "34da284c4522893b8ea3d00630986bb7d5a1e1b74c0e43ce970f78d13c12ae56.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 93, + 515, + 499, + 526 + ], + "lines": [ + { + "bbox": [ + 93, + 515, + 499, + 526 + ], + "spans": [ + { + "bbox": [ + 93, + 515, + 499, + 526 + ], + "type": "text", + "content": "Figure 4. Camera and object rotation. Our algorithm supports camera and object rotations for static scenarios." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 548, + 289, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 548, + 289, + 704 + ], + "spans": [ + { + "bbox": [ + 46, + 548, + 289, + 704 + ], + "type": "text", + "content": "sistency, because it purely relies on the cross-frame attention mechanism to attend to corresponding features. Token-Flow [10] can produce smoother results, but we find that it consistently produces blurry textures. This is in part due to the lack of DDIM inversion [17], which provides a decent level of feature similarities across frames. In our setting, this DDIM inversion is not meaningful and skipping the self-attention module to directly blend the features of keyframes means averaging less similar features, resulting in blurry appearance. Without accurate DDIM inversion and correct nearest neighbor matching, the blending oversmoothes the features and lead to blurry outputs of low quality. In addition, unlike nearest neighbor correspondences computed" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 548, + 547, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 548, + 547, + 608 + ], + "spans": [ + { + "bbox": [ + 305, + 548, + 547, + 608 + ], + "type": "text", + "content": "using DDIM inversion, UV correspondences are not guaranteed to be an exact match in the low-dimensional feature space, which is only " + }, + { + "bbox": [ + 305, + 548, + 547, + 608 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 305, + 548, + 547, + 608 + ], + "type": "text", + "content": " pixels in our model, especially around object edges. Consequently, a straight blending of the post-attention features also creates blurry results." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "Our method can handle camera rotation with reasonable quality in static object-level scenarios, as shown in Fig. 4. In Fig. 5, we provide additional results on the same input meshes but with different prompts. Our method shows robustness across a diverse set of prompts and styles. Our inflated attention modules can capture motion-specific traits, such as shadows (Row 3, Row 7), lighting changes (Row 1), and reasonable deformations that are not modeled by the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7616" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 542, + 133 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 542, + 133 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 542, + 133 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 542, + 133 + ], + "type": "image", + "image_path": "62a31c81aead55562e78dc8ee695237fda05241d8a94c0a991a141f4d7a9d304.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 213, + 133, + 380, + 140 + ], + "lines": [ + { + "bbox": [ + 213, + 133, + 380, + 140 + ], + "spans": [ + { + "bbox": [ + 213, + 133, + 380, + 140 + ], + "type": "text", + "content": "a shining diamond ball bouncing in a chamber" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 51, + 140, + 542, + 202 + ], + "blocks": [ + { + "bbox": [ + 51, + 140, + 542, + 202 + ], + "lines": [ + { + "bbox": [ + 51, + 140, + 542, + 202 + ], + "spans": [ + { + "bbox": [ + 51, + 140, + 542, + 202 + ], + "type": "image", + "image_path": "5ae8460dad3f11cd0a0712af65f36559a9aedde509b48522e15902f472ea0425.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 231, + 201, + 362, + 210 + ], + "lines": [ + { + "bbox": [ + 231, + 201, + 362, + 210 + ], + "spans": [ + { + "bbox": [ + 231, + 201, + 362, + 210 + ], + "type": "text", + "content": "a fox running in a lively green forest" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 51, + 210, + 542, + 272 + ], + "blocks": [ + { + "bbox": [ + 51, + 210, + 542, + 272 + ], + "lines": [ + { + "bbox": [ + 51, + 210, + 542, + 272 + ], + "spans": [ + { + "bbox": [ + 51, + 210, + 542, + 272 + ], + "type": "image", + "image_path": "a56ffbfad99db634f52080aac741a62ef4f99417af0791059c27e57a8e7e5ae2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 264, + 271, + 329, + 279 + ], + "lines": [ + { + "bbox": [ + 264, + 271, + 329, + 279 + ], + "spans": [ + { + "bbox": [ + 264, + 271, + 329, + 279 + ], + "type": "text", + "content": "the Joker dancing" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 51, + 279, + 542, + 342 + ], + "blocks": [ + { + "bbox": [ + 51, + 279, + 542, + 342 + ], + "lines": [ + { + "bbox": [ + 51, + 279, + 542, + 342 + ], + "spans": [ + { + "bbox": [ + 51, + 279, + 542, + 342 + ], + "type": "image", + "image_path": "83cc8ddd47be6de75b625f28ed4bdba481870dcc5a0b26f12d465c5298d17837.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 341, + 373, + 349 + ], + "lines": [ + { + "bbox": [ + 220, + 341, + 373, + 349 + ], + "spans": [ + { + "bbox": [ + 220, + 341, + 373, + 349 + ], + "type": "text", + "content": "a little girl walking around, Van Gogh style" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 51, + 349, + 542, + 410 + ], + "blocks": [ + { + "bbox": [ + 51, + 349, + 542, + 410 + ], + "lines": [ + { + "bbox": [ + 51, + 349, + 542, + 410 + ], + "spans": [ + { + "bbox": [ + 51, + 349, + 542, + 410 + ], + "type": "image", + "image_path": "7a846c22002c0546bf3281d5e63c540a22fbf2e07e90d2071ec98dcc710032f7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 410, + 370, + 418 + ], + "lines": [ + { + "bbox": [ + 223, + 410, + 370, + 418 + ], + "spans": [ + { + "bbox": [ + 223, + 410, + 370, + 418 + ], + "type": "text", + "content": "a little girl walking around, Da Vinci style" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 51, + 418, + 542, + 479 + ], + "blocks": [ + { + "bbox": [ + 51, + 418, + 542, + 479 + ], + "lines": [ + { + "bbox": [ + 51, + 418, + 542, + 479 + ], + "spans": [ + { + "bbox": [ + 51, + 418, + 542, + 479 + ], + "type": "image", + "image_path": "f9f6183bcf9dd417ebb16915e1c8d32ffd097e3a6b2aca8fc16078f4da1d6825.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 250, + 479, + 343, + 487 + ], + "lines": [ + { + "bbox": [ + 250, + 479, + 343, + 487 + ], + "spans": [ + { + "bbox": [ + 250, + 479, + 343, + 487 + ], + "type": "text", + "content": "a stormtrooper swimming" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 51, + 487, + 542, + 548 + ], + "blocks": [ + { + "bbox": [ + 51, + 487, + 542, + 548 + ], + "lines": [ + { + "bbox": [ + 51, + 487, + 542, + 548 + ], + "spans": [ + { + "bbox": [ + 51, + 487, + 542, + 548 + ], + "type": "image", + "image_path": "907f33d096ce5e1401d8cec552912f50bc0d6561d80ce2351e0c5bd1fdebc2e2.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 249, + 548, + 345, + 557 + ], + "lines": [ + { + "bbox": [ + 249, + 548, + 345, + 557 + ], + "spans": [ + { + "bbox": [ + 249, + 548, + 345, + 557 + ], + "type": "text", + "content": "three slendermen dancing" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 100, + 562, + 490, + 573 + ], + "lines": [ + { + "bbox": [ + 100, + 562, + 490, + 573 + ], + "spans": [ + { + "bbox": [ + 100, + 562, + 490, + 573 + ], + "type": "text", + "content": "Figure 5. Qualitative results. Our method generalizes to a diverse set of input meshes, styles, and prompts." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 594, + 251, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 594, + 251, + 605 + ], + "spans": [ + { + "bbox": [ + 47, + 594, + 251, + 605 + ], + "type": "text", + "content": "input low-fidelity meshes (Row 1 ground regions)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 618, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 713 + ], + "type": "text", + "content": "Quantitative results. We show quantitative evaluations for frame consistency and prompt fidelity in Tab. 1. We observe that our frame consistency is better than all baselines using T2I models. As expected, Pix2Video [6] suffers more from temporal changes due to the lack of a correspondence injection module. Even though TensorFlow's token propagation encourages smoothness over fidelity, our method still shows better consistency. Only the T2V model Gen1 has a better" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 594, + 547, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 702 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 702 + ], + "type": "text", + "content": "frame consistency, which is expected because it is trained on video data. All current T2V models, including Gen1, can only generate a few frames and are limited in other ways. Our approach is an attempt to overcome this deficiency using 2D T2I models exclusively. Fidelity-wise, we are slightly better than Pix2Video [6], because it reuses the set of pre-attention features from previous frames, while we only reuse a portion of the pre-attention features to generate each frame. Since TokenFlow [10]'s token propagation mechanism in" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7617" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 70, + 296, + 189 + ], + "blocks": [ + { + "bbox": [ + 52, + 70, + 296, + 189 + ], + "lines": [ + { + "bbox": [ + 52, + 70, + 296, + 189 + ], + "spans": [ + { + "bbox": [ + 52, + 70, + 296, + 189 + ], + "type": "image", + "image_path": "e6e68b1da0bc89dc23021d313a39e5e63e302acba770b95f8dce37886bd13aee.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 192, + 527, + 204 + ], + "lines": [ + { + "bbox": [ + 64, + 192, + 527, + 204 + ], + "spans": [ + { + "bbox": [ + 64, + 192, + 527, + 204 + ], + "type": "text", + "content": "Figure 6. Ablation study. We disabled several key building blocks and observe clear artifacts or inconsistencies in all ablations." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 299, + 70, + 543, + 189 + ], + "blocks": [ + { + "bbox": [ + 299, + 70, + 543, + 189 + ], + "lines": [ + { + "bbox": [ + 299, + 70, + 543, + 189 + ], + "spans": [ + { + "bbox": [ + 299, + 70, + 543, + 189 + ], + "type": "image", + "image_path": "f85826a741fc862bbae0d6869a5fd1a0de4ce57fe9fa26ef49b3ab5c5571bb01.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 48, + 222, + 291, + 308 + ], + "blocks": [ + { + "bbox": [ + 48, + 222, + 291, + 308 + ], + "lines": [ + { + "bbox": [ + 48, + 222, + 291, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 222, + 291, + 308 + ], + "type": "table", + "html": "
MethodFrame Consistency ↑Prompt Fidelity ↑
Per-Frame0.95470.3233
Pix2Video*0.96300.2983
TokenFlow*0.98220.2737
Gen10.99070.3029
Ours0.98450.3227
", + "image_path": "6f560122ea8d7d1be833d42190bc8929aa3cd044f6864ede6b0abfc0cf9741eb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 316, + 288, + 383 + ], + "lines": [ + { + "bbox": [ + 46, + 316, + 288, + 383 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 288, + 383 + ], + "type": "text", + "content": "Table 1. Quantitative evaluation. Our approach offers the best frame consistency among approaches using T2I models. As expected, only the proprietary T2V model Gen1 observes better temporal consistency. Our method also shows a higher prompt fidelity than relevant baselines that offer a reasonable amount of temporal consistency. *baseline adapted to our setting." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 403, + 288, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 288, + 501 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 288, + 501 + ], + "type": "text", + "content": "produces over-smoothing effects, its results are blurry for many of the samples, hence achieving lower fidelity scores. Our generative rendering utilizes the benefits of pre-attention inflation to gain overall global structural and semantic consistency, while also combining with post-attention feature warping as local guidance. Combined with our UV noise initialization, we outperform or are on par with all relevant baselines in the proposed setting." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": "Ablation studies. We perform a thorough ablation study to verify our design choices, shown in Fig. 6. We study three aspects of our pipeline: 1) comparing with pre-attention injection only, 2) comparing with post-attention injection only, and 3) the effectiveness of UV noise initialization. We observe that directly unifying and projecting the post-attention features without passing through the attention module produces chaotic results (severe artifacts in \"no pre-attn\", the joker's face and clothes regions). This observation is in line with our hypothesis for the poor performance of the adapted TokenFlow mechanism. When the motion is relatively large, geometrically matched features may not be good matches in the feature space. Therefore, directly warping a feature from one frame to another based on the geometrical correspondence has a high chance of not producing reasonable outputs. However, when we combine post-attention feature warping with pre-attention feature injection, the post-attention fea" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 224, + 547, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 224, + 547, + 370 + ], + "spans": [ + { + "bbox": [ + 304, + 224, + 547, + 370 + ], + "type": "text", + "content": "tures can act as guidance for the next network layer. Without post-attention feature warping, purely pre-attention feature injection does not guarantee pixel-wise consistency since the attention operation merely computes a correspondence based on feature similarities (in “no post-attn”, the joker’s clothes change over frames). Our UV noise initialization is well-aligned with our setup. If we use fixed noise instead, the blending of the features may cause unwanted artifacts (in “fixed noise”, the joker’s face has blending artifacts and is not recognizable in the third image). Our UV noise initialization serves generally and can work beyond our setup. Additional details on ablations are included in the supplementary." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 395, + 375, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 395, + 375, + 407 + ], + "spans": [ + { + "bbox": [ + 306, + 395, + 375, + 407 + ], + "type": "text", + "content": "5. Discussion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 420, + 547, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 420, + 547, + 588 + ], + "spans": [ + { + "bbox": [ + 304, + 420, + 547, + 588 + ], + "type": "text", + "content": "In this paper, we introduce Generative Rendering, a novel zero-shot pipeline based on 2D diffusion models for 4D-conditioned animation generation. Our method can animate creator-defined low-fidelity meshes and motion sequences, thereby bypassing steps requiring significant manual labor such as detailing, texturing, physical simulation, etc. The key idea of generative rendering is to utilize the prior within a depth-conditioned 2D diffusion model to provide the basic structure and physical fidelity to make convincing animations. This is accomplished by injecting the correspondences into the diffusion model using a combination of pre- and post-attention feature injection while unifying these features in the UV space. Our model demonstrates better frame consistency and prompt fidelity than relevant baselines." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 602, + 547, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 602, + 547, + 675 + ], + "spans": [ + { + "bbox": [ + 304, + 602, + 547, + 675 + ], + "type": "text", + "content": "Conclusion. Diffusion models have emerged as state-of-the-art generative methods across multiple domains, such as 2D, 3D, and video generation and editing. generative rendering adds unprecedented levels of control to video generation with T2I diffusion models by bridging the gap between the traditional rendering pipeline and diffusion models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "Acknowledgement. This project was in part supported by Samsung, Google and Stanford HAI." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7618" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 710 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 288, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 288, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 288, + 123 + ], + "type": "text", + "content": "[1] Yuval Alaluf, Daniel Garibi, Or Patashnik, Hadar Averbuch-Elor, and Daniel Cohen-Or. Cross-image attention for zero-shot appearance transfer. In arXiv, 2023. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 288, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 288, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 288, + 156 + ], + "type": "text", + "content": "[2] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In ECCV, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 158, + 288, + 200 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 288, + 200 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 288, + 200 + ], + "type": "text", + "content": "[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 1, 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 201, + 288, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 201, + 288, + 255 + ], + "spans": [ + { + "bbox": [ + 53, + 201, + 288, + 255 + ], + "type": "text", + "content": "[4] Shengqu Cai, Eric Ryan Chan, Songyou Peng, Mohamad Shahbazi, Anton Obukhov, Luc Van Gool, and Gordon Wetzstein. Diffdreamer: Towards consistent unsupervised single-view scene extrapolation with conditional diffusion models. In ICCV, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 257, + 288, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 257, + 288, + 289 + ], + "spans": [ + { + "bbox": [ + 53, + 257, + 288, + 289 + ], + "type": "text", + "content": "[5] Tianshi Cao, Karsten Kreis, Sanja Fidler, Nicholas Sharp, and Kangxue Yin. Texfusion: Synthesizing 3d textures with text-guided image diffusion models. In ICCV, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 290, + 288, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 290, + 288, + 322 + ], + "spans": [ + { + "bbox": [ + 53, + 290, + 288, + 322 + ], + "type": "text", + "content": "[6] Duygu Ceylan, Chun-Hao Huang, and Niloy J. Mitra. Pix2video: Video editing using image diffusion. In ICCV, 2023. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 323, + 288, + 366 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 323, + 288, + 366 + ], + "spans": [ + { + "bbox": [ + 53, + 323, + 288, + 366 + ], + "type": "text", + "content": "[7] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In ICCV, 2023. 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 368, + 288, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 368, + 288, + 399 + ], + "spans": [ + { + "bbox": [ + 53, + 368, + 288, + 399 + ], + "type": "text", + "content": "[8] Rafail Fridman, Amit Abecasis, Yoni Kasten, and Tali Dekel. Scenescape: Text-driven consistent scene generation. In arXiv, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 401, + 288, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 401, + 288, + 444 + ], + "spans": [ + { + "bbox": [ + 53, + 401, + 288, + 444 + ], + "type": "text", + "content": "[9] Songwei Ge, Seungjun Nah, Guilin Liu, Tyler Poon, Andrew Tao, Bryan Catanzaro, David Jacobs, Jia-Bin Huang, Ming-Yu Liu, and Yogesh Balaji. Preserve your own correlation: A noise prior for video diffusion models. In ICCV, 2023. 1, 2, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 445, + 288, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 288, + 477 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 288, + 477 + ], + "type": "text", + "content": "[10] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. In arXiv, 2023. 1, 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 478, + 288, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 478, + 288, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 478, + 288, + 521 + ], + "type": "text", + "content": "[11] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. In arXiv, 2023. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 522, + 288, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 288, + 565 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 288, + 565 + ], + "type": "text", + "content": "[12] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2022. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 567, + 288, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 288, + 620 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 288, + 620 + ], + "type": "text", + "content": "[13] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Image video: High definition video generation with diffusion models. In arXiv, 2022. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 621, + 288, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 621, + 288, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 621, + 288, + 643 + ], + "type": "text", + "content": "[14] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In arXiv, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 644, + 288, + 675 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 644, + 288, + 675 + ], + "spans": [ + { + "bbox": [ + 48, + 644, + 288, + 675 + ], + "type": "text", + "content": "[15] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. In arXiv, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 677, + 288, + 710 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 677, + 288, + 710 + ], + "spans": [ + { + "bbox": [ + 48, + 677, + 288, + 710 + ], + "type": "text", + "content": "[16] Lukas Hollein, Ang Cao, Andrew Owens, Justin Johnson, and Matthias Nießner. Text2room: Extracting textured 3d meshes from 2d text-to-image models. In ICCV, 2023. 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 700 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[17] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly ddpm noise space: Inversion and manipulations. In arXiv, 2023. 2, 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "type": "text", + "content": "[18] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In NeurIPS, 2022. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 141, + 545, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 163 + ], + "type": "text", + "content": "[19] Michael Kass and Davide Pesare. Coherent noise for nonphotorealistic rendering. In TOG, 2011. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 164, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 164, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 164, + 545, + 217 + ], + "type": "text", + "content": "[20] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. In arXiv, 2023. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 220, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 262 + ], + "type": "text", + "content": "[21] Xin Li, Wenqing Chu, Ye Wu, Weihang Yuan, Fanglong Liu, Qi Zhang, Fu Li, Haocheng Feng, Errui Ding, and Jingdong Wang. Videogen: A reference-guided latent diffusion approach for high definition text-to-video generation, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 264, + 545, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 307 + ], + "type": "text", + "content": "[22] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 309, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 352 + ], + "type": "text", + "content": "[23] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 354, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 398 + ], + "type": "text", + "content": "[24] Hao Ouyang, Qiuyu Wang, Yuxi Xiao, Qingyan Bai, Juntao Zhang, Kecheng Zheng, Xiaowei Zhou, Qifeng Chen, and Yujun Shen. Codef: Content deformation fields for temporally consistent video processing. In arXiv, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 399, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 430 + ], + "type": "text", + "content": "[25] Ryan Po and Gordon Wetzstein. Compositional 3d scene generation using locally conditioned diffusion. In arXiv, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 433, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 433, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 433, + 545, + 487 + ], + "type": "text", + "content": "[26] Ryan Po, Wang Yifan, Vladislav Golyanik, Kfir Aberman, Jonathan T Barron, Amit H Bermano, Eric Ryan Chan, Tali Dekel, Aleksander Holynski, Angjoo Kanazawa, et al. State of the art on diffusion models for visual computing. arXiv preprint arXiv:2310.07204, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 488, + 545, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 488, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 307, + 488, + 545, + 520 + ], + "type": "text", + "content": "[27] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 522, + 545, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 522, + 545, + 565 + ], + "spans": [ + { + "bbox": [ + 307, + 522, + 545, + 565 + ], + "type": "text", + "content": "[28] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. In ICCV, 2023. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 567, + 545, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 545, + 621 + ], + "type": "text", + "content": "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In CoRR, 2021. 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 623, + 545, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 623, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 307, + 623, + 545, + 666 + ], + "type": "text", + "content": "[30] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. In TPAMI, 2022. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 668, + 545, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 668, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 307, + 668, + 545, + 700 + ], + "type": "text", + "content": "[31] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. In SIGGRAPH, 2023. 3" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7619" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 274 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 115 + ], + "type": "text", + "content": "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 288, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 288, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 288, + 173 + ], + "type": "text", + "content": "[33] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-video: Text-to-video generation without text-video data. In arXiv, 2022. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 174, + 288, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 174, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 174, + 288, + 228 + ], + "type": "text", + "content": "[34] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1921-1930, June 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 229, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 229, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 229, + 288, + 262 + ], + "type": "text", + "content": "[35] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 263, + 288, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 263, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 263, + 288, + 274 + ], + "type": "text", + "content": "[36] Tianfu Wang, Menelaos Kanakis, Konrad Schindler, Luc Van" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 547, + 274 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 326, + 72, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 545, + 95 + ], + "type": "text", + "content": "Gool, and Anton Obukhov. Breathing new life into 3d assets with generative repainting. In BMVC, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 307, + 96, + 547, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 149 + ], + "type": "text", + "content": "[37] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In arXiv, 2022. 1, 2, 3, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 307, + 152, + 547, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 547, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 547, + 196 + ], + "type": "text", + "content": "[38] Sihyun Yu, Kihyuk Sohn, Subin Kim, and Jinwoo Shin. Video probabilistic diffusion models in projected latent space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18456-18466, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 307, + 197, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 229 + ], + "type": "text", + "content": "[39] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 1, 2, 3, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 308, + 231, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 231, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 231, + 545, + 274 + ], + "type": "text", + "content": "[40] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 1" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7620" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Generative Unlearning for Any Identity/c3c2d3ee-c7c2-4590-9220-bed4fbf9b764_content_list.json b/2024/Generative Unlearning for Any Identity/c3c2d3ee-c7c2-4590-9220-bed4fbf9b764_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5ac974de670c945c4f0c33babcf18706cb34cefe --- /dev/null +++ b/2024/Generative Unlearning for Any Identity/c3c2d3ee-c7c2-4590-9220-bed4fbf9b764_content_list.json @@ -0,0 +1,1510 @@ +[ + { + "type": "text", + "text": "Generative Unlearning for Any Identity", + "text_level": 1, + "bbox": [ + 282, + 130, + 687, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Juwon $\\mathrm{Seo}^{1*}$", + "bbox": [ + 161, + 179, + 264, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sung-Hoon Lee $^{1*}$ Tae-Young Lee $^{1*}$ Gyeong-Moon Park $^{1\\dagger}$", + "bbox": [ + 305, + 179, + 629, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Seungjun Moon²", + "bbox": [ + 669, + 180, + 807, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Kyung Hee University, Yongin, Republic of Korea \n2KLleon Tech., Seoul, Republic of Korea", + "bbox": [ + 282, + 232, + 689, + 270 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{jwseo001, sunghoonlee961, slcks1, gmpark}@khu.ac.kr seungjunmoon@klleon.io", + "bbox": [ + 254, + 277, + 712, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 345, + 313, + 363 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in generative models trained on large-scale datasets have made it possible to synthesize high-quality samples across various domains. Moreover, the emergence of strong inversion networks enables not only a reconstruction of real-world images but also the modification of attributes through various editing methods. However, in certain domains related to privacy issues, e.g., human faces, advanced generative models along with strong inversion methods can lead to potential misuses. In this paper, we propose an essential yet under-explored task called generative identity unlearning, which steers the model not to generate an image of specific identity. In the generative identity unlearning, we target the following objectives: (i) preventing the generation of images with a certain identity, and (ii) preserving the overall quality of the generative model. To satisfy these goals, we propose a novel framework, Generative Unlearning for Any IDENTITY (GUIDE), which prevents the reconstruction of a specific identity by unlearning the generator with only a single image. GUIDE consists of two parts: (i) finding a target point for optimization that un-identifies the source latent code and (ii) novel loss functions that facilitate the unlearning procedure while less affecting the learned distribution. Our extensive experiments demonstrate that our proposed method achieves state-of-the-art performance in the generative machine unlearning task. The code is available at https://github.com/KHU-AGI/GUIDE.", + "bbox": [ + 76, + 375, + 473, + 782 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 810, + 209, + 825 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, 2D or 3D Generative Adversarial Networks (GANs) [4, 19-21] pre-trained on large datasets, e.g., FFHQ", + "bbox": [ + 76, + 835, + 470, + 867 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/30461510649baa1c9b78705c4fac198e76a80a95d1ce761366dd47823862bd32.jpg", + "image_caption": [ + "Figure 1. Given a single source image containing a specific identity, we remove that identity from the pre-trained 3D generative adversarial network (e.g., EG3D [4]). Our method effectively unlearns identity even from in-the-wild images where the source image is absent in the pre-training dataset." + ], + "image_footnote": [], + "bbox": [ + 521, + 348, + 874, + 566 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "[19] or AFHQ [5], have drawn substantial attention due to their remarkable generation performance and highly disentangled representation space. However, their advancements have raised privacy concerns [15], especially regarding the potential misuse of generative models to represent and exploit individual identities. For instance, deepfakes [48, 49] can create very believable images or videos of people in made-up situations, causing major concerns about ethics and privacy.", + "bbox": [ + 496, + 672, + 892, + 809 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To alleviate privacy issues in generative models, machine unlearning task has been actively studied. Machine unlearning involves the process of selectively removing specific knowledge or erasing the influence of certain data from the training dataset of pre-trained models. It is beneficial especially when the data are harmful, private, or biased", + "bbox": [ + 496, + 810, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution", + "bbox": [ + 94, + 875, + 202, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author", + "bbox": [ + 96, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9151", + "bbox": [ + 480, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "[11, 12, 26, 31]. Despite a focus on discriminative tasks in most machine unlearning research, a few studies have ventured into generative models, attempting to erase high-level concepts such as socially inappropriate content or artistic styles that present copyright challenges [9, 25].", + "bbox": [ + 75, + 90, + 468, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Nevertheless, generative models still exhibit ongoing privacy issues. Even if an identity of someone is not used in the pre-training of the generative models, it can be easily reconstructed in the pre-trained models via GAN inversion models [29, 34, 35, 45, 46, 52]. Furthermore, the reconstructed image can be manipulated or edited easily via image editing methods [30, 39, 40]. To prevent potential exploits of an identity, it is necessary to erase a certain identity from the pre-trained generative models.", + "bbox": [ + 75, + 169, + 467, + 305 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To consider the above issue, we introduce an essential task of unlearning any identity from the pre-trained 2D or 3D GANs [4, 20], called generative identity unlearning. Unlike typical machine unlearning tasks, which focus on unlearning the training samples our generative identity unlearning task unlearns any identity on pre-trained GANs, even if it was not shown during the pre-training. Our goal is to remove the whole identity associated with a given single image from the generator while minimally impacting the overall performance of the pre-trained model.", + "bbox": [ + 75, + 306, + 467, + 458 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To achieve our goal, we propose a novel generative unlearning framework, Generative Unlearning for Any IDEntity, named GUIDE. GUIDE replaces the source identity with an anonymous target identity, erasing the original identity effectively. To this end, we propose a new exploration method to determine an effective target latent code, called Un-Identifying Face On latent space (UFO). UFO utilizes the GAN inversion method [52] to embed the given identity into the source latent, and then decides the target latent using both the source and the average latent codes. We empirically find that the proposed UFO can identify the promising target to erase any given source identity robustly.", + "bbox": [ + 75, + 460, + 467, + 642 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given the source and target latent code, we update the generator to shift from the source identity to the target identity. To this end, we propose three novel loss functions: (i) local unlearning loss, (ii) adjacency-aware unlearning loss, and (iii) global preservation loss. (i) guides our model directly shifting the source identity to the target identity. (ii) utilizes other latent codes adjacent to the source and target latent codes to effectively unlearn the entire identity from a single image. To minimize side effects from the unlearning process, (iii) additionally regularizes the generator to retain generation performance for latent codes relatively far from the source and target latent codes. Through comprehensive experiments on diverse identities, including Random, InD, and OOD, we confirm that GUIDE can successfully remove the identity of the source image from the pre-trained generative model, and shows qualitatively and quantitatively superior performances.", + "bbox": [ + 75, + 643, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 517, + 90, + 844, + 104 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- For the first time, we propose a novel task, generative identity unlearning, which tackles machine unlearning in generative models in the aspect of privacy protection. In our task, we aim to prevent the pre-trained generative models from synthesizing the given identity by utilizing only a single image.", + "- For the effective and robust elimination of the identity, we propose a novel method - Un-Identifying Face On Latent Space (UFO). We configure the unlearning procedure by formulating how to represent and shift the identity in the latent space. We find that setting the extrapolated latent code between the source and average latent codes as an optimization target facilitates the unlearning procedure.", + "- We propose three loss functions - local unlearning loss, adjacency-aware unlearning loss, and global preservation loss to effectively unlearn the identity from the pretrained model while less affecting the generation performance on other identity.", + "- We show that our proposed framework, GUIDE achieves state-of-the-art performance both qualitatively and quantitatively, through extensive experiments. We demonstrate that GUIDE can remove the specific identity successfully in the generative models while minimizing the negative effect on other identities." + ], + "bbox": [ + 500, + 107, + 890, + 468 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 482, + 638, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generative Models and Privacy Issue. In image synthesis field, GAN-based generative models have achieved remarkable performance not only in 2D [18-21, 32, 33, 36, 37] but also in 3D [4, 42, 47, 55, 56] domain. The application of various image editing methods [30, 40, 50] to strong generative models, people can easily generate edited images of specific individuals and various artistic styles [38, 41], as well as extract copyrighted content [3] without the permission of the individual or the original creator.", + "bbox": [ + 496, + 508, + 890, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, with the rise of the importance of AI ethics, several works have addressed this issue [9, 25, 28, 44, 53]. ESD [9] erases specific visual concepts from diffusion model by using negative guidance about the undesired concepts. Kumari et al. [25] modifies the conditional distribution of the model a specific target concepts to match the anchor concept. Forget-Me-Not [53] fine-tunes U-Net to minimize each of the intermediate attention associated with the target concepts to remove. Additional works in GANs [28, 44] focus on unlearning specific features, e.g., \"Bang\", \"Hat\" or \"Beard\" rather than forgetting specific identity. The above methods primarily concentrate on the elimination of specific concepts or high-level features. In other words, these cannot preclude models from generating specific individuals while maintaining the generation performance of realistic human faces. Unlike the existing works, our work targets to unlearn only specific individuals without", + "bbox": [ + 496, + 643, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9152", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "shifting the overall distribution of generated images.", + "bbox": [ + 76, + 90, + 421, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Machine Unlearning. Machine unlearning aims to selectively forget specific acquired knowledge or diminish the impact of certain training data subsets on a trained model. Since previous research [8, 22, 51] shows that machine learning models might accidentally share private information when faced with certain attacks or inputs, machine unlearning becomes crucial.", + "bbox": [ + 75, + 126, + 467, + 229 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While previous machine unlearning is mainly focused on supervised learning tasks [2, 6, 10-14, 43], the interest in unlearning techniques within unsupervised learning, i.e., generative models, is growing [24, 28, 44]. However, most of the existing methods need full dataset access for retraining, which is hard to acquire and computationally expensive [2, 10, 14, 43]. For example, Kong and Chaudhuri [24] utilizes data reduction and augmentation algorithms, which requires a full training dataset. Despite the existence of a feature unlearning model [28] which does not need full dataset access, unlearning only an individual feature is not enough to forget a whole specific identity. To this end, we propose an algorithm that enables forgetting the specific identity only with a single image. Furthermore, our approach distinguishes itself from existing research by applying unlearning to unseen images, enabling the erasure of specific identities without prior exposure to those images.", + "bbox": [ + 75, + 231, + 468, + 489 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 503, + 168, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Firstly, in Section 3.1, as shown in Figure 2, we introduce the problem we aim to address, named generative identity unlearning. In Section 3.2, we introduce un-identifying face on latent space, which designates an appropriate target latent for unlearning. Then, in Section 3.3, we introduce latent target unlearning, along with our proposed novel losses to unlearn the generator. The total overview of our method is illustrated in Figure 3.", + "bbox": [ + 75, + 523, + 468, + 647 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 76, + 656, + 279, + 671 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a set of images $\\mathbf{x}$ depicting a specific identity, we randomly select a single source image $x_{u} \\in \\mathbf{x}$ as an exemplar of the identity. Initially, using off-the-shelf inversion network [52] $E$ corresponding to the unconditional generator, i.e., EG3D [4], we embed $x_{u}$ to the source latent code $w_{u}$ in the latent space of EG3D:", + "bbox": [ + 75, + 680, + 468, + 771 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nw _ {u} = E \\left(x _ {u}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 223, + 781, + 468, + 797 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since EG3D [4] consists of the mapping network $Map(\\cdot)$ , StyleGAN2 [19] backbone $G(\\cdot)$ and the neural renderer with a super-resolution module $R(\\cdot)$ , we can denote the reconstructed image $\\hat{x}$ from $w_{u}$ as following:", + "bbox": [ + 75, + 809, + 468, + 869 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {x} = R \\left(G \\left(w _ {u}\\right); c\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 209, + 880, + 468, + 896 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/cde3a6799f641e5b379ca23096bc13eef2febd497281e44fb2fad14724196e95.jpg", + "image_caption": [ + "Figure 2. An illustration of generative identity unlearning. Upon GUIDE, the identity of the image generated from $w_{u}$ , i.e., inversion of the source image $x_{u}$ by inversion network $E$ , should exhibit a distinct identity when passed through the pre-trained generator $G_{s}$ compared to the unlearned generator $G_{u}$ . Furthermore, other images $x_{o}$ , not used in unlearning but sharing the same identity with $x_{u}$ , also should vary an identity through GUIDE." + ], + "image_footnote": [], + "bbox": [ + 514, + 90, + 879, + 244 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $c$ denotes camera poses. For convenience, we omit the explicit notation of camera poses in this paper, i.e., $\\hat{x} = (R \\circ G)(w_u)$ . We target to derive an unlearned $G$ , i.e., $G_u$ , from the pre-trained EG3D generator $G$ , i.e., $G_s$ , while fixing $Map$ and $R$ . With proper unlearning, an image generated by unlearned EG3D using $w_u$ , i.e., $\\hat{x}_u = (R \\circ G_u)(w_u)$ should have a distinct identity from the image generated by original EG3D using $w_u$ , i.e., $(R \\circ G_s)(w_u)$ .", + "bbox": [ + 496, + 381, + 890, + 502 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our task formulation, two considerations are paramount. First, we aim to eliminate the entire identity from the generator only utilizing a single image. To validate this, we incorporate other multiple images $x_{o} \\in \\mathbf{x}$ for testing and its corresponding latent code $w_{o} = E(x_{o})$ . By utilizing $x_{o}$ , we can verify whether $G_{u}$ has successfully unlearned the identity as a whole, rather than just unlearning the specific image $x_{u}$ . Second, we strive to maintain the generation performance of the pre-trained model. To assess this, we sample multiple images from fixed latent codes using both the unlearned and pre-trained generators. We then estimate the distribution shift between the images generated before and after the unlearning process.", + "bbox": [ + 496, + 503, + 892, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Un-Identifying Face On Latent Space", + "text_level": 1, + "bbox": [ + 500, + 710, + 823, + 726 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The successfully unlearned model should not generate the image with the identity of $\\mathbf{x}$ , even when $w_{u}$ is used as a latent. Consequently, we initiate our approach to manipulate $\\hat{x}_u$ to be another image rather than the image with identity in $\\mathbf{x}$ by unlearning $G$ . To design the objective function for unlearning, we first need to establish the objective for unlearning, which involves defining the target image $\\hat{x}_t$ , derived from $w_t$ , which the unlearned image $\\hat{x}_u$ , derived from $w_{u}$ , should mimic after unlearning. While there exist various options for setting $\\hat{x}_t$ , e.g., a random face or even a non-human image, we choose the mean face generated by", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9153", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ffa2618d4725731f575a1660beb56ea64b94f932c30b6ff96b23cdc5937bf14f.jpg", + "image_caption": [ + "Figure 3. An overview of GUIDE. Starting with a source image, we employ a GAN inversion network $E$ , specifically GOAE [52], to embed this image into the latent space of a pre-trained generative model, namely EG3D [4], obtaining the source latent code $w_{u}$ . The target latent code $w_{t}$ is designated through the UFO process. To facilitate identity removal in $w_{u}$ , we shift its identity to match that of $w_{t}$ with our Latent Target Unlearning (LTU) process. Three loss functions of LTU are designed for this purpose: (i) The generator is optimized to produce an image from the source latent code, denoted as $(R \\circ G_{u})(w_{u})$ , that is similar to the image from the target latent code, represented as $(R \\circ G_{s})(w_{t})$ . (ii) To achieve unlearning across the entire identity, we consider latent codes near both the source and target latent codes, denoted as $w_{u,a}$ and $w_{t,a}$ , respectively. (iii) To prevent model corruption during the unlearning process, we additionally sample latent codes from a random noise vector, represented as $w_{r,g}$ , and optimize $G_{u}$ to preserve its generation ability on $w_{r,g}$ ." + ], + "image_footnote": [], + "bbox": [ + 83, + 92, + 893, + 315 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a498b047462a15846977f2f41cc34875a1ab9f8e3342a8b5deab74778be983cf.jpg", + "image_caption": [ + "Figure 4. An illustration of Un-identifying Face On Latent Space (UFO). We define the identity of the source latent code by subtract it from the average latent code. We set the target latent code for our unlearning process by measuring an extrapolation between the source and average latent code with a fixed distance $d$ ." + ], + "image_footnote": [], + "bbox": [ + 119, + 467, + 428, + 579 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the mean latent $\\bar{w}$ of $Map$ , i.e., $(R \\circ G_s)(\\bar{w})$ , as $\\hat{x}_t$ . This setting is inspired by inversion methods [1, 29] which apply the identity to the mean face through inversion stages. Since our goal is to de-identify the image, we argue that it is intuitive and reasonable to send the image back to the mean face, which is opposite to the inversion.", + "bbox": [ + 75, + 672, + 468, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, this simple approach might be problematic when $w_{u}$ is close to $\\bar{w}$ , where the unlearned image might still resemble the original identity. To this end, we propose a novel method named Un-identifying Face On latent space (UFO), which can set $\\hat{x}_t$ robustly regardless of the distance between $w_{u}$ and $\\bar{w}$ , as shown in Figure 4. The following are the processes of UFO: First, in the $de$ -identification process, we extract the identity latent $w_{id} = w_{u} - \\bar{w}$ by subtracting $\\bar{w}$ from $w_{u}$ . Then we propose a process of en-identification,", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "setting the target in the opposite direction of the existing $w_{id}$ to foster a more pronounced change in identity, creating an entirely new identity. In other words, we chose $\\bar{w}$ as the stopover point and established the final target point at the extrapolation of $w_{u}$ and $\\bar{w}$ . The finalized target point can be expressed as:", + "bbox": [ + 498, + 465, + 890, + 556 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nw _ {t} = \\bar {w} - d \\cdot \\frac {w _ {i d}}{\\| w _ {i d} \\| _ {2}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 561, + 890, + 592 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $d$ is defined as the distance that balances un-identification with preservation of the source distribution, as determined by our empirical results. Finally, we can set $\\hat{x}_t = (R\\circ G_s)(w_t)$ . We empirically demonstrate that whether the given identity is in close proximity to the average latent code within the latent space or not, UFO sets the desirable target latent code for the generative unlearning.", + "bbox": [ + 496, + 603, + 890, + 709 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Latent Target Unlearning", + "text_level": 1, + "bbox": [ + 500, + 719, + 733, + 736 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After setting $\\hat{x}_t$ , we need to robustly design the unlearning process that can effectively make $\\hat{x}_u$ to be similar to $\\hat{x}_t$ , while maintaining the generation performance of $G_u$ . We coin this process as Latent Target Unlearning (LTU), which targets to unlearn images from the specific latent while keeping generation results from other latent codes. LTU utilizes the following three losses to achieve this goal:", + "bbox": [ + 496, + 743, + 890, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Local Unlearning Loss. To force $\\hat{x}_u$ to be $\\hat{x}_t$ , we use the widely-used reconstruction losses, i.e., Euclidean loss $\\mathcal{L}_2$ ,", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9154", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9db8382f578fb7d11023da21d017b3da352967bb5ef2e563a6f054bdafcc50d1.jpg", + "image_caption": [ + "Figure 5. An illustration of determining latent codes near a latent code $w$ in adjacency-aware unlearning loss. We first sample a latent code $w_{r,a}$ which is derived from a random noise vector $z_{r,a}$ via the mapping network $\\text{Map}(\\cdot)$ , i.e. $w_{r,a} = \\text{Map}(z_{r,a})$ . Next, we compute the direction between $w$ and $w_{r,a}$ , and we scale it to fall within range between 0 and $\\alpha_{max}$ . This yields the distance vector $\\Delta$ to compute the adjacent latent code $w_{a} = w + \\Delta$ ." + ], + "image_footnote": [], + "bbox": [ + 86, + 89, + 472, + 229 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "perceptual loss $\\mathcal{L}_{per}$ [54], and identity loss $\\mathcal{L}_{id}$ [7]. Using $\\mathcal{L}_{recon}$ , we compare the tri-plane features $F_{u} = G_{u}(w_{u})$ and $F_{t} = G_{s}(w_{t})$ , derived from source and target latent codes, respectively. The local unlearning loss is defined as:", + "bbox": [ + 76, + 361, + 468, + 422 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {l o c a l}} \\left(\\hat {x} _ {u}, \\hat {x} _ {t}\\right) = \\lambda_ {L 2} \\mathcal {L} _ {L 2} \\left(F _ {u}, F _ {t}\\right) \\tag {4} \\\\ + \\lambda_ {p e r} \\mathcal {L} _ {p e r} (\\hat {x} _ {u}, \\hat {x} _ {t}) + \\lambda_ {i d} \\mathcal {L} _ {i d} (\\hat {x} _ {u}, \\hat {x} _ {t}). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 441, + 468, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By adopting $\\mathcal{L}_{local}$ , we can successfully un-identify the given source identity in $\\hat{x}_t$ .", + "bbox": [ + 76, + 494, + 468, + 526 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Adjacency-Aware Unlearning Loss. The above equation considers only one pair of source and target latent codes. However, images of a similar identity to the source identity can be obtained by introducing marginal perturbations to the latent code. For the successful unlearning of the given identity, we need to consider the neighborhood of both the source and the target latent codes. Consequently, as shown in Figure 5, we sample $N_{a}$ latent codes in the vicinity of the $w_{u}$ . Specifically, with the scale $\\alpha^{i}$ sampled from the uniform distribution with hyperparameter $\\alpha_{max}$ , i.e. $\\alpha^{i} \\sim \\mathcal{U}(0, \\alpha_{max})$ , we define the distances $\\Delta$ to compute the adjacent latent codes as:", + "bbox": [ + 76, + 551, + 472, + 734 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta = \\left\\{\\alpha^ {i} \\cdot \\frac {w _ {r , a} ^ {i} - w _ {u}}{\\| w _ {r , a} ^ {i} - w _ {u} \\| _ {2}} \\right\\} _ {i = 1} ^ {N _ {a}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 168, + 751, + 468, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $w_{r,a}^{i}$ is a latent code sampled from the random noise vector $z_{r,a}^{i}$ . Using these distances $\\Delta$ , we can compute $N_{a}$ latent codes for both the source and the target latent codes. Similar to the local unlearning loss, we optimize the generated tri-plane features and images from $w_{u,a}^{i} = w_{u} + \\Delta^{i}$ and $w_{t,a}^{i} = w_{t} + \\Delta^{i}$ to be similar:", + "bbox": [ + 76, + 806, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {x} _ {u, a} ^ {i} = R \\left(F _ {u, a} ^ {i}\\right), \\hat {x} _ {t, a} ^ {i} = R \\left(F _ {u, a} ^ {i}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 116, + 890, + 136 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {a d j} \\left(w _ {u}, w _ {t}\\right) = \\frac {1}{N _ {a}} \\sum_ {i = 1} ^ {N _ {a}} \\mathcal {L} _ {\\text {l o c a l}} \\left(\\hat {x} _ {u, a} ^ {i}, \\hat {x} _ {t, a} ^ {i}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 552, + 138, + 890, + 180 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $F_{u,a}^{i} = G_{u}(w_{u,a}^{i}), F_{t,a}^{i} = G_{s}(w_{t,a}^{i})$ denotes for triplane features, and $\\mathcal{L}_{local}$ in Equation 4. From $\\mathcal{L}_{adj}$ , we can further consider possible variations of the source identity.", + "bbox": [ + 496, + 191, + 890, + 238 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Global Preservation Loss. While the local unlearning loss and adjacency-aware unlearning loss mentioned above facilitate the removal of the source identity, we propose a global preservation loss to mitigate side effects arising from these unlearning loss functions. In the global preservation loss, we constrain the generator to maintain generation performance for latent codes that are relatively distant from both the source and target latent codes.", + "bbox": [ + 496, + 255, + 890, + 376 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To be precise, we sample $N_{g}$ latent codes $\\{w_{r,g}^{i}\\}_{i = 1}^{N_{g}}$ from random noise vectors $\\{z_{r,g}^{i}\\}_{i = 1}^{N_{g}}$ . We ensure that these do not overlap with the adjacent latent codes used in the adjacency-aware unlearning loss. Unlike the unlearning loss functions, we find that adopting only $L_{per}$ achieves a balanced performance between identity shift and model preservation. The global preservation loss is computed as:", + "bbox": [ + 496, + 376, + 890, + 489 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {x} _ {u, g} ^ {i} = (R \\circ G _ {u}) (w _ {r, g} ^ {i}),\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 500, + 787, + 520 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {x} _ {s, g} ^ {i} = \\left(R \\circ G _ {s}\\right) \\left(w _ {r, g} ^ {i}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 522, + 890, + 549 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {g l o b a l} (G _ {u}, G _ {s}) = \\frac {1}{N _ {g}} \\sum_ {i = 1} ^ {N _ {g}} \\mathcal {L} _ {p e r} (\\hat {x} _ {u, g} ^ {i}, \\hat {x} _ {s, g} ^ {i}).\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 544, + 843, + 584 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In summary, our final objective is:", + "bbox": [ + 500, + 595, + 727, + 611 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {l o c a l}} + \\lambda_ {\\text {a d j}} \\mathcal {L} _ {\\text {a d j}} + \\lambda_ {\\text {g l o b a l}} \\mathcal {L} _ {\\text {g l o b a l}}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 623, + 890, + 640 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 652, + 633, + 669 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 500, + 676, + 689, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "**Baseline.** Since we propose generative identity unlearning task for the first time, to evaluate the effectiveness of GUIDE, we constructed a simple baseline. In the baseline, we used the target latent code as the average one for the unlearning. During the unlearning, we updated the pre-trained generator using $\\mathcal{L}_{local}$ as described in Equation 4.", + "bbox": [ + 496, + 700, + 890, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. We built GUIDE based on the 3D generative adversarial network [4] pre-trained on FFHQ dataset [19]. We used GOAE [52] as a GAN inversion network to obtain the latent code from an image. The image resolution we used in our experiments is $512 \\times 512$ with a rendering resolution of $128 \\times 128$ . We used Adam optimizer", + "bbox": [ + 496, + 809, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9155", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/1f9c0a39103959cae32f313595f8c200c1b3f48f14e32978b51a2ccc16493a3c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SourceRandomIn-Domain (FFHQ)Out-of-Domain (CelebAHQ)
BaselineOursBaselineOursBaselineOursBaselineOursBaselineOurs
Unlearned
", + "bbox": [ + 84, + 93, + 890, + 311 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "[23] with a learning rate of $10^{-4}$ in the unlearning procedure. The hyperparameters used in the experiments were: $d = 30$ , $\\alpha_{max} = 15$ , $\\lambda_{L2} = 10^{-2}$ , $\\lambda_{per} = 1$ , $\\lambda_{id} = 10^{-1}$ , $N_a = N_g = 2$ , and $\\lambda_{adj} = \\lambda_{global} = 1$ .", + "bbox": [ + 76, + 391, + 468, + 454 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset and Scenarios. We evaluated GUIDE in three scenarios: Random, where we set an unlearning target image from a randomly sampled noise vector; InD (domain), where we sampled an image from the FFHQ dataset [19] used for pre-training; and OOD (out-of-domain), where the unlearning target image was sampled from the CelebAHQ dataset [18]. For InD and OOD, we used the GAN inversion network to obtain corresponding latent codes. For OOD scenario, we also conducted multi-image test since there were multiple images with a same identity in CelebAHQ. On the other hand, we performed only single-image test in the Random and InD scenarios.", + "bbox": [ + 75, + 472, + 470, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. We evaluated GUIDE on two key aspects. Firstly, we estimated the efficacy of our approach in preventing the generator from producing images similar to the unlearning target. We quantitatively measured similarity of identities (ID) using face recognition network, CurricularFace [17], between images generated from the same latent codes before and after unlearning. Moreover, we utilized $\\mathrm{ID}_{\\mathrm{others}}$ to estimate the erasure of a identity from images which are unseen during training but containing the same identity of the source image. Secondly, we assessed whether our method preserves overall generation performance using the Fréchet Inception distance (FID) score [16]. Different from the existing usages, we utilized two variants of FID. First, we evaluated the distribution shift of generated images the pre-trained generator and the un-", + "bbox": [ + 75, + 672, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/98eb4066c843cf5e7dbf10a6c817e84367ed08a86a316b95ada8d29f1f722a01.jpg", + "image_caption": [ + "Figure 6. Qualitative results of GUIDE and the baseline in generative identity unlearning task. For the given source image each (the first row), GUIDE and the baseline tried to erase the identity in the pre-trained generator. The images in the second and third row are the target and unlearned images, respectively.", + "Figure 7. Qualitative results of GUIDE and the baseline on a multi-image test using CelebAHQ dataset. We additionally utilized images that are unseen during unlearning, to show how thoroughly erase the given identity." + ], + "image_footnote": [], + "bbox": [ + 537, + 390, + 864, + 597 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "learned generator via $\\mathrm{FID}_{\\mathrm{pre}}$ . Furthermore, we measured the distribution shift with respect to the real FFHQ images, which we denoted as $\\Delta \\mathrm{FID}_{\\mathrm{real}}$ .", + "bbox": [ + 498, + 676, + 890, + 723 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Main Experiment", + "text_level": 1, + "bbox": [ + 500, + 734, + 671, + 751 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.1 Qualitative Results", + "text_level": 1, + "bbox": [ + 500, + 758, + 687, + 773 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conducted a comparative analysis of GUIDE against the baseline in the generative identity unlearning task. Initiating from the provided source image, we aimed to eliminate the identity within the pre-trained generator, as illustrated in Figure 6. We presented the resulting unlearned image, along with the target image optimized in our loss functions. Notably, GUIDE effectively erases identities whether synthetic, presented during pre-training, or unseen.", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9156", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b16d41950d4932b684867f56533b114c7ec8e83f39877a17d3183961a207a665.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsRandomIn-Domain (FFHQ)Out-of-Domain (CelebAHQ)
ID (↓)FIDpre (↓)ΔFIDreal (↓)ID (↓)FIDpre (↓)ΔFIDreal (↓)ID (↓)FIDpre (↓)ΔFIDreal (↓)
Baseline0.19 ± 0.0911.73 ± 2.747.46 ± 2.200.16 ± 0.079.00 ± 1.154.15 ± 1.180.12 ± 0.069.52 ± 1.534.75 ± 0.89
+ extrapolated wt0.12 ± 0.0614.28 ± 3.349.63 ± 2.530.05 ± 0.0612.78 ± 1.826.76 ± 1.410.02 ± 0.0513.02 ± 3.207.31 ± 1.98
+ Ladj0.14 ± 0.0719.65 ± 4.9013.94 ± 3.590.04 ± 0.0613.53 ± 2.087.35 ± 1.700.01 ± 0.0513.63 ± 3.527.83 ± 2.19
+ Lglobal (GUIDE)0.14 ± 0.0610.80 ± 2.706.64 ± 1.600.06 ± 0.068.00 ± 1.203.05 ± 0.810.03 ± 0.057.88 ± 1.963.34 ± 1.10
", + "bbox": [ + 78, + 89, + 893, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4852243969db08bbcc2f30262ac7108a70c1fb4675d35e4b502a4f324d6091cd.jpg", + "image_caption": [ + "Figure 8. Qualitative comparison between GUIDE and the baseline on the preservation of the generation quality of other identities. GUIDE generates images almost identical to those synthesized by $G_{s}$ , whereas the baseline often results in noticeable changes, e.g., beard shape, hairstyle change, hat." + ], + "image_footnote": [], + "bbox": [ + 86, + 234, + 467, + 444 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate the thoroughness of identity removal, we performed a multi-image test using identities from the CelebAHQ dataset. This test involved assessing the ID similarity not only for the unlearned image derived from the source image but also for other images sharing the same identity. As shown in Figure 7, GUIDE showed superior generalization for unseen images compared to the baseline. This improvement is attributed to the adjacency-aware unlearning, which facilitated the unlearning process not just for the given images but also for their neighborhood.", + "bbox": [ + 75, + 544, + 468, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Figure 8, we conducted an experiment to assess the effect of the unlearning process on other identities. While the baseline had a significant impact on other identities through the unlearning, GUIDE showed a relatively lesser effect. We attribute this to the global preservation loss, which constrained the distribution shift on other latent codes.", + "bbox": [ + 75, + 696, + 468, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.2 Quantitative Results", + "text_level": 1, + "bbox": [ + 76, + 800, + 274, + 814 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Table 1, we compared GUIDE to the baseline by gradually applying the components of GUIDE. By configuring $w_{t}$ through extrapolation, we achieved performance improvements in ID similarity across three scenarios. Notably, we observed a significant difference in ID similarities in the", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2e7ed798aac777dc786c0a1113337fecc0372c6e9b63b32bdefce7a6b043405c.jpg", + "table_caption": [ + "Table 1. Quantitative results of GUIDE and the baseline in the generative identity unlearning task, tested in a single-image setting using one image per identity. Starting from the baseline, we gradually introduced components of GUIDE." + ], + "table_footnote": [], + "table_body": "
MethodsID (↓)ID others (↓)FID pre (↓)ΔFID real (↓)
Baseline0.12 ± 0.060.28 ± 0.089.52 ± 1.534.75 ± 0.89
+ extrapolated wt0.02 ± 0.050.15 ± 0.0713.02 ± 3.207.31 ± 1.98
+ Ladj0.01 ± 0.050.14 ± 0.0713.63 ± 3.527.83 ± 2.19
+ Lglobal (GUIDE)0.03 ± 0.050.17 ± 0.087.88 ± 1.963.34 ± 1.10
", + "bbox": [ + 500, + 236, + 893, + 303 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Quantitative results of GUIDE and the baseline in the generative identity unlearning in a multi-image setting, i.e., using a single image for unlearning and the other images for testing. We used CelebAHQ dataset for this test.", + "bbox": [ + 498, + 313, + 892, + 369 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "random scenario, indicating that in cases where a latent code was close to $\\bar{w}$ , i.e., as in the random scenario, there was insufficient removal of identity. The effectiveness of employing an extrapolation between the source latent code and the average latent code was evident in such instances.", + "bbox": [ + 496, + 382, + 890, + 458 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The adjacency-aware unlearning loss further enhanced the unlearning an identity. This loss was designed to cover the vicinity of the source latent code, thereby promoting unlearning on the source latent code itself. Finally, the application of the global preservation loss effectively reduced the estimated distribution shift using $\\mathrm{FID}_{\\mathrm{pre}}$ and $\\Delta \\mathrm{FID}_{\\mathrm{real}}$ .", + "bbox": [ + 496, + 458, + 890, + 549 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Moreover, we conducted a multi-image test in an OOD scenario. In this particular experiment, we introduced additional metric - $\\mathrm{ID}_{\\mathrm{others}}$ aimed at quantifying ID similarities for the unseen images associated with the source identity. As presented in Table 2, the introduction of the adjacency-aware unlearning loss resulted in a remarkable improvement in $\\mathrm{ID}_{\\mathrm{others}}$ , emphasizing the effectiveness of this unlearning approach for handling unseen images.", + "bbox": [ + 496, + 549, + 890, + 671 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 680, + 653, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of $d$ in Determination of $w_{t}$ . We conducted an ablation study by comparing target images derived from varied values of $d$ . Setting $d$ to 0 denotes utilizing the $\\bar{w}$ as $w_{t}$ in the unlearning process. For $d < 0$ , we designated $w_{t}$ as an interpolated latent code between the $w_{s}$ and the $\\bar{w}$ . Conversely, for $d > 0$ , we employed an extrapolated $w_{t}$ , as detailed in Section 3.2. As illustrated in Figure 9, when $d < 0$ , the target image closely aligns with the given source images. However, as $d$ deviates from 0, the quality of the target image rapidly deteriorates, resulting in a pronounced collapse in the distribution of the pre-trained generator. Consequently, the effectiveness of unlearning with such target images diminishes in removing identity from the", + "bbox": [ + 496, + 703, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9157", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/443c05e2d1b560cf48b1fc0c57d9e8313479e8f75c361064dd31632491e29ac0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SourceTarget
d = -10d = -5d = 0d = 30 (Ours)d = 60
Random
FFHQ
d = -70d = -35d = 0d = 30 (Ours)d = 60
", + "bbox": [ + 96, + 89, + 455, + 239 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d2622d79190a4fc8d2a3529e9f6ab875accf9065c3cf0c1f40cf5a695a7bebf8.jpg", + "table_caption": [ + "Figure 9. Ablation study to figure out the effectiveness of $d$ . We visualized target images corresponding to each source image with different values of $d$ . The target images were generated using target latent codes derived from interpolated latent codes, the average latent code ( $d = 0$ ), or extrapolated latent codes ( $d > 0$ ). Interpolation and extrapolation were carried out between the source and the average latent code. In the case of interpolation, the center between the source and the average latent code was computed." + ], + "table_footnote": [], + "table_body": "
αmaxID (↓)IDothers (↓)
00.1205 ± 0.06030.2754 ± 0.0791
100.0892 ± 0.06200.2123 ± 0.0762
150.0878 ± 0.03750.2094 ± 0.0692
200.0900 ± 0.05380.2105 ± 0.0924
300.0926 ± 0.05610.2111 ± 0.0653
", + "bbox": [ + 155, + 377, + 392, + 463 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "source images. Setting $d$ to 0 might suggest the use of $\\bar{w}$ as an effective target for erasing identity. However, our ablation studies indicate that when the source image closely aligns with $\\bar{w}$ , the unlearning procedure fails to thoroughly eliminate the identity. Conversely, when $d > 0$ , $w_{t}$ contains a distinct identity compared to the source image while maintaining a consistent distance from $\\bar{w}$ . Among the instances where $d > 0$ , our ablation studies reveal that setting $d = 30$ achieves a balanced performance between effective unlearning and preservation of the generation performance of the pre-trained model.", + "bbox": [ + 75, + 559, + 468, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of $\\alpha_{max}$ in $\\mathcal{L}_{adj}$ . In Table 3, we scrutinized the effectiveness of the adjacency-aware unlearning loss. To ensure a fair comparison, we employed $\\bar{w}$ as $w_{t}$ in this experiment, and we used $\\mathcal{L}_{local}$ and $\\mathcal{L}_{adj}$ in the unlearning procedure. Rows corresponding to $\\alpha_{max} = 0$ represent experimental results without the incorporation of $\\mathcal{L}_{adj}$ in the unlearning procedure. The introduction of $\\mathcal{L}_{adj}$ resulted in consistent performance gains in $\\mathrm{ID}_{\\mathrm{others}}$ . This observation highlights the efficacy of considering not only the pair of source and target latent codes but also their surroundings for unlearning the entire identity.", + "bbox": [ + 75, + 734, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9eb6ee6e7f81fb311d3bc4270918e89721b6396518ed1f010228c7f13200510e.jpg", + "table_caption": [ + "Table 3. Ablation study to figure out the effectiveness of $\\mathcal{L}_{adj}$ and $\\alpha_{max}$ . We compared the performance based on how successfully the given identity was erased, using ID and IDothers metric. The row where $\\alpha_{max} = 0$ denotes the baseline. We used CelebAHQ dataset in this experiment." + ], + "table_footnote": [], + "table_body": "
\\( \\mathcal{L}_{local} \\)\\( \\mathcal{L}_{global} \\)\\( \\mathrm{FID}_{\\text{pre}}(\\downarrow) \\)\\( \\Delta \\mathrm{FID}_{\\text{real}}(\\downarrow) \\)
✓9.52 ± 1.534.75 ± 0.89
✓✓4.63 ± 0.431.48 ± 0.29
", + "bbox": [ + 558, + 89, + 834, + 138 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablation study to figure our the effectiveness of $\\mathcal{L}_{global}$ . We compared how preserved the performance of the pre-trained model through the unlearning process, via $\\mathrm{FID}_{\\mathrm{pre}}$ and $\\Delta \\mathrm{FID}_{\\mathrm{real}}$ . We used CelebAHQ dataset in this experiment.", + "bbox": [ + 498, + 148, + 890, + 204 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of $\\mathcal{L}_{global}$ . To assess the effectiveness of the global preservation loss, a similar experiment was conducted as the previous experiment, i.e., setting $\\bar{w}$ as $w_{t}$ . The results are presented in Table 4. The application of $\\mathcal{L}_{global}$ demonstrated consistent performance improvements in both $\\mathrm{FID}_{\\mathrm{pre}}$ and $\\Delta \\mathrm{FID}_{\\mathrm{real}}$ . This suggests that imposing constraints on the generator to maintain its generation performance in latent codes distant from our primary focus is effective in reducing distribution shifts in generative models.", + "bbox": [ + 496, + 231, + 890, + 367 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 382, + 617, + 397 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduced a novel task, referred to as generative identity unlearning, designed to address privacy concerns in pre-trained generative adversarial networks. This task requires thoroughly removing the identity of a single source image from the pre-trained generator. To achieve this, we proposed a new framework, GUIDE (Generative Unlearning for any IDENTITY). To unlearn the single identity, we first defined the target latent code via extrapolation, moving away from the average latent by the pre-defined distance in the direction from the source to the average latent. Using this, our GUIDE successfully unlearned the given identity via Latent Target Unlearning (LTU), which optimized the pre-trained model to preserve the overall generative ability but not to generate the same identity within the local space. Experimental results demonstrated the effectiveness of GUIDE with promising outcomes. We anticipate that our work will be widely applied in research or the industry field, providing users with a sense of freedom from privacy concerns through identity removal.", + "bbox": [ + 496, + 407, + 890, + 695 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 709, + 658, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by MSIT (Ministry of Science and ICT), Korea, under the ITRC (Information Technology Research Center) support program (IITP-2024-RS-2023-00258649) supervised by the IITP (Institute for Information & Communications Technology Planning & Evaluation), and in part by NRF-2023S1A5A2A21083590, and in part by the IITP grant funded by the Korea Government (MSIT) (Artificial Intelligence Innovation Hub) under Grant 2021-0-02068, and by the IITP grant funded by the Korea government (MSIT) (No.RS-2022-00155911, Artificial Intelligence Convergence Innovation Human Resources Development (Kyung Hee University)).", + "bbox": [ + 496, + 732, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9158", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 84, + 174, + 99 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yuval Alaluf, Or Patashnik, and Daniel Cohen-Or. Restyle: A residual-based stylegan encoder via iterative refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6711–6720, 2021. 4", + "[2] Thomas Baumhauer, Pascal Schottle, and Matthias Zeppelinzauer. Machine unlearning: Linear filtration for logit-based classifiers. Machine Learning, 111(9):3203-3226, 2022. 3", + "[3] Nicolas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikash Sehwag, Florian Tramer, Borja Balle, Daphne Ippolito, and Eric Wallace. Extracting training data from diffusion models. In Proceedings of the 32nd USENIX Security Symposium (USENIX Security 23), pages 5253-5270, 2023. 2", + "[4] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 3, 4, 5", + "[5] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1", + "[6] Vikram S Chundawat, Ayush K Tarun, Murari Mandal, and Mohan Kankanhalli. Zero-shot machine unlearning. IEEE Transactions on Information Forensics and Security, 2023. 3", + "[7] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4690-4699, 2019. 5", + "[8] Matt Fredrikson, Somesh Jha, and Thomas Ristenpart. Model inversion attacks that exploit confidence information and basic countermeasures. In Proceedings of the 22nd ACM SIGSAC conference on computer and communications security, pages 1322-1333, 2015. 3", + "[9] Rohit Gandikota, Joanna Materzyńska, Jaden Fiotto-Kaufman, and David Bau. Erasing concepts from diffusion models. In Proceedings of the 2023 IEEE International Conference on Computer Vision (ICCV), pages 2426-2436, 2023. 2", + "[10] Antonio Ginart, Melody Guan, Gregory Valiant, and James Y Zou. Making ai forget you: Data deletion in machine learning. Advances in Neural Information Processing Systems, 32, 2019. 3", + "[11] Aditya Golatkar, Alessandro Achille, and Stefano Soatto. Eternal sunshine of the spotless net: Selective forgetting in deep networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9304–9312, 2020. 2", + "[12] Aditya Golatkar, Alessandro Achille, and Stefano Soatto. Forgetting outside the box: Scrubbing deep networks of information accessible from input-output observations. In Pro" + ], + "bbox": [ + 78, + 109, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ceedings of the European Conference on Computer Vision (ECCV), pages 383-398. Springer, 2020. 2", + "[13] Aditya Golatkar, Alessandro Achille, Avinash Ravichandran, Marzia Polito, and Stefano Soatto. Mixed-privacy forgetting in deep networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 792–801, 2021.", + "[14] Varun Gupta, Christopher Jung, Seth Neel, Aaron Roth, Saeed Sharifi-Malvajerdi, and Chris Waites. Adaptive machine unlearning. Advances in Neural Information Processing Systems, 34:16319-16330, 2021. 3", + "[15] Jamie Hayes, Luca Melis, George Danezis, and Emiliano De Cristofaro. Logan: Membership inference attacks against generative models. arXiv preprint arXiv:1705.07663, 2017. 1", + "[16] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in Neural Information Processing Systems, 30, 2017. 6", + "[17] Yuge Huang, Yuhan Wang, Ying Tai, Xiaoming Liu, Pengcheng Shen, Shaoxin Li, Jilin Li, and Feiyue Huang. Curricularface: adaptive curriculum learning loss for deep face recognition. In proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5901-5910, 2020. 6", + "[18] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability and variation. In Proceedings of the International Conference on Learning Representations (ICLR), 2018. 2, 6", + "[19] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 3, 5, 6", + "[20] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2", + "[21] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021. 1, 2", + "[22] Mahdi Khosravy, Kazuaki Nakamura, Yuki Hirose, Naoko Nitta, and Noboru Babaguchi. Model inversion attack by integration of deep generative models: Privacy-sensitive face generation from a face recognition system. IEEE Transactions on Information Forensics and Security, 17:357-372, 2022. 3", + "[23] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proceedings of the International Conference on Learning Representations (ICLR), 2015. 6", + "[24] Zhifeng Kong and Kamalika Chaudhuri. Data redaction from pre-trained gans. In Proceedings of the 2023 IEEE Conference on Secure and Trustworthy Machine Learning (SaTML), pages 638-677. IEEE, 2023. 3" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9159", + "bbox": [ + 480, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Nupur Kumari, Bingliang Zhang, Sheng-Yu Wang, Eli Shechtman, Richard Zhang, and Jun-Yan Zhu. Ablating concepts in text-to-image diffusion models. In Proceedings of the 2023 IEEE International Conference on Computer Vision (ICCV), pages 22691-22702, 2023. 2", + "[26] Ronak Mehta, Sourav Pal, Vikas Singh, and Sathya N Ravi. Deep unlearning via randomized conditionally independent hessian. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10422-10431, 2022. 2", + "[27] Jun-Yeong Moon, Keon-Hee Park, Jung Uk Kim, and Gyeong-Moon Park. Online class incremental learning on stochastic blurry task boundary via mask and visual prompt tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11731-11741, 2023.", + "[28] Saemi Moon, Seunghyuk Cho, and Dongwoo Kim. Feature unlearning for generative models via implicit feedback. arXiv preprint arXiv:2303.05699, 2023. 2, 3", + "[29] Seung-Jun Moon and Gyeong-Moon Park. Interstyle: Encoding an interest region for robust stylegan inversion. In Proceedings of the European Conference on Computer Vision (ECCV), pages 460-476. Springer, 2022. 2, 4", + "[30] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2085–2094, 2021. 2", + "[31] Alexandra Peste, Dan Alistarh, and Christoph H Lampert. Ssse: Efficiently erasing samples from trained machine learning models. corr. arXiv preprint arXiv:2107.03860, 1 (3), 2021. 2", + "[32] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In Proceedings of the International Conference on Machine Learning (ICML), pages 8821-8831. PMLR, 2021. 2", + "[33] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2", + "[34] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2287–2296, 2021. 2", + "[35] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on graphics (TOG), 42(1):1-13, 2022. 2", + "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 2", + "[37] Juwon Seo, Ji-Su Kang, and Gyeong-Moon Park. Lfs-gan: Lifelong few-shot image generation. In Proceedings of the" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE/CVF International Conference on Computer Vision (ICCV), pages 11356-11366, 2023. 2", + "[38] Shawn Shan, Jenna Cryan, Emily Wenger, Haitao Zheng, Rana Hanocka, and Ben Y Zhao. Glaze: Protecting artists from style mimicry by text-to-image models. arXiv preprint arXiv:2302.04222, 2023. 2", + "[39] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9243-9252, 2020. 2", + "[40] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. Interfacegan: Interpreting the disentangled face representation learned by gans. IEEE transactions on pattern analysis and machine intelligence, 44(4):2004-2018, 2020. 2", + "[41] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Diffusion art or digital forgery? investigating data replication in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6048-6058, 2023. 2", + "[42] Jingxiang Sun, Xuan Wang, Lizhen Wang, Xiaoyu Li, Yong Zhang, Hongwen Zhang, and Yebin Liu. Next3d: Generative neural texture rasterization for 3d-aware head avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20991-21002, 2023. 2", + "[43] Ayush K Tarun, Vikram S Chandawat, Murari Mandal, and Mohan Kankanhalli. Fast yet effective machine unlearning. IEEE Transactions on Neural Networks and Learning Systems, 2023. 3", + "[44] Piyush Tiwary, Atri Guha, Subhodip Panda, et al. Adapt then unlearn: Exploiting parameter space semantics for unlearning in generative adversarial networks. arXiv preprint arXiv:2309.14054, 2023. 2, 3", + "[45] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40(4): 1-14, 2021. 2", + "[46] Tengfei Wang, Yong Zhang, Yanbo Fan, Jue Wang, and Qifeng Chen. High-fidelity gan inversion for image attribute editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11379-11388, 2022. 2", + "[47] Yue Wu, Yu Deng, Jiaolong Yang, Fangyun Wei, Qifeng Chen, and Xin Tong. Anifacegan: Animatable 3d-aware face image generation for video avatars. Advances in Neural Information Processing Systems, 35:36188-36201, 2022. 2", + "[48] Yuting Xu, Jian Liang, Gengyun Jia, Ziming Yang, Yanhao Zhang, and Ran He. Tall: Thumbnail layout for deepfake video detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22658-22668, 2023. 1", + "[49] Zhiyuan Yan, Yong Zhang, Yanbo Fan, and Baoyuan Wu. Ucf: Uncovering common features for generalizable deepfake detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22412-22423, 2023. 1" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9160", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Jaejun Yoo, Youngjung Uh, Sanghyuk Chun, Byeongkyu Kang, and Jung-Woo Ha. Photorealistic style transfer via wavelet transforms. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9036-9045, 2019. 2", + "[51] Xiaoyong Yuan, Pan He, Qile Zhu, and Xiaolin Li. Adversarial examples: Attacks and defenses for deep learning. IEEE transactions on neural networks and learning systems, 30 (9):2805-2824, 2019. 3", + "[52] Ziyang Yuan, Yiming Zhu, Yu Li, Hongyu Liu, and Chun Yuan. Make encoder great again in 3d gan inversion through geometry and occlusion-aware encoding. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2437-2447, 2023. 2, 3, 4, 5", + "[53] Eric Zhang, Kai Wang, Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Forget-me-not: Learning to forget in text-to-image diffusion models. arXiv preprint arXiv:2303.17591, 2023. 2", + "[54] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 5", + "[55] Xuanmeng Zhang, Zhedong Zheng, Daiheng Gao, Bang Zhang, Yi Yang, and Tat-Seng Chua. Multi-view consistent generative adversarial networks for compositional 3d-aware image synthesis. International Journal of Computer Vision, pages 1-24, 2023. 2", + "[56] Xiaoming Zhao, Fangchang Ma, David Guera, Zhile Ren, Alexander G Schwing, and Alex Colburn. Generative multiplane images: Making a 2d gan 3d-aware. In Proceedings of the European Conference on Computer Vision (ECCV), pages 18-35. Springer, 2022. 2" + ], + "bbox": [ + 78, + 90, + 470, + 556 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "9161", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Generative Unlearning for Any Identity/c3c2d3ee-c7c2-4590-9220-bed4fbf9b764_model.json b/2024/Generative Unlearning for Any Identity/c3c2d3ee-c7c2-4590-9220-bed4fbf9b764_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7d324cbb879fa50fb0d31c488e4908631b4387f4 --- /dev/null +++ b/2024/Generative Unlearning for Any Identity/c3c2d3ee-c7c2-4590-9220-bed4fbf9b764_model.json @@ -0,0 +1,2191 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.284, + 0.131, + 0.688, + 0.154 + ], + "angle": 0, + "content": "Generative Unlearning for Any Identity" + }, + { + "type": "text", + "bbox": [ + 0.162, + 0.18, + 0.266, + 0.198 + ], + "angle": 0, + "content": "Juwon \\(\\mathrm{Seo}^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.307, + 0.18, + 0.63, + 0.218 + ], + "angle": 0, + "content": "Sung-Hoon Lee\\(^{1*}\\) Tae-Young Lee\\(^{1*}\\) Gyeong-Moon Park\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.67, + 0.181, + 0.808, + 0.2 + ], + "angle": 0, + "content": "Seungjun Moon²" + }, + { + "type": "text", + "bbox": [ + 0.283, + 0.233, + 0.69, + 0.271 + ], + "angle": 0, + "content": "1Kyung Hee University, Yongin, Republic of Korea \n2KLleon Tech., Seoul, Republic of Korea" + }, + { + "type": "text", + "bbox": [ + 0.255, + 0.279, + 0.714, + 0.313 + ], + "angle": 0, + "content": "{jwseo001, sunghoonlee961, slcks1, gmpark}@khu.ac.kr seungjunmoon@klleon.io" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.347, + 0.314, + 0.364 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.376, + 0.474, + 0.784 + ], + "angle": 0, + "content": "Recent advances in generative models trained on large-scale datasets have made it possible to synthesize high-quality samples across various domains. Moreover, the emergence of strong inversion networks enables not only a reconstruction of real-world images but also the modification of attributes through various editing methods. However, in certain domains related to privacy issues, e.g., human faces, advanced generative models along with strong inversion methods can lead to potential misuses. In this paper, we propose an essential yet under-explored task called generative identity unlearning, which steers the model not to generate an image of specific identity. In the generative identity unlearning, we target the following objectives: (i) preventing the generation of images with a certain identity, and (ii) preserving the overall quality of the generative model. To satisfy these goals, we propose a novel framework, Generative Unlearning for Any IDENTITY (GUIDE), which prevents the reconstruction of a specific identity by unlearning the generator with only a single image. GUIDE consists of two parts: (i) finding a target point for optimization that un-identifies the source latent code and (ii) novel loss functions that facilitate the unlearning procedure while less affecting the learned distribution. Our extensive experiments demonstrate that our proposed method achieves state-of-the-art performance in the generative machine unlearning task. The code is available at https://github.com/KHU-AGI/GUIDE." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.811, + 0.21, + 0.827 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.836, + 0.471, + 0.868 + ], + "angle": 0, + "content": "Recently, 2D or 3D Generative Adversarial Networks (GANs) [4, 19-21] pre-trained on large datasets, e.g., FFHQ" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.349, + 0.875, + 0.567 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.577, + 0.895, + 0.648 + ], + "angle": 0, + "content": "Figure 1. Given a single source image containing a specific identity, we remove that identity from the pre-trained 3D generative adversarial network (e.g., EG3D [4]). Our method effectively unlearns identity even from in-the-wild images where the source image is absent in the pre-training dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.893, + 0.81 + ], + "angle": 0, + "content": "[19] or AFHQ [5], have drawn substantial attention due to their remarkable generation performance and highly disentangled representation space. However, their advancements have raised privacy concerns [15], especially regarding the potential misuse of generative models to represent and exploit individual identities. For instance, deepfakes [48, 49] can create very believable images or videos of people in made-up situations, causing major concerns about ethics and privacy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.895, + 0.902 + ], + "angle": 0, + "content": "To alleviate privacy issues in generative models, machine unlearning task has been actively studied. Machine unlearning involves the process of selectively removing specific knowledge or erasing the influence of certain data from the training dataset of pre-trained models. It is beneficial especially when the data are harmful, private, or biased" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.204, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.888, + 0.221, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.221, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "9151" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.168 + ], + "angle": 0, + "content": "[11, 12, 26, 31]. Despite a focus on discriminative tasks in most machine unlearning research, a few studies have ventured into generative models, attempting to erase high-level concepts such as socially inappropriate content or artistic styles that present copyright challenges [9, 25]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.17, + 0.468, + 0.306 + ], + "angle": 0, + "content": "Nevertheless, generative models still exhibit ongoing privacy issues. Even if an identity of someone is not used in the pre-training of the generative models, it can be easily reconstructed in the pre-trained models via GAN inversion models [29, 34, 35, 45, 46, 52]. Furthermore, the reconstructed image can be manipulated or edited easily via image editing methods [30, 39, 40]. To prevent potential exploits of an identity, it is necessary to erase a certain identity from the pre-trained generative models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.308, + 0.468, + 0.459 + ], + "angle": 0, + "content": "To consider the above issue, we introduce an essential task of unlearning any identity from the pre-trained 2D or 3D GANs [4, 20], called generative identity unlearning. Unlike typical machine unlearning tasks, which focus on unlearning the training samples our generative identity unlearning task unlearns any identity on pre-trained GANs, even if it was not shown during the pre-training. Our goal is to remove the whole identity associated with a given single image from the generator while minimally impacting the overall performance of the pre-trained model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.461, + 0.468, + 0.643 + ], + "angle": 0, + "content": "To achieve our goal, we propose a novel generative unlearning framework, Generative Unlearning for Any IDEntity, named GUIDE. GUIDE replaces the source identity with an anonymous target identity, erasing the original identity effectively. To this end, we propose a new exploration method to determine an effective target latent code, called Un-Identifying Face On latent space (UFO). UFO utilizes the GAN inversion method [52] to embed the given identity into the source latent, and then decides the target latent using both the source and the average latent codes. We empirically find that the proposed UFO can identify the promising target to erase any given source identity robustly." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Given the source and target latent code, we update the generator to shift from the source identity to the target identity. To this end, we propose three novel loss functions: (i) local unlearning loss, (ii) adjacency-aware unlearning loss, and (iii) global preservation loss. (i) guides our model directly shifting the source identity to the target identity. (ii) utilizes other latent codes adjacent to the source and target latent codes to effectively unlearn the entire identity from a single image. To minimize side effects from the unlearning process, (iii) additionally regularizes the generator to retain generation performance for latent codes relatively far from the source and target latent codes. Through comprehensive experiments on diverse identities, including Random, InD, and OOD, we confirm that GUIDE can successfully remove the identity of the source image from the pre-trained generative model, and shows qualitatively and quantitatively superior performances." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.092, + 0.845, + 0.106 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.108, + 0.892, + 0.197 + ], + "angle": 0, + "content": "- For the first time, we propose a novel task, generative identity unlearning, which tackles machine unlearning in generative models in the aspect of privacy protection. In our task, we aim to prevent the pre-trained generative models from synthesizing the given identity by utilizing only a single image." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.198, + 0.892, + 0.303 + ], + "angle": 0, + "content": "- For the effective and robust elimination of the identity, we propose a novel method - Un-Identifying Face On Latent Space (UFO). We configure the unlearning procedure by formulating how to represent and shift the identity in the latent space. We find that setting the extrapolated latent code between the source and average latent codes as an optimization target facilitates the unlearning procedure." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.304, + 0.892, + 0.378 + ], + "angle": 0, + "content": "- We propose three loss functions - local unlearning loss, adjacency-aware unlearning loss, and global preservation loss to effectively unlearn the identity from the pretrained model while less affecting the generation performance on other identity." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.379, + 0.892, + 0.469 + ], + "angle": 0, + "content": "- We show that our proposed framework, GUIDE achieves state-of-the-art performance both qualitatively and quantitatively, through extensive experiments. We demonstrate that GUIDE can remove the specific identity successfully in the generative models while minimizing the negative effect on other identities." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.108, + 0.892, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.483, + 0.64, + 0.499 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.509, + 0.892, + 0.644 + ], + "angle": 0, + "content": "Generative Models and Privacy Issue. In image synthesis field, GAN-based generative models have achieved remarkable performance not only in 2D [18-21, 32, 33, 36, 37] but also in 3D [4, 42, 47, 55, 56] domain. The application of various image editing methods [30, 40, 50] to strong generative models, people can easily generate edited images of specific individuals and various artistic styles [38, 41], as well as extract copyrighted content [3] without the permission of the individual or the original creator." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Recently, with the rise of the importance of AI ethics, several works have addressed this issue [9, 25, 28, 44, 53]. ESD [9] erases specific visual concepts from diffusion model by using negative guidance about the undesired concepts. Kumari et al. [25] modifies the conditional distribution of the model a specific target concepts to match the anchor concept. Forget-Me-Not [53] fine-tunes U-Net to minimize each of the intermediate attention associated with the target concepts to remove. Additional works in GANs [28, 44] focus on unlearning specific features, e.g., \"Bang\", \"Hat\" or \"Beard\" rather than forgetting specific identity. The above methods primarily concentrate on the elimination of specific concepts or high-level features. In other words, these cannot preclude models from generating specific individuals while maintaining the generation performance of realistic human faces. Unlike the existing works, our work targets to unlearn only specific individuals without" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9152" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.423, + 0.108 + ], + "angle": 0, + "content": "shifting the overall distribution of generated images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.127, + 0.468, + 0.231 + ], + "angle": 0, + "content": "Machine Unlearning. Machine unlearning aims to selectively forget specific acquired knowledge or diminish the impact of certain training data subsets on a trained model. Since previous research [8, 22, 51] shows that machine learning models might accidentally share private information when faced with certain attacks or inputs, machine unlearning becomes crucial." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.232, + 0.47, + 0.491 + ], + "angle": 0, + "content": "While previous machine unlearning is mainly focused on supervised learning tasks [2, 6, 10-14, 43], the interest in unlearning techniques within unsupervised learning, i.e., generative models, is growing [24, 28, 44]. However, most of the existing methods need full dataset access for retraining, which is hard to acquire and computationally expensive [2, 10, 14, 43]. For example, Kong and Chaudhuri [24] utilizes data reduction and augmentation algorithms, which requires a full training dataset. Despite the existence of a feature unlearning model [28] which does not need full dataset access, unlearning only an individual feature is not enough to forget a whole specific identity. To this end, we propose an algorithm that enables forgetting the specific identity only with a single image. Furthermore, our approach distinguishes itself from existing research by applying unlearning to unseen images, enabling the erasure of specific identities without prior exposure to those images." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.504, + 0.169, + 0.519 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.525, + 0.469, + 0.648 + ], + "angle": 0, + "content": "Firstly, in Section 3.1, as shown in Figure 2, we introduce the problem we aim to address, named generative identity unlearning. In Section 3.2, we introduce un-identifying face on latent space, which designates an appropriate target latent for unlearning. Then, in Section 3.3, we introduce latent target unlearning, along with our proposed novel losses to unlearn the generator. The total overview of our method is illustrated in Figure 3." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.657, + 0.281, + 0.672 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.681, + 0.469, + 0.772 + ], + "angle": 0, + "content": "Given a set of images \\(\\mathbf{x}\\) depicting a specific identity, we randomly select a single source image \\(x_{u} \\in \\mathbf{x}\\) as an exemplar of the identity. Initially, using off-the-shelf inversion network [52] \\(E\\) corresponding to the unconditional generator, i.e., EG3D [4], we embed \\(x_{u}\\) to the source latent code \\(w_{u}\\) in the latent space of EG3D:" + }, + { + "type": "equation", + "bbox": [ + 0.225, + 0.782, + 0.469, + 0.798 + ], + "angle": 0, + "content": "\\[\nw _ {u} = E \\left(x _ {u}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.469, + 0.871 + ], + "angle": 0, + "content": "Since EG3D [4] consists of the mapping network \\(Map(\\cdot)\\), StyleGAN2 [19] backbone \\(G(\\cdot)\\) and the neural renderer with a super-resolution module \\(R(\\cdot)\\), we can denote the reconstructed image \\(\\hat{x}\\) from \\(w_{u}\\) as following:" + }, + { + "type": "equation", + "bbox": [ + 0.21, + 0.881, + 0.469, + 0.897 + ], + "angle": 0, + "content": "\\[\n\\hat {x} = R \\left(G \\left(w _ {u}\\right); c\\right), \\tag {2}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.091, + 0.88, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.256, + 0.892, + 0.355 + ], + "angle": 0, + "content": "Figure 2. An illustration of generative identity unlearning. Upon GUIDE, the identity of the image generated from \\( w_{u} \\), i.e., inversion of the source image \\( x_{u} \\) by inversion network \\( E \\), should exhibit a distinct identity when passed through the pre-trained generator \\( G_{s} \\) compared to the unlearned generator \\( G_{u} \\). Furthermore, other images \\( x_{o} \\), not used in unlearning but sharing the same identity with \\( x_{u} \\), also should vary an identity through GUIDE." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.382, + 0.892, + 0.503 + ], + "angle": 0, + "content": "where \\( c \\) denotes camera poses. For convenience, we omit the explicit notation of camera poses in this paper, i.e., \\( \\hat{x} = (R \\circ G)(w_u) \\). We target to derive an unlearned \\( G \\), i.e., \\( G_u \\), from the pre-trained EG3D generator \\( G \\), i.e., \\( G_s \\), while fixing \\( Map \\) and \\( R \\). With proper unlearning, an image generated by unlearned EG3D using \\( w_u \\), i.e., \\( \\hat{x}_u = (R \\circ G_u)(w_u) \\) should have a distinct identity from the image generated by original EG3D using \\( w_u \\), i.e., \\( (R \\circ G_s)(w_u) \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.504, + 0.893, + 0.701 + ], + "angle": 0, + "content": "In our task formulation, two considerations are paramount. First, we aim to eliminate the entire identity from the generator only utilizing a single image. To validate this, we incorporate other multiple images \\( x_{o} \\in \\mathbf{x} \\) for testing and its corresponding latent code \\( w_{o} = E(x_{o}) \\). By utilizing \\( x_{o} \\), we can verify whether \\( G_{u} \\) has successfully unlearned the identity as a whole, rather than just unlearning the specific image \\( x_{u} \\). Second, we strive to maintain the generation performance of the pre-trained model. To assess this, we sample multiple images from fixed latent codes using both the unlearned and pre-trained generators. We then estimate the distribution shift between the images generated before and after the unlearning process." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.711, + 0.824, + 0.727 + ], + "angle": 0, + "content": "3.2. Un-Identifying Face On Latent Space" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The successfully unlearned model should not generate the image with the identity of \\(\\mathbf{x}\\), even when \\(w_{u}\\) is used as a latent. Consequently, we initiate our approach to manipulate \\(\\hat{x}_u\\) to be another image rather than the image with identity in \\(\\mathbf{x}\\) by unlearning \\(G\\). To design the objective function for unlearning, we first need to establish the objective for unlearning, which involves defining the target image \\(\\hat{x}_t\\), derived from \\(w_t\\), which the unlearned image \\(\\hat{x}_u\\), derived from \\(w_{u}\\), should mimic after unlearning. While there exist various options for setting \\(\\hat{x}_t\\), e.g., a random face or even a non-human image, we choose the mean face generated by" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9153" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.093, + 0.895, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.328, + 0.895, + 0.442 + ], + "angle": 0, + "content": "Figure 3. An overview of GUIDE. Starting with a source image, we employ a GAN inversion network \\( E \\), specifically GOAE [52], to embed this image into the latent space of a pre-trained generative model, namely EG3D [4], obtaining the source latent code \\( w_{u} \\). The target latent code \\( w_{t} \\) is designated through the UFO process. To facilitate identity removal in \\( w_{u} \\), we shift its identity to match that of \\( w_{t} \\) with our Latent Target Unlearning (LTU) process. Three loss functions of LTU are designed for this purpose: (i) The generator is optimized to produce an image from the source latent code, denoted as \\( (R \\circ G_{u})(w_{u}) \\), that is similar to the image from the target latent code, represented as \\( (R \\circ G_{s})(w_{t}) \\). (ii) To achieve unlearning across the entire identity, we consider latent codes near both the source and target latent codes, denoted as \\( w_{u,a} \\) and \\( w_{t,a} \\), respectively. (iii) To prevent model corruption during the unlearning process, we additionally sample latent codes from a random noise vector, represented as \\( w_{r,g} \\), and optimize \\( G_{u} \\) to preserve its generation ability on \\( w_{r,g} \\)." + }, + { + "type": "image", + "bbox": [ + 0.12, + 0.468, + 0.429, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.592, + 0.471, + 0.661 + ], + "angle": 0, + "content": "Figure 4. An illustration of Un-identifying Face On Latent Space (UFO). We define the identity of the source latent code by subtract it from the average latent code. We set the target latent code for our unlearning process by measuring an extrapolation between the source and average latent code with a fixed distance \\( d \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.47, + 0.764 + ], + "angle": 0, + "content": "the mean latent \\(\\bar{w}\\) of \\(Map\\), i.e., \\((R \\circ G_s)(\\bar{w})\\), as \\(\\hat{x}_t\\). This setting is inspired by inversion methods [1, 29] which apply the identity to the mean face through inversion stages. Since our goal is to de-identify the image, we argue that it is intuitive and reasonable to send the image back to the mean face, which is opposite to the inversion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "However, this simple approach might be problematic when \\( w_{u} \\) is close to \\( \\bar{w} \\), where the unlearned image might still resemble the original identity. To this end, we propose a novel method named Un-identifying Face On latent space (UFO), which can set \\( \\hat{x}_t \\) robustly regardless of the distance between \\( w_{u} \\) and \\( \\bar{w} \\), as shown in Figure 4. The following are the processes of UFO: First, in the \\( de \\)-identification process, we extract the identity latent \\( w_{id} = w_{u} - \\bar{w} \\) by subtracting \\( \\bar{w} \\) from \\( w_{u} \\). Then we propose a process of en-identification," + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.467, + 0.892, + 0.557 + ], + "angle": 0, + "content": "setting the target in the opposite direction of the existing \\( w_{id} \\) to foster a more pronounced change in identity, creating an entirely new identity. In other words, we chose \\( \\bar{w} \\) as the stopover point and established the final target point at the extrapolation of \\( w_{u} \\) and \\( \\bar{w} \\). The finalized target point can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.619, + 0.563, + 0.892, + 0.593 + ], + "angle": 0, + "content": "\\[\nw _ {t} = \\bar {w} - d \\cdot \\frac {w _ {i d}}{\\| w _ {i d} \\| _ {2}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.604, + 0.892, + 0.71 + ], + "angle": 0, + "content": "where \\(d\\) is defined as the distance that balances un-identification with preservation of the source distribution, as determined by our empirical results. Finally, we can set \\(\\hat{x}_t = (R\\circ G_s)(w_t)\\). We empirically demonstrate that whether the given identity is in close proximity to the average latent code within the latent space or not, UFO sets the desirable target latent code for the generative unlearning." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.72, + 0.734, + 0.737 + ], + "angle": 0, + "content": "3.3. Latent Target Unlearning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.744, + 0.892, + 0.851 + ], + "angle": 0, + "content": "After setting \\(\\hat{x}_t\\), we need to robustly design the unlearning process that can effectively make \\(\\hat{x}_u\\) to be similar to \\(\\hat{x}_t\\), while maintaining the generation performance of \\(G_u\\). We coin this process as Latent Target Unlearning (LTU), which targets to unlearn images from the specific latent while keeping generation results from other latent codes. LTU utilizes the following three losses to achieve this goal:" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Local Unlearning Loss. To force \\(\\hat{x}_u\\) to be \\(\\hat{x}_t\\), we use the widely-used reconstruction losses, i.e., Euclidean loss \\(\\mathcal{L}_2\\)," + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9154" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.09, + 0.473, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.242, + 0.473, + 0.342 + ], + "angle": 0, + "content": "Figure 5. An illustration of determining latent codes near a latent code \\( w \\) in adjacency-aware unlearning loss. We first sample a latent code \\( w_{r,a} \\) which is derived from a random noise vector \\( z_{r,a} \\) via the mapping network \\( \\text{Map}(\\cdot) \\), i.e. \\( w_{r,a} = \\text{Map}(z_{r,a}) \\). Next, we compute the direction between \\( w \\) and \\( w_{r,a} \\), and we scale it to fall within range between 0 and \\( \\alpha_{max} \\). This yields the distance vector \\( \\Delta \\) to compute the adjacent latent code \\( w_{a} = w + \\Delta \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.362, + 0.47, + 0.424 + ], + "angle": 0, + "content": "perceptual loss \\(\\mathcal{L}_{per}\\) [54], and identity loss \\(\\mathcal{L}_{id}\\) [7]. Using \\(\\mathcal{L}_{recon}\\), we compare the tri-plane features \\(F_{u} = G_{u}(w_{u})\\) and \\(F_{t} = G_{s}(w_{t})\\), derived from source and target latent codes, respectively. The local unlearning loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.443, + 0.47, + 0.48 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {l o c a l}} \\left(\\hat {x} _ {u}, \\hat {x} _ {t}\\right) = \\lambda_ {L 2} \\mathcal {L} _ {L 2} \\left(F _ {u}, F _ {t}\\right) \\tag {4} \\\\ + \\lambda_ {p e r} \\mathcal {L} _ {p e r} (\\hat {x} _ {u}, \\hat {x} _ {t}) + \\lambda_ {i d} \\mathcal {L} _ {i d} (\\hat {x} _ {u}, \\hat {x} _ {t}). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.496, + 0.47, + 0.527 + ], + "angle": 0, + "content": "By adopting \\(\\mathcal{L}_{local}\\), we can successfully un-identify the given source identity in \\(\\hat{x}_t\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.553, + 0.473, + 0.736 + ], + "angle": 0, + "content": "Adjacency-Aware Unlearning Loss. The above equation considers only one pair of source and target latent codes. However, images of a similar identity to the source identity can be obtained by introducing marginal perturbations to the latent code. For the successful unlearning of the given identity, we need to consider the neighborhood of both the source and the target latent codes. Consequently, as shown in Figure 5, we sample \\( N_{a} \\) latent codes in the vicinity of the \\( w_{u} \\). Specifically, with the scale \\( \\alpha^{i} \\) sampled from the uniform distribution with hyperparameter \\( \\alpha_{max} \\), i.e. \\( \\alpha^{i} \\sim \\mathcal{U}(0, \\alpha_{max}) \\), we define the distances \\( \\Delta \\) to compute the adjacent latent codes as:" + }, + { + "type": "equation", + "bbox": [ + 0.169, + 0.752, + 0.47, + 0.79 + ], + "angle": 0, + "content": "\\[\n\\Delta = \\left\\{\\alpha^ {i} \\cdot \\frac {w _ {r , a} ^ {i} - w _ {u}}{\\| w _ {r , a} ^ {i} - w _ {u} \\| _ {2}} \\right\\} _ {i = 1} ^ {N _ {a}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.807, + 0.47, + 0.904 + ], + "angle": 0, + "content": "where \\( w_{r,a}^{i} \\) is a latent code sampled from the random noise vector \\( z_{r,a}^{i} \\). Using these distances \\( \\Delta \\), we can compute \\( N_{a} \\) latent codes for both the source and the target latent codes. Similar to the local unlearning loss, we optimize the generated tri-plane features and images from \\( w_{u,a}^{i} = w_{u} + \\Delta^{i} \\) and \\( w_{t,a}^{i} = w_{t} + \\Delta^{i} \\) to be similar:" + }, + { + "type": "equation", + "bbox": [ + 0.552, + 0.117, + 0.892, + 0.137 + ], + "angle": 0, + "content": "\\[\n\\hat {x} _ {u, a} ^ {i} = R \\left(F _ {u, a} ^ {i}\\right), \\hat {x} _ {t, a} ^ {i} = R \\left(F _ {u, a} ^ {i}\\right), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.553, + 0.14, + 0.892, + 0.181 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {a d j} \\left(w _ {u}, w _ {t}\\right) = \\frac {1}{N _ {a}} \\sum_ {i = 1} ^ {N _ {a}} \\mathcal {L} _ {\\text {l o c a l}} \\left(\\hat {x} _ {u, a} ^ {i}, \\hat {x} _ {t, a} ^ {i}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.192, + 0.892, + 0.239 + ], + "angle": 0, + "content": "where \\( F_{u,a}^{i} = G_{u}(w_{u,a}^{i}), F_{t,a}^{i} = G_{s}(w_{t,a}^{i}) \\) denotes for triplane features, and \\( \\mathcal{L}_{local} \\) in Equation 4. From \\( \\mathcal{L}_{adj} \\), we can further consider possible variations of the source identity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.256, + 0.892, + 0.377 + ], + "angle": 0, + "content": "Global Preservation Loss. While the local unlearning loss and adjacency-aware unlearning loss mentioned above facilitate the removal of the source identity, we propose a global preservation loss to mitigate side effects arising from these unlearning loss functions. In the global preservation loss, we constrain the generator to maintain generation performance for latent codes that are relatively distant from both the source and target latent codes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.377, + 0.892, + 0.49 + ], + "angle": 0, + "content": "To be precise, we sample \\(N_{g}\\) latent codes \\(\\{w_{r,g}^{i}\\}_{i = 1}^{N_{g}}\\) from random noise vectors \\(\\{z_{r,g}^{i}\\}_{i = 1}^{N_{g}}\\). We ensure that these do not overlap with the adjacent latent codes used in the adjacency-aware unlearning loss. Unlike the unlearning loss functions, we find that adopting only \\(L_{per}\\) achieves a balanced performance between identity shift and model preservation. The global preservation loss is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.624, + 0.501, + 0.789, + 0.521 + ], + "angle": 0, + "content": "\\[\n\\hat {x} _ {u, g} ^ {i} = (R \\circ G _ {u}) (w _ {r, g} ^ {i}),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.629, + 0.523, + 0.892, + 0.55 + ], + "angle": 0, + "content": "\\[\n\\hat {x} _ {s, g} ^ {i} = \\left(R \\circ G _ {s}\\right) \\left(w _ {r, g} ^ {i}\\right), \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.549, + 0.545, + 0.844, + 0.585 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {g l o b a l} (G _ {u}, G _ {s}) = \\frac {1}{N _ {g}} \\sum_ {i = 1} ^ {N _ {g}} \\mathcal {L} _ {p e r} (\\hat {x} _ {u, g} ^ {i}, \\hat {x} _ {s, g} ^ {i}).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.597, + 0.728, + 0.612 + ], + "angle": 0, + "content": "In summary, our final objective is:" + }, + { + "type": "equation", + "bbox": [ + 0.546, + 0.624, + 0.892, + 0.641 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {l o c a l}} + \\lambda_ {\\text {a d j}} \\mathcal {L} _ {\\text {a d j}} + \\lambda_ {\\text {g l o b a l}} \\mathcal {L} _ {\\text {g l o b a l}}. \\tag {9}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.653, + 0.634, + 0.67 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.678, + 0.691, + 0.695 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.701, + 0.892, + 0.793 + ], + "angle": 0, + "content": "**Baseline.** Since we propose generative identity unlearning task for the first time, to evaluate the effectiveness of GUIDE, we constructed a simple baseline. In the baseline, we used the target latent code as the average one for the unlearning. During the unlearning, we updated the pre-trained generator using \\(\\mathcal{L}_{local}\\) as described in Equation 4." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Implementation Details. We built GUIDE based on the 3D generative adversarial network [4] pre-trained on FFHQ dataset [19]. We used GOAE [52] as a GAN inversion network to obtain the latent code from an image. The image resolution we used in our experiments is \\(512 \\times 512\\) with a rendering resolution of \\(128 \\times 128\\). We used Adam optimizer" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9155" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.086, + 0.094, + 0.891, + 0.312 + ], + "angle": 0, + "content": "
SourceRandomIn-Domain (FFHQ)Out-of-Domain (CelebAHQ)
BaselineOursBaselineOursBaselineOursBaselineOursBaselineOurs
Unlearned
" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.323, + 0.893, + 0.366 + ], + "angle": 0, + "content": "Figure 6. Qualitative results of GUIDE and the baseline in generative identity unlearning task. For the given source image each (the first row), GUIDE and the baseline tried to erase the identity in the pre-trained generator. The images in the second and third row are the target and unlearned images, respectively." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.392, + 0.47, + 0.455 + ], + "angle": 0, + "content": "[23] with a learning rate of \\(10^{-4}\\) in the unlearning procedure. The hyperparameters used in the experiments were: \\(d = 30\\), \\(\\alpha_{max} = 15\\), \\(\\lambda_{L2} = 10^{-2}\\), \\(\\lambda_{per} = 1\\), \\(\\lambda_{id} = 10^{-1}\\), \\(N_a = N_g = 2\\), and \\(\\lambda_{adj} = \\lambda_{global} = 1\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.473, + 0.471, + 0.653 + ], + "angle": 0, + "content": "Dataset and Scenarios. We evaluated GUIDE in three scenarios: Random, where we set an unlearning target image from a randomly sampled noise vector; InD (domain), where we sampled an image from the FFHQ dataset [19] used for pre-training; and OOD (out-of-domain), where the unlearning target image was sampled from the CelebAHQ dataset [18]. For InD and OOD, we used the GAN inversion network to obtain corresponding latent codes. For OOD scenario, we also conducted multi-image test since there were multiple images with a same identity in CelebAHQ. On the other hand, we performed only single-image test in the Random and InD scenarios." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Evaluation Metrics. We evaluated GUIDE on two key aspects. Firstly, we estimated the efficacy of our approach in preventing the generator from producing images similar to the unlearning target. We quantitatively measured similarity of identities (ID) using face recognition network, CurricularFace [17], between images generated from the same latent codes before and after unlearning. Moreover, we utilized \\(\\mathrm{ID}_{\\mathrm{others}}\\) to estimate the erasure of a identity from images which are unseen during training but containing the same identity of the source image. Secondly, we assessed whether our method preserves overall generation performance using the Fréchet Inception distance (FID) score [16]. Different from the existing usages, we utilized two variants of FID. First, we evaluated the distribution shift of generated images the pre-trained generator and the un-" + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.391, + 0.865, + 0.598 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.608, + 0.892, + 0.664 + ], + "angle": 0, + "content": "Figure 7. Qualitative results of GUIDE and the baseline on a multi-image test using CelebAHQ dataset. We additionally utilized images that are unseen during unlearning, to show how thoroughly erase the given identity." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.678, + 0.892, + 0.724 + ], + "angle": 0, + "content": "learned generator via \\(\\mathrm{FID}_{\\mathrm{pre}}\\). Furthermore, we measured the distribution shift with respect to the real FFHQ images, which we denoted as \\(\\Delta \\mathrm{FID}_{\\mathrm{real}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.736, + 0.673, + 0.752 + ], + "angle": 0, + "content": "4.2. Main Experiment" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.76, + 0.688, + 0.774 + ], + "angle": 0, + "content": "4.2.1 Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We conducted a comparative analysis of GUIDE against the baseline in the generative identity unlearning task. Initiating from the provided source image, we aimed to eliminate the identity within the pre-trained generator, as illustrated in Figure 6. We presented the resulting unlearned image, along with the target image optimized in our loss functions. Notably, GUIDE effectively erases identities whether synthetic, presented during pre-training, or unseen." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9156" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.09, + 0.894, + 0.174 + ], + "angle": 0, + "content": "
MethodsRandomIn-Domain (FFHQ)Out-of-Domain (CelebAHQ)
ID (↓)FIDpre (↓)ΔFIDreal (↓)ID (↓)FIDpre (↓)ΔFIDreal (↓)ID (↓)FIDpre (↓)ΔFIDreal (↓)
Baseline0.19 ± 0.0911.73 ± 2.747.46 ± 2.200.16 ± 0.079.00 ± 1.154.15 ± 1.180.12 ± 0.069.52 ± 1.534.75 ± 0.89
+ extrapolated wt0.12 ± 0.0614.28 ± 3.349.63 ± 2.530.05 ± 0.0612.78 ± 1.826.76 ± 1.410.02 ± 0.0513.02 ± 3.207.31 ± 1.98
+ Ladj0.14 ± 0.0719.65 ± 4.9013.94 ± 3.590.04 ± 0.0613.53 ± 2.087.35 ± 1.700.01 ± 0.0513.63 ± 3.527.83 ± 2.19
+ Lglobal (GUIDE)0.14 ± 0.0610.80 ± 2.706.64 ± 1.600.06 ± 0.068.00 ± 1.203.05 ± 0.810.03 ± 0.057.88 ± 1.963.34 ± 1.10
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.185, + 0.893, + 0.214 + ], + "angle": 0, + "content": "Table 1. Quantitative results of GUIDE and the baseline in the generative identity unlearning task, tested in a single-image setting using one image per identity. Starting from the baseline, we gradually introduced components of GUIDE." + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.236, + 0.468, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.457, + 0.47, + 0.528 + ], + "angle": 0, + "content": "Figure 8. Qualitative comparison between GUIDE and the baseline on the preservation of the generation quality of other identities. GUIDE generates images almost identical to those synthesized by \\( G_{s} \\), whereas the baseline often results in noticeable changes, e.g., beard shape, hairstyle change, hat." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.545, + 0.469, + 0.697 + ], + "angle": 0, + "content": "To evaluate the thoroughness of identity removal, we performed a multi-image test using identities from the CelebAHQ dataset. This test involved assessing the ID similarity not only for the unlearned image derived from the source image but also for other images sharing the same identity. As shown in Figure 7, GUIDE showed superior generalization for unseen images compared to the baseline. This improvement is attributed to the adjacency-aware unlearning, which facilitated the unlearning process not just for the given images but also for their neighborhood." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.697, + 0.469, + 0.788 + ], + "angle": 0, + "content": "In Figure 8, we conducted an experiment to assess the effect of the unlearning process on other identities. While the baseline had a significant impact on other identities through the unlearning, GUIDE showed a relatively lesser effect. We attribute this to the global preservation loss, which constrained the distribution shift on other latent codes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.801, + 0.275, + 0.815 + ], + "angle": 0, + "content": "4.2.2 Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.901 + ], + "angle": 0, + "content": "In Table 1, we compared GUIDE to the baseline by gradually applying the components of GUIDE. By configuring \\( w_{t} \\) through extrapolation, we achieved performance improvements in ID similarity across three scenarios. Notably, we observed a significant difference in ID similarities in the" + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.237, + 0.894, + 0.304 + ], + "angle": 0, + "content": "
MethodsID (↓)ID others (↓)FID pre (↓)ΔFID real (↓)
Baseline0.12 ± 0.060.28 ± 0.089.52 ± 1.534.75 ± 0.89
+ extrapolated wt0.02 ± 0.050.15 ± 0.0713.02 ± 3.207.31 ± 1.98
+ Ladj0.01 ± 0.050.14 ± 0.0713.63 ± 3.527.83 ± 2.19
+ Lglobal (GUIDE)0.03 ± 0.050.17 ± 0.087.88 ± 1.963.34 ± 1.10
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.314, + 0.893, + 0.37 + ], + "angle": 0, + "content": "Table 2. Quantitative results of GUIDE and the baseline in the generative identity unlearning in a multi-image setting, i.e., using a single image for unlearning and the other images for testing. We used CelebAHQ dataset for this test." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.383, + 0.892, + 0.459 + ], + "angle": 0, + "content": "random scenario, indicating that in cases where a latent code was close to \\(\\bar{w}\\), i.e., as in the random scenario, there was insufficient removal of identity. The effectiveness of employing an extrapolation between the source latent code and the average latent code was evident in such instances." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.459, + 0.892, + 0.55 + ], + "angle": 0, + "content": "The adjacency-aware unlearning loss further enhanced the unlearning an identity. This loss was designed to cover the vicinity of the source latent code, thereby promoting unlearning on the source latent code itself. Finally, the application of the global preservation loss effectively reduced the estimated distribution shift using \\(\\mathrm{FID}_{\\mathrm{pre}}\\) and \\(\\Delta \\mathrm{FID}_{\\mathrm{real}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.55, + 0.892, + 0.672 + ], + "angle": 0, + "content": "Moreover, we conducted a multi-image test in an OOD scenario. In this particular experiment, we introduced additional metric - \\(\\mathrm{ID}_{\\mathrm{others}}\\) aimed at quantifying ID similarities for the unseen images associated with the source identity. As presented in Table 2, the introduction of the adjacency-aware unlearning loss resulted in a remarkable improvement in \\(\\mathrm{ID}_{\\mathrm{others}}\\), emphasizing the effectiveness of this unlearning approach for handling unseen images." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.681, + 0.654, + 0.697 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Effect of \\(d\\) in Determination of \\(w_{t}\\). We conducted an ablation study by comparing target images derived from varied values of \\(d\\). Setting \\(d\\) to 0 denotes utilizing the \\(\\bar{w}\\) as \\(w_{t}\\) in the unlearning process. For \\(d < 0\\), we designated \\(w_{t}\\) as an interpolated latent code between the \\(w_{s}\\) and the \\(\\bar{w}\\). Conversely, for \\(d > 0\\), we employed an extrapolated \\(w_{t}\\), as detailed in Section 3.2. As illustrated in Figure 9, when \\(d < 0\\), the target image closely aligns with the given source images. However, as \\(d\\) deviates from 0, the quality of the target image rapidly deteriorates, resulting in a pronounced collapse in the distribution of the pre-trained generator. Consequently, the effectiveness of unlearning with such target images diminishes in removing identity from the" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9157" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.098, + 0.09, + 0.457, + 0.241 + ], + "angle": 0, + "content": "
SourceTarget
d = -10d = -5d = 0d = 30 (Ours)d = 60
Random
FFHQ
d = -70d = -35d = 0d = 30 (Ours)d = 60
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.253, + 0.47, + 0.365 + ], + "angle": 0, + "content": "Figure 9. Ablation study to figure out the effectiveness of \\( d \\). We visualized target images corresponding to each source image with different values of \\( d \\). The target images were generated using target latent codes derived from interpolated latent codes, the average latent code (\\( d = 0 \\)), or extrapolated latent codes (\\( d > 0 \\)). Interpolation and extrapolation were carried out between the source and the average latent code. In the case of interpolation, the center between the source and the average latent code was computed." + }, + { + "type": "table", + "bbox": [ + 0.156, + 0.378, + 0.393, + 0.464 + ], + "angle": 0, + "content": "
αmaxID (↓)IDothers (↓)
00.1205 ± 0.06030.2754 ± 0.0791
100.0892 ± 0.06200.2123 ± 0.0762
150.0878 ± 0.03750.2094 ± 0.0692
200.0900 ± 0.05380.2105 ± 0.0924
300.0926 ± 0.05610.2111 ± 0.0653
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.473, + 0.47, + 0.543 + ], + "angle": 0, + "content": "Table 3. Ablation study to figure out the effectiveness of \\(\\mathcal{L}_{adj}\\) and \\(\\alpha_{max}\\). We compared the performance based on how successfully the given identity was erased, using ID and IDothers metric. The row where \\(\\alpha_{max} = 0\\) denotes the baseline. We used CelebAHQ dataset in this experiment." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.56, + 0.47, + 0.726 + ], + "angle": 0, + "content": "source images. Setting \\( d \\) to 0 might suggest the use of \\( \\bar{w} \\) as an effective target for erasing identity. However, our ablation studies indicate that when the source image closely aligns with \\( \\bar{w} \\), the unlearning procedure fails to thoroughly eliminate the identity. Conversely, when \\( d > 0 \\), \\( w_{t} \\) contains a distinct identity compared to the source image while maintaining a consistent distance from \\( \\bar{w} \\). Among the instances where \\( d > 0 \\), our ablation studies reveal that setting \\( d = 30 \\) achieves a balanced performance between effective unlearning and preservation of the generation performance of the pre-trained model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Effect of \\(\\alpha_{max}\\) in \\(\\mathcal{L}_{adj}\\). In Table 3, we scrutinized the effectiveness of the adjacency-aware unlearning loss. To ensure a fair comparison, we employed \\(\\bar{w}\\) as \\(w_{t}\\) in this experiment, and we used \\(\\mathcal{L}_{local}\\) and \\(\\mathcal{L}_{adj}\\) in the unlearning procedure. Rows corresponding to \\(\\alpha_{max} = 0\\) represent experimental results without the incorporation of \\(\\mathcal{L}_{adj}\\) in the unlearning procedure. The introduction of \\(\\mathcal{L}_{adj}\\) resulted in consistent performance gains in \\(\\mathrm{ID}_{\\mathrm{others}}\\). This observation highlights the efficacy of considering not only the pair of source and target latent codes but also their surroundings for unlearning the entire identity." + }, + { + "type": "table", + "bbox": [ + 0.56, + 0.09, + 0.835, + 0.139 + ], + "angle": 0, + "content": "
\\( \\mathcal{L}_{local} \\)\\( \\mathcal{L}_{global} \\)\\( \\mathrm{FID}_{\\text{pre}}(\\downarrow) \\)\\( \\Delta \\mathrm{FID}_{\\text{real}}(\\downarrow) \\)
✓9.52 ± 1.534.75 ± 0.89
✓✓4.63 ± 0.431.48 ± 0.29
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.149, + 0.892, + 0.205 + ], + "angle": 0, + "content": "Table 4. Ablation study to figure our the effectiveness of \\( \\mathcal{L}_{global} \\). We compared how preserved the performance of the pre-trained model through the unlearning process, via \\( \\mathrm{FID}_{\\mathrm{pre}} \\) and \\( \\Delta \\mathrm{FID}_{\\mathrm{real}} \\). We used CelebAHQ dataset in this experiment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.232, + 0.892, + 0.368 + ], + "angle": 0, + "content": "Effect of \\(\\mathcal{L}_{global}\\). To assess the effectiveness of the global preservation loss, a similar experiment was conducted as the previous experiment, i.e., setting \\(\\bar{w}\\) as \\(w_{t}\\). The results are presented in Table 4. The application of \\(\\mathcal{L}_{global}\\) demonstrated consistent performance improvements in both \\(\\mathrm{FID}_{\\mathrm{pre}}\\) and \\(\\Delta \\mathrm{FID}_{\\mathrm{real}}\\). This suggests that imposing constraints on the generator to maintain its generation performance in latent codes distant from our primary focus is effective in reducing distribution shifts in generative models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.383, + 0.619, + 0.398 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.408, + 0.892, + 0.696 + ], + "angle": 0, + "content": "In this paper, we introduced a novel task, referred to as generative identity unlearning, designed to address privacy concerns in pre-trained generative adversarial networks. This task requires thoroughly removing the identity of a single source image from the pre-trained generator. To achieve this, we proposed a new framework, GUIDE (Generative Unlearning for any IDENTITY). To unlearn the single identity, we first defined the target latent code via extrapolation, moving away from the average latent by the pre-defined distance in the direction from the source to the average latent. Using this, our GUIDE successfully unlearned the given identity via Latent Target Unlearning (LTU), which optimized the pre-trained model to preserve the overall generative ability but not to generate the same identity within the local space. Experimental results demonstrated the effectiveness of GUIDE with promising outcomes. We anticipate that our work will be widely applied in research or the industry field, providing users with a sense of freedom from privacy concerns through identity removal." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.71, + 0.66, + 0.726 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.733, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This work was supported by MSIT (Ministry of Science and ICT), Korea, under the ITRC (Information Technology Research Center) support program (IITP-2024-RS-2023-00258649) supervised by the IITP (Institute for Information & Communications Technology Planning & Evaluation), and in part by NRF-2023S1A5A2A21083590, and in part by the IITP grant funded by the Korea Government (MSIT) (Artificial Intelligence Innovation Hub) under Grant 2021-0-02068, and by the IITP grant funded by the Korea government (MSIT) (No.RS-2022-00155911, Artificial Intelligence Convergence Innovation Human Resources Development (Kyung Hee University))." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9158" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.085, + 0.176, + 0.1 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.11, + 0.47, + 0.166 + ], + "angle": 0, + "content": "[1] Yuval Alaluf, Or Patashnik, and Daniel Cohen-Or. Restyle: A residual-based stylegan encoder via iterative refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6711–6720, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.167, + 0.472, + 0.222 + ], + "angle": 0, + "content": "[2] Thomas Baumhauer, Pascal Schottle, and Matthias Zeppelinzauer. Machine unlearning: Linear filtration for logit-based classifiers. Machine Learning, 111(9):3203-3226, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.224, + 0.472, + 0.305 + ], + "angle": 0, + "content": "[3] Nicolas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikash Sehwag, Florian Tramer, Borja Balle, Daphne Ippolito, and Eric Wallace. Extracting training data from diffusion models. In Proceedings of the 32nd USENIX Security Symposium (USENIX Security 23), pages 5253-5270, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.308, + 0.472, + 0.405 + ], + "angle": 0, + "content": "[4] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.407, + 0.47, + 0.463 + ], + "angle": 0, + "content": "[5] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.464, + 0.47, + 0.517 + ], + "angle": 0, + "content": "[6] Vikram S Chundawat, Ayush K Tarun, Murari Mandal, and Mohan Kankanhalli. Zero-shot machine unlearning. IEEE Transactions on Information Forensics and Security, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.52, + 0.47, + 0.589 + ], + "angle": 0, + "content": "[7] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4690-4699, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.591, + 0.47, + 0.66 + ], + "angle": 0, + "content": "[8] Matt Fredrikson, Somesh Jha, and Thomas Ristenpart. Model inversion attacks that exploit confidence information and basic countermeasures. In Proceedings of the 22nd ACM SIGSAC conference on computer and communications security, pages 1322-1333, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.661, + 0.47, + 0.73 + ], + "angle": 0, + "content": "[9] Rohit Gandikota, Joanna Materzyńska, Jaden Fiotto-Kaufman, and David Bau. Erasing concepts from diffusion models. In Proceedings of the 2023 IEEE International Conference on Computer Vision (ICCV), pages 2426-2436, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.732, + 0.47, + 0.787 + ], + "angle": 0, + "content": "[10] Antonio Ginart, Melody Guan, Gregory Valiant, and James Y Zou. Making ai forget you: Data deletion in machine learning. Advances in Neural Information Processing Systems, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.789, + 0.47, + 0.857 + ], + "angle": 0, + "content": "[11] Aditya Golatkar, Alessandro Achille, and Stefano Soatto. Eternal sunshine of the spotless net: Selective forgetting in deep networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9304–9312, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.859, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[12] Aditya Golatkar, Alessandro Achille, and Stefano Soatto. Forgetting outside the box: Scrubbing deep networks of information accessible from input-output observations. In Pro" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.11, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "ceedings of the European Conference on Computer Vision (ECCV), pages 383-398. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.124, + 0.894, + 0.192 + ], + "angle": 0, + "content": "[13] Aditya Golatkar, Alessandro Achille, Avinash Ravichandran, Marzia Polito, and Stefano Soatto. Mixed-privacy forgetting in deep networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 792–801, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.195, + 0.892, + 0.25 + ], + "angle": 0, + "content": "[14] Varun Gupta, Christopher Jung, Seth Neel, Aaron Roth, Saeed Sharifi-Malvajerdi, and Chris Waites. Adaptive machine unlearning. Advances in Neural Information Processing Systems, 34:16319-16330, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.253, + 0.892, + 0.307 + ], + "angle": 0, + "content": "[15] Jamie Hayes, Luca Melis, George Danezis, and Emiliano De Cristofaro. Logan: Membership inference attacks against generative models. arXiv preprint arXiv:1705.07663, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.311, + 0.892, + 0.38 + ], + "angle": 0, + "content": "[16] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in Neural Information Processing Systems, 30, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.383, + 0.892, + 0.467 + ], + "angle": 0, + "content": "[17] Yuge Huang, Yuhan Wang, Ying Tai, Xiaoming Liu, Pengcheng Shen, Shaoxin Li, Jilin Li, and Feiyue Huang. Curricularface: adaptive curriculum learning loss for deep face recognition. In proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5901-5910, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.469, + 0.892, + 0.525 + ], + "angle": 0, + "content": "[18] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability and variation. In Proceedings of the International Conference on Learning Representations (ICLR), 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.527, + 0.892, + 0.583 + ], + "angle": 0, + "content": "[19] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.585, + 0.892, + 0.655 + ], + "angle": 0, + "content": "[20] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.657, + 0.892, + 0.713 + ], + "angle": 0, + "content": "[21] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.715, + 0.892, + 0.798 + ], + "angle": 0, + "content": "[22] Mahdi Khosravy, Kazuaki Nakamura, Yuki Hirose, Naoko Nitta, and Noboru Babaguchi. Model inversion attack by integration of deep generative models: Privacy-sensitive face generation from a face recognition system. IEEE Transactions on Information Forensics and Security, 17:357-372, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.801, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[23] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proceedings of the International Conference on Learning Representations (ICLR), 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[24] Zhifeng Kong and Kamalika Chaudhuri. Data redaction from pre-trained gans. In Proceedings of the 2023 IEEE Conference on Secure and Trustworthy Machine Learning (SaTML), pages 638-677. IEEE, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "9159" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.161 + ], + "angle": 0, + "content": "[25] Nupur Kumari, Bingliang Zhang, Sheng-Yu Wang, Eli Shechtman, Richard Zhang, and Jun-Yan Zhu. Ablating concepts in text-to-image diffusion models. In Proceedings of the 2023 IEEE International Conference on Computer Vision (ICCV), pages 22691-22702, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.472, + 0.232 + ], + "angle": 0, + "content": "[26] Ronak Mehta, Sourav Pal, Vikas Singh, and Sathya N Ravi. Deep unlearning via randomized conditionally independent hessian. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10422-10431, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.232, + 0.472, + 0.313 + ], + "angle": 0, + "content": "[27] Jun-Yeong Moon, Keon-Hee Park, Jung Uk Kim, and Gyeong-Moon Park. Online class incremental learning on stochastic blurry task boundary via mask and visual prompt tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11731-11741, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.315, + 0.472, + 0.357 + ], + "angle": 0, + "content": "[28] Saemi Moon, Seunghyuk Cho, and Dongwoo Kim. Feature unlearning for generative models via implicit feedback. arXiv preprint arXiv:2303.05699, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.358, + 0.472, + 0.413 + ], + "angle": 0, + "content": "[29] Seung-Jun Moon and Gyeong-Moon Park. Interstyle: Encoding an interest region for robust stylegan inversion. In Proceedings of the European Conference on Computer Vision (ECCV), pages 460-476. Springer, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.414, + 0.472, + 0.482 + ], + "angle": 0, + "content": "[30] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2085–2094, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.483, + 0.472, + 0.537 + ], + "angle": 0, + "content": "[31] Alexandra Peste, Dan Alistarh, and Christoph H Lampert. Ssse: Efficiently erasing samples from trained machine learning models. corr. arXiv preprint arXiv:2107.03860, 1 (3), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.538, + 0.472, + 0.607 + ], + "angle": 0, + "content": "[32] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In Proceedings of the International Conference on Machine Learning (ICML), pages 8821-8831. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.472, + 0.663 + ], + "angle": 0, + "content": "[33] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.472, + 0.746 + ], + "angle": 0, + "content": "[34] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2287–2296, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.472, + 0.802 + ], + "angle": 0, + "content": "[35] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on graphics (TOG), 42(1):1-13, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.472, + 0.873 + ], + "angle": 0, + "content": "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[37] Juwon Seo, Ji-Su Kang, and Gyeong-Moon Park. Lfs-gan: Lifelong few-shot image generation. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "IEEE/CVF International Conference on Computer Vision (ICCV), pages 11356-11366, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[38] Shawn Shan, Jenna Cryan, Emily Wenger, Haitao Zheng, Rana Hanocka, and Ben Y Zhao. Glaze: Protecting artists from style mimicry by text-to-image models. arXiv preprint arXiv:2302.04222, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[39] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9243-9252, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[40] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. Interfacegan: Interpreting the disentangled face representation learned by gans. IEEE transactions on pattern analysis and machine intelligence, 44(4):2004-2018, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.307, + 0.892, + 0.376 + ], + "angle": 0, + "content": "[41] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Diffusion art or digital forgery? investigating data replication in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6048-6058, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.377, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[42] Jingxiang Sun, Xuan Wang, Lizhen Wang, Xiaoyu Li, Yong Zhang, Hongwen Zhang, and Yebin Liu. Next3d: Generative neural texture rasterization for 3d-aware head avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20991-21002, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.461, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[43] Ayush K Tarun, Vikram S Chandawat, Murari Mandal, and Mohan Kankanhalli. Fast yet effective machine unlearning. IEEE Transactions on Neural Networks and Learning Systems, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.519, + 0.892, + 0.573 + ], + "angle": 0, + "content": "[44] Piyush Tiwary, Atri Guha, Subhodip Panda, et al. Adapt then unlearn: Exploiting parameter space semantics for unlearning in generative adversarial networks. arXiv preprint arXiv:2309.14054, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.576, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[45] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40(4): 1-14, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.632, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[46] Tengfei Wang, Yong Zhang, Yanbo Fan, Jue Wang, and Qifeng Chen. High-fidelity gan inversion for image attribute editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11379-11388, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.892, + 0.759 + ], + "angle": 0, + "content": "[47] Yue Wu, Yu Deng, Jiaolong Yang, Fangyun Wei, Qifeng Chen, and Xin Tong. Anifacegan: Animatable 3d-aware face image generation for video avatars. Advances in Neural Information Processing Systems, 35:36188-36201, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.76, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[48] Yuting Xu, Jian Liang, Gengyun Jia, Ziming Yang, Yanhao Zhang, and Ran He. Tall: Thumbnail layout for deepfake video detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22658-22668, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.831, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[49] Zhiyuan Yan, Yong Zhang, Yanbo Fan, and Baoyuan Wu. Ucf: Uncovering common features for generalizable deepfake detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22412-22423, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "9160" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[50] Jaejun Yoo, Youngjung Uh, Sanghyuk Chun, Byeongkyu Kang, and Jung-Woo Ha. Photorealistic style transfer via wavelet transforms. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9036-9045, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.219 + ], + "angle": 0, + "content": "[51] Xiaoyong Yuan, Pan He, Qile Zhu, and Xiaolin Li. Adversarial examples: Attacks and defenses for deep learning. IEEE transactions on neural networks and learning systems, 30 (9):2805-2824, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.22, + 0.471, + 0.29 + ], + "angle": 0, + "content": "[52] Ziyang Yuan, Yiming Zhu, Yu Li, Hongyu Liu, and Chun Yuan. Make encoder great again in 3d gan inversion through geometry and occlusion-aware encoding. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2437-2447, 2023. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.471, + 0.344 + ], + "angle": 0, + "content": "[53] Eric Zhang, Kai Wang, Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Forget-me-not: Learning to forget in text-to-image diffusion models. arXiv preprint arXiv:2303.17591, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.346, + 0.471, + 0.416 + ], + "angle": 0, + "content": "[54] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.418, + 0.47, + 0.487 + ], + "angle": 0, + "content": "[55] Xuanmeng Zhang, Zhedong Zheng, Daiheng Gao, Bang Zhang, Yi Yang, and Tat-Seng Chua. Multi-view consistent generative adversarial networks for compositional 3d-aware image synthesis. International Journal of Computer Vision, pages 1-24, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.488, + 0.47, + 0.558 + ], + "angle": 0, + "content": "[56] Xiaoming Zhao, Fangchang Ma, David Guera, Zhile Ren, Alexander G Schwing, and Alex Colburn. Generative multiplane images: Making a 2d gan 3d-aware. In Proceedings of the European Conference on Computer Vision (ECCV), pages 18-35. Springer, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.558 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.956 + ], + "angle": 0, + "content": "9161" + } + ] +] \ No newline at end of file diff --git a/2024/Generative Unlearning for Any Identity/c3c2d3ee-c7c2-4590-9220-bed4fbf9b764_origin.pdf b/2024/Generative Unlearning for Any Identity/c3c2d3ee-c7c2-4590-9220-bed4fbf9b764_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6fdf36dcc602da4486f5e31dfb4e9ee5afa3615a --- /dev/null +++ b/2024/Generative Unlearning for Any Identity/c3c2d3ee-c7c2-4590-9220-bed4fbf9b764_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e5607d48a1e22a335863697a8e870fe8df7ba817d9971de34efcd1dce8935162 +size 4512236 diff --git a/2024/Generative Unlearning for Any Identity/full.md b/2024/Generative Unlearning for Any Identity/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0afcb3bd1580800fb07a8b7f3a7d67af15d08ee8 --- /dev/null +++ b/2024/Generative Unlearning for Any Identity/full.md @@ -0,0 +1,308 @@ +# Generative Unlearning for Any Identity + +Juwon $\mathrm{Seo}^{1*}$ + +Sung-Hoon Lee $^{1*}$ Tae-Young Lee $^{1*}$ Gyeong-Moon Park $^{1\dagger}$ + +Seungjun Moon² + +1Kyung Hee University, Yongin, Republic of Korea +2KLleon Tech., Seoul, Republic of Korea + +{jwseo001, sunghoonlee961, slcks1, gmpark}@khu.ac.kr seungjunmoon@klleon.io + +# Abstract + +Recent advances in generative models trained on large-scale datasets have made it possible to synthesize high-quality samples across various domains. Moreover, the emergence of strong inversion networks enables not only a reconstruction of real-world images but also the modification of attributes through various editing methods. However, in certain domains related to privacy issues, e.g., human faces, advanced generative models along with strong inversion methods can lead to potential misuses. In this paper, we propose an essential yet under-explored task called generative identity unlearning, which steers the model not to generate an image of specific identity. In the generative identity unlearning, we target the following objectives: (i) preventing the generation of images with a certain identity, and (ii) preserving the overall quality of the generative model. To satisfy these goals, we propose a novel framework, Generative Unlearning for Any IDENTITY (GUIDE), which prevents the reconstruction of a specific identity by unlearning the generator with only a single image. GUIDE consists of two parts: (i) finding a target point for optimization that un-identifies the source latent code and (ii) novel loss functions that facilitate the unlearning procedure while less affecting the learned distribution. Our extensive experiments demonstrate that our proposed method achieves state-of-the-art performance in the generative machine unlearning task. The code is available at https://github.com/KHU-AGI/GUIDE. + +# 1. Introduction + +Recently, 2D or 3D Generative Adversarial Networks (GANs) [4, 19-21] pre-trained on large datasets, e.g., FFHQ + +![](images/30461510649baa1c9b78705c4fac198e76a80a95d1ce761366dd47823862bd32.jpg) +Figure 1. Given a single source image containing a specific identity, we remove that identity from the pre-trained 3D generative adversarial network (e.g., EG3D [4]). Our method effectively unlearns identity even from in-the-wild images where the source image is absent in the pre-training dataset. + +[19] or AFHQ [5], have drawn substantial attention due to their remarkable generation performance and highly disentangled representation space. However, their advancements have raised privacy concerns [15], especially regarding the potential misuse of generative models to represent and exploit individual identities. For instance, deepfakes [48, 49] can create very believable images or videos of people in made-up situations, causing major concerns about ethics and privacy. + +To alleviate privacy issues in generative models, machine unlearning task has been actively studied. Machine unlearning involves the process of selectively removing specific knowledge or erasing the influence of certain data from the training dataset of pre-trained models. It is beneficial especially when the data are harmful, private, or biased + +[11, 12, 26, 31]. Despite a focus on discriminative tasks in most machine unlearning research, a few studies have ventured into generative models, attempting to erase high-level concepts such as socially inappropriate content or artistic styles that present copyright challenges [9, 25]. + +Nevertheless, generative models still exhibit ongoing privacy issues. Even if an identity of someone is not used in the pre-training of the generative models, it can be easily reconstructed in the pre-trained models via GAN inversion models [29, 34, 35, 45, 46, 52]. Furthermore, the reconstructed image can be manipulated or edited easily via image editing methods [30, 39, 40]. To prevent potential exploits of an identity, it is necessary to erase a certain identity from the pre-trained generative models. + +To consider the above issue, we introduce an essential task of unlearning any identity from the pre-trained 2D or 3D GANs [4, 20], called generative identity unlearning. Unlike typical machine unlearning tasks, which focus on unlearning the training samples our generative identity unlearning task unlearns any identity on pre-trained GANs, even if it was not shown during the pre-training. Our goal is to remove the whole identity associated with a given single image from the generator while minimally impacting the overall performance of the pre-trained model. + +To achieve our goal, we propose a novel generative unlearning framework, Generative Unlearning for Any IDEntity, named GUIDE. GUIDE replaces the source identity with an anonymous target identity, erasing the original identity effectively. To this end, we propose a new exploration method to determine an effective target latent code, called Un-Identifying Face On latent space (UFO). UFO utilizes the GAN inversion method [52] to embed the given identity into the source latent, and then decides the target latent using both the source and the average latent codes. We empirically find that the proposed UFO can identify the promising target to erase any given source identity robustly. + +Given the source and target latent code, we update the generator to shift from the source identity to the target identity. To this end, we propose three novel loss functions: (i) local unlearning loss, (ii) adjacency-aware unlearning loss, and (iii) global preservation loss. (i) guides our model directly shifting the source identity to the target identity. (ii) utilizes other latent codes adjacent to the source and target latent codes to effectively unlearn the entire identity from a single image. To minimize side effects from the unlearning process, (iii) additionally regularizes the generator to retain generation performance for latent codes relatively far from the source and target latent codes. Through comprehensive experiments on diverse identities, including Random, InD, and OOD, we confirm that GUIDE can successfully remove the identity of the source image from the pre-trained generative model, and shows qualitatively and quantitatively superior performances. + +Our contributions can be summarized as follows: + +- For the first time, we propose a novel task, generative identity unlearning, which tackles machine unlearning in generative models in the aspect of privacy protection. In our task, we aim to prevent the pre-trained generative models from synthesizing the given identity by utilizing only a single image. +- For the effective and robust elimination of the identity, we propose a novel method - Un-Identifying Face On Latent Space (UFO). We configure the unlearning procedure by formulating how to represent and shift the identity in the latent space. We find that setting the extrapolated latent code between the source and average latent codes as an optimization target facilitates the unlearning procedure. +- We propose three loss functions - local unlearning loss, adjacency-aware unlearning loss, and global preservation loss to effectively unlearn the identity from the pretrained model while less affecting the generation performance on other identity. +- We show that our proposed framework, GUIDE achieves state-of-the-art performance both qualitatively and quantitatively, through extensive experiments. We demonstrate that GUIDE can remove the specific identity successfully in the generative models while minimizing the negative effect on other identities. + +# 2. Related Work + +Generative Models and Privacy Issue. In image synthesis field, GAN-based generative models have achieved remarkable performance not only in 2D [18-21, 32, 33, 36, 37] but also in 3D [4, 42, 47, 55, 56] domain. The application of various image editing methods [30, 40, 50] to strong generative models, people can easily generate edited images of specific individuals and various artistic styles [38, 41], as well as extract copyrighted content [3] without the permission of the individual or the original creator. + +Recently, with the rise of the importance of AI ethics, several works have addressed this issue [9, 25, 28, 44, 53]. ESD [9] erases specific visual concepts from diffusion model by using negative guidance about the undesired concepts. Kumari et al. [25] modifies the conditional distribution of the model a specific target concepts to match the anchor concept. Forget-Me-Not [53] fine-tunes U-Net to minimize each of the intermediate attention associated with the target concepts to remove. Additional works in GANs [28, 44] focus on unlearning specific features, e.g., "Bang", "Hat" or "Beard" rather than forgetting specific identity. The above methods primarily concentrate on the elimination of specific concepts or high-level features. In other words, these cannot preclude models from generating specific individuals while maintaining the generation performance of realistic human faces. Unlike the existing works, our work targets to unlearn only specific individuals without + +shifting the overall distribution of generated images. + +Machine Unlearning. Machine unlearning aims to selectively forget specific acquired knowledge or diminish the impact of certain training data subsets on a trained model. Since previous research [8, 22, 51] shows that machine learning models might accidentally share private information when faced with certain attacks or inputs, machine unlearning becomes crucial. + +While previous machine unlearning is mainly focused on supervised learning tasks [2, 6, 10-14, 43], the interest in unlearning techniques within unsupervised learning, i.e., generative models, is growing [24, 28, 44]. However, most of the existing methods need full dataset access for retraining, which is hard to acquire and computationally expensive [2, 10, 14, 43]. For example, Kong and Chaudhuri [24] utilizes data reduction and augmentation algorithms, which requires a full training dataset. Despite the existence of a feature unlearning model [28] which does not need full dataset access, unlearning only an individual feature is not enough to forget a whole specific identity. To this end, we propose an algorithm that enables forgetting the specific identity only with a single image. Furthermore, our approach distinguishes itself from existing research by applying unlearning to unseen images, enabling the erasure of specific identities without prior exposure to those images. + +# 3. Method + +Firstly, in Section 3.1, as shown in Figure 2, we introduce the problem we aim to address, named generative identity unlearning. In Section 3.2, we introduce un-identifying face on latent space, which designates an appropriate target latent for unlearning. Then, in Section 3.3, we introduce latent target unlearning, along with our proposed novel losses to unlearn the generator. The total overview of our method is illustrated in Figure 3. + +# 3.1. Problem Formulation + +Given a set of images $\mathbf{x}$ depicting a specific identity, we randomly select a single source image $x_{u} \in \mathbf{x}$ as an exemplar of the identity. Initially, using off-the-shelf inversion network [52] $E$ corresponding to the unconditional generator, i.e., EG3D [4], we embed $x_{u}$ to the source latent code $w_{u}$ in the latent space of EG3D: + +$$ +w _ {u} = E \left(x _ {u}\right). \tag {1} +$$ + +Since EG3D [4] consists of the mapping network $Map(\cdot)$ , StyleGAN2 [19] backbone $G(\cdot)$ and the neural renderer with a super-resolution module $R(\cdot)$ , we can denote the reconstructed image $\hat{x}$ from $w_{u}$ as following: + +$$ +\hat {x} = R \left(G \left(w _ {u}\right); c\right), \tag {2} +$$ + +![](images/cde3a6799f641e5b379ca23096bc13eef2febd497281e44fb2fad14724196e95.jpg) +Figure 2. An illustration of generative identity unlearning. Upon GUIDE, the identity of the image generated from $w_{u}$ , i.e., inversion of the source image $x_{u}$ by inversion network $E$ , should exhibit a distinct identity when passed through the pre-trained generator $G_{s}$ compared to the unlearned generator $G_{u}$ . Furthermore, other images $x_{o}$ , not used in unlearning but sharing the same identity with $x_{u}$ , also should vary an identity through GUIDE. + +where $c$ denotes camera poses. For convenience, we omit the explicit notation of camera poses in this paper, i.e., $\hat{x} = (R \circ G)(w_u)$ . We target to derive an unlearned $G$ , i.e., $G_u$ , from the pre-trained EG3D generator $G$ , i.e., $G_s$ , while fixing $Map$ and $R$ . With proper unlearning, an image generated by unlearned EG3D using $w_u$ , i.e., $\hat{x}_u = (R \circ G_u)(w_u)$ should have a distinct identity from the image generated by original EG3D using $w_u$ , i.e., $(R \circ G_s)(w_u)$ . + +In our task formulation, two considerations are paramount. First, we aim to eliminate the entire identity from the generator only utilizing a single image. To validate this, we incorporate other multiple images $x_{o} \in \mathbf{x}$ for testing and its corresponding latent code $w_{o} = E(x_{o})$ . By utilizing $x_{o}$ , we can verify whether $G_{u}$ has successfully unlearned the identity as a whole, rather than just unlearning the specific image $x_{u}$ . Second, we strive to maintain the generation performance of the pre-trained model. To assess this, we sample multiple images from fixed latent codes using both the unlearned and pre-trained generators. We then estimate the distribution shift between the images generated before and after the unlearning process. + +# 3.2. Un-Identifying Face On Latent Space + +The successfully unlearned model should not generate the image with the identity of $\mathbf{x}$ , even when $w_{u}$ is used as a latent. Consequently, we initiate our approach to manipulate $\hat{x}_u$ to be another image rather than the image with identity in $\mathbf{x}$ by unlearning $G$ . To design the objective function for unlearning, we first need to establish the objective for unlearning, which involves defining the target image $\hat{x}_t$ , derived from $w_t$ , which the unlearned image $\hat{x}_u$ , derived from $w_{u}$ , should mimic after unlearning. While there exist various options for setting $\hat{x}_t$ , e.g., a random face or even a non-human image, we choose the mean face generated by + +![](images/ffa2618d4725731f575a1660beb56ea64b94f932c30b6ff96b23cdc5937bf14f.jpg) +Figure 3. An overview of GUIDE. Starting with a source image, we employ a GAN inversion network $E$ , specifically GOAE [52], to embed this image into the latent space of a pre-trained generative model, namely EG3D [4], obtaining the source latent code $w_{u}$ . The target latent code $w_{t}$ is designated through the UFO process. To facilitate identity removal in $w_{u}$ , we shift its identity to match that of $w_{t}$ with our Latent Target Unlearning (LTU) process. Three loss functions of LTU are designed for this purpose: (i) The generator is optimized to produce an image from the source latent code, denoted as $(R \circ G_{u})(w_{u})$ , that is similar to the image from the target latent code, represented as $(R \circ G_{s})(w_{t})$ . (ii) To achieve unlearning across the entire identity, we consider latent codes near both the source and target latent codes, denoted as $w_{u,a}$ and $w_{t,a}$ , respectively. (iii) To prevent model corruption during the unlearning process, we additionally sample latent codes from a random noise vector, represented as $w_{r,g}$ , and optimize $G_{u}$ to preserve its generation ability on $w_{r,g}$ . + +![](images/a498b047462a15846977f2f41cc34875a1ab9f8e3342a8b5deab74778be983cf.jpg) +Figure 4. An illustration of Un-identifying Face On Latent Space (UFO). We define the identity of the source latent code by subtract it from the average latent code. We set the target latent code for our unlearning process by measuring an extrapolation between the source and average latent code with a fixed distance $d$ . + +the mean latent $\bar{w}$ of $Map$ , i.e., $(R \circ G_s)(\bar{w})$ , as $\hat{x}_t$ . This setting is inspired by inversion methods [1, 29] which apply the identity to the mean face through inversion stages. Since our goal is to de-identify the image, we argue that it is intuitive and reasonable to send the image back to the mean face, which is opposite to the inversion. + +However, this simple approach might be problematic when $w_{u}$ is close to $\bar{w}$ , where the unlearned image might still resemble the original identity. To this end, we propose a novel method named Un-identifying Face On latent space (UFO), which can set $\hat{x}_t$ robustly regardless of the distance between $w_{u}$ and $\bar{w}$ , as shown in Figure 4. The following are the processes of UFO: First, in the $de$ -identification process, we extract the identity latent $w_{id} = w_{u} - \bar{w}$ by subtracting $\bar{w}$ from $w_{u}$ . Then we propose a process of en-identification, + +setting the target in the opposite direction of the existing $w_{id}$ to foster a more pronounced change in identity, creating an entirely new identity. In other words, we chose $\bar{w}$ as the stopover point and established the final target point at the extrapolation of $w_{u}$ and $\bar{w}$ . The finalized target point can be expressed as: + +$$ +w _ {t} = \bar {w} - d \cdot \frac {w _ {i d}}{\| w _ {i d} \| _ {2}}, \tag {3} +$$ + +where $d$ is defined as the distance that balances un-identification with preservation of the source distribution, as determined by our empirical results. Finally, we can set $\hat{x}_t = (R\circ G_s)(w_t)$ . We empirically demonstrate that whether the given identity is in close proximity to the average latent code within the latent space or not, UFO sets the desirable target latent code for the generative unlearning. + +# 3.3. Latent Target Unlearning + +After setting $\hat{x}_t$ , we need to robustly design the unlearning process that can effectively make $\hat{x}_u$ to be similar to $\hat{x}_t$ , while maintaining the generation performance of $G_u$ . We coin this process as Latent Target Unlearning (LTU), which targets to unlearn images from the specific latent while keeping generation results from other latent codes. LTU utilizes the following three losses to achieve this goal: + +Local Unlearning Loss. To force $\hat{x}_u$ to be $\hat{x}_t$ , we use the widely-used reconstruction losses, i.e., Euclidean loss $\mathcal{L}_2$ , + +![](images/9db8382f578fb7d11023da21d017b3da352967bb5ef2e563a6f054bdafcc50d1.jpg) +Figure 5. An illustration of determining latent codes near a latent code $w$ in adjacency-aware unlearning loss. We first sample a latent code $w_{r,a}$ which is derived from a random noise vector $z_{r,a}$ via the mapping network $\text{Map}(\cdot)$ , i.e. $w_{r,a} = \text{Map}(z_{r,a})$ . Next, we compute the direction between $w$ and $w_{r,a}$ , and we scale it to fall within range between 0 and $\alpha_{max}$ . This yields the distance vector $\Delta$ to compute the adjacent latent code $w_{a} = w + \Delta$ . + +perceptual loss $\mathcal{L}_{per}$ [54], and identity loss $\mathcal{L}_{id}$ [7]. Using $\mathcal{L}_{recon}$ , we compare the tri-plane features $F_{u} = G_{u}(w_{u})$ and $F_{t} = G_{s}(w_{t})$ , derived from source and target latent codes, respectively. The local unlearning loss is defined as: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {l o c a l}} \left(\hat {x} _ {u}, \hat {x} _ {t}\right) = \lambda_ {L 2} \mathcal {L} _ {L 2} \left(F _ {u}, F _ {t}\right) \tag {4} \\ + \lambda_ {p e r} \mathcal {L} _ {p e r} (\hat {x} _ {u}, \hat {x} _ {t}) + \lambda_ {i d} \mathcal {L} _ {i d} (\hat {x} _ {u}, \hat {x} _ {t}). \\ \end{array} +$$ + +By adopting $\mathcal{L}_{local}$ , we can successfully un-identify the given source identity in $\hat{x}_t$ . + +Adjacency-Aware Unlearning Loss. The above equation considers only one pair of source and target latent codes. However, images of a similar identity to the source identity can be obtained by introducing marginal perturbations to the latent code. For the successful unlearning of the given identity, we need to consider the neighborhood of both the source and the target latent codes. Consequently, as shown in Figure 5, we sample $N_{a}$ latent codes in the vicinity of the $w_{u}$ . Specifically, with the scale $\alpha^{i}$ sampled from the uniform distribution with hyperparameter $\alpha_{max}$ , i.e. $\alpha^{i} \sim \mathcal{U}(0, \alpha_{max})$ , we define the distances $\Delta$ to compute the adjacent latent codes as: + +$$ +\Delta = \left\{\alpha^ {i} \cdot \frac {w _ {r , a} ^ {i} - w _ {u}}{\| w _ {r , a} ^ {i} - w _ {u} \| _ {2}} \right\} _ {i = 1} ^ {N _ {a}}, \tag {5} +$$ + +where $w_{r,a}^{i}$ is a latent code sampled from the random noise vector $z_{r,a}^{i}$ . Using these distances $\Delta$ , we can compute $N_{a}$ latent codes for both the source and the target latent codes. Similar to the local unlearning loss, we optimize the generated tri-plane features and images from $w_{u,a}^{i} = w_{u} + \Delta^{i}$ and $w_{t,a}^{i} = w_{t} + \Delta^{i}$ to be similar: + +$$ +\hat {x} _ {u, a} ^ {i} = R \left(F _ {u, a} ^ {i}\right), \hat {x} _ {t, a} ^ {i} = R \left(F _ {u, a} ^ {i}\right), \tag {6} +$$ + +$$ +\mathcal {L} _ {a d j} \left(w _ {u}, w _ {t}\right) = \frac {1}{N _ {a}} \sum_ {i = 1} ^ {N _ {a}} \mathcal {L} _ {\text {l o c a l}} \left(\hat {x} _ {u, a} ^ {i}, \hat {x} _ {t, a} ^ {i}\right), \tag {7} +$$ + +where $F_{u,a}^{i} = G_{u}(w_{u,a}^{i}), F_{t,a}^{i} = G_{s}(w_{t,a}^{i})$ denotes for triplane features, and $\mathcal{L}_{local}$ in Equation 4. From $\mathcal{L}_{adj}$ , we can further consider possible variations of the source identity. + +Global Preservation Loss. While the local unlearning loss and adjacency-aware unlearning loss mentioned above facilitate the removal of the source identity, we propose a global preservation loss to mitigate side effects arising from these unlearning loss functions. In the global preservation loss, we constrain the generator to maintain generation performance for latent codes that are relatively distant from both the source and target latent codes. + +To be precise, we sample $N_{g}$ latent codes $\{w_{r,g}^{i}\}_{i = 1}^{N_{g}}$ from random noise vectors $\{z_{r,g}^{i}\}_{i = 1}^{N_{g}}$ . We ensure that these do not overlap with the adjacent latent codes used in the adjacency-aware unlearning loss. Unlike the unlearning loss functions, we find that adopting only $L_{per}$ achieves a balanced performance between identity shift and model preservation. The global preservation loss is computed as: + +$$ +\hat {x} _ {u, g} ^ {i} = (R \circ G _ {u}) (w _ {r, g} ^ {i}), +$$ + +$$ +\hat {x} _ {s, g} ^ {i} = \left(R \circ G _ {s}\right) \left(w _ {r, g} ^ {i}\right), \tag {8} +$$ + +$$ +\mathcal {L} _ {g l o b a l} (G _ {u}, G _ {s}) = \frac {1}{N _ {g}} \sum_ {i = 1} ^ {N _ {g}} \mathcal {L} _ {p e r} (\hat {x} _ {u, g} ^ {i}, \hat {x} _ {s, g} ^ {i}). +$$ + +In summary, our final objective is: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {l o c a l}} + \lambda_ {\text {a d j}} \mathcal {L} _ {\text {a d j}} + \lambda_ {\text {g l o b a l}} \mathcal {L} _ {\text {g l o b a l}}. \tag {9} +$$ + +# 4. Experiments + +# 4.1. Experimental Setup + +**Baseline.** Since we propose generative identity unlearning task for the first time, to evaluate the effectiveness of GUIDE, we constructed a simple baseline. In the baseline, we used the target latent code as the average one for the unlearning. During the unlearning, we updated the pre-trained generator using $\mathcal{L}_{local}$ as described in Equation 4. + +Implementation Details. We built GUIDE based on the 3D generative adversarial network [4] pre-trained on FFHQ dataset [19]. We used GOAE [52] as a GAN inversion network to obtain the latent code from an image. The image resolution we used in our experiments is $512 \times 512$ with a rendering resolution of $128 \times 128$ . We used Adam optimizer + +
SourceRandomIn-Domain (FFHQ)Out-of-Domain (CelebAHQ)
BaselineOursBaselineOursBaselineOursBaselineOursBaselineOurs
Unlearned
+ +[23] with a learning rate of $10^{-4}$ in the unlearning procedure. The hyperparameters used in the experiments were: $d = 30$ , $\alpha_{max} = 15$ , $\lambda_{L2} = 10^{-2}$ , $\lambda_{per} = 1$ , $\lambda_{id} = 10^{-1}$ , $N_a = N_g = 2$ , and $\lambda_{adj} = \lambda_{global} = 1$ . + +Dataset and Scenarios. We evaluated GUIDE in three scenarios: Random, where we set an unlearning target image from a randomly sampled noise vector; InD (domain), where we sampled an image from the FFHQ dataset [19] used for pre-training; and OOD (out-of-domain), where the unlearning target image was sampled from the CelebAHQ dataset [18]. For InD and OOD, we used the GAN inversion network to obtain corresponding latent codes. For OOD scenario, we also conducted multi-image test since there were multiple images with a same identity in CelebAHQ. On the other hand, we performed only single-image test in the Random and InD scenarios. + +Evaluation Metrics. We evaluated GUIDE on two key aspects. Firstly, we estimated the efficacy of our approach in preventing the generator from producing images similar to the unlearning target. We quantitatively measured similarity of identities (ID) using face recognition network, CurricularFace [17], between images generated from the same latent codes before and after unlearning. Moreover, we utilized $\mathrm{ID}_{\mathrm{others}}$ to estimate the erasure of a identity from images which are unseen during training but containing the same identity of the source image. Secondly, we assessed whether our method preserves overall generation performance using the Fréchet Inception distance (FID) score [16]. Different from the existing usages, we utilized two variants of FID. First, we evaluated the distribution shift of generated images the pre-trained generator and the un- + +![](images/98eb4066c843cf5e7dbf10a6c817e84367ed08a86a316b95ada8d29f1f722a01.jpg) +Figure 6. Qualitative results of GUIDE and the baseline in generative identity unlearning task. For the given source image each (the first row), GUIDE and the baseline tried to erase the identity in the pre-trained generator. The images in the second and third row are the target and unlearned images, respectively. +Figure 7. Qualitative results of GUIDE and the baseline on a multi-image test using CelebAHQ dataset. We additionally utilized images that are unseen during unlearning, to show how thoroughly erase the given identity. + +learned generator via $\mathrm{FID}_{\mathrm{pre}}$ . Furthermore, we measured the distribution shift with respect to the real FFHQ images, which we denoted as $\Delta \mathrm{FID}_{\mathrm{real}}$ . + +# 4.2. Main Experiment + +# 4.2.1 Qualitative Results + +We conducted a comparative analysis of GUIDE against the baseline in the generative identity unlearning task. Initiating from the provided source image, we aimed to eliminate the identity within the pre-trained generator, as illustrated in Figure 6. We presented the resulting unlearned image, along with the target image optimized in our loss functions. Notably, GUIDE effectively erases identities whether synthetic, presented during pre-training, or unseen. + +
MethodsRandomIn-Domain (FFHQ)Out-of-Domain (CelebAHQ)
ID (↓)FIDpre (↓)ΔFIDreal (↓)ID (↓)FIDpre (↓)ΔFIDreal (↓)ID (↓)FIDpre (↓)ΔFIDreal (↓)
Baseline0.19 ± 0.0911.73 ± 2.747.46 ± 2.200.16 ± 0.079.00 ± 1.154.15 ± 1.180.12 ± 0.069.52 ± 1.534.75 ± 0.89
+ extrapolated wt0.12 ± 0.0614.28 ± 3.349.63 ± 2.530.05 ± 0.0612.78 ± 1.826.76 ± 1.410.02 ± 0.0513.02 ± 3.207.31 ± 1.98
+ Ladj0.14 ± 0.0719.65 ± 4.9013.94 ± 3.590.04 ± 0.0613.53 ± 2.087.35 ± 1.700.01 ± 0.0513.63 ± 3.527.83 ± 2.19
+ Lglobal (GUIDE)0.14 ± 0.0610.80 ± 2.706.64 ± 1.600.06 ± 0.068.00 ± 1.203.05 ± 0.810.03 ± 0.057.88 ± 1.963.34 ± 1.10
+ +![](images/4852243969db08bbcc2f30262ac7108a70c1fb4675d35e4b502a4f324d6091cd.jpg) +Figure 8. Qualitative comparison between GUIDE and the baseline on the preservation of the generation quality of other identities. GUIDE generates images almost identical to those synthesized by $G_{s}$ , whereas the baseline often results in noticeable changes, e.g., beard shape, hairstyle change, hat. + +To evaluate the thoroughness of identity removal, we performed a multi-image test using identities from the CelebAHQ dataset. This test involved assessing the ID similarity not only for the unlearned image derived from the source image but also for other images sharing the same identity. As shown in Figure 7, GUIDE showed superior generalization for unseen images compared to the baseline. This improvement is attributed to the adjacency-aware unlearning, which facilitated the unlearning process not just for the given images but also for their neighborhood. + +In Figure 8, we conducted an experiment to assess the effect of the unlearning process on other identities. While the baseline had a significant impact on other identities through the unlearning, GUIDE showed a relatively lesser effect. We attribute this to the global preservation loss, which constrained the distribution shift on other latent codes. + +# 4.2.2 Quantitative Results + +In Table 1, we compared GUIDE to the baseline by gradually applying the components of GUIDE. By configuring $w_{t}$ through extrapolation, we achieved performance improvements in ID similarity across three scenarios. Notably, we observed a significant difference in ID similarities in the + +Table 1. Quantitative results of GUIDE and the baseline in the generative identity unlearning task, tested in a single-image setting using one image per identity. Starting from the baseline, we gradually introduced components of GUIDE. + +
MethodsID (↓)ID others (↓)FID pre (↓)ΔFID real (↓)
Baseline0.12 ± 0.060.28 ± 0.089.52 ± 1.534.75 ± 0.89
+ extrapolated wt0.02 ± 0.050.15 ± 0.0713.02 ± 3.207.31 ± 1.98
+ Ladj0.01 ± 0.050.14 ± 0.0713.63 ± 3.527.83 ± 2.19
+ Lglobal (GUIDE)0.03 ± 0.050.17 ± 0.087.88 ± 1.963.34 ± 1.10
+ +Table 2. Quantitative results of GUIDE and the baseline in the generative identity unlearning in a multi-image setting, i.e., using a single image for unlearning and the other images for testing. We used CelebAHQ dataset for this test. + +random scenario, indicating that in cases where a latent code was close to $\bar{w}$ , i.e., as in the random scenario, there was insufficient removal of identity. The effectiveness of employing an extrapolation between the source latent code and the average latent code was evident in such instances. + +The adjacency-aware unlearning loss further enhanced the unlearning an identity. This loss was designed to cover the vicinity of the source latent code, thereby promoting unlearning on the source latent code itself. Finally, the application of the global preservation loss effectively reduced the estimated distribution shift using $\mathrm{FID}_{\mathrm{pre}}$ and $\Delta \mathrm{FID}_{\mathrm{real}}$ . + +Moreover, we conducted a multi-image test in an OOD scenario. In this particular experiment, we introduced additional metric - $\mathrm{ID}_{\mathrm{others}}$ aimed at quantifying ID similarities for the unseen images associated with the source identity. As presented in Table 2, the introduction of the adjacency-aware unlearning loss resulted in a remarkable improvement in $\mathrm{ID}_{\mathrm{others}}$ , emphasizing the effectiveness of this unlearning approach for handling unseen images. + +# 4.3. Ablation Study + +Effect of $d$ in Determination of $w_{t}$ . We conducted an ablation study by comparing target images derived from varied values of $d$ . Setting $d$ to 0 denotes utilizing the $\bar{w}$ as $w_{t}$ in the unlearning process. For $d < 0$ , we designated $w_{t}$ as an interpolated latent code between the $w_{s}$ and the $\bar{w}$ . Conversely, for $d > 0$ , we employed an extrapolated $w_{t}$ , as detailed in Section 3.2. As illustrated in Figure 9, when $d < 0$ , the target image closely aligns with the given source images. However, as $d$ deviates from 0, the quality of the target image rapidly deteriorates, resulting in a pronounced collapse in the distribution of the pre-trained generator. Consequently, the effectiveness of unlearning with such target images diminishes in removing identity from the + +
SourceTarget
d = -10d = -5d = 0d = 30 (Ours)d = 60
Random
FFHQ
d = -70d = -35d = 0d = 30 (Ours)d = 60
+ +Figure 9. Ablation study to figure out the effectiveness of $d$ . We visualized target images corresponding to each source image with different values of $d$ . The target images were generated using target latent codes derived from interpolated latent codes, the average latent code ( $d = 0$ ), or extrapolated latent codes ( $d > 0$ ). Interpolation and extrapolation were carried out between the source and the average latent code. In the case of interpolation, the center between the source and the average latent code was computed. + +
αmaxID (↓)IDothers (↓)
00.1205 ± 0.06030.2754 ± 0.0791
100.0892 ± 0.06200.2123 ± 0.0762
150.0878 ± 0.03750.2094 ± 0.0692
200.0900 ± 0.05380.2105 ± 0.0924
300.0926 ± 0.05610.2111 ± 0.0653
+ +source images. Setting $d$ to 0 might suggest the use of $\bar{w}$ as an effective target for erasing identity. However, our ablation studies indicate that when the source image closely aligns with $\bar{w}$ , the unlearning procedure fails to thoroughly eliminate the identity. Conversely, when $d > 0$ , $w_{t}$ contains a distinct identity compared to the source image while maintaining a consistent distance from $\bar{w}$ . Among the instances where $d > 0$ , our ablation studies reveal that setting $d = 30$ achieves a balanced performance between effective unlearning and preservation of the generation performance of the pre-trained model. + +Effect of $\alpha_{max}$ in $\mathcal{L}_{adj}$ . In Table 3, we scrutinized the effectiveness of the adjacency-aware unlearning loss. To ensure a fair comparison, we employed $\bar{w}$ as $w_{t}$ in this experiment, and we used $\mathcal{L}_{local}$ and $\mathcal{L}_{adj}$ in the unlearning procedure. Rows corresponding to $\alpha_{max} = 0$ represent experimental results without the incorporation of $\mathcal{L}_{adj}$ in the unlearning procedure. The introduction of $\mathcal{L}_{adj}$ resulted in consistent performance gains in $\mathrm{ID}_{\mathrm{others}}$ . This observation highlights the efficacy of considering not only the pair of source and target latent codes but also their surroundings for unlearning the entire identity. + +Table 3. Ablation study to figure out the effectiveness of $\mathcal{L}_{adj}$ and $\alpha_{max}$ . We compared the performance based on how successfully the given identity was erased, using ID and IDothers metric. The row where $\alpha_{max} = 0$ denotes the baseline. We used CelebAHQ dataset in this experiment. + +
\( \mathcal{L}_{local} \)\( \mathcal{L}_{global} \)\( \mathrm{FID}_{\text{pre}}(\downarrow) \)\( \Delta \mathrm{FID}_{\text{real}}(\downarrow) \)
✓9.52 ± 1.534.75 ± 0.89
✓✓4.63 ± 0.431.48 ± 0.29
+ +Table 4. Ablation study to figure our the effectiveness of $\mathcal{L}_{global}$ . We compared how preserved the performance of the pre-trained model through the unlearning process, via $\mathrm{FID}_{\mathrm{pre}}$ and $\Delta \mathrm{FID}_{\mathrm{real}}$ . We used CelebAHQ dataset in this experiment. + +Effect of $\mathcal{L}_{global}$ . To assess the effectiveness of the global preservation loss, a similar experiment was conducted as the previous experiment, i.e., setting $\bar{w}$ as $w_{t}$ . The results are presented in Table 4. The application of $\mathcal{L}_{global}$ demonstrated consistent performance improvements in both $\mathrm{FID}_{\mathrm{pre}}$ and $\Delta \mathrm{FID}_{\mathrm{real}}$ . This suggests that imposing constraints on the generator to maintain its generation performance in latent codes distant from our primary focus is effective in reducing distribution shifts in generative models. + +# 5. Conclusion + +In this paper, we introduced a novel task, referred to as generative identity unlearning, designed to address privacy concerns in pre-trained generative adversarial networks. This task requires thoroughly removing the identity of a single source image from the pre-trained generator. To achieve this, we proposed a new framework, GUIDE (Generative Unlearning for any IDENTITY). To unlearn the single identity, we first defined the target latent code via extrapolation, moving away from the average latent by the pre-defined distance in the direction from the source to the average latent. Using this, our GUIDE successfully unlearned the given identity via Latent Target Unlearning (LTU), which optimized the pre-trained model to preserve the overall generative ability but not to generate the same identity within the local space. Experimental results demonstrated the effectiveness of GUIDE with promising outcomes. We anticipate that our work will be widely applied in research or the industry field, providing users with a sense of freedom from privacy concerns through identity removal. + +# Acknowledgement + +This work was supported by MSIT (Ministry of Science and ICT), Korea, under the ITRC (Information Technology Research Center) support program (IITP-2024-RS-2023-00258649) supervised by the IITP (Institute for Information & Communications Technology Planning & Evaluation), and in part by NRF-2023S1A5A2A21083590, and in part by the IITP grant funded by the Korea Government (MSIT) (Artificial Intelligence Innovation Hub) under Grant 2021-0-02068, and by the IITP grant funded by the Korea government (MSIT) (No.RS-2022-00155911, Artificial Intelligence Convergence Innovation Human Resources Development (Kyung Hee University)). + +# References + +[1] Yuval Alaluf, Or Patashnik, and Daniel Cohen-Or. Restyle: A residual-based stylegan encoder via iterative refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6711–6720, 2021. 4 +[2] Thomas Baumhauer, Pascal Schottle, and Matthias Zeppelinzauer. Machine unlearning: Linear filtration for logit-based classifiers. Machine Learning, 111(9):3203-3226, 2022. 3 +[3] Nicolas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikash Sehwag, Florian Tramer, Borja Balle, Daphne Ippolito, and Eric Wallace. Extracting training data from diffusion models. In Proceedings of the 32nd USENIX Security Symposium (USENIX Security 23), pages 5253-5270, 2023. 2 +[4] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 3, 4, 5 +[5] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1 +[6] Vikram S Chundawat, Ayush K Tarun, Murari Mandal, and Mohan Kankanhalli. Zero-shot machine unlearning. IEEE Transactions on Information Forensics and Security, 2023. 3 +[7] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4690-4699, 2019. 5 +[8] Matt Fredrikson, Somesh Jha, and Thomas Ristenpart. Model inversion attacks that exploit confidence information and basic countermeasures. In Proceedings of the 22nd ACM SIGSAC conference on computer and communications security, pages 1322-1333, 2015. 3 +[9] Rohit Gandikota, Joanna Materzyńska, Jaden Fiotto-Kaufman, and David Bau. Erasing concepts from diffusion models. In Proceedings of the 2023 IEEE International Conference on Computer Vision (ICCV), pages 2426-2436, 2023. 2 +[10] Antonio Ginart, Melody Guan, Gregory Valiant, and James Y Zou. Making ai forget you: Data deletion in machine learning. Advances in Neural Information Processing Systems, 32, 2019. 3 +[11] Aditya Golatkar, Alessandro Achille, and Stefano Soatto. Eternal sunshine of the spotless net: Selective forgetting in deep networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9304–9312, 2020. 2 +[12] Aditya Golatkar, Alessandro Achille, and Stefano Soatto. Forgetting outside the box: Scrubbing deep networks of information accessible from input-output observations. In Pro + +ceedings of the European Conference on Computer Vision (ECCV), pages 383-398. Springer, 2020. 2 +[13] Aditya Golatkar, Alessandro Achille, Avinash Ravichandran, Marzia Polito, and Stefano Soatto. Mixed-privacy forgetting in deep networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 792–801, 2021. +[14] Varun Gupta, Christopher Jung, Seth Neel, Aaron Roth, Saeed Sharifi-Malvajerdi, and Chris Waites. Adaptive machine unlearning. Advances in Neural Information Processing Systems, 34:16319-16330, 2021. 3 +[15] Jamie Hayes, Luca Melis, George Danezis, and Emiliano De Cristofaro. Logan: Membership inference attacks against generative models. arXiv preprint arXiv:1705.07663, 2017. 1 +[16] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in Neural Information Processing Systems, 30, 2017. 6 +[17] Yuge Huang, Yuhan Wang, Ying Tai, Xiaoming Liu, Pengcheng Shen, Shaoxin Li, Jilin Li, and Feiyue Huang. Curricularface: adaptive curriculum learning loss for deep face recognition. In proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5901-5910, 2020. 6 +[18] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability and variation. In Proceedings of the International Conference on Learning Representations (ICLR), 2018. 2, 6 +[19] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 3, 5, 6 +[20] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2 +[21] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021. 1, 2 +[22] Mahdi Khosravy, Kazuaki Nakamura, Yuki Hirose, Naoko Nitta, and Noboru Babaguchi. Model inversion attack by integration of deep generative models: Privacy-sensitive face generation from a face recognition system. IEEE Transactions on Information Forensics and Security, 17:357-372, 2022. 3 +[23] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proceedings of the International Conference on Learning Representations (ICLR), 2015. 6 +[24] Zhifeng Kong and Kamalika Chaudhuri. Data redaction from pre-trained gans. In Proceedings of the 2023 IEEE Conference on Secure and Trustworthy Machine Learning (SaTML), pages 638-677. IEEE, 2023. 3 + +[25] Nupur Kumari, Bingliang Zhang, Sheng-Yu Wang, Eli Shechtman, Richard Zhang, and Jun-Yan Zhu. Ablating concepts in text-to-image diffusion models. In Proceedings of the 2023 IEEE International Conference on Computer Vision (ICCV), pages 22691-22702, 2023. 2 +[26] Ronak Mehta, Sourav Pal, Vikas Singh, and Sathya N Ravi. Deep unlearning via randomized conditionally independent hessian. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10422-10431, 2022. 2 +[27] Jun-Yeong Moon, Keon-Hee Park, Jung Uk Kim, and Gyeong-Moon Park. Online class incremental learning on stochastic blurry task boundary via mask and visual prompt tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11731-11741, 2023. +[28] Saemi Moon, Seunghyuk Cho, and Dongwoo Kim. Feature unlearning for generative models via implicit feedback. arXiv preprint arXiv:2303.05699, 2023. 2, 3 +[29] Seung-Jun Moon and Gyeong-Moon Park. Interstyle: Encoding an interest region for robust stylegan inversion. In Proceedings of the European Conference on Computer Vision (ECCV), pages 460-476. Springer, 2022. 2, 4 +[30] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2085–2094, 2021. 2 +[31] Alexandra Peste, Dan Alistarh, and Christoph H Lampert. Ssse: Efficiently erasing samples from trained machine learning models. corr. arXiv preprint arXiv:2107.03860, 1 (3), 2021. 2 +[32] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In Proceedings of the International Conference on Machine Learning (ICML), pages 8821-8831. PMLR, 2021. 2 +[33] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2 +[34] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2287–2296, 2021. 2 +[35] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on graphics (TOG), 42(1):1-13, 2022. 2 +[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 2 +[37] Juwon Seo, Ji-Su Kang, and Gyeong-Moon Park. Lfs-gan: Lifelong few-shot image generation. In Proceedings of the + +IEEE/CVF International Conference on Computer Vision (ICCV), pages 11356-11366, 2023. 2 +[38] Shawn Shan, Jenna Cryan, Emily Wenger, Haitao Zheng, Rana Hanocka, and Ben Y Zhao. Glaze: Protecting artists from style mimicry by text-to-image models. arXiv preprint arXiv:2302.04222, 2023. 2 +[39] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9243-9252, 2020. 2 +[40] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. Interfacegan: Interpreting the disentangled face representation learned by gans. IEEE transactions on pattern analysis and machine intelligence, 44(4):2004-2018, 2020. 2 +[41] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Diffusion art or digital forgery? investigating data replication in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6048-6058, 2023. 2 +[42] Jingxiang Sun, Xuan Wang, Lizhen Wang, Xiaoyu Li, Yong Zhang, Hongwen Zhang, and Yebin Liu. Next3d: Generative neural texture rasterization for 3d-aware head avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20991-21002, 2023. 2 +[43] Ayush K Tarun, Vikram S Chandawat, Murari Mandal, and Mohan Kankanhalli. Fast yet effective machine unlearning. IEEE Transactions on Neural Networks and Learning Systems, 2023. 3 +[44] Piyush Tiwary, Atri Guha, Subhodip Panda, et al. Adapt then unlearn: Exploiting parameter space semantics for unlearning in generative adversarial networks. arXiv preprint arXiv:2309.14054, 2023. 2, 3 +[45] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40(4): 1-14, 2021. 2 +[46] Tengfei Wang, Yong Zhang, Yanbo Fan, Jue Wang, and Qifeng Chen. High-fidelity gan inversion for image attribute editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11379-11388, 2022. 2 +[47] Yue Wu, Yu Deng, Jiaolong Yang, Fangyun Wei, Qifeng Chen, and Xin Tong. Anifacegan: Animatable 3d-aware face image generation for video avatars. Advances in Neural Information Processing Systems, 35:36188-36201, 2022. 2 +[48] Yuting Xu, Jian Liang, Gengyun Jia, Ziming Yang, Yanhao Zhang, and Ran He. Tall: Thumbnail layout for deepfake video detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22658-22668, 2023. 1 +[49] Zhiyuan Yan, Yong Zhang, Yanbo Fan, and Baoyuan Wu. Ucf: Uncovering common features for generalizable deepfake detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22412-22423, 2023. 1 + +[50] Jaejun Yoo, Youngjung Uh, Sanghyuk Chun, Byeongkyu Kang, and Jung-Woo Ha. Photorealistic style transfer via wavelet transforms. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9036-9045, 2019. 2 +[51] Xiaoyong Yuan, Pan He, Qile Zhu, and Xiaolin Li. Adversarial examples: Attacks and defenses for deep learning. IEEE transactions on neural networks and learning systems, 30 (9):2805-2824, 2019. 3 +[52] Ziyang Yuan, Yiming Zhu, Yu Li, Hongyu Liu, and Chun Yuan. Make encoder great again in 3d gan inversion through geometry and occlusion-aware encoding. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2437-2447, 2023. 2, 3, 4, 5 +[53] Eric Zhang, Kai Wang, Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Forget-me-not: Learning to forget in text-to-image diffusion models. arXiv preprint arXiv:2303.17591, 2023. 2 +[54] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 5 +[55] Xuanmeng Zhang, Zhedong Zheng, Daiheng Gao, Bang Zhang, Yi Yang, and Tat-Seng Chua. Multi-view consistent generative adversarial networks for compositional 3d-aware image synthesis. International Journal of Computer Vision, pages 1-24, 2023. 2 +[56] Xiaoming Zhao, Fangchang Ma, David Guera, Zhile Ren, Alexander G Schwing, and Alex Colburn. Generative multiplane images: Making a 2d gan 3d-aware. In Proceedings of the European Conference on Computer Vision (ECCV), pages 18-35. Springer, 2022. 2 \ No newline at end of file diff --git a/2024/Generative Unlearning for Any Identity/images.zip b/2024/Generative Unlearning for Any Identity/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5b3d90a8ed7d86c98253ebc4b3c1983ffbeb4665 --- /dev/null +++ b/2024/Generative Unlearning for Any Identity/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:41f31a8f5688973d58d39e6f74fdf88d34fc0ba62be522a4ef9acedb5b9466ab +size 631448 diff --git a/2024/Generative Unlearning for Any Identity/layout.json b/2024/Generative Unlearning for Any Identity/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b3491dae130110790ade54185749171505b56205 --- /dev/null +++ b/2024/Generative Unlearning for Any Identity/layout.json @@ -0,0 +1,10474 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 173, + 103, + 421, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 103, + 421, + 121 + ], + "spans": [ + { + "bbox": [ + 173, + 103, + 421, + 121 + ], + "type": "text", + "content": "Generative Unlearning for Any Identity" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 99, + 142, + 162, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 142, + 162, + 156 + ], + "spans": [ + { + "bbox": [ + 99, + 142, + 162, + 156 + ], + "type": "text", + "content": "Juwon " + }, + { + "bbox": [ + 99, + 142, + 162, + 156 + ], + "type": "inline_equation", + "content": "\\mathrm{Seo}^{1*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 187, + 142, + 385, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 142, + 385, + 172 + ], + "spans": [ + { + "bbox": [ + 187, + 142, + 385, + 172 + ], + "type": "text", + "content": "Sung-Hoon Lee" + }, + { + "bbox": [ + 187, + 142, + 385, + 172 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 187, + 142, + 385, + 172 + ], + "type": "text", + "content": " Tae-Young Lee" + }, + { + "bbox": [ + 187, + 142, + 385, + 172 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 187, + 142, + 385, + 172 + ], + "type": "text", + "content": " Gyeong-Moon Park" + }, + { + "bbox": [ + 187, + 142, + 385, + 172 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 410, + 143, + 494, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 410, + 143, + 494, + 158 + ], + "spans": [ + { + "bbox": [ + 410, + 143, + 494, + 158 + ], + "type": "text", + "content": "Seungjun Moon²" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 173, + 184, + 422, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 184, + 422, + 214 + ], + "spans": [ + { + "bbox": [ + 173, + 184, + 422, + 214 + ], + "type": "text", + "content": "1Kyung Hee University, Yongin, Republic of Korea \n2KLleon Tech., Seoul, Republic of Korea" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 156, + 220, + 436, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 220, + 436, + 247 + ], + "spans": [ + { + "bbox": [ + 156, + 220, + 436, + 247 + ], + "type": "text", + "content": "{jwseo001, sunghoonlee961, slcks1, gmpark}@khu.ac.kr seungjunmoon@klleon.io" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 274, + 192, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 274, + 192, + 288 + ], + "spans": [ + { + "bbox": [ + 143, + 274, + 192, + 288 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 297, + 290, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 297, + 290, + 620 + ], + "spans": [ + { + "bbox": [ + 47, + 297, + 290, + 620 + ], + "type": "text", + "content": "Recent advances in generative models trained on large-scale datasets have made it possible to synthesize high-quality samples across various domains. Moreover, the emergence of strong inversion networks enables not only a reconstruction of real-world images but also the modification of attributes through various editing methods. However, in certain domains related to privacy issues, e.g., human faces, advanced generative models along with strong inversion methods can lead to potential misuses. In this paper, we propose an essential yet under-explored task called generative identity unlearning, which steers the model not to generate an image of specific identity. In the generative identity unlearning, we target the following objectives: (i) preventing the generation of images with a certain identity, and (ii) preserving the overall quality of the generative model. To satisfy these goals, we propose a novel framework, Generative Unlearning for Any IDENTITY (GUIDE), which prevents the reconstruction of a specific identity by unlearning the generator with only a single image. GUIDE consists of two parts: (i) finding a target point for optimization that un-identifies the source latent code and (ii) novel loss functions that facilitate the unlearning procedure while less affecting the learned distribution. Our extensive experiments demonstrate that our proposed method achieves state-of-the-art performance in the generative machine unlearning task. The code is available at https://github.com/KHU-AGI/GUIDE." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 642, + 128, + 654 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 642, + 128, + 654 + ], + "spans": [ + { + "bbox": [ + 47, + 642, + 128, + 654 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 662, + 288, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 662, + 288, + 687 + ], + "spans": [ + { + "bbox": [ + 47, + 662, + 288, + 687 + ], + "type": "text", + "content": "Recently, 2D or 3D Generative Adversarial Networks (GANs) [4, 19-21] pre-trained on large datasets, e.g., FFHQ" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 319, + 276, + 535, + 449 + ], + "blocks": [ + { + "bbox": [ + 319, + 276, + 535, + 449 + ], + "lines": [ + { + "bbox": [ + 319, + 276, + 535, + 449 + ], + "spans": [ + { + "bbox": [ + 319, + 276, + 535, + 449 + ], + "type": "image", + "image_path": "30461510649baa1c9b78705c4fac198e76a80a95d1ce761366dd47823862bd32.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 456, + 547, + 513 + ], + "lines": [ + { + "bbox": [ + 305, + 456, + 547, + 513 + ], + "spans": [ + { + "bbox": [ + 305, + 456, + 547, + 513 + ], + "type": "text", + "content": "Figure 1. Given a single source image containing a specific identity, we remove that identity from the pre-trained 3D generative adversarial network (e.g., EG3D [4]). Our method effectively unlearns identity even from in-the-wild images where the source image is absent in the pre-training dataset." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 533, + 546, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 546, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 546, + 641 + ], + "type": "text", + "content": "[19] or AFHQ [5], have drawn substantial attention due to their remarkable generation performance and highly disentangled representation space. However, their advancements have raised privacy concerns [15], especially regarding the potential misuse of generative models to represent and exploit individual identities. For instance, deepfakes [48, 49] can create very believable images or videos of people in made-up situations, causing major concerns about ethics and privacy." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "type": "text", + "content": "To alleviate privacy issues in generative models, machine unlearning task has been actively studied. Machine unlearning involves the process of selectively removing specific knowledge or erasing the influence of certain data from the training dataset of pre-trained models. It is beneficial especially when the data are harmful, private, or biased" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 124, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 124, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 124, + 703 + ], + "type": "text", + "content": "*Equal contribution" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "text", + "content": "† Corresponding author" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "text", + "content": "9151" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": "[11, 12, 26, 31]. Despite a focus on discriminative tasks in most machine unlearning research, a few studies have ventured into generative models, attempting to erase high-level concepts such as socially inappropriate content or artistic styles that present copyright challenges [9, 25]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 134, + 286, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 134, + 286, + 242 + ], + "spans": [ + { + "bbox": [ + 46, + 134, + 286, + 242 + ], + "type": "text", + "content": "Nevertheless, generative models still exhibit ongoing privacy issues. Even if an identity of someone is not used in the pre-training of the generative models, it can be easily reconstructed in the pre-trained models via GAN inversion models [29, 34, 35, 45, 46, 52]. Furthermore, the reconstructed image can be manipulated or edited easily via image editing methods [30, 39, 40]. To prevent potential exploits of an identity, it is necessary to erase a certain identity from the pre-trained generative models." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 243, + 286, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 243, + 286, + 363 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 286, + 363 + ], + "type": "text", + "content": "To consider the above issue, we introduce an essential task of unlearning any identity from the pre-trained 2D or 3D GANs [4, 20], called generative identity unlearning. Unlike typical machine unlearning tasks, which focus on unlearning the training samples our generative identity unlearning task unlearns any identity on pre-trained GANs, even if it was not shown during the pre-training. Our goal is to remove the whole identity associated with a given single image from the generator while minimally impacting the overall performance of the pre-trained model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 365, + 286, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 286, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 286, + 509 + ], + "type": "text", + "content": "To achieve our goal, we propose a novel generative unlearning framework, Generative Unlearning for Any IDEntity, named GUIDE. GUIDE replaces the source identity with an anonymous target identity, erasing the original identity effectively. To this end, we propose a new exploration method to determine an effective target latent code, called Un-Identifying Face On latent space (UFO). UFO utilizes the GAN inversion method [52] to embed the given identity into the source latent, and then decides the target latent using both the source and the average latent codes. We empirically find that the proposed UFO can identify the promising target to erase any given source identity robustly." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 510, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 286, + 713 + ], + "type": "text", + "content": "Given the source and target latent code, we update the generator to shift from the source identity to the target identity. To this end, we propose three novel loss functions: (i) local unlearning loss, (ii) adjacency-aware unlearning loss, and (iii) global preservation loss. (i) guides our model directly shifting the source identity to the target identity. (ii) utilizes other latent codes adjacent to the source and target latent codes to effectively unlearn the entire identity from a single image. To minimize side effects from the unlearning process, (iii) additionally regularizes the generator to retain generation performance for latent codes relatively far from the source and target latent codes. Through comprehensive experiments on diverse identities, including Random, InD, and OOD, we confirm that GUIDE can successfully remove the identity of the source image from the pre-trained generative model, and shows qualitatively and quantitatively superior performances." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 72, + 517, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 72, + 517, + 83 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 517, + 83 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 85, + 545, + 371 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 306, + 85, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 85, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 306, + 85, + 545, + 156 + ], + "type": "text", + "content": "- For the first time, we propose a novel task, generative identity unlearning, which tackles machine unlearning in generative models in the aspect of privacy protection. In our task, we aim to prevent the pre-trained generative models from synthesizing the given identity by utilizing only a single image." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 156, + 545, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 156, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 306, + 156, + 545, + 239 + ], + "type": "text", + "content": "- For the effective and robust elimination of the identity, we propose a novel method - Un-Identifying Face On Latent Space (UFO). We configure the unlearning procedure by formulating how to represent and shift the identity in the latent space. We find that setting the extrapolated latent code between the source and average latent codes as an optimization target facilitates the unlearning procedure." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 240, + 545, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 240, + 545, + 299 + ], + "spans": [ + { + "bbox": [ + 306, + 240, + 545, + 299 + ], + "type": "text", + "content": "- We propose three loss functions - local unlearning loss, adjacency-aware unlearning loss, and global preservation loss to effectively unlearn the identity from the pretrained model while less affecting the generation performance on other identity." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 300, + 545, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 300, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 306, + 300, + 545, + 371 + ], + "type": "text", + "content": "- We show that our proposed framework, GUIDE achieves state-of-the-art performance both qualitatively and quantitatively, through extensive experiments. We demonstrate that GUIDE can remove the specific identity successfully in the generative models while minimizing the negative effect on other identities." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 382, + 391, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 382, + 391, + 395 + ], + "spans": [ + { + "bbox": [ + 306, + 382, + 391, + 395 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 403, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 545, + 510 + ], + "type": "text", + "content": "Generative Models and Privacy Issue. In image synthesis field, GAN-based generative models have achieved remarkable performance not only in 2D [18-21, 32, 33, 36, 37] but also in 3D [4, 42, 47, 55, 56] domain. The application of various image editing methods [30, 40, 50] to strong generative models, people can easily generate edited images of specific individuals and various artistic styles [38, 41], as well as extract copyrighted content [3] without the permission of the individual or the original creator." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "Recently, with the rise of the importance of AI ethics, several works have addressed this issue [9, 25, 28, 44, 53]. ESD [9] erases specific visual concepts from diffusion model by using negative guidance about the undesired concepts. Kumari et al. [25] modifies the conditional distribution of the model a specific target concepts to match the anchor concept. Forget-Me-Not [53] fine-tunes U-Net to minimize each of the intermediate attention associated with the target concepts to remove. Additional works in GANs [28, 44] focus on unlearning specific features, e.g., \"Bang\", \"Hat\" or \"Beard\" rather than forgetting specific identity. The above methods primarily concentrate on the elimination of specific concepts or high-level features. In other words, these cannot preclude models from generating specific individuals while maintaining the generation performance of realistic human faces. Unlike the existing works, our work targets to unlearn only specific individuals without" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9152" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 258, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 258, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 258, + 85 + ], + "type": "text", + "content": "shifting the overall distribution of generated images." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 100, + 286, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 100, + 286, + 182 + ], + "spans": [ + { + "bbox": [ + 46, + 100, + 286, + 182 + ], + "type": "text", + "content": "Machine Unlearning. Machine unlearning aims to selectively forget specific acquired knowledge or diminish the impact of certain training data subsets on a trained model. Since previous research [8, 22, 51] shows that machine learning models might accidentally share private information when faced with certain attacks or inputs, machine unlearning becomes crucial." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 183, + 287, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 183, + 287, + 388 + ], + "spans": [ + { + "bbox": [ + 46, + 183, + 287, + 388 + ], + "type": "text", + "content": "While previous machine unlearning is mainly focused on supervised learning tasks [2, 6, 10-14, 43], the interest in unlearning techniques within unsupervised learning, i.e., generative models, is growing [24, 28, 44]. However, most of the existing methods need full dataset access for retraining, which is hard to acquire and computationally expensive [2, 10, 14, 43]. For example, Kong and Chaudhuri [24] utilizes data reduction and augmentation algorithms, which requires a full training dataset. Despite the existence of a feature unlearning model [28] which does not need full dataset access, unlearning only an individual feature is not enough to forget a whole specific identity. To this end, we propose an algorithm that enables forgetting the specific identity only with a single image. Furthermore, our approach distinguishes itself from existing research by applying unlearning to unseen images, enabling the erasure of specific identities without prior exposure to those images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 399, + 103, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 399, + 103, + 411 + ], + "spans": [ + { + "bbox": [ + 47, + 399, + 103, + 411 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 415, + 287, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 287, + 513 + ], + "type": "text", + "content": "Firstly, in Section 3.1, as shown in Figure 2, we introduce the problem we aim to address, named generative identity unlearning. In Section 3.2, we introduce un-identifying face on latent space, which designates an appropriate target latent for unlearning. Then, in Section 3.3, we introduce latent target unlearning, along with our proposed novel losses to unlearn the generator. The total overview of our method is illustrated in Figure 3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 520, + 171, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 520, + 171, + 532 + ], + "spans": [ + { + "bbox": [ + 47, + 520, + 171, + 532 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 539, + 287, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 539, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 46, + 539, + 287, + 611 + ], + "type": "text", + "content": "Given a set of images " + }, + { + "bbox": [ + 46, + 539, + 287, + 611 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 539, + 287, + 611 + ], + "type": "text", + "content": " depicting a specific identity, we randomly select a single source image " + }, + { + "bbox": [ + 46, + 539, + 287, + 611 + ], + "type": "inline_equation", + "content": "x_{u} \\in \\mathbf{x}" + }, + { + "bbox": [ + 46, + 539, + 287, + 611 + ], + "type": "text", + "content": " as an exemplar of the identity. Initially, using off-the-shelf inversion network [52] " + }, + { + "bbox": [ + 46, + 539, + 287, + 611 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 46, + 539, + 287, + 611 + ], + "type": "text", + "content": " corresponding to the unconditional generator, i.e., EG3D [4], we embed " + }, + { + "bbox": [ + 46, + 539, + 287, + 611 + ], + "type": "inline_equation", + "content": "x_{u}" + }, + { + "bbox": [ + 46, + 539, + 287, + 611 + ], + "type": "text", + "content": " to the source latent code " + }, + { + "bbox": [ + 46, + 539, + 287, + 611 + ], + "type": "inline_equation", + "content": "w_{u}" + }, + { + "bbox": [ + 46, + 539, + 287, + 611 + ], + "type": "text", + "content": " in the latent space of EG3D:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 619, + 287, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 619, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 137, + 619, + 287, + 632 + ], + "type": "interline_equation", + "content": "w _ {u} = E \\left(x _ {u}\\right). \\tag {1}", + "image_path": "68049b8d6b677d6022d5a73c280918936f026f3ff8d44fb5550404dfdd096dfe.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 641, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 689 + ], + "type": "text", + "content": "Since EG3D [4] consists of the mapping network " + }, + { + "bbox": [ + 46, + 641, + 287, + 689 + ], + "type": "inline_equation", + "content": "Map(\\cdot)" + }, + { + "bbox": [ + 46, + 641, + 287, + 689 + ], + "type": "text", + "content": ", StyleGAN2 [19] backbone " + }, + { + "bbox": [ + 46, + 641, + 287, + 689 + ], + "type": "inline_equation", + "content": "G(\\cdot)" + }, + { + "bbox": [ + 46, + 641, + 287, + 689 + ], + "type": "text", + "content": " and the neural renderer with a super-resolution module " + }, + { + "bbox": [ + 46, + 641, + 287, + 689 + ], + "type": "inline_equation", + "content": "R(\\cdot)" + }, + { + "bbox": [ + 46, + 641, + 287, + 689 + ], + "type": "text", + "content": ", we can denote the reconstructed image " + }, + { + "bbox": [ + 46, + 641, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\hat{x}" + }, + { + "bbox": [ + 46, + 641, + 287, + 689 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 641, + 287, + 689 + ], + "type": "inline_equation", + "content": "w_{u}" + }, + { + "bbox": [ + 46, + 641, + 287, + 689 + ], + "type": "text", + "content": " as following:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 128, + 697, + 287, + 710 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 697, + 287, + 710 + ], + "spans": [ + { + "bbox": [ + 128, + 697, + 287, + 710 + ], + "type": "interline_equation", + "content": "\\hat {x} = R \\left(G \\left(w _ {u}\\right); c\\right), \\tag {2}", + "image_path": "ad7fdbab5971278034442637cea25cff12e093d2f2fe35eb4d7af5b5b39f831f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 315, + 72, + 538, + 194 + ], + "blocks": [ + { + "bbox": [ + 315, + 72, + 538, + 194 + ], + "lines": [ + { + "bbox": [ + 315, + 72, + 538, + 194 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 538, + 194 + ], + "type": "image", + "image_path": "cde3a6799f641e5b379ca23096bc13eef2febd497281e44fb2fad14724196e95.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 202, + 545, + 281 + ], + "lines": [ + { + "bbox": [ + 304, + 202, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 304, + 202, + 545, + 281 + ], + "type": "text", + "content": "Figure 2. An illustration of generative identity unlearning. Upon GUIDE, the identity of the image generated from " + }, + { + "bbox": [ + 304, + 202, + 545, + 281 + ], + "type": "inline_equation", + "content": "w_{u}" + }, + { + "bbox": [ + 304, + 202, + 545, + 281 + ], + "type": "text", + "content": ", i.e., inversion of the source image " + }, + { + "bbox": [ + 304, + 202, + 545, + 281 + ], + "type": "inline_equation", + "content": "x_{u}" + }, + { + "bbox": [ + 304, + 202, + 545, + 281 + ], + "type": "text", + "content": " by inversion network " + }, + { + "bbox": [ + 304, + 202, + 545, + 281 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 202, + 545, + 281 + ], + "type": "text", + "content": ", should exhibit a distinct identity when passed through the pre-trained generator " + }, + { + "bbox": [ + 304, + 202, + 545, + 281 + ], + "type": "inline_equation", + "content": "G_{s}" + }, + { + "bbox": [ + 304, + 202, + 545, + 281 + ], + "type": "text", + "content": " compared to the unlearned generator " + }, + { + "bbox": [ + 304, + 202, + 545, + 281 + ], + "type": "inline_equation", + "content": "G_{u}" + }, + { + "bbox": [ + 304, + 202, + 545, + 281 + ], + "type": "text", + "content": ". Furthermore, other images " + }, + { + "bbox": [ + 304, + 202, + 545, + 281 + ], + "type": "inline_equation", + "content": "x_{o}" + }, + { + "bbox": [ + 304, + 202, + 545, + 281 + ], + "type": "text", + "content": ", not used in unlearning but sharing the same identity with " + }, + { + "bbox": [ + 304, + 202, + 545, + 281 + ], + "type": "inline_equation", + "content": "x_{u}" + }, + { + "bbox": [ + 304, + 202, + 545, + 281 + ], + "type": "text", + "content": ", also should vary an identity through GUIDE." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 302, + 545, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 302, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 304, + 302, + 545, + 398 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 302, + 545, + 398 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 302, + 545, + 398 + ], + "type": "text", + "content": " denotes camera poses. For convenience, we omit the explicit notation of camera poses in this paper, i.e., " + }, + { + "bbox": [ + 304, + 302, + 545, + 398 + ], + "type": "inline_equation", + "content": "\\hat{x} = (R \\circ G)(w_u)" + }, + { + "bbox": [ + 304, + 302, + 545, + 398 + ], + "type": "text", + "content": ". We target to derive an unlearned " + }, + { + "bbox": [ + 304, + 302, + 545, + 398 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 302, + 545, + 398 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 304, + 302, + 545, + 398 + ], + "type": "inline_equation", + "content": "G_u" + }, + { + "bbox": [ + 304, + 302, + 545, + 398 + ], + "type": "text", + "content": ", from the pre-trained EG3D generator " + }, + { + "bbox": [ + 304, + 302, + 545, + 398 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 302, + 545, + 398 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 304, + 302, + 545, + 398 + ], + "type": "inline_equation", + "content": "G_s" + }, + { + "bbox": [ + 304, + 302, + 545, + 398 + ], + "type": "text", + "content": ", while fixing " + }, + { + "bbox": [ + 304, + 302, + 545, + 398 + ], + "type": "inline_equation", + "content": "Map" + }, + { + "bbox": [ + 304, + 302, + 545, + 398 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 302, + 545, + 398 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 302, + 545, + 398 + ], + "type": "text", + "content": ". With proper unlearning, an image generated by unlearned EG3D using " + }, + { + "bbox": [ + 304, + 302, + 545, + 398 + ], + "type": "inline_equation", + "content": "w_u" + }, + { + "bbox": [ + 304, + 302, + 545, + 398 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 304, + 302, + 545, + 398 + ], + "type": "inline_equation", + "content": "\\hat{x}_u = (R \\circ G_u)(w_u)" + }, + { + "bbox": [ + 304, + 302, + 545, + 398 + ], + "type": "text", + "content": " should have a distinct identity from the image generated by original EG3D using " + }, + { + "bbox": [ + 304, + 302, + 545, + 398 + ], + "type": "inline_equation", + "content": "w_u" + }, + { + "bbox": [ + 304, + 302, + 545, + 398 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 304, + 302, + 545, + 398 + ], + "type": "inline_equation", + "content": "(R \\circ G_s)(w_u)" + }, + { + "bbox": [ + 304, + 302, + 545, + 398 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 399, + 546, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 546, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 546, + 555 + ], + "type": "text", + "content": "In our task formulation, two considerations are paramount. First, we aim to eliminate the entire identity from the generator only utilizing a single image. To validate this, we incorporate other multiple images " + }, + { + "bbox": [ + 304, + 399, + 546, + 555 + ], + "type": "inline_equation", + "content": "x_{o} \\in \\mathbf{x}" + }, + { + "bbox": [ + 304, + 399, + 546, + 555 + ], + "type": "text", + "content": " for testing and its corresponding latent code " + }, + { + "bbox": [ + 304, + 399, + 546, + 555 + ], + "type": "inline_equation", + "content": "w_{o} = E(x_{o})" + }, + { + "bbox": [ + 304, + 399, + 546, + 555 + ], + "type": "text", + "content": ". By utilizing " + }, + { + "bbox": [ + 304, + 399, + 546, + 555 + ], + "type": "inline_equation", + "content": "x_{o}" + }, + { + "bbox": [ + 304, + 399, + 546, + 555 + ], + "type": "text", + "content": ", we can verify whether " + }, + { + "bbox": [ + 304, + 399, + 546, + 555 + ], + "type": "inline_equation", + "content": "G_{u}" + }, + { + "bbox": [ + 304, + 399, + 546, + 555 + ], + "type": "text", + "content": " has successfully unlearned the identity as a whole, rather than just unlearning the specific image " + }, + { + "bbox": [ + 304, + 399, + 546, + 555 + ], + "type": "inline_equation", + "content": "x_{u}" + }, + { + "bbox": [ + 304, + 399, + 546, + 555 + ], + "type": "text", + "content": ". Second, we strive to maintain the generation performance of the pre-trained model. To assess this, we sample multiple images from fixed latent codes using both the unlearned and pre-trained generators. We then estimate the distribution shift between the images generated before and after the unlearning process." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 563, + 504, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 504, + 575 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 504, + 575 + ], + "type": "text", + "content": "3.2. Un-Identifying Face On Latent Space" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "The successfully unlearned model should not generate the image with the identity of " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": ", even when " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "w_{u}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " is used as a latent. Consequently, we initiate our approach to manipulate " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{x}_u" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " to be another image rather than the image with identity in " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " by unlearning " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": ". To design the objective function for unlearning, we first need to establish the objective for unlearning, which involves defining the target image " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{x}_t" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": ", derived from " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "w_t" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": ", which the unlearned image " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{x}_u" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": ", derived from " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "w_{u}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": ", should mimic after unlearning. While there exist various options for setting " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{x}_t" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": ", e.g., a random face or even a non-human image, we choose the mean face generated by" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9153" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 73, + 547, + 250 + ], + "blocks": [ + { + "bbox": [ + 51, + 73, + 547, + 250 + ], + "lines": [ + { + "bbox": [ + 51, + 73, + 547, + 250 + ], + "spans": [ + { + "bbox": [ + 51, + 73, + 547, + 250 + ], + "type": "image", + "image_path": "ffa2618d4725731f575a1660beb56ea64b94f932c30b6ff96b23cdc5937bf14f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 259, + 547, + 350 + ], + "lines": [ + { + "bbox": [ + 46, + 259, + 547, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 259, + 547, + 350 + ], + "type": "text", + "content": "Figure 3. An overview of GUIDE. Starting with a source image, we employ a GAN inversion network " + }, + { + "bbox": [ + 46, + 259, + 547, + 350 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 46, + 259, + 547, + 350 + ], + "type": "text", + "content": ", specifically GOAE [52], to embed this image into the latent space of a pre-trained generative model, namely EG3D [4], obtaining the source latent code " + }, + { + "bbox": [ + 46, + 259, + 547, + 350 + ], + "type": "inline_equation", + "content": "w_{u}" + }, + { + "bbox": [ + 46, + 259, + 547, + 350 + ], + "type": "text", + "content": ". The target latent code " + }, + { + "bbox": [ + 46, + 259, + 547, + 350 + ], + "type": "inline_equation", + "content": "w_{t}" + }, + { + "bbox": [ + 46, + 259, + 547, + 350 + ], + "type": "text", + "content": " is designated through the UFO process. To facilitate identity removal in " + }, + { + "bbox": [ + 46, + 259, + 547, + 350 + ], + "type": "inline_equation", + "content": "w_{u}" + }, + { + "bbox": [ + 46, + 259, + 547, + 350 + ], + "type": "text", + "content": ", we shift its identity to match that of " + }, + { + "bbox": [ + 46, + 259, + 547, + 350 + ], + "type": "inline_equation", + "content": "w_{t}" + }, + { + "bbox": [ + 46, + 259, + 547, + 350 + ], + "type": "text", + "content": " with our Latent Target Unlearning (LTU) process. Three loss functions of LTU are designed for this purpose: (i) The generator is optimized to produce an image from the source latent code, denoted as " + }, + { + "bbox": [ + 46, + 259, + 547, + 350 + ], + "type": "inline_equation", + "content": "(R \\circ G_{u})(w_{u})" + }, + { + "bbox": [ + 46, + 259, + 547, + 350 + ], + "type": "text", + "content": ", that is similar to the image from the target latent code, represented as " + }, + { + "bbox": [ + 46, + 259, + 547, + 350 + ], + "type": "inline_equation", + "content": "(R \\circ G_{s})(w_{t})" + }, + { + "bbox": [ + 46, + 259, + 547, + 350 + ], + "type": "text", + "content": ". (ii) To achieve unlearning across the entire identity, we consider latent codes near both the source and target latent codes, denoted as " + }, + { + "bbox": [ + 46, + 259, + 547, + 350 + ], + "type": "inline_equation", + "content": "w_{u,a}" + }, + { + "bbox": [ + 46, + 259, + 547, + 350 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 259, + 547, + 350 + ], + "type": "inline_equation", + "content": "w_{t,a}" + }, + { + "bbox": [ + 46, + 259, + 547, + 350 + ], + "type": "text", + "content": ", respectively. (iii) To prevent model corruption during the unlearning process, we additionally sample latent codes from a random noise vector, represented as " + }, + { + "bbox": [ + 46, + 259, + 547, + 350 + ], + "type": "inline_equation", + "content": "w_{r,g}" + }, + { + "bbox": [ + 46, + 259, + 547, + 350 + ], + "type": "text", + "content": ", and optimize " + }, + { + "bbox": [ + 46, + 259, + 547, + 350 + ], + "type": "inline_equation", + "content": "G_{u}" + }, + { + "bbox": [ + 46, + 259, + 547, + 350 + ], + "type": "text", + "content": " to preserve its generation ability on " + }, + { + "bbox": [ + 46, + 259, + 547, + 350 + ], + "type": "inline_equation", + "content": "w_{r,g}" + }, + { + "bbox": [ + 46, + 259, + 547, + 350 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 73, + 370, + 262, + 459 + ], + "blocks": [ + { + "bbox": [ + 73, + 370, + 262, + 459 + ], + "lines": [ + { + "bbox": [ + 73, + 370, + 262, + 459 + ], + "spans": [ + { + "bbox": [ + 73, + 370, + 262, + 459 + ], + "type": "image", + "image_path": "a498b047462a15846977f2f41cc34875a1ab9f8e3342a8b5deab74778be983cf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 468, + 288, + 523 + ], + "lines": [ + { + "bbox": [ + 46, + 468, + 288, + 523 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 288, + 523 + ], + "type": "text", + "content": "Figure 4. An illustration of Un-identifying Face On Latent Space (UFO). We define the identity of the source latent code by subtract it from the average latent code. We set the target latent code for our unlearning process by measuring an extrapolation between the source and average latent code with a fixed distance " + }, + { + "bbox": [ + 46, + 468, + 288, + 523 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 468, + 288, + 523 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 533, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 605 + ], + "type": "text", + "content": "the mean latent " + }, + { + "bbox": [ + 46, + 533, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\bar{w}" + }, + { + "bbox": [ + 46, + 533, + 287, + 605 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 46, + 533, + 287, + 605 + ], + "type": "inline_equation", + "content": "Map" + }, + { + "bbox": [ + 46, + 533, + 287, + 605 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 46, + 533, + 287, + 605 + ], + "type": "inline_equation", + "content": "(R \\circ G_s)(\\bar{w})" + }, + { + "bbox": [ + 46, + 533, + 287, + 605 + ], + "type": "text", + "content": ", as " + }, + { + "bbox": [ + 46, + 533, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\hat{x}_t" + }, + { + "bbox": [ + 46, + 533, + 287, + 605 + ], + "type": "text", + "content": ". This setting is inspired by inversion methods [1, 29] which apply the identity to the mean face through inversion stages. Since our goal is to de-identify the image, we argue that it is intuitive and reasonable to send the image back to the mean face, which is opposite to the inversion." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "However, this simple approach might be problematic when " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "w_{u}" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " is close to " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\bar{w}" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": ", where the unlearned image might still resemble the original identity. To this end, we propose a novel method named Un-identifying Face On latent space (UFO), which can set " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\hat{x}_t" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " robustly regardless of the distance between " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "w_{u}" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\bar{w}" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": ", as shown in Figure 4. The following are the processes of UFO: First, in the " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "de" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "-identification process, we extract the identity latent " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "w_{id} = w_{u} - \\bar{w}" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " by subtracting " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\bar{w}" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "w_{u}" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": ". Then we propose a process of en-identification," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 369, + 545, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 369, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 305, + 369, + 545, + 441 + ], + "type": "text", + "content": "setting the target in the opposite direction of the existing " + }, + { + "bbox": [ + 305, + 369, + 545, + 441 + ], + "type": "inline_equation", + "content": "w_{id}" + }, + { + "bbox": [ + 305, + 369, + 545, + 441 + ], + "type": "text", + "content": " to foster a more pronounced change in identity, creating an entirely new identity. In other words, we chose " + }, + { + "bbox": [ + 305, + 369, + 545, + 441 + ], + "type": "inline_equation", + "content": "\\bar{w}" + }, + { + "bbox": [ + 305, + 369, + 545, + 441 + ], + "type": "text", + "content": " as the stopover point and established the final target point at the extrapolation of " + }, + { + "bbox": [ + 305, + 369, + 545, + 441 + ], + "type": "inline_equation", + "content": "w_{u}" + }, + { + "bbox": [ + 305, + 369, + 545, + 441 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 369, + 545, + 441 + ], + "type": "inline_equation", + "content": "\\bar{w}" + }, + { + "bbox": [ + 305, + 369, + 545, + 441 + ], + "type": "text", + "content": ". The finalized target point can be expressed as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 378, + 445, + 545, + 469 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 445, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 378, + 445, + 545, + 469 + ], + "type": "interline_equation", + "content": "w _ {t} = \\bar {w} - d \\cdot \\frac {w _ {i d}}{\\| w _ {i d} \\| _ {2}}, \\tag {3}", + "image_path": "c8430f0e55b2239377d0016e68973c6b08ef3484b6571acc4b8d47d19b79f40f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 478, + 545, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 478, + 545, + 562 + ], + "spans": [ + { + "bbox": [ + 304, + 478, + 545, + 562 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 478, + 545, + 562 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 478, + 545, + 562 + ], + "type": "text", + "content": " is defined as the distance that balances un-identification with preservation of the source distribution, as determined by our empirical results. Finally, we can set " + }, + { + "bbox": [ + 304, + 478, + 545, + 562 + ], + "type": "inline_equation", + "content": "\\hat{x}_t = (R\\circ G_s)(w_t)" + }, + { + "bbox": [ + 304, + 478, + 545, + 562 + ], + "type": "text", + "content": ". We empirically demonstrate that whether the given identity is in close proximity to the average latent code within the latent space or not, UFO sets the desirable target latent code for the generative unlearning." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 570, + 449, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 570, + 449, + 583 + ], + "spans": [ + { + "bbox": [ + 306, + 570, + 449, + 583 + ], + "type": "text", + "content": "3.3. Latent Target Unlearning" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 589, + 545, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 589, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 304, + 589, + 545, + 673 + ], + "type": "text", + "content": "After setting " + }, + { + "bbox": [ + 304, + 589, + 545, + 673 + ], + "type": "inline_equation", + "content": "\\hat{x}_t" + }, + { + "bbox": [ + 304, + 589, + 545, + 673 + ], + "type": "text", + "content": ", we need to robustly design the unlearning process that can effectively make " + }, + { + "bbox": [ + 304, + 589, + 545, + 673 + ], + "type": "inline_equation", + "content": "\\hat{x}_u" + }, + { + "bbox": [ + 304, + 589, + 545, + 673 + ], + "type": "text", + "content": " to be similar to " + }, + { + "bbox": [ + 304, + 589, + 545, + 673 + ], + "type": "inline_equation", + "content": "\\hat{x}_t" + }, + { + "bbox": [ + 304, + 589, + 545, + 673 + ], + "type": "text", + "content": ", while maintaining the generation performance of " + }, + { + "bbox": [ + 304, + 589, + 545, + 673 + ], + "type": "inline_equation", + "content": "G_u" + }, + { + "bbox": [ + 304, + 589, + 545, + 673 + ], + "type": "text", + "content": ". We coin this process as Latent Target Unlearning (LTU), which targets to unlearn images from the specific latent while keeping generation results from other latent codes. LTU utilizes the following three losses to achieve this goal:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "Local Unlearning Loss. To force " + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\hat{x}_u" + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": " to be " + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\hat{x}_t" + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": ", we use the widely-used reconstruction losses, i.e., Euclidean loss " + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_2" + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9154" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 71, + 289, + 182 + ], + "blocks": [ + { + "bbox": [ + 53, + 71, + 289, + 182 + ], + "lines": [ + { + "bbox": [ + 53, + 71, + 289, + 182 + ], + "spans": [ + { + "bbox": [ + 53, + 71, + 289, + 182 + ], + "type": "image", + "image_path": "9db8382f578fb7d11023da21d017b3da352967bb5ef2e563a6f054bdafcc50d1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 191, + 289, + 270 + ], + "lines": [ + { + "bbox": [ + 46, + 191, + 289, + 270 + ], + "spans": [ + { + "bbox": [ + 46, + 191, + 289, + 270 + ], + "type": "text", + "content": "Figure 5. An illustration of determining latent codes near a latent code " + }, + { + "bbox": [ + 46, + 191, + 289, + 270 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 191, + 289, + 270 + ], + "type": "text", + "content": " in adjacency-aware unlearning loss. We first sample a latent code " + }, + { + "bbox": [ + 46, + 191, + 289, + 270 + ], + "type": "inline_equation", + "content": "w_{r,a}" + }, + { + "bbox": [ + 46, + 191, + 289, + 270 + ], + "type": "text", + "content": " which is derived from a random noise vector " + }, + { + "bbox": [ + 46, + 191, + 289, + 270 + ], + "type": "inline_equation", + "content": "z_{r,a}" + }, + { + "bbox": [ + 46, + 191, + 289, + 270 + ], + "type": "text", + "content": " via the mapping network " + }, + { + "bbox": [ + 46, + 191, + 289, + 270 + ], + "type": "inline_equation", + "content": "\\text{Map}(\\cdot)" + }, + { + "bbox": [ + 46, + 191, + 289, + 270 + ], + "type": "text", + "content": ", i.e. " + }, + { + "bbox": [ + 46, + 191, + 289, + 270 + ], + "type": "inline_equation", + "content": "w_{r,a} = \\text{Map}(z_{r,a})" + }, + { + "bbox": [ + 46, + 191, + 289, + 270 + ], + "type": "text", + "content": ". Next, we compute the direction between " + }, + { + "bbox": [ + 46, + 191, + 289, + 270 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 191, + 289, + 270 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 191, + 289, + 270 + ], + "type": "inline_equation", + "content": "w_{r,a}" + }, + { + "bbox": [ + 46, + 191, + 289, + 270 + ], + "type": "text", + "content": ", and we scale it to fall within range between 0 and " + }, + { + "bbox": [ + 46, + 191, + 289, + 270 + ], + "type": "inline_equation", + "content": "\\alpha_{max}" + }, + { + "bbox": [ + 46, + 191, + 289, + 270 + ], + "type": "text", + "content": ". This yields the distance vector " + }, + { + "bbox": [ + 46, + 191, + 289, + 270 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 46, + 191, + 289, + 270 + ], + "type": "text", + "content": " to compute the adjacent latent code " + }, + { + "bbox": [ + 46, + 191, + 289, + 270 + ], + "type": "inline_equation", + "content": "w_{a} = w + \\Delta" + }, + { + "bbox": [ + 46, + 191, + 289, + 270 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 286, + 287, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 286, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 47, + 286, + 287, + 335 + ], + "type": "text", + "content": "perceptual loss " + }, + { + "bbox": [ + 47, + 286, + 287, + 335 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{per}" + }, + { + "bbox": [ + 47, + 286, + 287, + 335 + ], + "type": "text", + "content": " [54], and identity loss " + }, + { + "bbox": [ + 47, + 286, + 287, + 335 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{id}" + }, + { + "bbox": [ + 47, + 286, + 287, + 335 + ], + "type": "text", + "content": " [7]. Using " + }, + { + "bbox": [ + 47, + 286, + 287, + 335 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{recon}" + }, + { + "bbox": [ + 47, + 286, + 287, + 335 + ], + "type": "text", + "content": ", we compare the tri-plane features " + }, + { + "bbox": [ + 47, + 286, + 287, + 335 + ], + "type": "inline_equation", + "content": "F_{u} = G_{u}(w_{u})" + }, + { + "bbox": [ + 47, + 286, + 287, + 335 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 286, + 287, + 335 + ], + "type": "inline_equation", + "content": "F_{t} = G_{s}(w_{t})" + }, + { + "bbox": [ + 47, + 286, + 287, + 335 + ], + "type": "text", + "content": ", derived from source and target latent codes, respectively. The local unlearning loss is defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 52, + 350, + 287, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 350, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 52, + 350, + 287, + 380 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {l o c a l}} \\left(\\hat {x} _ {u}, \\hat {x} _ {t}\\right) = \\lambda_ {L 2} \\mathcal {L} _ {L 2} \\left(F _ {u}, F _ {t}\\right) \\tag {4} \\\\ + \\lambda_ {p e r} \\mathcal {L} _ {p e r} (\\hat {x} _ {u}, \\hat {x} _ {t}) + \\lambda_ {i d} \\mathcal {L} _ {i d} (\\hat {x} _ {u}, \\hat {x} _ {t}). \\\\ \\end{array}", + "image_path": "7c502ace34b57b7aa8e794a3c600aac29d9566ae82ff87e730735777f817a7f9.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 392, + 287, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 392, + 287, + 417 + ], + "spans": [ + { + "bbox": [ + 47, + 392, + 287, + 417 + ], + "type": "text", + "content": "By adopting " + }, + { + "bbox": [ + 47, + 392, + 287, + 417 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{local}" + }, + { + "bbox": [ + 47, + 392, + 287, + 417 + ], + "type": "text", + "content": ", we can successfully un-identify the given source identity in " + }, + { + "bbox": [ + 47, + 392, + 287, + 417 + ], + "type": "inline_equation", + "content": "\\hat{x}_t" + }, + { + "bbox": [ + 47, + 392, + 287, + 417 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 437, + 289, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 437, + 289, + 582 + ], + "spans": [ + { + "bbox": [ + 47, + 437, + 289, + 582 + ], + "type": "text", + "content": "Adjacency-Aware Unlearning Loss. The above equation considers only one pair of source and target latent codes. However, images of a similar identity to the source identity can be obtained by introducing marginal perturbations to the latent code. For the successful unlearning of the given identity, we need to consider the neighborhood of both the source and the target latent codes. Consequently, as shown in Figure 5, we sample " + }, + { + "bbox": [ + 47, + 437, + 289, + 582 + ], + "type": "inline_equation", + "content": "N_{a}" + }, + { + "bbox": [ + 47, + 437, + 289, + 582 + ], + "type": "text", + "content": " latent codes in the vicinity of the " + }, + { + "bbox": [ + 47, + 437, + 289, + 582 + ], + "type": "inline_equation", + "content": "w_{u}" + }, + { + "bbox": [ + 47, + 437, + 289, + 582 + ], + "type": "text", + "content": ". Specifically, with the scale " + }, + { + "bbox": [ + 47, + 437, + 289, + 582 + ], + "type": "inline_equation", + "content": "\\alpha^{i}" + }, + { + "bbox": [ + 47, + 437, + 289, + 582 + ], + "type": "text", + "content": " sampled from the uniform distribution with hyperparameter " + }, + { + "bbox": [ + 47, + 437, + 289, + 582 + ], + "type": "inline_equation", + "content": "\\alpha_{max}" + }, + { + "bbox": [ + 47, + 437, + 289, + 582 + ], + "type": "text", + "content": ", i.e. " + }, + { + "bbox": [ + 47, + 437, + 289, + 582 + ], + "type": "inline_equation", + "content": "\\alpha^{i} \\sim \\mathcal{U}(0, \\alpha_{max})" + }, + { + "bbox": [ + 47, + 437, + 289, + 582 + ], + "type": "text", + "content": ", we define the distances " + }, + { + "bbox": [ + 47, + 437, + 289, + 582 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 47, + 437, + 289, + 582 + ], + "type": "text", + "content": " to compute the adjacent latent codes as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 103, + 595, + 287, + 625 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 595, + 287, + 625 + ], + "spans": [ + { + "bbox": [ + 103, + 595, + 287, + 625 + ], + "type": "interline_equation", + "content": "\\Delta = \\left\\{\\alpha^ {i} \\cdot \\frac {w _ {r , a} ^ {i} - w _ {u}}{\\| w _ {r , a} ^ {i} - w _ {u} \\| _ {2}} \\right\\} _ {i = 1} ^ {N _ {a}}, \\tag {5}", + "image_path": "fec1d318e4b1293b360887ffdabb607b73be182d609dc99ae43854bdfa12a392.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 639, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 639, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 639, + 287, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 639, + 287, + 715 + ], + "type": "inline_equation", + "content": "w_{r,a}^{i}" + }, + { + "bbox": [ + 47, + 639, + 287, + 715 + ], + "type": "text", + "content": " is a latent code sampled from the random noise vector " + }, + { + "bbox": [ + 47, + 639, + 287, + 715 + ], + "type": "inline_equation", + "content": "z_{r,a}^{i}" + }, + { + "bbox": [ + 47, + 639, + 287, + 715 + ], + "type": "text", + "content": ". Using these distances " + }, + { + "bbox": [ + 47, + 639, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 47, + 639, + 287, + 715 + ], + "type": "text", + "content": ", we can compute " + }, + { + "bbox": [ + 47, + 639, + 287, + 715 + ], + "type": "inline_equation", + "content": "N_{a}" + }, + { + "bbox": [ + 47, + 639, + 287, + 715 + ], + "type": "text", + "content": " latent codes for both the source and the target latent codes. Similar to the local unlearning loss, we optimize the generated tri-plane features and images from " + }, + { + "bbox": [ + 47, + 639, + 287, + 715 + ], + "type": "inline_equation", + "content": "w_{u,a}^{i} = w_{u} + \\Delta^{i}" + }, + { + "bbox": [ + 47, + 639, + 287, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 639, + 287, + 715 + ], + "type": "inline_equation", + "content": "w_{t,a}^{i} = w_{t} + \\Delta^{i}" + }, + { + "bbox": [ + 47, + 639, + 287, + 715 + ], + "type": "text", + "content": " to be similar:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 337, + 92, + 545, + 108 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 92, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 337, + 92, + 545, + 108 + ], + "type": "interline_equation", + "content": "\\hat {x} _ {u, a} ^ {i} = R \\left(F _ {u, a} ^ {i}\\right), \\hat {x} _ {t, a} ^ {i} = R \\left(F _ {u, a} ^ {i}\\right), \\tag {6}", + "image_path": "68f5d472f04e229e3dbff3438cca32bdd0f9e21f7fe2ca855711990b5890d0d0.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 338, + 110, + 545, + 143 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 110, + 545, + 143 + ], + "spans": [ + { + "bbox": [ + 338, + 110, + 545, + 143 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {a d j} \\left(w _ {u}, w _ {t}\\right) = \\frac {1}{N _ {a}} \\sum_ {i = 1} ^ {N _ {a}} \\mathcal {L} _ {\\text {l o c a l}} \\left(\\hat {x} _ {u, a} ^ {i}, \\hat {x} _ {t, a} ^ {i}\\right), \\tag {7}", + "image_path": "af3ef95131e9647d2653ff5b561efa994bc0887c0839434f7fc8821e4b1fcfa7.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 152, + 545, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 152, + 545, + 189 + ], + "spans": [ + { + "bbox": [ + 304, + 152, + 545, + 189 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 152, + 545, + 189 + ], + "type": "inline_equation", + "content": "F_{u,a}^{i} = G_{u}(w_{u,a}^{i}), F_{t,a}^{i} = G_{s}(w_{t,a}^{i})" + }, + { + "bbox": [ + 304, + 152, + 545, + 189 + ], + "type": "text", + "content": " denotes for triplane features, and " + }, + { + "bbox": [ + 304, + 152, + 545, + 189 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{local}" + }, + { + "bbox": [ + 304, + 152, + 545, + 189 + ], + "type": "text", + "content": " in Equation 4. From " + }, + { + "bbox": [ + 304, + 152, + 545, + 189 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{adj}" + }, + { + "bbox": [ + 304, + 152, + 545, + 189 + ], + "type": "text", + "content": ", we can further consider possible variations of the source identity." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 202, + 545, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 202, + 545, + 298 + ], + "spans": [ + { + "bbox": [ + 304, + 202, + 545, + 298 + ], + "type": "text", + "content": "Global Preservation Loss. While the local unlearning loss and adjacency-aware unlearning loss mentioned above facilitate the removal of the source identity, we propose a global preservation loss to mitigate side effects arising from these unlearning loss functions. In the global preservation loss, we constrain the generator to maintain generation performance for latent codes that are relatively distant from both the source and target latent codes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 298, + 545, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 298, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 304, + 298, + 545, + 388 + ], + "type": "text", + "content": "To be precise, we sample " + }, + { + "bbox": [ + 304, + 298, + 545, + 388 + ], + "type": "inline_equation", + "content": "N_{g}" + }, + { + "bbox": [ + 304, + 298, + 545, + 388 + ], + "type": "text", + "content": " latent codes " + }, + { + "bbox": [ + 304, + 298, + 545, + 388 + ], + "type": "inline_equation", + "content": "\\{w_{r,g}^{i}\\}_{i = 1}^{N_{g}}" + }, + { + "bbox": [ + 304, + 298, + 545, + 388 + ], + "type": "text", + "content": " from random noise vectors " + }, + { + "bbox": [ + 304, + 298, + 545, + 388 + ], + "type": "inline_equation", + "content": "\\{z_{r,g}^{i}\\}_{i = 1}^{N_{g}}" + }, + { + "bbox": [ + 304, + 298, + 545, + 388 + ], + "type": "text", + "content": ". We ensure that these do not overlap with the adjacent latent codes used in the adjacency-aware unlearning loss. Unlike the unlearning loss functions, we find that adopting only " + }, + { + "bbox": [ + 304, + 298, + 545, + 388 + ], + "type": "inline_equation", + "content": "L_{per}" + }, + { + "bbox": [ + 304, + 298, + 545, + 388 + ], + "type": "text", + "content": " achieves a balanced performance between identity shift and model preservation. The global preservation loss is computed as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 381, + 396, + 482, + 412 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 396, + 482, + 412 + ], + "spans": [ + { + "bbox": [ + 381, + 396, + 482, + 412 + ], + "type": "interline_equation", + "content": "\\hat {x} _ {u, g} ^ {i} = (R \\circ G _ {u}) (w _ {r, g} ^ {i}),", + "image_path": "26b50c296142598d46c302755536f8987f7ac75a32199262bd9e3cef05ddfa89.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 384, + 414, + 545, + 435 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 414, + 545, + 435 + ], + "spans": [ + { + "bbox": [ + 384, + 414, + 545, + 435 + ], + "type": "interline_equation", + "content": "\\hat {x} _ {s, g} ^ {i} = \\left(R \\circ G _ {s}\\right) \\left(w _ {r, g} ^ {i}\\right), \\tag {8}", + "image_path": "dc1af88bfe540024cafe1bdc52ca615431e068285923b6e92f3cab074283600d.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 335, + 431, + 516, + 463 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 431, + 516, + 463 + ], + "spans": [ + { + "bbox": [ + 335, + 431, + 516, + 463 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {g l o b a l} (G _ {u}, G _ {s}) = \\frac {1}{N _ {g}} \\sum_ {i = 1} ^ {N _ {g}} \\mathcal {L} _ {p e r} (\\hat {x} _ {u, g} ^ {i}, \\hat {x} _ {s, g} ^ {i}).", + "image_path": "89e3864039bac69f84ccc27cf2576a964ab3ea8f086fa2543d6d53ad4d884099.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 472, + 445, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 472, + 445, + 484 + ], + "spans": [ + { + "bbox": [ + 306, + 472, + 445, + 484 + ], + "type": "text", + "content": "In summary, our final objective is:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 334, + 494, + 545, + 507 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 494, + 545, + 507 + ], + "spans": [ + { + "bbox": [ + 334, + 494, + 545, + 507 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {l o c a l}} + \\lambda_ {\\text {a d j}} \\mathcal {L} _ {\\text {a d j}} + \\lambda_ {\\text {g l o b a l}} \\mathcal {L} _ {\\text {g l o b a l}}. \\tag {9}", + "image_path": "f80c7a2e75c1d04c81e2cf0cc98e28c27dfb94c88b5216ba3be75379f1717717.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 517, + 388, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 517, + 388, + 530 + ], + "spans": [ + { + "bbox": [ + 306, + 517, + 388, + 530 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 536, + 422, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 536, + 422, + 550 + ], + "spans": [ + { + "bbox": [ + 306, + 536, + 422, + 550 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 555, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 545, + 628 + ], + "type": "text", + "content": "**Baseline.** Since we propose generative identity unlearning task for the first time, to evaluate the effectiveness of GUIDE, we constructed a simple baseline. In the baseline, we used the target latent code as the average one for the unlearning. During the unlearning, we updated the pre-trained generator using " + }, + { + "bbox": [ + 304, + 555, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{local}" + }, + { + "bbox": [ + 304, + 555, + 545, + 628 + ], + "type": "text", + "content": " as described in Equation 4." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": "Implementation Details. We built GUIDE based on the 3D generative adversarial network [4] pre-trained on FFHQ dataset [19]. We used GOAE [52] as a GAN inversion network to obtain the latent code from an image. The image resolution we used in our experiments is " + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": " with a rendering resolution of " + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": ". We used Adam optimizer" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9155" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 74, + 545, + 247 + ], + "blocks": [ + { + "bbox": [ + 52, + 74, + 545, + 247 + ], + "lines": [ + { + "bbox": [ + 52, + 74, + 545, + 247 + ], + "spans": [ + { + "bbox": [ + 52, + 74, + 545, + 247 + ], + "type": "table", + "html": "
SourceRandomIn-Domain (FFHQ)Out-of-Domain (CelebAHQ)
BaselineOursBaselineOursBaselineOursBaselineOursBaselineOurs
Unlearned
", + "image_path": "1f9c0a39103959cae32f313595f8c200c1b3f48f14e32978b51a2ccc16493a3c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 310, + 287, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 310, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 47, + 310, + 287, + 360 + ], + "type": "text", + "content": "[23] with a learning rate of " + }, + { + "bbox": [ + 47, + 310, + 287, + 360 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 47, + 310, + 287, + 360 + ], + "type": "text", + "content": " in the unlearning procedure. The hyperparameters used in the experiments were: " + }, + { + "bbox": [ + 47, + 310, + 287, + 360 + ], + "type": "inline_equation", + "content": "d = 30" + }, + { + "bbox": [ + 47, + 310, + 287, + 360 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 310, + 287, + 360 + ], + "type": "inline_equation", + "content": "\\alpha_{max} = 15" + }, + { + "bbox": [ + 47, + 310, + 287, + 360 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 310, + 287, + 360 + ], + "type": "inline_equation", + "content": "\\lambda_{L2} = 10^{-2}" + }, + { + "bbox": [ + 47, + 310, + 287, + 360 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 310, + 287, + 360 + ], + "type": "inline_equation", + "content": "\\lambda_{per} = 1" + }, + { + "bbox": [ + 47, + 310, + 287, + 360 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 310, + 287, + 360 + ], + "type": "inline_equation", + "content": "\\lambda_{id} = 10^{-1}" + }, + { + "bbox": [ + 47, + 310, + 287, + 360 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 310, + 287, + 360 + ], + "type": "inline_equation", + "content": "N_a = N_g = 2" + }, + { + "bbox": [ + 47, + 310, + 287, + 360 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 310, + 287, + 360 + ], + "type": "inline_equation", + "content": "\\lambda_{adj} = \\lambda_{global} = 1" + }, + { + "bbox": [ + 47, + 310, + 287, + 360 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 374, + 288, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 374, + 288, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 374, + 288, + 517 + ], + "type": "text", + "content": "Dataset and Scenarios. We evaluated GUIDE in three scenarios: Random, where we set an unlearning target image from a randomly sampled noise vector; InD (domain), where we sampled an image from the FFHQ dataset [19] used for pre-training; and OOD (out-of-domain), where the unlearning target image was sampled from the CelebAHQ dataset [18]. For InD and OOD, we used the GAN inversion network to obtain corresponding latent codes. For OOD scenario, we also conducted multi-image test since there were multiple images with a same identity in CelebAHQ. On the other hand, we performed only single-image test in the Random and InD scenarios." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 533, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 713 + ], + "type": "text", + "content": "Evaluation Metrics. We evaluated GUIDE on two key aspects. Firstly, we estimated the efficacy of our approach in preventing the generator from producing images similar to the unlearning target. We quantitatively measured similarity of identities (ID) using face recognition network, CurricularFace [17], between images generated from the same latent codes before and after unlearning. Moreover, we utilized " + }, + { + "bbox": [ + 46, + 533, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}_{\\mathrm{others}}" + }, + { + "bbox": [ + 46, + 533, + 287, + 713 + ], + "type": "text", + "content": " to estimate the erasure of a identity from images which are unseen during training but containing the same identity of the source image. Secondly, we assessed whether our method preserves overall generation performance using the Fréchet Inception distance (FID) score [16]. Different from the existing usages, we utilized two variants of FID. First, we evaluated the distribution shift of generated images the pre-trained generator and the un-" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 329, + 309, + 529, + 473 + ], + "blocks": [ + { + "bbox": [ + 46, + 255, + 546, + 289 + ], + "lines": [ + { + "bbox": [ + 46, + 255, + 546, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 546, + 289 + ], + "type": "text", + "content": "Figure 6. Qualitative results of GUIDE and the baseline in generative identity unlearning task. For the given source image each (the first row), GUIDE and the baseline tried to erase the identity in the pre-trained generator. The images in the second and third row are the target and unlearned images, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 329, + 309, + 529, + 473 + ], + "lines": [ + { + "bbox": [ + 329, + 309, + 529, + 473 + ], + "spans": [ + { + "bbox": [ + 329, + 309, + 529, + 473 + ], + "type": "image", + "image_path": "98eb4066c843cf5e7dbf10a6c817e84367ed08a86a316b95ada8d29f1f722a01.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 481, + 545, + 525 + ], + "lines": [ + { + "bbox": [ + 305, + 481, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 305, + 481, + 545, + 525 + ], + "type": "text", + "content": "Figure 7. Qualitative results of GUIDE and the baseline on a multi-image test using CelebAHQ dataset. We additionally utilized images that are unseen during unlearning, to show how thoroughly erase the given identity." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 536, + 545, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 536, + 545, + 573 + ], + "spans": [ + { + "bbox": [ + 305, + 536, + 545, + 573 + ], + "type": "text", + "content": "learned generator via " + }, + { + "bbox": [ + 305, + 536, + 545, + 573 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{\\mathrm{pre}}" + }, + { + "bbox": [ + 305, + 536, + 545, + 573 + ], + "type": "text", + "content": ". Furthermore, we measured the distribution shift with respect to the real FFHQ images, which we denoted as " + }, + { + "bbox": [ + 305, + 536, + 545, + 573 + ], + "type": "inline_equation", + "content": "\\Delta \\mathrm{FID}_{\\mathrm{real}}" + }, + { + "bbox": [ + 305, + 536, + 545, + 573 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 582, + 411, + 595 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 582, + 411, + 595 + ], + "spans": [ + { + "bbox": [ + 306, + 582, + 411, + 595 + ], + "type": "text", + "content": "4.2. Main Experiment" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 601, + 421, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 601, + 421, + 613 + ], + "spans": [ + { + "bbox": [ + 306, + 601, + 421, + 613 + ], + "type": "text", + "content": "4.2.1 Qualitative Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "We conducted a comparative analysis of GUIDE against the baseline in the generative identity unlearning task. Initiating from the provided source image, we aimed to eliminate the identity within the pre-trained generator, as illustrated in Figure 6. We presented the resulting unlearned image, along with the target image optimized in our loss functions. Notably, GUIDE effectively erases identities whether synthetic, presented during pre-training, or unseen." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9156" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 71, + 547, + 137 + ], + "blocks": [ + { + "bbox": [ + 48, + 71, + 547, + 137 + ], + "lines": [ + { + "bbox": [ + 48, + 71, + 547, + 137 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 547, + 137 + ], + "type": "table", + "html": "
MethodsRandomIn-Domain (FFHQ)Out-of-Domain (CelebAHQ)
ID (↓)FIDpre (↓)ΔFIDreal (↓)ID (↓)FIDpre (↓)ΔFIDreal (↓)ID (↓)FIDpre (↓)ΔFIDreal (↓)
Baseline0.19 ± 0.0911.73 ± 2.747.46 ± 2.200.16 ± 0.079.00 ± 1.154.15 ± 1.180.12 ± 0.069.52 ± 1.534.75 ± 0.89
+ extrapolated wt0.12 ± 0.0614.28 ± 3.349.63 ± 2.530.05 ± 0.0612.78 ± 1.826.76 ± 1.410.02 ± 0.0513.02 ± 3.207.31 ± 1.98
+ Ladj0.14 ± 0.0719.65 ± 4.9013.94 ± 3.590.04 ± 0.0613.53 ± 2.087.35 ± 1.700.01 ± 0.0513.63 ± 3.527.83 ± 2.19
+ Lglobal (GUIDE)0.14 ± 0.0610.80 ± 2.706.64 ± 1.600.06 ± 0.068.00 ± 1.203.05 ± 0.810.03 ± 0.057.88 ± 1.963.34 ± 1.10
", + "image_path": "b16d41950d4932b684867f56533b114c7ec8e83f39877a17d3183961a207a665.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 53, + 186, + 286, + 352 + ], + "blocks": [ + { + "bbox": [ + 53, + 186, + 286, + 352 + ], + "lines": [ + { + "bbox": [ + 53, + 186, + 286, + 352 + ], + "spans": [ + { + "bbox": [ + 53, + 186, + 286, + 352 + ], + "type": "image", + "image_path": "4852243969db08bbcc2f30262ac7108a70c1fb4675d35e4b502a4f324d6091cd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 361, + 287, + 418 + ], + "lines": [ + { + "bbox": [ + 46, + 361, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 287, + 418 + ], + "type": "text", + "content": "Figure 8. Qualitative comparison between GUIDE and the baseline on the preservation of the generation quality of other identities. GUIDE generates images almost identical to those synthesized by " + }, + { + "bbox": [ + 46, + 361, + 287, + 418 + ], + "type": "inline_equation", + "content": "G_{s}" + }, + { + "bbox": [ + 46, + 361, + 287, + 418 + ], + "type": "text", + "content": ", whereas the baseline often results in noticeable changes, e.g., beard shape, hairstyle change, hat." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 431, + 287, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 431, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 431, + 287, + 552 + ], + "type": "text", + "content": "To evaluate the thoroughness of identity removal, we performed a multi-image test using identities from the CelebAHQ dataset. This test involved assessing the ID similarity not only for the unlearned image derived from the source image but also for other images sharing the same identity. As shown in Figure 7, GUIDE showed superior generalization for unseen images compared to the baseline. This improvement is attributed to the adjacency-aware unlearning, which facilitated the unlearning process not just for the given images but also for their neighborhood." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 552, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 552, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 552, + 287, + 624 + ], + "type": "text", + "content": "In Figure 8, we conducted an experiment to assess the effect of the unlearning process on other identities. While the baseline had a significant impact on other identities through the unlearning, GUIDE showed a relatively lesser effect. We attribute this to the global preservation loss, which constrained the distribution shift on other latent codes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 634, + 168, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 168, + 645 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 168, + 645 + ], + "type": "text", + "content": "4.2.2 Quantitative Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "In Table 1, we compared GUIDE to the baseline by gradually applying the components of GUIDE. By configuring " + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "w_{t}" + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": " through extrapolation, we achieved performance improvements in ID similarity across three scenarios. Notably, we observed a significant difference in ID similarities in the" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 306, + 187, + 547, + 240 + ], + "blocks": [ + { + "bbox": [ + 46, + 146, + 546, + 169 + ], + "lines": [ + { + "bbox": [ + 46, + 146, + 546, + 169 + ], + "spans": [ + { + "bbox": [ + 46, + 146, + 546, + 169 + ], + "type": "text", + "content": "Table 1. Quantitative results of GUIDE and the baseline in the generative identity unlearning task, tested in a single-image setting using one image per identity. Starting from the baseline, we gradually introduced components of GUIDE." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 187, + 547, + 240 + ], + "lines": [ + { + "bbox": [ + 306, + 187, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 306, + 187, + 547, + 240 + ], + "type": "table", + "html": "
MethodsID (↓)ID others (↓)FID pre (↓)ΔFID real (↓)
Baseline0.12 ± 0.060.28 ± 0.089.52 ± 1.534.75 ± 0.89
+ extrapolated wt0.02 ± 0.050.15 ± 0.0713.02 ± 3.207.31 ± 1.98
+ Ladj0.01 ± 0.050.14 ± 0.0713.63 ± 3.527.83 ± 2.19
+ Lglobal (GUIDE)0.03 ± 0.050.17 ± 0.087.88 ± 1.963.34 ± 1.10
", + "image_path": "2e7ed798aac777dc786c0a1113337fecc0372c6e9b63b32bdefce7a6b043405c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 248, + 546, + 293 + ], + "lines": [ + { + "bbox": [ + 305, + 248, + 546, + 293 + ], + "spans": [ + { + "bbox": [ + 305, + 248, + 546, + 293 + ], + "type": "text", + "content": "Table 2. Quantitative results of GUIDE and the baseline in the generative identity unlearning in a multi-image setting, i.e., using a single image for unlearning and the other images for testing. We used CelebAHQ dataset for this test." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 303, + 545, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 303, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 304, + 303, + 545, + 363 + ], + "type": "text", + "content": "random scenario, indicating that in cases where a latent code was close to " + }, + { + "bbox": [ + 304, + 303, + 545, + 363 + ], + "type": "inline_equation", + "content": "\\bar{w}" + }, + { + "bbox": [ + 304, + 303, + 545, + 363 + ], + "type": "text", + "content": ", i.e., as in the random scenario, there was insufficient removal of identity. The effectiveness of employing an extrapolation between the source latent code and the average latent code was evident in such instances." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 363, + 545, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 363, + 545, + 435 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 545, + 435 + ], + "type": "text", + "content": "The adjacency-aware unlearning loss further enhanced the unlearning an identity. This loss was designed to cover the vicinity of the source latent code, thereby promoting unlearning on the source latent code itself. Finally, the application of the global preservation loss effectively reduced the estimated distribution shift using " + }, + { + "bbox": [ + 304, + 363, + 545, + 435 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{\\mathrm{pre}}" + }, + { + "bbox": [ + 304, + 363, + 545, + 435 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 363, + 545, + 435 + ], + "type": "inline_equation", + "content": "\\Delta \\mathrm{FID}_{\\mathrm{real}}" + }, + { + "bbox": [ + 304, + 363, + 545, + 435 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 435, + 545, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 435, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 545, + 532 + ], + "type": "text", + "content": "Moreover, we conducted a multi-image test in an OOD scenario. In this particular experiment, we introduced additional metric - " + }, + { + "bbox": [ + 304, + 435, + 545, + 532 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}_{\\mathrm{others}}" + }, + { + "bbox": [ + 304, + 435, + 545, + 532 + ], + "type": "text", + "content": " aimed at quantifying ID similarities for the unseen images associated with the source identity. As presented in Table 2, the introduction of the adjacency-aware unlearning loss resulted in a remarkable improvement in " + }, + { + "bbox": [ + 304, + 435, + 545, + 532 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}_{\\mathrm{others}}" + }, + { + "bbox": [ + 304, + 435, + 545, + 532 + ], + "type": "text", + "content": ", emphasizing the effectiveness of this unlearning approach for handling unseen images." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 539, + 400, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 539, + 400, + 552 + ], + "spans": [ + { + "bbox": [ + 306, + 539, + 400, + 552 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "content": "Effect of " + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "content": " in Determination of " + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "inline_equation", + "content": "w_{t}" + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "content": ". We conducted an ablation study by comparing target images derived from varied values of " + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "content": ". Setting " + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "content": " to 0 denotes utilizing the " + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\bar{w}" + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "inline_equation", + "content": "w_{t}" + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "content": " in the unlearning process. For " + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "inline_equation", + "content": "d < 0" + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "content": ", we designated " + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "inline_equation", + "content": "w_{t}" + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "content": " as an interpolated latent code between the " + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "inline_equation", + "content": "w_{s}" + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\bar{w}" + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "content": ". Conversely, for " + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "inline_equation", + "content": "d > 0" + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "content": ", we employed an extrapolated " + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "inline_equation", + "content": "w_{t}" + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "content": ", as detailed in Section 3.2. As illustrated in Figure 9, when " + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "inline_equation", + "content": "d < 0" + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "content": ", the target image closely aligns with the given source images. However, as " + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "content": " deviates from 0, the quality of the target image rapidly deteriorates, resulting in a pronounced collapse in the distribution of the pre-trained generator. Consequently, the effectiveness of unlearning with such target images diminishes in removing identity from the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9157" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 71, + 279, + 190 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 279, + 190 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 279, + 190 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 279, + 190 + ], + "type": "table", + "html": "
SourceTarget
d = -10d = -5d = 0d = 30 (Ours)d = 60
Random
FFHQ
d = -70d = -35d = 0d = 30 (Ours)d = 60
", + "image_path": "443c05e2d1b560cf48b1fc0c57d9e8313479e8f75c361064dd31632491e29ac0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 95, + 299, + 240, + 367 + ], + "blocks": [ + { + "bbox": [ + 46, + 200, + 287, + 289 + ], + "lines": [ + { + "bbox": [ + 46, + 200, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 200, + 287, + 289 + ], + "type": "text", + "content": "Figure 9. Ablation study to figure out the effectiveness of " + }, + { + "bbox": [ + 46, + 200, + 287, + 289 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 200, + 287, + 289 + ], + "type": "text", + "content": ". We visualized target images corresponding to each source image with different values of " + }, + { + "bbox": [ + 46, + 200, + 287, + 289 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 200, + 287, + 289 + ], + "type": "text", + "content": ". The target images were generated using target latent codes derived from interpolated latent codes, the average latent code (" + }, + { + "bbox": [ + 46, + 200, + 287, + 289 + ], + "type": "inline_equation", + "content": "d = 0" + }, + { + "bbox": [ + 46, + 200, + 287, + 289 + ], + "type": "text", + "content": "), or extrapolated latent codes (" + }, + { + "bbox": [ + 46, + 200, + 287, + 289 + ], + "type": "inline_equation", + "content": "d > 0" + }, + { + "bbox": [ + 46, + 200, + 287, + 289 + ], + "type": "text", + "content": "). Interpolation and extrapolation were carried out between the source and the average latent code. In the case of interpolation, the center between the source and the average latent code was computed." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 95, + 299, + 240, + 367 + ], + "lines": [ + { + "bbox": [ + 95, + 299, + 240, + 367 + ], + "spans": [ + { + "bbox": [ + 95, + 299, + 240, + 367 + ], + "type": "table", + "html": "
αmaxID (↓)IDothers (↓)
00.1205 ± 0.06030.2754 ± 0.0791
100.0892 ± 0.06200.2123 ± 0.0762
150.0878 ± 0.03750.2094 ± 0.0692
200.0900 ± 0.05380.2105 ± 0.0924
300.0926 ± 0.05610.2111 ± 0.0653
", + "image_path": "d2622d79190a4fc8d2a3529e9f6ab875accf9065c3cf0c1f40cf5a695a7bebf8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 443, + 287, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 443, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 443, + 287, + 574 + ], + "type": "text", + "content": "source images. Setting " + }, + { + "bbox": [ + 46, + 443, + 287, + 574 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 443, + 287, + 574 + ], + "type": "text", + "content": " to 0 might suggest the use of " + }, + { + "bbox": [ + 46, + 443, + 287, + 574 + ], + "type": "inline_equation", + "content": "\\bar{w}" + }, + { + "bbox": [ + 46, + 443, + 287, + 574 + ], + "type": "text", + "content": " as an effective target for erasing identity. However, our ablation studies indicate that when the source image closely aligns with " + }, + { + "bbox": [ + 46, + 443, + 287, + 574 + ], + "type": "inline_equation", + "content": "\\bar{w}" + }, + { + "bbox": [ + 46, + 443, + 287, + 574 + ], + "type": "text", + "content": ", the unlearning procedure fails to thoroughly eliminate the identity. Conversely, when " + }, + { + "bbox": [ + 46, + 443, + 287, + 574 + ], + "type": "inline_equation", + "content": "d > 0" + }, + { + "bbox": [ + 46, + 443, + 287, + 574 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 443, + 287, + 574 + ], + "type": "inline_equation", + "content": "w_{t}" + }, + { + "bbox": [ + 46, + 443, + 287, + 574 + ], + "type": "text", + "content": " contains a distinct identity compared to the source image while maintaining a consistent distance from " + }, + { + "bbox": [ + 46, + 443, + 287, + 574 + ], + "type": "inline_equation", + "content": "\\bar{w}" + }, + { + "bbox": [ + 46, + 443, + 287, + 574 + ], + "type": "text", + "content": ". Among the instances where " + }, + { + "bbox": [ + 46, + 443, + 287, + 574 + ], + "type": "inline_equation", + "content": "d > 0" + }, + { + "bbox": [ + 46, + 443, + 287, + 574 + ], + "type": "text", + "content": ", our ablation studies reveal that setting " + }, + { + "bbox": [ + 46, + 443, + 287, + 574 + ], + "type": "inline_equation", + "content": "d = 30" + }, + { + "bbox": [ + 46, + 443, + 287, + 574 + ], + "type": "text", + "content": " achieves a balanced performance between effective unlearning and preservation of the generation performance of the pre-trained model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": "Effect of " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\alpha_{max}" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{adj}" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": ". In Table 3, we scrutinized the effectiveness of the adjacency-aware unlearning loss. To ensure a fair comparison, we employed " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\bar{w}" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "w_{t}" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " in this experiment, and we used " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{local}" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{adj}" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " in the unlearning procedure. Rows corresponding to " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\alpha_{max} = 0" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " represent experimental results without the incorporation of " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{adj}" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " in the unlearning procedure. The introduction of " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{adj}" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " resulted in consistent performance gains in " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}_{\\mathrm{others}}" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": ". This observation highlights the efficacy of considering not only the pair of source and target latent codes but also their surroundings for unlearning the entire identity." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 342, + 71, + 511, + 110 + ], + "blocks": [ + { + "bbox": [ + 46, + 374, + 287, + 430 + ], + "lines": [ + { + "bbox": [ + 46, + 374, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 374, + 287, + 430 + ], + "type": "text", + "content": "Table 3. Ablation study to figure out the effectiveness of " + }, + { + "bbox": [ + 46, + 374, + 287, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{adj}" + }, + { + "bbox": [ + 46, + 374, + 287, + 430 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 374, + 287, + 430 + ], + "type": "inline_equation", + "content": "\\alpha_{max}" + }, + { + "bbox": [ + 46, + 374, + 287, + 430 + ], + "type": "text", + "content": ". We compared the performance based on how successfully the given identity was erased, using ID and IDothers metric. The row where " + }, + { + "bbox": [ + 46, + 374, + 287, + 430 + ], + "type": "inline_equation", + "content": "\\alpha_{max} = 0" + }, + { + "bbox": [ + 46, + 374, + 287, + 430 + ], + "type": "text", + "content": " denotes the baseline. We used CelebAHQ dataset in this experiment." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 342, + 71, + 511, + 110 + ], + "lines": [ + { + "bbox": [ + 342, + 71, + 511, + 110 + ], + "spans": [ + { + "bbox": [ + 342, + 71, + 511, + 110 + ], + "type": "table", + "html": "
\\( \\mathcal{L}_{local} \\)\\( \\mathcal{L}_{global} \\)\\( \\mathrm{FID}_{\\text{pre}}(\\downarrow) \\)\\( \\Delta \\mathrm{FID}_{\\text{real}}(\\downarrow) \\)
✓9.52 ± 1.534.75 ± 0.89
✓✓4.63 ± 0.431.48 ± 0.29
", + "image_path": "9eb6ee6e7f81fb311d3bc4270918e89721b6396518ed1f010228c7f13200510e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 118, + 545, + 162 + ], + "lines": [ + { + "bbox": [ + 305, + 118, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 305, + 118, + 545, + 162 + ], + "type": "text", + "content": "Table 4. Ablation study to figure our the effectiveness of " + }, + { + "bbox": [ + 305, + 118, + 545, + 162 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{global}" + }, + { + "bbox": [ + 305, + 118, + 545, + 162 + ], + "type": "text", + "content": ". We compared how preserved the performance of the pre-trained model through the unlearning process, via " + }, + { + "bbox": [ + 305, + 118, + 545, + 162 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{\\mathrm{pre}}" + }, + { + "bbox": [ + 305, + 118, + 545, + 162 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 118, + 545, + 162 + ], + "type": "inline_equation", + "content": "\\Delta \\mathrm{FID}_{\\mathrm{real}}" + }, + { + "bbox": [ + 305, + 118, + 545, + 162 + ], + "type": "text", + "content": ". We used CelebAHQ dataset in this experiment." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 183, + 545, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 183, + 545, + 291 + ], + "spans": [ + { + "bbox": [ + 304, + 183, + 545, + 291 + ], + "type": "text", + "content": "Effect of " + }, + { + "bbox": [ + 304, + 183, + 545, + 291 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{global}" + }, + { + "bbox": [ + 304, + 183, + 545, + 291 + ], + "type": "text", + "content": ". To assess the effectiveness of the global preservation loss, a similar experiment was conducted as the previous experiment, i.e., setting " + }, + { + "bbox": [ + 304, + 183, + 545, + 291 + ], + "type": "inline_equation", + "content": "\\bar{w}" + }, + { + "bbox": [ + 304, + 183, + 545, + 291 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 183, + 545, + 291 + ], + "type": "inline_equation", + "content": "w_{t}" + }, + { + "bbox": [ + 304, + 183, + 545, + 291 + ], + "type": "text", + "content": ". The results are presented in Table 4. The application of " + }, + { + "bbox": [ + 304, + 183, + 545, + 291 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{global}" + }, + { + "bbox": [ + 304, + 183, + 545, + 291 + ], + "type": "text", + "content": " demonstrated consistent performance improvements in both " + }, + { + "bbox": [ + 304, + 183, + 545, + 291 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{\\mathrm{pre}}" + }, + { + "bbox": [ + 304, + 183, + 545, + 291 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 183, + 545, + 291 + ], + "type": "inline_equation", + "content": "\\Delta \\mathrm{FID}_{\\mathrm{real}}" + }, + { + "bbox": [ + 304, + 183, + 545, + 291 + ], + "type": "text", + "content": ". This suggests that imposing constraints on the generator to maintain its generation performance in latent codes distant from our primary focus is effective in reducing distribution shifts in generative models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 303, + 378, + 315 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 303, + 378, + 315 + ], + "spans": [ + { + "bbox": [ + 306, + 303, + 378, + 315 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 323, + 545, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 323, + 545, + 551 + ], + "spans": [ + { + "bbox": [ + 304, + 323, + 545, + 551 + ], + "type": "text", + "content": "In this paper, we introduced a novel task, referred to as generative identity unlearning, designed to address privacy concerns in pre-trained generative adversarial networks. This task requires thoroughly removing the identity of a single source image from the pre-trained generator. To achieve this, we proposed a new framework, GUIDE (Generative Unlearning for any IDENTITY). To unlearn the single identity, we first defined the target latent code via extrapolation, moving away from the average latent by the pre-defined distance in the direction from the source to the average latent. Using this, our GUIDE successfully unlearned the given identity via Latent Target Unlearning (LTU), which optimized the pre-trained model to preserve the overall generative ability but not to generate the same identity within the local space. Experimental results demonstrated the effectiveness of GUIDE with promising outcomes. We anticipate that our work will be widely applied in research or the industry field, providing users with a sense of freedom from privacy concerns through identity removal." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 562, + 403, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 562, + 403, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 562, + 403, + 574 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 580, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 580, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 580, + 545, + 713 + ], + "type": "text", + "content": "This work was supported by MSIT (Ministry of Science and ICT), Korea, under the ITRC (Information Technology Research Center) support program (IITP-2024-RS-2023-00258649) supervised by the IITP (Institute for Information & Communications Technology Planning & Evaluation), and in part by NRF-2023S1A5A2A21083590, and in part by the IITP grant funded by the Korea Government (MSIT) (Artificial Intelligence Innovation Hub) under Grant 2021-0-02068, and by the IITP grant funded by the Korea government (MSIT) (No.RS-2022-00155911, Artificial Intelligence Convergence Innovation Human Resources Development (Kyung Hee University))." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9158" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 67, + 107, + 79 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 67, + 107, + 79 + ], + "spans": [ + { + "bbox": [ + 48, + 67, + 107, + 79 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 87, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 87, + 287, + 131 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 87, + 287, + 131 + ], + "spans": [ + { + "bbox": [ + 53, + 87, + 287, + 131 + ], + "type": "text", + "content": "[1] Yuval Alaluf, Or Patashnik, and Daniel Cohen-Or. Restyle: A residual-based stylegan encoder via iterative refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6711–6720, 2021. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 132, + 288, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 132, + 288, + 175 + ], + "spans": [ + { + "bbox": [ + 53, + 132, + 288, + 175 + ], + "type": "text", + "content": "[2] Thomas Baumhauer, Pascal Schottle, and Matthias Zeppelinzauer. Machine unlearning: Linear filtration for logit-based classifiers. Machine Learning, 111(9):3203-3226, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 177, + 288, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 177, + 288, + 241 + ], + "spans": [ + { + "bbox": [ + 53, + 177, + 288, + 241 + ], + "type": "text", + "content": "[3] Nicolas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikash Sehwag, Florian Tramer, Borja Balle, Daphne Ippolito, and Eric Wallace. Extracting training data from diffusion models. In Proceedings of the 32nd USENIX Security Symposium (USENIX Security 23), pages 5253-5270, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 243, + 288, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 243, + 288, + 320 + ], + "spans": [ + { + "bbox": [ + 53, + 243, + 288, + 320 + ], + "type": "text", + "content": "[4] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 322, + 287, + 366 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 322, + 287, + 366 + ], + "spans": [ + { + "bbox": [ + 53, + 322, + 287, + 366 + ], + "type": "text", + "content": "[5] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 367, + 287, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 367, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 53, + 367, + 287, + 409 + ], + "type": "text", + "content": "[6] Vikram S Chundawat, Ayush K Tarun, Murari Mandal, and Mohan Kankanhalli. Zero-shot machine unlearning. IEEE Transactions on Information Forensics and Security, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 411, + 287, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 411, + 287, + 466 + ], + "spans": [ + { + "bbox": [ + 53, + 411, + 287, + 466 + ], + "type": "text", + "content": "[7] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4690-4699, 2019. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 468, + 287, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 468, + 287, + 522 + ], + "spans": [ + { + "bbox": [ + 53, + 468, + 287, + 522 + ], + "type": "text", + "content": "[8] Matt Fredrikson, Somesh Jha, and Thomas Ristenpart. Model inversion attacks that exploit confidence information and basic countermeasures. In Proceedings of the 22nd ACM SIGSAC conference on computer and communications security, pages 1322-1333, 2015. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 523, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 523, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 53, + 523, + 287, + 578 + ], + "type": "text", + "content": "[9] Rohit Gandikota, Joanna Materzyńska, Jaden Fiotto-Kaufman, and David Bau. Erasing concepts from diffusion models. In Proceedings of the 2023 IEEE International Conference on Computer Vision (ICCV), pages 2426-2436, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 579, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 287, + 623 + ], + "type": "text", + "content": "[10] Antonio Ginart, Melody Guan, Gregory Valiant, and James Y Zou. Making ai forget you: Data deletion in machine learning. Advances in Neural Information Processing Systems, 32, 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 624, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 678 + ], + "type": "text", + "content": "[11] Aditya Golatkar, Alessandro Achille, and Stefano Soatto. Eternal sunshine of the spotless net: Selective forgetting in deep networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9304–9312, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "text", + "content": "[12] Aditya Golatkar, Alessandro Achille, and Stefano Soatto. Forgetting outside the box: Scrubbing deep networks of information accessible from input-output observations. In Pro" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "ceedings of the European Conference on Computer Vision (ECCV), pages 383-398. Springer, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 98, + 547, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 98, + 547, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 98, + 547, + 152 + ], + "type": "text", + "content": "[13] Aditya Golatkar, Alessandro Achille, Avinash Ravichandran, Marzia Polito, and Stefano Soatto. Mixed-privacy forgetting in deep networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 792–801, 2021." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 154, + 545, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 154, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 307, + 154, + 545, + 198 + ], + "type": "text", + "content": "[14] Varun Gupta, Christopher Jung, Seth Neel, Aaron Roth, Saeed Sharifi-Malvajerdi, and Chris Waites. Adaptive machine unlearning. Advances in Neural Information Processing Systems, 34:16319-16330, 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 200, + 545, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 200, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 307, + 200, + 545, + 243 + ], + "type": "text", + "content": "[15] Jamie Hayes, Luca Melis, George Danezis, and Emiliano De Cristofaro. Logan: Membership inference attacks against generative models. arXiv preprint arXiv:1705.07663, 2017. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 246, + 545, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 246, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 307, + 246, + 545, + 300 + ], + "type": "text", + "content": "[16] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in Neural Information Processing Systems, 30, 2017. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 303, + 545, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 303, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 307, + 303, + 545, + 369 + ], + "type": "text", + "content": "[17] Yuge Huang, Yuhan Wang, Ying Tai, Xiaoming Liu, Pengcheng Shen, Shaoxin Li, Jilin Li, and Feiyue Huang. Curricularface: adaptive curriculum learning loss for deep face recognition. In proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5901-5910, 2020. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 371, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 371, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 371, + 545, + 415 + ], + "type": "text", + "content": "[18] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability and variation. In Proceedings of the International Conference on Learning Representations (ICLR), 2018. 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 417, + 545, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 417, + 545, + 461 + ], + "spans": [ + { + "bbox": [ + 307, + 417, + 545, + 461 + ], + "type": "text", + "content": "[19] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 3, 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 463, + 545, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 463, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 307, + 463, + 545, + 518 + ], + "type": "text", + "content": "[20] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 520, + 545, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 520, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 307, + 520, + 545, + 564 + ], + "type": "text", + "content": "[21] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 566, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 566, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 307, + 566, + 545, + 632 + ], + "type": "text", + "content": "[22] Mahdi Khosravy, Kazuaki Nakamura, Yuki Hirose, Naoko Nitta, and Noboru Babaguchi. Model inversion attack by integration of deep generative models: Privacy-sensitive face generation from a face recognition system. IEEE Transactions on Information Forensics and Security, 17:357-372, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 634, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 634, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 634, + 545, + 667 + ], + "type": "text", + "content": "[23] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proceedings of the International Conference on Learning Representations (ICLR), 2015. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 669, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 714 + ], + "type": "text", + "content": "[24] Zhifeng Kong and Kamalika Chaudhuri. Data redaction from pre-trained gans. In Proceedings of the 2023 IEEE Conference on Secure and Trustworthy Machine Learning (SaTML), pages 638-677. IEEE, 2023. 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 316, + 757 + ], + "type": "text", + "content": "9159" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "text", + "content": "[25] Nupur Kumari, Bingliang Zhang, Sheng-Yu Wang, Eli Shechtman, Richard Zhang, and Jun-Yan Zhu. Ablating concepts in text-to-image diffusion models. In Proceedings of the 2023 IEEE International Conference on Computer Vision (ICCV), pages 22691-22702, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 288, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 288, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 288, + 183 + ], + "type": "text", + "content": "[26] Ronak Mehta, Sourav Pal, Vikas Singh, and Sathya N Ravi. Deep unlearning via randomized conditionally independent hessian. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10422-10431, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 183, + 288, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 183, + 288, + 247 + ], + "spans": [ + { + "bbox": [ + 48, + 183, + 288, + 247 + ], + "type": "text", + "content": "[27] Jun-Yeong Moon, Keon-Hee Park, Jung Uk Kim, and Gyeong-Moon Park. Online class incremental learning on stochastic blurry task boundary via mask and visual prompt tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11731-11741, 2023." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 249, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 249, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 249, + 288, + 282 + ], + "type": "text", + "content": "[28] Saemi Moon, Seunghyuk Cho, and Dongwoo Kim. Feature unlearning for generative models via implicit feedback. arXiv preprint arXiv:2303.05699, 2023. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 283, + 288, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 283, + 288, + 327 + ], + "spans": [ + { + "bbox": [ + 48, + 283, + 288, + 327 + ], + "type": "text", + "content": "[29] Seung-Jun Moon and Gyeong-Moon Park. Interstyle: Encoding an interest region for robust stylegan inversion. In Proceedings of the European Conference on Computer Vision (ECCV), pages 460-476. Springer, 2022. 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 327, + 288, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 288, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 288, + 381 + ], + "type": "text", + "content": "[30] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2085–2094, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 382, + 288, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 382, + 288, + 425 + ], + "spans": [ + { + "bbox": [ + 48, + 382, + 288, + 425 + ], + "type": "text", + "content": "[31] Alexandra Peste, Dan Alistarh, and Christoph H Lampert. Ssse: Efficiently erasing samples from trained machine learning models. corr. arXiv preprint arXiv:2107.03860, 1 (3), 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 426, + 288, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 288, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 288, + 480 + ], + "type": "text", + "content": "[32] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In Proceedings of the International Conference on Machine Learning (ICML), pages 8821-8831. PMLR, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 481, + 288, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 288, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 288, + 525 + ], + "type": "text", + "content": "[33] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 525, + 288, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 288, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 288, + 590 + ], + "type": "text", + "content": "[34] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2287–2296, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 592, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 288, + 635 + ], + "type": "text", + "content": "[35] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on graphics (TOG), 42(1):1-13, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 288, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 691 + ], + "type": "text", + "content": "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "type": "text", + "content": "[37] Juwon Seo, Ji-Su Kang, and Gyeong-Moon Park. Lfs-gan: Lifelong few-shot image generation. In Proceedings of the" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "IEEE/CVF International Conference on Computer Vision (ICCV), pages 11356-11366, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 139 + ], + "type": "text", + "content": "[38] Shawn Shan, Jenna Cryan, Emily Wenger, Haitao Zheng, Rana Hanocka, and Ben Y Zhao. Glaze: Protecting artists from style mimicry by text-to-image models. arXiv preprint arXiv:2302.04222, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 141, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 195 + ], + "type": "text", + "content": "[39] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9243-9252, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 197, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 241 + ], + "type": "text", + "content": "[40] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. Interfacegan: Interpreting the disentangled face representation learned by gans. IEEE transactions on pattern analysis and machine intelligence, 44(4):2004-2018, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 243, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 545, + 297 + ], + "type": "text", + "content": "[41] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Diffusion art or digital forgery? investigating data replication in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6048-6058, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 298, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 298, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 545, + 363 + ], + "type": "text", + "content": "[42] Jingxiang Sun, Xuan Wang, Lizhen Wang, Xiaoyu Li, Yong Zhang, Hongwen Zhang, and Yebin Liu. Next3d: Generative neural texture rasterization for 3d-aware head avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20991-21002, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 365, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 409 + ], + "type": "text", + "content": "[43] Ayush K Tarun, Vikram S Chandawat, Murari Mandal, and Mohan Kankanhalli. Fast yet effective machine unlearning. IEEE Transactions on Neural Networks and Learning Systems, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 411, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 453 + ], + "type": "text", + "content": "[44] Piyush Tiwary, Atri Guha, Subhodip Panda, et al. Adapt then unlearn: Exploiting parameter space semantics for unlearning in generative adversarial networks. arXiv preprint arXiv:2309.14054, 2023. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 456, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 498 + ], + "type": "text", + "content": "[45] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40(4): 1-14, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 500, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 500, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 500, + 545, + 555 + ], + "type": "text", + "content": "[46] Tengfei Wang, Yong Zhang, Yanbo Fan, Jue Wang, and Qifeng Chen. High-fidelity gan inversion for image attribute editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11379-11388, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 557, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 545, + 601 + ], + "type": "text", + "content": "[47] Yue Wu, Yu Deng, Jiaolong Yang, Fangyun Wei, Qifeng Chen, and Xin Tong. Anifacegan: Animatable 3d-aware face image generation for video avatars. Advances in Neural Information Processing Systems, 35:36188-36201, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 601, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 656 + ], + "type": "text", + "content": "[48] Yuting Xu, Jian Liang, Gengyun Jia, Ziming Yang, Yanhao Zhang, and Ran He. Tall: Thumbnail layout for deepfake video detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22658-22668, 2023. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "text", + "content": "[49] Zhiyuan Yan, Yong Zhang, Yanbo Fan, and Baoyuan Wu. Ucf: Uncovering common features for generalizable deepfake detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22412-22423, 2023. 1" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "9160" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 441 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[50] Jaejun Yoo, Youngjung Uh, Sanghyuk Chun, Byeongkyu Kang, and Jung-Woo Ha. Photorealistic style transfer via wavelet transforms. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9036-9045, 2019. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "text", + "content": "[51] Xiaoyong Yuan, Pan He, Qile Zhu, and Xiaolin Li. Adversarial examples: Attacks and defenses for deep learning. IEEE transactions on neural networks and learning systems, 30 (9):2805-2824, 2019. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 174, + 288, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 174, + 288, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 174, + 288, + 229 + ], + "type": "text", + "content": "[52] Ziyang Yuan, Yiming Zhu, Yu Li, Hongyu Liu, and Chun Yuan. Make encoder great again in 3d gan inversion through geometry and occlusion-aware encoding. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2437-2447, 2023. 2, 3, 4, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 230, + 288, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 288, + 272 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 288, + 272 + ], + "type": "text", + "content": "[53] Eric Zhang, Kai Wang, Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Forget-me-not: Learning to forget in text-to-image diffusion models. arXiv preprint arXiv:2303.17591, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 274, + 288, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 274, + 288, + 329 + ], + "spans": [ + { + "bbox": [ + 48, + 274, + 288, + 329 + ], + "type": "text", + "content": "[54] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 331, + 287, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 287, + 385 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 287, + 385 + ], + "type": "text", + "content": "[55] Xuanmeng Zhang, Zhedong Zheng, Daiheng Gao, Bang Zhang, Yi Yang, and Tat-Seng Chua. Multi-view consistent generative adversarial networks for compositional 3d-aware image synthesis. International Journal of Computer Vision, pages 1-24, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 386, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 386, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 48, + 386, + 287, + 441 + ], + "type": "text", + "content": "[56] Xiaoming Zhao, Fangchang Ma, David Guera, Zhile Ren, Alexander G Schwing, and Alex Colburn. Generative multiplane images: Making a 2d gan 3d-aware. In Proceedings of the European Conference on Computer Vision (ECCV), pages 18-35. Springer, 2022. 2" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "9161" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/GenesisTex_ Adapting Image Denoising Diffusion to Texture Space/ba009eff-00cf-485a-8674-0d1aa83b6a01_content_list.json b/2024/GenesisTex_ Adapting Image Denoising Diffusion to Texture Space/ba009eff-00cf-485a-8674-0d1aa83b6a01_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..847bb231798b973888713064ec1d1017ad070d60 --- /dev/null +++ b/2024/GenesisTex_ Adapting Image Denoising Diffusion to Texture Space/ba009eff-00cf-485a-8674-0d1aa83b6a01_content_list.json @@ -0,0 +1,1844 @@ +[ + { + "type": "text", + "text": "GenesisTex: Adapting Image Denoising Diffusion to Texture Space", + "text_level": 1, + "bbox": [ + 148, + 130, + 818, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chenjian Gao $^{1,2\\dagger}$ Boyan Jiang $^{2}$ Xinghui Li $^{2}$ Yingpeng Zhang $^{2\\#}$ Qian Yu $^{1*}$ $^{1}$ School of Software, Beihang University \n $^{2}$ R&D Efficiency and Capability Department, Tencent IEG \nhttps://cjeen.github.io/GenesisTexPaper/", + "bbox": [ + 142, + 178, + 825, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 271, + 286, + 349, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present GenesisTex, a novel method for synthesizing textures for 3D geometries from text descriptions. GenesisTex adapts the pretrained image diffusion model to texture space by texture space sampling. Specifically, we maintain a latent texture map for each viewpoint, which is updated with predicted noise on the rendering of the corresponding viewpoint. The sampled latent texture maps are then decoded into a final texture map. During the sampling process, we focus on both global and local consistency across multiple viewpoints: global consistency is achieved through the integration of style consistency mechanisms within the noise prediction network, and low-level consistency is achieved by dynamically aligning latent textures. Finally, we apply reference-based inpainting and img2img on denser views for texture refinement. Our approach overcomes the limitations of slow optimization in distillation-based methods and instability in inpainting-based methods. Experiments on meshes from various sources demonstrate that our method surpasses the baseline methods quantitatively and qualitatively.", + "bbox": [ + 75, + 316, + 473, + 621 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 647, + 209, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, with the development of deep learning, 3D content generation technology has made significant progress. The applications of 3D content generation are diverse, ranging from AR/VR to gaming and filmmaking. While there has been considerable research on deep learning-based geometric asset generation[40], there has been a notable industry demand for generating realistic textures for given geometries.", + "bbox": [ + 75, + 672, + 468, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, text-conditioned image diffusion models [34] have achieved impressive results in image generation. Some works [7, 19, 29] have leveraged text-conditioned image diffusion models to generate textured 3D assets by gener", + "bbox": [ + 75, + 794, + 468, + 854 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/375f4d1e1db8ab1323d0a75ed48f511ef8e3f009a1755fe6418f02c50d1e724f.jpg", + "image_caption": [ + "red volkswagon beetle, cartoon style", + "Figure 1. Texturing results of different methods. Score Distillation Sampling (SDS) based method produces blurred and oversaturated textures. Inpainting-based approach results in artifacts at the boundaries of inpainting masks. Texture space sampling concurrently generating content from multiple viewpoints, produces clean, clear and natural colored textures." + ], + "image_footnote": [], + "bbox": [ + 503, + 285, + 888, + 532 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ating content from multiple viewpoints, achieving notable performance. In this work, our focus is on generating high-quality textures for a given geometry, leveraging the priors provided by pre-trained text-to-image diffusion models. This task poses several challenges, including: 1) View consistency: ensuring cross-view constraints for maintaining low-level consistency; 2) High efficiency: generating textures for a model within a few minutes, enabling practical applications; 3) Zero-shot learning: achieving texture generation without requiring additional training or finetuning. Addressing these challenges is crucial for successful application of image diffusion models to the texture domain.", + "bbox": [ + 496, + 672, + 892, + 852 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Currently, the most prominent method for text-to-3D generation is Score Distillation Sampling (SDS) [7, 19, 29]. SDS utilizes the prior knowledge from image diffusion", + "bbox": [ + 498, + 854, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{\\dagger}$ Work was done during an internship at Tencent IEG.", + "bbox": [ + 99, + 862, + 383, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Equal contribution.", + "bbox": [ + 102, + 875, + 210, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author.", + "bbox": [ + 102, + 887, + 225, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4620", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "models to iteratively generate gradients for rendering images of 3D objects. However, SDS has limitations in terms of both efficiency and quality. It often takes several hours to generate a single textured 3D object, and the resulting textures can suffer from over-saturation due to the adoption of a large CFG (classifier-free guidance) scale, which reduces randomness during generation. Recent works [2, 6, 33] have introduced inpainting-based methods for multi-view generation. These methods establish a predetermined order of views, using the content of previously generated views as a condition for subsequent views. While inpainting-based approaches are faster and produce realistic colors, they are sensitive to the predefined view order. Additionally, they require texture segmentation to determine the source view for each texture pixel, which often results in artifacts around the borders of the segmented areas.", + "bbox": [ + 75, + 90, + 472, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the aforementioned issues, we propose GenesisTex, a novel approach that introduces texture space sampling. Specifically, we maintain multiple-view latent texture maps throughout the sampling process and perform denoising to improve their quality. Subsequently, the content is decoded from the latent space to obtain RGB textures. Our approach focuses on achieving two aspects of consistency during the sampling process. Firstly, we ensure global style consistency across multiple views by incorporating style consistency in the noise prediction network. This helps maintain a coherent style throughout the generated textures. Secondly, we employ dynamic alignment of latent textures to ensure low-level multi-view consistency, enhancing the overall quality of the generated textures. Due to memory limitations, texture space sampling operates on sparse views. However, to further enhance the quality, we leverage reference-based inpainting and Img2Img techniques on denser views for texture map refinement. GenesisTex can generate detailed, clean, and naturally colored textures for a given geometry within a few minutes. Unlike inpainting-based methods, our proposed texture space sampling concurrently generates content for multiple views without relying on a predefined order. This adaptability makes GenesisTex suitable for various geometries and results in fewer artifacts in the generated textures.", + "bbox": [ + 75, + 335, + 472, + 714 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 94, + 715, + 421, + 729 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- First, we present a novel method for texture generation, where the core is texture space sampling. This sampling technique allows for the concurrent denoising of latent textures associated with multiple viewpoints.", + "- Second, we introduce Style Consistency and Dynamic Alignment in texture space sampling for multi-view consistency.", + "- Third, we conduct a comprehensive study involving numerous 3D objects from various sources. The experimental results demonstrate the superiority of our method over baseline methods." + ], + "bbox": [ + 76, + 734, + 468, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 89, + 640, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Texture Synthesis. Generating textures over 3D surfaces is a challenging problem, as it requires attention to both colors and geometry. Earlier works like AUV-Net [8] and Texturify [38] embed the geometric prior into a UV map or mesh parameterization. Different from them, EG3D [3] and GET3D [11] directly train 3D StyleGANs [15] to generate geometries and textures jointly, where the textures are implicit texture fields. However, these methods either only work on a single category, or demand textured 3D shapes for training without text conditioning, which limits their broad applicability. Recently, [2, 6, 33] use the priors provided by the image diffusion model to synthesize textures. Text2Tex [6] and TEXTure [33] performs inpainting on multiview renderings. TexFusion [2] propose a sequential texture sampling method. Our method predicts multi-view noise concurrently, avoiding the following issues of sequential noise prediction: 1) Each iteration requires adding forward noise to visited regions to match unvisited areas, potentially leading to detail loss. 2) Consistency is only constrained between adjacent viewpoints, with no direct consistency for long-range viewpoints.", + "bbox": [ + 496, + 119, + 893, + 438 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-to-Image Diffusion Models. Over the past years, the development of several large-scale diffusion models [28, 32, 34, 35] has enabled the production of highly detailed and visually impressive images. These models generate images based on input text prompts. Specifically, Stable Diffusion is trained on a substantial text-image dataset [36] and incorporates a text encoder from CLIP [31] to understand the input prompts. Beyond the basic text conditioning, ControlNet [51] enables the model to condition its denoising network on additional input modalities, such as depth maps. In this work, we utilize ControlNet and Stable Diffusion to provide geometrically-conditioned image priors.", + "bbox": [ + 496, + 441, + 893, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-to-3D using 2D Image Diffusion Models. Early works [14, 26, 27] utilize the pretrained CLIP model to maximize the similarity between rendered images and text prompt. However, pioneering works DreamFusion [29] and SJC [46], on the other hand, propose to distill a 2D text-to-image generation model to generate 3D shapes from texts, and many follow-up works [19-21, 23, 25, 30, 47] follow such per-shape optimization scheme. Recently, several methods [4, 12, 13, 22, 24, 37, 42-44, 48] have been proposed to generate consistent multi-view images by using diffusion models. MVDream [37] and SyncDreamer [22] share similar ideas, generating consistent multi-view images via attention layers. However, existing approaches either endure slow optimization processes or depend on separately trained 3D priors, rendering them unsuitable for direct application in texture synthesis. In contrast, our method does not require additional training and can generate results within several minutes.", + "bbox": [ + 496, + 628, + 895, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4621", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 89, + 210, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our objective is to generate a texture map $\\mathcal{T}$ for a given 3D mesh $\\mathcal{M}$ , using the provided text as a descriptive condition. Our approach leverages the Stable Diffusion model as the image diffusion model. We begin by introducing some basic concept about Stable Diffusion and texture representation in Sec. 3.1. Then, we introduce a sampling algorithm in texture space for Stable Diffusion in Sec. 3.2. Furthermore, we propose multi-view consistency strategy in Sec. 3.3. Finally we perform refinement on the texture map, which includes Inpainting and Img2Img, as described in Sec. 3.4.", + "bbox": [ + 75, + 114, + 472, + 267 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 76, + 273, + 207, + 290 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Stable Diffusion. The diffusion model belongs to a class of generative models that generate data through iterative denoising from random noise. In the sampling process, $\\mathbf{z}_{i-1}$ can be sampled from $\\mathbf{z}_i$ using a DDIM sampler [39]:", + "bbox": [ + 75, + 297, + 470, + 359 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\hat {\\mathbf {z}} _ {i \\rightarrow 0} = (\\mathbf {z} _ {i} - \\sqrt {1 - \\alpha_ {i}} \\epsilon_ {\\theta} (\\mathbf {z} _ {i}, t _ {i}, \\mathbf {c} _ {t e x t})) / \\sqrt {\\alpha_ {i}}, \\\\ \\mathbf {z} _ {i - 1} = \\sqrt {\\alpha_ {i - 1}} \\hat {\\mathbf {z}} _ {i \\rightarrow 0} + \\sqrt {1 - \\alpha_ {i - 1} - \\sigma_ {i} ^ {2}} \\epsilon_ {\\theta} (\\mathbf {z} _ {i}, t _ {i}, \\mathbf {c} _ {t e x t}) \\\\ + \\sigma_ {i} \\varepsilon , \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 367, + 468, + 445 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\epsilon_{\\theta}$ is a pretrained U-Net for noise prediction and the initial $\\mathbf{z}_T \\sim \\mathcal{N}(0, \\mathbf{I})$ . In this paper, we specifically employ Stable Diffusion [34]. Stable Diffusion performs denoising in the latent space and employs an autoencoder $\\mathcal{D}(\\mathcal{E}(\\cdot))$ for the conversion between image and latent representations. So $\\mathbf{z}_0$ generated by the diffusion process is decoded to the image $\\mathcal{D}(\\mathbf{z}_0)$ finally.", + "bbox": [ + 75, + 446, + 468, + 550 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ControlNet. ControlNet [51] injects low-level control during the denoising process of Stable Diffusion. Our approach employs depth $\\mathbf{d}$ as the geometric control condition. The predicted noise of the U-Net with ControlNet is represented as $\\epsilon_{\\theta}(\\mathbf{z}_i, t_i, \\mathbf{c}_{text}, \\mathbf{d})$ .", + "bbox": [ + 75, + 551, + 468, + 626 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Texture Representation. In this paper, we utilize mesh as the 3D representation. Texture map is associated with an UV parameterization of the mesh $\\mathcal{M}$ . We use $x\\text{atlas}$ [50] to generate the UV parameterization.", + "bbox": [ + 75, + 627, + 468, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "**Rendering.** Given a mesh $\\mathcal{M}$ , a texture map $\\mathcal{T}$ and a viewpoint $C$ , we can use the rendering function $\\mathcal{R}$ to get the rendered image $\\mathbf{x}^{(img)} = \\mathcal{R}(\\mathcal{T};\\mathcal{M},C)$ . Similar to [2], we do not consider any lighting and rendering involves solely sampling colors from the texture map. The inverse rendering function $\\mathcal{R}^{-1}$ can inverse render the image $\\mathbf{x}^{(img)}$ back to a texture map $\\mathcal{T}' = \\mathcal{R}^{-1}(\\mathbf{x}^{(img)};\\mathcal{M},C)$ .", + "bbox": [ + 75, + 686, + 468, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Sampling in Texture Space", + "text_level": 1, + "bbox": [ + 76, + 801, + 318, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The most straightforward method for sampling on the texture map is to fine-tune the Stable Diffusion model directly on some texture maps, making its distribution more like the distribution of texture maps. However, the distribution of texture maps is much more complex than that of rendering", + "bbox": [ + 75, + 824, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Texture Space Sampling" + ], + "code_body": "Input: mesh $\\mathcal{M}$ ,cameras $\\{C_1,\\dots ,C_N\\}$ \nParameters:Denoising time schedule $\\{t_i\\}_{i = T}^0$ ,DDIM \nnoise schedule $\\{\\sigma_i\\}_{i = T}^0$ $\\{\\mathbf{z}_{T,n}^{(tex)}\\}_{n = 1}^N\\gets \\{\\mathbf{0}\\}$ $\\{\\mathbf{z}_n^{(bg)}\\}_{n = 1}^N\\sim \\{\\mathcal{N}(\\mathbf{0},\\mathbf{I})\\}$ $\\{\\hat{\\epsilon}_{T + 1,n}\\}_{n = 1}^N\\sim \\{\\mathcal{N}(\\mathbf{0},\\mathbf{I})\\}$ \ndenoising stage: \nfor $i\\in \\{T\\dots 1\\}$ do \nfor $n\\in \\{1\\dots N\\}$ do \n $\\varepsilon_{n}\\sim \\mathcal{N}(0,I)$ ${\\bf z}_{i,n}^{(img)}\\gets {\\bf M}_n(\\sqrt{\\alpha_i}{\\cal R}({\\bf z}_{i,n}^{(tex)}) + \\sqrt{1 - \\alpha_i - \\sigma_{i + 1}^2}\\hat{\\epsilon}_{i + 1,n})$ $+(1 - \\mathbf{M}_n)\\mathbf{z}_n^{(bg)} + \\sigma_{i + 1}\\varepsilon_n$ $\\hat{\\epsilon}_{i,n}\\gets \\epsilon_{\\theta}(\\mathbf{z}_{i,n}^{(img)},t_i,\\mathbf{c},\\mathbf{d}_n)$ $\\hat{\\mathbf{z}}_{0,n}^{(img)}\\gets (\\mathbf{z}_{i,n}^{(img)} - \\sqrt{1 - \\alpha_i}\\hat{\\epsilon}_{i,n}) / \\sqrt{\\alpha_i}$ ${\\bf z}_{i - 1,n}^{(tex)}\\gets \\mathcal{R}^{-1}(\\hat{\\mathbf{z}}_{0,n}^{(img)})$ ${\\bf z}_{n}^{(bg)}\\gets \\sqrt{\\alpha_{i - 1}}\\hat{\\mathbf{z}}_{0,n}^{(img)} + \\sqrt{1 - \\alpha_{i - 1} - \\sigma_i^2}\\cdot \\hat{\\epsilon}_{i,n}$ \nend for \n $\\{{\\bf z}_{i - 1,n}^{(tex)}\\}_{n = 1}^N = dynamic\\_ align(\\{{\\bf z}_{i - 1,n}^{(tex)}\\}_{n = 1}^N)$ \nend for \ndecoding stage: \nfor $n\\in \\{1\\dots N\\}$ do \n ${\\bf x}_n^{(img)} = \\mathcal{D}(\\hat{\\bf z}_{0,n}^{(img)})$ ${\\bf x}_n^{(tex)} = R^{-1}({\\bf x}_n^{(img)})$ \nend for \nT=merge({x(tex)})N \nreturn Texture map Tsampled", + "bbox": [ + 500, + 107, + 924, + 550 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "images since different UV parameterizations of the same 3D object correspond to different texture maps. Instead of fine-tuning existing image diffusion models, we adapt them to the texture space for texture map generation.", + "bbox": [ + 496, + 578, + 890, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Algorithm 1 presents our proposed texture space sampling algorithm. Firstly we define a set of $\\mathcal{C}^{(sampling)} = \\{C_1,..,C_N\\}$ camera views and render the corresponding depth map $\\mathbf{d}_n$ in each view. Since we employ a latent diffusion model, our texture space sampling method first performs denoising in the latent texture space and then decode the latent to texture image. In the denoising stage, a dedicated diffusion denoising algorithm is applied to a set of latent texture maps $\\{\\mathbf{z}_n^{(tex)}\\}_{n=1}^N$ . In the decoding stage, we recover the texture map $\\mathcal{T}_{sampled}$ from the latent space.", + "bbox": [ + 496, + 638, + 892, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Denoising Stage. In the denoising stage, we adapt the DDIM from the rendering space to the texture space and utilize the same parameters as DDIM, including denoising time schedule $\\{t_i\\}_{i = T}^0$ , DDIM noise schedule $\\{\\sigma_i\\}_{i = T}^0$ . For each viewpoint, we maintain a separate latent texture which dynamically interact during the denoising process. We begin with $\\{\\mathbf{z}_{T,n}^{(tex)}\\}_{n = 1}^{N} = \\{\\mathbf{0}\\}$ , the zero latent texture map", + "bbox": [ + 496, + 792, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4622", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9b1d8a67dfd9b827f94980ff551b3ef6522b236bf388670b26a3d08c2d688ffe.jpg", + "image_caption": [ + "Figure 2. Overview of GenesisTex. GenesisTex generates a texture map for a given mesh $\\mathcal{M}$ , based on a prompt. Texture Space Sampling samples a texture map using Stable Diffusion, introducing style consistency and dynamic alignment across multiple viewpoints. Furthermore, Inpainting and Img2Img are applied to fill in the blank regions and enhance the quality of texture map details, respectively." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 390, + 281 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f658c0f6c5646555c3e8d6eb1749aab590aa4831ff414b7bc70c790dfa9b67ce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 392, + 89, + 604, + 281 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/412af3afeb20fccbaee64f218c7e5ed5a7547a1fa13126e6915321f0c9f92dd5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 609, + 89, + 890, + 199 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/63dcb9d3bf64b3fd97c2ec66455bc3322d028e86115878d631fedd9975f57a3a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 609, + 202, + 890, + 281 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "corresponding to $N$ viewpoints. At each denoising step, our goal is to predict $\\mathbf{z}_{i - 1,n}^{(tex)}$ from $\\mathbf{z}_{i,n}^{(tex)}$ . We first render the latent texture map $\\mathbf{z}_{i,n}^{(tex)}$ to the rendering space using the rendering function $\\mathcal{R}$ , and then calculate the corresponding latent image $\\mathbf{z}_{i,n}^{(img)}$ :", + "bbox": [ + 75, + 364, + 468, + 453 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {z} _ {i, n} ^ {(i m g)} = \\mathbf {M} _ {n} (\\sqrt {\\alpha_ {i}} \\mathcal {R} (\\mathbf {z} _ {i, n} ^ {(t e x)}) + \\sqrt {1 - \\alpha_ {i} - \\sigma_ {i + 1} ^ {2}} \\hat {\\epsilon} _ {i + 1, n}) \\\\ + \\left(1 - \\mathbf {M} _ {n}\\right) \\mathbf {z} _ {n} ^ {(b g)} + \\sigma_ {i + 1} \\varepsilon_ {n} \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 83, + 465, + 468, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\mathbf{M}_n$ represents the foreground mask for viewpoint $C_n$ (downsampled to the same resolution as $\\mathbf{z}^{(img)})$ , where foreground pixels have a value of 1, and background pixels have a value of 0. $\\hat{\\epsilon}_{i + 1,n}$ represents the noise predicted by the noise prediction network at step $i + 1\\rightarrow i$ (notably, we define $\\hat{\\epsilon}_{T + 1,n}\\sim \\mathcal{N}(0,\\mathbf{I})$ ). $\\mathbf{z}_n^{(bg)}$ is the background of the latent image, which is set to Gaussian noise at initialization. $\\varepsilon_{n}\\sim \\mathcal{N}(0,\\mathbf{I})$ is a random Gaussian noise. Subsequently, we predict the corresponding noise $\\hat{\\epsilon}_{i,n} = \\epsilon_{\\theta}(\\mathbf{z}_{i,n}^{(img)},t_i,\\mathbf{c},\\mathbf{d}_n)$ for the latent image $\\mathbf{z}_{i,n}^{(img)}$ conditioned on the depth map $\\mathbf{d}_n$ . With the predicted noise, we can obtain the predicted $\\hat{\\mathbf{z}}_{0,n}^{(img)}$ :", + "bbox": [ + 75, + 523, + 468, + 715 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {z}} _ {0, n} ^ {(i m g)} = (\\mathbf {z} _ {i, n} ^ {(i m g)} - \\sqrt {1 - \\alpha_ {i}} \\hat {\\epsilon} _ {i, n}) / \\sqrt {\\alpha_ {i}}, \\qquad (3)\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 726, + 468, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, we inverse render $\\hat{\\mathbf{z}}_{0,n}^{(img)}$ to the texture space using $\\mathcal{R}^{-1}$ and update $\\mathbf{z}_n^{(bg)}$ :", + "bbox": [ + 76, + 761, + 468, + 799 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {i - 1, n} ^ {(t e x)} = \\mathcal {R} ^ {- 1} \\left(\\hat {\\mathbf {z}} _ {0, n} ^ {(i m g)}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 809, + 468, + 840 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {n} ^ {(b g)} = \\sqrt {\\alpha_ {i - 1}} \\hat {\\mathbf {z}} _ {0, n} ^ {(i m g)} + \\sqrt {1 - \\alpha_ {i - 1} - \\sigma_ {i} ^ {2}} \\cdot \\hat {\\epsilon} _ {i, n}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 830, + 421, + 859 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the denoising stage, we repeat the above process $T - 1$ times. After each denoising step, we impose consistency", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "constraints on latent texture $\\mathbf{z}_{i - 1,n}^{(tex)}$ , which will be introduced in Sec. 3.3. Notably, differing from DDIM where the latent includes noise, the latent texture in our method is an estimate of the noise-free version. This is because $\\mathcal{R}$ and $\\mathcal{R}^{-1}$ in the denoising process require interpolation of the latent, which may distort the noise.", + "bbox": [ + 496, + 364, + 890, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Decoding Stage. To recover the texture map $\\mathcal{T}_{\\text{sampled}}$ , we first decode each viewpoint's latent separately to obtain multi-view rendering. Subsequently, we inverse render the renderings into the texture space. After obtaining multiview texture maps, we merge them into a single texture map as the output. Similar to [2, 6, 33], we aim for each pixel on each texture map to come from the viewpoint where the corresponding point on the mesh is observed most directly. Therefore, our merge is defined as follows:", + "bbox": [ + 496, + 455, + 890, + 592 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {T} _ {\\text {s a m p l e d}} = \\sum_ {n = 1} ^ {N} \\operatorname {S o f t m a x} \\left(\\mathcal {R} ^ {- 1} \\left(\\mathbf {N} _ {\\mathbf {n}}\\right)\\right) \\times \\mathbf {x} _ {n} ^ {(t e x)} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 604, + 890, + 646 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathbf{N}_{\\mathbf{n}}$ is the similarity mask at viewpoint $C_n$ , where each pixel represents the cosine similarity between the normal vectors of the visible faces and the reversed view direction. We use Softmax instead of Max because Softmax can result in more natural transition in the texture maps.", + "bbox": [ + 496, + 657, + 890, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Consistency between Latent Texture Maps", + "text_level": 1, + "bbox": [ + 500, + 742, + 861, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1 Dynamic Alignment", + "text_level": 1, + "bbox": [ + 500, + 766, + 692, + 781 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During sampling in texture space, we maintain $N$ latent texture maps, now we introduce an alignment approach to ensure their local consistency. Firstly, we directly reduce these latent texture maps to obtain a uniform latent texture map:", + "bbox": [ + 496, + 790, + 890, + 851 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {i - 1} ^ {(u n i t e x)} = \\sum_ {n = 1} ^ {N} \\operatorname {S o f t m a x} \\left(\\mathcal {R} ^ {- 1} \\left(\\mathbf {N} _ {n}\\right)\\right) \\times \\mathbf {z} _ {i - 1, n} ^ {(t e x)} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 862, + 890, + 904 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4623", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Next, we blend each viewpoint's latent texture map $\\mathbf{z}_{i-1,n}^{(tex)}$ with the uniform latent texture map $\\mathbf{z}_{i-1}^{(unitex)}$ . We do not enforce strict alignment among the $N$ latent texture maps at every step. Since a latent pixel gains full meaning only when combined with the surrounding context, which is viewpoint-dependent, the latent texture map from each viewpoint should contain some independent information. Moreover, the content generated at different timestamps varies, so the alignment constraints should adapt flexibly. We introduce a signal $c(t)$ for dynamically regulating the variations in consistency during the sampling process:", + "bbox": [ + 76, + 88, + 472, + 263 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {i - 1, n} ^ {(c o n s i s t e x)} = c (t _ {i}) \\times \\mathbf {z} _ {i - 1} ^ {(u n i t e x)} + (1 - c (t _ {i})) \\times \\mathbf {z} _ {i - 1, n} ^ {(t e x)} (7)\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 273, + 468, + 296 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We empirically find that maintaining strong local consistency in the mid-term of denoising, while keeping weak local consistency in the early and late stages, leads to higher quality generation results.", + "bbox": [ + 76, + 305, + 468, + 367 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.2 Style Consistency", + "text_level": 1, + "bbox": [ + 76, + 385, + 253, + 401 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In texture space sampling, we use the noise estimation network $\\epsilon_{\\theta}$ to estimate noise for the multi-view latents, which is responsible for content generation. So we adapt $\\epsilon_{\\theta}$ to ensure style consistency of estimated noise across different views. Inspired by video diffusion models [16, 49] and multi-view diffusion model [37], we modify all the self-attention layers and group normalization layers in the noise prediction network to align the style of multi-view content. Adapted Self-Attention. In Stable Diffusion, self-attention facilitates long-range interactions among features within an image. We transform self-attention into cross-view attention to establish style consistency across multiple viewpoints, by using the key $\\mathbf{K}^{\\prime}$ and value $\\mathbf{V}^{\\prime}$ from all viewpoints,", + "bbox": [ + 76, + 410, + 468, + 619 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {K} ^ {\\prime} = \\mathbf {W} ^ {K} \\left[ \\mathbf {z} _ {i, 1} ^ {(i m g)}; \\dots ; \\mathbf {z} _ {i, n} ^ {(i m g)} \\right], \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 621, + 468, + 647 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {V} ^ {\\prime} = \\mathbf {W} ^ {V} [ \\mathbf {z} _ {i, 1} ^ {(i m g)}; \\dots ; \\mathbf {z} _ {i, n} ^ {(i m g)} ].\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 646, + 370, + 664 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{W}^K$ , $\\mathbf{W}^V$ are pretrained parameters.", + "bbox": [ + 76, + 671, + 362, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Adapted Group Normalization. Stable diffusion adopt group normalization as the normalization layer. Group normalization divides the channels into groups and computes within each group the mean and variance for normalization. We convert 2D group normalization to 3D by connecting all different views in the group normalization layer, i.e., the mean and variance are calculated within the channel group of all pixels across all viewpoints.", + "bbox": [ + 75, + 688, + 468, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By adjusting self-attention layer and group normalization layer, we achieve multi-view style consistency without any additional training. In addition to the noise prediction network $\\epsilon_{\\theta}$ , we apply the same modifications to the decoder $\\mathcal{D}$ to further ensure style consistency in the multi-view decoded RGB images.", + "bbox": [ + 76, + 810, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Texture Map Refinement", + "text_level": 1, + "bbox": [ + 500, + 90, + 728, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4.1 Post-processing: Inpainting", + "text_level": 1, + "bbox": [ + 500, + 114, + 746, + 131 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since the memory cost of texture space sampling increases with the number of viewpoints of $\\mathcal{C}_{\\text{sampling}}$ , the number of viewpoints $N$ is limited. Some regions in the texture map $\\mathcal{T}_{\\text{sampled}}$ may not be observed in any of the viewpoints in $\\mathcal{C}_{\\text{sampling}}$ . Therefore, after texture space sampling, we introduce an inpainting epoch to fill texture in areas not observed in $\\mathcal{C}$ .", + "bbox": [ + 496, + 138, + 893, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Firstly we define a denser set of viewpoints $\\mathcal{C}^{(inpainting)} = \\{C_1,..,C_{N_1}\\}$ as compared to $\\mathcal{C}^{(samping)}$ where $N_{1} > N$ . We use a mask $\\mathbf{M}_C^{blank}$ to indicate the rendered areas with blank textures at the viewpoint $C$ . The entire inpainting epoch comprises $N_{1}$ iterations. In each iteration, we first compute the viewpoint $C_n = \\arg \\max_C\\mathbf{M}_C^{blank}$ with the largest blank area. Then, we perform inpainting on the rendering $\\mathbf{x}_n^{(img)}$ using the pretrained depth conditioned Stable Diffusion model as described in [6, 33]. During inpainting, Stable Diffusion utilizes regions with textures as conditions to fill in blank areas without textures. To enhance the robustness of inpainting, we render an image using a viewpoint $C_0$ from the set $\\mathcal{C}^{(sampling)}$ . We then concatenate this image to the image requiring inpainting as a additional condition for inpainting. Compared to naive inpainting, reference-based inpainting provides a more comprehensive condition. To ensure a natural and smooth transition between blank areas and textured regions, we apply Gaussian blur to the mask $\\mathbf{M}_{C_n}^{blank}$ . Finally the inpainted rendering $\\mathbf{x}_n^{(img)'}$ is used to update the texture map:", + "bbox": [ + 496, + 246, + 892, + 573 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {T} _ {n} = \\mathcal {T} _ {n - 1} \\times (1 - \\mathcal {R} ^ {- 1} (\\mathbf {M} ^ {(b l a n k)}) \\tag {9} \\\\ + \\mathcal {R} ^ {- 1} \\left(\\mathbf {x} _ {n} ^ {(i m g) \\prime}\\right) \\times \\mathcal {R} ^ {- 1} \\left(\\mathbf {M} ^ {(b l a n k)}\\right) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 585, + 890, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We start with $\\mathcal{T}_0 = \\mathcal{T}_{\\text{sampled}}$ and iterate for $N_1$ times. Finally we get $\\mathcal{T}_{\\text{inpainted}} = \\mathcal{T}_{N_1}$ .", + "bbox": [ + 496, + 638, + 890, + 669 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It's worth noting that, unlike [2, 6, 33], which relies heavily on inpainting to maintain multi-view consistency, in our method inpainting is performed only to patch small areas. Therefore, the inpainting's performance does not significantly affect the final quality of the generated texture.", + "bbox": [ + 496, + 670, + 890, + 746 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4.2 Post-processing:Img2Img", + "text_level": 1, + "bbox": [ + 500, + 768, + 736, + 784 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "So far our generated texture map $\\mathcal{T}_{\\text {inpainted }}$ may still contain unnatural transitions in some regions due to multi-view conflicts and inpainting artifacts. To further improve the generation quality, we introduce an Img2Img epoch.", + "bbox": [ + 496, + 794, + 890, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Img2Img epoch, we also define a set of viewpoints $C^{(img2img)}$ include $N_{2}$ viewpoints and iterate through viewpoint from $C^{(img2img)}$ . In each iteration, we encode", + "bbox": [ + 500, + 854, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4624", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the rendering image $\\mathbf{x}_n^{(img)}$ to the latent space using the encoder $\\mathcal{E}$ and add some noise to the latent rendering. Then we perform denoising using the pretrained depth conditioned Stable Diffusion model. And finally we decode the denoised latent back to image space using $\\mathcal{D}$ . However, $\\mathcal{D}(\\mathcal{E}(\\cdot))$ is not an identity transformation because of the reconstruction distortion of the autoencoder. So we apply the color histogram matching to the foreground of the decoded image to alleviate the color distortion. Finally we update the texture map using the img2img result $\\mathbf{x}_n^{(img)'}$ based on view similarity $\\mathbf{N}_n$ :", + "bbox": [ + 75, + 89, + 472, + 261 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {T} _ {n} = \\mathcal {T} _ {n - 1} \\times (1 - \\mathcal {R} ^ {- 1} (\\mathbf {N} _ {n})) \\\\ + \\mathcal {R} ^ {- 1} \\left(\\mathbf {x} _ {n} ^ {(i m g) \\prime}\\right) \\times \\mathcal {R} ^ {- 1} \\left(\\mathbf {N} _ {n}\\right) \\tag {10} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 160, + 271, + 467, + 310 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We start with $\\mathcal{T}_0 = \\mathcal{T}_{\\text{inpainted}}$ . After $N_2$ iterations we get $\\mathcal{T}_{\\text{refined}} = \\mathcal{T}_{N_2}$ .", + "bbox": [ + 76, + 321, + 468, + 354 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 367, + 209, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Setup", + "text_level": 1, + "bbox": [ + 76, + 392, + 158, + 409 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. Our experiments are conducted on an NVIDIA A10 GPU. We utilize the official Stable Diffusion-v1.5 model with the ControlNet-v1.1 (depth). For Inpainting epoch and Img2Img epoch, we use DDIM [39] as the sampler. For all samplers, we set the number of iterations to 20 steps and the CFG scale (classifier-free guidance scale) to 7.5. We implement the rendering function using nvdiffrast [17] and make modifications to nvdiffrast to support inverse rendering. Texture space sampling takes about 3 minutes to generate textures for a single object. For more detailed parameter settings, please refer to the supplementary materials.", + "bbox": [ + 75, + 415, + 468, + 598 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. Similar to TexFusion [2], we utilize 35 meshes to test the texture generation performance, including 11 objects from Objaverse [10], 17 from ShapeNet [5], 1 from Text2Mesh [26], 2 from Turbosquid [1], 2 from Stanford 3D Scans [9, 45] and 2 from Three D Scans [18]. Each object has 1-4 text descriptions, resulting in a total of 80 (mesh, prompt) pairs in this collection.", + "bbox": [ + 75, + 598, + 468, + 703 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We conduct comparisons with state-of-the-art methods for generating textures from text: 1) Text2Mesh [26], a method that stylizes a 3D mesh by predicting color and local geometric details which conform to a target text prompt, harnessing the representational power of CLIP. 2) Latent-Paint [25], an approach leveraging SDS to obtain texture gradients from the Stable Diffusion model. 3) Text2Tex [6], a method that uses a depth-aware image inpainting diffusion model to incrementally produce partial textures from various viewpoints. 4) TEXTure [33], similar to Text2Tex, but utilizes a different region segmentation strategy. 5) TexFusion [2], a method sequentially inpainting latent texture. As TexFusion does not have publicly", + "bbox": [ + 75, + 704, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/92f9dcb881b28ce9d41bd7eddba908fba16b8ed645c77f3957ecfe4ba76997f6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodFID (↓)KID (↓) (×10-3)User study (%)
Visual Quality (↑)Align with Prompt (↑)
Latent-Paint110.1410.645.154.28
Text2Mesh121.6115.133.154.86
Text2Tex101.388.3518.2819.72
TEXTure100.479.2223.4224.86
Ours74.582.8950.0046.28
", + "bbox": [ + 504, + 88, + 898, + 213 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0d16068cfd6512d63d1279703ca5186d50b291f714f54b3564e89212c2533e75.jpg", + "table_caption": [ + "Table 1. Quantitative comparisons with baseline methods." + ], + "table_footnote": [], + "table_body": "
texture space samplingInpainting RoundImg2Img RoundFID(↓)KID(↓) (×10-3)
✓✗✗86.253.79
✓✗✓82.053.67
✓✓✗76.303.29
✓✓✓74.582.89
", + "bbox": [ + 504, + 253, + 898, + 361 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Effectiveness of texture refinement.", + "bbox": [ + 562, + 372, + 826, + 385 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "available source code, we extract some rendered images of generated results from their paper for comparison.", + "bbox": [ + 498, + 405, + 892, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Qualitative Analysis", + "text_level": 1, + "bbox": [ + 498, + 446, + 692, + 464 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 3 and Fig. 5 presents visual comparisons between our method and baseline approaches. Textures produced by Text2Mesh are notably lacking in detail. Latent-Paint yields textures with greater intricacy compared to Text2Mesh, but they still exhibit considerable blurriness due to SDS limitations. Inpainting-based approaches like Text2Tex and TEXTure introduce artifacts with unnatural transitions at inpainting mask boundaries, struggling to maintain consistency across multiple views because they depend heavily on inpainting. The output of TexFusion is compromised by blurry details, likely a consequence of repeated noise addition during sequential sampling. In contrast, our method stands out by producing textures that are not only clear but also remarkably coherent.", + "bbox": [ + 496, + 470, + 892, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 4 presents more results of our method. Benefiting from the powerful prior of Stable Diffusion, our method can generate textures of diverse styles for different geometries.", + "bbox": [ + 496, + 683, + 890, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Quantitative Comparisons", + "text_level": 1, + "bbox": [ + 498, + 739, + 738, + 757 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "FID & KID. Similar to TexFusion, we sample from pretrained image diffusion model to create ground truth labels. For each mesh, we generate depth maps from the five most common canonical viewpoints: front, back, top, and both sides. These depth maps, along with textual descriptions, serve as inputs to condition the Stable Diffusion model. To ensure the focus remains on the texture of the objects, we modify the background of the ground truth images to be white. We visualize the textures created by various methods", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4625", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fdc48c32a5f11f589a8403e43677db985a342b561a9f5fd1628db52124b015d8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 138, + 90, + 250, + 167 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d2b4d22b3d333ef1d7edae3141d9c9771ecd38c9bd6184a063e54638a0195a46.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 155, + 179, + 261, + 276 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cd9686633cb48e567e402d2f5a1e66ca050d98576870fb1a65fc642231d7b828.jpg", + "image_caption": [ + "red apple blue handbag with dilapidated Halloween silver trims tap pumpkin head" + ], + "image_footnote": [], + "bbox": [ + 137, + 277, + 238, + 376 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f66f2cec905dfc5ae96e8bae3daf9f318f78329bcf9d2dfd3cb9c6e61c7d4c31.jpg", + "image_caption": [ + "Geometry" + ], + "image_footnote": [], + "bbox": [ + 143, + 381, + 246, + 460 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3a88db337c4d909230a438642665240e74ccf6d986ec7814a7af27b32287adb9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 256, + 90, + 364, + 166 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3b06058ff15f1977324d52c080347bdcbe8f500089ab05a2202185df47364ca5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 179, + 379, + 276 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8ff9f5e1793f898017a1bc68710f842a758019abbf69896eef9bea814761c1eb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 254, + 277, + 356, + 375 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7945def590a373bcbc75c0fb67f16d2b1e5133ea559d029245dda8edbc2b3f83.jpg", + "image_caption": [ + "Text2Mesh" + ], + "image_footnote": [], + "bbox": [ + 261, + 381, + 362, + 460 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/67daf3b8d0a1cb7133e37f85e82ba0b30d3bdee3f6a8d82e2941c5c70d1d072a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 385, + 90, + 493, + 166 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/83261be0dd824e36413c37d8e7f574daa425ea68b3d4c2e6ce7795bd49e806ed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 403, + 179, + 509, + 276 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1f9823e5d109eb8e75ffdc9aa1dc3372594c171c1aa74e9150ac57c7d2e243ac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 277, + 483, + 375 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5331997ccdbd20e40b77c9a5cf15142ddbcfb028f99c89ddc14fdf87e38a3435.jpg", + "image_caption": [ + "Latent-Paint" + ], + "image_footnote": [], + "bbox": [ + 390, + 381, + 491, + 460 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3009c04219e35ab44fb8edcd5e209d390537087a45da4d44ac34789bc6d14312.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 522, + 90, + 630, + 166 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/daa7fdce027551be5438c7875d2230a42ee1ca694bcf779108c9ed074e7e7d43.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 179, + 643, + 276 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/98524417bb1e4c9adffb95fab197ffe25236610af7188078f2466a75d1613fbf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 277, + 620, + 375 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/096a9d24c767616a95b05312f84c308ec0cc7a2a557e4f4e1fd4bbdfbb47fefd.jpg", + "image_caption": [ + "Text2Tex" + ], + "image_footnote": [], + "bbox": [ + 526, + 381, + 629, + 460 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fa07765acbd0f50df1805ac367794bb54fddb0f5c3e57ea338641640a0a91a20.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 653, + 90, + 758, + 166 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/edcbf3c8900402c219af8ede23e95f6fbb7976280aef88e686e3539faeedf629.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 179, + 772, + 276 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/828c1245e3f85806bd29afdf6fc9b3679e5ae7c2cd69dec2ef792b8a82e60d95.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 648, + 277, + 750, + 375 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a72a7a4f53dee22f4d9d2c6be81a6962747f89409f477c43e87c5a19e303c1c4.jpg", + "image_caption": [ + "TEXTure" + ], + "image_footnote": [], + "bbox": [ + 656, + 381, + 756, + 460 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/07235b9d9c58d72508c23fdf78daa077cece9bc6d9679d97f5ea04b4883e8476.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 769, + 90, + 874, + 166 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4c567eede86508bf4de02ef0a5b81d56a68996781f88cae11fb1fb0c88614f37.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 785, + 179, + 890, + 276 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e84a26487879ef9aa4a5b34d08dfb19ae9962f45351b79752c93cef507cb3626.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 766, + 277, + 864, + 375 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/178bf64a2ad0b905a0aa74428bfc2d9d09743d1c125800715eb0b4a273a7de53.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 772, + 381, + 874, + 460 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5687d71dca2d7f983a1e3308812f90633d508936eed8e3c02bdf84e0086618b8.jpg", + "image_caption": [ + "Figure 3. Qualitative comparisons with Text2Mesh [26], Latent-Paint [25], Text2Tex [6] and TEXTure [33]. In comparison with the baselines, our GenesisTex exhibits richer details and fewer artifacts.", + "Figure 4. Texturing results with GenesisTex." + ], + "image_footnote": [], + "bbox": [ + 81, + 554, + 467, + 752 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "from the same five viewpoints and evaluate their quality by calculating the Fréchet Inception Distance (FID) and Kernel Inception Distance (KID). FID and KID measure the feature dissimilarity between two image collections, with feature extraction performed using the Inception V3 [41]. The results in Tab.1 demonstrate that the textures gener", + "bbox": [ + 75, + 810, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/da735ee0a098b3bb990e17955ad314fed34a91df832ffb3f5f255f720641bce9.jpg", + "image_caption": [ + "Figure 5. Qualitative Comparisons with TexFusion [2]. Images are extracted from their original paper." + ], + "image_footnote": [], + "bbox": [ + 506, + 554, + 880, + 726 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ated by GenesisTex are closer to the ground truth compared to those produced by baselines, indicating a superior synthesis quality.", + "bbox": [ + 496, + 794, + 890, + 840 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "User Study. We conduct a user study to compare the quality of textures generated by our method versus those produced by baseline methods. We render all textured mesh results into videos to facilitate the presentation to users. Each", + "bbox": [ + 496, + 840, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4626", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "video contains a rotating textured mesh along with the corresponding prompt. For each questionnaire, we randomly show the users 10 groups of rendered videos. Each group displays results for a specific (mesh, prompt) pair, comparing outputs from baseline methods and our approach. Each volunteer is required to answer two questions for a single group, including which one has the highest visual quality and which one aligns best with the prompt. Additional information about the questionnaire is available in the supplementary. Finally, we received 35 valid responses from the questionnaires, which included the results of comparisons from 350 groups. The results of the user study are shown in Tab. 1. In terms of both visual quality and alignment with the text prompt, our method is preferred by the most participants, with percentages reaching $50.00\\%$ and $46.28\\%$ , respectively. Furthermore, we conduct a pairwise comparison study and the results are included in the supplementary.", + "bbox": [ + 75, + 90, + 472, + 349 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 359, + 243, + 376 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Consistency in Texture Space Sampling. Our GenesisTex employs style consistency and dynamic alignment for maintaining multi-view consistency during texture space sampling. To investigate the impact of these two consistency strategies, we visualize the decoded multi-view images. We set $\\mathcal{C}(elevation, azimuth) = \\{(0^{\\circ}, 0^{\\circ}), (0^{\\circ}, 15^{\\circ}), (0^{\\circ}, 35^{\\circ}), (0^{\\circ}, 45^{\\circ})\\}$ to ensure that multiple viewpoints share sufficient content. Fig. 6 illustrates an example with the prompt penguin toy. We can observe that without any consistency constraints, the color of the toy varies significant across different viewpoints. After introducing style consistency, consistency for the toy color improves, but there are still inconsistencies in the hair and ears. Further we use dynamic alignment and we can see the", + "bbox": [ + 75, + 383, + 473, + 597 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/78b4fee5eb22640a4d3f53733c8ef35dcc4c98e20a3fba639f6199fc3b4685ff.jpg", + "image_caption": [ + "Figure 6. Ablation results on consistency in texture space sampling. Style Consistency + Dynamic Alignment (bottom row) achieves the best multi-view consistency." + ], + "image_footnote": [], + "bbox": [ + 81, + 621, + 464, + 843 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0be7867152068e75ea0142d33a1c55951a20f0b7e89b2e58cfcc6c9c45a70819.jpg", + "image_caption": [ + "Figure 7. Ablation results on texture refinement. Inpainting fills the blank regions, while Img2Img enhances the quality of details." + ], + "image_footnote": [], + "bbox": [ + 506, + 92, + 890, + 294 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "consistency of details has significantly improved. We include more results in the supplementary to further demonstrate the effectiveness of these two consistencies.", + "bbox": [ + 496, + 368, + 892, + 412 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of Texture Refinement. After texture space sampling, we conduct refinement on the texture, including Inpainting and Img2Img. We validate the effectiveness of these two modules and the results are presented in Fig. 7 and Tab. 2. Inpainting fills in areas lacking texture obtained during texture space sampling, contributing significantly to the improvement in visual quality. Img2Img further repairs artifacts and enhances details, leading to a certain improvement in visual quality as well.", + "bbox": [ + 496, + 416, + 893, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Limitations", + "text_level": 1, + "bbox": [ + 500, + 574, + 756, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose a novel method for text-based texture generation, named GenesisTex. GenesisTex leverages the prior of the pre-trained Stable Diffusion model by introducing texture space sampling. Texture space sampling concurrently generates multi-view content without relying on a predefined sequence of views. Our approach can generate high-quality textures for a given 3D model in several minutes. The primary limitation of our method is the significant memory cost associated with maintaining style consistency, caused by cross-view attention. This limitation restricts the number of viewpoints and necessitates post-processing steps, such as inpainting and img2img. Future work could investigate hierarchical style consistency approaches to reduce the computational costs of cross-view attention by iterating over a smaller set of viewpoints.", + "bbox": [ + 496, + 602, + 890, + 828 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement: This work is supported by the Young Elite Scientists Sponsorship Program by CAST (China Association for Science and Technology) and IEG Moonshot Program by Tencent.", + "bbox": [ + 496, + 829, + 893, + 887 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4627", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Turbosquid by shutterstock. In www.turbosquid.com. 6", + "[2] Tianshi Cao, Karsten Kreis, Sanja Fidler, Nicholas Sharp, and Kangxue Yin. Texfusion: Synthesizing 3d textures with text-guided image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4169-4181, 2023. 2, 3, 4, 5, 6, 7", + "[3] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 2", + "[4] Eric R Chan, Koki Nagano, Matthew A Chan, Alexander W Bergman, Jeong Joon Park, Axel Levy, Miika Aittala, Shalini De Mello, Tero Karras, and Gordon Wetzstein. Generative novel view synthesis with 3d-aware diffusion models. arXiv preprint arXiv:2304.02602, 2023. 2", + "[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 6", + "[6] Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text-driven texture synthesis via diffusion models. arXiv preprint arXiv:2303.11396, 2023. 2, 4, 5, 6, 7", + "[7] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. *Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation.* arXiv preprint arXiv:2303.13873, 2023. 1", + "[8] Zhiqin Chen, Kangxue Yin, and Sanja Fidler. Auv-net: Learning aligned uv maps for texture transfer and synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1465-1474, 2022. 2", + "[9] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 303-312, 1996. 6", + "[10] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 6", + "[11] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. Advances In Neural Information Processing Systems, 35:31841-31854, 2022. 2", + "[12] Jiatao Gu, Alex Trevithick, Kai-En Lin, Joshua M Susskind, Christian Theobalt, Lingjie Liu, and Ravi Ramamoorthi. Nerfdiff: Single-image view synthesis with nef-guided distillation from 3d-aware diffusion. In International Conference on Machine Learning, pages 11808-11826. PMLR, 2023. 2" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Zehuan Huang, Hao Wen, Junting Dong, Yaohui Wang, Yangguang Li, Xinyuan Chen, Yan-Pei Cao, Ding Liang, Yu Qiao, Bo Dai, and Lu Sheng. Epidiff: Enhancing multi-view synthesis via localized epipolar-constrained diffusion, 2023. 2", + "[14] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 867-876, 2022. 2", + "[15] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2", + "[16] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439, 2023. 5", + "[17] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics (TOG), 39(6):1-14, 2020. 6", + "[18] Oliver Laric. Three d scans. In threadscans.com, 2012. 6", + "[19] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300–309, 2023. 1, 2", + "[20] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Zexiang Xu, Hao Su, et al. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. arXiv preprint arXiv:2306.16928, 2023.", + "[21] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9298–9309, 2023. 2", + "[22] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 2", + "[23] Xiaoxiao Long, Cheng Lin, Peng Wang, Taku Komura, and Wenping Wang. Sparseneus: Fast generalizable neural surface reconstruction from sparse views. In European Conference on Computer Vision, pages 210-227. Springer, 2022. 2", + "[24] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. arXiv preprint arXiv:2310.15008, 2023. 2", + "[25] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12663–12673, 2023. 2, 6, 7" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4628", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 2, 6, 7", + "[27] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 conference papers, pages 1-8, 2022. 2", + "[28] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2", + "[29] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 1, 2", + "[30] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, et al. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. arXiv preprint arXiv:2306.17843, 2023. 2", + "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2", + "[32] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2", + "[33] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. arXiv preprint arXiv:2302.01721, 2023. 2, 4, 5, 6, 7", + "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 3", + "[35] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 2", + "[36] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2", + "[37] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 2, 5", + "[38] Yawar Siddiqui, Justus Thies, Fangchang Ma, Qi Shan, Matthias Nießner, and Angela Dai. Texturify: Generating" + ], + "bbox": [ + 78, + 90, + 468, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "textures on 3d shape surfaces. In European Conference on Computer Vision, pages 72-88. Springer, 2022. 2", + "[39] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3, 6", + "[40] Jia-Mu Sun, Tong Wu, and Lin Gao. Recent advances in implicit representation-based 3d shape generation. Visual Intelligence, 2(1):9, 2024. 1", + "[41] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2818-2826, 2016. 7", + "[42] Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi. Viewset diffusion:(0-) image-conditioned 3d generative models from 2d data. arXiv preprint arXiv:2306.07881, 2023. 2", + "[43] Shitao Tang, Fuyang Zhang, Jiacheng Chen, Peng Wang, and Yasutaka Furukawa. Mvdiffusion: Enabling holistic multiview image generation with correspondence-aware diffusion. arXiv preprint arXiv:2307.01097, 2023.", + "[44] Hung-Yu Tseng, Qinbo Li, Changil Kim, Suhib Alsisan, Jia-Bin Huang, and Johannes Kopf. Consistent view synthesis with pose-guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16773-16783, 2023. 2", + "[45] Greg Turk and Marc Levoy. Zippered polygon meshes from range images. In Proceedings of the 21st annual conference on Computer graphics and interactive techniques, pages 311-318, 1994. 6", + "[46] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12619-12629, 2023. 2", + "[47] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213, 2023. 2", + "[48] Jianfeng Xiang, Jiaolong Yang, Binbin Huang, and Xin Tong. 3d-aware image generation using 2d diffusion models. arXiv preprint arXiv:2303.17905, 2023. 2", + "[49] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. arXiv preprint arXiv:2306.07954, 2023. 5", + "[50] Jonathan Young. xatlas. In github.com/jpcy/xatlas, 2016. 3", + "[51] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3" + ], + "bbox": [ + 503, + 92, + 890, + 800 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4629", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/GenesisTex_ Adapting Image Denoising Diffusion to Texture Space/ba009eff-00cf-485a-8674-0d1aa83b6a01_model.json b/2024/GenesisTex_ Adapting Image Denoising Diffusion to Texture Space/ba009eff-00cf-485a-8674-0d1aa83b6a01_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9ee2f844307ba409062185f41dfc36223db4b2ec --- /dev/null +++ b/2024/GenesisTex_ Adapting Image Denoising Diffusion to Texture Space/ba009eff-00cf-485a-8674-0d1aa83b6a01_model.json @@ -0,0 +1,2464 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.149, + 0.131, + 0.82, + 0.154 + ], + "angle": 0, + "content": "GenesisTex: Adapting Image Denoising Diffusion to Texture Space" + }, + { + "type": "text", + "bbox": [ + 0.143, + 0.179, + 0.826, + 0.253 + ], + "angle": 0, + "content": "Chenjian Gao\\(^{1,2\\dagger}\\) Boyan Jiang\\(^{2}\\) Xinghui Li\\(^{2}\\) Yingpeng Zhang\\(^{2\\#}\\) Qian Yu\\(^{1*}\\) \n\\(^{1}\\)School of Software, Beihang University \n\\(^{2}\\)R&D Efficiency and Capability Department, Tencent IEG \nhttps://cjeen.github.io/GenesisTexPaper/" + }, + { + "type": "title", + "bbox": [ + 0.272, + 0.287, + 0.35, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.318, + 0.474, + 0.622 + ], + "angle": 0, + "content": "We present GenesisTex, a novel method for synthesizing textures for 3D geometries from text descriptions. GenesisTex adapts the pretrained image diffusion model to texture space by texture space sampling. Specifically, we maintain a latent texture map for each viewpoint, which is updated with predicted noise on the rendering of the corresponding viewpoint. The sampled latent texture maps are then decoded into a final texture map. During the sampling process, we focus on both global and local consistency across multiple viewpoints: global consistency is achieved through the integration of style consistency mechanisms within the noise prediction network, and low-level consistency is achieved by dynamically aligning latent textures. Finally, we apply reference-based inpainting and img2img on denser views for texture refinement. Our approach overcomes the limitations of slow optimization in distillation-based methods and instability in inpainting-based methods. Experiments on meshes from various sources demonstrate that our method surpasses the baseline methods quantitatively and qualitatively." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.648, + 0.21, + 0.663 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.47, + 0.794 + ], + "angle": 0, + "content": "In recent years, with the development of deep learning, 3D content generation technology has made significant progress. The applications of 3D content generation are diverse, ranging from AR/VR to gaming and filmmaking. While there has been considerable research on deep learning-based geometric asset generation[40], there has been a notable industry demand for generating realistic textures for given geometries." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.855 + ], + "angle": 0, + "content": "Recently, text-conditioned image diffusion models [34] have achieved impressive results in image generation. Some works [7, 19, 29] have leveraged text-conditioned image diffusion models to generate textured 3D assets by gener" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.286, + 0.889, + 0.534 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.595, + 0.535, + 0.799, + 0.548 + ], + "angle": 0, + "content": "red volkswagon beetle, cartoon style" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.56, + 0.894, + 0.644 + ], + "angle": 0, + "content": "Figure 1. Texturing results of different methods. Score Distillation Sampling (SDS) based method produces blurred and oversaturated textures. Inpainting-based approach results in artifacts at the boundaries of inpainting masks. Texture space sampling concurrently generating content from multiple viewpoints, produces clean, clear and natural colored textures." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.673, + 0.893, + 0.853 + ], + "angle": 0, + "content": "ating content from multiple viewpoints, achieving notable performance. In this work, our focus is on generating high-quality textures for a given geometry, leveraging the priors provided by pre-trained text-to-image diffusion models. This task poses several challenges, including: 1) View consistency: ensuring cross-view constraints for maintaining low-level consistency; 2) High efficiency: generating textures for a model within a few minutes, enabling practical applications; 3) Zero-shot learning: achieving texture generation without requiring additional training or finetuning. Addressing these challenges is crucial for successful application of image diffusion models to the texture domain." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Currently, the most prominent method for text-to-3D generation is Score Distillation Sampling (SDS) [7, 19, 29]. SDS utilizes the prior knowledge from image diffusion" + }, + { + "type": "page_footnote", + "bbox": [ + 0.1, + 0.863, + 0.385, + 0.876 + ], + "angle": 0, + "content": "\\(^{\\dagger}\\)Work was done during an internship at Tencent IEG." + }, + { + "type": "page_footnote", + "bbox": [ + 0.103, + 0.876, + 0.212, + 0.888 + ], + "angle": 0, + "content": "Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.103, + 0.888, + 0.227, + 0.9 + ], + "angle": 0, + "content": "* Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.863, + 0.385, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4620" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.333 + ], + "angle": 0, + "content": "models to iteratively generate gradients for rendering images of 3D objects. However, SDS has limitations in terms of both efficiency and quality. It often takes several hours to generate a single textured 3D object, and the resulting textures can suffer from over-saturation due to the adoption of a large CFG (classifier-free guidance) scale, which reduces randomness during generation. Recent works [2, 6, 33] have introduced inpainting-based methods for multi-view generation. These methods establish a predetermined order of views, using the content of previously generated views as a condition for subsequent views. While inpainting-based approaches are faster and produce realistic colors, they are sensitive to the predefined view order. Additionally, they require texture segmentation to determine the source view for each texture pixel, which often results in artifacts around the borders of the segmented areas." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.337, + 0.473, + 0.715 + ], + "angle": 0, + "content": "To address the aforementioned issues, we propose GenesisTex, a novel approach that introduces texture space sampling. Specifically, we maintain multiple-view latent texture maps throughout the sampling process and perform denoising to improve their quality. Subsequently, the content is decoded from the latent space to obtain RGB textures. Our approach focuses on achieving two aspects of consistency during the sampling process. Firstly, we ensure global style consistency across multiple views by incorporating style consistency in the noise prediction network. This helps maintain a coherent style throughout the generated textures. Secondly, we employ dynamic alignment of latent textures to ensure low-level multi-view consistency, enhancing the overall quality of the generated textures. Due to memory limitations, texture space sampling operates on sparse views. However, to further enhance the quality, we leverage reference-based inpainting and Img2Img techniques on denser views for texture map refinement. GenesisTex can generate detailed, clean, and naturally colored textures for a given geometry within a few minutes. Unlike inpainting-based methods, our proposed texture space sampling concurrently generates content for multiple views without relying on a predefined order. This adaptability makes GenesisTex suitable for various geometries and results in fewer artifacts in the generated textures." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.717, + 0.423, + 0.731 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.735, + 0.47, + 0.795 + ], + "angle": 0, + "content": "- First, we present a novel method for texture generation, where the core is texture space sampling. This sampling technique allows for the concurrent denoising of latent textures associated with multiple viewpoints." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.796, + 0.47, + 0.841 + ], + "angle": 0, + "content": "- Second, we introduce Style Consistency and Dynamic Alignment in texture space sampling for multi-view consistency." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.841, + 0.47, + 0.9 + ], + "angle": 0, + "content": "- Third, we conduct a comprehensive study involving numerous 3D objects from various sources. The experimental results demonstrate the superiority of our method over baseline methods." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.735, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.09, + 0.642, + 0.107 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.12, + 0.895, + 0.439 + ], + "angle": 0, + "content": "Texture Synthesis. Generating textures over 3D surfaces is a challenging problem, as it requires attention to both colors and geometry. Earlier works like AUV-Net [8] and Texturify [38] embed the geometric prior into a UV map or mesh parameterization. Different from them, EG3D [3] and GET3D [11] directly train 3D StyleGANs [15] to generate geometries and textures jointly, where the textures are implicit texture fields. However, these methods either only work on a single category, or demand textured 3D shapes for training without text conditioning, which limits their broad applicability. Recently, [2, 6, 33] use the priors provided by the image diffusion model to synthesize textures. Text2Tex [6] and TEXTure [33] performs inpainting on multiview renderings. TexFusion [2] propose a sequential texture sampling method. Our method predicts multi-view noise concurrently, avoiding the following issues of sequential noise prediction: 1) Each iteration requires adding forward noise to visited regions to match unvisited areas, potentially leading to detail loss. 2) Consistency is only constrained between adjacent viewpoints, with no direct consistency for long-range viewpoints." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.443, + 0.895, + 0.626 + ], + "angle": 0, + "content": "Text-to-Image Diffusion Models. Over the past years, the development of several large-scale diffusion models [28, 32, 34, 35] has enabled the production of highly detailed and visually impressive images. These models generate images based on input text prompts. Specifically, Stable Diffusion is trained on a substantial text-image dataset [36] and incorporates a text encoder from CLIP [31] to understand the input prompts. Beyond the basic text conditioning, ControlNet [51] enables the model to condition its denoising network on additional input modalities, such as depth maps. In this work, we utilize ControlNet and Stable Diffusion to provide geometrically-conditioned image priors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.896, + 0.901 + ], + "angle": 0, + "content": "Text-to-3D using 2D Image Diffusion Models. Early works [14, 26, 27] utilize the pretrained CLIP model to maximize the similarity between rendered images and text prompt. However, pioneering works DreamFusion [29] and SJC [46], on the other hand, propose to distill a 2D text-to-image generation model to generate 3D shapes from texts, and many follow-up works [19-21, 23, 25, 30, 47] follow such per-shape optimization scheme. Recently, several methods [4, 12, 13, 22, 24, 37, 42-44, 48] have been proposed to generate consistent multi-view images by using diffusion models. MVDream [37] and SyncDreamer [22] share similar ideas, generating consistent multi-view images via attention layers. However, existing approaches either endure slow optimization processes or depend on separately trained 3D priors, rendering them unsuitable for direct application in texture synthesis. In contrast, our method does not require additional training and can generate results within several minutes." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.958 + ], + "angle": 0, + "content": "4621" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.09, + 0.212, + 0.108 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.115, + 0.473, + 0.268 + ], + "angle": 0, + "content": "Our objective is to generate a texture map \\(\\mathcal{T}\\) for a given 3D mesh \\(\\mathcal{M}\\), using the provided text as a descriptive condition. Our approach leverages the Stable Diffusion model as the image diffusion model. We begin by introducing some basic concept about Stable Diffusion and texture representation in Sec. 3.1. Then, we introduce a sampling algorithm in texture space for Stable Diffusion in Sec. 3.2. Furthermore, we propose multi-view consistency strategy in Sec. 3.3. Finally we perform refinement on the texture map, which includes Inpainting and Img2Img, as described in Sec. 3.4." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.275, + 0.208, + 0.291 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.298, + 0.471, + 0.36 + ], + "angle": 0, + "content": "Stable Diffusion. The diffusion model belongs to a class of generative models that generate data through iterative denoising from random noise. In the sampling process, \\(\\mathbf{z}_{i-1}\\) can be sampled from \\(\\mathbf{z}_i\\) using a DDIM sampler [39]:" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.368, + 0.469, + 0.446 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\hat {\\mathbf {z}} _ {i \\rightarrow 0} = (\\mathbf {z} _ {i} - \\sqrt {1 - \\alpha_ {i}} \\epsilon_ {\\theta} (\\mathbf {z} _ {i}, t _ {i}, \\mathbf {c} _ {t e x t})) / \\sqrt {\\alpha_ {i}}, \\\\ \\mathbf {z} _ {i - 1} = \\sqrt {\\alpha_ {i - 1}} \\hat {\\mathbf {z}} _ {i \\rightarrow 0} + \\sqrt {1 - \\alpha_ {i - 1} - \\sigma_ {i} ^ {2}} \\epsilon_ {\\theta} (\\mathbf {z} _ {i}, t _ {i}, \\mathbf {c} _ {t e x t}) \\\\ + \\sigma_ {i} \\varepsilon , \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.447, + 0.47, + 0.551 + ], + "angle": 0, + "content": "where \\(\\epsilon_{\\theta}\\) is a pretrained U-Net for noise prediction and the initial \\(\\mathbf{z}_T \\sim \\mathcal{N}(0, \\mathbf{I})\\). In this paper, we specifically employ Stable Diffusion [34]. Stable Diffusion performs denoising in the latent space and employs an autoencoder \\(\\mathcal{D}(\\mathcal{E}(\\cdot))\\) for the conversion between image and latent representations. So \\(\\mathbf{z}_0\\) generated by the diffusion process is decoded to the image \\(\\mathcal{D}(\\mathbf{z}_0)\\) finally." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.552, + 0.47, + 0.627 + ], + "angle": 0, + "content": "ControlNet. ControlNet [51] injects low-level control during the denoising process of Stable Diffusion. Our approach employs depth \\(\\mathbf{d}\\) as the geometric control condition. The predicted noise of the U-Net with ControlNet is represented as \\(\\epsilon_{\\theta}(\\mathbf{z}_i, t_i, \\mathbf{c}_{text}, \\mathbf{d})\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.628, + 0.47, + 0.687 + ], + "angle": 0, + "content": "Texture Representation. In this paper, we utilize mesh as the 3D representation. Texture map is associated with an UV parameterization of the mesh \\(\\mathcal{M}\\). We use \\(x\\text{atlas}\\) [50] to generate the UV parameterization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.688, + 0.47, + 0.794 + ], + "angle": 0, + "content": "**Rendering.** Given a mesh \\(\\mathcal{M}\\), a texture map \\(\\mathcal{T}\\) and a viewpoint \\(C\\), we can use the rendering function \\(\\mathcal{R}\\) to get the rendered image \\(\\mathbf{x}^{(img)} = \\mathcal{R}(\\mathcal{T};\\mathcal{M},C)\\). Similar to [2], we do not consider any lighting and rendering involves solely sampling colors from the texture map. The inverse rendering function \\(\\mathcal{R}^{-1}\\) can inverse render the image \\(\\mathbf{x}^{(img)}\\) back to a texture map \\(\\mathcal{T}' = \\mathcal{R}^{-1}(\\mathbf{x}^{(img)};\\mathcal{M},C)\\)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.802, + 0.32, + 0.818 + ], + "angle": 0, + "content": "3.2. Sampling in Texture Space" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.902 + ], + "angle": 0, + "content": "The most straightforward method for sampling on the texture map is to fine-tune the Stable Diffusion model directly on some texture maps, making its distribution more like the distribution of texture maps. However, the distribution of texture maps is much more complex than that of rendering" + }, + { + "type": "code_caption", + "bbox": [ + 0.502, + 0.091, + 0.755, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Texture Space Sampling" + }, + { + "type": "algorithm", + "bbox": [ + 0.501, + 0.108, + 0.925, + 0.551 + ], + "angle": 0, + "content": "Input: mesh \\(\\mathcal{M}\\) ,cameras \\(\\{C_1,\\dots ,C_N\\}\\) \nParameters:Denoising time schedule \\(\\{t_i\\}_{i = T}^0\\) ,DDIM \nnoise schedule \\(\\{\\sigma_i\\}_{i = T}^0\\) \n\\(\\{\\mathbf{z}_{T,n}^{(tex)}\\}_{n = 1}^N\\gets \\{\\mathbf{0}\\}\\) \n\\(\\{\\mathbf{z}_n^{(bg)}\\}_{n = 1}^N\\sim \\{\\mathcal{N}(\\mathbf{0},\\mathbf{I})\\}\\) \n\\(\\{\\hat{\\epsilon}_{T + 1,n}\\}_{n = 1}^N\\sim \\{\\mathcal{N}(\\mathbf{0},\\mathbf{I})\\}\\) \ndenoising stage: \nfor \\(i\\in \\{T\\dots 1\\}\\) do \nfor \\(n\\in \\{1\\dots N\\}\\) do \n\\(\\varepsilon_{n}\\sim \\mathcal{N}(0,I)\\) \n\\({\\bf z}_{i,n}^{(img)}\\gets {\\bf M}_n(\\sqrt{\\alpha_i}{\\cal R}({\\bf z}_{i,n}^{(tex)}) + \\sqrt{1 - \\alpha_i - \\sigma_{i + 1}^2}\\hat{\\epsilon}_{i + 1,n})\\) \\(+(1 - \\mathbf{M}_n)\\mathbf{z}_n^{(bg)} + \\sigma_{i + 1}\\varepsilon_n\\) \n\\(\\hat{\\epsilon}_{i,n}\\gets \\epsilon_{\\theta}(\\mathbf{z}_{i,n}^{(img)},t_i,\\mathbf{c},\\mathbf{d}_n)\\) \n\\(\\hat{\\mathbf{z}}_{0,n}^{(img)}\\gets (\\mathbf{z}_{i,n}^{(img)} - \\sqrt{1 - \\alpha_i}\\hat{\\epsilon}_{i,n}) / \\sqrt{\\alpha_i}\\) \n\\({\\bf z}_{i - 1,n}^{(tex)}\\gets \\mathcal{R}^{-1}(\\hat{\\mathbf{z}}_{0,n}^{(img)})\\) \n\\({\\bf z}_{n}^{(bg)}\\gets \\sqrt{\\alpha_{i - 1}}\\hat{\\mathbf{z}}_{0,n}^{(img)} + \\sqrt{1 - \\alpha_{i - 1} - \\sigma_i^2}\\cdot \\hat{\\epsilon}_{i,n}\\) \nend for \n\\(\\{{\\bf z}_{i - 1,n}^{(tex)}\\}_{n = 1}^N = dynamic\\_ align(\\{{\\bf z}_{i - 1,n}^{(tex)}\\}_{n = 1}^N)\\) \nend for \ndecoding stage: \nfor \\(n\\in \\{1\\dots N\\}\\) do \n\\({\\bf x}_n^{(img)} = \\mathcal{D}(\\hat{\\bf z}_{0,n}^{(img)})\\) \n\\({\\bf x}_n^{(tex)} = R^{-1}({\\bf x}_n^{(img)})\\) \nend for \nT=merge({x(tex)})N \nreturn Texture map Tsampled" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.579, + 0.892, + 0.639 + ], + "angle": 0, + "content": "images since different UV parameterizations of the same 3D object correspond to different texture maps. Instead of fine-tuning existing image diffusion models, we adapt them to the texture space for texture map generation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.64, + 0.893, + 0.793 + ], + "angle": 0, + "content": "Algorithm 1 presents our proposed texture space sampling algorithm. Firstly we define a set of \\(\\mathcal{C}^{(sampling)} = \\{C_1,..,C_N\\}\\) camera views and render the corresponding depth map \\(\\mathbf{d}_n\\) in each view. Since we employ a latent diffusion model, our texture space sampling method first performs denoising in the latent texture space and then decode the latent to texture image. In the denoising stage, a dedicated diffusion denoising algorithm is applied to a set of latent texture maps \\(\\{\\mathbf{z}_n^{(tex)}\\}_{n=1}^N\\). In the decoding stage, we recover the texture map \\(\\mathcal{T}_{sampled}\\) from the latent space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.793, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Denoising Stage. In the denoising stage, we adapt the DDIM from the rendering space to the texture space and utilize the same parameters as DDIM, including denoising time schedule \\(\\{t_i\\}_{i = T}^0\\) , DDIM noise schedule \\(\\{\\sigma_i\\}_{i = T}^0\\) . For each viewpoint, we maintain a separate latent texture which dynamically interact during the denoising process. We begin with \\(\\{\\mathbf{z}_{T,n}^{(tex)}\\}_{n = 1}^{N} = \\{\\mathbf{0}\\}\\) , the zero latent texture map" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4622" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.391, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.393, + 0.09, + 0.605, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.09, + 0.892, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.203, + 0.891, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.298, + 0.894, + 0.342 + ], + "angle": 0, + "content": "Figure 2. Overview of GenesisTex. GenesisTex generates a texture map for a given mesh \\(\\mathcal{M}\\), based on a prompt. Texture Space Sampling samples a texture map using Stable Diffusion, introducing style consistency and dynamic alignment across multiple viewpoints. Furthermore, Inpainting and Img2Img are applied to fill in the blank regions and enhance the quality of texture map details, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.366, + 0.47, + 0.454 + ], + "angle": 0, + "content": "corresponding to \\(N\\) viewpoints. At each denoising step, our goal is to predict \\(\\mathbf{z}_{i - 1,n}^{(tex)}\\) from \\(\\mathbf{z}_{i,n}^{(tex)}\\). We first render the latent texture map \\(\\mathbf{z}_{i,n}^{(tex)}\\) to the rendering space using the rendering function \\(\\mathcal{R}\\), and then calculate the corresponding latent image \\(\\mathbf{z}_{i,n}^{(img)}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.084, + 0.466, + 0.469, + 0.525 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {z} _ {i, n} ^ {(i m g)} = \\mathbf {M} _ {n} (\\sqrt {\\alpha_ {i}} \\mathcal {R} (\\mathbf {z} _ {i, n} ^ {(t e x)}) + \\sqrt {1 - \\alpha_ {i} - \\sigma_ {i + 1} ^ {2}} \\hat {\\epsilon} _ {i + 1, n}) \\\\ + \\left(1 - \\mathbf {M} _ {n}\\right) \\mathbf {z} _ {n} ^ {(b g)} + \\sigma_ {i + 1} \\varepsilon_ {n} \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.525, + 0.47, + 0.716 + ], + "angle": 0, + "content": "Here, \\(\\mathbf{M}_n\\) represents the foreground mask for viewpoint \\(C_n\\) (downsampled to the same resolution as \\(\\mathbf{z}^{(img)})\\), where foreground pixels have a value of 1, and background pixels have a value of 0. \\(\\hat{\\epsilon}_{i + 1,n}\\) represents the noise predicted by the noise prediction network at step \\(i + 1\\rightarrow i\\) (notably, we define \\(\\hat{\\epsilon}_{T + 1,n}\\sim \\mathcal{N}(0,\\mathbf{I})\\)). \\(\\mathbf{z}_n^{(bg)}\\) is the background of the latent image, which is set to Gaussian noise at initialization. \\(\\varepsilon_{n}\\sim \\mathcal{N}(0,\\mathbf{I})\\) is a random Gaussian noise. Subsequently, we predict the corresponding noise \\(\\hat{\\epsilon}_{i,n} = \\epsilon_{\\theta}(\\mathbf{z}_{i,n}^{(img)},t_i,\\mathbf{c},\\mathbf{d}_n)\\) for the latent image \\(\\mathbf{z}_{i,n}^{(img)}\\) conditioned on the depth map \\(\\mathbf{d}_n\\). With the predicted noise, we can obtain the predicted \\(\\hat{\\mathbf{z}}_{0,n}^{(img)}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.727, + 0.469, + 0.749 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {z}} _ {0, n} ^ {(i m g)} = (\\mathbf {z} _ {i, n} ^ {(i m g)} - \\sqrt {1 - \\alpha_ {i}} \\hat {\\epsilon} _ {i, n}) / \\sqrt {\\alpha_ {i}}, \\qquad (3)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.762, + 0.469, + 0.8 + ], + "angle": 0, + "content": "Finally, we inverse render \\(\\hat{\\mathbf{z}}_{0,n}^{(img)}\\) to the texture space using \\(\\mathcal{R}^{-1}\\) and update \\(\\mathbf{z}_n^{(bg)}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.81, + 0.469, + 0.841 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {i - 1, n} ^ {(t e x)} = \\mathcal {R} ^ {- 1} \\left(\\hat {\\mathbf {z}} _ {0, n} ^ {(i m g)}\\right), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.832, + 0.422, + 0.86 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {n} ^ {(b g)} = \\sqrt {\\alpha_ {i - 1}} \\hat {\\mathbf {z}} _ {0, n} ^ {(i m g)} + \\sqrt {1 - \\alpha_ {i - 1} - \\sigma_ {i} ^ {2}} \\cdot \\hat {\\epsilon} _ {i, n}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "In the denoising stage, we repeat the above process \\( T - 1 \\) times. After each denoising step, we impose consistency" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.365, + 0.892, + 0.457 + ], + "angle": 0, + "content": "constraints on latent texture \\(\\mathbf{z}_{i - 1,n}^{(tex)}\\), which will be introduced in Sec. 3.3. Notably, differing from DDIM where the latent includes noise, the latent texture in our method is an estimate of the noise-free version. This is because \\(\\mathcal{R}\\) and \\(\\mathcal{R}^{-1}\\) in the denoising process require interpolation of the latent, which may distort the noise." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.457, + 0.892, + 0.593 + ], + "angle": 0, + "content": "Decoding Stage. To recover the texture map \\(\\mathcal{T}_{\\text{sampled}}\\), we first decode each viewpoint's latent separately to obtain multi-view rendering. Subsequently, we inverse render the renderings into the texture space. After obtaining multiview texture maps, we merge them into a single texture map as the output. Similar to [2, 6, 33], we aim for each pixel on each texture map to come from the viewpoint where the corresponding point on the mesh is observed most directly. Therefore, our merge is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.543, + 0.605, + 0.892, + 0.647 + ], + "angle": 0, + "content": "\\[\n\\mathcal {T} _ {\\text {s a m p l e d}} = \\sum_ {n = 1} ^ {N} \\operatorname {S o f t m a x} \\left(\\mathcal {R} ^ {- 1} \\left(\\mathbf {N} _ {\\mathbf {n}}\\right)\\right) \\times \\mathbf {x} _ {n} ^ {(t e x)} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.892, + 0.735 + ], + "angle": 0, + "content": "\\(\\mathbf{N}_{\\mathbf{n}}\\) is the similarity mask at viewpoint \\(C_n\\), where each pixel represents the cosine similarity between the normal vectors of the visible faces and the reversed view direction. We use Softmax instead of Max because Softmax can result in more natural transition in the texture maps." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.743, + 0.862, + 0.76 + ], + "angle": 0, + "content": "3.3. Consistency between Latent Texture Maps" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.767, + 0.694, + 0.782 + ], + "angle": 0, + "content": "3.3.1 Dynamic Alignment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.791, + 0.892, + 0.852 + ], + "angle": 0, + "content": "During sampling in texture space, we maintain \\(N\\) latent texture maps, now we introduce an alignment approach to ensure their local consistency. Firstly, we directly reduce these latent texture maps to obtain a uniform latent texture map:" + }, + { + "type": "equation", + "bbox": [ + 0.541, + 0.863, + 0.892, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {i - 1} ^ {(u n i t e x)} = \\sum_ {n = 1} ^ {N} \\operatorname {S o f t m a x} \\left(\\mathcal {R} ^ {- 1} \\left(\\mathbf {N} _ {n}\\right)\\right) \\times \\mathbf {z} _ {i - 1, n} ^ {(t e x)} \\tag {6}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4623" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.089, + 0.473, + 0.265 + ], + "angle": 0, + "content": "Next, we blend each viewpoint's latent texture map \\(\\mathbf{z}_{i-1,n}^{(tex)}\\) with the uniform latent texture map \\(\\mathbf{z}_{i-1}^{(unitex)}\\). We do not enforce strict alignment among the \\(N\\) latent texture maps at every step. Since a latent pixel gains full meaning only when combined with the surrounding context, which is viewpoint-dependent, the latent texture map from each viewpoint should contain some independent information. Moreover, the content generated at different timestamps varies, so the alignment constraints should adapt flexibly. We introduce a signal \\(c(t)\\) for dynamically regulating the variations in consistency during the sampling process:" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.274, + 0.47, + 0.297 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {i - 1, n} ^ {(c o n s i s t e x)} = c (t _ {i}) \\times \\mathbf {z} _ {i - 1} ^ {(u n i t e x)} + (1 - c (t _ {i})) \\times \\mathbf {z} _ {i - 1, n} ^ {(t e x)} (7)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.306, + 0.47, + 0.368 + ], + "angle": 0, + "content": "We empirically find that maintaining strong local consistency in the mid-term of denoising, while keeping weak local consistency in the early and late stages, leads to higher quality generation results." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.386, + 0.254, + 0.402 + ], + "angle": 0, + "content": "3.3.2 Style Consistency" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.411, + 0.47, + 0.621 + ], + "angle": 0, + "content": "In texture space sampling, we use the noise estimation network \\(\\epsilon_{\\theta}\\) to estimate noise for the multi-view latents, which is responsible for content generation. So we adapt \\(\\epsilon_{\\theta}\\) to ensure style consistency of estimated noise across different views. Inspired by video diffusion models [16, 49] and multi-view diffusion model [37], we modify all the self-attention layers and group normalization layers in the noise prediction network to align the style of multi-view content. Adapted Self-Attention. In Stable Diffusion, self-attention facilitates long-range interactions among features within an image. We transform self-attention into cross-view attention to establish style consistency across multiple viewpoints, by using the key \\(\\mathbf{K}^{\\prime}\\) and value \\(\\mathbf{V}^{\\prime}\\) from all viewpoints," + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.622, + 0.47, + 0.648 + ], + "angle": 0, + "content": "\\[\n\\mathbf {K} ^ {\\prime} = \\mathbf {W} ^ {K} \\left[ \\mathbf {z} _ {i, 1} ^ {(i m g)}; \\dots ; \\mathbf {z} _ {i, n} ^ {(i m g)} \\right], \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.647, + 0.371, + 0.665 + ], + "angle": 0, + "content": "\\[\n\\mathbf {V} ^ {\\prime} = \\mathbf {W} ^ {V} [ \\mathbf {z} _ {i, 1} ^ {(i m g)}; \\dots ; \\mathbf {z} _ {i, n} ^ {(i m g)} ].\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.672, + 0.364, + 0.689 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}^K\\), \\(\\mathbf{W}^V\\) are pretrained parameters." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.469, + 0.81 + ], + "angle": 0, + "content": "Adapted Group Normalization. Stable diffusion adopt group normalization as the normalization layer. Group normalization divides the channels into groups and computes within each group the mean and variance for normalization. We convert 2D group normalization to 3D by connecting all different views in the group normalization layer, i.e., the mean and variance are calculated within the channel group of all pixels across all viewpoints." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.811, + 0.47, + 0.902 + ], + "angle": 0, + "content": "By adjusting self-attention layer and group normalization layer, we achieve multi-view style consistency without any additional training. In addition to the noise prediction network \\(\\epsilon_{\\theta}\\), we apply the same modifications to the decoder \\(\\mathcal{D}\\) to further ensure style consistency in the multi-view decoded RGB images." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.729, + 0.107 + ], + "angle": 0, + "content": "3.4. Texture Map Refinement" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.115, + 0.747, + 0.132 + ], + "angle": 0, + "content": "3.4.1 Post-processing: Inpainting" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.14, + 0.894, + 0.246 + ], + "angle": 0, + "content": "Since the memory cost of texture space sampling increases with the number of viewpoints of \\( \\mathcal{C}_{\\text{sampling}} \\), the number of viewpoints \\( N \\) is limited. Some regions in the texture map \\( \\mathcal{T}_{\\text{sampled}} \\) may not be observed in any of the viewpoints in \\( \\mathcal{C}_{\\text{sampling}} \\). Therefore, after texture space sampling, we introduce an inpainting epoch to fill texture in areas not observed in \\( \\mathcal{C} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.247, + 0.893, + 0.574 + ], + "angle": 0, + "content": "Firstly we define a denser set of viewpoints \\(\\mathcal{C}^{(inpainting)} = \\{C_1,..,C_{N_1}\\}\\) as compared to \\(\\mathcal{C}^{(samping)}\\) where \\(N_{1} > N\\). We use a mask \\(\\mathbf{M}_C^{blank}\\) to indicate the rendered areas with blank textures at the viewpoint \\(C\\). The entire inpainting epoch comprises \\(N_{1}\\) iterations. In each iteration, we first compute the viewpoint \\(C_n = \\arg \\max_C\\mathbf{M}_C^{blank}\\) with the largest blank area. Then, we perform inpainting on the rendering \\(\\mathbf{x}_n^{(img)}\\) using the pretrained depth conditioned Stable Diffusion model as described in [6, 33]. During inpainting, Stable Diffusion utilizes regions with textures as conditions to fill in blank areas without textures. To enhance the robustness of inpainting, we render an image using a viewpoint \\(C_0\\) from the set \\(\\mathcal{C}^{(sampling)}\\). We then concatenate this image to the image requiring inpainting as a additional condition for inpainting. Compared to naive inpainting, reference-based inpainting provides a more comprehensive condition. To ensure a natural and smooth transition between blank areas and textured regions, we apply Gaussian blur to the mask \\(\\mathbf{M}_{C_n}^{blank}\\). Finally the inpainted rendering \\(\\mathbf{x}_n^{(img)'}\\) is used to update the texture map:" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.587, + 0.891, + 0.627 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {T} _ {n} = \\mathcal {T} _ {n - 1} \\times (1 - \\mathcal {R} ^ {- 1} (\\mathbf {M} ^ {(b l a n k)}) \\tag {9} \\\\ + \\mathcal {R} ^ {- 1} \\left(\\mathbf {x} _ {n} ^ {(i m g) \\prime}\\right) \\times \\mathcal {R} ^ {- 1} \\left(\\mathbf {M} ^ {(b l a n k)}\\right) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.639, + 0.892, + 0.67 + ], + "angle": 0, + "content": "We start with \\(\\mathcal{T}_0 = \\mathcal{T}_{\\text{sampled}}\\) and iterate for \\(N_1\\) times. Finally we get \\(\\mathcal{T}_{\\text{inpainted}} = \\mathcal{T}_{N_1}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.671, + 0.892, + 0.747 + ], + "angle": 0, + "content": "It's worth noting that, unlike [2, 6, 33], which relies heavily on inpainting to maintain multi-view consistency, in our method inpainting is performed only to patch small areas. Therefore, the inpainting's performance does not significantly affect the final quality of the generated texture." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.769, + 0.738, + 0.785 + ], + "angle": 0, + "content": "3.4.2 Post-processing:Img2Img" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.855 + ], + "angle": 0, + "content": "So far our generated texture map \\(\\mathcal{T}_{\\text {inpainted }}\\) may still contain unnatural transitions in some regions due to multi-view conflicts and inpainting artifacts. To further improve the generation quality, we introduce an Img2Img epoch." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In Img2Img epoch, we also define a set of viewpoints \\( C^{(img2img)} \\) include \\( N_{2} \\) viewpoints and iterate through viewpoint from \\( C^{(img2img)} \\). In each iteration, we encode" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4624" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.09, + 0.473, + 0.262 + ], + "angle": 0, + "content": "the rendering image \\(\\mathbf{x}_n^{(img)}\\) to the latent space using the encoder \\(\\mathcal{E}\\) and add some noise to the latent rendering. Then we perform denoising using the pretrained depth conditioned Stable Diffusion model. And finally we decode the denoised latent back to image space using \\(\\mathcal{D}\\). However, \\(\\mathcal{D}(\\mathcal{E}(\\cdot))\\) is not an identity transformation because of the reconstruction distortion of the autoencoder. So we apply the color histogram matching to the foreground of the decoded image to alleviate the color distortion. Finally we update the texture map using the img2img result \\(\\mathbf{x}_n^{(img)'}\\) based on view similarity \\(\\mathbf{N}_n\\):" + }, + { + "type": "equation", + "bbox": [ + 0.161, + 0.272, + 0.468, + 0.311 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {T} _ {n} = \\mathcal {T} _ {n - 1} \\times (1 - \\mathcal {R} ^ {- 1} (\\mathbf {N} _ {n})) \\\\ + \\mathcal {R} ^ {- 1} \\left(\\mathbf {x} _ {n} ^ {(i m g) \\prime}\\right) \\times \\mathcal {R} ^ {- 1} \\left(\\mathbf {N} _ {n}\\right) \\tag {10} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.322, + 0.47, + 0.356 + ], + "angle": 0, + "content": "We start with \\(\\mathcal{T}_0 = \\mathcal{T}_{\\text{inpainted}}\\). After \\(N_2\\) iterations we get \\(\\mathcal{T}_{\\text{refined}} = \\mathcal{T}_{N_2}\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.368, + 0.21, + 0.385 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.393, + 0.16, + 0.41 + ], + "angle": 0, + "content": "4.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.416, + 0.47, + 0.599 + ], + "angle": 0, + "content": "Implementation Details. Our experiments are conducted on an NVIDIA A10 GPU. We utilize the official Stable Diffusion-v1.5 model with the ControlNet-v1.1 (depth). For Inpainting epoch and Img2Img epoch, we use DDIM [39] as the sampler. For all samplers, we set the number of iterations to 20 steps and the CFG scale (classifier-free guidance scale) to 7.5. We implement the rendering function using nvdiffrast [17] and make modifications to nvdiffrast to support inverse rendering. Texture space sampling takes about 3 minutes to generate textures for a single object. For more detailed parameter settings, please refer to the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.47, + 0.704 + ], + "angle": 0, + "content": "Dataset. Similar to TexFusion [2], we utilize 35 meshes to test the texture generation performance, including 11 objects from Objaverse [10], 17 from ShapeNet [5], 1 from Text2Mesh [26], 2 from Turbosquid [1], 2 from Stanford 3D Scans [9, 45] and 2 from Three D Scans [18]. Each object has 1-4 text descriptions, resulting in a total of 80 (mesh, prompt) pairs in this collection." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Baselines. We conduct comparisons with state-of-the-art methods for generating textures from text: 1) Text2Mesh [26], a method that stylizes a 3D mesh by predicting color and local geometric details which conform to a target text prompt, harnessing the representational power of CLIP. 2) Latent-Paint [25], an approach leveraging SDS to obtain texture gradients from the Stable Diffusion model. 3) Text2Tex [6], a method that uses a depth-aware image inpainting diffusion model to incrementally produce partial textures from various viewpoints. 4) TEXTure [33], similar to Text2Tex, but utilizes a different region segmentation strategy. 5) TexFusion [2], a method sequentially inpainting latent texture. As TexFusion does not have publicly" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.089, + 0.9, + 0.214 + ], + "angle": 0, + "content": "
MethodFID (↓)KID (↓) (×10-3)User study (%)
Visual Quality (↑)Align with Prompt (↑)
Latent-Paint110.1410.645.154.28
Text2Mesh121.6115.133.154.86
Text2Tex101.388.3518.2819.72
TEXTure100.479.2223.4224.86
Ours74.582.8950.0046.28
" + }, + { + "type": "table_caption", + "bbox": [ + 0.523, + 0.224, + 0.868, + 0.239 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons with baseline methods." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.254, + 0.899, + 0.362 + ], + "angle": 0, + "content": "
texture space samplingInpainting RoundImg2Img RoundFID(↓)KID(↓) (×10-3)
✓✗✗86.253.79
✓✗✓82.053.67
✓✓✗76.303.29
✓✓✓74.582.89
" + }, + { + "type": "table_caption", + "bbox": [ + 0.563, + 0.373, + 0.828, + 0.386 + ], + "angle": 0, + "content": "Table 2. Effectiveness of texture refinement." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.406, + 0.893, + 0.438 + ], + "angle": 0, + "content": "available source code, we extract some rendered images of generated results from their paper for comparison." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.448, + 0.693, + 0.465 + ], + "angle": 0, + "content": "4.2. Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.472, + 0.893, + 0.683 + ], + "angle": 0, + "content": "Fig. 3 and Fig. 5 presents visual comparisons between our method and baseline approaches. Textures produced by Text2Mesh are notably lacking in detail. Latent-Paint yields textures with greater intricacy compared to Text2Mesh, but they still exhibit considerable blurriness due to SDS limitations. Inpainting-based approaches like Text2Tex and TEXTure introduce artifacts with unnatural transitions at inpainting mask boundaries, struggling to maintain consistency across multiple views because they depend heavily on inpainting. The output of TexFusion is compromised by blurry details, likely a consequence of repeated noise addition during sequential sampling. In contrast, our method stands out by producing textures that are not only clear but also remarkably coherent." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.684, + 0.892, + 0.73 + ], + "angle": 0, + "content": "Fig. 4 presents more results of our method. Benefiting from the powerful prior of Stable Diffusion, our method can generate textures of diverse styles for different geometries." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.741, + 0.74, + 0.758 + ], + "angle": 0, + "content": "4.3. Quantitative Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "FID & KID. Similar to TexFusion, we sample from pretrained image diffusion model to create ground truth labels. For each mesh, we generate depth maps from the five most common canonical viewpoints: front, back, top, and both sides. These depth maps, along with textual descriptions, serve as inputs to condition the Stable Diffusion model. To ensure the focus remains on the texture of the objects, we modify the background of the ground truth images to be white. We visualize the textures created by various methods" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4625" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.095, + 0.135, + 0.452 + ], + "angle": 270, + "content": "red apple blue handbag with dilapidated Halloween silver trims tap pumpkin head" + }, + { + "type": "image", + "bbox": [ + 0.139, + 0.091, + 0.25, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.156, + 0.18, + 0.262, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.138, + 0.278, + 0.24, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.145, + 0.382, + 0.247, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.158, + 0.478, + 0.226, + 0.493 + ], + "angle": 0, + "content": "Geometry" + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.091, + 0.365, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.18, + 0.38, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.255, + 0.278, + 0.357, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.262, + 0.382, + 0.364, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.279, + 0.478, + 0.353, + 0.492 + ], + "angle": 0, + "content": "Text2Mesh" + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.092, + 0.495, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.404, + 0.18, + 0.51, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.278, + 0.485, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.391, + 0.382, + 0.493, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.404, + 0.478, + 0.486, + 0.492 + ], + "angle": 0, + "content": "Latent-Paint" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.092, + 0.631, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.18, + 0.645, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.278, + 0.622, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.382, + 0.63, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.549, + 0.478, + 0.611, + 0.491 + ], + "angle": 0, + "content": "Text2Tex" + }, + { + "type": "image", + "bbox": [ + 0.654, + 0.092, + 0.759, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.18, + 0.774, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.65, + 0.278, + 0.75, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.657, + 0.382, + 0.758, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.674, + 0.478, + 0.737, + 0.491 + ], + "angle": 0, + "content": "TEXTure" + }, + { + "type": "image", + "bbox": [ + 0.77, + 0.092, + 0.875, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.787, + 0.18, + 0.891, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.767, + 0.278, + 0.866, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.774, + 0.382, + 0.875, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.805, + 0.478, + 0.843, + 0.491 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.503, + 0.893, + 0.531 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparisons with Text2Mesh [26], Latent-Paint [25], Text2Tex [6] and TEXTure [33]. In comparison with the baselines, our GenesisTex exhibits richer details and fewer artifacts." + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.555, + 0.468, + 0.753 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.14, + 0.766, + 0.406, + 0.78 + ], + "angle": 0, + "content": "Figure 4. Texturing results with GenesisTex." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.903 + ], + "angle": 0, + "content": "from the same five viewpoints and evaluate their quality by calculating the Fréchet Inception Distance (FID) and Kernel Inception Distance (KID). FID and KID measure the feature dissimilarity between two image collections, with feature extraction performed using the Inception V3 [41]. The results in Tab.1 demonstrate that the textures gener" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.555, + 0.882, + 0.727 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.739, + 0.892, + 0.768 + ], + "angle": 0, + "content": "Figure 5. Qualitative Comparisons with TexFusion [2]. Images are extracted from their original paper." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.841 + ], + "angle": 0, + "content": "ated by GenesisTex are closer to the ground truth compared to those produced by baselines, indicating a superior synthesis quality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.893, + 0.901 + ], + "angle": 0, + "content": "User Study. We conduct a user study to compare the quality of textures generated by our method versus those produced by baseline methods. We render all textured mesh results into videos to facilitate the presentation to users. Each" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4626" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.351 + ], + "angle": 0, + "content": "video contains a rotating textured mesh along with the corresponding prompt. For each questionnaire, we randomly show the users 10 groups of rendered videos. Each group displays results for a specific (mesh, prompt) pair, comparing outputs from baseline methods and our approach. Each volunteer is required to answer two questions for a single group, including which one has the highest visual quality and which one aligns best with the prompt. Additional information about the questionnaire is available in the supplementary. Finally, we received 35 valid responses from the questionnaires, which included the results of comparisons from 350 groups. The results of the user study are shown in Tab. 1. In terms of both visual quality and alignment with the text prompt, our method is preferred by the most participants, with percentages reaching \\(50.00\\%\\) and \\(46.28\\%\\), respectively. Furthermore, we conduct a pairwise comparison study and the results are included in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.361, + 0.245, + 0.377 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.385, + 0.474, + 0.598 + ], + "angle": 0, + "content": "Consistency in Texture Space Sampling. Our GenesisTex employs style consistency and dynamic alignment for maintaining multi-view consistency during texture space sampling. To investigate the impact of these two consistency strategies, we visualize the decoded multi-view images. We set \\(\\mathcal{C}(elevation, azimuth) = \\{(0^{\\circ}, 0^{\\circ}), (0^{\\circ}, 15^{\\circ}), (0^{\\circ}, 35^{\\circ}), (0^{\\circ}, 45^{\\circ})\\}\\) to ensure that multiple viewpoints share sufficient content. Fig. 6 illustrates an example with the prompt penguin toy. We can observe that without any consistency constraints, the color of the toy varies significant across different viewpoints. After introducing style consistency, consistency for the toy color improves, but there are still inconsistencies in the hair and ears. Further we use dynamic alignment and we can see the" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.622, + 0.465, + 0.844 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.855, + 0.47, + 0.898 + ], + "angle": 0, + "content": "Figure 6. Ablation results on consistency in texture space sampling. Style Consistency + Dynamic Alignment (bottom row) achieves the best multi-view consistency." + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.093, + 0.891, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.309, + 0.893, + 0.339 + ], + "angle": 0, + "content": "Figure 7. Ablation results on texture refinement. Inpainting fills the blank regions, while Img2Img enhances the quality of details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.369, + 0.893, + 0.414 + ], + "angle": 0, + "content": "consistency of details has significantly improved. We include more results in the supplementary to further demonstrate the effectiveness of these two consistencies." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.417, + 0.895, + 0.553 + ], + "angle": 0, + "content": "Effectiveness of Texture Refinement. After texture space sampling, we conduct refinement on the texture, including Inpainting and Img2Img. We validate the effectiveness of these two modules and the results are presented in Fig. 7 and Tab. 2. Inpainting fills in areas lacking texture obtained during texture space sampling, contributing significantly to the improvement in visual quality. Img2Img further repairs artifacts and enhances details, leading to a certain improvement in visual quality as well." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.575, + 0.757, + 0.59 + ], + "angle": 0, + "content": "5. Conclusion and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.603, + 0.892, + 0.829 + ], + "angle": 0, + "content": "In this work, we propose a novel method for text-based texture generation, named GenesisTex. GenesisTex leverages the prior of the pre-trained Stable Diffusion model by introducing texture space sampling. Texture space sampling concurrently generates multi-view content without relying on a predefined sequence of views. Our approach can generate high-quality textures for a given 3D model in several minutes. The primary limitation of our method is the significant memory cost associated with maintaining style consistency, caused by cross-view attention. This limitation restricts the number of viewpoints and necessitates post-processing steps, such as inpainting and img2img. Future work could investigate hierarchical style consistency approaches to reduce the computational costs of cross-view attention by iterating over a smaller set of viewpoints." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.83, + 0.894, + 0.888 + ], + "angle": 0, + "content": "Acknowledgement: This work is supported by the Young Elite Scientists Sponsorship Program by CAST (China Association for Science and Technology) and IEG Moonshot Program by Tencent." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4627" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.44, + 0.129 + ], + "angle": 0, + "content": "[1] Turbosquid by shutterstock. In www.turbosquid.com. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.131, + 0.47, + 0.201 + ], + "angle": 0, + "content": "[2] Tianshi Cao, Karsten Kreis, Sanja Fidler, Nicholas Sharp, and Kangxue Yin. Texfusion: Synthesizing 3d textures with text-guided image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4169-4181, 2023. 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.203, + 0.471, + 0.286 + ], + "angle": 0, + "content": "[3] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.288, + 0.471, + 0.357 + ], + "angle": 0, + "content": "[4] Eric R Chan, Koki Nagano, Matthew A Chan, Alexander W Bergman, Jeong Joon Park, Axel Levy, Miika Aittala, Shalini De Mello, Tero Karras, and Gordon Wetzstein. Generative novel view synthesis with 3d-aware diffusion models. arXiv preprint arXiv:2304.02602, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.36, + 0.47, + 0.428 + ], + "angle": 0, + "content": "[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.431, + 0.47, + 0.486 + ], + "angle": 0, + "content": "[6] Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text-driven texture synthesis via diffusion models. arXiv preprint arXiv:2303.11396, 2023. 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.488, + 0.47, + 0.543 + ], + "angle": 0, + "content": "[7] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. *Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation.* arXiv preprint arXiv:2303.13873, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.546, + 0.47, + 0.602 + ], + "angle": 0, + "content": "[8] Zhiqin Chen, Kangxue Yin, and Sanja Fidler. Auv-net: Learning aligned uv maps for texture transfer and synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1465-1474, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.604, + 0.471, + 0.659 + ], + "angle": 0, + "content": "[9] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 303-312, 1996. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.47, + 0.744 + ], + "angle": 0, + "content": "[10] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.47, + 0.815 + ], + "angle": 0, + "content": "[11] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. Advances In Neural Information Processing Systems, 35:31841-31854, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.47, + 0.899 + ], + "angle": 0, + "content": "[12] Jiatao Gu, Alex Trevithick, Kai-En Lin, Joshua M Susskind, Christian Theobalt, Lingjie Liu, and Ravi Ramamoorthi. Nerfdiff: Single-image view synthesis with nef-guided distillation from 3d-aware diffusion. In International Conference on Machine Learning, pages 11808-11826. PMLR, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[13] Zehuan Huang, Hao Wen, Junting Dong, Yaohui Wang, Yangguang Li, Xinyuan Chen, Yan-Pei Cao, Ding Liang, Yu Qiao, Bo Dai, and Lu Sheng. Epidiff: Enhancing multi-view synthesis via localized epipolar-constrained diffusion, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.231 + ], + "angle": 0, + "content": "[14] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 867-876, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.233, + 0.892, + 0.288 + ], + "angle": 0, + "content": "[15] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.289, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[16] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.358, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[17] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics (TOG), 39(6):1-14, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.414, + 0.875, + 0.427 + ], + "angle": 0, + "content": "[18] Oliver Laric. Three d scans. In threadscans.com, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.428, + 0.892, + 0.51 + ], + "angle": 0, + "content": "[19] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300–309, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.511, + 0.892, + 0.565 + ], + "angle": 0, + "content": "[20] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Zexiang Xu, Hao Su, et al. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. arXiv preprint arXiv:2306.16928, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.567, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[21] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9298–9309, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.637, + 0.892, + 0.692 + ], + "angle": 0, + "content": "[22] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.692, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[23] Xiaoxiao Long, Cheng Lin, Peng Wang, Taku Komura, and Wenping Wang. Sparseneus: Fast generalizable neural surface reconstruction from sparse views. In European Conference on Computer Vision, pages 210-227. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.763, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[24] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. arXiv preprint arXiv:2310.15008, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12663–12673, 2023. 2, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4628" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[26] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.47, + 0.218 + ], + "angle": 0, + "content": "[27] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 conference papers, pages 1-8, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.47, + 0.287 + ], + "angle": 0, + "content": "[28] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.469, + 0.329 + ], + "angle": 0, + "content": "[29] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.469, + 0.411 + ], + "angle": 0, + "content": "[30] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, et al. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. arXiv preprint arXiv:2306.17843, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.413, + 0.469, + 0.495 + ], + "angle": 0, + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.469, + 0.552 + ], + "angle": 0, + "content": "[32] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.553, + 0.469, + 0.594 + ], + "angle": 0, + "content": "[33] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. arXiv preprint arXiv:2302.01721, 2023. 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.47, + 0.663 + ], + "angle": 0, + "content": "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.469, + 0.747 + ], + "angle": 0, + "content": "[35] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.469, + 0.83 + ], + "angle": 0, + "content": "[36] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.469, + 0.872 + ], + "angle": 0, + "content": "[37] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[38] Yawar Siddiqui, Justus Thies, Fangchang Ma, Qi Shan, Matthias Nießner, and Angela Dai. Texturify: Generating" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "textures on 3d shape surfaces. In European Conference on Computer Vision, pages 72-88. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[39] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.163, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[40] Jia-Mu Sun, Tong Wu, and Lin Gao. Recent advances in implicit representation-based 3d shape generation. Visual Intelligence, 2(1):9, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[41] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2818-2826, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.277, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[42] Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi. Viewset diffusion:(0-) image-conditioned 3d generative models from 2d data. arXiv preprint arXiv:2306.07881, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.333, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[43] Shitao Tang, Fuyang Zhang, Jiacheng Chen, Peng Wang, and Yasutaka Furukawa. Mvdiffusion: Enabling holistic multiview image generation with correspondence-aware diffusion. arXiv preprint arXiv:2307.01097, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.391, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[44] Hung-Yu Tseng, Qinbo Li, Changil Kim, Suhib Alsisan, Jia-Bin Huang, and Johannes Kopf. Consistent view synthesis with pose-guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16773-16783, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.461, + 0.892, + 0.516 + ], + "angle": 0, + "content": "[45] Greg Turk and Marc Levoy. Zippered polygon meshes from range images. In Proceedings of the 21st annual conference on Computer graphics and interactive techniques, pages 311-318, 1994. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.518, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[46] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12619-12629, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.588, + 0.892, + 0.643 + ], + "angle": 0, + "content": "[47] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.645, + 0.892, + 0.686 + ], + "angle": 0, + "content": "[48] Jianfeng Xiang, Jiaolong Yang, Binbin Huang, and Xin Tong. 3d-aware image generation using 2d diffusion models. arXiv preprint arXiv:2303.17905, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.688, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[49] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. arXiv preprint arXiv:2306.07954, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.731, + 0.886, + 0.745 + ], + "angle": 0, + "content": "[50] Jonathan Young. xatlas. In github.com/jpcy/xatlas, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.746, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[51] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.801 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4629" + } + ] +] \ No newline at end of file diff --git a/2024/GenesisTex_ Adapting Image Denoising Diffusion to Texture Space/ba009eff-00cf-485a-8674-0d1aa83b6a01_origin.pdf b/2024/GenesisTex_ Adapting Image Denoising Diffusion to Texture Space/ba009eff-00cf-485a-8674-0d1aa83b6a01_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..81a52d1ad686c5ff58e249c456d3fb4dbb553f01 --- /dev/null +++ b/2024/GenesisTex_ Adapting Image Denoising Diffusion to Texture Space/ba009eff-00cf-485a-8674-0d1aa83b6a01_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7560f0f372d9cf9beb12dbd97dde51dd1dbd9a688ce364c5e96920fbb6a6b321 +size 4263539 diff --git a/2024/GenesisTex_ Adapting Image Denoising Diffusion to Texture Space/full.md b/2024/GenesisTex_ Adapting Image Denoising Diffusion to Texture Space/full.md new file mode 100644 index 0000000000000000000000000000000000000000..44ea5c556fdf53230d14c2a88216d5f0a2c5f1c4 --- /dev/null +++ b/2024/GenesisTex_ Adapting Image Denoising Diffusion to Texture Space/full.md @@ -0,0 +1,376 @@ +# GenesisTex: Adapting Image Denoising Diffusion to Texture Space + +Chenjian Gao $^{1,2\dagger}$ Boyan Jiang $^{2}$ Xinghui Li $^{2}$ Yingpeng Zhang $^{2\#}$ Qian Yu $^{1*}$ $^{1}$ School of Software, Beihang University + $^{2}$ R&D Efficiency and Capability Department, Tencent IEG +https://cjeen.github.io/GenesisTexPaper/ + +# Abstract + +We present GenesisTex, a novel method for synthesizing textures for 3D geometries from text descriptions. GenesisTex adapts the pretrained image diffusion model to texture space by texture space sampling. Specifically, we maintain a latent texture map for each viewpoint, which is updated with predicted noise on the rendering of the corresponding viewpoint. The sampled latent texture maps are then decoded into a final texture map. During the sampling process, we focus on both global and local consistency across multiple viewpoints: global consistency is achieved through the integration of style consistency mechanisms within the noise prediction network, and low-level consistency is achieved by dynamically aligning latent textures. Finally, we apply reference-based inpainting and img2img on denser views for texture refinement. Our approach overcomes the limitations of slow optimization in distillation-based methods and instability in inpainting-based methods. Experiments on meshes from various sources demonstrate that our method surpasses the baseline methods quantitatively and qualitatively. + +# 1. Introduction + +In recent years, with the development of deep learning, 3D content generation technology has made significant progress. The applications of 3D content generation are diverse, ranging from AR/VR to gaming and filmmaking. While there has been considerable research on deep learning-based geometric asset generation[40], there has been a notable industry demand for generating realistic textures for given geometries. + +Recently, text-conditioned image diffusion models [34] have achieved impressive results in image generation. Some works [7, 19, 29] have leveraged text-conditioned image diffusion models to generate textured 3D assets by gener + +![](images/375f4d1e1db8ab1323d0a75ed48f511ef8e3f009a1755fe6418f02c50d1e724f.jpg) +red volkswagon beetle, cartoon style +Figure 1. Texturing results of different methods. Score Distillation Sampling (SDS) based method produces blurred and oversaturated textures. Inpainting-based approach results in artifacts at the boundaries of inpainting masks. Texture space sampling concurrently generating content from multiple viewpoints, produces clean, clear and natural colored textures. + +ating content from multiple viewpoints, achieving notable performance. In this work, our focus is on generating high-quality textures for a given geometry, leveraging the priors provided by pre-trained text-to-image diffusion models. This task poses several challenges, including: 1) View consistency: ensuring cross-view constraints for maintaining low-level consistency; 2) High efficiency: generating textures for a model within a few minutes, enabling practical applications; 3) Zero-shot learning: achieving texture generation without requiring additional training or finetuning. Addressing these challenges is crucial for successful application of image diffusion models to the texture domain. + +Currently, the most prominent method for text-to-3D generation is Score Distillation Sampling (SDS) [7, 19, 29]. SDS utilizes the prior knowledge from image diffusion + +models to iteratively generate gradients for rendering images of 3D objects. However, SDS has limitations in terms of both efficiency and quality. It often takes several hours to generate a single textured 3D object, and the resulting textures can suffer from over-saturation due to the adoption of a large CFG (classifier-free guidance) scale, which reduces randomness during generation. Recent works [2, 6, 33] have introduced inpainting-based methods for multi-view generation. These methods establish a predetermined order of views, using the content of previously generated views as a condition for subsequent views. While inpainting-based approaches are faster and produce realistic colors, they are sensitive to the predefined view order. Additionally, they require texture segmentation to determine the source view for each texture pixel, which often results in artifacts around the borders of the segmented areas. + +To address the aforementioned issues, we propose GenesisTex, a novel approach that introduces texture space sampling. Specifically, we maintain multiple-view latent texture maps throughout the sampling process and perform denoising to improve their quality. Subsequently, the content is decoded from the latent space to obtain RGB textures. Our approach focuses on achieving two aspects of consistency during the sampling process. Firstly, we ensure global style consistency across multiple views by incorporating style consistency in the noise prediction network. This helps maintain a coherent style throughout the generated textures. Secondly, we employ dynamic alignment of latent textures to ensure low-level multi-view consistency, enhancing the overall quality of the generated textures. Due to memory limitations, texture space sampling operates on sparse views. However, to further enhance the quality, we leverage reference-based inpainting and Img2Img techniques on denser views for texture map refinement. GenesisTex can generate detailed, clean, and naturally colored textures for a given geometry within a few minutes. Unlike inpainting-based methods, our proposed texture space sampling concurrently generates content for multiple views without relying on a predefined order. This adaptability makes GenesisTex suitable for various geometries and results in fewer artifacts in the generated textures. + +Our contributions can be summarized as follows: + +- First, we present a novel method for texture generation, where the core is texture space sampling. This sampling technique allows for the concurrent denoising of latent textures associated with multiple viewpoints. +- Second, we introduce Style Consistency and Dynamic Alignment in texture space sampling for multi-view consistency. +- Third, we conduct a comprehensive study involving numerous 3D objects from various sources. The experimental results demonstrate the superiority of our method over baseline methods. + +# 2. Related Work + +Texture Synthesis. Generating textures over 3D surfaces is a challenging problem, as it requires attention to both colors and geometry. Earlier works like AUV-Net [8] and Texturify [38] embed the geometric prior into a UV map or mesh parameterization. Different from them, EG3D [3] and GET3D [11] directly train 3D StyleGANs [15] to generate geometries and textures jointly, where the textures are implicit texture fields. However, these methods either only work on a single category, or demand textured 3D shapes for training without text conditioning, which limits their broad applicability. Recently, [2, 6, 33] use the priors provided by the image diffusion model to synthesize textures. Text2Tex [6] and TEXTure [33] performs inpainting on multiview renderings. TexFusion [2] propose a sequential texture sampling method. Our method predicts multi-view noise concurrently, avoiding the following issues of sequential noise prediction: 1) Each iteration requires adding forward noise to visited regions to match unvisited areas, potentially leading to detail loss. 2) Consistency is only constrained between adjacent viewpoints, with no direct consistency for long-range viewpoints. + +Text-to-Image Diffusion Models. Over the past years, the development of several large-scale diffusion models [28, 32, 34, 35] has enabled the production of highly detailed and visually impressive images. These models generate images based on input text prompts. Specifically, Stable Diffusion is trained on a substantial text-image dataset [36] and incorporates a text encoder from CLIP [31] to understand the input prompts. Beyond the basic text conditioning, ControlNet [51] enables the model to condition its denoising network on additional input modalities, such as depth maps. In this work, we utilize ControlNet and Stable Diffusion to provide geometrically-conditioned image priors. + +Text-to-3D using 2D Image Diffusion Models. Early works [14, 26, 27] utilize the pretrained CLIP model to maximize the similarity between rendered images and text prompt. However, pioneering works DreamFusion [29] and SJC [46], on the other hand, propose to distill a 2D text-to-image generation model to generate 3D shapes from texts, and many follow-up works [19-21, 23, 25, 30, 47] follow such per-shape optimization scheme. Recently, several methods [4, 12, 13, 22, 24, 37, 42-44, 48] have been proposed to generate consistent multi-view images by using diffusion models. MVDream [37] and SyncDreamer [22] share similar ideas, generating consistent multi-view images via attention layers. However, existing approaches either endure slow optimization processes or depend on separately trained 3D priors, rendering them unsuitable for direct application in texture synthesis. In contrast, our method does not require additional training and can generate results within several minutes. + +# 3. Methodology + +Our objective is to generate a texture map $\mathcal{T}$ for a given 3D mesh $\mathcal{M}$ , using the provided text as a descriptive condition. Our approach leverages the Stable Diffusion model as the image diffusion model. We begin by introducing some basic concept about Stable Diffusion and texture representation in Sec. 3.1. Then, we introduce a sampling algorithm in texture space for Stable Diffusion in Sec. 3.2. Furthermore, we propose multi-view consistency strategy in Sec. 3.3. Finally we perform refinement on the texture map, which includes Inpainting and Img2Img, as described in Sec. 3.4. + +# 3.1. Preliminary + +Stable Diffusion. The diffusion model belongs to a class of generative models that generate data through iterative denoising from random noise. In the sampling process, $\mathbf{z}_{i-1}$ can be sampled from $\mathbf{z}_i$ using a DDIM sampler [39]: + +$$ +\begin{array}{l} \hat {\mathbf {z}} _ {i \rightarrow 0} = (\mathbf {z} _ {i} - \sqrt {1 - \alpha_ {i}} \epsilon_ {\theta} (\mathbf {z} _ {i}, t _ {i}, \mathbf {c} _ {t e x t})) / \sqrt {\alpha_ {i}}, \\ \mathbf {z} _ {i - 1} = \sqrt {\alpha_ {i - 1}} \hat {\mathbf {z}} _ {i \rightarrow 0} + \sqrt {1 - \alpha_ {i - 1} - \sigma_ {i} ^ {2}} \epsilon_ {\theta} (\mathbf {z} _ {i}, t _ {i}, \mathbf {c} _ {t e x t}) \\ + \sigma_ {i} \varepsilon , \tag {1} \\ \end{array} +$$ + +where $\epsilon_{\theta}$ is a pretrained U-Net for noise prediction and the initial $\mathbf{z}_T \sim \mathcal{N}(0, \mathbf{I})$ . In this paper, we specifically employ Stable Diffusion [34]. Stable Diffusion performs denoising in the latent space and employs an autoencoder $\mathcal{D}(\mathcal{E}(\cdot))$ for the conversion between image and latent representations. So $\mathbf{z}_0$ generated by the diffusion process is decoded to the image $\mathcal{D}(\mathbf{z}_0)$ finally. + +ControlNet. ControlNet [51] injects low-level control during the denoising process of Stable Diffusion. Our approach employs depth $\mathbf{d}$ as the geometric control condition. The predicted noise of the U-Net with ControlNet is represented as $\epsilon_{\theta}(\mathbf{z}_i, t_i, \mathbf{c}_{text}, \mathbf{d})$ . + +Texture Representation. In this paper, we utilize mesh as the 3D representation. Texture map is associated with an UV parameterization of the mesh $\mathcal{M}$ . We use $x\text{atlas}$ [50] to generate the UV parameterization. + +**Rendering.** Given a mesh $\mathcal{M}$ , a texture map $\mathcal{T}$ and a viewpoint $C$ , we can use the rendering function $\mathcal{R}$ to get the rendered image $\mathbf{x}^{(img)} = \mathcal{R}(\mathcal{T};\mathcal{M},C)$ . Similar to [2], we do not consider any lighting and rendering involves solely sampling colors from the texture map. The inverse rendering function $\mathcal{R}^{-1}$ can inverse render the image $\mathbf{x}^{(img)}$ back to a texture map $\mathcal{T}' = \mathcal{R}^{-1}(\mathbf{x}^{(img)};\mathcal{M},C)$ . + +# 3.2. Sampling in Texture Space + +The most straightforward method for sampling on the texture map is to fine-tune the Stable Diffusion model directly on some texture maps, making its distribution more like the distribution of texture maps. However, the distribution of texture maps is much more complex than that of rendering + +Algorithm 1 Texture Space Sampling +Input: mesh $\mathcal{M}$ ,cameras $\{C_1,\dots ,C_N\}$ +Parameters:Denoising time schedule $\{t_i\}_{i = T}^0$ ,DDIM +noise schedule $\{\sigma_i\}_{i = T}^0$ $\{\mathbf{z}_{T,n}^{(tex)}\}_{n = 1}^N\gets \{\mathbf{0}\}$ $\{\mathbf{z}_n^{(bg)}\}_{n = 1}^N\sim \{\mathcal{N}(\mathbf{0},\mathbf{I})\}$ $\{\hat{\epsilon}_{T + 1,n}\}_{n = 1}^N\sim \{\mathcal{N}(\mathbf{0},\mathbf{I})\}$ +denoising stage: +for $i\in \{T\dots 1\}$ do +for $n\in \{1\dots N\}$ do + $\varepsilon_{n}\sim \mathcal{N}(0,I)$ ${\bf z}_{i,n}^{(img)}\gets {\bf M}_n(\sqrt{\alpha_i}{\cal R}({\bf z}_{i,n}^{(tex)}) + \sqrt{1 - \alpha_i - \sigma_{i + 1}^2}\hat{\epsilon}_{i + 1,n})$ $+(1 - \mathbf{M}_n)\mathbf{z}_n^{(bg)} + \sigma_{i + 1}\varepsilon_n$ $\hat{\epsilon}_{i,n}\gets \epsilon_{\theta}(\mathbf{z}_{i,n}^{(img)},t_i,\mathbf{c},\mathbf{d}_n)$ $\hat{\mathbf{z}}_{0,n}^{(img)}\gets (\mathbf{z}_{i,n}^{(img)} - \sqrt{1 - \alpha_i}\hat{\epsilon}_{i,n}) / \sqrt{\alpha_i}$ ${\bf z}_{i - 1,n}^{(tex)}\gets \mathcal{R}^{-1}(\hat{\mathbf{z}}_{0,n}^{(img)})$ ${\bf z}_{n}^{(bg)}\gets \sqrt{\alpha_{i - 1}}\hat{\mathbf{z}}_{0,n}^{(img)} + \sqrt{1 - \alpha_{i - 1} - \sigma_i^2}\cdot \hat{\epsilon}_{i,n}$ +end for + $\{{\bf z}_{i - 1,n}^{(tex)}\}_{n = 1}^N = dynamic\_ align(\{{\bf z}_{i - 1,n}^{(tex)}\}_{n = 1}^N)$ +end for +decoding stage: +for $n\in \{1\dots N\}$ do + ${\bf x}_n^{(img)} = \mathcal{D}(\hat{\bf z}_{0,n}^{(img)})$ ${\bf x}_n^{(tex)} = R^{-1}({\bf x}_n^{(img)})$ +end for +T=merge({x(tex)})N +return Texture map Tsampled + +images since different UV parameterizations of the same 3D object correspond to different texture maps. Instead of fine-tuning existing image diffusion models, we adapt them to the texture space for texture map generation. + +Algorithm 1 presents our proposed texture space sampling algorithm. Firstly we define a set of $\mathcal{C}^{(sampling)} = \{C_1,..,C_N\}$ camera views and render the corresponding depth map $\mathbf{d}_n$ in each view. Since we employ a latent diffusion model, our texture space sampling method first performs denoising in the latent texture space and then decode the latent to texture image. In the denoising stage, a dedicated diffusion denoising algorithm is applied to a set of latent texture maps $\{\mathbf{z}_n^{(tex)}\}_{n=1}^N$ . In the decoding stage, we recover the texture map $\mathcal{T}_{sampled}$ from the latent space. + +Denoising Stage. In the denoising stage, we adapt the DDIM from the rendering space to the texture space and utilize the same parameters as DDIM, including denoising time schedule $\{t_i\}_{i = T}^0$ , DDIM noise schedule $\{\sigma_i\}_{i = T}^0$ . For each viewpoint, we maintain a separate latent texture which dynamically interact during the denoising process. We begin with $\{\mathbf{z}_{T,n}^{(tex)}\}_{n = 1}^{N} = \{\mathbf{0}\}$ , the zero latent texture map + +![](images/9b1d8a67dfd9b827f94980ff551b3ef6522b236bf388670b26a3d08c2d688ffe.jpg) +Figure 2. Overview of GenesisTex. GenesisTex generates a texture map for a given mesh $\mathcal{M}$ , based on a prompt. Texture Space Sampling samples a texture map using Stable Diffusion, introducing style consistency and dynamic alignment across multiple viewpoints. Furthermore, Inpainting and Img2Img are applied to fill in the blank regions and enhance the quality of texture map details, respectively. + +![](images/f658c0f6c5646555c3e8d6eb1749aab590aa4831ff414b7bc70c790dfa9b67ce.jpg) + +![](images/412af3afeb20fccbaee64f218c7e5ed5a7547a1fa13126e6915321f0c9f92dd5.jpg) + +![](images/63dcb9d3bf64b3fd97c2ec66455bc3322d028e86115878d631fedd9975f57a3a.jpg) + +corresponding to $N$ viewpoints. At each denoising step, our goal is to predict $\mathbf{z}_{i - 1,n}^{(tex)}$ from $\mathbf{z}_{i,n}^{(tex)}$ . We first render the latent texture map $\mathbf{z}_{i,n}^{(tex)}$ to the rendering space using the rendering function $\mathcal{R}$ , and then calculate the corresponding latent image $\mathbf{z}_{i,n}^{(img)}$ : + +$$ +\begin{array}{l} \mathbf {z} _ {i, n} ^ {(i m g)} = \mathbf {M} _ {n} (\sqrt {\alpha_ {i}} \mathcal {R} (\mathbf {z} _ {i, n} ^ {(t e x)}) + \sqrt {1 - \alpha_ {i} - \sigma_ {i + 1} ^ {2}} \hat {\epsilon} _ {i + 1, n}) \\ + \left(1 - \mathbf {M} _ {n}\right) \mathbf {z} _ {n} ^ {(b g)} + \sigma_ {i + 1} \varepsilon_ {n} \tag {2} \\ \end{array} +$$ + +Here, $\mathbf{M}_n$ represents the foreground mask for viewpoint $C_n$ (downsampled to the same resolution as $\mathbf{z}^{(img)})$ , where foreground pixels have a value of 1, and background pixels have a value of 0. $\hat{\epsilon}_{i + 1,n}$ represents the noise predicted by the noise prediction network at step $i + 1\rightarrow i$ (notably, we define $\hat{\epsilon}_{T + 1,n}\sim \mathcal{N}(0,\mathbf{I})$ ). $\mathbf{z}_n^{(bg)}$ is the background of the latent image, which is set to Gaussian noise at initialization. $\varepsilon_{n}\sim \mathcal{N}(0,\mathbf{I})$ is a random Gaussian noise. Subsequently, we predict the corresponding noise $\hat{\epsilon}_{i,n} = \epsilon_{\theta}(\mathbf{z}_{i,n}^{(img)},t_i,\mathbf{c},\mathbf{d}_n)$ for the latent image $\mathbf{z}_{i,n}^{(img)}$ conditioned on the depth map $\mathbf{d}_n$ . With the predicted noise, we can obtain the predicted $\hat{\mathbf{z}}_{0,n}^{(img)}$ : + +$$ +\hat {\mathbf {z}} _ {0, n} ^ {(i m g)} = (\mathbf {z} _ {i, n} ^ {(i m g)} - \sqrt {1 - \alpha_ {i}} \hat {\epsilon} _ {i, n}) / \sqrt {\alpha_ {i}}, \qquad (3) +$$ + +Finally, we inverse render $\hat{\mathbf{z}}_{0,n}^{(img)}$ to the texture space using $\mathcal{R}^{-1}$ and update $\mathbf{z}_n^{(bg)}$ : + +$$ +\mathbf {z} _ {i - 1, n} ^ {(t e x)} = \mathcal {R} ^ {- 1} \left(\hat {\mathbf {z}} _ {0, n} ^ {(i m g)}\right), \tag {4} +$$ + +$$ +\mathbf {z} _ {n} ^ {(b g)} = \sqrt {\alpha_ {i - 1}} \hat {\mathbf {z}} _ {0, n} ^ {(i m g)} + \sqrt {1 - \alpha_ {i - 1} - \sigma_ {i} ^ {2}} \cdot \hat {\epsilon} _ {i, n} +$$ + +In the denoising stage, we repeat the above process $T - 1$ times. After each denoising step, we impose consistency + +constraints on latent texture $\mathbf{z}_{i - 1,n}^{(tex)}$ , which will be introduced in Sec. 3.3. Notably, differing from DDIM where the latent includes noise, the latent texture in our method is an estimate of the noise-free version. This is because $\mathcal{R}$ and $\mathcal{R}^{-1}$ in the denoising process require interpolation of the latent, which may distort the noise. + +Decoding Stage. To recover the texture map $\mathcal{T}_{\text{sampled}}$ , we first decode each viewpoint's latent separately to obtain multi-view rendering. Subsequently, we inverse render the renderings into the texture space. After obtaining multiview texture maps, we merge them into a single texture map as the output. Similar to [2, 6, 33], we aim for each pixel on each texture map to come from the viewpoint where the corresponding point on the mesh is observed most directly. Therefore, our merge is defined as follows: + +$$ +\mathcal {T} _ {\text {s a m p l e d}} = \sum_ {n = 1} ^ {N} \operatorname {S o f t m a x} \left(\mathcal {R} ^ {- 1} \left(\mathbf {N} _ {\mathbf {n}}\right)\right) \times \mathbf {x} _ {n} ^ {(t e x)} \tag {5} +$$ + +$\mathbf{N}_{\mathbf{n}}$ is the similarity mask at viewpoint $C_n$ , where each pixel represents the cosine similarity between the normal vectors of the visible faces and the reversed view direction. We use Softmax instead of Max because Softmax can result in more natural transition in the texture maps. + +# 3.3. Consistency between Latent Texture Maps + +# 3.3.1 Dynamic Alignment + +During sampling in texture space, we maintain $N$ latent texture maps, now we introduce an alignment approach to ensure their local consistency. Firstly, we directly reduce these latent texture maps to obtain a uniform latent texture map: + +$$ +\mathbf {z} _ {i - 1} ^ {(u n i t e x)} = \sum_ {n = 1} ^ {N} \operatorname {S o f t m a x} \left(\mathcal {R} ^ {- 1} \left(\mathbf {N} _ {n}\right)\right) \times \mathbf {z} _ {i - 1, n} ^ {(t e x)} \tag {6} +$$ + +Next, we blend each viewpoint's latent texture map $\mathbf{z}_{i-1,n}^{(tex)}$ with the uniform latent texture map $\mathbf{z}_{i-1}^{(unitex)}$ . We do not enforce strict alignment among the $N$ latent texture maps at every step. Since a latent pixel gains full meaning only when combined with the surrounding context, which is viewpoint-dependent, the latent texture map from each viewpoint should contain some independent information. Moreover, the content generated at different timestamps varies, so the alignment constraints should adapt flexibly. We introduce a signal $c(t)$ for dynamically regulating the variations in consistency during the sampling process: + +$$ +\mathbf {z} _ {i - 1, n} ^ {(c o n s i s t e x)} = c (t _ {i}) \times \mathbf {z} _ {i - 1} ^ {(u n i t e x)} + (1 - c (t _ {i})) \times \mathbf {z} _ {i - 1, n} ^ {(t e x)} (7) +$$ + +We empirically find that maintaining strong local consistency in the mid-term of denoising, while keeping weak local consistency in the early and late stages, leads to higher quality generation results. + +# 3.3.2 Style Consistency + +In texture space sampling, we use the noise estimation network $\epsilon_{\theta}$ to estimate noise for the multi-view latents, which is responsible for content generation. So we adapt $\epsilon_{\theta}$ to ensure style consistency of estimated noise across different views. Inspired by video diffusion models [16, 49] and multi-view diffusion model [37], we modify all the self-attention layers and group normalization layers in the noise prediction network to align the style of multi-view content. Adapted Self-Attention. In Stable Diffusion, self-attention facilitates long-range interactions among features within an image. We transform self-attention into cross-view attention to establish style consistency across multiple viewpoints, by using the key $\mathbf{K}^{\prime}$ and value $\mathbf{V}^{\prime}$ from all viewpoints, + +$$ +\mathbf {K} ^ {\prime} = \mathbf {W} ^ {K} \left[ \mathbf {z} _ {i, 1} ^ {(i m g)}; \dots ; \mathbf {z} _ {i, n} ^ {(i m g)} \right], \tag {8} +$$ + +$$ +\mathbf {V} ^ {\prime} = \mathbf {W} ^ {V} [ \mathbf {z} _ {i, 1} ^ {(i m g)}; \dots ; \mathbf {z} _ {i, n} ^ {(i m g)} ]. +$$ + +where $\mathbf{W}^K$ , $\mathbf{W}^V$ are pretrained parameters. + +Adapted Group Normalization. Stable diffusion adopt group normalization as the normalization layer. Group normalization divides the channels into groups and computes within each group the mean and variance for normalization. We convert 2D group normalization to 3D by connecting all different views in the group normalization layer, i.e., the mean and variance are calculated within the channel group of all pixels across all viewpoints. + +By adjusting self-attention layer and group normalization layer, we achieve multi-view style consistency without any additional training. In addition to the noise prediction network $\epsilon_{\theta}$ , we apply the same modifications to the decoder $\mathcal{D}$ to further ensure style consistency in the multi-view decoded RGB images. + +# 3.4. Texture Map Refinement + +# 3.4.1 Post-processing: Inpainting + +Since the memory cost of texture space sampling increases with the number of viewpoints of $\mathcal{C}_{\text{sampling}}$ , the number of viewpoints $N$ is limited. Some regions in the texture map $\mathcal{T}_{\text{sampled}}$ may not be observed in any of the viewpoints in $\mathcal{C}_{\text{sampling}}$ . Therefore, after texture space sampling, we introduce an inpainting epoch to fill texture in areas not observed in $\mathcal{C}$ . + +Firstly we define a denser set of viewpoints $\mathcal{C}^{(inpainting)} = \{C_1,..,C_{N_1}\}$ as compared to $\mathcal{C}^{(samping)}$ where $N_{1} > N$ . We use a mask $\mathbf{M}_C^{blank}$ to indicate the rendered areas with blank textures at the viewpoint $C$ . The entire inpainting epoch comprises $N_{1}$ iterations. In each iteration, we first compute the viewpoint $C_n = \arg \max_C\mathbf{M}_C^{blank}$ with the largest blank area. Then, we perform inpainting on the rendering $\mathbf{x}_n^{(img)}$ using the pretrained depth conditioned Stable Diffusion model as described in [6, 33]. During inpainting, Stable Diffusion utilizes regions with textures as conditions to fill in blank areas without textures. To enhance the robustness of inpainting, we render an image using a viewpoint $C_0$ from the set $\mathcal{C}^{(sampling)}$ . We then concatenate this image to the image requiring inpainting as a additional condition for inpainting. Compared to naive inpainting, reference-based inpainting provides a more comprehensive condition. To ensure a natural and smooth transition between blank areas and textured regions, we apply Gaussian blur to the mask $\mathbf{M}_{C_n}^{blank}$ . Finally the inpainted rendering $\mathbf{x}_n^{(img)'}$ is used to update the texture map: + +$$ +\begin{array}{l} \mathcal {T} _ {n} = \mathcal {T} _ {n - 1} \times (1 - \mathcal {R} ^ {- 1} (\mathbf {M} ^ {(b l a n k)}) \tag {9} \\ + \mathcal {R} ^ {- 1} \left(\mathbf {x} _ {n} ^ {(i m g) \prime}\right) \times \mathcal {R} ^ {- 1} \left(\mathbf {M} ^ {(b l a n k)}\right) \\ \end{array} +$$ + +We start with $\mathcal{T}_0 = \mathcal{T}_{\text{sampled}}$ and iterate for $N_1$ times. Finally we get $\mathcal{T}_{\text{inpainted}} = \mathcal{T}_{N_1}$ . + +It's worth noting that, unlike [2, 6, 33], which relies heavily on inpainting to maintain multi-view consistency, in our method inpainting is performed only to patch small areas. Therefore, the inpainting's performance does not significantly affect the final quality of the generated texture. + +# 3.4.2 Post-processing:Img2Img + +So far our generated texture map $\mathcal{T}_{\text {inpainted }}$ may still contain unnatural transitions in some regions due to multi-view conflicts and inpainting artifacts. To further improve the generation quality, we introduce an Img2Img epoch. + +In Img2Img epoch, we also define a set of viewpoints $C^{(img2img)}$ include $N_{2}$ viewpoints and iterate through viewpoint from $C^{(img2img)}$ . In each iteration, we encode + +the rendering image $\mathbf{x}_n^{(img)}$ to the latent space using the encoder $\mathcal{E}$ and add some noise to the latent rendering. Then we perform denoising using the pretrained depth conditioned Stable Diffusion model. And finally we decode the denoised latent back to image space using $\mathcal{D}$ . However, $\mathcal{D}(\mathcal{E}(\cdot))$ is not an identity transformation because of the reconstruction distortion of the autoencoder. So we apply the color histogram matching to the foreground of the decoded image to alleviate the color distortion. Finally we update the texture map using the img2img result $\mathbf{x}_n^{(img)'}$ based on view similarity $\mathbf{N}_n$ : + +$$ +\begin{array}{l} \mathcal {T} _ {n} = \mathcal {T} _ {n - 1} \times (1 - \mathcal {R} ^ {- 1} (\mathbf {N} _ {n})) \\ + \mathcal {R} ^ {- 1} \left(\mathbf {x} _ {n} ^ {(i m g) \prime}\right) \times \mathcal {R} ^ {- 1} \left(\mathbf {N} _ {n}\right) \tag {10} \\ \end{array} +$$ + +We start with $\mathcal{T}_0 = \mathcal{T}_{\text{inpainted}}$ . After $N_2$ iterations we get $\mathcal{T}_{\text{refined}} = \mathcal{T}_{N_2}$ . + +# 4. Experiments + +# 4.1. Setup + +Implementation Details. Our experiments are conducted on an NVIDIA A10 GPU. We utilize the official Stable Diffusion-v1.5 model with the ControlNet-v1.1 (depth). For Inpainting epoch and Img2Img epoch, we use DDIM [39] as the sampler. For all samplers, we set the number of iterations to 20 steps and the CFG scale (classifier-free guidance scale) to 7.5. We implement the rendering function using nvdiffrast [17] and make modifications to nvdiffrast to support inverse rendering. Texture space sampling takes about 3 minutes to generate textures for a single object. For more detailed parameter settings, please refer to the supplementary materials. + +Dataset. Similar to TexFusion [2], we utilize 35 meshes to test the texture generation performance, including 11 objects from Objaverse [10], 17 from ShapeNet [5], 1 from Text2Mesh [26], 2 from Turbosquid [1], 2 from Stanford 3D Scans [9, 45] and 2 from Three D Scans [18]. Each object has 1-4 text descriptions, resulting in a total of 80 (mesh, prompt) pairs in this collection. + +Baselines. We conduct comparisons with state-of-the-art methods for generating textures from text: 1) Text2Mesh [26], a method that stylizes a 3D mesh by predicting color and local geometric details which conform to a target text prompt, harnessing the representational power of CLIP. 2) Latent-Paint [25], an approach leveraging SDS to obtain texture gradients from the Stable Diffusion model. 3) Text2Tex [6], a method that uses a depth-aware image inpainting diffusion model to incrementally produce partial textures from various viewpoints. 4) TEXTure [33], similar to Text2Tex, but utilizes a different region segmentation strategy. 5) TexFusion [2], a method sequentially inpainting latent texture. As TexFusion does not have publicly + +
MethodFID (↓)KID (↓) (×10-3)User study (%)
Visual Quality (↑)Align with Prompt (↑)
Latent-Paint110.1410.645.154.28
Text2Mesh121.6115.133.154.86
Text2Tex101.388.3518.2819.72
TEXTure100.479.2223.4224.86
Ours74.582.8950.0046.28
+ +Table 1. Quantitative comparisons with baseline methods. + +
texture space samplingInpainting RoundImg2Img RoundFID(↓)KID(↓) (×10-3)
✓✗✗86.253.79
✓✗✓82.053.67
✓✓✗76.303.29
✓✓✓74.582.89
+ +Table 2. Effectiveness of texture refinement. + +available source code, we extract some rendered images of generated results from their paper for comparison. + +# 4.2. Qualitative Analysis + +Fig. 3 and Fig. 5 presents visual comparisons between our method and baseline approaches. Textures produced by Text2Mesh are notably lacking in detail. Latent-Paint yields textures with greater intricacy compared to Text2Mesh, but they still exhibit considerable blurriness due to SDS limitations. Inpainting-based approaches like Text2Tex and TEXTure introduce artifacts with unnatural transitions at inpainting mask boundaries, struggling to maintain consistency across multiple views because they depend heavily on inpainting. The output of TexFusion is compromised by blurry details, likely a consequence of repeated noise addition during sequential sampling. In contrast, our method stands out by producing textures that are not only clear but also remarkably coherent. + +Fig. 4 presents more results of our method. Benefiting from the powerful prior of Stable Diffusion, our method can generate textures of diverse styles for different geometries. + +# 4.3. Quantitative Comparisons + +FID & KID. Similar to TexFusion, we sample from pretrained image diffusion model to create ground truth labels. For each mesh, we generate depth maps from the five most common canonical viewpoints: front, back, top, and both sides. These depth maps, along with textual descriptions, serve as inputs to condition the Stable Diffusion model. To ensure the focus remains on the texture of the objects, we modify the background of the ground truth images to be white. We visualize the textures created by various methods + +![](images/fdc48c32a5f11f589a8403e43677db985a342b561a9f5fd1628db52124b015d8.jpg) + +![](images/d2b4d22b3d333ef1d7edae3141d9c9771ecd38c9bd6184a063e54638a0195a46.jpg) + +![](images/cd9686633cb48e567e402d2f5a1e66ca050d98576870fb1a65fc642231d7b828.jpg) +red apple blue handbag with dilapidated Halloween silver trims tap pumpkin head + +![](images/f66f2cec905dfc5ae96e8bae3daf9f318f78329bcf9d2dfd3cb9c6e61c7d4c31.jpg) +Geometry + +![](images/3a88db337c4d909230a438642665240e74ccf6d986ec7814a7af27b32287adb9.jpg) + +![](images/3b06058ff15f1977324d52c080347bdcbe8f500089ab05a2202185df47364ca5.jpg) + +![](images/8ff9f5e1793f898017a1bc68710f842a758019abbf69896eef9bea814761c1eb.jpg) + +![](images/7945def590a373bcbc75c0fb67f16d2b1e5133ea559d029245dda8edbc2b3f83.jpg) +Text2Mesh + +![](images/67daf3b8d0a1cb7133e37f85e82ba0b30d3bdee3f6a8d82e2941c5c70d1d072a.jpg) + +![](images/83261be0dd824e36413c37d8e7f574daa425ea68b3d4c2e6ce7795bd49e806ed.jpg) + +![](images/1f9823e5d109eb8e75ffdc9aa1dc3372594c171c1aa74e9150ac57c7d2e243ac.jpg) + +![](images/5331997ccdbd20e40b77c9a5cf15142ddbcfb028f99c89ddc14fdf87e38a3435.jpg) +Latent-Paint + +![](images/3009c04219e35ab44fb8edcd5e209d390537087a45da4d44ac34789bc6d14312.jpg) + +![](images/daa7fdce027551be5438c7875d2230a42ee1ca694bcf779108c9ed074e7e7d43.jpg) + +![](images/98524417bb1e4c9adffb95fab197ffe25236610af7188078f2466a75d1613fbf.jpg) + +![](images/096a9d24c767616a95b05312f84c308ec0cc7a2a557e4f4e1fd4bbdfbb47fefd.jpg) +Text2Tex + +![](images/fa07765acbd0f50df1805ac367794bb54fddb0f5c3e57ea338641640a0a91a20.jpg) + +![](images/edcbf3c8900402c219af8ede23e95f6fbb7976280aef88e686e3539faeedf629.jpg) + +![](images/828c1245e3f85806bd29afdf6fc9b3679e5ae7c2cd69dec2ef792b8a82e60d95.jpg) + +![](images/a72a7a4f53dee22f4d9d2c6be81a6962747f89409f477c43e87c5a19e303c1c4.jpg) +TEXTure + +![](images/07235b9d9c58d72508c23fdf78daa077cece9bc6d9679d97f5ea04b4883e8476.jpg) + +![](images/4c567eede86508bf4de02ef0a5b81d56a68996781f88cae11fb1fb0c88614f37.jpg) + +![](images/e84a26487879ef9aa4a5b34d08dfb19ae9962f45351b79752c93cef507cb3626.jpg) + +![](images/178bf64a2ad0b905a0aa74428bfc2d9d09743d1c125800715eb0b4a273a7de53.jpg) +Ours + +![](images/5687d71dca2d7f983a1e3308812f90633d508936eed8e3c02bdf84e0086618b8.jpg) +Figure 3. Qualitative comparisons with Text2Mesh [26], Latent-Paint [25], Text2Tex [6] and TEXTure [33]. In comparison with the baselines, our GenesisTex exhibits richer details and fewer artifacts. +Figure 4. Texturing results with GenesisTex. + +from the same five viewpoints and evaluate their quality by calculating the Fréchet Inception Distance (FID) and Kernel Inception Distance (KID). FID and KID measure the feature dissimilarity between two image collections, with feature extraction performed using the Inception V3 [41]. The results in Tab.1 demonstrate that the textures gener + +![](images/da735ee0a098b3bb990e17955ad314fed34a91df832ffb3f5f255f720641bce9.jpg) +Figure 5. Qualitative Comparisons with TexFusion [2]. Images are extracted from their original paper. + +ated by GenesisTex are closer to the ground truth compared to those produced by baselines, indicating a superior synthesis quality. + +User Study. We conduct a user study to compare the quality of textures generated by our method versus those produced by baseline methods. We render all textured mesh results into videos to facilitate the presentation to users. Each + +video contains a rotating textured mesh along with the corresponding prompt. For each questionnaire, we randomly show the users 10 groups of rendered videos. Each group displays results for a specific (mesh, prompt) pair, comparing outputs from baseline methods and our approach. Each volunteer is required to answer two questions for a single group, including which one has the highest visual quality and which one aligns best with the prompt. Additional information about the questionnaire is available in the supplementary. Finally, we received 35 valid responses from the questionnaires, which included the results of comparisons from 350 groups. The results of the user study are shown in Tab. 1. In terms of both visual quality and alignment with the text prompt, our method is preferred by the most participants, with percentages reaching $50.00\%$ and $46.28\%$ , respectively. Furthermore, we conduct a pairwise comparison study and the results are included in the supplementary. + +# 4.4. Ablation Studies + +Consistency in Texture Space Sampling. Our GenesisTex employs style consistency and dynamic alignment for maintaining multi-view consistency during texture space sampling. To investigate the impact of these two consistency strategies, we visualize the decoded multi-view images. We set $\mathcal{C}(elevation, azimuth) = \{(0^{\circ}, 0^{\circ}), (0^{\circ}, 15^{\circ}), (0^{\circ}, 35^{\circ}), (0^{\circ}, 45^{\circ})\}$ to ensure that multiple viewpoints share sufficient content. Fig. 6 illustrates an example with the prompt penguin toy. We can observe that without any consistency constraints, the color of the toy varies significant across different viewpoints. After introducing style consistency, consistency for the toy color improves, but there are still inconsistencies in the hair and ears. Further we use dynamic alignment and we can see the + +![](images/78b4fee5eb22640a4d3f53733c8ef35dcc4c98e20a3fba639f6199fc3b4685ff.jpg) +Figure 6. Ablation results on consistency in texture space sampling. Style Consistency + Dynamic Alignment (bottom row) achieves the best multi-view consistency. + +![](images/0be7867152068e75ea0142d33a1c55951a20f0b7e89b2e58cfcc6c9c45a70819.jpg) +Figure 7. Ablation results on texture refinement. Inpainting fills the blank regions, while Img2Img enhances the quality of details. + +consistency of details has significantly improved. We include more results in the supplementary to further demonstrate the effectiveness of these two consistencies. + +Effectiveness of Texture Refinement. After texture space sampling, we conduct refinement on the texture, including Inpainting and Img2Img. We validate the effectiveness of these two modules and the results are presented in Fig. 7 and Tab. 2. Inpainting fills in areas lacking texture obtained during texture space sampling, contributing significantly to the improvement in visual quality. Img2Img further repairs artifacts and enhances details, leading to a certain improvement in visual quality as well. + +# 5. Conclusion and Limitations + +In this work, we propose a novel method for text-based texture generation, named GenesisTex. GenesisTex leverages the prior of the pre-trained Stable Diffusion model by introducing texture space sampling. Texture space sampling concurrently generates multi-view content without relying on a predefined sequence of views. Our approach can generate high-quality textures for a given 3D model in several minutes. The primary limitation of our method is the significant memory cost associated with maintaining style consistency, caused by cross-view attention. This limitation restricts the number of viewpoints and necessitates post-processing steps, such as inpainting and img2img. Future work could investigate hierarchical style consistency approaches to reduce the computational costs of cross-view attention by iterating over a smaller set of viewpoints. + +Acknowledgement: This work is supported by the Young Elite Scientists Sponsorship Program by CAST (China Association for Science and Technology) and IEG Moonshot Program by Tencent. + +# References + +[1] Turbosquid by shutterstock. In www.turbosquid.com. 6 +[2] Tianshi Cao, Karsten Kreis, Sanja Fidler, Nicholas Sharp, and Kangxue Yin. Texfusion: Synthesizing 3d textures with text-guided image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4169-4181, 2023. 2, 3, 4, 5, 6, 7 +[3] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 2 +[4] Eric R Chan, Koki Nagano, Matthew A Chan, Alexander W Bergman, Jeong Joon Park, Axel Levy, Miika Aittala, Shalini De Mello, Tero Karras, and Gordon Wetzstein. Generative novel view synthesis with 3d-aware diffusion models. arXiv preprint arXiv:2304.02602, 2023. 2 +[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 6 +[6] Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text-driven texture synthesis via diffusion models. arXiv preprint arXiv:2303.11396, 2023. 2, 4, 5, 6, 7 +[7] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. *Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation.* arXiv preprint arXiv:2303.13873, 2023. 1 +[8] Zhiqin Chen, Kangxue Yin, and Sanja Fidler. Auv-net: Learning aligned uv maps for texture transfer and synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1465-1474, 2022. 2 +[9] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 303-312, 1996. 6 +[10] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 6 +[11] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. Advances In Neural Information Processing Systems, 35:31841-31854, 2022. 2 +[12] Jiatao Gu, Alex Trevithick, Kai-En Lin, Joshua M Susskind, Christian Theobalt, Lingjie Liu, and Ravi Ramamoorthi. Nerfdiff: Single-image view synthesis with nef-guided distillation from 3d-aware diffusion. In International Conference on Machine Learning, pages 11808-11826. PMLR, 2023. 2 + +[13] Zehuan Huang, Hao Wen, Junting Dong, Yaohui Wang, Yangguang Li, Xinyuan Chen, Yan-Pei Cao, Ding Liang, Yu Qiao, Bo Dai, and Lu Sheng. Epidiff: Enhancing multi-view synthesis via localized epipolar-constrained diffusion, 2023. 2 +[14] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 867-876, 2022. 2 +[15] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2 +[16] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439, 2023. 5 +[17] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics (TOG), 39(6):1-14, 2020. 6 +[18] Oliver Laric. Three d scans. In threadscans.com, 2012. 6 +[19] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300–309, 2023. 1, 2 +[20] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Zexiang Xu, Hao Su, et al. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. arXiv preprint arXiv:2306.16928, 2023. +[21] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9298–9309, 2023. 2 +[22] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 2 +[23] Xiaoxiao Long, Cheng Lin, Peng Wang, Taku Komura, and Wenping Wang. Sparseneus: Fast generalizable neural surface reconstruction from sparse views. In European Conference on Computer Vision, pages 210-227. Springer, 2022. 2 +[24] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. arXiv preprint arXiv:2310.15008, 2023. 2 +[25] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12663–12673, 2023. 2, 6, 7 + +[26] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 2, 6, 7 +[27] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 conference papers, pages 1-8, 2022. 2 +[28] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2 +[29] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 1, 2 +[30] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, et al. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. arXiv preprint arXiv:2306.17843, 2023. 2 +[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2 +[32] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2 +[33] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. arXiv preprint arXiv:2302.01721, 2023. 2, 4, 5, 6, 7 +[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 3 +[35] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 2 +[36] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2 +[37] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 2, 5 +[38] Yawar Siddiqui, Justus Thies, Fangchang Ma, Qi Shan, Matthias Nießner, and Angela Dai. Texturify: Generating + +textures on 3d shape surfaces. In European Conference on Computer Vision, pages 72-88. Springer, 2022. 2 +[39] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3, 6 +[40] Jia-Mu Sun, Tong Wu, and Lin Gao. Recent advances in implicit representation-based 3d shape generation. Visual Intelligence, 2(1):9, 2024. 1 +[41] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2818-2826, 2016. 7 +[42] Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi. Viewset diffusion:(0-) image-conditioned 3d generative models from 2d data. arXiv preprint arXiv:2306.07881, 2023. 2 +[43] Shitao Tang, Fuyang Zhang, Jiacheng Chen, Peng Wang, and Yasutaka Furukawa. Mvdiffusion: Enabling holistic multiview image generation with correspondence-aware diffusion. arXiv preprint arXiv:2307.01097, 2023. +[44] Hung-Yu Tseng, Qinbo Li, Changil Kim, Suhib Alsisan, Jia-Bin Huang, and Johannes Kopf. Consistent view synthesis with pose-guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16773-16783, 2023. 2 +[45] Greg Turk and Marc Levoy. Zippered polygon meshes from range images. In Proceedings of the 21st annual conference on Computer graphics and interactive techniques, pages 311-318, 1994. 6 +[46] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12619-12629, 2023. 2 +[47] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213, 2023. 2 +[48] Jianfeng Xiang, Jiaolong Yang, Binbin Huang, and Xin Tong. 3d-aware image generation using 2d diffusion models. arXiv preprint arXiv:2303.17905, 2023. 2 +[49] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. arXiv preprint arXiv:2306.07954, 2023. 5 +[50] Jonathan Young. xatlas. In github.com/jpcy/xatlas, 2016. 3 +[51] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3 \ No newline at end of file diff --git a/2024/GenesisTex_ Adapting Image Denoising Diffusion to Texture Space/images.zip b/2024/GenesisTex_ Adapting Image Denoising Diffusion to Texture Space/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5ee3e8c9fb348cb332a3caab2097e631706c0689 --- /dev/null +++ b/2024/GenesisTex_ Adapting Image Denoising Diffusion to Texture Space/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:057ab16e81cba770f390012405b90c362291ec7e36fc57755bfbb98fb70df135 +size 538651 diff --git a/2024/GenesisTex_ Adapting Image Denoising Diffusion to Texture Space/layout.json b/2024/GenesisTex_ Adapting Image Denoising Diffusion to Texture Space/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0622adc9be5d8893743cf69ef147cd8e98ab2e79 --- /dev/null +++ b/2024/GenesisTex_ Adapting Image Denoising Diffusion to Texture Space/layout.json @@ -0,0 +1,10311 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 91, + 103, + 501, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 103, + 501, + 121 + ], + "spans": [ + { + "bbox": [ + 91, + 103, + 501, + 121 + ], + "type": "text", + "content": "GenesisTex: Adapting Image Denoising Diffusion to Texture Space" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 87, + 141, + 505, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 141, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 87, + 141, + 505, + 200 + ], + "type": "text", + "content": "Chenjian Gao" + }, + { + "bbox": [ + 87, + 141, + 505, + 200 + ], + "type": "inline_equation", + "content": "^{1,2\\dagger}" + }, + { + "bbox": [ + 87, + 141, + 505, + 200 + ], + "type": "text", + "content": " Boyan Jiang" + }, + { + "bbox": [ + 87, + 141, + 505, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 87, + 141, + 505, + 200 + ], + "type": "text", + "content": " Xinghui Li" + }, + { + "bbox": [ + 87, + 141, + 505, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 87, + 141, + 505, + 200 + ], + "type": "text", + "content": " Yingpeng Zhang" + }, + { + "bbox": [ + 87, + 141, + 505, + 200 + ], + "type": "inline_equation", + "content": "^{2\\#}" + }, + { + "bbox": [ + 87, + 141, + 505, + 200 + ], + "type": "text", + "content": " Qian Yu" + }, + { + "bbox": [ + 87, + 141, + 505, + 200 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 87, + 141, + 505, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 87, + 141, + 505, + 200 + ], + "type": "text", + "content": "School of Software, Beihang University \n" + }, + { + "bbox": [ + 87, + 141, + 505, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 87, + 141, + 505, + 200 + ], + "type": "text", + "content": "R&D Efficiency and Capability Department, Tencent IEG \nhttps://cjeen.github.io/GenesisTexPaper/" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 166, + 227, + 214, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 227, + 214, + 239 + ], + "spans": [ + { + "bbox": [ + 166, + 227, + 214, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 251, + 290, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 290, + 492 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 290, + 492 + ], + "type": "text", + "content": "We present GenesisTex, a novel method for synthesizing textures for 3D geometries from text descriptions. GenesisTex adapts the pretrained image diffusion model to texture space by texture space sampling. Specifically, we maintain a latent texture map for each viewpoint, which is updated with predicted noise on the rendering of the corresponding viewpoint. The sampled latent texture maps are then decoded into a final texture map. During the sampling process, we focus on both global and local consistency across multiple viewpoints: global consistency is achieved through the integration of style consistency mechanisms within the noise prediction network, and low-level consistency is achieved by dynamically aligning latent textures. Finally, we apply reference-based inpainting and img2img on denser views for texture refinement. Our approach overcomes the limitations of slow optimization in distillation-based methods and instability in inpainting-based methods. Experiments on meshes from various sources demonstrate that our method surpasses the baseline methods quantitatively and qualitatively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 513, + 128, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 513, + 128, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 513, + 128, + 525 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 533, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 628 + ], + "type": "text", + "content": "In recent years, with the development of deep learning, 3D content generation technology has made significant progress. The applications of 3D content generation are diverse, ranging from AR/VR to gaming and filmmaking. While there has been considerable research on deep learning-based geometric asset generation[40], there has been a notable industry demand for generating realistic textures for given geometries." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "text", + "content": "Recently, text-conditioned image diffusion models [34] have achieved impressive results in image generation. Some works [7, 19, 29] have leveraged text-conditioned image diffusion models to generate textured 3D assets by gener" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 308, + 226, + 544, + 422 + ], + "blocks": [ + { + "bbox": [ + 308, + 226, + 544, + 422 + ], + "lines": [ + { + "bbox": [ + 308, + 226, + 544, + 422 + ], + "spans": [ + { + "bbox": [ + 308, + 226, + 544, + 422 + ], + "type": "image", + "image_path": "375f4d1e1db8ab1323d0a75ed48f511ef8e3f009a1755fe6418f02c50d1e724f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 364, + 423, + 488, + 434 + ], + "lines": [ + { + "bbox": [ + 364, + 423, + 488, + 434 + ], + "spans": [ + { + "bbox": [ + 364, + 423, + 488, + 434 + ], + "type": "text", + "content": "red volkswagon beetle, cartoon style" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 443, + 547, + 510 + ], + "lines": [ + { + "bbox": [ + 304, + 443, + 547, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 443, + 547, + 510 + ], + "type": "text", + "content": "Figure 1. Texturing results of different methods. Score Distillation Sampling (SDS) based method produces blurred and oversaturated textures. Inpainting-based approach results in artifacts at the boundaries of inpainting masks. Texture space sampling concurrently generating content from multiple viewpoints, produces clean, clear and natural colored textures." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 533, + 546, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 546, + 675 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 546, + 675 + ], + "type": "text", + "content": "ating content from multiple viewpoints, achieving notable performance. In this work, our focus is on generating high-quality textures for a given geometry, leveraging the priors provided by pre-trained text-to-image diffusion models. This task poses several challenges, including: 1) View consistency: ensuring cross-view constraints for maintaining low-level consistency; 2) High efficiency: generating textures for a model within a few minutes, enabling practical applications; 3) Zero-shot learning: achieving texture generation without requiring additional training or finetuning. Addressing these challenges is crucial for successful application of image diffusion models to the texture domain." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "content": "Currently, the most prominent method for text-to-3D generation is Score Distillation Sampling (SDS) [7, 19, 29]. SDS utilizes the prior knowledge from image diffusion" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 683, + 235, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 683, + 235, + 693 + ], + "spans": [ + { + "bbox": [ + 61, + 683, + 235, + 693 + ], + "type": "inline_equation", + "content": "^{\\dagger}" + }, + { + "bbox": [ + 61, + 683, + 235, + 693 + ], + "type": "text", + "content": "Work was done during an internship at Tencent IEG." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 63, + 693, + 129, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 693, + 129, + 703 + ], + "spans": [ + { + "bbox": [ + 63, + 693, + 129, + 703 + ], + "type": "text", + "content": "Equal contribution." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 63, + 703, + 138, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 703, + 138, + 712 + ], + "spans": [ + { + "bbox": [ + 63, + 703, + 138, + 712 + ], + "type": "text", + "content": "* Corresponding author." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4620" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "text", + "content": "models to iteratively generate gradients for rendering images of 3D objects. However, SDS has limitations in terms of both efficiency and quality. It often takes several hours to generate a single textured 3D object, and the resulting textures can suffer from over-saturation due to the adoption of a large CFG (classifier-free guidance) scale, which reduces randomness during generation. Recent works [2, 6, 33] have introduced inpainting-based methods for multi-view generation. These methods establish a predetermined order of views, using the content of previously generated views as a condition for subsequent views. While inpainting-based approaches are faster and produce realistic colors, they are sensitive to the predefined view order. Additionally, they require texture segmentation to determine the source view for each texture pixel, which often results in artifacts around the borders of the segmented areas." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 266, + 289, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 289, + 566 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 289, + 566 + ], + "type": "text", + "content": "To address the aforementioned issues, we propose GenesisTex, a novel approach that introduces texture space sampling. Specifically, we maintain multiple-view latent texture maps throughout the sampling process and perform denoising to improve their quality. Subsequently, the content is decoded from the latent space to obtain RGB textures. Our approach focuses on achieving two aspects of consistency during the sampling process. Firstly, we ensure global style consistency across multiple views by incorporating style consistency in the noise prediction network. This helps maintain a coherent style throughout the generated textures. Secondly, we employ dynamic alignment of latent textures to ensure low-level multi-view consistency, enhancing the overall quality of the generated textures. Due to memory limitations, texture space sampling operates on sparse views. However, to further enhance the quality, we leverage reference-based inpainting and Img2Img techniques on denser views for texture map refinement. GenesisTex can generate detailed, clean, and naturally colored textures for a given geometry within a few minutes. Unlike inpainting-based methods, our proposed texture space sampling concurrently generates content for multiple views without relying on a predefined order. This adaptability makes GenesisTex suitable for various geometries and results in fewer artifacts in the generated textures." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 567, + 258, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 567, + 258, + 578 + ], + "spans": [ + { + "bbox": [ + 58, + 567, + 258, + 578 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 582, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 582, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 582, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 47, + 582, + 287, + 629 + ], + "type": "text", + "content": "- First, we present a novel method for texture generation, where the core is texture space sampling. This sampling technique allows for the concurrent denoising of latent textures associated with multiple viewpoints." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 630, + 287, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 630, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 47, + 630, + 287, + 666 + ], + "type": "text", + "content": "- Second, we introduce Style Consistency and Dynamic Alignment in texture space sampling for multi-view consistency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 666, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 666, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 666, + 287, + 712 + ], + "type": "text", + "content": "- Third, we conduct a comprehensive study involving numerous 3D objects from various sources. The experimental results demonstrate the superiority of our method over baseline methods." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 95, + 547, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 95, + 547, + 347 + ], + "spans": [ + { + "bbox": [ + 304, + 95, + 547, + 347 + ], + "type": "text", + "content": "Texture Synthesis. Generating textures over 3D surfaces is a challenging problem, as it requires attention to both colors and geometry. Earlier works like AUV-Net [8] and Texturify [38] embed the geometric prior into a UV map or mesh parameterization. Different from them, EG3D [3] and GET3D [11] directly train 3D StyleGANs [15] to generate geometries and textures jointly, where the textures are implicit texture fields. However, these methods either only work on a single category, or demand textured 3D shapes for training without text conditioning, which limits their broad applicability. Recently, [2, 6, 33] use the priors provided by the image diffusion model to synthesize textures. Text2Tex [6] and TEXTure [33] performs inpainting on multiview renderings. TexFusion [2] propose a sequential texture sampling method. Our method predicts multi-view noise concurrently, avoiding the following issues of sequential noise prediction: 1) Each iteration requires adding forward noise to visited regions to match unvisited areas, potentially leading to detail loss. 2) Consistency is only constrained between adjacent viewpoints, with no direct consistency for long-range viewpoints." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 350, + 547, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 350, + 547, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 547, + 495 + ], + "type": "text", + "content": "Text-to-Image Diffusion Models. Over the past years, the development of several large-scale diffusion models [28, 32, 34, 35] has enabled the production of highly detailed and visually impressive images. These models generate images based on input text prompts. Specifically, Stable Diffusion is trained on a substantial text-image dataset [36] and incorporates a text encoder from CLIP [31] to understand the input prompts. Beyond the basic text conditioning, ControlNet [51] enables the model to condition its denoising network on additional input modalities, such as depth maps. In this work, we utilize ControlNet and Stable Diffusion to provide geometrically-conditioned image priors." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 498, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 548, + 713 + ], + "type": "text", + "content": "Text-to-3D using 2D Image Diffusion Models. Early works [14, 26, 27] utilize the pretrained CLIP model to maximize the similarity between rendered images and text prompt. However, pioneering works DreamFusion [29] and SJC [46], on the other hand, propose to distill a 2D text-to-image generation model to generate 3D shapes from texts, and many follow-up works [19-21, 23, 25, 30, 47] follow such per-shape optimization scheme. Recently, several methods [4, 12, 13, 22, 24, 37, 42-44, 48] have been proposed to generate consistent multi-view images by using diffusion models. MVDream [37] and SyncDreamer [22] share similar ideas, generating consistent multi-view images via attention layers. However, existing approaches either endure slow optimization processes or depend on separately trained 3D priors, rendering them unsuitable for direct application in texture synthesis. In contrast, our method does not require additional training and can generate results within several minutes." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "4621" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 129, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 129, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 129, + 85 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 289, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 289, + 212 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 289, + 212 + ], + "type": "text", + "content": "Our objective is to generate a texture map " + }, + { + "bbox": [ + 46, + 91, + 289, + 212 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 46, + 91, + 289, + 212 + ], + "type": "text", + "content": " for a given 3D mesh " + }, + { + "bbox": [ + 46, + 91, + 289, + 212 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 91, + 289, + 212 + ], + "type": "text", + "content": ", using the provided text as a descriptive condition. Our approach leverages the Stable Diffusion model as the image diffusion model. We begin by introducing some basic concept about Stable Diffusion and texture representation in Sec. 3.1. Then, we introduce a sampling algorithm in texture space for Stable Diffusion in Sec. 3.2. Furthermore, we propose multi-view consistency strategy in Sec. 3.3. Finally we perform refinement on the texture map, which includes Inpainting and Img2Img, as described in Sec. 3.4." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 217, + 127, + 230 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 217, + 127, + 230 + ], + "spans": [ + { + "bbox": [ + 47, + 217, + 127, + 230 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 236, + 288, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 236, + 288, + 285 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 288, + 285 + ], + "type": "text", + "content": "Stable Diffusion. The diffusion model belongs to a class of generative models that generate data through iterative denoising from random noise. In the sampling process, " + }, + { + "bbox": [ + 46, + 236, + 288, + 285 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{i-1}" + }, + { + "bbox": [ + 46, + 236, + 288, + 285 + ], + "type": "text", + "content": " can be sampled from " + }, + { + "bbox": [ + 46, + 236, + 288, + 285 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i" + }, + { + "bbox": [ + 46, + 236, + 288, + 285 + ], + "type": "text", + "content": " using a DDIM sampler [39]:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 291, + 287, + 353 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 291, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 291, + 287, + 353 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\hat {\\mathbf {z}} _ {i \\rightarrow 0} = (\\mathbf {z} _ {i} - \\sqrt {1 - \\alpha_ {i}} \\epsilon_ {\\theta} (\\mathbf {z} _ {i}, t _ {i}, \\mathbf {c} _ {t e x t})) / \\sqrt {\\alpha_ {i}}, \\\\ \\mathbf {z} _ {i - 1} = \\sqrt {\\alpha_ {i - 1}} \\hat {\\mathbf {z}} _ {i \\rightarrow 0} + \\sqrt {1 - \\alpha_ {i - 1} - \\sigma_ {i} ^ {2}} \\epsilon_ {\\theta} (\\mathbf {z} _ {i}, t _ {i}, \\mathbf {c} _ {t e x t}) \\\\ + \\sigma_ {i} \\varepsilon , \\tag {1} \\\\ \\end{array}", + "image_path": "917ff3fae615e952c4f07fe2e03e92348a39d9870228f6900ac55c0d47652081.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 354, + 287, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 287, + 436 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 287, + 436 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 354, + 287, + 436 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 46, + 354, + 287, + 436 + ], + "type": "text", + "content": " is a pretrained U-Net for noise prediction and the initial " + }, + { + "bbox": [ + 46, + 354, + 287, + 436 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_T \\sim \\mathcal{N}(0, \\mathbf{I})" + }, + { + "bbox": [ + 46, + 354, + 287, + 436 + ], + "type": "text", + "content": ". In this paper, we specifically employ Stable Diffusion [34]. Stable Diffusion performs denoising in the latent space and employs an autoencoder " + }, + { + "bbox": [ + 46, + 354, + 287, + 436 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(\\mathcal{E}(\\cdot))" + }, + { + "bbox": [ + 46, + 354, + 287, + 436 + ], + "type": "text", + "content": " for the conversion between image and latent representations. So " + }, + { + "bbox": [ + 46, + 354, + 287, + 436 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + }, + { + "bbox": [ + 46, + 354, + 287, + 436 + ], + "type": "text", + "content": " generated by the diffusion process is decoded to the image " + }, + { + "bbox": [ + 46, + 354, + 287, + 436 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(\\mathbf{z}_0)" + }, + { + "bbox": [ + 46, + 354, + 287, + 436 + ], + "type": "text", + "content": " finally." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 437, + 287, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 496 + ], + "type": "text", + "content": "ControlNet. ControlNet [51] injects low-level control during the denoising process of Stable Diffusion. Our approach employs depth " + }, + { + "bbox": [ + 46, + 437, + 287, + 496 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 46, + 437, + 287, + 496 + ], + "type": "text", + "content": " as the geometric control condition. The predicted noise of the U-Net with ControlNet is represented as " + }, + { + "bbox": [ + 46, + 437, + 287, + 496 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{z}_i, t_i, \\mathbf{c}_{text}, \\mathbf{d})" + }, + { + "bbox": [ + 46, + 437, + 287, + 496 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 497, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 497, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 497, + 287, + 544 + ], + "type": "text", + "content": "Texture Representation. In this paper, we utilize mesh as the 3D representation. Texture map is associated with an UV parameterization of the mesh " + }, + { + "bbox": [ + 46, + 497, + 287, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 497, + 287, + 544 + ], + "type": "text", + "content": ". We use " + }, + { + "bbox": [ + 46, + 497, + 287, + 544 + ], + "type": "inline_equation", + "content": "x\\text{atlas}" + }, + { + "bbox": [ + 46, + 497, + 287, + 544 + ], + "type": "text", + "content": " [50] to generate the UV parameterization." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "type": "text", + "content": "**Rendering.** Given a mesh " + }, + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "type": "text", + "content": ", a texture map " + }, + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "type": "text", + "content": " and a viewpoint " + }, + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "type": "text", + "content": ", we can use the rendering function " + }, + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "type": "text", + "content": " to get the rendered image " + }, + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{(img)} = \\mathcal{R}(\\mathcal{T};\\mathcal{M},C)" + }, + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "type": "text", + "content": ". Similar to [2], we do not consider any lighting and rendering involves solely sampling colors from the texture map. The inverse rendering function " + }, + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{R}^{-1}" + }, + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "type": "text", + "content": " can inverse render the image " + }, + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{(img)}" + }, + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "type": "text", + "content": " back to a texture map " + }, + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{T}' = \\mathcal{R}^{-1}(\\mathbf{x}^{(img)};\\mathcal{M},C)" + }, + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 635, + 195, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 195, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 195, + 647 + ], + "type": "text", + "content": "3.2. Sampling in Texture Space" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "content": "The most straightforward method for sampling on the texture map is to fine-tune the Stable Diffusion model directly on some texture maps, making its distribution more like the distribution of texture maps. However, the distribution of texture maps is much more complex than that of rendering" + } + ] + } + ], + "index": 10 + }, + { + "type": "code", + "bbox": [ + 306, + 85, + 566, + 436 + ], + "blocks": [ + { + "bbox": [ + 307, + 72, + 462, + 84 + ], + "lines": [ + { + "bbox": [ + 307, + 72, + 462, + 84 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 462, + 84 + ], + "type": "text", + "content": "Algorithm 1 Texture Space Sampling" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 306, + 85, + 566, + 436 + ], + "lines": [ + { + "bbox": [ + 306, + 85, + 566, + 436 + ], + "spans": [ + { + "bbox": [ + 306, + 85, + 566, + 436 + ], + "type": "text", + "content": "Input: mesh " + }, + { + "bbox": [ + 306, + 85, + 566, + 436 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 306, + 85, + 566, + 436 + ], + "type": "text", + "content": " ,cameras " + }, + { + "bbox": [ + 306, + 85, + 566, + 436 + ], + "type": "inline_equation", + "content": "\\{C_1,\\dots ,C_N\\}" + }, + { + "bbox": [ + 306, + 85, + 566, + 436 + ], + "type": "text", + "content": " \nParameters:Denoising time schedule " + }, + { + "bbox": [ + 306, + 85, + 566, + 436 + ], + "type": "inline_equation", + "content": "\\{t_i\\}_{i = T}^0" + }, + { + "bbox": [ + 306, + 85, + 566, + 436 + ], + "type": "text", + "content": " ,DDIM \nnoise schedule " + }, + { + "bbox": [ + 306, + 85, + 566, + 436 + ], + "type": "inline_equation", + "content": "\\{\\sigma_i\\}_{i = T}^0" + }, + { + "bbox": [ + 306, + 85, + 566, + 436 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{z}_{T,n}^{(tex)}\\}_{n = 1}^N\\gets \\{\\mathbf{0}\\}" + }, + { + "bbox": [ + 306, + 85, + 566, + 436 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{z}_n^{(bg)}\\}_{n = 1}^N\\sim \\{\\mathcal{N}(\\mathbf{0},\\mathbf{I})\\}" + }, + { + "bbox": [ + 306, + 85, + 566, + 436 + ], + "type": "inline_equation", + "content": "\\{\\hat{\\epsilon}_{T + 1,n}\\}_{n = 1}^N\\sim \\{\\mathcal{N}(\\mathbf{0},\\mathbf{I})\\}" + }, + { + "bbox": [ + 306, + 85, + 566, + 436 + ], + "type": "text", + "content": " \ndenoising stage: \nfor " + }, + { + "bbox": [ + 306, + 85, + 566, + 436 + ], + "type": "inline_equation", + "content": "i\\in \\{T\\dots 1\\}" + }, + { + "bbox": [ + 306, + 85, + 566, + 436 + ], + "type": "text", + "content": " do \nfor " + }, + { + "bbox": [ + 306, + 85, + 566, + 436 + ], + "type": "inline_equation", + "content": "n\\in \\{1\\dots N\\}" + }, + { + "bbox": [ + 306, + 85, + 566, + 436 + ], + "type": "text", + "content": " do \n" + }, + { + "bbox": [ + 306, + 85, + 566, + 436 + ], + "type": "inline_equation", + "content": "\\varepsilon_{n}\\sim \\mathcal{N}(0,I)" + }, + { + "bbox": [ + 306, + 85, + 566, + 436 + ], + "type": "inline_equation", + "content": "{\\bf z}_{i,n}^{(img)}\\gets {\\bf M}_n(\\sqrt{\\alpha_i}{\\cal R}({\\bf z}_{i,n}^{(tex)}) + \\sqrt{1 - \\alpha_i - \\sigma_{i + 1}^2}\\hat{\\epsilon}_{i + 1,n})" + }, + { + "bbox": [ + 306, + 85, + 566, + 436 + ], + "type": "inline_equation", + "content": "+(1 - \\mathbf{M}_n)\\mathbf{z}_n^{(bg)} + \\sigma_{i + 1}\\varepsilon_n" + }, + { + "bbox": [ + 306, + 85, + 566, + 436 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_{i,n}\\gets \\epsilon_{\\theta}(\\mathbf{z}_{i,n}^{(img)},t_i,\\mathbf{c},\\mathbf{d}_n)" + }, + { + "bbox": [ + 306, + 85, + 566, + 436 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_{0,n}^{(img)}\\gets (\\mathbf{z}_{i,n}^{(img)} - \\sqrt{1 - \\alpha_i}\\hat{\\epsilon}_{i,n}) / \\sqrt{\\alpha_i}" + }, + { + "bbox": [ + 306, + 85, + 566, + 436 + ], + "type": "inline_equation", + "content": "{\\bf z}_{i - 1,n}^{(tex)}\\gets \\mathcal{R}^{-1}(\\hat{\\mathbf{z}}_{0,n}^{(img)})" + }, + { + "bbox": [ + 306, + 85, + 566, + 436 + ], + "type": "inline_equation", + "content": "{\\bf z}_{n}^{(bg)}\\gets \\sqrt{\\alpha_{i - 1}}\\hat{\\mathbf{z}}_{0,n}^{(img)} + \\sqrt{1 - \\alpha_{i - 1} - \\sigma_i^2}\\cdot \\hat{\\epsilon}_{i,n}" + }, + { + "bbox": [ + 306, + 85, + 566, + 436 + ], + "type": "text", + "content": " \nend for \n" + }, + { + "bbox": [ + 306, + 85, + 566, + 436 + ], + "type": "inline_equation", + "content": "\\{{\\bf z}_{i - 1,n}^{(tex)}\\}_{n = 1}^N = dynamic\\_ align(\\{{\\bf z}_{i - 1,n}^{(tex)}\\}_{n = 1}^N)" + }, + { + "bbox": [ + 306, + 85, + 566, + 436 + ], + "type": "text", + "content": " \nend for \ndecoding stage: \nfor " + }, + { + "bbox": [ + 306, + 85, + 566, + 436 + ], + "type": "inline_equation", + "content": "n\\in \\{1\\dots N\\}" + }, + { + "bbox": [ + 306, + 85, + 566, + 436 + ], + "type": "text", + "content": " do \n" + }, + { + "bbox": [ + 306, + 85, + 566, + 436 + ], + "type": "inline_equation", + "content": "{\\bf x}_n^{(img)} = \\mathcal{D}(\\hat{\\bf z}_{0,n}^{(img)})" + }, + { + "bbox": [ + 306, + 85, + 566, + 436 + ], + "type": "inline_equation", + "content": "{\\bf x}_n^{(tex)} = R^{-1}({\\bf x}_n^{(img)})" + }, + { + "bbox": [ + 306, + 85, + 566, + 436 + ], + "type": "text", + "content": " \nend for \nT=merge({x(tex)})N \nreturn Texture map Tsampled" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "code_body" + } + ], + "index": 12, + "sub_type": "algorithm" + }, + { + "bbox": [ + 304, + 458, + 545, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 458, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 458, + 545, + 506 + ], + "type": "text", + "content": "images since different UV parameterizations of the same 3D object correspond to different texture maps. Instead of fine-tuning existing image diffusion models, we adapt them to the texture space for texture map generation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 506, + 546, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 546, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 546, + 628 + ], + "type": "text", + "content": "Algorithm 1 presents our proposed texture space sampling algorithm. Firstly we define a set of " + }, + { + "bbox": [ + 304, + 506, + 546, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{(sampling)} = \\{C_1,..,C_N\\}" + }, + { + "bbox": [ + 304, + 506, + 546, + 628 + ], + "type": "text", + "content": " camera views and render the corresponding depth map " + }, + { + "bbox": [ + 304, + 506, + 546, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_n" + }, + { + "bbox": [ + 304, + 506, + 546, + 628 + ], + "type": "text", + "content": " in each view. Since we employ a latent diffusion model, our texture space sampling method first performs denoising in the latent texture space and then decode the latent to texture image. In the denoising stage, a dedicated diffusion denoising algorithm is applied to a set of latent texture maps " + }, + { + "bbox": [ + 304, + 506, + 546, + 628 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{z}_n^{(tex)}\\}_{n=1}^N" + }, + { + "bbox": [ + 304, + 506, + 546, + 628 + ], + "type": "text", + "content": ". In the decoding stage, we recover the texture map " + }, + { + "bbox": [ + 304, + 506, + 546, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{sampled}" + }, + { + "bbox": [ + 304, + 506, + 546, + 628 + ], + "type": "text", + "content": " from the latent space." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 628, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 628, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 628, + 547, + 715 + ], + "type": "text", + "content": "Denoising Stage. In the denoising stage, we adapt the DDIM from the rendering space to the texture space and utilize the same parameters as DDIM, including denoising time schedule " + }, + { + "bbox": [ + 304, + 628, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\{t_i\\}_{i = T}^0" + }, + { + "bbox": [ + 304, + 628, + 547, + 715 + ], + "type": "text", + "content": " , DDIM noise schedule " + }, + { + "bbox": [ + 304, + 628, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\{\\sigma_i\\}_{i = T}^0" + }, + { + "bbox": [ + 304, + 628, + 547, + 715 + ], + "type": "text", + "content": " . For each viewpoint, we maintain a separate latent texture which dynamically interact during the denoising process. We begin with " + }, + { + "bbox": [ + 304, + 628, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{z}_{T,n}^{(tex)}\\}_{n = 1}^{N} = \\{\\mathbf{0}\\}" + }, + { + "bbox": [ + 304, + 628, + 547, + 715 + ], + "type": "text", + "content": " , the zero latent texture map" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4622" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 239, + 223 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 239, + 223 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 239, + 223 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 239, + 223 + ], + "type": "image", + "image_path": "9b1d8a67dfd9b827f94980ff551b3ef6522b236bf388670b26a3d08c2d688ffe.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 236, + 547, + 270 + ], + "lines": [ + { + "bbox": [ + 46, + 236, + 547, + 270 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 547, + 270 + ], + "type": "text", + "content": "Figure 2. Overview of GenesisTex. GenesisTex generates a texture map for a given mesh " + }, + { + "bbox": [ + 46, + 236, + 547, + 270 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 236, + 547, + 270 + ], + "type": "text", + "content": ", based on a prompt. Texture Space Sampling samples a texture map using Stable Diffusion, introducing style consistency and dynamic alignment across multiple viewpoints. Furthermore, Inpainting and Img2Img are applied to fill in the blank regions and enhance the quality of texture map details, respectively." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 240, + 71, + 370, + 223 + ], + "blocks": [ + { + "bbox": [ + 240, + 71, + 370, + 223 + ], + "lines": [ + { + "bbox": [ + 240, + 71, + 370, + 223 + ], + "spans": [ + { + "bbox": [ + 240, + 71, + 370, + 223 + ], + "type": "image", + "image_path": "f658c0f6c5646555c3e8d6eb1749aab590aa4831ff414b7bc70c790dfa9b67ce.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 373, + 71, + 545, + 158 + ], + "blocks": [ + { + "bbox": [ + 373, + 71, + 545, + 158 + ], + "lines": [ + { + "bbox": [ + 373, + 71, + 545, + 158 + ], + "spans": [ + { + "bbox": [ + 373, + 71, + 545, + 158 + ], + "type": "image", + "image_path": "412af3afeb20fccbaee64f218c7e5ed5a7547a1fa13126e6915321f0c9f92dd5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 373, + 160, + 545, + 223 + ], + "blocks": [ + { + "bbox": [ + 373, + 160, + 545, + 223 + ], + "lines": [ + { + "bbox": [ + 373, + 160, + 545, + 223 + ], + "spans": [ + { + "bbox": [ + 373, + 160, + 545, + 223 + ], + "type": "image", + "image_path": "63dcb9d3bf64b3fd97c2ec66455bc3322d028e86115878d631fedd9975f57a3a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 289, + 287, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 287, + 359 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 287, + 359 + ], + "type": "text", + "content": "corresponding to " + }, + { + "bbox": [ + 46, + 289, + 287, + 359 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 289, + 287, + 359 + ], + "type": "text", + "content": " viewpoints. At each denoising step, our goal is to predict " + }, + { + "bbox": [ + 46, + 289, + 287, + 359 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{i - 1,n}^{(tex)}" + }, + { + "bbox": [ + 46, + 289, + 287, + 359 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 289, + 287, + 359 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{i,n}^{(tex)}" + }, + { + "bbox": [ + 46, + 289, + 287, + 359 + ], + "type": "text", + "content": ". We first render the latent texture map " + }, + { + "bbox": [ + 46, + 289, + 287, + 359 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{i,n}^{(tex)}" + }, + { + "bbox": [ + 46, + 289, + 287, + 359 + ], + "type": "text", + "content": " to the rendering space using the rendering function " + }, + { + "bbox": [ + 46, + 289, + 287, + 359 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 46, + 289, + 287, + 359 + ], + "type": "text", + "content": ", and then calculate the corresponding latent image " + }, + { + "bbox": [ + 46, + 289, + 287, + 359 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{i,n}^{(img)}" + }, + { + "bbox": [ + 46, + 289, + 287, + 359 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 51, + 369, + 287, + 415 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 369, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 51, + 369, + 287, + 415 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {z} _ {i, n} ^ {(i m g)} = \\mathbf {M} _ {n} (\\sqrt {\\alpha_ {i}} \\mathcal {R} (\\mathbf {z} _ {i, n} ^ {(t e x)}) + \\sqrt {1 - \\alpha_ {i} - \\sigma_ {i + 1} ^ {2}} \\hat {\\epsilon} _ {i + 1, n}) \\\\ + \\left(1 - \\mathbf {M} _ {n}\\right) \\mathbf {z} _ {n} ^ {(b g)} + \\sigma_ {i + 1} \\varepsilon_ {n} \\tag {2} \\\\ \\end{array}", + "image_path": "5ecdb123999f6a8650ff915d7f894fd87e3cf659e683aeca9fee6fdf3575f353.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 415, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 287, + 567 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 46, + 415, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_n" + }, + { + "bbox": [ + 46, + 415, + 287, + 567 + ], + "type": "text", + "content": " represents the foreground mask for viewpoint " + }, + { + "bbox": [ + 46, + 415, + 287, + 567 + ], + "type": "inline_equation", + "content": "C_n" + }, + { + "bbox": [ + 46, + 415, + 287, + 567 + ], + "type": "text", + "content": " (downsampled to the same resolution as " + }, + { + "bbox": [ + 46, + 415, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{(img)})" + }, + { + "bbox": [ + 46, + 415, + 287, + 567 + ], + "type": "text", + "content": ", where foreground pixels have a value of 1, and background pixels have a value of 0. " + }, + { + "bbox": [ + 46, + 415, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_{i + 1,n}" + }, + { + "bbox": [ + 46, + 415, + 287, + 567 + ], + "type": "text", + "content": " represents the noise predicted by the noise prediction network at step " + }, + { + "bbox": [ + 46, + 415, + 287, + 567 + ], + "type": "inline_equation", + "content": "i + 1\\rightarrow i" + }, + { + "bbox": [ + 46, + 415, + 287, + 567 + ], + "type": "text", + "content": " (notably, we define " + }, + { + "bbox": [ + 46, + 415, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_{T + 1,n}\\sim \\mathcal{N}(0,\\mathbf{I})" + }, + { + "bbox": [ + 46, + 415, + 287, + 567 + ], + "type": "text", + "content": "). " + }, + { + "bbox": [ + 46, + 415, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_n^{(bg)}" + }, + { + "bbox": [ + 46, + 415, + 287, + 567 + ], + "type": "text", + "content": " is the background of the latent image, which is set to Gaussian noise at initialization. " + }, + { + "bbox": [ + 46, + 415, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\varepsilon_{n}\\sim \\mathcal{N}(0,\\mathbf{I})" + }, + { + "bbox": [ + 46, + 415, + 287, + 567 + ], + "type": "text", + "content": " is a random Gaussian noise. Subsequently, we predict the corresponding noise " + }, + { + "bbox": [ + 46, + 415, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_{i,n} = \\epsilon_{\\theta}(\\mathbf{z}_{i,n}^{(img)},t_i,\\mathbf{c},\\mathbf{d}_n)" + }, + { + "bbox": [ + 46, + 415, + 287, + 567 + ], + "type": "text", + "content": " for the latent image " + }, + { + "bbox": [ + 46, + 415, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{i,n}^{(img)}" + }, + { + "bbox": [ + 46, + 415, + 287, + 567 + ], + "type": "text", + "content": " conditioned on the depth map " + }, + { + "bbox": [ + 46, + 415, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_n" + }, + { + "bbox": [ + 46, + 415, + 287, + 567 + ], + "type": "text", + "content": ". With the predicted noise, we can obtain the predicted " + }, + { + "bbox": [ + 46, + 415, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_{0,n}^{(img)}" + }, + { + "bbox": [ + 46, + 415, + 287, + 567 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 84, + 575, + 287, + 593 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 575, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 84, + 575, + 287, + 593 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {z}} _ {0, n} ^ {(i m g)} = (\\mathbf {z} _ {i, n} ^ {(i m g)} - \\sqrt {1 - \\alpha_ {i}} \\hat {\\epsilon} _ {i, n}) / \\sqrt {\\alpha_ {i}}, \\qquad (3)", + "image_path": "4e17ebc9c831ed19fcedc0c7466ebc7682cbbcf052c7647dc32cddfd3004d38a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 603, + 287, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 603, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 47, + 603, + 287, + 633 + ], + "type": "text", + "content": "Finally, we inverse render " + }, + { + "bbox": [ + 47, + 603, + 287, + 633 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_{0,n}^{(img)}" + }, + { + "bbox": [ + 47, + 603, + 287, + 633 + ], + "type": "text", + "content": " to the texture space using " + }, + { + "bbox": [ + 47, + 603, + 287, + 633 + ], + "type": "inline_equation", + "content": "\\mathcal{R}^{-1}" + }, + { + "bbox": [ + 47, + 603, + 287, + 633 + ], + "type": "text", + "content": " and update " + }, + { + "bbox": [ + 47, + 603, + 287, + 633 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_n^{(bg)}" + }, + { + "bbox": [ + 47, + 603, + 287, + 633 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 123, + 641, + 287, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 641, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 123, + 641, + 287, + 666 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {i - 1, n} ^ {(t e x)} = \\mathcal {R} ^ {- 1} \\left(\\hat {\\mathbf {z}} _ {0, n} ^ {(i m g)}\\right), \\tag {4}", + "image_path": "b8c87bdeac9a132fa4ca5421cb10b86f305cd65fdd0585b815541f6626f2b9b3.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 65, + 658, + 258, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 658, + 258, + 681 + ], + "spans": [ + { + "bbox": [ + 65, + 658, + 258, + 681 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {n} ^ {(b g)} = \\sqrt {\\alpha_ {i - 1}} \\hat {\\mathbf {z}} _ {0, n} ^ {(i m g)} + \\sqrt {1 - \\alpha_ {i - 1} - \\sigma_ {i} ^ {2}} \\cdot \\hat {\\epsilon} _ {i, n}", + "image_path": "e4705445432381db5e83bd879583abbd036d56916b5362566f043c6ac509e63c.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "In the denoising stage, we repeat the above process " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "T - 1" + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": " times. After each denoising step, we impose consistency" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 289, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 545, + 361 + ], + "type": "text", + "content": "constraints on latent texture " + }, + { + "bbox": [ + 304, + 289, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{i - 1,n}^{(tex)}" + }, + { + "bbox": [ + 304, + 289, + 545, + 361 + ], + "type": "text", + "content": ", which will be introduced in Sec. 3.3. Notably, differing from DDIM where the latent includes noise, the latent texture in our method is an estimate of the noise-free version. This is because " + }, + { + "bbox": [ + 304, + 289, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 304, + 289, + 545, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 289, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{R}^{-1}" + }, + { + "bbox": [ + 304, + 289, + 545, + 361 + ], + "type": "text", + "content": " in the denoising process require interpolation of the latent, which may distort the noise." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 361, + 545, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 545, + 469 + ], + "type": "text", + "content": "Decoding Stage. To recover the texture map " + }, + { + "bbox": [ + 304, + 361, + 545, + 469 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\text{sampled}}" + }, + { + "bbox": [ + 304, + 361, + 545, + 469 + ], + "type": "text", + "content": ", we first decode each viewpoint's latent separately to obtain multi-view rendering. Subsequently, we inverse render the renderings into the texture space. After obtaining multiview texture maps, we merge them into a single texture map as the output. Similar to [2, 6, 33], we aim for each pixel on each texture map to come from the viewpoint where the corresponding point on the mesh is observed most directly. Therefore, our merge is defined as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 332, + 479, + 545, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 479, + 545, + 512 + ], + "spans": [ + { + "bbox": [ + 332, + 479, + 545, + 512 + ], + "type": "interline_equation", + "content": "\\mathcal {T} _ {\\text {s a m p l e d}} = \\sum_ {n = 1} ^ {N} \\operatorname {S o f t m a x} \\left(\\mathcal {R} ^ {- 1} \\left(\\mathbf {N} _ {\\mathbf {n}}\\right)\\right) \\times \\mathbf {x} _ {n} ^ {(t e x)} \\tag {5}", + "image_path": "cf46cf52ee29535f9bda2ba4bf80e00677d041cbd179b42d12a5f8b9fe910cfd.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 521, + 545, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{N}_{\\mathbf{n}}" + }, + { + "bbox": [ + 304, + 521, + 545, + 582 + ], + "type": "text", + "content": " is the similarity mask at viewpoint " + }, + { + "bbox": [ + 304, + 521, + 545, + 582 + ], + "type": "inline_equation", + "content": "C_n" + }, + { + "bbox": [ + 304, + 521, + 545, + 582 + ], + "type": "text", + "content": ", where each pixel represents the cosine similarity between the normal vectors of the visible faces and the reversed view direction. We use Softmax instead of Max because Softmax can result in more natural transition in the texture maps." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 588, + 527, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 588, + 527, + 601 + ], + "spans": [ + { + "bbox": [ + 306, + 588, + 527, + 601 + ], + "type": "text", + "content": "3.3. Consistency between Latent Texture Maps" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 607, + 424, + 619 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 607, + 424, + 619 + ], + "spans": [ + { + "bbox": [ + 306, + 607, + 424, + 619 + ], + "type": "text", + "content": "3.3.1 Dynamic Alignment" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 626, + 545, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 626, + 545, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 626, + 545, + 674 + ], + "type": "text", + "content": "During sampling in texture space, we maintain " + }, + { + "bbox": [ + 304, + 626, + 545, + 674 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 626, + 545, + 674 + ], + "type": "text", + "content": " latent texture maps, now we introduce an alignment approach to ensure their local consistency. Firstly, we directly reduce these latent texture maps to obtain a uniform latent texture map:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 331, + 683, + 545, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 683, + 545, + 716 + ], + "spans": [ + { + "bbox": [ + 331, + 683, + 545, + 716 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {i - 1} ^ {(u n i t e x)} = \\sum_ {n = 1} ^ {N} \\operatorname {S o f t m a x} \\left(\\mathcal {R} ^ {- 1} \\left(\\mathbf {N} _ {n}\\right)\\right) \\times \\mathbf {z} _ {i - 1, n} ^ {(t e x)} \\tag {6}", + "image_path": "e8de0b2be48ff0efbcf3469caac7ec5cf82587188c6ffb0f453e1642a30b2724.jpg" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4623" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 70, + 289, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 70, + 289, + 209 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 289, + 209 + ], + "type": "text", + "content": "Next, we blend each viewpoint's latent texture map " + }, + { + "bbox": [ + 47, + 70, + 289, + 209 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{i-1,n}^{(tex)}" + }, + { + "bbox": [ + 47, + 70, + 289, + 209 + ], + "type": "text", + "content": " with the uniform latent texture map " + }, + { + "bbox": [ + 47, + 70, + 289, + 209 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{i-1}^{(unitex)}" + }, + { + "bbox": [ + 47, + 70, + 289, + 209 + ], + "type": "text", + "content": ". We do not enforce strict alignment among the " + }, + { + "bbox": [ + 47, + 70, + 289, + 209 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 70, + 289, + 209 + ], + "type": "text", + "content": " latent texture maps at every step. Since a latent pixel gains full meaning only when combined with the surrounding context, which is viewpoint-dependent, the latent texture map from each viewpoint should contain some independent information. Moreover, the content generated at different timestamps varies, so the alignment constraints should adapt flexibly. We introduce a signal " + }, + { + "bbox": [ + 47, + 70, + 289, + 209 + ], + "type": "inline_equation", + "content": "c(t)" + }, + { + "bbox": [ + 47, + 70, + 289, + 209 + ], + "type": "text", + "content": " for dynamically regulating the variations in consistency during the sampling process:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 52, + 217, + 287, + 235 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 217, + 287, + 235 + ], + "spans": [ + { + "bbox": [ + 52, + 217, + 287, + 235 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {i - 1, n} ^ {(c o n s i s t e x)} = c (t _ {i}) \\times \\mathbf {z} _ {i - 1} ^ {(u n i t e x)} + (1 - c (t _ {i})) \\times \\mathbf {z} _ {i - 1, n} ^ {(t e x)} (7)", + "image_path": "081492797f05fab3800af37804d9bd1e36fbebc205239e7e82262465c0a271d5.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 242, + 287, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 242, + 287, + 291 + ], + "spans": [ + { + "bbox": [ + 47, + 242, + 287, + 291 + ], + "type": "text", + "content": "We empirically find that maintaining strong local consistency in the mid-term of denoising, while keeping weak local consistency in the early and late stages, leads to higher quality generation results." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 305, + 155, + 318 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 305, + 155, + 318 + ], + "spans": [ + { + "bbox": [ + 47, + 305, + 155, + 318 + ], + "type": "text", + "content": "3.3.2 Style Consistency" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 325, + 287, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 325, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 325, + 287, + 491 + ], + "type": "text", + "content": "In texture space sampling, we use the noise estimation network " + }, + { + "bbox": [ + 47, + 325, + 287, + 491 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 47, + 325, + 287, + 491 + ], + "type": "text", + "content": " to estimate noise for the multi-view latents, which is responsible for content generation. So we adapt " + }, + { + "bbox": [ + 47, + 325, + 287, + 491 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 47, + 325, + 287, + 491 + ], + "type": "text", + "content": " to ensure style consistency of estimated noise across different views. Inspired by video diffusion models [16, 49] and multi-view diffusion model [37], we modify all the self-attention layers and group normalization layers in the noise prediction network to align the style of multi-view content. Adapted Self-Attention. In Stable Diffusion, self-attention facilitates long-range interactions among features within an image. We transform self-attention into cross-view attention to establish style consistency across multiple viewpoints, by using the key " + }, + { + "bbox": [ + 47, + 325, + 287, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{K}^{\\prime}" + }, + { + "bbox": [ + 47, + 325, + 287, + 491 + ], + "type": "text", + "content": " and value " + }, + { + "bbox": [ + 47, + 325, + 287, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^{\\prime}" + }, + { + "bbox": [ + 47, + 325, + 287, + 491 + ], + "type": "text", + "content": " from all viewpoints," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 492, + 287, + 513 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 492, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 106, + 492, + 287, + 513 + ], + "type": "interline_equation", + "content": "\\mathbf {K} ^ {\\prime} = \\mathbf {W} ^ {K} \\left[ \\mathbf {z} _ {i, 1} ^ {(i m g)}; \\dots ; \\mathbf {z} _ {i, n} ^ {(i m g)} \\right], \\tag {8}", + "image_path": "68c5247adbc449264bacd887c1894bcfc1b228f48bb8f2e2af5f6cb56c9d1953.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 107, + 512, + 227, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 512, + 227, + 526 + ], + "spans": [ + { + "bbox": [ + 107, + 512, + 227, + 526 + ], + "type": "interline_equation", + "content": "\\mathbf {V} ^ {\\prime} = \\mathbf {W} ^ {V} [ \\mathbf {z} _ {i, 1} ^ {(i m g)}; \\dots ; \\mathbf {z} _ {i, n} ^ {(i m g)} ].", + "image_path": "a017d3f63a3f231bc0fc02ca9b0f089b06b046cefeb0af0c0b2d235454d1cfc3.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 532, + 222, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 532, + 222, + 545 + ], + "spans": [ + { + "bbox": [ + 47, + 532, + 222, + 545 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 532, + 222, + 545 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^K" + }, + { + "bbox": [ + 47, + 532, + 222, + 545 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 532, + 222, + 545 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^V" + }, + { + "bbox": [ + 47, + 532, + 222, + 545 + ], + "type": "text", + "content": " are pretrained parameters." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 545, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 641 + ], + "type": "text", + "content": "Adapted Group Normalization. Stable diffusion adopt group normalization as the normalization layer. Group normalization divides the channels into groups and computes within each group the mean and variance for normalization. We convert 2D group normalization to 3D by connecting all different views in the group normalization layer, i.e., the mean and variance are calculated within the channel group of all pixels across all viewpoints." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 642, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 642, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 642, + 287, + 714 + ], + "type": "text", + "content": "By adjusting self-attention layer and group normalization layer, we achieve multi-view style consistency without any additional training. In addition to the noise prediction network " + }, + { + "bbox": [ + 47, + 642, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 47, + 642, + 287, + 714 + ], + "type": "text", + "content": ", we apply the same modifications to the decoder " + }, + { + "bbox": [ + 47, + 642, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 47, + 642, + 287, + 714 + ], + "type": "text", + "content": " to further ensure style consistency in the multi-view decoded RGB images." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 72, + 446, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 446, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 446, + 84 + ], + "type": "text", + "content": "3.4. Texture Map Refinement" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 91, + 457, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 91, + 457, + 104 + ], + "spans": [ + { + "bbox": [ + 306, + 91, + 457, + 104 + ], + "type": "text", + "content": "3.4.1 Post-processing: Inpainting" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 110, + 547, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 110, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 304, + 110, + 547, + 194 + ], + "type": "text", + "content": "Since the memory cost of texture space sampling increases with the number of viewpoints of " + }, + { + "bbox": [ + 304, + 110, + 547, + 194 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{\\text{sampling}}" + }, + { + "bbox": [ + 304, + 110, + 547, + 194 + ], + "type": "text", + "content": ", the number of viewpoints " + }, + { + "bbox": [ + 304, + 110, + 547, + 194 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 110, + 547, + 194 + ], + "type": "text", + "content": " is limited. Some regions in the texture map " + }, + { + "bbox": [ + 304, + 110, + 547, + 194 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\text{sampled}}" + }, + { + "bbox": [ + 304, + 110, + 547, + 194 + ], + "type": "text", + "content": " may not be observed in any of the viewpoints in " + }, + { + "bbox": [ + 304, + 110, + 547, + 194 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{\\text{sampling}}" + }, + { + "bbox": [ + 304, + 110, + 547, + 194 + ], + "type": "text", + "content": ". Therefore, after texture space sampling, we introduce an inpainting epoch to fill texture in areas not observed in " + }, + { + "bbox": [ + 304, + 110, + 547, + 194 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 304, + 110, + 547, + 194 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 195, + 546, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 195, + 546, + 454 + ], + "spans": [ + { + "bbox": [ + 304, + 195, + 546, + 454 + ], + "type": "text", + "content": "Firstly we define a denser set of viewpoints " + }, + { + "bbox": [ + 304, + 195, + 546, + 454 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{(inpainting)} = \\{C_1,..,C_{N_1}\\}" + }, + { + "bbox": [ + 304, + 195, + 546, + 454 + ], + "type": "text", + "content": " as compared to " + }, + { + "bbox": [ + 304, + 195, + 546, + 454 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{(samping)}" + }, + { + "bbox": [ + 304, + 195, + 546, + 454 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 195, + 546, + 454 + ], + "type": "inline_equation", + "content": "N_{1} > N" + }, + { + "bbox": [ + 304, + 195, + 546, + 454 + ], + "type": "text", + "content": ". We use a mask " + }, + { + "bbox": [ + 304, + 195, + 546, + 454 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_C^{blank}" + }, + { + "bbox": [ + 304, + 195, + 546, + 454 + ], + "type": "text", + "content": " to indicate the rendered areas with blank textures at the viewpoint " + }, + { + "bbox": [ + 304, + 195, + 546, + 454 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 195, + 546, + 454 + ], + "type": "text", + "content": ". The entire inpainting epoch comprises " + }, + { + "bbox": [ + 304, + 195, + 546, + 454 + ], + "type": "inline_equation", + "content": "N_{1}" + }, + { + "bbox": [ + 304, + 195, + 546, + 454 + ], + "type": "text", + "content": " iterations. In each iteration, we first compute the viewpoint " + }, + { + "bbox": [ + 304, + 195, + 546, + 454 + ], + "type": "inline_equation", + "content": "C_n = \\arg \\max_C\\mathbf{M}_C^{blank}" + }, + { + "bbox": [ + 304, + 195, + 546, + 454 + ], + "type": "text", + "content": " with the largest blank area. Then, we perform inpainting on the rendering " + }, + { + "bbox": [ + 304, + 195, + 546, + 454 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_n^{(img)}" + }, + { + "bbox": [ + 304, + 195, + 546, + 454 + ], + "type": "text", + "content": " using the pretrained depth conditioned Stable Diffusion model as described in [6, 33]. During inpainting, Stable Diffusion utilizes regions with textures as conditions to fill in blank areas without textures. To enhance the robustness of inpainting, we render an image using a viewpoint " + }, + { + "bbox": [ + 304, + 195, + 546, + 454 + ], + "type": "inline_equation", + "content": "C_0" + }, + { + "bbox": [ + 304, + 195, + 546, + 454 + ], + "type": "text", + "content": " from the set " + }, + { + "bbox": [ + 304, + 195, + 546, + 454 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{(sampling)}" + }, + { + "bbox": [ + 304, + 195, + 546, + 454 + ], + "type": "text", + "content": ". We then concatenate this image to the image requiring inpainting as a additional condition for inpainting. Compared to naive inpainting, reference-based inpainting provides a more comprehensive condition. To ensure a natural and smooth transition between blank areas and textured regions, we apply Gaussian blur to the mask " + }, + { + "bbox": [ + 304, + 195, + 546, + 454 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{C_n}^{blank}" + }, + { + "bbox": [ + 304, + 195, + 546, + 454 + ], + "type": "text", + "content": ". Finally the inpainted rendering " + }, + { + "bbox": [ + 304, + 195, + 546, + 454 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_n^{(img)'}" + }, + { + "bbox": [ + 304, + 195, + 546, + 454 + ], + "type": "text", + "content": " is used to update the texture map:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 347, + 464, + 545, + 496 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 464, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 347, + 464, + 545, + 496 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {T} _ {n} = \\mathcal {T} _ {n - 1} \\times (1 - \\mathcal {R} ^ {- 1} (\\mathbf {M} ^ {(b l a n k)}) \\tag {9} \\\\ + \\mathcal {R} ^ {- 1} \\left(\\mathbf {x} _ {n} ^ {(i m g) \\prime}\\right) \\times \\mathcal {R} ^ {- 1} \\left(\\mathbf {M} ^ {(b l a n k)}\\right) \\\\ \\end{array}", + "image_path": "d60e1d55a2b1fd8d1a1d8d4be544e03c6d660e506872fb5961dee3a0b7cc157d.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 506, + 545, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 545, + 530 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 545, + 530 + ], + "type": "text", + "content": "We start with " + }, + { + "bbox": [ + 304, + 506, + 545, + 530 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_0 = \\mathcal{T}_{\\text{sampled}}" + }, + { + "bbox": [ + 304, + 506, + 545, + 530 + ], + "type": "text", + "content": " and iterate for " + }, + { + "bbox": [ + 304, + 506, + 545, + 530 + ], + "type": "inline_equation", + "content": "N_1" + }, + { + "bbox": [ + 304, + 506, + 545, + 530 + ], + "type": "text", + "content": " times. Finally we get " + }, + { + "bbox": [ + 304, + 506, + 545, + 530 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\text{inpainted}} = \\mathcal{T}_{N_1}" + }, + { + "bbox": [ + 304, + 506, + 545, + 530 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 531, + 545, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 531, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 545, + 591 + ], + "type": "text", + "content": "It's worth noting that, unlike [2, 6, 33], which relies heavily on inpainting to maintain multi-view consistency, in our method inpainting is performed only to patch small areas. Therefore, the inpainting's performance does not significantly affect the final quality of the generated texture." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 609, + 451, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 609, + 451, + 621 + ], + "spans": [ + { + "bbox": [ + 306, + 609, + 451, + 621 + ], + "type": "text", + "content": "3.4.2 Post-processing:Img2Img" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 629, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 677 + ], + "type": "text", + "content": "So far our generated texture map " + }, + { + "bbox": [ + 304, + 629, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\text {inpainted }}" + }, + { + "bbox": [ + 304, + 629, + 545, + 677 + ], + "type": "text", + "content": " may still contain unnatural transitions in some regions due to multi-view conflicts and inpainting artifacts. To further improve the generation quality, we introduce an Img2Img epoch." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "text", + "content": "In Img2Img epoch, we also define a set of viewpoints " + }, + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "C^{(img2img)}" + }, + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "text", + "content": " include " + }, + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "N_{2}" + }, + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "text", + "content": " viewpoints and iterate through viewpoint from " + }, + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "C^{(img2img)}" + }, + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "text", + "content": ". In each iteration, we encode" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4624" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 71, + 289, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 71, + 289, + 207 + ], + "spans": [ + { + "bbox": [ + 46, + 71, + 289, + 207 + ], + "type": "text", + "content": "the rendering image " + }, + { + "bbox": [ + 46, + 71, + 289, + 207 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_n^{(img)}" + }, + { + "bbox": [ + 46, + 71, + 289, + 207 + ], + "type": "text", + "content": " to the latent space using the encoder " + }, + { + "bbox": [ + 46, + 71, + 289, + 207 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 46, + 71, + 289, + 207 + ], + "type": "text", + "content": " and add some noise to the latent rendering. Then we perform denoising using the pretrained depth conditioned Stable Diffusion model. And finally we decode the denoised latent back to image space using " + }, + { + "bbox": [ + 46, + 71, + 289, + 207 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 71, + 289, + 207 + ], + "type": "text", + "content": ". However, " + }, + { + "bbox": [ + 46, + 71, + 289, + 207 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(\\mathcal{E}(\\cdot))" + }, + { + "bbox": [ + 46, + 71, + 289, + 207 + ], + "type": "text", + "content": " is not an identity transformation because of the reconstruction distortion of the autoencoder. So we apply the color histogram matching to the foreground of the decoded image to alleviate the color distortion. Finally we update the texture map using the img2img result " + }, + { + "bbox": [ + 46, + 71, + 289, + 207 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_n^{(img)'}" + }, + { + "bbox": [ + 46, + 71, + 289, + 207 + ], + "type": "text", + "content": " based on view similarity " + }, + { + "bbox": [ + 46, + 71, + 289, + 207 + ], + "type": "inline_equation", + "content": "\\mathbf{N}_n" + }, + { + "bbox": [ + 46, + 71, + 289, + 207 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 98, + 215, + 286, + 246 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 215, + 286, + 246 + ], + "spans": [ + { + "bbox": [ + 98, + 215, + 286, + 246 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {T} _ {n} = \\mathcal {T} _ {n - 1} \\times (1 - \\mathcal {R} ^ {- 1} (\\mathbf {N} _ {n})) \\\\ + \\mathcal {R} ^ {- 1} \\left(\\mathbf {x} _ {n} ^ {(i m g) \\prime}\\right) \\times \\mathcal {R} ^ {- 1} \\left(\\mathbf {N} _ {n}\\right) \\tag {10} \\\\ \\end{array}", + "image_path": "51257a226fcb09dff56a42ef09a3f9e5ce52c2797c5bac761868f6931e124303.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 255, + 287, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 255, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 47, + 255, + 287, + 281 + ], + "type": "text", + "content": "We start with " + }, + { + "bbox": [ + 47, + 255, + 287, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_0 = \\mathcal{T}_{\\text{inpainted}}" + }, + { + "bbox": [ + 47, + 255, + 287, + 281 + ], + "type": "text", + "content": ". After " + }, + { + "bbox": [ + 47, + 255, + 287, + 281 + ], + "type": "inline_equation", + "content": "N_2" + }, + { + "bbox": [ + 47, + 255, + 287, + 281 + ], + "type": "text", + "content": " iterations we get " + }, + { + "bbox": [ + 47, + 255, + 287, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\text{refined}} = \\mathcal{T}_{N_2}" + }, + { + "bbox": [ + 47, + 255, + 287, + 281 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 291, + 128, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 291, + 128, + 304 + ], + "spans": [ + { + "bbox": [ + 47, + 291, + 128, + 304 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 311, + 97, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 311, + 97, + 324 + ], + "spans": [ + { + "bbox": [ + 47, + 311, + 97, + 324 + ], + "type": "text", + "content": "4.1. Setup" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 329, + 287, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 329, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 46, + 329, + 287, + 474 + ], + "type": "text", + "content": "Implementation Details. Our experiments are conducted on an NVIDIA A10 GPU. We utilize the official Stable Diffusion-v1.5 model with the ControlNet-v1.1 (depth). For Inpainting epoch and Img2Img epoch, we use DDIM [39] as the sampler. For all samplers, we set the number of iterations to 20 steps and the CFG scale (classifier-free guidance scale) to 7.5. We implement the rendering function using nvdiffrast [17] and make modifications to nvdiffrast to support inverse rendering. Texture space sampling takes about 3 minutes to generate textures for a single object. For more detailed parameter settings, please refer to the supplementary materials." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 474, + 287, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 287, + 557 + ], + "type": "text", + "content": "Dataset. Similar to TexFusion [2], we utilize 35 meshes to test the texture generation performance, including 11 objects from Objaverse [10], 17 from ShapeNet [5], 1 from Text2Mesh [26], 2 from Turbosquid [1], 2 from Stanford 3D Scans [9, 45] and 2 from Three D Scans [18]. Each object has 1-4 text descriptions, resulting in a total of 80 (mesh, prompt) pairs in this collection." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": "Baselines. We conduct comparisons with state-of-the-art methods for generating textures from text: 1) Text2Mesh [26], a method that stylizes a 3D mesh by predicting color and local geometric details which conform to a target text prompt, harnessing the representational power of CLIP. 2) Latent-Paint [25], an approach leveraging SDS to obtain texture gradients from the Stable Diffusion model. 3) Text2Tex [6], a method that uses a depth-aware image inpainting diffusion model to incrementally produce partial textures from various viewpoints. 4) TEXTure [33], similar to Text2Tex, but utilizes a different region segmentation strategy. 5) TexFusion [2], a method sequentially inpainting latent texture. As TexFusion does not have publicly" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 309, + 70, + 550, + 169 + ], + "blocks": [ + { + "bbox": [ + 309, + 70, + 550, + 169 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 550, + 169 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 550, + 169 + ], + "type": "table", + "html": "
MethodFID (↓)KID (↓) (×10-3)User study (%)
Visual Quality (↑)Align with Prompt (↑)
Latent-Paint110.1410.645.154.28
Text2Mesh121.6115.133.154.86
Text2Tex101.388.3518.2819.72
TEXTure100.479.2223.4224.86
Ours74.582.8950.0046.28
", + "image_path": "92f9dcb881b28ce9d41bd7eddba908fba16b8ed645c77f3957ecfe4ba76997f6.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 309, + 201, + 550, + 286 + ], + "blocks": [ + { + "bbox": [ + 320, + 177, + 531, + 189 + ], + "lines": [ + { + "bbox": [ + 320, + 177, + 531, + 189 + ], + "spans": [ + { + "bbox": [ + 320, + 177, + 531, + 189 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons with baseline methods." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 201, + 550, + 286 + ], + "lines": [ + { + "bbox": [ + 309, + 201, + 550, + 286 + ], + "spans": [ + { + "bbox": [ + 309, + 201, + 550, + 286 + ], + "type": "table", + "html": "
texture space samplingInpainting RoundImg2Img RoundFID(↓)KID(↓) (×10-3)
✓✗✗86.253.79
✓✗✓82.053.67
✓✓✗76.303.29
✓✓✓74.582.89
", + "image_path": "0d16068cfd6512d63d1279703ca5186d50b291f714f54b3564e89212c2533e75.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 344, + 295, + 506, + 305 + ], + "lines": [ + { + "bbox": [ + 344, + 295, + 506, + 305 + ], + "spans": [ + { + "bbox": [ + 344, + 295, + 506, + 305 + ], + "type": "text", + "content": "Table 2. Effectiveness of texture refinement." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 321, + 546, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 321, + 546, + 346 + ], + "spans": [ + { + "bbox": [ + 305, + 321, + 546, + 346 + ], + "type": "text", + "content": "available source code, we extract some rendered images of generated results from their paper for comparison." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 354, + 424, + 368 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 354, + 424, + 368 + ], + "spans": [ + { + "bbox": [ + 305, + 354, + 424, + 368 + ], + "type": "text", + "content": "4.2. Qualitative Analysis" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 373, + 546, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 546, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 546, + 540 + ], + "type": "text", + "content": "Fig. 3 and Fig. 5 presents visual comparisons between our method and baseline approaches. Textures produced by Text2Mesh are notably lacking in detail. Latent-Paint yields textures with greater intricacy compared to Text2Mesh, but they still exhibit considerable blurriness due to SDS limitations. Inpainting-based approaches like Text2Tex and TEXTure introduce artifacts with unnatural transitions at inpainting mask boundaries, struggling to maintain consistency across multiple views because they depend heavily on inpainting. The output of TexFusion is compromised by blurry details, likely a consequence of repeated noise addition during sequential sampling. In contrast, our method stands out by producing textures that are not only clear but also remarkably coherent." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 541, + 545, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 541, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 304, + 541, + 545, + 578 + ], + "type": "text", + "content": "Fig. 4 presents more results of our method. Benefiting from the powerful prior of Stable Diffusion, our method can generate textures of diverse styles for different geometries." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 586, + 452, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 586, + 452, + 600 + ], + "spans": [ + { + "bbox": [ + 305, + 586, + 452, + 600 + ], + "type": "text", + "content": "4.3. Quantitative Comparisons" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "FID & KID. Similar to TexFusion, we sample from pretrained image diffusion model to create ground truth labels. For each mesh, we generate depth maps from the five most common canonical viewpoints: front, back, top, and both sides. These depth maps, along with textual descriptions, serve as inputs to condition the Stable Diffusion model. To ensure the focus remains on the texture of the objects, we modify the background of the ground truth images to be white. We visualize the textures created by various methods" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4625" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 72, + 153, + 133 + ], + "blocks": [ + { + "bbox": [ + 85, + 72, + 153, + 133 + ], + "lines": [ + { + "bbox": [ + 85, + 72, + 153, + 133 + ], + "spans": [ + { + "bbox": [ + 85, + 72, + 153, + 133 + ], + "type": "image", + "image_path": "fdc48c32a5f11f589a8403e43677db985a342b561a9f5fd1628db52124b015d8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 95, + 142, + 160, + 219 + ], + "blocks": [ + { + "bbox": [ + 95, + 142, + 160, + 219 + ], + "lines": [ + { + "bbox": [ + 95, + 142, + 160, + 219 + ], + "spans": [ + { + "bbox": [ + 95, + 142, + 160, + 219 + ], + "type": "image", + "image_path": "d2b4d22b3d333ef1d7edae3141d9c9771ecd38c9bd6184a063e54638a0195a46.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 84, + 220, + 146, + 298 + ], + "blocks": [ + { + "bbox": [ + 55, + 75, + 82, + 357 + ], + "lines": [ + { + "bbox": [ + 55, + 75, + 82, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 75, + 82, + 357 + ], + "type": "text", + "content": "red apple blue handbag with dilapidated Halloween silver trims tap pumpkin head" + } + ] + } + ], + "index": 0, + "angle": 270, + "type": "image_caption" + }, + { + "bbox": [ + 84, + 220, + 146, + 298 + ], + "lines": [ + { + "bbox": [ + 84, + 220, + 146, + 298 + ], + "spans": [ + { + "bbox": [ + 84, + 220, + 146, + 298 + ], + "type": "image", + "image_path": "cd9686633cb48e567e402d2f5a1e66ca050d98576870fb1a65fc642231d7b828.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 88, + 302, + 151, + 365 + ], + "blocks": [ + { + "bbox": [ + 88, + 302, + 151, + 365 + ], + "lines": [ + { + "bbox": [ + 88, + 302, + 151, + 365 + ], + "spans": [ + { + "bbox": [ + 88, + 302, + 151, + 365 + ], + "type": "image", + "image_path": "f66f2cec905dfc5ae96e8bae3daf9f318f78329bcf9d2dfd3cb9c6e61c7d4c31.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 96, + 378, + 138, + 390 + ], + "lines": [ + { + "bbox": [ + 96, + 378, + 138, + 390 + ], + "spans": [ + { + "bbox": [ + 96, + 378, + 138, + 390 + ], + "type": "text", + "content": "Geometry" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 157, + 72, + 223, + 132 + ], + "blocks": [ + { + "bbox": [ + 157, + 72, + 223, + 132 + ], + "lines": [ + { + "bbox": [ + 157, + 72, + 223, + 132 + ], + "spans": [ + { + "bbox": [ + 157, + 72, + 223, + 132 + ], + "type": "image", + "image_path": "3a88db337c4d909230a438642665240e74ccf6d986ec7814a7af27b32287adb9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 168, + 142, + 232, + 219 + ], + "blocks": [ + { + "bbox": [ + 168, + 142, + 232, + 219 + ], + "lines": [ + { + "bbox": [ + 168, + 142, + 232, + 219 + ], + "spans": [ + { + "bbox": [ + 168, + 142, + 232, + 219 + ], + "type": "image", + "image_path": "3b06058ff15f1977324d52c080347bdcbe8f500089ab05a2202185df47364ca5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 156, + 220, + 218, + 297 + ], + "blocks": [ + { + "bbox": [ + 156, + 220, + 218, + 297 + ], + "lines": [ + { + "bbox": [ + 156, + 220, + 218, + 297 + ], + "spans": [ + { + "bbox": [ + 156, + 220, + 218, + 297 + ], + "type": "image", + "image_path": "8ff9f5e1793f898017a1bc68710f842a758019abbf69896eef9bea814761c1eb.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 160, + 302, + 222, + 365 + ], + "blocks": [ + { + "bbox": [ + 160, + 302, + 222, + 365 + ], + "lines": [ + { + "bbox": [ + 160, + 302, + 222, + 365 + ], + "spans": [ + { + "bbox": [ + 160, + 302, + 222, + 365 + ], + "type": "image", + "image_path": "7945def590a373bcbc75c0fb67f16d2b1e5133ea559d029245dda8edbc2b3f83.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 170, + 378, + 216, + 389 + ], + "lines": [ + { + "bbox": [ + 170, + 378, + 216, + 389 + ], + "spans": [ + { + "bbox": [ + 170, + 378, + 216, + 389 + ], + "type": "text", + "content": "Text2Mesh" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 236, + 72, + 302, + 132 + ], + "blocks": [ + { + "bbox": [ + 236, + 72, + 302, + 132 + ], + "lines": [ + { + "bbox": [ + 236, + 72, + 302, + 132 + ], + "spans": [ + { + "bbox": [ + 236, + 72, + 302, + 132 + ], + "type": "image", + "image_path": "67daf3b8d0a1cb7133e37f85e82ba0b30d3bdee3f6a8d82e2941c5c70d1d072a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 247, + 142, + 312, + 219 + ], + "blocks": [ + { + "bbox": [ + 247, + 142, + 312, + 219 + ], + "lines": [ + { + "bbox": [ + 247, + 142, + 312, + 219 + ], + "spans": [ + { + "bbox": [ + 247, + 142, + 312, + 219 + ], + "type": "image", + "image_path": "83261be0dd824e36413c37d8e7f574daa425ea68b3d4c2e6ce7795bd49e806ed.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 235, + 220, + 296, + 297 + ], + "blocks": [ + { + "bbox": [ + 235, + 220, + 296, + 297 + ], + "lines": [ + { + "bbox": [ + 235, + 220, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 235, + 220, + 296, + 297 + ], + "type": "image", + "image_path": "1f9823e5d109eb8e75ffdc9aa1dc3372594c171c1aa74e9150ac57c7d2e243ac.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 239, + 302, + 301, + 365 + ], + "blocks": [ + { + "bbox": [ + 239, + 302, + 301, + 365 + ], + "lines": [ + { + "bbox": [ + 239, + 302, + 301, + 365 + ], + "spans": [ + { + "bbox": [ + 239, + 302, + 301, + 365 + ], + "type": "image", + "image_path": "5331997ccdbd20e40b77c9a5cf15142ddbcfb028f99c89ddc14fdf87e38a3435.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 247, + 378, + 297, + 389 + ], + "lines": [ + { + "bbox": [ + 247, + 378, + 297, + 389 + ], + "spans": [ + { + "bbox": [ + 247, + 378, + 297, + 389 + ], + "type": "text", + "content": "Latent-Paint" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 320, + 72, + 386, + 132 + ], + "blocks": [ + { + "bbox": [ + 320, + 72, + 386, + 132 + ], + "lines": [ + { + "bbox": [ + 320, + 72, + 386, + 132 + ], + "spans": [ + { + "bbox": [ + 320, + 72, + 386, + 132 + ], + "type": "image", + "image_path": "3009c04219e35ab44fb8edcd5e209d390537087a45da4d44ac34789bc6d14312.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 331, + 142, + 394, + 219 + ], + "blocks": [ + { + "bbox": [ + 331, + 142, + 394, + 219 + ], + "lines": [ + { + "bbox": [ + 331, + 142, + 394, + 219 + ], + "spans": [ + { + "bbox": [ + 331, + 142, + 394, + 219 + ], + "type": "image", + "image_path": "daa7fdce027551be5438c7875d2230a42ee1ca694bcf779108c9ed074e7e7d43.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 318, + 220, + 380, + 297 + ], + "blocks": [ + { + "bbox": [ + 318, + 220, + 380, + 297 + ], + "lines": [ + { + "bbox": [ + 318, + 220, + 380, + 297 + ], + "spans": [ + { + "bbox": [ + 318, + 220, + 380, + 297 + ], + "type": "image", + "image_path": "98524417bb1e4c9adffb95fab197ffe25236610af7188078f2466a75d1613fbf.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 322, + 302, + 385, + 365 + ], + "blocks": [ + { + "bbox": [ + 322, + 302, + 385, + 365 + ], + "lines": [ + { + "bbox": [ + 322, + 302, + 385, + 365 + ], + "spans": [ + { + "bbox": [ + 322, + 302, + 385, + 365 + ], + "type": "image", + "image_path": "096a9d24c767616a95b05312f84c308ec0cc7a2a557e4f4e1fd4bbdfbb47fefd.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 378, + 373, + 388 + ], + "lines": [ + { + "bbox": [ + 335, + 378, + 373, + 388 + ], + "spans": [ + { + "bbox": [ + 335, + 378, + 373, + 388 + ], + "type": "text", + "content": "Text2Tex" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 400, + 72, + 464, + 132 + ], + "blocks": [ + { + "bbox": [ + 400, + 72, + 464, + 132 + ], + "lines": [ + { + "bbox": [ + 400, + 72, + 464, + 132 + ], + "spans": [ + { + "bbox": [ + 400, + 72, + 464, + 132 + ], + "type": "image", + "image_path": "fa07765acbd0f50df1805ac367794bb54fddb0f5c3e57ea338641640a0a91a20.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 410, + 142, + 473, + 219 + ], + "blocks": [ + { + "bbox": [ + 410, + 142, + 473, + 219 + ], + "lines": [ + { + "bbox": [ + 410, + 142, + 473, + 219 + ], + "spans": [ + { + "bbox": [ + 410, + 142, + 473, + 219 + ], + "type": "image", + "image_path": "edcbf3c8900402c219af8ede23e95f6fbb7976280aef88e686e3539faeedf629.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 397, + 220, + 459, + 297 + ], + "blocks": [ + { + "bbox": [ + 397, + 220, + 459, + 297 + ], + "lines": [ + { + "bbox": [ + 397, + 220, + 459, + 297 + ], + "spans": [ + { + "bbox": [ + 397, + 220, + 459, + 297 + ], + "type": "image", + "image_path": "828c1245e3f85806bd29afdf6fc9b3679e5ae7c2cd69dec2ef792b8a82e60d95.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 402, + 302, + 463, + 365 + ], + "blocks": [ + { + "bbox": [ + 402, + 302, + 463, + 365 + ], + "lines": [ + { + "bbox": [ + 402, + 302, + 463, + 365 + ], + "spans": [ + { + "bbox": [ + 402, + 302, + 463, + 365 + ], + "type": "image", + "image_path": "a72a7a4f53dee22f4d9d2c6be81a6962747f89409f477c43e87c5a19e303c1c4.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 412, + 378, + 451, + 388 + ], + "lines": [ + { + "bbox": [ + 412, + 378, + 451, + 388 + ], + "spans": [ + { + "bbox": [ + 412, + 378, + 451, + 388 + ], + "type": "text", + "content": "TEXTure" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 471, + 72, + 535, + 132 + ], + "blocks": [ + { + "bbox": [ + 471, + 72, + 535, + 132 + ], + "lines": [ + { + "bbox": [ + 471, + 72, + 535, + 132 + ], + "spans": [ + { + "bbox": [ + 471, + 72, + 535, + 132 + ], + "type": "image", + "image_path": "07235b9d9c58d72508c23fdf78daa077cece9bc6d9679d97f5ea04b4883e8476.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 481, + 142, + 545, + 219 + ], + "blocks": [ + { + "bbox": [ + 481, + 142, + 545, + 219 + ], + "lines": [ + { + "bbox": [ + 481, + 142, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 481, + 142, + 545, + 219 + ], + "type": "image", + "image_path": "4c567eede86508bf4de02ef0a5b81d56a68996781f88cae11fb1fb0c88614f37.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 469, + 220, + 529, + 297 + ], + "blocks": [ + { + "bbox": [ + 469, + 220, + 529, + 297 + ], + "lines": [ + { + "bbox": [ + 469, + 220, + 529, + 297 + ], + "spans": [ + { + "bbox": [ + 469, + 220, + 529, + 297 + ], + "type": "image", + "image_path": "e84a26487879ef9aa4a5b34d08dfb19ae9962f45351b79752c93cef507cb3626.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 473, + 302, + 535, + 365 + ], + "blocks": [ + { + "bbox": [ + 473, + 302, + 535, + 365 + ], + "lines": [ + { + "bbox": [ + 473, + 302, + 535, + 365 + ], + "spans": [ + { + "bbox": [ + 473, + 302, + 535, + 365 + ], + "type": "image", + "image_path": "178bf64a2ad0b905a0aa74428bfc2d9d09743d1c125800715eb0b4a273a7de53.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 492, + 378, + 515, + 388 + ], + "lines": [ + { + "bbox": [ + 492, + 378, + 515, + 388 + ], + "spans": [ + { + "bbox": [ + 492, + 378, + 515, + 388 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 50, + 439, + 286, + 596 + ], + "blocks": [ + { + "bbox": [ + 46, + 398, + 546, + 420 + ], + "lines": [ + { + "bbox": [ + 46, + 398, + 546, + 420 + ], + "spans": [ + { + "bbox": [ + 46, + 398, + 546, + 420 + ], + "type": "text", + "content": "Figure 3. Qualitative comparisons with Text2Mesh [26], Latent-Paint [25], Text2Tex [6] and TEXTure [33]. In comparison with the baselines, our GenesisTex exhibits richer details and fewer artifacts." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 439, + 286, + 596 + ], + "lines": [ + { + "bbox": [ + 50, + 439, + 286, + 596 + ], + "spans": [ + { + "bbox": [ + 50, + 439, + 286, + 596 + ], + "type": "image", + "image_path": "5687d71dca2d7f983a1e3308812f90633d508936eed8e3c02bdf84e0086618b8.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 85, + 606, + 248, + 617 + ], + "lines": [ + { + "bbox": [ + 85, + 606, + 248, + 617 + ], + "spans": [ + { + "bbox": [ + 85, + 606, + 248, + 617 + ], + "type": "text", + "content": "Figure 4. Texturing results with GenesisTex." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "text", + "content": "from the same five viewpoints and evaluate their quality by calculating the Fréchet Inception Distance (FID) and Kernel Inception Distance (KID). FID and KID measure the feature dissimilarity between two image collections, with feature extraction performed using the Inception V3 [41]. The results in Tab.1 demonstrate that the textures gener" + } + ] + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 310, + 439, + 539, + 575 + ], + "blocks": [ + { + "bbox": [ + 310, + 439, + 539, + 575 + ], + "lines": [ + { + "bbox": [ + 310, + 439, + 539, + 575 + ], + "spans": [ + { + "bbox": [ + 310, + 439, + 539, + 575 + ], + "type": "image", + "image_path": "da735ee0a098b3bb990e17955ad314fed34a91df832ffb3f5f255f720641bce9.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 585, + 545, + 608 + ], + "lines": [ + { + "bbox": [ + 305, + 585, + 545, + 608 + ], + "spans": [ + { + "bbox": [ + 305, + 585, + 545, + 608 + ], + "type": "text", + "content": "Figure 5. Qualitative Comparisons with TexFusion [2]. Images are extracted from their original paper." + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "bbox": [ + 304, + 629, + 545, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 666 + ], + "type": "text", + "content": "ated by GenesisTex are closer to the ground truth compared to those produced by baselines, indicating a superior synthesis quality." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 304, + 666, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 546, + 713 + ], + "type": "text", + "content": "User Study. We conduct a user study to compare the quality of textures generated by our method versus those produced by baseline methods. We render all textured mesh results into videos to facilitate the presentation to users. Each" + } + ] + } + ], + "index": 38 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4626" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 277 + ], + "type": "text", + "content": "video contains a rotating textured mesh along with the corresponding prompt. For each questionnaire, we randomly show the users 10 groups of rendered videos. Each group displays results for a specific (mesh, prompt) pair, comparing outputs from baseline methods and our approach. Each volunteer is required to answer two questions for a single group, including which one has the highest visual quality and which one aligns best with the prompt. Additional information about the questionnaire is available in the supplementary. Finally, we received 35 valid responses from the questionnaires, which included the results of comparisons from 350 groups. The results of the user study are shown in Tab. 1. In terms of both visual quality and alignment with the text prompt, our method is preferred by the most participants, with percentages reaching " + }, + { + "bbox": [ + 46, + 72, + 289, + 277 + ], + "type": "inline_equation", + "content": "50.00\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 277 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 289, + 277 + ], + "type": "inline_equation", + "content": "46.28\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 277 + ], + "type": "text", + "content": ", respectively. Furthermore, we conduct a pairwise comparison study and the results are included in the supplementary." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 285, + 149, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 285, + 149, + 298 + ], + "spans": [ + { + "bbox": [ + 47, + 285, + 149, + 298 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 304, + 290, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 290, + 473 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 290, + 473 + ], + "type": "text", + "content": "Consistency in Texture Space Sampling. Our GenesisTex employs style consistency and dynamic alignment for maintaining multi-view consistency during texture space sampling. To investigate the impact of these two consistency strategies, we visualize the decoded multi-view images. We set " + }, + { + "bbox": [ + 46, + 304, + 290, + 473 + ], + "type": "inline_equation", + "content": "\\mathcal{C}(elevation, azimuth) = \\{(0^{\\circ}, 0^{\\circ}), (0^{\\circ}, 15^{\\circ}), (0^{\\circ}, 35^{\\circ}), (0^{\\circ}, 45^{\\circ})\\}" + }, + { + "bbox": [ + 46, + 304, + 290, + 473 + ], + "type": "text", + "content": " to ensure that multiple viewpoints share sufficient content. Fig. 6 illustrates an example with the prompt penguin toy. We can observe that without any consistency constraints, the color of the toy varies significant across different viewpoints. After introducing style consistency, consistency for the toy color improves, but there are still inconsistencies in the hair and ears. Further we use dynamic alignment and we can see the" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 50, + 492, + 284, + 668 + ], + "blocks": [ + { + "bbox": [ + 50, + 492, + 284, + 668 + ], + "lines": [ + { + "bbox": [ + 50, + 492, + 284, + 668 + ], + "spans": [ + { + "bbox": [ + 50, + 492, + 284, + 668 + ], + "type": "image", + "image_path": "78b4fee5eb22640a4d3f53733c8ef35dcc4c98e20a3fba639f6199fc3b4685ff.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 677, + 287, + 711 + ], + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 711 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 711 + ], + "type": "text", + "content": "Figure 6. Ablation results on consistency in texture space sampling. Style Consistency + Dynamic Alignment (bottom row) achieves the best multi-view consistency." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 310, + 73, + 545, + 233 + ], + "blocks": [ + { + "bbox": [ + 310, + 73, + 545, + 233 + ], + "lines": [ + { + "bbox": [ + 310, + 73, + 545, + 233 + ], + "spans": [ + { + "bbox": [ + 310, + 73, + 545, + 233 + ], + "type": "image", + "image_path": "0be7867152068e75ea0142d33a1c55951a20f0b7e89b2e58cfcc6c9c45a70819.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 244, + 546, + 268 + ], + "lines": [ + { + "bbox": [ + 305, + 244, + 546, + 268 + ], + "spans": [ + { + "bbox": [ + 305, + 244, + 546, + 268 + ], + "type": "text", + "content": "Figure 7. Ablation results on texture refinement. Inpainting fills the blank regions, while Img2Img enhances the quality of details." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 292, + 546, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 292, + 546, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 292, + 546, + 327 + ], + "type": "text", + "content": "consistency of details has significantly improved. We include more results in the supplementary to further demonstrate the effectiveness of these two consistencies." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 330, + 547, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 330, + 547, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 330, + 547, + 437 + ], + "type": "text", + "content": "Effectiveness of Texture Refinement. After texture space sampling, we conduct refinement on the texture, including Inpainting and Img2Img. We validate the effectiveness of these two modules and the results are presented in Fig. 7 and Tab. 2. Inpainting fills in areas lacking texture obtained during texture space sampling, contributing significantly to the improvement in visual quality. Img2Img further repairs artifacts and enhances details, leading to a certain improvement in visual quality as well." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 455, + 463, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 455, + 463, + 467 + ], + "spans": [ + { + "bbox": [ + 306, + 455, + 463, + 467 + ], + "type": "text", + "content": "5. Conclusion and Limitations" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 477, + 545, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 477, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 304, + 477, + 545, + 656 + ], + "type": "text", + "content": "In this work, we propose a novel method for text-based texture generation, named GenesisTex. GenesisTex leverages the prior of the pre-trained Stable Diffusion model by introducing texture space sampling. Texture space sampling concurrently generates multi-view content without relying on a predefined sequence of views. Our approach can generate high-quality textures for a given 3D model in several minutes. The primary limitation of our method is the significant memory cost associated with maintaining style consistency, caused by cross-view attention. This limitation restricts the number of viewpoints and necessitates post-processing steps, such as inpainting and img2img. Future work could investigate hierarchical style consistency approaches to reduce the computational costs of cross-view attention by iterating over a smaller set of viewpoints." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 657, + 547, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 657, + 547, + 703 + ], + "spans": [ + { + "bbox": [ + 304, + 657, + 547, + 703 + ], + "type": "text", + "content": "Acknowledgement: This work is supported by the Young Elite Scientists Sponsorship Program by CAST (China Association for Science and Technology) and IEG Moonshot Program by Tencent." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4627" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 269, + 102 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 269, + 102 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 269, + 102 + ], + "type": "text", + "content": "[1] Turbosquid by shutterstock. In www.turbosquid.com. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 103, + 287, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 103, + 287, + 159 + ], + "spans": [ + { + "bbox": [ + 53, + 103, + 287, + 159 + ], + "type": "text", + "content": "[2] Tianshi Cao, Karsten Kreis, Sanja Fidler, Nicholas Sharp, and Kangxue Yin. Texfusion: Synthesizing 3d textures with text-guided image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4169-4181, 2023. 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 160, + 288, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 160, + 288, + 226 + ], + "spans": [ + { + "bbox": [ + 53, + 160, + 288, + 226 + ], + "type": "text", + "content": "[3] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 228, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 228, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 228, + 288, + 282 + ], + "type": "text", + "content": "[4] Eric R Chan, Koki Nagano, Matthew A Chan, Alexander W Bergman, Jeong Joon Park, Axel Levy, Miika Aittala, Shalini De Mello, Tero Karras, and Gordon Wetzstein. Generative novel view synthesis with 3d-aware diffusion models. arXiv preprint arXiv:2304.02602, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 285, + 287, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 285, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 53, + 285, + 287, + 338 + ], + "type": "text", + "content": "[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 341, + 287, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 341, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 341, + 287, + 384 + ], + "type": "text", + "content": "[6] Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text-driven texture synthesis via diffusion models. arXiv preprint arXiv:2303.11396, 2023. 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 386, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 386, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 386, + 287, + 430 + ], + "type": "text", + "content": "[7] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. *Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation.* arXiv preprint arXiv:2303.13873, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 432, + 287, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 432, + 287, + 476 + ], + "spans": [ + { + "bbox": [ + 53, + 432, + 287, + 476 + ], + "type": "text", + "content": "[8] Zhiqin Chen, Kangxue Yin, and Sanja Fidler. Auv-net: Learning aligned uv maps for texture transfer and synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1465-1474, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 478, + 288, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 478, + 288, + 521 + ], + "spans": [ + { + "bbox": [ + 53, + 478, + 288, + 521 + ], + "type": "text", + "content": "[9] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 303-312, 1996. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 523, + 287, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 589 + ], + "type": "text", + "content": "[10] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 591, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 287, + 645 + ], + "type": "text", + "content": "[11] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. Advances In Neural Information Processing Systems, 35:31841-31854, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "text", + "content": "[12] Jiatao Gu, Alex Trevithick, Kai-En Lin, Joshua M Susskind, Christian Theobalt, Lingjie Liu, and Ravi Ramamoorthi. Nerfdiff: Single-image view synthesis with nef-guided distillation from 3d-aware diffusion. In International Conference on Machine Learning, pages 11808-11826. PMLR, 2023. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[13] Zehuan Huang, Hao Wen, Junting Dong, Yaohui Wang, Yangguang Li, Xinyuan Chen, Yan-Pei Cao, Ding Liang, Yu Qiao, Bo Dai, and Lu Sheng. Epidiff: Enhancing multi-view synthesis via localized epipolar-constrained diffusion, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 545, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 182 + ], + "type": "text", + "content": "[14] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 867-876, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 184, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 184, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 184, + 545, + 228 + ], + "type": "text", + "content": "[15] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 228, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 545, + 282 + ], + "type": "text", + "content": "[16] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439, 2023. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 283, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 283, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 307, + 283, + 545, + 327 + ], + "type": "text", + "content": "[17] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics (TOG), 39(6):1-14, 2020. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 327, + 535, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 535, + 338 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 535, + 338 + ], + "type": "text", + "content": "[18] Oliver Laric. Three d scans. In threadscans.com, 2012. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 338, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 338, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 338, + 545, + 403 + ], + "type": "text", + "content": "[19] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300–309, 2023. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 404, + 545, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 404, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 307, + 404, + 545, + 447 + ], + "type": "text", + "content": "[20] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Zexiang Xu, Hao Su, et al. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. arXiv preprint arXiv:2306.16928, 2023." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 449, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 545, + 503 + ], + "type": "text", + "content": "[21] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9298–9309, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 504, + 545, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 504, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 307, + 504, + 545, + 548 + ], + "type": "text", + "content": "[22] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 548, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 548, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 307, + 548, + 545, + 602 + ], + "type": "text", + "content": "[23] Xiaoxiao Long, Cheng Lin, Peng Wang, Taku Komura, and Wenping Wang. Sparseneus: Fast generalizable neural surface reconstruction from sparse views. In European Conference on Computer Vision, pages 210-227. Springer, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 604, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 604, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 307, + 604, + 545, + 657 + ], + "type": "text", + "content": "[24] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. arXiv preprint arXiv:2310.15008, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "type": "text", + "content": "[25] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12663–12673, 2023. 2, 6, 7" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4628" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[26] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 2, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 287, + 172 + ], + "type": "text", + "content": "[27] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 conference papers, pages 1-8, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 287, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 287, + 227 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 287, + 227 + ], + "type": "text", + "content": "[28] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 228, + 287, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 287, + 260 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 287, + 260 + ], + "type": "text", + "content": "[29] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 261, + 287, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 287, + 325 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 287, + 325 + ], + "type": "text", + "content": "[30] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, et al. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. arXiv preprint arXiv:2306.17843, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 327, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 287, + 392 + ], + "type": "text", + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 393, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 287, + 437 + ], + "type": "text", + "content": "[32] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 437, + 287, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 287, + 470 + ], + "type": "text", + "content": "[33] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. arXiv preprint arXiv:2302.01721, 2023. 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 471, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 287, + 525 + ], + "type": "text", + "content": "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 525, + 287, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 591 + ], + "type": "text", + "content": "[35] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 592, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 287, + 657 + ], + "type": "text", + "content": "[36] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 658, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 690 + ], + "type": "text", + "content": "[37] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 2, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "text", + "content": "[38] Yawar Siddiqui, Justus Thies, Fangchang Ma, Qi Shan, Matthias Nießner, and Angela Dai. Texturify: Generating" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 634 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "textures on 3d shape surfaces. In European Conference on Computer Vision, pages 72-88. Springer, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 96, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 128 + ], + "type": "text", + "content": "[39] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 129, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 129, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 308, + 129, + 545, + 162 + ], + "type": "text", + "content": "[40] Jia-Mu Sun, Tong Wu, and Lin Gao. Recent advances in implicit representation-based 3d shape generation. Visual Intelligence, 2(1):9, 2024. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 163, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 217 + ], + "type": "text", + "content": "[41] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2818-2826, 2016. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 219, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 219, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 219, + 545, + 262 + ], + "type": "text", + "content": "[42] Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi. Viewset diffusion:(0-) image-conditioned 3d generative models from 2d data. arXiv preprint arXiv:2306.07881, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 263, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 263, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 263, + 545, + 308 + ], + "type": "text", + "content": "[43] Shitao Tang, Fuyang Zhang, Jiacheng Chen, Peng Wang, and Yasutaka Furukawa. Mvdiffusion: Enabling holistic multiview image generation with correspondence-aware diffusion. arXiv preprint arXiv:2307.01097, 2023." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 309, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 309, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 308, + 309, + 545, + 364 + ], + "type": "text", + "content": "[44] Hung-Yu Tseng, Qinbo Li, Changil Kim, Suhib Alsisan, Jia-Bin Huang, and Johannes Kopf. Consistent view synthesis with pose-guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16773-16783, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 365, + 545, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 545, + 408 + ], + "type": "text", + "content": "[45] Greg Turk and Marc Levoy. Zippered polygon meshes from range images. In Proceedings of the 21st annual conference on Computer graphics and interactive techniques, pages 311-318, 1994. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 410, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 410, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 308, + 410, + 545, + 464 + ], + "type": "text", + "content": "[46] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12619-12629, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 465, + 545, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 465, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 308, + 465, + 545, + 509 + ], + "type": "text", + "content": "[47] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 510, + 545, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 510, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 308, + 510, + 545, + 543 + ], + "type": "text", + "content": "[48] Jianfeng Xiang, Jiaolong Yang, Binbin Huang, and Xin Tong. 3d-aware image generation using 2d diffusion models. arXiv preprint arXiv:2303.17905, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 544, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 544, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 544, + 545, + 578 + ], + "type": "text", + "content": "[49] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. arXiv preprint arXiv:2306.07954, 2023. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 578, + 542, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 578, + 542, + 590 + ], + "spans": [ + { + "bbox": [ + 308, + 578, + 542, + 590 + ], + "type": "text", + "content": "[50] Jonathan Young. xatlas. In github.com/jpcy/xatlas, 2016. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 590, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 545, + 634 + ], + "type": "text", + "content": "[51] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4629" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Genuine Knowledge from Practice_ Diffusion Test-Time Adaptation for Video Adverse Weather Removal/254c789f-0e30-45ba-9906-00c14a08b4b6_content_list.json b/2024/Genuine Knowledge from Practice_ Diffusion Test-Time Adaptation for Video Adverse Weather Removal/254c789f-0e30-45ba-9906-00c14a08b4b6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..01fe8d77d7c62ad3bcaa1be402420265e058650e --- /dev/null +++ b/2024/Genuine Knowledge from Practice_ Diffusion Test-Time Adaptation for Video Adverse Weather Removal/254c789f-0e30-45ba-9906-00c14a08b4b6_content_list.json @@ -0,0 +1,1464 @@ +[ + { + "type": "text", + "text": "Genuine Knowledge from Practice: Diffusion Test-Time Adaptation for Video Adverse Weather Removal", + "text_level": 1, + "bbox": [ + 135, + 130, + 833, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yijun Yang $^{1}$ , Hongtao Wu $^{1}$ , Angelica I. Aviles-Rivero $^{2}$ , Yulun Zhang $^{3}$ , Jing Qin $^{4}$ , Lei Zhu $^{1,5*}$", + "bbox": [ + 119, + 202, + 854, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1The Hong Kong University of Science and Technology (Guangzhou) 2University of Cambridge 3ETH Zürich", + "bbox": [ + 183, + 223, + 785, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4The Hong Kong Polytechnic University 5The Hong Kong University of Science and Technology", + "bbox": [ + 223, + 241, + 750, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Project page: https://github.com/scott-yjyang/DiffTTA", + "bbox": [ + 243, + 257, + 723, + 272 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 306, + 313, + 321 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Real-world vision tasks frequently suffer from the appearance of unexpected adverse weather conditions, including rain, haze, snow, and raindrops. In the last decade, convolutional neural networks and vision transformers have yielded outstanding results in single-weather video removal. However, due to the absence of appropriate adaptation, most of them fail to generalize to other weather conditions. Although ViWS-Net is proposed to remove adverse weather conditions in videos with a single set of pretrained weights, it is seriously blinded by seen weather at train-time and degenerates when coming to unseen weather during test-time. In this work, we introduce test-time adaptation into adverse weather removal in videos, and propose the first framework that integrates test-time adaptation into the iterative diffusion reverse process. Specifically, we devise a diffusion-based network with a novel temporal noise model to efficiently explore frame-correlated information in degraded video clips at training stage. During inference stage, we introduce a proxy task named Diffusion Tubelet Self-Calibration to learn the primer distribution of test video stream and optimize the model by approximating the temporal noise model for online adaptation. Experimental results, on benchmark datasets, demonstrate that our Test-Time Adaptation method with Diffusion-based network(Diff-TTA) outperforms state-of-the-art methods in terms of restoring videos degraded by seen weather conditions. Its generalizable capability is validated with unseen weather conditions in synthesized and real-world videos.", + "bbox": [ + 75, + 339, + 472, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 791, + 207, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Adverse weather like rain, snow, and haze is very common in many existing outdoor videos. The presence of adverse weather usually decreases the visibility of the captured videos and significantly impairs the performance of", + "bbox": [ + 76, + 816, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/293ac8b6fdcba7f56490225645c32bd6b578be91b5edf728c832968484f0e142.jpg", + "image_caption": [ + "Figure 1. Overview of the existing all-in-one adverse weather removal methods. Our approach can achieve superior performance not only in seen weather conditions but also in unseen weather conditions with a single set of pre-trained weights by diffusion test-time adaptation. In particular, our Diff-TTA is $90 \\times$ more efficient than WeatherDiffusion." + ], + "image_footnote": [], + "bbox": [ + 504, + 305, + 893, + 491 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "subsequent high-level vision applications such as object detection, semantic segmentation and autonomous driving. To mitigate these negative effects, Convolutional Neural Networks (CNNs) and Transformers are frequently utilized in the literature. Also, Diffusion models have emerged as the de-facto generative model for image synthesis [14, 16, 36] and have been developed for single-image adverse weather removal [33]. While deep neural networks can achieve decent results on test points within the distribution, real-world applications often demand robustness. This is especially crucial when facing unknown input perturbations, changes in weather conditions, or other sources of distribution shift. Existing single-weather removal approaches [4, 20, 59] require follow-up training to adapt to specific domains. This leads to the switching between different parameter sets, which can make the pipeline cumbersome.", + "bbox": [ + 496, + 597, + 892, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, all-in-one adverse weather removal has attracted the attention of the community, which aims to handle diverse weather conditions with a generic model [11, 21, 38, 51]. However, the potential issue of frequently failing to", + "bbox": [ + 498, + 839, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Lei Zhu (leizhu@ust.hk) is the corresponding author.", + "bbox": [ + 94, + 886, + 382, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "25606", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ee45d853939569f13235d9fc562ce0f4228cfb244b9e8204dd1ba663bc36db4d.jpg", + "image_caption": [ + "Figure 2. Our Diff-TTA enables weather removal models to overcome unseen weather corruptions. We use t-SNE [39] to visualize features from the last feature extractor layer of each dataset. Obviously, unseen data points tend to approximate the seen ones after adaptation, which means Diff-TTA can categorize unknown degradation into known distribution. ('Real-world' contains video clips simultaneously degraded by fog and snow.)" + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 472, + 205 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "recover from out-of-distribution conditions remains largely unaddressed due to the finiteness of training data and the absence of a test data adaptation mechanism. As shown in Figure 1, generic models suffer significant degradation when handling all the possible weather conditions beyond the scope of their training data. Furthermore, a common aspect among many real-world applications is the necessity for online adaptation with limited data. This is particularly relevant in scenarios such as autonomous vehicles or drones, where a vision model must continually process a video stream and adjust its predictions. To achieve effective adaptation, it is crucial to leverage the incoming data in a way that is beneficial and does not introduce harmful biases, even when there are distribution shifts between the training and test data. Thus, akin to how humans update knowledge through practice, the task of removing any weather condition through online adaptation using a single set of pre-trained weights is both pragmatic and promising, albeit challenging.", + "bbox": [ + 75, + 325, + 473, + 613 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we construct the diffusion-based framework for video adverse weather removal. Yet, diffusion models usually entail significant training overhead, hindering the technique's broader adoption in the research community of video tasks. Instead of introducing massive parameters for temporal encoding, we design an efficient strategy based on DDIM [37] to preserve temporal correlation, specifically temporal noise model, in order to alleviate its training load. In contrast to the vanilla noise model, which independently samples noise for each frame from a standard Gaussian distribution, our temporal noise model takes into account the inter-frame relationship via time series models. More importantly, to promptly address unexpected weather conditions, we introduce test-time adaptation techniques by recalibrating the model's parameters using only the degraded test data. The mainstream of test-time adaptation methods [18, 40, 58] manipulate and iteratively update the normalization statistics of CNNs and Vision Transformer using specific proxy tasks at the test phase. However, these", + "bbox": [ + 75, + 613, + 473, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "methods often fail when transferring to low-level tasks. Additionally, incorporating such iterative adaptation into these models significantly increases the cost of inference, multiplying it by the number of iterations, compared to the original counterpart.", + "bbox": [ + 496, + 90, + 893, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast, we highlight the advantage of integrating the diffusion reverse process and test-time adaptation based on their common characteristics of iteration. As far as we know, there has been no work to emerge the possibility of Diffusion models and test-time adaptation. Based on our diffusion-based framework, we design a proxy task and efficiently incorporate it into each denoising iteration when reversing our temporal diffusion process. Specifically, we develop a surrogate task called Diffusion Tubelet Self-Calibration (Diff-TSC), which is supervised by temporal noise model. This is designed to be task-agnostic and does not require any alterations to the training procedure. Rather than directly traversing all fine-grained details, we extract a few tubelets to taste the primer distribution of the frames degraded by unknown weather. This approach facilitates rapid optimization of the diffusion model while also reducing computational costs. As displayed in Figure 2, test-time adaptation helps to improve the model's robustness in real-world scenarios by mitigating the gap between seen and unseen weather corruptions.", + "bbox": [ + 496, + 167, + 893, + 468 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as:", + "bbox": [ + 500, + 469, + 772, + 482 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce the first diffusion-based framework for all-in-one adverse weather removal in videos. This framework efficiently leverages temporal redundancy information through the novel temporal noise model.", + "- To improve the model's robustness against unknown weather, we are the first to introduce test-time adaptation by incorporating a proxy task into the diffusion reverse process to learn the primer distribution of the test data.", + "- We extensively evaluate our approach on seen weather (e.g., rain, haze, snow) and unseen weather (e.g., rain streak+raindrop, snow+fog). The superior performance over synthesized and real-world videos consistently validates its effectiveness and generalizable ability." + ], + "bbox": [ + 500, + 483, + 890, + 679 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 693, + 640, + 709 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "All-in-one Adverse Weather Removal. Over the past decade, The community of low-level vision has paid much attention to single weather removal, e.g., deraining [5, 20, 24, 42, 47, 49, 50, 53], dehazing [28, 35, 46, 59], desnowing [4, 7-9, 56], both on the image level and the video level. Due to the limitations of existing models in adapting to different weather conditions, researchers have recently been investigating the use of a single model instance for multiple adverse weather removal tasks. At the image level, following the pioneering work All-in-one [21], TransWeather [38] and TKL [11] were introduced to tackle many weather conditions with the single encoder and de", + "bbox": [ + 496, + 719, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "25607", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "coder. Zhu et al. [62] elaborated a two-stage training strategy to explore the weather-general and weather-specific information for the sophisticated removal of diverse weather conditions. Ye et al. [52] introduced high-quality codebook priors learned by a pre-trained VQGAN to restore texture details from adverse weather degradation. The aforementioned methods failed to capture redundant information from temporal space. While they can be adapted for frame-by-frame adverse weather removal, incorporating temporal data in video-level methods provides a more robust solution for dealing with multiple adverse weather conditions in videos. Very recently, Yang et al. [51] introduced 'ViWS-Net', the first video-level approach. This method uses weather messenger tokens to transfer weather knowledge across frames. It also employs adversarial learning to distinguish the weather type, thereby emphasizing weather-invariant information. However, they did not consider online adaptation of the deep model, making the potential collapse for unseen scenarios degraded by unknown weather.", + "bbox": [ + 76, + 90, + 472, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion Models. Diffusion model is well-known as a generative modeling approach for its superior achievements in image synthesis and generation tasks [14, 16, 27, 36, 57]. In essence, Denoising Diffusion Probabilistic Models (DDPM) [16] adopted parameterized Markov chain to optimize the lower variational bound on the likelihood function, which can simulate a diffusion process to iteratively improve the quality of target distribution than other generative models. For restoration tasks, Luo et al. [30] applied stochastic differential equations featuring mean reversion in diffusion models to improve low-quality images. DiffIR [45] provided a strong diffusion baseline for image restoration by estimating a compact IR prior representation. Zhang et al. [60] proposed a unified conditional framework based on diffusion models for image restoration by multi-source guidance of a lightweight U-Net. Ozdenizci et al. [33] designed a patch-based diffusion modeling approach called WeatherDiffusion to make the network size-agnostic. While this technique utilizes diffusion models for image weather removal, its very slow inference speed restricts practical use. Nevertheless, there is no work adapting diffusion models to video adverse weather removal.", + "bbox": [ + 76, + 382, + 472, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Test-time Adaptation. Recently, there has been a growing interest in adapting deep models using exclusively sequential, unlabeled test data. This contrasts with the traditional unsupervised domain adaptation, which asks the access to both training data and a sufficient amount of unlabeled test data. Test-time adaptation techniques allow for model updates using distributional statistics from either a single instance or batch of unlabeled test data. Typically, TENT [40] focuses on minimizing the model's prediction entropy on the target data. T3A [18] adjusts the classifier of a trained source model by computing a pseudo-prototype representation of different classes using unlabeled test data.", + "bbox": [ + 76, + 719, + 472, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "MEMO [58] aims to minimize marginal entropy, compelling the deep model to consistently predict across various augmentations. This approach enforces the invariances of these augmentations and ensures prediction confidence. Chen et al. [3] proposed to apply self-supervised contrastive learning for target feature learning, which was trained jointly with pseudo labeling. Although test-time adaptation methods are increasingly used in image classification [1, 31, 32, 41, 55] and segmentation [17, 25, 48] tasks, their potential in adverse weather removal tasks remains unexplored. In our work, the nature of the iterative sampling process in Diffusion model significantly promotes the adaptation to the target distribution. Video scenarios further allow for the accumulation of the adaptation knowledge for the subsequent frames.", + "bbox": [ + 496, + 90, + 893, + 318 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries: Diffusion Models", + "text_level": 1, + "bbox": [ + 498, + 327, + 794, + 343 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The diffusion model is composed of a forward and a reverse process. The forward process is defined as a discrete Markov chain of length $T$ : $q(x_{1:T}|x_0) = \\prod_{t=1}^{T} q(x_t|x_{t-1})$ . For each step $t \\in [1,T]$ in the forward process, a diffusion model adds noise $\\epsilon_t$ sampled from the Gaussian distribution $\\mathcal{N}(0,\\mathbf{I})$ to data $x_{t-1}$ and obtains disturbed data $x_t$ from $q(x_t|x_{t-1}) = \\mathcal{N}(x_t; \\sqrt{1 - \\beta_t} x_{t-1}, \\beta_t\\mathbf{I})$ . $\\beta_t$ decides the ratio of noise at timestep $t$ . Noticeably, instead of sampling sequentially along the Markov chain, we can sample $x_t$ at any time step $t$ in the closed form via $q(x_t|x_0) = \\mathcal{N}(x_t; \\sqrt{\\bar{\\alpha}_t} x_{t-1}, (1 - \\bar{\\alpha}_t)\\mathbf{I})$ , where $\\bar{\\alpha}_t = \\prod_{s=1}^t (1 - \\beta_s)$ . To parameterize Gaussian distribution, the neural network $\\epsilon_\\theta$ is introduced and optimized by the objective of DDPM [16]:", + "bbox": [ + 496, + 352, + 893, + 551 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {S} = \\left\\| \\epsilon - \\epsilon_ {\\theta} \\left(\\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , t\\right) \\right\\| _ {2} ^ {2}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 560, + 890, + 579 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the reverse process, the diffusion model gradually de-noises the randomly sampled Gaussian noise to the high-quality output $x_0$ through the predicted noise by the well-trained $\\epsilon_{\\theta}$ . This process is also defined as a Markov chain: $p_{\\theta}(x_{0:T}) = p(x_T)\\prod_{t=1}^{T}p_{\\theta}(x_{t-1}|x_t)$ , and $p_{\\theta}(x_{t-1}|x_t) = \\mathcal{N}(x_{t-1};\\mu_{\\theta}(x_t,t),\\sigma_t^2\\mathbf{I})$ . The mean and variance are $\\mu_{\\theta}(x_t,t) = \\frac{1}{\\sqrt{\\alpha_t}} (x_t - \\frac{\\beta_t}{\\sqrt{1 - \\bar{\\alpha}_t}}\\epsilon_{\\theta}(x_t,t))$ , $\\sigma_t^2 = \\frac{1 - \\bar{\\alpha}_{t-1}}{1 - \\bar{\\alpha}_t}\\beta_t$ , respectively.", + "bbox": [ + 496, + 590, + 893, + 714 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 500, + 726, + 591, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we devise the first diffusion-based model to remove arbitrary adverse weather in videos with one set of model parameters. We follow an end-to-end formulation of video adverse weather removal, which reads:", + "bbox": [ + 496, + 752, + 893, + 811 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {V}} _ {k} = \\phi (\\mathbf {V} _ {k}), \\mathbf {V} _ {k} = \\left\\{I _ {i} \\right\\} _ {i = 1} ^ {N _ {f}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 821, + 890, + 844 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{V}_k$ is the $k$ -th video clip with $N_{f}$ frames degraded by arbitrary weather type, $\\hat{\\mathbf{V}}_k$ is the restored clip. $\\phi$ is the denoising function derived from DDIM [37]. The denoising", + "bbox": [ + 496, + 852, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "25608", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0e1aa3868e40cc558991bf5c6e8e6a82e4bcfe43ca3e06ca7c240e30f341f7cc.jpg", + "image_caption": [ + "Figure 3. Overview of our Diff-TTA framework for video adverse weather removal. In the training stage, through the temporal diffusion process, we randomly add temporal noise on the clean video clip $\\mathbf{V}_{gt}$ from the mixed training set. Then, the denoising NAFNet $\\epsilon_{\\theta}$ is trained by estimating the applied noise, which is conditioned by the degraded counterpart $\\mathbf{V}_{lq}$ by seen weather types. In the test stage, the proxy task, Diff-TSC, where the cropped tubelets of the last restored pair $\\{\\mathbf{V}_{k-1}, \\hat{\\mathbf{V}}_{k-1}\\}$ are utilized to learn the primer distribution, is incorporated into each timestep of denoising the randomly sampled temporal noise $\\mathbf{V}_k^T$ for the iterative online adaptation." + ], + "image_footnote": [], + "bbox": [ + 106, + 89, + 867, + 441 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "network $\\epsilon_{\\theta}$ in the denoising function consists of a single encoder and decoder.", + "bbox": [ + 75, + 532, + 468, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the training stage, we follow the formulation of DDIM and extend the diffusion forward process to a temporal counterpart for frame-correlation modeling. The denoising network $\\epsilon_{\\theta}$ with NAFNet [6] as the backbone is trained for noise estimation, which enforces the output of the model to approximate the proposed temporal noise model. During the inference stage, we present a proxy task, Diff-TSC, to adapt the diffusion model to the target test data degraded by unknown weather based on the diffusion reverse process. We iteratively optimize the noise estimation of the cropped tubelets and calibrate the weight set of the model parameters for more robust weather removal. Next, we elaborate on our diffusion-based solution for adverse weather removal in videos and our diffusion test-time adaptation, which are also displayed in Figure 3.", + "bbox": [ + 75, + 564, + 470, + 790 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Temporal Diffusion Process", + "text_level": 1, + "bbox": [ + 76, + 801, + 326, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The traditional diffusion model is trained to denoise independent noise from a perturbed image. The noise $\\epsilon$ in Eq. (1) is sampled from an i.i.d. Gaussian distribution $\\epsilon \\sim \\mathcal{N}(0, \\mathbf{I})$ . Upon training the image diffusion model and utilizing it to reverse frames from a video into the noise", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "space individually, the significant correlation among the noise maps of these frames can be easily observed. Rather than introducing massive parameters into the denoising network, we extend the original diffusion process to the temporal counterpart by an ARMA-based temporal noise model to efficiently preserve the correlation between frames.", + "bbox": [ + 496, + 532, + 890, + 623 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Temporal noise model. ARMA (Auto-regressive Moving Average) [2] is a widely used statistical model for time series data, which combines the auto-regressive (AR) part and the moving average (MA) part. Typically, ARMA can be formulated as:", + "bbox": [ + 496, + 623, + 890, + 698 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nX _ {t} = c + \\varepsilon_ {t} + \\sum_ {i = 1} ^ {p} \\varphi_ {i} X _ {t - i} + \\sum_ {j = 1} ^ {q} \\tau_ {j} \\varepsilon_ {t - j}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 707, + 890, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $X$ is the value term, $\\varepsilon$ is the error term, $c$ is a constant term, $p,q$ are the auto-regressive order and moving average order while $\\varphi ,\\tau$ are their coefficients, respectively. Motivated by ARMA, we develop the original Gaussian noise model $\\mathcal{N}(0,\\mathbf{I})$ to the temporal noise model $\\mathcal{T}(0,\\mathbf{I})$ . Specifically, let $\\{\\bar{\\epsilon}_i\\}_{i = 1}^{N_f}$ denote the noise sequence corresponding to consecutive frames of a video clip. We first sample the noise sequence $\\{\\bar{\\epsilon}_i\\}_{i = 1}^{N_f}$ and the corresponding error sequence $\\{\\varepsilon_i\\}_{i = 1}^{N_f}$ from $\\mathcal{N}(0,\\mathbf{I})$ , respectively. The constant", + "bbox": [ + 496, + 758, + 892, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "25609", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "term $c$ should be consistent with the mean of Gaussian distribution, that is, zero. To take both the past and future information into consideration, we enforce the noise of the current frame affected not only by the noise and error term of the previous frame but also those of the next frame, and correspondingly the auto-regressive order and moving average order $p, q = 1$ . Consequently, the expression in Eq. (3) evolves into the convex combination of the noise term and error term:", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\epsilon} _ {i} = (1 - \\varphi - \\tau) \\varepsilon_ {i} + \\varphi \\frac {\\bar {\\epsilon} _ {i - 1} + \\bar {\\epsilon} _ {i + 1}}{2} + \\tau \\frac {\\varepsilon_ {i - 1} + \\varepsilon_ {i + 1}}{2}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 83, + 229, + 468, + 260 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\varphi, \\tau$ are empirically set as 0.6, 0.3, respectively. Algorithm 1 further illustrates our temporal noise model, where the noise sequence of each video clip is constructed with separate preorder and postorder steps.", + "bbox": [ + 76, + 265, + 468, + 327 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: Temporal Noise Model $\\mathcal{T}(0,\\mathbf{I})$" + ], + "code_body": "Input: $K$ clips: $\\{\\mathbf{V}_1,\\mathbf{V}_2,\\dots ,\\mathbf{V}_K\\}$ in a batch; The coefficients $\\varphi ,\\tau$ \nfor $k = 1,2,\\ldots ,K$ parallelly do \nfor Frame $i = 1,2,\\dots ,N_{f}$ parallelly do \nsample $\\bar{\\epsilon}_i,\\varepsilon_i\\in \\mathbb{R}^{1\\times c\\times h\\times w}$ from $\\mathcal{N}(0,\\mathbf{I})$ \nend \nfor Frame $i = 2,3,\\dots ,N_{f}$ sequentially do \n $\\bar{\\epsilon}_i = (1 - \\varphi -\\tau)\\varepsilon_i + \\varphi \\bar{\\epsilon}_{i - 1} + \\tau \\varepsilon_{i - 1}$ \nend \nfor Frame $i = N_{f} - 1,\\dots ,2,1$ sequentially do \n $\\bar{\\epsilon}_i = (1 - \\varphi -\\tau)\\varepsilon_i + \\varphi \\bar{\\epsilon}_{i + 1} + \\tau \\varepsilon_{i + 1}$ \nend \n $\\bar{\\epsilon} =$ normalize( $\\bar{\\epsilon})\\Rightarrow \\bar{\\epsilon}\\sim \\mathcal{T}(0,\\mathbf{I})$ \nend \nReturn $\\{\\{\\bar{\\epsilon}_i|\\bar{\\epsilon}_i\\in \\mathbb{R}^{1\\times c\\times h\\times w}\\}_{i = 1}^{N_f}\\}_{k = 1}^K$", + "bbox": [ + 78, + 354, + 434, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To enhance the capacity of our denoising network for capturing temporal dependencies, we have specifically modified the architecture by substituting only the first and last layers of the denoising NAFNet with the 3D convolutional layer. This tailored network, denoted as $\\epsilon_{\\theta}$ is then optimized as:", + "bbox": [ + 75, + 609, + 468, + 700 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {S} = \\left\\| \\bar {\\epsilon} - \\epsilon_ {\\theta} \\left(\\sqrt {\\bar {\\alpha} _ {t}} \\hat {\\mathbf {V}} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\bar {\\epsilon}, \\mathbf {V}, t\\right) \\right\\| _ {1}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 707, + 468, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where we utilize the temporal noise model as the supervision, and $\\| \\cdot \\|$ denotes $L_{1}$ norm.", + "bbox": [ + 75, + 733, + 468, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Diffusion Test-time Adaptation", + "text_level": 1, + "bbox": [ + 76, + 771, + 352, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Once trained by the proposed temporal noise model, the network $\\epsilon_{\\theta}$ can generate high-quality images. This is achieved by sampling a noisy state $\\mathbf{V}_T$ and then iteratively denoising it in the reverse process. Yang et al. [51] introduced adversarial learning for weather conditions, which require weather type annotations during the training stage. Additionally, their method may fail to address out-of-distribution", + "bbox": [ + 75, + 794, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "data, because including all real-world conditions in training, considering both weather types and image statistics, is impractical. Contrastively, we introduce test-time adaptation and update knowledge by practice to remove arbitrary weather conditions robustly. We design and incorporate Diffusion Tubelet Self-Calibration (Diff-TSC), a proxy task, into the iterative denoising of the diffusion reverse process to learn the primer distribution of the frames degraded by unknown weather.", + "bbox": [ + 496, + 90, + 893, + 227 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 2: Diffusion Test-Time Adaptation to unknown weather. $\\theta$ is the weight set of the trained network before adaptation, $\\delta$ is the learning rate for online adaptation." + ], + "code_body": "Input: $K$ overlapped clips: $\\{\\mathbf{V}_1,\\mathbf{V}_2,\\dots ,\\mathbf{V}_K\\}$ in one video stream \nfor $Clp k = 1,2,\\ldots ,K$ sequentially do \nInitialize the network $\\epsilon_{\\bar{\\theta}}$ with $\\theta$ \nInitialize $\\hat{\\mathbf{V}}_k$ by Algorithm 1 \nif $k = 1$ then \nfor step $t = T,\\dots ,2,1$ sequentially do \n $|\\hat{\\mathbf{V}}_k = ddim(\\mathbf{V}_k,\\hat{\\mathbf{V}}_k,\\epsilon_{\\bar{\\theta}},t)$ \nend \nelse \nA, $\\hat{\\mathbf{A}} =$ TubeletCrop(Vk-1,Vk-1,Na) for step $t = T,\\dots ,2,1$ sequentially do \nCompute $\\mathcal{L}_S(\\mathbf{A},\\hat{\\mathbf{A}},\\bar{\\epsilon},t)$ by Eq. (5) $\\bar{\\theta} = \\bar{\\theta} -\\delta \\nabla_{\\bar{\\theta}}\\mathcal{L}_{S}$ $\\hat{\\mathbf{V}}_k = ddim(\\mathbf{V}_k,\\hat{\\mathbf{V}}_k,\\epsilon_{\\bar{\\theta}},t)$ \nend \n $\\hat{\\mathbf{V}}_k =$ integrate( $\\hat{\\mathbf{V}}_k,\\hat{\\mathbf{V}}_{k - 1})$ \nend \nReturn the restored clips $\\{\\hat{\\mathbf{V}}_1,\\hat{\\mathbf{V}}_2,\\dots ,\\hat{\\mathbf{V}}_K\\}$", + "bbox": [ + 500, + 299, + 851, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Diffusion Tubelet Self-Calibration. We design a proxy task for self-supervised learning to transfer the diffusion model to the specific weather condition. Specifically, for each clip from the video stream, we first initialize the weight set of $\\epsilon_{\\bar{\\theta}}$ with $\\theta$ trained by source data. We leverage the previous clip $\\mathbf{V}_{k - 1}$ and its restored result $\\hat{\\mathbf{V}}_{k - 1}$ as the pair of condition and input for the online optimization when inferring the current clip $\\mathbf{V}_k$ . Rather than directly traversing all fine-grained details, we randomly crop $N_{a}$ tubelets from $\\mathbf{V}_{k - 1}$ as the primer distribution to avoid memory overhead and massive computational costs caused by high resolution. Following the process in Section 4.1, we conduct our temporal diffusion process on the restored tubelets $\\hat{\\mathbf{A}}$ and feed them together with the degraded tubelets $\\mathbf{A}$ into the denoising NAFNet for noise estimation. As suggested in [40], it is necessary to iteratively find a relatively optimal point for the target distribution during test-time adaptation. Inher", + "bbox": [ + 496, + 643, + 893, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "25610", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/912e1ff711fe7cf1878fe7782acf69fc168e6b017581369e25df42d223799ce8.jpg", + "table_caption": [ + "Table 1. Quantitative evaluation on seen weather conditions for video adverse weather removal. For Original Weather, these methods are trained on the weather-specific training set and tested on the weather-specific testing set. For Rain, Haze, and Snow, these methods are trained on a mixed training set and tested on the weather-specific testing set. The average performance is calculated on Rain, Haze, and Snow. PSNR and SSIM are adopted as our evaluation metrics. The best and second-best values are bold and underlined." + ], + "table_footnote": [], + "table_body": "
MethodsTypeSourceDatasets
Original WeatherRainHazeSnowAverage
DerainPReNet [34]ImageCVPR'1927.060.907726.800.881417.640.803028.570.940124.340.8748
SLDNet [50]VideoCVPR'2020.310.627221.240.712916.210.756122.010.855019.820.7747
S2VD [53]VideoCVPR'2124.090.794428.390.900619.650.860726.230.919024.760.8934
RDD-Net [42]VideoECCV'2231.820.942330.340.930018.360.843230.400.956026.370.9097
DehazeGDN [26]ImageICCV'1919.690.854529.960.937019.010.880531.020.951826.660.9231
MSBDN [15]ImageCVPR'2022.010.875926.700.914622.240.904727.070.934025.340.9178
VDHNet [35]VideoTIP'1916.640.813329.870.927216.850.821429.530.939525.420.8960
PM-Net [28]VideoMM'2223.830.895025.790.888023.570.914318.710.788122.690.8635
DesnowDesnowNet [29]ImageTIP'1828.300.953025.190.878616.430.790227.560.918123.060.8623
DDMSNET [56]ImageTIP'2132.550.961329.010.918819.500.861532.430.969426.980.9166
HDCW-W-Net [10]ImageICCV'2131.770.954228.100.905517.360.792131.050.948225.500.8819
SMGARN [13]ImageTCSVT'2233.240.972127.780.910017.850.807532.340.966825.990.8948
RestorationMPRNet [54]ImageCVPR'21————28.220.916520.250.893430.950.948226.470.9194
EDVR [43]VideoCVPR'19————31.100.937119.670.872430.270.944027.010.9178
RVRT [23]VideoNIPS'22————30.110.913221.160.894926.780.883426.020.8972
RTA [61]VideoCVPR'22————30.120.918620.750.891529.790.936726.890.9156
Multi-WeatherAll-in-one [21]ImageCVPR'20————26.620.894820.880.901030.090.943125.860.9130
UVRNet [19]ImageTMM'22————22.310.767820.820.857524.710.887322.610.8375
TransWeather [38]ImageCVPR'22————26.820.911822.170.902528.870.931325.950.9152
TKL [11]ImageCVPR'22————26.730.893522.080.904431.350.951526.720.9165
WeatherDiffusion [33]ImageTPAMI'23————25.860.912520.100.844226.400.911324.120.8893
WGWS-Net [62]ImageCVPR'23————29,640.931017.710.811331.580.952826.310.9265
ViWS-Net [51]VideoICCV'23————31.520.943324.510.918731.490.956229.170.9394
OursVideo——————32.430.957324.560.914831.860.964029.630.9453
", + "bbox": [ + 114, + 148, + 846, + 425 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ited from the characteristics of DDIM, such requirement can be easily achieved in the iterative diffusion reverse process, which needs more increments of inference time in the traditional CNNs and Transformers. For each timestep $t$ , we first obtain the noisy tubelet $\\hat{\\mathbf{A}}_t$ and then optimize the objective function given in Eq. (5) to update the weight set $\\bar{\\theta}$ . After the iterative adaptation to $\\hat{\\mathbf{V}}_k$ , we ensemble the overlapped part of it with that of the previous clip $\\hat{\\mathbf{V}}_{k - 1}$ to capture complementary information of pre-adaptation and post-adaptation. We optimize over the entire network except for the last layer suggested by SHOT [22]. The detailed procedure is described in Algorithm 2. We emphasize the synergy between diffusion models and test-time adaptation in accumulating frame-correlated knowledge for adverse weather removal. By attaching the proxy task to iterative denoising, our approach also avoids excessive computational overhead caused by iterative adaptation.", + "bbox": [ + 75, + 441, + 472, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 713, + 209, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Datasets", + "text_level": 1, + "bbox": [ + 76, + 739, + 179, + 753 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Various video adverse weather datasets are used in our experiments. Following the setting of [51], we adopt RainMotion [42], REVIDE [59] and KITTI-snow [51] as seen weather. For the training stage, we merge the training set of the three datasets as the mixed set to learn a generic model. For the testing stage, we evaluate our model on three testing sets, respectively. Additionally, we evaluate our approach on two datasets VRDS [44] and RVSD [4] to validate the generalization of our Diff-TTA on unseen", + "bbox": [ + 75, + 763, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d15cf52da9f5a6d01254807d850333723605f282c784470381d19dee6dd05fa4.jpg", + "table_caption": [ + "Table 2. Quantitative evaluation on unseen weather conditions for video adverse weather removal. The best and second-best performances are marked in bold and underlined, respectively." + ], + "table_footnote": [], + "table_body": "
MethodVRDS [44]RVSD [4]
PSNR ↑SSIM ↑PSNR ↑SSIM ↑
All-in-one [21]20.440.594419.790.7509
TransWeather [38]21.360.713620.250.7514
TKL [11]20.490.700319.710.7370
WeatherDiffusion [33]20.730.694318.080.6588
ViWS-Net [51]21.570.709419.830.7590
Diff-TTA (ours)22.570.728122.350.7719
", + "bbox": [ + 521, + 486, + 870, + 602 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "weather conditions. VRDS is a synthesized video dataset of joint rain streaks and raindrops with a total of 102 videos, while RVSD is a realistic video desnowing dataset with a total of 110 videos containing both snow and fog achieved by the rendering engine. Due to the large quantity of the two datasets, we randomly sample 8 videos as the test set, each of which has 20 frames, respectively. We also collect several real-world videos degraded by diverse weather conditions to prove its effectiveness on real-world applications.", + "bbox": [ + 496, + 611, + 890, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Implementation", + "text_level": 1, + "bbox": [ + 500, + 756, + 660, + 772 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training Details. The proposed framework was trained on NVIDIA RTX 4090 GPUs and implemented on the Pytorch platform. We adopted the Lion optimizer [12] ( $\\beta_{1} = 0.9$ , $\\beta_{2} = 0.99$ ) with the initial learning rate of $2 \\times 10^{-5}$ decayed to $2 \\times 10^{-7}$ by the Cosine scheduler. We randomly crop the video frames to $256 \\times 256$ for training. Our framework is empirically trained for $600k$ iterations in an end-to-end way with batch size of 4. Each batch consists of 4 video", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "25611", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c4ef1525402fa617c8c4fc37c9fc532693fa65f894364c12d715d6feb5a4cd38.jpg", + "image_caption": [ + "Input Frame" + ], + "image_footnote": [], + "bbox": [ + 84, + 90, + 183, + 250 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/367d8b11e4217833f1df827769e1dede565c28cd3cdb7e02e0c3262a33cbafe0.jpg", + "image_caption": [ + "EDVR" + ], + "image_footnote": [], + "bbox": [ + 184, + 90, + 284, + 250 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e4e0670ed6db40c128cae71d709d86e2d8697b607197ba7d2c10c6176427121b.jpg", + "image_caption": [ + "TransWeather" + ], + "image_footnote": [], + "bbox": [ + 285, + 90, + 383, + 250 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/06ec1ec474e98cf6df4aedb1e18c5c14342d37093ee7787396fe35fafa9ec250.jpg", + "image_caption": [ + "TKL", + "Figure 4. Qualitative Comparison of seen weather conditions between our approach and state-of-the-art algorithms. The competitive algorithms are selected to present the results on the example frames degraded by rain, haze, snow, respectively. The color box indicates the comparison of details. Please zoom in on the images for improved visualization." + ], + "image_footnote": [], + "bbox": [ + 385, + 90, + 483, + 250 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4c444ee7d709fc7617d4dbba15a6b83fb2c3c066d0d36d4b7b7ac59759ae87b0.jpg", + "image_caption": [ + "WeatherDiffusion" + ], + "image_footnote": [], + "bbox": [ + 486, + 90, + 584, + 250 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c51639beb9c788c6e37854cb62cc3cc81c46c2cd44bf59f7507afd04bc9f7c1c.jpg", + "image_caption": [ + "ViWS-Net" + ], + "image_footnote": [], + "bbox": [ + 586, + 90, + 684, + 250 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d71a8e9bcaff4b45039c96f8d47d118073fcc90da1686e6952c81ae61a106715.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 687, + 90, + 785, + 250 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/82476afc385bf30fd7b32d0c730b1503ab2b5f09f58730dc8f49414a48683f21.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 787, + 90, + 885, + 250 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "clips with 5 frames per clip.", + "bbox": [ + 76, + 324, + 263, + 338 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Method-specific Details. During the training, we used the diffusion process linearly increasing from $\\beta_{1} = 0.0001$ to $\\beta_{T} = 0.02$ with $T = 1000$ timesteps. During the testing, the inference step is reduced to 25 by deterministic implicit sampling. We utilize the same optimizer with a smaller learning rate of $1\\times 10^{-6}$ for the test-time adaptation. The number of cropped tubelets $N_{a}$ is empirically set to 5 while the crop size is $128\\times 128$ . Despite the incorporated adaptation, our Diff-TTA is particularly $90\\times$ more efficient than WeatherDiffusion. To process a video clip of 5 frames with the patch size of $256\\times 256$ , the average run time of Diff-TTA is 6.01s while that of WeatherDiffusion is 542.76s.", + "bbox": [ + 75, + 339, + 468, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Performance Evaluation", + "text_level": 1, + "bbox": [ + 76, + 530, + 302, + 544 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison methods. As shown in Table 1, we compared our proposed method against five kinds of state-of-the-art methods on the three seen weather conditions as [51] does, i.e., derain, dehaze, desnow, restoration, all-in-one adverse weather removal. For all-in-one adverse weather removal, we compared ours with the six representative single-image methods All-in-one [21], UVRNet [19], TransWeather [38], TKL [11], WeatherDiffusion [33], WGWS-Net [62], and the video-level method ViWS-Net [51].", + "bbox": [ + 75, + 551, + 468, + 688 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Comparison. For quantitative evaluation of the restored results, we apply the peak signal-to-noise ratio (PSNR) and the structural similarity (SSIM) as the metrics. Following the previous experimental setting, only the results of the model trained on the mixed training set are reported for all-in-one adverse weather removal models. To ensure the fairness of comparisons, we retrain each compared model implemented by the official codes based on our training dataset and report the best result. Observed from Table 1, our method achieves the best average performance across rain, haze and snow conditions among all kinds of methods by a considerable margin of 0.46, 0.0059 in PSNR, SSIM, respectively, than the second-best method ViWS-Net [51]. Though some single-weather", + "bbox": [ + 75, + 689, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "methods like DDMSNET [56] and SMGARN [13] achieve impressive results in their own weather condition, they frequently fail to remove other weather conditions, such as haze. Our method significantly surpasses other all-in-one adverse weather methods in rain and snow conditions, and obtains competitive results in haze conditions. It is worth noting that the generalizable ability is critical for such methods, which can be observed from the average performance. To further validate their generalization, we evaluate all-in-on adverse weather removal methods on unseen weather conditions, i.e., rain streak+raindrop and snow+fog, in Table 2. Our Diff-TTA exhibits remarkable superiority over state-of-the-art methods when confronted with these unseen weather conditions.", + "bbox": [ + 496, + 324, + 890, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visual comparisons on synthetic videos. To vividly demonstrate the effectiveness of our approach, Figure 4 shows the qualitative comparison under the rain, haze, and snow conditions between our method and five state-of-the-art methods. Our method demonstrates promising results in terms of visual quality across various weather types. Specifically, when it comes to rain and snow scenarios, our approach significantly reduces the presence of rain streaks and snow particles, outperforming other methods. In hazy scenarios, our method effectively removes residual haze and preserves the clean background with impressive results. More visual results of seen and unseen weather are displayed in the supplementary files.", + "bbox": [ + 496, + 536, + 890, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visual comparisons on real-world videos. To evaluate the universality of our Diff-TTA on real-world applications, we collect several real-world degraded videos by rain, haze, snow from Youtube website, and further compare our approach against other all-in-one adverse weather removal methods. Figure 5 displays the visual results produced by our network and the compared methods on real-world video frames, clearly showcasing our exceptional performance.", + "bbox": [ + 496, + 733, + 890, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation Study. We evaluate the effectiveness of each component including diffusion process, temporal noise model, and Diffusion test-time adaptation (Diff-TTA) as described", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "25612", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/26842280312f7349bfc056a3f08f4e53e497f162b172e0f200b8c8ec39cf9631.jpg", + "table_caption": [ + "Table 3. Ablation study of each critical component in the proposed framework on seen weather conditions. The top values are marked in bold font. \"Diff-TTA\" denote Diffusion Test-time Adaptation." + ], + "table_footnote": [], + "table_body": "
CombinationComponentDatasets
Diffusion ProcessTemporal NoiseDiff-TTARainHazeSnowAverage
M1---29.070.951422.640.893028.790.9350
M2✓--31.550.946623.580.904130.480.9520
M3✓✓-32.100.953024.310.912431.040.9621
Ours✓✓✓32.430.957324.560.914831.860.9640
", + "bbox": [ + 114, + 121, + 844, + 208 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1e0c27c42d2afe3f072b88bae1b21288b97b4be2c831ad7ed84afeac7a982c47.jpg", + "image_caption": [ + "Input Frame" + ], + "image_footnote": [], + "bbox": [ + 101, + 215, + 228, + 383 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/62b1cab1b0b1b1f9ca1d318d6547b8bb4df3ede760bf741977af5eff23bac9a6.jpg", + "image_caption": [ + "TransWeather" + ], + "image_footnote": [], + "bbox": [ + 228, + 215, + 356, + 383 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7951abc98c6172b2b2c0d5eb584958fe0fe1ca24a689d1836c1e229f6f1beb0d.jpg", + "image_caption": [ + "TKL" + ], + "image_footnote": [], + "bbox": [ + 357, + 215, + 483, + 383 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7b6d3a7397f89a7ad381452dae980d8886a93b5e119d01fa335b19497e19b340.jpg", + "image_caption": [ + "WeatherDiffusion" + ], + "image_footnote": [], + "bbox": [ + 483, + 215, + 611, + 383 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/83527c929b785e3c02ead6bed4f5ae637df96fa20edc43fa797b969155f139ab.jpg", + "image_caption": [ + "ViWS-Net" + ], + "image_footnote": [], + "bbox": [ + 611, + 215, + 736, + 383 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0de0f733029bbfbeb11d2510974acd38d2ffe067038064e88bfb1f2cfe8986de.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 736, + 215, + 864, + 383 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f3f626ef84e63794ca748132c2f7d9905f6dd84c482546657559fe53aa77cc67.jpg", + "image_caption": [ + "Figure 5. Visual comparison of all-in-one adverse weather removal methods on the selected real-world video sequences degraded by rain, haze, snow. Apparently, our network can more effectively remove rain streaks, haze, and snowflakes of input video frames than state-of-the-art methods. Please zoom in on the images for improved visualization." + ], + "image_footnote": [], + "bbox": [ + 88, + 453, + 267, + 601 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8aad17c2ef787788a2c858fcce686faa1bc258f89443ca9cdbe8462ac8f82d68.jpg", + "image_caption": [ + "Figure 6. Detailed results on each case of VRDS and RVSD by our proposed Diff-TTA. The proposed Diff-TTA brings consistent improvement across various cases." + ], + "image_footnote": [], + "bbox": [ + 269, + 453, + 455, + 601 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "in Table 3. We present the results trained on the mixed training set and tested on three weather-specific testing sets (rain, haze, snow) in view of PSNR and SSIM. The baseline model M1, i.e., the vanilla NAFNet, achieves the average performance on the three adverse weather datasets of 26.83, 0.9264 in PSNR, SSIM, respectively. M2 is equipped with diffusion process and trained by noise estimation, which advances M1 on the average performance by 1.69, 0.0078 of PSNR, SSIM, respectively. M3 introduces the proposed temporal noise model by replacing the vanilla noise model, and brings about an improvement of 0.63, 0.0083 in average PSNR, SSIM, respectively. Finally, our full model employs diffusion test-time adaptation (Diff-TTA) and gains a critical increase of 0.48, 0.0028 in PSNR, SSIM, respectively, compared to M3. Figure 6 further ablates our iterative adaptation on each case of the two unseen datasets, demonstrat", + "bbox": [ + 75, + 659, + 472, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ing consistent improvements.", + "bbox": [ + 500, + 455, + 694, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 492, + 617, + 507 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper presents Diff-TTA, a diffusion-based test-time adaptation method capable of restoring videos degraded by various, even unknown, weather conditions. The proposed approach involves constructing a diffusion-based network and utilizing a novel temporal noise model to establish correlations among video frames. More importantly, a proxy task Diff-TSC is incorporated into the diffusion reverse process to iteratively update the deep model. This allows for the efficient adaptation of vision systems in real-world scenarios without modifying the source training process, making it easily applicable to a wide range of off-the-shelf diffusion-based models for enhancing weather removal. Experimental results on both synthetic and real-world data, which encompass diverse weather conditions, confirm the effectiveness and generalization capability of the proposed approach. We believe Diff-TTA can be a compelling baseline that shifts the focus of the community from the complicated training process to online test-time adaptation.", + "bbox": [ + 496, + 520, + 893, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment. This work is supported by Guangzhou Municipal Science and Technology Project (Grant No. 2023A03J0671), Guangzhou-HKUST(GZ) Joint Funding Program (No. 2024A03J0618), a grant of General Research Fund from Hong Kong Research Grants Council (project no. 15218521), and the InnoHK funding launched by Innovation and Technology Commission, Hong Kong SAR.", + "bbox": [ + 496, + 795, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "25613", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Malik Boudiaf, Romain Mueller, Ismail Ben Ayed, and Luca Bertinetto. Parameter-free online test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8344-8353, 2022. 3", + "[2] George Box. Box and jenkins: time series analysis, forecasting and control. In A Very British Affair: Six Britons and the Development of Time Series Analysis During the 20th Century, pages 161-215. Springer, 2013. 4", + "[3] Dian Chen, Dequan Wang, Trevor Darrell, and Sayna Ebrahimi. Contrastive test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 295-305, 2022. 3", + "[4] Haoyu Chen, Jingjing Ren, Jinjin Gu, Hongtao Wu, Xuequan Lu, Haoming Cai, and Lei Zhu. Snow removal in video: A new dataset and a novel method. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13211-13222, 2023. 1, 2, 6", + "[5] Jie Chen, Cheen-Hau Tan, Junhui Hou, Lap-Pui Chau, and He Li. Robust video content alignment and compensation for rain removal in a cnn framework. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6286-6295, 2018. 2", + "[6] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European Conference on Computer Vision, pages 17-33. Springer, 2022. 4", + "[7] Sixiang Chen, Tian Ye, Yun Liu, Jinbin Bai, Haoyu Chen, Yunlong Lin, Jun Shi, and Erkang Chen. CPLformer: Cross-scale prototype learning transformer for image snow removal. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4228-4239, 2023. 2", + "[8] Sixiang Chen, Tian Ye, Chenghao Xue, Haoyu Chen, Yun Liu, Erkang Chen, and Lei Zhu. Uncertainty-driven dynamic degradation perceiving and background modeling for efficient single image desnowing. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4269-4280, 2023.", + "[9] Wei-Ting Chen, Hao-Yu Fang, Jian-Jiun Ding, Cheng-Che Tsai, and Sy-Yen Kuo. Jstasr: Joint size and transparency-aware snow removal algorithm based on modified partial convolution and veiling effect removal. In European Conference on Computer Vision, pages 754-770. Springer, 2020. 2", + "[10] Wei-Ting Chen, Hao-Yu Fang, Cheng-Lin Hsieh, Cheng-Che Tsai, I Chen, Jian-Jiun Ding, Sy-Yen Kuo, et al. All snow removed: Single image desnowing algorithm using hierarchical dual-tree complex wavelet representation and contradict channel loss. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4196-4205, 2021. 6", + "[11] Wei-Ting Chen, Zhi-Kai Huang, Cheng-Che Tsai, Hao-Hsiang Yang, Jian-Jiun Ding, and Sy-Yen Kuo. Learning multiple adverse weather removal via two-stage knowledge learning and multi-contrastive regularization: Toward a unified model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17653–17662, 2022. 1, 2, 6, 7" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Xiangning Chen, Chen Liang, Da Huang, Esteban Real, Kaiyuan Wang, Yao Liu, Hieu Pham, Xuanyi Dong, Thang Luong, Cho-Jui Hsieh, et al. Symbolic discovery of optimization algorithms. arXiv preprint arXiv:2302.06675, 2023.6", + "[13] Bodong Cheng, Juncheng Li, Ying Chen, Shuyi Zhang, and Tieyong Zeng. Snow mask guided adaptive residual network for image snow removal. arXiv preprint arXiv:2207.04754, 2022.6, 7", + "[14] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 1, 3", + "[15] Hang Dong, Jinshan Pan, Lei Xiang, Zhe Hu, Xinyi Zhang, Fei Wang, and Ming-Hsuan Yang. Multi-scale boosted de-hazing network with dense feature fusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2157-2167, 2020. 6", + "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 3", + "[17] Minhao Hu, Tao Song, Yujun Gu, Xiangde Luo, Jieneng Chen, Yinan Chen, Ya Zhang, and Shaoting Zhang. Fully test-time adaptation for image segmentation. In Medical Image Computing and Computer Assisted Intervention—MICCAI 2021: 24th International Conference, Strasbourg, France, September 27–October 1, 2021, Proceedings, Part III 24, pages 251–260. Springer, 2021. 3", + "[18] Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for model-agnostic domain generalization. Advances in Neural Information Processing Systems, 34:2427-2440, 2021. 2, 3", + "[19] Ashutosh Kulkarni, Prashant W Patil, Subrahmanyam Murala, and Sunil Gupta. Unified multi-weather visibility restoration. IEEE Transactions on Multimedia, 2022. 6, 7", + "[20] Minghan Li, Qi Xie, Qian Zhao, Wei Wei, Shuhang Gu, Jing Tao, and Deyu Meng. Video rain streak removal by multiscale convolutional sparse coding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6644-6653, 2018. 1, 2", + "[21] Ruoteng Li, Robby T Tan, and Loong-Fah Cheong. All in one bad weather removal using architectural search. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3175-3185, 2020. 1, 2, 6, 7", + "[22] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In International conference on machine learning, pages 6028-6039. PMLR, 2020. 6", + "[23] Jingyun Liang, Yuchen Fan, Xiaoyu Xiang, Rakesh Ranjan, Eddy Ilg, Simon Green, Jiezhang Cao, Kai Zhang, Radu Timofte, and Luc Van Gool. Recurrent video restoration transformer with guided deformable attention. arXiv preprint arXiv:2206.02146, 2022. 6", + "[24] Jiaying Liu, Wenhan Yang, Shuai Yang, and Zongming Guo. Erase or fill? deep joint recurrent rain removal and reconstruction in videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3233-3242, 2018. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "25614", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Quande Liu, Cheng Chen, Qi Dou, and Pheng-Ann Heng. Single-domain generalization in medical image segmentation via test-time adaptation from shape dictionary. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1756–1764, 2022. 3", + "[26] Xiaohong Liu, Yongrui Ma, Zhihao Shi, and Jun Chen. Grid-dehazenet: Attention-based multi-scale network for image dehazing. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7314-7323, 2019. 6", + "[27] Xihui Liu, Dong Huk Park, Samaneh Azadi, Gong Zhang, Arman Chopikyan, Yuxiao Hu, Humphrey Shi, Anna Rohrbach, and Trevor Darrell. More control for free! image synthesis with semantic diffusion guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 289-299, 2023. 3", + "[28] Ye Liu, Liang Wan, Huazhu Fu, Jing Qin, and Lei Zhu. Phase-based memory network for video dehazing. In Proceedings of the 30th ACM International Conference on Multimedia, pages 5427-5435, 2022. 2, 6", + "[29] Yun-Fu Liu, Da-Wei Jaw, Shih-Chia Huang, and Jenq-Neng Hwang. Desnownet: Context-aware deep network for snow removal. IEEE Transactions on Image Processing, 27(6): 3064–3073, 2018. 6", + "[30] Ziwei Luo, Fredrik K Gustafsson, Zheng Zhao, Jens Sjolund, and Thomas B Schon. Image restoration with meanreverting stochastic differential equations. arXiv preprint arXiv:2301.11699, 2023. 3", + "[31] A Tuan Nguyen, Thanh Nguyen-Tang, Ser-Nam Lim, and Philip HS Torr. Tipi: Test time adaptation with transformation invariance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24162-24171, 2023. 3", + "[32] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yafo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. In International conference on machine learning, pages 16888-16905. PMLR, 2022. 3", + "[33] Ozan Özdenizci and Robert Legenstein. Restoring vision in adverse weather conditions with patch-based denoising diffusion models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1, 3, 6, 7", + "[34] Dongwei Ren, Wangmeng Zuo, Qinghua Hu, Pengfei Zhu, and Deyu Meng. Progressive image deraining networks: A better and simpler baseline. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3937-3946, 2019. 6", + "[35] Wenqi Ren, Jingang Zhang, Xiangyu Xu, Lin Ma, Xiaochun Cao, Gaofeng Meng, and Wei Liu. Deep video dehazing with semantic segmentation. IEEE transactions on image processing, 28(4):1895-1908, 2018. 2, 6", + "[36] Jaskirat Singh, Stephen Gould, and Liang Zheng. High-fidelity guided image synthesis with latent diffusion models. arXiv preprint arXiv:2211.17084, 2022. 1, 3", + "[37] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2, 3" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Jeya Maria Jose Valanarasu, Rajeev Yasarla, and Vishal M Patel. Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2353-2363, 2022. 1, 2, 6, 7", + "[39] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 2", + "[40] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726, 2020. 2, 3, 5", + "[41] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7201-7211, 2022. 3", + "[42] Shuai Wang, Lei Zhu, Huazhu Fu, Jing Qin, Carola-Bibiane Schonlieb, Wei Feng, and Song Wang. Rethinking video rain streak removal: A new synthesis model and a deraining network with video rain prior. In European Conference on Computer Vision, pages 565-582. Springer, 2022. 2, 6", + "[43] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced deformable convolutional networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0-0, 2019. 6", + "[44] Hongtao Wu, Yijun Yang, Haoyu Chen, Jingjing Ren, and Lei Zhu. Mask-guided progressive network for joint raindrop and rain streak removal in videos. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7216-7225, 2023. 6", + "[45] Bin Xia, Yulun Zhang, Shiyin Wang, Yitong Wang, Xinglong Wu, Yapeng Tian, Wenming Yang, and Luc Van Gool. Diffir: Efficient diffusion model for image restoration. arXiv preprint arXiv:2303.09472, 2023. 3", + "[46] Jiaqi Xu, Xiaowei Hu, Lei Zhu, Qi Dou, Jifeng Dai, Yu Qiao, and Pheng-Ann Heng. Video dehazing via a multi-range temporal alignment network with physical prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18053–18062, 2023. 2", + "[47] Wending Yan, Robby T Tan, Wenhan Yang, and Dengxin Dai. Self-aligned video deraining with transmission-depth consistency. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11966-11976, 2021. 2", + "[48] Hongzheng Yang, Cheng Chen, Meirui Jiang, Quande Liu, Jianfeng Cao, Pheng Ann Heng, and Qi Dou. Dltta: Dynamic learning rate for test-time adaptation on cross-domain medical images. IEEE Transactions on Medical Imaging, 41 (12):3575-3586, 2022. 3", + "[49] Wenhan Yang, Jiaying Liu, and Jiashi Feng. Frameconsistent recurrent video deraining with dual-level flow. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1661-1670, 2019. 2", + "[50] Wenhan Yang, Robby T Tan, Shiqi Wang, and Jiaying Liu. Self-learning video rain streak removal: When cyclic consistency meets temporal correspondence. In Proceedings of" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "25615", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1720-1729, 2020. 2, 6", + "[51] Yijun Yang, Angelica I Aviles-Rivero, Huazhu Fu, Ye Liu, Weiming Wang, and Lei Zhu. Video adverse-weather-component suppression network via weather messenger and adversarial backpropagation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13200-13210, 2023. 1, 3, 5, 6, 7", + "[52] Tian Ye, Sixiang Chen, Jinbin Bai, Jun Shi, Chenghao Xue, Jingxia Jiang, Junjie Yin, Erkang Chen, and Yun Liu. Adverse weather removal with codebook priors. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12653-12664, 2023. 3", + "[53] Zongsheng Yue, Jianwen Xie, Qian Zhao, and Deyu Meng. Semi-supervised video deraining with dynamical rain generator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 642-652, 2021. 2, 6", + "[54] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14821-14831, 2021. 6", + "[55] Jian Zhang, Lei Qi, Yinghuan Shi, and Yang Gao. Domain adaptor: A novel approach to test-time adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18971-18981, 2023. 3", + "[56] Kaihao Zhang, Rongqing Li, Yanjiang Yu, Wenhan Luo, and Changsheng Li. Deep dense multi-scale network for snow removal using semantic and depth priors. IEEE Transactions on Image Processing, 30:7419-7431, 2021. 2, 6, 7", + "[57] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 3", + "[58] Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test time robustness via adaptation and augmentation. Advances in Neural Information Processing Systems, 35: 38629-38642, 2022. 2, 3", + "[59] Xinyi Zhang, Hang Dong, Jinshan Pan, Chao Zhu, Ying Tai, Chengjie Wang, Jilin Li, Feiyue Huang, and Fei Wang. Learning to restore hazy video: A new real-world dataset and a new method. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9239-9248, 2021. 1, 2, 6", + "[60] Yi Zhang, Xiaoyu Shi, Dasong Li, Xiaogang Wang, Jian Wang, and Hongsheng Li. A unified conditional framework for diffusion-based image restoration. arXiv preprint arXiv:2305.20049, 2023. 3", + "[61] Kun Zhou, Wenbo Li, Liying Lu, Xiaoguang Han, and Jiangbo Lu. Revisiting temporal alignment for video restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6053-6062, 2022. 6", + "[62] Yurui Zhu, Tianyu Wang, Xueyang Fu, Xuanyu Yang, Xin Guo, Jifeng Dai, Yu Qiao, and Xiaowei Hu. Learning weather-general and weather-specific features for image" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "restoration under multiple adverse weather conditions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21747-21758, 2023. 3, 6, 7", + "bbox": [ + 532, + 90, + 890, + 146 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "25616", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Genuine Knowledge from Practice_ Diffusion Test-Time Adaptation for Video Adverse Weather Removal/254c789f-0e30-45ba-9906-00c14a08b4b6_model.json b/2024/Genuine Knowledge from Practice_ Diffusion Test-Time Adaptation for Video Adverse Weather Removal/254c789f-0e30-45ba-9906-00c14a08b4b6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..dac1b21b25e6402f8ff59ef691be521fb49078f3 --- /dev/null +++ b/2024/Genuine Knowledge from Practice_ Diffusion Test-Time Adaptation for Video Adverse Weather Removal/254c789f-0e30-45ba-9906-00c14a08b4b6_model.json @@ -0,0 +1,2290 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.137, + 0.131, + 0.834, + 0.177 + ], + "angle": 0, + "content": "Genuine Knowledge from Practice: Diffusion Test-Time Adaptation for Video Adverse Weather Removal" + }, + { + "type": "text", + "bbox": [ + 0.12, + 0.203, + 0.856, + 0.223 + ], + "angle": 0, + "content": "Yijun Yang\\(^{1}\\), Hongtao Wu\\(^{1}\\), Angelica I. Aviles-Rivero\\(^{2}\\), Yulun Zhang\\(^{3}\\), Jing Qin\\(^{4}\\), Lei Zhu\\(^{1,5*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.184, + 0.224, + 0.786, + 0.239 + ], + "angle": 0, + "content": "1The Hong Kong University of Science and Technology (Guangzhou) 2University of Cambridge 3ETH Zürich" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.242, + 0.751, + 0.257 + ], + "angle": 0, + "content": "4The Hong Kong Polytechnic University 5The Hong Kong University of Science and Technology" + }, + { + "type": "text", + "bbox": [ + 0.244, + 0.258, + 0.725, + 0.273 + ], + "angle": 0, + "content": "Project page: https://github.com/scott-yjyang/DiffTTA" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.308, + 0.314, + 0.323 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.34, + 0.473, + 0.763 + ], + "angle": 0, + "content": "Real-world vision tasks frequently suffer from the appearance of unexpected adverse weather conditions, including rain, haze, snow, and raindrops. In the last decade, convolutional neural networks and vision transformers have yielded outstanding results in single-weather video removal. However, due to the absence of appropriate adaptation, most of them fail to generalize to other weather conditions. Although ViWS-Net is proposed to remove adverse weather conditions in videos with a single set of pretrained weights, it is seriously blinded by seen weather at train-time and degenerates when coming to unseen weather during test-time. In this work, we introduce test-time adaptation into adverse weather removal in videos, and propose the first framework that integrates test-time adaptation into the iterative diffusion reverse process. Specifically, we devise a diffusion-based network with a novel temporal noise model to efficiently explore frame-correlated information in degraded video clips at training stage. During inference stage, we introduce a proxy task named Diffusion Tubelet Self-Calibration to learn the primer distribution of test video stream and optimize the model by approximating the temporal noise model for online adaptation. Experimental results, on benchmark datasets, demonstrate that our Test-Time Adaptation method with Diffusion-based network(Diff-TTA) outperforms state-of-the-art methods in terms of restoring videos degraded by seen weather conditions. Its generalizable capability is validated with unseen weather conditions in synthesized and real-world videos." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.792, + 0.208, + 0.807 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.817, + 0.47, + 0.879 + ], + "angle": 0, + "content": "Adverse weather like rain, snow, and haze is very common in many existing outdoor videos. The presence of adverse weather usually decreases the visibility of the captured videos and significantly impairs the performance of" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.306, + 0.895, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.503, + 0.892, + 0.586 + ], + "angle": 0, + "content": "Figure 1. Overview of the existing all-in-one adverse weather removal methods. Our approach can achieve superior performance not only in seen weather conditions but also in unseen weather conditions with a single set of pre-trained weights by diffusion test-time adaptation. In particular, our Diff-TTA is \\(90 \\times\\) more efficient than WeatherDiffusion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.893, + 0.839 + ], + "angle": 0, + "content": "subsequent high-level vision applications such as object detection, semantic segmentation and autonomous driving. To mitigate these negative effects, Convolutional Neural Networks (CNNs) and Transformers are frequently utilized in the literature. Also, Diffusion models have emerged as the de-facto generative model for image synthesis [14, 16, 36] and have been developed for single-image adverse weather removal [33]. While deep neural networks can achieve decent results on test points within the distribution, real-world applications often demand robustness. This is especially crucial when facing unknown input perturbations, changes in weather conditions, or other sources of distribution shift. Existing single-weather removal approaches [4, 20, 59] require follow-up training to adapt to specific domains. This leads to the switching between different parameter sets, which can make the pipeline cumbersome." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Recently, all-in-one adverse weather removal has attracted the attention of the community, which aims to handle diverse weather conditions with a generic model [11, 21, 38, 51]. However, the potential issue of frequently failing to" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.383, + 0.901 + ], + "angle": 0, + "content": "*Lei Zhu (leizhu@ust.hk) is the corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25606" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.473, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.217, + 0.473, + 0.316 + ], + "angle": 0, + "content": "Figure 2. Our Diff-TTA enables weather removal models to overcome unseen weather corruptions. We use t-SNE [39] to visualize features from the last feature extractor layer of each dataset. Obviously, unseen data points tend to approximate the seen ones after adaptation, which means Diff-TTA can categorize unknown degradation into known distribution. ('Real-world' contains video clips simultaneously degraded by fog and snow.)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.326, + 0.474, + 0.614 + ], + "angle": 0, + "content": "recover from out-of-distribution conditions remains largely unaddressed due to the finiteness of training data and the absence of a test data adaptation mechanism. As shown in Figure 1, generic models suffer significant degradation when handling all the possible weather conditions beyond the scope of their training data. Furthermore, a common aspect among many real-world applications is the necessity for online adaptation with limited data. This is particularly relevant in scenarios such as autonomous vehicles or drones, where a vision model must continually process a video stream and adjust its predictions. To achieve effective adaptation, it is crucial to leverage the incoming data in a way that is beneficial and does not introduce harmful biases, even when there are distribution shifts between the training and test data. Thus, akin to how humans update knowledge through practice, the task of removing any weather condition through online adaptation using a single set of pre-trained weights is both pragmatic and promising, albeit challenging." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.474, + 0.903 + ], + "angle": 0, + "content": "In this paper, we construct the diffusion-based framework for video adverse weather removal. Yet, diffusion models usually entail significant training overhead, hindering the technique's broader adoption in the research community of video tasks. Instead of introducing massive parameters for temporal encoding, we design an efficient strategy based on DDIM [37] to preserve temporal correlation, specifically temporal noise model, in order to alleviate its training load. In contrast to the vanilla noise model, which independently samples noise for each frame from a standard Gaussian distribution, our temporal noise model takes into account the inter-frame relationship via time series models. More importantly, to promptly address unexpected weather conditions, we introduce test-time adaptation techniques by recalibrating the model's parameters using only the degraded test data. The mainstream of test-time adaptation methods [18, 40, 58] manipulate and iteratively update the normalization statistics of CNNs and Vision Transformer using specific proxy tasks at the test phase. However, these" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.168 + ], + "angle": 0, + "content": "methods often fail when transferring to low-level tasks. Additionally, incorporating such iterative adaptation into these models significantly increases the cost of inference, multiplying it by the number of iterations, compared to the original counterpart." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.168, + 0.895, + 0.469 + ], + "angle": 0, + "content": "In contrast, we highlight the advantage of integrating the diffusion reverse process and test-time adaptation based on their common characteristics of iteration. As far as we know, there has been no work to emerge the possibility of Diffusion models and test-time adaptation. Based on our diffusion-based framework, we design a proxy task and efficiently incorporate it into each denoising iteration when reversing our temporal diffusion process. Specifically, we develop a surrogate task called Diffusion Tubelet Self-Calibration (Diff-TSC), which is supervised by temporal noise model. This is designed to be task-agnostic and does not require any alterations to the training procedure. Rather than directly traversing all fine-grained details, we extract a few tubelets to taste the primer distribution of the frames degraded by unknown weather. This approach facilitates rapid optimization of the diffusion model while also reducing computational costs. As displayed in Figure 2, test-time adaptation helps to improve the model's robustness in real-world scenarios by mitigating the gap between seen and unseen weather corruptions." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.47, + 0.774, + 0.483 + ], + "angle": 0, + "content": "Our contributions can be summarized as:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.484, + 0.892, + 0.545 + ], + "angle": 0, + "content": "- We introduce the first diffusion-based framework for all-in-one adverse weather removal in videos. This framework efficiently leverages temporal redundancy information through the novel temporal noise model." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.545, + 0.892, + 0.605 + ], + "angle": 0, + "content": "- To improve the model's robustness against unknown weather, we are the first to introduce test-time adaptation by incorporating a proxy task into the diffusion reverse process to learn the primer distribution of the test data." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.606, + 0.892, + 0.68 + ], + "angle": 0, + "content": "- We extensively evaluate our approach on seen weather (e.g., rain, haze, snow) and unseen weather (e.g., rain streak+raindrop, snow+fog). The superior performance over synthesized and real-world videos consistently validates its effectiveness and generalizable ability." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.484, + 0.892, + 0.68 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.694, + 0.642, + 0.71 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.903 + ], + "angle": 0, + "content": "All-in-one Adverse Weather Removal. Over the past decade, The community of low-level vision has paid much attention to single weather removal, e.g., deraining [5, 20, 24, 42, 47, 49, 50, 53], dehazing [28, 35, 46, 59], desnowing [4, 7-9, 56], both on the image level and the video level. Due to the limitations of existing models in adapting to different weather conditions, researchers have recently been investigating the use of a single model instance for multiple adverse weather removal tasks. At the image level, following the pioneering work All-in-one [21], TransWeather [38] and TKL [11] were introduced to tackle many weather conditions with the single encoder and de" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25607" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.091, + 0.473, + 0.38 + ], + "angle": 0, + "content": "coder. Zhu et al. [62] elaborated a two-stage training strategy to explore the weather-general and weather-specific information for the sophisticated removal of diverse weather conditions. Ye et al. [52] introduced high-quality codebook priors learned by a pre-trained VQGAN to restore texture details from adverse weather degradation. The aforementioned methods failed to capture redundant information from temporal space. While they can be adapted for frame-by-frame adverse weather removal, incorporating temporal data in video-level methods provides a more robust solution for dealing with multiple adverse weather conditions in videos. Very recently, Yang et al. [51] introduced 'ViWS-Net', the first video-level approach. This method uses weather messenger tokens to transfer weather knowledge across frames. It also employs adversarial learning to distinguish the weather type, thereby emphasizing weather-invariant information. However, they did not consider online adaptation of the deep model, making the potential collapse for unseen scenarios degraded by unknown weather." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.383, + 0.473, + 0.716 + ], + "angle": 0, + "content": "Diffusion Models. Diffusion model is well-known as a generative modeling approach for its superior achievements in image synthesis and generation tasks [14, 16, 27, 36, 57]. In essence, Denoising Diffusion Probabilistic Models (DDPM) [16] adopted parameterized Markov chain to optimize the lower variational bound on the likelihood function, which can simulate a diffusion process to iteratively improve the quality of target distribution than other generative models. For restoration tasks, Luo et al. [30] applied stochastic differential equations featuring mean reversion in diffusion models to improve low-quality images. DiffIR [45] provided a strong diffusion baseline for image restoration by estimating a compact IR prior representation. Zhang et al. [60] proposed a unified conditional framework based on diffusion models for image restoration by multi-source guidance of a lightweight U-Net. Ozdenizci et al. [33] designed a patch-based diffusion modeling approach called WeatherDiffusion to make the network size-agnostic. While this technique utilizes diffusion models for image weather removal, its very slow inference speed restricts practical use. Nevertheless, there is no work adapting diffusion models to video adverse weather removal." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.72, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Test-time Adaptation. Recently, there has been a growing interest in adapting deep models using exclusively sequential, unlabeled test data. This contrasts with the traditional unsupervised domain adaptation, which asks the access to both training data and a sufficient amount of unlabeled test data. Test-time adaptation techniques allow for model updates using distributional statistics from either a single instance or batch of unlabeled test data. Typically, TENT [40] focuses on minimizing the model's prediction entropy on the target data. T3A [18] adjusts the classifier of a trained source model by computing a pseudo-prototype representation of different classes using unlabeled test data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.091, + 0.895, + 0.319 + ], + "angle": 0, + "content": "MEMO [58] aims to minimize marginal entropy, compelling the deep model to consistently predict across various augmentations. This approach enforces the invariances of these augmentations and ensures prediction confidence. Chen et al. [3] proposed to apply self-supervised contrastive learning for target feature learning, which was trained jointly with pseudo labeling. Although test-time adaptation methods are increasingly used in image classification [1, 31, 32, 41, 55] and segmentation [17, 25, 48] tasks, their potential in adverse weather removal tasks remains unexplored. In our work, the nature of the iterative sampling process in Diffusion model significantly promotes the adaptation to the target distribution. Video scenarios further allow for the accumulation of the adaptation knowledge for the subsequent frames." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.328, + 0.795, + 0.344 + ], + "angle": 0, + "content": "3. Preliminaries: Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.353, + 0.895, + 0.552 + ], + "angle": 0, + "content": "The diffusion model is composed of a forward and a reverse process. The forward process is defined as a discrete Markov chain of length \\( T \\): \\( q(x_{1:T}|x_0) = \\prod_{t=1}^{T} q(x_t|x_{t-1}) \\). For each step \\( t \\in [1,T] \\) in the forward process, a diffusion model adds noise \\( \\epsilon_t \\) sampled from the Gaussian distribution \\( \\mathcal{N}(0,\\mathbf{I}) \\) to data \\( x_{t-1} \\) and obtains disturbed data \\( x_t \\) from \\( q(x_t|x_{t-1}) = \\mathcal{N}(x_t; \\sqrt{1 - \\beta_t} x_{t-1}, \\beta_t\\mathbf{I}) \\). \\( \\beta_t \\) decides the ratio of noise at timestep \\( t \\). Noticeably, instead of sampling sequentially along the Markov chain, we can sample \\( x_t \\) at any time step \\( t \\) in the closed form via \\( q(x_t|x_0) = \\mathcal{N}(x_t; \\sqrt{\\bar{\\alpha}_t} x_{t-1}, (1 - \\bar{\\alpha}_t)\\mathbf{I}) \\), where \\( \\bar{\\alpha}_t = \\prod_{s=1}^t (1 - \\beta_s) \\). To parameterize Gaussian distribution, the neural network \\( \\epsilon_\\theta \\) is introduced and optimized by the objective of DDPM [16]:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.561, + 0.892, + 0.58 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {S} = \\left\\| \\epsilon - \\epsilon_ {\\theta} \\left(\\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , t\\right) \\right\\| _ {2} ^ {2}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.591, + 0.895, + 0.715 + ], + "angle": 0, + "content": "In the reverse process, the diffusion model gradually de-noises the randomly sampled Gaussian noise to the high-quality output \\(x_0\\) through the predicted noise by the well-trained \\(\\epsilon_{\\theta}\\). This process is also defined as a Markov chain: \\(p_{\\theta}(x_{0:T}) = p(x_T)\\prod_{t=1}^{T}p_{\\theta}(x_{t-1}|x_t)\\), and \\(p_{\\theta}(x_{t-1}|x_t) = \\mathcal{N}(x_{t-1};\\mu_{\\theta}(x_t,t),\\sigma_t^2\\mathbf{I})\\). The mean and variance are \\(\\mu_{\\theta}(x_t,t) = \\frac{1}{\\sqrt{\\alpha_t}} (x_t - \\frac{\\beta_t}{\\sqrt{1 - \\bar{\\alpha}_t}}\\epsilon_{\\theta}(x_t,t))\\), \\(\\sigma_t^2 = \\frac{1 - \\bar{\\alpha}_{t-1}}{1 - \\bar{\\alpha}_t}\\beta_t\\), respectively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.727, + 0.593, + 0.743 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.753, + 0.895, + 0.813 + ], + "angle": 0, + "content": "In this work, we devise the first diffusion-based model to remove arbitrary adverse weather in videos with one set of model parameters. We follow an end-to-end formulation of video adverse weather removal, which reads:" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.823, + 0.892, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {V}} _ {k} = \\phi (\\mathbf {V} _ {k}), \\mathbf {V} _ {k} = \\left\\{I _ {i} \\right\\} _ {i = 1} ^ {N _ {f}}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.853, + 0.895, + 0.903 + ], + "angle": 0, + "content": "where \\(\\mathbf{V}_k\\) is the \\(k\\)-th video clip with \\(N_{f}\\) frames degraded by arbitrary weather type, \\(\\hat{\\mathbf{V}}_k\\) is the restored clip. \\(\\phi\\) is the denoising function derived from DDIM [37]. The denoising" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25608" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.09, + 0.868, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.455, + 0.892, + 0.526 + ], + "angle": 0, + "content": "Figure 3. Overview of our Diff-TTA framework for video adverse weather removal. In the training stage, through the temporal diffusion process, we randomly add temporal noise on the clean video clip \\(\\mathbf{V}_{gt}\\) from the mixed training set. Then, the denoising NAFNet \\(\\epsilon_{\\theta}\\) is trained by estimating the applied noise, which is conditioned by the degraded counterpart \\(\\mathbf{V}_{lq}\\) by seen weather types. In the test stage, the proxy task, Diff-TSC, where the cropped tubelets of the last restored pair \\(\\{\\mathbf{V}_{k-1}, \\hat{\\mathbf{V}}_{k-1}\\}\\) are utilized to learn the primer distribution, is incorporated into each timestep of denoising the randomly sampled temporal noise \\(\\mathbf{V}_k^T\\) for the iterative online adaptation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.534, + 0.47, + 0.562 + ], + "angle": 0, + "content": "network \\(\\epsilon_{\\theta}\\) in the denoising function consists of a single encoder and decoder." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.565, + 0.471, + 0.791 + ], + "angle": 0, + "content": "In the training stage, we follow the formulation of DDIM and extend the diffusion forward process to a temporal counterpart for frame-correlation modeling. The denoising network \\(\\epsilon_{\\theta}\\) with NAFNet [6] as the backbone is trained for noise estimation, which enforces the output of the model to approximate the proposed temporal noise model. During the inference stage, we present a proxy task, Diff-TSC, to adapt the diffusion model to the target test data degraded by unknown weather based on the diffusion reverse process. We iteratively optimize the noise estimation of the cropped tubelets and calibrate the weight set of the model parameters for more robust weather removal. Next, we elaborate on our diffusion-based solution for adverse weather removal in videos and our diffusion test-time adaptation, which are also displayed in Figure 3." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.327, + 0.818 + ], + "angle": 0, + "content": "4.1. Temporal Diffusion Process" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.901 + ], + "angle": 0, + "content": "The traditional diffusion model is trained to denoise independent noise from a perturbed image. The noise \\(\\epsilon\\) in Eq. (1) is sampled from an i.i.d. Gaussian distribution \\(\\epsilon \\sim \\mathcal{N}(0, \\mathbf{I})\\). Upon training the image diffusion model and utilizing it to reverse frames from a video into the noise" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.534, + 0.892, + 0.624 + ], + "angle": 0, + "content": "space individually, the significant correlation among the noise maps of these frames can be easily observed. Rather than introducing massive parameters into the denoising network, we extend the original diffusion process to the temporal counterpart by an ARMA-based temporal noise model to efficiently preserve the correlation between frames." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.624, + 0.892, + 0.699 + ], + "angle": 0, + "content": "Temporal noise model. ARMA (Auto-regressive Moving Average) [2] is a widely used statistical model for time series data, which combines the auto-regressive (AR) part and the moving average (MA) part. Typically, ARMA can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.708, + 0.892, + 0.75 + ], + "angle": 0, + "content": "\\[\nX _ {t} = c + \\varepsilon_ {t} + \\sum_ {i = 1} ^ {p} \\varphi_ {i} X _ {t - i} + \\sum_ {j = 1} ^ {q} \\tau_ {j} \\varepsilon_ {t - j}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.759, + 0.893, + 0.903 + ], + "angle": 0, + "content": "where \\(X\\) is the value term, \\(\\varepsilon\\) is the error term, \\(c\\) is a constant term, \\(p,q\\) are the auto-regressive order and moving average order while \\(\\varphi ,\\tau\\) are their coefficients, respectively. Motivated by ARMA, we develop the original Gaussian noise model \\(\\mathcal{N}(0,\\mathbf{I})\\) to the temporal noise model \\(\\mathcal{T}(0,\\mathbf{I})\\). Specifically, let \\(\\{\\bar{\\epsilon}_i\\}_{i = 1}^{N_f}\\) denote the noise sequence corresponding to consecutive frames of a video clip. We first sample the noise sequence \\(\\{\\bar{\\epsilon}_i\\}_{i = 1}^{N_f}\\) and the corresponding error sequence \\(\\{\\varepsilon_i\\}_{i = 1}^{N_f}\\) from \\(\\mathcal{N}(0,\\mathbf{I})\\), respectively. The constant" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "25609" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "term \\( c \\) should be consistent with the mean of Gaussian distribution, that is, zero. To take both the past and future information into consideration, we enforce the noise of the current frame affected not only by the noise and error term of the previous frame but also those of the next frame, and correspondingly the auto-regressive order and moving average order \\( p, q = 1 \\). Consequently, the expression in Eq. (3) evolves into the convex combination of the noise term and error term:" + }, + { + "type": "equation", + "bbox": [ + 0.084, + 0.231, + 0.47, + 0.261 + ], + "angle": 0, + "content": "\\[\n\\bar {\\epsilon} _ {i} = (1 - \\varphi - \\tau) \\varepsilon_ {i} + \\varphi \\frac {\\bar {\\epsilon} _ {i - 1} + \\bar {\\epsilon} _ {i + 1}}{2} + \\tau \\frac {\\varepsilon_ {i - 1} + \\varepsilon_ {i + 1}}{2}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.266, + 0.47, + 0.328 + ], + "angle": 0, + "content": "where \\(\\varphi, \\tau\\) are empirically set as 0.6, 0.3, respectively. Algorithm 1 further illustrates our temporal noise model, where the noise sequence of each video clip is constructed with separate preorder and postorder steps." + }, + { + "type": "code_caption", + "bbox": [ + 0.086, + 0.337, + 0.39, + 0.353 + ], + "angle": 0, + "content": "Algorithm 1: Temporal Noise Model \\(\\mathcal{T}(0,\\mathbf{I})\\)" + }, + { + "type": "algorithm", + "bbox": [ + 0.08, + 0.355, + 0.435, + 0.59 + ], + "angle": 0, + "content": "Input: \\(K\\) clips: \\(\\{\\mathbf{V}_1,\\mathbf{V}_2,\\dots ,\\mathbf{V}_K\\}\\) in a batch; The coefficients \\(\\varphi ,\\tau\\) \nfor \\(k = 1,2,\\ldots ,K\\) parallelly do \nfor Frame \\(i = 1,2,\\dots ,N_{f}\\) parallelly do \nsample \\(\\bar{\\epsilon}_i,\\varepsilon_i\\in \\mathbb{R}^{1\\times c\\times h\\times w}\\) from \\(\\mathcal{N}(0,\\mathbf{I})\\) \nend \nfor Frame \\(i = 2,3,\\dots ,N_{f}\\) sequentially do \n\\(\\bar{\\epsilon}_i = (1 - \\varphi -\\tau)\\varepsilon_i + \\varphi \\bar{\\epsilon}_{i - 1} + \\tau \\varepsilon_{i - 1}\\) \nend \nfor Frame \\(i = N_{f} - 1,\\dots ,2,1\\) sequentially do \n\\(\\bar{\\epsilon}_i = (1 - \\varphi -\\tau)\\varepsilon_i + \\varphi \\bar{\\epsilon}_{i + 1} + \\tau \\varepsilon_{i + 1}\\) \nend \n\\(\\bar{\\epsilon} =\\) normalize( \\(\\bar{\\epsilon})\\Rightarrow \\bar{\\epsilon}\\sim \\mathcal{T}(0,\\mathbf{I})\\) \nend \nReturn \\(\\{\\{\\bar{\\epsilon}_i|\\bar{\\epsilon}_i\\in \\mathbb{R}^{1\\times c\\times h\\times w}\\}_{i = 1}^{N_f}\\}_{k = 1}^K\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.61, + 0.47, + 0.701 + ], + "angle": 0, + "content": "To enhance the capacity of our denoising network for capturing temporal dependencies, we have specifically modified the architecture by substituting only the first and last layers of the denoising NAFNet with the 3D convolutional layer. This tailored network, denoted as \\(\\epsilon_{\\theta}\\) is then optimized as:" + }, + { + "type": "equation", + "bbox": [ + 0.126, + 0.708, + 0.47, + 0.727 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {S} = \\left\\| \\bar {\\epsilon} - \\epsilon_ {\\theta} \\left(\\sqrt {\\bar {\\alpha} _ {t}} \\hat {\\mathbf {V}} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\bar {\\epsilon}, \\mathbf {V}, t\\right) \\right\\| _ {1}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.734, + 0.469, + 0.765 + ], + "angle": 0, + "content": "where we utilize the temporal noise model as the supervision, and \\(\\| \\cdot \\|\\) denotes \\(L_{1}\\) norm." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.772, + 0.353, + 0.789 + ], + "angle": 0, + "content": "4.2. Diffusion Test-time Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Once trained by the proposed temporal noise model, the network \\(\\epsilon_{\\theta}\\) can generate high-quality images. This is achieved by sampling a noisy state \\(\\mathbf{V}_T\\) and then iteratively denoising it in the reverse process. Yang et al. [51] introduced adversarial learning for weather conditions, which require weather type annotations during the training stage. Additionally, their method may fail to address out-of-distribution" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.228 + ], + "angle": 0, + "content": "data, because including all real-world conditions in training, considering both weather types and image statistics, is impractical. Contrastively, we introduce test-time adaptation and update knowledge by practice to remove arbitrary weather conditions robustly. We design and incorporate Diffusion Tubelet Self-Calibration (Diff-TSC), a proxy task, into the iterative denoising of the diffusion reverse process to learn the primer distribution of the frames degraded by unknown weather." + }, + { + "type": "code_caption", + "bbox": [ + 0.508, + 0.237, + 0.852, + 0.298 + ], + "angle": 0, + "content": "Algorithm 2: Diffusion Test-Time Adaptation to unknown weather. \\(\\theta\\) is the weight set of the trained network before adaptation, \\(\\delta\\) is the learning rate for online adaptation." + }, + { + "type": "algorithm", + "bbox": [ + 0.501, + 0.3, + 0.852, + 0.625 + ], + "angle": 0, + "content": "Input: \\(K\\) overlapped clips: \\(\\{\\mathbf{V}_1,\\mathbf{V}_2,\\dots ,\\mathbf{V}_K\\}\\) in one video stream \nfor \\(Clp k = 1,2,\\ldots ,K\\) sequentially do \nInitialize the network \\(\\epsilon_{\\bar{\\theta}}\\) with \\(\\theta\\) \nInitialize \\(\\hat{\\mathbf{V}}_k\\) by Algorithm 1 \nif \\(k = 1\\) then \nfor step \\(t = T,\\dots ,2,1\\) sequentially do \n\\(|\\hat{\\mathbf{V}}_k = ddim(\\mathbf{V}_k,\\hat{\\mathbf{V}}_k,\\epsilon_{\\bar{\\theta}},t)\\) \nend \nelse \nA, \\(\\hat{\\mathbf{A}} =\\) TubeletCrop(Vk-1,Vk-1,Na) for step \\(t = T,\\dots ,2,1\\) sequentially do \nCompute \\(\\mathcal{L}_S(\\mathbf{A},\\hat{\\mathbf{A}},\\bar{\\epsilon},t)\\) by Eq. (5) \\(\\bar{\\theta} = \\bar{\\theta} -\\delta \\nabla_{\\bar{\\theta}}\\mathcal{L}_{S}\\) \\(\\hat{\\mathbf{V}}_k = ddim(\\mathbf{V}_k,\\hat{\\mathbf{V}}_k,\\epsilon_{\\bar{\\theta}},t)\\) \nend \n\\(\\hat{\\mathbf{V}}_k =\\) integrate( \\(\\hat{\\mathbf{V}}_k,\\hat{\\mathbf{V}}_{k - 1})\\) \nend \nReturn the restored clips \\(\\{\\hat{\\mathbf{V}}_1,\\hat{\\mathbf{V}}_2,\\dots ,\\hat{\\mathbf{V}}_K\\}\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Diffusion Tubelet Self-Calibration. We design a proxy task for self-supervised learning to transfer the diffusion model to the specific weather condition. Specifically, for each clip from the video stream, we first initialize the weight set of \\(\\epsilon_{\\bar{\\theta}}\\) with \\(\\theta\\) trained by source data. We leverage the previous clip \\(\\mathbf{V}_{k - 1}\\) and its restored result \\(\\hat{\\mathbf{V}}_{k - 1}\\) as the pair of condition and input for the online optimization when inferring the current clip \\(\\mathbf{V}_k\\). Rather than directly traversing all fine-grained details, we randomly crop \\(N_{a}\\) tubelets from \\(\\mathbf{V}_{k - 1}\\) as the primer distribution to avoid memory overhead and massive computational costs caused by high resolution. Following the process in Section 4.1, we conduct our temporal diffusion process on the restored tubelets \\(\\hat{\\mathbf{A}}\\) and feed them together with the degraded tubelets \\(\\mathbf{A}\\) into the denoising NAFNet for noise estimation. As suggested in [40], it is necessary to iteratively find a relatively optimal point for the target distribution during test-time adaptation. Inher" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25610" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.09, + 0.894, + 0.145 + ], + "angle": 0, + "content": "Table 1. Quantitative evaluation on seen weather conditions for video adverse weather removal. For Original Weather, these methods are trained on the weather-specific training set and tested on the weather-specific testing set. For Rain, Haze, and Snow, these methods are trained on a mixed training set and tested on the weather-specific testing set. The average performance is calculated on Rain, Haze, and Snow. PSNR and SSIM are adopted as our evaluation metrics. The best and second-best values are bold and underlined." + }, + { + "type": "table", + "bbox": [ + 0.115, + 0.15, + 0.847, + 0.426 + ], + "angle": 0, + "content": "
MethodsTypeSourceDatasets
Original WeatherRainHazeSnowAverage
DerainPReNet [34]ImageCVPR'1927.060.907726.800.881417.640.803028.570.940124.340.8748
SLDNet [50]VideoCVPR'2020.310.627221.240.712916.210.756122.010.855019.820.7747
S2VD [53]VideoCVPR'2124.090.794428.390.900619.650.860726.230.919024.760.8934
RDD-Net [42]VideoECCV'2231.820.942330.340.930018.360.843230.400.956026.370.9097
DehazeGDN [26]ImageICCV'1919.690.854529.960.937019.010.880531.020.951826.660.9231
MSBDN [15]ImageCVPR'2022.010.875926.700.914622.240.904727.070.934025.340.9178
VDHNet [35]VideoTIP'1916.640.813329.870.927216.850.821429.530.939525.420.8960
PM-Net [28]VideoMM'2223.830.895025.790.888023.570.914318.710.788122.690.8635
DesnowDesnowNet [29]ImageTIP'1828.300.953025.190.878616.430.790227.560.918123.060.8623
DDMSNET [56]ImageTIP'2132.550.961329.010.918819.500.861532.430.969426.980.9166
HDCW-W-Net [10]ImageICCV'2131.770.954228.100.905517.360.792131.050.948225.500.8819
SMGARN [13]ImageTCSVT'2233.240.972127.780.910017.850.807532.340.966825.990.8948
RestorationMPRNet [54]ImageCVPR'21————28.220.916520.250.893430.950.948226.470.9194
EDVR [43]VideoCVPR'19————31.100.937119.670.872430.270.944027.010.9178
RVRT [23]VideoNIPS'22————30.110.913221.160.894926.780.883426.020.8972
RTA [61]VideoCVPR'22————30.120.918620.750.891529.790.936726.890.9156
Multi-WeatherAll-in-one [21]ImageCVPR'20————26.620.894820.880.901030.090.943125.860.9130
UVRNet [19]ImageTMM'22————22.310.767820.820.857524.710.887322.610.8375
TransWeather [38]ImageCVPR'22————26.820.911822.170.902528.870.931325.950.9152
TKL [11]ImageCVPR'22————26.730.893522.080.904431.350.951526.720.9165
WeatherDiffusion [33]ImageTPAMI'23————25.860.912520.100.844226.400.911324.120.8893
WGWS-Net [62]ImageCVPR'23————29,640.931017.710.811331.580.952826.310.9265
ViWS-Net [51]VideoICCV'23————31.520.943324.510.918731.490.956229.170.9394
OursVideo——————32.430.957324.560.914831.860.964029.630.9453
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.442, + 0.473, + 0.701 + ], + "angle": 0, + "content": "ited from the characteristics of DDIM, such requirement can be easily achieved in the iterative diffusion reverse process, which needs more increments of inference time in the traditional CNNs and Transformers. For each timestep \\( t \\), we first obtain the noisy tubelet \\( \\hat{\\mathbf{A}}_t \\) and then optimize the objective function given in Eq. (5) to update the weight set \\( \\bar{\\theta} \\). After the iterative adaptation to \\( \\hat{\\mathbf{V}}_k \\), we ensemble the overlapped part of it with that of the previous clip \\( \\hat{\\mathbf{V}}_{k - 1} \\) to capture complementary information of pre-adaptation and post-adaptation. We optimize over the entire network except for the last layer suggested by SHOT [22]. The detailed procedure is described in Algorithm 2. We emphasize the synergy between diffusion models and test-time adaptation in accumulating frame-correlated knowledge for adverse weather removal. By attaching the proxy task to iterative denoising, our approach also avoids excessive computational overhead caused by iterative adaptation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.714, + 0.21, + 0.731 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.74, + 0.181, + 0.755 + ], + "angle": 0, + "content": "5.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Various video adverse weather datasets are used in our experiments. Following the setting of [51], we adopt RainMotion [42], REVIDE [59] and KITTI-snow [51] as seen weather. For the training stage, we merge the training set of the three datasets as the mixed set to learn a generic model. For the testing stage, we evaluate our model on three testing sets, respectively. Additionally, we evaluate our approach on two datasets VRDS [44] and RVSD [4] to validate the generalization of our Diff-TTA on unseen" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.44, + 0.892, + 0.483 + ], + "angle": 0, + "content": "Table 2. Quantitative evaluation on unseen weather conditions for video adverse weather removal. The best and second-best performances are marked in bold and underlined, respectively." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.487, + 0.872, + 0.603 + ], + "angle": 0, + "content": "
MethodVRDS [44]RVSD [4]
PSNR ↑SSIM ↑PSNR ↑SSIM ↑
All-in-one [21]20.440.594419.790.7509
TransWeather [38]21.360.713620.250.7514
TKL [11]20.490.700319.710.7370
WeatherDiffusion [33]20.730.694318.080.6588
ViWS-Net [51]21.570.709419.830.7590
Diff-TTA (ours)22.570.728122.350.7719
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.892, + 0.749 + ], + "angle": 0, + "content": "weather conditions. VRDS is a synthesized video dataset of joint rain streaks and raindrops with a total of 102 videos, while RVSD is a realistic video desnowing dataset with a total of 110 videos containing both snow and fog achieved by the rendering engine. Due to the large quantity of the two datasets, we randomly sample 8 videos as the test set, each of which has 20 frames, respectively. We also collect several real-world videos degraded by diverse weather conditions to prove its effectiveness on real-world applications." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.757, + 0.661, + 0.773 + ], + "angle": 0, + "content": "5.2. Implementation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Training Details. The proposed framework was trained on NVIDIA RTX 4090 GPUs and implemented on the Pytorch platform. We adopted the Lion optimizer [12] (\\(\\beta_{1} = 0.9\\), \\(\\beta_{2} = 0.99\\)) with the initial learning rate of \\(2 \\times 10^{-5}\\) decayed to \\(2 \\times 10^{-7}\\) by the Cosine scheduler. We randomly crop the video frames to \\(256 \\times 256\\) for training. Our framework is empirically trained for \\(600k\\) iterations in an end-to-end way with batch size of 4. Each batch consists of 4 video" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "25611" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.091, + 0.184, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.252, + 0.172, + 0.263 + ], + "angle": 0, + "content": "Input Frame" + }, + { + "type": "image", + "bbox": [ + 0.186, + 0.091, + 0.285, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.252, + 0.252, + 0.262 + ], + "angle": 0, + "content": "EDVR" + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.092, + 0.384, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.296, + 0.253, + 0.378, + 0.263 + ], + "angle": 0, + "content": "TransWeather" + }, + { + "type": "image", + "bbox": [ + 0.386, + 0.092, + 0.485, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.425, + 0.253, + 0.447, + 0.262 + ], + "angle": 0, + "content": "TKL" + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.092, + 0.585, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.487, + 0.253, + 0.59, + 0.262 + ], + "angle": 0, + "content": "WeatherDiffusion" + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.092, + 0.686, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.609, + 0.253, + 0.666, + 0.262 + ], + "angle": 0, + "content": "ViWS-Net" + }, + { + "type": "image", + "bbox": [ + 0.688, + 0.092, + 0.786, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.724, + 0.254, + 0.752, + 0.262 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.788, + 0.092, + 0.886, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.829, + 0.254, + 0.845, + 0.262 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.274, + 0.893, + 0.316 + ], + "angle": 0, + "content": "Figure 4. Qualitative Comparison of seen weather conditions between our approach and state-of-the-art algorithms. The competitive algorithms are selected to present the results on the example frames degraded by rain, haze, snow, respectively. The color box indicates the comparison of details. Please zoom in on the images for improved visualization." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.325, + 0.264, + 0.339 + ], + "angle": 0, + "content": "clips with 5 frames per clip." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.34, + 0.47, + 0.52 + ], + "angle": 0, + "content": "Method-specific Details. During the training, we used the diffusion process linearly increasing from \\(\\beta_{1} = 0.0001\\) to \\(\\beta_{T} = 0.02\\) with \\(T = 1000\\) timesteps. During the testing, the inference step is reduced to 25 by deterministic implicit sampling. We utilize the same optimizer with a smaller learning rate of \\(1\\times 10^{-6}\\) for the test-time adaptation. The number of cropped tubelets \\(N_{a}\\) is empirically set to 5 while the crop size is \\(128\\times 128\\). Despite the incorporated adaptation, our Diff-TTA is particularly \\(90\\times\\) more efficient than WeatherDiffusion. To process a video clip of 5 frames with the patch size of \\(256\\times 256\\), the average run time of Diff-TTA is 6.01s while that of WeatherDiffusion is 542.76s." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.531, + 0.303, + 0.545 + ], + "angle": 0, + "content": "5.3. Performance Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.47, + 0.689 + ], + "angle": 0, + "content": "Comparison methods. As shown in Table 1, we compared our proposed method against five kinds of state-of-the-art methods on the three seen weather conditions as [51] does, i.e., derain, dehaze, desnow, restoration, all-in-one adverse weather removal. For all-in-one adverse weather removal, we compared ours with the six representative single-image methods All-in-one [21], UVRNet [19], TransWeather [38], TKL [11], WeatherDiffusion [33], WGWS-Net [62], and the video-level method ViWS-Net [51]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.69, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Quantitative Comparison. For quantitative evaluation of the restored results, we apply the peak signal-to-noise ratio (PSNR) and the structural similarity (SSIM) as the metrics. Following the previous experimental setting, only the results of the model trained on the mixed training set are reported for all-in-one adverse weather removal models. To ensure the fairness of comparisons, we retrain each compared model implemented by the official codes based on our training dataset and report the best result. Observed from Table 1, our method achieves the best average performance across rain, haze and snow conditions among all kinds of methods by a considerable margin of 0.46, 0.0059 in PSNR, SSIM, respectively, than the second-best method ViWS-Net [51]. Though some single-weather" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.325, + 0.892, + 0.534 + ], + "angle": 0, + "content": "methods like DDMSNET [56] and SMGARN [13] achieve impressive results in their own weather condition, they frequently fail to remove other weather conditions, such as haze. Our method significantly surpasses other all-in-one adverse weather methods in rain and snow conditions, and obtains competitive results in haze conditions. It is worth noting that the generalizable ability is critical for such methods, which can be observed from the average performance. To further validate their generalization, we evaluate all-in-on adverse weather removal methods on unseen weather conditions, i.e., rain streak+raindrop and snow+fog, in Table 2. Our Diff-TTA exhibits remarkable superiority over state-of-the-art methods when confronted with these unseen weather conditions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.537, + 0.892, + 0.734 + ], + "angle": 0, + "content": "Visual comparisons on synthetic videos. To vividly demonstrate the effectiveness of our approach, Figure 4 shows the qualitative comparison under the rain, haze, and snow conditions between our method and five state-of-the-art methods. Our method demonstrates promising results in terms of visual quality across various weather types. Specifically, when it comes to rain and snow scenarios, our approach significantly reduces the presence of rain streaks and snow particles, outperforming other methods. In hazy scenarios, our method effectively removes residual haze and preserves the clean background with impressive results. More visual results of seen and unseen weather are displayed in the supplementary files." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.734, + 0.892, + 0.854 + ], + "angle": 0, + "content": "Visual comparisons on real-world videos. To evaluate the universality of our Diff-TTA on real-world applications, we collect several real-world degraded videos by rain, haze, snow from Youtube website, and further compare our approach against other all-in-one adverse weather removal methods. Figure 5 displays the visual results produced by our network and the compared methods on real-world video frames, clearly showcasing our exceptional performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Ablation Study. We evaluate the effectiveness of each component including diffusion process, temporal noise model, and Diffusion test-time adaptation (Diff-TTA) as described" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25612" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.09, + 0.895, + 0.119 + ], + "angle": 0, + "content": "Table 3. Ablation study of each critical component in the proposed framework on seen weather conditions. The top values are marked in bold font. \"Diff-TTA\" denote Diffusion Test-time Adaptation." + }, + { + "type": "table", + "bbox": [ + 0.115, + 0.122, + 0.846, + 0.209 + ], + "angle": 0, + "content": "
CombinationComponentDatasets
Diffusion ProcessTemporal NoiseDiff-TTARainHazeSnowAverage
M1---29.070.951422.640.893028.790.9350
M2✓--31.550.946623.580.904130.480.9520
M3✓✓-32.100.953024.310.912431.040.9621
Ours✓✓✓32.430.957324.560.914831.860.9640
" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.216, + 0.23, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.129, + 0.386, + 0.201, + 0.396 + ], + "angle": 0, + "content": "Input Frame" + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.217, + 0.357, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.252, + 0.386, + 0.333, + 0.395 + ], + "angle": 0, + "content": "TransWeather" + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.217, + 0.484, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.408, + 0.386, + 0.431, + 0.395 + ], + "angle": 0, + "content": "TKL" + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.217, + 0.612, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.495, + 0.386, + 0.597, + 0.395 + ], + "angle": 0, + "content": "WeatherDiffusion" + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.217, + 0.738, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.645, + 0.386, + 0.703, + 0.395 + ], + "angle": 0, + "content": "ViWS-Net" + }, + { + "type": "image", + "bbox": [ + 0.738, + 0.217, + 0.866, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.786, + 0.386, + 0.816, + 0.395 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.401, + 0.895, + 0.444 + ], + "angle": 0, + "content": "Figure 5. Visual comparison of all-in-one adverse weather removal methods on the selected real-world video sequences degraded by rain, haze, snow. Apparently, our network can more effectively remove rain streaks, haze, and snowflakes of input video frames than state-of-the-art methods. Please zoom in on the images for improved visualization." + }, + { + "type": "image", + "bbox": [ + 0.089, + 0.454, + 0.268, + 0.602 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.271, + 0.454, + 0.456, + 0.602 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.607, + 0.471, + 0.648 + ], + "angle": 0, + "content": "Figure 6. Detailed results on each case of VRDS and RVSD by our proposed Diff-TTA. The proposed Diff-TTA brings consistent improvement across various cases." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.473, + 0.901 + ], + "angle": 0, + "content": "in Table 3. We present the results trained on the mixed training set and tested on three weather-specific testing sets (rain, haze, snow) in view of PSNR and SSIM. The baseline model M1, i.e., the vanilla NAFNet, achieves the average performance on the three adverse weather datasets of 26.83, 0.9264 in PSNR, SSIM, respectively. M2 is equipped with diffusion process and trained by noise estimation, which advances M1 on the average performance by 1.69, 0.0078 of PSNR, SSIM, respectively. M3 introduces the proposed temporal noise model by replacing the vanilla noise model, and brings about an improvement of 0.63, 0.0083 in average PSNR, SSIM, respectively. Finally, our full model employs diffusion test-time adaptation (Diff-TTA) and gains a critical increase of 0.48, 0.0028 in PSNR, SSIM, respectively, compared to M3. Figure 6 further ablates our iterative adaptation on each case of the two unseen datasets, demonstrat" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.456, + 0.696, + 0.471 + ], + "angle": 0, + "content": "ing consistent improvements." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.493, + 0.619, + 0.508 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.521, + 0.895, + 0.793 + ], + "angle": 0, + "content": "This paper presents Diff-TTA, a diffusion-based test-time adaptation method capable of restoring videos degraded by various, even unknown, weather conditions. The proposed approach involves constructing a diffusion-based network and utilizing a novel temporal noise model to establish correlations among video frames. More importantly, a proxy task Diff-TSC is incorporated into the diffusion reverse process to iteratively update the deep model. This allows for the efficient adaptation of vision systems in real-world scenarios without modifying the source training process, making it easily applicable to a wide range of off-the-shelf diffusion-based models for enhancing weather removal. Experimental results on both synthetic and real-world data, which encompass diverse weather conditions, confirm the effectiveness and generalization capability of the proposed approach. We believe Diff-TTA can be a compelling baseline that shifts the focus of the community from the complicated training process to online test-time adaptation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Acknowledgment. This work is supported by Guangzhou Municipal Science and Technology Project (Grant No. 2023A03J0671), Guangzhou-HKUST(GZ) Joint Funding Program (No. 2024A03J0618), a grant of General Research Fund from Hong Kong Research Grants Council (project no. 15218521), and the InnoHK funding launched by Innovation and Technology Commission, Hong Kong SAR." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "25613" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Malik Boudiaf, Romain Mueller, Ismail Ben Ayed, and Luca Bertinetto. Parameter-free online test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8344-8353, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.472, + 0.227 + ], + "angle": 0, + "content": "[2] George Box. Box and jenkins: time series analysis, forecasting and control. In A Very British Affair: Six Britons and the Development of Time Series Analysis During the 20th Century, pages 161-215. Springer, 2013. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.471, + 0.284 + ], + "angle": 0, + "content": "[3] Dian Chen, Dequan Wang, Trevor Darrell, and Sayna Ebrahimi. Contrastive test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 295-305, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.285, + 0.471, + 0.354 + ], + "angle": 0, + "content": "[4] Haoyu Chen, Jingjing Ren, Jinjin Gu, Hongtao Wu, Xuequan Lu, Haoming Cai, and Lei Zhu. Snow removal in video: A new dataset and a novel method. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13211-13222, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.355, + 0.471, + 0.424 + ], + "angle": 0, + "content": "[5] Jie Chen, Cheen-Hau Tan, Junhui Hou, Lap-Pui Chau, and He Li. Robust video content alignment and compensation for rain removal in a cnn framework. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6286-6295, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.425, + 0.471, + 0.479 + ], + "angle": 0, + "content": "[6] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European Conference on Computer Vision, pages 17-33. Springer, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.471, + 0.55 + ], + "angle": 0, + "content": "[7] Sixiang Chen, Tian Ye, Yun Liu, Jinbin Bai, Haoyu Chen, Yunlong Lin, Jun Shi, and Erkang Chen. CPLformer: Cross-scale prototype learning transformer for image snow removal. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4228-4239, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.551, + 0.471, + 0.634 + ], + "angle": 0, + "content": "[8] Sixiang Chen, Tian Ye, Chenghao Xue, Haoyu Chen, Yun Liu, Erkang Chen, and Lei Zhu. Uncertainty-driven dynamic degradation perceiving and background modeling for efficient single image desnowing. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4269-4280, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.635, + 0.471, + 0.718 + ], + "angle": 0, + "content": "[9] Wei-Ting Chen, Hao-Yu Fang, Jian-Jiun Ding, Cheng-Che Tsai, and Sy-Yen Kuo. Jstasr: Joint size and transparency-aware snow removal algorithm based on modified partial convolution and veiling effect removal. In European Conference on Computer Vision, pages 754-770. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.47, + 0.803 + ], + "angle": 0, + "content": "[10] Wei-Ting Chen, Hao-Yu Fang, Cheng-Lin Hsieh, Cheng-Che Tsai, I Chen, Jian-Jiun Ding, Sy-Yen Kuo, et al. All snow removed: Single image desnowing algorithm using hierarchical dual-tree complex wavelet representation and contradict channel loss. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4196-4205, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[11] Wei-Ting Chen, Zhi-Kai Huang, Cheng-Che Tsai, Hao-Hsiang Yang, Jian-Jiun Ding, and Sy-Yen Kuo. Learning multiple adverse weather removal via two-stage knowledge learning and multi-contrastive regularization: Toward a unified model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17653–17662, 2022. 1, 2, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[12] Xiangning Chen, Chen Liang, Da Huang, Esteban Real, Kaiyuan Wang, Yao Liu, Hieu Pham, Xuanyi Dong, Thang Luong, Cho-Jui Hsieh, et al. Symbolic discovery of optimization algorithms. arXiv preprint arXiv:2302.06675, 2023.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[13] Bodong Cheng, Juncheng Li, Ying Chen, Shuyi Zhang, and Tieyong Zeng. Snow mask guided adaptive residual network for image snow removal. arXiv preprint arXiv:2207.04754, 2022.6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.22, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[14] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.263, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[15] Hang Dong, Jinshan Pan, Lei Xiang, Zhe Hu, Xinyi Zhang, Fei Wang, and Ming-Hsuan Yang. Multi-scale boosted de-hazing network with dense feature fusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2157-2167, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.376 + ], + "angle": 0, + "content": "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.377, + 0.892, + 0.475 + ], + "angle": 0, + "content": "[17] Minhao Hu, Tao Song, Yujun Gu, Xiangde Luo, Jieneng Chen, Yinan Chen, Ya Zhang, and Shaoting Zhang. Fully test-time adaptation for image segmentation. In Medical Image Computing and Computer Assisted Intervention—MICCAI 2021: 24th International Conference, Strasbourg, France, September 27–October 1, 2021, Proceedings, Part III 24, pages 251–260. Springer, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.476, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[18] Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for model-agnostic domain generalization. Advances in Neural Information Processing Systems, 34:2427-2440, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.532, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[19] Ashutosh Kulkarni, Prashant W Patil, Subrahmanyam Murala, and Sunil Gupta. Unified multi-weather visibility restoration. IEEE Transactions on Multimedia, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.576, + 0.892, + 0.645 + ], + "angle": 0, + "content": "[20] Minghan Li, Qi Xie, Qian Zhao, Wei Wei, Shuhang Gu, Jing Tao, and Deyu Meng. Video rain streak removal by multiscale convolutional sparse coding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6644-6653, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.647, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[21] Ruoteng Li, Robby T Tan, and Loong-Fah Cheong. All in one bad weather removal using architectural search. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3175-3185, 2020. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.704, + 0.892, + 0.759 + ], + "angle": 0, + "content": "[22] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In International conference on machine learning, pages 6028-6039. PMLR, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.761, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[23] Jingyun Liang, Yuchen Fan, Xiaoyu Xiang, Rakesh Ranjan, Eddy Ilg, Simon Green, Jiezhang Cao, Kai Zhang, Radu Timofte, and Luc Van Gool. Recurrent video restoration transformer with guided deformable attention. arXiv preprint arXiv:2206.02146, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.831, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[24] Jiaying Liu, Wenhan Yang, Shuai Yang, and Zongming Guo. Erase or fill? deep joint recurrent rain removal and reconstruction in videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3233-3242, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "25614" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.162 + ], + "angle": 0, + "content": "[25] Quande Liu, Cheng Chen, Qi Dou, and Pheng-Ann Heng. Single-domain generalization in medical image segmentation via test-time adaptation from shape dictionary. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1756–1764, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.165, + 0.471, + 0.22 + ], + "angle": 0, + "content": "[26] Xiaohong Liu, Yongrui Ma, Zhihao Shi, and Jun Chen. Grid-dehazenet: Attention-based multi-scale network for image dehazing. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7314-7323, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.223, + 0.472, + 0.306 + ], + "angle": 0, + "content": "[27] Xihui Liu, Dong Huk Park, Samaneh Azadi, Gong Zhang, Arman Chopikyan, Yuxiao Hu, Humphrey Shi, Anna Rohrbach, and Trevor Darrell. More control for free! image synthesis with semantic diffusion guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 289-299, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.309, + 0.47, + 0.364 + ], + "angle": 0, + "content": "[28] Ye Liu, Liang Wan, Huazhu Fu, Jing Qin, and Lei Zhu. Phase-based memory network for video dehazing. In Proceedings of the 30th ACM International Conference on Multimedia, pages 5427-5435, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.367, + 0.469, + 0.421 + ], + "angle": 0, + "content": "[29] Yun-Fu Liu, Da-Wei Jaw, Shih-Chia Huang, and Jenq-Neng Hwang. Desnownet: Context-aware deep network for snow removal. IEEE Transactions on Image Processing, 27(6): 3064–3073, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.425, + 0.469, + 0.479 + ], + "angle": 0, + "content": "[30] Ziwei Luo, Fredrik K Gustafsson, Zheng Zhao, Jens Sjolund, and Thomas B Schon. Image restoration with meanreverting stochastic differential equations. arXiv preprint arXiv:2301.11699, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.483, + 0.469, + 0.55 + ], + "angle": 0, + "content": "[31] A Tuan Nguyen, Thanh Nguyen-Tang, Ser-Nam Lim, and Philip HS Torr. Tipi: Test time adaptation with transformation invariance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24162-24171, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.555, + 0.469, + 0.623 + ], + "angle": 0, + "content": "[32] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yafo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. In International conference on machine learning, pages 16888-16905. PMLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.627, + 0.469, + 0.682 + ], + "angle": 0, + "content": "[33] Ozan Özdenizci and Robert Legenstein. Restoring vision in adverse weather conditions with patch-based denoising diffusion models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.686, + 0.469, + 0.754 + ], + "angle": 0, + "content": "[34] Dongwei Ren, Wangmeng Zuo, Qinghua Hu, Pengfei Zhu, and Deyu Meng. Progressive image deraining networks: A better and simpler baseline. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3937-3946, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.757, + 0.469, + 0.813 + ], + "angle": 0, + "content": "[35] Wenqi Ren, Jingang Zhang, Xiangyu Xu, Lin Ma, Xiaochun Cao, Gaofeng Meng, and Wei Liu. Deep video dehazing with semantic segmentation. IEEE transactions on image processing, 28(4):1895-1908, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.469, + 0.856 + ], + "angle": 0, + "content": "[36] Jaskirat Singh, Stephen Gould, and Liang Zheng. High-fidelity guided image synthesis with latent diffusion models. arXiv preprint arXiv:2211.17084, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[37] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[38] Jeya Maria Jose Valanarasu, Rajeev Yasarla, and Vishal M Patel. Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2353-2363, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[39] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[40] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726, 2020. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.264, + 0.894, + 0.319 + ], + "angle": 0, + "content": "[41] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7201-7211, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.321, + 0.892, + 0.39 + ], + "angle": 0, + "content": "[42] Shuai Wang, Lei Zhu, Huazhu Fu, Jing Qin, Carola-Bibiane Schonlieb, Wei Feng, and Song Wang. Rethinking video rain streak removal: A new synthesis model and a deraining network with video rain prior. In European Conference on Computer Vision, pages 565-582. Springer, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.392, + 0.892, + 0.461 + ], + "angle": 0, + "content": "[43] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced deformable convolutional networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0-0, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.463, + 0.892, + 0.53 + ], + "angle": 0, + "content": "[44] Hongtao Wu, Yijun Yang, Haoyu Chen, Jingjing Ren, and Lei Zhu. Mask-guided progressive network for joint raindrop and rain streak removal in videos. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7216-7225, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.533, + 0.892, + 0.588 + ], + "angle": 0, + "content": "[45] Bin Xia, Yulun Zhang, Shiyin Wang, Yitong Wang, Xinglong Wu, Yapeng Tian, Wenming Yang, and Luc Van Gool. Diffir: Efficient diffusion model for image restoration. arXiv preprint arXiv:2303.09472, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.591, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[46] Jiaqi Xu, Xiaowei Hu, Lei Zhu, Qi Dou, Jifeng Dai, Yu Qiao, and Pheng-Ann Heng. Video dehazing via a multi-range temporal alignment network with physical prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18053–18062, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.661, + 0.892, + 0.729 + ], + "angle": 0, + "content": "[47] Wending Yan, Robby T Tan, Wenhan Yang, and Dengxin Dai. Self-aligned video deraining with transmission-depth consistency. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11966-11976, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.732, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[48] Hongzheng Yang, Cheng Chen, Meirui Jiang, Quande Liu, Jianfeng Cao, Pheng Ann Heng, and Qi Dou. Dltta: Dynamic learning rate for test-time adaptation on cross-domain medical images. IEEE Transactions on Medical Imaging, 41 (12):3575-3586, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.802, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[49] Wenhan Yang, Jiaying Liu, and Jiashi Feng. Frameconsistent recurrent video deraining with dual-level flow. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1661-1670, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[50] Wenhan Yang, Robby T Tan, Shiqi Wang, and Jiaying Liu. Self-learning video rain streak removal: When cyclic consistency meets temporal correspondence. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25615" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1720-1729, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[51] Yijun Yang, Angelica I Aviles-Rivero, Huazhu Fu, Ye Liu, Weiming Wang, and Lei Zhu. Video adverse-weather-component suppression network via weather messenger and adversarial backpropagation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13200-13210, 2023. 1, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.471, + 0.276 + ], + "angle": 0, + "content": "[52] Tian Ye, Sixiang Chen, Jinbin Bai, Jun Shi, Chenghao Xue, Jingxia Jiang, Junjie Yin, Erkang Chen, and Yun Liu. Adverse weather removal with codebook priors. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12653-12664, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.47, + 0.346 + ], + "angle": 0, + "content": "[53] Zongsheng Yue, Jianwen Xie, Qian Zhao, and Deyu Meng. Semi-supervised video deraining with dynamical rain generator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 642-652, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.47, + 0.417 + ], + "angle": 0, + "content": "[54] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14821-14831, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.419, + 0.469, + 0.473 + ], + "angle": 0, + "content": "[55] Jian Zhang, Lei Qi, Yinghuan Shi, and Yang Gao. Domain adaptor: A novel approach to test-time adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18971-18981, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.476, + 0.469, + 0.532 + ], + "angle": 0, + "content": "[56] Kaihao Zhang, Rongqing Li, Yanjiang Yu, Wenhan Luo, and Changsheng Li. Deep dense multi-scale network for snow removal using semantic and depth priors. IEEE Transactions on Image Processing, 30:7419-7431, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.534, + 0.469, + 0.588 + ], + "angle": 0, + "content": "[57] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.59, + 0.469, + 0.644 + ], + "angle": 0, + "content": "[58] Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test time robustness via adaptation and augmentation. Advances in Neural Information Processing Systems, 35: 38629-38642, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.469, + 0.729 + ], + "angle": 0, + "content": "[59] Xinyi Zhang, Hang Dong, Jinshan Pan, Chao Zhu, Ying Tai, Chengjie Wang, Jilin Li, Feiyue Huang, and Fei Wang. Learning to restore hazy video: A new real-world dataset and a new method. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9239-9248, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.469, + 0.786 + ], + "angle": 0, + "content": "[60] Yi Zhang, Xiaoyu Shi, Dasong Li, Xiaogang Wang, Jian Wang, and Hongsheng Li. A unified conditional framework for diffusion-based image restoration. arXiv preprint arXiv:2305.20049, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.856 + ], + "angle": 0, + "content": "[61] Kun Zhou, Wenbo Li, Liying Lu, Xiaoguang Han, and Jiangbo Lu. Revisiting temporal alignment for video restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6053-6062, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[62] Yurui Zhu, Tianyu Wang, Xueyang Fu, Xuanyu Yang, Xin Guo, Jifeng Dai, Yu Qiao, and Xiaowei Hu. Learning weather-general and weather-specific features for image" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.092, + 0.892, + 0.147 + ], + "angle": 0, + "content": "restoration under multiple adverse weather conditions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21747-21758, 2023. 3, 6, 7" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "25616" + } + ] +] \ No newline at end of file diff --git a/2024/Genuine Knowledge from Practice_ Diffusion Test-Time Adaptation for Video Adverse Weather Removal/254c789f-0e30-45ba-9906-00c14a08b4b6_origin.pdf b/2024/Genuine Knowledge from Practice_ Diffusion Test-Time Adaptation for Video Adverse Weather Removal/254c789f-0e30-45ba-9906-00c14a08b4b6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ac3bcc8250a8772b857f1e34ade40c40bd6999c3 --- /dev/null +++ b/2024/Genuine Knowledge from Practice_ Diffusion Test-Time Adaptation for Video Adverse Weather Removal/254c789f-0e30-45ba-9906-00c14a08b4b6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:403f2b02c9759b35dda8cd4f22498b6a2a11acfba9c8032de812c7836692f691 +size 5028495 diff --git a/2024/Genuine Knowledge from Practice_ Diffusion Test-Time Adaptation for Video Adverse Weather Removal/full.md b/2024/Genuine Knowledge from Practice_ Diffusion Test-Time Adaptation for Video Adverse Weather Removal/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9184cde05901feb80ca3e5e54e514eba17eedb16 --- /dev/null +++ b/2024/Genuine Knowledge from Practice_ Diffusion Test-Time Adaptation for Video Adverse Weather Removal/full.md @@ -0,0 +1,326 @@ +# Genuine Knowledge from Practice: Diffusion Test-Time Adaptation for Video Adverse Weather Removal + +Yijun Yang $^{1}$ , Hongtao Wu $^{1}$ , Angelica I. Aviles-Rivero $^{2}$ , Yulun Zhang $^{3}$ , Jing Qin $^{4}$ , Lei Zhu $^{1,5*}$ + +1The Hong Kong University of Science and Technology (Guangzhou) 2University of Cambridge 3ETH Zürich + +4The Hong Kong Polytechnic University 5The Hong Kong University of Science and Technology + +Project page: https://github.com/scott-yjyang/DiffTTA + +# Abstract + +Real-world vision tasks frequently suffer from the appearance of unexpected adverse weather conditions, including rain, haze, snow, and raindrops. In the last decade, convolutional neural networks and vision transformers have yielded outstanding results in single-weather video removal. However, due to the absence of appropriate adaptation, most of them fail to generalize to other weather conditions. Although ViWS-Net is proposed to remove adverse weather conditions in videos with a single set of pretrained weights, it is seriously blinded by seen weather at train-time and degenerates when coming to unseen weather during test-time. In this work, we introduce test-time adaptation into adverse weather removal in videos, and propose the first framework that integrates test-time adaptation into the iterative diffusion reverse process. Specifically, we devise a diffusion-based network with a novel temporal noise model to efficiently explore frame-correlated information in degraded video clips at training stage. During inference stage, we introduce a proxy task named Diffusion Tubelet Self-Calibration to learn the primer distribution of test video stream and optimize the model by approximating the temporal noise model for online adaptation. Experimental results, on benchmark datasets, demonstrate that our Test-Time Adaptation method with Diffusion-based network(Diff-TTA) outperforms state-of-the-art methods in terms of restoring videos degraded by seen weather conditions. Its generalizable capability is validated with unseen weather conditions in synthesized and real-world videos. + +# 1. Introduction + +Adverse weather like rain, snow, and haze is very common in many existing outdoor videos. The presence of adverse weather usually decreases the visibility of the captured videos and significantly impairs the performance of + +![](images/293ac8b6fdcba7f56490225645c32bd6b578be91b5edf728c832968484f0e142.jpg) +Figure 1. Overview of the existing all-in-one adverse weather removal methods. Our approach can achieve superior performance not only in seen weather conditions but also in unseen weather conditions with a single set of pre-trained weights by diffusion test-time adaptation. In particular, our Diff-TTA is $90 \times$ more efficient than WeatherDiffusion. + +subsequent high-level vision applications such as object detection, semantic segmentation and autonomous driving. To mitigate these negative effects, Convolutional Neural Networks (CNNs) and Transformers are frequently utilized in the literature. Also, Diffusion models have emerged as the de-facto generative model for image synthesis [14, 16, 36] and have been developed for single-image adverse weather removal [33]. While deep neural networks can achieve decent results on test points within the distribution, real-world applications often demand robustness. This is especially crucial when facing unknown input perturbations, changes in weather conditions, or other sources of distribution shift. Existing single-weather removal approaches [4, 20, 59] require follow-up training to adapt to specific domains. This leads to the switching between different parameter sets, which can make the pipeline cumbersome. + +Recently, all-in-one adverse weather removal has attracted the attention of the community, which aims to handle diverse weather conditions with a generic model [11, 21, 38, 51]. However, the potential issue of frequently failing to + +![](images/ee45d853939569f13235d9fc562ce0f4228cfb244b9e8204dd1ba663bc36db4d.jpg) +Figure 2. Our Diff-TTA enables weather removal models to overcome unseen weather corruptions. We use t-SNE [39] to visualize features from the last feature extractor layer of each dataset. Obviously, unseen data points tend to approximate the seen ones after adaptation, which means Diff-TTA can categorize unknown degradation into known distribution. ('Real-world' contains video clips simultaneously degraded by fog and snow.) + +recover from out-of-distribution conditions remains largely unaddressed due to the finiteness of training data and the absence of a test data adaptation mechanism. As shown in Figure 1, generic models suffer significant degradation when handling all the possible weather conditions beyond the scope of their training data. Furthermore, a common aspect among many real-world applications is the necessity for online adaptation with limited data. This is particularly relevant in scenarios such as autonomous vehicles or drones, where a vision model must continually process a video stream and adjust its predictions. To achieve effective adaptation, it is crucial to leverage the incoming data in a way that is beneficial and does not introduce harmful biases, even when there are distribution shifts between the training and test data. Thus, akin to how humans update knowledge through practice, the task of removing any weather condition through online adaptation using a single set of pre-trained weights is both pragmatic and promising, albeit challenging. + +In this paper, we construct the diffusion-based framework for video adverse weather removal. Yet, diffusion models usually entail significant training overhead, hindering the technique's broader adoption in the research community of video tasks. Instead of introducing massive parameters for temporal encoding, we design an efficient strategy based on DDIM [37] to preserve temporal correlation, specifically temporal noise model, in order to alleviate its training load. In contrast to the vanilla noise model, which independently samples noise for each frame from a standard Gaussian distribution, our temporal noise model takes into account the inter-frame relationship via time series models. More importantly, to promptly address unexpected weather conditions, we introduce test-time adaptation techniques by recalibrating the model's parameters using only the degraded test data. The mainstream of test-time adaptation methods [18, 40, 58] manipulate and iteratively update the normalization statistics of CNNs and Vision Transformer using specific proxy tasks at the test phase. However, these + +methods often fail when transferring to low-level tasks. Additionally, incorporating such iterative adaptation into these models significantly increases the cost of inference, multiplying it by the number of iterations, compared to the original counterpart. + +In contrast, we highlight the advantage of integrating the diffusion reverse process and test-time adaptation based on their common characteristics of iteration. As far as we know, there has been no work to emerge the possibility of Diffusion models and test-time adaptation. Based on our diffusion-based framework, we design a proxy task and efficiently incorporate it into each denoising iteration when reversing our temporal diffusion process. Specifically, we develop a surrogate task called Diffusion Tubelet Self-Calibration (Diff-TSC), which is supervised by temporal noise model. This is designed to be task-agnostic and does not require any alterations to the training procedure. Rather than directly traversing all fine-grained details, we extract a few tubelets to taste the primer distribution of the frames degraded by unknown weather. This approach facilitates rapid optimization of the diffusion model while also reducing computational costs. As displayed in Figure 2, test-time adaptation helps to improve the model's robustness in real-world scenarios by mitigating the gap between seen and unseen weather corruptions. + +Our contributions can be summarized as: + +- We introduce the first diffusion-based framework for all-in-one adverse weather removal in videos. This framework efficiently leverages temporal redundancy information through the novel temporal noise model. +- To improve the model's robustness against unknown weather, we are the first to introduce test-time adaptation by incorporating a proxy task into the diffusion reverse process to learn the primer distribution of the test data. +- We extensively evaluate our approach on seen weather (e.g., rain, haze, snow) and unseen weather (e.g., rain streak+raindrop, snow+fog). The superior performance over synthesized and real-world videos consistently validates its effectiveness and generalizable ability. + +# 2. Related Work + +All-in-one Adverse Weather Removal. Over the past decade, The community of low-level vision has paid much attention to single weather removal, e.g., deraining [5, 20, 24, 42, 47, 49, 50, 53], dehazing [28, 35, 46, 59], desnowing [4, 7-9, 56], both on the image level and the video level. Due to the limitations of existing models in adapting to different weather conditions, researchers have recently been investigating the use of a single model instance for multiple adverse weather removal tasks. At the image level, following the pioneering work All-in-one [21], TransWeather [38] and TKL [11] were introduced to tackle many weather conditions with the single encoder and de + +coder. Zhu et al. [62] elaborated a two-stage training strategy to explore the weather-general and weather-specific information for the sophisticated removal of diverse weather conditions. Ye et al. [52] introduced high-quality codebook priors learned by a pre-trained VQGAN to restore texture details from adverse weather degradation. The aforementioned methods failed to capture redundant information from temporal space. While they can be adapted for frame-by-frame adverse weather removal, incorporating temporal data in video-level methods provides a more robust solution for dealing with multiple adverse weather conditions in videos. Very recently, Yang et al. [51] introduced 'ViWS-Net', the first video-level approach. This method uses weather messenger tokens to transfer weather knowledge across frames. It also employs adversarial learning to distinguish the weather type, thereby emphasizing weather-invariant information. However, they did not consider online adaptation of the deep model, making the potential collapse for unseen scenarios degraded by unknown weather. + +Diffusion Models. Diffusion model is well-known as a generative modeling approach for its superior achievements in image synthesis and generation tasks [14, 16, 27, 36, 57]. In essence, Denoising Diffusion Probabilistic Models (DDPM) [16] adopted parameterized Markov chain to optimize the lower variational bound on the likelihood function, which can simulate a diffusion process to iteratively improve the quality of target distribution than other generative models. For restoration tasks, Luo et al. [30] applied stochastic differential equations featuring mean reversion in diffusion models to improve low-quality images. DiffIR [45] provided a strong diffusion baseline for image restoration by estimating a compact IR prior representation. Zhang et al. [60] proposed a unified conditional framework based on diffusion models for image restoration by multi-source guidance of a lightweight U-Net. Ozdenizci et al. [33] designed a patch-based diffusion modeling approach called WeatherDiffusion to make the network size-agnostic. While this technique utilizes diffusion models for image weather removal, its very slow inference speed restricts practical use. Nevertheless, there is no work adapting diffusion models to video adverse weather removal. + +Test-time Adaptation. Recently, there has been a growing interest in adapting deep models using exclusively sequential, unlabeled test data. This contrasts with the traditional unsupervised domain adaptation, which asks the access to both training data and a sufficient amount of unlabeled test data. Test-time adaptation techniques allow for model updates using distributional statistics from either a single instance or batch of unlabeled test data. Typically, TENT [40] focuses on minimizing the model's prediction entropy on the target data. T3A [18] adjusts the classifier of a trained source model by computing a pseudo-prototype representation of different classes using unlabeled test data. + +MEMO [58] aims to minimize marginal entropy, compelling the deep model to consistently predict across various augmentations. This approach enforces the invariances of these augmentations and ensures prediction confidence. Chen et al. [3] proposed to apply self-supervised contrastive learning for target feature learning, which was trained jointly with pseudo labeling. Although test-time adaptation methods are increasingly used in image classification [1, 31, 32, 41, 55] and segmentation [17, 25, 48] tasks, their potential in adverse weather removal tasks remains unexplored. In our work, the nature of the iterative sampling process in Diffusion model significantly promotes the adaptation to the target distribution. Video scenarios further allow for the accumulation of the adaptation knowledge for the subsequent frames. + +# 3. Preliminaries: Diffusion Models + +The diffusion model is composed of a forward and a reverse process. The forward process is defined as a discrete Markov chain of length $T$ : $q(x_{1:T}|x_0) = \prod_{t=1}^{T} q(x_t|x_{t-1})$ . For each step $t \in [1,T]$ in the forward process, a diffusion model adds noise $\epsilon_t$ sampled from the Gaussian distribution $\mathcal{N}(0,\mathbf{I})$ to data $x_{t-1}$ and obtains disturbed data $x_t$ from $q(x_t|x_{t-1}) = \mathcal{N}(x_t; \sqrt{1 - \beta_t} x_{t-1}, \beta_t\mathbf{I})$ . $\beta_t$ decides the ratio of noise at timestep $t$ . Noticeably, instead of sampling sequentially along the Markov chain, we can sample $x_t$ at any time step $t$ in the closed form via $q(x_t|x_0) = \mathcal{N}(x_t; \sqrt{\bar{\alpha}_t} x_{t-1}, (1 - \bar{\alpha}_t)\mathbf{I})$ , where $\bar{\alpha}_t = \prod_{s=1}^t (1 - \beta_s)$ . To parameterize Gaussian distribution, the neural network $\epsilon_\theta$ is introduced and optimized by the objective of DDPM [16]: + +$$ +\mathcal {L} _ {S} = \left\| \epsilon - \epsilon_ {\theta} \left(\sqrt {\bar {\alpha} _ {t}} x _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon , t\right) \right\| _ {2} ^ {2}. \tag {1} +$$ + +In the reverse process, the diffusion model gradually de-noises the randomly sampled Gaussian noise to the high-quality output $x_0$ through the predicted noise by the well-trained $\epsilon_{\theta}$ . This process is also defined as a Markov chain: $p_{\theta}(x_{0:T}) = p(x_T)\prod_{t=1}^{T}p_{\theta}(x_{t-1}|x_t)$ , and $p_{\theta}(x_{t-1}|x_t) = \mathcal{N}(x_{t-1};\mu_{\theta}(x_t,t),\sigma_t^2\mathbf{I})$ . The mean and variance are $\mu_{\theta}(x_t,t) = \frac{1}{\sqrt{\alpha_t}} (x_t - \frac{\beta_t}{\sqrt{1 - \bar{\alpha}_t}}\epsilon_{\theta}(x_t,t))$ , $\sigma_t^2 = \frac{1 - \bar{\alpha}_{t-1}}{1 - \bar{\alpha}_t}\beta_t$ , respectively. + +# 4. Method + +In this work, we devise the first diffusion-based model to remove arbitrary adverse weather in videos with one set of model parameters. We follow an end-to-end formulation of video adverse weather removal, which reads: + +$$ +\hat {\mathbf {V}} _ {k} = \phi (\mathbf {V} _ {k}), \mathbf {V} _ {k} = \left\{I _ {i} \right\} _ {i = 1} ^ {N _ {f}}, \tag {2} +$$ + +where $\mathbf{V}_k$ is the $k$ -th video clip with $N_{f}$ frames degraded by arbitrary weather type, $\hat{\mathbf{V}}_k$ is the restored clip. $\phi$ is the denoising function derived from DDIM [37]. The denoising + +![](images/0e1aa3868e40cc558991bf5c6e8e6a82e4bcfe43ca3e06ca7c240e30f341f7cc.jpg) +Figure 3. Overview of our Diff-TTA framework for video adverse weather removal. In the training stage, through the temporal diffusion process, we randomly add temporal noise on the clean video clip $\mathbf{V}_{gt}$ from the mixed training set. Then, the denoising NAFNet $\epsilon_{\theta}$ is trained by estimating the applied noise, which is conditioned by the degraded counterpart $\mathbf{V}_{lq}$ by seen weather types. In the test stage, the proxy task, Diff-TSC, where the cropped tubelets of the last restored pair $\{\mathbf{V}_{k-1}, \hat{\mathbf{V}}_{k-1}\}$ are utilized to learn the primer distribution, is incorporated into each timestep of denoising the randomly sampled temporal noise $\mathbf{V}_k^T$ for the iterative online adaptation. + +network $\epsilon_{\theta}$ in the denoising function consists of a single encoder and decoder. + +In the training stage, we follow the formulation of DDIM and extend the diffusion forward process to a temporal counterpart for frame-correlation modeling. The denoising network $\epsilon_{\theta}$ with NAFNet [6] as the backbone is trained for noise estimation, which enforces the output of the model to approximate the proposed temporal noise model. During the inference stage, we present a proxy task, Diff-TSC, to adapt the diffusion model to the target test data degraded by unknown weather based on the diffusion reverse process. We iteratively optimize the noise estimation of the cropped tubelets and calibrate the weight set of the model parameters for more robust weather removal. Next, we elaborate on our diffusion-based solution for adverse weather removal in videos and our diffusion test-time adaptation, which are also displayed in Figure 3. + +# 4.1. Temporal Diffusion Process + +The traditional diffusion model is trained to denoise independent noise from a perturbed image. The noise $\epsilon$ in Eq. (1) is sampled from an i.i.d. Gaussian distribution $\epsilon \sim \mathcal{N}(0, \mathbf{I})$ . Upon training the image diffusion model and utilizing it to reverse frames from a video into the noise + +space individually, the significant correlation among the noise maps of these frames can be easily observed. Rather than introducing massive parameters into the denoising network, we extend the original diffusion process to the temporal counterpart by an ARMA-based temporal noise model to efficiently preserve the correlation between frames. + +Temporal noise model. ARMA (Auto-regressive Moving Average) [2] is a widely used statistical model for time series data, which combines the auto-regressive (AR) part and the moving average (MA) part. Typically, ARMA can be formulated as: + +$$ +X _ {t} = c + \varepsilon_ {t} + \sum_ {i = 1} ^ {p} \varphi_ {i} X _ {t - i} + \sum_ {j = 1} ^ {q} \tau_ {j} \varepsilon_ {t - j}, \tag {3} +$$ + +where $X$ is the value term, $\varepsilon$ is the error term, $c$ is a constant term, $p,q$ are the auto-regressive order and moving average order while $\varphi ,\tau$ are their coefficients, respectively. Motivated by ARMA, we develop the original Gaussian noise model $\mathcal{N}(0,\mathbf{I})$ to the temporal noise model $\mathcal{T}(0,\mathbf{I})$ . Specifically, let $\{\bar{\epsilon}_i\}_{i = 1}^{N_f}$ denote the noise sequence corresponding to consecutive frames of a video clip. We first sample the noise sequence $\{\bar{\epsilon}_i\}_{i = 1}^{N_f}$ and the corresponding error sequence $\{\varepsilon_i\}_{i = 1}^{N_f}$ from $\mathcal{N}(0,\mathbf{I})$ , respectively. The constant + +term $c$ should be consistent with the mean of Gaussian distribution, that is, zero. To take both the past and future information into consideration, we enforce the noise of the current frame affected not only by the noise and error term of the previous frame but also those of the next frame, and correspondingly the auto-regressive order and moving average order $p, q = 1$ . Consequently, the expression in Eq. (3) evolves into the convex combination of the noise term and error term: + +$$ +\bar {\epsilon} _ {i} = (1 - \varphi - \tau) \varepsilon_ {i} + \varphi \frac {\bar {\epsilon} _ {i - 1} + \bar {\epsilon} _ {i + 1}}{2} + \tau \frac {\varepsilon_ {i - 1} + \varepsilon_ {i + 1}}{2}, \tag {4} +$$ + +where $\varphi, \tau$ are empirically set as 0.6, 0.3, respectively. Algorithm 1 further illustrates our temporal noise model, where the noise sequence of each video clip is constructed with separate preorder and postorder steps. + +Algorithm 1: Temporal Noise Model $\mathcal{T}(0,\mathbf{I})$ +Input: $K$ clips: $\{\mathbf{V}_1,\mathbf{V}_2,\dots ,\mathbf{V}_K\}$ in a batch; The coefficients $\varphi ,\tau$ +for $k = 1,2,\ldots ,K$ parallelly do +for Frame $i = 1,2,\dots ,N_{f}$ parallelly do +sample $\bar{\epsilon}_i,\varepsilon_i\in \mathbb{R}^{1\times c\times h\times w}$ from $\mathcal{N}(0,\mathbf{I})$ +end +for Frame $i = 2,3,\dots ,N_{f}$ sequentially do + $\bar{\epsilon}_i = (1 - \varphi -\tau)\varepsilon_i + \varphi \bar{\epsilon}_{i - 1} + \tau \varepsilon_{i - 1}$ +end +for Frame $i = N_{f} - 1,\dots ,2,1$ sequentially do + $\bar{\epsilon}_i = (1 - \varphi -\tau)\varepsilon_i + \varphi \bar{\epsilon}_{i + 1} + \tau \varepsilon_{i + 1}$ +end + $\bar{\epsilon} =$ normalize( $\bar{\epsilon})\Rightarrow \bar{\epsilon}\sim \mathcal{T}(0,\mathbf{I})$ +end +Return $\{\{\bar{\epsilon}_i|\bar{\epsilon}_i\in \mathbb{R}^{1\times c\times h\times w}\}_{i = 1}^{N_f}\}_{k = 1}^K$ + +To enhance the capacity of our denoising network for capturing temporal dependencies, we have specifically modified the architecture by substituting only the first and last layers of the denoising NAFNet with the 3D convolutional layer. This tailored network, denoted as $\epsilon_{\theta}$ is then optimized as: + +$$ +\mathcal {L} _ {S} = \left\| \bar {\epsilon} - \epsilon_ {\theta} \left(\sqrt {\bar {\alpha} _ {t}} \hat {\mathbf {V}} + \sqrt {1 - \bar {\alpha} _ {t}} \bar {\epsilon}, \mathbf {V}, t\right) \right\| _ {1}, \tag {5} +$$ + +where we utilize the temporal noise model as the supervision, and $\| \cdot \|$ denotes $L_{1}$ norm. + +# 4.2. Diffusion Test-time Adaptation + +Once trained by the proposed temporal noise model, the network $\epsilon_{\theta}$ can generate high-quality images. This is achieved by sampling a noisy state $\mathbf{V}_T$ and then iteratively denoising it in the reverse process. Yang et al. [51] introduced adversarial learning for weather conditions, which require weather type annotations during the training stage. Additionally, their method may fail to address out-of-distribution + +data, because including all real-world conditions in training, considering both weather types and image statistics, is impractical. Contrastively, we introduce test-time adaptation and update knowledge by practice to remove arbitrary weather conditions robustly. We design and incorporate Diffusion Tubelet Self-Calibration (Diff-TSC), a proxy task, into the iterative denoising of the diffusion reverse process to learn the primer distribution of the frames degraded by unknown weather. + +Algorithm 2: Diffusion Test-Time Adaptation to unknown weather. $\theta$ is the weight set of the trained network before adaptation, $\delta$ is the learning rate for online adaptation. +Input: $K$ overlapped clips: $\{\mathbf{V}_1,\mathbf{V}_2,\dots ,\mathbf{V}_K\}$ in one video stream +for $Clp k = 1,2,\ldots ,K$ sequentially do +Initialize the network $\epsilon_{\bar{\theta}}$ with $\theta$ +Initialize $\hat{\mathbf{V}}_k$ by Algorithm 1 +if $k = 1$ then +for step $t = T,\dots ,2,1$ sequentially do + $|\hat{\mathbf{V}}_k = ddim(\mathbf{V}_k,\hat{\mathbf{V}}_k,\epsilon_{\bar{\theta}},t)$ +end +else +A, $\hat{\mathbf{A}} =$ TubeletCrop(Vk-1,Vk-1,Na) for step $t = T,\dots ,2,1$ sequentially do +Compute $\mathcal{L}_S(\mathbf{A},\hat{\mathbf{A}},\bar{\epsilon},t)$ by Eq. (5) $\bar{\theta} = \bar{\theta} -\delta \nabla_{\bar{\theta}}\mathcal{L}_{S}$ $\hat{\mathbf{V}}_k = ddim(\mathbf{V}_k,\hat{\mathbf{V}}_k,\epsilon_{\bar{\theta}},t)$ +end + $\hat{\mathbf{V}}_k =$ integrate( $\hat{\mathbf{V}}_k,\hat{\mathbf{V}}_{k - 1})$ +end +Return the restored clips $\{\hat{\mathbf{V}}_1,\hat{\mathbf{V}}_2,\dots ,\hat{\mathbf{V}}_K\}$ + +Diffusion Tubelet Self-Calibration. We design a proxy task for self-supervised learning to transfer the diffusion model to the specific weather condition. Specifically, for each clip from the video stream, we first initialize the weight set of $\epsilon_{\bar{\theta}}$ with $\theta$ trained by source data. We leverage the previous clip $\mathbf{V}_{k - 1}$ and its restored result $\hat{\mathbf{V}}_{k - 1}$ as the pair of condition and input for the online optimization when inferring the current clip $\mathbf{V}_k$ . Rather than directly traversing all fine-grained details, we randomly crop $N_{a}$ tubelets from $\mathbf{V}_{k - 1}$ as the primer distribution to avoid memory overhead and massive computational costs caused by high resolution. Following the process in Section 4.1, we conduct our temporal diffusion process on the restored tubelets $\hat{\mathbf{A}}$ and feed them together with the degraded tubelets $\mathbf{A}$ into the denoising NAFNet for noise estimation. As suggested in [40], it is necessary to iteratively find a relatively optimal point for the target distribution during test-time adaptation. Inher + +Table 1. Quantitative evaluation on seen weather conditions for video adverse weather removal. For Original Weather, these methods are trained on the weather-specific training set and tested on the weather-specific testing set. For Rain, Haze, and Snow, these methods are trained on a mixed training set and tested on the weather-specific testing set. The average performance is calculated on Rain, Haze, and Snow. PSNR and SSIM are adopted as our evaluation metrics. The best and second-best values are bold and underlined. + +
MethodsTypeSourceDatasets
Original WeatherRainHazeSnowAverage
DerainPReNet [34]ImageCVPR'1927.060.907726.800.881417.640.803028.570.940124.340.8748
SLDNet [50]VideoCVPR'2020.310.627221.240.712916.210.756122.010.855019.820.7747
S2VD [53]VideoCVPR'2124.090.794428.390.900619.650.860726.230.919024.760.8934
RDD-Net [42]VideoECCV'2231.820.942330.340.930018.360.843230.400.956026.370.9097
DehazeGDN [26]ImageICCV'1919.690.854529.960.937019.010.880531.020.951826.660.9231
MSBDN [15]ImageCVPR'2022.010.875926.700.914622.240.904727.070.934025.340.9178
VDHNet [35]VideoTIP'1916.640.813329.870.927216.850.821429.530.939525.420.8960
PM-Net [28]VideoMM'2223.830.895025.790.888023.570.914318.710.788122.690.8635
DesnowDesnowNet [29]ImageTIP'1828.300.953025.190.878616.430.790227.560.918123.060.8623
DDMSNET [56]ImageTIP'2132.550.961329.010.918819.500.861532.430.969426.980.9166
HDCW-W-Net [10]ImageICCV'2131.770.954228.100.905517.360.792131.050.948225.500.8819
SMGARN [13]ImageTCSVT'2233.240.972127.780.910017.850.807532.340.966825.990.8948
RestorationMPRNet [54]ImageCVPR'21————28.220.916520.250.893430.950.948226.470.9194
EDVR [43]VideoCVPR'19————31.100.937119.670.872430.270.944027.010.9178
RVRT [23]VideoNIPS'22————30.110.913221.160.894926.780.883426.020.8972
RTA [61]VideoCVPR'22————30.120.918620.750.891529.790.936726.890.9156
Multi-WeatherAll-in-one [21]ImageCVPR'20————26.620.894820.880.901030.090.943125.860.9130
UVRNet [19]ImageTMM'22————22.310.767820.820.857524.710.887322.610.8375
TransWeather [38]ImageCVPR'22————26.820.911822.170.902528.870.931325.950.9152
TKL [11]ImageCVPR'22————26.730.893522.080.904431.350.951526.720.9165
WeatherDiffusion [33]ImageTPAMI'23————25.860.912520.100.844226.400.911324.120.8893
WGWS-Net [62]ImageCVPR'23————29,640.931017.710.811331.580.952826.310.9265
ViWS-Net [51]VideoICCV'23————31.520.943324.510.918731.490.956229.170.9394
OursVideo——————32.430.957324.560.914831.860.964029.630.9453
+ +ited from the characteristics of DDIM, such requirement can be easily achieved in the iterative diffusion reverse process, which needs more increments of inference time in the traditional CNNs and Transformers. For each timestep $t$ , we first obtain the noisy tubelet $\hat{\mathbf{A}}_t$ and then optimize the objective function given in Eq. (5) to update the weight set $\bar{\theta}$ . After the iterative adaptation to $\hat{\mathbf{V}}_k$ , we ensemble the overlapped part of it with that of the previous clip $\hat{\mathbf{V}}_{k - 1}$ to capture complementary information of pre-adaptation and post-adaptation. We optimize over the entire network except for the last layer suggested by SHOT [22]. The detailed procedure is described in Algorithm 2. We emphasize the synergy between diffusion models and test-time adaptation in accumulating frame-correlated knowledge for adverse weather removal. By attaching the proxy task to iterative denoising, our approach also avoids excessive computational overhead caused by iterative adaptation. + +# 5. Experiments + +# 5.1. Datasets + +Various video adverse weather datasets are used in our experiments. Following the setting of [51], we adopt RainMotion [42], REVIDE [59] and KITTI-snow [51] as seen weather. For the training stage, we merge the training set of the three datasets as the mixed set to learn a generic model. For the testing stage, we evaluate our model on three testing sets, respectively. Additionally, we evaluate our approach on two datasets VRDS [44] and RVSD [4] to validate the generalization of our Diff-TTA on unseen + +Table 2. Quantitative evaluation on unseen weather conditions for video adverse weather removal. The best and second-best performances are marked in bold and underlined, respectively. + +
MethodVRDS [44]RVSD [4]
PSNR ↑SSIM ↑PSNR ↑SSIM ↑
All-in-one [21]20.440.594419.790.7509
TransWeather [38]21.360.713620.250.7514
TKL [11]20.490.700319.710.7370
WeatherDiffusion [33]20.730.694318.080.6588
ViWS-Net [51]21.570.709419.830.7590
Diff-TTA (ours)22.570.728122.350.7719
+ +weather conditions. VRDS is a synthesized video dataset of joint rain streaks and raindrops with a total of 102 videos, while RVSD is a realistic video desnowing dataset with a total of 110 videos containing both snow and fog achieved by the rendering engine. Due to the large quantity of the two datasets, we randomly sample 8 videos as the test set, each of which has 20 frames, respectively. We also collect several real-world videos degraded by diverse weather conditions to prove its effectiveness on real-world applications. + +# 5.2. Implementation + +Training Details. The proposed framework was trained on NVIDIA RTX 4090 GPUs and implemented on the Pytorch platform. We adopted the Lion optimizer [12] ( $\beta_{1} = 0.9$ , $\beta_{2} = 0.99$ ) with the initial learning rate of $2 \times 10^{-5}$ decayed to $2 \times 10^{-7}$ by the Cosine scheduler. We randomly crop the video frames to $256 \times 256$ for training. Our framework is empirically trained for $600k$ iterations in an end-to-end way with batch size of 4. Each batch consists of 4 video + +![](images/c4ef1525402fa617c8c4fc37c9fc532693fa65f894364c12d715d6feb5a4cd38.jpg) +Input Frame + +![](images/367d8b11e4217833f1df827769e1dede565c28cd3cdb7e02e0c3262a33cbafe0.jpg) +EDVR + +![](images/e4e0670ed6db40c128cae71d709d86e2d8697b607197ba7d2c10c6176427121b.jpg) +TransWeather + +![](images/06ec1ec474e98cf6df4aedb1e18c5c14342d37093ee7787396fe35fafa9ec250.jpg) +TKL +Figure 4. Qualitative Comparison of seen weather conditions between our approach and state-of-the-art algorithms. The competitive algorithms are selected to present the results on the example frames degraded by rain, haze, snow, respectively. The color box indicates the comparison of details. Please zoom in on the images for improved visualization. + +![](images/4c444ee7d709fc7617d4dbba15a6b83fb2c3c066d0d36d4b7b7ac59759ae87b0.jpg) +WeatherDiffusion + +![](images/c51639beb9c788c6e37854cb62cc3cc81c46c2cd44bf59f7507afd04bc9f7c1c.jpg) +ViWS-Net + +![](images/d71a8e9bcaff4b45039c96f8d47d118073fcc90da1686e6952c81ae61a106715.jpg) +Ours + +![](images/82476afc385bf30fd7b32d0c730b1503ab2b5f09f58730dc8f49414a48683f21.jpg) +GT + +clips with 5 frames per clip. + +Method-specific Details. During the training, we used the diffusion process linearly increasing from $\beta_{1} = 0.0001$ to $\beta_{T} = 0.02$ with $T = 1000$ timesteps. During the testing, the inference step is reduced to 25 by deterministic implicit sampling. We utilize the same optimizer with a smaller learning rate of $1\times 10^{-6}$ for the test-time adaptation. The number of cropped tubelets $N_{a}$ is empirically set to 5 while the crop size is $128\times 128$ . Despite the incorporated adaptation, our Diff-TTA is particularly $90\times$ more efficient than WeatherDiffusion. To process a video clip of 5 frames with the patch size of $256\times 256$ , the average run time of Diff-TTA is 6.01s while that of WeatherDiffusion is 542.76s. + +# 5.3. Performance Evaluation + +Comparison methods. As shown in Table 1, we compared our proposed method against five kinds of state-of-the-art methods on the three seen weather conditions as [51] does, i.e., derain, dehaze, desnow, restoration, all-in-one adverse weather removal. For all-in-one adverse weather removal, we compared ours with the six representative single-image methods All-in-one [21], UVRNet [19], TransWeather [38], TKL [11], WeatherDiffusion [33], WGWS-Net [62], and the video-level method ViWS-Net [51]. + +Quantitative Comparison. For quantitative evaluation of the restored results, we apply the peak signal-to-noise ratio (PSNR) and the structural similarity (SSIM) as the metrics. Following the previous experimental setting, only the results of the model trained on the mixed training set are reported for all-in-one adverse weather removal models. To ensure the fairness of comparisons, we retrain each compared model implemented by the official codes based on our training dataset and report the best result. Observed from Table 1, our method achieves the best average performance across rain, haze and snow conditions among all kinds of methods by a considerable margin of 0.46, 0.0059 in PSNR, SSIM, respectively, than the second-best method ViWS-Net [51]. Though some single-weather + +methods like DDMSNET [56] and SMGARN [13] achieve impressive results in their own weather condition, they frequently fail to remove other weather conditions, such as haze. Our method significantly surpasses other all-in-one adverse weather methods in rain and snow conditions, and obtains competitive results in haze conditions. It is worth noting that the generalizable ability is critical for such methods, which can be observed from the average performance. To further validate their generalization, we evaluate all-in-on adverse weather removal methods on unseen weather conditions, i.e., rain streak+raindrop and snow+fog, in Table 2. Our Diff-TTA exhibits remarkable superiority over state-of-the-art methods when confronted with these unseen weather conditions. + +Visual comparisons on synthetic videos. To vividly demonstrate the effectiveness of our approach, Figure 4 shows the qualitative comparison under the rain, haze, and snow conditions between our method and five state-of-the-art methods. Our method demonstrates promising results in terms of visual quality across various weather types. Specifically, when it comes to rain and snow scenarios, our approach significantly reduces the presence of rain streaks and snow particles, outperforming other methods. In hazy scenarios, our method effectively removes residual haze and preserves the clean background with impressive results. More visual results of seen and unseen weather are displayed in the supplementary files. + +Visual comparisons on real-world videos. To evaluate the universality of our Diff-TTA on real-world applications, we collect several real-world degraded videos by rain, haze, snow from Youtube website, and further compare our approach against other all-in-one adverse weather removal methods. Figure 5 displays the visual results produced by our network and the compared methods on real-world video frames, clearly showcasing our exceptional performance. + +Ablation Study. We evaluate the effectiveness of each component including diffusion process, temporal noise model, and Diffusion test-time adaptation (Diff-TTA) as described + +Table 3. Ablation study of each critical component in the proposed framework on seen weather conditions. The top values are marked in bold font. "Diff-TTA" denote Diffusion Test-time Adaptation. + +
CombinationComponentDatasets
Diffusion ProcessTemporal NoiseDiff-TTARainHazeSnowAverage
M1---29.070.951422.640.893028.790.9350
M2✓--31.550.946623.580.904130.480.9520
M3✓✓-32.100.953024.310.912431.040.9621
Ours✓✓✓32.430.957324.560.914831.860.9640
+ +![](images/1e0c27c42d2afe3f072b88bae1b21288b97b4be2c831ad7ed84afeac7a982c47.jpg) +Input Frame + +![](images/62b1cab1b0b1b1f9ca1d318d6547b8bb4df3ede760bf741977af5eff23bac9a6.jpg) +TransWeather + +![](images/7951abc98c6172b2b2c0d5eb584958fe0fe1ca24a689d1836c1e229f6f1beb0d.jpg) +TKL + +![](images/7b6d3a7397f89a7ad381452dae980d8886a93b5e119d01fa335b19497e19b340.jpg) +WeatherDiffusion + +![](images/83527c929b785e3c02ead6bed4f5ae637df96fa20edc43fa797b969155f139ab.jpg) +ViWS-Net + +![](images/0de0f733029bbfbeb11d2510974acd38d2ffe067038064e88bfb1f2cfe8986de.jpg) +Ours + +![](images/f3f626ef84e63794ca748132c2f7d9905f6dd84c482546657559fe53aa77cc67.jpg) +Figure 5. Visual comparison of all-in-one adverse weather removal methods on the selected real-world video sequences degraded by rain, haze, snow. Apparently, our network can more effectively remove rain streaks, haze, and snowflakes of input video frames than state-of-the-art methods. Please zoom in on the images for improved visualization. + +![](images/8aad17c2ef787788a2c858fcce686faa1bc258f89443ca9cdbe8462ac8f82d68.jpg) +Figure 6. Detailed results on each case of VRDS and RVSD by our proposed Diff-TTA. The proposed Diff-TTA brings consistent improvement across various cases. + +in Table 3. We present the results trained on the mixed training set and tested on three weather-specific testing sets (rain, haze, snow) in view of PSNR and SSIM. The baseline model M1, i.e., the vanilla NAFNet, achieves the average performance on the three adverse weather datasets of 26.83, 0.9264 in PSNR, SSIM, respectively. M2 is equipped with diffusion process and trained by noise estimation, which advances M1 on the average performance by 1.69, 0.0078 of PSNR, SSIM, respectively. M3 introduces the proposed temporal noise model by replacing the vanilla noise model, and brings about an improvement of 0.63, 0.0083 in average PSNR, SSIM, respectively. Finally, our full model employs diffusion test-time adaptation (Diff-TTA) and gains a critical increase of 0.48, 0.0028 in PSNR, SSIM, respectively, compared to M3. Figure 6 further ablates our iterative adaptation on each case of the two unseen datasets, demonstrat + +ing consistent improvements. + +# 6. Conclusion + +This paper presents Diff-TTA, a diffusion-based test-time adaptation method capable of restoring videos degraded by various, even unknown, weather conditions. The proposed approach involves constructing a diffusion-based network and utilizing a novel temporal noise model to establish correlations among video frames. More importantly, a proxy task Diff-TSC is incorporated into the diffusion reverse process to iteratively update the deep model. This allows for the efficient adaptation of vision systems in real-world scenarios without modifying the source training process, making it easily applicable to a wide range of off-the-shelf diffusion-based models for enhancing weather removal. Experimental results on both synthetic and real-world data, which encompass diverse weather conditions, confirm the effectiveness and generalization capability of the proposed approach. We believe Diff-TTA can be a compelling baseline that shifts the focus of the community from the complicated training process to online test-time adaptation. + +Acknowledgment. This work is supported by Guangzhou Municipal Science and Technology Project (Grant No. 2023A03J0671), Guangzhou-HKUST(GZ) Joint Funding Program (No. 2024A03J0618), a grant of General Research Fund from Hong Kong Research Grants Council (project no. 15218521), and the InnoHK funding launched by Innovation and Technology Commission, Hong Kong SAR. + +# References + +[1] Malik Boudiaf, Romain Mueller, Ismail Ben Ayed, and Luca Bertinetto. Parameter-free online test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8344-8353, 2022. 3 +[2] George Box. Box and jenkins: time series analysis, forecasting and control. In A Very British Affair: Six Britons and the Development of Time Series Analysis During the 20th Century, pages 161-215. Springer, 2013. 4 +[3] Dian Chen, Dequan Wang, Trevor Darrell, and Sayna Ebrahimi. Contrastive test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 295-305, 2022. 3 +[4] Haoyu Chen, Jingjing Ren, Jinjin Gu, Hongtao Wu, Xuequan Lu, Haoming Cai, and Lei Zhu. Snow removal in video: A new dataset and a novel method. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13211-13222, 2023. 1, 2, 6 +[5] Jie Chen, Cheen-Hau Tan, Junhui Hou, Lap-Pui Chau, and He Li. Robust video content alignment and compensation for rain removal in a cnn framework. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6286-6295, 2018. 2 +[6] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European Conference on Computer Vision, pages 17-33. Springer, 2022. 4 +[7] Sixiang Chen, Tian Ye, Yun Liu, Jinbin Bai, Haoyu Chen, Yunlong Lin, Jun Shi, and Erkang Chen. CPLformer: Cross-scale prototype learning transformer for image snow removal. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4228-4239, 2023. 2 +[8] Sixiang Chen, Tian Ye, Chenghao Xue, Haoyu Chen, Yun Liu, Erkang Chen, and Lei Zhu. Uncertainty-driven dynamic degradation perceiving and background modeling for efficient single image desnowing. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4269-4280, 2023. +[9] Wei-Ting Chen, Hao-Yu Fang, Jian-Jiun Ding, Cheng-Che Tsai, and Sy-Yen Kuo. Jstasr: Joint size and transparency-aware snow removal algorithm based on modified partial convolution and veiling effect removal. In European Conference on Computer Vision, pages 754-770. Springer, 2020. 2 +[10] Wei-Ting Chen, Hao-Yu Fang, Cheng-Lin Hsieh, Cheng-Che Tsai, I Chen, Jian-Jiun Ding, Sy-Yen Kuo, et al. All snow removed: Single image desnowing algorithm using hierarchical dual-tree complex wavelet representation and contradict channel loss. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4196-4205, 2021. 6 +[11] Wei-Ting Chen, Zhi-Kai Huang, Cheng-Che Tsai, Hao-Hsiang Yang, Jian-Jiun Ding, and Sy-Yen Kuo. Learning multiple adverse weather removal via two-stage knowledge learning and multi-contrastive regularization: Toward a unified model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17653–17662, 2022. 1, 2, 6, 7 + +[12] Xiangning Chen, Chen Liang, Da Huang, Esteban Real, Kaiyuan Wang, Yao Liu, Hieu Pham, Xuanyi Dong, Thang Luong, Cho-Jui Hsieh, et al. Symbolic discovery of optimization algorithms. arXiv preprint arXiv:2302.06675, 2023.6 +[13] Bodong Cheng, Juncheng Li, Ying Chen, Shuyi Zhang, and Tieyong Zeng. Snow mask guided adaptive residual network for image snow removal. arXiv preprint arXiv:2207.04754, 2022.6, 7 +[14] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 1, 3 +[15] Hang Dong, Jinshan Pan, Lei Xiang, Zhe Hu, Xinyi Zhang, Fei Wang, and Ming-Hsuan Yang. Multi-scale boosted de-hazing network with dense feature fusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2157-2167, 2020. 6 +[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 3 +[17] Minhao Hu, Tao Song, Yujun Gu, Xiangde Luo, Jieneng Chen, Yinan Chen, Ya Zhang, and Shaoting Zhang. Fully test-time adaptation for image segmentation. In Medical Image Computing and Computer Assisted Intervention—MICCAI 2021: 24th International Conference, Strasbourg, France, September 27–October 1, 2021, Proceedings, Part III 24, pages 251–260. Springer, 2021. 3 +[18] Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for model-agnostic domain generalization. Advances in Neural Information Processing Systems, 34:2427-2440, 2021. 2, 3 +[19] Ashutosh Kulkarni, Prashant W Patil, Subrahmanyam Murala, and Sunil Gupta. Unified multi-weather visibility restoration. IEEE Transactions on Multimedia, 2022. 6, 7 +[20] Minghan Li, Qi Xie, Qian Zhao, Wei Wei, Shuhang Gu, Jing Tao, and Deyu Meng. Video rain streak removal by multiscale convolutional sparse coding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6644-6653, 2018. 1, 2 +[21] Ruoteng Li, Robby T Tan, and Loong-Fah Cheong. All in one bad weather removal using architectural search. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3175-3185, 2020. 1, 2, 6, 7 +[22] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In International conference on machine learning, pages 6028-6039. PMLR, 2020. 6 +[23] Jingyun Liang, Yuchen Fan, Xiaoyu Xiang, Rakesh Ranjan, Eddy Ilg, Simon Green, Jiezhang Cao, Kai Zhang, Radu Timofte, and Luc Van Gool. Recurrent video restoration transformer with guided deformable attention. arXiv preprint arXiv:2206.02146, 2022. 6 +[24] Jiaying Liu, Wenhan Yang, Shuai Yang, and Zongming Guo. Erase or fill? deep joint recurrent rain removal and reconstruction in videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3233-3242, 2018. 2 + +[25] Quande Liu, Cheng Chen, Qi Dou, and Pheng-Ann Heng. Single-domain generalization in medical image segmentation via test-time adaptation from shape dictionary. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1756–1764, 2022. 3 +[26] Xiaohong Liu, Yongrui Ma, Zhihao Shi, and Jun Chen. Grid-dehazenet: Attention-based multi-scale network for image dehazing. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7314-7323, 2019. 6 +[27] Xihui Liu, Dong Huk Park, Samaneh Azadi, Gong Zhang, Arman Chopikyan, Yuxiao Hu, Humphrey Shi, Anna Rohrbach, and Trevor Darrell. More control for free! image synthesis with semantic diffusion guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 289-299, 2023. 3 +[28] Ye Liu, Liang Wan, Huazhu Fu, Jing Qin, and Lei Zhu. Phase-based memory network for video dehazing. In Proceedings of the 30th ACM International Conference on Multimedia, pages 5427-5435, 2022. 2, 6 +[29] Yun-Fu Liu, Da-Wei Jaw, Shih-Chia Huang, and Jenq-Neng Hwang. Desnownet: Context-aware deep network for snow removal. IEEE Transactions on Image Processing, 27(6): 3064–3073, 2018. 6 +[30] Ziwei Luo, Fredrik K Gustafsson, Zheng Zhao, Jens Sjolund, and Thomas B Schon. Image restoration with meanreverting stochastic differential equations. arXiv preprint arXiv:2301.11699, 2023. 3 +[31] A Tuan Nguyen, Thanh Nguyen-Tang, Ser-Nam Lim, and Philip HS Torr. Tipi: Test time adaptation with transformation invariance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24162-24171, 2023. 3 +[32] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yafo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. In International conference on machine learning, pages 16888-16905. PMLR, 2022. 3 +[33] Ozan Özdenizci and Robert Legenstein. Restoring vision in adverse weather conditions with patch-based denoising diffusion models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1, 3, 6, 7 +[34] Dongwei Ren, Wangmeng Zuo, Qinghua Hu, Pengfei Zhu, and Deyu Meng. Progressive image deraining networks: A better and simpler baseline. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3937-3946, 2019. 6 +[35] Wenqi Ren, Jingang Zhang, Xiangyu Xu, Lin Ma, Xiaochun Cao, Gaofeng Meng, and Wei Liu. Deep video dehazing with semantic segmentation. IEEE transactions on image processing, 28(4):1895-1908, 2018. 2, 6 +[36] Jaskirat Singh, Stephen Gould, and Liang Zheng. High-fidelity guided image synthesis with latent diffusion models. arXiv preprint arXiv:2211.17084, 2022. 1, 3 +[37] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2, 3 + +[38] Jeya Maria Jose Valanarasu, Rajeev Yasarla, and Vishal M Patel. Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2353-2363, 2022. 1, 2, 6, 7 +[39] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 2 +[40] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726, 2020. 2, 3, 5 +[41] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7201-7211, 2022. 3 +[42] Shuai Wang, Lei Zhu, Huazhu Fu, Jing Qin, Carola-Bibiane Schonlieb, Wei Feng, and Song Wang. Rethinking video rain streak removal: A new synthesis model and a deraining network with video rain prior. In European Conference on Computer Vision, pages 565-582. Springer, 2022. 2, 6 +[43] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced deformable convolutional networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0-0, 2019. 6 +[44] Hongtao Wu, Yijun Yang, Haoyu Chen, Jingjing Ren, and Lei Zhu. Mask-guided progressive network for joint raindrop and rain streak removal in videos. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7216-7225, 2023. 6 +[45] Bin Xia, Yulun Zhang, Shiyin Wang, Yitong Wang, Xinglong Wu, Yapeng Tian, Wenming Yang, and Luc Van Gool. Diffir: Efficient diffusion model for image restoration. arXiv preprint arXiv:2303.09472, 2023. 3 +[46] Jiaqi Xu, Xiaowei Hu, Lei Zhu, Qi Dou, Jifeng Dai, Yu Qiao, and Pheng-Ann Heng. Video dehazing via a multi-range temporal alignment network with physical prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18053–18062, 2023. 2 +[47] Wending Yan, Robby T Tan, Wenhan Yang, and Dengxin Dai. Self-aligned video deraining with transmission-depth consistency. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11966-11976, 2021. 2 +[48] Hongzheng Yang, Cheng Chen, Meirui Jiang, Quande Liu, Jianfeng Cao, Pheng Ann Heng, and Qi Dou. Dltta: Dynamic learning rate for test-time adaptation on cross-domain medical images. IEEE Transactions on Medical Imaging, 41 (12):3575-3586, 2022. 3 +[49] Wenhan Yang, Jiaying Liu, and Jiashi Feng. Frameconsistent recurrent video deraining with dual-level flow. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1661-1670, 2019. 2 +[50] Wenhan Yang, Robby T Tan, Shiqi Wang, and Jiaying Liu. Self-learning video rain streak removal: When cyclic consistency meets temporal correspondence. In Proceedings of + +the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1720-1729, 2020. 2, 6 +[51] Yijun Yang, Angelica I Aviles-Rivero, Huazhu Fu, Ye Liu, Weiming Wang, and Lei Zhu. Video adverse-weather-component suppression network via weather messenger and adversarial backpropagation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13200-13210, 2023. 1, 3, 5, 6, 7 +[52] Tian Ye, Sixiang Chen, Jinbin Bai, Jun Shi, Chenghao Xue, Jingxia Jiang, Junjie Yin, Erkang Chen, and Yun Liu. Adverse weather removal with codebook priors. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12653-12664, 2023. 3 +[53] Zongsheng Yue, Jianwen Xie, Qian Zhao, and Deyu Meng. Semi-supervised video deraining with dynamical rain generator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 642-652, 2021. 2, 6 +[54] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14821-14831, 2021. 6 +[55] Jian Zhang, Lei Qi, Yinghuan Shi, and Yang Gao. Domain adaptor: A novel approach to test-time adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18971-18981, 2023. 3 +[56] Kaihao Zhang, Rongqing Li, Yanjiang Yu, Wenhan Luo, and Changsheng Li. Deep dense multi-scale network for snow removal using semantic and depth priors. IEEE Transactions on Image Processing, 30:7419-7431, 2021. 2, 6, 7 +[57] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 3 +[58] Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test time robustness via adaptation and augmentation. Advances in Neural Information Processing Systems, 35: 38629-38642, 2022. 2, 3 +[59] Xinyi Zhang, Hang Dong, Jinshan Pan, Chao Zhu, Ying Tai, Chengjie Wang, Jilin Li, Feiyue Huang, and Fei Wang. Learning to restore hazy video: A new real-world dataset and a new method. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9239-9248, 2021. 1, 2, 6 +[60] Yi Zhang, Xiaoyu Shi, Dasong Li, Xiaogang Wang, Jian Wang, and Hongsheng Li. A unified conditional framework for diffusion-based image restoration. arXiv preprint arXiv:2305.20049, 2023. 3 +[61] Kun Zhou, Wenbo Li, Liying Lu, Xiaoguang Han, and Jiangbo Lu. Revisiting temporal alignment for video restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6053-6062, 2022. 6 +[62] Yurui Zhu, Tianyu Wang, Xueyang Fu, Xuanyu Yang, Xin Guo, Jifeng Dai, Yu Qiao, and Xiaowei Hu. Learning weather-general and weather-specific features for image + +restoration under multiple adverse weather conditions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21747-21758, 2023. 3, 6, 7 \ No newline at end of file diff --git a/2024/Genuine Knowledge from Practice_ Diffusion Test-Time Adaptation for Video Adverse Weather Removal/images.zip b/2024/Genuine Knowledge from Practice_ Diffusion Test-Time Adaptation for Video Adverse Weather Removal/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3ab158a608d76eec4d2d647ae0f4d4f5f1bb3695 --- /dev/null +++ b/2024/Genuine Knowledge from Practice_ Diffusion Test-Time Adaptation for Video Adverse Weather Removal/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:31deab2c61d3b5d0a02c17dbdcc609a2e18d4ea4f4b1512f629addd53565bb62 +size 666031 diff --git a/2024/Genuine Knowledge from Practice_ Diffusion Test-Time Adaptation for Video Adverse Weather Removal/layout.json b/2024/Genuine Knowledge from Practice_ Diffusion Test-Time Adaptation for Video Adverse Weather Removal/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..dcb8a1f11f7d38d8db969ed4a1a7945f5d5c40c2 --- /dev/null +++ b/2024/Genuine Knowledge from Practice_ Diffusion Test-Time Adaptation for Video Adverse Weather Removal/layout.json @@ -0,0 +1,9519 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 83, + 103, + 510, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 103, + 510, + 140 + ], + "spans": [ + { + "bbox": [ + 83, + 103, + 510, + 140 + ], + "type": "text", + "content": "Genuine Knowledge from Practice: Diffusion Test-Time Adaptation for Video Adverse Weather Removal" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 73, + 160, + 523, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 160, + 523, + 176 + ], + "spans": [ + { + "bbox": [ + 73, + 160, + 523, + 176 + ], + "type": "text", + "content": "Yijun Yang" + }, + { + "bbox": [ + 73, + 160, + 523, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 73, + 160, + 523, + 176 + ], + "type": "text", + "content": ", Hongtao Wu" + }, + { + "bbox": [ + 73, + 160, + 523, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 73, + 160, + 523, + 176 + ], + "type": "text", + "content": ", Angelica I. Aviles-Rivero" + }, + { + "bbox": [ + 73, + 160, + 523, + 176 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 73, + 160, + 523, + 176 + ], + "type": "text", + "content": ", Yulun Zhang" + }, + { + "bbox": [ + 73, + 160, + 523, + 176 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 73, + 160, + 523, + 176 + ], + "type": "text", + "content": ", Jing Qin" + }, + { + "bbox": [ + 73, + 160, + 523, + 176 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 73, + 160, + 523, + 176 + ], + "type": "text", + "content": ", Lei Zhu" + }, + { + "bbox": [ + 73, + 160, + 523, + 176 + ], + "type": "inline_equation", + "content": "^{1,5*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 112, + 177, + 481, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 177, + 481, + 189 + ], + "spans": [ + { + "bbox": [ + 112, + 177, + 481, + 189 + ], + "type": "text", + "content": "1The Hong Kong University of Science and Technology (Guangzhou) 2University of Cambridge 3ETH Zürich" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 191, + 459, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 191, + 459, + 203 + ], + "spans": [ + { + "bbox": [ + 137, + 191, + 459, + 203 + ], + "type": "text", + "content": "4The Hong Kong Polytechnic University 5The Hong Kong University of Science and Technology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 149, + 204, + 443, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 204, + 443, + 216 + ], + "spans": [ + { + "bbox": [ + 149, + 204, + 443, + 216 + ], + "type": "text", + "content": "Project page: https://github.com/scott-yjyang/DiffTTA" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 243, + 192, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 243, + 192, + 255 + ], + "spans": [ + { + "bbox": [ + 143, + 243, + 192, + 255 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 269, + 289, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 269, + 289, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 269, + 289, + 604 + ], + "type": "text", + "content": "Real-world vision tasks frequently suffer from the appearance of unexpected adverse weather conditions, including rain, haze, snow, and raindrops. In the last decade, convolutional neural networks and vision transformers have yielded outstanding results in single-weather video removal. However, due to the absence of appropriate adaptation, most of them fail to generalize to other weather conditions. Although ViWS-Net is proposed to remove adverse weather conditions in videos with a single set of pretrained weights, it is seriously blinded by seen weather at train-time and degenerates when coming to unseen weather during test-time. In this work, we introduce test-time adaptation into adverse weather removal in videos, and propose the first framework that integrates test-time adaptation into the iterative diffusion reverse process. Specifically, we devise a diffusion-based network with a novel temporal noise model to efficiently explore frame-correlated information in degraded video clips at training stage. During inference stage, we introduce a proxy task named Diffusion Tubelet Self-Calibration to learn the primer distribution of test video stream and optimize the model by approximating the temporal noise model for online adaptation. Experimental results, on benchmark datasets, demonstrate that our Test-Time Adaptation method with Diffusion-based network(Diff-TTA) outperforms state-of-the-art methods in terms of restoring videos degraded by seen weather conditions. Its generalizable capability is validated with unseen weather conditions in synthesized and real-world videos." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 627, + 127, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 627, + 127, + 639 + ], + "spans": [ + { + "bbox": [ + 47, + 627, + 127, + 639 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 647, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 287, + 696 + ], + "type": "text", + "content": "Adverse weather like rain, snow, and haze is very common in many existing outdoor videos. The presence of adverse weather usually decreases the visibility of the captured videos and significantly impairs the performance of" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 309, + 242, + 547, + 389 + ], + "blocks": [ + { + "bbox": [ + 309, + 242, + 547, + 389 + ], + "lines": [ + { + "bbox": [ + 309, + 242, + 547, + 389 + ], + "spans": [ + { + "bbox": [ + 309, + 242, + 547, + 389 + ], + "type": "image", + "image_path": "293ac8b6fdcba7f56490225645c32bd6b578be91b5edf728c832968484f0e142.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 398, + 545, + 464 + ], + "lines": [ + { + "bbox": [ + 305, + 398, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 305, + 398, + 545, + 464 + ], + "type": "text", + "content": "Figure 1. Overview of the existing all-in-one adverse weather removal methods. Our approach can achieve superior performance not only in seen weather conditions but also in unseen weather conditions with a single set of pre-trained weights by diffusion test-time adaptation. In particular, our Diff-TTA is " + }, + { + "bbox": [ + 305, + 398, + 545, + 464 + ], + "type": "inline_equation", + "content": "90 \\times" + }, + { + "bbox": [ + 305, + 398, + 545, + 464 + ], + "type": "text", + "content": " more efficient than WeatherDiffusion." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 473, + 546, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 546, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 546, + 664 + ], + "type": "text", + "content": "subsequent high-level vision applications such as object detection, semantic segmentation and autonomous driving. To mitigate these negative effects, Convolutional Neural Networks (CNNs) and Transformers are frequently utilized in the literature. Also, Diffusion models have emerged as the de-facto generative model for image synthesis [14, 16, 36] and have been developed for single-image adverse weather removal [33]. While deep neural networks can achieve decent results on test points within the distribution, real-world applications often demand robustness. This is especially crucial when facing unknown input perturbations, changes in weather conditions, or other sources of distribution shift. Existing single-weather removal approaches [4, 20, 59] require follow-up training to adapt to specific domains. This leads to the switching between different parameter sets, which can make the pipeline cumbersome." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 665, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 547, + 714 + ], + "type": "text", + "content": "Recently, all-in-one adverse weather removal has attracted the attention of the community, which aims to handle diverse weather conditions with a generic model [11, 21, 38, 51]. However, the potential issue of frequently failing to" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 234, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 234, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 234, + 713 + ], + "type": "text", + "content": "*Lei Zhu (leizhu@ust.hk) is the corresponding author." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25606" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 289, + 163 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 289, + 163 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 289, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 289, + 163 + ], + "type": "image", + "image_path": "ee45d853939569f13235d9fc562ce0f4228cfb244b9e8204dd1ba663bc36db4d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 171, + 289, + 250 + ], + "lines": [ + { + "bbox": [ + 46, + 171, + 289, + 250 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 289, + 250 + ], + "type": "text", + "content": "Figure 2. Our Diff-TTA enables weather removal models to overcome unseen weather corruptions. We use t-SNE [39] to visualize features from the last feature extractor layer of each dataset. Obviously, unseen data points tend to approximate the seen ones after adaptation, which means Diff-TTA can categorize unknown degradation into known distribution. ('Real-world' contains video clips simultaneously degraded by fog and snow.)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 258, + 290, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 290, + 486 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 290, + 486 + ], + "type": "text", + "content": "recover from out-of-distribution conditions remains largely unaddressed due to the finiteness of training data and the absence of a test data adaptation mechanism. As shown in Figure 1, generic models suffer significant degradation when handling all the possible weather conditions beyond the scope of their training data. Furthermore, a common aspect among many real-world applications is the necessity for online adaptation with limited data. This is particularly relevant in scenarios such as autonomous vehicles or drones, where a vision model must continually process a video stream and adjust its predictions. To achieve effective adaptation, it is crucial to leverage the incoming data in a way that is beneficial and does not introduce harmful biases, even when there are distribution shifts between the training and test data. Thus, akin to how humans update knowledge through practice, the task of removing any weather condition through online adaptation using a single set of pre-trained weights is both pragmatic and promising, albeit challenging." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 486, + 290, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 290, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 290, + 715 + ], + "type": "text", + "content": "In this paper, we construct the diffusion-based framework for video adverse weather removal. Yet, diffusion models usually entail significant training overhead, hindering the technique's broader adoption in the research community of video tasks. Instead of introducing massive parameters for temporal encoding, we design an efficient strategy based on DDIM [37] to preserve temporal correlation, specifically temporal noise model, in order to alleviate its training load. In contrast to the vanilla noise model, which independently samples noise for each frame from a standard Gaussian distribution, our temporal noise model takes into account the inter-frame relationship via time series models. More importantly, to promptly address unexpected weather conditions, we introduce test-time adaptation techniques by recalibrating the model's parameters using only the degraded test data. The mainstream of test-time adaptation methods [18, 40, 58] manipulate and iteratively update the normalization statistics of CNNs and Vision Transformer using specific proxy tasks at the test phase. However, these" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "content": "methods often fail when transferring to low-level tasks. Additionally, incorporating such iterative adaptation into these models significantly increases the cost of inference, multiplying it by the number of iterations, compared to the original counterpart." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 133, + 547, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 547, + 371 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 547, + 371 + ], + "type": "text", + "content": "In contrast, we highlight the advantage of integrating the diffusion reverse process and test-time adaptation based on their common characteristics of iteration. As far as we know, there has been no work to emerge the possibility of Diffusion models and test-time adaptation. Based on our diffusion-based framework, we design a proxy task and efficiently incorporate it into each denoising iteration when reversing our temporal diffusion process. Specifically, we develop a surrogate task called Diffusion Tubelet Self-Calibration (Diff-TSC), which is supervised by temporal noise model. This is designed to be task-agnostic and does not require any alterations to the training procedure. Rather than directly traversing all fine-grained details, we extract a few tubelets to taste the primer distribution of the frames degraded by unknown weather. This approach facilitates rapid optimization of the diffusion model while also reducing computational costs. As displayed in Figure 2, test-time adaptation helps to improve the model's robustness in real-world scenarios by mitigating the gap between seen and unseen weather corruptions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 372, + 473, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 372, + 473, + 382 + ], + "spans": [ + { + "bbox": [ + 306, + 372, + 473, + 382 + ], + "type": "text", + "content": "Our contributions can be summarized as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 383, + 545, + 538 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 306, + 383, + 545, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 383, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 306, + 383, + 545, + 431 + ], + "type": "text", + "content": "- We introduce the first diffusion-based framework for all-in-one adverse weather removal in videos. This framework efficiently leverages temporal redundancy information through the novel temporal noise model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 431, + 545, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 431, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 306, + 431, + 545, + 479 + ], + "type": "text", + "content": "- To improve the model's robustness against unknown weather, we are the first to introduce test-time adaptation by incorporating a proxy task into the diffusion reverse process to learn the primer distribution of the test data." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 479, + 545, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 545, + 538 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 545, + 538 + ], + "type": "text", + "content": "- We extensively evaluate our approach on seen weather (e.g., rain, haze, snow) and unseen weather (e.g., rain streak+raindrop, snow+fog). The superior performance over synthesized and real-world videos consistently validates its effectiveness and generalizable ability." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 549, + 392, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 549, + 392, + 562 + ], + "spans": [ + { + "bbox": [ + 306, + 549, + 392, + 562 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": "All-in-one Adverse Weather Removal. Over the past decade, The community of low-level vision has paid much attention to single weather removal, e.g., deraining [5, 20, 24, 42, 47, 49, 50, 53], dehazing [28, 35, 46, 59], desnowing [4, 7-9, 56], both on the image level and the video level. Due to the limitations of existing models in adapting to different weather conditions, researchers have recently been investigating the use of a single model instance for multiple adverse weather removal tasks. At the image level, following the pioneering work All-in-one [21], TransWeather [38] and TKL [11] were introduced to tackle many weather conditions with the single encoder and de" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25607" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 300 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 300 + ], + "type": "text", + "content": "coder. Zhu et al. [62] elaborated a two-stage training strategy to explore the weather-general and weather-specific information for the sophisticated removal of diverse weather conditions. Ye et al. [52] introduced high-quality codebook priors learned by a pre-trained VQGAN to restore texture details from adverse weather degradation. The aforementioned methods failed to capture redundant information from temporal space. While they can be adapted for frame-by-frame adverse weather removal, incorporating temporal data in video-level methods provides a more robust solution for dealing with multiple adverse weather conditions in videos. Very recently, Yang et al. [51] introduced 'ViWS-Net', the first video-level approach. This method uses weather messenger tokens to transfer weather knowledge across frames. It also employs adversarial learning to distinguish the weather type, thereby emphasizing weather-invariant information. However, they did not consider online adaptation of the deep model, making the potential collapse for unseen scenarios degraded by unknown weather." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 303, + 289, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 303, + 289, + 567 + ], + "spans": [ + { + "bbox": [ + 47, + 303, + 289, + 567 + ], + "type": "text", + "content": "Diffusion Models. Diffusion model is well-known as a generative modeling approach for its superior achievements in image synthesis and generation tasks [14, 16, 27, 36, 57]. In essence, Denoising Diffusion Probabilistic Models (DDPM) [16] adopted parameterized Markov chain to optimize the lower variational bound on the likelihood function, which can simulate a diffusion process to iteratively improve the quality of target distribution than other generative models. For restoration tasks, Luo et al. [30] applied stochastic differential equations featuring mean reversion in diffusion models to improve low-quality images. DiffIR [45] provided a strong diffusion baseline for image restoration by estimating a compact IR prior representation. Zhang et al. [60] proposed a unified conditional framework based on diffusion models for image restoration by multi-source guidance of a lightweight U-Net. Ozdenizci et al. [33] designed a patch-based diffusion modeling approach called WeatherDiffusion to make the network size-agnostic. While this technique utilizes diffusion models for image weather removal, its very slow inference speed restricts practical use. Nevertheless, there is no work adapting diffusion models to video adverse weather removal." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 570, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 570, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 570, + 289, + 714 + ], + "type": "text", + "content": "Test-time Adaptation. Recently, there has been a growing interest in adapting deep models using exclusively sequential, unlabeled test data. This contrasts with the traditional unsupervised domain adaptation, which asks the access to both training data and a sufficient amount of unlabeled test data. Test-time adaptation techniques allow for model updates using distributional statistics from either a single instance or batch of unlabeled test data. Typically, TENT [40] focuses on minimizing the model's prediction entropy on the target data. T3A [18] adjusts the classifier of a trained source model by computing a pseudo-prototype representation of different classes using unlabeled test data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "text", + "content": "MEMO [58] aims to minimize marginal entropy, compelling the deep model to consistently predict across various augmentations. This approach enforces the invariances of these augmentations and ensures prediction confidence. Chen et al. [3] proposed to apply self-supervised contrastive learning for target feature learning, which was trained jointly with pseudo labeling. Although test-time adaptation methods are increasingly used in image classification [1, 31, 32, 41, 55] and segmentation [17, 25, 48] tasks, their potential in adverse weather removal tasks remains unexplored. In our work, the nature of the iterative sampling process in Diffusion model significantly promotes the adaptation to the target distribution. Video scenarios further allow for the accumulation of the adaptation knowledge for the subsequent frames." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 259, + 486, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 259, + 486, + 272 + ], + "spans": [ + { + "bbox": [ + 305, + 259, + 486, + 272 + ], + "type": "text", + "content": "3. Preliminaries: Diffusion Models" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 279, + 547, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 279, + 547, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 279, + 547, + 437 + ], + "type": "text", + "content": "The diffusion model is composed of a forward and a reverse process. The forward process is defined as a discrete Markov chain of length " + }, + { + "bbox": [ + 304, + 279, + 547, + 437 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 279, + 547, + 437 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 304, + 279, + 547, + 437 + ], + "type": "inline_equation", + "content": "q(x_{1:T}|x_0) = \\prod_{t=1}^{T} q(x_t|x_{t-1})" + }, + { + "bbox": [ + 304, + 279, + 547, + 437 + ], + "type": "text", + "content": ". For each step " + }, + { + "bbox": [ + 304, + 279, + 547, + 437 + ], + "type": "inline_equation", + "content": "t \\in [1,T]" + }, + { + "bbox": [ + 304, + 279, + 547, + 437 + ], + "type": "text", + "content": " in the forward process, a diffusion model adds noise " + }, + { + "bbox": [ + 304, + 279, + 547, + 437 + ], + "type": "inline_equation", + "content": "\\epsilon_t" + }, + { + "bbox": [ + 304, + 279, + 547, + 437 + ], + "type": "text", + "content": " sampled from the Gaussian distribution " + }, + { + "bbox": [ + 304, + 279, + 547, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,\\mathbf{I})" + }, + { + "bbox": [ + 304, + 279, + 547, + 437 + ], + "type": "text", + "content": " to data " + }, + { + "bbox": [ + 304, + 279, + 547, + 437 + ], + "type": "inline_equation", + "content": "x_{t-1}" + }, + { + "bbox": [ + 304, + 279, + 547, + 437 + ], + "type": "text", + "content": " and obtains disturbed data " + }, + { + "bbox": [ + 304, + 279, + 547, + 437 + ], + "type": "inline_equation", + "content": "x_t" + }, + { + "bbox": [ + 304, + 279, + 547, + 437 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 304, + 279, + 547, + 437 + ], + "type": "inline_equation", + "content": "q(x_t|x_{t-1}) = \\mathcal{N}(x_t; \\sqrt{1 - \\beta_t} x_{t-1}, \\beta_t\\mathbf{I})" + }, + { + "bbox": [ + 304, + 279, + 547, + 437 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 279, + 547, + 437 + ], + "type": "inline_equation", + "content": "\\beta_t" + }, + { + "bbox": [ + 304, + 279, + 547, + 437 + ], + "type": "text", + "content": " decides the ratio of noise at timestep " + }, + { + "bbox": [ + 304, + 279, + 547, + 437 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 279, + 547, + 437 + ], + "type": "text", + "content": ". Noticeably, instead of sampling sequentially along the Markov chain, we can sample " + }, + { + "bbox": [ + 304, + 279, + 547, + 437 + ], + "type": "inline_equation", + "content": "x_t" + }, + { + "bbox": [ + 304, + 279, + 547, + 437 + ], + "type": "text", + "content": " at any time step " + }, + { + "bbox": [ + 304, + 279, + 547, + 437 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 279, + 547, + 437 + ], + "type": "text", + "content": " in the closed form via " + }, + { + "bbox": [ + 304, + 279, + 547, + 437 + ], + "type": "inline_equation", + "content": "q(x_t|x_0) = \\mathcal{N}(x_t; \\sqrt{\\bar{\\alpha}_t} x_{t-1}, (1 - \\bar{\\alpha}_t)\\mathbf{I})" + }, + { + "bbox": [ + 304, + 279, + 547, + 437 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 279, + 547, + 437 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t = \\prod_{s=1}^t (1 - \\beta_s)" + }, + { + "bbox": [ + 304, + 279, + 547, + 437 + ], + "type": "text", + "content": ". To parameterize Gaussian distribution, the neural network " + }, + { + "bbox": [ + 304, + 279, + 547, + 437 + ], + "type": "inline_equation", + "content": "\\epsilon_\\theta" + }, + { + "bbox": [ + 304, + 279, + 547, + 437 + ], + "type": "text", + "content": " is introduced and optimized by the objective of DDPM [16]:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 341, + 444, + 545, + 459 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 444, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 341, + 444, + 545, + 459 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {S} = \\left\\| \\epsilon - \\epsilon_ {\\theta} \\left(\\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , t\\right) \\right\\| _ {2} ^ {2}. \\tag {1}", + "image_path": "77ca3eb6385b67e287b7ff288c73928e4342e7642f6918e24fcfc3fd5776b382.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 468, + 547, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 468, + 547, + 566 + ], + "spans": [ + { + "bbox": [ + 304, + 468, + 547, + 566 + ], + "type": "text", + "content": "In the reverse process, the diffusion model gradually de-noises the randomly sampled Gaussian noise to the high-quality output " + }, + { + "bbox": [ + 304, + 468, + 547, + 566 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 304, + 468, + 547, + 566 + ], + "type": "text", + "content": " through the predicted noise by the well-trained " + }, + { + "bbox": [ + 304, + 468, + 547, + 566 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 304, + 468, + 547, + 566 + ], + "type": "text", + "content": ". This process is also defined as a Markov chain: " + }, + { + "bbox": [ + 304, + 468, + 547, + 566 + ], + "type": "inline_equation", + "content": "p_{\\theta}(x_{0:T}) = p(x_T)\\prod_{t=1}^{T}p_{\\theta}(x_{t-1}|x_t)" + }, + { + "bbox": [ + 304, + 468, + 547, + 566 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 468, + 547, + 566 + ], + "type": "inline_equation", + "content": "p_{\\theta}(x_{t-1}|x_t) = \\mathcal{N}(x_{t-1};\\mu_{\\theta}(x_t,t),\\sigma_t^2\\mathbf{I})" + }, + { + "bbox": [ + 304, + 468, + 547, + 566 + ], + "type": "text", + "content": ". The mean and variance are " + }, + { + "bbox": [ + 304, + 468, + 547, + 566 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}(x_t,t) = \\frac{1}{\\sqrt{\\alpha_t}} (x_t - \\frac{\\beta_t}{\\sqrt{1 - \\bar{\\alpha}_t}}\\epsilon_{\\theta}(x_t,t))" + }, + { + "bbox": [ + 304, + 468, + 547, + 566 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 468, + 547, + 566 + ], + "type": "inline_equation", + "content": "\\sigma_t^2 = \\frac{1 - \\bar{\\alpha}_{t-1}}{1 - \\bar{\\alpha}_t}\\beta_t" + }, + { + "bbox": [ + 304, + 468, + 547, + 566 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 575, + 362, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 575, + 362, + 588 + ], + "spans": [ + { + "bbox": [ + 306, + 575, + 362, + 588 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 596, + 547, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 596, + 547, + 643 + ], + "spans": [ + { + "bbox": [ + 304, + 596, + 547, + 643 + ], + "type": "text", + "content": "In this work, we devise the first diffusion-based model to remove arbitrary adverse weather in videos with one set of model parameters. We follow an end-to-end formulation of video adverse weather removal, which reads:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 365, + 651, + 545, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 651, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 365, + 651, + 545, + 669 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {V}} _ {k} = \\phi (\\mathbf {V} _ {k}), \\mathbf {V} _ {k} = \\left\\{I _ {i} \\right\\} _ {i = 1} ^ {N _ {f}}, \\tag {2}", + "image_path": "91b0d6497370d14ea0f9a5258aa9adcdd93aa9494406f0a47fdfd5b88ca1a3ad.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 675, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 675, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 675, + 547, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 675, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_k" + }, + { + "bbox": [ + 304, + 675, + 547, + 715 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 304, + 675, + 547, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 675, + 547, + 715 + ], + "type": "text", + "content": "-th video clip with " + }, + { + "bbox": [ + 304, + 675, + 547, + 715 + ], + "type": "inline_equation", + "content": "N_{f}" + }, + { + "bbox": [ + 304, + 675, + 547, + 715 + ], + "type": "text", + "content": " frames degraded by arbitrary weather type, " + }, + { + "bbox": [ + 304, + 675, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{V}}_k" + }, + { + "bbox": [ + 304, + 675, + 547, + 715 + ], + "type": "text", + "content": " is the restored clip. " + }, + { + "bbox": [ + 304, + 675, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 675, + 547, + 715 + ], + "type": "text", + "content": " is the denoising function derived from DDIM [37]. The denoising" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25608" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 71, + 531, + 350 + ], + "blocks": [ + { + "bbox": [ + 65, + 71, + 531, + 350 + ], + "lines": [ + { + "bbox": [ + 65, + 71, + 531, + 350 + ], + "spans": [ + { + "bbox": [ + 65, + 71, + 531, + 350 + ], + "type": "image", + "image_path": "0e1aa3868e40cc558991bf5c6e8e6a82e4bcfe43ca3e06ca7c240e30f341f7cc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 360, + 545, + 416 + ], + "lines": [ + { + "bbox": [ + 46, + 360, + 545, + 416 + ], + "spans": [ + { + "bbox": [ + 46, + 360, + 545, + 416 + ], + "type": "text", + "content": "Figure 3. Overview of our Diff-TTA framework for video adverse weather removal. In the training stage, through the temporal diffusion process, we randomly add temporal noise on the clean video clip " + }, + { + "bbox": [ + 46, + 360, + 545, + 416 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{gt}" + }, + { + "bbox": [ + 46, + 360, + 545, + 416 + ], + "type": "text", + "content": " from the mixed training set. Then, the denoising NAFNet " + }, + { + "bbox": [ + 46, + 360, + 545, + 416 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 46, + 360, + 545, + 416 + ], + "type": "text", + "content": " is trained by estimating the applied noise, which is conditioned by the degraded counterpart " + }, + { + "bbox": [ + 46, + 360, + 545, + 416 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{lq}" + }, + { + "bbox": [ + 46, + 360, + 545, + 416 + ], + "type": "text", + "content": " by seen weather types. In the test stage, the proxy task, Diff-TSC, where the cropped tubelets of the last restored pair " + }, + { + "bbox": [ + 46, + 360, + 545, + 416 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{V}_{k-1}, \\hat{\\mathbf{V}}_{k-1}\\}" + }, + { + "bbox": [ + 46, + 360, + 545, + 416 + ], + "type": "text", + "content": " are utilized to learn the primer distribution, is incorporated into each timestep of denoising the randomly sampled temporal noise " + }, + { + "bbox": [ + 46, + 360, + 545, + 416 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_k^T" + }, + { + "bbox": [ + 46, + 360, + 545, + 416 + ], + "type": "text", + "content": " for the iterative online adaptation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 422, + 287, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 287, + 445 + ], + "type": "text", + "content": "network " + }, + { + "bbox": [ + 46, + 422, + 287, + 445 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 46, + 422, + 287, + 445 + ], + "type": "text", + "content": " in the denoising function consists of a single encoder and decoder." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 447, + 288, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 447, + 288, + 626 + ], + "spans": [ + { + "bbox": [ + 46, + 447, + 288, + 626 + ], + "type": "text", + "content": "In the training stage, we follow the formulation of DDIM and extend the diffusion forward process to a temporal counterpart for frame-correlation modeling. The denoising network " + }, + { + "bbox": [ + 46, + 447, + 288, + 626 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 46, + 447, + 288, + 626 + ], + "type": "text", + "content": " with NAFNet [6] as the backbone is trained for noise estimation, which enforces the output of the model to approximate the proposed temporal noise model. During the inference stage, we present a proxy task, Diff-TSC, to adapt the diffusion model to the target test data degraded by unknown weather based on the diffusion reverse process. We iteratively optimize the noise estimation of the cropped tubelets and calibrate the weight set of the model parameters for more robust weather removal. Next, we elaborate on our diffusion-based solution for adverse weather removal in videos and our diffusion test-time adaptation, which are also displayed in Figure 3." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 635, + 200, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 200, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 200, + 647 + ], + "type": "text", + "content": "4.1. Temporal Diffusion Process" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "The traditional diffusion model is trained to denoise independent noise from a perturbed image. The noise " + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": " in Eq. (1) is sampled from an i.i.d. Gaussian distribution " + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0, \\mathbf{I})" + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": ". Upon training the image diffusion model and utilizing it to reverse frames from a video into the noise" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 422, + 545, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 545, + 494 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 545, + 494 + ], + "type": "text", + "content": "space individually, the significant correlation among the noise maps of these frames can be easily observed. Rather than introducing massive parameters into the denoising network, we extend the original diffusion process to the temporal counterpart by an ARMA-based temporal noise model to efficiently preserve the correlation between frames." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 494, + 545, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 494, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 304, + 494, + 545, + 553 + ], + "type": "text", + "content": "Temporal noise model. ARMA (Auto-regressive Moving Average) [2] is a widely used statistical model for time series data, which combines the auto-regressive (AR) part and the moving average (MA) part. Typically, ARMA can be formulated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 341, + 560, + 545, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 560, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 341, + 560, + 545, + 594 + ], + "type": "interline_equation", + "content": "X _ {t} = c + \\varepsilon_ {t} + \\sum_ {i = 1} ^ {p} \\varphi_ {i} X _ {t - i} + \\sum_ {j = 1} ^ {q} \\tau_ {j} \\varepsilon_ {t - j}, \\tag {3}", + "image_path": "cc36675dcc80fc6cddde224bb7deeb9d48d6921c82ebc9a329275614f4479aac.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 601, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 601, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 601, + 546, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 601, + 546, + 715 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 601, + 546, + 715 + ], + "type": "text", + "content": " is the value term, " + }, + { + "bbox": [ + 304, + 601, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 304, + 601, + 546, + 715 + ], + "type": "text", + "content": " is the error term, " + }, + { + "bbox": [ + 304, + 601, + 546, + 715 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 601, + 546, + 715 + ], + "type": "text", + "content": " is a constant term, " + }, + { + "bbox": [ + 304, + 601, + 546, + 715 + ], + "type": "inline_equation", + "content": "p,q" + }, + { + "bbox": [ + 304, + 601, + 546, + 715 + ], + "type": "text", + "content": " are the auto-regressive order and moving average order while " + }, + { + "bbox": [ + 304, + 601, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\varphi ,\\tau" + }, + { + "bbox": [ + 304, + 601, + 546, + 715 + ], + "type": "text", + "content": " are their coefficients, respectively. Motivated by ARMA, we develop the original Gaussian noise model " + }, + { + "bbox": [ + 304, + 601, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,\\mathbf{I})" + }, + { + "bbox": [ + 304, + 601, + 546, + 715 + ], + "type": "text", + "content": " to the temporal noise model " + }, + { + "bbox": [ + 304, + 601, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(0,\\mathbf{I})" + }, + { + "bbox": [ + 304, + 601, + 546, + 715 + ], + "type": "text", + "content": ". Specifically, let " + }, + { + "bbox": [ + 304, + 601, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\{\\bar{\\epsilon}_i\\}_{i = 1}^{N_f}" + }, + { + "bbox": [ + 304, + 601, + 546, + 715 + ], + "type": "text", + "content": " denote the noise sequence corresponding to consecutive frames of a video clip. We first sample the noise sequence " + }, + { + "bbox": [ + 304, + 601, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\{\\bar{\\epsilon}_i\\}_{i = 1}^{N_f}" + }, + { + "bbox": [ + 304, + 601, + 546, + 715 + ], + "type": "text", + "content": " and the corresponding error sequence " + }, + { + "bbox": [ + 304, + 601, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\{\\varepsilon_i\\}_{i = 1}^{N_f}" + }, + { + "bbox": [ + 304, + 601, + 546, + 715 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 304, + 601, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,\\mathbf{I})" + }, + { + "bbox": [ + 304, + 601, + 546, + 715 + ], + "type": "text", + "content": ", respectively. The constant" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "25609" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "term " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": " should be consistent with the mean of Gaussian distribution, that is, zero. To take both the past and future information into consideration, we enforce the noise of the current frame affected not only by the noise and error term of the previous frame but also those of the next frame, and correspondingly the auto-regressive order and moving average order " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "p, q = 1" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": ". Consequently, the expression in Eq. (3) evolves into the convex combination of the noise term and error term:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 51, + 182, + 287, + 206 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 182, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 51, + 182, + 287, + 206 + ], + "type": "interline_equation", + "content": "\\bar {\\epsilon} _ {i} = (1 - \\varphi - \\tau) \\varepsilon_ {i} + \\varphi \\frac {\\bar {\\epsilon} _ {i - 1} + \\bar {\\epsilon} _ {i + 1}}{2} + \\tau \\frac {\\varepsilon_ {i - 1} + \\varepsilon_ {i + 1}}{2}, \\tag {4}", + "image_path": "3841b3a321eae102d973e384599e31ddbc41f3f5b13d32aeb03ed9d347b7f888.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 210, + 287, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 210, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 47, + 210, + 287, + 259 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 210, + 287, + 259 + ], + "type": "inline_equation", + "content": "\\varphi, \\tau" + }, + { + "bbox": [ + 47, + 210, + 287, + 259 + ], + "type": "text", + "content": " are empirically set as 0.6, 0.3, respectively. Algorithm 1 further illustrates our temporal noise model, where the noise sequence of each video clip is constructed with separate preorder and postorder steps." + } + ] + } + ], + "index": 2 + }, + { + "type": "code", + "bbox": [ + 48, + 281, + 266, + 467 + ], + "blocks": [ + { + "bbox": [ + 52, + 266, + 238, + 279 + ], + "lines": [ + { + "bbox": [ + 52, + 266, + 238, + 279 + ], + "spans": [ + { + "bbox": [ + 52, + 266, + 238, + 279 + ], + "type": "text", + "content": "Algorithm 1: Temporal Noise Model " + }, + { + "bbox": [ + 52, + 266, + 238, + 279 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(0,\\mathbf{I})" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 48, + 281, + 266, + 467 + ], + "lines": [ + { + "bbox": [ + 48, + 281, + 266, + 467 + ], + "spans": [ + { + "bbox": [ + 48, + 281, + 266, + 467 + ], + "type": "text", + "content": "Input: " + }, + { + "bbox": [ + 48, + 281, + 266, + 467 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 48, + 281, + 266, + 467 + ], + "type": "text", + "content": " clips: " + }, + { + "bbox": [ + 48, + 281, + 266, + 467 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{V}_1,\\mathbf{V}_2,\\dots ,\\mathbf{V}_K\\}" + }, + { + "bbox": [ + 48, + 281, + 266, + 467 + ], + "type": "text", + "content": " in a batch; The coefficients " + }, + { + "bbox": [ + 48, + 281, + 266, + 467 + ], + "type": "inline_equation", + "content": "\\varphi ,\\tau" + }, + { + "bbox": [ + 48, + 281, + 266, + 467 + ], + "type": "text", + "content": " \nfor " + }, + { + "bbox": [ + 48, + 281, + 266, + 467 + ], + "type": "inline_equation", + "content": "k = 1,2,\\ldots ,K" + }, + { + "bbox": [ + 48, + 281, + 266, + 467 + ], + "type": "text", + "content": " parallelly do \nfor Frame " + }, + { + "bbox": [ + 48, + 281, + 266, + 467 + ], + "type": "inline_equation", + "content": "i = 1,2,\\dots ,N_{f}" + }, + { + "bbox": [ + 48, + 281, + 266, + 467 + ], + "type": "text", + "content": " parallelly do \nsample " + }, + { + "bbox": [ + 48, + 281, + 266, + 467 + ], + "type": "inline_equation", + "content": "\\bar{\\epsilon}_i,\\varepsilon_i\\in \\mathbb{R}^{1\\times c\\times h\\times w}" + }, + { + "bbox": [ + 48, + 281, + 266, + 467 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 48, + 281, + 266, + 467 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,\\mathbf{I})" + }, + { + "bbox": [ + 48, + 281, + 266, + 467 + ], + "type": "text", + "content": " \nend \nfor Frame " + }, + { + "bbox": [ + 48, + 281, + 266, + 467 + ], + "type": "inline_equation", + "content": "i = 2,3,\\dots ,N_{f}" + }, + { + "bbox": [ + 48, + 281, + 266, + 467 + ], + "type": "text", + "content": " sequentially do \n" + }, + { + "bbox": [ + 48, + 281, + 266, + 467 + ], + "type": "inline_equation", + "content": "\\bar{\\epsilon}_i = (1 - \\varphi -\\tau)\\varepsilon_i + \\varphi \\bar{\\epsilon}_{i - 1} + \\tau \\varepsilon_{i - 1}" + }, + { + "bbox": [ + 48, + 281, + 266, + 467 + ], + "type": "text", + "content": " \nend \nfor Frame " + }, + { + "bbox": [ + 48, + 281, + 266, + 467 + ], + "type": "inline_equation", + "content": "i = N_{f} - 1,\\dots ,2,1" + }, + { + "bbox": [ + 48, + 281, + 266, + 467 + ], + "type": "text", + "content": " sequentially do \n" + }, + { + "bbox": [ + 48, + 281, + 266, + 467 + ], + "type": "inline_equation", + "content": "\\bar{\\epsilon}_i = (1 - \\varphi -\\tau)\\varepsilon_i + \\varphi \\bar{\\epsilon}_{i + 1} + \\tau \\varepsilon_{i + 1}" + }, + { + "bbox": [ + 48, + 281, + 266, + 467 + ], + "type": "text", + "content": " \nend \n" + }, + { + "bbox": [ + 48, + 281, + 266, + 467 + ], + "type": "inline_equation", + "content": "\\bar{\\epsilon} =" + }, + { + "bbox": [ + 48, + 281, + 266, + 467 + ], + "type": "text", + "content": " normalize( " + }, + { + "bbox": [ + 48, + 281, + 266, + 467 + ], + "type": "inline_equation", + "content": "\\bar{\\epsilon})\\Rightarrow \\bar{\\epsilon}\\sim \\mathcal{T}(0,\\mathbf{I})" + }, + { + "bbox": [ + 48, + 281, + 266, + 467 + ], + "type": "text", + "content": " \nend \nReturn " + }, + { + "bbox": [ + 48, + 281, + 266, + 467 + ], + "type": "inline_equation", + "content": "\\{\\{\\bar{\\epsilon}_i|\\bar{\\epsilon}_i\\in \\mathbb{R}^{1\\times c\\times h\\times w}\\}_{i = 1}^{N_f}\\}_{k = 1}^K" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "code_body" + } + ], + "index": 4, + "sub_type": "algorithm" + }, + { + "bbox": [ + 46, + 483, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 287, + 555 + ], + "type": "text", + "content": "To enhance the capacity of our denoising network for capturing temporal dependencies, we have specifically modified the architecture by substituting only the first and last layers of the denoising NAFNet with the 3D convolutional layer. This tailored network, denoted as " + }, + { + "bbox": [ + 46, + 483, + 287, + 555 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 46, + 483, + 287, + 555 + ], + "type": "text", + "content": " is then optimized as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 77, + 560, + 287, + 575 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 560, + 287, + 575 + ], + "spans": [ + { + "bbox": [ + 77, + 560, + 287, + 575 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {S} = \\left\\| \\bar {\\epsilon} - \\epsilon_ {\\theta} \\left(\\sqrt {\\bar {\\alpha} _ {t}} \\hat {\\mathbf {V}} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\bar {\\epsilon}, \\mathbf {V}, t\\right) \\right\\| _ {1}, \\tag {5}", + "image_path": "61f2f7ea515564860e5858bfcbebec57972cc1cfe6d2790c5b66e6375e68e701.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 581, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 581, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 581, + 287, + 605 + ], + "type": "text", + "content": "where we utilize the temporal noise model as the supervision, and " + }, + { + "bbox": [ + 46, + 581, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|" + }, + { + "bbox": [ + 46, + 581, + 287, + 605 + ], + "type": "text", + "content": " denotes " + }, + { + "bbox": [ + 46, + 581, + 287, + 605 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 46, + 581, + 287, + 605 + ], + "type": "text", + "content": " norm." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 611, + 216, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 216, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 216, + 624 + ], + "type": "text", + "content": "4.2. Diffusion Test-time Adaptation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "content": "Once trained by the proposed temporal noise model, the network " + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "content": " can generate high-quality images. This is achieved by sampling a noisy state " + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_T" + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "content": " and then iteratively denoising it in the reverse process. Yang et al. [51] introduced adversarial learning for weather conditions, which require weather type annotations during the training stage. Additionally, their method may fail to address out-of-distribution" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "content": "data, because including all real-world conditions in training, considering both weather types and image statistics, is impractical. Contrastively, we introduce test-time adaptation and update knowledge by practice to remove arbitrary weather conditions robustly. We design and incorporate Diffusion Tubelet Self-Calibration (Diff-TSC), a proxy task, into the iterative denoising of the diffusion reverse process to learn the primer distribution of the frames degraded by unknown weather." + } + ] + } + ], + "index": 10 + }, + { + "type": "code", + "bbox": [ + 306, + 237, + 521, + 495 + ], + "blocks": [ + { + "bbox": [ + 310, + 187, + 521, + 236 + ], + "lines": [ + { + "bbox": [ + 310, + 187, + 521, + 236 + ], + "spans": [ + { + "bbox": [ + 310, + 187, + 521, + 236 + ], + "type": "text", + "content": "Algorithm 2: Diffusion Test-Time Adaptation to unknown weather. " + }, + { + "bbox": [ + 310, + 187, + 521, + 236 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 310, + 187, + 521, + 236 + ], + "type": "text", + "content": " is the weight set of the trained network before adaptation, " + }, + { + "bbox": [ + 310, + 187, + 521, + 236 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 310, + 187, + 521, + 236 + ], + "type": "text", + "content": " is the learning rate for online adaptation." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 306, + 237, + 521, + 495 + ], + "lines": [ + { + "bbox": [ + 306, + 237, + 521, + 495 + ], + "spans": [ + { + "bbox": [ + 306, + 237, + 521, + 495 + ], + "type": "text", + "content": "Input: " + }, + { + "bbox": [ + 306, + 237, + 521, + 495 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 306, + 237, + 521, + 495 + ], + "type": "text", + "content": " overlapped clips: " + }, + { + "bbox": [ + 306, + 237, + 521, + 495 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{V}_1,\\mathbf{V}_2,\\dots ,\\mathbf{V}_K\\}" + }, + { + "bbox": [ + 306, + 237, + 521, + 495 + ], + "type": "text", + "content": " in one video stream \nfor " + }, + { + "bbox": [ + 306, + 237, + 521, + 495 + ], + "type": "inline_equation", + "content": "Clp k = 1,2,\\ldots ,K" + }, + { + "bbox": [ + 306, + 237, + 521, + 495 + ], + "type": "text", + "content": " sequentially do \nInitialize the network " + }, + { + "bbox": [ + 306, + 237, + 521, + 495 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\bar{\\theta}}" + }, + { + "bbox": [ + 306, + 237, + 521, + 495 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 306, + 237, + 521, + 495 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 306, + 237, + 521, + 495 + ], + "type": "text", + "content": " \nInitialize " + }, + { + "bbox": [ + 306, + 237, + 521, + 495 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{V}}_k" + }, + { + "bbox": [ + 306, + 237, + 521, + 495 + ], + "type": "text", + "content": " by Algorithm 1 \nif " + }, + { + "bbox": [ + 306, + 237, + 521, + 495 + ], + "type": "inline_equation", + "content": "k = 1" + }, + { + "bbox": [ + 306, + 237, + 521, + 495 + ], + "type": "text", + "content": " then \nfor step " + }, + { + "bbox": [ + 306, + 237, + 521, + 495 + ], + "type": "inline_equation", + "content": "t = T,\\dots ,2,1" + }, + { + "bbox": [ + 306, + 237, + 521, + 495 + ], + "type": "text", + "content": " sequentially do \n" + }, + { + "bbox": [ + 306, + 237, + 521, + 495 + ], + "type": "inline_equation", + "content": "|\\hat{\\mathbf{V}}_k = ddim(\\mathbf{V}_k,\\hat{\\mathbf{V}}_k,\\epsilon_{\\bar{\\theta}},t)" + }, + { + "bbox": [ + 306, + 237, + 521, + 495 + ], + "type": "text", + "content": " \nend \nelse \nA, " + }, + { + "bbox": [ + 306, + 237, + 521, + 495 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{A}} =" + }, + { + "bbox": [ + 306, + 237, + 521, + 495 + ], + "type": "text", + "content": " TubeletCrop(Vk-1,Vk-1,Na) for step " + }, + { + "bbox": [ + 306, + 237, + 521, + 495 + ], + "type": "inline_equation", + "content": "t = T,\\dots ,2,1" + }, + { + "bbox": [ + 306, + 237, + 521, + 495 + ], + "type": "text", + "content": " sequentially do \nCompute " + }, + { + "bbox": [ + 306, + 237, + 521, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_S(\\mathbf{A},\\hat{\\mathbf{A}},\\bar{\\epsilon},t)" + }, + { + "bbox": [ + 306, + 237, + 521, + 495 + ], + "type": "text", + "content": " by Eq. (5) " + }, + { + "bbox": [ + 306, + 237, + 521, + 495 + ], + "type": "inline_equation", + "content": "\\bar{\\theta} = \\bar{\\theta} -\\delta \\nabla_{\\bar{\\theta}}\\mathcal{L}_{S}" + }, + { + "bbox": [ + 306, + 237, + 521, + 495 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{V}}_k = ddim(\\mathbf{V}_k,\\hat{\\mathbf{V}}_k,\\epsilon_{\\bar{\\theta}},t)" + }, + { + "bbox": [ + 306, + 237, + 521, + 495 + ], + "type": "text", + "content": " \nend \n" + }, + { + "bbox": [ + 306, + 237, + 521, + 495 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{V}}_k =" + }, + { + "bbox": [ + 306, + 237, + 521, + 495 + ], + "type": "text", + "content": " integrate( " + }, + { + "bbox": [ + 306, + 237, + 521, + 495 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{V}}_k,\\hat{\\mathbf{V}}_{k - 1})" + }, + { + "bbox": [ + 306, + 237, + 521, + 495 + ], + "type": "text", + "content": " \nend \nReturn the restored clips " + }, + { + "bbox": [ + 306, + 237, + 521, + 495 + ], + "type": "inline_equation", + "content": "\\{\\hat{\\mathbf{V}}_1,\\hat{\\mathbf{V}}_2,\\dots ,\\hat{\\mathbf{V}}_K\\}" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "code_body" + } + ], + "index": 12, + "sub_type": "algorithm" + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "text", + "content": "Diffusion Tubelet Self-Calibration. We design a proxy task for self-supervised learning to transfer the diffusion model to the specific weather condition. Specifically, for each clip from the video stream, we first initialize the weight set of " + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\bar{\\theta}}" + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "text", + "content": " trained by source data. We leverage the previous clip " + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{k - 1}" + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "text", + "content": " and its restored result " + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{V}}_{k - 1}" + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "text", + "content": " as the pair of condition and input for the online optimization when inferring the current clip " + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_k" + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "text", + "content": ". Rather than directly traversing all fine-grained details, we randomly crop " + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "inline_equation", + "content": "N_{a}" + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "text", + "content": " tubelets from " + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{k - 1}" + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "text", + "content": " as the primer distribution to avoid memory overhead and massive computational costs caused by high resolution. Following the process in Section 4.1, we conduct our temporal diffusion process on the restored tubelets " + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{A}}" + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "text", + "content": " and feed them together with the degraded tubelets " + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "text", + "content": " into the denoising NAFNet for noise estimation. As suggested in [40], it is necessary to iteratively find a relatively optimal point for the target distribution during test-time adaptation. Inher" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25610" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 118, + 518, + 337 + ], + "blocks": [ + { + "bbox": [ + 46, + 71, + 547, + 114 + ], + "lines": [ + { + "bbox": [ + 46, + 71, + 547, + 114 + ], + "spans": [ + { + "bbox": [ + 46, + 71, + 547, + 114 + ], + "type": "text", + "content": "Table 1. Quantitative evaluation on seen weather conditions for video adverse weather removal. For Original Weather, these methods are trained on the weather-specific training set and tested on the weather-specific testing set. For Rain, Haze, and Snow, these methods are trained on a mixed training set and tested on the weather-specific testing set. The average performance is calculated on Rain, Haze, and Snow. PSNR and SSIM are adopted as our evaluation metrics. The best and second-best values are bold and underlined." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 118, + 518, + 337 + ], + "lines": [ + { + "bbox": [ + 70, + 118, + 518, + 337 + ], + "spans": [ + { + "bbox": [ + 70, + 118, + 518, + 337 + ], + "type": "table", + "html": "
MethodsTypeSourceDatasets
Original WeatherRainHazeSnowAverage
DerainPReNet [34]ImageCVPR'1927.060.907726.800.881417.640.803028.570.940124.340.8748
SLDNet [50]VideoCVPR'2020.310.627221.240.712916.210.756122.010.855019.820.7747
S2VD [53]VideoCVPR'2124.090.794428.390.900619.650.860726.230.919024.760.8934
RDD-Net [42]VideoECCV'2231.820.942330.340.930018.360.843230.400.956026.370.9097
DehazeGDN [26]ImageICCV'1919.690.854529.960.937019.010.880531.020.951826.660.9231
MSBDN [15]ImageCVPR'2022.010.875926.700.914622.240.904727.070.934025.340.9178
VDHNet [35]VideoTIP'1916.640.813329.870.927216.850.821429.530.939525.420.8960
PM-Net [28]VideoMM'2223.830.895025.790.888023.570.914318.710.788122.690.8635
DesnowDesnowNet [29]ImageTIP'1828.300.953025.190.878616.430.790227.560.918123.060.8623
DDMSNET [56]ImageTIP'2132.550.961329.010.918819.500.861532.430.969426.980.9166
HDCW-W-Net [10]ImageICCV'2131.770.954228.100.905517.360.792131.050.948225.500.8819
SMGARN [13]ImageTCSVT'2233.240.972127.780.910017.850.807532.340.966825.990.8948
RestorationMPRNet [54]ImageCVPR'21————28.220.916520.250.893430.950.948226.470.9194
EDVR [43]VideoCVPR'19————31.100.937119.670.872430.270.944027.010.9178
RVRT [23]VideoNIPS'22————30.110.913221.160.894926.780.883426.020.8972
RTA [61]VideoCVPR'22————30.120.918620.750.891529.790.936726.890.9156
Multi-WeatherAll-in-one [21]ImageCVPR'20————26.620.894820.880.901030.090.943125.860.9130
UVRNet [19]ImageTMM'22————22.310.767820.820.857524.710.887322.610.8375
TransWeather [38]ImageCVPR'22————26.820.911822.170.902528.870.931325.950.9152
TKL [11]ImageCVPR'22————26.730.893522.080.904431.350.951526.720.9165
WeatherDiffusion [33]ImageTPAMI'23————25.860.912520.100.844226.400.911324.120.8893
WGWS-Net [62]ImageCVPR'23————29,640.931017.710.811331.580.952826.310.9265
ViWS-Net [51]VideoICCV'23————31.520.943324.510.918731.490.956229.170.9394
OursVideo——————32.430.957324.560.914831.860.964029.630.9453
", + "image_path": "912e1ff711fe7cf1878fe7782acf69fc168e6b017581369e25df42d223799ce8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 350, + 289, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 350, + 289, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 289, + 555 + ], + "type": "text", + "content": "ited from the characteristics of DDIM, such requirement can be easily achieved in the iterative diffusion reverse process, which needs more increments of inference time in the traditional CNNs and Transformers. For each timestep " + }, + { + "bbox": [ + 46, + 350, + 289, + 555 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 350, + 289, + 555 + ], + "type": "text", + "content": ", we first obtain the noisy tubelet " + }, + { + "bbox": [ + 46, + 350, + 289, + 555 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{A}}_t" + }, + { + "bbox": [ + 46, + 350, + 289, + 555 + ], + "type": "text", + "content": " and then optimize the objective function given in Eq. (5) to update the weight set " + }, + { + "bbox": [ + 46, + 350, + 289, + 555 + ], + "type": "inline_equation", + "content": "\\bar{\\theta}" + }, + { + "bbox": [ + 46, + 350, + 289, + 555 + ], + "type": "text", + "content": ". After the iterative adaptation to " + }, + { + "bbox": [ + 46, + 350, + 289, + 555 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{V}}_k" + }, + { + "bbox": [ + 46, + 350, + 289, + 555 + ], + "type": "text", + "content": ", we ensemble the overlapped part of it with that of the previous clip " + }, + { + "bbox": [ + 46, + 350, + 289, + 555 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{V}}_{k - 1}" + }, + { + "bbox": [ + 46, + 350, + 289, + 555 + ], + "type": "text", + "content": " to capture complementary information of pre-adaptation and post-adaptation. We optimize over the entire network except for the last layer suggested by SHOT [22]. The detailed procedure is described in Algorithm 2. We emphasize the synergy between diffusion models and test-time adaptation in accumulating frame-correlated knowledge for adverse weather removal. By attaching the proxy task to iterative denoising, our approach also avoids excessive computational overhead caused by iterative adaptation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 565, + 128, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 565, + 128, + 578 + ], + "spans": [ + { + "bbox": [ + 47, + 565, + 128, + 578 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 586, + 110, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 110, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 110, + 597 + ], + "type": "text", + "content": "5.1. Datasets" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": "Various video adverse weather datasets are used in our experiments. Following the setting of [51], we adopt RainMotion [42], REVIDE [59] and KITTI-snow [51] as seen weather. For the training stage, we merge the training set of the three datasets as the mixed set to learn a generic model. For the testing stage, we evaluate our model on three testing sets, respectively. Additionally, we evaluate our approach on two datasets VRDS [44] and RVSD [4] to validate the generalization of our Diff-TTA on unseen" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 319, + 385, + 533, + 477 + ], + "blocks": [ + { + "bbox": [ + 305, + 348, + 545, + 382 + ], + "lines": [ + { + "bbox": [ + 305, + 348, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 305, + 348, + 545, + 382 + ], + "type": "text", + "content": "Table 2. Quantitative evaluation on unseen weather conditions for video adverse weather removal. The best and second-best performances are marked in bold and underlined, respectively." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 385, + 533, + 477 + ], + "lines": [ + { + "bbox": [ + 319, + 385, + 533, + 477 + ], + "spans": [ + { + "bbox": [ + 319, + 385, + 533, + 477 + ], + "type": "table", + "html": "
MethodVRDS [44]RVSD [4]
PSNR ↑SSIM ↑PSNR ↑SSIM ↑
All-in-one [21]20.440.594419.790.7509
TransWeather [38]21.360.713620.250.7514
TKL [11]20.490.700319.710.7370
WeatherDiffusion [33]20.730.694318.080.6588
ViWS-Net [51]21.570.709419.830.7590
Diff-TTA (ours)22.570.728122.350.7719
", + "image_path": "d15cf52da9f5a6d01254807d850333723605f282c784470381d19dee6dd05fa4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 484, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 545, + 593 + ], + "type": "text", + "content": "weather conditions. VRDS is a synthesized video dataset of joint rain streaks and raindrops with a total of 102 videos, while RVSD is a realistic video desnowing dataset with a total of 110 videos containing both snow and fog achieved by the rendering engine. Due to the large quantity of the two datasets, we randomly sample 8 videos as the test set, each of which has 20 frames, respectively. We also collect several real-world videos degraded by diverse weather conditions to prove its effectiveness on real-world applications." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 599, + 404, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 599, + 404, + 612 + ], + "spans": [ + { + "bbox": [ + 306, + 599, + 404, + 612 + ], + "type": "text", + "content": "5.2. Implementation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Training Details. The proposed framework was trained on NVIDIA RTX 4090 GPUs and implemented on the Pytorch platform. We adopted the Lion optimizer [12] (" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.99" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": ") with the initial learning rate of " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-5}" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " decayed to " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-7}" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " by the Cosine scheduler. We randomly crop the video frames to " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " for training. Our framework is empirically trained for " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "600k" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " iterations in an end-to-end way with batch size of 4. Each batch consists of 4 video" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "25611" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 72, + 112, + 198 + ], + "blocks": [ + { + "bbox": [ + 52, + 72, + 112, + 198 + ], + "lines": [ + { + "bbox": [ + 52, + 72, + 112, + 198 + ], + "spans": [ + { + "bbox": [ + 52, + 72, + 112, + 198 + ], + "type": "image", + "image_path": "c4ef1525402fa617c8c4fc37c9fc532693fa65f894364c12d715d6feb5a4cd38.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 199, + 105, + 208 + ], + "lines": [ + { + "bbox": [ + 61, + 199, + 105, + 208 + ], + "spans": [ + { + "bbox": [ + 61, + 199, + 105, + 208 + ], + "type": "text", + "content": "Input Frame" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 113, + 72, + 174, + 198 + ], + "blocks": [ + { + "bbox": [ + 113, + 72, + 174, + 198 + ], + "lines": [ + { + "bbox": [ + 113, + 72, + 174, + 198 + ], + "spans": [ + { + "bbox": [ + 113, + 72, + 174, + 198 + ], + "type": "image", + "image_path": "367d8b11e4217833f1df827769e1dede565c28cd3cdb7e02e0c3262a33cbafe0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 133, + 199, + 154, + 207 + ], + "lines": [ + { + "bbox": [ + 133, + 199, + 154, + 207 + ], + "spans": [ + { + "bbox": [ + 133, + 199, + 154, + 207 + ], + "type": "text", + "content": "EDVR" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 175, + 72, + 235, + 198 + ], + "blocks": [ + { + "bbox": [ + 175, + 72, + 235, + 198 + ], + "lines": [ + { + "bbox": [ + 175, + 72, + 235, + 198 + ], + "spans": [ + { + "bbox": [ + 175, + 72, + 235, + 198 + ], + "type": "image", + "image_path": "e4e0670ed6db40c128cae71d709d86e2d8697b607197ba7d2c10c6176427121b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 200, + 231, + 208 + ], + "lines": [ + { + "bbox": [ + 181, + 200, + 231, + 208 + ], + "spans": [ + { + "bbox": [ + 181, + 200, + 231, + 208 + ], + "type": "text", + "content": "TransWeather" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 236, + 72, + 296, + 198 + ], + "blocks": [ + { + "bbox": [ + 236, + 72, + 296, + 198 + ], + "lines": [ + { + "bbox": [ + 236, + 72, + 296, + 198 + ], + "spans": [ + { + "bbox": [ + 236, + 72, + 296, + 198 + ], + "type": "image", + "image_path": "06ec1ec474e98cf6df4aedb1e18c5c14342d37093ee7787396fe35fafa9ec250.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 260, + 200, + 273, + 207 + ], + "lines": [ + { + "bbox": [ + 260, + 200, + 273, + 207 + ], + "spans": [ + { + "bbox": [ + 260, + 200, + 273, + 207 + ], + "type": "text", + "content": "TKL" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 217, + 546, + 250 + ], + "lines": [ + { + "bbox": [ + 46, + 217, + 546, + 250 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 546, + 250 + ], + "type": "text", + "content": "Figure 4. Qualitative Comparison of seen weather conditions between our approach and state-of-the-art algorithms. The competitive algorithms are selected to present the results on the example frames degraded by rain, haze, snow, respectively. The color box indicates the comparison of details. Please zoom in on the images for improved visualization." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 298, + 72, + 358, + 198 + ], + "blocks": [ + { + "bbox": [ + 298, + 72, + 358, + 198 + ], + "lines": [ + { + "bbox": [ + 298, + 72, + 358, + 198 + ], + "spans": [ + { + "bbox": [ + 298, + 72, + 358, + 198 + ], + "type": "image", + "image_path": "4c444ee7d709fc7617d4dbba15a6b83fb2c3c066d0d36d4b7b7ac59759ae87b0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 298, + 200, + 361, + 207 + ], + "lines": [ + { + "bbox": [ + 298, + 200, + 361, + 207 + ], + "spans": [ + { + "bbox": [ + 298, + 200, + 361, + 207 + ], + "type": "text", + "content": "WeatherDiffusion" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 359, + 72, + 419, + 198 + ], + "blocks": [ + { + "bbox": [ + 359, + 72, + 419, + 198 + ], + "lines": [ + { + "bbox": [ + 359, + 72, + 419, + 198 + ], + "spans": [ + { + "bbox": [ + 359, + 72, + 419, + 198 + ], + "type": "image", + "image_path": "c51639beb9c788c6e37854cb62cc3cc81c46c2cd44bf59f7507afd04bc9f7c1c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 372, + 200, + 407, + 207 + ], + "lines": [ + { + "bbox": [ + 372, + 200, + 407, + 207 + ], + "spans": [ + { + "bbox": [ + 372, + 200, + 407, + 207 + ], + "type": "text", + "content": "ViWS-Net" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 421, + 72, + 481, + 198 + ], + "blocks": [ + { + "bbox": [ + 421, + 72, + 481, + 198 + ], + "lines": [ + { + "bbox": [ + 421, + 72, + 481, + 198 + ], + "spans": [ + { + "bbox": [ + 421, + 72, + 481, + 198 + ], + "type": "image", + "image_path": "d71a8e9bcaff4b45039c96f8d47d118073fcc90da1686e6952c81ae61a106715.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 443, + 201, + 460, + 207 + ], + "lines": [ + { + "bbox": [ + 443, + 201, + 460, + 207 + ], + "spans": [ + { + "bbox": [ + 443, + 201, + 460, + 207 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 482, + 72, + 542, + 198 + ], + "blocks": [ + { + "bbox": [ + 482, + 72, + 542, + 198 + ], + "lines": [ + { + "bbox": [ + 482, + 72, + 542, + 198 + ], + "spans": [ + { + "bbox": [ + 482, + 72, + 542, + 198 + ], + "type": "image", + "image_path": "82476afc385bf30fd7b32d0c730b1503ab2b5f09f58730dc8f49414a48683f21.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 507, + 201, + 517, + 207 + ], + "lines": [ + { + "bbox": [ + 507, + 201, + 517, + 207 + ], + "spans": [ + { + "bbox": [ + 507, + 201, + 517, + 207 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 257, + 161, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 257, + 161, + 268 + ], + "spans": [ + { + "bbox": [ + 47, + 257, + 161, + 268 + ], + "type": "text", + "content": "clips with 5 frames per clip." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 269, + 287, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 269, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 269, + 287, + 411 + ], + "type": "text", + "content": "Method-specific Details. During the training, we used the diffusion process linearly increasing from " + }, + { + "bbox": [ + 46, + 269, + 287, + 411 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.0001" + }, + { + "bbox": [ + 46, + 269, + 287, + 411 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 269, + 287, + 411 + ], + "type": "inline_equation", + "content": "\\beta_{T} = 0.02" + }, + { + "bbox": [ + 46, + 269, + 287, + 411 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 269, + 287, + 411 + ], + "type": "inline_equation", + "content": "T = 1000" + }, + { + "bbox": [ + 46, + 269, + 287, + 411 + ], + "type": "text", + "content": " timesteps. During the testing, the inference step is reduced to 25 by deterministic implicit sampling. We utilize the same optimizer with a smaller learning rate of " + }, + { + "bbox": [ + 46, + 269, + 287, + 411 + ], + "type": "inline_equation", + "content": "1\\times 10^{-6}" + }, + { + "bbox": [ + 46, + 269, + 287, + 411 + ], + "type": "text", + "content": " for the test-time adaptation. The number of cropped tubelets " + }, + { + "bbox": [ + 46, + 269, + 287, + 411 + ], + "type": "inline_equation", + "content": "N_{a}" + }, + { + "bbox": [ + 46, + 269, + 287, + 411 + ], + "type": "text", + "content": " is empirically set to 5 while the crop size is " + }, + { + "bbox": [ + 46, + 269, + 287, + 411 + ], + "type": "inline_equation", + "content": "128\\times 128" + }, + { + "bbox": [ + 46, + 269, + 287, + 411 + ], + "type": "text", + "content": ". Despite the incorporated adaptation, our Diff-TTA is particularly " + }, + { + "bbox": [ + 46, + 269, + 287, + 411 + ], + "type": "inline_equation", + "content": "90\\times" + }, + { + "bbox": [ + 46, + 269, + 287, + 411 + ], + "type": "text", + "content": " more efficient than WeatherDiffusion. To process a video clip of 5 frames with the patch size of " + }, + { + "bbox": [ + 46, + 269, + 287, + 411 + ], + "type": "inline_equation", + "content": "256\\times 256" + }, + { + "bbox": [ + 46, + 269, + 287, + 411 + ], + "type": "text", + "content": ", the average run time of Diff-TTA is 6.01s while that of WeatherDiffusion is 542.76s." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 47, + 420, + 185, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 420, + 185, + 431 + ], + "spans": [ + { + "bbox": [ + 47, + 420, + 185, + 431 + ], + "type": "text", + "content": "5.3. Performance Evaluation" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 437, + 287, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 545 + ], + "type": "text", + "content": "Comparison methods. As shown in Table 1, we compared our proposed method against five kinds of state-of-the-art methods on the three seen weather conditions as [51] does, i.e., derain, dehaze, desnow, restoration, all-in-one adverse weather removal. For all-in-one adverse weather removal, we compared ours with the six representative single-image methods All-in-one [21], UVRNet [19], TransWeather [38], TKL [11], WeatherDiffusion [33], WGWS-Net [62], and the video-level method ViWS-Net [51]." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "type": "text", + "content": "Quantitative Comparison. For quantitative evaluation of the restored results, we apply the peak signal-to-noise ratio (PSNR) and the structural similarity (SSIM) as the metrics. Following the previous experimental setting, only the results of the model trained on the mixed training set are reported for all-in-one adverse weather removal models. To ensure the fairness of comparisons, we retrain each compared model implemented by the official codes based on our training dataset and report the best result. Observed from Table 1, our method achieves the best average performance across rain, haze and snow conditions among all kinds of methods by a considerable margin of 0.46, 0.0059 in PSNR, SSIM, respectively, than the second-best method ViWS-Net [51]. Though some single-weather" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 257, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 257, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 257, + 545, + 422 + ], + "type": "text", + "content": "methods like DDMSNET [56] and SMGARN [13] achieve impressive results in their own weather condition, they frequently fail to remove other weather conditions, such as haze. Our method significantly surpasses other all-in-one adverse weather methods in rain and snow conditions, and obtains competitive results in haze conditions. It is worth noting that the generalizable ability is critical for such methods, which can be observed from the average performance. To further validate their generalization, we evaluate all-in-on adverse weather removal methods on unseen weather conditions, i.e., rain streak+raindrop and snow+fog, in Table 2. Our Diff-TTA exhibits remarkable superiority over state-of-the-art methods when confronted with these unseen weather conditions." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 425, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 425, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 425, + 545, + 581 + ], + "type": "text", + "content": "Visual comparisons on synthetic videos. To vividly demonstrate the effectiveness of our approach, Figure 4 shows the qualitative comparison under the rain, haze, and snow conditions between our method and five state-of-the-art methods. Our method demonstrates promising results in terms of visual quality across various weather types. Specifically, when it comes to rain and snow scenarios, our approach significantly reduces the presence of rain streaks and snow particles, outperforming other methods. In hazy scenarios, our method effectively removes residual haze and preserves the clean background with impressive results. More visual results of seen and unseen weather are displayed in the supplementary files." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 581, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 581, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 581, + 545, + 676 + ], + "type": "text", + "content": "Visual comparisons on real-world videos. To evaluate the universality of our Diff-TTA on real-world applications, we collect several real-world degraded videos by rain, haze, snow from Youtube website, and further compare our approach against other all-in-one adverse weather removal methods. Figure 5 displays the visual results produced by our network and the compared methods on real-world video frames, clearly showcasing our exceptional performance." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Ablation Study. We evaluate the effectiveness of each component including diffusion process, temporal noise model, and Diffusion test-time adaptation (Diff-TTA) as described" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25612" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 96, + 517, + 165 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 547, + 94 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 547, + 94 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 547, + 94 + ], + "type": "text", + "content": "Table 3. Ablation study of each critical component in the proposed framework on seen weather conditions. The top values are marked in bold font. \"Diff-TTA\" denote Diffusion Test-time Adaptation." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 96, + 517, + 165 + ], + "lines": [ + { + "bbox": [ + 70, + 96, + 517, + 165 + ], + "spans": [ + { + "bbox": [ + 70, + 96, + 517, + 165 + ], + "type": "table", + "html": "
CombinationComponentDatasets
Diffusion ProcessTemporal NoiseDiff-TTARainHazeSnowAverage
M1---29.070.951422.640.893028.790.9350
M2✓--31.550.946623.580.904130.480.9520
M3✓✓-32.100.953024.310.912431.040.9621
Ours✓✓✓32.430.957324.560.914831.860.9640
", + "image_path": "26842280312f7349bfc056a3f08f4e53e497f162b172e0f200b8c8ec39cf9631.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 62, + 171, + 140, + 304 + ], + "blocks": [ + { + "bbox": [ + 62, + 171, + 140, + 304 + ], + "lines": [ + { + "bbox": [ + 62, + 171, + 140, + 304 + ], + "spans": [ + { + "bbox": [ + 62, + 171, + 140, + 304 + ], + "type": "image", + "image_path": "1e0c27c42d2afe3f072b88bae1b21288b97b4be2c831ad7ed84afeac7a982c47.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 305, + 123, + 313 + ], + "lines": [ + { + "bbox": [ + 78, + 305, + 123, + 313 + ], + "spans": [ + { + "bbox": [ + 78, + 305, + 123, + 313 + ], + "type": "text", + "content": "Input Frame" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 140, + 171, + 218, + 304 + ], + "blocks": [ + { + "bbox": [ + 140, + 171, + 218, + 304 + ], + "lines": [ + { + "bbox": [ + 140, + 171, + 218, + 304 + ], + "spans": [ + { + "bbox": [ + 140, + 171, + 218, + 304 + ], + "type": "image", + "image_path": "62b1cab1b0b1b1f9ca1d318d6547b8bb4df3ede760bf741977af5eff23bac9a6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 154, + 305, + 203, + 312 + ], + "lines": [ + { + "bbox": [ + 154, + 305, + 203, + 312 + ], + "spans": [ + { + "bbox": [ + 154, + 305, + 203, + 312 + ], + "type": "text", + "content": "TransWeather" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 219, + 171, + 296, + 304 + ], + "blocks": [ + { + "bbox": [ + 219, + 171, + 296, + 304 + ], + "lines": [ + { + "bbox": [ + 219, + 171, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 219, + 171, + 296, + 304 + ], + "type": "image", + "image_path": "7951abc98c6172b2b2c0d5eb584958fe0fe1ca24a689d1836c1e229f6f1beb0d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 249, + 305, + 263, + 312 + ], + "lines": [ + { + "bbox": [ + 249, + 305, + 263, + 312 + ], + "spans": [ + { + "bbox": [ + 249, + 305, + 263, + 312 + ], + "type": "text", + "content": "TKL" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 296, + 171, + 374, + 304 + ], + "blocks": [ + { + "bbox": [ + 296, + 171, + 374, + 304 + ], + "lines": [ + { + "bbox": [ + 296, + 171, + 374, + 304 + ], + "spans": [ + { + "bbox": [ + 296, + 171, + 374, + 304 + ], + "type": "image", + "image_path": "7b6d3a7397f89a7ad381452dae980d8886a93b5e119d01fa335b19497e19b340.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 305, + 365, + 312 + ], + "lines": [ + { + "bbox": [ + 302, + 305, + 365, + 312 + ], + "spans": [ + { + "bbox": [ + 302, + 305, + 365, + 312 + ], + "type": "text", + "content": "WeatherDiffusion" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 374, + 171, + 451, + 304 + ], + "blocks": [ + { + "bbox": [ + 374, + 171, + 451, + 304 + ], + "lines": [ + { + "bbox": [ + 374, + 171, + 451, + 304 + ], + "spans": [ + { + "bbox": [ + 374, + 171, + 451, + 304 + ], + "type": "image", + "image_path": "83527c929b785e3c02ead6bed4f5ae637df96fa20edc43fa797b969155f139ab.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 394, + 305, + 430, + 312 + ], + "lines": [ + { + "bbox": [ + 394, + 305, + 430, + 312 + ], + "spans": [ + { + "bbox": [ + 394, + 305, + 430, + 312 + ], + "type": "text", + "content": "ViWS-Net" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 451, + 171, + 529, + 304 + ], + "blocks": [ + { + "bbox": [ + 451, + 171, + 529, + 304 + ], + "lines": [ + { + "bbox": [ + 451, + 171, + 529, + 304 + ], + "spans": [ + { + "bbox": [ + 451, + 171, + 529, + 304 + ], + "type": "image", + "image_path": "0de0f733029bbfbeb11d2510974acd38d2ffe067038064e88bfb1f2cfe8986de.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 481, + 305, + 499, + 312 + ], + "lines": [ + { + "bbox": [ + 481, + 305, + 499, + 312 + ], + "spans": [ + { + "bbox": [ + 481, + 305, + 499, + 312 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 54, + 359, + 164, + 476 + ], + "blocks": [ + { + "bbox": [ + 46, + 317, + 547, + 351 + ], + "lines": [ + { + "bbox": [ + 46, + 317, + 547, + 351 + ], + "spans": [ + { + "bbox": [ + 46, + 317, + 547, + 351 + ], + "type": "text", + "content": "Figure 5. Visual comparison of all-in-one adverse weather removal methods on the selected real-world video sequences degraded by rain, haze, snow. Apparently, our network can more effectively remove rain streaks, haze, and snowflakes of input video frames than state-of-the-art methods. Please zoom in on the images for improved visualization." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 359, + 164, + 476 + ], + "lines": [ + { + "bbox": [ + 54, + 359, + 164, + 476 + ], + "spans": [ + { + "bbox": [ + 54, + 359, + 164, + 476 + ], + "type": "image", + "image_path": "f3f626ef84e63794ca748132c2f7d9905f6dd84c482546657559fe53aa77cc67.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 165, + 359, + 279, + 476 + ], + "blocks": [ + { + "bbox": [ + 165, + 359, + 279, + 476 + ], + "lines": [ + { + "bbox": [ + 165, + 359, + 279, + 476 + ], + "spans": [ + { + "bbox": [ + 165, + 359, + 279, + 476 + ], + "type": "image", + "image_path": "8aad17c2ef787788a2c858fcce686faa1bc258f89443ca9cdbe8462ac8f82d68.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 480, + 288, + 513 + ], + "lines": [ + { + "bbox": [ + 46, + 480, + 288, + 513 + ], + "spans": [ + { + "bbox": [ + 46, + 480, + 288, + 513 + ], + "type": "text", + "content": "Figure 6. Detailed results on each case of VRDS and RVSD by our proposed Diff-TTA. The proposed Diff-TTA brings consistent improvement across various cases." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 522, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 289, + 713 + ], + "type": "text", + "content": "in Table 3. We present the results trained on the mixed training set and tested on three weather-specific testing sets (rain, haze, snow) in view of PSNR and SSIM. The baseline model M1, i.e., the vanilla NAFNet, achieves the average performance on the three adverse weather datasets of 26.83, 0.9264 in PSNR, SSIM, respectively. M2 is equipped with diffusion process and trained by noise estimation, which advances M1 on the average performance by 1.69, 0.0078 of PSNR, SSIM, respectively. M3 introduces the proposed temporal noise model by replacing the vanilla noise model, and brings about an improvement of 0.63, 0.0083 in average PSNR, SSIM, respectively. Finally, our full model employs diffusion test-time adaptation (Diff-TTA) and gains a critical increase of 0.48, 0.0028 in PSNR, SSIM, respectively, compared to M3. Figure 6 further ablates our iterative adaptation on each case of the two unseen datasets, demonstrat" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 361, + 425, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 361, + 425, + 373 + ], + "spans": [ + { + "bbox": [ + 306, + 361, + 425, + 373 + ], + "type": "text", + "content": "ing consistent improvements." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 390, + 378, + 402 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 390, + 378, + 402 + ], + "spans": [ + { + "bbox": [ + 306, + 390, + 378, + 402 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 412, + 547, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 412, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 412, + 547, + 628 + ], + "type": "text", + "content": "This paper presents Diff-TTA, a diffusion-based test-time adaptation method capable of restoring videos degraded by various, even unknown, weather conditions. The proposed approach involves constructing a diffusion-based network and utilizing a novel temporal noise model to establish correlations among video frames. More importantly, a proxy task Diff-TSC is incorporated into the diffusion reverse process to iteratively update the deep model. This allows for the efficient adaptation of vision systems in real-world scenarios without modifying the source training process, making it easily applicable to a wide range of off-the-shelf diffusion-based models for enhancing weather removal. Experimental results on both synthetic and real-world data, which encompass diverse weather conditions, confirm the effectiveness and generalization capability of the proposed approach. We believe Diff-TTA can be a compelling baseline that shifts the focus of the community from the complicated training process to online test-time adaptation." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": "Acknowledgment. This work is supported by Guangzhou Municipal Science and Technology Project (Grant No. 2023A03J0671), Guangzhou-HKUST(GZ) Joint Funding Program (No. 2024A03J0618), a grant of General Research Fund from Hong Kong Research Grants Council (project no. 15218521), and the InnoHK funding launched by Innovation and Technology Commission, Hong Kong SAR." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "25613" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Malik Boudiaf, Romain Mueller, Ismail Ben Ayed, and Luca Bertinetto. Parameter-free online test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8344-8353, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 288, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 288, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 288, + 179 + ], + "type": "text", + "content": "[2] George Box. Box and jenkins: time series analysis, forecasting and control. In A Very British Affair: Six Britons and the Development of Time Series Analysis During the 20th Century, pages 161-215. Springer, 2013. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 288, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 288, + 224 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 288, + 224 + ], + "type": "text", + "content": "[3] Dian Chen, Dequan Wang, Trevor Darrell, and Sayna Ebrahimi. Contrastive test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 295-305, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 225, + 288, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 225, + 288, + 280 + ], + "spans": [ + { + "bbox": [ + 53, + 225, + 288, + 280 + ], + "type": "text", + "content": "[4] Haoyu Chen, Jingjing Ren, Jinjin Gu, Hongtao Wu, Xuequan Lu, Haoming Cai, and Lei Zhu. Snow removal in video: A new dataset and a novel method. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13211-13222, 2023. 1, 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 281, + 288, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 281, + 288, + 335 + ], + "spans": [ + { + "bbox": [ + 53, + 281, + 288, + 335 + ], + "type": "text", + "content": "[5] Jie Chen, Cheen-Hau Tan, Junhui Hou, Lap-Pui Chau, and He Li. Robust video content alignment and compensation for rain removal in a cnn framework. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6286-6295, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 336, + 288, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 336, + 288, + 379 + ], + "spans": [ + { + "bbox": [ + 53, + 336, + 288, + 379 + ], + "type": "text", + "content": "[6] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European Conference on Computer Vision, pages 17-33. Springer, 2022. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 380, + 288, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 288, + 435 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 288, + 435 + ], + "type": "text", + "content": "[7] Sixiang Chen, Tian Ye, Yun Liu, Jinbin Bai, Haoyu Chen, Yunlong Lin, Jun Shi, and Erkang Chen. CPLformer: Cross-scale prototype learning transformer for image snow removal. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4228-4239, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 436, + 288, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 436, + 288, + 502 + ], + "spans": [ + { + "bbox": [ + 53, + 436, + 288, + 502 + ], + "type": "text", + "content": "[8] Sixiang Chen, Tian Ye, Chenghao Xue, Haoyu Chen, Yun Liu, Erkang Chen, and Lei Zhu. Uncertainty-driven dynamic degradation perceiving and background modeling for efficient single image desnowing. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4269-4280, 2023." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 502, + 288, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 502, + 288, + 568 + ], + "spans": [ + { + "bbox": [ + 53, + 502, + 288, + 568 + ], + "type": "text", + "content": "[9] Wei-Ting Chen, Hao-Yu Fang, Jian-Jiun Ding, Cheng-Che Tsai, and Sy-Yen Kuo. Jstasr: Joint size and transparency-aware snow removal algorithm based on modified partial convolution and veiling effect removal. In European Conference on Computer Vision, pages 754-770. Springer, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 570, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 635 + ], + "type": "text", + "content": "[10] Wei-Ting Chen, Hao-Yu Fang, Cheng-Lin Hsieh, Cheng-Che Tsai, I Chen, Jian-Jiun Ding, Sy-Yen Kuo, et al. All snow removed: Single image desnowing algorithm using hierarchical dual-tree complex wavelet representation and contradict channel loss. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4196-4205, 2021. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 636, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 288, + 712 + ], + "type": "text", + "content": "[11] Wei-Ting Chen, Zhi-Kai Huang, Cheng-Che Tsai, Hao-Hsiang Yang, Jian-Jiun Ding, and Sy-Yen Kuo. Learning multiple adverse weather removal via two-stage knowledge learning and multi-contrastive regularization: Toward a unified model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17653–17662, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[12] Xiangning Chen, Chen Liang, Da Huang, Esteban Real, Kaiyuan Wang, Yao Liu, Hieu Pham, Xuanyi Dong, Thang Luong, Cho-Jui Hsieh, et al. Symbolic discovery of optimization algorithms. arXiv preprint arXiv:2302.06675, 2023.6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 129, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 172 + ], + "type": "text", + "content": "[13] Bodong Cheng, Juncheng Li, Ying Chen, Shuyi Zhang, and Tieyong Zeng. Snow mask guided adaptive residual network for image snow removal. arXiv preprint arXiv:2207.04754, 2022.6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 174, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 174, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 174, + 545, + 207 + ], + "type": "text", + "content": "[14] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 1, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 208, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 208, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 208, + 545, + 262 + ], + "type": "text", + "content": "[15] Hang Dong, Jinshan Pan, Lei Xiang, Zhe Hu, Xinyi Zhang, Fei Wang, and Ming-Hsuan Yang. Multi-scale boosted de-hazing network with dense feature fusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2157-2167, 2020. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 264, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 297 + ], + "type": "text", + "content": "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 298, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 298, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 308, + 298, + 545, + 376 + ], + "type": "text", + "content": "[17] Minhao Hu, Tao Song, Yujun Gu, Xiangde Luo, Jieneng Chen, Yinan Chen, Ya Zhang, and Shaoting Zhang. Fully test-time adaptation for image segmentation. In Medical Image Computing and Computer Assisted Intervention—MICCAI 2021: 24th International Conference, Strasbourg, France, September 27–October 1, 2021, Proceedings, Part III 24, pages 251–260. Springer, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 376, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 545, + 420 + ], + "type": "text", + "content": "[18] Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for model-agnostic domain generalization. Advances in Neural Information Processing Systems, 34:2427-2440, 2021. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 421, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 421, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 308, + 421, + 545, + 454 + ], + "type": "text", + "content": "[19] Ashutosh Kulkarni, Prashant W Patil, Subrahmanyam Murala, and Sunil Gupta. Unified multi-weather visibility restoration. IEEE Transactions on Multimedia, 2022. 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 456, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 456, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 456, + 545, + 510 + ], + "type": "text", + "content": "[20] Minghan Li, Qi Xie, Qian Zhao, Wei Wei, Shuhang Gu, Jing Tao, and Deyu Meng. Video rain streak removal by multiscale convolutional sparse coding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6644-6653, 2018. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 512, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 512, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 512, + 545, + 555 + ], + "type": "text", + "content": "[21] Ruoteng Li, Robby T Tan, and Loong-Fah Cheong. All in one bad weather removal using architectural search. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3175-3185, 2020. 1, 2, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 557, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 557, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 308, + 557, + 545, + 601 + ], + "type": "text", + "content": "[22] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In International conference on machine learning, pages 6028-6039. PMLR, 2020. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 602, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 602, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 308, + 602, + 545, + 656 + ], + "type": "text", + "content": "[23] Jingyun Liang, Yuchen Fan, Xiaoyu Xiang, Rakesh Ranjan, Eddy Ilg, Simon Green, Jiezhang Cao, Kai Zhang, Radu Timofte, and Luc Van Gool. Recurrent video restoration transformer with guided deformable attention. arXiv preprint arXiv:2206.02146, 2022. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "text", + "content": "[24] Jiaying Liu, Wenhan Yang, Shuai Yang, and Zongming Guo. Erase or fill? deep joint recurrent rain removal and reconstruction in videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3233-3242, 2018. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "25614" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 128 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 128 + ], + "type": "text", + "content": "[25] Quande Liu, Cheng Chen, Qi Dou, and Pheng-Ann Heng. Single-domain generalization in medical image segmentation via test-time adaptation from shape dictionary. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1756–1764, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 130, + 288, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 130, + 288, + 174 + ], + "spans": [ + { + "bbox": [ + 49, + 130, + 288, + 174 + ], + "type": "text", + "content": "[26] Xiaohong Liu, Yongrui Ma, Zhihao Shi, and Jun Chen. Grid-dehazenet: Attention-based multi-scale network for image dehazing. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7314-7323, 2019. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 176, + 288, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 176, + 288, + 242 + ], + "spans": [ + { + "bbox": [ + 48, + 176, + 288, + 242 + ], + "type": "text", + "content": "[27] Xihui Liu, Dong Huk Park, Samaneh Azadi, Gong Zhang, Arman Chopikyan, Yuxiao Hu, Humphrey Shi, Anna Rohrbach, and Trevor Darrell. More control for free! image synthesis with semantic diffusion guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 289-299, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 244, + 287, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 244, + 287, + 288 + ], + "spans": [ + { + "bbox": [ + 48, + 244, + 287, + 288 + ], + "type": "text", + "content": "[28] Ye Liu, Liang Wan, Huazhu Fu, Jing Qin, and Lei Zhu. Phase-based memory network for video dehazing. In Proceedings of the 30th ACM International Conference on Multimedia, pages 5427-5435, 2022. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 290, + 287, + 333 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 290, + 287, + 333 + ], + "spans": [ + { + "bbox": [ + 48, + 290, + 287, + 333 + ], + "type": "text", + "content": "[29] Yun-Fu Liu, Da-Wei Jaw, Shih-Chia Huang, and Jenq-Neng Hwang. Desnownet: Context-aware deep network for snow removal. IEEE Transactions on Image Processing, 27(6): 3064–3073, 2018. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 336, + 287, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 336, + 287, + 379 + ], + "spans": [ + { + "bbox": [ + 48, + 336, + 287, + 379 + ], + "type": "text", + "content": "[30] Ziwei Luo, Fredrik K Gustafsson, Zheng Zhao, Jens Sjolund, and Thomas B Schon. Image restoration with meanreverting stochastic differential equations. arXiv preprint arXiv:2301.11699, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 382, + 287, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 382, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 48, + 382, + 287, + 435 + ], + "type": "text", + "content": "[31] A Tuan Nguyen, Thanh Nguyen-Tang, Ser-Nam Lim, and Philip HS Torr. Tipi: Test time adaptation with transformation invariance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24162-24171, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 439, + 287, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 439, + 287, + 493 + ], + "spans": [ + { + "bbox": [ + 48, + 439, + 287, + 493 + ], + "type": "text", + "content": "[32] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yafo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. In International conference on machine learning, pages 16888-16905. PMLR, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 496, + 287, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 496, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 48, + 496, + 287, + 540 + ], + "type": "text", + "content": "[33] Ozan Özdenizci and Robert Legenstein. Restoring vision in adverse weather conditions with patch-based denoising diffusion models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1, 3, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 543, + 287, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 543, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 48, + 543, + 287, + 597 + ], + "type": "text", + "content": "[34] Dongwei Ren, Wangmeng Zuo, Qinghua Hu, Pengfei Zhu, and Deyu Meng. Progressive image deraining networks: A better and simpler baseline. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3937-3946, 2019. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 599, + 287, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 599, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 599, + 287, + 643 + ], + "type": "text", + "content": "[35] Wenqi Ren, Jingang Zhang, Xiangyu Xu, Lin Ma, Xiaochun Cao, Gaofeng Meng, and Wei Liu. Deep video dehazing with semantic segmentation. IEEE transactions on image processing, 28(4):1895-1908, 2018. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 646, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 677 + ], + "type": "text", + "content": "[36] Jaskirat Singh, Stephen Gould, and Liang Zheng. High-fidelity guided image synthesis with latent diffusion models. arXiv preprint arXiv:2211.17084, 2022. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "text", + "content": "[37] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2, 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "text", + "content": "[38] Jeya Maria Jose Valanarasu, Rajeev Yasarla, and Vishal M Patel. Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2353-2363, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 161 + ], + "type": "text", + "content": "[39] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 163, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 206 + ], + "type": "text", + "content": "[40] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726, 2020. 2, 3, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 209, + 547, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 547, + 252 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 547, + 252 + ], + "type": "text", + "content": "[41] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7201-7211, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 254, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 254, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 254, + 545, + 308 + ], + "type": "text", + "content": "[42] Shuai Wang, Lei Zhu, Huazhu Fu, Jing Qin, Carola-Bibiane Schonlieb, Wei Feng, and Song Wang. Rethinking video rain streak removal: A new synthesis model and a deraining network with video rain prior. In European Conference on Computer Vision, pages 565-582. Springer, 2022. 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 310, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 310, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 308, + 310, + 545, + 365 + ], + "type": "text", + "content": "[43] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced deformable convolutional networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0-0, 2019. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 366, + 545, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 366, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 308, + 366, + 545, + 419 + ], + "type": "text", + "content": "[44] Hongtao Wu, Yijun Yang, Haoyu Chen, Jingjing Ren, and Lei Zhu. Mask-guided progressive network for joint raindrop and rain streak removal in videos. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7216-7225, 2023. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 422, + 545, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 545, + 465 + ], + "type": "text", + "content": "[45] Bin Xia, Yulun Zhang, Shiyin Wang, Yitong Wang, Xinglong Wu, Yapeng Tian, Wenming Yang, and Luc Van Gool. Diffir: Efficient diffusion model for image restoration. arXiv preprint arXiv:2303.09472, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 468, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 468, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 468, + 545, + 521 + ], + "type": "text", + "content": "[46] Jiaqi Xu, Xiaowei Hu, Lei Zhu, Qi Dou, Jifeng Dai, Yu Qiao, and Pheng-Ann Heng. Video dehazing via a multi-range temporal alignment network with physical prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18053–18062, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 523, + 545, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 523, + 545, + 577 + ], + "spans": [ + { + "bbox": [ + 308, + 523, + 545, + 577 + ], + "type": "text", + "content": "[47] Wending Yan, Robby T Tan, Wenhan Yang, and Dengxin Dai. Self-aligned video deraining with transmission-depth consistency. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11966-11976, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 579, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 579, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 308, + 579, + 545, + 633 + ], + "type": "text", + "content": "[48] Hongzheng Yang, Cheng Chen, Meirui Jiang, Quande Liu, Jianfeng Cao, Pheng Ann Heng, and Qi Dou. Dltta: Dynamic learning rate for test-time adaptation on cross-domain medical images. IEEE Transactions on Medical Imaging, 41 (12):3575-3586, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 635, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 679 + ], + "type": "text", + "content": "[49] Wenhan Yang, Jiaying Liu, and Jiashi Feng. Frameconsistent recurrent video deraining with dual-level flow. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1661-1670, 2019. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 681, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 547, + 713 + ], + "type": "text", + "content": "[50] Wenhan Yang, Robby T Tan, Shiqi Wang, and Jiaying Liu. Self-learning video rain streak removal: When cyclic consistency meets temporal correspondence. In Proceedings of" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25615" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "text", + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1720-1729, 2020. 2, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 162 + ], + "type": "text", + "content": "[51] Yijun Yang, Angelica I Aviles-Rivero, Huazhu Fu, Ye Liu, Weiming Wang, and Lei Zhu. Video adverse-weather-component suppression network via weather messenger and adversarial backpropagation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13200-13210, 2023. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 218 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 218 + ], + "type": "text", + "content": "[52] Tian Ye, Sixiang Chen, Jinbin Bai, Jun Shi, Chenghao Xue, Jingxia Jiang, Junjie Yin, Erkang Chen, and Yun Liu. Adverse weather removal with codebook priors. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12653-12664, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 220, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 274 + ], + "type": "text", + "content": "[53] Zongsheng Yue, Jianwen Xie, Qian Zhao, and Deyu Meng. Semi-supervised video deraining with dynamical rain generator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 642-652, 2021. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 276, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 287, + 330 + ], + "type": "text", + "content": "[54] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14821-14831, 2021. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 331, + 287, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 287, + 374 + ], + "type": "text", + "content": "[55] Jian Zhang, Lei Qi, Yinghuan Shi, and Yang Gao. Domain adaptor: A novel approach to test-time adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18971-18981, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 376, + 287, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 287, + 421 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 287, + 421 + ], + "type": "text", + "content": "[56] Kaihao Zhang, Rongqing Li, Yanjiang Yu, Wenhan Luo, and Changsheng Li. Deep dense multi-scale network for snow removal using semantic and depth priors. IEEE Transactions on Image Processing, 30:7419-7431, 2021. 2, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 422, + 287, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 465 + ], + "type": "text", + "content": "[57] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 467, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 467, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 287, + 510 + ], + "type": "text", + "content": "[58] Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test time robustness via adaptation and augmentation. Advances in Neural Information Processing Systems, 35: 38629-38642, 2022. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 512, + 287, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 287, + 577 + ], + "type": "text", + "content": "[59] Xinyi Zhang, Hang Dong, Jinshan Pan, Chao Zhu, Ying Tai, Chengjie Wang, Jilin Li, Feiyue Huang, and Fei Wang. Learning to restore hazy video: A new real-world dataset and a new method. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9239-9248, 2021. 1, 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 579, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 287, + 622 + ], + "type": "text", + "content": "[60] Yi Zhang, Xiaoyu Shi, Dasong Li, Xiaogang Wang, Jian Wang, and Hongsheng Li. A unified conditional framework for diffusion-based image restoration. arXiv preprint arXiv:2305.20049, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 624, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 677 + ], + "type": "text", + "content": "[61] Kun Zhou, Wenbo Li, Liying Lu, Xiaoguang Han, and Jiangbo Lu. Revisiting temporal alignment for video restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6053-6062, 2022. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[62] Yurui Zhu, Tianyu Wang, Xueyang Fu, Xuanyu Yang, Xin Guo, Jifeng Dai, Yu Qiao, and Xiaowei Hu. Learning weather-general and weather-specific features for image" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 326, + 72, + 545, + 116 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 72, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 545, + 116 + ], + "type": "text", + "content": "restoration under multiple adverse weather conditions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21747-21758, 2023. 3, 6, 7" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "25616" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/GeoAuxNet_ Towards Universal 3D Representation Learning for Multi-sensor Point Clouds/f223747a-d407-4fbb-9660-d95973ceefec_content_list.json b/2024/GeoAuxNet_ Towards Universal 3D Representation Learning for Multi-sensor Point Clouds/f223747a-d407-4fbb-9660-d95973ceefec_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..936d4a7daefb8e73156cd713d2f45d7a99f86303 --- /dev/null +++ b/2024/GeoAuxNet_ Towards Universal 3D Representation Learning for Multi-sensor Point Clouds/f223747a-d407-4fbb-9660-d95973ceefec_content_list.json @@ -0,0 +1,1805 @@ +[ + { + "type": "text", + "text": "GeoAuxNet: Towards Universal 3D Representation Learning for Multi-sensor Point Clouds", + "text_level": 1, + "bbox": [ + 94, + 128, + 875, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shengjun Zhang $^{1}$ , Xin Fei $^{2}$ , Yueqi Duan $^{1\\dagger}$", + "bbox": [ + 316, + 202, + 651, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Department of Electronic Engineering, Tsinghua University", + "bbox": [ + 243, + 220, + 727, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Department of Automation, Tsinghua University", + "bbox": [ + 285, + 238, + 683, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zhangsj23, feix21}@mails.tsinghua.edu.cn, duanyueqi@tsinghua.edu.cn", + "bbox": [ + 183, + 258, + 781, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/zhangshengjun2019/GeoAuxNet", + "bbox": [ + 256, + 282, + 712, + 299 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "S3DIS RGB-D", + "bbox": [ + 96, + 340, + 145, + 367 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ScanNet RGB-D", + "bbox": [ + 88, + 411, + 151, + 436 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f9fa253af30bdc16ecfba1378f7a3ddd9777ca60233a5fec21806f2f77f2721c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 153, + 306, + 274, + 388 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/035c1fc49a6ab46e0a0ec9be69ccabbf12b2f3e8730b1d4b0c3b72c947edc90c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 153, + 388, + 274, + 458 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1b6ecc3f2bddbef315988f6667e68e67d220548f9ed23eba17ec40434e367129.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 276, + 306, + 398, + 388 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0eff91df56662770d7c04c32fc45ff088170b7e0d937420805c3787e023b1bb0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 276, + 388, + 397, + 458 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4288e65a0a4405c27bf51de3bea7abbb0e6ef3cd4a35ea28e825a66c4f3df4c6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 398, + 306, + 519, + 388 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/621246fc57fdf4ad71411102708842118b71049542d1d8bb3901955463b1ff09.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 398, + 388, + 519, + 458 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ebf1348708d946905c1f90c3095b06f98e15f3af7902ee1c44789c6f78333c6b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 521, + 306, + 642, + 388 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/65d0de5372e84dc7f045cb588678bf106d5e6b38859d33b7b6415ff7585fbb21.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 521, + 388, + 642, + 458 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1775f9fac4575a54898da131783cf35497e2db0284522493a08ce95d1b0f55df.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 642, + 306, + 764, + 388 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2aa5d9a94727c261f3587f3d4d6c07f539c54f92b00cbd98b3f99cd156bb6440.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 642, + 388, + 764, + 458 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d2f7654ed92c096ba16bdc162a4848135cde7d76f667090f2b437b1751bc32c3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 306, + 887, + 388 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/256874226c954a1d8958a648d2b8fd3cfad654fa70f96541367e39458f901402.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 388, + 888, + 458 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Semantic KITTI LiDAR", + "bbox": [ + 86, + 481, + 153, + 522 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/69eef06fd72fbe39980ced831df56c9a5d39d7a6255f89f8b1225fe5e5942f78.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 153, + 459, + 274, + 540 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/77bb9faf76bbb8dce0e0ce68d26f103f03b94b1792ef833934bf2ffc86307ada.jpg", + "image_caption": [ + "MinkowskiNet" + ], + "image_footnote": [], + "bbox": [ + 276, + 459, + 397, + 540 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4a7d5b77ae2634146c7d64f9b78bf61e9005f0c302bbf4ebcdfd1ea6bd1350f1.jpg", + "image_caption": [ + "SPVCNN", + "Figure 1. Semantic segmentation results on S3DIS [1] and ScanNet [9] from RGB-D cameras and SemanticKITTI [2] from LiDAR. For all methods, we trained collectively on three datasets. Our method outperforms other methods with better detailed structures." + ], + "image_footnote": [], + "bbox": [ + 397, + 459, + 519, + 540 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8fcd188bf04145dd4b48723f0e5e564e0a22d61eeb83f3eed88dad17d206e0c7.jpg", + "image_caption": [ + "PPT" + ], + "image_footnote": [], + "bbox": [ + 519, + 459, + 642, + 540 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/458920fd3ebf2201c5a4080515d84a96c25ee7e09a2ca70f61a0c96cf29e7642.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 642, + 459, + 764, + 540 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ae239ccb8a8bb17d5d16ca105420545cc08c4a3557f86fcb16669813cd372c6f.jpg", + "image_caption": [ + "GroundTruth" + ], + "image_footnote": [], + "bbox": [ + 764, + 459, + 888, + 540 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 618, + 313, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Point clouds captured by different sensors such as RGB-D cameras and LiDAR possess non-negligible domain gaps. Most existing methods design different network architectures and train separately on point clouds from various sensors. Typically, point-based methods achieve outstanding performances on even-distributed dense point clouds from RGB-D cameras, while voxel-based methods are more efficient for large-range sparse LiDAR point clouds. In this paper, we propose geometry-to-voxel auxiliary learning to enable voxel representations to access point-level geometric information, which supports better generalisation of the voxel-based backbone with additional interpretations of multi-sensor point clouds. Specifically, we construct hierarchical geometry pools generated by a voxel-guided dynamic point network, which efficiently provide", + "bbox": [ + 73, + 650, + 470, + 877 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "auxiliary fine-grained geometric information adapted to different stages of voxel features. We conduct experiments on joint multi-sensor datasets to demonstrate the effectiveness of GeoAuxNet. Enjoying elaborate geometric information, our method outperforms other models collectively trained on multi-sensor datasets, and achieve competitive results with the-state-of-art experts on each single dataset.", + "bbox": [ + 500, + 619, + 890, + 726 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 755, + 632, + 770 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Point cloud analysis has attracted widespread attention due to its growing applications, such as autonomous driving [20, 45, 51] and robotics [4, 40]. Unlike images which are represented by regular pixels, 3D point clouds are irregular and unordered. These characteristics aggravate the inconsistency on density and sampling patterns of point clouds captured by different sensors, such as RGB-D cameras [1, 9, 33] and LiDAR [2, 5, 34].", + "bbox": [ + 496, + 780, + 890, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding author.", + "bbox": [ + 101, + 887, + 227, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20019", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Typically, points generated from RGB-D pictures are equally distributed and dense, while points scanned by LiDAR are sparse and uneven. The diversity on input data hinders the construction of universal network architectures. For point clouds from RGB-D cameras, point-based methods [28, 29, 54] usually gather local information via grouping and aggregating neighborhood features to extract detailed spatial information. Since neighbors are not stored contiguously in point representations, indexing them requires the costly searching operations [24], which limits the application of these methods on large scale LiDAR point clouds. To address this challenge, voxel-based methods [12, 25, 31] leverage the regular 3D convolution to obtain the contiguous memory access pattern. However, the resolution of voxels is constrained by the memory and becomes lower at deeper stages, resulting in loss of local information [35]. An intuitive idea is to ensemble both point-based and voxel-based methods [24, 35, 52], yet they are still confronted with either high time-consumption or the lack of detailed geometric features, which fail to simultaneously process large scale point clouds from different sensors with fine-grained spatial information while maintaining efficiency.", + "bbox": [ + 80, + 90, + 467, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose GeoAuxNet to provide point-level geometric information for voxel representations in an efficient manner with geometry-to-voxel auxiliary learning. The support of elaborate geometric features introduces additional point-level information, which cannot be fully exploited by voxel-based backbones. Motivated by the observation that the local geometric structures present high similarity at each stage of the network, we construct hierarchical geometry pools to learn auxiliary sensor-aware point-level geometric priors corresponding to different layers for sensor-agnostic voxel features. To update our geometry pools, we present a voxel-guided dynamic point network, where we leverage prior knowledge in voxel features to guide high quality spatial feature extraction. Then, we fuse elaborate geometric features in the pools into voxel representations via our designed geometry-to-voxel auxiliary mechanism. For large-scale point clouds, the hierarchical geometry pools store representative point-level features at different stages, which efficiently provide complementary geometric information without using the point-based networks during inference time.", + "bbox": [ + 80, + 443, + 467, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct extensive experiments on multi-sensor datasets, including S3DIS [1] and ScanNet [9] from RGB-D cameras as well as SemanticKITTI [2] captured by LiDAR, to demonstrate the efficiency and effectiveness of our method. With a shared weight of backbone, our method outperforms other models trained on joint datasets from different sensors and achieves competitive performance with experts on single dataset. As shown in Figure 1, we trained MinkowskiNet [8], SPVCNN [35], PPT [43] and", + "bbox": [ + 80, + 763, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "GeoAuxNet on the above three datasets collectively, where our method preserve better detailed structures for point clouds from various sensors.", + "bbox": [ + 503, + 92, + 890, + 133 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 503, + 148, + 645, + 162 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Scene Understanding. Since images are composed of regular pixels, point clouds generated from RGB-D cameras are well-distributed and dense. Voxelization on dense point clouds ignore detailed structure information. To directly learn from points, Qi et al. [27] first proposed PointNet to process point clouds through shared multi-layer perceptions (MLP) to get a global representation. Yet this method lacks the ability to comprehend local geometric features. Therefore, PointNet++ [28] built a hierarchical architecture to gather local information via grouping and aggregating neighborhood features. Following up MLP-based works [7, 29] focused on designing effective learning schemes to capture spatial features. Other methods also introduced specific convolution operations [3, 19, 36, 48], graph-based representations [22, 39] and self-attention mechanism [11, 16, 42, 49, 54] to explore the relationship between points. Apart from cameras, LiDAR is also a widely used sensor to collect point clouds. Some methods [15, 26, 41, 53] focused on projecting point clouds to 2D grids to utilize 2D CNN. However, 3D to 2D projection limits the representation quality of geometric structures [50] due to the sacrifice of 3D spatial information [48]. Point based methods have high time consumption because of the large scale and the sparsity of these point clouds. Voxel-based methods [12, 13, 17, 56, 57] show more robustness against the sparsity. In view of the sampling characteristic of LiDAR, Zhu et al. [56] introduced cylinder coordinate system, for nearby regions have greater density than farther away regions. Similarly, Lai et al. [17] proposed spherical transformer to aggregate information from dense close points to sparse distant ones. Voxel-based networks maintain efficiency when processing scene-level point clouds, but ignore some detailed spatial structures.", + "bbox": [ + 503, + 172, + 890, + 670 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Universal Learning from Multiple Data Sources. Large language models achieve remarkable progress on many natural language processing systems through pretraining on extremely large text corpora [10], which enables foundation models to process data across domains, involving multiple languages, various majors, diverse application scenarios and so on. In 2D computer vision, joint learning across domains improve model robustness for detection [38, 47, 55], depth estimation [30] and semantic segmentation [18, 37]. Wang et al. [38] proposed a universal object detector through a domain-attention mechanism. But they did not model the training differences between different datasets. Universal-RCNN [47] modeled the class relations by designing a inter-dataset attention module to in", + "bbox": [ + 503, + 689, + 890, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20020", + "bbox": [ + 480, + 946, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4e9e97d7b4d8c135ecae23d28cd421a78973969a9acfd51b27665b821a3fdea0.jpg", + "image_caption": [ + "Figure 2. The pipeline of our GeoAuxNet. For a complete scene point cloud $\\mathcal{P}^{\\mathcal{C}}$ and a local point patch $\\mathcal{P} \\subseteq \\mathcal{P}^{\\mathcal{C}}$ , our voxel-based backbone first voxelizes $\\mathcal{P}^{\\mathcal{C}}$ and conducts sparse convolutional operations. The voxel-guided hypernetwork takes relative positions, voxel features and a stage latent code as input to provide weights and biases for the point network. Then, we encode the spatial information for $\\mathcal{P}$ with the point network and aggregate local features to generate geometric feature candidates. Following the update strategy, we construct hierarchical geometry pools. The geometry-to-voxel mechanism fuses geometric features stored in the pools to enable voxel representations to access point-level geometric information. We repeat the above process several times to extract effective representation hierarchically and predict the results with a voxel decoder for the primary task and a point decoder for the auxiliary task. The dotted line stands for the course of the auxiliary learning which is ignored during inference to ensure efficiency. Geo-to-Vox is abbreviation of Geometry-to-Voxel." + ], + "image_footnote": [], + "bbox": [ + 76, + 87, + 890, + 338 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "corporates graph transfer learning for propagating relevant semantic information across multiple datasets. CDCL [37] pointed out that convolution filters can be shared across domains without accuracy loss, while normalization operators are not appropriate to share in view of the bias of statistics. Recently, Wu et al. [43] introduced point prompt training to optimize a learnable domain-specific prompt with language-guided categorical alignment, but processed point clouds from RGB-D and LiDAR separately. However, point clouds captured by different sensors have diverse density and distributions, which limits the exploration of the universal network on multiple data sources.", + "bbox": [ + 75, + 482, + 472, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Auxiliary Learning. Auxiliary learning aims to achieve outstanding performance on a single primary task with the assist of auxiliary tasks and has shown many successful applications in a wide range of fields, including knowledge distillation [46], transfer learning [14] and reinforcement learning [21]. For instance, DCAN [6] learned an auxiliary contour detection task for more accurate segmentation, while Araslanov investigated the joint learning of image classification and semantic segmentation. In 3D vision, MonoCon [23] introduced monocular contexts as auxiliary tasks in training for 3D object detection. Nevertheless, the inherent relationship between point-level features and voxel-level features is still under exploration to better leverage the advantages of two domains without any requirement", + "bbox": [ + 75, + 688, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of additional data.", + "bbox": [ + 500, + 482, + 624, + 496 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 510, + 591, + 526 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall framework of GeoAuxNet is illustrated in Figure 2. To generate point-level geometric features, we propose a voxel-guided hypernetwork to produce weights and biases for the point network. Then, point features are grouped and aggregated to extract fine-grained local features, which are employed to update geometry pools. For voxel-based backbone, we introduce geometry-to-voxel auxiliary mechanism to access elaborate spatial information in our hierarchical geometry pools.", + "bbox": [ + 496, + 535, + 890, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Voxel-guided Dynamic Point Network", + "text_level": 1, + "bbox": [ + 498, + 681, + 826, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Point-based methods are mostly confronted with high time consumption towards large scale point clouds, since the farthest point sampling (FPS) and k-nearest neighbors (kNN) algorithms in these networks have a time complexity of $\\mathcal{O}(n^2)$ . Therefore, to efficiently process large scale point clouds, input points are voxelized and passed through sparse convolution operations to extract voxel features in our voxel-based backbone.", + "bbox": [ + 496, + 704, + 890, + 823 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To generate point-level geometric information efficiently, we train our point network on local point patches instead of complete scene point clouds. However, less training data weakens the ability of the point network to extract local features. Considering that our voxel-based branch are", + "bbox": [ + 496, + 825, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20021", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "trained on complete voxelized point clouds, we treat voxel features as prior knowledge to instruct the learning of local geometric information. Yet, points inside a voxel share the same voxel feature, which ignores detailed difference between points. Thus, we also introduce relative coordinates to depict diversity of points within the same voxel. On account of the hierarchical architectures of both the point network and the voxel-based backbone with several stages, we optimize a learnable latent code $s$ for each stage during our auxiliary learning.", + "bbox": [ + 75, + 90, + 472, + 241 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a local point patch $\\mathcal{P} = \\{p_i, 1 \\leq i \\leq N\\} \\in \\mathbb{R}^{N \\times 3}$ with point features $\\mathcal{F}^{\\mathcal{P}} = \\{f_i^{\\mathcal{P}}, 1 \\leq i \\leq N\\}$ , each point $p_i$ belongs to a voxel $v$ with the voxel feature $f^{\\mathcal{V}}$ . Here $v \\in \\mathbb{R}^3$ stands for the center coordinate of the voxel. Then, we can obtain the weight and bias for point $p_i$ :", + "bbox": [ + 76, + 241, + 468, + 318 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nw _ {i} = h _ {w} \\left(h _ {p} \\left(p _ {i} - v\\right), f ^ {\\mathcal {V}}\\right) \\odot h _ {s} (s), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 325, + 468, + 344 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nb _ {i} = h _ {b} \\left(h _ {p} \\left(p _ {i} - v\\right), f ^ {\\mathcal {V}}\\right) \\odot h _ {s} (s), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 347, + 468, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $h_p$ and $h_s$ are MLPs to project relative positions and the stage latent code to specific dimensions, and $h_w$ and $h_b$ are hypernetworks to generate the weight and bias. Hence, new point features can be formulated as:", + "bbox": [ + 76, + 375, + 468, + 434 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {f} _ {i} ^ {\\mathcal {P}} = \\operatorname {M L P} \\left(w _ {i} \\odot f _ {i} ^ {\\mathcal {P}} + b _ {i}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 178, + 441, + 468, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To extract local geometric information, we use kNN algorithm to select $k$ nearest neighbor points and employ a max-pooling function to aggregate each local group. In this manner, we generate a set of geometric feature candidates for hierarchical geometry pools.", + "bbox": [ + 76, + 470, + 468, + 546 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Hierarchical Geometry Pools", + "text_level": 1, + "bbox": [ + 76, + 554, + 336, + 570 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The voxel branch ignores some detailed spatial structures due to voxelization. Motivated by the similarity of geometric structures in 3D space, we establish a pool to store typical geometric patterns. The pool contains $n$ typical spatial structure features $\\mathcal{F}^{\\mathcal{D}} = \\{f_i^{\\mathcal{D}}, 1 \\leq i \\leq n\\}$ . The max size of the pool is set to $N$ . In Section 3.1, we generate a set of geometric pattern candidates $\\mathcal{F}^S = \\{f_i^S, 1 \\leq i \\leq m\\}$ . When updating the pool, we first calculate the similarity between $\\mathcal{F}^S$ and $\\mathcal{F}^D$ :", + "bbox": [ + 76, + 577, + 468, + 713 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns _ {i} = \\max _ {1 \\leq j \\leq N} s _ {i, j}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 724, + 468, + 747 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 76, + 756, + 122, + 768 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns _ {i, j} = \\frac {f _ {i} ^ {\\mathcal {S}} \\cdot f _ {j} ^ {\\mathcal {D}}}{\\| f _ {i} ^ {\\mathcal {S}} \\| \\| f _ {j} ^ {\\mathcal {D}} \\|.} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 766, + 468, + 804 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, if $s_i \\geq \\epsilon$ , $f_i^S$ is merged into the geometry pool $\\mathcal{F}^{\\mathcal{D}}$ :", + "bbox": [ + 76, + 808, + 468, + 838 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nt = \\underset {1 \\leq j \\leq N} {\\operatorname {a r g m a x}} s _ {i, j}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 851, + 468, + 875 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {f} _ {t} ^ {\\mathcal {D}} = \\lambda f _ {t} ^ {\\mathcal {D}} + (1 - \\lambda) f _ {i} ^ {\\mathcal {S}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 878, + 468, + 897 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Update Pools" + ], + "code_body": "Input: New features $\\mathcal{F}^S$ , geometry pool $\\mathcal{F}^D$ , maximum size of the pool $N$ , threshold $\\epsilon$ , update rate $\\lambda$ \nOutput: New pool $\\hat{\\mathcal{F}}^D$ \nCompute $s_i$ for $1 \\leq i \\leq m$ according to Eq. 4 \nfor $i \\gets 1$ to $|\\mathcal{F}^S|$ do \n if $s_i \\geq \\epsilon$ then \n t ← arg max $_{1 \\leq j \\leq N}$ $s_i, j$ $f_t^D \\gets \\lambda f_t^D + (1 - \\lambda)f_i^S$ $\\mathcal{F}^S \\gets \\mathcal{F}^S \\setminus \\{f_i^S\\}$ \n end if \nend for \nif $|\\mathcal{F}^S| + |\\mathcal{F}^D| \\leq N$ then \n $\\mathcal{F}^D \\gets \\mathcal{F}^D \\cup \\mathcal{F}^S$ \nelse \n $\\hat{\\mathcal{F}}^S \\gets$ randomly select $N - |\\mathcal{F}^D|$ features from $\\mathcal{F}^S$ $\\mathcal{F}^S \\gets \\mathcal{F}^S \\setminus \\hat{\\mathcal{F}}^S$ $\\mathcal{F}^D \\gets \\mathcal{F}^D \\cup \\hat{\\mathcal{F}}^S$ \n for $i \\gets 1$ to $|\\mathcal{F}^S|$ do \n t ← arg max $_{1 \\leq j \\leq N}$ $s_i, j$ $f_t^D \\gets \\lambda f_t^D + (1 - \\lambda)f_i^S$ \n end for \nend if", + "bbox": [ + 500, + 109, + 890, + 445 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\epsilon$ is a threshold and $\\lambda$ stands for the update rate. Otherwise, we append the new features to the geometry pool. If the total number of features is more than $N$ , we randomly select $N - n$ new features and add them to the pool. The rest new features following Eq. 7 are merged into existing patterns. Our update strategy is illustrated in Algorithm 1.", + "bbox": [ + 496, + 474, + 890, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since voxel representations at various stages contains different level information, we introduce hierarchical geometry pools to store representative point-level features corresponding to different stages. We also enlarge the size of pools for deeper stages due to more complex geometric structures with larger receptive field sizes.", + "bbox": [ + 496, + 566, + 890, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Geometry-to-Voxel Auxiliary", + "text_level": 1, + "bbox": [ + 500, + 672, + 756, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After we update the geometric pool, we use geometric patterns stored in $\\mathcal{F}^{\\mathcal{D}}\\in \\mathbb{R}^{N\\times C}$ to generate fine-grained local information for each voxel following a cross-attention operation. Given voxel features $\\mathcal{F}^{\\nu} = \\{f_i^\\nu ,1\\leq i\\leq M\\} \\in$ $\\mathbb{R}^{M\\times D}$ , the geometric fine-grained feature $f_{i}^{\\mathcal{G}}$ for $f_{i}^{\\nu}\\in \\mathcal{F}^{\\nu}$ is formulated by:", + "bbox": [ + 496, + 696, + 890, + 787 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nq _ {i} = h _ {Q} \\left(f _ {i} ^ {\\mathcal {V}}\\right) \\in \\mathbb {R} ^ {1 \\times s}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 801, + 890, + 819 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nK = h _ {K} \\left(\\mathcal {F} ^ {\\mathcal {D}}\\right) \\in \\mathbb {R} ^ {N \\times s}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 821, + 890, + 840 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nV = h _ {V} \\left(\\mathcal {F} ^ {\\mathcal {D}}\\right) \\in \\mathbb {R} ^ {N \\times t}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 842, + 890, + 859 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {i} ^ {\\mathcal {G}} = \\operatorname {s o f t m a x} \\left(\\frac {q _ {i} K ^ {\\top}}{\\sqrt {s}}\\right) V \\in \\mathbb {R} ^ {1 \\times t}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 863, + 890, + 898 + ], + "page_idx": 3 + }, + { + "type": "footer", + "text": "20022", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $h_{Q}, h_{K}$ , and $h_{V}$ stand for projections. Then, we concatenate $f_{i}^{\\mathcal{G}}$ with $f_{i}^{\\mathcal{V}}$ as the new voxel feature and employ sparse convolution operations. In this manner, we can provide complementary geometric information without using the point-based networks during inference stage.", + "bbox": [ + 75, + 90, + 468, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Overall Architecture", + "text_level": 1, + "bbox": [ + 76, + 180, + 272, + 194 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given a point cloud $\\mathcal{P}_i^{\\mathcal{C}}$ with features $\\mathcal{F}_i^{\\mathcal{C}}$ and a local point patch $\\mathcal{P}_i\\subseteq \\mathcal{P}_i^{\\mathcal{C}}$ with $\\mathcal{F}_i^{\\mathcal{P}}\\subseteq \\mathcal{F}_i^{\\mathcal{C}}$ from sensor $S_{i}$ , we voxelize all points to obtain voxel coordinates $\\nu_{i}$ with voxel features $\\mathcal{F}_i^\\nu$ . Points obtained from RGB-D cameras usually possess colors and normal vectors, while points captured by LiDAR always possess intensity. Thus, we introduce sensor-aware input embedding. Specifically, we embed $\\mathcal{P}_i$ from sensor $S_{i}$ by:", + "bbox": [ + 75, + 204, + 468, + 325 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {F}} _ {i} ^ {\\mathcal {P}} = h _ {\\mathcal {S} _ {i}} ^ {\\mathcal {P}} \\left(\\mathcal {F} _ {i} ^ {P}\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 338, + 468, + 357 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {F}} _ {i} ^ {\\mathcal {V}} = h _ {\\mathcal {S} _ {i}} ^ {\\mathcal {V}} \\left(\\mathcal {F} _ {i} ^ {V}\\right), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 359, + 468, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $h_{\\mathcal{S}_i}^{\\mathcal{P}}$ means shared MLPs on points and $h_{\\mathcal{S}_i}^{\\mathcal{V}}$ means sparse convolution operations on voxels. Then, we learn voxel features and point features following the extraction strategy in Section 3.1 and update hierarchical geometry pools using Algorithm 1. We fuse point-level geometric information to voxel representation via geometry-to-voxel auxiliary mechanism. Our auxiliary task is to predict point labels for local point patches with point branch and point decoder. Therefore, the final loss function is formulated as:", + "bbox": [ + 75, + 392, + 468, + 529 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {v} + \\mu \\mathcal {L} _ {p}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 544, + 468, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_v$ and $\\mathcal{L}_p$ stand for per-voxel and per-point cross entropy loss and $\\mu$ is the weight of the point loss.", + "bbox": [ + 75, + 575, + 468, + 606 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Discussion", + "text_level": 1, + "bbox": [ + 76, + 619, + 194, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we highlight previous point-voxel networks [24, 35, 52] which combine a point-based branch and a voxel-based branch together to leverage the advantages of both efficient voxel representations and elaborate point features. Additionally, we compare our GeoAuxNet with these methods. PVCNN [24] and PVT [52] rely on point-based networks to extract high quality representations with time-consuming searching or attention mechanisms, while SPVCNN [35] does not explore detailed local geometric information. Since the interaction between two branches in these methods only contains voxelization and devoxelization operations, they fail to benefit from both voxel representations and point-level elaborate features. Our GeoAuxNet preserves hierarchical sensor-aware geometry pools as the bridge to enable voxel representations to absorb point-level local geometric information without using the point network during inference.", + "bbox": [ + 75, + 643, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 90, + 633, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We validate our proposed GeoAuxNet across point clouds from multiple sensors. In Section 4.1, we train various methods on joint datasets including S3DIS [1], ScanNet [9] and SemanticKITTI [2]. In Section 4.2, we analyze the effectiveness of our sensor-aware hierarchical geometry pools. In Section 4.3, we compare GeoAuxNet with typical point-based and voxel-based methods to demonstrate the efficiency of our designed geometry-to-voxel auxiliary learning. In Section 4.4, we ablate different design choices of our voxel-guided dynamic point network.", + "bbox": [ + 496, + 114, + 890, + 267 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Semantic Segmentation", + "text_level": 1, + "bbox": [ + 498, + 279, + 718, + 295 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. We conduct semantic segmentation experiments on the joint data of three datasets: S3DIS [1] and ScanNet [9], which are generated from RGB-D cameras, and SemanticKITTI [2], which is captured by LiDAR. The S3DIS dataset comprises 271 rooms from six areas in three distinct buildings, which are densely sampled on the surfaces of the meshes and annotated into 13 categories. Model performance evaluation is typically done using results from Area 5. The ScanNet dataset comprises 1,613 scene scans reconstructed from RGB-D frames. It is divided into 1,201 scenes for training, 312 scenes for validation, and 100 scenes for benchmark testing. SemanticKITTI consists of 22 point cloud sequences, where sequences 00 to 10, 08 and 11 to 21 are used for training, validation and testing. After merging classes with distinct moving status and discarding classes with very few points, a total of 19 classes are selected for training and evaluation.", + "bbox": [ + 496, + 303, + 890, + 560 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. Our voxel-based backbone is built on MinkowskiNet [8] with the batch normalization operations proposed by PPT [43]. The voxel size is set to $0.05\\mathrm{m}$ . We train our methods as well as MinkowskiNet SPVCNN [35] and PPT [43] on joint input data [1, 2, 9] captured by RGB-D cameras and LiDAR collectively. We also fine tune our method on three datasets separately. For the sake of fairness, we set the number of parameters of MinkowskiNet, SPVCNN, PPT and GeoAuxNet to about 60M. We use SGD optimizer and OneCycleLR [32] scheduler, with learning rate 0.05 and weight decay $10^{-4}$ . We adopt OneCycleLR [32] scheduler with $5\\%$ percentage of the cycle spent increasing the learning rate and a cosine annealing strategy. We train all models with data augmentation used in PPT [43]. We set $\\lambda$ in Eq. 7 to 0.1, $\\mu$ in Eq. 14 to 0.1, threshold for geometry pool update $\\epsilon$ to 0.9. The total number of iterations is equal to the sum of necessary iterations for all datasets. The sampling ratio across S3DIS, ScanNet and SemanticKITTI is set to 2:2:5. The number of points for the auxiliary branch does not exceed $2\\times 10^{4}$ . The max size of hierarchical geometry pools for each stage", + "bbox": [ + 496, + 583, + 892, + 901 + ], + "page_idx": 4 + }, + { + "type": "footer", + "text": "20023", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/2d64efa52801232fe321a2f66562aec1e35bda5bda3d747518ccafd06466d245.jpg", + "table_caption": [ + "Table 1. Semantic segmentation results on three benchmarks including S3DIS [1], ScanNet [9] and SemanticKITTI [2]. We train SPVCNN [35], PPT [43] and GeoAuxNet on the joint training data of three datasets and also compare with experts [16, 17, 29, 35, 54, 56] on each single dataset. We report the class-average accuracy (mAcc, %) and class-wise mean IoU (mIoU, %). The type stands for point-based $(p)$ or voxel-based $(\\nu)$ methods, where $p \\rightarrow \\nu$ means the geometry-to-voxel auxiliary learning." + ], + "table_footnote": [ + "† The original PPT with language-guided categorical alignment fails to converge on the joint training data. We use decoupled projection heads to employ PPT on multi-sensor datasets." + ], + "table_body": "
MethodTypeParams.S3DIS Area5ScanNetSemanticKITTI
Test mIoUTest mAccVal mIoUVal mAccVal mIoUVal mAcc
PointTransformer [54]p11.4M70.476.570.6-
StratifiedFromer [16]p18.8M72.078.174.3-
PointNeXt [29]p41.6M70.577.271.5-
SPVCNN [35]p+v21.8M63.8-
Cylinder3D [56]v26.1M64.3-
SphereFormer [17]v32.3M67.8-
MinkowskiNet [8]v60.9M18.821.325.630.536.041.2
SPVCNN [35]p+v61.0M58.662.356.760.452.056.7
PPT [43]v63.0M30.434.616.527.425.031.2
PPT †[43]v63.0M67.973.269.677.561.868.9
GeoAuxNetp→v64.7M69.5(+1.6)74.5(+1.3)71.3(+1.7)79.3(+1.8)63.8(+2.0)69.3(+0.4)
", + "bbox": [ + 76, + 155, + 901, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "is set to 32, 64, 128, 256 separately.", + "bbox": [ + 76, + 455, + 315, + 472 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results Comparison. As shown in Table 1, GeoAuxNet surpasses other methods which are trained on three datasets jointly. Our methods outperforms SPVCNN [35], which does not utilize its point branch to extract local information, by more than $10\\%$ in mIoU on each dataset. PPT [43] introduces different datasets as prompts to collaboratively train a single model on multiple datasets from the same sensor type. The language-guided alignment of PPT suffers from the large domain gap caused by different sensors, which fails to converge on the joint training datasets. Therefore, we replace it by the decoupled alignment with separate heads for each dataset, which is also defined in PPT [43]. We construct sensor-aware geometry pools to address this issue and outperform PPT by about $2\\%$ in mIoU on each dataset. Besides, we also compare with experts trained on each single dataset. We achieve competitive results with point-based methods on datasets from RGB-D cameras. More specifically, we even outperform PointTransformer in validation mIoU of ScanNet. Despite training on inter-domain datasets from different sensors, GeoAuxNet has a higher accuracy on SemanticKITTI than SPVCNN [35], and attain encouraging performances even compared with recent state-of-the-art expert [17]. Although we do not introduce any specific designs such as pointwise attention mechanism [16, 54] and cylinder coordinate system [56] for sparse LiDAR point clouds, the quantitative results demonstrate the effectiveness of geometry-to", + "bbox": [ + 75, + 492, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "voxel auxiliary learning by constructing hierarchical geometry pools to provide auxiliary sensor-aware point-level geometric priors for voxel representations.", + "bbox": [ + 496, + 455, + 890, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Hierarchical Geometry Pools", + "text_level": 1, + "bbox": [ + 498, + 511, + 759, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we further conduct experiments on our hierarchical geometry pools. First, we analyze the similarity of geometric structures in 3D space. Second, we discuss the appropriation to represent geometric information with our geometry pools. Finally, we study the construction of sensor-aware pools for different sensors.", + "bbox": [ + 496, + 534, + 893, + 626 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Similarity of Geometric Structures. Local geometric structures possess similarity in 3D space. For example, planar structures exist in desks, chairs and airplanes. These spatial structures are not unrelated but consistent. To further analyze the similarity of geometric structures learned by networks, we train our point encoder on ModelNet40 [44]. For each point, we use kNN to search 16 nearest neighbors and aggregate them via a max-pooling function to generate local geometric features. We select a center point marked as a purple star in Figure 3a and Figure 3c, and calculate the cosine similarity between its features and other point features. Specifically, the wing and the tail of the plane in Figure 3a have similar local structures, resulting in high similarity of features in Figure 3b. Figure 3d and Figure 3f show the same phenomenon at different stages in the network, where planar structures are similar to each other, while different from the chair legs. Since features from deeper lay", + "bbox": [ + 496, + 643, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20024", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ef135819b164502f5a8716d592ff894fe59ad82d2f7aeed5c4b39a9c5444a93f.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 112, + 89, + 228, + 170 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/18d61ff77136088f9be03e48200993f8e82bfe4e961130ea9af124df715c51de.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 240, + 89, + 356, + 170 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fae73d3bdaa19de92d09a6aaf24785df40088d025f09e9d392588855884bc009.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 362, + 90, + 475, + 170 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/028dbe805c154ec5e6f80049a3b872fe6c1a4f765d613e3622e9e92df32817dd.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 482, + 90, + 591, + 169 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/031ee3badb36e5675dd67436c80e7f951c71ec29cd4edb11a8b23a770c9f4d63.jpg", + "image_caption": [ + "(e)" + ], + "image_footnote": [], + "bbox": [ + 598, + 93, + 710, + 169 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d7112883770caceb463c692f7bb6835effa122c18891681a0aa58d5ace5163f5.jpg", + "image_caption": [ + "(f)" + ], + "image_footnote": [], + "bbox": [ + 718, + 90, + 831, + 170 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/38ad5ca00e8f04f0cf84398e28444d7880664b102b2882d808bd3fb2b6c093c9.jpg", + "image_caption": [ + "Figure 3. Visualization of the cosine similarity between features. The purple star is the selected point, and the green star is the point with a significantly similar feature to the purple star. We calculate the cosine similarity between the feature of the red star and other features and visualize them in image (b), (d) and (f). The nearest neighbors of the purple and green stars are marked red in image (a), (c) and (e). (a), (b), (c) and (e) are generated from the first stage of the point network and (e), (f) are from the second stage." + ], + "image_footnote": [], + "bbox": [ + 81, + 279, + 267, + 386 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d13192e61c6877c92ddf92b5d16c9fcec796fc89c388f8642a298cba067f4139.jpg", + "image_caption": [ + "(b) Stage Two" + ], + "image_footnote": [], + "bbox": [ + 281, + 280, + 464, + 386 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d0de33688bd881555a4a86954612f05813e3b50aa2e2025a4abd43351de8d607.jpg", + "image_caption": [ + "(a) Stage One", + "(c) Stage Three", + "Figure 4. Statistical results of cosine similarity between hierarchical geometry pools and point-level features extracted by the point network from point clouds in S3DIS [1] at different stages." + ], + "image_footnote": [], + "bbox": [ + 81, + 405, + 264, + 511 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a02125abbb347cdebb504f8b46b709584b0a330802314605e95ba701428174af.jpg", + "image_caption": [ + "(d) Stage Four", + "Figure 5. Cosine similarity between geometry pools. (a) shows the similarity between geometry pools of S3DIS [1] and ScanNet [9]. (b) shows the similarity between geometry pools of S3DIS [1] and SemanticKITTI [2]. Intra-sensor geometry pools in (a) have higher similarity than inter-sensor geometry pools in (b)." + ], + "image_footnote": [], + "bbox": [ + 281, + 405, + 465, + 511 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ers possess larger receptive field sizes due to grouping and aggregating operations and comprise more geometric information, our hierarchical pools contain more geometric features for deeper layers.", + "bbox": [ + 75, + 589, + 468, + 650 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Representativeness of Geometry Pools. To demonstrate the representativeness of our pools, we calculate cosine similarity between all features in a scene from S3DIS [1] extracted by the point encoder and the features stored in the pool. For each stage, given the geometry pool $\\mathcal{F}^D = \\{f_i^D, 1 \\leq i \\leq N\\}$ and all features in a scene $\\mathcal{F}^S = \\{f_i^S, 1 \\leq i \\leq M\\}$ , the similarity $s_i$ between $f_i^S$ and $\\mathcal{F}^D$ is formulated as:", + "bbox": [ + 76, + 667, + 468, + 787 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\ns _ {i} = \\max _ {1 \\leq j \\leq N} \\frac {f _ {i} ^ {S} \\cdot f _ {j} ^ {D}}{\\| f _ {i} ^ {S} \\| \\| f _ {j} ^ {D} \\|}, \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 792, + 468, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\|\\cdot\\|$ stands for L2 norm. The distribution of similarities is shown in Figure 4. From the statistical data, most features have cosine similarity higher than 0.85, which indicates that each feature can find a similar feature among our", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2c8e1ebf082ed4023da4b1ab056c0d195f1e69dcc78a63b020cc2bc34f0e79c5.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 509, + 281, + 674, + 405 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d73edf02bb3173414dad3b59e9c165baeb5a9dcb25623c2dabc9cc16f815bdcd.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 687, + 280, + 852, + 406 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/73fa5de8b4b2cae4d4f8f2c282e8019f431196bdb9c0e56579430f574b99f75f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 861, + 280, + 888, + 401 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "learned representatives in the pool. In this manner, we avoid using point-based network to process large scale points to generate fine-grained features for each voxel. Our hierarchical pools contain typical geometric features of scenes. Therefore, we can fuse point-level geometry information stored in our pools into voxel representations via geometry-to-voxel auxiliary mechanism efficiently.", + "bbox": [ + 498, + 518, + 892, + 626 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Sensor-aware Pools for Multiple Sensors. Motivated by the diversity of density and sampling patterns of point clouds from various sensors, we construct different geometry pools to preserve relevant geometric information for diverse sensors. We investigate another two strategies: (i) shared geometry pools across sensors and (ii) different geometry pools for different datasets. We calculate the cosine similarity between features from different pools. As shown in Figure 5, geometry pools from the same sensor possess higher similarity than those from different sensors. As shown in Table 2, training with shared geometry pools across sensors leads to a significant decrease on performance due to the inherent diversity of geometric structures from different sensors, while sharing geometry pools with datasets from the same sensor serves as data augmentation to improve the effectiveness of the geometry pools.", + "bbox": [ + 496, + 659, + 892, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20025", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/941e96d74f6f5887dde8c7fa815ce197cebee9ca1b77a6565cc595441c633ea2.jpg", + "table_caption": [ + "Table 2. Ablation study of sensor-aware geometry pools. \"Shared intra-sensor\" means the geometry pools which are shared for point clouds from the same sensor, while different for point clouds from different sensors. \"Shared inter-sensor\" means the geometry pools shared for different sensors. \"Shared intra-dataset\" means the geometry pools are different for different datasets. We report mIoU and mAcc on Area 5 of S3DIS [1]." + ], + "table_footnote": [], + "table_body": "
MethodsmIoU (%)mAcc (%)
Shared intra-sensor69.574.5
Shared inter-sensor61.4(−8.1)68.3(−6.2)
Shared intra-dataset67.5(−2.0)73.3(−1.2)
", + "bbox": [ + 107, + 196, + 436, + 275 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6d568c54cba8d6b0b759f7120bbca820e5e74b672d826d88588323cd3873f624.jpg", + "table_caption": [ + "Table 3. Efficiency analysis. We report the inference time and throughput on pre-processed data from S3DIS dataset." + ], + "table_footnote": [], + "table_body": "
MethodParams.Inference Time (ms)Throughput (ins./sec.)
PointNet++ [28]1.0M378.7108
PT [54]7.8M1079.734
PTv2 [42]3.9M4732.212
MinkowskiNet [8]60.9M45.4898
SPVCNN [35]61.0M48.4682
PPT [43]63.0M80.4251
GeoAuxNet64.7M51.8572
", + "bbox": [ + 78, + 330, + 478, + 483 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Efficiency of GeoAuxNet", + "text_level": 1, + "bbox": [ + 76, + 494, + 302, + 511 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We validate the efficiency of GeoAuxNet compared with other typical point-based [42, 54] and voxel-based methods [8, 43]. We focus on the mean inference time and throughput by employing a standardized pre-processed point cloud with instance labels from S3DIS Area 5 dataset as the input on a single NVIDIA A40 GPU. Initially, the device undergoes a \"warm-up\" phase, where the model is fed with the input point cloud ten times consecutively to eliminate any latency associated with data loading. Subsequently, the model is subjected to an additional 300 iterations of input feeding to calculate mean inference time. To effectively quantify the throughput, we measure the number of instances that a model processes per second. This systematic approach ensures a thorough and reliable assessment of our GeoAuxNet and other state-of-the-art methods.", + "bbox": [ + 75, + 521, + 468, + 747 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Table 3, we show the number of parameters, inference time and throughput for typical point-based and voxel-based networks. Specifically, our GeoAuxNet outperforms the expert model of PointTransformer [54] which is trained on ScanNet [9] individually in mIoU with $16.8 \\times$ throughput and $20.8 \\times$ inference speedup. Even compared with the simple PointNet++ [28] whose parameter size is $1/68$ of ours, GeoAuxNet has less inference time and higher throughput. Besides, our method has similar inference time and throughput to voxel-based methods but achieve", + "bbox": [ + 75, + 750, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/78244ea50d925776369570c9c7d3a3a42104e1f6a965588568caa149d7292417.jpg", + "table_caption": [ + "Table 4. Ablation study of our voxel-guided hypernetwork. We report the mIoU and mAcc on Area 5 of S3DIS [1]." + ], + "table_footnote": [], + "table_body": "
MethodsmIoU (%)mAcc (%)
Vanilla63.368.7
+ voxel-guided66.0(+2.7)72.5(+3.8)
+ relative positions66.8(+0.8)73.2(+0.7)
+ stage latent code67.6(+0.8)73.3(+0.1)
", + "bbox": [ + 529, + 128, + 856, + 220 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "better performance according to the results in Table 1. GeoAuxNet surpasses PPT [43] on each benchmark about $2\\%$ in mIoU with $2.3\\times$ throughput and $1.6\\times$ inference speedup. Our method maintains efficiency for large scale point clouds, while achieves encouraging performances on datasets from multiple sensors.", + "bbox": [ + 496, + 229, + 893, + 320 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Voxel-guided Hypernetworks", + "text_level": 1, + "bbox": [ + 498, + 330, + 759, + 347 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To investigate the architecture design of our voxel-guided hypernetwork, we conduct ablation studies on S3DIS semantic segmentation. We first introduce a vanilla point network which learns the weight and bias directly. Then, we adopt a hypernetwork taking voxel features as input to generate weights and biases for the point network. We further take account of the relative positions of points. Finally, we optimize learnable stage latent codes to instruct the information of different layers. We train these four models on S3DIS dataset. As shown in Table 4, the exploration of voxel features provides significant guidance for the point network to learn high quality geometric information, leading to an increase in test accuracy by $2.7\\%$ in mIoU and $3.8\\%$ in mAcc. Relative positions and stage latent codes also improve the performance of our methods.", + "bbox": [ + 496, + 354, + 890, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 498, + 595, + 617, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose GeoAuxNet for multi-sensor point clouds with designed geometry-to-voxel auxiliary learning. We construct hierarchical geometry pools to learn auxiliary sensor-aware point-level geometric priors at different layers for sensor-agnostic voxel features. To generate our geometry pools, we also introduce a voxel-guided dynamic point network to leverage voxel prior knowledge for elaborate point features extraction. Our proposed geometry-to-voxel auxiliary builds a bridge between point-level and voxel-level features in an efficient manner without using the point network during inference time. Experimental results have shown the effectiveness and efficiency of our methods. We hope this work will inspire future research towards universal representation learning for point clouds.", + "bbox": [ + 496, + 622, + 890, + 834 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported by the National Natural Science Foundation of China under Grant 62206147.", + "bbox": [ + 496, + 854, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20026", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Iro Armeni, Ozan Sener, Amir Roshan Zamir, Helen Jiang, Ioannis K. Brilakis, Martin Fischer, and Silvio Savarese. 3D Semantic Parsing of Large-Scale Indoor Spaces. CVPR, pages 1534-1543, 2016. 1, 2, 5, 6, 7, 8", + "[2] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, C. Stachniss, and Juergen Gall. SemanticKITTI: A Dataset for Semantic Scene Understanding of LiDAR Sequences. ICCV, pages 9296-9306, 2019. 1, 2, 5, 6, 7", + "[3] Alexandre Boulch. ConvPoint: Continuous convolutions for point cloud processing. CGF, 88:24-34, 2019. 2", + "[4] César Cadena, Anthony R. Dick, and Ian D. Reid. Multimodal Auto-Encoders as Joint Estimators for Robotics Scene Understanding. In RSS, 2016. 1", + "[5] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yuxin Pan, Giancarlo Baldan, and Oscar Beijbom. nuScenes: A Multimodal Dataset for Autonomous Driving. CVPR, pages 11618-11628, 2019. 1", + "[6] Hao Chen, Xiaojuan Qi, Lequan Yu, and Pheng-Ann Heng. DCAN: Deep Contour-Aware Networks for Accurate Gland Segmentation. CVPR, pages 2487-2496, 2016. 3", + "[7] Jaesung Choe, Chunghyun Park, Francois Rameau, Jaesik Park, and In So Kweon. PointMixer: MLP-Mixer for Point Cloud Understanding. ECCV, 2022. 2", + "[8] Christopher Bongsoo Choy, JunYoung Gwak, and Silvio Savarese. 4D Spatio-Temporal ConvNets: Minkowski Convolutional Neural Networks. CVPR, pages 3070-3079, 2019. 2, 5, 6, 8", + "[9] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas A. Funkhouser, and Matthias Nießner. ScanNet: Richly-Annotated 3D Reconstructions of Indoor Scenes. CVPR, pages 2432-2443, 2017. 1, 2, 5, 6, 7, 8", + "[10] Jesse Dodge, Ana Marasovic, Gabriel Ilharco, Dirk Groeneweld, Margaret Mitchell, and Matt Gardner. Documenting Large Webtext Corpora: A Case Study on the Colossal Clean Crawled Corpus. In EMNLP, 2021. 2", + "[11] Nico Engel, Vasileios Belagiannis, and Klaus C. J. Dietmayer. Point Transformer. IEEE, 9:134826-134840, 2020. 2", + "[12] Benjamin Graham, Martin Engelcke, and Laurens van der Maaten. 3D Semantic Segmentation with Submanifold Sparse Convolutional Networks. CVPR, pages 9224-9232, 2018. 2", + "[13] Lei Han, Tian Zheng, Lan Xu, and Lu Fang. OccuSeg: Occupancy-Aware 3D Instance Segmentation. CVPR, pages 2937-2946, 2020. 2", + "[14] Judy Hoffman, Eric Tzeng, Trevor Darrell, and Kate Saenko. Simultaneous Deep Transfer Across Domains and Tasks. ICCV, pages 4068-4076, 2015. 3", + "[15] Lingdong Kong, You-Chen Liu, Runnan Chen, Yuexin Ma, Xinge Zhu, Yikang Li, Yuenan Hou, Y. Qiao, and Ziwei Liu. Rethinking Range View Representation for LiDAR Segmentation. ArXiv, abs/2303.05367, 2023. 2", + "[16] Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, and Jiaya Jia. Stratified Trans-" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "former for 3D Point Cloud Segmentation. CVPR, pages 8490-8499, 2022. 2, 6", + "[17] Xin Lai, Yukang Chen, Fanbin Lu, Jianhui Liu, and Jiaya Jia. Spherical Transformer for LiDAR-Based 3D Recognition. CVPR, pages 17545-17555, 2023. 2, 6", + "[18] John Lambert, Zhuang Liu, Ozan Sener, James Hays, and Vladlen Koltun. MSeg: A Composite Dataset for Multi-Domain Semantic Segmentation. PAMI, 45:796-810, 2021. 2", + "[19] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. PointCNN: Convolution On X-Transformed Points. In NIPS, 2018. 2", + "[20] Yingwei Li, Adams Wei Yu, Tianjian Meng, Benjamin Caine, Jiquan Ngiam, Daiyi Peng, Junyang Shen, Bo-Xun Wu, Yifeng Lu, Denny Zhou, Quoc V. Le, Alan Loddon Yuille, and Mingxing Tan. DeepFusion: Lidar-Camera Deep Fusion for Multi-Modal 3D Object Detection. CVPR, pages 17161-17170, 2022. 1", + "[21] Xingyu Lin, Harjatin Singh Baweja, George A. Kantor, and David Held. Adaptive Auxiliary Task Weighting for Reinforcement Learning. In NIPS, 2019. 3", + "[22] Zhi-Hao Lin, Sheng Yu Huang, and Y. Wang. Convolution in the Cloud: Learning Deformable Kernels in 3D Graph Convolution Networks for Point Cloud Analysis. CVPR, pages 1797-1806, 2020. 2", + "[23] Xianpeng Liu, Nan Xue, and Tianfu Wu. Learning Auxiliary Monocular Contexts Helps Monocular 3D Object Detection. In AAAI, 2021. 3", + "[24] Zhijian Liu, Haotian Tang, Yujun Lin, and Song Han. PointVoxel CNN for Efficient 3D Deep Learning. In NIPS, 2019. 2, 5", + "[25] Daniel Maturana and Sebastian A. Scherer. VoxNet: A 3D Convolutional Neural Network for real-time object recognition. IROS, pages 922-928, 2015. 2", + "[26] Andres Milioto, Ignacio Vizzo, Jens Behley, and C. Stachniss. RangeNet ++: Fast and Accurate LiDAR Semantic Segmentation. IROS, pages 4213-4220, 2019. 2", + "[27] C. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. Point-Net: Deep Learning on Point Sets for 3D Classification and Segmentation. CVPR, pages 77–85, 2017. 2", + "[28] C. Qi, L. Yi, Hao Su, and Leonidas J. Guibas. PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space. In NIPS, 2017. 2, 8", + "[29] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. PointNeXt: Revisiting PointNet++ with Improved Training and Scaling Strategies. In NIPS, 2022. 2, 6", + "[30] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards Robust Monocular Depth Estimation: Mixing Datasets for Zero-Shot Cross-Dataset Transfer. PAMI, 44:1623-1637, 2019. 2", + "[31] Gernot Riegler, Ali O. Ulusoy, and Andreas Geiger. Oct-Net: Learning Deep 3D Representations at High Resolutions. CVPR, pages 6620-6629, 2017. 2", + "[32] Leslie N. Smith and Nicholay Topin. Super-Convergence: Very Fast Training of Residual Networks Using Large Learning Rates. ArXiv, abs/1708.07120, 2017. 5" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20027", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Shuran Song, Samuel P. Lichtenberg, and Jianxiong Xiao. SUN RGB-D: A RGB-D scene understanding benchmark suite. CVPR, pages 567-576, 2015. 1", + "[34] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott M. Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in Perception for Autonomous Driving: Waymo Open Dataset. CVPR, pages 2443-2451, 2019. 1", + "[35] Haotian* Tang, Zhijian* Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching Efficient 3D Architectures with Sparse Point-Voxel Convolution. In ECCV, 2020. 2, 5, 6, 8", + "[36] Hugues Thomas, C. Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J. Guibas. KPConv: Flexible and Deformable Convolution for Point Clouds. ICCV, pages 6410–6419, 2019. 2", + "[37] Li Wang, Dong Wei Li, Han Liu, Jinzhan Peng, Lu Tian, and Yi Shan. Cross-Dataset Collaborative Learning for Semantic Segmentation. AAAI, 2021. 2, 3", + "[38] Xudong Wang, Zhaowei Cai, Dashan Gao, and Nuno Vasconcelos. Towards Universal Object Detection by Domain Attention. CVPR, pages 7281-7290, 2019. 2", + "[39] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E. Sarma, Michael M. Bronstein, and Justin M. Solomon. Dynamic Graph CNN for Learning on Point Clouds. TOG, 38:1-12, 2018. 2", + "[40] Christian Wojek, Stefan Walk, Stefan Roth, and Bernt Schiele. Monocular 3D scene understanding with explicit occlusion reasoning. CVPR, pages 1993-2000, 2011. 1", + "[41] Bichen Wu, Alvin Wan, Xiangyu Yue, and Kurt Keutzer. SqueezeSeg: Convolutional Neural Nets with Recurrent CRF for Real-Time Road-Object Segmentation from 3D Li-DAR Point Cloud. ICRA, pages 1887–1893, 2018. 2", + "[42] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point Transformer V2: Grouped Vector Attention and Partition-based Pooling. In NIPS, 2022. 2, 8", + "[43] Xiaoyang Wu, Zhuotao Tian, Xin Wen, Bohao Peng, Xihui Liu, Kaicheng Yu, and Hengshuang Zhao. Towards Large-scale 3D Representation Learning with Multi-dataset Point Prompt Training. ArXiv, abs/2308.09718, 2023. 2, 3, 5, 6, 8", + "[44] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3D ShapeNets: A deep representation for volumetric shapes. CVPR, pages 1912-1920, 2015. 6", + "[45] Yichen Xie, Chenfeng Xu, Marie-Julie Rakotosaona, Patrick Rim, Federico Tombari, Kurt Keutzer, Masayoshi Tomizuka, and Wei Zhan. SparseFusion: Fusing Multi-Modal Sparse Representations for Multi-Sensor 3D Object Detection. ICCV, abs/2304.14340, 2023. 1", + "[46] Guodong Xu, Ziwei Liu, Xiaoxiao Li, and Chen Change Loy. Knowledge Distillation Meets Self-Supervision. In ECCV, 2020. 3", + "[47] Hang Xu, Linpu Fang, Xiaodan Liang, Wenxiong Kang, and Zhenguo Li. Universal-RCNN: Universal Object Detector via Transferable Graph R-CNN. In AAAI, 2020. 2" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[48] Mutian Xu, Runyu Ding, Hengshuang Zhao, and Xiaojuan Qi. PACnv: Position Adaptive Convolution with Dynamic Kernel Assembling on Point Clouds. CVPR, pages 3172-3181, 2021. 2", + "[49] Yu-Qi Yang, Yu-Xiao Guo, Jiangfeng Xiong, Yang Liu, Hao Pan, Peng-Shuai Wang, Xin Tong, and Baining Guo. Swin3D: A Pretrained Transformer Backbone for 3D Indoor Scene Understanding. ArXiv, abs/2304.06906, 2023. 2", + "[50] Zetong Yang, Yanan Sun, Shu Liu, Xiaoyong Shen, and Jiaya Jia. STD: Sparse-to-Dense 3D Object Detector for Point Cloud. ICCV, pages 1951-1960, 2019. 2", + "[51] Tianwei Yin, Xingyi Zhou, and Philipp Kähenbuhl. Center-based 3D Object Detection and Tracking. CVPR, pages 11779-11788, 2021. 1", + "[52] Cheng Zhang, Haocheng Wan, Xinyi Shen, and Zizhao Wu. PVT: Point-voxel transformer for point cloud learning. *IJIS*, 37:11985 - 12008, 2021. 2, 5", + "[53] Yang Zhang, Zixiang Zhou, Philip David, Xiangyu Yue, Zerong Xi, and Hassan Foroosh. PolarNet: An Improved Grid Representation for Online LiDAR Point Clouds Semantic Segmentation. CVPR, pages 9598-9607, 2020. 2", + "[54] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip Torr, and Vladlen Koltun. Point Transformer. In ICCV, 2021. 2, 6, 8", + "[55] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Simple Multi-dataset Detection. CVPR, pages 7561-7570, 2021. 2", + "[56] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and Asymmetrical 3D Convolution Networks for LiDAR Segmentation. CVPR, pages 9934-9943, 2021. 2, 6", + "[57] Özgün Çiçek, Ahmed Abdulkadir, Soeren S. Lienkamp, Thomas Brox, and Olaf Ronneberger. 3D U-Net: Learning Dense Volumetric Segmentation from Sparse Annotation. In MICCAI, 2016. 2" + ], + "bbox": [ + 501, + 90, + 890, + 585 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20028", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/GeoAuxNet_ Towards Universal 3D Representation Learning for Multi-sensor Point Clouds/f223747a-d407-4fbb-9660-d95973ceefec_model.json b/2024/GeoAuxNet_ Towards Universal 3D Representation Learning for Multi-sensor Point Clouds/f223747a-d407-4fbb-9660-d95973ceefec_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3717a84a550005e1721f32f5e8d0830c2ee1de45 --- /dev/null +++ b/2024/GeoAuxNet_ Towards Universal 3D Representation Learning for Multi-sensor Point Clouds/f223747a-d407-4fbb-9660-d95973ceefec_model.json @@ -0,0 +1,2552 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.095, + 0.13, + 0.877, + 0.175 + ], + "angle": 0, + "content": "GeoAuxNet: Towards Universal 3D Representation Learning for Multi-sensor Point Clouds" + }, + { + "type": "text", + "bbox": [ + 0.318, + 0.203, + 0.653, + 0.222 + ], + "angle": 0, + "content": "Shengjun Zhang\\(^{1}\\), Xin Fei\\(^{2}\\), Yueqi Duan\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.244, + 0.222, + 0.728, + 0.24 + ], + "angle": 0, + "content": "1 Department of Electronic Engineering, Tsinghua University" + }, + { + "type": "text", + "bbox": [ + 0.286, + 0.239, + 0.684, + 0.258 + ], + "angle": 0, + "content": "\\(^{2}\\) Department of Automation, Tsinghua University" + }, + { + "type": "text", + "bbox": [ + 0.184, + 0.26, + 0.782, + 0.275 + ], + "angle": 0, + "content": "{zhangsj23, feix21}@mails.tsinghua.edu.cn, duanyueqi@tsinghua.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.257, + 0.283, + 0.713, + 0.3 + ], + "angle": 0, + "content": "https://github.com/zhangshengjun2019/GeoAuxNet" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.342, + 0.146, + 0.368 + ], + "angle": 0, + "content": "S3DIS RGB-D" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.412, + 0.153, + 0.438 + ], + "angle": 0, + "content": "ScanNet RGB-D" + }, + { + "type": "image", + "bbox": [ + 0.155, + 0.307, + 0.276, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.155, + 0.389, + 0.276, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.307, + 0.399, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.389, + 0.398, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.307, + 0.521, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.389, + 0.52, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.307, + 0.643, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.389, + 0.643, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.307, + 0.765, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.389, + 0.765, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.766, + 0.307, + 0.888, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.766, + 0.389, + 0.889, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.482, + 0.155, + 0.523 + ], + "angle": 0, + "content": "Semantic KITTI LiDAR" + }, + { + "type": "image", + "bbox": [ + 0.155, + 0.46, + 0.276, + 0.541 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.199, + 0.543, + 0.237, + 0.558 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.46, + 0.398, + 0.541 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.284, + 0.543, + 0.387, + 0.556 + ], + "angle": 0, + "content": "MinkowskiNet" + }, + { + "type": "image", + "bbox": [ + 0.398, + 0.46, + 0.52, + 0.541 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.425, + 0.543, + 0.494, + 0.556 + ], + "angle": 0, + "content": "SPVCNN" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.46, + 0.643, + 0.541 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.566, + 0.543, + 0.597, + 0.556 + ], + "angle": 0, + "content": "PPT" + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.46, + 0.765, + 0.541 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.686, + 0.543, + 0.721, + 0.556 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.766, + 0.46, + 0.889, + 0.541 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.781, + 0.543, + 0.871, + 0.556 + ], + "angle": 0, + "content": "GroundTruth" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.57, + 0.893, + 0.599 + ], + "angle": 0, + "content": "Figure 1. Semantic segmentation results on S3DIS [1] and ScanNet [9] from RGB-D cameras and SemanticKITTI [2] from LiDAR. For all methods, we trained collectively on three datasets. Our method outperforms other methods with better detailed structures." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.619, + 0.314, + 0.635 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.651, + 0.471, + 0.878 + ], + "angle": 0, + "content": "Point clouds captured by different sensors such as RGB-D cameras and LiDAR possess non-negligible domain gaps. Most existing methods design different network architectures and train separately on point clouds from various sensors. Typically, point-based methods achieve outstanding performances on even-distributed dense point clouds from RGB-D cameras, while voxel-based methods are more efficient for large-range sparse LiDAR point clouds. In this paper, we propose geometry-to-voxel auxiliary learning to enable voxel representations to access point-level geometric information, which supports better generalisation of the voxel-based backbone with additional interpretations of multi-sensor point clouds. Specifically, we construct hierarchical geometry pools generated by a voxel-guided dynamic point network, which efficiently provide" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.621, + 0.892, + 0.727 + ], + "angle": 0, + "content": "auxiliary fine-grained geometric information adapted to different stages of voxel features. We conduct experiments on joint multi-sensor datasets to demonstrate the effectiveness of GeoAuxNet. Enjoying elaborate geometric information, our method outperforms other models collectively trained on multi-sensor datasets, and achieve competitive results with the-state-of-art experts on each single dataset." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.756, + 0.633, + 0.771 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Point cloud analysis has attracted widespread attention due to its growing applications, such as autonomous driving [20, 45, 51] and robotics [4, 40]. Unlike images which are represented by regular pixels, 3D point clouds are irregular and unordered. These characteristics aggravate the inconsistency on density and sampling patterns of point clouds captured by different sensors, such as RGB-D cameras [1, 9, 33] and LiDAR [2, 5, 34]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.102, + 0.888, + 0.228, + 0.902 + ], + "angle": 0, + "content": "†Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "20019" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.081, + 0.092, + 0.468, + 0.438 + ], + "angle": 0, + "content": "Typically, points generated from RGB-D pictures are equally distributed and dense, while points scanned by LiDAR are sparse and uneven. The diversity on input data hinders the construction of universal network architectures. For point clouds from RGB-D cameras, point-based methods [28, 29, 54] usually gather local information via grouping and aggregating neighborhood features to extract detailed spatial information. Since neighbors are not stored contiguously in point representations, indexing them requires the costly searching operations [24], which limits the application of these methods on large scale LiDAR point clouds. To address this challenge, voxel-based methods [12, 25, 31] leverage the regular 3D convolution to obtain the contiguous memory access pattern. However, the resolution of voxels is constrained by the memory and becomes lower at deeper stages, resulting in loss of local information [35]. An intuitive idea is to ensemble both point-based and voxel-based methods [24, 35, 52], yet they are still confronted with either high time-consumption or the lack of detailed geometric features, which fail to simultaneously process large scale point clouds from different sensors with fine-grained spatial information while maintaining efficiency." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.444, + 0.468, + 0.759 + ], + "angle": 0, + "content": "In this paper, we propose GeoAuxNet to provide point-level geometric information for voxel representations in an efficient manner with geometry-to-voxel auxiliary learning. The support of elaborate geometric features introduces additional point-level information, which cannot be fully exploited by voxel-based backbones. Motivated by the observation that the local geometric structures present high similarity at each stage of the network, we construct hierarchical geometry pools to learn auxiliary sensor-aware point-level geometric priors corresponding to different layers for sensor-agnostic voxel features. To update our geometry pools, we present a voxel-guided dynamic point network, where we leverage prior knowledge in voxel features to guide high quality spatial feature extraction. Then, we fuse elaborate geometric features in the pools into voxel representations via our designed geometry-to-voxel auxiliary mechanism. For large-scale point clouds, the hierarchical geometry pools store representative point-level features at different stages, which efficiently provide complementary geometric information without using the point-based networks during inference time." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.765, + 0.468, + 0.9 + ], + "angle": 0, + "content": "We conduct extensive experiments on multi-sensor datasets, including S3DIS [1] and ScanNet [9] from RGB-D cameras as well as SemanticKITTI [2] captured by LiDAR, to demonstrate the efficiency and effectiveness of our method. With a shared weight of backbone, our method outperforms other models trained on joint datasets from different sensors and achieves competitive performance with experts on single dataset. As shown in Figure 1, we trained MinkowskiNet [8], SPVCNN [35], PPT [43] and" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.093, + 0.891, + 0.135 + ], + "angle": 0, + "content": "GeoAuxNet on the above three datasets collectively, where our method preserve better detailed structures for point clouds from various sensors." + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.149, + 0.647, + 0.164 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.174, + 0.891, + 0.671 + ], + "angle": 0, + "content": "3D Scene Understanding. Since images are composed of regular pixels, point clouds generated from RGB-D cameras are well-distributed and dense. Voxelization on dense point clouds ignore detailed structure information. To directly learn from points, Qi et al. [27] first proposed PointNet to process point clouds through shared multi-layer perceptions (MLP) to get a global representation. Yet this method lacks the ability to comprehend local geometric features. Therefore, PointNet++ [28] built a hierarchical architecture to gather local information via grouping and aggregating neighborhood features. Following up MLP-based works [7, 29] focused on designing effective learning schemes to capture spatial features. Other methods also introduced specific convolution operations [3, 19, 36, 48], graph-based representations [22, 39] and self-attention mechanism [11, 16, 42, 49, 54] to explore the relationship between points. Apart from cameras, LiDAR is also a widely used sensor to collect point clouds. Some methods [15, 26, 41, 53] focused on projecting point clouds to 2D grids to utilize 2D CNN. However, 3D to 2D projection limits the representation quality of geometric structures [50] due to the sacrifice of 3D spatial information [48]. Point based methods have high time consumption because of the large scale and the sparsity of these point clouds. Voxel-based methods [12, 13, 17, 56, 57] show more robustness against the sparsity. In view of the sampling characteristic of LiDAR, Zhu et al. [56] introduced cylinder coordinate system, for nearby regions have greater density than farther away regions. Similarly, Lai et al. [17] proposed spherical transformer to aggregate information from dense close points to sparse distant ones. Voxel-based networks maintain efficiency when processing scene-level point clouds, but ignore some detailed spatial structures." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.69, + 0.891, + 0.9 + ], + "angle": 0, + "content": "Universal Learning from Multiple Data Sources. Large language models achieve remarkable progress on many natural language processing systems through pretraining on extremely large text corpora [10], which enables foundation models to process data across domains, involving multiple languages, various majors, diverse application scenarios and so on. In 2D computer vision, joint learning across domains improve model robustness for detection [38, 47, 55], depth estimation [30] and semantic segmentation [18, 37]. Wang et al. [38] proposed a universal object detector through a domain-attention mechanism. But they did not model the training differences between different datasets. Universal-RCNN [47] modeled the class relations by designing a inter-dataset attention module to in" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.947, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20020" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.088, + 0.891, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.348, + 0.896, + 0.473 + ], + "angle": 0, + "content": "Figure 2. The pipeline of our GeoAuxNet. For a complete scene point cloud \\(\\mathcal{P}^{\\mathcal{C}}\\) and a local point patch \\(\\mathcal{P} \\subseteq \\mathcal{P}^{\\mathcal{C}}\\), our voxel-based backbone first voxelizes \\(\\mathcal{P}^{\\mathcal{C}}\\) and conducts sparse convolutional operations. The voxel-guided hypernetwork takes relative positions, voxel features and a stage latent code as input to provide weights and biases for the point network. Then, we encode the spatial information for \\(\\mathcal{P}\\) with the point network and aggregate local features to generate geometric feature candidates. Following the update strategy, we construct hierarchical geometry pools. The geometry-to-voxel mechanism fuses geometric features stored in the pools to enable voxel representations to access point-level geometric information. We repeat the above process several times to extract effective representation hierarchically and predict the results with a voxel decoder for the primary task and a point decoder for the auxiliary task. The dotted line stands for the course of the auxiliary learning which is ignored during inference to ensure efficiency. Geo-to-Vox is abbreviation of Geometry-to-Voxel." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.483, + 0.473, + 0.664 + ], + "angle": 0, + "content": "corporates graph transfer learning for propagating relevant semantic information across multiple datasets. CDCL [37] pointed out that convolution filters can be shared across domains without accuracy loss, while normalization operators are not appropriate to share in view of the bias of statistics. Recently, Wu et al. [43] introduced point prompt training to optimize a learnable domain-specific prompt with language-guided categorical alignment, but processed point clouds from RGB-D and LiDAR separately. However, point clouds captured by different sensors have diverse density and distributions, which limits the exploration of the universal network on multiple data sources." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Auxiliary Learning. Auxiliary learning aims to achieve outstanding performance on a single primary task with the assist of auxiliary tasks and has shown many successful applications in a wide range of fields, including knowledge distillation [46], transfer learning [14] and reinforcement learning [21]. For instance, DCAN [6] learned an auxiliary contour detection task for more accurate segmentation, while Araslanov investigated the joint learning of image classification and semantic segmentation. In 3D vision, MonoCon [23] introduced monocular contexts as auxiliary tasks in training for 3D object detection. Nevertheless, the inherent relationship between point-level features and voxel-level features is still under exploration to better leverage the advantages of two domains without any requirement" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.483, + 0.625, + 0.497 + ], + "angle": 0, + "content": "of additional data." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.511, + 0.593, + 0.527 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.536, + 0.892, + 0.674 + ], + "angle": 0, + "content": "The overall framework of GeoAuxNet is illustrated in Figure 2. To generate point-level geometric features, we propose a voxel-guided hypernetwork to produce weights and biases for the point network. Then, point features are grouped and aggregated to extract fine-grained local features, which are employed to update geometry pools. For voxel-based backbone, we introduce geometry-to-voxel auxiliary mechanism to access elaborate spatial information in our hierarchical geometry pools." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.682, + 0.828, + 0.698 + ], + "angle": 0, + "content": "3.1. Voxel-guided Dynamic Point Network" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.824 + ], + "angle": 0, + "content": "Point-based methods are mostly confronted with high time consumption towards large scale point clouds, since the farthest point sampling (FPS) and k-nearest neighbors (kNN) algorithms in these networks have a time complexity of \\(\\mathcal{O}(n^2)\\). Therefore, to efficiently process large scale point clouds, input points are voxelized and passed through sparse convolution operations to extract voxel features in our voxel-based backbone." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.901 + ], + "angle": 0, + "content": "To generate point-level geometric information efficiently, we train our point network on local point patches instead of complete scene point clouds. However, less training data weakens the ability of the point network to extract local features. Considering that our voxel-based branch are" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "20021" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.242 + ], + "angle": 0, + "content": "trained on complete voxelized point clouds, we treat voxel features as prior knowledge to instruct the learning of local geometric information. Yet, points inside a voxel share the same voxel feature, which ignores detailed difference between points. Thus, we also introduce relative coordinates to depict diversity of points within the same voxel. On account of the hierarchical architectures of both the point network and the voxel-based backbone with several stages, we optimize a learnable latent code \\( s \\) for each stage during our auxiliary learning." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.242, + 0.47, + 0.319 + ], + "angle": 0, + "content": "Given a local point patch \\(\\mathcal{P} = \\{p_i, 1 \\leq i \\leq N\\} \\in \\mathbb{R}^{N \\times 3}\\) with point features \\(\\mathcal{F}^{\\mathcal{P}} = \\{f_i^{\\mathcal{P}}, 1 \\leq i \\leq N\\}\\), each point \\(p_i\\) belongs to a voxel \\(v\\) with the voxel feature \\(f^{\\mathcal{V}}\\). Here \\(v \\in \\mathbb{R}^3\\) stands for the center coordinate of the voxel. Then, we can obtain the weight and bias for point \\(p_i\\):" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.327, + 0.47, + 0.345 + ], + "angle": 0, + "content": "\\[\nw _ {i} = h _ {w} \\left(h _ {p} \\left(p _ {i} - v\\right), f ^ {\\mathcal {V}}\\right) \\odot h _ {s} (s), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.348, + 0.469, + 0.366 + ], + "angle": 0, + "content": "\\[\nb _ {i} = h _ {b} \\left(h _ {p} \\left(p _ {i} - v\\right), f ^ {\\mathcal {V}}\\right) \\odot h _ {s} (s), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.375, + 0.47, + 0.435 + ], + "angle": 0, + "content": "where \\( h_p \\) and \\( h_s \\) are MLPs to project relative positions and the stage latent code to specific dimensions, and \\( h_w \\) and \\( h_b \\) are hypernetworks to generate the weight and bias. Hence, new point features can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.179, + 0.443, + 0.469, + 0.462 + ], + "angle": 0, + "content": "\\[\n\\hat {f} _ {i} ^ {\\mathcal {P}} = \\operatorname {M L P} \\left(w _ {i} \\odot f _ {i} ^ {\\mathcal {P}} + b _ {i}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.471, + 0.469, + 0.547 + ], + "angle": 0, + "content": "To extract local geometric information, we use kNN algorithm to select \\(k\\) nearest neighbor points and employ a max-pooling function to aggregate each local group. In this manner, we generate a set of geometric feature candidates for hierarchical geometry pools." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.555, + 0.337, + 0.571 + ], + "angle": 0, + "content": "3.2. Hierarchical Geometry Pools" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.578, + 0.47, + 0.714 + ], + "angle": 0, + "content": "The voxel branch ignores some detailed spatial structures due to voxelization. Motivated by the similarity of geometric structures in 3D space, we establish a pool to store typical geometric patterns. The pool contains \\( n \\) typical spatial structure features \\( \\mathcal{F}^{\\mathcal{D}} = \\{f_i^{\\mathcal{D}}, 1 \\leq i \\leq n\\} \\). The max size of the pool is set to \\( N \\). In Section 3.1, we generate a set of geometric pattern candidates \\( \\mathcal{F}^S = \\{f_i^S, 1 \\leq i \\leq m\\} \\). When updating the pool, we first calculate the similarity between \\( \\mathcal{F}^S \\) and \\( \\mathcal{F}^D \\):" + }, + { + "type": "equation", + "bbox": [ + 0.216, + 0.726, + 0.469, + 0.748 + ], + "angle": 0, + "content": "\\[\ns _ {i} = \\max _ {1 \\leq j \\leq N} s _ {i, j}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.757, + 0.124, + 0.769 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.208, + 0.767, + 0.469, + 0.805 + ], + "angle": 0, + "content": "\\[\ns _ {i, j} = \\frac {f _ {i} ^ {\\mathcal {S}} \\cdot f _ {j} ^ {\\mathcal {D}}}{\\| f _ {i} ^ {\\mathcal {S}} \\| \\| f _ {j} ^ {\\mathcal {D}} \\|.} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.809, + 0.469, + 0.839 + ], + "angle": 0, + "content": "Specifically, if \\( s_i \\geq \\epsilon \\), \\( f_i^S \\) is merged into the geometry pool \\( \\mathcal{F}^{\\mathcal{D}} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.201, + 0.852, + 0.469, + 0.876 + ], + "angle": 0, + "content": "\\[\nt = \\underset {1 \\leq j \\leq N} {\\operatorname {a r g m a x}} s _ {i, j}, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.879, + 0.469, + 0.898 + ], + "angle": 0, + "content": "\\[\n\\hat {f} _ {t} ^ {\\mathcal {D}} = \\lambda f _ {t} ^ {\\mathcal {D}} + (1 - \\lambda) f _ {i} ^ {\\mathcal {S}}, \\tag {7}\n\\]" + }, + { + "type": "code_caption", + "bbox": [ + 0.501, + 0.091, + 0.682, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Update Pools" + }, + { + "type": "algorithm", + "bbox": [ + 0.5, + 0.111, + 0.892, + 0.446 + ], + "angle": 0, + "content": "Input: New features \\(\\mathcal{F}^S\\), geometry pool \\(\\mathcal{F}^D\\), maximum size of the pool \\(N\\), threshold \\(\\epsilon\\), update rate \\(\\lambda\\) \nOutput: New pool \\(\\hat{\\mathcal{F}}^D\\) \nCompute \\(s_i\\) for \\(1 \\leq i \\leq m\\) according to Eq. 4 \nfor \\(i \\gets 1\\) to \\(|\\mathcal{F}^S|\\) do \n if \\(s_i \\geq \\epsilon\\) then \n t ← arg max\\(_{1 \\leq j \\leq N}\\) \\(s_i, j\\) \n \\(f_t^D \\gets \\lambda f_t^D + (1 - \\lambda)f_i^S\\) \n \\(\\mathcal{F}^S \\gets \\mathcal{F}^S \\setminus \\{f_i^S\\}\\) \n end if \nend for \nif \\(|\\mathcal{F}^S| + |\\mathcal{F}^D| \\leq N\\) then \n \\(\\mathcal{F}^D \\gets \\mathcal{F}^D \\cup \\mathcal{F}^S\\) \nelse \n \\(\\hat{\\mathcal{F}}^S \\gets\\) randomly select \\(N - |\\mathcal{F}^D|\\) features from \\(\\mathcal{F}^S\\) \n \\(\\mathcal{F}^S \\gets \\mathcal{F}^S \\setminus \\hat{\\mathcal{F}}^S\\) \n \\(\\mathcal{F}^D \\gets \\mathcal{F}^D \\cup \\hat{\\mathcal{F}}^S\\) \n for \\(i \\gets 1\\) to \\(|\\mathcal{F}^S|\\) do \n t ← arg max\\(_{1 \\leq j \\leq N}\\) \\(s_i, j\\) \n \\(f_t^D \\gets \\lambda f_t^D + (1 - \\lambda)f_i^S\\) \n end for \nend if" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.476, + 0.892, + 0.567 + ], + "angle": 0, + "content": "where \\(\\epsilon\\) is a threshold and \\(\\lambda\\) stands for the update rate. Otherwise, we append the new features to the geometry pool. If the total number of features is more than \\(N\\), we randomly select \\(N - n\\) new features and add them to the pool. The rest new features following Eq. 7 are merged into existing patterns. Our update strategy is illustrated in Algorithm 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.568, + 0.892, + 0.659 + ], + "angle": 0, + "content": "Since voxel representations at various stages contains different level information, we introduce hierarchical geometry pools to store representative point-level features corresponding to different stages. We also enlarge the size of pools for deeper stages due to more complex geometric structures with larger receptive field sizes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.673, + 0.758, + 0.689 + ], + "angle": 0, + "content": "3.3. Geometry-to-Voxel Auxiliary" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.698, + 0.892, + 0.789 + ], + "angle": 0, + "content": "After we update the geometric pool, we use geometric patterns stored in \\(\\mathcal{F}^{\\mathcal{D}}\\in \\mathbb{R}^{N\\times C}\\) to generate fine-grained local information for each voxel following a cross-attention operation. Given voxel features \\(\\mathcal{F}^{\\nu} = \\{f_i^\\nu ,1\\leq i\\leq M\\} \\in\\) \\(\\mathbb{R}^{M\\times D}\\), the geometric fine-grained feature \\(f_{i}^{\\mathcal{G}}\\) for \\(f_{i}^{\\nu}\\in \\mathcal{F}^{\\nu}\\) is formulated by:" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.802, + 0.891, + 0.82 + ], + "angle": 0, + "content": "\\[\nq _ {i} = h _ {Q} \\left(f _ {i} ^ {\\mathcal {V}}\\right) \\in \\mathbb {R} ^ {1 \\times s}, \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.822, + 0.891, + 0.841 + ], + "angle": 0, + "content": "\\[\nK = h _ {K} \\left(\\mathcal {F} ^ {\\mathcal {D}}\\right) \\in \\mathbb {R} ^ {N \\times s}, \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.843, + 0.891, + 0.861 + ], + "angle": 0, + "content": "\\[\nV = h _ {V} \\left(\\mathcal {F} ^ {\\mathcal {D}}\\right) \\in \\mathbb {R} ^ {N \\times t}, \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.864, + 0.891, + 0.899 + ], + "angle": 0, + "content": "\\[\nf _ {i} ^ {\\mathcal {G}} = \\operatorname {s o f t m a x} \\left(\\frac {q _ {i} K ^ {\\top}}{\\sqrt {s}}\\right) V \\in \\mathbb {R} ^ {1 \\times t}, \\tag {11}\n\\]" + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20022" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.168 + ], + "angle": 0, + "content": "where \\( h_{Q}, h_{K} \\), and \\( h_{V} \\) stand for projections. Then, we concatenate \\( f_{i}^{\\mathcal{G}} \\) with \\( f_{i}^{\\mathcal{V}} \\) as the new voxel feature and employ sparse convolution operations. In this manner, we can provide complementary geometric information without using the point-based networks during inference stage." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.181, + 0.274, + 0.195 + ], + "angle": 0, + "content": "3.4. Overall Architecture" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.205, + 0.469, + 0.326 + ], + "angle": 0, + "content": "Given a point cloud \\(\\mathcal{P}_i^{\\mathcal{C}}\\) with features \\(\\mathcal{F}_i^{\\mathcal{C}}\\) and a local point patch \\(\\mathcal{P}_i\\subseteq \\mathcal{P}_i^{\\mathcal{C}}\\) with \\(\\mathcal{F}_i^{\\mathcal{P}}\\subseteq \\mathcal{F}_i^{\\mathcal{C}}\\) from sensor \\(S_{i}\\), we voxelize all points to obtain voxel coordinates \\(\\nu_{i}\\) with voxel features \\(\\mathcal{F}_i^\\nu\\). Points obtained from RGB-D cameras usually possess colors and normal vectors, while points captured by LiDAR always possess intensity. Thus, we introduce sensor-aware input embedding. Specifically, we embed \\(\\mathcal{P}_i\\) from sensor \\(S_{i}\\) by:" + }, + { + "type": "equation", + "bbox": [ + 0.214, + 0.339, + 0.469, + 0.358 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {F}} _ {i} ^ {\\mathcal {P}} = h _ {\\mathcal {S} _ {i}} ^ {\\mathcal {P}} \\left(\\mathcal {F} _ {i} ^ {P}\\right), \\tag {12}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.217, + 0.361, + 0.469, + 0.38 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {F}} _ {i} ^ {\\mathcal {V}} = h _ {\\mathcal {S} _ {i}} ^ {\\mathcal {V}} \\left(\\mathcal {F} _ {i} ^ {V}\\right), \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.393, + 0.469, + 0.53 + ], + "angle": 0, + "content": "where \\( h_{\\mathcal{S}_i}^{\\mathcal{P}} \\) means shared MLPs on points and \\( h_{\\mathcal{S}_i}^{\\mathcal{V}} \\) means sparse convolution operations on voxels. Then, we learn voxel features and point features following the extraction strategy in Section 3.1 and update hierarchical geometry pools using Algorithm 1. We fuse point-level geometric information to voxel representation via geometry-to-voxel auxiliary mechanism. Our auxiliary task is to predict point labels for local point patches with point branch and point decoder. Therefore, the final loss function is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.545, + 0.469, + 0.563 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {v} + \\mu \\mathcal {L} _ {p}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.576, + 0.469, + 0.607 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_v\\) and \\(\\mathcal{L}_p\\) stand for per-voxel and per-point cross entropy loss and \\(\\mu\\) is the weight of the point loss." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.62, + 0.195, + 0.635 + ], + "angle": 0, + "content": "3.5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.469, + 0.901 + ], + "angle": 0, + "content": "In this section, we highlight previous point-voxel networks [24, 35, 52] which combine a point-based branch and a voxel-based branch together to leverage the advantages of both efficient voxel representations and elaborate point features. Additionally, we compare our GeoAuxNet with these methods. PVCNN [24] and PVT [52] rely on point-based networks to extract high quality representations with time-consuming searching or attention mechanisms, while SPVCNN [35] does not explore detailed local geometric information. Since the interaction between two branches in these methods only contains voxelization and devoxelization operations, they fail to benefit from both voxel representations and point-level elaborate features. Our GeoAuxNet preserves hierarchical sensor-aware geometry pools as the bridge to enable voxel representations to absorb point-level local geometric information without using the point network during inference." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.634, + 0.108 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.116, + 0.892, + 0.268 + ], + "angle": 0, + "content": "We validate our proposed GeoAuxNet across point clouds from multiple sensors. In Section 4.1, we train various methods on joint datasets including S3DIS [1], ScanNet [9] and SemanticKITTI [2]. In Section 4.2, we analyze the effectiveness of our sensor-aware hierarchical geometry pools. In Section 4.3, we compare GeoAuxNet with typical point-based and voxel-based methods to demonstrate the efficiency of our designed geometry-to-voxel auxiliary learning. In Section 4.4, we ablate different design choices of our voxel-guided dynamic point network." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.28, + 0.719, + 0.296 + ], + "angle": 0, + "content": "4.1. Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.304, + 0.892, + 0.561 + ], + "angle": 0, + "content": "Dataset. We conduct semantic segmentation experiments on the joint data of three datasets: S3DIS [1] and ScanNet [9], which are generated from RGB-D cameras, and SemanticKITTI [2], which is captured by LiDAR. The S3DIS dataset comprises 271 rooms from six areas in three distinct buildings, which are densely sampled on the surfaces of the meshes and annotated into 13 categories. Model performance evaluation is typically done using results from Area 5. The ScanNet dataset comprises 1,613 scene scans reconstructed from RGB-D frames. It is divided into 1,201 scenes for training, 312 scenes for validation, and 100 scenes for benchmark testing. SemanticKITTI consists of 22 point cloud sequences, where sequences 00 to 10, 08 and 11 to 21 are used for training, validation and testing. After merging classes with distinct moving status and discarding classes with very few points, a total of 19 classes are selected for training and evaluation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Implementation Details. Our voxel-based backbone is built on MinkowskiNet [8] with the batch normalization operations proposed by PPT [43]. The voxel size is set to \\(0.05\\mathrm{m}\\). We train our methods as well as MinkowskiNet SPVCNN [35] and PPT [43] on joint input data [1, 2, 9] captured by RGB-D cameras and LiDAR collectively. We also fine tune our method on three datasets separately. For the sake of fairness, we set the number of parameters of MinkowskiNet, SPVCNN, PPT and GeoAuxNet to about 60M. We use SGD optimizer and OneCycleLR [32] scheduler, with learning rate 0.05 and weight decay \\(10^{-4}\\). We adopt OneCycleLR [32] scheduler with \\(5\\%\\) percentage of the cycle spent increasing the learning rate and a cosine annealing strategy. We train all models with data augmentation used in PPT [43]. We set \\(\\lambda\\) in Eq. 7 to 0.1, \\(\\mu\\) in Eq. 14 to 0.1, threshold for geometry pool update \\(\\epsilon\\) to 0.9. The total number of iterations is equal to the sum of necessary iterations for all datasets. The sampling ratio across S3DIS, ScanNet and SemanticKITTI is set to 2:2:5. The number of points for the auxiliary branch does not exceed \\(2\\times 10^{4}\\). The max size of hierarchical geometry pools for each stage" + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20023" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.904, + 0.148 + ], + "angle": 0, + "content": "Table 1. Semantic segmentation results on three benchmarks including S3DIS [1], ScanNet [9] and SemanticKITTI [2]. We train SPVCNN [35], PPT [43] and GeoAuxNet on the joint training data of three datasets and also compare with experts [16, 17, 29, 35, 54, 56] on each single dataset. We report the class-average accuracy (mAcc, %) and class-wise mean IoU (mIoU, %). The type stands for point-based \\((p)\\) or voxel-based \\((\\nu)\\) methods, where \\(p \\rightarrow \\nu\\) means the geometry-to-voxel auxiliary learning." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.156, + 0.902, + 0.403 + ], + "angle": 0, + "content": "
MethodTypeParams.S3DIS Area5ScanNetSemanticKITTI
Test mIoUTest mAccVal mIoUVal mAccVal mIoUVal mAcc
PointTransformer [54]p11.4M70.476.570.6-
StratifiedFromer [16]p18.8M72.078.174.3-
PointNeXt [29]p41.6M70.577.271.5-
SPVCNN [35]p+v21.8M63.8-
Cylinder3D [56]v26.1M64.3-
SphereFormer [17]v32.3M67.8-
MinkowskiNet [8]v60.9M18.821.325.630.536.041.2
SPVCNN [35]p+v61.0M58.662.356.760.452.056.7
PPT [43]v63.0M30.434.616.527.425.031.2
PPT †[43]v63.0M67.973.269.677.561.868.9
GeoAuxNetp→v64.7M69.5(+1.6)74.5(+1.3)71.3(+1.7)79.3(+1.8)63.8(+2.0)69.3(+0.4)
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.091, + 0.404, + 0.904, + 0.434 + ], + "angle": 0, + "content": "† The original PPT with language-guided categorical alignment fails to converge on the joint training data. We use decoupled projection heads to employ PPT on multi-sensor datasets." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.457, + 0.316, + 0.473 + ], + "angle": 0, + "content": "is set to 32, 64, 128, 256 separately." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.493, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Results Comparison. As shown in Table 1, GeoAuxNet surpasses other methods which are trained on three datasets jointly. Our methods outperforms SPVCNN [35], which does not utilize its point branch to extract local information, by more than \\(10\\%\\) in mIoU on each dataset. PPT [43] introduces different datasets as prompts to collaboratively train a single model on multiple datasets from the same sensor type. The language-guided alignment of PPT suffers from the large domain gap caused by different sensors, which fails to converge on the joint training datasets. Therefore, we replace it by the decoupled alignment with separate heads for each dataset, which is also defined in PPT [43]. We construct sensor-aware geometry pools to address this issue and outperform PPT by about \\(2\\%\\) in mIoU on each dataset. Besides, we also compare with experts trained on each single dataset. We achieve competitive results with point-based methods on datasets from RGB-D cameras. More specifically, we even outperform PointTransformer in validation mIoU of ScanNet. Despite training on inter-domain datasets from different sensors, GeoAuxNet has a higher accuracy on SemanticKITTI than SPVCNN [35], and attain encouraging performances even compared with recent state-of-the-art expert [17]. Although we do not introduce any specific designs such as pointwise attention mechanism [16, 54] and cylinder coordinate system [56] for sparse LiDAR point clouds, the quantitative results demonstrate the effectiveness of geometry-to" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.457, + 0.892, + 0.504 + ], + "angle": 0, + "content": "voxel auxiliary learning by constructing hierarchical geometry pools to provide auxiliary sensor-aware point-level geometric priors for voxel representations." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.512, + 0.761, + 0.529 + ], + "angle": 0, + "content": "4.2. Hierarchical Geometry Pools" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.895, + 0.627 + ], + "angle": 0, + "content": "In this section, we further conduct experiments on our hierarchical geometry pools. First, we analyze the similarity of geometric structures in 3D space. Second, we discuss the appropriation to represent geometric information with our geometry pools. Finally, we study the construction of sensor-aware pools for different sensors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Similarity of Geometric Structures. Local geometric structures possess similarity in 3D space. For example, planar structures exist in desks, chairs and airplanes. These spatial structures are not unrelated but consistent. To further analyze the similarity of geometric structures learned by networks, we train our point encoder on ModelNet40 [44]. For each point, we use kNN to search 16 nearest neighbors and aggregate them via a max-pooling function to generate local geometric features. We select a center point marked as a purple star in Figure 3a and Figure 3c, and calculate the cosine similarity between its features and other point features. Specifically, the wing and the tail of the plane in Figure 3a have similar local structures, resulting in high similarity of features in Figure 3b. Figure 3d and Figure 3f show the same phenomenon at different stages in the network, where planar structures are similar to each other, while different from the chair legs. Since features from deeper lay" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20024" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.114, + 0.09, + 0.23, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.162, + 0.175, + 0.18, + 0.187 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.09, + 0.357, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.286, + 0.175, + 0.304, + 0.187 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.364, + 0.092, + 0.477, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.408, + 0.175, + 0.425, + 0.187 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.483, + 0.092, + 0.592, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.525, + 0.175, + 0.544, + 0.187 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.094, + 0.711, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.643, + 0.175, + 0.661, + 0.187 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.091, + 0.832, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.762, + 0.175, + 0.778, + 0.187 + ], + "angle": 0, + "content": "(f)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.199, + 0.894, + 0.256 + ], + "angle": 0, + "content": "Figure 3. Visualization of the cosine similarity between features. The purple star is the selected point, and the green star is the point with a significantly similar feature to the purple star. We calculate the cosine similarity between the feature of the red star and other features and visualize them in image (b), (d) and (f). The nearest neighbors of the purple and green stars are marked red in image (a), (c) and (e). (a), (b), (c) and (e) are generated from the first stage of the point network and (e), (f) are from the second stage." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.28, + 0.268, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.137, + 0.391, + 0.212, + 0.404 + ], + "angle": 0, + "content": "(a) Stage One" + }, + { + "type": "image", + "bbox": [ + 0.282, + 0.281, + 0.465, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.336, + 0.391, + 0.413, + 0.404 + ], + "angle": 0, + "content": "(b) Stage Two" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.406, + 0.266, + 0.512 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.132, + 0.516, + 0.216, + 0.529 + ], + "angle": 0, + "content": "(c) Stage Three" + }, + { + "type": "image", + "bbox": [ + 0.282, + 0.406, + 0.466, + 0.512 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.335, + 0.516, + 0.413, + 0.529 + ], + "angle": 0, + "content": "(d) Stage Four" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.541, + 0.47, + 0.584 + ], + "angle": 0, + "content": "Figure 4. Statistical results of cosine similarity between hierarchical geometry pools and point-level features extracted by the point network from point clouds in S3DIS [1] at different stages." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.59, + 0.47, + 0.651 + ], + "angle": 0, + "content": "ers possess larger receptive field sizes due to grouping and aggregating operations and comprise more geometric information, our hierarchical pools contain more geometric features for deeper layers." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.668, + 0.47, + 0.789 + ], + "angle": 0, + "content": "Representativeness of Geometry Pools. To demonstrate the representativeness of our pools, we calculate cosine similarity between all features in a scene from S3DIS [1] extracted by the point encoder and the features stored in the pool. For each stage, given the geometry pool \\(\\mathcal{F}^D = \\{f_i^D, 1 \\leq i \\leq N\\}\\) and all features in a scene \\(\\mathcal{F}^S = \\{f_i^S, 1 \\leq i \\leq M\\}\\), the similarity \\(s_i\\) between \\(f_i^S\\) and \\(\\mathcal{F}^D\\) is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.794, + 0.47, + 0.833 + ], + "angle": 0, + "content": "\\[\ns _ {i} = \\max _ {1 \\leq j \\leq N} \\frac {f _ {i} ^ {S} \\cdot f _ {j} ^ {D}}{\\| f _ {i} ^ {S} \\| \\| f _ {j} ^ {D} \\|}, \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "where \\( \\|\\cdot\\| \\) stands for L2 norm. The distribution of similarities is shown in Figure 4. From the statistical data, most features have cosine similarity higher than 0.85, which indicates that each feature can find a similar feature among our" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.282, + 0.676, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.581, + 0.413, + 0.598, + 0.424 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.688, + 0.281, + 0.853, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.759, + 0.412, + 0.777, + 0.424 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.862, + 0.281, + 0.89, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.437, + 0.893, + 0.507 + ], + "angle": 0, + "content": "Figure 5. Cosine similarity between geometry pools. (a) shows the similarity between geometry pools of S3DIS [1] and ScanNet [9]. (b) shows the similarity between geometry pools of S3DIS [1] and SemanticKITTI [2]. Intra-sensor geometry pools in (a) have higher similarity than inter-sensor geometry pools in (b)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.52, + 0.893, + 0.627 + ], + "angle": 0, + "content": "learned representatives in the pool. In this manner, we avoid using point-based network to process large scale points to generate fine-grained features for each voxel. Our hierarchical pools contain typical geometric features of scenes. Therefore, we can fuse point-level geometry information stored in our pools into voxel representations via geometry-to-voxel auxiliary mechanism efficiently." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Sensor-aware Pools for Multiple Sensors. Motivated by the diversity of density and sampling patterns of point clouds from various sensors, we construct different geometry pools to preserve relevant geometric information for diverse sensors. We investigate another two strategies: (i) shared geometry pools across sensors and (ii) different geometry pools for different datasets. We calculate the cosine similarity between features from different pools. As shown in Figure 5, geometry pools from the same sensor possess higher similarity than those from different sensors. As shown in Table 2, training with shared geometry pools across sensors leads to a significant decrease on performance due to the inherent diversity of geometric structures from different sensors, while sharing geometry pools with datasets from the same sensor serves as data augmentation to improve the effectiveness of the geometry pools." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20025" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.472, + 0.187 + ], + "angle": 0, + "content": "Table 2. Ablation study of sensor-aware geometry pools. \"Shared intra-sensor\" means the geometry pools which are shared for point clouds from the same sensor, while different for point clouds from different sensors. \"Shared inter-sensor\" means the geometry pools shared for different sensors. \"Shared intra-dataset\" means the geometry pools are different for different datasets. We report mIoU and mAcc on Area 5 of S3DIS [1]." + }, + { + "type": "table", + "bbox": [ + 0.108, + 0.198, + 0.437, + 0.276 + ], + "angle": 0, + "content": "
MethodsmIoU (%)mAcc (%)
Shared intra-sensor69.574.5
Shared inter-sensor61.4(−8.1)68.3(−6.2)
Shared intra-dataset67.5(−2.0)73.3(−1.2)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.292, + 0.471, + 0.322 + ], + "angle": 0, + "content": "Table 3. Efficiency analysis. We report the inference time and throughput on pre-processed data from S3DIS dataset." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.332, + 0.479, + 0.484 + ], + "angle": 0, + "content": "
MethodParams.Inference Time (ms)Throughput (ins./sec.)
PointNet++ [28]1.0M378.7108
PT [54]7.8M1079.734
PTv2 [42]3.9M4732.212
MinkowskiNet [8]60.9M45.4898
SPVCNN [35]61.0M48.4682
PPT [43]63.0M80.4251
GeoAuxNet64.7M51.8572
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.496, + 0.303, + 0.512 + ], + "angle": 0, + "content": "4.3. Efficiency of GeoAuxNet" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.47, + 0.748 + ], + "angle": 0, + "content": "We validate the efficiency of GeoAuxNet compared with other typical point-based [42, 54] and voxel-based methods [8, 43]. We focus on the mean inference time and throughput by employing a standardized pre-processed point cloud with instance labels from S3DIS Area 5 dataset as the input on a single NVIDIA A40 GPU. Initially, the device undergoes a \"warm-up\" phase, where the model is fed with the input point cloud ten times consecutively to eliminate any latency associated with data loading. Subsequently, the model is subjected to an additional 300 iterations of input feeding to calculate mean inference time. To effectively quantify the throughput, we measure the number of instances that a model processes per second. This systematic approach ensures a thorough and reliable assessment of our GeoAuxNet and other state-of-the-art methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In Table 3, we show the number of parameters, inference time and throughput for typical point-based and voxel-based networks. Specifically, our GeoAuxNet outperforms the expert model of PointTransformer [54] which is trained on ScanNet [9] individually in mIoU with \\(16.8 \\times\\) throughput and \\(20.8 \\times\\) inference speedup. Even compared with the simple PointNet++ [28] whose parameter size is \\(1/68\\) of ours, GeoAuxNet has less inference time and higher throughput. Besides, our method has similar inference time and throughput to voxel-based methods but achieve" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.118 + ], + "angle": 0, + "content": "Table 4. Ablation study of our voxel-guided hypernetwork. We report the mIoU and mAcc on Area 5 of S3DIS [1]." + }, + { + "type": "table", + "bbox": [ + 0.53, + 0.129, + 0.857, + 0.221 + ], + "angle": 0, + "content": "
MethodsmIoU (%)mAcc (%)
Vanilla63.368.7
+ voxel-guided66.0(+2.7)72.5(+3.8)
+ relative positions66.8(+0.8)73.2(+0.7)
+ stage latent code67.6(+0.8)73.3(+0.1)
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.23, + 0.894, + 0.321 + ], + "angle": 0, + "content": "better performance according to the results in Table 1. GeoAuxNet surpasses PPT [43] on each benchmark about \\(2\\%\\) in mIoU with \\(2.3\\times\\) throughput and \\(1.6\\times\\) inference speedup. Our method maintains efficiency for large scale point clouds, while achieves encouraging performances on datasets from multiple sensors." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.332, + 0.761, + 0.348 + ], + "angle": 0, + "content": "4.4. Voxel-guided Hypernetworks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.355, + 0.892, + 0.582 + ], + "angle": 0, + "content": "To investigate the architecture design of our voxel-guided hypernetwork, we conduct ablation studies on S3DIS semantic segmentation. We first introduce a vanilla point network which learns the weight and bias directly. Then, we adopt a hypernetwork taking voxel features as input to generate weights and biases for the point network. We further take account of the relative positions of points. Finally, we optimize learnable stage latent codes to instruct the information of different layers. We train these four models on S3DIS dataset. As shown in Table 4, the exploration of voxel features provides significant guidance for the point network to learn high quality geometric information, leading to an increase in test accuracy by \\(2.7\\%\\) in mIoU and \\(3.8\\%\\) in mAcc. Relative positions and stage latent codes also improve the performance of our methods." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.597, + 0.619, + 0.613 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.623, + 0.892, + 0.835 + ], + "angle": 0, + "content": "In this paper, we propose GeoAuxNet for multi-sensor point clouds with designed geometry-to-voxel auxiliary learning. We construct hierarchical geometry pools to learn auxiliary sensor-aware point-level geometric priors at different layers for sensor-agnostic voxel features. To generate our geometry pools, we also introduce a voxel-guided dynamic point network to leverage voxel prior knowledge for elaborate point features extraction. Our proposed geometry-to-voxel auxiliary builds a bridge between point-level and voxel-level features in an efficient manner without using the point network during inference time. Experimental results have shown the effectiveness and efficiency of our methods. We hope this work will inspire future research towards universal representation learning for point clouds." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported by the National Natural Science Foundation of China under Grant 62206147." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20026" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Iro Armeni, Ozan Sener, Amir Roshan Zamir, Helen Jiang, Ioannis K. Brilakis, Martin Fischer, and Silvio Savarese. 3D Semantic Parsing of Large-Scale Indoor Spaces. CVPR, pages 1534-1543, 2016. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.472, + 0.24 + ], + "angle": 0, + "content": "[2] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, C. Stachniss, and Juergen Gall. SemanticKITTI: A Dataset for Semantic Scene Understanding of LiDAR Sequences. ICCV, pages 9296-9306, 2019. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.471, + 0.269 + ], + "angle": 0, + "content": "[3] Alexandre Boulch. ConvPoint: Continuous convolutions for point cloud processing. CGF, 88:24-34, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.27, + 0.47, + 0.311 + ], + "angle": 0, + "content": "[4] César Cadena, Anthony R. Dick, and Ian D. Reid. Multimodal Auto-Encoders as Joint Estimators for Robotics Scene Understanding. In RSS, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.312, + 0.47, + 0.38 + ], + "angle": 0, + "content": "[5] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yuxin Pan, Giancarlo Baldan, and Oscar Beijbom. nuScenes: A Multimodal Dataset for Autonomous Driving. CVPR, pages 11618-11628, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.469, + 0.424 + ], + "angle": 0, + "content": "[6] Hao Chen, Xiaojuan Qi, Lequan Yu, and Pheng-Ann Heng. DCAN: Deep Contour-Aware Networks for Accurate Gland Segmentation. CVPR, pages 2487-2496, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.425, + 0.469, + 0.465 + ], + "angle": 0, + "content": "[7] Jaesung Choe, Chunghyun Park, Francois Rameau, Jaesik Park, and In So Kweon. PointMixer: MLP-Mixer for Point Cloud Understanding. ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.467, + 0.469, + 0.52 + ], + "angle": 0, + "content": "[8] Christopher Bongsoo Choy, JunYoung Gwak, and Silvio Savarese. 4D Spatio-Temporal ConvNets: Minkowski Convolutional Neural Networks. CVPR, pages 3070-3079, 2019. 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.522, + 0.469, + 0.577 + ], + "angle": 0, + "content": "[9] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas A. Funkhouser, and Matthias Nießner. ScanNet: Richly-Annotated 3D Reconstructions of Indoor Scenes. CVPR, pages 2432-2443, 2017. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.579, + 0.469, + 0.633 + ], + "angle": 0, + "content": "[10] Jesse Dodge, Ana Marasovic, Gabriel Ilharco, Dirk Groeneweld, Margaret Mitchell, and Matt Gardner. Documenting Large Webtext Corpora: A Case Study on the Colossal Clean Crawled Corpus. In EMNLP, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.635, + 0.469, + 0.674 + ], + "angle": 0, + "content": "[11] Nico Engel, Vasileios Belagiannis, and Klaus C. J. Dietmayer. Point Transformer. IEEE, 9:134826-134840, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.676, + 0.469, + 0.731 + ], + "angle": 0, + "content": "[12] Benjamin Graham, Martin Engelcke, and Laurens van der Maaten. 3D Semantic Segmentation with Submanifold Sparse Convolutional Networks. CVPR, pages 9224-9232, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.469, + 0.773 + ], + "angle": 0, + "content": "[13] Lei Han, Tian Zheng, Lan Xu, and Lu Fang. OccuSeg: Occupancy-Aware 3D Instance Segmentation. CVPR, pages 2937-2946, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.469, + 0.816 + ], + "angle": 0, + "content": "[14] Judy Hoffman, Eric Tzeng, Trevor Darrell, and Kate Saenko. Simultaneous Deep Transfer Across Domains and Tasks. ICCV, pages 4068-4076, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[15] Lingdong Kong, You-Chen Liu, Runnan Chen, Yuexin Ma, Xinge Zhu, Yikang Li, Yuenan Hou, Y. Qiao, and Ziwei Liu. Rethinking Range View Representation for LiDAR Segmentation. ArXiv, abs/2303.05367, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[16] Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, and Jiaya Jia. Stratified Trans-" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "former for 3D Point Cloud Segmentation. CVPR, pages 8490-8499, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.163 + ], + "angle": 0, + "content": "[17] Xin Lai, Yukang Chen, Fanbin Lu, Jianhui Liu, and Jiaya Jia. Spherical Transformer for LiDAR-Based 3D Recognition. CVPR, pages 17545-17555, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.165, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[18] John Lambert, Zhuang Liu, Ozan Sener, James Hays, and Vladlen Koltun. MSeg: A Composite Dataset for Multi-Domain Semantic Segmentation. PAMI, 45:796-810, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.221, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[19] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. PointCNN: Convolution On X-Transformed Points. In NIPS, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.264, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[20] Yingwei Li, Adams Wei Yu, Tianjian Meng, Benjamin Caine, Jiquan Ngiam, Daiyi Peng, Junyang Shen, Bo-Xun Wu, Yifeng Lu, Denny Zhou, Quoc V. Le, Alan Loddon Yuille, and Mingxing Tan. DeepFusion: Lidar-Camera Deep Fusion for Multi-Modal 3D Object Detection. CVPR, pages 17161-17170, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.348, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[21] Xingyu Lin, Harjatin Singh Baweja, George A. Kantor, and David Held. Adaptive Auxiliary Task Weighting for Reinforcement Learning. In NIPS, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.391, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[22] Zhi-Hao Lin, Sheng Yu Huang, and Y. Wang. Convolution in the Cloud: Learning Deformable Kernels in 3D Graph Convolution Networks for Point Cloud Analysis. CVPR, pages 1797-1806, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.447, + 0.892, + 0.488 + ], + "angle": 0, + "content": "[23] Xianpeng Liu, Nan Xue, and Tianfu Wu. Learning Auxiliary Monocular Contexts Helps Monocular 3D Object Detection. In AAAI, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.49, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[24] Zhijian Liu, Haotian Tang, Yujun Lin, and Song Han. PointVoxel CNN for Efficient 3D Deep Learning. In NIPS, 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.533, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[25] Daniel Maturana and Sebastian A. Scherer. VoxNet: A 3D Convolutional Neural Network for real-time object recognition. IROS, pages 922-928, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.576, + 0.892, + 0.617 + ], + "angle": 0, + "content": "[26] Andres Milioto, Ignacio Vizzo, Jens Behley, and C. Stachniss. RangeNet ++: Fast and Accurate LiDAR Semantic Segmentation. IROS, pages 4213-4220, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.618, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[27] C. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. Point-Net: Deep Learning on Point Sets for 3D Classification and Segmentation. CVPR, pages 77–85, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.661, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[28] C. Qi, L. Yi, Hao Su, and Leonidas J. Guibas. PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space. In NIPS, 2017. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.704, + 0.892, + 0.759 + ], + "angle": 0, + "content": "[29] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. PointNeXt: Revisiting PointNet++ with Improved Training and Scaling Strategies. In NIPS, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.761, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[30] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards Robust Monocular Depth Estimation: Mixing Datasets for Zero-Shot Cross-Dataset Transfer. PAMI, 44:1623-1637, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.817, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[31] Gernot Riegler, Ali O. Ulusoy, and Andreas Geiger. Oct-Net: Learning Deep 3D Representations at High Resolutions. CVPR, pages 6620-6629, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[32] Leslie N. Smith and Nicholay Topin. Super-Convergence: Very Fast Training of Residual Networks Using Large Learning Rates. ArXiv, abs/1708.07120, 2017. 5" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20027" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "[33] Shuran Song, Samuel P. Lichtenberg, and Jianxiong Xiao. SUN RGB-D: A RGB-D scene understanding benchmark suite. CVPR, pages 567-576, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.469, + 0.245 + ], + "angle": 0, + "content": "[34] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott M. Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in Perception for Autonomous Driving: Waymo Open Dataset. CVPR, pages 2443-2451, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.246, + 0.469, + 0.3 + ], + "angle": 0, + "content": "[35] Haotian* Tang, Zhijian* Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching Efficient 3D Architectures with Sparse Point-Voxel Convolution. In ECCV, 2020. 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.301, + 0.469, + 0.357 + ], + "angle": 0, + "content": "[36] Hugues Thomas, C. Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J. Guibas. KPConv: Flexible and Deformable Convolution for Point Clouds. ICCV, pages 6410–6419, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.358, + 0.469, + 0.398 + ], + "angle": 0, + "content": "[37] Li Wang, Dong Wei Li, Han Liu, Jinzhan Peng, Lu Tian, and Yi Shan. Cross-Dataset Collaborative Learning for Semantic Segmentation. AAAI, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.469, + 0.44 + ], + "angle": 0, + "content": "[38] Xudong Wang, Zhaowei Cai, Dashan Gao, and Nuno Vasconcelos. Towards Universal Object Detection by Domain Attention. CVPR, pages 7281-7290, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.441, + 0.469, + 0.495 + ], + "angle": 0, + "content": "[39] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E. Sarma, Michael M. Bronstein, and Justin M. Solomon. Dynamic Graph CNN for Learning on Point Clouds. TOG, 38:1-12, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.469, + 0.538 + ], + "angle": 0, + "content": "[40] Christian Wojek, Stefan Walk, Stefan Roth, and Bernt Schiele. Monocular 3D scene understanding with explicit occlusion reasoning. CVPR, pages 1993-2000, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.469, + 0.594 + ], + "angle": 0, + "content": "[41] Bichen Wu, Alvin Wan, Xiangyu Yue, and Kurt Keutzer. SqueezeSeg: Convolutional Neural Nets with Recurrent CRF for Real-Time Road-Object Segmentation from 3D Li-DAR Point Cloud. ICRA, pages 1887–1893, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.469, + 0.636 + ], + "angle": 0, + "content": "[42] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point Transformer V2: Grouped Vector Attention and Partition-based Pooling. In NIPS, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.637, + 0.469, + 0.692 + ], + "angle": 0, + "content": "[43] Xiaoyang Wu, Zhuotao Tian, Xin Wen, Bohao Peng, Xihui Liu, Kaicheng Yu, and Hengshuang Zhao. Towards Large-scale 3D Representation Learning with Multi-dataset Point Prompt Training. ArXiv, abs/2308.09718, 2023. 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.469, + 0.748 + ], + "angle": 0, + "content": "[44] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3D ShapeNets: A deep representation for volumetric shapes. CVPR, pages 1912-1920, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.469, + 0.816 + ], + "angle": 0, + "content": "[45] Yichen Xie, Chenfeng Xu, Marie-Julie Rakotosaona, Patrick Rim, Federico Tombari, Kurt Keutzer, Masayoshi Tomizuka, and Wei Zhan. SparseFusion: Fusing Multi-Modal Sparse Representations for Multi-Sensor 3D Object Detection. ICCV, abs/2304.14340, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[46] Guodong Xu, Ziwei Liu, Xiaoxiao Li, and Chen Change Loy. Knowledge Distillation Meets Self-Supervision. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[47] Hang Xu, Linpu Fang, Xiaodan Liang, Wenxiong Kang, and Zhenguo Li. Universal-RCNN: Universal Object Detector via Transferable Graph R-CNN. In AAAI, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[48] Mutian Xu, Runyu Ding, Hengshuang Zhao, and Xiaojuan Qi. PACnv: Position Adaptive Convolution with Dynamic Kernel Assembling on Point Clouds. CVPR, pages 3172-3181, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[49] Yu-Qi Yang, Yu-Xiao Guo, Jiangfeng Xiong, Yang Liu, Hao Pan, Peng-Shuai Wang, Xin Tong, and Baining Guo. Swin3D: A Pretrained Transformer Backbone for 3D Indoor Scene Understanding. ArXiv, abs/2304.06906, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[50] Zetong Yang, Yanan Sun, Shu Liu, Xiaoyong Shen, and Jiaya Jia. STD: Sparse-to-Dense 3D Object Detector for Point Cloud. ICCV, pages 1951-1960, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[51] Tianwei Yin, Xingyi Zhou, and Philipp Kähenbuhl. Center-based 3D Object Detection and Tracking. CVPR, pages 11779-11788, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[52] Cheng Zhang, Haocheng Wan, Xinyi Shen, and Zizhao Wu. PVT: Point-voxel transformer for point cloud learning. *IJIS*, 37:11985 - 12008, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[53] Yang Zhang, Zixiang Zhou, Philip David, Xiangyu Yue, Zerong Xi, and Hassan Foroosh. PolarNet: An Improved Grid Representation for Online LiDAR Point Clouds Semantic Segmentation. CVPR, pages 9598-9607, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[54] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip Torr, and Vladlen Koltun. Point Transformer. In ICCV, 2021. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.892, + 0.474 + ], + "angle": 0, + "content": "[55] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Simple Multi-dataset Detection. CVPR, pages 7561-7570, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.892, + 0.532 + ], + "angle": 0, + "content": "[56] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and Asymmetrical 3D Convolution Networks for LiDAR Segmentation. CVPR, pages 9934-9943, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.533, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[57] Özgün Çiçek, Ahmed Abdulkadir, Soeren S. Lienkamp, Thomas Brox, and Olaf Ronneberger. 3D U-Net: Learning Dense Volumetric Segmentation from Sparse Annotation. In MICCAI, 2016. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.587 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20028" + } + ] +] \ No newline at end of file diff --git a/2024/GeoAuxNet_ Towards Universal 3D Representation Learning for Multi-sensor Point Clouds/f223747a-d407-4fbb-9660-d95973ceefec_origin.pdf b/2024/GeoAuxNet_ Towards Universal 3D Representation Learning for Multi-sensor Point Clouds/f223747a-d407-4fbb-9660-d95973ceefec_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1b878a7e30dcdcae5b95a19fc416a70a5d888c59 --- /dev/null +++ b/2024/GeoAuxNet_ Towards Universal 3D Representation Learning for Multi-sensor Point Clouds/f223747a-d407-4fbb-9660-d95973ceefec_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:840e6a5da07cd33816af50202a54921df0f4cb0ff6b2366ffa4cb5b9dfc3a6a7 +size 1654773 diff --git a/2024/GeoAuxNet_ Towards Universal 3D Representation Learning for Multi-sensor Point Clouds/full.md b/2024/GeoAuxNet_ Towards Universal 3D Representation Learning for Multi-sensor Point Clouds/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5b9fc03d5dc47b2dd1e70806a51585b8eab940cc --- /dev/null +++ b/2024/GeoAuxNet_ Towards Universal 3D Representation Learning for Multi-sensor Point Clouds/full.md @@ -0,0 +1,392 @@ +# GeoAuxNet: Towards Universal 3D Representation Learning for Multi-sensor Point Clouds + +Shengjun Zhang $^{1}$ , Xin Fei $^{2}$ , Yueqi Duan $^{1\dagger}$ + +1 Department of Electronic Engineering, Tsinghua University + +$^{2}$ Department of Automation, Tsinghua University + +{zhangsj23, feix21}@mails.tsinghua.edu.cn, duanyueqi@tsinghua.edu.cn + +https://github.com/zhangshengjun2019/GeoAuxNet + +S3DIS RGB-D + +ScanNet RGB-D + +![](images/f9fa253af30bdc16ecfba1378f7a3ddd9777ca60233a5fec21806f2f77f2721c.jpg) + +![](images/035c1fc49a6ab46e0a0ec9be69ccabbf12b2f3e8730b1d4b0c3b72c947edc90c.jpg) + +![](images/1b6ecc3f2bddbef315988f6667e68e67d220548f9ed23eba17ec40434e367129.jpg) + +![](images/0eff91df56662770d7c04c32fc45ff088170b7e0d937420805c3787e023b1bb0.jpg) + +![](images/4288e65a0a4405c27bf51de3bea7abbb0e6ef3cd4a35ea28e825a66c4f3df4c6.jpg) + +![](images/621246fc57fdf4ad71411102708842118b71049542d1d8bb3901955463b1ff09.jpg) + +![](images/ebf1348708d946905c1f90c3095b06f98e15f3af7902ee1c44789c6f78333c6b.jpg) + +![](images/65d0de5372e84dc7f045cb588678bf106d5e6b38859d33b7b6415ff7585fbb21.jpg) + +![](images/1775f9fac4575a54898da131783cf35497e2db0284522493a08ce95d1b0f55df.jpg) + +![](images/2aa5d9a94727c261f3587f3d4d6c07f539c54f92b00cbd98b3f99cd156bb6440.jpg) + +![](images/d2f7654ed92c096ba16bdc162a4848135cde7d76f667090f2b437b1751bc32c3.jpg) + +![](images/256874226c954a1d8958a648d2b8fd3cfad654fa70f96541367e39458f901402.jpg) + +Semantic KITTI LiDAR + +![](images/69eef06fd72fbe39980ced831df56c9a5d39d7a6255f89f8b1225fe5e5942f78.jpg) +Input + +![](images/77bb9faf76bbb8dce0e0ce68d26f103f03b94b1792ef833934bf2ffc86307ada.jpg) +MinkowskiNet + +![](images/4a7d5b77ae2634146c7d64f9b78bf61e9005f0c302bbf4ebcdfd1ea6bd1350f1.jpg) +SPVCNN +Figure 1. Semantic segmentation results on S3DIS [1] and ScanNet [9] from RGB-D cameras and SemanticKITTI [2] from LiDAR. For all methods, we trained collectively on three datasets. Our method outperforms other methods with better detailed structures. + +![](images/8fcd188bf04145dd4b48723f0e5e564e0a22d61eeb83f3eed88dad17d206e0c7.jpg) +PPT + +![](images/458920fd3ebf2201c5a4080515d84a96c25ee7e09a2ca70f61a0c96cf29e7642.jpg) +Ours + +![](images/ae239ccb8a8bb17d5d16ca105420545cc08c4a3557f86fcb16669813cd372c6f.jpg) +GroundTruth + +# Abstract + +Point clouds captured by different sensors such as RGB-D cameras and LiDAR possess non-negligible domain gaps. Most existing methods design different network architectures and train separately on point clouds from various sensors. Typically, point-based methods achieve outstanding performances on even-distributed dense point clouds from RGB-D cameras, while voxel-based methods are more efficient for large-range sparse LiDAR point clouds. In this paper, we propose geometry-to-voxel auxiliary learning to enable voxel representations to access point-level geometric information, which supports better generalisation of the voxel-based backbone with additional interpretations of multi-sensor point clouds. Specifically, we construct hierarchical geometry pools generated by a voxel-guided dynamic point network, which efficiently provide + +auxiliary fine-grained geometric information adapted to different stages of voxel features. We conduct experiments on joint multi-sensor datasets to demonstrate the effectiveness of GeoAuxNet. Enjoying elaborate geometric information, our method outperforms other models collectively trained on multi-sensor datasets, and achieve competitive results with the-state-of-art experts on each single dataset. + +# 1. Introduction + +Point cloud analysis has attracted widespread attention due to its growing applications, such as autonomous driving [20, 45, 51] and robotics [4, 40]. Unlike images which are represented by regular pixels, 3D point clouds are irregular and unordered. These characteristics aggravate the inconsistency on density and sampling patterns of point clouds captured by different sensors, such as RGB-D cameras [1, 9, 33] and LiDAR [2, 5, 34]. + +Typically, points generated from RGB-D pictures are equally distributed and dense, while points scanned by LiDAR are sparse and uneven. The diversity on input data hinders the construction of universal network architectures. For point clouds from RGB-D cameras, point-based methods [28, 29, 54] usually gather local information via grouping and aggregating neighborhood features to extract detailed spatial information. Since neighbors are not stored contiguously in point representations, indexing them requires the costly searching operations [24], which limits the application of these methods on large scale LiDAR point clouds. To address this challenge, voxel-based methods [12, 25, 31] leverage the regular 3D convolution to obtain the contiguous memory access pattern. However, the resolution of voxels is constrained by the memory and becomes lower at deeper stages, resulting in loss of local information [35]. An intuitive idea is to ensemble both point-based and voxel-based methods [24, 35, 52], yet they are still confronted with either high time-consumption or the lack of detailed geometric features, which fail to simultaneously process large scale point clouds from different sensors with fine-grained spatial information while maintaining efficiency. + +In this paper, we propose GeoAuxNet to provide point-level geometric information for voxel representations in an efficient manner with geometry-to-voxel auxiliary learning. The support of elaborate geometric features introduces additional point-level information, which cannot be fully exploited by voxel-based backbones. Motivated by the observation that the local geometric structures present high similarity at each stage of the network, we construct hierarchical geometry pools to learn auxiliary sensor-aware point-level geometric priors corresponding to different layers for sensor-agnostic voxel features. To update our geometry pools, we present a voxel-guided dynamic point network, where we leverage prior knowledge in voxel features to guide high quality spatial feature extraction. Then, we fuse elaborate geometric features in the pools into voxel representations via our designed geometry-to-voxel auxiliary mechanism. For large-scale point clouds, the hierarchical geometry pools store representative point-level features at different stages, which efficiently provide complementary geometric information without using the point-based networks during inference time. + +We conduct extensive experiments on multi-sensor datasets, including S3DIS [1] and ScanNet [9] from RGB-D cameras as well as SemanticKITTI [2] captured by LiDAR, to demonstrate the efficiency and effectiveness of our method. With a shared weight of backbone, our method outperforms other models trained on joint datasets from different sensors and achieves competitive performance with experts on single dataset. As shown in Figure 1, we trained MinkowskiNet [8], SPVCNN [35], PPT [43] and + +GeoAuxNet on the above three datasets collectively, where our method preserve better detailed structures for point clouds from various sensors. + +# 2. Related Works + +3D Scene Understanding. Since images are composed of regular pixels, point clouds generated from RGB-D cameras are well-distributed and dense. Voxelization on dense point clouds ignore detailed structure information. To directly learn from points, Qi et al. [27] first proposed PointNet to process point clouds through shared multi-layer perceptions (MLP) to get a global representation. Yet this method lacks the ability to comprehend local geometric features. Therefore, PointNet++ [28] built a hierarchical architecture to gather local information via grouping and aggregating neighborhood features. Following up MLP-based works [7, 29] focused on designing effective learning schemes to capture spatial features. Other methods also introduced specific convolution operations [3, 19, 36, 48], graph-based representations [22, 39] and self-attention mechanism [11, 16, 42, 49, 54] to explore the relationship between points. Apart from cameras, LiDAR is also a widely used sensor to collect point clouds. Some methods [15, 26, 41, 53] focused on projecting point clouds to 2D grids to utilize 2D CNN. However, 3D to 2D projection limits the representation quality of geometric structures [50] due to the sacrifice of 3D spatial information [48]. Point based methods have high time consumption because of the large scale and the sparsity of these point clouds. Voxel-based methods [12, 13, 17, 56, 57] show more robustness against the sparsity. In view of the sampling characteristic of LiDAR, Zhu et al. [56] introduced cylinder coordinate system, for nearby regions have greater density than farther away regions. Similarly, Lai et al. [17] proposed spherical transformer to aggregate information from dense close points to sparse distant ones. Voxel-based networks maintain efficiency when processing scene-level point clouds, but ignore some detailed spatial structures. + +Universal Learning from Multiple Data Sources. Large language models achieve remarkable progress on many natural language processing systems through pretraining on extremely large text corpora [10], which enables foundation models to process data across domains, involving multiple languages, various majors, diverse application scenarios and so on. In 2D computer vision, joint learning across domains improve model robustness for detection [38, 47, 55], depth estimation [30] and semantic segmentation [18, 37]. Wang et al. [38] proposed a universal object detector through a domain-attention mechanism. But they did not model the training differences between different datasets. Universal-RCNN [47] modeled the class relations by designing a inter-dataset attention module to in + +![](images/4e9e97d7b4d8c135ecae23d28cd421a78973969a9acfd51b27665b821a3fdea0.jpg) +Figure 2. The pipeline of our GeoAuxNet. For a complete scene point cloud $\mathcal{P}^{\mathcal{C}}$ and a local point patch $\mathcal{P} \subseteq \mathcal{P}^{\mathcal{C}}$ , our voxel-based backbone first voxelizes $\mathcal{P}^{\mathcal{C}}$ and conducts sparse convolutional operations. The voxel-guided hypernetwork takes relative positions, voxel features and a stage latent code as input to provide weights and biases for the point network. Then, we encode the spatial information for $\mathcal{P}$ with the point network and aggregate local features to generate geometric feature candidates. Following the update strategy, we construct hierarchical geometry pools. The geometry-to-voxel mechanism fuses geometric features stored in the pools to enable voxel representations to access point-level geometric information. We repeat the above process several times to extract effective representation hierarchically and predict the results with a voxel decoder for the primary task and a point decoder for the auxiliary task. The dotted line stands for the course of the auxiliary learning which is ignored during inference to ensure efficiency. Geo-to-Vox is abbreviation of Geometry-to-Voxel. + +corporates graph transfer learning for propagating relevant semantic information across multiple datasets. CDCL [37] pointed out that convolution filters can be shared across domains without accuracy loss, while normalization operators are not appropriate to share in view of the bias of statistics. Recently, Wu et al. [43] introduced point prompt training to optimize a learnable domain-specific prompt with language-guided categorical alignment, but processed point clouds from RGB-D and LiDAR separately. However, point clouds captured by different sensors have diverse density and distributions, which limits the exploration of the universal network on multiple data sources. + +Auxiliary Learning. Auxiliary learning aims to achieve outstanding performance on a single primary task with the assist of auxiliary tasks and has shown many successful applications in a wide range of fields, including knowledge distillation [46], transfer learning [14] and reinforcement learning [21]. For instance, DCAN [6] learned an auxiliary contour detection task for more accurate segmentation, while Araslanov investigated the joint learning of image classification and semantic segmentation. In 3D vision, MonoCon [23] introduced monocular contexts as auxiliary tasks in training for 3D object detection. Nevertheless, the inherent relationship between point-level features and voxel-level features is still under exploration to better leverage the advantages of two domains without any requirement + +of additional data. + +# 3. Method + +The overall framework of GeoAuxNet is illustrated in Figure 2. To generate point-level geometric features, we propose a voxel-guided hypernetwork to produce weights and biases for the point network. Then, point features are grouped and aggregated to extract fine-grained local features, which are employed to update geometry pools. For voxel-based backbone, we introduce geometry-to-voxel auxiliary mechanism to access elaborate spatial information in our hierarchical geometry pools. + +# 3.1. Voxel-guided Dynamic Point Network + +Point-based methods are mostly confronted with high time consumption towards large scale point clouds, since the farthest point sampling (FPS) and k-nearest neighbors (kNN) algorithms in these networks have a time complexity of $\mathcal{O}(n^2)$ . Therefore, to efficiently process large scale point clouds, input points are voxelized and passed through sparse convolution operations to extract voxel features in our voxel-based backbone. + +To generate point-level geometric information efficiently, we train our point network on local point patches instead of complete scene point clouds. However, less training data weakens the ability of the point network to extract local features. Considering that our voxel-based branch are + +trained on complete voxelized point clouds, we treat voxel features as prior knowledge to instruct the learning of local geometric information. Yet, points inside a voxel share the same voxel feature, which ignores detailed difference between points. Thus, we also introduce relative coordinates to depict diversity of points within the same voxel. On account of the hierarchical architectures of both the point network and the voxel-based backbone with several stages, we optimize a learnable latent code $s$ for each stage during our auxiliary learning. + +Given a local point patch $\mathcal{P} = \{p_i, 1 \leq i \leq N\} \in \mathbb{R}^{N \times 3}$ with point features $\mathcal{F}^{\mathcal{P}} = \{f_i^{\mathcal{P}}, 1 \leq i \leq N\}$ , each point $p_i$ belongs to a voxel $v$ with the voxel feature $f^{\mathcal{V}}$ . Here $v \in \mathbb{R}^3$ stands for the center coordinate of the voxel. Then, we can obtain the weight and bias for point $p_i$ : + +$$ +w _ {i} = h _ {w} \left(h _ {p} \left(p _ {i} - v\right), f ^ {\mathcal {V}}\right) \odot h _ {s} (s), \tag {1} +$$ + +$$ +b _ {i} = h _ {b} \left(h _ {p} \left(p _ {i} - v\right), f ^ {\mathcal {V}}\right) \odot h _ {s} (s), \tag {2} +$$ + +where $h_p$ and $h_s$ are MLPs to project relative positions and the stage latent code to specific dimensions, and $h_w$ and $h_b$ are hypernetworks to generate the weight and bias. Hence, new point features can be formulated as: + +$$ +\hat {f} _ {i} ^ {\mathcal {P}} = \operatorname {M L P} \left(w _ {i} \odot f _ {i} ^ {\mathcal {P}} + b _ {i}\right). \tag {3} +$$ + +To extract local geometric information, we use kNN algorithm to select $k$ nearest neighbor points and employ a max-pooling function to aggregate each local group. In this manner, we generate a set of geometric feature candidates for hierarchical geometry pools. + +# 3.2. Hierarchical Geometry Pools + +The voxel branch ignores some detailed spatial structures due to voxelization. Motivated by the similarity of geometric structures in 3D space, we establish a pool to store typical geometric patterns. The pool contains $n$ typical spatial structure features $\mathcal{F}^{\mathcal{D}} = \{f_i^{\mathcal{D}}, 1 \leq i \leq n\}$ . The max size of the pool is set to $N$ . In Section 3.1, we generate a set of geometric pattern candidates $\mathcal{F}^S = \{f_i^S, 1 \leq i \leq m\}$ . When updating the pool, we first calculate the similarity between $\mathcal{F}^S$ and $\mathcal{F}^D$ : + +$$ +s _ {i} = \max _ {1 \leq j \leq N} s _ {i, j}, \tag {4} +$$ + +where + +$$ +s _ {i, j} = \frac {f _ {i} ^ {\mathcal {S}} \cdot f _ {j} ^ {\mathcal {D}}}{\| f _ {i} ^ {\mathcal {S}} \| \| f _ {j} ^ {\mathcal {D}} \|.} \tag {5} +$$ + +Specifically, if $s_i \geq \epsilon$ , $f_i^S$ is merged into the geometry pool $\mathcal{F}^{\mathcal{D}}$ : + +$$ +t = \underset {1 \leq j \leq N} {\operatorname {a r g m a x}} s _ {i, j}, \tag {6} +$$ + +$$ +\hat {f} _ {t} ^ {\mathcal {D}} = \lambda f _ {t} ^ {\mathcal {D}} + (1 - \lambda) f _ {i} ^ {\mathcal {S}}, \tag {7} +$$ + +Algorithm 1 Update Pools +Input: New features $\mathcal{F}^S$ , geometry pool $\mathcal{F}^D$ , maximum size of the pool $N$ , threshold $\epsilon$ , update rate $\lambda$ +Output: New pool $\hat{\mathcal{F}}^D$ +Compute $s_i$ for $1 \leq i \leq m$ according to Eq. 4 +for $i \gets 1$ to $|\mathcal{F}^S|$ do + if $s_i \geq \epsilon$ then + t ← arg max $_{1 \leq j \leq N}$ $s_i, j$ $f_t^D \gets \lambda f_t^D + (1 - \lambda)f_i^S$ $\mathcal{F}^S \gets \mathcal{F}^S \setminus \{f_i^S\}$ + end if +end for +if $|\mathcal{F}^S| + |\mathcal{F}^D| \leq N$ then + $\mathcal{F}^D \gets \mathcal{F}^D \cup \mathcal{F}^S$ +else + $\hat{\mathcal{F}}^S \gets$ randomly select $N - |\mathcal{F}^D|$ features from $\mathcal{F}^S$ $\mathcal{F}^S \gets \mathcal{F}^S \setminus \hat{\mathcal{F}}^S$ $\mathcal{F}^D \gets \mathcal{F}^D \cup \hat{\mathcal{F}}^S$ + for $i \gets 1$ to $|\mathcal{F}^S|$ do + t ← arg max $_{1 \leq j \leq N}$ $s_i, j$ $f_t^D \gets \lambda f_t^D + (1 - \lambda)f_i^S$ + end for +end if + +where $\epsilon$ is a threshold and $\lambda$ stands for the update rate. Otherwise, we append the new features to the geometry pool. If the total number of features is more than $N$ , we randomly select $N - n$ new features and add them to the pool. The rest new features following Eq. 7 are merged into existing patterns. Our update strategy is illustrated in Algorithm 1. + +Since voxel representations at various stages contains different level information, we introduce hierarchical geometry pools to store representative point-level features corresponding to different stages. We also enlarge the size of pools for deeper stages due to more complex geometric structures with larger receptive field sizes. + +# 3.3. Geometry-to-Voxel Auxiliary + +After we update the geometric pool, we use geometric patterns stored in $\mathcal{F}^{\mathcal{D}}\in \mathbb{R}^{N\times C}$ to generate fine-grained local information for each voxel following a cross-attention operation. Given voxel features $\mathcal{F}^{\nu} = \{f_i^\nu ,1\leq i\leq M\} \in$ $\mathbb{R}^{M\times D}$ , the geometric fine-grained feature $f_{i}^{\mathcal{G}}$ for $f_{i}^{\nu}\in \mathcal{F}^{\nu}$ is formulated by: + +$$ +q _ {i} = h _ {Q} \left(f _ {i} ^ {\mathcal {V}}\right) \in \mathbb {R} ^ {1 \times s}, \tag {8} +$$ + +$$ +K = h _ {K} \left(\mathcal {F} ^ {\mathcal {D}}\right) \in \mathbb {R} ^ {N \times s}, \tag {9} +$$ + +$$ +V = h _ {V} \left(\mathcal {F} ^ {\mathcal {D}}\right) \in \mathbb {R} ^ {N \times t}, \tag {10} +$$ + +$$ +f _ {i} ^ {\mathcal {G}} = \operatorname {s o f t m a x} \left(\frac {q _ {i} K ^ {\top}}{\sqrt {s}}\right) V \in \mathbb {R} ^ {1 \times t}, \tag {11} +$$ + +where $h_{Q}, h_{K}$ , and $h_{V}$ stand for projections. Then, we concatenate $f_{i}^{\mathcal{G}}$ with $f_{i}^{\mathcal{V}}$ as the new voxel feature and employ sparse convolution operations. In this manner, we can provide complementary geometric information without using the point-based networks during inference stage. + +# 3.4. Overall Architecture + +Given a point cloud $\mathcal{P}_i^{\mathcal{C}}$ with features $\mathcal{F}_i^{\mathcal{C}}$ and a local point patch $\mathcal{P}_i\subseteq \mathcal{P}_i^{\mathcal{C}}$ with $\mathcal{F}_i^{\mathcal{P}}\subseteq \mathcal{F}_i^{\mathcal{C}}$ from sensor $S_{i}$ , we voxelize all points to obtain voxel coordinates $\nu_{i}$ with voxel features $\mathcal{F}_i^\nu$ . Points obtained from RGB-D cameras usually possess colors and normal vectors, while points captured by LiDAR always possess intensity. Thus, we introduce sensor-aware input embedding. Specifically, we embed $\mathcal{P}_i$ from sensor $S_{i}$ by: + +$$ +\hat {\mathcal {F}} _ {i} ^ {\mathcal {P}} = h _ {\mathcal {S} _ {i}} ^ {\mathcal {P}} \left(\mathcal {F} _ {i} ^ {P}\right), \tag {12} +$$ + +$$ +\hat {\mathcal {F}} _ {i} ^ {\mathcal {V}} = h _ {\mathcal {S} _ {i}} ^ {\mathcal {V}} \left(\mathcal {F} _ {i} ^ {V}\right), \tag {13} +$$ + +where $h_{\mathcal{S}_i}^{\mathcal{P}}$ means shared MLPs on points and $h_{\mathcal{S}_i}^{\mathcal{V}}$ means sparse convolution operations on voxels. Then, we learn voxel features and point features following the extraction strategy in Section 3.1 and update hierarchical geometry pools using Algorithm 1. We fuse point-level geometric information to voxel representation via geometry-to-voxel auxiliary mechanism. Our auxiliary task is to predict point labels for local point patches with point branch and point decoder. Therefore, the final loss function is formulated as: + +$$ +\mathcal {L} = \mathcal {L} _ {v} + \mu \mathcal {L} _ {p}, \tag {14} +$$ + +where $\mathcal{L}_v$ and $\mathcal{L}_p$ stand for per-voxel and per-point cross entropy loss and $\mu$ is the weight of the point loss. + +# 3.5. Discussion + +In this section, we highlight previous point-voxel networks [24, 35, 52] which combine a point-based branch and a voxel-based branch together to leverage the advantages of both efficient voxel representations and elaborate point features. Additionally, we compare our GeoAuxNet with these methods. PVCNN [24] and PVT [52] rely on point-based networks to extract high quality representations with time-consuming searching or attention mechanisms, while SPVCNN [35] does not explore detailed local geometric information. Since the interaction between two branches in these methods only contains voxelization and devoxelization operations, they fail to benefit from both voxel representations and point-level elaborate features. Our GeoAuxNet preserves hierarchical sensor-aware geometry pools as the bridge to enable voxel representations to absorb point-level local geometric information without using the point network during inference. + +# 4. Experiments + +We validate our proposed GeoAuxNet across point clouds from multiple sensors. In Section 4.1, we train various methods on joint datasets including S3DIS [1], ScanNet [9] and SemanticKITTI [2]. In Section 4.2, we analyze the effectiveness of our sensor-aware hierarchical geometry pools. In Section 4.3, we compare GeoAuxNet with typical point-based and voxel-based methods to demonstrate the efficiency of our designed geometry-to-voxel auxiliary learning. In Section 4.4, we ablate different design choices of our voxel-guided dynamic point network. + +# 4.1. Semantic Segmentation + +Dataset. We conduct semantic segmentation experiments on the joint data of three datasets: S3DIS [1] and ScanNet [9], which are generated from RGB-D cameras, and SemanticKITTI [2], which is captured by LiDAR. The S3DIS dataset comprises 271 rooms from six areas in three distinct buildings, which are densely sampled on the surfaces of the meshes and annotated into 13 categories. Model performance evaluation is typically done using results from Area 5. The ScanNet dataset comprises 1,613 scene scans reconstructed from RGB-D frames. It is divided into 1,201 scenes for training, 312 scenes for validation, and 100 scenes for benchmark testing. SemanticKITTI consists of 22 point cloud sequences, where sequences 00 to 10, 08 and 11 to 21 are used for training, validation and testing. After merging classes with distinct moving status and discarding classes with very few points, a total of 19 classes are selected for training and evaluation. + +Implementation Details. Our voxel-based backbone is built on MinkowskiNet [8] with the batch normalization operations proposed by PPT [43]. The voxel size is set to $0.05\mathrm{m}$ . We train our methods as well as MinkowskiNet SPVCNN [35] and PPT [43] on joint input data [1, 2, 9] captured by RGB-D cameras and LiDAR collectively. We also fine tune our method on three datasets separately. For the sake of fairness, we set the number of parameters of MinkowskiNet, SPVCNN, PPT and GeoAuxNet to about 60M. We use SGD optimizer and OneCycleLR [32] scheduler, with learning rate 0.05 and weight decay $10^{-4}$ . We adopt OneCycleLR [32] scheduler with $5\%$ percentage of the cycle spent increasing the learning rate and a cosine annealing strategy. We train all models with data augmentation used in PPT [43]. We set $\lambda$ in Eq. 7 to 0.1, $\mu$ in Eq. 14 to 0.1, threshold for geometry pool update $\epsilon$ to 0.9. The total number of iterations is equal to the sum of necessary iterations for all datasets. The sampling ratio across S3DIS, ScanNet and SemanticKITTI is set to 2:2:5. The number of points for the auxiliary branch does not exceed $2\times 10^{4}$ . The max size of hierarchical geometry pools for each stage + +Table 1. Semantic segmentation results on three benchmarks including S3DIS [1], ScanNet [9] and SemanticKITTI [2]. We train SPVCNN [35], PPT [43] and GeoAuxNet on the joint training data of three datasets and also compare with experts [16, 17, 29, 35, 54, 56] on each single dataset. We report the class-average accuracy (mAcc, %) and class-wise mean IoU (mIoU, %). The type stands for point-based $(p)$ or voxel-based $(\nu)$ methods, where $p \rightarrow \nu$ means the geometry-to-voxel auxiliary learning. + +
MethodTypeParams.S3DIS Area5ScanNetSemanticKITTI
Test mIoUTest mAccVal mIoUVal mAccVal mIoUVal mAcc
PointTransformer [54]p11.4M70.476.570.6-
StratifiedFromer [16]p18.8M72.078.174.3-
PointNeXt [29]p41.6M70.577.271.5-
SPVCNN [35]p+v21.8M63.8-
Cylinder3D [56]v26.1M64.3-
SphereFormer [17]v32.3M67.8-
MinkowskiNet [8]v60.9M18.821.325.630.536.041.2
SPVCNN [35]p+v61.0M58.662.356.760.452.056.7
PPT [43]v63.0M30.434.616.527.425.031.2
PPT †[43]v63.0M67.973.269.677.561.868.9
GeoAuxNetp→v64.7M69.5(+1.6)74.5(+1.3)71.3(+1.7)79.3(+1.8)63.8(+2.0)69.3(+0.4)
+ +† The original PPT with language-guided categorical alignment fails to converge on the joint training data. We use decoupled projection heads to employ PPT on multi-sensor datasets. + +is set to 32, 64, 128, 256 separately. + +Results Comparison. As shown in Table 1, GeoAuxNet surpasses other methods which are trained on three datasets jointly. Our methods outperforms SPVCNN [35], which does not utilize its point branch to extract local information, by more than $10\%$ in mIoU on each dataset. PPT [43] introduces different datasets as prompts to collaboratively train a single model on multiple datasets from the same sensor type. The language-guided alignment of PPT suffers from the large domain gap caused by different sensors, which fails to converge on the joint training datasets. Therefore, we replace it by the decoupled alignment with separate heads for each dataset, which is also defined in PPT [43]. We construct sensor-aware geometry pools to address this issue and outperform PPT by about $2\%$ in mIoU on each dataset. Besides, we also compare with experts trained on each single dataset. We achieve competitive results with point-based methods on datasets from RGB-D cameras. More specifically, we even outperform PointTransformer in validation mIoU of ScanNet. Despite training on inter-domain datasets from different sensors, GeoAuxNet has a higher accuracy on SemanticKITTI than SPVCNN [35], and attain encouraging performances even compared with recent state-of-the-art expert [17]. Although we do not introduce any specific designs such as pointwise attention mechanism [16, 54] and cylinder coordinate system [56] for sparse LiDAR point clouds, the quantitative results demonstrate the effectiveness of geometry-to + +voxel auxiliary learning by constructing hierarchical geometry pools to provide auxiliary sensor-aware point-level geometric priors for voxel representations. + +# 4.2. Hierarchical Geometry Pools + +In this section, we further conduct experiments on our hierarchical geometry pools. First, we analyze the similarity of geometric structures in 3D space. Second, we discuss the appropriation to represent geometric information with our geometry pools. Finally, we study the construction of sensor-aware pools for different sensors. + +Similarity of Geometric Structures. Local geometric structures possess similarity in 3D space. For example, planar structures exist in desks, chairs and airplanes. These spatial structures are not unrelated but consistent. To further analyze the similarity of geometric structures learned by networks, we train our point encoder on ModelNet40 [44]. For each point, we use kNN to search 16 nearest neighbors and aggregate them via a max-pooling function to generate local geometric features. We select a center point marked as a purple star in Figure 3a and Figure 3c, and calculate the cosine similarity between its features and other point features. Specifically, the wing and the tail of the plane in Figure 3a have similar local structures, resulting in high similarity of features in Figure 3b. Figure 3d and Figure 3f show the same phenomenon at different stages in the network, where planar structures are similar to each other, while different from the chair legs. Since features from deeper lay + +![](images/ef135819b164502f5a8716d592ff894fe59ad82d2f7aeed5c4b39a9c5444a93f.jpg) +(a) + +![](images/18d61ff77136088f9be03e48200993f8e82bfe4e961130ea9af124df715c51de.jpg) +(b) + +![](images/fae73d3bdaa19de92d09a6aaf24785df40088d025f09e9d392588855884bc009.jpg) +(c) + +![](images/028dbe805c154ec5e6f80049a3b872fe6c1a4f765d613e3622e9e92df32817dd.jpg) +(d) + +![](images/031ee3badb36e5675dd67436c80e7f951c71ec29cd4edb11a8b23a770c9f4d63.jpg) +(e) + +![](images/d7112883770caceb463c692f7bb6835effa122c18891681a0aa58d5ace5163f5.jpg) +(f) + +![](images/38ad5ca00e8f04f0cf84398e28444d7880664b102b2882d808bd3fb2b6c093c9.jpg) +Figure 3. Visualization of the cosine similarity between features. The purple star is the selected point, and the green star is the point with a significantly similar feature to the purple star. We calculate the cosine similarity between the feature of the red star and other features and visualize them in image (b), (d) and (f). The nearest neighbors of the purple and green stars are marked red in image (a), (c) and (e). (a), (b), (c) and (e) are generated from the first stage of the point network and (e), (f) are from the second stage. + +![](images/d13192e61c6877c92ddf92b5d16c9fcec796fc89c388f8642a298cba067f4139.jpg) +(b) Stage Two + +![](images/d0de33688bd881555a4a86954612f05813e3b50aa2e2025a4abd43351de8d607.jpg) +(a) Stage One +(c) Stage Three +Figure 4. Statistical results of cosine similarity between hierarchical geometry pools and point-level features extracted by the point network from point clouds in S3DIS [1] at different stages. + +![](images/a02125abbb347cdebb504f8b46b709584b0a330802314605e95ba701428174af.jpg) +(d) Stage Four +Figure 5. Cosine similarity between geometry pools. (a) shows the similarity between geometry pools of S3DIS [1] and ScanNet [9]. (b) shows the similarity between geometry pools of S3DIS [1] and SemanticKITTI [2]. Intra-sensor geometry pools in (a) have higher similarity than inter-sensor geometry pools in (b). + +ers possess larger receptive field sizes due to grouping and aggregating operations and comprise more geometric information, our hierarchical pools contain more geometric features for deeper layers. + +Representativeness of Geometry Pools. To demonstrate the representativeness of our pools, we calculate cosine similarity between all features in a scene from S3DIS [1] extracted by the point encoder and the features stored in the pool. For each stage, given the geometry pool $\mathcal{F}^D = \{f_i^D, 1 \leq i \leq N\}$ and all features in a scene $\mathcal{F}^S = \{f_i^S, 1 \leq i \leq M\}$ , the similarity $s_i$ between $f_i^S$ and $\mathcal{F}^D$ is formulated as: + +$$ +s _ {i} = \max _ {1 \leq j \leq N} \frac {f _ {i} ^ {S} \cdot f _ {j} ^ {D}}{\| f _ {i} ^ {S} \| \| f _ {j} ^ {D} \|}, \tag {15} +$$ + +where $\|\cdot\|$ stands for L2 norm. The distribution of similarities is shown in Figure 4. From the statistical data, most features have cosine similarity higher than 0.85, which indicates that each feature can find a similar feature among our + +![](images/2c8e1ebf082ed4023da4b1ab056c0d195f1e69dcc78a63b020cc2bc34f0e79c5.jpg) +(a) + +![](images/d73edf02bb3173414dad3b59e9c165baeb5a9dcb25623c2dabc9cc16f815bdcd.jpg) +(b) + +![](images/73fa5de8b4b2cae4d4f8f2c282e8019f431196bdb9c0e56579430f574b99f75f.jpg) + +learned representatives in the pool. In this manner, we avoid using point-based network to process large scale points to generate fine-grained features for each voxel. Our hierarchical pools contain typical geometric features of scenes. Therefore, we can fuse point-level geometry information stored in our pools into voxel representations via geometry-to-voxel auxiliary mechanism efficiently. + +Sensor-aware Pools for Multiple Sensors. Motivated by the diversity of density and sampling patterns of point clouds from various sensors, we construct different geometry pools to preserve relevant geometric information for diverse sensors. We investigate another two strategies: (i) shared geometry pools across sensors and (ii) different geometry pools for different datasets. We calculate the cosine similarity between features from different pools. As shown in Figure 5, geometry pools from the same sensor possess higher similarity than those from different sensors. As shown in Table 2, training with shared geometry pools across sensors leads to a significant decrease on performance due to the inherent diversity of geometric structures from different sensors, while sharing geometry pools with datasets from the same sensor serves as data augmentation to improve the effectiveness of the geometry pools. + +Table 2. Ablation study of sensor-aware geometry pools. "Shared intra-sensor" means the geometry pools which are shared for point clouds from the same sensor, while different for point clouds from different sensors. "Shared inter-sensor" means the geometry pools shared for different sensors. "Shared intra-dataset" means the geometry pools are different for different datasets. We report mIoU and mAcc on Area 5 of S3DIS [1]. + +
MethodsmIoU (%)mAcc (%)
Shared intra-sensor69.574.5
Shared inter-sensor61.4(−8.1)68.3(−6.2)
Shared intra-dataset67.5(−2.0)73.3(−1.2)
+ +Table 3. Efficiency analysis. We report the inference time and throughput on pre-processed data from S3DIS dataset. + +
MethodParams.Inference Time (ms)Throughput (ins./sec.)
PointNet++ [28]1.0M378.7108
PT [54]7.8M1079.734
PTv2 [42]3.9M4732.212
MinkowskiNet [8]60.9M45.4898
SPVCNN [35]61.0M48.4682
PPT [43]63.0M80.4251
GeoAuxNet64.7M51.8572
+ +# 4.3. Efficiency of GeoAuxNet + +We validate the efficiency of GeoAuxNet compared with other typical point-based [42, 54] and voxel-based methods [8, 43]. We focus on the mean inference time and throughput by employing a standardized pre-processed point cloud with instance labels from S3DIS Area 5 dataset as the input on a single NVIDIA A40 GPU. Initially, the device undergoes a "warm-up" phase, where the model is fed with the input point cloud ten times consecutively to eliminate any latency associated with data loading. Subsequently, the model is subjected to an additional 300 iterations of input feeding to calculate mean inference time. To effectively quantify the throughput, we measure the number of instances that a model processes per second. This systematic approach ensures a thorough and reliable assessment of our GeoAuxNet and other state-of-the-art methods. + +In Table 3, we show the number of parameters, inference time and throughput for typical point-based and voxel-based networks. Specifically, our GeoAuxNet outperforms the expert model of PointTransformer [54] which is trained on ScanNet [9] individually in mIoU with $16.8 \times$ throughput and $20.8 \times$ inference speedup. Even compared with the simple PointNet++ [28] whose parameter size is $1/68$ of ours, GeoAuxNet has less inference time and higher throughput. Besides, our method has similar inference time and throughput to voxel-based methods but achieve + +Table 4. Ablation study of our voxel-guided hypernetwork. We report the mIoU and mAcc on Area 5 of S3DIS [1]. + +
MethodsmIoU (%)mAcc (%)
Vanilla63.368.7
+ voxel-guided66.0(+2.7)72.5(+3.8)
+ relative positions66.8(+0.8)73.2(+0.7)
+ stage latent code67.6(+0.8)73.3(+0.1)
+ +better performance according to the results in Table 1. GeoAuxNet surpasses PPT [43] on each benchmark about $2\%$ in mIoU with $2.3\times$ throughput and $1.6\times$ inference speedup. Our method maintains efficiency for large scale point clouds, while achieves encouraging performances on datasets from multiple sensors. + +# 4.4. Voxel-guided Hypernetworks + +To investigate the architecture design of our voxel-guided hypernetwork, we conduct ablation studies on S3DIS semantic segmentation. We first introduce a vanilla point network which learns the weight and bias directly. Then, we adopt a hypernetwork taking voxel features as input to generate weights and biases for the point network. We further take account of the relative positions of points. Finally, we optimize learnable stage latent codes to instruct the information of different layers. We train these four models on S3DIS dataset. As shown in Table 4, the exploration of voxel features provides significant guidance for the point network to learn high quality geometric information, leading to an increase in test accuracy by $2.7\%$ in mIoU and $3.8\%$ in mAcc. Relative positions and stage latent codes also improve the performance of our methods. + +# 5. Conclusion + +In this paper, we propose GeoAuxNet for multi-sensor point clouds with designed geometry-to-voxel auxiliary learning. We construct hierarchical geometry pools to learn auxiliary sensor-aware point-level geometric priors at different layers for sensor-agnostic voxel features. To generate our geometry pools, we also introduce a voxel-guided dynamic point network to leverage voxel prior knowledge for elaborate point features extraction. Our proposed geometry-to-voxel auxiliary builds a bridge between point-level and voxel-level features in an efficient manner without using the point network during inference time. Experimental results have shown the effectiveness and efficiency of our methods. We hope this work will inspire future research towards universal representation learning for point clouds. + +Acknowledgements. This work was supported by the National Natural Science Foundation of China under Grant 62206147. + +# References + +[1] Iro Armeni, Ozan Sener, Amir Roshan Zamir, Helen Jiang, Ioannis K. Brilakis, Martin Fischer, and Silvio Savarese. 3D Semantic Parsing of Large-Scale Indoor Spaces. CVPR, pages 1534-1543, 2016. 1, 2, 5, 6, 7, 8 +[2] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, C. Stachniss, and Juergen Gall. SemanticKITTI: A Dataset for Semantic Scene Understanding of LiDAR Sequences. ICCV, pages 9296-9306, 2019. 1, 2, 5, 6, 7 +[3] Alexandre Boulch. ConvPoint: Continuous convolutions for point cloud processing. CGF, 88:24-34, 2019. 2 +[4] César Cadena, Anthony R. Dick, and Ian D. Reid. Multimodal Auto-Encoders as Joint Estimators for Robotics Scene Understanding. In RSS, 2016. 1 +[5] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yuxin Pan, Giancarlo Baldan, and Oscar Beijbom. nuScenes: A Multimodal Dataset for Autonomous Driving. CVPR, pages 11618-11628, 2019. 1 +[6] Hao Chen, Xiaojuan Qi, Lequan Yu, and Pheng-Ann Heng. DCAN: Deep Contour-Aware Networks for Accurate Gland Segmentation. CVPR, pages 2487-2496, 2016. 3 +[7] Jaesung Choe, Chunghyun Park, Francois Rameau, Jaesik Park, and In So Kweon. PointMixer: MLP-Mixer for Point Cloud Understanding. ECCV, 2022. 2 +[8] Christopher Bongsoo Choy, JunYoung Gwak, and Silvio Savarese. 4D Spatio-Temporal ConvNets: Minkowski Convolutional Neural Networks. CVPR, pages 3070-3079, 2019. 2, 5, 6, 8 +[9] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas A. Funkhouser, and Matthias Nießner. ScanNet: Richly-Annotated 3D Reconstructions of Indoor Scenes. CVPR, pages 2432-2443, 2017. 1, 2, 5, 6, 7, 8 +[10] Jesse Dodge, Ana Marasovic, Gabriel Ilharco, Dirk Groeneweld, Margaret Mitchell, and Matt Gardner. Documenting Large Webtext Corpora: A Case Study on the Colossal Clean Crawled Corpus. In EMNLP, 2021. 2 +[11] Nico Engel, Vasileios Belagiannis, and Klaus C. J. Dietmayer. Point Transformer. IEEE, 9:134826-134840, 2020. 2 +[12] Benjamin Graham, Martin Engelcke, and Laurens van der Maaten. 3D Semantic Segmentation with Submanifold Sparse Convolutional Networks. CVPR, pages 9224-9232, 2018. 2 +[13] Lei Han, Tian Zheng, Lan Xu, and Lu Fang. OccuSeg: Occupancy-Aware 3D Instance Segmentation. CVPR, pages 2937-2946, 2020. 2 +[14] Judy Hoffman, Eric Tzeng, Trevor Darrell, and Kate Saenko. Simultaneous Deep Transfer Across Domains and Tasks. ICCV, pages 4068-4076, 2015. 3 +[15] Lingdong Kong, You-Chen Liu, Runnan Chen, Yuexin Ma, Xinge Zhu, Yikang Li, Yuenan Hou, Y. Qiao, and Ziwei Liu. Rethinking Range View Representation for LiDAR Segmentation. ArXiv, abs/2303.05367, 2023. 2 +[16] Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, and Jiaya Jia. Stratified Trans- + +former for 3D Point Cloud Segmentation. CVPR, pages 8490-8499, 2022. 2, 6 +[17] Xin Lai, Yukang Chen, Fanbin Lu, Jianhui Liu, and Jiaya Jia. Spherical Transformer for LiDAR-Based 3D Recognition. CVPR, pages 17545-17555, 2023. 2, 6 +[18] John Lambert, Zhuang Liu, Ozan Sener, James Hays, and Vladlen Koltun. MSeg: A Composite Dataset for Multi-Domain Semantic Segmentation. PAMI, 45:796-810, 2021. 2 +[19] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. PointCNN: Convolution On X-Transformed Points. In NIPS, 2018. 2 +[20] Yingwei Li, Adams Wei Yu, Tianjian Meng, Benjamin Caine, Jiquan Ngiam, Daiyi Peng, Junyang Shen, Bo-Xun Wu, Yifeng Lu, Denny Zhou, Quoc V. Le, Alan Loddon Yuille, and Mingxing Tan. DeepFusion: Lidar-Camera Deep Fusion for Multi-Modal 3D Object Detection. CVPR, pages 17161-17170, 2022. 1 +[21] Xingyu Lin, Harjatin Singh Baweja, George A. Kantor, and David Held. Adaptive Auxiliary Task Weighting for Reinforcement Learning. In NIPS, 2019. 3 +[22] Zhi-Hao Lin, Sheng Yu Huang, and Y. Wang. Convolution in the Cloud: Learning Deformable Kernels in 3D Graph Convolution Networks for Point Cloud Analysis. CVPR, pages 1797-1806, 2020. 2 +[23] Xianpeng Liu, Nan Xue, and Tianfu Wu. Learning Auxiliary Monocular Contexts Helps Monocular 3D Object Detection. In AAAI, 2021. 3 +[24] Zhijian Liu, Haotian Tang, Yujun Lin, and Song Han. PointVoxel CNN for Efficient 3D Deep Learning. In NIPS, 2019. 2, 5 +[25] Daniel Maturana and Sebastian A. Scherer. VoxNet: A 3D Convolutional Neural Network for real-time object recognition. IROS, pages 922-928, 2015. 2 +[26] Andres Milioto, Ignacio Vizzo, Jens Behley, and C. Stachniss. RangeNet ++: Fast and Accurate LiDAR Semantic Segmentation. IROS, pages 4213-4220, 2019. 2 +[27] C. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. Point-Net: Deep Learning on Point Sets for 3D Classification and Segmentation. CVPR, pages 77–85, 2017. 2 +[28] C. Qi, L. Yi, Hao Su, and Leonidas J. Guibas. PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space. In NIPS, 2017. 2, 8 +[29] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. PointNeXt: Revisiting PointNet++ with Improved Training and Scaling Strategies. In NIPS, 2022. 2, 6 +[30] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards Robust Monocular Depth Estimation: Mixing Datasets for Zero-Shot Cross-Dataset Transfer. PAMI, 44:1623-1637, 2019. 2 +[31] Gernot Riegler, Ali O. Ulusoy, and Andreas Geiger. Oct-Net: Learning Deep 3D Representations at High Resolutions. CVPR, pages 6620-6629, 2017. 2 +[32] Leslie N. Smith and Nicholay Topin. Super-Convergence: Very Fast Training of Residual Networks Using Large Learning Rates. ArXiv, abs/1708.07120, 2017. 5 + +[33] Shuran Song, Samuel P. Lichtenberg, and Jianxiong Xiao. SUN RGB-D: A RGB-D scene understanding benchmark suite. CVPR, pages 567-576, 2015. 1 +[34] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott M. Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in Perception for Autonomous Driving: Waymo Open Dataset. CVPR, pages 2443-2451, 2019. 1 +[35] Haotian* Tang, Zhijian* Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching Efficient 3D Architectures with Sparse Point-Voxel Convolution. In ECCV, 2020. 2, 5, 6, 8 +[36] Hugues Thomas, C. Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J. Guibas. KPConv: Flexible and Deformable Convolution for Point Clouds. ICCV, pages 6410–6419, 2019. 2 +[37] Li Wang, Dong Wei Li, Han Liu, Jinzhan Peng, Lu Tian, and Yi Shan. Cross-Dataset Collaborative Learning for Semantic Segmentation. AAAI, 2021. 2, 3 +[38] Xudong Wang, Zhaowei Cai, Dashan Gao, and Nuno Vasconcelos. Towards Universal Object Detection by Domain Attention. CVPR, pages 7281-7290, 2019. 2 +[39] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E. Sarma, Michael M. Bronstein, and Justin M. Solomon. Dynamic Graph CNN for Learning on Point Clouds. TOG, 38:1-12, 2018. 2 +[40] Christian Wojek, Stefan Walk, Stefan Roth, and Bernt Schiele. Monocular 3D scene understanding with explicit occlusion reasoning. CVPR, pages 1993-2000, 2011. 1 +[41] Bichen Wu, Alvin Wan, Xiangyu Yue, and Kurt Keutzer. SqueezeSeg: Convolutional Neural Nets with Recurrent CRF for Real-Time Road-Object Segmentation from 3D Li-DAR Point Cloud. ICRA, pages 1887–1893, 2018. 2 +[42] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point Transformer V2: Grouped Vector Attention and Partition-based Pooling. In NIPS, 2022. 2, 8 +[43] Xiaoyang Wu, Zhuotao Tian, Xin Wen, Bohao Peng, Xihui Liu, Kaicheng Yu, and Hengshuang Zhao. Towards Large-scale 3D Representation Learning with Multi-dataset Point Prompt Training. ArXiv, abs/2308.09718, 2023. 2, 3, 5, 6, 8 +[44] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3D ShapeNets: A deep representation for volumetric shapes. CVPR, pages 1912-1920, 2015. 6 +[45] Yichen Xie, Chenfeng Xu, Marie-Julie Rakotosaona, Patrick Rim, Federico Tombari, Kurt Keutzer, Masayoshi Tomizuka, and Wei Zhan. SparseFusion: Fusing Multi-Modal Sparse Representations for Multi-Sensor 3D Object Detection. ICCV, abs/2304.14340, 2023. 1 +[46] Guodong Xu, Ziwei Liu, Xiaoxiao Li, and Chen Change Loy. Knowledge Distillation Meets Self-Supervision. In ECCV, 2020. 3 +[47] Hang Xu, Linpu Fang, Xiaodan Liang, Wenxiong Kang, and Zhenguo Li. Universal-RCNN: Universal Object Detector via Transferable Graph R-CNN. In AAAI, 2020. 2 + +[48] Mutian Xu, Runyu Ding, Hengshuang Zhao, and Xiaojuan Qi. PACnv: Position Adaptive Convolution with Dynamic Kernel Assembling on Point Clouds. CVPR, pages 3172-3181, 2021. 2 +[49] Yu-Qi Yang, Yu-Xiao Guo, Jiangfeng Xiong, Yang Liu, Hao Pan, Peng-Shuai Wang, Xin Tong, and Baining Guo. Swin3D: A Pretrained Transformer Backbone for 3D Indoor Scene Understanding. ArXiv, abs/2304.06906, 2023. 2 +[50] Zetong Yang, Yanan Sun, Shu Liu, Xiaoyong Shen, and Jiaya Jia. STD: Sparse-to-Dense 3D Object Detector for Point Cloud. ICCV, pages 1951-1960, 2019. 2 +[51] Tianwei Yin, Xingyi Zhou, and Philipp Kähenbuhl. Center-based 3D Object Detection and Tracking. CVPR, pages 11779-11788, 2021. 1 +[52] Cheng Zhang, Haocheng Wan, Xinyi Shen, and Zizhao Wu. PVT: Point-voxel transformer for point cloud learning. *IJIS*, 37:11985 - 12008, 2021. 2, 5 +[53] Yang Zhang, Zixiang Zhou, Philip David, Xiangyu Yue, Zerong Xi, and Hassan Foroosh. PolarNet: An Improved Grid Representation for Online LiDAR Point Clouds Semantic Segmentation. CVPR, pages 9598-9607, 2020. 2 +[54] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip Torr, and Vladlen Koltun. Point Transformer. In ICCV, 2021. 2, 6, 8 +[55] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Simple Multi-dataset Detection. CVPR, pages 7561-7570, 2021. 2 +[56] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and Asymmetrical 3D Convolution Networks for LiDAR Segmentation. CVPR, pages 9934-9943, 2021. 2, 6 +[57] Özgün Çiçek, Ahmed Abdulkadir, Soeren S. Lienkamp, Thomas Brox, and Olaf Ronneberger. 3D U-Net: Learning Dense Volumetric Segmentation from Sparse Annotation. In MICCAI, 2016. 2 \ No newline at end of file diff --git a/2024/GeoAuxNet_ Towards Universal 3D Representation Learning for Multi-sensor Point Clouds/images.zip b/2024/GeoAuxNet_ Towards Universal 3D Representation Learning for Multi-sensor Point Clouds/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..93b8d284696bad39d8921c5fd81544161acb61aa --- /dev/null +++ b/2024/GeoAuxNet_ Towards Universal 3D Representation Learning for Multi-sensor Point Clouds/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:064aa544a4bd5fcd1a93a40449812bbee55adcb231814bb200f3524a79ee6e91 +size 562598 diff --git a/2024/GeoAuxNet_ Towards Universal 3D Representation Learning for Multi-sensor Point Clouds/layout.json b/2024/GeoAuxNet_ Towards Universal 3D Representation Learning for Multi-sensor Point Clouds/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..854a50243d095b663dd4709fef1f7d7bbf3580c8 --- /dev/null +++ b/2024/GeoAuxNet_ Towards Universal 3D Representation Learning for Multi-sensor Point Clouds/layout.json @@ -0,0 +1,10214 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 58, + 102, + 536, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 102, + 536, + 138 + ], + "spans": [ + { + "bbox": [ + 58, + 102, + 536, + 138 + ], + "type": "text", + "content": "GeoAuxNet: Towards Universal 3D Representation Learning for Multi-sensor Point Clouds" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 194, + 160, + 399, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 160, + 399, + 175 + ], + "spans": [ + { + "bbox": [ + 194, + 160, + 399, + 175 + ], + "type": "text", + "content": "Shengjun Zhang" + }, + { + "bbox": [ + 194, + 160, + 399, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 194, + 160, + 399, + 175 + ], + "type": "text", + "content": ", Xin Fei" + }, + { + "bbox": [ + 194, + 160, + 399, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 194, + 160, + 399, + 175 + ], + "type": "text", + "content": ", Yueqi Duan" + }, + { + "bbox": [ + 194, + 160, + 399, + 175 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 149, + 175, + 445, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 175, + 445, + 190 + ], + "spans": [ + { + "bbox": [ + 149, + 175, + 445, + 190 + ], + "type": "text", + "content": "1 Department of Electronic Engineering, Tsinghua University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 175, + 189, + 418, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 189, + 418, + 204 + ], + "spans": [ + { + "bbox": [ + 175, + 189, + 418, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 175, + 189, + 418, + 204 + ], + "type": "text", + "content": " Department of Automation, Tsinghua University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 112, + 205, + 478, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 205, + 478, + 217 + ], + "spans": [ + { + "bbox": [ + 112, + 205, + 478, + 217 + ], + "type": "text", + "content": "{zhangsj23, feix21}@mails.tsinghua.edu.cn, duanyueqi@tsinghua.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 157, + 224, + 436, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 224, + 436, + 237 + ], + "spans": [ + { + "bbox": [ + 157, + 224, + 436, + 237 + ], + "type": "text", + "content": "https://github.com/zhangshengjun2019/GeoAuxNet" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 59, + 270, + 89, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 270, + 89, + 291 + ], + "spans": [ + { + "bbox": [ + 59, + 270, + 89, + 291 + ], + "type": "text", + "content": "S3DIS RGB-D" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 326, + 93, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 326, + 93, + 346 + ], + "spans": [ + { + "bbox": [ + 54, + 326, + 93, + 346 + ], + "type": "text", + "content": "ScanNet RGB-D" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 94, + 243, + 168, + 308 + ], + "blocks": [ + { + "bbox": [ + 94, + 243, + 168, + 308 + ], + "lines": [ + { + "bbox": [ + 94, + 243, + 168, + 308 + ], + "spans": [ + { + "bbox": [ + 94, + 243, + 168, + 308 + ], + "type": "image", + "image_path": "f9fa253af30bdc16ecfba1378f7a3ddd9777ca60233a5fec21806f2f77f2721c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 94, + 308, + 168, + 363 + ], + "blocks": [ + { + "bbox": [ + 94, + 308, + 168, + 363 + ], + "lines": [ + { + "bbox": [ + 94, + 308, + 168, + 363 + ], + "spans": [ + { + "bbox": [ + 94, + 308, + 168, + 363 + ], + "type": "image", + "image_path": "035c1fc49a6ab46e0a0ec9be69ccabbf12b2f3e8730b1d4b0c3b72c947edc90c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 169, + 243, + 244, + 308 + ], + "blocks": [ + { + "bbox": [ + 169, + 243, + 244, + 308 + ], + "lines": [ + { + "bbox": [ + 169, + 243, + 244, + 308 + ], + "spans": [ + { + "bbox": [ + 169, + 243, + 244, + 308 + ], + "type": "image", + "image_path": "1b6ecc3f2bddbef315988f6667e68e67d220548f9ed23eba17ec40434e367129.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 169, + 308, + 243, + 363 + ], + "blocks": [ + { + "bbox": [ + 169, + 308, + 243, + 363 + ], + "lines": [ + { + "bbox": [ + 169, + 308, + 243, + 363 + ], + "spans": [ + { + "bbox": [ + 169, + 308, + 243, + 363 + ], + "type": "image", + "image_path": "0eff91df56662770d7c04c32fc45ff088170b7e0d937420805c3787e023b1bb0.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 244, + 243, + 318, + 308 + ], + "blocks": [ + { + "bbox": [ + 244, + 243, + 318, + 308 + ], + "lines": [ + { + "bbox": [ + 244, + 243, + 318, + 308 + ], + "spans": [ + { + "bbox": [ + 244, + 243, + 318, + 308 + ], + "type": "image", + "image_path": "4288e65a0a4405c27bf51de3bea7abbb0e6ef3cd4a35ea28e825a66c4f3df4c6.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 244, + 308, + 318, + 363 + ], + "blocks": [ + { + "bbox": [ + 244, + 308, + 318, + 363 + ], + "lines": [ + { + "bbox": [ + 244, + 308, + 318, + 363 + ], + "spans": [ + { + "bbox": [ + 244, + 308, + 318, + 363 + ], + "type": "image", + "image_path": "621246fc57fdf4ad71411102708842118b71049542d1d8bb3901955463b1ff09.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 319, + 243, + 393, + 308 + ], + "blocks": [ + { + "bbox": [ + 319, + 243, + 393, + 308 + ], + "lines": [ + { + "bbox": [ + 319, + 243, + 393, + 308 + ], + "spans": [ + { + "bbox": [ + 319, + 243, + 393, + 308 + ], + "type": "image", + "image_path": "ebf1348708d946905c1f90c3095b06f98e15f3af7902ee1c44789c6f78333c6b.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 319, + 308, + 393, + 363 + ], + "blocks": [ + { + "bbox": [ + 319, + 308, + 393, + 363 + ], + "lines": [ + { + "bbox": [ + 319, + 308, + 393, + 363 + ], + "spans": [ + { + "bbox": [ + 319, + 308, + 393, + 363 + ], + "type": "image", + "image_path": "65d0de5372e84dc7f045cb588678bf106d5e6b38859d33b7b6415ff7585fbb21.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 393, + 243, + 468, + 308 + ], + "blocks": [ + { + "bbox": [ + 393, + 243, + 468, + 308 + ], + "lines": [ + { + "bbox": [ + 393, + 243, + 468, + 308 + ], + "spans": [ + { + "bbox": [ + 393, + 243, + 468, + 308 + ], + "type": "image", + "image_path": "1775f9fac4575a54898da131783cf35497e2db0284522493a08ce95d1b0f55df.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 393, + 308, + 468, + 363 + ], + "blocks": [ + { + "bbox": [ + 393, + 308, + 468, + 363 + ], + "lines": [ + { + "bbox": [ + 393, + 308, + 468, + 363 + ], + "spans": [ + { + "bbox": [ + 393, + 308, + 468, + 363 + ], + "type": "image", + "image_path": "2aa5d9a94727c261f3587f3d4d6c07f539c54f92b00cbd98b3f99cd156bb6440.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 468, + 243, + 543, + 308 + ], + "blocks": [ + { + "bbox": [ + 468, + 243, + 543, + 308 + ], + "lines": [ + { + "bbox": [ + 468, + 243, + 543, + 308 + ], + "spans": [ + { + "bbox": [ + 468, + 243, + 543, + 308 + ], + "type": "image", + "image_path": "d2f7654ed92c096ba16bdc162a4848135cde7d76f667090f2b437b1751bc32c3.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 468, + 308, + 544, + 363 + ], + "blocks": [ + { + "bbox": [ + 468, + 308, + 544, + 363 + ], + "lines": [ + { + "bbox": [ + 468, + 308, + 544, + 363 + ], + "spans": [ + { + "bbox": [ + 468, + 308, + 544, + 363 + ], + "type": "image", + "image_path": "256874226c954a1d8958a648d2b8fd3cfad654fa70f96541367e39458f901402.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 53, + 381, + 94, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 381, + 94, + 414 + ], + "spans": [ + { + "bbox": [ + 53, + 381, + 94, + 414 + ], + "type": "text", + "content": "Semantic KITTI LiDAR" + } + ] + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 94, + 364, + 168, + 428 + ], + "blocks": [ + { + "bbox": [ + 94, + 364, + 168, + 428 + ], + "lines": [ + { + "bbox": [ + 94, + 364, + 168, + 428 + ], + "spans": [ + { + "bbox": [ + 94, + 364, + 168, + 428 + ], + "type": "image", + "image_path": "69eef06fd72fbe39980ced831df56c9a5d39d7a6255f89f8b1225fe5e5942f78.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 121, + 430, + 145, + 441 + ], + "lines": [ + { + "bbox": [ + 121, + 430, + 145, + 441 + ], + "spans": [ + { + "bbox": [ + 121, + 430, + 145, + 441 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 169, + 364, + 243, + 428 + ], + "blocks": [ + { + "bbox": [ + 169, + 364, + 243, + 428 + ], + "lines": [ + { + "bbox": [ + 169, + 364, + 243, + 428 + ], + "spans": [ + { + "bbox": [ + 169, + 364, + 243, + 428 + ], + "type": "image", + "image_path": "77bb9faf76bbb8dce0e0ce68d26f103f03b94b1792ef833934bf2ffc86307ada.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 173, + 430, + 236, + 440 + ], + "lines": [ + { + "bbox": [ + 173, + 430, + 236, + 440 + ], + "spans": [ + { + "bbox": [ + 173, + 430, + 236, + 440 + ], + "type": "text", + "content": "MinkowskiNet" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 243, + 364, + 318, + 428 + ], + "blocks": [ + { + "bbox": [ + 243, + 364, + 318, + 428 + ], + "lines": [ + { + "bbox": [ + 243, + 364, + 318, + 428 + ], + "spans": [ + { + "bbox": [ + 243, + 364, + 318, + 428 + ], + "type": "image", + "image_path": "4a7d5b77ae2634146c7d64f9b78bf61e9005f0c302bbf4ebcdfd1ea6bd1350f1.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 260, + 430, + 302, + 440 + ], + "lines": [ + { + "bbox": [ + 260, + 430, + 302, + 440 + ], + "spans": [ + { + "bbox": [ + 260, + 430, + 302, + 440 + ], + "type": "text", + "content": "SPVCNN" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 451, + 546, + 474 + ], + "lines": [ + { + "bbox": [ + 46, + 451, + 546, + 474 + ], + "spans": [ + { + "bbox": [ + 46, + 451, + 546, + 474 + ], + "type": "text", + "content": "Figure 1. Semantic segmentation results on S3DIS [1] and ScanNet [9] from RGB-D cameras and SemanticKITTI [2] from LiDAR. For all methods, we trained collectively on three datasets. Our method outperforms other methods with better detailed structures." + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 318, + 364, + 393, + 428 + ], + "blocks": [ + { + "bbox": [ + 318, + 364, + 393, + 428 + ], + "lines": [ + { + "bbox": [ + 318, + 364, + 393, + 428 + ], + "spans": [ + { + "bbox": [ + 318, + 364, + 393, + 428 + ], + "type": "image", + "image_path": "8fcd188bf04145dd4b48723f0e5e564e0a22d61eeb83f3eed88dad17d206e0c7.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 346, + 430, + 365, + 440 + ], + "lines": [ + { + "bbox": [ + 346, + 430, + 365, + 440 + ], + "spans": [ + { + "bbox": [ + 346, + 430, + 365, + 440 + ], + "type": "text", + "content": "PPT" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 393, + 364, + 468, + 428 + ], + "blocks": [ + { + "bbox": [ + 393, + 364, + 468, + 428 + ], + "lines": [ + { + "bbox": [ + 393, + 364, + 468, + 428 + ], + "spans": [ + { + "bbox": [ + 393, + 364, + 468, + 428 + ], + "type": "image", + "image_path": "458920fd3ebf2201c5a4080515d84a96c25ee7e09a2ca70f61a0c96cf29e7642.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 419, + 430, + 441, + 440 + ], + "lines": [ + { + "bbox": [ + 419, + 430, + 441, + 440 + ], + "spans": [ + { + "bbox": [ + 419, + 430, + 441, + 440 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 468, + 364, + 544, + 428 + ], + "blocks": [ + { + "bbox": [ + 468, + 364, + 544, + 428 + ], + "lines": [ + { + "bbox": [ + 468, + 364, + 544, + 428 + ], + "spans": [ + { + "bbox": [ + 468, + 364, + 544, + 428 + ], + "type": "image", + "image_path": "ae239ccb8a8bb17d5d16ca105420545cc08c4a3557f86fcb16669813cd372c6f.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 477, + 430, + 533, + 440 + ], + "lines": [ + { + "bbox": [ + 477, + 430, + 533, + 440 + ], + "spans": [ + { + "bbox": [ + 477, + 430, + 533, + 440 + ], + "type": "text", + "content": "GroundTruth" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "bbox": [ + 143, + 490, + 192, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 490, + 192, + 502 + ], + "spans": [ + { + "bbox": [ + 143, + 490, + 192, + 502 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 45, + 515, + 288, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 515, + 288, + 695 + ], + "spans": [ + { + "bbox": [ + 45, + 515, + 288, + 695 + ], + "type": "text", + "content": "Point clouds captured by different sensors such as RGB-D cameras and LiDAR possess non-negligible domain gaps. Most existing methods design different network architectures and train separately on point clouds from various sensors. Typically, point-based methods achieve outstanding performances on even-distributed dense point clouds from RGB-D cameras, while voxel-based methods are more efficient for large-range sparse LiDAR point clouds. In this paper, we propose geometry-to-voxel auxiliary learning to enable voxel representations to access point-level geometric information, which supports better generalisation of the voxel-based backbone with additional interpretations of multi-sensor point clouds. Specifically, we construct hierarchical geometry pools generated by a voxel-guided dynamic point network, which efficiently provide" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 306, + 491, + 545, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 491, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 306, + 491, + 545, + 575 + ], + "type": "text", + "content": "auxiliary fine-grained geometric information adapted to different stages of voxel features. We conduct experiments on joint multi-sensor datasets to demonstrate the effectiveness of GeoAuxNet. Enjoying elaborate geometric information, our method outperforms other models collectively trained on multi-sensor datasets, and achieve competitive results with the-state-of-art experts on each single dataset." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 306, + 598, + 387, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 598, + 387, + 610 + ], + "spans": [ + { + "bbox": [ + 306, + 598, + 387, + 610 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 304, + 618, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 715 + ], + "type": "text", + "content": "Point cloud analysis has attracted widespread attention due to its growing applications, such as autonomous driving [20, 45, 51] and robotics [4, 40]. Unlike images which are represented by regular pixels, 3D point clouds are irregular and unordered. These characteristics aggravate the inconsistency on density and sampling patterns of point clouds captured by different sensors, such as RGB-D cameras [1, 9, 33] and LiDAR [2, 5, 34]." + } + ] + } + ], + "index": 40 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 703, + 139, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 703, + 139, + 714 + ], + "spans": [ + { + "bbox": [ + 62, + 703, + 139, + 714 + ], + "type": "text", + "content": "†Corresponding author." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20019" + } + ] + } + ], + "index": 42 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 49, + 72, + 286, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 72, + 286, + 346 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 286, + 346 + ], + "type": "text", + "content": "Typically, points generated from RGB-D pictures are equally distributed and dense, while points scanned by LiDAR are sparse and uneven. The diversity on input data hinders the construction of universal network architectures. For point clouds from RGB-D cameras, point-based methods [28, 29, 54] usually gather local information via grouping and aggregating neighborhood features to extract detailed spatial information. Since neighbors are not stored contiguously in point representations, indexing them requires the costly searching operations [24], which limits the application of these methods on large scale LiDAR point clouds. To address this challenge, voxel-based methods [12, 25, 31] leverage the regular 3D convolution to obtain the contiguous memory access pattern. However, the resolution of voxels is constrained by the memory and becomes lower at deeper stages, resulting in loss of local information [35]. An intuitive idea is to ensemble both point-based and voxel-based methods [24, 35, 52], yet they are still confronted with either high time-consumption or the lack of detailed geometric features, which fail to simultaneously process large scale point clouds from different sensors with fine-grained spatial information while maintaining efficiency." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 351, + 286, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 351, + 286, + 601 + ], + "spans": [ + { + "bbox": [ + 49, + 351, + 286, + 601 + ], + "type": "text", + "content": "In this paper, we propose GeoAuxNet to provide point-level geometric information for voxel representations in an efficient manner with geometry-to-voxel auxiliary learning. The support of elaborate geometric features introduces additional point-level information, which cannot be fully exploited by voxel-based backbones. Motivated by the observation that the local geometric structures present high similarity at each stage of the network, we construct hierarchical geometry pools to learn auxiliary sensor-aware point-level geometric priors corresponding to different layers for sensor-agnostic voxel features. To update our geometry pools, we present a voxel-guided dynamic point network, where we leverage prior knowledge in voxel features to guide high quality spatial feature extraction. Then, we fuse elaborate geometric features in the pools into voxel representations via our designed geometry-to-voxel auxiliary mechanism. For large-scale point clouds, the hierarchical geometry pools store representative point-level features at different stages, which efficiently provide complementary geometric information without using the point-based networks during inference time." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 605, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 605, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 605, + 286, + 712 + ], + "type": "text", + "content": "We conduct extensive experiments on multi-sensor datasets, including S3DIS [1] and ScanNet [9] from RGB-D cameras as well as SemanticKITTI [2] captured by LiDAR, to demonstrate the efficiency and effectiveness of our method. With a shared weight of backbone, our method outperforms other models trained on joint datasets from different sensors and achieves competitive performance with experts on single dataset. As shown in Figure 1, we trained MinkowskiNet [8], SPVCNN [35], PPT [43] and" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 308, + 73, + 545, + 106 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 308, + 73, + 545, + 106 + ], + "type": "text", + "content": "GeoAuxNet on the above three datasets collectively, where our method preserve better detailed structures for point clouds from various sensors." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 308, + 118, + 395, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 395, + 129 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 395, + 129 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 308, + 137, + 545, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 137, + 545, + 531 + ], + "spans": [ + { + "bbox": [ + 308, + 137, + 545, + 531 + ], + "type": "text", + "content": "3D Scene Understanding. Since images are composed of regular pixels, point clouds generated from RGB-D cameras are well-distributed and dense. Voxelization on dense point clouds ignore detailed structure information. To directly learn from points, Qi et al. [27] first proposed PointNet to process point clouds through shared multi-layer perceptions (MLP) to get a global representation. Yet this method lacks the ability to comprehend local geometric features. Therefore, PointNet++ [28] built a hierarchical architecture to gather local information via grouping and aggregating neighborhood features. Following up MLP-based works [7, 29] focused on designing effective learning schemes to capture spatial features. Other methods also introduced specific convolution operations [3, 19, 36, 48], graph-based representations [22, 39] and self-attention mechanism [11, 16, 42, 49, 54] to explore the relationship between points. Apart from cameras, LiDAR is also a widely used sensor to collect point clouds. Some methods [15, 26, 41, 53] focused on projecting point clouds to 2D grids to utilize 2D CNN. However, 3D to 2D projection limits the representation quality of geometric structures [50] due to the sacrifice of 3D spatial information [48]. Point based methods have high time consumption because of the large scale and the sparsity of these point clouds. Voxel-based methods [12, 13, 17, 56, 57] show more robustness against the sparsity. In view of the sampling characteristic of LiDAR, Zhu et al. [56] introduced cylinder coordinate system, for nearby regions have greater density than farther away regions. Similarly, Lai et al. [17] proposed spherical transformer to aggregate information from dense close points to sparse distant ones. Voxel-based networks maintain efficiency when processing scene-level point clouds, but ignore some detailed spatial structures." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 308, + 546, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 546, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 546, + 545, + 712 + ], + "type": "text", + "content": "Universal Learning from Multiple Data Sources. Large language models achieve remarkable progress on many natural language processing systems through pretraining on extremely large text corpora [10], which enables foundation models to process data across domains, involving multiple languages, various majors, diverse application scenarios and so on. In 2D computer vision, joint learning across domains improve model robustness for detection [38, 47, 55], depth estimation [30] and semantic segmentation [18, 37]. Wang et al. [38] proposed a universal object detector through a domain-attention mechanism. But they did not model the training differences between different datasets. Universal-RCNN [47] modeled the class relations by designing a inter-dataset attention module to in" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "type": "text", + "content": "20020" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 69, + 545, + 268 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 545, + 268 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 545, + 268 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 545, + 268 + ], + "type": "image", + "image_path": "4e9e97d7b4d8c135ecae23d28cd421a78973969a9acfd51b27665b821a3fdea0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 275, + 548, + 374 + ], + "lines": [ + { + "bbox": [ + 46, + 275, + 548, + 374 + ], + "spans": [ + { + "bbox": [ + 46, + 275, + 548, + 374 + ], + "type": "text", + "content": "Figure 2. The pipeline of our GeoAuxNet. For a complete scene point cloud " + }, + { + "bbox": [ + 46, + 275, + 548, + 374 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{\\mathcal{C}}" + }, + { + "bbox": [ + 46, + 275, + 548, + 374 + ], + "type": "text", + "content": " and a local point patch " + }, + { + "bbox": [ + 46, + 275, + 548, + 374 + ], + "type": "inline_equation", + "content": "\\mathcal{P} \\subseteq \\mathcal{P}^{\\mathcal{C}}" + }, + { + "bbox": [ + 46, + 275, + 548, + 374 + ], + "type": "text", + "content": ", our voxel-based backbone first voxelizes " + }, + { + "bbox": [ + 46, + 275, + 548, + 374 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{\\mathcal{C}}" + }, + { + "bbox": [ + 46, + 275, + 548, + 374 + ], + "type": "text", + "content": " and conducts sparse convolutional operations. The voxel-guided hypernetwork takes relative positions, voxel features and a stage latent code as input to provide weights and biases for the point network. Then, we encode the spatial information for " + }, + { + "bbox": [ + 46, + 275, + 548, + 374 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 275, + 548, + 374 + ], + "type": "text", + "content": " with the point network and aggregate local features to generate geometric feature candidates. Following the update strategy, we construct hierarchical geometry pools. The geometry-to-voxel mechanism fuses geometric features stored in the pools to enable voxel representations to access point-level geometric information. We repeat the above process several times to extract effective representation hierarchically and predict the results with a voxel decoder for the primary task and a point decoder for the auxiliary task. The dotted line stands for the course of the auxiliary learning which is ignored during inference to ensure efficiency. Geo-to-Vox is abbreviation of Geometry-to-Voxel." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 382, + 289, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 382, + 289, + 525 + ], + "spans": [ + { + "bbox": [ + 46, + 382, + 289, + 525 + ], + "type": "text", + "content": "corporates graph transfer learning for propagating relevant semantic information across multiple datasets. CDCL [37] pointed out that convolution filters can be shared across domains without accuracy loss, while normalization operators are not appropriate to share in view of the bias of statistics. Recently, Wu et al. [43] introduced point prompt training to optimize a learnable domain-specific prompt with language-guided categorical alignment, but processed point clouds from RGB-D and LiDAR separately. However, point clouds captured by different sensors have diverse density and distributions, which limits the exploration of the universal network on multiple data sources." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": "Auxiliary Learning. Auxiliary learning aims to achieve outstanding performance on a single primary task with the assist of auxiliary tasks and has shown many successful applications in a wide range of fields, including knowledge distillation [46], transfer learning [14] and reinforcement learning [21]. For instance, DCAN [6] learned an auxiliary contour detection task for more accurate segmentation, while Araslanov investigated the joint learning of image classification and semantic segmentation. In 3D vision, MonoCon [23] introduced monocular contexts as auxiliary tasks in training for 3D object detection. Nevertheless, the inherent relationship between point-level features and voxel-level features is still under exploration to better leverage the advantages of two domains without any requirement" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 382, + 382, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 382, + 382, + 393 + ], + "spans": [ + { + "bbox": [ + 306, + 382, + 382, + 393 + ], + "type": "text", + "content": "of additional data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 404, + 362, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 404, + 362, + 417 + ], + "spans": [ + { + "bbox": [ + 306, + 404, + 362, + 417 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 424, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 424, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 424, + 545, + 533 + ], + "type": "text", + "content": "The overall framework of GeoAuxNet is illustrated in Figure 2. To generate point-level geometric features, we propose a voxel-guided hypernetwork to produce weights and biases for the point network. Then, point features are grouped and aggregated to extract fine-grained local features, which are employed to update geometry pools. For voxel-based backbone, we introduce geometry-to-voxel auxiliary mechanism to access elaborate spatial information in our hierarchical geometry pools." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 540, + 506, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 540, + 506, + 552 + ], + "spans": [ + { + "bbox": [ + 305, + 540, + 506, + 552 + ], + "type": "text", + "content": "3.1. Voxel-guided Dynamic Point Network" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 558, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 652 + ], + "type": "text", + "content": "Point-based methods are mostly confronted with high time consumption towards large scale point clouds, since the farthest point sampling (FPS) and k-nearest neighbors (kNN) algorithms in these networks have a time complexity of " + }, + { + "bbox": [ + 304, + 558, + 545, + 652 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(n^2)" + }, + { + "bbox": [ + 304, + 558, + 545, + 652 + ], + "type": "text", + "content": ". Therefore, to efficiently process large scale point clouds, input points are voxelized and passed through sparse convolution operations to extract voxel features in our voxel-based backbone." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": "To generate point-level geometric information efficiently, we train our point network on local point patches instead of complete scene point clouds. However, less training data weakens the ability of the point network to extract local features. Considering that our voxel-based branch are" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "20021" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": "trained on complete voxelized point clouds, we treat voxel features as prior knowledge to instruct the learning of local geometric information. Yet, points inside a voxel share the same voxel feature, which ignores detailed difference between points. Thus, we also introduce relative coordinates to depict diversity of points within the same voxel. On account of the hierarchical architectures of both the point network and the voxel-based backbone with several stages, we optimize a learnable latent code " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": " for each stage during our auxiliary learning." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 191, + 287, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 191, + 287, + 252 + ], + "spans": [ + { + "bbox": [ + 47, + 191, + 287, + 252 + ], + "type": "text", + "content": "Given a local point patch " + }, + { + "bbox": [ + 47, + 191, + 287, + 252 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\{p_i, 1 \\leq i \\leq N\\} \\in \\mathbb{R}^{N \\times 3}" + }, + { + "bbox": [ + 47, + 191, + 287, + 252 + ], + "type": "text", + "content": " with point features " + }, + { + "bbox": [ + 47, + 191, + 287, + 252 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{\\mathcal{P}} = \\{f_i^{\\mathcal{P}}, 1 \\leq i \\leq N\\}" + }, + { + "bbox": [ + 47, + 191, + 287, + 252 + ], + "type": "text", + "content": ", each point " + }, + { + "bbox": [ + 47, + 191, + 287, + 252 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 47, + 191, + 287, + 252 + ], + "type": "text", + "content": " belongs to a voxel " + }, + { + "bbox": [ + 47, + 191, + 287, + 252 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 47, + 191, + 287, + 252 + ], + "type": "text", + "content": " with the voxel feature " + }, + { + "bbox": [ + 47, + 191, + 287, + 252 + ], + "type": "inline_equation", + "content": "f^{\\mathcal{V}}" + }, + { + "bbox": [ + 47, + 191, + 287, + 252 + ], + "type": "text", + "content": ". Here " + }, + { + "bbox": [ + 47, + 191, + 287, + 252 + ], + "type": "inline_equation", + "content": "v \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 47, + 191, + 287, + 252 + ], + "type": "text", + "content": " stands for the center coordinate of the voxel. Then, we can obtain the weight and bias for point " + }, + { + "bbox": [ + 47, + 191, + 287, + 252 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 47, + 191, + 287, + 252 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 94, + 258, + 287, + 273 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 258, + 287, + 273 + ], + "spans": [ + { + "bbox": [ + 94, + 258, + 287, + 273 + ], + "type": "interline_equation", + "content": "w _ {i} = h _ {w} \\left(h _ {p} \\left(p _ {i} - v\\right), f ^ {\\mathcal {V}}\\right) \\odot h _ {s} (s), \\tag {1}", + "image_path": "60e14e95d2716c7526401f5e1cd86682892b74198ebe1d24d776672f82b647e5.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 97, + 275, + 287, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 275, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 97, + 275, + 287, + 289 + ], + "type": "interline_equation", + "content": "b _ {i} = h _ {b} \\left(h _ {p} \\left(p _ {i} - v\\right), f ^ {\\mathcal {V}}\\right) \\odot h _ {s} (s), \\tag {2}", + "image_path": "530ef9f269e3ad9fa882acd907babe0610936724c1bdc3d151690d373d2eefeb.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 297, + 287, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 297, + 287, + 344 + ], + "spans": [ + { + "bbox": [ + 47, + 297, + 287, + 344 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 297, + 287, + 344 + ], + "type": "inline_equation", + "content": "h_p" + }, + { + "bbox": [ + 47, + 297, + 287, + 344 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 297, + 287, + 344 + ], + "type": "inline_equation", + "content": "h_s" + }, + { + "bbox": [ + 47, + 297, + 287, + 344 + ], + "type": "text", + "content": " are MLPs to project relative positions and the stage latent code to specific dimensions, and " + }, + { + "bbox": [ + 47, + 297, + 287, + 344 + ], + "type": "inline_equation", + "content": "h_w" + }, + { + "bbox": [ + 47, + 297, + 287, + 344 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 297, + 287, + 344 + ], + "type": "inline_equation", + "content": "h_b" + }, + { + "bbox": [ + 47, + 297, + 287, + 344 + ], + "type": "text", + "content": " are hypernetworks to generate the weight and bias. Hence, new point features can be formulated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 109, + 350, + 287, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 350, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 109, + 350, + 287, + 365 + ], + "type": "interline_equation", + "content": "\\hat {f} _ {i} ^ {\\mathcal {P}} = \\operatorname {M L P} \\left(w _ {i} \\odot f _ {i} ^ {\\mathcal {P}} + b _ {i}\\right). \\tag {3}", + "image_path": "14c198d17940e563b0a0191cc70374be11b4d2c872d3b1ba24af2bdc6d9a86dd.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 373, + 287, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 373, + 287, + 433 + ], + "spans": [ + { + "bbox": [ + 47, + 373, + 287, + 433 + ], + "type": "text", + "content": "To extract local geometric information, we use kNN algorithm to select " + }, + { + "bbox": [ + 47, + 373, + 287, + 433 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 373, + 287, + 433 + ], + "type": "text", + "content": " nearest neighbor points and employ a max-pooling function to aggregate each local group. In this manner, we generate a set of geometric feature candidates for hierarchical geometry pools." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 439, + 206, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 439, + 206, + 452 + ], + "spans": [ + { + "bbox": [ + 47, + 439, + 206, + 452 + ], + "type": "text", + "content": "3.2. Hierarchical Geometry Pools" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 457, + 287, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 457, + 287, + 565 + ], + "spans": [ + { + "bbox": [ + 47, + 457, + 287, + 565 + ], + "type": "text", + "content": "The voxel branch ignores some detailed spatial structures due to voxelization. Motivated by the similarity of geometric structures in 3D space, we establish a pool to store typical geometric patterns. The pool contains " + }, + { + "bbox": [ + 47, + 457, + 287, + 565 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 47, + 457, + 287, + 565 + ], + "type": "text", + "content": " typical spatial structure features " + }, + { + "bbox": [ + 47, + 457, + 287, + 565 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{\\mathcal{D}} = \\{f_i^{\\mathcal{D}}, 1 \\leq i \\leq n\\}" + }, + { + "bbox": [ + 47, + 457, + 287, + 565 + ], + "type": "text", + "content": ". The max size of the pool is set to " + }, + { + "bbox": [ + 47, + 457, + 287, + 565 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 457, + 287, + 565 + ], + "type": "text", + "content": ". In Section 3.1, we generate a set of geometric pattern candidates " + }, + { + "bbox": [ + 47, + 457, + 287, + 565 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^S = \\{f_i^S, 1 \\leq i \\leq m\\}" + }, + { + "bbox": [ + 47, + 457, + 287, + 565 + ], + "type": "text", + "content": ". When updating the pool, we first calculate the similarity between " + }, + { + "bbox": [ + 47, + 457, + 287, + 565 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^S" + }, + { + "bbox": [ + 47, + 457, + 287, + 565 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 457, + 287, + 565 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^D" + }, + { + "bbox": [ + 47, + 457, + 287, + 565 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 574, + 287, + 592 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 574, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 132, + 574, + 287, + 592 + ], + "type": "interline_equation", + "content": "s _ {i} = \\max _ {1 \\leq j \\leq N} s _ {i, j}, \\tag {4}", + "image_path": "24e9084a84011a74e902029ec2447e3079b8c1809836799ec28cbee2da8431ae.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 599, + 75, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 75, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 75, + 609 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 127, + 607, + 287, + 637 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 607, + 287, + 637 + ], + "spans": [ + { + "bbox": [ + 127, + 607, + 287, + 637 + ], + "type": "interline_equation", + "content": "s _ {i, j} = \\frac {f _ {i} ^ {\\mathcal {S}} \\cdot f _ {j} ^ {\\mathcal {D}}}{\\| f _ {i} ^ {\\mathcal {S}} \\| \\| f _ {j} ^ {\\mathcal {D}} \\|.} \\tag {5}", + "image_path": "a73ae85f976252f28c3e7e08a04764954a61fb36317df0455212d83f0b4f4314.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 640, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 640, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 47, + 640, + 287, + 664 + ], + "type": "text", + "content": "Specifically, if " + }, + { + "bbox": [ + 47, + 640, + 287, + 664 + ], + "type": "inline_equation", + "content": "s_i \\geq \\epsilon" + }, + { + "bbox": [ + 47, + 640, + 287, + 664 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 640, + 287, + 664 + ], + "type": "inline_equation", + "content": "f_i^S" + }, + { + "bbox": [ + 47, + 640, + 287, + 664 + ], + "type": "text", + "content": " is merged into the geometry pool " + }, + { + "bbox": [ + 47, + 640, + 287, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{\\mathcal{D}}" + }, + { + "bbox": [ + 47, + 640, + 287, + 664 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 123, + 674, + 287, + 693 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 674, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 123, + 674, + 287, + 693 + ], + "type": "interline_equation", + "content": "t = \\underset {1 \\leq j \\leq N} {\\operatorname {a r g m a x}} s _ {i, j}, \\tag {6}", + "image_path": "93ad7b3aec69af510fbccad8c4eac5ebc2daadc3dfb95bfee555741b934ae497.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 696, + 287, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 696, + 287, + 711 + ], + "spans": [ + { + "bbox": [ + 115, + 696, + 287, + 711 + ], + "type": "interline_equation", + "content": "\\hat {f} _ {t} ^ {\\mathcal {D}} = \\lambda f _ {t} ^ {\\mathcal {D}} + (1 - \\lambda) f _ {i} ^ {\\mathcal {S}}, \\tag {7}", + "image_path": "228e569815ea55727966b5c8f0671403e18a0767866f62f61e7274fbd92bdb72.jpg" + } + ] + } + ], + "index": 14 + }, + { + "type": "code", + "bbox": [ + 306, + 87, + 545, + 353 + ], + "blocks": [ + { + "bbox": [ + 306, + 72, + 417, + 84 + ], + "lines": [ + { + "bbox": [ + 306, + 72, + 417, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 417, + 84 + ], + "type": "text", + "content": "Algorithm 1 Update Pools" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 306, + 87, + 545, + 353 + ], + "lines": [ + { + "bbox": [ + 306, + 87, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 306, + 87, + 545, + 353 + ], + "type": "text", + "content": "Input: New features " + }, + { + "bbox": [ + 306, + 87, + 545, + 353 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^S" + }, + { + "bbox": [ + 306, + 87, + 545, + 353 + ], + "type": "text", + "content": ", geometry pool " + }, + { + "bbox": [ + 306, + 87, + 545, + 353 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^D" + }, + { + "bbox": [ + 306, + 87, + 545, + 353 + ], + "type": "text", + "content": ", maximum size of the pool " + }, + { + "bbox": [ + 306, + 87, + 545, + 353 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 306, + 87, + 545, + 353 + ], + "type": "text", + "content": ", threshold " + }, + { + "bbox": [ + 306, + 87, + 545, + 353 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 306, + 87, + 545, + 353 + ], + "type": "text", + "content": ", update rate " + }, + { + "bbox": [ + 306, + 87, + 545, + 353 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 306, + 87, + 545, + 353 + ], + "type": "text", + "content": " \nOutput: New pool " + }, + { + "bbox": [ + 306, + 87, + 545, + 353 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{F}}^D" + }, + { + "bbox": [ + 306, + 87, + 545, + 353 + ], + "type": "text", + "content": " \nCompute " + }, + { + "bbox": [ + 306, + 87, + 545, + 353 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 306, + 87, + 545, + 353 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 306, + 87, + 545, + 353 + ], + "type": "inline_equation", + "content": "1 \\leq i \\leq m" + }, + { + "bbox": [ + 306, + 87, + 545, + 353 + ], + "type": "text", + "content": " according to Eq. 4 \nfor " + }, + { + "bbox": [ + 306, + 87, + 545, + 353 + ], + "type": "inline_equation", + "content": "i \\gets 1" + }, + { + "bbox": [ + 306, + 87, + 545, + 353 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 306, + 87, + 545, + 353 + ], + "type": "inline_equation", + "content": "|\\mathcal{F}^S|" + }, + { + "bbox": [ + 306, + 87, + 545, + 353 + ], + "type": "text", + "content": " do \n if " + }, + { + "bbox": [ + 306, + 87, + 545, + 353 + ], + "type": "inline_equation", + "content": "s_i \\geq \\epsilon" + }, + { + "bbox": [ + 306, + 87, + 545, + 353 + ], + "type": "text", + "content": " then \n t ← arg max" + }, + { + "bbox": [ + 306, + 87, + 545, + 353 + ], + "type": "inline_equation", + "content": "_{1 \\leq j \\leq N}" + }, + { + "bbox": [ + 306, + 87, + 545, + 353 + ], + "type": "inline_equation", + "content": "s_i, j" + }, + { + "bbox": [ + 306, + 87, + 545, + 353 + ], + "type": "inline_equation", + "content": "f_t^D \\gets \\lambda f_t^D + (1 - \\lambda)f_i^S" + }, + { + "bbox": [ + 306, + 87, + 545, + 353 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^S \\gets \\mathcal{F}^S \\setminus \\{f_i^S\\}" + }, + { + "bbox": [ + 306, + 87, + 545, + 353 + ], + "type": "text", + "content": " \n end if \nend for \nif " + }, + { + "bbox": [ + 306, + 87, + 545, + 353 + ], + "type": "inline_equation", + "content": "|\\mathcal{F}^S| + |\\mathcal{F}^D| \\leq N" + }, + { + "bbox": [ + 306, + 87, + 545, + 353 + ], + "type": "text", + "content": " then \n " + }, + { + "bbox": [ + 306, + 87, + 545, + 353 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^D \\gets \\mathcal{F}^D \\cup \\mathcal{F}^S" + }, + { + "bbox": [ + 306, + 87, + 545, + 353 + ], + "type": "text", + "content": " \nelse \n " + }, + { + "bbox": [ + 306, + 87, + 545, + 353 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{F}}^S \\gets" + }, + { + "bbox": [ + 306, + 87, + 545, + 353 + ], + "type": "text", + "content": " randomly select " + }, + { + "bbox": [ + 306, + 87, + 545, + 353 + ], + "type": "inline_equation", + "content": "N - |\\mathcal{F}^D|" + }, + { + "bbox": [ + 306, + 87, + 545, + 353 + ], + "type": "text", + "content": " features from " + }, + { + "bbox": [ + 306, + 87, + 545, + 353 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^S" + }, + { + "bbox": [ + 306, + 87, + 545, + 353 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^S \\gets \\mathcal{F}^S \\setminus \\hat{\\mathcal{F}}^S" + }, + { + "bbox": [ + 306, + 87, + 545, + 353 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^D \\gets \\mathcal{F}^D \\cup \\hat{\\mathcal{F}}^S" + }, + { + "bbox": [ + 306, + 87, + 545, + 353 + ], + "type": "text", + "content": " \n for " + }, + { + "bbox": [ + 306, + 87, + 545, + 353 + ], + "type": "inline_equation", + "content": "i \\gets 1" + }, + { + "bbox": [ + 306, + 87, + 545, + 353 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 306, + 87, + 545, + 353 + ], + "type": "inline_equation", + "content": "|\\mathcal{F}^S|" + }, + { + "bbox": [ + 306, + 87, + 545, + 353 + ], + "type": "text", + "content": " do \n t ← arg max" + }, + { + "bbox": [ + 306, + 87, + 545, + 353 + ], + "type": "inline_equation", + "content": "_{1 \\leq j \\leq N}" + }, + { + "bbox": [ + 306, + 87, + 545, + 353 + ], + "type": "inline_equation", + "content": "s_i, j" + }, + { + "bbox": [ + 306, + 87, + 545, + 353 + ], + "type": "inline_equation", + "content": "f_t^D \\gets \\lambda f_t^D + (1 - \\lambda)f_i^S" + }, + { + "bbox": [ + 306, + 87, + 545, + 353 + ], + "type": "text", + "content": " \n end for \nend if" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "code_body" + } + ], + "index": 16, + "sub_type": "algorithm" + }, + { + "bbox": [ + 304, + 376, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 545, + 449 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 376, + 545, + 449 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 376, + 545, + 449 + ], + "type": "text", + "content": " is a threshold and " + }, + { + "bbox": [ + 304, + 376, + 545, + 449 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 376, + 545, + 449 + ], + "type": "text", + "content": " stands for the update rate. Otherwise, we append the new features to the geometry pool. If the total number of features is more than " + }, + { + "bbox": [ + 304, + 376, + 545, + 449 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 376, + 545, + 449 + ], + "type": "text", + "content": ", we randomly select " + }, + { + "bbox": [ + 304, + 376, + 545, + 449 + ], + "type": "inline_equation", + "content": "N - n" + }, + { + "bbox": [ + 304, + 376, + 545, + 449 + ], + "type": "text", + "content": " new features and add them to the pool. The rest new features following Eq. 7 are merged into existing patterns. Our update strategy is illustrated in Algorithm 1." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 449, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 521 + ], + "type": "text", + "content": "Since voxel representations at various stages contains different level information, we introduce hierarchical geometry pools to store representative point-level features corresponding to different stages. We also enlarge the size of pools for deeper stages due to more complex geometric structures with larger receptive field sizes." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 533, + 463, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 533, + 463, + 545 + ], + "spans": [ + { + "bbox": [ + 306, + 533, + 463, + 545 + ], + "type": "text", + "content": "3.3. Geometry-to-Voxel Auxiliary" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 552, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 552, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 552, + 545, + 624 + ], + "type": "text", + "content": "After we update the geometric pool, we use geometric patterns stored in " + }, + { + "bbox": [ + 304, + 552, + 545, + 624 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{\\mathcal{D}}\\in \\mathbb{R}^{N\\times C}" + }, + { + "bbox": [ + 304, + 552, + 545, + 624 + ], + "type": "text", + "content": " to generate fine-grained local information for each voxel following a cross-attention operation. Given voxel features " + }, + { + "bbox": [ + 304, + 552, + 545, + 624 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{\\nu} = \\{f_i^\\nu ,1\\leq i\\leq M\\} \\in" + }, + { + "bbox": [ + 304, + 552, + 545, + 624 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{M\\times D}" + }, + { + "bbox": [ + 304, + 552, + 545, + 624 + ], + "type": "text", + "content": ", the geometric fine-grained feature " + }, + { + "bbox": [ + 304, + 552, + 545, + 624 + ], + "type": "inline_equation", + "content": "f_{i}^{\\mathcal{G}}" + }, + { + "bbox": [ + 304, + 552, + 545, + 624 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 304, + 552, + 545, + 624 + ], + "type": "inline_equation", + "content": "f_{i}^{\\nu}\\in \\mathcal{F}^{\\nu}" + }, + { + "bbox": [ + 304, + 552, + 545, + 624 + ], + "type": "text", + "content": " is formulated by:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 351, + 635, + 545, + 649 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 635, + 545, + 649 + ], + "spans": [ + { + "bbox": [ + 351, + 635, + 545, + 649 + ], + "type": "interline_equation", + "content": "q _ {i} = h _ {Q} \\left(f _ {i} ^ {\\mathcal {V}}\\right) \\in \\mathbb {R} ^ {1 \\times s}, \\tag {8}", + "image_path": "b006ae1df5c9d2f5312526a3ef7719f60f8700a1e1f113b63ad8ba8ff4e50507.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 351, + 651, + 545, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 651, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 351, + 651, + 545, + 666 + ], + "type": "interline_equation", + "content": "K = h _ {K} \\left(\\mathcal {F} ^ {\\mathcal {D}}\\right) \\in \\mathbb {R} ^ {N \\times s}, \\tag {9}", + "image_path": "2a15ef87c56d14bb0a7f5c9d5e7382c6f3b7e81310bbccfb889e5deebe5919aa.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 352, + 667, + 545, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 667, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 352, + 667, + 545, + 681 + ], + "type": "interline_equation", + "content": "V = h _ {V} \\left(\\mathcal {F} ^ {\\mathcal {D}}\\right) \\in \\mathbb {R} ^ {N \\times t}, \\tag {10}", + "image_path": "d9f956f0ce475d07e56f017aa45de08244d903444e613d1f3966ab90a2bbc4ec.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 348, + 684, + 545, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 684, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 348, + 684, + 545, + 712 + ], + "type": "interline_equation", + "content": "f _ {i} ^ {\\mathcal {G}} = \\operatorname {s o f t m a x} \\left(\\frac {q _ {i} K ^ {\\top}}{\\sqrt {s}}\\right) V \\in \\mathbb {R} ^ {1 \\times t}, \\tag {11}", + "image_path": "3a50837420e9b6ebb67209727ca2318e975c8cb527cf6757140f697a441e3731.jpg" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20022" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "h_{Q}, h_{K}" + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "h_{V}" + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": " stand for projections. Then, we concatenate " + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "f_{i}^{\\mathcal{G}}" + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "f_{i}^{\\mathcal{V}}" + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": " as the new voxel feature and employ sparse convolution operations. In this manner, we can provide complementary geometric information without using the point-based networks during inference stage." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 143, + 167, + 154 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 143, + 167, + 154 + ], + "spans": [ + { + "bbox": [ + 47, + 143, + 167, + 154 + ], + "type": "text", + "content": "3.4. Overall Architecture" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 162, + 287, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 162, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 162, + 287, + 258 + ], + "type": "text", + "content": "Given a point cloud " + }, + { + "bbox": [ + 46, + 162, + 287, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i^{\\mathcal{C}}" + }, + { + "bbox": [ + 46, + 162, + 287, + 258 + ], + "type": "text", + "content": " with features " + }, + { + "bbox": [ + 46, + 162, + 287, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_i^{\\mathcal{C}}" + }, + { + "bbox": [ + 46, + 162, + 287, + 258 + ], + "type": "text", + "content": " and a local point patch " + }, + { + "bbox": [ + 46, + 162, + 287, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i\\subseteq \\mathcal{P}_i^{\\mathcal{C}}" + }, + { + "bbox": [ + 46, + 162, + 287, + 258 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 162, + 287, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_i^{\\mathcal{P}}\\subseteq \\mathcal{F}_i^{\\mathcal{C}}" + }, + { + "bbox": [ + 46, + 162, + 287, + 258 + ], + "type": "text", + "content": " from sensor " + }, + { + "bbox": [ + 46, + 162, + 287, + 258 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 46, + 162, + 287, + 258 + ], + "type": "text", + "content": ", we voxelize all points to obtain voxel coordinates " + }, + { + "bbox": [ + 46, + 162, + 287, + 258 + ], + "type": "inline_equation", + "content": "\\nu_{i}" + }, + { + "bbox": [ + 46, + 162, + 287, + 258 + ], + "type": "text", + "content": " with voxel features " + }, + { + "bbox": [ + 46, + 162, + 287, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_i^\\nu" + }, + { + "bbox": [ + 46, + 162, + 287, + 258 + ], + "type": "text", + "content": ". Points obtained from RGB-D cameras usually possess colors and normal vectors, while points captured by LiDAR always possess intensity. Thus, we introduce sensor-aware input embedding. Specifically, we embed " + }, + { + "bbox": [ + 46, + 162, + 287, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 46, + 162, + 287, + 258 + ], + "type": "text", + "content": " from sensor " + }, + { + "bbox": [ + 46, + 162, + 287, + 258 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 46, + 162, + 287, + 258 + ], + "type": "text", + "content": " by:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 268, + 287, + 283 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 268, + 287, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 268, + 287, + 283 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {F}} _ {i} ^ {\\mathcal {P}} = h _ {\\mathcal {S} _ {i}} ^ {\\mathcal {P}} \\left(\\mathcal {F} _ {i} ^ {P}\\right), \\tag {12}", + "image_path": "8cc7377aacae2c853a43eaa998dfe3031afd30da81e844f8e07a412913761341.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 285, + 287, + 300 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 287, + 300 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {F}} _ {i} ^ {\\mathcal {V}} = h _ {\\mathcal {S} _ {i}} ^ {\\mathcal {V}} \\left(\\mathcal {F} _ {i} ^ {V}\\right), \\tag {13}", + "image_path": "26874e3992b90a11963fdfb61aa3902948649cf195505b7dbf79416c0e271e83.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 311, + 287, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 311, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 46, + 311, + 287, + 419 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 311, + 287, + 419 + ], + "type": "inline_equation", + "content": "h_{\\mathcal{S}_i}^{\\mathcal{P}}" + }, + { + "bbox": [ + 46, + 311, + 287, + 419 + ], + "type": "text", + "content": " means shared MLPs on points and " + }, + { + "bbox": [ + 46, + 311, + 287, + 419 + ], + "type": "inline_equation", + "content": "h_{\\mathcal{S}_i}^{\\mathcal{V}}" + }, + { + "bbox": [ + 46, + 311, + 287, + 419 + ], + "type": "text", + "content": " means sparse convolution operations on voxels. Then, we learn voxel features and point features following the extraction strategy in Section 3.1 and update hierarchical geometry pools using Algorithm 1. We fuse point-level geometric information to voxel representation via geometry-to-voxel auxiliary mechanism. Our auxiliary task is to predict point labels for local point patches with point branch and point decoder. Therefore, the final loss function is formulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 431, + 287, + 445 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 431, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 133, + 431, + 287, + 445 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {v} + \\mu \\mathcal {L} _ {p}, \\tag {14}", + "image_path": "0379838dabb247f854ef47ac8b18ef3c7e7bb99163f2d000c1a1cd40b69ae344.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 456, + 287, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 287, + 480 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 456, + 287, + 480 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_v" + }, + { + "bbox": [ + 46, + 456, + 287, + 480 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 456, + 287, + 480 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_p" + }, + { + "bbox": [ + 46, + 456, + 287, + 480 + ], + "type": "text", + "content": " stand for per-voxel and per-point cross entropy loss and " + }, + { + "bbox": [ + 46, + 456, + 287, + 480 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 456, + 287, + 480 + ], + "type": "text", + "content": " is the weight of the point loss." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 491, + 119, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 491, + 119, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 491, + 119, + 502 + ], + "type": "text", + "content": "3.5. Discussion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "type": "text", + "content": "In this section, we highlight previous point-voxel networks [24, 35, 52] which combine a point-based branch and a voxel-based branch together to leverage the advantages of both efficient voxel representations and elaborate point features. Additionally, we compare our GeoAuxNet with these methods. PVCNN [24] and PVT [52] rely on point-based networks to extract high quality representations with time-consuming searching or attention mechanisms, while SPVCNN [35] does not explore detailed local geometric information. Since the interaction between two branches in these methods only contains voxelization and devoxelization operations, they fail to benefit from both voxel representations and point-level elaborate features. Our GeoAuxNet preserves hierarchical sensor-aware geometry pools as the bridge to enable voxel representations to absorb point-level local geometric information without using the point network during inference." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 72, + 388, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 388, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 388, + 85 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 91, + 545, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 545, + 212 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 545, + 212 + ], + "type": "text", + "content": "We validate our proposed GeoAuxNet across point clouds from multiple sensors. In Section 4.1, we train various methods on joint datasets including S3DIS [1], ScanNet [9] and SemanticKITTI [2]. In Section 4.2, we analyze the effectiveness of our sensor-aware hierarchical geometry pools. In Section 4.3, we compare GeoAuxNet with typical point-based and voxel-based methods to demonstrate the efficiency of our designed geometry-to-voxel auxiliary learning. In Section 4.4, we ablate different design choices of our voxel-guided dynamic point network." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 221, + 440, + 234 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 221, + 440, + 234 + ], + "spans": [ + { + "bbox": [ + 305, + 221, + 440, + 234 + ], + "type": "text", + "content": "4.1. Semantic Segmentation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 240, + 545, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 240, + 545, + 444 + ], + "spans": [ + { + "bbox": [ + 304, + 240, + 545, + 444 + ], + "type": "text", + "content": "Dataset. We conduct semantic segmentation experiments on the joint data of three datasets: S3DIS [1] and ScanNet [9], which are generated from RGB-D cameras, and SemanticKITTI [2], which is captured by LiDAR. The S3DIS dataset comprises 271 rooms from six areas in three distinct buildings, which are densely sampled on the surfaces of the meshes and annotated into 13 categories. Model performance evaluation is typically done using results from Area 5. The ScanNet dataset comprises 1,613 scene scans reconstructed from RGB-D frames. It is divided into 1,201 scenes for training, 312 scenes for validation, and 100 scenes for benchmark testing. SemanticKITTI consists of 22 point cloud sequences, where sequences 00 to 10, 08 and 11 to 21 are used for training, validation and testing. After merging classes with distinct moving status and discarding classes with very few points, a total of 19 classes are selected for training and evaluation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 462, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 546, + 714 + ], + "type": "text", + "content": "Implementation Details. Our voxel-based backbone is built on MinkowskiNet [8] with the batch normalization operations proposed by PPT [43]. The voxel size is set to " + }, + { + "bbox": [ + 304, + 462, + 546, + 714 + ], + "type": "inline_equation", + "content": "0.05\\mathrm{m}" + }, + { + "bbox": [ + 304, + 462, + 546, + 714 + ], + "type": "text", + "content": ". We train our methods as well as MinkowskiNet SPVCNN [35] and PPT [43] on joint input data [1, 2, 9] captured by RGB-D cameras and LiDAR collectively. We also fine tune our method on three datasets separately. For the sake of fairness, we set the number of parameters of MinkowskiNet, SPVCNN, PPT and GeoAuxNet to about 60M. We use SGD optimizer and OneCycleLR [32] scheduler, with learning rate 0.05 and weight decay " + }, + { + "bbox": [ + 304, + 462, + 546, + 714 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 304, + 462, + 546, + 714 + ], + "type": "text", + "content": ". We adopt OneCycleLR [32] scheduler with " + }, + { + "bbox": [ + 304, + 462, + 546, + 714 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 304, + 462, + 546, + 714 + ], + "type": "text", + "content": " percentage of the cycle spent increasing the learning rate and a cosine annealing strategy. We train all models with data augmentation used in PPT [43]. We set " + }, + { + "bbox": [ + 304, + 462, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 462, + 546, + 714 + ], + "type": "text", + "content": " in Eq. 7 to 0.1, " + }, + { + "bbox": [ + 304, + 462, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 462, + 546, + 714 + ], + "type": "text", + "content": " in Eq. 14 to 0.1, threshold for geometry pool update " + }, + { + "bbox": [ + 304, + 462, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 462, + 546, + 714 + ], + "type": "text", + "content": " to 0.9. The total number of iterations is equal to the sum of necessary iterations for all datasets. The sampling ratio across S3DIS, ScanNet and SemanticKITTI is set to 2:2:5. The number of points for the auxiliary branch does not exceed " + }, + { + "bbox": [ + 304, + 462, + 546, + 714 + ], + "type": "inline_equation", + "content": "2\\times 10^{4}" + }, + { + "bbox": [ + 304, + 462, + 546, + 714 + ], + "type": "text", + "content": ". The max size of hierarchical geometry pools for each stage" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20023" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 123, + 552, + 319 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 553, + 117 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 553, + 117 + ], + "type": "text", + "content": "Table 1. Semantic segmentation results on three benchmarks including S3DIS [1], ScanNet [9] and SemanticKITTI [2]. We train SPVCNN [35], PPT [43] and GeoAuxNet on the joint training data of three datasets and also compare with experts [16, 17, 29, 35, 54, 56] on each single dataset. We report the class-average accuracy (mAcc, %) and class-wise mean IoU (mIoU, %). The type stands for point-based " + }, + { + "bbox": [ + 46, + 70, + 553, + 117 + ], + "type": "inline_equation", + "content": "(p)" + }, + { + "bbox": [ + 46, + 70, + 553, + 117 + ], + "type": "text", + "content": " or voxel-based " + }, + { + "bbox": [ + 46, + 70, + 553, + 117 + ], + "type": "inline_equation", + "content": "(\\nu)" + }, + { + "bbox": [ + 46, + 70, + 553, + 117 + ], + "type": "text", + "content": " methods, where " + }, + { + "bbox": [ + 46, + 70, + 553, + 117 + ], + "type": "inline_equation", + "content": "p \\rightarrow \\nu" + }, + { + "bbox": [ + 46, + 70, + 553, + 117 + ], + "type": "text", + "content": " means the geometry-to-voxel auxiliary learning." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 123, + 552, + 319 + ], + "lines": [ + { + "bbox": [ + 47, + 123, + 552, + 319 + ], + "spans": [ + { + "bbox": [ + 47, + 123, + 552, + 319 + ], + "type": "table", + "html": "
MethodTypeParams.S3DIS Area5ScanNetSemanticKITTI
Test mIoUTest mAccVal mIoUVal mAccVal mIoUVal mAcc
PointTransformer [54]p11.4M70.476.570.6-
StratifiedFromer [16]p18.8M72.078.174.3-
PointNeXt [29]p41.6M70.577.271.5-
SPVCNN [35]p+v21.8M63.8-
Cylinder3D [56]v26.1M64.3-
SphereFormer [17]v32.3M67.8-
MinkowskiNet [8]v60.9M18.821.325.630.536.041.2
SPVCNN [35]p+v61.0M58.662.356.760.452.056.7
PPT [43]v63.0M30.434.616.527.425.031.2
PPT †[43]v63.0M67.973.269.677.561.868.9
GeoAuxNetp→v64.7M69.5(+1.6)74.5(+1.3)71.3(+1.7)79.3(+1.8)63.8(+2.0)69.3(+0.4)
", + "image_path": "2d64efa52801232fe321a2f66562aec1e35bda5bda3d747518ccafd06466d245.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 319, + 553, + 343 + ], + "lines": [ + { + "bbox": [ + 55, + 319, + 553, + 343 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 553, + 343 + ], + "type": "text", + "content": "† The original PPT with language-guided categorical alignment fails to converge on the joint training data. We use decoupled projection heads to employ PPT on multi-sensor datasets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 361, + 193, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 361, + 193, + 374 + ], + "spans": [ + { + "bbox": [ + 47, + 361, + 193, + 374 + ], + "type": "text", + "content": "is set to 32, 64, 128, 256 separately." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 390, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 390, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 390, + 289, + 715 + ], + "type": "text", + "content": "Results Comparison. As shown in Table 1, GeoAuxNet surpasses other methods which are trained on three datasets jointly. Our methods outperforms SPVCNN [35], which does not utilize its point branch to extract local information, by more than " + }, + { + "bbox": [ + 46, + 390, + 289, + 715 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 390, + 289, + 715 + ], + "type": "text", + "content": " in mIoU on each dataset. PPT [43] introduces different datasets as prompts to collaboratively train a single model on multiple datasets from the same sensor type. The language-guided alignment of PPT suffers from the large domain gap caused by different sensors, which fails to converge on the joint training datasets. Therefore, we replace it by the decoupled alignment with separate heads for each dataset, which is also defined in PPT [43]. We construct sensor-aware geometry pools to address this issue and outperform PPT by about " + }, + { + "bbox": [ + 46, + 390, + 289, + 715 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 46, + 390, + 289, + 715 + ], + "type": "text", + "content": " in mIoU on each dataset. Besides, we also compare with experts trained on each single dataset. We achieve competitive results with point-based methods on datasets from RGB-D cameras. More specifically, we even outperform PointTransformer in validation mIoU of ScanNet. Despite training on inter-domain datasets from different sensors, GeoAuxNet has a higher accuracy on SemanticKITTI than SPVCNN [35], and attain encouraging performances even compared with recent state-of-the-art expert [17]. Although we do not introduce any specific designs such as pointwise attention mechanism [16, 54] and cylinder coordinate system [56] for sparse LiDAR point clouds, the quantitative results demonstrate the effectiveness of geometry-to" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 361, + 545, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 545, + 399 + ], + "type": "text", + "content": "voxel auxiliary learning by constructing hierarchical geometry pools to provide auxiliary sensor-aware point-level geometric priors for voxel representations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 405, + 465, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 405, + 465, + 418 + ], + "spans": [ + { + "bbox": [ + 305, + 405, + 465, + 418 + ], + "type": "text", + "content": "4.2. Hierarchical Geometry Pools" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 423, + 547, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 547, + 496 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 547, + 496 + ], + "type": "text", + "content": "In this section, we further conduct experiments on our hierarchical geometry pools. First, we analyze the similarity of geometric structures in 3D space. Second, we discuss the appropriation to represent geometric information with our geometry pools. Finally, we study the construction of sensor-aware pools for different sensors." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "text", + "content": "Similarity of Geometric Structures. Local geometric structures possess similarity in 3D space. For example, planar structures exist in desks, chairs and airplanes. These spatial structures are not unrelated but consistent. To further analyze the similarity of geometric structures learned by networks, we train our point encoder on ModelNet40 [44]. For each point, we use kNN to search 16 nearest neighbors and aggregate them via a max-pooling function to generate local geometric features. We select a center point marked as a purple star in Figure 3a and Figure 3c, and calculate the cosine similarity between its features and other point features. Specifically, the wing and the tail of the plane in Figure 3a have similar local structures, resulting in high similarity of features in Figure 3b. Figure 3d and Figure 3f show the same phenomenon at different stages in the network, where planar structures are similar to each other, while different from the chair legs. Since features from deeper lay" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20024" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 71, + 140, + 135 + ], + "blocks": [ + { + "bbox": [ + 69, + 71, + 140, + 135 + ], + "lines": [ + { + "bbox": [ + 69, + 71, + 140, + 135 + ], + "spans": [ + { + "bbox": [ + 69, + 71, + 140, + 135 + ], + "type": "image", + "image_path": "ef135819b164502f5a8716d592ff894fe59ad82d2f7aeed5c4b39a9c5444a93f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 99, + 138, + 110, + 148 + ], + "lines": [ + { + "bbox": [ + 99, + 138, + 110, + 148 + ], + "spans": [ + { + "bbox": [ + 99, + 138, + 110, + 148 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 147, + 71, + 218, + 135 + ], + "blocks": [ + { + "bbox": [ + 147, + 71, + 218, + 135 + ], + "lines": [ + { + "bbox": [ + 147, + 71, + 218, + 135 + ], + "spans": [ + { + "bbox": [ + 147, + 71, + 218, + 135 + ], + "type": "image", + "image_path": "18d61ff77136088f9be03e48200993f8e82bfe4e961130ea9af124df715c51de.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 175, + 138, + 186, + 148 + ], + "lines": [ + { + "bbox": [ + 175, + 138, + 186, + 148 + ], + "spans": [ + { + "bbox": [ + 175, + 138, + 186, + 148 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 222, + 72, + 291, + 135 + ], + "blocks": [ + { + "bbox": [ + 222, + 72, + 291, + 135 + ], + "lines": [ + { + "bbox": [ + 222, + 72, + 291, + 135 + ], + "spans": [ + { + "bbox": [ + 222, + 72, + 291, + 135 + ], + "type": "image", + "image_path": "fae73d3bdaa19de92d09a6aaf24785df40088d025f09e9d392588855884bc009.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 249, + 138, + 260, + 148 + ], + "lines": [ + { + "bbox": [ + 249, + 138, + 260, + 148 + ], + "spans": [ + { + "bbox": [ + 249, + 138, + 260, + 148 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 295, + 72, + 362, + 134 + ], + "blocks": [ + { + "bbox": [ + 295, + 72, + 362, + 134 + ], + "lines": [ + { + "bbox": [ + 295, + 72, + 362, + 134 + ], + "spans": [ + { + "bbox": [ + 295, + 72, + 362, + 134 + ], + "type": "image", + "image_path": "028dbe805c154ec5e6f80049a3b872fe6c1a4f765d613e3622e9e92df32817dd.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 138, + 332, + 148 + ], + "lines": [ + { + "bbox": [ + 321, + 138, + 332, + 148 + ], + "spans": [ + { + "bbox": [ + 321, + 138, + 332, + 148 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 366, + 74, + 435, + 134 + ], + "blocks": [ + { + "bbox": [ + 366, + 74, + 435, + 134 + ], + "lines": [ + { + "bbox": [ + 366, + 74, + 435, + 134 + ], + "spans": [ + { + "bbox": [ + 366, + 74, + 435, + 134 + ], + "type": "image", + "image_path": "031ee3badb36e5675dd67436c80e7f951c71ec29cd4edb11a8b23a770c9f4d63.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 393, + 138, + 404, + 148 + ], + "lines": [ + { + "bbox": [ + 393, + 138, + 404, + 148 + ], + "spans": [ + { + "bbox": [ + 393, + 138, + 404, + 148 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 440, + 72, + 509, + 135 + ], + "blocks": [ + { + "bbox": [ + 440, + 72, + 509, + 135 + ], + "lines": [ + { + "bbox": [ + 440, + 72, + 509, + 135 + ], + "spans": [ + { + "bbox": [ + 440, + 72, + 509, + 135 + ], + "type": "image", + "image_path": "d7112883770caceb463c692f7bb6835effa122c18891681a0aa58d5ace5163f5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 466, + 138, + 476, + 148 + ], + "lines": [ + { + "bbox": [ + 466, + 138, + 476, + 148 + ], + "spans": [ + { + "bbox": [ + 466, + 138, + 476, + 148 + ], + "type": "text", + "content": "(f)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 50, + 221, + 164, + 306 + ], + "blocks": [ + { + "bbox": [ + 46, + 157, + 547, + 202 + ], + "lines": [ + { + "bbox": [ + 46, + 157, + 547, + 202 + ], + "spans": [ + { + "bbox": [ + 46, + 157, + 547, + 202 + ], + "type": "text", + "content": "Figure 3. Visualization of the cosine similarity between features. The purple star is the selected point, and the green star is the point with a significantly similar feature to the purple star. We calculate the cosine similarity between the feature of the red star and other features and visualize them in image (b), (d) and (f). The nearest neighbors of the purple and green stars are marked red in image (a), (c) and (e). (a), (b), (c) and (e) are generated from the first stage of the point network and (e), (f) are from the second stage." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 221, + 164, + 306 + ], + "lines": [ + { + "bbox": [ + 50, + 221, + 164, + 306 + ], + "spans": [ + { + "bbox": [ + 50, + 221, + 164, + 306 + ], + "type": "image", + "image_path": "38ad5ca00e8f04f0cf84398e28444d7880664b102b2882d808bd3fb2b6c093c9.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 172, + 222, + 284, + 306 + ], + "blocks": [ + { + "bbox": [ + 172, + 222, + 284, + 306 + ], + "lines": [ + { + "bbox": [ + 172, + 222, + 284, + 306 + ], + "spans": [ + { + "bbox": [ + 172, + 222, + 284, + 306 + ], + "type": "image", + "image_path": "d13192e61c6877c92ddf92b5d16c9fcec796fc89c388f8642a298cba067f4139.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 205, + 309, + 252, + 319 + ], + "lines": [ + { + "bbox": [ + 205, + 309, + 252, + 319 + ], + "spans": [ + { + "bbox": [ + 205, + 309, + 252, + 319 + ], + "type": "text", + "content": "(b) Stage Two" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 50, + 321, + 162, + 405 + ], + "blocks": [ + { + "bbox": [ + 83, + 309, + 129, + 319 + ], + "lines": [ + { + "bbox": [ + 83, + 309, + 129, + 319 + ], + "spans": [ + { + "bbox": [ + 83, + 309, + 129, + 319 + ], + "type": "text", + "content": "(a) Stage One" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 321, + 162, + 405 + ], + "lines": [ + { + "bbox": [ + 50, + 321, + 162, + 405 + ], + "spans": [ + { + "bbox": [ + 50, + 321, + 162, + 405 + ], + "type": "image", + "image_path": "d0de33688bd881555a4a86954612f05813e3b50aa2e2025a4abd43351de8d607.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 408, + 132, + 418 + ], + "lines": [ + { + "bbox": [ + 80, + 408, + 132, + 418 + ], + "spans": [ + { + "bbox": [ + 80, + 408, + 132, + 418 + ], + "type": "text", + "content": "(c) Stage Three" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 428, + 287, + 462 + ], + "lines": [ + { + "bbox": [ + 46, + 428, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 428, + 287, + 462 + ], + "type": "text", + "content": "Figure 4. Statistical results of cosine similarity between hierarchical geometry pools and point-level features extracted by the point network from point clouds in S3DIS [1] at different stages." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 172, + 321, + 285, + 405 + ], + "blocks": [ + { + "bbox": [ + 172, + 321, + 285, + 405 + ], + "lines": [ + { + "bbox": [ + 172, + 321, + 285, + 405 + ], + "spans": [ + { + "bbox": [ + 172, + 321, + 285, + 405 + ], + "type": "image", + "image_path": "a02125abbb347cdebb504f8b46b709584b0a330802314605e95ba701428174af.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 205, + 408, + 252, + 418 + ], + "lines": [ + { + "bbox": [ + 205, + 408, + 252, + 418 + ], + "spans": [ + { + "bbox": [ + 205, + 408, + 252, + 418 + ], + "type": "text", + "content": "(d) Stage Four" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 346, + 546, + 401 + ], + "lines": [ + { + "bbox": [ + 305, + 346, + 546, + 401 + ], + "spans": [ + { + "bbox": [ + 305, + 346, + 546, + 401 + ], + "type": "text", + "content": "Figure 5. Cosine similarity between geometry pools. (a) shows the similarity between geometry pools of S3DIS [1] and ScanNet [9]. (b) shows the similarity between geometry pools of S3DIS [1] and SemanticKITTI [2]. Intra-sensor geometry pools in (a) have higher similarity than inter-sensor geometry pools in (b)." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 467, + 287, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 467, + 287, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 467, + 287, + 515 + ], + "type": "text", + "content": "ers possess larger receptive field sizes due to grouping and aggregating operations and comprise more geometric information, our hierarchical pools contain more geometric features for deeper layers." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 47, + 529, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 529, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 529, + 287, + 624 + ], + "type": "text", + "content": "Representativeness of Geometry Pools. To demonstrate the representativeness of our pools, we calculate cosine similarity between all features in a scene from S3DIS [1] extracted by the point encoder and the features stored in the pool. For each stage, given the geometry pool " + }, + { + "bbox": [ + 47, + 529, + 287, + 624 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^D = \\{f_i^D, 1 \\leq i \\leq N\\}" + }, + { + "bbox": [ + 47, + 529, + 287, + 624 + ], + "type": "text", + "content": " and all features in a scene " + }, + { + "bbox": [ + 47, + 529, + 287, + 624 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^S = \\{f_i^S, 1 \\leq i \\leq M\\}" + }, + { + "bbox": [ + 47, + 529, + 287, + 624 + ], + "type": "text", + "content": ", the similarity " + }, + { + "bbox": [ + 47, + 529, + 287, + 624 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 47, + 529, + 287, + 624 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 47, + 529, + 287, + 624 + ], + "type": "inline_equation", + "content": "f_i^S" + }, + { + "bbox": [ + 47, + 529, + 287, + 624 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 529, + 287, + 624 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^D" + }, + { + "bbox": [ + 47, + 529, + 287, + 624 + ], + "type": "text", + "content": " is formulated as:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 628, + 287, + 659 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 628, + 287, + 659 + ], + "spans": [ + { + "bbox": [ + 115, + 628, + 287, + 659 + ], + "type": "interline_equation", + "content": "s _ {i} = \\max _ {1 \\leq j \\leq N} \\frac {f _ {i} ^ {S} \\cdot f _ {j} ^ {D}}{\\| f _ {i} ^ {S} \\| \\| f _ {j} ^ {D} \\|}, \\tag {15}", + "image_path": "3e3330a71da90ea96e7bce7e241ec7b0d74dddb3ab7b0f38b7bbb9c316ced5a2.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\|\\cdot\\|" + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": " stands for L2 norm. The distribution of similarities is shown in Figure 4. From the statistical data, most features have cosine similarity higher than 0.85, which indicates that each feature can find a similar feature among our" + } + ] + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 312, + 223, + 413, + 321 + ], + "blocks": [ + { + "bbox": [ + 312, + 223, + 413, + 321 + ], + "lines": [ + { + "bbox": [ + 312, + 223, + 413, + 321 + ], + "spans": [ + { + "bbox": [ + 312, + 223, + 413, + 321 + ], + "type": "image", + "image_path": "2c8e1ebf082ed4023da4b1ab056c0d195f1e69dcc78a63b020cc2bc34f0e79c5.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 355, + 327, + 365, + 335 + ], + "lines": [ + { + "bbox": [ + 355, + 327, + 365, + 335 + ], + "spans": [ + { + "bbox": [ + 355, + 327, + 365, + 335 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 421, + 222, + 522, + 322 + ], + "blocks": [ + { + "bbox": [ + 421, + 222, + 522, + 322 + ], + "lines": [ + { + "bbox": [ + 421, + 222, + 522, + 322 + ], + "spans": [ + { + "bbox": [ + 421, + 222, + 522, + 322 + ], + "type": "image", + "image_path": "d73edf02bb3173414dad3b59e9c165baeb5a9dcb25623c2dabc9cc16f815bdcd.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 464, + 326, + 475, + 335 + ], + "lines": [ + { + "bbox": [ + 464, + 326, + 475, + 335 + ], + "spans": [ + { + "bbox": [ + 464, + 326, + 475, + 335 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 527, + 222, + 544, + 318 + ], + "blocks": [ + { + "bbox": [ + 527, + 222, + 544, + 318 + ], + "lines": [ + { + "bbox": [ + 527, + 222, + 544, + 318 + ], + "spans": [ + { + "bbox": [ + 527, + 222, + 544, + 318 + ], + "type": "image", + "image_path": "73fa5de8b4b2cae4d4f8f2c282e8019f431196bdb9c0e56579430f574b99f75f.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "bbox": [ + 305, + 411, + 546, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 411, + 546, + 496 + ], + "spans": [ + { + "bbox": [ + 305, + 411, + 546, + 496 + ], + "type": "text", + "content": "learned representatives in the pool. In this manner, we avoid using point-based network to process large scale points to generate fine-grained features for each voxel. Our hierarchical pools contain typical geometric features of scenes. Therefore, we can fuse point-level geometry information stored in our pools into voxel representations via geometry-to-voxel auxiliary mechanism efficiently." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 522, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 546, + 714 + ], + "type": "text", + "content": "Sensor-aware Pools for Multiple Sensors. Motivated by the diversity of density and sampling patterns of point clouds from various sensors, we construct different geometry pools to preserve relevant geometric information for diverse sensors. We investigate another two strategies: (i) shared geometry pools across sensors and (ii) different geometry pools for different datasets. We calculate the cosine similarity between features from different pools. As shown in Figure 5, geometry pools from the same sensor possess higher similarity than those from different sensors. As shown in Table 2, training with shared geometry pools across sensors leads to a significant decrease on performance due to the inherent diversity of geometric structures from different sensors, while sharing geometry pools with datasets from the same sensor serves as data augmentation to improve the effectiveness of the geometry pools." + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20025" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 66, + 156, + 267, + 218 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 288, + 148 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 288, + 148 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 288, + 148 + ], + "type": "text", + "content": "Table 2. Ablation study of sensor-aware geometry pools. \"Shared intra-sensor\" means the geometry pools which are shared for point clouds from the same sensor, while different for point clouds from different sensors. \"Shared inter-sensor\" means the geometry pools shared for different sensors. \"Shared intra-dataset\" means the geometry pools are different for different datasets. We report mIoU and mAcc on Area 5 of S3DIS [1]." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 156, + 267, + 218 + ], + "lines": [ + { + "bbox": [ + 66, + 156, + 267, + 218 + ], + "spans": [ + { + "bbox": [ + 66, + 156, + 267, + 218 + ], + "type": "table", + "html": "
MethodsmIoU (%)mAcc (%)
Shared intra-sensor69.574.5
Shared inter-sensor61.4(−8.1)68.3(−6.2)
Shared intra-dataset67.5(−2.0)73.3(−1.2)
", + "image_path": "941e96d74f6f5887dde8c7fa815ce197cebee9ca1b77a6565cc595441c633ea2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 48, + 262, + 293, + 383 + ], + "blocks": [ + { + "bbox": [ + 47, + 231, + 288, + 255 + ], + "lines": [ + { + "bbox": [ + 47, + 231, + 288, + 255 + ], + "spans": [ + { + "bbox": [ + 47, + 231, + 288, + 255 + ], + "type": "text", + "content": "Table 3. Efficiency analysis. We report the inference time and throughput on pre-processed data from S3DIS dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 262, + 293, + 383 + ], + "lines": [ + { + "bbox": [ + 48, + 262, + 293, + 383 + ], + "spans": [ + { + "bbox": [ + 48, + 262, + 293, + 383 + ], + "type": "table", + "html": "
MethodParams.Inference Time (ms)Throughput (ins./sec.)
PointNet++ [28]1.0M378.7108
PT [54]7.8M1079.734
PTv2 [42]3.9M4732.212
MinkowskiNet [8]60.9M45.4898
SPVCNN [35]61.0M48.4682
PPT [43]63.0M80.4251
GeoAuxNet64.7M51.8572
", + "image_path": "6d568c54cba8d6b0b759f7120bbca820e5e74b672d826d88588323cd3873f624.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 392, + 185, + 405 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 392, + 185, + 405 + ], + "spans": [ + { + "bbox": [ + 47, + 392, + 185, + 405 + ], + "type": "text", + "content": "4.3. Efficiency of GeoAuxNet" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 413, + 287, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 413, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 287, + 592 + ], + "type": "text", + "content": "We validate the efficiency of GeoAuxNet compared with other typical point-based [42, 54] and voxel-based methods [8, 43]. We focus on the mean inference time and throughput by employing a standardized pre-processed point cloud with instance labels from S3DIS Area 5 dataset as the input on a single NVIDIA A40 GPU. Initially, the device undergoes a \"warm-up\" phase, where the model is fed with the input point cloud ten times consecutively to eliminate any latency associated with data loading. Subsequently, the model is subjected to an additional 300 iterations of input feeding to calculate mean inference time. To effectively quantify the throughput, we measure the number of instances that a model processes per second. This systematic approach ensures a thorough and reliable assessment of our GeoAuxNet and other state-of-the-art methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": "In Table 3, we show the number of parameters, inference time and throughput for typical point-based and voxel-based networks. Specifically, our GeoAuxNet outperforms the expert model of PointTransformer [54] which is trained on ScanNet [9] individually in mIoU with " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "16.8 \\times" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " throughput and " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "20.8 \\times" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " inference speedup. Even compared with the simple PointNet++ [28] whose parameter size is " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "1/68" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " of ours, GeoAuxNet has less inference time and higher throughput. Besides, our method has similar inference time and throughput to voxel-based methods but achieve" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 324, + 102, + 524, + 175 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "type": "text", + "content": "Table 4. Ablation study of our voxel-guided hypernetwork. We report the mIoU and mAcc on Area 5 of S3DIS [1]." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 324, + 102, + 524, + 175 + ], + "lines": [ + { + "bbox": [ + 324, + 102, + 524, + 175 + ], + "spans": [ + { + "bbox": [ + 324, + 102, + 524, + 175 + ], + "type": "table", + "html": "
MethodsmIoU (%)mAcc (%)
Vanilla63.368.7
+ voxel-guided66.0(+2.7)72.5(+3.8)
+ relative positions66.8(+0.8)73.2(+0.7)
+ stage latent code67.6(+0.8)73.3(+0.1)
", + "image_path": "78244ea50d925776369570c9c7d3a3a42104e1f6a965588568caa149d7292417.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 182, + 547, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 182, + 547, + 254 + ], + "spans": [ + { + "bbox": [ + 304, + 182, + 547, + 254 + ], + "type": "text", + "content": "better performance according to the results in Table 1. GeoAuxNet surpasses PPT [43] on each benchmark about " + }, + { + "bbox": [ + 304, + 182, + 547, + 254 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 304, + 182, + 547, + 254 + ], + "type": "text", + "content": " in mIoU with " + }, + { + "bbox": [ + 304, + 182, + 547, + 254 + ], + "type": "inline_equation", + "content": "2.3\\times" + }, + { + "bbox": [ + 304, + 182, + 547, + 254 + ], + "type": "text", + "content": " throughput and " + }, + { + "bbox": [ + 304, + 182, + 547, + 254 + ], + "type": "inline_equation", + "content": "1.6\\times" + }, + { + "bbox": [ + 304, + 182, + 547, + 254 + ], + "type": "text", + "content": " inference speedup. Our method maintains efficiency for large scale point clouds, while achieves encouraging performances on datasets from multiple sensors." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 262, + 465, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 262, + 465, + 275 + ], + "spans": [ + { + "bbox": [ + 305, + 262, + 465, + 275 + ], + "type": "text", + "content": "4.4. Voxel-guided Hypernetworks" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 281, + 545, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 545, + 460 + ], + "type": "text", + "content": "To investigate the architecture design of our voxel-guided hypernetwork, we conduct ablation studies on S3DIS semantic segmentation. We first introduce a vanilla point network which learns the weight and bias directly. Then, we adopt a hypernetwork taking voxel features as input to generate weights and biases for the point network. We further take account of the relative positions of points. Finally, we optimize learnable stage latent codes to instruct the information of different layers. We train these four models on S3DIS dataset. As shown in Table 4, the exploration of voxel features provides significant guidance for the point network to learn high quality geometric information, leading to an increase in test accuracy by " + }, + { + "bbox": [ + 304, + 281, + 545, + 460 + ], + "type": "inline_equation", + "content": "2.7\\%" + }, + { + "bbox": [ + 304, + 281, + 545, + 460 + ], + "type": "text", + "content": " in mIoU and " + }, + { + "bbox": [ + 304, + 281, + 545, + 460 + ], + "type": "inline_equation", + "content": "3.8\\%" + }, + { + "bbox": [ + 304, + 281, + 545, + 460 + ], + "type": "text", + "content": " in mAcc. Relative positions and stage latent codes also improve the performance of our methods." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 472, + 378, + 485 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 472, + 378, + 485 + ], + "spans": [ + { + "bbox": [ + 305, + 472, + 378, + 485 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 493, + 545, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 493, + 545, + 661 + ], + "spans": [ + { + "bbox": [ + 304, + 493, + 545, + 661 + ], + "type": "text", + "content": "In this paper, we propose GeoAuxNet for multi-sensor point clouds with designed geometry-to-voxel auxiliary learning. We construct hierarchical geometry pools to learn auxiliary sensor-aware point-level geometric priors at different layers for sensor-agnostic voxel features. To generate our geometry pools, we also introduce a voxel-guided dynamic point network to leverage voxel prior knowledge for elaborate point features extraction. Our proposed geometry-to-voxel auxiliary builds a bridge between point-level and voxel-level features in an efficient manner without using the point network during inference time. Experimental results have shown the effectiveness and efficiency of our methods. We hope this work will inspire future research towards universal representation learning for point clouds." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgements. This work was supported by the National Natural Science Foundation of China under Grant 62206147." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20026" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Iro Armeni, Ozan Sener, Amir Roshan Zamir, Helen Jiang, Ioannis K. Brilakis, Martin Fischer, and Silvio Savarese. 3D Semantic Parsing of Large-Scale Indoor Spaces. CVPR, pages 1534-1543, 2016. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 288, + 190 + ], + "type": "text", + "content": "[2] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, C. Stachniss, and Juergen Gall. SemanticKITTI: A Dataset for Semantic Scene Understanding of LiDAR Sequences. ICCV, pages 9296-9306, 2019. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 288, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 288, + 213 + ], + "type": "text", + "content": "[3] Alexandre Boulch. ConvPoint: Continuous convolutions for point cloud processing. CGF, 88:24-34, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 213, + 287, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 213, + 287, + 246 + ], + "spans": [ + { + "bbox": [ + 53, + 213, + 287, + 246 + ], + "type": "text", + "content": "[4] César Cadena, Anthony R. Dick, and Ian D. Reid. Multimodal Auto-Encoders as Joint Estimators for Robotics Scene Understanding. In RSS, 2016. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 247, + 287, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 287, + 300 + ], + "type": "text", + "content": "[5] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yuxin Pan, Giancarlo Baldan, and Oscar Beijbom. nuScenes: A Multimodal Dataset for Autonomous Driving. CVPR, pages 11618-11628, 2019. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 302, + 287, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 287, + 335 + ], + "type": "text", + "content": "[6] Hao Chen, Xiaojuan Qi, Lequan Yu, and Pheng-Ann Heng. DCAN: Deep Contour-Aware Networks for Accurate Gland Segmentation. CVPR, pages 2487-2496, 2016. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 336, + 287, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 336, + 287, + 368 + ], + "spans": [ + { + "bbox": [ + 53, + 336, + 287, + 368 + ], + "type": "text", + "content": "[7] Jaesung Choe, Chunghyun Park, Francois Rameau, Jaesik Park, and In So Kweon. PointMixer: MLP-Mixer for Point Cloud Understanding. ECCV, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 369, + 287, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 287, + 411 + ], + "type": "text", + "content": "[8] Christopher Bongsoo Choy, JunYoung Gwak, and Silvio Savarese. 4D Spatio-Temporal ConvNets: Minkowski Convolutional Neural Networks. CVPR, pages 3070-3079, 2019. 2, 5, 6, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 413, + 287, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 413, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 53, + 413, + 287, + 456 + ], + "type": "text", + "content": "[9] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas A. Funkhouser, and Matthias Nießner. ScanNet: Richly-Annotated 3D Reconstructions of Indoor Scenes. CVPR, pages 2432-2443, 2017. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 458, + 287, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 458, + 287, + 501 + ], + "spans": [ + { + "bbox": [ + 48, + 458, + 287, + 501 + ], + "type": "text", + "content": "[10] Jesse Dodge, Ana Marasovic, Gabriel Ilharco, Dirk Groeneweld, Margaret Mitchell, and Matt Gardner. Documenting Large Webtext Corpora: A Case Study on the Colossal Clean Crawled Corpus. In EMNLP, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 502, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 502, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 502, + 287, + 533 + ], + "type": "text", + "content": "[11] Nico Engel, Vasileios Belagiannis, and Klaus C. J. Dietmayer. Point Transformer. IEEE, 9:134826-134840, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 535, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 535, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 535, + 287, + 578 + ], + "type": "text", + "content": "[12] Benjamin Graham, Martin Engelcke, and Laurens van der Maaten. 3D Semantic Segmentation with Submanifold Sparse Convolutional Networks. CVPR, pages 9224-9232, 2018. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 580, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 287, + 612 + ], + "type": "text", + "content": "[13] Lei Han, Tian Zheng, Lan Xu, and Lu Fang. OccuSeg: Occupancy-Aware 3D Instance Segmentation. CVPR, pages 2937-2946, 2020. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 613, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 646 + ], + "type": "text", + "content": "[14] Judy Hoffman, Eric Tzeng, Trevor Darrell, and Kate Saenko. Simultaneous Deep Transfer Across Domains and Tasks. ICCV, pages 4068-4076, 2015. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[15] Lingdong Kong, You-Chen Liu, Runnan Chen, Yuexin Ma, Xinge Zhu, Yikang Li, Yuenan Hou, Y. Qiao, and Ziwei Liu. Rethinking Range View Representation for LiDAR Segmentation. ArXiv, abs/2303.05367, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[16] Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, and Jiaya Jia. Stratified Trans-" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "former for 3D Point Cloud Segmentation. CVPR, pages 8490-8499, 2022. 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "type": "text", + "content": "[17] Xin Lai, Yukang Chen, Fanbin Lu, Jianhui Liu, and Jiaya Jia. Spherical Transformer for LiDAR-Based 3D Recognition. CVPR, pages 17545-17555, 2023. 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 130, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 130, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 308, + 130, + 545, + 173 + ], + "type": "text", + "content": "[18] John Lambert, Zhuang Liu, Ozan Sener, James Hays, and Vladlen Koltun. MSeg: A Composite Dataset for Multi-Domain Semantic Segmentation. PAMI, 45:796-810, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 175, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 175, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 308, + 175, + 545, + 207 + ], + "type": "text", + "content": "[19] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. PointCNN: Convolution On X-Transformed Points. In NIPS, 2018. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 209, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 274 + ], + "type": "text", + "content": "[20] Yingwei Li, Adams Wei Yu, Tianjian Meng, Benjamin Caine, Jiquan Ngiam, Daiyi Peng, Junyang Shen, Bo-Xun Wu, Yifeng Lu, Denny Zhou, Quoc V. Le, Alan Loddon Yuille, and Mingxing Tan. DeepFusion: Lidar-Camera Deep Fusion for Multi-Modal 3D Object Detection. CVPR, pages 17161-17170, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 275, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 275, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 275, + 545, + 308 + ], + "type": "text", + "content": "[21] Xingyu Lin, Harjatin Singh Baweja, George A. Kantor, and David Held. Adaptive Auxiliary Task Weighting for Reinforcement Learning. In NIPS, 2019. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 309, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 309, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 308, + 309, + 545, + 352 + ], + "type": "text", + "content": "[22] Zhi-Hao Lin, Sheng Yu Huang, and Y. Wang. Convolution in the Cloud: Learning Deformable Kernels in 3D Graph Convolution Networks for Point Cloud Analysis. CVPR, pages 1797-1806, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 354, + 545, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 354, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 308, + 354, + 545, + 386 + ], + "type": "text", + "content": "[23] Xianpeng Liu, Nan Xue, and Tianfu Wu. Learning Auxiliary Monocular Contexts Helps Monocular 3D Object Detection. In AAAI, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 388, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 545, + 420 + ], + "type": "text", + "content": "[24] Zhijian Liu, Haotian Tang, Yujun Lin, and Song Han. PointVoxel CNN for Efficient 3D Deep Learning. In NIPS, 2019. 2, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 422, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 545, + 454 + ], + "type": "text", + "content": "[25] Daniel Maturana and Sebastian A. Scherer. VoxNet: A 3D Convolutional Neural Network for real-time object recognition. IROS, pages 922-928, 2015. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 456, + 545, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 456, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 308, + 456, + 545, + 488 + ], + "type": "text", + "content": "[26] Andres Milioto, Ignacio Vizzo, Jens Behley, and C. Stachniss. RangeNet ++: Fast and Accurate LiDAR Semantic Segmentation. IROS, pages 4213-4220, 2019. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 489, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 489, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 489, + 545, + 521 + ], + "type": "text", + "content": "[27] C. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. Point-Net: Deep Learning on Point Sets for 3D Classification and Segmentation. CVPR, pages 77–85, 2017. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 523, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 523, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 523, + 545, + 555 + ], + "type": "text", + "content": "[28] C. Qi, L. Yi, Hao Su, and Leonidas J. Guibas. PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space. In NIPS, 2017. 2, 8" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 557, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 557, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 308, + 557, + 545, + 601 + ], + "type": "text", + "content": "[29] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. PointNeXt: Revisiting PointNet++ with Improved Training and Scaling Strategies. In NIPS, 2022. 2, 6" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 602, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 602, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 308, + 602, + 545, + 644 + ], + "type": "text", + "content": "[30] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards Robust Monocular Depth Estimation: Mixing Datasets for Zero-Shot Cross-Dataset Transfer. PAMI, 44:1623-1637, 2019. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 647, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 679 + ], + "type": "text", + "content": "[31] Gernot Riegler, Ali O. Ulusoy, and Andreas Geiger. Oct-Net: Learning Deep 3D Representations at High Resolutions. CVPR, pages 6620-6629, 2017. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "text", + "content": "[32] Leslie N. Smith and Nicholay Topin. Super-Convergence: Very Fast Training of Residual Networks Using Large Learning Rates. ArXiv, abs/1708.07120, 2017. 5" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20027" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[33] Shuran Song, Samuel P. Lichtenberg, and Jianxiong Xiao. SUN RGB-D: A RGB-D scene understanding benchmark suite. CVPR, pages 567-576, 2015. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 194 + ], + "type": "text", + "content": "[34] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott M. Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in Perception for Autonomous Driving: Waymo Open Dataset. CVPR, pages 2443-2451, 2019. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 194, + 287, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 194, + 287, + 237 + ], + "spans": [ + { + "bbox": [ + 48, + 194, + 287, + 237 + ], + "type": "text", + "content": "[35] Haotian* Tang, Zhijian* Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching Efficient 3D Architectures with Sparse Point-Voxel Convolution. In ECCV, 2020. 2, 5, 6, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 238, + 287, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 238, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 238, + 287, + 282 + ], + "type": "text", + "content": "[36] Hugues Thomas, C. Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J. Guibas. KPConv: Flexible and Deformable Convolution for Point Clouds. ICCV, pages 6410–6419, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 283, + 287, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 283, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 48, + 283, + 287, + 315 + ], + "type": "text", + "content": "[37] Li Wang, Dong Wei Li, Han Liu, Jinzhan Peng, Lu Tian, and Yi Shan. Cross-Dataset Collaborative Learning for Semantic Segmentation. AAAI, 2021. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 316, + 287, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 287, + 348 + ], + "type": "text", + "content": "[38] Xudong Wang, Zhaowei Cai, Dashan Gao, and Nuno Vasconcelos. Towards Universal Object Detection by Domain Attention. CVPR, pages 7281-7290, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 349, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 349, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 349, + 287, + 392 + ], + "type": "text", + "content": "[39] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E. Sarma, Michael M. Bronstein, and Justin M. Solomon. Dynamic Graph CNN for Learning on Point Clouds. TOG, 38:1-12, 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 393, + 287, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 287, + 426 + ], + "type": "text", + "content": "[40] Christian Wojek, Stefan Walk, Stefan Roth, and Bernt Schiele. Monocular 3D scene understanding with explicit occlusion reasoning. CVPR, pages 1993-2000, 2011. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 426, + 287, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 287, + 470 + ], + "type": "text", + "content": "[41] Bichen Wu, Alvin Wan, Xiangyu Yue, and Kurt Keutzer. SqueezeSeg: Convolutional Neural Nets with Recurrent CRF for Real-Time Road-Object Segmentation from 3D Li-DAR Point Cloud. ICRA, pages 1887–1893, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 471, + 287, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 287, + 503 + ], + "type": "text", + "content": "[42] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point Transformer V2: Grouped Vector Attention and Partition-based Pooling. In NIPS, 2022. 2, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 504, + 287, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 504, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 48, + 504, + 287, + 548 + ], + "type": "text", + "content": "[43] Xiaoyang Wu, Zhuotao Tian, Xin Wen, Bohao Peng, Xihui Liu, Kaicheng Yu, and Hengshuang Zhao. Towards Large-scale 3D Representation Learning with Multi-dataset Point Prompt Training. ArXiv, abs/2308.09718, 2023. 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 548, + 287, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 592 + ], + "type": "text", + "content": "[44] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3D ShapeNets: A deep representation for volumetric shapes. CVPR, pages 1912-1920, 2015. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 592, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 287, + 646 + ], + "type": "text", + "content": "[45] Yichen Xie, Chenfeng Xu, Marie-Julie Rakotosaona, Patrick Rim, Federico Tombari, Kurt Keutzer, Masayoshi Tomizuka, and Wei Zhan. SparseFusion: Fusing Multi-Modal Sparse Representations for Multi-Sensor 3D Object Detection. ICCV, abs/2304.14340, 2023. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "type": "text", + "content": "[46] Guodong Xu, Ziwei Liu, Xiaoxiao Li, and Chen Change Loy. Knowledge Distillation Meets Self-Supervision. In ECCV, 2020. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[47] Hang Xu, Linpu Fang, Xiaodan Liang, Wenxiong Kang, and Zhenguo Li. Universal-RCNN: Universal Object Detector via Transferable Graph R-CNN. In AAAI, 2020. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 464 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 116 + ], + "type": "text", + "content": "[48] Mutian Xu, Runyu Ding, Hengshuang Zhao, and Xiaojuan Qi. PACnv: Position Adaptive Convolution with Dynamic Kernel Assembling on Point Clouds. CVPR, pages 3172-3181, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "text", + "content": "[49] Yu-Qi Yang, Yu-Xiao Guo, Jiangfeng Xiong, Yang Liu, Hao Pan, Peng-Shuai Wang, Xin Tong, and Baining Guo. Swin3D: A Pretrained Transformer Backbone for 3D Indoor Scene Understanding. ArXiv, abs/2304.06906, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 163, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 196 + ], + "type": "text", + "content": "[50] Zetong Yang, Yanan Sun, Shu Liu, Xiaoyong Shen, and Jiaya Jia. STD: Sparse-to-Dense 3D Object Detector for Point Cloud. ICCV, pages 1951-1960, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 197, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 228 + ], + "type": "text", + "content": "[51] Tianwei Yin, Xingyi Zhou, and Philipp Kähenbuhl. Center-based 3D Object Detection and Tracking. CVPR, pages 11779-11788, 2021. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 230, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 545, + 262 + ], + "type": "text", + "content": "[52] Cheng Zhang, Haocheng Wan, Xinyi Shen, and Zizhao Wu. PVT: Point-voxel transformer for point cloud learning. *IJIS*, 37:11985 - 12008, 2021. 2, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "type": "text", + "content": "[53] Yang Zhang, Zixiang Zhou, Philip David, Xiangyu Yue, Zerong Xi, and Hassan Foroosh. PolarNet: An Improved Grid Representation for Online LiDAR Point Clouds Semantic Segmentation. CVPR, pages 9598-9607, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 309, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 342 + ], + "type": "text", + "content": "[54] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip Torr, and Vladlen Koltun. Point Transformer. In ICCV, 2021. 2, 6, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 343, + 545, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 545, + 375 + ], + "type": "text", + "content": "[55] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Simple Multi-dataset Detection. CVPR, pages 7561-7570, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 376, + 545, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 421 + ], + "type": "text", + "content": "[56] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and Asymmetrical 3D Convolution Networks for LiDAR Segmentation. CVPR, pages 9934-9943, 2021. 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 422, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 545, + 464 + ], + "type": "text", + "content": "[57] Özgün Çiçek, Ahmed Abdulkadir, Soeren S. Lienkamp, Thomas Brox, and Olaf Ronneberger. 3D U-Net: Learning Dense Volumetric Segmentation from Sparse Annotation. In MICCAI, 2016. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20028" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/GeoChat_ Grounded Large Vision-Language Model for Remote Sensing/d9275ce8-58db-4c85-a4fc-e4300f4bc853_content_list.json b/2024/GeoChat_ Grounded Large Vision-Language Model for Remote Sensing/d9275ce8-58db-4c85-a4fc-e4300f4bc853_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a675290e1785d1d5fc0fb62b2dbe763d20eb467a --- /dev/null +++ b/2024/GeoChat_ Grounded Large Vision-Language Model for Remote Sensing/d9275ce8-58db-4c85-a4fc-e4300f4bc853_content_list.json @@ -0,0 +1,1661 @@ +[ + { + "type": "text", + "text": "GeoChat: Grounded Large Vision-Language Model for Remote Sensing", + "text_level": 1, + "bbox": [ + 106, + 122, + 864, + 157 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kartik Kuckreja $^{1,2*}$", + "bbox": [ + 166, + 183, + 323, + 202 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Muhammad Sohail Danish $^{1*}$", + "bbox": [ + 364, + 183, + 593, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Muzammal Naseer1", + "bbox": [ + 640, + 184, + 800, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abhijit Das", + "bbox": [ + 205, + 208, + 308, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Salman Khan1,3", + "bbox": [ + 385, + 208, + 516, + 224 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fahad Shahbaz Khan $^{1,4}$", + "bbox": [ + 571, + 208, + 761, + 224 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Mohamed bin Zayed University of AI, 2Birla Institute of Technology & Science, Hyderabad", + "bbox": [ + 210, + 228, + 761, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Australian National University, $^{4}$ Linköping University", + "bbox": [ + 320, + 247, + 648, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "kartik.kuckreja@mbzuai.ac.ae, muhammad.sohail@mbzuai.ac.ae", + "bbox": [ + 266, + 266, + 692, + 279 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 315, + 312, + 330 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in Large Vision-Language Models (VLMs) have shown great promise in natural image domains, allowing users to hold a dialogue about given visual content. However, such general-domain VLMs perform poorly for Remote Sensing (RS) scenarios, leading to inaccurate or fabricated information when presented with RS domain-specific queries. Such a behavior emerges due to the unique challenges introduced by RS imagery. For example, to handle high-resolution RS imagery with diverse scale changes across categories and many small objects, region-level reasoning is necessary alongside holistic scene interpretation. Furthermore, the lack of domain-specific multimodal instruction following data as well as strong backbone models for RS make it hard for the models to align their behavior with user queries. To address these limitations, we propose GeoChat - the first versatile remote sensing VLM that offers multitask conversational capabilities with high-resolution RS images. Specifically, GeoChat can not only answer image-level queries but also accepts region inputs to hold region-specific dialogue. Furthermore, it can visually ground objects in its responses by referring to their spatial coordinates. To address the lack of domain-specific datasets, we generate a novel RS multimodal instruction-following dataset by extending imagetext pairs from existing diverse RS datasets. We establish a comprehensive benchmark for RS multitask conversations and compare with a number of baseline methods. GeoChat demonstrates robust zero-shot performance on various RS tasks, e.g., image and region captioning, visual question answering, scene classification, visually grounded conversations and referring detection. Our code is available here.", + "bbox": [ + 75, + 344, + 472, + 814 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 827, + 209, + 842 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the natural image domain, the abundance of aligned image-text data sourced from web imagery or manual anno-", + "bbox": [ + 75, + 845, + 468, + 877 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0fad32ffbdf2710f079251f3c134ec5c87b4219e0774924ff853c756ca0a336c.jpg", + "image_caption": [ + "Figure 1. GeoChat can accomplish multiple tasks for remote-sensing (RS) image comprehension in a unified framework. Given suitable task tokens and user queries, the model can generate visually grounded responses (text with corresponding object locations - shown on top), visual question answering on images and regions (top left and bottom right, respectively) as well as scene classification (top right) and normal natural language conversations (bottom). This makes it the first RS VLM with grounding capability." + ], + "image_footnote": [], + "bbox": [ + 501, + 311, + 890, + 568 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tations facilitate effective self-supervised vision-language modeling, as demonstrated by multimodal GPT-4 [23] and open-source initiatives like LLaVA [19]. These vision-language models (VLMs), developed through generative pretraining and instruction-tuning, exhibit robust zero-shot task completion across various user-oriented multimodal tasks. The resulting capabilities open the door to the development of versatile multimodal conversational assistants with broad applications in real-world scenarios [12].", + "bbox": [ + 496, + 729, + 890, + 866 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, general-domain VLMs designed for natural images, exhibit poor performance when presented with re", + "bbox": [ + 500, + 869, + 890, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 808, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equally contributing first authors.", + "bbox": [ + 94, + 886, + 279, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "27831", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "motely sensed visual imagery. The performance disparity arises primarily from the distinct nature of content found in remote sensing image-text pairings compared to the publicly available web data. As a result, general-domain VLMs can provide inaccurate information or hallucinate when presented with spatial images from RS sensors. Although there has been significant progress in the field of remote sensing visual question answering (VQA) [39, 41], earlier methods have framed the task as a classification problem. Here, the model chooses answers from predetermined responses found in the training data. It limits their applicability to open-ended answer generation and instruction-following.", + "bbox": [ + 75, + 90, + 472, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we introduce GeoChat, an attempt to extend multimodal instruction-tuning to the remote sensing domain for training a multitask conversational assistant. However, remote-sensing domain lacks a multimodal instruction-tuning conversational dataset. Inspired by recent work in instruction-tuning [14, 19, 29, 42], GeoChat uses Vicuna-v1.5 [7] and an automated pipeline to generate diverse remote sensing multimodal instruction-following data comprising of nearly $318k$ instructions. We create the image-text pairs from various existing remote sensing datasets developed for diverse tasks. These include LRBEN for VQA [20], NWPU-RESISC-45 for scene classification [5] and SAMRS for object detection [31].", + "bbox": [ + 75, + 275, + 472, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A crucial capability of GeoChat is the unification of multiple image and region-level reasoning tasks for RS imagery within a single pipeline (see Fig. 1). We achieve this via distinct task tokens that help suitably direct the model's responses according to user requirements. In addition, the model uses spatial location representations in its inputs to seamlessly reason about local regions and can also generate object locations in its responses to visually ground objects. This enables a diverse set of tasks possible with GeoChat including referring expression detection, image/region captioning, scene classification, natural language conversations and VQA, besides visually grounded conversations.", + "bbox": [ + 75, + 473, + 472, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, this work has the following contributions:", + "bbox": [ + 94, + 657, + 459, + 672 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- RS multimodal instruction following dataset. We present a novel data generation pipeline, to leverage existing object detection dataset [31] to create short descriptions of the images, followed by using Vicuna-v1.5 [7] to create conversations using the generated text alone. Further, we add visual question-answering and scene classification abilities using their corresponding datasets [5, 20]. This results in a total of $318k$ instruction pairs for RS domain.", + "- GeoChat. Leveraging our dataset, we finetune LLaVA-1.5 [14] to create the remote sensing-domain vision-language model - GeoChat. Our LoRA [11] fine-tuning is efficient and avoids forgetting the necessary context embedded in fully-tuned LLaVA model, whose MLP projection is trained to align images into the word embedding space of the LLM (Vicuna-v1.5 [7]). This allows" + ], + "bbox": [ + 76, + 674, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "GeoChat to retain the conversation and instruction following abilities of LLaVA and extend its domain-knowledge to remote sensing tasks.", + "bbox": [ + 511, + 90, + 890, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We also address the lack of evaluation benchmarks to assess the capability of existing VLMs on remote-sensing conversations. To this end, we setup evaluation protocols for conversation grounding in RS, as well as a setup a suite of tasks to allow comparisons with future efforts in this direction. We show various supervised as well as zero-shot evaluations for different remote sensing tasks, including image captioning, visual question answering and scene classification to demonstrate the generalisability of GeoChat conversational VLM.", + "bbox": [ + 500, + 137, + 893, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 300, + 640, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Large Vision-Language Models. The typical architecture of instruction-following Vision Language Models (VLMs) consists of utilising a pre-trained visual backbone[9] to encode visual data, a large language model [7] for interpreting user instructions and generating responses, and a vision-language cross-modal connector, e.g., a linear projection layer [18, 42] or an MLP [17], for fusing visual information with language models. The results achieved with VLMs show great promise; for example, LLaVA [18], Instruct-BLIP [8], Otter [13] and MiniGPT-4 [42] show remarkable gains in language instruction following and visual reasoning ability for natural scenes. More recent studies have shown that these models can be adapted to other domains such as videos [22], biomedical [14, 30] and remote sensing [12].", + "bbox": [ + 496, + 325, + 890, + 537 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Remote Sensing VLMs. The application of generalized VLMs in remote sensing is comparatively sparse. The majority of research so far has neglected the semantic understanding of the items and their relationships towards a deep visual comprehension. Beyond merely identifying the objects in an image, vision-language models are also capable of generating natural language descriptions of the image and inferring the connections between the objects. This makes them more appropriate for tasks like text-based image retrieval, captioning images, and answering visual questions that call for both visual and linguistic knowledge. Although there has been progress in vision language models for remote sensing tasks, such as image captioning [43], zero-shot classification [16] and visual question answering [3, 39], these models can only perform a specific task they are trained for, lack conversational capability and do not possess generic semantic knowledge about the remote sensing images. A major gap exists in the remote sensing domain towards developing general-purpose models to solve all tasks together, while also maintaining conversation abilities. While RSGPT [12] is an initial effort that has shown good conversation ability along with solving multiple tasks, it requires finetuning the model for each task separately, which makes it cumbersome and not generalizable. Fur", + "bbox": [ + 496, + 537, + 892, + 901 + ], + "page_idx": 1 + }, + { + "type": "footer", + "text": "27832", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ther, RSGPT cannot work for region-level reasoning or visual grounding, which our work aims to address.", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. GeoChat: Grounded Remote Sensing VLM", + "text_level": 1, + "bbox": [ + 76, + 133, + 464, + 152 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Visually grounded conversations for remote sensing aim to generate textual responses interleaved with corresponding object locations. Further, a user can also provide visual prompts (e.g., a bounding box) besides natural language questions, and the model should be able to answer questions about the specified Region of Interest (RoI). Such seamless interplay between visual and language modalities necessitate a deep comprehension of linguistic constructions that denote particular objects or elements in a visual scene.", + "bbox": [ + 75, + 159, + 468, + 294 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As mentioned above, GeoChat is the first model capable of holding visually grounded conversations about remotely sensed images. By construction, GeoChat can address not only the challenging task of visually grounded conversations, but can also perform a spectrum of other spatial reasoning tasks that span varying levels of granularity in visual imagery understanding e.g., image/region captioning, referring object detection and image/region-level conversations about remotely sensed images. We formally outline the tasks possible with GeoChat below.", + "bbox": [ + 75, + 295, + 468, + 445 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "a) Image-Level Conversation Tasks. In this task, GeoChat processes an image $\\mathbf{x}$ and a user text query $\\mathbf{q}$ without any specific spatial coordinates in its inputs or outputs. The goal is to perform conversation-based tasks at a holistic level with image-wide context, such as visual question answering (VQA), scene classification and image captioning.", + "b) Region-Level Conversation Tasks. This task involves providing spatial box locations $\\pmb{b}$ in the input to GeoChat besides $\\pmb{x}$ and $\\pmb{q}$ . Region locations $\\pmb{b}$ guide the model's attention to specific regions within the image, so that the model can perform tasks such as region-level captioning, region-specific VQA or multi-turn conversation.", + "c) Grounded Conversation Tasks. With the use of special tokens, termed as task-specification tokens $t$ , GeoChat can be guided to provide object locations at different granularities, while maintaining conversation abilities. It helps in tasks including grounded image captioning/conversation, object grounding and referring expression detection." + ], + "bbox": [ + 76, + 446, + 468, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. GeoChat Architecture", + "text_level": 1, + "bbox": [ + 76, + 726, + 284, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "GeoChat follows the architecture as of LLaVA-v1.5 [17], which consists of three core components, i) Global Image encoder, ii) an MLP adaptor (two linear layers) and iii) LLM. Different to LLaVA, we add specific task prompt that indicates the type of task desired from the model i.e., grounding, image-level or region-level conversations. Additionally, we allow spatial positions within both inputs and outputs, enabling visual prompts as inputs and grounded objects in GeoChat outputs. Notably, the original LLaVA model cannot perform object grounding or accept region", + "bbox": [ + 75, + 750, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/f587cd81832e0824a88d1655bc82324be98073b3b330f859d46c2f43ef19435e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DataSizeResponse formatting prompts
Detailed Description30kDescribe the image in detail.
Multi-Round Conversation65k-
Complex Questions10k-
RSVQA-LRBEN[20]56kAnswer the question using a single word or phrase.
NWPU-RESISC-45[5]31.5k
Floodnet[25]4k
Grounding Description25k-grounding] Describe the image in detail.
Region Captioning40k[identify] {bx_left, by_top, bx_right, by_bottom|θ}
Referring Expression45k[refer] < p > Object < /p>
", + "bbox": [ + 503, + 88, + 890, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Instruction following data used to train GeoChat. Instruction types and format are shown. We use a $308\\mathrm{k}$ set for training and a separate $10\\mathrm{k}$ instruction-set for testing.", + "bbox": [ + 498, + 217, + 890, + 258 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "inputs. Further, the original LLaVA can not reason about remote sensing images which is enabled via our domain-specific dataset. We describe each component in the architecture as follows:", + "bbox": [ + 496, + 290, + 890, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Task Token: The unique quality of GeoChat is its ability to easily switch between different types of remote sensing visual interpretation tasks. To eliminate uncertainty among tasks, our approach assigns a unique task identification to each one. We suggest three distinct task identities, $t \\in \\{ \\text{grounding}, \\text{identify}, \\text{refer} \\}$ , each for grounded conversations, region captioning and referring expression comprehension. As for the case of visual question answering and scene classification, we directly ask the model to output the answer in a single word or phrase, as shown in Table 1. Our approach does not employ any task identification tokens for vision-irrelevant commands. This unified approach is supported by a modular design that efficiently integrates spatial data, giving the model flexibility in its reasoning about visual content.", + "bbox": [ + 496, + 353, + 890, + 579 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Spatial Location Representation. Our model must precisely identify the spatial position of the referenced items for tasks such as grounded conversations, referring expression generation, and comprehension. To this end, we represent the box locations in a textual format to express the geographical position: $\\mathbf{b} = \\{b_{\\mathrm{x\\_left}}, b_{\\mathrm{y\\_top}}, b_{\\mathrm{x\\_right}}, b_{\\mathrm{y\\_bottom}} | \\theta \\}$ . Here, $b_{\\mathrm{x\\_left}}, b_{\\mathrm{y\\_top}}$ denote the top left corner point of box while the $b_{\\mathrm{x\\_right}}, b_{\\mathrm{y\\_bottom}}$ represent the bottom right corner coordinates. The angle $\\theta$ represents the angle of rotation for the bounding box, from the lower edge. Numerical values normalised within the interval [0, 100] are used to represent the x and y coordinates. Region locations in this format are used to interact with the model via its inputs and outputs.", + "bbox": [ + 496, + 580, + 890, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Visual Backbone. GeoChat adapts the pretrained vision backbone of CLIP-ViT(L-14) [28], which has an input resolution of $336 \\times 336$ . This results in effectively 576 patches per image. Since this resolution is not sufficient to understand details presented in remote sensing imagery (e.g., small objects and object details), we interpolate the positional encoding in the transformer-based CLIP [28] model to scale with input image sizes of $504 \\times 504$ . Although this", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "27833", + "bbox": [ + 478, + 925, + 517, + 936 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9a86cc66af1c5179ab8404b3f8d6a880419e922f34ee1d777345ae23c6edf16c.jpg", + "image_caption": [ + "Figure 2. An overview of GeoChat - the first grounded large vision-language model for remote sensing. Given an image input together with a user query, a visual backbone is first used to encode patch-level tokens at a higher resolution via interpolating positional encodings. A multi-layer perceptron (MLP) is used to adapt vision-tokens to language space suitable for input to a Large Language Model (Vicuna 1.5). Besides visual inputs, region locations can also be input to the model together with task-specific prompts that specify the desired task required by the user. Given this context, the LLM can generate natural language responses interleaved with corresponding object locations. GeoChat can perform multiple tasks as shown on top e.g., scene classification, image/region captioning, VQA and grounded conversations.", + "Figure 3. Multi-task instruction template for GeoChat." + ], + "image_footnote": [], + "bbox": [ + 80, + 80, + 890, + 424 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "leads to an increase in the number of patches to almost double (i.e., 1296 per image), this enhanced resolution allows us to handle larger image sizes and also supports better visual grounding in high-resolution RS images.", + "bbox": [ + 75, + 542, + 470, + 603 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "MLP Cross-modal Adaptor. From the frozen CLIP-ViT[28], we project the output tokens $(\\in \\mathbb{R}^{1296\\times 1024})$ with dimensions 1024 onto the language model space, using an MLP adaptor with one hidden layer. The adaptor has an input dimensionality of 1024 and outputs a vector of size 4096, corresponding to the input size of the LLM [7]. A GeLU [10] is used as the activation function.", + "bbox": [ + 75, + 607, + 468, + 714 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Large Language Model. The open source Vicuna-v1.5(7B) [7] large language model is utilised as the foundation for GeoChat. The language model functions as a single interface for diverse vision-language inputs in our framework. To accomplish different vision-language tasks, we directly depend on the Vicuna-v1.5(7B) [7] language tokens. We explicitly interact with the language model to construct textual representations of bounding boxes to express their spatial coordinates for the visual grounding tasks that require the production of spatial locations. Similarly, the safe, aligned and effective behavior of LLM is ensured via system prompts appended together with given inputs. A", + "bbox": [ + 75, + 719, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "[USER] Image Features [Task Identifier] [ASSISTANT]", + "bbox": [ + 506, + 546, + 885, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Low-Rank Adaptation (LoRA) [11] based strategy is used for fine-tuning the LLM. While training, instead of finetuning all of the weights that comprise the weight matrix of the pre-trained Vicuna-v1.5[7], we finetune two smaller matrices in LoRA [11] that approximate the original larger matrix. After that, the fine-tuned adaptor is fed into the pretrained model and utilised for inference. The LoRA adaptation ensures faster training and avoids forgetting original knowledge embedded in the LLM trained and fine-tuned on generic natural language instructions. This is an important feature since it allows the model to bring in external context about generic object types, landmarks and affordances in the remote-sensing reasoning framework of GeoChat.", + "bbox": [ + 496, + 609, + 890, + 806 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Training Details", + "text_level": 1, + "bbox": [ + 500, + 816, + 661, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To enhance the effectiveness of our model on general visual tasks and optimize training efficiency, we employ a strategy that involves initializing the network with pre-trained weights and fine-tuning specific segments for remote sens", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "27834", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Object Attributes and Relationships", + "text_level": 1, + "bbox": [ + 176, + 95, + 411, + 108 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "1. Ship {white,small,top-left}, {grey, small, bottom-right}\n2. Harbor {small, right}, {large, bottom-left}, {small, top right}, {small, left}\n3. Trees {top-right}.", + "bbox": [ + 96, + 108, + 460, + 135 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Referring Expression", + "text_level": 1, + "bbox": [ + 222, + 148, + 361, + 160 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "1. Small Grey ship anchored at large harbor on bottom-right \n2. Some Trees at the top-right. \n3. White ship at the top of a small harbor on the right.", + "bbox": [ + 94, + 160, + 385, + 189 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Region Captioning", + "text_level": 1, + "bbox": [ + 235, + 199, + 351, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "1. Object located at 0.\n2. Size of object at 0 is small.", + "bbox": [ + 94, + 210, + 250, + 231 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/655d531fc8c6e616e4a3e4f7329a204de544a52a8a859b2955f40efe6e05214e.jpg", + "image_caption": [ + "Figure 4. Types of annotations available in the GeoChat instruction-set. Top-row: For a given RS image, we obtain object attribute and relationship information, referring expressions and region captions along with their corresponding region annotations (shown over the image). Bottom-row: This structured information is used to create the rich instruction-set with a total of 318k image-instruction pairs." + ], + "image_footnote": [], + "bbox": [ + 279, + 237, + 310, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "LLM", + "bbox": [ + 279, + 257, + 310, + 267 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MultiTurn Conversation", + "text_level": 1, + "bbox": [ + 102, + 299, + 267, + 311 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "USER: Where is the white ship present?", + "bbox": [ + 98, + 311, + 267, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ASSISTANT: White ship is present at the top left.", + "bbox": [ + 99, + 330, + 267, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "USER: What is the color of the ship anchored at the harbor?", + "bbox": [ + 99, + 349, + 267, + 367 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ASSISTANT: The ship anchored at the harbor is grey in color.", + "bbox": [ + 99, + 367, + 267, + 386 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Detailed Description", + "text_level": 1, + "bbox": [ + 307, + 300, + 462, + 313 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "USER: Please Describe the image in detail.", + "bbox": [ + 299, + 313, + 467, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ASSISTANT: This satellite image shows depicts a river scene with multiple ships and harbors present. There are a total of two ships and four harbors. The grey ship at the bottom right is anchored to the harbor", + "bbox": [ + 299, + 330, + 467, + 396 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Complex Reasoning", + "text_level": 1, + "bbox": [ + 509, + 300, + 656, + 313 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "USER: What does the presence of 4 harbors and 2 ships indicate?", + "bbox": [ + 498, + 313, + 666, + 332 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ASSISTANT: The presence of 4 harbors and 2 ships constitute a bustling maritime environment. The presence of such harbors may indicate a water area with the possibility of export /imports of goods as well as tourism.", + "bbox": [ + 498, + 332, + 666, + 396 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Grounded Description", + "text_level": 1, + "bbox": [ + 702, + 300, + 867, + 311 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "USER:[grounding] Describe the image in detail.", + "bbox": [ + 699, + 311, + 867, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ASSISTANT: This satellite image depicts a grey ship anchored at a large harbor. There is another harbor at the top right, which is adjacent to some trees. There is another white ship close to a harbor at the right.", + "bbox": [ + 697, + 330, + 867, + 396 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ing related tasks. We use a pre-trained CLIP-ViT(L-14) encoder[28], trained on large amounts of textual and visual data, a pretrained MLP adaptor[17], pretrained on a 558K subset of the LAION-CC-SBU [26] dataset with BLIP [15] captions, and Vicuna-v1.5[7] to initialize our model. To adapt our model to remote sensing images, we subsequently LoRA [11] fine-tune the LLM, while keeping the MLP adaptor and the CLIP encoder [28] frozen during training.", + "bbox": [ + 75, + 479, + 468, + 601 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. RS Multimodal Instruction Dataset", + "text_level": 1, + "bbox": [ + 76, + 616, + 397, + 632 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By using LLM Vicuna [7], we align the model to follow a range of instructions by presenting and curating varied instruction-following data with multi-round conversations regarding remote sensing imagery (Table 1). We specifically provide system instructions as prompts that ask Vicuna [7] to generate multi-round question and answer pairs in a manner as if it could visualize the image (although it only has access to the text). This is achieved by providing few-shot in-context examples manually composed within the prompt to show Vicuna [7] how to build high-quality instruction-response pairs based on the caption and information supplied. Specifically, from our short descriptions created using the below pipeline, we randomly sample 65k images to create multi-round conversations, 10k images to generate complex question answers and 30k images to generate detailed descriptions for the given short descriptions.", + "bbox": [ + 75, + 642, + 468, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In combination, after conversion to instruction format,", + "bbox": [ + 96, + 886, + 468, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "we obtain a total of nearly 308k image-instruction pairs for training and 10k for testing. Next, we outline the instruction-set creation process.", + "bbox": [ + 498, + 479, + 890, + 525 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c2813674588d2be2e65b77e2c5998e124f7f745428369c37c6cb84bb8dfab396.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetCategory# Classes# ImagesImage Size
DOTAObject Detection1817,4801024 × 1024
DIORObject Detection2023,463800 × 800
FAIR1MObject Detection3764,147600 × 600
LRBN(rsvqa)Visual Question Answering-600256 × 256
FloodnetVisual Question Answering-40563000 × 4000
NWPU-RESISC-45Scene Classification4531,500256 × 256
", + "bbox": [ + 503, + 529, + 890, + 606 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. List of datasets used to creat our remote-sensing instruction set for GeoChat VLM training. We include object detection, visual question answering and scene classification datasets with varying image sizes and types of classes to ensure diversity.", + "bbox": [ + 498, + 612, + 890, + 669 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Constituent Datasets: In the compilation of our instruction set, we incorporate three distinct types of datasets, encompassing the ones designed for object detection, scene classification, and visual question answering (VQA). Specifically, we integrate three object detection (DOTA [35], DIOR [6], and FAIR1M [27] which together form the SAMRS [31] dataset), one scene classification (NWPU-RESISC-45 [5]), one VQA (LRBEN[20]), and one flood detection [25] VQA dataset (see Table 2). The object detection datasets allow region-level reasoning capability as they offer segmentation masks along with bounding boxes.", + "bbox": [ + 496, + 686, + 890, + 853 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Addition of Missing Classes: Although a wide variety of object classes are included in the object detection databases, several essential categories like buildings, roads,", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "27835", + "bbox": [ + 478, + 925, + 517, + 936 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fad3bd06df2560e997c705ad1a6eb1679b4aa307aacdc1d0f38132556a3b03ae.jpg", + "image_caption": [ + "Figure 5. Qualitative results of GeoChat. (left-right) Results are shown on grounding, referring object detection, and disaster/damage detection. The user can provide task-specific tokens (e.g., [grounding]) to shape model responses according to the desired behavior. The model can generate textual responses (right), only visual grounding (center) and both text and object groundings interleaved together (left). The model can also specify object types, object counts, object attributes and object relationships." + ], + "image_footnote": [], + "bbox": [ + 76, + 80, + 890, + 345 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/037530172dc22d3b84a7c684fd516eedeb8d10b37f29a0c74df246c3163da7e0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
AttributeExample
a1category(e.g. “plane, ship”)
a2color(e.g. “gray, white”)
a3relative size(e.g. “small, large”)
a4relative location(e.g. “top right, bottom”)
a5relation(e.g. “parked at, driving through”)
", + "bbox": [ + 91, + 416, + 459, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and trees are missing. To address this, we propose to utilize ViTAE-RVSA [32] model, pre-trained on the LoveDA dataset [33], which encompasses the required important classes. The model [32] is used to infer these classes on the SAMRS [31] dataset, yielding pseudo labels. To mitigate potential noise in these predictions, we remove the predictions of ViTAE-RVSA [32] for which we already have ground truth from the SAMRS [31] dataset to refine the results.", + "bbox": [ + 75, + 574, + 468, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Attribute extraction: For referring expression annotations, it is important to derive a variety of attributes in RS images. To this end, we have selected five distinct types of attributes, as outlined in Table 3. Object category information can be directly obtained from the SAMRS dataset. For color extraction, we use the K-Means clustering algorithm. Specifically, we extract the object's pixels from the image using ground-truth box and cluster them into $K$ groups. The center of the largest cluster is then selected as the object's color. To specify the relative size of the object, we categorize objects into three sizes: small, normal, and large. This categorization is determined by measuring the area of all in", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/96ea952af3d1f0d8b908e3aed53f81ea9e65e9820a383bf32fed8c97cdb2e3b1.jpg", + "table_caption": [ + "Table 3. List of attributes collected for objects. Attributes are used to obtain referring expressions e.g., small-sized plane to the left." + ], + "table_footnote": [], + "table_body": "
CategoriesExample
Ships and Harbors(e.g. “anchored at, parked at”)
Track Field and Soccer Field(e.g. “Surrounded by, Inside”)
Vehicles, Bridge, Road, Roundabout(e.g. “passing through, passing through”)
Vehicles and Building(e.g. “parked”)
Airport and Plane(e.g. “parked”)
Ship and Helipad(e.g. “on, contains”)
", + "bbox": [ + 493, + 417, + 908, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. Example of relationships between different objects used in the proposed instruction dataset.", + "bbox": [ + 488, + 518, + 892, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "stances of a class in the entire dataset and assigning the $80^{th}$ percentile as the large label. Similarly, the $20^{th}$ percentile is designated as small size, with the remaining falling into the normal category. To determine the object's relative position within the images, we partition the entire image into a $3 \\times 3$ grid, defining regions such as Top Right, Top, Top Left, Left, Center, Right, Bottom Right, Bottom Left, and Bottom. Based on the object's center pixel coordinates, we assign its relative position accordingly.", + "bbox": [ + 496, + 574, + 890, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To define the relation between objects in a given image, we group different objects based on their distance between the bounding boxes, and for each sub-graph, we assign different relationships between objects based on their class labels. Table 4 presents various examples of object relationships. To establish relationships like \"surrounded by,\" we cross-reference pixel-level coordinates to verify if one object is entirely contained within another object.", + "bbox": [ + 496, + 715, + 890, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Expression Generation: To emulate natural language expressions, we employ predefined textual templates based on [40]. The phrase template encompasses the attributes $\\{\\mathrm{a1},\\dots,\\mathrm{a5}\\}$ from Table 3. The expression for a group of", + "bbox": [ + 496, + 840, + 892, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "27836", + "bbox": [ + 478, + 925, + 519, + 936 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7a900cdabd9c0b027ea7c4cd5f158720f7a9ef20d4a0ae24ca66e4ef4e2dc794.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelUCMercedAID
Qwen-VL [1]62.9052.60
MiniGPTv2 [4]4.7612.90
LLaVA-1.5 [17]68.0051.00
GeoChat84.4372.03
", + "bbox": [ + 125, + 88, + 426, + 171 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "objects of the same class is formulated as:", + "bbox": [ + 76, + 244, + 356, + 260 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "\"The/A $\\langle a3\\rangle \\langle a2\\rangle a1\\langle \\mathrm{in / on the}a4\\rangle$", + "bbox": [ + 145, + 271, + 395, + 287 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Attributes that may be absent are enclosed in $\\langle \\rangle$ , and attributes $\\{\\mathrm{a2}, \\mathrm{a3}\\}$ can be arranged in any sequence.", + "bbox": [ + 76, + 299, + 468, + 329 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Similarly, the sentence template incorporates the relational attributes a5 to establish connections between two objects through this structure:", + "bbox": [ + 76, + 329, + 468, + 375 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "\"The/A $\\langle a_i3\\rangle \\langle a_i2\\rangle a_i1a_i5a_j1\\langle \\mathrm{in / on}$ the $a_{j}4\\rangle$", + "bbox": [ + 112, + 385, + 433, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Here, the indices $i$ and $j$ represent the $i^{th}$ and $j^{th}$ object.", + "bbox": [ + 76, + 414, + 455, + 430 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visual Grounding: Although referring expression datasets are available in the natural image domain [37, 38], they lack for the remote sensing domain. To this end, we use our short descriptions as referring expressions to create three different kinds of question answering pairs, i.e. grounding image description, referring expression, and region level captioning, as described in Table 1.", + "bbox": [ + 75, + 430, + 468, + 535 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 547, + 209, + 565 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 571, + 294, + 588 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We initialize the weights of our model with the pretrained CLIP-ViT [24], and LLM (Vicuna-v1.5 [7] and apply LoRA [11] finetuning. Utilizing LoRA, we refine the parameters $W_{q}$ and $W_{v}$ through low-rank adaptation, with a designated rank $r$ set to 64 in our implementation. The model undergoes training consistently at an image resolution of $504 \\times 504$ throughout the whole process. Each training step incorporates specifically crafted multi-modal instructional templates designed for a variety of vision-language tasks during the training process. We use AdamW [21] optimizer with a cosine learning rate scheduler to train our model. We keep the global batch size as 144. We train our model in two stages, first, we train using all of our datasets for 1 epoch, correspondingly 2144 steps, followed by stage 2, where we only train on the grounding dataset for 1600 more steps(at 128 batch size).", + "bbox": [ + 75, + 595, + 468, + 837 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Scene Classification", + "text_level": 1, + "bbox": [ + 76, + 847, + 264, + 862 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Datasets for evaluation. For scene classification, we evaluate our model using AID [34] and UCMerced [36]. AID", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/36c293b1cd4ac72e6d6fe8235141d2ec77a1e516dd90ad690061f4bac0826512.jpg", + "table_caption": [ + "Table 5. Zero-shot scene classification accuracy comparison on AID [34] and UCMerced [36] datasets. In comparison to other generic VLMs, GeoChat performs favorably well." + ], + "table_footnote": [], + "table_body": "
MethodPresenceComparisonRural/UrbanAvg. Accuracy
LLaVA-1.5[17]55.4668.2059.0062.77
Qwen-vl-Chat [1]38.5767.5961.0055.35
MiniGPTv2 [4]55.1655.2239.0054.96
RSVQA[20]87.4781.5090.0086.32
EasyToHard[39]90.6687.4991.6789.94
Bi-Modal[2]91.0691.1692.6691.63
SHRNet [41]91.0390.4894.0091.84
RSGPT[12]91.1791.7094.0092.29
GeoChat91.0990.3394.0090.70
", + "bbox": [ + 504, + 88, + 890, + 237 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Comparisons with general zero-shot (top) and RS-VQA specialized (middle) models on RSVQA-LRBEN [20] dataset for VQA task. [1, 4, 17] are evaluated in zero-shot setting. GeoChat outperforms other zero-shot models and performs competitively to SoTA-supervised models like RSGPT which are specifically finetuned on target dataset (while ours is a generic model not specifically finetuned on target dataset).", + "bbox": [ + 496, + 242, + 890, + 339 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "[34] is a large-scale aerial image collection compiled from Google Earth imagery, with 30 classes, such as a river, dense residential area, etc. The images are labeled by specialists in the field of remote sensing image interpretation. In total, the AID [34] dataset has 10,000 images within 30 classes. The images have been taken from different countries as well as different weather conditions. For evaluation, we use a $20\\%$ split of the AID [34] dataset. UCMered [36] is a Land Use scene classification dataset, with 2,100 images and 21 classes. Each image is of size $256 \\times 256$ . We use the whole UCMered [36] dataset as a zero-shot test set.", + "bbox": [ + 496, + 367, + 890, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results. We prompt the models with all of the classes and prompt to classify the image using just one word/phrase. For example, we input a prompt like \"Classify the image within one of the given classes: dense residential area, ..., school. Answer with one word or short phrase.\" We calculate zero-shot accuracy on both AID and UCMered. GeoChat significantly outperforms other VLM's with an accuracy of $84.43\\%$ on UCMered [36] and $72.03\\%$ on AID [34], as presented in Table 5. Notably, the recent MiniGPT-4-v2[4] fails to follow the instructions provided for this specific task and returns unrelated classes that are not a part of the dataset. It's accuracy is close to $5\\%$ if we pass the answers from Vicuna-v1.5 [7] and ask it to check if the output sentence refers to the ground truth class or not. In comparison, Qwen-VL and LLaVa-1.5 perform well in instruction following, but fall short to GeoChat, due to lack of domain knowledge.", + "bbox": [ + 496, + 534, + 890, + 790 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Visual Question Answering", + "text_level": 1, + "bbox": [ + 500, + 801, + 745, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Datasets for evaluation. RSVQA-HRBEN [20] comprises 10,569 high-resolution photos and 1,066,316 question-answer pairs, with $61.5\\%$ , $11.2\\%$ , $20.5\\%$ , and $6.8\\%$ divided into training, validation, test 1, and test 2 sets, respectively. This dataset has three question types: presence,", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "27837", + "bbox": [ + 478, + 925, + 517, + 936 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/24e985e1d28a6e8cce537c52c82d7a0ef8671c8021e534263692304c8c953689.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelSmallMediumLargeSingle-object groundingMulti-object grounding[refer][grounding]Overall
MiniGPTv2 [4]1.79.921.99.13.68.22.67.6
GeoChat2.913.621.716.04.310.511.810.6
", + "bbox": [ + 81, + 88, + 893, + 148 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8329b583f3df5d019df7463856cddba8d8e2907f938ea1e23984fd86724ec666.jpg", + "table_caption": [ + "Table 7. Performance (acc@0.5%) comparison of GeoChat on our benchmark. Small, medium and large refer to the size of the objects based on the bounding box area. Single/multi-object refer to how many objects the question asks the model to predict. [refer]: object referenced using one attribute from a2, a3 or a4 in Table 3. [grounding]: objects referenced using a combination of attributes from a1-a5 in Table 3. Overall, GeoChat outperforms the baseline, but there is still significant room for further improvement on this complex task." + ], + "table_footnote": [], + "table_body": "
ModelPresenceComparisonAverage Accuracy
Qwen-VL[1]66.4460.4163.06
LLaVA-1.5[17]69.8367.2968.40
MiniGPTv2[4]40.7950.9146.46
GeoChat58.4583.1972.30
", + "bbox": [ + 81, + 218, + 478, + 306 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "comparison, and count. For evaluation, we use the test set-2 for RSVQA-HRBEN [20] with 47k question answer pairs. RSVQA-LR [20] is made up of 772 low-resolution images and 77,232 question-answer pairs, with $77.8\\%$ , $11.1\\%$ , and $11.1\\%$ used for training, validation, and testing, respectively. There are four different categories of questions: presence, comparison, rural/urban, and count. We omitted area and count questions during evaluation because the responses are numerical and quantifiable into numerous categories. In the RSVQA-LRBEN [20] dataset, for example, counting questions are quantified into five categories: 0, between 1 and 10, between 11 and 100, between 101 and 1000, and greater than 1000. For evaluation, we use the test set of RSVQA-LRBEN [20] with 7k question-answer pairs.", + "bbox": [ + 75, + 382, + 468, + 593 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results. To constrain the answers to a simple yes/no and for rural/urban question types, we add a suitable prompt at the end of each question. GeoChat performs close to the SOTA specialist models on RSVQA-LRBEN test set, which is RSGPT [12], finetuned on the target dataset for 5 iterations in comparison. For RSVQA-HRBEN, GeoChat outperforms other VLM's in zero-shot setting on average accuracy by $3.9\\%$ , while beating the Comparison subset by $15.9\\%$ on LLaVA-v1.5 [17], as shown in Table 8.", + "bbox": [ + 75, + 594, + 468, + 729 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Visual Grounding", + "text_level": 1, + "bbox": [ + 76, + 739, + 251, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Datasets for evaluation. For the evaluation of grounding tasks, we propose a new benchmark that contains different referring and grounding tasks. We use the validation set from [31] and used the same dataset creation pipeline as in Sec. 4 to construct the test benchmark. There are a total of 7593 [refer], 560 [grounding], and 495 grounding description questions, as well as 2793 for region captioning. We use accuracy@0.5 as the evaluation metric. Accuracy is calculated if the predicted box has an overlap of more than", + "bbox": [ + 75, + 763, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ba56e70d7759fd1218718272fd7496c711f7aa03dd0009a2821c32e33381e836.jpg", + "table_caption": [ + "Table 8. Comparison with other general ZS model's on RSVQA-HRBEN [20] dataset for visual qa. All models here have not been trained on the target dataset. GeoChat performs favorably well compared to generic VLMs." + ], + "table_footnote": [], + "table_body": "
Modelacc@0.5acc@.25METEOR
MiniGPTv2[4]10.830.916.4
GeoChat11.733.948.9
", + "bbox": [ + 522, + 218, + 870, + 273 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6fa7b205c5b8655e9709242f8c75ef7872633b98c1c0d635eb832cd58b0e5cd3.jpg", + "table_caption": [ + "Table 9. Results on grounding description task." + ], + "table_footnote": [], + "table_body": "
ModelROUGE-1ROUGE-LMETEOR
MiniGPTv2[4]32.131.210.0
GeoChat87.387.283.9
", + "bbox": [ + 524, + 299, + 870, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 10. Region level captioning performance.", + "bbox": [ + 553, + 359, + 836, + 375 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "0.5 IoU with the ground-truth box.", + "text_level": 1, + "bbox": [ + 500, + 390, + 730, + 405 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results. Table 7 shows the performance of our method and MiniGPT-4-v2 [4] on the proposed benchmark. Overall, the model performance is low on small objects or when it has to predict multiple boxes. Compared to MiniGPT-4-v2[4], our model works better on medium size images. On the grounding description task, we calculate both, the IoU for the multiple bounding boxes generated as well as the text answer generated. Our model provides a better description with slightly better box accuracy than MiniGPT-4-v2 [4] (Table 9). As for region-level captioning, we evaluate both models based on the text accuracy with ground truth region-level captions (Table 10). Our model significantly outperforms MiniGPT-4-v2 in terms of ROUGE and ME-TEOR score.", + "bbox": [ + 496, + 406, + 890, + 617 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 632, + 619, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Although recent advancements in large Vision-Language Models (VLMs) have shown promise in nature image domains, their performance in Remote Sensing (RS) scenarios is still limited due to the unique domain-specific challenges. Addressing this gap, we present GeoChat, the first unified remote sensing VLM that excels in multitask conversational capabilities with high-resolution RS images. GeoChat not only answers image-level queries but also engages in region-specific dialogue, grounding responses with precise spatial coordinates. We create a novel RS multimodal instruction-following dataset comprising of $318k$ image-instruction pairs with a diverse multitask format. GeoChat achieves robust zero-shot performance across various RS tasks including scene classification, VQA, multi-turn dialogue, visual grounding and referring object detection, thus establishing a comprehensive benchmark.", + "bbox": [ + 496, + 657, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "27838", + "bbox": [ + 478, + 925, + 519, + 936 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023.7, 8", + "[2] Yakoub Bazi, Mohamad Mahmoud Al Rahhal, Mohamed Lamine Mekhalfi, Mansour Abdulaziz Al Zuair, and Farid Melgani. Bi-modal transformer-based approach for visual question answering in remote sensing imagery. IEEE Transactions on Geoscience and Remote Sensing, 60:1-11, 2022. 7", + "[3] Christel Chappuis, Valérie Zermatten, Sylvain Lobry, Bertrand Le Saux, and Devis Tuia. Prompt-rsvqa: Prompting visual context to a language model for remote sensing visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1372–1381, 2022. 2", + "[4] Jun Chen, Deyao Zhu, Xiaogian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023. 7, 8", + "[5] Gong Cheng, Junwei Han, and Xiaoqiang Lu. Remote sensing image scene classification: Benchmark and state of the art. Proceedings of the IEEE, 105(10):1865-1883, 2017. 2, 3, 5", + "[6] Gong Cheng, Jiabao Wang, Ke Li, Xingxing Xie, Chunbo Lang, Yanqing Yao, and Junwei Han. Anchor-free oriented proposal generator for object detection. IEEE Transactions on Geoscience and Remote Sensing, 60:1-11, 2022. 5", + "[7] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ chatgpt quality, 2023. 2, 4, 5, 7", + "[8] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning, 2023. 2", + "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 2", + "[10] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 4", + "[11] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 2, 4, 5, 7", + "[12] Yuan Hu, Jianlong Yuan, Congcong Wen, Xiaonan Lu, and Xiang Li. Rsgpt: A remote sensing vision language model" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and benchmark. arXiv preprint arXiv:2307.15266, 2023. 1, 2, 7, 8", + "[13] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726, 2023. 2", + "[14] Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. Llava-med: Training a large language-and-vision assistant for biomedicine in one day. arXiv preprint arXiv:2306.00890, 2023. 2", + "[15] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 5", + "[16] Xiang Li, Congcong Wen, Yuan Hu, and Nan Zhou. Rs-clip: Zero shot remote sensing scene classification via contrastive vision-language supervision. International Journal of Applied Earth Observation and Geoinformation, 124:103497, 2023. 2", + "[17] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 2, 3, 5, 7, 8", + "[18] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 2", + "[19] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 1, 2", + "[20] Sylvain Lobry, Diego Marcos, Jesse Murray, and Devis Tuia. Rsvqa: Visual question answering for remote sensing data. IEEE Transactions on Geoscience and Remote Sensing, 58 (12):8555-8566, 2020. 2, 3, 5, 7, 8", + "[21] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 7", + "[22] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 2", + "[23] OpenAI. Gpt-4 technical report, 2023. 1", + "[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7", + "[25] Maryam Rahnemoonfar, Tashnim Chowdhury, Argho Sarkar, Debvrat Varshney, Masoud Yari, and Robin Murphy. Floodnet: A high resolution aerial imagery dataset for post flood scene understanding. arXiv preprint arXiv:2012.02951, 2020. 3, 5", + "[26] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 5" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "27839", + "bbox": [ + 478, + 925, + 519, + 936 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Xian Sun, Peijin Wang, Zhiyuan Yan, Feng Xu, Ruiping Wang, Wenhui Diao, Jin Chen, Jihao Li, Yingchao Feng, Tao Xu, Martin Weinmann, Stefan Hinz, Cheng Wang, and Kun Fu. Fair1m: A benchmark dataset for fine-grained object recognition in high-resolution remote sensing imagery. IS-PRS Journal of Photogrammetry and Remote Sensing, 184: 116–130, 2022. 5", + "[28] Yi Tay, Minh C Phan, Luu Anh Tuan, and Siu Cheung Hui. Learning to rank question answer pairs with holographic dual lstm architecture. In Proceedings of the 40th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 695-704. ACM, 2017. 3, 4, 5", + "[29] Omkar Thawakar, Ashmal Vayani, Salman Khan, Hisham Cholakal, Rao M Anwer, Michael Felsberg, Tim Baldwin, Eric P Xing, and Fahad Shahbaz Khan. Mobillama: Towards accurate and lightweight fully transparent gpt. arXiv preprint arXiv:2402.16840, 2024. 2", + "[30] Omkar Thawkar, Abdelrahman Shaker, Sahal Shaji Mullanlappilly, Hisham Cholakkal, Rao Muhammad Anwer, Salman Khan, Jorma Laaksonen, and Fahad Shahbaz Khan. Xraygpt: Chest radiographs summarization using large medical vision-language models. arXiv: 2306.07971, 2023. 2", + "[31] Di Wang, Jing Zhang, Bo Du, Dacheng Tao, and Liangpei Zhang. Scaling-up remote sensing segmentation dataset with segment anything model. In arxiv, 2023. 2, 5, 6, 8", + "[32] Di Wang, Qiming Zhang, Yufei Xu, Jing Zhang, Bo Du, Dacheng Tao, and Liangpei Zhang. Advancing plain vision transformer toward remote sensing foundation model. IEEE Transactions on Geoscience and Remote Sensing, 61:1-15, 2023. 6", + "[33] Junjue Wang, Zhuo Zheng, Ailong Ma, Xiaoyan Lu, and Yanfei Zhong. LoveDA: A remote sensing land-cover dataset for domain adaptive semantic segmentation, 2021. 6", + "[34] Gui-Song Xia, Jingwen Hu, Fan Hu, Baoguang Shi, Xiang Bai, Yanfei Zhong, Liangpei Zhang, and Xiaoqiang Lu. Aid: A benchmark data set for performance evaluation of aerial scene classification. IEEE Transactions on Geoscience and Remote Sensing, 55(7):3965-3981, 2017. 7", + "[35] Gui-Song Xia, Xiang Bai, Jian Ding, Zhen Zhu, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, and Liangpei Zhang. Dota: A large-scale dataset for object detection in aerial images. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 5", + "[36] Yi Yang and Shawn Newsam. Bag-of-visual-words and spatial extensions for land-use classification. In Proceedings of the 18th SIGSPATIAL international conference on advances in geographic information systems, pages 270–279, 2010. 7", + "[37] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014. 7", + "[38] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. 7" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Zhenghang Yuan, Lichao Mou, Qi Wang, and Xiao Xiang Zhu. From easy to hard: Learning language-guided curriculum for visual question answering on remote sensing data. IEEE Transactions on Geoscience and Remote Sensing, 60: 1-11, 2022. 2, 7", + "[40] Yang Zhan, Zhitong Xiong, and Yuan Yuan. Rsvg: Exploring data and models for visual grounding on remote sensing data. IEEE Transactions on Geoscience and Remote Sensing, 61: 1-13, 2023. 6", + "[41] Zixiao Zhang, Licheng Jiao, Lingling Li, Xu Liu, Puhua Chen, Fang Liu, Yuxuan Li, and Zhicheng Guo. A spatial hierarchical reasoning network for remote sensing visual question answering. IEEE Transactions on Geoscience and Remote Sensing, 61:1-15, 2023. 2, 7", + "[42] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2", + "[43] Usman Zia, M Mohsin Riaz, and Abdul Ghafoor. Transforming remote sensing images to textual descriptions. International Journal of Applied Earth Observation and Geoinformation, 108:102741, 2022. 2" + ], + "bbox": [ + 501, + 90, + 890, + 401 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "27840", + "bbox": [ + 478, + 926, + 519, + 936 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/GeoChat_ Grounded Large Vision-Language Model for Remote Sensing/d9275ce8-58db-4c85-a4fc-e4300f4bc853_model.json b/2024/GeoChat_ Grounded Large Vision-Language Model for Remote Sensing/d9275ce8-58db-4c85-a4fc-e4300f4bc853_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c567b22968decd6f6e25b3f2d0d77fe14b39a96a --- /dev/null +++ b/2024/GeoChat_ Grounded Large Vision-Language Model for Remote Sensing/d9275ce8-58db-4c85-a4fc-e4300f4bc853_model.json @@ -0,0 +1,2189 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.809, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.107, + 0.123, + 0.865, + 0.158 + ], + "angle": 0, + "content": "GeoChat: Grounded Large Vision-Language Model for Remote Sensing" + }, + { + "type": "text", + "bbox": [ + 0.167, + 0.184, + 0.325, + 0.203 + ], + "angle": 0, + "content": "Kartik Kuckreja\\(^{1,2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.366, + 0.184, + 0.594, + 0.202 + ], + "angle": 0, + "content": "Muhammad Sohail Danish\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.642, + 0.185, + 0.801, + 0.202 + ], + "angle": 0, + "content": "Muzammal Naseer1" + }, + { + "type": "text", + "bbox": [ + 0.207, + 0.209, + 0.309, + 0.228 + ], + "angle": 0, + "content": "Abhijit Das" + }, + { + "type": "text", + "bbox": [ + 0.387, + 0.209, + 0.517, + 0.226 + ], + "angle": 0, + "content": "Salman Khan1,3" + }, + { + "type": "text", + "bbox": [ + 0.573, + 0.209, + 0.763, + 0.226 + ], + "angle": 0, + "content": "Fahad Shahbaz Khan\\(^{1,4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.211, + 0.229, + 0.762, + 0.246 + ], + "angle": 0, + "content": "1Mohamed bin Zayed University of AI, 2Birla Institute of Technology & Science, Hyderabad" + }, + { + "type": "text", + "bbox": [ + 0.321, + 0.248, + 0.649, + 0.264 + ], + "angle": 0, + "content": "\\(^{3}\\)Australian National University, \\(^{4}\\)Linköping University" + }, + { + "type": "text", + "bbox": [ + 0.267, + 0.267, + 0.694, + 0.28 + ], + "angle": 0, + "content": "kartik.kuckreja@mbzuai.ac.ae, muhammad.sohail@mbzuai.ac.ae" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.316, + 0.313, + 0.331 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.345, + 0.473, + 0.815 + ], + "angle": 0, + "content": "Recent advancements in Large Vision-Language Models (VLMs) have shown great promise in natural image domains, allowing users to hold a dialogue about given visual content. However, such general-domain VLMs perform poorly for Remote Sensing (RS) scenarios, leading to inaccurate or fabricated information when presented with RS domain-specific queries. Such a behavior emerges due to the unique challenges introduced by RS imagery. For example, to handle high-resolution RS imagery with diverse scale changes across categories and many small objects, region-level reasoning is necessary alongside holistic scene interpretation. Furthermore, the lack of domain-specific multimodal instruction following data as well as strong backbone models for RS make it hard for the models to align their behavior with user queries. To address these limitations, we propose GeoChat - the first versatile remote sensing VLM that offers multitask conversational capabilities with high-resolution RS images. Specifically, GeoChat can not only answer image-level queries but also accepts region inputs to hold region-specific dialogue. Furthermore, it can visually ground objects in its responses by referring to their spatial coordinates. To address the lack of domain-specific datasets, we generate a novel RS multimodal instruction-following dataset by extending imagetext pairs from existing diverse RS datasets. We establish a comprehensive benchmark for RS multitask conversations and compare with a number of baseline methods. GeoChat demonstrates robust zero-shot performance on various RS tasks, e.g., image and region captioning, visual question answering, scene classification, visually grounded conversations and referring detection. Our code is available here." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.828, + 0.21, + 0.843 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.847, + 0.47, + 0.878 + ], + "angle": 0, + "content": "In the natural image domain, the abundance of aligned image-text data sourced from web imagery or manual anno-" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.313, + 0.892, + 0.569 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.582, + 0.892, + 0.694 + ], + "angle": 0, + "content": "Figure 1. GeoChat can accomplish multiple tasks for remote-sensing (RS) image comprehension in a unified framework. Given suitable task tokens and user queries, the model can generate visually grounded responses (text with corresponding object locations - shown on top), visual question answering on images and regions (top left and bottom right, respectively) as well as scene classification (top right) and normal natural language conversations (bottom). This makes it the first RS VLM with grounding capability." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.73, + 0.892, + 0.867 + ], + "angle": 0, + "content": "tations facilitate effective self-supervised vision-language modeling, as demonstrated by multimodal GPT-4 [23] and open-source initiatives like LLaVA [19]. These vision-language models (VLMs), developed through generative pretraining and instruction-tuning, exhibit robust zero-shot task completion across various user-oriented multimodal tasks. The resulting capabilities open the door to the development of versatile multimodal conversational assistants with broad applications in real-world scenarios [12]." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "However, general-domain VLMs designed for natural images, exhibit poor performance when presented with re" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.281, + 0.9 + ], + "angle": 0, + "content": "*Equally contributing first authors." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.925, + 0.518, + 0.937 + ], + "angle": 0, + "content": "27831" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.273 + ], + "angle": 0, + "content": "motely sensed visual imagery. The performance disparity arises primarily from the distinct nature of content found in remote sensing image-text pairings compared to the publicly available web data. As a result, general-domain VLMs can provide inaccurate information or hallucinate when presented with spatial images from RS sensors. Although there has been significant progress in the field of remote sensing visual question answering (VQA) [39, 41], earlier methods have framed the task as a classification problem. Here, the model chooses answers from predetermined responses found in the training data. It limits their applicability to open-ended answer generation and instruction-following." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.276, + 0.473, + 0.472 + ], + "angle": 0, + "content": "In this paper, we introduce GeoChat, an attempt to extend multimodal instruction-tuning to the remote sensing domain for training a multitask conversational assistant. However, remote-sensing domain lacks a multimodal instruction-tuning conversational dataset. Inspired by recent work in instruction-tuning [14, 19, 29, 42], GeoChat uses Vicuna-v1.5 [7] and an automated pipeline to generate diverse remote sensing multimodal instruction-following data comprising of nearly \\(318k\\) instructions. We create the image-text pairs from various existing remote sensing datasets developed for diverse tasks. These include LRBEN for VQA [20], NWPU-RESISC-45 for scene classification [5] and SAMRS for object detection [31]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.474, + 0.473, + 0.655 + ], + "angle": 0, + "content": "A crucial capability of GeoChat is the unification of multiple image and region-level reasoning tasks for RS imagery within a single pipeline (see Fig. 1). We achieve this via distinct task tokens that help suitably direct the model's responses according to user requirements. In addition, the model uses spatial location representations in its inputs to seamlessly reason about local regions and can also generate object locations in its responses to visually ground objects. This enables a diverse set of tasks possible with GeoChat including referring expression detection, image/region captioning, scene classification, natural language conversations and VQA, besides visually grounded conversations." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.658, + 0.46, + 0.673 + ], + "angle": 0, + "content": "In summary, this work has the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.675, + 0.47, + 0.795 + ], + "angle": 0, + "content": "- RS multimodal instruction following dataset. We present a novel data generation pipeline, to leverage existing object detection dataset [31] to create short descriptions of the images, followed by using Vicuna-v1.5 [7] to create conversations using the generated text alone. Further, we add visual question-answering and scene classification abilities using their corresponding datasets [5, 20]. This results in a total of \\(318k\\) instruction pairs for RS domain." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.796, + 0.471, + 0.902 + ], + "angle": 0, + "content": "- GeoChat. Leveraging our dataset, we finetune LLaVA-1.5 [14] to create the remote sensing-domain vision-language model - GeoChat. Our LoRA [11] fine-tuning is efficient and avoids forgetting the necessary context embedded in fully-tuned LLaVA model, whose MLP projection is trained to align images into the word embedding space of the LLM (Vicuna-v1.5 [7]). This allows" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.675, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "GeoChat to retain the conversation and instruction following abilities of LLaVA and extend its domain-knowledge to remote sensing tasks." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.138, + 0.895, + 0.289 + ], + "angle": 0, + "content": "- We also address the lack of evaluation benchmarks to assess the capability of existing VLMs on remote-sensing conversations. To this end, we setup evaluation protocols for conversation grounding in RS, as well as a setup a suite of tasks to allow comparisons with future efforts in this direction. We show various supervised as well as zero-shot evaluations for different remote sensing tasks, including image captioning, visual question answering and scene classification to demonstrate the generalisability of GeoChat conversational VLM." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.301, + 0.642, + 0.318 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.327, + 0.892, + 0.538 + ], + "angle": 0, + "content": "Large Vision-Language Models. The typical architecture of instruction-following Vision Language Models (VLMs) consists of utilising a pre-trained visual backbone[9] to encode visual data, a large language model [7] for interpreting user instructions and generating responses, and a vision-language cross-modal connector, e.g., a linear projection layer [18, 42] or an MLP [17], for fusing visual information with language models. The results achieved with VLMs show great promise; for example, LLaVA [18], Instruct-BLIP [8], Otter [13] and MiniGPT-4 [42] show remarkable gains in language instruction following and visual reasoning ability for natural scenes. More recent studies have shown that these models can be adapted to other domains such as videos [22], biomedical [14, 30] and remote sensing [12]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Remote Sensing VLMs. The application of generalized VLMs in remote sensing is comparatively sparse. The majority of research so far has neglected the semantic understanding of the items and their relationships towards a deep visual comprehension. Beyond merely identifying the objects in an image, vision-language models are also capable of generating natural language descriptions of the image and inferring the connections between the objects. This makes them more appropriate for tasks like text-based image retrieval, captioning images, and answering visual questions that call for both visual and linguistic knowledge. Although there has been progress in vision language models for remote sensing tasks, such as image captioning [43], zero-shot classification [16] and visual question answering [3, 39], these models can only perform a specific task they are trained for, lack conversational capability and do not possess generic semantic knowledge about the remote sensing images. A major gap exists in the remote sensing domain towards developing general-purpose models to solve all tasks together, while also maintaining conversation abilities. While RSGPT [12] is an initial effort that has shown good conversation ability along with solving multiple tasks, it requires finetuning the model for each task separately, which makes it cumbersome and not generalizable. Fur" + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.925, + 0.521, + 0.937 + ], + "angle": 0, + "content": "27832" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "ther, RSGPT cannot work for region-level reasoning or visual grounding, which our work aims to address." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.134, + 0.465, + 0.153 + ], + "angle": 0, + "content": "3. GeoChat: Grounded Remote Sensing VLM" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.16, + 0.47, + 0.295 + ], + "angle": 0, + "content": "Visually grounded conversations for remote sensing aim to generate textual responses interleaved with corresponding object locations. Further, a user can also provide visual prompts (e.g., a bounding box) besides natural language questions, and the model should be able to answer questions about the specified Region of Interest (RoI). Such seamless interplay between visual and language modalities necessitate a deep comprehension of linguistic constructions that denote particular objects or elements in a visual scene." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.296, + 0.469, + 0.446 + ], + "angle": 0, + "content": "As mentioned above, GeoChat is the first model capable of holding visually grounded conversations about remotely sensed images. By construction, GeoChat can address not only the challenging task of visually grounded conversations, but can also perform a spectrum of other spatial reasoning tasks that span varying levels of granularity in visual imagery understanding e.g., image/region captioning, referring object detection and image/region-level conversations about remotely sensed images. We formally outline the tasks possible with GeoChat below." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.447, + 0.469, + 0.538 + ], + "angle": 0, + "content": "a) Image-Level Conversation Tasks. In this task, GeoChat processes an image \\( \\mathbf{x} \\) and a user text query \\( \\mathbf{q} \\) without any specific spatial coordinates in its inputs or outputs. The goal is to perform conversation-based tasks at a holistic level with image-wide context, such as visual question answering (VQA), scene classification and image captioning." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.538, + 0.469, + 0.627 + ], + "angle": 0, + "content": "b) Region-Level Conversation Tasks. This task involves providing spatial box locations \\( \\pmb{b} \\) in the input to GeoChat besides \\( \\pmb{x} \\) and \\( \\pmb{q} \\). Region locations \\( \\pmb{b} \\) guide the model's attention to specific regions within the image, so that the model can perform tasks such as region-level captioning, region-specific VQA or multi-turn conversation." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.628, + 0.469, + 0.718 + ], + "angle": 0, + "content": "c) Grounded Conversation Tasks. With the use of special tokens, termed as task-specification tokens \\( t \\), GeoChat can be guided to provide object locations at different granularities, while maintaining conversation abilities. It helps in tasks including grounded image captioning/conversation, object grounding and referring expression detection." + }, + { + "type": "list", + "bbox": [ + 0.077, + 0.447, + 0.469, + 0.718 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.727, + 0.285, + 0.74 + ], + "angle": 0, + "content": "3.1. GeoChat Architecture" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.469, + 0.901 + ], + "angle": 0, + "content": "GeoChat follows the architecture as of LLaVA-v1.5 [17], which consists of three core components, i) Global Image encoder, ii) an MLP adaptor (two linear layers) and iii) LLM. Different to LLaVA, we add specific task prompt that indicates the type of task desired from the model i.e., grounding, image-level or region-level conversations. Additionally, we allow spatial positions within both inputs and outputs, enabling visual prompts as inputs and grounded objects in GeoChat outputs. Notably, the original LLaVA model cannot perform object grounding or accept region" + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.089, + 0.892, + 0.213 + ], + "angle": 0, + "content": "
DataSizeResponse formatting prompts
Detailed Description30kDescribe the image in detail.
Multi-Round Conversation65k-
Complex Questions10k-
RSVQA-LRBEN[20]56kAnswer the question using a single word or phrase.
NWPU-RESISC-45[5]31.5k
Floodnet[25]4k
Grounding Description25k-grounding] Describe the image in detail.
Region Captioning40k[identify] {bx_left, by_top, bx_right, by_bottom|θ}
Referring Expression45k[refer] < p > Object < /p>
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.218, + 0.892, + 0.26 + ], + "angle": 0, + "content": "Table 1. Instruction following data used to train GeoChat. Instruction types and format are shown. We use a \\(308\\mathrm{k}\\) set for training and a separate \\(10\\mathrm{k}\\) instruction-set for testing." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.291, + 0.892, + 0.351 + ], + "angle": 0, + "content": "inputs. Further, the original LLaVA can not reason about remote sensing images which is enabled via our domain-specific dataset. We describe each component in the architecture as follows:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.354, + 0.892, + 0.58 + ], + "angle": 0, + "content": "Task Token: The unique quality of GeoChat is its ability to easily switch between different types of remote sensing visual interpretation tasks. To eliminate uncertainty among tasks, our approach assigns a unique task identification to each one. We suggest three distinct task identities, \\( t \\in \\{ \\text{grounding}, \\text{identify}, \\text{refer} \\} \\), each for grounded conversations, region captioning and referring expression comprehension. As for the case of visual question answering and scene classification, we directly ask the model to output the answer in a single word or phrase, as shown in Table 1. Our approach does not employ any task identification tokens for vision-irrelevant commands. This unified approach is supported by a modular design that efficiently integrates spatial data, giving the model flexibility in its reasoning about visual content." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.582, + 0.892, + 0.779 + ], + "angle": 0, + "content": "Spatial Location Representation. Our model must precisely identify the spatial position of the referenced items for tasks such as grounded conversations, referring expression generation, and comprehension. To this end, we represent the box locations in a textual format to express the geographical position: \\( \\mathbf{b} = \\{b_{\\mathrm{x\\_left}}, b_{\\mathrm{y\\_top}}, b_{\\mathrm{x\\_right}}, b_{\\mathrm{y\\_bottom}} | \\theta \\} \\). Here, \\( b_{\\mathrm{x\\_left}}, b_{\\mathrm{y\\_top}} \\) denote the top left corner point of box while the \\( b_{\\mathrm{x\\_right}}, b_{\\mathrm{y\\_bottom}} \\) represent the bottom right corner coordinates. The angle \\( \\theta \\) represents the angle of rotation for the bounding box, from the lower edge. Numerical values normalised within the interval [0, 100] are used to represent the x and y coordinates. Region locations in this format are used to interact with the model via its inputs and outputs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Visual Backbone. GeoChat adapts the pretrained vision backbone of CLIP-ViT(L-14) [28], which has an input resolution of \\(336 \\times 336\\). This results in effectively 576 patches per image. Since this resolution is not sufficient to understand details presented in remote sensing imagery (e.g., small objects and object details), we interpolate the positional encoding in the transformer-based CLIP [28] model to scale with input image sizes of \\(504 \\times 504\\). Although this" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.926, + 0.519, + 0.938 + ], + "angle": 0, + "content": "27833" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.082, + 0.891, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.432, + 0.896, + 0.518 + ], + "angle": 0, + "content": "Figure 2. An overview of GeoChat - the first grounded large vision-language model for remote sensing. Given an image input together with a user query, a visual backbone is first used to encode patch-level tokens at a higher resolution via interpolating positional encodings. A multi-layer perceptron (MLP) is used to adapt vision-tokens to language space suitable for input to a Large Language Model (Vicuna 1.5). Besides visual inputs, region locations can also be input to the model together with task-specific prompts that specify the desired task required by the user. Given this context, the LLM can generate natural language responses interleaved with corresponding object locations. GeoChat can perform multiple tasks as shown on top e.g., scene classification, image/region captioning, VQA and grounded conversations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.543, + 0.471, + 0.604 + ], + "angle": 0, + "content": "leads to an increase in the number of patches to almost double (i.e., 1296 per image), this enhanced resolution allows us to handle larger image sizes and also supports better visual grounding in high-resolution RS images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.608, + 0.47, + 0.715 + ], + "angle": 0, + "content": "MLP Cross-modal Adaptor. From the frozen CLIP-ViT[28], we project the output tokens \\((\\in \\mathbb{R}^{1296\\times 1024})\\) with dimensions 1024 onto the language model space, using an MLP adaptor with one hidden layer. The adaptor has an input dimensionality of 1024 and outputs a vector of size 4096, corresponding to the input size of the LLM [7]. A GeLU [10] is used as the activation function." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Large Language Model. The open source Vicuna-v1.5(7B) [7] large language model is utilised as the foundation for GeoChat. The language model functions as a single interface for diverse vision-language inputs in our framework. To accomplish different vision-language tasks, we directly depend on the Vicuna-v1.5(7B) [7] language tokens. We explicitly interact with the language model to construct textual representations of bounding boxes to express their spatial coordinates for the visual grounding tasks that require the production of spatial locations. Similarly, the safe, aligned and effective behavior of LLM is ensured via system prompts appended together with given inputs. A" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.547, + 0.887, + 0.56 + ], + "angle": 0, + "content": "[USER] Image Features [Task Identifier] [ASSISTANT]" + }, + { + "type": "image_caption", + "bbox": [ + 0.534, + 0.57, + 0.859, + 0.585 + ], + "angle": 0, + "content": "Figure 3. Multi-task instruction template for GeoChat." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.611, + 0.892, + 0.807 + ], + "angle": 0, + "content": "Low-Rank Adaptation (LoRA) [11] based strategy is used for fine-tuning the LLM. While training, instead of finetuning all of the weights that comprise the weight matrix of the pre-trained Vicuna-v1.5[7], we finetune two smaller matrices in LoRA [11] that approximate the original larger matrix. After that, the fine-tuned adaptor is fed into the pretrained model and utilised for inference. The LoRA adaptation ensures faster training and avoids forgetting original knowledge embedded in the LLM trained and fine-tuned on generic natural language instructions. This is an important feature since it allows the model to bring in external context about generic object types, landmarks and affordances in the remote-sensing reasoning framework of GeoChat." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.817, + 0.663, + 0.833 + ], + "angle": 0, + "content": "3.2. Training Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.903 + ], + "angle": 0, + "content": "To enhance the effectiveness of our model on general visual tasks and optimize training efficiency, we employ a strategy that involves initializing the network with pre-trained weights and fine-tuning specific segments for remote sens" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.925, + 0.521, + 0.937 + ], + "angle": 0, + "content": "27834" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.177, + 0.096, + 0.413, + 0.109 + ], + "angle": 0, + "content": "Object Attributes and Relationships" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.109, + 0.461, + 0.136 + ], + "angle": 0, + "content": "1. Ship {white,small,top-left}, {grey, small, bottom-right}\n2. Harbor {small, right}, {large, bottom-left}, {small, top right}, {small, left}\n3. Trees {top-right}." + }, + { + "type": "title", + "bbox": [ + 0.223, + 0.149, + 0.362, + 0.161 + ], + "angle": 0, + "content": "Referring Expression" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.161, + 0.386, + 0.19 + ], + "angle": 0, + "content": "1. Small Grey ship anchored at large harbor on bottom-right \n2. Some Trees at the top-right. \n3. White ship at the top of a small harbor on the right." + }, + { + "type": "title", + "bbox": [ + 0.236, + 0.2, + 0.352, + 0.213 + ], + "angle": 0, + "content": "Region Captioning" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.212, + 0.25, + 0.232 + ], + "angle": 0, + "content": "1. Object located at 0.\n2. Size of object at 0 is small." + }, + { + "type": "image", + "bbox": [ + 0.281, + 0.238, + 0.311, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.281, + 0.258, + 0.311, + 0.268 + ], + "angle": 0, + "content": "LLM" + }, + { + "type": "title", + "bbox": [ + 0.103, + 0.3, + 0.269, + 0.313 + ], + "angle": 0, + "content": "MultiTurn Conversation" + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.313, + 0.268, + 0.331 + ], + "angle": 0, + "content": "USER: Where is the white ship present?" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.332, + 0.268, + 0.35 + ], + "angle": 0, + "content": "ASSISTANT: White ship is present at the top left." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.35, + 0.268, + 0.368 + ], + "angle": 0, + "content": "USER: What is the color of the ship anchored at the harbor?" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.368, + 0.268, + 0.387 + ], + "angle": 0, + "content": "ASSISTANT: The ship anchored at the harbor is grey in color." + }, + { + "type": "title", + "bbox": [ + 0.308, + 0.301, + 0.464, + 0.314 + ], + "angle": 0, + "content": "Detailed Description" + }, + { + "type": "text", + "bbox": [ + 0.3, + 0.314, + 0.468, + 0.332 + ], + "angle": 0, + "content": "USER: Please Describe the image in detail." + }, + { + "type": "text", + "bbox": [ + 0.3, + 0.332, + 0.468, + 0.397 + ], + "angle": 0, + "content": "ASSISTANT: This satellite image shows depicts a river scene with multiple ships and harbors present. There are a total of two ships and four harbors. The grey ship at the bottom right is anchored to the harbor" + }, + { + "type": "title", + "bbox": [ + 0.511, + 0.301, + 0.657, + 0.314 + ], + "angle": 0, + "content": "Complex Reasoning" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.314, + 0.668, + 0.333 + ], + "angle": 0, + "content": "USER: What does the presence of 4 harbors and 2 ships indicate?" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.333, + 0.668, + 0.397 + ], + "angle": 0, + "content": "ASSISTANT: The presence of 4 harbors and 2 ships constitute a bustling maritime environment. The presence of such harbors may indicate a water area with the possibility of export /imports of goods as well as tourism." + }, + { + "type": "title", + "bbox": [ + 0.703, + 0.301, + 0.868, + 0.313 + ], + "angle": 0, + "content": "Grounded Description" + }, + { + "type": "text", + "bbox": [ + 0.7, + 0.313, + 0.868, + 0.332 + ], + "angle": 0, + "content": "USER:[grounding] Describe the image in detail." + }, + { + "type": "text", + "bbox": [ + 0.699, + 0.332, + 0.868, + 0.397 + ], + "angle": 0, + "content": "ASSISTANT: This satellite image depicts a grey ship anchored at a large harbor. There is another harbor at the top right, which is adjacent to some trees. There is another white ship close to a harbor at the right." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.411, + 0.894, + 0.455 + ], + "angle": 0, + "content": "Figure 4. Types of annotations available in the GeoChat instruction-set. Top-row: For a given RS image, we obtain object attribute and relationship information, referring expressions and region captions along with their corresponding region annotations (shown over the image). Bottom-row: This structured information is used to create the rich instruction-set with a total of 318k image-instruction pairs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.48, + 0.47, + 0.602 + ], + "angle": 0, + "content": "ing related tasks. We use a pre-trained CLIP-ViT(L-14) encoder[28], trained on large amounts of textual and visual data, a pretrained MLP adaptor[17], pretrained on a 558K subset of the LAION-CC-SBU [26] dataset with BLIP [15] captions, and Vicuna-v1.5[7] to initialize our model. To adapt our model to remote sensing images, we subsequently LoRA [11] fine-tune the LLM, while keeping the MLP adaptor and the CLIP encoder [28] frozen during training." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.617, + 0.398, + 0.633 + ], + "angle": 0, + "content": "4. RS Multimodal Instruction Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.643, + 0.47, + 0.885 + ], + "angle": 0, + "content": "By using LLM Vicuna [7], we align the model to follow a range of instructions by presenting and curating varied instruction-following data with multi-round conversations regarding remote sensing imagery (Table 1). We specifically provide system instructions as prompts that ask Vicuna [7] to generate multi-round question and answer pairs in a manner as if it could visualize the image (although it only has access to the text). This is achieved by providing few-shot in-context examples manually composed within the prompt to show Vicuna [7] how to build high-quality instruction-response pairs based on the caption and information supplied. Specifically, from our short descriptions created using the below pipeline, we randomly sample 65k images to create multi-round conversations, 10k images to generate complex question answers and 30k images to generate detailed descriptions for the given short descriptions." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.887, + 0.469, + 0.9 + ], + "angle": 0, + "content": "In combination, after conversion to instruction format," + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.48, + 0.892, + 0.526 + ], + "angle": 0, + "content": "we obtain a total of nearly 308k image-instruction pairs for training and 10k for testing. Next, we outline the instruction-set creation process." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.53, + 0.892, + 0.607 + ], + "angle": 0, + "content": "
DatasetCategory# Classes# ImagesImage Size
DOTAObject Detection1817,4801024 × 1024
DIORObject Detection2023,463800 × 800
FAIR1MObject Detection3764,147600 × 600
LRBN(rsvqa)Visual Question Answering-600256 × 256
FloodnetVisual Question Answering-40563000 × 4000
NWPU-RESISC-45Scene Classification4531,500256 × 256
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.613, + 0.892, + 0.67 + ], + "angle": 0, + "content": "Table 2. List of datasets used to creat our remote-sensing instruction set for GeoChat VLM training. We include object detection, visual question answering and scene classification datasets with varying image sizes and types of classes to ensure diversity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.688, + 0.892, + 0.854 + ], + "angle": 0, + "content": "Constituent Datasets: In the compilation of our instruction set, we incorporate three distinct types of datasets, encompassing the ones designed for object detection, scene classification, and visual question answering (VQA). Specifically, we integrate three object detection (DOTA [35], DIOR [6], and FAIR1M [27] which together form the SAMRS [31] dataset), one scene classification (NWPU-RESISC-45 [5]), one VQA (LRBEN[20]), and one flood detection [25] VQA dataset (see Table 2). The object detection datasets allow region-level reasoning capability as they offer segmentation masks along with bounding boxes." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Addition of Missing Classes: Although a wide variety of object classes are included in the object detection databases, several essential categories like buildings, roads," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.926, + 0.519, + 0.938 + ], + "angle": 0, + "content": "27835" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.081, + 0.891, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.348, + 0.895, + 0.406 + ], + "angle": 0, + "content": "Figure 5. Qualitative results of GeoChat. (left-right) Results are shown on grounding, referring object detection, and disaster/damage detection. The user can provide task-specific tokens (e.g., [grounding]) to shape model responses according to the desired behavior. The model can generate textual responses (right), only visual grounding (center) and both text and object groundings interleaved together (left). The model can also specify object types, object counts, object attributes and object relationships." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.417, + 0.46, + 0.514 + ], + "angle": 0, + "content": "
AttributeExample
a1category(e.g. “plane, ship”)
a2color(e.g. “gray, white”)
a3relative size(e.g. “small, large”)
a4relative location(e.g. “top right, bottom”)
a5relation(e.g. “parked at, driving through”)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.52, + 0.47, + 0.549 + ], + "angle": 0, + "content": "Table 3. List of attributes collected for objects. Attributes are used to obtain referring expressions e.g., small-sized plane to the left." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.575, + 0.47, + 0.71 + ], + "angle": 0, + "content": "and trees are missing. To address this, we propose to utilize ViTAE-RVSA [32] model, pre-trained on the LoveDA dataset [33], which encompasses the required important classes. The model [32] is used to infer these classes on the SAMRS [31] dataset, yielding pseudo labels. To mitigate potential noise in these predictions, we remove the predictions of ViTAE-RVSA [32] for which we already have ground truth from the SAMRS [31] dataset to refine the results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Attribute extraction: For referring expression annotations, it is important to derive a variety of attributes in RS images. To this end, we have selected five distinct types of attributes, as outlined in Table 3. Object category information can be directly obtained from the SAMRS dataset. For color extraction, we use the K-Means clustering algorithm. Specifically, we extract the object's pixels from the image using ground-truth box and cluster them into \\(K\\) groups. The center of the largest cluster is then selected as the object's color. To specify the relative size of the object, we categorize objects into three sizes: small, normal, and large. This categorization is determined by measuring the area of all in" + }, + { + "type": "table", + "bbox": [ + 0.494, + 0.418, + 0.909, + 0.514 + ], + "angle": 0, + "content": "
CategoriesExample
Ships and Harbors(e.g. “anchored at, parked at”)
Track Field and Soccer Field(e.g. “Surrounded by, Inside”)
Vehicles, Bridge, Road, Roundabout(e.g. “passing through, passing through”)
Vehicles and Building(e.g. “parked”)
Airport and Plane(e.g. “parked”)
Ship and Helipad(e.g. “on, contains”)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.489, + 0.52, + 0.893, + 0.548 + ], + "angle": 0, + "content": "Table 4. Example of relationships between different objects used in the proposed instruction dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.575, + 0.892, + 0.711 + ], + "angle": 0, + "content": "stances of a class in the entire dataset and assigning the \\(80^{th}\\) percentile as the large label. Similarly, the \\(20^{th}\\) percentile is designated as small size, with the remaining falling into the normal category. To determine the object's relative position within the images, we partition the entire image into a \\(3 \\times 3\\) grid, defining regions such as Top Right, Top, Top Left, Left, Center, Right, Bottom Right, Bottom Left, and Bottom. Based on the object's center pixel coordinates, we assign its relative position accordingly." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.716, + 0.892, + 0.837 + ], + "angle": 0, + "content": "To define the relation between objects in a given image, we group different objects based on their distance between the bounding boxes, and for each sub-graph, we assign different relationships between objects based on their class labels. Table 4 presents various examples of object relationships. To establish relationships like \"surrounded by,\" we cross-reference pixel-level coordinates to verify if one object is entirely contained within another object." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Expression Generation: To emulate natural language expressions, we employ predefined textual templates based on [40]. The phrase template encompasses the attributes \\(\\{\\mathrm{a1},\\dots,\\mathrm{a5}\\}\\) from Table 3. The expression for a group of" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.926, + 0.52, + 0.938 + ], + "angle": 0, + "content": "27836" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.126, + 0.089, + 0.428, + 0.172 + ], + "angle": 0, + "content": "
ModelUCMercedAID
Qwen-VL [1]62.9052.60
MiniGPTv2 [4]4.7612.90
LLaVA-1.5 [17]68.0051.00
GeoChat84.4372.03
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.178, + 0.47, + 0.22 + ], + "angle": 0, + "content": "Table 5. Zero-shot scene classification accuracy comparison on AID [34] and UCMerced [36] datasets. In comparison to other generic VLMs, GeoChat performs favorably well." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.246, + 0.357, + 0.261 + ], + "angle": 0, + "content": "objects of the same class is formulated as:" + }, + { + "type": "text", + "bbox": [ + 0.146, + 0.272, + 0.396, + 0.289 + ], + "angle": 0, + "content": "\"The/A \\(\\langle a3\\rangle \\langle a2\\rangle a1\\langle \\mathrm{in / on the}a4\\rangle\\)" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.3, + 0.469, + 0.33 + ], + "angle": 0, + "content": "Attributes that may be absent are enclosed in \\(\\langle \\rangle\\), and attributes \\(\\{\\mathrm{a2}, \\mathrm{a3}\\}\\) can be arranged in any sequence." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.33, + 0.469, + 0.375 + ], + "angle": 0, + "content": "Similarly, the sentence template incorporates the relational attributes a5 to establish connections between two objects through this structure:" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.386, + 0.434, + 0.403 + ], + "angle": 0, + "content": "\"The/A \\(\\langle a_i3\\rangle \\langle a_i2\\rangle a_i1a_i5a_j1\\langle \\mathrm{in / on}\\) the \\(a_{j}4\\rangle\\)" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.415, + 0.457, + 0.431 + ], + "angle": 0, + "content": "Here, the indices \\(i\\) and \\(j\\) represent the \\(i^{th}\\) and \\(j^{th}\\) object." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.431, + 0.469, + 0.536 + ], + "angle": 0, + "content": "Visual Grounding: Although referring expression datasets are available in the natural image domain [37, 38], they lack for the remote sensing domain. To this end, we use our short descriptions as referring expressions to create three different kinds of question answering pairs, i.e. grounding image description, referring expression, and region level captioning, as described in Table 1." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.549, + 0.21, + 0.566 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.573, + 0.295, + 0.589 + ], + "angle": 0, + "content": "5.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.597, + 0.47, + 0.838 + ], + "angle": 0, + "content": "We initialize the weights of our model with the pretrained CLIP-ViT [24], and LLM (Vicuna-v1.5 [7] and apply LoRA [11] finetuning. Utilizing LoRA, we refine the parameters \\( W_{q} \\) and \\( W_{v} \\) through low-rank adaptation, with a designated rank \\( r \\) set to 64 in our implementation. The model undergoes training consistently at an image resolution of \\( 504 \\times 504 \\) throughout the whole process. Each training step incorporates specifically crafted multi-modal instructional templates designed for a variety of vision-language tasks during the training process. We use AdamW [21] optimizer with a cosine learning rate scheduler to train our model. We keep the global batch size as 144. We train our model in two stages, first, we train using all of our datasets for 1 epoch, correspondingly 2144 steps, followed by stage 2, where we only train on the grounding dataset for 1600 more steps(at 128 batch size)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.266, + 0.863 + ], + "angle": 0, + "content": "5.2. Scene Classification" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Datasets for evaluation. For scene classification, we evaluate our model using AID [34] and UCMerced [36]. AID" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.089, + 0.892, + 0.238 + ], + "angle": 0, + "content": "
MethodPresenceComparisonRural/UrbanAvg. Accuracy
LLaVA-1.5[17]55.4668.2059.0062.77
Qwen-vl-Chat [1]38.5767.5961.0055.35
MiniGPTv2 [4]55.1655.2239.0054.96
RSVQA[20]87.4781.5090.0086.32
EasyToHard[39]90.6687.4991.6789.94
Bi-Modal[2]91.0691.1692.6691.63
SHRNet [41]91.0390.4894.0091.84
RSGPT[12]91.1791.7094.0092.29
GeoChat91.0990.3394.0090.70
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.243, + 0.892, + 0.34 + ], + "angle": 0, + "content": "Table 6. Comparisons with general zero-shot (top) and RS-VQA specialized (middle) models on RSVQA-LRBEN [20] dataset for VQA task. [1, 4, 17] are evaluated in zero-shot setting. GeoChat outperforms other zero-shot models and performs competitively to SoTA-supervised models like RSGPT which are specifically finetuned on target dataset (while ours is a generic model not specifically finetuned on target dataset)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.368, + 0.892, + 0.534 + ], + "angle": 0, + "content": "[34] is a large-scale aerial image collection compiled from Google Earth imagery, with 30 classes, such as a river, dense residential area, etc. The images are labeled by specialists in the field of remote sensing image interpretation. In total, the AID [34] dataset has 10,000 images within 30 classes. The images have been taken from different countries as well as different weather conditions. For evaluation, we use a \\(20\\%\\) split of the AID [34] dataset. UCMered [36] is a Land Use scene classification dataset, with 2,100 images and 21 classes. Each image is of size \\(256 \\times 256\\). We use the whole UCMered [36] dataset as a zero-shot test set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.892, + 0.791 + ], + "angle": 0, + "content": "Results. We prompt the models with all of the classes and prompt to classify the image using just one word/phrase. For example, we input a prompt like \"Classify the image within one of the given classes: dense residential area, ..., school. Answer with one word or short phrase.\" We calculate zero-shot accuracy on both AID and UCMered. GeoChat significantly outperforms other VLM's with an accuracy of \\(84.43\\%\\) on UCMered [36] and \\(72.03\\%\\) on AID [34], as presented in Table 5. Notably, the recent MiniGPT-4-v2[4] fails to follow the instructions provided for this specific task and returns unrelated classes that are not a part of the dataset. It's accuracy is close to \\(5\\%\\) if we pass the answers from Vicuna-v1.5 [7] and ask it to check if the output sentence refers to the ground truth class or not. In comparison, Qwen-VL and LLaVa-1.5 perform well in instruction following, but fall short to GeoChat, due to lack of domain knowledge." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.746, + 0.818 + ], + "angle": 0, + "content": "5.3. Visual Question Answering" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Datasets for evaluation. RSVQA-HRBEN [20] comprises 10,569 high-resolution photos and 1,066,316 question-answer pairs, with \\(61.5\\%\\), \\(11.2\\%\\), \\(20.5\\%\\), and \\(6.8\\%\\) divided into training, validation, test 1, and test 2 sets, respectively. This dataset has three question types: presence," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.926, + 0.519, + 0.938 + ], + "angle": 0, + "content": "27837" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.894, + 0.149 + ], + "angle": 0, + "content": "
ModelSmallMediumLargeSingle-object groundingMulti-object grounding[refer][grounding]Overall
MiniGPTv2 [4]1.79.921.99.13.68.22.67.6
GeoChat2.913.621.716.04.310.511.810.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.154, + 0.895, + 0.212 + ], + "angle": 0, + "content": "Table 7. Performance (acc@0.5%) comparison of GeoChat on our benchmark. Small, medium and large refer to the size of the objects based on the bounding box area. Single/multi-object refer to how many objects the question asks the model to predict. [refer]: object referenced using one attribute from a2, a3 or a4 in Table 3. [grounding]: objects referenced using a combination of attributes from a1-a5 in Table 3. Overall, GeoChat outperforms the baseline, but there is still significant room for further improvement on this complex task." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.219, + 0.479, + 0.307 + ], + "angle": 0, + "content": "
ModelPresenceComparisonAverage Accuracy
Qwen-VL[1]66.4460.4163.06
LLaVA-1.5[17]69.8367.2968.40
MiniGPTv2[4]40.7950.9146.46
GeoChat58.4583.1972.30
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.312, + 0.471, + 0.368 + ], + "angle": 0, + "content": "Table 8. Comparison with other general ZS model's on RSVQA-HRBEN [20] dataset for visual qa. All models here have not been trained on the target dataset. GeoChat performs favorably well compared to generic VLMs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.383, + 0.47, + 0.594 + ], + "angle": 0, + "content": "comparison, and count. For evaluation, we use the test set-2 for RSVQA-HRBEN [20] with 47k question answer pairs. RSVQA-LR [20] is made up of 772 low-resolution images and 77,232 question-answer pairs, with \\(77.8\\%\\), \\(11.1\\%\\), and \\(11.1\\%\\) used for training, validation, and testing, respectively. There are four different categories of questions: presence, comparison, rural/urban, and count. We omitted area and count questions during evaluation because the responses are numerical and quantifiable into numerous categories. In the RSVQA-LRBEN [20] dataset, for example, counting questions are quantified into five categories: 0, between 1 and 10, between 11 and 100, between 101 and 1000, and greater than 1000. For evaluation, we use the test set of RSVQA-LRBEN [20] with 7k question-answer pairs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.595, + 0.47, + 0.731 + ], + "angle": 0, + "content": "Results. To constrain the answers to a simple yes/no and for rural/urban question types, we add a suitable prompt at the end of each question. GeoChat performs close to the SOTA specialist models on RSVQA-LRBEN test set, which is RSGPT [12], finetuned on the target dataset for 5 iterations in comparison. For RSVQA-HRBEN, GeoChat outperforms other VLM's in zero-shot setting on average accuracy by \\(3.9\\%\\), while beating the Comparison subset by \\(15.9\\%\\) on LLaVA-v1.5 [17], as shown in Table 8." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.741, + 0.253, + 0.759 + ], + "angle": 0, + "content": "5.4. Visual Grounding" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Datasets for evaluation. For the evaluation of grounding tasks, we propose a new benchmark that contains different referring and grounding tasks. We use the validation set from [31] and used the same dataset creation pipeline as in Sec. 4 to construct the test benchmark. There are a total of 7593 [refer], 560 [grounding], and 495 grounding description questions, as well as 2793 for region captioning. We use accuracy@0.5 as the evaluation metric. Accuracy is calculated if the predicted box has an overlap of more than" + }, + { + "type": "table", + "bbox": [ + 0.523, + 0.219, + 0.872, + 0.274 + ], + "angle": 0, + "content": "
Modelacc@0.5acc@.25METEOR
MiniGPTv2[4]10.830.916.4
GeoChat11.733.948.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.555, + 0.275, + 0.837, + 0.289 + ], + "angle": 0, + "content": "Table 9. Results on grounding description task." + }, + { + "type": "table", + "bbox": [ + 0.525, + 0.3, + 0.872, + 0.356 + ], + "angle": 0, + "content": "
ModelROUGE-1ROUGE-LMETEOR
MiniGPTv2[4]32.131.210.0
GeoChat87.387.283.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.554, + 0.361, + 0.838, + 0.376 + ], + "angle": 0, + "content": "Table 10. Region level captioning performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.391, + 0.731, + 0.406 + ], + "angle": 0, + "content": "0.5 IoU with the ground-truth box." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.407, + 0.892, + 0.618 + ], + "angle": 0, + "content": "Results. Table 7 shows the performance of our method and MiniGPT-4-v2 [4] on the proposed benchmark. Overall, the model performance is low on small objects or when it has to predict multiple boxes. Compared to MiniGPT-4-v2[4], our model works better on medium size images. On the grounding description task, we calculate both, the IoU for the multiple bounding boxes generated as well as the text answer generated. Our model provides a better description with slightly better box accuracy than MiniGPT-4-v2 [4] (Table 9). As for region-level captioning, we evaluate both models based on the text accuracy with ground truth region-level captions (Table 10). Our model significantly outperforms MiniGPT-4-v2 in terms of ROUGE and ME-TEOR score." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.633, + 0.62, + 0.649 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Although recent advancements in large Vision-Language Models (VLMs) have shown promise in nature image domains, their performance in Remote Sensing (RS) scenarios is still limited due to the unique domain-specific challenges. Addressing this gap, we present GeoChat, the first unified remote sensing VLM that excels in multitask conversational capabilities with high-resolution RS images. GeoChat not only answers image-level queries but also engages in region-specific dialogue, grounding responses with precise spatial coordinates. We create a novel RS multimodal instruction-following dataset comprising of \\(318k\\) image-instruction pairs with a diverse multitask format. GeoChat achieves robust zero-shot performance across various RS tasks including scene classification, VQA, multi-turn dialogue, visual grounding and referring object detection, thus establishing a comprehensive benchmark." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.926, + 0.521, + 0.938 + ], + "angle": 0, + "content": "27838" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.183 + ], + "angle": 0, + "content": "[1] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023.7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.472, + 0.268 + ], + "angle": 0, + "content": "[2] Yakoub Bazi, Mohamad Mahmoud Al Rahhal, Mohamed Lamine Mekhalfi, Mansour Abdulaziz Al Zuair, and Farid Melgani. Bi-modal transformer-based approach for visual question answering in remote sensing imagery. IEEE Transactions on Geoscience and Remote Sensing, 60:1-11, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.27, + 0.47, + 0.352 + ], + "angle": 0, + "content": "[3] Christel Chappuis, Valérie Zermatten, Sylvain Lobry, Bertrand Le Saux, and Devis Tuia. Prompt-rsvqa: Prompting visual context to a language model for remote sensing visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1372–1381, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.354, + 0.47, + 0.436 + ], + "angle": 0, + "content": "[4] Jun Chen, Deyao Zhu, Xiaogian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.47, + 0.491 + ], + "angle": 0, + "content": "[5] Gong Cheng, Junwei Han, and Xiaoqiang Lu. Remote sensing image scene classification: Benchmark and state of the art. Proceedings of the IEEE, 105(10):1865-1883, 2017. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.494, + 0.47, + 0.549 + ], + "angle": 0, + "content": "[6] Gong Cheng, Jiabao Wang, Ke Li, Xingxing Xie, Chunbo Lang, Yanqing Yao, and Junwei Han. Anchor-free oriented proposal generator for object detection. IEEE Transactions on Geoscience and Remote Sensing, 60:1-11, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.551, + 0.47, + 0.62 + ], + "angle": 0, + "content": "[7] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) chatgpt quality, 2023. 2, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.621, + 0.47, + 0.688 + ], + "angle": 0, + "content": "[8] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.69, + 0.47, + 0.773 + ], + "angle": 0, + "content": "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.47, + 0.803 + ], + "angle": 0, + "content": "[10] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[11] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 2, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[12] Yuan Hu, Jianlong Yuan, Congcong Wen, Xiaonan Lu, and Xiang Li. Rsgpt: A remote sensing vision language model" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "and benchmark. arXiv preprint arXiv:2307.15266, 2023. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.123, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[13] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.181, + 0.892, + 0.25 + ], + "angle": 0, + "content": "[14] Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. Llava-med: Training a large language-and-vision assistant for biomedicine in one day. arXiv preprint arXiv:2306.00890, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.252, + 0.892, + 0.32 + ], + "angle": 0, + "content": "[15] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.324, + 0.892, + 0.392 + ], + "angle": 0, + "content": "[16] Xiang Li, Congcong Wen, Yuan Hu, and Nan Zhou. Rs-clip: Zero shot remote sensing scene classification via contrastive vision-language supervision. International Journal of Applied Earth Observation and Geoinformation, 124:103497, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.395, + 0.892, + 0.437 + ], + "angle": 0, + "content": "[17] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 2, 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.439, + 0.892, + 0.479 + ], + "angle": 0, + "content": "[18] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.482, + 0.891, + 0.511 + ], + "angle": 0, + "content": "[19] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.513, + 0.892, + 0.568 + ], + "angle": 0, + "content": "[20] Sylvain Lobry, Diego Marcos, Jesse Murray, and Devis Tuia. Rsvqa: Visual question answering for remote sensing data. IEEE Transactions on Geoscience and Remote Sensing, 58 (12):8555-8566, 2020. 2, 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.571, + 0.892, + 0.599 + ], + "angle": 0, + "content": "[21] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.601, + 0.892, + 0.657 + ], + "angle": 0, + "content": "[22] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.659, + 0.776, + 0.673 + ], + "angle": 0, + "content": "[23] OpenAI. Gpt-4 technical report, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.675, + 0.892, + 0.757 + ], + "angle": 0, + "content": "[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.76, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[25] Maryam Rahnemoonfar, Tashnim Chowdhury, Argho Sarkar, Debvrat Varshney, Masoud Yari, and Robin Murphy. Floodnet: A high resolution aerial imagery dataset for post flood scene understanding. arXiv preprint arXiv:2012.02951, 2020. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[26] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.926, + 0.521, + 0.938 + ], + "angle": 0, + "content": "27839" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.188 + ], + "angle": 0, + "content": "[27] Xian Sun, Peijin Wang, Zhiyuan Yan, Feng Xu, Ruiping Wang, Wenhui Diao, Jin Chen, Jihao Li, Yingchao Feng, Tao Xu, Martin Weinmann, Stefan Hinz, Cheng Wang, and Kun Fu. Fair1m: A benchmark dataset for fine-grained object recognition in high-resolution remote sensing imagery. IS-PRS Journal of Photogrammetry and Remote Sensing, 184: 116–130, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.189, + 0.471, + 0.259 + ], + "angle": 0, + "content": "[28] Yi Tay, Minh C Phan, Luu Anh Tuan, and Siu Cheung Hui. Learning to rank question answer pairs with holographic dual lstm architecture. In Proceedings of the 40th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 695-704. ACM, 2017. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.26, + 0.471, + 0.328 + ], + "angle": 0, + "content": "[29] Omkar Thawakar, Ashmal Vayani, Salman Khan, Hisham Cholakal, Rao M Anwer, Michael Felsberg, Tim Baldwin, Eric P Xing, and Fahad Shahbaz Khan. Mobillama: Towards accurate and lightweight fully transparent gpt. arXiv preprint arXiv:2402.16840, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.329, + 0.471, + 0.399 + ], + "angle": 0, + "content": "[30] Omkar Thawkar, Abdelrahman Shaker, Sahal Shaji Mullanlappilly, Hisham Cholakkal, Rao Muhammad Anwer, Salman Khan, Jorma Laaksonen, and Fahad Shahbaz Khan. Xraygpt: Chest radiographs summarization using large medical vision-language models. arXiv: 2306.07971, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.471, + 0.44 + ], + "angle": 0, + "content": "[31] Di Wang, Jing Zhang, Bo Du, Dacheng Tao, and Liangpei Zhang. Scaling-up remote sensing segmentation dataset with segment anything model. In arxiv, 2023. 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.441, + 0.471, + 0.509 + ], + "angle": 0, + "content": "[32] Di Wang, Qiming Zhang, Yufei Xu, Jing Zhang, Bo Du, Dacheng Tao, and Liangpei Zhang. Advancing plain vision transformer toward remote sensing foundation model. IEEE Transactions on Geoscience and Remote Sensing, 61:1-15, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.471, + 0.552 + ], + "angle": 0, + "content": "[33] Junjue Wang, Zhuo Zheng, Ailong Ma, Xiaoyan Lu, and Yanfei Zhong. LoveDA: A remote sensing land-cover dataset for domain adaptive semantic segmentation, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.553, + 0.471, + 0.621 + ], + "angle": 0, + "content": "[34] Gui-Song Xia, Jingwen Hu, Fan Hu, Baoguang Shi, Xiang Bai, Yanfei Zhong, Liangpei Zhang, and Xiaoqiang Lu. Aid: A benchmark data set for performance evaluation of aerial scene classification. IEEE Transactions on Geoscience and Remote Sensing, 55(7):3965-3981, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.471, + 0.691 + ], + "angle": 0, + "content": "[35] Gui-Song Xia, Xiang Bai, Jian Ding, Zhen Zhu, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, and Liangpei Zhang. Dota: A large-scale dataset for object detection in aerial images. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.471, + 0.748 + ], + "angle": 0, + "content": "[36] Yi Yang and Shawn Newsam. Bag-of-visual-words and spatial extensions for land-use classification. In Proceedings of the 18th SIGSPATIAL international conference on advances in geographic information systems, pages 270–279, 2010. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.749, + 0.471, + 0.816 + ], + "angle": 0, + "content": "[37] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.471, + 0.899 + ], + "angle": 0, + "content": "[38] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[39] Zhenghang Yuan, Lichao Mou, Qi Wang, and Xiao Xiang Zhu. From easy to hard: Learning language-guided curriculum for visual question answering on remote sensing data. IEEE Transactions on Geoscience and Remote Sensing, 60: 1-11, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[40] Yang Zhan, Zhitong Xiong, and Yuan Yuan. Rsvg: Exploring data and models for visual grounding on remote sensing data. IEEE Transactions on Geoscience and Remote Sensing, 61: 1-13, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.22, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[41] Zixiao Zhang, Licheng Jiao, Lingling Li, Xu Liu, Puhua Chen, Fang Liu, Yuxuan Li, and Zhicheng Guo. A spatial hierarchical reasoning network for remote sensing visual question answering. IEEE Transactions on Geoscience and Remote Sensing, 61:1-15, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.29, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[42] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.347, + 0.892, + 0.402 + ], + "angle": 0, + "content": "[43] Usman Zia, M Mohsin Riaz, and Abdul Ghafoor. Transforming remote sensing images to textual descriptions. International Journal of Applied Earth Observation and Geoinformation, 108:102741, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.927, + 0.52, + 0.938 + ], + "angle": 0, + "content": "27840" + } + ] +] \ No newline at end of file diff --git a/2024/GeoChat_ Grounded Large Vision-Language Model for Remote Sensing/d9275ce8-58db-4c85-a4fc-e4300f4bc853_origin.pdf b/2024/GeoChat_ Grounded Large Vision-Language Model for Remote Sensing/d9275ce8-58db-4c85-a4fc-e4300f4bc853_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fb64883e355a8d842aa4787ab5d9711729ce42b0 --- /dev/null +++ b/2024/GeoChat_ Grounded Large Vision-Language Model for Remote Sensing/d9275ce8-58db-4c85-a4fc-e4300f4bc853_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fe838998e4ce623546a096663ebdf86de6d9168806741988a6a4d8f8e413e430 +size 2480710 diff --git a/2024/GeoChat_ Grounded Large Vision-Language Model for Remote Sensing/full.md b/2024/GeoChat_ Grounded Large Vision-Language Model for Remote Sensing/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e7ed79dfd9c8c8c49ef48ffccf129a98fbb8e07a --- /dev/null +++ b/2024/GeoChat_ Grounded Large Vision-Language Model for Remote Sensing/full.md @@ -0,0 +1,312 @@ +# GeoChat: Grounded Large Vision-Language Model for Remote Sensing + +Kartik Kuckreja $^{1,2*}$ + +Muhammad Sohail Danish $^{1*}$ + +Muzammal Naseer1 + +Abhijit Das + +Salman Khan1,3 + +Fahad Shahbaz Khan $^{1,4}$ + +1Mohamed bin Zayed University of AI, 2Birla Institute of Technology & Science, Hyderabad + +$^{3}$ Australian National University, $^{4}$ Linköping University + +kartik.kuckreja@mbzuai.ac.ae, muhammad.sohail@mbzuai.ac.ae + +# Abstract + +Recent advancements in Large Vision-Language Models (VLMs) have shown great promise in natural image domains, allowing users to hold a dialogue about given visual content. However, such general-domain VLMs perform poorly for Remote Sensing (RS) scenarios, leading to inaccurate or fabricated information when presented with RS domain-specific queries. Such a behavior emerges due to the unique challenges introduced by RS imagery. For example, to handle high-resolution RS imagery with diverse scale changes across categories and many small objects, region-level reasoning is necessary alongside holistic scene interpretation. Furthermore, the lack of domain-specific multimodal instruction following data as well as strong backbone models for RS make it hard for the models to align their behavior with user queries. To address these limitations, we propose GeoChat - the first versatile remote sensing VLM that offers multitask conversational capabilities with high-resolution RS images. Specifically, GeoChat can not only answer image-level queries but also accepts region inputs to hold region-specific dialogue. Furthermore, it can visually ground objects in its responses by referring to their spatial coordinates. To address the lack of domain-specific datasets, we generate a novel RS multimodal instruction-following dataset by extending imagetext pairs from existing diverse RS datasets. We establish a comprehensive benchmark for RS multitask conversations and compare with a number of baseline methods. GeoChat demonstrates robust zero-shot performance on various RS tasks, e.g., image and region captioning, visual question answering, scene classification, visually grounded conversations and referring detection. Our code is available here. + +# 1. Introduction + +In the natural image domain, the abundance of aligned image-text data sourced from web imagery or manual anno- + +![](images/0fad32ffbdf2710f079251f3c134ec5c87b4219e0774924ff853c756ca0a336c.jpg) +Figure 1. GeoChat can accomplish multiple tasks for remote-sensing (RS) image comprehension in a unified framework. Given suitable task tokens and user queries, the model can generate visually grounded responses (text with corresponding object locations - shown on top), visual question answering on images and regions (top left and bottom right, respectively) as well as scene classification (top right) and normal natural language conversations (bottom). This makes it the first RS VLM with grounding capability. + +tations facilitate effective self-supervised vision-language modeling, as demonstrated by multimodal GPT-4 [23] and open-source initiatives like LLaVA [19]. These vision-language models (VLMs), developed through generative pretraining and instruction-tuning, exhibit robust zero-shot task completion across various user-oriented multimodal tasks. The resulting capabilities open the door to the development of versatile multimodal conversational assistants with broad applications in real-world scenarios [12]. + +However, general-domain VLMs designed for natural images, exhibit poor performance when presented with re + +motely sensed visual imagery. The performance disparity arises primarily from the distinct nature of content found in remote sensing image-text pairings compared to the publicly available web data. As a result, general-domain VLMs can provide inaccurate information or hallucinate when presented with spatial images from RS sensors. Although there has been significant progress in the field of remote sensing visual question answering (VQA) [39, 41], earlier methods have framed the task as a classification problem. Here, the model chooses answers from predetermined responses found in the training data. It limits their applicability to open-ended answer generation and instruction-following. + +In this paper, we introduce GeoChat, an attempt to extend multimodal instruction-tuning to the remote sensing domain for training a multitask conversational assistant. However, remote-sensing domain lacks a multimodal instruction-tuning conversational dataset. Inspired by recent work in instruction-tuning [14, 19, 29, 42], GeoChat uses Vicuna-v1.5 [7] and an automated pipeline to generate diverse remote sensing multimodal instruction-following data comprising of nearly $318k$ instructions. We create the image-text pairs from various existing remote sensing datasets developed for diverse tasks. These include LRBEN for VQA [20], NWPU-RESISC-45 for scene classification [5] and SAMRS for object detection [31]. + +A crucial capability of GeoChat is the unification of multiple image and region-level reasoning tasks for RS imagery within a single pipeline (see Fig. 1). We achieve this via distinct task tokens that help suitably direct the model's responses according to user requirements. In addition, the model uses spatial location representations in its inputs to seamlessly reason about local regions and can also generate object locations in its responses to visually ground objects. This enables a diverse set of tasks possible with GeoChat including referring expression detection, image/region captioning, scene classification, natural language conversations and VQA, besides visually grounded conversations. + +In summary, this work has the following contributions: + +- RS multimodal instruction following dataset. We present a novel data generation pipeline, to leverage existing object detection dataset [31] to create short descriptions of the images, followed by using Vicuna-v1.5 [7] to create conversations using the generated text alone. Further, we add visual question-answering and scene classification abilities using their corresponding datasets [5, 20]. This results in a total of $318k$ instruction pairs for RS domain. +- GeoChat. Leveraging our dataset, we finetune LLaVA-1.5 [14] to create the remote sensing-domain vision-language model - GeoChat. Our LoRA [11] fine-tuning is efficient and avoids forgetting the necessary context embedded in fully-tuned LLaVA model, whose MLP projection is trained to align images into the word embedding space of the LLM (Vicuna-v1.5 [7]). This allows + +GeoChat to retain the conversation and instruction following abilities of LLaVA and extend its domain-knowledge to remote sensing tasks. + +- We also address the lack of evaluation benchmarks to assess the capability of existing VLMs on remote-sensing conversations. To this end, we setup evaluation protocols for conversation grounding in RS, as well as a setup a suite of tasks to allow comparisons with future efforts in this direction. We show various supervised as well as zero-shot evaluations for different remote sensing tasks, including image captioning, visual question answering and scene classification to demonstrate the generalisability of GeoChat conversational VLM. + +# 2. Related Work + +Large Vision-Language Models. The typical architecture of instruction-following Vision Language Models (VLMs) consists of utilising a pre-trained visual backbone[9] to encode visual data, a large language model [7] for interpreting user instructions and generating responses, and a vision-language cross-modal connector, e.g., a linear projection layer [18, 42] or an MLP [17], for fusing visual information with language models. The results achieved with VLMs show great promise; for example, LLaVA [18], Instruct-BLIP [8], Otter [13] and MiniGPT-4 [42] show remarkable gains in language instruction following and visual reasoning ability for natural scenes. More recent studies have shown that these models can be adapted to other domains such as videos [22], biomedical [14, 30] and remote sensing [12]. + +Remote Sensing VLMs. The application of generalized VLMs in remote sensing is comparatively sparse. The majority of research so far has neglected the semantic understanding of the items and their relationships towards a deep visual comprehension. Beyond merely identifying the objects in an image, vision-language models are also capable of generating natural language descriptions of the image and inferring the connections between the objects. This makes them more appropriate for tasks like text-based image retrieval, captioning images, and answering visual questions that call for both visual and linguistic knowledge. Although there has been progress in vision language models for remote sensing tasks, such as image captioning [43], zero-shot classification [16] and visual question answering [3, 39], these models can only perform a specific task they are trained for, lack conversational capability and do not possess generic semantic knowledge about the remote sensing images. A major gap exists in the remote sensing domain towards developing general-purpose models to solve all tasks together, while also maintaining conversation abilities. While RSGPT [12] is an initial effort that has shown good conversation ability along with solving multiple tasks, it requires finetuning the model for each task separately, which makes it cumbersome and not generalizable. Fur + +ther, RSGPT cannot work for region-level reasoning or visual grounding, which our work aims to address. + +# 3. GeoChat: Grounded Remote Sensing VLM + +Visually grounded conversations for remote sensing aim to generate textual responses interleaved with corresponding object locations. Further, a user can also provide visual prompts (e.g., a bounding box) besides natural language questions, and the model should be able to answer questions about the specified Region of Interest (RoI). Such seamless interplay between visual and language modalities necessitate a deep comprehension of linguistic constructions that denote particular objects or elements in a visual scene. + +As mentioned above, GeoChat is the first model capable of holding visually grounded conversations about remotely sensed images. By construction, GeoChat can address not only the challenging task of visually grounded conversations, but can also perform a spectrum of other spatial reasoning tasks that span varying levels of granularity in visual imagery understanding e.g., image/region captioning, referring object detection and image/region-level conversations about remotely sensed images. We formally outline the tasks possible with GeoChat below. + +a) Image-Level Conversation Tasks. In this task, GeoChat processes an image $\mathbf{x}$ and a user text query $\mathbf{q}$ without any specific spatial coordinates in its inputs or outputs. The goal is to perform conversation-based tasks at a holistic level with image-wide context, such as visual question answering (VQA), scene classification and image captioning. +b) Region-Level Conversation Tasks. This task involves providing spatial box locations $\pmb{b}$ in the input to GeoChat besides $\pmb{x}$ and $\pmb{q}$ . Region locations $\pmb{b}$ guide the model's attention to specific regions within the image, so that the model can perform tasks such as region-level captioning, region-specific VQA or multi-turn conversation. +c) Grounded Conversation Tasks. With the use of special tokens, termed as task-specification tokens $t$ , GeoChat can be guided to provide object locations at different granularities, while maintaining conversation abilities. It helps in tasks including grounded image captioning/conversation, object grounding and referring expression detection. + +# 3.1. GeoChat Architecture + +GeoChat follows the architecture as of LLaVA-v1.5 [17], which consists of three core components, i) Global Image encoder, ii) an MLP adaptor (two linear layers) and iii) LLM. Different to LLaVA, we add specific task prompt that indicates the type of task desired from the model i.e., grounding, image-level or region-level conversations. Additionally, we allow spatial positions within both inputs and outputs, enabling visual prompts as inputs and grounded objects in GeoChat outputs. Notably, the original LLaVA model cannot perform object grounding or accept region + +
DataSizeResponse formatting prompts
Detailed Description30kDescribe the image in detail.
Multi-Round Conversation65k-
Complex Questions10k-
RSVQA-LRBEN[20]56kAnswer the question using a single word or phrase.
NWPU-RESISC-45[5]31.5k
Floodnet[25]4k
Grounding Description25k-grounding] Describe the image in detail.
Region Captioning40k[identify] {bx_left, by_top, bx_right, by_bottom|θ}
Referring Expression45k[refer] < p > Object < /p>
+ +Table 1. Instruction following data used to train GeoChat. Instruction types and format are shown. We use a $308\mathrm{k}$ set for training and a separate $10\mathrm{k}$ instruction-set for testing. + +inputs. Further, the original LLaVA can not reason about remote sensing images which is enabled via our domain-specific dataset. We describe each component in the architecture as follows: + +Task Token: The unique quality of GeoChat is its ability to easily switch between different types of remote sensing visual interpretation tasks. To eliminate uncertainty among tasks, our approach assigns a unique task identification to each one. We suggest three distinct task identities, $t \in \{ \text{grounding}, \text{identify}, \text{refer} \}$ , each for grounded conversations, region captioning and referring expression comprehension. As for the case of visual question answering and scene classification, we directly ask the model to output the answer in a single word or phrase, as shown in Table 1. Our approach does not employ any task identification tokens for vision-irrelevant commands. This unified approach is supported by a modular design that efficiently integrates spatial data, giving the model flexibility in its reasoning about visual content. + +Spatial Location Representation. Our model must precisely identify the spatial position of the referenced items for tasks such as grounded conversations, referring expression generation, and comprehension. To this end, we represent the box locations in a textual format to express the geographical position: $\mathbf{b} = \{b_{\mathrm{x\_left}}, b_{\mathrm{y\_top}}, b_{\mathrm{x\_right}}, b_{\mathrm{y\_bottom}} | \theta \}$ . Here, $b_{\mathrm{x\_left}}, b_{\mathrm{y\_top}}$ denote the top left corner point of box while the $b_{\mathrm{x\_right}}, b_{\mathrm{y\_bottom}}$ represent the bottom right corner coordinates. The angle $\theta$ represents the angle of rotation for the bounding box, from the lower edge. Numerical values normalised within the interval [0, 100] are used to represent the x and y coordinates. Region locations in this format are used to interact with the model via its inputs and outputs. + +Visual Backbone. GeoChat adapts the pretrained vision backbone of CLIP-ViT(L-14) [28], which has an input resolution of $336 \times 336$ . This results in effectively 576 patches per image. Since this resolution is not sufficient to understand details presented in remote sensing imagery (e.g., small objects and object details), we interpolate the positional encoding in the transformer-based CLIP [28] model to scale with input image sizes of $504 \times 504$ . Although this + +![](images/9a86cc66af1c5179ab8404b3f8d6a880419e922f34ee1d777345ae23c6edf16c.jpg) +Figure 2. An overview of GeoChat - the first grounded large vision-language model for remote sensing. Given an image input together with a user query, a visual backbone is first used to encode patch-level tokens at a higher resolution via interpolating positional encodings. A multi-layer perceptron (MLP) is used to adapt vision-tokens to language space suitable for input to a Large Language Model (Vicuna 1.5). Besides visual inputs, region locations can also be input to the model together with task-specific prompts that specify the desired task required by the user. Given this context, the LLM can generate natural language responses interleaved with corresponding object locations. GeoChat can perform multiple tasks as shown on top e.g., scene classification, image/region captioning, VQA and grounded conversations. +Figure 3. Multi-task instruction template for GeoChat. + +leads to an increase in the number of patches to almost double (i.e., 1296 per image), this enhanced resolution allows us to handle larger image sizes and also supports better visual grounding in high-resolution RS images. + +MLP Cross-modal Adaptor. From the frozen CLIP-ViT[28], we project the output tokens $(\in \mathbb{R}^{1296\times 1024})$ with dimensions 1024 onto the language model space, using an MLP adaptor with one hidden layer. The adaptor has an input dimensionality of 1024 and outputs a vector of size 4096, corresponding to the input size of the LLM [7]. A GeLU [10] is used as the activation function. + +Large Language Model. The open source Vicuna-v1.5(7B) [7] large language model is utilised as the foundation for GeoChat. The language model functions as a single interface for diverse vision-language inputs in our framework. To accomplish different vision-language tasks, we directly depend on the Vicuna-v1.5(7B) [7] language tokens. We explicitly interact with the language model to construct textual representations of bounding boxes to express their spatial coordinates for the visual grounding tasks that require the production of spatial locations. Similarly, the safe, aligned and effective behavior of LLM is ensured via system prompts appended together with given inputs. A + +[USER] Image Features [Task Identifier] [ASSISTANT] + +Low-Rank Adaptation (LoRA) [11] based strategy is used for fine-tuning the LLM. While training, instead of finetuning all of the weights that comprise the weight matrix of the pre-trained Vicuna-v1.5[7], we finetune two smaller matrices in LoRA [11] that approximate the original larger matrix. After that, the fine-tuned adaptor is fed into the pretrained model and utilised for inference. The LoRA adaptation ensures faster training and avoids forgetting original knowledge embedded in the LLM trained and fine-tuned on generic natural language instructions. This is an important feature since it allows the model to bring in external context about generic object types, landmarks and affordances in the remote-sensing reasoning framework of GeoChat. + +# 3.2. Training Details + +To enhance the effectiveness of our model on general visual tasks and optimize training efficiency, we employ a strategy that involves initializing the network with pre-trained weights and fine-tuning specific segments for remote sens + +# Object Attributes and Relationships + +1. Ship {white,small,top-left}, {grey, small, bottom-right} +2. Harbor {small, right}, {large, bottom-left}, {small, top right}, {small, left} +3. Trees {top-right}. + +# Referring Expression + +1. Small Grey ship anchored at large harbor on bottom-right +2. Some Trees at the top-right. +3. White ship at the top of a small harbor on the right. + +# Region Captioning + +1. Object located at 0. +2. Size of object at 0 is small. + +![](images/655d531fc8c6e616e4a3e4f7329a204de544a52a8a859b2955f40efe6e05214e.jpg) +Figure 4. Types of annotations available in the GeoChat instruction-set. Top-row: For a given RS image, we obtain object attribute and relationship information, referring expressions and region captions along with their corresponding region annotations (shown over the image). Bottom-row: This structured information is used to create the rich instruction-set with a total of 318k image-instruction pairs. + +LLM + +# MultiTurn Conversation + +USER: Where is the white ship present? + +ASSISTANT: White ship is present at the top left. + +USER: What is the color of the ship anchored at the harbor? + +ASSISTANT: The ship anchored at the harbor is grey in color. + +# Detailed Description + +USER: Please Describe the image in detail. + +ASSISTANT: This satellite image shows depicts a river scene with multiple ships and harbors present. There are a total of two ships and four harbors. The grey ship at the bottom right is anchored to the harbor + +# Complex Reasoning + +USER: What does the presence of 4 harbors and 2 ships indicate? + +ASSISTANT: The presence of 4 harbors and 2 ships constitute a bustling maritime environment. The presence of such harbors may indicate a water area with the possibility of export /imports of goods as well as tourism. + +# Grounded Description + +USER:[grounding] Describe the image in detail. + +ASSISTANT: This satellite image depicts a grey ship anchored at a large harbor. There is another harbor at the top right, which is adjacent to some trees. There is another white ship close to a harbor at the right. + +ing related tasks. We use a pre-trained CLIP-ViT(L-14) encoder[28], trained on large amounts of textual and visual data, a pretrained MLP adaptor[17], pretrained on a 558K subset of the LAION-CC-SBU [26] dataset with BLIP [15] captions, and Vicuna-v1.5[7] to initialize our model. To adapt our model to remote sensing images, we subsequently LoRA [11] fine-tune the LLM, while keeping the MLP adaptor and the CLIP encoder [28] frozen during training. + +# 4. RS Multimodal Instruction Dataset + +By using LLM Vicuna [7], we align the model to follow a range of instructions by presenting and curating varied instruction-following data with multi-round conversations regarding remote sensing imagery (Table 1). We specifically provide system instructions as prompts that ask Vicuna [7] to generate multi-round question and answer pairs in a manner as if it could visualize the image (although it only has access to the text). This is achieved by providing few-shot in-context examples manually composed within the prompt to show Vicuna [7] how to build high-quality instruction-response pairs based on the caption and information supplied. Specifically, from our short descriptions created using the below pipeline, we randomly sample 65k images to create multi-round conversations, 10k images to generate complex question answers and 30k images to generate detailed descriptions for the given short descriptions. + +In combination, after conversion to instruction format, + +we obtain a total of nearly 308k image-instruction pairs for training and 10k for testing. Next, we outline the instruction-set creation process. + +
DatasetCategory# Classes# ImagesImage Size
DOTAObject Detection1817,4801024 × 1024
DIORObject Detection2023,463800 × 800
FAIR1MObject Detection3764,147600 × 600
LRBN(rsvqa)Visual Question Answering-600256 × 256
FloodnetVisual Question Answering-40563000 × 4000
NWPU-RESISC-45Scene Classification4531,500256 × 256
+ +Table 2. List of datasets used to creat our remote-sensing instruction set for GeoChat VLM training. We include object detection, visual question answering and scene classification datasets with varying image sizes and types of classes to ensure diversity. + +Constituent Datasets: In the compilation of our instruction set, we incorporate three distinct types of datasets, encompassing the ones designed for object detection, scene classification, and visual question answering (VQA). Specifically, we integrate three object detection (DOTA [35], DIOR [6], and FAIR1M [27] which together form the SAMRS [31] dataset), one scene classification (NWPU-RESISC-45 [5]), one VQA (LRBEN[20]), and one flood detection [25] VQA dataset (see Table 2). The object detection datasets allow region-level reasoning capability as they offer segmentation masks along with bounding boxes. + +Addition of Missing Classes: Although a wide variety of object classes are included in the object detection databases, several essential categories like buildings, roads, + +![](images/fad3bd06df2560e997c705ad1a6eb1679b4aa307aacdc1d0f38132556a3b03ae.jpg) +Figure 5. Qualitative results of GeoChat. (left-right) Results are shown on grounding, referring object detection, and disaster/damage detection. The user can provide task-specific tokens (e.g., [grounding]) to shape model responses according to the desired behavior. The model can generate textual responses (right), only visual grounding (center) and both text and object groundings interleaved together (left). The model can also specify object types, object counts, object attributes and object relationships. + +
AttributeExample
a1category(e.g. “plane, ship”)
a2color(e.g. “gray, white”)
a3relative size(e.g. “small, large”)
a4relative location(e.g. “top right, bottom”)
a5relation(e.g. “parked at, driving through”)
+ +and trees are missing. To address this, we propose to utilize ViTAE-RVSA [32] model, pre-trained on the LoveDA dataset [33], which encompasses the required important classes. The model [32] is used to infer these classes on the SAMRS [31] dataset, yielding pseudo labels. To mitigate potential noise in these predictions, we remove the predictions of ViTAE-RVSA [32] for which we already have ground truth from the SAMRS [31] dataset to refine the results. + +Attribute extraction: For referring expression annotations, it is important to derive a variety of attributes in RS images. To this end, we have selected five distinct types of attributes, as outlined in Table 3. Object category information can be directly obtained from the SAMRS dataset. For color extraction, we use the K-Means clustering algorithm. Specifically, we extract the object's pixels from the image using ground-truth box and cluster them into $K$ groups. The center of the largest cluster is then selected as the object's color. To specify the relative size of the object, we categorize objects into three sizes: small, normal, and large. This categorization is determined by measuring the area of all in + +Table 3. List of attributes collected for objects. Attributes are used to obtain referring expressions e.g., small-sized plane to the left. + +
CategoriesExample
Ships and Harbors(e.g. “anchored at, parked at”)
Track Field and Soccer Field(e.g. “Surrounded by, Inside”)
Vehicles, Bridge, Road, Roundabout(e.g. “passing through, passing through”)
Vehicles and Building(e.g. “parked”)
Airport and Plane(e.g. “parked”)
Ship and Helipad(e.g. “on, contains”)
+ +Table 4. Example of relationships between different objects used in the proposed instruction dataset. + +stances of a class in the entire dataset and assigning the $80^{th}$ percentile as the large label. Similarly, the $20^{th}$ percentile is designated as small size, with the remaining falling into the normal category. To determine the object's relative position within the images, we partition the entire image into a $3 \times 3$ grid, defining regions such as Top Right, Top, Top Left, Left, Center, Right, Bottom Right, Bottom Left, and Bottom. Based on the object's center pixel coordinates, we assign its relative position accordingly. + +To define the relation between objects in a given image, we group different objects based on their distance between the bounding boxes, and for each sub-graph, we assign different relationships between objects based on their class labels. Table 4 presents various examples of object relationships. To establish relationships like "surrounded by," we cross-reference pixel-level coordinates to verify if one object is entirely contained within another object. + +Expression Generation: To emulate natural language expressions, we employ predefined textual templates based on [40]. The phrase template encompasses the attributes $\{\mathrm{a1},\dots,\mathrm{a5}\}$ from Table 3. The expression for a group of + +
ModelUCMercedAID
Qwen-VL [1]62.9052.60
MiniGPTv2 [4]4.7612.90
LLaVA-1.5 [17]68.0051.00
GeoChat84.4372.03
+ +objects of the same class is formulated as: + +"The/A $\langle a3\rangle \langle a2\rangle a1\langle \mathrm{in / on the}a4\rangle$ + +Attributes that may be absent are enclosed in $\langle \rangle$ , and attributes $\{\mathrm{a2}, \mathrm{a3}\}$ can be arranged in any sequence. + +Similarly, the sentence template incorporates the relational attributes a5 to establish connections between two objects through this structure: + +"The/A $\langle a_i3\rangle \langle a_i2\rangle a_i1a_i5a_j1\langle \mathrm{in / on}$ the $a_{j}4\rangle$ + +Here, the indices $i$ and $j$ represent the $i^{th}$ and $j^{th}$ object. + +Visual Grounding: Although referring expression datasets are available in the natural image domain [37, 38], they lack for the remote sensing domain. To this end, we use our short descriptions as referring expressions to create three different kinds of question answering pairs, i.e. grounding image description, referring expression, and region level captioning, as described in Table 1. + +# 5. Experiments + +# 5.1. Implementation Details + +We initialize the weights of our model with the pretrained CLIP-ViT [24], and LLM (Vicuna-v1.5 [7] and apply LoRA [11] finetuning. Utilizing LoRA, we refine the parameters $W_{q}$ and $W_{v}$ through low-rank adaptation, with a designated rank $r$ set to 64 in our implementation. The model undergoes training consistently at an image resolution of $504 \times 504$ throughout the whole process. Each training step incorporates specifically crafted multi-modal instructional templates designed for a variety of vision-language tasks during the training process. We use AdamW [21] optimizer with a cosine learning rate scheduler to train our model. We keep the global batch size as 144. We train our model in two stages, first, we train using all of our datasets for 1 epoch, correspondingly 2144 steps, followed by stage 2, where we only train on the grounding dataset for 1600 more steps(at 128 batch size). + +# 5.2. Scene Classification + +Datasets for evaluation. For scene classification, we evaluate our model using AID [34] and UCMerced [36]. AID + +Table 5. Zero-shot scene classification accuracy comparison on AID [34] and UCMerced [36] datasets. In comparison to other generic VLMs, GeoChat performs favorably well. + +
MethodPresenceComparisonRural/UrbanAvg. Accuracy
LLaVA-1.5[17]55.4668.2059.0062.77
Qwen-vl-Chat [1]38.5767.5961.0055.35
MiniGPTv2 [4]55.1655.2239.0054.96
RSVQA[20]87.4781.5090.0086.32
EasyToHard[39]90.6687.4991.6789.94
Bi-Modal[2]91.0691.1692.6691.63
SHRNet [41]91.0390.4894.0091.84
RSGPT[12]91.1791.7094.0092.29
GeoChat91.0990.3394.0090.70
+ +Table 6. Comparisons with general zero-shot (top) and RS-VQA specialized (middle) models on RSVQA-LRBEN [20] dataset for VQA task. [1, 4, 17] are evaluated in zero-shot setting. GeoChat outperforms other zero-shot models and performs competitively to SoTA-supervised models like RSGPT which are specifically finetuned on target dataset (while ours is a generic model not specifically finetuned on target dataset). + +[34] is a large-scale aerial image collection compiled from Google Earth imagery, with 30 classes, such as a river, dense residential area, etc. The images are labeled by specialists in the field of remote sensing image interpretation. In total, the AID [34] dataset has 10,000 images within 30 classes. The images have been taken from different countries as well as different weather conditions. For evaluation, we use a $20\%$ split of the AID [34] dataset. UCMered [36] is a Land Use scene classification dataset, with 2,100 images and 21 classes. Each image is of size $256 \times 256$ . We use the whole UCMered [36] dataset as a zero-shot test set. + +Results. We prompt the models with all of the classes and prompt to classify the image using just one word/phrase. For example, we input a prompt like "Classify the image within one of the given classes: dense residential area, ..., school. Answer with one word or short phrase." We calculate zero-shot accuracy on both AID and UCMered. GeoChat significantly outperforms other VLM's with an accuracy of $84.43\%$ on UCMered [36] and $72.03\%$ on AID [34], as presented in Table 5. Notably, the recent MiniGPT-4-v2[4] fails to follow the instructions provided for this specific task and returns unrelated classes that are not a part of the dataset. It's accuracy is close to $5\%$ if we pass the answers from Vicuna-v1.5 [7] and ask it to check if the output sentence refers to the ground truth class or not. In comparison, Qwen-VL and LLaVa-1.5 perform well in instruction following, but fall short to GeoChat, due to lack of domain knowledge. + +# 5.3. Visual Question Answering + +Datasets for evaluation. RSVQA-HRBEN [20] comprises 10,569 high-resolution photos and 1,066,316 question-answer pairs, with $61.5\%$ , $11.2\%$ , $20.5\%$ , and $6.8\%$ divided into training, validation, test 1, and test 2 sets, respectively. This dataset has three question types: presence, + +
ModelSmallMediumLargeSingle-object groundingMulti-object grounding[refer][grounding]Overall
MiniGPTv2 [4]1.79.921.99.13.68.22.67.6
GeoChat2.913.621.716.04.310.511.810.6
+ +Table 7. Performance (acc@0.5%) comparison of GeoChat on our benchmark. Small, medium and large refer to the size of the objects based on the bounding box area. Single/multi-object refer to how many objects the question asks the model to predict. [refer]: object referenced using one attribute from a2, a3 or a4 in Table 3. [grounding]: objects referenced using a combination of attributes from a1-a5 in Table 3. Overall, GeoChat outperforms the baseline, but there is still significant room for further improvement on this complex task. + +
ModelPresenceComparisonAverage Accuracy
Qwen-VL[1]66.4460.4163.06
LLaVA-1.5[17]69.8367.2968.40
MiniGPTv2[4]40.7950.9146.46
GeoChat58.4583.1972.30
+ +comparison, and count. For evaluation, we use the test set-2 for RSVQA-HRBEN [20] with 47k question answer pairs. RSVQA-LR [20] is made up of 772 low-resolution images and 77,232 question-answer pairs, with $77.8\%$ , $11.1\%$ , and $11.1\%$ used for training, validation, and testing, respectively. There are four different categories of questions: presence, comparison, rural/urban, and count. We omitted area and count questions during evaluation because the responses are numerical and quantifiable into numerous categories. In the RSVQA-LRBEN [20] dataset, for example, counting questions are quantified into five categories: 0, between 1 and 10, between 11 and 100, between 101 and 1000, and greater than 1000. For evaluation, we use the test set of RSVQA-LRBEN [20] with 7k question-answer pairs. + +Results. To constrain the answers to a simple yes/no and for rural/urban question types, we add a suitable prompt at the end of each question. GeoChat performs close to the SOTA specialist models on RSVQA-LRBEN test set, which is RSGPT [12], finetuned on the target dataset for 5 iterations in comparison. For RSVQA-HRBEN, GeoChat outperforms other VLM's in zero-shot setting on average accuracy by $3.9\%$ , while beating the Comparison subset by $15.9\%$ on LLaVA-v1.5 [17], as shown in Table 8. + +# 5.4. Visual Grounding + +Datasets for evaluation. For the evaluation of grounding tasks, we propose a new benchmark that contains different referring and grounding tasks. We use the validation set from [31] and used the same dataset creation pipeline as in Sec. 4 to construct the test benchmark. There are a total of 7593 [refer], 560 [grounding], and 495 grounding description questions, as well as 2793 for region captioning. We use accuracy@0.5 as the evaluation metric. Accuracy is calculated if the predicted box has an overlap of more than + +Table 8. Comparison with other general ZS model's on RSVQA-HRBEN [20] dataset for visual qa. All models here have not been trained on the target dataset. GeoChat performs favorably well compared to generic VLMs. + +
Modelacc@0.5acc@.25METEOR
MiniGPTv2[4]10.830.916.4
GeoChat11.733.948.9
+ +Table 9. Results on grounding description task. + +
ModelROUGE-1ROUGE-LMETEOR
MiniGPTv2[4]32.131.210.0
GeoChat87.387.283.9
+ +Table 10. Region level captioning performance. + +# 0.5 IoU with the ground-truth box. + +Results. Table 7 shows the performance of our method and MiniGPT-4-v2 [4] on the proposed benchmark. Overall, the model performance is low on small objects or when it has to predict multiple boxes. Compared to MiniGPT-4-v2[4], our model works better on medium size images. On the grounding description task, we calculate both, the IoU for the multiple bounding boxes generated as well as the text answer generated. Our model provides a better description with slightly better box accuracy than MiniGPT-4-v2 [4] (Table 9). As for region-level captioning, we evaluate both models based on the text accuracy with ground truth region-level captions (Table 10). Our model significantly outperforms MiniGPT-4-v2 in terms of ROUGE and ME-TEOR score. + +# 6. Conclusion + +Although recent advancements in large Vision-Language Models (VLMs) have shown promise in nature image domains, their performance in Remote Sensing (RS) scenarios is still limited due to the unique domain-specific challenges. Addressing this gap, we present GeoChat, the first unified remote sensing VLM that excels in multitask conversational capabilities with high-resolution RS images. GeoChat not only answers image-level queries but also engages in region-specific dialogue, grounding responses with precise spatial coordinates. We create a novel RS multimodal instruction-following dataset comprising of $318k$ image-instruction pairs with a diverse multitask format. GeoChat achieves robust zero-shot performance across various RS tasks including scene classification, VQA, multi-turn dialogue, visual grounding and referring object detection, thus establishing a comprehensive benchmark. + +# References + +[1] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023.7, 8 +[2] Yakoub Bazi, Mohamad Mahmoud Al Rahhal, Mohamed Lamine Mekhalfi, Mansour Abdulaziz Al Zuair, and Farid Melgani. Bi-modal transformer-based approach for visual question answering in remote sensing imagery. IEEE Transactions on Geoscience and Remote Sensing, 60:1-11, 2022. 7 +[3] Christel Chappuis, Valérie Zermatten, Sylvain Lobry, Bertrand Le Saux, and Devis Tuia. Prompt-rsvqa: Prompting visual context to a language model for remote sensing visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1372–1381, 2022. 2 +[4] Jun Chen, Deyao Zhu, Xiaogian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023. 7, 8 +[5] Gong Cheng, Junwei Han, and Xiaoqiang Lu. Remote sensing image scene classification: Benchmark and state of the art. Proceedings of the IEEE, 105(10):1865-1883, 2017. 2, 3, 5 +[6] Gong Cheng, Jiabao Wang, Ke Li, Xingxing Xie, Chunbo Lang, Yanqing Yao, and Junwei Han. Anchor-free oriented proposal generator for object detection. IEEE Transactions on Geoscience and Remote Sensing, 60:1-11, 2022. 5 +[7] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ chatgpt quality, 2023. 2, 4, 5, 7 +[8] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning, 2023. 2 +[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 2 +[10] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 4 +[11] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 2, 4, 5, 7 +[12] Yuan Hu, Jianlong Yuan, Congcong Wen, Xiaonan Lu, and Xiang Li. Rsgpt: A remote sensing vision language model + +and benchmark. arXiv preprint arXiv:2307.15266, 2023. 1, 2, 7, 8 +[13] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726, 2023. 2 +[14] Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. Llava-med: Training a large language-and-vision assistant for biomedicine in one day. arXiv preprint arXiv:2306.00890, 2023. 2 +[15] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 5 +[16] Xiang Li, Congcong Wen, Yuan Hu, and Nan Zhou. Rs-clip: Zero shot remote sensing scene classification via contrastive vision-language supervision. International Journal of Applied Earth Observation and Geoinformation, 124:103497, 2023. 2 +[17] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 2, 3, 5, 7, 8 +[18] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 2 +[19] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 1, 2 +[20] Sylvain Lobry, Diego Marcos, Jesse Murray, and Devis Tuia. Rsvqa: Visual question answering for remote sensing data. IEEE Transactions on Geoscience and Remote Sensing, 58 (12):8555-8566, 2020. 2, 3, 5, 7, 8 +[21] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 7 +[22] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 2 +[23] OpenAI. Gpt-4 technical report, 2023. 1 +[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7 +[25] Maryam Rahnemoonfar, Tashnim Chowdhury, Argho Sarkar, Debvrat Varshney, Masoud Yari, and Robin Murphy. Floodnet: A high resolution aerial imagery dataset for post flood scene understanding. arXiv preprint arXiv:2012.02951, 2020. 3, 5 +[26] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 5 + +[27] Xian Sun, Peijin Wang, Zhiyuan Yan, Feng Xu, Ruiping Wang, Wenhui Diao, Jin Chen, Jihao Li, Yingchao Feng, Tao Xu, Martin Weinmann, Stefan Hinz, Cheng Wang, and Kun Fu. Fair1m: A benchmark dataset for fine-grained object recognition in high-resolution remote sensing imagery. IS-PRS Journal of Photogrammetry and Remote Sensing, 184: 116–130, 2022. 5 +[28] Yi Tay, Minh C Phan, Luu Anh Tuan, and Siu Cheung Hui. Learning to rank question answer pairs with holographic dual lstm architecture. In Proceedings of the 40th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 695-704. ACM, 2017. 3, 4, 5 +[29] Omkar Thawakar, Ashmal Vayani, Salman Khan, Hisham Cholakal, Rao M Anwer, Michael Felsberg, Tim Baldwin, Eric P Xing, and Fahad Shahbaz Khan. Mobillama: Towards accurate and lightweight fully transparent gpt. arXiv preprint arXiv:2402.16840, 2024. 2 +[30] Omkar Thawkar, Abdelrahman Shaker, Sahal Shaji Mullanlappilly, Hisham Cholakkal, Rao Muhammad Anwer, Salman Khan, Jorma Laaksonen, and Fahad Shahbaz Khan. Xraygpt: Chest radiographs summarization using large medical vision-language models. arXiv: 2306.07971, 2023. 2 +[31] Di Wang, Jing Zhang, Bo Du, Dacheng Tao, and Liangpei Zhang. Scaling-up remote sensing segmentation dataset with segment anything model. In arxiv, 2023. 2, 5, 6, 8 +[32] Di Wang, Qiming Zhang, Yufei Xu, Jing Zhang, Bo Du, Dacheng Tao, and Liangpei Zhang. Advancing plain vision transformer toward remote sensing foundation model. IEEE Transactions on Geoscience and Remote Sensing, 61:1-15, 2023. 6 +[33] Junjue Wang, Zhuo Zheng, Ailong Ma, Xiaoyan Lu, and Yanfei Zhong. LoveDA: A remote sensing land-cover dataset for domain adaptive semantic segmentation, 2021. 6 +[34] Gui-Song Xia, Jingwen Hu, Fan Hu, Baoguang Shi, Xiang Bai, Yanfei Zhong, Liangpei Zhang, and Xiaoqiang Lu. Aid: A benchmark data set for performance evaluation of aerial scene classification. IEEE Transactions on Geoscience and Remote Sensing, 55(7):3965-3981, 2017. 7 +[35] Gui-Song Xia, Xiang Bai, Jian Ding, Zhen Zhu, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, and Liangpei Zhang. Dota: A large-scale dataset for object detection in aerial images. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 5 +[36] Yi Yang and Shawn Newsam. Bag-of-visual-words and spatial extensions for land-use classification. In Proceedings of the 18th SIGSPATIAL international conference on advances in geographic information systems, pages 270–279, 2010. 7 +[37] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014. 7 +[38] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. 7 + +[39] Zhenghang Yuan, Lichao Mou, Qi Wang, and Xiao Xiang Zhu. From easy to hard: Learning language-guided curriculum for visual question answering on remote sensing data. IEEE Transactions on Geoscience and Remote Sensing, 60: 1-11, 2022. 2, 7 +[40] Yang Zhan, Zhitong Xiong, and Yuan Yuan. Rsvg: Exploring data and models for visual grounding on remote sensing data. IEEE Transactions on Geoscience and Remote Sensing, 61: 1-13, 2023. 6 +[41] Zixiao Zhang, Licheng Jiao, Lingling Li, Xu Liu, Puhua Chen, Fang Liu, Yuxuan Li, and Zhicheng Guo. A spatial hierarchical reasoning network for remote sensing visual question answering. IEEE Transactions on Geoscience and Remote Sensing, 61:1-15, 2023. 2, 7 +[42] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2 +[43] Usman Zia, M Mohsin Riaz, and Abdul Ghafoor. Transforming remote sensing images to textual descriptions. International Journal of Applied Earth Observation and Geoinformation, 108:102741, 2022. 2 \ No newline at end of file diff --git a/2024/GeoChat_ Grounded Large Vision-Language Model for Remote Sensing/images.zip b/2024/GeoChat_ Grounded Large Vision-Language Model for Remote Sensing/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..508fe6f1a84122cf48973c90392e2ddaefd55ffb --- /dev/null +++ b/2024/GeoChat_ Grounded Large Vision-Language Model for Remote Sensing/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5a9191e726cb33d78597223cea8311641b1e69dda5e6ff85773ad0862dffbf0e +size 535922 diff --git a/2024/GeoChat_ Grounded Large Vision-Language Model for Remote Sensing/layout.json b/2024/GeoChat_ Grounded Large Vision-Language Model for Remote Sensing/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..486080aa117b36ec96327ed4cc12bee40ded68fb --- /dev/null +++ b/2024/GeoChat_ Grounded Large Vision-Language Model for Remote Sensing/layout.json @@ -0,0 +1,7781 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 65, + 97, + 529, + 125 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 97, + 529, + 125 + ], + "spans": [ + { + "bbox": [ + 65, + 97, + 529, + 125 + ], + "type": "text", + "content": "GeoChat: Grounded Large Vision-Language Model for Remote Sensing" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 102, + 145, + 198, + 160 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 145, + 198, + 160 + ], + "spans": [ + { + "bbox": [ + 102, + 145, + 198, + 160 + ], + "type": "text", + "content": "Kartik Kuckreja" + }, + { + "bbox": [ + 102, + 145, + 198, + 160 + ], + "type": "inline_equation", + "content": "^{1,2*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 223, + 145, + 363, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 145, + 363, + 159 + ], + "spans": [ + { + "bbox": [ + 223, + 145, + 363, + 159 + ], + "type": "text", + "content": "Muhammad Sohail Danish" + }, + { + "bbox": [ + 223, + 145, + 363, + 159 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 392, + 146, + 490, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 146, + 490, + 159 + ], + "spans": [ + { + "bbox": [ + 392, + 146, + 490, + 159 + ], + "type": "text", + "content": "Muzammal Naseer1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 126, + 165, + 189, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 165, + 189, + 180 + ], + "spans": [ + { + "bbox": [ + 126, + 165, + 189, + 180 + ], + "type": "text", + "content": "Abhijit Das" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 236, + 165, + 316, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 165, + 316, + 178 + ], + "spans": [ + { + "bbox": [ + 236, + 165, + 316, + 178 + ], + "type": "text", + "content": "Salman Khan1,3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 350, + 165, + 466, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 165, + 466, + 178 + ], + "spans": [ + { + "bbox": [ + 350, + 165, + 466, + 178 + ], + "type": "text", + "content": "Fahad Shahbaz Khan" + }, + { + "bbox": [ + 350, + 165, + 466, + 178 + ], + "type": "inline_equation", + "content": "^{1,4}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 129, + 181, + 466, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 181, + 466, + 194 + ], + "spans": [ + { + "bbox": [ + 129, + 181, + 466, + 194 + ], + "type": "text", + "content": "1Mohamed bin Zayed University of AI, 2Birla Institute of Technology & Science, Hyderabad" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 196, + 196, + 397, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 196, + 397, + 209 + ], + "spans": [ + { + "bbox": [ + 196, + 196, + 397, + 209 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 196, + 196, + 397, + 209 + ], + "type": "text", + "content": "Australian National University, " + }, + { + "bbox": [ + 196, + 196, + 397, + 209 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 196, + 196, + 397, + 209 + ], + "type": "text", + "content": "Linköping University" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 163, + 211, + 424, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 163, + 211, + 424, + 221 + ], + "spans": [ + { + "bbox": [ + 163, + 211, + 424, + 221 + ], + "type": "text", + "content": "kartik.kuckreja@mbzuai.ac.ae, muhammad.sohail@mbzuai.ac.ae" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 143, + 250, + 191, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 250, + 191, + 262 + ], + "spans": [ + { + "bbox": [ + 143, + 250, + 191, + 262 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 273, + 289, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 273, + 289, + 645 + ], + "spans": [ + { + "bbox": [ + 46, + 273, + 289, + 645 + ], + "type": "text", + "content": "Recent advancements in Large Vision-Language Models (VLMs) have shown great promise in natural image domains, allowing users to hold a dialogue about given visual content. However, such general-domain VLMs perform poorly for Remote Sensing (RS) scenarios, leading to inaccurate or fabricated information when presented with RS domain-specific queries. Such a behavior emerges due to the unique challenges introduced by RS imagery. For example, to handle high-resolution RS imagery with diverse scale changes across categories and many small objects, region-level reasoning is necessary alongside holistic scene interpretation. Furthermore, the lack of domain-specific multimodal instruction following data as well as strong backbone models for RS make it hard for the models to align their behavior with user queries. To address these limitations, we propose GeoChat - the first versatile remote sensing VLM that offers multitask conversational capabilities with high-resolution RS images. Specifically, GeoChat can not only answer image-level queries but also accepts region inputs to hold region-specific dialogue. Furthermore, it can visually ground objects in its responses by referring to their spatial coordinates. To address the lack of domain-specific datasets, we generate a novel RS multimodal instruction-following dataset by extending imagetext pairs from existing diverse RS datasets. We establish a comprehensive benchmark for RS multitask conversations and compare with a number of baseline methods. GeoChat demonstrates robust zero-shot performance on various RS tasks, e.g., image and region captioning, visual question answering, scene classification, visually grounded conversations and referring detection. Our code is available here." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 655, + 128, + 667 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 655, + 128, + 667 + ], + "spans": [ + { + "bbox": [ + 47, + 655, + 128, + 667 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 670, + 287, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 670, + 287, + 695 + ], + "spans": [ + { + "bbox": [ + 46, + 670, + 287, + 695 + ], + "type": "text", + "content": "In the natural image domain, the abundance of aligned image-text data sourced from web imagery or manual anno-" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 307, + 247, + 545, + 450 + ], + "blocks": [ + { + "bbox": [ + 307, + 247, + 545, + 450 + ], + "lines": [ + { + "bbox": [ + 307, + 247, + 545, + 450 + ], + "spans": [ + { + "bbox": [ + 307, + 247, + 545, + 450 + ], + "type": "image", + "image_path": "0fad32ffbdf2710f079251f3c134ec5c87b4219e0774924ff853c756ca0a336c.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 460, + 545, + 549 + ], + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 549 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 549 + ], + "type": "text", + "content": "Figure 1. GeoChat can accomplish multiple tasks for remote-sensing (RS) image comprehension in a unified framework. Given suitable task tokens and user queries, the model can generate visually grounded responses (text with corresponding object locations - shown on top), visual question answering on images and regions (top left and bottom right, respectively) as well as scene classification (top right) and normal natural language conversations (bottom). This makes it the first RS VLM with grounding capability." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 578, + 545, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 578, + 545, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 578, + 545, + 686 + ], + "type": "text", + "content": "tations facilitate effective self-supervised vision-language modeling, as demonstrated by multimodal GPT-4 [23] and open-source initiatives like LLaVA [19]. These vision-language models (VLMs), developed through generative pretraining and instruction-tuning, exhibit robust zero-shot task completion across various user-oriented multimodal tasks. The resulting capabilities open the door to the development of versatile multimodal conversational assistants with broad applications in real-world scenarios [12]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": "However, general-domain VLMs designed for natural images, exhibit poor performance when presented with re" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 495, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 495, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 495, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 171, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 171, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 171, + 712 + ], + "type": "text", + "content": "*Equally contributing first authors." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "27831" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": "motely sensed visual imagery. The performance disparity arises primarily from the distinct nature of content found in remote sensing image-text pairings compared to the publicly available web data. As a result, general-domain VLMs can provide inaccurate information or hallucinate when presented with spatial images from RS sensors. Although there has been significant progress in the field of remote sensing visual question answering (VQA) [39, 41], earlier methods have framed the task as a classification problem. Here, the model chooses answers from predetermined responses found in the training data. It limits their applicability to open-ended answer generation and instruction-following." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 218, + 289, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 218, + 289, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 218, + 289, + 373 + ], + "type": "text", + "content": "In this paper, we introduce GeoChat, an attempt to extend multimodal instruction-tuning to the remote sensing domain for training a multitask conversational assistant. However, remote-sensing domain lacks a multimodal instruction-tuning conversational dataset. Inspired by recent work in instruction-tuning [14, 19, 29, 42], GeoChat uses Vicuna-v1.5 [7] and an automated pipeline to generate diverse remote sensing multimodal instruction-following data comprising of nearly " + }, + { + "bbox": [ + 46, + 218, + 289, + 373 + ], + "type": "inline_equation", + "content": "318k" + }, + { + "bbox": [ + 46, + 218, + 289, + 373 + ], + "type": "text", + "content": " instructions. We create the image-text pairs from various existing remote sensing datasets developed for diverse tasks. These include LRBEN for VQA [20], NWPU-RESISC-45 for scene classification [5] and SAMRS for object detection [31]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 375, + 289, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 375, + 289, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 375, + 289, + 518 + ], + "type": "text", + "content": "A crucial capability of GeoChat is the unification of multiple image and region-level reasoning tasks for RS imagery within a single pipeline (see Fig. 1). We achieve this via distinct task tokens that help suitably direct the model's responses according to user requirements. In addition, the model uses spatial location representations in its inputs to seamlessly reason about local regions and can also generate object locations in its responses to visually ground objects. This enables a diverse set of tasks possible with GeoChat including referring expression detection, image/region captioning, scene classification, natural language conversations and VQA, besides visually grounded conversations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 521, + 281, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 521, + 281, + 533 + ], + "spans": [ + { + "bbox": [ + 58, + 521, + 281, + 533 + ], + "type": "text", + "content": "In summary, this work has the following contributions:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 534, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 534, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 534, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 47, + 534, + 287, + 629 + ], + "type": "text", + "content": "- RS multimodal instruction following dataset. We present a novel data generation pipeline, to leverage existing object detection dataset [31] to create short descriptions of the images, followed by using Vicuna-v1.5 [7] to create conversations using the generated text alone. Further, we add visual question-answering and scene classification abilities using their corresponding datasets [5, 20]. This results in a total of " + }, + { + "bbox": [ + 47, + 534, + 287, + 629 + ], + "type": "inline_equation", + "content": "318k" + }, + { + "bbox": [ + 47, + 534, + 287, + 629 + ], + "type": "text", + "content": " instruction pairs for RS domain." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 630, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 630, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 630, + 288, + 714 + ], + "type": "text", + "content": "- GeoChat. Leveraging our dataset, we finetune LLaVA-1.5 [14] to create the remote sensing-domain vision-language model - GeoChat. Our LoRA [11] fine-tuning is efficient and avoids forgetting the necessary context embedded in fully-tuned LLaVA model, whose MLP projection is trained to align images into the word embedding space of the LLM (Vicuna-v1.5 [7]). This allows" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 545, + 108 + ], + "type": "text", + "content": "GeoChat to retain the conversation and instruction following abilities of LLaVA and extend its domain-knowledge to remote sensing tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 109, + 547, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 109, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 306, + 109, + 547, + 228 + ], + "type": "text", + "content": "- We also address the lack of evaluation benchmarks to assess the capability of existing VLMs on remote-sensing conversations. To this end, we setup evaluation protocols for conversation grounding in RS, as well as a setup a suite of tasks to allow comparisons with future efforts in this direction. We show various supervised as well as zero-shot evaluations for different remote sensing tasks, including image captioning, visual question answering and scene classification to demonstrate the generalisability of GeoChat conversational VLM." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 238, + 392, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 238, + 392, + 251 + ], + "spans": [ + { + "bbox": [ + 306, + 238, + 392, + 251 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 258, + 545, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 545, + 426 + ], + "type": "text", + "content": "Large Vision-Language Models. The typical architecture of instruction-following Vision Language Models (VLMs) consists of utilising a pre-trained visual backbone[9] to encode visual data, a large language model [7] for interpreting user instructions and generating responses, and a vision-language cross-modal connector, e.g., a linear projection layer [18, 42] or an MLP [17], for fusing visual information with language models. The results achieved with VLMs show great promise; for example, LLaVA [18], Instruct-BLIP [8], Otter [13] and MiniGPT-4 [42] show remarkable gains in language instruction following and visual reasoning ability for natural scenes. More recent studies have shown that these models can be adapted to other domains such as videos [22], biomedical [14, 30] and remote sensing [12]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 426, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 546, + 714 + ], + "type": "text", + "content": "Remote Sensing VLMs. The application of generalized VLMs in remote sensing is comparatively sparse. The majority of research so far has neglected the semantic understanding of the items and their relationships towards a deep visual comprehension. Beyond merely identifying the objects in an image, vision-language models are also capable of generating natural language descriptions of the image and inferring the connections between the objects. This makes them more appropriate for tasks like text-based image retrieval, captioning images, and answering visual questions that call for both visual and linguistic knowledge. Although there has been progress in vision language models for remote sensing tasks, such as image captioning [43], zero-shot classification [16] and visual question answering [3, 39], these models can only perform a specific task they are trained for, lack conversational capability and do not possess generic semantic knowledge about the remote sensing images. A major gap exists in the remote sensing domain towards developing general-purpose models to solve all tasks together, while also maintaining conversation abilities. While RSGPT [12] is an initial effort that has shown good conversation ability along with solving multiple tasks, it requires finetuning the model for each task separately, which makes it cumbersome and not generalizable. Fur" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "27832" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "ther, RSGPT cannot work for region-level reasoning or visual grounding, which our work aims to address." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 106, + 284, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 106, + 284, + 121 + ], + "spans": [ + { + "bbox": [ + 47, + 106, + 284, + 121 + ], + "type": "text", + "content": "3. GeoChat: Grounded Remote Sensing VLM" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 126, + 287, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 126, + 287, + 233 + ], + "spans": [ + { + "bbox": [ + 46, + 126, + 287, + 233 + ], + "type": "text", + "content": "Visually grounded conversations for remote sensing aim to generate textual responses interleaved with corresponding object locations. Further, a user can also provide visual prompts (e.g., a bounding box) besides natural language questions, and the model should be able to answer questions about the specified Region of Interest (RoI). Such seamless interplay between visual and language modalities necessitate a deep comprehension of linguistic constructions that denote particular objects or elements in a visual scene." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 234, + 287, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 234, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 46, + 234, + 287, + 353 + ], + "type": "text", + "content": "As mentioned above, GeoChat is the first model capable of holding visually grounded conversations about remotely sensed images. By construction, GeoChat can address not only the challenging task of visually grounded conversations, but can also perform a spectrum of other spatial reasoning tasks that span varying levels of granularity in visual imagery understanding e.g., image/region captioning, referring object detection and image/region-level conversations about remotely sensed images. We formally outline the tasks possible with GeoChat below." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 354, + 287, + 568 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 47, + 354, + 287, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 354, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 47, + 354, + 287, + 426 + ], + "type": "text", + "content": "a) Image-Level Conversation Tasks. In this task, GeoChat processes an image " + }, + { + "bbox": [ + 47, + 354, + 287, + 426 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 47, + 354, + 287, + 426 + ], + "type": "text", + "content": " and a user text query " + }, + { + "bbox": [ + 47, + 354, + 287, + 426 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 47, + 354, + 287, + 426 + ], + "type": "text", + "content": " without any specific spatial coordinates in its inputs or outputs. The goal is to perform conversation-based tasks at a holistic level with image-wide context, such as visual question answering (VQA), scene classification and image captioning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 426, + 287, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 426, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 47, + 426, + 287, + 496 + ], + "type": "text", + "content": "b) Region-Level Conversation Tasks. This task involves providing spatial box locations " + }, + { + "bbox": [ + 47, + 426, + 287, + 496 + ], + "type": "inline_equation", + "content": "\\pmb{b}" + }, + { + "bbox": [ + 47, + 426, + 287, + 496 + ], + "type": "text", + "content": " in the input to GeoChat besides " + }, + { + "bbox": [ + 47, + 426, + 287, + 496 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 47, + 426, + 287, + 496 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 426, + 287, + 496 + ], + "type": "inline_equation", + "content": "\\pmb{q}" + }, + { + "bbox": [ + 47, + 426, + 287, + 496 + ], + "type": "text", + "content": ". Region locations " + }, + { + "bbox": [ + 47, + 426, + 287, + 496 + ], + "type": "inline_equation", + "content": "\\pmb{b}" + }, + { + "bbox": [ + 47, + 426, + 287, + 496 + ], + "type": "text", + "content": " guide the model's attention to specific regions within the image, so that the model can perform tasks such as region-level captioning, region-specific VQA or multi-turn conversation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 497, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 497, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 47, + 497, + 287, + 568 + ], + "type": "text", + "content": "c) Grounded Conversation Tasks. With the use of special tokens, termed as task-specification tokens " + }, + { + "bbox": [ + 47, + 497, + 287, + 568 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 497, + 287, + 568 + ], + "type": "text", + "content": ", GeoChat can be guided to provide object locations at different granularities, while maintaining conversation abilities. It helps in tasks including grounded image captioning/conversation, object grounding and referring expression detection." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 575, + 174, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 174, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 174, + 586 + ], + "type": "text", + "content": "3.1. GeoChat Architecture" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": "GeoChat follows the architecture as of LLaVA-v1.5 [17], which consists of three core components, i) Global Image encoder, ii) an MLP adaptor (two linear layers) and iii) LLM. Different to LLaVA, we add specific task prompt that indicates the type of task desired from the model i.e., grounding, image-level or region-level conversations. Additionally, we allow spatial positions within both inputs and outputs, enabling visual prompts as inputs and grounded objects in GeoChat outputs. Notably, the original LLaVA model cannot perform object grounding or accept region" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 308, + 70, + 545, + 168 + ], + "blocks": [ + { + "bbox": [ + 308, + 70, + 545, + 168 + ], + "lines": [ + { + "bbox": [ + 308, + 70, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 308, + 70, + 545, + 168 + ], + "type": "table", + "html": "
DataSizeResponse formatting prompts
Detailed Description30kDescribe the image in detail.
Multi-Round Conversation65k-
Complex Questions10k-
RSVQA-LRBEN[20]56kAnswer the question using a single word or phrase.
NWPU-RESISC-45[5]31.5k
Floodnet[25]4k
Grounding Description25k-grounding] Describe the image in detail.
Region Captioning40k[identify] {bx_left, by_top, bx_right, by_bottom|θ}
Referring Expression45k[refer] < p > Object < /p>
", + "image_path": "f587cd81832e0824a88d1655bc82324be98073b3b330f859d46c2f43ef19435e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 172, + 545, + 205 + ], + "lines": [ + { + "bbox": [ + 305, + 172, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 305, + 172, + 545, + 205 + ], + "type": "text", + "content": "Table 1. Instruction following data used to train GeoChat. Instruction types and format are shown. We use a " + }, + { + "bbox": [ + 305, + 172, + 545, + 205 + ], + "type": "inline_equation", + "content": "308\\mathrm{k}" + }, + { + "bbox": [ + 305, + 172, + 545, + 205 + ], + "type": "text", + "content": " set for training and a separate " + }, + { + "bbox": [ + 305, + 172, + 545, + 205 + ], + "type": "inline_equation", + "content": "10\\mathrm{k}" + }, + { + "bbox": [ + 305, + 172, + 545, + 205 + ], + "type": "text", + "content": " instruction-set for testing." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 230, + 545, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 230, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 304, + 230, + 545, + 277 + ], + "type": "text", + "content": "inputs. Further, the original LLaVA can not reason about remote sensing images which is enabled via our domain-specific dataset. We describe each component in the architecture as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 280, + 545, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 280, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 304, + 280, + 545, + 459 + ], + "type": "text", + "content": "Task Token: The unique quality of GeoChat is its ability to easily switch between different types of remote sensing visual interpretation tasks. To eliminate uncertainty among tasks, our approach assigns a unique task identification to each one. We suggest three distinct task identities, " + }, + { + "bbox": [ + 304, + 280, + 545, + 459 + ], + "type": "inline_equation", + "content": "t \\in \\{ \\text{grounding}, \\text{identify}, \\text{refer} \\}" + }, + { + "bbox": [ + 304, + 280, + 545, + 459 + ], + "type": "text", + "content": ", each for grounded conversations, region captioning and referring expression comprehension. As for the case of visual question answering and scene classification, we directly ask the model to output the answer in a single word or phrase, as shown in Table 1. Our approach does not employ any task identification tokens for vision-irrelevant commands. This unified approach is supported by a modular design that efficiently integrates spatial data, giving the model flexibility in its reasoning about visual content." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 460, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 616 + ], + "type": "text", + "content": "Spatial Location Representation. Our model must precisely identify the spatial position of the referenced items for tasks such as grounded conversations, referring expression generation, and comprehension. To this end, we represent the box locations in a textual format to express the geographical position: " + }, + { + "bbox": [ + 304, + 460, + 545, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{b} = \\{b_{\\mathrm{x\\_left}}, b_{\\mathrm{y\\_top}}, b_{\\mathrm{x\\_right}}, b_{\\mathrm{y\\_bottom}} | \\theta \\}" + }, + { + "bbox": [ + 304, + 460, + 545, + 616 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 304, + 460, + 545, + 616 + ], + "type": "inline_equation", + "content": "b_{\\mathrm{x\\_left}}, b_{\\mathrm{y\\_top}}" + }, + { + "bbox": [ + 304, + 460, + 545, + 616 + ], + "type": "text", + "content": " denote the top left corner point of box while the " + }, + { + "bbox": [ + 304, + 460, + 545, + 616 + ], + "type": "inline_equation", + "content": "b_{\\mathrm{x\\_right}}, b_{\\mathrm{y\\_bottom}}" + }, + { + "bbox": [ + 304, + 460, + 545, + 616 + ], + "type": "text", + "content": " represent the bottom right corner coordinates. The angle " + }, + { + "bbox": [ + 304, + 460, + 545, + 616 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 460, + 545, + 616 + ], + "type": "text", + "content": " represents the angle of rotation for the bounding box, from the lower edge. Numerical values normalised within the interval [0, 100] are used to represent the x and y coordinates. Region locations in this format are used to interact with the model via its inputs and outputs." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Visual Backbone. GeoChat adapts the pretrained vision backbone of CLIP-ViT(L-14) [28], which has an input resolution of " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "336 \\times 336" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": ". This results in effectively 576 patches per image. Since this resolution is not sufficient to understand details presented in remote sensing imagery (e.g., small objects and object details), we interpolate the positional encoding in the transformer-based CLIP [28] model to scale with input image sizes of " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "504 \\times 504" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": ". Although this" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "type": "text", + "content": "27833" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 64, + 545, + 336 + ], + "blocks": [ + { + "bbox": [ + 49, + 64, + 545, + 336 + ], + "lines": [ + { + "bbox": [ + 49, + 64, + 545, + 336 + ], + "spans": [ + { + "bbox": [ + 49, + 64, + 545, + 336 + ], + "type": "image", + "image_path": "9a86cc66af1c5179ab8404b3f8d6a880419e922f34ee1d777345ae23c6edf16c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 342, + 548, + 410 + ], + "lines": [ + { + "bbox": [ + 46, + 342, + 548, + 410 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 548, + 410 + ], + "type": "text", + "content": "Figure 2. An overview of GeoChat - the first grounded large vision-language model for remote sensing. Given an image input together with a user query, a visual backbone is first used to encode patch-level tokens at a higher resolution via interpolating positional encodings. A multi-layer perceptron (MLP) is used to adapt vision-tokens to language space suitable for input to a Large Language Model (Vicuna 1.5). Besides visual inputs, region locations can also be input to the model together with task-specific prompts that specify the desired task required by the user. Given this context, the LLM can generate natural language responses interleaved with corresponding object locations. GeoChat can perform multiple tasks as shown on top e.g., scene classification, image/region captioning, VQA and grounded conversations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 326, + 451, + 525, + 463 + ], + "lines": [ + { + "bbox": [ + 326, + 451, + 525, + 463 + ], + "spans": [ + { + "bbox": [ + 326, + 451, + 525, + 463 + ], + "type": "text", + "content": "Figure 3. Multi-task instruction template for GeoChat." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 430, + 288, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 430, + 288, + 478 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 288, + 478 + ], + "type": "text", + "content": "leads to an increase in the number of patches to almost double (i.e., 1296 per image), this enhanced resolution allows us to handle larger image sizes and also supports better visual grounding in high-resolution RS images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 481, + 287, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 481, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 46, + 481, + 287, + 566 + ], + "type": "text", + "content": "MLP Cross-modal Adaptor. From the frozen CLIP-ViT[28], we project the output tokens " + }, + { + "bbox": [ + 46, + 481, + 287, + 566 + ], + "type": "inline_equation", + "content": "(\\in \\mathbb{R}^{1296\\times 1024})" + }, + { + "bbox": [ + 46, + 481, + 287, + 566 + ], + "type": "text", + "content": " with dimensions 1024 onto the language model space, using an MLP adaptor with one hidden layer. The adaptor has an input dimensionality of 1024 and outputs a vector of size 4096, corresponding to the input size of the LLM [7]. A GeLU [10] is used as the activation function." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 570, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 715 + ], + "type": "text", + "content": "Large Language Model. The open source Vicuna-v1.5(7B) [7] large language model is utilised as the foundation for GeoChat. The language model functions as a single interface for diverse vision-language inputs in our framework. To accomplish different vision-language tasks, we directly depend on the Vicuna-v1.5(7B) [7] language tokens. We explicitly interact with the language model to construct textual representations of bounding boxes to express their spatial coordinates for the visual grounding tasks that require the production of spatial locations. Similarly, the safe, aligned and effective behavior of LLM is ensured via system prompts appended together with given inputs. A" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 310, + 433, + 542, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 433, + 542, + 443 + ], + "spans": [ + { + "bbox": [ + 310, + 433, + 542, + 443 + ], + "type": "text", + "content": "[USER] Image Features [Task Identifier] [ASSISTANT]" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 483, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 545, + 639 + ], + "type": "text", + "content": "Low-Rank Adaptation (LoRA) [11] based strategy is used for fine-tuning the LLM. While training, instead of finetuning all of the weights that comprise the weight matrix of the pre-trained Vicuna-v1.5[7], we finetune two smaller matrices in LoRA [11] that approximate the original larger matrix. After that, the fine-tuned adaptor is fed into the pretrained model and utilised for inference. The LoRA adaptation ensures faster training and avoids forgetting original knowledge embedded in the LLM trained and fine-tuned on generic natural language instructions. This is an important feature since it allows the model to bring in external context about generic object types, landmarks and affordances in the remote-sensing reasoning framework of GeoChat." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 647, + 405, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 405, + 659 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 405, + 659 + ], + "type": "text", + "content": "3.2. Training Details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "To enhance the effectiveness of our model on general visual tasks and optimize training efficiency, we employ a strategy that involves initializing the network with pre-trained weights and fine-tuning specific segments for remote sens" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "27834" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 108, + 76, + 252, + 86 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 76, + 252, + 86 + ], + "spans": [ + { + "bbox": [ + 108, + 76, + 252, + 86 + ], + "type": "text", + "content": "Object Attributes and Relationships" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 59, + 86, + 282, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 86, + 282, + 107 + ], + "spans": [ + { + "bbox": [ + 59, + 86, + 282, + 107 + ], + "type": "text", + "content": "1. Ship {white,small,top-left}, {grey, small, bottom-right}\n2. Harbor {small, right}, {large, bottom-left}, {small, top right}, {small, left}\n3. Trees {top-right}." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 136, + 118, + 221, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 118, + 221, + 127 + ], + "spans": [ + { + "bbox": [ + 136, + 118, + 221, + 127 + ], + "type": "text", + "content": "Referring Expression" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 127, + 236, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 127, + 236, + 150 + ], + "spans": [ + { + "bbox": [ + 58, + 127, + 236, + 150 + ], + "type": "text", + "content": "1. Small Grey ship anchored at large harbor on bottom-right \n2. Some Trees at the top-right. \n3. White ship at the top of a small harbor on the right." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 144, + 158, + 215, + 168 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 158, + 215, + 168 + ], + "spans": [ + { + "bbox": [ + 144, + 158, + 215, + 168 + ], + "type": "text", + "content": "Region Captioning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 167, + 153, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 167, + 153, + 183 + ], + "spans": [ + { + "bbox": [ + 58, + 167, + 153, + 183 + ], + "type": "text", + "content": "1. Object located at 0.\n2. Size of object at 0 is small." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 171, + 188, + 190, + 205 + ], + "blocks": [ + { + "bbox": [ + 171, + 188, + 190, + 205 + ], + "lines": [ + { + "bbox": [ + 171, + 188, + 190, + 205 + ], + "spans": [ + { + "bbox": [ + 171, + 188, + 190, + 205 + ], + "type": "image", + "image_path": "655d531fc8c6e616e4a3e4f7329a204de544a52a8a859b2955f40efe6e05214e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 325, + 547, + 360 + ], + "lines": [ + { + "bbox": [ + 46, + 325, + 547, + 360 + ], + "spans": [ + { + "bbox": [ + 46, + 325, + 547, + 360 + ], + "type": "text", + "content": "Figure 4. Types of annotations available in the GeoChat instruction-set. Top-row: For a given RS image, we obtain object attribute and relationship information, referring expressions and region captions along with their corresponding region annotations (shown over the image). Bottom-row: This structured information is used to create the rich instruction-set with a total of 318k image-instruction pairs." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 171, + 204, + 190, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 204, + 190, + 212 + ], + "spans": [ + { + "bbox": [ + 171, + 204, + 190, + 212 + ], + "type": "text", + "content": "LLM" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 63, + 237, + 164, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 237, + 164, + 247 + ], + "spans": [ + { + "bbox": [ + 63, + 237, + 164, + 247 + ], + "type": "text", + "content": "MultiTurn Conversation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 60, + 247, + 164, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 247, + 164, + 262 + ], + "spans": [ + { + "bbox": [ + 60, + 247, + 164, + 262 + ], + "type": "text", + "content": "USER: Where is the white ship present?" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 61, + 262, + 164, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 262, + 164, + 277 + ], + "spans": [ + { + "bbox": [ + 61, + 262, + 164, + 277 + ], + "type": "text", + "content": "ASSISTANT: White ship is present at the top left." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 61, + 277, + 164, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 277, + 164, + 291 + ], + "spans": [ + { + "bbox": [ + 61, + 277, + 164, + 291 + ], + "type": "text", + "content": "USER: What is the color of the ship anchored at the harbor?" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 61, + 291, + 164, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 291, + 164, + 306 + ], + "spans": [ + { + "bbox": [ + 61, + 291, + 164, + 306 + ], + "type": "text", + "content": "ASSISTANT: The ship anchored at the harbor is grey in color." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 188, + 238, + 283, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 238, + 283, + 248 + ], + "spans": [ + { + "bbox": [ + 188, + 238, + 283, + 248 + ], + "type": "text", + "content": "Detailed Description" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 183, + 248, + 286, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 183, + 248, + 286, + 262 + ], + "spans": [ + { + "bbox": [ + 183, + 248, + 286, + 262 + ], + "type": "text", + "content": "USER: Please Describe the image in detail." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 183, + 262, + 286, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 183, + 262, + 286, + 314 + ], + "spans": [ + { + "bbox": [ + 183, + 262, + 286, + 314 + ], + "type": "text", + "content": "ASSISTANT: This satellite image shows depicts a river scene with multiple ships and harbors present. There are a total of two ships and four harbors. The grey ship at the bottom right is anchored to the harbor" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 238, + 402, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 238, + 402, + 248 + ], + "spans": [ + { + "bbox": [ + 312, + 238, + 402, + 248 + ], + "type": "text", + "content": "Complex Reasoning" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 248, + 408, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 248, + 408, + 263 + ], + "spans": [ + { + "bbox": [ + 305, + 248, + 408, + 263 + ], + "type": "text", + "content": "USER: What does the presence of 4 harbors and 2 ships indicate?" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 263, + 408, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 263, + 408, + 314 + ], + "spans": [ + { + "bbox": [ + 305, + 263, + 408, + 314 + ], + "type": "text", + "content": "ASSISTANT: The presence of 4 harbors and 2 ships constitute a bustling maritime environment. The presence of such harbors may indicate a water area with the possibility of export /imports of goods as well as tourism." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 430, + 238, + 531, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 430, + 238, + 531, + 247 + ], + "spans": [ + { + "bbox": [ + 430, + 238, + 531, + 247 + ], + "type": "text", + "content": "Grounded Description" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 428, + 247, + 531, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 428, + 247, + 531, + 262 + ], + "spans": [ + { + "bbox": [ + 428, + 247, + 531, + 262 + ], + "type": "text", + "content": "USER:[grounding] Describe the image in detail." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 427, + 262, + 531, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 427, + 262, + 531, + 314 + ], + "spans": [ + { + "bbox": [ + 427, + 262, + 531, + 314 + ], + "type": "text", + "content": "ASSISTANT: This satellite image depicts a grey ship anchored at a large harbor. There is another harbor at the top right, which is adjacent to some trees. There is another white ship close to a harbor at the right." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 380, + 287, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 380, + 287, + 476 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 287, + 476 + ], + "type": "text", + "content": "ing related tasks. We use a pre-trained CLIP-ViT(L-14) encoder[28], trained on large amounts of textual and visual data, a pretrained MLP adaptor[17], pretrained on a 558K subset of the LAION-CC-SBU [26] dataset with BLIP [15] captions, and Vicuna-v1.5[7] to initialize our model. To adapt our model to remote sensing images, we subsequently LoRA [11] fine-tune the LLM, while keeping the MLP adaptor and the CLIP encoder [28] frozen during training." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 47, + 488, + 243, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 488, + 243, + 501 + ], + "spans": [ + { + "bbox": [ + 47, + 488, + 243, + 501 + ], + "type": "text", + "content": "4. RS Multimodal Instruction Dataset" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 46, + 509, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 509, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 509, + 287, + 700 + ], + "type": "text", + "content": "By using LLM Vicuna [7], we align the model to follow a range of instructions by presenting and curating varied instruction-following data with multi-round conversations regarding remote sensing imagery (Table 1). We specifically provide system instructions as prompts that ask Vicuna [7] to generate multi-round question and answer pairs in a manner as if it could visualize the image (although it only has access to the text). This is achieved by providing few-shot in-context examples manually composed within the prompt to show Vicuna [7] how to build high-quality instruction-response pairs based on the caption and information supplied. Specifically, from our short descriptions created using the below pipeline, we randomly sample 65k images to create multi-round conversations, 10k images to generate complex question answers and 30k images to generate detailed descriptions for the given short descriptions." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 59, + 702, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 702, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 702, + 287, + 712 + ], + "type": "text", + "content": "In combination, after conversion to instruction format," + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 305, + 380, + 545, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 380, + 545, + 416 + ], + "spans": [ + { + "bbox": [ + 305, + 380, + 545, + 416 + ], + "type": "text", + "content": "we obtain a total of nearly 308k image-instruction pairs for training and 10k for testing. Next, we outline the instruction-set creation process." + } + ] + } + ], + "index": 27 + }, + { + "type": "table", + "bbox": [ + 308, + 419, + 545, + 480 + ], + "blocks": [ + { + "bbox": [ + 308, + 419, + 545, + 480 + ], + "lines": [ + { + "bbox": [ + 308, + 419, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 308, + 419, + 545, + 480 + ], + "type": "table", + "html": "
DatasetCategory# Classes# ImagesImage Size
DOTAObject Detection1817,4801024 × 1024
DIORObject Detection2023,463800 × 800
FAIR1MObject Detection3764,147600 × 600
LRBN(rsvqa)Visual Question Answering-600256 × 256
FloodnetVisual Question Answering-40563000 × 4000
NWPU-RESISC-45Scene Classification4531,500256 × 256
", + "image_path": "c2813674588d2be2e65b77e2c5998e124f7f745428369c37c6cb84bb8dfab396.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "table_body" + } + ], + "index": 28 + }, + { + "bbox": [ + 305, + 485, + 545, + 530 + ], + "lines": [ + { + "bbox": [ + 305, + 485, + 545, + 530 + ], + "spans": [ + { + "bbox": [ + 305, + 485, + 545, + 530 + ], + "type": "text", + "content": "Table 2. List of datasets used to creat our remote-sensing instruction set for GeoChat VLM training. We include object detection, visual question answering and scene classification datasets with varying image sizes and types of classes to ensure diversity." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 544, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 545, + 676 + ], + "type": "text", + "content": "Constituent Datasets: In the compilation of our instruction set, we incorporate three distinct types of datasets, encompassing the ones designed for object detection, scene classification, and visual question answering (VQA). Specifically, we integrate three object detection (DOTA [35], DIOR [6], and FAIR1M [27] which together form the SAMRS [31] dataset), one scene classification (NWPU-RESISC-45 [5]), one VQA (LRBEN[20]), and one flood detection [25] VQA dataset (see Table 2). The object detection datasets allow region-level reasoning capability as they offer segmentation masks along with bounding boxes." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "Addition of Missing Classes: Although a wide variety of object classes are included in the object detection databases, several essential categories like buildings, roads," + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "type": "text", + "content": "27835" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 64, + 545, + 274 + ], + "blocks": [ + { + "bbox": [ + 47, + 64, + 545, + 274 + ], + "lines": [ + { + "bbox": [ + 47, + 64, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 47, + 64, + 545, + 274 + ], + "type": "image", + "image_path": "fad3bd06df2560e997c705ad1a6eb1679b4aa307aacdc1d0f38132556a3b03ae.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 275, + 547, + 321 + ], + "lines": [ + { + "bbox": [ + 46, + 275, + 547, + 321 + ], + "spans": [ + { + "bbox": [ + 46, + 275, + 547, + 321 + ], + "type": "text", + "content": "Figure 5. Qualitative results of GeoChat. (left-right) Results are shown on grounding, referring object detection, and disaster/damage detection. The user can provide task-specific tokens (e.g., [grounding]) to shape model responses according to the desired behavior. The model can generate textual responses (right), only visual grounding (center) and both text and object groundings interleaved together (left). The model can also specify object types, object counts, object attributes and object relationships." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 330, + 281, + 407 + ], + "blocks": [ + { + "bbox": [ + 56, + 330, + 281, + 407 + ], + "lines": [ + { + "bbox": [ + 56, + 330, + 281, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 330, + 281, + 407 + ], + "type": "table", + "html": "
AttributeExample
a1category(e.g. “plane, ship”)
a2color(e.g. “gray, white”)
a3relative size(e.g. “small, large”)
a4relative location(e.g. “top right, bottom”)
a5relation(e.g. “parked at, driving through”)
", + "image_path": "037530172dc22d3b84a7c684fd516eedeb8d10b37f29a0c74df246c3163da7e0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 455, + 287, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 455, + 287, + 562 + ], + "spans": [ + { + "bbox": [ + 46, + 455, + 287, + 562 + ], + "type": "text", + "content": "and trees are missing. To address this, we propose to utilize ViTAE-RVSA [32] model, pre-trained on the LoveDA dataset [33], which encompasses the required important classes. The model [32] is used to infer these classes on the SAMRS [31] dataset, yielding pseudo labels. To mitigate potential noise in these predictions, we remove the predictions of ViTAE-RVSA [32] for which we already have ground truth from the SAMRS [31] dataset to refine the results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "Attribute extraction: For referring expression annotations, it is important to derive a variety of attributes in RS images. To this end, we have selected five distinct types of attributes, as outlined in Table 3. Object category information can be directly obtained from the SAMRS dataset. For color extraction, we use the K-Means clustering algorithm. Specifically, we extract the object's pixels from the image using ground-truth box and cluster them into " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " groups. The center of the largest cluster is then selected as the object's color. To specify the relative size of the object, we categorize objects into three sizes: small, normal, and large. This categorization is determined by measuring the area of all in" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 302, + 331, + 556, + 407 + ], + "blocks": [ + { + "bbox": [ + 46, + 411, + 287, + 434 + ], + "lines": [ + { + "bbox": [ + 46, + 411, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 287, + 434 + ], + "type": "text", + "content": "Table 3. List of attributes collected for objects. Attributes are used to obtain referring expressions e.g., small-sized plane to the left." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 302, + 331, + 556, + 407 + ], + "lines": [ + { + "bbox": [ + 302, + 331, + 556, + 407 + ], + "spans": [ + { + "bbox": [ + 302, + 331, + 556, + 407 + ], + "type": "table", + "html": "
CategoriesExample
Ships and Harbors(e.g. “anchored at, parked at”)
Track Field and Soccer Field(e.g. “Surrounded by, Inside”)
Vehicles, Bridge, Road, Roundabout(e.g. “passing through, passing through”)
Vehicles and Building(e.g. “parked”)
Airport and Plane(e.g. “parked”)
Ship and Helipad(e.g. “on, contains”)
", + "image_path": "96ea952af3d1f0d8b908e3aed53f81ea9e65e9820a383bf32fed8c97cdb2e3b1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 299, + 411, + 546, + 434 + ], + "lines": [ + { + "bbox": [ + 299, + 411, + 546, + 434 + ], + "spans": [ + { + "bbox": [ + 299, + 411, + 546, + 434 + ], + "type": "text", + "content": "Table 4. Example of relationships between different objects used in the proposed instruction dataset." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 455, + 545, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 455, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 304, + 455, + 545, + 563 + ], + "type": "text", + "content": "stances of a class in the entire dataset and assigning the " + }, + { + "bbox": [ + 304, + 455, + 545, + 563 + ], + "type": "inline_equation", + "content": "80^{th}" + }, + { + "bbox": [ + 304, + 455, + 545, + 563 + ], + "type": "text", + "content": " percentile as the large label. Similarly, the " + }, + { + "bbox": [ + 304, + 455, + 545, + 563 + ], + "type": "inline_equation", + "content": "20^{th}" + }, + { + "bbox": [ + 304, + 455, + 545, + 563 + ], + "type": "text", + "content": " percentile is designated as small size, with the remaining falling into the normal category. To determine the object's relative position within the images, we partition the entire image into a " + }, + { + "bbox": [ + 304, + 455, + 545, + 563 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 304, + 455, + 545, + 563 + ], + "type": "text", + "content": " grid, defining regions such as Top Right, Top, Top Left, Left, Center, Right, Bottom Right, Bottom Left, and Bottom. Based on the object's center pixel coordinates, we assign its relative position accordingly." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 567, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 567, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 567, + 545, + 662 + ], + "type": "text", + "content": "To define the relation between objects in a given image, we group different objects based on their distance between the bounding boxes, and for each sub-graph, we assign different relationships between objects based on their class labels. Table 4 presents various examples of object relationships. To establish relationships like \"surrounded by,\" we cross-reference pixel-level coordinates to verify if one object is entirely contained within another object." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 666, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 546, + 715 + ], + "type": "text", + "content": "Expression Generation: To emulate natural language expressions, we employ predefined textual templates based on [40]. The phrase template encompasses the attributes " + }, + { + "bbox": [ + 304, + 666, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{a1},\\dots,\\mathrm{a5}\\}" + }, + { + "bbox": [ + 304, + 666, + 546, + 715 + ], + "type": "text", + "content": " from Table 3. The expression for a group of" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "text", + "content": "27836" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 77, + 70, + 261, + 136 + ], + "blocks": [ + { + "bbox": [ + 77, + 70, + 261, + 136 + ], + "lines": [ + { + "bbox": [ + 77, + 70, + 261, + 136 + ], + "spans": [ + { + "bbox": [ + 77, + 70, + 261, + 136 + ], + "type": "table", + "html": "
ModelUCMercedAID
Qwen-VL [1]62.9052.60
MiniGPTv2 [4]4.7612.90
LLaVA-1.5 [17]68.0051.00
GeoChat84.4372.03
", + "image_path": "7a900cdabd9c0b027ea7c4cd5f158720f7a9ef20d4a0ae24ca66e4ef4e2dc794.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 194, + 218, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 194, + 218, + 206 + ], + "spans": [ + { + "bbox": [ + 47, + 194, + 218, + 206 + ], + "type": "text", + "content": "objects of the same class is formulated as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 215, + 242, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 215, + 242, + 228 + ], + "spans": [ + { + "bbox": [ + 89, + 215, + 242, + 228 + ], + "type": "text", + "content": "\"The/A " + }, + { + "bbox": [ + 89, + 215, + 242, + 228 + ], + "type": "inline_equation", + "content": "\\langle a3\\rangle \\langle a2\\rangle a1\\langle \\mathrm{in / on the}a4\\rangle" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 237, + 287, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 237, + 287, + 261 + ], + "spans": [ + { + "bbox": [ + 47, + 237, + 287, + 261 + ], + "type": "text", + "content": "Attributes that may be absent are enclosed in " + }, + { + "bbox": [ + 47, + 237, + 287, + 261 + ], + "type": "inline_equation", + "content": "\\langle \\rangle" + }, + { + "bbox": [ + 47, + 237, + 287, + 261 + ], + "type": "text", + "content": ", and attributes " + }, + { + "bbox": [ + 47, + 237, + 287, + 261 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{a2}, \\mathrm{a3}\\}" + }, + { + "bbox": [ + 47, + 237, + 287, + 261 + ], + "type": "text", + "content": " can be arranged in any sequence." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 261, + 287, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 261, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 47, + 261, + 287, + 297 + ], + "type": "text", + "content": "Similarly, the sentence template incorporates the relational attributes a5 to establish connections between two objects through this structure:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 305, + 265, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 305, + 265, + 319 + ], + "spans": [ + { + "bbox": [ + 69, + 305, + 265, + 319 + ], + "type": "text", + "content": "\"The/A " + }, + { + "bbox": [ + 69, + 305, + 265, + 319 + ], + "type": "inline_equation", + "content": "\\langle a_i3\\rangle \\langle a_i2\\rangle a_i1a_i5a_j1\\langle \\mathrm{in / on}" + }, + { + "bbox": [ + 69, + 305, + 265, + 319 + ], + "type": "text", + "content": " the " + }, + { + "bbox": [ + 69, + 305, + 265, + 319 + ], + "type": "inline_equation", + "content": "a_{j}4\\rangle" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 328, + 279, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 328, + 279, + 341 + ], + "spans": [ + { + "bbox": [ + 47, + 328, + 279, + 341 + ], + "type": "text", + "content": "Here, the indices " + }, + { + "bbox": [ + 47, + 328, + 279, + 341 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 328, + 279, + 341 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 328, + 279, + 341 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 328, + 279, + 341 + ], + "type": "text", + "content": " represent the " + }, + { + "bbox": [ + 47, + 328, + 279, + 341 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 47, + 328, + 279, + 341 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 328, + 279, + 341 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 47, + 328, + 279, + 341 + ], + "type": "text", + "content": " object." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 341, + 287, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 341, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 46, + 341, + 287, + 424 + ], + "type": "text", + "content": "Visual Grounding: Although referring expression datasets are available in the natural image domain [37, 38], they lack for the remote sensing domain. To this end, we use our short descriptions as referring expressions to create three different kinds of question answering pairs, i.e. grounding image description, referring expression, and region level captioning, as described in Table 1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 434, + 128, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 434, + 128, + 448 + ], + "spans": [ + { + "bbox": [ + 47, + 434, + 128, + 448 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 453, + 180, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 453, + 180, + 466 + ], + "spans": [ + { + "bbox": [ + 47, + 453, + 180, + 466 + ], + "type": "text", + "content": "5.1. Implementation Details" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 472, + 287, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 663 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 663 + ], + "type": "text", + "content": "We initialize the weights of our model with the pretrained CLIP-ViT [24], and LLM (Vicuna-v1.5 [7] and apply LoRA [11] finetuning. Utilizing LoRA, we refine the parameters " + }, + { + "bbox": [ + 46, + 472, + 287, + 663 + ], + "type": "inline_equation", + "content": "W_{q}" + }, + { + "bbox": [ + 46, + 472, + 287, + 663 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 472, + 287, + 663 + ], + "type": "inline_equation", + "content": "W_{v}" + }, + { + "bbox": [ + 46, + 472, + 287, + 663 + ], + "type": "text", + "content": " through low-rank adaptation, with a designated rank " + }, + { + "bbox": [ + 46, + 472, + 287, + 663 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 472, + 287, + 663 + ], + "type": "text", + "content": " set to 64 in our implementation. The model undergoes training consistently at an image resolution of " + }, + { + "bbox": [ + 46, + 472, + 287, + 663 + ], + "type": "inline_equation", + "content": "504 \\times 504" + }, + { + "bbox": [ + 46, + 472, + 287, + 663 + ], + "type": "text", + "content": " throughout the whole process. Each training step incorporates specifically crafted multi-modal instructional templates designed for a variety of vision-language tasks during the training process. We use AdamW [21] optimizer with a cosine learning rate scheduler to train our model. We keep the global batch size as 144. We train our model in two stages, first, we train using all of our datasets for 1 epoch, correspondingly 2144 steps, followed by stage 2, where we only train on the grounding dataset for 1600 more steps(at 128 batch size)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 671, + 162, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 162, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 162, + 683 + ], + "type": "text", + "content": "5.2. Scene Classification" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Datasets for evaluation. For scene classification, we evaluate our model using AID [34] and UCMerced [36]. AID" + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 309, + 70, + 545, + 188 + ], + "blocks": [ + { + "bbox": [ + 47, + 140, + 287, + 174 + ], + "lines": [ + { + "bbox": [ + 47, + 140, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 47, + 140, + 287, + 174 + ], + "type": "text", + "content": "Table 5. Zero-shot scene classification accuracy comparison on AID [34] and UCMerced [36] datasets. In comparison to other generic VLMs, GeoChat performs favorably well." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 70, + 545, + 188 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 545, + 188 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 545, + 188 + ], + "type": "table", + "html": "
MethodPresenceComparisonRural/UrbanAvg. Accuracy
LLaVA-1.5[17]55.4668.2059.0062.77
Qwen-vl-Chat [1]38.5767.5961.0055.35
MiniGPTv2 [4]55.1655.2239.0054.96
RSVQA[20]87.4781.5090.0086.32
EasyToHard[39]90.6687.4991.6789.94
Bi-Modal[2]91.0691.1692.6691.63
SHRNet [41]91.0390.4894.0091.84
RSGPT[12]91.1791.7094.0092.29
GeoChat91.0990.3394.0090.70
", + "image_path": "36c293b1cd4ac72e6d6fe8235141d2ec77a1e516dd90ad690061f4bac0826512.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 192, + 545, + 269 + ], + "lines": [ + { + "bbox": [ + 304, + 192, + 545, + 269 + ], + "spans": [ + { + "bbox": [ + 304, + 192, + 545, + 269 + ], + "type": "text", + "content": "Table 6. Comparisons with general zero-shot (top) and RS-VQA specialized (middle) models on RSVQA-LRBEN [20] dataset for VQA task. [1, 4, 17] are evaluated in zero-shot setting. GeoChat outperforms other zero-shot models and performs competitively to SoTA-supervised models like RSGPT which are specifically finetuned on target dataset (while ours is a generic model not specifically finetuned on target dataset)." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 291, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 291, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 291, + 545, + 422 + ], + "type": "text", + "content": "[34] is a large-scale aerial image collection compiled from Google Earth imagery, with 30 classes, such as a river, dense residential area, etc. The images are labeled by specialists in the field of remote sensing image interpretation. In total, the AID [34] dataset has 10,000 images within 30 classes. The images have been taken from different countries as well as different weather conditions. For evaluation, we use a " + }, + { + "bbox": [ + 304, + 291, + 545, + 422 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 291, + 545, + 422 + ], + "type": "text", + "content": " split of the AID [34] dataset. UCMered [36] is a Land Use scene classification dataset, with 2,100 images and 21 classes. Each image is of size " + }, + { + "bbox": [ + 304, + 291, + 545, + 422 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 291, + 545, + 422 + ], + "type": "text", + "content": ". We use the whole UCMered [36] dataset as a zero-shot test set." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 423, + 545, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 545, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 545, + 626 + ], + "type": "text", + "content": "Results. We prompt the models with all of the classes and prompt to classify the image using just one word/phrase. For example, we input a prompt like \"Classify the image within one of the given classes: dense residential area, ..., school. Answer with one word or short phrase.\" We calculate zero-shot accuracy on both AID and UCMered. GeoChat significantly outperforms other VLM's with an accuracy of " + }, + { + "bbox": [ + 304, + 423, + 545, + 626 + ], + "type": "inline_equation", + "content": "84.43\\%" + }, + { + "bbox": [ + 304, + 423, + 545, + 626 + ], + "type": "text", + "content": " on UCMered [36] and " + }, + { + "bbox": [ + 304, + 423, + 545, + 626 + ], + "type": "inline_equation", + "content": "72.03\\%" + }, + { + "bbox": [ + 304, + 423, + 545, + 626 + ], + "type": "text", + "content": " on AID [34], as presented in Table 5. Notably, the recent MiniGPT-4-v2[4] fails to follow the instructions provided for this specific task and returns unrelated classes that are not a part of the dataset. It's accuracy is close to " + }, + { + "bbox": [ + 304, + 423, + 545, + 626 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 304, + 423, + 545, + 626 + ], + "type": "text", + "content": " if we pass the answers from Vicuna-v1.5 [7] and ask it to check if the output sentence refers to the ground truth class or not. In comparison, Qwen-VL and LLaVa-1.5 perform well in instruction following, but fall short to GeoChat, due to lack of domain knowledge." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 635, + 456, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 456, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 456, + 647 + ], + "type": "text", + "content": "5.3. Visual Question Answering" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "Datasets for evaluation. RSVQA-HRBEN [20] comprises 10,569 high-resolution photos and 1,066,316 question-answer pairs, with " + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "61.5\\%" + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "11.2\\%" + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "20.5\\%" + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "6.8\\%" + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": " divided into training, validation, test 1, and test 2 sets, respectively. This dataset has three question types: presence," + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "type": "text", + "content": "27837" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 547, + 118 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 547, + 118 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 547, + 118 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 547, + 118 + ], + "type": "table", + "html": "
ModelSmallMediumLargeSingle-object groundingMulti-object grounding[refer][grounding]Overall
MiniGPTv2 [4]1.79.921.99.13.68.22.67.6
GeoChat2.913.621.716.04.310.511.810.6
", + "image_path": "24e985e1d28a6e8cce537c52c82d7a0ef8671c8021e534263692304c8c953689.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 173, + 293, + 243 + ], + "blocks": [ + { + "bbox": [ + 46, + 121, + 547, + 167 + ], + "lines": [ + { + "bbox": [ + 46, + 121, + 547, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 547, + 167 + ], + "type": "text", + "content": "Table 7. Performance (acc@0.5%) comparison of GeoChat on our benchmark. Small, medium and large refer to the size of the objects based on the bounding box area. Single/multi-object refer to how many objects the question asks the model to predict. [refer]: object referenced using one attribute from a2, a3 or a4 in Table 3. [grounding]: objects referenced using a combination of attributes from a1-a5 in Table 3. Overall, GeoChat outperforms the baseline, but there is still significant room for further improvement on this complex task." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 173, + 293, + 243 + ], + "lines": [ + { + "bbox": [ + 50, + 173, + 293, + 243 + ], + "spans": [ + { + "bbox": [ + 50, + 173, + 293, + 243 + ], + "type": "table", + "html": "
ModelPresenceComparisonAverage Accuracy
Qwen-VL[1]66.4460.4163.06
LLaVA-1.5[17]69.8367.2968.40
MiniGPTv2[4]40.7950.9146.46
GeoChat58.4583.1972.30
", + "image_path": "8329b583f3df5d019df7463856cddba8d8e2907f938ea1e23984fd86724ec666.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 303, + 287, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 303, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 46, + 303, + 287, + 470 + ], + "type": "text", + "content": "comparison, and count. For evaluation, we use the test set-2 for RSVQA-HRBEN [20] with 47k question answer pairs. RSVQA-LR [20] is made up of 772 low-resolution images and 77,232 question-answer pairs, with " + }, + { + "bbox": [ + 46, + 303, + 287, + 470 + ], + "type": "inline_equation", + "content": "77.8\\%" + }, + { + "bbox": [ + 46, + 303, + 287, + 470 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 303, + 287, + 470 + ], + "type": "inline_equation", + "content": "11.1\\%" + }, + { + "bbox": [ + 46, + 303, + 287, + 470 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 303, + 287, + 470 + ], + "type": "inline_equation", + "content": "11.1\\%" + }, + { + "bbox": [ + 46, + 303, + 287, + 470 + ], + "type": "text", + "content": " used for training, validation, and testing, respectively. There are four different categories of questions: presence, comparison, rural/urban, and count. We omitted area and count questions during evaluation because the responses are numerical and quantifiable into numerous categories. In the RSVQA-LRBEN [20] dataset, for example, counting questions are quantified into five categories: 0, between 1 and 10, between 11 and 100, between 101 and 1000, and greater than 1000. For evaluation, we use the test set of RSVQA-LRBEN [20] with 7k question-answer pairs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 471, + 287, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 471, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 46, + 471, + 287, + 578 + ], + "type": "text", + "content": "Results. To constrain the answers to a simple yes/no and for rural/urban question types, we add a suitable prompt at the end of each question. GeoChat performs close to the SOTA specialist models on RSVQA-LRBEN test set, which is RSGPT [12], finetuned on the target dataset for 5 iterations in comparison. For RSVQA-HRBEN, GeoChat outperforms other VLM's in zero-shot setting on average accuracy by " + }, + { + "bbox": [ + 46, + 471, + 287, + 578 + ], + "type": "inline_equation", + "content": "3.9\\%" + }, + { + "bbox": [ + 46, + 471, + 287, + 578 + ], + "type": "text", + "content": ", while beating the Comparison subset by " + }, + { + "bbox": [ + 46, + 471, + 287, + 578 + ], + "type": "inline_equation", + "content": "15.9\\%" + }, + { + "bbox": [ + 46, + 471, + 287, + 578 + ], + "type": "text", + "content": " on LLaVA-v1.5 [17], as shown in Table 8." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 586, + 154, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 154, + 601 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 154, + 601 + ], + "type": "text", + "content": "5.4. Visual Grounding" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": "Datasets for evaluation. For the evaluation of grounding tasks, we propose a new benchmark that contains different referring and grounding tasks. We use the validation set from [31] and used the same dataset creation pipeline as in Sec. 4 to construct the test benchmark. There are a total of 7593 [refer], 560 [grounding], and 495 grounding description questions, as well as 2793 for region captioning. We use accuracy@0.5 as the evaluation metric. Accuracy is calculated if the predicted box has an overlap of more than" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 320, + 173, + 533, + 217 + ], + "blocks": [ + { + "bbox": [ + 46, + 247, + 288, + 291 + ], + "lines": [ + { + "bbox": [ + 46, + 247, + 288, + 291 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 288, + 291 + ], + "type": "text", + "content": "Table 8. Comparison with other general ZS model's on RSVQA-HRBEN [20] dataset for visual qa. All models here have not been trained on the target dataset. GeoChat performs favorably well compared to generic VLMs." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 320, + 173, + 533, + 217 + ], + "lines": [ + { + "bbox": [ + 320, + 173, + 533, + 217 + ], + "spans": [ + { + "bbox": [ + 320, + 173, + 533, + 217 + ], + "type": "table", + "html": "
Modelacc@0.5acc@.25METEOR
MiniGPTv2[4]10.830.916.4
GeoChat11.733.948.9
", + "image_path": "ba56e70d7759fd1218718272fd7496c711f7aa03dd0009a2821c32e33381e836.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 321, + 237, + 533, + 281 + ], + "blocks": [ + { + "bbox": [ + 339, + 217, + 512, + 228 + ], + "lines": [ + { + "bbox": [ + 339, + 217, + 512, + 228 + ], + "spans": [ + { + "bbox": [ + 339, + 217, + 512, + 228 + ], + "type": "text", + "content": "Table 9. Results on grounding description task." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 237, + 533, + 281 + ], + "lines": [ + { + "bbox": [ + 321, + 237, + 533, + 281 + ], + "spans": [ + { + "bbox": [ + 321, + 237, + 533, + 281 + ], + "type": "table", + "html": "
ModelROUGE-1ROUGE-LMETEOR
MiniGPTv2[4]32.131.210.0
GeoChat87.387.283.9
", + "image_path": "6fa7b205c5b8655e9709242f8c75ef7872633b98c1c0d635eb832cd58b0e5cd3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 339, + 285, + 512, + 297 + ], + "lines": [ + { + "bbox": [ + 339, + 285, + 512, + 297 + ], + "spans": [ + { + "bbox": [ + 339, + 285, + 512, + 297 + ], + "type": "text", + "content": "Table 10. Region level captioning performance." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 309, + 447, + 321 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 309, + 447, + 321 + ], + "spans": [ + { + "bbox": [ + 306, + 309, + 447, + 321 + ], + "type": "text", + "content": "0.5 IoU with the ground-truth box." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 322, + 545, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 322, + 545, + 489 + ], + "spans": [ + { + "bbox": [ + 304, + 322, + 545, + 489 + ], + "type": "text", + "content": "Results. Table 7 shows the performance of our method and MiniGPT-4-v2 [4] on the proposed benchmark. Overall, the model performance is low on small objects or when it has to predict multiple boxes. Compared to MiniGPT-4-v2[4], our model works better on medium size images. On the grounding description task, we calculate both, the IoU for the multiple bounding boxes generated as well as the text answer generated. Our model provides a better description with slightly better box accuracy than MiniGPT-4-v2 [4] (Table 9). As for region-level captioning, we evaluate both models based on the text accuracy with ground truth region-level captions (Table 10). Our model significantly outperforms MiniGPT-4-v2 in terms of ROUGE and ME-TEOR score." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 501, + 379, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 501, + 379, + 514 + ], + "spans": [ + { + "bbox": [ + 306, + 501, + 379, + 514 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": "Although recent advancements in large Vision-Language Models (VLMs) have shown promise in nature image domains, their performance in Remote Sensing (RS) scenarios is still limited due to the unique domain-specific challenges. Addressing this gap, we present GeoChat, the first unified remote sensing VLM that excels in multitask conversational capabilities with high-resolution RS images. GeoChat not only answers image-level queries but also engages in region-specific dialogue, grounding responses with precise spatial coordinates. We create a novel RS multimodal instruction-following dataset comprising of " + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "inline_equation", + "content": "318k" + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": " image-instruction pairs with a diverse multitask format. GeoChat achieves robust zero-shot performance across various RS tasks including scene classification, VQA, multi-turn dialogue, visual grounding and referring object detection, thus establishing a comprehensive benchmark." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "text", + "content": "27838" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 144 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 144 + ], + "type": "text", + "content": "[1] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023.7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 288, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 288, + 212 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 288, + 212 + ], + "type": "text", + "content": "[2] Yakoub Bazi, Mohamad Mahmoud Al Rahhal, Mohamed Lamine Mekhalfi, Mansour Abdulaziz Al Zuair, and Farid Melgani. Bi-modal transformer-based approach for visual question answering in remote sensing imagery. IEEE Transactions on Geoscience and Remote Sensing, 60:1-11, 2022. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 213, + 287, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 213, + 287, + 278 + ], + "spans": [ + { + "bbox": [ + 53, + 213, + 287, + 278 + ], + "type": "text", + "content": "[3] Christel Chappuis, Valérie Zermatten, Sylvain Lobry, Bertrand Le Saux, and Devis Tuia. Prompt-rsvqa: Prompting visual context to a language model for remote sensing visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1372–1381, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 280, + 287, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 280, + 287, + 345 + ], + "spans": [ + { + "bbox": [ + 53, + 280, + 287, + 345 + ], + "type": "text", + "content": "[4] Jun Chen, Deyao Zhu, Xiaogian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023. 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 346, + 287, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 287, + 388 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 287, + 388 + ], + "type": "text", + "content": "[5] Gong Cheng, Junwei Han, and Xiaoqiang Lu. Remote sensing image scene classification: Benchmark and state of the art. Proceedings of the IEEE, 105(10):1865-1883, 2017. 2, 3, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 391, + 287, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 391, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 391, + 287, + 434 + ], + "type": "text", + "content": "[6] Gong Cheng, Jiabao Wang, Ke Li, Xingxing Xie, Chunbo Lang, Yanqing Yao, and Junwei Han. Anchor-free oriented proposal generator for object detection. IEEE Transactions on Geoscience and Remote Sensing, 60:1-11, 2022. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 436, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 436, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 53, + 436, + 287, + 491 + ], + "type": "text", + "content": "[7] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 53, + 436, + 287, + 491 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 53, + 436, + 287, + 491 + ], + "type": "text", + "content": " chatgpt quality, 2023. 2, 4, 5, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 491, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 491, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 53, + 491, + 287, + 544 + ], + "type": "text", + "content": "[8] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 546, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 546, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 53, + 546, + 287, + 612 + ], + "type": "text", + "content": "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 613, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 635 + ], + "type": "text", + "content": "[10] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "type": "text", + "content": "[11] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 2, 4, 5, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "text", + "content": "[12] Yuan Hu, Jianlong Yuan, Congcong Wen, Xiaonan Lu, and Xiang Li. Rsgpt: A remote sensing vision language model" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "and benchmark. arXiv preprint arXiv:2307.15266, 2023. 1, 2, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 97, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 97, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 97, + 545, + 140 + ], + "type": "text", + "content": "[13] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 143, + 545, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 143, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 307, + 143, + 545, + 198 + ], + "type": "text", + "content": "[14] Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. Llava-med: Training a large language-and-vision assistant for biomedicine in one day. arXiv preprint arXiv:2306.00890, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 199, + 545, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 199, + 545, + 253 + ], + "spans": [ + { + "bbox": [ + 308, + 199, + 545, + 253 + ], + "type": "text", + "content": "[15] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 256, + 545, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 256, + 545, + 310 + ], + "spans": [ + { + "bbox": [ + 308, + 256, + 545, + 310 + ], + "type": "text", + "content": "[16] Xiang Li, Congcong Wen, Yuan Hu, and Nan Zhou. Rs-clip: Zero shot remote sensing scene classification via contrastive vision-language supervision. International Journal of Applied Earth Observation and Geoinformation, 124:103497, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 312, + 545, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 312, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 307, + 312, + 545, + 346 + ], + "type": "text", + "content": "[17] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 2, 3, 5, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 347, + 545, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 347, + 545, + 379 + ], + "spans": [ + { + "bbox": [ + 308, + 347, + 545, + 379 + ], + "type": "text", + "content": "[18] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 381, + 545, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 381, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 308, + 381, + 545, + 404 + ], + "type": "text", + "content": "[19] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 406, + 545, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 406, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 308, + 406, + 545, + 449 + ], + "type": "text", + "content": "[20] Sylvain Lobry, Diego Marcos, Jesse Murray, and Devis Tuia. Rsvqa: Visual question answering for remote sensing data. IEEE Transactions on Geoscience and Remote Sensing, 58 (12):8555-8566, 2020. 2, 3, 5, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 452, + 545, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 452, + 545, + 474 + ], + "spans": [ + { + "bbox": [ + 307, + 452, + 545, + 474 + ], + "type": "text", + "content": "[21] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 475, + 545, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 475, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 308, + 475, + 545, + 520 + ], + "type": "text", + "content": "[22] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 521, + 474, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 521, + 474, + 533 + ], + "spans": [ + { + "bbox": [ + 308, + 521, + 474, + 533 + ], + "type": "text", + "content": "[23] OpenAI. Gpt-4 technical report, 2023. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 534, + 545, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 534, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 308, + 534, + 545, + 599 + ], + "type": "text", + "content": "[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 601, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 601, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 308, + 601, + 545, + 656 + ], + "type": "text", + "content": "[25] Maryam Rahnemoonfar, Tashnim Chowdhury, Argho Sarkar, Debvrat Varshney, Masoud Yari, and Robin Murphy. Floodnet: A high resolution aerial imagery dataset for post flood scene understanding. arXiv preprint arXiv:2012.02951, 2020. 3, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "text", + "content": "[26] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 5" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "text", + "content": "27839" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 148 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 148 + ], + "type": "text", + "content": "[27] Xian Sun, Peijin Wang, Zhiyuan Yan, Feng Xu, Ruiping Wang, Wenhui Diao, Jin Chen, Jihao Li, Yingchao Feng, Tao Xu, Martin Weinmann, Stefan Hinz, Cheng Wang, and Kun Fu. Fair1m: A benchmark dataset for fine-grained object recognition in high-resolution remote sensing imagery. IS-PRS Journal of Photogrammetry and Remote Sensing, 184: 116–130, 2022. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 149, + 288, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 149, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 149, + 288, + 205 + ], + "type": "text", + "content": "[28] Yi Tay, Minh C Phan, Luu Anh Tuan, and Siu Cheung Hui. Learning to rank question answer pairs with holographic dual lstm architecture. In Proceedings of the 40th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 695-704. ACM, 2017. 3, 4, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 205, + 288, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 205, + 288, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 205, + 288, + 259 + ], + "type": "text", + "content": "[29] Omkar Thawakar, Ashmal Vayani, Salman Khan, Hisham Cholakal, Rao M Anwer, Michael Felsberg, Tim Baldwin, Eric P Xing, and Fahad Shahbaz Khan. Mobillama: Towards accurate and lightweight fully transparent gpt. arXiv preprint arXiv:2402.16840, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 260, + 288, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 260, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 48, + 260, + 288, + 316 + ], + "type": "text", + "content": "[30] Omkar Thawkar, Abdelrahman Shaker, Sahal Shaji Mullanlappilly, Hisham Cholakkal, Rao Muhammad Anwer, Salman Khan, Jorma Laaksonen, and Fahad Shahbaz Khan. Xraygpt: Chest radiographs summarization using large medical vision-language models. arXiv: 2306.07971, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 316, + 288, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 288, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 288, + 348 + ], + "type": "text", + "content": "[31] Di Wang, Jing Zhang, Bo Du, Dacheng Tao, and Liangpei Zhang. Scaling-up remote sensing segmentation dataset with segment anything model. In arxiv, 2023. 2, 5, 6, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 349, + 288, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 349, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 349, + 288, + 403 + ], + "type": "text", + "content": "[32] Di Wang, Qiming Zhang, Yufei Xu, Jing Zhang, Bo Du, Dacheng Tao, and Liangpei Zhang. Advancing plain vision transformer toward remote sensing foundation model. IEEE Transactions on Geoscience and Remote Sensing, 61:1-15, 2023. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 404, + 288, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 288, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 288, + 437 + ], + "type": "text", + "content": "[33] Junjue Wang, Zhuo Zheng, Ailong Ma, Xiaoyan Lu, and Yanfei Zhong. LoveDA: A remote sensing land-cover dataset for domain adaptive semantic segmentation, 2021. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 437, + 288, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 288, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 288, + 491 + ], + "type": "text", + "content": "[34] Gui-Song Xia, Jingwen Hu, Fan Hu, Baoguang Shi, Xiang Bai, Yanfei Zhong, Liangpei Zhang, and Xiaoqiang Lu. Aid: A benchmark data set for performance evaluation of aerial scene classification. IEEE Transactions on Geoscience and Remote Sensing, 55(7):3965-3981, 2017. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 492, + 288, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 492, + 288, + 547 + ], + "spans": [ + { + "bbox": [ + 48, + 492, + 288, + 547 + ], + "type": "text", + "content": "[35] Gui-Song Xia, Xiang Bai, Jian Ding, Zhen Zhu, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, and Liangpei Zhang. Dota: A large-scale dataset for object detection in aerial images. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 548, + 288, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 288, + 592 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 288, + 592 + ], + "type": "text", + "content": "[36] Yi Yang and Shawn Newsam. Bag-of-visual-words and spatial extensions for land-use classification. In Proceedings of the 18th SIGSPATIAL international conference on advances in geographic information systems, pages 270–279, 2010. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 593, + 288, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 593, + 288, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 593, + 288, + 646 + ], + "type": "text", + "content": "[37] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 647, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 712 + ], + "type": "text", + "content": "[38] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. 7" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 318 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 127 + ], + "type": "text", + "content": "[39] Zhenghang Yuan, Lichao Mou, Qi Wang, and Xiao Xiang Zhu. From easy to hard: Learning language-guided curriculum for visual question answering on remote sensing data. IEEE Transactions on Geoscience and Remote Sensing, 60: 1-11, 2022. 2, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 129, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 172 + ], + "type": "text", + "content": "[40] Yang Zhan, Zhitong Xiong, and Yuan Yuan. Rsvg: Exploring data and models for visual grounding on remote sensing data. IEEE Transactions on Geoscience and Remote Sensing, 61: 1-13, 2023. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 174, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 174, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 174, + 545, + 228 + ], + "type": "text", + "content": "[41] Zixiao Zhang, Licheng Jiao, Lingling Li, Xu Liu, Puhua Chen, Fang Liu, Yuxuan Li, and Zhicheng Guo. A spatial hierarchical reasoning network for remote sensing visual question answering. IEEE Transactions on Geoscience and Remote Sensing, 61:1-15, 2023. 2, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 229, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 229, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 229, + 545, + 274 + ], + "type": "text", + "content": "[42] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 274, + 545, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 274, + 545, + 318 + ], + "spans": [ + { + "bbox": [ + 308, + 274, + 545, + 318 + ], + "type": "text", + "content": "[43] Usman Zia, M Mohsin Riaz, and Abdul Ghafoor. Transforming remote sensing images to textual descriptions. International Journal of Applied Earth Observation and Geoinformation, 108:102741, 2022. 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 734, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 734, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 734, + 318, + 742 + ], + "type": "text", + "content": "27840" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/GeoReF_ Geometric Alignment Across Shape Variation for Category-level Object Pose Refinement/966fabe9-a4db-4529-94bf-c710cc01e232_content_list.json b/2024/GeoReF_ Geometric Alignment Across Shape Variation for Category-level Object Pose Refinement/966fabe9-a4db-4529-94bf-c710cc01e232_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8a96f7f90927dc6c0f53d0d2a63be7aad2da2f5b --- /dev/null +++ b/2024/GeoReF_ Geometric Alignment Across Shape Variation for Category-level Object Pose Refinement/966fabe9-a4db-4529-94bf-c710cc01e232_content_list.json @@ -0,0 +1,1461 @@ +[ + { + "type": "text", + "text": "GeoReF: Geometric Alignment Across Shape Variation for Category-level Object Pose Refinement", + "text_level": 1, + "bbox": [ + 112, + 128, + 856, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Linfang Zheng $^{1,3}$ Tze Ho Elden Tse $^{*3}$ Chen Wang $^{*1,2}$ Yinghan Sun $^{1}$ Hua Chen $^{1}$ \nAles Leonardis $^{3}$ Wei Zhang $^{\\dagger 1}$ Hyung Jin Chang $^{3}$", + "bbox": [ + 94, + 200, + 872, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Shenzhen Key Laboratory of Control Theory and Intelligent Systems, School of System Design and Intelligent Manufacturing, Southern University of Science and Technology, China $^{2}$ Department of Computer Science, the University of Hong Kong, China $^{3}$ School of Computer Science, University of Birmingham, UK", + "bbox": [ + 88, + 255, + 880, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{1xz948, txt994}@student.bham.ac.uk, cwang5@cs.hku.hk, sunyh2021@mail.sustech.edu.cn {chenh6,zhangw3}@sustech.edu.cn,{a.leonadis,h.j.chang}@bham.ac.uk", + "bbox": [ + 112, + 329, + 854, + 362 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 396, + 313, + 412 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Object pose refinement is essential for robust object pose estimation. Previous work has made significant progress towards instance-level object pose refinement. Yet, category-level pose refinement is a more challenging problem due to large shape variations within a category and the discrepancies between the target object and the shape prior. To address these challenges, we introduce a novel architecture for category-level object pose refinement. Our approach integrates an HS-layer and learnable affine transformations, which aims to enhance the extraction and alignment of geometric information. Additionally, we introduce a crosscloud transformation mechanism that efficiently merges diverse data sources. Finally, we push the limits of our model by incorporating the shape prior information for translation and size error prediction. We conducted extensive experiments to demonstrate the effectiveness of the proposed framework. Through extensive quantitative experiments, we demonstrate significant improvement over the baseline method by a large margin across all metrics.", + "bbox": [ + 75, + 428, + 473, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 737, + 209, + 752 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Understanding an object's pose is crucial for a wide range of real-world applications, including robotic manipulation [18, 32, 52, 56], augmented reality [30, 33], and autonomous driving [19, 39]. Significant progress has been made for object pose estimation [5, 11, 15, 48, 53, 55] and pose refinement [2, 17, 22, 37, 51, 60] using the object's CAD", + "bbox": [ + 75, + 762, + 468, + 853 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/02dd435b036b416f9fac1ef5e657fb95dac3b3fcce496b602397162ee4ebbb56.jpg", + "image_caption": [ + "Camera" + ], + "image_footnote": [], + "bbox": [ + 586, + 396, + 666, + 454 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ab27819a1505b161a313cf61b448f90c0d8171b32f08746fe09be9bfa4341d2f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 687, + 396, + 763, + 452 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/64d9494ff06a4439fe985d710e0f582c261f07e4f6005a01240c2a0772aa711c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 787, + 397, + 870, + 452 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8b40e3ca3f05cb9dc6c6ca2b50af538cac51cef53f44afa8a4133c1dc1e5317f.jpg", + "image_caption": [ + "Mug", + "SP-m" + ], + "image_footnote": [], + "bbox": [ + 596, + 462, + 653, + 522 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e80710786aca18753def72fbe24cac96a512d803c7960d8f1c244b34adc9403f.jpg", + "image_caption": [ + "SP-1", + "Figure 1. Illustration of the shape variation. $SP - m$ represents the category's mean shape, $SP - 1$ and $SP - 2$ represents the randomly sampled object shapes from the CAMERA25 training set." + ], + "image_footnote": [], + "bbox": [ + 689, + 462, + 763, + 522 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2f7f16aba523964b2ffe813e816347b5f1bd6101bf27957cf85a0edede474ba2.jpg", + "image_caption": [ + "SP-2" + ], + "image_footnote": [], + "bbox": [ + 792, + 462, + 866, + 522 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "model. Despite the promising performance, the reliance on accurate instance-level CAD models limits their generalizability to everyday objects. Category-level methods [7, 23-25, 29, 59, 61] is therefore been proposed to overcome this limitation. The objective of this line of work focuses on estimating object poses within a category given category-level shape priors. As a result, they face unique challenges as there exist diverse shape variations in each object category. We illustrate these shape variations in Fig. 1.", + "bbox": [ + 496, + 597, + 890, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, there have been remarkable advancements in category-level object pose estimation [9, 10, 25, 58, 61], primarily due to effective utilization of geometric information through 3D graph convolution [26]. In applications that require high precision, it is common to employ an object pose refinement procedure in conjunction with pose estimation. This involves an initial pose estimation algorithm determining the object pose, followed by a refinement step to further enhance the accuracy of the initially estimated pose by predicting and correcting its error. However, while instance-level object pose refinement has been extensively studied,", + "bbox": [ + 496, + 734, + 892, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution, order by dice rolling.", + "bbox": [ + 94, + 862, + 318, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "The corresponding author.", + "bbox": [ + 96, + 875, + 241, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1Project page: https://lynne-zheng-linfang.github.io/georef.github.io", + "bbox": [ + 96, + 887, + 455, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10693", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "category-level pose refinement remained unexplored until the introduction of CATRE [29]. By leveraging initial object pose and size estimations, CATRE achieves category-level pose refinement by iteratively aligning the observed target object point cloud with the category-level shape prior. This pipeline is shown to be effective by improving the accuracy of the initial pose and size estimations.", + "bbox": [ + 75, + 90, + 468, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While CATRE has proven to be effective in many scenarios, it is limited by the reliance on the PointNet [35] encoder, which is primarily designed for classification and segmentation tasks. This design choice limits its ability to capture essential and fine-grained geometric relationships for accurate pose estimation and refinement. This ability is particularly important in category-level pose refinement as there exists diverse shape variations between inputs. Consequently, CATRE obtains suboptimal performance by direct application of 3D graph convolution. In addition, as CATRE treats the point cloud and the shape prior features separately until a later stage of the network, they potentially miss out on the benefits of integrating these features earlier. Moreover, their approach does not incorporate shape prior information into the translation and scale estimation module, which presents another area for potential improvement.", + "bbox": [ + 75, + 199, + 470, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we introduce a novel architecture for category-level object pose refinement which aims to address the limitations mentioned above. To better extract both local and global geometric information, we incorporate an HS layer into our feature extraction process. We apply learnable affine transformations to the features to address the geometric discrepancies between the observed point cloud and the shape prior. This enables the network to align these features more effectively. In addition, we propose a cross-cloud transformation mechanism that is specifically designed to enhance the merging of information between the observed point clouds and the shape prior. This mechanism enables more efficient integration of information between the two sources. Finally, we push the limit of our model by incorporating shape prior information to more accurately predict errors in translation and size estimation.", + "bbox": [ + 75, + 444, + 468, + 685 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our extensive experimental results on two category-level object pose datasets demonstrate that our proposed model to be effective in addressing the problem of shape variations in category-level object pose refinement, and consequently outperforms the state-of-the-art significantly. To the best of our knowledge, our proposed method is the first to successfully address the shape variation issue which is common in category-level pose refinement. Specifically, to enable graph convolution to be effective in capturing geometric relationships between different shapes, we propose an adaptive affine transformation matrix that aligns the observed point clouds and the shape prior. Additionally, the proposed cross-cloud transformation mechanism effectively fuses features from different input point clouds and brings", + "bbox": [ + 75, + 688, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "further performance improvements.", + "bbox": [ + 500, + 90, + 736, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are as follows:", + "bbox": [ + 517, + 107, + 738, + 119 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a novel architecture to specifically address the shape variations issue in category-level object pose refinement. Our proposed method results in consistent performance gain and exhibits better generalization ability.", + "- We propose a unique cross-cloud transformation mechanism which efficiently merges diverse information from observed point clouds and shape priors.", + "- We conduct extensive experiments on two category-level object pose datasets to validate our proposed method. On the REAL275 dataset, our method significantly outperforms SPD by $39.1\\%$ increase in the $5^{\\circ}5\\mathrm{cm}$ metric. Additionally, we achieve $10.5\\%$ improvement in the $10^{\\circ}2\\mathrm{cm}$ metric over the state-of-the-art method, CATRE." + ], + "bbox": [ + 500, + 122, + 890, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 332, + 640, + 348 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Instance-level object pose estimation and refinement. Instance-level approaches estimate the pose of the target object given known 3D CAD models. They can be briefly divided into correspondence matching methods and template matching methods. Correspondence matching methods [4, 8, 12, 13, 31, 36, 40, 41, 44, 46, 55, 62] matches the outstanding features of the observed object images with its model. Template matching methods [1, 14, 27, 34, 38, 42, 43, 47] compares the images or extracted features with the pre-generated templates. As the initial pose estimates can be noisy to various factors such as occlusions, object pose refinement [20, 22, 51] is shown to be useful in improving the performance of instance-level methods. Even though they achieved impressive over the target object, the reliance on object CAD models limited their generalizability for handling everyday objects. In this paper, we consider a more challenging problem setting where only the category-level shape prior is provided.", + "bbox": [ + 496, + 358, + 890, + 630 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Category-level object pose estimation and refinement. Both tasks mainly focus on addressing the shape variation between the objects. The pioneering work NOCS [49] tackles the shape discrepancy by recovering the normalized visible shape of the target object and achieving the pose by point cloud matching. A series of methods extend this structure by leveraging different information such as domain adaptation [21], different reconstruction space [3], shape prior [3, 16, 21, 45], and structural similarities [6, 25]. However, this line of work is often limited in speed due to the iterative point matching. Another series of work starts with FS-Net [9], which adopts 3D graph convolution (3D-GC) [26] to obtain geometric sensitivity. Due to its effectiveness and real-time performance, graph convolution is widely adopted in recent methods with an enhancement in directions including loss function [10], bounding box voting [58], and shape deformation [59]. HS-Pose [61] ex", + "bbox": [ + 496, + 643, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10694", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tends the geometric feature extraction from local to global, which enhances the capability to handle objects with complex shapes. The research on category-level refinement began recently with the proposal of CATRE [29]. It introduced an effective pipeline that leverages shape priors and a focalization strategy for pose refinement and effectively improves the initial pose estimations. In this paper, we extend the CATRE and tackle the geometric variation issue within the framework of category-level pose refinement.", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 239, + 210, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 76, + 263, + 279, + 279 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we tackle the problem of category-level object pose refinement. Given the initial pose and size estimation $(\\mathbf{R}_0,\\mathbf{t}_0,\\mathbf{s}_0)$ , along with the observed point cloud $\\mathcal{O}\\in \\mathbb{R}^{N^O\\times 3}$ and the shape prior $\\mathcal{P}\\in \\mathbb{R}^{N^P\\times 3}$ , we aim to predict the estimation error $(\\Delta \\mathbf{R},\\Delta \\mathbf{t},\\Delta \\mathbf{s})$ between the initial estimations and the ground truths. The pose refinement algorithm $\\phi$ can be described as:", + "bbox": [ + 75, + 287, + 468, + 396 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\Delta \\mathbf {R}, \\Delta \\mathbf {t}, \\Delta \\mathbf {s}\\right) = \\phi \\left(\\mathbf {R} _ {0}, \\mathbf {t} _ {0}, \\mathbf {s} _ {0}, \\mathcal {O}, \\mathcal {P}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 407, + 468, + 424 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This pose refinement algorithm $\\phi$ can be applied iteratively to improve the refinement performance.", + "bbox": [ + 75, + 433, + 468, + 464 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 472, + 217, + 487 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our proposed category-level object pose refinement framework builds upon two previous works, CATRE [29] and HS-layer [61], which we briefly review them in the following.", + "bbox": [ + 75, + 496, + 468, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CATRE. CATRE is the first framework that considers the problem of category-level pose refinement. It predicts the error between the ground truth and the estimated poses by aligning the input point clouds and the categorical shape priors. Specifically, the network architecture of CATRE consists of four components: a) point cloud focalization, b) shared encoder, c) rotation prediction, and d) translation and size prediction. In point clouds focalization, the observed point clouds $\\mathcal{O}$ and the shape prior $\\mathcal{P}$ are first aligned with the initial pose and size estimation $[\\mathbf{R}_0, \\mathbf{t}_0, \\mathbf{s}_0]$ :", + "bbox": [ + 75, + 556, + 468, + 708 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\hat {\\mathcal {O}} = \\left\\{\\hat {o} _ {i} | \\hat {o} _ {i} = o _ {i} - \\mathbf {t} _ {0}, o _ {i} \\in \\mathcal {O} \\right\\}, \\\\ \\hat {\\mathcal {P}} = \\left\\{\\hat {p} _ {i} | \\hat {p} _ {i} = \\operatorname {d i a g} \\left(\\mathbf {s} _ {0}\\right) \\mathbf {R} _ {0} p _ {i}, p _ {i} \\in \\mathcal {P} \\right\\}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 715, + 401, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathrm{diag}(\\cdot)$ converts a vector to a diagonal matrix. The focalized observed point cloud $\\hat{\\mathcal{O}}$ and the focalized shape prior points $\\hat{\\mathcal{P}}$ contain full information required to predict the estimation error $(\\Delta \\mathbf{R},\\Delta \\mathbf{t},\\Delta \\mathbf{s})$ . First, a PointNet-based shared encoder is used to extract features from the two focalized point clouds independently. Then, both the extracted features are used for $\\Delta \\mathbf{R}$ estimation, while the global feature of the focalized observed point cloud along with the $\\mathbf{s}_0$ are used for $\\Delta \\mathbf{t}$ and $\\Delta \\mathbf{s}$ . The initial estimates", + "bbox": [ + 75, + 763, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "are updated using the predicted error $(\\Delta \\mathbf{R}, \\Delta \\mathbf{t}, \\Delta \\mathbf{s})$ . Finally, the updated estimates are used to predict the error again in which this process is iterative and the estimations are refined progressively and continuously.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "HS-layer. The Hybrid-Scope Geometric Feature Extraction Layer (HS-layer) is a simple network structure based on 3D graph convolution. It consists of two parallel paths that extract different scopes of features from the point cloud. The first path encodes the size and translation information of the target object. Meanwhile, the second path extracts outlier-robust local and global geometric features by applying graph convolution with a strategy of Receptive Fields with Feature Distance (RF-F) metric, alongside an Outlier Robust Feature Extraction Layer (ORL). These properties are particularly beneficial for category-level object pose estimation tasks. For more details, please refer to [61].", + "bbox": [ + 496, + 157, + 890, + 339 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Overall Structure of GeoReF", + "text_level": 1, + "bbox": [ + 498, + 347, + 758, + 363 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall framework of our proposed object pose refinement approach is shown in Fig. 2. This framework comprises three principal components: 1) point cloud focalization, 2) feature extraction, and 3) pose error prediction. We follow CATRE and use the same point cloud focalization module. We apply focalization and extract features from both the observed point cloud and the shape prior by our Feature Extraction component. Then, we predict the estimation errors $(\\Delta \\mathbf{R}, \\Delta t, \\Delta s)$ using the extracted features in the pose error prediction component.", + "bbox": [ + 496, + 369, + 890, + 522 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.4. Graph Convolution with Learnable Affine Transformation (LAT)", + "text_level": 1, + "bbox": [ + 498, + 530, + 890, + 560 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Geometric structural information is effective in estimating an object's pose for category-level object pose estimations. However, as shown in ablation study [AS-1], directly applying the 3D graph convolutions (e.g., HS-Encoder [61] and 3DGCN-Encoder [26]) to category-level object pose refinement tasks results in poor performance. This is due to the differences in task nature between the pose estimation and the pose refinement. In the pose estimation task, the network only needs to extract geometric structural information from a single point cloud. However, in the pose refinement framework, it requires extracting the geometric structural information from the two input point clouds as well as establishing the geometric correspondences between different object shapes. This becomes challenging due to the issue of shape variation in category pose refinement.", + "bbox": [ + 496, + 568, + 890, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address the aforementioned problem, we propose to use learnable affine transformations (LATs). By employing LATs, the network can dynamically adjust the input point cloud and the point features which enables better establishment of geometric correspondences between the two different input shapes. Specifically, we apply three LATs (as shown in the bottom left of Fig. 2, where the Matrix Net", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10695", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/039b1a85f9060f9d5bce9cd1f04401193aeeb56fc69c4c463c05cae237fe2de1.jpg", + "image_caption": [ + "Figure 2. Overall structure of the proposed method. Our object pose refinement structure contains three main modules. Given the shape prior point cloud, the target object's observed point cloud, and the initial estimation, we first apply point cloud focalization on the input point clouds using the initial estimation. The focalized point clouds then go through a geometric-based feature extraction encoder to obtain geometric structural features. The extracted features are then fed into two branches for rotation error estimation, translation error, and size error estimation. Within the HS Feature Extractor, the Matrix Net models output the learnable affine transformations (LATs) for adaptive point and feature adjustment. The left output of the Matrix Net adjusts the input point clouds, while the right Matrix Net model outputs two affine transformations for adjusting the rotation features, and the translation and size features." + ], + "image_footnote": [], + "bbox": [ + 86, + 87, + 883, + 368 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "outputs the learnable affine transformations): The first LAT $M \\in \\mathbb{R}^3$ is applied to the input point cloud in the Euclidean space. The second LAT $M_{\\mathrm{ts}}$ is applied to the extracted translation and size features $f_{\\mathrm{ts}}$ . The third LAT $M_r$ is applied to the extracted rotation feature $f_r$ . With this approach, our method can better utilize the valuable geometric features in pose refinement.", + "bbox": [ + 75, + 453, + 472, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.5. Cross-Cloud Transformation (CCT) for Information Mixing", + "text_level": 1, + "bbox": [ + 76, + 565, + 468, + 597 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In pose refinement, effectively blending information from the focalized observed object and the shape prior is crucial for enabling the network to align them accurately. However, in CATRE, the data from the observed point cloud and the shape prior are processed independently until the late rotation prediction stage, where they are merely concatenated, limiting the effectiveness of the alignment. To address this problem, we introduce a novel cross-cloud transformation mechanism that effectively mixes the geometric information from the shape of prior features into the features of the observed point cloud. In particular, we use the feature transformation matrices $M_r^P$ , $M_{\\mathrm{ts}}^P$ of shape prior to transforming the features of the observed point cloud:", + "bbox": [ + 75, + 604, + 468, + 800 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {r} ^ {O} = M _ {r} ^ {P} f _ {r} ^ {O}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 803, + 468, + 821 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\mathrm {t s}} ^ {O} = M _ {\\mathrm {t s}} ^ {P} f _ {\\mathrm {t s}} ^ {O}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 223, + 823, + 468, + 842 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.6. Integrating Shape Prior in Pose Estimation", + "text_level": 1, + "bbox": [ + 76, + 847, + 444, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The information contained in the shape prior is crucial for the network to align the observed point cloud and the shape", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "prior. For the rotation error prediction, the information contained in the shape prior is the essential information. For the translation and size prediction, this information can also be utilized by the network to adjust the learned geometric features accordingly. Therefore, unlike CATRE, which relied solely on features extracted from the observed point cloud to predict $(\\Delta \\mathbf{t}, \\Delta \\mathbf{s})$ , our approach also incorporates the information from shape prior to predict them. In particular, we not only mix the information using previous CCT mechanism, but also concatenate the features from shape prior and observed point cloud to obtain mixed features in the similarly way as the rotation estimation. We utilize $f_{\\mathrm{t}}^{\\mathrm{OP}}$ and $f_{\\mathrm{s}}^{\\mathrm{OP}}$ which contain both information from shape prior and observed point cloud like $f_{r}^{\\mathrm{OP}}$ to predict $(\\Delta \\mathbf{t}, \\delta)$ .", + "bbox": [ + 496, + 453, + 890, + 666 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use two pose error predictors of the same network architecture to predict the rotation error and the translation and size error, respectively. Note that the weights of these two pose error predictors are not shared. The network structure of the pose error predictor is shown in Fig. 2. The pose predictor takes in two features, passes them through two same paths separately, and obtains two vectors in $\\mathbb{R}^3$ . In the translation and size branch, the pose error predictor takes in $f_{\\mathrm{t}}^{\\mathrm{OP}}$ and $f_{\\mathrm{s}}^{\\mathrm{OP}}$ and passes them through the two paths, and the output two vectors are regarded as $\\Delta \\mathbf{t}$ and $\\Delta \\mathbf{s}$ , respectively. For the rotation error prediction, the mixed rotation features $f_{r}^{\\mathrm{OP}}$ are copied and passed through the two paths in the pose error predictor. The two output vectors are regarded as $r_x$ and $r_y$ , where $r_x$ and $r_y$ are the first and second axes of the rotation error matrix $\\Delta \\mathbf{R}$ . The third column $r_z$ of $\\Delta \\mathbf{R}$ can", + "bbox": [ + 496, + 674, + 892, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10696", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "be found by:", + "bbox": [ + 76, + 90, + 166, + 104 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr _ {z} = r _ {x} \\times r _ {y}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 223, + 107, + 468, + 122 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 133, + 209, + 150 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. We implement and experiment with our method using an RTX 4090 GPU with a batch size of 12 and 150 training epochs. We follow CATRE [29] and adopt its loss functions and the basic data augmentation strategies including random dropping points, adding Gaussian noise, random pose perturbations, etc. We set the number of points for both the observed points and shape prior to be 512. We train the network using Ranger optimizer [28, 54, 57] with a base learning rate of $10^{-4}$ and anneal the learning rate from the $72\\%$ of the total epoch based on cosine schedule.", + "bbox": [ + 75, + 157, + 472, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We use CATRE as the baseline for our ablation study as it is the state-of-the-art category-level object refinement method. As CATRE did not provide the results of $\\mathrm{IoU}_{50}$ , we obtained them by using their official pre-trained model and kept the rest of the reported metric scores consistent with the corresponding paper. For fair comparisons, we use the same initial estimations as CATRE, which is the pose estimation results of SPD [45]. The result of replacing PointNet with the 3DGC-Encoder in the ablation study is provided by CATRE. We also apply our method to other state-of-the-art category-level object pose estimation approaches [10, 25, 61] to demonstrate the effectiveness of our proposed refinement method. For the pose refinement on HS-Pose [61] and RBP-Pose [58], we compute the initial estimations using their official pre-trained models. The results of other methods are taken directly from their paper.", + "bbox": [ + 75, + 339, + 472, + 582 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. As we focus on the problem of shape variation between input object point clouds, we choose two popular category-level object pose estimation benchmarks to verify our approach, i.e., REAL275 [49] and CAMERA25 [49]. They both contain 6 object categories with multiple levels of shape complexities, i.e., bowl, can, bottle, laptop, mug, and camera. REAL275 contains 36 objects in 13 real-world scenes with 7k RGB-D images in total. Among them, 16 objects in 7 scenes are used for training, resulting in 4.3k images in training. CAMERA25 is a large synthetic RGB-D dataset. It provides 1085 objects and 275k RGB-D images for training, and 184 objects and 25k images for testing.", + "bbox": [ + 75, + 595, + 470, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation metrics. Following [29, 61], we evaluate our method using: 1) The mean average precision (mAP) of the 3D Intersection over Union $(IoU)$ at different thresholds $(50\\%$ and $75\\%)$ to evaluate the pose and size estimation together. 2) The pose metric at $n^{\\circ}mcm$ defines a pose as", + "bbox": [ + 75, + 790, + 470, + 868 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "correct if the rotation error is below $n^{\\circ}$ and the translation error is below $m$ cm. Here, we use $5^{\\circ}, 10^{\\circ}, 2\\mathrm{cm}$ , and $5\\mathrm{cm}$ as the thresholds.", + "bbox": [ + 500, + 90, + 890, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Ablation study", + "text_level": 1, + "bbox": [ + 500, + 146, + 650, + 162 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To verify the proposed architecture, we conducted comprehensive ablation studies on the REAL275 dataset using the initial pose estimations from SPD [45]. We present a quantitative comparison of our method with various key components disabled to motivate our design choices in Table 1. Full results of the ablation study are reported in the supplementary materials.", + "bbox": [ + 498, + 169, + 893, + 276 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "[AS-1] Using geometric features directly. To illustrate the limitations of existing geometric-based encoder under the problem of shape variations, we replace the encoder of CATRE with two robust geometric-based point cloud convolutional structures, namely 3DGCN-Encoder [26] and HS-Encoder [61]. 3DGCN is a widely adopted graph convolution in existing category-level object pose estimation algorithms, while HS-Encoder is a recent architecture that achieves state-of-the-art performance in category-level object pose estimation. However, as shown in Table 1 [C0, C1], even though both HS-Encoder and 3DGCN-Encoder are powerful in finding an object's pose from individual inputs, they failed to manage the pose refinement scenarios when there exist shape variations between the target object and the shape prior. We observed both encoders result in a performance drop when compared to the original CATRE with IoU75 of 37.3% (HS-Encoder) vs. 43.7% (CATRE), 5°5cm of 43.4% (3DGCN-Encoder) vs. 53.3% (CATRE).", + "bbox": [ + 498, + 294, + 892, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "[AS-2] Use prior features in translation and scale estimation. To validate that the information of shape prior is also important in scale and translation error prediction, we add the features of the prior shape to the scale and translation branch by using the same network architecture as the rotation branch. As shown in Table 1 [D0], incorporating shape prior information in translation and size estimation enhanced the overall performance by $2.2\\%$ improvement on $5^{\\circ}2\\mathrm{cm}$ metric and $2.4\\%$ on $10^{\\circ}2\\mathrm{cm}$ metric.", + "bbox": [ + 498, + 585, + 890, + 722 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "[AS-3] Use learnable affine transformation (LAT) for geometric features. To demonstrate the effectiveness of LAT in addressing the shape variation issue, we conduct ablation studies on applying the proposed LAT to the input point cloud and the extracted geometric features in the feature space. The result is shown in Table 1 [E0]. Compared to using geometric features directly (Table 1 [C0]), LATs bring a significant boost on all the metrics, with $\\mathrm{IoU}_{75}$ improved by $20.9\\%$ , $5^{\\circ}2\\mathrm{cm}$ improved by $10.5\\%$ ,", + "bbox": [ + 498, + 739, + 893, + 878 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "2Note that there was a small mistake with the IoU computation from the original benchmark evaluation code [49], we follow [29] to recalculate", + "bbox": [ + 76, + 875, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "the IoU metrics for the SOTA methods.", + "bbox": [ + 500, + 886, + 710, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10697", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/8eeadb48bc79d2c9fa9976b80d83d882378ae48e1dc534739c189cb99b2a2b7f.jpg", + "table_caption": [ + "Table 1. Ablation studies on REAL275.", + "Higher score indicates better performance. In the 'Row' column, the code in bold means the strategies taken in the final structure. In the 'Method' column, the notation $X:Y$ denotes module $Y$ from structure $X$ , $X + Y$ means add module $Y$ to $X$ , and $X \\rightarrow Y$ indicates replacing $X$ with $Y$ ." + ], + "table_footnote": [], + "table_body": "
RowMethodIoU50IoU755°2cm5°5cm10°2cm10°5cm2cm5°
A0CATRE[45] (baseline)77.043.645.854.461.473.175.158.0
B0Ours: E0 + Cross-Cloud Transformation79.2 2.2↑51.8 8.2↑54.4 8.6↑60.3 5.9↑71.9 10.5↑79.4 6.3↑81.9 6.8↑64.3 6.3↑
C0A0: PointNet → HS-Encoder71.030.141.945.960.670.371.948.7
C1A0: PointNet → 3DGCN-Encoder-28.436.043.4--68.047.7
D0A0 + prior in ST branch77.145.848.054.663.872.577.959.2
E0D0: PointNet → HS-layer+LAT79.451.052.458.669.477.780.462.4
E1B0: No LAT on input points76.139.346.653.065.474.878.058.2
E2B0: No LAT on features78.548.847.453.067.475.080.457.4
E3B0: No LAT on the rotation feature79.850.650.456.268.676.380.260.8
F0E0+ Global Concatenation Fusion77.748.447.854.567.175.280.159.4
", + "bbox": [ + 76, + 128, + 890, + 316 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and $5^{\\circ}5\\mathrm{cm}$ improved by $12.7\\%$ . The resulted network also significantly outperforms the PointNet-based encoder (Table 1[D0]) on all the metrics with $\\mathrm{IoU}_{75}$ and $10^{\\circ}5\\mathrm{cm}$ improved by $5.2\\%$ , $5^{\\circ}2\\mathrm{cm}$ improved by $6.4\\%$ , and $10^{\\circ}2\\mathrm{cm}$ improved by $5.6\\%$ . These results verified the effectiveness of LAT on geometric features.", + "bbox": [ + 75, + 335, + 470, + 429 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "[AS-4] The influence of each learnable affine transformation (LAT). To further demonstrate the influences of each LAT, we conducted three experiments by gradually disabling LAT from the framework: 1) without the LAT on the input point cloud, 2) without applying LATs on features, and 3) without independent LATs on the rotation features, where a single LAT is used for the rotation, translation, and scale features. The results are shown in Table 1 [E1-E3]. Compared to the directly using geometric features (Table 1 [C0]), we show that LAT can significantly enhance the network by around $10\\%$ improvements on $\\mathrm{IoU}_{75}$ and $7\\%$ on $5^{\\circ}5\\mathrm{cm}$ metric in Table 1 [E1-E3]. We verify that the combination of them consistently results in better performance.", + "bbox": [ + 75, + 445, + 468, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "[AS-5] Cross-cloud transformation (CCT) based feature fusion To demonstrate that it is important to have a good feature fusion strategy in the feature extraction phase, we conducted experiments on two different feature fusion strategies. One of them is the widely adopted feature fusion strategy, where the global feature of one point cloud is concatenated with the features of another point cloud and then goes through convolutional layers for feature fusion. Another one is the proposed CCT-based feature fusion. As shown in Table 1 [F0], applying global concatenation-based fusion does not enhance the overall performance, and even results in worse performance in all the metrics with the $5^{\\circ}2\\mathrm{cm}$ significantly decreased by $4.6\\%$ , and $5^{\\circ}5\\mathrm{cm}$ decreased by $3.1\\%$ (compared to Table 1 [D0]). On the contrary, as shown in Table 1 [B0], our simple CCT-based feature fusion shows its effectiveness by improving all the pose", + "bbox": [ + 75, + 657, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/80e8d47f5b934956f66a7f331dad000b3ff466c9344aa828b320c4dcf51b0fa5.jpg", + "image_caption": [ + "Figure 3. Performance comparison of our method and CATRE under different shape priors. $SP - m$ denotes the category's mean shape, while $SP - 1$ and $SP - 2$ are randomly sampled object shapes from CAMERA25." + ], + "image_footnote": [], + "bbox": [ + 501, + 335, + 890, + 525 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "metrics by around $2.0\\%$ . In Supplementary, we also visualize the influence of CCT in feature space.", + "bbox": [ + 498, + 575, + 890, + 607 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "[AS-6] Handle shape variations. To demonstrate that the proposed method can handle the shape variations, we replaced the original shape prior with two randomly sampled models from CAMERA25 training set and trained on REAL275. Note that the original shape prior represents the mean shape of a category, and the new models are randomly sampled. Therefore, there are larger shape variations with certain target objects. As shown in Fig. 3, CATRE performs best when using the mean shape of the category, while its performance drops dramatically on the randomly sampled shape priors. In contrast, our method exhibits robustness to shape variations introduced by different shape priors and consistently delivers strong performance.", + "bbox": [ + 496, + 625, + 890, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "[AS-7] Refinement with different initial estimations. To demonstrate our model robustness to different initial estimations, we compare the proposed method with CATRE on different initial estimations generated by category-level", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10698", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1996b314fa576aac0fbcadfd6781397fddb6746998eb12cf02891ac3b40a9684.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison of proposed (row #3) and baseline (row #2) methods using SPD (row #1) as initial estimation. Ground truth shown with white lines. Note that the estimated rotations of symmetric objects (e.g. bowl, bottle, and can) are considered correct if the symmetry axis is aligned." + ], + "image_footnote": [], + "bbox": [ + 76, + 88, + 893, + 417 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "object pose estimation methods with varying performance, including HS-Pose [61], GPVPose [10], and the SelfDPDN [25]. As shown in Table 2, our method consistently improves the initial estimations by a large margin across different metrics. For example, we improved the SelfDPND and GPV-Pose on strict $5^{\\circ}2\\mathrm{cm}$ metric by $9.6\\%$ and $15.4\\%$ , respectively. We boosted the $\\mathrm{IoU}_{75}$ of the GPV-Pose and the HS-Pose by $26.5\\%$ and $15.2\\%$ , respectively. However, the baseline CATRE fails to refine the initial poses of SelfDPDN. Also, CATRE reaches its limit when refining high-accuracy initial estimations such as the initial poses generated by HS-Pose, resulting in performance drops on $10^{\\circ}5\\mathrm{cm}$ and $10^{\\circ}2\\mathrm{cm}$ . The comparison results demonstrated the robustness and capability of our method to different initial estimations. We report full details in the supplementary.", + "bbox": [ + 75, + 460, + 472, + 688 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Generalizably test on the CAMERA25 dataset", + "text_level": 1, + "bbox": [ + 76, + 696, + 470, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In real-world applications, category-level algorithms often need to generalize across diverse testing scenarios, encountering a larger number of objects than represented in their training sets. To simulate this problem setting, we choose CAMERA25 as it provides more than 25K RGB-D testing images. We train our model using only a mini set (2% and 4%) of the CAMERA25 training data, resulting in a training set of around 5K and 10K images from a total of 275K images. As shown in Table 3, the performance of the baseline method decreased dramatically when using 2% of the training images, with IoU75 dropped by 12.9% and 5°2cm dropped by 9.0% (see [B0]). On the contrary, our", + "bbox": [ + 75, + 719, + 472, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/92a30f3ad1d34b80c178baea0621fd847890f69ee8310dcab0cdc4f4213b70f8.jpg", + "table_caption": [ + "Table 2. Comparison of the pose refinement with the baseline method CATRE on REAL275 with different initial estimations. Each comparison group contains 3 methods: the initial pose estimation method, refinement using CATRE, and refinement using our method, respectively." + ], + "table_footnote": [], + "table_body": "
MethodIoU755°2cm5°5cm10°2cm10°5cm
Self-DPDN [25]42.244.350.965.178.6
Self-DPDN + CATRE0.422↓0.344.0↓5.145.8↓0.464.7↓6.971.7↓
Self-DPDN + Ours49.77.5↑53.99.6↑60.19.2↑75.09.9↑82.84.2↑
GPV-Pose [10]23.1*32.042.955.073.3
GPV-Pose + CATRE42.619.5↑39.77.7↑54.111.2↑57.12.1↑78.04.7↑
GPV-Pose + Ours49.626.5↑47.415.4↑57.814.9↑68.113.1↑81.27.9↑
HS-Pose [61]39.1*46.555.268.682.7
HS-Pose + CATRE47.18.0↑48.72.2↑59.13.9↑67.80.8↓81.21.5↓
HS-Pose + Ours54.315.2↑51.75.2↑59.64.4↑74.35.7↑83.81.1↑
", + "bbox": [ + 501, + 510, + 890, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "method exhibits a much higher performance when using small datasets for training. Specifically, our method can already outperforms the fully-trained CATRE with only $2\\%$ of images (see [B1]). We also report the performance using $4\\%$ images of the training set in Table 3. It can be seen that, despite CATRE use full CAMERA25 training set, our method outperforms it with $\\mathrm{IoU}_{75}$ and $5^{\\circ}2\\mathrm{cm}$ improved by $3.1\\%$ and $3.6\\%$ , respectively.", + "bbox": [ + 496, + 652, + 890, + 776 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Comparison with state-of-the-arts", + "text_level": 1, + "bbox": [ + 498, + 785, + 797, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "REAL275. We conduct pose refinement on SPD [45] using the proposed approach and compare the resulting performance with the state-of-the-art category-level object pose estimation and refinement methods. As shown in Table 4, our method significantly improves the performance of SPD on all the metrics, with $5^{\\circ}5\\mathrm{cm}$ enhanced", + "bbox": [ + 496, + 809, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10699", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5cfcde5a1af6fb3561c2c3eb03240d877abeec0e4da5c1fa47749dda64221d95.jpg", + "table_caption": [ + "Table 3. The generalizability test on the CAMERA25 dataset." + ], + "table_footnote": [], + "table_body": "
RowMethodTrain Data SizeIoU755°2cm5°5cm10°2cm10°5cm2cm5°
A0CATRE275K76.175.480.383.389.3--
B0CATRE5K63.266.472.379.487.488.873.3
B1Ours5K77.575.481.183.490.091.082.3
C0CATRE10K66.569.775.581.889.189.976.7
C1Ours10K79.277.984.083.890.592.085.4
", + "bbox": [ + 197, + 106, + 767, + 212 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/cee9af89f1e39cd5ded6e6b42e2326bfc3ed66791d4067de50935d3df7545e51.jpg", + "table_caption": [ + "Table 4. Comparison with other methods on REAL275." + ], + "table_footnote": [], + "table_body": "
MethodIoU755°2cm5°5cm10°2cm10°5cm
NOCS [49]9.47.210.013.825.2
DualPoseNet [24]30.829.335.950.066.8
CR-Net [50]33.227.834.347.260.8
SGPA [6]37.135.939.661.370.7
RBP-Pose [58]24.538.248.163.179.2
GPV-Pose [10]23.132.042.955.073.3
HS-Pose [61]39.146.555.268.682.7
SPD* [45]27.019.121.243.554.0
SPD*+CATRE [29]43.645.854.461.473.1
SPD*+Ours51.854.460.371.979.4
", + "bbox": [ + 76, + 250, + 468, + 412 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "by $39.1\\%$ , $5^{\\circ}5\\mathrm{cm}$ improved by $35.3\\%$ , $\\mathrm{IoU}_{75}$ enhanced by $24.8\\%$ , $10^{\\circ}2\\mathrm{cm}$ improved by $24.4\\%$ , and $10^{\\circ}5\\mathrm{cm}$ improve by $25.4\\%$ . In comparison to our baseline, CATRE, our proposed method demonstrates a substantial improvement across various performance metrics. Specifically, we observe a remarkable enhancement of $10.5\\%$ in $10^{\\circ}2\\mathrm{cm}$ , $8.6\\%$ in $5^{\\circ}2\\mathrm{cm}$ , $7.2\\%$ in $\\mathrm{IoU}_{75}$ , and $6.3\\%$ in $10^{\\circ}5\\mathrm{cm}$ . In addition, compared with SOTA pose estimation methods, the pose estimation results of applying our proposed refinement method on SPD significantly outperformed these methods's results by a large margin. Specifically, the estimation results of applying our method on SPD rank top on 4 out of 5 metrics and rank second on the rest metric, and achieved a $7.9\\%$ increase on the $5^{\\circ}2\\mathrm{cm}$ metric and $5.1\\%$ enhancement on $5^{\\circ}5\\mathrm{cm}$ . It is worth noting that the purpose of this section is not to compare refinement and estimation methods, as they are designed to address different problems. Instead, our objective here is to demonstrate how our proposed refinement approach can improve the performance of existing pose estimation methods. Therefore, even though our refinement on HS-Pose produced better performance, as mentioned in the ablation study, we choose to refine weaker initial estimations to show the capability of our approach.", + "bbox": [ + 75, + 426, + 472, + 776 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "CAMERA25. Our method outperforms the state-of-the-art methods using only $2\\%$ of the training image set. For more details, please refer to the supplementary.", + "bbox": [ + 75, + 784, + 470, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative examples. We provide qualitative comparisons of the pose estimation results by SPD, CATRE on SPD, and our method on SPD in Fig. 4 and 5. As shown in Fig. 4, our method on SPD achieves the best size and pose", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6e029e7926f2ec3140981561b68e0b5eb8de2f23237bac2089d17c2122c31d25.jpg", + "image_caption": [ + "CATRE", + "Figure 5. Comparison of proposed (row #2) and baseline (row #1) methods) during a complete refinement iteration, both using SPD as initial estimation. The ground truth is represented by white lines." + ], + "image_footnote": [], + "bbox": [ + 557, + 233, + 885, + 328 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "estimations. In particular, by considering the first column of Fig. 4, all of the comparison methods struggle to estimate the orientation of the camera category. We also provide qualitative examples on the iteration process in Fig. 5. Our method demonstrates faster convergence and more accurate final result than CATRE.", + "bbox": [ + 496, + 385, + 890, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 489, + 617, + 505 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we proposed a novel category-level object pose refinement method which targeted at addressing the challenge of shape variation. We shown that the geometric structural information can be aligned by our adaptive affine transformations. We also demonstrated that the crosscloud transformation mechanism can efficiently merges information from distinct point clouds. We further incorporated shape prior information and observed improvements in translation and size predictions. We verified that each of our technical components contributed meaningfully through extensive ablations. We believe our method sets a strong baseline for future study and opens up new possibilities to handling more complex shapes, i.e. articulated objects.", + "bbox": [ + 496, + 518, + 890, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 500, + 734, + 658, + 751 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This research was supported in part by the MSIT (Ministry of Science and ICT), Korea, under the ITRC (Information Technology Research Center) support program (IITP-2024-2020-0-01789) supervised by the IITP (Institute for Information & Communications Technology Planning & Evaluation), National Natural Science Foundation of China under Grant No.62073159, and the Shenzhen Key Laboratory of Control Theory and Intelligent Systems, under grant No.ZDSYS20220330161800001.", + "bbox": [ + 496, + 760, + 890, + 887 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10700", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Dingding Cai, Janne Heikkilä, and Esa Rahtu. Ove6d: Object viewpoint encoding for depth-based 6d object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6803-6813, 2022. 2", + "[2] Pedro Castro and Tae-Kyun Kim. Crt-6d: Fast 6d object pose estimation with cascaded refinement transformers. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 5746-5755, 2023. 1", + "[3] Dengsheng Chen, Jun Li, Zheng Wang, and Kai Xu. Learning canonical shape space for category-level 6d object pose and size estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2", + "[4] Hansheng Chen, Pichao Wang, Fan Wang, Wei Tian, Lu Xiong, and Hao Li. Epro-pnp: Generalized end-to-end probabilistic perspective-n-points for monocular object pose estimation, 2022. 2", + "[5] Hanzhi Chen, Fabian Manhardt, Nassir Navab, and Benjamin Busam. Texpose: Neural texture learning for self-supervised 6d object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4841-4852, 2023. 1", + "[6] Kai Chen and Qi Dou. SGPA: Structure-guided prior adaptation for category-level 6d object pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2773-2782, 2021. 2, 8", + "[7] Kai Chen, Stephen James, Congying Sui, Yun-Hui Liu, Pieter Abbeel, and Qi Dou. Stereopose: Category-level 6d transparent object pose estimation from stereo images via back-view nocs. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2855-2861, 2023. 1", + "[8] Wei Chen, Xi Jia, Hyung Jin Chang, Jinming Duan, and Ales Leonardis. G2L-Net: Global to Local Network for Real-Time 6D Pose Estimation With Embedding Vector Features. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2", + "[9] Wei Chen, Xi Jia, Hyung Jin Chang, Jinming Duan, Linlin Shen, and Ales Leonardis. Fs-net: Fast shape-based network for category-level 6d object pose estimation with decoupled rotation mechanism. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1581-1590, 2021. 1, 2", + "[10] Yan Di, Ruida Zhang, Zhiqiang Lou, Fabian Manhardt, Xiangyang Ji, Nassir Navab, and Federico Tombari. GPV-Pose: Category-level object pose estimation via geometry-guided point-wise voting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6781-6791, 2022. 1, 2, 5, 7, 8", + "[11] Yang Hai, Rui Song, Jiaojiao Li, Mathieu Salzmann, and Yinlin Hu. Rigidity-aware detection for 6d object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8927-8936, 2023. 1" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Rasmus Laurvig Haugaard and Anders Glent Buch. Surfemb: Dense and continuous correspondence distributions for object pose estimation with learnt surface embeddings. CoRR, abs/2111.13489, 2021. 2", + "[13] Yisheng He, Wei Sun, Haibin Huang, Jianran Liu, Haoqiang Fan, and Jian Sun. Pvn3d: A deep point-wise 3d keypoints voting network for 6 dof pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2", + "[14] Stefan Hinterstoisser, Vincent Lepetit, Naresh Rajkumar, and Kurt Konolige. Going further with point pair features. In Computer Vision – ECCV 2016, pages 834–848. Springer International Publishing, 2016. 2", + "[15] Tomas Hodan, Daniel Barath, and Jiri Matas. Epos: Estimating 6d pose of objects with symmetries. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1", + "[16] Muhammad Zubair Irshad, Sergey Zakharov, Rares Ambrus, Thomas Kollar, Zsolt Kira, and Adrien Gaidon. Shapo: Implicit representations for multi object shape appearance and pose optimization. 2022. 2", + "[17] Shun Iwase, Xingyu Liu, Rawal Khirodkar, Rio Yokota, and Kris M Kitani. Repose: Fast 6d object pose refinement via deep texture rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3303-3312, 2021. 1", + "[18] Daniel Kappler, Franziska Meier, Jan Issac, Jim Mainprice, Cristina Garcia Cifuentes, Manuel Wüthrich, Vincent Berenz, Stefan Schaal, Nathan Ratliff, and Jeannette Bohg. Real-time perception meets reactive motion generation. IEEE Robotics and Automation Letters, 3(3):1864-1871, 2018. 1", + "[19] Nikunj Kothari, Misha Gupta, Leena Vachhani, and Hemendra Arya. Pose estimation for an autonomous vehicle using monocular vision. pages 424-431, 2017. 1", + "[20] Yann Labbe, Justin Carpentier, Mathieu Aubry, and Josef Sivic. CosyPose: Consistent multi-view multi-object 6D pose estimation. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2", + "[21] Taeyeop Lee, Byeong-Uk Lee, Inkyu Shin, Jaesung Choe, Ukcheol Shin, In So Kweon, and Kuk-Jin Yoon. UDA-COPE: unsupervised domain adaptation for category-level object pose estimation. CoRR, abs/2111.12580, 2021. 2", + "[22] Yi Li, Gu Wang, Xiangyang Ji, Yu Xiang, and Dieter Fox. Deepim: Deep iterative matching for 6d pose estimation. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 1, 2", + "[23] Haitao Lin, Zichang Liu, Chilam Cheang, Yanwei Fu, Guodong Guo, and Xiangyang Xue. Sar-net: Shape alignment and recovery network for category-level 6d object pose and size estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6707-6717, 2022. 1", + "[24] Jiehong Lin, Zewei Wei, Zhihao Li, Songcen Xu, Kui Jia, and Yuanqing Li. DualPoseNet: Category-level 6d object pose and size estimation using dual pose network with refined learning of pose consistency. In Proceedings of the" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10701", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE/CVF International Conference on Computer Vision (ICCV), pages 3560-3569, 2021. 8", + "[25] Jiehong Lin, Zewei Wei, Changxing Ding, and Kui Jia. Category-level 6d object pose and size estimation using self-supervised deep prior deformation networks, 2022. 1, 2, 5, 7", + "[26] Zhi-Hao Lin, Sheng-Yu Huang, and Yu-Chiang Frank Wang. Convolution in the cloud: Learning deformable kernels in 3d graph convolution networks for point cloud analysis. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1797-1806, 2020. 1, 2, 3, 5", + "[27] Zheng Linfang, Leonardis Ales, Tse Tze Ho, Elden, Horanyi Nora, Chen Hua, Zhang Wei, and Chang Hyung Jin. Tp-ae: Temporally primed 6d object pose tracking with auto-encoders. In 2022 IEEE International Conference on Robotics and Automation (ICRA), 2022. 2", + "[28] Liyuan Liu, Haoming Jiang, Pengcheng He, Weizhu Chen, Xiaodong Liu, Jianfeng Gao, and Jiawei Han. On the variance of the adaptive learning rate and beyond, 2019. 5", + "[29] Xingyu Liu, Gu Wang, Yi Li, and Xiangyang Ji. CATRE: Iterative point clouds alignment for category-level object pose refinement. In European Conference on Computer Vision (ECCV), pages 499-516. Springer, 2022. 1, 2, 3, 5, 8", + "[30] Eitan Marder-Eppstein. Project tango. pages 25-25, 2016. 1", + "[31] Nathaniel Merrill, Yuliang Guo, Xingxing Zuo, Xinyu Huang, Stefan Leutenegger, Xi Peng, Liu Ren, and Guoquan Huang. Symmetry and uncertainty-aware object slam for 6 dof object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14901-14910, 2022. 2", + "[32] Andrew S. Morgan, Bowen Wen, Junchi Liang, Abdeslam Boularias, Aaron M. Dollar, and Kostas Bekris. Vision-driven Compliant Manipulation for Reliable; High-Precision Assembly Tasks. In Proceedings of Robotics: Science and Systems, Virtual, 2021. 1", + "[33] Andrew Nee, S K Ong, George Chryssolouris, and Dimitris Mourtzis. Augmented reality applications in design and manufacturing. CIRP Annals - Manufacturing Technology, 61:657-679, 2012. 1", + "[34] Van Nguyen Nguyen, Yinlin Hu, Yang Xiao, Mathieu Salzmann, and Vincent Lepetit. Templates for 3d object pose estimation revisited: Generalization to new objects and robustness to occlusions, 2022. 2", + "[35] Charles R. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2", + "[36] Mahdi Rad and Vincent Lepetit. BB8: A Scalable, Accurate, Robust to Partial Occlusion Method for Predicting the 3D Poses of Challenging Objects Without Using Depth. In IEEE International Conference on Computer Vision (ICCV), 2017.", + "[37] Aleksandr Segal, Dirk Haehnel, and Sebastian Thrun. Generalized-icp. In Robotics: science and systems (RSS), page 435. Seattle, WA, 2009. 1" + ], + "bbox": [ + 80, + 92, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Ivan Shugurov, Fu Li, Benjamin Busam, and Slobodan Ilic. Osop: A multi-stage one shot object pose estimation framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6835-6844, 2022. 2", + "[39] Yongzhi Su, Jason Rambach, Nareg Minaskan, Paul Lesur, Alain Pagani, and Didier Stricker. Deep multi-state object pose estimation for augmented reality assembly. In 2019 IEEE International Symposium on Mixed and Augmented Reality Adjunct (ISMAR-Adjunct), pages 222–227, 2019. 1", + "[40] Yongzhi Su, Mahdi Saleh, Torben Fetzer, Jason Rambach, Nassir Navab, Benjamin Busam, Didier Stricker, and Federico Tombari. Zebrapose: Coarse to fine surface encoding for 6 dof object pose estimation, 2022. 2", + "[41] Jiaming Sun, Zihao Wang, Siyu Zhang, Xingyi He, Hongcheng Zhao, Guofeng Zhang, and Xiaowei Zhou. Onepose: One-shot object pose estimation without cad models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6825-6834, 2022. 2", + "[42] Martin Sundermeyer, Maximilian Durner, En Yen Puang, Zoltan-Csaba Marton, Narunas Vaskevicius, Kai O. Arras, and Rudolph Triebel. Multi-path learning for object pose estimation across domains, 2019. 2", + "[43] Martin Sundermeyer, Zoltan Marton, Maximilian Durner, and Rudolph Triebel. Augmented Autoencoders: Implicit 3D Orientation Learning for 6D Object Detection. International Journal of Computer Vision (IJCV), 128, 2019. 2", + "[44] Bugra Tekin, Sudipta Sinha, and Pascal Fua. Real-time seamless single shot 6d object pose prediction. pages 292-301, 2018. 2", + "[45] Meng Tian, Marcelo H Ang, and Gim Hee Lee. Shape prior deformation for categorical 6d object pose and size estimation. In European Conference on Computer Vision (ECCV), pages 530-546. Springer, 2020. 2, 5, 6, 7, 8", + "[46] Jonathan Tremblay, Thang To, Balakumar Sundaralingam, Yu Xiang, Dieter Fox, and Stan Birchfield. Deep object pose estimation for semantic robotic grasping of household objects, 2018. 2", + "[47] Joel Vidal, Chyi-Yeu Lin, and Robert Martí. 6d pose estimation using an improved method based on point pair features, 2018. 2", + "[48] Chen Wang, Danfei Xu, Yuke Zhu, Roberto Martin-Martin, Cewu Lu, Li Fei-Fei, and Silvio Savarese. DenseFusion: 6D Object Pose Estimation by Iterative Dense Fusion. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1", + "[49] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J Guibas. Normalized object coordinate space for category-level 6d object pose and size estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2642-2651, 2019. 2, 5, 8", + "[50] Jiaze Wang, Kai Chen, and Qi Dou. Category-level 6d object pose estimation via cascaded relation and recurrent reconstruction networks. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4807-4814. IEEE, 2021. 8" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10702", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Bowen Wen, Chaitanya Mitash, Baozhang Ren, and Kostas Bekris. se(3)-TrackNet: Data-driven 6D Pose Tracking by Calibrating Image Residuals in Synthetic Domains. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2020. 1, 2", + "[52] Bowen Wen, Wenzhao Lian, Kostas Bekris, and Stefan Schaal. You Only Demonstrate Once: Category-Level Manipulation from Single Visual Demonstration. In Proceedings of Robotics: Science and Systems, New York City, NY, USA, 2022. 1", + "[53] Heng Yang and Marco Pavone. Object pose estimation with statistical guarantees: Conformal keypoint detection and geometric uncertainty propagation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8947-8958, 2023. 1", + "[54] Hongwei Yong, Jianqiang Huang, Xiansheng Hua, and Lei Zhang. Gradient centralization: A new optimization technique for deep neural networks, 2020. 5", + "[55] Sergey Zakharov, Ivan Shugurov, and Slobodan Ilic. DPOD: 6D Pose Object Detector and Refiner. In The IEEE International Conference on Computer Vision (ICCV), 2019. 1, 2", + "[56] Kun Zhang, Chen Wang, Hua Chen, Jia Pan, Michael Yu Wang, and Wei Zhang. Vision-based six-dimensional peg-in-hole for practical connector insertion. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 1771-1777, 2023. 1", + "[57] Michael R. Zhang, James Lucas, Geoffrey Hinton, and Jimmy Ba. Lookahead optimizer: k steps forward, 1 step back, 2019. 5", + "[58] Ruida Zhang, Yan Di, Zhiqiang Lou, Fabian Manhardt, Federico Tombari, and Xiangyang Ji. RBP-Pose: Residual bounding box projection for category-level pose estimation, 2022. 1, 2, 5, 8", + "[59] Ruida Zhang, Yan Di, Fabian Manhardt, Federico Tombari, and Xiangyang Ji. SSP-Pose: Symmetry-aware shape prior deformation for direct category-level object pose estimation, 2022. 1, 2", + "[60] Zhongqun Zhang, Wei Chen, Linfang Zheng, Ales Leonardis, and Hyung Jin Chang. Trans6d: Transformer-based 6d object pose estimation and refinement. In Computer Vision - ECCV 2022 Workshops, pages 112-128, Cham, 2023. Springer Nature Switzerland. 1", + "[61] Linfang Zheng, Chen Wang, Yinghan Sun, Esha Dasgupta, Hua Chen, Aleš Leonardis, Wei Zhang, and Hyung Jin Chang. HS-Pose: Hybrid scope feature extraction for category-level object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17163-17173, 2023. 1, 2, 3, 5, 7, 8", + "[62] Jun Zhou, Kai Chen, Linlin Xu, Qi Dou, and Jing Qin. Deep fusion transformer network with weighted vector-wise keypoints voting for robust 6d object pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13967-13977, 2023. 2" + ], + "bbox": [ + 78, + 90, + 470, + 868 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "10703", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/GeoReF_ Geometric Alignment Across Shape Variation for Category-level Object Pose Refinement/966fabe9-a4db-4529-94bf-c710cc01e232_model.json b/2024/GeoReF_ Geometric Alignment Across Shape Variation for Category-level Object Pose Refinement/966fabe9-a4db-4529-94bf-c710cc01e232_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1fc836d0bded375744fe34bf0412ec95eab5eac1 --- /dev/null +++ b/2024/GeoReF_ Geometric Alignment Across Shape Variation for Category-level Object Pose Refinement/966fabe9-a4db-4529-94bf-c710cc01e232_model.json @@ -0,0 +1,2235 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.13, + 0.857, + 0.175 + ], + "angle": 0, + "content": "GeoReF: Geometric Alignment Across Shape Variation for Category-level Object Pose Refinement" + }, + { + "type": "text", + "bbox": [ + 0.095, + 0.202, + 0.873, + 0.241 + ], + "angle": 0, + "content": "Linfang Zheng\\(^{1,3}\\) Tze Ho Elden Tse\\(^{*3}\\) Chen Wang\\(^{*1,2}\\) Yinghan Sun\\(^{1}\\) Hua Chen\\(^{1}\\) \nAles Leonardis\\(^{3}\\) Wei Zhang\\(^{\\dagger 1}\\) Hyung Jin Chang\\(^{3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.256, + 0.882, + 0.327 + ], + "angle": 0, + "content": "\\(^{1}\\)Shenzhen Key Laboratory of Control Theory and Intelligent Systems, School of System Design and Intelligent Manufacturing, Southern University of Science and Technology, China \\(^{2}\\)Department of Computer Science, the University of Hong Kong, China \\(^{3}\\)School of Computer Science, University of Birmingham, UK" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.33, + 0.856, + 0.363 + ], + "angle": 0, + "content": "{1xz948, txt994}@student.bham.ac.uk, cwang5@cs.hku.hk, sunyh2021@mail.sustech.edu.cn {chenh6,zhangw3}@sustech.edu.cn,{a.leonadis,h.j.chang}@bham.ac.uk" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.397, + 0.314, + 0.413 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.429, + 0.474, + 0.718 + ], + "angle": 0, + "content": "Object pose refinement is essential for robust object pose estimation. Previous work has made significant progress towards instance-level object pose refinement. Yet, category-level pose refinement is a more challenging problem due to large shape variations within a category and the discrepancies between the target object and the shape prior. To address these challenges, we introduce a novel architecture for category-level object pose refinement. Our approach integrates an HS-layer and learnable affine transformations, which aims to enhance the extraction and alignment of geometric information. Additionally, we introduce a crosscloud transformation mechanism that efficiently merges diverse data sources. Finally, we push the limits of our model by incorporating the shape prior information for translation and size error prediction. We conducted extensive experiments to demonstrate the effectiveness of the proposed framework. Through extensive quantitative experiments, we demonstrate significant improvement over the baseline method by a large margin across all metrics." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.738, + 0.21, + 0.753 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.763, + 0.47, + 0.854 + ], + "angle": 0, + "content": "Understanding an object's pose is crucial for a wide range of real-world applications, including robotic manipulation [18, 32, 52, 56], augmented reality [30, 33], and autonomous driving [19, 39]. Significant progress has been made for object pose estimation [5, 11, 15, 48, 53, 55] and pose refinement [2, 17, 22, 37, 51, 60] using the object's CAD" + }, + { + "type": "image_caption", + "bbox": [ + 0.52, + 0.42, + 0.564, + 0.432 + ], + "angle": 0, + "content": "Camera" + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.397, + 0.667, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.688, + 0.397, + 0.764, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.789, + 0.398, + 0.871, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.522, + 0.487, + 0.548, + 0.499 + ], + "angle": 0, + "content": "Mug" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.463, + 0.655, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.603, + 0.53, + 0.633, + 0.541 + ], + "angle": 0, + "content": "SP-m" + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.463, + 0.764, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.702, + 0.53, + 0.728, + 0.54 + ], + "angle": 0, + "content": "SP-1" + }, + { + "type": "image", + "bbox": [ + 0.793, + 0.463, + 0.867, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.807, + 0.53, + 0.834, + 0.54 + ], + "angle": 0, + "content": "SP-2" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.548, + 0.892, + 0.585 + ], + "angle": 0, + "content": "Figure 1. Illustration of the shape variation. \\(SP - m\\) represents the category's mean shape, \\(SP - 1\\) and \\(SP - 2\\) represents the randomly sampled object shapes from the CAMERA25 training set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.892, + 0.734 + ], + "angle": 0, + "content": "model. Despite the promising performance, the reliance on accurate instance-level CAD models limits their generalizability to everyday objects. Category-level methods [7, 23-25, 29, 59, 61] is therefore been proposed to overcome this limitation. The objective of this line of work focuses on estimating object poses within a category given category-level shape priors. As a result, they face unique challenges as there exist diverse shape variations in each object category. We illustrate these shape variations in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Recently, there have been remarkable advancements in category-level object pose estimation [9, 10, 25, 58, 61], primarily due to effective utilization of geometric information through 3D graph convolution [26]. In applications that require high precision, it is common to employ an object pose refinement procedure in conjunction with pose estimation. This involves an initial pose estimation algorithm determining the object pose, followed by a refinement step to further enhance the accuracy of the initially estimated pose by predicting and correcting its error. However, while instance-level object pose refinement has been extensively studied," + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.863, + 0.32, + 0.876 + ], + "angle": 0, + "content": "*Equal contribution, order by dice rolling." + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.876, + 0.243, + 0.888 + ], + "angle": 0, + "content": "The corresponding author." + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.888, + 0.456, + 0.901 + ], + "angle": 0, + "content": "1Project page: https://lynne-zheng-linfang.github.io/georef.github.io" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.863, + 0.456, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10693" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.198 + ], + "angle": 0, + "content": "category-level pose refinement remained unexplored until the introduction of CATRE [29]. By leveraging initial object pose and size estimations, CATRE achieves category-level pose refinement by iteratively aligning the observed target object point cloud with the category-level shape prior. This pipeline is shown to be effective by improving the accuracy of the initial pose and size estimations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.2, + 0.471, + 0.442 + ], + "angle": 0, + "content": "While CATRE has proven to be effective in many scenarios, it is limited by the reliance on the PointNet [35] encoder, which is primarily designed for classification and segmentation tasks. This design choice limits its ability to capture essential and fine-grained geometric relationships for accurate pose estimation and refinement. This ability is particularly important in category-level pose refinement as there exists diverse shape variations between inputs. Consequently, CATRE obtains suboptimal performance by direct application of 3D graph convolution. In addition, as CATRE treats the point cloud and the shape prior features separately until a later stage of the network, they potentially miss out on the benefits of integrating these features earlier. Moreover, their approach does not incorporate shape prior information into the translation and scale estimation module, which presents another area for potential improvement." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.445, + 0.47, + 0.686 + ], + "angle": 0, + "content": "In this paper, we introduce a novel architecture for category-level object pose refinement which aims to address the limitations mentioned above. To better extract both local and global geometric information, we incorporate an HS layer into our feature extraction process. We apply learnable affine transformations to the features to address the geometric discrepancies between the observed point cloud and the shape prior. This enables the network to align these features more effectively. In addition, we propose a cross-cloud transformation mechanism that is specifically designed to enhance the merging of information between the observed point clouds and the shape prior. This mechanism enables more efficient integration of information between the two sources. Finally, we push the limit of our model by incorporating shape prior information to more accurately predict errors in translation and size estimation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Our extensive experimental results on two category-level object pose datasets demonstrate that our proposed model to be effective in addressing the problem of shape variations in category-level object pose refinement, and consequently outperforms the state-of-the-art significantly. To the best of our knowledge, our proposed method is the first to successfully address the shape variation issue which is common in category-level pose refinement. Specifically, to enable graph convolution to be effective in capturing geometric relationships between different shapes, we propose an adaptive affine transformation matrix that aligns the observed point clouds and the shape prior. Additionally, the proposed cross-cloud transformation mechanism effectively fuses features from different input point clouds and brings" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.738, + 0.106 + ], + "angle": 0, + "content": "further performance improvements." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.108, + 0.739, + 0.121 + ], + "angle": 0, + "content": "Our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.123, + 0.892, + 0.183 + ], + "angle": 0, + "content": "- We introduce a novel architecture to specifically address the shape variations issue in category-level object pose refinement. Our proposed method results in consistent performance gain and exhibits better generalization ability." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.184, + 0.892, + 0.228 + ], + "angle": 0, + "content": "- We propose a unique cross-cloud transformation mechanism which efficiently merges diverse information from observed point clouds and shape priors." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.229, + 0.892, + 0.318 + ], + "angle": 0, + "content": "- We conduct extensive experiments on two category-level object pose datasets to validate our proposed method. On the REAL275 dataset, our method significantly outperforms SPD by \\(39.1\\%\\) increase in the \\(5^{\\circ}5\\mathrm{cm}\\) metric. Additionally, we achieve \\(10.5\\%\\) improvement in the \\(10^{\\circ}2\\mathrm{cm}\\) metric over the state-of-the-art method, CATRE." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.123, + 0.892, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.333, + 0.642, + 0.349 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.359, + 0.892, + 0.631 + ], + "angle": 0, + "content": "Instance-level object pose estimation and refinement. Instance-level approaches estimate the pose of the target object given known 3D CAD models. They can be briefly divided into correspondence matching methods and template matching methods. Correspondence matching methods [4, 8, 12, 13, 31, 36, 40, 41, 44, 46, 55, 62] matches the outstanding features of the observed object images with its model. Template matching methods [1, 14, 27, 34, 38, 42, 43, 47] compares the images or extracted features with the pre-generated templates. As the initial pose estimates can be noisy to various factors such as occlusions, object pose refinement [20, 22, 51] is shown to be useful in improving the performance of instance-level methods. Even though they achieved impressive over the target object, the reliance on object CAD models limited their generalizability for handling everyday objects. In this paper, we consider a more challenging problem setting where only the category-level shape prior is provided." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Category-level object pose estimation and refinement. Both tasks mainly focus on addressing the shape variation between the objects. The pioneering work NOCS [49] tackles the shape discrepancy by recovering the normalized visible shape of the target object and achieving the pose by point cloud matching. A series of methods extend this structure by leveraging different information such as domain adaptation [21], different reconstruction space [3], shape prior [3, 16, 21, 45], and structural similarities [6, 25]. However, this line of work is often limited in speed due to the iterative point matching. Another series of work starts with FS-Net [9], which adopts 3D graph convolution (3D-GC) [26] to obtain geometric sensitivity. Due to its effectiveness and real-time performance, graph convolution is widely adopted in recent methods with an enhancement in directions including loss function [10], bounding box voting [58], and shape deformation [59]. HS-Pose [61] ex" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10694" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "tends the geometric feature extraction from local to global, which enhances the capability to handle objects with complex shapes. The research on category-level refinement began recently with the proposal of CATRE [29]. It introduced an effective pipeline that leverages shape priors and a focalization strategy for pose refinement and effectively improves the initial pose estimations. In this paper, we extend the CATRE and tackle the geometric variation issue within the framework of category-level pose refinement." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.241, + 0.212, + 0.258 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.265, + 0.281, + 0.28 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.289, + 0.47, + 0.397 + ], + "angle": 0, + "content": "In this paper, we tackle the problem of category-level object pose refinement. Given the initial pose and size estimation \\((\\mathbf{R}_0,\\mathbf{t}_0,\\mathbf{s}_0)\\), along with the observed point cloud \\(\\mathcal{O}\\in \\mathbb{R}^{N^O\\times 3}\\) and the shape prior \\(\\mathcal{P}\\in \\mathbb{R}^{N^P\\times 3}\\), we aim to predict the estimation error \\((\\Delta \\mathbf{R},\\Delta \\mathbf{t},\\Delta \\mathbf{s})\\) between the initial estimations and the ground truths. The pose refinement algorithm \\(\\phi\\) can be described as:" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.408, + 0.47, + 0.425 + ], + "angle": 0, + "content": "\\[\n\\left(\\Delta \\mathbf {R}, \\Delta \\mathbf {t}, \\Delta \\mathbf {s}\\right) = \\phi \\left(\\mathbf {R} _ {0}, \\mathbf {t} _ {0}, \\mathbf {s} _ {0}, \\mathcal {O}, \\mathcal {P}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.434, + 0.469, + 0.465 + ], + "angle": 0, + "content": "This pose refinement algorithm \\(\\phi\\) can be applied iteratively to improve the refinement performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.473, + 0.218, + 0.488 + ], + "angle": 0, + "content": "3.2. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.497, + 0.469, + 0.543 + ], + "angle": 0, + "content": "Our proposed category-level object pose refinement framework builds upon two previous works, CATRE [29] and HS-layer [61], which we briefly review them in the following." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.557, + 0.47, + 0.709 + ], + "angle": 0, + "content": "CATRE. CATRE is the first framework that considers the problem of category-level pose refinement. It predicts the error between the ground truth and the estimated poses by aligning the input point clouds and the categorical shape priors. Specifically, the network architecture of CATRE consists of four components: a) point cloud focalization, b) shared encoder, c) rotation prediction, and d) translation and size prediction. In point clouds focalization, the observed point clouds \\(\\mathcal{O}\\) and the shape prior \\(\\mathcal{P}\\) are first aligned with the initial pose and size estimation \\([\\mathbf{R}_0, \\mathbf{t}_0, \\mathbf{s}_0]\\):" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.716, + 0.402, + 0.755 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\hat {\\mathcal {O}} = \\left\\{\\hat {o} _ {i} | \\hat {o} _ {i} = o _ {i} - \\mathbf {t} _ {0}, o _ {i} \\in \\mathcal {O} \\right\\}, \\\\ \\hat {\\mathcal {P}} = \\left\\{\\hat {p} _ {i} | \\hat {p} _ {i} = \\operatorname {d i a g} \\left(\\mathbf {s} _ {0}\\right) \\mathbf {R} _ {0} p _ {i}, p _ {i} \\in \\mathcal {P} \\right\\}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathrm{diag}(\\cdot)\\) converts a vector to a diagonal matrix. The focalized observed point cloud \\(\\hat{\\mathcal{O}}\\) and the focalized shape prior points \\(\\hat{\\mathcal{P}}\\) contain full information required to predict the estimation error \\((\\Delta \\mathbf{R},\\Delta \\mathbf{t},\\Delta \\mathbf{s})\\). First, a PointNet-based shared encoder is used to extract features from the two focalized point clouds independently. Then, both the extracted features are used for \\(\\Delta \\mathbf{R}\\) estimation, while the global feature of the focalized observed point cloud along with the \\(\\mathbf{s}_0\\) are used for \\(\\Delta \\mathbf{t}\\) and \\(\\Delta \\mathbf{s}\\). The initial estimates" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "are updated using the predicted error \\((\\Delta \\mathbf{R}, \\Delta \\mathbf{t}, \\Delta \\mathbf{s})\\). Finally, the updated estimates are used to predict the error again in which this process is iterative and the estimations are refined progressively and continuously." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.159, + 0.892, + 0.34 + ], + "angle": 0, + "content": "HS-layer. The Hybrid-Scope Geometric Feature Extraction Layer (HS-layer) is a simple network structure based on 3D graph convolution. It consists of two parallel paths that extract different scopes of features from the point cloud. The first path encodes the size and translation information of the target object. Meanwhile, the second path extracts outlier-robust local and global geometric features by applying graph convolution with a strategy of Receptive Fields with Feature Distance (RF-F) metric, alongside an Outlier Robust Feature Extraction Layer (ORL). These properties are particularly beneficial for category-level object pose estimation tasks. For more details, please refer to [61]." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.348, + 0.759, + 0.364 + ], + "angle": 0, + "content": "3.3. Overall Structure of GeoReF" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.371, + 0.892, + 0.523 + ], + "angle": 0, + "content": "The overall framework of our proposed object pose refinement approach is shown in Fig. 2. This framework comprises three principal components: 1) point cloud focalization, 2) feature extraction, and 3) pose error prediction. We follow CATRE and use the same point cloud focalization module. We apply focalization and extract features from both the observed point cloud and the shape prior by our Feature Extraction component. Then, we predict the estimation errors \\((\\Delta \\mathbf{R}, \\Delta t, \\Delta s)\\) using the extracted features in the pose error prediction component." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.531, + 0.892, + 0.561 + ], + "angle": 0, + "content": "3.4. Graph Convolution with Learnable Affine Transformation (LAT)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.892, + 0.795 + ], + "angle": 0, + "content": "Geometric structural information is effective in estimating an object's pose for category-level object pose estimations. However, as shown in ablation study [AS-1], directly applying the 3D graph convolutions (e.g., HS-Encoder [61] and 3DGCN-Encoder [26]) to category-level object pose refinement tasks results in poor performance. This is due to the differences in task nature between the pose estimation and the pose refinement. In the pose estimation task, the network only needs to extract geometric structural information from a single point cloud. However, in the pose refinement framework, it requires extracting the geometric structural information from the two input point clouds as well as establishing the geometric correspondences between different object shapes. This becomes challenging due to the issue of shape variation in category pose refinement." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To address the aforementioned problem, we propose to use learnable affine transformations (LATs). By employing LATs, the network can dynamically adjust the input point cloud and the point features which enables better establishment of geometric correspondences between the two different input shapes. Specifically, we apply three LATs (as shown in the bottom left of Fig. 2, where the Matrix Net" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10695" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.088, + 0.885, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.37, + 0.895, + 0.444 + ], + "angle": 0, + "content": "Figure 2. Overall structure of the proposed method. Our object pose refinement structure contains three main modules. Given the shape prior point cloud, the target object's observed point cloud, and the initial estimation, we first apply point cloud focalization on the input point clouds using the initial estimation. The focalized point clouds then go through a geometric-based feature extraction encoder to obtain geometric structural features. The extracted features are then fed into two branches for rotation error estimation, translation error, and size error estimation. Within the HS Feature Extractor, the Matrix Net models output the learnable affine transformations (LATs) for adaptive point and feature adjustment. The left output of the Matrix Net adjusts the input point clouds, while the right Matrix Net model outputs two affine transformations for adjusting the rotation features, and the translation and size features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.454, + 0.473, + 0.56 + ], + "angle": 0, + "content": "outputs the learnable affine transformations): The first LAT \\( M \\in \\mathbb{R}^3 \\) is applied to the input point cloud in the Euclidean space. The second LAT \\( M_{\\mathrm{ts}} \\) is applied to the extracted translation and size features \\( f_{\\mathrm{ts}} \\). The third LAT \\( M_r \\) is applied to the extracted rotation feature \\( f_r \\). With this approach, our method can better utilize the valuable geometric features in pose refinement." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.566, + 0.47, + 0.598 + ], + "angle": 0, + "content": "3.5. Cross-Cloud Transformation (CCT) for Information Mixing" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.605, + 0.47, + 0.801 + ], + "angle": 0, + "content": "In pose refinement, effectively blending information from the focalized observed object and the shape prior is crucial for enabling the network to align them accurately. However, in CATRE, the data from the observed point cloud and the shape prior are processed independently until the late rotation prediction stage, where they are merely concatenated, limiting the effectiveness of the alignment. To address this problem, we introduce a novel cross-cloud transformation mechanism that effectively mixes the geometric information from the shape of prior features into the features of the observed point cloud. In particular, we use the feature transformation matrices \\( M_r^P \\), \\( M_{\\mathrm{ts}}^P \\) of shape prior to transforming the features of the observed point cloud:" + }, + { + "type": "equation", + "bbox": [ + 0.222, + 0.804, + 0.469, + 0.822 + ], + "angle": 0, + "content": "\\[\nf _ {r} ^ {O} = M _ {r} ^ {P} f _ {r} ^ {O}, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.225, + 0.824, + 0.469, + 0.843 + ], + "angle": 0, + "content": "\\[\nf _ {\\mathrm {t s}} ^ {O} = M _ {\\mathrm {t s}} ^ {P} f _ {\\mathrm {t s}} ^ {O}. \\tag {4}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.445, + 0.865 + ], + "angle": 0, + "content": "3.6. Integrating Shape Prior in Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "The information contained in the shape prior is crucial for the network to align the observed point cloud and the shape" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.454, + 0.892, + 0.667 + ], + "angle": 0, + "content": "prior. For the rotation error prediction, the information contained in the shape prior is the essential information. For the translation and size prediction, this information can also be utilized by the network to adjust the learned geometric features accordingly. Therefore, unlike CATRE, which relied solely on features extracted from the observed point cloud to predict \\((\\Delta \\mathbf{t}, \\Delta \\mathbf{s})\\), our approach also incorporates the information from shape prior to predict them. In particular, we not only mix the information using previous CCT mechanism, but also concatenate the features from shape prior and observed point cloud to obtain mixed features in the similarly way as the rotation estimation. We utilize \\(f_{\\mathrm{t}}^{\\mathrm{OP}}\\) and \\(f_{\\mathrm{s}}^{\\mathrm{OP}}\\) which contain both information from shape prior and observed point cloud like \\(f_{r}^{\\mathrm{OP}}\\) to predict \\((\\Delta \\mathbf{t}, \\delta)\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.893, + 0.901 + ], + "angle": 0, + "content": "We use two pose error predictors of the same network architecture to predict the rotation error and the translation and size error, respectively. Note that the weights of these two pose error predictors are not shared. The network structure of the pose error predictor is shown in Fig. 2. The pose predictor takes in two features, passes them through two same paths separately, and obtains two vectors in \\(\\mathbb{R}^3\\). In the translation and size branch, the pose error predictor takes in \\(f_{\\mathrm{t}}^{\\mathrm{OP}}\\) and \\(f_{\\mathrm{s}}^{\\mathrm{OP}}\\) and passes them through the two paths, and the output two vectors are regarded as \\(\\Delta \\mathbf{t}\\) and \\(\\Delta \\mathbf{s}\\), respectively. For the rotation error prediction, the mixed rotation features \\(f_{r}^{\\mathrm{OP}}\\) are copied and passed through the two paths in the pose error predictor. The two output vectors are regarded as \\(r_x\\) and \\(r_y\\), where \\(r_x\\) and \\(r_y\\) are the first and second axes of the rotation error matrix \\(\\Delta \\mathbf{R}\\). The third column \\(r_z\\) of \\(\\Delta \\mathbf{R}\\) can" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10696" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.167, + 0.106 + ], + "angle": 0, + "content": "be found by:" + }, + { + "type": "equation", + "bbox": [ + 0.224, + 0.108, + 0.47, + 0.123 + ], + "angle": 0, + "content": "\\[\nr _ {z} = r _ {x} \\times r _ {y}. \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.134, + 0.21, + 0.151 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.159, + 0.473, + 0.326 + ], + "angle": 0, + "content": "Implementation details. We implement and experiment with our method using an RTX 4090 GPU with a batch size of 12 and 150 training epochs. We follow CATRE [29] and adopt its loss functions and the basic data augmentation strategies including random dropping points, adding Gaussian noise, random pose perturbations, etc. We set the number of points for both the observed points and shape prior to be 512. We train the network using Ranger optimizer [28, 54, 57] with a base learning rate of \\(10^{-4}\\) and anneal the learning rate from the \\(72\\%\\) of the total epoch based on cosine schedule." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.34, + 0.473, + 0.583 + ], + "angle": 0, + "content": "Baselines. We use CATRE as the baseline for our ablation study as it is the state-of-the-art category-level object refinement method. As CATRE did not provide the results of \\(\\mathrm{IoU}_{50}\\), we obtained them by using their official pre-trained model and kept the rest of the reported metric scores consistent with the corresponding paper. For fair comparisons, we use the same initial estimations as CATRE, which is the pose estimation results of SPD [45]. The result of replacing PointNet with the 3DGC-Encoder in the ablation study is provided by CATRE. We also apply our method to other state-of-the-art category-level object pose estimation approaches [10, 25, 61] to demonstrate the effectiveness of our proposed refinement method. For the pose refinement on HS-Pose [61] and RBP-Pose [58], we compute the initial estimations using their official pre-trained models. The results of other methods are taken directly from their paper." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.596, + 0.471, + 0.779 + ], + "angle": 0, + "content": "Datasets. As we focus on the problem of shape variation between input object point clouds, we choose two popular category-level object pose estimation benchmarks to verify our approach, i.e., REAL275 [49] and CAMERA25 [49]. They both contain 6 object categories with multiple levels of shape complexities, i.e., bowl, can, bottle, laptop, mug, and camera. REAL275 contains 36 objects in 13 real-world scenes with 7k RGB-D images in total. Among them, 16 objects in 7 scenes are used for training, resulting in 4.3k images in training. CAMERA25 is a large synthetic RGB-D dataset. It provides 1085 objects and 275k RGB-D images for training, and 184 objects and 25k images for testing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.791, + 0.472, + 0.869 + ], + "angle": 0, + "content": "Evaluation metrics. Following [29, 61], we evaluate our method using: 1) The mean average precision (mAP) of the 3D Intersection over Union \\((IoU)\\) at different thresholds \\((50\\%\\) and \\(75\\%)\\) to evaluate the pose and size estimation together. 2) The pose metric at \\(n^{\\circ}mcm\\) defines a pose as" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "correct if the rotation error is below \\( n^{\\circ} \\) and the translation error is below \\( m \\) cm. Here, we use \\( 5^{\\circ}, 10^{\\circ}, 2\\mathrm{cm} \\), and \\( 5\\mathrm{cm} \\) as the thresholds." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.147, + 0.651, + 0.164 + ], + "angle": 0, + "content": "4.1. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.17, + 0.894, + 0.277 + ], + "angle": 0, + "content": "To verify the proposed architecture, we conducted comprehensive ablation studies on the REAL275 dataset using the initial pose estimations from SPD [45]. We present a quantitative comparison of our method with various key components disabled to motivate our design choices in Table 1. Full results of the ablation study are reported in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.295, + 0.893, + 0.568 + ], + "angle": 0, + "content": "[AS-1] Using geometric features directly. To illustrate the limitations of existing geometric-based encoder under the problem of shape variations, we replace the encoder of CATRE with two robust geometric-based point cloud convolutional structures, namely 3DGCN-Encoder [26] and HS-Encoder [61]. 3DGCN is a widely adopted graph convolution in existing category-level object pose estimation algorithms, while HS-Encoder is a recent architecture that achieves state-of-the-art performance in category-level object pose estimation. However, as shown in Table 1 [C0, C1], even though both HS-Encoder and 3DGCN-Encoder are powerful in finding an object's pose from individual inputs, they failed to manage the pose refinement scenarios when there exist shape variations between the target object and the shape prior. We observed both encoders result in a performance drop when compared to the original CATRE with IoU75 of 37.3% (HS-Encoder) vs. 43.7% (CATRE), 5°5cm of 43.4% (3DGCN-Encoder) vs. 53.3% (CATRE)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.587, + 0.892, + 0.723 + ], + "angle": 0, + "content": "[AS-2] Use prior features in translation and scale estimation. To validate that the information of shape prior is also important in scale and translation error prediction, we add the features of the prior shape to the scale and translation branch by using the same network architecture as the rotation branch. As shown in Table 1 [D0], incorporating shape prior information in translation and size estimation enhanced the overall performance by \\(2.2\\%\\) improvement on \\(5^{\\circ}2\\mathrm{cm}\\) metric and \\(2.4\\%\\) on \\(10^{\\circ}2\\mathrm{cm}\\) metric." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.741, + 0.894, + 0.879 + ], + "angle": 0, + "content": "[AS-3] Use learnable affine transformation (LAT) for geometric features. To demonstrate the effectiveness of LAT in addressing the shape variation issue, we conduct ablation studies on applying the proposed LAT to the input point cloud and the extracted geometric features in the feature space. The result is shown in Table 1 [E0]. Compared to using geometric features directly (Table 1 [C0]), LATs bring a significant boost on all the metrics, with \\(\\mathrm{IoU}_{75}\\) improved by \\(20.9\\%\\), \\(5^{\\circ}2\\mathrm{cm}\\) improved by \\(10.5\\%\\)," + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.875, + 0.472, + 0.901 + ], + "angle": 0, + "content": "2Note that there was a small mistake with the IoU computation from the original benchmark evaluation code [49], we follow [29] to recalculate" + }, + { + "type": "page_footnote", + "bbox": [ + 0.501, + 0.887, + 0.712, + 0.9 + ], + "angle": 0, + "content": "the IoU metrics for the SOTA methods." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10697" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.372, + 0.09, + 0.598, + 0.102 + ], + "angle": 0, + "content": "Table 1. Ablation studies on REAL275." + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.103, + 0.892, + 0.127 + ], + "angle": 0, + "content": "Higher score indicates better performance. In the 'Row' column, the code in bold means the strategies taken in the final structure. In the 'Method' column, the notation \\( X:Y \\) denotes module \\( Y \\) from structure \\( X \\), \\( X + Y \\) means add module \\( Y \\) to \\( X \\), and \\( X \\rightarrow Y \\) indicates replacing \\( X \\) with \\( Y \\)." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.129, + 0.892, + 0.317 + ], + "angle": 0, + "content": "
RowMethodIoU50IoU755°2cm5°5cm10°2cm10°5cm2cm5°
A0CATRE[45] (baseline)77.043.645.854.461.473.175.158.0
B0Ours: E0 + Cross-Cloud Transformation79.2 2.2↑51.8 8.2↑54.4 8.6↑60.3 5.9↑71.9 10.5↑79.4 6.3↑81.9 6.8↑64.3 6.3↑
C0A0: PointNet → HS-Encoder71.030.141.945.960.670.371.948.7
C1A0: PointNet → 3DGCN-Encoder-28.436.043.4--68.047.7
D0A0 + prior in ST branch77.145.848.054.663.872.577.959.2
E0D0: PointNet → HS-layer+LAT79.451.052.458.669.477.780.462.4
E1B0: No LAT on input points76.139.346.653.065.474.878.058.2
E2B0: No LAT on features78.548.847.453.067.475.080.457.4
E3B0: No LAT on the rotation feature79.850.650.456.268.676.380.260.8
F0E0+ Global Concatenation Fusion77.748.447.854.567.175.280.159.4
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.337, + 0.471, + 0.43 + ], + "angle": 0, + "content": "and \\(5^{\\circ}5\\mathrm{cm}\\) improved by \\(12.7\\%\\). The resulted network also significantly outperforms the PointNet-based encoder (Table 1[D0]) on all the metrics with \\(\\mathrm{IoU}_{75}\\) and \\(10^{\\circ}5\\mathrm{cm}\\) improved by \\(5.2\\%\\), \\(5^{\\circ}2\\mathrm{cm}\\) improved by \\(6.4\\%\\), and \\(10^{\\circ}2\\mathrm{cm}\\) improved by \\(5.6\\%\\). These results verified the effectiveness of LAT on geometric features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.446, + 0.47, + 0.643 + ], + "angle": 0, + "content": "[AS-4] The influence of each learnable affine transformation (LAT). To further demonstrate the influences of each LAT, we conducted three experiments by gradually disabling LAT from the framework: 1) without the LAT on the input point cloud, 2) without applying LATs on features, and 3) without independent LATs on the rotation features, where a single LAT is used for the rotation, translation, and scale features. The results are shown in Table 1 [E1-E3]. Compared to the directly using geometric features (Table 1 [C0]), we show that LAT can significantly enhance the network by around \\(10\\%\\) improvements on \\(\\mathrm{IoU}_{75}\\) and \\(7\\%\\) on \\(5^{\\circ}5\\mathrm{cm}\\) metric in Table 1 [E1-E3]. We verify that the combination of them consistently results in better performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.471, + 0.903 + ], + "angle": 0, + "content": "[AS-5] Cross-cloud transformation (CCT) based feature fusion To demonstrate that it is important to have a good feature fusion strategy in the feature extraction phase, we conducted experiments on two different feature fusion strategies. One of them is the widely adopted feature fusion strategy, where the global feature of one point cloud is concatenated with the features of another point cloud and then goes through convolutional layers for feature fusion. Another one is the proposed CCT-based feature fusion. As shown in Table 1 [F0], applying global concatenation-based fusion does not enhance the overall performance, and even results in worse performance in all the metrics with the \\(5^{\\circ}2\\mathrm{cm}\\) significantly decreased by \\(4.6\\%\\), and \\(5^{\\circ}5\\mathrm{cm}\\) decreased by \\(3.1\\%\\) (compared to Table 1 [D0]). On the contrary, as shown in Table 1 [B0], our simple CCT-based feature fusion shows its effectiveness by improving all the pose" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.337, + 0.892, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.529, + 0.892, + 0.567 + ], + "angle": 0, + "content": "Figure 3. Performance comparison of our method and CATRE under different shape priors. \\(SP - m\\) denotes the category's mean shape, while \\(SP - 1\\) and \\(SP - 2\\) are randomly sampled object shapes from CAMERA25." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.577, + 0.891, + 0.608 + ], + "angle": 0, + "content": "metrics by around \\(2.0\\%\\). In Supplementary, we also visualize the influence of CCT in feature space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.626, + 0.892, + 0.822 + ], + "angle": 0, + "content": "[AS-6] Handle shape variations. To demonstrate that the proposed method can handle the shape variations, we replaced the original shape prior with two randomly sampled models from CAMERA25 training set and trained on REAL275. Note that the original shape prior represents the mean shape of a category, and the new models are randomly sampled. Therefore, there are larger shape variations with certain target objects. As shown in Fig. 3, CATRE performs best when using the mean shape of the category, while its performance drops dramatically on the randomly sampled shape priors. In contrast, our method exhibits robustness to shape variations introduced by different shape priors and consistently delivers strong performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.903 + ], + "angle": 0, + "content": "[AS-7] Refinement with different initial estimations. To demonstrate our model robustness to different initial estimations, we compare the proposed method with CATRE on different initial estimations generated by category-level" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10698" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.089, + 0.895, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.422, + 0.893, + 0.448 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison of proposed (row #3) and baseline (row #2) methods using SPD (row #1) as initial estimation. Ground truth shown with white lines. Note that the estimated rotations of symmetric objects (e.g. bowl, bottle, and can) are considered correct if the symmetry axis is aligned." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.461, + 0.473, + 0.689 + ], + "angle": 0, + "content": "object pose estimation methods with varying performance, including HS-Pose [61], GPVPose [10], and the SelfDPDN [25]. As shown in Table 2, our method consistently improves the initial estimations by a large margin across different metrics. For example, we improved the SelfDPND and GPV-Pose on strict \\(5^{\\circ}2\\mathrm{cm}\\) metric by \\(9.6\\%\\) and \\(15.4\\%\\), respectively. We boosted the \\(\\mathrm{IoU}_{75}\\) of the GPV-Pose and the HS-Pose by \\(26.5\\%\\) and \\(15.2\\%\\), respectively. However, the baseline CATRE fails to refine the initial poses of SelfDPDN. Also, CATRE reaches its limit when refining high-accuracy initial estimations such as the initial poses generated by HS-Pose, resulting in performance drops on \\(10^{\\circ}5\\mathrm{cm}\\) and \\(10^{\\circ}2\\mathrm{cm}\\). The comparison results demonstrated the robustness and capability of our method to different initial estimations. We report full details in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.697, + 0.471, + 0.713 + ], + "angle": 0, + "content": "4.2. Generalizably test on the CAMERA25 dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.473, + 0.902 + ], + "angle": 0, + "content": "In real-world applications, category-level algorithms often need to generalize across diverse testing scenarios, encountering a larger number of objects than represented in their training sets. To simulate this problem setting, we choose CAMERA25 as it provides more than 25K RGB-D testing images. We train our model using only a mini set (2% and 4%) of the CAMERA25 training data, resulting in a training set of around 5K and 10K images from a total of 275K images. As shown in Table 3, the performance of the baseline method decreased dramatically when using 2% of the training images, with IoU75 dropped by 12.9% and 5°2cm dropped by 9.0% (see [B0]). On the contrary, our" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.459, + 0.892, + 0.509 + ], + "angle": 0, + "content": "Table 2. Comparison of the pose refinement with the baseline method CATRE on REAL275 with different initial estimations. Each comparison group contains 3 methods: the initial pose estimation method, refinement using CATRE, and refinement using our method, respectively." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.511, + 0.892, + 0.639 + ], + "angle": 0, + "content": "
MethodIoU755°2cm5°5cm10°2cm10°5cm
Self-DPDN [25]42.244.350.965.178.6
Self-DPDN + CATRE0.422↓0.344.0↓5.145.8↓0.464.7↓6.971.7↓
Self-DPDN + Ours49.77.5↑53.99.6↑60.19.2↑75.09.9↑82.84.2↑
GPV-Pose [10]23.1*32.042.955.073.3
GPV-Pose + CATRE42.619.5↑39.77.7↑54.111.2↑57.12.1↑78.04.7↑
GPV-Pose + Ours49.626.5↑47.415.4↑57.814.9↑68.113.1↑81.27.9↑
HS-Pose [61]39.1*46.555.268.682.7
HS-Pose + CATRE47.18.0↑48.72.2↑59.13.9↑67.80.8↓81.21.5↓
HS-Pose + Ours54.315.2↑51.75.2↑59.64.4↑74.35.7↑83.81.1↑
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.654, + 0.892, + 0.777 + ], + "angle": 0, + "content": "method exhibits a much higher performance when using small datasets for training. Specifically, our method can already outperforms the fully-trained CATRE with only \\(2\\%\\) of images (see [B1]). We also report the performance using \\(4\\%\\) images of the training set in Table 3. It can be seen that, despite CATRE use full CAMERA25 training set, our method outperforms it with \\(\\mathrm{IoU}_{75}\\) and \\(5^{\\circ}2\\mathrm{cm}\\) improved by \\(3.1\\%\\) and \\(3.6\\%\\), respectively." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.786, + 0.798, + 0.803 + ], + "angle": 0, + "content": "4.3. Comparison with state-of-the-arts" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.901 + ], + "angle": 0, + "content": "REAL275. We conduct pose refinement on SPD [45] using the proposed approach and compare the resulting performance with the state-of-the-art category-level object pose estimation and refinement methods. As shown in Table 4, our method significantly improves the performance of SPD on all the metrics, with \\(5^{\\circ}5\\mathrm{cm}\\) enhanced" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10699" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.31, + 0.09, + 0.66, + 0.103 + ], + "angle": 0, + "content": "Table 3. The generalizability test on the CAMERA25 dataset." + }, + { + "type": "table", + "bbox": [ + 0.198, + 0.107, + 0.768, + 0.213 + ], + "angle": 0, + "content": "
RowMethodTrain Data SizeIoU755°2cm5°5cm10°2cm10°5cm2cm5°
A0CATRE275K76.175.480.383.389.3--
B0CATRE5K63.266.472.379.487.488.873.3
B1Ours5K77.575.481.183.490.091.082.3
C0CATRE10K66.569.775.581.889.189.976.7
C1Ours10K79.277.984.083.890.592.085.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.116, + 0.235, + 0.431, + 0.248 + ], + "angle": 0, + "content": "Table 4. Comparison with other methods on REAL275." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.251, + 0.47, + 0.413 + ], + "angle": 0, + "content": "
MethodIoU755°2cm5°5cm10°2cm10°5cm
NOCS [49]9.47.210.013.825.2
DualPoseNet [24]30.829.335.950.066.8
CR-Net [50]33.227.834.347.260.8
SGPA [6]37.135.939.661.370.7
RBP-Pose [58]24.538.248.163.179.2
GPV-Pose [10]23.132.042.955.073.3
HS-Pose [61]39.146.555.268.682.7
SPD* [45]27.019.121.243.554.0
SPD*+CATRE [29]43.645.854.461.473.1
SPD*+Ours51.854.460.371.979.4
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.428, + 0.473, + 0.777 + ], + "angle": 0, + "content": "by \\(39.1\\%\\), \\(5^{\\circ}5\\mathrm{cm}\\) improved by \\(35.3\\%\\), \\(\\mathrm{IoU}_{75}\\) enhanced by \\(24.8\\%\\), \\(10^{\\circ}2\\mathrm{cm}\\) improved by \\(24.4\\%\\), and \\(10^{\\circ}5\\mathrm{cm}\\) improve by \\(25.4\\%\\). In comparison to our baseline, CATRE, our proposed method demonstrates a substantial improvement across various performance metrics. Specifically, we observe a remarkable enhancement of \\(10.5\\%\\) in \\(10^{\\circ}2\\mathrm{cm}\\), \\(8.6\\%\\) in \\(5^{\\circ}2\\mathrm{cm}\\), \\(7.2\\%\\) in \\(\\mathrm{IoU}_{75}\\), and \\(6.3\\%\\) in \\(10^{\\circ}5\\mathrm{cm}\\). In addition, compared with SOTA pose estimation methods, the pose estimation results of applying our proposed refinement method on SPD significantly outperformed these methods's results by a large margin. Specifically, the estimation results of applying our method on SPD rank top on 4 out of 5 metrics and rank second on the rest metric, and achieved a \\(7.9\\%\\) increase on the \\(5^{\\circ}2\\mathrm{cm}\\) metric and \\(5.1\\%\\) enhancement on \\(5^{\\circ}5\\mathrm{cm}\\). It is worth noting that the purpose of this section is not to compare refinement and estimation methods, as they are designed to address different problems. Instead, our objective here is to demonstrate how our proposed refinement approach can improve the performance of existing pose estimation methods. Therefore, even though our refinement on HS-Pose produced better performance, as mentioned in the ablation study, we choose to refine weaker initial estimations to show the capability of our approach." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.785, + 0.471, + 0.832 + ], + "angle": 0, + "content": "CAMERA25. Our method outperforms the state-of-the-art methods using only \\(2\\%\\) of the training image set. For more details, please refer to the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Qualitative examples. We provide qualitative comparisons of the pose estimation results by SPD, CATRE on SPD, and our method on SPD in Fig. 4 and 5. As shown in Fig. 4, our method on SPD achieves the best size and pose" + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.25, + 0.553, + 0.262 + ], + "angle": 0, + "content": "CATRE" + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.234, + 0.887, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.332, + 0.892, + 0.37 + ], + "angle": 0, + "content": "Figure 5. Comparison of proposed (row #2) and baseline (row #1) methods) during a complete refinement iteration, both using SPD as initial estimation. The ground truth is represented by white lines." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.386, + 0.892, + 0.476 + ], + "angle": 0, + "content": "estimations. In particular, by considering the first column of Fig. 4, all of the comparison methods struggle to estimate the orientation of the camera category. We also provide qualitative examples on the iteration process in Fig. 5. Our method demonstrates faster convergence and more accurate final result than CATRE." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.491, + 0.619, + 0.506 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.519, + 0.892, + 0.716 + ], + "angle": 0, + "content": "In this work, we proposed a novel category-level object pose refinement method which targeted at addressing the challenge of shape variation. We shown that the geometric structural information can be aligned by our adaptive affine transformations. We also demonstrated that the crosscloud transformation mechanism can efficiently merges information from distinct point clouds. We further incorporated shape prior information and observed improvements in translation and size predictions. We verified that each of our technical components contributed meaningfully through extensive ablations. We believe our method sets a strong baseline for future study and opens up new possibilities to handling more complex shapes, i.e. articulated objects." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.735, + 0.66, + 0.752 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.761, + 0.892, + 0.888 + ], + "angle": 0, + "content": "This research was supported in part by the MSIT (Ministry of Science and ICT), Korea, under the ITRC (Information Technology Research Center) support program (IITP-2024-2020-0-01789) supervised by the IITP (Institute for Information & Communications Technology Planning & Evaluation), National Natural Science Foundation of China under Grant No.62073159, and the Shenzhen Key Laboratory of Control Theory and Intelligent Systems, under grant No.ZDSYS20220330161800001." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10700" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Dingding Cai, Janne Heikkilä, and Esa Rahtu. Ove6d: Object viewpoint encoding for depth-based 6d object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6803-6813, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.256 + ], + "angle": 0, + "content": "[2] Pedro Castro and Tae-Kyun Kim. Crt-6d: Fast 6d object pose estimation with cascaded refinement transformers. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 5746-5755, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.472, + 0.327 + ], + "angle": 0, + "content": "[3] Dengsheng Chen, Jun Li, Zheng Wang, and Kai Xu. Learning canonical shape space for category-level 6d object pose and size estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.33, + 0.47, + 0.386 + ], + "angle": 0, + "content": "[4] Hansheng Chen, Pichao Wang, Fan Wang, Wei Tian, Lu Xiong, and Hao Li. Epro-pnp: Generalized end-to-end probabilistic perspective-n-points for monocular object pose estimation, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.388, + 0.472, + 0.457 + ], + "angle": 0, + "content": "[5] Hanzhi Chen, Fabian Manhardt, Nassir Navab, and Benjamin Busam. Texpose: Neural texture learning for self-supervised 6d object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4841-4852, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.46, + 0.47, + 0.516 + ], + "angle": 0, + "content": "[6] Kai Chen and Qi Dou. SGPA: Structure-guided prior adaptation for category-level 6d object pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2773-2782, 2021. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.518, + 0.47, + 0.587 + ], + "angle": 0, + "content": "[7] Kai Chen, Stephen James, Congying Sui, Yun-Hui Liu, Pieter Abbeel, and Qi Dou. Stereopose: Category-level 6d transparent object pose estimation from stereo images via back-view nocs. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2855-2861, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.59, + 0.47, + 0.658 + ], + "angle": 0, + "content": "[8] Wei Chen, Xi Jia, Hyung Jin Chang, Jinming Duan, and Ales Leonardis. G2L-Net: Global to Local Network for Real-Time 6D Pose Estimation With Embedding Vector Features. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.661, + 0.47, + 0.744 + ], + "angle": 0, + "content": "[9] Wei Chen, Xi Jia, Hyung Jin Chang, Jinming Duan, Linlin Shen, and Ales Leonardis. Fs-net: Fast shape-based network for category-level 6d object pose estimation with decoupled rotation mechanism. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1581-1590, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.47, + 0.829 + ], + "angle": 0, + "content": "[10] Yan Di, Ruida Zhang, Zhiqiang Lou, Fabian Manhardt, Xiangyang Ji, Nassir Navab, and Federico Tombari. GPV-Pose: Category-level object pose estimation via geometry-guided point-wise voting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6781-6791, 2022. 1, 2, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[11] Yang Hai, Rui Song, Jiaojiao Li, Mathieu Salzmann, and Yinlin Hu. Rigidity-aware detection for 6d object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8927-8936, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[12] Rasmus Laurvig Haugaard and Anders Glent Buch. Surfemb: Dense and continuous correspondence distributions for object pose estimation with learnt surface embeddings. CoRR, abs/2111.13489, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.894, + 0.219 + ], + "angle": 0, + "content": "[13] Yisheng He, Wei Sun, Haibin Huang, Jianran Liu, Haoqiang Fan, and Jian Sun. Pvn3d: A deep point-wise 3d keypoints voting network for 6 dof pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.276 + ], + "angle": 0, + "content": "[14] Stefan Hinterstoisser, Vincent Lepetit, Naresh Rajkumar, and Kurt Konolige. Going further with point pair features. In Computer Vision – ECCV 2016, pages 834–848. Springer International Publishing, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.895, + 0.332 + ], + "angle": 0, + "content": "[15] Tomas Hodan, Daniel Barath, and Jiri Matas. Epos: Estimating 6d pose of objects with symmetries. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.39 + ], + "angle": 0, + "content": "[16] Muhammad Zubair Irshad, Sergey Zakharov, Rares Ambrus, Thomas Kollar, Zsolt Kira, and Adrien Gaidon. Shapo: Implicit representations for multi object shape appearance and pose optimization. 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.392, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[17] Shun Iwase, Xingyu Liu, Rawal Khirodkar, Rio Yokota, and Kris M Kitani. Repose: Fast 6d object pose refinement via deep texture rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3303-3312, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.462, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[18] Daniel Kappler, Franziska Meier, Jan Issac, Jim Mainprice, Cristina Garcia Cifuentes, Manuel Wüthrich, Vincent Berenz, Stefan Schaal, Nathan Ratliff, and Jeannette Bohg. Real-time perception meets reactive motion generation. IEEE Robotics and Automation Letters, 3(3):1864-1871, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.547, + 0.892, + 0.589 + ], + "angle": 0, + "content": "[19] Nikunj Kothari, Misha Gupta, Leena Vachhani, and Hemendra Arya. Pose estimation for an autonomous vehicle using monocular vision. pages 424-431, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.591, + 0.892, + 0.645 + ], + "angle": 0, + "content": "[20] Yann Labbe, Justin Carpentier, Mathieu Aubry, and Josef Sivic. CosyPose: Consistent multi-view multi-object 6D pose estimation. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.648, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[21] Taeyeop Lee, Byeong-Uk Lee, Inkyu Shin, Jaesung Choe, Ukcheol Shin, In So Kweon, and Kuk-Jin Yoon. UDA-COPE: unsupervised domain adaptation for category-level object pose estimation. CoRR, abs/2111.12580, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.704, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[22] Yi Li, Gu Wang, Xiangyang Ji, Yu Xiang, and Dieter Fox. Deepim: Deep iterative matching for 6d pose estimation. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.761, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[23] Haitao Lin, Zichang Liu, Chilam Cheang, Yanwei Fu, Guodong Guo, and Xiangyang Xue. Sar-net: Shape alignment and recovery network for category-level 6d object pose and size estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6707-6717, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[24] Jiehong Lin, Zewei Wei, Zhihao Li, Songcen Xu, Kui Jia, and Yuanqing Li. DualPoseNet: Category-level 6d object pose and size estimation using dual pose network with refined learning of pose consistency. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.895, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10701" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.093, + 0.469, + 0.12 + ], + "angle": 0, + "content": "IEEE/CVF International Conference on Computer Vision (ICCV), pages 3560-3569, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.123, + 0.469, + 0.177 + ], + "angle": 0, + "content": "[25] Jiehong Lin, Zewei Wei, Changxing Ding, and Kui Jia. Category-level 6d object pose and size estimation using self-supervised deep prior deformation networks, 2022. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.182, + 0.469, + 0.263 + ], + "angle": 0, + "content": "[26] Zhi-Hao Lin, Sheng-Yu Huang, and Yu-Chiang Frank Wang. Convolution in the cloud: Learning deformable kernels in 3d graph convolution networks for point cloud analysis. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1797-1806, 2020. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.267, + 0.469, + 0.335 + ], + "angle": 0, + "content": "[27] Zheng Linfang, Leonardis Ales, Tse Tze Ho, Elden, Horanyi Nora, Chen Hua, Zhang Wei, and Chang Hyung Jin. Tp-ae: Temporally primed 6d object pose tracking with auto-encoders. In 2022 IEEE International Conference on Robotics and Automation (ICRA), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.339, + 0.469, + 0.38 + ], + "angle": 0, + "content": "[28] Liyuan Liu, Haoming Jiang, Pengcheng He, Weizhu Chen, Xiaodong Liu, Jianfeng Gao, and Jiawei Han. On the variance of the adaptive learning rate and beyond, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.383, + 0.469, + 0.438 + ], + "angle": 0, + "content": "[29] Xingyu Liu, Gu Wang, Yi Li, and Xiangyang Ji. CATRE: Iterative point clouds alignment for category-level object pose refinement. In European Conference on Computer Vision (ECCV), pages 499-516. Springer, 2022. 1, 2, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.441, + 0.468, + 0.454 + ], + "angle": 0, + "content": "[30] Eitan Marder-Eppstein. Project tango. pages 25-25, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.457, + 0.469, + 0.539 + ], + "angle": 0, + "content": "[31] Nathaniel Merrill, Yuliang Guo, Xingxing Zuo, Xinyu Huang, Stefan Leutenegger, Xi Peng, Liu Ren, and Guoquan Huang. Symmetry and uncertainty-aware object slam for 6 dof object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14901-14910, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.543, + 0.469, + 0.61 + ], + "angle": 0, + "content": "[32] Andrew S. Morgan, Bowen Wen, Junchi Liang, Abdeslam Boularias, Aaron M. Dollar, and Kostas Bekris. Vision-driven Compliant Manipulation for Reliable; High-Precision Assembly Tasks. In Proceedings of Robotics: Science and Systems, Virtual, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.615, + 0.469, + 0.668 + ], + "angle": 0, + "content": "[33] Andrew Nee, S K Ong, George Chryssolouris, and Dimitris Mourtzis. Augmented reality applications in design and manufacturing. CIRP Annals - Manufacturing Technology, 61:657-679, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.673, + 0.469, + 0.726 + ], + "angle": 0, + "content": "[34] Van Nguyen Nguyen, Yinlin Hu, Yang Xiao, Mathieu Salzmann, and Vincent Lepetit. Templates for 3d object pose estimation revisited: Generalization to new objects and robustness to occlusions, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.731, + 0.469, + 0.784 + ], + "angle": 0, + "content": "[35] Charles R. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.788, + 0.469, + 0.854 + ], + "angle": 0, + "content": "[36] Mahdi Rad and Vincent Lepetit. BB8: A Scalable, Accurate, Robust to Partial Occlusion Method for Predicting the 3D Poses of Challenging Objects Without Using Depth. In IEEE International Conference on Computer Vision (ICCV), 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[37] Aleksandr Segal, Dirk Haehnel, and Sebastian Thrun. Generalized-icp. In Robotics: science and systems (RSS), page 435. Seattle, WA, 2009. 1" + }, + { + "type": "list", + "bbox": [ + 0.081, + 0.093, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.16 + ], + "angle": 0, + "content": "[38] Ivan Shugurov, Fu Li, Benjamin Busam, and Slobodan Ilic. Osop: A multi-stage one shot object pose estimation framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6835-6844, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[39] Yongzhi Su, Jason Rambach, Nareg Minaskan, Paul Lesur, Alain Pagani, and Didier Stricker. Deep multi-state object pose estimation for augmented reality assembly. In 2019 IEEE International Symposium on Mixed and Augmented Reality Adjunct (ISMAR-Adjunct), pages 222–227, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.234, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[40] Yongzhi Su, Mahdi Saleh, Torben Fetzer, Jason Rambach, Nassir Navab, Benjamin Busam, Didier Stricker, and Federico Tombari. Zebrapose: Coarse to fine surface encoding for 6 dof object pose estimation, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.289, + 0.892, + 0.369 + ], + "angle": 0, + "content": "[41] Jiaming Sun, Zihao Wang, Siyu Zhang, Xingyi He, Hongcheng Zhao, Guofeng Zhang, and Xiaowei Zhou. Onepose: One-shot object pose estimation without cad models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6825-6834, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.373, + 0.892, + 0.425 + ], + "angle": 0, + "content": "[42] Martin Sundermeyer, Maximilian Durner, En Yen Puang, Zoltan-Csaba Marton, Narunas Vaskevicius, Kai O. Arras, and Rudolph Triebel. Multi-path learning for object pose estimation across domains, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.428, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[43] Martin Sundermeyer, Zoltan Marton, Maximilian Durner, and Rudolph Triebel. Augmented Autoencoders: Implicit 3D Orientation Learning for 6D Object Detection. International Journal of Computer Vision (IJCV), 128, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.485, + 0.892, + 0.523 + ], + "angle": 0, + "content": "[44] Bugra Tekin, Sudipta Sinha, and Pascal Fua. Real-time seamless single shot 6d object pose prediction. pages 292-301, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.526, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[45] Meng Tian, Marcelo H Ang, and Gim Hee Lee. Shape prior deformation for categorical 6d object pose and size estimation. In European Conference on Computer Vision (ECCV), pages 530-546. Springer, 2020. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.582, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[46] Jonathan Tremblay, Thang To, Balakumar Sundaralingam, Yu Xiang, Dieter Fox, and Stan Birchfield. Deep object pose estimation for semantic robotic grasping of household objects, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.638, + 0.892, + 0.676 + ], + "angle": 0, + "content": "[47] Joel Vidal, Chyi-Yeu Lin, and Robert Martí. 6d pose estimation using an improved method based on point pair features, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.892, + 0.747 + ], + "angle": 0, + "content": "[48] Chen Wang, Danfei Xu, Yuke Zhu, Roberto Martin-Martin, Cewu Lu, Li Fei-Fei, and Silvio Savarese. DenseFusion: 6D Object Pose Estimation by Iterative Dense Fusion. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.749, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[49] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J Guibas. Normalized object coordinate space for category-level 6d object pose and size estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2642-2651, 2019. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.833, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[50] Jiaze Wang, Kai Chen, and Qi Dou. Category-level 6d object pose estimation via cascaded relation and recurrent reconstruction networks. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4807-4814. IEEE, 2021. 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10702" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[51] Bowen Wen, Chaitanya Mitash, Baozhang Ren, and Kostas Bekris. se(3)-TrackNet: Data-driven 6D Pose Tracking by Calibrating Image Residuals in Synthetic Domains. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.232 + ], + "angle": 0, + "content": "[52] Bowen Wen, Wenzhao Lian, Kostas Bekris, and Stefan Schaal. You Only Demonstrate Once: Category-Level Manipulation from Single Visual Demonstration. In Proceedings of Robotics: Science and Systems, New York City, NY, USA, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.471, + 0.303 + ], + "angle": 0, + "content": "[53] Heng Yang and Marco Pavone. Object pose estimation with statistical guarantees: Conformal keypoint detection and geometric uncertainty propagation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8947-8958, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.304, + 0.47, + 0.346 + ], + "angle": 0, + "content": "[54] Hongwei Yong, Jianqiang Huang, Xiansheng Hua, and Lei Zhang. Gradient centralization: A new optimization technique for deep neural networks, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.346, + 0.47, + 0.401 + ], + "angle": 0, + "content": "[55] Sergey Zakharov, Ivan Shugurov, and Slobodan Ilic. DPOD: 6D Pose Object Detector and Refiner. In The IEEE International Conference on Computer Vision (ICCV), 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.403, + 0.47, + 0.473 + ], + "angle": 0, + "content": "[56] Kun Zhang, Chen Wang, Hua Chen, Jia Pan, Michael Yu Wang, and Wei Zhang. Vision-based six-dimensional peg-in-hole for practical connector insertion. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 1771-1777, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.474, + 0.47, + 0.514 + ], + "angle": 0, + "content": "[57] Michael R. Zhang, James Lucas, Geoffrey Hinton, and Jimmy Ba. Lookahead optimizer: k steps forward, 1 step back, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.516, + 0.47, + 0.572 + ], + "angle": 0, + "content": "[58] Ruida Zhang, Yan Di, Zhiqiang Lou, Fabian Manhardt, Federico Tombari, and Xiangyang Ji. RBP-Pose: Residual bounding box projection for category-level pose estimation, 2022. 1, 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.573, + 0.47, + 0.627 + ], + "angle": 0, + "content": "[59] Ruida Zhang, Yan Di, Fabian Manhardt, Federico Tombari, and Xiangyang Ji. SSP-Pose: Symmetry-aware shape prior deformation for direct category-level object pose estimation, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.63, + 0.47, + 0.699 + ], + "angle": 0, + "content": "[60] Zhongqun Zhang, Wei Chen, Linfang Zheng, Ales Leonardis, and Hyung Jin Chang. Trans6d: Transformer-based 6d object pose estimation and refinement. In Computer Vision - ECCV 2022 Workshops, pages 112-128, Cham, 2023. Springer Nature Switzerland. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.7, + 0.471, + 0.796 + ], + "angle": 0, + "content": "[61] Linfang Zheng, Chen Wang, Yinghan Sun, Esha Dasgupta, Hua Chen, Aleš Leonardis, Wei Zhang, and Hyung Jin Chang. HS-Pose: Hybrid scope feature extraction for category-level object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17163-17173, 2023. 1, 2, 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.798, + 0.47, + 0.869 + ], + "angle": 0, + "content": "[62] Jun Zhou, Kai Chen, Linlin Xu, Qi Dou, and Jing Qin. Deep fusion transformer network with weighted vector-wise keypoints voting for robust 6d object pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13967-13977, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.869 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10703" + } + ] +] \ No newline at end of file diff --git a/2024/GeoReF_ Geometric Alignment Across Shape Variation for Category-level Object Pose Refinement/966fabe9-a4db-4529-94bf-c710cc01e232_origin.pdf b/2024/GeoReF_ Geometric Alignment Across Shape Variation for Category-level Object Pose Refinement/966fabe9-a4db-4529-94bf-c710cc01e232_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c9359c39b519d0e20da67058f73d49793043de5d --- /dev/null +++ b/2024/GeoReF_ Geometric Alignment Across Shape Variation for Category-level Object Pose Refinement/966fabe9-a4db-4529-94bf-c710cc01e232_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:15b54c8523042eb2b55f653c6b28bf5c6903a1a65d681fc77a54d51f12f160de +size 1161945 diff --git a/2024/GeoReF_ Geometric Alignment Across Shape Variation for Category-level Object Pose Refinement/full.md b/2024/GeoReF_ Geometric Alignment Across Shape Variation for Category-level Object Pose Refinement/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4609637e2c997876e5118f59a64f7227a08f63d3 --- /dev/null +++ b/2024/GeoReF_ Geometric Alignment Across Shape Variation for Category-level Object Pose Refinement/full.md @@ -0,0 +1,290 @@ +# GeoReF: Geometric Alignment Across Shape Variation for Category-level Object Pose Refinement + +Linfang Zheng $^{1,3}$ Tze Ho Elden Tse $^{*3}$ Chen Wang $^{*1,2}$ Yinghan Sun $^{1}$ Hua Chen $^{1}$ +Ales Leonardis $^{3}$ Wei Zhang $^{\dagger 1}$ Hyung Jin Chang $^{3}$ + +$^{1}$ Shenzhen Key Laboratory of Control Theory and Intelligent Systems, School of System Design and Intelligent Manufacturing, Southern University of Science and Technology, China $^{2}$ Department of Computer Science, the University of Hong Kong, China $^{3}$ School of Computer Science, University of Birmingham, UK + +{1xz948, txt994}@student.bham.ac.uk, cwang5@cs.hku.hk, sunyh2021@mail.sustech.edu.cn {chenh6,zhangw3}@sustech.edu.cn,{a.leonadis,h.j.chang}@bham.ac.uk + +# Abstract + +Object pose refinement is essential for robust object pose estimation. Previous work has made significant progress towards instance-level object pose refinement. Yet, category-level pose refinement is a more challenging problem due to large shape variations within a category and the discrepancies between the target object and the shape prior. To address these challenges, we introduce a novel architecture for category-level object pose refinement. Our approach integrates an HS-layer and learnable affine transformations, which aims to enhance the extraction and alignment of geometric information. Additionally, we introduce a crosscloud transformation mechanism that efficiently merges diverse data sources. Finally, we push the limits of our model by incorporating the shape prior information for translation and size error prediction. We conducted extensive experiments to demonstrate the effectiveness of the proposed framework. Through extensive quantitative experiments, we demonstrate significant improvement over the baseline method by a large margin across all metrics. + +# 1. Introduction + +Understanding an object's pose is crucial for a wide range of real-world applications, including robotic manipulation [18, 32, 52, 56], augmented reality [30, 33], and autonomous driving [19, 39]. Significant progress has been made for object pose estimation [5, 11, 15, 48, 53, 55] and pose refinement [2, 17, 22, 37, 51, 60] using the object's CAD + +![](images/02dd435b036b416f9fac1ef5e657fb95dac3b3fcce496b602397162ee4ebbb56.jpg) +Camera + +![](images/ab27819a1505b161a313cf61b448f90c0d8171b32f08746fe09be9bfa4341d2f.jpg) + +![](images/64d9494ff06a4439fe985d710e0f582c261f07e4f6005a01240c2a0772aa711c.jpg) + +![](images/8b40e3ca3f05cb9dc6c6ca2b50af538cac51cef53f44afa8a4133c1dc1e5317f.jpg) +Mug +SP-m + +![](images/e80710786aca18753def72fbe24cac96a512d803c7960d8f1c244b34adc9403f.jpg) +SP-1 +Figure 1. Illustration of the shape variation. $SP - m$ represents the category's mean shape, $SP - 1$ and $SP - 2$ represents the randomly sampled object shapes from the CAMERA25 training set. + +![](images/2f7f16aba523964b2ffe813e816347b5f1bd6101bf27957cf85a0edede474ba2.jpg) +SP-2 + +model. Despite the promising performance, the reliance on accurate instance-level CAD models limits their generalizability to everyday objects. Category-level methods [7, 23-25, 29, 59, 61] is therefore been proposed to overcome this limitation. The objective of this line of work focuses on estimating object poses within a category given category-level shape priors. As a result, they face unique challenges as there exist diverse shape variations in each object category. We illustrate these shape variations in Fig. 1. + +Recently, there have been remarkable advancements in category-level object pose estimation [9, 10, 25, 58, 61], primarily due to effective utilization of geometric information through 3D graph convolution [26]. In applications that require high precision, it is common to employ an object pose refinement procedure in conjunction with pose estimation. This involves an initial pose estimation algorithm determining the object pose, followed by a refinement step to further enhance the accuracy of the initially estimated pose by predicting and correcting its error. However, while instance-level object pose refinement has been extensively studied, + +category-level pose refinement remained unexplored until the introduction of CATRE [29]. By leveraging initial object pose and size estimations, CATRE achieves category-level pose refinement by iteratively aligning the observed target object point cloud with the category-level shape prior. This pipeline is shown to be effective by improving the accuracy of the initial pose and size estimations. + +While CATRE has proven to be effective in many scenarios, it is limited by the reliance on the PointNet [35] encoder, which is primarily designed for classification and segmentation tasks. This design choice limits its ability to capture essential and fine-grained geometric relationships for accurate pose estimation and refinement. This ability is particularly important in category-level pose refinement as there exists diverse shape variations between inputs. Consequently, CATRE obtains suboptimal performance by direct application of 3D graph convolution. In addition, as CATRE treats the point cloud and the shape prior features separately until a later stage of the network, they potentially miss out on the benefits of integrating these features earlier. Moreover, their approach does not incorporate shape prior information into the translation and scale estimation module, which presents another area for potential improvement. + +In this paper, we introduce a novel architecture for category-level object pose refinement which aims to address the limitations mentioned above. To better extract both local and global geometric information, we incorporate an HS layer into our feature extraction process. We apply learnable affine transformations to the features to address the geometric discrepancies between the observed point cloud and the shape prior. This enables the network to align these features more effectively. In addition, we propose a cross-cloud transformation mechanism that is specifically designed to enhance the merging of information between the observed point clouds and the shape prior. This mechanism enables more efficient integration of information between the two sources. Finally, we push the limit of our model by incorporating shape prior information to more accurately predict errors in translation and size estimation. + +Our extensive experimental results on two category-level object pose datasets demonstrate that our proposed model to be effective in addressing the problem of shape variations in category-level object pose refinement, and consequently outperforms the state-of-the-art significantly. To the best of our knowledge, our proposed method is the first to successfully address the shape variation issue which is common in category-level pose refinement. Specifically, to enable graph convolution to be effective in capturing geometric relationships between different shapes, we propose an adaptive affine transformation matrix that aligns the observed point clouds and the shape prior. Additionally, the proposed cross-cloud transformation mechanism effectively fuses features from different input point clouds and brings + +further performance improvements. + +Our contributions are as follows: + +- We introduce a novel architecture to specifically address the shape variations issue in category-level object pose refinement. Our proposed method results in consistent performance gain and exhibits better generalization ability. +- We propose a unique cross-cloud transformation mechanism which efficiently merges diverse information from observed point clouds and shape priors. +- We conduct extensive experiments on two category-level object pose datasets to validate our proposed method. On the REAL275 dataset, our method significantly outperforms SPD by $39.1\%$ increase in the $5^{\circ}5\mathrm{cm}$ metric. Additionally, we achieve $10.5\%$ improvement in the $10^{\circ}2\mathrm{cm}$ metric over the state-of-the-art method, CATRE. + +# 2. Related Work + +Instance-level object pose estimation and refinement. Instance-level approaches estimate the pose of the target object given known 3D CAD models. They can be briefly divided into correspondence matching methods and template matching methods. Correspondence matching methods [4, 8, 12, 13, 31, 36, 40, 41, 44, 46, 55, 62] matches the outstanding features of the observed object images with its model. Template matching methods [1, 14, 27, 34, 38, 42, 43, 47] compares the images or extracted features with the pre-generated templates. As the initial pose estimates can be noisy to various factors such as occlusions, object pose refinement [20, 22, 51] is shown to be useful in improving the performance of instance-level methods. Even though they achieved impressive over the target object, the reliance on object CAD models limited their generalizability for handling everyday objects. In this paper, we consider a more challenging problem setting where only the category-level shape prior is provided. + +Category-level object pose estimation and refinement. Both tasks mainly focus on addressing the shape variation between the objects. The pioneering work NOCS [49] tackles the shape discrepancy by recovering the normalized visible shape of the target object and achieving the pose by point cloud matching. A series of methods extend this structure by leveraging different information such as domain adaptation [21], different reconstruction space [3], shape prior [3, 16, 21, 45], and structural similarities [6, 25]. However, this line of work is often limited in speed due to the iterative point matching. Another series of work starts with FS-Net [9], which adopts 3D graph convolution (3D-GC) [26] to obtain geometric sensitivity. Due to its effectiveness and real-time performance, graph convolution is widely adopted in recent methods with an enhancement in directions including loss function [10], bounding box voting [58], and shape deformation [59]. HS-Pose [61] ex + +tends the geometric feature extraction from local to global, which enhances the capability to handle objects with complex shapes. The research on category-level refinement began recently with the proposal of CATRE [29]. It introduced an effective pipeline that leverages shape priors and a focalization strategy for pose refinement and effectively improves the initial pose estimations. In this paper, we extend the CATRE and tackle the geometric variation issue within the framework of category-level pose refinement. + +# 3. Methodology + +# 3.1. Problem Formulation + +In this paper, we tackle the problem of category-level object pose refinement. Given the initial pose and size estimation $(\mathbf{R}_0,\mathbf{t}_0,\mathbf{s}_0)$ , along with the observed point cloud $\mathcal{O}\in \mathbb{R}^{N^O\times 3}$ and the shape prior $\mathcal{P}\in \mathbb{R}^{N^P\times 3}$ , we aim to predict the estimation error $(\Delta \mathbf{R},\Delta \mathbf{t},\Delta \mathbf{s})$ between the initial estimations and the ground truths. The pose refinement algorithm $\phi$ can be described as: + +$$ +\left(\Delta \mathbf {R}, \Delta \mathbf {t}, \Delta \mathbf {s}\right) = \phi \left(\mathbf {R} _ {0}, \mathbf {t} _ {0}, \mathbf {s} _ {0}, \mathcal {O}, \mathcal {P}\right). \tag {1} +$$ + +This pose refinement algorithm $\phi$ can be applied iteratively to improve the refinement performance. + +# 3.2. Preliminaries + +Our proposed category-level object pose refinement framework builds upon two previous works, CATRE [29] and HS-layer [61], which we briefly review them in the following. + +CATRE. CATRE is the first framework that considers the problem of category-level pose refinement. It predicts the error between the ground truth and the estimated poses by aligning the input point clouds and the categorical shape priors. Specifically, the network architecture of CATRE consists of four components: a) point cloud focalization, b) shared encoder, c) rotation prediction, and d) translation and size prediction. In point clouds focalization, the observed point clouds $\mathcal{O}$ and the shape prior $\mathcal{P}$ are first aligned with the initial pose and size estimation $[\mathbf{R}_0, \mathbf{t}_0, \mathbf{s}_0]$ : + +$$ +\begin{array}{l} \hat {\mathcal {O}} = \left\{\hat {o} _ {i} | \hat {o} _ {i} = o _ {i} - \mathbf {t} _ {0}, o _ {i} \in \mathcal {O} \right\}, \\ \hat {\mathcal {P}} = \left\{\hat {p} _ {i} | \hat {p} _ {i} = \operatorname {d i a g} \left(\mathbf {s} _ {0}\right) \mathbf {R} _ {0} p _ {i}, p _ {i} \in \mathcal {P} \right\}, \\ \end{array} +$$ + +where $\mathrm{diag}(\cdot)$ converts a vector to a diagonal matrix. The focalized observed point cloud $\hat{\mathcal{O}}$ and the focalized shape prior points $\hat{\mathcal{P}}$ contain full information required to predict the estimation error $(\Delta \mathbf{R},\Delta \mathbf{t},\Delta \mathbf{s})$ . First, a PointNet-based shared encoder is used to extract features from the two focalized point clouds independently. Then, both the extracted features are used for $\Delta \mathbf{R}$ estimation, while the global feature of the focalized observed point cloud along with the $\mathbf{s}_0$ are used for $\Delta \mathbf{t}$ and $\Delta \mathbf{s}$ . The initial estimates + +are updated using the predicted error $(\Delta \mathbf{R}, \Delta \mathbf{t}, \Delta \mathbf{s})$ . Finally, the updated estimates are used to predict the error again in which this process is iterative and the estimations are refined progressively and continuously. + +HS-layer. The Hybrid-Scope Geometric Feature Extraction Layer (HS-layer) is a simple network structure based on 3D graph convolution. It consists of two parallel paths that extract different scopes of features from the point cloud. The first path encodes the size and translation information of the target object. Meanwhile, the second path extracts outlier-robust local and global geometric features by applying graph convolution with a strategy of Receptive Fields with Feature Distance (RF-F) metric, alongside an Outlier Robust Feature Extraction Layer (ORL). These properties are particularly beneficial for category-level object pose estimation tasks. For more details, please refer to [61]. + +# 3.3. Overall Structure of GeoReF + +The overall framework of our proposed object pose refinement approach is shown in Fig. 2. This framework comprises three principal components: 1) point cloud focalization, 2) feature extraction, and 3) pose error prediction. We follow CATRE and use the same point cloud focalization module. We apply focalization and extract features from both the observed point cloud and the shape prior by our Feature Extraction component. Then, we predict the estimation errors $(\Delta \mathbf{R}, \Delta t, \Delta s)$ using the extracted features in the pose error prediction component. + +# 3.4. Graph Convolution with Learnable Affine Transformation (LAT) + +Geometric structural information is effective in estimating an object's pose for category-level object pose estimations. However, as shown in ablation study [AS-1], directly applying the 3D graph convolutions (e.g., HS-Encoder [61] and 3DGCN-Encoder [26]) to category-level object pose refinement tasks results in poor performance. This is due to the differences in task nature between the pose estimation and the pose refinement. In the pose estimation task, the network only needs to extract geometric structural information from a single point cloud. However, in the pose refinement framework, it requires extracting the geometric structural information from the two input point clouds as well as establishing the geometric correspondences between different object shapes. This becomes challenging due to the issue of shape variation in category pose refinement. + +To address the aforementioned problem, we propose to use learnable affine transformations (LATs). By employing LATs, the network can dynamically adjust the input point cloud and the point features which enables better establishment of geometric correspondences between the two different input shapes. Specifically, we apply three LATs (as shown in the bottom left of Fig. 2, where the Matrix Net + +![](images/039b1a85f9060f9d5bce9cd1f04401193aeeb56fc69c4c463c05cae237fe2de1.jpg) +Figure 2. Overall structure of the proposed method. Our object pose refinement structure contains three main modules. Given the shape prior point cloud, the target object's observed point cloud, and the initial estimation, we first apply point cloud focalization on the input point clouds using the initial estimation. The focalized point clouds then go through a geometric-based feature extraction encoder to obtain geometric structural features. The extracted features are then fed into two branches for rotation error estimation, translation error, and size error estimation. Within the HS Feature Extractor, the Matrix Net models output the learnable affine transformations (LATs) for adaptive point and feature adjustment. The left output of the Matrix Net adjusts the input point clouds, while the right Matrix Net model outputs two affine transformations for adjusting the rotation features, and the translation and size features. + +outputs the learnable affine transformations): The first LAT $M \in \mathbb{R}^3$ is applied to the input point cloud in the Euclidean space. The second LAT $M_{\mathrm{ts}}$ is applied to the extracted translation and size features $f_{\mathrm{ts}}$ . The third LAT $M_r$ is applied to the extracted rotation feature $f_r$ . With this approach, our method can better utilize the valuable geometric features in pose refinement. + +# 3.5. Cross-Cloud Transformation (CCT) for Information Mixing + +In pose refinement, effectively blending information from the focalized observed object and the shape prior is crucial for enabling the network to align them accurately. However, in CATRE, the data from the observed point cloud and the shape prior are processed independently until the late rotation prediction stage, where they are merely concatenated, limiting the effectiveness of the alignment. To address this problem, we introduce a novel cross-cloud transformation mechanism that effectively mixes the geometric information from the shape of prior features into the features of the observed point cloud. In particular, we use the feature transformation matrices $M_r^P$ , $M_{\mathrm{ts}}^P$ of shape prior to transforming the features of the observed point cloud: + +$$ +f _ {r} ^ {O} = M _ {r} ^ {P} f _ {r} ^ {O}, \tag {3} +$$ + +$$ +f _ {\mathrm {t s}} ^ {O} = M _ {\mathrm {t s}} ^ {P} f _ {\mathrm {t s}} ^ {O}. \tag {4} +$$ + +# 3.6. Integrating Shape Prior in Pose Estimation + +The information contained in the shape prior is crucial for the network to align the observed point cloud and the shape + +prior. For the rotation error prediction, the information contained in the shape prior is the essential information. For the translation and size prediction, this information can also be utilized by the network to adjust the learned geometric features accordingly. Therefore, unlike CATRE, which relied solely on features extracted from the observed point cloud to predict $(\Delta \mathbf{t}, \Delta \mathbf{s})$ , our approach also incorporates the information from shape prior to predict them. In particular, we not only mix the information using previous CCT mechanism, but also concatenate the features from shape prior and observed point cloud to obtain mixed features in the similarly way as the rotation estimation. We utilize $f_{\mathrm{t}}^{\mathrm{OP}}$ and $f_{\mathrm{s}}^{\mathrm{OP}}$ which contain both information from shape prior and observed point cloud like $f_{r}^{\mathrm{OP}}$ to predict $(\Delta \mathbf{t}, \delta)$ . + +We use two pose error predictors of the same network architecture to predict the rotation error and the translation and size error, respectively. Note that the weights of these two pose error predictors are not shared. The network structure of the pose error predictor is shown in Fig. 2. The pose predictor takes in two features, passes them through two same paths separately, and obtains two vectors in $\mathbb{R}^3$ . In the translation and size branch, the pose error predictor takes in $f_{\mathrm{t}}^{\mathrm{OP}}$ and $f_{\mathrm{s}}^{\mathrm{OP}}$ and passes them through the two paths, and the output two vectors are regarded as $\Delta \mathbf{t}$ and $\Delta \mathbf{s}$ , respectively. For the rotation error prediction, the mixed rotation features $f_{r}^{\mathrm{OP}}$ are copied and passed through the two paths in the pose error predictor. The two output vectors are regarded as $r_x$ and $r_y$ , where $r_x$ and $r_y$ are the first and second axes of the rotation error matrix $\Delta \mathbf{R}$ . The third column $r_z$ of $\Delta \mathbf{R}$ can + +be found by: + +$$ +r _ {z} = r _ {x} \times r _ {y}. \tag {5} +$$ + +# 4. Experiments + +Implementation details. We implement and experiment with our method using an RTX 4090 GPU with a batch size of 12 and 150 training epochs. We follow CATRE [29] and adopt its loss functions and the basic data augmentation strategies including random dropping points, adding Gaussian noise, random pose perturbations, etc. We set the number of points for both the observed points and shape prior to be 512. We train the network using Ranger optimizer [28, 54, 57] with a base learning rate of $10^{-4}$ and anneal the learning rate from the $72\%$ of the total epoch based on cosine schedule. + +Baselines. We use CATRE as the baseline for our ablation study as it is the state-of-the-art category-level object refinement method. As CATRE did not provide the results of $\mathrm{IoU}_{50}$ , we obtained them by using their official pre-trained model and kept the rest of the reported metric scores consistent with the corresponding paper. For fair comparisons, we use the same initial estimations as CATRE, which is the pose estimation results of SPD [45]. The result of replacing PointNet with the 3DGC-Encoder in the ablation study is provided by CATRE. We also apply our method to other state-of-the-art category-level object pose estimation approaches [10, 25, 61] to demonstrate the effectiveness of our proposed refinement method. For the pose refinement on HS-Pose [61] and RBP-Pose [58], we compute the initial estimations using their official pre-trained models. The results of other methods are taken directly from their paper. + +Datasets. As we focus on the problem of shape variation between input object point clouds, we choose two popular category-level object pose estimation benchmarks to verify our approach, i.e., REAL275 [49] and CAMERA25 [49]. They both contain 6 object categories with multiple levels of shape complexities, i.e., bowl, can, bottle, laptop, mug, and camera. REAL275 contains 36 objects in 13 real-world scenes with 7k RGB-D images in total. Among them, 16 objects in 7 scenes are used for training, resulting in 4.3k images in training. CAMERA25 is a large synthetic RGB-D dataset. It provides 1085 objects and 275k RGB-D images for training, and 184 objects and 25k images for testing. + +Evaluation metrics. Following [29, 61], we evaluate our method using: 1) The mean average precision (mAP) of the 3D Intersection over Union $(IoU)$ at different thresholds $(50\%$ and $75\%)$ to evaluate the pose and size estimation together. 2) The pose metric at $n^{\circ}mcm$ defines a pose as + +correct if the rotation error is below $n^{\circ}$ and the translation error is below $m$ cm. Here, we use $5^{\circ}, 10^{\circ}, 2\mathrm{cm}$ , and $5\mathrm{cm}$ as the thresholds. + +# 4.1. Ablation study + +To verify the proposed architecture, we conducted comprehensive ablation studies on the REAL275 dataset using the initial pose estimations from SPD [45]. We present a quantitative comparison of our method with various key components disabled to motivate our design choices in Table 1. Full results of the ablation study are reported in the supplementary materials. + +[AS-1] Using geometric features directly. To illustrate the limitations of existing geometric-based encoder under the problem of shape variations, we replace the encoder of CATRE with two robust geometric-based point cloud convolutional structures, namely 3DGCN-Encoder [26] and HS-Encoder [61]. 3DGCN is a widely adopted graph convolution in existing category-level object pose estimation algorithms, while HS-Encoder is a recent architecture that achieves state-of-the-art performance in category-level object pose estimation. However, as shown in Table 1 [C0, C1], even though both HS-Encoder and 3DGCN-Encoder are powerful in finding an object's pose from individual inputs, they failed to manage the pose refinement scenarios when there exist shape variations between the target object and the shape prior. We observed both encoders result in a performance drop when compared to the original CATRE with IoU75 of 37.3% (HS-Encoder) vs. 43.7% (CATRE), 5°5cm of 43.4% (3DGCN-Encoder) vs. 53.3% (CATRE). + +[AS-2] Use prior features in translation and scale estimation. To validate that the information of shape prior is also important in scale and translation error prediction, we add the features of the prior shape to the scale and translation branch by using the same network architecture as the rotation branch. As shown in Table 1 [D0], incorporating shape prior information in translation and size estimation enhanced the overall performance by $2.2\%$ improvement on $5^{\circ}2\mathrm{cm}$ metric and $2.4\%$ on $10^{\circ}2\mathrm{cm}$ metric. + +[AS-3] Use learnable affine transformation (LAT) for geometric features. To demonstrate the effectiveness of LAT in addressing the shape variation issue, we conduct ablation studies on applying the proposed LAT to the input point cloud and the extracted geometric features in the feature space. The result is shown in Table 1 [E0]. Compared to using geometric features directly (Table 1 [C0]), LATs bring a significant boost on all the metrics, with $\mathrm{IoU}_{75}$ improved by $20.9\%$ , $5^{\circ}2\mathrm{cm}$ improved by $10.5\%$ , + +Table 1. Ablation studies on REAL275. +Higher score indicates better performance. In the 'Row' column, the code in bold means the strategies taken in the final structure. In the 'Method' column, the notation $X:Y$ denotes module $Y$ from structure $X$ , $X + Y$ means add module $Y$ to $X$ , and $X \rightarrow Y$ indicates replacing $X$ with $Y$ . + +
RowMethodIoU50IoU755°2cm5°5cm10°2cm10°5cm2cm5°
A0CATRE[45] (baseline)77.043.645.854.461.473.175.158.0
B0Ours: E0 + Cross-Cloud Transformation79.2 2.2↑51.8 8.2↑54.4 8.6↑60.3 5.9↑71.9 10.5↑79.4 6.3↑81.9 6.8↑64.3 6.3↑
C0A0: PointNet → HS-Encoder71.030.141.945.960.670.371.948.7
C1A0: PointNet → 3DGCN-Encoder-28.436.043.4--68.047.7
D0A0 + prior in ST branch77.145.848.054.663.872.577.959.2
E0D0: PointNet → HS-layer+LAT79.451.052.458.669.477.780.462.4
E1B0: No LAT on input points76.139.346.653.065.474.878.058.2
E2B0: No LAT on features78.548.847.453.067.475.080.457.4
E3B0: No LAT on the rotation feature79.850.650.456.268.676.380.260.8
F0E0+ Global Concatenation Fusion77.748.447.854.567.175.280.159.4
+ +and $5^{\circ}5\mathrm{cm}$ improved by $12.7\%$ . The resulted network also significantly outperforms the PointNet-based encoder (Table 1[D0]) on all the metrics with $\mathrm{IoU}_{75}$ and $10^{\circ}5\mathrm{cm}$ improved by $5.2\%$ , $5^{\circ}2\mathrm{cm}$ improved by $6.4\%$ , and $10^{\circ}2\mathrm{cm}$ improved by $5.6\%$ . These results verified the effectiveness of LAT on geometric features. + +[AS-4] The influence of each learnable affine transformation (LAT). To further demonstrate the influences of each LAT, we conducted three experiments by gradually disabling LAT from the framework: 1) without the LAT on the input point cloud, 2) without applying LATs on features, and 3) without independent LATs on the rotation features, where a single LAT is used for the rotation, translation, and scale features. The results are shown in Table 1 [E1-E3]. Compared to the directly using geometric features (Table 1 [C0]), we show that LAT can significantly enhance the network by around $10\%$ improvements on $\mathrm{IoU}_{75}$ and $7\%$ on $5^{\circ}5\mathrm{cm}$ metric in Table 1 [E1-E3]. We verify that the combination of them consistently results in better performance. + +[AS-5] Cross-cloud transformation (CCT) based feature fusion To demonstrate that it is important to have a good feature fusion strategy in the feature extraction phase, we conducted experiments on two different feature fusion strategies. One of them is the widely adopted feature fusion strategy, where the global feature of one point cloud is concatenated with the features of another point cloud and then goes through convolutional layers for feature fusion. Another one is the proposed CCT-based feature fusion. As shown in Table 1 [F0], applying global concatenation-based fusion does not enhance the overall performance, and even results in worse performance in all the metrics with the $5^{\circ}2\mathrm{cm}$ significantly decreased by $4.6\%$ , and $5^{\circ}5\mathrm{cm}$ decreased by $3.1\%$ (compared to Table 1 [D0]). On the contrary, as shown in Table 1 [B0], our simple CCT-based feature fusion shows its effectiveness by improving all the pose + +![](images/80e8d47f5b934956f66a7f331dad000b3ff466c9344aa828b320c4dcf51b0fa5.jpg) +Figure 3. Performance comparison of our method and CATRE under different shape priors. $SP - m$ denotes the category's mean shape, while $SP - 1$ and $SP - 2$ are randomly sampled object shapes from CAMERA25. + +metrics by around $2.0\%$ . In Supplementary, we also visualize the influence of CCT in feature space. + +[AS-6] Handle shape variations. To demonstrate that the proposed method can handle the shape variations, we replaced the original shape prior with two randomly sampled models from CAMERA25 training set and trained on REAL275. Note that the original shape prior represents the mean shape of a category, and the new models are randomly sampled. Therefore, there are larger shape variations with certain target objects. As shown in Fig. 3, CATRE performs best when using the mean shape of the category, while its performance drops dramatically on the randomly sampled shape priors. In contrast, our method exhibits robustness to shape variations introduced by different shape priors and consistently delivers strong performance. + +[AS-7] Refinement with different initial estimations. To demonstrate our model robustness to different initial estimations, we compare the proposed method with CATRE on different initial estimations generated by category-level + +![](images/1996b314fa576aac0fbcadfd6781397fddb6746998eb12cf02891ac3b40a9684.jpg) +Figure 4. Qualitative comparison of proposed (row #3) and baseline (row #2) methods using SPD (row #1) as initial estimation. Ground truth shown with white lines. Note that the estimated rotations of symmetric objects (e.g. bowl, bottle, and can) are considered correct if the symmetry axis is aligned. + +object pose estimation methods with varying performance, including HS-Pose [61], GPVPose [10], and the SelfDPDN [25]. As shown in Table 2, our method consistently improves the initial estimations by a large margin across different metrics. For example, we improved the SelfDPND and GPV-Pose on strict $5^{\circ}2\mathrm{cm}$ metric by $9.6\%$ and $15.4\%$ , respectively. We boosted the $\mathrm{IoU}_{75}$ of the GPV-Pose and the HS-Pose by $26.5\%$ and $15.2\%$ , respectively. However, the baseline CATRE fails to refine the initial poses of SelfDPDN. Also, CATRE reaches its limit when refining high-accuracy initial estimations such as the initial poses generated by HS-Pose, resulting in performance drops on $10^{\circ}5\mathrm{cm}$ and $10^{\circ}2\mathrm{cm}$ . The comparison results demonstrated the robustness and capability of our method to different initial estimations. We report full details in the supplementary. + +# 4.2. Generalizably test on the CAMERA25 dataset + +In real-world applications, category-level algorithms often need to generalize across diverse testing scenarios, encountering a larger number of objects than represented in their training sets. To simulate this problem setting, we choose CAMERA25 as it provides more than 25K RGB-D testing images. We train our model using only a mini set (2% and 4%) of the CAMERA25 training data, resulting in a training set of around 5K and 10K images from a total of 275K images. As shown in Table 3, the performance of the baseline method decreased dramatically when using 2% of the training images, with IoU75 dropped by 12.9% and 5°2cm dropped by 9.0% (see [B0]). On the contrary, our + +Table 2. Comparison of the pose refinement with the baseline method CATRE on REAL275 with different initial estimations. Each comparison group contains 3 methods: the initial pose estimation method, refinement using CATRE, and refinement using our method, respectively. + +
MethodIoU755°2cm5°5cm10°2cm10°5cm
Self-DPDN [25]42.244.350.965.178.6
Self-DPDN + CATRE0.422↓0.344.0↓5.145.8↓0.464.7↓6.971.7↓
Self-DPDN + Ours49.77.5↑53.99.6↑60.19.2↑75.09.9↑82.84.2↑
GPV-Pose [10]23.1*32.042.955.073.3
GPV-Pose + CATRE42.619.5↑39.77.7↑54.111.2↑57.12.1↑78.04.7↑
GPV-Pose + Ours49.626.5↑47.415.4↑57.814.9↑68.113.1↑81.27.9↑
HS-Pose [61]39.1*46.555.268.682.7
HS-Pose + CATRE47.18.0↑48.72.2↑59.13.9↑67.80.8↓81.21.5↓
HS-Pose + Ours54.315.2↑51.75.2↑59.64.4↑74.35.7↑83.81.1↑
+ +method exhibits a much higher performance when using small datasets for training. Specifically, our method can already outperforms the fully-trained CATRE with only $2\%$ of images (see [B1]). We also report the performance using $4\%$ images of the training set in Table 3. It can be seen that, despite CATRE use full CAMERA25 training set, our method outperforms it with $\mathrm{IoU}_{75}$ and $5^{\circ}2\mathrm{cm}$ improved by $3.1\%$ and $3.6\%$ , respectively. + +# 4.3. Comparison with state-of-the-arts + +REAL275. We conduct pose refinement on SPD [45] using the proposed approach and compare the resulting performance with the state-of-the-art category-level object pose estimation and refinement methods. As shown in Table 4, our method significantly improves the performance of SPD on all the metrics, with $5^{\circ}5\mathrm{cm}$ enhanced + +Table 3. The generalizability test on the CAMERA25 dataset. + +
RowMethodTrain Data SizeIoU755°2cm5°5cm10°2cm10°5cm2cm5°
A0CATRE275K76.175.480.383.389.3--
B0CATRE5K63.266.472.379.487.488.873.3
B1Ours5K77.575.481.183.490.091.082.3
C0CATRE10K66.569.775.581.889.189.976.7
C1Ours10K79.277.984.083.890.592.085.4
+ +Table 4. Comparison with other methods on REAL275. + +
MethodIoU755°2cm5°5cm10°2cm10°5cm
NOCS [49]9.47.210.013.825.2
DualPoseNet [24]30.829.335.950.066.8
CR-Net [50]33.227.834.347.260.8
SGPA [6]37.135.939.661.370.7
RBP-Pose [58]24.538.248.163.179.2
GPV-Pose [10]23.132.042.955.073.3
HS-Pose [61]39.146.555.268.682.7
SPD* [45]27.019.121.243.554.0
SPD*+CATRE [29]43.645.854.461.473.1
SPD*+Ours51.854.460.371.979.4
+ +by $39.1\%$ , $5^{\circ}5\mathrm{cm}$ improved by $35.3\%$ , $\mathrm{IoU}_{75}$ enhanced by $24.8\%$ , $10^{\circ}2\mathrm{cm}$ improved by $24.4\%$ , and $10^{\circ}5\mathrm{cm}$ improve by $25.4\%$ . In comparison to our baseline, CATRE, our proposed method demonstrates a substantial improvement across various performance metrics. Specifically, we observe a remarkable enhancement of $10.5\%$ in $10^{\circ}2\mathrm{cm}$ , $8.6\%$ in $5^{\circ}2\mathrm{cm}$ , $7.2\%$ in $\mathrm{IoU}_{75}$ , and $6.3\%$ in $10^{\circ}5\mathrm{cm}$ . In addition, compared with SOTA pose estimation methods, the pose estimation results of applying our proposed refinement method on SPD significantly outperformed these methods's results by a large margin. Specifically, the estimation results of applying our method on SPD rank top on 4 out of 5 metrics and rank second on the rest metric, and achieved a $7.9\%$ increase on the $5^{\circ}2\mathrm{cm}$ metric and $5.1\%$ enhancement on $5^{\circ}5\mathrm{cm}$ . It is worth noting that the purpose of this section is not to compare refinement and estimation methods, as they are designed to address different problems. Instead, our objective here is to demonstrate how our proposed refinement approach can improve the performance of existing pose estimation methods. Therefore, even though our refinement on HS-Pose produced better performance, as mentioned in the ablation study, we choose to refine weaker initial estimations to show the capability of our approach. + +CAMERA25. Our method outperforms the state-of-the-art methods using only $2\%$ of the training image set. For more details, please refer to the supplementary. + +Qualitative examples. We provide qualitative comparisons of the pose estimation results by SPD, CATRE on SPD, and our method on SPD in Fig. 4 and 5. As shown in Fig. 4, our method on SPD achieves the best size and pose + +![](images/6e029e7926f2ec3140981561b68e0b5eb8de2f23237bac2089d17c2122c31d25.jpg) +CATRE +Figure 5. Comparison of proposed (row #2) and baseline (row #1) methods) during a complete refinement iteration, both using SPD as initial estimation. The ground truth is represented by white lines. + +estimations. In particular, by considering the first column of Fig. 4, all of the comparison methods struggle to estimate the orientation of the camera category. We also provide qualitative examples on the iteration process in Fig. 5. Our method demonstrates faster convergence and more accurate final result than CATRE. + +# 5. Conclusion + +In this work, we proposed a novel category-level object pose refinement method which targeted at addressing the challenge of shape variation. We shown that the geometric structural information can be aligned by our adaptive affine transformations. We also demonstrated that the crosscloud transformation mechanism can efficiently merges information from distinct point clouds. We further incorporated shape prior information and observed improvements in translation and size predictions. We verified that each of our technical components contributed meaningfully through extensive ablations. We believe our method sets a strong baseline for future study and opens up new possibilities to handling more complex shapes, i.e. articulated objects. + +# Acknowledgments + +This research was supported in part by the MSIT (Ministry of Science and ICT), Korea, under the ITRC (Information Technology Research Center) support program (IITP-2024-2020-0-01789) supervised by the IITP (Institute for Information & Communications Technology Planning & Evaluation), National Natural Science Foundation of China under Grant No.62073159, and the Shenzhen Key Laboratory of Control Theory and Intelligent Systems, under grant No.ZDSYS20220330161800001. + +# References + +[1] Dingding Cai, Janne Heikkilä, and Esa Rahtu. Ove6d: Object viewpoint encoding for depth-based 6d object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6803-6813, 2022. 2 +[2] Pedro Castro and Tae-Kyun Kim. Crt-6d: Fast 6d object pose estimation with cascaded refinement transformers. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 5746-5755, 2023. 1 +[3] Dengsheng Chen, Jun Li, Zheng Wang, and Kai Xu. Learning canonical shape space for category-level 6d object pose and size estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2 +[4] Hansheng Chen, Pichao Wang, Fan Wang, Wei Tian, Lu Xiong, and Hao Li. Epro-pnp: Generalized end-to-end probabilistic perspective-n-points for monocular object pose estimation, 2022. 2 +[5] Hanzhi Chen, Fabian Manhardt, Nassir Navab, and Benjamin Busam. Texpose: Neural texture learning for self-supervised 6d object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4841-4852, 2023. 1 +[6] Kai Chen and Qi Dou. SGPA: Structure-guided prior adaptation for category-level 6d object pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2773-2782, 2021. 2, 8 +[7] Kai Chen, Stephen James, Congying Sui, Yun-Hui Liu, Pieter Abbeel, and Qi Dou. Stereopose: Category-level 6d transparent object pose estimation from stereo images via back-view nocs. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2855-2861, 2023. 1 +[8] Wei Chen, Xi Jia, Hyung Jin Chang, Jinming Duan, and Ales Leonardis. G2L-Net: Global to Local Network for Real-Time 6D Pose Estimation With Embedding Vector Features. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2 +[9] Wei Chen, Xi Jia, Hyung Jin Chang, Jinming Duan, Linlin Shen, and Ales Leonardis. Fs-net: Fast shape-based network for category-level 6d object pose estimation with decoupled rotation mechanism. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1581-1590, 2021. 1, 2 +[10] Yan Di, Ruida Zhang, Zhiqiang Lou, Fabian Manhardt, Xiangyang Ji, Nassir Navab, and Federico Tombari. GPV-Pose: Category-level object pose estimation via geometry-guided point-wise voting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6781-6791, 2022. 1, 2, 5, 7, 8 +[11] Yang Hai, Rui Song, Jiaojiao Li, Mathieu Salzmann, and Yinlin Hu. Rigidity-aware detection for 6d object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8927-8936, 2023. 1 + +[12] Rasmus Laurvig Haugaard and Anders Glent Buch. Surfemb: Dense and continuous correspondence distributions for object pose estimation with learnt surface embeddings. CoRR, abs/2111.13489, 2021. 2 +[13] Yisheng He, Wei Sun, Haibin Huang, Jianran Liu, Haoqiang Fan, and Jian Sun. Pvn3d: A deep point-wise 3d keypoints voting network for 6 dof pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2 +[14] Stefan Hinterstoisser, Vincent Lepetit, Naresh Rajkumar, and Kurt Konolige. Going further with point pair features. In Computer Vision – ECCV 2016, pages 834–848. Springer International Publishing, 2016. 2 +[15] Tomas Hodan, Daniel Barath, and Jiri Matas. Epos: Estimating 6d pose of objects with symmetries. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1 +[16] Muhammad Zubair Irshad, Sergey Zakharov, Rares Ambrus, Thomas Kollar, Zsolt Kira, and Adrien Gaidon. Shapo: Implicit representations for multi object shape appearance and pose optimization. 2022. 2 +[17] Shun Iwase, Xingyu Liu, Rawal Khirodkar, Rio Yokota, and Kris M Kitani. Repose: Fast 6d object pose refinement via deep texture rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3303-3312, 2021. 1 +[18] Daniel Kappler, Franziska Meier, Jan Issac, Jim Mainprice, Cristina Garcia Cifuentes, Manuel Wüthrich, Vincent Berenz, Stefan Schaal, Nathan Ratliff, and Jeannette Bohg. Real-time perception meets reactive motion generation. IEEE Robotics and Automation Letters, 3(3):1864-1871, 2018. 1 +[19] Nikunj Kothari, Misha Gupta, Leena Vachhani, and Hemendra Arya. Pose estimation for an autonomous vehicle using monocular vision. pages 424-431, 2017. 1 +[20] Yann Labbe, Justin Carpentier, Mathieu Aubry, and Josef Sivic. CosyPose: Consistent multi-view multi-object 6D pose estimation. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2 +[21] Taeyeop Lee, Byeong-Uk Lee, Inkyu Shin, Jaesung Choe, Ukcheol Shin, In So Kweon, and Kuk-Jin Yoon. UDA-COPE: unsupervised domain adaptation for category-level object pose estimation. CoRR, abs/2111.12580, 2021. 2 +[22] Yi Li, Gu Wang, Xiangyang Ji, Yu Xiang, and Dieter Fox. Deepim: Deep iterative matching for 6d pose estimation. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 1, 2 +[23] Haitao Lin, Zichang Liu, Chilam Cheang, Yanwei Fu, Guodong Guo, and Xiangyang Xue. Sar-net: Shape alignment and recovery network for category-level 6d object pose and size estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6707-6717, 2022. 1 +[24] Jiehong Lin, Zewei Wei, Zhihao Li, Songcen Xu, Kui Jia, and Yuanqing Li. DualPoseNet: Category-level 6d object pose and size estimation using dual pose network with refined learning of pose consistency. In Proceedings of the + +IEEE/CVF International Conference on Computer Vision (ICCV), pages 3560-3569, 2021. 8 +[25] Jiehong Lin, Zewei Wei, Changxing Ding, and Kui Jia. Category-level 6d object pose and size estimation using self-supervised deep prior deformation networks, 2022. 1, 2, 5, 7 +[26] Zhi-Hao Lin, Sheng-Yu Huang, and Yu-Chiang Frank Wang. Convolution in the cloud: Learning deformable kernels in 3d graph convolution networks for point cloud analysis. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1797-1806, 2020. 1, 2, 3, 5 +[27] Zheng Linfang, Leonardis Ales, Tse Tze Ho, Elden, Horanyi Nora, Chen Hua, Zhang Wei, and Chang Hyung Jin. Tp-ae: Temporally primed 6d object pose tracking with auto-encoders. In 2022 IEEE International Conference on Robotics and Automation (ICRA), 2022. 2 +[28] Liyuan Liu, Haoming Jiang, Pengcheng He, Weizhu Chen, Xiaodong Liu, Jianfeng Gao, and Jiawei Han. On the variance of the adaptive learning rate and beyond, 2019. 5 +[29] Xingyu Liu, Gu Wang, Yi Li, and Xiangyang Ji. CATRE: Iterative point clouds alignment for category-level object pose refinement. In European Conference on Computer Vision (ECCV), pages 499-516. Springer, 2022. 1, 2, 3, 5, 8 +[30] Eitan Marder-Eppstein. Project tango. pages 25-25, 2016. 1 +[31] Nathaniel Merrill, Yuliang Guo, Xingxing Zuo, Xinyu Huang, Stefan Leutenegger, Xi Peng, Liu Ren, and Guoquan Huang. Symmetry and uncertainty-aware object slam for 6 dof object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14901-14910, 2022. 2 +[32] Andrew S. Morgan, Bowen Wen, Junchi Liang, Abdeslam Boularias, Aaron M. Dollar, and Kostas Bekris. Vision-driven Compliant Manipulation for Reliable; High-Precision Assembly Tasks. In Proceedings of Robotics: Science and Systems, Virtual, 2021. 1 +[33] Andrew Nee, S K Ong, George Chryssolouris, and Dimitris Mourtzis. Augmented reality applications in design and manufacturing. CIRP Annals - Manufacturing Technology, 61:657-679, 2012. 1 +[34] Van Nguyen Nguyen, Yinlin Hu, Yang Xiao, Mathieu Salzmann, and Vincent Lepetit. Templates for 3d object pose estimation revisited: Generalization to new objects and robustness to occlusions, 2022. 2 +[35] Charles R. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2 +[36] Mahdi Rad and Vincent Lepetit. BB8: A Scalable, Accurate, Robust to Partial Occlusion Method for Predicting the 3D Poses of Challenging Objects Without Using Depth. In IEEE International Conference on Computer Vision (ICCV), 2017. +[37] Aleksandr Segal, Dirk Haehnel, and Sebastian Thrun. Generalized-icp. In Robotics: science and systems (RSS), page 435. Seattle, WA, 2009. 1 + +[38] Ivan Shugurov, Fu Li, Benjamin Busam, and Slobodan Ilic. Osop: A multi-stage one shot object pose estimation framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6835-6844, 2022. 2 +[39] Yongzhi Su, Jason Rambach, Nareg Minaskan, Paul Lesur, Alain Pagani, and Didier Stricker. Deep multi-state object pose estimation for augmented reality assembly. In 2019 IEEE International Symposium on Mixed and Augmented Reality Adjunct (ISMAR-Adjunct), pages 222–227, 2019. 1 +[40] Yongzhi Su, Mahdi Saleh, Torben Fetzer, Jason Rambach, Nassir Navab, Benjamin Busam, Didier Stricker, and Federico Tombari. Zebrapose: Coarse to fine surface encoding for 6 dof object pose estimation, 2022. 2 +[41] Jiaming Sun, Zihao Wang, Siyu Zhang, Xingyi He, Hongcheng Zhao, Guofeng Zhang, and Xiaowei Zhou. Onepose: One-shot object pose estimation without cad models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6825-6834, 2022. 2 +[42] Martin Sundermeyer, Maximilian Durner, En Yen Puang, Zoltan-Csaba Marton, Narunas Vaskevicius, Kai O. Arras, and Rudolph Triebel. Multi-path learning for object pose estimation across domains, 2019. 2 +[43] Martin Sundermeyer, Zoltan Marton, Maximilian Durner, and Rudolph Triebel. Augmented Autoencoders: Implicit 3D Orientation Learning for 6D Object Detection. International Journal of Computer Vision (IJCV), 128, 2019. 2 +[44] Bugra Tekin, Sudipta Sinha, and Pascal Fua. Real-time seamless single shot 6d object pose prediction. pages 292-301, 2018. 2 +[45] Meng Tian, Marcelo H Ang, and Gim Hee Lee. Shape prior deformation for categorical 6d object pose and size estimation. In European Conference on Computer Vision (ECCV), pages 530-546. Springer, 2020. 2, 5, 6, 7, 8 +[46] Jonathan Tremblay, Thang To, Balakumar Sundaralingam, Yu Xiang, Dieter Fox, and Stan Birchfield. Deep object pose estimation for semantic robotic grasping of household objects, 2018. 2 +[47] Joel Vidal, Chyi-Yeu Lin, and Robert Martí. 6d pose estimation using an improved method based on point pair features, 2018. 2 +[48] Chen Wang, Danfei Xu, Yuke Zhu, Roberto Martin-Martin, Cewu Lu, Li Fei-Fei, and Silvio Savarese. DenseFusion: 6D Object Pose Estimation by Iterative Dense Fusion. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1 +[49] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J Guibas. Normalized object coordinate space for category-level 6d object pose and size estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2642-2651, 2019. 2, 5, 8 +[50] Jiaze Wang, Kai Chen, and Qi Dou. Category-level 6d object pose estimation via cascaded relation and recurrent reconstruction networks. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4807-4814. IEEE, 2021. 8 + +[51] Bowen Wen, Chaitanya Mitash, Baozhang Ren, and Kostas Bekris. se(3)-TrackNet: Data-driven 6D Pose Tracking by Calibrating Image Residuals in Synthetic Domains. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2020. 1, 2 +[52] Bowen Wen, Wenzhao Lian, Kostas Bekris, and Stefan Schaal. You Only Demonstrate Once: Category-Level Manipulation from Single Visual Demonstration. In Proceedings of Robotics: Science and Systems, New York City, NY, USA, 2022. 1 +[53] Heng Yang and Marco Pavone. Object pose estimation with statistical guarantees: Conformal keypoint detection and geometric uncertainty propagation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8947-8958, 2023. 1 +[54] Hongwei Yong, Jianqiang Huang, Xiansheng Hua, and Lei Zhang. Gradient centralization: A new optimization technique for deep neural networks, 2020. 5 +[55] Sergey Zakharov, Ivan Shugurov, and Slobodan Ilic. DPOD: 6D Pose Object Detector and Refiner. In The IEEE International Conference on Computer Vision (ICCV), 2019. 1, 2 +[56] Kun Zhang, Chen Wang, Hua Chen, Jia Pan, Michael Yu Wang, and Wei Zhang. Vision-based six-dimensional peg-in-hole for practical connector insertion. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 1771-1777, 2023. 1 +[57] Michael R. Zhang, James Lucas, Geoffrey Hinton, and Jimmy Ba. Lookahead optimizer: k steps forward, 1 step back, 2019. 5 +[58] Ruida Zhang, Yan Di, Zhiqiang Lou, Fabian Manhardt, Federico Tombari, and Xiangyang Ji. RBP-Pose: Residual bounding box projection for category-level pose estimation, 2022. 1, 2, 5, 8 +[59] Ruida Zhang, Yan Di, Fabian Manhardt, Federico Tombari, and Xiangyang Ji. SSP-Pose: Symmetry-aware shape prior deformation for direct category-level object pose estimation, 2022. 1, 2 +[60] Zhongqun Zhang, Wei Chen, Linfang Zheng, Ales Leonardis, and Hyung Jin Chang. Trans6d: Transformer-based 6d object pose estimation and refinement. In Computer Vision - ECCV 2022 Workshops, pages 112-128, Cham, 2023. Springer Nature Switzerland. 1 +[61] Linfang Zheng, Chen Wang, Yinghan Sun, Esha Dasgupta, Hua Chen, Aleš Leonardis, Wei Zhang, and Hyung Jin Chang. HS-Pose: Hybrid scope feature extraction for category-level object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17163-17173, 2023. 1, 2, 3, 5, 7, 8 +[62] Jun Zhou, Kai Chen, Linlin Xu, Qi Dou, and Jing Qin. Deep fusion transformer network with weighted vector-wise keypoints voting for robust 6d object pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13967-13977, 2023. 2 \ No newline at end of file diff --git a/2024/GeoReF_ Geometric Alignment Across Shape Variation for Category-level Object Pose Refinement/images.zip b/2024/GeoReF_ Geometric Alignment Across Shape Variation for Category-level Object Pose Refinement/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1e6b3b3cb0d058fa1f177e294cd94e2883881b1e --- /dev/null +++ b/2024/GeoReF_ Geometric Alignment Across Shape Variation for Category-level Object Pose Refinement/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f76f59776202cedd7fad2450ec9f900855308cd3cc87e14c79c312172e3a5f86 +size 591222 diff --git a/2024/GeoReF_ Geometric Alignment Across Shape Variation for Category-level Object Pose Refinement/layout.json b/2024/GeoReF_ Geometric Alignment Across Shape Variation for Category-level Object Pose Refinement/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..304e2a8daf079f649f1867eaf46ec9ffee1c0dbe --- /dev/null +++ b/2024/GeoReF_ Geometric Alignment Across Shape Variation for Category-level Object Pose Refinement/layout.json @@ -0,0 +1,9730 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 69, + 102, + 524, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 102, + 524, + 138 + ], + "spans": [ + { + "bbox": [ + 69, + 102, + 524, + 138 + ], + "type": "text", + "content": "GeoReF: Geometric Alignment Across Shape Variation for Category-level Object Pose Refinement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 159, + 534, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 159, + 534, + 190 + ], + "spans": [ + { + "bbox": [ + 58, + 159, + 534, + 190 + ], + "type": "text", + "content": "Linfang Zheng" + }, + { + "bbox": [ + 58, + 159, + 534, + 190 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 58, + 159, + 534, + 190 + ], + "type": "text", + "content": " Tze Ho Elden Tse" + }, + { + "bbox": [ + 58, + 159, + 534, + 190 + ], + "type": "inline_equation", + "content": "^{*3}" + }, + { + "bbox": [ + 58, + 159, + 534, + 190 + ], + "type": "text", + "content": " Chen Wang" + }, + { + "bbox": [ + 58, + 159, + 534, + 190 + ], + "type": "inline_equation", + "content": "^{*1,2}" + }, + { + "bbox": [ + 58, + 159, + 534, + 190 + ], + "type": "text", + "content": " Yinghan Sun" + }, + { + "bbox": [ + 58, + 159, + 534, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 58, + 159, + 534, + 190 + ], + "type": "text", + "content": " Hua Chen" + }, + { + "bbox": [ + 58, + 159, + 534, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 58, + 159, + 534, + 190 + ], + "type": "text", + "content": " \nAles Leonardis" + }, + { + "bbox": [ + 58, + 159, + 534, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 58, + 159, + 534, + 190 + ], + "type": "text", + "content": " Wei Zhang" + }, + { + "bbox": [ + 58, + 159, + 534, + 190 + ], + "type": "inline_equation", + "content": "^{\\dagger 1}" + }, + { + "bbox": [ + 58, + 159, + 534, + 190 + ], + "type": "text", + "content": " Hyung Jin Chang" + }, + { + "bbox": [ + 58, + 159, + 534, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 202, + 539, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 202, + 539, + 258 + ], + "spans": [ + { + "bbox": [ + 54, + 202, + 539, + 258 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 54, + 202, + 539, + 258 + ], + "type": "text", + "content": "Shenzhen Key Laboratory of Control Theory and Intelligent Systems, School of System Design and Intelligent Manufacturing, Southern University of Science and Technology, China " + }, + { + "bbox": [ + 54, + 202, + 539, + 258 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 54, + 202, + 539, + 258 + ], + "type": "text", + "content": "Department of Computer Science, the University of Hong Kong, China " + }, + { + "bbox": [ + 54, + 202, + 539, + 258 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 54, + 202, + 539, + 258 + ], + "type": "text", + "content": "School of Computer Science, University of Birmingham, UK" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 261, + 523, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 261, + 523, + 287 + ], + "spans": [ + { + "bbox": [ + 69, + 261, + 523, + 287 + ], + "type": "text", + "content": "{1xz948, txt994}@student.bham.ac.uk, cwang5@cs.hku.hk, sunyh2021@mail.sustech.edu.cn {chenh6,zhangw3}@sustech.edu.cn,{a.leonadis,h.j.chang}@bham.ac.uk" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 314, + 192, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 314, + 192, + 327 + ], + "spans": [ + { + "bbox": [ + 143, + 314, + 192, + 327 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 339, + 290, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 339, + 290, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 339, + 290, + 568 + ], + "type": "text", + "content": "Object pose refinement is essential for robust object pose estimation. Previous work has made significant progress towards instance-level object pose refinement. Yet, category-level pose refinement is a more challenging problem due to large shape variations within a category and the discrepancies between the target object and the shape prior. To address these challenges, we introduce a novel architecture for category-level object pose refinement. Our approach integrates an HS-layer and learnable affine transformations, which aims to enhance the extraction and alignment of geometric information. Additionally, we introduce a crosscloud transformation mechanism that efficiently merges diverse data sources. Finally, we push the limits of our model by incorporating the shape prior information for translation and size error prediction. We conducted extensive experiments to demonstrate the effectiveness of the proposed framework. Through extensive quantitative experiments, we demonstrate significant improvement over the baseline method by a large margin across all metrics." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 584, + 128, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 584, + 128, + 596 + ], + "spans": [ + { + "bbox": [ + 47, + 584, + 128, + 596 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 604, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 604, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 604, + 287, + 676 + ], + "type": "text", + "content": "Understanding an object's pose is crucial for a wide range of real-world applications, including robotic manipulation [18, 32, 52, 56], augmented reality [30, 33], and autonomous driving [19, 39]. Significant progress has been made for object pose estimation [5, 11, 15, 48, 53, 55] and pose refinement [2, 17, 22, 37, 51, 60] using the object's CAD" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 359, + 314, + 408, + 360 + ], + "blocks": [ + { + "bbox": [ + 318, + 332, + 345, + 342 + ], + "lines": [ + { + "bbox": [ + 318, + 332, + 345, + 342 + ], + "spans": [ + { + "bbox": [ + 318, + 332, + 345, + 342 + ], + "type": "text", + "content": "Camera" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 359, + 314, + 408, + 360 + ], + "lines": [ + { + "bbox": [ + 359, + 314, + 408, + 360 + ], + "spans": [ + { + "bbox": [ + 359, + 314, + 408, + 360 + ], + "type": "image", + "image_path": "02dd435b036b416f9fac1ef5e657fb95dac3b3fcce496b602397162ee4ebbb56.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 421, + 314, + 467, + 358 + ], + "blocks": [ + { + "bbox": [ + 421, + 314, + 467, + 358 + ], + "lines": [ + { + "bbox": [ + 421, + 314, + 467, + 358 + ], + "spans": [ + { + "bbox": [ + 421, + 314, + 467, + 358 + ], + "type": "image", + "image_path": "ab27819a1505b161a313cf61b448f90c0d8171b32f08746fe09be9bfa4341d2f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 482, + 315, + 533, + 358 + ], + "blocks": [ + { + "bbox": [ + 482, + 315, + 533, + 358 + ], + "lines": [ + { + "bbox": [ + 482, + 315, + 533, + 358 + ], + "spans": [ + { + "bbox": [ + 482, + 315, + 533, + 358 + ], + "type": "image", + "image_path": "64d9494ff06a4439fe985d710e0f582c261f07e4f6005a01240c2a0772aa711c.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 365, + 366, + 400, + 414 + ], + "blocks": [ + { + "bbox": [ + 319, + 385, + 335, + 395 + ], + "lines": [ + { + "bbox": [ + 319, + 385, + 335, + 395 + ], + "spans": [ + { + "bbox": [ + 319, + 385, + 335, + 395 + ], + "type": "text", + "content": "Mug" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 365, + 366, + 400, + 414 + ], + "lines": [ + { + "bbox": [ + 365, + 366, + 400, + 414 + ], + "spans": [ + { + "bbox": [ + 365, + 366, + 400, + 414 + ], + "type": "image", + "image_path": "8b40e3ca3f05cb9dc6c6ca2b50af538cac51cef53f44afa8a4133c1dc1e5317f.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 369, + 419, + 387, + 428 + ], + "lines": [ + { + "bbox": [ + 369, + 419, + 387, + 428 + ], + "spans": [ + { + "bbox": [ + 369, + 419, + 387, + 428 + ], + "type": "text", + "content": "SP-m" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 422, + 366, + 467, + 414 + ], + "blocks": [ + { + "bbox": [ + 422, + 366, + 467, + 414 + ], + "lines": [ + { + "bbox": [ + 422, + 366, + 467, + 414 + ], + "spans": [ + { + "bbox": [ + 422, + 366, + 467, + 414 + ], + "type": "image", + "image_path": "e80710786aca18753def72fbe24cac96a512d803c7960d8f1c244b34adc9403f.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 429, + 419, + 445, + 427 + ], + "lines": [ + { + "bbox": [ + 429, + 419, + 445, + 427 + ], + "spans": [ + { + "bbox": [ + 429, + 419, + 445, + 427 + ], + "type": "text", + "content": "SP-1" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 434, + 545, + 463 + ], + "lines": [ + { + "bbox": [ + 306, + 434, + 545, + 463 + ], + "spans": [ + { + "bbox": [ + 306, + 434, + 545, + 463 + ], + "type": "text", + "content": "Figure 1. Illustration of the shape variation. " + }, + { + "bbox": [ + 306, + 434, + 545, + 463 + ], + "type": "inline_equation", + "content": "SP - m" + }, + { + "bbox": [ + 306, + 434, + 545, + 463 + ], + "type": "text", + "content": " represents the category's mean shape, " + }, + { + "bbox": [ + 306, + 434, + 545, + 463 + ], + "type": "inline_equation", + "content": "SP - 1" + }, + { + "bbox": [ + 306, + 434, + 545, + 463 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 434, + 545, + 463 + ], + "type": "inline_equation", + "content": "SP - 2" + }, + { + "bbox": [ + 306, + 434, + 545, + 463 + ], + "type": "text", + "content": " represents the randomly sampled object shapes from the CAMERA25 training set." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 485, + 366, + 530, + 414 + ], + "blocks": [ + { + "bbox": [ + 485, + 366, + 530, + 414 + ], + "lines": [ + { + "bbox": [ + 485, + 366, + 530, + 414 + ], + "spans": [ + { + "bbox": [ + 485, + 366, + 530, + 414 + ], + "type": "image", + "image_path": "2f7f16aba523964b2ffe813e816347b5f1bd6101bf27957cf85a0edede474ba2.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 493, + 419, + 510, + 427 + ], + "lines": [ + { + "bbox": [ + 493, + 419, + 510, + 427 + ], + "spans": [ + { + "bbox": [ + 493, + 419, + 510, + 427 + ], + "type": "text", + "content": "SP-2" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 473, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 545, + 581 + ], + "type": "text", + "content": "model. Despite the promising performance, the reliance on accurate instance-level CAD models limits their generalizability to everyday objects. Category-level methods [7, 23-25, 29, 59, 61] is therefore been proposed to overcome this limitation. The objective of this line of work focuses on estimating object poses within a category given category-level shape priors. As a result, they face unique challenges as there exist diverse shape variations in each object category. We illustrate these shape variations in Fig. 1." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "text", + "content": "Recently, there have been remarkable advancements in category-level object pose estimation [9, 10, 25, 58, 61], primarily due to effective utilization of geometric information through 3D graph convolution [26]. In applications that require high precision, it is common to employ an object pose refinement procedure in conjunction with pose estimation. This involves an initial pose estimation algorithm determining the object pose, followed by a refinement step to further enhance the accuracy of the initially estimated pose by predicting and correcting its error. However, while instance-level object pose refinement has been extensively studied," + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 683, + 195, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 683, + 195, + 693 + ], + "spans": [ + { + "bbox": [ + 58, + 683, + 195, + 693 + ], + "type": "text", + "content": "*Equal contribution, order by dice rolling." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 59, + 693, + 148, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 693, + 148, + 703 + ], + "spans": [ + { + "bbox": [ + 59, + 693, + 148, + 703 + ], + "type": "text", + "content": "The corresponding author." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 59, + 703, + 279, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 279, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 279, + 713 + ], + "type": "text", + "content": "1Project page: https://lynne-zheng-linfang.github.io/georef.github.io" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10693" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": "category-level pose refinement remained unexplored until the introduction of CATRE [29]. By leveraging initial object pose and size estimations, CATRE achieves category-level pose refinement by iteratively aligning the observed target object point cloud with the category-level shape prior. This pipeline is shown to be effective by improving the accuracy of the initial pose and size estimations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 158, + 288, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 158, + 288, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 158, + 288, + 350 + ], + "type": "text", + "content": "While CATRE has proven to be effective in many scenarios, it is limited by the reliance on the PointNet [35] encoder, which is primarily designed for classification and segmentation tasks. This design choice limits its ability to capture essential and fine-grained geometric relationships for accurate pose estimation and refinement. This ability is particularly important in category-level pose refinement as there exists diverse shape variations between inputs. Consequently, CATRE obtains suboptimal performance by direct application of 3D graph convolution. In addition, as CATRE treats the point cloud and the shape prior features separately until a later stage of the network, they potentially miss out on the benefits of integrating these features earlier. Moreover, their approach does not incorporate shape prior information into the translation and scale estimation module, which presents another area for potential improvement." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 352, + 287, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 352, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 46, + 352, + 287, + 543 + ], + "type": "text", + "content": "In this paper, we introduce a novel architecture for category-level object pose refinement which aims to address the limitations mentioned above. To better extract both local and global geometric information, we incorporate an HS layer into our feature extraction process. We apply learnable affine transformations to the features to address the geometric discrepancies between the observed point cloud and the shape prior. This enables the network to align these features more effectively. In addition, we propose a cross-cloud transformation mechanism that is specifically designed to enhance the merging of information between the observed point clouds and the shape prior. This mechanism enables more efficient integration of information between the two sources. Finally, we push the limit of our model by incorporating shape prior information to more accurately predict errors in translation and size estimation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 545, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 713 + ], + "type": "text", + "content": "Our extensive experimental results on two category-level object pose datasets demonstrate that our proposed model to be effective in addressing the problem of shape variations in category-level object pose refinement, and consequently outperforms the state-of-the-art significantly. To the best of our knowledge, our proposed method is the first to successfully address the shape variation issue which is common in category-level pose refinement. Specifically, to enable graph convolution to be effective in capturing geometric relationships between different shapes, we propose an adaptive affine transformation matrix that aligns the observed point clouds and the shape prior. Additionally, the proposed cross-cloud transformation mechanism effectively fuses features from different input point clouds and brings" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 72, + 451, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 451, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 451, + 83 + ], + "type": "text", + "content": "further performance improvements." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 85, + 452, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 85, + 452, + 95 + ], + "spans": [ + { + "bbox": [ + 317, + 85, + 452, + 95 + ], + "type": "text", + "content": "Our contributions are as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 97, + 545, + 251 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 306, + 97, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 97, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 306, + 97, + 545, + 144 + ], + "type": "text", + "content": "- We introduce a novel architecture to specifically address the shape variations issue in category-level object pose refinement. Our proposed method results in consistent performance gain and exhibits better generalization ability." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 145, + 545, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 145, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 306, + 145, + 545, + 180 + ], + "type": "text", + "content": "- We propose a unique cross-cloud transformation mechanism which efficiently merges diverse information from observed point clouds and shape priors." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 181, + 545, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 181, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 306, + 181, + 545, + 251 + ], + "type": "text", + "content": "- We conduct extensive experiments on two category-level object pose datasets to validate our proposed method. On the REAL275 dataset, our method significantly outperforms SPD by " + }, + { + "bbox": [ + 306, + 181, + 545, + 251 + ], + "type": "inline_equation", + "content": "39.1\\%" + }, + { + "bbox": [ + 306, + 181, + 545, + 251 + ], + "type": "text", + "content": " increase in the " + }, + { + "bbox": [ + 306, + 181, + 545, + 251 + ], + "type": "inline_equation", + "content": "5^{\\circ}5\\mathrm{cm}" + }, + { + "bbox": [ + 306, + 181, + 545, + 251 + ], + "type": "text", + "content": " metric. Additionally, we achieve " + }, + { + "bbox": [ + 306, + 181, + 545, + 251 + ], + "type": "inline_equation", + "content": "10.5\\%" + }, + { + "bbox": [ + 306, + 181, + 545, + 251 + ], + "type": "text", + "content": " improvement in the " + }, + { + "bbox": [ + 306, + 181, + 545, + 251 + ], + "type": "inline_equation", + "content": "10^{\\circ}2\\mathrm{cm}" + }, + { + "bbox": [ + 306, + 181, + 545, + 251 + ], + "type": "text", + "content": " metric over the state-of-the-art method, CATRE." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 263, + 392, + 276 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 263, + 392, + 276 + ], + "spans": [ + { + "bbox": [ + 306, + 263, + 392, + 276 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 284, + 545, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 284, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 304, + 284, + 545, + 499 + ], + "type": "text", + "content": "Instance-level object pose estimation and refinement. Instance-level approaches estimate the pose of the target object given known 3D CAD models. They can be briefly divided into correspondence matching methods and template matching methods. Correspondence matching methods [4, 8, 12, 13, 31, 36, 40, 41, 44, 46, 55, 62] matches the outstanding features of the observed object images with its model. Template matching methods [1, 14, 27, 34, 38, 42, 43, 47] compares the images or extracted features with the pre-generated templates. As the initial pose estimates can be noisy to various factors such as occlusions, object pose refinement [20, 22, 51] is shown to be useful in improving the performance of instance-level methods. Even though they achieved impressive over the target object, the reliance on object CAD models limited their generalizability for handling everyday objects. In this paper, we consider a more challenging problem setting where only the category-level shape prior is provided." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "Category-level object pose estimation and refinement. Both tasks mainly focus on addressing the shape variation between the objects. The pioneering work NOCS [49] tackles the shape discrepancy by recovering the normalized visible shape of the target object and achieving the pose by point cloud matching. A series of methods extend this structure by leveraging different information such as domain adaptation [21], different reconstruction space [3], shape prior [3, 16, 21, 45], and structural similarities [6, 25]. However, this line of work is often limited in speed due to the iterative point matching. Another series of work starts with FS-Net [9], which adopts 3D graph convolution (3D-GC) [26] to obtain geometric sensitivity. Due to its effectiveness and real-time performance, graph convolution is widely adopted in recent methods with an enhancement in directions including loss function [10], bounding box voting [58], and shape deformation [59]. HS-Pose [61] ex" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10694" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "tends the geometric feature extraction from local to global, which enhances the capability to handle objects with complex shapes. The research on category-level refinement began recently with the proposal of CATRE [29]. It introduced an effective pipeline that leverages shape priors and a focalization strategy for pose refinement and effectively improves the initial pose estimations. In this paper, we extend the CATRE and tackle the geometric variation issue within the framework of category-level pose refinement." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 190, + 129, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 190, + 129, + 204 + ], + "spans": [ + { + "bbox": [ + 47, + 190, + 129, + 204 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 209, + 171, + 221 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 209, + 171, + 221 + ], + "spans": [ + { + "bbox": [ + 47, + 209, + 171, + 221 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 228, + 287, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 287, + 314 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 287, + 314 + ], + "type": "text", + "content": "In this paper, we tackle the problem of category-level object pose refinement. Given the initial pose and size estimation " + }, + { + "bbox": [ + 46, + 228, + 287, + 314 + ], + "type": "inline_equation", + "content": "(\\mathbf{R}_0,\\mathbf{t}_0,\\mathbf{s}_0)" + }, + { + "bbox": [ + 46, + 228, + 287, + 314 + ], + "type": "text", + "content": ", along with the observed point cloud " + }, + { + "bbox": [ + 46, + 228, + 287, + 314 + ], + "type": "inline_equation", + "content": "\\mathcal{O}\\in \\mathbb{R}^{N^O\\times 3}" + }, + { + "bbox": [ + 46, + 228, + 287, + 314 + ], + "type": "text", + "content": " and the shape prior " + }, + { + "bbox": [ + 46, + 228, + 287, + 314 + ], + "type": "inline_equation", + "content": "\\mathcal{P}\\in \\mathbb{R}^{N^P\\times 3}" + }, + { + "bbox": [ + 46, + 228, + 287, + 314 + ], + "type": "text", + "content": ", we aim to predict the estimation error " + }, + { + "bbox": [ + 46, + 228, + 287, + 314 + ], + "type": "inline_equation", + "content": "(\\Delta \\mathbf{R},\\Delta \\mathbf{t},\\Delta \\mathbf{s})" + }, + { + "bbox": [ + 46, + 228, + 287, + 314 + ], + "type": "text", + "content": " between the initial estimations and the ground truths. The pose refinement algorithm " + }, + { + "bbox": [ + 46, + 228, + 287, + 314 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 228, + 287, + 314 + ], + "type": "text", + "content": " can be described as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 89, + 323, + 287, + 336 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 323, + 287, + 336 + ], + "spans": [ + { + "bbox": [ + 89, + 323, + 287, + 336 + ], + "type": "interline_equation", + "content": "\\left(\\Delta \\mathbf {R}, \\Delta \\mathbf {t}, \\Delta \\mathbf {s}\\right) = \\phi \\left(\\mathbf {R} _ {0}, \\mathbf {t} _ {0}, \\mathbf {s} _ {0}, \\mathcal {O}, \\mathcal {P}\\right). \\tag {1}", + "image_path": "8a6627dd6ac18ba24c679bc565cccbca8a93a669aa3c48c8ad2ae0cc80722adc.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 343, + 287, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 343, + 287, + 368 + ], + "spans": [ + { + "bbox": [ + 46, + 343, + 287, + 368 + ], + "type": "text", + "content": "This pose refinement algorithm " + }, + { + "bbox": [ + 46, + 343, + 287, + 368 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 343, + 287, + 368 + ], + "type": "text", + "content": " can be applied iteratively to improve the refinement performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 374, + 133, + 386 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 374, + 133, + 386 + ], + "spans": [ + { + "bbox": [ + 47, + 374, + 133, + 386 + ], + "type": "text", + "content": "3.2. Preliminaries" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 393, + 287, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 393, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 393, + 287, + 430 + ], + "type": "text", + "content": "Our proposed category-level object pose refinement framework builds upon two previous works, CATRE [29] and HS-layer [61], which we briefly review them in the following." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 441, + 287, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 441, + 287, + 561 + ], + "spans": [ + { + "bbox": [ + 46, + 441, + 287, + 561 + ], + "type": "text", + "content": "CATRE. CATRE is the first framework that considers the problem of category-level pose refinement. It predicts the error between the ground truth and the estimated poses by aligning the input point clouds and the categorical shape priors. Specifically, the network architecture of CATRE consists of four components: a) point cloud focalization, b) shared encoder, c) rotation prediction, and d) translation and size prediction. In point clouds focalization, the observed point clouds " + }, + { + "bbox": [ + 46, + 441, + 287, + 561 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 46, + 441, + 287, + 561 + ], + "type": "text", + "content": " and the shape prior " + }, + { + "bbox": [ + 46, + 441, + 287, + 561 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 441, + 287, + 561 + ], + "type": "text", + "content": " are first aligned with the initial pose and size estimation " + }, + { + "bbox": [ + 46, + 441, + 287, + 561 + ], + "type": "inline_equation", + "content": "[\\mathbf{R}_0, \\mathbf{t}_0, \\mathbf{s}_0]" + }, + { + "bbox": [ + 46, + 441, + 287, + 561 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 89, + 567, + 246, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 567, + 246, + 597 + ], + "spans": [ + { + "bbox": [ + 89, + 567, + 246, + 597 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\hat {\\mathcal {O}} = \\left\\{\\hat {o} _ {i} | \\hat {o} _ {i} = o _ {i} - \\mathbf {t} _ {0}, o _ {i} \\in \\mathcal {O} \\right\\}, \\\\ \\hat {\\mathcal {P}} = \\left\\{\\hat {p} _ {i} | \\hat {p} _ {i} = \\operatorname {d i a g} \\left(\\mathbf {s} _ {0}\\right) \\mathbf {R} _ {0} p _ {i}, p _ {i} \\in \\mathcal {P} \\right\\}, \\\\ \\end{array}", + "image_path": "c11a12241a7128c3e49d876be2244fabaf31607ef37deeccacb8c44d45475a5f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{diag}(\\cdot)" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " converts a vector to a diagonal matrix. The focalized observed point cloud " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{O}}" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " and the focalized shape prior points " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{P}}" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " contain full information required to predict the estimation error " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "(\\Delta \\mathbf{R},\\Delta \\mathbf{t},\\Delta \\mathbf{s})" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": ". First, a PointNet-based shared encoder is used to extract features from the two focalized point clouds independently. Then, both the extracted features are used for " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{R}" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " estimation, while the global feature of the focalized observed point cloud along with the " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_0" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " are used for " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{t}" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{s}" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": ". The initial estimates" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "are updated using the predicted error " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "(\\Delta \\mathbf{R}, \\Delta \\mathbf{t}, \\Delta \\mathbf{s})" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": ". Finally, the updated estimates are used to predict the error again in which this process is iterative and the estimations are refined progressively and continuously." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 125, + 545, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 125, + 545, + 269 + ], + "spans": [ + { + "bbox": [ + 304, + 125, + 545, + 269 + ], + "type": "text", + "content": "HS-layer. The Hybrid-Scope Geometric Feature Extraction Layer (HS-layer) is a simple network structure based on 3D graph convolution. It consists of two parallel paths that extract different scopes of features from the point cloud. The first path encodes the size and translation information of the target object. Meanwhile, the second path extracts outlier-robust local and global geometric features by applying graph convolution with a strategy of Receptive Fields with Feature Distance (RF-F) metric, alongside an Outlier Robust Feature Extraction Layer (ORL). These properties are particularly beneficial for category-level object pose estimation tasks. For more details, please refer to [61]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 275, + 464, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 275, + 464, + 288 + ], + "spans": [ + { + "bbox": [ + 305, + 275, + 464, + 288 + ], + "type": "text", + "content": "3.3. Overall Structure of GeoReF" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 293, + 545, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 545, + 414 + ], + "type": "text", + "content": "The overall framework of our proposed object pose refinement approach is shown in Fig. 2. This framework comprises three principal components: 1) point cloud focalization, 2) feature extraction, and 3) pose error prediction. We follow CATRE and use the same point cloud focalization module. We apply focalization and extract features from both the observed point cloud and the shape prior by our Feature Extraction component. Then, we predict the estimation errors " + }, + { + "bbox": [ + 304, + 293, + 545, + 414 + ], + "type": "inline_equation", + "content": "(\\Delta \\mathbf{R}, \\Delta t, \\Delta s)" + }, + { + "bbox": [ + 304, + 293, + 545, + 414 + ], + "type": "text", + "content": " using the extracted features in the pose error prediction component." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 420, + 545, + 444 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 420, + 545, + 444 + ], + "spans": [ + { + "bbox": [ + 305, + 420, + 545, + 444 + ], + "type": "text", + "content": "3.4. Graph Convolution with Learnable Affine Transformation (LAT)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 450, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 545, + 629 + ], + "type": "text", + "content": "Geometric structural information is effective in estimating an object's pose for category-level object pose estimations. However, as shown in ablation study [AS-1], directly applying the 3D graph convolutions (e.g., HS-Encoder [61] and 3DGCN-Encoder [26]) to category-level object pose refinement tasks results in poor performance. This is due to the differences in task nature between the pose estimation and the pose refinement. In the pose estimation task, the network only needs to extract geometric structural information from a single point cloud. However, in the pose refinement framework, it requires extracting the geometric structural information from the two input point clouds as well as establishing the geometric correspondences between different object shapes. This becomes challenging due to the issue of shape variation in category pose refinement." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "To address the aforementioned problem, we propose to use learnable affine transformations (LATs). By employing LATs, the network can dynamically adjust the input point cloud and the point features which enables better establishment of geometric correspondences between the two different input shapes. Specifically, we apply three LATs (as shown in the bottom left of Fig. 2, where the Matrix Net" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10695" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 69, + 541, + 292 + ], + "blocks": [ + { + "bbox": [ + 53, + 69, + 541, + 292 + ], + "lines": [ + { + "bbox": [ + 53, + 69, + 541, + 292 + ], + "spans": [ + { + "bbox": [ + 53, + 69, + 541, + 292 + ], + "type": "image", + "image_path": "039b1a85f9060f9d5bce9cd1f04401193aeeb56fc69c4c463c05cae237fe2de1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 293, + 547, + 351 + ], + "lines": [ + { + "bbox": [ + 46, + 293, + 547, + 351 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 547, + 351 + ], + "type": "text", + "content": "Figure 2. Overall structure of the proposed method. Our object pose refinement structure contains three main modules. Given the shape prior point cloud, the target object's observed point cloud, and the initial estimation, we first apply point cloud focalization on the input point clouds using the initial estimation. The focalized point clouds then go through a geometric-based feature extraction encoder to obtain geometric structural features. The extracted features are then fed into two branches for rotation error estimation, translation error, and size error estimation. Within the HS Feature Extractor, the Matrix Net models output the learnable affine transformations (LATs) for adaptive point and feature adjustment. The left output of the Matrix Net adjusts the input point clouds, while the right Matrix Net model outputs two affine transformations for adjusting the rotation features, and the translation and size features." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 359, + 289, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 359, + 289, + 443 + ], + "spans": [ + { + "bbox": [ + 46, + 359, + 289, + 443 + ], + "type": "text", + "content": "outputs the learnable affine transformations): The first LAT " + }, + { + "bbox": [ + 46, + 359, + 289, + 443 + ], + "type": "inline_equation", + "content": "M \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 359, + 289, + 443 + ], + "type": "text", + "content": " is applied to the input point cloud in the Euclidean space. The second LAT " + }, + { + "bbox": [ + 46, + 359, + 289, + 443 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{ts}}" + }, + { + "bbox": [ + 46, + 359, + 289, + 443 + ], + "type": "text", + "content": " is applied to the extracted translation and size features " + }, + { + "bbox": [ + 46, + 359, + 289, + 443 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{ts}}" + }, + { + "bbox": [ + 46, + 359, + 289, + 443 + ], + "type": "text", + "content": ". The third LAT " + }, + { + "bbox": [ + 46, + 359, + 289, + 443 + ], + "type": "inline_equation", + "content": "M_r" + }, + { + "bbox": [ + 46, + 359, + 289, + 443 + ], + "type": "text", + "content": " is applied to the extracted rotation feature " + }, + { + "bbox": [ + 46, + 359, + 289, + 443 + ], + "type": "inline_equation", + "content": "f_r" + }, + { + "bbox": [ + 46, + 359, + 289, + 443 + ], + "type": "text", + "content": ". With this approach, our method can better utilize the valuable geometric features in pose refinement." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 448, + 287, + 473 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 448, + 287, + 473 + ], + "spans": [ + { + "bbox": [ + 47, + 448, + 287, + 473 + ], + "type": "text", + "content": "3.5. Cross-Cloud Transformation (CCT) for Information Mixing" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 479, + 287, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 287, + 634 + ], + "type": "text", + "content": "In pose refinement, effectively blending information from the focalized observed object and the shape prior is crucial for enabling the network to align them accurately. However, in CATRE, the data from the observed point cloud and the shape prior are processed independently until the late rotation prediction stage, where they are merely concatenated, limiting the effectiveness of the alignment. To address this problem, we introduce a novel cross-cloud transformation mechanism that effectively mixes the geometric information from the shape of prior features into the features of the observed point cloud. In particular, we use the feature transformation matrices " + }, + { + "bbox": [ + 46, + 479, + 287, + 634 + ], + "type": "inline_equation", + "content": "M_r^P" + }, + { + "bbox": [ + 46, + 479, + 287, + 634 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 479, + 287, + 634 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{ts}}^P" + }, + { + "bbox": [ + 46, + 479, + 287, + 634 + ], + "type": "text", + "content": " of shape prior to transforming the features of the observed point cloud:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 135, + 636, + 287, + 651 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 636, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 135, + 636, + 287, + 651 + ], + "type": "interline_equation", + "content": "f _ {r} ^ {O} = M _ {r} ^ {P} f _ {r} ^ {O}, \\tag {3}", + "image_path": "ba4d3cbe61c1e00e7f36b0438d95314dd36d82e31300d7bc7f207e1365a864eb.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 652, + 287, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 652, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 137, + 652, + 287, + 667 + ], + "type": "interline_equation", + "content": "f _ {\\mathrm {t s}} ^ {O} = M _ {\\mathrm {t s}} ^ {P} f _ {\\mathrm {t s}} ^ {O}. \\tag {4}", + "image_path": "ebd3ab0dfb942307348b05d988536218d0bff80dfc74ea4f9bc31bbf48eb16b5.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 671, + 272, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 272, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 272, + 685 + ], + "type": "text", + "content": "3.6. Integrating Shape Prior in Pose Estimation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "The information contained in the shape prior is crucial for the network to align the observed point cloud and the shape" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 359, + 545, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 359, + 545, + 528 + ], + "spans": [ + { + "bbox": [ + 304, + 359, + 545, + 528 + ], + "type": "text", + "content": "prior. For the rotation error prediction, the information contained in the shape prior is the essential information. For the translation and size prediction, this information can also be utilized by the network to adjust the learned geometric features accordingly. Therefore, unlike CATRE, which relied solely on features extracted from the observed point cloud to predict " + }, + { + "bbox": [ + 304, + 359, + 545, + 528 + ], + "type": "inline_equation", + "content": "(\\Delta \\mathbf{t}, \\Delta \\mathbf{s})" + }, + { + "bbox": [ + 304, + 359, + 545, + 528 + ], + "type": "text", + "content": ", our approach also incorporates the information from shape prior to predict them. In particular, we not only mix the information using previous CCT mechanism, but also concatenate the features from shape prior and observed point cloud to obtain mixed features in the similarly way as the rotation estimation. We utilize " + }, + { + "bbox": [ + 304, + 359, + 545, + 528 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{t}}^{\\mathrm{OP}}" + }, + { + "bbox": [ + 304, + 359, + 545, + 528 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 359, + 545, + 528 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{s}}^{\\mathrm{OP}}" + }, + { + "bbox": [ + 304, + 359, + 545, + 528 + ], + "type": "text", + "content": " which contain both information from shape prior and observed point cloud like " + }, + { + "bbox": [ + 304, + 359, + 545, + 528 + ], + "type": "inline_equation", + "content": "f_{r}^{\\mathrm{OP}}" + }, + { + "bbox": [ + 304, + 359, + 545, + 528 + ], + "type": "text", + "content": " to predict " + }, + { + "bbox": [ + 304, + 359, + 545, + 528 + ], + "type": "inline_equation", + "content": "(\\Delta \\mathbf{t}, \\delta)" + }, + { + "bbox": [ + 304, + 359, + 545, + 528 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": "We use two pose error predictors of the same network architecture to predict the rotation error and the translation and size error, respectively. Note that the weights of these two pose error predictors are not shared. The network structure of the pose error predictor is shown in Fig. 2. The pose predictor takes in two features, passes them through two same paths separately, and obtains two vectors in " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": ". In the translation and size branch, the pose error predictor takes in " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{t}}^{\\mathrm{OP}}" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{s}}^{\\mathrm{OP}}" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": " and passes them through the two paths, and the output two vectors are regarded as " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{t}" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{s}" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": ", respectively. For the rotation error prediction, the mixed rotation features " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "f_{r}^{\\mathrm{OP}}" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": " are copied and passed through the two paths in the pose error predictor. The two output vectors are regarded as " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "r_x" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "r_y" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "r_x" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "r_y" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": " are the first and second axes of the rotation error matrix " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{R}" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": ". The third column " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "r_z" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{R}" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": " can" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10696" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 102, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 102, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 102, + 83 + ], + "type": "text", + "content": "be found by:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 137, + 85, + 287, + 97 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 85, + 287, + 97 + ], + "spans": [ + { + "bbox": [ + 137, + 85, + 287, + 97 + ], + "type": "interline_equation", + "content": "r _ {z} = r _ {x} \\times r _ {y}. \\tag {5}", + "image_path": "47f3ac47ad47014344653d68831f9be3f952df1ec1e377d04ad59859fbf6e6f0.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 106, + 128, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 106, + 128, + 119 + ], + "spans": [ + { + "bbox": [ + 47, + 106, + 128, + 119 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 125, + 289, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 125, + 289, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 125, + 289, + 258 + ], + "type": "text", + "content": "Implementation details. We implement and experiment with our method using an RTX 4090 GPU with a batch size of 12 and 150 training epochs. We follow CATRE [29] and adopt its loss functions and the basic data augmentation strategies including random dropping points, adding Gaussian noise, random pose perturbations, etc. We set the number of points for both the observed points and shape prior to be 512. We train the network using Ranger optimizer [28, 54, 57] with a base learning rate of " + }, + { + "bbox": [ + 46, + 125, + 289, + 258 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 46, + 125, + 289, + 258 + ], + "type": "text", + "content": " and anneal the learning rate from the " + }, + { + "bbox": [ + 46, + 125, + 289, + 258 + ], + "type": "inline_equation", + "content": "72\\%" + }, + { + "bbox": [ + 46, + 125, + 289, + 258 + ], + "type": "text", + "content": " of the total epoch based on cosine schedule." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 269, + 289, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 269, + 289, + 461 + ], + "spans": [ + { + "bbox": [ + 46, + 269, + 289, + 461 + ], + "type": "text", + "content": "Baselines. We use CATRE as the baseline for our ablation study as it is the state-of-the-art category-level object refinement method. As CATRE did not provide the results of " + }, + { + "bbox": [ + 46, + 269, + 289, + 461 + ], + "type": "inline_equation", + "content": "\\mathrm{IoU}_{50}" + }, + { + "bbox": [ + 46, + 269, + 289, + 461 + ], + "type": "text", + "content": ", we obtained them by using their official pre-trained model and kept the rest of the reported metric scores consistent with the corresponding paper. For fair comparisons, we use the same initial estimations as CATRE, which is the pose estimation results of SPD [45]. The result of replacing PointNet with the 3DGC-Encoder in the ablation study is provided by CATRE. We also apply our method to other state-of-the-art category-level object pose estimation approaches [10, 25, 61] to demonstrate the effectiveness of our proposed refinement method. For the pose refinement on HS-Pose [61] and RBP-Pose [58], we compute the initial estimations using their official pre-trained models. The results of other methods are taken directly from their paper." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 472, + 288, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 288, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 288, + 616 + ], + "type": "text", + "content": "Datasets. As we focus on the problem of shape variation between input object point clouds, we choose two popular category-level object pose estimation benchmarks to verify our approach, i.e., REAL275 [49] and CAMERA25 [49]. They both contain 6 object categories with multiple levels of shape complexities, i.e., bowl, can, bottle, laptop, mug, and camera. REAL275 contains 36 objects in 13 real-world scenes with 7k RGB-D images in total. Among them, 16 objects in 7 scenes are used for training, resulting in 4.3k images in training. CAMERA25 is a large synthetic RGB-D dataset. It provides 1085 objects and 275k RGB-D images for training, and 184 objects and 25k images for testing." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 626, + 288, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 626, + 288, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 626, + 288, + 688 + ], + "type": "text", + "content": "Evaluation metrics. Following [29, 61], we evaluate our method using: 1) The mean average precision (mAP) of the 3D Intersection over Union " + }, + { + "bbox": [ + 46, + 626, + 288, + 688 + ], + "type": "inline_equation", + "content": "(IoU)" + }, + { + "bbox": [ + 46, + 626, + 288, + 688 + ], + "type": "text", + "content": " at different thresholds " + }, + { + "bbox": [ + 46, + 626, + 288, + 688 + ], + "type": "inline_equation", + "content": "(50\\%" + }, + { + "bbox": [ + 46, + 626, + 288, + 688 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 626, + 288, + 688 + ], + "type": "inline_equation", + "content": "75\\%)" + }, + { + "bbox": [ + 46, + 626, + 288, + 688 + ], + "type": "text", + "content": " to evaluate the pose and size estimation together. 2) The pose metric at " + }, + { + "bbox": [ + 46, + 626, + 288, + 688 + ], + "type": "inline_equation", + "content": "n^{\\circ}mcm" + }, + { + "bbox": [ + 46, + 626, + 288, + 688 + ], + "type": "text", + "content": " defines a pose as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 108 + ], + "type": "text", + "content": "correct if the rotation error is below " + }, + { + "bbox": [ + 306, + 72, + 545, + 108 + ], + "type": "inline_equation", + "content": "n^{\\circ}" + }, + { + "bbox": [ + 306, + 72, + 545, + 108 + ], + "type": "text", + "content": " and the translation error is below " + }, + { + "bbox": [ + 306, + 72, + 545, + 108 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 306, + 72, + 545, + 108 + ], + "type": "text", + "content": " cm. Here, we use " + }, + { + "bbox": [ + 306, + 72, + 545, + 108 + ], + "type": "inline_equation", + "content": "5^{\\circ}, 10^{\\circ}, 2\\mathrm{cm}" + }, + { + "bbox": [ + 306, + 72, + 545, + 108 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 306, + 72, + 545, + 108 + ], + "type": "inline_equation", + "content": "5\\mathrm{cm}" + }, + { + "bbox": [ + 306, + 72, + 545, + 108 + ], + "type": "text", + "content": " as the thresholds." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 116, + 398, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 116, + 398, + 129 + ], + "spans": [ + { + "bbox": [ + 306, + 116, + 398, + 129 + ], + "type": "text", + "content": "4.1. Ablation study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 134, + 547, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 134, + 547, + 219 + ], + "spans": [ + { + "bbox": [ + 305, + 134, + 547, + 219 + ], + "type": "text", + "content": "To verify the proposed architecture, we conducted comprehensive ablation studies on the REAL275 dataset using the initial pose estimations from SPD [45]. We present a quantitative comparison of our method with various key components disabled to motivate our design choices in Table 1. Full results of the ablation study are reported in the supplementary materials." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 233, + 546, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 233, + 546, + 449 + ], + "spans": [ + { + "bbox": [ + 305, + 233, + 546, + 449 + ], + "type": "text", + "content": "[AS-1] Using geometric features directly. To illustrate the limitations of existing geometric-based encoder under the problem of shape variations, we replace the encoder of CATRE with two robust geometric-based point cloud convolutional structures, namely 3DGCN-Encoder [26] and HS-Encoder [61]. 3DGCN is a widely adopted graph convolution in existing category-level object pose estimation algorithms, while HS-Encoder is a recent architecture that achieves state-of-the-art performance in category-level object pose estimation. However, as shown in Table 1 [C0, C1], even though both HS-Encoder and 3DGCN-Encoder are powerful in finding an object's pose from individual inputs, they failed to manage the pose refinement scenarios when there exist shape variations between the target object and the shape prior. We observed both encoders result in a performance drop when compared to the original CATRE with IoU75 of 37.3% (HS-Encoder) vs. 43.7% (CATRE), 5°5cm of 43.4% (3DGCN-Encoder) vs. 53.3% (CATRE)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 464, + 545, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 464, + 545, + 572 + ], + "spans": [ + { + "bbox": [ + 305, + 464, + 545, + 572 + ], + "type": "text", + "content": "[AS-2] Use prior features in translation and scale estimation. To validate that the information of shape prior is also important in scale and translation error prediction, we add the features of the prior shape to the scale and translation branch by using the same network architecture as the rotation branch. As shown in Table 1 [D0], incorporating shape prior information in translation and size estimation enhanced the overall performance by " + }, + { + "bbox": [ + 305, + 464, + 545, + 572 + ], + "type": "inline_equation", + "content": "2.2\\%" + }, + { + "bbox": [ + 305, + 464, + 545, + 572 + ], + "type": "text", + "content": " improvement on " + }, + { + "bbox": [ + 305, + 464, + 545, + 572 + ], + "type": "inline_equation", + "content": "5^{\\circ}2\\mathrm{cm}" + }, + { + "bbox": [ + 305, + 464, + 545, + 572 + ], + "type": "text", + "content": " metric and " + }, + { + "bbox": [ + 305, + 464, + 545, + 572 + ], + "type": "inline_equation", + "content": "2.4\\%" + }, + { + "bbox": [ + 305, + 464, + 545, + 572 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 305, + 464, + 545, + 572 + ], + "type": "inline_equation", + "content": "10^{\\circ}2\\mathrm{cm}" + }, + { + "bbox": [ + 305, + 464, + 545, + 572 + ], + "type": "text", + "content": " metric." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 586, + 547, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 586, + 547, + 696 + ], + "spans": [ + { + "bbox": [ + 305, + 586, + 547, + 696 + ], + "type": "text", + "content": "[AS-3] Use learnable affine transformation (LAT) for geometric features. To demonstrate the effectiveness of LAT in addressing the shape variation issue, we conduct ablation studies on applying the proposed LAT to the input point cloud and the extracted geometric features in the feature space. The result is shown in Table 1 [E0]. Compared to using geometric features directly (Table 1 [C0]), LATs bring a significant boost on all the metrics, with " + }, + { + "bbox": [ + 305, + 586, + 547, + 696 + ], + "type": "inline_equation", + "content": "\\mathrm{IoU}_{75}" + }, + { + "bbox": [ + 305, + 586, + 547, + 696 + ], + "type": "text", + "content": " improved by " + }, + { + "bbox": [ + 305, + 586, + 547, + 696 + ], + "type": "inline_equation", + "content": "20.9\\%" + }, + { + "bbox": [ + 305, + 586, + 547, + 696 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 586, + 547, + 696 + ], + "type": "inline_equation", + "content": "5^{\\circ}2\\mathrm{cm}" + }, + { + "bbox": [ + 305, + 586, + 547, + 696 + ], + "type": "text", + "content": " improved by " + }, + { + "bbox": [ + 305, + 586, + 547, + 696 + ], + "type": "inline_equation", + "content": "10.5\\%" + }, + { + "bbox": [ + 305, + 586, + 547, + 696 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 288, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 288, + 713 + ], + "type": "text", + "content": "2Note that there was a small mistake with the IoU computation from the original benchmark evaluation code [49], we follow [29] to recalculate" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 702, + 435, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 702, + 435, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 702, + 435, + 712 + ], + "type": "text", + "content": "the IoU metrics for the SOTA methods." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10697" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 102, + 545, + 251 + ], + "blocks": [ + { + "bbox": [ + 227, + 71, + 365, + 80 + ], + "lines": [ + { + "bbox": [ + 227, + 71, + 365, + 80 + ], + "spans": [ + { + "bbox": [ + 227, + 71, + 365, + 80 + ], + "type": "text", + "content": "Table 1. Ablation studies on REAL275." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 81, + 545, + 100 + ], + "lines": [ + { + "bbox": [ + 47, + 81, + 545, + 100 + ], + "spans": [ + { + "bbox": [ + 47, + 81, + 545, + 100 + ], + "type": "text", + "content": "Higher score indicates better performance. In the 'Row' column, the code in bold means the strategies taken in the final structure. In the 'Method' column, the notation " + }, + { + "bbox": [ + 47, + 81, + 545, + 100 + ], + "type": "inline_equation", + "content": "X:Y" + }, + { + "bbox": [ + 47, + 81, + 545, + 100 + ], + "type": "text", + "content": " denotes module " + }, + { + "bbox": [ + 47, + 81, + 545, + 100 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 47, + 81, + 545, + 100 + ], + "type": "text", + "content": " from structure " + }, + { + "bbox": [ + 47, + 81, + 545, + 100 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 47, + 81, + 545, + 100 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 81, + 545, + 100 + ], + "type": "inline_equation", + "content": "X + Y" + }, + { + "bbox": [ + 47, + 81, + 545, + 100 + ], + "type": "text", + "content": " means add module " + }, + { + "bbox": [ + 47, + 81, + 545, + 100 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 47, + 81, + 545, + 100 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 47, + 81, + 545, + 100 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 47, + 81, + 545, + 100 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 81, + 545, + 100 + ], + "type": "inline_equation", + "content": "X \\rightarrow Y" + }, + { + "bbox": [ + 47, + 81, + 545, + 100 + ], + "type": "text", + "content": " indicates replacing " + }, + { + "bbox": [ + 47, + 81, + 545, + 100 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 47, + 81, + 545, + 100 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 47, + 81, + 545, + 100 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 47, + 81, + 545, + 100 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 102, + 545, + 251 + ], + "lines": [ + { + "bbox": [ + 47, + 102, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 47, + 102, + 545, + 251 + ], + "type": "table", + "html": "
RowMethodIoU50IoU755°2cm5°5cm10°2cm10°5cm2cm5°
A0CATRE[45] (baseline)77.043.645.854.461.473.175.158.0
B0Ours: E0 + Cross-Cloud Transformation79.2 2.2↑51.8 8.2↑54.4 8.6↑60.3 5.9↑71.9 10.5↑79.4 6.3↑81.9 6.8↑64.3 6.3↑
C0A0: PointNet → HS-Encoder71.030.141.945.960.670.371.948.7
C1A0: PointNet → 3DGCN-Encoder-28.436.043.4--68.047.7
D0A0 + prior in ST branch77.145.848.054.663.872.577.959.2
E0D0: PointNet → HS-layer+LAT79.451.052.458.669.477.780.462.4
E1B0: No LAT on input points76.139.346.653.065.474.878.058.2
E2B0: No LAT on features78.548.847.453.067.475.080.457.4
E3B0: No LAT on the rotation feature79.850.650.456.268.676.380.260.8
F0E0+ Global Concatenation Fusion77.748.447.854.567.175.280.159.4
", + "image_path": "8eeadb48bc79d2c9fa9976b80d83d882378ae48e1dc534739c189cb99b2a2b7f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 266, + 288, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 288, + 340 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 288, + 340 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 46, + 266, + 288, + 340 + ], + "type": "inline_equation", + "content": "5^{\\circ}5\\mathrm{cm}" + }, + { + "bbox": [ + 46, + 266, + 288, + 340 + ], + "type": "text", + "content": " improved by " + }, + { + "bbox": [ + 46, + 266, + 288, + 340 + ], + "type": "inline_equation", + "content": "12.7\\%" + }, + { + "bbox": [ + 46, + 266, + 288, + 340 + ], + "type": "text", + "content": ". The resulted network also significantly outperforms the PointNet-based encoder (Table 1[D0]) on all the metrics with " + }, + { + "bbox": [ + 46, + 266, + 288, + 340 + ], + "type": "inline_equation", + "content": "\\mathrm{IoU}_{75}" + }, + { + "bbox": [ + 46, + 266, + 288, + 340 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 266, + 288, + 340 + ], + "type": "inline_equation", + "content": "10^{\\circ}5\\mathrm{cm}" + }, + { + "bbox": [ + 46, + 266, + 288, + 340 + ], + "type": "text", + "content": " improved by " + }, + { + "bbox": [ + 46, + 266, + 288, + 340 + ], + "type": "inline_equation", + "content": "5.2\\%" + }, + { + "bbox": [ + 46, + 266, + 288, + 340 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 266, + 288, + 340 + ], + "type": "inline_equation", + "content": "5^{\\circ}2\\mathrm{cm}" + }, + { + "bbox": [ + 46, + 266, + 288, + 340 + ], + "type": "text", + "content": " improved by " + }, + { + "bbox": [ + 46, + 266, + 288, + 340 + ], + "type": "inline_equation", + "content": "6.4\\%" + }, + { + "bbox": [ + 46, + 266, + 288, + 340 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 266, + 288, + 340 + ], + "type": "inline_equation", + "content": "10^{\\circ}2\\mathrm{cm}" + }, + { + "bbox": [ + 46, + 266, + 288, + 340 + ], + "type": "text", + "content": " improved by " + }, + { + "bbox": [ + 46, + 266, + 288, + 340 + ], + "type": "inline_equation", + "content": "5.6\\%" + }, + { + "bbox": [ + 46, + 266, + 288, + 340 + ], + "type": "text", + "content": ". These results verified the effectiveness of LAT on geometric features." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 353, + 287, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 353, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 353, + 287, + 509 + ], + "type": "text", + "content": "[AS-4] The influence of each learnable affine transformation (LAT). To further demonstrate the influences of each LAT, we conducted three experiments by gradually disabling LAT from the framework: 1) without the LAT on the input point cloud, 2) without applying LATs on features, and 3) without independent LATs on the rotation features, where a single LAT is used for the rotation, translation, and scale features. The results are shown in Table 1 [E1-E3]. Compared to the directly using geometric features (Table 1 [C0]), we show that LAT can significantly enhance the network by around " + }, + { + "bbox": [ + 46, + 353, + 287, + 509 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 353, + 287, + 509 + ], + "type": "text", + "content": " improvements on " + }, + { + "bbox": [ + 46, + 353, + 287, + 509 + ], + "type": "inline_equation", + "content": "\\mathrm{IoU}_{75}" + }, + { + "bbox": [ + 46, + 353, + 287, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 353, + 287, + 509 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 46, + 353, + 287, + 509 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 46, + 353, + 287, + 509 + ], + "type": "inline_equation", + "content": "5^{\\circ}5\\mathrm{cm}" + }, + { + "bbox": [ + 46, + 353, + 287, + 509 + ], + "type": "text", + "content": " metric in Table 1 [E1-E3]. We verify that the combination of them consistently results in better performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 521, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 288, + 715 + ], + "type": "text", + "content": "[AS-5] Cross-cloud transformation (CCT) based feature fusion To demonstrate that it is important to have a good feature fusion strategy in the feature extraction phase, we conducted experiments on two different feature fusion strategies. One of them is the widely adopted feature fusion strategy, where the global feature of one point cloud is concatenated with the features of another point cloud and then goes through convolutional layers for feature fusion. Another one is the proposed CCT-based feature fusion. As shown in Table 1 [F0], applying global concatenation-based fusion does not enhance the overall performance, and even results in worse performance in all the metrics with the " + }, + { + "bbox": [ + 46, + 521, + 288, + 715 + ], + "type": "inline_equation", + "content": "5^{\\circ}2\\mathrm{cm}" + }, + { + "bbox": [ + 46, + 521, + 288, + 715 + ], + "type": "text", + "content": " significantly decreased by " + }, + { + "bbox": [ + 46, + 521, + 288, + 715 + ], + "type": "inline_equation", + "content": "4.6\\%" + }, + { + "bbox": [ + 46, + 521, + 288, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 521, + 288, + 715 + ], + "type": "inline_equation", + "content": "5^{\\circ}5\\mathrm{cm}" + }, + { + "bbox": [ + 46, + 521, + 288, + 715 + ], + "type": "text", + "content": " decreased by " + }, + { + "bbox": [ + 46, + 521, + 288, + 715 + ], + "type": "inline_equation", + "content": "3.1\\%" + }, + { + "bbox": [ + 46, + 521, + 288, + 715 + ], + "type": "text", + "content": " (compared to Table 1 [D0]). On the contrary, as shown in Table 1 [B0], our simple CCT-based feature fusion shows its effectiveness by improving all the pose" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 307, + 266, + 545, + 416 + ], + "blocks": [ + { + "bbox": [ + 307, + 266, + 545, + 416 + ], + "lines": [ + { + "bbox": [ + 307, + 266, + 545, + 416 + ], + "spans": [ + { + "bbox": [ + 307, + 266, + 545, + 416 + ], + "type": "image", + "image_path": "80e8d47f5b934956f66a7f331dad000b3ff466c9344aa828b320c4dcf51b0fa5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 418, + 545, + 449 + ], + "lines": [ + { + "bbox": [ + 305, + 418, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 305, + 418, + 545, + 449 + ], + "type": "text", + "content": "Figure 3. Performance comparison of our method and CATRE under different shape priors. " + }, + { + "bbox": [ + 305, + 418, + 545, + 449 + ], + "type": "inline_equation", + "content": "SP - m" + }, + { + "bbox": [ + 305, + 418, + 545, + 449 + ], + "type": "text", + "content": " denotes the category's mean shape, while " + }, + { + "bbox": [ + 305, + 418, + 545, + 449 + ], + "type": "inline_equation", + "content": "SP - 1" + }, + { + "bbox": [ + 305, + 418, + 545, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 418, + 545, + 449 + ], + "type": "inline_equation", + "content": "SP - 2" + }, + { + "bbox": [ + 305, + 418, + 545, + 449 + ], + "type": "text", + "content": " are randomly sampled object shapes from CAMERA25." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 456, + 545, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 456, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 305, + 456, + 545, + 481 + ], + "type": "text", + "content": "metrics by around " + }, + { + "bbox": [ + 305, + 456, + 545, + 481 + ], + "type": "inline_equation", + "content": "2.0\\%" + }, + { + "bbox": [ + 305, + 456, + 545, + 481 + ], + "type": "text", + "content": ". In Supplementary, we also visualize the influence of CCT in feature space." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 495, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 545, + 651 + ], + "type": "text", + "content": "[AS-6] Handle shape variations. To demonstrate that the proposed method can handle the shape variations, we replaced the original shape prior with two randomly sampled models from CAMERA25 training set and trained on REAL275. Note that the original shape prior represents the mean shape of a category, and the new models are randomly sampled. Therefore, there are larger shape variations with certain target objects. As shown in Fig. 3, CATRE performs best when using the mean shape of the category, while its performance drops dramatically on the randomly sampled shape priors. In contrast, our method exhibits robustness to shape variations introduced by different shape priors and consistently delivers strong performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "[AS-7] Refinement with different initial estimations. To demonstrate our model robustness to different initial estimations, we compare the proposed method with CATRE on different initial estimations generated by category-level" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10698" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 70, + 547, + 331 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 547, + 331 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 547, + 331 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 547, + 331 + ], + "type": "image", + "image_path": "1996b314fa576aac0fbcadfd6781397fddb6746998eb12cf02891ac3b40a9684.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 334, + 546, + 354 + ], + "lines": [ + { + "bbox": [ + 46, + 334, + 546, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 334, + 546, + 354 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison of proposed (row #3) and baseline (row #2) methods using SPD (row #1) as initial estimation. Ground truth shown with white lines. Note that the estimated rotations of symmetric objects (e.g. bowl, bottle, and can) are considered correct if the symmetry axis is aligned." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 365, + 289, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 289, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 289, + 545 + ], + "type": "text", + "content": "object pose estimation methods with varying performance, including HS-Pose [61], GPVPose [10], and the SelfDPDN [25]. As shown in Table 2, our method consistently improves the initial estimations by a large margin across different metrics. For example, we improved the SelfDPND and GPV-Pose on strict " + }, + { + "bbox": [ + 46, + 365, + 289, + 545 + ], + "type": "inline_equation", + "content": "5^{\\circ}2\\mathrm{cm}" + }, + { + "bbox": [ + 46, + 365, + 289, + 545 + ], + "type": "text", + "content": " metric by " + }, + { + "bbox": [ + 46, + 365, + 289, + 545 + ], + "type": "inline_equation", + "content": "9.6\\%" + }, + { + "bbox": [ + 46, + 365, + 289, + 545 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 365, + 289, + 545 + ], + "type": "inline_equation", + "content": "15.4\\%" + }, + { + "bbox": [ + 46, + 365, + 289, + 545 + ], + "type": "text", + "content": ", respectively. We boosted the " + }, + { + "bbox": [ + 46, + 365, + 289, + 545 + ], + "type": "inline_equation", + "content": "\\mathrm{IoU}_{75}" + }, + { + "bbox": [ + 46, + 365, + 289, + 545 + ], + "type": "text", + "content": " of the GPV-Pose and the HS-Pose by " + }, + { + "bbox": [ + 46, + 365, + 289, + 545 + ], + "type": "inline_equation", + "content": "26.5\\%" + }, + { + "bbox": [ + 46, + 365, + 289, + 545 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 365, + 289, + 545 + ], + "type": "inline_equation", + "content": "15.2\\%" + }, + { + "bbox": [ + 46, + 365, + 289, + 545 + ], + "type": "text", + "content": ", respectively. However, the baseline CATRE fails to refine the initial poses of SelfDPDN. Also, CATRE reaches its limit when refining high-accuracy initial estimations such as the initial poses generated by HS-Pose, resulting in performance drops on " + }, + { + "bbox": [ + 46, + 365, + 289, + 545 + ], + "type": "inline_equation", + "content": "10^{\\circ}5\\mathrm{cm}" + }, + { + "bbox": [ + 46, + 365, + 289, + 545 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 365, + 289, + 545 + ], + "type": "inline_equation", + "content": "10^{\\circ}2\\mathrm{cm}" + }, + { + "bbox": [ + 46, + 365, + 289, + 545 + ], + "type": "text", + "content": ". The comparison results demonstrated the robustness and capability of our method to different initial estimations. We report full details in the supplementary." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 552, + 288, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 288, + 564 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 288, + 564 + ], + "type": "text", + "content": "4.2. Generalizably test on the CAMERA25 dataset" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": "In real-world applications, category-level algorithms often need to generalize across diverse testing scenarios, encountering a larger number of objects than represented in their training sets. To simulate this problem setting, we choose CAMERA25 as it provides more than 25K RGB-D testing images. We train our model using only a mini set (2% and 4%) of the CAMERA25 training data, resulting in a training set of around 5K and 10K images from a total of 275K images. As shown in Table 3, the performance of the baseline method decreased dramatically when using 2% of the training images, with IoU75 dropped by 12.9% and 5°2cm dropped by 9.0% (see [B0]). On the contrary, our" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 307, + 404, + 545, + 506 + ], + "blocks": [ + { + "bbox": [ + 305, + 363, + 545, + 403 + ], + "lines": [ + { + "bbox": [ + 305, + 363, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 305, + 363, + 545, + 403 + ], + "type": "text", + "content": "Table 2. Comparison of the pose refinement with the baseline method CATRE on REAL275 with different initial estimations. Each comparison group contains 3 methods: the initial pose estimation method, refinement using CATRE, and refinement using our method, respectively." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 404, + 545, + 506 + ], + "lines": [ + { + "bbox": [ + 307, + 404, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 307, + 404, + 545, + 506 + ], + "type": "table", + "html": "
MethodIoU755°2cm5°5cm10°2cm10°5cm
Self-DPDN [25]42.244.350.965.178.6
Self-DPDN + CATRE0.422↓0.344.0↓5.145.8↓0.464.7↓6.971.7↓
Self-DPDN + Ours49.77.5↑53.99.6↑60.19.2↑75.09.9↑82.84.2↑
GPV-Pose [10]23.1*32.042.955.073.3
GPV-Pose + CATRE42.619.5↑39.77.7↑54.111.2↑57.12.1↑78.04.7↑
GPV-Pose + Ours49.626.5↑47.415.4↑57.814.9↑68.113.1↑81.27.9↑
HS-Pose [61]39.1*46.555.268.682.7
HS-Pose + CATRE47.18.0↑48.72.2↑59.13.9↑67.80.8↓81.21.5↓
HS-Pose + Ours54.315.2↑51.75.2↑59.64.4↑74.35.7↑83.81.1↑
", + "image_path": "92a30f3ad1d34b80c178baea0621fd847890f69ee8310dcab0cdc4f4213b70f8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 517, + 545, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 517, + 545, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 517, + 545, + 615 + ], + "type": "text", + "content": "method exhibits a much higher performance when using small datasets for training. Specifically, our method can already outperforms the fully-trained CATRE with only " + }, + { + "bbox": [ + 304, + 517, + 545, + 615 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 304, + 517, + 545, + 615 + ], + "type": "text", + "content": " of images (see [B1]). We also report the performance using " + }, + { + "bbox": [ + 304, + 517, + 545, + 615 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 304, + 517, + 545, + 615 + ], + "type": "text", + "content": " images of the training set in Table 3. It can be seen that, despite CATRE use full CAMERA25 training set, our method outperforms it with " + }, + { + "bbox": [ + 304, + 517, + 545, + 615 + ], + "type": "inline_equation", + "content": "\\mathrm{IoU}_{75}" + }, + { + "bbox": [ + 304, + 517, + 545, + 615 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 517, + 545, + 615 + ], + "type": "inline_equation", + "content": "5^{\\circ}2\\mathrm{cm}" + }, + { + "bbox": [ + 304, + 517, + 545, + 615 + ], + "type": "text", + "content": " improved by " + }, + { + "bbox": [ + 304, + 517, + 545, + 615 + ], + "type": "inline_equation", + "content": "3.1\\%" + }, + { + "bbox": [ + 304, + 517, + 545, + 615 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 517, + 545, + 615 + ], + "type": "inline_equation", + "content": "3.6\\%" + }, + { + "bbox": [ + 304, + 517, + 545, + 615 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 622, + 488, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 622, + 488, + 635 + ], + "spans": [ + { + "bbox": [ + 305, + 622, + 488, + 635 + ], + "type": "text", + "content": "4.3. Comparison with state-of-the-arts" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "content": "REAL275. We conduct pose refinement on SPD [45] using the proposed approach and compare the resulting performance with the state-of-the-art category-level object pose estimation and refinement methods. As shown in Table 4, our method significantly improves the performance of SPD on all the metrics, with " + }, + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "inline_equation", + "content": "5^{\\circ}5\\mathrm{cm}" + }, + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "content": " enhanced" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10699" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 121, + 84, + 470, + 168 + ], + "blocks": [ + { + "bbox": [ + 189, + 71, + 403, + 81 + ], + "lines": [ + { + "bbox": [ + 189, + 71, + 403, + 81 + ], + "spans": [ + { + "bbox": [ + 189, + 71, + 403, + 81 + ], + "type": "text", + "content": "Table 3. The generalizability test on the CAMERA25 dataset." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 121, + 84, + 470, + 168 + ], + "lines": [ + { + "bbox": [ + 121, + 84, + 470, + 168 + ], + "spans": [ + { + "bbox": [ + 121, + 84, + 470, + 168 + ], + "type": "table", + "html": "
RowMethodTrain Data SizeIoU755°2cm5°5cm10°2cm10°5cm2cm5°
A0CATRE275K76.175.480.383.389.3--
B0CATRE5K63.266.472.379.487.488.873.3
B1Ours5K77.575.481.183.490.091.082.3
C0CATRE10K66.569.775.581.889.189.976.7
C1Ours10K79.277.984.083.890.592.085.4
", + "image_path": "5cfcde5a1af6fb3561c2c3eb03240d877abeec0e4da5c1fa47749dda64221d95.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 47, + 198, + 287, + 327 + ], + "blocks": [ + { + "bbox": [ + 70, + 186, + 263, + 196 + ], + "lines": [ + { + "bbox": [ + 70, + 186, + 263, + 196 + ], + "spans": [ + { + "bbox": [ + 70, + 186, + 263, + 196 + ], + "type": "text", + "content": "Table 4. Comparison with other methods on REAL275." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 198, + 287, + 327 + ], + "lines": [ + { + "bbox": [ + 47, + 198, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 47, + 198, + 287, + 327 + ], + "type": "table", + "html": "
MethodIoU755°2cm5°5cm10°2cm10°5cm
NOCS [49]9.47.210.013.825.2
DualPoseNet [24]30.829.335.950.066.8
CR-Net [50]33.227.834.347.260.8
SGPA [6]37.135.939.661.370.7
RBP-Pose [58]24.538.248.163.179.2
GPV-Pose [10]23.132.042.955.073.3
HS-Pose [61]39.146.555.268.682.7
SPD* [45]27.019.121.243.554.0
SPD*+CATRE [29]43.645.854.461.473.1
SPD*+Ours51.854.460.371.979.4
", + "image_path": "cee9af89f1e39cd5ded6e6b42e2326bfc3ed66791d4067de50935d3df7545e51.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 338, + 289, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 338, + 289, + 615 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 289, + 615 + ], + "type": "text", + "content": "by " + }, + { + "bbox": [ + 46, + 338, + 289, + 615 + ], + "type": "inline_equation", + "content": "39.1\\%" + }, + { + "bbox": [ + 46, + 338, + 289, + 615 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 338, + 289, + 615 + ], + "type": "inline_equation", + "content": "5^{\\circ}5\\mathrm{cm}" + }, + { + "bbox": [ + 46, + 338, + 289, + 615 + ], + "type": "text", + "content": " improved by " + }, + { + "bbox": [ + 46, + 338, + 289, + 615 + ], + "type": "inline_equation", + "content": "35.3\\%" + }, + { + "bbox": [ + 46, + 338, + 289, + 615 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 338, + 289, + 615 + ], + "type": "inline_equation", + "content": "\\mathrm{IoU}_{75}" + }, + { + "bbox": [ + 46, + 338, + 289, + 615 + ], + "type": "text", + "content": " enhanced by " + }, + { + "bbox": [ + 46, + 338, + 289, + 615 + ], + "type": "inline_equation", + "content": "24.8\\%" + }, + { + "bbox": [ + 46, + 338, + 289, + 615 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 338, + 289, + 615 + ], + "type": "inline_equation", + "content": "10^{\\circ}2\\mathrm{cm}" + }, + { + "bbox": [ + 46, + 338, + 289, + 615 + ], + "type": "text", + "content": " improved by " + }, + { + "bbox": [ + 46, + 338, + 289, + 615 + ], + "type": "inline_equation", + "content": "24.4\\%" + }, + { + "bbox": [ + 46, + 338, + 289, + 615 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 338, + 289, + 615 + ], + "type": "inline_equation", + "content": "10^{\\circ}5\\mathrm{cm}" + }, + { + "bbox": [ + 46, + 338, + 289, + 615 + ], + "type": "text", + "content": " improve by " + }, + { + "bbox": [ + 46, + 338, + 289, + 615 + ], + "type": "inline_equation", + "content": "25.4\\%" + }, + { + "bbox": [ + 46, + 338, + 289, + 615 + ], + "type": "text", + "content": ". In comparison to our baseline, CATRE, our proposed method demonstrates a substantial improvement across various performance metrics. Specifically, we observe a remarkable enhancement of " + }, + { + "bbox": [ + 46, + 338, + 289, + 615 + ], + "type": "inline_equation", + "content": "10.5\\%" + }, + { + "bbox": [ + 46, + 338, + 289, + 615 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 46, + 338, + 289, + 615 + ], + "type": "inline_equation", + "content": "10^{\\circ}2\\mathrm{cm}" + }, + { + "bbox": [ + 46, + 338, + 289, + 615 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 338, + 289, + 615 + ], + "type": "inline_equation", + "content": "8.6\\%" + }, + { + "bbox": [ + 46, + 338, + 289, + 615 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 46, + 338, + 289, + 615 + ], + "type": "inline_equation", + "content": "5^{\\circ}2\\mathrm{cm}" + }, + { + "bbox": [ + 46, + 338, + 289, + 615 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 338, + 289, + 615 + ], + "type": "inline_equation", + "content": "7.2\\%" + }, + { + "bbox": [ + 46, + 338, + 289, + 615 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 46, + 338, + 289, + 615 + ], + "type": "inline_equation", + "content": "\\mathrm{IoU}_{75}" + }, + { + "bbox": [ + 46, + 338, + 289, + 615 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 338, + 289, + 615 + ], + "type": "inline_equation", + "content": "6.3\\%" + }, + { + "bbox": [ + 46, + 338, + 289, + 615 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 46, + 338, + 289, + 615 + ], + "type": "inline_equation", + "content": "10^{\\circ}5\\mathrm{cm}" + }, + { + "bbox": [ + 46, + 338, + 289, + 615 + ], + "type": "text", + "content": ". In addition, compared with SOTA pose estimation methods, the pose estimation results of applying our proposed refinement method on SPD significantly outperformed these methods's results by a large margin. Specifically, the estimation results of applying our method on SPD rank top on 4 out of 5 metrics and rank second on the rest metric, and achieved a " + }, + { + "bbox": [ + 46, + 338, + 289, + 615 + ], + "type": "inline_equation", + "content": "7.9\\%" + }, + { + "bbox": [ + 46, + 338, + 289, + 615 + ], + "type": "text", + "content": " increase on the " + }, + { + "bbox": [ + 46, + 338, + 289, + 615 + ], + "type": "inline_equation", + "content": "5^{\\circ}2\\mathrm{cm}" + }, + { + "bbox": [ + 46, + 338, + 289, + 615 + ], + "type": "text", + "content": " metric and " + }, + { + "bbox": [ + 46, + 338, + 289, + 615 + ], + "type": "inline_equation", + "content": "5.1\\%" + }, + { + "bbox": [ + 46, + 338, + 289, + 615 + ], + "type": "text", + "content": " enhancement on " + }, + { + "bbox": [ + 46, + 338, + 289, + 615 + ], + "type": "inline_equation", + "content": "5^{\\circ}5\\mathrm{cm}" + }, + { + "bbox": [ + 46, + 338, + 289, + 615 + ], + "type": "text", + "content": ". It is worth noting that the purpose of this section is not to compare refinement and estimation methods, as they are designed to address different problems. Instead, our objective here is to demonstrate how our proposed refinement approach can improve the performance of existing pose estimation methods. Therefore, even though our refinement on HS-Pose produced better performance, as mentioned in the ablation study, we choose to refine weaker initial estimations to show the capability of our approach." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 621, + 288, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 621, + 288, + 658 + ], + "spans": [ + { + "bbox": [ + 46, + 621, + 288, + 658 + ], + "type": "text", + "content": "CAMERA25. Our method outperforms the state-of-the-art methods using only " + }, + { + "bbox": [ + 46, + 621, + 288, + 658 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 46, + 621, + 288, + 658 + ], + "type": "text", + "content": " of the training image set. For more details, please refer to the supplementary." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "Qualitative examples. We provide qualitative comparisons of the pose estimation results by SPD, CATRE on SPD, and our method on SPD in Fig. 4 and 5. As shown in Fig. 4, our method on SPD achieves the best size and pose" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 341, + 185, + 542, + 260 + ], + "blocks": [ + { + "bbox": [ + 310, + 198, + 338, + 207 + ], + "lines": [ + { + "bbox": [ + 310, + 198, + 338, + 207 + ], + "spans": [ + { + "bbox": [ + 310, + 198, + 338, + 207 + ], + "type": "text", + "content": "CATRE" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 341, + 185, + 542, + 260 + ], + "lines": [ + { + "bbox": [ + 341, + 185, + 542, + 260 + ], + "spans": [ + { + "bbox": [ + 341, + 185, + 542, + 260 + ], + "type": "image", + "image_path": "6e029e7926f2ec3140981561b68e0b5eb8de2f23237bac2089d17c2122c31d25.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 262, + 545, + 293 + ], + "lines": [ + { + "bbox": [ + 306, + 262, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 306, + 262, + 545, + 293 + ], + "type": "text", + "content": "Figure 5. Comparison of proposed (row #2) and baseline (row #1) methods) during a complete refinement iteration, both using SPD as initial estimation. The ground truth is represented by white lines." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 305, + 545, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 305, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 304, + 305, + 545, + 376 + ], + "type": "text", + "content": "estimations. In particular, by considering the first column of Fig. 4, all of the comparison methods struggle to estimate the orientation of the camera category. We also provide qualitative examples on the iteration process in Fig. 5. Our method demonstrates faster convergence and more accurate final result than CATRE." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 388, + 378, + 400 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 388, + 378, + 400 + ], + "spans": [ + { + "bbox": [ + 306, + 388, + 378, + 400 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 411, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 545, + 567 + ], + "type": "text", + "content": "In this work, we proposed a novel category-level object pose refinement method which targeted at addressing the challenge of shape variation. We shown that the geometric structural information can be aligned by our adaptive affine transformations. We also demonstrated that the crosscloud transformation mechanism can efficiently merges information from distinct point clouds. We further incorporated shape prior information and observed improvements in translation and size predictions. We verified that each of our technical components contributed meaningfully through extensive ablations. We believe our method sets a strong baseline for future study and opens up new possibilities to handling more complex shapes, i.e. articulated objects." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 582, + 403, + 595 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 582, + 403, + 595 + ], + "spans": [ + { + "bbox": [ + 306, + 582, + 403, + 595 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 602, + 545, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 602, + 545, + 703 + ], + "spans": [ + { + "bbox": [ + 304, + 602, + 545, + 703 + ], + "type": "text", + "content": "This research was supported in part by the MSIT (Ministry of Science and ICT), Korea, under the ITRC (Information Technology Research Center) support program (IITP-2024-2020-0-01789) supervised by the IITP (Institute for Information & Communications Technology Planning & Evaluation), National Natural Science Foundation of China under Grant No.62073159, and the Shenzhen Key Laboratory of Control Theory and Intelligent Systems, under grant No.ZDSYS20220330161800001." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10700" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Dingding Cai, Janne Heikkilä, and Esa Rahtu. Ove6d: Object viewpoint encoding for depth-based 6d object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6803-6813, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 202 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 202 + ], + "type": "text", + "content": "[2] Pedro Castro and Tae-Kyun Kim. Crt-6d: Fast 6d object pose estimation with cascaded refinement transformers. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 5746-5755, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 205, + 288, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 288, + 258 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 288, + 258 + ], + "type": "text", + "content": "[3] Dengsheng Chen, Jun Li, Zheng Wang, and Kai Xu. Learning canonical shape space for category-level 6d object pose and size estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 261, + 287, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 261, + 287, + 305 + ], + "spans": [ + { + "bbox": [ + 53, + 261, + 287, + 305 + ], + "type": "text", + "content": "[4] Hansheng Chen, Pichao Wang, Fan Wang, Wei Tian, Lu Xiong, and Hao Li. Epro-pnp: Generalized end-to-end probabilistic perspective-n-points for monocular object pose estimation, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 307, + 288, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 307, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 53, + 307, + 288, + 361 + ], + "type": "text", + "content": "[5] Hanzhi Chen, Fabian Manhardt, Nassir Navab, and Benjamin Busam. Texpose: Neural texture learning for self-supervised 6d object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4841-4852, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 364, + 287, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 364, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 53, + 364, + 287, + 408 + ], + "type": "text", + "content": "[6] Kai Chen and Qi Dou. SGPA: Structure-guided prior adaptation for category-level 6d object pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2773-2782, 2021. 2, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 410, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 410, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 53, + 410, + 287, + 464 + ], + "type": "text", + "content": "[7] Kai Chen, Stephen James, Congying Sui, Yun-Hui Liu, Pieter Abbeel, and Qi Dou. Stereopose: Category-level 6d transparent object pose estimation from stereo images via back-view nocs. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2855-2861, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 467, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 467, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 53, + 467, + 287, + 521 + ], + "type": "text", + "content": "[8] Wei Chen, Xi Jia, Hyung Jin Chang, Jinming Duan, and Ales Leonardis. G2L-Net: Global to Local Network for Real-Time 6D Pose Estimation With Embedding Vector Features. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 523, + 287, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 523, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 53, + 523, + 287, + 589 + ], + "type": "text", + "content": "[9] Wei Chen, Xi Jia, Hyung Jin Chang, Jinming Duan, Linlin Shen, and Ales Leonardis. Fs-net: Fast shape-based network for category-level 6d object pose estimation with decoupled rotation mechanism. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1581-1590, 2021. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 591, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 287, + 656 + ], + "type": "text", + "content": "[10] Yan Di, Ruida Zhang, Zhiqiang Lou, Fabian Manhardt, Xiangyang Ji, Nassir Navab, and Federico Tombari. GPV-Pose: Category-level object pose estimation via geometry-guided point-wise voting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6781-6791, 2022. 1, 2, 5, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[11] Yang Hai, Rui Song, Jiaojiao Li, Mathieu Salzmann, and Yinlin Hu. Rigidity-aware detection for 6d object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8927-8936, 2023. 1" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[12] Rasmus Laurvig Haugaard and Anders Glent Buch. Surfemb: Dense and continuous correspondence distributions for object pose estimation with learnt surface embeddings. CoRR, abs/2111.13489, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 118, + 547, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 173 + ], + "type": "text", + "content": "[13] Yisheng He, Wei Sun, Haibin Huang, Jianran Liu, Haoqiang Fan, and Jian Sun. Pvn3d: A deep point-wise 3d keypoints voting network for 6 dof pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 175, + 545, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 218 + ], + "type": "text", + "content": "[14] Stefan Hinterstoisser, Vincent Lepetit, Naresh Rajkumar, and Kurt Konolige. Going further with point pair features. In Computer Vision – ECCV 2016, pages 834–848. Springer International Publishing, 2016. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 220, + 547, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 547, + 262 + ], + "type": "text", + "content": "[15] Tomas Hodan, Daniel Barath, and Jiri Matas. Epos: Estimating 6d pose of objects with symmetries. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "type": "text", + "content": "[16] Muhammad Zubair Irshad, Sergey Zakharov, Rares Ambrus, Thomas Kollar, Zsolt Kira, and Adrien Gaidon. Shapo: Implicit representations for multi object shape appearance and pose optimization. 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 310, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 310, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 308, + 310, + 545, + 363 + ], + "type": "text", + "content": "[17] Shun Iwase, Xingyu Liu, Rawal Khirodkar, Rio Yokota, and Kris M Kitani. Repose: Fast 6d object pose refinement via deep texture rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3303-3312, 2021. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 365, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 545, + 430 + ], + "type": "text", + "content": "[18] Daniel Kappler, Franziska Meier, Jan Issac, Jim Mainprice, Cristina Garcia Cifuentes, Manuel Wüthrich, Vincent Berenz, Stefan Schaal, Nathan Ratliff, and Jeannette Bohg. Real-time perception meets reactive motion generation. IEEE Robotics and Automation Letters, 3(3):1864-1871, 2018. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 433, + 545, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 433, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 307, + 433, + 545, + 466 + ], + "type": "text", + "content": "[19] Nikunj Kothari, Misha Gupta, Leena Vachhani, and Hemendra Arya. Pose estimation for an autonomous vehicle using monocular vision. pages 424-431, 2017. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 468, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 468, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 468, + 545, + 510 + ], + "type": "text", + "content": "[20] Yann Labbe, Justin Carpentier, Mathieu Aubry, and Josef Sivic. CosyPose: Consistent multi-view multi-object 6D pose estimation. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 513, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 513, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 513, + 545, + 555 + ], + "type": "text", + "content": "[21] Taeyeop Lee, Byeong-Uk Lee, Inkyu Shin, Jaesung Choe, Ukcheol Shin, In So Kweon, and Kuk-Jin Yoon. UDA-COPE: unsupervised domain adaptation for category-level object pose estimation. CoRR, abs/2111.12580, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 557, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 557, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 308, + 557, + 545, + 600 + ], + "type": "text", + "content": "[22] Yi Li, Gu Wang, Xiangyang Ji, Yu Xiang, and Dieter Fox. Deepim: Deep iterative matching for 6d pose estimation. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 602, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 602, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 308, + 602, + 545, + 668 + ], + "type": "text", + "content": "[23] Haitao Lin, Zichang Liu, Chilam Cheang, Yanwei Fu, Guodong Guo, and Xiangyang Xue. Sar-net: Shape alignment and recovery network for category-level 6d object pose and size estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6707-6717, 2022. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "text", + "content": "[24] Jiehong Lin, Zewei Wei, Zhihao Li, Songcen Xu, Kui Jia, and Yuanqing Li. DualPoseNet: Category-level 6d object pose and size estimation using dual pose network with refined learning of pose consistency. In Proceedings of the" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10701" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 49, + 73, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "text", + "content": "IEEE/CVF International Conference on Computer Vision (ICCV), pages 3560-3569, 2021. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 97, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 97, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 49, + 97, + 287, + 140 + ], + "type": "text", + "content": "[25] Jiehong Lin, Zewei Wei, Changxing Ding, and Kui Jia. Category-level 6d object pose and size estimation using self-supervised deep prior deformation networks, 2022. 1, 2, 5, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 144, + 287, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 144, + 287, + 208 + ], + "spans": [ + { + "bbox": [ + 49, + 144, + 287, + 208 + ], + "type": "text", + "content": "[26] Zhi-Hao Lin, Sheng-Yu Huang, and Yu-Chiang Frank Wang. Convolution in the cloud: Learning deformable kernels in 3d graph convolution networks for point cloud analysis. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1797-1806, 2020. 1, 2, 3, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 211, + 287, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 211, + 287, + 265 + ], + "spans": [ + { + "bbox": [ + 49, + 211, + 287, + 265 + ], + "type": "text", + "content": "[27] Zheng Linfang, Leonardis Ales, Tse Tze Ho, Elden, Horanyi Nora, Chen Hua, Zhang Wei, and Chang Hyung Jin. Tp-ae: Temporally primed 6d object pose tracking with auto-encoders. In 2022 IEEE International Conference on Robotics and Automation (ICRA), 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 268, + 287, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 268, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 49, + 268, + 287, + 300 + ], + "type": "text", + "content": "[28] Liyuan Liu, Haoming Jiang, Pengcheng He, Weizhu Chen, Xiaodong Liu, Jianfeng Gao, and Jiawei Han. On the variance of the adaptive learning rate and beyond, 2019. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 303, + 287, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 303, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 49, + 303, + 287, + 346 + ], + "type": "text", + "content": "[29] Xingyu Liu, Gu Wang, Yi Li, and Xiangyang Ji. CATRE: Iterative point clouds alignment for category-level object pose refinement. In European Conference on Computer Vision (ECCV), pages 499-516. Springer, 2022. 1, 2, 3, 5, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 349, + 286, + 359 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 349, + 286, + 359 + ], + "spans": [ + { + "bbox": [ + 49, + 349, + 286, + 359 + ], + "type": "text", + "content": "[30] Eitan Marder-Eppstein. Project tango. pages 25-25, 2016. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 361, + 287, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 361, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 49, + 361, + 287, + 426 + ], + "type": "text", + "content": "[31] Nathaniel Merrill, Yuliang Guo, Xingxing Zuo, Xinyu Huang, Stefan Leutenegger, Xi Peng, Liu Ren, and Guoquan Huang. Symmetry and uncertainty-aware object slam for 6 dof object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14901-14910, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 430, + 287, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 430, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 49, + 430, + 287, + 483 + ], + "type": "text", + "content": "[32] Andrew S. Morgan, Bowen Wen, Junchi Liang, Abdeslam Boularias, Aaron M. Dollar, and Kostas Bekris. Vision-driven Compliant Manipulation for Reliable; High-Precision Assembly Tasks. In Proceedings of Robotics: Science and Systems, Virtual, 2021. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 487, + 287, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 487, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 49, + 487, + 287, + 529 + ], + "type": "text", + "content": "[33] Andrew Nee, S K Ong, George Chryssolouris, and Dimitris Mourtzis. Augmented reality applications in design and manufacturing. CIRP Annals - Manufacturing Technology, 61:657-679, 2012. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 533, + 287, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 533, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 49, + 533, + 287, + 574 + ], + "type": "text", + "content": "[34] Van Nguyen Nguyen, Yinlin Hu, Yang Xiao, Mathieu Salzmann, and Vincent Lepetit. Templates for 3d object pose estimation revisited: Generalization to new objects and robustness to occlusions, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 578, + 287, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 578, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 49, + 578, + 287, + 620 + ], + "type": "text", + "content": "[35] Charles R. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 624, + 287, + 676 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 624, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 49, + 624, + 287, + 676 + ], + "type": "text", + "content": "[36] Mahdi Rad and Vincent Lepetit. BB8: A Scalable, Accurate, Robust to Partial Occlusion Method for Predicting the 3D Poses of Challenging Objects Without Using Depth. In IEEE International Conference on Computer Vision (ICCV), 2017." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 681, + 287, + 713 + ], + "type": "text", + "content": "[37] Aleksandr Segal, Dirk Haehnel, and Sebastian Thrun. Generalized-icp. In Robotics: science and systems (RSS), page 435. Seattle, WA, 2009. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "type": "text", + "content": "[38] Ivan Shugurov, Fu Li, Benjamin Busam, and Slobodan Ilic. Osop: A multi-stage one shot object pose estimation framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6835-6844, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "text", + "content": "[39] Yongzhi Su, Jason Rambach, Nareg Minaskan, Paul Lesur, Alain Pagani, and Didier Stricker. Deep multi-state object pose estimation for augmented reality assembly. In 2019 IEEE International Symposium on Mixed and Augmented Reality Adjunct (ISMAR-Adjunct), pages 222–227, 2019. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 185, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 185, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 308, + 185, + 545, + 227 + ], + "type": "text", + "content": "[40] Yongzhi Su, Mahdi Saleh, Torben Fetzer, Jason Rambach, Nassir Navab, Benjamin Busam, Didier Stricker, and Federico Tombari. Zebrapose: Coarse to fine surface encoding for 6 dof object pose estimation, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 228, + 545, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 292 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 292 + ], + "type": "text", + "content": "[41] Jiaming Sun, Zihao Wang, Siyu Zhang, Xingyi He, Hongcheng Zhao, Guofeng Zhang, and Xiaowei Zhou. Onepose: One-shot object pose estimation without cad models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6825-6834, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 295, + 545, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 295, + 545, + 336 + ], + "spans": [ + { + "bbox": [ + 308, + 295, + 545, + 336 + ], + "type": "text", + "content": "[42] Martin Sundermeyer, Maximilian Durner, En Yen Puang, Zoltan-Csaba Marton, Narunas Vaskevicius, Kai O. Arras, and Rudolph Triebel. Multi-path learning for object pose estimation across domains, 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 338, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 545, + 381 + ], + "type": "text", + "content": "[43] Martin Sundermeyer, Zoltan Marton, Maximilian Durner, and Rudolph Triebel. Augmented Autoencoders: Implicit 3D Orientation Learning for 6D Object Detection. International Journal of Computer Vision (IJCV), 128, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 384, + 545, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 384, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 308, + 384, + 545, + 414 + ], + "type": "text", + "content": "[44] Bugra Tekin, Sudipta Sinha, and Pascal Fua. Real-time seamless single shot 6d object pose prediction. pages 292-301, 2018. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 416, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 416, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 308, + 416, + 545, + 459 + ], + "type": "text", + "content": "[45] Meng Tian, Marcelo H Ang, and Gim Hee Lee. Shape prior deformation for categorical 6d object pose and size estimation. In European Conference on Computer Vision (ECCV), pages 530-546. Springer, 2020. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 460, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 502 + ], + "type": "text", + "content": "[46] Jonathan Tremblay, Thang To, Balakumar Sundaralingam, Yu Xiang, Dieter Fox, and Stan Birchfield. Deep object pose estimation for semantic robotic grasping of household objects, 2018. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 505, + 545, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 505, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 308, + 505, + 545, + 535 + ], + "type": "text", + "content": "[47] Joel Vidal, Chyi-Yeu Lin, and Robert Martí. 6d pose estimation using an improved method based on point pair features, 2018. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 537, + 545, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 545, + 591 + ], + "type": "text", + "content": "[48] Chen Wang, Danfei Xu, Yuke Zhu, Roberto Martin-Martin, Cewu Lu, Li Fei-Fei, and Silvio Savarese. DenseFusion: 6D Object Pose Estimation by Iterative Dense Fusion. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 593, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 593, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 593, + 545, + 657 + ], + "type": "text", + "content": "[49] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J Guibas. Normalized object coordinate space for category-level 6d object pose and size estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2642-2651, 2019. 2, 5, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 659, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 659, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 659, + 545, + 712 + ], + "type": "text", + "content": "[50] Jiaze Wang, Kai Chen, and Qi Dou. Category-level 6d object pose estimation via cascaded relation and recurrent reconstruction networks. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4807-4814. IEEE, 2021. 8" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "10702" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 688 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[51] Bowen Wen, Chaitanya Mitash, Baozhang Ren, and Kostas Bekris. se(3)-TrackNet: Data-driven 6D Pose Tracking by Calibrating Image Residuals in Synthetic Domains. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2020. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 183 + ], + "type": "text", + "content": "[52] Bowen Wen, Wenzhao Lian, Kostas Bekris, and Stefan Schaal. You Only Demonstrate Once: Category-Level Manipulation from Single Visual Demonstration. In Proceedings of Robotics: Science and Systems, New York City, NY, USA, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 185, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 288, + 239 + ], + "type": "text", + "content": "[53] Heng Yang and Marco Pavone. Object pose estimation with statistical guarantees: Conformal keypoint detection and geometric uncertainty propagation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8947-8958, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 240, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 240, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 240, + 287, + 274 + ], + "type": "text", + "content": "[54] Hongwei Yong, Jianqiang Huang, Xiansheng Hua, and Lei Zhang. Gradient centralization: A new optimization technique for deep neural networks, 2020. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 274, + 287, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 274, + 287, + 317 + ], + "spans": [ + { + "bbox": [ + 48, + 274, + 287, + 317 + ], + "type": "text", + "content": "[55] Sergey Zakharov, Ivan Shugurov, and Slobodan Ilic. DPOD: 6D Pose Object Detector and Refiner. In The IEEE International Conference on Computer Vision (ICCV), 2019. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 319, + 287, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 319, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 319, + 287, + 374 + ], + "type": "text", + "content": "[56] Kun Zhang, Chen Wang, Hua Chen, Jia Pan, Michael Yu Wang, and Wei Zhang. Vision-based six-dimensional peg-in-hole for practical connector insertion. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 1771-1777, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 375, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 375, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 48, + 375, + 287, + 407 + ], + "type": "text", + "content": "[57] Michael R. Zhang, James Lucas, Geoffrey Hinton, and Jimmy Ba. Lookahead optimizer: k steps forward, 1 step back, 2019. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 408, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 408, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 408, + 287, + 453 + ], + "type": "text", + "content": "[58] Ruida Zhang, Yan Di, Zhiqiang Lou, Fabian Manhardt, Federico Tombari, and Xiangyang Ji. RBP-Pose: Residual bounding box projection for category-level pose estimation, 2022. 1, 2, 5, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 453, + 287, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 453, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 48, + 453, + 287, + 496 + ], + "type": "text", + "content": "[59] Ruida Zhang, Yan Di, Fabian Manhardt, Federico Tombari, and Xiangyang Ji. SSP-Pose: Symmetry-aware shape prior deformation for direct category-level object pose estimation, 2022. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 498, + 287, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 498, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 48, + 498, + 287, + 553 + ], + "type": "text", + "content": "[60] Zhongqun Zhang, Wei Chen, Linfang Zheng, Ales Leonardis, and Hyung Jin Chang. Trans6d: Transformer-based 6d object pose estimation and refinement. In Computer Vision - ECCV 2022 Workshops, pages 112-128, Cham, 2023. Springer Nature Switzerland. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 554, + 288, + 630 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 554, + 288, + 630 + ], + "spans": [ + { + "bbox": [ + 48, + 554, + 288, + 630 + ], + "type": "text", + "content": "[61] Linfang Zheng, Chen Wang, Yinghan Sun, Esha Dasgupta, Hua Chen, Aleš Leonardis, Wei Zhang, and Hyung Jin Chang. HS-Pose: Hybrid scope feature extraction for category-level object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17163-17173, 2023. 1, 2, 3, 5, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 632, + 287, + 688 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 632, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 48, + 632, + 287, + 688 + ], + "type": "text", + "content": "[62] Jun Zhou, Kai Chen, Linlin Xu, Qi Dou, and Jing Qin. Deep fusion transformer network with weighted vector-wise keypoints voting for robust 6d object pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13967-13977, 2023. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10703" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file